diff --git a/.gitattributes b/.gitattributes index 250c692e90754f85251ec406700accaea8bbb908..ca8458f060bfbf41c824a962ec1b8fb4d6f3dbe7 100644 --- a/.gitattributes +++ b/.gitattributes @@ -5448,3 +5448,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text 2025/FADA_[[:space:]]Fast[[:space:]]Diffusion[[:space:]]Avatar[[:space:]]Synthesis[[:space:]]with[[:space:]]Mixed-Supervised[[:space:]]Multi-CFG[[:space:]]Distillation/eae624fb-184e-49bf-afbc-d98a98aeb9b5_origin.pdf filter=lfs diff=lfs merge=lfs -text 2025/FADE_[[:space:]]Frequency-Aware[[:space:]]Diffusion[[:space:]]Model[[:space:]]Factorization[[:space:]]for[[:space:]]Video[[:space:]]Editing/37abe74e-aca7-48e5-911c-4b7009fcf3ba_origin.pdf filter=lfs diff=lfs merge=lfs -text 2025/FALCON_[[:space:]]Fairness[[:space:]]Learning[[:space:]]via[[:space:]]Contrastive[[:space:]]Attention[[:space:]]Approach[[:space:]]to[[:space:]]Continual[[:space:]]Semantic[[:space:]]Scene[[:space:]]Understanding/2e34452d-f7d4-4d46-8b4f-8391c1e12974_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/HoVLE_[[:space:]]Unleashing[[:space:]]the[[:space:]]Power[[:space:]]of[[:space:]]Monolithic[[:space:]]Vision-Language[[:space:]]Models[[:space:]]with[[:space:]]Holistic[[:space:]]Vision-Language[[:space:]]Embedding/126721ba-0c45-4655-8cc5-9017b39955e6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Holmes-VAU_[[:space:]]Towards[[:space:]]Long-term[[:space:]]Video[[:space:]]Anomaly[[:space:]]Understanding[[:space:]]at[[:space:]]Any[[:space:]]Granularity/8b3a9750-2575-4fe7-bb78-a5c96dc5aac4_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/HomoGen_[[:space:]]Enhanced[[:space:]]Video[[:space:]]Inpainting[[:space:]]via[[:space:]]Homography[[:space:]]Propagation[[:space:]]and[[:space:]]Diffusion/c0acdc17-6ead-4f9d-8631-4f1535650ac6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Homogeneous[[:space:]]Dynamics[[:space:]]Space[[:space:]]for[[:space:]]Heterogeneous[[:space:]]Humans/a999c957-db24-46d0-83c3-e6da7809465e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Horizon-GS_[[:space:]]Unified[[:space:]]3D[[:space:]]Gaussian[[:space:]]Splatting[[:space:]]for[[:space:]]Large-Scale[[:space:]]Aerial-to-Ground[[:space:]]Scenes/65e9755f-fe8e-4a32-88df-f20ee7cf9c9c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/HotSpot_[[:space:]]Signed[[:space:]]Distance[[:space:]]Function[[:space:]]Optimization[[:space:]]with[[:space:]]an[[:space:]]Asymptotically[[:space:]]Sufficient[[:space:]]Condition/733f46a7-f78c-4e53-bba4-ead7eb705bb7_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/How[[:space:]]Do[[:space:]]I[[:space:]]Do[[:space:]]That_[[:space:]]Synthesizing[[:space:]]3D[[:space:]]Hand[[:space:]]Motion[[:space:]]and[[:space:]]Contacts[[:space:]]for[[:space:]]Everyday[[:space:]]Interactions/4f85bf33-0c00-4386-bb62-720463efe772_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/How[[:space:]]to[[:space:]]Merge[[:space:]]Your[[:space:]]Multimodal[[:space:]]Models[[:space:]]Over[[:space:]]Time_/76e2370a-2b50-436a-9bdc-9b8af8c8e7ca_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/HuMoCon_[[:space:]]Concept[[:space:]]Discovery[[:space:]]for[[:space:]]Human[[:space:]]Motion[[:space:]]Understanding/b5dd425e-4f3e-4b2d-9c88-36e4f0bafd85_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/HuPerFlow_[[:space:]]A[[:space:]]Comprehensive[[:space:]]Benchmark[[:space:]]for[[:space:]]Human[[:space:]]vs.[[:space:]]Machine[[:space:]]Motion[[:space:]]Estimation[[:space:]]Comparison/5a790d2d-155e-42dc-bc07-eb0a483c9c11_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Human[[:space:]]Motion[[:space:]]Instruction[[:space:]]Tuning/8c08c8b0-9aae-431a-b68e-699a5ddd5c32_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Human-centered[[:space:]]Interactive[[:space:]]Learning[[:space:]]via[[:space:]]MLLMs[[:space:]]for[[:space:]]Text-to-Image[[:space:]]Person[[:space:]]Re-identification/5cc75be5-e122-4a4f-ba40-7450ab436d6a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/HumanDreamer_[[:space:]]Generating[[:space:]]Controllable[[:space:]]Human-Motion[[:space:]]Videos[[:space:]]via[[:space:]]Decoupled[[:space:]]Generation/8f919b44-e98e-44f2-a774-268fc39345bc_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/HumanMM_[[:space:]]Global[[:space:]]Human[[:space:]]Motion[[:space:]]Recovery[[:space:]]from[[:space:]]Multi-shot[[:space:]]Videos/8ae05aea-0260-4aa4-bbb0-352a7f188bbe_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/HumanRig_[[:space:]]Learning[[:space:]]Automatic[[:space:]]Rigging[[:space:]]for[[:space:]]Humanoid[[:space:]]Character[[:space:]]in[[:space:]]a[[:space:]]Large[[:space:]]Scale[[:space:]]Dataset/a5721584-3cc2-4564-a9c0-b5eefdb73fbc_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/HunyuanPortrait_[[:space:]]Implicit[[:space:]]Condition[[:space:]]Control[[:space:]]for[[:space:]]Enhanced[[:space:]]Portrait[[:space:]]Animation/a6cc1e88-c49b-4877-897d-6a6632aadecd_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Hybrid[[:space:]]Concept[[:space:]]Bottleneck[[:space:]]Models/7a729be8-7e97-4b64-9b6a-2c339ff64815_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Hybrid[[:space:]]Global-Local[[:space:]]Representation[[:space:]]with[[:space:]]Augmented[[:space:]]Spatial[[:space:]]Guidance[[:space:]]for[[:space:]]Zero-Shot[[:space:]]Referring[[:space:]]Image[[:space:]]Segmentation/9c88ce4f-9647-4320-bdd2-f6fc8a88e98e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Hybrid[[:space:]]Reciprocal[[:space:]]Transformer[[:space:]]with[[:space:]]Triplet[[:space:]]Feature[[:space:]]Alignment[[:space:]]for[[:space:]]Scene[[:space:]]Graph[[:space:]]Generation/11e7a635-4306-4d78-8190-6be9b1f3acac_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Hybrid-Level[[:space:]]Instruction[[:space:]]Injection[[:space:]]for[[:space:]]Video[[:space:]]Token[[:space:]]Compression[[:space:]]in[[:space:]]Multi-modal[[:space:]]Large[[:space:]]Language[[:space:]]Models/a0ce533d-bc87-47ce-aab6-707ed1ae09bc_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/HybridGS_[[:space:]]Decoupling[[:space:]]Transients[[:space:]]and[[:space:]]Statics[[:space:]]with[[:space:]]2D[[:space:]]and[[:space:]]3D[[:space:]]Gaussian[[:space:]]Splatting/347531dd-b02e-491d-be81-0620c6e2bfcd_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/HybridMQA_[[:space:]]Exploring[[:space:]]Geometry-Texture[[:space:]]Interactions[[:space:]]for[[:space:]]Colored[[:space:]]Mesh[[:space:]]Quality[[:space:]]Assessment/a45af5e0-e0d7-4402-9367-2386d583248b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/HyperFree_[[:space:]]A[[:space:]]Channel-adaptive[[:space:]]and[[:space:]]Tuning-free[[:space:]]Foundation[[:space:]]Model[[:space:]]for[[:space:]]Hyperspectral[[:space:]]Remote[[:space:]]Sensing[[:space:]]Imagery/9236da20-05f2-4327-b226-ca754d2dead5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/HyperGLM_[[:space:]]HyperGraph[[:space:]]for[[:space:]]Video[[:space:]]Scene[[:space:]]Graph[[:space:]]Generation[[:space:]]and[[:space:]]Anticipation/07619f38-6a65-46d1-ae2a-3885ad0e2f6d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/HyperGS_[[:space:]]Hyperspectral[[:space:]]3D[[:space:]]Gaussian[[:space:]]Splatting/8e6d9c3d-45db-41da-98c2-e1299df7de27_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/HyperLoRA_[[:space:]]Parameter-Efficient[[:space:]]Adaptive[[:space:]]Generation[[:space:]]for[[:space:]]Portrait[[:space:]]Synthesis/0dcd811d-bb8b-4fcc-9377-9711c569d3e6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/HyperNVD_[[:space:]]Accelerating[[:space:]]Neural[[:space:]]Video[[:space:]]Decomposition[[:space:]]via[[:space:]]Hypernetworks/f624369a-7a34-4be6-8ddf-7450e2fca80c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/HyperNet[[:space:]]Fields_[[:space:]]Efficiently[[:space:]]Training[[:space:]]Hypernetworks[[:space:]]without[[:space:]]Ground[[:space:]]Truth[[:space:]]by[[:space:]]Learning[[:space:]]Weight[[:space:]]Trajectories/68fd503f-6b57-47fa-8aca-593407ee13b9_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/HyperPose_[[:space:]]Hypernetwork-Infused[[:space:]]Camera[[:space:]]Pose[[:space:]]Localization[[:space:]]and[[:space:]]an[[:space:]]Extended[[:space:]]Cambridge[[:space:]]Landmarks[[:space:]]Dataset/dd4d95b8-93a1-4dc9-acfe-357b62a866ca_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/HyperSeg_[[:space:]]Hybrid[[:space:]]Segmentation[[:space:]]Assistant[[:space:]]with[[:space:]]Fine-grained[[:space:]]Visual[[:space:]]Perceiver/d3d1df0c-7574-4e1e-b447-81df1bf28440_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Hyperbolic[[:space:]]Category[[:space:]]Discovery/103e7fd4-0c6b-413e-8aae-04fa73b4770b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Hyperbolic[[:space:]]Safety-Aware[[:space:]]Vision-Language[[:space:]]Models/5ca9a6d6-45f1-4b74-b091-1c79b0b891b6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Hyperbolic[[:space:]]Uncertainty-Aware[[:space:]]Few-Shot[[:space:]]Incremental[[:space:]]Point[[:space:]]Cloud[[:space:]]Segmentation/1247e066-a517-4ffe-b97e-a9bff43b89c3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Hyperdimensional[[:space:]]Uncertainty[[:space:]]Quantification[[:space:]]for[[:space:]]Multimodal[[:space:]]Uncertainty[[:space:]]Fusion[[:space:]]in[[:space:]]Autonomous[[:space:]]Vehicles[[:space:]]Perception/8d8555dd-a579-45c4-ad75-fee4bbfad5af_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Hypergraph[[:space:]]Vision[[:space:]]Transformers_[[:space:]]Images[[:space:]]are[[:space:]]More[[:space:]]than[[:space:]]Nodes,[[:space:]]More[[:space:]]than[[:space:]]Edges/b3c66b72-bf5c-47e0-a296-e834f5e49c0f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Hyperspectral[[:space:]]Pansharpening[[:space:]]via[[:space:]]Diffusion[[:space:]]Models[[:space:]]with[[:space:]]Iteratively[[:space:]]Zero-Shot[[:space:]]Guidance/f79e03e7-ea12-4e70-879c-79258b8c5e82_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/I2VGuard_[[:space:]]Safeguarding[[:space:]]Images[[:space:]]against[[:space:]]Misuse[[:space:]]in[[:space:]]Diffusion-based[[:space:]]Image-to-Video[[:space:]]Models/612ebd29-cda5-46e3-b035-7c7393e6f819_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/IAAO_[[:space:]]Interactive[[:space:]]Affordance[[:space:]]Learning[[:space:]]for[[:space:]]Articulated[[:space:]]Objects[[:space:]]in[[:space:]]3D[[:space:]]Environments/c94d31b4-1877-4bc2-8b0f-9d4d8e45806a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/ICE_[[:space:]]Intrinsic[[:space:]]Concept[[:space:]]Extraction[[:space:]]from[[:space:]]a[[:space:]]Single[[:space:]]Image[[:space:]]via[[:space:]]Diffusion[[:space:]]Models/e9edff49-034e-4138-82ed-91d11ac96969_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/ICP_[[:space:]]Immediate[[:space:]]Compensation[[:space:]]Pruning[[:space:]]for[[:space:]]Mid-to-high[[:space:]]Sparsity/9ac0c560-9a5b-4372-9074-16698b23f1e1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/ICT_[[:space:]]Image-Object[[:space:]]Cross-Level[[:space:]]Trusted[[:space:]]Intervention[[:space:]]for[[:space:]]Mitigating[[:space:]]Object[[:space:]]Hallucination[[:space:]]in[[:space:]]Large[[:space:]]Vision-Language[[:space:]]Models/a90828a4-8b1b-4072-b5af-bed317d47267_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/ID-Patch_[[:space:]]Robust[[:space:]]ID[[:space:]]Association[[:space:]]for[[:space:]]Group[[:space:]]Photo[[:space:]]Personalization/a37d3ef3-36c1-46f2-886c-58bf6589bcc4_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/IDEA-Bench_[[:space:]]How[[:space:]]Far[[:space:]]are[[:space:]]Generative[[:space:]]Models[[:space:]]from[[:space:]]Professional[[:space:]]Designing_/970924c3-54e6-4027-b252-42ee55ce3b43_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/IDEA_[[:space:]]Inverted[[:space:]]Text[[:space:]]with[[:space:]]Cooperative[[:space:]]Deformable[[:space:]]Aggregation[[:space:]]for[[:space:]]Multi-modal[[:space:]]Object[[:space:]]Re-Identification/b7176393-d8d0-4ced-ba0c-43acd09bc650_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/IDOL_[[:space:]]Instant[[:space:]]Photorealistic[[:space:]]3D[[:space:]]Human[[:space:]]Creation[[:space:]]from[[:space:]]a[[:space:]]Single[[:space:]]Image/c28e8ffb-3109-4ced-b9e1-553dbeace189_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/IDProtector_[[:space:]]An[[:space:]]Adversarial[[:space:]]Noise[[:space:]]Encoder[[:space:]]to[[:space:]]Protect[[:space:]]Against[[:space:]]ID-Preserving[[:space:]]Image[[:space:]]Generation/cbe88f8c-13ab-4200-b1b9-bb453b62ad1d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/ILIAS_[[:space:]]Instance-Level[[:space:]]Image[[:space:]]retrieval[[:space:]]At[[:space:]]Scale/bd262cec-535c-47b6-b08f-5430c3f91bd5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/IM-Portrait_[[:space:]]Learning[[:space:]]3D-aware[[:space:]]Video[[:space:]]Diffusion[[:space:]]for[[:space:]]Photorealistic[[:space:]]Talking[[:space:]]Heads[[:space:]]from[[:space:]]Monocular[[:space:]]VideosC/d89a7508-cd12-421c-9424-93ad28adceb7_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/IM-Zero_[[:space:]]Instance-level[[:space:]]Motion[[:space:]]Controllable[[:space:]]Video[[:space:]]Generation[[:space:]]in[[:space:]]a[[:space:]]Zero-shot[[:space:]]Manner/a669fc0a-17c9-4001-97e0-7737b0fc24de_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/IMFine_[[:space:]]3D[[:space:]]Inpainting[[:space:]]via[[:space:]]Geometry-guided[[:space:]]Multi-view[[:space:]]Refinement/0becb03e-65d1-4a50-a09c-58afb744258c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/INFP_[[:space:]]Audio-Driven[[:space:]]Interactive[[:space:]]Head[[:space:]]Generation[[:space:]]in[[:space:]]Dyadic[[:space:]]Conversations/4440ddb1-10e9-45a7-942b-205ac25fbaaa_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/IRGS_[[:space:]]Inter-Reflective[[:space:]]Gaussian[[:space:]]Splatting[[:space:]]with[[:space:]]2D[[:space:]]Gaussian[[:space:]]Ray[[:space:]]Tracing/ff8cd318-c293-446d-b3d4-b22f9fc186ef_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/IRIS_[[:space:]]Inverse[[:space:]]Rendering[[:space:]]of[[:space:]]Indoor[[:space:]]Scenes[[:space:]]from[[:space:]]Low[[:space:]]Dynamic[[:space:]]Range[[:space:]]Images/53c93036-c714-4f00-9960-2f72fce37288_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/ITA-MDT_[[:space:]]Image-Timestep-Adaptive[[:space:]]Masked[[:space:]]Diffusion[[:space:]]Transformer[[:space:]]Framework[[:space:]]for[[:space:]]Image-Based[[:space:]]Virtual[[:space:]]Try-On/e6b13099-74a4-4157-8bc0-2e7511ad889f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/IceDiff_[[:space:]]High[[:space:]]Resolution[[:space:]]and[[:space:]]High-Quality[[:space:]]Arctic[[:space:]]Sea[[:space:]]Ice[[:space:]]Forecasting[[:space:]]with[[:space:]]Generative[[:space:]]Diffusion[[:space:]]Prior/3f2f5e81-8b45-4fda-b37e-2ea2e3413d80_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Identifying[[:space:]]and[[:space:]]Mitigating[[:space:]]Position[[:space:]]Bias[[:space:]]of[[:space:]]Multi-image[[:space:]]Vision-Language[[:space:]]Models/17ea1e19-b848-462e-8296-051308ed2dd4_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Identifying[[:space:]]and[[:space:]]Mitigating[[:space:]]Spurious[[:space:]]Correlation[[:space:]]in[[:space:]]Multi-Task[[:space:]]Learning/67faa86d-7895-441b-b372-45b9fee04763_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Identity-Clothing[[:space:]]Similarity[[:space:]]Modeling[[:space:]]for[[:space:]]Unsupervised[[:space:]]Clothing[[:space:]]Change[[:space:]]Person[[:space:]]Re-Identification/42258662-090f-4600-a2a7-5bf4fc16321d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Identity-Preserving[[:space:]]Text-to-Video[[:space:]]Generation[[:space:]]by[[:space:]]Frequency[[:space:]]Decomposition/16b4d9e1-d6f9-4df5-9c6a-3949d63cad2d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Identity-preserving[[:space:]]Distillation[[:space:]]Sampling[[:space:]]by[[:space:]]Fixed-Point[[:space:]]Iterator/8dc89aa9-e649-490d-ba0b-e87a936deb98_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Illumination[[:space:]]Spectrum[[:space:]]Estimation[[:space:]]for[[:space:]]Multispectral[[:space:]]Images[[:space:]]via[[:space:]]Surface[[:space:]]Reflectance[[:space:]]Modeling[[:space:]]and[[:space:]]Spatial-Spectral[[:space:]]Feature[[:space:]]Generation/afff26f0-40a0-4b11-8350-630ae1c85fad_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/ImViD_[[:space:]]Immersive[[:space:]]Volumetric[[:space:]]Videos[[:space:]]for[[:space:]]Enhanced[[:space:]]VR[[:space:]]Engagement/6da376dc-6436-436c-b26c-23ba201328ef_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Image[[:space:]]Generation[[:space:]]Diversity[[:space:]]Issues[[:space:]]and[[:space:]]How[[:space:]]to[[:space:]]Tame[[:space:]]Them/7493d326-112b-4a6d-8c2c-7e4449ad1384_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Image[[:space:]]Over[[:space:]]Text_[[:space:]]Transforming[[:space:]]Formula[[:space:]]Recognition[[:space:]]Evaluation[[:space:]]with[[:space:]]Character[[:space:]]Detection[[:space:]]Matching/e8290078-d6a1-49c9-b28d-cb4de981b7cd_origin.pdf filter=lfs diff=lfs merge=lfs -text diff --git a/2025/HoVLE_ Unleashing the Power of Monolithic Vision-Language Models with Holistic Vision-Language Embedding/126721ba-0c45-4655-8cc5-9017b39955e6_content_list.json b/2025/HoVLE_ Unleashing the Power of Monolithic Vision-Language Models with Holistic Vision-Language Embedding/126721ba-0c45-4655-8cc5-9017b39955e6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..30d2202172d2b91edf4872c67695cdb28531fa76 --- /dev/null +++ b/2025/HoVLE_ Unleashing the Power of Monolithic Vision-Language Models with Holistic Vision-Language Embedding/126721ba-0c45-4655-8cc5-9017b39955e6_content_list.json @@ -0,0 +1,1653 @@ +[ + { + "type": "text", + "text": "HoVLE: Unleashing the Power of Monolithic Vision-Language Models with Holistic Vision-Language Embedding", + "text_level": 1, + "bbox": [ + 117, + 128, + 879, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chenxin Tao $^{1,3*}$ , Shiqian Su $^{1,3*}$ , Xizhou Zhu $^{1,2*}$ , Chenyu Zhang $^{1,2}$ , Zhe Chen $^{4,2}$ , Jiawen Liu $^{6}$ , Wenhai Wang $^{5,2}$ , Lewei Lu $^{3}$ , Gao Huang $^{1}$ , Yu Qiao $^{2}$ , Jifeng Dai $^{1,2}$", + "bbox": [ + 125, + 199, + 872, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Tsinghua University $^{2}$ Shanghai Artificial Intelligence Laboratory", + "bbox": [ + 269, + 246, + 727, + 263 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ SenseTime Research $^{4}$ Nanjing University", + "bbox": [ + 349, + 263, + 648, + 277 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "5The Chinese University of Hong Kong 6 Johns Hopkins University", + "bbox": [ + 267, + 277, + 728, + 294 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Model: https://huggingface.co/OpenGVLab/HoVLE", + "bbox": [ + 303, + 296, + 732, + 311 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 328, + 326, + 344 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The rapid advance of Large Language Models (LLMs) has catalyzed the development of Vision-Language Models (VLMs). Monolithic VLMs, which avoid modality-specific encoders, offer a promising alternative to the compositional ones but face the challenge of inferior performance. Most existing monolithic VLMs require tuning pre-trained LLMs to acquire vision abilities, which may degrade their language capabilities. To address this dilemma, this paper presents a novel high-performance monolithic VLM named HoVLE. We note that LLMs have been shown to be capable of interpreting images when image embeddings are aligned with text embeddings. The challenge for current monolithic VLMs actually lies in the lack of a holistic embedding module for both vision and language inputs. Therefore, HoVLE introduces a holistic embedding module that converts visual and textual inputs into a shared space, allowing LLMs to process images in the same way as texts. Furthermore, a multi-stage training strategy is carefully designed to empower the holistic embedding module. It is first trained to distill visual features from a pre-trained vision encoder and text embeddings from the LLM, enabling largescale training with unpaired random images and text tokens. The whole model further undergoes next-token prediction on multi-modal data to align the embeddings. Finally, an instruction-tuning stage is incorporated. Our experiments show that HoVLE achieves performance close to leading compositional models on various benchmarks, outperforming previous monolithic models by a large margin.", + "bbox": [ + 86, + 359, + 483, + 785 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 810, + 220, + 825 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The rapid advancement of Large Language Models (LLMs) [2, 4, 7, 45, 56] has significantly accelerated the", + "bbox": [ + 89, + 835, + 482, + 867 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4d9029fafad4ac06bce0a77104698f6f2b9e9c48ca65696f7cb967e1e5e28bb3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 542, + 311, + 906, + 501 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1665665ec4ecccc53813df1560e205fefa93613dcc1393d7d742d6d0a58098c7.jpg", + "image_caption": [ + "Figure 1. Performance comparison on different benchmarks between compositional VLMs (dashed lines) and monolithic VLMs (solid lines). Previous monolithic VLMs exhibit a significant performance gap compared to compositional VLMs, while Our HoVLE demonstrate competitive capabilities with state-of-the-art compositional VLMs." + ], + "image_footnote": [], + "bbox": [ + 513, + 510, + 906, + 566 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "development of Vision-Language Models (VLMs) [26, 48, 58], which have demonstrated powerful visual-text processing capabilities. Compositional VLMs are currently the de facto approach for building high-performance VLMs [11, 27, 46], which integrate pre-trained vision encoders with LLMs, as shown in Fig. 2(a). By combining pre-trained visual understanding capabilities with language modeling capabilities, compositional VLMs achieve impressive multimodal performance. On the other hand, there is growing interest in developing VLMs without modality-specific encoders, referred to as monolithic VLMs in this paper. These models process image and text inputs as a unified whole, directly feeding them into the LLM, as illustrated in Fig. 2(b). Due to their monolithic design, these VLMs feature simpler architectures, offering greater potential for", + "bbox": [ + 509, + 672, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 104, + 875, + 220, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Corresponding Author: daijifeng@tsinghua.edu.cn.", + "bbox": [ + 104, + 887, + 392, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14559", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/34dede1f800a6182dfb96d2c82fec367cdbc76f4baa6dce686ff6b2420b21d36.jpg", + "image_caption": [ + "(a) Compositional VLMs" + ], + "image_footnote": [], + "bbox": [ + 99, + 87, + 215, + 234 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3926d4724e775c32cc46513c6434edaca47ebd944fed416d3fb147658cf874b9.jpg", + "image_caption": [ + "(b) Existing Monolithic VLMs" + ], + "image_footnote": [], + "bbox": [ + 241, + 88, + 349, + 233 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f07db8dd1e502ad993949f7b4f036b3f9b7317289c19c15d086e1bab869973f7.jpg", + "image_caption": [ + "(c) Our HoVLE", + "Figure 2. Comparison of VLM architectures. (a) Compositional VLMs integrate pre-trained vision encoders with LLMs, using an extra connector to align image and text embeddings. (b) Existing Monolithic VLMs directly feed image and text inputs into LLMs, which require continual pre-training to gain visual abilities. (c) HoVLE uses a holistic embedding module to project image and text input to a unified embedding space, enabling LLMs to interpret images in a text-like manner. Blocks with the same color have the same Transformer layer architecture." + ], + "image_footnote": [], + "bbox": [ + 375, + 88, + 480, + 233 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "multimodal tasks, such as unifying generation and recognition [43, 49, 57]. Monolithic VLMs are gaining increasing attention and have shown their potential across many studies [5, 8, 12].", + "bbox": [ + 89, + 397, + 482, + 457 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, monolithic VLMs still lag behind compositional VLMs in terms of performance. Some researches [5, 6, 43, 57] explore training monolithic VLMs from scratch using a combination of multi-modal and pure text data. Such practice fails to leverage existing pre-trained models, resulting in substantial computational demands for training. Another research direction [8, 12] proposes injecting visual understanding capabilities into pre-trained LLMs by conducting continual pre-training [16] with large-scale and high-quality multi-modal data. Nevertheless, such continual-tuning methods may interfere with the language knowledge already acquired by pre-trained LLMs, potentially limiting their multi-modal performance [12].", + "bbox": [ + 89, + 459, + 482, + 655 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The performance limitations of current monolithic VLMs mainly stem from the need to tune pre-trained LLMs. However, the success of compositional VLMs indicates that pre-trained LLMs can interpret visual information without further pre-training, as long as image embeddings are properly aligned with text embeddings. Thus, the reason for such LLM-tuning dilemma of current monolithic VLMs is the lack of a holistic embedding module that can coordinate different modalities.", + "bbox": [ + 89, + 657, + 482, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we develop a modality-sharing holistic embedding module as shown in Fig. 2(c), which maps mixed visual and language inputs into a shared embedding space. It can enable pre-trained LLMs to interpret visual input in the same way as text. Similar to compositional VLMs, this not only avoids tuning pre-trained LLMs but also preserves their language proficiency while successfully inte", + "bbox": [ + 89, + 795, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "grating visual capabilities. Our experiments demonstrate that such a holistic embedding module can indeed achieve high-performance monolithic VLMs.", + "bbox": [ + 511, + 90, + 903, + 135 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we introduce HoVLE, a novel monolithic vision-language model that learns a Holistic Vision-Language Embedding module to extend LLMs with vision capabilities. The holistic embedding module consists of the same causal Transformer layers as the LLM. It accepts both images and texts as input and projects them into a unified embedding space. These embeddings are then forwarded into the LLM, constituting a monolithic VLM. To empower the holistic embedding module with vision and language encoding abilities, a multi-stage training strategy is proposed. First, the module is trained to distill visual features from a pre-trained vision encoder and text embeddings from the LLM. The distillation stage does not require image-text pairs, but instead uses large-scale unpaired random images and text tokens as inputs. After distillation, the holistic embedding module is combined with the LLM to perform next token prediction on multi-modal data, so that the output embeddings can be further aligned into a unified space. Finally, the whole model is enhanced with an instruction tuning stage. The alignment and instruction tuning stages are similar to those used in compositional VLMs, eliminating the need for extra supervised data.", + "bbox": [ + 511, + 136, + 906, + 468 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Extensive experiments show that our HoVLE is able to achieve close performance with these leading compositional VLMs on 17 multi-modal benchmarks. It also surpasses previous monolithic VLMs by a large margin, e.g., $\\sim 15$ points on MMBench [30], showing the effectiveness of HoVLE. In conclusion, our contributions can be summarized as follows:", + "bbox": [ + 511, + 468, + 905, + 571 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- HoVLE is proposed as a novel monolithic VLM that learns a holistic embedding module to expand LLMs with vision capabilities. By projecting vision and language inputs to a shared embedding space, HoVLE can understand images without compromising language proficiency.", + "- A novel multi-stage training strategy is proposed to empower the holistic embedding module with vision and language encoding abilities. Our training procedure enables HoVLE to develop strong vision and language capabilities without requiring additional supervised data compared to compositional VLMs.", + "- HoVLE demonstrates comparable performance with compositional VLMs, and surpasses previous monolithic VLMs by a large margin." + ], + "bbox": [ + 511, + 574, + 905, + 801 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 814, + 653, + 830 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Compositional Vision-Language Models. The rapid advancements and success of Large Language Models (LLMs) have sparked researchers' interest in multi-modal models, with Vision-Language Models (VLMs) emerging", + "bbox": [ + 511, + 839, + 905, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14560", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "as a prominent focus [11, 12, 22, 26, 28, 41]. Compositional VLMs have become the dominant paradigm in the field, where image tokens are typically generated through visual encoders and an additional connector (e.g., MLP projection). Many notable commercial models, including GPT-4V [52], the Gemini series [44], Claude-3V [3], Grok-1.5V [50], and MM1 [36], have demonstrated impressive performance across various multi-modal tasks. In the open-source domain, models like LLaVA [26, 28], BLIP [22], Emu2 [41], InternVL [11], and Qwen-VL [4] stand out. However, compositional VLMs face several challenges, such as complexities in deployment and optimization due to their heterogeneous architecture, trade-offs in the design of network capacity between image and text processing, and potential limitations in cross-modal understanding capabilities [6, 8, 12, 43].", + "bbox": [ + 89, + 90, + 480, + 333 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Monolithic Vision-Language Models. In contrast to compositional VLMs, monolithic VLMs do not rely on a specific vision encoder. Some works explore training a monolithic VLM from scratch. Fuyu-8B [5] directly inputs image pixel patches into the backbone network using a simple linear projection, while Chameleon [43], Transfusion [57], and Emu3 [49] adopt image tokenizers to map images to discrete tokens as inputs. Nevertheless, these methods require collecting massive pure text and paired image-text data, as well as intensive training computational resources to achieve good performance. Other approaches propose to extend existing pre-trained LLMs to VLMs. SOLO [8] uses a simple linear projection to handle raw image pixels and fine-tunes the pre-trained LLM to develop vision abilities. EVE [12] employs a patch embedding layer with cross-attention to enhance image representation, alongside a patch aligning layer to provide fine-grained visual supervision with pre-trained vision encoders. However, these methods require continually pre-training LLMs, which may diminish their language proficiency. Current monolithic VLMs still underperform in image-text understanding tasks when compared to their compositional counterparts.", + "bbox": [ + 91, + 343, + 483, + 675 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In contrast, by introducing a holistic embedding module, HoVLE can achieve comparable performance with state-of-the-art compositional VLMs, significantly surpassing other monolithic VLMs.", + "bbox": [ + 89, + 678, + 482, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Knowledge Distillation. Knowledge distillation is an effective technique for transferring knowledge and abilities from a teacher model to a student model. In the realm of VLMs, knowledge distillation has also been successfully incorporated into various tasks. DistillVLM [13] and EfficientVLM [47] use knowledge distillation to compress the model size and accelerate inference speed. VPD [17] and LSKD [37] leverage tools or LLMs to synthesize multimodal data so that VLMs can be trained to distill the skills in specialist models and LLMs. Our work employs knowl", + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "edge distillation to endow the holistic embedding module with general vision and language encoding abilities. Since distillation targets can be computed independently, we can use unpaired random image and text data, eliminating the need to collect additional multi-modal data.", + "bbox": [ + 511, + 90, + 903, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 511, + 179, + 604, + 194 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we present HoVLE, a monolithic vision-language model with decent vision and language capabilities. In the following sections, we introduce the network architecture and the training procedures for such a unified model. The details are shown in Figure 3.", + "bbox": [ + 511, + 204, + 903, + 279 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Network Architecture", + "text_level": 1, + "bbox": [ + 511, + 287, + 718, + 301 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The success of compositional VLMs demonstrates that LLMs are capable of understanding visual information when image embeddings are properly aligned with text embeddings. Inspired by this, HoVLE proposes that a monolithic VLM can be built using a holistic embedding module. This module takes mixed raw images and texts as inputs and projects them into a shared embedding space. In this way, LLMs can interpret images in the same way as texts, resulting in a monolithic VLM.", + "bbox": [ + 511, + 310, + 903, + 446 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Holistic Embedding Module. We first introduce the holistic embedding module to project vision and language inputs to a shared embedding space. Specifically, given an input image $I$ and texts $T$ , we first transform them into vector forms via simple tokenization and linear projection:", + "bbox": [ + 511, + 453, + 905, + 529 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nx _ {I} = \\operatorname {P a t c h E m b e d} (\\operatorname {D y n P r o c e s s} (I)) + \\mathrm {P E}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 537, + 903, + 553 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nx _ {T} = \\operatorname {T e x t E m b e d} (\\operatorname {T o k e n i z e r} (T)), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 556, + 903, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where DynProcess(\\cdot) is the dynamic high-resolution strategy from InternVL [11]. This process divides the image into $448 \\times 448$ tiles and appends a global thumbnail to provide overall context. Similar operations have been adopted in many compositional and monolithic VLMs [12, 18, 51] and have proved beneficial for model performance. PatchEmbed(\\cdot) refers to a patch embedding layer with stride $s$ and converts the image tiles into a sequence $x_{I}$ of $n_I$ image tokens with dimension $c$ . $\\mathrm{PE} \\in \\mathbb{R}^{n_I \\times c}$ is a learnable position embedding. Tokenizer(\\cdot) converts raw texts into token indices and TextEmbed(\\cdot) is a word embedding layer with vocabulary size of $v$ and dimension of $c$ . $x_{T}$ is a sequence of $n_T$ text tokens.", + "bbox": [ + 511, + 580, + 905, + 775 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Afterward, these image tokens and text tokens are combined as the input tokens $x \\in \\mathbb{R}^{n \\times c}$ . The holistic embedding module utilizes a stack of causal Transformer layers to project these tokens to a shared embedding space:", + "bbox": [ + 511, + 777, + 903, + 838 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {x} = \\text {C a u s a l T r a n s f o r m e r} (x), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 845, + 903, + 862 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\hat{x}$ denotes final embeddings of the input. The architecture of each Transformer layer is the same as the LLM's", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14561", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5169ed9ce8be5ff8959ff8e84eafce012c32ca9fc19e612313ae56f1f254e43f.jpg", + "image_caption": [ + "(a) Architecture", + "Figure 3. (a) The architecture of HoVLE. HoVLE initially segments the input images into patches dynamically and tokenizes input texts. The holistic embedding module then projects them into a unified space. Finally, the LLM processes these unified embeddings to produce the final outputs. (b) The training strategies of HoVLE. Distillation stage trains the holistic embedding to distill a pre-trained vision encoder and text embeddings of the LLM using unpaired random images and texts. Alignment stage combines the holistic embedding module with a frozen LLM, conducting auto-regressive training to align the vision-language embeddings. Instruction tuning further enhance HoVLE's overall ability by tuning the whole model." + ], + "image_footnote": [], + "bbox": [ + 122, + 87, + 449, + 452 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/bee8b4481e490d536919aaa562bc4d8c664184d0b2b69941f2a66a9abd788368.jpg", + "image_caption": [ + "(b) Training Strategies" + ], + "image_footnote": [], + "bbox": [ + 460, + 88, + 867, + 452 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "for simplicity and unity. Similar to LLMs, these Transformer layers use causal attention for both image and text tokens. The effectiveness of this architecture suggests that 1D causal modeling works as well as non-causal modeling for images in multi-modal tasks, aligning with recent findings in [42].", + "bbox": [ + 89, + 570, + 482, + 661 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Monolithic Vision-Language Model. By combining the holistic embedding module with a pre-trained LLM, a monolithic VLM can be naturally built. After the holistic embedding module generates input embeddings, the LLM takes them as input and performs next-token prediction:", + "bbox": [ + 89, + 669, + 483, + 744 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np _ {n + 1} = \\operatorname {L L M} (\\hat {x}; \\theta_ {\\text {L L M}}), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 761, + 482, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\theta_{\\mathrm{LLM}}$ refers to parameters of LLM, $p_{n + 1}\\in \\mathbb{R}^v$ is the probability of the next token.", + "bbox": [ + 89, + 792, + 482, + 823 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this way, we replace the modality-specific encoders in compositional VLMs with a holistic embedding module. This module has the same Transformer layer as the LLM, so the whole model presents a simple architecture, leading to our monolithic VLM.", + "bbox": [ + 89, + 825, + 483, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Training Procedure", + "text_level": 1, + "bbox": [ + 513, + 569, + 700, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To empower the holistic embedding module with decent vision and language encoding abilities, a multi-stage training procedure is carefully designed. The first stage trains the holistic embedding module to distill the image feature from a pre-trained visual encoder and the text embeddings from an LLM, providing general encoding abilities. The second stage combines the holistic embedding module with the LLM to perform auto-regressive training, aligning different modalities to a shared embedding space.", + "bbox": [ + 511, + 593, + 906, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Distillation Stage. This stage aims to provide basic vision and language encoding abilities with the holistic embedding module. We extract image and text embeddings from the holistic embedding module output, denoted as $\\hat{x}_I$ and $\\hat{x}_T$ , respectively. A pre-trained visual encoder and an LLM embedding layer are utilized to compute the target image feature $z_I$ and text feature $z_T$ . The training objective is", + "bbox": [ + 511, + 734, + 906, + 842 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta_ {\\text {E m b e d}}} \\mathcal {L} _ {\\text {d i s t i l l}} \\left(\\hat {x} _ {I}, z _ {I}; \\theta_ {\\text {E m b e d}}\\right) + \\mathcal {L} _ {\\text {d i s t i l l}} \\left(\\hat {x} _ {T}, z _ {T}; \\theta_ {\\text {E m b e d}}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 852, + 906, + 875 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{distill}}$ is the negative cosine similarity between two", + "bbox": [ + 513, + 885, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14562", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "features, $\\theta_{\\mathrm{Embed}}$ refers to the parameters of the holistic embedding module.", + "bbox": [ + 89, + 90, + 482, + 121 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Unpaired Random Images and Texts for Distillation. The distillation stage does not put any requirement on input images and texts, because their target features are computed independently. Therefore, we propose to combine random images with random texts as input. In fact, this strategy proves to be more effective than using paired image-text inputs, potentially due to better vocabulary coverage during distillation, as demonstrated in Section 4.3.", + "bbox": [ + 89, + 127, + 482, + 247 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In practice, we randomly sample 500M images from LAION-2B [39] dataset and randomly sample 100 text tokens from the vocabulary for each image as training data. Each sample consists of a user query and a model response. We place the image at the beginning of the query, followed by the text content for the query and response, each with 50 random text tokens. Unlike previous monolithic VLMs, HoVLE can benefit from large-scale vision pre-training, unleashing the full potential of the monolithic VLM.", + "bbox": [ + 89, + 248, + 483, + 383 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Alignment Stage. To align the image and text embeddings, the holistic embedding module is further combined with the LLM to perform auto-regressive training. The training objective is", + "bbox": [ + 89, + 390, + 483, + 450 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta_ {\\text {E m b e d}}} \\sum_ {i} \\mathcal {L} _ {\\text {a l i g n}} \\left(p _ {i}, y _ {i}; \\theta_ {\\text {E m b e d}}, \\theta_ {\\text {L L M}}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 455, + 482, + 487 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{align}}$ is the cross-entropy loss, $p_i$ and $y_{i}$ denote the output probability and target index of the $i^{\\mathrm{th}}$ text token. During training, the LLM is frozen, and only the holistic embedding module is optimized.", + "bbox": [ + 89, + 493, + 482, + 553 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This stage is similar to the alignment stage used in compositional VLMs, but we do not have an extra MLP connector. During alignment, the model is trained on 45M multi-modal data from diverse domains, such as captioning, detection, and OCR data. Please refer to Appendix A for details.", + "bbox": [ + 89, + 554, + 483, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Instruction Tuning", + "text_level": 1, + "bbox": [ + 89, + 651, + 272, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The holistic embedding module is now empowered with strong vision and language encoding abilities. To further strengthen the whole VLM to follow instructions, a visual instruction tuning stage is incorporated.", + "bbox": [ + 89, + 674, + 482, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this stage, the whole model is optimized using the same auto-regressive training objective as in the alignment stage. The training data contains around 5M data, covering visual question answering, mathematics, world knowledge, etc. Please refer to Appendix A for details.", + "bbox": [ + 89, + 734, + 482, + 810 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 89, + 821, + 215, + 838 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 847, + 307, + 863 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Network Architecture. HoVLE adopts InternLM2-1.8B [7] as the LLM backbone. The holistic embedding is", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "composed of 8 causal Transformer layers, with hidden size 2048 and 16 attention heads, as in InternLM2-1.8B. The image patch embedding layer has a stride of $s = 28$ and output dimension $c = 2048$ . The text embedding has a vocabulary of 92,553. To demonstrate the high-resolution image processing ability of our model, like other works [5, 12, 24], we also provide HoVLE (HD). It uses a patch embedding with stride $s = 14$ and applies pixel shuffle [11] after the holistic embedding to reduce the number of visual tokens to 1/4. All other details are kept the same.", + "bbox": [ + 511, + 90, + 903, + 242 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Procedure. During distillation, we randomly sample 500M images from LAION-2B [39] with 100 random text indices for each image. We use InternViT-300M [9, 11] and the text embedding from InternLM2-1.8B as the teacher model for image and text tokens, respectively. For alignment, a collection of datasets containing 50M samples is used, as shown in Appendix A. For instruction tuning, a collection of datasets containing 5M samples is used, as shown in Appendix A. We use AdamW as the optimizer for all three stages, with learning rates 3e-4, 5e-5, and 4e-5, respectively. The batch size is set to 4096 for all stages. Please refer to Appendix A for more hyper-parameters.", + "bbox": [ + 511, + 250, + 903, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Benchmarks. We evaluate our model on 17 public benchmarks. MMbench [30], MME [14], MMMU [55], MM-Vet [54], and SEED [21] focus on multimodal perception and reasoning, providing a comprehensive assessment of VLM capabilities. POPE [23] and HallB [15] specifically target hallucination detection. For OCR and OCR-related visual question answering, TextVQA [40], ChartQA [33], InfoVQA [35], DocVQA [34], and OCRBench [29] are used. GQA [19] evaluates visual scene understanding, while ScienceQA [32] and AI2D [20] assess scientific image understanding. MathVista focuses on mathematical reasoning. For specific metric values shown in the experiments, MMbench-EN results are counted and presented as the MMB score. The POPE score is averaged across its three categories, the MME score is calculated as the sum of perception and cognition scores, and the SEED score reflects image accuracy specifically.", + "bbox": [ + 511, + 438, + 906, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Main Results", + "text_level": 1, + "bbox": [ + 513, + 724, + 651, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We compare HoVLE with previous compositional and monolithic VLMs in Table 1 and Table 2. We highlight several key observations as follows: (1) There are significant performance gaps between previous monolithic VLMs and compositional VLMs, indicating the challenges in building a monolithic VLM with state-of-the-art performance. (2) Compared with existing monolithic VLMs, our HoVLE demonstrates substantially better performance with fewer parameters. For example, HoVLE surpasses the previous best monolithic VLM by $\\sim 15$ points on MMBench, which", + "bbox": [ + 511, + 750, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14563", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/ae9a083c9e10ca6170e39490821ffa7f2fdf0734ae1bd086e4a17670173ddfa1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method# ParamsMMBMMEMMMUMM-VetPOPESEEDMathVistaHallBCCB
Compositional VLMs
VILA1.5-3B [25]3.1B63.4164833.335.485.967.931.631.224.1
DeepSeek-VL-1.3B [31]2.0B64.6153232.234.887.666.731.127.637.6
PaliGemma-3B [6]2.9B71.0168634.933.187.069.628.732.229.6
MM1-3B-Chat [36]3B67.8176233.943.787.468.832.0--
MiniCPM-V [53]2.8B64.1165038.331.179.565.628.936.241.4
MiniCPM-V-2 [53]2.8B69.1180938.241.086.367.138.736.145.3
InternVL1.5 [9]2.2B70.9190234.639.387.169.841.137.563.5
†InternVL2 [10]2.2B73.2187734.344.688.371.646.437.974.7
Qwen2-VL-2B [46]2.2B74.9187241.149.5--43.041.7-
Phi-3 Vision [1]4.2B80.5-40.4-85.8-44.5--
Monolithic VLMs
Fuyu-8B (HD) [5]8B10.7--21.474.1----
SOLO [8]7B-1260---64.434.4--
*Chameleon-7B [43]7B31.117025.48.3-30.622.317.13.5
EVE-7B [12]7B49.5148332.325.683.661.325.221.112.4
EVE-7B (HD) [12]7B52.3162832.625.785.064.634.226.416.3
Emu3 [49]8B58.5-31.637.285.268.2---
Our HoVLE2.6B71.9186433.744.387.670.746.239.675.3
Our HoVLE (HD)2.6B73.3186232.243.887.470.949.238.474.3
", + "bbox": [ + 181, + 88, + 812, + 385 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Comparison with existing VLMs on general VLM benchmarks. The highest scores in monolithic VLMs are highlighted in bold. *We test Chameleon-7B with its released model, but it frequently refuses to perform the task with a response of \"I can't help you with this\", thus resulting in poor performance. ${}^{ \\dagger }$ InternVL2 2.2B adopts the same LLM and supervised data with our HoVLE,sowe mark it as the compositional counterpart.", + "bbox": [ + 89, + 392, + 906, + 446 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "is a comprehensive multi-modal benchmark covering a variety of fine-grained abilities. This substantial improvement is attributed to the holistic embedding, which prevents continual pre-training from disrupting language knowledge stored in pre-trained LLMs. (3) By incorporating higher resolution, HoVLE (HD) consistently improves the performance on VQA benchmarks, demonstrating the versatility of our holistic embedding in handling different resolutions. (4) Compared with previous compositional VLMs, HoVLE (HD) is able to deliver competitive results with many of these models. This shows that the introduction of holistic embedding provides an effective encoding that allows the LLM to understand images while preserving its language abilities. While HoVLE (HD) still falls short of state-of-the-art compositional VLMs, like Qwen2-VL2B [46] and Phi-3 Vision [1], it narrows the gap between monolithic and compositional VLMs, suggesting the feasibility of a high-performance monolithic VLM.", + "bbox": [ + 88, + 450, + 482, + 723 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 734, + 253, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we aim to display and discuss the design choices that affect our HoVLE. Ablation experiments train the model with 300M and 8M data during the distillation and alignment stages respectively, unless specified.", + "bbox": [ + 89, + 757, + 482, + 819 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Holistic Embedding Depth. To validate the effectiveness of the holistic embedding, we first compare models with different holistic embedding depths in Table 3. With 0 Transformer layer, the model degenerates to a VLM without holistic embedding and the poor performance suggests", + "bbox": [ + 89, + 824, + 482, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the importance of this module. The model performs consistently better as the depth goes from 4 to 12. We speculate that deep embedding is beneficial for extracting visual representation. HoVLE uses 8 layers to achieve a balance between performance and model complexity.", + "bbox": [ + 511, + 450, + 906, + 527 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training Strategy. Table 4 displays the influence of different training stages. The results show that omitting any stage will result in a performance drop. Among these, removing the distillation stage leads to the most severe degradation, highlighting its importance. We observe that models after distillation exhibit a faster decrease in loss during subsequent training stages, likely because the distillation provides general abilities of representation extraction, which is helpful for accelerating training. The improvement from instruction tuning is greater than that of the alignment stage. We assume this is because the LLM has more trainable parameters, allowing the whole model to fit the data well.", + "bbox": [ + 511, + 530, + 908, + 712 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Distillation Data Scaling. Figure 6 illustrates how the performance changes as distillation data scales up. Overall, the model performs better with more distillation data, implying the importance of comprehensive training. Specifically, the model's performance improves rapidly with 300M data and continues to increase at a more gradual pace as more data is added. We use 500M distillation data as our default setting. Please refer to Appendix B for scaling performance on more benchmarks.", + "bbox": [ + 511, + 715, + 908, + 851 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Text Data for Distillation. Table 5 presents the effects of different text input choices during distillation. The LAION [39] dataset provides a short text caption for each", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14564", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8f2d2cf52876d36c033e6acf61ad3046bb4810ffc940940cb5751c9af9c1f47a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method# ParamsTextVQAChartVQAInfoVQADocVQAOCRbenchGQASQA-IAI2D
Compositional VLMs
VILA1.5-3B [25]3.1B60.4---43761.569.0-
DeepSeek-VL-1.3B [31]2.0B57.8---409--51.5
PaliGemma-3B [6]2.9B68.1---614--68.3
MM1-3B-Chat [36]3B71.9-----69.4-
MiniCPM-V [53]2.8B60.6--38.2366--56.3
MiniCPM-V-2 [53]2.8B74.1--71.9605--62.9
InternVL1.5 [9]2.2B70.574.855.485.065461.684.969.8
†InternVL2 [10]2.2B73.476.257.785.978461.084.974.1
Qwen2-VL-2B [46]2.2B79.773.565.590.1794---
Phi3-Vision [1]4.2B70.981.4----90.876.7
Monolithic VLMs
Fuyu-8B (HD) [5]8B-------64.5
SOLO [8]7B------73.361.4
*Chameleon-7B [43]7B4.82.95.01.57-47.246.0
EVE-7B [12]7B51.919.520.022.032760.863.048.5
EVE-7B (HD) [12]7B56.859.125.053.039862.664.959.1
Emu3 [49]8B64.768.643.876.368760.389.270.0
Our HoVLE2.6B66.077.652.184.773563.694.873.7
Our HoVLE (HD)2.6B70.978.655.786.174064.994.873.0
", + "bbox": [ + 168, + 88, + 826, + 385 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/92f875096f322e90090b082c3facfe66c6277160608b67f0f0b75987a352cdd5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 112, + 421, + 302, + 525 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d50f9ad8eae487fb8b973c2434a686aa556f2ff6e22950649d71fb29cd95f35e.jpg", + "image_caption": [ + "Figure 4. Distillation data scaling performance." + ], + "image_footnote": [], + "bbox": [ + 305, + 421, + 493, + 523 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4d1605567c118e5b519892f419bdf0820702a124d90fab752e0c35a5656b77dc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 421, + 687, + 523 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/253e95810dc5ced324554bbda1376815d0d2e8c5db25aebb9111f9ad32fb97a4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 691, + 421, + 879, + 525 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/94b837e9eca76f663c3eda822bff19e4f1328822d043315623c06894ae5a1c69.jpg", + "table_caption": [ + "Table 2. Comparison with existing VLMs on visual question answering benchmarks. The highest scores in monolithic VLMs are highlighted in bold. Same as Tab. 1, *Chameleon-7B frequently refuses to response, and ${}^{ \\dagger }$ InternVL2 2.2B is our compositional counterpart." + ], + "table_footnote": [], + "table_body": "
DepthMMBMMESEEDTextVQAInfoVQADocVQA
035.4114946.011.319.112.2
467.8180168.047.742.673.6
872.0186270.962.051.482.1
1272.3188270.965.552.584.7
", + "bbox": [ + 114, + 547, + 459, + 631 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/df8d1deac675cc843e4af3c1ba60f477d49b8b3a87949e5c34587c5faf88e8b1.jpg", + "table_caption": [ + "Table 3. Ablation on holistic embedding depth." + ], + "table_footnote": [], + "table_body": "
Dist. Align. I.T.MMBMMESEEDTextVQAInfoVQADocVQA
✓✓37.3127446.411.219.3
✓✓71.3182870.560.149.2
✓✓66.1169866.755.043.0
✓✓✓72.0186270.962.051.4
", + "bbox": [ + 101, + 651, + 475, + 731 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6f47121e389921b3ae89a38f2784a304dae19e6127d96472bbda826fef117158.jpg", + "table_caption": [ + "Table 4. Ablation on training strategies." + ], + "table_footnote": [], + "table_body": "
Text DataMMBMMESEEDTextVQAInfoVQADocVQA
Paired71.7176870.561.349.481.5
Random72.0186270.962.051.482.1
", + "bbox": [ + 102, + 752, + 468, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Ablation on text data for distillation.", + "bbox": [ + 143, + 808, + 428, + 819 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "image, so we can also use image-text pairs as input. However, this approach yields relatively lower results, and the model continues to exhibit high loss values during alignment. We hypothesize that the text data from the LAION", + "bbox": [ + 89, + 840, + 480, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "dataset lacks sufficient diversity to cover the full vocabulary, suggesting that paired data may require higher quality and variety in texts. In contrast, we propose using random texts as input. Table 5 shows that this strategy outperforms paired text data. We also observe lower loss during alignment. By leveraging random text, we reduce the high-quality need for image-text pair collection, thereby enhancing data utilization for large-scale training.", + "bbox": [ + 511, + 551, + 906, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Analysis", + "text_level": 1, + "bbox": [ + 511, + 681, + 614, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visualization. To better understand the effects of our introduced holistic embedding, we visualize the attention map of the LLM in Figure 5. We find that existing monolithic models, such as EVE [12] and Emu3 [49], exhibit denser text-to-image attention in their initial layers, while the compositional InternVL2 [10] and our HoVLE model maintain sparse attention patterns throughout. All models demonstrate sparse attention patterns in the later layers (with the final layers shown). Please refer to Appendix C for more attention maps in the middle layers.", + "bbox": [ + 511, + 703, + 906, + 854 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We hypothesize that sparser text-to-image attention indicates that visual features are effectively extracted and well-aligned with text tokens, allowing the text tokens to focus", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14565", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2699d803a1aad92a6bcca7e9f35c86d5a743bc99d9cbdefc38ceab6755bf659b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 158, + 88, + 316, + 222 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1d76128a5a51400e1a75e7954d9846d1ad48833416451fb0e580fd0eee6bed82.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 334, + 88, + 472, + 220 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/39a9cdb1effd2a6b2a66132b55f294617849ba9899c5dd765ae02a0dde89e23a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 495, + 88, + 630, + 220 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b1ed64c45eab64b2819afccb46f204fcfe159456cad018663c541d6f47847ba0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 89, + 785, + 220 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2cdd9aeb0759e22cb2fb3cb5a3330d9e8008a1021ecd47469371e38e63992505.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 794, + 98, + 839, + 335 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/873bf3203cc994facceac438dd30b9ea4eccf04207f4d967abf28ad4387bb017.jpg", + "image_caption": [ + "Figure 5. Attention Maps for EVE, Emu3, InternVL2 and our HoVLE at the first and last layers of LLM backbones. Y-axis represents query tokens, and X-axis represents key tokens, with text modality tokens in gray and image modality tokens in yellow. All four models share the same input, but the sequence lengths of input tokens are different due to different image pre-processing. We highlight text-to-image attention below each full attention map. Our HoVLE, like the compositional InternVL2, has sparse attention across all network layers, while other monolithic models Emu3 and EVE have denser attention in shallow layers." + ], + "image_footnote": [], + "bbox": [ + 156, + 223, + 313, + 343 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/24475e612ea96d33a8c1df1105e18bf343cce755f1af31d70de9a5bc384472f6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 336, + 223, + 472, + 343 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2a743b2267515adae460315f409e22a74599db4cb1386e0d17a39ce353cda6f8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 223, + 630, + 343 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d2ddb53de9b6f618ac6bcc77f5fdd57be9e45d9fb61d0d9430f7d2f5b7c20cfb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 653, + 223, + 787, + 343 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/619503184fe418f96609e5b727f2ff3a38c33cdb96c986109e948fdedda484a5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelTile ResolutionFLOPs(T)↓TTFT↓TPS↑Avg. Score↑
InternVL2448 × 4481.811.3728165.8
HoVLE (HD)336 × 3361.511.2424365.7
HoVLE (HD)392 × 3921.891.5222466.1
HoVLE (HD)448 × 4482.181.8420666.8
", + "bbox": [ + 94, + 417, + 488, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Inference speed-performance trade-off of HoVLE (HD). \"TTFT\" and \"TPS\" denotes the time to first token in seconds and throughput in tokens per second, respectively. \"Avg. Score\" refers to the average performance over 17 benchmarks. Please refer to Appendix for detailed results.", + "bbox": [ + 89, + 520, + 485, + 589 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "only on the most relevant visual information. In contrast, denser text-to-image attention may suggest that the model is still in the process of extracting visual features within the LLM component. Therefore, by introducing holistic embedding, the LLM may process vision and language inputs more effectively.", + "bbox": [ + 89, + 595, + 482, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Speed-Performance Trade-off. We report the speed-performance trade-off of HoVLE (HD) in Table 6. For evaluation, we use one image and 256 text tokens as input, with the model restricted to output 120 tokens during testing. InternVL2 serves as the main comparison model, as it uses the same LLM. When the tile resolution of DynProcess( $\\cdot$ ) is 336, HoVLE (HD) achieves comparable performance to InternVL2, while also showing lower computational demand. As the input resolution increases, HoVLE's performance improves consistently, but at the cost of slower speed. Additionally, we observe that using the KV cache [38] during inference significantly reduces computational time, making kernel launch the main bottleneck in TPS. Optimizing the kernel could further enhance processing speed.", + "bbox": [ + 89, + 688, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 419, + 633, + 435 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work presents HoVLE, a monolithic Vision-Language Model (VLM) that processes images and texts in a unified manner. HoVLE introduces a holistic embedding module that projects image and text inputs into a shared embedding space. This allows the Large Language Model (LLM) to interpret images in the same way as texts. Furthermore, a multi-stage training procedure is designed to strengthen the encoding abilities of the holistic embedding. A distillation stage first provides general vision and language encoding capabilities, while an alignment stage aligns different modalities. An instruction tuning stage enhances the overall abilities in the end. Extensive experiments show that HoVLE surpasses previous monolithic VLMs and demonstrates competitive performance with compositional VLMs. This work focuses on enhancing monolithic VLMs by introducing a holistic embedding module, aiming to elevate their capabilities rather than rival those of compositional VLMs. This work narrows the gap between monolithic and compositional VLMs, providing a promising direction for the development of monolithic VLMs.", + "bbox": [ + 511, + 444, + 906, + 746 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. Due to limited computational resources, our experiments are conducted on 2B parameter scale. Scaling to larger scales can provide further insight into the model scalability of our approach, which we hope to explore in future work.", + "bbox": [ + 511, + 747, + 906, + 821 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work is supported by the National Key R&D Program of China (NO. 2022ZD0161300), by the National Natural Science Foundation of China (62376134).", + "bbox": [ + 511, + 821, + 906, + 883 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14566", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Marah Abdin, Jyoti Aneja, Hany Awadalla, Ahmed Awadallah, Ammar Ahmad Awan, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Jianmin Bao, Harkirat Behl, et al. Phi-3 technical report: A highly capable language model locally on your phone. arxiv:2404.14219, 2024. 6, 7", + "[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arxiv:2303.08774, 2023. 1", + "[3] AI Anthropic. The claude 3 model family: Opus, sonnet, haiku. Claude-3 Model Card, 2024. 3", + "[4] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arxiv:2308.12966, 2023. 1, 3", + "[5] Rohan Bavishi, Erich Elsen, Curtis Hawthorne, Maxwell Nye, Augustus Odena, Arushi Somani, and Sagnak Tasirlar. Introducing our multimodal models. https://www. adept.ai/blog/fuyu-8b, 2023. 2, 3, 5, 6, 7", + "[6] Lucas Beyer, Andreas Steiner, André Susano Pinto, Alexander Kolesnikov, Xiao Wang, Daniel Salz, Maxim Neumann, Ibrahim Alabdulmohsin, Michael Tschannen, Emanuele Bugliarello, et al. Paligemma: A versatile 3b vlm for transfer. arxiv:2407.07726, 2024. 2, 3, 6, 7", + "[7] Zheng Cai, Maosong Cao, Haojiong Chen, Kai Chen, Keyu Chen, Xin Chen, Xun Chen, Zehui Chen, Zhi Chen, Pei Chu, et al. Internl m2 technical report. arxiv:2403.17297, 2024. 1, 5", + "[8] Yangyi Chen, Xingyao Wang, Hao Peng, and Heng Ji. A single transformer for scalable vision-language modeling. arxiv:2407.06438, 2024. 2, 3, 6, 7", + "[9] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arxiv:2404.16821, 2024. 5, 6, 7", + "[10] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. Internv12: Better than the best—expanding performance boundaries of open-source multimodal models with the progressive scaling strategy. https://internvl.github.io/blog/2024-07-02-InternVL-2.0/, 2024.6.7", + "[11] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In CVPR, pages 24185–24198, 2024. 1, 3, 5", + "[12] Haiwen Diao, Yufeng Cui, Xiaotong Li, Yueze Wang, Huchuan Lu, and Xinlong Wang. Unveiling encoder-free vision-language models. arxiv:2406.11832, 2024. 2, 3, 5, 6, 7", + "[13] Zhiyuan Fang, Jianfeng Wang, Xiaowei Hu, Lijuan Wang, Yezhou Yang, and Zicheng Liu. Compressing visual-linguistic model via knowledge distillation. In ICCV, pages 1428-1438, 2021. 3" + ], + "bbox": [ + 93, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. arxiv:2306.13394, 2023.5", + "[15] Tianrui Guan, Fuxiao Liu, Xiyang Wu, Ruiqi Xian, Zongxia Li, Xiaoyu Liu, Xijun Wang, Lichang Chen, Furong Huang, Yaser Yacoob, et al. Hallusionbench: An advanced diagnostic suite for entangled language hallucination and visual illusion in large vision-language models. arxiv:2310.14566, 2023. 5", + "[16] Suchin Gururangan, Ana Marasovic, Swabha Swayamdipta, Kyle Lo, Iz Beltagy, Doug Downey, and Noah A Smith. Don't stop pretraining: Adapt language models to domains and tasks. In ACL, pages 8342-8360, 2020. 2", + "[17] Yushi Hu, Otilia Stretcu, Chun-Ta Lu, Krishnamurthy Viswanathan, Kenji Hata, Enming Luo, Ranjay Krishna, and Ariel Fuxman. Visual program distillation: Distilling tools and programmatic reasoning into vision-language models. In CVPR, pages 9590-9601, 2024. 3", + "[18] Runhui Huang, Xinpeng Ding, Chunwei Wang, Jianhua Han, Yulong Liu, Hengshuang Zhao, Hang Xu, Lu Hou, Wei Zhang, and Xiaodan Liang. Hires-llava: Restoring fragmentation input in high-resolution large vision-language models. arXiv:2407.08706, 2024.3", + "[19] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In CVPR, pages 6700-6709, 2019. 5, 1", + "[20] Aniruddha Kembhavi, Mike Salvato, Eric Kolve, Minjoon Seo, Hannaneh Hajishirzi, and Ali Farhadi. A diagram is worth a dozen images. In ECCV, pages 235-251. Springer, 2016. 5, 1", + "[21] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arxiv:2307.16125, 2023. 5", + "[22] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, pages 12888-12900. PMLR, 2022. 3", + "[23] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. In EMNLP, pages 292-305, 2023. 5", + "[24] Zhang Li, Biao Yang, Qiang Liu, Zhiyin Ma, Shuo Zhang, Jingxu Yang, Yabo Sun, Yuliang Liu, and Xiang Bai. Monkey: Image resolution and text label are important things for large multi-modal models. In CVPR, pages 26763-26773, 2024. 5", + "[25] Ji Lin, Hongxu Yin, Wei Ping, Pavlo Molchanov, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models. In CVPR, pages 26689-26699, 2024. 6, 7", + "[26] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In CVPR, pages 26296-26306, 2024. 1, 3", + "[27] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge. https :" + ], + "bbox": [ + 516, + 92, + 903, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14567", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "//llava-v1.github.io/blog/2024-01-30-llava-last/, 2024.1", + "[28] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2024. 3", + "[29] Yuliang Liu, Zhang Li, Biao Yang, Chunyuan Li, Xucheng Yin, Cheng-lin Liu, Lianwen Jin, and Xiang Bai. On the hidden mystery ofOCR in large multimodal models. arxiv:2305.07895, 2023. 5", + "[30] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? In ECCV, pages 216-233. Springer, 2025. 2, 5", + "[31] Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhuoshu Li, Hao Yang, et al. Deepseek-vl: towards real-world vision-language understanding. arxiv:2403.05525, 2024. 6, 7", + "[32] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. In NeurIPS, pages 2507–2521, 2022. 5, 1", + "[33] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arxiv:2203.10244, 2022. 5, 1", + "[34] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In WACV, pages 2200-2209, 2021. 5", + "[35] Minesh Mathew, Viraj Bagal, Ruben Tito, Dimosthenis Karatzas, Ernest Valveny, and CV Jawahar. Infographicvqa. In WACV, pages 1697-1706, 2022. 5, 1", + "[36] Brandon McKinzie, Zhe Gan, Jean-Philippe Fauconnier, Sam Dodge, Bowen Zhang, Philipp Dufter, Dhruti Shah, Xianzhi Du, Futang Peng, Floris Weers, et al. Mm1: Methods, analysis & insights from multimodal llm pre-training. arxiv:2403.09611, 2024. 3, 6, 7", + "[37] Jae Sung Park, Jack Hessel, Khyathi Chandu, Paul Pu Liang, Ximing Lu, Peter West, Youngjae Yu, Qiuyuan Huang, Jianfeng Gao, Ali Farhadi, et al. Localized symbolic knowledge distillation for visual commonsense models. In NeurIPS, 2024. 3", + "[38] Reiner Pope, Sholto Douglas, Aakanksha Chowdhery, Jacob Devlin, James Bradbury, Jonathan Heek, Kefan Xiao, Shivani Agrawal, and Jeff Dean. Efficiently scaling transformer inference. Proceedings of Machine Learning and Systems, 5: 606-624, 2023. 8", + "[39] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS, pages 25278–25294, 2022. 5, 6, 1", + "[40] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In CVPR, pages 8317-8326, 2019. 5, 1" + ], + "bbox": [ + 91, + 90, + 480, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Quan Sun, Yufeng Cui, Xiaosong Zhang, Fan Zhang, Qiying Yu, Yueze Wang, Yongming Rao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative multimodal models are in-context learners. In CVPR, pages 14398-14409, 2024. 3", + "[42] Chenxin Tao, Xizhou Zhu, Shiqian Su, Lewei Lu, Changyao Tian, Xuan Luo, Gao Huang, Hongsheng Li, Yu Qiao, Jie Zhou, et al. Learning 1d causal visual representation with de-focus attention networks. arxiv:2406.04342, 2024. 4", + "[43] Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models. arxiv:2405.09818, 2024. 2, 3, 6, 7", + "[44] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arxiv:2312.11805, 2023. 3", + "[45] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arxiv:2302.13971, 2023. 1", + "[46] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arxiv:2409.12191, 2024. 1, 6, 7", + "[47] Tiannan Wang, Wangchunshu Zhou, Yan Zeng, and Xinsong Zhang. Efficientvlm: Fast and accurate vision-language models via knowledge distillation and modal-adaptive pruning. arxiv:2210.07795, 2022. 3", + "[48] Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong Lu, Jie Zhou, Yu Qiao, et al. Vision: Large language model is also an open-ended decoder for vision-centric tasks. In NeurIPS, 2024. 1", + "[49] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arxiv:2409.18869, 2024. 2, 3, 6, 7", + "[50] X.ai. Grok-1.5 vision preview. https://x.ai/blog/grok-1.5v, 2024.3", + "[51] Ruyi Xu, Yuan Yao, Zonghao Guo, Junbo Cui, Zanlin Ni, Chunjiang Ge, Tat-Seng Chua, Zhiyuan Liu, Maosong Sun, and Gao Huang. Llava-uhd: an Imm perceiving any aspect ratio and high-resolution images. arXiv:2403.11703, 2024. 3", + "[52] Zhengyuan Yang, Linjie Li, Kevin Lin, Jianfeng Wang, Chung-Ching Lin, Zicheng Liu, and Lijuan Wang. The dawn of Imms: Preliminary explorations with gpt-4v (ision). arxiv:2309.17421, 9(1):1, 2023. 3", + "[53] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arxiv:2408.01800, 2024.6, 7", + "[54] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. Mm-vet: Evaluating large multimodal models for integrated capabilities. arxiv:2308.02490, 2023. 5" + ], + "bbox": [ + 514, + 92, + 903, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14568", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[55] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In CVPR, pages 9556-9567, 2024. 5", + "[56] Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. A survey of large language models. arxiv:2303.18223, 2023. 1", + "[57] Chunting Zhou, Lili Yu, Arun Babu, Kushal Tirumala, Michihiro Yasunaga, Leonid Shamis, Jacob Kahn, Xuezhe Ma, Luke Zettlemoyer, and Omer Levy. Transfusion: Predict the next token and diffuse images with one multi-modal model. arxiv:2408.11039, 2024. 2, 3", + "[58] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arxiv:2304.10592, 2023. 1" + ], + "bbox": [ + 91, + 92, + 480, + 343 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "14569", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/HoVLE_ Unleashing the Power of Monolithic Vision-Language Models with Holistic Vision-Language Embedding/126721ba-0c45-4655-8cc5-9017b39955e6_model.json b/2025/HoVLE_ Unleashing the Power of Monolithic Vision-Language Models with Holistic Vision-Language Embedding/126721ba-0c45-4655-8cc5-9017b39955e6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5abb68d225ae4898e585a2ae61b829c1811ee1 --- /dev/null +++ b/2025/HoVLE_ Unleashing the Power of Monolithic Vision-Language Models with Holistic Vision-Language Embedding/126721ba-0c45-4655-8cc5-9017b39955e6_model.json @@ -0,0 +1,2334 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.119, + 0.13, + 0.88, + 0.177 + ], + "angle": 0, + "content": "HoVLE: Unleashing the Power of Monolithic Vision-Language Models with Holistic Vision-Language Embedding" + }, + { + "type": "text", + "bbox": [ + 0.126, + 0.2, + 0.873, + 0.238 + ], + "angle": 0, + "content": "Chenxin Tao\\(^{1,3*}\\), Shiqian Su\\(^{1,3*}\\), Xizhou Zhu\\(^{1,2*}\\), Chenyu Zhang\\(^{1,2}\\), Zhe Chen\\(^{4,2}\\), Jiawen Liu\\(^{6}\\), Wenhai Wang\\(^{5,2}\\), Lewei Lu\\(^{3}\\), Gao Huang\\(^{1}\\), Yu Qiao\\(^{2}\\), Jifeng Dai\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.27, + 0.247, + 0.728, + 0.264 + ], + "angle": 0, + "content": "\\(^{1}\\)Tsinghua University \\(^{2}\\)Shanghai Artificial Intelligence Laboratory" + }, + { + "type": "text", + "bbox": [ + 0.35, + 0.264, + 0.649, + 0.279 + ], + "angle": 0, + "content": "\\(^{3}\\)SenseTime Research \\(^{4}\\)Nanjing University" + }, + { + "type": "text", + "bbox": [ + 0.269, + 0.279, + 0.73, + 0.295 + ], + "angle": 0, + "content": "5The Chinese University of Hong Kong 6 Johns Hopkins University" + }, + { + "type": "text", + "bbox": [ + 0.305, + 0.297, + 0.733, + 0.312 + ], + "angle": 0, + "content": "Model: https://huggingface.co/OpenGVLab/HoVLE" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.329, + 0.327, + 0.345 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.361, + 0.485, + 0.786 + ], + "angle": 0, + "content": "The rapid advance of Large Language Models (LLMs) has catalyzed the development of Vision-Language Models (VLMs). Monolithic VLMs, which avoid modality-specific encoders, offer a promising alternative to the compositional ones but face the challenge of inferior performance. Most existing monolithic VLMs require tuning pre-trained LLMs to acquire vision abilities, which may degrade their language capabilities. To address this dilemma, this paper presents a novel high-performance monolithic VLM named HoVLE. We note that LLMs have been shown to be capable of interpreting images when image embeddings are aligned with text embeddings. The challenge for current monolithic VLMs actually lies in the lack of a holistic embedding module for both vision and language inputs. Therefore, HoVLE introduces a holistic embedding module that converts visual and textual inputs into a shared space, allowing LLMs to process images in the same way as texts. Furthermore, a multi-stage training strategy is carefully designed to empower the holistic embedding module. It is first trained to distill visual features from a pre-trained vision encoder and text embeddings from the LLM, enabling largescale training with unpaired random images and text tokens. The whole model further undergoes next-token prediction on multi-modal data to align the embeddings. Finally, an instruction-tuning stage is incorporated. Our experiments show that HoVLE achieves performance close to leading compositional models on various benchmarks, outperforming previous monolithic models by a large margin." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.811, + 0.222, + 0.827 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.837, + 0.483, + 0.868 + ], + "angle": 0, + "content": "The rapid advancement of Large Language Models (LLMs) [2, 4, 7, 45, 56] has significantly accelerated the" + }, + { + "type": "image", + "bbox": [ + 0.543, + 0.313, + 0.907, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.511, + 0.907, + 0.568 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.578, + 0.907, + 0.661 + ], + "angle": 0, + "content": "Figure 1. Performance comparison on different benchmarks between compositional VLMs (dashed lines) and monolithic VLMs (solid lines). Previous monolithic VLMs exhibit a significant performance gap compared to compositional VLMs, while Our HoVLE demonstrate competitive capabilities with state-of-the-art compositional VLMs." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.674, + 0.909, + 0.903 + ], + "angle": 0, + "content": "development of Vision-Language Models (VLMs) [26, 48, 58], which have demonstrated powerful visual-text processing capabilities. Compositional VLMs are currently the de facto approach for building high-performance VLMs [11, 27, 46], which integrate pre-trained vision encoders with LLMs, as shown in Fig. 2(a). By combining pre-trained visual understanding capabilities with language modeling capabilities, compositional VLMs achieve impressive multimodal performance. On the other hand, there is growing interest in developing VLMs without modality-specific encoders, referred to as monolithic VLMs in this paper. These models process image and text inputs as a unified whole, directly feeding them into the LLM, as illustrated in Fig. 2(b). Due to their monolithic design, these VLMs feature simpler architectures, offering greater potential for" + }, + { + "type": "page_footnote", + "bbox": [ + 0.106, + 0.876, + 0.221, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.106, + 0.888, + 0.393, + 0.9 + ], + "angle": 0, + "content": "Corresponding Author: daijifeng@tsinghua.edu.cn." + }, + { + "type": "list", + "bbox": [ + 0.106, + 0.876, + 0.393, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14559" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.088, + 0.216, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.094, + 0.238, + 0.212, + 0.248 + ], + "angle": 0, + "content": "(a) Compositional VLMs" + }, + { + "type": "image", + "bbox": [ + 0.242, + 0.089, + 0.351, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.222, + 0.238, + 0.366, + 0.248 + ], + "angle": 0, + "content": "(b) Existing Monolithic VLMs" + }, + { + "type": "image", + "bbox": [ + 0.377, + 0.089, + 0.482, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.389, + 0.238, + 0.464, + 0.248 + ], + "angle": 0, + "content": "(c) Our HoVLE" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.256, + 0.483, + 0.381 + ], + "angle": 0, + "content": "Figure 2. Comparison of VLM architectures. (a) Compositional VLMs integrate pre-trained vision encoders with LLMs, using an extra connector to align image and text embeddings. (b) Existing Monolithic VLMs directly feed image and text inputs into LLMs, which require continual pre-training to gain visual abilities. (c) HoVLE uses a holistic embedding module to project image and text input to a unified embedding space, enabling LLMs to interpret images in a text-like manner. Blocks with the same color have the same Transformer layer architecture." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.398, + 0.483, + 0.458 + ], + "angle": 0, + "content": "multimodal tasks, such as unifying generation and recognition [43, 49, 57]. Monolithic VLMs are gaining increasing attention and have shown their potential across many studies [5, 8, 12]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.46, + 0.483, + 0.656 + ], + "angle": 0, + "content": "However, monolithic VLMs still lag behind compositional VLMs in terms of performance. Some researches [5, 6, 43, 57] explore training monolithic VLMs from scratch using a combination of multi-modal and pure text data. Such practice fails to leverage existing pre-trained models, resulting in substantial computational demands for training. Another research direction [8, 12] proposes injecting visual understanding capabilities into pre-trained LLMs by conducting continual pre-training [16] with large-scale and high-quality multi-modal data. Nevertheless, such continual-tuning methods may interfere with the language knowledge already acquired by pre-trained LLMs, potentially limiting their multi-modal performance [12]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.658, + 0.483, + 0.793 + ], + "angle": 0, + "content": "The performance limitations of current monolithic VLMs mainly stem from the need to tune pre-trained LLMs. However, the success of compositional VLMs indicates that pre-trained LLMs can interpret visual information without further pre-training, as long as image embeddings are properly aligned with text embeddings. Thus, the reason for such LLM-tuning dilemma of current monolithic VLMs is the lack of a holistic embedding module that can coordinate different modalities." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To this end, we develop a modality-sharing holistic embedding module as shown in Fig. 2(c), which maps mixed visual and language inputs into a shared embedding space. It can enable pre-trained LLMs to interpret visual input in the same way as text. Similar to compositional VLMs, this not only avoids tuning pre-trained LLMs but also preserves their language proficiency while successfully inte" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.136 + ], + "angle": 0, + "content": "grating visual capabilities. Our experiments demonstrate that such a holistic embedding module can indeed achieve high-performance monolithic VLMs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.137, + 0.907, + 0.469 + ], + "angle": 0, + "content": "In this work, we introduce HoVLE, a novel monolithic vision-language model that learns a Holistic Vision-Language Embedding module to extend LLMs with vision capabilities. The holistic embedding module consists of the same causal Transformer layers as the LLM. It accepts both images and texts as input and projects them into a unified embedding space. These embeddings are then forwarded into the LLM, constituting a monolithic VLM. To empower the holistic embedding module with vision and language encoding abilities, a multi-stage training strategy is proposed. First, the module is trained to distill visual features from a pre-trained vision encoder and text embeddings from the LLM. The distillation stage does not require image-text pairs, but instead uses large-scale unpaired random images and text tokens as inputs. After distillation, the holistic embedding module is combined with the LLM to perform next token prediction on multi-modal data, so that the output embeddings can be further aligned into a unified space. Finally, the whole model is enhanced with an instruction tuning stage. The alignment and instruction tuning stages are similar to those used in compositional VLMs, eliminating the need for extra supervised data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.469, + 0.906, + 0.573 + ], + "angle": 0, + "content": "Extensive experiments show that our HoVLE is able to achieve close performance with these leading compositional VLMs on 17 multi-modal benchmarks. It also surpasses previous monolithic VLMs by a large margin, e.g., \\(\\sim 15\\) points on MMBench [30], showing the effectiveness of HoVLE. In conclusion, our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.575, + 0.906, + 0.666 + ], + "angle": 0, + "content": "- HoVLE is proposed as a novel monolithic VLM that learns a holistic embedding module to expand LLMs with vision capabilities. By projecting vision and language inputs to a shared embedding space, HoVLE can understand images without compromising language proficiency." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.666, + 0.905, + 0.756 + ], + "angle": 0, + "content": "- A novel multi-stage training strategy is proposed to empower the holistic embedding module with vision and language encoding abilities. Our training procedure enables HoVLE to develop strong vision and language capabilities without requiring additional supervised data compared to compositional VLMs." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.756, + 0.905, + 0.802 + ], + "angle": 0, + "content": "- HoVLE demonstrates comparable performance with compositional VLMs, and surpasses previous monolithic VLMs by a large margin." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.575, + 0.906, + 0.802 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.815, + 0.655, + 0.831 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Compositional Vision-Language Models. The rapid advancements and success of Large Language Models (LLMs) have sparked researchers' interest in multi-modal models, with Vision-Language Models (VLMs) emerging" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14560" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.334 + ], + "angle": 0, + "content": "as a prominent focus [11, 12, 22, 26, 28, 41]. Compositional VLMs have become the dominant paradigm in the field, where image tokens are typically generated through visual encoders and an additional connector (e.g., MLP projection). Many notable commercial models, including GPT-4V [52], the Gemini series [44], Claude-3V [3], Grok-1.5V [50], and MM1 [36], have demonstrated impressive performance across various multi-modal tasks. In the open-source domain, models like LLaVA [26, 28], BLIP [22], Emu2 [41], InternVL [11], and Qwen-VL [4] stand out. However, compositional VLMs face several challenges, such as complexities in deployment and optimization due to their heterogeneous architecture, trade-offs in the design of network capacity between image and text processing, and potential limitations in cross-modal understanding capabilities [6, 8, 12, 43]." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.344, + 0.485, + 0.676 + ], + "angle": 0, + "content": "Monolithic Vision-Language Models. In contrast to compositional VLMs, monolithic VLMs do not rely on a specific vision encoder. Some works explore training a monolithic VLM from scratch. Fuyu-8B [5] directly inputs image pixel patches into the backbone network using a simple linear projection, while Chameleon [43], Transfusion [57], and Emu3 [49] adopt image tokenizers to map images to discrete tokens as inputs. Nevertheless, these methods require collecting massive pure text and paired image-text data, as well as intensive training computational resources to achieve good performance. Other approaches propose to extend existing pre-trained LLMs to VLMs. SOLO [8] uses a simple linear projection to handle raw image pixels and fine-tunes the pre-trained LLM to develop vision abilities. EVE [12] employs a patch embedding layer with cross-attention to enhance image representation, alongside a patch aligning layer to provide fine-grained visual supervision with pre-trained vision encoders. However, these methods require continually pre-training LLMs, which may diminish their language proficiency. Current monolithic VLMs still underperform in image-text understanding tasks when compared to their compositional counterparts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.679, + 0.483, + 0.741 + ], + "angle": 0, + "content": "In contrast, by introducing a holistic embedding module, HoVLE can achieve comparable performance with state-of-the-art compositional VLMs, significantly surpassing other monolithic VLMs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Knowledge Distillation. Knowledge distillation is an effective technique for transferring knowledge and abilities from a teacher model to a student model. In the realm of VLMs, knowledge distillation has also been successfully incorporated into various tasks. DistillVLM [13] and EfficientVLM [47] use knowledge distillation to compress the model size and accelerate inference speed. VPD [17] and LSKD [37] leverage tools or LLMs to synthesize multimodal data so that VLMs can be trained to distill the skills in specialist models and LLMs. Our work employs knowl" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.168 + ], + "angle": 0, + "content": "edge distillation to endow the holistic embedding module with general vision and language encoding abilities. Since distillation targets can be computed independently, we can use unpaired random image and text data, eliminating the need to collect additional multi-modal data." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.18, + 0.605, + 0.195 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.205, + 0.905, + 0.28 + ], + "angle": 0, + "content": "In this section, we present HoVLE, a monolithic vision-language model with decent vision and language capabilities. In the following sections, we introduce the network architecture and the training procedures for such a unified model. The details are shown in Figure 3." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.288, + 0.719, + 0.303 + ], + "angle": 0, + "content": "3.1. Network Architecture" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.311, + 0.905, + 0.448 + ], + "angle": 0, + "content": "The success of compositional VLMs demonstrates that LLMs are capable of understanding visual information when image embeddings are properly aligned with text embeddings. Inspired by this, HoVLE proposes that a monolithic VLM can be built using a holistic embedding module. This module takes mixed raw images and texts as inputs and projects them into a shared embedding space. In this way, LLMs can interpret images in the same way as texts, resulting in a monolithic VLM." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.454, + 0.906, + 0.53 + ], + "angle": 0, + "content": "Holistic Embedding Module. We first introduce the holistic embedding module to project vision and language inputs to a shared embedding space. Specifically, given an input image \\(I\\) and texts \\(T\\), we first transform them into vector forms via simple tokenization and linear projection:" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.539, + 0.904, + 0.554 + ], + "angle": 0, + "content": "\\[\nx _ {I} = \\operatorname {P a t c h E m b e d} (\\operatorname {D y n P r o c e s s} (I)) + \\mathrm {P E}, \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.558, + 0.905, + 0.573 + ], + "angle": 0, + "content": "\\[\nx _ {T} = \\operatorname {T e x t E m b e d} (\\operatorname {T o k e n i z e r} (T)), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.581, + 0.906, + 0.776 + ], + "angle": 0, + "content": "where DynProcess(\\cdot) is the dynamic high-resolution strategy from InternVL [11]. This process divides the image into \\(448 \\times 448\\) tiles and appends a global thumbnail to provide overall context. Similar operations have been adopted in many compositional and monolithic VLMs [12, 18, 51] and have proved beneficial for model performance. PatchEmbed(\\cdot) refers to a patch embedding layer with stride \\(s\\) and converts the image tiles into a sequence \\(x_{I}\\) of \\(n_I\\) image tokens with dimension \\(c\\). \\(\\mathrm{PE} \\in \\mathbb{R}^{n_I \\times c}\\) is a learnable position embedding. Tokenizer(\\cdot) converts raw texts into token indices and TextEmbed(\\cdot) is a word embedding layer with vocabulary size of \\(v\\) and dimension of \\(c\\). \\(x_{T}\\) is a sequence of \\(n_T\\) text tokens." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.778, + 0.905, + 0.839 + ], + "angle": 0, + "content": "Afterward, these image tokens and text tokens are combined as the input tokens \\( x \\in \\mathbb{R}^{n \\times c} \\). The holistic embedding module utilizes a stack of causal Transformer layers to project these tokens to a shared embedding space:" + }, + { + "type": "equation", + "bbox": [ + 0.616, + 0.847, + 0.905, + 0.863 + ], + "angle": 0, + "content": "\\[\n\\hat {x} = \\text {C a u s a l T r a n s f o r m e r} (x), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "where \\(\\hat{x}\\) denotes final embeddings of the input. The architecture of each Transformer layer is the same as the LLM's" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14561" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.123, + 0.088, + 0.45, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.251, + 0.463, + 0.333, + 0.475 + ], + "angle": 0, + "content": "(a) Architecture" + }, + { + "type": "image", + "bbox": [ + 0.462, + 0.089, + 0.868, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.611, + 0.463, + 0.725, + 0.476 + ], + "angle": 0, + "content": "(b) Training Strategies" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.479, + 0.907, + 0.563 + ], + "angle": 0, + "content": "Figure 3. (a) The architecture of HoVLE. HoVLE initially segments the input images into patches dynamically and tokenizes input texts. The holistic embedding module then projects them into a unified space. Finally, the LLM processes these unified embeddings to produce the final outputs. (b) The training strategies of HoVLE. Distillation stage trains the holistic embedding to distill a pre-trained vision encoder and text embeddings of the LLM using unpaired random images and texts. Alignment stage combines the holistic embedding module with a frozen LLM, conducting auto-regressive training to align the vision-language embeddings. Instruction tuning further enhance HoVLE's overall ability by tuning the whole model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.571, + 0.483, + 0.662 + ], + "angle": 0, + "content": "for simplicity and unity. Similar to LLMs, these Transformer layers use causal attention for both image and text tokens. The effectiveness of this architecture suggests that 1D causal modeling works as well as non-causal modeling for images in multi-modal tasks, aligning with recent findings in [42]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.67, + 0.484, + 0.746 + ], + "angle": 0, + "content": "Monolithic Vision-Language Model. By combining the holistic embedding module with a pre-trained LLM, a monolithic VLM can be naturally built. After the holistic embedding module generates input embeddings, the LLM takes them as input and performs next-token prediction:" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.762, + 0.483, + 0.779 + ], + "angle": 0, + "content": "\\[\np _ {n + 1} = \\operatorname {L L M} (\\hat {x}; \\theta_ {\\text {L L M}}), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.793, + 0.483, + 0.824 + ], + "angle": 0, + "content": "where \\(\\theta_{\\mathrm{LLM}}\\) refers to parameters of LLM, \\(p_{n + 1}\\in \\mathbb{R}^v\\) is the probability of the next token." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.9 + ], + "angle": 0, + "content": "In this way, we replace the modality-specific encoders in compositional VLMs with a holistic embedding module. This module has the same Transformer layer as the LLM, so the whole model presents a simple architecture, leading to our monolithic VLM." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.57, + 0.701, + 0.587 + ], + "angle": 0, + "content": "3.2. Training Procedure" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.594, + 0.907, + 0.731 + ], + "angle": 0, + "content": "To empower the holistic embedding module with decent vision and language encoding abilities, a multi-stage training procedure is carefully designed. The first stage trains the holistic embedding module to distill the image feature from a pre-trained visual encoder and the text embeddings from an LLM, providing general encoding abilities. The second stage combines the holistic embedding module with the LLM to perform auto-regressive training, aligning different modalities to a shared embedding space." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.736, + 0.907, + 0.843 + ], + "angle": 0, + "content": "Distillation Stage. This stage aims to provide basic vision and language encoding abilities with the holistic embedding module. We extract image and text embeddings from the holistic embedding module output, denoted as \\(\\hat{x}_I\\) and \\(\\hat{x}_T\\), respectively. A pre-trained visual encoder and an LLM embedding layer are utilized to compute the target image feature \\(z_I\\) and text feature \\(z_T\\). The training objective is" + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.853, + 0.907, + 0.876 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta_ {\\text {E m b e d}}} \\mathcal {L} _ {\\text {d i s t i l l}} \\left(\\hat {x} _ {I}, z _ {I}; \\theta_ {\\text {E m b e d}}\\right) + \\mathcal {L} _ {\\text {d i s t i l l}} \\left(\\hat {x} _ {T}, z _ {T}; \\theta_ {\\text {E m b e d}}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.886, + 0.907, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{distill}}\\) is the negative cosine similarity between two" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14562" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.122 + ], + "angle": 0, + "content": "features, \\(\\theta_{\\mathrm{Embed}}\\) refers to the parameters of the holistic embedding module." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.128, + 0.483, + 0.248 + ], + "angle": 0, + "content": "Unpaired Random Images and Texts for Distillation. The distillation stage does not put any requirement on input images and texts, because their target features are computed independently. Therefore, we propose to combine random images with random texts as input. In fact, this strategy proves to be more effective than using paired image-text inputs, potentially due to better vocabulary coverage during distillation, as demonstrated in Section 4.3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.249, + 0.484, + 0.385 + ], + "angle": 0, + "content": "In practice, we randomly sample 500M images from LAION-2B [39] dataset and randomly sample 100 text tokens from the vocabulary for each image as training data. Each sample consists of a user query and a model response. We place the image at the beginning of the query, followed by the text content for the query and response, each with 50 random text tokens. Unlike previous monolithic VLMs, HoVLE can benefit from large-scale vision pre-training, unleashing the full potential of the monolithic VLM." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.391, + 0.484, + 0.451 + ], + "angle": 0, + "content": "Alignment Stage. To align the image and text embeddings, the holistic embedding module is further combined with the LLM to perform auto-regressive training. The training objective is" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.457, + 0.483, + 0.488 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta_ {\\text {E m b e d}}} \\sum_ {i} \\mathcal {L} _ {\\text {a l i g n}} \\left(p _ {i}, y _ {i}; \\theta_ {\\text {E m b e d}}, \\theta_ {\\text {L L M}}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.494, + 0.483, + 0.554 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{align}}\\) is the cross-entropy loss, \\(p_i\\) and \\(y_{i}\\) denote the output probability and target index of the \\(i^{\\mathrm{th}}\\) text token. During training, the LLM is frozen, and only the holistic embedding module is optimized." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.555, + 0.484, + 0.645 + ], + "angle": 0, + "content": "This stage is similar to the alignment stage used in compositional VLMs, but we do not have an extra MLP connector. During alignment, the model is trained on 45M multi-modal data from diverse domains, such as captioning, detection, and OCR data. Please refer to Appendix A for details." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.652, + 0.273, + 0.669 + ], + "angle": 0, + "content": "3.3. Instruction Tuning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.736 + ], + "angle": 0, + "content": "The holistic embedding module is now empowered with strong vision and language encoding abilities. To further strengthen the whole VLM to follow instructions, a visual instruction tuning stage is incorporated." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.736, + 0.483, + 0.811 + ], + "angle": 0, + "content": "In this stage, the whole model is optimized using the same auto-regressive training objective as in the alignment stage. The training data contains around 5M data, covering visual question answering, mathematics, world knowledge, etc. Please refer to Appendix A for details." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.822, + 0.216, + 0.839 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.848, + 0.308, + 0.864 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Network Architecture. HoVLE adopts InternLM2-1.8B [7] as the LLM backbone. The holistic embedding is" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.243 + ], + "angle": 0, + "content": "composed of 8 causal Transformer layers, with hidden size 2048 and 16 attention heads, as in InternLM2-1.8B. The image patch embedding layer has a stride of \\( s = 28 \\) and output dimension \\( c = 2048 \\). The text embedding has a vocabulary of 92,553. To demonstrate the high-resolution image processing ability of our model, like other works [5, 12, 24], we also provide HoVLE (HD). It uses a patch embedding with stride \\( s = 14 \\) and applies pixel shuffle [11] after the holistic embedding to reduce the number of visual tokens to 1/4. All other details are kept the same." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.25, + 0.905, + 0.432 + ], + "angle": 0, + "content": "Training Procedure. During distillation, we randomly sample 500M images from LAION-2B [39] with 100 random text indices for each image. We use InternViT-300M [9, 11] and the text embedding from InternLM2-1.8B as the teacher model for image and text tokens, respectively. For alignment, a collection of datasets containing 50M samples is used, as shown in Appendix A. For instruction tuning, a collection of datasets containing 5M samples is used, as shown in Appendix A. We use AdamW as the optimizer for all three stages, with learning rates 3e-4, 5e-5, and 4e-5, respectively. The batch size is set to 4096 for all stages. Please refer to Appendix A for more hyper-parameters." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.439, + 0.907, + 0.712 + ], + "angle": 0, + "content": "Evaluation Benchmarks. We evaluate our model on 17 public benchmarks. MMbench [30], MME [14], MMMU [55], MM-Vet [54], and SEED [21] focus on multimodal perception and reasoning, providing a comprehensive assessment of VLM capabilities. POPE [23] and HallB [15] specifically target hallucination detection. For OCR and OCR-related visual question answering, TextVQA [40], ChartQA [33], InfoVQA [35], DocVQA [34], and OCRBench [29] are used. GQA [19] evaluates visual scene understanding, while ScienceQA [32] and AI2D [20] assess scientific image understanding. MathVista focuses on mathematical reasoning. For specific metric values shown in the experiments, MMbench-EN results are counted and presented as the MMB score. The POPE score is averaged across its three categories, the MME score is calculated as the sum of perception and cognition scores, and the SEED score reflects image accuracy specifically." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.725, + 0.652, + 0.74 + ], + "angle": 0, + "content": "4.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We compare HoVLE with previous compositional and monolithic VLMs in Table 1 and Table 2. We highlight several key observations as follows: (1) There are significant performance gaps between previous monolithic VLMs and compositional VLMs, indicating the challenges in building a monolithic VLM with state-of-the-art performance. (2) Compared with existing monolithic VLMs, our HoVLE demonstrates substantially better performance with fewer parameters. For example, HoVLE surpasses the previous best monolithic VLM by \\(\\sim 15\\) points on MMBench, which" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14563" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.182, + 0.089, + 0.813, + 0.386 + ], + "angle": 0, + "content": "
Method# ParamsMMBMMEMMMUMM-VetPOPESEEDMathVistaHallBCCB
Compositional VLMs
VILA1.5-3B [25]3.1B63.4164833.335.485.967.931.631.224.1
DeepSeek-VL-1.3B [31]2.0B64.6153232.234.887.666.731.127.637.6
PaliGemma-3B [6]2.9B71.0168634.933.187.069.628.732.229.6
MM1-3B-Chat [36]3B67.8176233.943.787.468.832.0--
MiniCPM-V [53]2.8B64.1165038.331.179.565.628.936.241.4
MiniCPM-V-2 [53]2.8B69.1180938.241.086.367.138.736.145.3
InternVL1.5 [9]2.2B70.9190234.639.387.169.841.137.563.5
†InternVL2 [10]2.2B73.2187734.344.688.371.646.437.974.7
Qwen2-VL-2B [46]2.2B74.9187241.149.5--43.041.7-
Phi-3 Vision [1]4.2B80.5-40.4-85.8-44.5--
Monolithic VLMs
Fuyu-8B (HD) [5]8B10.7--21.474.1----
SOLO [8]7B-1260---64.434.4--
*Chameleon-7B [43]7B31.117025.48.3-30.622.317.13.5
EVE-7B [12]7B49.5148332.325.683.661.325.221.112.4
EVE-7B (HD) [12]7B52.3162832.625.785.064.634.226.416.3
Emu3 [49]8B58.5-31.637.285.268.2---
Our HoVLE2.6B71.9186433.744.387.670.746.239.675.3
Our HoVLE (HD)2.6B73.3186232.243.887.470.949.238.474.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.393, + 0.907, + 0.448 + ], + "angle": 0, + "content": "Table 1. Comparison with existing VLMs on general VLM benchmarks. The highest scores in monolithic VLMs are highlighted in bold. *We test Chameleon-7B with its released model, but it frequently refuses to perform the task with a response of \"I can't help you with this\", thus resulting in poor performance. \\( {}^{ \\dagger } \\) InternVL2 2.2B adopts the same LLM and supervised data with our HoVLE,sowe mark it as the compositional counterpart." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.452, + 0.483, + 0.724 + ], + "angle": 0, + "content": "is a comprehensive multi-modal benchmark covering a variety of fine-grained abilities. This substantial improvement is attributed to the holistic embedding, which prevents continual pre-training from disrupting language knowledge stored in pre-trained LLMs. (3) By incorporating higher resolution, HoVLE (HD) consistently improves the performance on VQA benchmarks, demonstrating the versatility of our holistic embedding in handling different resolutions. (4) Compared with previous compositional VLMs, HoVLE (HD) is able to deliver competitive results with many of these models. This shows that the introduction of holistic embedding provides an effective encoding that allows the LLM to understand images while preserving its language abilities. While HoVLE (HD) still falls short of state-of-the-art compositional VLMs, like Qwen2-VL2B [46] and Phi-3 Vision [1], it narrows the gap between monolithic and compositional VLMs, suggesting the feasibility of a high-performance monolithic VLM." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.735, + 0.254, + 0.75 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.758, + 0.483, + 0.82 + ], + "angle": 0, + "content": "In this section, we aim to display and discuss the design choices that affect our HoVLE. Ablation experiments train the model with 300M and 8M data during the distillation and alignment stages respectively, unless specified." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Holistic Embedding Depth. To validate the effectiveness of the holistic embedding, we first compare models with different holistic embedding depths in Table 3. With 0 Transformer layer, the model degenerates to a VLM without holistic embedding and the poor performance suggests" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.452, + 0.907, + 0.528 + ], + "angle": 0, + "content": "the importance of this module. The model performs consistently better as the depth goes from 4 to 12. We speculate that deep embedding is beneficial for extracting visual representation. HoVLE uses 8 layers to achieve a balance between performance and model complexity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.531, + 0.909, + 0.713 + ], + "angle": 0, + "content": "Training Strategy. Table 4 displays the influence of different training stages. The results show that omitting any stage will result in a performance drop. Among these, removing the distillation stage leads to the most severe degradation, highlighting its importance. We observe that models after distillation exhibit a faster decrease in loss during subsequent training stages, likely because the distillation provides general abilities of representation extraction, which is helpful for accelerating training. The improvement from instruction tuning is greater than that of the alignment stage. We assume this is because the LLM has more trainable parameters, allowing the whole model to fit the data well." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.716, + 0.909, + 0.852 + ], + "angle": 0, + "content": "Distillation Data Scaling. Figure 6 illustrates how the performance changes as distillation data scales up. Overall, the model performs better with more distillation data, implying the importance of comprehensive training. Specifically, the model's performance improves rapidly with 300M data and continues to increase at a more gradual pace as more data is added. We use 500M distillation data as our default setting. Please refer to Appendix B for scaling performance on more benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Text Data for Distillation. Table 5 presents the effects of different text input choices during distillation. The LAION [39] dataset provides a short text caption for each" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14564" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.169, + 0.089, + 0.827, + 0.386 + ], + "angle": 0, + "content": "
Method# ParamsTextVQAChartVQAInfoVQADocVQAOCRbenchGQASQA-IAI2D
Compositional VLMs
VILA1.5-3B [25]3.1B60.4---43761.569.0-
DeepSeek-VL-1.3B [31]2.0B57.8---409--51.5
PaliGemma-3B [6]2.9B68.1---614--68.3
MM1-3B-Chat [36]3B71.9-----69.4-
MiniCPM-V [53]2.8B60.6--38.2366--56.3
MiniCPM-V-2 [53]2.8B74.1--71.9605--62.9
InternVL1.5 [9]2.2B70.574.855.485.065461.684.969.8
†InternVL2 [10]2.2B73.476.257.785.978461.084.974.1
Qwen2-VL-2B [46]2.2B79.773.565.590.1794---
Phi3-Vision [1]4.2B70.981.4----90.876.7
Monolithic VLMs
Fuyu-8B (HD) [5]8B-------64.5
SOLO [8]7B------73.361.4
*Chameleon-7B [43]7B4.82.95.01.57-47.246.0
EVE-7B [12]7B51.919.520.022.032760.863.048.5
EVE-7B (HD) [12]7B56.859.125.053.039862.664.959.1
Emu3 [49]8B64.768.643.876.368760.389.270.0
Our HoVLE2.6B66.077.652.184.773563.694.873.7
Our HoVLE (HD)2.6B70.978.655.786.174064.994.873.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.387, + 0.907, + 0.416 + ], + "angle": 0, + "content": "Table 2. Comparison with existing VLMs on visual question answering benchmarks. The highest scores in monolithic VLMs are highlighted in bold. Same as Tab. 1, *Chameleon-7B frequently refuses to response, and \\( {}^{ \\dagger } \\) InternVL2 2.2B is our compositional counterpart." + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.422, + 0.303, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.306, + 0.422, + 0.495, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.422, + 0.688, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.692, + 0.422, + 0.88, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.35, + 0.525, + 0.647, + 0.538 + ], + "angle": 0, + "content": "Figure 4. Distillation data scaling performance." + }, + { + "type": "table", + "bbox": [ + 0.115, + 0.549, + 0.46, + 0.632 + ], + "angle": 0, + "content": "
DepthMMBMMESEEDTextVQAInfoVQADocVQA
035.4114946.011.319.112.2
467.8180168.047.742.673.6
872.0186270.962.051.482.1
1272.3188270.965.552.584.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.14, + 0.632, + 0.434, + 0.645 + ], + "angle": 0, + "content": "Table 3. Ablation on holistic embedding depth." + }, + { + "type": "table", + "bbox": [ + 0.102, + 0.652, + 0.477, + 0.732 + ], + "angle": 0, + "content": "
Dist. Align. I.T.MMBMMESEEDTextVQAInfoVQADocVQA
✓✓37.3127446.411.219.3
✓✓71.3182870.560.149.2
✓✓66.1169866.755.043.0
✓✓✓72.0186270.962.051.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.162, + 0.734, + 0.412, + 0.747 + ], + "angle": 0, + "content": "Table 4. Ablation on training strategies." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.753, + 0.47, + 0.808 + ], + "angle": 0, + "content": "
Text DataMMBMMESEEDTextVQAInfoVQADocVQA
Paired71.7176870.561.349.481.5
Random72.0186270.962.051.482.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.145, + 0.809, + 0.429, + 0.82 + ], + "angle": 0, + "content": "Table 5. Ablation on text data for distillation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.482, + 0.901 + ], + "angle": 0, + "content": "image, so we can also use image-text pairs as input. However, this approach yields relatively lower results, and the model continues to exhibit high loss values during alignment. We hypothesize that the text data from the LAION" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.907, + 0.674 + ], + "angle": 0, + "content": "dataset lacks sufficient diversity to cover the full vocabulary, suggesting that paired data may require higher quality and variety in texts. In contrast, we propose using random texts as input. Table 5 shows that this strategy outperforms paired text data. We also observe lower loss during alignment. By leveraging random text, we reduce the high-quality need for image-text pair collection, thereby enhancing data utilization for large-scale training." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.682, + 0.615, + 0.698 + ], + "angle": 0, + "content": "4.4. Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.907, + 0.856 + ], + "angle": 0, + "content": "Visualization. To better understand the effects of our introduced holistic embedding, we visualize the attention map of the LLM in Figure 5. We find that existing monolithic models, such as EVE [12] and Emu3 [49], exhibit denser text-to-image attention in their initial layers, while the compositional InternVL2 [10] and our HoVLE model maintain sparse attention patterns throughout. All models demonstrate sparse attention patterns in the later layers (with the final layers shown). Please refer to Appendix C for more attention maps in the middle layers." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "We hypothesize that sparser text-to-image attention indicates that visual features are effectively extracted and well-aligned with text tokens, allowing the text tokens to focus" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14565" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.159, + 0.089, + 0.318, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.089, + 0.473, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.496, + 0.089, + 0.631, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.09, + 0.787, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.795, + 0.099, + 0.84, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.158, + 0.224, + 0.315, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.338, + 0.224, + 0.473, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.224, + 0.631, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.655, + 0.224, + 0.788, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.345, + 0.907, + 0.415 + ], + "angle": 0, + "content": "Figure 5. Attention Maps for EVE, Emu3, InternVL2 and our HoVLE at the first and last layers of LLM backbones. Y-axis represents query tokens, and X-axis represents key tokens, with text modality tokens in gray and image modality tokens in yellow. All four models share the same input, but the sequence lengths of input tokens are different due to different image pre-processing. We highlight text-to-image attention below each full attention map. Our HoVLE, like the compositional InternVL2, has sparse attention across all network layers, while other monolithic models Emu3 and EVE have denser attention in shallow layers." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.419, + 0.49, + 0.515 + ], + "angle": 0, + "content": "
ModelTile ResolutionFLOPs(T)↓TTFT↓TPS↑Avg. Score↑
InternVL2448 × 4481.811.3728165.8
HoVLE (HD)336 × 3361.511.2424365.7
HoVLE (HD)392 × 3921.891.5222466.1
HoVLE (HD)448 × 4482.181.8420666.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.521, + 0.486, + 0.59 + ], + "angle": 0, + "content": "Table 6. Inference speed-performance trade-off of HoVLE (HD). \"TTFT\" and \"TPS\" denotes the time to first token in seconds and throughput in tokens per second, respectively. \"Avg. Score\" refers to the average performance over 17 benchmarks. Please refer to Appendix for detailed results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.597, + 0.483, + 0.688 + ], + "angle": 0, + "content": "only on the most relevant visual information. In contrast, denser text-to-image attention may suggest that the model is still in the process of extracting visual features within the LLM component. Therefore, by introducing holistic embedding, the LLM may process vision and language inputs more effectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Speed-Performance Trade-off. We report the speed-performance trade-off of HoVLE (HD) in Table 6. For evaluation, we use one image and 256 text tokens as input, with the model restricted to output 120 tokens during testing. InternVL2 serves as the main comparison model, as it uses the same LLM. When the tile resolution of DynProcess(\\(\\cdot\\)) is 336, HoVLE (HD) achieves comparable performance to InternVL2, while also showing lower computational demand. As the input resolution increases, HoVLE's performance improves consistently, but at the cost of slower speed. Additionally, we observe that using the KV cache [38] during inference significantly reduces computational time, making kernel launch the main bottleneck in TPS. Optimizing the kernel could further enhance processing speed." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.42, + 0.634, + 0.436 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.445, + 0.907, + 0.747 + ], + "angle": 0, + "content": "This work presents HoVLE, a monolithic Vision-Language Model (VLM) that processes images and texts in a unified manner. HoVLE introduces a holistic embedding module that projects image and text inputs into a shared embedding space. This allows the Large Language Model (LLM) to interpret images in the same way as texts. Furthermore, a multi-stage training procedure is designed to strengthen the encoding abilities of the holistic embedding. A distillation stage first provides general vision and language encoding capabilities, while an alignment stage aligns different modalities. An instruction tuning stage enhances the overall abilities in the end. Extensive experiments show that HoVLE surpasses previous monolithic VLMs and demonstrates competitive performance with compositional VLMs. This work focuses on enhancing monolithic VLMs by introducing a holistic embedding module, aiming to elevate their capabilities rather than rival those of compositional VLMs. This work narrows the gap between monolithic and compositional VLMs, providing a promising direction for the development of monolithic VLMs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.748, + 0.907, + 0.822 + ], + "angle": 0, + "content": "Limitations. Due to limited computational resources, our experiments are conducted on 2B parameter scale. Scaling to larger scales can provide further insight into the model scalability of our approach, which we hope to explore in future work." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.823, + 0.907, + 0.884 + ], + "angle": 0, + "content": "Acknowledgements. This work is supported by the National Key R&D Program of China (NO. 2022ZD0161300), by the National Natural Science Foundation of China (62376134)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14566" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Marah Abdin, Jyoti Aneja, Hany Awadalla, Ahmed Awadallah, Ammar Ahmad Awan, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Jianmin Bao, Harkirat Behl, et al. Phi-3 technical report: A highly capable language model locally on your phone. arxiv:2404.14219, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.186, + 0.484, + 0.241 + ], + "angle": 0, + "content": "[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arxiv:2303.08774, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.242, + 0.482, + 0.269 + ], + "angle": 0, + "content": "[3] AI Anthropic. The claude 3 model family: Opus, sonnet, haiku. Claude-3 Model Card, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.271, + 0.482, + 0.325 + ], + "angle": 0, + "content": "[4] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arxiv:2308.12966, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.327, + 0.482, + 0.382 + ], + "angle": 0, + "content": "[5] Rohan Bavishi, Erich Elsen, Curtis Hawthorne, Maxwell Nye, Augustus Odena, Arushi Somani, and Sagnak Tasirlar. Introducing our multimodal models. https://www. adept.ai/blog/fuyu-8b, 2023. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.383, + 0.482, + 0.451 + ], + "angle": 0, + "content": "[6] Lucas Beyer, Andreas Steiner, André Susano Pinto, Alexander Kolesnikov, Xiao Wang, Daniel Salz, Maxim Neumann, Ibrahim Alabdulmohsin, Michael Tschannen, Emanuele Bugliarello, et al. Paligemma: A versatile 3b vlm for transfer. arxiv:2407.07726, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.453, + 0.482, + 0.507 + ], + "angle": 0, + "content": "[7] Zheng Cai, Maosong Cao, Haojiong Chen, Kai Chen, Keyu Chen, Xin Chen, Xun Chen, Zehui Chen, Zhi Chen, Pei Chu, et al. Internl m2 technical report. arxiv:2403.17297, 2024. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.509, + 0.482, + 0.55 + ], + "angle": 0, + "content": "[8] Yangyi Chen, Xingyao Wang, Hao Peng, and Heng Ji. A single transformer for scalable vision-language modeling. arxiv:2407.06438, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.552, + 0.482, + 0.62 + ], + "angle": 0, + "content": "[9] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arxiv:2404.16821, 2024. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.622, + 0.482, + 0.718 + ], + "angle": 0, + "content": "[10] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. Internv12: Better than the best—expanding performance boundaries of open-source multimodal models with the progressive scaling strategy. https://internvl.github.io/blog/2024-07-02-InternVL-2.0/, 2024.6.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.72, + 0.482, + 0.789 + ], + "angle": 0, + "content": "[11] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In CVPR, pages 24185–24198, 2024. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.79, + 0.482, + 0.844 + ], + "angle": 0, + "content": "[12] Haiwen Diao, Yufeng Cui, Xiaotong Li, Yueze Wang, Huchuan Lu, and Xinlong Wang. Unveiling encoder-free vision-language models. arxiv:2406.11832, 2024. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[13] Zhiyuan Fang, Jianfeng Wang, Xiaowei Hu, Lijuan Wang, Yezhou Yang, and Zicheng Liu. Compressing visual-linguistic model via knowledge distillation. In ICCV, pages 1428-1438, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[14] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. arxiv:2306.13394, 2023.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.163, + 0.905, + 0.244 + ], + "angle": 0, + "content": "[15] Tianrui Guan, Fuxiao Liu, Xiyang Wu, Ruiqi Xian, Zongxia Li, Xiaoyu Liu, Xijun Wang, Lichang Chen, Furong Huang, Yaser Yacoob, et al. Hallusionbench: An advanced diagnostic suite for entangled language hallucination and visual illusion in large vision-language models. arxiv:2310.14566, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.246, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[16] Suchin Gururangan, Ana Marasovic, Swabha Swayamdipta, Kyle Lo, Iz Beltagy, Doug Downey, and Noah A Smith. Don't stop pretraining: Adapt language models to domains and tasks. In ACL, pages 8342-8360, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.303, + 0.905, + 0.37 + ], + "angle": 0, + "content": "[17] Yushi Hu, Otilia Stretcu, Chun-Ta Lu, Krishnamurthy Viswanathan, Kenji Hata, Enming Luo, Ranjay Krishna, and Ariel Fuxman. Visual program distillation: Distilling tools and programmatic reasoning into vision-language models. In CVPR, pages 9590-9601, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.372, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[18] Runhui Huang, Xinpeng Ding, Chunwei Wang, Jianhua Han, Yulong Liu, Hengshuang Zhao, Hang Xu, Lu Hou, Wei Zhang, and Xiaodan Liang. Hires-llava: Restoring fragmentation input in high-resolution large vision-language models. arXiv:2407.08706, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.442, + 0.905, + 0.483 + ], + "angle": 0, + "content": "[19] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In CVPR, pages 6700-6709, 2019. 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.484, + 0.905, + 0.537 + ], + "angle": 0, + "content": "[20] Aniruddha Kembhavi, Mike Salvato, Eric Kolve, Minjoon Seo, Hannaneh Hajishirzi, and Ali Farhadi. A diagram is worth a dozen images. In ECCV, pages 235-251. Springer, 2016. 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.539, + 0.905, + 0.593 + ], + "angle": 0, + "content": "[21] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arxiv:2307.16125, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.595, + 0.905, + 0.65 + ], + "angle": 0, + "content": "[22] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, pages 12888-12900. PMLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.651, + 0.905, + 0.704 + ], + "angle": 0, + "content": "[23] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. In EMNLP, pages 292-305, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.706, + 0.905, + 0.773 + ], + "angle": 0, + "content": "[24] Zhang Li, Biao Yang, Qiang Liu, Zhiyin Ma, Shuo Zhang, Jingxu Yang, Yabo Sun, Yuliang Liu, and Xiang Bai. Monkey: Image resolution and text label are important things for large multi-modal models. In CVPR, pages 26763-26773, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.776, + 0.905, + 0.817 + ], + "angle": 0, + "content": "[25] Ji Lin, Hongxu Yin, Wei Ping, Pavlo Molchanov, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models. In CVPR, pages 26689-26699, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.818, + 0.905, + 0.859 + ], + "angle": 0, + "content": "[26] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In CVPR, pages 26296-26306, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.86, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[27] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge. https :" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14567" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "//llava-v1.github.io/blog/2024-01-30-llava-last/, 2024.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.122, + 0.482, + 0.149 + ], + "angle": 0, + "content": "[28] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.151, + 0.482, + 0.207 + ], + "angle": 0, + "content": "[29] Yuliang Liu, Zhang Li, Biao Yang, Chunyuan Li, Xucheng Yin, Cheng-lin Liu, Lianwen Jin, and Xiang Bai. On the hidden mystery ofOCR in large multimodal models. arxiv:2305.07895, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.208, + 0.482, + 0.276 + ], + "angle": 0, + "content": "[30] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? In ECCV, pages 216-233. Springer, 2025. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.334 + ], + "angle": 0, + "content": "[31] Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhuoshu Li, Hao Yang, et al. Deepseek-vl: towards real-world vision-language understanding. arxiv:2403.05525, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.336, + 0.482, + 0.405 + ], + "angle": 0, + "content": "[32] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. In NeurIPS, pages 2507–2521, 2022. 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.407, + 0.482, + 0.462 + ], + "angle": 0, + "content": "[33] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arxiv:2203.10244, 2022. 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.463, + 0.482, + 0.506 + ], + "angle": 0, + "content": "[34] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In WACV, pages 2200-2209, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.507, + 0.482, + 0.548 + ], + "angle": 0, + "content": "[35] Minesh Mathew, Viraj Bagal, Ruben Tito, Dimosthenis Karatzas, Ernest Valveny, and CV Jawahar. Infographicvqa. In WACV, pages 1697-1706, 2022. 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.549, + 0.482, + 0.619 + ], + "angle": 0, + "content": "[36] Brandon McKinzie, Zhe Gan, Jean-Philippe Fauconnier, Sam Dodge, Bowen Zhang, Philipp Dufter, Dhruti Shah, Xianzhi Du, Futang Peng, Floris Weers, et al. Mm1: Methods, analysis & insights from multimodal llm pre-training. arxiv:2403.09611, 2024. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.62, + 0.482, + 0.689 + ], + "angle": 0, + "content": "[37] Jae Sung Park, Jack Hessel, Khyathi Chandu, Paul Pu Liang, Ximing Lu, Peter West, Youngjae Yu, Qiuyuan Huang, Jianfeng Gao, Ali Farhadi, et al. Localized symbolic knowledge distillation for visual commonsense models. In NeurIPS, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.69, + 0.482, + 0.759 + ], + "angle": 0, + "content": "[38] Reiner Pope, Sholto Douglas, Aakanksha Chowdhery, Jacob Devlin, James Bradbury, Jonathan Heek, Kefan Xiao, Shivani Agrawal, and Jeff Dean. Efficiently scaling transformer inference. Proceedings of Machine Learning and Systems, 5: 606-624, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.482, + 0.844 + ], + "angle": 0, + "content": "[39] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS, pages 25278–25294, 2022. 5, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[40] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In CVPR, pages 8317-8326, 2019. 5, 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[41] Quan Sun, Yufeng Cui, Xiaosong Zhang, Fan Zhang, Qiying Yu, Yueze Wang, Yongming Rao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative multimodal models are in-context learners. In CVPR, pages 14398-14409, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[42] Chenxin Tao, Xizhou Zhu, Shiqian Su, Lewei Lu, Changyao Tian, Xuan Luo, Gao Huang, Hongsheng Li, Yu Qiao, Jie Zhou, et al. Learning 1d causal visual representation with de-focus attention networks. arxiv:2406.04342, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[43] Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models. arxiv:2405.09818, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[44] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arxiv:2312.11805, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[45] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arxiv:2302.13971, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.46 + ], + "angle": 0, + "content": "[46] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arxiv:2409.12191, 2024. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.462, + 0.905, + 0.517 + ], + "angle": 0, + "content": "[47] Tiannan Wang, Wangchunshu Zhou, Yan Zeng, and Xinsong Zhang. Efficientvlm: Fast and accurate vision-language models via knowledge distillation and modal-adaptive pruning. arxiv:2210.07795, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.518, + 0.905, + 0.574 + ], + "angle": 0, + "content": "[48] Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong Lu, Jie Zhou, Yu Qiao, et al. Vision: Large language model is also an open-ended decoder for vision-centric tasks. In NeurIPS, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.905, + 0.631 + ], + "angle": 0, + "content": "[49] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arxiv:2409.18869, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.632, + 0.905, + 0.659 + ], + "angle": 0, + "content": "[50] X.ai. Grok-1.5 vision preview. https://x.ai/blog/grok-1.5v, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.905, + 0.729 + ], + "angle": 0, + "content": "[51] Ruyi Xu, Yuan Yao, Zonghao Guo, Junbo Cui, Zanlin Ni, Chunjiang Ge, Tat-Seng Chua, Zhiyuan Liu, Maosong Sun, and Gao Huang. Llava-uhd: an Imm perceiving any aspect ratio and high-resolution images. arXiv:2403.11703, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.732, + 0.905, + 0.787 + ], + "angle": 0, + "content": "[52] Zhengyuan Yang, Linjie Li, Kevin Lin, Jianfeng Wang, Chung-Ching Lin, Zicheng Liu, and Lijuan Wang. The dawn of Imms: Preliminary explorations with gpt-4v (ision). arxiv:2309.17421, 9(1):1, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[53] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arxiv:2408.01800, 2024.6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[54] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. Mm-vet: Evaluating large multimodal models for integrated capabilities. arxiv:2308.02490, 2023. 5" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.093, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14568" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[55] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In CVPR, pages 9556-9567, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[56] Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. A survey of large language models. arxiv:2303.18223, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.288 + ], + "angle": 0, + "content": "[57] Chunting Zhou, Lili Yu, Arun Babu, Kushal Tirumala, Michihiro Yasunaga, Leonid Shamis, Jacob Kahn, Xuezhe Ma, Luke Zettlemoyer, and Omer Levy. Transfusion: Predict the next token and diffuse images with one multi-modal model. arxiv:2408.11039, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.482, + 0.344 + ], + "angle": 0, + "content": "[58] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arxiv:2304.10592, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "14569" + } + ] +] \ No newline at end of file diff --git a/2025/HoVLE_ Unleashing the Power of Monolithic Vision-Language Models with Holistic Vision-Language Embedding/126721ba-0c45-4655-8cc5-9017b39955e6_origin.pdf b/2025/HoVLE_ Unleashing the Power of Monolithic Vision-Language Models with Holistic Vision-Language Embedding/126721ba-0c45-4655-8cc5-9017b39955e6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6eb119051a6553c0352f8dedec3194427f7eac09 --- /dev/null +++ b/2025/HoVLE_ Unleashing the Power of Monolithic Vision-Language Models with Holistic Vision-Language Embedding/126721ba-0c45-4655-8cc5-9017b39955e6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2dbac3186fb08c0efaa7b6b5869a3318fe1ebba7f2b6dc0d982105d78213d61c +size 3607195 diff --git a/2025/HoVLE_ Unleashing the Power of Monolithic Vision-Language Models with Holistic Vision-Language Embedding/full.md b/2025/HoVLE_ Unleashing the Power of Monolithic Vision-Language Models with Holistic Vision-Language Embedding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..25c90aa189421762aa8f54971ddbaf6f62f5fd38 --- /dev/null +++ b/2025/HoVLE_ Unleashing the Power of Monolithic Vision-Language Models with Holistic Vision-Language Embedding/full.md @@ -0,0 +1,321 @@ +# HoVLE: Unleashing the Power of Monolithic Vision-Language Models with Holistic Vision-Language Embedding + +Chenxin Tao $^{1,3*}$ , Shiqian Su $^{1,3*}$ , Xizhou Zhu $^{1,2*}$ , Chenyu Zhang $^{1,2}$ , Zhe Chen $^{4,2}$ , Jiawen Liu $^{6}$ , Wenhai Wang $^{5,2}$ , Lewei Lu $^{3}$ , Gao Huang $^{1}$ , Yu Qiao $^{2}$ , Jifeng Dai $^{1,2}$ + +$^{1}$ Tsinghua University $^{2}$ Shanghai Artificial Intelligence Laboratory + +$^{3}$ SenseTime Research $^{4}$ Nanjing University + +5The Chinese University of Hong Kong 6 Johns Hopkins University + +Model: https://huggingface.co/OpenGVLab/HoVLE + +# Abstract + +The rapid advance of Large Language Models (LLMs) has catalyzed the development of Vision-Language Models (VLMs). Monolithic VLMs, which avoid modality-specific encoders, offer a promising alternative to the compositional ones but face the challenge of inferior performance. Most existing monolithic VLMs require tuning pre-trained LLMs to acquire vision abilities, which may degrade their language capabilities. To address this dilemma, this paper presents a novel high-performance monolithic VLM named HoVLE. We note that LLMs have been shown to be capable of interpreting images when image embeddings are aligned with text embeddings. The challenge for current monolithic VLMs actually lies in the lack of a holistic embedding module for both vision and language inputs. Therefore, HoVLE introduces a holistic embedding module that converts visual and textual inputs into a shared space, allowing LLMs to process images in the same way as texts. Furthermore, a multi-stage training strategy is carefully designed to empower the holistic embedding module. It is first trained to distill visual features from a pre-trained vision encoder and text embeddings from the LLM, enabling largescale training with unpaired random images and text tokens. The whole model further undergoes next-token prediction on multi-modal data to align the embeddings. Finally, an instruction-tuning stage is incorporated. Our experiments show that HoVLE achieves performance close to leading compositional models on various benchmarks, outperforming previous monolithic models by a large margin. + +# 1. Introduction + +The rapid advancement of Large Language Models (LLMs) [2, 4, 7, 45, 56] has significantly accelerated the + +![](images/4d9029fafad4ac06bce0a77104698f6f2b9e9c48ca65696f7cb967e1e5e28bb3.jpg) + +![](images/1665665ec4ecccc53813df1560e205fefa93613dcc1393d7d742d6d0a58098c7.jpg) +Figure 1. Performance comparison on different benchmarks between compositional VLMs (dashed lines) and monolithic VLMs (solid lines). Previous monolithic VLMs exhibit a significant performance gap compared to compositional VLMs, while Our HoVLE demonstrate competitive capabilities with state-of-the-art compositional VLMs. + +development of Vision-Language Models (VLMs) [26, 48, 58], which have demonstrated powerful visual-text processing capabilities. Compositional VLMs are currently the de facto approach for building high-performance VLMs [11, 27, 46], which integrate pre-trained vision encoders with LLMs, as shown in Fig. 2(a). By combining pre-trained visual understanding capabilities with language modeling capabilities, compositional VLMs achieve impressive multimodal performance. On the other hand, there is growing interest in developing VLMs without modality-specific encoders, referred to as monolithic VLMs in this paper. These models process image and text inputs as a unified whole, directly feeding them into the LLM, as illustrated in Fig. 2(b). Due to their monolithic design, these VLMs feature simpler architectures, offering greater potential for + +![](images/34dede1f800a6182dfb96d2c82fec367cdbc76f4baa6dce686ff6b2420b21d36.jpg) +(a) Compositional VLMs + +![](images/3926d4724e775c32cc46513c6434edaca47ebd944fed416d3fb147658cf874b9.jpg) +(b) Existing Monolithic VLMs + +![](images/f07db8dd1e502ad993949f7b4f036b3f9b7317289c19c15d086e1bab869973f7.jpg) +(c) Our HoVLE +Figure 2. Comparison of VLM architectures. (a) Compositional VLMs integrate pre-trained vision encoders with LLMs, using an extra connector to align image and text embeddings. (b) Existing Monolithic VLMs directly feed image and text inputs into LLMs, which require continual pre-training to gain visual abilities. (c) HoVLE uses a holistic embedding module to project image and text input to a unified embedding space, enabling LLMs to interpret images in a text-like manner. Blocks with the same color have the same Transformer layer architecture. + +multimodal tasks, such as unifying generation and recognition [43, 49, 57]. Monolithic VLMs are gaining increasing attention and have shown their potential across many studies [5, 8, 12]. + +However, monolithic VLMs still lag behind compositional VLMs in terms of performance. Some researches [5, 6, 43, 57] explore training monolithic VLMs from scratch using a combination of multi-modal and pure text data. Such practice fails to leverage existing pre-trained models, resulting in substantial computational demands for training. Another research direction [8, 12] proposes injecting visual understanding capabilities into pre-trained LLMs by conducting continual pre-training [16] with large-scale and high-quality multi-modal data. Nevertheless, such continual-tuning methods may interfere with the language knowledge already acquired by pre-trained LLMs, potentially limiting their multi-modal performance [12]. + +The performance limitations of current monolithic VLMs mainly stem from the need to tune pre-trained LLMs. However, the success of compositional VLMs indicates that pre-trained LLMs can interpret visual information without further pre-training, as long as image embeddings are properly aligned with text embeddings. Thus, the reason for such LLM-tuning dilemma of current monolithic VLMs is the lack of a holistic embedding module that can coordinate different modalities. + +To this end, we develop a modality-sharing holistic embedding module as shown in Fig. 2(c), which maps mixed visual and language inputs into a shared embedding space. It can enable pre-trained LLMs to interpret visual input in the same way as text. Similar to compositional VLMs, this not only avoids tuning pre-trained LLMs but also preserves their language proficiency while successfully inte + +grating visual capabilities. Our experiments demonstrate that such a holistic embedding module can indeed achieve high-performance monolithic VLMs. + +In this work, we introduce HoVLE, a novel monolithic vision-language model that learns a Holistic Vision-Language Embedding module to extend LLMs with vision capabilities. The holistic embedding module consists of the same causal Transformer layers as the LLM. It accepts both images and texts as input and projects them into a unified embedding space. These embeddings are then forwarded into the LLM, constituting a monolithic VLM. To empower the holistic embedding module with vision and language encoding abilities, a multi-stage training strategy is proposed. First, the module is trained to distill visual features from a pre-trained vision encoder and text embeddings from the LLM. The distillation stage does not require image-text pairs, but instead uses large-scale unpaired random images and text tokens as inputs. After distillation, the holistic embedding module is combined with the LLM to perform next token prediction on multi-modal data, so that the output embeddings can be further aligned into a unified space. Finally, the whole model is enhanced with an instruction tuning stage. The alignment and instruction tuning stages are similar to those used in compositional VLMs, eliminating the need for extra supervised data. + +Extensive experiments show that our HoVLE is able to achieve close performance with these leading compositional VLMs on 17 multi-modal benchmarks. It also surpasses previous monolithic VLMs by a large margin, e.g., $\sim 15$ points on MMBench [30], showing the effectiveness of HoVLE. In conclusion, our contributions can be summarized as follows: + +- HoVLE is proposed as a novel monolithic VLM that learns a holistic embedding module to expand LLMs with vision capabilities. By projecting vision and language inputs to a shared embedding space, HoVLE can understand images without compromising language proficiency. +- A novel multi-stage training strategy is proposed to empower the holistic embedding module with vision and language encoding abilities. Our training procedure enables HoVLE to develop strong vision and language capabilities without requiring additional supervised data compared to compositional VLMs. +- HoVLE demonstrates comparable performance with compositional VLMs, and surpasses previous monolithic VLMs by a large margin. + +# 2. Related Work + +Compositional Vision-Language Models. The rapid advancements and success of Large Language Models (LLMs) have sparked researchers' interest in multi-modal models, with Vision-Language Models (VLMs) emerging + +as a prominent focus [11, 12, 22, 26, 28, 41]. Compositional VLMs have become the dominant paradigm in the field, where image tokens are typically generated through visual encoders and an additional connector (e.g., MLP projection). Many notable commercial models, including GPT-4V [52], the Gemini series [44], Claude-3V [3], Grok-1.5V [50], and MM1 [36], have demonstrated impressive performance across various multi-modal tasks. In the open-source domain, models like LLaVA [26, 28], BLIP [22], Emu2 [41], InternVL [11], and Qwen-VL [4] stand out. However, compositional VLMs face several challenges, such as complexities in deployment and optimization due to their heterogeneous architecture, trade-offs in the design of network capacity between image and text processing, and potential limitations in cross-modal understanding capabilities [6, 8, 12, 43]. + +Monolithic Vision-Language Models. In contrast to compositional VLMs, monolithic VLMs do not rely on a specific vision encoder. Some works explore training a monolithic VLM from scratch. Fuyu-8B [5] directly inputs image pixel patches into the backbone network using a simple linear projection, while Chameleon [43], Transfusion [57], and Emu3 [49] adopt image tokenizers to map images to discrete tokens as inputs. Nevertheless, these methods require collecting massive pure text and paired image-text data, as well as intensive training computational resources to achieve good performance. Other approaches propose to extend existing pre-trained LLMs to VLMs. SOLO [8] uses a simple linear projection to handle raw image pixels and fine-tunes the pre-trained LLM to develop vision abilities. EVE [12] employs a patch embedding layer with cross-attention to enhance image representation, alongside a patch aligning layer to provide fine-grained visual supervision with pre-trained vision encoders. However, these methods require continually pre-training LLMs, which may diminish their language proficiency. Current monolithic VLMs still underperform in image-text understanding tasks when compared to their compositional counterparts. + +In contrast, by introducing a holistic embedding module, HoVLE can achieve comparable performance with state-of-the-art compositional VLMs, significantly surpassing other monolithic VLMs. + +Knowledge Distillation. Knowledge distillation is an effective technique for transferring knowledge and abilities from a teacher model to a student model. In the realm of VLMs, knowledge distillation has also been successfully incorporated into various tasks. DistillVLM [13] and EfficientVLM [47] use knowledge distillation to compress the model size and accelerate inference speed. VPD [17] and LSKD [37] leverage tools or LLMs to synthesize multimodal data so that VLMs can be trained to distill the skills in specialist models and LLMs. Our work employs knowl + +edge distillation to endow the holistic embedding module with general vision and language encoding abilities. Since distillation targets can be computed independently, we can use unpaired random image and text data, eliminating the need to collect additional multi-modal data. + +# 3. Method + +In this section, we present HoVLE, a monolithic vision-language model with decent vision and language capabilities. In the following sections, we introduce the network architecture and the training procedures for such a unified model. The details are shown in Figure 3. + +# 3.1. Network Architecture + +The success of compositional VLMs demonstrates that LLMs are capable of understanding visual information when image embeddings are properly aligned with text embeddings. Inspired by this, HoVLE proposes that a monolithic VLM can be built using a holistic embedding module. This module takes mixed raw images and texts as inputs and projects them into a shared embedding space. In this way, LLMs can interpret images in the same way as texts, resulting in a monolithic VLM. + +Holistic Embedding Module. We first introduce the holistic embedding module to project vision and language inputs to a shared embedding space. Specifically, given an input image $I$ and texts $T$ , we first transform them into vector forms via simple tokenization and linear projection: + +$$ +x _ {I} = \operatorname {P a t c h E m b e d} (\operatorname {D y n P r o c e s s} (I)) + \mathrm {P E}, \tag {1} +$$ + +$$ +x _ {T} = \operatorname {T e x t E m b e d} (\operatorname {T o k e n i z e r} (T)), \tag {2} +$$ + +where DynProcess(\cdot) is the dynamic high-resolution strategy from InternVL [11]. This process divides the image into $448 \times 448$ tiles and appends a global thumbnail to provide overall context. Similar operations have been adopted in many compositional and monolithic VLMs [12, 18, 51] and have proved beneficial for model performance. PatchEmbed(\cdot) refers to a patch embedding layer with stride $s$ and converts the image tiles into a sequence $x_{I}$ of $n_I$ image tokens with dimension $c$ . $\mathrm{PE} \in \mathbb{R}^{n_I \times c}$ is a learnable position embedding. Tokenizer(\cdot) converts raw texts into token indices and TextEmbed(\cdot) is a word embedding layer with vocabulary size of $v$ and dimension of $c$ . $x_{T}$ is a sequence of $n_T$ text tokens. + +Afterward, these image tokens and text tokens are combined as the input tokens $x \in \mathbb{R}^{n \times c}$ . The holistic embedding module utilizes a stack of causal Transformer layers to project these tokens to a shared embedding space: + +$$ +\hat {x} = \text {C a u s a l T r a n s f o r m e r} (x), \tag {3} +$$ + +where $\hat{x}$ denotes final embeddings of the input. The architecture of each Transformer layer is the same as the LLM's + +![](images/5169ed9ce8be5ff8959ff8e84eafce012c32ca9fc19e612313ae56f1f254e43f.jpg) +(a) Architecture +Figure 3. (a) The architecture of HoVLE. HoVLE initially segments the input images into patches dynamically and tokenizes input texts. The holistic embedding module then projects them into a unified space. Finally, the LLM processes these unified embeddings to produce the final outputs. (b) The training strategies of HoVLE. Distillation stage trains the holistic embedding to distill a pre-trained vision encoder and text embeddings of the LLM using unpaired random images and texts. Alignment stage combines the holistic embedding module with a frozen LLM, conducting auto-regressive training to align the vision-language embeddings. Instruction tuning further enhance HoVLE's overall ability by tuning the whole model. + +![](images/bee8b4481e490d536919aaa562bc4d8c664184d0b2b69941f2a66a9abd788368.jpg) +(b) Training Strategies + +for simplicity and unity. Similar to LLMs, these Transformer layers use causal attention for both image and text tokens. The effectiveness of this architecture suggests that 1D causal modeling works as well as non-causal modeling for images in multi-modal tasks, aligning with recent findings in [42]. + +Monolithic Vision-Language Model. By combining the holistic embedding module with a pre-trained LLM, a monolithic VLM can be naturally built. After the holistic embedding module generates input embeddings, the LLM takes them as input and performs next-token prediction: + +$$ +p _ {n + 1} = \operatorname {L L M} (\hat {x}; \theta_ {\text {L L M}}), \tag {4} +$$ + +where $\theta_{\mathrm{LLM}}$ refers to parameters of LLM, $p_{n + 1}\in \mathbb{R}^v$ is the probability of the next token. + +In this way, we replace the modality-specific encoders in compositional VLMs with a holistic embedding module. This module has the same Transformer layer as the LLM, so the whole model presents a simple architecture, leading to our monolithic VLM. + +# 3.2. Training Procedure + +To empower the holistic embedding module with decent vision and language encoding abilities, a multi-stage training procedure is carefully designed. The first stage trains the holistic embedding module to distill the image feature from a pre-trained visual encoder and the text embeddings from an LLM, providing general encoding abilities. The second stage combines the holistic embedding module with the LLM to perform auto-regressive training, aligning different modalities to a shared embedding space. + +Distillation Stage. This stage aims to provide basic vision and language encoding abilities with the holistic embedding module. We extract image and text embeddings from the holistic embedding module output, denoted as $\hat{x}_I$ and $\hat{x}_T$ , respectively. A pre-trained visual encoder and an LLM embedding layer are utilized to compute the target image feature $z_I$ and text feature $z_T$ . The training objective is + +$$ +\min _ {\theta_ {\text {E m b e d}}} \mathcal {L} _ {\text {d i s t i l l}} \left(\hat {x} _ {I}, z _ {I}; \theta_ {\text {E m b e d}}\right) + \mathcal {L} _ {\text {d i s t i l l}} \left(\hat {x} _ {T}, z _ {T}; \theta_ {\text {E m b e d}}\right), \tag {5} +$$ + +where $\mathcal{L}_{\mathrm{distill}}$ is the negative cosine similarity between two + +features, $\theta_{\mathrm{Embed}}$ refers to the parameters of the holistic embedding module. + +Unpaired Random Images and Texts for Distillation. The distillation stage does not put any requirement on input images and texts, because their target features are computed independently. Therefore, we propose to combine random images with random texts as input. In fact, this strategy proves to be more effective than using paired image-text inputs, potentially due to better vocabulary coverage during distillation, as demonstrated in Section 4.3. + +In practice, we randomly sample 500M images from LAION-2B [39] dataset and randomly sample 100 text tokens from the vocabulary for each image as training data. Each sample consists of a user query and a model response. We place the image at the beginning of the query, followed by the text content for the query and response, each with 50 random text tokens. Unlike previous monolithic VLMs, HoVLE can benefit from large-scale vision pre-training, unleashing the full potential of the monolithic VLM. + +Alignment Stage. To align the image and text embeddings, the holistic embedding module is further combined with the LLM to perform auto-regressive training. The training objective is + +$$ +\min _ {\theta_ {\text {E m b e d}}} \sum_ {i} \mathcal {L} _ {\text {a l i g n}} \left(p _ {i}, y _ {i}; \theta_ {\text {E m b e d}}, \theta_ {\text {L L M}}\right), \tag {6} +$$ + +where $\mathcal{L}_{\mathrm{align}}$ is the cross-entropy loss, $p_i$ and $y_{i}$ denote the output probability and target index of the $i^{\mathrm{th}}$ text token. During training, the LLM is frozen, and only the holistic embedding module is optimized. + +This stage is similar to the alignment stage used in compositional VLMs, but we do not have an extra MLP connector. During alignment, the model is trained on 45M multi-modal data from diverse domains, such as captioning, detection, and OCR data. Please refer to Appendix A for details. + +# 3.3. Instruction Tuning + +The holistic embedding module is now empowered with strong vision and language encoding abilities. To further strengthen the whole VLM to follow instructions, a visual instruction tuning stage is incorporated. + +In this stage, the whole model is optimized using the same auto-regressive training objective as in the alignment stage. The training data contains around 5M data, covering visual question answering, mathematics, world knowledge, etc. Please refer to Appendix A for details. + +# 4. Experiment + +# 4.1. Implementation Details + +Network Architecture. HoVLE adopts InternLM2-1.8B [7] as the LLM backbone. The holistic embedding is + +composed of 8 causal Transformer layers, with hidden size 2048 and 16 attention heads, as in InternLM2-1.8B. The image patch embedding layer has a stride of $s = 28$ and output dimension $c = 2048$ . The text embedding has a vocabulary of 92,553. To demonstrate the high-resolution image processing ability of our model, like other works [5, 12, 24], we also provide HoVLE (HD). It uses a patch embedding with stride $s = 14$ and applies pixel shuffle [11] after the holistic embedding to reduce the number of visual tokens to 1/4. All other details are kept the same. + +Training Procedure. During distillation, we randomly sample 500M images from LAION-2B [39] with 100 random text indices for each image. We use InternViT-300M [9, 11] and the text embedding from InternLM2-1.8B as the teacher model for image and text tokens, respectively. For alignment, a collection of datasets containing 50M samples is used, as shown in Appendix A. For instruction tuning, a collection of datasets containing 5M samples is used, as shown in Appendix A. We use AdamW as the optimizer for all three stages, with learning rates 3e-4, 5e-5, and 4e-5, respectively. The batch size is set to 4096 for all stages. Please refer to Appendix A for more hyper-parameters. + +Evaluation Benchmarks. We evaluate our model on 17 public benchmarks. MMbench [30], MME [14], MMMU [55], MM-Vet [54], and SEED [21] focus on multimodal perception and reasoning, providing a comprehensive assessment of VLM capabilities. POPE [23] and HallB [15] specifically target hallucination detection. For OCR and OCR-related visual question answering, TextVQA [40], ChartQA [33], InfoVQA [35], DocVQA [34], and OCRBench [29] are used. GQA [19] evaluates visual scene understanding, while ScienceQA [32] and AI2D [20] assess scientific image understanding. MathVista focuses on mathematical reasoning. For specific metric values shown in the experiments, MMbench-EN results are counted and presented as the MMB score. The POPE score is averaged across its three categories, the MME score is calculated as the sum of perception and cognition scores, and the SEED score reflects image accuracy specifically. + +# 4.2. Main Results + +We compare HoVLE with previous compositional and monolithic VLMs in Table 1 and Table 2. We highlight several key observations as follows: (1) There are significant performance gaps between previous monolithic VLMs and compositional VLMs, indicating the challenges in building a monolithic VLM with state-of-the-art performance. (2) Compared with existing monolithic VLMs, our HoVLE demonstrates substantially better performance with fewer parameters. For example, HoVLE surpasses the previous best monolithic VLM by $\sim 15$ points on MMBench, which + +
Method# ParamsMMBMMEMMMUMM-VetPOPESEEDMathVistaHallBCCB
Compositional VLMs
VILA1.5-3B [25]3.1B63.4164833.335.485.967.931.631.224.1
DeepSeek-VL-1.3B [31]2.0B64.6153232.234.887.666.731.127.637.6
PaliGemma-3B [6]2.9B71.0168634.933.187.069.628.732.229.6
MM1-3B-Chat [36]3B67.8176233.943.787.468.832.0--
MiniCPM-V [53]2.8B64.1165038.331.179.565.628.936.241.4
MiniCPM-V-2 [53]2.8B69.1180938.241.086.367.138.736.145.3
InternVL1.5 [9]2.2B70.9190234.639.387.169.841.137.563.5
†InternVL2 [10]2.2B73.2187734.344.688.371.646.437.974.7
Qwen2-VL-2B [46]2.2B74.9187241.149.5--43.041.7-
Phi-3 Vision [1]4.2B80.5-40.4-85.8-44.5--
Monolithic VLMs
Fuyu-8B (HD) [5]8B10.7--21.474.1----
SOLO [8]7B-1260---64.434.4--
*Chameleon-7B [43]7B31.117025.48.3-30.622.317.13.5
EVE-7B [12]7B49.5148332.325.683.661.325.221.112.4
EVE-7B (HD) [12]7B52.3162832.625.785.064.634.226.416.3
Emu3 [49]8B58.5-31.637.285.268.2---
Our HoVLE2.6B71.9186433.744.387.670.746.239.675.3
Our HoVLE (HD)2.6B73.3186232.243.887.470.949.238.474.3
+ +Table 1. Comparison with existing VLMs on general VLM benchmarks. The highest scores in monolithic VLMs are highlighted in bold. *We test Chameleon-7B with its released model, but it frequently refuses to perform the task with a response of "I can't help you with this", thus resulting in poor performance. ${}^{ \dagger }$ InternVL2 2.2B adopts the same LLM and supervised data with our HoVLE,sowe mark it as the compositional counterpart. + +is a comprehensive multi-modal benchmark covering a variety of fine-grained abilities. This substantial improvement is attributed to the holistic embedding, which prevents continual pre-training from disrupting language knowledge stored in pre-trained LLMs. (3) By incorporating higher resolution, HoVLE (HD) consistently improves the performance on VQA benchmarks, demonstrating the versatility of our holistic embedding in handling different resolutions. (4) Compared with previous compositional VLMs, HoVLE (HD) is able to deliver competitive results with many of these models. This shows that the introduction of holistic embedding provides an effective encoding that allows the LLM to understand images while preserving its language abilities. While HoVLE (HD) still falls short of state-of-the-art compositional VLMs, like Qwen2-VL2B [46] and Phi-3 Vision [1], it narrows the gap between monolithic and compositional VLMs, suggesting the feasibility of a high-performance monolithic VLM. + +# 4.3. Ablation Studies + +In this section, we aim to display and discuss the design choices that affect our HoVLE. Ablation experiments train the model with 300M and 8M data during the distillation and alignment stages respectively, unless specified. + +Holistic Embedding Depth. To validate the effectiveness of the holistic embedding, we first compare models with different holistic embedding depths in Table 3. With 0 Transformer layer, the model degenerates to a VLM without holistic embedding and the poor performance suggests + +the importance of this module. The model performs consistently better as the depth goes from 4 to 12. We speculate that deep embedding is beneficial for extracting visual representation. HoVLE uses 8 layers to achieve a balance between performance and model complexity. + +Training Strategy. Table 4 displays the influence of different training stages. The results show that omitting any stage will result in a performance drop. Among these, removing the distillation stage leads to the most severe degradation, highlighting its importance. We observe that models after distillation exhibit a faster decrease in loss during subsequent training stages, likely because the distillation provides general abilities of representation extraction, which is helpful for accelerating training. The improvement from instruction tuning is greater than that of the alignment stage. We assume this is because the LLM has more trainable parameters, allowing the whole model to fit the data well. + +Distillation Data Scaling. Figure 6 illustrates how the performance changes as distillation data scales up. Overall, the model performs better with more distillation data, implying the importance of comprehensive training. Specifically, the model's performance improves rapidly with 300M data and continues to increase at a more gradual pace as more data is added. We use 500M distillation data as our default setting. Please refer to Appendix B for scaling performance on more benchmarks. + +Text Data for Distillation. Table 5 presents the effects of different text input choices during distillation. The LAION [39] dataset provides a short text caption for each + +
Method# ParamsTextVQAChartVQAInfoVQADocVQAOCRbenchGQASQA-IAI2D
Compositional VLMs
VILA1.5-3B [25]3.1B60.4---43761.569.0-
DeepSeek-VL-1.3B [31]2.0B57.8---409--51.5
PaliGemma-3B [6]2.9B68.1---614--68.3
MM1-3B-Chat [36]3B71.9-----69.4-
MiniCPM-V [53]2.8B60.6--38.2366--56.3
MiniCPM-V-2 [53]2.8B74.1--71.9605--62.9
InternVL1.5 [9]2.2B70.574.855.485.065461.684.969.8
†InternVL2 [10]2.2B73.476.257.785.978461.084.974.1
Qwen2-VL-2B [46]2.2B79.773.565.590.1794---
Phi3-Vision [1]4.2B70.981.4----90.876.7
Monolithic VLMs
Fuyu-8B (HD) [5]8B-------64.5
SOLO [8]7B------73.361.4
*Chameleon-7B [43]7B4.82.95.01.57-47.246.0
EVE-7B [12]7B51.919.520.022.032760.863.048.5
EVE-7B (HD) [12]7B56.859.125.053.039862.664.959.1
Emu3 [49]8B64.768.643.876.368760.389.270.0
Our HoVLE2.6B66.077.652.184.773563.694.873.7
Our HoVLE (HD)2.6B70.978.655.786.174064.994.873.0
+ +![](images/92f875096f322e90090b082c3facfe66c6277160608b67f0f0b75987a352cdd5.jpg) + +![](images/d50f9ad8eae487fb8b973c2434a686aa556f2ff6e22950649d71fb29cd95f35e.jpg) +Figure 4. Distillation data scaling performance. + +![](images/4d1605567c118e5b519892f419bdf0820702a124d90fab752e0c35a5656b77dc.jpg) + +![](images/253e95810dc5ced324554bbda1376815d0d2e8c5db25aebb9111f9ad32fb97a4.jpg) + +Table 2. Comparison with existing VLMs on visual question answering benchmarks. The highest scores in monolithic VLMs are highlighted in bold. Same as Tab. 1, *Chameleon-7B frequently refuses to response, and ${}^{ \dagger }$ InternVL2 2.2B is our compositional counterpart. + +
DepthMMBMMESEEDTextVQAInfoVQADocVQA
035.4114946.011.319.112.2
467.8180168.047.742.673.6
872.0186270.962.051.482.1
1272.3188270.965.552.584.7
+ +Table 3. Ablation on holistic embedding depth. + +
Dist. Align. I.T.MMBMMESEEDTextVQAInfoVQADocVQA
✓✓37.3127446.411.219.3
✓✓71.3182870.560.149.2
✓✓66.1169866.755.043.0
✓✓✓72.0186270.962.051.4
+ +Table 4. Ablation on training strategies. + +
Text DataMMBMMESEEDTextVQAInfoVQADocVQA
Paired71.7176870.561.349.481.5
Random72.0186270.962.051.482.1
+ +Table 5. Ablation on text data for distillation. + +image, so we can also use image-text pairs as input. However, this approach yields relatively lower results, and the model continues to exhibit high loss values during alignment. We hypothesize that the text data from the LAION + +dataset lacks sufficient diversity to cover the full vocabulary, suggesting that paired data may require higher quality and variety in texts. In contrast, we propose using random texts as input. Table 5 shows that this strategy outperforms paired text data. We also observe lower loss during alignment. By leveraging random text, we reduce the high-quality need for image-text pair collection, thereby enhancing data utilization for large-scale training. + +# 4.4. Analysis + +Visualization. To better understand the effects of our introduced holistic embedding, we visualize the attention map of the LLM in Figure 5. We find that existing monolithic models, such as EVE [12] and Emu3 [49], exhibit denser text-to-image attention in their initial layers, while the compositional InternVL2 [10] and our HoVLE model maintain sparse attention patterns throughout. All models demonstrate sparse attention patterns in the later layers (with the final layers shown). Please refer to Appendix C for more attention maps in the middle layers. + +We hypothesize that sparser text-to-image attention indicates that visual features are effectively extracted and well-aligned with text tokens, allowing the text tokens to focus + +![](images/2699d803a1aad92a6bcca7e9f35c86d5a743bc99d9cbdefc38ceab6755bf659b.jpg) + +![](images/1d76128a5a51400e1a75e7954d9846d1ad48833416451fb0e580fd0eee6bed82.jpg) + +![](images/39a9cdb1effd2a6b2a66132b55f294617849ba9899c5dd765ae02a0dde89e23a.jpg) + +![](images/b1ed64c45eab64b2819afccb46f204fcfe159456cad018663c541d6f47847ba0.jpg) + +![](images/2cdd9aeb0759e22cb2fb3cb5a3330d9e8008a1021ecd47469371e38e63992505.jpg) + +![](images/873bf3203cc994facceac438dd30b9ea4eccf04207f4d967abf28ad4387bb017.jpg) +Figure 5. Attention Maps for EVE, Emu3, InternVL2 and our HoVLE at the first and last layers of LLM backbones. Y-axis represents query tokens, and X-axis represents key tokens, with text modality tokens in gray and image modality tokens in yellow. All four models share the same input, but the sequence lengths of input tokens are different due to different image pre-processing. We highlight text-to-image attention below each full attention map. Our HoVLE, like the compositional InternVL2, has sparse attention across all network layers, while other monolithic models Emu3 and EVE have denser attention in shallow layers. + +![](images/24475e612ea96d33a8c1df1105e18bf343cce755f1af31d70de9a5bc384472f6.jpg) + +![](images/2a743b2267515adae460315f409e22a74599db4cb1386e0d17a39ce353cda6f8.jpg) + +![](images/d2ddb53de9b6f618ac6bcc77f5fdd57be9e45d9fb61d0d9430f7d2f5b7c20cfb.jpg) + +
ModelTile ResolutionFLOPs(T)↓TTFT↓TPS↑Avg. Score↑
InternVL2448 × 4481.811.3728165.8
HoVLE (HD)336 × 3361.511.2424365.7
HoVLE (HD)392 × 3921.891.5222466.1
HoVLE (HD)448 × 4482.181.8420666.8
+ +Table 6. Inference speed-performance trade-off of HoVLE (HD). "TTFT" and "TPS" denotes the time to first token in seconds and throughput in tokens per second, respectively. "Avg. Score" refers to the average performance over 17 benchmarks. Please refer to Appendix for detailed results. + +only on the most relevant visual information. In contrast, denser text-to-image attention may suggest that the model is still in the process of extracting visual features within the LLM component. Therefore, by introducing holistic embedding, the LLM may process vision and language inputs more effectively. + +Speed-Performance Trade-off. We report the speed-performance trade-off of HoVLE (HD) in Table 6. For evaluation, we use one image and 256 text tokens as input, with the model restricted to output 120 tokens during testing. InternVL2 serves as the main comparison model, as it uses the same LLM. When the tile resolution of DynProcess( $\cdot$ ) is 336, HoVLE (HD) achieves comparable performance to InternVL2, while also showing lower computational demand. As the input resolution increases, HoVLE's performance improves consistently, but at the cost of slower speed. Additionally, we observe that using the KV cache [38] during inference significantly reduces computational time, making kernel launch the main bottleneck in TPS. Optimizing the kernel could further enhance processing speed. + +# 5. Conclusion + +This work presents HoVLE, a monolithic Vision-Language Model (VLM) that processes images and texts in a unified manner. HoVLE introduces a holistic embedding module that projects image and text inputs into a shared embedding space. This allows the Large Language Model (LLM) to interpret images in the same way as texts. Furthermore, a multi-stage training procedure is designed to strengthen the encoding abilities of the holistic embedding. A distillation stage first provides general vision and language encoding capabilities, while an alignment stage aligns different modalities. An instruction tuning stage enhances the overall abilities in the end. Extensive experiments show that HoVLE surpasses previous monolithic VLMs and demonstrates competitive performance with compositional VLMs. This work focuses on enhancing monolithic VLMs by introducing a holistic embedding module, aiming to elevate their capabilities rather than rival those of compositional VLMs. This work narrows the gap between monolithic and compositional VLMs, providing a promising direction for the development of monolithic VLMs. + +Limitations. Due to limited computational resources, our experiments are conducted on 2B parameter scale. Scaling to larger scales can provide further insight into the model scalability of our approach, which we hope to explore in future work. + +Acknowledgements. This work is supported by the National Key R&D Program of China (NO. 2022ZD0161300), by the National Natural Science Foundation of China (62376134). + +# References + +[1] Marah Abdin, Jyoti Aneja, Hany Awadalla, Ahmed Awadallah, Ammar Ahmad Awan, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Jianmin Bao, Harkirat Behl, et al. Phi-3 technical report: A highly capable language model locally on your phone. arxiv:2404.14219, 2024. 6, 7 +[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arxiv:2303.08774, 2023. 1 +[3] AI Anthropic. The claude 3 model family: Opus, sonnet, haiku. Claude-3 Model Card, 2024. 3 +[4] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arxiv:2308.12966, 2023. 1, 3 +[5] Rohan Bavishi, Erich Elsen, Curtis Hawthorne, Maxwell Nye, Augustus Odena, Arushi Somani, and Sagnak Tasirlar. Introducing our multimodal models. https://www. adept.ai/blog/fuyu-8b, 2023. 2, 3, 5, 6, 7 +[6] Lucas Beyer, Andreas Steiner, André Susano Pinto, Alexander Kolesnikov, Xiao Wang, Daniel Salz, Maxim Neumann, Ibrahim Alabdulmohsin, Michael Tschannen, Emanuele Bugliarello, et al. Paligemma: A versatile 3b vlm for transfer. arxiv:2407.07726, 2024. 2, 3, 6, 7 +[7] Zheng Cai, Maosong Cao, Haojiong Chen, Kai Chen, Keyu Chen, Xin Chen, Xun Chen, Zehui Chen, Zhi Chen, Pei Chu, et al. Internl m2 technical report. arxiv:2403.17297, 2024. 1, 5 +[8] Yangyi Chen, Xingyao Wang, Hao Peng, and Heng Ji. A single transformer for scalable vision-language modeling. arxiv:2407.06438, 2024. 2, 3, 6, 7 +[9] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arxiv:2404.16821, 2024. 5, 6, 7 +[10] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. Internv12: Better than the best—expanding performance boundaries of open-source multimodal models with the progressive scaling strategy. https://internvl.github.io/blog/2024-07-02-InternVL-2.0/, 2024.6.7 +[11] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In CVPR, pages 24185–24198, 2024. 1, 3, 5 +[12] Haiwen Diao, Yufeng Cui, Xiaotong Li, Yueze Wang, Huchuan Lu, and Xinlong Wang. Unveiling encoder-free vision-language models. arxiv:2406.11832, 2024. 2, 3, 5, 6, 7 +[13] Zhiyuan Fang, Jianfeng Wang, Xiaowei Hu, Lijuan Wang, Yezhou Yang, and Zicheng Liu. Compressing visual-linguistic model via knowledge distillation. In ICCV, pages 1428-1438, 2021. 3 + +[14] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. arxiv:2306.13394, 2023.5 +[15] Tianrui Guan, Fuxiao Liu, Xiyang Wu, Ruiqi Xian, Zongxia Li, Xiaoyu Liu, Xijun Wang, Lichang Chen, Furong Huang, Yaser Yacoob, et al. Hallusionbench: An advanced diagnostic suite for entangled language hallucination and visual illusion in large vision-language models. arxiv:2310.14566, 2023. 5 +[16] Suchin Gururangan, Ana Marasovic, Swabha Swayamdipta, Kyle Lo, Iz Beltagy, Doug Downey, and Noah A Smith. Don't stop pretraining: Adapt language models to domains and tasks. In ACL, pages 8342-8360, 2020. 2 +[17] Yushi Hu, Otilia Stretcu, Chun-Ta Lu, Krishnamurthy Viswanathan, Kenji Hata, Enming Luo, Ranjay Krishna, and Ariel Fuxman. Visual program distillation: Distilling tools and programmatic reasoning into vision-language models. In CVPR, pages 9590-9601, 2024. 3 +[18] Runhui Huang, Xinpeng Ding, Chunwei Wang, Jianhua Han, Yulong Liu, Hengshuang Zhao, Hang Xu, Lu Hou, Wei Zhang, and Xiaodan Liang. Hires-llava: Restoring fragmentation input in high-resolution large vision-language models. arXiv:2407.08706, 2024.3 +[19] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In CVPR, pages 6700-6709, 2019. 5, 1 +[20] Aniruddha Kembhavi, Mike Salvato, Eric Kolve, Minjoon Seo, Hannaneh Hajishirzi, and Ali Farhadi. A diagram is worth a dozen images. In ECCV, pages 235-251. Springer, 2016. 5, 1 +[21] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arxiv:2307.16125, 2023. 5 +[22] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, pages 12888-12900. PMLR, 2022. 3 +[23] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. In EMNLP, pages 292-305, 2023. 5 +[24] Zhang Li, Biao Yang, Qiang Liu, Zhiyin Ma, Shuo Zhang, Jingxu Yang, Yabo Sun, Yuliang Liu, and Xiang Bai. Monkey: Image resolution and text label are important things for large multi-modal models. In CVPR, pages 26763-26773, 2024. 5 +[25] Ji Lin, Hongxu Yin, Wei Ping, Pavlo Molchanov, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models. In CVPR, pages 26689-26699, 2024. 6, 7 +[26] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In CVPR, pages 26296-26306, 2024. 1, 3 +[27] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge. https : + +//llava-v1.github.io/blog/2024-01-30-llava-last/, 2024.1 +[28] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2024. 3 +[29] Yuliang Liu, Zhang Li, Biao Yang, Chunyuan Li, Xucheng Yin, Cheng-lin Liu, Lianwen Jin, and Xiang Bai. On the hidden mystery ofOCR in large multimodal models. arxiv:2305.07895, 2023. 5 +[30] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? In ECCV, pages 216-233. Springer, 2025. 2, 5 +[31] Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhuoshu Li, Hao Yang, et al. Deepseek-vl: towards real-world vision-language understanding. arxiv:2403.05525, 2024. 6, 7 +[32] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. In NeurIPS, pages 2507–2521, 2022. 5, 1 +[33] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arxiv:2203.10244, 2022. 5, 1 +[34] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In WACV, pages 2200-2209, 2021. 5 +[35] Minesh Mathew, Viraj Bagal, Ruben Tito, Dimosthenis Karatzas, Ernest Valveny, and CV Jawahar. Infographicvqa. In WACV, pages 1697-1706, 2022. 5, 1 +[36] Brandon McKinzie, Zhe Gan, Jean-Philippe Fauconnier, Sam Dodge, Bowen Zhang, Philipp Dufter, Dhruti Shah, Xianzhi Du, Futang Peng, Floris Weers, et al. Mm1: Methods, analysis & insights from multimodal llm pre-training. arxiv:2403.09611, 2024. 3, 6, 7 +[37] Jae Sung Park, Jack Hessel, Khyathi Chandu, Paul Pu Liang, Ximing Lu, Peter West, Youngjae Yu, Qiuyuan Huang, Jianfeng Gao, Ali Farhadi, et al. Localized symbolic knowledge distillation for visual commonsense models. In NeurIPS, 2024. 3 +[38] Reiner Pope, Sholto Douglas, Aakanksha Chowdhery, Jacob Devlin, James Bradbury, Jonathan Heek, Kefan Xiao, Shivani Agrawal, and Jeff Dean. Efficiently scaling transformer inference. Proceedings of Machine Learning and Systems, 5: 606-624, 2023. 8 +[39] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS, pages 25278–25294, 2022. 5, 6, 1 +[40] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In CVPR, pages 8317-8326, 2019. 5, 1 + +[41] Quan Sun, Yufeng Cui, Xiaosong Zhang, Fan Zhang, Qiying Yu, Yueze Wang, Yongming Rao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative multimodal models are in-context learners. In CVPR, pages 14398-14409, 2024. 3 +[42] Chenxin Tao, Xizhou Zhu, Shiqian Su, Lewei Lu, Changyao Tian, Xuan Luo, Gao Huang, Hongsheng Li, Yu Qiao, Jie Zhou, et al. Learning 1d causal visual representation with de-focus attention networks. arxiv:2406.04342, 2024. 4 +[43] Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models. arxiv:2405.09818, 2024. 2, 3, 6, 7 +[44] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arxiv:2312.11805, 2023. 3 +[45] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arxiv:2302.13971, 2023. 1 +[46] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arxiv:2409.12191, 2024. 1, 6, 7 +[47] Tiannan Wang, Wangchunshu Zhou, Yan Zeng, and Xinsong Zhang. Efficientvlm: Fast and accurate vision-language models via knowledge distillation and modal-adaptive pruning. arxiv:2210.07795, 2022. 3 +[48] Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong Lu, Jie Zhou, Yu Qiao, et al. Vision: Large language model is also an open-ended decoder for vision-centric tasks. In NeurIPS, 2024. 1 +[49] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arxiv:2409.18869, 2024. 2, 3, 6, 7 +[50] X.ai. Grok-1.5 vision preview. https://x.ai/blog/grok-1.5v, 2024.3 +[51] Ruyi Xu, Yuan Yao, Zonghao Guo, Junbo Cui, Zanlin Ni, Chunjiang Ge, Tat-Seng Chua, Zhiyuan Liu, Maosong Sun, and Gao Huang. Llava-uhd: an Imm perceiving any aspect ratio and high-resolution images. arXiv:2403.11703, 2024. 3 +[52] Zhengyuan Yang, Linjie Li, Kevin Lin, Jianfeng Wang, Chung-Ching Lin, Zicheng Liu, and Lijuan Wang. The dawn of Imms: Preliminary explorations with gpt-4v (ision). arxiv:2309.17421, 9(1):1, 2023. 3 +[53] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arxiv:2408.01800, 2024.6, 7 +[54] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. Mm-vet: Evaluating large multimodal models for integrated capabilities. arxiv:2308.02490, 2023. 5 + +[55] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In CVPR, pages 9556-9567, 2024. 5 +[56] Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. A survey of large language models. arxiv:2303.18223, 2023. 1 +[57] Chunting Zhou, Lili Yu, Arun Babu, Kushal Tirumala, Michihiro Yasunaga, Leonid Shamis, Jacob Kahn, Xuezhe Ma, Luke Zettlemoyer, and Omer Levy. Transfusion: Predict the next token and diffuse images with one multi-modal model. arxiv:2408.11039, 2024. 2, 3 +[58] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arxiv:2304.10592, 2023. 1 \ No newline at end of file diff --git a/2025/HoVLE_ Unleashing the Power of Monolithic Vision-Language Models with Holistic Vision-Language Embedding/images.zip b/2025/HoVLE_ Unleashing the Power of Monolithic Vision-Language Models with Holistic Vision-Language Embedding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..cbd9777e5eb9370fcc5cb9e2e41f90e171100e51 --- /dev/null +++ b/2025/HoVLE_ Unleashing the Power of Monolithic Vision-Language Models with Holistic Vision-Language Embedding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:71ac961c9509ed9dfa4637d68e3ff52ff627fbaec8318cf73e62cdb18d07a2bf +size 671710 diff --git a/2025/HoVLE_ Unleashing the Power of Monolithic Vision-Language Models with Holistic Vision-Language Embedding/layout.json b/2025/HoVLE_ Unleashing the Power of Monolithic Vision-Language Models with Holistic Vision-Language Embedding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9dc1874434a7597e303fb10f6a485b3f96969336 --- /dev/null +++ b/2025/HoVLE_ Unleashing the Power of Monolithic Vision-Language Models with Holistic Vision-Language Embedding/layout.json @@ -0,0 +1,8184 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 72, + 102, + 538, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 102, + 538, + 140 + ], + "spans": [ + { + "bbox": [ + 72, + 102, + 538, + 140 + ], + "type": "text", + "content": "HoVLE: Unleashing the Power of Monolithic Vision-Language Models with Holistic Vision-Language Embedding" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 77, + 158, + 534, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 158, + 534, + 188 + ], + "spans": [ + { + "bbox": [ + 77, + 158, + 534, + 188 + ], + "type": "text", + "content": "Chenxin Tao" + }, + { + "bbox": [ + 77, + 158, + 534, + 188 + ], + "type": "inline_equation", + "content": "^{1,3*}" + }, + { + "bbox": [ + 77, + 158, + 534, + 188 + ], + "type": "text", + "content": ", Shiqian Su" + }, + { + "bbox": [ + 77, + 158, + 534, + 188 + ], + "type": "inline_equation", + "content": "^{1,3*}" + }, + { + "bbox": [ + 77, + 158, + 534, + 188 + ], + "type": "text", + "content": ", Xizhou Zhu" + }, + { + "bbox": [ + 77, + 158, + 534, + 188 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 77, + 158, + 534, + 188 + ], + "type": "text", + "content": ", Chenyu Zhang" + }, + { + "bbox": [ + 77, + 158, + 534, + 188 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 77, + 158, + 534, + 188 + ], + "type": "text", + "content": ", Zhe Chen" + }, + { + "bbox": [ + 77, + 158, + 534, + 188 + ], + "type": "inline_equation", + "content": "^{4,2}" + }, + { + "bbox": [ + 77, + 158, + 534, + 188 + ], + "type": "text", + "content": ", Jiawen Liu" + }, + { + "bbox": [ + 77, + 158, + 534, + 188 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 77, + 158, + 534, + 188 + ], + "type": "text", + "content": ", Wenhai Wang" + }, + { + "bbox": [ + 77, + 158, + 534, + 188 + ], + "type": "inline_equation", + "content": "^{5,2}" + }, + { + "bbox": [ + 77, + 158, + 534, + 188 + ], + "type": "text", + "content": ", Lewei Lu" + }, + { + "bbox": [ + 77, + 158, + 534, + 188 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 77, + 158, + 534, + 188 + ], + "type": "text", + "content": ", Gao Huang" + }, + { + "bbox": [ + 77, + 158, + 534, + 188 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 77, + 158, + 534, + 188 + ], + "type": "text", + "content": ", Yu Qiao" + }, + { + "bbox": [ + 77, + 158, + 534, + 188 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 77, + 158, + 534, + 188 + ], + "type": "text", + "content": ", Jifeng Dai" + }, + { + "bbox": [ + 77, + 158, + 534, + 188 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 165, + 195, + 445, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 195, + 445, + 209 + ], + "spans": [ + { + "bbox": [ + 165, + 195, + 445, + 209 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 165, + 195, + 445, + 209 + ], + "type": "text", + "content": "Tsinghua University " + }, + { + "bbox": [ + 165, + 195, + 445, + 209 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 165, + 195, + 445, + 209 + ], + "type": "text", + "content": "Shanghai Artificial Intelligence Laboratory" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 214, + 209, + 397, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 209, + 397, + 220 + ], + "spans": [ + { + "bbox": [ + 214, + 209, + 397, + 220 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 214, + 209, + 397, + 220 + ], + "type": "text", + "content": "SenseTime Research " + }, + { + "bbox": [ + 214, + 209, + 397, + 220 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 214, + 209, + 397, + 220 + ], + "type": "text", + "content": "Nanjing University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 164, + 220, + 446, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 220, + 446, + 233 + ], + "spans": [ + { + "bbox": [ + 164, + 220, + 446, + 233 + ], + "type": "text", + "content": "5The Chinese University of Hong Kong 6 Johns Hopkins University" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 186, + 235, + 448, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 235, + 448, + 247 + ], + "spans": [ + { + "bbox": [ + 186, + 235, + 448, + 247 + ], + "type": "text", + "content": "Model: https://huggingface.co/OpenGVLab/HoVLE" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 151, + 260, + 200, + 273 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 260, + 200, + 273 + ], + "spans": [ + { + "bbox": [ + 151, + 260, + 200, + 273 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 285, + 296, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 285, + 296, + 622 + ], + "spans": [ + { + "bbox": [ + 53, + 285, + 296, + 622 + ], + "type": "text", + "content": "The rapid advance of Large Language Models (LLMs) has catalyzed the development of Vision-Language Models (VLMs). Monolithic VLMs, which avoid modality-specific encoders, offer a promising alternative to the compositional ones but face the challenge of inferior performance. Most existing monolithic VLMs require tuning pre-trained LLMs to acquire vision abilities, which may degrade their language capabilities. To address this dilemma, this paper presents a novel high-performance monolithic VLM named HoVLE. We note that LLMs have been shown to be capable of interpreting images when image embeddings are aligned with text embeddings. The challenge for current monolithic VLMs actually lies in the lack of a holistic embedding module for both vision and language inputs. Therefore, HoVLE introduces a holistic embedding module that converts visual and textual inputs into a shared space, allowing LLMs to process images in the same way as texts. Furthermore, a multi-stage training strategy is carefully designed to empower the holistic embedding module. It is first trained to distill visual features from a pre-trained vision encoder and text embeddings from the LLM, enabling largescale training with unpaired random images and text tokens. The whole model further undergoes next-token prediction on multi-modal data to align the embeddings. Finally, an instruction-tuning stage is incorporated. Our experiments show that HoVLE achieves performance close to leading compositional models on various benchmarks, outperforming previous monolithic models by a large margin." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 642, + 135, + 654 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 642, + 135, + 654 + ], + "spans": [ + { + "bbox": [ + 56, + 642, + 135, + 654 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 662, + 295, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 662, + 295, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 662, + 295, + 687 + ], + "type": "text", + "content": "The rapid advancement of Large Language Models (LLMs) [2, 4, 7, 45, 56] has significantly accelerated the" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 332, + 247, + 555, + 397 + ], + "blocks": [ + { + "bbox": [ + 332, + 247, + 555, + 397 + ], + "lines": [ + { + "bbox": [ + 332, + 247, + 555, + 397 + ], + "spans": [ + { + "bbox": [ + 332, + 247, + 555, + 397 + ], + "type": "image", + "image_path": "4d9029fafad4ac06bce0a77104698f6f2b9e9c48ca65696f7cb967e1e5e28bb3.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 314, + 404, + 555, + 449 + ], + "blocks": [ + { + "bbox": [ + 314, + 404, + 555, + 449 + ], + "lines": [ + { + "bbox": [ + 314, + 404, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 314, + 404, + 555, + 449 + ], + "type": "image", + "image_path": "1665665ec4ecccc53813df1560e205fefa93613dcc1393d7d742d6d0a58098c7.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 457, + 555, + 523 + ], + "lines": [ + { + "bbox": [ + 313, + 457, + 555, + 523 + ], + "spans": [ + { + "bbox": [ + 313, + 457, + 555, + 523 + ], + "type": "text", + "content": "Figure 1. Performance comparison on different benchmarks between compositional VLMs (dashed lines) and monolithic VLMs (solid lines). Previous monolithic VLMs exhibit a significant performance gap compared to compositional VLMs, while Our HoVLE demonstrate competitive capabilities with state-of-the-art compositional VLMs." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 533, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 533, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 533, + 556, + 715 + ], + "type": "text", + "content": "development of Vision-Language Models (VLMs) [26, 48, 58], which have demonstrated powerful visual-text processing capabilities. Compositional VLMs are currently the de facto approach for building high-performance VLMs [11, 27, 46], which integrate pre-trained vision encoders with LLMs, as shown in Fig. 2(a). By combining pre-trained visual understanding capabilities with language modeling capabilities, compositional VLMs achieve impressive multimodal performance. On the other hand, there is growing interest in developing VLMs without modality-specific encoders, referred to as monolithic VLMs in this paper. These models process image and text inputs as a unified whole, directly feeding them into the LLM, as illustrated in Fig. 2(b). Due to their monolithic design, these VLMs feature simpler architectures, offering greater potential for" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 64, + 693, + 135, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 693, + 135, + 703 + ], + "spans": [ + { + "bbox": [ + 64, + 693, + 135, + 703 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 64, + 703, + 240, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 703, + 240, + 712 + ], + "spans": [ + { + "bbox": [ + 64, + 703, + 240, + 712 + ], + "type": "text", + "content": "Corresponding Author: daijifeng@tsinghua.edu.cn." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14559" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 69, + 132, + 186 + ], + "blocks": [ + { + "bbox": [ + 61, + 69, + 132, + 186 + ], + "lines": [ + { + "bbox": [ + 61, + 69, + 132, + 186 + ], + "spans": [ + { + "bbox": [ + 61, + 69, + 132, + 186 + ], + "type": "image", + "image_path": "34dede1f800a6182dfb96d2c82fec367cdbc76f4baa6dce686ff6b2420b21d36.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 57, + 188, + 129, + 196 + ], + "lines": [ + { + "bbox": [ + 57, + 188, + 129, + 196 + ], + "spans": [ + { + "bbox": [ + 57, + 188, + 129, + 196 + ], + "type": "text", + "content": "(a) Compositional VLMs" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 148, + 70, + 214, + 185 + ], + "blocks": [ + { + "bbox": [ + 148, + 70, + 214, + 185 + ], + "lines": [ + { + "bbox": [ + 148, + 70, + 214, + 185 + ], + "spans": [ + { + "bbox": [ + 148, + 70, + 214, + 185 + ], + "type": "image", + "image_path": "3926d4724e775c32cc46513c6434edaca47ebd944fed416d3fb147658cf874b9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 135, + 188, + 223, + 196 + ], + "lines": [ + { + "bbox": [ + 135, + 188, + 223, + 196 + ], + "spans": [ + { + "bbox": [ + 135, + 188, + 223, + 196 + ], + "type": "text", + "content": "(b) Existing Monolithic VLMs" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 230, + 70, + 294, + 185 + ], + "blocks": [ + { + "bbox": [ + 230, + 70, + 294, + 185 + ], + "lines": [ + { + "bbox": [ + 230, + 70, + 294, + 185 + ], + "spans": [ + { + "bbox": [ + 230, + 70, + 294, + 185 + ], + "type": "image", + "image_path": "f07db8dd1e502ad993949f7b4f036b3f9b7317289c19c15d086e1bab869973f7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 238, + 188, + 283, + 196 + ], + "lines": [ + { + "bbox": [ + 238, + 188, + 283, + 196 + ], + "spans": [ + { + "bbox": [ + 238, + 188, + 283, + 196 + ], + "type": "text", + "content": "(c) Our HoVLE" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 202, + 295, + 301 + ], + "lines": [ + { + "bbox": [ + 55, + 202, + 295, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 202, + 295, + 301 + ], + "type": "text", + "content": "Figure 2. Comparison of VLM architectures. (a) Compositional VLMs integrate pre-trained vision encoders with LLMs, using an extra connector to align image and text embeddings. (b) Existing Monolithic VLMs directly feed image and text inputs into LLMs, which require continual pre-training to gain visual abilities. (c) HoVLE uses a holistic embedding module to project image and text input to a unified embedding space, enabling LLMs to interpret images in a text-like manner. Blocks with the same color have the same Transformer layer architecture." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 315, + 295, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 315, + 295, + 362 + ], + "spans": [ + { + "bbox": [ + 55, + 315, + 295, + 362 + ], + "type": "text", + "content": "multimodal tasks, such as unifying generation and recognition [43, 49, 57]. Monolithic VLMs are gaining increasing attention and have shown their potential across many studies [5, 8, 12]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 364, + 295, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 364, + 295, + 519 + ], + "spans": [ + { + "bbox": [ + 55, + 364, + 295, + 519 + ], + "type": "text", + "content": "However, monolithic VLMs still lag behind compositional VLMs in terms of performance. Some researches [5, 6, 43, 57] explore training monolithic VLMs from scratch using a combination of multi-modal and pure text data. Such practice fails to leverage existing pre-trained models, resulting in substantial computational demands for training. Another research direction [8, 12] proposes injecting visual understanding capabilities into pre-trained LLMs by conducting continual pre-training [16] with large-scale and high-quality multi-modal data. Nevertheless, such continual-tuning methods may interfere with the language knowledge already acquired by pre-trained LLMs, potentially limiting their multi-modal performance [12]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 521, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 628 + ], + "type": "text", + "content": "The performance limitations of current monolithic VLMs mainly stem from the need to tune pre-trained LLMs. However, the success of compositional VLMs indicates that pre-trained LLMs can interpret visual information without further pre-training, as long as image embeddings are properly aligned with text embeddings. Thus, the reason for such LLM-tuning dilemma of current monolithic VLMs is the lack of a holistic embedding module that can coordinate different modalities." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "content": "To this end, we develop a modality-sharing holistic embedding module as shown in Fig. 2(c), which maps mixed visual and language inputs into a shared embedding space. It can enable pre-trained LLMs to interpret visual input in the same way as text. Similar to compositional VLMs, this not only avoids tuning pre-trained LLMs but also preserves their language proficiency while successfully inte" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "text", + "content": "grating visual capabilities. Our experiments demonstrate that such a holistic embedding module can indeed achieve high-performance monolithic VLMs." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 108, + 555, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 108, + 555, + 371 + ], + "spans": [ + { + "bbox": [ + 313, + 108, + 555, + 371 + ], + "type": "text", + "content": "In this work, we introduce HoVLE, a novel monolithic vision-language model that learns a Holistic Vision-Language Embedding module to extend LLMs with vision capabilities. The holistic embedding module consists of the same causal Transformer layers as the LLM. It accepts both images and texts as input and projects them into a unified embedding space. These embeddings are then forwarded into the LLM, constituting a monolithic VLM. To empower the holistic embedding module with vision and language encoding abilities, a multi-stage training strategy is proposed. First, the module is trained to distill visual features from a pre-trained vision encoder and text embeddings from the LLM. The distillation stage does not require image-text pairs, but instead uses large-scale unpaired random images and text tokens as inputs. After distillation, the holistic embedding module is combined with the LLM to perform next token prediction on multi-modal data, so that the output embeddings can be further aligned into a unified space. Finally, the whole model is enhanced with an instruction tuning stage. The alignment and instruction tuning stages are similar to those used in compositional VLMs, eliminating the need for extra supervised data." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 371, + 554, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 371, + 554, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 554, + 453 + ], + "type": "text", + "content": "Extensive experiments show that our HoVLE is able to achieve close performance with these leading compositional VLMs on 17 multi-modal benchmarks. It also surpasses previous monolithic VLMs by a large margin, e.g., " + }, + { + "bbox": [ + 313, + 371, + 554, + 453 + ], + "type": "inline_equation", + "content": "\\sim 15" + }, + { + "bbox": [ + 313, + 371, + 554, + 453 + ], + "type": "text", + "content": " points on MMBench [30], showing the effectiveness of HoVLE. In conclusion, our contributions can be summarized as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 455, + 554, + 635 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 314, + 455, + 554, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 455, + 554, + 527 + ], + "spans": [ + { + "bbox": [ + 314, + 455, + 554, + 527 + ], + "type": "text", + "content": "- HoVLE is proposed as a novel monolithic VLM that learns a holistic embedding module to expand LLMs with vision capabilities. By projecting vision and language inputs to a shared embedding space, HoVLE can understand images without compromising language proficiency." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 527, + 553, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 527, + 553, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 527, + 553, + 598 + ], + "type": "text", + "content": "- A novel multi-stage training strategy is proposed to empower the holistic embedding module with vision and language encoding abilities. Our training procedure enables HoVLE to develop strong vision and language capabilities without requiring additional supervised data compared to compositional VLMs." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 598, + 553, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 598, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 598, + 553, + 635 + ], + "type": "text", + "content": "- HoVLE demonstrates comparable performance with compositional VLMs, and surpasses previous monolithic VLMs by a large margin." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 645, + 400, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 645, + 400, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 645, + 400, + 658 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": "Compositional Vision-Language Models. The rapid advancements and success of Large Language Models (LLMs) have sparked researchers' interest in multi-modal models, with Vision-Language Models (VLMs) emerging" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14560" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 264 + ], + "type": "text", + "content": "as a prominent focus [11, 12, 22, 26, 28, 41]. Compositional VLMs have become the dominant paradigm in the field, where image tokens are typically generated through visual encoders and an additional connector (e.g., MLP projection). Many notable commercial models, including GPT-4V [52], the Gemini series [44], Claude-3V [3], Grok-1.5V [50], and MM1 [36], have demonstrated impressive performance across various multi-modal tasks. In the open-source domain, models like LLaVA [26, 28], BLIP [22], Emu2 [41], InternVL [11], and Qwen-VL [4] stand out. However, compositional VLMs face several challenges, such as complexities in deployment and optimization due to their heterogeneous architecture, trade-offs in the design of network capacity between image and text processing, and potential limitations in cross-modal understanding capabilities [6, 8, 12, 43]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 272, + 296, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 296, + 535 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 296, + 535 + ], + "type": "text", + "content": "Monolithic Vision-Language Models. In contrast to compositional VLMs, monolithic VLMs do not rely on a specific vision encoder. Some works explore training a monolithic VLM from scratch. Fuyu-8B [5] directly inputs image pixel patches into the backbone network using a simple linear projection, while Chameleon [43], Transfusion [57], and Emu3 [49] adopt image tokenizers to map images to discrete tokens as inputs. Nevertheless, these methods require collecting massive pure text and paired image-text data, as well as intensive training computational resources to achieve good performance. Other approaches propose to extend existing pre-trained LLMs to VLMs. SOLO [8] uses a simple linear projection to handle raw image pixels and fine-tunes the pre-trained LLM to develop vision abilities. EVE [12] employs a patch embedding layer with cross-attention to enhance image representation, alongside a patch aligning layer to provide fine-grained visual supervision with pre-trained vision encoders. However, these methods require continually pre-training LLMs, which may diminish their language proficiency. Current monolithic VLMs still underperform in image-text understanding tasks when compared to their compositional counterparts." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 537, + 295, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 537, + 295, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 537, + 295, + 586 + ], + "type": "text", + "content": "In contrast, by introducing a holistic embedding module, HoVLE can achieve comparable performance with state-of-the-art compositional VLMs, significantly surpassing other monolithic VLMs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": "Knowledge Distillation. Knowledge distillation is an effective technique for transferring knowledge and abilities from a teacher model to a student model. In the realm of VLMs, knowledge distillation has also been successfully incorporated into various tasks. DistillVLM [13] and EfficientVLM [47] use knowledge distillation to compress the model size and accelerate inference speed. VPD [17] and LSKD [37] leverage tools or LLMs to synthesize multimodal data so that VLMs can be trained to distill the skills in specialist models and LLMs. Our work employs knowl" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "edge distillation to endow the holistic embedding module with general vision and language encoding abilities. Since distillation targets can be computed independently, we can use unpaired random image and text data, eliminating the need to collect additional multi-modal data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 142, + 370, + 154 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 142, + 370, + 154 + ], + "spans": [ + { + "bbox": [ + 313, + 142, + 370, + 154 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 162, + 553, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 162, + 553, + 221 + ], + "spans": [ + { + "bbox": [ + 313, + 162, + 553, + 221 + ], + "type": "text", + "content": "In this section, we present HoVLE, a monolithic vision-language model with decent vision and language capabilities. In the following sections, we introduce the network architecture and the training procedures for such a unified model. The details are shown in Figure 3." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 228, + 440, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 440, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 440, + 239 + ], + "type": "text", + "content": "3.1. Network Architecture" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 246, + 553, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 246, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 246, + 553, + 354 + ], + "type": "text", + "content": "The success of compositional VLMs demonstrates that LLMs are capable of understanding visual information when image embeddings are properly aligned with text embeddings. Inspired by this, HoVLE proposes that a monolithic VLM can be built using a holistic embedding module. This module takes mixed raw images and texts as inputs and projects them into a shared embedding space. In this way, LLMs can interpret images in the same way as texts, resulting in a monolithic VLM." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 359, + 554, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 359, + 554, + 419 + ], + "spans": [ + { + "bbox": [ + 313, + 359, + 554, + 419 + ], + "type": "text", + "content": "Holistic Embedding Module. We first introduce the holistic embedding module to project vision and language inputs to a shared embedding space. Specifically, given an input image " + }, + { + "bbox": [ + 313, + 359, + 554, + 419 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 359, + 554, + 419 + ], + "type": "text", + "content": " and texts " + }, + { + "bbox": [ + 313, + 359, + 554, + 419 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 359, + 554, + 419 + ], + "type": "text", + "content": ", we first transform them into vector forms via simple tokenization and linear projection:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 351, + 426, + 553, + 438 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 426, + 553, + 438 + ], + "spans": [ + { + "bbox": [ + 351, + 426, + 553, + 438 + ], + "type": "interline_equation", + "content": "x _ {I} = \\operatorname {P a t c h E m b e d} (\\operatorname {D y n P r o c e s s} (I)) + \\mathrm {P E}, \\tag {1}", + "image_path": "7f00edc5b81c3c2ff86aaeaff5c1630e742dd9999aff13f36a1e70023546aaf6.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 348, + 441, + 553, + 453 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 441, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 348, + 441, + 553, + 453 + ], + "type": "interline_equation", + "content": "x _ {T} = \\operatorname {T e x t E m b e d} (\\operatorname {T o k e n i z e r} (T)), \\tag {2}", + "image_path": "e56bcaacab27aedda47253b5ddd2b302ddb09535c7664c521089805f9848c899.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 460, + 554, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 554, + 614 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 554, + 614 + ], + "type": "text", + "content": "where DynProcess(\\cdot) is the dynamic high-resolution strategy from InternVL [11]. This process divides the image into " + }, + { + "bbox": [ + 313, + 460, + 554, + 614 + ], + "type": "inline_equation", + "content": "448 \\times 448" + }, + { + "bbox": [ + 313, + 460, + 554, + 614 + ], + "type": "text", + "content": " tiles and appends a global thumbnail to provide overall context. Similar operations have been adopted in many compositional and monolithic VLMs [12, 18, 51] and have proved beneficial for model performance. PatchEmbed(\\cdot) refers to a patch embedding layer with stride " + }, + { + "bbox": [ + 313, + 460, + 554, + 614 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 460, + 554, + 614 + ], + "type": "text", + "content": " and converts the image tiles into a sequence " + }, + { + "bbox": [ + 313, + 460, + 554, + 614 + ], + "type": "inline_equation", + "content": "x_{I}" + }, + { + "bbox": [ + 313, + 460, + 554, + 614 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 460, + 554, + 614 + ], + "type": "inline_equation", + "content": "n_I" + }, + { + "bbox": [ + 313, + 460, + 554, + 614 + ], + "type": "text", + "content": " image tokens with dimension " + }, + { + "bbox": [ + 313, + 460, + 554, + 614 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 460, + 554, + 614 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 460, + 554, + 614 + ], + "type": "inline_equation", + "content": "\\mathrm{PE} \\in \\mathbb{R}^{n_I \\times c}" + }, + { + "bbox": [ + 313, + 460, + 554, + 614 + ], + "type": "text", + "content": " is a learnable position embedding. Tokenizer(\\cdot) converts raw texts into token indices and TextEmbed(\\cdot) is a word embedding layer with vocabulary size of " + }, + { + "bbox": [ + 313, + 460, + 554, + 614 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 313, + 460, + 554, + 614 + ], + "type": "text", + "content": " and dimension of " + }, + { + "bbox": [ + 313, + 460, + 554, + 614 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 460, + 554, + 614 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 460, + 554, + 614 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 313, + 460, + 554, + 614 + ], + "type": "text", + "content": " is a sequence of " + }, + { + "bbox": [ + 313, + 460, + 554, + 614 + ], + "type": "inline_equation", + "content": "n_T" + }, + { + "bbox": [ + 313, + 460, + 554, + 614 + ], + "type": "text", + "content": " text tokens." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 616, + 553, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 616, + 553, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 616, + 553, + 664 + ], + "type": "text", + "content": "Afterward, these image tokens and text tokens are combined as the input tokens " + }, + { + "bbox": [ + 313, + 616, + 553, + 664 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}^{n \\times c}" + }, + { + "bbox": [ + 313, + 616, + 553, + 664 + ], + "type": "text", + "content": ". The holistic embedding module utilizes a stack of causal Transformer layers to project these tokens to a shared embedding space:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 376, + 670, + 553, + 683 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 670, + 553, + 683 + ], + "spans": [ + { + "bbox": [ + 376, + 670, + 553, + 683 + ], + "type": "interline_equation", + "content": "\\hat {x} = \\text {C a u s a l T r a n s f o r m e r} (x), \\tag {3}", + "image_path": "9d2e1d73628057ef50265fe52a90f6a4edd25728b4321ccc1c48c5e140f5b248.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\hat{x}" + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": " denotes final embeddings of the input. The architecture of each Transformer layer is the same as the LLM's" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14561" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 75, + 69, + 275, + 358 + ], + "blocks": [ + { + "bbox": [ + 75, + 69, + 275, + 358 + ], + "lines": [ + { + "bbox": [ + 75, + 69, + 275, + 358 + ], + "spans": [ + { + "bbox": [ + 75, + 69, + 275, + 358 + ], + "type": "image", + "image_path": "5169ed9ce8be5ff8959ff8e84eafce012c32ca9fc19e612313ae56f1f254e43f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 153, + 366, + 203, + 376 + ], + "lines": [ + { + "bbox": [ + 153, + 366, + 203, + 376 + ], + "spans": [ + { + "bbox": [ + 153, + 366, + 203, + 376 + ], + "type": "text", + "content": "(a) Architecture" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 379, + 555, + 445 + ], + "lines": [ + { + "bbox": [ + 54, + 379, + 555, + 445 + ], + "spans": [ + { + "bbox": [ + 54, + 379, + 555, + 445 + ], + "type": "text", + "content": "Figure 3. (a) The architecture of HoVLE. HoVLE initially segments the input images into patches dynamically and tokenizes input texts. The holistic embedding module then projects them into a unified space. Finally, the LLM processes these unified embeddings to produce the final outputs. (b) The training strategies of HoVLE. Distillation stage trains the holistic embedding to distill a pre-trained vision encoder and text embeddings of the LLM using unpaired random images and texts. Alignment stage combines the holistic embedding module with a frozen LLM, conducting auto-regressive training to align the vision-language embeddings. Instruction tuning further enhance HoVLE's overall ability by tuning the whole model." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 282, + 70, + 531, + 358 + ], + "blocks": [ + { + "bbox": [ + 282, + 70, + 531, + 358 + ], + "lines": [ + { + "bbox": [ + 282, + 70, + 531, + 358 + ], + "spans": [ + { + "bbox": [ + 282, + 70, + 531, + 358 + ], + "type": "image", + "image_path": "bee8b4481e490d536919aaa562bc4d8c664184d0b2b69941f2a66a9abd788368.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 373, + 366, + 443, + 376 + ], + "lines": [ + { + "bbox": [ + 373, + 366, + 443, + 376 + ], + "spans": [ + { + "bbox": [ + 373, + 366, + 443, + 376 + ], + "type": "text", + "content": "(b) Training Strategies" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 452, + 295, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 452, + 295, + 524 + ], + "spans": [ + { + "bbox": [ + 55, + 452, + 295, + 524 + ], + "type": "text", + "content": "for simplicity and unity. Similar to LLMs, these Transformer layers use causal attention for both image and text tokens. The effectiveness of this architecture suggests that 1D causal modeling works as well as non-causal modeling for images in multi-modal tasks, aligning with recent findings in [42]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 530, + 296, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 530, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 530, + 296, + 590 + ], + "type": "text", + "content": "Monolithic Vision-Language Model. By combining the holistic embedding module with a pre-trained LLM, a monolithic VLM can be naturally built. After the holistic embedding module generates input embeddings, the LLM takes them as input and performs next-token prediction:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 125, + 603, + 295, + 616 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 603, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 125, + 603, + 295, + 616 + ], + "type": "interline_equation", + "content": "p _ {n + 1} = \\operatorname {L L M} (\\hat {x}; \\theta_ {\\text {L L M}}), \\tag {4}", + "image_path": "2eacb2e400486ddd14c5c8495b5618970d89e5ddceff669340b5082a4b3c606a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 628, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 628, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 295, + 652 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 628, + 295, + 652 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{LLM}}" + }, + { + "bbox": [ + 55, + 628, + 295, + 652 + ], + "type": "text", + "content": " refers to parameters of LLM, " + }, + { + "bbox": [ + 55, + 628, + 295, + 652 + ], + "type": "inline_equation", + "content": "p_{n + 1}\\in \\mathbb{R}^v" + }, + { + "bbox": [ + 55, + 628, + 295, + 652 + ], + "type": "text", + "content": " is the probability of the next token." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "type": "text", + "content": "In this way, we replace the modality-specific encoders in compositional VLMs with a holistic embedding module. This module has the same Transformer layer as the LLM, so the whole model presents a simple architecture, leading to our monolithic VLM." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 451, + 429, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 451, + 429, + 464 + ], + "spans": [ + { + "bbox": [ + 314, + 451, + 429, + 464 + ], + "type": "text", + "content": "3.2. Training Procedure" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 470, + 555, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 470, + 555, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 470, + 555, + 578 + ], + "type": "text", + "content": "To empower the holistic embedding module with decent vision and language encoding abilities, a multi-stage training procedure is carefully designed. The first stage trains the holistic embedding module to distill the image feature from a pre-trained visual encoder and the text embeddings from an LLM, providing general encoding abilities. The second stage combines the holistic embedding module with the LLM to perform auto-regressive training, aligning different modalities to a shared embedding space." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 582, + 555, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 667 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 667 + ], + "type": "text", + "content": "Distillation Stage. This stage aims to provide basic vision and language encoding abilities with the holistic embedding module. We extract image and text embeddings from the holistic embedding module output, denoted as " + }, + { + "bbox": [ + 313, + 582, + 555, + 667 + ], + "type": "inline_equation", + "content": "\\hat{x}_I" + }, + { + "bbox": [ + 313, + 582, + 555, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 582, + 555, + 667 + ], + "type": "inline_equation", + "content": "\\hat{x}_T" + }, + { + "bbox": [ + 313, + 582, + 555, + 667 + ], + "type": "text", + "content": ", respectively. A pre-trained visual encoder and an LLM embedding layer are utilized to compute the target image feature " + }, + { + "bbox": [ + 313, + 582, + 555, + 667 + ], + "type": "inline_equation", + "content": "z_I" + }, + { + "bbox": [ + 313, + 582, + 555, + 667 + ], + "type": "text", + "content": " and text feature " + }, + { + "bbox": [ + 313, + 582, + 555, + 667 + ], + "type": "inline_equation", + "content": "z_T" + }, + { + "bbox": [ + 313, + 582, + 555, + 667 + ], + "type": "text", + "content": ". The training objective is" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 323, + 675, + 555, + 693 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 675, + 555, + 693 + ], + "spans": [ + { + "bbox": [ + 323, + 675, + 555, + 693 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta_ {\\text {E m b e d}}} \\mathcal {L} _ {\\text {d i s t i l l}} \\left(\\hat {x} _ {I}, z _ {I}; \\theta_ {\\text {E m b e d}}\\right) + \\mathcal {L} _ {\\text {d i s t i l l}} \\left(\\hat {x} _ {T}, z _ {T}; \\theta_ {\\text {E m b e d}}\\right), \\tag {5}", + "image_path": "35c2ee47ab1c0ab2ebbfc7a3d59c8c8d1ea65f755ec43126cc1ec857c32b2e53.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 701, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 701, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 314, + 701, + 555, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 314, + 701, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{distill}}" + }, + { + "bbox": [ + 314, + 701, + 555, + 714 + ], + "type": "text", + "content": " is the negative cosine similarity between two" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14562" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": "features, " + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{Embed}}" + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": " refers to the parameters of the holistic embedding module." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 101, + 295, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 101, + 295, + 196 + ], + "spans": [ + { + "bbox": [ + 55, + 101, + 295, + 196 + ], + "type": "text", + "content": "Unpaired Random Images and Texts for Distillation. The distillation stage does not put any requirement on input images and texts, because their target features are computed independently. Therefore, we propose to combine random images with random texts as input. In fact, this strategy proves to be more effective than using paired image-text inputs, potentially due to better vocabulary coverage during distillation, as demonstrated in Section 4.3." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 197, + 296, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 197, + 296, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 197, + 296, + 304 + ], + "type": "text", + "content": "In practice, we randomly sample 500M images from LAION-2B [39] dataset and randomly sample 100 text tokens from the vocabulary for each image as training data. Each sample consists of a user query and a model response. We place the image at the beginning of the query, followed by the text content for the query and response, each with 50 random text tokens. Unlike previous monolithic VLMs, HoVLE can benefit from large-scale vision pre-training, unleashing the full potential of the monolithic VLM." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 309, + 296, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 309, + 296, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 309, + 296, + 357 + ], + "type": "text", + "content": "Alignment Stage. To align the image and text embeddings, the holistic embedding module is further combined with the LLM to perform auto-regressive training. The training objective is" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 102, + 361, + 295, + 386 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 361, + 295, + 386 + ], + "spans": [ + { + "bbox": [ + 102, + 361, + 295, + 386 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta_ {\\text {E m b e d}}} \\sum_ {i} \\mathcal {L} _ {\\text {a l i g n}} \\left(p _ {i}, y _ {i}; \\theta_ {\\text {E m b e d}}, \\theta_ {\\text {L L M}}\\right), \\tag {6}", + "image_path": "598b71830e2a81eeef348d9a8ee9295ea4d33f8298bf839b3f368ea39e86c892.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 391, + 295, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 391, + 295, + 438 + ], + "spans": [ + { + "bbox": [ + 55, + 391, + 295, + 438 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 391, + 295, + 438 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{align}}" + }, + { + "bbox": [ + 55, + 391, + 295, + 438 + ], + "type": "text", + "content": " is the cross-entropy loss, " + }, + { + "bbox": [ + 55, + 391, + 295, + 438 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 55, + 391, + 295, + 438 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 391, + 295, + 438 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 55, + 391, + 295, + 438 + ], + "type": "text", + "content": " denote the output probability and target index of the " + }, + { + "bbox": [ + 55, + 391, + 295, + 438 + ], + "type": "inline_equation", + "content": "i^{\\mathrm{th}}" + }, + { + "bbox": [ + 55, + 391, + 295, + 438 + ], + "type": "text", + "content": " text token. During training, the LLM is frozen, and only the holistic embedding module is optimized." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 439, + 296, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 439, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 439, + 296, + 510 + ], + "type": "text", + "content": "This stage is similar to the alignment stage used in compositional VLMs, but we do not have an extra MLP connector. During alignment, the model is trained on 45M multi-modal data from diverse domains, such as captioning, detection, and OCR data. Please refer to Appendix A for details." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 516, + 167, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 516, + 167, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 516, + 167, + 529 + ], + "type": "text", + "content": "3.3. Instruction Tuning" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 534, + 295, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 582 + ], + "type": "text", + "content": "The holistic embedding module is now empowered with strong vision and language encoding abilities. To further strengthen the whole VLM to follow instructions, a visual instruction tuning stage is incorporated." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 582, + 295, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 642 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 642 + ], + "type": "text", + "content": "In this stage, the whole model is optimized using the same auto-regressive training objective as in the alignment stage. The training data contains around 5M data, covering visual question answering, mathematics, world knowledge, etc. Please refer to Appendix A for details." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 651, + 132, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 651, + 132, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 651, + 132, + 664 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 671, + 188, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 188, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 188, + 684 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Network Architecture. HoVLE adopts InternLM2-1.8B [7] as the LLM backbone. The holistic embedding is" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "content": "composed of 8 causal Transformer layers, with hidden size 2048 and 16 attention heads, as in InternLM2-1.8B. The image patch embedding layer has a stride of " + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "inline_equation", + "content": "s = 28" + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "content": " and output dimension " + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "inline_equation", + "content": "c = 2048" + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "content": ". The text embedding has a vocabulary of 92,553. To demonstrate the high-resolution image processing ability of our model, like other works [5, 12, 24], we also provide HoVLE (HD). It uses a patch embedding with stride " + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "inline_equation", + "content": "s = 14" + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "content": " and applies pixel shuffle [11] after the holistic embedding to reduce the number of visual tokens to 1/4. All other details are kept the same." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 198, + 553, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 198, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 553, + 342 + ], + "type": "text", + "content": "Training Procedure. During distillation, we randomly sample 500M images from LAION-2B [39] with 100 random text indices for each image. We use InternViT-300M [9, 11] and the text embedding from InternLM2-1.8B as the teacher model for image and text tokens, respectively. For alignment, a collection of datasets containing 50M samples is used, as shown in Appendix A. For instruction tuning, a collection of datasets containing 5M samples is used, as shown in Appendix A. We use AdamW as the optimizer for all three stages, with learning rates 3e-4, 5e-5, and 4e-5, respectively. The batch size is set to 4096 for all stages. Please refer to Appendix A for more hyper-parameters." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 347, + 555, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 347, + 555, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 347, + 555, + 563 + ], + "type": "text", + "content": "Evaluation Benchmarks. We evaluate our model on 17 public benchmarks. MMbench [30], MME [14], MMMU [55], MM-Vet [54], and SEED [21] focus on multimodal perception and reasoning, providing a comprehensive assessment of VLM capabilities. POPE [23] and HallB [15] specifically target hallucination detection. For OCR and OCR-related visual question answering, TextVQA [40], ChartQA [33], InfoVQA [35], DocVQA [34], and OCRBench [29] are used. GQA [19] evaluates visual scene understanding, while ScienceQA [32] and AI2D [20] assess scientific image understanding. MathVista focuses on mathematical reasoning. For specific metric values shown in the experiments, MMbench-EN results are counted and presented as the MMB score. The POPE score is averaged across its three categories, the MME score is calculated as the sum of perception and cognition scores, and the SEED score reflects image accuracy specifically." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 574, + 399, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 574, + 399, + 586 + ], + "spans": [ + { + "bbox": [ + 314, + 574, + 399, + 586 + ], + "type": "text", + "content": "4.2. Main Results" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "We compare HoVLE with previous compositional and monolithic VLMs in Table 1 and Table 2. We highlight several key observations as follows: (1) There are significant performance gaps between previous monolithic VLMs and compositional VLMs, indicating the challenges in building a monolithic VLM with state-of-the-art performance. (2) Compared with existing monolithic VLMs, our HoVLE demonstrates substantially better performance with fewer parameters. For example, HoVLE surpasses the previous best monolithic VLM by " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\sim 15" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": " points on MMBench, which" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14563" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 111, + 70, + 497, + 305 + ], + "blocks": [ + { + "bbox": [ + 111, + 70, + 497, + 305 + ], + "lines": [ + { + "bbox": [ + 111, + 70, + 497, + 305 + ], + "spans": [ + { + "bbox": [ + 111, + 70, + 497, + 305 + ], + "type": "table", + "html": "
Method# ParamsMMBMMEMMMUMM-VetPOPESEEDMathVistaHallBCCB
Compositional VLMs
VILA1.5-3B [25]3.1B63.4164833.335.485.967.931.631.224.1
DeepSeek-VL-1.3B [31]2.0B64.6153232.234.887.666.731.127.637.6
PaliGemma-3B [6]2.9B71.0168634.933.187.069.628.732.229.6
MM1-3B-Chat [36]3B67.8176233.943.787.468.832.0--
MiniCPM-V [53]2.8B64.1165038.331.179.565.628.936.241.4
MiniCPM-V-2 [53]2.8B69.1180938.241.086.367.138.736.145.3
InternVL1.5 [9]2.2B70.9190234.639.387.169.841.137.563.5
†InternVL2 [10]2.2B73.2187734.344.688.371.646.437.974.7
Qwen2-VL-2B [46]2.2B74.9187241.149.5--43.041.7-
Phi-3 Vision [1]4.2B80.5-40.4-85.8-44.5--
Monolithic VLMs
Fuyu-8B (HD) [5]8B10.7--21.474.1----
SOLO [8]7B-1260---64.434.4--
*Chameleon-7B [43]7B31.117025.48.3-30.622.317.13.5
EVE-7B [12]7B49.5148332.325.683.661.325.221.112.4
EVE-7B (HD) [12]7B52.3162832.625.785.064.634.226.416.3
Emu3 [49]8B58.5-31.637.285.268.2---
Our HoVLE2.6B71.9186433.744.387.670.746.239.675.3
Our HoVLE (HD)2.6B73.3186232.243.887.470.949.238.474.3
", + "image_path": "ae9a083c9e10ca6170e39490821ffa7f2fdf0734ae1bd086e4a17670173ddfa1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 311, + 555, + 354 + ], + "lines": [ + { + "bbox": [ + 55, + 311, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 311, + 555, + 354 + ], + "type": "text", + "content": "Table 1. Comparison with existing VLMs on general VLM benchmarks. The highest scores in monolithic VLMs are highlighted in bold. *We test Chameleon-7B with its released model, but it frequently refuses to perform the task with a response of \"I can't help you with this\", thus resulting in poor performance. " + }, + { + "bbox": [ + 55, + 311, + 555, + 354 + ], + "type": "inline_equation", + "content": "{}^{ \\dagger }" + }, + { + "bbox": [ + 55, + 311, + 555, + 354 + ], + "type": "text", + "content": " InternVL2 2.2B adopts the same LLM and supervised data with our HoVLE,sowe mark it as the compositional counterpart." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 357, + 295, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 357, + 295, + 573 + ], + "spans": [ + { + "bbox": [ + 54, + 357, + 295, + 573 + ], + "type": "text", + "content": "is a comprehensive multi-modal benchmark covering a variety of fine-grained abilities. This substantial improvement is attributed to the holistic embedding, which prevents continual pre-training from disrupting language knowledge stored in pre-trained LLMs. (3) By incorporating higher resolution, HoVLE (HD) consistently improves the performance on VQA benchmarks, demonstrating the versatility of our holistic embedding in handling different resolutions. (4) Compared with previous compositional VLMs, HoVLE (HD) is able to deliver competitive results with many of these models. This shows that the introduction of holistic embedding provides an effective encoding that allows the LLM to understand images while preserving its language abilities. While HoVLE (HD) still falls short of state-of-the-art compositional VLMs, like Qwen2-VL2B [46] and Phi-3 Vision [1], it narrows the gap between monolithic and compositional VLMs, suggesting the feasibility of a high-performance monolithic VLM." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 582, + 155, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 155, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 155, + 594 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 600, + 295, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 295, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 295, + 649 + ], + "type": "text", + "content": "In this section, we aim to display and discuss the design choices that affect our HoVLE. Ablation experiments train the model with 300M and 8M data during the distillation and alignment stages respectively, unless specified." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 653, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 715 + ], + "type": "text", + "content": "Holistic Embedding Depth. To validate the effectiveness of the holistic embedding, we first compare models with different holistic embedding depths in Table 3. With 0 Transformer layer, the model degenerates to a VLM without holistic embedding and the poor performance suggests" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 357, + 555, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 555, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 555, + 418 + ], + "type": "text", + "content": "the importance of this module. The model performs consistently better as the depth goes from 4 to 12. We speculate that deep embedding is beneficial for extracting visual representation. HoVLE uses 8 layers to achieve a balance between performance and model complexity." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 420, + 556, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 420, + 556, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 420, + 556, + 564 + ], + "type": "text", + "content": "Training Strategy. Table 4 displays the influence of different training stages. The results show that omitting any stage will result in a performance drop. Among these, removing the distillation stage leads to the most severe degradation, highlighting its importance. We observe that models after distillation exhibit a faster decrease in loss during subsequent training stages, likely because the distillation provides general abilities of representation extraction, which is helpful for accelerating training. The improvement from instruction tuning is greater than that of the alignment stage. We assume this is because the LLM has more trainable parameters, allowing the whole model to fit the data well." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 567, + 556, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 556, + 674 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 556, + 674 + ], + "type": "text", + "content": "Distillation Data Scaling. Figure 6 illustrates how the performance changes as distillation data scales up. Overall, the model performs better with more distillation data, implying the importance of comprehensive training. Specifically, the model's performance improves rapidly with 300M data and continues to increase at a more gradual pace as more data is added. We use 500M distillation data as our default setting. Please refer to Appendix B for scaling performance on more benchmarks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "Text Data for Distillation. Table 5 presents the effects of different text input choices during distillation. The LAION [39] dataset provides a short text caption for each" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14564" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 103, + 70, + 506, + 305 + ], + "blocks": [ + { + "bbox": [ + 103, + 70, + 506, + 305 + ], + "lines": [ + { + "bbox": [ + 103, + 70, + 506, + 305 + ], + "spans": [ + { + "bbox": [ + 103, + 70, + 506, + 305 + ], + "type": "table", + "html": "
Method# ParamsTextVQAChartVQAInfoVQADocVQAOCRbenchGQASQA-IAI2D
Compositional VLMs
VILA1.5-3B [25]3.1B60.4---43761.569.0-
DeepSeek-VL-1.3B [31]2.0B57.8---409--51.5
PaliGemma-3B [6]2.9B68.1---614--68.3
MM1-3B-Chat [36]3B71.9-----69.4-
MiniCPM-V [53]2.8B60.6--38.2366--56.3
MiniCPM-V-2 [53]2.8B74.1--71.9605--62.9
InternVL1.5 [9]2.2B70.574.855.485.065461.684.969.8
†InternVL2 [10]2.2B73.476.257.785.978461.084.974.1
Qwen2-VL-2B [46]2.2B79.773.565.590.1794---
Phi3-Vision [1]4.2B70.981.4----90.876.7
Monolithic VLMs
Fuyu-8B (HD) [5]8B-------64.5
SOLO [8]7B------73.361.4
*Chameleon-7B [43]7B4.82.95.01.57-47.246.0
EVE-7B [12]7B51.919.520.022.032760.863.048.5
EVE-7B (HD) [12]7B56.859.125.053.039862.664.959.1
Emu3 [49]8B64.768.643.876.368760.389.270.0
Our HoVLE2.6B66.077.652.184.773563.694.873.7
Our HoVLE (HD)2.6B70.978.655.786.174064.994.873.0
", + "image_path": "8f2d2cf52876d36c033e6acf61ad3046bb4810ffc940940cb5751c9af9c1f47a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 69, + 334, + 185, + 416 + ], + "blocks": [ + { + "bbox": [ + 69, + 334, + 185, + 416 + ], + "lines": [ + { + "bbox": [ + 69, + 334, + 185, + 416 + ], + "spans": [ + { + "bbox": [ + 69, + 334, + 185, + 416 + ], + "type": "image", + "image_path": "92f875096f322e90090b082c3facfe66c6277160608b67f0f0b75987a352cdd5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 187, + 334, + 302, + 415 + ], + "blocks": [ + { + "bbox": [ + 187, + 334, + 302, + 415 + ], + "lines": [ + { + "bbox": [ + 187, + 334, + 302, + 415 + ], + "spans": [ + { + "bbox": [ + 187, + 334, + 302, + 415 + ], + "type": "image", + "image_path": "d50f9ad8eae487fb8b973c2434a686aa556f2ff6e22950649d71fb29cd95f35e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 214, + 415, + 395, + 426 + ], + "lines": [ + { + "bbox": [ + 214, + 415, + 395, + 426 + ], + "spans": [ + { + "bbox": [ + 214, + 415, + 395, + 426 + ], + "type": "text", + "content": "Figure 4. Distillation data scaling performance." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 306, + 334, + 421, + 415 + ], + "blocks": [ + { + "bbox": [ + 306, + 334, + 421, + 415 + ], + "lines": [ + { + "bbox": [ + 306, + 334, + 421, + 415 + ], + "spans": [ + { + "bbox": [ + 306, + 334, + 421, + 415 + ], + "type": "image", + "image_path": "4d1605567c118e5b519892f419bdf0820702a124d90fab752e0c35a5656b77dc.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 423, + 334, + 538, + 416 + ], + "blocks": [ + { + "bbox": [ + 423, + 334, + 538, + 416 + ], + "lines": [ + { + "bbox": [ + 423, + 334, + 538, + 416 + ], + "spans": [ + { + "bbox": [ + 423, + 334, + 538, + 416 + ], + "type": "image", + "image_path": "253e95810dc5ced324554bbda1376815d0d2e8c5db25aebb9111f9ad32fb97a4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 70, + 434, + 281, + 500 + ], + "blocks": [ + { + "bbox": [ + 55, + 306, + 555, + 329 + ], + "lines": [ + { + "bbox": [ + 55, + 306, + 555, + 329 + ], + "spans": [ + { + "bbox": [ + 55, + 306, + 555, + 329 + ], + "type": "text", + "content": "Table 2. Comparison with existing VLMs on visual question answering benchmarks. The highest scores in monolithic VLMs are highlighted in bold. Same as Tab. 1, *Chameleon-7B frequently refuses to response, and " + }, + { + "bbox": [ + 55, + 306, + 555, + 329 + ], + "type": "inline_equation", + "content": "{}^{ \\dagger }" + }, + { + "bbox": [ + 55, + 306, + 555, + 329 + ], + "type": "text", + "content": " InternVL2 2.2B is our compositional counterpart." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 70, + 434, + 281, + 500 + ], + "lines": [ + { + "bbox": [ + 70, + 434, + 281, + 500 + ], + "spans": [ + { + "bbox": [ + 70, + 434, + 281, + 500 + ], + "type": "table", + "html": "
DepthMMBMMESEEDTextVQAInfoVQADocVQA
035.4114946.011.319.112.2
467.8180168.047.742.673.6
872.0186270.962.051.482.1
1272.3188270.965.552.584.7
", + "image_path": "94b837e9eca76f663c3eda822bff19e4f1328822d043315623c06894ae5a1c69.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 62, + 516, + 291, + 579 + ], + "blocks": [ + { + "bbox": [ + 85, + 500, + 265, + 510 + ], + "lines": [ + { + "bbox": [ + 85, + 500, + 265, + 510 + ], + "spans": [ + { + "bbox": [ + 85, + 500, + 265, + 510 + ], + "type": "text", + "content": "Table 3. Ablation on holistic embedding depth." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 62, + 516, + 291, + 579 + ], + "lines": [ + { + "bbox": [ + 62, + 516, + 291, + 579 + ], + "spans": [ + { + "bbox": [ + 62, + 516, + 291, + 579 + ], + "type": "table", + "html": "
Dist. Align. I.T.MMBMMESEEDTextVQAInfoVQADocVQA
✓✓37.3127446.411.219.3
✓✓71.3182870.560.149.2
✓✓66.1169866.755.043.0
✓✓✓72.0186270.962.051.4
", + "image_path": "df8d1deac675cc843e4af3c1ba60f477d49b8b3a87949e5c34587c5faf88e8b1.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 63, + 596, + 287, + 639 + ], + "blocks": [ + { + "bbox": [ + 99, + 581, + 252, + 591 + ], + "lines": [ + { + "bbox": [ + 99, + 581, + 252, + 591 + ], + "spans": [ + { + "bbox": [ + 99, + 581, + 252, + 591 + ], + "type": "text", + "content": "Table 4. Ablation on training strategies." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 596, + 287, + 639 + ], + "lines": [ + { + "bbox": [ + 63, + 596, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 63, + 596, + 287, + 639 + ], + "type": "table", + "html": "
Text DataMMBMMESEEDTextVQAInfoVQADocVQA
Paired71.7176870.561.349.481.5
Random72.0186270.962.051.482.1
", + "image_path": "6f47121e389921b3ae89a38f2784a304dae19e6127d96472bbda826fef117158.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 88, + 640, + 262, + 649 + ], + "lines": [ + { + "bbox": [ + 88, + 640, + 262, + 649 + ], + "spans": [ + { + "bbox": [ + 88, + 640, + 262, + 649 + ], + "type": "text", + "content": "Table 5. Ablation on text data for distillation." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 666, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 294, + 713 + ], + "type": "text", + "content": "image, so we can also use image-text pairs as input. However, this approach yields relatively lower results, and the model continues to exhibit high loss values during alignment. We hypothesize that the text data from the LAION" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 437, + 555, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 555, + 533 + ], + "type": "text", + "content": "dataset lacks sufficient diversity to cover the full vocabulary, suggesting that paired data may require higher quality and variety in texts. In contrast, we propose using random texts as input. Table 5 shows that this strategy outperforms paired text data. We also observe lower loss during alignment. By leveraging random text, we reduce the high-quality need for image-text pair collection, thereby enhancing data utilization for large-scale training." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 540, + 376, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 376, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 376, + 552 + ], + "type": "text", + "content": "4.4. Analysis" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 557, + 555, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 555, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 555, + 677 + ], + "type": "text", + "content": "Visualization. To better understand the effects of our introduced holistic embedding, we visualize the attention map of the LLM in Figure 5. We find that existing monolithic models, such as EVE [12] and Emu3 [49], exhibit denser text-to-image attention in their initial layers, while the compositional InternVL2 [10] and our HoVLE model maintain sparse attention patterns throughout. All models demonstrate sparse attention patterns in the later layers (with the final layers shown). Please refer to Appendix C for more attention maps in the middle layers." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "We hypothesize that sparser text-to-image attention indicates that visual features are effectively extracted and well-aligned with text tokens, allowing the text tokens to focus" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14565" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 97, + 70, + 194, + 176 + ], + "blocks": [ + { + "bbox": [ + 97, + 70, + 194, + 176 + ], + "lines": [ + { + "bbox": [ + 97, + 70, + 194, + 176 + ], + "spans": [ + { + "bbox": [ + 97, + 70, + 194, + 176 + ], + "type": "image", + "image_path": "2699d803a1aad92a6bcca7e9f35c86d5a743bc99d9cbdefc38ceab6755bf659b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 205, + 70, + 289, + 175 + ], + "blocks": [ + { + "bbox": [ + 205, + 70, + 289, + 175 + ], + "lines": [ + { + "bbox": [ + 205, + 70, + 289, + 175 + ], + "spans": [ + { + "bbox": [ + 205, + 70, + 289, + 175 + ], + "type": "image", + "image_path": "1d76128a5a51400e1a75e7954d9846d1ad48833416451fb0e580fd0eee6bed82.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 303, + 70, + 386, + 175 + ], + "blocks": [ + { + "bbox": [ + 303, + 70, + 386, + 175 + ], + "lines": [ + { + "bbox": [ + 303, + 70, + 386, + 175 + ], + "spans": [ + { + "bbox": [ + 303, + 70, + 386, + 175 + ], + "type": "image", + "image_path": "39a9cdb1effd2a6b2a66132b55f294617849ba9899c5dd765ae02a0dde89e23a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 399, + 71, + 481, + 175 + ], + "blocks": [ + { + "bbox": [ + 399, + 71, + 481, + 175 + ], + "lines": [ + { + "bbox": [ + 399, + 71, + 481, + 175 + ], + "spans": [ + { + "bbox": [ + 399, + 71, + 481, + 175 + ], + "type": "image", + "image_path": "b1ed64c45eab64b2819afccb46f204fcfe159456cad018663c541d6f47847ba0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 486, + 78, + 514, + 266 + ], + "blocks": [ + { + "bbox": [ + 486, + 78, + 514, + 266 + ], + "lines": [ + { + "bbox": [ + 486, + 78, + 514, + 266 + ], + "spans": [ + { + "bbox": [ + 486, + 78, + 514, + 266 + ], + "type": "image", + "image_path": "2cdd9aeb0759e22cb2fb3cb5a3330d9e8008a1021ecd47469371e38e63992505.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 96, + 177, + 192, + 272 + ], + "blocks": [ + { + "bbox": [ + 96, + 177, + 192, + 272 + ], + "lines": [ + { + "bbox": [ + 96, + 177, + 192, + 272 + ], + "spans": [ + { + "bbox": [ + 96, + 177, + 192, + 272 + ], + "type": "image", + "image_path": "873bf3203cc994facceac438dd30b9ea4eccf04207f4d967abf28ad4387bb017.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 273, + 555, + 328 + ], + "lines": [ + { + "bbox": [ + 55, + 273, + 555, + 328 + ], + "spans": [ + { + "bbox": [ + 55, + 273, + 555, + 328 + ], + "type": "text", + "content": "Figure 5. Attention Maps for EVE, Emu3, InternVL2 and our HoVLE at the first and last layers of LLM backbones. Y-axis represents query tokens, and X-axis represents key tokens, with text modality tokens in gray and image modality tokens in yellow. All four models share the same input, but the sequence lengths of input tokens are different due to different image pre-processing. We highlight text-to-image attention below each full attention map. Our HoVLE, like the compositional InternVL2, has sparse attention across all network layers, while other monolithic models Emu3 and EVE have denser attention in shallow layers." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 206, + 177, + 289, + 272 + ], + "blocks": [ + { + "bbox": [ + 206, + 177, + 289, + 272 + ], + "lines": [ + { + "bbox": [ + 206, + 177, + 289, + 272 + ], + "spans": [ + { + "bbox": [ + 206, + 177, + 289, + 272 + ], + "type": "image", + "image_path": "24475e612ea96d33a8c1df1105e18bf343cce755f1af31d70de9a5bc384472f6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 304, + 177, + 386, + 272 + ], + "blocks": [ + { + "bbox": [ + 304, + 177, + 386, + 272 + ], + "lines": [ + { + "bbox": [ + 304, + 177, + 386, + 272 + ], + "spans": [ + { + "bbox": [ + 304, + 177, + 386, + 272 + ], + "type": "image", + "image_path": "2a743b2267515adae460315f409e22a74599db4cb1386e0d17a39ce353cda6f8.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 400, + 177, + 482, + 272 + ], + "blocks": [ + { + "bbox": [ + 400, + 177, + 482, + 272 + ], + "lines": [ + { + "bbox": [ + 400, + 177, + 482, + 272 + ], + "spans": [ + { + "bbox": [ + 400, + 177, + 482, + 272 + ], + "type": "image", + "image_path": "d2ddb53de9b6f618ac6bcc77f5fdd57be9e45d9fb61d0d9430f7d2f5b7c20cfb.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 58, + 331, + 299, + 407 + ], + "blocks": [ + { + "bbox": [ + 58, + 331, + 299, + 407 + ], + "lines": [ + { + "bbox": [ + 58, + 331, + 299, + 407 + ], + "spans": [ + { + "bbox": [ + 58, + 331, + 299, + 407 + ], + "type": "table", + "html": "
ModelTile ResolutionFLOPs(T)↓TTFT↓TPS↑Avg. Score↑
InternVL2448 × 4481.811.3728165.8
HoVLE (HD)336 × 3361.511.2424365.7
HoVLE (HD)392 × 3921.891.5222466.1
HoVLE (HD)448 × 4482.181.8420666.8
", + "image_path": "619503184fe418f96609e5b727f2ff3a38c33cdb96c986109e948fdedda484a5.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 412, + 297, + 467 + ], + "lines": [ + { + "bbox": [ + 55, + 412, + 297, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 412, + 297, + 467 + ], + "type": "text", + "content": "Table 6. Inference speed-performance trade-off of HoVLE (HD). \"TTFT\" and \"TPS\" denotes the time to first token in seconds and throughput in tokens per second, respectively. \"Avg. Score\" refers to the average performance over 17 benchmarks. Please refer to Appendix for detailed results." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 472, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 295, + 544 + ], + "type": "text", + "content": "only on the most relevant visual information. In contrast, denser text-to-image attention may suggest that the model is still in the process of extracting visual features within the LLM component. Therefore, by introducing holistic embedding, the LLM may process vision and language inputs more effectively." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 545, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 714 + ], + "type": "text", + "content": "Speed-Performance Trade-off. We report the speed-performance trade-off of HoVLE (HD) in Table 6. For evaluation, we use one image and 256 text tokens as input, with the model restricted to output 120 tokens during testing. InternVL2 serves as the main comparison model, as it uses the same LLM. When the tile resolution of DynProcess(" + }, + { + "bbox": [ + 55, + 545, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 55, + 545, + 295, + 714 + ], + "type": "text", + "content": ") is 336, HoVLE (HD) achieves comparable performance to InternVL2, while also showing lower computational demand. As the input resolution increases, HoVLE's performance improves consistently, but at the cost of slower speed. Additionally, we observe that using the KV cache [38] during inference significantly reduces computational time, making kernel launch the main bottleneck in TPS. Optimizing the kernel could further enhance processing speed." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 332, + 388, + 345 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 332, + 388, + 345 + ], + "spans": [ + { + "bbox": [ + 314, + 332, + 388, + 345 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 352, + 555, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 352, + 555, + 591 + ], + "spans": [ + { + "bbox": [ + 313, + 352, + 555, + 591 + ], + "type": "text", + "content": "This work presents HoVLE, a monolithic Vision-Language Model (VLM) that processes images and texts in a unified manner. HoVLE introduces a holistic embedding module that projects image and text inputs into a shared embedding space. This allows the Large Language Model (LLM) to interpret images in the same way as texts. Furthermore, a multi-stage training procedure is designed to strengthen the encoding abilities of the holistic embedding. A distillation stage first provides general vision and language encoding capabilities, while an alignment stage aligns different modalities. An instruction tuning stage enhances the overall abilities in the end. Extensive experiments show that HoVLE surpasses previous monolithic VLMs and demonstrates competitive performance with compositional VLMs. This work focuses on enhancing monolithic VLMs by introducing a holistic embedding module, aiming to elevate their capabilities rather than rival those of compositional VLMs. This work narrows the gap between monolithic and compositional VLMs, providing a promising direction for the development of monolithic VLMs." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 592, + 555, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 592, + 555, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 592, + 555, + 651 + ], + "type": "text", + "content": "Limitations. Due to limited computational resources, our experiments are conducted on 2B parameter scale. Scaling to larger scales can provide further insight into the model scalability of our approach, which we hope to explore in future work." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 651, + 555, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 651, + 555, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 651, + 555, + 700 + ], + "type": "text", + "content": "Acknowledgements. This work is supported by the National Key R&D Program of China (NO. 2022ZD0161300), by the National Natural Science Foundation of China (62376134)." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14566" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Marah Abdin, Jyoti Aneja, Hany Awadalla, Ahmed Awadallah, Ammar Ahmad Awan, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Jianmin Bao, Harkirat Behl, et al. Phi-3 technical report: A highly capable language model locally on your phone. arxiv:2404.14219, 2024. 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 147, + 296, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 147, + 296, + 190 + ], + "spans": [ + { + "bbox": [ + 61, + 147, + 296, + 190 + ], + "type": "text", + "content": "[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arxiv:2303.08774, 2023. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 191, + 294, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 191, + 294, + 213 + ], + "spans": [ + { + "bbox": [ + 62, + 191, + 294, + 213 + ], + "type": "text", + "content": "[3] AI Anthropic. The claude 3 model family: Opus, sonnet, haiku. Claude-3 Model Card, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 214, + 294, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 214, + 294, + 257 + ], + "spans": [ + { + "bbox": [ + 62, + 214, + 294, + 257 + ], + "type": "text", + "content": "[4] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arxiv:2308.12966, 2023. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 258, + 294, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 258, + 294, + 302 + ], + "spans": [ + { + "bbox": [ + 62, + 258, + 294, + 302 + ], + "type": "text", + "content": "[5] Rohan Bavishi, Erich Elsen, Curtis Hawthorne, Maxwell Nye, Augustus Odena, Arushi Somani, and Sagnak Tasirlar. Introducing our multimodal models. https://www. adept.ai/blog/fuyu-8b, 2023. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 303, + 294, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 303, + 294, + 357 + ], + "spans": [ + { + "bbox": [ + 62, + 303, + 294, + 357 + ], + "type": "text", + "content": "[6] Lucas Beyer, Andreas Steiner, André Susano Pinto, Alexander Kolesnikov, Xiao Wang, Daniel Salz, Maxim Neumann, Ibrahim Alabdulmohsin, Michael Tschannen, Emanuele Bugliarello, et al. Paligemma: A versatile 3b vlm for transfer. arxiv:2407.07726, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 358, + 294, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 358, + 294, + 401 + ], + "spans": [ + { + "bbox": [ + 62, + 358, + 294, + 401 + ], + "type": "text", + "content": "[7] Zheng Cai, Maosong Cao, Haojiong Chen, Kai Chen, Keyu Chen, Xin Chen, Xun Chen, Zehui Chen, Zhi Chen, Pei Chu, et al. Internl m2 technical report. arxiv:2403.17297, 2024. 1, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 403, + 294, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 403, + 294, + 435 + ], + "spans": [ + { + "bbox": [ + 62, + 403, + 294, + 435 + ], + "type": "text", + "content": "[8] Yangyi Chen, Xingyao Wang, Hao Peng, and Heng Ji. A single transformer for scalable vision-language modeling. arxiv:2407.06438, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 437, + 294, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 437, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 62, + 437, + 294, + 491 + ], + "type": "text", + "content": "[9] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arxiv:2404.16821, 2024. 5, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 492, + 294, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 492, + 294, + 568 + ], + "spans": [ + { + "bbox": [ + 57, + 492, + 294, + 568 + ], + "type": "text", + "content": "[10] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. Internv12: Better than the best—expanding performance boundaries of open-source multimodal models with the progressive scaling strategy. https://internvl.github.io/blog/2024-07-02-InternVL-2.0/, 2024.6.7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 570, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 570, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 570, + 294, + 624 + ], + "type": "text", + "content": "[11] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In CVPR, pages 24185–24198, 2024. 1, 3, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 625, + 294, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 625, + 294, + 668 + ], + "spans": [ + { + "bbox": [ + 57, + 625, + 294, + 668 + ], + "type": "text", + "content": "[12] Haiwen Diao, Yufeng Cui, Xiaotong Li, Yueze Wang, Huchuan Lu, and Xinlong Wang. Unveiling encoder-free vision-language models. arxiv:2406.11832, 2024. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "type": "text", + "content": "[13] Zhiyuan Fang, Jianfeng Wang, Xiaowei Hu, Lijuan Wang, Yezhou Yang, and Zicheng Liu. Compressing visual-linguistic model via knowledge distillation. In ICCV, pages 1428-1438, 2021. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "text", + "content": "[14] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. arxiv:2306.13394, 2023.5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 129, + 553, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 553, + 193 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 553, + 193 + ], + "type": "text", + "content": "[15] Tianrui Guan, Fuxiao Liu, Xiyang Wu, Ruiqi Xian, Zongxia Li, Xiaoyu Liu, Xijun Wang, Lichang Chen, Furong Huang, Yaser Yacoob, et al. Hallusionbench: An advanced diagnostic suite for entangled language hallucination and visual illusion in large vision-language models. arxiv:2310.14566, 2023. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 194, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 194, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 317, + 194, + 553, + 239 + ], + "type": "text", + "content": "[16] Suchin Gururangan, Ana Marasovic, Swabha Swayamdipta, Kyle Lo, Iz Beltagy, Doug Downey, and Noah A Smith. Don't stop pretraining: Adapt language models to domains and tasks. In ACL, pages 8342-8360, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 239, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 239, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 317, + 239, + 553, + 293 + ], + "type": "text", + "content": "[17] Yushi Hu, Otilia Stretcu, Chun-Ta Lu, Krishnamurthy Viswanathan, Kenji Hata, Enming Luo, Ranjay Krishna, and Ariel Fuxman. Visual program distillation: Distilling tools and programmatic reasoning into vision-language models. In CVPR, pages 9590-9601, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 294, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 294, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 317, + 294, + 553, + 348 + ], + "type": "text", + "content": "[18] Runhui Huang, Xinpeng Ding, Chunwei Wang, Jianhua Han, Yulong Liu, Hengshuang Zhao, Hang Xu, Lu Hou, Wei Zhang, and Xiaodan Liang. Hires-llava: Restoring fragmentation input in high-resolution large vision-language models. arXiv:2407.08706, 2024.3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 350, + 553, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 350, + 553, + 382 + ], + "spans": [ + { + "bbox": [ + 317, + 350, + 553, + 382 + ], + "type": "text", + "content": "[19] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In CVPR, pages 6700-6709, 2019. 5, 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 383, + 553, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 383, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 317, + 383, + 553, + 425 + ], + "type": "text", + "content": "[20] Aniruddha Kembhavi, Mike Salvato, Eric Kolve, Minjoon Seo, Hannaneh Hajishirzi, and Ali Farhadi. A diagram is worth a dozen images. In ECCV, pages 235-251. Springer, 2016. 5, 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 426, + 553, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 426, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 317, + 426, + 553, + 469 + ], + "type": "text", + "content": "[21] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arxiv:2307.16125, 2023. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 471, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 471, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 317, + 471, + 553, + 514 + ], + "type": "text", + "content": "[22] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, pages 12888-12900. PMLR, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 515, + 553, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 515, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 317, + 515, + 553, + 557 + ], + "type": "text", + "content": "[23] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. In EMNLP, pages 292-305, 2023. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 559, + 553, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 559, + 553, + 612 + ], + "spans": [ + { + "bbox": [ + 317, + 559, + 553, + 612 + ], + "type": "text", + "content": "[24] Zhang Li, Biao Yang, Qiang Liu, Zhiyin Ma, Shuo Zhang, Jingxu Yang, Yabo Sun, Yuliang Liu, and Xiang Bai. Monkey: Image resolution and text label are important things for large multi-modal models. In CVPR, pages 26763-26773, 2024. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 614, + 553, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 614, + 553, + 647 + ], + "spans": [ + { + "bbox": [ + 317, + 614, + 553, + 647 + ], + "type": "text", + "content": "[25] Ji Lin, Hongxu Yin, Wei Ping, Pavlo Molchanov, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models. In CVPR, pages 26689-26699, 2024. 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 647, + 553, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 553, + 680 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 553, + 680 + ], + "type": "text", + "content": "[26] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In CVPR, pages 26296-26306, 2024. 1, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 681, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 681, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 317, + 681, + 553, + 714 + ], + "type": "text", + "content": "[27] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge. https :" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14567" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "type": "text", + "content": "//llava-v1.github.io/blog/2024-01-30-llava-last/, 2024.1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 96, + 294, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 96, + 294, + 118 + ], + "spans": [ + { + "bbox": [ + 57, + 96, + 294, + 118 + ], + "type": "text", + "content": "[28] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 119, + 294, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 119, + 294, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 119, + 294, + 163 + ], + "type": "text", + "content": "[29] Yuliang Liu, Zhang Li, Biao Yang, Chunyuan Li, Xucheng Yin, Cheng-lin Liu, Lianwen Jin, and Xiang Bai. On the hidden mystery ofOCR in large multimodal models. arxiv:2305.07895, 2023. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 164, + 294, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 164, + 294, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 294, + 218 + ], + "type": "text", + "content": "[30] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? In ECCV, pages 216-233. Springer, 2025. 2, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 220, + 294, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 264 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 264 + ], + "type": "text", + "content": "[31] Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhuoshu Li, Hao Yang, et al. Deepseek-vl: towards real-world vision-language understanding. arxiv:2403.05525, 2024. 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 266, + 294, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 266, + 294, + 320 + ], + "spans": [ + { + "bbox": [ + 56, + 266, + 294, + 320 + ], + "type": "text", + "content": "[32] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. In NeurIPS, pages 2507–2521, 2022. 5, 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 322, + 294, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 322, + 294, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 322, + 294, + 365 + ], + "type": "text", + "content": "[33] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arxiv:2203.10244, 2022. 5, 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 366, + 294, + 400 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 366, + 294, + 400 + ], + "spans": [ + { + "bbox": [ + 56, + 366, + 294, + 400 + ], + "type": "text", + "content": "[34] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In WACV, pages 2200-2209, 2021. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 401, + 294, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 401, + 294, + 434 + ], + "spans": [ + { + "bbox": [ + 56, + 401, + 294, + 434 + ], + "type": "text", + "content": "[35] Minesh Mathew, Viraj Bagal, Ruben Tito, Dimosthenis Karatzas, Ernest Valveny, and CV Jawahar. Infographicvqa. In WACV, pages 1697-1706, 2022. 5, 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 434, + 294, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 294, + 490 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 294, + 490 + ], + "type": "text", + "content": "[36] Brandon McKinzie, Zhe Gan, Jean-Philippe Fauconnier, Sam Dodge, Bowen Zhang, Philipp Dufter, Dhruti Shah, Xianzhi Du, Futang Peng, Floris Weers, et al. Mm1: Methods, analysis & insights from multimodal llm pre-training. arxiv:2403.09611, 2024. 3, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 491, + 294, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 491, + 294, + 545 + ], + "spans": [ + { + "bbox": [ + 56, + 491, + 294, + 545 + ], + "type": "text", + "content": "[37] Jae Sung Park, Jack Hessel, Khyathi Chandu, Paul Pu Liang, Ximing Lu, Peter West, Youngjae Yu, Qiuyuan Huang, Jianfeng Gao, Ali Farhadi, et al. Localized symbolic knowledge distillation for visual commonsense models. In NeurIPS, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 546, + 294, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 546, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 546, + 294, + 601 + ], + "type": "text", + "content": "[38] Reiner Pope, Sholto Douglas, Aakanksha Chowdhery, Jacob Devlin, James Bradbury, Jonathan Heek, Kefan Xiao, Shivani Agrawal, and Jeff Dean. Efficiently scaling transformer inference. Proceedings of Machine Learning and Systems, 5: 606-624, 2023. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 602, + 294, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 294, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 294, + 668 + ], + "type": "text", + "content": "[39] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS, pages 25278–25294, 2022. 5, 6, 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 669, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 294, + 714 + ], + "type": "text", + "content": "[40] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In CVPR, pages 8317-8326, 2019. 5, 1" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 315, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 315, + 73, + 553, + 127 + ], + "type": "text", + "content": "[41] Quan Sun, Yufeng Cui, Xiaosong Zhang, Fan Zhang, Qiying Yu, Yueze Wang, Yongming Rao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative multimodal models are in-context learners. In CVPR, pages 14398-14409, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "type": "text", + "content": "[42] Chenxin Tao, Xizhou Zhu, Shiqian Su, Lewei Lu, Changyao Tian, Xuan Luo, Gao Huang, Hongsheng Li, Yu Qiao, Jie Zhou, et al. Learning 1d causal visual representation with de-focus attention networks. arxiv:2406.04342, 2024. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 175, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 196 + ], + "type": "text", + "content": "[43] Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models. arxiv:2405.09818, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 198, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 251 + ], + "type": "text", + "content": "[44] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arxiv:2312.11805, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 254, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 553, + 308 + ], + "type": "text", + "content": "[45] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arxiv:2302.13971, 2023. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 309, + 553, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 364 + ], + "type": "text", + "content": "[46] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arxiv:2409.12191, 2024. 1, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 365, + 553, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 409 + ], + "type": "text", + "content": "[47] Tiannan Wang, Wangchunshu Zhou, Yan Zeng, and Xinsong Zhang. Efficientvlm: Fast and accurate vision-language models via knowledge distillation and modal-adaptive pruning. arxiv:2210.07795, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 410, + 553, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 410, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 316, + 410, + 553, + 454 + ], + "type": "text", + "content": "[48] Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong Lu, Jie Zhou, Yu Qiao, et al. Vision: Large language model is also an open-ended decoder for vision-centric tasks. In NeurIPS, 2024. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 456, + 553, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 499 + ], + "type": "text", + "content": "[49] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arxiv:2409.18869, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 500, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 500, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 500, + 553, + 521 + ], + "type": "text", + "content": "[50] X.ai. Grok-1.5 vision preview. https://x.ai/blog/grok-1.5v, 2024.3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 523, + 553, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 553, + 577 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 553, + 577 + ], + "type": "text", + "content": "[51] Ruyi Xu, Yuan Yao, Zonghao Guo, Junbo Cui, Zanlin Ni, Chunjiang Ge, Tat-Seng Chua, Zhiyuan Liu, Maosong Sun, and Gao Huang. Llava-uhd: an Imm perceiving any aspect ratio and high-resolution images. arXiv:2403.11703, 2024. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 579, + 553, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 579, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 316, + 579, + 553, + 623 + ], + "type": "text", + "content": "[52] Zhengyuan Yang, Linjie Li, Kevin Lin, Jianfeng Wang, Chung-Ching Lin, Zicheng Liu, and Lijuan Wang. The dawn of Imms: Preliminary explorations with gpt-4v (ision). arxiv:2309.17421, 9(1):1, 2023. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 624, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 668 + ], + "type": "text", + "content": "[53] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arxiv:2408.01800, 2024.6, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 669, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 553, + 714 + ], + "type": "text", + "content": "[54] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. Mm-vet: Evaluating large multimodal models for integrated capabilities. arxiv:2308.02490, 2023. 5" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14568" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 272 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "text", + "content": "[55] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In CVPR, pages 9556-9567, 2024. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 172 + ], + "type": "text", + "content": "[56] Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. A survey of large language models. arxiv:2303.18223, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "type": "text", + "content": "[57] Chunting Zhou, Lili Yu, Arun Babu, Kushal Tirumala, Michihiro Yasunaga, Leonid Shamis, Jacob Kahn, Xuezhe Ma, Luke Zettlemoyer, and Omer Levy. Transfusion: Predict the next token and diffuse images with one multi-modal model. arxiv:2408.11039, 2024. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 294, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 294, + 272 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 294, + 272 + ], + "type": "text", + "content": "[58] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arxiv:2304.10592, 2023. 1" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "14569" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Holmes-VAU_ Towards Long-term Video Anomaly Understanding at Any Granularity/8b3a9750-2575-4fe7-bb78-a5c96dc5aac4_content_list.json b/2025/Holmes-VAU_ Towards Long-term Video Anomaly Understanding at Any Granularity/8b3a9750-2575-4fe7-bb78-a5c96dc5aac4_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..20f21a702381f25dfdccc068236aed8a782deae2 --- /dev/null +++ b/2025/Holmes-VAU_ Towards Long-term Video Anomaly Understanding at Any Granularity/8b3a9750-2575-4fe7-bb78-a5c96dc5aac4_content_list.json @@ -0,0 +1,1817 @@ +[ + { + "type": "text", + "text": "Holmes-VAU: Towards Long-term Video Anomaly Understanding at Any Granularity", + "text_level": 1, + "bbox": [ + 166, + 130, + 831, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Huaxin Zhang $^{1}$ Xiaohao Xu $^{2}$ Xiang Wang $^{1}$ Jialong Zuo $^{1}$ Xiaonan Huang $^{2}$ Changxin Gao $^{1}$ Shanjun Zhang $^{3}$ Li Yu $^{1}$ Nong Sang $^{1*}$ $^{1}$ Key Laboratory of Image Processing and Intelligent Control, School of Artificial Intelligence and Automation, Huazhong University of Science and Technology \n $^{2}$ University of Michigan, Ann Arbor \n $^{3}$ Kanagawa University", + "bbox": [ + 127, + 202, + 864, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zhanghuaxin,wxiang,cgao,hustlyu,nsang}@hust.edu.cn,{xiaohaox,xiaonanh}@umich.edu,{chiyoz01}@kanagawa-u.ac.jp", + "bbox": [ + 117, + 296, + 879, + 309 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/91680e9686dac8a1c7c39831e48df448d8514ddd8efc7f8e41225746cb194df8.jpg", + "image_caption": [ + "Figure 1. Motivation. Left: Existing datasets lack the hierarchical structure to capture transient and sustained anomalies across varying temporal scales. Our HIVAU-70k dataset addresses this by providing multi-granularity annotations—clip, event, and video levels—that enable detailed anomaly analysis in complex real-world scenarios. Right: Inspired by Sherlock Holmes's knack for zeroing in on critical details, our Holmes-VAU method integrates an Anomaly-focused Temporal Sampler with a multi-modal LLM, directing model attention to anomaly-rich segments, which enables models to decode complex, long-term video anomalies efficiently." + ], + "image_footnote": [], + "bbox": [ + 96, + 316, + 658, + 556 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/20f4aec5a3fd4a30388dd8ab4138d26c5614df3c98528ba32bf241494842a176.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 681, + 319, + 901, + 556 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 643, + 326, + 659 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "How can we enable models to comprehend video anomalies occurring over varying temporal scales and contexts? Traditional Video Anomaly Understanding (VAU) methods focus on frame-level anomaly prediction, often missing the interpretability of complex and diverse real-world anomalies. Recent multimodal approaches leverage visual and textual data but lack hierarchical annotations that capture both short-term and long-term anomalies. To address this challenge, we introduce HIVAU-70k, a large-scale benchmark for hierarchical video anomaly understanding across any granularity. We develop a semi-automated annotation engine that efficiently scales high-quality annotations by combining manual video segmentation with recursive free-text", + "bbox": [ + 86, + 676, + 485, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "annotation using large language models (LLMs). This results in over 70,000 multi-granular annotations organized at clip-level, event-level, and video-level segments. For efficient anomaly detection in long videos, we propose the Anomaly-focused Temporal Sampler (ATS). ATS integrates an anomaly scorer with a density-aware sampler to adaptively select frames based on anomaly scores, ensuring that the multimodal LLM concentrates on anomaly-rich regions, which significantly enhances both efficiency and accuracy. Extensive experiments demonstrate that our hierarchical instruction data markedly improves anomaly comprehension. The integrated ATS and visual-language model outperform traditional methods in processing long videos. Our benchmark and model are publicly available at https://github.com/pipixin321/HolmesVAU.", + "bbox": [ + 511, + 643, + 908, + 872 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 109, + 887, + 235, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13843", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 89, + 222, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video Anomaly Understanding (VAU) is crucial for applications such as video surveillance [46], violent content analysis [56], and autonomous driving [63]. Detecting deviations from normal patterns aids in hazard prevention and real-time decision-making. Traditional methods [14, 49, 74] mainly focus on frame-level predefined closed-set anomaly prediction, assigning an anomaly score to each frame. However, these approaches often fail to describe and understand complex anomalies in the real world.", + "bbox": [ + 89, + 114, + 480, + 251 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this gap, open-world anomaly understanding [57] embraces the diversity and unpredictability of real-world anomalies. Recent work integrates multimodal approaches, combining visual data with textual descriptions [42, 58, 62, 65], while advances in multimodal visual-language models (VLMs) [6, 21, 27, 29, 68] have enabled more nuanced understanding through anomaly-related instruction tuning and text generation [9, 36, 47, 67].", + "bbox": [ + 89, + 251, + 482, + 371 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite these advancements, a significant gap remains in models' ability to comprehend anomalies across multiple temporal scales. For instance, while anomalies such as explosions or fights may be captured in a single frame, more complex events like theft or arson require understanding long-term contextual patterns. Existing VAU datasets [46, 56] typically provide annotations at a single level of granularity, limiting models to understanding either immediate perceptual anomalies or those requiring extended contextual reasoning. The lack of datasets with hierarchical annotations—encompassing both short-term and long-term anomalies—hinders models' capacity to reason about anomalies with diverse temporal characteristics. Moreover, constructing datasets that encapsulate this hierarchical complexity poses significant challenges in scalability and annotation quality.", + "bbox": [ + 89, + 372, + 482, + 613 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these issues, we develop a semi-automated annotation engine that efficiently scales high-quality annotation by combining manual video segmentation with recursive free-text annotation using large language models (LLMs). The process involves three key stages: 1) hierarchical video decoupling, where we manually identify anomaly events and segment them into shorter clips; 2) hierarchical free-text annotation, where captions for each clip are generated through human effort or video captioning models, then summarized at the event and video levels via LLMs; and 3) hierarchical instruction construction, where the textual data is transformed into question-answer instruction prompts by combining captions and summaries with designed prompts, creating a dataset with rich annotations for training and evaluating models.", + "bbox": [ + 89, + 614, + 482, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Utilizing the annotation engine, we introduce HIVAU-70k, a large-scale video anomaly understanding benchmark with hierarchical instructions. Our dataset comprises over 70,000 multi-granular instruction data organized", + "bbox": [ + 89, + 839, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "across clip-level, event-level, and video-level segments as shown in Fig. 1. This hierarchical structure empowers models to detect immediate anomalies, e.g., sudden explosions or fighting, as well as complex events that require an understanding of long-term context, like theft or arson. By annotating at multiple temporal levels, HIVAU-70k provides diverse anomalies in open-world scenarios.", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Towards long-term VAU, efficiency remains a critical challenge. Previous methods [9, 47, 67] often rely on uniform frame sampling, which can either miss crucial anomaly frames or incur large computational costs [25, 47, 67]. To address this, we propose the Holmes-VAU method, which combines the proposed Anomaly-focused Temporal Sampler (ATS) with the multimodal visual-language model for efficient long-term video anomaly understanding (See Fig. 1). The ATS combines a anomaly scorer with a density-aware sampler that adaptively selects frames by their anomaly scores. This integration ensures that the visual-language model concentrates on anomaly-rich regions, enhancing both efficiency and accuracy.", + "bbox": [ + 511, + 198, + 903, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are threefold: 1) We introduce HIVAU-70k, a large-scale, multi-granular benchmark for hierarchical video anomaly understanding. 2) We propose the Holmes-VAU method, which combines the proposed Anomaly-focused Temporal Sampler (ATS) to boost the efficiency and accuracy of inference on long-term videos. 3) We conduct extensive experiments demonstrating the effectiveness of hierarchical instruction data in enhancing anomaly comprehension and validate the performance gains provided by the integrated ATS and visual-language model in processing long videos.", + "bbox": [ + 511, + 396, + 903, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 511, + 580, + 663, + 597 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video Anomaly Detection. This task aims to temporally detect abnormal frames in a long untrimmed video [1, 14, 23, 34, 40, 52]. Early attempts are based on handcrafted features [1, 19, 23, 34, 40, 72]. Recently, deep learning approaches [14, 37, 61] have become dominant, broadly classified into unsupervised, weakly-supervised, and fully-supervised methods. Unsupervised methods train only on normal videos to learn normal patterns and are often designed as reconstruction-based [13, 14, 59, 61], prediction-based [31], or a combination [33]. Some methods [48, 50, 66] also explore a fully unsupervised setting, including both normal and abnormal videos in training set without real labels. Weakly-supervised methods [11, 22, 37, 46, 49, 55, 56, 70, 73, 74] use both normal and abnormal videos with video-level annotations. Fully-supervised methods [20, 30] are less studied due to the high cost of precise frame-level annotations.", + "bbox": [ + 511, + 608, + 903, + 864 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-modal Video Anomaly Understanding. Large-scale visual-language pre-trained models such as CLIP [44]", + "bbox": [ + 511, + 869, + 903, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13844", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "serve as a bridge between visual and textual modalities. Some recent works [17, 42, 58, 62] in the realm of video anomaly detection have leveraged textual information as prompts to enhance the model's anomaly representation. Based on this, [57] firstly proposed the open vocabulary VAD task, [65] introduced a multimodal video anomaly dataset composed of dense clip captions. Furthermore, [67] extracted captions from video frames and designed prompts for LLMs to provide anomaly scores, [9] and [47] construct diverse and interactive instruction data at the video level. However, these datasets consider only a single temporal level of anomaly understanding data construction, i.e. clip-level [65] or video-level [9, 47]. Unlike these methods, we focus on building large-scale hierarchical video anomaly understanding data for multimodal instruction tuning.", + "bbox": [ + 89, + 90, + 483, + 332 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Hierarchical Video Understanding. Video understanding is a challenging task due to its temporal-scale diversity. To better comprehend videos, many previous works have focused on both datasets and models in hierarchical video understanding. For example, [3, 24, 32, 35, 45, 54] proposed fine-grained action recognition and localization datasets, [16] provided free-form hierarchical captions for hour-long videos at multiple temporal scales, and [8, 18, 38, 43, 60, 69] trained models on hierarchical levels to obtain better video feature representation. Recently, to assess the capability of video vision-language models (VLMs) in handling challenges in real-world scenarios, several video benchmarks [10, 12] have been built, incorporating data at multiple temporal scales for evaluation. Unlike these works, we dive into the field of Video Anomaly Understanding, thus filling the gap in multi-scale annotations in this area.", + "bbox": [ + 89, + 337, + 483, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. HIVAU-70k Benchmark", + "text_level": 1, + "bbox": [ + 89, + 595, + 316, + 611 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first define the video anomaly understanding task in Sec.3.1. Then, the construction process of the HIVAU-70k benchmark will be elaborated in Sec.3.2. Finally, we will present HIVAU-70k's statistical information in Sec. 3.3.", + "bbox": [ + 89, + 622, + 483, + 683 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Task Description", + "text_level": 1, + "bbox": [ + 89, + 695, + 256, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This work focuses on video anomaly understanding, involving temporal anomaly detection and anomaly explainability. Temporal anomaly detection aims to predict an anomaly score for each frame in a video $\\mathcal{V}$ , represented as $S\\in \\mathbb{R}^T$ , where $T$ is the total number of frames. Building on this, we explore the model's ability to generate explanatory text outputs related to anomalies based on video input and user queries. Specifically, given a video $\\mathcal{V}$ and a text query $\\mathcal{Q}$ , the model generates an anomaly-related response $\\mathcal{A}$ . We consider two key abilities: 1) visual perception, which involves recognizing main entities in the video, and 2) anomaly reasoning, which encompasses the", + "bbox": [ + 89, + 719, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/716735a97f013fd3347cdca1ef45bb25a82e5313a822f0e352fb671715079360.jpg", + "image_caption": [ + "Figure 2. Data Engine. We present a structured workflow for generating hierarchical annotations across video, event, and clip levels. Clips are first captioned, then processed through a large language model (LLM) with prompts for event summarization. The outputs include clip captions, event summaries, and video summaries, followed by manual checking and refinement. This multi-step approach enriches the dataset with detailed judgments, descriptions, and analyses of anomalies, enabling robust contextual understanding at varying granularities." + ], + "image_footnote": [], + "bbox": [ + 519, + 93, + 903, + 324 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "model's judgment and analysis of the anomaly content.", + "bbox": [ + 511, + 474, + 877, + 489 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. LLM-Empowered Data Engine", + "text_level": 1, + "bbox": [ + 511, + 500, + 790, + 515 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Fig.2, we develop a semi-automated annotation engine that efficiently scales high-quality annotations, which consists of three main steps: 1) hierarchical video decoupling, 2) hierarchical free-text annotation, and 3) hierarchical instruction construction.", + "bbox": [ + 511, + 522, + 905, + 597 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Hierarchical Video Decoupling. Our video sources include the training set of the UCF-Crime [46] dataset and the XD-Violence [56] dataset, which contains videos of varying durations and diverse real-world anomalies. For abnormal videos, we first manually obtain the temporal boundaries of each anomaly event in the video. Non-continuous anomalous states are considered separate events. Then, we divide each event into clips of random lengths. For normal videos, we apply random sampling to obtain corresponding segments of varying granularities. Ultimately, we obtained 5,443 videos, 11,076 events, and 55,806 clips. This process took 5 annotators approximately 20 hours to complete. More details can be found in the appendix.", + "bbox": [ + 511, + 598, + 905, + 794 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Hierarchical Free-text Annotation. To fully extract semantic information from the clip-level videos, we utilize a powerful off-the-shell video perception model LLaVA-Next-Video [28] to generate detailed captions for each clip. We also include the UCA dataset [64], which provides manually annotated captions for video clips in the UCF-Crime [46] dataset. Then, we use an LLM [2] to consoli", + "bbox": [ + 511, + 795, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13845", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b5288291184c2c45051df10e1fc4281661cd7533f284234b3570a9825728988e.jpg", + "image_caption": [ + "Q: Give a short and clear caption of the video clip." + ], + "image_footnote": [], + "bbox": [ + 96, + 92, + 361, + 200 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/afe828ef6ae55dc7a7bca169325cd44c0a08c5ac8c2dabb99be098c56c997894.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 375, + 90, + 630, + 200 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fb2b43b8e06015e5e6f484b9c008b3e8e3a78acc326bb6e6148738050b275d0d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 90, + 900, + 200 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5fa90769cc6e549f8ba56fbcbed5274d7753f89b8efe0b2300fe0d9a028f65b5.jpg", + "image_caption": [ + "A: A group of policemen dragged a man red out of the gate.", + "Q: Are there any anomaly events detected in the video?", + "A: An anomaly exists, specifically an Arrest anomaly.", + "Q: Describe the anomaly events in the video. A: The anomaly event involves the three policemen turning the man in red over and moving him to the wall, indicating a possible arrest or detention.", + "Q: Why is this event as an anomaly? A: The unusual and potentially forceful interaction between the policemen and the man in red, deviating from normal pedestrian behavior in the scene", + "Figure 3. HVAU-70k dataset. (a) Duration distributions for clips, events, and full videos, showing dominance of short clips. (b) Hierarchical data organization from clip-level to video-level, enabling perception-to-reasoning insights. (c) Word count variations across annotation levels, with more detailed descriptions at the video level. (d) Sample annotations capturing captioning, judgment, description, and anomaly analysis, highlighting nuanced understanding of anomaly events in complex scenes." + ], + "image_footnote": [], + "bbox": [ + 372, + 207, + 629, + 318 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "date all clip captions within an event, generating an event-level video summary. Specifically, we design prompts1 to guide the LLM to produce three parts of content for each anomalous event summary: 1) Judgment: A determination of whether an anomaly exists and its specific category, 2) Description: A detailed description of the anomalous or normal event, 3) Analysis: The reasoning behind the anomaly judgment, including causal analysis. To guide the LLM in generating reliable responses, we also inject the event's category label (e.g., \"Shooting\", \"Explosion\") into the LLM prompt. We consolidate all event-level summaries to obtain the video-level summary. This results in free-text annotations across multiple temporal scales, including short-term visual perception (clip-level) and long-term anomaly reasoning (event-level, video-level).", + "bbox": [ + 88, + 400, + 483, + 626 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Hierarchical Instruction Data Construction. The ability of VLMs to follow user instructions and generate responses is achieved through instruction-tuning [21, 25, 27, 29]. The training data format typically consists of the following:", + "bbox": [ + 89, + 627, + 483, + 688 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\{\\mathcal{Q}$ :user instruction, $\\mathcal{A}$ :model response.}", + "bbox": [ + 109, + 689, + 393, + 705 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To build instruction-tuning data for VLMs in the domain of Video Anomaly Understanding, we matched free-text annotations with pre-designed anomaly-related user instructions. Specifically, for clip-level segments, we only construct instructions related to captions, as it is challenging to obtain anomaly-related analysis for short videos. For event-level and video-level segments, we construct instruction data from the perspectives of Judgment, Description, and Analysis. Typical examples are shown in the Fig. 3 (d). Manual Checking for Data Quality Control. To ensure dataset quality, we implemented several human inspec", + "bbox": [ + 89, + 705, + 483, + 873 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tion and curation strategies. First, we labeled the temporal boundaries of abnormal event segments and reviewed them at the second level. Next, anomaly labels were incorporated during summary generation using LLMs, directing focus on relevant entities. Finally, manual reviews were performed to correct low-quality instruction data.", + "bbox": [ + 511, + 400, + 906, + 491 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Data Statistic of HIVAU-70k", + "text_level": 1, + "bbox": [ + 511, + 502, + 767, + 517 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Utilizing the proposed annotation engine, we introduce HIVAU-70k, as shown in Fig. 3, a large-scale benchmark designed for hierarchical instruction-based video anomaly understanding. As shown in Fig. 3(b), HIVAU-70k contains over 70,000 multi-granular annotations organized at clip-level, event-level, and video-level segments, achieving a progression from perception to reasoning. As shown in Fig. 3(a) and (c), the durations of segments and the word numbers of text annotations at different granularities exhibit significant distributional differences. As shown in Fig. 3(d), HIVAU-70k's instruction data covers Caption, Judgment, Description, and Analysis for real-world anomalies, which guide the model to develop both short-term and long-term video anomaly understanding capabilities.", + "bbox": [ + 511, + 526, + 906, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Method: Holmes-VAU", + "text_level": 1, + "bbox": [ + 511, + 753, + 723, + 768 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Long-term video anomaly understanding with LLMs/VLMs has traditionally been hindered by frame redundancy, complicating accurate anomaly detection. Previous VAU approaches struggle with focus: methods like dense window sampling [67] add redundancy, and uniform frame sampling [9, 47] often misses key anomalies, limiting application to short videos. We introduce the Anomaly-focused Temporal Sampler (ATS) to address this,", + "bbox": [ + 511, + 779, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "For detailed prompts, please refer to the appendix.", + "bbox": [ + 107, + 886, + 380, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13846", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f67777cea1f190d73ce83a9f2a9cff77e319b69aae4387130340c2fb515e0dca.jpg", + "image_caption": [ + "Figure 4. Holmes-VAU: a multi-modal-LLM-based video anomaly detection framework with adaptive anomaly focus." + ], + "image_footnote": [], + "bbox": [ + 98, + 88, + 480, + 366 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "integrating it into the VLM, and fine-tuning it via instruction on HIVAU-70k to form our Holmes-VAU model.", + "bbox": [ + 89, + 416, + 483, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Pipeline Overview", + "text_level": 1, + "bbox": [ + 89, + 455, + 266, + 472 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The overall pipeline of our Holmes-VAU model is shown in Fig. 4. Video frames are processed by a visual encoder, creating visual tokens. These tokens are analyzed by an Anomaly-focused Temporal Sampler using an anomaly scorer and cumulative sum ( $S_{cumsum}$ ) to select keyframes. A text encoder processes a prompt (e.g., 'Describe the abnormal events in the video'). Visual and textual representations are combined in a pre-trained language model, fine-tuned with LoRA, to generate a description of detected anomalies, such as a 'collision between a car and a motorcycle' and 'traffic accident' indicators.", + "bbox": [ + 89, + 479, + 483, + 647 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Model Architecture", + "text_level": 1, + "bbox": [ + 89, + 655, + 277, + 670 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Visual and Text embedding. We utilize the frozen visual encoder in InternVL2 [6], which inherits the ViT structure from CLIP [44], we refer to it as $\\phi_v$ . Following previous VAU works [55, 58, 67, 74], we sample dense video frames at an interval of 16 frames from the input video. Each video frame is then processed through the visual encoder to obtain the corresponding visual tokens. Given the input video frame sequence $\\mathcal{V} \\in \\mathbb{R}^{T \\times H \\times W \\times C}$ , the output features of $i$ -th frame can be denoted as:", + "bbox": [ + 89, + 676, + 483, + 813 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nV _ {i} = \\left\\{v _ {i} ^ {c l s}, v _ {i} ^ {1}, v _ {i} ^ {2}, \\dots , v _ {i} ^ {N _ {p}} \\right\\} = \\phi_ {v} \\left(\\mathcal {V} _ {i}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 160, + 821, + 482, + 843 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $v_{i}^{cls}$ indicates the class token, $v_{i}^{j}$ ( $j \\in \\{1,2,\\dots,N_{p}\\}$ ) denotes the patch tokens, and $N_{p}$ represents the number of patches. The text encoder $\\phi_t$ is also initialized from [6],", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "which includes a tokenizer and an embedding layer. The prompt text $\\mathcal{Q}$ is converted into text tokens through the text encoder: $X_{q} = \\phi_{t}(\\mathcal{Q})$", + "bbox": [ + 511, + 90, + 903, + 136 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Anomaly-focused Temporal Sampler (ATS). ATS consists of two components: the anomaly scorer and the density-aware sampler. The anomaly scorer $\\phi_s$ is a feature-based VAD network which estimates the anomaly score for each frame. We follow the network architecture in [74] due to its simplicity and good performance. Given the class token of the video frames $\\{v_1^{cls}, v_2^{cls}, \\dots, v_T^{cls}\\}$ , the anomaly scores can be obtained: $s_i = \\phi_s(v_i^{cls})$ , where $s_i$ denotes the anomaly score of the $i$ -th frame.", + "bbox": [ + 511, + 136, + 905, + 271 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Anomalous frames typically contain more information and exhibit greater variation than normal frames [49]. This observation motivates us to sample more frames in regions with higher anomaly scores while reducing the sampling in areas with lower anomaly scores. As shown in Fig. 4, to achieve non-uniform sampling, we propose density-aware sampler to selectively choose $N$ frames from a total of $T$ input frames. Specifically, we treat the anomaly scores $S \\in \\mathbb{R}^T$ as a probability mass function and first accumulate them along the temporal dimension to obtain the cumulative distribution function, denoted as $S_{cumsum}$ :", + "bbox": [ + 511, + 272, + 906, + 439 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS _ {c u m s u m} (t) = \\sum_ {i = 1} ^ {t} \\left(s _ {i} + \\tau\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 449, + 903, + 491 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We uniformly sample $N$ points along the cumulative axis, then map these points to the cumulative distribution $S_{cumsum}$ , the corresponding $N$ timestamps on the time axis are mapped to the closest frame index and finally form the sampled frame indices, denoted as $\\mathcal{G}$ . $\\tau$ is used to control the uniformity of the sampling.", + "bbox": [ + 511, + 500, + 905, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Projector and LLM. We select the tokens corresponding to the sampled frame, i.e., $\\mathcal{G}$ , as the visual embedding. A projector $\\phi_p$ is then used to map the visual embedding to the language feature space. Finally, we concatenate these embeddings with the text embeddings, input them into the pre-trained large language model, and compute the probability of the target answers $X_a$ . To obtain an initial visual-language alignment space, we initialize the projector and LLM parameters from [6], with the parameters kept frozen. The above process can be expressed as follows:", + "bbox": [ + 511, + 590, + 905, + 742 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nX _ {i n s} = \\operatorname {c a t} \\left[ \\phi_ {p} \\left(\\operatorname {c a t} \\left[ V _ {i} \\right]\\right), X _ {q} \\right] \\quad i \\in \\mathcal {G} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 753, + 903, + 771 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np \\left(X _ {a} \\mid X _ {i n s}\\right) = \\prod_ {i = 1} ^ {L} p _ {\\theta} \\left(\\left(x _ {i} \\mid X _ {i n s, < i}, X _ {a, < i}\\right)\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 792, + 903, + 834 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\operatorname{cat}[\\cdot]$ represents the concatenation operation, $\\theta$ is the trainable parameters, $X_{ins, < i}, X_{a, < i}$ are the instruction and answer tokens in all turns before the current prediction token $x_i$ , $L$ is the length of sequence, respectively.", + "bbox": [ + 511, + 839, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13847", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Training and Testing", + "text_level": 1, + "bbox": [ + 89, + 90, + 287, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training. We train the model following two steps. In the first step, we use the video data and annotated frame-level label $(\\hat{y} \\in \\mathbb{R}^T)$ from HIVAU-70k to train the anomaly scorer, which provides more accurate anomaly supervision compared to previous unsupervised and weakly-supervised methods [14, 46, 55, 74].", + "bbox": [ + 89, + 114, + 483, + 205 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {A S} = - \\sum_ {i = 1} ^ {T} \\left(s _ {i} \\log \\left(\\hat {y} _ {i}\\right) + \\left(1 - s _ {i}\\right) \\log \\left(1 - \\hat {y} _ {i}\\right)\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 217, + 482, + 258 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the second step, we keep the anomaly scorer fixed, and use all the instruction data from HIVAU-70k to train the model. To achieve more efficient fine-tuning without disrupting the original capabilities of the LLM, we employ LoRA [15] for fine-tuning, optimizing the cross entropy loss between the predicted and the ground truth tokens.", + "bbox": [ + 89, + 271, + 483, + 362 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Testing. During testing, the user inputs a video and text prompts; the model will generate the corresponding text response following the user's instruction.", + "bbox": [ + 89, + 363, + 483, + 409 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 425, + 223, + 443 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experiment Setup", + "text_level": 1, + "bbox": [ + 89, + 450, + 267, + 468 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset. Our HIVAU-70k is built upon two large-scale real-world datasets, i.e., UCF-Crime [46] and XD-Violence [56], they provide a diverse range of videos with anomalous events. UCF-Crime [46] comprises 1,900 untrimmed videos totaling 128 hours from outdoor and indoor surveillance cameras. It encompasses 13 classes of real-world anomalies, including Abuse, Explosion, Fighting, and Shooting. XD-Violence [56] is the largest VAD benchmark, comprising 4,754 videos totaling 217 hours sourced from surveillance, movies, car cameras, and games. It encompasses 6 anomaly classes: Abuse, Car Accidents, Explosions, Fighting, Riots, and Shooting.", + "bbox": [ + 89, + 474, + 483, + 656 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metric. We assess the anomaly understanding ability from two aspects: anomaly detection and reasoning. 1) For anomaly detection, we use the anomaly scores output by the Anomaly Scorer as the prediction and perform the evaluation. Following [14, 46, 67, 74], we use AUC and AP to quantify detection performance, which is evaluated only on the video level. 2) For anomaly reasoning, we annotate instruction data from the UCF-Crime and XD-Violence test sets, which have been carefully reviewed and filtered by annotators. We finally collected 3,300 test samples at multiple granularities. The test set contains 2200/732/398 samples at clip/event/video levels. We calculate metrics including BLEU [41], CIDEr [51], METEOR [4] and ROUGE [26] to measure the quality of the reasoning text output by the model, comparing with the annotated ground truth text.", + "bbox": [ + 89, + 657, + 482, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. For the proposed Holmes-VAU", + "bbox": [ + 89, + 885, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1fffd0e8301545be9613af8604b5afe3d1b9cb40631198f730b7c29d199aa20f.jpg", + "table_caption": [ + "Table 1. Comparison of detection performance with state-of-the-art Video Anomaly Detection approaches. We include the results of explainable and non-explainable methods. \\*\\* represents the result reported in [67]." + ], + "table_footnote": [], + "table_body": "
MethodsBackboneXD-ViolenceUCF-Crime
AP/%AUC/%
Non-explainable VAD
Conv-AE [14] (CVPR'16)-27.2550.60
GODS [52] (ICCV'19)I3DN/A70.46
GCL [66] (CVPR'22)ResNextN/A71.04
DYANNET [48] (WACV'23)I3DN/A84.50
MIST [11] (CVPR'21)I3DN/A82.30
Wu et al. [56] (ECCV'20)I3D78.6482.44
RTFM [49] (ICCV'21)I3D77.8184.30
MSL [22] (AAAI'22)I3D78.2885.30
S3R [55] (ECCV'22)I3D80.2685.99
MGFN [5] (AAAI'23)I3D79.1986.98
UR-DMU [74] (AAAI'23)I3D81.6686.97
CLIP-TSA [17] (ICIP'23)ViT82.1987.58
VadCLIP [58] (AAAI'24)ViT84.5188.02
Yang et al. [62] (CVPR'24)ViT83.6887.79
Wu et al. [57] (CVPR'24)ViT66.5386.40
Explainable Multi-modal VAD
Zero-Shot CLIP [44]*ViT17.8353.16
LLAVA-1.5 [27]*ViT50.2672.84
LAVAD [67] (CVPR'24)ViT62.0180.28
Holmes-VAU (Ours)ViT87.6888.96
", + "bbox": [ + 519, + 147, + 898, + 425 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "method, we initialize the Multimodal LLM with InternVL2-2B [6]. To optimize the Anomaly-focused Temporal Sampler, we adopt the Adam optimizer with a learning rate of 1e-4. Note that when evaluating detection performance on XD-Violence and UCF-Crime, only videos in the corresponding training sets are used to train our model for fair comparisons. For instruction tuning, we train with a batch size of 512 for 1 epoch, using the AdamW optimizer with cosine learning rate decay and a warm-up period. The LoRA [15] parameters are set as: $r = 64$ , $\\alpha = 128$ , and learning rate=4e-5. During testing, $\\tau$ in Eq. 2 is set to 0.1. Experiments are conducted on 2 NVIDIA A100 GPUs.", + "bbox": [ + 511, + 443, + 906, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Main Results", + "text_level": 1, + "bbox": [ + 511, + 635, + 653, + 648 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Anomaly Detection Results. We compare our method with state-of-the-art methods, including semi-supervised methods [14, 52], unsupervised methods [48, 66], weakly-supervised methods [17, 22, 49, 55, 58, 74] and recently training-free method [67]. We have indicated their backbones and performance on the UCF-Crime and XD-Violence datasets, as shown in Table 1. Our method has an AP of $87.68\\%$ on XD-Violence and an AUC of $88.96\\%$ on UCF-Crime, significantly outperforming the prior state-of-the-art methods, which demonstrates that our method can generate less biased anomaly scores. It is worth noting that while achieving precise localization of anomalies, Holmes-VAU is also capable of providing explanations and analysis for the detected anomalies by the model, a feature unavailable in existing non-explainable VAD methods. Although LAVAD [67] has explainability, the training-free large lan", + "bbox": [ + 511, + 657, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13848", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/cc583d64780c65bda3f6d9249aede0002b59ddedde52a33b8d1737fdec3aa11b.jpg", + "table_caption": [ + "Table 2. Comparison of reasoning performance with state-of-the-art Multimodal Large Language Models (MLLMs). 'BLEU' refers to the cumulative values from BLEU-1 to BLEU-4. We evaluate the quality of the generated text at different granularities, including clip-level (C), event-level (E), and video-level (V)." + ], + "table_footnote": [], + "table_body": "
MethodParamsBLEU [41](↑)CIDEr [51](↑)METEOR [4](↑)ROUGE [26](↑)
CEVCEVCEVCEV
Video-ChatGPT [39]7B0.1520.0680.0660.0330.0110.0130.1020.0690.0440.1530.0480.079
Video-LLaMA [68]7B0.1510.0790.1040.0240.0140.0170.1120.0760.0570.1560.0670.090
Video-LLaVA [25]7B0.1640.0460.0550.0320.0090.0130.0970.0220.0140.1320.0230.045
LLaVA-Next-Video [71]7B0.4350.0910.1200.1020.0150.0310.1170.0850.0960.1980.0800.106
QwenVL2 [53]7B0.3120.0820.1550.0440.0200.0440.1330.0920.1120.1630.0810.137
InternVL2 [6]8B0.3310.1010.1450.0520.0220.0350.1410.0950.1010.1820.1020.122
Holmes-VAU (Ours)2B0.9130.8040.5660.4671.5191.4370.1900.1650.1210.3290.3700.355
", + "bbox": [ + 94, + 133, + 903, + 256 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/87cc6987fe0258a1eed8290e62f96fa28f0a33cc46ca981b51f3aac92b4539b4.jpg", + "table_caption": [ + "Table 3. Ablation of hierarchical instruction data. During the instruction tuning phase, we combined training data of different granularities, including clip (C), event (E), and video (V) levels." + ], + "table_footnote": [], + "table_body": "
Training DataBLEU(↑)CIDEr(↑)
CEVCEVCEV
✓0.9840.2610.3510.4590.1200.106
✓0.5080.5760.2920.0971.1830.872
✓0.2800.2220.2790.0390.7080.884
✓✓0.8890.7410.3490.4701.2850.889
✓✓0.9060.3410.5220.4720.9621.093
✓✓0.3940.7970.5050.0811.4721.074
✓✓✓0.9130.8040.5660.4671.5191.437
", + "bbox": [ + 94, + 309, + 498, + 434 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "guage model lacks an understanding of anomaly knowledge due to the limitation of insufficient supervised data.", + "bbox": [ + 89, + 454, + 482, + 483 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Anomaly Reasoning Results. We compare the anomaly-related text quality generated by Holmes-VAU with that produced by state-of-the-art general Multimodal Large Language Models (MLLMs), and presented the results at different temporal granularities, including clip-level, event-level, and video-level, as shown in Table 2. Earlier MLLMs such as Video-ChatGPT [21] and Video-LLaMA [68], struggled with basic visual perception and instruction-following capabilities. Recent MLLMs [6, 7, 71] trained on larger and higher-quality video instruction data have made significant progress in general video understanding, with noticeable improvements at the clip-level perception task. However, due to the absence of learning from complex, real-world anomaly data, their reasoning abilities at the event-level and video-level are still lacking. Our Holmes-VAU, however, shows significant improvements in video understanding across all temporal granularities compared to existing general MLLMs, highlighting the importance of injecting anomaly-related knowledge through instruction tuning on high-quality Video Anomaly Understanding benchmarks.", + "bbox": [ + 89, + 484, + 482, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Analytic Results", + "text_level": 1, + "bbox": [ + 89, + 800, + 253, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Influence of Hierarchical Instruction. To explore the impact of different granularity video training data on the model's anomaly reasoning ability, we designed various training data combinations during the instruction tuning phase and evaluated the model's performance, as shown", + "bbox": [ + 89, + 824, + 482, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8cdf84ff07e3edfb89100c945b5ef5e72bbd712f920c4b2b182b3be9a07c28c6.jpg", + "table_caption": [ + "Table 4. Ablation study of sampling methods and the number of sampled frames. We compare the proposed Anomaly-focused Temporal Sampler (ATS) with other sampling methods under different frame sampling numbers, including Uniform and Top-K. Latency is the time delay in generating the first token." + ], + "table_footnote": [], + "table_body": "
Frames (N)SamplerLatency (ms)Video-level
BLEU(↑)CIDEr(↑)
8Top-K2440.4621.229
Uniform0.4911.276
ATS (Ours)0.5141.324
16Top-K5660.4761.302
Uniform0.5111.345
ATS (Ours)0.5661.437
32Top-K14020.4811.332
Uniform0.5581.357
ATS (Ours)0.5761.460
", + "bbox": [ + 517, + 338, + 903, + 477 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "in Table 3. The inclusion of clip-level data primarily enhanced the model's basic visual perception abilities regarding actions and scenes within the video. Adding event-level data improved the model's ability to judge and understand complete anomaly events. Furthermore, the involvement of video-level data further enhanced the model's ability to analyze and summarize anomaly-related information across longer-span videos. The hierarchical instruction data structure facilitated a comprehensive and complementary improvement in the model's anomaly-related perception-to-reasoning capabilities.", + "bbox": [ + 511, + 503, + 906, + 670 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Influence of different sampling methods and the number of sampled frames. Our ATS (Anomaly-focused Temporal Sampler) is designed to adaptively sample frames input to the LLM based on the frame-level anomaly scores. To validate its advantages, we compared ATS with other sampling methods at various sample frame counts, including Uniform and Top-K sampling. In Uniform sampling, N frames are uniformly sampled from all frames, while Top-K sampling selects the frames with the top $N$ highest anomaly scores. As shown in Table 4, ATS consistently outperforms other sampling methods, regardless of the sample count. We believe that Uniform sampling tends to overlook key anomaly frames, though this issue lessens as more frames are sampled. Besides, Top-K sampling tends to overly focus on local anomaly frames, missing contextual frame informa", + "bbox": [ + 511, + 674, + 908, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13849", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a3539e036593faaac8034fc67e9939e4aa357ae675c99186e6486495ab7caf89.jpg", + "image_caption": [ + "··" + ], + "image_footnote": [], + "bbox": [ + 102, + 95, + 225, + 132 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/81c4fc8f13a8690ef1e239037ca10d2daaa0281908bb60f59303889bbb67ae61.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 254, + 97, + 483, + 133 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Prompt: Determine whether there are anomaly events in the video and the specific name of the anomaly.", + "bbox": [ + 101, + 136, + 277, + 162 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ground Truth", + "text_level": 1, + "bbox": [ + 102, + 172, + 163, + 181 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The anomaly exists and is specifically identified as a Burglary event.", + "bbox": [ + 102, + 185, + 290, + 202 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Holmes-VAU (Ours)", + "text_level": 1, + "bbox": [ + 102, + 220, + 186, + 229 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Yes, there is an anomaly in the video, specifically a break-in burglary.", + "bbox": [ + 102, + 232, + 287, + 258 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "InternVL2-BB [Incorrect Judgement] Yes, there are anomaly events in the video. The specific anomaly is that the man in the video is seen throwing a punch at the other man.", + "bbox": [ + 305, + 145, + 477, + 189 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "QwenVL2-7B [Incorrect Judgement] There are no anomaly events in the video.", + "bbox": [ + 305, + 205, + 475, + 234 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2c898df8c75fe0d6e6131670fcaea798172033a90f1443db56cc54474af81ebd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 102, + 268, + 225, + 306 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/00195e2b01988744be8bd79d766165fedfe371e83651df04f37d07ef36992356.jpg", + "image_caption": [ + "Bulgary037_x264" + ], + "image_footnote": [], + "bbox": [ + 220, + 271, + 303, + 306 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cf38f80e93e5239daed104c387edede3cf469af887927484dc5aa55c1baaa202.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 303, + 270, + 480, + 306 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Prompt: Identify any abnormal events depicted in the video.", + "bbox": [ + 102, + 311, + 282, + 329 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "InternVL2-8B [Incorrect Judgement] The video shows a white truck parked on the side of the road, which is an abnormal behavior as it is not supposed to be parked on the road.", + "bbox": [ + 303, + 318, + 477, + 369 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ground Truth", + "text_level": 1, + "bbox": [ + 102, + 347, + 163, + 354 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "No anomaly exists in the video", + "bbox": [ + 102, + 359, + 254, + 369 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Holmes-VAU (Ours)", + "text_level": 1, + "bbox": [ + 102, + 393, + 187, + 402 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "There is no anomaly in the video.", + "bbox": [ + 102, + 409, + 264, + 417 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "QwenVL2-7B", + "text_level": 1, + "bbox": [ + 305, + 378, + 362, + 387 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The video does not show any abnormal events.", + "bbox": [ + 305, + 391, + 475, + 407 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Normal_Videos_902_x264", + "bbox": [ + 330, + 420, + 457, + 430 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c4a55d808cdb7b4123c4a65bfcc83b439904c3a5536425786e201a75908c248d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 95, + 890, + 133 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Prompt: How would you describe the particular anomaly events in the video? What evidence do you have to support your judgment of this event as an anomaly?", + "bbox": [ + 504, + 137, + 803, + 162 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ground Truth", + "text_level": 1, + "bbox": [ + 504, + 172, + 566, + 180 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The video anomaly event depicts chaotic street scenes where people are running, police officers are clad in riot gear, and confrontations or protests are underway, with scenes of apprehension, restraint, and property damage unfolding from start to end. The presence of violent behavior, law enforcement intervention, and disruption of public order in the urban environment depicted in the video, which collectively indicate a significant deviation from normal, peaceful civic activity.", + "bbox": [ + 504, + 180, + 885, + 238 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Holmes-VAU (Ours)", + "text_level": 1, + "bbox": [ + 504, + 244, + 589, + 252 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The anomaly events involve a series of chaotic and violent scenes, including police officers in riot gear, people running and pushing against barriers, and a helicopter flying over the scene, suggesting a high level of tension and unrest. The presence of destructive behavior, including property damage, property destruction, and physical altercations, which deviates from normal and peaceful urban activity.", + "bbox": [ + 504, + 253, + 885, + 304 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "InternVL2-8B [Visual Hallucination, Incomplete Analysis]", + "text_level": 1, + "bbox": [ + 504, + 311, + 741, + 321 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The video shows a group of people walking down a street, and a man is seen walking down the street with a skateboard. The man is wearing a black jacket and a black hat. The event depicted in the video is classified as an anomaly due to the presence of a large group of people engaging in violent behavior.", + "bbox": [ + 504, + 321, + 885, + 354 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "QwenVL2-7B [Incomplete Description, Incomplete Analysis]", + "text_level": 1, + "bbox": [ + 504, + 363, + 751, + 372 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The video shows a group of people walking down the street, and suddenly, a car crashes into a crowd of people. The car is on fire, and the people are running away. The video does not provide any context or explanation for this event, but it is clear that it is an anomaly. The car crash is a violent and unexpected event that is not related to the rest of the video.", + "bbox": [ + 504, + 372, + 883, + 412 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "v=3WmMhircZ0c__#1_label_B4-0-0", + "bbox": [ + 606, + 421, + 777, + 431 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5f553fb092e3feaf3b37f7f59db441378fa98a7f126f589dc496b44745634815.jpg", + "image_caption": [ + "Figure 5. Qualitative comparison of anomaly understanding explanation. Compared with state-of-the-art general MLLMs, i.e., InternVL2 [6] and QwenVL2 [53], our proposed Holmes-VAU demonstrates more accurate anomaly judgment, along with more detailed and comprehensive anomaly-related descriptions and analysis. Correct and wrong explanations are highlighted in green and red, respectively.", + "Figure 6. Ablation study of trainable parameters. (a) Loss curve during instruction-tuning. (b) We tuned the LoRA [15] parameter $r$ to control trainable parameters, evaluating its impact on VAU capability, and General performance on Video-MME [12]." + ], + "image_footnote": [], + "bbox": [ + 94, + 500, + 254, + 618 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2e799acaa4580cdb49022072ce0ded081fc1bc24a2b2833a33c72c5c42a4c316.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 254, + 501, + 480, + 612 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tion. Our proposed ATS mitigates both issues. To balance inference efficiency and performance, we set $N = 16$ as the default sample frame number.", + "bbox": [ + 89, + 700, + 482, + 744 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Instruction Tuning Parameters. We conducted an ablation study on the parameter $r$ in LoRA [15] to explore how the trainable parameters affects both the model's VAU performance and its general capability. We use VideoMME [12] to evaluate the model's general capability. The results are shown in Fig. 6, as $r$ increases, the model gradually adapts to the VAU task. However, when $r$ becomes too large, the model's general capability decreases. To retain the original general video understanding capability of the MLLM, we set $r = 64$ as the default value.", + "bbox": [ + 89, + 750, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Qualitative Comparison", + "text_level": 1, + "bbox": [ + 513, + 497, + 738, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We provide qualitative comparisons between Holmes-VAU and existing MLLMs in Fig. 5. The results demonstrate that Holmes-VAU can accurately identify anomalies in videos and provide accurate and complete explanations, highlight the effectiveness and advantage of Holmes-VAU in perceiving video events and analyzing anomalies.", + "bbox": [ + 511, + 523, + 906, + 614 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 632, + 632, + 647 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In conclusion, this work pushes the boundaries of video anomaly understanding by introducing hierarchical anomaly detection across diverse temporal scales, from momentary clips to extended events. The HIVAU-70k benchmark, with over 70,000 multi-level annotations, addresses a critical gap in the field, enabling comprehensive anomaly analysis in real-world scenarios. Our Anomaly-focused Temporal Sampler (ATS) strategically enhances focus on anomaly-dense segments, optimizing both efficiency and accuracy in long-term anomaly detection. Extensive experiments demonstrate that our hierarchical dataset and ATS-enhanced VLM achieve significant performance gains over conventional methods, proving robust for open-world anomaly understanding. This work sets a new standard for multi-granular anomaly comprehension, paving the way for more fine-grained video anomaly understanding.", + "bbox": [ + 511, + 659, + 906, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13850", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 164, + 187, + 179 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Amit Adam, Ehud Rivlin, Ilan Shimshoni, and Daviv Reinitz. Robust real-time unusual event detection using multiple fixed-location monitors. IEEE transactions on pattern analysis and machine intelligence, 30(3):555-560, 2008. 2", + "[2] AI@Meta. Llama 3 model card. 2024. 3", + "[3] Kumar Ashutosh, Rohit Girdhar, Lorenzo Torresani, and Kristen Grauman. Hiervl: Learning hierarchical videolanguage embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23066-23078, 2023. 3", + "[4] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65-72, 2005. 6, 7", + "[5] Yingxian Chen, Zhengzhe Liu, Baoheng Zhang, Wilton Fok, Xiaojuan Qi, and Yik-Chung Wu. Mgfn: Magnitude-contrastive glance-and-focus network for weakly-supervised video anomaly detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 387-395, 2023. 6", + "[6] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 2, 5, 6, 7, 8", + "[7] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024. 7", + "[8] Jisheng Dang, Huicheng Zheng, Xiaohao Xu, Longguang Wang, Qingyong Hu, and Yulan Guo. Adaptive sparse memory networks for efficient and robust video object segmentation. IEEE Transactions on Neural Networks and Learning Systems, 2024. 3", + "[9] Hang Du, Sicheng Zhang, Binzhu Xie, Guoshun Nan, Jiayang Zhang, Junrui Xu, Hangyu Liu, Sicong Leng, Jiangming Liu, Hehe Fan, et al. Uncovering what why and how: A comprehensive benchmark for causation understanding of video anomaly. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18793-18803, 2024. 2, 3, 4", + "[10] Xinyu Fang, Kangrui Mao, Haodong Duan, Xiangyu Zhao, Yining Li, Dahua Lin, and Kai Chen. Mmbench-video: A long-form multi-shot benchmark for holistic video understanding. arXiv preprint arXiv:2406.14515, 2024. 3", + "[11] Jia-Chang Feng, Fa-Ting Hong, and Wei-Shi Zheng. Mist: Multiple instance self-training framework for video anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14009-14018, 2021. 2, 6" + ], + "bbox": [ + 93, + 188, + 482, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Acknowledgement This work is supported by the National Natural Science Foundation of China under grants U22B2053 and 623B2039, and in part by the Interdisciplinary Research Program of HUST (2024JCYJ034).", + "[12] Chaoyou Fu, Yuhan Dai, Yondong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 3, 8", + "[13] Dong Gong, Lingqiao Liu, Vuong Le, Budhaditya Saha, Moussa Reda Mansour, Svetha Venkatesh, and Anton van den Hengel. Memorizing normality to detect anomaly: Memory-augmented deep autoencoder for unsupervised anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1705-1714, 2019. 2", + "[14] Mahmudul Hasan, Jonghyun Choi, Jan Neumann, Amit K Roy-Chowdhury, and Larry S Davis. Learning temporal regularity in video sequences. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 733-742, 2016. 2, 6", + "[15] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 6, 8", + "[16] Md Mohaiminul Islam, Ngan Ho, Xitong Yang, Tushar Nagarajan, Lorenzo Torresani, and Gedas Bertasius. Video recap: Recursive captioning of hour-long videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18198-18208, 2024. 3", + "[17] Hyekang Kevin Joo, Khoa Vo, Kashu Yamazaki, and Ngan Le. Clip-tsa: Clip-assisted temporal self-attention for weakly-supervised video anomaly detection. In 2023 IEEE International Conference on Image Processing (ICIP), pages 3230–3234. IEEE, 2023. 3, 6", + "[18] Kumara Kahatapitiya and Michael S Ryoo. Coarse-fine networks for temporal activity detection in videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8385-8394, 2021. 3", + "[19] Jaechul Kim and Kristen Grauman. Observe locally, infer globally: a space-time mrf for detecting abnormal activities with incremental updates. In 2009 IEEE conference on computer vision and pattern recognition, pages 2921-2928. IEEE, 2009. 2", + "[20] Federico Landi, Cees GM Snoek, and Rita Cucchiara. Anomaly locality in video surveillance. arXiv preprint arXiv:1901.10364, 2019. 2", + "[21] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 2, 4, 7", + "[22] Shuo Li, Fang Liu, and Licheng Jiao. Self-training multisession learning with transformer for weakly supervised video anomaly detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1395-1403, 2022. 2, 6", + "[23] Weixin Li, Vijay Mahadevan, and Nuno Vasconcelos. Anomaly detection and localization in crowded scenes. IEEE transactions on pattern analysis and machine intelligence, 36(1):18-32, 2013. 2", + "[24] Yingwei Li, Yi Li, and Nuno Vasconcelos. Resound: Towards action recognition without representation bias. In Pro" + ], + "bbox": [ + 93, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13851", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ceedings of the European Conference on Computer Vision (ECCV), pages 513-528, 2018. 3", + "[25] Bin Lin, Bin Zhu, Yang Ye, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 2, 4, 7", + "[26] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 6, 7", + "[27] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 2, 4, 6", + "[28] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024.3", + "[29] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2, 4", + "[30] Kun Liu and Huadong Ma. Exploring background-bias for anomaly detection in surveillance videos. In Proceedings of the 27th ACM International Conference on Multimedia, pages 1490-1499, 2019. 2", + "[31] Wen Liu, Weixin Luo, Dongze Lian, and Shenghua Gao. Future frame prediction for anomaly detection-a new baseline. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6536-6545, 2018. 2", + "[32] Yi Liu, Limin Wang, Yali Wang, Xiao Ma, and Yu Qiao. Fineaction: A fine-grained video dataset for temporal action localization. IEEE transactions on image processing, 31: 6937-6950, 2022. 3", + "[33] Zhian Liu, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. A hybrid video anomaly detection framework via memory-augmented flow reconstruction and flow-guided frame prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 13588-13597, 2021. 2", + "[34] Cewu Lu, Jianping Shi, and Jiaya Jia. Abnormal event detection at 150 fps in matlab. In Proceedings of the IEEE international conference on computer vision, pages 2720-2727, 2013. 2", + "[35] Jian Lu, Xuanfeng Li, Bo Zhao, and Jian Zhou. A review of skeleton-based human action recognition. Journal of Image and Graphics, 28(12):3651-3669, 2023. 3", + "[36] Hui Lv and Qianru Sun. Video anomaly detection and explanation via large language models. arXiv preprint arXiv:2401.05702, 2024. 2", + "[37] Hui Lv, Zhongqi Yue, Qianru Sun, Bin Luo, Zhen Cui, and Hanwang Zhang. Unbiased multiple instance learning for weakly supervised video anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8022-8031, 2023. 2", + "[38] Baiteng Ma, Shiwei Zhang, Changxin Gao, and Nong Sang. Temporal global correlation network for end-to-end action proposal generation. Acta Electronica Sinica, 50(10):2452-2461, 2022. 3", + "[39] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "understanding via large vision and language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (ACL 2024), 2024. 7", + "[40] Ramin Mehran, Alexis Oyama, and Mubarak Shah. Abnormal crowd behavior detection using social force model. In 2009 IEEE conference on computer vision and pattern recognition, pages 935-942. IEEE, 2009. 2", + "[41] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311–318, 2002. 6, 7", + "[42] Yujiang Pu, Xiaoyu Wu, Lulu Yang, and Shengjin Wang. Learning prompt-enhanced context features for weakly-supervised video anomaly detection. IEEE Transactions on Image Processing, 2024. 2, 3", + "[43] Zhiwu Qing, Shiwei Zhang, Ziyuan Huang, Yi Xu, Xiang Wang, Mingqian Tang, Changxin Gao, Rong Jin, and Nong Sang. Learning from untrimmed videos: Self-supervised video representation learning with hierarchical consistency. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13821-13831, 2022. 3", + "[44] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 5, 6", + "[45] Dian Shao, Yue Zhao, Bo Dai, and Dahua Lin. Finegym: A hierarchical video dataset for fine-grained action understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2616-2625, 2020. 3", + "[46] Waqas Sultani, Chen Chen, and Mubarak Shah. Real-world anomaly detection in surveillance videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6479-6488, 2018. 2, 3, 6", + "[47] Jiaqi Tang, Hao Lu, Ruizheng Wu, Xiaogang Xu, Ke Ma, Cheng Fang, Bin Guo, Jiangbo Lu, Qifeng Chen, and YingCong Chen. Hawk: Learning to understand open-world video anomalies. arXiv preprint arXiv:2405.16886, 2024. 2,3,4", + "[48] Kamalakar Vijay Thakare, Yash Raghuwanshi, Debi Prosad Dogra, Heeseung Choi, and Ig-Jae Kim. Dyannet: A scene dynamicity guided self-trained video anomaly detection network. In Proceedings of the IEEE/CVF Winter conference on applications of computer vision, pages 5541-5550, 2023. 2, 6", + "[49] Yu Tian, Guansong Pang, Yuanhong Chen, Rajvinder Singh, Johan W Verjans, and Gustavo Carneiro. Weakly-supervised video anomaly detection with robust temporal feature magnitude learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4975-4986, 2021. 2, 5, 6", + "[50] Anil Osman Tur, Nicola Dall'Asen, Cigdem Bayan, and Elisa Ricci. Exploring diffusion models for unsupervised" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13852", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "video anomaly detection. In 2023 IEEE International Conference on Image Processing (ICIP), pages 2540-2544. IEEE, 2023. 2", + "[51] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4566-4575, 2015. 6, 7", + "[52] Jue Wang and Anoop Cherian. Gods: Generalized one-class discriminative subspaces for anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8201-8211, 2019. 2, 6", + "[53] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 7, 8", + "[54] SC Wang, Q Huang, YF Zhang, X Li, YQ Nie, and GC Luo. Review of action recognition based on multimodal data. Image Graph, 27(11):3139-3159, 2022. 3", + "[55] Jhih-Ciang Wu, He-Yen Hsieh, Ding-Jie Chen, Chiou-Shann Fuh, and Tyng-Luh Liu. Self-supervised sparse representation for video anomaly detection. In European Conference on Computer Vision, pages 729-745. Springer, 2022. 2, 5, 6", + "[56] Peng Wu, Jing Liu, Yujia Shi, Yujia Sun, Fangtao Shao, Zhaoyang Wu, and Zhiwei Yang. Not only look, but also listen: Learning multimodal violence detection under weak supervision. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXX 16, pages 322-339. Springer, 2020. 2, 3, 6", + "[57] Peng Wu, Xuerong Zhou, Guansong Pang, Yujia Sun, Jing Liu, Peng Wang, and Yanning Zhang. Open-vocabulary video anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18297-18307, 2024. 2, 3, 6", + "[58] Peng Wu, Xuerong Zhou, Guansong Pang, Lingru Zhou, Qingsen Yan, Peng Wang, and Yanning Zhang. Vadclip: Adapting vision-language models for weakly supervised video anomaly detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6074-6082, 2024. 2, 3, 5, 6", + "[59] Dan Xu, Yan Yan, Elisa Ricci, and Nicu Sebe. Detecting anomalous events in videos by learning deep representations of appearance and motion. Computer Vision and Image Understanding, 156:117-127, 2017. 2", + "[60] Xiaohao Xu, Jinglu Wang, Xiang Ming, and Yan Lu. Towards robust video object segmentation with adaptive object calibration. In Proceedings of the 30th ACM International Conference on Multimedia, pages 2709-2718, 2022. 3", + "[61] Zhiwei Yang, Jing Liu, Zhaoyang Wu, Peng Wu, and Xiaotao Liu. Video event restoration based on keyframes for video anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14592-14601, 2023. 2", + "[62] Zhiwei Yang, Jing Liu, and Peng Wu. Text prompt with normality guidance for weakly supervised video anomaly detec" + ], + "bbox": [ + 91, + 90, + 483, + 902 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18899-18908, 2024. 2, 3, 6", + "[63] Yu Yao, Xizi Wang, Mingze Xu, Zelin Pu, Yuchen Wang, Ella Atkins, and David J Crandall. Dota: Unsupervised detection of traffic anomaly in driving videos. IEEE transactions on pattern analysis and machine intelligence, 45(1): 444-459, 2022. 2", + "[64] Tongtong Yuan, Xuange Zhang, Kun Liu, Bo Liu, Chen Chen, Jian Jin, and Zhenzhen Jiao. Towards surveillance video-and-language understanding: New dataset, baselines, and challenges, 2023. 3", + "[65] Tongtong Yuan, Xuange Zhang, Kun Liu, Bo Liu, Chen Chen, Jian Jin, and Zhenzhen Jiao. Towards surveillance video-and-language understanding: New dataset baselines and challenges. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22052-22061, 2024. 2, 3", + "[66] M Zaigham Zaheer, Arif Mahmood, M Haris Khan, Mattia Segu, Fisher Yu, and Seung-Ik Lee. Generative cooperative learning for unsupervised video anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14744-14754, 2022. 2, 6", + "[67] Luca Zanella, Willi Menapace, Massimiliano Mancini, Yiming Wang, and Elisa Ricci. Harnessing large language models for training-free video anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18527-18536, 2024. 2, 3, 4, 5, 6", + "[68] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 2, 7", + "[69] Huaxin Zhang, Xiang Wang, Xiaohao Xu, Zhiwu Qing, Changxin Gao, and Nong Sang. Hr-pro: Point-supervised temporal action localization via hierarchical reliability propagation. arXiv preprint arXiv:2308.12608, 2023. 3", + "[70] Huaxin Zhang, Xiang Wang, Xiaohao Xu, Xiaonan Huang, Chuchu Han, Yuehuan Wang, Changxin Gao, Shanjun Zhang, and Nong Sang. Glancevad: Exploring glance supervision for label-efficient video anomaly detection. arXiv preprint arXiv:2403.06154, 2024. 2", + "[71] Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, and Chunyuan Li. Video instruction tuning with synthetic data. arXiv preprint arXiv:2410.02713, 2024. 7", + "[72] Bin Zhao, Li Fei-Fei, and Eric P Xing. Online detection of unusual events in videos via dynamic sparse coding. In CVPR 2011, pages 3313-3320. IEEE, 2011. 2", + "[73] Jia-Xing Zhong, Nannan Li, Weijie Kong, Shan Liu, Thomas H Li, and Ge Li. Graph convolutional label noise cleaner: Train a plug-and-play action classifier for anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1237-1246, 2019. 2", + "[74] Hang Zhou, Junqing Yu, and Wei Yang. Dual memory units with uncertainty regulation for weakly supervised video anomaly detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3769-3777, 2023. 2, 5, 6" + ], + "bbox": [ + 516, + 90, + 903, + 883 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "13853", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Holmes-VAU_ Towards Long-term Video Anomaly Understanding at Any Granularity/8b3a9750-2575-4fe7-bb78-a5c96dc5aac4_model.json b/2025/Holmes-VAU_ Towards Long-term Video Anomaly Understanding at Any Granularity/8b3a9750-2575-4fe7-bb78-a5c96dc5aac4_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ec222e9b700684effdbb34f75e3d235b32c0ba5e --- /dev/null +++ b/2025/Holmes-VAU_ Towards Long-term Video Anomaly Understanding at Any Granularity/8b3a9750-2575-4fe7-bb78-a5c96dc5aac4_model.json @@ -0,0 +1,2708 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.167, + 0.131, + 0.833, + 0.177 + ], + "angle": 0, + "content": "Holmes-VAU: Towards Long-term Video Anomaly Understanding at Any Granularity" + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.203, + 0.866, + 0.293 + ], + "angle": 0, + "content": "Huaxin Zhang\\(^{1}\\) Xiaohao Xu\\(^{2}\\) Xiang Wang\\(^{1}\\) Jialong Zuo\\(^{1}\\) Xiaonan Huang\\(^{2}\\) Changxin Gao\\(^{1}\\) Shanjun Zhang\\(^{3}\\) Li Yu\\(^{1}\\) Nong Sang\\(^{1*}\\) \n\\(^{1}\\)Key Laboratory of Image Processing and Intelligent Control, School of Artificial Intelligence and Automation, Huazhong University of Science and Technology \n\\(^{2}\\)University of Michigan, Ann Arbor \n\\(^{3}\\)Kanagawa University" + }, + { + "type": "text", + "bbox": [ + 0.118, + 0.297, + 0.88, + 0.31 + ], + "angle": 0, + "content": "{zhanghuaxin,wxiang,cgao,hustlyu,nsang}@hust.edu.cn,{xiaohaox,xiaonanh}@umich.edu,{chiyoz01}@kanagawa-u.ac.jp" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.318, + 0.659, + 0.558 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.683, + 0.32, + 0.902, + 0.558 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.562, + 0.907, + 0.633 + ], + "angle": 0, + "content": "Figure 1. Motivation. Left: Existing datasets lack the hierarchical structure to capture transient and sustained anomalies across varying temporal scales. Our HIVAU-70k dataset addresses this by providing multi-granularity annotations—clip, event, and video levels—that enable detailed anomaly analysis in complex real-world scenarios. Right: Inspired by Sherlock Holmes's knack for zeroing in on critical details, our Holmes-VAU method integrates an Anomaly-focused Temporal Sampler with a multi-modal LLM, directing model attention to anomaly-rich segments, which enables models to decode complex, long-term video anomalies efficiently." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.644, + 0.327, + 0.66 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.678, + 0.486, + 0.875 + ], + "angle": 0, + "content": "How can we enable models to comprehend video anomalies occurring over varying temporal scales and contexts? Traditional Video Anomaly Understanding (VAU) methods focus on frame-level anomaly prediction, often missing the interpretability of complex and diverse real-world anomalies. Recent multimodal approaches leverage visual and textual data but lack hierarchical annotations that capture both short-term and long-term anomalies. To address this challenge, we introduce HIVAU-70k, a large-scale benchmark for hierarchical video anomaly understanding across any granularity. We develop a semi-automated annotation engine that efficiently scales high-quality annotations by combining manual video segmentation with recursive free-text" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.909, + 0.873 + ], + "angle": 0, + "content": "annotation using large language models (LLMs). This results in over 70,000 multi-granular annotations organized at clip-level, event-level, and video-level segments. For efficient anomaly detection in long videos, we propose the Anomaly-focused Temporal Sampler (ATS). ATS integrates an anomaly scorer with a density-aware sampler to adaptively select frames based on anomaly scores, ensuring that the multimodal LLM concentrates on anomaly-rich regions, which significantly enhances both efficiency and accuracy. Extensive experiments demonstrate that our hierarchical instruction data markedly improves anomaly comprehension. The integrated ATS and visual-language model outperform traditional methods in processing long videos. Our benchmark and model are publicly available at https://github.com/pipixin321/HolmesVAU." + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.236, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13843" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.223, + 0.107 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.482, + 0.252 + ], + "angle": 0, + "content": "Video Anomaly Understanding (VAU) is crucial for applications such as video surveillance [46], violent content analysis [56], and autonomous driving [63]. Detecting deviations from normal patterns aids in hazard prevention and real-time decision-making. Traditional methods [14, 49, 74] mainly focus on frame-level predefined closed-set anomaly prediction, assigning an anomaly score to each frame. However, these approaches often fail to describe and understand complex anomalies in the real world." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.252, + 0.483, + 0.372 + ], + "angle": 0, + "content": "To address this gap, open-world anomaly understanding [57] embraces the diversity and unpredictability of real-world anomalies. Recent work integrates multimodal approaches, combining visual data with textual descriptions [42, 58, 62, 65], while advances in multimodal visual-language models (VLMs) [6, 21, 27, 29, 68] have enabled more nuanced understanding through anomaly-related instruction tuning and text generation [9, 36, 47, 67]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.373, + 0.483, + 0.614 + ], + "angle": 0, + "content": "Despite these advancements, a significant gap remains in models' ability to comprehend anomalies across multiple temporal scales. For instance, while anomalies such as explosions or fights may be captured in a single frame, more complex events like theft or arson require understanding long-term contextual patterns. Existing VAU datasets [46, 56] typically provide annotations at a single level of granularity, limiting models to understanding either immediate perceptual anomalies or those requiring extended contextual reasoning. The lack of datasets with hierarchical annotations—encompassing both short-term and long-term anomalies—hinders models' capacity to reason about anomalies with diverse temporal characteristics. Moreover, constructing datasets that encapsulate this hierarchical complexity poses significant challenges in scalability and annotation quality." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.615, + 0.483, + 0.84 + ], + "angle": 0, + "content": "To address these issues, we develop a semi-automated annotation engine that efficiently scales high-quality annotation by combining manual video segmentation with recursive free-text annotation using large language models (LLMs). The process involves three key stages: 1) hierarchical video decoupling, where we manually identify anomaly events and segment them into shorter clips; 2) hierarchical free-text annotation, where captions for each clip are generated through human effort or video captioning models, then summarized at the event and video levels via LLMs; and 3) hierarchical instruction construction, where the textual data is transformed into question-answer instruction prompts by combining captions and summaries with designed prompts, creating a dataset with rich annotations for training and evaluating models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Utilizing the annotation engine, we introduce HIVAU-70k, a large-scale video anomaly understanding benchmark with hierarchical instructions. Our dataset comprises over 70,000 multi-granular instruction data organized" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "across clip-level, event-level, and video-level segments as shown in Fig. 1. This hierarchical structure empowers models to detect immediate anomalies, e.g., sudden explosions or fighting, as well as complex events that require an understanding of long-term context, like theft or arson. By annotating at multiple temporal levels, HIVAU-70k provides diverse anomalies in open-world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.199, + 0.905, + 0.395 + ], + "angle": 0, + "content": "Towards long-term VAU, efficiency remains a critical challenge. Previous methods [9, 47, 67] often rely on uniform frame sampling, which can either miss crucial anomaly frames or incur large computational costs [25, 47, 67]. To address this, we propose the Holmes-VAU method, which combines the proposed Anomaly-focused Temporal Sampler (ATS) with the multimodal visual-language model for efficient long-term video anomaly understanding (See Fig. 1). The ATS combines a anomaly scorer with a density-aware sampler that adaptively selects frames by their anomaly scores. This integration ensures that the visual-language model concentrates on anomaly-rich regions, enhancing both efficiency and accuracy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.397, + 0.905, + 0.563 + ], + "angle": 0, + "content": "Our contributions are threefold: 1) We introduce HIVAU-70k, a large-scale, multi-granular benchmark for hierarchical video anomaly understanding. 2) We propose the Holmes-VAU method, which combines the proposed Anomaly-focused Temporal Sampler (ATS) to boost the efficiency and accuracy of inference on long-term videos. 3) We conduct extensive experiments demonstrating the effectiveness of hierarchical instruction data in enhancing anomaly comprehension and validate the performance gains provided by the integrated ATS and visual-language model in processing long videos." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.581, + 0.665, + 0.598 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.609, + 0.905, + 0.866 + ], + "angle": 0, + "content": "Video Anomaly Detection. This task aims to temporally detect abnormal frames in a long untrimmed video [1, 14, 23, 34, 40, 52]. Early attempts are based on handcrafted features [1, 19, 23, 34, 40, 72]. Recently, deep learning approaches [14, 37, 61] have become dominant, broadly classified into unsupervised, weakly-supervised, and fully-supervised methods. Unsupervised methods train only on normal videos to learn normal patterns and are often designed as reconstruction-based [13, 14, 59, 61], prediction-based [31], or a combination [33]. Some methods [48, 50, 66] also explore a fully unsupervised setting, including both normal and abnormal videos in training set without real labels. Weakly-supervised methods [11, 22, 37, 46, 49, 55, 56, 70, 73, 74] use both normal and abnormal videos with video-level annotations. Fully-supervised methods [20, 30] are less studied due to the high cost of precise frame-level annotations." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Multi-modal Video Anomaly Understanding. Large-scale visual-language pre-trained models such as CLIP [44]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13844" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.333 + ], + "angle": 0, + "content": "serve as a bridge between visual and textual modalities. Some recent works [17, 42, 58, 62] in the realm of video anomaly detection have leveraged textual information as prompts to enhance the model's anomaly representation. Based on this, [57] firstly proposed the open vocabulary VAD task, [65] introduced a multimodal video anomaly dataset composed of dense clip captions. Furthermore, [67] extracted captions from video frames and designed prompts for LLMs to provide anomaly scores, [9] and [47] construct diverse and interactive instruction data at the video level. However, these datasets consider only a single temporal level of anomaly understanding data construction, i.e. clip-level [65] or video-level [9, 47]. Unlike these methods, we focus on building large-scale hierarchical video anomaly understanding data for multimodal instruction tuning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.338, + 0.484, + 0.581 + ], + "angle": 0, + "content": "Hierarchical Video Understanding. Video understanding is a challenging task due to its temporal-scale diversity. To better comprehend videos, many previous works have focused on both datasets and models in hierarchical video understanding. For example, [3, 24, 32, 35, 45, 54] proposed fine-grained action recognition and localization datasets, [16] provided free-form hierarchical captions for hour-long videos at multiple temporal scales, and [8, 18, 38, 43, 60, 69] trained models on hierarchical levels to obtain better video feature representation. Recently, to assess the capability of video vision-language models (VLMs) in handling challenges in real-world scenarios, several video benchmarks [10, 12] have been built, incorporating data at multiple temporal scales for evaluation. Unlike these works, we dive into the field of Video Anomaly Understanding, thus filling the gap in multi-scale annotations in this area." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.596, + 0.318, + 0.612 + ], + "angle": 0, + "content": "3. HIVAU-70k Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.623, + 0.484, + 0.684 + ], + "angle": 0, + "content": "We first define the video anomaly understanding task in Sec.3.1. Then, the construction process of the HIVAU-70k benchmark will be elaborated in Sec.3.2. Finally, we will present HIVAU-70k's statistical information in Sec. 3.3." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.696, + 0.258, + 0.712 + ], + "angle": 0, + "content": "3.1. Task Description" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.901 + ], + "angle": 0, + "content": "This work focuses on video anomaly understanding, involving temporal anomaly detection and anomaly explainability. Temporal anomaly detection aims to predict an anomaly score for each frame in a video \\(\\mathcal{V}\\), represented as \\(S\\in \\mathbb{R}^T\\), where \\(T\\) is the total number of frames. Building on this, we explore the model's ability to generate explanatory text outputs related to anomalies based on video input and user queries. Specifically, given a video \\(\\mathcal{V}\\) and a text query \\(\\mathcal{Q}\\), the model generates an anomaly-related response \\(\\mathcal{A}\\). We consider two key abilities: 1) visual perception, which involves recognizing main entities in the video, and 2) anomaly reasoning, which encompasses the" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.094, + 0.905, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.33, + 0.908, + 0.456 + ], + "angle": 0, + "content": "Figure 2. Data Engine. We present a structured workflow for generating hierarchical annotations across video, event, and clip levels. Clips are first captioned, then processed through a large language model (LLM) with prompts for event summarization. The outputs include clip captions, event summaries, and video summaries, followed by manual checking and refinement. This multi-step approach enriches the dataset with detailed judgments, descriptions, and analyses of anomalies, enabling robust contextual understanding at varying granularities." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.475, + 0.878, + 0.49 + ], + "angle": 0, + "content": "model's judgment and analysis of the anomaly content." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.5, + 0.792, + 0.516 + ], + "angle": 0, + "content": "3.2. LLM-Empowered Data Engine" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.523, + 0.906, + 0.598 + ], + "angle": 0, + "content": "As shown in Fig.2, we develop a semi-automated annotation engine that efficiently scales high-quality annotations, which consists of three main steps: 1) hierarchical video decoupling, 2) hierarchical free-text annotation, and 3) hierarchical instruction construction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.906, + 0.795 + ], + "angle": 0, + "content": "Hierarchical Video Decoupling. Our video sources include the training set of the UCF-Crime [46] dataset and the XD-Violence [56] dataset, which contains videos of varying durations and diverse real-world anomalies. For abnormal videos, we first manually obtain the temporal boundaries of each anomaly event in the video. Non-continuous anomalous states are considered separate events. Then, we divide each event into clips of random lengths. For normal videos, we apply random sampling to obtain corresponding segments of varying granularities. Ultimately, we obtained 5,443 videos, 11,076 events, and 55,806 clips. This process took 5 annotators approximately 20 hours to complete. More details can be found in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Hierarchical Free-text Annotation. To fully extract semantic information from the clip-level videos, we utilize a powerful off-the-shell video perception model LLaVA-Next-Video [28] to generate detailed captions for each clip. We also include the UCA dataset [64], which provides manually annotated captions for video clips in the UCF-Crime [46] dataset. Then, we use an LLM [2] to consoli" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13845" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.093, + 0.362, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.377, + 0.092, + 0.631, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.092, + 0.901, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.106, + 0.207, + 0.345, + 0.226 + ], + "angle": 0, + "content": "Q: Give a short and clear caption of the video clip." + }, + { + "type": "image_caption", + "bbox": [ + 0.106, + 0.233, + 0.34, + 0.25 + ], + "angle": 0, + "content": "A: A group of policemen dragged a man red out of the gate." + }, + { + "type": "image_caption", + "bbox": [ + 0.107, + 0.264, + 0.345, + 0.281 + ], + "angle": 0, + "content": "Q: Are there any anomaly events detected in the video?" + }, + { + "type": "image_caption", + "bbox": [ + 0.107, + 0.286, + 0.326, + 0.304 + ], + "angle": 0, + "content": "A: An anomaly exists, specifically an Arrest anomaly." + }, + { + "type": "image", + "bbox": [ + 0.374, + 0.208, + 0.63, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.636, + 0.209, + 0.895, + 0.255 + ], + "angle": 0, + "content": "Q: Describe the anomaly events in the video. A: The anomaly event involves the three policemen turning the man in red over and moving him to the wall, indicating a possible arrest or detention." + }, + { + "type": "image_caption", + "bbox": [ + 0.637, + 0.263, + 0.895, + 0.311 + ], + "angle": 0, + "content": "Q: Why is this event as an anomaly? A: The unusual and potentially forceful interaction between the policemen and the man in red, deviating from normal pedestrian behavior in the scene" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.325, + 0.907, + 0.382 + ], + "angle": 0, + "content": "Figure 3. HVAU-70k dataset. (a) Duration distributions for clips, events, and full videos, showing dominance of short clips. (b) Hierarchical data organization from clip-level to video-level, enabling perception-to-reasoning insights. (c) Word count variations across annotation levels, with more detailed descriptions at the video level. (d) Sample annotations capturing captioning, judgment, description, and anomaly analysis, highlighting nuanced understanding of anomaly events in complex scenes." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.401, + 0.484, + 0.627 + ], + "angle": 0, + "content": "date all clip captions within an event, generating an event-level video summary. Specifically, we design prompts1 to guide the LLM to produce three parts of content for each anomalous event summary: 1) Judgment: A determination of whether an anomaly exists and its specific category, 2) Description: A detailed description of the anomalous or normal event, 3) Analysis: The reasoning behind the anomaly judgment, including causal analysis. To guide the LLM in generating reliable responses, we also inject the event's category label (e.g., \"Shooting\", \"Explosion\") into the LLM prompt. We consolidate all event-level summaries to obtain the video-level summary. This results in free-text annotations across multiple temporal scales, including short-term visual perception (clip-level) and long-term anomaly reasoning (event-level, video-level)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.628, + 0.484, + 0.689 + ], + "angle": 0, + "content": "Hierarchical Instruction Data Construction. The ability of VLMs to follow user instructions and generate responses is achieved through instruction-tuning [21, 25, 27, 29]. The training data format typically consists of the following:" + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.69, + 0.395, + 0.706 + ], + "angle": 0, + "content": "\\(\\{\\mathcal{Q}\\) :user instruction, \\(\\mathcal{A}\\) :model response.}" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.706, + 0.484, + 0.874 + ], + "angle": 0, + "content": "To build instruction-tuning data for VLMs in the domain of Video Anomaly Understanding, we matched free-text annotations with pre-designed anomaly-related user instructions. Specifically, for clip-level segments, we only construct instructions related to captions, as it is challenging to obtain anomaly-related analysis for short videos. For event-level and video-level segments, we construct instruction data from the perspectives of Judgment, Description, and Analysis. Typical examples are shown in the Fig. 3 (d). Manual Checking for Data Quality Control. To ensure dataset quality, we implemented several human inspec" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.401, + 0.907, + 0.492 + ], + "angle": 0, + "content": "tion and curation strategies. First, we labeled the temporal boundaries of abnormal event segments and reviewed them at the second level. Next, anomaly labels were incorporated during summary generation using LLMs, directing focus on relevant entities. Finally, manual reviews were performed to correct low-quality instruction data." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.503, + 0.768, + 0.518 + ], + "angle": 0, + "content": "3.3. Data Statistic of HIVAU-70k" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.527, + 0.907, + 0.74 + ], + "angle": 0, + "content": "Utilizing the proposed annotation engine, we introduce HIVAU-70k, as shown in Fig. 3, a large-scale benchmark designed for hierarchical instruction-based video anomaly understanding. As shown in Fig. 3(b), HIVAU-70k contains over 70,000 multi-granular annotations organized at clip-level, event-level, and video-level segments, achieving a progression from perception to reasoning. As shown in Fig. 3(a) and (c), the durations of segments and the word numbers of text annotations at different granularities exhibit significant distributional differences. As shown in Fig. 3(d), HIVAU-70k's instruction data covers Caption, Judgment, Description, and Analysis for real-world anomalies, which guide the model to develop both short-term and long-term video anomaly understanding capabilities." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.754, + 0.725, + 0.77 + ], + "angle": 0, + "content": "4. Method: Holmes-VAU" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Long-term video anomaly understanding with LLMs/VLMs has traditionally been hindered by frame redundancy, complicating accurate anomaly detection. Previous VAU approaches struggle with focus: methods like dense window sampling [67] add redundancy, and uniform frame sampling [9, 47] often misses key anomalies, limiting application to short videos. We introduce the Anomaly-focused Temporal Sampler (ATS) to address this," + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.382, + 0.901 + ], + "angle": 0, + "content": "For detailed prompts, please refer to the appendix." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13846" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.089, + 0.482, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.37, + 0.483, + 0.4 + ], + "angle": 0, + "content": "Figure 4. Holmes-VAU: a multi-modal-LLM-based video anomaly detection framework with adaptive anomaly focus." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.417, + 0.484, + 0.448 + ], + "angle": 0, + "content": "integrating it into the VLM, and fine-tuning it via instruction on HIVAU-70k to form our Holmes-VAU model." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.457, + 0.267, + 0.473 + ], + "angle": 0, + "content": "4.1. Pipeline Overview" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.48, + 0.484, + 0.648 + ], + "angle": 0, + "content": "The overall pipeline of our Holmes-VAU model is shown in Fig. 4. Video frames are processed by a visual encoder, creating visual tokens. These tokens are analyzed by an Anomaly-focused Temporal Sampler using an anomaly scorer and cumulative sum (\\( S_{cumsum} \\)) to select keyframes. A text encoder processes a prompt (e.g., 'Describe the abnormal events in the video'). Visual and textual representations are combined in a pre-trained language model, fine-tuned with LoRA, to generate a description of detected anomalies, such as a 'collision between a car and a motorcycle' and 'traffic accident' indicators." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.656, + 0.279, + 0.671 + ], + "angle": 0, + "content": "4.2. Model Architecture" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.678, + 0.484, + 0.814 + ], + "angle": 0, + "content": "Visual and Text embedding. We utilize the frozen visual encoder in InternVL2 [6], which inherits the ViT structure from CLIP [44], we refer to it as \\(\\phi_v\\). Following previous VAU works [55, 58, 67, 74], we sample dense video frames at an interval of 16 frames from the input video. Each video frame is then processed through the visual encoder to obtain the corresponding visual tokens. Given the input video frame sequence \\(\\mathcal{V} \\in \\mathbb{R}^{T \\times H \\times W \\times C}\\), the output features of \\(i\\)-th frame can be denoted as:" + }, + { + "type": "equation", + "bbox": [ + 0.161, + 0.823, + 0.483, + 0.844 + ], + "angle": 0, + "content": "\\[\nV _ {i} = \\left\\{v _ {i} ^ {c l s}, v _ {i} ^ {1}, v _ {i} ^ {2}, \\dots , v _ {i} ^ {N _ {p}} \\right\\} = \\phi_ {v} \\left(\\mathcal {V} _ {i}\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.855, + 0.484, + 0.903 + ], + "angle": 0, + "content": "where \\( v_{i}^{cls} \\) indicates the class token, \\( v_{i}^{j} \\) (\\( j \\in \\{1,2,\\dots,N_{p}\\} \\)) denotes the patch tokens, and \\( N_{p} \\) represents the number of patches. The text encoder \\( \\phi_t \\) is also initialized from [6]," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "which includes a tokenizer and an embedding layer. The prompt text \\(\\mathcal{Q}\\) is converted into text tokens through the text encoder: \\(X_{q} = \\phi_{t}(\\mathcal{Q})\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.137, + 0.906, + 0.272 + ], + "angle": 0, + "content": "Anomaly-focused Temporal Sampler (ATS). ATS consists of two components: the anomaly scorer and the density-aware sampler. The anomaly scorer \\(\\phi_s\\) is a feature-based VAD network which estimates the anomaly score for each frame. We follow the network architecture in [74] due to its simplicity and good performance. Given the class token of the video frames \\(\\{v_1^{cls}, v_2^{cls}, \\dots, v_T^{cls}\\}\\), the anomaly scores can be obtained: \\(s_i = \\phi_s(v_i^{cls})\\), where \\(s_i\\) denotes the anomaly score of the \\(i\\)-th frame." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.273, + 0.907, + 0.44 + ], + "angle": 0, + "content": "Anomalous frames typically contain more information and exhibit greater variation than normal frames [49]. This observation motivates us to sample more frames in regions with higher anomaly scores while reducing the sampling in areas with lower anomaly scores. As shown in Fig. 4, to achieve non-uniform sampling, we propose density-aware sampler to selectively choose \\( N \\) frames from a total of \\( T \\) input frames. Specifically, we treat the anomaly scores \\( S \\in \\mathbb{R}^T \\) as a probability mass function and first accumulate them along the temporal dimension to obtain the cumulative distribution function, denoted as \\( S_{cumsum} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.45, + 0.905, + 0.492 + ], + "angle": 0, + "content": "\\[\nS _ {c u m s u m} (t) = \\sum_ {i = 1} ^ {t} \\left(s _ {i} + \\tau\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.5, + 0.906, + 0.59 + ], + "angle": 0, + "content": "We uniformly sample \\(N\\) points along the cumulative axis, then map these points to the cumulative distribution \\(S_{cumsum}\\), the corresponding \\(N\\) timestamps on the time axis are mapped to the closest frame index and finally form the sampled frame indices, denoted as \\(\\mathcal{G}\\). \\(\\tau\\) is used to control the uniformity of the sampling." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.592, + 0.906, + 0.743 + ], + "angle": 0, + "content": "Projector and LLM. We select the tokens corresponding to the sampled frame, i.e., \\(\\mathcal{G}\\), as the visual embedding. A projector \\(\\phi_p\\) is then used to map the visual embedding to the language feature space. Finally, we concatenate these embeddings with the text embeddings, input them into the pre-trained large language model, and compute the probability of the target answers \\(X_a\\). To obtain an initial visual-language alignment space, we initialize the projector and LLM parameters from [6], with the parameters kept frozen. The above process can be expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.754, + 0.905, + 0.772 + ], + "angle": 0, + "content": "\\[\nX _ {i n s} = \\operatorname {c a t} \\left[ \\phi_ {p} \\left(\\operatorname {c a t} \\left[ V _ {i} \\right]\\right), X _ {q} \\right] \\quad i \\in \\mathcal {G} \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.794, + 0.905, + 0.835 + ], + "angle": 0, + "content": "\\[\np \\left(X _ {a} \\mid X _ {i n s}\\right) = \\prod_ {i = 1} ^ {L} p _ {\\theta} \\left(\\left(x _ {i} \\mid X _ {i n s, < i}, X _ {a, < i}\\right)\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.902 + ], + "angle": 0, + "content": "where \\(\\operatorname{cat}[\\cdot]\\) represents the concatenation operation, \\(\\theta\\) is the trainable parameters, \\(X_{ins, < i}, X_{a, < i}\\) are the instruction and answer tokens in all turns before the current prediction token \\(x_i\\), \\(L\\) is the length of sequence, respectively." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13847" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.289, + 0.108 + ], + "angle": 0, + "content": "4.3. Training and Testing" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.207 + ], + "angle": 0, + "content": "Training. We train the model following two steps. In the first step, we use the video data and annotated frame-level label \\((\\hat{y} \\in \\mathbb{R}^T)\\) from HIVAU-70k to train the anomaly scorer, which provides more accurate anomaly supervision compared to previous unsupervised and weakly-supervised methods [14, 46, 55, 74]." + }, + { + "type": "equation", + "bbox": [ + 0.118, + 0.218, + 0.483, + 0.26 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {A S} = - \\sum_ {i = 1} ^ {T} \\left(s _ {i} \\log \\left(\\hat {y} _ {i}\\right) + \\left(1 - s _ {i}\\right) \\log \\left(1 - \\hat {y} _ {i}\\right)\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.272, + 0.484, + 0.363 + ], + "angle": 0, + "content": "In the second step, we keep the anomaly scorer fixed, and use all the instruction data from HIVAU-70k to train the model. To achieve more efficient fine-tuning without disrupting the original capabilities of the LLM, we employ LoRA [15] for fine-tuning, optimizing the cross entropy loss between the predicted and the ground truth tokens." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.364, + 0.484, + 0.41 + ], + "angle": 0, + "content": "Testing. During testing, the user inputs a video and text prompts; the model will generate the corresponding text response following the user's instruction." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.426, + 0.224, + 0.444 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.452, + 0.268, + 0.469 + ], + "angle": 0, + "content": "5.1. Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.476, + 0.484, + 0.657 + ], + "angle": 0, + "content": "Dataset. Our HIVAU-70k is built upon two large-scale real-world datasets, i.e., UCF-Crime [46] and XD-Violence [56], they provide a diverse range of videos with anomalous events. UCF-Crime [46] comprises 1,900 untrimmed videos totaling 128 hours from outdoor and indoor surveillance cameras. It encompasses 13 classes of real-world anomalies, including Abuse, Explosion, Fighting, and Shooting. XD-Violence [56] is the largest VAD benchmark, comprising 4,754 videos totaling 217 hours sourced from surveillance, movies, car cameras, and games. It encompasses 6 anomaly classes: Abuse, Car Accidents, Explosions, Fighting, Riots, and Shooting." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.483, + 0.884 + ], + "angle": 0, + "content": "Metric. We assess the anomaly understanding ability from two aspects: anomaly detection and reasoning. 1) For anomaly detection, we use the anomaly scores output by the Anomaly Scorer as the prediction and perform the evaluation. Following [14, 46, 67, 74], we use AUC and AP to quantify detection performance, which is evaluated only on the video level. 2) For anomaly reasoning, we annotate instruction data from the UCF-Crime and XD-Violence test sets, which have been carefully reviewed and filtered by annotators. We finally collected 3,300 test samples at multiple granularities. The test set contains 2200/732/398 samples at clip/event/video levels. We calculate metrics including BLEU [41], CIDEr [51], METEOR [4] and ROUGE [26] to measure the quality of the reasoning text output by the model, comparing with the annotated ground truth text." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Implementation Details. For the proposed Holmes-VAU" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.905, + 0.145 + ], + "angle": 0, + "content": "Table 1. Comparison of detection performance with state-of-the-art Video Anomaly Detection approaches. We include the results of explainable and non-explainable methods. \\*\\* represents the result reported in [67]." + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.148, + 0.9, + 0.426 + ], + "angle": 0, + "content": "
MethodsBackboneXD-ViolenceUCF-Crime
AP/%AUC/%
Non-explainable VAD
Conv-AE [14] (CVPR'16)-27.2550.60
GODS [52] (ICCV'19)I3DN/A70.46
GCL [66] (CVPR'22)ResNextN/A71.04
DYANNET [48] (WACV'23)I3DN/A84.50
MIST [11] (CVPR'21)I3DN/A82.30
Wu et al. [56] (ECCV'20)I3D78.6482.44
RTFM [49] (ICCV'21)I3D77.8184.30
MSL [22] (AAAI'22)I3D78.2885.30
S3R [55] (ECCV'22)I3D80.2685.99
MGFN [5] (AAAI'23)I3D79.1986.98
UR-DMU [74] (AAAI'23)I3D81.6686.97
CLIP-TSA [17] (ICIP'23)ViT82.1987.58
VadCLIP [58] (AAAI'24)ViT84.5188.02
Yang et al. [62] (CVPR'24)ViT83.6887.79
Wu et al. [57] (CVPR'24)ViT66.5386.40
Explainable Multi-modal VAD
Zero-Shot CLIP [44]*ViT17.8353.16
LLAVA-1.5 [27]*ViT50.2672.84
LAVAD [67] (CVPR'24)ViT62.0180.28
Holmes-VAU (Ours)ViT87.6888.96
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.444, + 0.907, + 0.626 + ], + "angle": 0, + "content": "method, we initialize the Multimodal LLM with InternVL2-2B [6]. To optimize the Anomaly-focused Temporal Sampler, we adopt the Adam optimizer with a learning rate of 1e-4. Note that when evaluating detection performance on XD-Violence and UCF-Crime, only videos in the corresponding training sets are used to train our model for fair comparisons. For instruction tuning, we train with a batch size of 512 for 1 epoch, using the AdamW optimizer with cosine learning rate decay and a warm-up period. The LoRA [15] parameters are set as: \\( r = 64 \\), \\( \\alpha = 128 \\), and learning rate=4e-5. During testing, \\( \\tau \\) in Eq. 2 is set to 0.1. Experiments are conducted on 2 NVIDIA A100 GPUs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.636, + 0.655, + 0.65 + ], + "angle": 0, + "content": "5.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Anomaly Detection Results. We compare our method with state-of-the-art methods, including semi-supervised methods [14, 52], unsupervised methods [48, 66], weakly-supervised methods [17, 22, 49, 55, 58, 74] and recently training-free method [67]. We have indicated their backbones and performance on the UCF-Crime and XD-Violence datasets, as shown in Table 1. Our method has an AP of \\(87.68\\%\\) on XD-Violence and an AUC of \\(88.96\\%\\) on UCF-Crime, significantly outperforming the prior state-of-the-art methods, which demonstrates that our method can generate less biased anomaly scores. It is worth noting that while achieving precise localization of anomalies, Holmes-VAU is also capable of providing explanations and analysis for the detected anomalies by the model, a feature unavailable in existing non-explainable VAD methods. Although LAVAD [67] has explainability, the training-free large lan" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13848" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.131 + ], + "angle": 0, + "content": "Table 2. Comparison of reasoning performance with state-of-the-art Multimodal Large Language Models (MLLMs). 'BLEU' refers to the cumulative values from BLEU-1 to BLEU-4. We evaluate the quality of the generated text at different granularities, including clip-level (C), event-level (E), and video-level (V)." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.134, + 0.904, + 0.257 + ], + "angle": 0, + "content": "
MethodParamsBLEU [41](↑)CIDEr [51](↑)METEOR [4](↑)ROUGE [26](↑)
CEVCEVCEVCEV
Video-ChatGPT [39]7B0.1520.0680.0660.0330.0110.0130.1020.0690.0440.1530.0480.079
Video-LLaMA [68]7B0.1510.0790.1040.0240.0140.0170.1120.0760.0570.1560.0670.090
Video-LLaVA [25]7B0.1640.0460.0550.0320.0090.0130.0970.0220.0140.1320.0230.045
LLaVA-Next-Video [71]7B0.4350.0910.1200.1020.0150.0310.1170.0850.0960.1980.0800.106
QwenVL2 [53]7B0.3120.0820.1550.0440.0200.0440.1330.0920.1120.1630.0810.137
InternVL2 [6]8B0.3310.1010.1450.0520.0220.0350.1410.0950.1010.1820.1020.122
Holmes-VAU (Ours)2B0.9130.8040.5660.4671.5191.4370.1900.1650.1210.3290.3700.355
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.265, + 0.483, + 0.307 + ], + "angle": 0, + "content": "Table 3. Ablation of hierarchical instruction data. During the instruction tuning phase, we combined training data of different granularities, including clip (C), event (E), and video (V) levels." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.31, + 0.499, + 0.435 + ], + "angle": 0, + "content": "
Training DataBLEU(↑)CIDEr(↑)
CEVCEVCEV
✓0.9840.2610.3510.4590.1200.106
✓0.5080.5760.2920.0971.1830.872
✓0.2800.2220.2790.0390.7080.884
✓✓0.8890.7410.3490.4701.2850.889
✓✓0.9060.3410.5220.4720.9621.093
✓✓0.3940.7970.5050.0811.4721.074
✓✓✓0.9130.8040.5660.4671.5191.437
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.455, + 0.483, + 0.484 + ], + "angle": 0, + "content": "guage model lacks an understanding of anomaly knowledge due to the limitation of insufficient supervised data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.486, + 0.483, + 0.789 + ], + "angle": 0, + "content": "Anomaly Reasoning Results. We compare the anomaly-related text quality generated by Holmes-VAU with that produced by state-of-the-art general Multimodal Large Language Models (MLLMs), and presented the results at different temporal granularities, including clip-level, event-level, and video-level, as shown in Table 2. Earlier MLLMs such as Video-ChatGPT [21] and Video-LLaMA [68], struggled with basic visual perception and instruction-following capabilities. Recent MLLMs [6, 7, 71] trained on larger and higher-quality video instruction data have made significant progress in general video understanding, with noticeable improvements at the clip-level perception task. However, due to the absence of learning from complex, real-world anomaly data, their reasoning abilities at the event-level and video-level are still lacking. Our Holmes-VAU, however, shows significant improvements in video understanding across all temporal granularities compared to existing general MLLMs, highlighting the importance of injecting anomaly-related knowledge through instruction tuning on high-quality Video Anomaly Understanding benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.801, + 0.254, + 0.817 + ], + "angle": 0, + "content": "5.3. Analytic Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Influence of Hierarchical Instruction. To explore the impact of different granularity video training data on the model's anomaly reasoning ability, we designed various training data combinations during the instruction tuning phase and evaluated the model's performance, as shown" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.265, + 0.907, + 0.334 + ], + "angle": 0, + "content": "Table 4. Ablation study of sampling methods and the number of sampled frames. We compare the proposed Anomaly-focused Temporal Sampler (ATS) with other sampling methods under different frame sampling numbers, including Uniform and Top-K. Latency is the time delay in generating the first token." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.339, + 0.905, + 0.478 + ], + "angle": 0, + "content": "
Frames (N)SamplerLatency (ms)Video-level
BLEU(↑)CIDEr(↑)
8Top-K2440.4621.229
Uniform0.4911.276
ATS (Ours)0.5141.324
16Top-K5660.4761.302
Uniform0.5111.345
ATS (Ours)0.5661.437
32Top-K14020.4811.332
Uniform0.5581.357
ATS (Ours)0.5761.460
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.504, + 0.907, + 0.671 + ], + "angle": 0, + "content": "in Table 3. The inclusion of clip-level data primarily enhanced the model's basic visual perception abilities regarding actions and scenes within the video. Adding event-level data improved the model's ability to judge and understand complete anomaly events. Furthermore, the involvement of video-level data further enhanced the model's ability to analyze and summarize anomaly-related information across longer-span videos. The hierarchical instruction data structure facilitated a comprehensive and complementary improvement in the model's anomaly-related perception-to-reasoning capabilities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Influence of different sampling methods and the number of sampled frames. Our ATS (Anomaly-focused Temporal Sampler) is designed to adaptively sample frames input to the LLM based on the frame-level anomaly scores. To validate its advantages, we compared ATS with other sampling methods at various sample frame counts, including Uniform and Top-K sampling. In Uniform sampling, N frames are uniformly sampled from all frames, while Top-K sampling selects the frames with the top \\(N\\) highest anomaly scores. As shown in Table 4, ATS consistently outperforms other sampling methods, regardless of the sample count. We believe that Uniform sampling tends to overlook key anomaly frames, though this issue lessens as more frames are sampled. Besides, Top-K sampling tends to overly focus on local anomaly frames, missing contextual frame informa" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13849" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.097, + 0.226, + 0.133 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.23, + 0.111, + 0.246, + 0.117 + ], + "angle": 0, + "content": "··" + }, + { + "type": "image", + "bbox": [ + 0.256, + 0.098, + 0.484, + 0.134 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.137, + 0.278, + 0.164 + ], + "angle": 0, + "content": "Prompt: Determine whether there are anomaly events in the video and the specific name of the anomaly." + }, + { + "type": "title", + "bbox": [ + 0.104, + 0.174, + 0.164, + 0.183 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.186, + 0.292, + 0.203 + ], + "angle": 0, + "content": "The anomaly exists and is specifically identified as a Burglary event." + }, + { + "type": "title", + "bbox": [ + 0.104, + 0.221, + 0.187, + 0.23 + ], + "angle": 0, + "content": "Holmes-VAU (Ours)" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.233, + 0.288, + 0.259 + ], + "angle": 0, + "content": "Yes, there is an anomaly in the video, specifically a break-in burglary." + }, + { + "type": "text", + "bbox": [ + 0.306, + 0.146, + 0.478, + 0.19 + ], + "angle": 0, + "content": "InternVL2-BB [Incorrect Judgement] Yes, there are anomaly events in the video. The specific anomaly is that the man in the video is seen throwing a punch at the other man." + }, + { + "type": "text", + "bbox": [ + 0.306, + 0.206, + 0.477, + 0.235 + ], + "angle": 0, + "content": "QwenVL2-7B [Incorrect Judgement] There are no anomaly events in the video." + }, + { + "type": "image_caption", + "bbox": [ + 0.348, + 0.247, + 0.441, + 0.257 + ], + "angle": 0, + "content": "Bulgary037_x264" + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.27, + 0.226, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.272, + 0.304, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.305, + 0.271, + 0.482, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.312, + 0.284, + 0.33 + ], + "angle": 0, + "content": "Prompt: Identify any abnormal events depicted in the video." + }, + { + "type": "text", + "bbox": [ + 0.305, + 0.319, + 0.478, + 0.371 + ], + "angle": 0, + "content": "InternVL2-8B [Incorrect Judgement] The video shows a white truck parked on the side of the road, which is an abnormal behavior as it is not supposed to be parked on the road." + }, + { + "type": "title", + "bbox": [ + 0.104, + 0.348, + 0.164, + 0.356 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.361, + 0.255, + 0.37 + ], + "angle": 0, + "content": "No anomaly exists in the video" + }, + { + "type": "title", + "bbox": [ + 0.104, + 0.395, + 0.188, + 0.404 + ], + "angle": 0, + "content": "Holmes-VAU (Ours)" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.41, + 0.265, + 0.419 + ], + "angle": 0, + "content": "There is no anomaly in the video." + }, + { + "type": "title", + "bbox": [ + 0.306, + 0.379, + 0.364, + 0.388 + ], + "angle": 0, + "content": "QwenVL2-7B" + }, + { + "type": "text", + "bbox": [ + 0.306, + 0.392, + 0.476, + 0.409 + ], + "angle": 0, + "content": "The video does not show any abnormal events." + }, + { + "type": "text", + "bbox": [ + 0.331, + 0.421, + 0.458, + 0.431 + ], + "angle": 0, + "content": "Normal_Videos_902_x264" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.097, + 0.892, + 0.134 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.138, + 0.804, + 0.164 + ], + "angle": 0, + "content": "Prompt: How would you describe the particular anomaly events in the video? What evidence do you have to support your judgment of this event as an anomaly?" + }, + { + "type": "title", + "bbox": [ + 0.506, + 0.173, + 0.567, + 0.181 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.181, + 0.887, + 0.239 + ], + "angle": 0, + "content": "The video anomaly event depicts chaotic street scenes where people are running, police officers are clad in riot gear, and confrontations or protests are underway, with scenes of apprehension, restraint, and property damage unfolding from start to end. The presence of violent behavior, law enforcement intervention, and disruption of public order in the urban environment depicted in the video, which collectively indicate a significant deviation from normal, peaceful civic activity." + }, + { + "type": "title", + "bbox": [ + 0.506, + 0.245, + 0.591, + 0.253 + ], + "angle": 0, + "content": "Holmes-VAU (Ours)" + }, + { + "type": "text", + "bbox": [ + 0.505, + 0.255, + 0.887, + 0.305 + ], + "angle": 0, + "content": "The anomaly events involve a series of chaotic and violent scenes, including police officers in riot gear, people running and pushing against barriers, and a helicopter flying over the scene, suggesting a high level of tension and unrest. The presence of destructive behavior, including property damage, property destruction, and physical altercations, which deviates from normal and peaceful urban activity." + }, + { + "type": "title", + "bbox": [ + 0.506, + 0.313, + 0.742, + 0.322 + ], + "angle": 0, + "content": "InternVL2-8B [Visual Hallucination, Incomplete Analysis]" + }, + { + "type": "text", + "bbox": [ + 0.505, + 0.323, + 0.886, + 0.356 + ], + "angle": 0, + "content": "The video shows a group of people walking down a street, and a man is seen walking down the street with a skateboard. The man is wearing a black jacket and a black hat. The event depicted in the video is classified as an anomaly due to the presence of a large group of people engaging in violent behavior." + }, + { + "type": "title", + "bbox": [ + 0.506, + 0.364, + 0.752, + 0.373 + ], + "angle": 0, + "content": "QwenVL2-7B [Incomplete Description, Incomplete Analysis]" + }, + { + "type": "text", + "bbox": [ + 0.505, + 0.373, + 0.884, + 0.414 + ], + "angle": 0, + "content": "The video shows a group of people walking down the street, and suddenly, a car crashes into a crowd of people. The car is on fire, and the people are running away. The video does not provide any context or explanation for this event, but it is clear that it is an anomaly. The car crash is a violent and unexpected event that is not related to the rest of the video." + }, + { + "type": "text", + "bbox": [ + 0.607, + 0.422, + 0.779, + 0.432 + ], + "angle": 0, + "content": "v=3WmMhircZ0c__#1_label_B4-0-0" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.445, + 0.907, + 0.487 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison of anomaly understanding explanation. Compared with state-of-the-art general MLLMs, i.e., InternVL2 [6] and QwenVL2 [53], our proposed Holmes-VAU demonstrates more accurate anomaly judgment, along with more detailed and comprehensive anomaly-related descriptions and analysis. Correct and wrong explanations are highlighted in green and red, respectively." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.5, + 0.256, + 0.619 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.256, + 0.502, + 0.482, + 0.613 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.621, + 0.483, + 0.677 + ], + "angle": 0, + "content": "Figure 6. Ablation study of trainable parameters. (a) Loss curve during instruction-tuning. (b) We tuned the LoRA [15] parameter \\( r \\) to control trainable parameters, evaluating its impact on VAU capability, and General performance on Video-MME [12]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.701, + 0.483, + 0.746 + ], + "angle": 0, + "content": "tion. Our proposed ATS mitigates both issues. To balance inference efficiency and performance, we set \\( N = 16 \\) as the default sample frame number." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Instruction Tuning Parameters. We conducted an ablation study on the parameter \\( r \\) in LoRA [15] to explore how the trainable parameters affects both the model's VAU performance and its general capability. We use VideoMME [12] to evaluate the model's general capability. The results are shown in Fig. 6, as \\( r \\) increases, the model gradually adapts to the VAU task. However, when \\( r \\) becomes too large, the model's general capability decreases. To retain the original general video understanding capability of the MLLM, we set \\( r = 64 \\) as the default value." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.498, + 0.74, + 0.515 + ], + "angle": 0, + "content": "5.4. Qualitative Comparison" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.907, + 0.615 + ], + "angle": 0, + "content": "We provide qualitative comparisons between Holmes-VAU and existing MLLMs in Fig. 5. The results demonstrate that Holmes-VAU can accurately identify anomalies in videos and provide accurate and complete explanations, highlight the effectiveness and advantage of Holmes-VAU in perceiving video events and analyzing anomalies." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.633, + 0.633, + 0.648 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.907, + 0.902 + ], + "angle": 0, + "content": "In conclusion, this work pushes the boundaries of video anomaly understanding by introducing hierarchical anomaly detection across diverse temporal scales, from momentary clips to extended events. The HIVAU-70k benchmark, with over 70,000 multi-level annotations, addresses a critical gap in the field, enabling comprehensive anomaly analysis in real-world scenarios. Our Anomaly-focused Temporal Sampler (ATS) strategically enhances focus on anomaly-dense segments, optimizing both efficiency and accuracy in long-term anomaly detection. Extensive experiments demonstrate that our hierarchical dataset and ATS-enhanced VLM achieve significant performance gains over conventional methods, proving robust for open-world anomaly understanding. This work sets a new standard for multi-granular anomaly comprehension, paving the way for more fine-grained video anomaly understanding." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13850" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "Acknowledgement This work is supported by the National Natural Science Foundation of China under grants U22B2053 and 623B2039, and in part by the Interdisciplinary Research Program of HUST (2024JCYJ034)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.165, + 0.188, + 0.18 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.189, + 0.482, + 0.245 + ], + "angle": 0, + "content": "[1] Amit Adam, Ehud Rivlin, Ilan Shimshoni, and Daviv Reinitz. Robust real-time unusual event detection using multiple fixed-location monitors. IEEE transactions on pattern analysis and machine intelligence, 30(3):555-560, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.246, + 0.37, + 0.258 + ], + "angle": 0, + "content": "[2] AI@Meta. Llama 3 model card. 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.26, + 0.483, + 0.328 + ], + "angle": 0, + "content": "[3] Kumar Ashutosh, Rohit Girdhar, Lorenzo Torresani, and Kristen Grauman. Hiervl: Learning hierarchical videolanguage embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23066-23078, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.33, + 0.483, + 0.398 + ], + "angle": 0, + "content": "[4] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65-72, 2005. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.4, + 0.482, + 0.468 + ], + "angle": 0, + "content": "[5] Yingxian Chen, Zhengzhe Liu, Baoheng Zhang, Wilton Fok, Xiaojuan Qi, and Yik-Chung Wu. Mgfn: Magnitude-contrastive glance-and-focus network for weakly-supervised video anomaly detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 387-395, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.47, + 0.482, + 0.538 + ], + "angle": 0, + "content": "[6] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.54, + 0.482, + 0.607 + ], + "angle": 0, + "content": "[7] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.609, + 0.482, + 0.677 + ], + "angle": 0, + "content": "[8] Jisheng Dang, Huicheng Zheng, Xiaohao Xu, Longguang Wang, Qingyong Hu, and Yulan Guo. Adaptive sparse memory networks for efficient and robust video object segmentation. IEEE Transactions on Neural Networks and Learning Systems, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.679, + 0.482, + 0.774 + ], + "angle": 0, + "content": "[9] Hang Du, Sicheng Zhang, Binzhu Xie, Guoshun Nan, Jiayang Zhang, Junrui Xu, Hangyu Liu, Sicong Leng, Jiangming Liu, Hehe Fan, et al. Uncovering what why and how: A comprehensive benchmark for causation understanding of video anomaly. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18793-18803, 2024. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.776, + 0.482, + 0.831 + ], + "angle": 0, + "content": "[10] Xinyu Fang, Kangrui Mao, Haodong Duan, Xiangyu Zhao, Yining Li, Dahua Lin, and Kai Chen. Mmbench-video: A long-form multi-shot benchmark for holistic video understanding. arXiv preprint arXiv:2406.14515, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[11] Jia-Chang Feng, Fa-Ting Hong, and Wei-Shi Zheng. Mist: Multiple instance self-training framework for video anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14009-14018, 2021. 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.189, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[12] Chaoyou Fu, Yuhan Dai, Yondong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.163, + 0.905, + 0.258 + ], + "angle": 0, + "content": "[13] Dong Gong, Lingqiao Liu, Vuong Le, Budhaditya Saha, Moussa Reda Mansour, Svetha Venkatesh, and Anton van den Hengel. Memorizing normality to detect anomaly: Memory-augmented deep autoencoder for unsupervised anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1705-1714, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.261, + 0.905, + 0.327 + ], + "angle": 0, + "content": "[14] Mahmudul Hasan, Jonghyun Choi, Jan Neumann, Amit K Roy-Chowdhury, and Larry S Davis. Learning temporal regularity in video sequences. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 733-742, 2016. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.905, + 0.385 + ], + "angle": 0, + "content": "[15] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.386, + 0.905, + 0.455 + ], + "angle": 0, + "content": "[16] Md Mohaiminul Islam, Ngan Ho, Xitong Yang, Tushar Nagarajan, Lorenzo Torresani, and Gedas Bertasius. Video recap: Recursive captioning of hour-long videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18198-18208, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.456, + 0.905, + 0.523 + ], + "angle": 0, + "content": "[17] Hyekang Kevin Joo, Khoa Vo, Kashu Yamazaki, and Ngan Le. Clip-tsa: Clip-assisted temporal self-attention for weakly-supervised video anomaly detection. In 2023 IEEE International Conference on Image Processing (ICIP), pages 3230–3234. IEEE, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.525, + 0.905, + 0.579 + ], + "angle": 0, + "content": "[18] Kumara Kahatapitiya and Michael S Ryoo. Coarse-fine networks for temporal activity detection in videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8385-8394, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.581, + 0.905, + 0.648 + ], + "angle": 0, + "content": "[19] Jaechul Kim and Kristen Grauman. Observe locally, infer globally: a space-time mrf for detecting abnormal activities with incremental updates. In 2009 IEEE conference on computer vision and pattern recognition, pages 2921-2928. IEEE, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.65, + 0.905, + 0.69 + ], + "angle": 0, + "content": "[20] Federico Landi, Cees GM Snoek, and Rita Cucchiara. Anomaly locality in video surveillance. arXiv preprint arXiv:1901.10364, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.692, + 0.905, + 0.747 + ], + "angle": 0, + "content": "[21] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.749, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[22] Shuo Li, Fang Liu, and Licheng Jiao. Self-training multisession learning with transformer for weakly supervised video anomaly detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1395-1403, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[23] Weixin Li, Vijay Mahadevan, and Nuno Vasconcelos. Anomaly detection and localization in crowded scenes. IEEE transactions on pattern analysis and machine intelligence, 36(1):18-32, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[24] Yingwei Li, Yi Li, and Nuno Vasconcelos. Resound: Towards action recognition without representation bias. In Pro" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13851" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "ceedings of the European Conference on Computer Vision (ECCV), pages 513-528, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.122, + 0.483, + 0.177 + ], + "angle": 0, + "content": "[25] Bin Lin, Bin Zhu, Yang Ye, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[26] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.483, + 0.262 + ], + "angle": 0, + "content": "[27] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.482, + 0.305 + ], + "angle": 0, + "content": "[28] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.307, + 0.482, + 0.348 + ], + "angle": 0, + "content": "[29] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.35, + 0.482, + 0.405 + ], + "angle": 0, + "content": "[30] Kun Liu and Huadong Ma. Exploring background-bias for anomaly detection in surveillance videos. In Proceedings of the 27th ACM International Conference on Multimedia, pages 1490-1499, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.407, + 0.482, + 0.462 + ], + "angle": 0, + "content": "[31] Wen Liu, Weixin Luo, Dongze Lian, and Shenghua Gao. Future frame prediction for anomaly detection-a new baseline. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6536-6545, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.464, + 0.482, + 0.517 + ], + "angle": 0, + "content": "[32] Yi Liu, Limin Wang, Yali Wang, Xiao Ma, and Yu Qiao. Fineaction: A fine-grained video dataset for temporal action localization. IEEE transactions on image processing, 31: 6937-6950, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.482, + 0.601 + ], + "angle": 0, + "content": "[33] Zhian Liu, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. A hybrid video anomaly detection framework via memory-augmented flow reconstruction and flow-guided frame prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 13588-13597, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.482, + 0.657 + ], + "angle": 0, + "content": "[34] Cewu Lu, Jianping Shi, and Jiaya Jia. Abnormal event detection at 150 fps in matlab. In Proceedings of the IEEE international conference on computer vision, pages 2720-2727, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.482, + 0.702 + ], + "angle": 0, + "content": "[35] Jian Lu, Xuanfeng Li, Bo Zhao, and Jian Zhou. A review of skeleton-based human action recognition. Journal of Image and Graphics, 28(12):3651-3669, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.482, + 0.744 + ], + "angle": 0, + "content": "[36] Hui Lv and Qianru Sun. Video anomaly detection and explanation via large language models. arXiv preprint arXiv:2401.05702, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.747, + 0.482, + 0.815 + ], + "angle": 0, + "content": "[37] Hui Lv, Zhongqi Yue, Qianru Sun, Bin Luo, Zhen Cui, and Hanwang Zhang. Unbiased multiple instance learning for weakly supervised video anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8022-8031, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[38] Baiteng Ma, Shiwei Zhang, Changxin Gao, and Nong Sang. Temporal global correlation network for end-to-end action proposal generation. Acta Electronica Sinica, 50(10):2452-2461, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[39] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "understanding via large vision and language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (ACL 2024), 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.137, + 0.905, + 0.191 + ], + "angle": 0, + "content": "[40] Ramin Mehran, Alexis Oyama, and Mubarak Shah. Abnormal crowd behavior detection using social force model. In 2009 IEEE conference on computer vision and pattern recognition, pages 935-942. IEEE, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.194, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[41] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311–318, 2002. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[42] Yujiang Pu, Xiaoyu Wu, Lulu Yang, and Shengjin Wang. Learning prompt-enhanced context features for weakly-supervised video anomaly detection. IEEE Transactions on Image Processing, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.322, + 0.905, + 0.416 + ], + "angle": 0, + "content": "[43] Zhiwu Qing, Shiwei Zhang, Ziyuan Huang, Yi Xu, Xiang Wang, Mingqian Tang, Changxin Gao, Rong Jin, and Nong Sang. Learning from untrimmed videos: Self-supervised video representation learning with hierarchical consistency. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13821-13831, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.905, + 0.502 + ], + "angle": 0, + "content": "[44] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.505, + 0.905, + 0.572 + ], + "angle": 0, + "content": "[45] Dian Shao, Yue Zhao, Bo Dai, and Dahua Lin. Finegym: A hierarchical video dataset for fine-grained action understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2616-2625, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.907, + 0.631 + ], + "angle": 0, + "content": "[46] Waqas Sultani, Chen Chen, and Mubarak Shah. Real-world anomaly detection in surveillance videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6479-6488, 2018. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.633, + 0.905, + 0.701 + ], + "angle": 0, + "content": "[47] Jiaqi Tang, Hao Lu, Ruizheng Wu, Xiaogang Xu, Ke Ma, Cheng Fang, Bin Guo, Jiangbo Lu, Qifeng Chen, and YingCong Chen. Hawk: Learning to understand open-world video anomalies. arXiv preprint arXiv:2405.16886, 2024. 2,3,4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.704, + 0.905, + 0.786 + ], + "angle": 0, + "content": "[48] Kamalakar Vijay Thakare, Yash Raghuwanshi, Debi Prosad Dogra, Heeseung Choi, and Ig-Jae Kim. Dyannet: A scene dynamicity guided self-trained video anomaly detection network. In Proceedings of the IEEE/CVF Winter conference on applications of computer vision, pages 5541-5550, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[49] Yu Tian, Guansong Pang, Yuanhong Chen, Rajvinder Singh, Johan W Verjans, and Gustavo Carneiro. Weakly-supervised video anomaly detection with robust temporal feature magnitude learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4975-4986, 2021. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[50] Anil Osman Tur, Nicola Dall'Asen, Cigdem Bayan, and Elisa Ricci. Exploring diffusion models for unsupervised" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13852" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.092, + 0.484, + 0.133 + ], + "angle": 0, + "content": "video anomaly detection. In 2023 IEEE International Conference on Image Processing (ICIP), pages 2540-2544. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.484, + 0.19 + ], + "angle": 0, + "content": "[51] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4566-4575, 2015. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.483, + 0.246 + ], + "angle": 0, + "content": "[52] Jue Wang and Anoop Cherian. Gods: Generalized one-class discriminative subspaces for anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8201-8211, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.246, + 0.483, + 0.342 + ], + "angle": 0, + "content": "[53] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.342, + 0.483, + 0.385 + ], + "angle": 0, + "content": "[54] SC Wang, Q Huang, YF Zhang, X Li, YQ Nie, and GC Luo. Review of action recognition based on multimodal data. Image Graph, 27(11):3139-3159, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.385, + 0.483, + 0.442 + ], + "angle": 0, + "content": "[55] Jhih-Ciang Wu, He-Yen Hsieh, Ding-Jie Chen, Chiou-Shann Fuh, and Tyng-Luh Liu. Self-supervised sparse representation for video anomaly detection. In European Conference on Computer Vision, pages 729-745. Springer, 2022. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.483, + 0.537 + ], + "angle": 0, + "content": "[56] Peng Wu, Jing Liu, Yujia Shi, Yujia Sun, Fangtao Shao, Zhaoyang Wu, and Zhiwei Yang. Not only look, but also listen: Learning multimodal violence detection under weak supervision. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXX 16, pages 322-339. Springer, 2020. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.538, + 0.483, + 0.609 + ], + "angle": 0, + "content": "[57] Peng Wu, Xuerong Zhou, Guansong Pang, Yujia Sun, Jing Liu, Peng Wang, and Yanning Zhang. Open-vocabulary video anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18297-18307, 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.608, + 0.483, + 0.691 + ], + "angle": 0, + "content": "[58] Peng Wu, Xuerong Zhou, Guansong Pang, Lingru Zhou, Qingsen Yan, Peng Wang, and Yanning Zhang. Vadclip: Adapting vision-language models for weakly supervised video anomaly detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6074-6082, 2024. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.691, + 0.483, + 0.748 + ], + "angle": 0, + "content": "[59] Dan Xu, Yan Yan, Elisa Ricci, and Nicu Sebe. Detecting anomalous events in videos by learning deep representations of appearance and motion. Computer Vision and Image Understanding, 156:117-127, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.483, + 0.804 + ], + "angle": 0, + "content": "[60] Xiaohao Xu, Jinglu Wang, Xiang Ming, and Yan Lu. Towards robust video object segmentation with adaptive object calibration. In Proceedings of the 30th ACM International Conference on Multimedia, pages 2709-2718, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.483, + 0.872 + ], + "angle": 0, + "content": "[61] Zhiwei Yang, Jing Liu, Zhaoyang Wu, Peng Wu, and Xiaotao Liu. Video event restoration based on keyframes for video anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14592-14601, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.872, + 0.483, + 0.903 + ], + "angle": 0, + "content": "[62] Zhiwei Yang, Jing Liu, and Peng Wu. Text prompt with normality guidance for weakly supervised video anomaly detec" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.092, + 0.905, + 0.134 + ], + "angle": 0, + "content": "tion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18899-18908, 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.135, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[63] Yu Yao, Xizi Wang, Mingze Xu, Zelin Pu, Yuchen Wang, Ella Atkins, and David J Crandall. Dota: Unsupervised detection of traffic anomaly in driving videos. IEEE transactions on pattern analysis and machine intelligence, 45(1): 444-459, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.205, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[64] Tongtong Yuan, Xuange Zhang, Kun Liu, Bo Liu, Chen Chen, Jian Jin, and Zhenzhen Jiao. Towards surveillance video-and-language understanding: New dataset, baselines, and challenges, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.262, + 0.905, + 0.346 + ], + "angle": 0, + "content": "[65] Tongtong Yuan, Xuange Zhang, Kun Liu, Bo Liu, Chen Chen, Jian Jin, and Zhenzhen Jiao. Towards surveillance video-and-language understanding: New dataset baselines and challenges. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22052-22061, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.346, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[66] M Zaigham Zaheer, Arif Mahmood, M Haris Khan, Mattia Segu, Fisher Yu, and Seung-Ik Lee. Generative cooperative learning for unsupervised video anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14744-14754, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.417, + 0.905, + 0.488 + ], + "angle": 0, + "content": "[67] Luca Zanella, Willi Menapace, Massimiliano Mancini, Yiming Wang, and Elisa Ricci. Harnessing large language models for training-free video anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18527-18536, 2024. 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.488, + 0.905, + 0.531 + ], + "angle": 0, + "content": "[68] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.531, + 0.905, + 0.587 + ], + "angle": 0, + "content": "[69] Huaxin Zhang, Xiang Wang, Xiaohao Xu, Zhiwu Qing, Changxin Gao, and Nong Sang. Hr-pro: Point-supervised temporal action localization via hierarchical reliability propagation. arXiv preprint arXiv:2308.12608, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.587, + 0.905, + 0.657 + ], + "angle": 0, + "content": "[70] Huaxin Zhang, Xiang Wang, Xiaohao Xu, Xiaonan Huang, Chuchu Han, Yuehuan Wang, Changxin Gao, Shanjun Zhang, and Nong Sang. Glancevad: Exploring glance supervision for label-efficient video anomaly detection. arXiv preprint arXiv:2403.06154, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.657, + 0.905, + 0.699 + ], + "angle": 0, + "content": "[71] Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, and Chunyuan Li. Video instruction tuning with synthetic data. arXiv preprint arXiv:2410.02713, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.7, + 0.905, + 0.743 + ], + "angle": 0, + "content": "[72] Bin Zhao, Li Fei-Fei, and Eric P Xing. Online detection of unusual events in videos via dynamic sparse coding. In CVPR 2011, pages 3313-3320. IEEE, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.743, + 0.905, + 0.826 + ], + "angle": 0, + "content": "[73] Jia-Xing Zhong, Nannan Li, Weijie Kong, Shan Liu, Thomas H Li, and Ge Li. Graph convolutional label noise cleaner: Train a plug-and-play action classifier for anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1237-1246, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.826, + 0.905, + 0.884 + ], + "angle": 0, + "content": "[74] Hang Zhou, Junqing Yu, and Wei Yang. Dual memory units with uncertainty regulation for weakly supervised video anomaly detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3769-3777, 2023. 2, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.884 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13853" + } + ] +] \ No newline at end of file diff --git a/2025/Holmes-VAU_ Towards Long-term Video Anomaly Understanding at Any Granularity/8b3a9750-2575-4fe7-bb78-a5c96dc5aac4_origin.pdf b/2025/Holmes-VAU_ Towards Long-term Video Anomaly Understanding at Any Granularity/8b3a9750-2575-4fe7-bb78-a5c96dc5aac4_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..41ff0fe6161baef7c491c39409fbbb7a993849ef --- /dev/null +++ b/2025/Holmes-VAU_ Towards Long-term Video Anomaly Understanding at Any Granularity/8b3a9750-2575-4fe7-bb78-a5c96dc5aac4_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:faf52a7950d9e6eee95640deed907f5633fdf200c1e42a1a4ae3051caeb72053 +size 1394528 diff --git a/2025/Holmes-VAU_ Towards Long-term Video Anomaly Understanding at Any Granularity/full.md b/2025/Holmes-VAU_ Towards Long-term Video Anomaly Understanding at Any Granularity/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a666a39d7beaca035f467aacef7ca7ad58cf97a4 --- /dev/null +++ b/2025/Holmes-VAU_ Towards Long-term Video Anomaly Understanding at Any Granularity/full.md @@ -0,0 +1,370 @@ +# Holmes-VAU: Towards Long-term Video Anomaly Understanding at Any Granularity + +Huaxin Zhang $^{1}$ Xiaohao Xu $^{2}$ Xiang Wang $^{1}$ Jialong Zuo $^{1}$ Xiaonan Huang $^{2}$ Changxin Gao $^{1}$ Shanjun Zhang $^{3}$ Li Yu $^{1}$ Nong Sang $^{1*}$ $^{1}$ Key Laboratory of Image Processing and Intelligent Control, School of Artificial Intelligence and Automation, Huazhong University of Science and Technology + $^{2}$ University of Michigan, Ann Arbor + $^{3}$ Kanagawa University + +{zhanghuaxin,wxiang,cgao,hustlyu,nsang}@hust.edu.cn,{xiaohaox,xiaonanh}@umich.edu,{chiyoz01}@kanagawa-u.ac.jp + +![](images/91680e9686dac8a1c7c39831e48df448d8514ddd8efc7f8e41225746cb194df8.jpg) +Figure 1. Motivation. Left: Existing datasets lack the hierarchical structure to capture transient and sustained anomalies across varying temporal scales. Our HIVAU-70k dataset addresses this by providing multi-granularity annotations—clip, event, and video levels—that enable detailed anomaly analysis in complex real-world scenarios. Right: Inspired by Sherlock Holmes's knack for zeroing in on critical details, our Holmes-VAU method integrates an Anomaly-focused Temporal Sampler with a multi-modal LLM, directing model attention to anomaly-rich segments, which enables models to decode complex, long-term video anomalies efficiently. + +![](images/20f4aec5a3fd4a30388dd8ab4138d26c5614df3c98528ba32bf241494842a176.jpg) + +# Abstract + +How can we enable models to comprehend video anomalies occurring over varying temporal scales and contexts? Traditional Video Anomaly Understanding (VAU) methods focus on frame-level anomaly prediction, often missing the interpretability of complex and diverse real-world anomalies. Recent multimodal approaches leverage visual and textual data but lack hierarchical annotations that capture both short-term and long-term anomalies. To address this challenge, we introduce HIVAU-70k, a large-scale benchmark for hierarchical video anomaly understanding across any granularity. We develop a semi-automated annotation engine that efficiently scales high-quality annotations by combining manual video segmentation with recursive free-text + +annotation using large language models (LLMs). This results in over 70,000 multi-granular annotations organized at clip-level, event-level, and video-level segments. For efficient anomaly detection in long videos, we propose the Anomaly-focused Temporal Sampler (ATS). ATS integrates an anomaly scorer with a density-aware sampler to adaptively select frames based on anomaly scores, ensuring that the multimodal LLM concentrates on anomaly-rich regions, which significantly enhances both efficiency and accuracy. Extensive experiments demonstrate that our hierarchical instruction data markedly improves anomaly comprehension. The integrated ATS and visual-language model outperform traditional methods in processing long videos. Our benchmark and model are publicly available at https://github.com/pipixin321/HolmesVAU. + +# 1. Introduction + +Video Anomaly Understanding (VAU) is crucial for applications such as video surveillance [46], violent content analysis [56], and autonomous driving [63]. Detecting deviations from normal patterns aids in hazard prevention and real-time decision-making. Traditional methods [14, 49, 74] mainly focus on frame-level predefined closed-set anomaly prediction, assigning an anomaly score to each frame. However, these approaches often fail to describe and understand complex anomalies in the real world. + +To address this gap, open-world anomaly understanding [57] embraces the diversity and unpredictability of real-world anomalies. Recent work integrates multimodal approaches, combining visual data with textual descriptions [42, 58, 62, 65], while advances in multimodal visual-language models (VLMs) [6, 21, 27, 29, 68] have enabled more nuanced understanding through anomaly-related instruction tuning and text generation [9, 36, 47, 67]. + +Despite these advancements, a significant gap remains in models' ability to comprehend anomalies across multiple temporal scales. For instance, while anomalies such as explosions or fights may be captured in a single frame, more complex events like theft or arson require understanding long-term contextual patterns. Existing VAU datasets [46, 56] typically provide annotations at a single level of granularity, limiting models to understanding either immediate perceptual anomalies or those requiring extended contextual reasoning. The lack of datasets with hierarchical annotations—encompassing both short-term and long-term anomalies—hinders models' capacity to reason about anomalies with diverse temporal characteristics. Moreover, constructing datasets that encapsulate this hierarchical complexity poses significant challenges in scalability and annotation quality. + +To address these issues, we develop a semi-automated annotation engine that efficiently scales high-quality annotation by combining manual video segmentation with recursive free-text annotation using large language models (LLMs). The process involves three key stages: 1) hierarchical video decoupling, where we manually identify anomaly events and segment them into shorter clips; 2) hierarchical free-text annotation, where captions for each clip are generated through human effort or video captioning models, then summarized at the event and video levels via LLMs; and 3) hierarchical instruction construction, where the textual data is transformed into question-answer instruction prompts by combining captions and summaries with designed prompts, creating a dataset with rich annotations for training and evaluating models. + +Utilizing the annotation engine, we introduce HIVAU-70k, a large-scale video anomaly understanding benchmark with hierarchical instructions. Our dataset comprises over 70,000 multi-granular instruction data organized + +across clip-level, event-level, and video-level segments as shown in Fig. 1. This hierarchical structure empowers models to detect immediate anomalies, e.g., sudden explosions or fighting, as well as complex events that require an understanding of long-term context, like theft or arson. By annotating at multiple temporal levels, HIVAU-70k provides diverse anomalies in open-world scenarios. + +Towards long-term VAU, efficiency remains a critical challenge. Previous methods [9, 47, 67] often rely on uniform frame sampling, which can either miss crucial anomaly frames or incur large computational costs [25, 47, 67]. To address this, we propose the Holmes-VAU method, which combines the proposed Anomaly-focused Temporal Sampler (ATS) with the multimodal visual-language model for efficient long-term video anomaly understanding (See Fig. 1). The ATS combines a anomaly scorer with a density-aware sampler that adaptively selects frames by their anomaly scores. This integration ensures that the visual-language model concentrates on anomaly-rich regions, enhancing both efficiency and accuracy. + +Our contributions are threefold: 1) We introduce HIVAU-70k, a large-scale, multi-granular benchmark for hierarchical video anomaly understanding. 2) We propose the Holmes-VAU method, which combines the proposed Anomaly-focused Temporal Sampler (ATS) to boost the efficiency and accuracy of inference on long-term videos. 3) We conduct extensive experiments demonstrating the effectiveness of hierarchical instruction data in enhancing anomaly comprehension and validate the performance gains provided by the integrated ATS and visual-language model in processing long videos. + +# 2. Related Works + +Video Anomaly Detection. This task aims to temporally detect abnormal frames in a long untrimmed video [1, 14, 23, 34, 40, 52]. Early attempts are based on handcrafted features [1, 19, 23, 34, 40, 72]. Recently, deep learning approaches [14, 37, 61] have become dominant, broadly classified into unsupervised, weakly-supervised, and fully-supervised methods. Unsupervised methods train only on normal videos to learn normal patterns and are often designed as reconstruction-based [13, 14, 59, 61], prediction-based [31], or a combination [33]. Some methods [48, 50, 66] also explore a fully unsupervised setting, including both normal and abnormal videos in training set without real labels. Weakly-supervised methods [11, 22, 37, 46, 49, 55, 56, 70, 73, 74] use both normal and abnormal videos with video-level annotations. Fully-supervised methods [20, 30] are less studied due to the high cost of precise frame-level annotations. + +Multi-modal Video Anomaly Understanding. Large-scale visual-language pre-trained models such as CLIP [44] + +serve as a bridge between visual and textual modalities. Some recent works [17, 42, 58, 62] in the realm of video anomaly detection have leveraged textual information as prompts to enhance the model's anomaly representation. Based on this, [57] firstly proposed the open vocabulary VAD task, [65] introduced a multimodal video anomaly dataset composed of dense clip captions. Furthermore, [67] extracted captions from video frames and designed prompts for LLMs to provide anomaly scores, [9] and [47] construct diverse and interactive instruction data at the video level. However, these datasets consider only a single temporal level of anomaly understanding data construction, i.e. clip-level [65] or video-level [9, 47]. Unlike these methods, we focus on building large-scale hierarchical video anomaly understanding data for multimodal instruction tuning. + +Hierarchical Video Understanding. Video understanding is a challenging task due to its temporal-scale diversity. To better comprehend videos, many previous works have focused on both datasets and models in hierarchical video understanding. For example, [3, 24, 32, 35, 45, 54] proposed fine-grained action recognition and localization datasets, [16] provided free-form hierarchical captions for hour-long videos at multiple temporal scales, and [8, 18, 38, 43, 60, 69] trained models on hierarchical levels to obtain better video feature representation. Recently, to assess the capability of video vision-language models (VLMs) in handling challenges in real-world scenarios, several video benchmarks [10, 12] have been built, incorporating data at multiple temporal scales for evaluation. Unlike these works, we dive into the field of Video Anomaly Understanding, thus filling the gap in multi-scale annotations in this area. + +# 3. HIVAU-70k Benchmark + +We first define the video anomaly understanding task in Sec.3.1. Then, the construction process of the HIVAU-70k benchmark will be elaborated in Sec.3.2. Finally, we will present HIVAU-70k's statistical information in Sec. 3.3. + +# 3.1. Task Description + +This work focuses on video anomaly understanding, involving temporal anomaly detection and anomaly explainability. Temporal anomaly detection aims to predict an anomaly score for each frame in a video $\mathcal{V}$ , represented as $S\in \mathbb{R}^T$ , where $T$ is the total number of frames. Building on this, we explore the model's ability to generate explanatory text outputs related to anomalies based on video input and user queries. Specifically, given a video $\mathcal{V}$ and a text query $\mathcal{Q}$ , the model generates an anomaly-related response $\mathcal{A}$ . We consider two key abilities: 1) visual perception, which involves recognizing main entities in the video, and 2) anomaly reasoning, which encompasses the + +![](images/716735a97f013fd3347cdca1ef45bb25a82e5313a822f0e352fb671715079360.jpg) +Figure 2. Data Engine. We present a structured workflow for generating hierarchical annotations across video, event, and clip levels. Clips are first captioned, then processed through a large language model (LLM) with prompts for event summarization. The outputs include clip captions, event summaries, and video summaries, followed by manual checking and refinement. This multi-step approach enriches the dataset with detailed judgments, descriptions, and analyses of anomalies, enabling robust contextual understanding at varying granularities. + +model's judgment and analysis of the anomaly content. + +# 3.2. LLM-Empowered Data Engine + +As shown in Fig.2, we develop a semi-automated annotation engine that efficiently scales high-quality annotations, which consists of three main steps: 1) hierarchical video decoupling, 2) hierarchical free-text annotation, and 3) hierarchical instruction construction. + +Hierarchical Video Decoupling. Our video sources include the training set of the UCF-Crime [46] dataset and the XD-Violence [56] dataset, which contains videos of varying durations and diverse real-world anomalies. For abnormal videos, we first manually obtain the temporal boundaries of each anomaly event in the video. Non-continuous anomalous states are considered separate events. Then, we divide each event into clips of random lengths. For normal videos, we apply random sampling to obtain corresponding segments of varying granularities. Ultimately, we obtained 5,443 videos, 11,076 events, and 55,806 clips. This process took 5 annotators approximately 20 hours to complete. More details can be found in the appendix. + +Hierarchical Free-text Annotation. To fully extract semantic information from the clip-level videos, we utilize a powerful off-the-shell video perception model LLaVA-Next-Video [28] to generate detailed captions for each clip. We also include the UCA dataset [64], which provides manually annotated captions for video clips in the UCF-Crime [46] dataset. Then, we use an LLM [2] to consoli + +![](images/b5288291184c2c45051df10e1fc4281661cd7533f284234b3570a9825728988e.jpg) +Q: Give a short and clear caption of the video clip. + +![](images/afe828ef6ae55dc7a7bca169325cd44c0a08c5ac8c2dabb99be098c56c997894.jpg) + +![](images/fb2b43b8e06015e5e6f484b9c008b3e8e3a78acc326bb6e6148738050b275d0d.jpg) + +![](images/5fa90769cc6e549f8ba56fbcbed5274d7753f89b8efe0b2300fe0d9a028f65b5.jpg) +A: A group of policemen dragged a man red out of the gate. +Q: Are there any anomaly events detected in the video? +A: An anomaly exists, specifically an Arrest anomaly. +Q: Describe the anomaly events in the video. A: The anomaly event involves the three policemen turning the man in red over and moving him to the wall, indicating a possible arrest or detention. +Q: Why is this event as an anomaly? A: The unusual and potentially forceful interaction between the policemen and the man in red, deviating from normal pedestrian behavior in the scene +Figure 3. HVAU-70k dataset. (a) Duration distributions for clips, events, and full videos, showing dominance of short clips. (b) Hierarchical data organization from clip-level to video-level, enabling perception-to-reasoning insights. (c) Word count variations across annotation levels, with more detailed descriptions at the video level. (d) Sample annotations capturing captioning, judgment, description, and anomaly analysis, highlighting nuanced understanding of anomaly events in complex scenes. + +date all clip captions within an event, generating an event-level video summary. Specifically, we design prompts1 to guide the LLM to produce three parts of content for each anomalous event summary: 1) Judgment: A determination of whether an anomaly exists and its specific category, 2) Description: A detailed description of the anomalous or normal event, 3) Analysis: The reasoning behind the anomaly judgment, including causal analysis. To guide the LLM in generating reliable responses, we also inject the event's category label (e.g., "Shooting", "Explosion") into the LLM prompt. We consolidate all event-level summaries to obtain the video-level summary. This results in free-text annotations across multiple temporal scales, including short-term visual perception (clip-level) and long-term anomaly reasoning (event-level, video-level). + +Hierarchical Instruction Data Construction. The ability of VLMs to follow user instructions and generate responses is achieved through instruction-tuning [21, 25, 27, 29]. The training data format typically consists of the following: + +$\{\mathcal{Q}$ :user instruction, $\mathcal{A}$ :model response.} + +To build instruction-tuning data for VLMs in the domain of Video Anomaly Understanding, we matched free-text annotations with pre-designed anomaly-related user instructions. Specifically, for clip-level segments, we only construct instructions related to captions, as it is challenging to obtain anomaly-related analysis for short videos. For event-level and video-level segments, we construct instruction data from the perspectives of Judgment, Description, and Analysis. Typical examples are shown in the Fig. 3 (d). Manual Checking for Data Quality Control. To ensure dataset quality, we implemented several human inspec + +tion and curation strategies. First, we labeled the temporal boundaries of abnormal event segments and reviewed them at the second level. Next, anomaly labels were incorporated during summary generation using LLMs, directing focus on relevant entities. Finally, manual reviews were performed to correct low-quality instruction data. + +# 3.3. Data Statistic of HIVAU-70k + +Utilizing the proposed annotation engine, we introduce HIVAU-70k, as shown in Fig. 3, a large-scale benchmark designed for hierarchical instruction-based video anomaly understanding. As shown in Fig. 3(b), HIVAU-70k contains over 70,000 multi-granular annotations organized at clip-level, event-level, and video-level segments, achieving a progression from perception to reasoning. As shown in Fig. 3(a) and (c), the durations of segments and the word numbers of text annotations at different granularities exhibit significant distributional differences. As shown in Fig. 3(d), HIVAU-70k's instruction data covers Caption, Judgment, Description, and Analysis for real-world anomalies, which guide the model to develop both short-term and long-term video anomaly understanding capabilities. + +# 4. Method: Holmes-VAU + +Long-term video anomaly understanding with LLMs/VLMs has traditionally been hindered by frame redundancy, complicating accurate anomaly detection. Previous VAU approaches struggle with focus: methods like dense window sampling [67] add redundancy, and uniform frame sampling [9, 47] often misses key anomalies, limiting application to short videos. We introduce the Anomaly-focused Temporal Sampler (ATS) to address this, + +![](images/f67777cea1f190d73ce83a9f2a9cff77e319b69aae4387130340c2fb515e0dca.jpg) +Figure 4. Holmes-VAU: a multi-modal-LLM-based video anomaly detection framework with adaptive anomaly focus. + +integrating it into the VLM, and fine-tuning it via instruction on HIVAU-70k to form our Holmes-VAU model. + +# 4.1. Pipeline Overview + +The overall pipeline of our Holmes-VAU model is shown in Fig. 4. Video frames are processed by a visual encoder, creating visual tokens. These tokens are analyzed by an Anomaly-focused Temporal Sampler using an anomaly scorer and cumulative sum ( $S_{cumsum}$ ) to select keyframes. A text encoder processes a prompt (e.g., 'Describe the abnormal events in the video'). Visual and textual representations are combined in a pre-trained language model, fine-tuned with LoRA, to generate a description of detected anomalies, such as a 'collision between a car and a motorcycle' and 'traffic accident' indicators. + +# 4.2. Model Architecture + +Visual and Text embedding. We utilize the frozen visual encoder in InternVL2 [6], which inherits the ViT structure from CLIP [44], we refer to it as $\phi_v$ . Following previous VAU works [55, 58, 67, 74], we sample dense video frames at an interval of 16 frames from the input video. Each video frame is then processed through the visual encoder to obtain the corresponding visual tokens. Given the input video frame sequence $\mathcal{V} \in \mathbb{R}^{T \times H \times W \times C}$ , the output features of $i$ -th frame can be denoted as: + +$$ +V _ {i} = \left\{v _ {i} ^ {c l s}, v _ {i} ^ {1}, v _ {i} ^ {2}, \dots , v _ {i} ^ {N _ {p}} \right\} = \phi_ {v} \left(\mathcal {V} _ {i}\right) \tag {1} +$$ + +where $v_{i}^{cls}$ indicates the class token, $v_{i}^{j}$ ( $j \in \{1,2,\dots,N_{p}\}$ ) denotes the patch tokens, and $N_{p}$ represents the number of patches. The text encoder $\phi_t$ is also initialized from [6], + +which includes a tokenizer and an embedding layer. The prompt text $\mathcal{Q}$ is converted into text tokens through the text encoder: $X_{q} = \phi_{t}(\mathcal{Q})$ + +Anomaly-focused Temporal Sampler (ATS). ATS consists of two components: the anomaly scorer and the density-aware sampler. The anomaly scorer $\phi_s$ is a feature-based VAD network which estimates the anomaly score for each frame. We follow the network architecture in [74] due to its simplicity and good performance. Given the class token of the video frames $\{v_1^{cls}, v_2^{cls}, \dots, v_T^{cls}\}$ , the anomaly scores can be obtained: $s_i = \phi_s(v_i^{cls})$ , where $s_i$ denotes the anomaly score of the $i$ -th frame. + +Anomalous frames typically contain more information and exhibit greater variation than normal frames [49]. This observation motivates us to sample more frames in regions with higher anomaly scores while reducing the sampling in areas with lower anomaly scores. As shown in Fig. 4, to achieve non-uniform sampling, we propose density-aware sampler to selectively choose $N$ frames from a total of $T$ input frames. Specifically, we treat the anomaly scores $S \in \mathbb{R}^T$ as a probability mass function and first accumulate them along the temporal dimension to obtain the cumulative distribution function, denoted as $S_{cumsum}$ : + +$$ +S _ {c u m s u m} (t) = \sum_ {i = 1} ^ {t} \left(s _ {i} + \tau\right) \tag {2} +$$ + +We uniformly sample $N$ points along the cumulative axis, then map these points to the cumulative distribution $S_{cumsum}$ , the corresponding $N$ timestamps on the time axis are mapped to the closest frame index and finally form the sampled frame indices, denoted as $\mathcal{G}$ . $\tau$ is used to control the uniformity of the sampling. + +Projector and LLM. We select the tokens corresponding to the sampled frame, i.e., $\mathcal{G}$ , as the visual embedding. A projector $\phi_p$ is then used to map the visual embedding to the language feature space. Finally, we concatenate these embeddings with the text embeddings, input them into the pre-trained large language model, and compute the probability of the target answers $X_a$ . To obtain an initial visual-language alignment space, we initialize the projector and LLM parameters from [6], with the parameters kept frozen. The above process can be expressed as follows: + +$$ +X _ {i n s} = \operatorname {c a t} \left[ \phi_ {p} \left(\operatorname {c a t} \left[ V _ {i} \right]\right), X _ {q} \right] \quad i \in \mathcal {G} \tag {3} +$$ + +$$ +p \left(X _ {a} \mid X _ {i n s}\right) = \prod_ {i = 1} ^ {L} p _ {\theta} \left(\left(x _ {i} \mid X _ {i n s, < i}, X _ {a, < i}\right)\right) \tag {4} +$$ + +where $\operatorname{cat}[\cdot]$ represents the concatenation operation, $\theta$ is the trainable parameters, $X_{ins, < i}, X_{a, < i}$ are the instruction and answer tokens in all turns before the current prediction token $x_i$ , $L$ is the length of sequence, respectively. + +# 4.3. Training and Testing + +Training. We train the model following two steps. In the first step, we use the video data and annotated frame-level label $(\hat{y} \in \mathbb{R}^T)$ from HIVAU-70k to train the anomaly scorer, which provides more accurate anomaly supervision compared to previous unsupervised and weakly-supervised methods [14, 46, 55, 74]. + +$$ +\mathcal {L} _ {A S} = - \sum_ {i = 1} ^ {T} \left(s _ {i} \log \left(\hat {y} _ {i}\right) + \left(1 - s _ {i}\right) \log \left(1 - \hat {y} _ {i}\right)\right) \tag {5} +$$ + +In the second step, we keep the anomaly scorer fixed, and use all the instruction data from HIVAU-70k to train the model. To achieve more efficient fine-tuning without disrupting the original capabilities of the LLM, we employ LoRA [15] for fine-tuning, optimizing the cross entropy loss between the predicted and the ground truth tokens. + +Testing. During testing, the user inputs a video and text prompts; the model will generate the corresponding text response following the user's instruction. + +# 5. Experiments + +# 5.1. Experiment Setup + +Dataset. Our HIVAU-70k is built upon two large-scale real-world datasets, i.e., UCF-Crime [46] and XD-Violence [56], they provide a diverse range of videos with anomalous events. UCF-Crime [46] comprises 1,900 untrimmed videos totaling 128 hours from outdoor and indoor surveillance cameras. It encompasses 13 classes of real-world anomalies, including Abuse, Explosion, Fighting, and Shooting. XD-Violence [56] is the largest VAD benchmark, comprising 4,754 videos totaling 217 hours sourced from surveillance, movies, car cameras, and games. It encompasses 6 anomaly classes: Abuse, Car Accidents, Explosions, Fighting, Riots, and Shooting. + +Metric. We assess the anomaly understanding ability from two aspects: anomaly detection and reasoning. 1) For anomaly detection, we use the anomaly scores output by the Anomaly Scorer as the prediction and perform the evaluation. Following [14, 46, 67, 74], we use AUC and AP to quantify detection performance, which is evaluated only on the video level. 2) For anomaly reasoning, we annotate instruction data from the UCF-Crime and XD-Violence test sets, which have been carefully reviewed and filtered by annotators. We finally collected 3,300 test samples at multiple granularities. The test set contains 2200/732/398 samples at clip/event/video levels. We calculate metrics including BLEU [41], CIDEr [51], METEOR [4] and ROUGE [26] to measure the quality of the reasoning text output by the model, comparing with the annotated ground truth text. + +Implementation Details. For the proposed Holmes-VAU + +Table 1. Comparison of detection performance with state-of-the-art Video Anomaly Detection approaches. We include the results of explainable and non-explainable methods. \*\* represents the result reported in [67]. + +
MethodsBackboneXD-ViolenceUCF-Crime
AP/%AUC/%
Non-explainable VAD
Conv-AE [14] (CVPR'16)-27.2550.60
GODS [52] (ICCV'19)I3DN/A70.46
GCL [66] (CVPR'22)ResNextN/A71.04
DYANNET [48] (WACV'23)I3DN/A84.50
MIST [11] (CVPR'21)I3DN/A82.30
Wu et al. [56] (ECCV'20)I3D78.6482.44
RTFM [49] (ICCV'21)I3D77.8184.30
MSL [22] (AAAI'22)I3D78.2885.30
S3R [55] (ECCV'22)I3D80.2685.99
MGFN [5] (AAAI'23)I3D79.1986.98
UR-DMU [74] (AAAI'23)I3D81.6686.97
CLIP-TSA [17] (ICIP'23)ViT82.1987.58
VadCLIP [58] (AAAI'24)ViT84.5188.02
Yang et al. [62] (CVPR'24)ViT83.6887.79
Wu et al. [57] (CVPR'24)ViT66.5386.40
Explainable Multi-modal VAD
Zero-Shot CLIP [44]*ViT17.8353.16
LLAVA-1.5 [27]*ViT50.2672.84
LAVAD [67] (CVPR'24)ViT62.0180.28
Holmes-VAU (Ours)ViT87.6888.96
+ +method, we initialize the Multimodal LLM with InternVL2-2B [6]. To optimize the Anomaly-focused Temporal Sampler, we adopt the Adam optimizer with a learning rate of 1e-4. Note that when evaluating detection performance on XD-Violence and UCF-Crime, only videos in the corresponding training sets are used to train our model for fair comparisons. For instruction tuning, we train with a batch size of 512 for 1 epoch, using the AdamW optimizer with cosine learning rate decay and a warm-up period. The LoRA [15] parameters are set as: $r = 64$ , $\alpha = 128$ , and learning rate=4e-5. During testing, $\tau$ in Eq. 2 is set to 0.1. Experiments are conducted on 2 NVIDIA A100 GPUs. + +# 5.2. Main Results + +Anomaly Detection Results. We compare our method with state-of-the-art methods, including semi-supervised methods [14, 52], unsupervised methods [48, 66], weakly-supervised methods [17, 22, 49, 55, 58, 74] and recently training-free method [67]. We have indicated their backbones and performance on the UCF-Crime and XD-Violence datasets, as shown in Table 1. Our method has an AP of $87.68\%$ on XD-Violence and an AUC of $88.96\%$ on UCF-Crime, significantly outperforming the prior state-of-the-art methods, which demonstrates that our method can generate less biased anomaly scores. It is worth noting that while achieving precise localization of anomalies, Holmes-VAU is also capable of providing explanations and analysis for the detected anomalies by the model, a feature unavailable in existing non-explainable VAD methods. Although LAVAD [67] has explainability, the training-free large lan + +Table 2. Comparison of reasoning performance with state-of-the-art Multimodal Large Language Models (MLLMs). 'BLEU' refers to the cumulative values from BLEU-1 to BLEU-4. We evaluate the quality of the generated text at different granularities, including clip-level (C), event-level (E), and video-level (V). + +
MethodParamsBLEU [41](↑)CIDEr [51](↑)METEOR [4](↑)ROUGE [26](↑)
CEVCEVCEVCEV
Video-ChatGPT [39]7B0.1520.0680.0660.0330.0110.0130.1020.0690.0440.1530.0480.079
Video-LLaMA [68]7B0.1510.0790.1040.0240.0140.0170.1120.0760.0570.1560.0670.090
Video-LLaVA [25]7B0.1640.0460.0550.0320.0090.0130.0970.0220.0140.1320.0230.045
LLaVA-Next-Video [71]7B0.4350.0910.1200.1020.0150.0310.1170.0850.0960.1980.0800.106
QwenVL2 [53]7B0.3120.0820.1550.0440.0200.0440.1330.0920.1120.1630.0810.137
InternVL2 [6]8B0.3310.1010.1450.0520.0220.0350.1410.0950.1010.1820.1020.122
Holmes-VAU (Ours)2B0.9130.8040.5660.4671.5191.4370.1900.1650.1210.3290.3700.355
+ +Table 3. Ablation of hierarchical instruction data. During the instruction tuning phase, we combined training data of different granularities, including clip (C), event (E), and video (V) levels. + +
Training DataBLEU(↑)CIDEr(↑)
CEVCEVCEV
✓0.9840.2610.3510.4590.1200.106
✓0.5080.5760.2920.0971.1830.872
✓0.2800.2220.2790.0390.7080.884
✓✓0.8890.7410.3490.4701.2850.889
✓✓0.9060.3410.5220.4720.9621.093
✓✓0.3940.7970.5050.0811.4721.074
✓✓✓0.9130.8040.5660.4671.5191.437
+ +guage model lacks an understanding of anomaly knowledge due to the limitation of insufficient supervised data. + +Anomaly Reasoning Results. We compare the anomaly-related text quality generated by Holmes-VAU with that produced by state-of-the-art general Multimodal Large Language Models (MLLMs), and presented the results at different temporal granularities, including clip-level, event-level, and video-level, as shown in Table 2. Earlier MLLMs such as Video-ChatGPT [21] and Video-LLaMA [68], struggled with basic visual perception and instruction-following capabilities. Recent MLLMs [6, 7, 71] trained on larger and higher-quality video instruction data have made significant progress in general video understanding, with noticeable improvements at the clip-level perception task. However, due to the absence of learning from complex, real-world anomaly data, their reasoning abilities at the event-level and video-level are still lacking. Our Holmes-VAU, however, shows significant improvements in video understanding across all temporal granularities compared to existing general MLLMs, highlighting the importance of injecting anomaly-related knowledge through instruction tuning on high-quality Video Anomaly Understanding benchmarks. + +# 5.3. Analytic Results + +Influence of Hierarchical Instruction. To explore the impact of different granularity video training data on the model's anomaly reasoning ability, we designed various training data combinations during the instruction tuning phase and evaluated the model's performance, as shown + +Table 4. Ablation study of sampling methods and the number of sampled frames. We compare the proposed Anomaly-focused Temporal Sampler (ATS) with other sampling methods under different frame sampling numbers, including Uniform and Top-K. Latency is the time delay in generating the first token. + +
Frames (N)SamplerLatency (ms)Video-level
BLEU(↑)CIDEr(↑)
8Top-K2440.4621.229
Uniform0.4911.276
ATS (Ours)0.5141.324
16Top-K5660.4761.302
Uniform0.5111.345
ATS (Ours)0.5661.437
32Top-K14020.4811.332
Uniform0.5581.357
ATS (Ours)0.5761.460
+ +in Table 3. The inclusion of clip-level data primarily enhanced the model's basic visual perception abilities regarding actions and scenes within the video. Adding event-level data improved the model's ability to judge and understand complete anomaly events. Furthermore, the involvement of video-level data further enhanced the model's ability to analyze and summarize anomaly-related information across longer-span videos. The hierarchical instruction data structure facilitated a comprehensive and complementary improvement in the model's anomaly-related perception-to-reasoning capabilities. + +Influence of different sampling methods and the number of sampled frames. Our ATS (Anomaly-focused Temporal Sampler) is designed to adaptively sample frames input to the LLM based on the frame-level anomaly scores. To validate its advantages, we compared ATS with other sampling methods at various sample frame counts, including Uniform and Top-K sampling. In Uniform sampling, N frames are uniformly sampled from all frames, while Top-K sampling selects the frames with the top $N$ highest anomaly scores. As shown in Table 4, ATS consistently outperforms other sampling methods, regardless of the sample count. We believe that Uniform sampling tends to overlook key anomaly frames, though this issue lessens as more frames are sampled. Besides, Top-K sampling tends to overly focus on local anomaly frames, missing contextual frame informa + +![](images/a3539e036593faaac8034fc67e9939e4aa357ae675c99186e6486495ab7caf89.jpg) +·· + +![](images/81c4fc8f13a8690ef1e239037ca10d2daaa0281908bb60f59303889bbb67ae61.jpg) + +Prompt: Determine whether there are anomaly events in the video and the specific name of the anomaly. + +# Ground Truth + +The anomaly exists and is specifically identified as a Burglary event. + +# Holmes-VAU (Ours) + +Yes, there is an anomaly in the video, specifically a break-in burglary. + +InternVL2-BB [Incorrect Judgement] Yes, there are anomaly events in the video. The specific anomaly is that the man in the video is seen throwing a punch at the other man. + +QwenVL2-7B [Incorrect Judgement] There are no anomaly events in the video. + +![](images/2c898df8c75fe0d6e6131670fcaea798172033a90f1443db56cc54474af81ebd.jpg) + +![](images/00195e2b01988744be8bd79d766165fedfe371e83651df04f37d07ef36992356.jpg) +Bulgary037_x264 + +![](images/cf38f80e93e5239daed104c387edede3cf469af887927484dc5aa55c1baaa202.jpg) + +Prompt: Identify any abnormal events depicted in the video. + +InternVL2-8B [Incorrect Judgement] The video shows a white truck parked on the side of the road, which is an abnormal behavior as it is not supposed to be parked on the road. + +# Ground Truth + +No anomaly exists in the video + +# Holmes-VAU (Ours) + +There is no anomaly in the video. + +# QwenVL2-7B + +The video does not show any abnormal events. + +Normal_Videos_902_x264 + +![](images/c4a55d808cdb7b4123c4a65bfcc83b439904c3a5536425786e201a75908c248d.jpg) + +Prompt: How would you describe the particular anomaly events in the video? What evidence do you have to support your judgment of this event as an anomaly? + +# Ground Truth + +The video anomaly event depicts chaotic street scenes where people are running, police officers are clad in riot gear, and confrontations or protests are underway, with scenes of apprehension, restraint, and property damage unfolding from start to end. The presence of violent behavior, law enforcement intervention, and disruption of public order in the urban environment depicted in the video, which collectively indicate a significant deviation from normal, peaceful civic activity. + +# Holmes-VAU (Ours) + +The anomaly events involve a series of chaotic and violent scenes, including police officers in riot gear, people running and pushing against barriers, and a helicopter flying over the scene, suggesting a high level of tension and unrest. The presence of destructive behavior, including property damage, property destruction, and physical altercations, which deviates from normal and peaceful urban activity. + +# InternVL2-8B [Visual Hallucination, Incomplete Analysis] + +The video shows a group of people walking down a street, and a man is seen walking down the street with a skateboard. The man is wearing a black jacket and a black hat. The event depicted in the video is classified as an anomaly due to the presence of a large group of people engaging in violent behavior. + +# QwenVL2-7B [Incomplete Description, Incomplete Analysis] + +The video shows a group of people walking down the street, and suddenly, a car crashes into a crowd of people. The car is on fire, and the people are running away. The video does not provide any context or explanation for this event, but it is clear that it is an anomaly. The car crash is a violent and unexpected event that is not related to the rest of the video. + +v=3WmMhircZ0c__#1_label_B4-0-0 + +![](images/5f553fb092e3feaf3b37f7f59db441378fa98a7f126f589dc496b44745634815.jpg) +Figure 5. Qualitative comparison of anomaly understanding explanation. Compared with state-of-the-art general MLLMs, i.e., InternVL2 [6] and QwenVL2 [53], our proposed Holmes-VAU demonstrates more accurate anomaly judgment, along with more detailed and comprehensive anomaly-related descriptions and analysis. Correct and wrong explanations are highlighted in green and red, respectively. +Figure 6. Ablation study of trainable parameters. (a) Loss curve during instruction-tuning. (b) We tuned the LoRA [15] parameter $r$ to control trainable parameters, evaluating its impact on VAU capability, and General performance on Video-MME [12]. + +![](images/2e799acaa4580cdb49022072ce0ded081fc1bc24a2b2833a33c72c5c42a4c316.jpg) + +tion. Our proposed ATS mitigates both issues. To balance inference efficiency and performance, we set $N = 16$ as the default sample frame number. + +Instruction Tuning Parameters. We conducted an ablation study on the parameter $r$ in LoRA [15] to explore how the trainable parameters affects both the model's VAU performance and its general capability. We use VideoMME [12] to evaluate the model's general capability. The results are shown in Fig. 6, as $r$ increases, the model gradually adapts to the VAU task. However, when $r$ becomes too large, the model's general capability decreases. To retain the original general video understanding capability of the MLLM, we set $r = 64$ as the default value. + +# 5.4. Qualitative Comparison + +We provide qualitative comparisons between Holmes-VAU and existing MLLMs in Fig. 5. The results demonstrate that Holmes-VAU can accurately identify anomalies in videos and provide accurate and complete explanations, highlight the effectiveness and advantage of Holmes-VAU in perceiving video events and analyzing anomalies. + +# 6. Conclusion + +In conclusion, this work pushes the boundaries of video anomaly understanding by introducing hierarchical anomaly detection across diverse temporal scales, from momentary clips to extended events. The HIVAU-70k benchmark, with over 70,000 multi-level annotations, addresses a critical gap in the field, enabling comprehensive anomaly analysis in real-world scenarios. Our Anomaly-focused Temporal Sampler (ATS) strategically enhances focus on anomaly-dense segments, optimizing both efficiency and accuracy in long-term anomaly detection. Extensive experiments demonstrate that our hierarchical dataset and ATS-enhanced VLM achieve significant performance gains over conventional methods, proving robust for open-world anomaly understanding. This work sets a new standard for multi-granular anomaly comprehension, paving the way for more fine-grained video anomaly understanding. + +# References + +[1] Amit Adam, Ehud Rivlin, Ilan Shimshoni, and Daviv Reinitz. Robust real-time unusual event detection using multiple fixed-location monitors. IEEE transactions on pattern analysis and machine intelligence, 30(3):555-560, 2008. 2 +[2] AI@Meta. Llama 3 model card. 2024. 3 +[3] Kumar Ashutosh, Rohit Girdhar, Lorenzo Torresani, and Kristen Grauman. Hiervl: Learning hierarchical videolanguage embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23066-23078, 2023. 3 +[4] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65-72, 2005. 6, 7 +[5] Yingxian Chen, Zhengzhe Liu, Baoheng Zhang, Wilton Fok, Xiaojuan Qi, and Yik-Chung Wu. Mgfn: Magnitude-contrastive glance-and-focus network for weakly-supervised video anomaly detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 387-395, 2023. 6 +[6] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 2, 5, 6, 7, 8 +[7] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024. 7 +[8] Jisheng Dang, Huicheng Zheng, Xiaohao Xu, Longguang Wang, Qingyong Hu, and Yulan Guo. Adaptive sparse memory networks for efficient and robust video object segmentation. IEEE Transactions on Neural Networks and Learning Systems, 2024. 3 +[9] Hang Du, Sicheng Zhang, Binzhu Xie, Guoshun Nan, Jiayang Zhang, Junrui Xu, Hangyu Liu, Sicong Leng, Jiangming Liu, Hehe Fan, et al. Uncovering what why and how: A comprehensive benchmark for causation understanding of video anomaly. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18793-18803, 2024. 2, 3, 4 +[10] Xinyu Fang, Kangrui Mao, Haodong Duan, Xiangyu Zhao, Yining Li, Dahua Lin, and Kai Chen. Mmbench-video: A long-form multi-shot benchmark for holistic video understanding. arXiv preprint arXiv:2406.14515, 2024. 3 +[11] Jia-Chang Feng, Fa-Ting Hong, and Wei-Shi Zheng. Mist: Multiple instance self-training framework for video anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14009-14018, 2021. 2, 6 + +Acknowledgement This work is supported by the National Natural Science Foundation of China under grants U22B2053 and 623B2039, and in part by the Interdisciplinary Research Program of HUST (2024JCYJ034). +[12] Chaoyou Fu, Yuhan Dai, Yondong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 3, 8 +[13] Dong Gong, Lingqiao Liu, Vuong Le, Budhaditya Saha, Moussa Reda Mansour, Svetha Venkatesh, and Anton van den Hengel. Memorizing normality to detect anomaly: Memory-augmented deep autoencoder for unsupervised anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1705-1714, 2019. 2 +[14] Mahmudul Hasan, Jonghyun Choi, Jan Neumann, Amit K Roy-Chowdhury, and Larry S Davis. Learning temporal regularity in video sequences. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 733-742, 2016. 2, 6 +[15] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 6, 8 +[16] Md Mohaiminul Islam, Ngan Ho, Xitong Yang, Tushar Nagarajan, Lorenzo Torresani, and Gedas Bertasius. Video recap: Recursive captioning of hour-long videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18198-18208, 2024. 3 +[17] Hyekang Kevin Joo, Khoa Vo, Kashu Yamazaki, and Ngan Le. Clip-tsa: Clip-assisted temporal self-attention for weakly-supervised video anomaly detection. In 2023 IEEE International Conference on Image Processing (ICIP), pages 3230–3234. IEEE, 2023. 3, 6 +[18] Kumara Kahatapitiya and Michael S Ryoo. Coarse-fine networks for temporal activity detection in videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8385-8394, 2021. 3 +[19] Jaechul Kim and Kristen Grauman. Observe locally, infer globally: a space-time mrf for detecting abnormal activities with incremental updates. In 2009 IEEE conference on computer vision and pattern recognition, pages 2921-2928. IEEE, 2009. 2 +[20] Federico Landi, Cees GM Snoek, and Rita Cucchiara. Anomaly locality in video surveillance. arXiv preprint arXiv:1901.10364, 2019. 2 +[21] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 2, 4, 7 +[22] Shuo Li, Fang Liu, and Licheng Jiao. Self-training multisession learning with transformer for weakly supervised video anomaly detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1395-1403, 2022. 2, 6 +[23] Weixin Li, Vijay Mahadevan, and Nuno Vasconcelos. Anomaly detection and localization in crowded scenes. IEEE transactions on pattern analysis and machine intelligence, 36(1):18-32, 2013. 2 +[24] Yingwei Li, Yi Li, and Nuno Vasconcelos. Resound: Towards action recognition without representation bias. In Pro + +ceedings of the European Conference on Computer Vision (ECCV), pages 513-528, 2018. 3 +[25] Bin Lin, Bin Zhu, Yang Ye, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 2, 4, 7 +[26] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 6, 7 +[27] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 2, 4, 6 +[28] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024.3 +[29] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2, 4 +[30] Kun Liu and Huadong Ma. Exploring background-bias for anomaly detection in surveillance videos. In Proceedings of the 27th ACM International Conference on Multimedia, pages 1490-1499, 2019. 2 +[31] Wen Liu, Weixin Luo, Dongze Lian, and Shenghua Gao. Future frame prediction for anomaly detection-a new baseline. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6536-6545, 2018. 2 +[32] Yi Liu, Limin Wang, Yali Wang, Xiao Ma, and Yu Qiao. Fineaction: A fine-grained video dataset for temporal action localization. IEEE transactions on image processing, 31: 6937-6950, 2022. 3 +[33] Zhian Liu, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. A hybrid video anomaly detection framework via memory-augmented flow reconstruction and flow-guided frame prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 13588-13597, 2021. 2 +[34] Cewu Lu, Jianping Shi, and Jiaya Jia. Abnormal event detection at 150 fps in matlab. In Proceedings of the IEEE international conference on computer vision, pages 2720-2727, 2013. 2 +[35] Jian Lu, Xuanfeng Li, Bo Zhao, and Jian Zhou. A review of skeleton-based human action recognition. Journal of Image and Graphics, 28(12):3651-3669, 2023. 3 +[36] Hui Lv and Qianru Sun. Video anomaly detection and explanation via large language models. arXiv preprint arXiv:2401.05702, 2024. 2 +[37] Hui Lv, Zhongqi Yue, Qianru Sun, Bin Luo, Zhen Cui, and Hanwang Zhang. Unbiased multiple instance learning for weakly supervised video anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8022-8031, 2023. 2 +[38] Baiteng Ma, Shiwei Zhang, Changxin Gao, and Nong Sang. Temporal global correlation network for end-to-end action proposal generation. Acta Electronica Sinica, 50(10):2452-2461, 2022. 3 +[39] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video + +understanding via large vision and language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (ACL 2024), 2024. 7 +[40] Ramin Mehran, Alexis Oyama, and Mubarak Shah. Abnormal crowd behavior detection using social force model. In 2009 IEEE conference on computer vision and pattern recognition, pages 935-942. IEEE, 2009. 2 +[41] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311–318, 2002. 6, 7 +[42] Yujiang Pu, Xiaoyu Wu, Lulu Yang, and Shengjin Wang. Learning prompt-enhanced context features for weakly-supervised video anomaly detection. IEEE Transactions on Image Processing, 2024. 2, 3 +[43] Zhiwu Qing, Shiwei Zhang, Ziyuan Huang, Yi Xu, Xiang Wang, Mingqian Tang, Changxin Gao, Rong Jin, and Nong Sang. Learning from untrimmed videos: Self-supervised video representation learning with hierarchical consistency. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13821-13831, 2022. 3 +[44] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 5, 6 +[45] Dian Shao, Yue Zhao, Bo Dai, and Dahua Lin. Finegym: A hierarchical video dataset for fine-grained action understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2616-2625, 2020. 3 +[46] Waqas Sultani, Chen Chen, and Mubarak Shah. Real-world anomaly detection in surveillance videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6479-6488, 2018. 2, 3, 6 +[47] Jiaqi Tang, Hao Lu, Ruizheng Wu, Xiaogang Xu, Ke Ma, Cheng Fang, Bin Guo, Jiangbo Lu, Qifeng Chen, and YingCong Chen. Hawk: Learning to understand open-world video anomalies. arXiv preprint arXiv:2405.16886, 2024. 2,3,4 +[48] Kamalakar Vijay Thakare, Yash Raghuwanshi, Debi Prosad Dogra, Heeseung Choi, and Ig-Jae Kim. Dyannet: A scene dynamicity guided self-trained video anomaly detection network. In Proceedings of the IEEE/CVF Winter conference on applications of computer vision, pages 5541-5550, 2023. 2, 6 +[49] Yu Tian, Guansong Pang, Yuanhong Chen, Rajvinder Singh, Johan W Verjans, and Gustavo Carneiro. Weakly-supervised video anomaly detection with robust temporal feature magnitude learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4975-4986, 2021. 2, 5, 6 +[50] Anil Osman Tur, Nicola Dall'Asen, Cigdem Bayan, and Elisa Ricci. Exploring diffusion models for unsupervised + +video anomaly detection. In 2023 IEEE International Conference on Image Processing (ICIP), pages 2540-2544. IEEE, 2023. 2 +[51] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4566-4575, 2015. 6, 7 +[52] Jue Wang and Anoop Cherian. Gods: Generalized one-class discriminative subspaces for anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8201-8211, 2019. 2, 6 +[53] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 7, 8 +[54] SC Wang, Q Huang, YF Zhang, X Li, YQ Nie, and GC Luo. Review of action recognition based on multimodal data. Image Graph, 27(11):3139-3159, 2022. 3 +[55] Jhih-Ciang Wu, He-Yen Hsieh, Ding-Jie Chen, Chiou-Shann Fuh, and Tyng-Luh Liu. Self-supervised sparse representation for video anomaly detection. In European Conference on Computer Vision, pages 729-745. Springer, 2022. 2, 5, 6 +[56] Peng Wu, Jing Liu, Yujia Shi, Yujia Sun, Fangtao Shao, Zhaoyang Wu, and Zhiwei Yang. Not only look, but also listen: Learning multimodal violence detection under weak supervision. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXX 16, pages 322-339. Springer, 2020. 2, 3, 6 +[57] Peng Wu, Xuerong Zhou, Guansong Pang, Yujia Sun, Jing Liu, Peng Wang, and Yanning Zhang. Open-vocabulary video anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18297-18307, 2024. 2, 3, 6 +[58] Peng Wu, Xuerong Zhou, Guansong Pang, Lingru Zhou, Qingsen Yan, Peng Wang, and Yanning Zhang. Vadclip: Adapting vision-language models for weakly supervised video anomaly detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6074-6082, 2024. 2, 3, 5, 6 +[59] Dan Xu, Yan Yan, Elisa Ricci, and Nicu Sebe. Detecting anomalous events in videos by learning deep representations of appearance and motion. Computer Vision and Image Understanding, 156:117-127, 2017. 2 +[60] Xiaohao Xu, Jinglu Wang, Xiang Ming, and Yan Lu. Towards robust video object segmentation with adaptive object calibration. In Proceedings of the 30th ACM International Conference on Multimedia, pages 2709-2718, 2022. 3 +[61] Zhiwei Yang, Jing Liu, Zhaoyang Wu, Peng Wu, and Xiaotao Liu. Video event restoration based on keyframes for video anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14592-14601, 2023. 2 +[62] Zhiwei Yang, Jing Liu, and Peng Wu. Text prompt with normality guidance for weakly supervised video anomaly detec + +tion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18899-18908, 2024. 2, 3, 6 +[63] Yu Yao, Xizi Wang, Mingze Xu, Zelin Pu, Yuchen Wang, Ella Atkins, and David J Crandall. Dota: Unsupervised detection of traffic anomaly in driving videos. IEEE transactions on pattern analysis and machine intelligence, 45(1): 444-459, 2022. 2 +[64] Tongtong Yuan, Xuange Zhang, Kun Liu, Bo Liu, Chen Chen, Jian Jin, and Zhenzhen Jiao. Towards surveillance video-and-language understanding: New dataset, baselines, and challenges, 2023. 3 +[65] Tongtong Yuan, Xuange Zhang, Kun Liu, Bo Liu, Chen Chen, Jian Jin, and Zhenzhen Jiao. Towards surveillance video-and-language understanding: New dataset baselines and challenges. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22052-22061, 2024. 2, 3 +[66] M Zaigham Zaheer, Arif Mahmood, M Haris Khan, Mattia Segu, Fisher Yu, and Seung-Ik Lee. Generative cooperative learning for unsupervised video anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14744-14754, 2022. 2, 6 +[67] Luca Zanella, Willi Menapace, Massimiliano Mancini, Yiming Wang, and Elisa Ricci. Harnessing large language models for training-free video anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18527-18536, 2024. 2, 3, 4, 5, 6 +[68] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 2, 7 +[69] Huaxin Zhang, Xiang Wang, Xiaohao Xu, Zhiwu Qing, Changxin Gao, and Nong Sang. Hr-pro: Point-supervised temporal action localization via hierarchical reliability propagation. arXiv preprint arXiv:2308.12608, 2023. 3 +[70] Huaxin Zhang, Xiang Wang, Xiaohao Xu, Xiaonan Huang, Chuchu Han, Yuehuan Wang, Changxin Gao, Shanjun Zhang, and Nong Sang. Glancevad: Exploring glance supervision for label-efficient video anomaly detection. arXiv preprint arXiv:2403.06154, 2024. 2 +[71] Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, and Chunyuan Li. Video instruction tuning with synthetic data. arXiv preprint arXiv:2410.02713, 2024. 7 +[72] Bin Zhao, Li Fei-Fei, and Eric P Xing. Online detection of unusual events in videos via dynamic sparse coding. In CVPR 2011, pages 3313-3320. IEEE, 2011. 2 +[73] Jia-Xing Zhong, Nannan Li, Weijie Kong, Shan Liu, Thomas H Li, and Ge Li. Graph convolutional label noise cleaner: Train a plug-and-play action classifier for anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1237-1246, 2019. 2 +[74] Hang Zhou, Junqing Yu, and Wei Yang. Dual memory units with uncertainty regulation for weakly supervised video anomaly detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3769-3777, 2023. 2, 5, 6 \ No newline at end of file diff --git a/2025/Holmes-VAU_ Towards Long-term Video Anomaly Understanding at Any Granularity/images.zip b/2025/Holmes-VAU_ Towards Long-term Video Anomaly Understanding at Any Granularity/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..82be4e538a84330560086c5d35487b674d63cad5 --- /dev/null +++ b/2025/Holmes-VAU_ Towards Long-term Video Anomaly Understanding at Any Granularity/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d85b282188ee1f28eeedcf54045f4f1bf11d007c8e0f33321555bbf87b8bd10a +size 603570 diff --git a/2025/Holmes-VAU_ Towards Long-term Video Anomaly Understanding at Any Granularity/layout.json b/2025/Holmes-VAU_ Towards Long-term Video Anomaly Understanding at Any Granularity/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b2d203f36e841f612c37e90a388d6edb81eccfca --- /dev/null +++ b/2025/Holmes-VAU_ Towards Long-term Video Anomaly Understanding at Any Granularity/layout.json @@ -0,0 +1,9604 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 102, + 103, + 509, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 103, + 509, + 140 + ], + "spans": [ + { + "bbox": [ + 102, + 103, + 509, + 140 + ], + "type": "text", + "content": "Holmes-VAU: Towards Long-term Video Anomaly Understanding at Any Granularity" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 78, + 160, + 529, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 160, + 529, + 232 + ], + "spans": [ + { + "bbox": [ + 78, + 160, + 529, + 232 + ], + "type": "text", + "content": "Huaxin Zhang" + }, + { + "bbox": [ + 78, + 160, + 529, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 78, + 160, + 529, + 232 + ], + "type": "text", + "content": " Xiaohao Xu" + }, + { + "bbox": [ + 78, + 160, + 529, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 78, + 160, + 529, + 232 + ], + "type": "text", + "content": " Xiang Wang" + }, + { + "bbox": [ + 78, + 160, + 529, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 78, + 160, + 529, + 232 + ], + "type": "text", + "content": " Jialong Zuo" + }, + { + "bbox": [ + 78, + 160, + 529, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 78, + 160, + 529, + 232 + ], + "type": "text", + "content": " Xiaonan Huang" + }, + { + "bbox": [ + 78, + 160, + 529, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 78, + 160, + 529, + 232 + ], + "type": "text", + "content": " Changxin Gao" + }, + { + "bbox": [ + 78, + 160, + 529, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 78, + 160, + 529, + 232 + ], + "type": "text", + "content": " Shanjun Zhang" + }, + { + "bbox": [ + 78, + 160, + 529, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 78, + 160, + 529, + 232 + ], + "type": "text", + "content": " Li Yu" + }, + { + "bbox": [ + 78, + 160, + 529, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 78, + 160, + 529, + 232 + ], + "type": "text", + "content": " Nong Sang" + }, + { + "bbox": [ + 78, + 160, + 529, + 232 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 78, + 160, + 529, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 78, + 160, + 529, + 232 + ], + "type": "text", + "content": "Key Laboratory of Image Processing and Intelligent Control, School of Artificial Intelligence and Automation, Huazhong University of Science and Technology \n" + }, + { + "bbox": [ + 78, + 160, + 529, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 78, + 160, + 529, + 232 + ], + "type": "text", + "content": "University of Michigan, Ann Arbor \n" + }, + { + "bbox": [ + 78, + 160, + 529, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 78, + 160, + 529, + 232 + ], + "type": "text", + "content": "Kanagawa University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 72, + 235, + 538, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 235, + 538, + 245 + ], + "spans": [ + { + "bbox": [ + 72, + 235, + 538, + 245 + ], + "type": "text", + "content": "{zhanghuaxin,wxiang,cgao,hustlyu,nsang}@hust.edu.cn,{xiaohaox,xiaonanh}@umich.edu,{chiyoz01}@kanagawa-u.ac.jp" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 59, + 251, + 403, + 441 + ], + "blocks": [ + { + "bbox": [ + 59, + 251, + 403, + 441 + ], + "lines": [ + { + "bbox": [ + 59, + 251, + 403, + 441 + ], + "spans": [ + { + "bbox": [ + 59, + 251, + 403, + 441 + ], + "type": "image", + "image_path": "91680e9686dac8a1c7c39831e48df448d8514ddd8efc7f8e41225746cb194df8.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 445, + 555, + 501 + ], + "lines": [ + { + "bbox": [ + 54, + 445, + 555, + 501 + ], + "spans": [ + { + "bbox": [ + 54, + 445, + 555, + 501 + ], + "type": "text", + "content": "Figure 1. Motivation. Left: Existing datasets lack the hierarchical structure to capture transient and sustained anomalies across varying temporal scales. Our HIVAU-70k dataset addresses this by providing multi-granularity annotations—clip, event, and video levels—that enable detailed anomaly analysis in complex real-world scenarios. Right: Inspired by Sherlock Holmes's knack for zeroing in on critical details, our Holmes-VAU method integrates an Anomaly-focused Temporal Sampler with a multi-modal LLM, directing model attention to anomaly-rich segments, which enables models to decode complex, long-term video anomalies efficiently." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 417, + 253, + 552, + 441 + ], + "blocks": [ + { + "bbox": [ + 417, + 253, + 552, + 441 + ], + "lines": [ + { + "bbox": [ + 417, + 253, + 552, + 441 + ], + "spans": [ + { + "bbox": [ + 417, + 253, + 552, + 441 + ], + "type": "image", + "image_path": "20f4aec5a3fd4a30388dd8ab4138d26c5614df3c98528ba32bf241494842a176.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 510, + 200, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 510, + 200, + 522 + ], + "spans": [ + { + "bbox": [ + 151, + 510, + 200, + 522 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 536, + 297, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 536, + 297, + 693 + ], + "spans": [ + { + "bbox": [ + 53, + 536, + 297, + 693 + ], + "type": "text", + "content": "How can we enable models to comprehend video anomalies occurring over varying temporal scales and contexts? Traditional Video Anomaly Understanding (VAU) methods focus on frame-level anomaly prediction, often missing the interpretability of complex and diverse real-world anomalies. Recent multimodal approaches leverage visual and textual data but lack hierarchical annotations that capture both short-term and long-term anomalies. To address this challenge, we introduce HIVAU-70k, a large-scale benchmark for hierarchical video anomaly understanding across any granularity. We develop a semi-automated annotation engine that efficiently scales high-quality annotations by combining manual video segmentation with recursive free-text" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 510, + 556, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 691 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 691 + ], + "type": "text", + "content": "annotation using large language models (LLMs). This results in over 70,000 multi-granular annotations organized at clip-level, event-level, and video-level segments. For efficient anomaly detection in long videos, we propose the Anomaly-focused Temporal Sampler (ATS). ATS integrates an anomaly scorer with a density-aware sampler to adaptively select frames based on anomaly scores, ensuring that the multimodal LLM concentrates on anomaly-rich regions, which significantly enhances both efficiency and accuracy. Extensive experiments demonstrate that our hierarchical instruction data markedly improves anomaly comprehension. The integrated ATS and visual-language model outperform traditional methods in processing long videos. Our benchmark and model are publicly available at https://github.com/pipixin321/HolmesVAU." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13843" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 136, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 136, + 84 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 136, + 84 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 294, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 294, + 199 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 294, + 199 + ], + "type": "text", + "content": "Video Anomaly Understanding (VAU) is crucial for applications such as video surveillance [46], violent content analysis [56], and autonomous driving [63]. Detecting deviations from normal patterns aids in hazard prevention and real-time decision-making. Traditional methods [14, 49, 74] mainly focus on frame-level predefined closed-set anomaly prediction, assigning an anomaly score to each frame. However, these approaches often fail to describe and understand complex anomalies in the real world." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 199, + 295, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 199, + 295, + 294 + ], + "spans": [ + { + "bbox": [ + 55, + 199, + 295, + 294 + ], + "type": "text", + "content": "To address this gap, open-world anomaly understanding [57] embraces the diversity and unpredictability of real-world anomalies. Recent work integrates multimodal approaches, combining visual data with textual descriptions [42, 58, 62, 65], while advances in multimodal visual-language models (VLMs) [6, 21, 27, 29, 68] have enabled more nuanced understanding through anomaly-related instruction tuning and text generation [9, 36, 47, 67]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 295, + 295, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 295, + 295, + 486 + ], + "spans": [ + { + "bbox": [ + 55, + 295, + 295, + 486 + ], + "type": "text", + "content": "Despite these advancements, a significant gap remains in models' ability to comprehend anomalies across multiple temporal scales. For instance, while anomalies such as explosions or fights may be captured in a single frame, more complex events like theft or arson require understanding long-term contextual patterns. Existing VAU datasets [46, 56] typically provide annotations at a single level of granularity, limiting models to understanding either immediate perceptual anomalies or those requiring extended contextual reasoning. The lack of datasets with hierarchical annotations—encompassing both short-term and long-term anomalies—hinders models' capacity to reason about anomalies with diverse temporal characteristics. Moreover, constructing datasets that encapsulate this hierarchical complexity poses significant challenges in scalability and annotation quality." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 487, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 295, + 665 + ], + "type": "text", + "content": "To address these issues, we develop a semi-automated annotation engine that efficiently scales high-quality annotation by combining manual video segmentation with recursive free-text annotation using large language models (LLMs). The process involves three key stages: 1) hierarchical video decoupling, where we manually identify anomaly events and segment them into shorter clips; 2) hierarchical free-text annotation, where captions for each clip are generated through human effort or video captioning models, then summarized at the event and video levels via LLMs; and 3) hierarchical instruction construction, where the textual data is transformed into question-answer instruction prompts by combining captions and summaries with designed prompts, creating a dataset with rich annotations for training and evaluating models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "content": "Utilizing the annotation engine, we introduce HIVAU-70k, a large-scale video anomaly understanding benchmark with hierarchical instructions. Our dataset comprises over 70,000 multi-granular instruction data organized" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "across clip-level, event-level, and video-level segments as shown in Fig. 1. This hierarchical structure empowers models to detect immediate anomalies, e.g., sudden explosions or fighting, as well as complex events that require an understanding of long-term context, like theft or arson. By annotating at multiple temporal levels, HIVAU-70k provides diverse anomalies in open-world scenarios." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 157, + 553, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 157, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 553, + 312 + ], + "type": "text", + "content": "Towards long-term VAU, efficiency remains a critical challenge. Previous methods [9, 47, 67] often rely on uniform frame sampling, which can either miss crucial anomaly frames or incur large computational costs [25, 47, 67]. To address this, we propose the Holmes-VAU method, which combines the proposed Anomaly-focused Temporal Sampler (ATS) with the multimodal visual-language model for efficient long-term video anomaly understanding (See Fig. 1). The ATS combines a anomaly scorer with a density-aware sampler that adaptively selects frames by their anomaly scores. This integration ensures that the visual-language model concentrates on anomaly-rich regions, enhancing both efficiency and accuracy." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 314, + 553, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 314, + 553, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 314, + 553, + 445 + ], + "type": "text", + "content": "Our contributions are threefold: 1) We introduce HIVAU-70k, a large-scale, multi-granular benchmark for hierarchical video anomaly understanding. 2) We propose the Holmes-VAU method, which combines the proposed Anomaly-focused Temporal Sampler (ATS) to boost the efficiency and accuracy of inference on long-term videos. 3) We conduct extensive experiments demonstrating the effectiveness of hierarchical instruction data in enhancing anomaly comprehension and validate the performance gains provided by the integrated ATS and visual-language model in processing long videos." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 460, + 406, + 473 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 406, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 406, + 473 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 482, + 553, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 482, + 553, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 482, + 553, + 685 + ], + "type": "text", + "content": "Video Anomaly Detection. This task aims to temporally detect abnormal frames in a long untrimmed video [1, 14, 23, 34, 40, 52]. Early attempts are based on handcrafted features [1, 19, 23, 34, 40, 72]. Recently, deep learning approaches [14, 37, 61] have become dominant, broadly classified into unsupervised, weakly-supervised, and fully-supervised methods. Unsupervised methods train only on normal videos to learn normal patterns and are often designed as reconstruction-based [13, 14, 59, 61], prediction-based [31], or a combination [33]. Some methods [48, 50, 66] also explore a fully unsupervised setting, including both normal and abnormal videos in training set without real labels. Weakly-supervised methods [11, 22, 37, 46, 49, 55, 56, 70, 73, 74] use both normal and abnormal videos with video-level annotations. Fully-supervised methods [20, 30] are less studied due to the high cost of precise frame-level annotations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": "Multi-modal Video Anomaly Understanding. Large-scale visual-language pre-trained models such as CLIP [44]" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13844" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 263 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 263 + ], + "type": "text", + "content": "serve as a bridge between visual and textual modalities. Some recent works [17, 42, 58, 62] in the realm of video anomaly detection have leveraged textual information as prompts to enhance the model's anomaly representation. Based on this, [57] firstly proposed the open vocabulary VAD task, [65] introduced a multimodal video anomaly dataset composed of dense clip captions. Furthermore, [67] extracted captions from video frames and designed prompts for LLMs to provide anomaly scores, [9] and [47] construct diverse and interactive instruction data at the video level. However, these datasets consider only a single temporal level of anomaly understanding data construction, i.e. clip-level [65] or video-level [9, 47]. Unlike these methods, we focus on building large-scale hierarchical video anomaly understanding data for multimodal instruction tuning." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 267, + 296, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 267, + 296, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 267, + 296, + 460 + ], + "type": "text", + "content": "Hierarchical Video Understanding. Video understanding is a challenging task due to its temporal-scale diversity. To better comprehend videos, many previous works have focused on both datasets and models in hierarchical video understanding. For example, [3, 24, 32, 35, 45, 54] proposed fine-grained action recognition and localization datasets, [16] provided free-form hierarchical captions for hour-long videos at multiple temporal scales, and [8, 18, 38, 43, 60, 69] trained models on hierarchical levels to obtain better video feature representation. Recently, to assess the capability of video vision-language models (VLMs) in handling challenges in real-world scenarios, several video benchmarks [10, 12] have been built, incorporating data at multiple temporal scales for evaluation. Unlike these works, we dive into the field of Video Anomaly Understanding, thus filling the gap in multi-scale annotations in this area." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 472, + 194, + 484 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 194, + 484 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 194, + 484 + ], + "type": "text", + "content": "3. HIVAU-70k Benchmark" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 493, + 296, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 493, + 296, + 541 + ], + "spans": [ + { + "bbox": [ + 55, + 493, + 296, + 541 + ], + "type": "text", + "content": "We first define the video anomaly understanding task in Sec.3.1. Then, the construction process of the HIVAU-70k benchmark will be elaborated in Sec.3.2. Finally, we will present HIVAU-70k's statistical information in Sec. 3.3." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 551, + 157, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 551, + 157, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 551, + 157, + 563 + ], + "type": "text", + "content": "3.1. Task Description" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": "This work focuses on video anomaly understanding, involving temporal anomaly detection and anomaly explainability. Temporal anomaly detection aims to predict an anomaly score for each frame in a video " + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": ", represented as " + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "S\\in \\mathbb{R}^T" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": " is the total number of frames. Building on this, we explore the model's ability to generate explanatory text outputs related to anomalies based on video input and user queries. Specifically, given a video " + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": " and a text query " + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": ", the model generates an anomaly-related response " + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": ". We consider two key abilities: 1) visual perception, which involves recognizing main entities in the video, and 2) anomaly reasoning, which encompasses the" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 318, + 74, + 553, + 257 + ], + "blocks": [ + { + "bbox": [ + 318, + 74, + 553, + 257 + ], + "lines": [ + { + "bbox": [ + 318, + 74, + 553, + 257 + ], + "spans": [ + { + "bbox": [ + 318, + 74, + 553, + 257 + ], + "type": "image", + "image_path": "716735a97f013fd3347cdca1ef45bb25a82e5313a822f0e352fb671715079360.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 261, + 555, + 361 + ], + "lines": [ + { + "bbox": [ + 313, + 261, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 261, + 555, + 361 + ], + "type": "text", + "content": "Figure 2. Data Engine. We present a structured workflow for generating hierarchical annotations across video, event, and clip levels. Clips are first captioned, then processed through a large language model (LLM) with prompts for event summarization. The outputs include clip captions, event summaries, and video summaries, followed by manual checking and refinement. This multi-step approach enriches the dataset with detailed judgments, descriptions, and analyses of anomalies, enabling robust contextual understanding at varying granularities." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 376, + 537, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 537, + 388 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 537, + 388 + ], + "type": "text", + "content": "model's judgment and analysis of the anomaly content." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 396, + 484, + 408 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 484, + 408 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 484, + 408 + ], + "type": "text", + "content": "3.2. LLM-Empowered Data Engine" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 414, + 554, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 414, + 554, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 414, + 554, + 473 + ], + "type": "text", + "content": "As shown in Fig.2, we develop a semi-automated annotation engine that efficiently scales high-quality annotations, which consists of three main steps: 1) hierarchical video decoupling, 2) hierarchical free-text annotation, and 3) hierarchical instruction construction." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 474, + 554, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 554, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 554, + 629 + ], + "type": "text", + "content": "Hierarchical Video Decoupling. Our video sources include the training set of the UCF-Crime [46] dataset and the XD-Violence [56] dataset, which contains videos of varying durations and diverse real-world anomalies. For abnormal videos, we first manually obtain the temporal boundaries of each anomaly event in the video. Non-continuous anomalous states are considered separate events. Then, we divide each event into clips of random lengths. For normal videos, we apply random sampling to obtain corresponding segments of varying granularities. Ultimately, we obtained 5,443 videos, 11,076 events, and 55,806 clips. This process took 5 annotators approximately 20 hours to complete. More details can be found in the appendix." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": "Hierarchical Free-text Annotation. To fully extract semantic information from the clip-level videos, we utilize a powerful off-the-shell video perception model LLaVA-Next-Video [28] to generate detailed captions for each clip. We also include the UCA dataset [64], which provides manually annotated captions for video clips in the UCF-Crime [46] dataset. Then, we use an LLM [2] to consoli" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13845" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 73, + 221, + 159 + ], + "blocks": [ + { + "bbox": [ + 59, + 73, + 221, + 159 + ], + "lines": [ + { + "bbox": [ + 59, + 73, + 221, + 159 + ], + "spans": [ + { + "bbox": [ + 59, + 73, + 221, + 159 + ], + "type": "image", + "image_path": "b5288291184c2c45051df10e1fc4281661cd7533f284234b3570a9825728988e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 163, + 211, + 178 + ], + "lines": [ + { + "bbox": [ + 64, + 163, + 211, + 178 + ], + "spans": [ + { + "bbox": [ + 64, + 163, + 211, + 178 + ], + "type": "text", + "content": "Q: Give a short and clear caption of the video clip." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 230, + 72, + 386, + 159 + ], + "blocks": [ + { + "bbox": [ + 230, + 72, + 386, + 159 + ], + "lines": [ + { + "bbox": [ + 230, + 72, + 386, + 159 + ], + "spans": [ + { + "bbox": [ + 230, + 72, + 386, + 159 + ], + "type": "image", + "image_path": "afe828ef6ae55dc7a7bca169325cd44c0a08c5ac8c2dabb99be098c56c997894.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 389, + 72, + 551, + 159 + ], + "blocks": [ + { + "bbox": [ + 389, + 72, + 551, + 159 + ], + "lines": [ + { + "bbox": [ + 389, + 72, + 551, + 159 + ], + "spans": [ + { + "bbox": [ + 389, + 72, + 551, + 159 + ], + "type": "image", + "image_path": "fb2b43b8e06015e5e6f484b9c008b3e8e3a78acc326bb6e6148738050b275d0d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 228, + 164, + 385, + 252 + ], + "blocks": [ + { + "bbox": [ + 64, + 184, + 208, + 198 + ], + "lines": [ + { + "bbox": [ + 64, + 184, + 208, + 198 + ], + "spans": [ + { + "bbox": [ + 64, + 184, + 208, + 198 + ], + "type": "text", + "content": "A: A group of policemen dragged a man red out of the gate." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 65, + 209, + 211, + 222 + ], + "lines": [ + { + "bbox": [ + 65, + 209, + 211, + 222 + ], + "spans": [ + { + "bbox": [ + 65, + 209, + 211, + 222 + ], + "type": "text", + "content": "Q: Are there any anomaly events detected in the video?" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 65, + 226, + 199, + 240 + ], + "lines": [ + { + "bbox": [ + 65, + 226, + 199, + 240 + ], + "spans": [ + { + "bbox": [ + 65, + 226, + 199, + 240 + ], + "type": "text", + "content": "A: An anomaly exists, specifically an Arrest anomaly." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 228, + 164, + 385, + 252 + ], + "lines": [ + { + "bbox": [ + 228, + 164, + 385, + 252 + ], + "spans": [ + { + "bbox": [ + 228, + 164, + 385, + 252 + ], + "type": "image", + "image_path": "5fa90769cc6e549f8ba56fbcbed5274d7753f89b8efe0b2300fe0d9a028f65b5.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 389, + 165, + 547, + 201 + ], + "lines": [ + { + "bbox": [ + 389, + 165, + 547, + 201 + ], + "spans": [ + { + "bbox": [ + 389, + 165, + 547, + 201 + ], + "type": "text", + "content": "Q: Describe the anomaly events in the video. A: The anomaly event involves the three policemen turning the man in red over and moving him to the wall, indicating a possible arrest or detention." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 389, + 208, + 547, + 246 + ], + "lines": [ + { + "bbox": [ + 389, + 208, + 547, + 246 + ], + "spans": [ + { + "bbox": [ + 389, + 208, + 547, + 246 + ], + "type": "text", + "content": "Q: Why is this event as an anomaly? A: The unusual and potentially forceful interaction between the policemen and the man in red, deviating from normal pedestrian behavior in the scene" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 257, + 555, + 302 + ], + "lines": [ + { + "bbox": [ + 55, + 257, + 555, + 302 + ], + "spans": [ + { + "bbox": [ + 55, + 257, + 555, + 302 + ], + "type": "text", + "content": "Figure 3. HVAU-70k dataset. (a) Duration distributions for clips, events, and full videos, showing dominance of short clips. (b) Hierarchical data organization from clip-level to video-level, enabling perception-to-reasoning insights. (c) Word count variations across annotation levels, with more detailed descriptions at the video level. (d) Sample annotations capturing captioning, judgment, description, and anomaly analysis, highlighting nuanced understanding of anomaly events in complex scenes." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 317, + 296, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 317, + 296, + 496 + ], + "spans": [ + { + "bbox": [ + 54, + 317, + 296, + 496 + ], + "type": "text", + "content": "date all clip captions within an event, generating an event-level video summary. Specifically, we design prompts1 to guide the LLM to produce three parts of content for each anomalous event summary: 1) Judgment: A determination of whether an anomaly exists and its specific category, 2) Description: A detailed description of the anomalous or normal event, 3) Analysis: The reasoning behind the anomaly judgment, including causal analysis. To guide the LLM in generating reliable responses, we also inject the event's category label (e.g., \"Shooting\", \"Explosion\") into the LLM prompt. We consolidate all event-level summaries to obtain the video-level summary. This results in free-text annotations across multiple temporal scales, including short-term visual perception (clip-level) and long-term anomaly reasoning (event-level, video-level)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 497, + 296, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 497, + 296, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 497, + 296, + 545 + ], + "type": "text", + "content": "Hierarchical Instruction Data Construction. The ability of VLMs to follow user instructions and generate responses is achieved through instruction-tuning [21, 25, 27, 29]. The training data format typically consists of the following:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 67, + 546, + 241, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 546, + 241, + 559 + ], + "spans": [ + { + "bbox": [ + 67, + 546, + 241, + 559 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{Q}" + }, + { + "bbox": [ + 67, + 546, + 241, + 559 + ], + "type": "text", + "content": " :user instruction, " + }, + { + "bbox": [ + 67, + 546, + 241, + 559 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 67, + 546, + 241, + 559 + ], + "type": "text", + "content": " :model response.}" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 559, + 296, + 692 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 559, + 296, + 692 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 296, + 692 + ], + "type": "text", + "content": "To build instruction-tuning data for VLMs in the domain of Video Anomaly Understanding, we matched free-text annotations with pre-designed anomaly-related user instructions. Specifically, for clip-level segments, we only construct instructions related to captions, as it is challenging to obtain anomaly-related analysis for short videos. For event-level and video-level segments, we construct instruction data from the perspectives of Judgment, Description, and Analysis. Typical examples are shown in the Fig. 3 (d). Manual Checking for Data Quality Control. To ensure dataset quality, we implemented several human inspec" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 317, + 555, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 317, + 555, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 317, + 555, + 389 + ], + "type": "text", + "content": "tion and curation strategies. First, we labeled the temporal boundaries of abnormal event segments and reviewed them at the second level. Next, anomaly labels were incorporated during summary generation using LLMs, directing focus on relevant entities. Finally, manual reviews were performed to correct low-quality instruction data." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 398, + 470, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 398, + 470, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 398, + 470, + 410 + ], + "type": "text", + "content": "3.3. Data Statistic of HIVAU-70k" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 417, + 555, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 417, + 555, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 417, + 555, + 586 + ], + "type": "text", + "content": "Utilizing the proposed annotation engine, we introduce HIVAU-70k, as shown in Fig. 3, a large-scale benchmark designed for hierarchical instruction-based video anomaly understanding. As shown in Fig. 3(b), HIVAU-70k contains over 70,000 multi-granular annotations organized at clip-level, event-level, and video-level segments, achieving a progression from perception to reasoning. As shown in Fig. 3(a) and (c), the durations of segments and the word numbers of text annotations at different granularities exhibit significant distributional differences. As shown in Fig. 3(d), HIVAU-70k's instruction data covers Caption, Judgment, Description, and Analysis for real-world anomalies, which guide the model to develop both short-term and long-term video anomaly understanding capabilities." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 597, + 443, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 597, + 443, + 609 + ], + "spans": [ + { + "bbox": [ + 313, + 597, + 443, + 609 + ], + "type": "text", + "content": "4. Method: Holmes-VAU" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": "Long-term video anomaly understanding with LLMs/VLMs has traditionally been hindered by frame redundancy, complicating accurate anomaly detection. Previous VAU approaches struggle with focus: methods like dense window sampling [67] add redundancy, and uniform frame sampling [9, 47] often misses key anomalies, limiting application to short videos. We introduce the Anomaly-focused Temporal Sampler (ATS) to address this," + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 66, + 702, + 233, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 233, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 233, + 713 + ], + "type": "text", + "content": "For detailed prompts, please refer to the appendix." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13846" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 70, + 294, + 290 + ], + "blocks": [ + { + "bbox": [ + 60, + 70, + 294, + 290 + ], + "lines": [ + { + "bbox": [ + 60, + 70, + 294, + 290 + ], + "spans": [ + { + "bbox": [ + 60, + 70, + 294, + 290 + ], + "type": "image", + "image_path": "f67777cea1f190d73ce83a9f2a9cff77e319b69aae4387130340c2fb515e0dca.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 293, + 295, + 316 + ], + "lines": [ + { + "bbox": [ + 55, + 293, + 295, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 295, + 316 + ], + "type": "text", + "content": "Figure 4. Holmes-VAU: a multi-modal-LLM-based video anomaly detection framework with adaptive anomaly focus." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 330, + 296, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 330, + 296, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 330, + 296, + 354 + ], + "type": "text", + "content": "integrating it into the VLM, and fine-tuning it via instruction on HIVAU-70k to form our Holmes-VAU model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 361, + 163, + 374 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 163, + 374 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 163, + 374 + ], + "type": "text", + "content": "4.1. Pipeline Overview" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 380, + 296, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 380, + 296, + 513 + ], + "spans": [ + { + "bbox": [ + 55, + 380, + 296, + 513 + ], + "type": "text", + "content": "The overall pipeline of our Holmes-VAU model is shown in Fig. 4. Video frames are processed by a visual encoder, creating visual tokens. These tokens are analyzed by an Anomaly-focused Temporal Sampler using an anomaly scorer and cumulative sum (" + }, + { + "bbox": [ + 55, + 380, + 296, + 513 + ], + "type": "inline_equation", + "content": "S_{cumsum}" + }, + { + "bbox": [ + 55, + 380, + 296, + 513 + ], + "type": "text", + "content": ") to select keyframes. A text encoder processes a prompt (e.g., 'Describe the abnormal events in the video'). Visual and textual representations are combined in a pre-trained language model, fine-tuned with LoRA, to generate a description of detected anomalies, such as a 'collision between a car and a motorcycle' and 'traffic accident' indicators." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 519, + 170, + 531 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 519, + 170, + 531 + ], + "spans": [ + { + "bbox": [ + 55, + 519, + 170, + 531 + ], + "type": "text", + "content": "4.2. Model Architecture" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 536, + 296, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 536, + 296, + 644 + ], + "spans": [ + { + "bbox": [ + 55, + 536, + 296, + 644 + ], + "type": "text", + "content": "Visual and Text embedding. We utilize the frozen visual encoder in InternVL2 [6], which inherits the ViT structure from CLIP [44], we refer to it as " + }, + { + "bbox": [ + 55, + 536, + 296, + 644 + ], + "type": "inline_equation", + "content": "\\phi_v" + }, + { + "bbox": [ + 55, + 536, + 296, + 644 + ], + "type": "text", + "content": ". Following previous VAU works [55, 58, 67, 74], we sample dense video frames at an interval of 16 frames from the input video. Each video frame is then processed through the visual encoder to obtain the corresponding visual tokens. Given the input video frame sequence " + }, + { + "bbox": [ + 55, + 536, + 296, + 644 + ], + "type": "inline_equation", + "content": "\\mathcal{V} \\in \\mathbb{R}^{T \\times H \\times W \\times C}" + }, + { + "bbox": [ + 55, + 536, + 296, + 644 + ], + "type": "text", + "content": ", the output features of " + }, + { + "bbox": [ + 55, + 536, + 296, + 644 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 536, + 296, + 644 + ], + "type": "text", + "content": "-th frame can be denoted as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 98, + 651, + 295, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 98, + 651, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 98, + 651, + 295, + 668 + ], + "type": "interline_equation", + "content": "V _ {i} = \\left\\{v _ {i} ^ {c l s}, v _ {i} ^ {1}, v _ {i} ^ {2}, \\dots , v _ {i} ^ {N _ {p}} \\right\\} = \\phi_ {v} \\left(\\mathcal {V} _ {i}\\right) \\tag {1}", + "image_path": "a1e61250c345dd1d084a74ed3bfd177ac11dca58f2f71adbde6c656f516180bd.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "v_{i}^{cls}" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": " indicates the class token, " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "v_{i}^{j}" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "j \\in \\{1,2,\\dots,N_{p}\\}" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": ") denotes the patch tokens, and " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "N_{p}" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": " represents the number of patches. The text encoder " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\phi_t" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": " is also initialized from [6]," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": "which includes a tokenizer and an embedding layer. The prompt text " + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": " is converted into text tokens through the text encoder: " + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "inline_equation", + "content": "X_{q} = \\phi_{t}(\\mathcal{Q})" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 108, + 554, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 108, + 554, + 215 + ], + "spans": [ + { + "bbox": [ + 313, + 108, + 554, + 215 + ], + "type": "text", + "content": "Anomaly-focused Temporal Sampler (ATS). ATS consists of two components: the anomaly scorer and the density-aware sampler. The anomaly scorer " + }, + { + "bbox": [ + 313, + 108, + 554, + 215 + ], + "type": "inline_equation", + "content": "\\phi_s" + }, + { + "bbox": [ + 313, + 108, + 554, + 215 + ], + "type": "text", + "content": " is a feature-based VAD network which estimates the anomaly score for each frame. We follow the network architecture in [74] due to its simplicity and good performance. Given the class token of the video frames " + }, + { + "bbox": [ + 313, + 108, + 554, + 215 + ], + "type": "inline_equation", + "content": "\\{v_1^{cls}, v_2^{cls}, \\dots, v_T^{cls}\\}" + }, + { + "bbox": [ + 313, + 108, + 554, + 215 + ], + "type": "text", + "content": ", the anomaly scores can be obtained: " + }, + { + "bbox": [ + 313, + 108, + 554, + 215 + ], + "type": "inline_equation", + "content": "s_i = \\phi_s(v_i^{cls})" + }, + { + "bbox": [ + 313, + 108, + 554, + 215 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 108, + 554, + 215 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 313, + 108, + 554, + 215 + ], + "type": "text", + "content": " denotes the anomaly score of the " + }, + { + "bbox": [ + 313, + 108, + 554, + 215 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 108, + 554, + 215 + ], + "type": "text", + "content": "-th frame." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 216, + 555, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 216, + 555, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 555, + 348 + ], + "type": "text", + "content": "Anomalous frames typically contain more information and exhibit greater variation than normal frames [49]. This observation motivates us to sample more frames in regions with higher anomaly scores while reducing the sampling in areas with lower anomaly scores. As shown in Fig. 4, to achieve non-uniform sampling, we propose density-aware sampler to selectively choose " + }, + { + "bbox": [ + 313, + 216, + 555, + 348 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 216, + 555, + 348 + ], + "type": "text", + "content": " frames from a total of " + }, + { + "bbox": [ + 313, + 216, + 555, + 348 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 216, + 555, + 348 + ], + "type": "text", + "content": " input frames. Specifically, we treat the anomaly scores " + }, + { + "bbox": [ + 313, + 216, + 555, + 348 + ], + "type": "inline_equation", + "content": "S \\in \\mathbb{R}^T" + }, + { + "bbox": [ + 313, + 216, + 555, + 348 + ], + "type": "text", + "content": " as a probability mass function and first accumulate them along the temporal dimension to obtain the cumulative distribution function, denoted as " + }, + { + "bbox": [ + 313, + 216, + 555, + 348 + ], + "type": "inline_equation", + "content": "S_{cumsum}" + }, + { + "bbox": [ + 313, + 216, + 555, + 348 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 377, + 356, + 553, + 389 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 356, + 553, + 389 + ], + "spans": [ + { + "bbox": [ + 377, + 356, + 553, + 389 + ], + "type": "interline_equation", + "content": "S _ {c u m s u m} (t) = \\sum_ {i = 1} ^ {t} \\left(s _ {i} + \\tau\\right) \\tag {2}", + "image_path": "b709a4b15081dcd737284148c53fd9f2573c49353b19dfc23f0b1579b674e799.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 396, + 554, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 554, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 554, + 467 + ], + "type": "text", + "content": "We uniformly sample " + }, + { + "bbox": [ + 313, + 396, + 554, + 467 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 396, + 554, + 467 + ], + "type": "text", + "content": " points along the cumulative axis, then map these points to the cumulative distribution " + }, + { + "bbox": [ + 313, + 396, + 554, + 467 + ], + "type": "inline_equation", + "content": "S_{cumsum}" + }, + { + "bbox": [ + 313, + 396, + 554, + 467 + ], + "type": "text", + "content": ", the corresponding " + }, + { + "bbox": [ + 313, + 396, + 554, + 467 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 396, + 554, + 467 + ], + "type": "text", + "content": " timestamps on the time axis are mapped to the closest frame index and finally form the sampled frame indices, denoted as " + }, + { + "bbox": [ + 313, + 396, + 554, + 467 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 313, + 396, + 554, + 467 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 396, + 554, + 467 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 396, + 554, + 467 + ], + "type": "text", + "content": " is used to control the uniformity of the sampling." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 468, + 554, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 554, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 554, + 588 + ], + "type": "text", + "content": "Projector and LLM. We select the tokens corresponding to the sampled frame, i.e., " + }, + { + "bbox": [ + 313, + 468, + 554, + 588 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 313, + 468, + 554, + 588 + ], + "type": "text", + "content": ", as the visual embedding. A projector " + }, + { + "bbox": [ + 313, + 468, + 554, + 588 + ], + "type": "inline_equation", + "content": "\\phi_p" + }, + { + "bbox": [ + 313, + 468, + 554, + 588 + ], + "type": "text", + "content": " is then used to map the visual embedding to the language feature space. Finally, we concatenate these embeddings with the text embeddings, input them into the pre-trained large language model, and compute the probability of the target answers " + }, + { + "bbox": [ + 313, + 468, + 554, + 588 + ], + "type": "inline_equation", + "content": "X_a" + }, + { + "bbox": [ + 313, + 468, + 554, + 588 + ], + "type": "text", + "content": ". To obtain an initial visual-language alignment space, we initialize the projector and LLM parameters from [6], with the parameters kept frozen. The above process can be expressed as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 355, + 597, + 553, + 611 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 597, + 553, + 611 + ], + "spans": [ + { + "bbox": [ + 355, + 597, + 553, + 611 + ], + "type": "interline_equation", + "content": "X _ {i n s} = \\operatorname {c a t} \\left[ \\phi_ {p} \\left(\\operatorname {c a t} \\left[ V _ {i} \\right]\\right), X _ {q} \\right] \\quad i \\in \\mathcal {G} \\tag {3}", + "image_path": "871f71460777190787c1d789bb6f0d076d69f675a67b43d5c6b1955f6a7ef41f.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 347, + 628, + 553, + 661 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 628, + 553, + 661 + ], + "spans": [ + { + "bbox": [ + 347, + 628, + 553, + 661 + ], + "type": "interline_equation", + "content": "p \\left(X _ {a} \\mid X _ {i n s}\\right) = \\prod_ {i = 1} ^ {L} p _ {\\theta} \\left(\\left(x _ {i} \\mid X _ {i n s, < i}, X _ {a, < i}\\right)\\right) \\tag {4}", + "image_path": "85455c9f2742c5e8a821bf4dee179dc252b40f96bcd9ffe611a0c573c379c420.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\operatorname{cat}[\\cdot]" + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "content": " represents the concatenation operation, " + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "content": " is the trainable parameters, " + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "inline_equation", + "content": "X_{ins, < i}, X_{a, < i}" + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "content": " are the instruction and answer tokens in all turns before the current prediction token " + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "content": " is the length of sequence, respectively." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13847" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 176, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 176, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 176, + 85 + ], + "type": "text", + "content": "4.3. Training and Testing" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "content": "Training. We train the model following two steps. In the first step, we use the video data and annotated frame-level label " + }, + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "inline_equation", + "content": "(\\hat{y} \\in \\mathbb{R}^T)" + }, + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "content": " from HIVAU-70k to train the anomaly scorer, which provides more accurate anomaly supervision compared to previous unsupervised and weakly-supervised methods [14, 46, 55, 74]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 72, + 172, + 295, + 205 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 172, + 295, + 205 + ], + "spans": [ + { + "bbox": [ + 72, + 172, + 295, + 205 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {A S} = - \\sum_ {i = 1} ^ {T} \\left(s _ {i} \\log \\left(\\hat {y} _ {i}\\right) + \\left(1 - s _ {i}\\right) \\log \\left(1 - \\hat {y} _ {i}\\right)\\right) \\tag {5}", + "image_path": "3f1f5e9ede8225fa00d4b4ffdb967c7e4ea52aecab60be02f751c7c2a4141728.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 215, + 296, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 215, + 296, + 287 + ], + "spans": [ + { + "bbox": [ + 55, + 215, + 296, + 287 + ], + "type": "text", + "content": "In the second step, we keep the anomaly scorer fixed, and use all the instruction data from HIVAU-70k to train the model. To achieve more efficient fine-tuning without disrupting the original capabilities of the LLM, we employ LoRA [15] for fine-tuning, optimizing the cross entropy loss between the predicted and the ground truth tokens." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 288, + 296, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 288, + 296, + 324 + ], + "spans": [ + { + "bbox": [ + 55, + 288, + 296, + 324 + ], + "type": "text", + "content": "Testing. During testing, the user inputs a video and text prompts; the model will generate the corresponding text response following the user's instruction." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 337, + 137, + 351 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 337, + 137, + 351 + ], + "spans": [ + { + "bbox": [ + 55, + 337, + 137, + 351 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 357, + 164, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 357, + 164, + 371 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 164, + 371 + ], + "type": "text", + "content": "5.1. Experiment Setup" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 376, + 296, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 296, + 520 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 296, + 520 + ], + "type": "text", + "content": "Dataset. Our HIVAU-70k is built upon two large-scale real-world datasets, i.e., UCF-Crime [46] and XD-Violence [56], they provide a diverse range of videos with anomalous events. UCF-Crime [46] comprises 1,900 untrimmed videos totaling 128 hours from outdoor and indoor surveillance cameras. It encompasses 13 classes of real-world anomalies, including Abuse, Explosion, Fighting, and Shooting. XD-Violence [56] is the largest VAD benchmark, comprising 4,754 videos totaling 217 hours sourced from surveillance, movies, car cameras, and games. It encompasses 6 anomaly classes: Abuse, Car Accidents, Explosions, Fighting, Riots, and Shooting." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 521, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 700 + ], + "type": "text", + "content": "Metric. We assess the anomaly understanding ability from two aspects: anomaly detection and reasoning. 1) For anomaly detection, we use the anomaly scores output by the Anomaly Scorer as the prediction and perform the evaluation. Following [14, 46, 67, 74], we use AUC and AP to quantify detection performance, which is evaluated only on the video level. 2) For anomaly reasoning, we annotate instruction data from the UCF-Crime and XD-Violence test sets, which have been carefully reviewed and filtered by annotators. We finally collected 3,300 test samples at multiple granularities. The test set contains 2200/732/398 samples at clip/event/video levels. We calculate metrics including BLEU [41], CIDEr [51], METEOR [4] and ROUGE [26] to measure the quality of the reasoning text output by the model, comparing with the annotated ground truth text." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "content": "Implementation Details. For the proposed Holmes-VAU" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 318, + 117, + 550, + 337 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 553, + 114 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 553, + 114 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 553, + 114 + ], + "type": "text", + "content": "Table 1. Comparison of detection performance with state-of-the-art Video Anomaly Detection approaches. We include the results of explainable and non-explainable methods. \\*\\* represents the result reported in [67]." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 117, + 550, + 337 + ], + "lines": [ + { + "bbox": [ + 318, + 117, + 550, + 337 + ], + "spans": [ + { + "bbox": [ + 318, + 117, + 550, + 337 + ], + "type": "table", + "html": "
MethodsBackboneXD-ViolenceUCF-Crime
AP/%AUC/%
Non-explainable VAD
Conv-AE [14] (CVPR'16)-27.2550.60
GODS [52] (ICCV'19)I3DN/A70.46
GCL [66] (CVPR'22)ResNextN/A71.04
DYANNET [48] (WACV'23)I3DN/A84.50
MIST [11] (CVPR'21)I3DN/A82.30
Wu et al. [56] (ECCV'20)I3D78.6482.44
RTFM [49] (ICCV'21)I3D77.8184.30
MSL [22] (AAAI'22)I3D78.2885.30
S3R [55] (ECCV'22)I3D80.2685.99
MGFN [5] (AAAI'23)I3D79.1986.98
UR-DMU [74] (AAAI'23)I3D81.6686.97
CLIP-TSA [17] (ICIP'23)ViT82.1987.58
VadCLIP [58] (AAAI'24)ViT84.5188.02
Yang et al. [62] (CVPR'24)ViT83.6887.79
Wu et al. [57] (CVPR'24)ViT66.5386.40
Explainable Multi-modal VAD
Zero-Shot CLIP [44]*ViT17.8353.16
LLAVA-1.5 [27]*ViT50.2672.84
LAVAD [67] (CVPR'24)ViT62.0180.28
Holmes-VAU (Ours)ViT87.6888.96
", + "image_path": "1fffd0e8301545be9613af8604b5afe3d1b9cb40631198f730b7c29d199aa20f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 351, + 555, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 351, + 555, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 351, + 555, + 495 + ], + "type": "text", + "content": "method, we initialize the Multimodal LLM with InternVL2-2B [6]. To optimize the Anomaly-focused Temporal Sampler, we adopt the Adam optimizer with a learning rate of 1e-4. Note that when evaluating detection performance on XD-Violence and UCF-Crime, only videos in the corresponding training sets are used to train our model for fair comparisons. For instruction tuning, we train with a batch size of 512 for 1 epoch, using the AdamW optimizer with cosine learning rate decay and a warm-up period. The LoRA [15] parameters are set as: " + }, + { + "bbox": [ + 313, + 351, + 555, + 495 + ], + "type": "inline_equation", + "content": "r = 64" + }, + { + "bbox": [ + 313, + 351, + 555, + 495 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 351, + 555, + 495 + ], + "type": "inline_equation", + "content": "\\alpha = 128" + }, + { + "bbox": [ + 313, + 351, + 555, + 495 + ], + "type": "text", + "content": ", and learning rate=4e-5. During testing, " + }, + { + "bbox": [ + 313, + 351, + 555, + 495 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 351, + 555, + 495 + ], + "type": "text", + "content": " in Eq. 2 is set to 0.1. Experiments are conducted on 2 NVIDIA A100 GPUs." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 503, + 400, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 503, + 400, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 503, + 400, + 514 + ], + "type": "text", + "content": "5.2. Main Results" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": "Anomaly Detection Results. We compare our method with state-of-the-art methods, including semi-supervised methods [14, 52], unsupervised methods [48, 66], weakly-supervised methods [17, 22, 49, 55, 58, 74] and recently training-free method [67]. We have indicated their backbones and performance on the UCF-Crime and XD-Violence datasets, as shown in Table 1. Our method has an AP of " + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "inline_equation", + "content": "87.68\\%" + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": " on XD-Violence and an AUC of " + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "inline_equation", + "content": "88.96\\%" + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": " on UCF-Crime, significantly outperforming the prior state-of-the-art methods, which demonstrates that our method can generate less biased anomaly scores. It is worth noting that while achieving precise localization of anomalies, Holmes-VAU is also capable of providing explanations and analysis for the detected anomalies by the model, a feature unavailable in existing non-explainable VAD methods. Although LAVAD [67] has explainability, the training-free large lan" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13848" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 106, + 553, + 203 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "type": "text", + "content": "Table 2. Comparison of reasoning performance with state-of-the-art Multimodal Large Language Models (MLLMs). 'BLEU' refers to the cumulative values from BLEU-1 to BLEU-4. We evaluate the quality of the generated text at different granularities, including clip-level (C), event-level (E), and video-level (V)." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 106, + 553, + 203 + ], + "lines": [ + { + "bbox": [ + 58, + 106, + 553, + 203 + ], + "spans": [ + { + "bbox": [ + 58, + 106, + 553, + 203 + ], + "type": "table", + "html": "
MethodParamsBLEU [41](↑)CIDEr [51](↑)METEOR [4](↑)ROUGE [26](↑)
CEVCEVCEVCEV
Video-ChatGPT [39]7B0.1520.0680.0660.0330.0110.0130.1020.0690.0440.1530.0480.079
Video-LLaMA [68]7B0.1510.0790.1040.0240.0140.0170.1120.0760.0570.1560.0670.090
Video-LLaVA [25]7B0.1640.0460.0550.0320.0090.0130.0970.0220.0140.1320.0230.045
LLaVA-Next-Video [71]7B0.4350.0910.1200.1020.0150.0310.1170.0850.0960.1980.0800.106
QwenVL2 [53]7B0.3120.0820.1550.0440.0200.0440.1330.0920.1120.1630.0810.137
InternVL2 [6]8B0.3310.1010.1450.0520.0220.0350.1410.0950.1010.1820.1020.122
Holmes-VAU (Ours)2B0.9130.8040.5660.4671.5191.4370.1900.1650.1210.3290.3700.355
", + "image_path": "cc583d64780c65bda3f6d9249aede0002b59ddedde52a33b8d1737fdec3aa11b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 58, + 245, + 305, + 344 + ], + "blocks": [ + { + "bbox": [ + 55, + 209, + 295, + 243 + ], + "lines": [ + { + "bbox": [ + 55, + 209, + 295, + 243 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 295, + 243 + ], + "type": "text", + "content": "Table 3. Ablation of hierarchical instruction data. During the instruction tuning phase, we combined training data of different granularities, including clip (C), event (E), and video (V) levels." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 245, + 305, + 344 + ], + "lines": [ + { + "bbox": [ + 58, + 245, + 305, + 344 + ], + "spans": [ + { + "bbox": [ + 58, + 245, + 305, + 344 + ], + "type": "table", + "html": "
Training DataBLEU(↑)CIDEr(↑)
CEVCEVCEV
✓0.9840.2610.3510.4590.1200.106
✓0.5080.5760.2920.0971.1830.872
✓0.2800.2220.2790.0390.7080.884
✓✓0.8890.7410.3490.4701.2850.889
✓✓0.9060.3410.5220.4720.9621.093
✓✓0.3940.7970.5050.0811.4721.074
✓✓✓0.9130.8040.5660.4671.5191.437
", + "image_path": "87cc6987fe0258a1eed8290e62f96fa28f0a33cc46ca981b51f3aac92b4539b4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 360, + 295, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 360, + 295, + 383 + ], + "spans": [ + { + "bbox": [ + 55, + 360, + 295, + 383 + ], + "type": "text", + "content": "guage model lacks an understanding of anomaly knowledge due to the limitation of insufficient supervised data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 384, + 295, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 295, + 624 + ], + "type": "text", + "content": "Anomaly Reasoning Results. We compare the anomaly-related text quality generated by Holmes-VAU with that produced by state-of-the-art general Multimodal Large Language Models (MLLMs), and presented the results at different temporal granularities, including clip-level, event-level, and video-level, as shown in Table 2. Earlier MLLMs such as Video-ChatGPT [21] and Video-LLaMA [68], struggled with basic visual perception and instruction-following capabilities. Recent MLLMs [6, 7, 71] trained on larger and higher-quality video instruction data have made significant progress in general video understanding, with noticeable improvements at the clip-level perception task. However, due to the absence of learning from complex, real-world anomaly data, their reasoning abilities at the event-level and video-level are still lacking. Our Holmes-VAU, however, shows significant improvements in video understanding across all temporal granularities compared to existing general MLLMs, highlighting the importance of injecting anomaly-related knowledge through instruction tuning on high-quality Video Anomaly Understanding benchmarks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 634, + 155, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 634, + 155, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 634, + 155, + 647 + ], + "type": "text", + "content": "5.3. Analytic Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 653, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 715 + ], + "type": "text", + "content": "Influence of Hierarchical Instruction. To explore the impact of different granularity video training data on the model's anomaly reasoning ability, we designed various training data combinations during the instruction tuning phase and evaluated the model's performance, as shown" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 317, + 268, + 553, + 378 + ], + "blocks": [ + { + "bbox": [ + 313, + 209, + 555, + 264 + ], + "lines": [ + { + "bbox": [ + 313, + 209, + 555, + 264 + ], + "spans": [ + { + "bbox": [ + 313, + 209, + 555, + 264 + ], + "type": "text", + "content": "Table 4. Ablation study of sampling methods and the number of sampled frames. We compare the proposed Anomaly-focused Temporal Sampler (ATS) with other sampling methods under different frame sampling numbers, including Uniform and Top-K. Latency is the time delay in generating the first token." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 268, + 553, + 378 + ], + "lines": [ + { + "bbox": [ + 317, + 268, + 553, + 378 + ], + "spans": [ + { + "bbox": [ + 317, + 268, + 553, + 378 + ], + "type": "table", + "html": "
Frames (N)SamplerLatency (ms)Video-level
BLEU(↑)CIDEr(↑)
8Top-K2440.4621.229
Uniform0.4911.276
ATS (Ours)0.5141.324
16Top-K5660.4761.302
Uniform0.5111.345
ATS (Ours)0.5661.437
32Top-K14020.4811.332
Uniform0.5581.357
ATS (Ours)0.5761.460
", + "image_path": "8cdf84ff07e3edfb89100c945b5ef5e72bbd712f920c4b2b182b3be9a07c28c6.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 399, + 555, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 555, + 531 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 555, + 531 + ], + "type": "text", + "content": "in Table 3. The inclusion of clip-level data primarily enhanced the model's basic visual perception abilities regarding actions and scenes within the video. Adding event-level data improved the model's ability to judge and understand complete anomaly events. Furthermore, the involvement of video-level data further enhanced the model's ability to analyze and summarize anomaly-related information across longer-span videos. The hierarchical instruction data structure facilitated a comprehensive and complementary improvement in the model's anomaly-related perception-to-reasoning capabilities." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "text", + "content": "Influence of different sampling methods and the number of sampled frames. Our ATS (Anomaly-focused Temporal Sampler) is designed to adaptively sample frames input to the LLM based on the frame-level anomaly scores. To validate its advantages, we compared ATS with other sampling methods at various sample frame counts, including Uniform and Top-K sampling. In Uniform sampling, N frames are uniformly sampled from all frames, while Top-K sampling selects the frames with the top " + }, + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "text", + "content": " highest anomaly scores. As shown in Table 4, ATS consistently outperforms other sampling methods, regardless of the sample count. We believe that Uniform sampling tends to overlook key anomaly frames, though this issue lessens as more frames are sampled. Besides, Top-K sampling tends to overly focus on local anomaly frames, missing contextual frame informa" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13849" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 76, + 138, + 105 + ], + "blocks": [ + { + "bbox": [ + 63, + 76, + 138, + 105 + ], + "lines": [ + { + "bbox": [ + 63, + 76, + 138, + 105 + ], + "spans": [ + { + "bbox": [ + 63, + 76, + 138, + 105 + ], + "type": "image", + "image_path": "a3539e036593faaac8034fc67e9939e4aa357ae675c99186e6486495ab7caf89.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 140, + 87, + 150, + 92 + ], + "lines": [ + { + "bbox": [ + 140, + 87, + 150, + 92 + ], + "spans": [ + { + "bbox": [ + 140, + 87, + 150, + 92 + ], + "type": "text", + "content": "··" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 156, + 77, + 296, + 106 + ], + "blocks": [ + { + "bbox": [ + 156, + 77, + 296, + 106 + ], + "lines": [ + { + "bbox": [ + 156, + 77, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 156, + 77, + 296, + 106 + ], + "type": "image", + "image_path": "81c4fc8f13a8690ef1e239037ca10d2daaa0281908bb60f59303889bbb67ae61.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 108, + 170, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 108, + 170, + 129 + ], + "spans": [ + { + "bbox": [ + 62, + 108, + 170, + 129 + ], + "type": "text", + "content": "Prompt: Determine whether there are anomaly events in the video and the specific name of the anomaly." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 63, + 137, + 100, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 137, + 100, + 144 + ], + "spans": [ + { + "bbox": [ + 63, + 137, + 100, + 144 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 63, + 147, + 178, + 160 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 147, + 178, + 160 + ], + "spans": [ + { + "bbox": [ + 63, + 147, + 178, + 160 + ], + "type": "text", + "content": "The anomaly exists and is specifically identified as a Burglary event." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 63, + 175, + 114, + 182 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 175, + 114, + 182 + ], + "spans": [ + { + "bbox": [ + 63, + 175, + 114, + 182 + ], + "type": "text", + "content": "Holmes-VAU (Ours)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 63, + 184, + 176, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 184, + 176, + 205 + ], + "spans": [ + { + "bbox": [ + 63, + 184, + 176, + 205 + ], + "type": "text", + "content": "Yes, there is an anomaly in the video, specifically a break-in burglary." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 187, + 115, + 292, + 150 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 115, + 292, + 150 + ], + "spans": [ + { + "bbox": [ + 187, + 115, + 292, + 150 + ], + "type": "text", + "content": "InternVL2-BB [Incorrect Judgement] Yes, there are anomaly events in the video. The specific anomaly is that the man in the video is seen throwing a punch at the other man." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 187, + 163, + 291, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 163, + 291, + 186 + ], + "spans": [ + { + "bbox": [ + 187, + 163, + 291, + 186 + ], + "type": "text", + "content": "QwenVL2-7B [Incorrect Judgement] There are no anomaly events in the video." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 63, + 213, + 138, + 243 + ], + "blocks": [ + { + "bbox": [ + 63, + 213, + 138, + 243 + ], + "lines": [ + { + "bbox": [ + 63, + 213, + 138, + 243 + ], + "spans": [ + { + "bbox": [ + 63, + 213, + 138, + 243 + ], + "type": "image", + "image_path": "2c898df8c75fe0d6e6131670fcaea798172033a90f1443db56cc54474af81ebd.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 135, + 215, + 186, + 243 + ], + "blocks": [ + { + "bbox": [ + 212, + 195, + 269, + 203 + ], + "lines": [ + { + "bbox": [ + 212, + 195, + 269, + 203 + ], + "spans": [ + { + "bbox": [ + 212, + 195, + 269, + 203 + ], + "type": "text", + "content": "Bulgary037_x264" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 135, + 215, + 186, + 243 + ], + "lines": [ + { + "bbox": [ + 135, + 215, + 186, + 243 + ], + "spans": [ + { + "bbox": [ + 135, + 215, + 186, + 243 + ], + "type": "image", + "image_path": "00195e2b01988744be8bd79d766165fedfe371e83651df04f37d07ef36992356.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 186, + 214, + 294, + 243 + ], + "blocks": [ + { + "bbox": [ + 186, + 214, + 294, + 243 + ], + "lines": [ + { + "bbox": [ + 186, + 214, + 294, + 243 + ], + "spans": [ + { + "bbox": [ + 186, + 214, + 294, + 243 + ], + "type": "image", + "image_path": "cf38f80e93e5239daed104c387edede3cf469af887927484dc5aa55c1baaa202.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 63, + 247, + 173, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 247, + 173, + 261 + ], + "spans": [ + { + "bbox": [ + 63, + 247, + 173, + 261 + ], + "type": "text", + "content": "Prompt: Identify any abnormal events depicted in the video." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 186, + 252, + 292, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 252, + 292, + 293 + ], + "spans": [ + { + "bbox": [ + 186, + 252, + 292, + 293 + ], + "type": "text", + "content": "InternVL2-8B [Incorrect Judgement] The video shows a white truck parked on the side of the road, which is an abnormal behavior as it is not supposed to be parked on the road." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 63, + 275, + 100, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 275, + 100, + 281 + ], + "spans": [ + { + "bbox": [ + 63, + 275, + 100, + 281 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 63, + 285, + 156, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 285, + 156, + 293 + ], + "spans": [ + { + "bbox": [ + 63, + 285, + 156, + 293 + ], + "type": "text", + "content": "No anomaly exists in the video" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 63, + 312, + 115, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 312, + 115, + 319 + ], + "spans": [ + { + "bbox": [ + 63, + 312, + 115, + 319 + ], + "type": "text", + "content": "Holmes-VAU (Ours)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 63, + 324, + 162, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 324, + 162, + 331 + ], + "spans": [ + { + "bbox": [ + 63, + 324, + 162, + 331 + ], + "type": "text", + "content": "There is no anomaly in the video." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 187, + 300, + 222, + 307 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 300, + 222, + 307 + ], + "spans": [ + { + "bbox": [ + 187, + 300, + 222, + 307 + ], + "type": "text", + "content": "QwenVL2-7B" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 187, + 310, + 291, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 310, + 291, + 323 + ], + "spans": [ + { + "bbox": [ + 187, + 310, + 291, + 323 + ], + "type": "text", + "content": "The video does not show any abnormal events." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 202, + 333, + 280, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 333, + 280, + 341 + ], + "spans": [ + { + "bbox": [ + 202, + 333, + 280, + 341 + ], + "type": "text", + "content": "Normal_Videos_902_x264" + } + ] + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 309, + 76, + 545, + 106 + ], + "blocks": [ + { + "bbox": [ + 309, + 76, + 545, + 106 + ], + "lines": [ + { + "bbox": [ + 309, + 76, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 309, + 76, + 545, + 106 + ], + "type": "image", + "image_path": "c4a55d808cdb7b4123c4a65bfcc83b439904c3a5536425786e201a75908c248d.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "bbox": [ + 309, + 109, + 492, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 109, + 492, + 129 + ], + "spans": [ + { + "bbox": [ + 309, + 109, + 492, + 129 + ], + "type": "text", + "content": "Prompt: How would you describe the particular anomaly events in the video? What evidence do you have to support your judgment of this event as an anomaly?" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 309, + 137, + 347, + 143 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 137, + 347, + 143 + ], + "spans": [ + { + "bbox": [ + 309, + 137, + 347, + 143 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 309, + 143, + 542, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 143, + 542, + 189 + ], + "spans": [ + { + "bbox": [ + 309, + 143, + 542, + 189 + ], + "type": "text", + "content": "The video anomaly event depicts chaotic street scenes where people are running, police officers are clad in riot gear, and confrontations or protests are underway, with scenes of apprehension, restraint, and property damage unfolding from start to end. The presence of violent behavior, law enforcement intervention, and disruption of public order in the urban environment depicted in the video, which collectively indicate a significant deviation from normal, peaceful civic activity." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 309, + 194, + 361, + 200 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 194, + 361, + 200 + ], + "spans": [ + { + "bbox": [ + 309, + 194, + 361, + 200 + ], + "type": "text", + "content": "Holmes-VAU (Ours)" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 309, + 201, + 542, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 201, + 542, + 241 + ], + "spans": [ + { + "bbox": [ + 309, + 201, + 542, + 241 + ], + "type": "text", + "content": "The anomaly events involve a series of chaotic and violent scenes, including police officers in riot gear, people running and pushing against barriers, and a helicopter flying over the scene, suggesting a high level of tension and unrest. The presence of destructive behavior, including property damage, property destruction, and physical altercations, which deviates from normal and peaceful urban activity." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 309, + 247, + 454, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 247, + 454, + 255 + ], + "spans": [ + { + "bbox": [ + 309, + 247, + 454, + 255 + ], + "type": "text", + "content": "InternVL2-8B [Visual Hallucination, Incomplete Analysis]" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 309, + 255, + 542, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 255, + 542, + 281 + ], + "spans": [ + { + "bbox": [ + 309, + 255, + 542, + 281 + ], + "type": "text", + "content": "The video shows a group of people walking down a street, and a man is seen walking down the street with a skateboard. The man is wearing a black jacket and a black hat. The event depicted in the video is classified as an anomaly due to the presence of a large group of people engaging in violent behavior." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 309, + 288, + 460, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 288, + 460, + 295 + ], + "spans": [ + { + "bbox": [ + 309, + 288, + 460, + 295 + ], + "type": "text", + "content": "QwenVL2-7B [Incomplete Description, Incomplete Analysis]" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 309, + 295, + 541, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 295, + 541, + 327 + ], + "spans": [ + { + "bbox": [ + 309, + 295, + 541, + 327 + ], + "type": "text", + "content": "The video shows a group of people walking down the street, and suddenly, a car crashes into a crowd of people. The car is on fire, and the people are running away. The video does not provide any context or explanation for this event, but it is clear that it is an anomaly. The car crash is a violent and unexpected event that is not related to the rest of the video." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 371, + 334, + 476, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 371, + 334, + 476, + 342 + ], + "spans": [ + { + "bbox": [ + 371, + 334, + 476, + 342 + ], + "type": "text", + "content": "v=3WmMhircZ0c__#1_label_B4-0-0" + } + ] + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 58, + 396, + 156, + 490 + ], + "blocks": [ + { + "bbox": [ + 55, + 352, + 555, + 385 + ], + "lines": [ + { + "bbox": [ + 55, + 352, + 555, + 385 + ], + "spans": [ + { + "bbox": [ + 55, + 352, + 555, + 385 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison of anomaly understanding explanation. Compared with state-of-the-art general MLLMs, i.e., InternVL2 [6] and QwenVL2 [53], our proposed Holmes-VAU demonstrates more accurate anomaly judgment, along with more detailed and comprehensive anomaly-related descriptions and analysis. Correct and wrong explanations are highlighted in green and red, respectively." + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 396, + 156, + 490 + ], + "lines": [ + { + "bbox": [ + 58, + 396, + 156, + 490 + ], + "spans": [ + { + "bbox": [ + 58, + 396, + 156, + 490 + ], + "type": "image", + "image_path": "5f553fb092e3feaf3b37f7f59db441378fa98a7f126f589dc496b44745634815.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 491, + 295, + 536 + ], + "lines": [ + { + "bbox": [ + 55, + 491, + 295, + 536 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 295, + 536 + ], + "type": "text", + "content": "Figure 6. Ablation study of trainable parameters. (a) Loss curve during instruction-tuning. (b) We tuned the LoRA [15] parameter " + }, + { + "bbox": [ + 55, + 491, + 295, + 536 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 491, + 295, + 536 + ], + "type": "text", + "content": " to control trainable parameters, evaluating its impact on VAU capability, and General performance on Video-MME [12]." + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 156, + 397, + 294, + 485 + ], + "blocks": [ + { + "bbox": [ + 156, + 397, + 294, + 485 + ], + "lines": [ + { + "bbox": [ + 156, + 397, + 294, + 485 + ], + "spans": [ + { + "bbox": [ + 156, + 397, + 294, + 485 + ], + "type": "image", + "image_path": "2e799acaa4580cdb49022072ce0ded081fc1bc24a2b2833a33c72c5c42a4c316.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "bbox": [ + 55, + 555, + 295, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 295, + 590 + ], + "type": "text", + "content": "tion. Our proposed ATS mitigates both issues. To balance inference efficiency and performance, we set " + }, + { + "bbox": [ + 55, + 555, + 295, + 590 + ], + "type": "inline_equation", + "content": "N = 16" + }, + { + "bbox": [ + 55, + 555, + 295, + 590 + ], + "type": "text", + "content": " as the default sample frame number." + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": "Instruction Tuning Parameters. We conducted an ablation study on the parameter " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " in LoRA [15] to explore how the trainable parameters affects both the model's VAU performance and its general capability. We use VideoMME [12] to evaluate the model's general capability. The results are shown in Fig. 6, as " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " increases, the model gradually adapts to the VAU task. However, when " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " becomes too large, the model's general capability decreases. To retain the original general video understanding capability of the MLLM, we set " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "r = 64" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " as the default value." + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 314, + 394, + 452, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 394, + 452, + 407 + ], + "spans": [ + { + "bbox": [ + 314, + 394, + 452, + 407 + ], + "type": "text", + "content": "5.4. Qualitative Comparison" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 313, + 415, + 555, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 555, + 487 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 555, + 487 + ], + "type": "text", + "content": "We provide qualitative comparisons between Holmes-VAU and existing MLLMs in Fig. 5. The results demonstrate that Holmes-VAU can accurately identify anomalies in videos and provide accurate and complete explanations, highlight the effectiveness and advantage of Holmes-VAU in perceiving video events and analyzing anomalies." + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 314, + 501, + 387, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 501, + 387, + 513 + ], + "spans": [ + { + "bbox": [ + 314, + 501, + 387, + 513 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 313, + 522, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 555, + 714 + ], + "type": "text", + "content": "In conclusion, this work pushes the boundaries of video anomaly understanding by introducing hierarchical anomaly detection across diverse temporal scales, from momentary clips to extended events. The HIVAU-70k benchmark, with over 70,000 multi-level annotations, addresses a critical gap in the field, enabling comprehensive anomaly analysis in real-world scenarios. Our Anomaly-focused Temporal Sampler (ATS) strategically enhances focus on anomaly-dense segments, optimizing both efficiency and accuracy in long-term anomaly detection. Extensive experiments demonstrate that our hierarchical dataset and ATS-enhanced VLM achieve significant performance gains over conventional methods, proving robust for open-world anomaly understanding. This work sets a new standard for multi-granular anomaly comprehension, paving the way for more fine-grained video anomaly understanding." + } + ] + } + ], + "index": 43 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13850" + } + ] + } + ], + "index": 44 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 130, + 115, + 142 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 130, + 115, + 142 + ], + "spans": [ + { + "bbox": [ + 56, + 130, + 115, + 142 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 149, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 149, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 149, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 61, + 149, + 294, + 194 + ], + "type": "text", + "content": "[1] Amit Adam, Ehud Rivlin, Ilan Shimshoni, and Daviv Reinitz. Robust real-time unusual event detection using multiple fixed-location monitors. IEEE transactions on pattern analysis and machine intelligence, 30(3):555-560, 2008. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 194, + 226, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 194, + 226, + 204 + ], + "spans": [ + { + "bbox": [ + 62, + 194, + 226, + 204 + ], + "type": "text", + "content": "[2] AI@Meta. Llama 3 model card. 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 205, + 295, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 205, + 295, + 259 + ], + "spans": [ + { + "bbox": [ + 62, + 205, + 295, + 259 + ], + "type": "text", + "content": "[3] Kumar Ashutosh, Rohit Girdhar, Lorenzo Torresani, and Kristen Grauman. Hiervl: Learning hierarchical videolanguage embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23066-23078, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 261, + 295, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 261, + 295, + 315 + ], + "spans": [ + { + "bbox": [ + 62, + 261, + 295, + 315 + ], + "type": "text", + "content": "[4] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65-72, 2005. 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 316, + 294, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 316, + 294, + 370 + ], + "spans": [ + { + "bbox": [ + 62, + 316, + 294, + 370 + ], + "type": "text", + "content": "[5] Yingxian Chen, Zhengzhe Liu, Baoheng Zhang, Wilton Fok, Xiaojuan Qi, and Yik-Chung Wu. Mgfn: Magnitude-contrastive glance-and-focus network for weakly-supervised video anomaly detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 387-395, 2023. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 372, + 294, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 372, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 62, + 372, + 294, + 426 + ], + "type": "text", + "content": "[6] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 427, + 294, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 427, + 294, + 480 + ], + "spans": [ + { + "bbox": [ + 62, + 427, + 294, + 480 + ], + "type": "text", + "content": "[7] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 482, + 294, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 482, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 62, + 482, + 294, + 536 + ], + "type": "text", + "content": "[8] Jisheng Dang, Huicheng Zheng, Xiaohao Xu, Longguang Wang, Qingyong Hu, and Yulan Guo. Adaptive sparse memory networks for efficient and robust video object segmentation. IEEE Transactions on Neural Networks and Learning Systems, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 537, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 537, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 62, + 537, + 294, + 613 + ], + "type": "text", + "content": "[9] Hang Du, Sicheng Zhang, Binzhu Xie, Guoshun Nan, Jiayang Zhang, Junrui Xu, Hangyu Liu, Sicong Leng, Jiangming Liu, Hehe Fan, et al. Uncovering what why and how: A comprehensive benchmark for causation understanding of video anomaly. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18793-18803, 2024. 2, 3, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 614, + 294, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 614, + 294, + 658 + ], + "spans": [ + { + "bbox": [ + 57, + 614, + 294, + 658 + ], + "type": "text", + "content": "[10] Xinyu Fang, Kangrui Mao, Haodong Duan, Xiangyu Zhao, Yining Li, Dahua Lin, and Kai Chen. Mmbench-video: A long-form multi-shot benchmark for holistic video understanding. arXiv preprint arXiv:2406.14515, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 294, + 712 + ], + "type": "text", + "content": "[11] Jia-Chang Feng, Fa-Ting Hong, and Wei-Shi Zheng. Mist: Multiple instance self-training framework for video anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14009-14018, 2021. 2, 6" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 57, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "Acknowledgement This work is supported by the National Natural Science Foundation of China under grants U22B2053 and 623B2039, and in part by the Interdisciplinary Research Program of HUST (2024JCYJ034)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "text", + "content": "[12] Chaoyou Fu, Yuhan Dai, Yondong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 3, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 129, + 553, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 553, + 204 + ], + "type": "text", + "content": "[13] Dong Gong, Lingqiao Liu, Vuong Le, Budhaditya Saha, Moussa Reda Mansour, Svetha Venkatesh, and Anton van den Hengel. Memorizing normality to detect anomaly: Memory-augmented deep autoencoder for unsupervised anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1705-1714, 2019. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 206, + 553, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 206, + 553, + 258 + ], + "spans": [ + { + "bbox": [ + 317, + 206, + 553, + 258 + ], + "type": "text", + "content": "[14] Mahmudul Hasan, Jonghyun Choi, Jan Neumann, Amit K Roy-Chowdhury, and Larry S Davis. Learning temporal regularity in video sequences. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 733-742, 2016. 2, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 261, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 553, + 304 + ], + "type": "text", + "content": "[15] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 6, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 305, + 553, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 305, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 316, + 305, + 553, + 360 + ], + "type": "text", + "content": "[16] Md Mohaiminul Islam, Ngan Ho, Xitong Yang, Tushar Nagarajan, Lorenzo Torresani, and Gedas Bertasius. Video recap: Recursive captioning of hour-long videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18198-18208, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 361, + 553, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 361, + 553, + 414 + ], + "spans": [ + { + "bbox": [ + 317, + 361, + 553, + 414 + ], + "type": "text", + "content": "[17] Hyekang Kevin Joo, Khoa Vo, Kashu Yamazaki, and Ngan Le. Clip-tsa: Clip-assisted temporal self-attention for weakly-supervised video anomaly detection. In 2023 IEEE International Conference on Image Processing (ICIP), pages 3230–3234. IEEE, 2023. 3, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 415, + 553, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 415, + 553, + 458 + ], + "spans": [ + { + "bbox": [ + 317, + 415, + 553, + 458 + ], + "type": "text", + "content": "[18] Kumara Kahatapitiya and Michael S Ryoo. Coarse-fine networks for temporal activity detection in videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8385-8394, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 460, + 553, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 460, + 553, + 513 + ], + "spans": [ + { + "bbox": [ + 317, + 460, + 553, + 513 + ], + "type": "text", + "content": "[19] Jaechul Kim and Kristen Grauman. Observe locally, infer globally: a space-time mrf for detecting abnormal activities with incremental updates. In 2009 IEEE conference on computer vision and pattern recognition, pages 2921-2928. IEEE, 2009. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 514, + 553, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 514, + 553, + 546 + ], + "spans": [ + { + "bbox": [ + 317, + 514, + 553, + 546 + ], + "type": "text", + "content": "[20] Federico Landi, Cees GM Snoek, and Rita Cucchiara. Anomaly locality in video surveillance. arXiv preprint arXiv:1901.10364, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 548, + 553, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 548, + 553, + 591 + ], + "spans": [ + { + "bbox": [ + 317, + 548, + 553, + 591 + ], + "type": "text", + "content": "[21] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 2, 4, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 593, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 593, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 317, + 593, + 553, + 645 + ], + "type": "text", + "content": "[22] Shuo Li, Fang Liu, and Licheng Jiao. Self-training multisession learning with transformer for weakly supervised video anomaly detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1395-1403, 2022. 2, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "text", + "content": "[23] Weixin Li, Vijay Mahadevan, and Nuno Vasconcelos. Anomaly detection and localization in crowded scenes. IEEE transactions on pattern analysis and machine intelligence, 36(1):18-32, 2013. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[24] Yingwei Li, Yi Li, and Nuno Vasconcelos. Resound: Towards action recognition without representation bias. In Pro" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "13851" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "text", + "content": "ceedings of the European Conference on Computer Vision (ECCV), pages 513-528, 2018. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 96, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 96, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 57, + 96, + 295, + 140 + ], + "type": "text", + "content": "[25] Bin Lin, Bin Zhu, Yang Ye, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 2, 4, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 173 + ], + "type": "text", + "content": "[26] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 175, + 295, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 295, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 295, + 207 + ], + "type": "text", + "content": "[27] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 2, 4, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 209, + 294, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 241 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 241 + ], + "type": "text", + "content": "[28] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024.3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 243, + 294, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 294, + 275 + ], + "type": "text", + "content": "[29] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 277, + 294, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 277, + 294, + 320 + ], + "spans": [ + { + "bbox": [ + 56, + 277, + 294, + 320 + ], + "type": "text", + "content": "[30] Kun Liu and Huadong Ma. Exploring background-bias for anomaly detection in surveillance videos. In Proceedings of the 27th ACM International Conference on Multimedia, pages 1490-1499, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 322, + 294, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 322, + 294, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 322, + 294, + 365 + ], + "type": "text", + "content": "[31] Wen Liu, Weixin Luo, Dongze Lian, and Shenghua Gao. Future frame prediction for anomaly detection-a new baseline. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6536-6545, 2018. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 367, + 294, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 367, + 294, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 367, + 294, + 409 + ], + "type": "text", + "content": "[32] Yi Liu, Limin Wang, Yali Wang, Xiao Ma, and Yu Qiao. Fineaction: A fine-grained video dataset for temporal action localization. IEEE transactions on image processing, 31: 6937-6950, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 411, + 294, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 475 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 475 + ], + "type": "text", + "content": "[33] Zhian Liu, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. A hybrid video anomaly detection framework via memory-augmented flow reconstruction and flow-guided frame prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 13588-13597, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 478, + 294, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 294, + 520 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 294, + 520 + ], + "type": "text", + "content": "[34] Cewu Lu, Jianping Shi, and Jiaya Jia. Abnormal event detection at 150 fps in matlab. In Proceedings of the IEEE international conference on computer vision, pages 2720-2727, 2013. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 523, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 294, + 555 + ], + "type": "text", + "content": "[35] Jian Lu, Xuanfeng Li, Bo Zhao, and Jian Zhou. A review of skeleton-based human action recognition. Journal of Image and Graphics, 28(12):3651-3669, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 557, + 294, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 294, + 589 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 294, + 589 + ], + "type": "text", + "content": "[36] Hui Lv and Qianru Sun. Video anomaly detection and explanation via large language models. arXiv preprint arXiv:2401.05702, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 591, + 294, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 591, + 294, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 591, + 294, + 645 + ], + "type": "text", + "content": "[37] Hui Lv, Zhongqi Yue, Qianru Sun, Bin Luo, Zhen Cui, and Hanwang Zhang. Unbiased multiple instance learning for weakly supervised video anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8022-8031, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "text", + "content": "[38] Baiteng Ma, Shiwei Zhang, Changxin Gao, and Nong Sang. Temporal global correlation network for end-to-end action proposal generation. Acta Electronica Sinica, 50(10):2452-2461, 2022. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[39] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "text", + "content": "understanding via large vision and language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (ACL 2024), 2024. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 108, + 553, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 108, + 553, + 151 + ], + "spans": [ + { + "bbox": [ + 316, + 108, + 553, + 151 + ], + "type": "text", + "content": "[40] Ramin Mehran, Alexis Oyama, and Mubarak Shah. Abnormal crowd behavior detection using social force model. In 2009 IEEE conference on computer vision and pattern recognition, pages 935-942. IEEE, 2009. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 153, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 153, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 153, + 553, + 206 + ], + "type": "text", + "content": "[41] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311–318, 2002. 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 209, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 553, + 252 + ], + "type": "text", + "content": "[42] Yujiang Pu, Xiaoyu Wu, Lulu Yang, and Shengjin Wang. Learning prompt-enhanced context features for weakly-supervised video anomaly detection. IEEE Transactions on Image Processing, 2024. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 255, + 553, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 255, + 553, + 329 + ], + "spans": [ + { + "bbox": [ + 316, + 255, + 553, + 329 + ], + "type": "text", + "content": "[43] Zhiwu Qing, Shiwei Zhang, Ziyuan Huang, Yi Xu, Xiang Wang, Mingqian Tang, Changxin Gao, Rong Jin, and Nong Sang. Learning from untrimmed videos: Self-supervised video representation learning with hierarchical consistency. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13821-13831, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 332, + 553, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 553, + 397 + ], + "type": "text", + "content": "[44] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 399, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 453 + ], + "type": "text", + "content": "[45] Dian Shao, Yue Zhao, Bo Dai, and Dahua Lin. Finegym: A hierarchical video dataset for fine-grained action understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2616-2625, 2020. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 456, + 555, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 555, + 499 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 555, + 499 + ], + "type": "text", + "content": "[46] Waqas Sultani, Chen Chen, and Mubarak Shah. Real-world anomaly detection in surveillance videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6479-6488, 2018. 2, 3, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 501, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 501, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 501, + 553, + 555 + ], + "type": "text", + "content": "[47] Jiaqi Tang, Hao Lu, Ruizheng Wu, Xiaogang Xu, Ke Ma, Cheng Fang, Bin Guo, Jiangbo Lu, Qifeng Chen, and YingCong Chen. Hawk: Learning to understand open-world video anomalies. arXiv preprint arXiv:2405.16886, 2024. 2,3,4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 557, + 553, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 557, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 316, + 557, + 553, + 622 + ], + "type": "text", + "content": "[48] Kamalakar Vijay Thakare, Yash Raghuwanshi, Debi Prosad Dogra, Heeseung Choi, and Ig-Jae Kim. Dyannet: A scene dynamicity guided self-trained video anomaly detection network. In Proceedings of the IEEE/CVF Winter conference on applications of computer vision, pages 5541-5550, 2023. 2, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 624, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 689 + ], + "type": "text", + "content": "[49] Yu Tian, Guansong Pang, Yuanhong Chen, Rajvinder Singh, Johan W Verjans, and Gustavo Carneiro. Weakly-supervised video anomaly detection with robust temporal feature magnitude learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4975-4986, 2021. 2, 5, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[50] Anil Osman Tur, Nicola Dall'Asen, Cigdem Bayan, and Elisa Ricci. Exploring diffusion models for unsupervised" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13852" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 715 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 77, + 72, + 296, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 72, + 296, + 105 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 296, + 105 + ], + "type": "text", + "content": "video anomaly detection. In 2023 IEEE International Conference on Image Processing (ICIP), pages 2540-2544. IEEE, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 296, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 296, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 296, + 150 + ], + "type": "text", + "content": "[51] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4566-4575, 2015. 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 295, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 295, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 295, + 194 + ], + "type": "text", + "content": "[52] Jue Wang and Anoop Cherian. Gods: Generalized one-class discriminative subspaces for anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8201-8211, 2019. 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 194, + 295, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 194, + 295, + 270 + ], + "spans": [ + { + "bbox": [ + 56, + 194, + 295, + 270 + ], + "type": "text", + "content": "[53] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 270, + 295, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 270, + 295, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 270, + 295, + 304 + ], + "type": "text", + "content": "[54] SC Wang, Q Huang, YF Zhang, X Li, YQ Nie, and GC Luo. Review of action recognition based on multimodal data. Image Graph, 27(11):3139-3159, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 304, + 295, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 304, + 295, + 350 + ], + "spans": [ + { + "bbox": [ + 56, + 304, + 295, + 350 + ], + "type": "text", + "content": "[55] Jhih-Ciang Wu, He-Yen Hsieh, Ding-Jie Chen, Chiou-Shann Fuh, and Tyng-Luh Liu. Self-supervised sparse representation for video anomaly detection. In European Conference on Computer Vision, pages 729-745. Springer, 2022. 2, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 350, + 295, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 295, + 425 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 295, + 425 + ], + "type": "text", + "content": "[56] Peng Wu, Jing Liu, Yujia Shi, Yujia Sun, Fangtao Shao, Zhaoyang Wu, and Zhiwei Yang. Not only look, but also listen: Learning multimodal violence detection under weak supervision. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXX 16, pages 322-339. Springer, 2020. 2, 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 426, + 295, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 295, + 482 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 295, + 482 + ], + "type": "text", + "content": "[57] Peng Wu, Xuerong Zhou, Guansong Pang, Yujia Sun, Jing Liu, Peng Wang, and Yanning Zhang. Open-vocabulary video anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18297-18307, 2024. 2, 3, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 481, + 295, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 481, + 295, + 547 + ], + "spans": [ + { + "bbox": [ + 56, + 481, + 295, + 547 + ], + "type": "text", + "content": "[58] Peng Wu, Xuerong Zhou, Guansong Pang, Lingru Zhou, Qingsen Yan, Peng Wang, and Yanning Zhang. Vadclip: Adapting vision-language models for weakly supervised video anomaly detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6074-6082, 2024. 2, 3, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 547, + 295, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 547, + 295, + 592 + ], + "spans": [ + { + "bbox": [ + 56, + 547, + 295, + 592 + ], + "type": "text", + "content": "[59] Dan Xu, Yan Yan, Elisa Ricci, and Nicu Sebe. Detecting anomalous events in videos by learning deep representations of appearance and motion. Computer Vision and Image Understanding, 156:117-127, 2017. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 592, + 295, + 636 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 295, + 636 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 295, + 636 + ], + "type": "text", + "content": "[60] Xiaohao Xu, Jinglu Wang, Xiang Ming, and Yan Lu. Towards robust video object segmentation with adaptive object calibration. In Proceedings of the 30th ACM International Conference on Multimedia, pages 2709-2718, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 636, + 295, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 295, + 690 + ], + "type": "text", + "content": "[61] Zhiwei Yang, Jing Liu, Zhaoyang Wu, Peng Wu, and Xiaotao Liu. Video event restoration based on keyframes for video anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14592-14601, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 690, + 295, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 690, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 56, + 690, + 295, + 715 + ], + "type": "text", + "content": "[62] Zhiwei Yang, Jing Liu, and Peng Wu. Text prompt with normality guidance for weakly supervised video anomaly detec" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 700 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 72, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 72, + 553, + 106 + ], + "type": "text", + "content": "tion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18899-18908, 2024. 2, 3, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 106, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 106, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 106, + 553, + 162 + ], + "type": "text", + "content": "[63] Yu Yao, Xizi Wang, Mingze Xu, Zelin Pu, Yuchen Wang, Ella Atkins, and David J Crandall. Dota: Unsupervised detection of traffic anomaly in driving videos. IEEE transactions on pattern analysis and machine intelligence, 45(1): 444-459, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 162, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 162, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 162, + 553, + 207 + ], + "type": "text", + "content": "[64] Tongtong Yuan, Xuange Zhang, Kun Liu, Bo Liu, Chen Chen, Jian Jin, and Zhenzhen Jiao. Towards surveillance video-and-language understanding: New dataset, baselines, and challenges, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 207, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 207, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 207, + 553, + 274 + ], + "type": "text", + "content": "[65] Tongtong Yuan, Xuange Zhang, Kun Liu, Bo Liu, Chen Chen, Jian Jin, and Zhenzhen Jiao. Towards surveillance video-and-language understanding: New dataset baselines and challenges. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22052-22061, 2024. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 274, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 274, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 274, + 553, + 330 + ], + "type": "text", + "content": "[66] M Zaigham Zaheer, Arif Mahmood, M Haris Khan, Mattia Segu, Fisher Yu, and Seung-Ik Lee. Generative cooperative learning for unsupervised video anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14744-14754, 2022. 2, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 330, + 553, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 330, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 316, + 330, + 553, + 386 + ], + "type": "text", + "content": "[67] Luca Zanella, Willi Menapace, Massimiliano Mancini, Yiming Wang, and Elisa Ricci. Harnessing large language models for training-free video anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18527-18536, 2024. 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 386, + 553, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 386, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 386, + 553, + 420 + ], + "type": "text", + "content": "[68] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 2, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 420, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 420, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 420, + 553, + 464 + ], + "type": "text", + "content": "[69] Huaxin Zhang, Xiang Wang, Xiaohao Xu, Zhiwu Qing, Changxin Gao, and Nong Sang. Hr-pro: Point-supervised temporal action localization via hierarchical reliability propagation. arXiv preprint arXiv:2308.12608, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 464, + 553, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 464, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 316, + 464, + 553, + 520 + ], + "type": "text", + "content": "[70] Huaxin Zhang, Xiang Wang, Xiaohao Xu, Xiaonan Huang, Chuchu Han, Yuehuan Wang, Changxin Gao, Shanjun Zhang, and Nong Sang. Glancevad: Exploring glance supervision for label-efficient video anomaly detection. arXiv preprint arXiv:2403.06154, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 520, + 553, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 520, + 553, + 553 + ], + "spans": [ + { + "bbox": [ + 316, + 520, + 553, + 553 + ], + "type": "text", + "content": "[71] Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, and Chunyuan Li. Video instruction tuning with synthetic data. arXiv preprint arXiv:2410.02713, 2024. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 554, + 553, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 554, + 553, + 588 + ], + "spans": [ + { + "bbox": [ + 316, + 554, + 553, + 588 + ], + "type": "text", + "content": "[72] Bin Zhao, Li Fei-Fei, and Eric P Xing. Online detection of unusual events in videos via dynamic sparse coding. In CVPR 2011, pages 3313-3320. IEEE, 2011. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 588, + 553, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 588, + 553, + 654 + ], + "spans": [ + { + "bbox": [ + 316, + 588, + 553, + 654 + ], + "type": "text", + "content": "[73] Jia-Xing Zhong, Nannan Li, Weijie Kong, Shan Liu, Thomas H Li, and Ge Li. Graph convolutional label noise cleaner: Train a plug-and-play action classifier for anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1237-1246, 2019. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 654, + 553, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 654, + 553, + 700 + ], + "spans": [ + { + "bbox": [ + 316, + 654, + 553, + 700 + ], + "type": "text", + "content": "[74] Hang Zhou, Junqing Yu, and Wei Yang. Dual memory units with uncertainty regulation for weakly supervised video anomaly detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3769-3777, 2023. 2, 5, 6" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "13853" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/HomoGen_ Enhanced Video Inpainting via Homography Propagation and Diffusion/c0acdc17-6ead-4f9d-8631-4f1535650ac6_content_list.json b/2025/HomoGen_ Enhanced Video Inpainting via Homography Propagation and Diffusion/c0acdc17-6ead-4f9d-8631-4f1535650ac6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..cad384640117dbff763617e65e8d4bd0264946f6 --- /dev/null +++ b/2025/HomoGen_ Enhanced Video Inpainting via Homography Propagation and Diffusion/c0acdc17-6ead-4f9d-8631-4f1535650ac6_content_list.json @@ -0,0 +1,1664 @@ +[ + { + "type": "text", + "text": "HomoGen: Enhanced Video Inpainting via Homography Propagation and Diffusion", + "text_level": 1, + "bbox": [ + 212, + 130, + 784, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ding Ding $^{1*}$ Yueming Pan $^{2}$ Ruoyu Feng $^{3}$ Qi Dai $^{4}$ Kai Qiu $^{4}$ Jianmin Bao $^{4}$ Chong Luo $^{4\\dagger}$ Zhenzhong Chen $^{1}$ $^{1}$ Wuhan University $^{2}$ Xi'an Jiaotong University $^{3}$ University of Science and Technology of China $^{4}$ Microsoft Research Asia", + "bbox": [ + 99, + 200, + 903, + 273 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bd84f48f97310edc40f31c56d8b40d7befc8a29a83109163af082a82ceeae6d0.jpg", + "image_caption": [ + "Figure 1. Two-stage pipelines of HomoGen and the state-of-the-art method. Previous flow-based propagation warps pixels across frames, followed by a video transformer (VT) that generates remaining missing content. In contrast, HomoGen's homography-based propagation warps pixels as priors to control content generation in a video diffusion model (VDM) that filters out artifacts in the priors during denoising." + ], + "image_footnote": [], + "bbox": [ + 93, + 292, + 903, + 508 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 577, + 326, + 592 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we present HomoGen, an enhanced video inpainting method based on homography propagation and diffusion models. HomoGen leverages homography registration to propagate contextual pixels as priors for generating missing content in corrupted videos. Unlike previous flow-based propagation methods, which introduce local distortions due to point-to-point optical flows, homography-induced artifacts are typically global structural distortions that preserve semantic integrity. To effectively utilize these priors for generation, we employ a video diffusion model that inherently prioritizes semantic information within the priors over pixel-level details. A content-adaptive control mechanism is proposed to scale and inject the priors into intermediate video latents during iterative denoising. In contrast to existing transformer-based networks that often suffer from artifacts within priors, leading to error accumulation and unrealistic results, our denoising diffusion net-", + "bbox": [ + 88, + 609, + 485, + 864 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "work can smooth out artifacts and ensure natural outputs. Extensive experiments demonstrate the effectiveness of the proposed method qualitatively and quantitatively.", + "bbox": [ + 511, + 578, + 906, + 625 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 648, + 643, + 664 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video inpainting aims to fill corrupted regions in a video with realistic and spatiotemporally coherent content [2]. It has diverse applications including object removal [7], video restoration [43], autonomous driving [23], etc. Despite significant progress in image inpainting [29, 35, 57], video inpainting remains full of challenges due to the complexity of video scenarios. The primary difficulty lies in identifying spatiotemporal guidance for generating content within corrupted regions, particularly in videos with diverse camera movements and object motion. To tackle this, previous work has explored various technologies, such as 3D convolutions [4, 13] and attentions [21, 24, 58], to leverage spatiotemporal context and preserve coherent across frames.", + "bbox": [ + 509, + 674, + 906, + 871 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, two-stage methods combining flow-based propagation and generation networks have shown impres", + "bbox": [ + 511, + 872, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*This work was done when Ding Ding was an intern at MSRA.", + "bbox": [ + 107, + 875, + 442, + 888 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 109, + 888, + 235, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22953", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sive performance [22, 59, 63]. In flow-based propagation, optical flows are utilized to propagate pixels from adjacent frames into corrupted regions via warping. For generation networks, powerful video transformers are employed to generate remaining missing content. However, such methods still face challenges: the warping operation introduces artifacts such as blur and distortion in the propagated pixels, especially in scenarios characterized by significant object or camera motion and lighting changes [9]. As shown in Fig. 1, the flow-based propagation method in the state-of-the-art ProPainter [63] incorrectly warps a tree into the road center in the first stage. Besides, the propagated pixels are considered partial inpainting results to guide remaining content generation. Such artifacts cause error accumulation in video transformers, leading to unrealistic results [22, 59]. To mitigate artifacts in the propagated pixels, recent methods propose the flow consistency check to avoid inaccurate flows [56]. However, the consistency check discards partial flows, leaving the propagated pixels insufficient and semantically incoherent. As shown in Fig. 1, a subset of road pixels is removed from the propagated content after the consistency check. Therefore, video transformers struggle to generate high-fidelity content without sufficient guidance.", + "bbox": [ + 89, + 90, + 480, + 438 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Regarding the challenges in balancing retaining more propagated pixels while avoiding artifacts, diffusion models [12, 35] offer a promising solution as generation networks in the two-stage pipeline. Diffusion models prioritize semantic information within input conditions, such as stroke paintings and unnatural images [25, 28, 61], over pixel-level details. During denoising, diffusion models could generate realistic content while filtering out noisy low-level features from the conditions. Besides, homography registration for pixel propagation is more compatible with diffusion models. Specifically, artifacts from homography differ from those caused by flows. Flows capture point-to-point correlations [44], and flow-induced artifacts often considerably distort parts of objects in complex scenes, disrupting the semantics of propagated content. In contrast, homography imposes projective geometric constraints based on the global structure between frames [42], and homography-induced artifacts are typically structural distortions while preserving semantic integrity. As shown in Fig. 1, in the first stage, although several road pixels experience overall distortions, the structural integrity of objects remains largely intact. In the second stage, the video diffusion model generates more natural content while filtering out the artifacts.", + "bbox": [ + 91, + 443, + 482, + 789 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose an enhanced video inpainting framework, named HomoGen, which enables pixel propagation via Homography registration and designs a video diffusion model to Generate missing content. For pixel propagation, we register paired frames via homography matrices to explore valuable contextual pixels. After warping the reference frame, the propagated pixels located in the", + "bbox": [ + 89, + 795, + 480, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "corrupted regions of the target frame are extracted as priors. We iterate the propagation process across all frames in a bidirectional manner, where the reference frames are progressively updated with the propagated pixels at each step. For content generation, we design a prior controlled video diffusion model that captures the semantics within the bidirectional priors. Built upon a text-guided image inpainting model from Stable Diffusion, we integrate temporal layers into it to construct a 3D network. To effectively utilize the priors, we propose a content-adaptive control mechanism that scales and injects the priors into the diffusion model. Specifically, we design a lite ControlNet [61] to embed the priors into multi-scale features that are then fed into the 3D U-Net via dense injection. Learned scalers adjust the multi-scale features before injection, adapting to the spatiotemporal context of priors and the sampling steps. Simultaneously, the denoising process filters out artifacts within the priors during iterative sampling, thus preserving high-fidelity results across inpainted regions and visible pixels.", + "bbox": [ + 511, + 90, + 903, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Qualitative and quantitative experiment results demonstrate the superiority of HomoGen both in video completion and object removal. Compared to the state-of-the-art methods, HomoGen can generate more realistic content and improve performance by $23.81\\%$ in VFID.", + "bbox": [ + 511, + 378, + 903, + 455 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A propagate-and-diffuse framework is proposed for video inpainting, which leverages homography registration to propagate contextual pixels as priors to control the generation of missing content in a denoising diffusion network.", + "- A content-adaptive control mechanism is designed to inject prior information into the diffusion model. The dense injection learns to adapt to the spatiotemporal context of priors and the sampling steps, while artifacts within the priors are filtered out during iterative denoising.", + "- Extensive experiments show that HomoGen surpasses the state-of-the-art methods and generates more realistic content in video completion and object removal tasks." + ], + "bbox": [ + 513, + 455, + 903, + 636 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 654, + 653, + 669 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Diffusion-based Image and Video Generation", + "text_level": 1, + "bbox": [ + 511, + 681, + 895, + 696 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion models have exhibited remarkable abilities in visual generation [3, 6, 15, 31, 35, 45, 49, 51, 53, 54]. These models learn to approximate data distributions through the iterative denoising of diffused inputs [12, 41]. Their practicality is enhanced by the incorporation of various conditions, e.g., text prompts, unnatural images, segmentation maps, etc., allowing for controlled outputs over the generative process. Controllable generation methods particularly benefit from additional structure modules, e.g., ControlNet [61], T2I-adapter [26], etc., which manage the injection of conditions. Such models leverage large-scale datasets [1, 39, 62] to achieve impressive results not only in generation but in editing [8, 11, 32, 46]. Conditional diffusion can", + "bbox": [ + 511, + 703, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22954", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "also be applied to video inpainting, a widely explored video editing task, where the visible pixels serve as conditions to guide the generation of missing content in corrupted videos.", + "bbox": [ + 89, + 90, + 482, + 137 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Video Inpainting", + "text_level": 1, + "bbox": [ + 89, + 145, + 259, + 162 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Previous video inpainting methods can be broadly divided into two categories: one-stage and two-stage approaches. One-stage approaches employ 3D CNNs and video transformers to aggregate spatiotemporal context for generating missing content in videos [4, 21, 24, 34, 58]. These methods usually struggle to capture spatiotemporal details due to limited receptive fields, which leads to blurry and inconsistent outcomes in complex scenes. Two-stage approaches have gradually become mainstream choices for video inpainting. The first stage primarily extracts prior information from the video itself, such as motion information [17, 43, 56], homography [20], low-resolution results [47], and proposal sets [13]. Guided by the priors, the second stage generates missing content based on well-designed networks, such as image inpainting models [9, 56, 60], 3D CNNs [17] and transformers [59]. Recent methods capitalize on the advantages of optical flows for motion capture and video transformers for spatiotemporal context modeling [9, 63]. As a typical approach, ProPainter [63] begins with flow extraction and completion, and utilizes completed flows to propagate pixels into corrupted regions. A flow-guided video transformer is designed to refine the propagated pixels and generate remaining missing content.", + "bbox": [ + 89, + 167, + 483, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, such methods still suffer from blur and distortion in the results, primarily caused by the warping operation in pixel propagation and error accumulation in video transformers. To address this issue, we design a video diffusion model as the generation network that effectively removes artifacts within the priors during iterative denoising. Additionally, we employ homography-based pixel propagation to preserve the semantic integrity of warped content, providing fine-grained controls for conditional diffusion.", + "bbox": [ + 89, + 515, + 483, + 650 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 662, + 225, + 680 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "HomoGen consists of two main components: Homography-based Pixel Propagation (HPP) and Prior Controlled Video Diffusion Model (PCVDM), as shown in Fig. 2. In HPP, we estimate homography between paired frames for propagating pixels as contextual priors. PCVDM then generates missing content controlled by these priors.", + "bbox": [ + 89, + 688, + 483, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 89, + 787, + 220, + 804 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Consider a corrupted video sequence denoted as $\\mathbf{x} = \\{\\mathbf{x}_i\\mid$ $\\mathbf{x}_i\\in \\mathbb{R}^{3\\times H\\times W},1\\leq i\\leq N\\}$ comprising $N$ frames, accompanied by a binary mask sequence $\\mathbf{m} = \\{\\mathbf{m}_i\\mid \\mathbf{m}_i\\in$ $\\mathbb{R}^{1\\times H\\times W},1\\leq i\\leq N\\}$ identifying corrupted regions in all frames. Text prompts provide visual descriptions of the inpainted video. With text prompts as additional conditions,", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "text-guided video inpainting aims to generate visually faithful and coherent content within the corrupted regions.", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As text-guided video inpainting encounters diffusion models (DMs), latent diffusion models (LDMs) [31, 35] are effective choices. LDMs leverage a pre-trained Variational Autoencoder (VAE) that consists of an encoder $\\mathcal{E}$ and a decoder $\\mathcal{D}$ to perform denoising in the latent space where the latent variable $\\mathbf{z} = \\mathcal{E}(\\mathbf{x})$ . For video inpainting, conditions $\\mathbf{c}$ are injected into a denoising network $\\epsilon_{\\theta}$ to predict $\\mathbf{z}^t$ , a denoised variant of a noise, where $t = 0, 1, 2, \\dots, T$ denotes the sampling step. The optimization objective is formulated as", + "bbox": [ + 511, + 121, + 906, + 268 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} _ {\\mathbf {z} ^ {0}, t, \\epsilon \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I})} \\left[ \\left\\| \\epsilon - \\epsilon_ {\\theta} \\left(\\mathbf {z} ^ {t}, t, \\mathbf {c}\\right) \\right\\| _ {2} ^ {2} \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 268, + 903, + 292 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{c} = (\\mathbf{x},\\mathbf{m},\\tau_{\\theta}(y))$ $y$ represents the text prompts and $\\tau_{\\theta}(\\cdot)$ denotes a text encoder.", + "bbox": [ + 511, + 297, + 905, + 329 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Homography-based Pixel Propagation", + "text_level": 1, + "bbox": [ + 511, + 338, + 841, + 354 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Pixel propagation aligns content within corrupted videos by propagating pixels from reference frames into corrupted regions in the target frame. Flow-based propagation typically warps pixels based on the trajectories of optical flows [17, 59]. However, they face challenges in balancing more propagated pixels and fewer distortions that cause error accumulation within the generation network. To overcome this, we employ homography-based propagation for content alignment. Compared to flows that capture point-to-point correlations, homography registration offers greater robustness since homography matrices for warping impose projective geometric constraints based on global structures between frames [42]. Besides, homography registration usually preserves the semantic integrity without altering parts of objects in complex scenes, as discussed in Sec. 1.", + "bbox": [ + 511, + 359, + 906, + 585 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To search for valuable pixels within a large temporal window, we estimate the forward and backward homography and iteratively propagate pixels across frames in a bidirectional manner. For simplicity, we only describe the forward propagation process. The propagated pixels $\\mathbf{x}_{i,i + 1}^p$ in frame $\\mathbf{x}_{i + 1}$ are warped from frame $\\mathbf{x}_i$ as follows:", + "bbox": [ + 511, + 587, + 905, + 676 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {i, i + 1} ^ {p} = \\left(1 - \\mathbf {m} _ {i, i + 1} ^ {p}\\right) \\odot \\mathbf {x} _ {i, i + 1}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 688, + 903, + 707 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {i, i + 1} = \\mathcal {W} \\left(\\mathbf {x} _ {i} + \\mathbf {x} _ {i - 1, i} ^ {p}, \\mathcal {H} _ {i, i + 1}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 709, + 903, + 727 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {m} _ {i, i + 1} ^ {p} = 1 - \\mathbf {m} _ {i, i + 1} \\odot \\mathbf {m} _ {i + 1}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 729, + 903, + 748 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {m} _ {i, i + 1} = \\mathcal {W} \\left(\\mathbf {m} _ {i} \\odot \\mathbf {m} _ {i - 1, i} ^ {p}, \\mathcal {H} _ {i, i + 1}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 751, + 903, + 770 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{x}_{i,i + 1}$ and $\\mathbf{m}_{i,i + 1}$ denotes warped $\\mathbf{x}_i$ and $\\mathbf{m}_i$ registered with $\\mathbf{x}_{i + 1}$ and $\\mathbf{m}_{i + 1}$ , respectively. $\\mathbf{m}_{i + 1}^p$ denotes the mask indicating the regions of the propagated pixels. $\\odot$ denotes the Hadamard product. $\\mathcal{W}(\\cdot ,\\cdot)$ represents the warping operation. $\\mathcal{H}_{i,i + 1}$ denotes the homography matrix for projecting the plane of $\\mathbf{x}_i$ onto $\\mathbf{x}_{i + 1}$ . $\\mathcal{H}_{i,i + 1}$ is estimated by RANSAC on ORB feature matches [37]. During propagation, each reference frame is registered to the target frame.", + "bbox": [ + 511, + 780, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22955", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ddf6d486effa5817a165933c4503e8f1b630d985dfe509cec76d289d51e5c19c.jpg", + "image_caption": [ + "Figure 2. Overview of the proposed HomoGen framework, comprising two primary components: Homography-based Pixel Propagation (HPP) and Prior Controlled Video Diffusion Model (PCVDM). In this two-stage pipeline, HPP leverages homography between frames to propagate pixels as contextual priors that control missing content generation in PCVDM." + ], + "image_footnote": [], + "bbox": [ + 94, + 92, + 903, + 354 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d8eeff36c4947b82c41e32db08a1c1b12c56feb5635c0e448dfb27d74ad8cb0f.jpg", + "image_caption": [ + "(a) Flow-based propagation" + ], + "image_footnote": [], + "bbox": [ + 91, + 412, + 285, + 489 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fd7268ed549ef194b14fbad67b15b4bbb6eecd5debb1fd55e294f50aedfaeb6b.jpg", + "image_caption": [ + "(b) HPP", + "Figure 3. Statistics of warping error and propagated pixel count on the YouTube-VOS test set in video completion. Warping error denotes the Mean Absolute Error (MAE) between the ground truth (normalized to $[-1,1]$ ) and the propagated pixels within a video. Pixel count denotes the proportion of propagated pixels in the corrupted regions. The grid color indicates the number of videos." + ], + "image_footnote": [], + "bbox": [ + 289, + 412, + 480, + 489 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After warping based on homography matrices, the propagated pixels located in the corrupted regions are extracted as priors. We iterate the propagation process across all frames, where the reference frames are progressively updated with the propagated pixels at each step.", + "bbox": [ + 89, + 607, + 483, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Valuable pixels are propagated under homography constraints. For each frame, we blend forward and backward propagated pixels $\\mathbf{x}_{i - 1,i}^p$ and $\\mathbf{x}_{i + 1,i}^p$ as follows:", + "bbox": [ + 89, + 681, + 483, + 729 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {i} ^ {p} = \\mathbf {x} _ {i - 1, i} ^ {p} + \\left(1 - \\mathbf {m} _ {i - 1, i} ^ {p}\\right) \\odot \\mathbf {m} _ {i + 1, i} ^ {p} \\odot \\mathbf {x} _ {i + 1, i} ^ {p}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 737, + 482, + 756 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{x}_i^p$ denotes the bidirectional propagated pixels as the priors to control missing content generation in PCVDM.", + "bbox": [ + 89, + 763, + 482, + 795 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For further analysis, we present statistics on warping error and the count of propagated pixels for flow-based propagation and HPP. As shown in Fig. 3, the flow-based method propagates fewer pixels with lower warping error due to the consistency check, while HPP increases the pixel count at the cost of slightly higher warping error. Despite this added error, PCVDM benefits from richer spatiotemporal context", + "bbox": [ + 89, + 796, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and complete semantic information while filtering out artifacts in such priors during iterative denoising.", + "bbox": [ + 511, + 416, + 906, + 446 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Prior Controlled Video Diffusion Model", + "text_level": 1, + "bbox": [ + 511, + 457, + 856, + 472 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the priors from HPP, PCVDM aims to generate spatiotemporal coherent and realistic content in the corrupted regions. We propose a video diffusion model based on Stable Diffusion inpainting [35] and design a content-adaptive control mechanism to inject the priors into DM.", + "bbox": [ + 511, + 479, + 906, + 556 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.1. Inflating Image Inpainter to Video Inpainter", + "text_level": 1, + "bbox": [ + 511, + 565, + 870, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The proposed video inpainting model is extended from the image inpainting model in Stable Diffusion. Following previous work [10, 40], we extend the 2D DM into a 3D structure. In VAE decoder [18] and U-Net [36], original 2D layers are converted to pseudo-3D layers to learn temporal correlations, with detailed structures available in the supplementary. Besides, we employ an OpenCLIP image encoder and perceive attention layers to extract GoP-wise CLIP latents from the corrupted video [5, 14], and then inject the latents into the 3D U-Net via cross-attention. GoP refers to a Group of Picture, i.e., a video clip processed per inference. During training, the binary mask $\\mathbf{m}$ , the noised video latent $\\mathbf{z}^t$ , and the corrupted video latent $\\mathbf{z} = \\mathcal{E}(\\mathbf{x})$ are concatenated in the latent space and fed into the 3D U-Net. During inference, $\\mathbf{z}^t$ is initialized as Gaussian noise. The 3D U-Net generates the final sampled latent $\\hat{\\mathbf{z}}$ that is decoded by the 3D VAE decoder as the completed video $\\hat{\\mathbf{x}}$ .", + "bbox": [ + 511, + 584, + 906, + 843 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.2. Content-adaptive Control Mechanism", + "text_level": 1, + "bbox": [ + 511, + 849, + 823, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The priors are essential for generating missing content in corrupted videos. Recent work fills the propagated pix", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22956", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c6a798cf1d5a1c00938b66e0a67fb844ddc816eb4f2cd4a65bbd27f21c92b99f.jpg", + "image_caption": [ + "Figure 4. Illustration of content-adaptive control mechanism that injects prior information into PCVDM. Prior scalers in the dense injection are learned from the spatiotemporal context of priors and the sampling steps, enabling fine-grained controls." + ], + "image_footnote": [], + "bbox": [ + 106, + 88, + 483, + 397 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "els into the corrupted regions as partial inpainting results and uses transformers to generate remaining missing content [22, 59, 63]. However, due to inevitable artifacts introduced by warping, pixel-level error accumulates during token prediction in transformers guided by noisy tokens from the priors, as discussed in Sec. 1. A trade-off solution is to discard inaccurate propagated pixels based on the consistency check [56] while limiting sufficient priors and semantic integrity. Unlike transformers' sensitivity to low-level features, DMs prioritize semantic information within input conditions over pixel-level details. Hence, to ensure the semantic integrity of the propagated content, we retained all propagated pixels without any consistency check. Besides, we consider the priors as conditions instead of partial inpainting results pasted with original visible pixels, which also helps the network unambiguously perceive artifacts from the priors. During iterative sampling, the denoising process can smooth out such undesirable artifacts.", + "bbox": [ + 88, + 472, + 482, + 743 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To effectively utilize the priors, we propose a content-adaptive control mechanism to scale and inject them into PCVDM for controlled missing content generation. The priors $\\mathbf{x}^p = \\{\\mathbf{x}_i^p\\mid \\mathbf{x}_i^p\\in \\mathbb{R}^{3\\times H\\times W},1\\leq i\\leq N\\}$ are first encoded into latents $\\mathbf{z}^p = \\{\\mathbf{z}_i^p\\mid \\mathbf{z}_i^p\\in \\mathbb{R}^{4\\times \\frac{H}{8}\\times \\frac{W}{8}},1\\leq i\\leq N\\}$ . We first design a LiteControlNet to extract multi-scale features from $\\mathbf{z}^p$ for further injection. Modified from ControlNet [61], LiteControlNet features a lightweight structure with one convolutional layer at each scale. Detailed structures of LiteControlNet refer to the supplementary. The", + "bbox": [ + 89, + 747, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "prior feature at each scale is connected to the 3D U-Net encoder through a zero convolutional layer and a learned projector, i.e., a 3D convolutional layer with a kernel size of $1 \\times 1 \\times 1$ . Rather than using additive operation, we introduce a spatiotemporal adaptive de-normalization strategy to inject each feature into the corresponding normalization layer [28]. As shown in Fig. 4, after the $j$ -th latent normalized by 3D group normalization [52] in the 3D U-Net, we compute the output activation $\\mathbf{h}_j$ by de-normalizing the normalized latent $\\bar{\\mathbf{h}}_j$ according to the prior feature $\\mathbf{f}_j$ , formulated as:", + "bbox": [ + 511, + 90, + 906, + 243 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {h} _ {j} = \\gamma_ {j} \\odot \\bar {\\mathbf {h}} _ {j} + \\beta_ {j} + \\bar {\\mathbf {h}} _ {j}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 619, + 247, + 906, + 267 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\gamma_{j}$ and $\\beta_{j}$ are obtained through 3D convolutions of $\\mathbf{f}_j$ . Equipped with this design, $\\mathbf{f}_j$ is scaled and injected into the intermediate latent, where the generation of missing content adapts to the spatiotemporal context of the priors. Additionally, $\\mathbf{f}_j$ is injected after attention blocks and yielded as the $(j + 1)$ -th latent $\\mathbf{h}_{j + 1}$ , formulated as:", + "bbox": [ + 511, + 268, + 906, + 361 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {h} _ {j + 1} = \\zeta (t) \\mathbf {f} _ {j} + \\mathbf {h} _ {j}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 637, + 366, + 906, + 385 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\zeta(\\cdot)$ represents two linear layers interleaved with a SiLU activation [33] in between. $\\zeta(\\cdot)$ encodes the sampling step $t$ [35] into Fourier features, which allows the scaling strength to vary with $t$ for fine-grained controls.", + "bbox": [ + 511, + 388, + 905, + 448 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During iterative denoising, the content in the corrupted regions progressively generates, approaching seamless integration with original visible pixels. After decoding the final sampled latent by the 3D VAE decoder, the generated and corrupted videos are blended with a blurred mask to improve coherence along the mask boundary [16].", + "bbox": [ + 511, + 449, + 905, + 540 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Inpainting for Long Videos", + "text_level": 1, + "bbox": [ + 511, + 546, + 758, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Video inpainting faces the challenge of maintaining temporal coherence across GoPs in long videos that span several seconds. Generative models are often restricted to processing all frames at once due to memory limitations, which leads to variability when sequentially inferring all GoPs. To address this, we introduce temporal references from adjacent GoPs to assist in inpainting the current GoP.", + "bbox": [ + 511, + 568, + 905, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Let $N$ denote the length of the GoP that HomoGen processes per inference. For a video containing $L$ frames, we segment the video into $\\left\\lceil \\frac{L}{N} \\right\\rceil$ GoPs, each with an overlap of $N_{\\text{overlap}}$ frames, and process all GoPs sequentially. For the $i$ -th GoP, the first and last frames are set to the center frames of the $(i-1)$ -th and $(i+1)$ -th GoP, respectively, while $N_{\\text{overlap}} - 2$ frames from the second frame onward are set to the last frames of the $(i-1)$ -th GoP. In the initial inference, the first GoP undergoes inpainting. Before $i$ -th inference $(i > 1)$ , the beginning $N_{\\text{overlap}} - 1$ frames of the $i$ -th GoP are replaced by the corresponding frames from the inpainted $(i-1)$ -th GoP. During $i$ -th inference, HomoGen can capture long-term temporal references for inpainting the $i$ -th GoP. This process iterates until all GoPs are inpainted, which ensures content consistency across the entire video.", + "bbox": [ + 511, + 674, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22957", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/f57e74de431146e3f90729fde5a7af389e42e9bc5dafd1ff91abe0acf8a0425f.jpg", + "table_caption": [ + "Table 1. Quantitative comparisons on YouTube-VOS, DAVIS, and RORD datasets. $\\uparrow$ indicates higher is better. $\\downarrow$ indicates lower is better. $E_{w}^{*}$ denotes $E_{\\text{warp}} \\times 10^{-3}$ . The best performances are highlighted in bold. All methods are evaluated following their default settings." + ], + "table_footnote": [], + "table_body": "
YouTube-VOSDAVISRORD
ModelsPSNR ↑SSIM ↑VFID ↓Ew* ↓PSNR ↑SSIM ↑VFID ↓Ew* ↓PSNR ↑SSIM ↑VFID ↓Ew* ↓
DFVI [56]29.160.94290.0661.50928.810.94040.1871.60831.840.95710.0651.137
CPNet [20]31.580.96070.0711.42330.280.95210.1821.51935.070.97400.0390.879
FGVC [9]29.670.94030.0641.02230.800.94970.1651.58632.760.96530.0521.005
STTN [58]32.570.96800.0490.90730.740.95900.1531.66134.940.97180.0440.872
FuseFormer [24]33.290.96810.0530.89132.540.97020.1381.33635.610.98110.0310.843
FGT [59]32.170.95990.0541.02532.860.96500.1291.32334.210.97240.0380.867
ISVI [60]30.380.94710.0771.31132.200.96260.1661.29233.450.96910.0460.902
E2FGVI [22]33.710.97000.0460.86433.010.97210.1161.31536.040.98000.0270.797
ProPainter [63]34.460.97220.0420.87234.190.97590.0981.19536.760.98670.0280.813
HomoGen34.810.97350.0320.85834.710.97630.0771.23437.290.98720.0230.793
", + "bbox": [ + 94, + 119, + 906, + 321 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4becd39a5eb8747985cf7fe7aa91375719b5d6e523456739e44e9cb36f027baf.jpg", + "image_caption": [ + "(a) PSNR vs. mask ratio", + "Figure 5. Comparisons on different mask ratios." + ], + "image_footnote": [], + "bbox": [ + 91, + 323, + 285, + 435 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4838083238cf1141b23851635ff3c653e02081c431a925397b5e6a791a967362.jpg", + "image_caption": [ + "(b) VFID vs. mask ratio" + ], + "image_footnote": [], + "bbox": [ + 285, + 321, + 480, + 435 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/afad4e4cc0212255f4262b2392d4e8844a12645be75bcb86adf6cf90eabf279d.jpg", + "image_caption": [ + "(a) PSNR vs. motion magnitude", + "Figure 6. Comparisons on different motion magnitude." + ], + "image_footnote": [], + "bbox": [ + 514, + 323, + 707, + 436 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/47e2665b9cf4430ceaaee5b1effe8058644f697c19637b80da99128a310f4bb9.jpg", + "image_caption": [ + "(b) VFID vs. motion magnitude" + ], + "image_footnote": [], + "bbox": [ + 712, + 321, + 903, + 436 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 484, + 223, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 511, + 307, + 527 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.1. Datasets", + "text_level": 1, + "bbox": [ + 89, + 535, + 196, + 549 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use an internal video dataset including 5 million video sequences from Shutterstock1 as the training set. To simulate masks in video completion and object removal tasks, we synthesize stationary and object-aware masks in a random fashion, which are then pasted onto frames [17, 22, 63]. For evaluation, we use two video object segmentation datasets: DAVIS [30] and YouTube-VOS [55], which consist of 508 and 90 video sequences in test sets, respectively. For the DAVIS test set, following ProPainter [63], we use 50 video sequences for metric calculations. For mask settings, stationary masks are used to calculate objective metrics for the video completion task, and object-aware masks extracted from segmentation annotations are adopted for qualitative comparisons in the object removal task. Additionally, we adopt the Real-world Object Removal Dataset (RORD) [38] to evaluate the object removal task. The RORD test set includes 343 video pairs with and without objects, along with object mask annotations. This dataset addresses the lack of references for computing objective metrics in terms of the object removal task in previous work. For all datasets, video frames are sized to $240 \\times 432$ for evaluation. Text prompts", + "bbox": [ + 89, + 554, + 483, + 875 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "for test sets are generated by GPT-4o [27] using the center frame of each corrupted video.", + "bbox": [ + 511, + 484, + 906, + 516 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.2. Metrics", + "text_level": 1, + "bbox": [ + 511, + 525, + 612, + 539 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following widely adopted protocols, we use PSNR, SSIM [50], VFID [48], and warping error $E_{\\text{warp}}$ [19] as evaluation metrics. Specifically, PSNR and SSIM are frequently used for distortion-oriented video assessment. VFID measures the perceptual similarity between input videos and outputs. $E_{\\text{warp}}$ assesses the temporal stability.", + "bbox": [ + 511, + 544, + 906, + 635 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Detailed configurations of the proposed model for training and inference refer to the supplementary.", + "bbox": [ + 511, + 635, + 905, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparisons", + "text_level": 1, + "bbox": [ + 511, + 676, + 653, + 693 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.1. Quantitative Results", + "text_level": 1, + "bbox": [ + 511, + 699, + 702, + 714 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We report quantitative results on three datasets, using stationary masks for the YouTube-VOS and DAVIS datasets, and object-aware masks for the RORD dataset. We compare the proposed HomoGen with previous video inpainting methods, including DFVI [56], CPNet [20], FGVC [9], STTN [58], FuseFormer [24], FGT [59], ISVI [60], $\\mathrm{E}^2\\mathrm{FGVI}$ [22], and ProPainter [63]. As shown in Tab. 1, HomoGen outperforms previous methods across most quantitative metrics. Notably, HomoGen significantly surpasses the state-of-the-art method by $23.81\\%$ , $21.43\\%$ , and $17.86\\%$ in VFID on the YouTube-VOS, DAVIS, and RORD datasets, respectively, demonstrating its ability to generate", + "bbox": [ + 511, + 719, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "1https://wwwCRETstorstock.com", + "bbox": [ + 107, + 886, + 339, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22958", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0ca96261336b8c6677971935e684d72b1734e2012eb516d3fe41a3a528d28db4.jpg", + "image_caption": [ + "Figure 7. Qualitative results compared with FuseFormer [24], $\\mathrm{E}^2\\mathrm{FGVI}$ [22], and ProPainter [63]." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 903, + 388 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/89c92b7ef84712dd74a1c2352c361e41be8bc4f8ebd6ff61b3edeb4fcc67ec95.jpg", + "image_caption": [ + "Figure 8. Comparative user study on HomoGen and ProPainter." + ], + "image_footnote": [], + "bbox": [ + 96, + 417, + 478, + 526 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "videos with more visually plausible and realistic content.", + "bbox": [ + 89, + 556, + 467, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To assess the model performance across different scenarios, we evaluate HomoGen on videos with varying mask ratios, as well as camera and object motion. We select ProPainter for comparisons, given its best performance among the baselines. In evaluation for different mask ratios, we partition the mask ratio range from $[0, 0.5]$ into five equal intervals, each corresponding to a synthesized mask set. We randomly sample 30 video sequences for each mask set. As shown in Fig. 5, HomoGen performs on par with ProPainter when the mask ratio is low. As the mask ratio increases, ProPainter's performance degrades more substantially than HomoGen's. In evaluation for different camera and object motion, we define the average pixel displacement $D^{p}$ as a measure of motion, formulated as:", + "bbox": [ + 89, + 571, + 483, + 782 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nD ^ {p} = \\frac {1}{N H W} \\sum_ {n = 1} ^ {N} \\sum_ {i = 1} ^ {H} \\sum_ {j = 1} ^ {W} \\sqrt {\\left(d x _ {i , j , n}\\right) ^ {2} + \\left(d y _ {i , j , n}\\right) ^ {2}}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 789, + 483, + 832 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $dx$ and $dy$ denote the pixel displacement in the vertical and horizontal directions across frames based on the optical flows estimated by RAFT [44]. Similar to mask ratios, we select five video sets, each containing 30 video", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/04c693f55e15b532bddafca674f278d6983b25bbe4cd8cf851cb7a6c857a01fa.jpg", + "table_caption": [ + "Table 2. Ablation studies on various components in the two-stage pipeline. \"HP\" denotes the homography-based priors. \"FP\" denotes the flow-based priors. \"FCC\" denotes the flow consistency check. \"VT\" denotes the video transformer in ProPainter. \"VDM\" denotes PCVDM in the proposed model. (e) corresponds to ProPainter, and (f) corresponds to HomoGen." + ], + "table_footnote": [], + "table_body": "
ComponentsPSNRSSIMVFID
FPFCCHPVTVDM
(a)XXXX✓33.160.96150.042
(b)✓✓XX✓34.050.96840.036
(c)✓XXX✓34.230.96990.035
(d)XX✓✓X33.880.96530.044
(e)✓✓X✓X34.460.97220.042
(f)XX✓X✓34.810.97350.032
", + "bbox": [ + 517, + 500, + 893, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sequences, for various values of $D^p$ . As shown in Fig. 6, HomoGen outperforms ProPainter in most motion conditions, particularly those with minimal movements. The two quantitative results demonstrate that HomoGen has superior inpainting capability across various video scenarios, especially in more hard cases involving larger masks or less motion where the spatiotemporal context is limited.", + "bbox": [ + 511, + 640, + 906, + 747 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.2. Qualitative Results", + "text_level": 1, + "bbox": [ + 511, + 758, + 692, + 773 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 7 shows qualitative results of HomoGen compared to the state-of-the-art methods, including FuseFormer, $\\mathrm{E}^2\\mathrm{FGVI}$ , and ProPainter on both video completion and object removal tasks. Generally, HomoGen can generate more faithful textural and structural information, as well as preserve the semantic integrity of objects. Nevertheless, other compared methods tend to fail or produce artifacts such as blur, distortion, black hazy regions, etc.", + "bbox": [ + 511, + 779, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22959", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/efda1db770dacccf3d0d7515438f60adc0442bf7a5b8265a92117f67d5501e16.jpg", + "table_caption": [ + "Table 3. Ablation studies on various components in HomoGen." + ], + "table_footnote": [], + "table_body": "
ModelPSNRSSIMVFID
w/o text prompts34.070.97160.039
w/o GoP-wise CLIP embeddings34.620.97220.033
HomoGen34.810.97350.032
", + "bbox": [ + 96, + 103, + 470, + 181 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/09c9d9c6a23ab68fa30556273e19c2a1f7509f657dc9934d9e9df307366cb785.jpg", + "image_caption": [ + "(i) line (a)" + ], + "image_footnote": [], + "bbox": [ + 93, + 186, + 173, + 250 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/580a118ac8bfe3c6689093529a29496086e55c721a878ae05f4ee5fa37b04365.jpg", + "image_caption": [ + "(ii) line (b)" + ], + "image_footnote": [], + "bbox": [ + 174, + 186, + 254, + 250 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b96559d4740395cf54a9f40e6fd00294f67104e0a63d329e0164d41b8de22831.jpg", + "image_caption": [ + "(iii) line (c)" + ], + "image_footnote": [], + "bbox": [ + 254, + 186, + 334, + 250 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a25aa309304e1ec060923652d6a242b8da1a274ddd05109ce0ecf7caea6fc103.jpg", + "image_caption": [ + "(iv) Input video" + ], + "image_footnote": [], + "bbox": [ + 336, + 186, + 480, + 250 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c56238cca498c9f453af8fe09619e462da2cc9c323b87be5b60eb48f3304ca80.jpg", + "image_caption": [ + "(v) line (d)" + ], + "image_footnote": [], + "bbox": [ + 93, + 265, + 173, + 328 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e183d3c1701e32f22d58281478113fcd92cf18b9c3ba6bf9db900a8eb82bb49c.jpg", + "image_caption": [ + "(vi) line (e)" + ], + "image_footnote": [], + "bbox": [ + 174, + 265, + 254, + 329 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f549e3df9d6139a34a152d34a2d54aad7a8ded8eb40543056ec7a571c033d363.jpg", + "image_caption": [ + "(vii) line (f)" + ], + "image_footnote": [], + "bbox": [ + 254, + 265, + 336, + 329 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/db651aafdff4061ad50cc897629d59a120a74dce80b1c4ebf9201a500f732693.jpg", + "image_caption": [ + "(viii) Input mask", + "Figure 9. Qualitative comparisons of the ablation studies in Tab. 2." + ], + "image_footnote": [], + "bbox": [ + 336, + 265, + 480, + 329 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Furthermore, we conduct a user study with 15 participants who are shown randomly sampled 45 video triplets. Each triplet contains one original video, one generated by HomoGen, and one generated by ProPainter. Volunteers are asked to select a visually better inpainting video or indicate a tie for each triplet. As shown in Fig. 8, volunteers prefer our results in terms of spatial coherence, temporal consistency, visual quality, and overall impression, demonstrating that HomoGen can generate more visually pleasant results.", + "bbox": [ + 89, + 376, + 483, + 512 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablations", + "text_level": 1, + "bbox": [ + 89, + 521, + 202, + 535 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We perform ablation studies on different technologies employed in HomoGen on the YouTube-VOS dataset.", + "bbox": [ + 89, + 542, + 482, + 573 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3.1. Study of Two-stage Pipeline", + "text_level": 1, + "bbox": [ + 89, + 580, + 330, + 594 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We investigate the importance of HPP tailored for PCVDM. As discussed in Sec. 1, various technologies have been designed for the two-stage pipeline, i.e., pixel propagation and content generation, among recent video inpainting methods. We compare different combinations of technologies within the two stages. For the first stage, we select HPP and the typical flow-based propagation proposed in ProPainter for comparisons. For the second stage, we select PCVDM and the video transformer proposed in ProPainter. As shown in line (a), Tab. 2, by removing HPP, temporal context is no longer provided to control content generation, resulting in a large performance decrease. The corresponding qualitative result in Fig. 9 shows artifacts and unrealistic content (the horse's leg). As demonstrated in line (b) and line (c), integrating flow-based priors with PCVDM, but without the consistency check, further enhances performance, which highlights the critical role of semantic integrity for DMs. Results in line (d) and line (e) indicate that the video transformer is not robust to artifacts in the priors and requires cleaner priors from pixel propagation. We also in", + "bbox": [ + 89, + 598, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/888e60f8a6543e621461b835a2f0de384b77335855e3763e6767e438264d27b9.jpg", + "table_caption": [ + "Table 4. Ablation studies on various prior injection mechanisms." + ], + "table_footnote": [], + "table_body": "
ModelPSNRSSIMVFID
Pasting31.740.96100.053
Concatenation32.860.96430.043
Cross-attention34.140.97180.035
Noise prior33.520.96570.036
Content-adaptive controls34.810.97350.032
", + "bbox": [ + 519, + 104, + 893, + 208 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "vestigate flow-based or homography-based priors combined with PCVDM. Results in line (c) and line (f) demonstrate that PCVDM exhibits greater tolerance to artifacts in these priors when coupled with HPP that provides more priors.", + "bbox": [ + 511, + 218, + 903, + 279 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3.2. Study of Prior Controlled Video Diffusion Model", + "text_level": 1, + "bbox": [ + 511, + 285, + 898, + 300 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We investigate the effectiveness of various components employed in PCVDM. We first remove the text prompts by replacing the visual descriptions with blank strings, for which the PSNR value is decreased by a large margin, i.e., 0.74 dB, as shown in Tab. 3. In contrast to the model without GoP-wise CLIP embeddings, the performance exhibits a non-trivial improvement. Furthermore, we compare the proposed content-adaptive control mechanism with other homography-based prior injection strategies, including pasting, concatenation, and cross-attention. Pasting represents directly pasting the priors into the corrupted regions, and others have been widely adopted in previous work [3, 25, 35, 49, 61]. Detailed implementations of different injection strategies refer to the supplementary. As shown in Tab. 4, pasting the priors into the corrupted regions raises a significant performance drop, which can be attributed to the fact that placing propagated and visible pixels on the same plane, making it difficult for DMs to differentiate between them. Compared to the second-best strategy employing cross-attention, using the content-adaptive controls improves performance by a non-trivial margin, i.e., 0.67 dB in terms of PSNR. This suggests that the content-adaptive control mechanism learns fine-grained dense injection and further enhances the control ability of PCVDM.", + "bbox": [ + 511, + 304, + 906, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 679, + 633, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our work introduces a propagate-and-diffuse framework for video inpainting, named HomoGen, which utilizes homography registration to propagate contextual pixels for generating missing content in a video diffusion model. Experiments not only show HomoGen's superiority in generating realistic content but prove the synergy between homography registration and video diffusion models for inpainting.", + "bbox": [ + 511, + 704, + 905, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. There are several potential limitations in this study. For instance, generating identity-preserved content in videos with scene alternations and efficiently processing high-resolution videos remain challenges for both HomoGen and previous methods. These issues warrant further investigation and improvement in future work.", + "bbox": [ + 511, + 810, + 905, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22960", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In ICCV, 2021. 2", + "[2] Marcelo Bertalmio, Andrea L Bertozzi, and Guillermo Sapiro. Navier-stokes, fluid dynamics, and image and video inpainting. In CVPR, 2001. 1", + "[3] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In CVPR, 2023. 2, 8", + "[4] Ya-Liang Chang, Zhe Yu Liu, Kuan-Ying Lee, and Winston Hsu. Free-form video inpainting with 3d gated convolution and temporal patchgan. In ICCV, 2019. 1, 3", + "[5] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In CVPR, 2023. 4", + "[6] Prafulla Dhariwal and Alexander Quinn Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021. 2", + "[7] Mounira Ebdelli, Olivier Le Meur, and Christine Guillemot. Video inpainting with short-term windows: application to object removal and error concealment. IEEE TIP, 2015. 1", + "[8] Ruoyu Feng, Wenming Weng, Yanhui Wang, Yuhui Yuan, Jianmin Bao, Chong Luo, Zhibo Chen, and Baining Guo. Credit: Creative and controllable video editing via diffusion models. In CVPR, 2024. 2", + "[9] Chen Gao, Ayush Saraf, Jia-Bin Huang, and Johannes Kopf. Flow-edge guided video completion. In ECCV, 2020. 2, 3, 6", + "[10] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. In ICLR, 2024. 4", + "[11] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross-attention control. In ICLR, 2023. 2", + "[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 2", + "[13] Yuan-Ting Hu, Heng Wang, Nicolas Ballas, Kristen Grauman, and Alexander G Schwing. Proposal-based video completion. In ECCV, 2020. 1, 3", + "[14] Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and Joao Carreira. Perceiver: General perception with iterative attention. In ICML, 2021. 4", + "[15] Wonjoon Jin, Qi Dai, Chong Luo, Seung-Hwan Baek, and Sunghyun Cho. Flovd: Optical flow meets video diffusion model for enhanced camera-controlled video synthesis. In CVPR, 2025. 2", + "[16] Xuan Ju, Xian Liu, Xintao Wang, Yuxuan Bian, Ying Shan, and Qiang Xu. Brushnet: A plug-and-play image inpainting model with decomposed dual-branch diffusion. In ECCV, 2024.5", + "[17] Dahun Kim, Sanghyun Woo, Joon-Young Lee, and In So Kweon. Deep video inpainting. In CVPR, 2019. 3, 6" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[18] Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. In ICLR, 2014. 4", + "[19] Wei-Sheng Lai, Jia-Bin Huang, Oliver Wang, Eli Shechtman, Ersin Yumer, and Ming-Hsuan Yang. Learning blind video temporal consistency. In ECCV, 2018. 6", + "[20] Sungho Lee, Seoung Wug Oh, DaeYeun Won, and Seon Joo Kim. Copy-and-paste networks for deep video inpainting. In ICCV, 2019. 3, 6", + "[21] Ang Li, Shanshan Zhao, Xingjun Ma, Mingming Gong, Jianzhong Qi, Rui Zhang, Dacheng Tao, and Ramamohanarao Kotagiri. Short-term and long-term context aggregation network for video inpainting. In ECCV, 2020. 1, 3", + "[22] Zhen Li, Cheng-Ze Lu, Jianhua Qin, Chun-Le Guo, and Ming-Ming Cheng. Towards an end-to-end framework for flow-guided video inpainting. In CVPR, 2022. 2, 5, 6, 7", + "[23] Miao Liao, Feixiang Lu, Dingfu Zhou, Sibo Zhang, Wei Li, and Ruigang Yang. Dvi: Depth guided video inpainting for autonomous driving. In ECCV, 2020. 1", + "[24] Rui Liu, Hanming Deng, Yangyi Huang, Xiaoyu Shi, Lewei Lu, Wenxiu Sun, Xiaogang Wang, Jifeng Dai, and Hongsheng Li. Fuseformer: Fusing fine-grained information in transformers for video inpainting. In ICCV, 2021. 1, 3, 6, 7", + "[25] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. In ICLR, 2022. 2, 8", + "[26] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In AAAI, 2024. 2", + "[27] OpenAI. Gpt-4o contributions. https://openai.com/gpt-4o-contributions, 2024.6", + "[28] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In CVPR, 2019. 2, 5", + "[29] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In CVPR, 2016. 1", + "[30] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In CVPR, 2016. 6", + "[31] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: improving latent diffusion models for high-resolution image synthesis. In ICLR, 2024. 2, 3", + "[32] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. In ICCV, 2023. 2", + "[33] Prajit Ramachandran, Barret Zoph, and Quoc V. Le. Searching for activation functions. In ICLR Workshops, 2018. 5", + "[34] Jingjing Ren, Qingqing Zheng, Yuanyuan Zhao, Xuemiao Xu, and Chen Li. Dlformer: Discrete latent transformer for video inpainting. In CVPR, 2022. 3" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22961", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[35] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 2, 3, 4, 5, 8", + "[36] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 4", + "[37] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary Bradski. Orb: An efficient alternative to sift or surf. In ICCV. IEEE, 2011. 3", + "[38] Min-Cheol Sagong, Yoon-Jae Yeo, Seung-Won Jung, and Sung-Jea Ko. Rord: A real-world object removal dataset. In BMVC, 2022. 6", + "[39] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 2", + "[40] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-a-video: Text-to-video generation without text-video data. In ICLR, 2023. 4", + "[41] Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 2", + "[42] Richard Szeliski. Image alignment and stitching: A tutorial. Found. Trends Comput. Graph. Vis., 2006. 2, 3", + "[43] Nick C Tang, Chiou-Ting Hsu, Chih-Wen Su, Timothy K Shih, and Hong-Yuan Mark Liao. Video inpainting on digitized vintage films via maintaining spatiotemporal continuity. IEEE TMM, 2011. 1, 3", + "[44] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In ECCV, 2020. 2, 7", + "[45] Rui Tian, Qi Dai, Jianmin Bao, Kai Qiu, Yifan Yang, Chong Luo, Zuxuan Wu, and Yu-Gang Jiang. Reducio! generating $1024 \\times 1024$ video within 16 seconds using extremely compressed motion latents. arXiv preprint arXiv:2411.13552, 2024. 2", + "[46] Shuyuan Tu, Qi Dai, Zhi-Qi Cheng, Han Hu, Xintong Han, Zuxuan Wu, and Yu-Gang Jiang. Motioneditor: Editing video motion via content-aware diffusion. In CVPR, 2024. 2", + "[47] Chuan Wang, Haibin Huang, Xiaoguang Han, and Jue Wang. Video inpainting by jointly learning temporal structure and spatial details. In AAAI, 2019. 3", + "[48] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Nikolai Yakovenko, Andrew Tao, Jan Kautz, and Bryan Catanzaro. Video-to-video synthesis. In NeurIPS, 2018. 6", + "[49] Yanhui Wang, Jianmin Bao, Wenming Weng, Ruoyu Feng, Dacheng Yin, Tao Yang, Jingxu Zhang, Qi Dai, Zhiyuan Zhao, Chunyu Wang, et al. Microcinema: A divide-and-conquer approach for text-to-video generation. In CVPR, 2024. 2, 8", + "[50] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE TIP, 2004. 6" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Wenming Weng, Ruoyu Feng, Yanhui Wang, Qi Dai, Chunyu Wang, Dacheng Yin, Zhiyuan Zhao, Kai Qiu, Jian-min Bao, Yuhui Yuan, et al. Art-v: Auto-regressive text-to-video generation with diffusion models. In CVPR Workshops, 2024. 2", + "[52] Yuxin Wu and Kaiming He. Group normalization. In ECCV, 2018. 5", + "[53] Zhen Xing, Qi Dai, Han Hu, Zuxuan Wu, and Yu-Gang Jiang. Simda: Simple diffusion adapter for efficient video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7827-7839, 2024. 2", + "[54] Zhen Xing, Qijun Feng, Haoran Chen, Qi Dai, Han Hu, Hang Xu, Zuxuan Wu, and Yu-Gang Jiang. A survey on video diffusion models. ACM Computing Surveys, 57(2):1-42, 2024. 2", + "[55] Ning Xu, Linjie Yang, Yuchen Fan, Jianchao Yang, Dingcheng Yue, Yuchen Liang, Brian Price, Scott Cohen, and Thomas Huang. Youtube-vos: Sequence-to-sequence video object segmentation. In ECCV, 2018. 6", + "[56] Rui Xu, Xiaoxiao Li, Bolei Zhou, and Chen Change Loy. Deep flow-guided video inpainting. In CVPR, 2019. 2, 3, 5, 6", + "[57] Jiahui Yu, Zhe Lin, Jamei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Generative image inpainting with contextual attention. In CVPR, 2018. 1", + "[58] Yanhong Zeng, Jianlong Fu, and Hongyang Chao. Learning joint spatial-temporal transformations for video inpainting. In ECCV, 2020. 1, 3, 6", + "[59] Kaidong Zhang, Jingjing Fu, and Dong Liu. Flow-guided transformer for video inpainting. In ECCV, 2022. 2, 3, 5, 6", + "[60] Kaidong Zhang, Jingjing Fu, and Dong Liu. Inertia-guided flow completion and style fusion for video inpainting. In CVPR, 2022. 3, 6", + "[61] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 2, 5, 8", + "[62] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In CVPR, 2017. 2", + "[63] Shangchen Zhou, Chongyi Li, Kelvin CK Chan, and Chen Change Loy. Propainter: Improving propagation and transformer for video inpainting. In ICCV, 2023. 2, 3, 5, 6, 7" + ], + "bbox": [ + 516, + 92, + 903, + 714 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22962", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/HomoGen_ Enhanced Video Inpainting via Homography Propagation and Diffusion/c0acdc17-6ead-4f9d-8631-4f1535650ac6_model.json b/2025/HomoGen_ Enhanced Video Inpainting via Homography Propagation and Diffusion/c0acdc17-6ead-4f9d-8631-4f1535650ac6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ec870e623e2dfd344e779da367b3c682e843e23b --- /dev/null +++ b/2025/HomoGen_ Enhanced Video Inpainting via Homography Propagation and Diffusion/c0acdc17-6ead-4f9d-8631-4f1535650ac6_model.json @@ -0,0 +1,2508 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.131, + 0.785, + 0.177 + ], + "angle": 0, + "content": "HomoGen: Enhanced Video Inpainting via Homography Propagation and Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.202, + 0.905, + 0.275 + ], + "angle": 0, + "content": "Ding Ding\\(^{1*}\\) Yueming Pan\\(^{2}\\) Ruoyu Feng\\(^{3}\\) Qi Dai\\(^{4}\\) Kai Qiu\\(^{4}\\) Jianmin Bao\\(^{4}\\) Chong Luo\\(^{4\\dagger}\\) Zhenzhong Chen\\(^{1}\\) \n\\(^{1}\\)Wuhan University \\(^{2}\\)Xi'an Jiaotong University \\(^{3}\\)University of Science and Technology of China \\(^{4}\\)Microsoft Research Asia" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.294, + 0.905, + 0.509 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.515, + 0.908, + 0.559 + ], + "angle": 0, + "content": "Figure 1. Two-stage pipelines of HomoGen and the state-of-the-art method. Previous flow-based propagation warps pixels across frames, followed by a video transformer (VT) that generates remaining missing content. In contrast, HomoGen's homography-based propagation warps pixels as priors to control content generation in a video diffusion model (VDM) that filters out artifacts in the priors during denoising." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.578, + 0.327, + 0.593 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.61, + 0.486, + 0.866 + ], + "angle": 0, + "content": "In this paper, we present HomoGen, an enhanced video inpainting method based on homography propagation and diffusion models. HomoGen leverages homography registration to propagate contextual pixels as priors for generating missing content in corrupted videos. Unlike previous flow-based propagation methods, which introduce local distortions due to point-to-point optical flows, homography-induced artifacts are typically global structural distortions that preserve semantic integrity. To effectively utilize these priors for generation, we employ a video diffusion model that inherently prioritizes semantic information within the priors over pixel-level details. A content-adaptive control mechanism is proposed to scale and inject the priors into intermediate video latents during iterative denoising. In contrast to existing transformer-based networks that often suffer from artifacts within priors, leading to error accumulation and unrealistic results, our denoising diffusion net-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.579, + 0.907, + 0.626 + ], + "angle": 0, + "content": "work can smooth out artifacts and ensure natural outputs. Extensive experiments demonstrate the effectiveness of the proposed method qualitatively and quantitatively." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.649, + 0.645, + 0.665 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.675, + 0.907, + 0.872 + ], + "angle": 0, + "content": "Video inpainting aims to fill corrupted regions in a video with realistic and spatiotemporally coherent content [2]. It has diverse applications including object removal [7], video restoration [43], autonomous driving [23], etc. Despite significant progress in image inpainting [29, 35, 57], video inpainting remains full of challenges due to the complexity of video scenarios. The primary difficulty lies in identifying spatiotemporal guidance for generating content within corrupted regions, particularly in videos with diverse camera movements and object motion. To tackle this, previous work has explored various technologies, such as 3D convolutions [4, 13] and attentions [21, 24, 58], to leverage spatiotemporal context and preserve coherent across frames." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.873, + 0.907, + 0.904 + ], + "angle": 0, + "content": "Recently, two-stage methods combining flow-based propagation and generation networks have shown impres" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.876, + 0.443, + 0.889 + ], + "angle": 0, + "content": "*This work was done when Ding Ding was an intern at MSRA." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.889, + 0.236, + 0.902 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.876, + 0.443, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22953" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.439 + ], + "angle": 0, + "content": "sive performance [22, 59, 63]. In flow-based propagation, optical flows are utilized to propagate pixels from adjacent frames into corrupted regions via warping. For generation networks, powerful video transformers are employed to generate remaining missing content. However, such methods still face challenges: the warping operation introduces artifacts such as blur and distortion in the propagated pixels, especially in scenarios characterized by significant object or camera motion and lighting changes [9]. As shown in Fig. 1, the flow-based propagation method in the state-of-the-art ProPainter [63] incorrectly warps a tree into the road center in the first stage. Besides, the propagated pixels are considered partial inpainting results to guide remaining content generation. Such artifacts cause error accumulation in video transformers, leading to unrealistic results [22, 59]. To mitigate artifacts in the propagated pixels, recent methods propose the flow consistency check to avoid inaccurate flows [56]. However, the consistency check discards partial flows, leaving the propagated pixels insufficient and semantically incoherent. As shown in Fig. 1, a subset of road pixels is removed from the propagated content after the consistency check. Therefore, video transformers struggle to generate high-fidelity content without sufficient guidance." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.444, + 0.483, + 0.79 + ], + "angle": 0, + "content": "Regarding the challenges in balancing retaining more propagated pixels while avoiding artifacts, diffusion models [12, 35] offer a promising solution as generation networks in the two-stage pipeline. Diffusion models prioritize semantic information within input conditions, such as stroke paintings and unnatural images [25, 28, 61], over pixel-level details. During denoising, diffusion models could generate realistic content while filtering out noisy low-level features from the conditions. Besides, homography registration for pixel propagation is more compatible with diffusion models. Specifically, artifacts from homography differ from those caused by flows. Flows capture point-to-point correlations [44], and flow-induced artifacts often considerably distort parts of objects in complex scenes, disrupting the semantics of propagated content. In contrast, homography imposes projective geometric constraints based on the global structure between frames [42], and homography-induced artifacts are typically structural distortions while preserving semantic integrity. As shown in Fig. 1, in the first stage, although several road pixels experience overall distortions, the structural integrity of objects remains largely intact. In the second stage, the video diffusion model generates more natural content while filtering out the artifacts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.482, + 0.901 + ], + "angle": 0, + "content": "In this paper, we propose an enhanced video inpainting framework, named HomoGen, which enables pixel propagation via Homography registration and designs a video diffusion model to Generate missing content. For pixel propagation, we register paired frames via homography matrices to explore valuable contextual pixels. After warping the reference frame, the propagated pixels located in the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.379 + ], + "angle": 0, + "content": "corrupted regions of the target frame are extracted as priors. We iterate the propagation process across all frames in a bidirectional manner, where the reference frames are progressively updated with the propagated pixels at each step. For content generation, we design a prior controlled video diffusion model that captures the semantics within the bidirectional priors. Built upon a text-guided image inpainting model from Stable Diffusion, we integrate temporal layers into it to construct a 3D network. To effectively utilize the priors, we propose a content-adaptive control mechanism that scales and injects the priors into the diffusion model. Specifically, we design a lite ControlNet [61] to embed the priors into multi-scale features that are then fed into the 3D U-Net via dense injection. Learned scalers adjust the multi-scale features before injection, adapting to the spatiotemporal context of priors and the sampling steps. Simultaneously, the denoising process filters out artifacts within the priors during iterative sampling, thus preserving high-fidelity results across inpainted regions and visible pixels." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.38, + 0.905, + 0.456 + ], + "angle": 0, + "content": "Qualitative and quantitative experiment results demonstrate the superiority of HomoGen both in video completion and object removal. Compared to the state-of-the-art methods, HomoGen can generate more realistic content and improve performance by \\(23.81\\%\\) in VFID." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.457, + 0.904, + 0.516 + ], + "angle": 0, + "content": "- A propagate-and-diffuse framework is proposed for video inpainting, which leverages homography registration to propagate contextual pixels as priors to control the generation of missing content in a denoising diffusion network." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.517, + 0.905, + 0.591 + ], + "angle": 0, + "content": "- A content-adaptive control mechanism is designed to inject prior information into the diffusion model. The dense injection learns to adapt to the spatiotemporal context of priors and the sampling steps, while artifacts within the priors are filtered out during iterative denoising." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.593, + 0.904, + 0.637 + ], + "angle": 0, + "content": "- Extensive experiments show that HomoGen surpasses the state-of-the-art methods and generates more realistic content in video completion and object removal tasks." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.457, + 0.905, + 0.637 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.655, + 0.655, + 0.67 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.682, + 0.896, + 0.697 + ], + "angle": 0, + "content": "2.1. Diffusion-based Image and Video Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Diffusion models have exhibited remarkable abilities in visual generation [3, 6, 15, 31, 35, 45, 49, 51, 53, 54]. These models learn to approximate data distributions through the iterative denoising of diffused inputs [12, 41]. Their practicality is enhanced by the incorporation of various conditions, e.g., text prompts, unnatural images, segmentation maps, etc., allowing for controlled outputs over the generative process. Controllable generation methods particularly benefit from additional structure modules, e.g., ControlNet [61], T2I-adapter [26], etc., which manage the injection of conditions. Such models leverage large-scale datasets [1, 39, 62] to achieve impressive results not only in generation but in editing [8, 11, 32, 46]. Conditional diffusion can" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22954" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.138 + ], + "angle": 0, + "content": "also be applied to video inpainting, a widely explored video editing task, where the visible pixels serve as conditions to guide the generation of missing content in corrupted videos." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.146, + 0.26, + 0.163 + ], + "angle": 0, + "content": "2.2. Video Inpainting" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.168, + 0.484, + 0.515 + ], + "angle": 0, + "content": "Previous video inpainting methods can be broadly divided into two categories: one-stage and two-stage approaches. One-stage approaches employ 3D CNNs and video transformers to aggregate spatiotemporal context for generating missing content in videos [4, 21, 24, 34, 58]. These methods usually struggle to capture spatiotemporal details due to limited receptive fields, which leads to blurry and inconsistent outcomes in complex scenes. Two-stage approaches have gradually become mainstream choices for video inpainting. The first stage primarily extracts prior information from the video itself, such as motion information [17, 43, 56], homography [20], low-resolution results [47], and proposal sets [13]. Guided by the priors, the second stage generates missing content based on well-designed networks, such as image inpainting models [9, 56, 60], 3D CNNs [17] and transformers [59]. Recent methods capitalize on the advantages of optical flows for motion capture and video transformers for spatiotemporal context modeling [9, 63]. As a typical approach, ProPainter [63] begins with flow extraction and completion, and utilizes completed flows to propagate pixels into corrupted regions. A flow-guided video transformer is designed to refine the propagated pixels and generate remaining missing content." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.516, + 0.484, + 0.651 + ], + "angle": 0, + "content": "However, such methods still suffer from blur and distortion in the results, primarily caused by the warping operation in pixel propagation and error accumulation in video transformers. To address this issue, we design a video diffusion model as the generation network that effectively removes artifacts within the priors during iterative denoising. Additionally, we employ homography-based pixel propagation to preserve the semantic integrity of warped content, providing fine-grained controls for conditional diffusion." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.664, + 0.226, + 0.681 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.484, + 0.78 + ], + "angle": 0, + "content": "HomoGen consists of two main components: Homography-based Pixel Propagation (HPP) and Prior Controlled Video Diffusion Model (PCVDM), as shown in Fig. 2. In HPP, we estimate homography between paired frames for propagating pixels as contextual priors. PCVDM then generates missing content controlled by these priors." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.789, + 0.221, + 0.805 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Consider a corrupted video sequence denoted as \\(\\mathbf{x} = \\{\\mathbf{x}_i\\mid\\) \\(\\mathbf{x}_i\\in \\mathbb{R}^{3\\times H\\times W},1\\leq i\\leq N\\}\\) comprising \\(N\\) frames, accompanied by a binary mask sequence \\(\\mathbf{m} = \\{\\mathbf{m}_i\\mid \\mathbf{m}_i\\in\\) \\(\\mathbb{R}^{1\\times H\\times W},1\\leq i\\leq N\\}\\) identifying corrupted regions in all frames. Text prompts provide visual descriptions of the inpainted video. With text prompts as additional conditions," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "text-guided video inpainting aims to generate visually faithful and coherent content within the corrupted regions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.122, + 0.907, + 0.269 + ], + "angle": 0, + "content": "As text-guided video inpainting encounters diffusion models (DMs), latent diffusion models (LDMs) [31, 35] are effective choices. LDMs leverage a pre-trained Variational Autoencoder (VAE) that consists of an encoder \\(\\mathcal{E}\\) and a decoder \\(\\mathcal{D}\\) to perform denoising in the latent space where the latent variable \\(\\mathbf{z} = \\mathcal{E}(\\mathbf{x})\\). For video inpainting, conditions \\(\\mathbf{c}\\) are injected into a denoising network \\(\\epsilon_{\\theta}\\) to predict \\(\\mathbf{z}^t\\), a denoised variant of a noise, where \\(t = 0, 1, 2, \\dots, T\\) denotes the sampling step. The optimization objective is formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.269, + 0.905, + 0.294 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} _ {\\mathbf {z} ^ {0}, t, \\epsilon \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I})} \\left[ \\left\\| \\epsilon - \\epsilon_ {\\theta} \\left(\\mathbf {z} ^ {t}, t, \\mathbf {c}\\right) \\right\\| _ {2} ^ {2} \\right], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.299, + 0.906, + 0.33 + ], + "angle": 0, + "content": "where \\(\\mathbf{c} = (\\mathbf{x},\\mathbf{m},\\tau_{\\theta}(y))\\) \\(y\\) represents the text prompts and \\(\\tau_{\\theta}(\\cdot)\\) denotes a text encoder." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.339, + 0.843, + 0.355 + ], + "angle": 0, + "content": "3.2. Homography-based Pixel Propagation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.361, + 0.907, + 0.587 + ], + "angle": 0, + "content": "Pixel propagation aligns content within corrupted videos by propagating pixels from reference frames into corrupted regions in the target frame. Flow-based propagation typically warps pixels based on the trajectories of optical flows [17, 59]. However, they face challenges in balancing more propagated pixels and fewer distortions that cause error accumulation within the generation network. To overcome this, we employ homography-based propagation for content alignment. Compared to flows that capture point-to-point correlations, homography registration offers greater robustness since homography matrices for warping impose projective geometric constraints based on global structures between frames [42]. Besides, homography registration usually preserves the semantic integrity without altering parts of objects in complex scenes, as discussed in Sec. 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.588, + 0.906, + 0.678 + ], + "angle": 0, + "content": "To search for valuable pixels within a large temporal window, we estimate the forward and backward homography and iteratively propagate pixels across frames in a bidirectional manner. For simplicity, we only describe the forward propagation process. The propagated pixels \\(\\mathbf{x}_{i,i + 1}^p\\) in frame \\(\\mathbf{x}_{i + 1}\\) are warped from frame \\(\\mathbf{x}_i\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.689, + 0.905, + 0.708 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {i, i + 1} ^ {p} = \\left(1 - \\mathbf {m} _ {i, i + 1} ^ {p}\\right) \\odot \\mathbf {x} _ {i, i + 1}, \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.71, + 0.905, + 0.728 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {i, i + 1} = \\mathcal {W} \\left(\\mathbf {x} _ {i} + \\mathbf {x} _ {i - 1, i} ^ {p}, \\mathcal {H} _ {i, i + 1}\\right), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.731, + 0.905, + 0.749 + ], + "angle": 0, + "content": "\\[\n\\mathbf {m} _ {i, i + 1} ^ {p} = 1 - \\mathbf {m} _ {i, i + 1} \\odot \\mathbf {m} _ {i + 1}, \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.752, + 0.905, + 0.771 + ], + "angle": 0, + "content": "\\[\n\\mathbf {m} _ {i, i + 1} = \\mathcal {W} \\left(\\mathbf {m} _ {i} \\odot \\mathbf {m} _ {i - 1, i} ^ {p}, \\mathcal {H} _ {i, i + 1}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.907, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathbf{x}_{i,i + 1}\\) and \\(\\mathbf{m}_{i,i + 1}\\) denotes warped \\(\\mathbf{x}_i\\) and \\(\\mathbf{m}_i\\) registered with \\(\\mathbf{x}_{i + 1}\\) and \\(\\mathbf{m}_{i + 1}\\), respectively. \\(\\mathbf{m}_{i + 1}^p\\) denotes the mask indicating the regions of the propagated pixels. \\(\\odot\\) denotes the Hadamard product. \\(\\mathcal{W}(\\cdot ,\\cdot)\\) represents the warping operation. \\(\\mathcal{H}_{i,i + 1}\\) denotes the homography matrix for projecting the plane of \\(\\mathbf{x}_i\\) onto \\(\\mathbf{x}_{i + 1}\\). \\(\\mathcal{H}_{i,i + 1}\\) is estimated by RANSAC on ORB feature matches [37]. During propagation, each reference frame is registered to the target frame." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22955" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.093, + 0.904, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.361, + 0.908, + 0.404 + ], + "angle": 0, + "content": "Figure 2. Overview of the proposed HomoGen framework, comprising two primary components: Homography-based Pixel Propagation (HPP) and Prior Controlled Video Diffusion Model (PCVDM). In this two-stage pipeline, HPP leverages homography between frames to propagate pixels as contextual priors that control missing content generation in PCVDM." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.414, + 0.287, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.493, + 0.263, + 0.505 + ], + "angle": 0, + "content": "(a) Flow-based propagation" + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.414, + 0.482, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.363, + 0.493, + 0.408, + 0.504 + ], + "angle": 0, + "content": "(b) HPP" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.511, + 0.484, + 0.595 + ], + "angle": 0, + "content": "Figure 3. Statistics of warping error and propagated pixel count on the YouTube-VOS test set in video completion. Warping error denotes the Mean Absolute Error (MAE) between the ground truth (normalized to \\([-1,1]\\)) and the propagated pixels within a video. Pixel count denotes the proportion of propagated pixels in the corrupted regions. The grid color indicates the number of videos." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.608, + 0.484, + 0.682 + ], + "angle": 0, + "content": "After warping based on homography matrices, the propagated pixels located in the corrupted regions are extracted as priors. We iterate the propagation process across all frames, where the reference frames are progressively updated with the propagated pixels at each step." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.683, + 0.484, + 0.731 + ], + "angle": 0, + "content": "Valuable pixels are propagated under homography constraints. For each frame, we blend forward and backward propagated pixels \\(\\mathbf{x}_{i - 1,i}^p\\) and \\(\\mathbf{x}_{i + 1,i}^p\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.115, + 0.738, + 0.483, + 0.757 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {i} ^ {p} = \\mathbf {x} _ {i - 1, i} ^ {p} + \\left(1 - \\mathbf {m} _ {i - 1, i} ^ {p}\\right) \\odot \\mathbf {m} _ {i + 1, i} ^ {p} \\odot \\mathbf {x} _ {i + 1, i} ^ {p}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.796 + ], + "angle": 0, + "content": "where \\(\\mathbf{x}_i^p\\) denotes the bidirectional propagated pixels as the priors to control missing content generation in PCVDM." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.797, + 0.484, + 0.902 + ], + "angle": 0, + "content": "For further analysis, we present statistics on warping error and the count of propagated pixels for flow-based propagation and HPP. As shown in Fig. 3, the flow-based method propagates fewer pixels with lower warping error due to the consistency check, while HPP increases the pixel count at the cost of slightly higher warping error. Despite this added error, PCVDM benefits from richer spatiotemporal context" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.417, + 0.907, + 0.448 + ], + "angle": 0, + "content": "and complete semantic information while filtering out artifacts in such priors during iterative denoising." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.458, + 0.857, + 0.473 + ], + "angle": 0, + "content": "3.3. Prior Controlled Video Diffusion Model" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.481, + 0.907, + 0.557 + ], + "angle": 0, + "content": "Based on the priors from HPP, PCVDM aims to generate spatiotemporal coherent and realistic content in the corrupted regions. We propose a video diffusion model based on Stable Diffusion inpainting [35] and design a content-adaptive control mechanism to inject the priors into DM." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.566, + 0.871, + 0.581 + ], + "angle": 0, + "content": "3.3.1. Inflating Image Inpainter to Video Inpainter" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.585, + 0.907, + 0.844 + ], + "angle": 0, + "content": "The proposed video inpainting model is extended from the image inpainting model in Stable Diffusion. Following previous work [10, 40], we extend the 2D DM into a 3D structure. In VAE decoder [18] and U-Net [36], original 2D layers are converted to pseudo-3D layers to learn temporal correlations, with detailed structures available in the supplementary. Besides, we employ an OpenCLIP image encoder and perceive attention layers to extract GoP-wise CLIP latents from the corrupted video [5, 14], and then inject the latents into the 3D U-Net via cross-attention. GoP refers to a Group of Picture, i.e., a video clip processed per inference. During training, the binary mask \\(\\mathbf{m}\\), the noised video latent \\(\\mathbf{z}^t\\), and the corrupted video latent \\(\\mathbf{z} = \\mathcal{E}(\\mathbf{x})\\) are concatenated in the latent space and fed into the 3D U-Net. During inference, \\(\\mathbf{z}^t\\) is initialized as Gaussian noise. The 3D U-Net generates the final sampled latent \\(\\hat{\\mathbf{z}}\\) that is decoded by the 3D VAE decoder as the completed video \\(\\hat{\\mathbf{x}}\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.851, + 0.825, + 0.866 + ], + "angle": 0, + "content": "3.3.2. Content-adaptive Control Mechanism" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "The priors are essential for generating missing content in corrupted videos. Recent work fills the propagated pix" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22956" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.089, + 0.485, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.403, + 0.484, + 0.462 + ], + "angle": 0, + "content": "Figure 4. Illustration of content-adaptive control mechanism that injects prior information into PCVDM. Prior scalers in the dense injection are learned from the spatiotemporal context of priors and the sampling steps, enabling fine-grained controls." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.473, + 0.483, + 0.744 + ], + "angle": 0, + "content": "els into the corrupted regions as partial inpainting results and uses transformers to generate remaining missing content [22, 59, 63]. However, due to inevitable artifacts introduced by warping, pixel-level error accumulates during token prediction in transformers guided by noisy tokens from the priors, as discussed in Sec. 1. A trade-off solution is to discard inaccurate propagated pixels based on the consistency check [56] while limiting sufficient priors and semantic integrity. Unlike transformers' sensitivity to low-level features, DMs prioritize semantic information within input conditions over pixel-level details. Hence, to ensure the semantic integrity of the propagated content, we retained all propagated pixels without any consistency check. Besides, we consider the priors as conditions instead of partial inpainting results pasted with original visible pixels, which also helps the network unambiguously perceive artifacts from the priors. During iterative sampling, the denoising process can smooth out such undesirable artifacts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.748, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To effectively utilize the priors, we propose a content-adaptive control mechanism to scale and inject them into PCVDM for controlled missing content generation. The priors \\(\\mathbf{x}^p = \\{\\mathbf{x}_i^p\\mid \\mathbf{x}_i^p\\in \\mathbb{R}^{3\\times H\\times W},1\\leq i\\leq N\\}\\) are first encoded into latents \\(\\mathbf{z}^p = \\{\\mathbf{z}_i^p\\mid \\mathbf{z}_i^p\\in \\mathbb{R}^{4\\times \\frac{H}{8}\\times \\frac{W}{8}},1\\leq i\\leq N\\}\\). We first design a LiteControlNet to extract multi-scale features from \\(\\mathbf{z}^p\\) for further injection. Modified from ControlNet [61], LiteControlNet features a lightweight structure with one convolutional layer at each scale. Detailed structures of LiteControlNet refer to the supplementary. The" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.244 + ], + "angle": 0, + "content": "prior feature at each scale is connected to the 3D U-Net encoder through a zero convolutional layer and a learned projector, i.e., a 3D convolutional layer with a kernel size of \\(1 \\times 1 \\times 1\\). Rather than using additive operation, we introduce a spatiotemporal adaptive de-normalization strategy to inject each feature into the corresponding normalization layer [28]. As shown in Fig. 4, after the \\(j\\)-th latent normalized by 3D group normalization [52] in the 3D U-Net, we compute the output activation \\(\\mathbf{h}_j\\) by de-normalizing the normalized latent \\(\\bar{\\mathbf{h}}_j\\) according to the prior feature \\(\\mathbf{f}_j\\), formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.62, + 0.248, + 0.907, + 0.268 + ], + "angle": 0, + "content": "\\[\n\\mathbf {h} _ {j} = \\gamma_ {j} \\odot \\bar {\\mathbf {h}} _ {j} + \\beta_ {j} + \\bar {\\mathbf {h}} _ {j}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.27, + 0.907, + 0.362 + ], + "angle": 0, + "content": "where \\(\\gamma_{j}\\) and \\(\\beta_{j}\\) are obtained through 3D convolutions of \\(\\mathbf{f}_j\\). Equipped with this design, \\(\\mathbf{f}_j\\) is scaled and injected into the intermediate latent, where the generation of missing content adapts to the spatiotemporal context of the priors. Additionally, \\(\\mathbf{f}_j\\) is injected after attention blocks and yielded as the \\((j + 1)\\)-th latent \\(\\mathbf{h}_{j + 1}\\), formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.638, + 0.367, + 0.907, + 0.386 + ], + "angle": 0, + "content": "\\[\n\\mathbf {h} _ {j + 1} = \\zeta (t) \\mathbf {f} _ {j} + \\mathbf {h} _ {j}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.389, + 0.906, + 0.449 + ], + "angle": 0, + "content": "where \\(\\zeta(\\cdot)\\) represents two linear layers interleaved with a SiLU activation [33] in between. \\(\\zeta(\\cdot)\\) encodes the sampling step \\(t\\) [35] into Fourier features, which allows the scaling strength to vary with \\(t\\) for fine-grained controls." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.45, + 0.906, + 0.541 + ], + "angle": 0, + "content": "During iterative denoising, the content in the corrupted regions progressively generates, approaching seamless integration with original visible pixels. After decoding the final sampled latent by the 3D VAE decoder, the generated and corrupted videos are blended with a blurred mask to improve coherence along the mask boundary [16]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.547, + 0.759, + 0.563 + ], + "angle": 0, + "content": "3.4. Inpainting for Long Videos" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.906, + 0.674 + ], + "angle": 0, + "content": "Video inpainting faces the challenge of maintaining temporal coherence across GoPs in long videos that span several seconds. Generative models are often restricted to processing all frames at once due to memory limitations, which leads to variability when sequentially inferring all GoPs. To address this, we introduce temporal references from adjacent GoPs to assist in inpainting the current GoP." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Let \\(N\\) denote the length of the GoP that HomoGen processes per inference. For a video containing \\(L\\) frames, we segment the video into \\(\\left\\lceil \\frac{L}{N} \\right\\rceil\\) GoPs, each with an overlap of \\(N_{\\text{overlap}}\\) frames, and process all GoPs sequentially. For the \\(i\\)-th GoP, the first and last frames are set to the center frames of the \\((i-1)\\)-th and \\((i+1)\\)-th GoP, respectively, while \\(N_{\\text{overlap}} - 2\\) frames from the second frame onward are set to the last frames of the \\((i-1)\\)-th GoP. In the initial inference, the first GoP undergoes inpainting. Before \\(i\\)-th inference \\((i > 1)\\), the beginning \\(N_{\\text{overlap}} - 1\\) frames of the \\(i\\)-th GoP are replaced by the corresponding frames from the inpainted \\((i-1)\\)-th GoP. During \\(i\\)-th inference, HomoGen can capture long-term temporal references for inpainting the \\(i\\)-th GoP. This process iterates until all GoPs are inpainted, which ensures content consistency across the entire video." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "22957" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 1. Quantitative comparisons on YouTube-VOS, DAVIS, and RORD datasets. \\(\\uparrow\\) indicates higher is better. \\(\\downarrow\\) indicates lower is better. \\(E_{w}^{*}\\) denotes \\(E_{\\text{warp}} \\times 10^{-3}\\). The best performances are highlighted in bold. All methods are evaluated following their default settings." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.121, + 0.907, + 0.322 + ], + "angle": 0, + "content": "
YouTube-VOSDAVISRORD
ModelsPSNR ↑SSIM ↑VFID ↓Ew* ↓PSNR ↑SSIM ↑VFID ↓Ew* ↓PSNR ↑SSIM ↑VFID ↓Ew* ↓
DFVI [56]29.160.94290.0661.50928.810.94040.1871.60831.840.95710.0651.137
CPNet [20]31.580.96070.0711.42330.280.95210.1821.51935.070.97400.0390.879
FGVC [9]29.670.94030.0641.02230.800.94970.1651.58632.760.96530.0521.005
STTN [58]32.570.96800.0490.90730.740.95900.1531.66134.940.97180.0440.872
FuseFormer [24]33.290.96810.0530.89132.540.97020.1381.33635.610.98110.0310.843
FGT [59]32.170.95990.0541.02532.860.96500.1291.32334.210.97240.0380.867
ISVI [60]30.380.94710.0771.31132.200.96260.1661.29233.450.96910.0460.902
E2FGVI [22]33.710.97000.0460.86433.010.97210.1161.31536.040.98000.0270.797
ProPainter [63]34.460.97220.0420.87234.190.97590.0981.19536.760.98670.0280.813
HomoGen34.810.97350.0320.85834.710.97630.0771.23437.290.98720.0230.793
" + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.324, + 0.287, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.126, + 0.439, + 0.255, + 0.451 + ], + "angle": 0, + "content": "(a) PSNR vs. mask ratio" + }, + { + "type": "image", + "bbox": [ + 0.287, + 0.323, + 0.482, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.321, + 0.439, + 0.45, + 0.451 + ], + "angle": 0, + "content": "(b) VFID vs. mask ratio" + }, + { + "type": "image_caption", + "bbox": [ + 0.143, + 0.457, + 0.431, + 0.471 + ], + "angle": 0, + "content": "Figure 5. Comparisons on different mask ratios." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.324, + 0.709, + 0.437 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.528, + 0.439, + 0.698, + 0.452 + ], + "angle": 0, + "content": "(a) PSNR vs. motion magnitude" + }, + { + "type": "image", + "bbox": [ + 0.713, + 0.323, + 0.904, + 0.437 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.724, + 0.439, + 0.892, + 0.451 + ], + "angle": 0, + "content": "(b) VFID vs. motion magnitude" + }, + { + "type": "image_caption", + "bbox": [ + 0.546, + 0.457, + 0.874, + 0.472 + ], + "angle": 0, + "content": "Figure 6. Comparisons on different motion magnitude." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.486, + 0.224, + 0.504 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.512, + 0.308, + 0.529 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.536, + 0.197, + 0.55 + ], + "angle": 0, + "content": "4.1.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.555, + 0.485, + 0.875 + ], + "angle": 0, + "content": "We use an internal video dataset including 5 million video sequences from Shutterstock1 as the training set. To simulate masks in video completion and object removal tasks, we synthesize stationary and object-aware masks in a random fashion, which are then pasted onto frames [17, 22, 63]. For evaluation, we use two video object segmentation datasets: DAVIS [30] and YouTube-VOS [55], which consist of 508 and 90 video sequences in test sets, respectively. For the DAVIS test set, following ProPainter [63], we use 50 video sequences for metric calculations. For mask settings, stationary masks are used to calculate objective metrics for the video completion task, and object-aware masks extracted from segmentation annotations are adopted for qualitative comparisons in the object removal task. Additionally, we adopt the Real-world Object Removal Dataset (RORD) [38] to evaluate the object removal task. The RORD test set includes 343 video pairs with and without objects, along with object mask annotations. This dataset addresses the lack of references for computing objective metrics in terms of the object removal task in previous work. For all datasets, video frames are sized to \\(240 \\times 432\\) for evaluation. Text prompts" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.486, + 0.907, + 0.517 + ], + "angle": 0, + "content": "for test sets are generated by GPT-4o [27] using the center frame of each corrupted video." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.526, + 0.614, + 0.54 + ], + "angle": 0, + "content": "4.1.2. Metrics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.545, + 0.907, + 0.636 + ], + "angle": 0, + "content": "Following widely adopted protocols, we use PSNR, SSIM [50], VFID [48], and warping error \\( E_{\\text{warp}} \\) [19] as evaluation metrics. Specifically, PSNR and SSIM are frequently used for distortion-oriented video assessment. VFID measures the perceptual similarity between input videos and outputs. \\( E_{\\text{warp}} \\) assesses the temporal stability." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.636, + 0.906, + 0.668 + ], + "angle": 0, + "content": "Detailed configurations of the proposed model for training and inference refer to the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.678, + 0.654, + 0.694 + ], + "angle": 0, + "content": "4.2. Comparisons" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.7, + 0.703, + 0.715 + ], + "angle": 0, + "content": "4.2.1. Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We report quantitative results on three datasets, using stationary masks for the YouTube-VOS and DAVIS datasets, and object-aware masks for the RORD dataset. We compare the proposed HomoGen with previous video inpainting methods, including DFVI [56], CPNet [20], FGVC [9], STTN [58], FuseFormer [24], FGT [59], ISVI [60], \\(\\mathrm{E}^2\\mathrm{FGVI}\\) [22], and ProPainter [63]. As shown in Tab. 1, HomoGen outperforms previous methods across most quantitative metrics. Notably, HomoGen significantly surpasses the state-of-the-art method by \\(23.81\\%\\), \\(21.43\\%\\), and \\(17.86\\%\\) in VFID on the YouTube-VOS, DAVIS, and RORD datasets, respectively, demonstrating its ability to generate" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.341, + 0.9 + ], + "angle": 0, + "content": "1https://wwwCRETstorstock.com" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "22958" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.087, + 0.905, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.211, + 0.393, + 0.786, + 0.409 + ], + "angle": 0, + "content": "Figure 7. Qualitative results compared with FuseFormer [24], \\(\\mathrm{E}^2\\mathrm{FGVI}\\) [22], and ProPainter [63]." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.419, + 0.479, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.097, + 0.531, + 0.475, + 0.546 + ], + "angle": 0, + "content": "Figure 8. Comparative user study on HomoGen and ProPainter." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.557, + 0.468, + 0.572 + ], + "angle": 0, + "content": "videos with more visually plausible and realistic content." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.572, + 0.484, + 0.783 + ], + "angle": 0, + "content": "To assess the model performance across different scenarios, we evaluate HomoGen on videos with varying mask ratios, as well as camera and object motion. We select ProPainter for comparisons, given its best performance among the baselines. In evaluation for different mask ratios, we partition the mask ratio range from \\([0, 0.5]\\) into five equal intervals, each corresponding to a synthesized mask set. We randomly sample 30 video sequences for each mask set. As shown in Fig. 5, HomoGen performs on par with ProPainter when the mask ratio is low. As the mask ratio increases, ProPainter's performance degrades more substantially than HomoGen's. In evaluation for different camera and object motion, we define the average pixel displacement \\(D^{p}\\) as a measure of motion, formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.79, + 0.484, + 0.833 + ], + "angle": 0, + "content": "\\[\nD ^ {p} = \\frac {1}{N H W} \\sum_ {n = 1} ^ {N} \\sum_ {i = 1} ^ {H} \\sum_ {j = 1} ^ {W} \\sqrt {\\left(d x _ {i , j , n}\\right) ^ {2} + \\left(d y _ {i , j , n}\\right) ^ {2}}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.485, + 0.902 + ], + "angle": 0, + "content": "where \\( dx \\) and \\( dy \\) denote the pixel displacement in the vertical and horizontal directions across frames based on the optical flows estimated by RAFT [44]. Similar to mask ratios, we select five video sets, each containing 30 video" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.416, + 0.907, + 0.499 + ], + "angle": 0, + "content": "Table 2. Ablation studies on various components in the two-stage pipeline. \"HP\" denotes the homography-based priors. \"FP\" denotes the flow-based priors. \"FCC\" denotes the flow consistency check. \"VT\" denotes the video transformer in ProPainter. \"VDM\" denotes PCVDM in the proposed model. (e) corresponds to ProPainter, and (f) corresponds to HomoGen." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.501, + 0.895, + 0.626 + ], + "angle": 0, + "content": "
ComponentsPSNRSSIMVFID
FPFCCHPVTVDM
(a)XXXX✓33.160.96150.042
(b)✓✓XX✓34.050.96840.036
(c)✓XXX✓34.230.96990.035
(d)XX✓✓X33.880.96530.044
(e)✓✓X✓X34.460.97220.042
(f)XX✓X✓34.810.97350.032
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.641, + 0.907, + 0.748 + ], + "angle": 0, + "content": "sequences, for various values of \\( D^p \\). As shown in Fig. 6, HomoGen outperforms ProPainter in most motion conditions, particularly those with minimal movements. The two quantitative results demonstrate that HomoGen has superior inpainting capability across various video scenarios, especially in more hard cases involving larger masks or less motion where the spatiotemporal context is limited." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.759, + 0.693, + 0.774 + ], + "angle": 0, + "content": "4.2.2. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Fig. 7 shows qualitative results of HomoGen compared to the state-of-the-art methods, including FuseFormer, \\(\\mathrm{E}^2\\mathrm{FGVI}\\), and ProPainter on both video completion and object removal tasks. Generally, HomoGen can generate more faithful textural and structural information, as well as preserve the semantic integrity of objects. Nevertheless, other compared methods tend to fail or produce artifacts such as blur, distortion, black hazy regions, etc." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22959" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.1, + 0.09, + 0.472, + 0.103 + ], + "angle": 0, + "content": "Table 3. Ablation studies on various components in HomoGen." + }, + { + "type": "table", + "bbox": [ + 0.098, + 0.104, + 0.472, + 0.182 + ], + "angle": 0, + "content": "
ModelPSNRSSIMVFID
w/o text prompts34.070.97160.039
w/o GoP-wise CLIP embeddings34.620.97220.033
HomoGen34.810.97350.032
" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.187, + 0.174, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.107, + 0.253, + 0.163, + 0.264 + ], + "angle": 0, + "content": "(i) line (a)" + }, + { + "type": "image", + "bbox": [ + 0.175, + 0.187, + 0.255, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.186, + 0.253, + 0.245, + 0.264 + ], + "angle": 0, + "content": "(ii) line (b)" + }, + { + "type": "image", + "bbox": [ + 0.256, + 0.187, + 0.336, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.266, + 0.253, + 0.327, + 0.264 + ], + "angle": 0, + "content": "(iii) line (c)" + }, + { + "type": "image", + "bbox": [ + 0.338, + 0.187, + 0.482, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.368, + 0.253, + 0.451, + 0.264 + ], + "angle": 0, + "content": "(iv) Input video" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.266, + 0.174, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.105, + 0.332, + 0.164, + 0.343 + ], + "angle": 0, + "content": "(v) line (d)" + }, + { + "type": "image", + "bbox": [ + 0.175, + 0.266, + 0.255, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.185, + 0.332, + 0.246, + 0.343 + ], + "angle": 0, + "content": "(vi) line (e)" + }, + { + "type": "image", + "bbox": [ + 0.256, + 0.266, + 0.337, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.266, + 0.332, + 0.328, + 0.343 + ], + "angle": 0, + "content": "(vii) line (f)" + }, + { + "type": "image", + "bbox": [ + 0.338, + 0.266, + 0.482, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.365, + 0.332, + 0.454, + 0.343 + ], + "angle": 0, + "content": "(viii) Input mask" + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.35, + 0.482, + 0.364 + ], + "angle": 0, + "content": "Figure 9. Qualitative comparisons of the ablation studies in Tab. 2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.377, + 0.484, + 0.513 + ], + "angle": 0, + "content": "Furthermore, we conduct a user study with 15 participants who are shown randomly sampled 45 video triplets. Each triplet contains one original video, one generated by HomoGen, and one generated by ProPainter. Volunteers are asked to select a visually better inpainting video or indicate a tie for each triplet. As shown in Fig. 8, volunteers prefer our results in terms of spatial coherence, temporal consistency, visual quality, and overall impression, demonstrating that HomoGen can generate more visually pleasant results." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.522, + 0.203, + 0.536 + ], + "angle": 0, + "content": "4.3. Ablations" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.543, + 0.483, + 0.574 + ], + "angle": 0, + "content": "We perform ablation studies on different technologies employed in HomoGen on the YouTube-VOS dataset." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.581, + 0.331, + 0.595 + ], + "angle": 0, + "content": "4.3.1. Study of Two-stage Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We investigate the importance of HPP tailored for PCVDM. As discussed in Sec. 1, various technologies have been designed for the two-stage pipeline, i.e., pixel propagation and content generation, among recent video inpainting methods. We compare different combinations of technologies within the two stages. For the first stage, we select HPP and the typical flow-based propagation proposed in ProPainter for comparisons. For the second stage, we select PCVDM and the video transformer proposed in ProPainter. As shown in line (a), Tab. 2, by removing HPP, temporal context is no longer provided to control content generation, resulting in a large performance decrease. The corresponding qualitative result in Fig. 9 shows artifacts and unrealistic content (the horse's leg). As demonstrated in line (b) and line (c), integrating flow-based priors with PCVDM, but without the consistency check, further enhances performance, which highlights the critical role of semantic integrity for DMs. Results in line (d) and line (e) indicate that the video transformer is not robust to artifacts in the priors and requires cleaner priors from pixel propagation. We also in" + }, + { + "type": "table_caption", + "bbox": [ + 0.518, + 0.09, + 0.901, + 0.104 + ], + "angle": 0, + "content": "Table 4. Ablation studies on various prior injection mechanisms." + }, + { + "type": "table", + "bbox": [ + 0.52, + 0.105, + 0.895, + 0.209 + ], + "angle": 0, + "content": "
ModelPSNRSSIMVFID
Pasting31.740.96100.053
Concatenation32.860.96430.043
Cross-attention34.140.97180.035
Noise prior33.520.96570.036
Content-adaptive controls34.810.97350.032
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.219, + 0.905, + 0.28 + ], + "angle": 0, + "content": "vestigate flow-based or homography-based priors combined with PCVDM. Results in line (c) and line (f) demonstrate that PCVDM exhibits greater tolerance to artifacts in these priors when coupled with HPP that provides more priors." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.286, + 0.9, + 0.301 + ], + "angle": 0, + "content": "4.3.2. Study of Prior Controlled Video Diffusion Model" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.305, + 0.907, + 0.668 + ], + "angle": 0, + "content": "We investigate the effectiveness of various components employed in PCVDM. We first remove the text prompts by replacing the visual descriptions with blank strings, for which the PSNR value is decreased by a large margin, i.e., 0.74 dB, as shown in Tab. 3. In contrast to the model without GoP-wise CLIP embeddings, the performance exhibits a non-trivial improvement. Furthermore, we compare the proposed content-adaptive control mechanism with other homography-based prior injection strategies, including pasting, concatenation, and cross-attention. Pasting represents directly pasting the priors into the corrupted regions, and others have been widely adopted in previous work [3, 25, 35, 49, 61]. Detailed implementations of different injection strategies refer to the supplementary. As shown in Tab. 4, pasting the priors into the corrupted regions raises a significant performance drop, which can be attributed to the fact that placing propagated and visible pixels on the same plane, making it difficult for DMs to differentiate between them. Compared to the second-best strategy employing cross-attention, using the content-adaptive controls improves performance by a non-trivial margin, i.e., 0.67 dB in terms of PSNR. This suggests that the content-adaptive control mechanism learns fine-grained dense injection and further enhances the control ability of PCVDM." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.68, + 0.634, + 0.695 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.906, + 0.81 + ], + "angle": 0, + "content": "Our work introduces a propagate-and-diffuse framework for video inpainting, named HomoGen, which utilizes homography registration to propagate contextual pixels for generating missing content in a video diffusion model. Experiments not only show HomoGen's superiority in generating realistic content but prove the synergy between homography registration and video diffusion models for inpainting." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.9 + ], + "angle": 0, + "content": "Limitations. There are several potential limitations in this study. For instance, generating identity-preserved content in videos with scene alternations and efficiently processing high-resolution videos remain challenges for both HomoGen and previous methods. These issues warrant further investigation and improvement in future work." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22960" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.156 + ], + "angle": 0, + "content": "[1] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.159, + 0.484, + 0.2 + ], + "angle": 0, + "content": "[2] Marcelo Bertalmio, Andrea L Bertozzi, and Guillermo Sapiro. Navier-stokes, fluid dynamics, and image and video inpainting. In CVPR, 2001. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.202, + 0.483, + 0.257 + ], + "angle": 0, + "content": "[3] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In CVPR, 2023. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.259, + 0.483, + 0.3 + ], + "angle": 0, + "content": "[4] Ya-Liang Chang, Zhe Yu Liu, Kuan-Ying Lee, and Winston Hsu. Free-form video inpainting with 3d gated convolution and temporal patchgan. In ICCV, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.302, + 0.483, + 0.369 + ], + "angle": 0, + "content": "[5] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In CVPR, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.372, + 0.483, + 0.4 + ], + "angle": 0, + "content": "[6] Prafulla Dhariwal and Alexander Quinn Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.402, + 0.483, + 0.442 + ], + "angle": 0, + "content": "[7] Mounira Ebdelli, Olivier Le Meur, and Christine Guillemot. Video inpainting with short-term windows: application to object removal and error concealment. IEEE TIP, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.445, + 0.483, + 0.499 + ], + "angle": 0, + "content": "[8] Ruoyu Feng, Wenming Weng, Yanhui Wang, Yuhui Yuan, Jianmin Bao, Chong Luo, Zhibo Chen, and Baining Guo. Credit: Creative and controllable video editing via diffusion models. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.501, + 0.483, + 0.529 + ], + "angle": 0, + "content": "[9] Chen Gao, Ayush Saraf, Jia-Bin Huang, and Johannes Kopf. Flow-edge guided video completion. In ECCV, 2020. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.531, + 0.483, + 0.598 + ], + "angle": 0, + "content": "[10] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. In ICLR, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.602, + 0.483, + 0.642 + ], + "angle": 0, + "content": "[11] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross-attention control. In ICLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.645, + 0.483, + 0.672 + ], + "angle": 0, + "content": "[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.674, + 0.483, + 0.715 + ], + "angle": 0, + "content": "[13] Yuan-Ting Hu, Heng Wang, Nicolas Ballas, Kristen Grauman, and Alexander G Schwing. Proposal-based video completion. In ECCV, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.717, + 0.483, + 0.758 + ], + "angle": 0, + "content": "[14] Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and Joao Carreira. Perceiver: General perception with iterative attention. In ICML, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.76, + 0.483, + 0.814 + ], + "angle": 0, + "content": "[15] Wonjoon Jin, Qi Dai, Chong Luo, Seung-Hwan Baek, and Sunghyun Cho. Flovd: Optical flow meets video diffusion model for enhanced camera-controlled video synthesis. In CVPR, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.483, + 0.87 + ], + "angle": 0, + "content": "[16] Xuan Ju, Xian Liu, Xintao Wang, Yuxuan Bian, Ying Shan, and Qiang Xu. Brushnet: A plug-and-play image inpainting model with decomposed dual-branch diffusion. In ECCV, 2024.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[17] Dahun Kim, Sanghyun Woo, Joon-Young Lee, and In So Kweon. Deep video inpainting. In CVPR, 2019. 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "[18] Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. In ICLR, 2014. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.123, + 0.905, + 0.164 + ], + "angle": 0, + "content": "[19] Wei-Sheng Lai, Jia-Bin Huang, Oliver Wang, Eli Shechtman, Ersin Yumer, and Ming-Hsuan Yang. Learning blind video temporal consistency. In ECCV, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.167, + 0.905, + 0.207 + ], + "angle": 0, + "content": "[20] Sungho Lee, Seoung Wug Oh, DaeYeun Won, and Seon Joo Kim. Copy-and-paste networks for deep video inpainting. In ICCV, 2019. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.21, + 0.905, + 0.265 + ], + "angle": 0, + "content": "[21] Ang Li, Shanshan Zhao, Xingjun Ma, Mingming Gong, Jianzhong Qi, Rui Zhang, Dacheng Tao, and Ramamohanarao Kotagiri. Short-term and long-term context aggregation network for video inpainting. In ECCV, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.268, + 0.905, + 0.309 + ], + "angle": 0, + "content": "[22] Zhen Li, Cheng-Ze Lu, Jianhua Qin, Chun-Le Guo, and Ming-Ming Cheng. Towards an end-to-end framework for flow-guided video inpainting. In CVPR, 2022. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.311, + 0.905, + 0.352 + ], + "angle": 0, + "content": "[23] Miao Liao, Feixiang Lu, Dingfu Zhou, Sibo Zhang, Wei Li, and Ruigang Yang. Dvi: Depth guided video inpainting for autonomous driving. In ECCV, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.355, + 0.905, + 0.41 + ], + "angle": 0, + "content": "[24] Rui Liu, Hanming Deng, Yangyi Huang, Xiaoyu Shi, Lewei Lu, Wenxiu Sun, Xiaogang Wang, Jifeng Dai, and Hongsheng Li. Fuseformer: Fusing fine-grained information in transformers for video inpainting. In ICCV, 2021. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.412, + 0.905, + 0.466 + ], + "angle": 0, + "content": "[25] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. In ICLR, 2022. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.905, + 0.524 + ], + "angle": 0, + "content": "[26] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In AAAI, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.527, + 0.905, + 0.554 + ], + "angle": 0, + "content": "[27] OpenAI. Gpt-4o contributions. https://openai.com/gpt-4o-contributions, 2024.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.557, + 0.905, + 0.597 + ], + "angle": 0, + "content": "[28] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In CVPR, 2019. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.601, + 0.905, + 0.641 + ], + "angle": 0, + "content": "[29] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In CVPR, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.644, + 0.905, + 0.699 + ], + "angle": 0, + "content": "[30] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In CVPR, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.702, + 0.905, + 0.768 + ], + "angle": 0, + "content": "[31] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: improving latent diffusion models for high-resolution image synthesis. In ICLR, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.773, + 0.905, + 0.827 + ], + "angle": 0, + "content": "[32] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[33] Prajit Ramachandran, Barret Zoph, and Quoc V. Le. Searching for activation functions. In ICLR Workshops, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[34] Jingjing Ren, Qingqing Zheng, Yuanyuan Zhao, Xuemiao Xu, and Chen Li. Dlformer: Discrete latent transformer for video inpainting. In CVPR, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22961" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[35] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 2, 3, 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.19 + ], + "angle": 0, + "content": "[36] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[37] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary Bradski. Orb: An efficient alternative to sift or surf. In ICCV. IEEE, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.482, + 0.275 + ], + "angle": 0, + "content": "[38] Min-Cheol Sagong, Yoon-Jae Yeo, Seung-Won Jung, and Sung-Jea Ko. Rord: A real-world object removal dataset. In BMVC, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.347 + ], + "angle": 0, + "content": "[39] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[40] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-a-video: Text-to-video generation without text-video data. In ICLR, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.419, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[41] Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.475, + 0.482, + 0.503 + ], + "angle": 0, + "content": "[42] Richard Szeliski. Image alignment and stitching: A tutorial. Found. Trends Comput. Graph. Vis., 2006. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.505, + 0.482, + 0.56 + ], + "angle": 0, + "content": "[43] Nick C Tang, Chiou-Ting Hsu, Chih-Wen Su, Timothy K Shih, and Hong-Yuan Mark Liao. Video inpainting on digitized vintage films via maintaining spatiotemporal continuity. IEEE TMM, 2011. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.482, + 0.589 + ], + "angle": 0, + "content": "[44] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In ECCV, 2020. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.591, + 0.482, + 0.657 + ], + "angle": 0, + "content": "[45] Rui Tian, Qi Dai, Jianmin Bao, Kai Qiu, Yifan Yang, Chong Luo, Zuxuan Wu, and Yu-Gang Jiang. Reducio! generating \\(1024 \\times 1024\\) video within 16 seconds using extremely compressed motion latents. arXiv preprint arXiv:2411.13552, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.482, + 0.702 + ], + "angle": 0, + "content": "[46] Shuyuan Tu, Qi Dai, Zhi-Qi Cheng, Han Hu, Xintong Han, Zuxuan Wu, and Yu-Gang Jiang. Motioneditor: Editing video motion via content-aware diffusion. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.482, + 0.744 + ], + "angle": 0, + "content": "[47] Chuan Wang, Haibin Huang, Xiaoguang Han, and Jue Wang. Video inpainting by jointly learning temporal structure and spatial details. In AAAI, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.747, + 0.482, + 0.788 + ], + "angle": 0, + "content": "[48] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Nikolai Yakovenko, Andrew Tao, Jan Kautz, and Bryan Catanzaro. Video-to-video synthesis. In NeurIPS, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[49] Yanhui Wang, Jianmin Bao, Wenming Weng, Ruoyu Feng, Dacheng Yin, Tao Yang, Jingxu Zhang, Qi Dai, Zhiyuan Zhao, Chunyu Wang, et al. Microcinema: A divide-and-conquer approach for text-to-video generation. In CVPR, 2024. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[50] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE TIP, 2004. 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[51] Wenming Weng, Ruoyu Feng, Yanhui Wang, Qi Dai, Chunyu Wang, Dacheng Yin, Zhiyuan Zhao, Kai Qiu, Jian-min Bao, Yuhui Yuan, et al. Art-v: Auto-regressive text-to-video generation with diffusion models. In CVPR Workshops, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[52] Yuxin Wu and Kaiming He. Group normalization. In ECCV, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.26 + ], + "angle": 0, + "content": "[53] Zhen Xing, Qi Dai, Han Hu, Zuxuan Wu, and Yu-Gang Jiang. Simda: Simple diffusion adapter for efficient video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7827-7839, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.262, + 0.905, + 0.316 + ], + "angle": 0, + "content": "[54] Zhen Xing, Qijun Feng, Haoran Chen, Qi Dai, Han Hu, Hang Xu, Zuxuan Wu, and Yu-Gang Jiang. A survey on video diffusion models. ACM Computing Surveys, 57(2):1-42, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.319, + 0.905, + 0.374 + ], + "angle": 0, + "content": "[55] Ning Xu, Linjie Yang, Yuchen Fan, Jianchao Yang, Dingcheng Yue, Yuchen Liang, Brian Price, Scott Cohen, and Thomas Huang. Youtube-vos: Sequence-to-sequence video object segmentation. In ECCV, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.376, + 0.905, + 0.416 + ], + "angle": 0, + "content": "[56] Rui Xu, Xiaoxiao Li, Bolei Zhou, and Chen Change Loy. Deep flow-guided video inpainting. In CVPR, 2019. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.418, + 0.905, + 0.46 + ], + "angle": 0, + "content": "[57] Jiahui Yu, Zhe Lin, Jamei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Generative image inpainting with contextual attention. In CVPR, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.462, + 0.905, + 0.502 + ], + "angle": 0, + "content": "[58] Yanhong Zeng, Jianlong Fu, and Hongyang Chao. Learning joint spatial-temporal transformations for video inpainting. In ECCV, 2020. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.905, + 0.532 + ], + "angle": 0, + "content": "[59] Kaidong Zhang, Jingjing Fu, and Dong Liu. Flow-guided transformer for video inpainting. In ECCV, 2022. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.534, + 0.905, + 0.574 + ], + "angle": 0, + "content": "[60] Kaidong Zhang, Jingjing Fu, and Dong Liu. Inertia-guided flow completion and style fusion for video inpainting. In CVPR, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.905, + 0.616 + ], + "angle": 0, + "content": "[61] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.619, + 0.905, + 0.659 + ], + "angle": 0, + "content": "[62] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In CVPR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.905, + 0.715 + ], + "angle": 0, + "content": "[63] Shangchen Zhou, Chongyi Li, Kelvin CK Chan, and Chen Change Loy. Propainter: Improving propagation and transformer for video inpainting. In ICCV, 2023. 2, 3, 5, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.715 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22962" + } + ] +] \ No newline at end of file diff --git a/2025/HomoGen_ Enhanced Video Inpainting via Homography Propagation and Diffusion/c0acdc17-6ead-4f9d-8631-4f1535650ac6_origin.pdf b/2025/HomoGen_ Enhanced Video Inpainting via Homography Propagation and Diffusion/c0acdc17-6ead-4f9d-8631-4f1535650ac6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..38b76bb39c80cc81f1a6d17a4a9f7d7c8beec99a --- /dev/null +++ b/2025/HomoGen_ Enhanced Video Inpainting via Homography Propagation and Diffusion/c0acdc17-6ead-4f9d-8631-4f1535650ac6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c021dbc6e40bf4a4f02823dc7e359c330114a75313579fcb8f5493d0f049f6ef +size 2557353 diff --git a/2025/HomoGen_ Enhanced Video Inpainting via Homography Propagation and Diffusion/full.md b/2025/HomoGen_ Enhanced Video Inpainting via Homography Propagation and Diffusion/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c9b5d5023e9c0277ad2eda057bf2f59fd535f6f9 --- /dev/null +++ b/2025/HomoGen_ Enhanced Video Inpainting via Homography Propagation and Diffusion/full.md @@ -0,0 +1,343 @@ +# HomoGen: Enhanced Video Inpainting via Homography Propagation and Diffusion + +Ding Ding $^{1*}$ Yueming Pan $^{2}$ Ruoyu Feng $^{3}$ Qi Dai $^{4}$ Kai Qiu $^{4}$ Jianmin Bao $^{4}$ Chong Luo $^{4\dagger}$ Zhenzhong Chen $^{1}$ $^{1}$ Wuhan University $^{2}$ Xi'an Jiaotong University $^{3}$ University of Science and Technology of China $^{4}$ Microsoft Research Asia + +![](images/bd84f48f97310edc40f31c56d8b40d7befc8a29a83109163af082a82ceeae6d0.jpg) +Figure 1. Two-stage pipelines of HomoGen and the state-of-the-art method. Previous flow-based propagation warps pixels across frames, followed by a video transformer (VT) that generates remaining missing content. In contrast, HomoGen's homography-based propagation warps pixels as priors to control content generation in a video diffusion model (VDM) that filters out artifacts in the priors during denoising. + +# Abstract + +In this paper, we present HomoGen, an enhanced video inpainting method based on homography propagation and diffusion models. HomoGen leverages homography registration to propagate contextual pixels as priors for generating missing content in corrupted videos. Unlike previous flow-based propagation methods, which introduce local distortions due to point-to-point optical flows, homography-induced artifacts are typically global structural distortions that preserve semantic integrity. To effectively utilize these priors for generation, we employ a video diffusion model that inherently prioritizes semantic information within the priors over pixel-level details. A content-adaptive control mechanism is proposed to scale and inject the priors into intermediate video latents during iterative denoising. In contrast to existing transformer-based networks that often suffer from artifacts within priors, leading to error accumulation and unrealistic results, our denoising diffusion net- + +work can smooth out artifacts and ensure natural outputs. Extensive experiments demonstrate the effectiveness of the proposed method qualitatively and quantitatively. + +# 1. Introduction + +Video inpainting aims to fill corrupted regions in a video with realistic and spatiotemporally coherent content [2]. It has diverse applications including object removal [7], video restoration [43], autonomous driving [23], etc. Despite significant progress in image inpainting [29, 35, 57], video inpainting remains full of challenges due to the complexity of video scenarios. The primary difficulty lies in identifying spatiotemporal guidance for generating content within corrupted regions, particularly in videos with diverse camera movements and object motion. To tackle this, previous work has explored various technologies, such as 3D convolutions [4, 13] and attentions [21, 24, 58], to leverage spatiotemporal context and preserve coherent across frames. + +Recently, two-stage methods combining flow-based propagation and generation networks have shown impres + +sive performance [22, 59, 63]. In flow-based propagation, optical flows are utilized to propagate pixels from adjacent frames into corrupted regions via warping. For generation networks, powerful video transformers are employed to generate remaining missing content. However, such methods still face challenges: the warping operation introduces artifacts such as blur and distortion in the propagated pixels, especially in scenarios characterized by significant object or camera motion and lighting changes [9]. As shown in Fig. 1, the flow-based propagation method in the state-of-the-art ProPainter [63] incorrectly warps a tree into the road center in the first stage. Besides, the propagated pixels are considered partial inpainting results to guide remaining content generation. Such artifacts cause error accumulation in video transformers, leading to unrealistic results [22, 59]. To mitigate artifacts in the propagated pixels, recent methods propose the flow consistency check to avoid inaccurate flows [56]. However, the consistency check discards partial flows, leaving the propagated pixels insufficient and semantically incoherent. As shown in Fig. 1, a subset of road pixels is removed from the propagated content after the consistency check. Therefore, video transformers struggle to generate high-fidelity content without sufficient guidance. + +Regarding the challenges in balancing retaining more propagated pixels while avoiding artifacts, diffusion models [12, 35] offer a promising solution as generation networks in the two-stage pipeline. Diffusion models prioritize semantic information within input conditions, such as stroke paintings and unnatural images [25, 28, 61], over pixel-level details. During denoising, diffusion models could generate realistic content while filtering out noisy low-level features from the conditions. Besides, homography registration for pixel propagation is more compatible with diffusion models. Specifically, artifacts from homography differ from those caused by flows. Flows capture point-to-point correlations [44], and flow-induced artifacts often considerably distort parts of objects in complex scenes, disrupting the semantics of propagated content. In contrast, homography imposes projective geometric constraints based on the global structure between frames [42], and homography-induced artifacts are typically structural distortions while preserving semantic integrity. As shown in Fig. 1, in the first stage, although several road pixels experience overall distortions, the structural integrity of objects remains largely intact. In the second stage, the video diffusion model generates more natural content while filtering out the artifacts. + +In this paper, we propose an enhanced video inpainting framework, named HomoGen, which enables pixel propagation via Homography registration and designs a video diffusion model to Generate missing content. For pixel propagation, we register paired frames via homography matrices to explore valuable contextual pixels. After warping the reference frame, the propagated pixels located in the + +corrupted regions of the target frame are extracted as priors. We iterate the propagation process across all frames in a bidirectional manner, where the reference frames are progressively updated with the propagated pixels at each step. For content generation, we design a prior controlled video diffusion model that captures the semantics within the bidirectional priors. Built upon a text-guided image inpainting model from Stable Diffusion, we integrate temporal layers into it to construct a 3D network. To effectively utilize the priors, we propose a content-adaptive control mechanism that scales and injects the priors into the diffusion model. Specifically, we design a lite ControlNet [61] to embed the priors into multi-scale features that are then fed into the 3D U-Net via dense injection. Learned scalers adjust the multi-scale features before injection, adapting to the spatiotemporal context of priors and the sampling steps. Simultaneously, the denoising process filters out artifacts within the priors during iterative sampling, thus preserving high-fidelity results across inpainted regions and visible pixels. + +Qualitative and quantitative experiment results demonstrate the superiority of HomoGen both in video completion and object removal. Compared to the state-of-the-art methods, HomoGen can generate more realistic content and improve performance by $23.81\%$ in VFID. + +- A propagate-and-diffuse framework is proposed for video inpainting, which leverages homography registration to propagate contextual pixels as priors to control the generation of missing content in a denoising diffusion network. +- A content-adaptive control mechanism is designed to inject prior information into the diffusion model. The dense injection learns to adapt to the spatiotemporal context of priors and the sampling steps, while artifacts within the priors are filtered out during iterative denoising. +- Extensive experiments show that HomoGen surpasses the state-of-the-art methods and generates more realistic content in video completion and object removal tasks. + +# 2. Related Work + +# 2.1. Diffusion-based Image and Video Generation + +Diffusion models have exhibited remarkable abilities in visual generation [3, 6, 15, 31, 35, 45, 49, 51, 53, 54]. These models learn to approximate data distributions through the iterative denoising of diffused inputs [12, 41]. Their practicality is enhanced by the incorporation of various conditions, e.g., text prompts, unnatural images, segmentation maps, etc., allowing for controlled outputs over the generative process. Controllable generation methods particularly benefit from additional structure modules, e.g., ControlNet [61], T2I-adapter [26], etc., which manage the injection of conditions. Such models leverage large-scale datasets [1, 39, 62] to achieve impressive results not only in generation but in editing [8, 11, 32, 46]. Conditional diffusion can + +also be applied to video inpainting, a widely explored video editing task, where the visible pixels serve as conditions to guide the generation of missing content in corrupted videos. + +# 2.2. Video Inpainting + +Previous video inpainting methods can be broadly divided into two categories: one-stage and two-stage approaches. One-stage approaches employ 3D CNNs and video transformers to aggregate spatiotemporal context for generating missing content in videos [4, 21, 24, 34, 58]. These methods usually struggle to capture spatiotemporal details due to limited receptive fields, which leads to blurry and inconsistent outcomes in complex scenes. Two-stage approaches have gradually become mainstream choices for video inpainting. The first stage primarily extracts prior information from the video itself, such as motion information [17, 43, 56], homography [20], low-resolution results [47], and proposal sets [13]. Guided by the priors, the second stage generates missing content based on well-designed networks, such as image inpainting models [9, 56, 60], 3D CNNs [17] and transformers [59]. Recent methods capitalize on the advantages of optical flows for motion capture and video transformers for spatiotemporal context modeling [9, 63]. As a typical approach, ProPainter [63] begins with flow extraction and completion, and utilizes completed flows to propagate pixels into corrupted regions. A flow-guided video transformer is designed to refine the propagated pixels and generate remaining missing content. + +However, such methods still suffer from blur and distortion in the results, primarily caused by the warping operation in pixel propagation and error accumulation in video transformers. To address this issue, we design a video diffusion model as the generation network that effectively removes artifacts within the priors during iterative denoising. Additionally, we employ homography-based pixel propagation to preserve the semantic integrity of warped content, providing fine-grained controls for conditional diffusion. + +# 3. Methodology + +HomoGen consists of two main components: Homography-based Pixel Propagation (HPP) and Prior Controlled Video Diffusion Model (PCVDM), as shown in Fig. 2. In HPP, we estimate homography between paired frames for propagating pixels as contextual priors. PCVDM then generates missing content controlled by these priors. + +# 3.1. Preliminary + +Consider a corrupted video sequence denoted as $\mathbf{x} = \{\mathbf{x}_i\mid$ $\mathbf{x}_i\in \mathbb{R}^{3\times H\times W},1\leq i\leq N\}$ comprising $N$ frames, accompanied by a binary mask sequence $\mathbf{m} = \{\mathbf{m}_i\mid \mathbf{m}_i\in$ $\mathbb{R}^{1\times H\times W},1\leq i\leq N\}$ identifying corrupted regions in all frames. Text prompts provide visual descriptions of the inpainted video. With text prompts as additional conditions, + +text-guided video inpainting aims to generate visually faithful and coherent content within the corrupted regions. + +As text-guided video inpainting encounters diffusion models (DMs), latent diffusion models (LDMs) [31, 35] are effective choices. LDMs leverage a pre-trained Variational Autoencoder (VAE) that consists of an encoder $\mathcal{E}$ and a decoder $\mathcal{D}$ to perform denoising in the latent space where the latent variable $\mathbf{z} = \mathcal{E}(\mathbf{x})$ . For video inpainting, conditions $\mathbf{c}$ are injected into a denoising network $\epsilon_{\theta}$ to predict $\mathbf{z}^t$ , a denoised variant of a noise, where $t = 0, 1, 2, \dots, T$ denotes the sampling step. The optimization objective is formulated as + +$$ +\mathbb {E} _ {\mathbf {z} ^ {0}, t, \epsilon \sim \mathcal {N} (\mathbf {0}, \mathbf {I})} \left[ \left\| \epsilon - \epsilon_ {\theta} \left(\mathbf {z} ^ {t}, t, \mathbf {c}\right) \right\| _ {2} ^ {2} \right], \tag {1} +$$ + +where $\mathbf{c} = (\mathbf{x},\mathbf{m},\tau_{\theta}(y))$ $y$ represents the text prompts and $\tau_{\theta}(\cdot)$ denotes a text encoder. + +# 3.2. Homography-based Pixel Propagation + +Pixel propagation aligns content within corrupted videos by propagating pixels from reference frames into corrupted regions in the target frame. Flow-based propagation typically warps pixels based on the trajectories of optical flows [17, 59]. However, they face challenges in balancing more propagated pixels and fewer distortions that cause error accumulation within the generation network. To overcome this, we employ homography-based propagation for content alignment. Compared to flows that capture point-to-point correlations, homography registration offers greater robustness since homography matrices for warping impose projective geometric constraints based on global structures between frames [42]. Besides, homography registration usually preserves the semantic integrity without altering parts of objects in complex scenes, as discussed in Sec. 1. + +To search for valuable pixels within a large temporal window, we estimate the forward and backward homography and iteratively propagate pixels across frames in a bidirectional manner. For simplicity, we only describe the forward propagation process. The propagated pixels $\mathbf{x}_{i,i + 1}^p$ in frame $\mathbf{x}_{i + 1}$ are warped from frame $\mathbf{x}_i$ as follows: + +$$ +\mathbf {x} _ {i, i + 1} ^ {p} = \left(1 - \mathbf {m} _ {i, i + 1} ^ {p}\right) \odot \mathbf {x} _ {i, i + 1}, \tag {2} +$$ + +$$ +\mathbf {x} _ {i, i + 1} = \mathcal {W} \left(\mathbf {x} _ {i} + \mathbf {x} _ {i - 1, i} ^ {p}, \mathcal {H} _ {i, i + 1}\right), \tag {3} +$$ + +$$ +\mathbf {m} _ {i, i + 1} ^ {p} = 1 - \mathbf {m} _ {i, i + 1} \odot \mathbf {m} _ {i + 1}, \tag {4} +$$ + +$$ +\mathbf {m} _ {i, i + 1} = \mathcal {W} \left(\mathbf {m} _ {i} \odot \mathbf {m} _ {i - 1, i} ^ {p}, \mathcal {H} _ {i, i + 1}\right), \tag {5} +$$ + +where $\mathbf{x}_{i,i + 1}$ and $\mathbf{m}_{i,i + 1}$ denotes warped $\mathbf{x}_i$ and $\mathbf{m}_i$ registered with $\mathbf{x}_{i + 1}$ and $\mathbf{m}_{i + 1}$ , respectively. $\mathbf{m}_{i + 1}^p$ denotes the mask indicating the regions of the propagated pixels. $\odot$ denotes the Hadamard product. $\mathcal{W}(\cdot ,\cdot)$ represents the warping operation. $\mathcal{H}_{i,i + 1}$ denotes the homography matrix for projecting the plane of $\mathbf{x}_i$ onto $\mathbf{x}_{i + 1}$ . $\mathcal{H}_{i,i + 1}$ is estimated by RANSAC on ORB feature matches [37]. During propagation, each reference frame is registered to the target frame. + +![](images/ddf6d486effa5817a165933c4503e8f1b630d985dfe509cec76d289d51e5c19c.jpg) +Figure 2. Overview of the proposed HomoGen framework, comprising two primary components: Homography-based Pixel Propagation (HPP) and Prior Controlled Video Diffusion Model (PCVDM). In this two-stage pipeline, HPP leverages homography between frames to propagate pixels as contextual priors that control missing content generation in PCVDM. + +![](images/d8eeff36c4947b82c41e32db08a1c1b12c56feb5635c0e448dfb27d74ad8cb0f.jpg) +(a) Flow-based propagation + +![](images/fd7268ed549ef194b14fbad67b15b4bbb6eecd5debb1fd55e294f50aedfaeb6b.jpg) +(b) HPP +Figure 3. Statistics of warping error and propagated pixel count on the YouTube-VOS test set in video completion. Warping error denotes the Mean Absolute Error (MAE) between the ground truth (normalized to $[-1,1]$ ) and the propagated pixels within a video. Pixel count denotes the proportion of propagated pixels in the corrupted regions. The grid color indicates the number of videos. + +After warping based on homography matrices, the propagated pixels located in the corrupted regions are extracted as priors. We iterate the propagation process across all frames, where the reference frames are progressively updated with the propagated pixels at each step. + +Valuable pixels are propagated under homography constraints. For each frame, we blend forward and backward propagated pixels $\mathbf{x}_{i - 1,i}^p$ and $\mathbf{x}_{i + 1,i}^p$ as follows: + +$$ +\mathbf {x} _ {i} ^ {p} = \mathbf {x} _ {i - 1, i} ^ {p} + \left(1 - \mathbf {m} _ {i - 1, i} ^ {p}\right) \odot \mathbf {m} _ {i + 1, i} ^ {p} \odot \mathbf {x} _ {i + 1, i} ^ {p}, \tag {6} +$$ + +where $\mathbf{x}_i^p$ denotes the bidirectional propagated pixels as the priors to control missing content generation in PCVDM. + +For further analysis, we present statistics on warping error and the count of propagated pixels for flow-based propagation and HPP. As shown in Fig. 3, the flow-based method propagates fewer pixels with lower warping error due to the consistency check, while HPP increases the pixel count at the cost of slightly higher warping error. Despite this added error, PCVDM benefits from richer spatiotemporal context + +and complete semantic information while filtering out artifacts in such priors during iterative denoising. + +# 3.3. Prior Controlled Video Diffusion Model + +Based on the priors from HPP, PCVDM aims to generate spatiotemporal coherent and realistic content in the corrupted regions. We propose a video diffusion model based on Stable Diffusion inpainting [35] and design a content-adaptive control mechanism to inject the priors into DM. + +# 3.3.1. Inflating Image Inpainter to Video Inpainter + +The proposed video inpainting model is extended from the image inpainting model in Stable Diffusion. Following previous work [10, 40], we extend the 2D DM into a 3D structure. In VAE decoder [18] and U-Net [36], original 2D layers are converted to pseudo-3D layers to learn temporal correlations, with detailed structures available in the supplementary. Besides, we employ an OpenCLIP image encoder and perceive attention layers to extract GoP-wise CLIP latents from the corrupted video [5, 14], and then inject the latents into the 3D U-Net via cross-attention. GoP refers to a Group of Picture, i.e., a video clip processed per inference. During training, the binary mask $\mathbf{m}$ , the noised video latent $\mathbf{z}^t$ , and the corrupted video latent $\mathbf{z} = \mathcal{E}(\mathbf{x})$ are concatenated in the latent space and fed into the 3D U-Net. During inference, $\mathbf{z}^t$ is initialized as Gaussian noise. The 3D U-Net generates the final sampled latent $\hat{\mathbf{z}}$ that is decoded by the 3D VAE decoder as the completed video $\hat{\mathbf{x}}$ . + +# 3.3.2. Content-adaptive Control Mechanism + +The priors are essential for generating missing content in corrupted videos. Recent work fills the propagated pix + +![](images/c6a798cf1d5a1c00938b66e0a67fb844ddc816eb4f2cd4a65bbd27f21c92b99f.jpg) +Figure 4. Illustration of content-adaptive control mechanism that injects prior information into PCVDM. Prior scalers in the dense injection are learned from the spatiotemporal context of priors and the sampling steps, enabling fine-grained controls. + +els into the corrupted regions as partial inpainting results and uses transformers to generate remaining missing content [22, 59, 63]. However, due to inevitable artifacts introduced by warping, pixel-level error accumulates during token prediction in transformers guided by noisy tokens from the priors, as discussed in Sec. 1. A trade-off solution is to discard inaccurate propagated pixels based on the consistency check [56] while limiting sufficient priors and semantic integrity. Unlike transformers' sensitivity to low-level features, DMs prioritize semantic information within input conditions over pixel-level details. Hence, to ensure the semantic integrity of the propagated content, we retained all propagated pixels without any consistency check. Besides, we consider the priors as conditions instead of partial inpainting results pasted with original visible pixels, which also helps the network unambiguously perceive artifacts from the priors. During iterative sampling, the denoising process can smooth out such undesirable artifacts. + +To effectively utilize the priors, we propose a content-adaptive control mechanism to scale and inject them into PCVDM for controlled missing content generation. The priors $\mathbf{x}^p = \{\mathbf{x}_i^p\mid \mathbf{x}_i^p\in \mathbb{R}^{3\times H\times W},1\leq i\leq N\}$ are first encoded into latents $\mathbf{z}^p = \{\mathbf{z}_i^p\mid \mathbf{z}_i^p\in \mathbb{R}^{4\times \frac{H}{8}\times \frac{W}{8}},1\leq i\leq N\}$ . We first design a LiteControlNet to extract multi-scale features from $\mathbf{z}^p$ for further injection. Modified from ControlNet [61], LiteControlNet features a lightweight structure with one convolutional layer at each scale. Detailed structures of LiteControlNet refer to the supplementary. The + +prior feature at each scale is connected to the 3D U-Net encoder through a zero convolutional layer and a learned projector, i.e., a 3D convolutional layer with a kernel size of $1 \times 1 \times 1$ . Rather than using additive operation, we introduce a spatiotemporal adaptive de-normalization strategy to inject each feature into the corresponding normalization layer [28]. As shown in Fig. 4, after the $j$ -th latent normalized by 3D group normalization [52] in the 3D U-Net, we compute the output activation $\mathbf{h}_j$ by de-normalizing the normalized latent $\bar{\mathbf{h}}_j$ according to the prior feature $\mathbf{f}_j$ , formulated as: + +$$ +\mathbf {h} _ {j} = \gamma_ {j} \odot \bar {\mathbf {h}} _ {j} + \beta_ {j} + \bar {\mathbf {h}} _ {j}, \tag {7} +$$ + +where $\gamma_{j}$ and $\beta_{j}$ are obtained through 3D convolutions of $\mathbf{f}_j$ . Equipped with this design, $\mathbf{f}_j$ is scaled and injected into the intermediate latent, where the generation of missing content adapts to the spatiotemporal context of the priors. Additionally, $\mathbf{f}_j$ is injected after attention blocks and yielded as the $(j + 1)$ -th latent $\mathbf{h}_{j + 1}$ , formulated as: + +$$ +\mathbf {h} _ {j + 1} = \zeta (t) \mathbf {f} _ {j} + \mathbf {h} _ {j}, \tag {8} +$$ + +where $\zeta(\cdot)$ represents two linear layers interleaved with a SiLU activation [33] in between. $\zeta(\cdot)$ encodes the sampling step $t$ [35] into Fourier features, which allows the scaling strength to vary with $t$ for fine-grained controls. + +During iterative denoising, the content in the corrupted regions progressively generates, approaching seamless integration with original visible pixels. After decoding the final sampled latent by the 3D VAE decoder, the generated and corrupted videos are blended with a blurred mask to improve coherence along the mask boundary [16]. + +# 3.4. Inpainting for Long Videos + +Video inpainting faces the challenge of maintaining temporal coherence across GoPs in long videos that span several seconds. Generative models are often restricted to processing all frames at once due to memory limitations, which leads to variability when sequentially inferring all GoPs. To address this, we introduce temporal references from adjacent GoPs to assist in inpainting the current GoP. + +Let $N$ denote the length of the GoP that HomoGen processes per inference. For a video containing $L$ frames, we segment the video into $\left\lceil \frac{L}{N} \right\rceil$ GoPs, each with an overlap of $N_{\text{overlap}}$ frames, and process all GoPs sequentially. For the $i$ -th GoP, the first and last frames are set to the center frames of the $(i-1)$ -th and $(i+1)$ -th GoP, respectively, while $N_{\text{overlap}} - 2$ frames from the second frame onward are set to the last frames of the $(i-1)$ -th GoP. In the initial inference, the first GoP undergoes inpainting. Before $i$ -th inference $(i > 1)$ , the beginning $N_{\text{overlap}} - 1$ frames of the $i$ -th GoP are replaced by the corresponding frames from the inpainted $(i-1)$ -th GoP. During $i$ -th inference, HomoGen can capture long-term temporal references for inpainting the $i$ -th GoP. This process iterates until all GoPs are inpainted, which ensures content consistency across the entire video. + +Table 1. Quantitative comparisons on YouTube-VOS, DAVIS, and RORD datasets. $\uparrow$ indicates higher is better. $\downarrow$ indicates lower is better. $E_{w}^{*}$ denotes $E_{\text{warp}} \times 10^{-3}$ . The best performances are highlighted in bold. All methods are evaluated following their default settings. + +
YouTube-VOSDAVISRORD
ModelsPSNR ↑SSIM ↑VFID ↓Ew* ↓PSNR ↑SSIM ↑VFID ↓Ew* ↓PSNR ↑SSIM ↑VFID ↓Ew* ↓
DFVI [56]29.160.94290.0661.50928.810.94040.1871.60831.840.95710.0651.137
CPNet [20]31.580.96070.0711.42330.280.95210.1821.51935.070.97400.0390.879
FGVC [9]29.670.94030.0641.02230.800.94970.1651.58632.760.96530.0521.005
STTN [58]32.570.96800.0490.90730.740.95900.1531.66134.940.97180.0440.872
FuseFormer [24]33.290.96810.0530.89132.540.97020.1381.33635.610.98110.0310.843
FGT [59]32.170.95990.0541.02532.860.96500.1291.32334.210.97240.0380.867
ISVI [60]30.380.94710.0771.31132.200.96260.1661.29233.450.96910.0460.902
E2FGVI [22]33.710.97000.0460.86433.010.97210.1161.31536.040.98000.0270.797
ProPainter [63]34.460.97220.0420.87234.190.97590.0981.19536.760.98670.0280.813
HomoGen34.810.97350.0320.85834.710.97630.0771.23437.290.98720.0230.793
+ +![](images/4becd39a5eb8747985cf7fe7aa91375719b5d6e523456739e44e9cb36f027baf.jpg) +(a) PSNR vs. mask ratio +Figure 5. Comparisons on different mask ratios. + +![](images/4838083238cf1141b23851635ff3c653e02081c431a925397b5e6a791a967362.jpg) +(b) VFID vs. mask ratio + +![](images/afad4e4cc0212255f4262b2392d4e8844a12645be75bcb86adf6cf90eabf279d.jpg) +(a) PSNR vs. motion magnitude +Figure 6. Comparisons on different motion magnitude. + +![](images/47e2665b9cf4430ceaaee5b1effe8058644f697c19637b80da99128a310f4bb9.jpg) +(b) VFID vs. motion magnitude + +# 4. Experiments + +# 4.1. Implementation Details + +# 4.1.1. Datasets + +We use an internal video dataset including 5 million video sequences from Shutterstock1 as the training set. To simulate masks in video completion and object removal tasks, we synthesize stationary and object-aware masks in a random fashion, which are then pasted onto frames [17, 22, 63]. For evaluation, we use two video object segmentation datasets: DAVIS [30] and YouTube-VOS [55], which consist of 508 and 90 video sequences in test sets, respectively. For the DAVIS test set, following ProPainter [63], we use 50 video sequences for metric calculations. For mask settings, stationary masks are used to calculate objective metrics for the video completion task, and object-aware masks extracted from segmentation annotations are adopted for qualitative comparisons in the object removal task. Additionally, we adopt the Real-world Object Removal Dataset (RORD) [38] to evaluate the object removal task. The RORD test set includes 343 video pairs with and without objects, along with object mask annotations. This dataset addresses the lack of references for computing objective metrics in terms of the object removal task in previous work. For all datasets, video frames are sized to $240 \times 432$ for evaluation. Text prompts + +for test sets are generated by GPT-4o [27] using the center frame of each corrupted video. + +# 4.1.2. Metrics + +Following widely adopted protocols, we use PSNR, SSIM [50], VFID [48], and warping error $E_{\text{warp}}$ [19] as evaluation metrics. Specifically, PSNR and SSIM are frequently used for distortion-oriented video assessment. VFID measures the perceptual similarity between input videos and outputs. $E_{\text{warp}}$ assesses the temporal stability. + +Detailed configurations of the proposed model for training and inference refer to the supplementary. + +# 4.2. Comparisons + +# 4.2.1. Quantitative Results + +We report quantitative results on three datasets, using stationary masks for the YouTube-VOS and DAVIS datasets, and object-aware masks for the RORD dataset. We compare the proposed HomoGen with previous video inpainting methods, including DFVI [56], CPNet [20], FGVC [9], STTN [58], FuseFormer [24], FGT [59], ISVI [60], $\mathrm{E}^2\mathrm{FGVI}$ [22], and ProPainter [63]. As shown in Tab. 1, HomoGen outperforms previous methods across most quantitative metrics. Notably, HomoGen significantly surpasses the state-of-the-art method by $23.81\%$ , $21.43\%$ , and $17.86\%$ in VFID on the YouTube-VOS, DAVIS, and RORD datasets, respectively, demonstrating its ability to generate + +![](images/0ca96261336b8c6677971935e684d72b1734e2012eb516d3fe41a3a528d28db4.jpg) +Figure 7. Qualitative results compared with FuseFormer [24], $\mathrm{E}^2\mathrm{FGVI}$ [22], and ProPainter [63]. + +![](images/89c92b7ef84712dd74a1c2352c361e41be8bc4f8ebd6ff61b3edeb4fcc67ec95.jpg) +Figure 8. Comparative user study on HomoGen and ProPainter. + +videos with more visually plausible and realistic content. + +To assess the model performance across different scenarios, we evaluate HomoGen on videos with varying mask ratios, as well as camera and object motion. We select ProPainter for comparisons, given its best performance among the baselines. In evaluation for different mask ratios, we partition the mask ratio range from $[0, 0.5]$ into five equal intervals, each corresponding to a synthesized mask set. We randomly sample 30 video sequences for each mask set. As shown in Fig. 5, HomoGen performs on par with ProPainter when the mask ratio is low. As the mask ratio increases, ProPainter's performance degrades more substantially than HomoGen's. In evaluation for different camera and object motion, we define the average pixel displacement $D^{p}$ as a measure of motion, formulated as: + +$$ +D ^ {p} = \frac {1}{N H W} \sum_ {n = 1} ^ {N} \sum_ {i = 1} ^ {H} \sum_ {j = 1} ^ {W} \sqrt {\left(d x _ {i , j , n}\right) ^ {2} + \left(d y _ {i , j , n}\right) ^ {2}}, \tag {9} +$$ + +where $dx$ and $dy$ denote the pixel displacement in the vertical and horizontal directions across frames based on the optical flows estimated by RAFT [44]. Similar to mask ratios, we select five video sets, each containing 30 video + +Table 2. Ablation studies on various components in the two-stage pipeline. "HP" denotes the homography-based priors. "FP" denotes the flow-based priors. "FCC" denotes the flow consistency check. "VT" denotes the video transformer in ProPainter. "VDM" denotes PCVDM in the proposed model. (e) corresponds to ProPainter, and (f) corresponds to HomoGen. + +
ComponentsPSNRSSIMVFID
FPFCCHPVTVDM
(a)XXXX✓33.160.96150.042
(b)✓✓XX✓34.050.96840.036
(c)✓XXX✓34.230.96990.035
(d)XX✓✓X33.880.96530.044
(e)✓✓X✓X34.460.97220.042
(f)XX✓X✓34.810.97350.032
+ +sequences, for various values of $D^p$ . As shown in Fig. 6, HomoGen outperforms ProPainter in most motion conditions, particularly those with minimal movements. The two quantitative results demonstrate that HomoGen has superior inpainting capability across various video scenarios, especially in more hard cases involving larger masks or less motion where the spatiotemporal context is limited. + +# 4.2.2. Qualitative Results + +Fig. 7 shows qualitative results of HomoGen compared to the state-of-the-art methods, including FuseFormer, $\mathrm{E}^2\mathrm{FGVI}$ , and ProPainter on both video completion and object removal tasks. Generally, HomoGen can generate more faithful textural and structural information, as well as preserve the semantic integrity of objects. Nevertheless, other compared methods tend to fail or produce artifacts such as blur, distortion, black hazy regions, etc. + +Table 3. Ablation studies on various components in HomoGen. + +
ModelPSNRSSIMVFID
w/o text prompts34.070.97160.039
w/o GoP-wise CLIP embeddings34.620.97220.033
HomoGen34.810.97350.032
+ +![](images/09c9d9c6a23ab68fa30556273e19c2a1f7509f657dc9934d9e9df307366cb785.jpg) +(i) line (a) + +![](images/580a118ac8bfe3c6689093529a29496086e55c721a878ae05f4ee5fa37b04365.jpg) +(ii) line (b) + +![](images/b96559d4740395cf54a9f40e6fd00294f67104e0a63d329e0164d41b8de22831.jpg) +(iii) line (c) + +![](images/a25aa309304e1ec060923652d6a242b8da1a274ddd05109ce0ecf7caea6fc103.jpg) +(iv) Input video + +![](images/c56238cca498c9f453af8fe09619e462da2cc9c323b87be5b60eb48f3304ca80.jpg) +(v) line (d) + +![](images/e183d3c1701e32f22d58281478113fcd92cf18b9c3ba6bf9db900a8eb82bb49c.jpg) +(vi) line (e) + +![](images/f549e3df9d6139a34a152d34a2d54aad7a8ded8eb40543056ec7a571c033d363.jpg) +(vii) line (f) + +![](images/db651aafdff4061ad50cc897629d59a120a74dce80b1c4ebf9201a500f732693.jpg) +(viii) Input mask +Figure 9. Qualitative comparisons of the ablation studies in Tab. 2. + +Furthermore, we conduct a user study with 15 participants who are shown randomly sampled 45 video triplets. Each triplet contains one original video, one generated by HomoGen, and one generated by ProPainter. Volunteers are asked to select a visually better inpainting video or indicate a tie for each triplet. As shown in Fig. 8, volunteers prefer our results in terms of spatial coherence, temporal consistency, visual quality, and overall impression, demonstrating that HomoGen can generate more visually pleasant results. + +# 4.3. Ablations + +We perform ablation studies on different technologies employed in HomoGen on the YouTube-VOS dataset. + +# 4.3.1. Study of Two-stage Pipeline + +We investigate the importance of HPP tailored for PCVDM. As discussed in Sec. 1, various technologies have been designed for the two-stage pipeline, i.e., pixel propagation and content generation, among recent video inpainting methods. We compare different combinations of technologies within the two stages. For the first stage, we select HPP and the typical flow-based propagation proposed in ProPainter for comparisons. For the second stage, we select PCVDM and the video transformer proposed in ProPainter. As shown in line (a), Tab. 2, by removing HPP, temporal context is no longer provided to control content generation, resulting in a large performance decrease. The corresponding qualitative result in Fig. 9 shows artifacts and unrealistic content (the horse's leg). As demonstrated in line (b) and line (c), integrating flow-based priors with PCVDM, but without the consistency check, further enhances performance, which highlights the critical role of semantic integrity for DMs. Results in line (d) and line (e) indicate that the video transformer is not robust to artifacts in the priors and requires cleaner priors from pixel propagation. We also in + +Table 4. Ablation studies on various prior injection mechanisms. + +
ModelPSNRSSIMVFID
Pasting31.740.96100.053
Concatenation32.860.96430.043
Cross-attention34.140.97180.035
Noise prior33.520.96570.036
Content-adaptive controls34.810.97350.032
+ +vestigate flow-based or homography-based priors combined with PCVDM. Results in line (c) and line (f) demonstrate that PCVDM exhibits greater tolerance to artifacts in these priors when coupled with HPP that provides more priors. + +# 4.3.2. Study of Prior Controlled Video Diffusion Model + +We investigate the effectiveness of various components employed in PCVDM. We first remove the text prompts by replacing the visual descriptions with blank strings, for which the PSNR value is decreased by a large margin, i.e., 0.74 dB, as shown in Tab. 3. In contrast to the model without GoP-wise CLIP embeddings, the performance exhibits a non-trivial improvement. Furthermore, we compare the proposed content-adaptive control mechanism with other homography-based prior injection strategies, including pasting, concatenation, and cross-attention. Pasting represents directly pasting the priors into the corrupted regions, and others have been widely adopted in previous work [3, 25, 35, 49, 61]. Detailed implementations of different injection strategies refer to the supplementary. As shown in Tab. 4, pasting the priors into the corrupted regions raises a significant performance drop, which can be attributed to the fact that placing propagated and visible pixels on the same plane, making it difficult for DMs to differentiate between them. Compared to the second-best strategy employing cross-attention, using the content-adaptive controls improves performance by a non-trivial margin, i.e., 0.67 dB in terms of PSNR. This suggests that the content-adaptive control mechanism learns fine-grained dense injection and further enhances the control ability of PCVDM. + +# 5. Conclusion + +Our work introduces a propagate-and-diffuse framework for video inpainting, named HomoGen, which utilizes homography registration to propagate contextual pixels for generating missing content in a video diffusion model. Experiments not only show HomoGen's superiority in generating realistic content but prove the synergy between homography registration and video diffusion models for inpainting. + +Limitations. There are several potential limitations in this study. For instance, generating identity-preserved content in videos with scene alternations and efficiently processing high-resolution videos remain challenges for both HomoGen and previous methods. These issues warrant further investigation and improvement in future work. + +# References + +[1] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In ICCV, 2021. 2 +[2] Marcelo Bertalmio, Andrea L Bertozzi, and Guillermo Sapiro. Navier-stokes, fluid dynamics, and image and video inpainting. In CVPR, 2001. 1 +[3] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In CVPR, 2023. 2, 8 +[4] Ya-Liang Chang, Zhe Yu Liu, Kuan-Ying Lee, and Winston Hsu. Free-form video inpainting with 3d gated convolution and temporal patchgan. In ICCV, 2019. 1, 3 +[5] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In CVPR, 2023. 4 +[6] Prafulla Dhariwal and Alexander Quinn Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021. 2 +[7] Mounira Ebdelli, Olivier Le Meur, and Christine Guillemot. Video inpainting with short-term windows: application to object removal and error concealment. IEEE TIP, 2015. 1 +[8] Ruoyu Feng, Wenming Weng, Yanhui Wang, Yuhui Yuan, Jianmin Bao, Chong Luo, Zhibo Chen, and Baining Guo. Credit: Creative and controllable video editing via diffusion models. In CVPR, 2024. 2 +[9] Chen Gao, Ayush Saraf, Jia-Bin Huang, and Johannes Kopf. Flow-edge guided video completion. In ECCV, 2020. 2, 3, 6 +[10] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. In ICLR, 2024. 4 +[11] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross-attention control. In ICLR, 2023. 2 +[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 2 +[13] Yuan-Ting Hu, Heng Wang, Nicolas Ballas, Kristen Grauman, and Alexander G Schwing. Proposal-based video completion. In ECCV, 2020. 1, 3 +[14] Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and Joao Carreira. Perceiver: General perception with iterative attention. In ICML, 2021. 4 +[15] Wonjoon Jin, Qi Dai, Chong Luo, Seung-Hwan Baek, and Sunghyun Cho. Flovd: Optical flow meets video diffusion model for enhanced camera-controlled video synthesis. In CVPR, 2025. 2 +[16] Xuan Ju, Xian Liu, Xintao Wang, Yuxuan Bian, Ying Shan, and Qiang Xu. Brushnet: A plug-and-play image inpainting model with decomposed dual-branch diffusion. In ECCV, 2024.5 +[17] Dahun Kim, Sanghyun Woo, Joon-Young Lee, and In So Kweon. Deep video inpainting. In CVPR, 2019. 3, 6 + +[18] Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. In ICLR, 2014. 4 +[19] Wei-Sheng Lai, Jia-Bin Huang, Oliver Wang, Eli Shechtman, Ersin Yumer, and Ming-Hsuan Yang. Learning blind video temporal consistency. In ECCV, 2018. 6 +[20] Sungho Lee, Seoung Wug Oh, DaeYeun Won, and Seon Joo Kim. Copy-and-paste networks for deep video inpainting. In ICCV, 2019. 3, 6 +[21] Ang Li, Shanshan Zhao, Xingjun Ma, Mingming Gong, Jianzhong Qi, Rui Zhang, Dacheng Tao, and Ramamohanarao Kotagiri. Short-term and long-term context aggregation network for video inpainting. In ECCV, 2020. 1, 3 +[22] Zhen Li, Cheng-Ze Lu, Jianhua Qin, Chun-Le Guo, and Ming-Ming Cheng. Towards an end-to-end framework for flow-guided video inpainting. In CVPR, 2022. 2, 5, 6, 7 +[23] Miao Liao, Feixiang Lu, Dingfu Zhou, Sibo Zhang, Wei Li, and Ruigang Yang. Dvi: Depth guided video inpainting for autonomous driving. In ECCV, 2020. 1 +[24] Rui Liu, Hanming Deng, Yangyi Huang, Xiaoyu Shi, Lewei Lu, Wenxiu Sun, Xiaogang Wang, Jifeng Dai, and Hongsheng Li. Fuseformer: Fusing fine-grained information in transformers for video inpainting. In ICCV, 2021. 1, 3, 6, 7 +[25] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. In ICLR, 2022. 2, 8 +[26] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In AAAI, 2024. 2 +[27] OpenAI. Gpt-4o contributions. https://openai.com/gpt-4o-contributions, 2024.6 +[28] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In CVPR, 2019. 2, 5 +[29] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In CVPR, 2016. 1 +[30] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In CVPR, 2016. 6 +[31] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: improving latent diffusion models for high-resolution image synthesis. In ICLR, 2024. 2, 3 +[32] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. In ICCV, 2023. 2 +[33] Prajit Ramachandran, Barret Zoph, and Quoc V. Le. Searching for activation functions. In ICLR Workshops, 2018. 5 +[34] Jingjing Ren, Qingqing Zheng, Yuanyuan Zhao, Xuemiao Xu, and Chen Li. Dlformer: Discrete latent transformer for video inpainting. In CVPR, 2022. 3 + +[35] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 2, 3, 4, 5, 8 +[36] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 4 +[37] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary Bradski. Orb: An efficient alternative to sift or surf. In ICCV. IEEE, 2011. 3 +[38] Min-Cheol Sagong, Yoon-Jae Yeo, Seung-Won Jung, and Sung-Jea Ko. Rord: A real-world object removal dataset. In BMVC, 2022. 6 +[39] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 2 +[40] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-a-video: Text-to-video generation without text-video data. In ICLR, 2023. 4 +[41] Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 2 +[42] Richard Szeliski. Image alignment and stitching: A tutorial. Found. Trends Comput. Graph. Vis., 2006. 2, 3 +[43] Nick C Tang, Chiou-Ting Hsu, Chih-Wen Su, Timothy K Shih, and Hong-Yuan Mark Liao. Video inpainting on digitized vintage films via maintaining spatiotemporal continuity. IEEE TMM, 2011. 1, 3 +[44] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In ECCV, 2020. 2, 7 +[45] Rui Tian, Qi Dai, Jianmin Bao, Kai Qiu, Yifan Yang, Chong Luo, Zuxuan Wu, and Yu-Gang Jiang. Reducio! generating $1024 \times 1024$ video within 16 seconds using extremely compressed motion latents. arXiv preprint arXiv:2411.13552, 2024. 2 +[46] Shuyuan Tu, Qi Dai, Zhi-Qi Cheng, Han Hu, Xintong Han, Zuxuan Wu, and Yu-Gang Jiang. Motioneditor: Editing video motion via content-aware diffusion. In CVPR, 2024. 2 +[47] Chuan Wang, Haibin Huang, Xiaoguang Han, and Jue Wang. Video inpainting by jointly learning temporal structure and spatial details. In AAAI, 2019. 3 +[48] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Nikolai Yakovenko, Andrew Tao, Jan Kautz, and Bryan Catanzaro. Video-to-video synthesis. In NeurIPS, 2018. 6 +[49] Yanhui Wang, Jianmin Bao, Wenming Weng, Ruoyu Feng, Dacheng Yin, Tao Yang, Jingxu Zhang, Qi Dai, Zhiyuan Zhao, Chunyu Wang, et al. Microcinema: A divide-and-conquer approach for text-to-video generation. In CVPR, 2024. 2, 8 +[50] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE TIP, 2004. 6 + +[51] Wenming Weng, Ruoyu Feng, Yanhui Wang, Qi Dai, Chunyu Wang, Dacheng Yin, Zhiyuan Zhao, Kai Qiu, Jian-min Bao, Yuhui Yuan, et al. Art-v: Auto-regressive text-to-video generation with diffusion models. In CVPR Workshops, 2024. 2 +[52] Yuxin Wu and Kaiming He. Group normalization. In ECCV, 2018. 5 +[53] Zhen Xing, Qi Dai, Han Hu, Zuxuan Wu, and Yu-Gang Jiang. Simda: Simple diffusion adapter for efficient video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7827-7839, 2024. 2 +[54] Zhen Xing, Qijun Feng, Haoran Chen, Qi Dai, Han Hu, Hang Xu, Zuxuan Wu, and Yu-Gang Jiang. A survey on video diffusion models. ACM Computing Surveys, 57(2):1-42, 2024. 2 +[55] Ning Xu, Linjie Yang, Yuchen Fan, Jianchao Yang, Dingcheng Yue, Yuchen Liang, Brian Price, Scott Cohen, and Thomas Huang. Youtube-vos: Sequence-to-sequence video object segmentation. In ECCV, 2018. 6 +[56] Rui Xu, Xiaoxiao Li, Bolei Zhou, and Chen Change Loy. Deep flow-guided video inpainting. In CVPR, 2019. 2, 3, 5, 6 +[57] Jiahui Yu, Zhe Lin, Jamei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Generative image inpainting with contextual attention. In CVPR, 2018. 1 +[58] Yanhong Zeng, Jianlong Fu, and Hongyang Chao. Learning joint spatial-temporal transformations for video inpainting. In ECCV, 2020. 1, 3, 6 +[59] Kaidong Zhang, Jingjing Fu, and Dong Liu. Flow-guided transformer for video inpainting. In ECCV, 2022. 2, 3, 5, 6 +[60] Kaidong Zhang, Jingjing Fu, and Dong Liu. Inertia-guided flow completion and style fusion for video inpainting. In CVPR, 2022. 3, 6 +[61] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 2, 5, 8 +[62] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In CVPR, 2017. 2 +[63] Shangchen Zhou, Chongyi Li, Kelvin CK Chan, and Chen Change Loy. Propainter: Improving propagation and transformer for video inpainting. In ICCV, 2023. 2, 3, 5, 6, 7 \ No newline at end of file diff --git a/2025/HomoGen_ Enhanced Video Inpainting via Homography Propagation and Diffusion/images.zip b/2025/HomoGen_ Enhanced Video Inpainting via Homography Propagation and Diffusion/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f750e3e12a208de96fed40a4b9bc18773208d8c7 --- /dev/null +++ b/2025/HomoGen_ Enhanced Video Inpainting via Homography Propagation and Diffusion/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dee5988a1dd046e9a3d5a91ef16b05071ad61cd15f7b2c7de09437fea7f1872b +size 857908 diff --git a/2025/HomoGen_ Enhanced Video Inpainting via Homography Propagation and Diffusion/layout.json b/2025/HomoGen_ Enhanced Video Inpainting via Homography Propagation and Diffusion/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f043f81bf6c7cd2d888aa4b7493877f56ebbe595 --- /dev/null +++ b/2025/HomoGen_ Enhanced Video Inpainting via Homography Propagation and Diffusion/layout.json @@ -0,0 +1,9855 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 130, + 103, + 480, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 103, + 480, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 103, + 480, + 140 + ], + "type": "text", + "content": "HomoGen: Enhanced Video Inpainting via Homography Propagation and Diffusion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 159, + 553, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 159, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 61, + 159, + 553, + 217 + ], + "type": "text", + "content": "Ding Ding" + }, + { + "bbox": [ + 61, + 159, + 553, + 217 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 61, + 159, + 553, + 217 + ], + "type": "text", + "content": " Yueming Pan" + }, + { + "bbox": [ + 61, + 159, + 553, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 159, + 553, + 217 + ], + "type": "text", + "content": " Ruoyu Feng" + }, + { + "bbox": [ + 61, + 159, + 553, + 217 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 61, + 159, + 553, + 217 + ], + "type": "text", + "content": " Qi Dai" + }, + { + "bbox": [ + 61, + 159, + 553, + 217 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 61, + 159, + 553, + 217 + ], + "type": "text", + "content": " Kai Qiu" + }, + { + "bbox": [ + 61, + 159, + 553, + 217 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 61, + 159, + 553, + 217 + ], + "type": "text", + "content": " Jianmin Bao" + }, + { + "bbox": [ + 61, + 159, + 553, + 217 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 61, + 159, + 553, + 217 + ], + "type": "text", + "content": " Chong Luo" + }, + { + "bbox": [ + 61, + 159, + 553, + 217 + ], + "type": "inline_equation", + "content": "^{4\\dagger}" + }, + { + "bbox": [ + 61, + 159, + 553, + 217 + ], + "type": "text", + "content": " Zhenzhong Chen" + }, + { + "bbox": [ + 61, + 159, + 553, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 159, + 553, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 159, + 553, + 217 + ], + "type": "text", + "content": "Wuhan University " + }, + { + "bbox": [ + 61, + 159, + 553, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 159, + 553, + 217 + ], + "type": "text", + "content": "Xi'an Jiaotong University " + }, + { + "bbox": [ + 61, + 159, + 553, + 217 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 61, + 159, + 553, + 217 + ], + "type": "text", + "content": "University of Science and Technology of China " + }, + { + "bbox": [ + 61, + 159, + 553, + 217 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 61, + 159, + 553, + 217 + ], + "type": "text", + "content": "Microsoft Research Asia" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 57, + 232, + 553, + 403 + ], + "blocks": [ + { + "bbox": [ + 57, + 232, + 553, + 403 + ], + "lines": [ + { + "bbox": [ + 57, + 232, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 57, + 232, + 553, + 403 + ], + "type": "image", + "image_path": "bd84f48f97310edc40f31c56d8b40d7befc8a29a83109163af082a82ceeae6d0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 407, + 555, + 442 + ], + "lines": [ + { + "bbox": [ + 55, + 407, + 555, + 442 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 555, + 442 + ], + "type": "text", + "content": "Figure 1. Two-stage pipelines of HomoGen and the state-of-the-art method. Previous flow-based propagation warps pixels across frames, followed by a video transformer (VT) that generates remaining missing content. In contrast, HomoGen's homography-based propagation warps pixels as priors to control content generation in a video diffusion model (VDM) that filters out artifacts in the priors during denoising." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 457, + 200, + 469 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 457, + 200, + 469 + ], + "spans": [ + { + "bbox": [ + 151, + 457, + 200, + 469 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 483, + 297, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 483, + 297, + 685 + ], + "spans": [ + { + "bbox": [ + 54, + 483, + 297, + 685 + ], + "type": "text", + "content": "In this paper, we present HomoGen, an enhanced video inpainting method based on homography propagation and diffusion models. HomoGen leverages homography registration to propagate contextual pixels as priors for generating missing content in corrupted videos. Unlike previous flow-based propagation methods, which introduce local distortions due to point-to-point optical flows, homography-induced artifacts are typically global structural distortions that preserve semantic integrity. To effectively utilize these priors for generation, we employ a video diffusion model that inherently prioritizes semantic information within the priors over pixel-level details. A content-adaptive control mechanism is proposed to scale and inject the priors into intermediate video latents during iterative denoising. In contrast to existing transformer-based networks that often suffer from artifacts within priors, leading to error accumulation and unrealistic results, our denoising diffusion net-" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 458, + 555, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 458, + 555, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 458, + 555, + 495 + ], + "type": "text", + "content": "work can smooth out artifacts and ensure natural outputs. Extensive experiments demonstrate the effectiveness of the proposed method qualitatively and quantitatively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 514, + 394, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 514, + 394, + 526 + ], + "spans": [ + { + "bbox": [ + 314, + 514, + 394, + 526 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 534, + 555, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 534, + 555, + 690 + ], + "spans": [ + { + "bbox": [ + 312, + 534, + 555, + 690 + ], + "type": "text", + "content": "Video inpainting aims to fill corrupted regions in a video with realistic and spatiotemporally coherent content [2]. It has diverse applications including object removal [7], video restoration [43], autonomous driving [23], etc. Despite significant progress in image inpainting [29, 35, 57], video inpainting remains full of challenges due to the complexity of video scenarios. The primary difficulty lies in identifying spatiotemporal guidance for generating content within corrupted regions, particularly in videos with diverse camera movements and object motion. To tackle this, previous work has explored various technologies, such as 3D convolutions [4, 13] and attentions [21, 24, 58], to leverage spatiotemporal context and preserve coherent across frames." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 691, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 691, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 691, + 555, + 715 + ], + "type": "text", + "content": "Recently, two-stage methods combining flow-based propagation and generation networks have shown impres" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 271, + 704 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 271, + 704 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 271, + 704 + ], + "type": "text", + "content": "*This work was done when Ding Ding was an intern at MSRA." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 67, + 704, + 144, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 704, + 144, + 714 + ], + "spans": [ + { + "bbox": [ + 67, + 704, + 144, + 714 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22953" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 347 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 347 + ], + "type": "text", + "content": "sive performance [22, 59, 63]. In flow-based propagation, optical flows are utilized to propagate pixels from adjacent frames into corrupted regions via warping. For generation networks, powerful video transformers are employed to generate remaining missing content. However, such methods still face challenges: the warping operation introduces artifacts such as blur and distortion in the propagated pixels, especially in scenarios characterized by significant object or camera motion and lighting changes [9]. As shown in Fig. 1, the flow-based propagation method in the state-of-the-art ProPainter [63] incorrectly warps a tree into the road center in the first stage. Besides, the propagated pixels are considered partial inpainting results to guide remaining content generation. Such artifacts cause error accumulation in video transformers, leading to unrealistic results [22, 59]. To mitigate artifacts in the propagated pixels, recent methods propose the flow consistency check to avoid inaccurate flows [56]. However, the consistency check discards partial flows, leaving the propagated pixels insufficient and semantically incoherent. As shown in Fig. 1, a subset of road pixels is removed from the propagated content after the consistency check. Therefore, video transformers struggle to generate high-fidelity content without sufficient guidance." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 351, + 295, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 351, + 295, + 625 + ], + "spans": [ + { + "bbox": [ + 56, + 351, + 295, + 625 + ], + "type": "text", + "content": "Regarding the challenges in balancing retaining more propagated pixels while avoiding artifacts, diffusion models [12, 35] offer a promising solution as generation networks in the two-stage pipeline. Diffusion models prioritize semantic information within input conditions, such as stroke paintings and unnatural images [25, 28, 61], over pixel-level details. During denoising, diffusion models could generate realistic content while filtering out noisy low-level features from the conditions. Besides, homography registration for pixel propagation is more compatible with diffusion models. Specifically, artifacts from homography differ from those caused by flows. Flows capture point-to-point correlations [44], and flow-induced artifacts often considerably distort parts of objects in complex scenes, disrupting the semantics of propagated content. In contrast, homography imposes projective geometric constraints based on the global structure between frames [42], and homography-induced artifacts are typically structural distortions while preserving semantic integrity. As shown in Fig. 1, in the first stage, although several road pixels experience overall distortions, the structural integrity of objects remains largely intact. In the second stage, the video diffusion model generates more natural content while filtering out the artifacts." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 630, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 294, + 713 + ], + "type": "text", + "content": "In this paper, we propose an enhanced video inpainting framework, named HomoGen, which enables pixel propagation via Homography registration and designs a video diffusion model to Generate missing content. For pixel propagation, we register paired frames via homography matrices to explore valuable contextual pixels. After warping the reference frame, the propagated pixels located in the" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 553, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 300 + ], + "type": "text", + "content": "corrupted regions of the target frame are extracted as priors. We iterate the propagation process across all frames in a bidirectional manner, where the reference frames are progressively updated with the propagated pixels at each step. For content generation, we design a prior controlled video diffusion model that captures the semantics within the bidirectional priors. Built upon a text-guided image inpainting model from Stable Diffusion, we integrate temporal layers into it to construct a 3D network. To effectively utilize the priors, we propose a content-adaptive control mechanism that scales and injects the priors into the diffusion model. Specifically, we design a lite ControlNet [61] to embed the priors into multi-scale features that are then fed into the 3D U-Net via dense injection. Learned scalers adjust the multi-scale features before injection, adapting to the spatiotemporal context of priors and the sampling steps. Simultaneously, the denoising process filters out artifacts within the priors during iterative sampling, thus preserving high-fidelity results across inpainted regions and visible pixels." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 300, + 553, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 553, + 361 + ], + "type": "text", + "content": "Qualitative and quantitative experiment results demonstrate the superiority of HomoGen both in video completion and object removal. Compared to the state-of-the-art methods, HomoGen can generate more realistic content and improve performance by " + }, + { + "bbox": [ + 313, + 300, + 553, + 361 + ], + "type": "inline_equation", + "content": "23.81\\%" + }, + { + "bbox": [ + 313, + 300, + 553, + 361 + ], + "type": "text", + "content": " in VFID." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 361, + 553, + 504 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 314, + 361, + 553, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 361, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 314, + 361, + 553, + 408 + ], + "type": "text", + "content": "- A propagate-and-diffuse framework is proposed for video inpainting, which leverages homography registration to propagate contextual pixels as priors to control the generation of missing content in a denoising diffusion network." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 315, + 409, + 553, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 409, + 553, + 468 + ], + "spans": [ + { + "bbox": [ + 315, + 409, + 553, + 468 + ], + "type": "text", + "content": "- A content-adaptive control mechanism is designed to inject prior information into the diffusion model. The dense injection learns to adapt to the spatiotemporal context of priors and the sampling steps, while artifacts within the priors are filtered out during iterative denoising." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 315, + 469, + 553, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 469, + 553, + 504 + ], + "spans": [ + { + "bbox": [ + 315, + 469, + 553, + 504 + ], + "type": "text", + "content": "- Extensive experiments show that HomoGen surpasses the state-of-the-art methods and generates more realistic content in video completion and object removal tasks." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 518, + 400, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 518, + 400, + 530 + ], + "spans": [ + { + "bbox": [ + 314, + 518, + 400, + 530 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 540, + 548, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 548, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 548, + 552 + ], + "type": "text", + "content": "2.1. Diffusion-based Image and Video Generation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 557, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 553, + 713 + ], + "type": "text", + "content": "Diffusion models have exhibited remarkable abilities in visual generation [3, 6, 15, 31, 35, 45, 49, 51, 53, 54]. These models learn to approximate data distributions through the iterative denoising of diffused inputs [12, 41]. Their practicality is enhanced by the incorporation of various conditions, e.g., text prompts, unnatural images, segmentation maps, etc., allowing for controlled outputs over the generative process. Controllable generation methods particularly benefit from additional structure modules, e.g., ControlNet [61], T2I-adapter [26], etc., which manage the injection of conditions. Such models leverage large-scale datasets [1, 39, 62] to achieve impressive results not only in generation but in editing [8, 11, 32, 46]. Conditional diffusion can" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22954" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "content": "also be applied to video inpainting, a widely explored video editing task, where the visible pixels serve as conditions to guide the generation of missing content in corrupted videos." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 115, + 159, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 115, + 159, + 129 + ], + "spans": [ + { + "bbox": [ + 55, + 115, + 159, + 129 + ], + "type": "text", + "content": "2.2. Video Inpainting" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 133, + 296, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 296, + 407 + ], + "type": "text", + "content": "Previous video inpainting methods can be broadly divided into two categories: one-stage and two-stage approaches. One-stage approaches employ 3D CNNs and video transformers to aggregate spatiotemporal context for generating missing content in videos [4, 21, 24, 34, 58]. These methods usually struggle to capture spatiotemporal details due to limited receptive fields, which leads to blurry and inconsistent outcomes in complex scenes. Two-stage approaches have gradually become mainstream choices for video inpainting. The first stage primarily extracts prior information from the video itself, such as motion information [17, 43, 56], homography [20], low-resolution results [47], and proposal sets [13]. Guided by the priors, the second stage generates missing content based on well-designed networks, such as image inpainting models [9, 56, 60], 3D CNNs [17] and transformers [59]. Recent methods capitalize on the advantages of optical flows for motion capture and video transformers for spatiotemporal context modeling [9, 63]. As a typical approach, ProPainter [63] begins with flow extraction and completion, and utilizes completed flows to propagate pixels into corrupted regions. A flow-guided video transformer is designed to refine the propagated pixels and generate remaining missing content." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 408, + 296, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 408, + 296, + 515 + ], + "spans": [ + { + "bbox": [ + 55, + 408, + 296, + 515 + ], + "type": "text", + "content": "However, such methods still suffer from blur and distortion in the results, primarily caused by the warping operation in pixel propagation and error accumulation in video transformers. To address this issue, we design a video diffusion model as the generation network that effectively removes artifacts within the priors during iterative denoising. Additionally, we employ homography-based pixel propagation to preserve the semantic integrity of warped content, providing fine-grained controls for conditional diffusion." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 525, + 138, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 138, + 539 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 138, + 539 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 545, + 296, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 617 + ], + "type": "text", + "content": "HomoGen consists of two main components: Homography-based Pixel Propagation (HPP) and Prior Controlled Video Diffusion Model (PCVDM), as shown in Fig. 2. In HPP, we estimate homography between paired frames for propagating pixels as contextual priors. PCVDM then generates missing content controlled by these priors." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 624, + 135, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 135, + 637 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 135, + 637 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "Consider a corrupted video sequence denoted as " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x} = \\{\\mathbf{x}_i\\mid" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i\\in \\mathbb{R}^{3\\times H\\times W},1\\leq i\\leq N\\}" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " comprising " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " frames, accompanied by a binary mask sequence " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{m} = \\{\\mathbf{m}_i\\mid \\mathbf{m}_i\\in" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{1\\times H\\times W},1\\leq i\\leq N\\}" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " identifying corrupted regions in all frames. Text prompts provide visual descriptions of the inpainted video. With text prompts as additional conditions," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "text-guided video inpainting aims to generate visually faithful and coherent content within the corrupted regions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 96, + 555, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 96, + 555, + 213 + ], + "spans": [ + { + "bbox": [ + 313, + 96, + 555, + 213 + ], + "type": "text", + "content": "As text-guided video inpainting encounters diffusion models (DMs), latent diffusion models (LDMs) [31, 35] are effective choices. LDMs leverage a pre-trained Variational Autoencoder (VAE) that consists of an encoder " + }, + { + "bbox": [ + 313, + 96, + 555, + 213 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 313, + 96, + 555, + 213 + ], + "type": "text", + "content": " and a decoder " + }, + { + "bbox": [ + 313, + 96, + 555, + 213 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 96, + 555, + 213 + ], + "type": "text", + "content": " to perform denoising in the latent space where the latent variable " + }, + { + "bbox": [ + 313, + 96, + 555, + 213 + ], + "type": "inline_equation", + "content": "\\mathbf{z} = \\mathcal{E}(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 96, + 555, + 213 + ], + "type": "text", + "content": ". For video inpainting, conditions " + }, + { + "bbox": [ + 313, + 96, + 555, + 213 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 313, + 96, + 555, + 213 + ], + "type": "text", + "content": " are injected into a denoising network " + }, + { + "bbox": [ + 313, + 96, + 555, + 213 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 313, + 96, + 555, + 213 + ], + "type": "text", + "content": " to predict " + }, + { + "bbox": [ + 313, + 96, + 555, + 213 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^t" + }, + { + "bbox": [ + 313, + 96, + 555, + 213 + ], + "type": "text", + "content": ", a denoised variant of a noise, where " + }, + { + "bbox": [ + 313, + 96, + 555, + 213 + ], + "type": "inline_equation", + "content": "t = 0, 1, 2, \\dots, T" + }, + { + "bbox": [ + 313, + 96, + 555, + 213 + ], + "type": "text", + "content": " denotes the sampling step. The optimization objective is formulated as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 358, + 213, + 553, + 232 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 213, + 553, + 232 + ], + "spans": [ + { + "bbox": [ + 358, + 213, + 553, + 232 + ], + "type": "interline_equation", + "content": "\\mathbb {E} _ {\\mathbf {z} ^ {0}, t, \\epsilon \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I})} \\left[ \\left\\| \\epsilon - \\epsilon_ {\\theta} \\left(\\mathbf {z} ^ {t}, t, \\mathbf {c}\\right) \\right\\| _ {2} ^ {2} \\right], \\tag {1}", + "image_path": "77a4d60b60a94d533927a636ac24b77c7ccbd4a7870105fb542344ebb58ef105.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 236, + 554, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 554, + 261 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 554, + 261 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 236, + 554, + 261 + ], + "type": "inline_equation", + "content": "\\mathbf{c} = (\\mathbf{x},\\mathbf{m},\\tau_{\\theta}(y))" + }, + { + "bbox": [ + 313, + 236, + 554, + 261 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 236, + 554, + 261 + ], + "type": "text", + "content": " represents the text prompts and " + }, + { + "bbox": [ + 313, + 236, + 554, + 261 + ], + "type": "inline_equation", + "content": "\\tau_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 313, + 236, + 554, + 261 + ], + "type": "text", + "content": " denotes a text encoder." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 268, + 515, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 268, + 515, + 281 + ], + "spans": [ + { + "bbox": [ + 313, + 268, + 515, + 281 + ], + "type": "text", + "content": "3.2. Homography-based Pixel Propagation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 285, + 555, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 555, + 464 + ], + "type": "text", + "content": "Pixel propagation aligns content within corrupted videos by propagating pixels from reference frames into corrupted regions in the target frame. Flow-based propagation typically warps pixels based on the trajectories of optical flows [17, 59]. However, they face challenges in balancing more propagated pixels and fewer distortions that cause error accumulation within the generation network. To overcome this, we employ homography-based propagation for content alignment. Compared to flows that capture point-to-point correlations, homography registration offers greater robustness since homography matrices for warping impose projective geometric constraints based on global structures between frames [42]. Besides, homography registration usually preserves the semantic integrity without altering parts of objects in complex scenes, as discussed in Sec. 1." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 465, + 554, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 465, + 554, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 465, + 554, + 536 + ], + "type": "text", + "content": "To search for valuable pixels within a large temporal window, we estimate the forward and backward homography and iteratively propagate pixels across frames in a bidirectional manner. For simplicity, we only describe the forward propagation process. The propagated pixels " + }, + { + "bbox": [ + 313, + 465, + 554, + 536 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{i,i + 1}^p" + }, + { + "bbox": [ + 313, + 465, + 554, + 536 + ], + "type": "text", + "content": " in frame " + }, + { + "bbox": [ + 313, + 465, + 554, + 536 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{i + 1}" + }, + { + "bbox": [ + 313, + 465, + 554, + 536 + ], + "type": "text", + "content": " are warped from frame " + }, + { + "bbox": [ + 313, + 465, + 554, + 536 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 313, + 465, + 554, + 536 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 361, + 545, + 553, + 560 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 545, + 553, + 560 + ], + "spans": [ + { + "bbox": [ + 361, + 545, + 553, + 560 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {i, i + 1} ^ {p} = \\left(1 - \\mathbf {m} _ {i, i + 1} ^ {p}\\right) \\odot \\mathbf {x} _ {i, i + 1}, \\tag {2}", + "image_path": "32408f9ff0fc2a76f34a35c62afaf7dddc859a807947e6f5f2698503431089a4.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 361, + 562, + 553, + 576 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 562, + 553, + 576 + ], + "spans": [ + { + "bbox": [ + 361, + 562, + 553, + 576 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {i, i + 1} = \\mathcal {W} \\left(\\mathbf {x} _ {i} + \\mathbf {x} _ {i - 1, i} ^ {p}, \\mathcal {H} _ {i, i + 1}\\right), \\tag {3}", + "image_path": "3a5d39b1e4d488d546782dcd369eed648aa9208c3d5c6e037c2ff1acb6ed022a.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 359, + 578, + 553, + 593 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 578, + 553, + 593 + ], + "spans": [ + { + "bbox": [ + 359, + 578, + 553, + 593 + ], + "type": "interline_equation", + "content": "\\mathbf {m} _ {i, i + 1} ^ {p} = 1 - \\mathbf {m} _ {i, i + 1} \\odot \\mathbf {m} _ {i + 1}, \\tag {4}", + "image_path": "bace714ce2d4b79abe74d1eec35da131470ff4d3c1c1554663f345a4914195da.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 359, + 595, + 553, + 610 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 595, + 553, + 610 + ], + "spans": [ + { + "bbox": [ + 359, + 595, + 553, + 610 + ], + "type": "interline_equation", + "content": "\\mathbf {m} _ {i, i + 1} = \\mathcal {W} \\left(\\mathbf {m} _ {i} \\odot \\mathbf {m} _ {i - 1, i} ^ {p}, \\mathcal {H} _ {i, i + 1}\\right), \\tag {5}", + "image_path": "834f5ef0baaa50b8cd17ccb7c5309d7784e8b42a9e77f3bdec7ce072789fa7ac.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{i,i + 1}" + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_{i,i + 1}" + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "text", + "content": " denotes warped " + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_i" + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "text", + "content": " registered with " + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{i + 1}" + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_{i + 1}" + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "text", + "content": ", respectively. " + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_{i + 1}^p" + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "text", + "content": " denotes the mask indicating the regions of the propagated pixels. " + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "text", + "content": " denotes the Hadamard product. " + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{W}(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "text", + "content": " represents the warping operation. " + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{i,i + 1}" + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "text", + "content": " denotes the homography matrix for projecting the plane of " + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "text", + "content": " onto " + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{i + 1}" + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{i,i + 1}" + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "text", + "content": " is estimated by RANSAC on ORB feature matches [37]. During propagation, each reference frame is registered to the target frame." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22955" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 73, + 553, + 281 + ], + "blocks": [ + { + "bbox": [ + 58, + 73, + 553, + 281 + ], + "lines": [ + { + "bbox": [ + 58, + 73, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 58, + 73, + 553, + 281 + ], + "type": "image", + "image_path": "ddf6d486effa5817a165933c4503e8f1b630d985dfe509cec76d289d51e5c19c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 285, + 555, + 319 + ], + "lines": [ + { + "bbox": [ + 55, + 285, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 555, + 319 + ], + "type": "text", + "content": "Figure 2. Overview of the proposed HomoGen framework, comprising two primary components: Homography-based Pixel Propagation (HPP) and Prior Controlled Video Diffusion Model (PCVDM). In this two-stage pipeline, HPP leverages homography between frames to propagate pixels as contextual priors that control missing content generation in PCVDM." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 56, + 327, + 175, + 388 + ], + "blocks": [ + { + "bbox": [ + 56, + 327, + 175, + 388 + ], + "lines": [ + { + "bbox": [ + 56, + 327, + 175, + 388 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 175, + 388 + ], + "type": "image", + "image_path": "d8eeff36c4947b82c41e32db08a1c1b12c56feb5635c0e448dfb27d74ad8cb0f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 71, + 390, + 160, + 399 + ], + "lines": [ + { + "bbox": [ + 71, + 390, + 160, + 399 + ], + "spans": [ + { + "bbox": [ + 71, + 390, + 160, + 399 + ], + "type": "text", + "content": "(a) Flow-based propagation" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 177, + 327, + 294, + 388 + ], + "blocks": [ + { + "bbox": [ + 177, + 327, + 294, + 388 + ], + "lines": [ + { + "bbox": [ + 177, + 327, + 294, + 388 + ], + "spans": [ + { + "bbox": [ + 177, + 327, + 294, + 388 + ], + "type": "image", + "image_path": "fd7268ed549ef194b14fbad67b15b4bbb6eecd5debb1fd55e294f50aedfaeb6b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 222, + 390, + 249, + 399 + ], + "lines": [ + { + "bbox": [ + 222, + 390, + 249, + 399 + ], + "spans": [ + { + "bbox": [ + 222, + 390, + 249, + 399 + ], + "type": "text", + "content": "(b) HPP" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 404, + 296, + 471 + ], + "lines": [ + { + "bbox": [ + 55, + 404, + 296, + 471 + ], + "spans": [ + { + "bbox": [ + 55, + 404, + 296, + 471 + ], + "type": "text", + "content": "Figure 3. Statistics of warping error and propagated pixel count on the YouTube-VOS test set in video completion. Warping error denotes the Mean Absolute Error (MAE) between the ground truth (normalized to " + }, + { + "bbox": [ + 55, + 404, + 296, + 471 + ], + "type": "inline_equation", + "content": "[-1,1]" + }, + { + "bbox": [ + 55, + 404, + 296, + 471 + ], + "type": "text", + "content": ") and the propagated pixels within a video. Pixel count denotes the proportion of propagated pixels in the corrupted regions. The grid color indicates the number of videos." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 481, + 296, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 481, + 296, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 481, + 296, + 540 + ], + "type": "text", + "content": "After warping based on homography matrices, the propagated pixels located in the corrupted regions are extracted as priors. We iterate the propagation process across all frames, where the reference frames are progressively updated with the propagated pixels at each step." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 540, + 296, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 296, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 296, + 578 + ], + "type": "text", + "content": "Valuable pixels are propagated under homography constraints. For each frame, we blend forward and backward propagated pixels " + }, + { + "bbox": [ + 55, + 540, + 296, + 578 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{i - 1,i}^p" + }, + { + "bbox": [ + 55, + 540, + 296, + 578 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 540, + 296, + 578 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{i + 1,i}^p" + }, + { + "bbox": [ + 55, + 540, + 296, + 578 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 70, + 584, + 295, + 599 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 584, + 295, + 599 + ], + "spans": [ + { + "bbox": [ + 70, + 584, + 295, + 599 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {i} ^ {p} = \\mathbf {x} _ {i - 1, i} ^ {p} + \\left(1 - \\mathbf {m} _ {i - 1, i} ^ {p}\\right) \\odot \\mathbf {m} _ {i + 1, i} ^ {p} \\odot \\mathbf {x} _ {i + 1, i} ^ {p}, \\tag {6}", + "image_path": "eebbba6d6d7e12bed9cecf221a708aea55a312d667643a1cd36ea9cc9c8fe32c.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 605, + 295, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 630 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 630 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 605, + 295, + 630 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i^p" + }, + { + "bbox": [ + 55, + 605, + 295, + 630 + ], + "type": "text", + "content": " denotes the bidirectional propagated pixels as the priors to control missing content generation in PCVDM." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 631, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 631, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 631, + 296, + 714 + ], + "type": "text", + "content": "For further analysis, we present statistics on warping error and the count of propagated pixels for flow-based propagation and HPP. As shown in Fig. 3, the flow-based method propagates fewer pixels with lower warping error due to the consistency check, while HPP increases the pixel count at the cost of slightly higher warping error. Despite this added error, PCVDM benefits from richer spatiotemporal context" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 330, + 555, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 330, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 330, + 555, + 354 + ], + "type": "text", + "content": "and complete semantic information while filtering out artifacts in such priors during iterative denoising." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 362, + 524, + 374 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 362, + 524, + 374 + ], + "spans": [ + { + "bbox": [ + 313, + 362, + 524, + 374 + ], + "type": "text", + "content": "3.3. Prior Controlled Video Diffusion Model" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 380, + 555, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 555, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 555, + 441 + ], + "type": "text", + "content": "Based on the priors from HPP, PCVDM aims to generate spatiotemporal coherent and realistic content in the corrupted regions. We propose a video diffusion model based on Stable Diffusion inpainting [35] and design a content-adaptive control mechanism to inject the priors into DM." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 448, + 533, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 448, + 533, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 448, + 533, + 460 + ], + "type": "text", + "content": "3.3.1. Inflating Image Inpainter to Video Inpainter" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 463, + 555, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 463, + 555, + 668 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 555, + 668 + ], + "type": "text", + "content": "The proposed video inpainting model is extended from the image inpainting model in Stable Diffusion. Following previous work [10, 40], we extend the 2D DM into a 3D structure. In VAE decoder [18] and U-Net [36], original 2D layers are converted to pseudo-3D layers to learn temporal correlations, with detailed structures available in the supplementary. Besides, we employ an OpenCLIP image encoder and perceive attention layers to extract GoP-wise CLIP latents from the corrupted video [5, 14], and then inject the latents into the 3D U-Net via cross-attention. GoP refers to a Group of Picture, i.e., a video clip processed per inference. During training, the binary mask " + }, + { + "bbox": [ + 313, + 463, + 555, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{m}" + }, + { + "bbox": [ + 313, + 463, + 555, + 668 + ], + "type": "text", + "content": ", the noised video latent " + }, + { + "bbox": [ + 313, + 463, + 555, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^t" + }, + { + "bbox": [ + 313, + 463, + 555, + 668 + ], + "type": "text", + "content": ", and the corrupted video latent " + }, + { + "bbox": [ + 313, + 463, + 555, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{z} = \\mathcal{E}(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 463, + 555, + 668 + ], + "type": "text", + "content": " are concatenated in the latent space and fed into the 3D U-Net. During inference, " + }, + { + "bbox": [ + 313, + 463, + 555, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^t" + }, + { + "bbox": [ + 313, + 463, + 555, + 668 + ], + "type": "text", + "content": " is initialized as Gaussian noise. The 3D U-Net generates the final sampled latent " + }, + { + "bbox": [ + 313, + 463, + 555, + 668 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}" + }, + { + "bbox": [ + 313, + 463, + 555, + 668 + ], + "type": "text", + "content": " that is decoded by the 3D VAE decoder as the completed video " + }, + { + "bbox": [ + 313, + 463, + 555, + 668 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}" + }, + { + "bbox": [ + 313, + 463, + 555, + 668 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 673, + 504, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 673, + 504, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 673, + 504, + 685 + ], + "type": "text", + "content": "3.3.2. Content-adaptive Control Mechanism" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "The priors are essential for generating missing content in corrupted videos. Recent work fills the propagated pix" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22956" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 70, + 296, + 315 + ], + "blocks": [ + { + "bbox": [ + 65, + 70, + 296, + 315 + ], + "lines": [ + { + "bbox": [ + 65, + 70, + 296, + 315 + ], + "spans": [ + { + "bbox": [ + 65, + 70, + 296, + 315 + ], + "type": "image", + "image_path": "c6a798cf1d5a1c00938b66e0a67fb844ddc816eb4f2cd4a65bbd27f21c92b99f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 319, + 296, + 365 + ], + "lines": [ + { + "bbox": [ + 55, + 319, + 296, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 296, + 365 + ], + "type": "text", + "content": "Figure 4. Illustration of content-adaptive control mechanism that injects prior information into PCVDM. Prior scalers in the dense injection are learned from the spatiotemporal context of priors and the sampling steps, enabling fine-grained controls." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 374, + 295, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 374, + 295, + 589 + ], + "spans": [ + { + "bbox": [ + 54, + 374, + 295, + 589 + ], + "type": "text", + "content": "els into the corrupted regions as partial inpainting results and uses transformers to generate remaining missing content [22, 59, 63]. However, due to inevitable artifacts introduced by warping, pixel-level error accumulates during token prediction in transformers guided by noisy tokens from the priors, as discussed in Sec. 1. A trade-off solution is to discard inaccurate propagated pixels based on the consistency check [56] while limiting sufficient priors and semantic integrity. Unlike transformers' sensitivity to low-level features, DMs prioritize semantic information within input conditions over pixel-level details. Hence, to ensure the semantic integrity of the propagated content, we retained all propagated pixels without any consistency check. Besides, we consider the priors as conditions instead of partial inpainting results pasted with original visible pixels, which also helps the network unambiguously perceive artifacts from the priors. During iterative sampling, the denoising process can smooth out such undesirable artifacts." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 592, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 592, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 592, + 296, + 714 + ], + "type": "text", + "content": "To effectively utilize the priors, we propose a content-adaptive control mechanism to scale and inject them into PCVDM for controlled missing content generation. The priors " + }, + { + "bbox": [ + 55, + 592, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^p = \\{\\mathbf{x}_i^p\\mid \\mathbf{x}_i^p\\in \\mathbb{R}^{3\\times H\\times W},1\\leq i\\leq N\\}" + }, + { + "bbox": [ + 55, + 592, + 296, + 714 + ], + "type": "text", + "content": " are first encoded into latents " + }, + { + "bbox": [ + 55, + 592, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^p = \\{\\mathbf{z}_i^p\\mid \\mathbf{z}_i^p\\in \\mathbb{R}^{4\\times \\frac{H}{8}\\times \\frac{W}{8}},1\\leq i\\leq N\\}" + }, + { + "bbox": [ + 55, + 592, + 296, + 714 + ], + "type": "text", + "content": ". We first design a LiteControlNet to extract multi-scale features from " + }, + { + "bbox": [ + 55, + 592, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^p" + }, + { + "bbox": [ + 55, + 592, + 296, + 714 + ], + "type": "text", + "content": " for further injection. Modified from ControlNet [61], LiteControlNet features a lightweight structure with one convolutional layer at each scale. Detailed structures of LiteControlNet refer to the supplementary. The" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": "prior feature at each scale is connected to the 3D U-Net encoder through a zero convolutional layer and a learned projector, i.e., a 3D convolutional layer with a kernel size of " + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "inline_equation", + "content": "1 \\times 1 \\times 1" + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": ". Rather than using additive operation, we introduce a spatiotemporal adaptive de-normalization strategy to inject each feature into the corresponding normalization layer [28]. As shown in Fig. 4, after the " + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": "-th latent normalized by 3D group normalization [52] in the 3D U-Net, we compute the output activation " + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_j" + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": " by de-normalizing the normalized latent " + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{h}}_j" + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": " according to the prior feature " + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_j" + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": ", formulated as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 379, + 196, + 555, + 212 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 379, + 196, + 555, + 212 + ], + "spans": [ + { + "bbox": [ + 379, + 196, + 555, + 212 + ], + "type": "interline_equation", + "content": "\\mathbf {h} _ {j} = \\gamma_ {j} \\odot \\bar {\\mathbf {h}} _ {j} + \\beta_ {j} + \\bar {\\mathbf {h}} _ {j}, \\tag {7}", + "image_path": "a1e53f97ee8fb8a4a910b45c30e6fd55c274bcba54804373bf8115877ea28dba.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 213, + 555, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 555, + 286 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 555, + 286 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 213, + 555, + 286 + ], + "type": "inline_equation", + "content": "\\gamma_{j}" + }, + { + "bbox": [ + 313, + 213, + 555, + 286 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 213, + 555, + 286 + ], + "type": "inline_equation", + "content": "\\beta_{j}" + }, + { + "bbox": [ + 313, + 213, + 555, + 286 + ], + "type": "text", + "content": " are obtained through 3D convolutions of " + }, + { + "bbox": [ + 313, + 213, + 555, + 286 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_j" + }, + { + "bbox": [ + 313, + 213, + 555, + 286 + ], + "type": "text", + "content": ". Equipped with this design, " + }, + { + "bbox": [ + 313, + 213, + 555, + 286 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_j" + }, + { + "bbox": [ + 313, + 213, + 555, + 286 + ], + "type": "text", + "content": " is scaled and injected into the intermediate latent, where the generation of missing content adapts to the spatiotemporal context of the priors. Additionally, " + }, + { + "bbox": [ + 313, + 213, + 555, + 286 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_j" + }, + { + "bbox": [ + 313, + 213, + 555, + 286 + ], + "type": "text", + "content": " is injected after attention blocks and yielded as the " + }, + { + "bbox": [ + 313, + 213, + 555, + 286 + ], + "type": "inline_equation", + "content": "(j + 1)" + }, + { + "bbox": [ + 313, + 213, + 555, + 286 + ], + "type": "text", + "content": "-th latent " + }, + { + "bbox": [ + 313, + 213, + 555, + 286 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_{j + 1}" + }, + { + "bbox": [ + 313, + 213, + 555, + 286 + ], + "type": "text", + "content": ", formulated as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 390, + 290, + 555, + 305 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 290, + 555, + 305 + ], + "spans": [ + { + "bbox": [ + 390, + 290, + 555, + 305 + ], + "type": "interline_equation", + "content": "\\mathbf {h} _ {j + 1} = \\zeta (t) \\mathbf {f} _ {j} + \\mathbf {h} _ {j}, \\tag {8}", + "image_path": "24b63e349c9086f451402a36e566ac9d5457d96cc9f55c79f368192e11923c81.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 308, + 554, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 554, + 355 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 554, + 355 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 308, + 554, + 355 + ], + "type": "inline_equation", + "content": "\\zeta(\\cdot)" + }, + { + "bbox": [ + 313, + 308, + 554, + 355 + ], + "type": "text", + "content": " represents two linear layers interleaved with a SiLU activation [33] in between. " + }, + { + "bbox": [ + 313, + 308, + 554, + 355 + ], + "type": "inline_equation", + "content": "\\zeta(\\cdot)" + }, + { + "bbox": [ + 313, + 308, + 554, + 355 + ], + "type": "text", + "content": " encodes the sampling step " + }, + { + "bbox": [ + 313, + 308, + 554, + 355 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 308, + 554, + 355 + ], + "type": "text", + "content": " [35] into Fourier features, which allows the scaling strength to vary with " + }, + { + "bbox": [ + 313, + 308, + 554, + 355 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 308, + 554, + 355 + ], + "type": "text", + "content": " for fine-grained controls." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 356, + 554, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 356, + 554, + 428 + ], + "spans": [ + { + "bbox": [ + 313, + 356, + 554, + 428 + ], + "type": "text", + "content": "During iterative denoising, the content in the corrupted regions progressively generates, approaching seamless integration with original visible pixels. After decoding the final sampled latent by the 3D VAE decoder, the generated and corrupted videos are blended with a blurred mask to improve coherence along the mask boundary [16]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 433, + 464, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 433, + 464, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 433, + 464, + 445 + ], + "type": "text", + "content": "3.4. Inpainting for Long Videos" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 450, + 554, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 554, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 554, + 533 + ], + "type": "text", + "content": "Video inpainting faces the challenge of maintaining temporal coherence across GoPs in long videos that span several seconds. Generative models are often restricted to processing all frames at once due to memory limitations, which leads to variability when sequentially inferring all GoPs. To address this, we introduce temporal references from adjacent GoPs to assist in inpainting the current GoP." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": " denote the length of the GoP that HomoGen processes per inference. For a video containing " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": " frames, we segment the video into " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\left\\lceil \\frac{L}{N} \\right\\rceil" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": " GoPs, each with an overlap of " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "N_{\\text{overlap}}" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": " frames, and process all GoPs sequentially. For the " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": "-th GoP, the first and last frames are set to the center frames of the " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "(i-1)" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": "-th and " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "(i+1)" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": "-th GoP, respectively, while " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "N_{\\text{overlap}} - 2" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": " frames from the second frame onward are set to the last frames of the " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "(i-1)" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": "-th GoP. In the initial inference, the first GoP undergoes inpainting. Before " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": "-th inference " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "(i > 1)" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": ", the beginning " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "N_{\\text{overlap}} - 1" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": " frames of the " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": "-th GoP are replaced by the corresponding frames from the inpainted " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "(i-1)" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": "-th GoP. During " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": "-th inference, HomoGen can capture long-term temporal references for inpainting the " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": "-th GoP. This process iterates until all GoPs are inpainted, which ensures content consistency across the entire video." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "22957" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 95, + 555, + 255 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 1. Quantitative comparisons on YouTube-VOS, DAVIS, and RORD datasets. " + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": " indicates higher is better. " + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": " indicates lower is better. " + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "inline_equation", + "content": "E_{w}^{*}" + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": " denotes " + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "inline_equation", + "content": "E_{\\text{warp}} \\times 10^{-3}" + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": ". The best performances are highlighted in bold. All methods are evaluated following their default settings." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 95, + 555, + 255 + ], + "lines": [ + { + "bbox": [ + 58, + 95, + 555, + 255 + ], + "spans": [ + { + "bbox": [ + 58, + 95, + 555, + 255 + ], + "type": "table", + "html": "
YouTube-VOSDAVISRORD
ModelsPSNR ↑SSIM ↑VFID ↓Ew* ↓PSNR ↑SSIM ↑VFID ↓Ew* ↓PSNR ↑SSIM ↑VFID ↓Ew* ↓
DFVI [56]29.160.94290.0661.50928.810.94040.1871.60831.840.95710.0651.137
CPNet [20]31.580.96070.0711.42330.280.95210.1821.51935.070.97400.0390.879
FGVC [9]29.670.94030.0641.02230.800.94970.1651.58632.760.96530.0521.005
STTN [58]32.570.96800.0490.90730.740.95900.1531.66134.940.97180.0440.872
FuseFormer [24]33.290.96810.0530.89132.540.97020.1381.33635.610.98110.0310.843
FGT [59]32.170.95990.0541.02532.860.96500.1291.32334.210.97240.0380.867
ISVI [60]30.380.94710.0771.31132.200.96260.1661.29233.450.96910.0460.902
E2FGVI [22]33.710.97000.0460.86433.010.97210.1161.31536.040.98000.0270.797
ProPainter [63]34.460.97220.0420.87234.190.97590.0981.19536.760.98670.0280.813
HomoGen34.810.97350.0320.85834.710.97630.0771.23437.290.98720.0230.793
", + "image_path": "f57e74de431146e3f90729fde5a7af389e42e9bc5dafd1ff91abe0acf8a0425f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 56, + 256, + 175, + 345 + ], + "blocks": [ + { + "bbox": [ + 56, + 256, + 175, + 345 + ], + "lines": [ + { + "bbox": [ + 56, + 256, + 175, + 345 + ], + "spans": [ + { + "bbox": [ + 56, + 256, + 175, + 345 + ], + "type": "image", + "image_path": "4becd39a5eb8747985cf7fe7aa91375719b5d6e523456739e44e9cb36f027baf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 77, + 347, + 156, + 357 + ], + "lines": [ + { + "bbox": [ + 77, + 347, + 156, + 357 + ], + "spans": [ + { + "bbox": [ + 77, + 347, + 156, + 357 + ], + "type": "text", + "content": "(a) PSNR vs. mask ratio" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 87, + 361, + 263, + 373 + ], + "lines": [ + { + "bbox": [ + 87, + 361, + 263, + 373 + ], + "spans": [ + { + "bbox": [ + 87, + 361, + 263, + 373 + ], + "type": "text", + "content": "Figure 5. Comparisons on different mask ratios." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 175, + 255, + 294, + 345 + ], + "blocks": [ + { + "bbox": [ + 175, + 255, + 294, + 345 + ], + "lines": [ + { + "bbox": [ + 175, + 255, + 294, + 345 + ], + "spans": [ + { + "bbox": [ + 175, + 255, + 294, + 345 + ], + "type": "image", + "image_path": "4838083238cf1141b23851635ff3c653e02081c431a925397b5e6a791a967362.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 196, + 347, + 275, + 357 + ], + "lines": [ + { + "bbox": [ + 196, + 347, + 275, + 357 + ], + "spans": [ + { + "bbox": [ + 196, + 347, + 275, + 357 + ], + "type": "text", + "content": "(b) VFID vs. mask ratio" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 315, + 256, + 433, + 346 + ], + "blocks": [ + { + "bbox": [ + 315, + 256, + 433, + 346 + ], + "lines": [ + { + "bbox": [ + 315, + 256, + 433, + 346 + ], + "spans": [ + { + "bbox": [ + 315, + 256, + 433, + 346 + ], + "type": "image", + "image_path": "afad4e4cc0212255f4262b2392d4e8844a12645be75bcb86adf6cf90eabf279d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 347, + 427, + 357 + ], + "lines": [ + { + "bbox": [ + 323, + 347, + 427, + 357 + ], + "spans": [ + { + "bbox": [ + 323, + 347, + 427, + 357 + ], + "type": "text", + "content": "(a) PSNR vs. motion magnitude" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 334, + 361, + 534, + 373 + ], + "lines": [ + { + "bbox": [ + 334, + 361, + 534, + 373 + ], + "spans": [ + { + "bbox": [ + 334, + 361, + 534, + 373 + ], + "type": "text", + "content": "Figure 6. Comparisons on different motion magnitude." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 436, + 255, + 553, + 346 + ], + "blocks": [ + { + "bbox": [ + 436, + 255, + 553, + 346 + ], + "lines": [ + { + "bbox": [ + 436, + 255, + 553, + 346 + ], + "spans": [ + { + "bbox": [ + 436, + 255, + 553, + 346 + ], + "type": "image", + "image_path": "47e2665b9cf4430ceaaee5b1effe8058644f697c19637b80da99128a310f4bb9.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 443, + 347, + 545, + 357 + ], + "lines": [ + { + "bbox": [ + 443, + 347, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 443, + 347, + 545, + 357 + ], + "type": "text", + "content": "(b) VFID vs. motion magnitude" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 384, + 137, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 137, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 137, + 399 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 405, + 188, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 405, + 188, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 405, + 188, + 418 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 424, + 120, + 435 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 424, + 120, + 435 + ], + "spans": [ + { + "bbox": [ + 55, + 424, + 120, + 435 + ], + "type": "text", + "content": "4.1.1. Datasets" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 439, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 439, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 439, + 296, + 693 + ], + "type": "text", + "content": "We use an internal video dataset including 5 million video sequences from Shutterstock1 as the training set. To simulate masks in video completion and object removal tasks, we synthesize stationary and object-aware masks in a random fashion, which are then pasted onto frames [17, 22, 63]. For evaluation, we use two video object segmentation datasets: DAVIS [30] and YouTube-VOS [55], which consist of 508 and 90 video sequences in test sets, respectively. For the DAVIS test set, following ProPainter [63], we use 50 video sequences for metric calculations. For mask settings, stationary masks are used to calculate objective metrics for the video completion task, and object-aware masks extracted from segmentation annotations are adopted for qualitative comparisons in the object removal task. Additionally, we adopt the Real-world Object Removal Dataset (RORD) [38] to evaluate the object removal task. The RORD test set includes 343 video pairs with and without objects, along with object mask annotations. This dataset addresses the lack of references for computing objective metrics in terms of the object removal task in previous work. For all datasets, video frames are sized to " + }, + { + "bbox": [ + 55, + 439, + 296, + 693 + ], + "type": "inline_equation", + "content": "240 \\times 432" + }, + { + "bbox": [ + 55, + 439, + 296, + 693 + ], + "type": "text", + "content": " for evaluation. Text prompts" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 384, + 555, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 555, + 409 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 555, + 409 + ], + "type": "text", + "content": "for test sets are generated by GPT-4o [27] using the center frame of each corrupted video." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 416, + 375, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 416, + 375, + 427 + ], + "spans": [ + { + "bbox": [ + 313, + 416, + 375, + 427 + ], + "type": "text", + "content": "4.1.2. Metrics" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 431, + 555, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 555, + 503 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 555, + 503 + ], + "type": "text", + "content": "Following widely adopted protocols, we use PSNR, SSIM [50], VFID [48], and warping error " + }, + { + "bbox": [ + 313, + 431, + 555, + 503 + ], + "type": "inline_equation", + "content": "E_{\\text{warp}}" + }, + { + "bbox": [ + 313, + 431, + 555, + 503 + ], + "type": "text", + "content": " [19] as evaluation metrics. Specifically, PSNR and SSIM are frequently used for distortion-oriented video assessment. VFID measures the perceptual similarity between input videos and outputs. " + }, + { + "bbox": [ + 313, + 431, + 555, + 503 + ], + "type": "inline_equation", + "content": "E_{\\text{warp}}" + }, + { + "bbox": [ + 313, + 431, + 555, + 503 + ], + "type": "text", + "content": " assesses the temporal stability." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 503, + 554, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 503, + 554, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 503, + 554, + 529 + ], + "type": "text", + "content": "Detailed configurations of the proposed model for training and inference refer to the supplementary." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 536, + 400, + 549 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 536, + 400, + 549 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 400, + 549 + ], + "type": "text", + "content": "4.2. Comparisons" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 554, + 430, + 566 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 554, + 430, + 566 + ], + "spans": [ + { + "bbox": [ + 313, + 554, + 430, + 566 + ], + "type": "text", + "content": "4.2.1. Quantitative Results" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "type": "text", + "content": "We report quantitative results on three datasets, using stationary masks for the YouTube-VOS and DAVIS datasets, and object-aware masks for the RORD dataset. We compare the proposed HomoGen with previous video inpainting methods, including DFVI [56], CPNet [20], FGVC [9], STTN [58], FuseFormer [24], FGT [59], ISVI [60], " + }, + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{FGVI}" + }, + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "type": "text", + "content": " [22], and ProPainter [63]. As shown in Tab. 1, HomoGen outperforms previous methods across most quantitative metrics. Notably, HomoGen significantly surpasses the state-of-the-art method by " + }, + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "type": "inline_equation", + "content": "23.81\\%" + }, + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "type": "inline_equation", + "content": "21.43\\%" + }, + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "type": "inline_equation", + "content": "17.86\\%" + }, + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "type": "text", + "content": " in VFID on the YouTube-VOS, DAVIS, and RORD datasets, respectively, demonstrating its ability to generate" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 66, + 702, + 208, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 208, + 712 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 208, + 712 + ], + "type": "text", + "content": "1https://wwwCRETstorstock.com" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22958" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 553, + 308 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 553, + 308 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 553, + 308 + ], + "type": "image", + "image_path": "0ca96261336b8c6677971935e684d72b1734e2012eb516d3fe41a3a528d28db4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 129, + 311, + 481, + 323 + ], + "lines": [ + { + "bbox": [ + 129, + 311, + 481, + 323 + ], + "spans": [ + { + "bbox": [ + 129, + 311, + 481, + 323 + ], + "type": "text", + "content": "Figure 7. Qualitative results compared with FuseFormer [24], " + }, + { + "bbox": [ + 129, + 311, + 481, + 323 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{FGVI}" + }, + { + "bbox": [ + 129, + 311, + 481, + 323 + ], + "type": "text", + "content": " [22], and ProPainter [63]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 331, + 293, + 417 + ], + "blocks": [ + { + "bbox": [ + 59, + 331, + 293, + 417 + ], + "lines": [ + { + "bbox": [ + 59, + 331, + 293, + 417 + ], + "spans": [ + { + "bbox": [ + 59, + 331, + 293, + 417 + ], + "type": "image", + "image_path": "89c92b7ef84712dd74a1c2352c361e41be8bc4f8ebd6ff61b3edeb4fcc67ec95.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 420, + 290, + 432 + ], + "lines": [ + { + "bbox": [ + 59, + 420, + 290, + 432 + ], + "spans": [ + { + "bbox": [ + 59, + 420, + 290, + 432 + ], + "type": "text", + "content": "Figure 8. Comparative user study on HomoGen and ProPainter." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 441, + 286, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 286, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 286, + 453 + ], + "type": "text", + "content": "videos with more visually plausible and realistic content." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 453, + 296, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 453, + 296, + 620 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 296, + 620 + ], + "type": "text", + "content": "To assess the model performance across different scenarios, we evaluate HomoGen on videos with varying mask ratios, as well as camera and object motion. We select ProPainter for comparisons, given its best performance among the baselines. In evaluation for different mask ratios, we partition the mask ratio range from " + }, + { + "bbox": [ + 55, + 453, + 296, + 620 + ], + "type": "inline_equation", + "content": "[0, 0.5]" + }, + { + "bbox": [ + 55, + 453, + 296, + 620 + ], + "type": "text", + "content": " into five equal intervals, each corresponding to a synthesized mask set. We randomly sample 30 video sequences for each mask set. As shown in Fig. 5, HomoGen performs on par with ProPainter when the mask ratio is low. As the mask ratio increases, ProPainter's performance degrades more substantially than HomoGen's. In evaluation for different camera and object motion, we define the average pixel displacement " + }, + { + "bbox": [ + 55, + 453, + 296, + 620 + ], + "type": "inline_equation", + "content": "D^{p}" + }, + { + "bbox": [ + 55, + 453, + 296, + 620 + ], + "type": "text", + "content": " as a measure of motion, formulated as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 625, + 296, + 659 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 625, + 296, + 659 + ], + "spans": [ + { + "bbox": [ + 62, + 625, + 296, + 659 + ], + "type": "interline_equation", + "content": "D ^ {p} = \\frac {1}{N H W} \\sum_ {n = 1} ^ {N} \\sum_ {i = 1} ^ {H} \\sum_ {j = 1} ^ {W} \\sqrt {\\left(d x _ {i , j , n}\\right) ^ {2} + \\left(d y _ {i , j , n}\\right) ^ {2}}, \\tag {9}", + "image_path": "4edf69b958023f0f519f9bf2a7dda9882e2e07ae86991386af3588054df01dce.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "dx" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "dy" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " denote the pixel displacement in the vertical and horizontal directions across frames based on the optical flows estimated by RAFT [44]. Similar to mask ratios, we select five video sets, each containing 30 video" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 317, + 396, + 547, + 495 + ], + "blocks": [ + { + "bbox": [ + 313, + 329, + 555, + 395 + ], + "lines": [ + { + "bbox": [ + 313, + 329, + 555, + 395 + ], + "spans": [ + { + "bbox": [ + 313, + 329, + 555, + 395 + ], + "type": "text", + "content": "Table 2. Ablation studies on various components in the two-stage pipeline. \"HP\" denotes the homography-based priors. \"FP\" denotes the flow-based priors. \"FCC\" denotes the flow consistency check. \"VT\" denotes the video transformer in ProPainter. \"VDM\" denotes PCVDM in the proposed model. (e) corresponds to ProPainter, and (f) corresponds to HomoGen." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 396, + 547, + 495 + ], + "lines": [ + { + "bbox": [ + 317, + 396, + 547, + 495 + ], + "spans": [ + { + "bbox": [ + 317, + 396, + 547, + 495 + ], + "type": "table", + "html": "
ComponentsPSNRSSIMVFID
FPFCCHPVTVDM
(a)XXXX✓33.160.96150.042
(b)✓✓XX✓34.050.96840.036
(c)✓XXX✓34.230.96990.035
(d)XX✓✓X33.880.96530.044
(e)✓✓X✓X34.460.97220.042
(f)XX✓X✓34.810.97350.032
", + "image_path": "04c693f55e15b532bddafca674f278d6983b25bbe4cd8cf851cb7a6c857a01fa.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 507, + 555, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 507, + 555, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 507, + 555, + 592 + ], + "type": "text", + "content": "sequences, for various values of " + }, + { + "bbox": [ + 313, + 507, + 555, + 592 + ], + "type": "inline_equation", + "content": "D^p" + }, + { + "bbox": [ + 313, + 507, + 555, + 592 + ], + "type": "text", + "content": ". As shown in Fig. 6, HomoGen outperforms ProPainter in most motion conditions, particularly those with minimal movements. The two quantitative results demonstrate that HomoGen has superior inpainting capability across various video scenarios, especially in more hard cases involving larger masks or less motion where the spatiotemporal context is limited." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 601, + 424, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 601, + 424, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 601, + 424, + 613 + ], + "type": "text", + "content": "4.2.2. Qualitative Results" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": "Fig. 7 shows qualitative results of HomoGen compared to the state-of-the-art methods, including FuseFormer, " + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{FGVI}" + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": ", and ProPainter on both video completion and object removal tasks. Generally, HomoGen can generate more faithful textural and structural information, as well as preserve the semantic integrity of objects. Nevertheless, other compared methods tend to fail or produce artifacts such as blur, distortion, black hazy regions, etc." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22959" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 82, + 288, + 144 + ], + "blocks": [ + { + "bbox": [ + 61, + 71, + 288, + 81 + ], + "lines": [ + { + "bbox": [ + 61, + 71, + 288, + 81 + ], + "spans": [ + { + "bbox": [ + 61, + 71, + 288, + 81 + ], + "type": "text", + "content": "Table 3. Ablation studies on various components in HomoGen." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 82, + 288, + 144 + ], + "lines": [ + { + "bbox": [ + 59, + 82, + 288, + 144 + ], + "spans": [ + { + "bbox": [ + 59, + 82, + 288, + 144 + ], + "type": "table", + "html": "
ModelPSNRSSIMVFID
w/o text prompts34.070.97160.039
w/o GoP-wise CLIP embeddings34.620.97220.033
HomoGen34.810.97350.032
", + "image_path": "efda1db770dacccf3d0d7515438f60adc0442bf7a5b8265a92117f67d5501e16.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 57, + 148, + 106, + 198 + ], + "blocks": [ + { + "bbox": [ + 57, + 148, + 106, + 198 + ], + "lines": [ + { + "bbox": [ + 57, + 148, + 106, + 198 + ], + "spans": [ + { + "bbox": [ + 57, + 148, + 106, + 198 + ], + "type": "image", + "image_path": "09c9d9c6a23ab68fa30556273e19c2a1f7509f657dc9934d9e9df307366cb785.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 65, + 200, + 99, + 209 + ], + "lines": [ + { + "bbox": [ + 65, + 200, + 99, + 209 + ], + "spans": [ + { + "bbox": [ + 65, + 200, + 99, + 209 + ], + "type": "text", + "content": "(i) line (a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 107, + 148, + 156, + 198 + ], + "blocks": [ + { + "bbox": [ + 107, + 148, + 156, + 198 + ], + "lines": [ + { + "bbox": [ + 107, + 148, + 156, + 198 + ], + "spans": [ + { + "bbox": [ + 107, + 148, + 156, + 198 + ], + "type": "image", + "image_path": "580a118ac8bfe3c6689093529a29496086e55c721a878ae05f4ee5fa37b04365.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 113, + 200, + 149, + 209 + ], + "lines": [ + { + "bbox": [ + 113, + 200, + 149, + 209 + ], + "spans": [ + { + "bbox": [ + 113, + 200, + 149, + 209 + ], + "type": "text", + "content": "(ii) line (b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 156, + 148, + 205, + 198 + ], + "blocks": [ + { + "bbox": [ + 156, + 148, + 205, + 198 + ], + "lines": [ + { + "bbox": [ + 156, + 148, + 205, + 198 + ], + "spans": [ + { + "bbox": [ + 156, + 148, + 205, + 198 + ], + "type": "image", + "image_path": "b96559d4740395cf54a9f40e6fd00294f67104e0a63d329e0164d41b8de22831.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 162, + 200, + 200, + 209 + ], + "lines": [ + { + "bbox": [ + 162, + 200, + 200, + 209 + ], + "spans": [ + { + "bbox": [ + 162, + 200, + 200, + 209 + ], + "type": "text", + "content": "(iii) line (c)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 206, + 148, + 294, + 198 + ], + "blocks": [ + { + "bbox": [ + 206, + 148, + 294, + 198 + ], + "lines": [ + { + "bbox": [ + 206, + 148, + 294, + 198 + ], + "spans": [ + { + "bbox": [ + 206, + 148, + 294, + 198 + ], + "type": "image", + "image_path": "a25aa309304e1ec060923652d6a242b8da1a274ddd05109ce0ecf7caea6fc103.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 225, + 200, + 276, + 209 + ], + "lines": [ + { + "bbox": [ + 225, + 200, + 276, + 209 + ], + "spans": [ + { + "bbox": [ + 225, + 200, + 276, + 209 + ], + "type": "text", + "content": "(iv) Input video" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 57, + 210, + 106, + 260 + ], + "blocks": [ + { + "bbox": [ + 57, + 210, + 106, + 260 + ], + "lines": [ + { + "bbox": [ + 57, + 210, + 106, + 260 + ], + "spans": [ + { + "bbox": [ + 57, + 210, + 106, + 260 + ], + "type": "image", + "image_path": "c56238cca498c9f453af8fe09619e462da2cc9c323b87be5b60eb48f3304ca80.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 262, + 100, + 271 + ], + "lines": [ + { + "bbox": [ + 64, + 262, + 100, + 271 + ], + "spans": [ + { + "bbox": [ + 64, + 262, + 100, + 271 + ], + "type": "text", + "content": "(v) line (d)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 107, + 210, + 156, + 261 + ], + "blocks": [ + { + "bbox": [ + 107, + 210, + 156, + 261 + ], + "lines": [ + { + "bbox": [ + 107, + 210, + 156, + 261 + ], + "spans": [ + { + "bbox": [ + 107, + 210, + 156, + 261 + ], + "type": "image", + "image_path": "e183d3c1701e32f22d58281478113fcd92cf18b9c3ba6bf9db900a8eb82bb49c.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 113, + 262, + 150, + 271 + ], + "lines": [ + { + "bbox": [ + 113, + 262, + 150, + 271 + ], + "spans": [ + { + "bbox": [ + 113, + 262, + 150, + 271 + ], + "type": "text", + "content": "(vi) line (e)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 156, + 210, + 206, + 261 + ], + "blocks": [ + { + "bbox": [ + 156, + 210, + 206, + 261 + ], + "lines": [ + { + "bbox": [ + 156, + 210, + 206, + 261 + ], + "spans": [ + { + "bbox": [ + 156, + 210, + 206, + 261 + ], + "type": "image", + "image_path": "f549e3df9d6139a34a152d34a2d54aad7a8ded8eb40543056ec7a571c033d363.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 162, + 262, + 200, + 271 + ], + "lines": [ + { + "bbox": [ + 162, + 262, + 200, + 271 + ], + "spans": [ + { + "bbox": [ + 162, + 262, + 200, + 271 + ], + "type": "text", + "content": "(vii) line (f)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 206, + 210, + 294, + 261 + ], + "blocks": [ + { + "bbox": [ + 206, + 210, + 294, + 261 + ], + "lines": [ + { + "bbox": [ + 206, + 210, + 294, + 261 + ], + "spans": [ + { + "bbox": [ + 206, + 210, + 294, + 261 + ], + "type": "image", + "image_path": "db651aafdff4061ad50cc897629d59a120a74dce80b1c4ebf9201a500f732693.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 223, + 262, + 277, + 271 + ], + "lines": [ + { + "bbox": [ + 223, + 262, + 277, + 271 + ], + "spans": [ + { + "bbox": [ + 223, + 262, + 277, + 271 + ], + "type": "text", + "content": "(viii) Input mask" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 277, + 294, + 288 + ], + "lines": [ + { + "bbox": [ + 55, + 277, + 294, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 294, + 288 + ], + "type": "text", + "content": "Figure 9. Qualitative comparisons of the ablation studies in Tab. 2." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 298, + 296, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 298, + 296, + 406 + ], + "spans": [ + { + "bbox": [ + 55, + 298, + 296, + 406 + ], + "type": "text", + "content": "Furthermore, we conduct a user study with 15 participants who are shown randomly sampled 45 video triplets. Each triplet contains one original video, one generated by HomoGen, and one generated by ProPainter. Volunteers are asked to select a visually better inpainting video or indicate a tie for each triplet. As shown in Fig. 8, volunteers prefer our results in terms of spatial coherence, temporal consistency, visual quality, and overall impression, demonstrating that HomoGen can generate more visually pleasant results." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 413, + 124, + 424 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 413, + 124, + 424 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 124, + 424 + ], + "type": "text", + "content": "4.3. Ablations" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 430, + 295, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 430, + 295, + 454 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 295, + 454 + ], + "type": "text", + "content": "We perform ablation studies on different technologies employed in HomoGen on the YouTube-VOS dataset." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 55, + 460, + 202, + 471 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 202, + 471 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 202, + 471 + ], + "type": "text", + "content": "4.3.1. Study of Two-stage Pipeline" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 55, + 474, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 296, + 714 + ], + "type": "text", + "content": "We investigate the importance of HPP tailored for PCVDM. As discussed in Sec. 1, various technologies have been designed for the two-stage pipeline, i.e., pixel propagation and content generation, among recent video inpainting methods. We compare different combinations of technologies within the two stages. For the first stage, we select HPP and the typical flow-based propagation proposed in ProPainter for comparisons. For the second stage, we select PCVDM and the video transformer proposed in ProPainter. As shown in line (a), Tab. 2, by removing HPP, temporal context is no longer provided to control content generation, resulting in a large performance decrease. The corresponding qualitative result in Fig. 9 shows artifacts and unrealistic content (the horse's leg). As demonstrated in line (b) and line (c), integrating flow-based priors with PCVDM, but without the consistency check, further enhances performance, which highlights the critical role of semantic integrity for DMs. Results in line (d) and line (e) indicate that the video transformer is not robust to artifacts in the priors and requires cleaner priors from pixel propagation. We also in" + } + ] + } + ], + "index": 23 + }, + { + "type": "table", + "bbox": [ + 318, + 83, + 547, + 165 + ], + "blocks": [ + { + "bbox": [ + 317, + 71, + 551, + 82 + ], + "lines": [ + { + "bbox": [ + 317, + 71, + 551, + 82 + ], + "spans": [ + { + "bbox": [ + 317, + 71, + 551, + 82 + ], + "type": "text", + "content": "Table 4. Ablation studies on various prior injection mechanisms." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 83, + 547, + 165 + ], + "lines": [ + { + "bbox": [ + 318, + 83, + 547, + 165 + ], + "spans": [ + { + "bbox": [ + 318, + 83, + 547, + 165 + ], + "type": "table", + "html": "
ModelPSNRSSIMVFID
Pasting31.740.96100.053
Concatenation32.860.96430.043
Cross-attention34.140.97180.035
Noise prior33.520.96570.036
Content-adaptive controls34.810.97350.032
", + "image_path": "888e60f8a6543e621461b835a2f0de384b77335855e3763e6767e438264d27b9.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "table_body" + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 173, + 553, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 173, + 553, + 221 + ], + "spans": [ + { + "bbox": [ + 313, + 173, + 553, + 221 + ], + "type": "text", + "content": "vestigate flow-based or homography-based priors combined with PCVDM. Results in line (c) and line (f) demonstrate that PCVDM exhibits greater tolerance to artifacts in these priors when coupled with HPP that provides more priors." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 226, + 550, + 238 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 226, + 550, + 238 + ], + "spans": [ + { + "bbox": [ + 313, + 226, + 550, + 238 + ], + "type": "text", + "content": "4.3.2. Study of Prior Controlled Video Diffusion Model" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 241, + 555, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 241, + 555, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 241, + 555, + 529 + ], + "type": "text", + "content": "We investigate the effectiveness of various components employed in PCVDM. We first remove the text prompts by replacing the visual descriptions with blank strings, for which the PSNR value is decreased by a large margin, i.e., 0.74 dB, as shown in Tab. 3. In contrast to the model without GoP-wise CLIP embeddings, the performance exhibits a non-trivial improvement. Furthermore, we compare the proposed content-adaptive control mechanism with other homography-based prior injection strategies, including pasting, concatenation, and cross-attention. Pasting represents directly pasting the priors into the corrupted regions, and others have been widely adopted in previous work [3, 25, 35, 49, 61]. Detailed implementations of different injection strategies refer to the supplementary. As shown in Tab. 4, pasting the priors into the corrupted regions raises a significant performance drop, which can be attributed to the fact that placing propagated and visible pixels on the same plane, making it difficult for DMs to differentiate between them. Compared to the second-best strategy employing cross-attention, using the content-adaptive controls improves performance by a non-trivial margin, i.e., 0.67 dB in terms of PSNR. This suggests that the content-adaptive control mechanism learns fine-grained dense injection and further enhances the control ability of PCVDM." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 538, + 388, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 538, + 388, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 538, + 388, + 550 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 558, + 554, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 554, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 554, + 641 + ], + "type": "text", + "content": "Our work introduces a propagate-and-diffuse framework for video inpainting, named HomoGen, which utilizes homography registration to propagate contextual pixels for generating missing content in a video diffusion model. Experiments not only show HomoGen's superiority in generating realistic content but prove the synergy between homography registration and video diffusion models for inpainting." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 642, + 554, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 712 + ], + "type": "text", + "content": "Limitations. There are several potential limitations in this study. For instance, generating identity-preserved content in videos with scene alternations and efficiently processing high-resolution videos remain challenges for both HomoGen and previous methods. These issues warrant further investigation and improvement in future work." + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22960" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "type": "text", + "content": "[1] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In ICCV, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 296, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 296, + 158 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 296, + 158 + ], + "type": "text", + "content": "[2] Marcelo Bertalmio, Andrea L Bertozzi, and Guillermo Sapiro. Navier-stokes, fluid dynamics, and image and video inpainting. In CVPR, 2001. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 159, + 295, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 159, + 295, + 203 + ], + "spans": [ + { + "bbox": [ + 62, + 159, + 295, + 203 + ], + "type": "text", + "content": "[3] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In CVPR, 2023. 2, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 205, + 295, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 205, + 295, + 237 + ], + "spans": [ + { + "bbox": [ + 62, + 205, + 295, + 237 + ], + "type": "text", + "content": "[4] Ya-Liang Chang, Zhe Yu Liu, Kuan-Ying Lee, and Winston Hsu. Free-form video inpainting with 3d gated convolution and temporal patchgan. In ICCV, 2019. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 239, + 295, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 239, + 295, + 292 + ], + "spans": [ + { + "bbox": [ + 62, + 239, + 295, + 292 + ], + "type": "text", + "content": "[5] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In CVPR, 2023. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 294, + 295, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 294, + 295, + 316 + ], + "spans": [ + { + "bbox": [ + 62, + 294, + 295, + 316 + ], + "type": "text", + "content": "[6] Prafulla Dhariwal and Alexander Quinn Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 318, + 295, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 318, + 295, + 350 + ], + "spans": [ + { + "bbox": [ + 62, + 318, + 295, + 350 + ], + "type": "text", + "content": "[7] Mounira Ebdelli, Olivier Le Meur, and Christine Guillemot. Video inpainting with short-term windows: application to object removal and error concealment. IEEE TIP, 2015. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 352, + 295, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 352, + 295, + 395 + ], + "spans": [ + { + "bbox": [ + 62, + 352, + 295, + 395 + ], + "type": "text", + "content": "[8] Ruoyu Feng, Wenming Weng, Yanhui Wang, Yuhui Yuan, Jianmin Bao, Chong Luo, Zhibo Chen, and Baining Guo. Credit: Creative and controllable video editing via diffusion models. In CVPR, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 396, + 295, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 396, + 295, + 418 + ], + "spans": [ + { + "bbox": [ + 62, + 396, + 295, + 418 + ], + "type": "text", + "content": "[9] Chen Gao, Ayush Saraf, Jia-Bin Huang, and Johannes Kopf. Flow-edge guided video completion. In ECCV, 2020. 2, 3, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 420, + 295, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 420, + 295, + 473 + ], + "spans": [ + { + "bbox": [ + 57, + 420, + 295, + 473 + ], + "type": "text", + "content": "[10] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. In ICLR, 2024. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 476, + 295, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 476, + 295, + 508 + ], + "spans": [ + { + "bbox": [ + 57, + 476, + 295, + 508 + ], + "type": "text", + "content": "[11] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross-attention control. In ICLR, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 510, + 295, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 510, + 295, + 532 + ], + "spans": [ + { + "bbox": [ + 57, + 510, + 295, + 532 + ], + "type": "text", + "content": "[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 533, + 295, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 533, + 295, + 566 + ], + "spans": [ + { + "bbox": [ + 57, + 533, + 295, + 566 + ], + "type": "text", + "content": "[13] Yuan-Ting Hu, Heng Wang, Nicolas Ballas, Kristen Grauman, and Alexander G Schwing. Proposal-based video completion. In ECCV, 2020. 1, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 567, + 295, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 567, + 295, + 600 + ], + "spans": [ + { + "bbox": [ + 57, + 567, + 295, + 600 + ], + "type": "text", + "content": "[14] Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and Joao Carreira. Perceiver: General perception with iterative attention. In ICML, 2021. 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 601, + 295, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 601, + 295, + 644 + ], + "spans": [ + { + "bbox": [ + 57, + 601, + 295, + 644 + ], + "type": "text", + "content": "[15] Wonjoon Jin, Qi Dai, Chong Luo, Seung-Hwan Baek, and Sunghyun Cho. Flovd: Optical flow meets video diffusion model for enhanced camera-controlled video synthesis. In CVPR, 2025. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "type": "text", + "content": "[16] Xuan Ju, Xian Liu, Xintao Wang, Yuxuan Bian, Ying Shan, and Qiang Xu. Brushnet: A plug-and-play image inpainting model with decomposed dual-branch diffusion. In ECCV, 2024.5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "text", + "content": "[17] Dahun Kim, Sanghyun Woo, Joon-Young Lee, and In So Kweon. Deep video inpainting. In CVPR, 2019. 3, 6" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "type": "text", + "content": "[18] Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. In ICLR, 2014. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 97, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 97, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 317, + 97, + 553, + 129 + ], + "type": "text", + "content": "[19] Wei-Sheng Lai, Jia-Bin Huang, Oliver Wang, Eli Shechtman, Ersin Yumer, and Ming-Hsuan Yang. Learning blind video temporal consistency. In ECCV, 2018. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 132, + 553, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 132, + 553, + 163 + ], + "spans": [ + { + "bbox": [ + 317, + 132, + 553, + 163 + ], + "type": "text", + "content": "[20] Sungho Lee, Seoung Wug Oh, DaeYeun Won, and Seon Joo Kim. Copy-and-paste networks for deep video inpainting. In ICCV, 2019. 3, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 166, + 553, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 166, + 553, + 209 + ], + "spans": [ + { + "bbox": [ + 317, + 166, + 553, + 209 + ], + "type": "text", + "content": "[21] Ang Li, Shanshan Zhao, Xingjun Ma, Mingming Gong, Jianzhong Qi, Rui Zhang, Dacheng Tao, and Ramamohanarao Kotagiri. Short-term and long-term context aggregation network for video inpainting. In ECCV, 2020. 1, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 212, + 553, + 244 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 212, + 553, + 244 + ], + "spans": [ + { + "bbox": [ + 317, + 212, + 553, + 244 + ], + "type": "text", + "content": "[22] Zhen Li, Cheng-Ze Lu, Jianhua Qin, Chun-Le Guo, and Ming-Ming Cheng. Towards an end-to-end framework for flow-guided video inpainting. In CVPR, 2022. 2, 5, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 246, + 553, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 246, + 553, + 278 + ], + "spans": [ + { + "bbox": [ + 317, + 246, + 553, + 278 + ], + "type": "text", + "content": "[23] Miao Liao, Feixiang Lu, Dingfu Zhou, Sibo Zhang, Wei Li, and Ruigang Yang. Dvi: Depth guided video inpainting for autonomous driving. In ECCV, 2020. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 281, + 553, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 281, + 553, + 324 + ], + "spans": [ + { + "bbox": [ + 316, + 281, + 553, + 324 + ], + "type": "text", + "content": "[24] Rui Liu, Hanming Deng, Yangyi Huang, Xiaoyu Shi, Lewei Lu, Wenxiu Sun, Xiaogang Wang, Jifeng Dai, and Hongsheng Li. Fuseformer: Fusing fine-grained information in transformers for video inpainting. In ICCV, 2021. 1, 3, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 326, + 553, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 326, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 316, + 326, + 553, + 369 + ], + "type": "text", + "content": "[25] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. In ICLR, 2022. 2, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 372, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 553, + 415 + ], + "type": "text", + "content": "[26] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In AAAI, 2024. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 417, + 553, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 417, + 553, + 438 + ], + "spans": [ + { + "bbox": [ + 316, + 417, + 553, + 438 + ], + "type": "text", + "content": "[27] OpenAI. Gpt-4o contributions. https://openai.com/gpt-4o-contributions, 2024.6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 441, + 553, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 441, + 553, + 472 + ], + "spans": [ + { + "bbox": [ + 316, + 441, + 553, + 472 + ], + "type": "text", + "content": "[28] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In CVPR, 2019. 2, 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 475, + 553, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 475, + 553, + 507 + ], + "spans": [ + { + "bbox": [ + 316, + 475, + 553, + 507 + ], + "type": "text", + "content": "[29] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In CVPR, 2016. 1" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 510, + 553, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 510, + 553, + 553 + ], + "spans": [ + { + "bbox": [ + 316, + 510, + 553, + 553 + ], + "type": "text", + "content": "[30] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In CVPR, 2016. 6" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 317, + 555, + 553, + 608 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 555, + 553, + 608 + ], + "spans": [ + { + "bbox": [ + 317, + 555, + 553, + 608 + ], + "type": "text", + "content": "[31] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: improving latent diffusion models for high-resolution image synthesis. In ICLR, 2024. 2, 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 317, + 612, + 553, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 612, + 553, + 654 + ], + "spans": [ + { + "bbox": [ + 317, + 612, + 553, + 654 + ], + "type": "text", + "content": "[32] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. In ICCV, 2023. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 658, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 678 + ], + "type": "text", + "content": "[33] Prajit Ramachandran, Barret Zoph, and Quoc V. Le. Searching for activation functions. In ICLR Workshops, 2018. 5" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[34] Jingjing Ren, Qingqing Zheng, Yuanyuan Zhao, Xuemiao Xu, and Chen Li. Dlformer: Discrete latent transformer for video inpainting. In CVPR, 2022. 3" + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22961" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[35] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 2, 3, 4, 5, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "type": "text", + "content": "[36] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 184 + ], + "type": "text", + "content": "[37] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary Bradski. Orb: An efficient alternative to sift or surf. In ICCV. IEEE, 2011. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 217 + ], + "type": "text", + "content": "[38] Min-Cheol Sagong, Yoon-Jae Yeo, Seung-Won Jung, and Sung-Jea Ko. Rord: A real-world object removal dataset. In BMVC, 2022. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 220, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 274 + ], + "type": "text", + "content": "[39] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 276, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 330 + ], + "type": "text", + "content": "[40] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-a-video: Text-to-video generation without text-video data. In ICLR, 2023. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 331, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 294, + 374 + ], + "type": "text", + "content": "[41] Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 376, + 294, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 294, + 398 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 294, + 398 + ], + "type": "text", + "content": "[42] Richard Szeliski. Image alignment and stitching: A tutorial. Found. Trends Comput. Graph. Vis., 2006. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 399, + 294, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 294, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 294, + 443 + ], + "type": "text", + "content": "[43] Nick C Tang, Chiou-Ting Hsu, Chih-Wen Su, Timothy K Shih, and Hong-Yuan Mark Liao. Video inpainting on digitized vintage films via maintaining spatiotemporal continuity. IEEE TMM, 2011. 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 445, + 294, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 466 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 466 + ], + "type": "text", + "content": "[44] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In ECCV, 2020. 2, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 468, + 294, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 468, + 294, + 520 + ], + "spans": [ + { + "bbox": [ + 56, + 468, + 294, + 520 + ], + "type": "text", + "content": "[45] Rui Tian, Qi Dai, Jianmin Bao, Kai Qiu, Yifan Yang, Chong Luo, Zuxuan Wu, and Yu-Gang Jiang. Reducio! generating " + }, + { + "bbox": [ + 56, + 468, + 294, + 520 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 56, + 468, + 294, + 520 + ], + "type": "text", + "content": " video within 16 seconds using extremely compressed motion latents. arXiv preprint arXiv:2411.13552, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 523, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 294, + 555 + ], + "type": "text", + "content": "[46] Shuyuan Tu, Qi Dai, Zhi-Qi Cheng, Han Hu, Xintong Han, Zuxuan Wu, and Yu-Gang Jiang. Motioneditor: Editing video motion via content-aware diffusion. In CVPR, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 557, + 294, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 294, + 589 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 294, + 589 + ], + "type": "text", + "content": "[47] Chuan Wang, Haibin Huang, Xiaoguang Han, and Jue Wang. Video inpainting by jointly learning temporal structure and spatial details. In AAAI, 2019. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 591, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 591, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 591, + 294, + 624 + ], + "type": "text", + "content": "[48] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Nikolai Yakovenko, Andrew Tao, Jan Kautz, and Bryan Catanzaro. Video-to-video synthesis. In NeurIPS, 2018. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 625, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 294, + 678 + ], + "type": "text", + "content": "[49] Yanhui Wang, Jianmin Bao, Wenming Weng, Ruoyu Feng, Dacheng Yin, Tao Yang, Jingxu Zhang, Qi Dai, Zhiyuan Zhao, Chunyu Wang, et al. Microcinema: A divide-and-conquer approach for text-to-video generation. In CVPR, 2024. 2, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "text", + "content": "[50] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE TIP, 2004. 6" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 566 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "text", + "content": "[51] Wenming Weng, Ruoyu Feng, Yanhui Wang, Qi Dai, Chunyu Wang, Dacheng Yin, Zhiyuan Zhao, Kai Qiu, Jian-min Bao, Yuhui Yuan, et al. Art-v: Auto-regressive text-to-video generation with diffusion models. In CVPR Workshops, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 129, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 150 + ], + "type": "text", + "content": "[52] Yuxin Wu and Kaiming He. Group normalization. In ECCV, 2018. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 152, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 205 + ], + "type": "text", + "content": "[53] Zhen Xing, Qi Dai, Han Hu, Zuxuan Wu, and Yu-Gang Jiang. Simda: Simple diffusion adapter for efficient video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7827-7839, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 207, + 553, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 207, + 553, + 250 + ], + "spans": [ + { + "bbox": [ + 316, + 207, + 553, + 250 + ], + "type": "text", + "content": "[54] Zhen Xing, Qijun Feng, Haoran Chen, Qi Dai, Han Hu, Hang Xu, Zuxuan Wu, and Yu-Gang Jiang. A survey on video diffusion models. ACM Computing Surveys, 57(2):1-42, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 252, + 553, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 252, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 316, + 252, + 553, + 296 + ], + "type": "text", + "content": "[55] Ning Xu, Linjie Yang, Yuchen Fan, Jianchao Yang, Dingcheng Yue, Yuchen Liang, Brian Price, Scott Cohen, and Thomas Huang. Youtube-vos: Sequence-to-sequence video object segmentation. In ECCV, 2018. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 297, + 553, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 329 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 329 + ], + "type": "text", + "content": "[56] Rui Xu, Xiaoxiao Li, Bolei Zhou, and Chen Change Loy. Deep flow-guided video inpainting. In CVPR, 2019. 2, 3, 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 331, + 553, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 364 + ], + "type": "text", + "content": "[57] Jiahui Yu, Zhe Lin, Jamei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Generative image inpainting with contextual attention. In CVPR, 2018. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 365, + 553, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 397 + ], + "type": "text", + "content": "[58] Yanhong Zeng, Jianlong Fu, and Hongyang Chao. Learning joint spatial-temporal transformations for video inpainting. In ECCV, 2020. 1, 3, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 399, + 553, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 421 + ], + "type": "text", + "content": "[59] Kaidong Zhang, Jingjing Fu, and Dong Liu. Flow-guided transformer for video inpainting. In ECCV, 2022. 2, 3, 5, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 422, + 553, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 553, + 454 + ], + "type": "text", + "content": "[60] Kaidong Zhang, Jingjing Fu, and Dong Liu. Inertia-guided flow completion and style fusion for video inpainting. In CVPR, 2022. 3, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 456, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 487 + ], + "type": "text", + "content": "[61] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 2, 5, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 490, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 490, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 490, + 553, + 521 + ], + "type": "text", + "content": "[62] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In CVPR, 2017. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 523, + 553, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 553, + 566 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 553, + 566 + ], + "type": "text", + "content": "[63] Shangchen Zhou, Chongyi Li, Kelvin CK Chan, and Chen Change Loy. Propainter: Improving propagation and transformer for video inpainting. In ICCV, 2023. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "22962" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Homogeneous Dynamics Space for Heterogeneous Humans/a999c957-db24-46d0-83c3-e6da7809465e_content_list.json b/2025/Homogeneous Dynamics Space for Heterogeneous Humans/a999c957-db24-46d0-83c3-e6da7809465e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9bbdb0cc2ad665a2ebb1ba25bf68abe983a56071 --- /dev/null +++ b/2025/Homogeneous Dynamics Space for Heterogeneous Humans/a999c957-db24-46d0-83c3-e6da7809465e_content_list.json @@ -0,0 +1,1782 @@ +[ + { + "type": "text", + "text": "Homogeneous Dynamics Space for Heterogeneous Humans", + "text_level": 1, + "bbox": [ + 200, + 130, + 797, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xinpeng Liu $^{1,2}$ , Junxuan Liang $^{1}$ , Chenshuo Zhang $^{1}$ , Zixuan Cai $^{3}$ , Cewu Lu $^{1,2*}$ , Yong-Lu Li $^{1,2*}$", + "bbox": [ + 119, + 179, + 883, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Shanghai Jiao Tong University, 2Shanghai Innovation Institute, 3Soochow University xinpengliu0907@gmail.com, zxcai@stu.suda.edu.cn, {whitefork, zhangchenshuo, lucewu, yonglu.li}@sjtu.edu.cn", + "bbox": [ + 133, + 200, + 862, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 286, + 326, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Analyses of human motion kinematics have achieved tremendous advances. However, the production mechanism, known as human dynamics, is still undercovered. In this paper, we aim to push the understanding of data-driven human dynamics forward. We identify a major obstacle to this as the heterogeneity of existing human motion understanding efforts. Specifically, heterogeneity exists in not only the diverse kinematics representations and hierarchical dynamics representations but also the data from different domains, namely biomechanics and reinforcement learning. With an in-depth analysis of the existing heterogeneity, we propose to emphasize the beneath homogeneity: all of them represent the homogeneous fact of human motion, though from different perspectives. Given this, we propose Homogeneous Dynamics Space (HDyS) as a fundamental space for human dynamics by aggregating heterogeneous data and training a homogeneous latent space with inspiration from the inverse-forward dynamics procedure. HDyS achieves decent mapping between human kinematics and dynamics by leveraging the heterogeneous representations and datasets. We demonstrate the feasibility of HDyS with extensive experiments and applications. The project page is https://foruck.github.io/HDyS.", + "bbox": [ + 88, + 318, + 483, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 696, + 220, + 712 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Analyses on human motion have a wide range of applications, including animation [11, 60], healthcare [10, 55], and robotics [57, 70]. The computer vision community has made tremendous progress in understanding human kinematics with tasks like human reconstruction [24, 27, 34], action recognition [25, 26, 28, 29, 51], and motion generation [30, 32, 60]. However, the production mechanism hidden beneath human motion, known as human dynamics, is still limitedly explored.", + "bbox": [ + 89, + 722, + 482, + 859 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we aim to push the understanding of human", + "bbox": [ + 109, + 859, + 482, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e12f39a53841ebb868ad4c4bbeda49a01aa9df498208085ca756edd870cc5d3f.jpg", + "image_caption": [ + "Figure 1. The representation heterogeneity exists for both kinematics and dynamics." + ], + "image_footnote": [], + "bbox": [ + 517, + 286, + 901, + 458 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "dynamics forward with data-driven methodologies. Specifically, we pursue to build a bidirectional mapping between human kinematics and dynamics. We identify the major obstacle to this as the heterogeneity of existing human motion understanding efforts, which could be two-fold.", + "bbox": [ + 511, + 520, + 906, + 595 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "First, representation heterogeneity exists for both kinematics and dynamics, as shown in Figure 1. A typical optical motion capture pipeline, which produces most existing available data for human kinematics [20, 40], involves tracking the markers placed on the surface of humans. Then, different human template models are fitted to the tracked markers, resulting in parameters of musculoskeletal models [52] for biomechanics uses or SMPL-like models [36] for general CV/CG uses. Skeletons with pre-defined kinematic trees are sometimes preferred as a simple and intuitive representation. Different datasets tend to include only representations preferred by their designed goal. Besides the kinematics representation heterogeneity, the dynamics representation heterogeneity exists more severely. Human dynamics function in a hierarchical manner. The closest dynamics hierarchy to the kinematics is the joint torques, which are further produced by muscle forces from muscle actions according to the muscle contraction dynamics. Furthermore, electric signals from the neural system activate the muscle, which surface electromyogra", + "bbox": [ + 511, + 599, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding authors.", + "bbox": [ + 109, + 887, + 241, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "27782", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "phy (sEMG) could detect. Effective conversions between these different representations are not always available. For example, for kinematics representations, the conversion between the musculoskeletal model and SMPL becomes available in very recent advances [20], which requires time-consuming optimization. For dynamics, the bi-directional mapping between joint torques, muscle actions, and electric signals is even more difficult.", + "bbox": [ + 89, + 90, + 480, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Second, domain heterogeneity also exists. One crucial characteristic of human dynamics is the difficulty of non-intrusive measurement. In the biomechanics community, optimization-based methods are adopted to solve the Euler-Newton equation for joint torques [64] and muscle actions [53]. For electric signals, it is typically measured by sEMG sensors [3]. In the meantime, from the learning community, there emerge efforts [33] utilizing Reinforcement Learning (RL) to imitate existing motion captures with physically simulated humanoids and record the simulated human dynamics data in a fully synthetic manner. These different data sources introduce domain heterogeneity in three aspects. First, biomechanics data are more deeply rooted in reality with well-defined computation models and real human motions. Instead, RL data could diverge from real-world situations with unnatural motion imitations and inaccurate physics simulations, known as the sim2real domain gap. Second, biomechanics data, due to the optimization-based manner, is sensitive to variations and usually limited to strictly controlled laboratory setups with relatively simple motions, resulting in limited motion domain coverage. In contrast, RL data cover a wide range of motion with more diversity. Third, biomechanics data and RL data tend to adopt heterogeneous kinematics representations with different kinematic trees, making it hard to transfer directly from one domain to another.", + "bbox": [ + 91, + 215, + 483, + 607 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite the heterogeneity, we emphasize the homogeneous essence behind it: all of them represent the homogeneous fact of human motion, each from a different perspective. Given this, we demonstrate the feasibility of unifying these heterogeneous human representations and exploiting the underlying homogeneous knowledge of human dynamics. Low-level Cartesian kinematics representations, like markers and joints, are less heterogeneous than representations like joint angles. Also, though joint torques, muscle actions, and electric signals are not directly transferable to each other, they could share similar motor knowledge like coordination. To this end, we propose Homogeneous Dynamics Space for heterogeneous humans (HDyS).", + "bbox": [ + 89, + 609, + 482, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To achieve this, we aggregate human dynamics data from both RL [33] and biomechanics [3, 53, 65] with different kinematics and dynamics representations, covering human dynamics hierarchies, including joint torques, muscle actions, and electric signals. Then, HDyS is designed as an aggregation of multiple auto-encoders corresponding to the", + "bbox": [ + 89, + 810, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "inverse-forward dynamics procedure. We supervise HDyS with reconstruction and alignment losses. Our proposed HDyS has two major merits. First, by unifying heterogeneous human representation in the same latent space, it generalizes across different representations seamlessly while taking advantage of each. Second, by aggregating largescale heterogeneous motion data, it is empowered with general human motor knowledge from a wide span of motion, functioning as a reusable knowledge source for downstream dynamics-related applications. To demonstrate the efficacy of HDyS, we first evaluate it on human inverse dynamics. Then, we showcase how HDyS could facilitate downstream dynamics-related applications like ground reaction force prediction, biomechanics human simulation, and physics-simulated character control.", + "bbox": [ + 511, + 90, + 903, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, our contribution includes: 1) We analyzed the heterogeneity issue that hinders an in-depth understanding of human dynamics. 2) We highlighted the homogeneity beneath the heterogeneity and proposed a fundamental reusable space HDyS for human dynamics by unifying the heterogeneity. 3) We demonstrated the feasibility of digging homogeneity out from heterogeneity with extensive experiments and applications of HDyS.", + "bbox": [ + 511, + 316, + 906, + 438 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 450, + 663, + 465 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Human Dynamics", + "text_level": 1, + "bbox": [ + 513, + 474, + 687, + 491 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "By human dynamics, we mean the production mechanism of human motion, which the biomechanics community has actively explored. To produce a certain motion, neural commands are sent to activate the muscles. After receiving the activation signals, muscles contract and produce muscle forces. Multiple muscle forces form the joint torques according to certain musculoskeletal geometry, and the joint torques drive the accelerations that accumulate into movements. Thus, understanding human dynamics typically involves two heterogeneous hierarchies: joint torques and muscle activations. However, both are hard to measure nonintrusively. In the literature, to obtain them, an optimization problem is typically introduced as", + "bbox": [ + 511, + 500, + 906, + 696 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\min \\| a \\|, s. t. 0 \\leq a \\leq 1, \\tau = A (q) F (a), \\\\ M (q) \\ddot {q} + C (q, \\dot {q}) + G (q) = J \\lambda + \\tau . \\tag {1} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 704, + 903, + 739 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "with the generalized human inertia matrix $M(q)$ w.r.t. generalized coordinate $q$ , Coriolis and centrifugal forces $C(q,\\dot{q})$ , gravity $G(q)$ , Jacobian matrix $J$ mapping external forces $\\lambda$ to the generalized coordinates, and muscle activations $a$ . The joint torques $\\tau$ could be obtained by $\\tau = A(q)F(a)$ , where $A(q)$ maps muscle forces into joint torques and $F(a)$ maps $a$ into muscle forces usually with the hill-type function [15, 74]. Mature software [4, 5, 63] were developed for this purpose. However, the optimization quality is tightly bonded to the precision of external", + "bbox": [ + 511, + 750, + 906, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "27783", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "force measurement, which could be expensive. Therefore, the applications are mostly limited to simple motions like gaits in laboratory settings. Some efforts exploited wearable devices [22, 23] for more general applications. In addition, the optimization is deeply coupled with the adopted human models [36, 52], which vary with the application preferences. Fitting raw motion capture data to a specific human model could be unstable and time-consuming. The conversion between different human models is also nontrivial even with recent advances [20]. This way, transferring dynamics from one human model to another is limitedly explored. Besides torques and muscle actions that are typically obtained via optimization, Surface Electromyography (sEMG) is adopted as an indirect representation of human dynamics which could be directly measured. Efforts were made to build the accurate mapping from sEMG to muscle actions [19, 66], joint torques [54, 79], and human poses [35, 55]. Though progress was made, sEMG patterns might suffer from noises and vary drastically among subjects [61], hindering the generalization.", + "bbox": [ + 89, + 90, + 480, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Learning-based Human Dynamics", + "text_level": 1, + "bbox": [ + 89, + 406, + 390, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Joint Torques. Early efforts were made on ML-based joint torque analysis for certain human body parts [17, 43, 69]. Lv et al. [39] developed a Gaussian mixture framework for whole-body joint torque estimation. Other architectures like k-nearest neighbor-based regression [75, 77], random forests [76], and neural networks [78] were also adopted for the estimation of joint torques. However, most of these efforts suffered from limited data scale, which hampered learning methods from exploiting their full potential. The recent emergence of AddBiomechanics [65], which aggregated multiple biomechanics datasets, considerably boosted the data scale. However, most of the collected sequences contained only gaits with limited diversity. Another line of work adopted reinforcement learning to simulate motion in physics simulators, and the joint torques could be obtained in simulation. Though generalizability was limited at first [1, 49, 50, 67, 68], emerging efforts [37, 38, 73] managed to replicate a wide span of motion in simulators. The paradigm is further incorporated into recent MoCap systems for simultaneous estimation of motion and the joint torques [8, 12, 16, 62, 72, 80]. However, due to the involvement of simulators, the sim2real gap exists. The learned dynamics could be restricted to certain simulators or human models, which might diverge from real humans.", + "bbox": [ + 89, + 431, + 482, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Muscle Actions. sEMG is usually adopted as a proxy measurement of muscle actions and has been extensively studied in biomechanics [3, 6, 7, 14, 41, 44, 48]. There have been efforts delving into predicting sEMG signals given either joint torques [31, 56, 58], goniometers [59], motion captures [18, 45, 71], videos [3, 48], or point clouds [46]. However, each effort could be small in data scale, mo", + "bbox": [ + 89, + 795, + 480, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tion variations, and muscle coverage. Unifying these efforts could be hindered by the heterogeneous settings and data formats adopted, thus under-covered. Recently, musculoskeletal human simulation has gained attention. Multiple efficient simulators were developed [2, 9, 81], opening the potential for more accessible muscle action analysis. Specifically, MinT [53] was proposed by simulating motion sequences from AMASS [40] in OpenSim [5] and attaching simulated muscle actions to the raw motion sequence, enabling muscle action learning in scale.", + "bbox": [ + 511, + 90, + 903, + 242 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 511, + 258, + 604, + 273 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We introduce the proposed Homogeneous Dynamics Space (HDyS). We first cover the involved kinematics and dynamics representations in Section 3.1, 3.2. Then, the model architecture and designed losses are introduced in Section 3.3.", + "bbox": [ + 511, + 286, + 903, + 347 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Kinematics Representations", + "text_level": 1, + "bbox": [ + 511, + 359, + 764, + 375 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Fig. 2, we adopt four types of kinematics representations: Cartesian representations of markers and skeleton key-points, joint angles of Rajagopal's model [52] from the biomechanics community, and SMPL [36] which is widely adopted for CV and CG applications.", + "bbox": [ + 511, + 383, + 903, + 458 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Markers, placed on the human body surfaces, are typically the raw data for optical motion capture. In practice, most other representations are calculated by fitting certain human prior models to the marker observations, making it easy to obtain for heterogeneous datasets and representations. Thus, markers are expected to be a generalizable representation across heterogeneous datasets, though they could also be rather low-level and thus hard to learn. We define the marker representation at timestamp $t$ as $x_{m}^{t} = (m^{t},\\dot{m}^{t},\\dot{m})\\in \\mathcal{R}^{N_{m}\\times 9}$ , which is composed of marker Cartesian coordinates $m^t$ , finite-differentiated velocities $\\dot{m}^t$ and finite-differentiated accelerations $\\dot{m}^t$ with $N_{m}$ markers.", + "bbox": [ + 511, + 460, + 905, + 641 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Skeletal Keypoints, compared to markers, are less generalizable due to their reliance on pre-defined kinematic trees. However, with its Cartesian coordinates, common sense of human topology, and easy access, the gap between different kinematic trees can be mitigated. We define the joint representation at timestamp $t$ as $x_{k}^{t} = (k^{t},\\dot{k}^{t},\\dot{k})\\in \\mathcal{R}^{N_{k}\\times 9}$ , which is composed of joint coordinates $k^t$ , finite-differentiated velocities $\\dot{k}^t$ and finite-differentiated accelerations $\\dot{k}^t$ with $N_{k}$ skeletal joints.", + "bbox": [ + 511, + 643, + 905, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Joint Angles are preferred for clinical analyses which more faithfully preserves the biomechanics information of human kinematics. We adopt the Rajagopal's model [52] used in AddBiomechanics [65], and define the joint angle representation at timestamp $t$ as $x_{a}^{t} = (a^{t},\\dot{a}^{t},\\dot{a})\\in \\mathcal{R}^{3N_{j}}$ , containing joint angles $a^t$ , finite-differentiated velocities $\\dot{a}^t$ and accelerations $\\dot{a}^t$ with $N_{j}$ joints in Rajagopal's model (only 23 lower-body joints are used).", + "bbox": [ + 511, + 780, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "27784", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b98c47e8f12e3bb45e35434811d4eea3e2a9dc2220e3b5c3e3ef708495ada98f.jpg", + "image_caption": [ + "Figure 2. The overall architecture of HDyS." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 480, + 343 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SMPL [36] is a human parameter model widely adopted in the CV/CG community. We define the SMPL representation at timestamp $t$ as $x_{s}^{t} = (s^{t},\\dot{s}^{t},\\dot{s})\\in \\mathcal{R}^{3N_{j}}$ , which is composed of SMPL parameters $s^t$ , finite-differentiated velocities $\\dot{s}^t$ and finite-differentiated accelerations $\\dot{s}^t$ with $N_{j}$ joints of SMPL (75 used, 3 translational joints at root and 72 revolving joints).", + "bbox": [ + 88, + 377, + 482, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Dynamics Representations", + "text_level": 1, + "bbox": [ + 89, + 494, + 331, + 510 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We adopt three different types of dynamic representations: joint torques, muscle actions, and sEMGs.", + "bbox": [ + 88, + 517, + 482, + 547 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Joint Torques are the net torques exerted at each joint to drive the motion, represented as $\\tau_{t} \\in \\mathcal{R}^{N_{t}}$ . $N_{t}$ is coupled with the number of joints $N_{j}$ as $N_{t} = N_{j} - 6$ , with the 6-DoF free root joint unactuated. For clarity, we denote the joint torques corresponding to Rajagopal's model as $\\tau_{tr}$ and those corresponding to SMPL as $\\tau_{ts}$ .", + "bbox": [ + 88, + 547, + 482, + 638 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Muscle Actions represent the activation level of each muscle, resulting in the muscle forces that produce joint torques. We denote muscle actions as $\\tau_{m} \\in \\mathcal{R}^{N_{m}}$ with $N_{m}$ muscles. sEMG detects the electric potential generated on the body surface by activated muscle cells, which is closely related to muscle actions. We represent sEMGs at timestamp $t$ as $\\tau_{e}^{t} \\in \\mathcal{R}^{N_{e}}$ with $N_{e}$ sEMG channels.", + "bbox": [ + 89, + 638, + 482, + 746 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Homogeneous Dynamics Space", + "text_level": 1, + "bbox": [ + 89, + 755, + 366, + 772 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With the heterogeneous representations, we construct our homogeneous dynamics space (HDyS) as an aggregation of multiple auto-encoders corresponding to the inverseforward dynamics procedure. Then, reconstruction and alignment losses are adopted for supervision. For simplicity, all superscripts $t$ are omitted.", + "bbox": [ + 89, + 779, + 482, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inverse-Dynamics Auto-Encoder (IDAE) encodes the kinematics into the latent space and decodes the dynamics", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "from the latent. For markers $x_{m}$ and skeletal key points $x_{k}$ , we adopt three-layer transformer encoders with no positional embedding to encode them into latent $z_{m}, z_{k} \\in \\mathcal{R}^{d}$ of dimension $d$ . This enables the encoding of an arbitrary number of markers/points. For joint angles $x_{a}$ and SMPL parameters $x_{s}$ , we adopt simple three-layer MLPs to encode them into latent $z_{a}, z_{s} \\in \\mathcal{R}^{d}$ . The decoder is designed as a shared transformer followed by separate MLP regression heads. The transformer takes latent vectors from consecutive frames as input, outputting per-frame latent vectors refined with temporal contexts. Then, separate MLP decoders decode the per-frame dynamics as $\\hat{\\tau}_{tr}, \\hat{\\tau}_{ts}, \\hat{\\tau}_{m}, \\hat{\\tau}_{e}$ .", + "bbox": [ + 511, + 90, + 903, + 272 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Forward-Dynamics Auto-Encoder (FDAE) encodes dynamics and the kinematics except for accelerations (denoted as $\\tilde{x}$ ) into the latent space and decodes the accelerations. Similar to IDAE, markers $\\tilde{x}_m$ and skeletal key-points $\\tilde{x}_k$ are encoded by transformers as $\\tilde{z}_m,\\tilde{z}_k$ , while joint angles $\\tilde{x}_a$ and SMPL parameters $\\tilde{x}_s$ are encoded by MLPs as $\\tilde{z}_a,\\tilde{z}_p$ . Then, joint torques $\\tau_{tr},\\tau_{ts}$ , muscle activations $\\tau_m$ , and sEMGs $\\tau_e$ are also encoded by separate MLP encoders as dynamics latent vectors $\\tilde{z}^{tr},\\tilde{z}^{ts},\\tilde{z}^m,\\tilde{z}^e$ . $\\tilde{z}_m,\\tilde{z}_k,\\tilde{z}_a,\\tilde{z}_p$ is then concatenated with available dynamics latent vectors, fed into a shared MLP composer, resulted in latent vectors $z^{\\cdot}$ , with subscripts representing the dynamics. Although arbitrary combinations are feasible, in practice, only one dynamic latent vector typically exists due to the data limitations. Finally, separate MLP decoders decode the latent vectors $z_{\\{kin\\}}^{(dyn)}$ into skeletal key point accelerations $\\ddot{k}$ , SMPL accelerations $\\ddot{s}$ , and joint angle accelerations $\\ddot{a}$ , where $kin\\in \\{m,k,s,a\\}$ , $dyn\\in \\{m,e,ts,tr\\}$ . The marker accelerations are not predicted since their variable sizes could bring unnecessary complexity.", + "bbox": [ + 511, + 277, + 906, + 598 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Loss Terms. To train HDyS, we adopt reconstruction losses and alignment losses. For reconstruction losses, we simply calculate the L1 loss for $\\tau$ and $\\ddot{a}$ as", + "bbox": [ + 511, + 602, + 905, + 647 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {r e c o n} = \\| \\tau - \\hat {\\tau} \\| _ {1} + \\| \\ddot {a} - \\hat {\\dot {a}} \\|. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 667, + 906, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For alignment losses, we align the latent vectors $z$ of the same frame and separate $zs$ of different frames using InfoNCE [47]. Given a batch of $B$ frames, the obtained latent vectors is denoted as $Z = \\{z_{\\{kin\\}}, z_{\\{kin\\}}^{(dyn)}\\}$ , with $kin \\in \\{m, k, s, a\\}, dyn \\in \\{m, e, ts, tr\\}$ . The loss is", + "bbox": [ + 511, + 705, + 905, + 787 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {a l i g n} = \\sum_ {z _ {2}, z _ {2} \\in Z} \\sum_ {i = 1} ^ {B} - \\log \\left(\\frac {\\exp \\left(\\left\\langle z _ {1} ^ {(i)} , z _ {2} ^ {(i)} \\right\\rangle\\right)}{\\sum_ {j = 1} ^ {B} \\exp \\left(\\left\\langle z _ {1} ^ {(i)} , z _ {2} ^ {(j)} \\right\\rangle\\right)}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 806, + 906, + 849 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $(i),(j)$ indicate the batch index. The overall loss is as $\\mathcal{L} = \\alpha_{1}L_{recon} + \\alpha_{2}L_{align}$ with coefficients $\\alpha_{1},\\alpha_{2}$", + "bbox": [ + 511, + 869, + 903, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "27785", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/f40df5d834a56571d1001ca7b833c803a074e0f2f7d2e88a78f69be952bc115b.jpg", + "table_caption": [ + "Table 1. Quantitative results of HDyS with ablation studies. For HDyS, the results of averaged predictions and the best prediction among all representations are reported." + ], + "table_footnote": [], + "table_body": "
MethodsImDy mPJE↓ avg/bstAddBiomechanics mPJE↓ avg/bstMinTMiA PCC↑ avg/bst
RMSE↓ avg/bstPCC↑ avg/bstRMSE↓ avg/bst
ImDyS [33]0.63000.1626----
MiA[3]----13.3-
HDyS0.5765/0.46740.1189/0.12430.0614/0.06150.7420/0.740211.8/11.60.7232/0.7261
ImDy-only HDyS0.6854/0.5403-----
AddBiomechanics-only HDyS-0.1695/0.1691----
MinT-only HDyS--0.0637/0.06400.7179/0.7127--
MiA-only HDyS----13.6/13.50.6557/0.6421
HDyS w/o ImDy-0.1214/0.13860.0608/0.06170.7470/0.741715.8/15.60.6523/0.6497
HDyS w/o AddBiomechanics0.5787/0.4742-0.0616/0.06140.7408/0.738617.1/17.00.5769/0.5638
HDyS w/o MinT0.5730/0.46810.1197/0.1296--16.9/16.80.5213/0.5313
HDyS w/o MiA0.5890/0.47880.1200/0.12840.0616/0.06180.7395/0.7375--
HDyS w/o AMASS0.5797/0.47860.1217/0.13190.0613/0.06170.7419/0.738014.7/14.90.5704/0.5632
HDyS w/o Lalign0.6575/0.50190.1270/0.13290.0626/0.06300.7318/0.723813.7/13.40.6464/0.6402
HDyS w/o FDAE0.5776/0.48490.1198/0.12610.0617/0.06170.7388/0.737513.6/13.30.6517/0.6699
HDyS-32D0.7390/0.64500.1401/0.14200.0650/0.06480.6980/0.701016.7/16.30.5524/0.5614
HDyS-64D0.6505/0.54100.1295/0.13540.0627/0.06290.7272/0.725415.1/14.50.6034/0.6187
", + "bbox": [ + 132, + 117, + 861, + 359 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 364, + 223, + 382 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 89, + 397, + 192, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "AddBiomechanics [64] contains over 50 hours of human motion data with joint torques from Nimble [63] simulation. We follow the setting in [64] with the armless part of AddBiomechanics. Markers, key points, and joint angles are used as kinematics representations, and joint torques corresponding to [52] represent the dynamics.", + "bbox": [ + 89, + 428, + 483, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Muscles in Times (MinT) [53] simulates part of AMASS in OpenSim to obtain the actions of 402 muscles. We randomly split them into 906 sequences for training and 227 sequences for testing. Markers, key points, and SMPL parameters are used as kinematics representations, and muscle actions are the dynamics representations.", + "bbox": [ + 89, + 527, + 483, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Muscles in Act (MiA) [3] consists of 12.8 hours of human exercise motion reconstructed from videos with VIBE [21] and sEMG data of 8 muscles. Following [3], we split them into 19,563 training sequences and 3,053 testing sequences. Markers and key points are kinematics representations, and sEMGs are dynamics representations.", + "bbox": [ + 89, + 627, + 483, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ImDy [33] adopted PHC [38] to imitate motion sequences from AMASS [40] and KIT [42], resulting in over 150 hours of human dynamics data, with 27,501 sequences for training and 3,055 sequences for evaluation. Markers, key points, and SMPL are kinematics representations, and SMPL torques are dynamics representations.", + "bbox": [ + 89, + 724, + 483, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "AMASS [40] contains over 11,000 motion sequences represented in SMPL [36]. Though not paired with human dynamics, we included it as an extra knowledge base for training. Markers, key points, and SMPL parameters are used as kinematics representations.", + "bbox": [ + 89, + 824, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 364, + 730, + 381 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We adopt a compact design for HDyS with a latent dimension of 128. For joint angle and SMPL parameters, three-layer MLPs with hidden unit sizes of 256 and 128 are adopted as encoders, projecting the input to the 128-D latent space. For markers and key-points, three-layer transformer encoders with 2 heads are adopted as encoders. The transformer in the ID decoder is of 4 layers, 4 heads, and a dimension of 128. All MLPs in decoders are two-layer, with a hidden size of 32 for Rajagopal's torque, sEMG, joint angle accelerations, and key-point accelerations, and a hidden size of 64 for SMPL torque, SMPL accelerations, and muscle actions. For training, we adopt an AdamW optimizer with a learning rate of 1e-3 and a batch size of 9,600 frames for 1,000 epochs. Loss weights are set as $\\alpha_{1} = 0.01$ , $\\alpha_{2} = 0.05$ . Also, during training, a balanced sampling strategy is adopted to minimize the scale influence of different datasets. That is, we randomly sample 3,000 sequences per dataset for each training epoch. In this way, the comparison is fair in terms of the number of seen samples under different dataset settings. All experiments are conducted on a single NVIDIA Titan Xp GPU.", + "bbox": [ + 511, + 388, + 906, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Results on Inverse Dynamics", + "text_level": 1, + "bbox": [ + 511, + 714, + 769, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metric. We report mPJE and RMSE as", + "bbox": [ + 511, + 737, + 781, + 753 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nm P J E _ {\\tau .} = \\frac {1}{J} \\sum_ {j = 1} ^ {J} | \\tau . - \\hat {\\tau}. | _ {2}, R M S E _ {\\tau .} = | \\tau . - \\hat {\\tau}. | _ {2}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 763, + 906, + 806 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "PCC is also reported due to its invariance against scale and offset, which helps evaluate muscle actions and sEMG prediction. PCC is computed as", + "bbox": [ + 511, + 815, + 905, + 861 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP C C _ {\\tau_ {1}} = \\frac {\\operatorname {c o v} \\left(\\tau , \\hat {\\tau}\\right)}{\\sigma_ {\\tau_ {1}} \\sigma_ {\\hat {\\tau}}} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 869, + 906, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "27786", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $cov(\\tau, \\hat{\\tau})$ is the covariance of $\\tau$ , $\\hat{\\tau}$ , and $\\sigma$ . indicates the standard deviation. For ImDy and AddBiomechanics, we report mPJE normalized by body weight. For MinT and MiA, RMSE and PCC are reported.", + "bbox": [ + 89, + 90, + 483, + 151 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3.1 Quantitative results", + "text_level": 1, + "bbox": [ + 89, + 176, + 282, + 191 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative results are shown in Tab. 1, where HDyS is reported with the averaged predictions from different input kinematics representations and the best predictions among all representations. Compared to previous methods [3, 33], the proposed HDyS provides superior performances with substantial improvements on all datasets. We further analyze the performance of HDyS with three questions.", + "bbox": [ + 89, + 203, + 483, + 309 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "How do the heterogeneous datasets contribute to HDyS?", + "bbox": [ + 89, + 310, + 483, + 324 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare HDyS with its single-dataset variants and drop-one-out variants in Tab. 1. HDyS outperforms the corresponding single-dataset variants on all datasets, validating the existence of homogeneous human dynamics knowledge behind these heterogeneous datasets. According to the drop-one-out results, it is noticeable that datasets with similar dynamics representations are more cooperative. That is, muscle-related datasets (MiA and MinT) tend to benefit more from each other and less from torque-related datasets (AddBiomechanics and ImDy), and vice versa. Moreover, mutually harmful effects could be observed for ImDy and MinT. These are consistent with the gaps between these datasets: the sim2real gap from ImDy to others and the torque-to-muscle gap from AddBiomechanics and ImDy to MiA and MinT. AMASS, though not paired with dynamics information, is shown to be beneficial for inverse dynamics with its diverse and high-quality kinematics.", + "bbox": [ + 89, + 325, + 483, + 582 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The heterogeneous datasets introduce both increased data scale and heterogeneous knowledge. Given this, we try to investigate further the source of the improvement in Tab. 1. Trying to decompose the contributions of scale and heterogeneity, we compared the performance of the following three models on a single target test set in Tab. 2. HDyS-Single-50 denotes single-dataset HDyS with $50\\%$ of the data from the target dataset. HDyS-50/50 denotes HDyS using $50\\%$ of the data from the target dataset and $50\\%$ of the data from the other datasets, maintaining the same data scale as the target dataset. HDyS-Single represents the corresponding single-dataset variants of HDyS in Tab. 1. The performance of the best representation is reported on AddBiomechanics and MiA, considering their suitable scale and realistic nature. As shown, HDyS-50/50 consistently outperforms HDyS-Single-50, verifying the existence of homogeneous knowledge in heterogeneous data. Moreover, on AddBiomechanics, HDyS-50/50 even outperforms HDyS-Single, indicating that the diversity from heterogeneous data could sometimes benefit more than simply increasing seemingly homogeneous data.", + "bbox": [ + 89, + 584, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0e16789ec12e9676c56efde538758cfe3f41ca2faff0edc590db2db3e7c2fdc5.jpg", + "table_caption": [ + "Table 2. Results on scale-heterogeneity decomposition." + ], + "table_footnote": [], + "table_body": "
DatasetHDyS-Single-50HDyS-50/50HDyS-Single
AddBiomechanics mPJE ↓0.17070.12840.1695
MiA RMSE ↓16.214.513.5
", + "bbox": [ + 517, + 103, + 903, + 145 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/699fc81e0cac1bd36565cab198d4a647cc513fc92cd0901af8ac570d4080f877.jpg", + "table_caption": [ + "Table 3. Results of different kinematic representations in HDyS." + ], + "table_footnote": [], + "table_body": "
MethodsImDy mPJE↓AddBiomechanics mPJE↓MinTMiA
RMSE↓PCC↑RMSE↓PCC↑
HDys-marker0.81630.14550.06460.696812.30.7088
HDys-keypoint0.80840.13240.06400.710311.60.7261
HDys-angle-0.1243----
HDys-SMPL0.4674-0.06150.7402--
", + "bbox": [ + 517, + 162, + 903, + 234 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3c9f68164f38b8075ff6174233b461c36845759336cac95042a508218c559d09.jpg", + "image_caption": [ + "Heterogeneous Representations" + ], + "image_footnote": [], + "bbox": [ + 524, + 246, + 730, + 364 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1a9e8fb895845f39a4673e9f0c0ea54b30022429f10b1f2537ca9a573b8196be.jpg", + "image_caption": [ + "Heterogeneous Datasets" + ], + "image_footnote": [], + "bbox": [ + 738, + 246, + 901, + 364 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c944f285a4618bfc85270fe41c207c63756ae4492db652941e34227b808d5c71.jpg", + "image_caption": [ + "Figure 3. Feature visualization for HDyS w/ and w/o $L_{align}$ ." + ], + "image_footnote": [], + "bbox": [ + 516, + 368, + 730, + 460 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/81c5146fdcf80211a3b3c4ce4d5b1b782d035dda3aac2f149e57e487e741ca12.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 738, + 368, + 901, + 460 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "How do the heterogeneous representations contribute to HDyS? We report the results from individual kinematics representations in Tab. 3. Typically, the representations in joint space like SMPL and joint angles produce more accurate predictions compared to Cartesian representations with considerable performance gaps, which explains the performance drop when taking the average for ImDy and MiA. The markers are the worst-performed representation of all.", + "bbox": [ + 511, + 506, + 903, + 626 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Tab. 1, it is noticeable that MinT and AddBiomechanics take advantage of averaging predictions from different representations. For AddBiomechanics, this might be because the best-performing representation is joint angles, which are not available in other datasets. Therefore, the averaging could effectively aggregate knowledge from other datasets with other representations. For MinT, it is observable that different representations produce close results, preventing the ensemble from being held back by markers.", + "bbox": [ + 511, + 627, + 905, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Furthermore, we examine the alignment of heterogeneous representations by removing the alignment loss $L_{\\text{align}}$ in Tab. 1. Considerable performance drops could be observed. The drop is consistent for both best and average performances, indicating the cross-modality alignment manages to aggregate knowledge from heterogeneous representations. We also visualize HDyS features w/ and w/o $L_{\\text{align}}$ in Fig. 3. As shown, $L_{\\text{align}}$ prevents the HDyS feature from being dominated by the representation hetero-", + "bbox": [ + 511, + 763, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "27787", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c338a3d62238cc2d69ce42e615c3b8fb8faee7744c9fbf3b7ca30145a7145f6f.jpg", + "image_caption": [ + "Figure 4. Inverse dynamics visualization on ImDy." + ], + "image_footnote": [], + "bbox": [ + 93, + 87, + 480, + 277 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "geneity, encouraging the learning of homogeneous knowledge. Interestingly, though not intended to do so, $L_{\\text{align}}$ also helps to mitigate the gap between datasets, alleviating the domain heterogeneity.", + "bbox": [ + 89, + 313, + 482, + 375 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "How does the model design contribute to HDyS? We first evaluate the FDAE in Tab. 1. A substantial performance drop could be observed, demonstrating the benefit of informing HDyS with the forward dynamics procedure. Moreover, we also report HDyS with different dimensions in Tab. 1. Though HDyS's performance degrades with lower dimensions, it could still outperform some single-dataset variants with higher dimensions, validating the efficacy of heterogeneous datasets again.", + "bbox": [ + 89, + 375, + 483, + 511 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.2 Qualitative results", + "text_level": 1, + "bbox": [ + 89, + 531, + 272, + 545 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We visualize the inverse dynamics results of different datasets. As shown in Fig. 4-5, the joint torques could be faithfully reconstructed for different actions from either synthetic and jittering ImDy or realistic AddBiomechanics with the same HDyS. For the muscle actions shown in Fig. 6, HDyS produces reasonable estimations for lower-body muscles when walking. Moreover, as in the right half, though the lower-body kinematics are less significant, HDyS could capture the ankle muscle actions with high fidelity. The predictions are also coherent with GT for upper-body muscles like the internal oblique and levator scapulae. More details are in the appendix.", + "bbox": [ + 89, + 555, + 482, + 736 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Results on Downstream Tasks", + "text_level": 1, + "bbox": [ + 89, + 747, + 356, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4.1 Ground Reaction Force Estimation", + "text_level": 1, + "bbox": [ + 89, + 770, + 387, + 784 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate the effectiveness of HDyS on GRF estimation with GroundLink [13], which contains 1.5-hour motion with GRF recordings. We finetune HDyS with subjects 1-6 and evaluate it on subject 7. mPJE for GRF at both feet normalized by body weight is reported following Eq. 4. GroundLinkNet [13] and HDyS trained on GroundLink from scratch are also compared. As shown in Tab. 4, the", + "bbox": [ + 89, + 794, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/448c23ac1e11f1c1637d97dadcd0dad8e1c6c17e3b9eb388ada01297c71e3f12.jpg", + "image_caption": [ + "Figure 5. Inverse dynamics visualization on AddBiomechanics. Table 4. Results on GroundLink." + ], + "image_footnote": [], + "bbox": [ + 532, + 90, + 883, + 284 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6ee49300657fe733b2a8f4e46207aa15dffebbbaa11faed0dd7570b3e08bc783.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsGroundLinkNetGroundLink-onlyHDySHDyS
L-Foot mPJE ↓0.07110.05840.0514
R-Foot mPJE ↓0.09120.07510.0694
", + "bbox": [ + 517, + 316, + 903, + 364 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/13134cead5f08e4039c34aecb3c752f4e8f56ce7129572086539a8892f41b74f.jpg", + "table_caption": [ + "Table 5. Biomechanical human simulation results reported in per-frame MSE." + ], + "table_footnote": [], + "table_body": "
k12345
Optimized0.000630.19091.83062.01062.6027
HDyS0.000610.18601.71181.86512.2233
", + "bbox": [ + 535, + 400, + 883, + 448 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "finetuned HDyS outperforms its counterparts, indicating the efficacy of the aggregated homogeneous knowledge. Also, the GroundLink-only HDyS outperforms GroundLinkNet, reflecting the feasibility of unifying heterogeneous representations for better dynamics knowledge.", + "bbox": [ + 511, + 460, + 905, + 536 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4.2 Biomechanical Human Simulation", + "text_level": 1, + "bbox": [ + 511, + 558, + 805, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "HDyS could also be adopted for biomechanical human simulation. We start with the armless Rajagopal's model [52] in Nimble [63]. Given a motion sequence, we first adopt HDyS to estimate the joint torques. Then, we use the predicted torques to reproduce the motion in Nimble. Starting from the current state, we feed the predicted torques for $k$ steps and compare the simulated joint angles $\\hat{q}$ with the real joint angles $q$ . The simulation is performed at 90FPS. We report the per-frame MSE of joint angles.", + "bbox": [ + 511, + 582, + 905, + 717 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results. We demonstrated the results in Tab. 5 and Fig. 7. The results with optimized torques are also reported as a reference. Surprisingly, for different $k$ , simulation with HDyS is superior with better stability. However, with $k$ increasing, MSE also increases noticeably, indicating the accumulation of drifting errors. Further enhancing the forward-dynamics compatibility of HDyS would be a promising goal.", + "bbox": [ + 511, + 718, + 905, + 825 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4.3 Physical Character Control", + "text_level": 1, + "bbox": [ + 511, + 845, + 759, + 859 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We adopt HDyS for physical character control following the setting of PHC [38], with 140 testing sequences and the rest", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "27788", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0ab4675e5c355f18458edf063e5773631eb42ff28d997c2969f63b646fdf3f49.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 133, + 88, + 426, + 130 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/de9a76612f8d11a892a9cb0b33d88f7f164d4afbcbec1415a46b35074afb2284.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 436, + 88, + 560, + 126 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a20d64388348ebf1278ac27c378ff72c90f9cc0ecab68900d23db74d784a6156.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 573, + 89, + 846, + 128 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f0ea7a9623921a6c62d080b9ef8be3394b4d16b4dc1e4c21b48f6fa03c4a38d5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 135, + 132, + 424, + 202 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/930bd05651b006f5601ab61f5cdd2e86634c999a666b68cac522e5143d4c3467.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 426, + 132, + 859, + 202 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/04852060261f962f044e6d12cfa74873885975b5f1be74cc1d6dd2de5ea5c819.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 135, + 205, + 424, + 242 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/47e448ca6753e3bbdd9e68c4d385dc6ff0bd6cc1ad2166c9d86aeddbdeaf4fe3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 426, + 205, + 857, + 242 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6f3373ffaf9c1b2d3cc53659bcbe2263882375315ddbd973df17b97f85258a8d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 135, + 244, + 424, + 281 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b4cc7e30a2a58b8c7009d63846f7184b470a94a1434c400b2239f423d5fb4c37.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 426, + 244, + 857, + 281 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2bb7e2f6e9992dcbd3c5a41957ff84379d758c8e15d4d303eac2f81eec048c71.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 135, + 282, + 424, + 319 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/04e0433ae86687811d3b2f0e401b14d344cffedb2095cc4e4bfe8c780fbbd692.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 426, + 282, + 857, + 319 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0243863b503b6d33a9f026b8db299300a42caaa036b727fc336d072d20197d24.jpg", + "image_caption": [ + "Figure 6. Inverse dynamics results on MinT." + ], + "image_footnote": [], + "bbox": [ + 135, + 320, + 424, + 356 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/07afc5c80888e1f19e2dc422c51add60356027182e82c147a7cdff20b05ff460.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 426, + 320, + 857, + 356 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fed51097ebb0a64b17e3d3fcbe7fe49b093000590aa60372d7a2d12891a1560e.jpg", + "image_caption": [ + "Figure 7. Biomechanical human simulation visualization." + ], + "image_footnote": [], + "bbox": [ + 109, + 385, + 459, + 686 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/624d6c4ef3b258b7883b1a5e6719892e0790ee0e256f84ae18d6b35d8095028d.jpg", + "table_caption": [ + "Table 6. Results of HDyS for character control." + ], + "table_footnote": [], + "table_body": "
MethodsmPJPEg↓mPJEL↓Evel↓Eacc↓
Baseline74.51449.02911.56214.243
Baseline + HDyS72.69247.92011.30513.757
", + "bbox": [ + 112, + 737, + 460, + 782 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "for training. We train the baseline as a primitive in PHC with a batch size of 700 for 10k steps. Then, the HDyS latents of key points are inserted as extra observations. We report the global and local mPJPE and the errors of velocity and acceleration in Tab. 6. HDyS manages to improve its performance, validating its efficacy.", + "bbox": [ + 89, + 809, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Discussion", + "text_level": 1, + "bbox": [ + 513, + 383, + 627, + 398 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Despite the impressive results of HDyS, it could be improved. First, it is noticeable that HDyS performs better for the lower body than the higher body in Fig. 6. This might be due to the imbalanced focus on lower body dynamics in data adopted by HDyS. For all datasets, data on lower-body dynamics like gaits are dominating. For AddBiomechanics, we only adopted its armless part. Enhancing HDyS with more upper-body dynamics would be helpful. Second, for muscle actions, HDys could omit minor changes, and the magnitudes could sometimes diverge from the real. Mitigating it with more curated models and loss terms is desirable. Third, as a first step toward homogeneous human dynamics learning, HDyS is primarily instantiated with five initial datasets and an intuitive model. Scaling HDyS up with more datasets and more human priors would be a meaningful goal. Finally, we demonstrate the potential of HDyS with some simple applications in Sec. 4.4, while more sophisticated use cases for musculoskeletal human simulation, humanoid control, and human-robot transfer learning would be promising as future works.", + "bbox": [ + 511, + 407, + 906, + 710 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 724, + 633, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We analyzed the heterogeneity issue existing for human dynamics learning and highlighted the homogeneity beneath it. To fully exploit the homogeneity, we proposed HDyS as a homogeneous human dynamics space. Extensive experiments were conducted to validate the feasibility of digging homogeneity out from heterogeneity for human dynamics with detailed analyses of the contribution of heterogeneous components. We further demonstrated the potential of HDyS for downstream applications. We believe HDyS could shed new light on human dynamics understanding.", + "bbox": [ + 511, + 750, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "27789", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is supported in part by the National Natural Science Foundation of China under Grant No.62306175, CCF-Tencent Rhino-Bird Open Research Fund.", + "bbox": [ + 89, + 114, + 485, + 161 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 172, + 187, + 188 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Kevin Bergamin, Simon Clavet, Daniel Holden, and James Richard Forbes. Drecon: data-driven responsive control of physics-based characters. ACM Transactions On Graphics (TOG), 38(6):1-11, 2019. 3", + "[2] Vittorio Caggiano, Huawei Wang, Guillaume Durandau, Massimo Sartori, and Vikash Kumar. Myosuite: A contact-rich simulation suite for musculoskeletal motor control. In Learning for Dynamics and Control Conference, pages 492-507. PMLR, 2022. 3", + "[3] Mia Chiquier and Carl Vondrick. Muscles in action. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22091-22101, 2023. 2, 3, 5, 6", + "[4] Michael Damsgaard, John Rasmussen, Søren Tørholm Christensen, Egidijus Surma, and Mark De Zee. Analysis of musculoskeletal systems in the anybody modeling system. Simulation Modelling Practice and Theory, 14(8): 1100-1111, 2006. 2", + "[5] Scott L Delp, Frank C Anderson, Allison S Arnold, Peter Loan, Ayman Habib, Chand T John, Eran Guendelman, and Darryl G Thelen. Opensim: open-source software to create and analyze dynamic simulations of movement. IEEE transactions on biomedical engineering, 54(11):1940-1950, 2007. 2, 3", + "[6] Benedikt Feldotto, Cristian Soare, Alois Knoll, Piyanee Sriya, Sarah Astill, Marc de Kamps, and Samit Chakrabarty. Evaluating muscle synergies with emg data and physics simulation in the neurorobotics platform. Frontiers in Neurorobotics, 16:856797, 2022. 3", + "[7] Mariusz P Furmanek, Madhur Mangalam, Mathew Yarossi, Kyle Lockwood, and Eugene Tunik. A kinematic and emg dataset of online adjustment of reach-to-grasp movements to visual perturbations. Scientific data, 9(1):23, 2022. 3", + "[8] E. Gartner, M. Andriluka, E. Coumans, and C. Sminchisescu. Differentiable dynamics for articulated 3d human motion reconstruction. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13180-13190, Los Alamitos, CA, USA, 2022. IEEE Computer Society. 3", + "[9] Thomas Geijtenbeek. The Hyfydy simulation software, 2021. https://hyfydy.com.3", + "[10] Daniel FN Gordon, Andreas Christou, Theodoros Stouraitis, Michael Gienger, and Sethu Vijayakumar. Learning personalised human sit-to-stand motion strategies via inverse musculoskeletal optimal control. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 10497–10503. IEEE, 2023. 1", + "[11] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages" + ], + "bbox": [ + 93, + 198, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "5142-5151, Los Alamitos, CA, USA, 2022. IEEE Computer Society. 1", + "[12] Erik Gartner, Mykhaylo Andriluka, Hongyi Xu, and Cristian Sminchisescu. Trajectory optimization for physics-based reconstruction of 3d human pose from monocular video. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13096-13105, Los Alamitos, CA, USA, 2022. IEEE Computer Society. 3", + "[13] Xingjian Han, Ben Senderling, Stanley To, Deepak Kumar, Emily Whiting, and Jun Saito. Groundlink: A dataset unifying human body movement and ground reaction dynamics. In SIGGRAPH Asia 2023 Conference Papers, New York, NY, USA, 2023. Association for Computing Machinery. 7", + "[14] Oscar G Hernández, Jose M Lopez-Castellanos, Carlos A Jara, Gabriel J Garcia, Andres Ubeda, Vicente Morell-Gimenez, and Francisco Gomez-Donoso. A kinematic, imaging and electromyography dataset for human muscular manipulability index prediction. Scientific Data, 10(1):132, 2023. 3", + "[15] Archibald Vivian Hill. The heat of shortening and the dynamic constants of muscle. Proceedings of the Royal Society of London. Series B-Biological Sciences, 126(843):136-195, 1938. 2", + "[16] B. Huang, L. Pan, Y. Yang, J. Ju, and Y. Wang. Neural mo- con: Neural motion control for physically plausible human motion capture. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6407- 6416, Los Alamitos, CA, USA, 2022. IEEE Computer Society. 3", + "[17] Leif Johnson and Dana H. Ballard. Efficient codes for inverse dynamics during walking. In Proceedings of the Twenty-Eighth AAAI Conference on Artificial Intelligence, pages 343-349, Quebec City, Quebec, Canada, 2014. AAAI Press. 3", + "[18] Lise A Johnson and Andrew J Fuglevand. Evaluation of probabilistic methods to predict muscle activity: implications for neuroprosthetics. Journal of neural engineering, 6(5):055008, 2009. 3", + "[19] Kimoon Kang, Kiwon Rhee, and Hyun-Chool Shin. Event detection of muscle activation using an electromyogram. Applied Sciences, 10(16):5593, 2020. 3", + "[20] Marilyn Keller, Keenon Werling, Soyong Shin, Scott Delp, Sergi Pujades, C Karen Liu, and Michael J Black. From skin to skeleton: Towards biomechanically accurate 3d digital humans. ACM Transactions on Graphics (TOG), 42(6): 1-12, 2023. 1, 2, 3", + "[21] Muhammed Kocabas, Nikos Athanasiou, and Michael J Black. Vibe: Video inference for human body pose and shape estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5253-5263, 2020. 5", + "[22] Claudia Latella, Naveen Kuppuswamy, Francesco Romano, Silvio Traversaro, and Francesco Nori. Whole-body human inverse dynamics with distributed micro-accelerometers, gyros and force sensing. Sensors, 16(5), 2016. 3", + "[23] Claudia Latella, Silvio Traversaro, Diego Ferigo, Yeshasvi Tirupachuri, Lorenzo Rapetti, Francisco Javier An" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "27790", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "drade Chavez, Francesco Nori, and Daniele Pucci. Simultaneous floating-base estimation of human kinematics and joint torques. Sensors, 19(12), 2019. 3", + "[24] Jiefeng Li, Chao Xu, Zhicun Chen, Siyuan Bian, Lixin Yang, and Cewu Lu. Hybrik: A hybrid analytical-neural inverse kinematics solution for 3d human pose and shape estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3383-3393, 2021. 1", + "[25] Sheng-Wei Li, Zi-Xiang Wei, Wei-Jie Chen, Yi-Hsin Yu, Chih-Yuan Yang, and Jane Yung-jen Hsu. Sa-dvae: Improving zero-shot skeleton-based action recognition by disentangled variational autoencoders. In European Conference on Computer Vision, pages 447–462. Springer, 2024. 1", + "[26] Yong-Lu Li, Siyuan Zhou, Xijie Huang, Liang Xu, Ze Ma, Hao-Shu Fang, Yanfeng Wang, and Cewu Lu. Transferable interactiveness knowledge for human-object interaction detection. In CVPR, 2019. 1", + "[27] Yong-Lu Li, Xinpeng Liu, Han Lu, Shiyi Wang, Junqi Liu, Jiefeng Li, and Cewu Lu. Detailed 2d-3d joint representation for human-object interaction. In CVPR, 2020. 1", + "[28] Yong-Lu Li, Liang Xu, Xinpeng Liu, Xijie Huang, Yue Xu, Shiyi Wang, Hao-Shu Fang, Ze Ma, Mingyang Chen, and Cewu Lu. Pastanet: Toward human activity knowledge engine. In CVPR, 2020. 1", + "[29] Yong-Lu Li, Xinpeng Liu, Xiaogian Wu, Yizhuo Li, Zuoyu Qiu, Liang Xu, Yue Xu, Hao-Shu Fang, and Cewu Lu. Hake: A knowledge engine foundation for human activity understanding. TPAMI, 2022. 1", + "[30] Yong-Lu Li, Xiaogian Wu, Xinpeng Liu, Yiming Dou, Yikun Ji, Junyi Zhang, Yixing Li, Jingru Tan, Xudong Lu, and Cewu Lu. From isolated islands to pangea: Unifying semantic space for human action understanding. In CVPR, 2024. 1", + "[31] Zhan Li, David Guiraud, and Mitsuhiro Hayashibe. Inverse estimation of multiple muscle activations from joint moment with muscle synergy extraction. IEEE journal of biomedical and health informatics, 19(1):64-73, 2014. 3", + "[32] Xinpeng Liu, Haowen Hou, Yanchao Yang, Yong-Lu Li, and Cewu Lu. Revisit human-scene interaction via space occupancy. In European Conference on Computer Vision, pages 1-19. Springer, 2024. 1", + "[33] Xinpeng Liu, Junxuan Liang, Zili Lin, Haowen Hou, Yong-Lu Li, and Cewu Lu. Imdy: Human inverse dynamics from imitated observations. In The Thirteenth International Conference on Learning Representations, 2025. 2, 5, 6", + "[34] Xiaoyang Liu, Boran Wen, Xinpeng Liu, Zizheng Zhou, Hongwei Fan, Cewu Lu, Lizhuang Ma, Yulong Chen, and Yong-Lu Li. Interacted object grounding in spatio-temporal human-object interactions. In AAAI, 2025. 1", + "[35] Yilin Liu, Shijia Zhang, and Mahanth Gowda. Neuropose: 3d hand pose tracking using emg wearables. In Proceedings of the Web Conference 2021, pages 1471-1482, 2021. 3", + "[36] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. Smpl: a skinned multiperson linear model. In ACM Transactions on Graphics, New York, NY, USA, 2015. Association for Computing Machinery. 1, 3, 4, 5" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Zhengyi Luo, Ryo Hachiuma, Ye Yuan, and Kris Kitani. Dynamics-regulated kinematic policy for egocentric pose estimation. Advances in Neural Information Processing Systems, 34:25019-25032, 2021. 3", + "[38] Z. Luo, J. Cao, A. Winkler, K. Kitani, and W. Xu. Perpetual humanoid control for real-time simulated avatars. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 10861-10870, Los Alamitos, CA, USA, 2023. IEEE Computer Society. 3, 5, 7", + "[39] Xiaolei Lv, Jinxiang Chai, and Shihong Xia. Data-driven inverse dynamics for human motion. ACM Transactions on Graphics (TOG), 35(6):1-12, 2016. 3", + "[40] N. Mahmood, N. Ghorbani, N. F. Troje, G. Pons-Moll, and M. Black. Amass: Archive of motion capture as surface shapes. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 5441-5450, Los Alamitos, CA, USA, 2019. IEEE Computer Society. 1, 3, 5", + "[41] Nebojša Malešević, Alexander Olsson, Paulina Sager, Elin Andersson, Christian Cipriani, Marco Controzzi, Anders Björkman, and Christian Antfolk. A database of high-density surface electromyogram signals comprising 65 isometric hand gestures. Scientific Data, 8(1):63, 2021. 3", + "[42] Christian Mandery, Ömer Terlemez, Martin Do, Nikolaus Vahrenkamp, and Tamim Asfour. Unifying representations and large-scale whole-body motion databases for studying human motion. IEEE Transactions on Robotics, 32(4):796-809, 2016. 5", + "[43] Mykhailo Manukian, Serhii Bahdasariants, and Sergiy Yakovenko. Artificial physics engine for real-time inverse dynamics of arm and hand movement. Plos one, 18(12): e0295750, 2023. 3", + "[44] Luís Moreira, Joana Figueiredo, Pedro Fonseca, João P Vilas-Boas, and Cristina P Santos. Lower limb kinematic, kinetic, and emg data from young healthy humans during walking at controlled speeds. Scientific data, 8(1):103, 2021. 3", + "[45] Yoshihiko Nakamura, Katsu Yamane, Yusuke Fujita, and Ichiro Suzuki. Somatosensory computation for man-machine interface from motion-capture data and musculoskeletal human model. IEEE Transactions on Robotics, 21(1):58-66, 2005. 3", + "[46] Hui Niu, Takahiro Ito, Damien Desclaux, Ko Ayusawa, Yusuke Yoshiyasu, Ryusuke Sagawa, and Eiichi Yoshida. Estimating muscle activity from the deformation of a sequential 3d point cloud. Journal of Imaging, 8(6):168, 2022. 3", + "[47] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 4", + "[48] Kunyu Peng, David Schneider, Alina Roitberg, Kailun Yang, Jiaming Zhang, Chen Deng, Kaiyu Zhang, M Saqib Sarfraz, and Rainer Stiefelhagen. Towards video-based activated muscle group estimation in the wild. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 4495-4504, 2024. 3", + "[49] Xue Bin Peng, Ze Ma, Pieter Abbeel, Sergey Levine, and Angjoo Kanazawa. Amp: Adversarial motion priors for stylized physics-based character control. ACM Transactions on Graphics (ToG), 40(4):1-20, 2021. 3" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "27791", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[50] Xue Bin Peng, Yunrong Guo, Lina Halper, Sergey Levine, and Sanja Fidler. Ase: Large-scale reusable adversarial skill embeddings for physically simulated characters. ACM Transactions On Graphics (TOG), 41(4):1-17, 2022. 3", + "[51] A. R. Punnakkal, A. Chandrasekaran, N. Athanasiou, A. Quiros-Ramirez, and M. J. Black. Babel: Bodies, action and behavior with english labels. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 722-731, Los Alamitos, CA, USA, 2021. IEEE Computer Society. 1", + "[52] Apoorva Rajagopal, Christopher L Dembia, Matthew S Demers, Denny D Delp, Jennifer L Hicks, and Scott L Delp. Full-body musculoskeletal model for muscle-driven simulation of human gait. IEEE transactions on biomedical engineering, 63(10):2068-2079, 2016. 1, 3, 5, 7", + "[53] David Schneider, Simon Reiß, Marco Kugler, Alexander Jaus, Kunyu Peng, Susanne Sutschet, Muhammad Saqib Sarfraz, Sven Matthiesen, and Rainer Stiefelhagen. Muscles in time: Learning to understand human motion in-depth by simulating muscle activations. Advances in Neural Information Processing Systems, 2025. 2, 3, 5", + "[54] Robert V Schulte, Marijke Zondag, Jaap H Buurke, and Erik C Prinsen. Multi-day emg-based knee joint torque estimation using hybrid neuromusculoskeletal modelling and convolutional neural networks. Frontiers in Robotics and AI, 9:869476, 2022. 3", + "[55] Paniz Sedighi, Xingyu Li, and Mahdi Tavakoli. Emg-based intention detection using deep learning for shared control in upper-limb assistive exoskeletons. IEEE Robotics and Automation Letters, 2023. 1, 3", + "[56] Masashi Sekiya, Sho Sakaino, and Tsuji Toshiaki. Linear logistic regression for estimation of lower limb muscle activations. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 27(3):523-532, 2019. 3", + "[57] Antun Skuric, Vincent Padois, Nasser Rezzoug, and David Daney. On-line feasible wrench polytope evaluation based on human musculoskeletal models: an iterative convex hull method. IEEE robotics and automation letters, 7(2):5206-5213, 2022. 1", + "[58] Hyungeun Song and Yoichi Hori. Inverse muscle group activity estimation based on neuromusculoskeletal system model. In TENCON 2015-2015 IEEE Region 10 Conference, pages 1-5. IEEE, 2015. 3", + "[59] Yokhesh K Tamilselvam, Jacky Ganguly, Rajni V Patel, and Mandar Jog. Musculoskeletal model to predict muscle activity during upper limb movement. *Ieee Access*, 9:111472–111485, 2021. 3", + "[60] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit Haim Bermano. Human motion diffusion model. In The Eleventh International Conference on Learning Representations ICLR 2023, Kigali, Rwanda, 2023. OpenReview.net. 1", + "[61] Elly Trepman, Richard E Gellman, Lyle J Micheli, and CARLO J De Luca. Electromyographic analysis of grandplie in ballet and modern dancers. Medicine and science in sports and exercise, 30(12):1708-1720, 1998. 3", + "[62] J. Wang, Y. Yuan, Z. Luo, K. Xie, D. Lin, U. Iqbal, S. Fidler, and S. Khamis. Learning human dynamics in autonomous" + ], + "bbox": [ + 91, + 92, + 483, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "driving scenarios. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 20739-20749, Los Alamitos, CA, USA, 2023. IEEE Computer Society. 3", + "[63] Keenon Werling, Dalton Omens, Jeongseok Lee, Ioannis Exarchos, and C. Karen Liu. Fast and Feature-Complete Differentiable Physics Engine for Articulated Rigid Bodies with Contact Constraints. In Robotics: Science and Systems XVII, Virtual Event, July 12-16, 2021, Virtual, 2021. RSS Foundation. 2, 5, 7", + "[64] Keenon Werling, Nicholas A Bianco, Michael Raitor, Jon Stingel, Jennifer L Hicks, Steven H Collins, Scott L Delp, and C Karen Liu. Addbiomechanics: Automating model scaling, inverse kinematics, and inverse dynamics from human motion data through sequential optimization. Plos one, 18(11):e0295152, 2023. 2, 5", + "[65] Keenon Werling, Janelle Kaneda, Tian Tan, Rishi Agarwal, Six Skov, Tom Van Wouwe, Scott Uhlrich, Nicholas Bianco, Carmichael Ong, Antoine Falisse, et al. Addbiomechanics dataset: Capturing the physics of human motion at scale. In European Conference on Computer Vision, pages 490-508. Springer, 2024. 2, 3", + "[66] Lahiru N Wimalasena, Jonas F Braun, Mohammad Reza Keshtkaran, David Hofmann, Juan Álvaro Gallego, Cristiano Alessandro, Matthew C Tresch, Lee E Miller, and Chethan Pandarinath. Estimating muscle activation from emg using deep learning-based dynamical systems models. Journal of neural engineering, 19(3):036013, 2022. 3", + "[67] Jungdam Won, Deepak Gopinath, and Jessica Hodgins. Control strategies for physically simulated characters performing two-player competitive sports. ACM Transactions on Graphics (TOG), 40(4):1-11, 2021. 3", + "[68] Jungdam Won, Deepak Gopinath, and Jessica Hodgins. Physics-based character controllers using conditional vaes. ACM Transactions on Graphics (TOG), 41(4):1-12, 2022. 3", + "[69] Baoping Xiong, Nianyin Zeng, Han Li, Yuan Yang, Yurong Li, Meilan Huang, Wuxiang Shi, Min Du, and Yudong Zhang. Intelligent prediction of human lower extremity joint moment: an artificial neural network approach. IEEE Access, 7:29973-29980, 2019. 3", + "[70] Aya Yaacoub, Vincent Thomas, Francis Colas, and Pauline Maurice. A probabilistic model for cobot decision making to mitigate human fatigue in repetitive co-manipulation tasks. IEEE Robotics and Automation Letters, 2023. 1", + "[71] Katsu Yamane, Akihiko Murai, Sadahiro Takaya, and Yoshihiko Nakamura. Muscle tension database for contact-free estimation of human somatosensory information. In 2009 IEEE International Conference on Robotics and Automation, pages 633-638. IEEE, 2009. 3", + "[72] X. Yi, Y. Zhou, M. Habermann, S. Shimada, V. Golyanik, C. Theobalt, and F. Xu. Physical inertial poser (pip): Physics-aware real-time human motion tracking from sparse inertial sensors. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13157-13168, Los Alamitos, CA, USA, 2022. IEEE Computer Society. 3", + "[73] Y. Yuan and K. Kitani. Ego-posed estimation and forecasting as real-time pd control. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "27792", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "10081-10091, Los Alamitos, CA, USA, 2019. IEEE Computer Society. 3", + "[74] Felix E Zajac. Muscle and tendon: properties, models, scaling, and application to biomechanics and motor control. Critical reviews in biomedical engineering, 17(4):359-411, 1989. 2", + "[75] Petrissa Zell and Bodo Rosenhahn. A physics-based statistical model for human gait analysis. In Pattern Recognition, pages 169-180, Cham, 2015. Springer International Publishing. 3", + "[76] P. Zell and B. Rosenhahn. Learning-based inverse dynamics of human motion. In 2017 IEEE International Conference on Computer Vision Workshop (ICCVW), pages 842–850, Los Alamitos, CA, USA, 2017. IEEE Computer Society. 3", + "[77] P. Zell, B. Wandt, and B. Rosenhahn. Joint 3d human motion capture and physical analysis from monocular videos. In 2017 IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 17-26, Los Alamitos, CA, USA, 2017. IEEE Computer Society. 3", + "[78] Petrissa Zell, Bodo Rosenhahn, and Bastian Wandt. Weakly-supervised learning of human dynamics. In Computer Vision - ECCV 2020, pages 68-84, Cham, 2020. Springer International Publishing. 3", + "[79] Longbin Zhang, Davit Soselia, Ruoli Wang, and Elena M Gutierrez-Farewik. Estimation of joint torque by emg-driven neuromusculoskeletal models and lstm networks. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 2023. 3", + "[80] Yufei Zhang, Jeffrey O Kephart, and Qiang Ji. Incorporating physics principles for precise human motion prediction. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 6164–6174, 2024. 3", + "[81] Chenhui Zuo, Kaibo He, Jing Shao, and Yanan Sui. Self model for embodied intelligence: Modeling full-body human musculoskeletal system and locomotion control with hierarchical low-dimensional representation. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 13062-13069. IEEE, 2024. 3" + ], + "bbox": [ + 91, + 90, + 482, + 628 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "27793", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/Homogeneous Dynamics Space for Heterogeneous Humans/a999c957-db24-46d0-83c3-e6da7809465e_model.json b/2025/Homogeneous Dynamics Space for Heterogeneous Humans/a999c957-db24-46d0-83c3-e6da7809465e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..577c946b237c90be757fded71c95573da26ab6f6 --- /dev/null +++ b/2025/Homogeneous Dynamics Space for Heterogeneous Humans/a999c957-db24-46d0-83c3-e6da7809465e_model.json @@ -0,0 +1,2677 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.202, + 0.131, + 0.798, + 0.154 + ], + "angle": 0, + "content": "Homogeneous Dynamics Space for Heterogeneous Humans" + }, + { + "type": "text", + "bbox": [ + 0.12, + 0.18, + 0.884, + 0.201 + ], + "angle": 0, + "content": "Xinpeng Liu\\(^{1,2}\\), Junxuan Liang\\(^{1}\\), Chenshuo Zhang\\(^{1}\\), Zixuan Cai\\(^{3}\\), Cewu Lu\\(^{1,2*}\\), Yong-Lu Li\\(^{1,2*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.134, + 0.201, + 0.864, + 0.253 + ], + "angle": 0, + "content": "1Shanghai Jiao Tong University, 2Shanghai Innovation Institute, 3Soochow University xinpengliu0907@gmail.com, zxcai@stu.suda.edu.cn, {whitefork, zhangchenshuo, lucewu, yonglu.li}@sjtu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.287, + 0.327, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.319, + 0.485, + 0.668 + ], + "angle": 0, + "content": "Analyses of human motion kinematics have achieved tremendous advances. However, the production mechanism, known as human dynamics, is still undercovered. In this paper, we aim to push the understanding of data-driven human dynamics forward. We identify a major obstacle to this as the heterogeneity of existing human motion understanding efforts. Specifically, heterogeneity exists in not only the diverse kinematics representations and hierarchical dynamics representations but also the data from different domains, namely biomechanics and reinforcement learning. With an in-depth analysis of the existing heterogeneity, we propose to emphasize the beneath homogeneity: all of them represent the homogeneous fact of human motion, though from different perspectives. Given this, we propose Homogeneous Dynamics Space (HDyS) as a fundamental space for human dynamics by aggregating heterogeneous data and training a homogeneous latent space with inspiration from the inverse-forward dynamics procedure. HDyS achieves decent mapping between human kinematics and dynamics by leveraging the heterogeneous representations and datasets. We demonstrate the feasibility of HDyS with extensive experiments and applications. The project page is https://foruck.github.io/HDyS." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.698, + 0.222, + 0.713 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.723, + 0.483, + 0.86 + ], + "angle": 0, + "content": "Analyses on human motion have a wide range of applications, including animation [11, 60], healthcare [10, 55], and robotics [57, 70]. The computer vision community has made tremendous progress in understanding human kinematics with tasks like human reconstruction [24, 27, 34], action recognition [25, 26, 28, 29, 51], and motion generation [30, 32, 60]. However, the production mechanism hidden beneath human motion, known as human dynamics, is still limitedly explored." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.861, + 0.483, + 0.876 + ], + "angle": 0, + "content": "In this paper, we aim to push the understanding of human" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.287, + 0.903, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.473, + 0.907, + 0.502 + ], + "angle": 0, + "content": "Figure 1. The representation heterogeneity exists for both kinematics and dynamics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.521, + 0.907, + 0.597 + ], + "angle": 0, + "content": "dynamics forward with data-driven methodologies. Specifically, we pursue to build a bidirectional mapping between human kinematics and dynamics. We identify the major obstacle to this as the heterogeneity of existing human motion understanding efforts, which could be two-fold." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.6, + 0.909, + 0.903 + ], + "angle": 0, + "content": "First, representation heterogeneity exists for both kinematics and dynamics, as shown in Figure 1. A typical optical motion capture pipeline, which produces most existing available data for human kinematics [20, 40], involves tracking the markers placed on the surface of humans. Then, different human template models are fitted to the tracked markers, resulting in parameters of musculoskeletal models [52] for biomechanics uses or SMPL-like models [36] for general CV/CG uses. Skeletons with pre-defined kinematic trees are sometimes preferred as a simple and intuitive representation. Different datasets tend to include only representations preferred by their designed goal. Besides the kinematics representation heterogeneity, the dynamics representation heterogeneity exists more severely. Human dynamics function in a hierarchical manner. The closest dynamics hierarchy to the kinematics is the joint torques, which are further produced by muscle forces from muscle actions according to the muscle contraction dynamics. Furthermore, electric signals from the neural system activate the muscle, which surface electromyogra" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.242, + 0.9 + ], + "angle": 0, + "content": "*Corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "27782" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.212 + ], + "angle": 0, + "content": "phy (sEMG) could detect. Effective conversions between these different representations are not always available. For example, for kinematics representations, the conversion between the musculoskeletal model and SMPL becomes available in very recent advances [20], which requires time-consuming optimization. For dynamics, the bi-directional mapping between joint torques, muscle actions, and electric signals is even more difficult." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.216, + 0.485, + 0.608 + ], + "angle": 0, + "content": "Second, domain heterogeneity also exists. One crucial characteristic of human dynamics is the difficulty of non-intrusive measurement. In the biomechanics community, optimization-based methods are adopted to solve the Euler-Newton equation for joint torques [64] and muscle actions [53]. For electric signals, it is typically measured by sEMG sensors [3]. In the meantime, from the learning community, there emerge efforts [33] utilizing Reinforcement Learning (RL) to imitate existing motion captures with physically simulated humanoids and record the simulated human dynamics data in a fully synthetic manner. These different data sources introduce domain heterogeneity in three aspects. First, biomechanics data are more deeply rooted in reality with well-defined computation models and real human motions. Instead, RL data could diverge from real-world situations with unnatural motion imitations and inaccurate physics simulations, known as the sim2real domain gap. Second, biomechanics data, due to the optimization-based manner, is sensitive to variations and usually limited to strictly controlled laboratory setups with relatively simple motions, resulting in limited motion domain coverage. In contrast, RL data cover a wide range of motion with more diversity. Third, biomechanics data and RL data tend to adopt heterogeneous kinematics representations with different kinematic trees, making it hard to transfer directly from one domain to another." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.611, + 0.483, + 0.808 + ], + "angle": 0, + "content": "Despite the heterogeneity, we emphasize the homogeneous essence behind it: all of them represent the homogeneous fact of human motion, each from a different perspective. Given this, we demonstrate the feasibility of unifying these heterogeneous human representations and exploiting the underlying homogeneous knowledge of human dynamics. Low-level Cartesian kinematics representations, like markers and joints, are less heterogeneous than representations like joint angles. Also, though joint torques, muscle actions, and electric signals are not directly transferable to each other, they could share similar motor knowledge like coordination. To this end, we propose Homogeneous Dynamics Space for heterogeneous humans (HDyS)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.902 + ], + "angle": 0, + "content": "To achieve this, we aggregate human dynamics data from both RL [33] and biomechanics [3, 53, 65] with different kinematics and dynamics representations, covering human dynamics hierarchies, including joint torques, muscle actions, and electric signals. Then, HDyS is designed as an aggregation of multiple auto-encoders corresponding to the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.317 + ], + "angle": 0, + "content": "inverse-forward dynamics procedure. We supervise HDyS with reconstruction and alignment losses. Our proposed HDyS has two major merits. First, by unifying heterogeneous human representation in the same latent space, it generalizes across different representations seamlessly while taking advantage of each. Second, by aggregating largescale heterogeneous motion data, it is empowered with general human motor knowledge from a wide span of motion, functioning as a reusable knowledge source for downstream dynamics-related applications. To demonstrate the efficacy of HDyS, we first evaluate it on human inverse dynamics. Then, we showcase how HDyS could facilitate downstream dynamics-related applications like ground reaction force prediction, biomechanics human simulation, and physics-simulated character control." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.318, + 0.907, + 0.439 + ], + "angle": 0, + "content": "To summarize, our contribution includes: 1) We analyzed the heterogeneity issue that hinders an in-depth understanding of human dynamics. 2) We highlighted the homogeneity beneath the heterogeneity and proposed a fundamental reusable space HDyS for human dynamics by unifying the heterogeneity. 3) We demonstrated the feasibility of digging homogeneity out from heterogeneity with extensive experiments and applications of HDyS." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.452, + 0.665, + 0.467 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.476, + 0.689, + 0.492 + ], + "angle": 0, + "content": "2.1. Human Dynamics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.5, + 0.907, + 0.697 + ], + "angle": 0, + "content": "By human dynamics, we mean the production mechanism of human motion, which the biomechanics community has actively explored. To produce a certain motion, neural commands are sent to activate the muscles. After receiving the activation signals, muscles contract and produce muscle forces. Multiple muscle forces form the joint torques according to certain musculoskeletal geometry, and the joint torques drive the accelerations that accumulate into movements. Thus, understanding human dynamics typically involves two heterogeneous hierarchies: joint torques and muscle activations. However, both are hard to measure nonintrusively. In the literature, to obtain them, an optimization problem is typically introduced as" + }, + { + "type": "equation", + "bbox": [ + 0.537, + 0.705, + 0.905, + 0.74 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\min \\| a \\|, s. t. 0 \\leq a \\leq 1, \\tau = A (q) F (a), \\\\ M (q) \\ddot {q} + C (q, \\dot {q}) + G (q) = J \\lambda + \\tau . \\tag {1} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.902 + ], + "angle": 0, + "content": "with the generalized human inertia matrix \\( M(q) \\) w.r.t. generalized coordinate \\( q \\), Coriolis and centrifugal forces \\( C(q,\\dot{q}) \\), gravity \\( G(q) \\), Jacobian matrix \\( J \\) mapping external forces \\( \\lambda \\) to the generalized coordinates, and muscle activations \\( a \\). The joint torques \\( \\tau \\) could be obtained by \\( \\tau = A(q)F(a) \\), where \\( A(q) \\) maps muscle forces into joint torques and \\( F(a) \\) maps \\( a \\) into muscle forces usually with the hill-type function [15, 74]. Mature software [4, 5, 63] were developed for this purpose. However, the optimization quality is tightly bonded to the precision of external" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27783" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.394 + ], + "angle": 0, + "content": "force measurement, which could be expensive. Therefore, the applications are mostly limited to simple motions like gaits in laboratory settings. Some efforts exploited wearable devices [22, 23] for more general applications. In addition, the optimization is deeply coupled with the adopted human models [36, 52], which vary with the application preferences. Fitting raw motion capture data to a specific human model could be unstable and time-consuming. The conversion between different human models is also nontrivial even with recent advances [20]. This way, transferring dynamics from one human model to another is limitedly explored. Besides torques and muscle actions that are typically obtained via optimization, Surface Electromyography (sEMG) is adopted as an indirect representation of human dynamics which could be directly measured. Efforts were made to build the accurate mapping from sEMG to muscle actions [19, 66], joint torques [54, 79], and human poses [35, 55]. Though progress was made, sEMG patterns might suffer from noises and vary drastically among subjects [61], hindering the generalization." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.407, + 0.391, + 0.423 + ], + "angle": 0, + "content": "2.2. Learning-based Human Dynamics" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.432, + 0.483, + 0.793 + ], + "angle": 0, + "content": "Joint Torques. Early efforts were made on ML-based joint torque analysis for certain human body parts [17, 43, 69]. Lv et al. [39] developed a Gaussian mixture framework for whole-body joint torque estimation. Other architectures like k-nearest neighbor-based regression [75, 77], random forests [76], and neural networks [78] were also adopted for the estimation of joint torques. However, most of these efforts suffered from limited data scale, which hampered learning methods from exploiting their full potential. The recent emergence of AddBiomechanics [65], which aggregated multiple biomechanics datasets, considerably boosted the data scale. However, most of the collected sequences contained only gaits with limited diversity. Another line of work adopted reinforcement learning to simulate motion in physics simulators, and the joint torques could be obtained in simulation. Though generalizability was limited at first [1, 49, 50, 67, 68], emerging efforts [37, 38, 73] managed to replicate a wide span of motion in simulators. The paradigm is further incorporated into recent MoCap systems for simultaneous estimation of motion and the joint torques [8, 12, 16, 62, 72, 80]. However, due to the involvement of simulators, the sim2real gap exists. The learned dynamics could be restricted to certain simulators or human models, which might diverge from real humans." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.482, + 0.901 + ], + "angle": 0, + "content": "Muscle Actions. sEMG is usually adopted as a proxy measurement of muscle actions and has been extensively studied in biomechanics [3, 6, 7, 14, 41, 44, 48]. There have been efforts delving into predicting sEMG signals given either joint torques [31, 56, 58], goniometers [59], motion captures [18, 45, 71], videos [3, 48], or point clouds [46]. However, each effort could be small in data scale, mo" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.243 + ], + "angle": 0, + "content": "tion variations, and muscle coverage. Unifying these efforts could be hindered by the heterogeneous settings and data formats adopted, thus under-covered. Recently, musculoskeletal human simulation has gained attention. Multiple efficient simulators were developed [2, 9, 81], opening the potential for more accessible muscle action analysis. Specifically, MinT [53] was proposed by simulating motion sequences from AMASS [40] in OpenSim [5] and attaching simulated muscle actions to the raw motion sequence, enabling muscle action learning in scale." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.26, + 0.605, + 0.275 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.287, + 0.905, + 0.348 + ], + "angle": 0, + "content": "We introduce the proposed Homogeneous Dynamics Space (HDyS). We first cover the involved kinematics and dynamics representations in Section 3.1, 3.2. Then, the model architecture and designed losses are introduced in Section 3.3." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.36, + 0.766, + 0.375 + ], + "angle": 0, + "content": "3.1. Kinematics Representations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.384, + 0.905, + 0.459 + ], + "angle": 0, + "content": "As shown in Fig. 2, we adopt four types of kinematics representations: Cartesian representations of markers and skeleton key-points, joint angles of Rajagopal's model [52] from the biomechanics community, and SMPL [36] which is widely adopted for CV and CG applications." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.461, + 0.906, + 0.642 + ], + "angle": 0, + "content": "Markers, placed on the human body surfaces, are typically the raw data for optical motion capture. In practice, most other representations are calculated by fitting certain human prior models to the marker observations, making it easy to obtain for heterogeneous datasets and representations. Thus, markers are expected to be a generalizable representation across heterogeneous datasets, though they could also be rather low-level and thus hard to learn. We define the marker representation at timestamp \\( t \\) as \\( x_{m}^{t} = (m^{t},\\dot{m}^{t},\\dot{m})\\in \\mathcal{R}^{N_{m}\\times 9} \\), which is composed of marker Cartesian coordinates \\( m^t \\), finite-differentiated velocities \\( \\dot{m}^t \\) and finite-differentiated accelerations \\( \\dot{m}^t \\) with \\( N_{m} \\) markers." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.906, + 0.78 + ], + "angle": 0, + "content": "Skeletal Keypoints, compared to markers, are less generalizable due to their reliance on pre-defined kinematic trees. However, with its Cartesian coordinates, common sense of human topology, and easy access, the gap between different kinematic trees can be mitigated. We define the joint representation at timestamp \\( t \\) as \\( x_{k}^{t} = (k^{t},\\dot{k}^{t},\\dot{k})\\in \\mathcal{R}^{N_{k}\\times 9} \\), which is composed of joint coordinates \\( k^t \\), finite-differentiated velocities \\( \\dot{k}^t \\) and finite-differentiated accelerations \\( \\dot{k}^t \\) with \\( N_{k} \\) skeletal joints." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Joint Angles are preferred for clinical analyses which more faithfully preserves the biomechanics information of human kinematics. We adopt the Rajagopal's model [52] used in AddBiomechanics [65], and define the joint angle representation at timestamp \\( t \\) as \\( x_{a}^{t} = (a^{t},\\dot{a}^{t},\\dot{a})\\in \\mathcal{R}^{3N_{j}} \\), containing joint angles \\( a^t \\), finite-differentiated velocities \\( \\dot{a}^t \\) and accelerations \\( \\dot{a}^t \\) with \\( N_{j} \\) joints in Rajagopal's model (only 23 lower-body joints are used)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "27784" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.096, + 0.089, + 0.482, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.155, + 0.356, + 0.419, + 0.371 + ], + "angle": 0, + "content": "Figure 2. The overall architecture of HDyS." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.378, + 0.483, + 0.485 + ], + "angle": 0, + "content": "SMPL [36] is a human parameter model widely adopted in the CV/CG community. We define the SMPL representation at timestamp \\( t \\) as \\( x_{s}^{t} = (s^{t},\\dot{s}^{t},\\dot{s})\\in \\mathcal{R}^{3N_{j}} \\), which is composed of SMPL parameters \\( s^t \\), finite-differentiated velocities \\( \\dot{s}^t \\) and finite-differentiated accelerations \\( \\dot{s}^t \\) with \\( N_{j} \\) joints of SMPL (75 used, 3 translational joints at root and 72 revolving joints)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.495, + 0.332, + 0.511 + ], + "angle": 0, + "content": "3.2. Dynamics Representations" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.518, + 0.483, + 0.548 + ], + "angle": 0, + "content": "We adopt three different types of dynamic representations: joint torques, muscle actions, and sEMGs." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.549, + 0.483, + 0.64 + ], + "angle": 0, + "content": "Joint Torques are the net torques exerted at each joint to drive the motion, represented as \\(\\tau_{t} \\in \\mathcal{R}^{N_{t}}\\). \\(N_{t}\\) is coupled with the number of joints \\(N_{j}\\) as \\(N_{t} = N_{j} - 6\\), with the 6-DoF free root joint unactuated. For clarity, we denote the joint torques corresponding to Rajagopal's model as \\(\\tau_{tr}\\) and those corresponding to SMPL as \\(\\tau_{ts}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.64, + 0.483, + 0.747 + ], + "angle": 0, + "content": "Muscle Actions represent the activation level of each muscle, resulting in the muscle forces that produce joint torques. We denote muscle actions as \\(\\tau_{m} \\in \\mathcal{R}^{N_{m}}\\) with \\(N_{m}\\) muscles. sEMG detects the electric potential generated on the body surface by activated muscle cells, which is closely related to muscle actions. We represent sEMGs at timestamp \\(t\\) as \\(\\tau_{e}^{t} \\in \\mathcal{R}^{N_{e}}\\) with \\(N_{e}\\) sEMG channels." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.756, + 0.367, + 0.773 + ], + "angle": 0, + "content": "3.3. Homogeneous Dynamics Space" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.87 + ], + "angle": 0, + "content": "With the heterogeneous representations, we construct our homogeneous dynamics space (HDyS) as an aggregation of multiple auto-encoders corresponding to the inverseforward dynamics procedure. Then, reconstruction and alignment losses are adopted for supervision. For simplicity, all superscripts \\(t\\) are omitted." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Inverse-Dynamics Auto-Encoder (IDAE) encodes the kinematics into the latent space and decodes the dynamics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.273 + ], + "angle": 0, + "content": "from the latent. For markers \\( x_{m} \\) and skeletal key points \\( x_{k} \\), we adopt three-layer transformer encoders with no positional embedding to encode them into latent \\( z_{m}, z_{k} \\in \\mathcal{R}^{d} \\) of dimension \\( d \\). This enables the encoding of an arbitrary number of markers/points. For joint angles \\( x_{a} \\) and SMPL parameters \\( x_{s} \\), we adopt simple three-layer MLPs to encode them into latent \\( z_{a}, z_{s} \\in \\mathcal{R}^{d} \\). The decoder is designed as a shared transformer followed by separate MLP regression heads. The transformer takes latent vectors from consecutive frames as input, outputting per-frame latent vectors refined with temporal contexts. Then, separate MLP decoders decode the per-frame dynamics as \\( \\hat{\\tau}_{tr}, \\hat{\\tau}_{ts}, \\hat{\\tau}_{m}, \\hat{\\tau}_{e} \\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.278, + 0.907, + 0.599 + ], + "angle": 0, + "content": "Forward-Dynamics Auto-Encoder (FDAE) encodes dynamics and the kinematics except for accelerations (denoted as \\(\\tilde{x}\\)) into the latent space and decodes the accelerations. Similar to IDAE, markers \\(\\tilde{x}_m\\) and skeletal key-points \\(\\tilde{x}_k\\) are encoded by transformers as \\(\\tilde{z}_m,\\tilde{z}_k\\), while joint angles \\(\\tilde{x}_a\\) and SMPL parameters \\(\\tilde{x}_s\\) are encoded by MLPs as \\(\\tilde{z}_a,\\tilde{z}_p\\). Then, joint torques \\(\\tau_{tr},\\tau_{ts}\\), muscle activations \\(\\tau_m\\), and sEMGs \\(\\tau_e\\) are also encoded by separate MLP encoders as dynamics latent vectors \\(\\tilde{z}^{tr},\\tilde{z}^{ts},\\tilde{z}^m,\\tilde{z}^e\\). \\(\\tilde{z}_m,\\tilde{z}_k,\\tilde{z}_a,\\tilde{z}_p\\) is then concatenated with available dynamics latent vectors, fed into a shared MLP composer, resulted in latent vectors \\(z^{\\cdot}\\), with subscripts representing the dynamics. Although arbitrary combinations are feasible, in practice, only one dynamic latent vector typically exists due to the data limitations. Finally, separate MLP decoders decode the latent vectors \\(z_{\\{kin\\}}^{(dyn)}\\) into skeletal key point accelerations \\(\\ddot{k}\\), SMPL accelerations \\(\\ddot{s}\\), and joint angle accelerations \\(\\ddot{a}\\), where \\(kin\\in \\{m,k,s,a\\}\\), \\(dyn\\in \\{m,e,ts,tr\\}\\). The marker accelerations are not predicted since their variable sizes could bring unnecessary complexity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.603, + 0.906, + 0.648 + ], + "angle": 0, + "content": "Loss Terms. To train HDyS, we adopt reconstruction losses and alignment losses. For reconstruction losses, we simply calculate the L1 loss for \\(\\tau\\) and \\(\\ddot{a}\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.668, + 0.907, + 0.687 + ], + "angle": 0, + "content": "\\[\nL _ {r e c o n} = \\| \\tau - \\hat {\\tau} \\| _ {1} + \\| \\ddot {a} - \\hat {\\dot {a}} \\|. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.707, + 0.906, + 0.788 + ], + "angle": 0, + "content": "For alignment losses, we align the latent vectors \\(z\\) of the same frame and separate \\(zs\\) of different frames using InfoNCE [47]. Given a batch of \\(B\\) frames, the obtained latent vectors is denoted as \\(Z = \\{z_{\\{kin\\}}, z_{\\{kin\\}}^{(dyn)}\\}\\), with \\(kin \\in \\{m, k, s, a\\}, dyn \\in \\{m, e, ts, tr\\}\\). The loss is" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.807, + 0.907, + 0.851 + ], + "angle": 0, + "content": "\\[\nL _ {a l i g n} = \\sum_ {z _ {2}, z _ {2} \\in Z} \\sum_ {i = 1} ^ {B} - \\log \\left(\\frac {\\exp \\left(\\left\\langle z _ {1} ^ {(i)} , z _ {2} ^ {(i)} \\right\\rangle\\right)}{\\sum_ {j = 1} ^ {B} \\exp \\left(\\left\\langle z _ {1} ^ {(i)} , z _ {2} ^ {(j)} \\right\\rangle\\right)}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.903 + ], + "angle": 0, + "content": "where \\((i),(j)\\) indicate the batch index. The overall loss is as \\(\\mathcal{L} = \\alpha_{1}L_{recon} + \\alpha_{2}L_{align}\\) with coefficients \\(\\alpha_{1},\\alpha_{2}\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27785" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.117 + ], + "angle": 0, + "content": "Table 1. Quantitative results of HDyS with ablation studies. For HDyS, the results of averaged predictions and the best prediction among all representations are reported." + }, + { + "type": "table", + "bbox": [ + 0.133, + 0.118, + 0.862, + 0.361 + ], + "angle": 0, + "content": "
MethodsImDy mPJE↓ avg/bstAddBiomechanics mPJE↓ avg/bstMinTMiA PCC↑ avg/bst
RMSE↓ avg/bstPCC↑ avg/bstRMSE↓ avg/bst
ImDyS [33]0.63000.1626----
MiA[3]----13.3-
HDyS0.5765/0.46740.1189/0.12430.0614/0.06150.7420/0.740211.8/11.60.7232/0.7261
ImDy-only HDyS0.6854/0.5403-----
AddBiomechanics-only HDyS-0.1695/0.1691----
MinT-only HDyS--0.0637/0.06400.7179/0.7127--
MiA-only HDyS----13.6/13.50.6557/0.6421
HDyS w/o ImDy-0.1214/0.13860.0608/0.06170.7470/0.741715.8/15.60.6523/0.6497
HDyS w/o AddBiomechanics0.5787/0.4742-0.0616/0.06140.7408/0.738617.1/17.00.5769/0.5638
HDyS w/o MinT0.5730/0.46810.1197/0.1296--16.9/16.80.5213/0.5313
HDyS w/o MiA0.5890/0.47880.1200/0.12840.0616/0.06180.7395/0.7375--
HDyS w/o AMASS0.5797/0.47860.1217/0.13190.0613/0.06170.7419/0.738014.7/14.90.5704/0.5632
HDyS w/o Lalign0.6575/0.50190.1270/0.13290.0626/0.06300.7318/0.723813.7/13.40.6464/0.6402
HDyS w/o FDAE0.5776/0.48490.1198/0.12610.0617/0.06170.7388/0.737513.6/13.30.6517/0.6699
HDyS-32D0.7390/0.64500.1401/0.14200.0650/0.06480.6980/0.701016.7/16.30.5524/0.5614
HDyS-64D0.6505/0.54100.1295/0.13540.0627/0.06290.7272/0.725415.1/14.50.6034/0.6187
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.365, + 0.224, + 0.383 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.398, + 0.194, + 0.413 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.429, + 0.484, + 0.52 + ], + "angle": 0, + "content": "AddBiomechanics [64] contains over 50 hours of human motion data with joint torques from Nimble [63] simulation. We follow the setting in [64] with the armless part of AddBiomechanics. Markers, key points, and joint angles are used as kinematics representations, and joint torques corresponding to [52] represent the dynamics." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.528, + 0.484, + 0.62 + ], + "angle": 0, + "content": "Muscles in Times (MinT) [53] simulates part of AMASS in OpenSim to obtain the actions of 402 muscles. We randomly split them into 906 sequences for training and 227 sequences for testing. Markers, key points, and SMPL parameters are used as kinematics representations, and muscle actions are the dynamics representations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.628, + 0.484, + 0.719 + ], + "angle": 0, + "content": "Muscles in Act (MiA) [3] consists of 12.8 hours of human exercise motion reconstructed from videos with VIBE [21] and sEMG data of 8 muscles. Following [3], we split them into 19,563 training sequences and 3,053 testing sequences. Markers and key points are kinematics representations, and sEMGs are dynamics representations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.726, + 0.484, + 0.817 + ], + "angle": 0, + "content": "ImDy [33] adopted PHC [38] to imitate motion sequences from AMASS [40] and KIT [42], resulting in over 150 hours of human dynamics data, with 27,501 sequences for training and 3,055 sequences for evaluation. Markers, key points, and SMPL are kinematics representations, and SMPL torques are dynamics representations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.902 + ], + "angle": 0, + "content": "AMASS [40] contains over 11,000 motion sequences represented in SMPL [36]. Though not paired with human dynamics, we included it as an extra knowledge base for training. Markers, key points, and SMPL parameters are used as kinematics representations." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.366, + 0.731, + 0.382 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.389, + 0.907, + 0.707 + ], + "angle": 0, + "content": "We adopt a compact design for HDyS with a latent dimension of 128. For joint angle and SMPL parameters, three-layer MLPs with hidden unit sizes of 256 and 128 are adopted as encoders, projecting the input to the 128-D latent space. For markers and key-points, three-layer transformer encoders with 2 heads are adopted as encoders. The transformer in the ID decoder is of 4 layers, 4 heads, and a dimension of 128. All MLPs in decoders are two-layer, with a hidden size of 32 for Rajagopal's torque, sEMG, joint angle accelerations, and key-point accelerations, and a hidden size of 64 for SMPL torque, SMPL accelerations, and muscle actions. For training, we adopt an AdamW optimizer with a learning rate of 1e-3 and a batch size of 9,600 frames for 1,000 epochs. Loss weights are set as \\(\\alpha_{1} = 0.01\\), \\(\\alpha_{2} = 0.05\\). Also, during training, a balanced sampling strategy is adopted to minimize the scale influence of different datasets. That is, we randomly sample 3,000 sequences per dataset for each training epoch. In this way, the comparison is fair in terms of the number of seen samples under different dataset settings. All experiments are conducted on a single NVIDIA Titan Xp GPU." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.715, + 0.771, + 0.731 + ], + "angle": 0, + "content": "4.3. Results on Inverse Dynamics" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.738, + 0.782, + 0.754 + ], + "angle": 0, + "content": "Metric. We report mPJE and RMSE as" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.764, + 0.907, + 0.807 + ], + "angle": 0, + "content": "\\[\nm P J E _ {\\tau .} = \\frac {1}{J} \\sum_ {j = 1} ^ {J} | \\tau . - \\hat {\\tau}. | _ {2}, R M S E _ {\\tau .} = | \\tau . - \\hat {\\tau}. | _ {2}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.816, + 0.906, + 0.862 + ], + "angle": 0, + "content": "PCC is also reported due to its invariance against scale and offset, which helps evaluate muscle actions and sEMG prediction. PCC is computed as" + }, + { + "type": "equation", + "bbox": [ + 0.634, + 0.871, + 0.907, + 0.904 + ], + "angle": 0, + "content": "\\[\nP C C _ {\\tau_ {1}} = \\frac {\\operatorname {c o v} \\left(\\tau , \\hat {\\tau}\\right)}{\\sigma_ {\\tau_ {1}} \\sigma_ {\\hat {\\tau}}} \\tag {5}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "27786" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.485, + 0.152 + ], + "angle": 0, + "content": "where \\(cov(\\tau, \\hat{\\tau})\\) is the covariance of \\(\\tau\\), \\(\\hat{\\tau}\\), and \\(\\sigma\\). indicates the standard deviation. For ImDy and AddBiomechanics, we report mPJE normalized by body weight. For MinT and MiA, RMSE and PCC are reported." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.178, + 0.284, + 0.193 + ], + "angle": 0, + "content": "4.3.1 Quantitative results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.204, + 0.484, + 0.31 + ], + "angle": 0, + "content": "Quantitative results are shown in Tab. 1, where HDyS is reported with the averaged predictions from different input kinematics representations and the best predictions among all representations. Compared to previous methods [3, 33], the proposed HDyS provides superior performances with substantial improvements on all datasets. We further analyze the performance of HDyS with three questions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.311, + 0.484, + 0.325 + ], + "angle": 0, + "content": "How do the heterogeneous datasets contribute to HDyS?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.326, + 0.484, + 0.583 + ], + "angle": 0, + "content": "We compare HDyS with its single-dataset variants and drop-one-out variants in Tab. 1. HDyS outperforms the corresponding single-dataset variants on all datasets, validating the existence of homogeneous human dynamics knowledge behind these heterogeneous datasets. According to the drop-one-out results, it is noticeable that datasets with similar dynamics representations are more cooperative. That is, muscle-related datasets (MiA and MinT) tend to benefit more from each other and less from torque-related datasets (AddBiomechanics and ImDy), and vice versa. Moreover, mutually harmful effects could be observed for ImDy and MinT. These are consistent with the gaps between these datasets: the sim2real gap from ImDy to others and the torque-to-muscle gap from AddBiomechanics and ImDy to MiA and MinT. AMASS, though not paired with dynamics information, is shown to be beneficial for inverse dynamics with its diverse and high-quality kinematics." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.585, + 0.484, + 0.902 + ], + "angle": 0, + "content": "The heterogeneous datasets introduce both increased data scale and heterogeneous knowledge. Given this, we try to investigate further the source of the improvement in Tab. 1. Trying to decompose the contributions of scale and heterogeneity, we compared the performance of the following three models on a single target test set in Tab. 2. HDyS-Single-50 denotes single-dataset HDyS with \\(50\\%\\) of the data from the target dataset. HDyS-50/50 denotes HDyS using \\(50\\%\\) of the data from the target dataset and \\(50\\%\\) of the data from the other datasets, maintaining the same data scale as the target dataset. HDyS-Single represents the corresponding single-dataset variants of HDyS in Tab. 1. The performance of the best representation is reported on AddBiomechanics and MiA, considering their suitable scale and realistic nature. As shown, HDyS-50/50 consistently outperforms HDyS-Single-50, verifying the existence of homogeneous knowledge in heterogeneous data. Moreover, on AddBiomechanics, HDyS-50/50 even outperforms HDyS-Single, indicating that the diversity from heterogeneous data could sometimes benefit more than simply increasing seemingly homogeneous data." + }, + { + "type": "table_caption", + "bbox": [ + 0.545, + 0.09, + 0.874, + 0.103 + ], + "angle": 0, + "content": "Table 2. Results on scale-heterogeneity decomposition." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.104, + 0.905, + 0.146 + ], + "angle": 0, + "content": "
DatasetHDyS-Single-50HDyS-50/50HDyS-Single
AddBiomechanics mPJE ↓0.17070.12840.1695
MiA RMSE ↓16.214.513.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.52, + 0.149, + 0.902, + 0.163 + ], + "angle": 0, + "content": "Table 3. Results of different kinematic representations in HDyS." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.164, + 0.905, + 0.235 + ], + "angle": 0, + "content": "
MethodsImDy mPJE↓AddBiomechanics mPJE↓MinTMiA
RMSE↓PCC↑RMSE↓PCC↑
HDys-marker0.81630.14550.06460.696812.30.7088
HDys-keypoint0.80840.13240.06400.710311.60.7261
HDys-angle-0.1243----
HDys-SMPL0.4674-0.06150.7402--
" + }, + { + "type": "image_caption", + "bbox": [ + 0.564, + 0.235, + 0.744, + 0.246 + ], + "angle": 0, + "content": "Heterogeneous Representations" + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.247, + 0.732, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.758, + 0.235, + 0.897, + 0.246 + ], + "angle": 0, + "content": "Heterogeneous Datasets" + }, + { + "type": "image", + "bbox": [ + 0.739, + 0.247, + 0.903, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.369, + 0.732, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.739, + 0.369, + 0.903, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.528, + 0.476, + 0.89, + 0.491 + ], + "angle": 0, + "content": "Figure 3. Feature visualization for HDyS w/ and w/o \\( L_{align} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.507, + 0.905, + 0.627 + ], + "angle": 0, + "content": "How do the heterogeneous representations contribute to HDyS? We report the results from individual kinematics representations in Tab. 3. Typically, the representations in joint space like SMPL and joint angles produce more accurate predictions compared to Cartesian representations with considerable performance gaps, which explains the performance drop when taking the average for ImDy and MiA. The markers are the worst-performed representation of all." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.906, + 0.764 + ], + "angle": 0, + "content": "In Tab. 1, it is noticeable that MinT and AddBiomechanics take advantage of averaging predictions from different representations. For AddBiomechanics, this might be because the best-performing representation is joint angles, which are not available in other datasets. Therefore, the averaging could effectively aggregate knowledge from other datasets with other representations. For MinT, it is observable that different representations produce close results, preventing the ensemble from being held back by markers." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Furthermore, we examine the alignment of heterogeneous representations by removing the alignment loss \\( L_{\\text{align}} \\) in Tab. 1. Considerable performance drops could be observed. The drop is consistent for both best and average performances, indicating the cross-modality alignment manages to aggregate knowledge from heterogeneous representations. We also visualize HDyS features w/ and w/o \\( L_{\\text{align}} \\) in Fig. 3. As shown, \\( L_{\\text{align}} \\) prevents the HDyS feature from being dominated by the representation hetero-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27787" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.088, + 0.482, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.136, + 0.291, + 0.438, + 0.307 + ], + "angle": 0, + "content": "Figure 4. Inverse dynamics visualization on ImDy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.314, + 0.483, + 0.375 + ], + "angle": 0, + "content": "geneity, encouraging the learning of homogeneous knowledge. Interestingly, though not intended to do so, \\( L_{\\text{align}} \\) also helps to mitigate the gap between datasets, alleviating the domain heterogeneity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.375, + 0.484, + 0.512 + ], + "angle": 0, + "content": "How does the model design contribute to HDyS? We first evaluate the FDAE in Tab. 1. A substantial performance drop could be observed, demonstrating the benefit of informing HDyS with the forward dynamics procedure. Moreover, we also report HDyS with different dimensions in Tab. 1. Though HDyS's performance degrades with lower dimensions, it could still outperform some single-dataset variants with higher dimensions, validating the efficacy of heterogeneous datasets again." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.532, + 0.273, + 0.546 + ], + "angle": 0, + "content": "4.3.2 Qualitative results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.556, + 0.483, + 0.737 + ], + "angle": 0, + "content": "We visualize the inverse dynamics results of different datasets. As shown in Fig. 4-5, the joint torques could be faithfully reconstructed for different actions from either synthetic and jittering ImDy or realistic AddBiomechanics with the same HDyS. For the muscle actions shown in Fig. 6, HDyS produces reasonable estimations for lower-body muscles when walking. Moreover, as in the right half, though the lower-body kinematics are less significant, HDyS could capture the ankle muscle actions with high fidelity. The predictions are also coherent with GT for upper-body muscles like the internal oblique and levator scapulae. More details are in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.748, + 0.357, + 0.763 + ], + "angle": 0, + "content": "4.4. Results on Downstream Tasks" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.771, + 0.388, + 0.785 + ], + "angle": 0, + "content": "4.4.1 Ground Reaction Force Estimation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We evaluate the effectiveness of HDyS on GRF estimation with GroundLink [13], which contains 1.5-hour motion with GRF recordings. We finetune HDyS with subjects 1-6 and evaluate it on subject 7. mPJE for GRF at both feet normalized by body weight is reported following Eq. 4. GroundLinkNet [13] and HDyS trained on GroundLink from scratch are also compared. As shown in Tab. 4, the" + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.091, + 0.885, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.288, + 0.898, + 0.312 + ], + "angle": 0, + "content": "Figure 5. Inverse dynamics visualization on AddBiomechanics. Table 4. Results on GroundLink." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.318, + 0.905, + 0.365 + ], + "angle": 0, + "content": "
MethodsGroundLinkNetGroundLink-onlyHDySHDyS
L-Foot mPJE ↓0.07110.05840.0514
R-Foot mPJE ↓0.09120.07510.0694
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.368, + 0.905, + 0.395 + ], + "angle": 0, + "content": "Table 5. Biomechanical human simulation results reported in per-frame MSE." + }, + { + "type": "table", + "bbox": [ + 0.536, + 0.401, + 0.885, + 0.449 + ], + "angle": 0, + "content": "
k12345
Optimized0.000630.19091.83062.01062.6027
HDyS0.000610.18601.71181.86512.2233
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.462, + 0.906, + 0.537 + ], + "angle": 0, + "content": "finetuned HDyS outperforms its counterparts, indicating the efficacy of the aggregated homogeneous knowledge. Also, the GroundLink-only HDyS outperforms GroundLinkNet, reflecting the feasibility of unifying heterogeneous representations for better dynamics knowledge." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.559, + 0.807, + 0.572 + ], + "angle": 0, + "content": "4.4.2 Biomechanical Human Simulation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.906, + 0.718 + ], + "angle": 0, + "content": "HDyS could also be adopted for biomechanical human simulation. We start with the armless Rajagopal's model [52] in Nimble [63]. Given a motion sequence, we first adopt HDyS to estimate the joint torques. Then, we use the predicted torques to reproduce the motion in Nimble. Starting from the current state, we feed the predicted torques for \\( k \\) steps and compare the simulated joint angles \\( \\hat{q} \\) with the real joint angles \\( q \\). The simulation is performed at 90FPS. We report the per-frame MSE of joint angles." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.719, + 0.906, + 0.826 + ], + "angle": 0, + "content": "Results. We demonstrated the results in Tab. 5 and Fig. 7. The results with optimized torques are also reported as a reference. Surprisingly, for different \\( k \\), simulation with HDyS is superior with better stability. However, with \\( k \\) increasing, MSE also increases noticeably, indicating the accumulation of drifting errors. Further enhancing the forward-dynamics compatibility of HDyS would be a promising goal." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.846, + 0.76, + 0.861 + ], + "angle": 0, + "content": "4.4.3 Physical Character Control" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "We adopt HDyS for physical character control following the setting of PHC [38], with 140 testing sequences and the rest" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "27788" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.135, + 0.089, + 0.428, + 0.131 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.437, + 0.089, + 0.561, + 0.127 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.575, + 0.09, + 0.848, + 0.13 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.137, + 0.133, + 0.425, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.427, + 0.133, + 0.86, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.137, + 0.206, + 0.425, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.427, + 0.206, + 0.859, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.137, + 0.245, + 0.425, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.427, + 0.245, + 0.859, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.137, + 0.283, + 0.425, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.427, + 0.283, + 0.859, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.137, + 0.321, + 0.425, + 0.357 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.427, + 0.321, + 0.859, + 0.357 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.366, + 0.358, + 0.63, + 0.371 + ], + "angle": 0, + "content": "Figure 6. Inverse dynamics results on MinT." + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.386, + 0.46, + 0.688 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.116, + 0.688, + 0.458, + 0.702 + ], + "angle": 0, + "content": "Figure 7. Biomechanical human simulation visualization." + }, + { + "type": "table_caption", + "bbox": [ + 0.145, + 0.719, + 0.428, + 0.733 + ], + "angle": 0, + "content": "Table 6. Results of HDyS for character control." + }, + { + "type": "table", + "bbox": [ + 0.114, + 0.738, + 0.462, + 0.784 + ], + "angle": 0, + "content": "
MethodsmPJPEg↓mPJEL↓Evel↓Eacc↓
Baseline74.51449.02911.56214.243
Baseline + HDyS72.69247.92011.30513.757
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.903 + ], + "angle": 0, + "content": "for training. We train the baseline as a primitive in PHC with a batch size of 700 for 10k steps. Then, the HDyS latents of key points are inserted as extra observations. We report the global and local mPJPE and the errors of velocity and acceleration in Tab. 6. HDyS manages to improve its performance, validating its efficacy." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.384, + 0.628, + 0.399 + ], + "angle": 0, + "content": "5. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.409, + 0.908, + 0.712 + ], + "angle": 0, + "content": "Despite the impressive results of HDyS, it could be improved. First, it is noticeable that HDyS performs better for the lower body than the higher body in Fig. 6. This might be due to the imbalanced focus on lower body dynamics in data adopted by HDyS. For all datasets, data on lower-body dynamics like gaits are dominating. For AddBiomechanics, we only adopted its armless part. Enhancing HDyS with more upper-body dynamics would be helpful. Second, for muscle actions, HDys could omit minor changes, and the magnitudes could sometimes diverge from the real. Mitigating it with more curated models and loss terms is desirable. Third, as a first step toward homogeneous human dynamics learning, HDyS is primarily instantiated with five initial datasets and an intuitive model. Scaling HDyS up with more datasets and more human priors would be a meaningful goal. Finally, we demonstrate the potential of HDyS with some simple applications in Sec. 4.4, while more sophisticated use cases for musculoskeletal human simulation, humanoid control, and human-robot transfer learning would be promising as future works." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.725, + 0.634, + 0.74 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We analyzed the heterogeneity issue existing for human dynamics learning and highlighted the homogeneity beneath it. To fully exploit the homogeneity, we proposed HDyS as a homogeneous human dynamics space. Extensive experiments were conducted to validate the feasibility of digging homogeneity out from heterogeneity for human dynamics with detailed analyses of the contribution of heterogeneous components. We further demonstrated the potential of HDyS for downstream applications. We believe HDyS could shed new light on human dynamics understanding." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27789" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.486, + 0.162 + ], + "angle": 0, + "content": "This work is supported in part by the National Natural Science Foundation of China under Grant No.62306175, CCF-Tencent Rhino-Bird Open Research Fund." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.174, + 0.188, + 0.189 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.199, + 0.484, + 0.254 + ], + "angle": 0, + "content": "[1] Kevin Bergamin, Simon Clavet, Daniel Holden, and James Richard Forbes. Drecon: data-driven responsive control of physics-based characters. ACM Transactions On Graphics (TOG), 38(6):1-11, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.256, + 0.484, + 0.323 + ], + "angle": 0, + "content": "[2] Vittorio Caggiano, Huawei Wang, Guillaume Durandau, Massimo Sartori, and Vikash Kumar. Myosuite: A contact-rich simulation suite for musculoskeletal motor control. In Learning for Dynamics and Control Conference, pages 492-507. PMLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.326, + 0.483, + 0.367 + ], + "angle": 0, + "content": "[3] Mia Chiquier and Carl Vondrick. Muscles in action. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22091-22101, 2023. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.368, + 0.483, + 0.436 + ], + "angle": 0, + "content": "[4] Michael Damsgaard, John Rasmussen, Søren Tørholm Christensen, Egidijus Surma, and Mark De Zee. Analysis of musculoskeletal systems in the anybody modeling system. Simulation Modelling Practice and Theory, 14(8): 1100-1111, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.439, + 0.484, + 0.52 + ], + "angle": 0, + "content": "[5] Scott L Delp, Frank C Anderson, Allison S Arnold, Peter Loan, Ayman Habib, Chand T John, Eran Guendelman, and Darryl G Thelen. Opensim: open-source software to create and analyze dynamic simulations of movement. IEEE transactions on biomedical engineering, 54(11):1940-1950, 2007. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.523, + 0.483, + 0.59 + ], + "angle": 0, + "content": "[6] Benedikt Feldotto, Cristian Soare, Alois Knoll, Piyanee Sriya, Sarah Astill, Marc de Kamps, and Samit Chakrabarty. Evaluating muscle synergies with emg data and physics simulation in the neurorobotics platform. Frontiers in Neurorobotics, 16:856797, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.593, + 0.483, + 0.647 + ], + "angle": 0, + "content": "[7] Mariusz P Furmanek, Madhur Mangalam, Mathew Yarossi, Kyle Lockwood, and Eugene Tunik. A kinematic and emg dataset of online adjustment of reach-to-grasp movements to visual perturbations. Scientific data, 9(1):23, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.649, + 0.483, + 0.732 + ], + "angle": 0, + "content": "[8] E. Gartner, M. Andriluka, E. Coumans, and C. Sminchisescu. Differentiable dynamics for articulated 3d human motion reconstruction. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13180-13190, Los Alamitos, CA, USA, 2022. IEEE Computer Society. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.734, + 0.483, + 0.76 + ], + "angle": 0, + "content": "[9] Thomas Geijtenbeek. The Hyfydy simulation software, 2021. https://hyfydy.com.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.762, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[10] Daniel FN Gordon, Andreas Christou, Theodoros Stouraitis, Michael Gienger, and Sethu Vijayakumar. Learning personalised human sit-to-stand motion strategies via inverse musculoskeletal optimal control. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 10497–10503. IEEE, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[11] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.199, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.906, + 0.12 + ], + "angle": 0, + "content": "5142-5151, Los Alamitos, CA, USA, 2022. IEEE Computer Society. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.906, + 0.205 + ], + "angle": 0, + "content": "[12] Erik Gartner, Mykhaylo Andriluka, Hongyi Xu, and Cristian Sminchisescu. Trajectory optimization for physics-based reconstruction of 3d human pose from monocular video. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13096-13105, Los Alamitos, CA, USA, 2022. IEEE Computer Society. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.207, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[13] Xingjian Han, Ben Senderling, Stanley To, Deepak Kumar, Emily Whiting, and Jun Saito. Groundlink: A dataset unifying human body movement and ground reaction dynamics. In SIGGRAPH Asia 2023 Conference Papers, New York, NY, USA, 2023. Association for Computing Machinery. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.905, + 0.358 + ], + "angle": 0, + "content": "[14] Oscar G Hernández, Jose M Lopez-Castellanos, Carlos A Jara, Gabriel J Garcia, Andres Ubeda, Vicente Morell-Gimenez, and Francisco Gomez-Donoso. A kinematic, imaging and electromyography dataset for human muscular manipulability index prediction. Scientific Data, 10(1):132, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.416 + ], + "angle": 0, + "content": "[15] Archibald Vivian Hill. The heat of shortening and the dynamic constants of muscle. Proceedings of the Royal Society of London. Series B-Biological Sciences, 126(843):136-195, 1938. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.42, + 0.905, + 0.502 + ], + "angle": 0, + "content": "[16] B. Huang, L. Pan, Y. Yang, J. Ju, and Y. Wang. Neural mo- con: Neural motion control for physically plausible human motion capture. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6407- 6416, Los Alamitos, CA, USA, 2022. IEEE Computer Society. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.504, + 0.905, + 0.572 + ], + "angle": 0, + "content": "[17] Leif Johnson and Dana H. Ballard. Efficient codes for inverse dynamics during walking. In Proceedings of the Twenty-Eighth AAAI Conference on Artificial Intelligence, pages 343-349, Quebec City, Quebec, Canada, 2014. AAAI Press. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.575, + 0.906, + 0.629 + ], + "angle": 0, + "content": "[18] Lise A Johnson and Andrew J Fuglevand. Evaluation of probabilistic methods to predict muscle activity: implications for neuroprosthetics. Journal of neural engineering, 6(5):055008, 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.632, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[19] Kimoon Kang, Kiwon Rhee, and Hyun-Chool Shin. Event detection of muscle activation using an electromyogram. Applied Sciences, 10(16):5593, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.676, + 0.905, + 0.743 + ], + "angle": 0, + "content": "[20] Marilyn Keller, Keenon Werling, Soyong Shin, Scott Delp, Sergi Pujades, C Karen Liu, and Michael J Black. From skin to skeleton: Towards biomechanically accurate 3d digital humans. ACM Transactions on Graphics (TOG), 42(6): 1-12, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.746, + 0.905, + 0.813 + ], + "angle": 0, + "content": "[21] Muhammed Kocabas, Nikos Athanasiou, and Michael J Black. Vibe: Video inference for human body pose and shape estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5253-5263, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[22] Claudia Latella, Naveen Kuppuswamy, Francesco Romano, Silvio Traversaro, and Francesco Nori. Whole-body human inverse dynamics with distributed micro-accelerometers, gyros and force sensing. Sensors, 16(5), 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[23] Claudia Latella, Silvio Traversaro, Diego Ferigo, Yeshasvi Tirupachuri, Lorenzo Rapetti, Francisco Javier An" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "27790" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "drade Chavez, Francesco Nori, and Daniele Pucci. Simultaneous floating-base estimation of human kinematics and joint torques. Sensors, 19(12), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.205 + ], + "angle": 0, + "content": "[24] Jiefeng Li, Chao Xu, Zhicun Chen, Siyuan Bian, Lixin Yang, and Cewu Lu. Hybrik: A hybrid analytical-neural inverse kinematics solution for 3d human pose and shape estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3383-3393, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.276 + ], + "angle": 0, + "content": "[25] Sheng-Wei Li, Zi-Xiang Wei, Wei-Jie Chen, Yi-Hsin Yu, Chih-Yuan Yang, and Jane Yung-jen Hsu. Sa-dvae: Improving zero-shot skeleton-based action recognition by disentangled variational autoencoders. In European Conference on Computer Vision, pages 447–462. Springer, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[26] Yong-Lu Li, Siyuan Zhou, Xijie Huang, Liang Xu, Ze Ma, Hao-Shu Fang, Yanfeng Wang, and Cewu Lu. Transferable interactiveness knowledge for human-object interaction detection. In CVPR, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[27] Yong-Lu Li, Xinpeng Liu, Han Lu, Shiyi Wang, Junqi Liu, Jiefeng Li, and Cewu Lu. Detailed 2d-3d joint representation for human-object interaction. In CVPR, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.377, + 0.482, + 0.432 + ], + "angle": 0, + "content": "[28] Yong-Lu Li, Liang Xu, Xinpeng Liu, Xijie Huang, Yue Xu, Shiyi Wang, Hao-Shu Fang, Ze Ma, Mingyang Chen, and Cewu Lu. Pastanet: Toward human activity knowledge engine. In CVPR, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.482, + 0.489 + ], + "angle": 0, + "content": "[29] Yong-Lu Li, Xinpeng Liu, Xiaogian Wu, Yizhuo Li, Zuoyu Qiu, Liang Xu, Yue Xu, Hao-Shu Fang, and Cewu Lu. Hake: A knowledge engine foundation for human activity understanding. TPAMI, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.482, + 0.558 + ], + "angle": 0, + "content": "[30] Yong-Lu Li, Xiaogian Wu, Xinpeng Liu, Yiming Dou, Yikun Ji, Junyi Zhang, Yixing Li, Jingru Tan, Xudong Lu, and Cewu Lu. From isolated islands to pangea: Unifying semantic space for human action understanding. In CVPR, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.482, + 0.616 + ], + "angle": 0, + "content": "[31] Zhan Li, David Guiraud, and Mitsuhiro Hayashibe. Inverse estimation of multiple muscle activations from joint moment with muscle synergy extraction. IEEE journal of biomedical and health informatics, 19(1):64-73, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.619, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[32] Xinpeng Liu, Haowen Hou, Yanchao Yang, Yong-Lu Li, and Cewu Lu. Revisit human-scene interaction via space occupancy. In European Conference on Computer Vision, pages 1-19. Springer, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.482, + 0.731 + ], + "angle": 0, + "content": "[33] Xinpeng Liu, Junxuan Liang, Zili Lin, Haowen Hou, Yong-Lu Li, and Cewu Lu. Imdy: Human inverse dynamics from imitated observations. In The Thirteenth International Conference on Learning Representations, 2025. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.733, + 0.482, + 0.787 + ], + "angle": 0, + "content": "[34] Xiaoyang Liu, Boran Wen, Xinpeng Liu, Zizheng Zhou, Hongwei Fan, Cewu Lu, Lizhuang Ma, Yulong Chen, and Yong-Lu Li. Interacted object grounding in spatio-temporal human-object interactions. In AAAI, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[35] Yilin Liu, Shijia Zhang, and Mahanth Gowda. Neuropose: 3d hand pose tracking using emg wearables. In Proceedings of the Web Conference 2021, pages 1471-1482, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[36] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. Smpl: a skinned multiperson linear model. In ACM Transactions on Graphics, New York, NY, USA, 2015. Association for Computing Machinery. 1, 3, 4, 5" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[37] Zhengyi Luo, Ryo Hachiuma, Ye Yuan, and Kris Kitani. Dynamics-regulated kinematic policy for egocentric pose estimation. Advances in Neural Information Processing Systems, 34:25019-25032, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.217 + ], + "angle": 0, + "content": "[38] Z. Luo, J. Cao, A. Winkler, K. Kitani, and W. Xu. Perpetual humanoid control for real-time simulated avatars. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 10861-10870, Los Alamitos, CA, USA, 2023. IEEE Computer Society. 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.219, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[39] Xiaolei Lv, Jinxiang Chai, and Shihong Xia. Data-driven inverse dynamics for human motion. ACM Transactions on Graphics (TOG), 35(6):1-12, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.329 + ], + "angle": 0, + "content": "[40] N. Mahmood, N. Ghorbani, N. F. Troje, G. Pons-Moll, and M. Black. Amass: Archive of motion capture as surface shapes. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 5441-5450, Los Alamitos, CA, USA, 2019. IEEE Computer Society. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.331, + 0.905, + 0.399 + ], + "angle": 0, + "content": "[41] Nebojša Malešević, Alexander Olsson, Paulina Sager, Elin Andersson, Christian Cipriani, Marco Controzzi, Anders Björkman, and Christian Antfolk. A database of high-density surface electromyogram signals comprising 65 isometric hand gestures. Scientific Data, 8(1):63, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.401, + 0.905, + 0.467 + ], + "angle": 0, + "content": "[42] Christian Mandery, Ömer Terlemez, Martin Do, Nikolaus Vahrenkamp, and Tamim Asfour. Unifying representations and large-scale whole-body motion databases for studying human motion. IEEE Transactions on Robotics, 32(4):796-809, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.905, + 0.523 + ], + "angle": 0, + "content": "[43] Mykhailo Manukian, Serhii Bahdasariants, and Sergiy Yakovenko. Artificial physics engine for real-time inverse dynamics of arm and hand movement. Plos one, 18(12): e0295750, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.905, + 0.592 + ], + "angle": 0, + "content": "[44] Luís Moreira, Joana Figueiredo, Pedro Fonseca, João P Vilas-Boas, and Cristina P Santos. Lower limb kinematic, kinetic, and emg data from young healthy humans during walking at controlled speeds. Scientific data, 8(1):103, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.596, + 0.905, + 0.662 + ], + "angle": 0, + "content": "[45] Yoshihiko Nakamura, Katsu Yamane, Yusuke Fujita, and Ichiro Suzuki. Somatosensory computation for man-machine interface from motion-capture data and musculoskeletal human model. IEEE Transactions on Robotics, 21(1):58-66, 2005. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.665, + 0.905, + 0.719 + ], + "angle": 0, + "content": "[46] Hui Niu, Takahiro Ito, Damien Desclaux, Ko Ayusawa, Yusuke Yoshiyasu, Ryusuke Sagawa, and Eiichi Yoshida. Estimating muscle activity from the deformation of a sequential 3d point cloud. Journal of Imaging, 8(6):168, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.721, + 0.905, + 0.762 + ], + "angle": 0, + "content": "[47] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.763, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[48] Kunyu Peng, David Schneider, Alina Roitberg, Kailun Yang, Jiaming Zhang, Chen Deng, Kaiyu Zhang, M Saqib Sarfraz, and Rainer Stiefelhagen. Towards video-based activated muscle group estimation in the wild. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 4495-4504, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.847, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[49] Xue Bin Peng, Ze Ma, Pieter Abbeel, Sergey Levine, and Angjoo Kanazawa. Amp: Adversarial motion priors for stylized physics-based character control. ACM Transactions on Graphics (ToG), 40(4):1-20, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27791" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.147 + ], + "angle": 0, + "content": "[50] Xue Bin Peng, Yunrong Guo, Lina Halper, Sergey Levine, and Sanja Fidler. Ase: Large-scale reusable adversarial skill embeddings for physically simulated characters. ACM Transactions On Graphics (TOG), 41(4):1-17, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.484, + 0.232 + ], + "angle": 0, + "content": "[51] A. R. Punnakkal, A. Chandrasekaran, N. Athanasiou, A. Quiros-Ramirez, and M. J. Black. Babel: Bodies, action and behavior with english labels. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 722-731, Los Alamitos, CA, USA, 2021. IEEE Computer Society. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.483, + 0.301 + ], + "angle": 0, + "content": "[52] Apoorva Rajagopal, Christopher L Dembia, Matthew S Demers, Denny D Delp, Jennifer L Hicks, and Scott L Delp. Full-body musculoskeletal model for muscle-driven simulation of human gait. IEEE transactions on biomedical engineering, 63(10):2068-2079, 2016. 1, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.483, + 0.384 + ], + "angle": 0, + "content": "[53] David Schneider, Simon Reiß, Marco Kugler, Alexander Jaus, Kunyu Peng, Susanne Sutschet, Muhammad Saqib Sarfraz, Sven Matthiesen, and Rainer Stiefelhagen. Muscles in time: Learning to understand human motion in-depth by simulating muscle activations. Advances in Neural Information Processing Systems, 2025. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.385, + 0.483, + 0.453 + ], + "angle": 0, + "content": "[54] Robert V Schulte, Marijke Zondag, Jaap H Buurke, and Erik C Prinsen. Multi-day emg-based knee joint torque estimation using hybrid neuromusculoskeletal modelling and convolutional neural networks. Frontiers in Robotics and AI, 9:869476, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.455, + 0.483, + 0.509 + ], + "angle": 0, + "content": "[55] Paniz Sedighi, Xingyu Li, and Mahdi Tavakoli. Emg-based intention detection using deep learning for shared control in upper-limb assistive exoskeletons. IEEE Robotics and Automation Letters, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.483, + 0.566 + ], + "angle": 0, + "content": "[56] Masashi Sekiya, Sho Sakaino, and Tsuji Toshiaki. Linear logistic regression for estimation of lower limb muscle activations. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 27(3):523-532, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.483, + 0.635 + ], + "angle": 0, + "content": "[57] Antun Skuric, Vincent Padois, Nasser Rezzoug, and David Daney. On-line feasible wrench polytope evaluation based on human musculoskeletal models: an iterative convex hull method. IEEE robotics and automation letters, 7(2):5206-5213, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.637, + 0.483, + 0.691 + ], + "angle": 0, + "content": "[58] Hyungeun Song and Yoichi Hori. Inverse muscle group activity estimation based on neuromusculoskeletal system model. In TENCON 2015-2015 IEEE Region 10 Conference, pages 1-5. IEEE, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.483, + 0.746 + ], + "angle": 0, + "content": "[59] Yokhesh K Tamilselvam, Jacky Ganguly, Rajni V Patel, and Mandar Jog. Musculoskeletal model to predict muscle activity during upper limb movement. *Ieee Access*, 9:111472–111485, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.483, + 0.816 + ], + "angle": 0, + "content": "[60] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit Haim Bermano. Human motion diffusion model. In The Eleventh International Conference on Learning Representations ICLR 2023, Kigali, Rwanda, 2023. OpenReview.net. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.483, + 0.873 + ], + "angle": 0, + "content": "[61] Elly Trepman, Richard E Gellman, Lyle J Micheli, and CARLO J De Luca. Electromyographic analysis of grandplie in ballet and modern dancers. Medicine and science in sports and exercise, 30(12):1708-1720, 1998. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[62] J. Wang, Y. Yuan, Z. Luo, K. Xie, D. Lin, U. Iqbal, S. Fidler, and S. Khamis. Learning human dynamics in autonomous" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "driving scenarios. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 20739-20749, Los Alamitos, CA, USA, 2023. IEEE Computer Society. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.137, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[63] Keenon Werling, Dalton Omens, Jeongseok Lee, Ioannis Exarchos, and C. Karen Liu. Fast and Feature-Complete Differentiable Physics Engine for Articulated Rigid Bodies with Contact Constraints. In Robotics: Science and Systems XVII, Virtual Event, July 12-16, 2021, Virtual, 2021. RSS Foundation. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[64] Keenon Werling, Nicholas A Bianco, Michael Raitor, Jon Stingel, Jennifer L Hicks, Steven H Collins, Scott L Delp, and C Karen Liu. Addbiomechanics: Automating model scaling, inverse kinematics, and inverse dynamics from human motion data through sequential optimization. Plos one, 18(11):e0295152, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.388 + ], + "angle": 0, + "content": "[65] Keenon Werling, Janelle Kaneda, Tian Tan, Rishi Agarwal, Six Skov, Tom Van Wouwe, Scott Uhlrich, Nicholas Bianco, Carmichael Ong, Antoine Falisse, et al. Addbiomechanics dataset: Capturing the physics of human motion at scale. In European Conference on Computer Vision, pages 490-508. Springer, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.907, + 0.473 + ], + "angle": 0, + "content": "[66] Lahiru N Wimalasena, Jonas F Braun, Mohammad Reza Keshtkaran, David Hofmann, Juan Álvaro Gallego, Cristiano Alessandro, Matthew C Tresch, Lee E Miller, and Chethan Pandarinath. Estimating muscle activation from emg using deep learning-based dynamical systems models. Journal of neural engineering, 19(3):036013, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.475, + 0.905, + 0.531 + ], + "angle": 0, + "content": "[67] Jungdam Won, Deepak Gopinath, and Jessica Hodgins. Control strategies for physically simulated characters performing two-player competitive sports. ACM Transactions on Graphics (TOG), 40(4):1-11, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.532, + 0.905, + 0.574 + ], + "angle": 0, + "content": "[68] Jungdam Won, Deepak Gopinath, and Jessica Hodgins. Physics-based character controllers using conditional vaes. ACM Transactions on Graphics (TOG), 41(4):1-12, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.905, + 0.644 + ], + "angle": 0, + "content": "[69] Baoping Xiong, Nianyin Zeng, Han Li, Yuan Yang, Yurong Li, Meilan Huang, Wuxiang Shi, Min Du, and Yudong Zhang. Intelligent prediction of human lower extremity joint moment: an artificial neural network approach. IEEE Access, 7:29973-29980, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.646, + 0.905, + 0.701 + ], + "angle": 0, + "content": "[70] Aya Yaacoub, Vincent Thomas, Francis Colas, and Pauline Maurice. A probabilistic model for cobot decision making to mitigate human fatigue in repetitive co-manipulation tasks. IEEE Robotics and Automation Letters, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.703, + 0.905, + 0.773 + ], + "angle": 0, + "content": "[71] Katsu Yamane, Akihiko Murai, Sadahiro Takaya, and Yoshihiko Nakamura. Muscle tension database for contact-free estimation of human somatosensory information. In 2009 IEEE International Conference on Robotics and Automation, pages 633-638. IEEE, 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.775, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[72] X. Yi, Y. Zhou, M. Habermann, S. Shimada, V. Golyanik, C. Theobalt, and F. Xu. Physical inertial poser (pip): Physics-aware real-time human motion tracking from sparse inertial sensors. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13157-13168, Los Alamitos, CA, USA, 2022. IEEE Computer Society. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[73] Y. Yuan and K. Kitani. Ego-posed estimation and forecasting as real-time pd control. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "27792" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "10081-10091, Los Alamitos, CA, USA, 2019. IEEE Computer Society. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.176 + ], + "angle": 0, + "content": "[74] Felix E Zajac. Muscle and tendon: properties, models, scaling, and application to biomechanics and motor control. Critical reviews in biomedical engineering, 17(4):359-411, 1989. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[75] Petrissa Zell and Bodo Rosenhahn. A physics-based statistical model for human gait analysis. In Pattern Recognition, pages 169-180, Cham, 2015. Springer International Publishing. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.482, + 0.29 + ], + "angle": 0, + "content": "[76] P. Zell and B. Rosenhahn. Learning-based inverse dynamics of human motion. In 2017 IEEE International Conference on Computer Vision Workshop (ICCVW), pages 842–850, Los Alamitos, CA, USA, 2017. IEEE Computer Society. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.292, + 0.482, + 0.361 + ], + "angle": 0, + "content": "[77] P. Zell, B. Wandt, and B. Rosenhahn. Joint 3d human motion capture and physical analysis from monocular videos. In 2017 IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 17-26, Los Alamitos, CA, USA, 2017. IEEE Computer Society. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[78] Petrissa Zell, Bodo Rosenhahn, and Bastian Wandt. Weakly-supervised learning of human dynamics. In Computer Vision - ECCV 2020, pages 68-84, Cham, 2020. Springer International Publishing. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.419, + 0.482, + 0.487 + ], + "angle": 0, + "content": "[79] Longbin Zhang, Davit Soselia, Ruoli Wang, and Elena M Gutierrez-Farewik. Estimation of joint torque by emg-driven neuromusculoskeletal models and lstm networks. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.489, + 0.482, + 0.544 + ], + "angle": 0, + "content": "[80] Yufei Zhang, Jeffrey O Kephart, and Qiang Ji. Incorporating physics principles for precise human motion prediction. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 6164–6174, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.546, + 0.482, + 0.629 + ], + "angle": 0, + "content": "[81] Chenhui Zuo, Kaibo He, Jing Shao, and Yanan Sui. Self model for embodied intelligence: Modeling full-body human musculoskeletal system and locomotion control with hierarchical low-dimensional representation. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 13062-13069. IEEE, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.629 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "27793" + } + ] +] \ No newline at end of file diff --git a/2025/Homogeneous Dynamics Space for Heterogeneous Humans/a999c957-db24-46d0-83c3-e6da7809465e_origin.pdf b/2025/Homogeneous Dynamics Space for Heterogeneous Humans/a999c957-db24-46d0-83c3-e6da7809465e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..cce234edc67b05cffe55b24cc70fe0bfa47aebb7 --- /dev/null +++ b/2025/Homogeneous Dynamics Space for Heterogeneous Humans/a999c957-db24-46d0-83c3-e6da7809465e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0a7432de4c71438b7cca39a0c5eeb73809b9e1cb474d0af98949b41874e26a79 +size 3807990 diff --git a/2025/Homogeneous Dynamics Space for Heterogeneous Humans/full.md b/2025/Homogeneous Dynamics Space for Heterogeneous Humans/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2a8e6b366265f892434210f45a7b324fb6082c27 --- /dev/null +++ b/2025/Homogeneous Dynamics Space for Heterogeneous Humans/full.md @@ -0,0 +1,360 @@ +# Homogeneous Dynamics Space for Heterogeneous Humans + +Xinpeng Liu $^{1,2}$ , Junxuan Liang $^{1}$ , Chenshuo Zhang $^{1}$ , Zixuan Cai $^{3}$ , Cewu Lu $^{1,2*}$ , Yong-Lu Li $^{1,2*}$ + +1Shanghai Jiao Tong University, 2Shanghai Innovation Institute, 3Soochow University xinpengliu0907@gmail.com, zxcai@stu.suda.edu.cn, {whitefork, zhangchenshuo, lucewu, yonglu.li}@sjtu.edu.cn + +# Abstract + +Analyses of human motion kinematics have achieved tremendous advances. However, the production mechanism, known as human dynamics, is still undercovered. In this paper, we aim to push the understanding of data-driven human dynamics forward. We identify a major obstacle to this as the heterogeneity of existing human motion understanding efforts. Specifically, heterogeneity exists in not only the diverse kinematics representations and hierarchical dynamics representations but also the data from different domains, namely biomechanics and reinforcement learning. With an in-depth analysis of the existing heterogeneity, we propose to emphasize the beneath homogeneity: all of them represent the homogeneous fact of human motion, though from different perspectives. Given this, we propose Homogeneous Dynamics Space (HDyS) as a fundamental space for human dynamics by aggregating heterogeneous data and training a homogeneous latent space with inspiration from the inverse-forward dynamics procedure. HDyS achieves decent mapping between human kinematics and dynamics by leveraging the heterogeneous representations and datasets. We demonstrate the feasibility of HDyS with extensive experiments and applications. The project page is https://foruck.github.io/HDyS. + +# 1. Introduction + +Analyses on human motion have a wide range of applications, including animation [11, 60], healthcare [10, 55], and robotics [57, 70]. The computer vision community has made tremendous progress in understanding human kinematics with tasks like human reconstruction [24, 27, 34], action recognition [25, 26, 28, 29, 51], and motion generation [30, 32, 60]. However, the production mechanism hidden beneath human motion, known as human dynamics, is still limitedly explored. + +In this paper, we aim to push the understanding of human + +![](images/e12f39a53841ebb868ad4c4bbeda49a01aa9df498208085ca756edd870cc5d3f.jpg) +Figure 1. The representation heterogeneity exists for both kinematics and dynamics. + +dynamics forward with data-driven methodologies. Specifically, we pursue to build a bidirectional mapping between human kinematics and dynamics. We identify the major obstacle to this as the heterogeneity of existing human motion understanding efforts, which could be two-fold. + +First, representation heterogeneity exists for both kinematics and dynamics, as shown in Figure 1. A typical optical motion capture pipeline, which produces most existing available data for human kinematics [20, 40], involves tracking the markers placed on the surface of humans. Then, different human template models are fitted to the tracked markers, resulting in parameters of musculoskeletal models [52] for biomechanics uses or SMPL-like models [36] for general CV/CG uses. Skeletons with pre-defined kinematic trees are sometimes preferred as a simple and intuitive representation. Different datasets tend to include only representations preferred by their designed goal. Besides the kinematics representation heterogeneity, the dynamics representation heterogeneity exists more severely. Human dynamics function in a hierarchical manner. The closest dynamics hierarchy to the kinematics is the joint torques, which are further produced by muscle forces from muscle actions according to the muscle contraction dynamics. Furthermore, electric signals from the neural system activate the muscle, which surface electromyogra + +phy (sEMG) could detect. Effective conversions between these different representations are not always available. For example, for kinematics representations, the conversion between the musculoskeletal model and SMPL becomes available in very recent advances [20], which requires time-consuming optimization. For dynamics, the bi-directional mapping between joint torques, muscle actions, and electric signals is even more difficult. + +Second, domain heterogeneity also exists. One crucial characteristic of human dynamics is the difficulty of non-intrusive measurement. In the biomechanics community, optimization-based methods are adopted to solve the Euler-Newton equation for joint torques [64] and muscle actions [53]. For electric signals, it is typically measured by sEMG sensors [3]. In the meantime, from the learning community, there emerge efforts [33] utilizing Reinforcement Learning (RL) to imitate existing motion captures with physically simulated humanoids and record the simulated human dynamics data in a fully synthetic manner. These different data sources introduce domain heterogeneity in three aspects. First, biomechanics data are more deeply rooted in reality with well-defined computation models and real human motions. Instead, RL data could diverge from real-world situations with unnatural motion imitations and inaccurate physics simulations, known as the sim2real domain gap. Second, biomechanics data, due to the optimization-based manner, is sensitive to variations and usually limited to strictly controlled laboratory setups with relatively simple motions, resulting in limited motion domain coverage. In contrast, RL data cover a wide range of motion with more diversity. Third, biomechanics data and RL data tend to adopt heterogeneous kinematics representations with different kinematic trees, making it hard to transfer directly from one domain to another. + +Despite the heterogeneity, we emphasize the homogeneous essence behind it: all of them represent the homogeneous fact of human motion, each from a different perspective. Given this, we demonstrate the feasibility of unifying these heterogeneous human representations and exploiting the underlying homogeneous knowledge of human dynamics. Low-level Cartesian kinematics representations, like markers and joints, are less heterogeneous than representations like joint angles. Also, though joint torques, muscle actions, and electric signals are not directly transferable to each other, they could share similar motor knowledge like coordination. To this end, we propose Homogeneous Dynamics Space for heterogeneous humans (HDyS). + +To achieve this, we aggregate human dynamics data from both RL [33] and biomechanics [3, 53, 65] with different kinematics and dynamics representations, covering human dynamics hierarchies, including joint torques, muscle actions, and electric signals. Then, HDyS is designed as an aggregation of multiple auto-encoders corresponding to the + +inverse-forward dynamics procedure. We supervise HDyS with reconstruction and alignment losses. Our proposed HDyS has two major merits. First, by unifying heterogeneous human representation in the same latent space, it generalizes across different representations seamlessly while taking advantage of each. Second, by aggregating largescale heterogeneous motion data, it is empowered with general human motor knowledge from a wide span of motion, functioning as a reusable knowledge source for downstream dynamics-related applications. To demonstrate the efficacy of HDyS, we first evaluate it on human inverse dynamics. Then, we showcase how HDyS could facilitate downstream dynamics-related applications like ground reaction force prediction, biomechanics human simulation, and physics-simulated character control. + +To summarize, our contribution includes: 1) We analyzed the heterogeneity issue that hinders an in-depth understanding of human dynamics. 2) We highlighted the homogeneity beneath the heterogeneity and proposed a fundamental reusable space HDyS for human dynamics by unifying the heterogeneity. 3) We demonstrated the feasibility of digging homogeneity out from heterogeneity with extensive experiments and applications of HDyS. + +# 2. Related Works + +# 2.1. Human Dynamics + +By human dynamics, we mean the production mechanism of human motion, which the biomechanics community has actively explored. To produce a certain motion, neural commands are sent to activate the muscles. After receiving the activation signals, muscles contract and produce muscle forces. Multiple muscle forces form the joint torques according to certain musculoskeletal geometry, and the joint torques drive the accelerations that accumulate into movements. Thus, understanding human dynamics typically involves two heterogeneous hierarchies: joint torques and muscle activations. However, both are hard to measure nonintrusively. In the literature, to obtain them, an optimization problem is typically introduced as + +$$ +\begin{array}{l} \min \| a \|, s. t. 0 \leq a \leq 1, \tau = A (q) F (a), \\ M (q) \ddot {q} + C (q, \dot {q}) + G (q) = J \lambda + \tau . \tag {1} \\ \end{array} +$$ + +with the generalized human inertia matrix $M(q)$ w.r.t. generalized coordinate $q$ , Coriolis and centrifugal forces $C(q,\dot{q})$ , gravity $G(q)$ , Jacobian matrix $J$ mapping external forces $\lambda$ to the generalized coordinates, and muscle activations $a$ . The joint torques $\tau$ could be obtained by $\tau = A(q)F(a)$ , where $A(q)$ maps muscle forces into joint torques and $F(a)$ maps $a$ into muscle forces usually with the hill-type function [15, 74]. Mature software [4, 5, 63] were developed for this purpose. However, the optimization quality is tightly bonded to the precision of external + +force measurement, which could be expensive. Therefore, the applications are mostly limited to simple motions like gaits in laboratory settings. Some efforts exploited wearable devices [22, 23] for more general applications. In addition, the optimization is deeply coupled with the adopted human models [36, 52], which vary with the application preferences. Fitting raw motion capture data to a specific human model could be unstable and time-consuming. The conversion between different human models is also nontrivial even with recent advances [20]. This way, transferring dynamics from one human model to another is limitedly explored. Besides torques and muscle actions that are typically obtained via optimization, Surface Electromyography (sEMG) is adopted as an indirect representation of human dynamics which could be directly measured. Efforts were made to build the accurate mapping from sEMG to muscle actions [19, 66], joint torques [54, 79], and human poses [35, 55]. Though progress was made, sEMG patterns might suffer from noises and vary drastically among subjects [61], hindering the generalization. + +# 2.2. Learning-based Human Dynamics + +Joint Torques. Early efforts were made on ML-based joint torque analysis for certain human body parts [17, 43, 69]. Lv et al. [39] developed a Gaussian mixture framework for whole-body joint torque estimation. Other architectures like k-nearest neighbor-based regression [75, 77], random forests [76], and neural networks [78] were also adopted for the estimation of joint torques. However, most of these efforts suffered from limited data scale, which hampered learning methods from exploiting their full potential. The recent emergence of AddBiomechanics [65], which aggregated multiple biomechanics datasets, considerably boosted the data scale. However, most of the collected sequences contained only gaits with limited diversity. Another line of work adopted reinforcement learning to simulate motion in physics simulators, and the joint torques could be obtained in simulation. Though generalizability was limited at first [1, 49, 50, 67, 68], emerging efforts [37, 38, 73] managed to replicate a wide span of motion in simulators. The paradigm is further incorporated into recent MoCap systems for simultaneous estimation of motion and the joint torques [8, 12, 16, 62, 72, 80]. However, due to the involvement of simulators, the sim2real gap exists. The learned dynamics could be restricted to certain simulators or human models, which might diverge from real humans. + +Muscle Actions. sEMG is usually adopted as a proxy measurement of muscle actions and has been extensively studied in biomechanics [3, 6, 7, 14, 41, 44, 48]. There have been efforts delving into predicting sEMG signals given either joint torques [31, 56, 58], goniometers [59], motion captures [18, 45, 71], videos [3, 48], or point clouds [46]. However, each effort could be small in data scale, mo + +tion variations, and muscle coverage. Unifying these efforts could be hindered by the heterogeneous settings and data formats adopted, thus under-covered. Recently, musculoskeletal human simulation has gained attention. Multiple efficient simulators were developed [2, 9, 81], opening the potential for more accessible muscle action analysis. Specifically, MinT [53] was proposed by simulating motion sequences from AMASS [40] in OpenSim [5] and attaching simulated muscle actions to the raw motion sequence, enabling muscle action learning in scale. + +# 3. Method + +We introduce the proposed Homogeneous Dynamics Space (HDyS). We first cover the involved kinematics and dynamics representations in Section 3.1, 3.2. Then, the model architecture and designed losses are introduced in Section 3.3. + +# 3.1. Kinematics Representations + +As shown in Fig. 2, we adopt four types of kinematics representations: Cartesian representations of markers and skeleton key-points, joint angles of Rajagopal's model [52] from the biomechanics community, and SMPL [36] which is widely adopted for CV and CG applications. + +Markers, placed on the human body surfaces, are typically the raw data for optical motion capture. In practice, most other representations are calculated by fitting certain human prior models to the marker observations, making it easy to obtain for heterogeneous datasets and representations. Thus, markers are expected to be a generalizable representation across heterogeneous datasets, though they could also be rather low-level and thus hard to learn. We define the marker representation at timestamp $t$ as $x_{m}^{t} = (m^{t},\dot{m}^{t},\dot{m})\in \mathcal{R}^{N_{m}\times 9}$ , which is composed of marker Cartesian coordinates $m^t$ , finite-differentiated velocities $\dot{m}^t$ and finite-differentiated accelerations $\dot{m}^t$ with $N_{m}$ markers. + +Skeletal Keypoints, compared to markers, are less generalizable due to their reliance on pre-defined kinematic trees. However, with its Cartesian coordinates, common sense of human topology, and easy access, the gap between different kinematic trees can be mitigated. We define the joint representation at timestamp $t$ as $x_{k}^{t} = (k^{t},\dot{k}^{t},\dot{k})\in \mathcal{R}^{N_{k}\times 9}$ , which is composed of joint coordinates $k^t$ , finite-differentiated velocities $\dot{k}^t$ and finite-differentiated accelerations $\dot{k}^t$ with $N_{k}$ skeletal joints. + +Joint Angles are preferred for clinical analyses which more faithfully preserves the biomechanics information of human kinematics. We adopt the Rajagopal's model [52] used in AddBiomechanics [65], and define the joint angle representation at timestamp $t$ as $x_{a}^{t} = (a^{t},\dot{a}^{t},\dot{a})\in \mathcal{R}^{3N_{j}}$ , containing joint angles $a^t$ , finite-differentiated velocities $\dot{a}^t$ and accelerations $\dot{a}^t$ with $N_{j}$ joints in Rajagopal's model (only 23 lower-body joints are used). + +![](images/b98c47e8f12e3bb45e35434811d4eea3e2a9dc2220e3b5c3e3ef708495ada98f.jpg) +Figure 2. The overall architecture of HDyS. + +SMPL [36] is a human parameter model widely adopted in the CV/CG community. We define the SMPL representation at timestamp $t$ as $x_{s}^{t} = (s^{t},\dot{s}^{t},\dot{s})\in \mathcal{R}^{3N_{j}}$ , which is composed of SMPL parameters $s^t$ , finite-differentiated velocities $\dot{s}^t$ and finite-differentiated accelerations $\dot{s}^t$ with $N_{j}$ joints of SMPL (75 used, 3 translational joints at root and 72 revolving joints). + +# 3.2. Dynamics Representations + +We adopt three different types of dynamic representations: joint torques, muscle actions, and sEMGs. + +Joint Torques are the net torques exerted at each joint to drive the motion, represented as $\tau_{t} \in \mathcal{R}^{N_{t}}$ . $N_{t}$ is coupled with the number of joints $N_{j}$ as $N_{t} = N_{j} - 6$ , with the 6-DoF free root joint unactuated. For clarity, we denote the joint torques corresponding to Rajagopal's model as $\tau_{tr}$ and those corresponding to SMPL as $\tau_{ts}$ . + +Muscle Actions represent the activation level of each muscle, resulting in the muscle forces that produce joint torques. We denote muscle actions as $\tau_{m} \in \mathcal{R}^{N_{m}}$ with $N_{m}$ muscles. sEMG detects the electric potential generated on the body surface by activated muscle cells, which is closely related to muscle actions. We represent sEMGs at timestamp $t$ as $\tau_{e}^{t} \in \mathcal{R}^{N_{e}}$ with $N_{e}$ sEMG channels. + +# 3.3. Homogeneous Dynamics Space + +With the heterogeneous representations, we construct our homogeneous dynamics space (HDyS) as an aggregation of multiple auto-encoders corresponding to the inverseforward dynamics procedure. Then, reconstruction and alignment losses are adopted for supervision. For simplicity, all superscripts $t$ are omitted. + +Inverse-Dynamics Auto-Encoder (IDAE) encodes the kinematics into the latent space and decodes the dynamics + +from the latent. For markers $x_{m}$ and skeletal key points $x_{k}$ , we adopt three-layer transformer encoders with no positional embedding to encode them into latent $z_{m}, z_{k} \in \mathcal{R}^{d}$ of dimension $d$ . This enables the encoding of an arbitrary number of markers/points. For joint angles $x_{a}$ and SMPL parameters $x_{s}$ , we adopt simple three-layer MLPs to encode them into latent $z_{a}, z_{s} \in \mathcal{R}^{d}$ . The decoder is designed as a shared transformer followed by separate MLP regression heads. The transformer takes latent vectors from consecutive frames as input, outputting per-frame latent vectors refined with temporal contexts. Then, separate MLP decoders decode the per-frame dynamics as $\hat{\tau}_{tr}, \hat{\tau}_{ts}, \hat{\tau}_{m}, \hat{\tau}_{e}$ . + +Forward-Dynamics Auto-Encoder (FDAE) encodes dynamics and the kinematics except for accelerations (denoted as $\tilde{x}$ ) into the latent space and decodes the accelerations. Similar to IDAE, markers $\tilde{x}_m$ and skeletal key-points $\tilde{x}_k$ are encoded by transformers as $\tilde{z}_m,\tilde{z}_k$ , while joint angles $\tilde{x}_a$ and SMPL parameters $\tilde{x}_s$ are encoded by MLPs as $\tilde{z}_a,\tilde{z}_p$ . Then, joint torques $\tau_{tr},\tau_{ts}$ , muscle activations $\tau_m$ , and sEMGs $\tau_e$ are also encoded by separate MLP encoders as dynamics latent vectors $\tilde{z}^{tr},\tilde{z}^{ts},\tilde{z}^m,\tilde{z}^e$ . $\tilde{z}_m,\tilde{z}_k,\tilde{z}_a,\tilde{z}_p$ is then concatenated with available dynamics latent vectors, fed into a shared MLP composer, resulted in latent vectors $z^{\cdot}$ , with subscripts representing the dynamics. Although arbitrary combinations are feasible, in practice, only one dynamic latent vector typically exists due to the data limitations. Finally, separate MLP decoders decode the latent vectors $z_{\{kin\}}^{(dyn)}$ into skeletal key point accelerations $\ddot{k}$ , SMPL accelerations $\ddot{s}$ , and joint angle accelerations $\ddot{a}$ , where $kin\in \{m,k,s,a\}$ , $dyn\in \{m,e,ts,tr\}$ . The marker accelerations are not predicted since their variable sizes could bring unnecessary complexity. + +Loss Terms. To train HDyS, we adopt reconstruction losses and alignment losses. For reconstruction losses, we simply calculate the L1 loss for $\tau$ and $\ddot{a}$ as + +$$ +L _ {r e c o n} = \| \tau - \hat {\tau} \| _ {1} + \| \ddot {a} - \hat {\dot {a}} \|. \tag {2} +$$ + +For alignment losses, we align the latent vectors $z$ of the same frame and separate $zs$ of different frames using InfoNCE [47]. Given a batch of $B$ frames, the obtained latent vectors is denoted as $Z = \{z_{\{kin\}}, z_{\{kin\}}^{(dyn)}\}$ , with $kin \in \{m, k, s, a\}, dyn \in \{m, e, ts, tr\}$ . The loss is + +$$ +L _ {a l i g n} = \sum_ {z _ {2}, z _ {2} \in Z} \sum_ {i = 1} ^ {B} - \log \left(\frac {\exp \left(\left\langle z _ {1} ^ {(i)} , z _ {2} ^ {(i)} \right\rangle\right)}{\sum_ {j = 1} ^ {B} \exp \left(\left\langle z _ {1} ^ {(i)} , z _ {2} ^ {(j)} \right\rangle\right)}\right), \tag {3} +$$ + +where $(i),(j)$ indicate the batch index. The overall loss is as $\mathcal{L} = \alpha_{1}L_{recon} + \alpha_{2}L_{align}$ with coefficients $\alpha_{1},\alpha_{2}$ + +Table 1. Quantitative results of HDyS with ablation studies. For HDyS, the results of averaged predictions and the best prediction among all representations are reported. + +
MethodsImDy mPJE↓ avg/bstAddBiomechanics mPJE↓ avg/bstMinTMiA PCC↑ avg/bst
RMSE↓ avg/bstPCC↑ avg/bstRMSE↓ avg/bst
ImDyS [33]0.63000.1626----
MiA[3]----13.3-
HDyS0.5765/0.46740.1189/0.12430.0614/0.06150.7420/0.740211.8/11.60.7232/0.7261
ImDy-only HDyS0.6854/0.5403-----
AddBiomechanics-only HDyS-0.1695/0.1691----
MinT-only HDyS--0.0637/0.06400.7179/0.7127--
MiA-only HDyS----13.6/13.50.6557/0.6421
HDyS w/o ImDy-0.1214/0.13860.0608/0.06170.7470/0.741715.8/15.60.6523/0.6497
HDyS w/o AddBiomechanics0.5787/0.4742-0.0616/0.06140.7408/0.738617.1/17.00.5769/0.5638
HDyS w/o MinT0.5730/0.46810.1197/0.1296--16.9/16.80.5213/0.5313
HDyS w/o MiA0.5890/0.47880.1200/0.12840.0616/0.06180.7395/0.7375--
HDyS w/o AMASS0.5797/0.47860.1217/0.13190.0613/0.06170.7419/0.738014.7/14.90.5704/0.5632
HDyS w/o Lalign0.6575/0.50190.1270/0.13290.0626/0.06300.7318/0.723813.7/13.40.6464/0.6402
HDyS w/o FDAE0.5776/0.48490.1198/0.12610.0617/0.06170.7388/0.737513.6/13.30.6517/0.6699
HDyS-32D0.7390/0.64500.1401/0.14200.0650/0.06480.6980/0.701016.7/16.30.5524/0.5614
HDyS-64D0.6505/0.54100.1295/0.13540.0627/0.06290.7272/0.725415.1/14.50.6034/0.6187
+ +# 4. Experiments + +# 4.1. Datasets + +AddBiomechanics [64] contains over 50 hours of human motion data with joint torques from Nimble [63] simulation. We follow the setting in [64] with the armless part of AddBiomechanics. Markers, key points, and joint angles are used as kinematics representations, and joint torques corresponding to [52] represent the dynamics. + +Muscles in Times (MinT) [53] simulates part of AMASS in OpenSim to obtain the actions of 402 muscles. We randomly split them into 906 sequences for training and 227 sequences for testing. Markers, key points, and SMPL parameters are used as kinematics representations, and muscle actions are the dynamics representations. + +Muscles in Act (MiA) [3] consists of 12.8 hours of human exercise motion reconstructed from videos with VIBE [21] and sEMG data of 8 muscles. Following [3], we split them into 19,563 training sequences and 3,053 testing sequences. Markers and key points are kinematics representations, and sEMGs are dynamics representations. + +ImDy [33] adopted PHC [38] to imitate motion sequences from AMASS [40] and KIT [42], resulting in over 150 hours of human dynamics data, with 27,501 sequences for training and 3,055 sequences for evaluation. Markers, key points, and SMPL are kinematics representations, and SMPL torques are dynamics representations. + +AMASS [40] contains over 11,000 motion sequences represented in SMPL [36]. Though not paired with human dynamics, we included it as an extra knowledge base for training. Markers, key points, and SMPL parameters are used as kinematics representations. + +# 4.2. Implementation Details + +We adopt a compact design for HDyS with a latent dimension of 128. For joint angle and SMPL parameters, three-layer MLPs with hidden unit sizes of 256 and 128 are adopted as encoders, projecting the input to the 128-D latent space. For markers and key-points, three-layer transformer encoders with 2 heads are adopted as encoders. The transformer in the ID decoder is of 4 layers, 4 heads, and a dimension of 128. All MLPs in decoders are two-layer, with a hidden size of 32 for Rajagopal's torque, sEMG, joint angle accelerations, and key-point accelerations, and a hidden size of 64 for SMPL torque, SMPL accelerations, and muscle actions. For training, we adopt an AdamW optimizer with a learning rate of 1e-3 and a batch size of 9,600 frames for 1,000 epochs. Loss weights are set as $\alpha_{1} = 0.01$ , $\alpha_{2} = 0.05$ . Also, during training, a balanced sampling strategy is adopted to minimize the scale influence of different datasets. That is, we randomly sample 3,000 sequences per dataset for each training epoch. In this way, the comparison is fair in terms of the number of seen samples under different dataset settings. All experiments are conducted on a single NVIDIA Titan Xp GPU. + +# 4.3. Results on Inverse Dynamics + +Metric. We report mPJE and RMSE as + +$$ +m P J E _ {\tau .} = \frac {1}{J} \sum_ {j = 1} ^ {J} | \tau . - \hat {\tau}. | _ {2}, R M S E _ {\tau .} = | \tau . - \hat {\tau}. | _ {2}. \tag {4} +$$ + +PCC is also reported due to its invariance against scale and offset, which helps evaluate muscle actions and sEMG prediction. PCC is computed as + +$$ +P C C _ {\tau_ {1}} = \frac {\operatorname {c o v} \left(\tau , \hat {\tau}\right)}{\sigma_ {\tau_ {1}} \sigma_ {\hat {\tau}}} \tag {5} +$$ + +where $cov(\tau, \hat{\tau})$ is the covariance of $\tau$ , $\hat{\tau}$ , and $\sigma$ . indicates the standard deviation. For ImDy and AddBiomechanics, we report mPJE normalized by body weight. For MinT and MiA, RMSE and PCC are reported. + +# 4.3.1 Quantitative results + +Quantitative results are shown in Tab. 1, where HDyS is reported with the averaged predictions from different input kinematics representations and the best predictions among all representations. Compared to previous methods [3, 33], the proposed HDyS provides superior performances with substantial improvements on all datasets. We further analyze the performance of HDyS with three questions. + +How do the heterogeneous datasets contribute to HDyS? + +We compare HDyS with its single-dataset variants and drop-one-out variants in Tab. 1. HDyS outperforms the corresponding single-dataset variants on all datasets, validating the existence of homogeneous human dynamics knowledge behind these heterogeneous datasets. According to the drop-one-out results, it is noticeable that datasets with similar dynamics representations are more cooperative. That is, muscle-related datasets (MiA and MinT) tend to benefit more from each other and less from torque-related datasets (AddBiomechanics and ImDy), and vice versa. Moreover, mutually harmful effects could be observed for ImDy and MinT. These are consistent with the gaps between these datasets: the sim2real gap from ImDy to others and the torque-to-muscle gap from AddBiomechanics and ImDy to MiA and MinT. AMASS, though not paired with dynamics information, is shown to be beneficial for inverse dynamics with its diverse and high-quality kinematics. + +The heterogeneous datasets introduce both increased data scale and heterogeneous knowledge. Given this, we try to investigate further the source of the improvement in Tab. 1. Trying to decompose the contributions of scale and heterogeneity, we compared the performance of the following three models on a single target test set in Tab. 2. HDyS-Single-50 denotes single-dataset HDyS with $50\%$ of the data from the target dataset. HDyS-50/50 denotes HDyS using $50\%$ of the data from the target dataset and $50\%$ of the data from the other datasets, maintaining the same data scale as the target dataset. HDyS-Single represents the corresponding single-dataset variants of HDyS in Tab. 1. The performance of the best representation is reported on AddBiomechanics and MiA, considering their suitable scale and realistic nature. As shown, HDyS-50/50 consistently outperforms HDyS-Single-50, verifying the existence of homogeneous knowledge in heterogeneous data. Moreover, on AddBiomechanics, HDyS-50/50 even outperforms HDyS-Single, indicating that the diversity from heterogeneous data could sometimes benefit more than simply increasing seemingly homogeneous data. + +Table 2. Results on scale-heterogeneity decomposition. + +
DatasetHDyS-Single-50HDyS-50/50HDyS-Single
AddBiomechanics mPJE ↓0.17070.12840.1695
MiA RMSE ↓16.214.513.5
+ +Table 3. Results of different kinematic representations in HDyS. + +
MethodsImDy mPJE↓AddBiomechanics mPJE↓MinTMiA
RMSE↓PCC↑RMSE↓PCC↑
HDys-marker0.81630.14550.06460.696812.30.7088
HDys-keypoint0.80840.13240.06400.710311.60.7261
HDys-angle-0.1243----
HDys-SMPL0.4674-0.06150.7402--
+ +![](images/3c9f68164f38b8075ff6174233b461c36845759336cac95042a508218c559d09.jpg) +Heterogeneous Representations + +![](images/1a9e8fb895845f39a4673e9f0c0ea54b30022429f10b1f2537ca9a573b8196be.jpg) +Heterogeneous Datasets + +![](images/c944f285a4618bfc85270fe41c207c63756ae4492db652941e34227b808d5c71.jpg) +Figure 3. Feature visualization for HDyS w/ and w/o $L_{align}$ . + +![](images/81c5146fdcf80211a3b3c4ce4d5b1b782d035dda3aac2f149e57e487e741ca12.jpg) + +How do the heterogeneous representations contribute to HDyS? We report the results from individual kinematics representations in Tab. 3. Typically, the representations in joint space like SMPL and joint angles produce more accurate predictions compared to Cartesian representations with considerable performance gaps, which explains the performance drop when taking the average for ImDy and MiA. The markers are the worst-performed representation of all. + +In Tab. 1, it is noticeable that MinT and AddBiomechanics take advantage of averaging predictions from different representations. For AddBiomechanics, this might be because the best-performing representation is joint angles, which are not available in other datasets. Therefore, the averaging could effectively aggregate knowledge from other datasets with other representations. For MinT, it is observable that different representations produce close results, preventing the ensemble from being held back by markers. + +Furthermore, we examine the alignment of heterogeneous representations by removing the alignment loss $L_{\text{align}}$ in Tab. 1. Considerable performance drops could be observed. The drop is consistent for both best and average performances, indicating the cross-modality alignment manages to aggregate knowledge from heterogeneous representations. We also visualize HDyS features w/ and w/o $L_{\text{align}}$ in Fig. 3. As shown, $L_{\text{align}}$ prevents the HDyS feature from being dominated by the representation hetero- + +![](images/c338a3d62238cc2d69ce42e615c3b8fb8faee7744c9fbf3b7ca30145a7145f6f.jpg) +Figure 4. Inverse dynamics visualization on ImDy. + +geneity, encouraging the learning of homogeneous knowledge. Interestingly, though not intended to do so, $L_{\text{align}}$ also helps to mitigate the gap between datasets, alleviating the domain heterogeneity. + +How does the model design contribute to HDyS? We first evaluate the FDAE in Tab. 1. A substantial performance drop could be observed, demonstrating the benefit of informing HDyS with the forward dynamics procedure. Moreover, we also report HDyS with different dimensions in Tab. 1. Though HDyS's performance degrades with lower dimensions, it could still outperform some single-dataset variants with higher dimensions, validating the efficacy of heterogeneous datasets again. + +# 4.3.2 Qualitative results + +We visualize the inverse dynamics results of different datasets. As shown in Fig. 4-5, the joint torques could be faithfully reconstructed for different actions from either synthetic and jittering ImDy or realistic AddBiomechanics with the same HDyS. For the muscle actions shown in Fig. 6, HDyS produces reasonable estimations for lower-body muscles when walking. Moreover, as in the right half, though the lower-body kinematics are less significant, HDyS could capture the ankle muscle actions with high fidelity. The predictions are also coherent with GT for upper-body muscles like the internal oblique and levator scapulae. More details are in the appendix. + +# 4.4. Results on Downstream Tasks + +# 4.4.1 Ground Reaction Force Estimation + +We evaluate the effectiveness of HDyS on GRF estimation with GroundLink [13], which contains 1.5-hour motion with GRF recordings. We finetune HDyS with subjects 1-6 and evaluate it on subject 7. mPJE for GRF at both feet normalized by body weight is reported following Eq. 4. GroundLinkNet [13] and HDyS trained on GroundLink from scratch are also compared. As shown in Tab. 4, the + +![](images/448c23ac1e11f1c1637d97dadcd0dad8e1c6c17e3b9eb388ada01297c71e3f12.jpg) +Figure 5. Inverse dynamics visualization on AddBiomechanics. Table 4. Results on GroundLink. + +
MethodsGroundLinkNetGroundLink-onlyHDySHDyS
L-Foot mPJE ↓0.07110.05840.0514
R-Foot mPJE ↓0.09120.07510.0694
+ +Table 5. Biomechanical human simulation results reported in per-frame MSE. + +
k12345
Optimized0.000630.19091.83062.01062.6027
HDyS0.000610.18601.71181.86512.2233
+ +finetuned HDyS outperforms its counterparts, indicating the efficacy of the aggregated homogeneous knowledge. Also, the GroundLink-only HDyS outperforms GroundLinkNet, reflecting the feasibility of unifying heterogeneous representations for better dynamics knowledge. + +# 4.4.2 Biomechanical Human Simulation + +HDyS could also be adopted for biomechanical human simulation. We start with the armless Rajagopal's model [52] in Nimble [63]. Given a motion sequence, we first adopt HDyS to estimate the joint torques. Then, we use the predicted torques to reproduce the motion in Nimble. Starting from the current state, we feed the predicted torques for $k$ steps and compare the simulated joint angles $\hat{q}$ with the real joint angles $q$ . The simulation is performed at 90FPS. We report the per-frame MSE of joint angles. + +Results. We demonstrated the results in Tab. 5 and Fig. 7. The results with optimized torques are also reported as a reference. Surprisingly, for different $k$ , simulation with HDyS is superior with better stability. However, with $k$ increasing, MSE also increases noticeably, indicating the accumulation of drifting errors. Further enhancing the forward-dynamics compatibility of HDyS would be a promising goal. + +# 4.4.3 Physical Character Control + +We adopt HDyS for physical character control following the setting of PHC [38], with 140 testing sequences and the rest + +![](images/0ab4675e5c355f18458edf063e5773631eb42ff28d997c2969f63b646fdf3f49.jpg) + +![](images/de9a76612f8d11a892a9cb0b33d88f7f164d4afbcbec1415a46b35074afb2284.jpg) + +![](images/a20d64388348ebf1278ac27c378ff72c90f9cc0ecab68900d23db74d784a6156.jpg) + +![](images/f0ea7a9623921a6c62d080b9ef8be3394b4d16b4dc1e4c21b48f6fa03c4a38d5.jpg) + +![](images/930bd05651b006f5601ab61f5cdd2e86634c999a666b68cac522e5143d4c3467.jpg) + +![](images/04852060261f962f044e6d12cfa74873885975b5f1be74cc1d6dd2de5ea5c819.jpg) + +![](images/47e448ca6753e3bbdd9e68c4d385dc6ff0bd6cc1ad2166c9d86aeddbdeaf4fe3.jpg) + +![](images/6f3373ffaf9c1b2d3cc53659bcbe2263882375315ddbd973df17b97f85258a8d.jpg) + +![](images/b4cc7e30a2a58b8c7009d63846f7184b470a94a1434c400b2239f423d5fb4c37.jpg) + +![](images/2bb7e2f6e9992dcbd3c5a41957ff84379d758c8e15d4d303eac2f81eec048c71.jpg) + +![](images/04e0433ae86687811d3b2f0e401b14d344cffedb2095cc4e4bfe8c780fbbd692.jpg) + +![](images/0243863b503b6d33a9f026b8db299300a42caaa036b727fc336d072d20197d24.jpg) +Figure 6. Inverse dynamics results on MinT. + +![](images/07afc5c80888e1f19e2dc422c51add60356027182e82c147a7cdff20b05ff460.jpg) + +![](images/fed51097ebb0a64b17e3d3fcbe7fe49b093000590aa60372d7a2d12891a1560e.jpg) +Figure 7. Biomechanical human simulation visualization. + +Table 6. Results of HDyS for character control. + +
MethodsmPJPEg↓mPJEL↓Evel↓Eacc↓
Baseline74.51449.02911.56214.243
Baseline + HDyS72.69247.92011.30513.757
+ +for training. We train the baseline as a primitive in PHC with a batch size of 700 for 10k steps. Then, the HDyS latents of key points are inserted as extra observations. We report the global and local mPJPE and the errors of velocity and acceleration in Tab. 6. HDyS manages to improve its performance, validating its efficacy. + +# 5. Discussion + +Despite the impressive results of HDyS, it could be improved. First, it is noticeable that HDyS performs better for the lower body than the higher body in Fig. 6. This might be due to the imbalanced focus on lower body dynamics in data adopted by HDyS. For all datasets, data on lower-body dynamics like gaits are dominating. For AddBiomechanics, we only adopted its armless part. Enhancing HDyS with more upper-body dynamics would be helpful. Second, for muscle actions, HDys could omit minor changes, and the magnitudes could sometimes diverge from the real. Mitigating it with more curated models and loss terms is desirable. Third, as a first step toward homogeneous human dynamics learning, HDyS is primarily instantiated with five initial datasets and an intuitive model. Scaling HDyS up with more datasets and more human priors would be a meaningful goal. Finally, we demonstrate the potential of HDyS with some simple applications in Sec. 4.4, while more sophisticated use cases for musculoskeletal human simulation, humanoid control, and human-robot transfer learning would be promising as future works. + +# 6. Conclusion + +We analyzed the heterogeneity issue existing for human dynamics learning and highlighted the homogeneity beneath it. To fully exploit the homogeneity, we proposed HDyS as a homogeneous human dynamics space. Extensive experiments were conducted to validate the feasibility of digging homogeneity out from heterogeneity for human dynamics with detailed analyses of the contribution of heterogeneous components. We further demonstrated the potential of HDyS for downstream applications. We believe HDyS could shed new light on human dynamics understanding. + +# Acknowledgements + +This work is supported in part by the National Natural Science Foundation of China under Grant No.62306175, CCF-Tencent Rhino-Bird Open Research Fund. + +# References + +[1] Kevin Bergamin, Simon Clavet, Daniel Holden, and James Richard Forbes. Drecon: data-driven responsive control of physics-based characters. ACM Transactions On Graphics (TOG), 38(6):1-11, 2019. 3 +[2] Vittorio Caggiano, Huawei Wang, Guillaume Durandau, Massimo Sartori, and Vikash Kumar. Myosuite: A contact-rich simulation suite for musculoskeletal motor control. In Learning for Dynamics and Control Conference, pages 492-507. PMLR, 2022. 3 +[3] Mia Chiquier and Carl Vondrick. Muscles in action. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22091-22101, 2023. 2, 3, 5, 6 +[4] Michael Damsgaard, John Rasmussen, Søren Tørholm Christensen, Egidijus Surma, and Mark De Zee. Analysis of musculoskeletal systems in the anybody modeling system. Simulation Modelling Practice and Theory, 14(8): 1100-1111, 2006. 2 +[5] Scott L Delp, Frank C Anderson, Allison S Arnold, Peter Loan, Ayman Habib, Chand T John, Eran Guendelman, and Darryl G Thelen. Opensim: open-source software to create and analyze dynamic simulations of movement. IEEE transactions on biomedical engineering, 54(11):1940-1950, 2007. 2, 3 +[6] Benedikt Feldotto, Cristian Soare, Alois Knoll, Piyanee Sriya, Sarah Astill, Marc de Kamps, and Samit Chakrabarty. Evaluating muscle synergies with emg data and physics simulation in the neurorobotics platform. Frontiers in Neurorobotics, 16:856797, 2022. 3 +[7] Mariusz P Furmanek, Madhur Mangalam, Mathew Yarossi, Kyle Lockwood, and Eugene Tunik. A kinematic and emg dataset of online adjustment of reach-to-grasp movements to visual perturbations. Scientific data, 9(1):23, 2022. 3 +[8] E. Gartner, M. Andriluka, E. Coumans, and C. Sminchisescu. Differentiable dynamics for articulated 3d human motion reconstruction. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13180-13190, Los Alamitos, CA, USA, 2022. IEEE Computer Society. 3 +[9] Thomas Geijtenbeek. The Hyfydy simulation software, 2021. https://hyfydy.com.3 +[10] Daniel FN Gordon, Andreas Christou, Theodoros Stouraitis, Michael Gienger, and Sethu Vijayakumar. Learning personalised human sit-to-stand motion strategies via inverse musculoskeletal optimal control. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 10497–10503. IEEE, 2023. 1 +[11] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages + +5142-5151, Los Alamitos, CA, USA, 2022. IEEE Computer Society. 1 +[12] Erik Gartner, Mykhaylo Andriluka, Hongyi Xu, and Cristian Sminchisescu. Trajectory optimization for physics-based reconstruction of 3d human pose from monocular video. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13096-13105, Los Alamitos, CA, USA, 2022. IEEE Computer Society. 3 +[13] Xingjian Han, Ben Senderling, Stanley To, Deepak Kumar, Emily Whiting, and Jun Saito. Groundlink: A dataset unifying human body movement and ground reaction dynamics. In SIGGRAPH Asia 2023 Conference Papers, New York, NY, USA, 2023. Association for Computing Machinery. 7 +[14] Oscar G Hernández, Jose M Lopez-Castellanos, Carlos A Jara, Gabriel J Garcia, Andres Ubeda, Vicente Morell-Gimenez, and Francisco Gomez-Donoso. A kinematic, imaging and electromyography dataset for human muscular manipulability index prediction. Scientific Data, 10(1):132, 2023. 3 +[15] Archibald Vivian Hill. The heat of shortening and the dynamic constants of muscle. Proceedings of the Royal Society of London. Series B-Biological Sciences, 126(843):136-195, 1938. 2 +[16] B. Huang, L. Pan, Y. Yang, J. Ju, and Y. Wang. Neural mo- con: Neural motion control for physically plausible human motion capture. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6407- 6416, Los Alamitos, CA, USA, 2022. IEEE Computer Society. 3 +[17] Leif Johnson and Dana H. Ballard. Efficient codes for inverse dynamics during walking. In Proceedings of the Twenty-Eighth AAAI Conference on Artificial Intelligence, pages 343-349, Quebec City, Quebec, Canada, 2014. AAAI Press. 3 +[18] Lise A Johnson and Andrew J Fuglevand. Evaluation of probabilistic methods to predict muscle activity: implications for neuroprosthetics. Journal of neural engineering, 6(5):055008, 2009. 3 +[19] Kimoon Kang, Kiwon Rhee, and Hyun-Chool Shin. Event detection of muscle activation using an electromyogram. Applied Sciences, 10(16):5593, 2020. 3 +[20] Marilyn Keller, Keenon Werling, Soyong Shin, Scott Delp, Sergi Pujades, C Karen Liu, and Michael J Black. From skin to skeleton: Towards biomechanically accurate 3d digital humans. ACM Transactions on Graphics (TOG), 42(6): 1-12, 2023. 1, 2, 3 +[21] Muhammed Kocabas, Nikos Athanasiou, and Michael J Black. Vibe: Video inference for human body pose and shape estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5253-5263, 2020. 5 +[22] Claudia Latella, Naveen Kuppuswamy, Francesco Romano, Silvio Traversaro, and Francesco Nori. Whole-body human inverse dynamics with distributed micro-accelerometers, gyros and force sensing. Sensors, 16(5), 2016. 3 +[23] Claudia Latella, Silvio Traversaro, Diego Ferigo, Yeshasvi Tirupachuri, Lorenzo Rapetti, Francisco Javier An + +drade Chavez, Francesco Nori, and Daniele Pucci. Simultaneous floating-base estimation of human kinematics and joint torques. Sensors, 19(12), 2019. 3 +[24] Jiefeng Li, Chao Xu, Zhicun Chen, Siyuan Bian, Lixin Yang, and Cewu Lu. Hybrik: A hybrid analytical-neural inverse kinematics solution for 3d human pose and shape estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3383-3393, 2021. 1 +[25] Sheng-Wei Li, Zi-Xiang Wei, Wei-Jie Chen, Yi-Hsin Yu, Chih-Yuan Yang, and Jane Yung-jen Hsu. Sa-dvae: Improving zero-shot skeleton-based action recognition by disentangled variational autoencoders. In European Conference on Computer Vision, pages 447–462. Springer, 2024. 1 +[26] Yong-Lu Li, Siyuan Zhou, Xijie Huang, Liang Xu, Ze Ma, Hao-Shu Fang, Yanfeng Wang, and Cewu Lu. Transferable interactiveness knowledge for human-object interaction detection. In CVPR, 2019. 1 +[27] Yong-Lu Li, Xinpeng Liu, Han Lu, Shiyi Wang, Junqi Liu, Jiefeng Li, and Cewu Lu. Detailed 2d-3d joint representation for human-object interaction. In CVPR, 2020. 1 +[28] Yong-Lu Li, Liang Xu, Xinpeng Liu, Xijie Huang, Yue Xu, Shiyi Wang, Hao-Shu Fang, Ze Ma, Mingyang Chen, and Cewu Lu. Pastanet: Toward human activity knowledge engine. In CVPR, 2020. 1 +[29] Yong-Lu Li, Xinpeng Liu, Xiaogian Wu, Yizhuo Li, Zuoyu Qiu, Liang Xu, Yue Xu, Hao-Shu Fang, and Cewu Lu. Hake: A knowledge engine foundation for human activity understanding. TPAMI, 2022. 1 +[30] Yong-Lu Li, Xiaogian Wu, Xinpeng Liu, Yiming Dou, Yikun Ji, Junyi Zhang, Yixing Li, Jingru Tan, Xudong Lu, and Cewu Lu. From isolated islands to pangea: Unifying semantic space for human action understanding. In CVPR, 2024. 1 +[31] Zhan Li, David Guiraud, and Mitsuhiro Hayashibe. Inverse estimation of multiple muscle activations from joint moment with muscle synergy extraction. IEEE journal of biomedical and health informatics, 19(1):64-73, 2014. 3 +[32] Xinpeng Liu, Haowen Hou, Yanchao Yang, Yong-Lu Li, and Cewu Lu. Revisit human-scene interaction via space occupancy. In European Conference on Computer Vision, pages 1-19. Springer, 2024. 1 +[33] Xinpeng Liu, Junxuan Liang, Zili Lin, Haowen Hou, Yong-Lu Li, and Cewu Lu. Imdy: Human inverse dynamics from imitated observations. In The Thirteenth International Conference on Learning Representations, 2025. 2, 5, 6 +[34] Xiaoyang Liu, Boran Wen, Xinpeng Liu, Zizheng Zhou, Hongwei Fan, Cewu Lu, Lizhuang Ma, Yulong Chen, and Yong-Lu Li. Interacted object grounding in spatio-temporal human-object interactions. In AAAI, 2025. 1 +[35] Yilin Liu, Shijia Zhang, and Mahanth Gowda. Neuropose: 3d hand pose tracking using emg wearables. In Proceedings of the Web Conference 2021, pages 1471-1482, 2021. 3 +[36] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. Smpl: a skinned multiperson linear model. In ACM Transactions on Graphics, New York, NY, USA, 2015. Association for Computing Machinery. 1, 3, 4, 5 + +[37] Zhengyi Luo, Ryo Hachiuma, Ye Yuan, and Kris Kitani. Dynamics-regulated kinematic policy for egocentric pose estimation. Advances in Neural Information Processing Systems, 34:25019-25032, 2021. 3 +[38] Z. Luo, J. Cao, A. Winkler, K. Kitani, and W. Xu. Perpetual humanoid control for real-time simulated avatars. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 10861-10870, Los Alamitos, CA, USA, 2023. IEEE Computer Society. 3, 5, 7 +[39] Xiaolei Lv, Jinxiang Chai, and Shihong Xia. Data-driven inverse dynamics for human motion. ACM Transactions on Graphics (TOG), 35(6):1-12, 2016. 3 +[40] N. Mahmood, N. Ghorbani, N. F. Troje, G. Pons-Moll, and M. Black. Amass: Archive of motion capture as surface shapes. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 5441-5450, Los Alamitos, CA, USA, 2019. IEEE Computer Society. 1, 3, 5 +[41] Nebojša Malešević, Alexander Olsson, Paulina Sager, Elin Andersson, Christian Cipriani, Marco Controzzi, Anders Björkman, and Christian Antfolk. A database of high-density surface electromyogram signals comprising 65 isometric hand gestures. Scientific Data, 8(1):63, 2021. 3 +[42] Christian Mandery, Ömer Terlemez, Martin Do, Nikolaus Vahrenkamp, and Tamim Asfour. Unifying representations and large-scale whole-body motion databases for studying human motion. IEEE Transactions on Robotics, 32(4):796-809, 2016. 5 +[43] Mykhailo Manukian, Serhii Bahdasariants, and Sergiy Yakovenko. Artificial physics engine for real-time inverse dynamics of arm and hand movement. Plos one, 18(12): e0295750, 2023. 3 +[44] Luís Moreira, Joana Figueiredo, Pedro Fonseca, João P Vilas-Boas, and Cristina P Santos. Lower limb kinematic, kinetic, and emg data from young healthy humans during walking at controlled speeds. Scientific data, 8(1):103, 2021. 3 +[45] Yoshihiko Nakamura, Katsu Yamane, Yusuke Fujita, and Ichiro Suzuki. Somatosensory computation for man-machine interface from motion-capture data and musculoskeletal human model. IEEE Transactions on Robotics, 21(1):58-66, 2005. 3 +[46] Hui Niu, Takahiro Ito, Damien Desclaux, Ko Ayusawa, Yusuke Yoshiyasu, Ryusuke Sagawa, and Eiichi Yoshida. Estimating muscle activity from the deformation of a sequential 3d point cloud. Journal of Imaging, 8(6):168, 2022. 3 +[47] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 4 +[48] Kunyu Peng, David Schneider, Alina Roitberg, Kailun Yang, Jiaming Zhang, Chen Deng, Kaiyu Zhang, M Saqib Sarfraz, and Rainer Stiefelhagen. Towards video-based activated muscle group estimation in the wild. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 4495-4504, 2024. 3 +[49] Xue Bin Peng, Ze Ma, Pieter Abbeel, Sergey Levine, and Angjoo Kanazawa. Amp: Adversarial motion priors for stylized physics-based character control. ACM Transactions on Graphics (ToG), 40(4):1-20, 2021. 3 + +[50] Xue Bin Peng, Yunrong Guo, Lina Halper, Sergey Levine, and Sanja Fidler. Ase: Large-scale reusable adversarial skill embeddings for physically simulated characters. ACM Transactions On Graphics (TOG), 41(4):1-17, 2022. 3 +[51] A. R. Punnakkal, A. Chandrasekaran, N. Athanasiou, A. Quiros-Ramirez, and M. J. Black. Babel: Bodies, action and behavior with english labels. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 722-731, Los Alamitos, CA, USA, 2021. IEEE Computer Society. 1 +[52] Apoorva Rajagopal, Christopher L Dembia, Matthew S Demers, Denny D Delp, Jennifer L Hicks, and Scott L Delp. Full-body musculoskeletal model for muscle-driven simulation of human gait. IEEE transactions on biomedical engineering, 63(10):2068-2079, 2016. 1, 3, 5, 7 +[53] David Schneider, Simon Reiß, Marco Kugler, Alexander Jaus, Kunyu Peng, Susanne Sutschet, Muhammad Saqib Sarfraz, Sven Matthiesen, and Rainer Stiefelhagen. Muscles in time: Learning to understand human motion in-depth by simulating muscle activations. Advances in Neural Information Processing Systems, 2025. 2, 3, 5 +[54] Robert V Schulte, Marijke Zondag, Jaap H Buurke, and Erik C Prinsen. Multi-day emg-based knee joint torque estimation using hybrid neuromusculoskeletal modelling and convolutional neural networks. Frontiers in Robotics and AI, 9:869476, 2022. 3 +[55] Paniz Sedighi, Xingyu Li, and Mahdi Tavakoli. Emg-based intention detection using deep learning for shared control in upper-limb assistive exoskeletons. IEEE Robotics and Automation Letters, 2023. 1, 3 +[56] Masashi Sekiya, Sho Sakaino, and Tsuji Toshiaki. Linear logistic regression for estimation of lower limb muscle activations. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 27(3):523-532, 2019. 3 +[57] Antun Skuric, Vincent Padois, Nasser Rezzoug, and David Daney. On-line feasible wrench polytope evaluation based on human musculoskeletal models: an iterative convex hull method. IEEE robotics and automation letters, 7(2):5206-5213, 2022. 1 +[58] Hyungeun Song and Yoichi Hori. Inverse muscle group activity estimation based on neuromusculoskeletal system model. In TENCON 2015-2015 IEEE Region 10 Conference, pages 1-5. IEEE, 2015. 3 +[59] Yokhesh K Tamilselvam, Jacky Ganguly, Rajni V Patel, and Mandar Jog. Musculoskeletal model to predict muscle activity during upper limb movement. *Ieee Access*, 9:111472–111485, 2021. 3 +[60] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit Haim Bermano. Human motion diffusion model. In The Eleventh International Conference on Learning Representations ICLR 2023, Kigali, Rwanda, 2023. OpenReview.net. 1 +[61] Elly Trepman, Richard E Gellman, Lyle J Micheli, and CARLO J De Luca. Electromyographic analysis of grandplie in ballet and modern dancers. Medicine and science in sports and exercise, 30(12):1708-1720, 1998. 3 +[62] J. Wang, Y. Yuan, Z. Luo, K. Xie, D. Lin, U. Iqbal, S. Fidler, and S. Khamis. Learning human dynamics in autonomous + +driving scenarios. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 20739-20749, Los Alamitos, CA, USA, 2023. IEEE Computer Society. 3 +[63] Keenon Werling, Dalton Omens, Jeongseok Lee, Ioannis Exarchos, and C. Karen Liu. Fast and Feature-Complete Differentiable Physics Engine for Articulated Rigid Bodies with Contact Constraints. In Robotics: Science and Systems XVII, Virtual Event, July 12-16, 2021, Virtual, 2021. RSS Foundation. 2, 5, 7 +[64] Keenon Werling, Nicholas A Bianco, Michael Raitor, Jon Stingel, Jennifer L Hicks, Steven H Collins, Scott L Delp, and C Karen Liu. Addbiomechanics: Automating model scaling, inverse kinematics, and inverse dynamics from human motion data through sequential optimization. Plos one, 18(11):e0295152, 2023. 2, 5 +[65] Keenon Werling, Janelle Kaneda, Tian Tan, Rishi Agarwal, Six Skov, Tom Van Wouwe, Scott Uhlrich, Nicholas Bianco, Carmichael Ong, Antoine Falisse, et al. Addbiomechanics dataset: Capturing the physics of human motion at scale. In European Conference on Computer Vision, pages 490-508. Springer, 2024. 2, 3 +[66] Lahiru N Wimalasena, Jonas F Braun, Mohammad Reza Keshtkaran, David Hofmann, Juan Álvaro Gallego, Cristiano Alessandro, Matthew C Tresch, Lee E Miller, and Chethan Pandarinath. Estimating muscle activation from emg using deep learning-based dynamical systems models. Journal of neural engineering, 19(3):036013, 2022. 3 +[67] Jungdam Won, Deepak Gopinath, and Jessica Hodgins. Control strategies for physically simulated characters performing two-player competitive sports. ACM Transactions on Graphics (TOG), 40(4):1-11, 2021. 3 +[68] Jungdam Won, Deepak Gopinath, and Jessica Hodgins. Physics-based character controllers using conditional vaes. ACM Transactions on Graphics (TOG), 41(4):1-12, 2022. 3 +[69] Baoping Xiong, Nianyin Zeng, Han Li, Yuan Yang, Yurong Li, Meilan Huang, Wuxiang Shi, Min Du, and Yudong Zhang. Intelligent prediction of human lower extremity joint moment: an artificial neural network approach. IEEE Access, 7:29973-29980, 2019. 3 +[70] Aya Yaacoub, Vincent Thomas, Francis Colas, and Pauline Maurice. A probabilistic model for cobot decision making to mitigate human fatigue in repetitive co-manipulation tasks. IEEE Robotics and Automation Letters, 2023. 1 +[71] Katsu Yamane, Akihiko Murai, Sadahiro Takaya, and Yoshihiko Nakamura. Muscle tension database for contact-free estimation of human somatosensory information. In 2009 IEEE International Conference on Robotics and Automation, pages 633-638. IEEE, 2009. 3 +[72] X. Yi, Y. Zhou, M. Habermann, S. Shimada, V. Golyanik, C. Theobalt, and F. Xu. Physical inertial poser (pip): Physics-aware real-time human motion tracking from sparse inertial sensors. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13157-13168, Los Alamitos, CA, USA, 2022. IEEE Computer Society. 3 +[73] Y. Yuan and K. Kitani. Ego-posed estimation and forecasting as real-time pd control. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages + +10081-10091, Los Alamitos, CA, USA, 2019. IEEE Computer Society. 3 +[74] Felix E Zajac. Muscle and tendon: properties, models, scaling, and application to biomechanics and motor control. Critical reviews in biomedical engineering, 17(4):359-411, 1989. 2 +[75] Petrissa Zell and Bodo Rosenhahn. A physics-based statistical model for human gait analysis. In Pattern Recognition, pages 169-180, Cham, 2015. Springer International Publishing. 3 +[76] P. Zell and B. Rosenhahn. Learning-based inverse dynamics of human motion. In 2017 IEEE International Conference on Computer Vision Workshop (ICCVW), pages 842–850, Los Alamitos, CA, USA, 2017. IEEE Computer Society. 3 +[77] P. Zell, B. Wandt, and B. Rosenhahn. Joint 3d human motion capture and physical analysis from monocular videos. In 2017 IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 17-26, Los Alamitos, CA, USA, 2017. IEEE Computer Society. 3 +[78] Petrissa Zell, Bodo Rosenhahn, and Bastian Wandt. Weakly-supervised learning of human dynamics. In Computer Vision - ECCV 2020, pages 68-84, Cham, 2020. Springer International Publishing. 3 +[79] Longbin Zhang, Davit Soselia, Ruoli Wang, and Elena M Gutierrez-Farewik. Estimation of joint torque by emg-driven neuromusculoskeletal models and lstm networks. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 2023. 3 +[80] Yufei Zhang, Jeffrey O Kephart, and Qiang Ji. Incorporating physics principles for precise human motion prediction. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 6164–6174, 2024. 3 +[81] Chenhui Zuo, Kaibo He, Jing Shao, and Yanan Sui. Self model for embodied intelligence: Modeling full-body human musculoskeletal system and locomotion control with hierarchical low-dimensional representation. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 13062-13069. IEEE, 2024. 3 \ No newline at end of file diff --git a/2025/Homogeneous Dynamics Space for Heterogeneous Humans/images.zip b/2025/Homogeneous Dynamics Space for Heterogeneous Humans/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d33d1d20743a9fc0687bd6a451c20b7f34237c41 --- /dev/null +++ b/2025/Homogeneous Dynamics Space for Heterogeneous Humans/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ebde7a2e53847b8ae58ee1feeadcf0ed383968b492ce6fc9637b917e4875f6d8 +size 617788 diff --git a/2025/Homogeneous Dynamics Space for Heterogeneous Humans/layout.json b/2025/Homogeneous Dynamics Space for Heterogeneous Humans/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c9b38c74f6f3f083bb8a7eed0883548b6738175a --- /dev/null +++ b/2025/Homogeneous Dynamics Space for Heterogeneous Humans/layout.json @@ -0,0 +1,10485 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 123, + 103, + 488, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 103, + 488, + 121 + ], + "spans": [ + { + "bbox": [ + 123, + 103, + 488, + 121 + ], + "type": "text", + "content": "Homogeneous Dynamics Space for Heterogeneous Humans" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 73, + 142, + 541, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 142, + 541, + 159 + ], + "spans": [ + { + "bbox": [ + 73, + 142, + 541, + 159 + ], + "type": "text", + "content": "Xinpeng Liu" + }, + { + "bbox": [ + 73, + 142, + 541, + 159 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 73, + 142, + 541, + 159 + ], + "type": "text", + "content": ", Junxuan Liang" + }, + { + "bbox": [ + 73, + 142, + 541, + 159 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 73, + 142, + 541, + 159 + ], + "type": "text", + "content": ", Chenshuo Zhang" + }, + { + "bbox": [ + 73, + 142, + 541, + 159 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 73, + 142, + 541, + 159 + ], + "type": "text", + "content": ", Zixuan Cai" + }, + { + "bbox": [ + 73, + 142, + 541, + 159 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 73, + 142, + 541, + 159 + ], + "type": "text", + "content": ", Cewu Lu" + }, + { + "bbox": [ + 73, + 142, + 541, + 159 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 73, + 142, + 541, + 159 + ], + "type": "text", + "content": ", Yong-Lu Li" + }, + { + "bbox": [ + 73, + 142, + 541, + 159 + ], + "type": "inline_equation", + "content": "^{1,2*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 82, + 159, + 528, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 159, + 528, + 200 + ], + "spans": [ + { + "bbox": [ + 82, + 159, + 528, + 200 + ], + "type": "text", + "content": "1Shanghai Jiao Tong University, 2Shanghai Innovation Institute, 3Soochow University xinpengliu0907@gmail.com, zxcai@stu.suda.edu.cn, {whitefork, zhangchenshuo, lucewu, yonglu.li}@sjtu.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "spans": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 252, + 296, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 252, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 54, + 252, + 296, + 529 + ], + "type": "text", + "content": "Analyses of human motion kinematics have achieved tremendous advances. However, the production mechanism, known as human dynamics, is still undercovered. In this paper, we aim to push the understanding of data-driven human dynamics forward. We identify a major obstacle to this as the heterogeneity of existing human motion understanding efforts. Specifically, heterogeneity exists in not only the diverse kinematics representations and hierarchical dynamics representations but also the data from different domains, namely biomechanics and reinforcement learning. With an in-depth analysis of the existing heterogeneity, we propose to emphasize the beneath homogeneity: all of them represent the homogeneous fact of human motion, though from different perspectives. Given this, we propose Homogeneous Dynamics Space (HDyS) as a fundamental space for human dynamics by aggregating heterogeneous data and training a homogeneous latent space with inspiration from the inverse-forward dynamics procedure. HDyS achieves decent mapping between human kinematics and dynamics by leveraging the heterogeneous representations and datasets. We demonstrate the feasibility of HDyS with extensive experiments and applications. The project page is https://foruck.github.io/HDyS." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 552, + 135, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 135, + 564 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 135, + 564 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 572, + 295, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 572, + 295, + 681 + ], + "spans": [ + { + "bbox": [ + 55, + 572, + 295, + 681 + ], + "type": "text", + "content": "Analyses on human motion have a wide range of applications, including animation [11, 60], healthcare [10, 55], and robotics [57, 70]. The computer vision community has made tremendous progress in understanding human kinematics with tasks like human reconstruction [24, 27, 34], action recognition [25, 26, 28, 29, 51], and motion generation [30, 32, 60]. However, the production mechanism hidden beneath human motion, known as human dynamics, is still limitedly explored." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 681, + 295, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 681, + 295, + 693 + ], + "spans": [ + { + "bbox": [ + 67, + 681, + 295, + 693 + ], + "type": "text", + "content": "In this paper, we aim to push the understanding of human" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 227, + 552, + 363 + ], + "blocks": [ + { + "bbox": [ + 317, + 227, + 552, + 363 + ], + "lines": [ + { + "bbox": [ + 317, + 227, + 552, + 363 + ], + "spans": [ + { + "bbox": [ + 317, + 227, + 552, + 363 + ], + "type": "image", + "image_path": "e12f39a53841ebb868ad4c4bbeda49a01aa9df498208085ca756edd870cc5d3f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 374, + 555, + 397 + ], + "lines": [ + { + "bbox": [ + 313, + 374, + 555, + 397 + ], + "spans": [ + { + "bbox": [ + 313, + 374, + 555, + 397 + ], + "type": "text", + "content": "Figure 1. The representation heterogeneity exists for both kinematics and dynamics." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 412, + 555, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 412, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 412, + 555, + 472 + ], + "type": "text", + "content": "dynamics forward with data-driven methodologies. Specifically, we pursue to build a bidirectional mapping between human kinematics and dynamics. We identify the major obstacle to this as the heterogeneity of existing human motion understanding efforts, which could be two-fold." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 475, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 556, + 715 + ], + "type": "text", + "content": "First, representation heterogeneity exists for both kinematics and dynamics, as shown in Figure 1. A typical optical motion capture pipeline, which produces most existing available data for human kinematics [20, 40], involves tracking the markers placed on the surface of humans. Then, different human template models are fitted to the tracked markers, resulting in parameters of musculoskeletal models [52] for biomechanics uses or SMPL-like models [36] for general CV/CG uses. Skeletons with pre-defined kinematic trees are sometimes preferred as a simple and intuitive representation. Different datasets tend to include only representations preferred by their designed goal. Besides the kinematics representation heterogeneity, the dynamics representation heterogeneity exists more severely. Human dynamics function in a hierarchical manner. The closest dynamics hierarchy to the kinematics is the joint torques, which are further produced by muscle forces from muscle actions according to the muscle contraction dynamics. Furthermore, electric signals from the neural system activate the muscle, which surface electromyogra" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 148, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 148, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 148, + 712 + ], + "type": "text", + "content": "*Corresponding authors." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27782" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": "phy (sEMG) could detect. Effective conversions between these different representations are not always available. For example, for kinematics representations, the conversion between the musculoskeletal model and SMPL becomes available in very recent advances [20], which requires time-consuming optimization. For dynamics, the bi-directional mapping between joint torques, muscle actions, and electric signals is even more difficult." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 171, + 296, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 171, + 296, + 481 + ], + "spans": [ + { + "bbox": [ + 56, + 171, + 296, + 481 + ], + "type": "text", + "content": "Second, domain heterogeneity also exists. One crucial characteristic of human dynamics is the difficulty of non-intrusive measurement. In the biomechanics community, optimization-based methods are adopted to solve the Euler-Newton equation for joint torques [64] and muscle actions [53]. For electric signals, it is typically measured by sEMG sensors [3]. In the meantime, from the learning community, there emerge efforts [33] utilizing Reinforcement Learning (RL) to imitate existing motion captures with physically simulated humanoids and record the simulated human dynamics data in a fully synthetic manner. These different data sources introduce domain heterogeneity in three aspects. First, biomechanics data are more deeply rooted in reality with well-defined computation models and real human motions. Instead, RL data could diverge from real-world situations with unnatural motion imitations and inaccurate physics simulations, known as the sim2real domain gap. Second, biomechanics data, due to the optimization-based manner, is sensitive to variations and usually limited to strictly controlled laboratory setups with relatively simple motions, resulting in limited motion domain coverage. In contrast, RL data cover a wide range of motion with more diversity. Third, biomechanics data and RL data tend to adopt heterogeneous kinematics representations with different kinematic trees, making it hard to transfer directly from one domain to another." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 483, + 295, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 295, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 295, + 639 + ], + "type": "text", + "content": "Despite the heterogeneity, we emphasize the homogeneous essence behind it: all of them represent the homogeneous fact of human motion, each from a different perspective. Given this, we demonstrate the feasibility of unifying these heterogeneous human representations and exploiting the underlying homogeneous knowledge of human dynamics. Low-level Cartesian kinematics representations, like markers and joints, are less heterogeneous than representations like joint angles. Also, though joint torques, muscle actions, and electric signals are not directly transferable to each other, they could share similar motor knowledge like coordination. To this end, we propose Homogeneous Dynamics Space for heterogeneous humans (HDyS)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": "To achieve this, we aggregate human dynamics data from both RL [33] and biomechanics [3, 53, 65] with different kinematics and dynamics representations, covering human dynamics hierarchies, including joint torques, muscle actions, and electric signals. Then, HDyS is designed as an aggregation of multiple auto-encoders corresponding to the" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "text", + "content": "inverse-forward dynamics procedure. We supervise HDyS with reconstruction and alignment losses. Our proposed HDyS has two major merits. First, by unifying heterogeneous human representation in the same latent space, it generalizes across different representations seamlessly while taking advantage of each. Second, by aggregating largescale heterogeneous motion data, it is empowered with general human motor knowledge from a wide span of motion, functioning as a reusable knowledge source for downstream dynamics-related applications. To demonstrate the efficacy of HDyS, we first evaluate it on human inverse dynamics. Then, we showcase how HDyS could facilitate downstream dynamics-related applications like ground reaction force prediction, biomechanics human simulation, and physics-simulated character control." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 251, + 555, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 555, + 347 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 555, + 347 + ], + "type": "text", + "content": "To summarize, our contribution includes: 1) We analyzed the heterogeneity issue that hinders an in-depth understanding of human dynamics. 2) We highlighted the homogeneity beneath the heterogeneity and proposed a fundamental reusable space HDyS for human dynamics by unifying the heterogeneity. 3) We demonstrated the feasibility of digging homogeneity out from heterogeneity with extensive experiments and applications of HDyS." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 357, + 406, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 357, + 406, + 369 + ], + "spans": [ + { + "bbox": [ + 314, + 357, + 406, + 369 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 376, + 421, + 389 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 376, + 421, + 389 + ], + "spans": [ + { + "bbox": [ + 314, + 376, + 421, + 389 + ], + "type": "text", + "content": "2.1. Human Dynamics" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 396, + 555, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 555, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 555, + 552 + ], + "type": "text", + "content": "By human dynamics, we mean the production mechanism of human motion, which the biomechanics community has actively explored. To produce a certain motion, neural commands are sent to activate the muscles. After receiving the activation signals, muscles contract and produce muscle forces. Multiple muscle forces form the joint torques according to certain musculoskeletal geometry, and the joint torques drive the accelerations that accumulate into movements. Thus, understanding human dynamics typically involves two heterogeneous hierarchies: joint torques and muscle activations. However, both are hard to measure nonintrusively. In the literature, to obtain them, an optimization problem is typically introduced as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 328, + 558, + 553, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 558, + 553, + 586 + ], + "spans": [ + { + "bbox": [ + 328, + 558, + 553, + 586 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\min \\| a \\|, s. t. 0 \\leq a \\leq 1, \\tau = A (q) F (a), \\\\ M (q) \\ddot {q} + C (q, \\dot {q}) + G (q) = J \\lambda + \\tau . \\tag {1} \\\\ \\end{array}", + "image_path": "ce0f766617dd8beb7ac0ab96614d9a29f47fd85bfe62fd5d7a73f78c0fa3c3ec.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": "with the generalized human inertia matrix " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "M(q)" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " w.r.t. generalized coordinate " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": ", Coriolis and centrifugal forces " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "C(q,\\dot{q})" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": ", gravity " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "G(q)" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": ", Jacobian matrix " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " mapping external forces " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " to the generalized coordinates, and muscle activations " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": ". The joint torques " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " could be obtained by " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\tau = A(q)F(a)" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "A(q)" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " maps muscle forces into joint torques and " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "F(a)" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " maps " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " into muscle forces usually with the hill-type function [15, 74]. Mature software [4, 5, 63] were developed for this purpose. However, the optimization quality is tightly bonded to the precision of external" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27783" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 312 + ], + "type": "text", + "content": "force measurement, which could be expensive. Therefore, the applications are mostly limited to simple motions like gaits in laboratory settings. Some efforts exploited wearable devices [22, 23] for more general applications. In addition, the optimization is deeply coupled with the adopted human models [36, 52], which vary with the application preferences. Fitting raw motion capture data to a specific human model could be unstable and time-consuming. The conversion between different human models is also nontrivial even with recent advances [20]. This way, transferring dynamics from one human model to another is limitedly explored. Besides torques and muscle actions that are typically obtained via optimization, Surface Electromyography (sEMG) is adopted as an indirect representation of human dynamics which could be directly measured. Efforts were made to build the accurate mapping from sEMG to muscle actions [19, 66], joint torques [54, 79], and human poses [35, 55]. Though progress was made, sEMG patterns might suffer from noises and vary drastically among subjects [61], hindering the generalization." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 322, + 239, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 322, + 239, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 322, + 239, + 335 + ], + "type": "text", + "content": "2.2. Learning-based Human Dynamics" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 342, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 295, + 628 + ], + "type": "text", + "content": "Joint Torques. Early efforts were made on ML-based joint torque analysis for certain human body parts [17, 43, 69]. Lv et al. [39] developed a Gaussian mixture framework for whole-body joint torque estimation. Other architectures like k-nearest neighbor-based regression [75, 77], random forests [76], and neural networks [78] were also adopted for the estimation of joint torques. However, most of these efforts suffered from limited data scale, which hampered learning methods from exploiting their full potential. The recent emergence of AddBiomechanics [65], which aggregated multiple biomechanics datasets, considerably boosted the data scale. However, most of the collected sequences contained only gaits with limited diversity. Another line of work adopted reinforcement learning to simulate motion in physics simulators, and the joint torques could be obtained in simulation. Though generalizability was limited at first [1, 49, 50, 67, 68], emerging efforts [37, 38, 73] managed to replicate a wide span of motion in simulators. The paradigm is further incorporated into recent MoCap systems for simultaneous estimation of motion and the joint torques [8, 12, 16, 62, 72, 80]. However, due to the involvement of simulators, the sim2real gap exists. The learned dynamics could be restricted to certain simulators or human models, which might diverge from real humans." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 630, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 294, + 713 + ], + "type": "text", + "content": "Muscle Actions. sEMG is usually adopted as a proxy measurement of muscle actions and has been extensively studied in biomechanics [3, 6, 7, 14, 41, 44, 48]. There have been efforts delving into predicting sEMG signals given either joint torques [31, 56, 58], goniometers [59], motion captures [18, 45, 71], videos [3, 48], or point clouds [46]. However, each effort could be small in data scale, mo" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "content": "tion variations, and muscle coverage. Unifying these efforts could be hindered by the heterogeneous settings and data formats adopted, thus under-covered. Recently, musculoskeletal human simulation has gained attention. Multiple efficient simulators were developed [2, 9, 81], opening the potential for more accessible muscle action analysis. Specifically, MinT [53] was proposed by simulating motion sequences from AMASS [40] in OpenSim [5] and attaching simulated muscle actions to the raw motion sequence, enabling muscle action learning in scale." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 205, + 370, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 370, + 217 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 370, + 217 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 227, + 553, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 227, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 313, + 227, + 553, + 275 + ], + "type": "text", + "content": "We introduce the proposed Homogeneous Dynamics Space (HDyS). We first cover the involved kinematics and dynamics representations in Section 3.1, 3.2. Then, the model architecture and designed losses are introduced in Section 3.3." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 285, + 468, + 297 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 468, + 297 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 468, + 297 + ], + "type": "text", + "content": "3.1. Kinematics Representations" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 304, + 553, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 553, + 363 + ], + "type": "text", + "content": "As shown in Fig. 2, we adopt four types of kinematics representations: Cartesian representations of markers and skeleton key-points, joint angles of Rajagopal's model [52] from the biomechanics community, and SMPL [36] which is widely adopted for CV and CG applications." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 365, + 554, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 554, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 554, + 508 + ], + "type": "text", + "content": "Markers, placed on the human body surfaces, are typically the raw data for optical motion capture. In practice, most other representations are calculated by fitting certain human prior models to the marker observations, making it easy to obtain for heterogeneous datasets and representations. Thus, markers are expected to be a generalizable representation across heterogeneous datasets, though they could also be rather low-level and thus hard to learn. We define the marker representation at timestamp " + }, + { + "bbox": [ + 313, + 365, + 554, + 508 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 365, + 554, + 508 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 365, + 554, + 508 + ], + "type": "inline_equation", + "content": "x_{m}^{t} = (m^{t},\\dot{m}^{t},\\dot{m})\\in \\mathcal{R}^{N_{m}\\times 9}" + }, + { + "bbox": [ + 313, + 365, + 554, + 508 + ], + "type": "text", + "content": ", which is composed of marker Cartesian coordinates " + }, + { + "bbox": [ + 313, + 365, + 554, + 508 + ], + "type": "inline_equation", + "content": "m^t" + }, + { + "bbox": [ + 313, + 365, + 554, + 508 + ], + "type": "text", + "content": ", finite-differentiated velocities " + }, + { + "bbox": [ + 313, + 365, + 554, + 508 + ], + "type": "inline_equation", + "content": "\\dot{m}^t" + }, + { + "bbox": [ + 313, + 365, + 554, + 508 + ], + "type": "text", + "content": " and finite-differentiated accelerations " + }, + { + "bbox": [ + 313, + 365, + 554, + 508 + ], + "type": "inline_equation", + "content": "\\dot{m}^t" + }, + { + "bbox": [ + 313, + 365, + 554, + 508 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 365, + 554, + 508 + ], + "type": "inline_equation", + "content": "N_{m}" + }, + { + "bbox": [ + 313, + 365, + 554, + 508 + ], + "type": "text", + "content": " markers." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 510, + 554, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 554, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 554, + 617 + ], + "type": "text", + "content": "Skeletal Keypoints, compared to markers, are less generalizable due to their reliance on pre-defined kinematic trees. However, with its Cartesian coordinates, common sense of human topology, and easy access, the gap between different kinematic trees can be mitigated. We define the joint representation at timestamp " + }, + { + "bbox": [ + 313, + 510, + 554, + 617 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 510, + 554, + 617 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 510, + 554, + 617 + ], + "type": "inline_equation", + "content": "x_{k}^{t} = (k^{t},\\dot{k}^{t},\\dot{k})\\in \\mathcal{R}^{N_{k}\\times 9}" + }, + { + "bbox": [ + 313, + 510, + 554, + 617 + ], + "type": "text", + "content": ", which is composed of joint coordinates " + }, + { + "bbox": [ + 313, + 510, + 554, + 617 + ], + "type": "inline_equation", + "content": "k^t" + }, + { + "bbox": [ + 313, + 510, + 554, + 617 + ], + "type": "text", + "content": ", finite-differentiated velocities " + }, + { + "bbox": [ + 313, + 510, + 554, + 617 + ], + "type": "inline_equation", + "content": "\\dot{k}^t" + }, + { + "bbox": [ + 313, + 510, + 554, + 617 + ], + "type": "text", + "content": " and finite-differentiated accelerations " + }, + { + "bbox": [ + 313, + 510, + 554, + 617 + ], + "type": "inline_equation", + "content": "\\dot{k}^t" + }, + { + "bbox": [ + 313, + 510, + 554, + 617 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 510, + 554, + 617 + ], + "type": "inline_equation", + "content": "N_{k}" + }, + { + "bbox": [ + 313, + 510, + 554, + 617 + ], + "type": "text", + "content": " skeletal joints." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "type": "text", + "content": "Joint Angles are preferred for clinical analyses which more faithfully preserves the biomechanics information of human kinematics. We adopt the Rajagopal's model [52] used in AddBiomechanics [65], and define the joint angle representation at timestamp " + }, + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "type": "inline_equation", + "content": "x_{a}^{t} = (a^{t},\\dot{a}^{t},\\dot{a})\\in \\mathcal{R}^{3N_{j}}" + }, + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "type": "text", + "content": ", containing joint angles " + }, + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "type": "inline_equation", + "content": "a^t" + }, + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "type": "text", + "content": ", finite-differentiated velocities " + }, + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\dot{a}^t" + }, + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "type": "text", + "content": " and accelerations " + }, + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\dot{a}^t" + }, + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "type": "inline_equation", + "content": "N_{j}" + }, + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "type": "text", + "content": " joints in Rajagopal's model (only 23 lower-body joints are used)." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27784" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 294, + 272 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 294, + 272 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 294, + 272 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 294, + 272 + ], + "type": "image", + "image_path": "b98c47e8f12e3bb45e35434811d4eea3e2a9dc2220e3b5c3e3ef708495ada98f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 94, + 281, + 256, + 293 + ], + "lines": [ + { + "bbox": [ + 94, + 281, + 256, + 293 + ], + "spans": [ + { + "bbox": [ + 94, + 281, + 256, + 293 + ], + "type": "text", + "content": "Figure 2. The overall architecture of HDyS." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 299, + 295, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 299, + 295, + 384 + ], + "spans": [ + { + "bbox": [ + 54, + 299, + 295, + 384 + ], + "type": "text", + "content": "SMPL [36] is a human parameter model widely adopted in the CV/CG community. We define the SMPL representation at timestamp " + }, + { + "bbox": [ + 54, + 299, + 295, + 384 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 54, + 299, + 295, + 384 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 54, + 299, + 295, + 384 + ], + "type": "inline_equation", + "content": "x_{s}^{t} = (s^{t},\\dot{s}^{t},\\dot{s})\\in \\mathcal{R}^{3N_{j}}" + }, + { + "bbox": [ + 54, + 299, + 295, + 384 + ], + "type": "text", + "content": ", which is composed of SMPL parameters " + }, + { + "bbox": [ + 54, + 299, + 295, + 384 + ], + "type": "inline_equation", + "content": "s^t" + }, + { + "bbox": [ + 54, + 299, + 295, + 384 + ], + "type": "text", + "content": ", finite-differentiated velocities " + }, + { + "bbox": [ + 54, + 299, + 295, + 384 + ], + "type": "inline_equation", + "content": "\\dot{s}^t" + }, + { + "bbox": [ + 54, + 299, + 295, + 384 + ], + "type": "text", + "content": " and finite-differentiated accelerations " + }, + { + "bbox": [ + 54, + 299, + 295, + 384 + ], + "type": "inline_equation", + "content": "\\dot{s}^t" + }, + { + "bbox": [ + 54, + 299, + 295, + 384 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 54, + 299, + 295, + 384 + ], + "type": "inline_equation", + "content": "N_{j}" + }, + { + "bbox": [ + 54, + 299, + 295, + 384 + ], + "type": "text", + "content": " joints of SMPL (75 used, 3 translational joints at root and 72 revolving joints)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 392, + 203, + 404 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 392, + 203, + 404 + ], + "spans": [ + { + "bbox": [ + 55, + 392, + 203, + 404 + ], + "type": "text", + "content": "3.2. Dynamics Representations" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 410, + 295, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 410, + 295, + 434 + ], + "spans": [ + { + "bbox": [ + 54, + 410, + 295, + 434 + ], + "type": "text", + "content": "We adopt three different types of dynamic representations: joint torques, muscle actions, and sEMGs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 434, + 295, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 434, + 295, + 506 + ], + "spans": [ + { + "bbox": [ + 54, + 434, + 295, + 506 + ], + "type": "text", + "content": "Joint Torques are the net torques exerted at each joint to drive the motion, represented as " + }, + { + "bbox": [ + 54, + 434, + 295, + 506 + ], + "type": "inline_equation", + "content": "\\tau_{t} \\in \\mathcal{R}^{N_{t}}" + }, + { + "bbox": [ + 54, + 434, + 295, + 506 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 54, + 434, + 295, + 506 + ], + "type": "inline_equation", + "content": "N_{t}" + }, + { + "bbox": [ + 54, + 434, + 295, + 506 + ], + "type": "text", + "content": " is coupled with the number of joints " + }, + { + "bbox": [ + 54, + 434, + 295, + 506 + ], + "type": "inline_equation", + "content": "N_{j}" + }, + { + "bbox": [ + 54, + 434, + 295, + 506 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 54, + 434, + 295, + 506 + ], + "type": "inline_equation", + "content": "N_{t} = N_{j} - 6" + }, + { + "bbox": [ + 54, + 434, + 295, + 506 + ], + "type": "text", + "content": ", with the 6-DoF free root joint unactuated. For clarity, we denote the joint torques corresponding to Rajagopal's model as " + }, + { + "bbox": [ + 54, + 434, + 295, + 506 + ], + "type": "inline_equation", + "content": "\\tau_{tr}" + }, + { + "bbox": [ + 54, + 434, + 295, + 506 + ], + "type": "text", + "content": " and those corresponding to SMPL as " + }, + { + "bbox": [ + 54, + 434, + 295, + 506 + ], + "type": "inline_equation", + "content": "\\tau_{ts}" + }, + { + "bbox": [ + 54, + 434, + 295, + 506 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 506, + 295, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 506, + 295, + 591 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 295, + 591 + ], + "type": "text", + "content": "Muscle Actions represent the activation level of each muscle, resulting in the muscle forces that produce joint torques. We denote muscle actions as " + }, + { + "bbox": [ + 55, + 506, + 295, + 591 + ], + "type": "inline_equation", + "content": "\\tau_{m} \\in \\mathcal{R}^{N_{m}}" + }, + { + "bbox": [ + 55, + 506, + 295, + 591 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 506, + 295, + 591 + ], + "type": "inline_equation", + "content": "N_{m}" + }, + { + "bbox": [ + 55, + 506, + 295, + 591 + ], + "type": "text", + "content": " muscles. sEMG detects the electric potential generated on the body surface by activated muscle cells, which is closely related to muscle actions. We represent sEMGs at timestamp " + }, + { + "bbox": [ + 55, + 506, + 295, + 591 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 506, + 295, + 591 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 506, + 295, + 591 + ], + "type": "inline_equation", + "content": "\\tau_{e}^{t} \\in \\mathcal{R}^{N_{e}}" + }, + { + "bbox": [ + 55, + 506, + 295, + 591 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 506, + 295, + 591 + ], + "type": "inline_equation", + "content": "N_{e}" + }, + { + "bbox": [ + 55, + 506, + 295, + 591 + ], + "type": "text", + "content": " sEMG channels." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 598, + 224, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 598, + 224, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 598, + 224, + 612 + ], + "type": "text", + "content": "3.3. Homogeneous Dynamics Space" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 617, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 689 + ], + "type": "text", + "content": "With the heterogeneous representations, we construct our homogeneous dynamics space (HDyS) as an aggregation of multiple auto-encoders corresponding to the inverseforward dynamics procedure. Then, reconstruction and alignment losses are adopted for supervision. For simplicity, all superscripts " + }, + { + "bbox": [ + 55, + 617, + 295, + 689 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 617, + 295, + 689 + ], + "type": "text", + "content": " are omitted." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "Inverse-Dynamics Auto-Encoder (IDAE) encodes the kinematics into the latent space and decodes the dynamics" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": "from the latent. For markers " + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "inline_equation", + "content": "x_{m}" + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": " and skeletal key points " + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "inline_equation", + "content": "x_{k}" + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": ", we adopt three-layer transformer encoders with no positional embedding to encode them into latent " + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "inline_equation", + "content": "z_{m}, z_{k} \\in \\mathcal{R}^{d}" + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": " of dimension " + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": ". This enables the encoding of an arbitrary number of markers/points. For joint angles " + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "inline_equation", + "content": "x_{a}" + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": " and SMPL parameters " + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "inline_equation", + "content": "x_{s}" + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": ", we adopt simple three-layer MLPs to encode them into latent " + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "inline_equation", + "content": "z_{a}, z_{s} \\in \\mathcal{R}^{d}" + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": ". The decoder is designed as a shared transformer followed by separate MLP regression heads. The transformer takes latent vectors from consecutive frames as input, outputting per-frame latent vectors refined with temporal contexts. Then, separate MLP decoders decode the per-frame dynamics as " + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "inline_equation", + "content": "\\hat{\\tau}_{tr}, \\hat{\\tau}_{ts}, \\hat{\\tau}_{m}, \\hat{\\tau}_{e}" + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 220, + 555, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 220, + 555, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 220, + 555, + 474 + ], + "type": "text", + "content": "Forward-Dynamics Auto-Encoder (FDAE) encodes dynamics and the kinematics except for accelerations (denoted as " + }, + { + "bbox": [ + 313, + 220, + 555, + 474 + ], + "type": "inline_equation", + "content": "\\tilde{x}" + }, + { + "bbox": [ + 313, + 220, + 555, + 474 + ], + "type": "text", + "content": ") into the latent space and decodes the accelerations. Similar to IDAE, markers " + }, + { + "bbox": [ + 313, + 220, + 555, + 474 + ], + "type": "inline_equation", + "content": "\\tilde{x}_m" + }, + { + "bbox": [ + 313, + 220, + 555, + 474 + ], + "type": "text", + "content": " and skeletal key-points " + }, + { + "bbox": [ + 313, + 220, + 555, + 474 + ], + "type": "inline_equation", + "content": "\\tilde{x}_k" + }, + { + "bbox": [ + 313, + 220, + 555, + 474 + ], + "type": "text", + "content": " are encoded by transformers as " + }, + { + "bbox": [ + 313, + 220, + 555, + 474 + ], + "type": "inline_equation", + "content": "\\tilde{z}_m,\\tilde{z}_k" + }, + { + "bbox": [ + 313, + 220, + 555, + 474 + ], + "type": "text", + "content": ", while joint angles " + }, + { + "bbox": [ + 313, + 220, + 555, + 474 + ], + "type": "inline_equation", + "content": "\\tilde{x}_a" + }, + { + "bbox": [ + 313, + 220, + 555, + 474 + ], + "type": "text", + "content": " and SMPL parameters " + }, + { + "bbox": [ + 313, + 220, + 555, + 474 + ], + "type": "inline_equation", + "content": "\\tilde{x}_s" + }, + { + "bbox": [ + 313, + 220, + 555, + 474 + ], + "type": "text", + "content": " are encoded by MLPs as " + }, + { + "bbox": [ + 313, + 220, + 555, + 474 + ], + "type": "inline_equation", + "content": "\\tilde{z}_a,\\tilde{z}_p" + }, + { + "bbox": [ + 313, + 220, + 555, + 474 + ], + "type": "text", + "content": ". Then, joint torques " + }, + { + "bbox": [ + 313, + 220, + 555, + 474 + ], + "type": "inline_equation", + "content": "\\tau_{tr},\\tau_{ts}" + }, + { + "bbox": [ + 313, + 220, + 555, + 474 + ], + "type": "text", + "content": ", muscle activations " + }, + { + "bbox": [ + 313, + 220, + 555, + 474 + ], + "type": "inline_equation", + "content": "\\tau_m" + }, + { + "bbox": [ + 313, + 220, + 555, + 474 + ], + "type": "text", + "content": ", and sEMGs " + }, + { + "bbox": [ + 313, + 220, + 555, + 474 + ], + "type": "inline_equation", + "content": "\\tau_e" + }, + { + "bbox": [ + 313, + 220, + 555, + 474 + ], + "type": "text", + "content": " are also encoded by separate MLP encoders as dynamics latent vectors " + }, + { + "bbox": [ + 313, + 220, + 555, + 474 + ], + "type": "inline_equation", + "content": "\\tilde{z}^{tr},\\tilde{z}^{ts},\\tilde{z}^m,\\tilde{z}^e" + }, + { + "bbox": [ + 313, + 220, + 555, + 474 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 220, + 555, + 474 + ], + "type": "inline_equation", + "content": "\\tilde{z}_m,\\tilde{z}_k,\\tilde{z}_a,\\tilde{z}_p" + }, + { + "bbox": [ + 313, + 220, + 555, + 474 + ], + "type": "text", + "content": " is then concatenated with available dynamics latent vectors, fed into a shared MLP composer, resulted in latent vectors " + }, + { + "bbox": [ + 313, + 220, + 555, + 474 + ], + "type": "inline_equation", + "content": "z^{\\cdot}" + }, + { + "bbox": [ + 313, + 220, + 555, + 474 + ], + "type": "text", + "content": ", with subscripts representing the dynamics. Although arbitrary combinations are feasible, in practice, only one dynamic latent vector typically exists due to the data limitations. Finally, separate MLP decoders decode the latent vectors " + }, + { + "bbox": [ + 313, + 220, + 555, + 474 + ], + "type": "inline_equation", + "content": "z_{\\{kin\\}}^{(dyn)}" + }, + { + "bbox": [ + 313, + 220, + 555, + 474 + ], + "type": "text", + "content": " into skeletal key point accelerations " + }, + { + "bbox": [ + 313, + 220, + 555, + 474 + ], + "type": "inline_equation", + "content": "\\ddot{k}" + }, + { + "bbox": [ + 313, + 220, + 555, + 474 + ], + "type": "text", + "content": ", SMPL accelerations " + }, + { + "bbox": [ + 313, + 220, + 555, + 474 + ], + "type": "inline_equation", + "content": "\\ddot{s}" + }, + { + "bbox": [ + 313, + 220, + 555, + 474 + ], + "type": "text", + "content": ", and joint angle accelerations " + }, + { + "bbox": [ + 313, + 220, + 555, + 474 + ], + "type": "inline_equation", + "content": "\\ddot{a}" + }, + { + "bbox": [ + 313, + 220, + 555, + 474 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 220, + 555, + 474 + ], + "type": "inline_equation", + "content": "kin\\in \\{m,k,s,a\\}" + }, + { + "bbox": [ + 313, + 220, + 555, + 474 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 220, + 555, + 474 + ], + "type": "inline_equation", + "content": "dyn\\in \\{m,e,ts,tr\\}" + }, + { + "bbox": [ + 313, + 220, + 555, + 474 + ], + "type": "text", + "content": ". The marker accelerations are not predicted since their variable sizes could bring unnecessary complexity." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 477, + 554, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 477, + 554, + 513 + ], + "spans": [ + { + "bbox": [ + 313, + 477, + 554, + 513 + ], + "type": "text", + "content": "Loss Terms. To train HDyS, we adopt reconstruction losses and alignment losses. For reconstruction losses, we simply calculate the L1 loss for " + }, + { + "bbox": [ + 313, + 477, + 554, + 513 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 477, + 554, + 513 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 477, + 554, + 513 + ], + "type": "inline_equation", + "content": "\\ddot{a}" + }, + { + "bbox": [ + 313, + 477, + 554, + 513 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 369, + 529, + 555, + 544 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 529, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 369, + 529, + 555, + 544 + ], + "type": "interline_equation", + "content": "L _ {r e c o n} = \\| \\tau - \\hat {\\tau} \\| _ {1} + \\| \\ddot {a} - \\hat {\\dot {a}} \\|. \\tag {2}", + "image_path": "43127267a84b8255c4e3d77f12af5b6ae79601404f9c3205cf22e38206c32332.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 559, + 554, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 559, + 554, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 559, + 554, + 624 + ], + "type": "text", + "content": "For alignment losses, we align the latent vectors " + }, + { + "bbox": [ + 313, + 559, + 554, + 624 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 559, + 554, + 624 + ], + "type": "text", + "content": " of the same frame and separate " + }, + { + "bbox": [ + 313, + 559, + 554, + 624 + ], + "type": "inline_equation", + "content": "zs" + }, + { + "bbox": [ + 313, + 559, + 554, + 624 + ], + "type": "text", + "content": " of different frames using InfoNCE [47]. Given a batch of " + }, + { + "bbox": [ + 313, + 559, + 554, + 624 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 559, + 554, + 624 + ], + "type": "text", + "content": " frames, the obtained latent vectors is denoted as " + }, + { + "bbox": [ + 313, + 559, + 554, + 624 + ], + "type": "inline_equation", + "content": "Z = \\{z_{\\{kin\\}}, z_{\\{kin\\}}^{(dyn)}\\}" + }, + { + "bbox": [ + 313, + 559, + 554, + 624 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 313, + 559, + 554, + 624 + ], + "type": "inline_equation", + "content": "kin \\in \\{m, k, s, a\\}, dyn \\in \\{m, e, ts, tr\\}" + }, + { + "bbox": [ + 313, + 559, + 554, + 624 + ], + "type": "text", + "content": ". The loss is" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 319, + 639, + 555, + 673 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 639, + 555, + 673 + ], + "spans": [ + { + "bbox": [ + 319, + 639, + 555, + 673 + ], + "type": "interline_equation", + "content": "L _ {a l i g n} = \\sum_ {z _ {2}, z _ {2} \\in Z} \\sum_ {i = 1} ^ {B} - \\log \\left(\\frac {\\exp \\left(\\left\\langle z _ {1} ^ {(i)} , z _ {2} ^ {(i)} \\right\\rangle\\right)}{\\sum_ {j = 1} ^ {B} \\exp \\left(\\left\\langle z _ {1} ^ {(i)} , z _ {2} ^ {(j)} \\right\\rangle\\right)}\\right), \\tag {3}", + "image_path": "4f7d678bc1c00186e8dd3971d8cd15a2eb35bbb0c71cd0b4e14f5d182b0d9d8c.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 689, + 553, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 689, + 553, + 715 + ], + "type": "inline_equation", + "content": "(i),(j)" + }, + { + "bbox": [ + 313, + 689, + 553, + 715 + ], + "type": "text", + "content": " indicate the batch index. The overall loss is as " + }, + { + "bbox": [ + 313, + 689, + 553, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\alpha_{1}L_{recon} + \\alpha_{2}L_{align}" + }, + { + "bbox": [ + 313, + 689, + 553, + 715 + ], + "type": "text", + "content": " with coefficients " + }, + { + "bbox": [ + 313, + 689, + 553, + 715 + ], + "type": "inline_equation", + "content": "\\alpha_{1},\\alpha_{2}" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27785" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 81, + 93, + 527, + 285 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 92 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 92 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 92 + ], + "type": "text", + "content": "Table 1. Quantitative results of HDyS with ablation studies. For HDyS, the results of averaged predictions and the best prediction among all representations are reported." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 81, + 93, + 527, + 285 + ], + "lines": [ + { + "bbox": [ + 81, + 93, + 527, + 285 + ], + "spans": [ + { + "bbox": [ + 81, + 93, + 527, + 285 + ], + "type": "table", + "html": "
MethodsImDy mPJE↓ avg/bstAddBiomechanics mPJE↓ avg/bstMinTMiA PCC↑ avg/bst
RMSE↓ avg/bstPCC↑ avg/bstRMSE↓ avg/bst
ImDyS [33]0.63000.1626----
MiA[3]----13.3-
HDyS0.5765/0.46740.1189/0.12430.0614/0.06150.7420/0.740211.8/11.60.7232/0.7261
ImDy-only HDyS0.6854/0.5403-----
AddBiomechanics-only HDyS-0.1695/0.1691----
MinT-only HDyS--0.0637/0.06400.7179/0.7127--
MiA-only HDyS----13.6/13.50.6557/0.6421
HDyS w/o ImDy-0.1214/0.13860.0608/0.06170.7470/0.741715.8/15.60.6523/0.6497
HDyS w/o AddBiomechanics0.5787/0.4742-0.0616/0.06140.7408/0.738617.1/17.00.5769/0.5638
HDyS w/o MinT0.5730/0.46810.1197/0.1296--16.9/16.80.5213/0.5313
HDyS w/o MiA0.5890/0.47880.1200/0.12840.0616/0.06180.7395/0.7375--
HDyS w/o AMASS0.5797/0.47860.1217/0.13190.0613/0.06170.7419/0.738014.7/14.90.5704/0.5632
HDyS w/o Lalign0.6575/0.50190.1270/0.13290.0626/0.06300.7318/0.723813.7/13.40.6464/0.6402
HDyS w/o FDAE0.5776/0.48490.1198/0.12610.0617/0.06170.7388/0.737513.6/13.30.6517/0.6699
HDyS-32D0.7390/0.64500.1401/0.14200.0650/0.06480.6980/0.701016.7/16.30.5524/0.5614
HDyS-64D0.6505/0.54100.1295/0.13540.0627/0.06290.7272/0.725415.1/14.50.6034/0.6187
", + "image_path": "f40df5d834a56571d1001ca7b833c803a074e0f2f7d2e88a78f69be952bc115b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 289, + 137, + 303 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 137, + 303 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 137, + 303 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 315, + 118, + 327 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 315, + 118, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 315, + 118, + 327 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 339, + 296, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 339, + 296, + 411 + ], + "spans": [ + { + "bbox": [ + 55, + 339, + 296, + 411 + ], + "type": "text", + "content": "AddBiomechanics [64] contains over 50 hours of human motion data with joint torques from Nimble [63] simulation. We follow the setting in [64] with the armless part of AddBiomechanics. Markers, key points, and joint angles are used as kinematics representations, and joint torques corresponding to [52] represent the dynamics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 418, + 296, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 296, + 491 + ], + "type": "text", + "content": "Muscles in Times (MinT) [53] simulates part of AMASS in OpenSim to obtain the actions of 402 muscles. We randomly split them into 906 sequences for training and 227 sequences for testing. Markers, key points, and SMPL parameters are used as kinematics representations, and muscle actions are the dynamics representations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 497, + 296, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 497, + 296, + 569 + ], + "spans": [ + { + "bbox": [ + 55, + 497, + 296, + 569 + ], + "type": "text", + "content": "Muscles in Act (MiA) [3] consists of 12.8 hours of human exercise motion reconstructed from videos with VIBE [21] and sEMG data of 8 muscles. Following [3], we split them into 19,563 training sequences and 3,053 testing sequences. Markers and key points are kinematics representations, and sEMGs are dynamics representations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 574, + 296, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 574, + 296, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 296, + 647 + ], + "type": "text", + "content": "ImDy [33] adopted PHC [38] to imitate motion sequences from AMASS [40] and KIT [42], resulting in over 150 hours of human dynamics data, with 27,501 sequences for training and 3,055 sequences for evaluation. Markers, key points, and SMPL are kinematics representations, and SMPL torques are dynamics representations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": "AMASS [40] contains over 11,000 motion sequences represented in SMPL [36]. Though not paired with human dynamics, we included it as an extra knowledge base for training. Markers, key points, and SMPL parameters are used as kinematics representations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 289, + 447, + 302 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 447, + 302 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 447, + 302 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 308, + 555, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 555, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 555, + 559 + ], + "type": "text", + "content": "We adopt a compact design for HDyS with a latent dimension of 128. For joint angle and SMPL parameters, three-layer MLPs with hidden unit sizes of 256 and 128 are adopted as encoders, projecting the input to the 128-D latent space. For markers and key-points, three-layer transformer encoders with 2 heads are adopted as encoders. The transformer in the ID decoder is of 4 layers, 4 heads, and a dimension of 128. All MLPs in decoders are two-layer, with a hidden size of 32 for Rajagopal's torque, sEMG, joint angle accelerations, and key-point accelerations, and a hidden size of 64 for SMPL torque, SMPL accelerations, and muscle actions. For training, we adopt an AdamW optimizer with a learning rate of 1e-3 and a batch size of 9,600 frames for 1,000 epochs. Loss weights are set as " + }, + { + "bbox": [ + 313, + 308, + 555, + 559 + ], + "type": "inline_equation", + "content": "\\alpha_{1} = 0.01" + }, + { + "bbox": [ + 313, + 308, + 555, + 559 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 308, + 555, + 559 + ], + "type": "inline_equation", + "content": "\\alpha_{2} = 0.05" + }, + { + "bbox": [ + 313, + 308, + 555, + 559 + ], + "type": "text", + "content": ". Also, during training, a balanced sampling strategy is adopted to minimize the scale influence of different datasets. That is, we randomly sample 3,000 sequences per dataset for each training epoch. In this way, the comparison is fair in terms of the number of seen samples under different dataset settings. All experiments are conducted on a single NVIDIA Titan Xp GPU." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 566, + 471, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 566, + 471, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 566, + 471, + 578 + ], + "type": "text", + "content": "4.3. Results on Inverse Dynamics" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 584, + 478, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 584, + 478, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 584, + 478, + 597 + ], + "type": "text", + "content": "Metric. We report mPJE and RMSE as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 321, + 605, + 555, + 639 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 605, + 555, + 639 + ], + "spans": [ + { + "bbox": [ + 321, + 605, + 555, + 639 + ], + "type": "interline_equation", + "content": "m P J E _ {\\tau .} = \\frac {1}{J} \\sum_ {j = 1} ^ {J} | \\tau . - \\hat {\\tau}. | _ {2}, R M S E _ {\\tau .} = | \\tau . - \\hat {\\tau}. | _ {2}. \\tag {4}", + "image_path": "5a86d348f0f7bcea67786280baf9d389a52d0a8c17d9d18f6f5191d9c8f2cfa1.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 646, + 554, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 646, + 554, + 682 + ], + "spans": [ + { + "bbox": [ + 313, + 646, + 554, + 682 + ], + "type": "text", + "content": "PCC is also reported due to its invariance against scale and offset, which helps evaluate muscle actions and sEMG prediction. PCC is computed as" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 388, + 689, + 555, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 388, + 689, + 555, + 715 + ], + "type": "interline_equation", + "content": "P C C _ {\\tau_ {1}} = \\frac {\\operatorname {c o v} \\left(\\tau , \\hat {\\tau}\\right)}{\\sigma_ {\\tau_ {1}} \\sigma_ {\\hat {\\tau}}} \\tag {5}", + "image_path": "bcf09342c31545d835f13106efedcb037fd5d2620053d48a38aa77e3a87ed1b3.jpg" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27786" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "inline_equation", + "content": "cov(\\tau, \\hat{\\tau})" + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": " is the covariance of " + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "inline_equation", + "content": "\\hat{\\tau}" + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": ". indicates the standard deviation. For ImDy and AddBiomechanics, we report mPJE normalized by body weight. For MinT and MiA, RMSE and PCC are reported." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 140, + 173, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 140, + 173, + 152 + ], + "spans": [ + { + "bbox": [ + 55, + 140, + 173, + 152 + ], + "type": "text", + "content": "4.3.1 Quantitative results" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 161, + 296, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 161, + 296, + 245 + ], + "spans": [ + { + "bbox": [ + 55, + 161, + 296, + 245 + ], + "type": "text", + "content": "Quantitative results are shown in Tab. 1, where HDyS is reported with the averaged predictions from different input kinematics representations and the best predictions among all representations. Compared to previous methods [3, 33], the proposed HDyS provides superior performances with substantial improvements on all datasets. We further analyze the performance of HDyS with three questions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 246, + 296, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 246, + 296, + 257 + ], + "spans": [ + { + "bbox": [ + 55, + 246, + 296, + 257 + ], + "type": "text", + "content": "How do the heterogeneous datasets contribute to HDyS?" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 258, + 296, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 296, + 461 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 296, + 461 + ], + "type": "text", + "content": "We compare HDyS with its single-dataset variants and drop-one-out variants in Tab. 1. HDyS outperforms the corresponding single-dataset variants on all datasets, validating the existence of homogeneous human dynamics knowledge behind these heterogeneous datasets. According to the drop-one-out results, it is noticeable that datasets with similar dynamics representations are more cooperative. That is, muscle-related datasets (MiA and MinT) tend to benefit more from each other and less from torque-related datasets (AddBiomechanics and ImDy), and vice versa. Moreover, mutually harmful effects could be observed for ImDy and MinT. These are consistent with the gaps between these datasets: the sim2real gap from ImDy to others and the torque-to-muscle gap from AddBiomechanics and ImDy to MiA and MinT. AMASS, though not paired with dynamics information, is shown to be beneficial for inverse dynamics with its diverse and high-quality kinematics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 463, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 463, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 463, + 296, + 714 + ], + "type": "text", + "content": "The heterogeneous datasets introduce both increased data scale and heterogeneous knowledge. Given this, we try to investigate further the source of the improvement in Tab. 1. Trying to decompose the contributions of scale and heterogeneity, we compared the performance of the following three models on a single target test set in Tab. 2. HDyS-Single-50 denotes single-dataset HDyS with " + }, + { + "bbox": [ + 55, + 463, + 296, + 714 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 463, + 296, + 714 + ], + "type": "text", + "content": " of the data from the target dataset. HDyS-50/50 denotes HDyS using " + }, + { + "bbox": [ + 55, + 463, + 296, + 714 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 463, + 296, + 714 + ], + "type": "text", + "content": " of the data from the target dataset and " + }, + { + "bbox": [ + 55, + 463, + 296, + 714 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 463, + 296, + 714 + ], + "type": "text", + "content": " of the data from the other datasets, maintaining the same data scale as the target dataset. HDyS-Single represents the corresponding single-dataset variants of HDyS in Tab. 1. The performance of the best representation is reported on AddBiomechanics and MiA, considering their suitable scale and realistic nature. As shown, HDyS-50/50 consistently outperforms HDyS-Single-50, verifying the existence of homogeneous knowledge in heterogeneous data. Moreover, on AddBiomechanics, HDyS-50/50 even outperforms HDyS-Single, indicating that the diversity from heterogeneous data could sometimes benefit more than simply increasing seemingly homogeneous data." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 317, + 82, + 553, + 115 + ], + "blocks": [ + { + "bbox": [ + 333, + 71, + 534, + 81 + ], + "lines": [ + { + "bbox": [ + 333, + 71, + 534, + 81 + ], + "spans": [ + { + "bbox": [ + 333, + 71, + 534, + 81 + ], + "type": "text", + "content": "Table 2. Results on scale-heterogeneity decomposition." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 82, + 553, + 115 + ], + "lines": [ + { + "bbox": [ + 317, + 82, + 553, + 115 + ], + "spans": [ + { + "bbox": [ + 317, + 82, + 553, + 115 + ], + "type": "table", + "html": "
DatasetHDyS-Single-50HDyS-50/50HDyS-Single
AddBiomechanics mPJE ↓0.17070.12840.1695
MiA RMSE ↓16.214.513.5
", + "image_path": "0e16789ec12e9676c56efde538758cfe3f41ca2faff0edc590db2db3e7c2fdc5.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 317, + 129, + 553, + 186 + ], + "blocks": [ + { + "bbox": [ + 318, + 118, + 552, + 129 + ], + "lines": [ + { + "bbox": [ + 318, + 118, + 552, + 129 + ], + "spans": [ + { + "bbox": [ + 318, + 118, + 552, + 129 + ], + "type": "text", + "content": "Table 3. Results of different kinematic representations in HDyS." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 129, + 553, + 186 + ], + "lines": [ + { + "bbox": [ + 317, + 129, + 553, + 186 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 553, + 186 + ], + "type": "table", + "html": "
MethodsImDy mPJE↓AddBiomechanics mPJE↓MinTMiA
RMSE↓PCC↑RMSE↓PCC↑
HDys-marker0.81630.14550.06460.696812.30.7088
HDys-keypoint0.80840.13240.06400.710311.60.7261
HDys-angle-0.1243----
HDys-SMPL0.4674-0.06150.7402--
", + "image_path": "699fc81e0cac1bd36565cab198d4a647cc513fc92cd0901af8ac570d4080f877.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 321, + 195, + 447, + 289 + ], + "blocks": [ + { + "bbox": [ + 345, + 186, + 455, + 194 + ], + "lines": [ + { + "bbox": [ + 345, + 186, + 455, + 194 + ], + "spans": [ + { + "bbox": [ + 345, + 186, + 455, + 194 + ], + "type": "text", + "content": "Heterogeneous Representations" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 321, + 195, + 447, + 289 + ], + "lines": [ + { + "bbox": [ + 321, + 195, + 447, + 289 + ], + "spans": [ + { + "bbox": [ + 321, + 195, + 447, + 289 + ], + "type": "image", + "image_path": "3c9f68164f38b8075ff6174233b461c36845759336cac95042a508218c559d09.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 452, + 195, + 552, + 289 + ], + "blocks": [ + { + "bbox": [ + 463, + 186, + 548, + 194 + ], + "lines": [ + { + "bbox": [ + 463, + 186, + 548, + 194 + ], + "spans": [ + { + "bbox": [ + 463, + 186, + 548, + 194 + ], + "type": "text", + "content": "Heterogeneous Datasets" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 452, + 195, + 552, + 289 + ], + "lines": [ + { + "bbox": [ + 452, + 195, + 552, + 289 + ], + "spans": [ + { + "bbox": [ + 452, + 195, + 552, + 289 + ], + "type": "image", + "image_path": "1a9e8fb895845f39a4673e9f0c0ea54b30022429f10b1f2537ca9a573b8196be.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 316, + 292, + 447, + 365 + ], + "blocks": [ + { + "bbox": [ + 316, + 292, + 447, + 365 + ], + "lines": [ + { + "bbox": [ + 316, + 292, + 447, + 365 + ], + "spans": [ + { + "bbox": [ + 316, + 292, + 447, + 365 + ], + "type": "image", + "image_path": "c944f285a4618bfc85270fe41c207c63756ae4492db652941e34227b808d5c71.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 376, + 544, + 388 + ], + "lines": [ + { + "bbox": [ + 323, + 376, + 544, + 388 + ], + "spans": [ + { + "bbox": [ + 323, + 376, + 544, + 388 + ], + "type": "text", + "content": "Figure 3. Feature visualization for HDyS w/ and w/o " + }, + { + "bbox": [ + 323, + 376, + 544, + 388 + ], + "type": "inline_equation", + "content": "L_{align}" + }, + { + "bbox": [ + 323, + 376, + 544, + 388 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 452, + 292, + 552, + 365 + ], + "blocks": [ + { + "bbox": [ + 452, + 292, + 552, + 365 + ], + "lines": [ + { + "bbox": [ + 452, + 292, + 552, + 365 + ], + "spans": [ + { + "bbox": [ + 452, + 292, + 552, + 365 + ], + "type": "image", + "image_path": "81c5146fdcf80211a3b3c4ce4d5b1b782d035dda3aac2f149e57e487e741ca12.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 401, + 553, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 553, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 553, + 496 + ], + "type": "text", + "content": "How do the heterogeneous representations contribute to HDyS? We report the results from individual kinematics representations in Tab. 3. Typically, the representations in joint space like SMPL and joint angles produce more accurate predictions compared to Cartesian representations with considerable performance gaps, which explains the performance drop when taking the average for ImDy and MiA. The markers are the worst-performed representation of all." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 497, + 554, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 554, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 554, + 605 + ], + "type": "text", + "content": "In Tab. 1, it is noticeable that MinT and AddBiomechanics take advantage of averaging predictions from different representations. For AddBiomechanics, this might be because the best-performing representation is joint angles, which are not available in other datasets. Therefore, the averaging could effectively aggregate knowledge from other datasets with other representations. For MinT, it is observable that different representations produce close results, preventing the ensemble from being held back by markers." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": "Furthermore, we examine the alignment of heterogeneous representations by removing the alignment loss " + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "inline_equation", + "content": "L_{\\text{align}}" + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": " in Tab. 1. Considerable performance drops could be observed. The drop is consistent for both best and average performances, indicating the cross-modality alignment manages to aggregate knowledge from heterogeneous representations. We also visualize HDyS features w/ and w/o " + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "inline_equation", + "content": "L_{\\text{align}}" + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": " in Fig. 3. As shown, " + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "inline_equation", + "content": "L_{\\text{align}}" + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": " prevents the HDyS feature from being dominated by the representation hetero-" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27787" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 69, + 294, + 220 + ], + "blocks": [ + { + "bbox": [ + 57, + 69, + 294, + 220 + ], + "lines": [ + { + "bbox": [ + 57, + 69, + 294, + 220 + ], + "spans": [ + { + "bbox": [ + 57, + 69, + 294, + 220 + ], + "type": "image", + "image_path": "c338a3d62238cc2d69ce42e615c3b8fb8faee7744c9fbf3b7ca30145a7145f6f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 83, + 230, + 268, + 243 + ], + "lines": [ + { + "bbox": [ + 83, + 230, + 268, + 243 + ], + "spans": [ + { + "bbox": [ + 83, + 230, + 268, + 243 + ], + "type": "text", + "content": "Figure 4. Inverse dynamics visualization on ImDy." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 248, + 295, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 248, + 295, + 297 + ], + "spans": [ + { + "bbox": [ + 55, + 248, + 295, + 297 + ], + "type": "text", + "content": "geneity, encouraging the learning of homogeneous knowledge. Interestingly, though not intended to do so, " + }, + { + "bbox": [ + 55, + 248, + 295, + 297 + ], + "type": "inline_equation", + "content": "L_{\\text{align}}" + }, + { + "bbox": [ + 55, + 248, + 295, + 297 + ], + "type": "text", + "content": " also helps to mitigate the gap between datasets, alleviating the domain heterogeneity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 297, + 296, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 297, + 296, + 405 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 296, + 405 + ], + "type": "text", + "content": "How does the model design contribute to HDyS? We first evaluate the FDAE in Tab. 1. A substantial performance drop could be observed, demonstrating the benefit of informing HDyS with the forward dynamics procedure. Moreover, we also report HDyS with different dimensions in Tab. 1. Though HDyS's performance degrades with lower dimensions, it could still outperform some single-dataset variants with higher dimensions, validating the efficacy of heterogeneous datasets again." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 421, + 167, + 432 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 421, + 167, + 432 + ], + "spans": [ + { + "bbox": [ + 55, + 421, + 167, + 432 + ], + "type": "text", + "content": "4.3.2 Qualitative results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 440, + 295, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 440, + 295, + 583 + ], + "spans": [ + { + "bbox": [ + 55, + 440, + 295, + 583 + ], + "type": "text", + "content": "We visualize the inverse dynamics results of different datasets. As shown in Fig. 4-5, the joint torques could be faithfully reconstructed for different actions from either synthetic and jittering ImDy or realistic AddBiomechanics with the same HDyS. For the muscle actions shown in Fig. 6, HDyS produces reasonable estimations for lower-body muscles when walking. Moreover, as in the right half, though the lower-body kinematics are less significant, HDyS could capture the ankle muscle actions with high fidelity. The predictions are also coherent with GT for upper-body muscles like the internal oblique and levator scapulae. More details are in the appendix." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 592, + 218, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 592, + 218, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 592, + 218, + 604 + ], + "type": "text", + "content": "4.4. Results on Downstream Tasks" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 610, + 237, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 610, + 237, + 621 + ], + "spans": [ + { + "bbox": [ + 55, + 610, + 237, + 621 + ], + "type": "text", + "content": "4.4.1 Ground Reaction Force Estimation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "content": "We evaluate the effectiveness of HDyS on GRF estimation with GroundLink [13], which contains 1.5-hour motion with GRF recordings. We finetune HDyS with subjects 1-6 and evaluate it on subject 7. mPJE for GRF at both feet normalized by body weight is reported following Eq. 4. GroundLinkNet [13] and HDyS trained on GroundLink from scratch are also compared. As shown in Tab. 4, the" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 326, + 72, + 541, + 225 + ], + "blocks": [ + { + "bbox": [ + 326, + 72, + 541, + 225 + ], + "lines": [ + { + "bbox": [ + 326, + 72, + 541, + 225 + ], + "spans": [ + { + "bbox": [ + 326, + 72, + 541, + 225 + ], + "type": "image", + "image_path": "448c23ac1e11f1c1637d97dadcd0dad8e1c6c17e3b9eb388ada01297c71e3f12.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 228, + 549, + 247 + ], + "lines": [ + { + "bbox": [ + 318, + 228, + 549, + 247 + ], + "spans": [ + { + "bbox": [ + 318, + 228, + 549, + 247 + ], + "type": "text", + "content": "Figure 5. Inverse dynamics visualization on AddBiomechanics. Table 4. Results on GroundLink." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 317, + 251, + 553, + 289 + ], + "blocks": [ + { + "bbox": [ + 317, + 251, + 553, + 289 + ], + "lines": [ + { + "bbox": [ + 317, + 251, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 317, + 251, + 553, + 289 + ], + "type": "table", + "html": "
MethodsGroundLinkNetGroundLink-onlyHDySHDyS
L-Foot mPJE ↓0.07110.05840.0514
R-Foot mPJE ↓0.09120.07510.0694
", + "image_path": "6ee49300657fe733b2a8f4e46207aa15dffebbbaa11faed0dd7570b3e08bc783.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 328, + 317, + 541, + 355 + ], + "blocks": [ + { + "bbox": [ + 313, + 291, + 553, + 312 + ], + "lines": [ + { + "bbox": [ + 313, + 291, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 291, + 553, + 312 + ], + "type": "text", + "content": "Table 5. Biomechanical human simulation results reported in per-frame MSE." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 328, + 317, + 541, + 355 + ], + "lines": [ + { + "bbox": [ + 328, + 317, + 541, + 355 + ], + "spans": [ + { + "bbox": [ + 328, + 317, + 541, + 355 + ], + "type": "table", + "html": "
k12345
Optimized0.000630.19091.83062.01062.6027
HDyS0.000610.18601.71181.86512.2233
", + "image_path": "13134cead5f08e4039c34aecb3c752f4e8f56ce7129572086539a8892f41b74f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 365, + 554, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 554, + 425 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 554, + 425 + ], + "type": "text", + "content": "finetuned HDyS outperforms its counterparts, indicating the efficacy of the aggregated homogeneous knowledge. Also, the GroundLink-only HDyS outperforms GroundLinkNet, reflecting the feasibility of unifying heterogeneous representations for better dynamics knowledge." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 442, + 493, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 442, + 493, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 442, + 493, + 453 + ], + "type": "text", + "content": "4.4.2 Biomechanical Human Simulation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 461, + 554, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 554, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 554, + 568 + ], + "type": "text", + "content": "HDyS could also be adopted for biomechanical human simulation. We start with the armless Rajagopal's model [52] in Nimble [63]. Given a motion sequence, we first adopt HDyS to estimate the joint torques. Then, we use the predicted torques to reproduce the motion in Nimble. Starting from the current state, we feed the predicted torques for " + }, + { + "bbox": [ + 313, + 461, + 554, + 568 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 461, + 554, + 568 + ], + "type": "text", + "content": " steps and compare the simulated joint angles " + }, + { + "bbox": [ + 313, + 461, + 554, + 568 + ], + "type": "inline_equation", + "content": "\\hat{q}" + }, + { + "bbox": [ + 313, + 461, + 554, + 568 + ], + "type": "text", + "content": " with the real joint angles " + }, + { + "bbox": [ + 313, + 461, + 554, + 568 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 313, + 461, + 554, + 568 + ], + "type": "text", + "content": ". The simulation is performed at 90FPS. We report the per-frame MSE of joint angles." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 569, + 554, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 569, + 554, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 569, + 554, + 654 + ], + "type": "text", + "content": "Results. We demonstrated the results in Tab. 5 and Fig. 7. The results with optimized torques are also reported as a reference. Surprisingly, for different " + }, + { + "bbox": [ + 313, + 569, + 554, + 654 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 569, + 554, + 654 + ], + "type": "text", + "content": ", simulation with HDyS is superior with better stability. However, with " + }, + { + "bbox": [ + 313, + 569, + 554, + 654 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 569, + 554, + 654 + ], + "type": "text", + "content": " increasing, MSE also increases noticeably, indicating the accumulation of drifting errors. Further enhancing the forward-dynamics compatibility of HDyS would be a promising goal." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 670, + 465, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 670, + 465, + 681 + ], + "spans": [ + { + "bbox": [ + 313, + 670, + 465, + 681 + ], + "type": "text", + "content": "4.4.3 Physical Character Control" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "We adopt HDyS for physical character control following the setting of PHC [38], with 140 testing sequences and the rest" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27788" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 70, + 261, + 103 + ], + "blocks": [ + { + "bbox": [ + 82, + 70, + 261, + 103 + ], + "lines": [ + { + "bbox": [ + 82, + 70, + 261, + 103 + ], + "spans": [ + { + "bbox": [ + 82, + 70, + 261, + 103 + ], + "type": "image", + "image_path": "0ab4675e5c355f18458edf063e5773631eb42ff28d997c2969f63b646fdf3f49.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 267, + 70, + 343, + 100 + ], + "blocks": [ + { + "bbox": [ + 267, + 70, + 343, + 100 + ], + "lines": [ + { + "bbox": [ + 267, + 70, + 343, + 100 + ], + "spans": [ + { + "bbox": [ + 267, + 70, + 343, + 100 + ], + "type": "image", + "image_path": "de9a76612f8d11a892a9cb0b33d88f7f164d4afbcbec1415a46b35074afb2284.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 351, + 71, + 518, + 102 + ], + "blocks": [ + { + "bbox": [ + 351, + 71, + 518, + 102 + ], + "lines": [ + { + "bbox": [ + 351, + 71, + 518, + 102 + ], + "spans": [ + { + "bbox": [ + 351, + 71, + 518, + 102 + ], + "type": "image", + "image_path": "a20d64388348ebf1278ac27c378ff72c90f9cc0ecab68900d23db74d784a6156.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 83, + 105, + 260, + 160 + ], + "blocks": [ + { + "bbox": [ + 83, + 105, + 260, + 160 + ], + "lines": [ + { + "bbox": [ + 83, + 105, + 260, + 160 + ], + "spans": [ + { + "bbox": [ + 83, + 105, + 260, + 160 + ], + "type": "image", + "image_path": "f0ea7a9623921a6c62d080b9ef8be3394b4d16b4dc1e4c21b48f6fa03c4a38d5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 261, + 105, + 526, + 160 + ], + "blocks": [ + { + "bbox": [ + 261, + 105, + 526, + 160 + ], + "lines": [ + { + "bbox": [ + 261, + 105, + 526, + 160 + ], + "spans": [ + { + "bbox": [ + 261, + 105, + 526, + 160 + ], + "type": "image", + "image_path": "930bd05651b006f5601ab61f5cdd2e86634c999a666b68cac522e5143d4c3467.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 83, + 163, + 260, + 192 + ], + "blocks": [ + { + "bbox": [ + 83, + 163, + 260, + 192 + ], + "lines": [ + { + "bbox": [ + 83, + 163, + 260, + 192 + ], + "spans": [ + { + "bbox": [ + 83, + 163, + 260, + 192 + ], + "type": "image", + "image_path": "04852060261f962f044e6d12cfa74873885975b5f1be74cc1d6dd2de5ea5c819.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 261, + 163, + 525, + 192 + ], + "blocks": [ + { + "bbox": [ + 261, + 163, + 525, + 192 + ], + "lines": [ + { + "bbox": [ + 261, + 163, + 525, + 192 + ], + "spans": [ + { + "bbox": [ + 261, + 163, + 525, + 192 + ], + "type": "image", + "image_path": "47e448ca6753e3bbdd9e68c4d385dc6ff0bd6cc1ad2166c9d86aeddbdeaf4fe3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 83, + 194, + 260, + 223 + ], + "blocks": [ + { + "bbox": [ + 83, + 194, + 260, + 223 + ], + "lines": [ + { + "bbox": [ + 83, + 194, + 260, + 223 + ], + "spans": [ + { + "bbox": [ + 83, + 194, + 260, + 223 + ], + "type": "image", + "image_path": "6f3373ffaf9c1b2d3cc53659bcbe2263882375315ddbd973df17b97f85258a8d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 261, + 194, + 525, + 223 + ], + "blocks": [ + { + "bbox": [ + 261, + 194, + 525, + 223 + ], + "lines": [ + { + "bbox": [ + 261, + 194, + 525, + 223 + ], + "spans": [ + { + "bbox": [ + 261, + 194, + 525, + 223 + ], + "type": "image", + "image_path": "b4cc7e30a2a58b8c7009d63846f7184b470a94a1434c400b2239f423d5fb4c37.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 83, + 224, + 260, + 253 + ], + "blocks": [ + { + "bbox": [ + 83, + 224, + 260, + 253 + ], + "lines": [ + { + "bbox": [ + 83, + 224, + 260, + 253 + ], + "spans": [ + { + "bbox": [ + 83, + 224, + 260, + 253 + ], + "type": "image", + "image_path": "2bb7e2f6e9992dcbd3c5a41957ff84379d758c8e15d4d303eac2f81eec048c71.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 261, + 224, + 525, + 253 + ], + "blocks": [ + { + "bbox": [ + 261, + 224, + 525, + 253 + ], + "lines": [ + { + "bbox": [ + 261, + 224, + 525, + 253 + ], + "spans": [ + { + "bbox": [ + 261, + 224, + 525, + 253 + ], + "type": "image", + "image_path": "04e0433ae86687811d3b2f0e401b14d344cffedb2095cc4e4bfe8c780fbbd692.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 83, + 254, + 260, + 282 + ], + "blocks": [ + { + "bbox": [ + 83, + 254, + 260, + 282 + ], + "lines": [ + { + "bbox": [ + 83, + 254, + 260, + 282 + ], + "spans": [ + { + "bbox": [ + 83, + 254, + 260, + 282 + ], + "type": "image", + "image_path": "0243863b503b6d33a9f026b8db299300a42caaa036b727fc336d072d20197d24.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 223, + 283, + 385, + 293 + ], + "lines": [ + { + "bbox": [ + 223, + 283, + 385, + 293 + ], + "spans": [ + { + "bbox": [ + 223, + 283, + 385, + 293 + ], + "type": "text", + "content": "Figure 6. Inverse dynamics results on MinT." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 261, + 254, + 525, + 282 + ], + "blocks": [ + { + "bbox": [ + 261, + 254, + 525, + 282 + ], + "lines": [ + { + "bbox": [ + 261, + 254, + 525, + 282 + ], + "spans": [ + { + "bbox": [ + 261, + 254, + 525, + 282 + ], + "type": "image", + "image_path": "07afc5c80888e1f19e2dc422c51add60356027182e82c147a7cdff20b05ff460.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 67, + 305, + 281, + 544 + ], + "blocks": [ + { + "bbox": [ + 67, + 305, + 281, + 544 + ], + "lines": [ + { + "bbox": [ + 67, + 305, + 281, + 544 + ], + "spans": [ + { + "bbox": [ + 67, + 305, + 281, + 544 + ], + "type": "image", + "image_path": "fed51097ebb0a64b17e3d3fcbe7fe49b093000590aa60372d7a2d12891a1560e.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 70, + 544, + 280, + 555 + ], + "lines": [ + { + "bbox": [ + 70, + 544, + 280, + 555 + ], + "spans": [ + { + "bbox": [ + 70, + 544, + 280, + 555 + ], + "type": "text", + "content": "Figure 7. Biomechanical human simulation visualization." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 69, + 584, + 282, + 620 + ], + "blocks": [ + { + "bbox": [ + 88, + 569, + 261, + 580 + ], + "lines": [ + { + "bbox": [ + 88, + 569, + 261, + 580 + ], + "spans": [ + { + "bbox": [ + 88, + 569, + 261, + 580 + ], + "type": "text", + "content": "Table 6. Results of HDyS for character control." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 584, + 282, + 620 + ], + "lines": [ + { + "bbox": [ + 69, + 584, + 282, + 620 + ], + "spans": [ + { + "bbox": [ + 69, + 584, + 282, + 620 + ], + "type": "table", + "html": "
MethodsmPJPEg↓mPJEL↓Evel↓Eacc↓
Baseline74.51449.02911.56214.243
Baseline + HDyS72.69247.92011.30513.757
", + "image_path": "624d6c4ef3b258b7883b1a5e6719892e0790ee0e256f84ae18d6b35d8095028d.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "type": "text", + "content": "for training. We train the baseline as a primitive in PHC with a batch size of 700 for 10k steps. Then, the HDyS latents of key points are inserted as extra observations. We report the global and local mPJPE and the errors of velocity and acceleration in Tab. 6. HDyS manages to improve its performance, validating its efficacy." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 314, + 304, + 384, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 304, + 384, + 316 + ], + "spans": [ + { + "bbox": [ + 314, + 304, + 384, + 316 + ], + "type": "text", + "content": "5. Discussion" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 323, + 555, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 555, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 555, + 563 + ], + "type": "text", + "content": "Despite the impressive results of HDyS, it could be improved. First, it is noticeable that HDyS performs better for the lower body than the higher body in Fig. 6. This might be due to the imbalanced focus on lower body dynamics in data adopted by HDyS. For all datasets, data on lower-body dynamics like gaits are dominating. For AddBiomechanics, we only adopted its armless part. Enhancing HDyS with more upper-body dynamics would be helpful. Second, for muscle actions, HDys could omit minor changes, and the magnitudes could sometimes diverge from the real. Mitigating it with more curated models and loss terms is desirable. Third, as a first step toward homogeneous human dynamics learning, HDyS is primarily instantiated with five initial datasets and an intuitive model. Scaling HDyS up with more datasets and more human priors would be a meaningful goal. Finally, we demonstrate the potential of HDyS with some simple applications in Sec. 4.4, while more sophisticated use cases for musculoskeletal human simulation, humanoid control, and human-robot transfer learning would be promising as future works." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 314, + 574, + 388, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 574, + 388, + 586 + ], + "spans": [ + { + "bbox": [ + 314, + 574, + 388, + 586 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": "We analyzed the heterogeneity issue existing for human dynamics learning and highlighted the homogeneity beneath it. To fully exploit the homogeneity, we proposed HDyS as a homogeneous human dynamics space. Extensive experiments were conducted to validate the feasibility of digging homogeneity out from heterogeneity for human dynamics with detailed analyses of the contribution of heterogeneous components. We further demonstrated the potential of HDyS for downstream applications. We believe HDyS could shed new light on human dynamics understanding." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27789" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 128 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 128 + ], + "type": "text", + "content": "This work is supported in part by the National Natural Science Foundation of China under Grant No.62306175, CCF-Tencent Rhino-Bird Open Research Fund." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 137, + 115, + 149 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 137, + 115, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 137, + 115, + 149 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 157, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 157, + 296, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 157, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 61, + 157, + 296, + 201 + ], + "type": "text", + "content": "[1] Kevin Bergamin, Simon Clavet, Daniel Holden, and James Richard Forbes. Drecon: data-driven responsive control of physics-based characters. ACM Transactions On Graphics (TOG), 38(6):1-11, 2019. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 202, + 296, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 202, + 296, + 255 + ], + "spans": [ + { + "bbox": [ + 61, + 202, + 296, + 255 + ], + "type": "text", + "content": "[2] Vittorio Caggiano, Huawei Wang, Guillaume Durandau, Massimo Sartori, and Vikash Kumar. Myosuite: A contact-rich simulation suite for musculoskeletal motor control. In Learning for Dynamics and Control Conference, pages 492-507. PMLR, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 258, + 295, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 258, + 295, + 290 + ], + "spans": [ + { + "bbox": [ + 62, + 258, + 295, + 290 + ], + "type": "text", + "content": "[3] Mia Chiquier and Carl Vondrick. Muscles in action. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22091-22101, 2023. 2, 3, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 291, + 295, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 291, + 295, + 345 + ], + "spans": [ + { + "bbox": [ + 62, + 291, + 295, + 345 + ], + "type": "text", + "content": "[4] Michael Damsgaard, John Rasmussen, Søren Tørholm Christensen, Egidijus Surma, and Mark De Zee. Analysis of musculoskeletal systems in the anybody modeling system. Simulation Modelling Practice and Theory, 14(8): 1100-1111, 2006. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 347, + 296, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 347, + 296, + 411 + ], + "spans": [ + { + "bbox": [ + 62, + 347, + 296, + 411 + ], + "type": "text", + "content": "[5] Scott L Delp, Frank C Anderson, Allison S Arnold, Peter Loan, Ayman Habib, Chand T John, Eran Guendelman, and Darryl G Thelen. Opensim: open-source software to create and analyze dynamic simulations of movement. IEEE transactions on biomedical engineering, 54(11):1940-1950, 2007. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 414, + 295, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 414, + 295, + 467 + ], + "spans": [ + { + "bbox": [ + 62, + 414, + 295, + 467 + ], + "type": "text", + "content": "[6] Benedikt Feldotto, Cristian Soare, Alois Knoll, Piyanee Sriya, Sarah Astill, Marc de Kamps, and Samit Chakrabarty. Evaluating muscle synergies with emg data and physics simulation in the neurorobotics platform. Frontiers in Neurorobotics, 16:856797, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 469, + 295, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 469, + 295, + 512 + ], + "spans": [ + { + "bbox": [ + 62, + 469, + 295, + 512 + ], + "type": "text", + "content": "[7] Mariusz P Furmanek, Madhur Mangalam, Mathew Yarossi, Kyle Lockwood, and Eugene Tunik. A kinematic and emg dataset of online adjustment of reach-to-grasp movements to visual perturbations. Scientific data, 9(1):23, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 514, + 295, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 514, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 62, + 514, + 295, + 579 + ], + "type": "text", + "content": "[8] E. Gartner, M. Andriluka, E. Coumans, and C. Sminchisescu. Differentiable dynamics for articulated 3d human motion reconstruction. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13180-13190, Los Alamitos, CA, USA, 2022. IEEE Computer Society. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 581, + 295, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 581, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 62, + 581, + 295, + 601 + ], + "type": "text", + "content": "[9] Thomas Geijtenbeek. The Hyfydy simulation software, 2021. https://hyfydy.com.3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 603, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 603, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 57, + 603, + 295, + 668 + ], + "type": "text", + "content": "[10] Daniel FN Gordon, Andreas Christou, Theodoros Stouraitis, Michael Gienger, and Sethu Vijayakumar. Learning personalised human sit-to-stand motion strategies via inverse musculoskeletal optimal control. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 10497–10503. IEEE, 2023. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "text", + "content": "[11] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 554, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 554, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 554, + 95 + ], + "type": "text", + "content": "5142-5151, Los Alamitos, CA, USA, 2022. IEEE Computer Society. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 97, + 554, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 554, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 554, + 162 + ], + "type": "text", + "content": "[12] Erik Gartner, Mykhaylo Andriluka, Hongyi Xu, and Cristian Sminchisescu. Trajectory optimization for physics-based reconstruction of 3d human pose from monocular video. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13096-13105, Los Alamitos, CA, USA, 2022. IEEE Computer Society. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 163, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 553, + 217 + ], + "type": "text", + "content": "[13] Xingjian Han, Ben Senderling, Stanley To, Deepak Kumar, Emily Whiting, and Jun Saito. Groundlink: A dataset unifying human body movement and ground reaction dynamics. In SIGGRAPH Asia 2023 Conference Papers, New York, NY, USA, 2023. Association for Computing Machinery. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 220, + 553, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 283 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 283 + ], + "type": "text", + "content": "[14] Oscar G Hernández, Jose M Lopez-Castellanos, Carlos A Jara, Gabriel J Garcia, Andres Ubeda, Vicente Morell-Gimenez, and Francisco Gomez-Donoso. A kinematic, imaging and electromyography dataset for human muscular manipulability index prediction. Scientific Data, 10(1):132, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 286, + 553, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 329 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 329 + ], + "type": "text", + "content": "[15] Archibald Vivian Hill. The heat of shortening and the dynamic constants of muscle. Proceedings of the Royal Society of London. Series B-Biological Sciences, 126(843):136-195, 1938. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 332, + 553, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 332, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 317, + 332, + 553, + 397 + ], + "type": "text", + "content": "[16] B. Huang, L. Pan, Y. Yang, J. Ju, and Y. Wang. Neural mo- con: Neural motion control for physically plausible human motion capture. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6407- 6416, Los Alamitos, CA, USA, 2022. IEEE Computer Society. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 399, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 399, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 553, + 453 + ], + "type": "text", + "content": "[17] Leif Johnson and Dana H. Ballard. Efficient codes for inverse dynamics during walking. In Proceedings of the Twenty-Eighth AAAI Conference on Artificial Intelligence, pages 343-349, Quebec City, Quebec, Canada, 2014. AAAI Press. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 455, + 554, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 455, + 554, + 498 + ], + "spans": [ + { + "bbox": [ + 317, + 455, + 554, + 498 + ], + "type": "text", + "content": "[18] Lise A Johnson and Andrew J Fuglevand. Evaluation of probabilistic methods to predict muscle activity: implications for neuroprosthetics. Journal of neural engineering, 6(5):055008, 2009. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 500, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 500, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 317, + 500, + 553, + 533 + ], + "type": "text", + "content": "[19] Kimoon Kang, Kiwon Rhee, and Hyun-Chool Shin. Event detection of muscle activation using an electromyogram. Applied Sciences, 10(16):5593, 2020. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 535, + 553, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 535, + 553, + 588 + ], + "spans": [ + { + "bbox": [ + 317, + 535, + 553, + 588 + ], + "type": "text", + "content": "[20] Marilyn Keller, Keenon Werling, Soyong Shin, Scott Delp, Sergi Pujades, C Karen Liu, and Michael J Black. From skin to skeleton: Towards biomechanically accurate 3d digital humans. ACM Transactions on Graphics (TOG), 42(6): 1-12, 2023. 1, 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 590, + 553, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 590, + 553, + 643 + ], + "spans": [ + { + "bbox": [ + 317, + 590, + 553, + 643 + ], + "type": "text", + "content": "[21] Muhammed Kocabas, Nikos Athanasiou, and Michael J Black. Vibe: Video inference for human body pose and shape estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5253-5263, 2020. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "text", + "content": "[22] Claudia Latella, Naveen Kuppuswamy, Francesco Romano, Silvio Traversaro, and Francesco Nori. Whole-body human inverse dynamics with distributed micro-accelerometers, gyros and force sensing. Sensors, 16(5), 2016. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[23] Claudia Latella, Silvio Traversaro, Diego Ferigo, Yeshasvi Tirupachuri, Lorenzo Rapetti, Francisco Javier An" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27790" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "text", + "content": "drade Chavez, Francesco Nori, and Daniele Pucci. Simultaneous floating-base estimation of human kinematics and joint torques. Sensors, 19(12), 2019. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 162 + ], + "type": "text", + "content": "[24] Jiefeng Li, Chao Xu, Zhicun Chen, Siyuan Bian, Lixin Yang, and Cewu Lu. Hybrik: A hybrid analytical-neural inverse kinematics solution for 3d human pose and shape estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3383-3393, 2021. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 218 + ], + "type": "text", + "content": "[25] Sheng-Wei Li, Zi-Xiang Wei, Wei-Jie Chen, Yi-Hsin Yu, Chih-Yuan Yang, and Jane Yung-jen Hsu. Sa-dvae: Improving zero-shot skeleton-based action recognition by disentangled variational autoencoders. In European Conference on Computer Vision, pages 447–462. Springer, 2024. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 262 + ], + "type": "text", + "content": "[26] Yong-Lu Li, Siyuan Zhou, Xijie Huang, Liang Xu, Ze Ma, Hao-Shu Fang, Yanfeng Wang, and Cewu Lu. Transferable interactiveness knowledge for human-object interaction detection. In CVPR, 2019. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "type": "text", + "content": "[27] Yong-Lu Li, Xinpeng Liu, Han Lu, Shiyi Wang, Junqi Liu, Jiefeng Li, and Cewu Lu. Detailed 2d-3d joint representation for human-object interaction. In CVPR, 2020. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 298, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 298, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 294, + 342 + ], + "type": "text", + "content": "[28] Yong-Lu Li, Liang Xu, Xinpeng Liu, Xijie Huang, Yue Xu, Shiyi Wang, Hao-Shu Fang, Ze Ma, Mingyang Chen, and Cewu Lu. Pastanet: Toward human activity knowledge engine. In CVPR, 2020. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 343, + 294, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 294, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 294, + 387 + ], + "type": "text", + "content": "[29] Yong-Lu Li, Xinpeng Liu, Xiaogian Wu, Yizhuo Li, Zuoyu Qiu, Liang Xu, Yue Xu, Hao-Shu Fang, and Cewu Lu. Hake: A knowledge engine foundation for human activity understanding. TPAMI, 2022. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 388, + 294, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 441 + ], + "type": "text", + "content": "[30] Yong-Lu Li, Xiaogian Wu, Xinpeng Liu, Yiming Dou, Yikun Ji, Junyi Zhang, Yixing Li, Jingru Tan, Xudong Lu, and Cewu Lu. From isolated islands to pangea: Unifying semantic space for human action understanding. In CVPR, 2024. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 445, + 294, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 487 + ], + "type": "text", + "content": "[31] Zhan Li, David Guiraud, and Mitsuhiro Hayashibe. Inverse estimation of multiple muscle activations from joint moment with muscle synergy extraction. IEEE journal of biomedical and health informatics, 19(1):64-73, 2014. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 490, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 490, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 490, + 294, + 533 + ], + "type": "text", + "content": "[32] Xinpeng Liu, Haowen Hou, Yanchao Yang, Yong-Lu Li, and Cewu Lu. Revisit human-scene interaction via space occupancy. In European Conference on Computer Vision, pages 1-19. Springer, 2024. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 535, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 294, + 578 + ], + "type": "text", + "content": "[33] Xinpeng Liu, Junxuan Liang, Zili Lin, Haowen Hou, Yong-Lu Li, and Cewu Lu. Imdy: Human inverse dynamics from imitated observations. In The Thirteenth International Conference on Learning Representations, 2025. 2, 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 580, + 294, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 580, + 294, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 580, + 294, + 623 + ], + "type": "text", + "content": "[34] Xiaoyang Liu, Boran Wen, Xinpeng Liu, Zizheng Zhou, Hongwei Fan, Cewu Lu, Lizhuang Ma, Yulong Chen, and Yong-Lu Li. Interacted object grounding in spatio-temporal human-object interactions. In AAAI, 2025. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 624, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 657 + ], + "type": "text", + "content": "[35] Yilin Liu, Shijia Zhang, and Mahanth Gowda. Neuropose: 3d hand pose tracking using emg wearables. In Proceedings of the Web Conference 2021, pages 1471-1482, 2021. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "text", + "content": "[36] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. Smpl: a skinned multiperson linear model. In ACM Transactions on Graphics, New York, NY, USA, 2015. Association for Computing Machinery. 1, 3, 4, 5" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[37] Zhengyi Luo, Ryo Hachiuma, Ye Yuan, and Kris Kitani. Dynamics-regulated kinematic policy for egocentric pose estimation. Advances in Neural Information Processing Systems, 34:25019-25032, 2021. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 171 + ], + "type": "text", + "content": "[38] Z. Luo, J. Cao, A. Winkler, K. Kitani, and W. Xu. Perpetual humanoid control for real-time simulated avatars. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 10861-10870, Los Alamitos, CA, USA, 2023. IEEE Computer Society. 3, 5, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 173, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 173, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 173, + 553, + 205 + ], + "type": "text", + "content": "[39] Xiaolei Lv, Jinxiang Chai, and Shihong Xia. Data-driven inverse dynamics for human motion. ACM Transactions on Graphics (TOG), 35(6):1-12, 2016. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 206, + 553, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 260 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 260 + ], + "type": "text", + "content": "[40] N. Mahmood, N. Ghorbani, N. F. Troje, G. Pons-Moll, and M. Black. Amass: Archive of motion capture as surface shapes. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 5441-5450, Los Alamitos, CA, USA, 2019. IEEE Computer Society. 1, 3, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 262, + 553, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 262, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 262, + 553, + 316 + ], + "type": "text", + "content": "[41] Nebojša Malešević, Alexander Olsson, Paulina Sager, Elin Andersson, Christian Cipriani, Marco Controzzi, Anders Björkman, and Christian Antfolk. A database of high-density surface electromyogram signals comprising 65 isometric hand gestures. Scientific Data, 8(1):63, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 317, + 553, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 317, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 316, + 317, + 553, + 369 + ], + "type": "text", + "content": "[42] Christian Mandery, Ömer Terlemez, Martin Do, Nikolaus Vahrenkamp, and Tamim Asfour. Unifying representations and large-scale whole-body motion databases for studying human motion. IEEE Transactions on Robotics, 32(4):796-809, 2016. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 372, + 553, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 553, + 414 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 553, + 414 + ], + "type": "text", + "content": "[43] Mykhailo Manukian, Serhii Bahdasariants, and Sergiy Yakovenko. Artificial physics engine for real-time inverse dynamics of arm and hand movement. Plos one, 18(12): e0295750, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 415, + 553, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 553, + 468 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 553, + 468 + ], + "type": "text", + "content": "[44] Luís Moreira, Joana Figueiredo, Pedro Fonseca, João P Vilas-Boas, and Cristina P Santos. Lower limb kinematic, kinetic, and emg data from young healthy humans during walking at controlled speeds. Scientific data, 8(1):103, 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 472, + 553, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 472, + 553, + 524 + ], + "spans": [ + { + "bbox": [ + 316, + 472, + 553, + 524 + ], + "type": "text", + "content": "[45] Yoshihiko Nakamura, Katsu Yamane, Yusuke Fujita, and Ichiro Suzuki. Somatosensory computation for man-machine interface from motion-capture data and musculoskeletal human model. IEEE Transactions on Robotics, 21(1):58-66, 2005. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 526, + 553, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 526, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 316, + 526, + 553, + 569 + ], + "type": "text", + "content": "[46] Hui Niu, Takahiro Ito, Damien Desclaux, Ko Ayusawa, Yusuke Yoshiyasu, Ryusuke Sagawa, and Eiichi Yoshida. Estimating muscle activity from the deformation of a sequential 3d point cloud. Journal of Imaging, 8(6):168, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 571, + 553, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 571, + 553, + 603 + ], + "spans": [ + { + "bbox": [ + 316, + 571, + 553, + 603 + ], + "type": "text", + "content": "[47] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 604, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 604, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 604, + 553, + 669 + ], + "type": "text", + "content": "[48] Kunyu Peng, David Schneider, Alina Roitberg, Kailun Yang, Jiaming Zhang, Chen Deng, Kaiyu Zhang, M Saqib Sarfraz, and Rainer Stiefelhagen. Towards video-based activated muscle group estimation in the wild. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 4495-4504, 2024. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[49] Xue Bin Peng, Ze Ma, Pieter Abbeel, Sergey Levine, and Angjoo Kanazawa. Amp: Adversarial motion priors for stylized physics-based character control. ACM Transactions on Graphics (ToG), 40(4):1-20, 2021. 3" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27791" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 296, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 296, + 116 + ], + "type": "text", + "content": "[50] Xue Bin Peng, Yunrong Guo, Lina Halper, Sergey Levine, and Sanja Fidler. Ase: Large-scale reusable adversarial skill embeddings for physically simulated characters. ACM Transactions On Graphics (TOG), 41(4):1-17, 2022. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 296, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 296, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 296, + 183 + ], + "type": "text", + "content": "[51] A. R. Punnakkal, A. Chandrasekaran, N. Athanasiou, A. Quiros-Ramirez, and M. J. Black. Babel: Bodies, action and behavior with english labels. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 722-731, Los Alamitos, CA, USA, 2021. IEEE Computer Society. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 184, + 295, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 295, + 238 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 295, + 238 + ], + "type": "text", + "content": "[52] Apoorva Rajagopal, Christopher L Dembia, Matthew S Demers, Denny D Delp, Jennifer L Hicks, and Scott L Delp. Full-body musculoskeletal model for muscle-driven simulation of human gait. IEEE transactions on biomedical engineering, 63(10):2068-2079, 2016. 1, 3, 5, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 239, + 295, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 295, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 295, + 304 + ], + "type": "text", + "content": "[53] David Schneider, Simon Reiß, Marco Kugler, Alexander Jaus, Kunyu Peng, Susanne Sutschet, Muhammad Saqib Sarfraz, Sven Matthiesen, and Rainer Stiefelhagen. Muscles in time: Learning to understand human motion in-depth by simulating muscle activations. Advances in Neural Information Processing Systems, 2025. 2, 3, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 304, + 295, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 304, + 295, + 358 + ], + "spans": [ + { + "bbox": [ + 56, + 304, + 295, + 358 + ], + "type": "text", + "content": "[54] Robert V Schulte, Marijke Zondag, Jaap H Buurke, and Erik C Prinsen. Multi-day emg-based knee joint torque estimation using hybrid neuromusculoskeletal modelling and convolutional neural networks. Frontiers in Robotics and AI, 9:869476, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 360, + 295, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 360, + 295, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 360, + 295, + 403 + ], + "type": "text", + "content": "[55] Paniz Sedighi, Xingyu Li, and Mahdi Tavakoli. Emg-based intention detection using deep learning for shared control in upper-limb assistive exoskeletons. IEEE Robotics and Automation Letters, 2023. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 404, + 295, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 295, + 448 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 295, + 448 + ], + "type": "text", + "content": "[56] Masashi Sekiya, Sho Sakaino, and Tsuji Toshiaki. Linear logistic regression for estimation of lower limb muscle activations. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 27(3):523-532, 2019. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 449, + 295, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 295, + 502 + ], + "type": "text", + "content": "[57] Antun Skuric, Vincent Padois, Nasser Rezzoug, and David Daney. On-line feasible wrench polytope evaluation based on human musculoskeletal models: an iterative convex hull method. IEEE robotics and automation letters, 7(2):5206-5213, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 504, + 295, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 295, + 547 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 295, + 547 + ], + "type": "text", + "content": "[58] Hyungeun Song and Yoichi Hori. Inverse muscle group activity estimation based on neuromusculoskeletal system model. In TENCON 2015-2015 IEEE Region 10 Conference, pages 1-5. IEEE, 2015. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 548, + 295, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 295, + 590 + ], + "type": "text", + "content": "[59] Yokhesh K Tamilselvam, Jacky Ganguly, Rajni V Patel, and Mandar Jog. Musculoskeletal model to predict muscle activity during upper limb movement. *Ieee Access*, 9:111472–111485, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 592, + 295, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 295, + 646 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 295, + 646 + ], + "type": "text", + "content": "[60] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit Haim Bermano. Human motion diffusion model. In The Eleventh International Conference on Learning Representations ICLR 2023, Kigali, Rwanda, 2023. OpenReview.net. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 647, + 295, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 295, + 691 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 295, + 691 + ], + "type": "text", + "content": "[61] Elly Trepman, Richard E Gellman, Lyle J Micheli, and CARLO J De Luca. Electromyographic analysis of grandplie in ballet and modern dancers. Medicine and science in sports and exercise, 30(12):1708-1720, 1998. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 692, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 295, + 713 + ], + "type": "text", + "content": "[62] J. Wang, Y. Yuan, Z. Luo, K. Xie, D. Lin, U. Iqbal, S. Fidler, and S. Khamis. Learning human dynamics in autonomous" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "driving scenarios. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 20739-20749, Los Alamitos, CA, USA, 2023. IEEE Computer Society. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 108, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 108, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 108, + 553, + 173 + ], + "type": "text", + "content": "[63] Keenon Werling, Dalton Omens, Jeongseok Lee, Ioannis Exarchos, and C. Karen Liu. Fast and Feature-Complete Differentiable Physics Engine for Articulated Rigid Bodies with Contact Constraints. In Robotics: Science and Systems XVII, Virtual Event, July 12-16, 2021, Virtual, 2021. RSS Foundation. 2, 5, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 175, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 239 + ], + "type": "text", + "content": "[64] Keenon Werling, Nicholas A Bianco, Michael Raitor, Jon Stingel, Jennifer L Hicks, Steven H Collins, Scott L Delp, and C Karen Liu. Addbiomechanics: Automating model scaling, inverse kinematics, and inverse dynamics from human motion data through sequential optimization. Plos one, 18(11):e0295152, 2023. 2, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 241, + 553, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 307 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 307 + ], + "type": "text", + "content": "[65] Keenon Werling, Janelle Kaneda, Tian Tan, Rishi Agarwal, Six Skov, Tom Van Wouwe, Scott Uhlrich, Nicholas Bianco, Carmichael Ong, Antoine Falisse, et al. Addbiomechanics dataset: Capturing the physics of human motion at scale. In European Conference on Computer Vision, pages 490-508. Springer, 2024. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 308, + 555, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 555, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 555, + 374 + ], + "type": "text", + "content": "[66] Lahiru N Wimalasena, Jonas F Braun, Mohammad Reza Keshtkaran, David Hofmann, Juan Álvaro Gallego, Cristiano Alessandro, Matthew C Tresch, Lee E Miller, and Chethan Pandarinath. Estimating muscle activation from emg using deep learning-based dynamical systems models. Journal of neural engineering, 19(3):036013, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 376, + 553, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 420 + ], + "type": "text", + "content": "[67] Jungdam Won, Deepak Gopinath, and Jessica Hodgins. Control strategies for physically simulated characters performing two-player competitive sports. ACM Transactions on Graphics (TOG), 40(4):1-11, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 421, + 553, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 421, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 316, + 421, + 553, + 454 + ], + "type": "text", + "content": "[68] Jungdam Won, Deepak Gopinath, and Jessica Hodgins. Physics-based character controllers using conditional vaes. ACM Transactions on Graphics (TOG), 41(4):1-12, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 456, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 510 + ], + "type": "text", + "content": "[69] Baoping Xiong, Nianyin Zeng, Han Li, Yuan Yang, Yurong Li, Meilan Huang, Wuxiang Shi, Min Du, and Yudong Zhang. Intelligent prediction of human lower extremity joint moment: an artificial neural network approach. IEEE Access, 7:29973-29980, 2019. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 511, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 511, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 511, + 553, + 555 + ], + "type": "text", + "content": "[70] Aya Yaacoub, Vincent Thomas, Francis Colas, and Pauline Maurice. A probabilistic model for cobot decision making to mitigate human fatigue in repetitive co-manipulation tasks. IEEE Robotics and Automation Letters, 2023. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 556, + 553, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 556, + 553, + 612 + ], + "spans": [ + { + "bbox": [ + 316, + 556, + 553, + 612 + ], + "type": "text", + "content": "[71] Katsu Yamane, Akihiko Murai, Sadahiro Takaya, and Yoshihiko Nakamura. Muscle tension database for contact-free estimation of human somatosensory information. In 2009 IEEE International Conference on Robotics and Automation, pages 633-638. IEEE, 2009. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 613, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 679 + ], + "type": "text", + "content": "[72] X. Yi, Y. Zhou, M. Habermann, S. Shimada, V. Golyanik, C. Theobalt, and F. Xu. Physical inertial poser (pip): Physics-aware real-time human motion tracking from sparse inertial sensors. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13157-13168, Los Alamitos, CA, USA, 2022. IEEE Computer Society. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[73] Y. Yuan and K. Kitani. Ego-posed estimation and forecasting as real-time pd control. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27792" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 498 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "10081-10091, Los Alamitos, CA, USA, 2019. IEEE Computer Society. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 139 + ], + "type": "text", + "content": "[74] Felix E Zajac. Muscle and tendon: properties, models, scaling, and application to biomechanics and motor control. Critical reviews in biomedical engineering, 17(4):359-411, 1989. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 184 + ], + "type": "text", + "content": "[75] Petrissa Zell and Bodo Rosenhahn. A physics-based statistical model for human gait analysis. In Pattern Recognition, pages 169-180, Cham, 2015. Springer International Publishing. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 294, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 229 + ], + "type": "text", + "content": "[76] P. Zell and B. Rosenhahn. Learning-based inverse dynamics of human motion. In 2017 IEEE International Conference on Computer Vision Workshop (ICCVW), pages 842–850, Los Alamitos, CA, USA, 2017. IEEE Computer Society. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 231, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 294, + 285 + ], + "type": "text", + "content": "[77] P. Zell, B. Wandt, and B. Rosenhahn. Joint 3d human motion capture and physical analysis from monocular videos. In 2017 IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 17-26, Los Alamitos, CA, USA, 2017. IEEE Computer Society. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 286, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 294, + 330 + ], + "type": "text", + "content": "[78] Petrissa Zell, Bodo Rosenhahn, and Bastian Wandt. Weakly-supervised learning of human dynamics. In Computer Vision - ECCV 2020, pages 68-84, Cham, 2020. Springer International Publishing. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 331, + 294, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 294, + 385 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 294, + 385 + ], + "type": "text", + "content": "[79] Longbin Zhang, Davit Soselia, Ruoli Wang, and Elena M Gutierrez-Farewik. Estimation of joint torque by emg-driven neuromusculoskeletal models and lstm networks. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 387, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 387, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 387, + 294, + 430 + ], + "type": "text", + "content": "[80] Yufei Zhang, Jeffrey O Kephart, and Qiang Ji. Incorporating physics principles for precise human motion prediction. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 6164–6174, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 432, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 432, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 432, + 294, + 498 + ], + "type": "text", + "content": "[81] Chenhui Zuo, Kaibo He, Jing Shao, and Yanan Sui. Self model for embodied intelligence: Modeling full-body human musculoskeletal system and locomotion control with hierarchical low-dimensional representation. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 13062-13069. IEEE, 2024. 3" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27793" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Horizon-GS_ Unified 3D Gaussian Splatting for Large-Scale Aerial-to-Ground Scenes/65e9755f-fe8e-4a32-88df-f20ee7cf9c9c_content_list.json b/2025/Horizon-GS_ Unified 3D Gaussian Splatting for Large-Scale Aerial-to-Ground Scenes/65e9755f-fe8e-4a32-88df-f20ee7cf9c9c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..53b7e2d836dbabc056485a9f7bb7f33b767e5e7f --- /dev/null +++ b/2025/Horizon-GS_ Unified 3D Gaussian Splatting for Large-Scale Aerial-to-Ground Scenes/65e9755f-fe8e-4a32-88df-f20ee7cf9c9c_content_list.json @@ -0,0 +1,1500 @@ +[ + { + "type": "text", + "text": "Horizon-GS: Unified 3D Gaussian Splatting for Large-Scale Aerial-to-Ground Scenes", + "text_level": 1, + "bbox": [ + 197, + 130, + 802, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lihan Jiang $^{1,3*}$ Kerui Ren $^{2,3*}$ Mulin Yu $^{3}$ Linning Xu $^{4}$ \nJunting Dong $^{3}$ Tao Lu $^{5}$ Feng Zhao $^{1\\dagger}$ Dahua Lin $^{3,4,7}$ Bo Dai $^{6\\dagger}$", + "bbox": [ + 217, + 202, + 777, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of Science and Technology of China, 2Shanghai Jiao Tong University, 3Shanghai Artificial Intelligence Laboratory, 4The Chinese University of Hong Kong, 5Brown University, 6The University of Hong Kong, 7CPII", + "bbox": [ + 243, + 247, + 753, + 299 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/06d221dd08c6e9d6305ff87fff7af2ba62b1a5a30101d6039365b7c968f93648.jpg", + "image_caption": [ + "Figure 1. Horizon-GS enables high-quality rendering and reconstruction of aerial-to-ground scenes with unprecedented fidelity across scales, supporting drastic view changes. The colored camera trajectories in the center illustrate the novel viewpoints, while the reconstructed mesh is overlaid on the scene. The surrounding images show the corresponding predicted images for each viewpoint. Explore more demos on our project page: https://city-super.github.io/horizon-gs/." + ], + "image_footnote": [], + "bbox": [ + 96, + 324, + 897, + 594 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 681, + 326, + 698 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Seamless integration of both aerial and street view images remains a significant challenge in neural scene reconstruction and rendering. Existing methods predominantly focus on single domain, limiting their applications in immersive environments, which demand extensive free view exploration with large view changes both horizontally and vertically. We introduce Horizon-GS, a novel approach built upon Gaussian Splatting techniques, tackles the unified reconstruction and rendering for aerial and street views. Our method addresses the key challenges of combining these", + "bbox": [ + 89, + 715, + 485, + 868 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "perspectives with a new training strategy, overcoming viewpoint discrepancies to generate high-fidelity scenes. We also curate a high-quality aerial-to-ground views dataset encompassing both synthetic and real-world scene to advance further research. Experiments across diverse urban scene datasets confirm the effectiveness of our method.", + "bbox": [ + 511, + 669, + 908, + 761 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 768, + 645, + 784 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Modeling and rendering large-scale scenes has become essential across domains such as Embodied AI, digital twins, and autonomous driving. Neural radiance fields [19] have revealed the potential to realize this vision. Recently, 3D Gaussian Splatting (3D-GS) [8] has advanced this field further, achieving exceptional visual quality alongside real-time rendering speeds. These capabilities markedly accel", + "bbox": [ + 511, + 794, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution. †Corresponding author.", + "bbox": [ + 114, + 886, + 354, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "26789", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "erate the transformation of the physical world into digital spaces, supporting applications from real-to-sim simulation and autonomous navigation to immersive VR/AR experiences and the metaverse.", + "bbox": [ + 89, + 90, + 480, + 150 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite recent advancements in rendering aerial [2, 12, 16] and street views [9] with 3D-GS, existing methods remain limited to training on a single view type, resulting in a disjointed experience. We aim to bridge this gap by integrating both perspectives, enabling a seamless and unified experience across downstream applications. Our preliminary experiments on state-of-the-art Gaussian Splatting methods reveal two primary challenges in integrating aerial and street views: 1) conflicts in Gaussian densification arising from the substantial disparity between aerial and street views, and 2) an uneven distribution of camera data, with a bias toward street views. During the densification process, limited visibility from street views impedes consistent gradient accumulation, thereby disrupting the Gaussian instantiation process across the scene. Moreover, the variation in detail between aerial and street views makes a single static set of Gaussian primitives insufficiently flexible, restricting smooth transitions between viewpoints.", + "bbox": [ + 89, + 154, + 482, + 425 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address above challenges and achieve a seamless integration for both aerial and street views, we propose a novel method, Horizon-GS, that overcomes the limitations of existing approaches. Specifically, we propose a coarse-to-fine training strategy that explicitly divides the entire training process into two stages: the first stage benefits more from aerial images to establish a coarse geometry, and the second stage focuses on adding finer details mainly from street views. Combined with a camera distribution balance strategy, our method achieves state-of-the-art quality for both aerial and street views, delivering superior rendering quality and geometry accuracy. Moreover, our approach is easily adaptable for large-scale reconstruction. We also notice that, the lack of calibrated aerial and street datasets is a critical issue that hinders unified model training. To advance the field and evaluate our method, we construct a large-scale dataset consisting of five synthetic and two real-world scenes, encompassing both aerial and street views.", + "bbox": [ + 89, + 428, + 482, + 699 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, the main contributions of our method are:", + "bbox": [ + 109, + 702, + 472, + 715 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We tackle the challenging and important task of unified large-scale scene reconstruction from both aerial and street views, starting with an in-depth analysis of the core conflicts that hinder integration across these perspectives.", + "- We propose an efficient framework that addresses the inherent discrepancies and unifies the contribution from aerial and street views. Our framework is also adaptable to both 3D Gaussians for realistic novel view synthesis and 2D Gaussians for accurate geometry reconstruction.", + "- We present a high-quality and diverse dataset that includes both synthetic and real-world data, specifically curated to support balanced, cross-view reconstruction." + ], + "bbox": [ + 91, + 719, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Our approach achieves state-of-the-art performance in small-scale and large-scale scene reconstruction, setting a new benchmark for unified aerial and street modeling.", + "bbox": [ + 513, + 90, + 903, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 513, + 150, + 651, + 165 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural Scene Representations. Neural Radiance Fields (NeRF) [19] have sparked significantly interest in novel view synthesis due to their photorealistic visual quality. Recent research [1, 5, 15, 20, 25, 33] has introduced various hybrid grid representations, improving efficiency by reducing sampling points and MLP layers, thus accelerating training and rendering. More recently, 3D Gaussian Splatting (3D-GS) [8] has enhanced visual quality and training efficiency by utilizing anisotropic 3D Gaussians and enabling real-time rendering with tile-based splatting. It's widely applied in various fields, especially in many large-scale applications, such as autonomous driving [3, 31, 37, 44], geometry reconstruction [6, 39, 40], and 3D generative tasks [27, 28, 36, 41].", + "bbox": [ + 511, + 175, + 903, + 387 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Large-scale Scene Modeling. NeRF and 3D Gaussian Splatting have sparked a new trend in neural modeling for large-scale scenes, pushing the boundaries of visual fidelity and real-time performance. Extensive neural representation methods have been developed for large-scale city reconstruction and rendering. However, most methods target either aerial [2, 12, 16, 18, 23, 29, 32, 35] or street views [9, 14, 22, 26]. For aerial views, models like Vast-Gaussian [12] and CityGaussian [16] use data partitioning for parallel training, while DoGaussian [2] employs a global node to ensure consistent inference. For street scenes, Hierarchical-3DGS [9] optimizes and renders Gaussians in an LOD hierarchy. However, these methods do not naturally support a unified reconstruction from both aerial and street views. The work closest to ours is UC-GS [43], which incorporates cross-view uncertainty to enhance details in extrapolated views, but it still struggles with significant perspective and granularity differences.", + "bbox": [ + 511, + 393, + 903, + 664 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another major barrier to unified aerial and street modeling is the scarcity of datasets with calibrated images from both perspectives. Current large-scale neural rendering datasets, such as UrbanScene3D [13], Mill19 [29], Quad 6K [4], and MatrixCity [10], Waymo Block-NeRF [26], KITTI-360 [11], and Hierarchical-3DGS [9], typically cater to a single view type. MatrixCity is the only dataset with both views, but it is synthetic and lacks diversity. To fill this gap, we introduce a new dataset with a variety of synthetic and real-world scenes, supporting research on large-scale, unified reconstruction.", + "bbox": [ + 511, + 665, + 903, + 830 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Horizon-GS", + "text_level": 1, + "bbox": [ + 513, + 844, + 638, + 859 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While substantial progress has been made in reconstructing urban scenes from either aerial or street images [2, 9,", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "26790", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d66834cdb19543f17dba28bbe38c670431c73f324810969f8ee47ceae26c3650.jpg", + "image_caption": [ + "Figure 2. Pipeline of Horizon-GS. We divide large-scale scenes into chunks. For each chunk, we initialize LOD-structured anchors and conduct the coarse-to-fine training process. Specifically, the coarse stage reconstructs the overall scene, while the fine stage enhances street view details (highlighted in purple). We can derive RGB, depth, and normal images by utilizing different primitive attributes (2D/3D Gaussians) with a single shared underlying structure." + ], + "image_footnote": [], + "bbox": [ + 96, + 89, + 537, + 251 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d44a7aedbbb32dceea44916445ab9124412d7e1cf511ec70cdf103a213ce1cce.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 539, + 88, + 777, + 252 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5df718579db3f61ecb3f1c7cd9a9d8c39573134fd1cee8ce9116a4dcfee920da.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 785, + 89, + 892, + 251 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "12, 16], the challenge of combining both perspectives into a unified model remains unsolved and largely overlooked. This task is non-trivial, requiring solutions to bridge the significant disparities in scale, perspective, and appearance between aerial and street views. In this section, we first analyze the core challenges in Section 3.1, briefly describe the base modules in Section 3.2, and then detail the proposed framework that leverages a two-stage training strategy and the enhanced 3D Gaussian densification operators in Section 3.3. Additionally, how to apply the framework to large-scale scenes is discussed in Section 3.4. Finally, we describe the loss functions and regularization in Section 3.5.", + "bbox": [ + 88, + 340, + 483, + 521 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Challenges", + "text_level": 1, + "bbox": [ + 89, + 531, + 212, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Before presenting our framework, we first analyze the inherent conflicts in using 3D Gaussians for simultaneous training on aerial and street views. We apply the state-of-the-art Scaffold-GS [17] on our collected real-world scene which features both aerial and street views. Naturally, we aim to combine these two classes of views during reconstruction, as they complement each other. But we notice that reconstructing either aerial or street views alone has achieved much better results than training them together, as shown in Fig. 3. Combining these perspectives presents two major challenges: (1) conflicting gradient accumulation and (2) disparity in view information coverage.", + "bbox": [ + 88, + 553, + 482, + 733 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The first challenge arises from the differences between aerial and street views, which lead to conflicting gradient updates during training. While street views cause the densification policy to remove irrelevant, occluded Gaussians, aerial views promote regrowth in these regions. This interference disrupts the Gaussian densification process, making it unreliable and ultimately degrading reconstruction quality. Fig. 3 confirms that joint training across both domains consistently underperformed separate training, even with extensive iterations, and with more observed views, the number of final Gaussian primitives are even less.", + "bbox": [ + 88, + 734, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5b559bcf92091ab3382c7598b4eb8f857873e275a7420edf7dcd775e9a3142c9.jpg", + "image_caption": [ + "Figure 3. (a) Test curves for PSNR and the number of Gaussian primitives across aerial only, street only, and merged views from 15k to 100k iterations on our proposed Road scene. (b) Gradient conflicts restrict the optimization of Gaussian primitives because street views tend to exclude blue Gaussian primitives due to their lower contribution, while aerial views do the opposite." + ], + "image_footnote": [], + "bbox": [ + 522, + 338, + 678, + 496 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/327d3bd8480888f947fdac6c603a15da665b42fcc214d54cbc1f89ddf719ed40.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 679, + 338, + 903, + 498 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The second challenge is the imbalance in camera coverage: street views capture detailed, nearly 360-degree scenes, while aerial views are downward-focused and less frequent. This uneven distribution shifts the reconstruction focus toward street details, preventing a balanced integration of both perspectives. Together, these challenges emphasize the complexity of unified aerial-street modeling and the necessity for specialized solutions.", + "bbox": [ + 511, + 611, + 906, + 733 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Gaussian Splitting Base Modules", + "text_level": 1, + "bbox": [ + 511, + 742, + 805, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Gaussian Splatting (3D-GS) [8] represents scenes using anisotropic 3D Gaussian primitives, achieving high rendering quality at significantly faster speeds. Each Gaussian $G$ is defined by parameters: a center $\\mu \\in \\mathbb{R}^3$ , a rotation quaternion $q \\in \\mathbb{R}^4$ , scaling vector $s \\in \\mathbb{R}^3$ , opacity $\\sigma \\in [0,1]$ , and a color feature $c$ either encoded with spherical harmonics or direct RGB feature for view-dependent/independent color. To render images, the 3D Gaussians are projected onto the image plane as 2D Gaussians $G'(x)$ , ordered front-to-back,", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "26791", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and $\\alpha$ -blended for pixel color calculation.", + "bbox": [ + 89, + 90, + 367, + 104 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Scaffold-GS [17] further introduces anchors to reduce redundancy and enhance storage efficiency. From each anchor, $k$ neural Gaussians are emitted, with centers determined by learnable offsets. Properties are decoded by MLPs from anchor features and viewing positions, creating a more robust and compact representation.", + "bbox": [ + 89, + 107, + 483, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2D Gaussian Splating (2D-GS) [7] collapses 3D Gaussians into 2D oriented planar Gaussian disks for surface accuracy. An adaptive rasterizer efficiently renders these 2D Gaussians, and depth distortion and normal consistency losses are added to improve reconstruction quality. Finally, a mesh is extracted using a truncated signed distance function (TSDF), which fuses multi-view depth maps from the optimized 2D-GS field.", + "bbox": [ + 89, + 198, + 483, + 318 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For enhanced robustness, we adopt the anchor design from Scaffold-GS [17], incorporating a flexible neural Gaussian representation. Specifically, the properties of different neural Gaussians, both 3D and 2D, can be generated by learnable MLPs. We utilize neural 3D Gaussians to improve rendering quality and neural 2D Gaussians for high-fidelity geometry reconstruction.", + "bbox": [ + 89, + 320, + 483, + 426 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Aerial-Street Joint Reconstruction", + "text_level": 1, + "bbox": [ + 89, + 438, + 390, + 452 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Coarse-to-Fine Training. The significant difference in aerial and street views lead to performance challenges when training them together, necessitating separate handling for each modality. Inspired by BungeeNeRF [32], we propose a two-stage coarse-to-fine training strategy where street views refine the details of aerial views. In the first stage, we focus primarily on the aerial views to establish a coarse geometry, accumulating gradients for densification from the aerial images alone and allowing the Gaussian primitives to fully develop and cover a broader feature space. Aerial views guide the initial placement of fine-grained 3D Gaussians, ensuring their alignment with the global scene. In the second stage, we freeze the MLP weights and attributes of the Gaussian primitives from the first stage to preserve the skeleton structure. Street views then play a key role in refining this skeleton, adding finer details and achieving a balanced integration of both views while efficiently resolving their inherent conflicts.", + "bbox": [ + 89, + 460, + 483, + 732 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We modify the densification policy in the second stage to enhance the capture of fine-grained details in largescale scenes with sparse images. The original Scaffolds-GS method averages screen-space gradients of neural Gaussians within each voxel. While effective for object-centric scenes, it struggles to model local details in sparsely observed areas. As a result, subtle features remain underestimated, leading to blurred renderings.", + "bbox": [ + 89, + 733, + 482, + 853 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here we consider neural Gaussians with higher gradients, greater opacity, and larger projected screen size as more significant, inspired by Hierarchical-3DGS [9].", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, we compute the max gradients $\\nabla_g$ , average opacity $\\sigma$ and max projected radius $r$ of the included neural Gaussians for each $N$ iterations. Those neural Gaussians satisfying:", + "bbox": [ + 511, + 90, + 905, + 150 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {g} \\cdot r \\cdot \\sigma^ {\\tau_ {\\sigma}} > \\tau_ {g}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 647, + 151, + 903, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\tau_{\\sigma}$ and $\\tau_{g}$ are pre-defined thresholds, are treated as significant and used as new anchors.", + "bbox": [ + 511, + 172, + 903, + 202 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Balanced Camera Distribution. As previously mentioned, the imbalanced distribution of cameras can lead to 3D-GS overfitting to street view details during joint training, causing blurring and loss of structural detail in aerial views. Inspired by NeRF Director [34], we emphasize the importance of view selection and introduce a straightforward yet effective mechanism for balancing camera distribution. Specifically, we randomly select a value from the interval $[0,1]$ . If the value is smaller than $\\frac{R}{R + 1}$ , an aerial image is chosen for training; otherwise, a street image is picked. In detail, we set $R = 2$ in the first stage to ensure thorough training of the aerial views and reduce it to 1 in the second stage.", + "bbox": [ + 511, + 208, + 906, + 404 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multi-resolution LOD Construction. Considering the differing levels of detail in aerial and street view images, and to enable real-time rendering, we adopt a Level-of-Detail (LOD) strategy, inspired by Octree-GS [23]. Specifically, for content with varying details between aerial and street views, we set different LOD levels: $K_{\\text{aerial}}$ for aerial views and $K$ for the whole views, corresponding to the two-phase training process.", + "bbox": [ + 511, + 409, + 905, + 531 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nK _ {\\text {a e r i a l}} = \\left\\lfloor \\log_ {2} \\left(D _ {\\text {a e r i a l}} / d _ {\\text {a e r i a l}}\\right) \\right\\rfloor + 1, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 539, + 903, + 563 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nK = K _ {\\text {a e r i a l}} + \\left\\lfloor \\log_ {2} \\left(d _ {\\text {a e r i a l}} / d _ {\\text {s t r e e t}}\\right) \\right\\rfloor .\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 558, + 839, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $D_{\\mathrm{aerial}}$ and $d_{\\mathrm{aerial}}$ are the $r_d$ -th largest and $r_d$ -th smallest aerial distances, while $d_{\\mathrm{street}}$ is the $r_d$ -th smallest street distance. Note that, our model is set to $K$ layers and $K_{\\mathrm{aerial}}$ is only used in the first training stage. Specifically, in the first stage, we only activate $K_{\\mathrm{aerial}}$ layers and add new anchors to the same LOD level as the significant source anchors. In the second stage, we open all layers and add the new anchors to the next LOD level to capture the higher-frequency information of street views.", + "bbox": [ + 511, + 583, + 905, + 719 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Large-scale Scene Training", + "text_level": 1, + "bbox": [ + 511, + 727, + 758, + 743 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we present a partitioning strategy to adapt our framework for large-scale urban scenes, supporting both model scalability and rendering fidelity.", + "bbox": [ + 511, + 750, + 905, + 795 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Gaussian-based methods often suffer from blurring artifacts when reconstructing large-scale scenes, mainly due to challenges in densifying extensive image sets that require fine-tuned parameters. Additionally, limited GPU memory presents a hurdle when training on large-scale scenes. A common solution is the divide-and-conquer strategy, where a whole scene is split into smaller chunks.", + "bbox": [ + 511, + 795, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "26792", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f27342e61e4c116f8653ed681e2cce2ae36885cb587c80539594c13e8ac884aa.jpg", + "image_caption": [ + "Figure 4. Visualization of our constructed dataset. All the 7 scenes contain calibrated aerial and street view images. We illustrate the scenes with the point clouds and the corresponding image capture poses. The trajectory of aerial views is shown in purple, while street views are represented in yellow. Our dataset contains 5 synthetic scenes (a-e) and 2 real scenes (f-g)." + ], + "image_footnote": [], + "bbox": [ + 98, + 89, + 903, + 311 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For instance, VastGaussian [12] progressive partitioning for aerial views but faces projection errors with street views during visibility-based camera selection. To better accommodate the reconstruction of large-scale scenes involving both aerial and street views, we make the following adjustments. Following VastGaussian, we first partition the scene into $m \\times n$ chunks based on ground-projected camera positions, then expand the original boundaries of each chunk by a threshold to ensure sufficient overlap. Next, we apply different strategies for aerial and street views: for aerial views, we augment each chunk with additional visible cameras and accessory points, for street views, which experience significant occlusion, we generate point clouds from depth maps to ensure comprehensive training coverage. All scenes follow the Manhattan world assumption, with the z-axis perpendicular to the ground plane.", + "bbox": [ + 88, + 376, + 480, + 618 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Each chunk is trained independently using the strategy discussed in Section 3.3, then merged into a single model by discarding Gaussians outside the original boundaries and concatenating Gaussian attributes across chunks. To enhance rendering speed, we convert the hybrid representation to a fully explicit one by removing view inputs from each MLP and replacing the MLPs for color with SH predictions.", + "bbox": [ + 89, + 619, + 482, + 723 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Loss Function and Regularization", + "text_level": 1, + "bbox": [ + 89, + 731, + 387, + 747 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Photometric Loss. We use different losses for rendering and surface reconstruction. For rendering, in addition to the standard L1, SSIM Loss, we also employ volume regularization $\\mathcal{L}_{\\mathrm{vol}}$ from Scaffold-GS [17] to ensure Gaussian primitives effectively cover the entire scene.", + "bbox": [ + 89, + 753, + 482, + 828 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Geometry Regularizations. For geometry awareness, we introduce depth supervision $\\mathcal{L}_d = \\frac{1}{hw}\\sum \\left|\\hat{D} - D\\right|$ , where $h$ and $w$ the rendering image dimensions, $D$ is the inverse of the rendered depth map and $D_s$ is the pseudo ground", + "bbox": [ + 89, + 834, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "truth disparity map generated by either rendering engines (for synthetic data) or a scale-aligned pretrained monocular depth model [38] (for real data). For surface reconstruction, we also incorporate the normal loss $\\mathcal{L}_n$ , as used in 2D-GS [7], to enforce normal consistency.", + "bbox": [ + 511, + 376, + 906, + 452 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Mask Regularization. To mitigate the effects of moving pedestrians, vehicles, and the sky, which is challenging for vanilla GS methods, we employ an opacity regularization", + "bbox": [ + 511, + 458, + 906, + 503 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {o} = - \\frac {1}{h w} \\sum \\sigma \\cdot \\log \\sigma - \\frac {1}{h w} \\sum M \\cdot \\log (1 - \\sigma), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 513, + 906, + 544 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "which encourages the opacity values toward 0 or 1, where $M$ denotes the mask region. Finally, the combined rendering and surface reconstruction losses, $\\mathcal{L}_R$ and $\\mathcal{L}_S$ , are defined as follows respectively:", + "bbox": [ + 511, + 553, + 905, + 613 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {R} = \\mathcal {L} _ {1} + \\lambda_ {\\text {s s i m}} \\mathcal {L} _ {\\text {s s i m}} + \\lambda_ {\\text {v o l}} \\mathcal {L} _ {\\text {v o l}} + \\lambda_ {d} \\mathcal {L} _ {d} + \\lambda_ {o} \\mathcal {L} _ {o}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 625, + 903, + 650 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {S} = \\mathcal {L} _ {R} + \\lambda_ {n} \\mathcal {L} _ {n}.\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 648, + 656, + 659 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Dataset Curation", + "text_level": 1, + "bbox": [ + 511, + 672, + 684, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As previously mentioned, the lack of calibrated aerial and street datasets is a crucial issue that restricts the research on the unified scene reconstruction. To advance the field and evaluate our method, we construct a large-scale dataset with aerial and street views calibrated, which contains five synthetic scenes and two real scenes, as illustrated in Fig. 4. More details are listed in the supplementary materials.", + "bbox": [ + 511, + 696, + 906, + 803 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Synthetic Data. Following MatrixCity [10] curation process, we create 5 additional synthetic scenes to simulate real-world data using custom data collection plugins. To ensure diversity, these synthetic datasets encompass various scales, seasons, and environments, ranging from realistic cityscapes to imaginative gaming scenarios. In addition", + "bbox": [ + 511, + 809, + 908, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "26793", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/cbad9d9ca94b44941a0c02d446785270c1eed1b32e84a5b9ea2d3db44b492d51.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SceneBlock_SmallSyntheticReal
AerialStreetAerialStreetAerialStreet
MethodMetricsPSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
2D-GS [7]24.520.7430.36522.460.7630.36024.750.7780.37423.910.7950.29319.690.5040.58520.590.6030.422
Our-2D-GS29.600.8990.12123.600.8370.21630.390.9250.13625.440.8480.21622.570.6870.35721.520.6550.344
3D-GS [8]25.440.7810.32521.810.7440.37125.510.7980.35523.990.8090.27720.090.5270.56421.410.6270.398
Scaffold-GS [17]28.440.8630.19123.840.8190.27128.790.8910.19625.140.8330.24720.180.5390.54921.220.6260.394
Hier-GS [9]28.310.8610.18923.750.8230.22028.160.8660.24425.580.8610.19621.430.6390.43022.530.6860.303
Ours30.590.9130.09423.800.8390.20931.600.9380.10125.690.8620.19023.230.7290.32122.040.6690.324
", + "bbox": [ + 94, + 88, + 906, + 227 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Quantitative comparison on small-scale datasets. Horizon-GS consistently achieves superior rendering quality compared to baselines in both aerial and street views.", + "bbox": [ + 89, + 232, + 906, + 260 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e66befcd1581aba12574b2908c1bbf67cd21706c16393d02dd7535bd7d3f1908.jpg", + "image_caption": [ + "Figure 5. Qualitative comparisons of Horizon-GS against baselines [7-9, 17] across (a) small-scale and (b) large-scale scenes." + ], + "image_footnote": [], + "bbox": [ + 94, + 266, + 903, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "to RGB images, we render depth maps for geometry supervision and initial point clouds for 3D-GS training.", + "bbox": [ + 89, + 550, + 482, + 579 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Real Data. To evaluate Horizon-GS in real-world data, we collect 2 additional scenes. We use a DJI drone equipped with five cameras to capture aerial images and a custom-designed helmet equipped with six Action4 cameras to capture street data. To enhance the performance of the structure-from-motion (SfM) technique, we include transitional views to increase image overlap and utilize a powerful commercial software, ContextCapture1. Depth maps are generated using Depth-Anything-V2 [38], and scale/offset are estimated by aligning inverse depth of SfM points. Moving objects, such as humans and cars, are removed by Grounded-SAM [24], while sky masking is handled by the pretrained SkyRemoval model [21]. More details about the dataset are provided in the supplementary materials.", + "bbox": [ + 88, + 580, + 482, + 791 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 803, + 223, + 819 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 827, + 281, + 843 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets and Metrics. We conduct comprehensive evaluations across 11 scenes containing both aerial and street", + "bbox": [ + 89, + 849, + 482, + 878 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "views, sourced from the MatrixCity dataset [10], the UC-GS dataset [43], and our captured dataset. Specifically, in the MatrixCity dataset, we evaluate two synthetic scenes: Block_Small and Block_A. For the UC-GS dataset, we select two synthetic scenes, New York City (NYC) and San Francisco (SF). Consistent with the original UC-GS settings, evaluations are performed solely on street views. For our newly captured dataset, we select one out of every 32 images for evaluation. It is worth noting that previous works only evaluate scenes from a single perspective. To improve GPU compatibility and achieve satisfactory visual results, we divide Block_A into $4 \\times 2$ chunks and divide Hillside_Morning and Hillside_Summer into $6 \\times 2$ chunks. We use image resolution $1920 \\times 1080$ for synthetic data; $1600 \\times 1066$ and $1600 \\times 900$ for real-world aerial and street view images.", + "bbox": [ + 511, + 550, + 906, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results are evaluated using standard visual quality metrics: PSNR, SSIM [30], and LPIPS [42]. Aerial and street view images are evaluated separately. To minimize the impact of pedestrians, cars, and the sky, we mask out those moving entities out when comparing with ground truths.", + "bbox": [ + 511, + 792, + 906, + 868 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We compare our method against 3D-GS [8], Scaffold-GS [17], UC-GS [43] and Hierarchical-3DGS [9]", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "1 https://www.daspatial.com/cn/gcluster", + "bbox": [ + 107, + 886, + 320, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "26794", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0e3ffbdc799d200352b43f8f635954de6b74e5818ca3708337b32e50c223ab7b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SceneBlock_AHillside_MorningHillside_Summer
AerialStreetAerialStreetAerialStreet
MethodMetricsPSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
2D-GS [7]20.630.5850.59519.570.6350.47722.020.6290.51319.200.6370.47020.870.5600.52418.700.6140.481
Our-2D-GS*28.520.8860.16623.000.7890.30129.980.9000.13520.440.6860.40428.880.8890.12620.230.6740.404
3D-GS [8]21.020.5900.59519.060.6280.47722.870.6620.48919.350.6430.46621.500.5920.50318.630.6170.480
Scaffold-GS [17]22.980.6540.52320.890.6810.43623.050.6880.43320.080.6560.46122.020.6340.43619.360.6290.473
Scaffold-GS*27.620.8600.20623.100.8080.27725.460.7960.27120.660.6850.41623.910.7520.27920.300.6670.423
Vast-GS [12]25.850.8070.29521.460.7380.35324.890.7650.29520.220.6770.42423.200.7400.29219.820.6570.429
City-GS [16]26.860.8350.24722.230.7760.30625.560.7880.26720.390.6830.42023.350.7420.28419.980.6600.427
Hier-GS* [9]26.420.7920.34023.590.8110.27525.410.7830.26921.130.6920.39823.530.7350.27320.470.6730.402
Ours*28.890.8900.15123.660.8160.25531.090.9180.10720.620.6890.39929.340.9000.11020.340.6810.393
", + "bbox": [ + 94, + 88, + 906, + 265 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/24f2b2ff49177b3d11a435aa42b1c1dfae1ca392f5b21a2460dbcbebd368ca3f.jpg", + "image_caption": [ + "Figure 6. Qualitative comparisons of Horizon-GS against UC-GS [43]. The first row shows view shifting in the SF scene, while the second row shows view shifting and rotation in the NYC scene." + ], + "image_footnote": [], + "bbox": [ + 94, + 301, + 480, + 402 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "for rendering, as well as 2D-GS [7] for surface reconstruction. To ensure full convergence, we extend the total training iterations from $30\\mathrm{k}$ to $100\\mathrm{k}$ , with densification continuing until the $50\\mathrm{k}$ iteration. For all baselines, we run their official code on all datasets except for the UC-GS dataset, where we adopt the metrics from their original paper to maintain consistency with their settings. For Hierarchical-3DGS, we report the best visual quality results after optimizing the hierarchy (leaves). For each baseline, we also add depth supervision same as our method. In the large-scale setting, we partition the scenes and train each chunk in parallel, using $*$ to denote the strategy discussed in Sec. 3.4.", + "bbox": [ + 88, + 474, + 482, + 656 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implementation Details. For a fair comparison, we train the first stage for 60k iterations and the second stage for 40k iterations, with densification stopping at 30k and 20k iterations, respectively. In the first stage, we follow the default settings of Scaffold-GS [17] for learning rate and densification. In the second stage, we reduce the learning rate of the offset by a factor of 10 and set $\\tau_{\\sigma}$ , $\\tau_{g}$ and $N$ to 0.2, 0.15 and 100, respectively. In LOD construction stage, $r_d$ is set to 0.999. The loss function parameters are $\\lambda_{\\mathrm{ssim}} = 0.2$ , $\\lambda_{\\mathrm{vol}} = 0.01$ , $\\lambda_{o} = 0.05$ , and $\\lambda_{n} = 0.05$ . The weight for $\\mathcal{L}_d$ is exponentially decayed from 1 to 0.01 over both stages. Depth supervision is activated after 500 iterations, and normal supervision after 7k iterations. We denote neural Gaussians that MLP decode into 2D as 'Our-2D-GS' and denote 'Ours' if the primitives are generated into 3D. All experi", + "bbox": [ + 89, + 672, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/cf18522c60de7cb3cafc59a25ea57c1518db573364c552a3e4c9c3b4caf5ebba.jpg", + "table_caption": [ + "Table 2. Quantitative comparison on large-scale datasets. Horizon-GS achieves superior rendering quality compared to baselines." + ], + "table_footnote": [], + "table_body": "
Scene Method MetricsHeld-outView(+1m)View(+1m 5°down)
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
3D-GS [8]23.470.6680.40620.830.6050.44021.250.6430.402
Scaffold-GS [17]25.400.7440.32022.620.6710.37523.280.7110.334
UC-GS [17]25.950.7630.29123.520.7020.34024.150.7410.298
Ours25.350.7570.28025.460.7600.28025.370.7610.278
", + "bbox": [ + 516, + 300, + 905, + 381 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Quantitative comparison on UC-GS dataset [43]. The metrics are extracted from the original paper, in which all methods are trained for 900k iterations.", + "bbox": [ + 511, + 388, + 906, + 429 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ments are conducted on NVIDIA A100 80G GPUs.", + "bbox": [ + 511, + 438, + 854, + 452 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Results Analysis", + "text_level": 1, + "bbox": [ + 511, + 465, + 674, + 481 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our evaluation covers diverse scenes, including aerial and street views, intricate and large-scale scenes, both synthetic and real-world. We demonstrate that Horizon-GS preserves fine details and significantly improves the quality of surface reconstruction, as shown in Tab. 1, 2, 3, and Fig. 5, 6, 7.", + "bbox": [ + 511, + 488, + 906, + 564 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "**Rendering Performance Analysis.** Our method reconstructs a unified Gaussian model, delivering excellent visual quality in both aerial and street views. Compared to 2D-GS [7], Our-2D-GS consistently outperforms by 4 dB in aerial PSNR and 1.5 dB in street PSNR. Furthermore, compared to all the baselines, Horizon-GS exceeds the performance of all baselines, except Hierarchical-3DGS, which is competitive in street views due to its street-focused design. As shown in the highlighted patches and arrows above of Fig. 5 (a), our method excels in fine details (1st row) and reflective regions (2nd row). In the UC-GS setting (Tab. 3), under the held-out viewpoints condition where test views share the same heights as training views, Horizon-GS, trained for 100k iterations, shows slightly lower PSNR and SSIM values compared to UC-GS trained for 900k iterations. However, our approach produces richer details, as indicated by better LPIPS metrics. Moreover, when evaluating shifted and rotated views, our method consistently outperforms the baselines, as shown in Fig. 6. Compared with synthetic data, the quality of real data declined due to performance differences caused by challenges such as pose estimation, sensor noise, and texture complexity.", + "bbox": [ + 511, + 568, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "26795", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1bdec2e4a9490e24a44b1ca1a890fd06131b76bfbcf8495ce8ba2ca75236a88e.jpg", + "image_caption": [ + "Figure 7. Qualitative comparisons of our method with 2D-GS [7]. We visualize the rendered images and geometry results (depth and world-space normal) in the Block_Small scene from both aerial and street views." + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 901, + 239 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Large-scale Rendering Performance. For the challenging large-scale scenes, we conduct two types of experiments: direct training without chunking and chunked training. As shown in Table 2, partitioning and then merging proves effective, especially when comparing Scaffold-GS and Scaffold-GS*. Horizon-GS consistently delivers excellent rendering quality in large-scale scenes compared to other baselines. In contrast, Hierarchical-3DGS [9] tends to overfit the training views, producing artifacts in novel views, while other baselines fail to capture fine-grained details (Figure 5 (b)). Each scene is divided into chunks and trained in parallel, with a total training time of about four hours per scene. Due to the LOD design, our method achieves real-time rendering speeds of 51.5 FPS in the Hillside_Morning and Hillside_Summer scenes, which are the largest scenes among our captured dataset.", + "bbox": [ + 88, + 277, + 485, + 518 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Surface Reconstruction Performance. As shown in Fig.7, our method reconstructs more complete and detailed geometry compared to 2D-GS [7]. 2D-GS produces artifacts, resulting in incomplete and lackluster geometry. In contrast, our method, using the two-stage training approach and enhanced densification, delivers detailed, geometrically accurate, and artifact-free reconstruction.", + "bbox": [ + 89, + 523, + 483, + 630 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 640, + 256, + 655 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we ablate each individual module to validate their effectiveness. We use the scenes from our proposed real dataset, Road and Park. Quantitative results can be found in Tab. 4.", + "bbox": [ + 89, + 662, + 482, + 720 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Balanced Camera Distribution. To assess the effectiveness of the balanced camera distribution (Sec. 3.3), we conduct an ablation study with randomly selected training views. The results in Tab. 4 reveal a notable decline in the visual quality of aerial views, while the quality of street views remains similar. This suggests that a well-distributed set of observation perspectives enhances the model performance, rather than converging to a single-level view.", + "bbox": [ + 89, + 728, + 482, + 849 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Balanced Densification. As outlined in Sec. 3.3, we meticulously design the densification strategy in the two training stages. Specifically, gradients of the Gaussian", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/316f49a264f15e21253c0d86dcbc188590fb882876196ea9bb83b2d6c536b39d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Scene\nMethod | MetricsAerialStreet
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Ours w/o camera bal.22.490.6840.37522.000.6710.323
Ours w/o densify bal.23.020.7210.33221.980.6700.325
Ours w/o multi LOD23.050.7170.33722.030.6720.321
Ours w/o densify poli.22.830.6900.36921.680.6550.346
Ours23.230.7290.32122.040.6690.324
", + "bbox": [ + 517, + 273, + 905, + 375 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Ablations of our method on our proposed real datasets.", + "bbox": [ + 519, + 380, + 898, + 395 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "primitives are accumulated from aerial views, while in the second stage, they are derived solely from street views. To assess the effectiveness of this approach, we compute the gradients of the Gaussian primitives across all the images in both stages. The results reveal conflicts during the densification process, highlighting the need to explicitly separate the densification steps during optimization.", + "bbox": [ + 511, + 402, + 906, + 510 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Multi-Resolution LOD. To assess the impact of LOD design, we conduct an experiment without LOD. The quality of aerial views significantly decreases, indicating that the strategy enhances the capture of fine-scale details.", + "bbox": [ + 511, + 518, + 906, + 579 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Densification Policy. An ablation of the densification policy (Eq. 1) replacing it with Scaffold-GS in the second stage shows a significant reduction in rendering details for both cross views, highlighting that average gradients alone are insufficient for capturing fine-grained local details.", + "bbox": [ + 511, + 587, + 905, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 678, + 633, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce Horizon-GS, a novel framework for large-scale aerial-to-ground scene reconstruction. We first explore the challenges of unified scene reconstruction from both aerial and street views, and propose a end-to-end coarse-to-fine training framework that mitigates inherent conflicts and is adaptable for large-scale reconstruction. Additionally, we present a comprehensive dataset with two cross views. Our experiments show that Horizon-GS outperforms existing methods in visual quality and geometry accuracy. In the future, we aim to minimize input dependency by leveraging advanced techniques such as pretrained 3D scene foundation models and enable more systematic solutions, such as distributional training.", + "bbox": [ + 511, + 704, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "26796", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments This work was supported in part by the National Key R&D Program of China (2022ZD0160201), Shanghai Artificial Intelligence Laboratory, the HKU Startup Fund, the HKU Shanghai Intelligent Computing Research Center, the center for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK, and the Anhui Provincial Natural Science Foundation under Grant 2108085UD12.", + "bbox": [ + 89, + 90, + 485, + 218 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 243, + 187, + 258 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXII, pages 333-350. Springer, 2022. 2", + "[2] Yu Chen and Gim Hee Lee. Dogaussian: Distributed-oriented gaussian splattering for large-scale 3d reconstruction via gaussian consensus. CoRR, abs/2405.13943, 2024. 2", + "[3] Yurui Chen, Chun Gu, Junzhe Jiang, Xiatian Zhu, and Li Zhang. Periodic vibration gaussian: Dynamic urban scene reconstruction and real-time rendering. CoRR, abs/2311.18561, 2023. 2", + "[4] David J. Crandall, Andrew Owens, Noah Snavely, and Daniel P. Huttenlocher. Sfm with mrfs: Discrete-continuous optimization for large-scale structure from motion. IEEE Trans. Pattern Anal. Mach. Intell., 35(12):2841-2853, 2013. 2", + "[5] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 5491-5500. IEEE, 2022. 2", + "[6] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 5354-5363. IEEE, 2024. 2", + "[7] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, SIGGRAPH 2024, Denver, CO, USA, 27 July 2024-1 August 2024, page 32. ACM, 2024. 4, 5, 6, 7, 8", + "[8] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139:1-139:14, 2023. 1, 2, 3, 6, 7", + "[9] Bernhard Kerbl, Andreas Meuleman, Georgios Kopanas, Michael Wimmer, Alexandre Lanvin, and George Drettakis. A hierarchical 3d gaussian representation for real-time rendering of very large datasets. ACM Trans. Graph., 43(4): 62:1-62:15, 2024. 2, 4, 6, 7, 8", + "[10] Yixuan Li, Lihan Jiang, Linning Xu, Yuanbo Xiangli, Zhenzhi Wang, Dahua Lin, and Bo Dai. Matrixcity: A large-" + ], + "bbox": [ + 93, + 268, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "scale city dataset for city-scale neural rendering and beyond. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 3182-3192. IEEE, 2023. 2, 5, 6", + "[11] Yiyi Liao, Jun Xie, and Andreas Geiger. KITTI-360: A novel dataset and benchmarks for urban scene understanding in 2d and 3d. IEEE Trans. Pattern Anal. Mach. Intell., 45(3):3292–3310, 2023. 2", + "[12] Jiaqi Lin, Zhihao Li, Xiao Tang, Jianzhuang Liu, Shiyong Liu, Jiayue Liu, Yangdi Lu, Xiaofei Wu, Songcen Xu, Youliang Yan, and Wenming Yang. Vastgaussian: Vast 3d gaussians for large scene reconstruction. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 5166-5175. IEEE, 2024. 2, 3, 5, 7", + "[13] Liqiang Lin, Yilin Liu, Yue Hu, Xingguang Yan, Ke Xie, and Hui Huang. Capturing, reconstructing, and simulating: The urbanscene3d dataset. In Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part VIII, pages 93-109. Springer, 2022. 2", + "[14] Jeffrey Yunfan Liu, Yun Chen, Ze Yang, Jingkang Wang, Sivabalan Manivasagam, and Raquel Urtasun. Real-time neural rasterization for large scenes. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 8382-8393. IEEE, 2023. 2", + "[15] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020. 2", + "[16] Yang Liu, Chuanchen Luo, Lue Fan, Naiyan Wang, Junran Peng, and Zhaoxiang Zhang. Citygaussian: Real-time high-quality large-scale scene rendering with gaussians. In Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part XVI, pages 265-282. Springer, 2024. 2, 3, 7", + "[17] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 20654-20664. IEEE, 2024. 3, 4, 5, 6, 7", + "[18] Zhenxing Mi and Dan Xu. Switch-nerf: Learning scene decomposition with mixture of experts for large-scale neural radiance fields. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. 2", + "[19] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I, pages 405-421. Springer, 2020. 1, 2", + "[20] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multires" + ], + "bbox": [ + 516, + 92, + 906, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "26797", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "olution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 2", + "[21] OpenDroneMap. Skyremoval. https://github.com/OpenDroneMap/SkyRemoval/, 2022.6", + "[22] Konstantinos Rematas, Andrew Liu, Pratul P. Srinivasan, Jonathan T. Barron, Andrea Tagliasacchi, Thomas A. Funkhouser, and Vittorio Ferrari. Urban radiance fields. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 12922-12932. IEEE, 2022. 2", + "[23] Kerui Ren, Lihan Jiang, Tao Lu, Mulin Yu, Linning Xu, Zhangkai Ni, and Bo Dai. Octree-gs: Towards consistent real-time rendering with lod-structured 3d gaussians. CoRR, abs/2403.17898, 2024. 2, 4", + "[24] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, Zhaoyang Zeng, Hao Zhang, Feng Li, Jie Yang, Hongyang Li, Qing Jiang, and Lei Zhang. Grounded SAM: assembling open-world models for diverse visual tasks. CoRR, abs/2401.14159, 2024. 6", + "[25] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 5449-5459. IEEE, 2022. 2", + "[26] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben P. Mildenhall, Pratul P. Srinivasan, Jonathan T. Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 8238-8248. IEEE, 2022. 2", + "[27] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. LGM: large multi-view gaussian model for high-resolution 3d content creation. In Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part IV, pages 1-18. Springer, 2024. 2", + "[28] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024. 2", + "[29] Haithem Turki, Deva Ramanan, and Mahadev Satyanarayanan. Mega-nerf: Scalable construction of large-scale nerfs for virtual fly- throughs. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 12912-12921. IEEE, 2022. 2", + "[30] Zhou Wang, Alan C. Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Trans. Image Process., 13(4): 600-612, 2004. 6", + "[31] Yuxi Wei, Zi Wang, Yifan Lu, Chenxin Xu, Changxing Liu, Hao Zhao, Siheng Chen, and Yanfeng Wang. Editable scene simulation for autonomous driving via collaborative llm-agents. In IEEE/CVF Conference on Computer Vision and" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 15077-15087. IEEE, 2024. 2", + "[32] Yuanbo Xiangli, Linning Xu, Xingang Pan, Nanxuan Zhao, Anyi Rao, Christian Theobalt, Bo Dai, and Dahua Lin. Bungeenerf: Progressive neural radiance field for extreme multi-scale scene rendering. In Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXII, pages 106-122. Springer, 2022. 2, 4", + "[33] Yuanbo Xiangli, Linning Xu, Xingang Pan, Nanxuan Zhao, Bo Dai, and Dahua Lin. Assetfield: Assets mining and reconfiguration in ground feature plane representation. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 3228-3238. IEEE, 2023. 2", + "[34] Wenhui Xiao, Rodrigo Santa Cruz, David Ahmed-Aristizabal, Olivier Salvado, Clinton Fookes, and Léo Lebrat. Nerf director: Revisiting view selection in neural volume rendering. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 20742-20751. IEEE, 2024. 4", + "[35] Linning Xu, Yuanbo Xiangli, Sida Peng, Xingang Pan, Nanxuan Zhao, Christian Theobalt, Bo Dai, and Dahua Lin. Grid-guided neural radiance fields for large urban scenes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 8296-8306. IEEE, 2023. 2", + "[36] Yinghao Xu, Zifan Shi, Yifan Wang, Hansheng Chen, Ceyuan Yang, Sida Peng, Yujun Shen, and Gordon Wetzstein. GRM: large gaussian reconstruction model for efficient 3d reconstruction and generation. In Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part XV, pages 1-20. Springer, 2024. 2", + "[37] Yunzhi Yan, Haotong Lin, Chenxu Zhou, Weijie Wang, Haiyang Sun, Kun Zhan, Xianpeng Lang, Xiaowei Zhou, and Sida Peng. Street gaussians for modeling dynamic urban scenes. CoRR, abs/2401.01339, 2024. 2", + "[38] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything V2. In Advances in Neural Information Processing Systems 38: Annual Conference on Neural Information Processing Systems 2024, NeurIPS 2024, Vancouver, BC, Canada, December 10 - 15, 2024, 2024. 5, 6", + "[39] Mulin Yu, Tao Lu, Linning Xu, Lihan Jiang, Yuanbo Xiangli, and Bo Dai. GSDF: 3dgs meets SDF for improved neural rendering and reconstruction. In Advances in Neural Information Processing Systems 38: Annual Conference on Neural Information Processing Systems 2024, NeurIPS 2024, Vancouver, BC, Canada, December 10 - 15, 2024, 2024. 2", + "[40] Zehao Yu, Torsten Sattler, and Andreas Geiger. Gaussian opacity fields: Efficient adaptive surface reconstruction in unbounded scenes. ACM Transactions on Graphics, 2024. 2", + "[41] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. GS-LRM: large reconstruction model for 3d gaussian splatting. In Computer Vision - ECCV 2024 - 18th European Conference, Mi" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "26798", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "lan, Italy, September 29-October 4, 2024, Proceedings, Part XXII, pages 1-19. Springer, 2024. 2", + "[42] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 586-595. Computer Vision Foundation / IEEE Computer Society, 2018. 6", + "[43] Saining Zhang, Baijun Ye, Xiaoxue Chen, Yuantao Chen, Zongzheng Zhang, Cheng Peng, Yongliang Shi, and Hao Zhao. Drone-assisted road gaussian splatting with cross-view uncertainty. CoRR, abs/2408.15242, 2024. 2, 6, 7", + "[44] Xiaoyu Zhou, Zhiwei Lin, Xiaojun Shan, Yongtao Wang, Deqing Sun, and Ming-Hsuan Yang. Drivinggaussian: Composite gaussian splatting for surrounding dynamic autonomous driving scenes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 21634-21643. IEEE, 2024. 2" + ], + "bbox": [ + 91, + 90, + 483, + 369 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "26799", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Horizon-GS_ Unified 3D Gaussian Splatting for Large-Scale Aerial-to-Ground Scenes/65e9755f-fe8e-4a32-88df-f20ee7cf9c9c_model.json b/2025/Horizon-GS_ Unified 3D Gaussian Splatting for Large-Scale Aerial-to-Ground Scenes/65e9755f-fe8e-4a32-88df-f20ee7cf9c9c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..dbb7c11960c584bb6775b6142936d25342c30ef9 --- /dev/null +++ b/2025/Horizon-GS_ Unified 3D Gaussian Splatting for Large-Scale Aerial-to-Ground Scenes/65e9755f-fe8e-4a32-88df-f20ee7cf9c9c_model.json @@ -0,0 +1,2037 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.198, + 0.131, + 0.803, + 0.175 + ], + "angle": 0, + "content": "Horizon-GS: Unified 3D Gaussian Splatting for Large-Scale Aerial-to-Ground Scenes" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.203, + 0.778, + 0.24 + ], + "angle": 0, + "content": "Lihan Jiang\\(^{1,3*}\\) Kerui Ren\\(^{2,3*}\\) Mulin Yu\\(^{3}\\) Linning Xu\\(^{4}\\) \nJunting Dong\\(^{3}\\) Tao Lu\\(^{5}\\) Feng Zhao\\(^{1\\dagger}\\) Dahua Lin\\(^{3,4,7}\\) Bo Dai\\(^{6\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.248, + 0.754, + 0.3 + ], + "angle": 0, + "content": "1University of Science and Technology of China, 2Shanghai Jiao Tong University, 3Shanghai Artificial Intelligence Laboratory, 4The Chinese University of Hong Kong, 5Brown University, 6The University of Hong Kong, 7CPII" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.325, + 0.898, + 0.595 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.599, + 0.908, + 0.656 + ], + "angle": 0, + "content": "Figure 1. Horizon-GS enables high-quality rendering and reconstruction of aerial-to-ground scenes with unprecedented fidelity across scales, supporting drastic view changes. The colored camera trajectories in the center illustrate the novel viewpoints, while the reconstructed mesh is overlaid on the scene. The surrounding images show the corresponding predicted images for each viewpoint. Explore more demos on our project page: https://city-super.github.io/horizon-gs/." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.683, + 0.327, + 0.699 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.717, + 0.486, + 0.869 + ], + "angle": 0, + "content": "Seamless integration of both aerial and street view images remains a significant challenge in neural scene reconstruction and rendering. Existing methods predominantly focus on single domain, limiting their applications in immersive environments, which demand extensive free view exploration with large view changes both horizontally and vertically. We introduce Horizon-GS, a novel approach built upon Gaussian Splatting techniques, tackles the unified reconstruction and rendering for aerial and street views. Our method addresses the key challenges of combining these" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.67, + 0.909, + 0.762 + ], + "angle": 0, + "content": "perspectives with a new training strategy, overcoming viewpoint discrepancies to generate high-fidelity scenes. We also curate a high-quality aerial-to-ground views dataset encompassing both synthetic and real-world scene to advance further research. Experiments across diverse urban scene datasets confirm the effectiveness of our method." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.77, + 0.646, + 0.785 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Modeling and rendering large-scale scenes has become essential across domains such as Embodied AI, digital twins, and autonomous driving. Neural radiance fields [19] have revealed the potential to realize this vision. Recently, 3D Gaussian Splatting (3D-GS) [8] has advanced this field further, achieving exceptional visual quality alongside real-time rendering speeds. These capabilities markedly accel" + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.887, + 0.355, + 0.901 + ], + "angle": 0, + "content": "*Equal contribution. †Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26789" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.151 + ], + "angle": 0, + "content": "erate the transformation of the physical world into digital spaces, supporting applications from real-to-sim simulation and autonomous navigation to immersive VR/AR experiences and the metaverse." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.155, + 0.483, + 0.426 + ], + "angle": 0, + "content": "Despite recent advancements in rendering aerial [2, 12, 16] and street views [9] with 3D-GS, existing methods remain limited to training on a single view type, resulting in a disjointed experience. We aim to bridge this gap by integrating both perspectives, enabling a seamless and unified experience across downstream applications. Our preliminary experiments on state-of-the-art Gaussian Splatting methods reveal two primary challenges in integrating aerial and street views: 1) conflicts in Gaussian densification arising from the substantial disparity between aerial and street views, and 2) an uneven distribution of camera data, with a bias toward street views. During the densification process, limited visibility from street views impedes consistent gradient accumulation, thereby disrupting the Gaussian instantiation process across the scene. Moreover, the variation in detail between aerial and street views makes a single static set of Gaussian primitives insufficiently flexible, restricting smooth transitions between viewpoints." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.429, + 0.483, + 0.7 + ], + "angle": 0, + "content": "To address above challenges and achieve a seamless integration for both aerial and street views, we propose a novel method, Horizon-GS, that overcomes the limitations of existing approaches. Specifically, we propose a coarse-to-fine training strategy that explicitly divides the entire training process into two stages: the first stage benefits more from aerial images to establish a coarse geometry, and the second stage focuses on adding finer details mainly from street views. Combined with a camera distribution balance strategy, our method achieves state-of-the-art quality for both aerial and street views, delivering superior rendering quality and geometry accuracy. Moreover, our approach is easily adaptable for large-scale reconstruction. We also notice that, the lack of calibrated aerial and street datasets is a critical issue that hinders unified model training. To advance the field and evaluate our method, we construct a large-scale dataset consisting of five synthetic and two real-world scenes, encompassing both aerial and street views." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.703, + 0.473, + 0.717 + ], + "angle": 0, + "content": "In summary, the main contributions of our method are:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.72, + 0.483, + 0.78 + ], + "angle": 0, + "content": "- We tackle the challenging and important task of unified large-scale scene reconstruction from both aerial and street views, starting with an in-depth analysis of the core conflicts that hinder integration across these perspectives." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.781, + 0.483, + 0.856 + ], + "angle": 0, + "content": "- We propose an efficient framework that addresses the inherent discrepancies and unifies the contribution from aerial and street views. Our framework is also adaptable to both 3D Gaussians for realistic novel view synthesis and 2D Gaussians for accurate geometry reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.857, + 0.483, + 0.901 + ], + "angle": 0, + "content": "- We present a high-quality and diverse dataset that includes both synthetic and real-world data, specifically curated to support balanced, cross-view reconstruction." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.72, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "- Our approach achieves state-of-the-art performance in small-scale and large-scale scene reconstruction, setting a new benchmark for unified aerial and street modeling." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.151, + 0.652, + 0.166 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.176, + 0.905, + 0.388 + ], + "angle": 0, + "content": "Neural Scene Representations. Neural Radiance Fields (NeRF) [19] have sparked significantly interest in novel view synthesis due to their photorealistic visual quality. Recent research [1, 5, 15, 20, 25, 33] has introduced various hybrid grid representations, improving efficiency by reducing sampling points and MLP layers, thus accelerating training and rendering. More recently, 3D Gaussian Splatting (3D-GS) [8] has enhanced visual quality and training efficiency by utilizing anisotropic 3D Gaussians and enabling real-time rendering with tile-based splatting. It's widely applied in various fields, especially in many large-scale applications, such as autonomous driving [3, 31, 37, 44], geometry reconstruction [6, 39, 40], and 3D generative tasks [27, 28, 36, 41]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.394, + 0.905, + 0.665 + ], + "angle": 0, + "content": "Large-scale Scene Modeling. NeRF and 3D Gaussian Splatting have sparked a new trend in neural modeling for large-scale scenes, pushing the boundaries of visual fidelity and real-time performance. Extensive neural representation methods have been developed for large-scale city reconstruction and rendering. However, most methods target either aerial [2, 12, 16, 18, 23, 29, 32, 35] or street views [9, 14, 22, 26]. For aerial views, models like Vast-Gaussian [12] and CityGaussian [16] use data partitioning for parallel training, while DoGaussian [2] employs a global node to ensure consistent inference. For street scenes, Hierarchical-3DGS [9] optimizes and renders Gaussians in an LOD hierarchy. However, these methods do not naturally support a unified reconstruction from both aerial and street views. The work closest to ours is UC-GS [43], which incorporates cross-view uncertainty to enhance details in extrapolated views, but it still struggles with significant perspective and granularity differences." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.666, + 0.905, + 0.831 + ], + "angle": 0, + "content": "Another major barrier to unified aerial and street modeling is the scarcity of datasets with calibrated images from both perspectives. Current large-scale neural rendering datasets, such as UrbanScene3D [13], Mill19 [29], Quad 6K [4], and MatrixCity [10], Waymo Block-NeRF [26], KITTI-360 [11], and Hierarchical-3DGS [9], typically cater to a single view type. MatrixCity is the only dataset with both views, but it is synthetic and lacks diversity. To fill this gap, we introduce a new dataset with a variety of synthetic and real-world scenes, supporting research on large-scale, unified reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.845, + 0.64, + 0.86 + ], + "angle": 0, + "content": "3. Horizon-GS" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "While substantial progress has been made in reconstructing urban scenes from either aerial or street images [2, 9," + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26790" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.09, + 0.538, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.54, + 0.089, + 0.778, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.787, + 0.09, + 0.893, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.267, + 0.908, + 0.325 + ], + "angle": 0, + "content": "Figure 2. Pipeline of Horizon-GS. We divide large-scale scenes into chunks. For each chunk, we initialize LOD-structured anchors and conduct the coarse-to-fine training process. Specifically, the coarse stage reconstructs the overall scene, while the fine stage enhances street view details (highlighted in purple). We can derive RGB, depth, and normal images by utilizing different primitive attributes (2D/3D Gaussians) with a single shared underlying structure." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.341, + 0.485, + 0.522 + ], + "angle": 0, + "content": "12, 16], the challenge of combining both perspectives into a unified model remains unsolved and largely overlooked. This task is non-trivial, requiring solutions to bridge the significant disparities in scale, perspective, and appearance between aerial and street views. In this section, we first analyze the core challenges in Section 3.1, briefly describe the base modules in Section 3.2, and then detail the proposed framework that leverages a two-stage training strategy and the enhanced 3D Gaussian densification operators in Section 3.3. Additionally, how to apply the framework to large-scale scenes is discussed in Section 3.4. Finally, we describe the loss functions and regularization in Section 3.5." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.532, + 0.214, + 0.549 + ], + "angle": 0, + "content": "3.1. Challenges" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.554, + 0.483, + 0.734 + ], + "angle": 0, + "content": "Before presenting our framework, we first analyze the inherent conflicts in using 3D Gaussians for simultaneous training on aerial and street views. We apply the state-of-the-art Scaffold-GS [17] on our collected real-world scene which features both aerial and street views. Naturally, we aim to combine these two classes of views during reconstruction, as they complement each other. But we notice that reconstructing either aerial or street views alone has achieved much better results than training them together, as shown in Fig. 3. Combining these perspectives presents two major challenges: (1) conflicting gradient accumulation and (2) disparity in view information coverage." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.736, + 0.483, + 0.902 + ], + "angle": 0, + "content": "The first challenge arises from the differences between aerial and street views, which lead to conflicting gradient updates during training. While street views cause the densification policy to remove irrelevant, occluded Gaussians, aerial views promote regrowth in these regions. This interference disrupts the Gaussian densification process, making it unreliable and ultimately degrading reconstruction quality. Fig. 3 confirms that joint training across both domains consistently underperformed separate training, even with extensive iterations, and with more observed views, the number of final Gaussian primitives are even less." + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.339, + 0.679, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.68, + 0.339, + 0.904, + 0.499 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.51, + 0.907, + 0.594 + ], + "angle": 0, + "content": "Figure 3. (a) Test curves for PSNR and the number of Gaussian primitives across aerial only, street only, and merged views from 15k to 100k iterations on our proposed Road scene. (b) Gradient conflicts restrict the optimization of Gaussian primitives because street views tend to exclude blue Gaussian primitives due to their lower contribution, while aerial views do the opposite." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.612, + 0.907, + 0.734 + ], + "angle": 0, + "content": "The second challenge is the imbalance in camera coverage: street views capture detailed, nearly 360-degree scenes, while aerial views are downward-focused and less frequent. This uneven distribution shifts the reconstruction focus toward street details, preventing a balanced integration of both perspectives. Together, these challenges emphasize the complexity of unified aerial-street modeling and the necessity for specialized solutions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.743, + 0.807, + 0.759 + ], + "angle": 0, + "content": "3.2. Gaussian Splitting Base Modules" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "3D Gaussian Splatting (3D-GS) [8] represents scenes using anisotropic 3D Gaussian primitives, achieving high rendering quality at significantly faster speeds. Each Gaussian \\(G\\) is defined by parameters: a center \\(\\mu \\in \\mathbb{R}^3\\), a rotation quaternion \\(q \\in \\mathbb{R}^4\\), scaling vector \\(s \\in \\mathbb{R}^3\\), opacity \\(\\sigma \\in [0,1]\\), and a color feature \\(c\\) either encoded with spherical harmonics or direct RGB feature for view-dependent/independent color. To render images, the 3D Gaussians are projected onto the image plane as 2D Gaussians \\(G'(x)\\), ordered front-to-back," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26791" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.368, + 0.106 + ], + "angle": 0, + "content": "and \\(\\alpha\\)-blended for pixel color calculation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.108, + 0.484, + 0.198 + ], + "angle": 0, + "content": "Scaffold-GS [17] further introduces anchors to reduce redundancy and enhance storage efficiency. From each anchor, \\(k\\) neural Gaussians are emitted, with centers determined by learnable offsets. Properties are decoded by MLPs from anchor features and viewing positions, creating a more robust and compact representation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.199, + 0.484, + 0.319 + ], + "angle": 0, + "content": "2D Gaussian Splating (2D-GS) [7] collapses 3D Gaussians into 2D oriented planar Gaussian disks for surface accuracy. An adaptive rasterizer efficiently renders these 2D Gaussians, and depth distortion and normal consistency losses are added to improve reconstruction quality. Finally, a mesh is extracted using a truncated signed distance function (TSDF), which fuses multi-view depth maps from the optimized 2D-GS field." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.321, + 0.484, + 0.427 + ], + "angle": 0, + "content": "For enhanced robustness, we adopt the anchor design from Scaffold-GS [17], incorporating a flexible neural Gaussian representation. Specifically, the properties of different neural Gaussians, both 3D and 2D, can be generated by learnable MLPs. We utilize neural 3D Gaussians to improve rendering quality and neural 2D Gaussians for high-fidelity geometry reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.439, + 0.392, + 0.453 + ], + "angle": 0, + "content": "3.3. Aerial-Street Joint Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.461, + 0.484, + 0.733 + ], + "angle": 0, + "content": "Coarse-to-Fine Training. The significant difference in aerial and street views lead to performance challenges when training them together, necessitating separate handling for each modality. Inspired by BungeeNeRF [32], we propose a two-stage coarse-to-fine training strategy where street views refine the details of aerial views. In the first stage, we focus primarily on the aerial views to establish a coarse geometry, accumulating gradients for densification from the aerial images alone and allowing the Gaussian primitives to fully develop and cover a broader feature space. Aerial views guide the initial placement of fine-grained 3D Gaussians, ensuring their alignment with the global scene. In the second stage, we freeze the MLP weights and attributes of the Gaussian primitives from the first stage to preserve the skeleton structure. Street views then play a key role in refining this skeleton, adding finer details and achieving a balanced integration of both views while efficiently resolving their inherent conflicts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.734, + 0.483, + 0.854 + ], + "angle": 0, + "content": "We modify the densification policy in the second stage to enhance the capture of fine-grained details in largescale scenes with sparse images. The original Scaffolds-GS method averages screen-space gradients of neural Gaussians within each voxel. While effective for object-centric scenes, it struggles to model local details in sparsely observed areas. As a result, subtle features remain underestimated, leading to blurred renderings." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Here we consider neural Gaussians with higher gradients, greater opacity, and larger projected screen size as more significant, inspired by Hierarchical-3DGS [9]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.151 + ], + "angle": 0, + "content": "Specifically, we compute the max gradients \\(\\nabla_g\\), average opacity \\(\\sigma\\) and max projected radius \\(r\\) of the included neural Gaussians for each \\(N\\) iterations. Those neural Gaussians satisfying:" + }, + { + "type": "equation", + "bbox": [ + 0.648, + 0.152, + 0.905, + 0.169 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {g} \\cdot r \\cdot \\sigma^ {\\tau_ {\\sigma}} > \\tau_ {g}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.173, + 0.905, + 0.203 + ], + "angle": 0, + "content": "where \\(\\tau_{\\sigma}\\) and \\(\\tau_{g}\\) are pre-defined thresholds, are treated as significant and used as new anchors." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.209, + 0.907, + 0.405 + ], + "angle": 0, + "content": "Balanced Camera Distribution. As previously mentioned, the imbalanced distribution of cameras can lead to 3D-GS overfitting to street view details during joint training, causing blurring and loss of structural detail in aerial views. Inspired by NeRF Director [34], we emphasize the importance of view selection and introduce a straightforward yet effective mechanism for balancing camera distribution. Specifically, we randomly select a value from the interval \\([0,1]\\). If the value is smaller than \\(\\frac{R}{R + 1}\\), an aerial image is chosen for training; otherwise, a street image is picked. In detail, we set \\(R = 2\\) in the first stage to ensure thorough training of the aerial views and reduce it to 1 in the second stage." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.41, + 0.906, + 0.532 + ], + "angle": 0, + "content": "Multi-resolution LOD Construction. Considering the differing levels of detail in aerial and street view images, and to enable real-time rendering, we adopt a Level-of-Detail (LOD) strategy, inspired by Octree-GS [23]. Specifically, for content with varying details between aerial and street views, we set different LOD levels: \\( K_{\\text{aerial}} \\) for aerial views and \\( K \\) for the whole views, corresponding to the two-phase training process." + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.54, + 0.905, + 0.564 + ], + "angle": 0, + "content": "\\[\nK _ {\\text {a e r i a l}} = \\left\\lfloor \\log_ {2} \\left(D _ {\\text {a e r i a l}} / d _ {\\text {a e r i a l}}\\right) \\right\\rfloor + 1, \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.606, + 0.559, + 0.84, + 0.576 + ], + "angle": 0, + "content": "\\[\nK = K _ {\\text {a e r i a l}} + \\left\\lfloor \\log_ {2} \\left(d _ {\\text {a e r i a l}} / d _ {\\text {s t r e e t}}\\right) \\right\\rfloor .\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.906, + 0.72 + ], + "angle": 0, + "content": "where \\(D_{\\mathrm{aerial}}\\) and \\(d_{\\mathrm{aerial}}\\) are the \\(r_d\\)-th largest and \\(r_d\\)-th smallest aerial distances, while \\(d_{\\mathrm{street}}\\) is the \\(r_d\\)-th smallest street distance. Note that, our model is set to \\(K\\) layers and \\(K_{\\mathrm{aerial}}\\) is only used in the first training stage. Specifically, in the first stage, we only activate \\(K_{\\mathrm{aerial}}\\) layers and add new anchors to the same LOD level as the significant source anchors. In the second stage, we open all layers and add the new anchors to the next LOD level to capture the higher-frequency information of street views." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.728, + 0.759, + 0.744 + ], + "angle": 0, + "content": "3.4. Large-scale Scene Training" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.906, + 0.796 + ], + "angle": 0, + "content": "In this section, we present a partitioning strategy to adapt our framework for large-scale urban scenes, supporting both model scalability and rendering fidelity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Gaussian-based methods often suffer from blurring artifacts when reconstructing large-scale scenes, mainly due to challenges in densifying extensive image sets that require fine-tuned parameters. Additionally, limited GPU memory presents a hurdle when training on large-scale scenes. A common solution is the divide-and-conquer strategy, where a whole scene is split into smaller chunks." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26792" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.09, + 0.905, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.315, + 0.908, + 0.358 + ], + "angle": 0, + "content": "Figure 4. Visualization of our constructed dataset. All the 7 scenes contain calibrated aerial and street view images. We illustrate the scenes with the point clouds and the corresponding image capture poses. The trajectory of aerial views is shown in purple, while street views are represented in yellow. Our dataset contains 5 synthetic scenes (a-e) and 2 real scenes (f-g)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.377, + 0.482, + 0.619 + ], + "angle": 0, + "content": "For instance, VastGaussian [12] progressive partitioning for aerial views but faces projection errors with street views during visibility-based camera selection. To better accommodate the reconstruction of large-scale scenes involving both aerial and street views, we make the following adjustments. Following VastGaussian, we first partition the scene into \\( m \\times n \\) chunks based on ground-projected camera positions, then expand the original boundaries of each chunk by a threshold to ensure sufficient overlap. Next, we apply different strategies for aerial and street views: for aerial views, we augment each chunk with additional visible cameras and accessory points, for street views, which experience significant occlusion, we generate point clouds from depth maps to ensure comprehensive training coverage. All scenes follow the Manhattan world assumption, with the z-axis perpendicular to the ground plane." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.62, + 0.483, + 0.724 + ], + "angle": 0, + "content": "Each chunk is trained independently using the strategy discussed in Section 3.3, then merged into a single model by discarding Gaussians outside the original boundaries and concatenating Gaussian attributes across chunks. To enhance rendering speed, we convert the hybrid representation to a fully explicit one by removing view inputs from each MLP and replacing the MLPs for color with SH predictions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.732, + 0.388, + 0.748 + ], + "angle": 0, + "content": "3.5. Loss Function and Regularization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.754, + 0.483, + 0.829 + ], + "angle": 0, + "content": "Photometric Loss. We use different losses for rendering and surface reconstruction. For rendering, in addition to the standard L1, SSIM Loss, we also employ volume regularization \\(\\mathcal{L}_{\\mathrm{vol}}\\) from Scaffold-GS [17] to ensure Gaussian primitives effectively cover the entire scene." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.835, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Geometry Regularizations. For geometry awareness, we introduce depth supervision \\(\\mathcal{L}_d = \\frac{1}{hw}\\sum \\left|\\hat{D} - D\\right|\\), where \\(h\\) and \\(w\\) the rendering image dimensions, \\(D\\) is the inverse of the rendered depth map and \\(D_s\\) is the pseudo ground" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.377, + 0.907, + 0.453 + ], + "angle": 0, + "content": "truth disparity map generated by either rendering engines (for synthetic data) or a scale-aligned pretrained monocular depth model [38] (for real data). For surface reconstruction, we also incorporate the normal loss \\(\\mathcal{L}_n\\), as used in 2D-GS [7], to enforce normal consistency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.459, + 0.907, + 0.505 + ], + "angle": 0, + "content": "Mask Regularization. To mitigate the effects of moving pedestrians, vehicles, and the sky, which is challenging for vanilla GS methods, we employ an opacity regularization" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.515, + 0.907, + 0.545 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {o} = - \\frac {1}{h w} \\sum \\sigma \\cdot \\log \\sigma - \\frac {1}{h w} \\sum M \\cdot \\log (1 - \\sigma), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.554, + 0.906, + 0.614 + ], + "angle": 0, + "content": "which encourages the opacity values toward 0 or 1, where \\(M\\) denotes the mask region. Finally, the combined rendering and surface reconstruction losses, \\(\\mathcal{L}_R\\) and \\(\\mathcal{L}_S\\), are defined as follows respectively:" + }, + { + "type": "equation", + "bbox": [ + 0.524, + 0.626, + 0.905, + 0.651 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {R} = \\mathcal {L} _ {1} + \\lambda_ {\\text {s s i m}} \\mathcal {L} _ {\\text {s s i m}} + \\lambda_ {\\text {v o l}} \\mathcal {L} _ {\\text {v o l}} + \\lambda_ {d} \\mathcal {L} _ {d} + \\lambda_ {o} \\mathcal {L} _ {o}, \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.529, + 0.65, + 0.658, + 0.66 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {S} = \\mathcal {L} _ {R} + \\lambda_ {n} \\mathcal {L} _ {n}.\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.673, + 0.685, + 0.689 + ], + "angle": 0, + "content": "4. Dataset Curation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.698, + 0.907, + 0.804 + ], + "angle": 0, + "content": "As previously mentioned, the lack of calibrated aerial and street datasets is a crucial issue that restricts the research on the unified scene reconstruction. To advance the field and evaluate our method, we construct a large-scale dataset with aerial and street views calibrated, which contains five synthetic scenes and two real scenes, as illustrated in Fig. 4. More details are listed in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Synthetic Data. Following MatrixCity [10] curation process, we create 5 additional synthetic scenes to simulate real-world data using custom data collection plugins. To ensure diversity, these synthetic datasets encompass various scales, seasons, and environments, ranging from realistic cityscapes to imaginative gaming scenarios. In addition" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26793" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.907, + 0.228 + ], + "angle": 0, + "content": "
SceneBlock_SmallSyntheticReal
AerialStreetAerialStreetAerialStreet
MethodMetricsPSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
2D-GS [7]24.520.7430.36522.460.7630.36024.750.7780.37423.910.7950.29319.690.5040.58520.590.6030.422
Our-2D-GS29.600.8990.12123.600.8370.21630.390.9250.13625.440.8480.21622.570.6870.35721.520.6550.344
3D-GS [8]25.440.7810.32521.810.7440.37125.510.7980.35523.990.8090.27720.090.5270.56421.410.6270.398
Scaffold-GS [17]28.440.8630.19123.840.8190.27128.790.8910.19625.140.8330.24720.180.5390.54921.220.6260.394
Hier-GS [9]28.310.8610.18923.750.8230.22028.160.8660.24425.580.8610.19621.430.6390.43022.530.6860.303
Ours30.590.9130.09423.800.8390.20931.600.9380.10125.690.8620.19023.230.7290.32122.040.6690.324
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.233, + 0.907, + 0.261 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison on small-scale datasets. Horizon-GS consistently achieves superior rendering quality compared to baselines in both aerial and street views." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.267, + 0.905, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.127, + 0.519, + 0.868, + 0.532 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparisons of Horizon-GS against baselines [7-9, 17] across (a) small-scale and (b) large-scale scenes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.551, + 0.483, + 0.58 + ], + "angle": 0, + "content": "to RGB images, we render depth maps for geometry supervision and initial point clouds for 3D-GS training." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.581, + 0.483, + 0.792 + ], + "angle": 0, + "content": "Real Data. To evaluate Horizon-GS in real-world data, we collect 2 additional scenes. We use a DJI drone equipped with five cameras to capture aerial images and a custom-designed helmet equipped with six Action4 cameras to capture street data. To enhance the performance of the structure-from-motion (SfM) technique, we include transitional views to increase image overlap and utilize a powerful commercial software, ContextCapture1. Depth maps are generated using Depth-Anything-V2 [38], and scale/offset are estimated by aligning inverse depth of SfM points. Moving objects, such as humans and cars, are removed by Grounded-SAM [24], while sky masking is handled by the pretrained SkyRemoval model [21]. More details about the dataset are provided in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.804, + 0.224, + 0.82 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.828, + 0.282, + 0.844 + ], + "angle": 0, + "content": "5.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.85, + 0.483, + 0.88 + ], + "angle": 0, + "content": "Datasets and Metrics. We conduct comprehensive evaluations across 11 scenes containing both aerial and street" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.551, + 0.907, + 0.793 + ], + "angle": 0, + "content": "views, sourced from the MatrixCity dataset [10], the UC-GS dataset [43], and our captured dataset. Specifically, in the MatrixCity dataset, we evaluate two synthetic scenes: Block_Small and Block_A. For the UC-GS dataset, we select two synthetic scenes, New York City (NYC) and San Francisco (SF). Consistent with the original UC-GS settings, evaluations are performed solely on street views. For our newly captured dataset, we select one out of every 32 images for evaluation. It is worth noting that previous works only evaluate scenes from a single perspective. To improve GPU compatibility and achieve satisfactory visual results, we divide Block_A into \\(4 \\times 2\\) chunks and divide Hillside_Morning and Hillside_Summer into \\(6 \\times 2\\) chunks. We use image resolution \\(1920 \\times 1080\\) for synthetic data; \\(1600 \\times 1066\\) and \\(1600 \\times 900\\) for real-world aerial and street view images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.793, + 0.907, + 0.869 + ], + "angle": 0, + "content": "Results are evaluated using standard visual quality metrics: PSNR, SSIM [30], and LPIPS [42]. Aerial and street view images are evaluated separately. To minimize the impact of pedestrians, cars, and the sky, we mask out those moving entities out when comparing with ground truths." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Baselines. We compare our method against 3D-GS [8], Scaffold-GS [17], UC-GS [43] and Hierarchical-3DGS [9]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.321, + 0.901 + ], + "angle": 0, + "content": "1 https://www.daspatial.com/cn/gcluster" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "26794" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.907, + 0.266 + ], + "angle": 0, + "content": "
SceneBlock_AHillside_MorningHillside_Summer
AerialStreetAerialStreetAerialStreet
MethodMetricsPSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
2D-GS [7]20.630.5850.59519.570.6350.47722.020.6290.51319.200.6370.47020.870.5600.52418.700.6140.481
Our-2D-GS*28.520.8860.16623.000.7890.30129.980.9000.13520.440.6860.40428.880.8890.12620.230.6740.404
3D-GS [8]21.020.5900.59519.060.6280.47722.870.6620.48919.350.6430.46621.500.5920.50318.630.6170.480
Scaffold-GS [17]22.980.6540.52320.890.6810.43623.050.6880.43320.080.6560.46122.020.6340.43619.360.6290.473
Scaffold-GS*27.620.8600.20623.100.8080.27725.460.7960.27120.660.6850.41623.910.7520.27920.300.6670.423
Vast-GS [12]25.850.8070.29521.460.7380.35324.890.7650.29520.220.6770.42423.200.7400.29219.820.6570.429
City-GS [16]26.860.8350.24722.230.7760.30625.560.7880.26720.390.6830.42023.350.7420.28419.980.6600.427
Hier-GS* [9]26.420.7920.34023.590.8110.27525.410.7830.26921.130.6920.39823.530.7350.27320.470.6730.402
Ours*28.890.8900.15123.660.8160.25531.090.9180.10720.620.6890.39929.340.9000.11020.340.6810.393
" + }, + { + "type": "table_caption", + "bbox": [ + 0.116, + 0.27, + 0.879, + 0.285 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison on large-scale datasets. Horizon-GS achieves superior rendering quality compared to baselines." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.302, + 0.481, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.415, + 0.483, + 0.457 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparisons of Horizon-GS against UC-GS [43]. The first row shows view shifting in the SF scene, while the second row shows view shifting and rotation in the NYC scene." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.475, + 0.483, + 0.657 + ], + "angle": 0, + "content": "for rendering, as well as 2D-GS [7] for surface reconstruction. To ensure full convergence, we extend the total training iterations from \\(30\\mathrm{k}\\) to \\(100\\mathrm{k}\\), with densification continuing until the \\(50\\mathrm{k}\\) iteration. For all baselines, we run their official code on all datasets except for the UC-GS dataset, where we adopt the metrics from their original paper to maintain consistency with their settings. For Hierarchical-3DGS, we report the best visual quality results after optimizing the hierarchy (leaves). For each baseline, we also add depth supervision same as our method. In the large-scale setting, we partition the scenes and train each chunk in parallel, using \\(*\\) to denote the strategy discussed in Sec. 3.4." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Implementation Details. For a fair comparison, we train the first stage for 60k iterations and the second stage for 40k iterations, with densification stopping at 30k and 20k iterations, respectively. In the first stage, we follow the default settings of Scaffold-GS [17] for learning rate and densification. In the second stage, we reduce the learning rate of the offset by a factor of 10 and set \\(\\tau_{\\sigma}\\), \\(\\tau_{g}\\) and \\(N\\) to 0.2, 0.15 and 100, respectively. In LOD construction stage, \\(r_d\\) is set to 0.999. The loss function parameters are \\(\\lambda_{\\mathrm{ssim}} = 0.2\\), \\(\\lambda_{\\mathrm{vol}} = 0.01\\), \\(\\lambda_{o} = 0.05\\), and \\(\\lambda_{n} = 0.05\\). The weight for \\(\\mathcal{L}_d\\) is exponentially decayed from 1 to 0.01 over both stages. Depth supervision is activated after 500 iterations, and normal supervision after 7k iterations. We denote neural Gaussians that MLP decode into 2D as 'Our-2D-GS' and denote 'Ours' if the primitives are generated into 3D. All experi" + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.301, + 0.906, + 0.382 + ], + "angle": 0, + "content": "
Scene Method MetricsHeld-outView(+1m)View(+1m 5°down)
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
3D-GS [8]23.470.6680.40620.830.6050.44021.250.6430.402
Scaffold-GS [17]25.400.7440.32022.620.6710.37523.280.7110.334
UC-GS [17]25.950.7630.29123.520.7020.34024.150.7410.298
Ours25.350.7570.28025.460.7600.28025.370.7610.278
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.389, + 0.907, + 0.43 + ], + "angle": 0, + "content": "Table 3. Quantitative comparison on UC-GS dataset [43]. The metrics are extracted from the original paper, in which all methods are trained for 900k iterations." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.439, + 0.855, + 0.453 + ], + "angle": 0, + "content": "ments are conducted on NVIDIA A100 80G GPUs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.466, + 0.676, + 0.482 + ], + "angle": 0, + "content": "5.2. Results Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.489, + 0.907, + 0.565 + ], + "angle": 0, + "content": "Our evaluation covers diverse scenes, including aerial and street views, intricate and large-scale scenes, both synthetic and real-world. We demonstrate that Horizon-GS preserves fine details and significantly improves the quality of surface reconstruction, as shown in Tab. 1, 2, 3, and Fig. 5, 6, 7." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.909, + 0.903 + ], + "angle": 0, + "content": "**Rendering Performance Analysis.** Our method reconstructs a unified Gaussian model, delivering excellent visual quality in both aerial and street views. Compared to 2D-GS [7], Our-2D-GS consistently outperforms by 4 dB in aerial PSNR and 1.5 dB in street PSNR. Furthermore, compared to all the baselines, Horizon-GS exceeds the performance of all baselines, except Hierarchical-3DGS, which is competitive in street views due to its street-focused design. As shown in the highlighted patches and arrows above of Fig. 5 (a), our method excels in fine details (1st row) and reflective regions (2nd row). In the UC-GS setting (Tab. 3), under the held-out viewpoints condition where test views share the same heights as training views, Horizon-GS, trained for 100k iterations, shows slightly lower PSNR and SSIM values compared to UC-GS trained for 900k iterations. However, our approach produces richer details, as indicated by better LPIPS metrics. Moreover, when evaluating shifted and rotated views, our method consistently outperforms the baselines, as shown in Fig. 6. Compared with synthetic data, the quality of real data declined due to performance differences caused by challenges such as pose estimation, sensor noise, and texture complexity." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26795" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.088, + 0.903, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.242, + 0.907, + 0.271 + ], + "angle": 0, + "content": "Figure 7. Qualitative comparisons of our method with 2D-GS [7]. We visualize the rendered images and geometry results (depth and world-space normal) in the Block_Small scene from both aerial and street views." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.278, + 0.486, + 0.52 + ], + "angle": 0, + "content": "Large-scale Rendering Performance. For the challenging large-scale scenes, we conduct two types of experiments: direct training without chunking and chunked training. As shown in Table 2, partitioning and then merging proves effective, especially when comparing Scaffold-GS and Scaffold-GS*. Horizon-GS consistently delivers excellent rendering quality in large-scale scenes compared to other baselines. In contrast, Hierarchical-3DGS [9] tends to overfit the training views, producing artifacts in novel views, while other baselines fail to capture fine-grained details (Figure 5 (b)). Each scene is divided into chunks and trained in parallel, with a total training time of about four hours per scene. Due to the LOD design, our method achieves real-time rendering speeds of 51.5 FPS in the Hillside_Morning and Hillside_Summer scenes, which are the largest scenes among our captured dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.525, + 0.485, + 0.631 + ], + "angle": 0, + "content": "Surface Reconstruction Performance. As shown in Fig.7, our method reconstructs more complete and detailed geometry compared to 2D-GS [7]. 2D-GS produces artifacts, resulting in incomplete and lackluster geometry. In contrast, our method, using the two-stage training approach and enhanced densification, delivers detailed, geometrically accurate, and artifact-free reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.641, + 0.258, + 0.656 + ], + "angle": 0, + "content": "5.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.663, + 0.483, + 0.722 + ], + "angle": 0, + "content": "In this section, we ablate each individual module to validate their effectiveness. We use the scenes from our proposed real dataset, Road and Park. Quantitative results can be found in Tab. 4." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.729, + 0.483, + 0.85 + ], + "angle": 0, + "content": "Balanced Camera Distribution. To assess the effectiveness of the balanced camera distribution (Sec. 3.3), we conduct an ablation study with randomly selected training views. The results in Tab. 4 reveal a notable decline in the visual quality of aerial views, while the quality of street views remains similar. This suggests that a well-distributed set of observation perspectives enhances the model performance, rather than converging to a single-level view." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Balanced Densification. As outlined in Sec. 3.3, we meticulously design the densification strategy in the two training stages. Specifically, gradients of the Gaussian" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.275, + 0.906, + 0.376 + ], + "angle": 0, + "content": "
Scene\nMethod | MetricsAerialStreet
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Ours w/o camera bal.22.490.6840.37522.000.6710.323
Ours w/o densify bal.23.020.7210.33221.980.6700.325
Ours w/o multi LOD23.050.7170.33722.030.6720.321
Ours w/o densify poli.22.830.6900.36921.680.6550.346
Ours23.230.7290.32122.040.6690.324
" + }, + { + "type": "table_caption", + "bbox": [ + 0.52, + 0.381, + 0.9, + 0.396 + ], + "angle": 0, + "content": "Table 4. Ablations of our method on our proposed real datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.404, + 0.907, + 0.511 + ], + "angle": 0, + "content": "primitives are accumulated from aerial views, while in the second stage, they are derived solely from street views. To assess the effectiveness of this approach, we compute the gradients of the Gaussian primitives across all the images in both stages. The results reveal conflicts during the densification process, highlighting the need to explicitly separate the densification steps during optimization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.519, + 0.907, + 0.58 + ], + "angle": 0, + "content": "Multi-Resolution LOD. To assess the impact of LOD design, we conduct an experiment without LOD. The quality of aerial views significantly decreases, indicating that the strategy enhances the capture of fine-scale details." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.588, + 0.906, + 0.664 + ], + "angle": 0, + "content": "Densification Policy. An ablation of the densification policy (Eq. 1) replacing it with Scaffold-GS in the second stage shows a significant reduction in rendering details for both cross views, highlighting that average gradients alone are insufficient for capturing fine-grained local details." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.679, + 0.634, + 0.695 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We introduce Horizon-GS, a novel framework for large-scale aerial-to-ground scene reconstruction. We first explore the challenges of unified scene reconstruction from both aerial and street views, and propose a end-to-end coarse-to-fine training framework that mitigates inherent conflicts and is adaptable for large-scale reconstruction. Additionally, we present a comprehensive dataset with two cross views. Our experiments show that Horizon-GS outperforms existing methods in visual quality and geometry accuracy. In the future, we aim to minimize input dependency by leveraging advanced techniques such as pretrained 3D scene foundation models and enable more systematic solutions, such as distributional training." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26796" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.219 + ], + "angle": 0, + "content": "Acknowledgments This work was supported in part by the National Key R&D Program of China (2022ZD0160201), Shanghai Artificial Intelligence Laboratory, the HKU Startup Fund, the HKU Shanghai Intelligent Computing Research Center, the center for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK, and the Anhui Provincial Natural Science Foundation under Grant 2108085UD12." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.244, + 0.188, + 0.26 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.269, + 0.484, + 0.338 + ], + "angle": 0, + "content": "[1] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXII, pages 333-350. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.34, + 0.484, + 0.381 + ], + "angle": 0, + "content": "[2] Yu Chen and Gim Hee Lee. Dogaussian: Distributed-oriented gaussian splattering for large-scale 3d reconstruction via gaussian consensus. CoRR, abs/2405.13943, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.382, + 0.484, + 0.436 + ], + "angle": 0, + "content": "[3] Yurui Chen, Chun Gu, Junzhe Jiang, Xiatian Zhu, and Li Zhang. Periodic vibration gaussian: Dynamic urban scene reconstruction and real-time rendering. CoRR, abs/2311.18561, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.438, + 0.484, + 0.506 + ], + "angle": 0, + "content": "[4] David J. Crandall, Andrew Owens, Noah Snavely, and Daniel P. Huttenlocher. Sfm with mrfs: Discrete-continuous optimization for large-scale structure from motion. IEEE Trans. Pattern Anal. Mach. Intell., 35(12):2841-2853, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.509, + 0.484, + 0.591 + ], + "angle": 0, + "content": "[5] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 5491-5500. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.593, + 0.484, + 0.675 + ], + "angle": 0, + "content": "[6] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 5354-5363. IEEE, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.677, + 0.484, + 0.746 + ], + "angle": 0, + "content": "[7] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, SIGGRAPH 2024, Denver, CO, USA, 27 July 2024-1 August 2024, page 32. ACM, 2024. 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.747, + 0.484, + 0.802 + ], + "angle": 0, + "content": "[8] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139:1-139:14, 2023. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.804, + 0.484, + 0.871 + ], + "angle": 0, + "content": "[9] Bernhard Kerbl, Andreas Meuleman, Georgios Kopanas, Michael Wimmer, Alexandre Lanvin, and George Drettakis. A hierarchical 3d gaussian representation for real-time rendering of very large datasets. ACM Trans. Graph., 43(4): 62:1-62:15, 2024. 2, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[10] Yixuan Li, Lihan Jiang, Linning Xu, Yuanbo Xiangli, Zhenzhi Wang, Dahua Lin, and Bo Dai. Matrixcity: A large-" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.269, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.147 + ], + "angle": 0, + "content": "scale city dataset for city-scale neural rendering and beyond. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 3182-3192. IEEE, 2023. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.907, + 0.205 + ], + "angle": 0, + "content": "[11] Yiyi Liao, Jun Xie, and Andreas Geiger. KITTI-360: A novel dataset and benchmarks for urban scene understanding in 2d and 3d. IEEE Trans. Pattern Anal. Mach. Intell., 45(3):3292–3310, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.208, + 0.906, + 0.303 + ], + "angle": 0, + "content": "[12] Jiaqi Lin, Zhihao Li, Xiao Tang, Jianzhuang Liu, Shiyong Liu, Jiayue Liu, Yangdi Lu, Xiaofei Wu, Songcen Xu, Youliang Yan, and Wenming Yang. Vastgaussian: Vast 3d gaussians for large scene reconstruction. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 5166-5175. IEEE, 2024. 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.306, + 0.906, + 0.387 + ], + "angle": 0, + "content": "[13] Liqiang Lin, Yilin Liu, Yue Hu, Xingguang Yan, Ke Xie, and Hui Huang. Capturing, reconstructing, and simulating: The urbanscene3d dataset. In Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part VIII, pages 93-109. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.906, + 0.472 + ], + "angle": 0, + "content": "[14] Jeffrey Yunfan Liu, Yun Chen, Ze Yang, Jingkang Wang, Sivabalan Manivasagam, and Raquel Urtasun. Real-time neural rasterization for large scenes. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 8382-8393. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.906, + 0.545 + ], + "angle": 0, + "content": "[15] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.906, + 0.63 + ], + "angle": 0, + "content": "[16] Yang Liu, Chuanchen Luo, Lue Fan, Naiyan Wang, Junran Peng, and Zhaoxiang Zhang. Citygaussian: Real-time high-quality large-scale scene rendering with gaussians. In Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part XVI, pages 265-282. Springer, 2024. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.632, + 0.906, + 0.714 + ], + "angle": 0, + "content": "[17] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 20654-20664. IEEE, 2024. 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.717, + 0.906, + 0.786 + ], + "angle": 0, + "content": "[18] Zhenxing Mi and Dan Xu. Switch-nerf: Learning scene decomposition with mixture of experts for large-scale neural radiance fields. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.906, + 0.871 + ], + "angle": 0, + "content": "[19] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I, pages 405-421. Springer, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.906, + 0.902 + ], + "angle": 0, + "content": "[20] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multires" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26797" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.092, + 0.483, + 0.12 + ], + "angle": 0, + "content": "olution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.122, + 0.482, + 0.149 + ], + "angle": 0, + "content": "[21] OpenDroneMap. Skyremoval. https://github.com/OpenDroneMap/SkyRemoval/, 2022.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.152, + 0.483, + 0.234 + ], + "angle": 0, + "content": "[22] Konstantinos Rematas, Andrew Liu, Pratul P. Srinivasan, Jonathan T. Barron, Andrea Tagliasacchi, Thomas A. Funkhouser, and Vittorio Ferrari. Urban radiance fields. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 12922-12932. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.237, + 0.483, + 0.291 + ], + "angle": 0, + "content": "[23] Kerui Ren, Lihan Jiang, Tao Lu, Mulin Yu, Linning Xu, Zhangkai Ni, and Bo Dai. Octree-gs: Towards consistent real-time rendering with lod-structured 3d gaussians. CoRR, abs/2403.17898, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.293, + 0.482, + 0.376 + ], + "angle": 0, + "content": "[24] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, Zhaoyang Zeng, Hao Zhang, Feng Li, Jie Yang, Hongyang Li, Qing Jiang, and Lei Zhang. Grounded SAM: assembling open-world models for diverse visual tasks. CoRR, abs/2401.14159, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.379, + 0.482, + 0.447 + ], + "angle": 0, + "content": "[25] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 5449-5459. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.449, + 0.482, + 0.545 + ], + "angle": 0, + "content": "[26] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben P. Mildenhall, Pratul P. Srinivasan, Jonathan T. Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 8238-8248. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.482, + 0.631 + ], + "angle": 0, + "content": "[27] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. LGM: large multi-view gaussian model for high-resolution 3d content creation. In Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part IV, pages 1-18. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.482, + 0.702 + ], + "angle": 0, + "content": "[28] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.482, + 0.785 + ], + "angle": 0, + "content": "[29] Haithem Turki, Deva Ramanan, and Mahadev Satyanarayanan. Mega-nerf: Scalable construction of large-scale nerfs for virtual fly- throughs. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 12912-12921. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.788, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[30] Zhou Wang, Alan C. Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Trans. Image Process., 13(4): 600-612, 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[31] Yuxi Wei, Zi Wang, Yifan Lu, Chenxin Xu, Changxing Liu, Hao Zhao, Siheng Chen, and Yanfeng Wang. Editable scene simulation for autonomous driving via collaborative llm-agents. In IEEE/CVF Conference on Computer Vision and" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 15077-15087. IEEE, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[32] Yuanbo Xiangli, Linning Xu, Xingang Pan, Nanxuan Zhao, Anyi Rao, Christian Theobalt, Bo Dai, and Dahua Lin. Bungeenerf: Progressive neural radiance field for extreme multi-scale scene rendering. In Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXII, pages 106-122. Springer, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[33] Yuanbo Xiangli, Linning Xu, Xingang Pan, Nanxuan Zhao, Bo Dai, and Dahua Lin. Assetfield: Assets mining and reconfiguration in ground feature plane representation. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 3228-3238. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.307, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[34] Wenhui Xiao, Rodrigo Santa Cruz, David Ahmed-Aristizabal, Olivier Salvado, Clinton Fookes, and Léo Lebrat. Nerf director: Revisiting view selection in neural volume rendering. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 20742-20751. IEEE, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[35] Linning Xu, Yuanbo Xiangli, Sida Peng, Xingang Pan, Nanxuan Zhao, Christian Theobalt, Bo Dai, and Dahua Lin. Grid-guided neural radiance fields for large urban scenes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 8296-8306. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.905, + 0.573 + ], + "angle": 0, + "content": "[36] Yinghao Xu, Zifan Shi, Yifan Wang, Hansheng Chen, Ceyuan Yang, Sida Peng, Yujun Shen, and Gordon Wetzstein. GRM: large gaussian reconstruction model for efficient 3d reconstruction and generation. In Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part XV, pages 1-20. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.575, + 0.905, + 0.63 + ], + "angle": 0, + "content": "[37] Yunzhi Yan, Haotong Lin, Chenxu Zhou, Weijie Wang, Haiyang Sun, Kun Zhan, Xianpeng Lang, Xiaowei Zhou, and Sida Peng. Street gaussians for modeling dynamic urban scenes. CoRR, abs/2401.01339, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.632, + 0.905, + 0.714 + ], + "angle": 0, + "content": "[38] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything V2. In Advances in Neural Information Processing Systems 38: Annual Conference on Neural Information Processing Systems 2024, NeurIPS 2024, Vancouver, BC, Canada, December 10 - 15, 2024, 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.717, + 0.905, + 0.799 + ], + "angle": 0, + "content": "[39] Mulin Yu, Tao Lu, Linning Xu, Lihan Jiang, Yuanbo Xiangli, and Bo Dai. GSDF: 3dgs meets SDF for improved neural rendering and reconstruction. In Advances in Neural Information Processing Systems 38: Annual Conference on Neural Information Processing Systems 2024, NeurIPS 2024, Vancouver, BC, Canada, December 10 - 15, 2024, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[40] Zehao Yu, Torsten Sattler, and Andreas Geiger. Gaussian opacity fields: Efficient adaptive surface reconstruction in unbounded scenes. ACM Transactions on Graphics, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[41] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. GS-LRM: large reconstruction model for 3d gaussian splatting. In Computer Vision - ECCV 2024 - 18th European Conference, Mi" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26798" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.124, + 0.092, + 0.484, + 0.12 + ], + "angle": 0, + "content": "lan, Italy, September 29-October 4, 2024, Proceedings, Part XXII, pages 1-19. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.217 + ], + "angle": 0, + "content": "[42] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 586-595. Computer Vision Foundation / IEEE Computer Society, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.219, + 0.483, + 0.275 + ], + "angle": 0, + "content": "[43] Saining Zhang, Baijun Ye, Xiaoxue Chen, Yuantao Chen, Zongzheng Zhang, Cheng Peng, Yongliang Shi, and Hao Zhao. Drone-assisted road gaussian splatting with cross-view uncertainty. CoRR, abs/2408.15242, 2024. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.483, + 0.371 + ], + "angle": 0, + "content": "[44] Xiaoyu Zhou, Zhiwei Lin, Xiaojun Shan, Yongtao Wang, Deqing Sun, and Ming-Hsuan Yang. Drivinggaussian: Composite gaussian splatting for surrounding dynamic autonomous driving scenes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 21634-21643. IEEE, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "26799" + } + ] +] \ No newline at end of file diff --git a/2025/Horizon-GS_ Unified 3D Gaussian Splatting for Large-Scale Aerial-to-Ground Scenes/65e9755f-fe8e-4a32-88df-f20ee7cf9c9c_origin.pdf b/2025/Horizon-GS_ Unified 3D Gaussian Splatting for Large-Scale Aerial-to-Ground Scenes/65e9755f-fe8e-4a32-88df-f20ee7cf9c9c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1c3dac83eb21791e1724ccf09a5b7a626e36c150 --- /dev/null +++ b/2025/Horizon-GS_ Unified 3D Gaussian Splatting for Large-Scale Aerial-to-Ground Scenes/65e9755f-fe8e-4a32-88df-f20ee7cf9c9c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:84435a9c6cb14e0f894abe87d4303436069785fe8b5ea53ff30bc00f0407ad14 +size 9385802 diff --git a/2025/Horizon-GS_ Unified 3D Gaussian Splatting for Large-Scale Aerial-to-Ground Scenes/full.md b/2025/Horizon-GS_ Unified 3D Gaussian Splatting for Large-Scale Aerial-to-Ground Scenes/full.md new file mode 100644 index 0000000000000000000000000000000000000000..55d0627af3dc699497b3b224b17fcb7f4c0e4515 --- /dev/null +++ b/2025/Horizon-GS_ Unified 3D Gaussian Splatting for Large-Scale Aerial-to-Ground Scenes/full.md @@ -0,0 +1,285 @@ +# Horizon-GS: Unified 3D Gaussian Splatting for Large-Scale Aerial-to-Ground Scenes + +Lihan Jiang $^{1,3*}$ Kerui Ren $^{2,3*}$ Mulin Yu $^{3}$ Linning Xu $^{4}$ +Junting Dong $^{3}$ Tao Lu $^{5}$ Feng Zhao $^{1\dagger}$ Dahua Lin $^{3,4,7}$ Bo Dai $^{6\dagger}$ + +1University of Science and Technology of China, 2Shanghai Jiao Tong University, 3Shanghai Artificial Intelligence Laboratory, 4The Chinese University of Hong Kong, 5Brown University, 6The University of Hong Kong, 7CPII + +![](images/06d221dd08c6e9d6305ff87fff7af2ba62b1a5a30101d6039365b7c968f93648.jpg) +Figure 1. Horizon-GS enables high-quality rendering and reconstruction of aerial-to-ground scenes with unprecedented fidelity across scales, supporting drastic view changes. The colored camera trajectories in the center illustrate the novel viewpoints, while the reconstructed mesh is overlaid on the scene. The surrounding images show the corresponding predicted images for each viewpoint. Explore more demos on our project page: https://city-super.github.io/horizon-gs/. + +# Abstract + +Seamless integration of both aerial and street view images remains a significant challenge in neural scene reconstruction and rendering. Existing methods predominantly focus on single domain, limiting their applications in immersive environments, which demand extensive free view exploration with large view changes both horizontally and vertically. We introduce Horizon-GS, a novel approach built upon Gaussian Splatting techniques, tackles the unified reconstruction and rendering for aerial and street views. Our method addresses the key challenges of combining these + +perspectives with a new training strategy, overcoming viewpoint discrepancies to generate high-fidelity scenes. We also curate a high-quality aerial-to-ground views dataset encompassing both synthetic and real-world scene to advance further research. Experiments across diverse urban scene datasets confirm the effectiveness of our method. + +# 1. Introduction + +Modeling and rendering large-scale scenes has become essential across domains such as Embodied AI, digital twins, and autonomous driving. Neural radiance fields [19] have revealed the potential to realize this vision. Recently, 3D Gaussian Splatting (3D-GS) [8] has advanced this field further, achieving exceptional visual quality alongside real-time rendering speeds. These capabilities markedly accel + +erate the transformation of the physical world into digital spaces, supporting applications from real-to-sim simulation and autonomous navigation to immersive VR/AR experiences and the metaverse. + +Despite recent advancements in rendering aerial [2, 12, 16] and street views [9] with 3D-GS, existing methods remain limited to training on a single view type, resulting in a disjointed experience. We aim to bridge this gap by integrating both perspectives, enabling a seamless and unified experience across downstream applications. Our preliminary experiments on state-of-the-art Gaussian Splatting methods reveal two primary challenges in integrating aerial and street views: 1) conflicts in Gaussian densification arising from the substantial disparity between aerial and street views, and 2) an uneven distribution of camera data, with a bias toward street views. During the densification process, limited visibility from street views impedes consistent gradient accumulation, thereby disrupting the Gaussian instantiation process across the scene. Moreover, the variation in detail between aerial and street views makes a single static set of Gaussian primitives insufficiently flexible, restricting smooth transitions between viewpoints. + +To address above challenges and achieve a seamless integration for both aerial and street views, we propose a novel method, Horizon-GS, that overcomes the limitations of existing approaches. Specifically, we propose a coarse-to-fine training strategy that explicitly divides the entire training process into two stages: the first stage benefits more from aerial images to establish a coarse geometry, and the second stage focuses on adding finer details mainly from street views. Combined with a camera distribution balance strategy, our method achieves state-of-the-art quality for both aerial and street views, delivering superior rendering quality and geometry accuracy. Moreover, our approach is easily adaptable for large-scale reconstruction. We also notice that, the lack of calibrated aerial and street datasets is a critical issue that hinders unified model training. To advance the field and evaluate our method, we construct a large-scale dataset consisting of five synthetic and two real-world scenes, encompassing both aerial and street views. + +In summary, the main contributions of our method are: + +- We tackle the challenging and important task of unified large-scale scene reconstruction from both aerial and street views, starting with an in-depth analysis of the core conflicts that hinder integration across these perspectives. +- We propose an efficient framework that addresses the inherent discrepancies and unifies the contribution from aerial and street views. Our framework is also adaptable to both 3D Gaussians for realistic novel view synthesis and 2D Gaussians for accurate geometry reconstruction. +- We present a high-quality and diverse dataset that includes both synthetic and real-world data, specifically curated to support balanced, cross-view reconstruction. + +- Our approach achieves state-of-the-art performance in small-scale and large-scale scene reconstruction, setting a new benchmark for unified aerial and street modeling. + +# 2. Related work + +Neural Scene Representations. Neural Radiance Fields (NeRF) [19] have sparked significantly interest in novel view synthesis due to their photorealistic visual quality. Recent research [1, 5, 15, 20, 25, 33] has introduced various hybrid grid representations, improving efficiency by reducing sampling points and MLP layers, thus accelerating training and rendering. More recently, 3D Gaussian Splatting (3D-GS) [8] has enhanced visual quality and training efficiency by utilizing anisotropic 3D Gaussians and enabling real-time rendering with tile-based splatting. It's widely applied in various fields, especially in many large-scale applications, such as autonomous driving [3, 31, 37, 44], geometry reconstruction [6, 39, 40], and 3D generative tasks [27, 28, 36, 41]. + +Large-scale Scene Modeling. NeRF and 3D Gaussian Splatting have sparked a new trend in neural modeling for large-scale scenes, pushing the boundaries of visual fidelity and real-time performance. Extensive neural representation methods have been developed for large-scale city reconstruction and rendering. However, most methods target either aerial [2, 12, 16, 18, 23, 29, 32, 35] or street views [9, 14, 22, 26]. For aerial views, models like Vast-Gaussian [12] and CityGaussian [16] use data partitioning for parallel training, while DoGaussian [2] employs a global node to ensure consistent inference. For street scenes, Hierarchical-3DGS [9] optimizes and renders Gaussians in an LOD hierarchy. However, these methods do not naturally support a unified reconstruction from both aerial and street views. The work closest to ours is UC-GS [43], which incorporates cross-view uncertainty to enhance details in extrapolated views, but it still struggles with significant perspective and granularity differences. + +Another major barrier to unified aerial and street modeling is the scarcity of datasets with calibrated images from both perspectives. Current large-scale neural rendering datasets, such as UrbanScene3D [13], Mill19 [29], Quad 6K [4], and MatrixCity [10], Waymo Block-NeRF [26], KITTI-360 [11], and Hierarchical-3DGS [9], typically cater to a single view type. MatrixCity is the only dataset with both views, but it is synthetic and lacks diversity. To fill this gap, we introduce a new dataset with a variety of synthetic and real-world scenes, supporting research on large-scale, unified reconstruction. + +# 3. Horizon-GS + +While substantial progress has been made in reconstructing urban scenes from either aerial or street images [2, 9, + +![](images/d66834cdb19543f17dba28bbe38c670431c73f324810969f8ee47ceae26c3650.jpg) +Figure 2. Pipeline of Horizon-GS. We divide large-scale scenes into chunks. For each chunk, we initialize LOD-structured anchors and conduct the coarse-to-fine training process. Specifically, the coarse stage reconstructs the overall scene, while the fine stage enhances street view details (highlighted in purple). We can derive RGB, depth, and normal images by utilizing different primitive attributes (2D/3D Gaussians) with a single shared underlying structure. + +![](images/d44a7aedbbb32dceea44916445ab9124412d7e1cf511ec70cdf103a213ce1cce.jpg) + +![](images/5df718579db3f61ecb3f1c7cd9a9d8c39573134fd1cee8ce9116a4dcfee920da.jpg) + +12, 16], the challenge of combining both perspectives into a unified model remains unsolved and largely overlooked. This task is non-trivial, requiring solutions to bridge the significant disparities in scale, perspective, and appearance between aerial and street views. In this section, we first analyze the core challenges in Section 3.1, briefly describe the base modules in Section 3.2, and then detail the proposed framework that leverages a two-stage training strategy and the enhanced 3D Gaussian densification operators in Section 3.3. Additionally, how to apply the framework to large-scale scenes is discussed in Section 3.4. Finally, we describe the loss functions and regularization in Section 3.5. + +# 3.1. Challenges + +Before presenting our framework, we first analyze the inherent conflicts in using 3D Gaussians for simultaneous training on aerial and street views. We apply the state-of-the-art Scaffold-GS [17] on our collected real-world scene which features both aerial and street views. Naturally, we aim to combine these two classes of views during reconstruction, as they complement each other. But we notice that reconstructing either aerial or street views alone has achieved much better results than training them together, as shown in Fig. 3. Combining these perspectives presents two major challenges: (1) conflicting gradient accumulation and (2) disparity in view information coverage. + +The first challenge arises from the differences between aerial and street views, which lead to conflicting gradient updates during training. While street views cause the densification policy to remove irrelevant, occluded Gaussians, aerial views promote regrowth in these regions. This interference disrupts the Gaussian densification process, making it unreliable and ultimately degrading reconstruction quality. Fig. 3 confirms that joint training across both domains consistently underperformed separate training, even with extensive iterations, and with more observed views, the number of final Gaussian primitives are even less. + +![](images/5b559bcf92091ab3382c7598b4eb8f857873e275a7420edf7dcd775e9a3142c9.jpg) +Figure 3. (a) Test curves for PSNR and the number of Gaussian primitives across aerial only, street only, and merged views from 15k to 100k iterations on our proposed Road scene. (b) Gradient conflicts restrict the optimization of Gaussian primitives because street views tend to exclude blue Gaussian primitives due to their lower contribution, while aerial views do the opposite. + +![](images/327d3bd8480888f947fdac6c603a15da665b42fcc214d54cbc1f89ddf719ed40.jpg) + +The second challenge is the imbalance in camera coverage: street views capture detailed, nearly 360-degree scenes, while aerial views are downward-focused and less frequent. This uneven distribution shifts the reconstruction focus toward street details, preventing a balanced integration of both perspectives. Together, these challenges emphasize the complexity of unified aerial-street modeling and the necessity for specialized solutions. + +# 3.2. Gaussian Splitting Base Modules + +3D Gaussian Splatting (3D-GS) [8] represents scenes using anisotropic 3D Gaussian primitives, achieving high rendering quality at significantly faster speeds. Each Gaussian $G$ is defined by parameters: a center $\mu \in \mathbb{R}^3$ , a rotation quaternion $q \in \mathbb{R}^4$ , scaling vector $s \in \mathbb{R}^3$ , opacity $\sigma \in [0,1]$ , and a color feature $c$ either encoded with spherical harmonics or direct RGB feature for view-dependent/independent color. To render images, the 3D Gaussians are projected onto the image plane as 2D Gaussians $G'(x)$ , ordered front-to-back, + +and $\alpha$ -blended for pixel color calculation. + +Scaffold-GS [17] further introduces anchors to reduce redundancy and enhance storage efficiency. From each anchor, $k$ neural Gaussians are emitted, with centers determined by learnable offsets. Properties are decoded by MLPs from anchor features and viewing positions, creating a more robust and compact representation. + +2D Gaussian Splating (2D-GS) [7] collapses 3D Gaussians into 2D oriented planar Gaussian disks for surface accuracy. An adaptive rasterizer efficiently renders these 2D Gaussians, and depth distortion and normal consistency losses are added to improve reconstruction quality. Finally, a mesh is extracted using a truncated signed distance function (TSDF), which fuses multi-view depth maps from the optimized 2D-GS field. + +For enhanced robustness, we adopt the anchor design from Scaffold-GS [17], incorporating a flexible neural Gaussian representation. Specifically, the properties of different neural Gaussians, both 3D and 2D, can be generated by learnable MLPs. We utilize neural 3D Gaussians to improve rendering quality and neural 2D Gaussians for high-fidelity geometry reconstruction. + +# 3.3. Aerial-Street Joint Reconstruction + +Coarse-to-Fine Training. The significant difference in aerial and street views lead to performance challenges when training them together, necessitating separate handling for each modality. Inspired by BungeeNeRF [32], we propose a two-stage coarse-to-fine training strategy where street views refine the details of aerial views. In the first stage, we focus primarily on the aerial views to establish a coarse geometry, accumulating gradients for densification from the aerial images alone and allowing the Gaussian primitives to fully develop and cover a broader feature space. Aerial views guide the initial placement of fine-grained 3D Gaussians, ensuring their alignment with the global scene. In the second stage, we freeze the MLP weights and attributes of the Gaussian primitives from the first stage to preserve the skeleton structure. Street views then play a key role in refining this skeleton, adding finer details and achieving a balanced integration of both views while efficiently resolving their inherent conflicts. + +We modify the densification policy in the second stage to enhance the capture of fine-grained details in largescale scenes with sparse images. The original Scaffolds-GS method averages screen-space gradients of neural Gaussians within each voxel. While effective for object-centric scenes, it struggles to model local details in sparsely observed areas. As a result, subtle features remain underestimated, leading to blurred renderings. + +Here we consider neural Gaussians with higher gradients, greater opacity, and larger projected screen size as more significant, inspired by Hierarchical-3DGS [9]. + +Specifically, we compute the max gradients $\nabla_g$ , average opacity $\sigma$ and max projected radius $r$ of the included neural Gaussians for each $N$ iterations. Those neural Gaussians satisfying: + +$$ +\nabla_ {g} \cdot r \cdot \sigma^ {\tau_ {\sigma}} > \tau_ {g}, \tag {1} +$$ + +where $\tau_{\sigma}$ and $\tau_{g}$ are pre-defined thresholds, are treated as significant and used as new anchors. + +Balanced Camera Distribution. As previously mentioned, the imbalanced distribution of cameras can lead to 3D-GS overfitting to street view details during joint training, causing blurring and loss of structural detail in aerial views. Inspired by NeRF Director [34], we emphasize the importance of view selection and introduce a straightforward yet effective mechanism for balancing camera distribution. Specifically, we randomly select a value from the interval $[0,1]$ . If the value is smaller than $\frac{R}{R + 1}$ , an aerial image is chosen for training; otherwise, a street image is picked. In detail, we set $R = 2$ in the first stage to ensure thorough training of the aerial views and reduce it to 1 in the second stage. + +Multi-resolution LOD Construction. Considering the differing levels of detail in aerial and street view images, and to enable real-time rendering, we adopt a Level-of-Detail (LOD) strategy, inspired by Octree-GS [23]. Specifically, for content with varying details between aerial and street views, we set different LOD levels: $K_{\text{aerial}}$ for aerial views and $K$ for the whole views, corresponding to the two-phase training process. + +$$ +K _ {\text {a e r i a l}} = \left\lfloor \log_ {2} \left(D _ {\text {a e r i a l}} / d _ {\text {a e r i a l}}\right) \right\rfloor + 1, \tag {2} +$$ + +$$ +K = K _ {\text {a e r i a l}} + \left\lfloor \log_ {2} \left(d _ {\text {a e r i a l}} / d _ {\text {s t r e e t}}\right) \right\rfloor . +$$ + +where $D_{\mathrm{aerial}}$ and $d_{\mathrm{aerial}}$ are the $r_d$ -th largest and $r_d$ -th smallest aerial distances, while $d_{\mathrm{street}}$ is the $r_d$ -th smallest street distance. Note that, our model is set to $K$ layers and $K_{\mathrm{aerial}}$ is only used in the first training stage. Specifically, in the first stage, we only activate $K_{\mathrm{aerial}}$ layers and add new anchors to the same LOD level as the significant source anchors. In the second stage, we open all layers and add the new anchors to the next LOD level to capture the higher-frequency information of street views. + +# 3.4. Large-scale Scene Training + +In this section, we present a partitioning strategy to adapt our framework for large-scale urban scenes, supporting both model scalability and rendering fidelity. + +Gaussian-based methods often suffer from blurring artifacts when reconstructing large-scale scenes, mainly due to challenges in densifying extensive image sets that require fine-tuned parameters. Additionally, limited GPU memory presents a hurdle when training on large-scale scenes. A common solution is the divide-and-conquer strategy, where a whole scene is split into smaller chunks. + +![](images/f27342e61e4c116f8653ed681e2cce2ae36885cb587c80539594c13e8ac884aa.jpg) +Figure 4. Visualization of our constructed dataset. All the 7 scenes contain calibrated aerial and street view images. We illustrate the scenes with the point clouds and the corresponding image capture poses. The trajectory of aerial views is shown in purple, while street views are represented in yellow. Our dataset contains 5 synthetic scenes (a-e) and 2 real scenes (f-g). + +For instance, VastGaussian [12] progressive partitioning for aerial views but faces projection errors with street views during visibility-based camera selection. To better accommodate the reconstruction of large-scale scenes involving both aerial and street views, we make the following adjustments. Following VastGaussian, we first partition the scene into $m \times n$ chunks based on ground-projected camera positions, then expand the original boundaries of each chunk by a threshold to ensure sufficient overlap. Next, we apply different strategies for aerial and street views: for aerial views, we augment each chunk with additional visible cameras and accessory points, for street views, which experience significant occlusion, we generate point clouds from depth maps to ensure comprehensive training coverage. All scenes follow the Manhattan world assumption, with the z-axis perpendicular to the ground plane. + +Each chunk is trained independently using the strategy discussed in Section 3.3, then merged into a single model by discarding Gaussians outside the original boundaries and concatenating Gaussian attributes across chunks. To enhance rendering speed, we convert the hybrid representation to a fully explicit one by removing view inputs from each MLP and replacing the MLPs for color with SH predictions. + +# 3.5. Loss Function and Regularization + +Photometric Loss. We use different losses for rendering and surface reconstruction. For rendering, in addition to the standard L1, SSIM Loss, we also employ volume regularization $\mathcal{L}_{\mathrm{vol}}$ from Scaffold-GS [17] to ensure Gaussian primitives effectively cover the entire scene. + +Geometry Regularizations. For geometry awareness, we introduce depth supervision $\mathcal{L}_d = \frac{1}{hw}\sum \left|\hat{D} - D\right|$ , where $h$ and $w$ the rendering image dimensions, $D$ is the inverse of the rendered depth map and $D_s$ is the pseudo ground + +truth disparity map generated by either rendering engines (for synthetic data) or a scale-aligned pretrained monocular depth model [38] (for real data). For surface reconstruction, we also incorporate the normal loss $\mathcal{L}_n$ , as used in 2D-GS [7], to enforce normal consistency. + +Mask Regularization. To mitigate the effects of moving pedestrians, vehicles, and the sky, which is challenging for vanilla GS methods, we employ an opacity regularization + +$$ +\mathcal {L} _ {o} = - \frac {1}{h w} \sum \sigma \cdot \log \sigma - \frac {1}{h w} \sum M \cdot \log (1 - \sigma), \tag {3} +$$ + +which encourages the opacity values toward 0 or 1, where $M$ denotes the mask region. Finally, the combined rendering and surface reconstruction losses, $\mathcal{L}_R$ and $\mathcal{L}_S$ , are defined as follows respectively: + +$$ +\mathcal {L} _ {R} = \mathcal {L} _ {1} + \lambda_ {\text {s s i m}} \mathcal {L} _ {\text {s s i m}} + \lambda_ {\text {v o l}} \mathcal {L} _ {\text {v o l}} + \lambda_ {d} \mathcal {L} _ {d} + \lambda_ {o} \mathcal {L} _ {o}, \tag {4} +$$ + +$$ +\mathcal {L} _ {S} = \mathcal {L} _ {R} + \lambda_ {n} \mathcal {L} _ {n}. +$$ + +# 4. Dataset Curation + +As previously mentioned, the lack of calibrated aerial and street datasets is a crucial issue that restricts the research on the unified scene reconstruction. To advance the field and evaluate our method, we construct a large-scale dataset with aerial and street views calibrated, which contains five synthetic scenes and two real scenes, as illustrated in Fig. 4. More details are listed in the supplementary materials. + +Synthetic Data. Following MatrixCity [10] curation process, we create 5 additional synthetic scenes to simulate real-world data using custom data collection plugins. To ensure diversity, these synthetic datasets encompass various scales, seasons, and environments, ranging from realistic cityscapes to imaginative gaming scenarios. In addition + +
SceneBlock_SmallSyntheticReal
AerialStreetAerialStreetAerialStreet
MethodMetricsPSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
2D-GS [7]24.520.7430.36522.460.7630.36024.750.7780.37423.910.7950.29319.690.5040.58520.590.6030.422
Our-2D-GS29.600.8990.12123.600.8370.21630.390.9250.13625.440.8480.21622.570.6870.35721.520.6550.344
3D-GS [8]25.440.7810.32521.810.7440.37125.510.7980.35523.990.8090.27720.090.5270.56421.410.6270.398
Scaffold-GS [17]28.440.8630.19123.840.8190.27128.790.8910.19625.140.8330.24720.180.5390.54921.220.6260.394
Hier-GS [9]28.310.8610.18923.750.8230.22028.160.8660.24425.580.8610.19621.430.6390.43022.530.6860.303
Ours30.590.9130.09423.800.8390.20931.600.9380.10125.690.8620.19023.230.7290.32122.040.6690.324
+ +Table 1. Quantitative comparison on small-scale datasets. Horizon-GS consistently achieves superior rendering quality compared to baselines in both aerial and street views. + +![](images/e66befcd1581aba12574b2908c1bbf67cd21706c16393d02dd7535bd7d3f1908.jpg) +Figure 5. Qualitative comparisons of Horizon-GS against baselines [7-9, 17] across (a) small-scale and (b) large-scale scenes. + +to RGB images, we render depth maps for geometry supervision and initial point clouds for 3D-GS training. + +Real Data. To evaluate Horizon-GS in real-world data, we collect 2 additional scenes. We use a DJI drone equipped with five cameras to capture aerial images and a custom-designed helmet equipped with six Action4 cameras to capture street data. To enhance the performance of the structure-from-motion (SfM) technique, we include transitional views to increase image overlap and utilize a powerful commercial software, ContextCapture1. Depth maps are generated using Depth-Anything-V2 [38], and scale/offset are estimated by aligning inverse depth of SfM points. Moving objects, such as humans and cars, are removed by Grounded-SAM [24], while sky masking is handled by the pretrained SkyRemoval model [21]. More details about the dataset are provided in the supplementary materials. + +# 5. Experiments + +# 5.1. Experimental Setup + +Datasets and Metrics. We conduct comprehensive evaluations across 11 scenes containing both aerial and street + +views, sourced from the MatrixCity dataset [10], the UC-GS dataset [43], and our captured dataset. Specifically, in the MatrixCity dataset, we evaluate two synthetic scenes: Block_Small and Block_A. For the UC-GS dataset, we select two synthetic scenes, New York City (NYC) and San Francisco (SF). Consistent with the original UC-GS settings, evaluations are performed solely on street views. For our newly captured dataset, we select one out of every 32 images for evaluation. It is worth noting that previous works only evaluate scenes from a single perspective. To improve GPU compatibility and achieve satisfactory visual results, we divide Block_A into $4 \times 2$ chunks and divide Hillside_Morning and Hillside_Summer into $6 \times 2$ chunks. We use image resolution $1920 \times 1080$ for synthetic data; $1600 \times 1066$ and $1600 \times 900$ for real-world aerial and street view images. + +Results are evaluated using standard visual quality metrics: PSNR, SSIM [30], and LPIPS [42]. Aerial and street view images are evaluated separately. To minimize the impact of pedestrians, cars, and the sky, we mask out those moving entities out when comparing with ground truths. + +Baselines. We compare our method against 3D-GS [8], Scaffold-GS [17], UC-GS [43] and Hierarchical-3DGS [9] + +
SceneBlock_AHillside_MorningHillside_Summer
AerialStreetAerialStreetAerialStreet
MethodMetricsPSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
2D-GS [7]20.630.5850.59519.570.6350.47722.020.6290.51319.200.6370.47020.870.5600.52418.700.6140.481
Our-2D-GS*28.520.8860.16623.000.7890.30129.980.9000.13520.440.6860.40428.880.8890.12620.230.6740.404
3D-GS [8]21.020.5900.59519.060.6280.47722.870.6620.48919.350.6430.46621.500.5920.50318.630.6170.480
Scaffold-GS [17]22.980.6540.52320.890.6810.43623.050.6880.43320.080.6560.46122.020.6340.43619.360.6290.473
Scaffold-GS*27.620.8600.20623.100.8080.27725.460.7960.27120.660.6850.41623.910.7520.27920.300.6670.423
Vast-GS [12]25.850.8070.29521.460.7380.35324.890.7650.29520.220.6770.42423.200.7400.29219.820.6570.429
City-GS [16]26.860.8350.24722.230.7760.30625.560.7880.26720.390.6830.42023.350.7420.28419.980.6600.427
Hier-GS* [9]26.420.7920.34023.590.8110.27525.410.7830.26921.130.6920.39823.530.7350.27320.470.6730.402
Ours*28.890.8900.15123.660.8160.25531.090.9180.10720.620.6890.39929.340.9000.11020.340.6810.393
+ +![](images/24f2b2ff49177b3d11a435aa42b1c1dfae1ca392f5b21a2460dbcbebd368ca3f.jpg) +Figure 6. Qualitative comparisons of Horizon-GS against UC-GS [43]. The first row shows view shifting in the SF scene, while the second row shows view shifting and rotation in the NYC scene. + +for rendering, as well as 2D-GS [7] for surface reconstruction. To ensure full convergence, we extend the total training iterations from $30\mathrm{k}$ to $100\mathrm{k}$ , with densification continuing until the $50\mathrm{k}$ iteration. For all baselines, we run their official code on all datasets except for the UC-GS dataset, where we adopt the metrics from their original paper to maintain consistency with their settings. For Hierarchical-3DGS, we report the best visual quality results after optimizing the hierarchy (leaves). For each baseline, we also add depth supervision same as our method. In the large-scale setting, we partition the scenes and train each chunk in parallel, using $*$ to denote the strategy discussed in Sec. 3.4. + +Implementation Details. For a fair comparison, we train the first stage for 60k iterations and the second stage for 40k iterations, with densification stopping at 30k and 20k iterations, respectively. In the first stage, we follow the default settings of Scaffold-GS [17] for learning rate and densification. In the second stage, we reduce the learning rate of the offset by a factor of 10 and set $\tau_{\sigma}$ , $\tau_{g}$ and $N$ to 0.2, 0.15 and 100, respectively. In LOD construction stage, $r_d$ is set to 0.999. The loss function parameters are $\lambda_{\mathrm{ssim}} = 0.2$ , $\lambda_{\mathrm{vol}} = 0.01$ , $\lambda_{o} = 0.05$ , and $\lambda_{n} = 0.05$ . The weight for $\mathcal{L}_d$ is exponentially decayed from 1 to 0.01 over both stages. Depth supervision is activated after 500 iterations, and normal supervision after 7k iterations. We denote neural Gaussians that MLP decode into 2D as 'Our-2D-GS' and denote 'Ours' if the primitives are generated into 3D. All experi + +Table 2. Quantitative comparison on large-scale datasets. Horizon-GS achieves superior rendering quality compared to baselines. + +
Scene Method MetricsHeld-outView(+1m)View(+1m 5°down)
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
3D-GS [8]23.470.6680.40620.830.6050.44021.250.6430.402
Scaffold-GS [17]25.400.7440.32022.620.6710.37523.280.7110.334
UC-GS [17]25.950.7630.29123.520.7020.34024.150.7410.298
Ours25.350.7570.28025.460.7600.28025.370.7610.278
+ +Table 3. Quantitative comparison on UC-GS dataset [43]. The metrics are extracted from the original paper, in which all methods are trained for 900k iterations. + +ments are conducted on NVIDIA A100 80G GPUs. + +# 5.2. Results Analysis + +Our evaluation covers diverse scenes, including aerial and street views, intricate and large-scale scenes, both synthetic and real-world. We demonstrate that Horizon-GS preserves fine details and significantly improves the quality of surface reconstruction, as shown in Tab. 1, 2, 3, and Fig. 5, 6, 7. + +**Rendering Performance Analysis.** Our method reconstructs a unified Gaussian model, delivering excellent visual quality in both aerial and street views. Compared to 2D-GS [7], Our-2D-GS consistently outperforms by 4 dB in aerial PSNR and 1.5 dB in street PSNR. Furthermore, compared to all the baselines, Horizon-GS exceeds the performance of all baselines, except Hierarchical-3DGS, which is competitive in street views due to its street-focused design. As shown in the highlighted patches and arrows above of Fig. 5 (a), our method excels in fine details (1st row) and reflective regions (2nd row). In the UC-GS setting (Tab. 3), under the held-out viewpoints condition where test views share the same heights as training views, Horizon-GS, trained for 100k iterations, shows slightly lower PSNR and SSIM values compared to UC-GS trained for 900k iterations. However, our approach produces richer details, as indicated by better LPIPS metrics. Moreover, when evaluating shifted and rotated views, our method consistently outperforms the baselines, as shown in Fig. 6. Compared with synthetic data, the quality of real data declined due to performance differences caused by challenges such as pose estimation, sensor noise, and texture complexity. + +![](images/1bdec2e4a9490e24a44b1ca1a890fd06131b76bfbcf8495ce8ba2ca75236a88e.jpg) +Figure 7. Qualitative comparisons of our method with 2D-GS [7]. We visualize the rendered images and geometry results (depth and world-space normal) in the Block_Small scene from both aerial and street views. + +Large-scale Rendering Performance. For the challenging large-scale scenes, we conduct two types of experiments: direct training without chunking and chunked training. As shown in Table 2, partitioning and then merging proves effective, especially when comparing Scaffold-GS and Scaffold-GS*. Horizon-GS consistently delivers excellent rendering quality in large-scale scenes compared to other baselines. In contrast, Hierarchical-3DGS [9] tends to overfit the training views, producing artifacts in novel views, while other baselines fail to capture fine-grained details (Figure 5 (b)). Each scene is divided into chunks and trained in parallel, with a total training time of about four hours per scene. Due to the LOD design, our method achieves real-time rendering speeds of 51.5 FPS in the Hillside_Morning and Hillside_Summer scenes, which are the largest scenes among our captured dataset. + +Surface Reconstruction Performance. As shown in Fig.7, our method reconstructs more complete and detailed geometry compared to 2D-GS [7]. 2D-GS produces artifacts, resulting in incomplete and lackluster geometry. In contrast, our method, using the two-stage training approach and enhanced densification, delivers detailed, geometrically accurate, and artifact-free reconstruction. + +# 5.3. Ablation Studies + +In this section, we ablate each individual module to validate their effectiveness. We use the scenes from our proposed real dataset, Road and Park. Quantitative results can be found in Tab. 4. + +Balanced Camera Distribution. To assess the effectiveness of the balanced camera distribution (Sec. 3.3), we conduct an ablation study with randomly selected training views. The results in Tab. 4 reveal a notable decline in the visual quality of aerial views, while the quality of street views remains similar. This suggests that a well-distributed set of observation perspectives enhances the model performance, rather than converging to a single-level view. + +Balanced Densification. As outlined in Sec. 3.3, we meticulously design the densification strategy in the two training stages. Specifically, gradients of the Gaussian + +
Scene +Method | MetricsAerialStreet
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Ours w/o camera bal.22.490.6840.37522.000.6710.323
Ours w/o densify bal.23.020.7210.33221.980.6700.325
Ours w/o multi LOD23.050.7170.33722.030.6720.321
Ours w/o densify poli.22.830.6900.36921.680.6550.346
Ours23.230.7290.32122.040.6690.324
+ +Table 4. Ablations of our method on our proposed real datasets. + +primitives are accumulated from aerial views, while in the second stage, they are derived solely from street views. To assess the effectiveness of this approach, we compute the gradients of the Gaussian primitives across all the images in both stages. The results reveal conflicts during the densification process, highlighting the need to explicitly separate the densification steps during optimization. + +Multi-Resolution LOD. To assess the impact of LOD design, we conduct an experiment without LOD. The quality of aerial views significantly decreases, indicating that the strategy enhances the capture of fine-scale details. + +Densification Policy. An ablation of the densification policy (Eq. 1) replacing it with Scaffold-GS in the second stage shows a significant reduction in rendering details for both cross views, highlighting that average gradients alone are insufficient for capturing fine-grained local details. + +# 6. Conclusion + +We introduce Horizon-GS, a novel framework for large-scale aerial-to-ground scene reconstruction. We first explore the challenges of unified scene reconstruction from both aerial and street views, and propose a end-to-end coarse-to-fine training framework that mitigates inherent conflicts and is adaptable for large-scale reconstruction. Additionally, we present a comprehensive dataset with two cross views. Our experiments show that Horizon-GS outperforms existing methods in visual quality and geometry accuracy. In the future, we aim to minimize input dependency by leveraging advanced techniques such as pretrained 3D scene foundation models and enable more systematic solutions, such as distributional training. + +Acknowledgments This work was supported in part by the National Key R&D Program of China (2022ZD0160201), Shanghai Artificial Intelligence Laboratory, the HKU Startup Fund, the HKU Shanghai Intelligent Computing Research Center, the center for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK, and the Anhui Provincial Natural Science Foundation under Grant 2108085UD12. + +# References + +[1] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXII, pages 333-350. Springer, 2022. 2 +[2] Yu Chen and Gim Hee Lee. Dogaussian: Distributed-oriented gaussian splattering for large-scale 3d reconstruction via gaussian consensus. CoRR, abs/2405.13943, 2024. 2 +[3] Yurui Chen, Chun Gu, Junzhe Jiang, Xiatian Zhu, and Li Zhang. Periodic vibration gaussian: Dynamic urban scene reconstruction and real-time rendering. CoRR, abs/2311.18561, 2023. 2 +[4] David J. Crandall, Andrew Owens, Noah Snavely, and Daniel P. Huttenlocher. Sfm with mrfs: Discrete-continuous optimization for large-scale structure from motion. IEEE Trans. Pattern Anal. Mach. Intell., 35(12):2841-2853, 2013. 2 +[5] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 5491-5500. IEEE, 2022. 2 +[6] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 5354-5363. IEEE, 2024. 2 +[7] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, SIGGRAPH 2024, Denver, CO, USA, 27 July 2024-1 August 2024, page 32. ACM, 2024. 4, 5, 6, 7, 8 +[8] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139:1-139:14, 2023. 1, 2, 3, 6, 7 +[9] Bernhard Kerbl, Andreas Meuleman, Georgios Kopanas, Michael Wimmer, Alexandre Lanvin, and George Drettakis. A hierarchical 3d gaussian representation for real-time rendering of very large datasets. ACM Trans. Graph., 43(4): 62:1-62:15, 2024. 2, 4, 6, 7, 8 +[10] Yixuan Li, Lihan Jiang, Linning Xu, Yuanbo Xiangli, Zhenzhi Wang, Dahua Lin, and Bo Dai. Matrixcity: A large- + +scale city dataset for city-scale neural rendering and beyond. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 3182-3192. IEEE, 2023. 2, 5, 6 +[11] Yiyi Liao, Jun Xie, and Andreas Geiger. KITTI-360: A novel dataset and benchmarks for urban scene understanding in 2d and 3d. IEEE Trans. Pattern Anal. Mach. Intell., 45(3):3292–3310, 2023. 2 +[12] Jiaqi Lin, Zhihao Li, Xiao Tang, Jianzhuang Liu, Shiyong Liu, Jiayue Liu, Yangdi Lu, Xiaofei Wu, Songcen Xu, Youliang Yan, and Wenming Yang. Vastgaussian: Vast 3d gaussians for large scene reconstruction. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 5166-5175. IEEE, 2024. 2, 3, 5, 7 +[13] Liqiang Lin, Yilin Liu, Yue Hu, Xingguang Yan, Ke Xie, and Hui Huang. Capturing, reconstructing, and simulating: The urbanscene3d dataset. In Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part VIII, pages 93-109. Springer, 2022. 2 +[14] Jeffrey Yunfan Liu, Yun Chen, Ze Yang, Jingkang Wang, Sivabalan Manivasagam, and Raquel Urtasun. Real-time neural rasterization for large scenes. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 8382-8393. IEEE, 2023. 2 +[15] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020. 2 +[16] Yang Liu, Chuanchen Luo, Lue Fan, Naiyan Wang, Junran Peng, and Zhaoxiang Zhang. Citygaussian: Real-time high-quality large-scale scene rendering with gaussians. In Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part XVI, pages 265-282. Springer, 2024. 2, 3, 7 +[17] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 20654-20664. IEEE, 2024. 3, 4, 5, 6, 7 +[18] Zhenxing Mi and Dan Xu. Switch-nerf: Learning scene decomposition with mixture of experts for large-scale neural radiance fields. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. 2 +[19] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I, pages 405-421. Springer, 2020. 1, 2 +[20] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multires + +olution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 2 +[21] OpenDroneMap. Skyremoval. https://github.com/OpenDroneMap/SkyRemoval/, 2022.6 +[22] Konstantinos Rematas, Andrew Liu, Pratul P. Srinivasan, Jonathan T. Barron, Andrea Tagliasacchi, Thomas A. Funkhouser, and Vittorio Ferrari. Urban radiance fields. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 12922-12932. IEEE, 2022. 2 +[23] Kerui Ren, Lihan Jiang, Tao Lu, Mulin Yu, Linning Xu, Zhangkai Ni, and Bo Dai. Octree-gs: Towards consistent real-time rendering with lod-structured 3d gaussians. CoRR, abs/2403.17898, 2024. 2, 4 +[24] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, Zhaoyang Zeng, Hao Zhang, Feng Li, Jie Yang, Hongyang Li, Qing Jiang, and Lei Zhang. Grounded SAM: assembling open-world models for diverse visual tasks. CoRR, abs/2401.14159, 2024. 6 +[25] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 5449-5459. IEEE, 2022. 2 +[26] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben P. Mildenhall, Pratul P. Srinivasan, Jonathan T. Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 8238-8248. IEEE, 2022. 2 +[27] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. LGM: large multi-view gaussian model for high-resolution 3d content creation. In Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part IV, pages 1-18. Springer, 2024. 2 +[28] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024. 2 +[29] Haithem Turki, Deva Ramanan, and Mahadev Satyanarayanan. Mega-nerf: Scalable construction of large-scale nerfs for virtual fly- throughs. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 12912-12921. IEEE, 2022. 2 +[30] Zhou Wang, Alan C. Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Trans. Image Process., 13(4): 600-612, 2004. 6 +[31] Yuxi Wei, Zi Wang, Yifan Lu, Chenxin Xu, Changxing Liu, Hao Zhao, Siheng Chen, and Yanfeng Wang. Editable scene simulation for autonomous driving via collaborative llm-agents. In IEEE/CVF Conference on Computer Vision and + +Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 15077-15087. IEEE, 2024. 2 +[32] Yuanbo Xiangli, Linning Xu, Xingang Pan, Nanxuan Zhao, Anyi Rao, Christian Theobalt, Bo Dai, and Dahua Lin. Bungeenerf: Progressive neural radiance field for extreme multi-scale scene rendering. In Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXII, pages 106-122. Springer, 2022. 2, 4 +[33] Yuanbo Xiangli, Linning Xu, Xingang Pan, Nanxuan Zhao, Bo Dai, and Dahua Lin. Assetfield: Assets mining and reconfiguration in ground feature plane representation. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 3228-3238. IEEE, 2023. 2 +[34] Wenhui Xiao, Rodrigo Santa Cruz, David Ahmed-Aristizabal, Olivier Salvado, Clinton Fookes, and Léo Lebrat. Nerf director: Revisiting view selection in neural volume rendering. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 20742-20751. IEEE, 2024. 4 +[35] Linning Xu, Yuanbo Xiangli, Sida Peng, Xingang Pan, Nanxuan Zhao, Christian Theobalt, Bo Dai, and Dahua Lin. Grid-guided neural radiance fields for large urban scenes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 8296-8306. IEEE, 2023. 2 +[36] Yinghao Xu, Zifan Shi, Yifan Wang, Hansheng Chen, Ceyuan Yang, Sida Peng, Yujun Shen, and Gordon Wetzstein. GRM: large gaussian reconstruction model for efficient 3d reconstruction and generation. In Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part XV, pages 1-20. Springer, 2024. 2 +[37] Yunzhi Yan, Haotong Lin, Chenxu Zhou, Weijie Wang, Haiyang Sun, Kun Zhan, Xianpeng Lang, Xiaowei Zhou, and Sida Peng. Street gaussians for modeling dynamic urban scenes. CoRR, abs/2401.01339, 2024. 2 +[38] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything V2. In Advances in Neural Information Processing Systems 38: Annual Conference on Neural Information Processing Systems 2024, NeurIPS 2024, Vancouver, BC, Canada, December 10 - 15, 2024, 2024. 5, 6 +[39] Mulin Yu, Tao Lu, Linning Xu, Lihan Jiang, Yuanbo Xiangli, and Bo Dai. GSDF: 3dgs meets SDF for improved neural rendering and reconstruction. In Advances in Neural Information Processing Systems 38: Annual Conference on Neural Information Processing Systems 2024, NeurIPS 2024, Vancouver, BC, Canada, December 10 - 15, 2024, 2024. 2 +[40] Zehao Yu, Torsten Sattler, and Andreas Geiger. Gaussian opacity fields: Efficient adaptive surface reconstruction in unbounded scenes. ACM Transactions on Graphics, 2024. 2 +[41] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. GS-LRM: large reconstruction model for 3d gaussian splatting. In Computer Vision - ECCV 2024 - 18th European Conference, Mi + +lan, Italy, September 29-October 4, 2024, Proceedings, Part XXII, pages 1-19. Springer, 2024. 2 +[42] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 586-595. Computer Vision Foundation / IEEE Computer Society, 2018. 6 +[43] Saining Zhang, Baijun Ye, Xiaoxue Chen, Yuantao Chen, Zongzheng Zhang, Cheng Peng, Yongliang Shi, and Hao Zhao. Drone-assisted road gaussian splatting with cross-view uncertainty. CoRR, abs/2408.15242, 2024. 2, 6, 7 +[44] Xiaoyu Zhou, Zhiwei Lin, Xiaojun Shan, Yongtao Wang, Deqing Sun, and Ming-Hsuan Yang. Drivinggaussian: Composite gaussian splatting for surrounding dynamic autonomous driving scenes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 21634-21643. IEEE, 2024. 2 \ No newline at end of file diff --git a/2025/Horizon-GS_ Unified 3D Gaussian Splatting for Large-Scale Aerial-to-Ground Scenes/images.zip b/2025/Horizon-GS_ Unified 3D Gaussian Splatting for Large-Scale Aerial-to-Ground Scenes/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5372a65b307ba7321824ad5571efc371d71989cc --- /dev/null +++ b/2025/Horizon-GS_ Unified 3D Gaussian Splatting for Large-Scale Aerial-to-Ground Scenes/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e5c6b335c4c354a9b54528b6f80ed7815c37ae00719f69ece0821bec67decf18 +size 1086085 diff --git a/2025/Horizon-GS_ Unified 3D Gaussian Splatting for Large-Scale Aerial-to-Ground Scenes/layout.json b/2025/Horizon-GS_ Unified 3D Gaussian Splatting for Large-Scale Aerial-to-Ground Scenes/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9ec0f61d7bd7368aceacac7c9b104c8507364ecb --- /dev/null +++ b/2025/Horizon-GS_ Unified 3D Gaussian Splatting for Large-Scale Aerial-to-Ground Scenes/layout.json @@ -0,0 +1,7560 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 121, + 103, + 491, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 103, + 491, + 138 + ], + "spans": [ + { + "bbox": [ + 121, + 103, + 491, + 138 + ], + "type": "text", + "content": "Horizon-GS: Unified 3D Gaussian Splatting for Large-Scale Aerial-to-Ground Scenes" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 160, + 476, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 160, + 476, + 190 + ], + "spans": [ + { + "bbox": [ + 133, + 160, + 476, + 190 + ], + "type": "text", + "content": "Lihan Jiang" + }, + { + "bbox": [ + 133, + 160, + 476, + 190 + ], + "type": "inline_equation", + "content": "^{1,3*}" + }, + { + "bbox": [ + 133, + 160, + 476, + 190 + ], + "type": "text", + "content": " Kerui Ren" + }, + { + "bbox": [ + 133, + 160, + 476, + 190 + ], + "type": "inline_equation", + "content": "^{2,3*}" + }, + { + "bbox": [ + 133, + 160, + 476, + 190 + ], + "type": "text", + "content": " Mulin Yu" + }, + { + "bbox": [ + 133, + 160, + 476, + 190 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 133, + 160, + 476, + 190 + ], + "type": "text", + "content": " Linning Xu" + }, + { + "bbox": [ + 133, + 160, + 476, + 190 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 133, + 160, + 476, + 190 + ], + "type": "text", + "content": " \nJunting Dong" + }, + { + "bbox": [ + 133, + 160, + 476, + 190 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 133, + 160, + 476, + 190 + ], + "type": "text", + "content": " Tao Lu" + }, + { + "bbox": [ + 133, + 160, + 476, + 190 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 133, + 160, + 476, + 190 + ], + "type": "text", + "content": " Feng Zhao" + }, + { + "bbox": [ + 133, + 160, + 476, + 190 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 133, + 160, + 476, + 190 + ], + "type": "text", + "content": " Dahua Lin" + }, + { + "bbox": [ + 133, + 160, + 476, + 190 + ], + "type": "inline_equation", + "content": "^{3,4,7}" + }, + { + "bbox": [ + 133, + 160, + 476, + 190 + ], + "type": "text", + "content": " Bo Dai" + }, + { + "bbox": [ + 133, + 160, + 476, + 190 + ], + "type": "inline_equation", + "content": "^{6\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 149, + 196, + 461, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 196, + 461, + 237 + ], + "spans": [ + { + "bbox": [ + 149, + 196, + 461, + 237 + ], + "type": "text", + "content": "1University of Science and Technology of China, 2Shanghai Jiao Tong University, 3Shanghai Artificial Intelligence Laboratory, 4The Chinese University of Hong Kong, 5Brown University, 6The University of Hong Kong, 7CPII" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 59, + 257, + 549, + 471 + ], + "blocks": [ + { + "bbox": [ + 59, + 257, + 549, + 471 + ], + "lines": [ + { + "bbox": [ + 59, + 257, + 549, + 471 + ], + "spans": [ + { + "bbox": [ + 59, + 257, + 549, + 471 + ], + "type": "image", + "image_path": "06d221dd08c6e9d6305ff87fff7af2ba62b1a5a30101d6039365b7c968f93648.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 474, + 555, + 519 + ], + "lines": [ + { + "bbox": [ + 55, + 474, + 555, + 519 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 555, + 519 + ], + "type": "text", + "content": "Figure 1. Horizon-GS enables high-quality rendering and reconstruction of aerial-to-ground scenes with unprecedented fidelity across scales, supporting drastic view changes. The colored camera trajectories in the center illustrate the novel viewpoints, while the reconstructed mesh is overlaid on the scene. The surrounding images show the corresponding predicted images for each viewpoint. Explore more demos on our project page: https://city-super.github.io/horizon-gs/." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 540, + 200, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 540, + 200, + 553 + ], + "spans": [ + { + "bbox": [ + 151, + 540, + 200, + 553 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 567, + 297, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 567, + 297, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 297, + 688 + ], + "type": "text", + "content": "Seamless integration of both aerial and street view images remains a significant challenge in neural scene reconstruction and rendering. Existing methods predominantly focus on single domain, limiting their applications in immersive environments, which demand extensive free view exploration with large view changes both horizontally and vertically. We introduce Horizon-GS, a novel approach built upon Gaussian Splatting techniques, tackles the unified reconstruction and rendering for aerial and street views. Our method addresses the key challenges of combining these" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 530, + 556, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 530, + 556, + 603 + ], + "spans": [ + { + "bbox": [ + 313, + 530, + 556, + 603 + ], + "type": "text", + "content": "perspectives with a new training strategy, overcoming viewpoint discrepancies to generate high-fidelity scenes. We also curate a high-quality aerial-to-ground views dataset encompassing both synthetic and real-world scene to advance further research. Experiments across diverse urban scene datasets confirm the effectiveness of our method." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 609, + 395, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 609, + 395, + 621 + ], + "spans": [ + { + "bbox": [ + 314, + 609, + 395, + 621 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": "Modeling and rendering large-scale scenes has become essential across domains such as Embodied AI, digital twins, and autonomous driving. Neural radiance fields [19] have revealed the potential to realize this vision. Recently, 3D Gaussian Splatting (3D-GS) [8] has advanced this field further, achieving exceptional visual quality alongside real-time rendering speeds. These capabilities markedly accel" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 702, + 217, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 702, + 217, + 713 + ], + "spans": [ + { + "bbox": [ + 70, + 702, + 217, + 713 + ], + "type": "text", + "content": "*Equal contribution. †Corresponding author." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26789" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "text", + "content": "erate the transformation of the physical world into digital spaces, supporting applications from real-to-sim simulation and autonomous navigation to immersive VR/AR experiences and the metaverse." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 122, + 295, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 122, + 295, + 337 + ], + "spans": [ + { + "bbox": [ + 55, + 122, + 295, + 337 + ], + "type": "text", + "content": "Despite recent advancements in rendering aerial [2, 12, 16] and street views [9] with 3D-GS, existing methods remain limited to training on a single view type, resulting in a disjointed experience. We aim to bridge this gap by integrating both perspectives, enabling a seamless and unified experience across downstream applications. Our preliminary experiments on state-of-the-art Gaussian Splatting methods reveal two primary challenges in integrating aerial and street views: 1) conflicts in Gaussian densification arising from the substantial disparity between aerial and street views, and 2) an uneven distribution of camera data, with a bias toward street views. During the densification process, limited visibility from street views impedes consistent gradient accumulation, thereby disrupting the Gaussian instantiation process across the scene. Moreover, the variation in detail between aerial and street views makes a single static set of Gaussian primitives insufficiently flexible, restricting smooth transitions between viewpoints." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 339, + 295, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 339, + 295, + 554 + ], + "spans": [ + { + "bbox": [ + 55, + 339, + 295, + 554 + ], + "type": "text", + "content": "To address above challenges and achieve a seamless integration for both aerial and street views, we propose a novel method, Horizon-GS, that overcomes the limitations of existing approaches. Specifically, we propose a coarse-to-fine training strategy that explicitly divides the entire training process into two stages: the first stage benefits more from aerial images to establish a coarse geometry, and the second stage focuses on adding finer details mainly from street views. Combined with a camera distribution balance strategy, our method achieves state-of-the-art quality for both aerial and street views, delivering superior rendering quality and geometry accuracy. Moreover, our approach is easily adaptable for large-scale reconstruction. We also notice that, the lack of calibrated aerial and street datasets is a critical issue that hinders unified model training. To advance the field and evaluate our method, we construct a large-scale dataset consisting of five synthetic and two real-world scenes, encompassing both aerial and street views." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 556, + 289, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 556, + 289, + 567 + ], + "spans": [ + { + "bbox": [ + 67, + 556, + 289, + 567 + ], + "type": "text", + "content": "In summary, the main contributions of our method are:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 570, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 56, + 570, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 295, + 617 + ], + "type": "text", + "content": "- We tackle the challenging and important task of unified large-scale scene reconstruction from both aerial and street views, starting with an in-depth analysis of the core conflicts that hinder integration across these perspectives." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 618, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 618, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 56, + 618, + 295, + 677 + ], + "type": "text", + "content": "- We propose an efficient framework that addresses the inherent discrepancies and unifies the contribution from aerial and street views. Our framework is also adaptable to both 3D Gaussians for realistic novel view synthesis and 2D Gaussians for accurate geometry reconstruction." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 678, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 678, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 678, + 295, + 713 + ], + "type": "text", + "content": "- We present a high-quality and diverse dataset that includes both synthetic and real-world data, specifically curated to support balanced, cross-view reconstruction." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 553, + 108 + ], + "type": "text", + "content": "- Our approach achieves state-of-the-art performance in small-scale and large-scale scene reconstruction, setting a new benchmark for unified aerial and street modeling." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 119, + 399, + 131 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 119, + 399, + 131 + ], + "spans": [ + { + "bbox": [ + 314, + 119, + 399, + 131 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 139, + 553, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 139, + 553, + 307 + ], + "spans": [ + { + "bbox": [ + 313, + 139, + 553, + 307 + ], + "type": "text", + "content": "Neural Scene Representations. Neural Radiance Fields (NeRF) [19] have sparked significantly interest in novel view synthesis due to their photorealistic visual quality. Recent research [1, 5, 15, 20, 25, 33] has introduced various hybrid grid representations, improving efficiency by reducing sampling points and MLP layers, thus accelerating training and rendering. More recently, 3D Gaussian Splatting (3D-GS) [8] has enhanced visual quality and training efficiency by utilizing anisotropic 3D Gaussians and enabling real-time rendering with tile-based splatting. It's widely applied in various fields, especially in many large-scale applications, such as autonomous driving [3, 31, 37, 44], geometry reconstruction [6, 39, 40], and 3D generative tasks [27, 28, 36, 41]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 312, + 553, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 553, + 526 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 553, + 526 + ], + "type": "text", + "content": "Large-scale Scene Modeling. NeRF and 3D Gaussian Splatting have sparked a new trend in neural modeling for large-scale scenes, pushing the boundaries of visual fidelity and real-time performance. Extensive neural representation methods have been developed for large-scale city reconstruction and rendering. However, most methods target either aerial [2, 12, 16, 18, 23, 29, 32, 35] or street views [9, 14, 22, 26]. For aerial views, models like Vast-Gaussian [12] and CityGaussian [16] use data partitioning for parallel training, while DoGaussian [2] employs a global node to ensure consistent inference. For street scenes, Hierarchical-3DGS [9] optimizes and renders Gaussians in an LOD hierarchy. However, these methods do not naturally support a unified reconstruction from both aerial and street views. The work closest to ours is UC-GS [43], which incorporates cross-view uncertainty to enhance details in extrapolated views, but it still struggles with significant perspective and granularity differences." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 527, + 553, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 527, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 527, + 553, + 658 + ], + "type": "text", + "content": "Another major barrier to unified aerial and street modeling is the scarcity of datasets with calibrated images from both perspectives. Current large-scale neural rendering datasets, such as UrbanScene3D [13], Mill19 [29], Quad 6K [4], and MatrixCity [10], Waymo Block-NeRF [26], KITTI-360 [11], and Hierarchical-3DGS [9], typically cater to a single view type. MatrixCity is the only dataset with both views, but it is synthetic and lacks diversity. To fill this gap, we introduce a new dataset with a variety of synthetic and real-world scenes, supporting research on large-scale, unified reconstruction." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 669, + 391, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 669, + 391, + 681 + ], + "spans": [ + { + "bbox": [ + 314, + 669, + 391, + 681 + ], + "type": "text", + "content": "3. Horizon-GS" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "While substantial progress has been made in reconstructing urban scenes from either aerial or street images [2, 9," + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26790" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 329, + 199 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 329, + 199 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 329, + 199 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 329, + 199 + ], + "type": "image", + "image_path": "d66834cdb19543f17dba28bbe38c670431c73f324810969f8ee47ceae26c3650.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 211, + 555, + 257 + ], + "lines": [ + { + "bbox": [ + 54, + 211, + 555, + 257 + ], + "spans": [ + { + "bbox": [ + 54, + 211, + 555, + 257 + ], + "type": "text", + "content": "Figure 2. Pipeline of Horizon-GS. We divide large-scale scenes into chunks. For each chunk, we initialize LOD-structured anchors and conduct the coarse-to-fine training process. Specifically, the coarse stage reconstructs the overall scene, while the fine stage enhances street view details (highlighted in purple). We can derive RGB, depth, and normal images by utilizing different primitive attributes (2D/3D Gaussians) with a single shared underlying structure." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 330, + 70, + 476, + 200 + ], + "blocks": [ + { + "bbox": [ + 330, + 70, + 476, + 200 + ], + "lines": [ + { + "bbox": [ + 330, + 70, + 476, + 200 + ], + "spans": [ + { + "bbox": [ + 330, + 70, + 476, + 200 + ], + "type": "image", + "image_path": "d44a7aedbbb32dceea44916445ab9124412d7e1cf511ec70cdf103a213ce1cce.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 481, + 71, + 546, + 199 + ], + "blocks": [ + { + "bbox": [ + 481, + 71, + 546, + 199 + ], + "lines": [ + { + "bbox": [ + 481, + 71, + 546, + 199 + ], + "spans": [ + { + "bbox": [ + 481, + 71, + 546, + 199 + ], + "type": "image", + "image_path": "5df718579db3f61ecb3f1c7cd9a9d8c39573134fd1cee8ce9116a4dcfee920da.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 270, + 296, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 270, + 296, + 413 + ], + "spans": [ + { + "bbox": [ + 54, + 270, + 296, + 413 + ], + "type": "text", + "content": "12, 16], the challenge of combining both perspectives into a unified model remains unsolved and largely overlooked. This task is non-trivial, requiring solutions to bridge the significant disparities in scale, perspective, and appearance between aerial and street views. In this section, we first analyze the core challenges in Section 3.1, briefly describe the base modules in Section 3.2, and then detail the proposed framework that leverages a two-stage training strategy and the enhanced 3D Gaussian densification operators in Section 3.3. Additionally, how to apply the framework to large-scale scenes is discussed in Section 3.4. Finally, we describe the loss functions and regularization in Section 3.5." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 421, + 130, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 421, + 130, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 421, + 130, + 434 + ], + "type": "text", + "content": "3.1. Challenges" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 438, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 438, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 54, + 438, + 295, + 581 + ], + "type": "text", + "content": "Before presenting our framework, we first analyze the inherent conflicts in using 3D Gaussians for simultaneous training on aerial and street views. We apply the state-of-the-art Scaffold-GS [17] on our collected real-world scene which features both aerial and street views. Naturally, we aim to combine these two classes of views during reconstruction, as they complement each other. But we notice that reconstructing either aerial or street views alone has achieved much better results than training them together, as shown in Fig. 3. Combining these perspectives presents two major challenges: (1) conflicting gradient accumulation and (2) disparity in view information coverage." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "type": "text", + "content": "The first challenge arises from the differences between aerial and street views, which lead to conflicting gradient updates during training. While street views cause the densification policy to remove irrelevant, occluded Gaussians, aerial views promote regrowth in these regions. This interference disrupts the Gaussian densification process, making it unreliable and ultimately degrading reconstruction quality. Fig. 3 confirms that joint training across both domains consistently underperformed separate training, even with extensive iterations, and with more observed views, the number of final Gaussian primitives are even less." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 320, + 268, + 415, + 393 + ], + "blocks": [ + { + "bbox": [ + 320, + 268, + 415, + 393 + ], + "lines": [ + { + "bbox": [ + 320, + 268, + 415, + 393 + ], + "spans": [ + { + "bbox": [ + 320, + 268, + 415, + 393 + ], + "type": "image", + "image_path": "5b559bcf92091ab3382c7598b4eb8f857873e275a7420edf7dcd775e9a3142c9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 403, + 555, + 470 + ], + "lines": [ + { + "bbox": [ + 313, + 403, + 555, + 470 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 555, + 470 + ], + "type": "text", + "content": "Figure 3. (a) Test curves for PSNR and the number of Gaussian primitives across aerial only, street only, and merged views from 15k to 100k iterations on our proposed Road scene. (b) Gradient conflicts restrict the optimization of Gaussian primitives because street views tend to exclude blue Gaussian primitives due to their lower contribution, while aerial views do the opposite." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 416, + 268, + 553, + 395 + ], + "blocks": [ + { + "bbox": [ + 416, + 268, + 553, + 395 + ], + "lines": [ + { + "bbox": [ + 416, + 268, + 553, + 395 + ], + "spans": [ + { + "bbox": [ + 416, + 268, + 553, + 395 + ], + "type": "image", + "image_path": "327d3bd8480888f947fdac6c603a15da665b42fcc214d54cbc1f89ddf719ed40.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 484, + 555, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 484, + 555, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 555, + 581 + ], + "type": "text", + "content": "The second challenge is the imbalance in camera coverage: street views capture detailed, nearly 360-degree scenes, while aerial views are downward-focused and less frequent. This uneven distribution shifts the reconstruction focus toward street details, preventing a balanced integration of both perspectives. Together, these challenges emphasize the complexity of unified aerial-street modeling and the necessity for specialized solutions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 588, + 493, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 588, + 493, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 493, + 601 + ], + "type": "text", + "content": "3.2. Gaussian Splitting Base Modules" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "3D Gaussian Splatting (3D-GS) [8] represents scenes using anisotropic 3D Gaussian primitives, achieving high rendering quality at significantly faster speeds. Each Gaussian " + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": " is defined by parameters: a center " + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mu \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": ", a rotation quaternion " + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "inline_equation", + "content": "q \\in \\mathbb{R}^4" + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": ", scaling vector " + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "inline_equation", + "content": "s \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": ", opacity " + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\sigma \\in [0,1]" + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": ", and a color feature " + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": " either encoded with spherical harmonics or direct RGB feature for view-dependent/independent color. To render images, the 3D Gaussians are projected onto the image plane as 2D Gaussians " + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "inline_equation", + "content": "G'(x)" + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": ", ordered front-to-back," + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26791" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 225, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 225, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 225, + 83 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 55, + 72, + 225, + 83 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 72, + 225, + 83 + ], + "type": "text", + "content": "-blended for pixel color calculation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 85, + 296, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 85, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 85, + 296, + 156 + ], + "type": "text", + "content": "Scaffold-GS [17] further introduces anchors to reduce redundancy and enhance storage efficiency. From each anchor, " + }, + { + "bbox": [ + 55, + 85, + 296, + 156 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 85, + 296, + 156 + ], + "type": "text", + "content": " neural Gaussians are emitted, with centers determined by learnable offsets. Properties are decoded by MLPs from anchor features and viewing positions, creating a more robust and compact representation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 157, + 296, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 157, + 296, + 252 + ], + "spans": [ + { + "bbox": [ + 55, + 157, + 296, + 252 + ], + "type": "text", + "content": "2D Gaussian Splating (2D-GS) [7] collapses 3D Gaussians into 2D oriented planar Gaussian disks for surface accuracy. An adaptive rasterizer efficiently renders these 2D Gaussians, and depth distortion and normal consistency losses are added to improve reconstruction quality. Finally, a mesh is extracted using a truncated signed distance function (TSDF), which fuses multi-view depth maps from the optimized 2D-GS field." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 254, + 296, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 254, + 296, + 338 + ], + "spans": [ + { + "bbox": [ + 55, + 254, + 296, + 338 + ], + "type": "text", + "content": "For enhanced robustness, we adopt the anchor design from Scaffold-GS [17], incorporating a flexible neural Gaussian representation. Specifically, the properties of different neural Gaussians, both 3D and 2D, can be generated by learnable MLPs. We utilize neural 3D Gaussians to improve rendering quality and neural 2D Gaussians for high-fidelity geometry reconstruction." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 347, + 239, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 347, + 239, + 358 + ], + "spans": [ + { + "bbox": [ + 55, + 347, + 239, + 358 + ], + "type": "text", + "content": "3.3. Aerial-Street Joint Reconstruction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 365, + 296, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 296, + 580 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 296, + 580 + ], + "type": "text", + "content": "Coarse-to-Fine Training. The significant difference in aerial and street views lead to performance challenges when training them together, necessitating separate handling for each modality. Inspired by BungeeNeRF [32], we propose a two-stage coarse-to-fine training strategy where street views refine the details of aerial views. In the first stage, we focus primarily on the aerial views to establish a coarse geometry, accumulating gradients for densification from the aerial images alone and allowing the Gaussian primitives to fully develop and cover a broader feature space. Aerial views guide the initial placement of fine-grained 3D Gaussians, ensuring their alignment with the global scene. In the second stage, we freeze the MLP weights and attributes of the Gaussian primitives from the first stage to preserve the skeleton structure. Street views then play a key role in refining this skeleton, adding finer details and achieving a balanced integration of both views while efficiently resolving their inherent conflicts." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 581, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 581, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 581, + 295, + 676 + ], + "type": "text", + "content": "We modify the densification policy in the second stage to enhance the capture of fine-grained details in largescale scenes with sparse images. The original Scaffolds-GS method averages screen-space gradients of neural Gaussians within each voxel. While effective for object-centric scenes, it struggles to model local details in sparsely observed areas. As a result, subtle features remain underestimated, leading to blurred renderings." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Here we consider neural Gaussians with higher gradients, greater opacity, and larger projected screen size as more significant, inspired by Hierarchical-3DGS [9]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 554, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 119 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 119 + ], + "type": "text", + "content": "Specifically, we compute the max gradients " + }, + { + "bbox": [ + 313, + 72, + 554, + 119 + ], + "type": "inline_equation", + "content": "\\nabla_g" + }, + { + "bbox": [ + 313, + 72, + 554, + 119 + ], + "type": "text", + "content": ", average opacity " + }, + { + "bbox": [ + 313, + 72, + 554, + 119 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 72, + 554, + 119 + ], + "type": "text", + "content": " and max projected radius " + }, + { + "bbox": [ + 313, + 72, + 554, + 119 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 72, + 554, + 119 + ], + "type": "text", + "content": " of the included neural Gaussians for each " + }, + { + "bbox": [ + 313, + 72, + 554, + 119 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 72, + 554, + 119 + ], + "type": "text", + "content": " iterations. Those neural Gaussians satisfying:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 396, + 120, + 553, + 133 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 396, + 120, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 396, + 120, + 553, + 133 + ], + "type": "interline_equation", + "content": "\\nabla_ {g} \\cdot r \\cdot \\sigma^ {\\tau_ {\\sigma}} > \\tau_ {g}, \\tag {1}", + "image_path": "4ff9b177a54f09c1b8a64cb64efce779f3dc17b60e0c9167c244a291e9015c6e.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 137, + 553, + 160 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 137, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 313, + 137, + 553, + 160 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 137, + 553, + 160 + ], + "type": "inline_equation", + "content": "\\tau_{\\sigma}" + }, + { + "bbox": [ + 313, + 137, + 553, + 160 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 137, + 553, + 160 + ], + "type": "inline_equation", + "content": "\\tau_{g}" + }, + { + "bbox": [ + 313, + 137, + 553, + 160 + ], + "type": "text", + "content": " are pre-defined thresholds, are treated as significant and used as new anchors." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 165, + 555, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 165, + 555, + 320 + ], + "spans": [ + { + "bbox": [ + 313, + 165, + 555, + 320 + ], + "type": "text", + "content": "Balanced Camera Distribution. As previously mentioned, the imbalanced distribution of cameras can lead to 3D-GS overfitting to street view details during joint training, causing blurring and loss of structural detail in aerial views. Inspired by NeRF Director [34], we emphasize the importance of view selection and introduce a straightforward yet effective mechanism for balancing camera distribution. Specifically, we randomly select a value from the interval " + }, + { + "bbox": [ + 313, + 165, + 555, + 320 + ], + "type": "inline_equation", + "content": "[0,1]" + }, + { + "bbox": [ + 313, + 165, + 555, + 320 + ], + "type": "text", + "content": ". If the value is smaller than " + }, + { + "bbox": [ + 313, + 165, + 555, + 320 + ], + "type": "inline_equation", + "content": "\\frac{R}{R + 1}" + }, + { + "bbox": [ + 313, + 165, + 555, + 320 + ], + "type": "text", + "content": ", an aerial image is chosen for training; otherwise, a street image is picked. In detail, we set " + }, + { + "bbox": [ + 313, + 165, + 555, + 320 + ], + "type": "inline_equation", + "content": "R = 2" + }, + { + "bbox": [ + 313, + 165, + 555, + 320 + ], + "type": "text", + "content": " in the first stage to ensure thorough training of the aerial views and reduce it to 1 in the second stage." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 324, + 554, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 554, + 421 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 554, + 421 + ], + "type": "text", + "content": "Multi-resolution LOD Construction. Considering the differing levels of detail in aerial and street view images, and to enable real-time rendering, we adopt a Level-of-Detail (LOD) strategy, inspired by Octree-GS [23]. Specifically, for content with varying details between aerial and street views, we set different LOD levels: " + }, + { + "bbox": [ + 313, + 324, + 554, + 421 + ], + "type": "inline_equation", + "content": "K_{\\text{aerial}}" + }, + { + "bbox": [ + 313, + 324, + 554, + 421 + ], + "type": "text", + "content": " for aerial views and " + }, + { + "bbox": [ + 313, + 324, + 554, + 421 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 324, + 554, + 421 + ], + "type": "text", + "content": " for the whole views, corresponding to the two-phase training process." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 353, + 427, + 553, + 446 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 427, + 553, + 446 + ], + "spans": [ + { + "bbox": [ + 353, + 427, + 553, + 446 + ], + "type": "interline_equation", + "content": "K _ {\\text {a e r i a l}} = \\left\\lfloor \\log_ {2} \\left(D _ {\\text {a e r i a l}} / d _ {\\text {a e r i a l}}\\right) \\right\\rfloor + 1, \\tag {2}", + "image_path": "c67ba62245e02c1c7478a49f312c2c9389334d73db64992a4899bea0635acaf4.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 370, + 442, + 514, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 442, + 514, + 456 + ], + "spans": [ + { + "bbox": [ + 370, + 442, + 514, + 456 + ], + "type": "interline_equation", + "content": "K = K _ {\\text {a e r i a l}} + \\left\\lfloor \\log_ {2} \\left(d _ {\\text {a e r i a l}} / d _ {\\text {s t r e e t}}\\right) \\right\\rfloor .", + "image_path": "c36b5186b77d2330928d11d6362ff7bfbea7e636b20fbf98eaa8bf97ec3c9a86.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 462, + 554, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 554, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 554, + 570 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 462, + 554, + 570 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{aerial}}" + }, + { + "bbox": [ + 313, + 462, + 554, + 570 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 462, + 554, + 570 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{aerial}}" + }, + { + "bbox": [ + 313, + 462, + 554, + 570 + ], + "type": "text", + "content": " are the " + }, + { + "bbox": [ + 313, + 462, + 554, + 570 + ], + "type": "inline_equation", + "content": "r_d" + }, + { + "bbox": [ + 313, + 462, + 554, + 570 + ], + "type": "text", + "content": "-th largest and " + }, + { + "bbox": [ + 313, + 462, + 554, + 570 + ], + "type": "inline_equation", + "content": "r_d" + }, + { + "bbox": [ + 313, + 462, + 554, + 570 + ], + "type": "text", + "content": "-th smallest aerial distances, while " + }, + { + "bbox": [ + 313, + 462, + 554, + 570 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{street}}" + }, + { + "bbox": [ + 313, + 462, + 554, + 570 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 313, + 462, + 554, + 570 + ], + "type": "inline_equation", + "content": "r_d" + }, + { + "bbox": [ + 313, + 462, + 554, + 570 + ], + "type": "text", + "content": "-th smallest street distance. Note that, our model is set to " + }, + { + "bbox": [ + 313, + 462, + 554, + 570 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 462, + 554, + 570 + ], + "type": "text", + "content": " layers and " + }, + { + "bbox": [ + 313, + 462, + 554, + 570 + ], + "type": "inline_equation", + "content": "K_{\\mathrm{aerial}}" + }, + { + "bbox": [ + 313, + 462, + 554, + 570 + ], + "type": "text", + "content": " is only used in the first training stage. Specifically, in the first stage, we only activate " + }, + { + "bbox": [ + 313, + 462, + 554, + 570 + ], + "type": "inline_equation", + "content": "K_{\\mathrm{aerial}}" + }, + { + "bbox": [ + 313, + 462, + 554, + 570 + ], + "type": "text", + "content": " layers and add new anchors to the same LOD level as the significant source anchors. In the second stage, we open all layers and add the new anchors to the next LOD level to capture the higher-frequency information of street views." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 576, + 464, + 589 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 576, + 464, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 576, + 464, + 589 + ], + "type": "text", + "content": "3.4. Large-scale Scene Training" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 594, + 554, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 554, + 630 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 554, + 630 + ], + "type": "text", + "content": "In this section, we present a partitioning strategy to adapt our framework for large-scale urban scenes, supporting both model scalability and rendering fidelity." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": "Gaussian-based methods often suffer from blurring artifacts when reconstructing large-scale scenes, mainly due to challenges in densifying extensive image sets that require fine-tuned parameters. Additionally, limited GPU memory presents a hurdle when training on large-scale scenes. A common solution is the divide-and-conquer strategy, where a whole scene is split into smaller chunks." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26792" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 71, + 553, + 247 + ], + "blocks": [ + { + "bbox": [ + 60, + 71, + 553, + 247 + ], + "lines": [ + { + "bbox": [ + 60, + 71, + 553, + 247 + ], + "spans": [ + { + "bbox": [ + 60, + 71, + 553, + 247 + ], + "type": "image", + "image_path": "f27342e61e4c116f8653ed681e2cce2ae36885cb587c80539594c13e8ac884aa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 249, + 555, + 283 + ], + "lines": [ + { + "bbox": [ + 55, + 249, + 555, + 283 + ], + "spans": [ + { + "bbox": [ + 55, + 249, + 555, + 283 + ], + "type": "text", + "content": "Figure 4. Visualization of our constructed dataset. All the 7 scenes contain calibrated aerial and street view images. We illustrate the scenes with the point clouds and the corresponding image capture poses. The trajectory of aerial views is shown in purple, while street views are represented in yellow. Our dataset contains 5 synthetic scenes (a-e) and 2 real scenes (f-g)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 298, + 294, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 298, + 294, + 490 + ], + "spans": [ + { + "bbox": [ + 54, + 298, + 294, + 490 + ], + "type": "text", + "content": "For instance, VastGaussian [12] progressive partitioning for aerial views but faces projection errors with street views during visibility-based camera selection. To better accommodate the reconstruction of large-scale scenes involving both aerial and street views, we make the following adjustments. Following VastGaussian, we first partition the scene into " + }, + { + "bbox": [ + 54, + 298, + 294, + 490 + ], + "type": "inline_equation", + "content": "m \\times n" + }, + { + "bbox": [ + 54, + 298, + 294, + 490 + ], + "type": "text", + "content": " chunks based on ground-projected camera positions, then expand the original boundaries of each chunk by a threshold to ensure sufficient overlap. Next, we apply different strategies for aerial and street views: for aerial views, we augment each chunk with additional visible cameras and accessory points, for street views, which experience significant occlusion, we generate point clouds from depth maps to ensure comprehensive training coverage. All scenes follow the Manhattan world assumption, with the z-axis perpendicular to the ground plane." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 491, + 295, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 295, + 573 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 295, + 573 + ], + "type": "text", + "content": "Each chunk is trained independently using the strategy discussed in Section 3.3, then merged into a single model by discarding Gaussians outside the original boundaries and concatenating Gaussian attributes across chunks. To enhance rendering speed, we convert the hybrid representation to a fully explicit one by removing view inputs from each MLP and replacing the MLPs for color with SH predictions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 579, + 237, + 592 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 579, + 237, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 579, + 237, + 592 + ], + "type": "text", + "content": "3.5. Loss Function and Regularization" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 597, + 295, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 295, + 656 + ], + "type": "text", + "content": "Photometric Loss. We use different losses for rendering and surface reconstruction. For rendering, in addition to the standard L1, SSIM Loss, we also employ volume regularization " + }, + { + "bbox": [ + 55, + 597, + 295, + 656 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{vol}}" + }, + { + "bbox": [ + 55, + 597, + 295, + 656 + ], + "type": "text", + "content": " from Scaffold-GS [17] to ensure Gaussian primitives effectively cover the entire scene." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 661, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 661, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 661, + 296, + 714 + ], + "type": "text", + "content": "Geometry Regularizations. For geometry awareness, we introduce depth supervision " + }, + { + "bbox": [ + 55, + 661, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_d = \\frac{1}{hw}\\sum \\left|\\hat{D} - D\\right|" + }, + { + "bbox": [ + 55, + 661, + 296, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 661, + 296, + 714 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 55, + 661, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 661, + 296, + 714 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 661, + 296, + 714 + ], + "type": "text", + "content": " the rendering image dimensions, " + }, + { + "bbox": [ + 55, + 661, + 296, + 714 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 661, + 296, + 714 + ], + "type": "text", + "content": " is the inverse of the rendered depth map and " + }, + { + "bbox": [ + 55, + 661, + 296, + 714 + ], + "type": "inline_equation", + "content": "D_s" + }, + { + "bbox": [ + 55, + 661, + 296, + 714 + ], + "type": "text", + "content": " is the pseudo ground" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 298, + 555, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 298, + 555, + 358 + ], + "spans": [ + { + "bbox": [ + 313, + 298, + 555, + 358 + ], + "type": "text", + "content": "truth disparity map generated by either rendering engines (for synthetic data) or a scale-aligned pretrained monocular depth model [38] (for real data). For surface reconstruction, we also incorporate the normal loss " + }, + { + "bbox": [ + 313, + 298, + 555, + 358 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_n" + }, + { + "bbox": [ + 313, + 298, + 555, + 358 + ], + "type": "text", + "content": ", as used in 2D-GS [7], to enforce normal consistency." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 363, + 555, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 363, + 555, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 555, + 399 + ], + "type": "text", + "content": "Mask Regularization. To mitigate the effects of moving pedestrians, vehicles, and the sky, which is challenging for vanilla GS methods, we employ an opacity regularization" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 321, + 407, + 555, + 431 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 407, + 555, + 431 + ], + "spans": [ + { + "bbox": [ + 321, + 407, + 555, + 431 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {o} = - \\frac {1}{h w} \\sum \\sigma \\cdot \\log \\sigma - \\frac {1}{h w} \\sum M \\cdot \\log (1 - \\sigma), \\tag {3}", + "image_path": "e5338f54ce9929c185477e1702785b2fc51000579c0269a643fc744f7e278e5f.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 438, + 554, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 554, + 486 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 554, + 486 + ], + "type": "text", + "content": "which encourages the opacity values toward 0 or 1, where " + }, + { + "bbox": [ + 313, + 438, + 554, + 486 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 438, + 554, + 486 + ], + "type": "text", + "content": " denotes the mask region. Finally, the combined rendering and surface reconstruction losses, " + }, + { + "bbox": [ + 313, + 438, + 554, + 486 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_R" + }, + { + "bbox": [ + 313, + 438, + 554, + 486 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 438, + 554, + 486 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_S" + }, + { + "bbox": [ + 313, + 438, + 554, + 486 + ], + "type": "text", + "content": ", are defined as follows respectively:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 320, + 495, + 553, + 515 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 495, + 553, + 515 + ], + "spans": [ + { + "bbox": [ + 320, + 495, + 553, + 515 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {R} = \\mathcal {L} _ {1} + \\lambda_ {\\text {s s i m}} \\mathcal {L} _ {\\text {s s i m}} + \\lambda_ {\\text {v o l}} \\mathcal {L} _ {\\text {v o l}} + \\lambda_ {d} \\mathcal {L} _ {d} + \\lambda_ {o} \\mathcal {L} _ {o}, \\tag {4}", + "image_path": "4ea269f0538ba34a34eb43acbac0a15def2a982e805601a107c97ac031fbf00f.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 323, + 514, + 402, + 522 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 514, + 402, + 522 + ], + "spans": [ + { + "bbox": [ + 323, + 514, + 402, + 522 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {S} = \\mathcal {L} _ {R} + \\lambda_ {n} \\mathcal {L} _ {n}.", + "image_path": "1bec0d45ba80fdf82bc3e4a7bf0e64b1259c08eebd45ebd9ddfed93655b76552.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 533, + 419, + 545 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 419, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 419, + 545 + ], + "type": "text", + "content": "4. Dataset Curation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 552, + 555, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 555, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 555, + 636 + ], + "type": "text", + "content": "As previously mentioned, the lack of calibrated aerial and street datasets is a crucial issue that restricts the research on the unified scene reconstruction. To advance the field and evaluate our method, we construct a large-scale dataset with aerial and street views calibrated, which contains five synthetic scenes and two real scenes, as illustrated in Fig. 4. More details are listed in the supplementary materials." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 641, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 713 + ], + "type": "text", + "content": "Synthetic Data. Following MatrixCity [10] curation process, we create 5 additional synthetic scenes to simulate real-world data using custom data collection plugins. To ensure diversity, these synthetic datasets encompass various scales, seasons, and environments, ranging from realistic cityscapes to imaginative gaming scenarios. In addition" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26793" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 555, + 180 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 555, + 180 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 555, + 180 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 555, + 180 + ], + "type": "table", + "html": "
SceneBlock_SmallSyntheticReal
AerialStreetAerialStreetAerialStreet
MethodMetricsPSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
2D-GS [7]24.520.7430.36522.460.7630.36024.750.7780.37423.910.7950.29319.690.5040.58520.590.6030.422
Our-2D-GS29.600.8990.12123.600.8370.21630.390.9250.13625.440.8480.21622.570.6870.35721.520.6550.344
3D-GS [8]25.440.7810.32521.810.7440.37125.510.7980.35523.990.8090.27720.090.5270.56421.410.6270.398
Scaffold-GS [17]28.440.8630.19123.840.8190.27128.790.8910.19625.140.8330.24720.180.5390.54921.220.6260.394
Hier-GS [9]28.310.8610.18923.750.8230.22028.160.8660.24425.580.8610.19621.430.6390.43022.530.6860.303
Ours30.590.9130.09423.800.8390.20931.600.9380.10125.690.8620.19023.230.7290.32122.040.6690.324
", + "image_path": "cbad9d9ca94b44941a0c02d446785270c1eed1b32e84a5b9ea2d3db44b492d51.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 184, + 555, + 206 + ], + "lines": [ + { + "bbox": [ + 55, + 184, + 555, + 206 + ], + "spans": [ + { + "bbox": [ + 55, + 184, + 555, + 206 + ], + "type": "text", + "content": "Table 1. Quantitative comparison on small-scale datasets. Horizon-GS consistently achieves superior rendering quality compared to baselines in both aerial and street views." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 58, + 211, + 553, + 407 + ], + "blocks": [ + { + "bbox": [ + 58, + 211, + 553, + 407 + ], + "lines": [ + { + "bbox": [ + 58, + 211, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 58, + 211, + 553, + 407 + ], + "type": "image", + "image_path": "e66befcd1581aba12574b2908c1bbf67cd21706c16393d02dd7535bd7d3f1908.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 77, + 411, + 531, + 421 + ], + "lines": [ + { + "bbox": [ + 77, + 411, + 531, + 421 + ], + "spans": [ + { + "bbox": [ + 77, + 411, + 531, + 421 + ], + "type": "text", + "content": "Figure 5. Qualitative comparisons of Horizon-GS against baselines [7-9, 17] across (a) small-scale and (b) large-scale scenes." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 436, + 295, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 436, + 295, + 459 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 295, + 459 + ], + "type": "text", + "content": "to RGB images, we render depth maps for geometry supervision and initial point clouds for 3D-GS training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 460, + 295, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 460, + 295, + 627 + ], + "spans": [ + { + "bbox": [ + 54, + 460, + 295, + 627 + ], + "type": "text", + "content": "Real Data. To evaluate Horizon-GS in real-world data, we collect 2 additional scenes. We use a DJI drone equipped with five cameras to capture aerial images and a custom-designed helmet equipped with six Action4 cameras to capture street data. To enhance the performance of the structure-from-motion (SfM) technique, we include transitional views to increase image overlap and utilize a powerful commercial software, ContextCapture1. Depth maps are generated using Depth-Anything-V2 [38], and scale/offset are estimated by aligning inverse depth of SfM points. Moving objects, such as humans and cars, are removed by Grounded-SAM [24], while sky masking is handled by the pretrained SkyRemoval model [21]. More details about the dataset are provided in the supplementary materials." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 636, + 137, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 636, + 137, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 636, + 137, + 649 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 655, + 172, + 668 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 655, + 172, + 668 + ], + "spans": [ + { + "bbox": [ + 55, + 655, + 172, + 668 + ], + "type": "text", + "content": "5.1. Experimental Setup" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 673, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 673, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 673, + 295, + 696 + ], + "type": "text", + "content": "Datasets and Metrics. We conduct comprehensive evaluations across 11 scenes containing both aerial and street" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 436, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 436, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 436, + 555, + 628 + ], + "type": "text", + "content": "views, sourced from the MatrixCity dataset [10], the UC-GS dataset [43], and our captured dataset. Specifically, in the MatrixCity dataset, we evaluate two synthetic scenes: Block_Small and Block_A. For the UC-GS dataset, we select two synthetic scenes, New York City (NYC) and San Francisco (SF). Consistent with the original UC-GS settings, evaluations are performed solely on street views. For our newly captured dataset, we select one out of every 32 images for evaluation. It is worth noting that previous works only evaluate scenes from a single perspective. To improve GPU compatibility and achieve satisfactory visual results, we divide Block_A into " + }, + { + "bbox": [ + 313, + 436, + 555, + 628 + ], + "type": "inline_equation", + "content": "4 \\times 2" + }, + { + "bbox": [ + 313, + 436, + 555, + 628 + ], + "type": "text", + "content": " chunks and divide Hillside_Morning and Hillside_Summer into " + }, + { + "bbox": [ + 313, + 436, + 555, + 628 + ], + "type": "inline_equation", + "content": "6 \\times 2" + }, + { + "bbox": [ + 313, + 436, + 555, + 628 + ], + "type": "text", + "content": " chunks. We use image resolution " + }, + { + "bbox": [ + 313, + 436, + 555, + 628 + ], + "type": "inline_equation", + "content": "1920 \\times 1080" + }, + { + "bbox": [ + 313, + 436, + 555, + 628 + ], + "type": "text", + "content": " for synthetic data; " + }, + { + "bbox": [ + 313, + 436, + 555, + 628 + ], + "type": "inline_equation", + "content": "1600 \\times 1066" + }, + { + "bbox": [ + 313, + 436, + 555, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 436, + 555, + 628 + ], + "type": "inline_equation", + "content": "1600 \\times 900" + }, + { + "bbox": [ + 313, + 436, + 555, + 628 + ], + "type": "text", + "content": " for real-world aerial and street view images." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 628, + 555, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 628, + 555, + 688 + ], + "spans": [ + { + "bbox": [ + 313, + 628, + 555, + 688 + ], + "type": "text", + "content": "Results are evaluated using standard visual quality metrics: PSNR, SSIM [30], and LPIPS [42]. Aerial and street view images are evaluated separately. To minimize the impact of pedestrians, cars, and the sky, we mask out those moving entities out when comparing with ground truths." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "Baselines. We compare our method against 3D-GS [8], Scaffold-GS [17], UC-GS [43] and Hierarchical-3DGS [9]" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 66, + 702, + 196, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 196, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 196, + 713 + ], + "type": "text", + "content": "1 https://www.daspatial.com/cn/gcluster" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26794" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 555, + 210 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 555, + 210 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 555, + 210 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 555, + 210 + ], + "type": "table", + "html": "
SceneBlock_AHillside_MorningHillside_Summer
AerialStreetAerialStreetAerialStreet
MethodMetricsPSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
2D-GS [7]20.630.5850.59519.570.6350.47722.020.6290.51319.200.6370.47020.870.5600.52418.700.6140.481
Our-2D-GS*28.520.8860.16623.000.7890.30129.980.9000.13520.440.6860.40428.880.8890.12620.230.6740.404
3D-GS [8]21.020.5900.59519.060.6280.47722.870.6620.48919.350.6430.46621.500.5920.50318.630.6170.480
Scaffold-GS [17]22.980.6540.52320.890.6810.43623.050.6880.43320.080.6560.46122.020.6340.43619.360.6290.473
Scaffold-GS*27.620.8600.20623.100.8080.27725.460.7960.27120.660.6850.41623.910.7520.27920.300.6670.423
Vast-GS [12]25.850.8070.29521.460.7380.35324.890.7650.29520.220.6770.42423.200.7400.29219.820.6570.429
City-GS [16]26.860.8350.24722.230.7760.30625.560.7880.26720.390.6830.42023.350.7420.28419.980.6600.427
Hier-GS* [9]26.420.7920.34023.590.8110.27525.410.7830.26921.130.6920.39823.530.7350.27320.470.6730.402
Ours*28.890.8900.15123.660.8160.25531.090.9180.10720.620.6890.39929.340.9000.11020.340.6810.393
", + "image_path": "0e3ffbdc799d200352b43f8f635954de6b74e5818ca3708337b32e50c223ab7b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 239, + 294, + 319 + ], + "blocks": [ + { + "bbox": [ + 58, + 239, + 294, + 319 + ], + "lines": [ + { + "bbox": [ + 58, + 239, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 58, + 239, + 294, + 319 + ], + "type": "image", + "image_path": "24f2b2ff49177b3d11a435aa42b1c1dfae1ca392f5b21a2460dbcbebd368ca3f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 328, + 295, + 361 + ], + "lines": [ + { + "bbox": [ + 55, + 328, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 328, + 295, + 361 + ], + "type": "text", + "content": "Figure 6. Qualitative comparisons of Horizon-GS against UC-GS [43]. The first row shows view shifting in the SF scene, while the second row shows view shifting and rotation in the NYC scene." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 376, + 295, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 376, + 295, + 520 + ], + "spans": [ + { + "bbox": [ + 54, + 376, + 295, + 520 + ], + "type": "text", + "content": "for rendering, as well as 2D-GS [7] for surface reconstruction. To ensure full convergence, we extend the total training iterations from " + }, + { + "bbox": [ + 54, + 376, + 295, + 520 + ], + "type": "inline_equation", + "content": "30\\mathrm{k}" + }, + { + "bbox": [ + 54, + 376, + 295, + 520 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 54, + 376, + 295, + 520 + ], + "type": "inline_equation", + "content": "100\\mathrm{k}" + }, + { + "bbox": [ + 54, + 376, + 295, + 520 + ], + "type": "text", + "content": ", with densification continuing until the " + }, + { + "bbox": [ + 54, + 376, + 295, + 520 + ], + "type": "inline_equation", + "content": "50\\mathrm{k}" + }, + { + "bbox": [ + 54, + 376, + 295, + 520 + ], + "type": "text", + "content": " iteration. For all baselines, we run their official code on all datasets except for the UC-GS dataset, where we adopt the metrics from their original paper to maintain consistency with their settings. For Hierarchical-3DGS, we report the best visual quality results after optimizing the hierarchy (leaves). For each baseline, we also add depth supervision same as our method. In the large-scale setting, we partition the scenes and train each chunk in parallel, using " + }, + { + "bbox": [ + 54, + 376, + 295, + 520 + ], + "type": "inline_equation", + "content": "*" + }, + { + "bbox": [ + 54, + 376, + 295, + 520 + ], + "type": "text", + "content": " to denote the strategy discussed in Sec. 3.4." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "text", + "content": "Implementation Details. For a fair comparison, we train the first stage for 60k iterations and the second stage for 40k iterations, with densification stopping at 30k and 20k iterations, respectively. In the first stage, we follow the default settings of Scaffold-GS [17] for learning rate and densification. In the second stage, we reduce the learning rate of the offset by a factor of 10 and set " + }, + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\tau_{\\sigma}" + }, + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\tau_{g}" + }, + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "text", + "content": " to 0.2, 0.15 and 100, respectively. In LOD construction stage, " + }, + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "inline_equation", + "content": "r_d" + }, + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "text", + "content": " is set to 0.999. The loss function parameters are " + }, + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{ssim}} = 0.2" + }, + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{vol}} = 0.01" + }, + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{o} = 0.05" + }, + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{n} = 0.05" + }, + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "text", + "content": ". The weight for " + }, + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_d" + }, + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "text", + "content": " is exponentially decayed from 1 to 0.01 over both stages. Depth supervision is activated after 500 iterations, and normal supervision after 7k iterations. We denote neural Gaussians that MLP decode into 2D as 'Our-2D-GS' and denote 'Ours' if the primitives are generated into 3D. All experi" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 316, + 238, + 554, + 302 + ], + "blocks": [ + { + "bbox": [ + 70, + 213, + 537, + 225 + ], + "lines": [ + { + "bbox": [ + 70, + 213, + 537, + 225 + ], + "spans": [ + { + "bbox": [ + 70, + 213, + 537, + 225 + ], + "type": "text", + "content": "Table 2. Quantitative comparison on large-scale datasets. Horizon-GS achieves superior rendering quality compared to baselines." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 238, + 554, + 302 + ], + "lines": [ + { + "bbox": [ + 316, + 238, + 554, + 302 + ], + "spans": [ + { + "bbox": [ + 316, + 238, + 554, + 302 + ], + "type": "table", + "html": "
Scene Method MetricsHeld-outView(+1m)View(+1m 5°down)
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
3D-GS [8]23.470.6680.40620.830.6050.44021.250.6430.402
Scaffold-GS [17]25.400.7440.32022.620.6710.37523.280.7110.334
UC-GS [17]25.950.7630.29123.520.7020.34024.150.7410.298
Ours25.350.7570.28025.460.7600.28025.370.7610.278
", + "image_path": "cf18522c60de7cb3cafc59a25ea57c1518db573364c552a3e4c9c3b4caf5ebba.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 308, + 555, + 340 + ], + "lines": [ + { + "bbox": [ + 313, + 308, + 555, + 340 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 555, + 340 + ], + "type": "text", + "content": "Table 3. Quantitative comparison on UC-GS dataset [43]. The metrics are extracted from the original paper, in which all methods are trained for 900k iterations." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 347, + 523, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 347, + 523, + 358 + ], + "spans": [ + { + "bbox": [ + 313, + 347, + 523, + 358 + ], + "type": "text", + "content": "ments are conducted on NVIDIA A100 80G GPUs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 369, + 413, + 381 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 413, + 381 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 413, + 381 + ], + "type": "text", + "content": "5.2. Results Analysis" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 387, + 555, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 387, + 555, + 447 + ], + "spans": [ + { + "bbox": [ + 313, + 387, + 555, + 447 + ], + "type": "text", + "content": "Our evaluation covers diverse scenes, including aerial and street views, intricate and large-scale scenes, both synthetic and real-world. We demonstrate that Horizon-GS preserves fine details and significantly improves the quality of surface reconstruction, as shown in Tab. 1, 2, 3, and Fig. 5, 6, 7." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "type": "text", + "content": "**Rendering Performance Analysis.** Our method reconstructs a unified Gaussian model, delivering excellent visual quality in both aerial and street views. Compared to 2D-GS [7], Our-2D-GS consistently outperforms by 4 dB in aerial PSNR and 1.5 dB in street PSNR. Furthermore, compared to all the baselines, Horizon-GS exceeds the performance of all baselines, except Hierarchical-3DGS, which is competitive in street views due to its street-focused design. As shown in the highlighted patches and arrows above of Fig. 5 (a), our method excels in fine details (1st row) and reflective regions (2nd row). In the UC-GS setting (Tab. 3), under the held-out viewpoints condition where test views share the same heights as training views, Horizon-GS, trained for 100k iterations, shows slightly lower PSNR and SSIM values compared to UC-GS trained for 900k iterations. However, our approach produces richer details, as indicated by better LPIPS metrics. Moreover, when evaluating shifted and rotated views, our method consistently outperforms the baselines, as shown in Fig. 6. Compared with synthetic data, the quality of real data declined due to performance differences caused by challenges such as pose estimation, sensor noise, and texture complexity." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26795" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 552, + 190 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 552, + 190 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 552, + 190 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 552, + 190 + ], + "type": "image", + "image_path": "1bdec2e4a9490e24a44b1ca1a890fd06131b76bfbcf8495ce8ba2ca75236a88e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 191, + 555, + 214 + ], + "lines": [ + { + "bbox": [ + 55, + 191, + 555, + 214 + ], + "spans": [ + { + "bbox": [ + 55, + 191, + 555, + 214 + ], + "type": "text", + "content": "Figure 7. Qualitative comparisons of our method with 2D-GS [7]. We visualize the rendered images and geometry results (depth and world-space normal) in the Block_Small scene from both aerial and street views." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 220, + 297, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 220, + 297, + 411 + ], + "spans": [ + { + "bbox": [ + 54, + 220, + 297, + 411 + ], + "type": "text", + "content": "Large-scale Rendering Performance. For the challenging large-scale scenes, we conduct two types of experiments: direct training without chunking and chunked training. As shown in Table 2, partitioning and then merging proves effective, especially when comparing Scaffold-GS and Scaffold-GS*. Horizon-GS consistently delivers excellent rendering quality in large-scale scenes compared to other baselines. In contrast, Hierarchical-3DGS [9] tends to overfit the training views, producing artifacts in novel views, while other baselines fail to capture fine-grained details (Figure 5 (b)). Each scene is divided into chunks and trained in parallel, with a total training time of about four hours per scene. Due to the LOD design, our method achieves real-time rendering speeds of 51.5 FPS in the Hillside_Morning and Hillside_Summer scenes, which are the largest scenes among our captured dataset." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 415, + 296, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 296, + 499 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 296, + 499 + ], + "type": "text", + "content": "Surface Reconstruction Performance. As shown in Fig.7, our method reconstructs more complete and detailed geometry compared to 2D-GS [7]. 2D-GS produces artifacts, resulting in incomplete and lackluster geometry. In contrast, our method, using the two-stage training approach and enhanced densification, delivers detailed, geometrically accurate, and artifact-free reconstruction." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 507, + 157, + 519 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 507, + 157, + 519 + ], + "spans": [ + { + "bbox": [ + 55, + 507, + 157, + 519 + ], + "type": "text", + "content": "5.3. Ablation Studies" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 525, + 295, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 295, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 295, + 571 + ], + "type": "text", + "content": "In this section, we ablate each individual module to validate their effectiveness. We use the scenes from our proposed real dataset, Road and Park. Quantitative results can be found in Tab. 4." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 577, + 295, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 577, + 295, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 577, + 295, + 673 + ], + "type": "text", + "content": "Balanced Camera Distribution. To assess the effectiveness of the balanced camera distribution (Sec. 3.3), we conduct an ablation study with randomly selected training views. The results in Tab. 4 reveal a notable decline in the visual quality of aerial views, while the quality of street views remains similar. This suggests that a well-distributed set of observation perspectives enhances the model performance, rather than converging to a single-level view." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "Balanced Densification. As outlined in Sec. 3.3, we meticulously design the densification strategy in the two training stages. Specifically, gradients of the Gaussian" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 317, + 217, + 554, + 297 + ], + "blocks": [ + { + "bbox": [ + 317, + 217, + 554, + 297 + ], + "lines": [ + { + "bbox": [ + 317, + 217, + 554, + 297 + ], + "spans": [ + { + "bbox": [ + 317, + 217, + 554, + 297 + ], + "type": "table", + "html": "
Scene\nMethod | MetricsAerialStreet
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Ours w/o camera bal.22.490.6840.37522.000.6710.323
Ours w/o densify bal.23.020.7210.33221.980.6700.325
Ours w/o multi LOD23.050.7170.33722.030.6720.321
Ours w/o densify poli.22.830.6900.36921.680.6550.346
Ours23.230.7290.32122.040.6690.324
", + "image_path": "316f49a264f15e21253c0d86dcbc188590fb882876196ea9bb83b2d6c536b39d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 318, + 301, + 550, + 313 + ], + "lines": [ + { + "bbox": [ + 318, + 301, + 550, + 313 + ], + "spans": [ + { + "bbox": [ + 318, + 301, + 550, + 313 + ], + "type": "text", + "content": "Table 4. Ablations of our method on our proposed real datasets." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 319, + 555, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 555, + 404 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 555, + 404 + ], + "type": "text", + "content": "primitives are accumulated from aerial views, while in the second stage, they are derived solely from street views. To assess the effectiveness of this approach, we compute the gradients of the Gaussian primitives across all the images in both stages. The results reveal conflicts during the densification process, highlighting the need to explicitly separate the densification steps during optimization." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 411, + 555, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 555, + 459 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 555, + 459 + ], + "type": "text", + "content": "Multi-Resolution LOD. To assess the impact of LOD design, we conduct an experiment without LOD. The quality of aerial views significantly decreases, indicating that the strategy enhances the capture of fine-scale details." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 465, + 554, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 465, + 554, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 465, + 554, + 525 + ], + "type": "text", + "content": "Densification Policy. An ablation of the densification policy (Eq. 1) replacing it with Scaffold-GS in the second stage shows a significant reduction in rendering details for both cross views, highlighting that average gradients alone are insufficient for capturing fine-grained local details." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 537, + 388, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 537, + 388, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 537, + 388, + 550 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": "We introduce Horizon-GS, a novel framework for large-scale aerial-to-ground scene reconstruction. We first explore the challenges of unified scene reconstruction from both aerial and street views, and propose a end-to-end coarse-to-fine training framework that mitigates inherent conflicts and is adaptable for large-scale reconstruction. Additionally, we present a comprehensive dataset with two cross views. Our experiments show that Horizon-GS outperforms existing methods in visual quality and geometry accuracy. In the future, we aim to minimize input dependency by leveraging advanced techniques such as pretrained 3D scene foundation models and enable more systematic solutions, such as distributional training." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26796" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 173 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 173 + ], + "type": "text", + "content": "Acknowledgments This work was supported in part by the National Key R&D Program of China (2022ZD0160201), Shanghai Artificial Intelligence Laboratory, the HKU Startup Fund, the HKU Shanghai Intelligent Computing Research Center, the center for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK, and the Anhui Provincial Natural Science Foundation under Grant 2108085UD12." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 193, + 115, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 193, + 115, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 193, + 115, + 205 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 213, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 213, + 296, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 213, + 296, + 267 + ], + "spans": [ + { + "bbox": [ + 61, + 213, + 296, + 267 + ], + "type": "text", + "content": "[1] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXII, pages 333-350. Springer, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 269, + 296, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 269, + 296, + 301 + ], + "spans": [ + { + "bbox": [ + 62, + 269, + 296, + 301 + ], + "type": "text", + "content": "[2] Yu Chen and Gim Hee Lee. Dogaussian: Distributed-oriented gaussian splattering for large-scale 3d reconstruction via gaussian consensus. CoRR, abs/2405.13943, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 302, + 296, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 302, + 296, + 345 + ], + "spans": [ + { + "bbox": [ + 62, + 302, + 296, + 345 + ], + "type": "text", + "content": "[3] Yurui Chen, Chun Gu, Junzhe Jiang, Xiatian Zhu, and Li Zhang. Periodic vibration gaussian: Dynamic urban scene reconstruction and real-time rendering. CoRR, abs/2311.18561, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 346, + 296, + 400 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 346, + 296, + 400 + ], + "spans": [ + { + "bbox": [ + 62, + 346, + 296, + 400 + ], + "type": "text", + "content": "[4] David J. Crandall, Andrew Owens, Noah Snavely, and Daniel P. Huttenlocher. Sfm with mrfs: Discrete-continuous optimization for large-scale structure from motion. IEEE Trans. Pattern Anal. Mach. Intell., 35(12):2841-2853, 2013. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 403, + 296, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 403, + 296, + 468 + ], + "spans": [ + { + "bbox": [ + 62, + 403, + 296, + 468 + ], + "type": "text", + "content": "[5] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 5491-5500. IEEE, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 469, + 296, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 469, + 296, + 534 + ], + "spans": [ + { + "bbox": [ + 62, + 469, + 296, + 534 + ], + "type": "text", + "content": "[6] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 5354-5363. IEEE, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 536, + 296, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 536, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 62, + 536, + 296, + 590 + ], + "type": "text", + "content": "[7] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, SIGGRAPH 2024, Denver, CO, USA, 27 July 2024-1 August 2024, page 32. ACM, 2024. 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 591, + 296, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 591, + 296, + 635 + ], + "spans": [ + { + "bbox": [ + 62, + 591, + 296, + 635 + ], + "type": "text", + "content": "[8] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139:1-139:14, 2023. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 636, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 636, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 62, + 636, + 296, + 689 + ], + "type": "text", + "content": "[9] Bernhard Kerbl, Andreas Meuleman, Georgios Kopanas, Michael Wimmer, Alexandre Lanvin, and George Drettakis. A hierarchical 3d gaussian representation for real-time rendering of very large datasets. ACM Trans. Graph., 43(4): 62:1-62:15, 2024. 2, 4, 6, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "type": "text", + "content": "[10] Yixuan Li, Lihan Jiang, Linning Xu, Yuanbo Xiangli, Zhenzhi Wang, Dahua Lin, and Bo Dai. Matrixcity: A large-" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 116 + ], + "type": "text", + "content": "scale city dataset for city-scale neural rendering and beyond. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 3182-3192. IEEE, 2023. 2, 5, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 118, + 555, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 555, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 555, + 162 + ], + "type": "text", + "content": "[11] Yiyi Liao, Jun Xie, and Andreas Geiger. KITTI-360: A novel dataset and benchmarks for urban scene understanding in 2d and 3d. IEEE Trans. Pattern Anal. Mach. Intell., 45(3):3292–3310, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 164, + 554, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 164, + 554, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 164, + 554, + 239 + ], + "type": "text", + "content": "[12] Jiaqi Lin, Zhihao Li, Xiao Tang, Jianzhuang Liu, Shiyong Liu, Jiayue Liu, Yangdi Lu, Xiaofei Wu, Songcen Xu, Youliang Yan, and Wenming Yang. Vastgaussian: Vast 3d gaussians for large scene reconstruction. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 5166-5175. IEEE, 2024. 2, 3, 5, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 242, + 554, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 242, + 554, + 306 + ], + "spans": [ + { + "bbox": [ + 316, + 242, + 554, + 306 + ], + "type": "text", + "content": "[13] Liqiang Lin, Yilin Liu, Yue Hu, Xingguang Yan, Ke Xie, and Hui Huang. Capturing, reconstructing, and simulating: The urbanscene3d dataset. In Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part VIII, pages 93-109. Springer, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 309, + 554, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 554, + 373 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 554, + 373 + ], + "type": "text", + "content": "[14] Jeffrey Yunfan Liu, Yun Chen, Ze Yang, Jingkang Wang, Sivabalan Manivasagam, and Raquel Urtasun. Real-time neural rasterization for large scenes. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 8382-8393. IEEE, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 376, + 554, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 554, + 431 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 554, + 431 + ], + "type": "text", + "content": "[15] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 434, + 554, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 554, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 554, + 498 + ], + "type": "text", + "content": "[16] Yang Liu, Chuanchen Luo, Lue Fan, Naiyan Wang, Junran Peng, and Zhaoxiang Zhang. Citygaussian: Real-time high-quality large-scale scene rendering with gaussians. In Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part XVI, pages 265-282. Springer, 2024. 2, 3, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 500, + 554, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 500, + 554, + 565 + ], + "spans": [ + { + "bbox": [ + 316, + 500, + 554, + 565 + ], + "type": "text", + "content": "[17] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 20654-20664. IEEE, 2024. 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 567, + 554, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 567, + 554, + 622 + ], + "spans": [ + { + "bbox": [ + 316, + 567, + 554, + 622 + ], + "type": "text", + "content": "[18] Zhenxing Mi and Dan Xu. Switch-nerf: Learning scene decomposition with mixture of experts for large-scale neural radiance fields. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 624, + 554, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 554, + 689 + ], + "type": "text", + "content": "[19] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I, pages 405-421. Springer, 2020. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 691, + 554, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 554, + 714 + ], + "type": "text", + "content": "[20] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multires" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26797" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 77, + 72, + 295, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 295, + 95 + ], + "type": "text", + "content": "olution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 96, + 294, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 96, + 294, + 118 + ], + "spans": [ + { + "bbox": [ + 57, + 96, + 294, + 118 + ], + "type": "text", + "content": "[21] OpenDroneMap. Skyremoval. https://github.com/OpenDroneMap/SkyRemoval/, 2022.6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 120, + 295, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 120, + 295, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 120, + 295, + 185 + ], + "type": "text", + "content": "[22] Konstantinos Rematas, Andrew Liu, Pratul P. Srinivasan, Jonathan T. Barron, Andrea Tagliasacchi, Thomas A. Funkhouser, and Vittorio Ferrari. Urban radiance fields. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 12922-12932. IEEE, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 187, + 295, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 187, + 295, + 230 + ], + "spans": [ + { + "bbox": [ + 57, + 187, + 295, + 230 + ], + "type": "text", + "content": "[23] Kerui Ren, Lihan Jiang, Tao Lu, Mulin Yu, Linning Xu, Zhangkai Ni, and Bo Dai. Octree-gs: Towards consistent real-time rendering with lod-structured 3d gaussians. CoRR, abs/2403.17898, 2024. 2, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 232, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 232, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 232, + 294, + 297 + ], + "type": "text", + "content": "[24] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, Zhaoyang Zeng, Hao Zhang, Feng Li, Jie Yang, Hongyang Li, Qing Jiang, and Lei Zhang. Grounded SAM: assembling open-world models for diverse visual tasks. CoRR, abs/2401.14159, 2024. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 300, + 294, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 300, + 294, + 354 + ], + "spans": [ + { + "bbox": [ + 56, + 300, + 294, + 354 + ], + "type": "text", + "content": "[25] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 5449-5459. IEEE, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 355, + 294, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 355, + 294, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 355, + 294, + 431 + ], + "type": "text", + "content": "[26] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben P. Mildenhall, Pratul P. Srinivasan, Jonathan T. Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 8238-8248. IEEE, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 434, + 294, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 294, + 499 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 294, + 499 + ], + "type": "text", + "content": "[27] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. LGM: large multi-view gaussian model for high-resolution 3d content creation. In Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part IV, pages 1-18. Springer, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 501, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 294, + 555 + ], + "type": "text", + "content": "[28] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 557, + 294, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 294, + 621 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 294, + 621 + ], + "type": "text", + "content": "[29] Haithem Turki, Deva Ramanan, and Mahadev Satyanarayanan. Mega-nerf: Scalable construction of large-scale nerfs for virtual fly- throughs. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 12912-12921. IEEE, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 624, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 667 + ], + "type": "text", + "content": "[30] Zhou Wang, Alan C. Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Trans. Image Process., 13(4): 600-612, 2004. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 669, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 294, + 713 + ], + "type": "text", + "content": "[31] Yuxi Wei, Zi Wang, Yifan Lu, Chenxin Xu, Changxing Liu, Hao Zhao, Siheng Chen, and Yanfeng Wang. Editable scene simulation for autonomous driving via collaborative llm-agents. In IEEE/CVF Conference on Computer Vision and" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 15077-15087. IEEE, 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 97, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 553, + 173 + ], + "type": "text", + "content": "[32] Yuanbo Xiangli, Linning Xu, Xingang Pan, Nanxuan Zhao, Anyi Rao, Christian Theobalt, Bo Dai, and Dahua Lin. Bungeenerf: Progressive neural radiance field for extreme multi-scale scene rendering. In Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXII, pages 106-122. Springer, 2022. 2, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 175, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 239 + ], + "type": "text", + "content": "[33] Yuanbo Xiangli, Linning Xu, Xingang Pan, Nanxuan Zhao, Bo Dai, and Dahua Lin. Assetfield: Assets mining and reconfiguration in ground feature plane representation. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 3228-3238. IEEE, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 243, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 553, + 308 + ], + "type": "text", + "content": "[34] Wenhui Xiao, Rodrigo Santa Cruz, David Ahmed-Aristizabal, Olivier Salvado, Clinton Fookes, and Léo Lebrat. Nerf director: Revisiting view selection in neural volume rendering. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 20742-20751. IEEE, 2024. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 309, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 374 + ], + "type": "text", + "content": "[35] Linning Xu, Yuanbo Xiangli, Sida Peng, Xingang Pan, Nanxuan Zhao, Christian Theobalt, Bo Dai, and Dahua Lin. Grid-guided neural radiance fields for large urban scenes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 8296-8306. IEEE, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 376, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 453 + ], + "type": "text", + "content": "[36] Yinghao Xu, Zifan Shi, Yifan Wang, Hansheng Chen, Ceyuan Yang, Sida Peng, Yujun Shen, and Gordon Wetzstein. GRM: large gaussian reconstruction model for efficient 3d reconstruction and generation. In Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part XV, pages 1-20. Springer, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 455, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 455, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 455, + 553, + 498 + ], + "type": "text", + "content": "[37] Yunzhi Yan, Haotong Lin, Chenxu Zhou, Weijie Wang, Haiyang Sun, Kun Zhan, Xianpeng Lang, Xiaowei Zhou, and Sida Peng. Street gaussians for modeling dynamic urban scenes. CoRR, abs/2401.01339, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 500, + 553, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 500, + 553, + 565 + ], + "spans": [ + { + "bbox": [ + 316, + 500, + 553, + 565 + ], + "type": "text", + "content": "[38] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything V2. In Advances in Neural Information Processing Systems 38: Annual Conference on Neural Information Processing Systems 2024, NeurIPS 2024, Vancouver, BC, Canada, December 10 - 15, 2024, 2024. 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 567, + 553, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 567, + 553, + 632 + ], + "spans": [ + { + "bbox": [ + 316, + 567, + 553, + 632 + ], + "type": "text", + "content": "[39] Mulin Yu, Tao Lu, Linning Xu, Lihan Jiang, Yuanbo Xiangli, and Bo Dai. GSDF: 3dgs meets SDF for improved neural rendering and reconstruction. In Advances in Neural Information Processing Systems 38: Annual Conference on Neural Information Processing Systems 2024, NeurIPS 2024, Vancouver, BC, Canada, December 10 - 15, 2024, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 635, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 667 + ], + "type": "text", + "content": "[40] Zehao Yu, Torsten Sattler, and Andreas Geiger. Gaussian opacity fields: Efficient adaptive surface reconstruction in unbounded scenes. ACM Transactions on Graphics, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "type": "text", + "content": "[41] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. GS-LRM: large reconstruction model for 3d gaussian splatting. In Computer Vision - ECCV 2024 - 18th European Conference, Mi" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26798" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 293 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 75, + 72, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 72, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 296, + 95 + ], + "type": "text", + "content": "lan, Italy, September 29-October 4, 2024, Proceedings, Part XXII, pages 1-19. Springer, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 171 + ], + "type": "text", + "content": "[42] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 586-595. Computer Vision Foundation / IEEE Computer Society, 2018. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 173, + 295, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 173, + 295, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 173, + 295, + 217 + ], + "type": "text", + "content": "[43] Saining Zhang, Baijun Ye, Xiaoxue Chen, Yuantao Chen, Zongzheng Zhang, Cheng Peng, Yongliang Shi, and Hao Zhao. Drone-assisted road gaussian splatting with cross-view uncertainty. CoRR, abs/2408.15242, 2024. 2, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 295, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 295, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 295, + 293 + ], + "type": "text", + "content": "[44] Xiaoyu Zhou, Zhiwei Lin, Xiaojun Shan, Yongtao Wang, Deqing Sun, and Ming-Hsuan Yang. Drivinggaussian: Composite gaussian splatting for surrounding dynamic autonomous driving scenes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 21634-21643. IEEE, 2024. 2" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "26799" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/HotSpot_ Signed Distance Function Optimization with an Asymptotically Sufficient Condition/733f46a7-f78c-4e53-bba4-ead7eb705bb7_content_list.json b/2025/HotSpot_ Signed Distance Function Optimization with an Asymptotically Sufficient Condition/733f46a7-f78c-4e53-bba4-ead7eb705bb7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3b2b7c865fe5dca004b0d5866ab55bd4e4c70e48 --- /dev/null +++ b/2025/HotSpot_ Signed Distance Function Optimization with an Asymptotically Sufficient Condition/733f46a7-f78c-4e53-bba4-ead7eb705bb7_content_list.json @@ -0,0 +1,2136 @@ +[ + { + "type": "text", + "text": "HotSpot: Signed Distance Function Optimization with an Asymptotically Sufficient Condition", + "text_level": 1, + "bbox": [ + 117, + 128, + 852, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zimo Wang* Cheng Wang* Taiki Yoshino Sirui Tao Ziyang Fu Tzu-Mao Li UC San Diego", + "bbox": [ + 143, + 202, + 823, + 239 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/74e2cc9cebace942b87899a27c1858b46acc7b83ef83fb6f154ef3b51db6ab2f.jpg", + "image_caption": [ + "Figure 1. We propose HOTSPOT, a neural signed distance function optimization method that establishes an asymptotic sufficient condition to guarantee convergence to a true distance function, enabling precise surface reconstruction and level set representation for complex shapes. Here we show a reconstruction from a point cloud sampled from the reference bunny (taken from Mehta et al. [1]) on the right. In the inset, we visualize the recovered signed distance function on a horizontal slice, using warm colors for positive values and cool for negative (zoom in for details). Our reconstruction is significantly more accurate than prior works (SAL [2], DiGS [3], and StEik [4])." + ], + "image_footnote": [], + "bbox": [ + 83, + 255, + 888, + 393 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 496, + 313, + 512 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We propose a method, HOTSPOT, for optimizing neural signed distance functions. Existing losses, such as the eikonal loss, act as necessary but insufficient constraints and cannot guarantee that the recovered implicit function represents a true distance function, even if the output minimizes these losses almost everywhere. Furthermore, the eikonal loss suffers from stability issues in optimization. Finally, in conventional methods, regularization losses that penalize surface area distort the reconstructed signed distance function. We address these challenges by designing a loss function using the solution of a screened Poisson equation. Our loss, when minimized, provides an asymptotically sufficient condition to ensure the output converges to a true distance function. Our loss also leads to stable optimization and naturally penalizes large surface areas. We present theoretical analysis and experiments on both challenging 2D and 3D datasets and show that our method provides better surface reconstruction and a more accurate distance approximation.", + "bbox": [ + 75, + 527, + 472, + 815 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 820, + 209, + 835 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Learning neural signed distance function requires designing loss functions so that the zero crossings of the implicit", + "bbox": [ + 75, + 845, + 470, + 876 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "function are at the desired locations (e.g., adhere to a sparse input point cloud), and that the implicit function returns the correct signed distance. Designing the loss function proves to be challenging for shapes with complex topology and details (Fig. 1). We propose HOTSPOT, a method based on a relation between the screened Poisson equation and distance that enjoys both theoretical and practical benefits.", + "bbox": [ + 500, + 497, + 890, + 603 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A major challenge in signed distance function optimization is ensuring the implicit function follows the true distance. A common regularization loss used is the eikonal equation, which constrains the gradient norm of an implicit function to be 1 almost everywhere. If the implicit function is a signed distance function, then it satisfies the eikonal equation. However, the converse is not true. That is, satisfying the eikonal equation is only a necessary condition for the signed distance function, not a sufficient condition [5]. Fig. 2 presents the logic and Fig. 3 shows an example.", + "bbox": [ + 496, + 607, + 892, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Another key challenge lies in the optimization. It is known that the eikonal equation regularization is unstable [4], and the instability can make optimization converge to suboptimal results or even diverge.", + "bbox": [ + 496, + 761, + 890, + 821 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Furthermore, to address optimization ill-posedness and instability, existing methods penalize the surface area of the shape. We prove this degrades the reconstruction quality and distorts the output level sets, thereby leading to a delicate balance between over- and under-smoothing.", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 94, + 886, + 205, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1276", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{c} \\min \\{\\text {h e a t l o s s} \\} \\\\ \\lambda \\to \\infty \\end{array} \\Leftrightarrow \\begin{array}{l} \\text {D i s t a n c e} \\quad \\text {a . e .} \\\\ \\text {F u n c t i o n} \\quad \\stackrel {{\\Rightarrow}} {{\\Leftarrow}} \\end{array} \\left\\{ \\begin{array}{l} \\text {E i k o n a l c o n d i t i o n} \\\\ \\text {Z e r o g r a d i e n t d i v e r g e n c e} \\\\ \\text {S i n g u l a r H e s s i a n m a t r i x} \\\\ \\dots \\end{array} \\right.\n$$\n", + "text_format": "latex", + "bbox": [ + 83, + 89, + 464, + 142 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We propose a simple model that can address all the challenges above. Our model, HOTSPOT, is based on a screened Poisson equation and a classical relation between heat transfer and distance [6] that was popular for approximating both Euclidean and geodesic distance [7-9]. Using this relation, we design a loss function for optimizing neural signed distance functions. As our contributions, we show, both theoretically and empirically, that our model", + "bbox": [ + 75, + 272, + 468, + 393 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- serves as a sufficient condition ensuring the difference between the minimizer and the actual distance is bounded and decreases linearly as the parameter $\\lambda$ increases (Fig. 2 and Fig. 3, right).", + "- is stable both spatially and temporally, in the sense that a slight perturbation of the implicit function only introduces changes in a local region, and that the dynamical system formed by the gradient flow will converge in the long run.", + "- penalizes the surface area naturally, while remaining an undistorted signed distance function.", + "- works well in both 2D and 3D shapes, and it excels at approaching signed distance functions while optimizing for complex, high-genus shapes." + ], + "bbox": [ + 76, + 393, + 468, + 604 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 618, + 218, + 633 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Implicit surfaces. Modeling and reconstructing surfaces using implicit functions has a long history in computer vision and graphics [11-15]. Recently, there is a surge of interest to model surfaces using neural implicit functions [16-20]. Neural implicit representations are favored for their compactness, ability to represent detailed surfaces, and compatibility to gradient-based optimization [1].", + "bbox": [ + 75, + 643, + 468, + 750 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural signed distance functions. Among the classes of implicit surfaces, signed distance functions, where the implicit function outputs the distance to the closest point on the surface, and the sign represents whether the point is inside or outside the solid object, is one of the most popular variants. Signed distance functions enable an efficient rendering algorithm called sphere tracing [21-24] and helps geometry compression [25]. They can also be used for collision detection [26-28] and numerical simulation [29]. Traditionally, signed distance functions are often represented", + "bbox": [ + 75, + 750, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b06f6a187e16cd03ca5cd591c9a9b1a69f7c7a7b9b8bab6a8d473ec08cb4d766.jpg", + "image_caption": [ + "Figure 2. Existing constraints are only necessary but insufficient conditions. That is, there exist solutions, like Fig. 3, that satisfy them almost everywhere yet fail to be distance functions. In contrast, we use a heat loss, modeled by a screened Poisson equation with parameter $\\lambda$ . As the parameter $\\lambda$ goes to infinity, the minimizer converges to the distance function and excludes other solutions. Thus, our method is asymptotically both necessary and sufficient.", + "Figure 3. We illustrate a 1D neural signed distance function optimization using the classical eikonal loss [10] and our model. The $x$ axis is the domain and the $y$ axis shows the output of the implicit function. The middle and bottom rows show the intermediate and final states of the optimization respectively. The eikonal loss, as a necessary but insufficient condition, is incapable of converging to the actual signed distance function (dashed line), even when the function satisfies the eikonal equation almost everywhere." + ], + "image_footnote": [], + "bbox": [ + 550, + 89, + 836, + 339 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "as grids [30-32]. More recently, neural signed distance functions have become an appealing geometric representation [33-35]. (Neural) Signed distance functions are also often used in inverse rendering [36-42], for their flexibility to obtain surfaces with complex topology.", + "bbox": [ + 496, + 470, + 890, + 546 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Regularizing and initializing neural signed distance functions. A commonly used loss for optimizing neural signed distance functions is the eikonal regularization, which encourages the norm of the gradient of the implicit function to be one [10, 39]. However, the eikonal regularization alone is far from sufficient, as 1) minimizing eikonal loss does not necessarily lead to a distance field [5, 30], and 2) its optimization is unstable [4]. Previous works have proposed initialization to bootstrap the optimization with a simple shape [3, 10]. Recent works show that minimizing the divergence and the surface area of the implicit function can lead to better reconstruction [3, 43] and stabilizes the optimization [4]. Second-order information is also often used [43-46] for regularization. All losses above are based on necessary properties of real distance but neither establish convergence nor exclude non-SDF solutions.", + "bbox": [ + 496, + 549, + 892, + 791 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specialized architectures have been proposed to limit the Lipschitz constant of the implicit function [35], satisfy $p$ -Poisson equation [47], or represent high-frequency details [25, 34, 48]. Some research [49, 50] has shown a link between occupancy function optimization and an approximate signed distance field regularization. Our experiments further reveal that some implementations [2, 51, 52]", + "bbox": [ + 496, + 794, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1277", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "optimizing SDFs with the closest point may struggle when querying near the interpolated surface.", + "bbox": [ + 75, + 90, + 468, + 119 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose a new loss and method for optimizing signed distance functions. While derived from very different mathematics, our loss turns out to have a close connection to Lipman's PHASE loss [49]. We show that our theory leads to significantly more numerically stable optimization, and a different conclusion in the parameter setting that leads to drastically different results.", + "bbox": [ + 75, + 121, + 468, + 224 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Distance approximation using heat transfer. Our derivation is inspired by a relation between heat transfer and approximated distance [6]. This relation has been used in computer graphics for efficiently computing the geodesic distance on surfaces [7-9, 53, 54]. We apply the relation for optimizing neural signed distance functions, and analyze the properties of the resulting optimization in this context.", + "bbox": [ + 75, + 227, + 468, + 332 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Background and Motivation", + "text_level": 1, + "bbox": [ + 76, + 345, + 338, + 361 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given an input point cloud $\\pmb{x}_i\\in \\mathbb{R}^d$ (we focus on 2D and 3D in this work) with $N$ points without normal information that lies on an unknown surface $\\Gamma$ of a solid object $S$ , our goal is to find a signed distance function $f(\\pmb {x})$ , such that", + "bbox": [ + 75, + 369, + 468, + 431 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf (\\boldsymbol {x}) = \\left\\{ \\begin{array}{l l} - d _ {\\Gamma} (\\boldsymbol {x}) & \\text {i f} \\boldsymbol {x} \\in S \\\\ d _ {\\Gamma} (\\boldsymbol {x}) & \\text {o t h e r w i s e ,} \\end{array} \\right. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 440, + 468, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $d_{\\Gamma}(\\pmb{x})$ denotes the closest distance from point $\\pmb{x}$ to the surface $\\Gamma$ . In practice, we approximate the signed distance function $f(\\pmb{x})$ using a neural network $u(\\pmb{x}): \\mathbb{R}^d \\to \\mathbb{R}$ parameterized by its weights and use gradient-based optimization to figure out the weights. Previous work (e.g., [10]) often uses a linear combination of a boundary loss $L_{\\mathrm{boundary}}$ and a eikonal loss $L_{\\mathrm{eikonal}}$ :", + "bbox": [ + 75, + 488, + 468, + 595 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {b o u n d a r y}} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} | u (\\boldsymbol {x} _ {i}) | ^ {p}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 98, + 604, + 468, + 643 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {e i k o n a l}} = \\int_ {\\Omega} | \\| \\nabla u (\\boldsymbol {x}) \\| - 1 | ^ {p} \\mathrm {d} \\boldsymbol {x}, \\quad p = 1 \\text {o r} 2, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 648, + 468, + 680 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\Omega \\subset \\mathbb{R}^d$ is our domain of interest. The boundary loss encourages the implicit function $u$ to output 0 at the boundary, and the eikonal loss is based on that observation that if a function $f$ is a distance function, the $L^2$ norm of its gradient $\\| \\nabla f(\\pmb{x})\\| = 1$ holds almost everywhere. Thus the approximation $u$ is encouraged to behave the same.", + "bbox": [ + 75, + 688, + 468, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, as noted, simply optimizing for the two terms above does not guarantee that $u$ would be a distance function even when satisfied almost everywhere (Figs. 2 and 3), and the extra regularization previous work added to minimize the area and divergence [3, 4] could further tamper with the distance approximation quality. We derive an additional loss that can be used together with both of the losses above, and analyze our loss theoretically and empirically.", + "bbox": [ + 75, + 780, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5dfab8084f9c63b1361b5c2879bf2d56d0ffec1bd6d0b09daced5bc564ec7969.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 88, + 694, + 203 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9d07dff7616f91a37d1a713146f880761e116e3bcd78cc309afe6228b64be931.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 696, + 88, + 885, + 203 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9ff077163feb3e8b536152bf928464003c5f4d0b8f6bfaa6c50f62203c8997b2.jpg", + "image_caption": [ + "(c) HOUSE" + ], + "image_footnote": [], + "bbox": [ + 506, + 222, + 593, + 292 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5c75ab31f618615213776769f9e9fd3672bdc3296f16f4db456dc7f8c55753bb.jpg", + "image_caption": [ + "(a) 1D Heat Simulation via Eq. (4)", + "(d) $\\lambda = 1$", + "Figure 4. We show an illustration of the relation between screened Poisson equation and a distance field. The top shows a 1D example, where on the left we show a solution to Eq. (4) with different absorption $\\lambda$ , and on the right we show the reconstructed distance $-\\frac{1}{\\lambda}\\ln(h)$ (Eq. (5)). Boundary loss for real boundary points acts as an isothermal heat source, while heat loss diffuses the heat. When $\\lambda$ is increased, the heat decays faster, and the error of the reconstructed function approaches 0. The bottom shows a 2D example." + ], + "image_footnote": [], + "bbox": [ + 598, + 220, + 694, + 295 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/474199d8dbc3f6e92fa3f367cdeb6155f8e54a98a2fd0adfd608ac741a0a8232.jpg", + "image_caption": [ + "(b) Distance Reconstruction", + "(e) $\\lambda = 5$" + ], + "image_footnote": [], + "bbox": [ + 696, + 220, + 792, + 295 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/169c024f19af33225d597fd7b78d5b377a35da7298ba1ac6ee7880f21995eb89.jpg", + "image_caption": [ + "(f) $\\lambda = 50$" + ], + "image_footnote": [], + "bbox": [ + 794, + 220, + 887, + 294 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We will be using a key relation between the screen Poisson equation and the distance. We first define a screened Poisson equation over a heat field $h(\\pmb{x}): \\mathbb{R}^d \\to \\mathbb{R}$ , with an absorption coefficient $\\lambda$ , and set a Dirichlet boundary condition to be 1 on our target surface $\\Gamma$ :", + "bbox": [ + 498, + 436, + 890, + 512 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\begin{array}{l}\\nabla^ {2} h (\\boldsymbol {x}) - \\lambda^ {2} h (\\boldsymbol {x}) = 0 \\quad \\forall \\boldsymbol {x} \\in \\mathbb {R} ^ {d} \\backslash \\Gamma\\\\\\lim _ {| | \\boldsymbol {x} | | \\rightarrow \\infty} h (\\boldsymbol {x}) = 0, \\quad h (\\boldsymbol {x}) = 1 \\quad \\forall \\boldsymbol {x} \\in \\Gamma .\\end{array}\\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 518, + 890, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "If a heat field $h_\\lambda$ is a solution to the partial differential equation above, then the following equation holds [6]:", + "bbox": [ + 500, + 571, + 888, + 603 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\lim _ {\\lambda \\rightarrow \\infty} \\frac {1}{\\lambda} \\ln \\left(h _ {\\lambda} (\\boldsymbol {x})\\right) = - d _ {\\Gamma} (\\boldsymbol {x}). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 593, + 609, + 890, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fig. 4 shows the intuition. The screened Poisson equation simulates heat transfer until an equilibrium is reached, under absorption controlled by the $\\lambda$ coefficient. As we increase the absorption, the reconstructed distance $\\frac{1}{\\lambda}\\ln (h_{\\lambda})$ would approach the real distance. In practice, we could only obtain a discrete point set $\\Gamma_0$ . Provided that the neighbor points are close enough, if there are some boundaries connecting them, the boundary condition of Eq. (4) still holds. This can be interpreted as a superset $\\Gamma$ replacing $\\Gamma_0$ , effectively transforming the boundary from a discrete point cloud into a continuous surface, which is required and precisely the goal of the signed distance function reconstruction. Obtaining a distance function to the point cloud rather than the interpolated surface is not our objective. The reconstructed distance can still faithfully represent the distance to $\\Gamma$ because the approximation in Eq. (5) still holds. Later, we will prove that the measure of $\\Gamma$ is bounded by our loss.", + "bbox": [ + 496, + 643, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1278", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 76, + 89, + 169, + 104 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we introduce our method, HOTSPOT, based on the screened Poisson equation (Eq. (4)) and derive a theoretical analysis. We use the screened Poisson equation to derive a loss we can directly apply to an implicit function $u$ , and prove that there is a bound to the solution of the screened Poisson equation to the true distance (Section 4.1). We further derive a stability analysis of our loss on both the spatial stability over small perturbations and the temporal convergence (Section 4.2). Next, we show that our loss can penalize surfaces with large area (Section 4.3). Finally, we discuss the relation to prior work, PHASE (Section 4.4).", + "bbox": [ + 75, + 114, + 472, + 282 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Screened Poisson Equation Informed Signed Distance Functions Optimization", + "text_level": 1, + "bbox": [ + 76, + 289, + 470, + 321 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given an implicit function $u$ , we want to derive a loss based on the screened Poisson equation (Eq. (4)) and its relation to the distance (Eq. (5)). We achieve this by applying the following substitution based on Eq. (5):", + "bbox": [ + 75, + 327, + 468, + 388 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nh (\\boldsymbol {x}) = e ^ {- \\lambda | u (\\boldsymbol {x}) |}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 209, + 396, + 468, + 416 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We want to design a loss such that its minimizer satisfies the screened Poisson equation. This can be achieved by minimizing the energy functional $\\frac{1}{2}\\int_{\\Omega} \\| \\nabla h(\\pmb{x}) \\|^2 + \\lambda^2 h(\\pmb{x})^2 \\, \\mathrm{d}\\pmb{x}$ . After substitution and removing of the constant factor $\\lambda$ , we obtain our heat loss:", + "bbox": [ + 76, + 424, + 468, + 500 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {h e a t}} = \\frac {1}{2} \\int_ {\\Omega} e ^ {- 2 \\lambda | u (\\boldsymbol {x}) |} \\left(\\| \\nabla u (\\boldsymbol {x}) \\| ^ {2} + 1\\right) d \\boldsymbol {x}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 507, + 468, + 540 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "If we take the derivative of $L_{\\mathrm{heat}}$ with respect to the heat field $h$ , we recover the screened Poisson equation $\\nabla^2 h - \\lambda^2 h = 0$ . The boundary condition can be enforced by the standard boundary loss (Eq. (2)), by observing that $u(\\pmb{x}) = 0$ implies $h(\\pmb{x}) = 1$ . Together with the eikonal loss, we optimize:", + "bbox": [ + 76, + 547, + 468, + 625 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL = w _ {b} L _ {\\text {b o u n d a r y}} + w _ {e} L _ {\\text {e i k o n a l}} + w _ {h} L _ {\\text {h e a t}}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 633, + 468, + 650 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While prior work [6] has studied the limiting behavior of the absorption coefficient $\\lambda$ , to our knowledge, the bounds of the approximation and the convergence speed remain unknown. In the supplementary material A.3, we prove that, for a boundary $\\Gamma$ formed by a discrete set of points and a small volume around them, the distance approximation obtained from the solution of the screened Poisson equation converges linearly with respect to $\\lambda$ :", + "bbox": [ + 75, + 657, + 468, + 780 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {1}{\\lambda} C _ {\\text {l o w e r}} (\\boldsymbol {x}) \\leqslant \\frac {1}{\\lambda} \\ln \\left(h _ {\\lambda} (\\boldsymbol {x})\\right) + d _ {\\Gamma} (\\boldsymbol {x}) \\leqslant \\frac {1}{\\lambda} C _ {\\text {u p p e r}} (\\boldsymbol {x}). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 787, + 468, + 818 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, both $C_{\\mathrm{upper}}(\\pmb{x})$ and $C_{\\mathrm{lower}}(\\pmb{x})$ are scalar fields independent of $\\lambda$ . The inequalities are tight, which means that one can find a case where the equality holds.", + "bbox": [ + 75, + 824, + 468, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Compared to the eikonal loss (Eq. (3)), where the distance error can be unbounded even when the equation holds", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/10116da4676066829bfb6317112af6411f3b4a5a054c8bd4e130bdcdbb90e036.jpg", + "image_caption": [ + "Figure 5. We visualize the direction of the changes of the implicit function during one iteration of optimization. Left shows the result with only the eikonal loss and right shows the result after adding our loss. Our loss helps to escape the local minimum in Fig. 3. The heat diffusion encourages the reconstructed absolute distance to increase monotonically away from the boundary." + ], + "image_footnote": [], + "bbox": [ + 504, + 88, + 887, + 204 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "almost everywhere, Eq. (9) shows that even with a finite absorption $\\lambda$ , our loss can lead to much more accurate distance. In fact, any loss that only depends on the norm of the gradients $\\|\\nabla u\\|$ is not a sufficient condition of the distance function (Fig. 2), due to the discontinuous jumps of the gradient. Adding our loss (Eq. (7)) ensures that the absolute value of the implicit function converges to the actual distance. Fig. 5 visualizes the direction of how an implicit function would evolve when using only the eikonal loss and after adding our loss. Our loss ensures that the implicit function evolves towards the actual distance.", + "bbox": [ + 496, + 309, + 893, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Stability Analysis", + "text_level": 1, + "bbox": [ + 498, + 487, + 671, + 503 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our heat loss enjoys other theoretical benefits apart from the convergence to an actual distance. Here, we show that our loss leads to easier optimization. We separately analyze spatial stability and temporal stability, where the spatial stability studies the effect of small perturbation error added to the corresponding partial differential equation, and the temporal stability studies the convergence of the gradient flow.", + "bbox": [ + 496, + 511, + 893, + 617 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2.1 Spatial Stability", + "text_level": 1, + "bbox": [ + 498, + 637, + 666, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We show that for the screened Poisson equation, an error introduced to a single point in the solution decays exponentially and does not propagate to infinitely distant positions. On the contrary, for the eikonal equation, the error can propagate to infinitely far. The property of the screened Poisson equation ensures that the perturbation introduced during optimization does not lead to a drastically different target and stabilizes the optimization.", + "bbox": [ + 496, + 662, + 890, + 781 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "First, we show that for the eikonal equation, an additive error introduced to a point can presist along a line segment or a ray.", + "bbox": [ + 496, + 782, + 890, + 829 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Proposition 1. Given a field $u(\\pmb{x})$ that satisfies the eikonal equation $\\| \\nabla u \\| = 1$ , we introduce an additive error $u_{0e}$ at the point $\\pmb{x}_0$ where $u(\\pmb{x}_0) = u_0$ to obtain a perturbed field $u'$ where $u'(\\pmb{x}_0) = u_0 + u_{0e}$ and $\\| \\nabla u' \\| = 1$ . In the", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1279", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "perturbed field $u'$ , there exists a line segment parameterized by $\\pmb{x}(s)$ starting from $\\pmb{x}_0$ and $s \\in [0, M]$ , where $M > 0$ and can be infinite, such that every point along the line has the same error:", + "bbox": [ + 75, + 90, + 472, + 151 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nu ^ {\\prime} (\\boldsymbol {x} (s)) - u (\\boldsymbol {x} (s)) = u _ {0 e} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 160, + 468, + 178 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The boundary points can contain errors when approximating the true boundaries, and these errors remain constant along the line segment. As a result, the propagated error can become significant and potentially detrimental.", + "bbox": [ + 76, + 186, + 468, + 246 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Next, we demonstrate that the error field $h_e$ in Eq. (4) decays exponentially from the perturbed point $x_0$ .", + "bbox": [ + 76, + 246, + 468, + 277 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proposition 2. In 3D space, consider an additive error $h_{0e}$ within a small ball $B(\\pmb{x}_0,\\epsilon)$ around the point $\\pmb{x}_0$ , where the original field $h(\\pmb{x})$ and the perturbed field $h^{\\prime}(\\pmb{x})$ , $\\pmb{x} \\in \\mathbb{R}^n\\backslash B$ both satisfy the screened Poisson equation in Eq. (4). The resulting error field $h_e = h' - h$ is radially symmetric with respect to $x_0$ within some maximum radius $R$ , and is given by", + "bbox": [ + 76, + 284, + 468, + 388 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nh _ {e} (r) = \\frac {\\epsilon}{r} h _ {0 e} e ^ {\\lambda (\\epsilon - r)}, \\forall r > \\epsilon , \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 388, + 468, + 414 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $r = \\left\\| \\boldsymbol{x} - \\boldsymbol{x}_0 \\right\\|$ is the radial distance from $\\boldsymbol{x}_0$ .", + "bbox": [ + 76, + 417, + 421, + 434 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The proofs of these two propositions are in the supplementary material A.1 and A.2. By comparing the error fields, we observe that the partial differential equation governed by the screened Poisson equation exhibits greater spatial robustness than the one based on the eikonal equation.", + "bbox": [ + 76, + 441, + 468, + 516 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.2 Temporal Stability", + "text_level": 1, + "bbox": [ + 76, + 534, + 261, + 549 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We next demonstrate the temporal stability of the heat loss and show that the dynamical system of the gradient update is stable, that is, the solution of the system converges as the time $t$ approaches infinity. This is supported by von Neumann stability analysis, similar to Yang et al.'s work [4].", + "bbox": [ + 76, + 556, + 468, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Yang et al. showed that for the eikonal loss (Eq. (3)), the temporal gradient flow of the optimization is:", + "bbox": [ + 76, + 633, + 468, + 662 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\frac {\\partial u}{\\partial t} = - \\frac {\\delta L _ {\\mathrm {e i k o n a l}}}{\\delta u} = \\nabla \\cdot (\\kappa_ {e} \\nabla u) \\approx \\kappa_ {e} \\nabla^ {2} u, \\text {w h e r e} \\\\ \\kappa_ {e} (\\boldsymbol {x}) = \\left\\{ \\begin{array}{l l} \\frac {1}{| \\nabla u (\\boldsymbol {x}) |} \\operatorname {s g n} (| \\nabla u (\\boldsymbol {x}) | - 1) & p = 1 \\\\ 1 - | \\nabla u (\\boldsymbol {x}) | ^ {- 1} & p = 2 \\end{array} \\right. \\tag {12} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 98, + 671, + 468, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "They analyze the stability by applying Fourier transform:", + "bbox": [ + 96, + 750, + 467, + 765 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial \\hat {u}}{\\partial t} (t, \\omega) = - \\kappa_ {e} | \\omega | ^ {2} \\hat {u} (t, \\omega) \\Rightarrow \\hat {u} (t, \\omega) \\propto e ^ {- \\kappa_ {e} | \\omega | ^ {2} t}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 772, + 468, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\omega = (\\omega_{1},\\dots ,\\omega_{n})$ denotes the frequency variable and $\\hat{u}$ represents the Fourier transform of $u$ . In the frequency domain, if the original function converges, the term $\\frac{\\partial\\hat{u}}{\\partial t}$ should also converge. However, when $\\kappa_{e} < 0$ , this process corresponds to backward diffusion, which diverges and is inherently unstable regardless of the numerical implementation.", + "bbox": [ + 76, + 809, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We extend the analysis to our heat loss (Eq. (7)). The gradient flow is", + "bbox": [ + 500, + 90, + 890, + 121 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial h}{\\partial t} = - \\frac {\\delta L _ {\\text {h e a t}}}{\\delta h} = \\nabla^ {2} h - \\lambda^ {2} h \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 128, + 890, + 159 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This is exactly the heat equation, and is known to be stable. We can further verify the stability by applying Fourier transform to both sides:", + "bbox": [ + 498, + 165, + 890, + 209 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial \\hat {h}}{\\partial t} (t, \\omega) = - \\left(| \\omega | ^ {2} + \\lambda^ {2}\\right) \\hat {h} (t, \\omega) \\Rightarrow \\hat {h} (t, \\omega) \\propto e ^ {- \\left(| \\omega | ^ {2} + \\lambda^ {2}\\right) t} \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 217, + 890, + 260 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This analysis reveals that the original process $h(\\pmb{x}, t)$ converges over time and remains stable. Due to the constraint $|u| = -\\ln(h) / \\lambda$ , the term $|u(\\pmb{x}, t)|$ also converges during the optimization process. Since $u$ is continuous, it converges as well. Thus, our method exhibits greater spatial and temporal stability compared to the eikonal equation.", + "bbox": [ + 498, + 261, + 890, + 351 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Surface Area Regularization", + "text_level": 1, + "bbox": [ + 500, + 359, + 754, + 376 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We next show that our heat loss (Eq. (7)) penalizes large area of the reconstructed surfaces, by showing its relation to the coarea loss used by Pumarola et al. [30]. Pumarola et al. showed that the loss $\\int_{\\Omega} e^{-\\lambda |u|} \\| \\nabla u \\| \\, \\mathrm{d}\\pmb{x}$ approximates the surface area of the zero level set of the implicit function $u$ . We then form the following inequality:", + "bbox": [ + 498, + 383, + 890, + 473 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\int_ {\\Omega} e ^ {- \\lambda | u |} \\| \\nabla u \\| d \\boldsymbol {x} < \\int_ {\\Omega} \\sqrt {e ^ {- 2 \\lambda | u |} \\left(\\| \\nabla u \\| ^ {2} + 1\\right)} d \\boldsymbol {x} (16) \\\\ \\leqslant \\sqrt {| \\Omega |} \\cdot \\sqrt {\\int_ {\\Omega} e ^ {- 2 \\lambda | u |} \\left(\\| \\nabla u \\| ^ {2} + 1\\right) d x} (16) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 481, + 890, + 556 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Therefore, minimizing the right hand side (our heat loss) leads to a reduction in the upper bound of the left hand side.", + "bbox": [ + 498, + 564, + 890, + 593 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Furthermore, we can show that the coarea loss is too aggressive for our purpose of signed distance field reconstruction. If we directly minimize the coarea loss (the left hand side in Eq. (16)), according to Euler-Lagrange equation, the minimizer would satisfy:", + "bbox": [ + 498, + 594, + 890, + 669 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n1 - \\nabla^ {2} u = 0, \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 647, + 678, + 890, + 694 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "which does not relate to the regularization parameter $\\lambda$ . Unfortunately, this partial differential equation system formed by the coarea loss above has a unique solution under Dirichlet boundary condition, and the gradient of its solution does not have a unit norm. This makes the coarea loss unable to accurately shape a signed distance function. By contrast, our heat loss has bounds for its distance approximation and converges to the true distance as we increase the parameter $\\lambda$ (Eq. (9)). Another simple way to verify our convergence and the compatibility with the eikonal equation is to rewrite Eq. (4) in terms of $u$ following Eq. (6):", + "bbox": [ + 498, + 704, + 890, + 869 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left| \\left| \\nabla u \\right| \\right| ^ {2} - 1 - \\frac {\\operatorname {s g n} (\\mathrm {u})}{\\lambda} \\nabla^ {2} u = 0. \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 873, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1280", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e4e678153ee695bcdeac6ecac31d2699d70d595a44f3ffd69c80a350da2b3fd2.jpg", + "image_caption": [ + "(a) GT" + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 153, + 146 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/64868793a5d0bb8cc56a12997271fffcde9068ca761fb3e30a9b5d66e2d2596f.jpg", + "image_caption": [ + "(b) $w_{h} = 0.1$" + ], + "image_footnote": [], + "bbox": [ + 156, + 90, + 225, + 146 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7f79dd38777fbf7a13513caf181ee5d8a1150015a7603b3717186b1e5019e149.jpg", + "image_caption": [ + "(c) $w_{h} = 0.2$" + ], + "image_footnote": [], + "bbox": [ + 227, + 90, + 299, + 146 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/576c0944855c1f97579431ecff819857a01688dcaf87b0f5e001ca47f43c43e1.jpg", + "image_caption": [ + "(d) $w_{h} = 0.3$" + ], + "image_footnote": [], + "bbox": [ + 302, + 90, + 372, + 146 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5af4162badb2fe6729de3b2e334b6929406fde9c157c75ddc9721877c97ac73d.jpg", + "image_caption": [ + "(e) $w_{b} = 0.5$" + ], + "image_footnote": [], + "bbox": [ + 374, + 90, + 452, + 146 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a686cc57f472ad0cc6a198247c85c9291ccb7197f3626d0699de39ebe985c1ab.jpg", + "image_caption": [ + "(f) $w_{b} = 1$" + ], + "image_footnote": [], + "bbox": [ + 81, + 161, + 153, + 218 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d1ac9ae1207d822f36749d7c7c7869cc233e0ddcf303aea168543f8e006515ee.jpg", + "image_caption": [ + "(g) $w_{b} = 2$" + ], + "image_footnote": [], + "bbox": [ + 156, + 161, + 225, + 218 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5d43e0c22b7d8cb496a46aad7f495b65a4ee24f39712a55a32880777cccdb404.jpg", + "image_caption": [ + "(h) $w_{b} = 5$" + ], + "image_footnote": [], + "bbox": [ + 227, + 161, + 299, + 218 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/55af226a5059f90e1e84c1f23f25d4b17874d50ae39d371e172f0722f155573e.jpg", + "image_caption": [ + "(i) $w_{b} = 10$", + "Figure 6. We reproduce the PHASE model [49] and show the log-transformed occupancy field, i.e., the optimized signed distance function. PHASE recommends boundary weight range $w_{b} \\in (0.10, 0.32)$ based on their $\\Gamma$ -convergence theory. But in practice, only a stronger boundary weight constrains the boundary to follow the input points. We found that this phenomenon often occurs and show more examples in the supplementary material." + ], + "image_footnote": [], + "bbox": [ + 300, + 161, + 372, + 218 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2bc27a3330646b9d3cb33371d3c5026e0dbfc1ab0254d79b1cfd420bebd3fea3.jpg", + "image_caption": [ + "(j) $w_{b} = 20$" + ], + "image_footnote": [], + "bbox": [ + 374, + 161, + 452, + 218 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Relation to PHASE [49]", + "text_level": 1, + "bbox": [ + 76, + 340, + 295, + 357 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "While derived from different mathematical principles and motivations, our method is closely related to the PHASE model proposed by Lipman. It is motivated by smoothly approximating an occupancy function $o(x)$ , which is defined as 1 outside, -1 inside, and 0 on the surface.", + "bbox": [ + 75, + 363, + 468, + 438 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the supplement, we show that there is a close relation between our heat field $h$ and the smoothed occupancy function $o$ . We also show that our theory and derivation provide additional insights and significant differences in implementation and results. In particular, PHASE's theory introduces a condition stating that the boundary loss weight $w_{b}$ should be small, whereas our theory emphasizes the need for a firm boundary condition (Eq. (4) and Fig. 6).", + "bbox": [ + 75, + 439, + 468, + 559 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We further found that optimizing occupancy field in the PHASE model, when used together with the eikonal loss, can lead to unbounded network weight derivatives, impeding the optimization. Our method avoids this issue.", + "bbox": [ + 75, + 560, + 468, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 633, + 209, + 650 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We experiment on both 2D and 3D datasets. We construct a 2D dataset containing 14 shapes, ranging from simple geometric shapes to complex nested structures. We also evaluate on two 3D benchmarks: a subset of ShapeNet [55] and the Surface Reconstruction Benchmark (SRB) [56]. Additionally, we generate a dataset with high-genus geometries from Mehta et al. [1]. In all of our experiments, we use only the position information of the point cloud.", + "bbox": [ + 75, + 657, + 468, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We report the Intersection over Union (IoU), Chamfer distance, and Hausdorff distance between reconstructed shapes and ground truth shapes as metrics for surface reconstruction. In addition, we report Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and Symmetric Mean Absolute Percentage Error (SMAPE) as metrics for distance field evaluation. We further evaluate non-zero level set quality near the surface, and render the zero level", + "bbox": [ + 75, + 780, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "set of the signed distance functions via sphere tracing.", + "bbox": [ + 498, + 90, + 854, + 106 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use a network with 5 hidden layers and 128 hidden channels for our method in all experiments. The hyperparameters are kept consistent within each dataset. We additionally apply importance sampling to our heat loss integral. Full training details can be found in the supplement.", + "bbox": [ + 496, + 106, + 890, + 181 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1.2D Dataset", + "text_level": 1, + "bbox": [ + 500, + 191, + 619, + 205 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "An overview of our 2D dataset is attached in the supplementary material. For 2D experiments, we simply use a linear network and geometric initialization, as in DiGS [3]. We compare our method with DiGS [3] and StEik [4] with the hyperparameters and training configuration provided in their released code, and report the surface reconstruction metrics and distance field metrics in Table 1. Our method outperforms theirs across all metrics with only 8,192 sample points, whereas both DiGS [3] and StEik [4] use 15,000 sample points based on their default settings.", + "bbox": [ + 496, + 214, + 890, + 366 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Ablation Study on 2D Dataset", + "text_level": 1, + "bbox": [ + 498, + 375, + 764, + 391 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To more comprehensively compare over different losses, we conduct an ablation study on our 2D dataset using the same network architecture, parameter initialization, and sampling strategy. The only difference lies in the loss functions used. Quantitative results are shown in Table 2.", + "bbox": [ + 496, + 398, + 890, + 473 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The last row corresponds to our HOTSPOT model, combining boundary, eikonal, and heat losses. Our method reconstructs the correct topology for all 14 shapes, while others struggle with the harder ones. It outperforms across nearly all metrics, demonstrating superior reconstruction quality and distance query accuracy. Although HOTSPOT slightly trails SAL [2] in RMSE and MAE, it captures the complex topology where SAL fails. A visualization is shown in Fig. 7. Full qualitative results and detailed experiment parameters are provided in the supplementary material. Models relying heavily on area losses [3, 4] often face a trade-off between preserving original details and eliminating redundant boundaries, which distorts the distance field (Section 4.3). Moreover, methods without our heat loss fall in local optimum easily.", + "bbox": [ + 496, + 474, + 890, + 702 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. 3D Datasets", + "text_level": 1, + "bbox": [ + 500, + 710, + 627, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate our method on a processed subset [16, 19, 57] of ShapeNet [55], which provides surface points sampled on preprocessed watertight meshes for 260 shapes across 13 categories. To ensure a reasonable spatial density for passing the heat, we scale the point cloud so that $70\\%$ of the points lie within a sphere centered at the origin with a radius of 0.45. For a fair comparison, we transform our outcomes back into the coordinate system used in StEik [4] and DiGS [3]. Additionally, we implement a scheduler to gradually increase the absorption parameter $\\lambda$ , which enhances the representation of level set details.", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1281", + "bbox": [ + 483, + 944, + 513, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c558a5a646cd28ab71a02d4c0ed77d18f5aaa4e2585436c26a4a50cadd03d8dc.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
IoU ↑dC ↓dH ↓
DiGS [3]0.78820.00550.1267
StEik [4]0.66200.00730.1425
Ours0.98700.00140.0153
RMSE ↓MAE ↓SMAPE ↓
DiGS [3]0.05970.03150.3363
StEik [4]0.07250.04190.4222
Ours0.01990.01010.0699
", + "bbox": [ + 102, + 88, + 352, + 202 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e2858aeeda5e2529951057ec424b0f9b3ae40abd69940bbe9c6be6fbab3f8d23.jpg", + "table_caption": [ + "Table 1. Comparison of 2D reconstruction and distance query. $d_{C}$ and $d_{H}$ stand for the Chamfer and Hausdorff distances respectively. Our method outperforms DiGS [3] and StEik [4] on all metrics." + ], + "table_footnote": [], + "table_body": "
LBLELALDLDDLSLH| IoU ↑dC ↓dH ↓RMSE ↓MAE ↓SMAPE ↓
✓✓0.81920.00680.07120.03770.01580.1315
✓✓0.89360.00290.05790.01650.00810.1205
✓✓✓0.39070.08430.55980.29660.23881.7283
✓✓✓✓0.63380.05660.42210.20580.16161.0919
✓✓✓0.29380.03770.34440.29900.23551.6958
✓✓✓✓0.60370.04140.36560.13040.09370.7792
✓✓✓0.98510.00160.01600.01990.01050.0754
", + "bbox": [ + 369, + 89, + 911, + 186 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Ablation study of different loss combinations and metrics. From top to bottom, the loss combinations are from: IGR [10], SAL [2], DiGS w/o area loss, DiGS [3], StEik w/o loss, StEik [4], and our HOTSPOT. From left to right, the losses are boundary loss, eikonal loss, area loss, divergence loss, directional divergence loss, SAL loss, and heat loss. Bold and underlined data: optimal; bold only: suboptimal. Same below.", + "bbox": [ + 367, + 191, + 870, + 273 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/94247f3a293ba9f4c00834afb622f0e093198f5fe9c1f91ffd97e23b4b2bd9d0.jpg", + "image_caption": [ + "Figure 7. Our heat loss achieves the best reconstruction of triangle fragments. Dashed lines show real boundaries, while the boldest line represents reconstructed ones. (c) and (f) are from the original models by DiGS [3] and StEik [4]; the rest are from our ablation study with different loss combinations. Other combinations fail to remove extra boundaries or accurately capture the topology. More fundamentally, (c), (f), (g), and (h) fall into local optima, leading to incorrect depressions and elevations." + ], + "image_footnote": [], + "bbox": [ + 117, + 282, + 426, + 539 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We present the quantitative results in Table 3 and visual results in Fig. 9. More baselines [46, 51] are compared in the supplement. Our HOTSPOT model outperforms current state-of-the-art models across all surface reconstruction metrics, and achieves near-top accuracy in distance queries. While SAL [2] uses the closest point distance in their loss function, leading to slightly lower total absolute error, their performance drops significantly in relative error and near-surface metrics. This suggests that the closest point distance alone is unreliable near the surface. In contrast, as mentioned at the end of Section 3, after the interpolation driven by spectral bias [58], our loss remains effective in faithfully capturing the distance to the interpolated surface.", + "bbox": [ + 75, + 670, + 467, + 866 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 1 and Fig. 8 show high-genus surface reconstruction with shapes from Mehta et al. [1]. Existing methods strug", + "bbox": [ + 76, + 869, + 467, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/94f5f2ec8c7d661f1a9ad20120447a46903ab1372fc13043b58d72c717fb34e9.jpg", + "image_caption": [ + "DiGS" + ], + "image_footnote": [], + "bbox": [ + 501, + 279, + 589, + 334 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e13d43aa3c5556beb76b1652d1eec0209d81ec4c2c70d0f5955d645630cb935f.jpg", + "image_caption": [ + "StEi" + ], + "image_footnote": [], + "bbox": [ + 596, + 277, + 661, + 333 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6b046b7715874413a0a47c7cc94400126fd77442358555c004bbb85c872f3fed.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 669, + 277, + 736, + 333 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/51473219e1974c3d7d89b5eda798f76b03294d3ccfb47189da02b7379c8299f8.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 743, + 277, + 808, + 333 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4cc90aa9a24d5ef8849bd094c0e4db3f967f0f19c53c8a8aded6dd61ec77dcd2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 818, + 277, + 883, + 333 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0e38d075dfaa22550ddf3e77464741a018f298be9719ab07e82e09ee7a23ea2d.jpg", + "image_caption": [ + "Figure 8. Visual examples on high genus dataset. Our method uses significantly fewer iterations but reconstructs the topology correctly for all shapes, whereas others generate extra boundaries." + ], + "image_footnote": [], + "bbox": [ + 501, + 335, + 591, + 382 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/06b660072fafac8a3561dee2c69f91c20118edae6c7674cc45e0a6841c2ddcdf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 596, + 335, + 663, + 383 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/97216776a9c905a4e26418fc18109823757d679276a448d746dc55173ce82563.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 335, + 736, + 383 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b0dc0dca7a8ae1435404d6da9cc989eecaa1ce3deb3fa2595a29a74e56d8affa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 743, + 335, + 810, + 383 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6613d601fe160001a5bef3c7278680164b4b79e260668de84a5f166bcdfddfd6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 818, + 335, + 885, + 383 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "gle with the complex topology, while our method accurately reconstructs both the surface and the level set. In the supplement, we present evaluations on other high-genus shapes from Mehta et al. and the SRB dataset.", + "bbox": [ + 498, + 443, + 890, + 502 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Sphere Tracing", + "text_level": 1, + "bbox": [ + 500, + 512, + 653, + 529 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 10 visualizes the number of steps a sphere tracer [21] takes to render the shapes obtained by SAL, DiGS, StEik, and our method. Our method has the fewest iterations on average, since it provides the most accurate level sets. More results can be found in the supplement.", + "bbox": [ + 498, + 536, + 890, + 611 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6. Discussion", + "text_level": 1, + "bbox": [ + 500, + 619, + 612, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Efficiency. Our method uses a loss function of only the neural network output and its first-order derivatives, and thus can be more efficient than previous methods which involve second-order derivatives. We report the runtime of our method and StEik [4] for training a model from ShapeNet [55] over 100 iterations using a single NVIDIA A10 GPU with varying network sizes, while keeping all other settings identical, as shown in Tab. 4.", + "bbox": [ + 496, + 643, + 890, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Determining absorption coefficient $\\lambda$ . The absorption coefficient $\\lambda$ has the units of $[\\mathrm{L}]^{-1}$ , where $\\mathrm{L}$ denotes a length unit. Rescaling the space is equivalent to rescaling $\\lambda$ . While a larger $\\lambda$ can improve accuracy, it introduces two challenges: (1) our heat loss includes the factor $e^{-2\\lambda |u|}$ , which may become smaller than the precision limit of floating-point numbers, leading to round-off errors; (2) even if round-off errors are ignored, a large $\\lambda |u|$ may reduce the optimizer's ability to effectively shape the field. Never", + "bbox": [ + 496, + 765, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1282", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c9cdbbe5acae0741015467a61070f3a6dc308ec795d0542b217896e9e4e51dbf.jpg", + "image_caption": [ + "Figure 9. Visual results of DiGS [3], StEik [4], and HOTSPOT on ShapeNet [55]. Previous methods generate extra boundaries and fail to escape local optima, while our method successfully reconstructs the correct topology." + ], + "image_footnote": [], + "bbox": [ + 163, + 88, + 803, + 284 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f0d2b96dc1acd0982419a390263ffd82eabf51617c1c117380061c8b1dcde589.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
IoU ↑dC ↓dH ↓RMSE ↓MAE ↓SMAPE ↓RMSE0.1MAE0.1SMAPE0.1
SAL [2]0.74000.00740.08510.02510.01420.13440.02450.01820.6848
SIREN wo/ n [34]0.48740.00510.05580.50090.42611.26940.05130.03820.8858
DiGS [3]0.96360.00310.04350.11940.07250.21400.01520.00810.1760
StEik [4]0.96410.00320.03680.03870.02480.09310.01470.00810.1770
Ours0.97960.00290.02500.02810.01760.05400.00940.00470.1206
", + "bbox": [ + 163, + 320, + 805, + 398 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a68e5841683847a7d1fcf5c211e241cf47ccfe1a61181eebf86458bd14e70734.jpg", + "image_caption": [ + "Figure 10. Visualizations and statistics of sphere tracing [21] iteration counts required to locate the surface. In the visual results brighter pixels indicate higher iteration counts. The histograms show the distribution of iterations for all pixels across 10 different camera poses. HOTSPOT finds the surface in fewer steps." + ], + "image_footnote": [], + "bbox": [ + 81, + 455, + 467, + 712 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "theless, our loss still pushes $u$ away from 0, preventing unwanted surfaces where $\\lambda |u|$ should be large. We compensate by combining eikonal loss with heat losses. When $\\lambda |u|$ is large, the eikonal loss becomes dominant. Our $\\lambda$ scheduler also effectively shapes the distant regions. Spatially-adaptive parameters may also improve the results.", + "bbox": [ + 75, + 792, + 467, + 883 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Necessity of firm boundary condition. The heat diffu", + "bbox": [ + 76, + 885, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/876ba49726d52f85772e558399d44153ef6c35f4562c120a0b7ad059d67a528f.jpg", + "table_caption": [ + "Table 3. Surface reconstruction metrics on ShapeNet [55], where $d_{C}$ and $d_{H}$ stand for the Chamfer and Hausdorff distances, RMSE, MAE, and SMAPE are distance field metrics, and $\\mathrm{RMSE}_{0.1}$ , $\\mathrm{MAE}_{0.1}$ , and $\\mathrm{SMAPE}_{0.1}$ are the same distance field metrics but only for points within distance 0.1 of the surface. See the supplementary material for detailed results with more baselines." + ], + "table_footnote": [], + "table_body": "
Structure5 × 1285 × 2568 × 1288 × 256
StEik [4]5.65s10.62s8.78s17.00s
Ours4.49s8.43s6.79s12.86s
", + "bbox": [ + 539, + 450, + 849, + 497 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Runtime for 100 iterations with different network sizes.", + "bbox": [ + 501, + 503, + 885, + 517 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "sion of the screened Poisson equation is based on a well-established boundary condition. We enforce the boundary condition through the boundary loss over discrete points, with neural networks interpolating through spectral bias [58]. When the input points are sparse and the absorption coefficient $\\lambda$ is high, an overly strong heat loss ( $w_{b}$ being too small) can tear the boundary, causing the signed distance to collapse into an unsigned one like Fig. 6. Thus, our theory highlights the importance of a high boundary weight $w_{b}$ , while absorption $\\lambda$ is adjusted to match the input point density or by scaling the point cloud itself. Our experiments show that proper parameter setting and rescaling prevent the collapse. Future research is required for very sparse boundaries, and applications to inverse rendering.", + "bbox": [ + 496, + 523, + 890, + 736 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 743, + 617, + 757 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose a new model for neural signed distance function optimization based on the screened Poisson equation. We analyze our loss theoretically and show that it is an asymptotically sufficient condition to the true distance, is stable both to a small perturbation and in the temporal dynamics, and penalizes large surface area. Our experiments show that we reconstruct both better surfaces and better distance approximations compared to many existing methods, especially on complex and high-genus shapes.", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1283", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work was supported in part by NSF grants 2127544, 2238839, 2100237, 2120019, and gifts from Adobe and Google. Additionally, we would like to thank Bing Xu, Zilu Li, Yash Belhe, Ishit Mehta, and Jianan Xiao for their suggestions, discussions, proofreading, and providing reference materials.", + "bbox": [ + 76, + 90, + 472, + 183 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 193, + 174, + 209 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Ishit Mehta, Manmohan Chandraker, and Ravi Ramamoorthi. A level set theory for neural implicit evolution under explicit flows. In ECCV, pages 711-729, 2022. 1, 2, 6, 7, 8, 10", + "[2] Matan Atzmon and Yaron Lipman. Sal: Sign agnostic learning of shapes from raw data. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2565-2574, 2020. 1, 2, 6, 7, 8, 3, 5", + "[3] Yizhak Ben-Shabat, Chamin Hewa Koneputugodage, and Stephen Gould. DiGS: Divergence guided shape implicit neural representation for unoriented point clouds. In IEEE Conf. Comput. Vis. Pattern Recog., pages 19323-19332, 2022. 1, 2, 3, 6, 7, 8, 5", + "[4] Huizong Yang, Yuxin Sun, Ganesh Sundaramoorthi, and Anthony Yezzi. StEik: stabilizing the optimization of neural signed distance functions and finer shape representation. arXiv preprint arXiv:2305.18414, 2023. 1, 2, 3, 5, 6, 7, 8", + "[5] Zoë Marschner, Silvia Sellán, Hsueh-Ti Derek Liu, and Alec Jacobson. Constructive solid geometry on neural signed distance fields. In SIGGRAPH Asia 2023 Conference Papers, pages 1-12, 2023. 1, 2, 3", + "[6] Sathamangalam R Srinivasa Varadhan. On the behavior of the fundamental solution of the heat equation with variable coefficients. Communications on Pure and Applied Mathematics, 20(2):431-455, 1967. 2, 3, 4", + "[7] Keenan Crane, Clarisse Weischedel, and Max Wardetzky. Geodesics in heat: A new approach to computing distance based on heat flow. ACM TOG, 32(5):1-11, 2013. 2, 3", + "[8] Alexander G Belyaev and Pierre-Alain Fayolle. On variational and pde-based distance function approximations. In Computer Graphics Forum, volume 34, pages 104-118. Wiley Online Library, 2015.", + "[9] Alexander Belyaev and Pierre-Alain Fayolle. An ADMM-based scheme for distance function approximation. Numerical Algorithms, 84:983-996, 2020. 2, 3", + "[10] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. arXiv preprint arXiv:2002.10099, 2020. 2, 3, 7", + "[11] James F Blinn. A generalization of algebraic surface drawing. ACM TOG, 1(3):235-256, 1982. 2", + "[12] Jules Bloomenthal and Brian Wyvill. Interactive techniques for implicit modeling. Comput. Graph. (Proc. SIGGRAPH), 24(2):109-116, 1990.", + "[13] Hugues Hoppe, Tony DeRose, Tom Duchamp, John McDonald, and Werner Stuetzle. Surface reconstruction from unorganized points. Comput. Graph. (Proc. SIGGRAPH), 26(2):71-78, 1992.", + "[14] Brian Curless and Marc Levoy. A volumetric method for building complex models from range images. In SIGGRAPH, page 303-312, 1996." + ], + "bbox": [ + 78, + 218, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Shahram Izadi, David Kim, Otmar Hilliges, David Molyneaux, Richard Newcombe, Pushmeet Kohli, Jamie Shotton, Steve Hodges, Dustin Freeman, and Andrew Davison. KinectFusion: real-time 3D reconstruction and interaction using a moving depth camera. In ACM Symposium on User Interface Software and Technology, pages 559-568, 2011. 2", + "[16] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4460-4470, 2019. 2, 6, 7", + "[17] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5939-5948, 2019.", + "[18] Thomas Davies, Derek Nowrouzezahrai, and Alec Jacobson. On the effectiveness of weight-encoded neural implicit 3d shapes. arXiv preprint arXiv:2009.09808, 2020.", + "[19] Francis Williams, Matthew Trager, Joan Bruna, and Denis Zorin. Neural Splines: Fitting 3D surfaces with infinitely-wide neural networks. In IEEE Conf. Comput. Vis. Pattern Recog., pages 9949-9958, 2021. 6", + "[20] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. 41(2):641-676, 2022. 2", + "[21] John C Hart. Sphere tracing: A geometric method for the antialiased ray tracing of implicit surfaces. The Vis. Comput., 12(10):527-545, 1996. 2, 7, 8", + "[22] Dario Seyb, Alec Jacobson, Derek Nowrouzezahrai, and Wojciech Jarosz. Non-linear sphere tracing for rendering deformed signed distance fields. ACM Trans. Graph. (Proc. SIGGRAPH Asia), 38(6), 2019.", + "[23] Eric Galin, Eric Guérin, Axel Paris, and Adrien Peytavie. Segment tracing using local Lipschitz bounds, 2020.", + "[24] Nicholas Sharp and Alec Jacobson. Spelunking the deep: guaranteed queries on general neural implicit surfaces via range analysis. ACM Trans. Graph. (Proc. SIGGRAPH), 41(4), 2022. 2", + "[25] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3D shapes. In IEEE Conf. Comput. Vis. Pattern Recog., pages 11358-11367, 2021. 2", + "[26] Arnulph Fuhrmann, Gerrit Sobotka, and Clemens Groß. Distance fields for rapid collision detection in physically based modeling. In Proceedings of GraphiCon, volume 2003, pages 58-65, 2003. 2", + "[27] Eran Guendelman, Robert Bridson, and Ronald Fedkiw. Nonconvex rigid bodies with stacking. ACM TOG, 22(3):871-878, 2003.", + "[28] Miles Macklin, Kenny Erleben, Matthias Müller, Nuttapong Chentanez, Stefan Jeschke, and Zach Corse. Local optimization for robust signed distance field collision. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 3(1):1-17, 2020. 2", + "[29] Rohan Sawhney and Keenan Crane. Monte Carlo geometry processing: a grid-free approach to PDE-based methods on" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1284", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "volumetric domains. ACM Trans. Graph. (Proc. SIGGRAPH), 39(4), 2020. 2", + "[30] Albert Pumarola, Artsiom Sanakoyeu, Lior Yariv, Ali Thabet, and Yaron Lipman. Visco grids: Surface reconstruction with viscosity and coarea grids. Advances in Neural Information Processing Systems, 35:18060-18071, 2022. 2, 5", + "[31] Songyou Peng, Chiyu Jiang, Yiyi Liao, Michael Niemeyer, Marc Pollefeys, and Andreas Geiger. Shape as points: A differentiable poisson solver. Advances in Neural Information Processing Systems, 34:13032-13044, 2021.", + "[32] Chao Chen, Yu-Shen Liu, and Zhizhong Han. Gridpull: Towards scalability in learning implicit representations from 3d point clouds. In Proceedings of the IEEE/cvf international conference on computer vision, pages 18322-18334, 2023. 2", + "[33] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. DeepSDF: Learning continuous signed distance functions for shape representation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 165-174, 2019. 2", + "[34] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in neural information processing systems, 33:7462-7473, 2020. 2, 8, 5, 7", + "[35] Guillaume Coiffier and Louis Béthune. 1-Lipschitz neural distance fields. Comput. Graph. Forum (Eurographics STARs), 43(5):e15128, 2024. 2", + "[36] Yue Jiang, Dantong Ji, Zhizhong Han, and Matthias Zwicker. SDFDiff: Differentiable rendering of signed distance fields for 3D shape optimization. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1251-1261, 2020. 2", + "[37] Petr Kellnhofer, Lars C Jebe, Andrew Jones, Ryan Spicer, Kari Pulli, and Gordon Wetzstein. Neural lumigraph rendering. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4287-4297, 2021.", + "[38] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3D representations without 3D supervision. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3504-3515, 2020.", + "[39] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. In Advances in Neural Information Processing Systems, volume 33, pages 2492-2502, 2020. 2, 8", + "[40] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snavely. PhySG: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5453-5462, 2021.", + "[41] Sai Bangaru, Michael Gharbi, Tzu-Mao Li, Fujun Luan, Kalyan Sunkavalli, Milos Hasan, Sai Bi, Zexiang Xu, Gilbert Bernstein, and Fredo Durand. Differentiable rendering of neural sdfs through reparameterization. In SIGGRAPH Asia Conference Proceedings, 2022.", + "[42] Delio Vicini, Sébastien Speierer, and Wenzel Jakob. Differentiable signed distance function rendering. ACM Trans. Graph. (Proc. SIGGRAPH), 41(4):1-18, 2022. 2" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[43] Jingyang Zhang, Yao Yao, Shiwei Li, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Critical regularizations for neural surface reconstruction in the wild. In IEEE Conf. Comput. Vis. Pattern Recog., pages 6270-6279, 2022. 2", + "[44] Ruian Wang, Zixiong Wang, Yunxiao Zhang, Shuangmin Chen, Shiqing Xin, Changhe Tu, and Wenping Wang. Aligning gradient and hessian for neural signed distance function. In Advances in Neural Information Processing Systems, volume 36, 2024.", + "[45] Zixiong Wang, Yunxiao Zhang, Rui Xu, Fan Zhang, Peng-Shuai Wang, Shuangmin Chen, Shiqing Xin, Wenping Wang, and Changhe Tu. Neural-singular-hessian: Implicit neural representation of unoriented point clouds by enforcing singular hessian. ACM Transactions on Graphics (TOG), 42(6):1-14, 2023. 5, 7, 8", + "[46] Qiujie Dong, Rui Xu, Pengfei Wang, Shuangmin Chen, Shiqing Xin, Xiaohong Jia, Wenping Wang, and Changhe Tu. Neurcadrecon: Neural representation for reconstructing cad surfaces by enforcing zero gaussian curvature. arXiv preprint arXiv:2404.13420, 2024. 2, 7", + "[47] Yesom Park, Taekyung Lee, Jooyoung Hahn, and Myungjoo Kang. $p$ -poisson surface reconstruction in curl-free flow from point clouds. In Advances in Neural Information Processing Systems, volume 36, 2024. 2", + "[48] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In IEEE Conf. Comput. Vis. Pattern Recog., pages 8456-8465, 2023. 2", + "[49] Yaron Lipman. Phase transitions, distance functions, and implicit neural representations. In International Conference on Machine Learning, volume 139, pages 6702-6712, 2021. 2, 3, 6, 8, 9", + "[50] Angela Dai and Matthias Nießner. Neural poisson: Indicator functions for neural fields. arXiv preprint arXiv:2211.14249, 2022. 2", + "[51] Baorui Ma, Zhizhong Han, Yu-Shen Liu, and Matthias Zwicker. Neural-pull: Learning signed distance functions from point clouds by learning to pull space onto surfaces. arXiv preprint arXiv:2011.13495, 2020. 2, 7, 3, 5, 8", + "[52] Matan Atzmon and Yaron Lipman. SALD: Sign agnostic learning with derivatives. In ICLR, 2021. 2, 3", + "[53] Nicholas Sharp, Yousuf Soliman, and Keenan Crane. The vector heat method. ACM TOG, 38(3):1-19, 2019. 3", + "[54] Nicole Feng and Keenan Crane. A heat method for generalized signed distance. ACM Trans. Graph. (Proc. SIGGRAPH), 43(4), 2024. 3", + "[55] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 6, 7, 8, 5, 9", + "[56] Matthew Berger, Joshua A Levine, Luis Gustavo Nonato, Gabriel Taubin, and Claudio T Silva. A benchmark for surface reconstruction. ACM Transactions on Graphics (TOG), 32(2):1-17, 2013. 6, 8" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1285", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[57] David Stutz and Andreas Geiger. Learning 3d shape completion under weak supervision. International Journal of Computer Vision, 128:1162-1181, 2020. 6, 7", + "[58] Nasim Rahaman, Aristide Baratin, Devansh Arpit, Felix Draxler, Min Lin, Fred Hamprecht, Yoshua Bengio, and Aaron Courville. On the spectral bias of neural networks. In International Conference on Machine Learning, pages 5301-5310, 2019. 7, 8, 3, 5, 11", + "[59] Dmitry S Kulyabov, Anna V Korolkova, Tatiana R Velieva, and Migran N Gevorkyan. Numerical analysis of eikonal equation. In Saratov Fall Meeting 2018: Laser Physics, Photonic Technologies, and Molecular Modeling, volume 11066, pages 188-195. SPIE, 2019. 1", + "[60] Zhen-Hang Yang and Yu-Ming Chu. On approximating the modified bessel function of the second kind. Journal of inequalities and applications, 2017:1-8, 2017. 2", + "[61] Luciano Modica. The gradient theory of phase transitions and the minimal interface criterion. Archive for Rational Mechanics and Analysis, 98:123-142, 1987. 10", + "[62] Peter Sternberg. The effect of a singular perturbation on nonconvex variational problems. Archive for Rational Mechanics and Analysis, 101:209-260, 1988. 10" + ], + "bbox": [ + 78, + 90, + 467, + 401 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "1286", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/HotSpot_ Signed Distance Function Optimization with an Asymptotically Sufficient Condition/733f46a7-f78c-4e53-bba4-ead7eb705bb7_model.json b/2025/HotSpot_ Signed Distance Function Optimization with an Asymptotically Sufficient Condition/733f46a7-f78c-4e53-bba4-ead7eb705bb7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..dca730eaf690b031336c7fed97a9e8876db763fc --- /dev/null +++ b/2025/HotSpot_ Signed Distance Function Optimization with an Asymptotically Sufficient Condition/733f46a7-f78c-4e53-bba4-ead7eb705bb7_model.json @@ -0,0 +1,2994 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.119, + 0.13, + 0.853, + 0.175 + ], + "angle": 0, + "content": "HotSpot: Signed Distance Function Optimization with an Asymptotically Sufficient Condition" + }, + { + "type": "text", + "bbox": [ + 0.145, + 0.203, + 0.825, + 0.241 + ], + "angle": 0, + "content": "Zimo Wang* Cheng Wang* Taiki Yoshino Sirui Tao Ziyang Fu Tzu-Mao Li UC San Diego" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.256, + 0.889, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.405, + 0.893, + 0.476 + ], + "angle": 0, + "content": "Figure 1. We propose HOTSPOT, a neural signed distance function optimization method that establishes an asymptotic sufficient condition to guarantee convergence to a true distance function, enabling precise surface reconstruction and level set representation for complex shapes. Here we show a reconstruction from a point cloud sampled from the reference bunny (taken from Mehta et al. [1]) on the right. In the inset, we visualize the recovered signed distance function on a horizontal slice, using warm colors for positive values and cool for negative (zoom in for details). Our reconstruction is significantly more accurate than prior works (SAL [2], DiGS [3], and StEik [4])." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.497, + 0.314, + 0.513 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.529, + 0.473, + 0.816 + ], + "angle": 0, + "content": "We propose a method, HOTSPOT, for optimizing neural signed distance functions. Existing losses, such as the eikonal loss, act as necessary but insufficient constraints and cannot guarantee that the recovered implicit function represents a true distance function, even if the output minimizes these losses almost everywhere. Furthermore, the eikonal loss suffers from stability issues in optimization. Finally, in conventional methods, regularization losses that penalize surface area distort the reconstructed signed distance function. We address these challenges by designing a loss function using the solution of a screened Poisson equation. Our loss, when minimized, provides an asymptotically sufficient condition to ensure the output converges to a true distance function. Our loss also leads to stable optimization and naturally penalizes large surface areas. We present theoretical analysis and experiments on both challenging 2D and 3D datasets and show that our method provides better surface reconstruction and a more accurate distance approximation." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.821, + 0.21, + 0.836 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.846, + 0.471, + 0.877 + ], + "angle": 0, + "content": "Learning neural signed distance function requires designing loss functions so that the zero crossings of the implicit" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.498, + 0.892, + 0.604 + ], + "angle": 0, + "content": "function are at the desired locations (e.g., adhere to a sparse input point cloud), and that the implicit function returns the correct signed distance. Designing the loss function proves to be challenging for shapes with complex topology and details (Fig. 1). We propose HOTSPOT, a method based on a relation between the screened Poisson equation and distance that enjoys both theoretical and practical benefits." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.608, + 0.893, + 0.759 + ], + "angle": 0, + "content": "A major challenge in signed distance function optimization is ensuring the implicit function follows the true distance. A common regularization loss used is the eikonal equation, which constrains the gradient norm of an implicit function to be 1 almost everywhere. If the implicit function is a signed distance function, then it satisfies the eikonal equation. However, the converse is not true. That is, satisfying the eikonal equation is only a necessary condition for the signed distance function, not a sufficient condition [5]. Fig. 2 presents the logic and Fig. 3 shows an example." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.762, + 0.892, + 0.822 + ], + "angle": 0, + "content": "Another key challenge lies in the optimization. It is known that the eikonal equation regularization is unstable [4], and the instability can make optimization converge to suboptimal results or even diverge." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Furthermore, to address optimization ill-posedness and instability, existing methods penalize the surface area of the shape. We prove this degrades the reconstruction quality and distorts the output level sets, thereby leading to a delicate balance between over- and under-smoothing." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.206, + 0.9 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1276" + } + ], + [ + { + "type": "equation", + "bbox": [ + 0.084, + 0.09, + 0.465, + 0.143 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{c} \\min \\{\\text {h e a t l o s s} \\} \\\\ \\lambda \\to \\infty \\end{array} \\Leftrightarrow \\begin{array}{l} \\text {D i s t a n c e} \\quad \\text {a . e .} \\\\ \\text {F u n c t i o n} \\quad \\stackrel {{\\Rightarrow}} {{\\Leftarrow}} \\end{array} \\left\\{ \\begin{array}{l} \\text {E i k o n a l c o n d i t i o n} \\\\ \\text {Z e r o g r a d i e n t d i v e r g e n c e} \\\\ \\text {S i n g u l a r H e s s i a n m a t r i x} \\\\ \\dots \\end{array} \\right.\n\\]" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.147, + 0.471, + 0.258 + ], + "angle": 0, + "content": "Figure 2. Existing constraints are only necessary but insufficient conditions. That is, there exist solutions, like Fig. 3, that satisfy them almost everywhere yet fail to be distance functions. In contrast, we use a heat loss, modeled by a screened Poisson equation with parameter \\(\\lambda\\). As the parameter \\(\\lambda\\) goes to infinity, the minimizer converges to the distance function and excludes other solutions. Thus, our method is asymptotically both necessary and sufficient." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.273, + 0.47, + 0.394 + ], + "angle": 0, + "content": "We propose a simple model that can address all the challenges above. Our model, HOTSPOT, is based on a screened Poisson equation and a classical relation between heat transfer and distance [6] that was popular for approximating both Euclidean and geodesic distance [7-9]. Using this relation, we design a loss function for optimizing neural signed distance functions. As our contributions, we show, both theoretically and empirically, that our model" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.394, + 0.47, + 0.453 + ], + "angle": 0, + "content": "- serves as a sufficient condition ensuring the difference between the minimizer and the actual distance is bounded and decreases linearly as the parameter \\(\\lambda\\) increases (Fig. 2 and Fig. 3, right)." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.454, + 0.47, + 0.529 + ], + "angle": 0, + "content": "- is stable both spatially and temporally, in the sense that a slight perturbation of the implicit function only introduces changes in a local region, and that the dynamical system formed by the gradient flow will converge in the long run." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.53, + 0.469, + 0.559 + ], + "angle": 0, + "content": "- penalizes the surface area naturally, while remaining an undistorted signed distance function." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.56, + 0.469, + 0.606 + ], + "angle": 0, + "content": "- works well in both 2D and 3D shapes, and it excels at approaching signed distance functions while optimizing for complex, high-genus shapes." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.394, + 0.47, + 0.606 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.619, + 0.22, + 0.635 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.469, + 0.75 + ], + "angle": 0, + "content": "Implicit surfaces. Modeling and reconstructing surfaces using implicit functions has a long history in computer vision and graphics [11-15]. Recently, there is a surge of interest to model surfaces using neural implicit functions [16-20]. Neural implicit representations are favored for their compactness, ability to represent detailed surfaces, and compatibility to gradient-based optimization [1]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Neural signed distance functions. Among the classes of implicit surfaces, signed distance functions, where the implicit function outputs the distance to the closest point on the surface, and the sign represents whether the point is inside or outside the solid object, is one of the most popular variants. Signed distance functions enable an efficient rendering algorithm called sphere tracing [21-24] and helps geometry compression [25]. They can also be used for collision detection [26-28] and numerical simulation [29]. Traditionally, signed distance functions are often represented" + }, + { + "type": "image", + "bbox": [ + 0.551, + 0.09, + 0.837, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.345, + 0.892, + 0.457 + ], + "angle": 0, + "content": "Figure 3. We illustrate a 1D neural signed distance function optimization using the classical eikonal loss [10] and our model. The \\( x \\) axis is the domain and the \\( y \\) axis shows the output of the implicit function. The middle and bottom rows show the intermediate and final states of the optimization respectively. The eikonal loss, as a necessary but insufficient condition, is incapable of converging to the actual signed distance function (dashed line), even when the function satisfies the eikonal equation almost everywhere." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.471, + 0.892, + 0.547 + ], + "angle": 0, + "content": "as grids [30-32]. More recently, neural signed distance functions have become an appealing geometric representation [33-35]. (Neural) Signed distance functions are also often used in inverse rendering [36-42], for their flexibility to obtain surfaces with complex topology." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.55, + 0.893, + 0.792 + ], + "angle": 0, + "content": "Regularizing and initializing neural signed distance functions. A commonly used loss for optimizing neural signed distance functions is the eikonal regularization, which encourages the norm of the gradient of the implicit function to be one [10, 39]. However, the eikonal regularization alone is far from sufficient, as 1) minimizing eikonal loss does not necessarily lead to a distance field [5, 30], and 2) its optimization is unstable [4]. Previous works have proposed initialization to bootstrap the optimization with a simple shape [3, 10]. Recent works show that minimizing the divergence and the surface area of the implicit function can lead to better reconstruction [3, 43] and stabilizes the optimization [4]. Second-order information is also often used [43-46] for regularization. All losses above are based on necessary properties of real distance but neither establish convergence nor exclude non-SDF solutions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Specialized architectures have been proposed to limit the Lipschitz constant of the implicit function [35], satisfy \\( p \\)-Poisson equation [47], or represent high-frequency details [25, 34, 48]. Some research [49, 50] has shown a link between occupancy function optimization and an approximate signed distance field regularization. Our experiments further reveal that some implementations [2, 51, 52]" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1277" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.121 + ], + "angle": 0, + "content": "optimizing SDFs with the closest point may struggle when querying near the interpolated surface." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.122, + 0.469, + 0.226 + ], + "angle": 0, + "content": "We propose a new loss and method for optimizing signed distance functions. While derived from very different mathematics, our loss turns out to have a close connection to Lipman's PHASE loss [49]. We show that our theory leads to significantly more numerically stable optimization, and a different conclusion in the parameter setting that leads to drastically different results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.228, + 0.469, + 0.333 + ], + "angle": 0, + "content": "Distance approximation using heat transfer. Our derivation is inspired by a relation between heat transfer and approximated distance [6]. This relation has been used in computer graphics for efficiently computing the geodesic distance on surfaces [7-9, 53, 54]. We apply the relation for optimizing neural signed distance functions, and analyze the properties of the resulting optimization in this context." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.346, + 0.339, + 0.362 + ], + "angle": 0, + "content": "3. Background and Motivation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.371, + 0.47, + 0.432 + ], + "angle": 0, + "content": "Given an input point cloud \\(\\pmb{x}_i\\in \\mathbb{R}^d\\) (we focus on 2D and 3D in this work) with \\(N\\) points without normal information that lies on an unknown surface \\(\\Gamma\\) of a solid object \\(S\\), our goal is to find a signed distance function \\(f(\\pmb {x})\\), such that" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.441, + 0.469, + 0.481 + ], + "angle": 0, + "content": "\\[\nf (\\boldsymbol {x}) = \\left\\{ \\begin{array}{l l} - d _ {\\Gamma} (\\boldsymbol {x}) & \\text {i f} \\boldsymbol {x} \\in S \\\\ d _ {\\Gamma} (\\boldsymbol {x}) & \\text {o t h e r w i s e ,} \\end{array} \\right. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.489, + 0.47, + 0.596 + ], + "angle": 0, + "content": "where \\(d_{\\Gamma}(\\pmb{x})\\) denotes the closest distance from point \\(\\pmb{x}\\) to the surface \\(\\Gamma\\). In practice, we approximate the signed distance function \\(f(\\pmb{x})\\) using a neural network \\(u(\\pmb{x}): \\mathbb{R}^d \\to \\mathbb{R}\\) parameterized by its weights and use gradient-based optimization to figure out the weights. Previous work (e.g., [10]) often uses a linear combination of a boundary loss \\(L_{\\mathrm{boundary}}\\) and a eikonal loss \\(L_{\\mathrm{eikonal}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.099, + 0.605, + 0.47, + 0.645 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {b o u n d a r y}} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} | u (\\boldsymbol {x} _ {i}) | ^ {p}, \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.649, + 0.469, + 0.681 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {e i k o n a l}} = \\int_ {\\Omega} | \\| \\nabla u (\\boldsymbol {x}) \\| - 1 | ^ {p} \\mathrm {d} \\boldsymbol {x}, \\quad p = 1 \\text {o r} 2, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.469, + 0.78 + ], + "angle": 0, + "content": "where \\(\\Omega \\subset \\mathbb{R}^d\\) is our domain of interest. The boundary loss encourages the implicit function \\(u\\) to output 0 at the boundary, and the eikonal loss is based on that observation that if a function \\(f\\) is a distance function, the \\(L^2\\) norm of its gradient \\(\\| \\nabla f(\\pmb{x})\\| = 1\\) holds almost everywhere. Thus the approximation \\(u\\) is encouraged to behave the same." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.469, + 0.901 + ], + "angle": 0, + "content": "However, as noted, simply optimizing for the two terms above does not guarantee that \\( u \\) would be a distance function even when satisfied almost everywhere (Figs. 2 and 3), and the extra regularization previous work added to minimize the area and divergence [3, 4] could further tamper with the distance approximation quality. We derive an additional loss that can be used together with both of the losses above, and analyze our loss theoretically and empirically." + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.089, + 0.695, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.209, + 0.692, + 0.221 + ], + "angle": 0, + "content": "(a) 1D Heat Simulation via Eq. (4)" + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.089, + 0.887, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.72, + 0.209, + 0.868, + 0.221 + ], + "angle": 0, + "content": "(b) Distance Reconstruction" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.223, + 0.594, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.299, + 0.58, + 0.31 + ], + "angle": 0, + "content": "(c) HOUSE" + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.222, + 0.695, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.62, + 0.299, + 0.674, + 0.31 + ], + "angle": 0, + "content": "(d) \\(\\lambda = 1\\)" + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.222, + 0.793, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.717, + 0.299, + 0.77, + 0.31 + ], + "angle": 0, + "content": "(e) \\(\\lambda = 5\\)" + }, + { + "type": "image", + "bbox": [ + 0.795, + 0.222, + 0.888, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.811, + 0.299, + 0.869, + 0.31 + ], + "angle": 0, + "content": "(f) \\(\\lambda = 50\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.316, + 0.892, + 0.428 + ], + "angle": 0, + "content": "Figure 4. We show an illustration of the relation between screened Poisson equation and a distance field. The top shows a 1D example, where on the left we show a solution to Eq. (4) with different absorption \\(\\lambda\\), and on the right we show the reconstructed distance \\(-\\frac{1}{\\lambda}\\ln(h)\\) (Eq. (5)). Boundary loss for real boundary points acts as an isothermal heat source, while heat loss diffuses the heat. When \\(\\lambda\\) is increased, the heat decays faster, and the error of the reconstructed function approaches 0. The bottom shows a 2D example." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.438, + 0.892, + 0.513 + ], + "angle": 0, + "content": "We will be using a key relation between the screen Poisson equation and the distance. We first define a screened Poisson equation over a heat field \\( h(\\pmb{x}): \\mathbb{R}^d \\to \\mathbb{R} \\), with an absorption coefficient \\( \\lambda \\), and set a Dirichlet boundary condition to be 1 on our target surface \\( \\Gamma \\):" + }, + { + "type": "equation", + "bbox": [ + 0.549, + 0.52, + 0.891, + 0.568 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\begin{array}{l}\\nabla^ {2} h (\\boldsymbol {x}) - \\lambda^ {2} h (\\boldsymbol {x}) = 0 \\quad \\forall \\boldsymbol {x} \\in \\mathbb {R} ^ {d} \\backslash \\Gamma\\\\\\lim _ {| | \\boldsymbol {x} | | \\rightarrow \\infty} h (\\boldsymbol {x}) = 0, \\quad h (\\boldsymbol {x}) = 1 \\quad \\forall \\boldsymbol {x} \\in \\Gamma .\\end{array}\\right. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.573, + 0.89, + 0.604 + ], + "angle": 0, + "content": "If a heat field \\( h_\\lambda \\) is a solution to the partial differential equation above, then the following equation holds [6]:" + }, + { + "type": "equation", + "bbox": [ + 0.594, + 0.61, + 0.891, + 0.639 + ], + "angle": 0, + "content": "\\[\n\\lim _ {\\lambda \\rightarrow \\infty} \\frac {1}{\\lambda} \\ln \\left(h _ {\\lambda} (\\boldsymbol {x})\\right) = - d _ {\\Gamma} (\\boldsymbol {x}). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Fig. 4 shows the intuition. The screened Poisson equation simulates heat transfer until an equilibrium is reached, under absorption controlled by the \\(\\lambda\\) coefficient. As we increase the absorption, the reconstructed distance \\(\\frac{1}{\\lambda}\\ln (h_{\\lambda})\\) would approach the real distance. In practice, we could only obtain a discrete point set \\(\\Gamma_0\\). Provided that the neighbor points are close enough, if there are some boundaries connecting them, the boundary condition of Eq. (4) still holds. This can be interpreted as a superset \\(\\Gamma\\) replacing \\(\\Gamma_0\\), effectively transforming the boundary from a discrete point cloud into a continuous surface, which is required and precisely the goal of the signed distance function reconstruction. Obtaining a distance function to the point cloud rather than the interpolated surface is not our objective. The reconstructed distance can still faithfully represent the distance to \\(\\Gamma\\) because the approximation in Eq. (5) still holds. Later, we will prove that the measure of \\(\\Gamma\\) is bounded by our loss." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1278" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.09, + 0.17, + 0.105 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.115, + 0.473, + 0.283 + ], + "angle": 0, + "content": "In this section, we introduce our method, HOTSPOT, based on the screened Poisson equation (Eq. (4)) and derive a theoretical analysis. We use the screened Poisson equation to derive a loss we can directly apply to an implicit function \\( u \\), and prove that there is a bound to the solution of the screened Poisson equation to the true distance (Section 4.1). We further derive a stability analysis of our loss on both the spatial stability over small perturbations and the temporal convergence (Section 4.2). Next, we show that our loss can penalize surfaces with large area (Section 4.3). Finally, we discuss the relation to prior work, PHASE (Section 4.4)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.29, + 0.471, + 0.322 + ], + "angle": 0, + "content": "4.1. Screened Poisson Equation Informed Signed Distance Functions Optimization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.328, + 0.47, + 0.39 + ], + "angle": 0, + "content": "Given an implicit function \\( u \\), we want to derive a loss based on the screened Poisson equation (Eq. (4)) and its relation to the distance (Eq. (5)). We achieve this by applying the following substitution based on Eq. (5):" + }, + { + "type": "equation", + "bbox": [ + 0.21, + 0.397, + 0.47, + 0.417 + ], + "angle": 0, + "content": "\\[\nh (\\boldsymbol {x}) = e ^ {- \\lambda | u (\\boldsymbol {x}) |}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.425, + 0.47, + 0.501 + ], + "angle": 0, + "content": "We want to design a loss such that its minimizer satisfies the screened Poisson equation. This can be achieved by minimizing the energy functional \\(\\frac{1}{2}\\int_{\\Omega} \\| \\nabla h(\\pmb{x}) \\|^2 + \\lambda^2 h(\\pmb{x})^2 \\, \\mathrm{d}\\pmb{x}\\). After substitution and removing of the constant factor \\(\\lambda\\), we obtain our heat loss:" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.508, + 0.47, + 0.541 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {h e a t}} = \\frac {1}{2} \\int_ {\\Omega} e ^ {- 2 \\lambda | u (\\boldsymbol {x}) |} \\left(\\| \\nabla u (\\boldsymbol {x}) \\| ^ {2} + 1\\right) d \\boldsymbol {x}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.548, + 0.47, + 0.625 + ], + "angle": 0, + "content": "If we take the derivative of \\(L_{\\mathrm{heat}}\\) with respect to the heat field \\(h\\), we recover the screened Poisson equation \\(\\nabla^2 h - \\lambda^2 h = 0\\). The boundary condition can be enforced by the standard boundary loss (Eq. (2)), by observing that \\(u(\\pmb{x}) = 0\\) implies \\(h(\\pmb{x}) = 1\\). Together with the eikonal loss, we optimize:" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.634, + 0.469, + 0.651 + ], + "angle": 0, + "content": "\\[\nL = w _ {b} L _ {\\text {b o u n d a r y}} + w _ {e} L _ {\\text {e i k o n a l}} + w _ {h} L _ {\\text {h e a t}}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.47, + 0.781 + ], + "angle": 0, + "content": "While prior work [6] has studied the limiting behavior of the absorption coefficient \\(\\lambda\\), to our knowledge, the bounds of the approximation and the convergence speed remain unknown. In the supplementary material A.3, we prove that, for a boundary \\(\\Gamma\\) formed by a discrete set of points and a small volume around them, the distance approximation obtained from the solution of the screened Poisson equation converges linearly with respect to \\(\\lambda\\):" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.788, + 0.47, + 0.819 + ], + "angle": 0, + "content": "\\[\n\\frac {1}{\\lambda} C _ {\\text {l o w e r}} (\\boldsymbol {x}) \\leqslant \\frac {1}{\\lambda} \\ln \\left(h _ {\\lambda} (\\boldsymbol {x})\\right) + d _ {\\Gamma} (\\boldsymbol {x}) \\leqslant \\frac {1}{\\lambda} C _ {\\text {u p p e r}} (\\boldsymbol {x}). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.871 + ], + "angle": 0, + "content": "Here, both \\( C_{\\mathrm{upper}}(\\pmb{x}) \\) and \\( C_{\\mathrm{lower}}(\\pmb{x}) \\) are scalar fields independent of \\( \\lambda \\). The inequalities are tight, which means that one can find a case where the equality holds." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Compared to the eikonal loss (Eq. (3)), where the distance error can be unbounded even when the equation holds" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.089, + 0.888, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.212, + 0.895, + 0.296 + ], + "angle": 0, + "content": "Figure 5. We visualize the direction of the changes of the implicit function during one iteration of optimization. Left shows the result with only the eikonal loss and right shows the result after adding our loss. Our loss helps to escape the local minimum in Fig. 3. The heat diffusion encourages the reconstructed absolute distance to increase monotonically away from the boundary." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.31, + 0.895, + 0.476 + ], + "angle": 0, + "content": "almost everywhere, Eq. (9) shows that even with a finite absorption \\(\\lambda\\), our loss can lead to much more accurate distance. In fact, any loss that only depends on the norm of the gradients \\(\\|\\nabla u\\|\\) is not a sufficient condition of the distance function (Fig. 2), due to the discontinuous jumps of the gradient. Adding our loss (Eq. (7)) ensures that the absolute value of the implicit function converges to the actual distance. Fig. 5 visualizes the direction of how an implicit function would evolve when using only the eikonal loss and after adding our loss. Our loss ensures that the implicit function evolves towards the actual distance." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.488, + 0.673, + 0.505 + ], + "angle": 0, + "content": "4.2. Stability Analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.512, + 0.894, + 0.618 + ], + "angle": 0, + "content": "Our heat loss enjoys other theoretical benefits apart from the convergence to an actual distance. Here, we show that our loss leads to easier optimization. We separately analyze spatial stability and temporal stability, where the spatial stability studies the effect of small perturbation error added to the corresponding partial differential equation, and the temporal stability studies the convergence of the gradient flow." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.638, + 0.667, + 0.654 + ], + "angle": 0, + "content": "4.2.1 Spatial Stability" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.663, + 0.892, + 0.782 + ], + "angle": 0, + "content": "We show that for the screened Poisson equation, an error introduced to a single point in the solution decays exponentially and does not propagate to infinitely distant positions. On the contrary, for the eikonal equation, the error can propagate to infinitely far. The property of the screened Poisson equation ensures that the perturbation introduced during optimization does not lead to a drastically different target and stabilizes the optimization." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.784, + 0.892, + 0.83 + ], + "angle": 0, + "content": "First, we show that for the eikonal equation, an additive error introduced to a point can presist along a line segment or a ray." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Proposition 1. Given a field \\( u(\\pmb{x}) \\) that satisfies the eikonal equation \\( \\| \\nabla u \\| = 1 \\), we introduce an additive error \\( u_{0e} \\) at the point \\( \\pmb{x}_0 \\) where \\( u(\\pmb{x}_0) = u_0 \\) to obtain a perturbed field \\( u' \\) where \\( u'(\\pmb{x}_0) = u_0 + u_{0e} \\) and \\( \\| \\nabla u' \\| = 1 \\). In the" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1279" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.152 + ], + "angle": 0, + "content": "perturbed field \\( u' \\), there exists a line segment parameterized by \\( \\pmb{x}(s) \\) starting from \\( \\pmb{x}_0 \\) and \\( s \\in [0, M] \\), where \\( M > 0 \\) and can be infinite, such that every point along the line has the same error:" + }, + { + "type": "equation", + "bbox": [ + 0.183, + 0.161, + 0.47, + 0.179 + ], + "angle": 0, + "content": "\\[\nu ^ {\\prime} (\\boldsymbol {x} (s)) - u (\\boldsymbol {x} (s)) = u _ {0 e} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.187, + 0.47, + 0.247 + ], + "angle": 0, + "content": "The boundary points can contain errors when approximating the true boundaries, and these errors remain constant along the line segment. As a result, the propagated error can become significant and potentially detrimental." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.247, + 0.469, + 0.278 + ], + "angle": 0, + "content": "Next, we demonstrate that the error field \\( h_e \\) in Eq. (4) decays exponentially from the perturbed point \\( x_0 \\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.285, + 0.47, + 0.389 + ], + "angle": 0, + "content": "Proposition 2. In 3D space, consider an additive error \\( h_{0e} \\) within a small ball \\( B(\\pmb{x}_0,\\epsilon) \\) around the point \\( \\pmb{x}_0 \\), where the original field \\( h(\\pmb{x}) \\) and the perturbed field \\( h^{\\prime}(\\pmb{x}) \\), \\( \\pmb{x} \\in \\mathbb{R}^n\\backslash B \\) both satisfy the screened Poisson equation in Eq. (4). The resulting error field \\( h_e = h' - h \\) is radially symmetric with respect to \\( x_0 \\) within some maximum radius \\( R \\), and is given by" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.389, + 0.469, + 0.415 + ], + "angle": 0, + "content": "\\[\nh _ {e} (r) = \\frac {\\epsilon}{r} h _ {0 e} e ^ {\\lambda (\\epsilon - r)}, \\forall r > \\epsilon , \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.418, + 0.423, + 0.435 + ], + "angle": 0, + "content": "where \\(r = \\left\\| \\boldsymbol{x} - \\boldsymbol{x}_0 \\right\\|\\) is the radial distance from \\(\\boldsymbol{x}_0\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.442, + 0.469, + 0.517 + ], + "angle": 0, + "content": "The proofs of these two propositions are in the supplementary material A.1 and A.2. By comparing the error fields, we observe that the partial differential equation governed by the screened Poisson equation exhibits greater spatial robustness than the one based on the eikonal equation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.535, + 0.262, + 0.55 + ], + "angle": 0, + "content": "4.2.2 Temporal Stability" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.558, + 0.469, + 0.634 + ], + "angle": 0, + "content": "We next demonstrate the temporal stability of the heat loss and show that the dynamical system of the gradient update is stable, that is, the solution of the system converges as the time \\( t \\) approaches infinity. This is supported by von Neumann stability analysis, similar to Yang et al.'s work [4]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.634, + 0.469, + 0.664 + ], + "angle": 0, + "content": "Yang et al. showed that for the eikonal loss (Eq. (3)), the temporal gradient flow of the optimization is:" + }, + { + "type": "equation", + "bbox": [ + 0.099, + 0.672, + 0.469, + 0.75 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\frac {\\partial u}{\\partial t} = - \\frac {\\delta L _ {\\mathrm {e i k o n a l}}}{\\delta u} = \\nabla \\cdot (\\kappa_ {e} \\nabla u) \\approx \\kappa_ {e} \\nabla^ {2} u, \\text {w h e r e} \\\\ \\kappa_ {e} (\\boldsymbol {x}) = \\left\\{ \\begin{array}{l l} \\frac {1}{| \\nabla u (\\boldsymbol {x}) |} \\operatorname {s g n} (| \\nabla u (\\boldsymbol {x}) | - 1) & p = 1 \\\\ 1 - | \\nabla u (\\boldsymbol {x}) | ^ {- 1} & p = 2 \\end{array} \\right. \\tag {12} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.751, + 0.468, + 0.766 + ], + "angle": 0, + "content": "They analyze the stability by applying Fourier transform:" + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.773, + 0.469, + 0.803 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial \\hat {u}}{\\partial t} (t, \\omega) = - \\kappa_ {e} | \\omega | ^ {2} \\hat {u} (t, \\omega) \\Rightarrow \\hat {u} (t, \\omega) \\propto e ^ {- \\kappa_ {e} | \\omega | ^ {2} t}, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.81, + 0.469, + 0.901 + ], + "angle": 0, + "content": "where \\(\\omega = (\\omega_{1},\\dots ,\\omega_{n})\\) denotes the frequency variable and \\(\\hat{u}\\) represents the Fourier transform of \\(u\\). In the frequency domain, if the original function converges, the term \\(\\frac{\\partial\\hat{u}}{\\partial t}\\) should also converge. However, when \\(\\kappa_{e} < 0\\), this process corresponds to backward diffusion, which diverges and is inherently unstable regardless of the numerical implementation." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "We extend the analysis to our heat loss (Eq. (7)). The gradient flow is" + }, + { + "type": "equation", + "bbox": [ + 0.595, + 0.129, + 0.892, + 0.16 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial h}{\\partial t} = - \\frac {\\delta L _ {\\text {h e a t}}}{\\delta h} = \\nabla^ {2} h - \\lambda^ {2} h \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.166, + 0.892, + 0.21 + ], + "angle": 0, + "content": "This is exactly the heat equation, and is known to be stable. We can further verify the stability by applying Fourier transform to both sides:" + }, + { + "type": "equation", + "bbox": [ + 0.501, + 0.218, + 0.892, + 0.261 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial \\hat {h}}{\\partial t} (t, \\omega) = - \\left(| \\omega | ^ {2} + \\lambda^ {2}\\right) \\hat {h} (t, \\omega) \\Rightarrow \\hat {h} (t, \\omega) \\propto e ^ {- \\left(| \\omega | ^ {2} + \\lambda^ {2}\\right) t} \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.262, + 0.892, + 0.352 + ], + "angle": 0, + "content": "This analysis reveals that the original process \\( h(\\pmb{x}, t) \\) converges over time and remains stable. Due to the constraint \\( |u| = -\\ln(h) / \\lambda \\), the term \\( |u(\\pmb{x}, t)| \\) also converges during the optimization process. Since \\( u \\) is continuous, it converges as well. Thus, our method exhibits greater spatial and temporal stability compared to the eikonal equation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.36, + 0.756, + 0.377 + ], + "angle": 0, + "content": "4.3. Surface Area Regularization" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.384, + 0.892, + 0.474 + ], + "angle": 0, + "content": "We next show that our heat loss (Eq. (7)) penalizes large area of the reconstructed surfaces, by showing its relation to the coarea loss used by Pumarola et al. [30]. Pumarola et al. showed that the loss \\(\\int_{\\Omega} e^{-\\lambda |u|} \\| \\nabla u \\| \\, \\mathrm{d}\\pmb{x}\\) approximates the surface area of the zero level set of the implicit function \\(u\\). We then form the following inequality:" + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.482, + 0.892, + 0.557 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\int_ {\\Omega} e ^ {- \\lambda | u |} \\| \\nabla u \\| d \\boldsymbol {x} < \\int_ {\\Omega} \\sqrt {e ^ {- 2 \\lambda | u |} \\left(\\| \\nabla u \\| ^ {2} + 1\\right)} d \\boldsymbol {x} (16) \\\\ \\leqslant \\sqrt {| \\Omega |} \\cdot \\sqrt {\\int_ {\\Omega} e ^ {- 2 \\lambda | u |} \\left(\\| \\nabla u \\| ^ {2} + 1\\right) d x} (16) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.565, + 0.892, + 0.594 + ], + "angle": 0, + "content": "Therefore, minimizing the right hand side (our heat loss) leads to a reduction in the upper bound of the left hand side." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.595, + 0.892, + 0.67 + ], + "angle": 0, + "content": "Furthermore, we can show that the coarea loss is too aggressive for our purpose of signed distance field reconstruction. If we directly minimize the coarea loss (the left hand side in Eq. (16)), according to Euler-Lagrange equation, the minimizer would satisfy:" + }, + { + "type": "equation", + "bbox": [ + 0.648, + 0.679, + 0.891, + 0.695 + ], + "angle": 0, + "content": "\\[\n1 - \\nabla^ {2} u = 0, \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.705, + 0.892, + 0.871 + ], + "angle": 0, + "content": "which does not relate to the regularization parameter \\(\\lambda\\). Unfortunately, this partial differential equation system formed by the coarea loss above has a unique solution under Dirichlet boundary condition, and the gradient of its solution does not have a unit norm. This makes the coarea loss unable to accurately shape a signed distance function. By contrast, our heat loss has bounds for its distance approximation and converges to the true distance as we increase the parameter \\(\\lambda\\) (Eq. (9)). Another simple way to verify our convergence and the compatibility with the eikonal equation is to rewrite Eq. (4) in terms of \\(u\\) following Eq. (6):" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.874, + 0.891, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\left| \\left| \\nabla u \\right| \\right| ^ {2} - 1 - \\frac {\\operatorname {s g n} (\\mathrm {u})}{\\lambda} \\nabla^ {2} u = 0. \\tag {18}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1280" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.155, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.104, + 0.149, + 0.132, + 0.156 + ], + "angle": 0, + "content": "(a) GT" + }, + { + "type": "image", + "bbox": [ + 0.157, + 0.091, + 0.227, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.165, + 0.149, + 0.218, + 0.157 + ], + "angle": 0, + "content": "(b) \\(w_{h} = 0.1\\)" + }, + { + "type": "image", + "bbox": [ + 0.228, + 0.091, + 0.3, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.238, + 0.149, + 0.292, + 0.157 + ], + "angle": 0, + "content": "(c) \\(w_{h} = 0.2\\)" + }, + { + "type": "image", + "bbox": [ + 0.303, + 0.091, + 0.373, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.311, + 0.149, + 0.365, + 0.157 + ], + "angle": 0, + "content": "(d) \\(w_{h} = 0.3\\)" + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.091, + 0.454, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.385, + 0.149, + 0.437, + 0.157 + ], + "angle": 0, + "content": "(e) \\(w_{b} = 0.5\\)" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.162, + 0.155, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.096, + 0.22, + 0.14, + 0.229 + ], + "angle": 0, + "content": "(f) \\(w_{b} = 1\\)" + }, + { + "type": "image", + "bbox": [ + 0.157, + 0.162, + 0.227, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.168, + 0.22, + 0.215, + 0.229 + ], + "angle": 0, + "content": "(g) \\(w_{b} = 2\\)" + }, + { + "type": "image", + "bbox": [ + 0.228, + 0.162, + 0.3, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.241, + 0.22, + 0.288, + 0.229 + ], + "angle": 0, + "content": "(h) \\(w_{b} = 5\\)" + }, + { + "type": "image", + "bbox": [ + 0.302, + 0.162, + 0.373, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.312, + 0.22, + 0.362, + 0.229 + ], + "angle": 0, + "content": "(i) \\(w_{b} = 10\\)" + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.162, + 0.454, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.386, + 0.22, + 0.435, + 0.229 + ], + "angle": 0, + "content": "(j) \\(w_{b} = 20\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.231, + 0.47, + 0.328 + ], + "angle": 0, + "content": "Figure 6. We reproduce the PHASE model [49] and show the log-transformed occupancy field, i.e., the optimized signed distance function. PHASE recommends boundary weight range \\( w_{b} \\in (0.10, 0.32) \\) based on their \\( \\Gamma \\)-convergence theory. But in practice, only a stronger boundary weight constrains the boundary to follow the input points. We found that this phenomenon often occurs and show more examples in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.341, + 0.296, + 0.358 + ], + "angle": 0, + "content": "4.4. Relation to PHASE [49]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.364, + 0.469, + 0.439 + ], + "angle": 0, + "content": "While derived from different mathematical principles and motivations, our method is closely related to the PHASE model proposed by Lipman. It is motivated by smoothly approximating an occupancy function \\( o(x) \\), which is defined as 1 outside, -1 inside, and 0 on the surface." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.44, + 0.469, + 0.56 + ], + "angle": 0, + "content": "In the supplement, we show that there is a close relation between our heat field \\( h \\) and the smoothed occupancy function \\( o \\). We also show that our theory and derivation provide additional insights and significant differences in implementation and results. In particular, PHASE's theory introduces a condition stating that the boundary loss weight \\( w_{b} \\) should be small, whereas our theory emphasizes the need for a firm boundary condition (Eq. (4) and Fig. 6)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.561, + 0.469, + 0.621 + ], + "angle": 0, + "content": "We further found that optimizing occupancy field in the PHASE model, when used together with the eikonal loss, can lead to unbounded network weight derivatives, impeding the optimization. Our method avoids this issue." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.634, + 0.21, + 0.651 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.469, + 0.78 + ], + "angle": 0, + "content": "We experiment on both 2D and 3D datasets. We construct a 2D dataset containing 14 shapes, ranging from simple geometric shapes to complex nested structures. We also evaluate on two 3D benchmarks: a subset of ShapeNet [55] and the Surface Reconstruction Benchmark (SRB) [56]. Additionally, we generate a dataset with high-genus geometries from Mehta et al. [1]. In all of our experiments, we use only the position information of the point cloud." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.469, + 0.901 + ], + "angle": 0, + "content": "We report the Intersection over Union (IoU), Chamfer distance, and Hausdorff distance between reconstructed shapes and ground truth shapes as metrics for surface reconstruction. In addition, we report Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and Symmetric Mean Absolute Percentage Error (SMAPE) as metrics for distance field evaluation. We further evaluate non-zero level set quality near the surface, and render the zero level" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.856, + 0.107 + ], + "angle": 0, + "content": "set of the signed distance functions via sphere tracing." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.107, + 0.892, + 0.183 + ], + "angle": 0, + "content": "We use a network with 5 hidden layers and 128 hidden channels for our method in all experiments. The hyperparameters are kept consistent within each dataset. We additionally apply importance sampling to our heat loss integral. Full training details can be found in the supplement." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.192, + 0.62, + 0.206 + ], + "angle": 0, + "content": "5.1.2D Dataset" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.215, + 0.892, + 0.367 + ], + "angle": 0, + "content": "An overview of our 2D dataset is attached in the supplementary material. For 2D experiments, we simply use a linear network and geometric initialization, as in DiGS [3]. We compare our method with DiGS [3] and StEik [4] with the hyperparameters and training configuration provided in their released code, and report the surface reconstruction metrics and distance field metrics in Table 1. Our method outperforms theirs across all metrics with only 8,192 sample points, whereas both DiGS [3] and StEik [4] use 15,000 sample points based on their default settings." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.376, + 0.766, + 0.392 + ], + "angle": 0, + "content": "5.2. Ablation Study on 2D Dataset" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.399, + 0.892, + 0.474 + ], + "angle": 0, + "content": "To more comprehensively compare over different losses, we conduct an ablation study on our 2D dataset using the same network architecture, parameter initialization, and sampling strategy. The only difference lies in the loss functions used. Quantitative results are shown in Table 2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.476, + 0.892, + 0.703 + ], + "angle": 0, + "content": "The last row corresponds to our HOTSPOT model, combining boundary, eikonal, and heat losses. Our method reconstructs the correct topology for all 14 shapes, while others struggle with the harder ones. It outperforms across nearly all metrics, demonstrating superior reconstruction quality and distance query accuracy. Although HOTSPOT slightly trails SAL [2] in RMSE and MAE, it captures the complex topology where SAL fails. A visualization is shown in Fig. 7. Full qualitative results and detailed experiment parameters are provided in the supplementary material. Models relying heavily on area losses [3, 4] often face a trade-off between preserving original details and eliminating redundant boundaries, which distorts the distance field (Section 4.3). Moreover, methods without our heat loss fall in local optimum easily." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.711, + 0.628, + 0.726 + ], + "angle": 0, + "content": "5.3. 3D Datasets" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We evaluate our method on a processed subset [16, 19, 57] of ShapeNet [55], which provides surface points sampled on preprocessed watertight meshes for 260 shapes across 13 categories. To ensure a reasonable spatial density for passing the heat, we scale the point cloud so that \\(70\\%\\) of the points lie within a sphere centered at the origin with a radius of 0.45. For a fair comparison, we transform our outcomes back into the coordinate system used in StEik [4] and DiGS [3]. Additionally, we implement a scheduler to gradually increase the absorption parameter \\(\\lambda\\), which enhances the representation of level set details." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1281" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.104, + 0.089, + 0.354, + 0.203 + ], + "angle": 0, + "content": "
IoU ↑dC ↓dH ↓
DiGS [3]0.78820.00550.1267
StEik [4]0.66200.00730.1425
Ours0.98700.00140.0153
RMSE ↓MAE ↓SMAPE ↓
DiGS [3]0.05970.03150.3363
StEik [4]0.07250.04190.4222
Ours0.01990.01010.0699
" + }, + { + "type": "table_caption", + "bbox": [ + 0.097, + 0.207, + 0.361, + 0.276 + ], + "angle": 0, + "content": "Table 1. Comparison of 2D reconstruction and distance query. \\(d_{C}\\) and \\(d_{H}\\) stand for the Chamfer and Hausdorff distances respectively. Our method outperforms DiGS [3] and StEik [4] on all metrics." + }, + { + "type": "table", + "bbox": [ + 0.37, + 0.09, + 0.913, + 0.187 + ], + "angle": 0, + "content": "
LBLELALDLDDLSLH| IoU ↑dC ↓dH ↓RMSE ↓MAE ↓SMAPE ↓
✓✓0.81920.00680.07120.03770.01580.1315
✓✓0.89360.00290.05790.01650.00810.1205
✓✓✓0.39070.08430.55980.29660.23881.7283
✓✓✓✓0.63380.05660.42210.20580.16161.0919
✓✓✓0.29380.03770.34440.29900.23551.6958
✓✓✓✓0.60370.04140.36560.13040.09370.7792
✓✓✓0.98510.00160.01600.01990.01050.0754
" + }, + { + "type": "table_caption", + "bbox": [ + 0.368, + 0.192, + 0.871, + 0.274 + ], + "angle": 0, + "content": "Table 2. Ablation study of different loss combinations and metrics. From top to bottom, the loss combinations are from: IGR [10], SAL [2], DiGS w/o area loss, DiGS [3], StEik w/o loss, StEik [4], and our HOTSPOT. From left to right, the losses are boundary loss, eikonal loss, area loss, divergence loss, directional divergence loss, SAL loss, and heat loss. Bold and underlined data: optimal; bold only: suboptimal. Same below." + }, + { + "type": "image", + "bbox": [ + 0.118, + 0.284, + 0.428, + 0.54 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.547, + 0.47, + 0.658 + ], + "angle": 0, + "content": "Figure 7. Our heat loss achieves the best reconstruction of triangle fragments. Dashed lines show real boundaries, while the boldest line represents reconstructed ones. (c) and (f) are from the original models by DiGS [3] and StEik [4]; the rest are from our ablation study with different loss combinations. Other combinations fail to remove extra boundaries or accurately capture the topology. More fundamentally, (c), (f), (g), and (h) fall into local optima, leading to incorrect depressions and elevations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.671, + 0.468, + 0.867 + ], + "angle": 0, + "content": "We present the quantitative results in Table 3 and visual results in Fig. 9. More baselines [46, 51] are compared in the supplement. Our HOTSPOT model outperforms current state-of-the-art models across all surface reconstruction metrics, and achieves near-top accuracy in distance queries. While SAL [2] uses the closest point distance in their loss function, leading to slightly lower total absolute error, their performance drops significantly in relative error and near-surface metrics. This suggests that the closest point distance alone is unreliable near the surface. In contrast, as mentioned at the end of Section 3, after the interpolation driven by spectral bias [58], our loss remains effective in faithfully capturing the distance to the interpolated surface." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.468, + 0.902 + ], + "angle": 0, + "content": "Fig. 1 and Fig. 8 show high-genus surface reconstruction with shapes from Mehta et al. [1]. Existing methods strug" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.28, + 0.59, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.595, + 0.265, + 0.645, + 0.274 + ], + "angle": 0, + "content": "DiGS" + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.278, + 0.663, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.668, + 0.264, + 0.714, + 0.274 + ], + "angle": 0, + "content": "StEi" + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.278, + 0.738, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.743, + 0.265, + 0.79, + 0.274 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.744, + 0.278, + 0.81, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.82, + 0.265, + 0.882, + 0.274 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.819, + 0.278, + 0.885, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.336, + 0.593, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.337, + 0.665, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.67, + 0.337, + 0.738, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.744, + 0.337, + 0.811, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.819, + 0.337, + 0.887, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.397, + 0.892, + 0.439 + ], + "angle": 0, + "content": "Figure 8. Visual examples on high genus dataset. Our method uses significantly fewer iterations but reconstructs the topology correctly for all shapes, whereas others generate extra boundaries." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.444, + 0.892, + 0.503 + ], + "angle": 0, + "content": "gle with the complex topology, while our method accurately reconstructs both the surface and the level set. In the supplement, we present evaluations on other high-genus shapes from Mehta et al. and the SRB dataset." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.513, + 0.655, + 0.53 + ], + "angle": 0, + "content": "5.4. Sphere Tracing" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.537, + 0.892, + 0.612 + ], + "angle": 0, + "content": "Fig. 10 visualizes the number of steps a sphere tracer [21] takes to render the shapes obtained by SAL, DiGS, StEik, and our method. Our method has the fewest iterations on average, since it provides the most accurate level sets. More results can be found in the supplement." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.62, + 0.614, + 0.635 + ], + "angle": 0, + "content": "6. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.764 + ], + "angle": 0, + "content": "Efficiency. Our method uses a loss function of only the neural network output and its first-order derivatives, and thus can be more efficient than previous methods which involve second-order derivatives. We report the runtime of our method and StEik [4] for training a model from ShapeNet [55] over 100 iterations using a single NVIDIA A10 GPU with varying network sizes, while keeping all other settings identical, as shown in Tab. 4." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.766, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Determining absorption coefficient \\(\\lambda\\). The absorption coefficient \\(\\lambda\\) has the units of \\([\\mathrm{L}]^{-1}\\), where \\(\\mathrm{L}\\) denotes a length unit. Rescaling the space is equivalent to rescaling \\(\\lambda\\). While a larger \\(\\lambda\\) can improve accuracy, it introduces two challenges: (1) our heat loss includes the factor \\(e^{-2\\lambda |u|}\\), which may become smaller than the precision limit of floating-point numbers, leading to round-off errors; (2) even if round-off errors are ignored, a large \\(\\lambda |u|\\) may reduce the optimizer's ability to effectively shape the field. Never" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1282" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.164, + 0.089, + 0.805, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.286, + 0.893, + 0.315 + ], + "angle": 0, + "content": "Figure 9. Visual results of DiGS [3], StEik [4], and HOTSPOT on ShapeNet [55]. Previous methods generate extra boundaries and fail to escape local optima, while our method successfully reconstructs the correct topology." + }, + { + "type": "table", + "bbox": [ + 0.164, + 0.321, + 0.807, + 0.399 + ], + "angle": 0, + "content": "
IoU ↑dC ↓dH ↓RMSE ↓MAE ↓SMAPE ↓RMSE0.1MAE0.1SMAPE0.1
SAL [2]0.74000.00740.08510.02510.01420.13440.02450.01820.6848
SIREN wo/ n [34]0.48740.00510.05580.50090.42611.26940.05130.03820.8858
DiGS [3]0.96360.00310.04350.11940.07250.21400.01520.00810.1760
StEik [4]0.96410.00320.03680.03870.02480.09310.01470.00810.1770
Ours0.97960.00290.02500.02810.01760.05400.00940.00470.1206
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.403, + 0.893, + 0.446 + ], + "angle": 0, + "content": "Table 3. Surface reconstruction metrics on ShapeNet [55], where \\(d_{C}\\) and \\(d_{H}\\) stand for the Chamfer and Hausdorff distances, RMSE, MAE, and SMAPE are distance field metrics, and \\(\\mathrm{RMSE}_{0.1}\\), \\(\\mathrm{MAE}_{0.1}\\), and \\(\\mathrm{SMAPE}_{0.1}\\) are the same distance field metrics but only for points within distance 0.1 of the surface. See the supplementary material for detailed results with more baselines." + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.457, + 0.468, + 0.713 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.715, + 0.47, + 0.786 + ], + "angle": 0, + "content": "Figure 10. Visualizations and statistics of sphere tracing [21] iteration counts required to locate the surface. In the visual results brighter pixels indicate higher iteration counts. The histograms show the distribution of iterations for all pixels across 10 different camera poses. HOTSPOT finds the surface in fewer steps." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.794, + 0.468, + 0.884 + ], + "angle": 0, + "content": "theless, our loss still pushes \\(u\\) away from 0, preventing unwanted surfaces where \\(\\lambda |u|\\) should be large. We compensate by combining eikonal loss with heat losses. When \\(\\lambda |u|\\) is large, the eikonal loss becomes dominant. Our \\(\\lambda\\) scheduler also effectively shapes the distant regions. Spatially-adaptive parameters may also improve the results." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Necessity of firm boundary condition. The heat diffu" + }, + { + "type": "table", + "bbox": [ + 0.54, + 0.452, + 0.851, + 0.498 + ], + "angle": 0, + "content": "
Structure5 × 1285 × 2568 × 1288 × 256
StEik [4]5.65s10.62s8.78s17.00s
Ours4.49s8.43s6.79s12.86s
" + }, + { + "type": "table_caption", + "bbox": [ + 0.503, + 0.505, + 0.887, + 0.518 + ], + "angle": 0, + "content": "Table 4. Runtime for 100 iterations with different network sizes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.525, + 0.892, + 0.737 + ], + "angle": 0, + "content": "sion of the screened Poisson equation is based on a well-established boundary condition. We enforce the boundary condition through the boundary loss over discrete points, with neural networks interpolating through spectral bias [58]. When the input points are sparse and the absorption coefficient \\(\\lambda\\) is high, an overly strong heat loss (\\(w_{b}\\) being too small) can tear the boundary, causing the signed distance to collapse into an unsigned one like Fig. 6. Thus, our theory highlights the importance of a high boundary weight \\(w_{b}\\), while absorption \\(\\lambda\\) is adjusted to match the input point density or by scaling the point cloud itself. Our experiments show that proper parameter setting and rescaling prevent the collapse. Future research is required for very sparse boundaries, and applications to inverse rendering." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.744, + 0.619, + 0.758 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We propose a new model for neural signed distance function optimization based on the screened Poisson equation. We analyze our loss theoretically and show that it is an asymptotically sufficient condition to the true distance, is stable both to a small perturbation and in the temporal dynamics, and penalizes large surface area. Our experiments show that we reconstruct both better surfaces and better distance approximations compared to many existing methods, especially on complex and high-genus shapes." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1283" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.184 + ], + "angle": 0, + "content": "Acknowledgements. This work was supported in part by NSF grants 2127544, 2238839, 2100237, 2120019, and gifts from Adobe and Google. Additionally, we would like to thank Bing Xu, Zilu Li, Yash Belhe, Ishit Mehta, and Jianan Xiao for their suggestions, discussions, proofreading, and providing reference materials." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.194, + 0.175, + 0.21 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.219, + 0.47, + 0.261 + ], + "angle": 0, + "content": "[1] Ishit Mehta, Manmohan Chandraker, and Ravi Ramamoorthi. A level set theory for neural implicit evolution under explicit flows. In ECCV, pages 711-729, 2022. 1, 2, 6, 7, 8, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.262, + 0.472, + 0.316 + ], + "angle": 0, + "content": "[2] Matan Atzmon and Yaron Lipman. Sal: Sign agnostic learning of shapes from raw data. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2565-2574, 2020. 1, 2, 6, 7, 8, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.317, + 0.471, + 0.385 + ], + "angle": 0, + "content": "[3] Yizhak Ben-Shabat, Chamin Hewa Koneputugodage, and Stephen Gould. DiGS: Divergence guided shape implicit neural representation for unoriented point clouds. In IEEE Conf. Comput. Vis. Pattern Recog., pages 19323-19332, 2022. 1, 2, 3, 6, 7, 8, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.387, + 0.47, + 0.442 + ], + "angle": 0, + "content": "[4] Huizong Yang, Yuxin Sun, Ganesh Sundaramoorthi, and Anthony Yezzi. StEik: stabilizing the optimization of neural signed distance functions and finer shape representation. arXiv preprint arXiv:2305.18414, 2023. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.443, + 0.47, + 0.497 + ], + "angle": 0, + "content": "[5] Zoë Marschner, Silvia Sellán, Hsueh-Ti Derek Liu, and Alec Jacobson. Constructive solid geometry on neural signed distance fields. In SIGGRAPH Asia 2023 Conference Papers, pages 1-12, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.498, + 0.47, + 0.552 + ], + "angle": 0, + "content": "[6] Sathamangalam R Srinivasa Varadhan. On the behavior of the fundamental solution of the heat equation with variable coefficients. Communications on Pure and Applied Mathematics, 20(2):431-455, 1967. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.553, + 0.469, + 0.594 + ], + "angle": 0, + "content": "[7] Keenan Crane, Clarisse Weischedel, and Max Wardetzky. Geodesics in heat: A new approach to computing distance based on heat flow. ACM TOG, 32(5):1-11, 2013. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.47, + 0.65 + ], + "angle": 0, + "content": "[8] Alexander G Belyaev and Pierre-Alain Fayolle. On variational and pde-based distance function approximations. In Computer Graphics Forum, volume 34, pages 104-118. Wiley Online Library, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.651, + 0.47, + 0.692 + ], + "angle": 0, + "content": "[9] Alexander Belyaev and Pierre-Alain Fayolle. An ADMM-based scheme for distance function approximation. Numerical Algorithms, 84:983-996, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.693, + 0.47, + 0.734 + ], + "angle": 0, + "content": "[10] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. arXiv preprint arXiv:2002.10099, 2020. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.735, + 0.469, + 0.762 + ], + "angle": 0, + "content": "[11] James F Blinn. A generalization of algebraic surface drawing. ACM TOG, 1(3):235-256, 1982. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.763, + 0.469, + 0.803 + ], + "angle": 0, + "content": "[12] Jules Bloomenthal and Brian Wyvill. Interactive techniques for implicit modeling. Comput. Graph. (Proc. SIGGRAPH), 24(2):109-116, 1990." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.805, + 0.469, + 0.858 + ], + "angle": 0, + "content": "[13] Hugues Hoppe, Tony DeRose, Tom Duchamp, John McDonald, and Werner Stuetzle. Surface reconstruction from unorganized points. Comput. Graph. (Proc. SIGGRAPH), 26(2):71-78, 1992." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[14] Brian Curless and Marc Levoy. A volumetric method for building complex models from range images. In SIGGRAPH, page 303-312, 1996." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.219, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.188 + ], + "angle": 0, + "content": "[15] Shahram Izadi, David Kim, Otmar Hilliges, David Molyneaux, Richard Newcombe, Pushmeet Kohli, Jamie Shotton, Steve Hodges, Dustin Freeman, and Andrew Davison. KinectFusion: real-time 3D reconstruction and interaction using a moving depth camera. In ACM Symposium on User Interface Software and Technology, pages 559-568, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.19, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[16] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4460-4470, 2019. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.246, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[17] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5939-5948, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.288, + 0.892, + 0.329 + ], + "angle": 0, + "content": "[18] Thomas Davies, Derek Nowrouzezahrai, and Alec Jacobson. On the effectiveness of weight-encoded neural implicit 3d shapes. arXiv preprint arXiv:2009.09808, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.892, + 0.385 + ], + "angle": 0, + "content": "[19] Francis Williams, Matthew Trager, Joan Bruna, and Denis Zorin. Neural Splines: Fitting 3D surfaces with infinitely-wide neural networks. In IEEE Conf. Comput. Vis. Pattern Recog., pages 9949-9958, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.386, + 0.892, + 0.441 + ], + "angle": 0, + "content": "[20] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. 41(2):641-676, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.442, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[21] John C Hart. Sphere tracing: A geometric method for the antialiased ray tracing of implicit surfaces. The Vis. Comput., 12(10):527-545, 1996. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.892, + 0.538 + ], + "angle": 0, + "content": "[22] Dario Seyb, Alec Jacobson, Derek Nowrouzezahrai, and Wojciech Jarosz. Non-linear sphere tracing for rendering deformed signed distance fields. ACM Trans. Graph. (Proc. SIGGRAPH Asia), 38(6), 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.539, + 0.892, + 0.566 + ], + "angle": 0, + "content": "[23] Eric Galin, Eric Guérin, Axel Paris, and Adrien Peytavie. Segment tracing using local Lipschitz bounds, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.567, + 0.892, + 0.621 + ], + "angle": 0, + "content": "[24] Nicholas Sharp and Alec Jacobson. Spelunking the deep: guaranteed queries on general neural implicit surfaces via range analysis. ACM Trans. Graph. (Proc. SIGGRAPH), 41(4), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.622, + 0.892, + 0.704 + ], + "angle": 0, + "content": "[25] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3D shapes. In IEEE Conf. Comput. Vis. Pattern Recog., pages 11358-11367, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.706, + 0.892, + 0.76 + ], + "angle": 0, + "content": "[26] Arnulph Fuhrmann, Gerrit Sobotka, and Clemens Groß. Distance fields for rapid collision detection in physically based modeling. In Proceedings of GraphiCon, volume 2003, pages 58-65, 2003. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.761, + 0.892, + 0.802 + ], + "angle": 0, + "content": "[27] Eran Guendelman, Robert Bridson, and Ronald Fedkiw. Nonconvex rigid bodies with stacking. ACM TOG, 22(3):871-878, 2003." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.803, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[28] Miles Macklin, Kenny Erleben, Matthias Müller, Nuttapong Chentanez, Stefan Jeschke, and Zach Corse. Local optimization for robust signed distance field collision. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 3(1):1-17, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[29] Rohan Sawhney and Keenan Crane. Monte Carlo geometry processing: a grid-free approach to PDE-based methods on" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1284" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "volumetric domains. ACM Trans. Graph. (Proc. SIGGRAPH), 39(4), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.176 + ], + "angle": 0, + "content": "[30] Albert Pumarola, Artsiom Sanakoyeu, Lior Yariv, Ali Thabet, and Yaron Lipman. Visco grids: Surface reconstruction with viscosity and coarea grids. Advances in Neural Information Processing Systems, 35:18060-18071, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.178, + 0.469, + 0.232 + ], + "angle": 0, + "content": "[31] Songyou Peng, Chiyu Jiang, Yiyi Liao, Michael Niemeyer, Marc Pollefeys, and Andreas Geiger. Shape as points: A differentiable poisson solver. Advances in Neural Information Processing Systems, 34:13032-13044, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.234, + 0.469, + 0.288 + ], + "angle": 0, + "content": "[32] Chao Chen, Yu-Shen Liu, and Zhizhong Han. Gridpull: Towards scalability in learning implicit representations from 3d point clouds. In Proceedings of the IEEE/cvf international conference on computer vision, pages 18322-18334, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.29, + 0.469, + 0.356 + ], + "angle": 0, + "content": "[33] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. DeepSDF: Learning continuous signed distance functions for shape representation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 165-174, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.359, + 0.469, + 0.425 + ], + "angle": 0, + "content": "[34] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in neural information processing systems, 33:7462-7473, 2020. 2, 8, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.428, + 0.469, + 0.469 + ], + "angle": 0, + "content": "[35] Guillaume Coiffier and Louis Béthune. 1-Lipschitz neural distance fields. Comput. Graph. Forum (Eurographics STARs), 43(5):e15128, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.471, + 0.469, + 0.525 + ], + "angle": 0, + "content": "[36] Yue Jiang, Dantong Ji, Zhizhong Han, and Matthias Zwicker. SDFDiff: Differentiable rendering of signed distance fields for 3D shape optimization. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1251-1261, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.527, + 0.469, + 0.58 + ], + "angle": 0, + "content": "[37] Petr Kellnhofer, Lars C Jebe, Andrew Jones, Ryan Spicer, Kari Pulli, and Gordon Wetzstein. Neural lumigraph rendering. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4287-4297, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.582, + 0.469, + 0.649 + ], + "angle": 0, + "content": "[38] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3D representations without 3D supervision. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3504-3515, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.651, + 0.469, + 0.72 + ], + "angle": 0, + "content": "[39] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. In Advances in Neural Information Processing Systems, volume 33, pages 2492-2502, 2020. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.721, + 0.469, + 0.788 + ], + "angle": 0, + "content": "[40] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snavely. PhySG: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5453-5462, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.469, + 0.859 + ], + "angle": 0, + "content": "[41] Sai Bangaru, Michael Gharbi, Tzu-Mao Li, Fujun Luan, Kalyan Sunkavalli, Milos Hasan, Sai Bi, Zexiang Xu, Gilbert Bernstein, and Fredo Durand. Differentiable rendering of neural sdfs through reparameterization. In SIGGRAPH Asia Conference Proceedings, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.861, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[42] Delio Vicini, Sébastien Speierer, and Wenzel Jakob. Differentiable signed distance function rendering. ACM Trans. Graph. (Proc. SIGGRAPH), 41(4):1-18, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[43] Jingyang Zhang, Yao Yao, Shiwei Li, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Critical regularizations for neural surface reconstruction in the wild. In IEEE Conf. Comput. Vis. Pattern Recog., pages 6270-6279, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.151, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[44] Ruian Wang, Zixiong Wang, Yunxiao Zhang, Shuangmin Chen, Shiqing Xin, Changhe Tu, and Wenping Wang. Aligning gradient and hessian for neural signed distance function. In Advances in Neural Information Processing Systems, volume 36, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.223, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[45] Zixiong Wang, Yunxiao Zhang, Rui Xu, Fan Zhang, Peng-Shuai Wang, Shuangmin Chen, Shiqing Xin, Wenping Wang, and Changhe Tu. Neural-singular-hessian: Implicit neural representation of unoriented point clouds by enforcing singular hessian. ACM Transactions on Graphics (TOG), 42(6):1-14, 2023. 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.308, + 0.892, + 0.376 + ], + "angle": 0, + "content": "[46] Qiujie Dong, Rui Xu, Pengfei Wang, Shuangmin Chen, Shiqing Xin, Xiaohong Jia, Wenping Wang, and Changhe Tu. Neurcadrecon: Neural representation for reconstructing cad surfaces by enforcing zero gaussian curvature. arXiv preprint arXiv:2404.13420, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.38, + 0.892, + 0.435 + ], + "angle": 0, + "content": "[47] Yesom Park, Taekyung Lee, Jooyoung Hahn, and Myungjoo Kang. \\(p\\)-poisson surface reconstruction in curl-free flow from point clouds. In Advances in Neural Information Processing Systems, volume 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.438, + 0.892, + 0.506 + ], + "angle": 0, + "content": "[48] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In IEEE Conf. Comput. Vis. Pattern Recog., pages 8456-8465, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.51, + 0.892, + 0.564 + ], + "angle": 0, + "content": "[49] Yaron Lipman. Phase transitions, distance functions, and implicit neural representations. In International Conference on Machine Learning, volume 139, pages 6702-6712, 2021. 2, 3, 6, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.568, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[50] Angela Dai and Matthias Nießner. Neural poisson: Indicator functions for neural fields. arXiv preprint arXiv:2211.14249, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.612, + 0.892, + 0.667 + ], + "angle": 0, + "content": "[51] Baorui Ma, Zhizhong Han, Yu-Shen Liu, and Matthias Zwicker. Neural-pull: Learning signed distance functions from point clouds by learning to pull space onto surfaces. arXiv preprint arXiv:2011.13495, 2020. 2, 7, 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.67, + 0.892, + 0.696 + ], + "angle": 0, + "content": "[52] Matan Atzmon and Yaron Lipman. SALD: Sign agnostic learning with derivatives. In ICLR, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.7, + 0.892, + 0.727 + ], + "angle": 0, + "content": "[53] Nicholas Sharp, Yousuf Soliman, and Keenan Crane. The vector heat method. ACM TOG, 38(3):1-19, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.731, + 0.892, + 0.771 + ], + "angle": 0, + "content": "[54] Nicole Feng and Keenan Crane. A heat method for generalized signed distance. ACM Trans. Graph. (Proc. SIGGRAPH), 43(4), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.774, + 0.892, + 0.842 + ], + "angle": 0, + "content": "[55] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 6, 7, 8, 5, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[56] Matthew Berger, Joshua A Levine, Luis Gustavo Nonato, Gabriel Taubin, and Claudio T Silva. A benchmark for surface reconstruction. ACM Transactions on Graphics (TOG), 32(2):1-17, 2013. 6, 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.515, + 0.957 + ], + "angle": 0, + "content": "1285" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.134 + ], + "angle": 0, + "content": "[57] David Stutz and Andreas Geiger. Learning 3d shape completion under weak supervision. International Journal of Computer Vision, 128:1162-1181, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.468, + 0.205 + ], + "angle": 0, + "content": "[58] Nasim Rahaman, Aristide Baratin, Devansh Arpit, Felix Draxler, Min Lin, Fred Hamprecht, Yoshua Bengio, and Aaron Courville. On the spectral bias of neural networks. In International Conference on Machine Learning, pages 5301-5310, 2019. 7, 8, 3, 5, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.207, + 0.468, + 0.275 + ], + "angle": 0, + "content": "[59] Dmitry S Kulyabov, Anna V Korolkova, Tatiana R Velieva, and Migran N Gevorkyan. Numerical analysis of eikonal equation. In Saratov Fall Meeting 2018: Laser Physics, Photonic Technologies, and Molecular Modeling, volume 11066, pages 188-195. SPIE, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.277, + 0.468, + 0.318 + ], + "angle": 0, + "content": "[60] Zhen-Hang Yang and Yu-Ming Chu. On approximating the modified bessel function of the second kind. Journal of inequalities and applications, 2017:1-8, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.32, + 0.468, + 0.36 + ], + "angle": 0, + "content": "[61] Luciano Modica. The gradient theory of phase transitions and the minimal interface criterion. Archive for Rational Mechanics and Analysis, 98:123-142, 1987. 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.362, + 0.468, + 0.402 + ], + "angle": 0, + "content": "[62] Peter Sternberg. The effect of a singular perturbation on nonconvex variational problems. Archive for Rational Mechanics and Analysis, 101:209-260, 1988. 10" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1286" + } + ] +] \ No newline at end of file diff --git a/2025/HotSpot_ Signed Distance Function Optimization with an Asymptotically Sufficient Condition/733f46a7-f78c-4e53-bba4-ead7eb705bb7_origin.pdf b/2025/HotSpot_ Signed Distance Function Optimization with an Asymptotically Sufficient Condition/733f46a7-f78c-4e53-bba4-ead7eb705bb7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9dbfcb54159957585d39c517a27556537fc1ba21 --- /dev/null +++ b/2025/HotSpot_ Signed Distance Function Optimization with an Asymptotically Sufficient Condition/733f46a7-f78c-4e53-bba4-ead7eb705bb7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:553d5918669a1f4e15df151549870b58c35c5412830b1c7889b1cb6a4d2df100 +size 8063466 diff --git a/2025/HotSpot_ Signed Distance Function Optimization with an Asymptotically Sufficient Condition/full.md b/2025/HotSpot_ Signed Distance Function Optimization with an Asymptotically Sufficient Condition/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4be421f39e1b72d9451210e774857254eb493b13 --- /dev/null +++ b/2025/HotSpot_ Signed Distance Function Optimization with an Asymptotically Sufficient Condition/full.md @@ -0,0 +1,446 @@ +# HotSpot: Signed Distance Function Optimization with an Asymptotically Sufficient Condition + +Zimo Wang* Cheng Wang* Taiki Yoshino Sirui Tao Ziyang Fu Tzu-Mao Li UC San Diego + +![](images/74e2cc9cebace942b87899a27c1858b46acc7b83ef83fb6f154ef3b51db6ab2f.jpg) +Figure 1. We propose HOTSPOT, a neural signed distance function optimization method that establishes an asymptotic sufficient condition to guarantee convergence to a true distance function, enabling precise surface reconstruction and level set representation for complex shapes. Here we show a reconstruction from a point cloud sampled from the reference bunny (taken from Mehta et al. [1]) on the right. In the inset, we visualize the recovered signed distance function on a horizontal slice, using warm colors for positive values and cool for negative (zoom in for details). Our reconstruction is significantly more accurate than prior works (SAL [2], DiGS [3], and StEik [4]). + +# Abstract + +We propose a method, HOTSPOT, for optimizing neural signed distance functions. Existing losses, such as the eikonal loss, act as necessary but insufficient constraints and cannot guarantee that the recovered implicit function represents a true distance function, even if the output minimizes these losses almost everywhere. Furthermore, the eikonal loss suffers from stability issues in optimization. Finally, in conventional methods, regularization losses that penalize surface area distort the reconstructed signed distance function. We address these challenges by designing a loss function using the solution of a screened Poisson equation. Our loss, when minimized, provides an asymptotically sufficient condition to ensure the output converges to a true distance function. Our loss also leads to stable optimization and naturally penalizes large surface areas. We present theoretical analysis and experiments on both challenging 2D and 3D datasets and show that our method provides better surface reconstruction and a more accurate distance approximation. + +# 1. Introduction + +Learning neural signed distance function requires designing loss functions so that the zero crossings of the implicit + +function are at the desired locations (e.g., adhere to a sparse input point cloud), and that the implicit function returns the correct signed distance. Designing the loss function proves to be challenging for shapes with complex topology and details (Fig. 1). We propose HOTSPOT, a method based on a relation between the screened Poisson equation and distance that enjoys both theoretical and practical benefits. + +A major challenge in signed distance function optimization is ensuring the implicit function follows the true distance. A common regularization loss used is the eikonal equation, which constrains the gradient norm of an implicit function to be 1 almost everywhere. If the implicit function is a signed distance function, then it satisfies the eikonal equation. However, the converse is not true. That is, satisfying the eikonal equation is only a necessary condition for the signed distance function, not a sufficient condition [5]. Fig. 2 presents the logic and Fig. 3 shows an example. + +Another key challenge lies in the optimization. It is known that the eikonal equation regularization is unstable [4], and the instability can make optimization converge to suboptimal results or even diverge. + +Furthermore, to address optimization ill-posedness and instability, existing methods penalize the surface area of the shape. We prove this degrades the reconstruction quality and distorts the output level sets, thereby leading to a delicate balance between over- and under-smoothing. + +$$ +\begin{array}{c} \min \{\text {h e a t l o s s} \} \\ \lambda \to \infty \end{array} \Leftrightarrow \begin{array}{l} \text {D i s t a n c e} \quad \text {a . e .} \\ \text {F u n c t i o n} \quad \stackrel {{\Rightarrow}} {{\Leftarrow}} \end{array} \left\{ \begin{array}{l} \text {E i k o n a l c o n d i t i o n} \\ \text {Z e r o g r a d i e n t d i v e r g e n c e} \\ \text {S i n g u l a r H e s s i a n m a t r i x} \\ \dots \end{array} \right. +$$ + +We propose a simple model that can address all the challenges above. Our model, HOTSPOT, is based on a screened Poisson equation and a classical relation between heat transfer and distance [6] that was popular for approximating both Euclidean and geodesic distance [7-9]. Using this relation, we design a loss function for optimizing neural signed distance functions. As our contributions, we show, both theoretically and empirically, that our model + +- serves as a sufficient condition ensuring the difference between the minimizer and the actual distance is bounded and decreases linearly as the parameter $\lambda$ increases (Fig. 2 and Fig. 3, right). +- is stable both spatially and temporally, in the sense that a slight perturbation of the implicit function only introduces changes in a local region, and that the dynamical system formed by the gradient flow will converge in the long run. +- penalizes the surface area naturally, while remaining an undistorted signed distance function. +- works well in both 2D and 3D shapes, and it excels at approaching signed distance functions while optimizing for complex, high-genus shapes. + +# 2. Related Work + +Implicit surfaces. Modeling and reconstructing surfaces using implicit functions has a long history in computer vision and graphics [11-15]. Recently, there is a surge of interest to model surfaces using neural implicit functions [16-20]. Neural implicit representations are favored for their compactness, ability to represent detailed surfaces, and compatibility to gradient-based optimization [1]. + +Neural signed distance functions. Among the classes of implicit surfaces, signed distance functions, where the implicit function outputs the distance to the closest point on the surface, and the sign represents whether the point is inside or outside the solid object, is one of the most popular variants. Signed distance functions enable an efficient rendering algorithm called sphere tracing [21-24] and helps geometry compression [25]. They can also be used for collision detection [26-28] and numerical simulation [29]. Traditionally, signed distance functions are often represented + +![](images/b06f6a187e16cd03ca5cd591c9a9b1a69f7c7a7b9b8bab6a8d473ec08cb4d766.jpg) +Figure 2. Existing constraints are only necessary but insufficient conditions. That is, there exist solutions, like Fig. 3, that satisfy them almost everywhere yet fail to be distance functions. In contrast, we use a heat loss, modeled by a screened Poisson equation with parameter $\lambda$ . As the parameter $\lambda$ goes to infinity, the minimizer converges to the distance function and excludes other solutions. Thus, our method is asymptotically both necessary and sufficient. +Figure 3. We illustrate a 1D neural signed distance function optimization using the classical eikonal loss [10] and our model. The $x$ axis is the domain and the $y$ axis shows the output of the implicit function. The middle and bottom rows show the intermediate and final states of the optimization respectively. The eikonal loss, as a necessary but insufficient condition, is incapable of converging to the actual signed distance function (dashed line), even when the function satisfies the eikonal equation almost everywhere. + +as grids [30-32]. More recently, neural signed distance functions have become an appealing geometric representation [33-35]. (Neural) Signed distance functions are also often used in inverse rendering [36-42], for their flexibility to obtain surfaces with complex topology. + +Regularizing and initializing neural signed distance functions. A commonly used loss for optimizing neural signed distance functions is the eikonal regularization, which encourages the norm of the gradient of the implicit function to be one [10, 39]. However, the eikonal regularization alone is far from sufficient, as 1) minimizing eikonal loss does not necessarily lead to a distance field [5, 30], and 2) its optimization is unstable [4]. Previous works have proposed initialization to bootstrap the optimization with a simple shape [3, 10]. Recent works show that minimizing the divergence and the surface area of the implicit function can lead to better reconstruction [3, 43] and stabilizes the optimization [4]. Second-order information is also often used [43-46] for regularization. All losses above are based on necessary properties of real distance but neither establish convergence nor exclude non-SDF solutions. + +Specialized architectures have been proposed to limit the Lipschitz constant of the implicit function [35], satisfy $p$ -Poisson equation [47], or represent high-frequency details [25, 34, 48]. Some research [49, 50] has shown a link between occupancy function optimization and an approximate signed distance field regularization. Our experiments further reveal that some implementations [2, 51, 52] + +optimizing SDFs with the closest point may struggle when querying near the interpolated surface. + +We propose a new loss and method for optimizing signed distance functions. While derived from very different mathematics, our loss turns out to have a close connection to Lipman's PHASE loss [49]. We show that our theory leads to significantly more numerically stable optimization, and a different conclusion in the parameter setting that leads to drastically different results. + +Distance approximation using heat transfer. Our derivation is inspired by a relation between heat transfer and approximated distance [6]. This relation has been used in computer graphics for efficiently computing the geodesic distance on surfaces [7-9, 53, 54]. We apply the relation for optimizing neural signed distance functions, and analyze the properties of the resulting optimization in this context. + +# 3. Background and Motivation + +Given an input point cloud $\pmb{x}_i\in \mathbb{R}^d$ (we focus on 2D and 3D in this work) with $N$ points without normal information that lies on an unknown surface $\Gamma$ of a solid object $S$ , our goal is to find a signed distance function $f(\pmb {x})$ , such that + +$$ +f (\boldsymbol {x}) = \left\{ \begin{array}{l l} - d _ {\Gamma} (\boldsymbol {x}) & \text {i f} \boldsymbol {x} \in S \\ d _ {\Gamma} (\boldsymbol {x}) & \text {o t h e r w i s e ,} \end{array} \right. \tag {1} +$$ + +where $d_{\Gamma}(\pmb{x})$ denotes the closest distance from point $\pmb{x}$ to the surface $\Gamma$ . In practice, we approximate the signed distance function $f(\pmb{x})$ using a neural network $u(\pmb{x}): \mathbb{R}^d \to \mathbb{R}$ parameterized by its weights and use gradient-based optimization to figure out the weights. Previous work (e.g., [10]) often uses a linear combination of a boundary loss $L_{\mathrm{boundary}}$ and a eikonal loss $L_{\mathrm{eikonal}}$ : + +$$ +L _ {\text {b o u n d a r y}} = \frac {1}{N} \sum_ {i = 1} ^ {N} | u (\boldsymbol {x} _ {i}) | ^ {p}, \tag {2} +$$ + +$$ +L _ {\text {e i k o n a l}} = \int_ {\Omega} | \| \nabla u (\boldsymbol {x}) \| - 1 | ^ {p} \mathrm {d} \boldsymbol {x}, \quad p = 1 \text {o r} 2, \tag {3} +$$ + +where $\Omega \subset \mathbb{R}^d$ is our domain of interest. The boundary loss encourages the implicit function $u$ to output 0 at the boundary, and the eikonal loss is based on that observation that if a function $f$ is a distance function, the $L^2$ norm of its gradient $\| \nabla f(\pmb{x})\| = 1$ holds almost everywhere. Thus the approximation $u$ is encouraged to behave the same. + +However, as noted, simply optimizing for the two terms above does not guarantee that $u$ would be a distance function even when satisfied almost everywhere (Figs. 2 and 3), and the extra regularization previous work added to minimize the area and divergence [3, 4] could further tamper with the distance approximation quality. We derive an additional loss that can be used together with both of the losses above, and analyze our loss theoretically and empirically. + +![](images/5dfab8084f9c63b1361b5c2879bf2d56d0ffec1bd6d0b09daced5bc564ec7969.jpg) + +![](images/9d07dff7616f91a37d1a713146f880761e116e3bcd78cc309afe6228b64be931.jpg) + +![](images/9ff077163feb3e8b536152bf928464003c5f4d0b8f6bfaa6c50f62203c8997b2.jpg) +(c) HOUSE + +![](images/5c75ab31f618615213776769f9e9fd3672bdc3296f16f4db456dc7f8c55753bb.jpg) +(a) 1D Heat Simulation via Eq. (4) +(d) $\lambda = 1$ +Figure 4. We show an illustration of the relation between screened Poisson equation and a distance field. The top shows a 1D example, where on the left we show a solution to Eq. (4) with different absorption $\lambda$ , and on the right we show the reconstructed distance $-\frac{1}{\lambda}\ln(h)$ (Eq. (5)). Boundary loss for real boundary points acts as an isothermal heat source, while heat loss diffuses the heat. When $\lambda$ is increased, the heat decays faster, and the error of the reconstructed function approaches 0. The bottom shows a 2D example. + +![](images/474199d8dbc3f6e92fa3f367cdeb6155f8e54a98a2fd0adfd608ac741a0a8232.jpg) +(b) Distance Reconstruction +(e) $\lambda = 5$ + +![](images/169c024f19af33225d597fd7b78d5b377a35da7298ba1ac6ee7880f21995eb89.jpg) +(f) $\lambda = 50$ + +We will be using a key relation between the screen Poisson equation and the distance. We first define a screened Poisson equation over a heat field $h(\pmb{x}): \mathbb{R}^d \to \mathbb{R}$ , with an absorption coefficient $\lambda$ , and set a Dirichlet boundary condition to be 1 on our target surface $\Gamma$ : + +$$ +\left\{\begin{array}{l}\nabla^ {2} h (\boldsymbol {x}) - \lambda^ {2} h (\boldsymbol {x}) = 0 \quad \forall \boldsymbol {x} \in \mathbb {R} ^ {d} \backslash \Gamma\\\lim _ {| | \boldsymbol {x} | | \rightarrow \infty} h (\boldsymbol {x}) = 0, \quad h (\boldsymbol {x}) = 1 \quad \forall \boldsymbol {x} \in \Gamma .\end{array}\right. \tag {4} +$$ + +If a heat field $h_\lambda$ is a solution to the partial differential equation above, then the following equation holds [6]: + +$$ +\lim _ {\lambda \rightarrow \infty} \frac {1}{\lambda} \ln \left(h _ {\lambda} (\boldsymbol {x})\right) = - d _ {\Gamma} (\boldsymbol {x}). \tag {5} +$$ + +Fig. 4 shows the intuition. The screened Poisson equation simulates heat transfer until an equilibrium is reached, under absorption controlled by the $\lambda$ coefficient. As we increase the absorption, the reconstructed distance $\frac{1}{\lambda}\ln (h_{\lambda})$ would approach the real distance. In practice, we could only obtain a discrete point set $\Gamma_0$ . Provided that the neighbor points are close enough, if there are some boundaries connecting them, the boundary condition of Eq. (4) still holds. This can be interpreted as a superset $\Gamma$ replacing $\Gamma_0$ , effectively transforming the boundary from a discrete point cloud into a continuous surface, which is required and precisely the goal of the signed distance function reconstruction. Obtaining a distance function to the point cloud rather than the interpolated surface is not our objective. The reconstructed distance can still faithfully represent the distance to $\Gamma$ because the approximation in Eq. (5) still holds. Later, we will prove that the measure of $\Gamma$ is bounded by our loss. + +# 4. Method + +In this section, we introduce our method, HOTSPOT, based on the screened Poisson equation (Eq. (4)) and derive a theoretical analysis. We use the screened Poisson equation to derive a loss we can directly apply to an implicit function $u$ , and prove that there is a bound to the solution of the screened Poisson equation to the true distance (Section 4.1). We further derive a stability analysis of our loss on both the spatial stability over small perturbations and the temporal convergence (Section 4.2). Next, we show that our loss can penalize surfaces with large area (Section 4.3). Finally, we discuss the relation to prior work, PHASE (Section 4.4). + +# 4.1. Screened Poisson Equation Informed Signed Distance Functions Optimization + +Given an implicit function $u$ , we want to derive a loss based on the screened Poisson equation (Eq. (4)) and its relation to the distance (Eq. (5)). We achieve this by applying the following substitution based on Eq. (5): + +$$ +h (\boldsymbol {x}) = e ^ {- \lambda | u (\boldsymbol {x}) |}. \tag {6} +$$ + +We want to design a loss such that its minimizer satisfies the screened Poisson equation. This can be achieved by minimizing the energy functional $\frac{1}{2}\int_{\Omega} \| \nabla h(\pmb{x}) \|^2 + \lambda^2 h(\pmb{x})^2 \, \mathrm{d}\pmb{x}$ . After substitution and removing of the constant factor $\lambda$ , we obtain our heat loss: + +$$ +L _ {\text {h e a t}} = \frac {1}{2} \int_ {\Omega} e ^ {- 2 \lambda | u (\boldsymbol {x}) |} \left(\| \nabla u (\boldsymbol {x}) \| ^ {2} + 1\right) d \boldsymbol {x}. \tag {7} +$$ + +If we take the derivative of $L_{\mathrm{heat}}$ with respect to the heat field $h$ , we recover the screened Poisson equation $\nabla^2 h - \lambda^2 h = 0$ . The boundary condition can be enforced by the standard boundary loss (Eq. (2)), by observing that $u(\pmb{x}) = 0$ implies $h(\pmb{x}) = 1$ . Together with the eikonal loss, we optimize: + +$$ +L = w _ {b} L _ {\text {b o u n d a r y}} + w _ {e} L _ {\text {e i k o n a l}} + w _ {h} L _ {\text {h e a t}}. \tag {8} +$$ + +While prior work [6] has studied the limiting behavior of the absorption coefficient $\lambda$ , to our knowledge, the bounds of the approximation and the convergence speed remain unknown. In the supplementary material A.3, we prove that, for a boundary $\Gamma$ formed by a discrete set of points and a small volume around them, the distance approximation obtained from the solution of the screened Poisson equation converges linearly with respect to $\lambda$ : + +$$ +\frac {1}{\lambda} C _ {\text {l o w e r}} (\boldsymbol {x}) \leqslant \frac {1}{\lambda} \ln \left(h _ {\lambda} (\boldsymbol {x})\right) + d _ {\Gamma} (\boldsymbol {x}) \leqslant \frac {1}{\lambda} C _ {\text {u p p e r}} (\boldsymbol {x}). \tag {9} +$$ + +Here, both $C_{\mathrm{upper}}(\pmb{x})$ and $C_{\mathrm{lower}}(\pmb{x})$ are scalar fields independent of $\lambda$ . The inequalities are tight, which means that one can find a case where the equality holds. + +Compared to the eikonal loss (Eq. (3)), where the distance error can be unbounded even when the equation holds + +![](images/10116da4676066829bfb6317112af6411f3b4a5a054c8bd4e130bdcdbb90e036.jpg) +Figure 5. We visualize the direction of the changes of the implicit function during one iteration of optimization. Left shows the result with only the eikonal loss and right shows the result after adding our loss. Our loss helps to escape the local minimum in Fig. 3. The heat diffusion encourages the reconstructed absolute distance to increase monotonically away from the boundary. + +almost everywhere, Eq. (9) shows that even with a finite absorption $\lambda$ , our loss can lead to much more accurate distance. In fact, any loss that only depends on the norm of the gradients $\|\nabla u\|$ is not a sufficient condition of the distance function (Fig. 2), due to the discontinuous jumps of the gradient. Adding our loss (Eq. (7)) ensures that the absolute value of the implicit function converges to the actual distance. Fig. 5 visualizes the direction of how an implicit function would evolve when using only the eikonal loss and after adding our loss. Our loss ensures that the implicit function evolves towards the actual distance. + +# 4.2. Stability Analysis + +Our heat loss enjoys other theoretical benefits apart from the convergence to an actual distance. Here, we show that our loss leads to easier optimization. We separately analyze spatial stability and temporal stability, where the spatial stability studies the effect of small perturbation error added to the corresponding partial differential equation, and the temporal stability studies the convergence of the gradient flow. + +# 4.2.1 Spatial Stability + +We show that for the screened Poisson equation, an error introduced to a single point in the solution decays exponentially and does not propagate to infinitely distant positions. On the contrary, for the eikonal equation, the error can propagate to infinitely far. The property of the screened Poisson equation ensures that the perturbation introduced during optimization does not lead to a drastically different target and stabilizes the optimization. + +First, we show that for the eikonal equation, an additive error introduced to a point can presist along a line segment or a ray. + +Proposition 1. Given a field $u(\pmb{x})$ that satisfies the eikonal equation $\| \nabla u \| = 1$ , we introduce an additive error $u_{0e}$ at the point $\pmb{x}_0$ where $u(\pmb{x}_0) = u_0$ to obtain a perturbed field $u'$ where $u'(\pmb{x}_0) = u_0 + u_{0e}$ and $\| \nabla u' \| = 1$ . In the + +perturbed field $u'$ , there exists a line segment parameterized by $\pmb{x}(s)$ starting from $\pmb{x}_0$ and $s \in [0, M]$ , where $M > 0$ and can be infinite, such that every point along the line has the same error: + +$$ +u ^ {\prime} (\boldsymbol {x} (s)) - u (\boldsymbol {x} (s)) = u _ {0 e} \tag {10} +$$ + +The boundary points can contain errors when approximating the true boundaries, and these errors remain constant along the line segment. As a result, the propagated error can become significant and potentially detrimental. + +Next, we demonstrate that the error field $h_e$ in Eq. (4) decays exponentially from the perturbed point $x_0$ . + +Proposition 2. In 3D space, consider an additive error $h_{0e}$ within a small ball $B(\pmb{x}_0,\epsilon)$ around the point $\pmb{x}_0$ , where the original field $h(\pmb{x})$ and the perturbed field $h^{\prime}(\pmb{x})$ , $\pmb{x} \in \mathbb{R}^n\backslash B$ both satisfy the screened Poisson equation in Eq. (4). The resulting error field $h_e = h' - h$ is radially symmetric with respect to $x_0$ within some maximum radius $R$ , and is given by + +$$ +h _ {e} (r) = \frac {\epsilon}{r} h _ {0 e} e ^ {\lambda (\epsilon - r)}, \forall r > \epsilon , \tag {11} +$$ + +where $r = \left\| \boldsymbol{x} - \boldsymbol{x}_0 \right\|$ is the radial distance from $\boldsymbol{x}_0$ . + +The proofs of these two propositions are in the supplementary material A.1 and A.2. By comparing the error fields, we observe that the partial differential equation governed by the screened Poisson equation exhibits greater spatial robustness than the one based on the eikonal equation. + +# 4.2.2 Temporal Stability + +We next demonstrate the temporal stability of the heat loss and show that the dynamical system of the gradient update is stable, that is, the solution of the system converges as the time $t$ approaches infinity. This is supported by von Neumann stability analysis, similar to Yang et al.'s work [4]. + +Yang et al. showed that for the eikonal loss (Eq. (3)), the temporal gradient flow of the optimization is: + +$$ +\begin{array}{l} \frac {\partial u}{\partial t} = - \frac {\delta L _ {\mathrm {e i k o n a l}}}{\delta u} = \nabla \cdot (\kappa_ {e} \nabla u) \approx \kappa_ {e} \nabla^ {2} u, \text {w h e r e} \\ \kappa_ {e} (\boldsymbol {x}) = \left\{ \begin{array}{l l} \frac {1}{| \nabla u (\boldsymbol {x}) |} \operatorname {s g n} (| \nabla u (\boldsymbol {x}) | - 1) & p = 1 \\ 1 - | \nabla u (\boldsymbol {x}) | ^ {- 1} & p = 2 \end{array} \right. \tag {12} \\ \end{array} +$$ + +They analyze the stability by applying Fourier transform: + +$$ +\frac {\partial \hat {u}}{\partial t} (t, \omega) = - \kappa_ {e} | \omega | ^ {2} \hat {u} (t, \omega) \Rightarrow \hat {u} (t, \omega) \propto e ^ {- \kappa_ {e} | \omega | ^ {2} t}, \tag {13} +$$ + +where $\omega = (\omega_{1},\dots ,\omega_{n})$ denotes the frequency variable and $\hat{u}$ represents the Fourier transform of $u$ . In the frequency domain, if the original function converges, the term $\frac{\partial\hat{u}}{\partial t}$ should also converge. However, when $\kappa_{e} < 0$ , this process corresponds to backward diffusion, which diverges and is inherently unstable regardless of the numerical implementation. + +We extend the analysis to our heat loss (Eq. (7)). The gradient flow is + +$$ +\frac {\partial h}{\partial t} = - \frac {\delta L _ {\text {h e a t}}}{\delta h} = \nabla^ {2} h - \lambda^ {2} h \tag {14} +$$ + +This is exactly the heat equation, and is known to be stable. We can further verify the stability by applying Fourier transform to both sides: + +$$ +\frac {\partial \hat {h}}{\partial t} (t, \omega) = - \left(| \omega | ^ {2} + \lambda^ {2}\right) \hat {h} (t, \omega) \Rightarrow \hat {h} (t, \omega) \propto e ^ {- \left(| \omega | ^ {2} + \lambda^ {2}\right) t} \tag {15} +$$ + +This analysis reveals that the original process $h(\pmb{x}, t)$ converges over time and remains stable. Due to the constraint $|u| = -\ln(h) / \lambda$ , the term $|u(\pmb{x}, t)|$ also converges during the optimization process. Since $u$ is continuous, it converges as well. Thus, our method exhibits greater spatial and temporal stability compared to the eikonal equation. + +# 4.3. Surface Area Regularization + +We next show that our heat loss (Eq. (7)) penalizes large area of the reconstructed surfaces, by showing its relation to the coarea loss used by Pumarola et al. [30]. Pumarola et al. showed that the loss $\int_{\Omega} e^{-\lambda |u|} \| \nabla u \| \, \mathrm{d}\pmb{x}$ approximates the surface area of the zero level set of the implicit function $u$ . We then form the following inequality: + +$$ +\begin{array}{l} \int_ {\Omega} e ^ {- \lambda | u |} \| \nabla u \| d \boldsymbol {x} < \int_ {\Omega} \sqrt {e ^ {- 2 \lambda | u |} \left(\| \nabla u \| ^ {2} + 1\right)} d \boldsymbol {x} (16) \\ \leqslant \sqrt {| \Omega |} \cdot \sqrt {\int_ {\Omega} e ^ {- 2 \lambda | u |} \left(\| \nabla u \| ^ {2} + 1\right) d x} (16) \\ \end{array} +$$ + +Therefore, minimizing the right hand side (our heat loss) leads to a reduction in the upper bound of the left hand side. + +Furthermore, we can show that the coarea loss is too aggressive for our purpose of signed distance field reconstruction. If we directly minimize the coarea loss (the left hand side in Eq. (16)), according to Euler-Lagrange equation, the minimizer would satisfy: + +$$ +1 - \nabla^ {2} u = 0, \tag {17} +$$ + +which does not relate to the regularization parameter $\lambda$ . Unfortunately, this partial differential equation system formed by the coarea loss above has a unique solution under Dirichlet boundary condition, and the gradient of its solution does not have a unit norm. This makes the coarea loss unable to accurately shape a signed distance function. By contrast, our heat loss has bounds for its distance approximation and converges to the true distance as we increase the parameter $\lambda$ (Eq. (9)). Another simple way to verify our convergence and the compatibility with the eikonal equation is to rewrite Eq. (4) in terms of $u$ following Eq. (6): + +$$ +\left| \left| \nabla u \right| \right| ^ {2} - 1 - \frac {\operatorname {s g n} (\mathrm {u})}{\lambda} \nabla^ {2} u = 0. \tag {18} +$$ + +![](images/e4e678153ee695bcdeac6ecac31d2699d70d595a44f3ffd69c80a350da2b3fd2.jpg) +(a) GT + +![](images/64868793a5d0bb8cc56a12997271fffcde9068ca761fb3e30a9b5d66e2d2596f.jpg) +(b) $w_{h} = 0.1$ + +![](images/7f79dd38777fbf7a13513caf181ee5d8a1150015a7603b3717186b1e5019e149.jpg) +(c) $w_{h} = 0.2$ + +![](images/576c0944855c1f97579431ecff819857a01688dcaf87b0f5e001ca47f43c43e1.jpg) +(d) $w_{h} = 0.3$ + +![](images/5af4162badb2fe6729de3b2e334b6929406fde9c157c75ddc9721877c97ac73d.jpg) +(e) $w_{b} = 0.5$ + +![](images/a686cc57f472ad0cc6a198247c85c9291ccb7197f3626d0699de39ebe985c1ab.jpg) +(f) $w_{b} = 1$ + +![](images/d1ac9ae1207d822f36749d7c7c7869cc233e0ddcf303aea168543f8e006515ee.jpg) +(g) $w_{b} = 2$ + +![](images/5d43e0c22b7d8cb496a46aad7f495b65a4ee24f39712a55a32880777cccdb404.jpg) +(h) $w_{b} = 5$ + +![](images/55af226a5059f90e1e84c1f23f25d4b17874d50ae39d371e172f0722f155573e.jpg) +(i) $w_{b} = 10$ +Figure 6. We reproduce the PHASE model [49] and show the log-transformed occupancy field, i.e., the optimized signed distance function. PHASE recommends boundary weight range $w_{b} \in (0.10, 0.32)$ based on their $\Gamma$ -convergence theory. But in practice, only a stronger boundary weight constrains the boundary to follow the input points. We found that this phenomenon often occurs and show more examples in the supplementary material. + +![](images/2bc27a3330646b9d3cb33371d3c5026e0dbfc1ab0254d79b1cfd420bebd3fea3.jpg) +(j) $w_{b} = 20$ + +# 4.4. Relation to PHASE [49] + +While derived from different mathematical principles and motivations, our method is closely related to the PHASE model proposed by Lipman. It is motivated by smoothly approximating an occupancy function $o(x)$ , which is defined as 1 outside, -1 inside, and 0 on the surface. + +In the supplement, we show that there is a close relation between our heat field $h$ and the smoothed occupancy function $o$ . We also show that our theory and derivation provide additional insights and significant differences in implementation and results. In particular, PHASE's theory introduces a condition stating that the boundary loss weight $w_{b}$ should be small, whereas our theory emphasizes the need for a firm boundary condition (Eq. (4) and Fig. 6). + +We further found that optimizing occupancy field in the PHASE model, when used together with the eikonal loss, can lead to unbounded network weight derivatives, impeding the optimization. Our method avoids this issue. + +# 5. Experiments + +We experiment on both 2D and 3D datasets. We construct a 2D dataset containing 14 shapes, ranging from simple geometric shapes to complex nested structures. We also evaluate on two 3D benchmarks: a subset of ShapeNet [55] and the Surface Reconstruction Benchmark (SRB) [56]. Additionally, we generate a dataset with high-genus geometries from Mehta et al. [1]. In all of our experiments, we use only the position information of the point cloud. + +We report the Intersection over Union (IoU), Chamfer distance, and Hausdorff distance between reconstructed shapes and ground truth shapes as metrics for surface reconstruction. In addition, we report Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and Symmetric Mean Absolute Percentage Error (SMAPE) as metrics for distance field evaluation. We further evaluate non-zero level set quality near the surface, and render the zero level + +set of the signed distance functions via sphere tracing. + +We use a network with 5 hidden layers and 128 hidden channels for our method in all experiments. The hyperparameters are kept consistent within each dataset. We additionally apply importance sampling to our heat loss integral. Full training details can be found in the supplement. + +# 5.1.2D Dataset + +An overview of our 2D dataset is attached in the supplementary material. For 2D experiments, we simply use a linear network and geometric initialization, as in DiGS [3]. We compare our method with DiGS [3] and StEik [4] with the hyperparameters and training configuration provided in their released code, and report the surface reconstruction metrics and distance field metrics in Table 1. Our method outperforms theirs across all metrics with only 8,192 sample points, whereas both DiGS [3] and StEik [4] use 15,000 sample points based on their default settings. + +# 5.2. Ablation Study on 2D Dataset + +To more comprehensively compare over different losses, we conduct an ablation study on our 2D dataset using the same network architecture, parameter initialization, and sampling strategy. The only difference lies in the loss functions used. Quantitative results are shown in Table 2. + +The last row corresponds to our HOTSPOT model, combining boundary, eikonal, and heat losses. Our method reconstructs the correct topology for all 14 shapes, while others struggle with the harder ones. It outperforms across nearly all metrics, demonstrating superior reconstruction quality and distance query accuracy. Although HOTSPOT slightly trails SAL [2] in RMSE and MAE, it captures the complex topology where SAL fails. A visualization is shown in Fig. 7. Full qualitative results and detailed experiment parameters are provided in the supplementary material. Models relying heavily on area losses [3, 4] often face a trade-off between preserving original details and eliminating redundant boundaries, which distorts the distance field (Section 4.3). Moreover, methods without our heat loss fall in local optimum easily. + +# 5.3. 3D Datasets + +We evaluate our method on a processed subset [16, 19, 57] of ShapeNet [55], which provides surface points sampled on preprocessed watertight meshes for 260 shapes across 13 categories. To ensure a reasonable spatial density for passing the heat, we scale the point cloud so that $70\%$ of the points lie within a sphere centered at the origin with a radius of 0.45. For a fair comparison, we transform our outcomes back into the coordinate system used in StEik [4] and DiGS [3]. Additionally, we implement a scheduler to gradually increase the absorption parameter $\lambda$ , which enhances the representation of level set details. + +
IoU ↑dC ↓dH ↓
DiGS [3]0.78820.00550.1267
StEik [4]0.66200.00730.1425
Ours0.98700.00140.0153
RMSE ↓MAE ↓SMAPE ↓
DiGS [3]0.05970.03150.3363
StEik [4]0.07250.04190.4222
Ours0.01990.01010.0699
+ +Table 1. Comparison of 2D reconstruction and distance query. $d_{C}$ and $d_{H}$ stand for the Chamfer and Hausdorff distances respectively. Our method outperforms DiGS [3] and StEik [4] on all metrics. + +
LBLELALDLDDLSLH| IoU ↑dC ↓dH ↓RMSE ↓MAE ↓SMAPE ↓
✓✓0.81920.00680.07120.03770.01580.1315
✓✓0.89360.00290.05790.01650.00810.1205
✓✓✓0.39070.08430.55980.29660.23881.7283
✓✓✓✓0.63380.05660.42210.20580.16161.0919
✓✓✓0.29380.03770.34440.29900.23551.6958
✓✓✓✓0.60370.04140.36560.13040.09370.7792
✓✓✓0.98510.00160.01600.01990.01050.0754
+ +Table 2. Ablation study of different loss combinations and metrics. From top to bottom, the loss combinations are from: IGR [10], SAL [2], DiGS w/o area loss, DiGS [3], StEik w/o loss, StEik [4], and our HOTSPOT. From left to right, the losses are boundary loss, eikonal loss, area loss, divergence loss, directional divergence loss, SAL loss, and heat loss. Bold and underlined data: optimal; bold only: suboptimal. Same below. + +![](images/94247f3a293ba9f4c00834afb622f0e093198f5fe9c1f91ffd97e23b4b2bd9d0.jpg) +Figure 7. Our heat loss achieves the best reconstruction of triangle fragments. Dashed lines show real boundaries, while the boldest line represents reconstructed ones. (c) and (f) are from the original models by DiGS [3] and StEik [4]; the rest are from our ablation study with different loss combinations. Other combinations fail to remove extra boundaries or accurately capture the topology. More fundamentally, (c), (f), (g), and (h) fall into local optima, leading to incorrect depressions and elevations. + +We present the quantitative results in Table 3 and visual results in Fig. 9. More baselines [46, 51] are compared in the supplement. Our HOTSPOT model outperforms current state-of-the-art models across all surface reconstruction metrics, and achieves near-top accuracy in distance queries. While SAL [2] uses the closest point distance in their loss function, leading to slightly lower total absolute error, their performance drops significantly in relative error and near-surface metrics. This suggests that the closest point distance alone is unreliable near the surface. In contrast, as mentioned at the end of Section 3, after the interpolation driven by spectral bias [58], our loss remains effective in faithfully capturing the distance to the interpolated surface. + +Fig. 1 and Fig. 8 show high-genus surface reconstruction with shapes from Mehta et al. [1]. Existing methods strug + +![](images/94f5f2ec8c7d661f1a9ad20120447a46903ab1372fc13043b58d72c717fb34e9.jpg) +DiGS + +![](images/e13d43aa3c5556beb76b1652d1eec0209d81ec4c2c70d0f5955d645630cb935f.jpg) +StEi + +![](images/6b046b7715874413a0a47c7cc94400126fd77442358555c004bbb85c872f3fed.jpg) +Ours + +![](images/51473219e1974c3d7d89b5eda798f76b03294d3ccfb47189da02b7379c8299f8.jpg) +Ground Truth + +![](images/4cc90aa9a24d5ef8849bd094c0e4db3f967f0f19c53c8a8aded6dd61ec77dcd2.jpg) + +![](images/0e38d075dfaa22550ddf3e77464741a018f298be9719ab07e82e09ee7a23ea2d.jpg) +Figure 8. Visual examples on high genus dataset. Our method uses significantly fewer iterations but reconstructs the topology correctly for all shapes, whereas others generate extra boundaries. + +![](images/06b660072fafac8a3561dee2c69f91c20118edae6c7674cc45e0a6841c2ddcdf.jpg) + +![](images/97216776a9c905a4e26418fc18109823757d679276a448d746dc55173ce82563.jpg) + +![](images/b0dc0dca7a8ae1435404d6da9cc989eecaa1ce3deb3fa2595a29a74e56d8affa.jpg) + +![](images/6613d601fe160001a5bef3c7278680164b4b79e260668de84a5f166bcdfddfd6.jpg) + +gle with the complex topology, while our method accurately reconstructs both the surface and the level set. In the supplement, we present evaluations on other high-genus shapes from Mehta et al. and the SRB dataset. + +# 5.4. Sphere Tracing + +Fig. 10 visualizes the number of steps a sphere tracer [21] takes to render the shapes obtained by SAL, DiGS, StEik, and our method. Our method has the fewest iterations on average, since it provides the most accurate level sets. More results can be found in the supplement. + +# 6. Discussion + +Efficiency. Our method uses a loss function of only the neural network output and its first-order derivatives, and thus can be more efficient than previous methods which involve second-order derivatives. We report the runtime of our method and StEik [4] for training a model from ShapeNet [55] over 100 iterations using a single NVIDIA A10 GPU with varying network sizes, while keeping all other settings identical, as shown in Tab. 4. + +Determining absorption coefficient $\lambda$ . The absorption coefficient $\lambda$ has the units of $[\mathrm{L}]^{-1}$ , where $\mathrm{L}$ denotes a length unit. Rescaling the space is equivalent to rescaling $\lambda$ . While a larger $\lambda$ can improve accuracy, it introduces two challenges: (1) our heat loss includes the factor $e^{-2\lambda |u|}$ , which may become smaller than the precision limit of floating-point numbers, leading to round-off errors; (2) even if round-off errors are ignored, a large $\lambda |u|$ may reduce the optimizer's ability to effectively shape the field. Never + +![](images/c9cdbbe5acae0741015467a61070f3a6dc308ec795d0542b217896e9e4e51dbf.jpg) +Figure 9. Visual results of DiGS [3], StEik [4], and HOTSPOT on ShapeNet [55]. Previous methods generate extra boundaries and fail to escape local optima, while our method successfully reconstructs the correct topology. + +
IoU ↑dC ↓dH ↓RMSE ↓MAE ↓SMAPE ↓RMSE0.1MAE0.1SMAPE0.1
SAL [2]0.74000.00740.08510.02510.01420.13440.02450.01820.6848
SIREN wo/ n [34]0.48740.00510.05580.50090.42611.26940.05130.03820.8858
DiGS [3]0.96360.00310.04350.11940.07250.21400.01520.00810.1760
StEik [4]0.96410.00320.03680.03870.02480.09310.01470.00810.1770
Ours0.97960.00290.02500.02810.01760.05400.00940.00470.1206
+ +![](images/a68e5841683847a7d1fcf5c211e241cf47ccfe1a61181eebf86458bd14e70734.jpg) +Figure 10. Visualizations and statistics of sphere tracing [21] iteration counts required to locate the surface. In the visual results brighter pixels indicate higher iteration counts. The histograms show the distribution of iterations for all pixels across 10 different camera poses. HOTSPOT finds the surface in fewer steps. + +theless, our loss still pushes $u$ away from 0, preventing unwanted surfaces where $\lambda |u|$ should be large. We compensate by combining eikonal loss with heat losses. When $\lambda |u|$ is large, the eikonal loss becomes dominant. Our $\lambda$ scheduler also effectively shapes the distant regions. Spatially-adaptive parameters may also improve the results. + +Necessity of firm boundary condition. The heat diffu + +Table 3. Surface reconstruction metrics on ShapeNet [55], where $d_{C}$ and $d_{H}$ stand for the Chamfer and Hausdorff distances, RMSE, MAE, and SMAPE are distance field metrics, and $\mathrm{RMSE}_{0.1}$ , $\mathrm{MAE}_{0.1}$ , and $\mathrm{SMAPE}_{0.1}$ are the same distance field metrics but only for points within distance 0.1 of the surface. See the supplementary material for detailed results with more baselines. + +
Structure5 × 1285 × 2568 × 1288 × 256
StEik [4]5.65s10.62s8.78s17.00s
Ours4.49s8.43s6.79s12.86s
+ +Table 4. Runtime for 100 iterations with different network sizes. + +sion of the screened Poisson equation is based on a well-established boundary condition. We enforce the boundary condition through the boundary loss over discrete points, with neural networks interpolating through spectral bias [58]. When the input points are sparse and the absorption coefficient $\lambda$ is high, an overly strong heat loss ( $w_{b}$ being too small) can tear the boundary, causing the signed distance to collapse into an unsigned one like Fig. 6. Thus, our theory highlights the importance of a high boundary weight $w_{b}$ , while absorption $\lambda$ is adjusted to match the input point density or by scaling the point cloud itself. Our experiments show that proper parameter setting and rescaling prevent the collapse. Future research is required for very sparse boundaries, and applications to inverse rendering. + +# 7. Conclusion + +We propose a new model for neural signed distance function optimization based on the screened Poisson equation. We analyze our loss theoretically and show that it is an asymptotically sufficient condition to the true distance, is stable both to a small perturbation and in the temporal dynamics, and penalizes large surface area. Our experiments show that we reconstruct both better surfaces and better distance approximations compared to many existing methods, especially on complex and high-genus shapes. + +Acknowledgements. This work was supported in part by NSF grants 2127544, 2238839, 2100237, 2120019, and gifts from Adobe and Google. Additionally, we would like to thank Bing Xu, Zilu Li, Yash Belhe, Ishit Mehta, and Jianan Xiao for their suggestions, discussions, proofreading, and providing reference materials. + +# References + +[1] Ishit Mehta, Manmohan Chandraker, and Ravi Ramamoorthi. A level set theory for neural implicit evolution under explicit flows. In ECCV, pages 711-729, 2022. 1, 2, 6, 7, 8, 10 +[2] Matan Atzmon and Yaron Lipman. Sal: Sign agnostic learning of shapes from raw data. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2565-2574, 2020. 1, 2, 6, 7, 8, 3, 5 +[3] Yizhak Ben-Shabat, Chamin Hewa Koneputugodage, and Stephen Gould. DiGS: Divergence guided shape implicit neural representation for unoriented point clouds. In IEEE Conf. Comput. Vis. Pattern Recog., pages 19323-19332, 2022. 1, 2, 3, 6, 7, 8, 5 +[4] Huizong Yang, Yuxin Sun, Ganesh Sundaramoorthi, and Anthony Yezzi. StEik: stabilizing the optimization of neural signed distance functions and finer shape representation. arXiv preprint arXiv:2305.18414, 2023. 1, 2, 3, 5, 6, 7, 8 +[5] Zoë Marschner, Silvia Sellán, Hsueh-Ti Derek Liu, and Alec Jacobson. Constructive solid geometry on neural signed distance fields. In SIGGRAPH Asia 2023 Conference Papers, pages 1-12, 2023. 1, 2, 3 +[6] Sathamangalam R Srinivasa Varadhan. On the behavior of the fundamental solution of the heat equation with variable coefficients. Communications on Pure and Applied Mathematics, 20(2):431-455, 1967. 2, 3, 4 +[7] Keenan Crane, Clarisse Weischedel, and Max Wardetzky. Geodesics in heat: A new approach to computing distance based on heat flow. ACM TOG, 32(5):1-11, 2013. 2, 3 +[8] Alexander G Belyaev and Pierre-Alain Fayolle. On variational and pde-based distance function approximations. In Computer Graphics Forum, volume 34, pages 104-118. Wiley Online Library, 2015. +[9] Alexander Belyaev and Pierre-Alain Fayolle. An ADMM-based scheme for distance function approximation. Numerical Algorithms, 84:983-996, 2020. 2, 3 +[10] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. arXiv preprint arXiv:2002.10099, 2020. 2, 3, 7 +[11] James F Blinn. A generalization of algebraic surface drawing. ACM TOG, 1(3):235-256, 1982. 2 +[12] Jules Bloomenthal and Brian Wyvill. Interactive techniques for implicit modeling. Comput. Graph. (Proc. SIGGRAPH), 24(2):109-116, 1990. +[13] Hugues Hoppe, Tony DeRose, Tom Duchamp, John McDonald, and Werner Stuetzle. Surface reconstruction from unorganized points. Comput. Graph. (Proc. SIGGRAPH), 26(2):71-78, 1992. +[14] Brian Curless and Marc Levoy. A volumetric method for building complex models from range images. In SIGGRAPH, page 303-312, 1996. + +[15] Shahram Izadi, David Kim, Otmar Hilliges, David Molyneaux, Richard Newcombe, Pushmeet Kohli, Jamie Shotton, Steve Hodges, Dustin Freeman, and Andrew Davison. KinectFusion: real-time 3D reconstruction and interaction using a moving depth camera. In ACM Symposium on User Interface Software and Technology, pages 559-568, 2011. 2 +[16] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4460-4470, 2019. 2, 6, 7 +[17] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5939-5948, 2019. +[18] Thomas Davies, Derek Nowrouzezahrai, and Alec Jacobson. On the effectiveness of weight-encoded neural implicit 3d shapes. arXiv preprint arXiv:2009.09808, 2020. +[19] Francis Williams, Matthew Trager, Joan Bruna, and Denis Zorin. Neural Splines: Fitting 3D surfaces with infinitely-wide neural networks. In IEEE Conf. Comput. Vis. Pattern Recog., pages 9949-9958, 2021. 6 +[20] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. 41(2):641-676, 2022. 2 +[21] John C Hart. Sphere tracing: A geometric method for the antialiased ray tracing of implicit surfaces. The Vis. Comput., 12(10):527-545, 1996. 2, 7, 8 +[22] Dario Seyb, Alec Jacobson, Derek Nowrouzezahrai, and Wojciech Jarosz. Non-linear sphere tracing for rendering deformed signed distance fields. ACM Trans. Graph. (Proc. SIGGRAPH Asia), 38(6), 2019. +[23] Eric Galin, Eric Guérin, Axel Paris, and Adrien Peytavie. Segment tracing using local Lipschitz bounds, 2020. +[24] Nicholas Sharp and Alec Jacobson. Spelunking the deep: guaranteed queries on general neural implicit surfaces via range analysis. ACM Trans. Graph. (Proc. SIGGRAPH), 41(4), 2022. 2 +[25] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3D shapes. In IEEE Conf. Comput. Vis. Pattern Recog., pages 11358-11367, 2021. 2 +[26] Arnulph Fuhrmann, Gerrit Sobotka, and Clemens Groß. Distance fields for rapid collision detection in physically based modeling. In Proceedings of GraphiCon, volume 2003, pages 58-65, 2003. 2 +[27] Eran Guendelman, Robert Bridson, and Ronald Fedkiw. Nonconvex rigid bodies with stacking. ACM TOG, 22(3):871-878, 2003. +[28] Miles Macklin, Kenny Erleben, Matthias Müller, Nuttapong Chentanez, Stefan Jeschke, and Zach Corse. Local optimization for robust signed distance field collision. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 3(1):1-17, 2020. 2 +[29] Rohan Sawhney and Keenan Crane. Monte Carlo geometry processing: a grid-free approach to PDE-based methods on + +volumetric domains. ACM Trans. Graph. (Proc. SIGGRAPH), 39(4), 2020. 2 +[30] Albert Pumarola, Artsiom Sanakoyeu, Lior Yariv, Ali Thabet, and Yaron Lipman. Visco grids: Surface reconstruction with viscosity and coarea grids. Advances in Neural Information Processing Systems, 35:18060-18071, 2022. 2, 5 +[31] Songyou Peng, Chiyu Jiang, Yiyi Liao, Michael Niemeyer, Marc Pollefeys, and Andreas Geiger. Shape as points: A differentiable poisson solver. Advances in Neural Information Processing Systems, 34:13032-13044, 2021. +[32] Chao Chen, Yu-Shen Liu, and Zhizhong Han. Gridpull: Towards scalability in learning implicit representations from 3d point clouds. In Proceedings of the IEEE/cvf international conference on computer vision, pages 18322-18334, 2023. 2 +[33] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. DeepSDF: Learning continuous signed distance functions for shape representation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 165-174, 2019. 2 +[34] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in neural information processing systems, 33:7462-7473, 2020. 2, 8, 5, 7 +[35] Guillaume Coiffier and Louis Béthune. 1-Lipschitz neural distance fields. Comput. Graph. Forum (Eurographics STARs), 43(5):e15128, 2024. 2 +[36] Yue Jiang, Dantong Ji, Zhizhong Han, and Matthias Zwicker. SDFDiff: Differentiable rendering of signed distance fields for 3D shape optimization. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1251-1261, 2020. 2 +[37] Petr Kellnhofer, Lars C Jebe, Andrew Jones, Ryan Spicer, Kari Pulli, and Gordon Wetzstein. Neural lumigraph rendering. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4287-4297, 2021. +[38] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3D representations without 3D supervision. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3504-3515, 2020. +[39] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. In Advances in Neural Information Processing Systems, volume 33, pages 2492-2502, 2020. 2, 8 +[40] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snavely. PhySG: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5453-5462, 2021. +[41] Sai Bangaru, Michael Gharbi, Tzu-Mao Li, Fujun Luan, Kalyan Sunkavalli, Milos Hasan, Sai Bi, Zexiang Xu, Gilbert Bernstein, and Fredo Durand. Differentiable rendering of neural sdfs through reparameterization. In SIGGRAPH Asia Conference Proceedings, 2022. +[42] Delio Vicini, Sébastien Speierer, and Wenzel Jakob. Differentiable signed distance function rendering. ACM Trans. Graph. (Proc. SIGGRAPH), 41(4):1-18, 2022. 2 + +[43] Jingyang Zhang, Yao Yao, Shiwei Li, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Critical regularizations for neural surface reconstruction in the wild. In IEEE Conf. Comput. Vis. Pattern Recog., pages 6270-6279, 2022. 2 +[44] Ruian Wang, Zixiong Wang, Yunxiao Zhang, Shuangmin Chen, Shiqing Xin, Changhe Tu, and Wenping Wang. Aligning gradient and hessian for neural signed distance function. In Advances in Neural Information Processing Systems, volume 36, 2024. +[45] Zixiong Wang, Yunxiao Zhang, Rui Xu, Fan Zhang, Peng-Shuai Wang, Shuangmin Chen, Shiqing Xin, Wenping Wang, and Changhe Tu. Neural-singular-hessian: Implicit neural representation of unoriented point clouds by enforcing singular hessian. ACM Transactions on Graphics (TOG), 42(6):1-14, 2023. 5, 7, 8 +[46] Qiujie Dong, Rui Xu, Pengfei Wang, Shuangmin Chen, Shiqing Xin, Xiaohong Jia, Wenping Wang, and Changhe Tu. Neurcadrecon: Neural representation for reconstructing cad surfaces by enforcing zero gaussian curvature. arXiv preprint arXiv:2404.13420, 2024. 2, 7 +[47] Yesom Park, Taekyung Lee, Jooyoung Hahn, and Myungjoo Kang. $p$ -poisson surface reconstruction in curl-free flow from point clouds. In Advances in Neural Information Processing Systems, volume 36, 2024. 2 +[48] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In IEEE Conf. Comput. Vis. Pattern Recog., pages 8456-8465, 2023. 2 +[49] Yaron Lipman. Phase transitions, distance functions, and implicit neural representations. In International Conference on Machine Learning, volume 139, pages 6702-6712, 2021. 2, 3, 6, 8, 9 +[50] Angela Dai and Matthias Nießner. Neural poisson: Indicator functions for neural fields. arXiv preprint arXiv:2211.14249, 2022. 2 +[51] Baorui Ma, Zhizhong Han, Yu-Shen Liu, and Matthias Zwicker. Neural-pull: Learning signed distance functions from point clouds by learning to pull space onto surfaces. arXiv preprint arXiv:2011.13495, 2020. 2, 7, 3, 5, 8 +[52] Matan Atzmon and Yaron Lipman. SALD: Sign agnostic learning with derivatives. In ICLR, 2021. 2, 3 +[53] Nicholas Sharp, Yousuf Soliman, and Keenan Crane. The vector heat method. ACM TOG, 38(3):1-19, 2019. 3 +[54] Nicole Feng and Keenan Crane. A heat method for generalized signed distance. ACM Trans. Graph. (Proc. SIGGRAPH), 43(4), 2024. 3 +[55] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 6, 7, 8, 5, 9 +[56] Matthew Berger, Joshua A Levine, Luis Gustavo Nonato, Gabriel Taubin, and Claudio T Silva. A benchmark for surface reconstruction. ACM Transactions on Graphics (TOG), 32(2):1-17, 2013. 6, 8 + +[57] David Stutz and Andreas Geiger. Learning 3d shape completion under weak supervision. International Journal of Computer Vision, 128:1162-1181, 2020. 6, 7 +[58] Nasim Rahaman, Aristide Baratin, Devansh Arpit, Felix Draxler, Min Lin, Fred Hamprecht, Yoshua Bengio, and Aaron Courville. On the spectral bias of neural networks. In International Conference on Machine Learning, pages 5301-5310, 2019. 7, 8, 3, 5, 11 +[59] Dmitry S Kulyabov, Anna V Korolkova, Tatiana R Velieva, and Migran N Gevorkyan. Numerical analysis of eikonal equation. In Saratov Fall Meeting 2018: Laser Physics, Photonic Technologies, and Molecular Modeling, volume 11066, pages 188-195. SPIE, 2019. 1 +[60] Zhen-Hang Yang and Yu-Ming Chu. On approximating the modified bessel function of the second kind. Journal of inequalities and applications, 2017:1-8, 2017. 2 +[61] Luciano Modica. The gradient theory of phase transitions and the minimal interface criterion. Archive for Rational Mechanics and Analysis, 98:123-142, 1987. 10 +[62] Peter Sternberg. The effect of a singular perturbation on nonconvex variational problems. Archive for Rational Mechanics and Analysis, 101:209-260, 1988. 10 \ No newline at end of file diff --git a/2025/HotSpot_ Signed Distance Function Optimization with an Asymptotically Sufficient Condition/images.zip b/2025/HotSpot_ Signed Distance Function Optimization with an Asymptotically Sufficient Condition/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..77986953471c6423d633b695b830bfba94e695eb --- /dev/null +++ b/2025/HotSpot_ Signed Distance Function Optimization with an Asymptotically Sufficient Condition/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c906bbb2d22721221044b6137ee6c0241d9a909d4696b4cc6e74288f6d9e7c4d +size 602976 diff --git a/2025/HotSpot_ Signed Distance Function Optimization with an Asymptotically Sufficient Condition/layout.json b/2025/HotSpot_ Signed Distance Function Optimization with an Asymptotically Sufficient Condition/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..834da5b4e72630095d8b326aeab37d2249813e78 --- /dev/null +++ b/2025/HotSpot_ Signed Distance Function Optimization with an Asymptotically Sufficient Condition/layout.json @@ -0,0 +1,12078 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 72, + 102, + 522, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 102, + 522, + 138 + ], + "spans": [ + { + "bbox": [ + 72, + 102, + 522, + 138 + ], + "type": "text", + "content": "HotSpot: Signed Distance Function Optimization with an Asymptotically Sufficient Condition" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 88, + 160, + 504, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 160, + 504, + 190 + ], + "spans": [ + { + "bbox": [ + 88, + 160, + 504, + 190 + ], + "type": "text", + "content": "Zimo Wang* Cheng Wang* Taiki Yoshino Sirui Tao Ziyang Fu Tzu-Mao Li UC San Diego" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 51, + 202, + 544, + 312 + ], + "blocks": [ + { + "bbox": [ + 51, + 202, + 544, + 312 + ], + "lines": [ + { + "bbox": [ + 51, + 202, + 544, + 312 + ], + "spans": [ + { + "bbox": [ + 51, + 202, + 544, + 312 + ], + "type": "image", + "image_path": "74e2cc9cebace942b87899a27c1858b46acc7b83ef83fb6f154ef3b51db6ab2f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 320, + 546, + 376 + ], + "lines": [ + { + "bbox": [ + 46, + 320, + 546, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 320, + 546, + 376 + ], + "type": "text", + "content": "Figure 1. We propose HOTSPOT, a neural signed distance function optimization method that establishes an asymptotic sufficient condition to guarantee convergence to a true distance function, enabling precise surface reconstruction and level set representation for complex shapes. Here we show a reconstruction from a point cloud sampled from the reference bunny (taken from Mehta et al. [1]) on the right. In the inset, we visualize the recovered signed distance function on a horizontal slice, using warm colors for positive values and cool for negative (zoom in for details). Our reconstruction is significantly more accurate than prior works (SAL [2], DiGS [3], and StEik [4])." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 393, + 192, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 393, + 192, + 406 + ], + "spans": [ + { + "bbox": [ + 143, + 393, + 192, + 406 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 418, + 289, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 418, + 289, + 646 + ], + "spans": [ + { + "bbox": [ + 46, + 418, + 289, + 646 + ], + "type": "text", + "content": "We propose a method, HOTSPOT, for optimizing neural signed distance functions. Existing losses, such as the eikonal loss, act as necessary but insufficient constraints and cannot guarantee that the recovered implicit function represents a true distance function, even if the output minimizes these losses almost everywhere. Furthermore, the eikonal loss suffers from stability issues in optimization. Finally, in conventional methods, regularization losses that penalize surface area distort the reconstructed signed distance function. We address these challenges by designing a loss function using the solution of a screened Poisson equation. Our loss, when minimized, provides an asymptotically sufficient condition to ensure the output converges to a true distance function. Our loss also leads to stable optimization and naturally penalizes large surface areas. We present theoretical analysis and experiments on both challenging 2D and 3D datasets and show that our method provides better surface reconstruction and a more accurate distance approximation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 650, + 128, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 650, + 128, + 662 + ], + "spans": [ + { + "bbox": [ + 47, + 650, + 128, + 662 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 670, + 288, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 670, + 288, + 694 + ], + "spans": [ + { + "bbox": [ + 46, + 670, + 288, + 694 + ], + "type": "text", + "content": "Learning neural signed distance function requires designing loss functions so that the zero crossings of the implicit" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 394, + 545, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 394, + 545, + 478 + ], + "spans": [ + { + "bbox": [ + 306, + 394, + 545, + 478 + ], + "type": "text", + "content": "function are at the desired locations (e.g., adhere to a sparse input point cloud), and that the implicit function returns the correct signed distance. Designing the loss function proves to be challenging for shapes with complex topology and details (Fig. 1). We propose HOTSPOT, a method based on a relation between the screened Poisson equation and distance that enjoys both theoretical and practical benefits." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 481, + 546, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 481, + 546, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 481, + 546, + 601 + ], + "type": "text", + "content": "A major challenge in signed distance function optimization is ensuring the implicit function follows the true distance. A common regularization loss used is the eikonal equation, which constrains the gradient norm of an implicit function to be 1 almost everywhere. If the implicit function is a signed distance function, then it satisfies the eikonal equation. However, the converse is not true. That is, satisfying the eikonal equation is only a necessary condition for the signed distance function, not a sufficient condition [5]. Fig. 2 presents the logic and Fig. 3 shows an example." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 603, + 545, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 603, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 603, + 545, + 651 + ], + "type": "text", + "content": "Another key challenge lies in the optimization. It is known that the eikonal equation regularization is unstable [4], and the instability can make optimization converge to suboptimal results or even diverge." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "Furthermore, to address optimization ill-posedness and instability, existing methods penalize the surface area of the shape. We prove this degrades the reconstruction quality and distorts the output level sets, thereby leading to a delicate balance between over- and under-smoothing." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 126, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 126, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 126, + 712 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1276" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 51, + 71, + 284, + 113 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 71, + 284, + 113 + ], + "spans": [ + { + "bbox": [ + 51, + 71, + 284, + 113 + ], + "type": "interline_equation", + "content": "\\begin{array}{c} \\min \\{\\text {h e a t l o s s} \\} \\\\ \\lambda \\to \\infty \\end{array} \\Leftrightarrow \\begin{array}{l} \\text {D i s t a n c e} \\quad \\text {a . e .} \\\\ \\text {F u n c t i o n} \\quad \\stackrel {{\\Rightarrow}} {{\\Leftarrow}} \\end{array} \\left\\{ \\begin{array}{l} \\text {E i k o n a l c o n d i t i o n} \\\\ \\text {Z e r o g r a d i e n t d i v e r g e n c e} \\\\ \\text {S i n g u l a r H e s s i a n m a t r i x} \\\\ \\dots \\end{array} \\right.", + "image_path": "02635dd2be0d9ac643b746e07062537c893fb4d5b779164b9562078ff778463d.jpg" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 216, + 287, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 216, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 216, + 287, + 312 + ], + "type": "text", + "content": "We propose a simple model that can address all the challenges above. Our model, HOTSPOT, is based on a screened Poisson equation and a classical relation between heat transfer and distance [6] that was popular for approximating both Euclidean and geodesic distance [7-9]. Using this relation, we design a loss function for optimizing neural signed distance functions. As our contributions, we show, both theoretically and empirically, that our model" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 312, + 287, + 479 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 47, + 312, + 287, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 312, + 287, + 358 + ], + "spans": [ + { + "bbox": [ + 47, + 312, + 287, + 358 + ], + "type": "text", + "content": "- serves as a sufficient condition ensuring the difference between the minimizer and the actual distance is bounded and decreases linearly as the parameter " + }, + { + "bbox": [ + 47, + 312, + 287, + 358 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 47, + 312, + 287, + 358 + ], + "type": "text", + "content": " increases (Fig. 2 and Fig. 3, right)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 359, + 287, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 359, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 47, + 359, + 287, + 418 + ], + "type": "text", + "content": "- is stable both spatially and temporally, in the sense that a slight perturbation of the implicit function only introduces changes in a local region, and that the dynamical system formed by the gradient flow will converge in the long run." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 419, + 287, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 419, + 287, + 442 + ], + "spans": [ + { + "bbox": [ + 47, + 419, + 287, + 442 + ], + "type": "text", + "content": "- penalizes the surface area naturally, while remaining an undistorted signed distance function." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 443, + 287, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 443, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 47, + 443, + 287, + 479 + ], + "type": "text", + "content": "- works well in both 2D and 3D shapes, and it excels at approaching signed distance functions while optimizing for complex, high-genus shapes." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 490, + 134, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 490, + 134, + 502 + ], + "spans": [ + { + "bbox": [ + 47, + 490, + 134, + 502 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 510, + 287, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 594 + ], + "type": "text", + "content": "Implicit surfaces. Modeling and reconstructing surfaces using implicit functions has a long history in computer vision and graphics [11-15]. Recently, there is a surge of interest to model surfaces using neural implicit functions [16-20]. Neural implicit representations are favored for their compactness, ability to represent detailed surfaces, and compatibility to gradient-based optimization [1]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": "Neural signed distance functions. Among the classes of implicit surfaces, signed distance functions, where the implicit function outputs the distance to the closest point on the surface, and the sign represents whether the point is inside or outside the solid object, is one of the most popular variants. Signed distance functions enable an efficient rendering algorithm called sphere tracing [21-24] and helps geometry compression [25]. They can also be used for collision detection [26-28] and numerical simulation [29]. Traditionally, signed distance functions are often represented" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 337, + 71, + 512, + 269 + ], + "blocks": [ + { + "bbox": [ + 46, + 116, + 288, + 204 + ], + "lines": [ + { + "bbox": [ + 46, + 116, + 288, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 116, + 288, + 204 + ], + "type": "text", + "content": "Figure 2. Existing constraints are only necessary but insufficient conditions. That is, there exist solutions, like Fig. 3, that satisfy them almost everywhere yet fail to be distance functions. In contrast, we use a heat loss, modeled by a screened Poisson equation with parameter " + }, + { + "bbox": [ + 46, + 116, + 288, + 204 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 116, + 288, + 204 + ], + "type": "text", + "content": ". As the parameter " + }, + { + "bbox": [ + 46, + 116, + 288, + 204 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 116, + 288, + 204 + ], + "type": "text", + "content": " goes to infinity, the minimizer converges to the distance function and excludes other solutions. Thus, our method is asymptotically both necessary and sufficient." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 337, + 71, + 512, + 269 + ], + "lines": [ + { + "bbox": [ + 337, + 71, + 512, + 269 + ], + "spans": [ + { + "bbox": [ + 337, + 71, + 512, + 269 + ], + "type": "image", + "image_path": "b06f6a187e16cd03ca5cd591c9a9b1a69f7c7a7b9b8bab6a8d473ec08cb4d766.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 273, + 545, + 361 + ], + "lines": [ + { + "bbox": [ + 305, + 273, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 305, + 273, + 545, + 361 + ], + "type": "text", + "content": "Figure 3. We illustrate a 1D neural signed distance function optimization using the classical eikonal loss [10] and our model. The " + }, + { + "bbox": [ + 305, + 273, + 545, + 361 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 305, + 273, + 545, + 361 + ], + "type": "text", + "content": " axis is the domain and the " + }, + { + "bbox": [ + 305, + 273, + 545, + 361 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 305, + 273, + 545, + 361 + ], + "type": "text", + "content": " axis shows the output of the implicit function. The middle and bottom rows show the intermediate and final states of the optimization respectively. The eikonal loss, as a necessary but insufficient condition, is incapable of converging to the actual signed distance function (dashed line), even when the function satisfies the eikonal equation almost everywhere." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 373, + 545, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 545, + 433 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 545, + 433 + ], + "type": "text", + "content": "as grids [30-32]. More recently, neural signed distance functions have become an appealing geometric representation [33-35]. (Neural) Signed distance functions are also often used in inverse rendering [36-42], for their flexibility to obtain surfaces with complex topology." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 435, + 546, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 435, + 546, + 627 + ], + "spans": [ + { + "bbox": [ + 304, + 435, + 546, + 627 + ], + "type": "text", + "content": "Regularizing and initializing neural signed distance functions. A commonly used loss for optimizing neural signed distance functions is the eikonal regularization, which encourages the norm of the gradient of the implicit function to be one [10, 39]. However, the eikonal regularization alone is far from sufficient, as 1) minimizing eikonal loss does not necessarily lead to a distance field [5, 30], and 2) its optimization is unstable [4]. Previous works have proposed initialization to bootstrap the optimization with a simple shape [3, 10]. Recent works show that minimizing the divergence and the surface area of the implicit function can lead to better reconstruction [3, 43] and stabilizes the optimization [4]. Second-order information is also often used [43-46] for regularization. All losses above are based on necessary properties of real distance but neither establish convergence nor exclude non-SDF solutions." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": "Specialized architectures have been proposed to limit the Lipschitz constant of the implicit function [35], satisfy " + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": "-Poisson equation [47], or represent high-frequency details [25, 34, 48]. Some research [49, 50] has shown a link between occupancy function optimization and an approximate signed distance field regularization. Our experiments further reveal that some implementations [2, 51, 52]" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1277" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "content": "optimizing SDFs with the closest point may struggle when querying near the interpolated surface." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 96, + 287, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 96, + 287, + 178 + ], + "spans": [ + { + "bbox": [ + 46, + 96, + 287, + 178 + ], + "type": "text", + "content": "We propose a new loss and method for optimizing signed distance functions. While derived from very different mathematics, our loss turns out to have a close connection to Lipman's PHASE loss [49]. We show that our theory leads to significantly more numerically stable optimization, and a different conclusion in the parameter setting that leads to drastically different results." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 180, + 287, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 180, + 287, + 263 + ], + "spans": [ + { + "bbox": [ + 46, + 180, + 287, + 263 + ], + "type": "text", + "content": "Distance approximation using heat transfer. Our derivation is inspired by a relation between heat transfer and approximated distance [6]. This relation has been used in computer graphics for efficiently computing the geodesic distance on surfaces [7-9, 53, 54]. We apply the relation for optimizing neural signed distance functions, and analyze the properties of the resulting optimization in this context." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 274, + 207, + 286 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 274, + 207, + 286 + ], + "spans": [ + { + "bbox": [ + 47, + 274, + 207, + 286 + ], + "type": "text", + "content": "3. Background and Motivation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 293, + 287, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 293, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 287, + 342 + ], + "type": "text", + "content": "Given an input point cloud " + }, + { + "bbox": [ + 46, + 293, + 287, + 342 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i\\in \\mathbb{R}^d" + }, + { + "bbox": [ + 46, + 293, + 287, + 342 + ], + "type": "text", + "content": " (we focus on 2D and 3D in this work) with " + }, + { + "bbox": [ + 46, + 293, + 287, + 342 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 293, + 287, + 342 + ], + "type": "text", + "content": " points without normal information that lies on an unknown surface " + }, + { + "bbox": [ + 46, + 293, + 287, + 342 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 46, + 293, + 287, + 342 + ], + "type": "text", + "content": " of a solid object " + }, + { + "bbox": [ + 46, + 293, + 287, + 342 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 293, + 287, + 342 + ], + "type": "text", + "content": ", our goal is to find a signed distance function " + }, + { + "bbox": [ + 46, + 293, + 287, + 342 + ], + "type": "inline_equation", + "content": "f(\\pmb {x})" + }, + { + "bbox": [ + 46, + 293, + 287, + 342 + ], + "type": "text", + "content": ", such that" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 99, + 349, + 287, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 349, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 99, + 349, + 287, + 380 + ], + "type": "interline_equation", + "content": "f (\\boldsymbol {x}) = \\left\\{ \\begin{array}{l l} - d _ {\\Gamma} (\\boldsymbol {x}) & \\text {i f} \\boldsymbol {x} \\in S \\\\ d _ {\\Gamma} (\\boldsymbol {x}) & \\text {o t h e r w i s e ,} \\end{array} \\right. \\tag {1}", + "image_path": "002d7003bd9753902a9fbc1b34185504197b06d87ed0ce13ab835bc0062bd395.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 387, + 287, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 387, + 287, + 472 + ], + "spans": [ + { + "bbox": [ + 46, + 387, + 287, + 472 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 387, + 287, + 472 + ], + "type": "inline_equation", + "content": "d_{\\Gamma}(\\pmb{x})" + }, + { + "bbox": [ + 46, + 387, + 287, + 472 + ], + "type": "text", + "content": " denotes the closest distance from point " + }, + { + "bbox": [ + 46, + 387, + 287, + 472 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 46, + 387, + 287, + 472 + ], + "type": "text", + "content": " to the surface " + }, + { + "bbox": [ + 46, + 387, + 287, + 472 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 46, + 387, + 287, + 472 + ], + "type": "text", + "content": ". In practice, we approximate the signed distance function " + }, + { + "bbox": [ + 46, + 387, + 287, + 472 + ], + "type": "inline_equation", + "content": "f(\\pmb{x})" + }, + { + "bbox": [ + 46, + 387, + 287, + 472 + ], + "type": "text", + "content": " using a neural network " + }, + { + "bbox": [ + 46, + 387, + 287, + 472 + ], + "type": "inline_equation", + "content": "u(\\pmb{x}): \\mathbb{R}^d \\to \\mathbb{R}" + }, + { + "bbox": [ + 46, + 387, + 287, + 472 + ], + "type": "text", + "content": " parameterized by its weights and use gradient-based optimization to figure out the weights. Previous work (e.g., [10]) often uses a linear combination of a boundary loss " + }, + { + "bbox": [ + 46, + 387, + 287, + 472 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{boundary}}" + }, + { + "bbox": [ + 46, + 387, + 287, + 472 + ], + "type": "text", + "content": " and a eikonal loss " + }, + { + "bbox": [ + 46, + 387, + 287, + 472 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{eikonal}}" + }, + { + "bbox": [ + 46, + 387, + 287, + 472 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 60, + 479, + 287, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 479, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 60, + 479, + 287, + 510 + ], + "type": "interline_equation", + "content": "L _ {\\text {b o u n d a r y}} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} | u (\\boldsymbol {x} _ {i}) | ^ {p}, \\tag {2}", + "image_path": "e48af5e7554b5f344b39732c21853a4b20380d41ad865979320e023ac368ba8b.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 514, + 287, + 539 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 514, + 287, + 539 + ], + "spans": [ + { + "bbox": [ + 67, + 514, + 287, + 539 + ], + "type": "interline_equation", + "content": "L _ {\\text {e i k o n a l}} = \\int_ {\\Omega} | \\| \\nabla u (\\boldsymbol {x}) \\| - 1 | ^ {p} \\mathrm {d} \\boldsymbol {x}, \\quad p = 1 \\text {o r} 2, \\tag {3}", + "image_path": "7558fb66d85c02d7e5ad90dd8caf600f8b0756186703429a24774e5a48910be9.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 545, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 617 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 545, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\Omega \\subset \\mathbb{R}^d" + }, + { + "bbox": [ + 46, + 545, + 287, + 617 + ], + "type": "text", + "content": " is our domain of interest. The boundary loss encourages the implicit function " + }, + { + "bbox": [ + 46, + 545, + 287, + 617 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 46, + 545, + 287, + 617 + ], + "type": "text", + "content": " to output 0 at the boundary, and the eikonal loss is based on that observation that if a function " + }, + { + "bbox": [ + 46, + 545, + 287, + 617 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 545, + 287, + 617 + ], + "type": "text", + "content": " is a distance function, the " + }, + { + "bbox": [ + 46, + 545, + 287, + 617 + ], + "type": "inline_equation", + "content": "L^2" + }, + { + "bbox": [ + 46, + 545, + 287, + 617 + ], + "type": "text", + "content": " norm of its gradient " + }, + { + "bbox": [ + 46, + 545, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\| \\nabla f(\\pmb{x})\\| = 1" + }, + { + "bbox": [ + 46, + 545, + 287, + 617 + ], + "type": "text", + "content": " holds almost everywhere. Thus the approximation " + }, + { + "bbox": [ + 46, + 545, + 287, + 617 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 46, + 545, + 287, + 617 + ], + "type": "text", + "content": " is encouraged to behave the same." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "type": "text", + "content": "However, as noted, simply optimizing for the two terms above does not guarantee that " + }, + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "type": "text", + "content": " would be a distance function even when satisfied almost everywhere (Figs. 2 and 3), and the extra regularization previous work added to minimize the area and divergence [3, 4] could further tamper with the distance approximation quality. We derive an additional loss that can be used together with both of the losses above, and analyze our loss theoretically and empirically." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 310, + 70, + 425, + 161 + ], + "blocks": [ + { + "bbox": [ + 310, + 70, + 425, + 161 + ], + "lines": [ + { + "bbox": [ + 310, + 70, + 425, + 161 + ], + "spans": [ + { + "bbox": [ + 310, + 70, + 425, + 161 + ], + "type": "image", + "image_path": "5dfab8084f9c63b1361b5c2879bf2d56d0ffec1bd6d0b09daced5bc564ec7969.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 426, + 70, + 542, + 161 + ], + "blocks": [ + { + "bbox": [ + 426, + 70, + 542, + 161 + ], + "lines": [ + { + "bbox": [ + 426, + 70, + 542, + 161 + ], + "spans": [ + { + "bbox": [ + 426, + 70, + 542, + 161 + ], + "type": "image", + "image_path": "9d07dff7616f91a37d1a713146f880761e116e3bcd78cc309afe6228b64be931.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 310, + 176, + 363, + 232 + ], + "blocks": [ + { + "bbox": [ + 310, + 176, + 363, + 232 + ], + "lines": [ + { + "bbox": [ + 310, + 176, + 363, + 232 + ], + "spans": [ + { + "bbox": [ + 310, + 176, + 363, + 232 + ], + "type": "image", + "image_path": "9ff077163feb3e8b536152bf928464003c5f4d0b8f6bfaa6c50f62203c8997b2.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 236, + 354, + 245 + ], + "lines": [ + { + "bbox": [ + 318, + 236, + 354, + 245 + ], + "spans": [ + { + "bbox": [ + 318, + 236, + 354, + 245 + ], + "type": "text", + "content": "(c) HOUSE" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 366, + 175, + 425, + 234 + ], + "blocks": [ + { + "bbox": [ + 311, + 165, + 423, + 175 + ], + "lines": [ + { + "bbox": [ + 311, + 165, + 423, + 175 + ], + "spans": [ + { + "bbox": [ + 311, + 165, + 423, + 175 + ], + "type": "text", + "content": "(a) 1D Heat Simulation via Eq. (4)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 366, + 175, + 425, + 234 + ], + "lines": [ + { + "bbox": [ + 366, + 175, + 425, + 234 + ], + "spans": [ + { + "bbox": [ + 366, + 175, + 425, + 234 + ], + "type": "image", + "image_path": "5c75ab31f618615213776769f9e9fd3672bdc3296f16f4db456dc7f8c55753bb.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 379, + 236, + 412, + 245 + ], + "lines": [ + { + "bbox": [ + 379, + 236, + 412, + 245 + ], + "spans": [ + { + "bbox": [ + 379, + 236, + 412, + 245 + ], + "type": "text", + "content": "(d) " + }, + { + "bbox": [ + 379, + 236, + 412, + 245 + ], + "type": "inline_equation", + "content": "\\lambda = 1" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 250, + 545, + 338 + ], + "lines": [ + { + "bbox": [ + 305, + 250, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 305, + 250, + 545, + 338 + ], + "type": "text", + "content": "Figure 4. We show an illustration of the relation between screened Poisson equation and a distance field. The top shows a 1D example, where on the left we show a solution to Eq. (4) with different absorption " + }, + { + "bbox": [ + 305, + 250, + 545, + 338 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 305, + 250, + 545, + 338 + ], + "type": "text", + "content": ", and on the right we show the reconstructed distance " + }, + { + "bbox": [ + 305, + 250, + 545, + 338 + ], + "type": "inline_equation", + "content": "-\\frac{1}{\\lambda}\\ln(h)" + }, + { + "bbox": [ + 305, + 250, + 545, + 338 + ], + "type": "text", + "content": " (Eq. (5)). Boundary loss for real boundary points acts as an isothermal heat source, while heat loss diffuses the heat. When " + }, + { + "bbox": [ + 305, + 250, + 545, + 338 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 305, + 250, + 545, + 338 + ], + "type": "text", + "content": " is increased, the heat decays faster, and the error of the reconstructed function approaches 0. The bottom shows a 2D example." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 426, + 175, + 485, + 234 + ], + "blocks": [ + { + "bbox": [ + 440, + 165, + 531, + 175 + ], + "lines": [ + { + "bbox": [ + 440, + 165, + 531, + 175 + ], + "spans": [ + { + "bbox": [ + 440, + 165, + 531, + 175 + ], + "type": "text", + "content": "(b) Distance Reconstruction" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 426, + 175, + 485, + 234 + ], + "lines": [ + { + "bbox": [ + 426, + 175, + 485, + 234 + ], + "spans": [ + { + "bbox": [ + 426, + 175, + 485, + 234 + ], + "type": "image", + "image_path": "474199d8dbc3f6e92fa3f367cdeb6155f8e54a98a2fd0adfd608ac741a0a8232.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 438, + 236, + 471, + 245 + ], + "lines": [ + { + "bbox": [ + 438, + 236, + 471, + 245 + ], + "spans": [ + { + "bbox": [ + 438, + 236, + 471, + 245 + ], + "type": "text", + "content": "(e) " + }, + { + "bbox": [ + 438, + 236, + 471, + 245 + ], + "type": "inline_equation", + "content": "\\lambda = 5" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 486, + 175, + 543, + 233 + ], + "blocks": [ + { + "bbox": [ + 486, + 175, + 543, + 233 + ], + "lines": [ + { + "bbox": [ + 486, + 175, + 543, + 233 + ], + "spans": [ + { + "bbox": [ + 486, + 175, + 543, + 233 + ], + "type": "image", + "image_path": "169c024f19af33225d597fd7b78d5b377a35da7298ba1ac6ee7880f21995eb89.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 496, + 236, + 531, + 245 + ], + "lines": [ + { + "bbox": [ + 496, + 236, + 531, + 245 + ], + "spans": [ + { + "bbox": [ + 496, + 236, + 531, + 245 + ], + "type": "text", + "content": "(f) " + }, + { + "bbox": [ + 496, + 236, + 531, + 245 + ], + "type": "inline_equation", + "content": "\\lambda = 50" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 346, + 545, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 346, + 545, + 406 + ], + "spans": [ + { + "bbox": [ + 305, + 346, + 545, + 406 + ], + "type": "text", + "content": "We will be using a key relation between the screen Poisson equation and the distance. We first define a screened Poisson equation over a heat field " + }, + { + "bbox": [ + 305, + 346, + 545, + 406 + ], + "type": "inline_equation", + "content": "h(\\pmb{x}): \\mathbb{R}^d \\to \\mathbb{R}" + }, + { + "bbox": [ + 305, + 346, + 545, + 406 + ], + "type": "text", + "content": ", with an absorption coefficient " + }, + { + "bbox": [ + 305, + 346, + 545, + 406 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 305, + 346, + 545, + 406 + ], + "type": "text", + "content": ", and set a Dirichlet boundary condition to be 1 on our target surface " + }, + { + "bbox": [ + 305, + 346, + 545, + 406 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 305, + 346, + 545, + 406 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 335, + 411, + 545, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 411, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 335, + 411, + 545, + 449 + ], + "type": "interline_equation", + "content": "\\left\\{\\begin{array}{l}\\nabla^ {2} h (\\boldsymbol {x}) - \\lambda^ {2} h (\\boldsymbol {x}) = 0 \\quad \\forall \\boldsymbol {x} \\in \\mathbb {R} ^ {d} \\backslash \\Gamma\\\\\\lim _ {| | \\boldsymbol {x} | | \\rightarrow \\infty} h (\\boldsymbol {x}) = 0, \\quad h (\\boldsymbol {x}) = 1 \\quad \\forall \\boldsymbol {x} \\in \\Gamma .\\end{array}\\right. \\tag {4}", + "image_path": "d878b9e34a5914261541f5246367b6e52bf382da787f5d2bbda8f381283f1b98.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 306, + 453, + 544, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 453, + 544, + 478 + ], + "spans": [ + { + "bbox": [ + 306, + 453, + 544, + 478 + ], + "type": "text", + "content": "If a heat field " + }, + { + "bbox": [ + 306, + 453, + 544, + 478 + ], + "type": "inline_equation", + "content": "h_\\lambda" + }, + { + "bbox": [ + 306, + 453, + 544, + 478 + ], + "type": "text", + "content": " is a solution to the partial differential equation above, then the following equation holds [6]:" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 363, + 483, + 545, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 483, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 363, + 483, + 545, + 506 + ], + "type": "interline_equation", + "content": "\\lim _ {\\lambda \\rightarrow \\infty} \\frac {1}{\\lambda} \\ln \\left(h _ {\\lambda} (\\boldsymbol {x})\\right) = - d _ {\\Gamma} (\\boldsymbol {x}). \\tag {5}", + "image_path": "345ec812d89279158806260578d00ff5a4289354ae9d9c6a860563549e77a885.jpg" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": "Fig. 4 shows the intuition. The screened Poisson equation simulates heat transfer until an equilibrium is reached, under absorption controlled by the " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " coefficient. As we increase the absorption, the reconstructed distance " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\frac{1}{\\lambda}\\ln (h_{\\lambda})" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " would approach the real distance. In practice, we could only obtain a discrete point set " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\Gamma_0" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": ". Provided that the neighbor points are close enough, if there are some boundaries connecting them, the boundary condition of Eq. (4) still holds. This can be interpreted as a superset " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " replacing " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\Gamma_0" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": ", effectively transforming the boundary from a discrete point cloud into a continuous surface, which is required and precisely the goal of the signed distance function reconstruction. Obtaining a distance function to the point cloud rather than the interpolated surface is not our objective. The reconstructed distance can still faithfully represent the distance to " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " because the approximation in Eq. (5) still holds. Later, we will prove that the measure of " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " is bounded by our loss." + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1278" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 104, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 104, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 104, + 83 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 289, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 289, + 224 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 289, + 224 + ], + "type": "text", + "content": "In this section, we introduce our method, HOTSPOT, based on the screened Poisson equation (Eq. (4)) and derive a theoretical analysis. We use the screened Poisson equation to derive a loss we can directly apply to an implicit function " + }, + { + "bbox": [ + 46, + 91, + 289, + 224 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 46, + 91, + 289, + 224 + ], + "type": "text", + "content": ", and prove that there is a bound to the solution of the screened Poisson equation to the true distance (Section 4.1). We further derive a stability analysis of our loss on both the spatial stability over small perturbations and the temporal convergence (Section 4.2). Next, we show that our loss can penalize surfaces with large area (Section 4.3). Finally, we discuss the relation to prior work, PHASE (Section 4.4)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 229, + 288, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 229, + 288, + 255 + ], + "spans": [ + { + "bbox": [ + 47, + 229, + 288, + 255 + ], + "type": "text", + "content": "4.1. Screened Poisson Equation Informed Signed Distance Functions Optimization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 259, + 287, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 259, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 46, + 259, + 287, + 308 + ], + "type": "text", + "content": "Given an implicit function " + }, + { + "bbox": [ + 46, + 259, + 287, + 308 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 46, + 259, + 287, + 308 + ], + "type": "text", + "content": ", we want to derive a loss based on the screened Poisson equation (Eq. (4)) and its relation to the distance (Eq. (5)). We achieve this by applying the following substitution based on Eq. (5):" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 128, + 314, + 287, + 330 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 314, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 128, + 314, + 287, + 330 + ], + "type": "interline_equation", + "content": "h (\\boldsymbol {x}) = e ^ {- \\lambda | u (\\boldsymbol {x}) |}. \\tag {6}", + "image_path": "832f97f5b076a48a1bbd776328870ada619155edcd06de15ea2cda88ca65376c.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 336, + 287, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 336, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 47, + 336, + 287, + 396 + ], + "type": "text", + "content": "We want to design a loss such that its minimizer satisfies the screened Poisson equation. This can be achieved by minimizing the energy functional " + }, + { + "bbox": [ + 47, + 336, + 287, + 396 + ], + "type": "inline_equation", + "content": "\\frac{1}{2}\\int_{\\Omega} \\| \\nabla h(\\pmb{x}) \\|^2 + \\lambda^2 h(\\pmb{x})^2 \\, \\mathrm{d}\\pmb{x}" + }, + { + "bbox": [ + 47, + 336, + 287, + 396 + ], + "type": "text", + "content": ". After substitution and removing of the constant factor " + }, + { + "bbox": [ + 47, + 336, + 287, + 396 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 47, + 336, + 287, + 396 + ], + "type": "text", + "content": ", we obtain our heat loss:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 73, + 402, + 287, + 428 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 402, + 287, + 428 + ], + "spans": [ + { + "bbox": [ + 73, + 402, + 287, + 428 + ], + "type": "interline_equation", + "content": "L _ {\\text {h e a t}} = \\frac {1}{2} \\int_ {\\Omega} e ^ {- 2 \\lambda | u (\\boldsymbol {x}) |} \\left(\\| \\nabla u (\\boldsymbol {x}) \\| ^ {2} + 1\\right) d \\boldsymbol {x}. \\tag {7}", + "image_path": "cc3cc256cd7ff805ca066290fbe05c46291d7abe0fde4c84d515f66b04b9a856.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 434, + 287, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 434, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 47, + 434, + 287, + 495 + ], + "type": "text", + "content": "If we take the derivative of " + }, + { + "bbox": [ + 47, + 434, + 287, + 495 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{heat}}" + }, + { + "bbox": [ + 47, + 434, + 287, + 495 + ], + "type": "text", + "content": " with respect to the heat field " + }, + { + "bbox": [ + 47, + 434, + 287, + 495 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 47, + 434, + 287, + 495 + ], + "type": "text", + "content": ", we recover the screened Poisson equation " + }, + { + "bbox": [ + 47, + 434, + 287, + 495 + ], + "type": "inline_equation", + "content": "\\nabla^2 h - \\lambda^2 h = 0" + }, + { + "bbox": [ + 47, + 434, + 287, + 495 + ], + "type": "text", + "content": ". The boundary condition can be enforced by the standard boundary loss (Eq. (2)), by observing that " + }, + { + "bbox": [ + 47, + 434, + 287, + 495 + ], + "type": "inline_equation", + "content": "u(\\pmb{x}) = 0" + }, + { + "bbox": [ + 47, + 434, + 287, + 495 + ], + "type": "text", + "content": " implies " + }, + { + "bbox": [ + 47, + 434, + 287, + 495 + ], + "type": "inline_equation", + "content": "h(\\pmb{x}) = 1" + }, + { + "bbox": [ + 47, + 434, + 287, + 495 + ], + "type": "text", + "content": ". Together with the eikonal loss, we optimize:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 84, + 502, + 287, + 515 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 502, + 287, + 515 + ], + "spans": [ + { + "bbox": [ + 84, + 502, + 287, + 515 + ], + "type": "interline_equation", + "content": "L = w _ {b} L _ {\\text {b o u n d a r y}} + w _ {e} L _ {\\text {e i k o n a l}} + w _ {h} L _ {\\text {h e a t}}. \\tag {8}", + "image_path": "d93c539f16afa71fec0bfd19833d1669862c5db41cb9f10fa94d88586993016c.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 521, + 287, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 618 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 618 + ], + "type": "text", + "content": "While prior work [6] has studied the limiting behavior of the absorption coefficient " + }, + { + "bbox": [ + 46, + 521, + 287, + 618 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 521, + 287, + 618 + ], + "type": "text", + "content": ", to our knowledge, the bounds of the approximation and the convergence speed remain unknown. In the supplementary material A.3, we prove that, for a boundary " + }, + { + "bbox": [ + 46, + 521, + 287, + 618 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 46, + 521, + 287, + 618 + ], + "type": "text", + "content": " formed by a discrete set of points and a small volume around them, the distance approximation obtained from the solution of the screened Poisson equation converges linearly with respect to " + }, + { + "bbox": [ + 46, + 521, + 287, + 618 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 521, + 287, + 618 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 624, + 287, + 648 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 287, + 648 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 287, + 648 + ], + "type": "interline_equation", + "content": "\\frac {1}{\\lambda} C _ {\\text {l o w e r}} (\\boldsymbol {x}) \\leqslant \\frac {1}{\\lambda} \\ln \\left(h _ {\\lambda} (\\boldsymbol {x})\\right) + d _ {\\Gamma} (\\boldsymbol {x}) \\leqslant \\frac {1}{\\lambda} C _ {\\text {u p p e r}} (\\boldsymbol {x}). \\tag {9}", + "image_path": "f98f04fc813b88139152fa4c25872706544c239d211efa49f198a87ffbcc75a2.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 653, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 689 + ], + "type": "text", + "content": "Here, both " + }, + { + "bbox": [ + 46, + 653, + 287, + 689 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{upper}}(\\pmb{x})" + }, + { + "bbox": [ + 46, + 653, + 287, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 653, + 287, + 689 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{lower}}(\\pmb{x})" + }, + { + "bbox": [ + 46, + 653, + 287, + 689 + ], + "type": "text", + "content": " are scalar fields independent of " + }, + { + "bbox": [ + 46, + 653, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 653, + 287, + 689 + ], + "type": "text", + "content": ". The inequalities are tight, which means that one can find a case where the equality holds." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "Compared to the eikonal loss (Eq. (3)), where the distance error can be unbounded even when the equation holds" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 309, + 70, + 543, + 162 + ], + "blocks": [ + { + "bbox": [ + 309, + 70, + 543, + 162 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 543, + 162 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 543, + 162 + ], + "type": "image", + "image_path": "10116da4676066829bfb6317112af6411f3b4a5a054c8bd4e130bdcdbb90e036.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 167, + 547, + 234 + ], + "lines": [ + { + "bbox": [ + 304, + 167, + 547, + 234 + ], + "spans": [ + { + "bbox": [ + 304, + 167, + 547, + 234 + ], + "type": "text", + "content": "Figure 5. We visualize the direction of the changes of the implicit function during one iteration of optimization. Left shows the result with only the eikonal loss and right shows the result after adding our loss. Our loss helps to escape the local minimum in Fig. 3. The heat diffusion encourages the reconstructed absolute distance to increase monotonically away from the boundary." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 245, + 547, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 245, + 547, + 376 + ], + "spans": [ + { + "bbox": [ + 304, + 245, + 547, + 376 + ], + "type": "text", + "content": "almost everywhere, Eq. (9) shows that even with a finite absorption " + }, + { + "bbox": [ + 304, + 245, + 547, + 376 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 245, + 547, + 376 + ], + "type": "text", + "content": ", our loss can lead to much more accurate distance. In fact, any loss that only depends on the norm of the gradients " + }, + { + "bbox": [ + 304, + 245, + 547, + 376 + ], + "type": "inline_equation", + "content": "\\|\\nabla u\\|" + }, + { + "bbox": [ + 304, + 245, + 547, + 376 + ], + "type": "text", + "content": " is not a sufficient condition of the distance function (Fig. 2), due to the discontinuous jumps of the gradient. Adding our loss (Eq. (7)) ensures that the absolute value of the implicit function converges to the actual distance. Fig. 5 visualizes the direction of how an implicit function would evolve when using only the eikonal loss and after adding our loss. Our loss ensures that the implicit function evolves towards the actual distance." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 386, + 411, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 386, + 411, + 399 + ], + "spans": [ + { + "bbox": [ + 305, + 386, + 411, + 399 + ], + "type": "text", + "content": "4.2. Stability Analysis" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 405, + 547, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 405, + 547, + 489 + ], + "spans": [ + { + "bbox": [ + 304, + 405, + 547, + 489 + ], + "type": "text", + "content": "Our heat loss enjoys other theoretical benefits apart from the convergence to an actual distance. Here, we show that our loss leads to easier optimization. We separately analyze spatial stability and temporal stability, where the spatial stability studies the effect of small perturbation error added to the corresponding partial differential equation, and the temporal stability studies the convergence of the gradient flow." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 505, + 408, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 505, + 408, + 517 + ], + "spans": [ + { + "bbox": [ + 305, + 505, + 408, + 517 + ], + "type": "text", + "content": "4.2.1 Spatial Stability" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 525, + 545, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 525, + 545, + 619 + ], + "spans": [ + { + "bbox": [ + 304, + 525, + 545, + 619 + ], + "type": "text", + "content": "We show that for the screened Poisson equation, an error introduced to a single point in the solution decays exponentially and does not propagate to infinitely distant positions. On the contrary, for the eikonal equation, the error can propagate to infinitely far. The property of the screened Poisson equation ensures that the perturbation introduced during optimization does not lead to a drastically different target and stabilizes the optimization." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 620, + 545, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 620, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 304, + 620, + 545, + 657 + ], + "type": "text", + "content": "First, we show that for the eikonal equation, an additive error introduced to a point can presist along a line segment or a ray." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "Proposition 1. Given a field " + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "inline_equation", + "content": "u(\\pmb{x})" + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": " that satisfies the eikonal equation " + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\| \\nabla u \\| = 1" + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": ", we introduce an additive error " + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "inline_equation", + "content": "u_{0e}" + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": " at the point " + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0" + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "inline_equation", + "content": "u(\\pmb{x}_0) = u_0" + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": " to obtain a perturbed field " + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "inline_equation", + "content": "u'" + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "inline_equation", + "content": "u'(\\pmb{x}_0) = u_0 + u_{0e}" + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\| \\nabla u' \\| = 1" + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": ". In the" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1279" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 120 + ], + "type": "text", + "content": "perturbed field " + }, + { + "bbox": [ + 46, + 72, + 289, + 120 + ], + "type": "inline_equation", + "content": "u'" + }, + { + "bbox": [ + 46, + 72, + 289, + 120 + ], + "type": "text", + "content": ", there exists a line segment parameterized by " + }, + { + "bbox": [ + 46, + 72, + 289, + 120 + ], + "type": "inline_equation", + "content": "\\pmb{x}(s)" + }, + { + "bbox": [ + 46, + 72, + 289, + 120 + ], + "type": "text", + "content": " starting from " + }, + { + "bbox": [ + 46, + 72, + 289, + 120 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0" + }, + { + "bbox": [ + 46, + 72, + 289, + 120 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 289, + 120 + ], + "type": "inline_equation", + "content": "s \\in [0, M]" + }, + { + "bbox": [ + 46, + 72, + 289, + 120 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 72, + 289, + 120 + ], + "type": "inline_equation", + "content": "M > 0" + }, + { + "bbox": [ + 46, + 72, + 289, + 120 + ], + "type": "text", + "content": " and can be infinite, such that every point along the line has the same error:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 111, + 127, + 287, + 141 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 127, + 287, + 141 + ], + "spans": [ + { + "bbox": [ + 111, + 127, + 287, + 141 + ], + "type": "interline_equation", + "content": "u ^ {\\prime} (\\boldsymbol {x} (s)) - u (\\boldsymbol {x} (s)) = u _ {0 e} \\tag {10}", + "image_path": "7ce56ffb5d2b0f895f121dec30ef0bb9e3dcd3bf450e7d25fe5a3fe09d98d9b3.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 148, + 287, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 148, + 287, + 195 + ], + "spans": [ + { + "bbox": [ + 47, + 148, + 287, + 195 + ], + "type": "text", + "content": "The boundary points can contain errors when approximating the true boundaries, and these errors remain constant along the line segment. As a result, the propagated error can become significant and potentially detrimental." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 195, + 287, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 195, + 287, + 220 + ], + "spans": [ + { + "bbox": [ + 47, + 195, + 287, + 220 + ], + "type": "text", + "content": "Next, we demonstrate that the error field " + }, + { + "bbox": [ + 47, + 195, + 287, + 220 + ], + "type": "inline_equation", + "content": "h_e" + }, + { + "bbox": [ + 47, + 195, + 287, + 220 + ], + "type": "text", + "content": " in Eq. (4) decays exponentially from the perturbed point " + }, + { + "bbox": [ + 47, + 195, + 287, + 220 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 47, + 195, + 287, + 220 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 225, + 287, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 225, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 47, + 225, + 287, + 308 + ], + "type": "text", + "content": "Proposition 2. In 3D space, consider an additive error " + }, + { + "bbox": [ + 47, + 225, + 287, + 308 + ], + "type": "inline_equation", + "content": "h_{0e}" + }, + { + "bbox": [ + 47, + 225, + 287, + 308 + ], + "type": "text", + "content": " within a small ball " + }, + { + "bbox": [ + 47, + 225, + 287, + 308 + ], + "type": "inline_equation", + "content": "B(\\pmb{x}_0,\\epsilon)" + }, + { + "bbox": [ + 47, + 225, + 287, + 308 + ], + "type": "text", + "content": " around the point " + }, + { + "bbox": [ + 47, + 225, + 287, + 308 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0" + }, + { + "bbox": [ + 47, + 225, + 287, + 308 + ], + "type": "text", + "content": ", where the original field " + }, + { + "bbox": [ + 47, + 225, + 287, + 308 + ], + "type": "inline_equation", + "content": "h(\\pmb{x})" + }, + { + "bbox": [ + 47, + 225, + 287, + 308 + ], + "type": "text", + "content": " and the perturbed field " + }, + { + "bbox": [ + 47, + 225, + 287, + 308 + ], + "type": "inline_equation", + "content": "h^{\\prime}(\\pmb{x})" + }, + { + "bbox": [ + 47, + 225, + 287, + 308 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 225, + 287, + 308 + ], + "type": "inline_equation", + "content": "\\pmb{x} \\in \\mathbb{R}^n\\backslash B" + }, + { + "bbox": [ + 47, + 225, + 287, + 308 + ], + "type": "text", + "content": " both satisfy the screened Poisson equation in Eq. (4). The resulting error field " + }, + { + "bbox": [ + 47, + 225, + 287, + 308 + ], + "type": "inline_equation", + "content": "h_e = h' - h" + }, + { + "bbox": [ + 47, + 225, + 287, + 308 + ], + "type": "text", + "content": " is radially symmetric with respect to " + }, + { + "bbox": [ + 47, + 225, + 287, + 308 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 47, + 225, + 287, + 308 + ], + "type": "text", + "content": " within some maximum radius " + }, + { + "bbox": [ + 47, + 225, + 287, + 308 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 47, + 225, + 287, + 308 + ], + "type": "text", + "content": ", and is given by" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 308, + 287, + 328 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 308, + 287, + 328 + ], + "spans": [ + { + "bbox": [ + 104, + 308, + 287, + 328 + ], + "type": "interline_equation", + "content": "h _ {e} (r) = \\frac {\\epsilon}{r} h _ {0 e} e ^ {\\lambda (\\epsilon - r)}, \\forall r > \\epsilon , \\tag {11}", + "image_path": "ec8f09621f1153653514cc7a9139b76be2ef4103b080bcb3dd607e19541b5816.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 331, + 258, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 331, + 258, + 344 + ], + "spans": [ + { + "bbox": [ + 47, + 331, + 258, + 344 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 331, + 258, + 344 + ], + "type": "inline_equation", + "content": "r = \\left\\| \\boldsymbol{x} - \\boldsymbol{x}_0 \\right\\|" + }, + { + "bbox": [ + 47, + 331, + 258, + 344 + ], + "type": "text", + "content": " is the radial distance from " + }, + { + "bbox": [ + 47, + 331, + 258, + 344 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_0" + }, + { + "bbox": [ + 47, + 331, + 258, + 344 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 350, + 287, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 350, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 47, + 350, + 287, + 409 + ], + "type": "text", + "content": "The proofs of these two propositions are in the supplementary material A.1 and A.2. By comparing the error fields, we observe that the partial differential equation governed by the screened Poisson equation exhibits greater spatial robustness than the one based on the eikonal equation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 423, + 160, + 435 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 423, + 160, + 435 + ], + "spans": [ + { + "bbox": [ + 47, + 423, + 160, + 435 + ], + "type": "text", + "content": "4.2.2 Temporal Stability" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 441, + 287, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 441, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 47, + 441, + 287, + 502 + ], + "type": "text", + "content": "We next demonstrate the temporal stability of the heat loss and show that the dynamical system of the gradient update is stable, that is, the solution of the system converges as the time " + }, + { + "bbox": [ + 47, + 441, + 287, + 502 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 441, + 287, + 502 + ], + "type": "text", + "content": " approaches infinity. This is supported by von Neumann stability analysis, similar to Yang et al.'s work [4]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 502, + 287, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 502, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 47, + 502, + 287, + 525 + ], + "type": "text", + "content": "Yang et al. showed that for the eikonal loss (Eq. (3)), the temporal gradient flow of the optimization is:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 60, + 532, + 287, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 532, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 60, + 532, + 287, + 594 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\frac {\\partial u}{\\partial t} = - \\frac {\\delta L _ {\\mathrm {e i k o n a l}}}{\\delta u} = \\nabla \\cdot (\\kappa_ {e} \\nabla u) \\approx \\kappa_ {e} \\nabla^ {2} u, \\text {w h e r e} \\\\ \\kappa_ {e} (\\boldsymbol {x}) = \\left\\{ \\begin{array}{l l} \\frac {1}{| \\nabla u (\\boldsymbol {x}) |} \\operatorname {s g n} (| \\nabla u (\\boldsymbol {x}) | - 1) & p = 1 \\\\ 1 - | \\nabla u (\\boldsymbol {x}) | ^ {- 1} & p = 2 \\end{array} \\right. \\tag {12} \\\\ \\end{array}", + "image_path": "f388b4c0a7384585500dd270aa8c64e634be1195066b3e714beecd068575a9b1.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 59, + 594, + 286, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 594, + 286, + 606 + ], + "spans": [ + { + "bbox": [ + 59, + 594, + 286, + 606 + ], + "type": "text", + "content": "They analyze the stability by applying Fourier transform:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 612, + 287, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 612, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 53, + 612, + 287, + 635 + ], + "type": "interline_equation", + "content": "\\frac {\\partial \\hat {u}}{\\partial t} (t, \\omega) = - \\kappa_ {e} | \\omega | ^ {2} \\hat {u} (t, \\omega) \\Rightarrow \\hat {u} (t, \\omega) \\propto e ^ {- \\kappa_ {e} | \\omega | ^ {2} t}, \\tag {13}", + "image_path": "08565b90afe229972e9213e9ec7160e997dee0357735ffb82828cf5e8e1a85ca.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 641, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 641, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 641, + 287, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 641, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\omega = (\\omega_{1},\\dots ,\\omega_{n})" + }, + { + "bbox": [ + 47, + 641, + 287, + 713 + ], + "type": "text", + "content": " denotes the frequency variable and " + }, + { + "bbox": [ + 47, + 641, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\hat{u}" + }, + { + "bbox": [ + 47, + 641, + 287, + 713 + ], + "type": "text", + "content": " represents the Fourier transform of " + }, + { + "bbox": [ + 47, + 641, + 287, + 713 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 47, + 641, + 287, + 713 + ], + "type": "text", + "content": ". In the frequency domain, if the original function converges, the term " + }, + { + "bbox": [ + 47, + 641, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\frac{\\partial\\hat{u}}{\\partial t}" + }, + { + "bbox": [ + 47, + 641, + 287, + 713 + ], + "type": "text", + "content": " should also converge. However, when " + }, + { + "bbox": [ + 47, + 641, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\kappa_{e} < 0" + }, + { + "bbox": [ + 47, + 641, + 287, + 713 + ], + "type": "text", + "content": ", this process corresponds to backward diffusion, which diverges and is inherently unstable regardless of the numerical implementation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 96 + ], + "type": "text", + "content": "We extend the analysis to our heat loss (Eq. (7)). The gradient flow is" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 364, + 102, + 545, + 126 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 102, + 545, + 126 + ], + "spans": [ + { + "bbox": [ + 364, + 102, + 545, + 126 + ], + "type": "interline_equation", + "content": "\\frac {\\partial h}{\\partial t} = - \\frac {\\delta L _ {\\text {h e a t}}}{\\delta h} = \\nabla^ {2} h - \\lambda^ {2} h \\tag {14}", + "image_path": "0f3b4ba0dbe33b29ca6fccdb47cd7885ff3ddb9c5b3477579d26075ea99e9a17.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 131, + 545, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 131, + 545, + 166 + ], + "spans": [ + { + "bbox": [ + 305, + 131, + 545, + 166 + ], + "type": "text", + "content": "This is exactly the heat equation, and is known to be stable. We can further verify the stability by applying Fourier transform to both sides:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 172, + 545, + 206 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 172, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 306, + 172, + 545, + 206 + ], + "type": "interline_equation", + "content": "\\frac {\\partial \\hat {h}}{\\partial t} (t, \\omega) = - \\left(| \\omega | ^ {2} + \\lambda^ {2}\\right) \\hat {h} (t, \\omega) \\Rightarrow \\hat {h} (t, \\omega) \\propto e ^ {- \\left(| \\omega | ^ {2} + \\lambda^ {2}\\right) t} \\tag {15}", + "image_path": "23e5847b7cf53e7a66bf66064874419625f396a21f4c06353f12670411c703ca.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 207, + 545, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 207, + 545, + 278 + ], + "spans": [ + { + "bbox": [ + 305, + 207, + 545, + 278 + ], + "type": "text", + "content": "This analysis reveals that the original process " + }, + { + "bbox": [ + 305, + 207, + 545, + 278 + ], + "type": "inline_equation", + "content": "h(\\pmb{x}, t)" + }, + { + "bbox": [ + 305, + 207, + 545, + 278 + ], + "type": "text", + "content": " converges over time and remains stable. Due to the constraint " + }, + { + "bbox": [ + 305, + 207, + 545, + 278 + ], + "type": "inline_equation", + "content": "|u| = -\\ln(h) / \\lambda" + }, + { + "bbox": [ + 305, + 207, + 545, + 278 + ], + "type": "text", + "content": ", the term " + }, + { + "bbox": [ + 305, + 207, + 545, + 278 + ], + "type": "inline_equation", + "content": "|u(\\pmb{x}, t)|" + }, + { + "bbox": [ + 305, + 207, + 545, + 278 + ], + "type": "text", + "content": " also converges during the optimization process. Since " + }, + { + "bbox": [ + 305, + 207, + 545, + 278 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 305, + 207, + 545, + 278 + ], + "type": "text", + "content": " is continuous, it converges as well. Thus, our method exhibits greater spatial and temporal stability compared to the eikonal equation." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 285, + 462, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 285, + 462, + 298 + ], + "spans": [ + { + "bbox": [ + 306, + 285, + 462, + 298 + ], + "type": "text", + "content": "4.3. Surface Area Regularization" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 304, + 545, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 304, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 305, + 304, + 545, + 375 + ], + "type": "text", + "content": "We next show that our heat loss (Eq. (7)) penalizes large area of the reconstructed surfaces, by showing its relation to the coarea loss used by Pumarola et al. [30]. Pumarola et al. showed that the loss " + }, + { + "bbox": [ + 305, + 304, + 545, + 375 + ], + "type": "inline_equation", + "content": "\\int_{\\Omega} e^{-\\lambda |u|} \\| \\nabla u \\| \\, \\mathrm{d}\\pmb{x}" + }, + { + "bbox": [ + 305, + 304, + 545, + 375 + ], + "type": "text", + "content": " approximates the surface area of the zero level set of the implicit function " + }, + { + "bbox": [ + 305, + 304, + 545, + 375 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 305, + 304, + 545, + 375 + ], + "type": "text", + "content": ". We then form the following inequality:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 310, + 381, + 545, + 441 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 381, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 310, + 381, + 545, + 441 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\int_ {\\Omega} e ^ {- \\lambda | u |} \\| \\nabla u \\| d \\boldsymbol {x} < \\int_ {\\Omega} \\sqrt {e ^ {- 2 \\lambda | u |} \\left(\\| \\nabla u \\| ^ {2} + 1\\right)} d \\boldsymbol {x} (16) \\\\ \\leqslant \\sqrt {| \\Omega |} \\cdot \\sqrt {\\int_ {\\Omega} e ^ {- 2 \\lambda | u |} \\left(\\| \\nabla u \\| ^ {2} + 1\\right) d x} (16) \\\\ \\end{array}", + "image_path": "cae8684cc77f28df1f23e299c0b80734e8073d48ab145e29a9caa12db3e352ec.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 447, + 545, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 447, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 305, + 447, + 545, + 470 + ], + "type": "text", + "content": "Therefore, minimizing the right hand side (our heat loss) leads to a reduction in the upper bound of the left hand side." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 305, + 471, + 545, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 471, + 545, + 530 + ], + "spans": [ + { + "bbox": [ + 305, + 471, + 545, + 530 + ], + "type": "text", + "content": "Furthermore, we can show that the coarea loss is too aggressive for our purpose of signed distance field reconstruction. If we directly minimize the coarea loss (the left hand side in Eq. (16)), according to Euler-Lagrange equation, the minimizer would satisfy:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 396, + 537, + 545, + 550 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 396, + 537, + 545, + 550 + ], + "spans": [ + { + "bbox": [ + 396, + 537, + 545, + 550 + ], + "type": "interline_equation", + "content": "1 - \\nabla^ {2} u = 0, \\tag {17}", + "image_path": "9ccd7b2cce35712b3b5a8012b5e965c92b4234c3876ca012088e48411663613a.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 305, + 558, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 558, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 305, + 558, + 545, + 689 + ], + "type": "text", + "content": "which does not relate to the regularization parameter " + }, + { + "bbox": [ + 305, + 558, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 305, + 558, + 545, + 689 + ], + "type": "text", + "content": ". Unfortunately, this partial differential equation system formed by the coarea loss above has a unique solution under Dirichlet boundary condition, and the gradient of its solution does not have a unit norm. This makes the coarea loss unable to accurately shape a signed distance function. By contrast, our heat loss has bounds for its distance approximation and converges to the true distance as we increase the parameter " + }, + { + "bbox": [ + 305, + 558, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 305, + 558, + 545, + 689 + ], + "type": "text", + "content": " (Eq. (9)). Another simple way to verify our convergence and the compatibility with the eikonal equation is to rewrite Eq. (4) in terms of " + }, + { + "bbox": [ + 305, + 558, + 545, + 689 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 305, + 558, + 545, + 689 + ], + "type": "text", + "content": " following Eq. (6):" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 361, + 692, + 545, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 692, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 361, + 692, + 545, + 714 + ], + "type": "interline_equation", + "content": "\\left| \\left| \\nabla u \\right| \\right| ^ {2} - 1 - \\frac {\\operatorname {s g n} (\\mathrm {u})}{\\lambda} \\nabla^ {2} u = 0. \\tag {18}", + "image_path": "ba2fed892b3b7643473cf78b0cf9dcc65806b1f02725ffab47c139dd220b86dc.jpg" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1280" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 94, + 116 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 94, + 116 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 94, + 116 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 94, + 116 + ], + "type": "image", + "image_path": "e4e678153ee695bcdeac6ecac31d2699d70d595a44f3ffd69c80a350da2b3fd2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 63, + 118, + 80, + 123 + ], + "lines": [ + { + "bbox": [ + 63, + 118, + 80, + 123 + ], + "spans": [ + { + "bbox": [ + 63, + 118, + 80, + 123 + ], + "type": "text", + "content": "(a) GT" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 96, + 72, + 138, + 116 + ], + "blocks": [ + { + "bbox": [ + 96, + 72, + 138, + 116 + ], + "lines": [ + { + "bbox": [ + 96, + 72, + 138, + 116 + ], + "spans": [ + { + "bbox": [ + 96, + 72, + 138, + 116 + ], + "type": "image", + "image_path": "64868793a5d0bb8cc56a12997271fffcde9068ca761fb3e30a9b5d66e2d2596f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 100, + 118, + 133, + 124 + ], + "lines": [ + { + "bbox": [ + 100, + 118, + 133, + 124 + ], + "spans": [ + { + "bbox": [ + 100, + 118, + 133, + 124 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 100, + 118, + 133, + 124 + ], + "type": "inline_equation", + "content": "w_{h} = 0.1" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 139, + 72, + 183, + 116 + ], + "blocks": [ + { + "bbox": [ + 139, + 72, + 183, + 116 + ], + "lines": [ + { + "bbox": [ + 139, + 72, + 183, + 116 + ], + "spans": [ + { + "bbox": [ + 139, + 72, + 183, + 116 + ], + "type": "image", + "image_path": "7f79dd38777fbf7a13513caf181ee5d8a1150015a7603b3717186b1e5019e149.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 145, + 118, + 178, + 124 + ], + "lines": [ + { + "bbox": [ + 145, + 118, + 178, + 124 + ], + "spans": [ + { + "bbox": [ + 145, + 118, + 178, + 124 + ], + "type": "text", + "content": "(c) " + }, + { + "bbox": [ + 145, + 118, + 178, + 124 + ], + "type": "inline_equation", + "content": "w_{h} = 0.2" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 185, + 72, + 228, + 116 + ], + "blocks": [ + { + "bbox": [ + 185, + 72, + 228, + 116 + ], + "lines": [ + { + "bbox": [ + 185, + 72, + 228, + 116 + ], + "spans": [ + { + "bbox": [ + 185, + 72, + 228, + 116 + ], + "type": "image", + "image_path": "576c0944855c1f97579431ecff819857a01688dcaf87b0f5e001ca47f43c43e1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 190, + 118, + 223, + 124 + ], + "lines": [ + { + "bbox": [ + 190, + 118, + 223, + 124 + ], + "spans": [ + { + "bbox": [ + 190, + 118, + 223, + 124 + ], + "type": "text", + "content": "(d) " + }, + { + "bbox": [ + 190, + 118, + 223, + 124 + ], + "type": "inline_equation", + "content": "w_{h} = 0.3" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 229, + 72, + 277, + 116 + ], + "blocks": [ + { + "bbox": [ + 229, + 72, + 277, + 116 + ], + "lines": [ + { + "bbox": [ + 229, + 72, + 277, + 116 + ], + "spans": [ + { + "bbox": [ + 229, + 72, + 277, + 116 + ], + "type": "image", + "image_path": "5af4162badb2fe6729de3b2e334b6929406fde9c157c75ddc9721877c97ac73d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 235, + 118, + 267, + 124 + ], + "lines": [ + { + "bbox": [ + 235, + 118, + 267, + 124 + ], + "spans": [ + { + "bbox": [ + 235, + 118, + 267, + 124 + ], + "type": "text", + "content": "(e) " + }, + { + "bbox": [ + 235, + 118, + 267, + 124 + ], + "type": "inline_equation", + "content": "w_{b} = 0.5" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 50, + 128, + 94, + 173 + ], + "blocks": [ + { + "bbox": [ + 50, + 128, + 94, + 173 + ], + "lines": [ + { + "bbox": [ + 50, + 128, + 94, + 173 + ], + "spans": [ + { + "bbox": [ + 50, + 128, + 94, + 173 + ], + "type": "image", + "image_path": "a686cc57f472ad0cc6a198247c85c9291ccb7197f3626d0699de39ebe985c1ab.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 58, + 174, + 85, + 181 + ], + "lines": [ + { + "bbox": [ + 58, + 174, + 85, + 181 + ], + "spans": [ + { + "bbox": [ + 58, + 174, + 85, + 181 + ], + "type": "text", + "content": "(f) " + }, + { + "bbox": [ + 58, + 174, + 85, + 181 + ], + "type": "inline_equation", + "content": "w_{b} = 1" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 96, + 128, + 138, + 173 + ], + "blocks": [ + { + "bbox": [ + 96, + 128, + 138, + 173 + ], + "lines": [ + { + "bbox": [ + 96, + 128, + 138, + 173 + ], + "spans": [ + { + "bbox": [ + 96, + 128, + 138, + 173 + ], + "type": "image", + "image_path": "d1ac9ae1207d822f36749d7c7c7869cc233e0ddcf303aea168543f8e006515ee.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 102, + 174, + 131, + 181 + ], + "lines": [ + { + "bbox": [ + 102, + 174, + 131, + 181 + ], + "spans": [ + { + "bbox": [ + 102, + 174, + 131, + 181 + ], + "type": "text", + "content": "(g) " + }, + { + "bbox": [ + 102, + 174, + 131, + 181 + ], + "type": "inline_equation", + "content": "w_{b} = 2" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 139, + 128, + 183, + 173 + ], + "blocks": [ + { + "bbox": [ + 139, + 128, + 183, + 173 + ], + "lines": [ + { + "bbox": [ + 139, + 128, + 183, + 173 + ], + "spans": [ + { + "bbox": [ + 139, + 128, + 183, + 173 + ], + "type": "image", + "image_path": "5d43e0c22b7d8cb496a46aad7f495b65a4ee24f39712a55a32880777cccdb404.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 147, + 174, + 176, + 181 + ], + "lines": [ + { + "bbox": [ + 147, + 174, + 176, + 181 + ], + "spans": [ + { + "bbox": [ + 147, + 174, + 176, + 181 + ], + "type": "text", + "content": "(h) " + }, + { + "bbox": [ + 147, + 174, + 176, + 181 + ], + "type": "inline_equation", + "content": "w_{b} = 5" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 184, + 128, + 228, + 173 + ], + "blocks": [ + { + "bbox": [ + 184, + 128, + 228, + 173 + ], + "lines": [ + { + "bbox": [ + 184, + 128, + 228, + 173 + ], + "spans": [ + { + "bbox": [ + 184, + 128, + 228, + 173 + ], + "type": "image", + "image_path": "55af226a5059f90e1e84c1f23f25d4b17874d50ae39d371e172f0722f155573e.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 190, + 174, + 221, + 181 + ], + "lines": [ + { + "bbox": [ + 190, + 174, + 221, + 181 + ], + "spans": [ + { + "bbox": [ + 190, + 174, + 221, + 181 + ], + "type": "text", + "content": "(i) " + }, + { + "bbox": [ + 190, + 174, + 221, + 181 + ], + "type": "inline_equation", + "content": "w_{b} = 10" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 182, + 287, + 259 + ], + "lines": [ + { + "bbox": [ + 46, + 182, + 287, + 259 + ], + "spans": [ + { + "bbox": [ + 46, + 182, + 287, + 259 + ], + "type": "text", + "content": "Figure 6. We reproduce the PHASE model [49] and show the log-transformed occupancy field, i.e., the optimized signed distance function. PHASE recommends boundary weight range " + }, + { + "bbox": [ + 46, + 182, + 287, + 259 + ], + "type": "inline_equation", + "content": "w_{b} \\in (0.10, 0.32)" + }, + { + "bbox": [ + 46, + 182, + 287, + 259 + ], + "type": "text", + "content": " based on their " + }, + { + "bbox": [ + 46, + 182, + 287, + 259 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 46, + 182, + 287, + 259 + ], + "type": "text", + "content": "-convergence theory. But in practice, only a stronger boundary weight constrains the boundary to follow the input points. We found that this phenomenon often occurs and show more examples in the supplementary material." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 229, + 128, + 277, + 173 + ], + "blocks": [ + { + "bbox": [ + 229, + 128, + 277, + 173 + ], + "lines": [ + { + "bbox": [ + 229, + 128, + 277, + 173 + ], + "spans": [ + { + "bbox": [ + 229, + 128, + 277, + 173 + ], + "type": "image", + "image_path": "2bc27a3330646b9d3cb33371d3c5026e0dbfc1ab0254d79b1cfd420bebd3fea3.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 236, + 174, + 266, + 181 + ], + "lines": [ + { + "bbox": [ + 236, + 174, + 266, + 181 + ], + "spans": [ + { + "bbox": [ + 236, + 174, + 266, + 181 + ], + "type": "text", + "content": "(j) " + }, + { + "bbox": [ + 236, + 174, + 266, + 181 + ], + "type": "inline_equation", + "content": "w_{b} = 20" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 47, + 270, + 181, + 283 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 270, + 181, + 283 + ], + "spans": [ + { + "bbox": [ + 47, + 270, + 181, + 283 + ], + "type": "text", + "content": "4.4. Relation to PHASE [49]" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 46, + 288, + 287, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 288, + 287, + 347 + ], + "spans": [ + { + "bbox": [ + 46, + 288, + 287, + 347 + ], + "type": "text", + "content": "While derived from different mathematical principles and motivations, our method is closely related to the PHASE model proposed by Lipman. It is motivated by smoothly approximating an occupancy function " + }, + { + "bbox": [ + 46, + 288, + 287, + 347 + ], + "type": "inline_equation", + "content": "o(x)" + }, + { + "bbox": [ + 46, + 288, + 287, + 347 + ], + "type": "text", + "content": ", which is defined as 1 outside, -1 inside, and 0 on the surface." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 46, + 348, + 287, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 348, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 46, + 348, + 287, + 443 + ], + "type": "text", + "content": "In the supplement, we show that there is a close relation between our heat field " + }, + { + "bbox": [ + 46, + 348, + 287, + 443 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 46, + 348, + 287, + 443 + ], + "type": "text", + "content": " and the smoothed occupancy function " + }, + { + "bbox": [ + 46, + 348, + 287, + 443 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 46, + 348, + 287, + 443 + ], + "type": "text", + "content": ". We also show that our theory and derivation provide additional insights and significant differences in implementation and results. In particular, PHASE's theory introduces a condition stating that the boundary loss weight " + }, + { + "bbox": [ + 46, + 348, + 287, + 443 + ], + "type": "inline_equation", + "content": "w_{b}" + }, + { + "bbox": [ + 46, + 348, + 287, + 443 + ], + "type": "text", + "content": " should be small, whereas our theory emphasizes the need for a firm boundary condition (Eq. (4) and Fig. 6)." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 46, + 444, + 287, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 444, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 444, + 287, + 491 + ], + "type": "text", + "content": "We further found that optimizing occupancy field in the PHASE model, when used together with the eikonal loss, can lead to unbounded network weight derivatives, impeding the optimization. Our method avoids this issue." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 47, + 502, + 128, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 502, + 128, + 515 + ], + "spans": [ + { + "bbox": [ + 47, + 502, + 128, + 515 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 46, + 521, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 617 + ], + "type": "text", + "content": "We experiment on both 2D and 3D datasets. We construct a 2D dataset containing 14 shapes, ranging from simple geometric shapes to complex nested structures. We also evaluate on two 3D benchmarks: a subset of ShapeNet [55] and the Surface Reconstruction Benchmark (SRB) [56]. Additionally, we generate a dataset with high-genus geometries from Mehta et al. [1]. In all of our experiments, we use only the position information of the point cloud." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "type": "text", + "content": "We report the Intersection over Union (IoU), Chamfer distance, and Hausdorff distance between reconstructed shapes and ground truth shapes as metrics for surface reconstruction. In addition, we report Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and Symmetric Mean Absolute Percentage Error (SMAPE) as metrics for distance field evaluation. We further evaluate non-zero level set quality near the surface, and render the zero level" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 305, + 72, + 523, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 523, + 84 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 523, + 84 + ], + "type": "text", + "content": "set of the signed distance functions via sphere tracing." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 84, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 84, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 84, + 545, + 144 + ], + "type": "text", + "content": "We use a network with 5 hidden layers and 128 hidden channels for our method in all experiments. The hyperparameters are kept consistent within each dataset. We additionally apply importance sampling to our heat loss integral. Full training details can be found in the supplement." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 306, + 152, + 379, + 163 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 152, + 379, + 163 + ], + "spans": [ + { + "bbox": [ + 306, + 152, + 379, + 163 + ], + "type": "text", + "content": "5.1.2D Dataset" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 304, + 170, + 545, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 170, + 545, + 290 + ], + "spans": [ + { + "bbox": [ + 304, + 170, + 545, + 290 + ], + "type": "text", + "content": "An overview of our 2D dataset is attached in the supplementary material. For 2D experiments, we simply use a linear network and geometric initialization, as in DiGS [3]. We compare our method with DiGS [3] and StEik [4] with the hyperparameters and training configuration provided in their released code, and report the surface reconstruction metrics and distance field metrics in Table 1. Our method outperforms theirs across all metrics with only 8,192 sample points, whereas both DiGS [3] and StEik [4] use 15,000 sample points based on their default settings." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 305, + 297, + 468, + 310 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 297, + 468, + 310 + ], + "spans": [ + { + "bbox": [ + 305, + 297, + 468, + 310 + ], + "type": "text", + "content": "5.2. Ablation Study on 2D Dataset" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 304, + 316, + 545, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 316, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 304, + 316, + 545, + 375 + ], + "type": "text", + "content": "To more comprehensively compare over different losses, we conduct an ablation study on our 2D dataset using the same network architecture, parameter initialization, and sampling strategy. The only difference lies in the loss functions used. Quantitative results are shown in Table 2." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 304, + 376, + 545, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 376, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 304, + 376, + 545, + 556 + ], + "type": "text", + "content": "The last row corresponds to our HOTSPOT model, combining boundary, eikonal, and heat losses. Our method reconstructs the correct topology for all 14 shapes, while others struggle with the harder ones. It outperforms across nearly all metrics, demonstrating superior reconstruction quality and distance query accuracy. Although HOTSPOT slightly trails SAL [2] in RMSE and MAE, it captures the complex topology where SAL fails. A visualization is shown in Fig. 7. Full qualitative results and detailed experiment parameters are provided in the supplementary material. Models relying heavily on area losses [3, 4] often face a trade-off between preserving original details and eliminating redundant boundaries, which distorts the distance field (Section 4.3). Moreover, methods without our heat loss fall in local optimum easily." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 306, + 563, + 384, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 563, + 384, + 574 + ], + "spans": [ + { + "bbox": [ + 306, + 563, + 384, + 574 + ], + "type": "text", + "content": "5.3. 3D Datasets" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "We evaluate our method on a processed subset [16, 19, 57] of ShapeNet [55], which provides surface points sampled on preprocessed watertight meshes for 260 shapes across 13 categories. To ensure a reasonable spatial density for passing the heat, we scale the point cloud so that " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " of the points lie within a sphere centered at the origin with a radius of 0.45. For a fair comparison, we transform our outcomes back into the coordinate system used in StEik [4] and DiGS [3]. Additionally, we implement a scheduler to gradually increase the absorption parameter " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": ", which enhances the representation of level set details." + } + ] + } + ], + "index": 36 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "text", + "content": "1281" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 63, + 70, + 216, + 160 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 216, + 160 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 216, + 160 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 216, + 160 + ], + "type": "table", + "html": "
IoU ↑dC ↓dH ↓
DiGS [3]0.78820.00550.1267
StEik [4]0.66200.00730.1425
Ours0.98700.00140.0153
RMSE ↓MAE ↓SMAPE ↓
DiGS [3]0.05970.03150.3363
StEik [4]0.07250.04190.4222
Ours0.01990.01010.0699
", + "image_path": "c558a5a646cd28ab71a02d4c0ed77d18f5aaa4e2585436c26a4a50cadd03d8dc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 226, + 71, + 558, + 148 + ], + "blocks": [ + { + "bbox": [ + 59, + 163, + 220, + 218 + ], + "lines": [ + { + "bbox": [ + 59, + 163, + 220, + 218 + ], + "spans": [ + { + "bbox": [ + 59, + 163, + 220, + 218 + ], + "type": "text", + "content": "Table 1. Comparison of 2D reconstruction and distance query. " + }, + { + "bbox": [ + 59, + 163, + 220, + 218 + ], + "type": "inline_equation", + "content": "d_{C}" + }, + { + "bbox": [ + 59, + 163, + 220, + 218 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 59, + 163, + 220, + 218 + ], + "type": "inline_equation", + "content": "d_{H}" + }, + { + "bbox": [ + 59, + 163, + 220, + 218 + ], + "type": "text", + "content": " stand for the Chamfer and Hausdorff distances respectively. Our method outperforms DiGS [3] and StEik [4] on all metrics." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 226, + 71, + 558, + 148 + ], + "lines": [ + { + "bbox": [ + 226, + 71, + 558, + 148 + ], + "spans": [ + { + "bbox": [ + 226, + 71, + 558, + 148 + ], + "type": "table", + "html": "
LBLELALDLDDLSLH| IoU ↑dC ↓dH ↓RMSE ↓MAE ↓SMAPE ↓
✓✓0.81920.00680.07120.03770.01580.1315
✓✓0.89360.00290.05790.01650.00810.1205
✓✓✓0.39070.08430.55980.29660.23881.7283
✓✓✓✓0.63380.05660.42210.20580.16161.0919
✓✓✓0.29380.03770.34440.29900.23551.6958
✓✓✓✓0.60370.04140.36560.13040.09370.7792
✓✓✓0.98510.00160.01600.01990.01050.0754
", + "image_path": "e2858aeeda5e2529951057ec424b0f9b3ae40abd69940bbe9c6be6fbab3f8d23.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 225, + 152, + 533, + 217 + ], + "lines": [ + { + "bbox": [ + 225, + 152, + 533, + 217 + ], + "spans": [ + { + "bbox": [ + 225, + 152, + 533, + 217 + ], + "type": "text", + "content": "Table 2. Ablation study of different loss combinations and metrics. From top to bottom, the loss combinations are from: IGR [10], SAL [2], DiGS w/o area loss, DiGS [3], StEik w/o loss, StEik [4], and our HOTSPOT. From left to right, the losses are boundary loss, eikonal loss, area loss, divergence loss, directional divergence loss, SAL loss, and heat loss. Bold and underlined data: optimal; bold only: suboptimal. Same below." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 72, + 224, + 261, + 427 + ], + "blocks": [ + { + "bbox": [ + 72, + 224, + 261, + 427 + ], + "lines": [ + { + "bbox": [ + 72, + 224, + 261, + 427 + ], + "spans": [ + { + "bbox": [ + 72, + 224, + 261, + 427 + ], + "type": "image", + "image_path": "94247f3a293ba9f4c00834afb622f0e093198f5fe9c1f91ffd97e23b4b2bd9d0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 433, + 287, + 521 + ], + "lines": [ + { + "bbox": [ + 46, + 433, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 433, + 287, + 521 + ], + "type": "text", + "content": "Figure 7. Our heat loss achieves the best reconstruction of triangle fragments. Dashed lines show real boundaries, while the boldest line represents reconstructed ones. (c) and (f) are from the original models by DiGS [3] and StEik [4]; the rest are from our ablation study with different loss combinations. Other combinations fail to remove extra boundaries or accurately capture the topology. More fundamentally, (c), (f), (g), and (h) fall into local optima, leading to incorrect depressions and elevations." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 531, + 286, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 531, + 286, + 686 + ], + "spans": [ + { + "bbox": [ + 46, + 531, + 286, + 686 + ], + "type": "text", + "content": "We present the quantitative results in Table 3 and visual results in Fig. 9. More baselines [46, 51] are compared in the supplement. Our HOTSPOT model outperforms current state-of-the-art models across all surface reconstruction metrics, and achieves near-top accuracy in distance queries. While SAL [2] uses the closest point distance in their loss function, leading to slightly lower total absolute error, their performance drops significantly in relative error and near-surface metrics. This suggests that the closest point distance alone is unreliable near the surface. In contrast, as mentioned at the end of Section 3, after the interpolation driven by spectral bias [58], our loss remains effective in faithfully capturing the distance to the interpolated surface." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 689, + 286, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 286, + 714 + ], + "type": "text", + "content": "Fig. 1 and Fig. 8 show high-genus surface reconstruction with shapes from Mehta et al. [1]. Existing methods strug" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 221, + 361, + 265 + ], + "blocks": [ + { + "bbox": [ + 307, + 221, + 361, + 265 + ], + "lines": [ + { + "bbox": [ + 307, + 221, + 361, + 265 + ], + "spans": [ + { + "bbox": [ + 307, + 221, + 361, + 265 + ], + "type": "image", + "image_path": "94f5f2ec8c7d661f1a9ad20120447a46903ab1372fc13043b58d72c717fb34e9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 364, + 209, + 394, + 217 + ], + "lines": [ + { + "bbox": [ + 364, + 209, + 394, + 217 + ], + "spans": [ + { + "bbox": [ + 364, + 209, + 394, + 217 + ], + "type": "text", + "content": "DiGS" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 365, + 220, + 405, + 264 + ], + "blocks": [ + { + "bbox": [ + 365, + 220, + 405, + 264 + ], + "lines": [ + { + "bbox": [ + 365, + 220, + 405, + 264 + ], + "spans": [ + { + "bbox": [ + 365, + 220, + 405, + 264 + ], + "type": "image", + "image_path": "e13d43aa3c5556beb76b1652d1eec0209d81ec4c2c70d0f5955d645630cb935f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 408, + 209, + 436, + 217 + ], + "lines": [ + { + "bbox": [ + 408, + 209, + 436, + 217 + ], + "spans": [ + { + "bbox": [ + 408, + 209, + 436, + 217 + ], + "type": "text", + "content": "StEi" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 410, + 220, + 451, + 264 + ], + "blocks": [ + { + "bbox": [ + 410, + 220, + 451, + 264 + ], + "lines": [ + { + "bbox": [ + 410, + 220, + 451, + 264 + ], + "spans": [ + { + "bbox": [ + 410, + 220, + 451, + 264 + ], + "type": "image", + "image_path": "6b046b7715874413a0a47c7cc94400126fd77442358555c004bbb85c872f3fed.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 454, + 209, + 483, + 217 + ], + "lines": [ + { + "bbox": [ + 454, + 209, + 483, + 217 + ], + "spans": [ + { + "bbox": [ + 454, + 209, + 483, + 217 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 455, + 220, + 495, + 264 + ], + "blocks": [ + { + "bbox": [ + 455, + 220, + 495, + 264 + ], + "lines": [ + { + "bbox": [ + 455, + 220, + 495, + 264 + ], + "spans": [ + { + "bbox": [ + 455, + 220, + 495, + 264 + ], + "type": "image", + "image_path": "51473219e1974c3d7d89b5eda798f76b03294d3ccfb47189da02b7379c8299f8.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 501, + 209, + 539, + 217 + ], + "lines": [ + { + "bbox": [ + 501, + 209, + 539, + 217 + ], + "spans": [ + { + "bbox": [ + 501, + 209, + 539, + 217 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 501, + 220, + 541, + 264 + ], + "blocks": [ + { + "bbox": [ + 501, + 220, + 541, + 264 + ], + "lines": [ + { + "bbox": [ + 501, + 220, + 541, + 264 + ], + "spans": [ + { + "bbox": [ + 501, + 220, + 541, + 264 + ], + "type": "image", + "image_path": "4cc90aa9a24d5ef8849bd094c0e4db3f967f0f19c53c8a8aded6dd61ec77dcd2.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 307, + 266, + 362, + 303 + ], + "blocks": [ + { + "bbox": [ + 307, + 266, + 362, + 303 + ], + "lines": [ + { + "bbox": [ + 307, + 266, + 362, + 303 + ], + "spans": [ + { + "bbox": [ + 307, + 266, + 362, + 303 + ], + "type": "image", + "image_path": "0e38d075dfaa22550ddf3e77464741a018f298be9719ab07e82e09ee7a23ea2d.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 314, + 545, + 347 + ], + "lines": [ + { + "bbox": [ + 305, + 314, + 545, + 347 + ], + "spans": [ + { + "bbox": [ + 305, + 314, + 545, + 347 + ], + "type": "text", + "content": "Figure 8. Visual examples on high genus dataset. Our method uses significantly fewer iterations but reconstructs the topology correctly for all shapes, whereas others generate extra boundaries." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 365, + 266, + 406, + 304 + ], + "blocks": [ + { + "bbox": [ + 365, + 266, + 406, + 304 + ], + "lines": [ + { + "bbox": [ + 365, + 266, + 406, + 304 + ], + "spans": [ + { + "bbox": [ + 365, + 266, + 406, + 304 + ], + "type": "image", + "image_path": "06b660072fafac8a3561dee2c69f91c20118edae6c7674cc45e0a6841c2ddcdf.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 410, + 266, + 451, + 304 + ], + "blocks": [ + { + "bbox": [ + 410, + 266, + 451, + 304 + ], + "lines": [ + { + "bbox": [ + 410, + 266, + 451, + 304 + ], + "spans": [ + { + "bbox": [ + 410, + 266, + 451, + 304 + ], + "type": "image", + "image_path": "97216776a9c905a4e26418fc18109823757d679276a448d746dc55173ce82563.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 455, + 266, + 496, + 304 + ], + "blocks": [ + { + "bbox": [ + 455, + 266, + 496, + 304 + ], + "lines": [ + { + "bbox": [ + 455, + 266, + 496, + 304 + ], + "spans": [ + { + "bbox": [ + 455, + 266, + 496, + 304 + ], + "type": "image", + "image_path": "b0dc0dca7a8ae1435404d6da9cc989eecaa1ce3deb3fa2595a29a74e56d8affa.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 501, + 266, + 542, + 304 + ], + "blocks": [ + { + "bbox": [ + 501, + 266, + 542, + 304 + ], + "lines": [ + { + "bbox": [ + 501, + 266, + 542, + 304 + ], + "spans": [ + { + "bbox": [ + 501, + 266, + 542, + 304 + ], + "type": "image", + "image_path": "6613d601fe160001a5bef3c7278680164b4b79e260668de84a5f166bcdfddfd6.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 351, + 545, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 351, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 305, + 351, + 545, + 398 + ], + "type": "text", + "content": "gle with the complex topology, while our method accurately reconstructs both the surface and the level set. In the supplement, we present evaluations on other high-genus shapes from Mehta et al. and the SRB dataset." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 406, + 400, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 406, + 400, + 419 + ], + "spans": [ + { + "bbox": [ + 306, + 406, + 400, + 419 + ], + "type": "text", + "content": "5.4. Sphere Tracing" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 305, + 425, + 545, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 425, + 545, + 484 + ], + "spans": [ + { + "bbox": [ + 305, + 425, + 545, + 484 + ], + "type": "text", + "content": "Fig. 10 visualizes the number of steps a sphere tracer [21] takes to render the shapes obtained by SAL, DiGS, StEik, and our method. Our method has the fewest iterations on average, since it provides the most accurate level sets. More results can be found in the supplement." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 306, + 491, + 375, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 491, + 375, + 502 + ], + "spans": [ + { + "bbox": [ + 306, + 491, + 375, + 502 + ], + "type": "text", + "content": "6. Discussion" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 510, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 605 + ], + "type": "text", + "content": "Efficiency. Our method uses a loss function of only the neural network output and its first-order derivatives, and thus can be more efficient than previous methods which involve second-order derivatives. We report the runtime of our method and StEik [4] for training a model from ShapeNet [55] over 100 iterations using a single NVIDIA A10 GPU with varying network sizes, while keeping all other settings identical, as shown in Tab. 4." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 606, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 606, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 606, + 545, + 713 + ], + "type": "text", + "content": "Determining absorption coefficient " + }, + { + "bbox": [ + 304, + 606, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 606, + 545, + 713 + ], + "type": "text", + "content": ". The absorption coefficient " + }, + { + "bbox": [ + 304, + 606, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 606, + 545, + 713 + ], + "type": "text", + "content": " has the units of " + }, + { + "bbox": [ + 304, + 606, + 545, + 713 + ], + "type": "inline_equation", + "content": "[\\mathrm{L}]^{-1}" + }, + { + "bbox": [ + 304, + 606, + 545, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 606, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{L}" + }, + { + "bbox": [ + 304, + 606, + 545, + 713 + ], + "type": "text", + "content": " denotes a length unit. Rescaling the space is equivalent to rescaling " + }, + { + "bbox": [ + 304, + 606, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 606, + 545, + 713 + ], + "type": "text", + "content": ". While a larger " + }, + { + "bbox": [ + 304, + 606, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 606, + 545, + 713 + ], + "type": "text", + "content": " can improve accuracy, it introduces two challenges: (1) our heat loss includes the factor " + }, + { + "bbox": [ + 304, + 606, + 545, + 713 + ], + "type": "inline_equation", + "content": "e^{-2\\lambda |u|}" + }, + { + "bbox": [ + 304, + 606, + 545, + 713 + ], + "type": "text", + "content": ", which may become smaller than the precision limit of floating-point numbers, leading to round-off errors; (2) even if round-off errors are ignored, a large " + }, + { + "bbox": [ + 304, + 606, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\lambda |u|" + }, + { + "bbox": [ + 304, + 606, + 545, + 713 + ], + "type": "text", + "content": " may reduce the optimizer's ability to effectively shape the field. Never" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1282" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 100, + 70, + 492, + 225 + ], + "blocks": [ + { + "bbox": [ + 100, + 70, + 492, + 225 + ], + "lines": [ + { + "bbox": [ + 100, + 70, + 492, + 225 + ], + "spans": [ + { + "bbox": [ + 100, + 70, + 492, + 225 + ], + "type": "image", + "image_path": "c9cdbbe5acae0741015467a61070f3a6dc308ec795d0542b217896e9e4e51dbf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 226, + 546, + 249 + ], + "lines": [ + { + "bbox": [ + 46, + 226, + 546, + 249 + ], + "spans": [ + { + "bbox": [ + 46, + 226, + 546, + 249 + ], + "type": "text", + "content": "Figure 9. Visual results of DiGS [3], StEik [4], and HOTSPOT on ShapeNet [55]. Previous methods generate extra boundaries and fail to escape local optima, while our method successfully reconstructs the correct topology." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 100, + 254, + 493, + 316 + ], + "blocks": [ + { + "bbox": [ + 100, + 254, + 493, + 316 + ], + "lines": [ + { + "bbox": [ + 100, + 254, + 493, + 316 + ], + "spans": [ + { + "bbox": [ + 100, + 254, + 493, + 316 + ], + "type": "table", + "html": "
IoU ↑dC ↓dH ↓RMSE ↓MAE ↓SMAPE ↓RMSE0.1MAE0.1SMAPE0.1
SAL [2]0.74000.00740.08510.02510.01420.13440.02450.01820.6848
SIREN wo/ n [34]0.48740.00510.05580.50090.42611.26940.05130.03820.8858
DiGS [3]0.96360.00310.04350.11940.07250.21400.01520.00810.1760
StEik [4]0.96410.00320.03680.03870.02480.09310.01470.00810.1770
Ours0.97960.00290.02500.02810.01760.05400.00940.00470.1206
", + "image_path": "f0d2b96dc1acd0982419a390263ffd82eabf51617c1c117380061c8b1dcde589.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 50, + 361, + 286, + 564 + ], + "blocks": [ + { + "bbox": [ + 50, + 361, + 286, + 564 + ], + "lines": [ + { + "bbox": [ + 50, + 361, + 286, + 564 + ], + "spans": [ + { + "bbox": [ + 50, + 361, + 286, + 564 + ], + "type": "image", + "image_path": "a68e5841683847a7d1fcf5c211e241cf47ccfe1a61181eebf86458bd14e70734.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 566, + 287, + 622 + ], + "lines": [ + { + "bbox": [ + 46, + 566, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 46, + 566, + 287, + 622 + ], + "type": "text", + "content": "Figure 10. Visualizations and statistics of sphere tracing [21] iteration counts required to locate the surface. In the visual results brighter pixels indicate higher iteration counts. The histograms show the distribution of iterations for all pixels across 10 different camera poses. HOTSPOT finds the surface in fewer steps." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 628, + 286, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 628, + 286, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 628, + 286, + 700 + ], + "type": "text", + "content": "theless, our loss still pushes " + }, + { + "bbox": [ + 46, + 628, + 286, + 700 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 46, + 628, + 286, + 700 + ], + "type": "text", + "content": " away from 0, preventing unwanted surfaces where " + }, + { + "bbox": [ + 46, + 628, + 286, + 700 + ], + "type": "inline_equation", + "content": "\\lambda |u|" + }, + { + "bbox": [ + 46, + 628, + 286, + 700 + ], + "type": "text", + "content": " should be large. We compensate by combining eikonal loss with heat losses. When " + }, + { + "bbox": [ + 46, + 628, + 286, + 700 + ], + "type": "inline_equation", + "content": "\\lambda |u|" + }, + { + "bbox": [ + 46, + 628, + 286, + 700 + ], + "type": "text", + "content": " is large, the eikonal loss becomes dominant. Our " + }, + { + "bbox": [ + 46, + 628, + 286, + 700 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 628, + 286, + 700 + ], + "type": "text", + "content": " scheduler also effectively shapes the distant regions. Spatially-adaptive parameters may also improve the results." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 701, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 286, + 713 + ], + "type": "text", + "content": "Necessity of firm boundary condition. The heat diffu" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 330, + 357, + 520, + 394 + ], + "blocks": [ + { + "bbox": [ + 46, + 319, + 546, + 353 + ], + "lines": [ + { + "bbox": [ + 46, + 319, + 546, + 353 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 546, + 353 + ], + "type": "text", + "content": "Table 3. Surface reconstruction metrics on ShapeNet [55], where " + }, + { + "bbox": [ + 46, + 319, + 546, + 353 + ], + "type": "inline_equation", + "content": "d_{C}" + }, + { + "bbox": [ + 46, + 319, + 546, + 353 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 319, + 546, + 353 + ], + "type": "inline_equation", + "content": "d_{H}" + }, + { + "bbox": [ + 46, + 319, + 546, + 353 + ], + "type": "text", + "content": " stand for the Chamfer and Hausdorff distances, RMSE, MAE, and SMAPE are distance field metrics, and " + }, + { + "bbox": [ + 46, + 319, + 546, + 353 + ], + "type": "inline_equation", + "content": "\\mathrm{RMSE}_{0.1}" + }, + { + "bbox": [ + 46, + 319, + 546, + 353 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 319, + 546, + 353 + ], + "type": "inline_equation", + "content": "\\mathrm{MAE}_{0.1}" + }, + { + "bbox": [ + 46, + 319, + 546, + 353 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 319, + 546, + 353 + ], + "type": "inline_equation", + "content": "\\mathrm{SMAPE}_{0.1}" + }, + { + "bbox": [ + 46, + 319, + 546, + 353 + ], + "type": "text", + "content": " are the same distance field metrics but only for points within distance 0.1 of the surface. See the supplementary material for detailed results with more baselines." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 330, + 357, + 520, + 394 + ], + "lines": [ + { + "bbox": [ + 330, + 357, + 520, + 394 + ], + "spans": [ + { + "bbox": [ + 330, + 357, + 520, + 394 + ], + "type": "table", + "html": "
Structure5 × 1285 × 2568 × 1288 × 256
StEik [4]5.65s10.62s8.78s17.00s
Ours4.49s8.43s6.79s12.86s
", + "image_path": "876ba49726d52f85772e558399d44153ef6c35f4562c120a0b7ad059d67a528f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 307, + 399, + 542, + 410 + ], + "lines": [ + { + "bbox": [ + 307, + 399, + 542, + 410 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 542, + 410 + ], + "type": "text", + "content": "Table 4. Runtime for 100 iterations with different network sizes." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 415, + 545, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 545, + 583 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 545, + 583 + ], + "type": "text", + "content": "sion of the screened Poisson equation is based on a well-established boundary condition. We enforce the boundary condition through the boundary loss over discrete points, with neural networks interpolating through spectral bias [58]. When the input points are sparse and the absorption coefficient " + }, + { + "bbox": [ + 304, + 415, + 545, + 583 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 415, + 545, + 583 + ], + "type": "text", + "content": " is high, an overly strong heat loss (" + }, + { + "bbox": [ + 304, + 415, + 545, + 583 + ], + "type": "inline_equation", + "content": "w_{b}" + }, + { + "bbox": [ + 304, + 415, + 545, + 583 + ], + "type": "text", + "content": " being too small) can tear the boundary, causing the signed distance to collapse into an unsigned one like Fig. 6. Thus, our theory highlights the importance of a high boundary weight " + }, + { + "bbox": [ + 304, + 415, + 545, + 583 + ], + "type": "inline_equation", + "content": "w_{b}" + }, + { + "bbox": [ + 304, + 415, + 545, + 583 + ], + "type": "text", + "content": ", while absorption " + }, + { + "bbox": [ + 304, + 415, + 545, + 583 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 415, + 545, + 583 + ], + "type": "text", + "content": " is adjusted to match the input point density or by scaling the point cloud itself. Our experiments show that proper parameter setting and rescaling prevent the collapse. Future research is required for very sparse boundaries, and applications to inverse rendering." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 589, + 378, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 589, + 378, + 600 + ], + "spans": [ + { + "bbox": [ + 306, + 589, + 378, + 600 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "We propose a new model for neural signed distance function optimization based on the screened Poisson equation. We analyze our loss theoretically and show that it is an asymptotically sufficient condition to the true distance, is stable both to a small perturbation and in the temporal dynamics, and penalizes large surface area. Our experiments show that we reconstruct both better surfaces and better distance approximations compared to many existing methods, especially on complex and high-genus shapes." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1283" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "text", + "content": "Acknowledgements. This work was supported in part by NSF grants 2127544, 2238839, 2100237, 2120019, and gifts from Adobe and Google. Additionally, we would like to thank Bing Xu, Zilu Li, Yash Belhe, Ishit Mehta, and Jianan Xiao for their suggestions, discussions, proofreading, and providing reference materials." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 153, + 107, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 153, + 107, + 166 + ], + "spans": [ + { + "bbox": [ + 48, + 153, + 107, + 166 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 173, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 173, + 287, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 173, + 287, + 206 + ], + "spans": [ + { + "bbox": [ + 48, + 173, + 287, + 206 + ], + "type": "text", + "content": "[1] Ishit Mehta, Manmohan Chandraker, and Ravi Ramamoorthi. A level set theory for neural implicit evolution under explicit flows. In ECCV, pages 711-729, 2022. 1, 2, 6, 7, 8, 10" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 207, + 288, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 207, + 288, + 250 + ], + "spans": [ + { + "bbox": [ + 48, + 207, + 288, + 250 + ], + "type": "text", + "content": "[2] Matan Atzmon and Yaron Lipman. Sal: Sign agnostic learning of shapes from raw data. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2565-2574, 2020. 1, 2, 6, 7, 8, 3, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 251, + 288, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 251, + 288, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 251, + 288, + 304 + ], + "type": "text", + "content": "[3] Yizhak Ben-Shabat, Chamin Hewa Koneputugodage, and Stephen Gould. DiGS: Divergence guided shape implicit neural representation for unoriented point clouds. In IEEE Conf. Comput. Vis. Pattern Recog., pages 19323-19332, 2022. 1, 2, 3, 6, 7, 8, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 306, + 287, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 306, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 48, + 306, + 287, + 350 + ], + "type": "text", + "content": "[4] Huizong Yang, Yuxin Sun, Ganesh Sundaramoorthi, and Anthony Yezzi. StEik: stabilizing the optimization of neural signed distance functions and finer shape representation. arXiv preprint arXiv:2305.18414, 2023. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 350, + 287, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 350, + 287, + 393 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 287, + 393 + ], + "type": "text", + "content": "[5] Zoë Marschner, Silvia Sellán, Hsueh-Ti Derek Liu, and Alec Jacobson. Constructive solid geometry on neural signed distance fields. In SIGGRAPH Asia 2023 Conference Papers, pages 1-12, 2023. 1, 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 394, + 287, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 394, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 48, + 394, + 287, + 437 + ], + "type": "text", + "content": "[6] Sathamangalam R Srinivasa Varadhan. On the behavior of the fundamental solution of the heat equation with variable coefficients. Communications on Pure and Applied Mathematics, 20(2):431-455, 1967. 2, 3, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 437, + 287, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 287, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 287, + 470 + ], + "type": "text", + "content": "[7] Keenan Crane, Clarisse Weischedel, and Max Wardetzky. Geodesics in heat: A new approach to computing distance based on heat flow. ACM TOG, 32(5):1-11, 2013. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 471, + 287, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 287, + 514 + ], + "type": "text", + "content": "[8] Alexander G Belyaev and Pierre-Alain Fayolle. On variational and pde-based distance function approximations. In Computer Graphics Forum, volume 34, pages 104-118. Wiley Online Library, 2015." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 515, + 287, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 515, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 48, + 515, + 287, + 548 + ], + "type": "text", + "content": "[9] Alexander Belyaev and Pierre-Alain Fayolle. An ADMM-based scheme for distance function approximation. Numerical Algorithms, 84:983-996, 2020. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 548, + 287, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 287, + 581 + ], + "type": "text", + "content": "[10] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. arXiv preprint arXiv:2002.10099, 2020. 2, 3, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 582, + 287, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 582, + 287, + 603 + ], + "spans": [ + { + "bbox": [ + 48, + 582, + 287, + 603 + ], + "type": "text", + "content": "[11] James F Blinn. A generalization of algebraic surface drawing. ACM TOG, 1(3):235-256, 1982. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 604, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 604, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 604, + 287, + 635 + ], + "type": "text", + "content": "[12] Jules Bloomenthal and Brian Wyvill. Interactive techniques for implicit modeling. Comput. Graph. (Proc. SIGGRAPH), 24(2):109-116, 1990." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 637, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 637, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 637, + 287, + 679 + ], + "type": "text", + "content": "[13] Hugues Hoppe, Tony DeRose, Tom Duchamp, John McDonald, and Werner Stuetzle. Surface reconstruction from unorganized points. Comput. Graph. (Proc. SIGGRAPH), 26(2):71-78, 1992." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[14] Brian Curless and Marc Levoy. A volumetric method for building complex models from range images. In SIGGRAPH, page 303-312, 1996." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 148 + ], + "type": "text", + "content": "[15] Shahram Izadi, David Kim, Otmar Hilliges, David Molyneaux, Richard Newcombe, Pushmeet Kohli, Jamie Shotton, Steve Hodges, Dustin Freeman, and Andrew Davison. KinectFusion: real-time 3D reconstruction and interaction using a moving depth camera. In ACM Symposium on User Interface Software and Technology, pages 559-568, 2011. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 150, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 150, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 150, + 545, + 194 + ], + "type": "text", + "content": "[16] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4460-4470, 2019. 2, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 194, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 194, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 307, + 194, + 545, + 227 + ], + "type": "text", + "content": "[17] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5939-5948, 2019." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 228, + 545, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 260 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 260 + ], + "type": "text", + "content": "[18] Thomas Davies, Derek Nowrouzezahrai, and Alec Jacobson. On the effectiveness of weight-encoded neural implicit 3d shapes. arXiv preprint arXiv:2009.09808, 2020." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 261, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 545, + 304 + ], + "type": "text", + "content": "[19] Francis Williams, Matthew Trager, Joan Bruna, and Denis Zorin. Neural Splines: Fitting 3D surfaces with infinitely-wide neural networks. In IEEE Conf. Comput. Vis. Pattern Recog., pages 9949-9958, 2021. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 305, + 545, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 305, + 545, + 349 + ], + "spans": [ + { + "bbox": [ + 308, + 305, + 545, + 349 + ], + "type": "text", + "content": "[20] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. 41(2):641-676, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 350, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 381 + ], + "type": "text", + "content": "[21] John C Hart. Sphere tracing: A geometric method for the antialiased ray tracing of implicit surfaces. The Vis. Comput., 12(10):527-545, 1996. 2, 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 383, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 545, + 426 + ], + "type": "text", + "content": "[22] Dario Seyb, Alec Jacobson, Derek Nowrouzezahrai, and Wojciech Jarosz. Non-linear sphere tracing for rendering deformed signed distance fields. ACM Trans. Graph. (Proc. SIGGRAPH Asia), 38(6), 2019." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 426, + 545, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 545, + 448 + ], + "type": "text", + "content": "[23] Eric Galin, Eric Guérin, Axel Paris, and Adrien Peytavie. Segment tracing using local Lipschitz bounds, 2020." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 449, + 545, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 491 + ], + "type": "text", + "content": "[24] Nicholas Sharp and Alec Jacobson. Spelunking the deep: guaranteed queries on general neural implicit surfaces via range analysis. ACM Trans. Graph. (Proc. SIGGRAPH), 41(4), 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 492, + 545, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 492, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 308, + 492, + 545, + 557 + ], + "type": "text", + "content": "[25] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3D shapes. In IEEE Conf. Comput. Vis. Pattern Recog., pages 11358-11367, 2021. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 559, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 601 + ], + "type": "text", + "content": "[26] Arnulph Fuhrmann, Gerrit Sobotka, and Clemens Groß. Distance fields for rapid collision detection in physically based modeling. In Proceedings of GraphiCon, volume 2003, pages 58-65, 2003. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 602, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 602, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 602, + 545, + 635 + ], + "type": "text", + "content": "[27] Eran Guendelman, Robert Bridson, and Ronald Fedkiw. Nonconvex rigid bodies with stacking. ACM TOG, 22(3):871-878, 2003." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 635, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 545, + 690 + ], + "type": "text", + "content": "[28] Miles Macklin, Kenny Erleben, Matthias Müller, Nuttapong Chentanez, Stefan Jeschke, and Zach Corse. Local optimization for robust signed distance field collision. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 3(1):1-17, 2020. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[29] Rohan Sawhney and Keenan Crane. Monte Carlo geometry processing: a grid-free approach to PDE-based methods on" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1284" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 63, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 63, + 72, + 286, + 95 + ], + "type": "text", + "content": "volumetric domains. ACM Trans. Graph. (Proc. SIGGRAPH), 39(4), 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "type": "text", + "content": "[30] Albert Pumarola, Artsiom Sanakoyeu, Lior Yariv, Ali Thabet, and Yaron Lipman. Visco grids: Surface reconstruction with viscosity and coarea grids. Advances in Neural Information Processing Systems, 35:18060-18071, 2022. 2, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 140, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 287, + 183 + ], + "type": "text", + "content": "[31] Songyou Peng, Chiyu Jiang, Yiyi Liao, Michael Niemeyer, Marc Pollefeys, and Andreas Geiger. Shape as points: A differentiable poisson solver. Advances in Neural Information Processing Systems, 34:13032-13044, 2021." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 185, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 185, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 185, + 287, + 228 + ], + "type": "text", + "content": "[32] Chao Chen, Yu-Shen Liu, and Zhizhong Han. Gridpull: Towards scalability in learning implicit representations from 3d point clouds. In Proceedings of the IEEE/cvf international conference on computer vision, pages 18322-18334, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 229, + 287, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 229, + 287, + 281 + ], + "spans": [ + { + "bbox": [ + 48, + 229, + 287, + 281 + ], + "type": "text", + "content": "[33] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. DeepSDF: Learning continuous signed distance functions for shape representation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 165-174, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 284, + 287, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 284, + 287, + 336 + ], + "spans": [ + { + "bbox": [ + 48, + 284, + 287, + 336 + ], + "type": "text", + "content": "[34] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in neural information processing systems, 33:7462-7473, 2020. 2, 8, 5, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 338, + 287, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 287, + 371 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 287, + 371 + ], + "type": "text", + "content": "[35] Guillaume Coiffier and Louis Béthune. 1-Lipschitz neural distance fields. Comput. Graph. Forum (Eurographics STARs), 43(5):e15128, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 373, + 287, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 373, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 373, + 287, + 415 + ], + "type": "text", + "content": "[36] Yue Jiang, Dantong Ji, Zhizhong Han, and Matthias Zwicker. SDFDiff: Differentiable rendering of signed distance fields for 3D shape optimization. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1251-1261, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 417, + 287, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 417, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 417, + 287, + 459 + ], + "type": "text", + "content": "[37] Petr Kellnhofer, Lars C Jebe, Andrew Jones, Ryan Spicer, Kari Pulli, and Gordon Wetzstein. Neural lumigraph rendering. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4287-4297, 2021." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 460, + 287, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 287, + 514 + ], + "type": "text", + "content": "[38] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3D representations without 3D supervision. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3504-3515, 2020." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 515, + 287, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 515, + 287, + 570 + ], + "spans": [ + { + "bbox": [ + 48, + 515, + 287, + 570 + ], + "type": "text", + "content": "[39] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. In Advances in Neural Information Processing Systems, volume 33, pages 2492-2502, 2020. 2, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 571, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 571, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 571, + 287, + 624 + ], + "type": "text", + "content": "[40] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snavely. PhySG: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5453-5462, 2021." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 625, + 287, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 680 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 680 + ], + "type": "text", + "content": "[41] Sai Bangaru, Michael Gharbi, Tzu-Mao Li, Fujun Luan, Kalyan Sunkavalli, Milos Hasan, Sai Bi, Zexiang Xu, Gilbert Bernstein, and Fredo Durand. Differentiable rendering of neural sdfs through reparameterization. In SIGGRAPH Asia Conference Proceedings, 2022." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[42] Delio Vicini, Sébastien Speierer, and Wenzel Jakob. Differentiable signed distance function rendering. ACM Trans. Graph. (Proc. SIGGRAPH), 41(4):1-18, 2022. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "text", + "content": "[43] Jingyang Zhang, Yao Yao, Shiwei Li, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Critical regularizations for neural surface reconstruction in the wild. In IEEE Conf. Comput. Vis. Pattern Recog., pages 6270-6279, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 119, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 119, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 119, + 545, + 173 + ], + "type": "text", + "content": "[44] Ruian Wang, Zixiong Wang, Yunxiao Zhang, Shuangmin Chen, Shiqing Xin, Changhe Tu, and Wenping Wang. Aligning gradient and hessian for neural signed distance function. In Advances in Neural Information Processing Systems, volume 36, 2024." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 176, + 545, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 176, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 307, + 176, + 545, + 241 + ], + "type": "text", + "content": "[45] Zixiong Wang, Yunxiao Zhang, Rui Xu, Fan Zhang, Peng-Shuai Wang, Shuangmin Chen, Shiqing Xin, Wenping Wang, and Changhe Tu. Neural-singular-hessian: Implicit neural representation of unoriented point clouds by enforcing singular hessian. ACM Transactions on Graphics (TOG), 42(6):1-14, 2023. 5, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 243, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 243, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 243, + 545, + 297 + ], + "type": "text", + "content": "[46] Qiujie Dong, Rui Xu, Pengfei Wang, Shuangmin Chen, Shiqing Xin, Xiaohong Jia, Wenping Wang, and Changhe Tu. Neurcadrecon: Neural representation for reconstructing cad surfaces by enforcing zero gaussian curvature. arXiv preprint arXiv:2404.13420, 2024. 2, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 300, + 545, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 300, + 545, + 344 + ], + "spans": [ + { + "bbox": [ + 307, + 300, + 545, + 344 + ], + "type": "text", + "content": "[47] Yesom Park, Taekyung Lee, Jooyoung Hahn, and Myungjoo Kang. " + }, + { + "bbox": [ + 307, + 300, + 545, + 344 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 307, + 300, + 545, + 344 + ], + "type": "text", + "content": "-poisson surface reconstruction in curl-free flow from point clouds. In Advances in Neural Information Processing Systems, volume 36, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 346, + 545, + 400 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 346, + 545, + 400 + ], + "spans": [ + { + "bbox": [ + 307, + 346, + 545, + 400 + ], + "type": "text", + "content": "[48] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In IEEE Conf. Comput. Vis. Pattern Recog., pages 8456-8465, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 403, + 545, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 403, + 545, + 446 + ], + "spans": [ + { + "bbox": [ + 307, + 403, + 545, + 446 + ], + "type": "text", + "content": "[49] Yaron Lipman. Phase transitions, distance functions, and implicit neural representations. In International Conference on Machine Learning, volume 139, pages 6702-6712, 2021. 2, 3, 6, 8, 9" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 449, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 449, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 307, + 449, + 545, + 480 + ], + "type": "text", + "content": "[50] Angela Dai and Matthias Nießner. Neural poisson: Indicator functions for neural fields. arXiv preprint arXiv:2211.14249, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 484, + 545, + 528 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 484, + 545, + 528 + ], + "spans": [ + { + "bbox": [ + 307, + 484, + 545, + 528 + ], + "type": "text", + "content": "[51] Baorui Ma, Zhizhong Han, Yu-Shen Liu, and Matthias Zwicker. Neural-pull: Learning signed distance functions from point clouds by learning to pull space onto surfaces. arXiv preprint arXiv:2011.13495, 2020. 2, 7, 3, 5, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 530, + 545, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 530, + 545, + 551 + ], + "spans": [ + { + "bbox": [ + 307, + 530, + 545, + 551 + ], + "type": "text", + "content": "[52] Matan Atzmon and Yaron Lipman. SALD: Sign agnostic learning with derivatives. In ICLR, 2021. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 554, + 545, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 554, + 545, + 575 + ], + "spans": [ + { + "bbox": [ + 307, + 554, + 545, + 575 + ], + "type": "text", + "content": "[53] Nicholas Sharp, Yousuf Soliman, and Keenan Crane. The vector heat method. ACM TOG, 38(3):1-19, 2019. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 578, + 545, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 578, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 307, + 578, + 545, + 610 + ], + "type": "text", + "content": "[54] Nicole Feng and Keenan Crane. A heat method for generalized signed distance. ACM Trans. Graph. (Proc. SIGGRAPH), 43(4), 2024. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 613, + 545, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 666 + ], + "type": "text", + "content": "[55] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 6, 7, 8, 5, 9" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "text", + "content": "[56] Matthew Berger, Joshua A Levine, Luis Gustavo Nonato, Gabriel Taubin, and Claudio T Silva. A benchmark for surface reconstruction. ACM Transactions on Graphics (TOG), 32(2):1-17, 2013. 6, 8" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1285" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 318 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "text", + "content": "[57] David Stutz and Andreas Geiger. Learning 3d shape completion under weak supervision. International Journal of Computer Vision, 128:1162-1181, 2020. 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 286, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 286, + 162 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 286, + 162 + ], + "type": "text", + "content": "[58] Nasim Rahaman, Aristide Baratin, Devansh Arpit, Felix Draxler, Min Lin, Fred Hamprecht, Yoshua Bengio, and Aaron Courville. On the spectral bias of neural networks. In International Conference on Machine Learning, pages 5301-5310, 2019. 7, 8, 3, 5, 11" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 163, + 286, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 163, + 286, + 217 + ], + "spans": [ + { + "bbox": [ + 49, + 163, + 286, + 217 + ], + "type": "text", + "content": "[59] Dmitry S Kulyabov, Anna V Korolkova, Tatiana R Velieva, and Migran N Gevorkyan. Numerical analysis of eikonal equation. In Saratov Fall Meeting 2018: Laser Physics, Photonic Technologies, and Molecular Modeling, volume 11066, pages 188-195. SPIE, 2019. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 219, + 286, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 219, + 286, + 251 + ], + "spans": [ + { + "bbox": [ + 49, + 219, + 286, + 251 + ], + "type": "text", + "content": "[60] Zhen-Hang Yang and Yu-Ming Chu. On approximating the modified bessel function of the second kind. Journal of inequalities and applications, 2017:1-8, 2017. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 253, + 286, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 253, + 286, + 285 + ], + "spans": [ + { + "bbox": [ + 49, + 253, + 286, + 285 + ], + "type": "text", + "content": "[61] Luciano Modica. The gradient theory of phase transitions and the minimal interface criterion. Archive for Rational Mechanics and Analysis, 98:123-142, 1987. 10" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 286, + 286, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 286, + 286, + 318 + ], + "spans": [ + { + "bbox": [ + 49, + 286, + 286, + 318 + ], + "type": "text", + "content": "[62] Peter Sternberg. The effect of a singular perturbation on nonconvex variational problems. Archive for Rational Mechanics and Analysis, 101:209-260, 1988. 10" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1286" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/How Do I Do That_ Synthesizing 3D Hand Motion and Contacts for Everyday Interactions/4f85bf33-0c00-4386-bb62-720463efe772_content_list.json b/2025/How Do I Do That_ Synthesizing 3D Hand Motion and Contacts for Everyday Interactions/4f85bf33-0c00-4386-bb62-720463efe772_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..290a2ef185b01fc70c7f180dab28e90abaa6cc1e --- /dev/null +++ b/2025/How Do I Do That_ Synthesizing 3D Hand Motion and Contacts for Everyday Interactions/4f85bf33-0c00-4386-bb62-720463efe772_content_list.json @@ -0,0 +1,1349 @@ +[ + { + "type": "text", + "text": "How Do I Do That? Synthesizing 3D Hand Motion and Contacts for Everyday Interactions", + "text_level": 1, + "bbox": [ + 143, + 130, + 854, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Aditya Prakash\\*1 Benjamin Lundell\\*2 Dmitry Andreychuk\\* David Forsyth\\*1 Saurabh Gupta\\*1 Harpreet Sawhney\\*2 University of Illinois Urbana-Champaign \\*2Microsoft https://bit.ly/LatentAct", + "bbox": [ + 145, + 202, + 849, + 273 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4fd603c9ecbd617d8a6fbbd53cb2d9e00cc626c4340a90083783b61e102ff2d8.jpg", + "image_caption": [ + "Figure 1. Interaction Trajectories. We tackle the novel task of predicting future 3D hand poses & contact maps, i.e., interaction trajectories, from a single image showing the object, text describing the action and a 3D contact point on the object, in everyday activities. We show the trajectory predicted by our method (LatentAct) & ground truth (GT) for 3 future timesteps along with the contact point. We consider 2 settings (top) Forecasting: single RGB view, action text & 3D contact points as input, (bottom) Interpolation: goal image is also provided." + ], + "image_footnote": [], + "bbox": [ + 93, + 290, + 491, + 589 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2a87569ab56d8d97375a246d6ccd9a6594d3766d4e9fdc1260255f37ad1a2d8b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 292, + 893, + 590 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 674, + 326, + 690 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We tackle the novel problem of predicting 3D hand motion and contact maps (or Interaction Trajectories) given a single RGB view, action text, and a 3D contact point on the object as input. Our approach consists of (1) Interaction Codebook: a VQVAE model to learn a latent codebook of hand poses and contact points, effectively tokenizing interaction trajectories, (2) Interaction Predictor: a transformer-decoder module to predict the interaction trajectory from test time inputs by using an indexer module to retrieve a latent affordance from the learned codebook. To train our model,", + "bbox": [ + 88, + 707, + 485, + 859 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "we develop a data engine that extracts 3D hand poses and contact trajectories from the diverse HoloAssist dataset. We evaluate our model on a benchmark that is $2.5 - 10 \\times$ larger than existing works, in terms of diversity of objects and interactions observed, and test for generalization of the model across object categories, action categories, tasks, and scenes. Experimental results show the effectiveness of our approach over transformer & diffusion baselines across all settings.", + "bbox": [ + 511, + 675, + 908, + 797 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 829, + 645, + 845 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hands interact with objects in diverse ways, often requiring varying skill levels to complete different tasks. Consider teaching someone to repair a bike. It is challenging to specify", + "bbox": [ + 511, + 856, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Part of the work was done during an internship at Microsoft", + "bbox": [ + 107, + 875, + 431, + 888 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Joint last authors, indicates equal contribution", + "bbox": [ + 109, + 888, + 357, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7026", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "what part to manipulate [41], what axis to flip a part over such that it inserts well, etc. Instead, it is easier to directly show how the hand should move and what parts of it should make contact with the object over time to complete the task. In this work, we study this problem. Given a single image of an object, a 3D contact point, and a text indicating the high-level action, we predict Interaction Trajectories, the sequence of 3D hand poses and contact maps that specify the initial and evolving hand interactions (Fig. 1), interpretable by a human. We target interaction with everyday objects that are often small, thin, transparent, occluded, and deformable (Fig. 1) for which 3D object models may not be available. Our work is a first step towards learning a motion prior for diverse interactions in everyday activities, that can be combined with 3D object models in the future.", + "bbox": [ + 89, + 90, + 480, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the absence of a 3D model, explicitly specifying contact points in the scene grounds the hand-object interaction, especially when there are multiple objects & 3D models are unavailable. This information is provided in the form of a 3D contact point. Moreover, we observe that while objects are diverse, hands interact with them in stylized ways [39]. There are only a few prototypical ways in which hands act in everyday interactions. To exploit this property, we learn a latent codebook of trajectories of 3D hand motions & contact points (Interaction Codebook, Sec. 3.1), effectively tokenizing interaction trajectories using a VQVAE [48]. Intuitively, the codebook learns the prototypical motion involved in (say) screwing in a screw that can be repurposed into motion involved in (say) closing a bottle cap. The codebook is indexed using visual inputs at test time to retrieve the closest codebook entry (Indexer module, Sec. 3.2), and aligned with the scene using the 3D contact point & visual context from the image input (transformer-decoder based Interaction Predictor, Sec. 3.3). We refer to our approach as LatentAct.", + "bbox": [ + 91, + 321, + 485, + 607 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Since existing HOI datasets with 3D annotations have constrained setups, we use egocentric videos from the HoloAssist dataset, showing everyday interactions. We leverage recent advances in 3D hand pose estimation [30] and segmentation in videos [36] to build a semi-automatic data engine (Sec. 3.4) to extract trajectories of 3D hand poses and contact maps from HoloAssist for 800 tasks performed by ordinary users across 120 object categories and 24 action categories. Our setup is $2.5\\mathrm{X} - 10\\mathrm{X}$ larger than existing works, e.g. HOT3D [2] (33 rigid objects), ARCTIC [11] (10 articulated objects), GRAB [43] (51 objects), HOI4D [28] (16 object categories), in terms of diversity of objects and interactions observed in everyday activities.", + "bbox": [ + 89, + 609, + 483, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For evaluation, we focus on generalization to 4 aspects: novel object categories, action categories, tasks and scenes. We show results on 2 settings: (1) Forecasting: the model generated predictions conditioned on the textual description of the action, current image showing the object of interaction and the 3D contact point at the current timestep. (2) Inter-", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "polation: Here, we also provide the goal image, along with the previous inputs, showing the final state of interaction. For each of these settings, we consider 2 variants: (a) hand visible in the image, (b) hand is absent in the image (often the case at start of interactions). Across this large-scale experiment setting, we find our approach to generalize better than transformer & diffusion baselines that directly learn to predict interaction trajectories from images (Sec. 4.2).", + "bbox": [ + 511, + 90, + 906, + 212 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 224, + 653, + 239 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generating HOI sequences: Recent works in HOI [3, 7, 13, 26, 44, 54, 55] have studied generating hand-object motions in different settings, e.g. conditioned on textual descriptions [7, 13], body poses [3, 13, 26, 44] for both rigid [3, 7, 13, 26] and articulated objects [4, 55]. While they focus on pick & place tasks involving simple & constrained interactions, we consider everyday activities, often involving small, thin, textureless, occluded, articulated, deformable and dynamic objects in contact for which obtaining 3D models is non-trivial. Existing works in this domain use either object-centric [4, 7] or hand-centric [44, 54] frame to represent trajectories, this becomes a limitation in our setting since 3D objects are unknown and hands may not be present in the image (often the case at the start of the interaction). Instead, we focus on predicting future 3D hand poses & contact maps from a single image showing the object, text describing the action & a 3D contact point to localize the interaction, for diverse interactions in everyday activities.", + "bbox": [ + 511, + 250, + 908, + 522 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "HOI datasets: Existing works use datasets [11, 17, 18, 20, 24, 28] collected in controlled settings which provide 3D ground truth using MoCap [11, 43] or multi-camera setups [17, 18, 20, 24, 28, 29]. However, the diversity of objects and interactions in these datasets is limited due to constrained nature of the capture setup. While recent datasets [6, 9, 40] have explored natural interactions in egocentric videos [8, 15], they only provide 2D annotations, e.g. segmentation masks [6, 9], 2D bounding boxes [40] and grasp labels [6]. To mitigate this issue, we design a data engine to extract 3D hand poses and contact maps (Sec. 3.4) from egocentric videos showing everyday interactions. Specifically, we extract 3D trajectories from HoloAssist for 800 tasks performed by ordinary users across 120 object categories and 24 action categories, which is $2.5 \\times 10 \\times$ larger than existing datasets [2, 11, 28, 43].", + "bbox": [ + 511, + 523, + 908, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "HOI representation: Hands are often represented using a parametric mesh model, e.g., MANO [37], along with the global translation and rotation of the wrist [7, 11, 13, 30, 31, 33, 51, 53, 54]. Given an object mesh, its 6DoF pose can be denoted using translation & rotation [13, 55], signed distance field [54] or using Basis Point Set (BPS [34]) distances [7, 26]. For hand-object contact, existing approaches typically consider a canonicalized representation with the hand [4, 54] or object [7] at the origin & use relative dis", + "bbox": [ + 511, + 765, + 910, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7027", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/603f40e6e920e5ab8c0ad6144b250ab3944488914530f194c78d79c380d2f865.jpg", + "image_caption": [ + "Figure 2. We represent contact points as binary masks on the hand mesh vertices. The hand mesh is represented in the camera coordinate system, consisting of the local hand pose in the MANO [38] coordinate frame and a global transformation from the MANO frame to the camera frame. These hand poses and contact points over several timesteps form interaction trajectories." + ], + "image_footnote": [], + "bbox": [ + 96, + 92, + 472, + 273 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "stances [7, 11, 12, 32, 44] or a distance field [54] from hand joints to represent contact. Since 3D hand-object motion annotations are unavailable for diverse objects, e.g., thin, occluded, deformable, we instead focus on sequences of 3D hand poses and contact maps, i.e., interaction trajectories.", + "bbox": [ + 88, + 390, + 482, + 465 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Affordances: Gibson [14] developed the concept of affordances as a set of functionalities that the environment furnishes to an agent, which can be a human, a robot, an animal, or hands. Affordances are typically represented as the contact regions on the objects [1, 5, 23, 25, 27, 28, 43, 50] & the agent [1, 19, 23, 42, 43, 47, 53, 53, 54]. They have been explored in domains like robotics [1, 42], human interactions [19, 43, 47, 49] & hand grasp synthesis [23, 42, 43, 53, 54]. Recent works have also proposed generative models for affordances synthesis [49, 53, 54].", + "bbox": [ + 88, + 465, + 482, + 617 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 632, + 181, + 647 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a single RGB view, action text, and a 3D contact point as input, we focus on predicting 3D hand poses & contact points, i.e., interaction trajectories (Fig. 1), for future timesteps. Our approach consists of: (1) Interaction Codebook: learns a latent codebook of affordances using a VQVAE (Fig. 3a), (2) Learned Indexer: maps the action text, image (showing an object) & contact point inputs to the codebook indices to obtain the corresponding embeddings (Fig. 3b), (3) Interaction Predictor: takes the queried embeddings along with text, image and contact point inputs to output the 3D interaction trajectory (Fig. 3b).", + "bbox": [ + 89, + 657, + 482, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Hand mesh: We use the MANO [37] model to represent the hand. It consists of pose $\\theta$ and shape $\\beta$ parameters, which can be converted to a 3D hand mesh. These mesh vertices lie in the MANO coordinate system, which can be transformed to the camera coordinate system at the first", + "bbox": [ + 89, + 824, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "timestep (reference frame) using a rotation and translation. This gives us the hand trajectory in the reference frame.", + "bbox": [ + 511, + 90, + 906, + 121 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Contact map: We represent the contact points as a binary mask over the MANO hand mesh vertices, i.e., a 778 dimensional binary vector for each timestep, referred to as contact map in the paper (Fig. 2). The contact points may also change over time depending on the action, e.g. twisting.", + "bbox": [ + 511, + 122, + 906, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Once the 3D hand mesh is transformed into reference frame, we estimate the trajectory of the contact points by computing the centroid of the contact points at each timestep. The contact centroid in the reference frame is used as the 3D contact point input during training. We assume a 3D contact point to be available in the first frame at test time. To input this point, we create a 3D voxel grid of resolution $16 \\times 16 \\times 16$ . We take minimum and maximum values of 3D hand trajectory (along each axis, in the reference frame) across all training sequences. This span is divided into 16 equal parts per axis, associating each voxel to 3D metric range. It is kept fixed at train & test time. Thus, the center of each voxel represents a 3D location in metric space, i.e., a 3D metric grid, in the camera coordinate system. We create a 3D gaussian heatmap centered at 3D contact point in this grid, referred to as voxelized contact points in the paper. This changes as the trajectory evolves. This is different than the contact map (binary mask over mesh vertices).", + "bbox": [ + 511, + 198, + 908, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Interaction Codebook (InterCode)", + "text_level": 1, + "bbox": [ + 511, + 482, + 812, + 497 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This is implemented as a VQVAE with a transformer encoder-decoder architecture [16]. It consists of 3 modules:", + "bbox": [ + 511, + 505, + 908, + 534 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Encoder: The hand and contact map trajectories for a fixed time horizon $T$ are passed to a transformer encoder to get the feature embedding. These features are used to query the codebook to obtain the relevant embeddings using L2 distance. Specifically, the input trajectory, consisting of MANO parameters & contact maps, is first passed to a 1-layer MLP to match the feature dimension of the transformer. We use a 1-layer transformer encoder with 1 head & feature dimension = 512. It outputs a $T \\times 512$ dimensional feature.", + "bbox": [ + 511, + 535, + 908, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Latent codebook: It is initialized as a $K \\times E$ dimensional matrix and is updated during training using exponential moving average of the embeddings output from the encoder, following recent works [16]. Concretely, we use a multilayer residual VQVAE architecture [16]. We set the number of quantizers for VQVAE to 6, $K = 512$ and $E = 512$ . When querying the codebook during training, sampling is done using the Gumbel-SoftMax trick [21] to preserve differentiability. We set the temperature of Gumbel-Softmax to 0.5 & inject noise at training for stochastic sampling. At test time, argmax is used to get the closest codebook entry.", + "bbox": [ + 511, + 672, + 908, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Decoder: It takes multiple inputs: (1) the latent embeddings queried from the codebook $(T \\times 512)$ , (2) CLIP [35] embeddings of the text describing the action (512 dimensional feature extended for Timesteps), (3) DeiT [46] embeddings", + "bbox": [ + 511, + 839, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7028", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/aecf9241ff992cc297975b0a79be9bf9f0f79944f5ad80ef75d575a7213ac360.jpg", + "image_caption": [ + "(a) Interaction Codebook is a VQVAE model that learns a latent codebook of 3D interaction trajectories, consisting of a transformer encoder-decoder architecture. The encoder features are used to sample codebook indices and corresponding embeddings which are passed to a decoder that also takes in the video input and text describing the action and reconstructs the 3D interaction trajectories." + ], + "image_footnote": [], + "bbox": [ + 98, + 87, + 480, + 212 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a489bae29a12395225aed3b557f007ab2e69ccf0305c4770b44cf9479adcbc53.jpg", + "image_caption": [ + "(b) We train an Indexer module to map the inputs at test time to the codebook indices to extract the relevant embeddings. These are then passed to a interaction predictor module that outputs the 3D interaction trajectories. We consider 2 settings: (1) Forecasting: text describing the action, single image & 3D contact point, (2) Interpolation: also providing the goal image showing the final state of the interaction (not shown here for clarity).", + "Figure 3. Overview. Our framework involves a 2-stage training procedure: (left) Interaction Codebook: to learn a latent codebook of hand poses and contact points, i.e., tokenizing interaction trajectories, (right) a learned Indexer & an Interaction Predictor module to predict the interaction trajectories from single image, action text & 3D contact point. We use pretrained features for images (from DeiT [46]) and text (from CLIP [35]). 3D contact point is input as a 3D gaussian heatmap in a 3D voxel grid (omitted here for clarity)." + ], + "image_footnote": [], + "bbox": [ + 511, + 88, + 903, + 207 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "of the video showing the actions $(T \\times 768)$ , (4) voxelized contact points at each timestep, which are passed through four 3D convolution layers. We also pass in a 4D grid with each voxel containing the 3D location of its center in metric space, and process with 3D convolution layers. This gives us a 32 dimensional feature vector from the contact module. Concatenating all these input features, we get a 1824-dimensional joint feature that is passed through a 1-layer MLP to change the dimensionality to 512 and then fed to a transformer decoder (1-layer with 1 head & dropout of 0.2). The query embeddings for the transformer decoder are set as trainable parameters. It outputs a $T \\times 512$ feature map which is passed to separate decoders for predicting the MANO parameters and the contact map. We use a 2-layer MLP to predict the MANO parameters and a 3-layer MLP to predict the contact maps for the entire time horizon $T$ .", + "bbox": [ + 88, + 369, + 485, + 611 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training: The loss function consists of several terms: (1) Smooth L1 loss on the MANO parameters, (2) L1 loss on the contact centroid, (3) L1 loss on the translation between the future frame and first frame, (4) L1 loss on the 6D rotation representation between the future frame and first frame, (5) binary cross-entropy loss on the contact map predictions.", + "bbox": [ + 89, + 611, + 483, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Learned Indexer", + "text_level": 1, + "bbox": [ + 89, + 712, + 256, + 726 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "At test time, we need to query the learned codebook to get the latent interaction embeddings. This is done using a learned module which takes the test time inputs, i.e. text describing an action, image showing an object & a 3D contact point, and outputs a probability distribution over the codebook indices. The indices and corresponding embeddings can be sampled using the predicted probabilities. Note that the test time inputs only consist of a single image and a 3D contact point, unlike the entire trajectory used for training InterCode.", + "bbox": [ + 89, + 734, + 483, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, we first compute features from different inputs: CLIP embedding of text (512), DeiT embeddings of", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "images (768), features from voxelized contact point, in the reference frame, after passing through 4 3D convolutional layers (32) and the 3D contact point. We concatenate these features and process through a 2-layer MLP (1024 & 512 nodes) to match the input dimensions of the transformer classifier module. It is implemented as a 1 layer, 1 head transformer decoder layer that takes in learnable query embeddings and aforementioned features from different modalities to output $T \\times 512$ dimensional features. These features are then passed to a 3-layer MLP to output a probability distribution over the indices of the codebook.", + "bbox": [ + 511, + 369, + 908, + 535 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training: It consists of a cross-entropy loss with ground truth computed as the closest codebook indices (L2 distance) from the encoder features of the InterCode module.", + "bbox": [ + 511, + 536, + 908, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Interaction Predictor (InterPred)", + "text_level": 1, + "bbox": [ + 511, + 590, + 805, + 606 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since the inputs at test time differ from the decoder inputs of the InterCode module during training, we need a separate learned module to predict the interaction trajectories. It consists of a transformer decoder module with the following inputs: (1) the latent embeddings queried from the codebook using the learned prior, (2) CLIP embeddings of the text describing the action, (3) DeiT embeddings of the image showing the object, and (4) voxelized contact point in the reference frame. These inputs are processed in the same way as for Learned Indexer before passing to the transformer decoder. The architecture is same as the decoder in InterCode. While the inputs are different than InterCode, they are passed to an MLP to match transformer dimension.", + "bbox": [ + 511, + 613, + 908, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training: The loss function consists of several terms: (1) Smooth L1 loss on the MANO parameters, (2) L1 loss on the contact centroid, (3) L1 loss on the translation between the future frame and first frame, (4) L1 loss on the 6D rotation representation between the future frame and first frame, (5) binary cross-entropy loss on the contact map predictions.", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7029", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/545cd36b3f1354f3f05c212dd7e401b3661fc3b8615379fc74767c4c5b169365.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 93, + 87, + 478, + 167 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/83221cfffeb490e8dee0df08fa3c171e73b2485445b4bc94f5743b95ad0e35a2.jpg", + "image_caption": [ + "Generated annotations for the task pull lever", + "Figure 4. Data engine. (top) Object masks are extracted using SAMv2 [36], 3D hand poses & masks (2D rendering of mesh) are from HaMeR [30], contact points are computed by projecting the 3D hand points into the 2D contact region (intersection of hand & object masks). (bottom) Generated object masks (highlighted in white), 3D hand mesh & contact points for 3 timesteps." + ], + "image_footnote": [], + "bbox": [ + 96, + 185, + 478, + 251 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Data Engine", + "text_level": 1, + "bbox": [ + 89, + 359, + 223, + 376 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Existing lab datasets are limited to simple interactions in constrained setups. Instead, we use egocentric videos from HoloAssist showing everyday interactions with rigid, articulated & deformable objects, e.g. operating machines, tool use, kitchen tasks. It consists of $2\\mathrm{k}+$ videos with temporally localized atomic actions, e.g. grab, open, screw, mix, rotate, align, slide, etc. with diverse objects. We extract these clips of atomic actions and convert them into the required format for training & evaluation. We consider 24 action categories with 120 object categories spanning 800 tasks. This is 2.5- $10\\times$ larger than existing datasets, e.g. HOT3D [2] (33 rigid objects), ARCTIC [11] (10 articulated objects), GRAB [43] (51 objects), HOI4D [28] (16 object categories), in terms of objects & diversity of interactions. While HoloAssist contains diverse interactions, it does not provide 3D annotations for hand poses & contact points. To extract these annotations, we design a semi-automatic data pipeline (Fig. 4) using 2D segmentation masks, 3D hand poses & 2D contact region.", + "bbox": [ + 89, + 383, + 483, + 655 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Segmentation masks: We adopt a semi-automatic approach for object masks, that involves clicking a point on the target object in the first frame of a video & tracking the object mask across the video using SAMv2 [36]. For 2D hand masks, we render the 3D hand mesh (described next) into the image.", + "bbox": [ + 89, + 657, + 482, + 732 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3D hand poses: We use the off-the-shelf HaMeR [30] model to get 3D hand meshes in each video frame. HaMeR takes an image crop around the hand as input (estimated from Hands23 model [6]). The predicted 3D mesh in each frame is transformed to the reference frame using known intrinsics & extrinsics available in the HoloAssist dataset.", + "bbox": [ + 89, + 733, + 483, + 823 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3D Contact points: We first compute a 2D contact region in an image as the overlap between the hand & object masks. We add gaussian noise at the boundary to obtain a dense 2D contact region. We back project the 2D contact region into 3D hand mesh to obtain 3D contact points on the hand mesh.", + "bbox": [ + 89, + 825, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 513, + 89, + 643, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Protocol", + "text_level": 1, + "bbox": [ + 513, + 114, + 614, + 130 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Task: We predict trajectories of MANO hand parameters & contact map (a binary mask over 778 vertices of the MANO [38] hand mesh) in 2 settings: (1) Forecasting: input consists of a textual description of the action, a current image showing the object of interaction, and a 3D contact point. (2) Interpolation: We provide the goal image, along with the previous inputs, showing the final state of interaction. For each setting, we 2 two variants: (a) current image with a hand, (b) current image without a hand (removed via in-painting [52]) since the hand may not be visible before the start of the interaction in many practical scenarios. We only consider right hand motion in this work.", + "bbox": [ + 511, + 138, + 906, + 319 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Generalization aspects: We study generalization capabilities for 4 different settings: (1) Object categories, (2) Action categories, (3) Novel tasks: a combination of object & action categories, (4) Novel scene: we hold out videos from 1 location (data is collected in 2 locations). For each setting, we create train, validation & test splits in the ratio 80:10:10.", + "bbox": [ + 511, + 321, + 906, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics: (1) MPJPE: Mean Per Joint Position Error between predicted & ground truth hand joints, averaged over all future timesteps. (2) MPJPE-PA is the procrustus aligned variant, which optimizes for a single rotation, translation & scale to align the entire trajectory with ground truth trajectory, before computing MPJPE. (3) F1 score: harmonic mean of precision & recall to measure accuracy of contact maps.", + "bbox": [ + 511, + 412, + 906, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines: Since no prior works predict interaction trajectories from image input, we adapt 2 recent baselines from human pose literature to work in our setting: (1) HCTFormer: The input image is encoded using a ViT [10] encoder followed by a transformer decoder, similar to the pose decoder used in T2P [22]. TCP decoder takes in trajectory query (for global intent), pose query & pose embeddings. In our case, intent comes from text. We replace pose embeddings with {image, contact, text} features, pass them to an MLP & then feed to {TRM, MLP} modules, as in TCP. (2) HCTDiff: We modify the motion diffusion model (MDM [45]) for hand trajectory and contact map prediction. For this, we change conditioning to use {image, text, contact} features but diffusion architecture is same. Both HCTFormer and HCTDiff are trained in the same setting as ours for fair comparisons.", + "bbox": [ + 511, + 520, + 908, + 747 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Existing methods that use object-centric representations [4, 7] can not be applied in our setting since we do not have the 3D object models. Hand-centric representations [44, 54] are also not suitable in our setting since the hand is not always visible in the input image at test time.", + "bbox": [ + 511, + 748, + 906, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our approach involves a 2-stage training procedure: InterCode followed by Indexer & InterPred, from different inputs. We train with both transformer & diffusion variants for InterPred, referred to as LatentAct & LatentAct-Diff respectively. LatentAct-Diff uses same InterCode & Indexer as", + "bbox": [ + 511, + 825, + 908, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7030", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/315f8a4fe9627890556c1ed5b267a30cebe823a9dc7a35a32e2fb59e0669e4f1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTask-levelObject-levelAction-levelScene-level
M-PE(cm)↓M-PA(cm)↓F1↑M-PE(cm)↓M-PA(cm)↓F1↑M-PE(cm)↓M-PA(cm)↓F1↑M-PE(cm)↓M-PA(cm)↓F1↑
Hand visibleForecasting
HCTFormer8.323.060.728.703.190.728.593.250.738.523.090.72
HCTDiff8.422.750.729.012.880.739.113.100.749.492.840.71
LatentAct (Ours)7.612.990.758.093.060.758.143.320.777.873.040.74
Interpolation
HCTFormer7.523.020.768.103.100.767.773.290.787.543.000.76
HCTDiff8.302.720.778.522.910.788.383.160.788.602.810.77
LatentAct (Ours)6.722.870.807.483.020.787.253.190.807.272.930.78
No handsForecasting
HCTFormer8.342.950.728.853.020.738.543.210.748.533.050.72
HCTDiff8.852.650.729.742.810.729.503.110.749.212.740.71
LatentAct (Ours)7.932.970.768.263.010.768.193.280.777.903.000.75
Interpolation
HCTFormer7.302.950.768.423.150.768.013.300.777.292.990.77
HCTDiff8.272.760.768.682.910.769.333.100.7810.022.820.76
LatentAct (Ours)7.132.890.797.242.970.797.443.210.807.122.890.78
", + "bbox": [ + 94, + 88, + 906, + 333 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "LatentAct but different InterPred, i.e., an iterative denoising model from MDM instead of a single step prediction. The denoising is done in latent space of the codebook followed by a decoder to generate trajectories. It is trained using L2 loss on denoised embedding and loss in Sec. 3.3.", + "bbox": [ + 89, + 414, + 480, + 491 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Results", + "text_level": 1, + "bbox": [ + 89, + 500, + 184, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Generalization results: First, we study if the model can generate trajectories from the text describing the action, single image & 3D contact point, i.e., Forecasting task. In the absence of a goal, this is quite challenging and inherently multimodal in nature. We compare with the GT trajectories for all 4 generalization aspects: novel objects, actions, tasks & scene. We report MPJPE, MPJPE-PA for hand trajectories & F1-score for contact maps. We observe that our approach leads to consistent gains in absolute hand poses & contact maps. We show results (Tab. 1) on 2 variants of the task, (1) hand visible in the image, (2) hand absent in the image, often the case at the start of interactions. The hand visible in the image makes the tasks slightly easier since the initial pose is visible, as also evident from the results.", + "bbox": [ + 88, + 522, + 482, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Next, we also provide the goal image, along with the previous inputs, showing the final state of interaction, referred to as goal-conditioned interpolation. This is slightly easier than the Forecasting task since the final state of the interaction is also provided. The results (Tab. 1) on 2 varaints of the task, (1) hand visible in the image, (2) hand absent in the image, show similar benefits as in Forecasting.", + "bbox": [ + 89, + 734, + 482, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "These results are for $T = 30$ timesteps. We also report results for $T = 16$ in the supplementary with similar trends.", + "bbox": [ + 89, + 839, + 482, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results on ARCTIC: We also show results (Tab. 2) on ARCTIC [11] to check if LatentAct leads to gains on existing", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e44fe53372e239988e18729858f9f93d3ae0c2cf4f5cd985012d77ddf000b4d9.jpg", + "table_caption": [ + "Table 1. Generalization results. We report MPJPE (M-PE: cm), MPJPE-PA (M-PA: cm) & F1-score in 4 generalization settings: novel tasks, objects, actions & scene, to measure the accuracy of the predicted trajectories. We adapt two recent methods from human pose literature to work with image inputs, i.e., HCTFormer: ViT encoded image features passed to a transformer decoder (similar to pose decoder of T2P [22]), HCTDiff: MDM [45] modified to take image features as well. LatentAct leads to better absolute hand poses & contact maps." + ], + "table_footnote": [], + "table_body": "
MethodMPJPE↓MPJPE-PA↓F1↑
HCTFormer: trained on ARCTIC15.763.610.36
LatentAct: trained on ARCTIC14.773.760.41
LatentAct: zero-shot from Holo15.723.710.19
LatentAct: Codebook & Indexer from Holo15.363.580.45
", + "bbox": [ + 514, + 412, + 903, + 487 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. HoloAssist is significantly larger than ARCTIC in terms of contact sequences leading to benefits in zero-shot generalization & transferring models trained on HoloAssist to ARCTIC.", + "bbox": [ + 511, + 497, + 906, + 540 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "lab datasets. ARCTIC has accurate 3D labels but is limited in scale ( $10 \\times$ smaller than HoloAssist for contact sequences). When trained only on ARCTIC, we find the trends to transfer from HoloAssist, i.e., LatentAct is better than HCTFormer. Using codebook & indexer trained on HoloAssist further benefits models trained on ARCTIC. Zero-shot transfer from HoloAssist is competitive to models trained on ARCTIC.", + "bbox": [ + 511, + 550, + 906, + 655 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Benefits of Interaction Codebook: We verify if the 2-stage training procedure is beneficial by comparing with single-stage training methods HCTFormer & HCTDiff in Tab. 4 and observe consistent benefits. Another option is to retrieve an interaction trajectory from the training set, conditioned on the input. However, this does not scale well with the training data & is computationally expensive, e.g. the learned codebook in InterCode has 512 entries whereas the training dataset has $\\sim 15\\mathrm{K}$ sequences, which is significantly larger.", + "bbox": [ + 511, + 656, + 908, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Contact maps helps hand predictions: Here, we train the InterCode & InterPred modules with only the hand pose loss, i.e., removing the loss on contact maps. In Tab. 5, we observe that loss on contact maps helps with hand predictions.", + "bbox": [ + 511, + 794, + 906, + 854 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Trends with LatentAct-Diff.: Training InterPred with diffusion losses leads to better hand poses but worse contact maps compared to LatentAct, across all settings (Tab. 3).", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7031", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/99ebcb5415c45814e09171271b1dbf84db2f4f428f4c4e56bc9c0442e4d5ecca.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTask-levelObject-levelAction-levelScene-level
M-PE(cm)↓M-PA(cm)↓F1↑M-PE(cm)↓M-PA(cm)↓F1↑M-PE(cm)↓M-PA(cm)↓F1↑M-PE(cm)↓M-PA(cm)↓F1↑
Hand visibleForecasting
LatentAct7.612.990.758.093.060.758.143.320.777.873.040.74
LatentAct-Diff7.903.030.718.493.200.728.293.340.728.323.100.72
Interpolation
LatentAct6.722.870.807.483.020.787.253.190.807.272.930.78
LatentAct-Diff6.532.620.787.232.810.787.113.100.796.702.840.78
No handsForecasting
LatentAct7.932.970.768.263.010.768.193.280.777.903.000.75
LatentAct-Diff7.822.890.718.633.060.718.193.150.727.932.960.73
Interpolation
LatentAct7.132.890.797.242.970.797.443.210.807.122.890.78
LatentAct-Diff6.602.650.767.732.780.767.423.040.786.892.710.77
", + "bbox": [ + 94, + 88, + 903, + 287 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4b72fc47b89c6054713915770dafe41844ae10b6b661076e84c571abd9a192dd.jpg", + "table_caption": [ + "Table 3. LatentAct-Diff trends. Training InterPred with diffusion loss leads to better hand poses but worse contact maps than LatentAct." + ], + "table_footnote": [], + "table_body": "
MethodMPJPE↓MPJPE-PA↓F1↑
HCTFormer7.523.020.76
HCTFormer + InterCode6.722.870.80
HCTDiff8.302.720.77
HCTDiff + InterCode6.532.620.78
", + "bbox": [ + 106, + 324, + 470, + 406 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1a55bf12358673256d2af4439faa8ca1c6ea5c68fbb8f295467f405f0dccfb73.jpg", + "image_caption": [ + "Figure 5. More training data helps both LatentAct & HCTFormer." + ], + "image_footnote": [], + "bbox": [ + 94, + 467, + 478, + 545 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Dataset scale: In Fig. 5, we see that both LatentAct & HCTFormer improve with larger training dataset size.", + "bbox": [ + 89, + 582, + 483, + 613 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Visualizations", + "text_level": 1, + "bbox": [ + 89, + 622, + 235, + 636 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare the predictions of LatentAct & the best baseline on a few examples from task level generalization setting (Fig. 6) for both forecasting & interpolation for 3 timesteps $(t = 5, 15, 25)$ . The left column shows the input image with the contact point (projected in the image in cyan blob), goal image (for interpolation task) & text describing the action. The other columns show: (a) Camera View: predicted hand in the camera view with the contact point, this captures the placement of the hand around the contact point, (b) Another View: this better visualizes the hand articulation from a different camera viewpoint, (c) Contact Maps are shown as red parts of the hand mesh. LatentAct leads to better orientation of hand & sharper contact maps than the baseline. Note that these tasks are not seen during training. More visualizations & analysis are provided in the supplementary. Limitations: We do not predict the object state change. While 3D object models are not available, alternatives like", + "bbox": [ + 88, + 643, + 485, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/16081278a5f7f2890cd19a4c15c8a69d26ca25370bf3a80523f1f690e7278801.jpg", + "table_caption": [ + "Table 4. Two stage training with InterCode improves over single-stage training methods HCTFormer & HCTDiff. Note that HCT-Former + InterCode is same as LatentAct." + ], + "table_footnote": [], + "table_body": "
MethodMPJPE↓MPJPE-PA↓F1↑
HCTFormer7.523.020.76
- no contact map7.652.89-
LatentAct6.722.870.80
- no contact map7.762.91-
HCTDiff8.302.720.77
- no contact map8.432.78-
LatentAct-Diff6.532.620.78
- no contact map7.092.90-
", + "bbox": [ + 527, + 324, + 893, + 469 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Loss on contact maps helps hand predictions for all models.", + "bbox": [ + 511, + 479, + 903, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "predicting 2D object masks could be explored. While we assume a 3D contact point to be available, it can also be estimated using off-the-shelf depth models & ground truth intrinsics to get the 3D location of a 2D point that can be prompted by a user (e.g. by clicking a point on the image). Our experiments measure whether the predicted interaction trajectory is accurate, but it does not account for the multimodal nature of future prediction. The annotations in our work are programatically generated and may not be accurate in all cases. While we filter out the erroneous cases, using a better annotation tool [47] would be beneficial.", + "bbox": [ + 511, + 511, + 908, + 678 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 693, + 633, + 709 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our model predicts interaction trajectories, i.e. 3D hand motion & contact maps, from single RGB view, action text & 3D contact point as input. It consists of: (1) Interaction Codebook: a VQVAE to learn a latent codebook of 3D hand poses & contact points, i.e., tokenizing interactions. (2) Interaction Predictor: a transformer-decoder module to predict the interaction trajectory from test time inputs, by using an indexer module to retrieve a latent interaction from the learned codebook. Our large-scale experiments on the diverse HoloAssist dataset show benefits across 4 generalization settings. Our work is a first step towards learning a motion prior on interaction sequences and can be augmented", + "bbox": [ + 511, + 719, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7032", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/cf32ca31901cabed8b9368942629133b7bba977a85ba0a23923a3d4cdca3f66c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Input Image (t=0)Camera ViewAnother ViewContact Map
mix/stir coffeeBaselinet=5t=15t=25t=5t=15t=25t=5t=15t=25
GT.........
.........
rotate lensBaseline.........
GT.........
.........
pull batteryBaseline.........
GT.........
.........
slide controllerBaseline.........
GT.........
", + "bbox": [ + 127, + 87, + 883, + 767 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 6. Visualizations. We compare the predictions of LatentAct & the best baseline on a few examples from the task level generalization setting for both forecasting & interpolation for 3 timesteps (t=5, 15, 25). The left column shows the input image with the contact point (projected in the image in cyan blob), goal image (for interpolation task) & text describing the action. The other columns show: (a) Camera View: predicted hand in the camera view with the contact point, this captures the placement of the hand around the contact point, (b) Another View: this better visualizes the hand pose from a different camera viewpoint, (c) Contact Maps are shown as red parts of the hand mesh. LatentAct leads to better orientation of hand & sharper contact maps than the baseline. Note that these tasks are not seen during training.", + "bbox": [ + 89, + 776, + 906, + 861 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "with object trajectories with better 3D models in the future.", + "bbox": [ + 89, + 869, + 485, + 883 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our data pipeline is modular and can also incorporate better", + "bbox": [ + 89, + 883, + 483, + 898 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "hand & segmentation models in future.", + "bbox": [ + 513, + 869, + 772, + 883 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7033", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments: This material is based upon work supported by an NSF CAREER Award (IIS2143873), USDA-NIFA AIFARMS National AI Institute (USDA#2020-67021-32799) and NSF (IIS-2007035). Part of the work was done when Aditya Prakash was an intern at Microsoft. We acknowledge compute support by NVIDIA Academic Hardware Grant and a DURIP grant. We thank Shaowei Liu and Arjun Gupta for feedback on the draft.", + "bbox": [ + 89, + 90, + 485, + 212 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 223, + 187, + 239 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Shikhar Bahl, Russell Mendonca, Lili Chen, Unnat Jain, and Deepak Pathak. Affordances from human videos as a versatile representation for robotics. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3", + "[2] Prithviraj Banerjee, Sindi Shkodrani, Pierre Moulon, Shreyas Hampali, Fan Zhang, Jade Fountain, Edward Miller, Selen Basol, Richard Newcombe, Robert Wang, Jakob Julian Engel, and Tomas Hodan. Introducing hot3d: An egocentric dataset for 3d hand and object tracking. arXiv, 2406.09598, 2024. 2, 5", + "[3] Jona Braun, Sammy Joe Christen, Muhammed Kocabas, Emre Aksan, and Otmar Hilliges. Physically plausible full-body hand-object interaction synthesis. arXiv, 2309.07907, 2023. 2", + "[4] Junuk Cha, Jihyeon Kim, Jae Shin Yoon, and Seungryul Baek. Text2hoi: Text-guided 3d motion generation for handobject interaction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 5", + "[5] Matthew Chang, Aditya Prakash, and Saurabh Gupta. Look ma, no hands! agent-environment factorization of egocentric videos. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 3", + "[6] Tianyi Cheng, Dandan Shan, Ayda Sultan Hassen, Richard Ely Locke Higgins, and David Fouhey. Towards a richer 2d understanding of hands at scale. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 2, 5", + "[7] Sammy Christen, Shreyas Hampali, Fadime Sener, Edoardo Remelli, Tomas Hodan, Eric Sauser, Shugao Ma, and Bugra Tekin. Diffh2o: Diffusion-based synthesis of hand-object interactions from textual descriptions. arXiv, 2403.17827, 2024. 2, 3, 5", + "[8] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Sanja Fidler, Antonino Furnari, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, et al. Scaling egocentric vision: The epic-kitchens dataset. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 2", + "[9] Ahmad Darkhalil, Dandan Shan, Bin Zhu, Jian Ma, Amlan Kar, Richard Higgins, Sanja Fidler, David Fouhey, and Dima Damen. Epic-kitchens visor benchmark: Video segmentations and object relations. In NeurIPS Track on Datasets and Benchmarks, 2022. 2", + "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image" + ], + "bbox": [ + 93, + 247, + 485, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "is worth 16x16 words: Transformers for image recognition at scale. In Proceedings of the International Conference on Learning Representations (ICLR), 2021. 5", + "[11] Zicong Fan, Omid Taheri, Dimitrios Tzionas, Muhammed Kocabas, Manuel Kaufmann, Michael J. Black, and Otmar Hilliges. ARCTIC: A dataset for dexterous bimanual handobject manipulation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3, 5, 6", + "[12] Zicong Fan, Takehiko Ohkawa, Linlin Yang, Nie Lin, Zhishan Zhou, Shihao Zhou, Jiajun Liang, Zhong Gao, Xuanyang Zhang, Xue Zhang, Fei Li, Zheng Liu, Feng Lu, Karim Abou Zeid, Bastian Leibe, Jeongwan On, Seungryul Baek, Aditya Prakash, Saurabh Gupta, Kun He, Yoichi Sato, Otmar Hilliges, Hyung Jin Chang, and Angela Yao. Benchmarks and challenges in pose estimation for egocentric hand interactions with objects. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 3", + "[13] Anindita Ghosh, Rishabh Dabral, Vladislav Golyanik, Christian Theobalt, and Philipp Slusallek. Imos: Intent-driven full-body motion synthesis for human-object interactions. EuroGraphics, 42(2):1-12, 2023. 2", + "[14] J.J. Gibson. The ecological approach to visual perception, 1979. 3", + "[15] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[16] Chuan Guo, Yuxuan Mu, Muhammad Gohar Javed, Sen Wang, and Li Cheng. Momask: Generative masked modeling of 3d human motions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3", + "[17] Shreyas Hampali, Mahdi Rad, Markus Oberweger, and Vincent Lepetit. Honnotate: A method for 3d annotation of hand and object poses. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2", + "[18] Shreyas Hampali, Sayan Deb Sarkar, Mahdi Rad, and Vincent Lepetit. Keypoint transformer: Solving joint identification in challenging hands and object interactions for accurate 3d pose estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[19] Mohamed Hassan, Partha Ghosh, Joachim Tesch, Dimitrios Tzionas, and Michael J. Black. Populating 3d scenes by learning human-scene interaction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3", + "[20] Di Huang, Xiaopeng Ji, Xingyi He, Jiaming Sun, Tong He, Qing Shuai, Wanli Ouyang, and Xiaowei Zhou. Reconstructing hand-held objects from monocular video. In ACM Transactions on Graphics, 2022. 2", + "[21] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. In Proceedings of the International Conference on Learning Representations (ICLR), 2017. 3", + "[22] Jaewoo Jeong, Daehee Park, and Kuk-Jin Yoon. Multi-agent long-term 3d human pose forecasting via interaction-aware" + ], + "bbox": [ + 516, + 92, + 906, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7034", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "trajectory conditioning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 5, 6", + "[23] Hanwen Jiang, Shaowei Liu, Jiashun Wang, and Xiaolong Wang. Hand-object contact consistency reasoning for human grasps generation. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. 3", + "[24] Taein Kwon, Bugra Tekin, Jan Stühmer, Federica Bogo, and Marc Pollefeys. H2o: Two hands manipulating objects for first person interaction recognition. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. 2", + "[25] Gen Li, Varun Jampani, Deqing Sun, and Laura Sevilla-Lara. LOCATE: localize and transfer object parts for weakly supervised affordance grounding. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3", + "[26] Quanzhou Li, Jingbo Wang, Chen Change Loy, and Bo Dai. Task-oriented human-object interactions generation with implicit neural representations. In Proceedings of the IEEE Winter Conference on Applications of Computer Vision (WACV), 2024. 2", + "[27] Shaowei Liu, Yang Zhou, Jimei Yang, Saurabh Gupta, and Shenlong Wang. Contactgen: Generative contact modeling for grasp generation. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2023. 3", + "[28] Yunze Liu, Yun Liu, Che Jiang, Kangbo Lyu, Weikang Wan, Hao Shen, Boqiang Liang, Zhoujie Fu, He Wang, and Li Yi. HOI4D: A 4d egocentric dataset for category-level human-object interaction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 5", + "[29] Takehiko Ohkawa, Kun He, Fadime Sener, Tomas Hodan, Luan Tran, and Cem Keskin. Assemblyhands: Towards egocentric activity understanding via 3d hand pose estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 12999-13008, 2023. 2", + "[30] Georgios Pavlakos, Dandan Shan, Iija Radosavovic, Angjoo Kanazawa, David Fouhey, and Jitendra Malik. Reconstructing hands in 3d with transformers. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 5", + "[31] Aditya Prakash, Matthew Chang, Matthew Jin, and Saurabh Gupta. 3d reconstruction of objects in hands without real world 3d supervision. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 2", + "[32] Aditya Prakash, Arjun Gupta, and Saurabh Gupta. Mitigating perspective distortion-induced shape ambiguity in image crops. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 3", + "[33] Aditya Prakash, Ruisen Tu, Matthew Chang, and Saurabh Gupta. 3d hand pose estimation in everyday egocentric images. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 2", + "[34] Sergey Prokudin, Christoph Lassner, and Javier Romero. Efficient learning on point clouds with basis point sets. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2019. 2" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the International Conference on Machine Learning (ICML), 2021. 3, 4", + "[36] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dollár, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos. arXiv, 2408.00714, 2024. 2, 5", + "[37] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: modeling and capturing hands and bodies together. ACM Transactions on Graphics, 36(6):245:1-245:17, 2017. 2, 3", + "[38] Javier Romero, Dimitrios Tzionas, and Michael J Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Transactions on Graphics (ToG), 2017. 3, 5", + "[39] Rómer Rosales and Stan Sclaroff. Combining generative and discriminative models in a framework for articulated pose estimation. International Journal of Computer Vision, 67: 251-276, 2006. 2", + "[40] Dandan Shan, Jiaqi Geng, Michelle Shu, and David F Fouhey. Understanding human hands in contact at internet scale. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2", + "[41] Rajinder S Sodhi, Brett R Jones, David Forsyth, Brian P Bailey, and Giuliano Maciocci. Bethere: 3d mobile collaboration with spatial input. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, pages 179-188, 2013. 2", + "[42] Mohan Kumar Srirama, Sudeep Dasari, Shikhhar Bahl, and Abhinav Gupta. HRP: human affordances for robotic pretraining. arxiv, 2407.18911, 2024. 3", + "[43] Omid Taheri, Nima Ghorbani, Michael J Black, and Dimitrios Tzionas. GRAB: A dataset of whole-body human grasping of objects. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2, 3, 5", + "[44] Omid Taheri, Yi Zhou, Dimitrios Tzionas, Yang Zhou, Duygu Ceylan, Soren Pirk, and Michael J. Black. GRIP: Generating interaction poses using latent consistency and spatial cues. In Proceedings of the International Conference on 3D Vision (3DV), 2024. 2, 3, 5", + "[45] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit Haim Bermano. Human motion diffusion model. In Proceedings of the International Conference on Learning Representations (ICLR), 2023. 5, 6", + "[46] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In Proceedings of the International Conference on Machine Learning (ICML), 2021. 3, 4", + "[47] Shashank Tripathi, Agniv Chatterjee, Jean-Claude Passy, Hongwei Yi, Dimitrios Tzionas, and Michael J. Black. DECO:" + ], + "bbox": [ + 516, + 90, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7035", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "dense estimation of 3d human-scene contact in the wild. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2023. 3, 7", + "[48] Aäron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu. Neural discrete representation learning. In Advances in Neural Information Processing Systems (NeurIPS), 2017. 2", + "[49] Sirui Xu, Zhengyuan Li, Yu-Xiong Wang, and Liang-Yan Gui. Interdiff: Generating 3d human-object interactions with physics-informed diffusion. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2023. 3", + "[50] Yuhang Yang, Wei Zhai, Hongchen Luo, Yang Cao, Jiebo Luo, and Zheng-Jun Zha. Grounding 3d object affordance from 2d interactions in images. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2023. 3", + "[51] Yufei Ye, Abhinav Gupta, and Shubham Tulsiani. What's in your hands? 3D reconstruction of generic objects in hands. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[52] Yufei Ye, Poorvi Hebbar, Abhinav Gupta, and Shubham Tulsiani. Diffusion-guided reconstruction of everyday handobject interaction clips. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2023. 5", + "[53] Yufei Ye, Xueting Li, Abhinav Gupta, Shalini De Mello, Stan Birchfield, Jiaming Song, Shubham Tulsiani, and Sifei Liu. Affordance diffusion: Synthesizing hand-object interactions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3", + "[54] Yufei Ye, Abhinav Gupta, Kris Kitani, and Shubham Tulsiani. G-hop: Generative hand-object prior for interaction reconstruction and grasp synthesis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 3, 5", + "[55] Hui Zhang, Sammy Joe Christen, Zicong Fan, Luocheng Zheng, Jemin Hwangbo, Jie Song, and Otmar Hilliges. Artigrasp: Physically plausible synthesis of bi-manual dexterous grasping and articulation. arXiv, 2309.03891, 2023. 2" + ], + "bbox": [ + 91, + 90, + 483, + 628 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "7036", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/How Do I Do That_ Synthesizing 3D Hand Motion and Contacts for Everyday Interactions/4f85bf33-0c00-4386-bb62-720463efe772_model.json b/2025/How Do I Do That_ Synthesizing 3D Hand Motion and Contacts for Everyday Interactions/4f85bf33-0c00-4386-bb62-720463efe772_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7b8b3246c2f60953b043b2631b1f4141b8b44c2a --- /dev/null +++ b/2025/How Do I Do That_ Synthesizing 3D Hand Motion and Contacts for Everyday Interactions/4f85bf33-0c00-4386-bb62-720463efe772_model.json @@ -0,0 +1,2004 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.144, + 0.131, + 0.856, + 0.177 + ], + "angle": 0, + "content": "How Do I Do That? Synthesizing 3D Hand Motion and Contacts for Everyday Interactions" + }, + { + "type": "text", + "bbox": [ + 0.147, + 0.203, + 0.85, + 0.275 + ], + "angle": 0, + "content": "Aditya Prakash\\*1 Benjamin Lundell\\*2 Dmitry Andreychuk\\* David Forsyth\\*1 Saurabh Gupta\\*1 Harpreet Sawhney\\*2 University of Illinois Urbana-Champaign \\*2Microsoft https://bit.ly/LatentAct" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.291, + 0.492, + 0.59 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.293, + 0.895, + 0.591 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.599, + 0.908, + 0.657 + ], + "angle": 0, + "content": "Figure 1. Interaction Trajectories. We tackle the novel task of predicting future 3D hand poses & contact maps, i.e., interaction trajectories, from a single image showing the object, text describing the action and a 3D contact point on the object, in everyday activities. We show the trajectory predicted by our method (LatentAct) & ground truth (GT) for 3 future timesteps along with the contact point. We consider 2 settings (top) Forecasting: single RGB view, action text & 3D contact points as input, (bottom) Interpolation: goal image is also provided." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.675, + 0.327, + 0.691 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.708, + 0.486, + 0.86 + ], + "angle": 0, + "content": "We tackle the novel problem of predicting 3D hand motion and contact maps (or Interaction Trajectories) given a single RGB view, action text, and a 3D contact point on the object as input. Our approach consists of (1) Interaction Codebook: a VQVAE model to learn a latent codebook of hand poses and contact points, effectively tokenizing interaction trajectories, (2) Interaction Predictor: a transformer-decoder module to predict the interaction trajectory from test time inputs by using an indexer module to retrieve a latent affordance from the learned codebook. To train our model," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.676, + 0.909, + 0.799 + ], + "angle": 0, + "content": "we develop a data engine that extracts 3D hand poses and contact trajectories from the diverse HoloAssist dataset. We evaluate our model on a benchmark that is \\(2.5 - 10 \\times\\) larger than existing works, in terms of diversity of objects and interactions observed, and test for generalization of the model across object categories, action categories, tasks, and scenes. Experimental results show the effectiveness of our approach over transformer & diffusion baselines across all settings." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.83, + 0.646, + 0.846 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.857, + 0.908, + 0.904 + ], + "angle": 0, + "content": "Hands interact with objects in diverse ways, often requiring varying skill levels to complete different tasks. Consider teaching someone to repair a bike. It is challenging to specify" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.876, + 0.432, + 0.889 + ], + "angle": 0, + "content": "*Part of the work was done during an internship at Microsoft" + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.889, + 0.359, + 0.902 + ], + "angle": 0, + "content": "† Joint last authors, indicates equal contribution" + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.876, + 0.432, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7026" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.318 + ], + "angle": 0, + "content": "what part to manipulate [41], what axis to flip a part over such that it inserts well, etc. Instead, it is easier to directly show how the hand should move and what parts of it should make contact with the object over time to complete the task. In this work, we study this problem. Given a single image of an object, a 3D contact point, and a text indicating the high-level action, we predict Interaction Trajectories, the sequence of 3D hand poses and contact maps that specify the initial and evolving hand interactions (Fig. 1), interpretable by a human. We target interaction with everyday objects that are often small, thin, transparent, occluded, and deformable (Fig. 1) for which 3D object models may not be available. Our work is a first step towards learning a motion prior for diverse interactions in everyday activities, that can be combined with 3D object models in the future." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.322, + 0.486, + 0.608 + ], + "angle": 0, + "content": "In the absence of a 3D model, explicitly specifying contact points in the scene grounds the hand-object interaction, especially when there are multiple objects & 3D models are unavailable. This information is provided in the form of a 3D contact point. Moreover, we observe that while objects are diverse, hands interact with them in stylized ways [39]. There are only a few prototypical ways in which hands act in everyday interactions. To exploit this property, we learn a latent codebook of trajectories of 3D hand motions & contact points (Interaction Codebook, Sec. 3.1), effectively tokenizing interaction trajectories using a VQVAE [48]. Intuitively, the codebook learns the prototypical motion involved in (say) screwing in a screw that can be repurposed into motion involved in (say) closing a bottle cap. The codebook is indexed using visual inputs at test time to retrieve the closest codebook entry (Indexer module, Sec. 3.2), and aligned with the scene using the 3D contact point & visual context from the image input (transformer-decoder based Interaction Predictor, Sec. 3.3). We refer to our approach as LatentAct." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.611, + 0.485, + 0.807 + ], + "angle": 0, + "content": "Since existing HOI datasets with 3D annotations have constrained setups, we use egocentric videos from the HoloAssist dataset, showing everyday interactions. We leverage recent advances in 3D hand pose estimation [30] and segmentation in videos [36] to build a semi-automatic data engine (Sec. 3.4) to extract trajectories of 3D hand poses and contact maps from HoloAssist for 800 tasks performed by ordinary users across 120 object categories and 24 action categories. Our setup is \\(2.5\\mathrm{X} - 10\\mathrm{X}\\) larger than existing works, e.g. HOT3D [2] (33 rigid objects), ARCTIC [11] (10 articulated objects), GRAB [43] (51 objects), HOI4D [28] (16 object categories), in terms of diversity of objects and interactions observed in everyday activities." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.485, + 0.902 + ], + "angle": 0, + "content": "For evaluation, we focus on generalization to 4 aspects: novel object categories, action categories, tasks and scenes. We show results on 2 settings: (1) Forecasting: the model generated predictions conditioned on the textual description of the action, current image showing the object of interaction and the 3D contact point at the current timestep. (2) Inter-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.213 + ], + "angle": 0, + "content": "polation: Here, we also provide the goal image, along with the previous inputs, showing the final state of interaction. For each of these settings, we consider 2 variants: (a) hand visible in the image, (b) hand is absent in the image (often the case at start of interactions). Across this large-scale experiment setting, we find our approach to generalize better than transformer & diffusion baselines that directly learn to predict interaction trajectories from images (Sec. 4.2)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.226, + 0.655, + 0.241 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.251, + 0.909, + 0.523 + ], + "angle": 0, + "content": "Generating HOI sequences: Recent works in HOI [3, 7, 13, 26, 44, 54, 55] have studied generating hand-object motions in different settings, e.g. conditioned on textual descriptions [7, 13], body poses [3, 13, 26, 44] for both rigid [3, 7, 13, 26] and articulated objects [4, 55]. While they focus on pick & place tasks involving simple & constrained interactions, we consider everyday activities, often involving small, thin, textureless, occluded, articulated, deformable and dynamic objects in contact for which obtaining 3D models is non-trivial. Existing works in this domain use either object-centric [4, 7] or hand-centric [44, 54] frame to represent trajectories, this becomes a limitation in our setting since 3D objects are unknown and hands may not be present in the image (often the case at the start of the interaction). Instead, we focus on predicting future 3D hand poses & contact maps from a single image showing the object, text describing the action & a 3D contact point to localize the interaction, for diverse interactions in everyday activities." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.524, + 0.909, + 0.765 + ], + "angle": 0, + "content": "HOI datasets: Existing works use datasets [11, 17, 18, 20, 24, 28] collected in controlled settings which provide 3D ground truth using MoCap [11, 43] or multi-camera setups [17, 18, 20, 24, 28, 29]. However, the diversity of objects and interactions in these datasets is limited due to constrained nature of the capture setup. While recent datasets [6, 9, 40] have explored natural interactions in egocentric videos [8, 15], they only provide 2D annotations, e.g. segmentation masks [6, 9], 2D bounding boxes [40] and grasp labels [6]. To mitigate this issue, we design a data engine to extract 3D hand poses and contact maps (Sec. 3.4) from egocentric videos showing everyday interactions. Specifically, we extract 3D trajectories from HoloAssist for 800 tasks performed by ordinary users across 120 object categories and 24 action categories, which is \\(2.5 \\times 10 \\times\\) larger than existing datasets [2, 11, 28, 43]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.911, + 0.902 + ], + "angle": 0, + "content": "HOI representation: Hands are often represented using a parametric mesh model, e.g., MANO [37], along with the global translation and rotation of the wrist [7, 11, 13, 30, 31, 33, 51, 53, 54]. Given an object mesh, its 6DoF pose can be denoted using translation & rotation [13, 55], signed distance field [54] or using Basis Point Set (BPS [34]) distances [7, 26]. For hand-object contact, existing approaches typically consider a canonicalized representation with the hand [4, 54] or object [7] at the origin & use relative dis" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7027" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.093, + 0.473, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.291, + 0.484, + 0.377 + ], + "angle": 0, + "content": "Figure 2. We represent contact points as binary masks on the hand mesh vertices. The hand mesh is represented in the camera coordinate system, consisting of the local hand pose in the MANO [38] coordinate frame and a global transformation from the MANO frame to the camera frame. These hand poses and contact points over several timesteps form interaction trajectories." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.391, + 0.483, + 0.467 + ], + "angle": 0, + "content": "stances [7, 11, 12, 32, 44] or a distance field [54] from hand joints to represent contact. Since 3D hand-object motion annotations are unavailable for diverse objects, e.g., thin, occluded, deformable, we instead focus on sequences of 3D hand poses and contact maps, i.e., interaction trajectories." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.467, + 0.483, + 0.618 + ], + "angle": 0, + "content": "Affordances: Gibson [14] developed the concept of affordances as a set of functionalities that the environment furnishes to an agent, which can be a human, a robot, an animal, or hands. Affordances are typically represented as the contact regions on the objects [1, 5, 23, 25, 27, 28, 43, 50] & the agent [1, 19, 23, 42, 43, 47, 53, 53, 54]. They have been explored in domains like robotics [1, 42], human interactions [19, 43, 47, 49] & hand grasp synthesis [23, 42, 43, 53, 54]. Recent works have also proposed generative models for affordances synthesis [49, 53, 54]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.633, + 0.182, + 0.648 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.483, + 0.825 + ], + "angle": 0, + "content": "Given a single RGB view, action text, and a 3D contact point as input, we focus on predicting 3D hand poses & contact points, i.e., interaction trajectories (Fig. 1), for future timesteps. Our approach consists of: (1) Interaction Codebook: learns a latent codebook of affordances using a VQVAE (Fig. 3a), (2) Learned Indexer: maps the action text, image (showing an object) & contact point inputs to the codebook indices to obtain the corresponding embeddings (Fig. 3b), (3) Interaction Predictor: takes the queried embeddings along with text, image and contact point inputs to output the 3D interaction trajectory (Fig. 3b)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Hand mesh: We use the MANO [37] model to represent the hand. It consists of pose \\(\\theta\\) and shape \\(\\beta\\) parameters, which can be converted to a 3D hand mesh. These mesh vertices lie in the MANO coordinate system, which can be transformed to the camera coordinate system at the first" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.908, + 0.122 + ], + "angle": 0, + "content": "timestep (reference frame) using a rotation and translation. This gives us the hand trajectory in the reference frame." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.123, + 0.907, + 0.198 + ], + "angle": 0, + "content": "Contact map: We represent the contact points as a binary mask over the MANO hand mesh vertices, i.e., a 778 dimensional binary vector for each timestep, referred to as contact map in the paper (Fig. 2). The contact points may also change over time depending on the action, e.g. twisting." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.199, + 0.909, + 0.472 + ], + "angle": 0, + "content": "Once the 3D hand mesh is transformed into reference frame, we estimate the trajectory of the contact points by computing the centroid of the contact points at each timestep. The contact centroid in the reference frame is used as the 3D contact point input during training. We assume a 3D contact point to be available in the first frame at test time. To input this point, we create a 3D voxel grid of resolution \\(16 \\times 16 \\times 16\\). We take minimum and maximum values of 3D hand trajectory (along each axis, in the reference frame) across all training sequences. This span is divided into 16 equal parts per axis, associating each voxel to 3D metric range. It is kept fixed at train & test time. Thus, the center of each voxel represents a 3D location in metric space, i.e., a 3D metric grid, in the camera coordinate system. We create a 3D gaussian heatmap centered at 3D contact point in this grid, referred to as voxelized contact points in the paper. This changes as the trajectory evolves. This is different than the contact map (binary mask over mesh vertices)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.483, + 0.813, + 0.498 + ], + "angle": 0, + "content": "3.1. Interaction Codebook (InterCode)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.506, + 0.909, + 0.535 + ], + "angle": 0, + "content": "This is implemented as a VQVAE with a transformer encoder-decoder architecture [16]. It consists of 3 modules:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.536, + 0.909, + 0.673 + ], + "angle": 0, + "content": "Encoder: The hand and contact map trajectories for a fixed time horizon \\( T \\) are passed to a transformer encoder to get the feature embedding. These features are used to query the codebook to obtain the relevant embeddings using L2 distance. Specifically, the input trajectory, consisting of MANO parameters & contact maps, is first passed to a 1-layer MLP to match the feature dimension of the transformer. We use a 1-layer transformer encoder with 1 head & feature dimension = 512. It outputs a \\( T \\times 512 \\) dimensional feature." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.673, + 0.909, + 0.84 + ], + "angle": 0, + "content": "Latent codebook: It is initialized as a \\( K \\times E \\) dimensional matrix and is updated during training using exponential moving average of the embeddings output from the encoder, following recent works [16]. Concretely, we use a multilayer residual VQVAE architecture [16]. We set the number of quantizers for VQVAE to 6, \\( K = 512 \\) and \\( E = 512 \\). When querying the codebook during training, sampling is done using the Gumbel-SoftMax trick [21] to preserve differentiability. We set the temperature of Gumbel-Softmax to 0.5 & inject noise at training for stochastic sampling. At test time, argmax is used to get the closest codebook entry." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.908, + 0.901 + ], + "angle": 0, + "content": "Decoder: It takes multiple inputs: (1) the latent embeddings queried from the codebook \\((T \\times 512)\\), (2) CLIP [35] embeddings of the text describing the action (512 dimensional feature extended for Timesteps), (3) DeiT [46] embeddings" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7028" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.088, + 0.482, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.215, + 0.488, + 0.287 + ], + "angle": 0, + "content": "(a) Interaction Codebook is a VQVAE model that learns a latent codebook of 3D interaction trajectories, consisting of a transformer encoder-decoder architecture. The encoder features are used to sample codebook indices and corresponding embeddings which are passed to a decoder that also takes in the video input and text describing the action and reconstructs the 3D interaction trajectories." + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.089, + 0.905, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.511, + 0.214, + 0.909, + 0.288 + ], + "angle": 0, + "content": "(b) We train an Indexer module to map the inputs at test time to the codebook indices to extract the relevant embeddings. These are then passed to a interaction predictor module that outputs the 3D interaction trajectories. We consider 2 settings: (1) Forecasting: text describing the action, single image & 3D contact point, (2) Interpolation: also providing the goal image showing the final state of the interaction (not shown here for clarity)." + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.298, + 0.908, + 0.356 + ], + "angle": 0, + "content": "Figure 3. Overview. Our framework involves a 2-stage training procedure: (left) Interaction Codebook: to learn a latent codebook of hand poses and contact points, i.e., tokenizing interaction trajectories, (right) a learned Indexer & an Interaction Predictor module to predict the interaction trajectories from single image, action text & 3D contact point. We use pretrained features for images (from DeiT [46]) and text (from CLIP [35]). 3D contact point is input as a 3D gaussian heatmap in a 3D voxel grid (omitted here for clarity)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.37, + 0.486, + 0.612 + ], + "angle": 0, + "content": "of the video showing the actions \\((T \\times 768)\\), (4) voxelized contact points at each timestep, which are passed through four 3D convolution layers. We also pass in a 4D grid with each voxel containing the 3D location of its center in metric space, and process with 3D convolution layers. This gives us a 32 dimensional feature vector from the contact module. Concatenating all these input features, we get a 1824-dimensional joint feature that is passed through a 1-layer MLP to change the dimensionality to 512 and then fed to a transformer decoder (1-layer with 1 head & dropout of 0.2). The query embeddings for the transformer decoder are set as trainable parameters. It outputs a \\(T \\times 512\\) feature map which is passed to separate decoders for predicting the MANO parameters and the contact map. We use a 2-layer MLP to predict the MANO parameters and a 3-layer MLP to predict the contact maps for the entire time horizon \\(T\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.612, + 0.485, + 0.704 + ], + "angle": 0, + "content": "Training: The loss function consists of several terms: (1) Smooth L1 loss on the MANO parameters, (2) L1 loss on the contact centroid, (3) L1 loss on the translation between the future frame and first frame, (4) L1 loss on the 6D rotation representation between the future frame and first frame, (5) binary cross-entropy loss on the contact map predictions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.713, + 0.258, + 0.727 + ], + "angle": 0, + "content": "3.2. Learned Indexer" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.87 + ], + "angle": 0, + "content": "At test time, we need to query the learned codebook to get the latent interaction embeddings. This is done using a learned module which takes the test time inputs, i.e. text describing an action, image showing an object & a 3D contact point, and outputs a probability distribution over the codebook indices. The indices and corresponding embeddings can be sampled using the predicted probabilities. Note that the test time inputs only consist of a single image and a 3D contact point, unlike the entire trajectory used for training InterCode." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Specifically, we first compute features from different inputs: CLIP embedding of text (512), DeiT embeddings of" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.37, + 0.909, + 0.536 + ], + "angle": 0, + "content": "images (768), features from voxelized contact point, in the reference frame, after passing through 4 3D convolutional layers (32) and the 3D contact point. We concatenate these features and process through a 2-layer MLP (1024 & 512 nodes) to match the input dimensions of the transformer classifier module. It is implemented as a 1 layer, 1 head transformer decoder layer that takes in learnable query embeddings and aforementioned features from different modalities to output \\( T \\times 512 \\) dimensional features. These features are then passed to a 3-layer MLP to output a probability distribution over the indices of the codebook." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.537, + 0.909, + 0.582 + ], + "angle": 0, + "content": "Training: It consists of a cross-entropy loss with ground truth computed as the closest codebook indices (L2 distance) from the encoder features of the InterCode module." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.592, + 0.807, + 0.607 + ], + "angle": 0, + "content": "3.3. Interaction Predictor (InterPred)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.909, + 0.809 + ], + "angle": 0, + "content": "Since the inputs at test time differ from the decoder inputs of the InterCode module during training, we need a separate learned module to predict the interaction trajectories. It consists of a transformer decoder module with the following inputs: (1) the latent embeddings queried from the codebook using the learned prior, (2) CLIP embeddings of the text describing the action, (3) DeiT embeddings of the image showing the object, and (4) voxelized contact point in the reference frame. These inputs are processed in the same way as for Learned Indexer before passing to the transformer decoder. The architecture is same as the decoder in InterCode. While the inputs are different than InterCode, they are passed to an MLP to match transformer dimension." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Training: The loss function consists of several terms: (1) Smooth L1 loss on the MANO parameters, (2) L1 loss on the contact centroid, (3) L1 loss on the translation between the future frame and first frame, (4) L1 loss on the 6D rotation representation between the future frame and first frame, (5) binary cross-entropy loss on the contact map predictions." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7029" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.088, + 0.48, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.1, + 0.173, + 0.356, + 0.185 + ], + "angle": 0, + "content": "Generated annotations for the task pull lever" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.186, + 0.48, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.261, + 0.483, + 0.345 + ], + "angle": 0, + "content": "Figure 4. Data engine. (top) Object masks are extracted using SAMv2 [36], 3D hand poses & masks (2D rendering of mesh) are from HaMeR [30], contact points are computed by projecting the 3D hand points into the 2D contact region (intersection of hand & object masks). (bottom) Generated object masks (highlighted in white), 3D hand mesh & contact points for 3 timesteps." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.361, + 0.224, + 0.377 + ], + "angle": 0, + "content": "3.4. Data Engine" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.384, + 0.484, + 0.656 + ], + "angle": 0, + "content": "Existing lab datasets are limited to simple interactions in constrained setups. Instead, we use egocentric videos from HoloAssist showing everyday interactions with rigid, articulated & deformable objects, e.g. operating machines, tool use, kitchen tasks. It consists of \\(2\\mathrm{k}+\\) videos with temporally localized atomic actions, e.g. grab, open, screw, mix, rotate, align, slide, etc. with diverse objects. We extract these clips of atomic actions and convert them into the required format for training & evaluation. We consider 24 action categories with 120 object categories spanning 800 tasks. This is 2.5- \\(10\\times\\) larger than existing datasets, e.g. HOT3D [2] (33 rigid objects), ARCTIC [11] (10 articulated objects), GRAB [43] (51 objects), HOI4D [28] (16 object categories), in terms of objects & diversity of interactions. While HoloAssist contains diverse interactions, it does not provide 3D annotations for hand poses & contact points. To extract these annotations, we design a semi-automatic data pipeline (Fig. 4) using 2D segmentation masks, 3D hand poses & 2D contact region." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.658, + 0.483, + 0.733 + ], + "angle": 0, + "content": "Segmentation masks: We adopt a semi-automatic approach for object masks, that involves clicking a point on the target object in the first frame of a video & tracking the object mask across the video using SAMv2 [36]. For 2D hand masks, we render the 3D hand mesh (described next) into the image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.734, + 0.484, + 0.824 + ], + "angle": 0, + "content": "3D hand poses: We use the off-the-shelf HaMeR [30] model to get 3D hand meshes in each video frame. HaMeR takes an image crop around the hand as input (estimated from Hands23 model [6]). The predicted 3D mesh in each frame is transformed to the reference frame using known intrinsics & extrinsics available in the HoloAssist dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.901 + ], + "angle": 0, + "content": "3D Contact points: We first compute a 2D contact region in an image as the overlap between the hand & object masks. We add gaussian noise at the boundary to obtain a dense 2D contact region. We back project the 2D contact region into 3D hand mesh to obtain 3D contact points on the hand mesh." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.645, + 0.107 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.116, + 0.615, + 0.131 + ], + "angle": 0, + "content": "4.1. Protocol" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.139, + 0.907, + 0.32 + ], + "angle": 0, + "content": "Task: We predict trajectories of MANO hand parameters & contact map (a binary mask over 778 vertices of the MANO [38] hand mesh) in 2 settings: (1) Forecasting: input consists of a textual description of the action, a current image showing the object of interaction, and a 3D contact point. (2) Interpolation: We provide the goal image, along with the previous inputs, showing the final state of interaction. For each setting, we 2 two variants: (a) current image with a hand, (b) current image without a hand (removed via in-painting [52]) since the hand may not be visible before the start of the interaction in many practical scenarios. We only consider right hand motion in this work." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.322, + 0.908, + 0.413 + ], + "angle": 0, + "content": "Generalization aspects: We study generalization capabilities for 4 different settings: (1) Object categories, (2) Action categories, (3) Novel tasks: a combination of object & action categories, (4) Novel scene: we hold out videos from 1 location (data is collected in 2 locations). For each setting, we create train, validation & test splits in the ratio 80:10:10." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.414, + 0.907, + 0.52 + ], + "angle": 0, + "content": "Metrics: (1) MPJPE: Mean Per Joint Position Error between predicted & ground truth hand joints, averaged over all future timesteps. (2) MPJPE-PA is the procrustus aligned variant, which optimizes for a single rotation, translation & scale to align the entire trajectory with ground truth trajectory, before computing MPJPE. (3) F1 score: harmonic mean of precision & recall to measure accuracy of contact maps." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.521, + 0.909, + 0.748 + ], + "angle": 0, + "content": "Baselines: Since no prior works predict interaction trajectories from image input, we adapt 2 recent baselines from human pose literature to work in our setting: (1) HCTFormer: The input image is encoded using a ViT [10] encoder followed by a transformer decoder, similar to the pose decoder used in T2P [22]. TCP decoder takes in trajectory query (for global intent), pose query & pose embeddings. In our case, intent comes from text. We replace pose embeddings with {image, contact, text} features, pass them to an MLP & then feed to {TRM, MLP} modules, as in TCP. (2) HCTDiff: We modify the motion diffusion model (MDM [45]) for hand trajectory and contact map prediction. For this, we change conditioning to use {image, text, contact} features but diffusion architecture is same. Both HCTFormer and HCTDiff are trained in the same setting as ours for fair comparisons." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.749, + 0.908, + 0.825 + ], + "angle": 0, + "content": "Existing methods that use object-centric representations [4, 7] can not be applied in our setting since we do not have the 3D object models. Hand-centric representations [44, 54] are also not suitable in our setting since the hand is not always visible in the input image at test time." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Our approach involves a 2-stage training procedure: InterCode followed by Indexer & InterPred, from different inputs. We train with both transformer & diffusion variants for InterPred, referred to as LatentAct & LatentAct-Diff respectively. LatentAct-Diff uses same InterCode & Indexer as" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7030" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.907, + 0.334 + ], + "angle": 0, + "content": "
MethodTask-levelObject-levelAction-levelScene-level
M-PE(cm)↓M-PA(cm)↓F1↑M-PE(cm)↓M-PA(cm)↓F1↑M-PE(cm)↓M-PA(cm)↓F1↑M-PE(cm)↓M-PA(cm)↓F1↑
Hand visibleForecasting
HCTFormer8.323.060.728.703.190.728.593.250.738.523.090.72
HCTDiff8.422.750.729.012.880.739.113.100.749.492.840.71
LatentAct (Ours)7.612.990.758.093.060.758.143.320.777.873.040.74
Interpolation
HCTFormer7.523.020.768.103.100.767.773.290.787.543.000.76
HCTDiff8.302.720.778.522.910.788.383.160.788.602.810.77
LatentAct (Ours)6.722.870.807.483.020.787.253.190.807.272.930.78
No handsForecasting
HCTFormer8.342.950.728.853.020.738.543.210.748.533.050.72
HCTDiff8.852.650.729.742.810.729.503.110.749.212.740.71
LatentAct (Ours)7.932.970.768.263.010.768.193.280.777.903.000.75
Interpolation
HCTFormer7.302.950.768.423.150.768.013.300.777.292.990.77
HCTDiff8.272.760.768.682.910.769.333.100.7810.022.820.76
LatentAct (Ours)7.132.890.797.242.970.797.443.210.807.122.890.78
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.343, + 0.907, + 0.4 + ], + "angle": 0, + "content": "Table 1. Generalization results. We report MPJPE (M-PE: cm), MPJPE-PA (M-PA: cm) & F1-score in 4 generalization settings: novel tasks, objects, actions & scene, to measure the accuracy of the predicted trajectories. We adapt two recent methods from human pose literature to work with image inputs, i.e., HCTFormer: ViT encoded image features passed to a transformer decoder (similar to pose decoder of T2P [22]), HCTDiff: MDM [45] modified to take image features as well. LatentAct leads to better absolute hand poses & contact maps." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.415, + 0.482, + 0.492 + ], + "angle": 0, + "content": "LatentAct but different InterPred, i.e., an iterative denoising model from MDM instead of a single step prediction. The denoising is done in latent space of the codebook followed by a decoder to generate trajectories. It is trained using L2 loss on denoised embedding and loss in Sec. 3.3." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.501, + 0.185, + 0.515 + ], + "angle": 0, + "content": "4.2. Results" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.523, + 0.483, + 0.734 + ], + "angle": 0, + "content": "Generalization results: First, we study if the model can generate trajectories from the text describing the action, single image & 3D contact point, i.e., Forecasting task. In the absence of a goal, this is quite challenging and inherently multimodal in nature. We compare with the GT trajectories for all 4 generalization aspects: novel objects, actions, tasks & scene. We report MPJPE, MPJPE-PA for hand trajectories & F1-score for contact maps. We observe that our approach leads to consistent gains in absolute hand poses & contact maps. We show results (Tab. 1) on 2 variants of the task, (1) hand visible in the image, (2) hand absent in the image, often the case at the start of interactions. The hand visible in the image makes the tasks slightly easier since the initial pose is visible, as also evident from the results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.84 + ], + "angle": 0, + "content": "Next, we also provide the goal image, along with the previous inputs, showing the final state of interaction, referred to as goal-conditioned interpolation. This is slightly easier than the Forecasting task since the final state of the interaction is also provided. The results (Tab. 1) on 2 varaints of the task, (1) hand visible in the image, (2) hand absent in the image, show similar benefits as in Forecasting." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.871 + ], + "angle": 0, + "content": "These results are for \\( T = 30 \\) timesteps. We also report results for \\( T = 16 \\) in the supplementary with similar trends." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Results on ARCTIC: We also show results (Tab. 2) on ARCTIC [11] to check if LatentAct leads to gains on existing" + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.413, + 0.905, + 0.488 + ], + "angle": 0, + "content": "
MethodMPJPE↓MPJPE-PA↓F1↑
HCTFormer: trained on ARCTIC15.763.610.36
LatentAct: trained on ARCTIC14.773.760.41
LatentAct: zero-shot from Holo15.723.710.19
LatentAct: Codebook & Indexer from Holo15.363.580.45
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.498, + 0.907, + 0.541 + ], + "angle": 0, + "content": "Table 2. HoloAssist is significantly larger than ARCTIC in terms of contact sequences leading to benefits in zero-shot generalization & transferring models trained on HoloAssist to ARCTIC." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.551, + 0.907, + 0.656 + ], + "angle": 0, + "content": "lab datasets. ARCTIC has accurate 3D labels but is limited in scale (\\(10 \\times\\) smaller than HoloAssist for contact sequences). When trained only on ARCTIC, we find the trends to transfer from HoloAssist, i.e., LatentAct is better than HCTFormer. Using codebook & indexer trained on HoloAssist further benefits models trained on ARCTIC. Zero-shot transfer from HoloAssist is competitive to models trained on ARCTIC." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.657, + 0.909, + 0.794 + ], + "angle": 0, + "content": "Benefits of Interaction Codebook: We verify if the 2-stage training procedure is beneficial by comparing with single-stage training methods HCTFormer & HCTDiff in Tab. 4 and observe consistent benefits. Another option is to retrieve an interaction trajectory from the training set, conditioned on the input. However, this does not scale well with the training data & is computationally expensive, e.g. the learned codebook in InterCode has 512 entries whereas the training dataset has \\(\\sim 15\\mathrm{K}\\) sequences, which is significantly larger." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.908, + 0.855 + ], + "angle": 0, + "content": "Contact maps helps hand predictions: Here, we train the InterCode & InterPred modules with only the hand pose loss, i.e., removing the loss on contact maps. In Tab. 5, we observe that loss on contact maps helps with hand predictions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.901 + ], + "angle": 0, + "content": "Trends with LatentAct-Diff.: Training InterPred with diffusion losses leads to better hand poses but worse contact maps compared to LatentAct, across all settings (Tab. 3)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7031" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.905, + 0.288 + ], + "angle": 0, + "content": "
MethodTask-levelObject-levelAction-levelScene-level
M-PE(cm)↓M-PA(cm)↓F1↑M-PE(cm)↓M-PA(cm)↓F1↑M-PE(cm)↓M-PA(cm)↓F1↑M-PE(cm)↓M-PA(cm)↓F1↑
Hand visibleForecasting
LatentAct7.612.990.758.093.060.758.143.320.777.873.040.74
LatentAct-Diff7.903.030.718.493.200.728.293.340.728.323.100.72
Interpolation
LatentAct6.722.870.807.483.020.787.253.190.807.272.930.78
LatentAct-Diff6.532.620.787.232.810.787.113.100.796.702.840.78
No handsForecasting
LatentAct7.932.970.768.263.010.768.193.280.777.903.000.75
LatentAct-Diff7.822.890.718.633.060.718.193.150.727.932.960.73
Interpolation
LatentAct7.132.890.797.242.970.797.443.210.807.122.890.78
LatentAct-Diff6.602.650.767.732.780.767.423.040.786.892.710.77
" + }, + { + "type": "table_caption", + "bbox": [ + 0.092, + 0.298, + 0.905, + 0.313 + ], + "angle": 0, + "content": "Table 3. LatentAct-Diff trends. Training InterPred with diffusion loss leads to better hand poses but worse contact maps than LatentAct." + }, + { + "type": "table", + "bbox": [ + 0.107, + 0.325, + 0.472, + 0.407 + ], + "angle": 0, + "content": "
MethodMPJPE↓MPJPE-PA↓F1↑
HCTFormer7.523.020.76
HCTFormer + InterCode6.722.870.80
HCTDiff8.302.720.77
HCTDiff + InterCode6.532.620.78
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.417, + 0.486, + 0.459 + ], + "angle": 0, + "content": "Table 4. Two stage training with InterCode improves over single-stage training methods HCTFormer & HCTDiff. Note that HCT-Former + InterCode is same as LatentAct." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.468, + 0.48, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.56, + 0.483, + 0.575 + ], + "angle": 0, + "content": "Figure 5. More training data helps both LatentAct & HCTFormer." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.583, + 0.484, + 0.614 + ], + "angle": 0, + "content": "Dataset scale: In Fig. 5, we see that both LatentAct & HCTFormer improve with larger training dataset size." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.623, + 0.236, + 0.637 + ], + "angle": 0, + "content": "4.3. Visualizations" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.644, + 0.486, + 0.901 + ], + "angle": 0, + "content": "We compare the predictions of LatentAct & the best baseline on a few examples from task level generalization setting (Fig. 6) for both forecasting & interpolation for 3 timesteps \\((t = 5, 15, 25)\\). The left column shows the input image with the contact point (projected in the image in cyan blob), goal image (for interpolation task) & text describing the action. The other columns show: (a) Camera View: predicted hand in the camera view with the contact point, this captures the placement of the hand around the contact point, (b) Another View: this better visualizes the hand articulation from a different camera viewpoint, (c) Contact Maps are shown as red parts of the hand mesh. LatentAct leads to better orientation of hand & sharper contact maps than the baseline. Note that these tasks are not seen during training. More visualizations & analysis are provided in the supplementary. Limitations: We do not predict the object state change. While 3D object models are not available, alternatives like" + }, + { + "type": "table", + "bbox": [ + 0.529, + 0.325, + 0.895, + 0.47 + ], + "angle": 0, + "content": "
MethodMPJPE↓MPJPE-PA↓F1↑
HCTFormer7.523.020.76
- no contact map7.652.89-
LatentAct6.722.870.80
- no contact map7.762.91-
HCTDiff8.302.720.77
- no contact map8.432.78-
LatentAct-Diff6.532.620.78
- no contact map7.092.90-
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.481, + 0.905, + 0.495 + ], + "angle": 0, + "content": "Table 5. Loss on contact maps helps hand predictions for all models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.512, + 0.909, + 0.679 + ], + "angle": 0, + "content": "predicting 2D object masks could be explored. While we assume a 3D contact point to be available, it can also be estimated using off-the-shelf depth models & ground truth intrinsics to get the 3D location of a 2D point that can be prompted by a user (e.g. by clicking a point on the image). Our experiments measure whether the predicted interaction trajectory is accurate, but it does not account for the multimodal nature of future prediction. The annotations in our work are programatically generated and may not be accurate in all cases. While we filter out the erroneous cases, using a better annotation tool [47] would be beneficial." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.694, + 0.634, + 0.71 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Our model predicts interaction trajectories, i.e. 3D hand motion & contact maps, from single RGB view, action text & 3D contact point as input. It consists of: (1) Interaction Codebook: a VQVAE to learn a latent codebook of 3D hand poses & contact points, i.e., tokenizing interactions. (2) Interaction Predictor: a transformer-decoder module to predict the interaction trajectory from test time inputs, by using an indexer module to retrieve a latent interaction from the learned codebook. Our large-scale experiments on the diverse HoloAssist dataset show benefits across 4 generalization settings. Our work is a first step towards learning a motion prior on interaction sequences and can be augmented" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7032" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.128, + 0.088, + 0.885, + 0.768 + ], + "angle": 0, + "content": "
Input Image (t=0)Camera ViewAnother ViewContact Map
mix/stir coffeeBaselinet=5t=15t=25t=5t=15t=25t=5t=15t=25
GT.........
.........
rotate lensBaseline.........
GT.........
.........
pull batteryBaseline.........
GT.........
.........
slide controllerBaseline.........
GT.........
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.777, + 0.907, + 0.862 + ], + "angle": 0, + "content": "Figure 6. Visualizations. We compare the predictions of LatentAct & the best baseline on a few examples from the task level generalization setting for both forecasting & interpolation for 3 timesteps (t=5, 15, 25). The left column shows the input image with the contact point (projected in the image in cyan blob), goal image (for interpolation task) & text describing the action. The other columns show: (a) Camera View: predicted hand in the camera view with the contact point, this captures the placement of the hand around the contact point, (b) Another View: this better visualizes the hand pose from a different camera viewpoint, (c) Contact Maps are shown as red parts of the hand mesh. LatentAct leads to better orientation of hand & sharper contact maps than the baseline. Note that these tasks are not seen during training." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.87, + 0.486, + 0.884 + ], + "angle": 0, + "content": "with object trajectories with better 3D models in the future." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.885, + 0.485, + 0.9 + ], + "angle": 0, + "content": "Our data pipeline is modular and can also incorporate better" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.87, + 0.774, + 0.884 + ], + "angle": 0, + "content": "hand & segmentation models in future." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7033" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.213 + ], + "angle": 0, + "content": "Acknowledgments: This material is based upon work supported by an NSF CAREER Award (IIS2143873), USDA-NIFA AIFARMS National AI Institute (USDA#2020-67021-32799) and NSF (IIS-2007035). Part of the work was done when Aditya Prakash was an intern at Microsoft. We acknowledge compute support by NVIDIA Academic Hardware Grant and a DURIP grant. We thank Shaowei Liu and Arjun Gupta for feedback on the draft." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.224, + 0.188, + 0.24 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.248, + 0.486, + 0.317 + ], + "angle": 0, + "content": "[1] Shikhar Bahl, Russell Mendonca, Lili Chen, Unnat Jain, and Deepak Pathak. Affordances from human videos as a versatile representation for robotics. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.318, + 0.485, + 0.399 + ], + "angle": 0, + "content": "[2] Prithviraj Banerjee, Sindi Shkodrani, Pierre Moulon, Shreyas Hampali, Fan Zhang, Jade Fountain, Edward Miller, Selen Basol, Richard Newcombe, Robert Wang, Jakob Julian Engel, and Tomas Hodan. Introducing hot3d: An egocentric dataset for 3d hand and object tracking. arXiv, 2406.09598, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.402, + 0.484, + 0.455 + ], + "angle": 0, + "content": "[3] Jona Braun, Sammy Joe Christen, Muhammed Kocabas, Emre Aksan, and Otmar Hilliges. Physically plausible full-body hand-object interaction synthesis. arXiv, 2309.07907, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.457, + 0.484, + 0.513 + ], + "angle": 0, + "content": "[4] Junuk Cha, Jihyeon Kim, Jae Shin Yoon, and Seungryul Baek. Text2hoi: Text-guided 3d motion generation for handobject interaction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.513, + 0.484, + 0.568 + ], + "angle": 0, + "content": "[5] Matthew Chang, Aditya Prakash, and Saurabh Gupta. Look ma, no hands! agent-environment factorization of egocentric videos. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.568, + 0.484, + 0.624 + ], + "angle": 0, + "content": "[6] Tianyi Cheng, Dandan Shan, Ayda Sultan Hassen, Richard Ely Locke Higgins, and David Fouhey. Towards a richer 2d understanding of hands at scale. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.624, + 0.484, + 0.692 + ], + "angle": 0, + "content": "[7] Sammy Christen, Shreyas Hampali, Fadime Sener, Edoardo Remelli, Tomas Hodan, Eric Sauser, Shugao Ma, and Bugra Tekin. Diffh2o: Diffusion-based synthesis of hand-object interactions from textual descriptions. arXiv, 2403.17827, 2024. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.693, + 0.484, + 0.775 + ], + "angle": 0, + "content": "[8] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Sanja Fidler, Antonino Furnari, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, et al. Scaling egocentric vision: The epic-kitchens dataset. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.776, + 0.484, + 0.845 + ], + "angle": 0, + "content": "[9] Ahmad Darkhalil, Dandan Shan, Bin Zhu, Jian Ma, Amlan Kar, Richard Higgins, Sanja Fidler, David Fouhey, and Dima Damen. Epic-kitchens visor benchmark: Video segmentations and object relations. In NeurIPS Track on Datasets and Benchmarks, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.248, + 0.486, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.134 + ], + "angle": 0, + "content": "is worth 16x16 words: Transformers for image recognition at scale. In Proceedings of the International Conference on Learning Representations (ICLR), 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.135, + 0.908, + 0.217 + ], + "angle": 0, + "content": "[11] Zicong Fan, Omid Taheri, Dimitrios Tzionas, Muhammed Kocabas, Manuel Kaufmann, Michael J. Black, and Otmar Hilliges. ARCTIC: A dataset for dexterous bimanual handobject manipulation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.218, + 0.907, + 0.342 + ], + "angle": 0, + "content": "[12] Zicong Fan, Takehiko Ohkawa, Linlin Yang, Nie Lin, Zhishan Zhou, Shihao Zhou, Jiajun Liang, Zhong Gao, Xuanyang Zhang, Xue Zhang, Fei Li, Zheng Liu, Feng Lu, Karim Abou Zeid, Bastian Leibe, Jeongwan On, Seungryul Baek, Aditya Prakash, Saurabh Gupta, Kun He, Yoichi Sato, Otmar Hilliges, Hyung Jin Chang, and Angela Yao. Benchmarks and challenges in pose estimation for egocentric hand interactions with objects. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.343, + 0.907, + 0.398 + ], + "angle": 0, + "content": "[13] Anindita Ghosh, Rishabh Dabral, Vladislav Golyanik, Christian Theobalt, and Philipp Slusallek. Imos: Intent-driven full-body motion synthesis for human-object interactions. EuroGraphics, 42(2):1-12, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.399, + 0.907, + 0.425 + ], + "angle": 0, + "content": "[14] J.J. Gibson. The ecological approach to visual perception, 1979. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.427, + 0.906, + 0.509 + ], + "angle": 0, + "content": "[15] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.51, + 0.907, + 0.566 + ], + "angle": 0, + "content": "[16] Chuan Guo, Yuxuan Mu, Muhammad Gohar Javed, Sen Wang, and Li Cheng. Momask: Generative masked modeling of 3d human motions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.567, + 0.907, + 0.621 + ], + "angle": 0, + "content": "[17] Shreyas Hampali, Mahdi Rad, Markus Oberweger, and Vincent Lepetit. Honnotate: A method for 3d annotation of hand and object poses. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.622, + 0.906, + 0.691 + ], + "angle": 0, + "content": "[18] Shreyas Hampali, Sayan Deb Sarkar, Mahdi Rad, and Vincent Lepetit. Keypoint transformer: Solving joint identification in challenging hands and object interactions for accurate 3d pose estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.692, + 0.907, + 0.76 + ], + "angle": 0, + "content": "[19] Mohamed Hassan, Partha Ghosh, Joachim Tesch, Dimitrios Tzionas, and Michael J. Black. Populating 3d scenes by learning human-scene interaction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.761, + 0.907, + 0.816 + ], + "angle": 0, + "content": "[20] Di Huang, Xiaopeng Ji, Xingyi He, Jiaming Sun, Tong He, Qing Shuai, Wanli Ouyang, and Xiaowei Zhou. Reconstructing hand-held objects from monocular video. In ACM Transactions on Graphics, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.907, + 0.871 + ], + "angle": 0, + "content": "[21] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. In Proceedings of the International Conference on Learning Representations (ICLR), 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.872, + 0.906, + 0.902 + ], + "angle": 0, + "content": "[22] Jaewoo Jeong, Daehee Park, and Kuk-Jin Yoon. Multi-agent long-term 3d human pose forecasting via interaction-aware" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7034" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.133 + ], + "angle": 0, + "content": "trajectory conditioning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.484, + 0.19 + ], + "angle": 0, + "content": "[23] Hanwen Jiang, Shaowei Liu, Jiashun Wang, and Xiaolong Wang. Hand-object contact consistency reasoning for human grasps generation. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.484, + 0.259 + ], + "angle": 0, + "content": "[24] Taein Kwon, Bugra Tekin, Jan Stühmer, Federica Bogo, and Marc Pollefeys. H2o: Two hands manipulating objects for first person interaction recognition. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.26, + 0.484, + 0.328 + ], + "angle": 0, + "content": "[25] Gen Li, Varun Jampani, Deqing Sun, and Laura Sevilla-Lara. LOCATE: localize and transfer object parts for weakly supervised affordance grounding. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.484, + 0.398 + ], + "angle": 0, + "content": "[26] Quanzhou Li, Jingbo Wang, Chen Change Loy, and Bo Dai. Task-oriented human-object interactions generation with implicit neural representations. In Proceedings of the IEEE Winter Conference on Applications of Computer Vision (WACV), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.484, + 0.455 + ], + "angle": 0, + "content": "[27] Shaowei Liu, Yang Zhou, Jimei Yang, Saurabh Gupta, and Shenlong Wang. Contactgen: Generative contact modeling for grasp generation. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.456, + 0.484, + 0.537 + ], + "angle": 0, + "content": "[28] Yunze Liu, Yun Liu, Che Jiang, Kangbo Lyu, Weikang Wan, Hao Shen, Boqiang Liang, Zhoujie Fu, He Wang, and Li Yi. HOI4D: A 4d egocentric dataset for category-level human-object interaction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.484, + 0.609 + ], + "angle": 0, + "content": "[29] Takehiko Ohkawa, Kun He, Fadime Sener, Tomas Hodan, Luan Tran, and Cem Keskin. Assemblyhands: Towards egocentric activity understanding via 3d hand pose estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 12999-13008, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.609, + 0.484, + 0.678 + ], + "angle": 0, + "content": "[30] Georgios Pavlakos, Dandan Shan, Iija Radosavovic, Angjoo Kanazawa, David Fouhey, and Jitendra Malik. Reconstructing hands in 3d with transformers. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.484, + 0.734 + ], + "angle": 0, + "content": "[31] Aditya Prakash, Matthew Chang, Matthew Jin, and Saurabh Gupta. 3d reconstruction of objects in hands without real world 3d supervision. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.484, + 0.789 + ], + "angle": 0, + "content": "[32] Aditya Prakash, Arjun Gupta, and Saurabh Gupta. Mitigating perspective distortion-induced shape ambiguity in image crops. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.484, + 0.845 + ], + "angle": 0, + "content": "[33] Aditya Prakash, Ruisen Tu, Matthew Chang, and Saurabh Gupta. 3d hand pose estimation in everyday egocentric images. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[34] Sergey Prokudin, Christoph Lassner, and Javier Romero. Efficient learning on point clouds with basis point sets. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.188 + ], + "angle": 0, + "content": "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the International Conference on Machine Learning (ICML), 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.19, + 0.908, + 0.288 + ], + "angle": 0, + "content": "[36] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dollár, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos. arXiv, 2408.00714, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.29, + 0.908, + 0.344 + ], + "angle": 0, + "content": "[37] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: modeling and capturing hands and bodies together. ACM Transactions on Graphics, 36(6):245:1-245:17, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.346, + 0.908, + 0.389 + ], + "angle": 0, + "content": "[38] Javier Romero, Dimitrios Tzionas, and Michael J Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Transactions on Graphics (ToG), 2017. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.908, + 0.445 + ], + "angle": 0, + "content": "[39] Rómer Rosales and Stan Sclaroff. Combining generative and discriminative models in a framework for articulated pose estimation. International Journal of Computer Vision, 67: 251-276, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.446, + 0.908, + 0.503 + ], + "angle": 0, + "content": "[40] Dandan Shan, Jiaqi Geng, Michelle Shu, and David F Fouhey. Understanding human hands in contact at internet scale. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.908, + 0.572 + ], + "angle": 0, + "content": "[41] Rajinder S Sodhi, Brett R Jones, David Forsyth, Brian P Bailey, and Giuliano Maciocci. Bethere: 3d mobile collaboration with spatial input. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, pages 179-188, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.575, + 0.908, + 0.616 + ], + "angle": 0, + "content": "[42] Mohan Kumar Srirama, Sudeep Dasari, Shikhhar Bahl, and Abhinav Gupta. HRP: human affordances for robotic pretraining. arxiv, 2407.18911, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.617, + 0.908, + 0.673 + ], + "angle": 0, + "content": "[43] Omid Taheri, Nima Ghorbani, Michael J Black, and Dimitrios Tzionas. GRAB: A dataset of whole-body human grasping of objects. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.675, + 0.908, + 0.743 + ], + "angle": 0, + "content": "[44] Omid Taheri, Yi Zhou, Dimitrios Tzionas, Yang Zhou, Duygu Ceylan, Soren Pirk, and Michael J. Black. GRIP: Generating interaction poses using latent consistency and spatial cues. In Proceedings of the International Conference on 3D Vision (3DV), 2024. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.745, + 0.908, + 0.801 + ], + "angle": 0, + "content": "[45] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit Haim Bermano. Human motion diffusion model. In Proceedings of the International Conference on Learning Representations (ICLR), 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.908, + 0.872 + ], + "angle": 0, + "content": "[46] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In Proceedings of the International Conference on Machine Learning (ICML), 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.908, + 0.901 + ], + "angle": 0, + "content": "[47] Shashank Tripathi, Agniv Chatterjee, Jean-Claude Passy, Hongwei Yi, Dimitrios Tzionas, and Michael J. Black. DECO:" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7035" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.134 + ], + "angle": 0, + "content": "dense estimation of 3d human-scene contact in the wild. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.485, + 0.177 + ], + "angle": 0, + "content": "[48] Aäron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu. Neural discrete representation learning. In Advances in Neural Information Processing Systems (NeurIPS), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.484, + 0.246 + ], + "angle": 0, + "content": "[49] Sirui Xu, Zhengyuan Li, Yu-Xiong Wang, and Liang-Yan Gui. Interdiff: Generating 3d human-object interactions with physics-informed diffusion. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.483, + 0.316 + ], + "angle": 0, + "content": "[50] Yuhang Yang, Wei Zhai, Hongchen Luo, Yang Cao, Jiebo Luo, and Zheng-Jun Zha. Grounding 3d object affordance from 2d interactions in images. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.319, + 0.484, + 0.374 + ], + "angle": 0, + "content": "[51] Yufei Ye, Abhinav Gupta, and Shubham Tulsiani. What's in your hands? 3D reconstruction of generic objects in hands. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.376, + 0.484, + 0.431 + ], + "angle": 0, + "content": "[52] Yufei Ye, Poorvi Hebbar, Abhinav Gupta, and Shubham Tulsiani. Diffusion-guided reconstruction of everyday handobject interaction clips. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.433, + 0.484, + 0.5 + ], + "angle": 0, + "content": "[53] Yufei Ye, Xueting Li, Abhinav Gupta, Shalini De Mello, Stan Birchfield, Jiaming Song, Shubham Tulsiani, and Sifei Liu. Affordance diffusion: Synthesizing hand-object interactions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.503, + 0.484, + 0.572 + ], + "angle": 0, + "content": "[54] Yufei Ye, Abhinav Gupta, Kris Kitani, and Shubham Tulsiani. G-hop: Generative hand-object prior for interaction reconstruction and grasp synthesis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.574, + 0.484, + 0.629 + ], + "angle": 0, + "content": "[55] Hui Zhang, Sammy Joe Christen, Zicong Fan, Luocheng Zheng, Jemin Hwangbo, Jie Song, and Otmar Hilliges. Artigrasp: Physically plausible synthesis of bi-manual dexterous grasping and articulation. arXiv, 2309.03891, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.485, + 0.629 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "7036" + } + ] +] \ No newline at end of file diff --git a/2025/How Do I Do That_ Synthesizing 3D Hand Motion and Contacts for Everyday Interactions/4f85bf33-0c00-4386-bb62-720463efe772_origin.pdf b/2025/How Do I Do That_ Synthesizing 3D Hand Motion and Contacts for Everyday Interactions/4f85bf33-0c00-4386-bb62-720463efe772_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9d82c1e505f80e127a85a622e31edbd6632ba56b --- /dev/null +++ b/2025/How Do I Do That_ Synthesizing 3D Hand Motion and Contacts for Everyday Interactions/4f85bf33-0c00-4386-bb62-720463efe772_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6b88598b5b3782ed5ced6c313207f04a3dbd900fcda3b1703427545232a73e66 +size 4300986 diff --git a/2025/How Do I Do That_ Synthesizing 3D Hand Motion and Contacts for Everyday Interactions/full.md b/2025/How Do I Do That_ Synthesizing 3D Hand Motion and Contacts for Everyday Interactions/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3128c41f558b0d4a065a272b521af162b4b6c7ae --- /dev/null +++ b/2025/How Do I Do That_ Synthesizing 3D Hand Motion and Contacts for Everyday Interactions/full.md @@ -0,0 +1,256 @@ +# How Do I Do That? Synthesizing 3D Hand Motion and Contacts for Everyday Interactions + +Aditya Prakash\*1 Benjamin Lundell\*2 Dmitry Andreychuk\* David Forsyth\*1 Saurabh Gupta\*1 Harpreet Sawhney\*2 University of Illinois Urbana-Champaign \*2Microsoft https://bit.ly/LatentAct + +![](images/4fd603c9ecbd617d8a6fbbd53cb2d9e00cc626c4340a90083783b61e102ff2d8.jpg) +Figure 1. Interaction Trajectories. We tackle the novel task of predicting future 3D hand poses & contact maps, i.e., interaction trajectories, from a single image showing the object, text describing the action and a 3D contact point on the object, in everyday activities. We show the trajectory predicted by our method (LatentAct) & ground truth (GT) for 3 future timesteps along with the contact point. We consider 2 settings (top) Forecasting: single RGB view, action text & 3D contact points as input, (bottom) Interpolation: goal image is also provided. + +![](images/2a87569ab56d8d97375a246d6ccd9a6594d3766d4e9fdc1260255f37ad1a2d8b.jpg) + +# Abstract + +We tackle the novel problem of predicting 3D hand motion and contact maps (or Interaction Trajectories) given a single RGB view, action text, and a 3D contact point on the object as input. Our approach consists of (1) Interaction Codebook: a VQVAE model to learn a latent codebook of hand poses and contact points, effectively tokenizing interaction trajectories, (2) Interaction Predictor: a transformer-decoder module to predict the interaction trajectory from test time inputs by using an indexer module to retrieve a latent affordance from the learned codebook. To train our model, + +we develop a data engine that extracts 3D hand poses and contact trajectories from the diverse HoloAssist dataset. We evaluate our model on a benchmark that is $2.5 - 10 \times$ larger than existing works, in terms of diversity of objects and interactions observed, and test for generalization of the model across object categories, action categories, tasks, and scenes. Experimental results show the effectiveness of our approach over transformer & diffusion baselines across all settings. + +# 1. Introduction + +Hands interact with objects in diverse ways, often requiring varying skill levels to complete different tasks. Consider teaching someone to repair a bike. It is challenging to specify + +what part to manipulate [41], what axis to flip a part over such that it inserts well, etc. Instead, it is easier to directly show how the hand should move and what parts of it should make contact with the object over time to complete the task. In this work, we study this problem. Given a single image of an object, a 3D contact point, and a text indicating the high-level action, we predict Interaction Trajectories, the sequence of 3D hand poses and contact maps that specify the initial and evolving hand interactions (Fig. 1), interpretable by a human. We target interaction with everyday objects that are often small, thin, transparent, occluded, and deformable (Fig. 1) for which 3D object models may not be available. Our work is a first step towards learning a motion prior for diverse interactions in everyday activities, that can be combined with 3D object models in the future. + +In the absence of a 3D model, explicitly specifying contact points in the scene grounds the hand-object interaction, especially when there are multiple objects & 3D models are unavailable. This information is provided in the form of a 3D contact point. Moreover, we observe that while objects are diverse, hands interact with them in stylized ways [39]. There are only a few prototypical ways in which hands act in everyday interactions. To exploit this property, we learn a latent codebook of trajectories of 3D hand motions & contact points (Interaction Codebook, Sec. 3.1), effectively tokenizing interaction trajectories using a VQVAE [48]. Intuitively, the codebook learns the prototypical motion involved in (say) screwing in a screw that can be repurposed into motion involved in (say) closing a bottle cap. The codebook is indexed using visual inputs at test time to retrieve the closest codebook entry (Indexer module, Sec. 3.2), and aligned with the scene using the 3D contact point & visual context from the image input (transformer-decoder based Interaction Predictor, Sec. 3.3). We refer to our approach as LatentAct. + +Since existing HOI datasets with 3D annotations have constrained setups, we use egocentric videos from the HoloAssist dataset, showing everyday interactions. We leverage recent advances in 3D hand pose estimation [30] and segmentation in videos [36] to build a semi-automatic data engine (Sec. 3.4) to extract trajectories of 3D hand poses and contact maps from HoloAssist for 800 tasks performed by ordinary users across 120 object categories and 24 action categories. Our setup is $2.5\mathrm{X} - 10\mathrm{X}$ larger than existing works, e.g. HOT3D [2] (33 rigid objects), ARCTIC [11] (10 articulated objects), GRAB [43] (51 objects), HOI4D [28] (16 object categories), in terms of diversity of objects and interactions observed in everyday activities. + +For evaluation, we focus on generalization to 4 aspects: novel object categories, action categories, tasks and scenes. We show results on 2 settings: (1) Forecasting: the model generated predictions conditioned on the textual description of the action, current image showing the object of interaction and the 3D contact point at the current timestep. (2) Inter- + +polation: Here, we also provide the goal image, along with the previous inputs, showing the final state of interaction. For each of these settings, we consider 2 variants: (a) hand visible in the image, (b) hand is absent in the image (often the case at start of interactions). Across this large-scale experiment setting, we find our approach to generalize better than transformer & diffusion baselines that directly learn to predict interaction trajectories from images (Sec. 4.2). + +# 2. Related Work + +Generating HOI sequences: Recent works in HOI [3, 7, 13, 26, 44, 54, 55] have studied generating hand-object motions in different settings, e.g. conditioned on textual descriptions [7, 13], body poses [3, 13, 26, 44] for both rigid [3, 7, 13, 26] and articulated objects [4, 55]. While they focus on pick & place tasks involving simple & constrained interactions, we consider everyday activities, often involving small, thin, textureless, occluded, articulated, deformable and dynamic objects in contact for which obtaining 3D models is non-trivial. Existing works in this domain use either object-centric [4, 7] or hand-centric [44, 54] frame to represent trajectories, this becomes a limitation in our setting since 3D objects are unknown and hands may not be present in the image (often the case at the start of the interaction). Instead, we focus on predicting future 3D hand poses & contact maps from a single image showing the object, text describing the action & a 3D contact point to localize the interaction, for diverse interactions in everyday activities. + +HOI datasets: Existing works use datasets [11, 17, 18, 20, 24, 28] collected in controlled settings which provide 3D ground truth using MoCap [11, 43] or multi-camera setups [17, 18, 20, 24, 28, 29]. However, the diversity of objects and interactions in these datasets is limited due to constrained nature of the capture setup. While recent datasets [6, 9, 40] have explored natural interactions in egocentric videos [8, 15], they only provide 2D annotations, e.g. segmentation masks [6, 9], 2D bounding boxes [40] and grasp labels [6]. To mitigate this issue, we design a data engine to extract 3D hand poses and contact maps (Sec. 3.4) from egocentric videos showing everyday interactions. Specifically, we extract 3D trajectories from HoloAssist for 800 tasks performed by ordinary users across 120 object categories and 24 action categories, which is $2.5 \times 10 \times$ larger than existing datasets [2, 11, 28, 43]. + +HOI representation: Hands are often represented using a parametric mesh model, e.g., MANO [37], along with the global translation and rotation of the wrist [7, 11, 13, 30, 31, 33, 51, 53, 54]. Given an object mesh, its 6DoF pose can be denoted using translation & rotation [13, 55], signed distance field [54] or using Basis Point Set (BPS [34]) distances [7, 26]. For hand-object contact, existing approaches typically consider a canonicalized representation with the hand [4, 54] or object [7] at the origin & use relative dis + +![](images/603f40e6e920e5ab8c0ad6144b250ab3944488914530f194c78d79c380d2f865.jpg) +Figure 2. We represent contact points as binary masks on the hand mesh vertices. The hand mesh is represented in the camera coordinate system, consisting of the local hand pose in the MANO [38] coordinate frame and a global transformation from the MANO frame to the camera frame. These hand poses and contact points over several timesteps form interaction trajectories. + +stances [7, 11, 12, 32, 44] or a distance field [54] from hand joints to represent contact. Since 3D hand-object motion annotations are unavailable for diverse objects, e.g., thin, occluded, deformable, we instead focus on sequences of 3D hand poses and contact maps, i.e., interaction trajectories. + +Affordances: Gibson [14] developed the concept of affordances as a set of functionalities that the environment furnishes to an agent, which can be a human, a robot, an animal, or hands. Affordances are typically represented as the contact regions on the objects [1, 5, 23, 25, 27, 28, 43, 50] & the agent [1, 19, 23, 42, 43, 47, 53, 53, 54]. They have been explored in domains like robotics [1, 42], human interactions [19, 43, 47, 49] & hand grasp synthesis [23, 42, 43, 53, 54]. Recent works have also proposed generative models for affordances synthesis [49, 53, 54]. + +# 3. Method + +Given a single RGB view, action text, and a 3D contact point as input, we focus on predicting 3D hand poses & contact points, i.e., interaction trajectories (Fig. 1), for future timesteps. Our approach consists of: (1) Interaction Codebook: learns a latent codebook of affordances using a VQVAE (Fig. 3a), (2) Learned Indexer: maps the action text, image (showing an object) & contact point inputs to the codebook indices to obtain the corresponding embeddings (Fig. 3b), (3) Interaction Predictor: takes the queried embeddings along with text, image and contact point inputs to output the 3D interaction trajectory (Fig. 3b). + +Hand mesh: We use the MANO [37] model to represent the hand. It consists of pose $\theta$ and shape $\beta$ parameters, which can be converted to a 3D hand mesh. These mesh vertices lie in the MANO coordinate system, which can be transformed to the camera coordinate system at the first + +timestep (reference frame) using a rotation and translation. This gives us the hand trajectory in the reference frame. + +Contact map: We represent the contact points as a binary mask over the MANO hand mesh vertices, i.e., a 778 dimensional binary vector for each timestep, referred to as contact map in the paper (Fig. 2). The contact points may also change over time depending on the action, e.g. twisting. + +Once the 3D hand mesh is transformed into reference frame, we estimate the trajectory of the contact points by computing the centroid of the contact points at each timestep. The contact centroid in the reference frame is used as the 3D contact point input during training. We assume a 3D contact point to be available in the first frame at test time. To input this point, we create a 3D voxel grid of resolution $16 \times 16 \times 16$ . We take minimum and maximum values of 3D hand trajectory (along each axis, in the reference frame) across all training sequences. This span is divided into 16 equal parts per axis, associating each voxel to 3D metric range. It is kept fixed at train & test time. Thus, the center of each voxel represents a 3D location in metric space, i.e., a 3D metric grid, in the camera coordinate system. We create a 3D gaussian heatmap centered at 3D contact point in this grid, referred to as voxelized contact points in the paper. This changes as the trajectory evolves. This is different than the contact map (binary mask over mesh vertices). + +# 3.1. Interaction Codebook (InterCode) + +This is implemented as a VQVAE with a transformer encoder-decoder architecture [16]. It consists of 3 modules: + +Encoder: The hand and contact map trajectories for a fixed time horizon $T$ are passed to a transformer encoder to get the feature embedding. These features are used to query the codebook to obtain the relevant embeddings using L2 distance. Specifically, the input trajectory, consisting of MANO parameters & contact maps, is first passed to a 1-layer MLP to match the feature dimension of the transformer. We use a 1-layer transformer encoder with 1 head & feature dimension = 512. It outputs a $T \times 512$ dimensional feature. + +Latent codebook: It is initialized as a $K \times E$ dimensional matrix and is updated during training using exponential moving average of the embeddings output from the encoder, following recent works [16]. Concretely, we use a multilayer residual VQVAE architecture [16]. We set the number of quantizers for VQVAE to 6, $K = 512$ and $E = 512$ . When querying the codebook during training, sampling is done using the Gumbel-SoftMax trick [21] to preserve differentiability. We set the temperature of Gumbel-Softmax to 0.5 & inject noise at training for stochastic sampling. At test time, argmax is used to get the closest codebook entry. + +Decoder: It takes multiple inputs: (1) the latent embeddings queried from the codebook $(T \times 512)$ , (2) CLIP [35] embeddings of the text describing the action (512 dimensional feature extended for Timesteps), (3) DeiT [46] embeddings + +![](images/aecf9241ff992cc297975b0a79be9bf9f0f79944f5ad80ef75d575a7213ac360.jpg) +(a) Interaction Codebook is a VQVAE model that learns a latent codebook of 3D interaction trajectories, consisting of a transformer encoder-decoder architecture. The encoder features are used to sample codebook indices and corresponding embeddings which are passed to a decoder that also takes in the video input and text describing the action and reconstructs the 3D interaction trajectories. + +![](images/a489bae29a12395225aed3b557f007ab2e69ccf0305c4770b44cf9479adcbc53.jpg) +(b) We train an Indexer module to map the inputs at test time to the codebook indices to extract the relevant embeddings. These are then passed to a interaction predictor module that outputs the 3D interaction trajectories. We consider 2 settings: (1) Forecasting: text describing the action, single image & 3D contact point, (2) Interpolation: also providing the goal image showing the final state of the interaction (not shown here for clarity). +Figure 3. Overview. Our framework involves a 2-stage training procedure: (left) Interaction Codebook: to learn a latent codebook of hand poses and contact points, i.e., tokenizing interaction trajectories, (right) a learned Indexer & an Interaction Predictor module to predict the interaction trajectories from single image, action text & 3D contact point. We use pretrained features for images (from DeiT [46]) and text (from CLIP [35]). 3D contact point is input as a 3D gaussian heatmap in a 3D voxel grid (omitted here for clarity). + +of the video showing the actions $(T \times 768)$ , (4) voxelized contact points at each timestep, which are passed through four 3D convolution layers. We also pass in a 4D grid with each voxel containing the 3D location of its center in metric space, and process with 3D convolution layers. This gives us a 32 dimensional feature vector from the contact module. Concatenating all these input features, we get a 1824-dimensional joint feature that is passed through a 1-layer MLP to change the dimensionality to 512 and then fed to a transformer decoder (1-layer with 1 head & dropout of 0.2). The query embeddings for the transformer decoder are set as trainable parameters. It outputs a $T \times 512$ feature map which is passed to separate decoders for predicting the MANO parameters and the contact map. We use a 2-layer MLP to predict the MANO parameters and a 3-layer MLP to predict the contact maps for the entire time horizon $T$ . + +Training: The loss function consists of several terms: (1) Smooth L1 loss on the MANO parameters, (2) L1 loss on the contact centroid, (3) L1 loss on the translation between the future frame and first frame, (4) L1 loss on the 6D rotation representation between the future frame and first frame, (5) binary cross-entropy loss on the contact map predictions. + +# 3.2. Learned Indexer + +At test time, we need to query the learned codebook to get the latent interaction embeddings. This is done using a learned module which takes the test time inputs, i.e. text describing an action, image showing an object & a 3D contact point, and outputs a probability distribution over the codebook indices. The indices and corresponding embeddings can be sampled using the predicted probabilities. Note that the test time inputs only consist of a single image and a 3D contact point, unlike the entire trajectory used for training InterCode. + +Specifically, we first compute features from different inputs: CLIP embedding of text (512), DeiT embeddings of + +images (768), features from voxelized contact point, in the reference frame, after passing through 4 3D convolutional layers (32) and the 3D contact point. We concatenate these features and process through a 2-layer MLP (1024 & 512 nodes) to match the input dimensions of the transformer classifier module. It is implemented as a 1 layer, 1 head transformer decoder layer that takes in learnable query embeddings and aforementioned features from different modalities to output $T \times 512$ dimensional features. These features are then passed to a 3-layer MLP to output a probability distribution over the indices of the codebook. + +Training: It consists of a cross-entropy loss with ground truth computed as the closest codebook indices (L2 distance) from the encoder features of the InterCode module. + +# 3.3. Interaction Predictor (InterPred) + +Since the inputs at test time differ from the decoder inputs of the InterCode module during training, we need a separate learned module to predict the interaction trajectories. It consists of a transformer decoder module with the following inputs: (1) the latent embeddings queried from the codebook using the learned prior, (2) CLIP embeddings of the text describing the action, (3) DeiT embeddings of the image showing the object, and (4) voxelized contact point in the reference frame. These inputs are processed in the same way as for Learned Indexer before passing to the transformer decoder. The architecture is same as the decoder in InterCode. While the inputs are different than InterCode, they are passed to an MLP to match transformer dimension. + +Training: The loss function consists of several terms: (1) Smooth L1 loss on the MANO parameters, (2) L1 loss on the contact centroid, (3) L1 loss on the translation between the future frame and first frame, (4) L1 loss on the 6D rotation representation between the future frame and first frame, (5) binary cross-entropy loss on the contact map predictions. + +![](images/545cd36b3f1354f3f05c212dd7e401b3661fc3b8615379fc74767c4c5b169365.jpg) + +![](images/83221cfffeb490e8dee0df08fa3c171e73b2485445b4bc94f5743b95ad0e35a2.jpg) +Generated annotations for the task pull lever +Figure 4. Data engine. (top) Object masks are extracted using SAMv2 [36], 3D hand poses & masks (2D rendering of mesh) are from HaMeR [30], contact points are computed by projecting the 3D hand points into the 2D contact region (intersection of hand & object masks). (bottom) Generated object masks (highlighted in white), 3D hand mesh & contact points for 3 timesteps. + +# 3.4. Data Engine + +Existing lab datasets are limited to simple interactions in constrained setups. Instead, we use egocentric videos from HoloAssist showing everyday interactions with rigid, articulated & deformable objects, e.g. operating machines, tool use, kitchen tasks. It consists of $2\mathrm{k}+$ videos with temporally localized atomic actions, e.g. grab, open, screw, mix, rotate, align, slide, etc. with diverse objects. We extract these clips of atomic actions and convert them into the required format for training & evaluation. We consider 24 action categories with 120 object categories spanning 800 tasks. This is 2.5- $10\times$ larger than existing datasets, e.g. HOT3D [2] (33 rigid objects), ARCTIC [11] (10 articulated objects), GRAB [43] (51 objects), HOI4D [28] (16 object categories), in terms of objects & diversity of interactions. While HoloAssist contains diverse interactions, it does not provide 3D annotations for hand poses & contact points. To extract these annotations, we design a semi-automatic data pipeline (Fig. 4) using 2D segmentation masks, 3D hand poses & 2D contact region. + +Segmentation masks: We adopt a semi-automatic approach for object masks, that involves clicking a point on the target object in the first frame of a video & tracking the object mask across the video using SAMv2 [36]. For 2D hand masks, we render the 3D hand mesh (described next) into the image. + +3D hand poses: We use the off-the-shelf HaMeR [30] model to get 3D hand meshes in each video frame. HaMeR takes an image crop around the hand as input (estimated from Hands23 model [6]). The predicted 3D mesh in each frame is transformed to the reference frame using known intrinsics & extrinsics available in the HoloAssist dataset. + +3D Contact points: We first compute a 2D contact region in an image as the overlap between the hand & object masks. We add gaussian noise at the boundary to obtain a dense 2D contact region. We back project the 2D contact region into 3D hand mesh to obtain 3D contact points on the hand mesh. + +# 4. Experiments + +# 4.1. Protocol + +Task: We predict trajectories of MANO hand parameters & contact map (a binary mask over 778 vertices of the MANO [38] hand mesh) in 2 settings: (1) Forecasting: input consists of a textual description of the action, a current image showing the object of interaction, and a 3D contact point. (2) Interpolation: We provide the goal image, along with the previous inputs, showing the final state of interaction. For each setting, we 2 two variants: (a) current image with a hand, (b) current image without a hand (removed via in-painting [52]) since the hand may not be visible before the start of the interaction in many practical scenarios. We only consider right hand motion in this work. + +Generalization aspects: We study generalization capabilities for 4 different settings: (1) Object categories, (2) Action categories, (3) Novel tasks: a combination of object & action categories, (4) Novel scene: we hold out videos from 1 location (data is collected in 2 locations). For each setting, we create train, validation & test splits in the ratio 80:10:10. + +Metrics: (1) MPJPE: Mean Per Joint Position Error between predicted & ground truth hand joints, averaged over all future timesteps. (2) MPJPE-PA is the procrustus aligned variant, which optimizes for a single rotation, translation & scale to align the entire trajectory with ground truth trajectory, before computing MPJPE. (3) F1 score: harmonic mean of precision & recall to measure accuracy of contact maps. + +Baselines: Since no prior works predict interaction trajectories from image input, we adapt 2 recent baselines from human pose literature to work in our setting: (1) HCTFormer: The input image is encoded using a ViT [10] encoder followed by a transformer decoder, similar to the pose decoder used in T2P [22]. TCP decoder takes in trajectory query (for global intent), pose query & pose embeddings. In our case, intent comes from text. We replace pose embeddings with {image, contact, text} features, pass them to an MLP & then feed to {TRM, MLP} modules, as in TCP. (2) HCTDiff: We modify the motion diffusion model (MDM [45]) for hand trajectory and contact map prediction. For this, we change conditioning to use {image, text, contact} features but diffusion architecture is same. Both HCTFormer and HCTDiff are trained in the same setting as ours for fair comparisons. + +Existing methods that use object-centric representations [4, 7] can not be applied in our setting since we do not have the 3D object models. Hand-centric representations [44, 54] are also not suitable in our setting since the hand is not always visible in the input image at test time. + +Our approach involves a 2-stage training procedure: InterCode followed by Indexer & InterPred, from different inputs. We train with both transformer & diffusion variants for InterPred, referred to as LatentAct & LatentAct-Diff respectively. LatentAct-Diff uses same InterCode & Indexer as + +
MethodTask-levelObject-levelAction-levelScene-level
M-PE(cm)↓M-PA(cm)↓F1↑M-PE(cm)↓M-PA(cm)↓F1↑M-PE(cm)↓M-PA(cm)↓F1↑M-PE(cm)↓M-PA(cm)↓F1↑
Hand visibleForecasting
HCTFormer8.323.060.728.703.190.728.593.250.738.523.090.72
HCTDiff8.422.750.729.012.880.739.113.100.749.492.840.71
LatentAct (Ours)7.612.990.758.093.060.758.143.320.777.873.040.74
Interpolation
HCTFormer7.523.020.768.103.100.767.773.290.787.543.000.76
HCTDiff8.302.720.778.522.910.788.383.160.788.602.810.77
LatentAct (Ours)6.722.870.807.483.020.787.253.190.807.272.930.78
No handsForecasting
HCTFormer8.342.950.728.853.020.738.543.210.748.533.050.72
HCTDiff8.852.650.729.742.810.729.503.110.749.212.740.71
LatentAct (Ours)7.932.970.768.263.010.768.193.280.777.903.000.75
Interpolation
HCTFormer7.302.950.768.423.150.768.013.300.777.292.990.77
HCTDiff8.272.760.768.682.910.769.333.100.7810.022.820.76
LatentAct (Ours)7.132.890.797.242.970.797.443.210.807.122.890.78
+ +LatentAct but different InterPred, i.e., an iterative denoising model from MDM instead of a single step prediction. The denoising is done in latent space of the codebook followed by a decoder to generate trajectories. It is trained using L2 loss on denoised embedding and loss in Sec. 3.3. + +# 4.2. Results + +Generalization results: First, we study if the model can generate trajectories from the text describing the action, single image & 3D contact point, i.e., Forecasting task. In the absence of a goal, this is quite challenging and inherently multimodal in nature. We compare with the GT trajectories for all 4 generalization aspects: novel objects, actions, tasks & scene. We report MPJPE, MPJPE-PA for hand trajectories & F1-score for contact maps. We observe that our approach leads to consistent gains in absolute hand poses & contact maps. We show results (Tab. 1) on 2 variants of the task, (1) hand visible in the image, (2) hand absent in the image, often the case at the start of interactions. The hand visible in the image makes the tasks slightly easier since the initial pose is visible, as also evident from the results. + +Next, we also provide the goal image, along with the previous inputs, showing the final state of interaction, referred to as goal-conditioned interpolation. This is slightly easier than the Forecasting task since the final state of the interaction is also provided. The results (Tab. 1) on 2 varaints of the task, (1) hand visible in the image, (2) hand absent in the image, show similar benefits as in Forecasting. + +These results are for $T = 30$ timesteps. We also report results for $T = 16$ in the supplementary with similar trends. + +Results on ARCTIC: We also show results (Tab. 2) on ARCTIC [11] to check if LatentAct leads to gains on existing + +Table 1. Generalization results. We report MPJPE (M-PE: cm), MPJPE-PA (M-PA: cm) & F1-score in 4 generalization settings: novel tasks, objects, actions & scene, to measure the accuracy of the predicted trajectories. We adapt two recent methods from human pose literature to work with image inputs, i.e., HCTFormer: ViT encoded image features passed to a transformer decoder (similar to pose decoder of T2P [22]), HCTDiff: MDM [45] modified to take image features as well. LatentAct leads to better absolute hand poses & contact maps. + +
MethodMPJPE↓MPJPE-PA↓F1↑
HCTFormer: trained on ARCTIC15.763.610.36
LatentAct: trained on ARCTIC14.773.760.41
LatentAct: zero-shot from Holo15.723.710.19
LatentAct: Codebook & Indexer from Holo15.363.580.45
+ +Table 2. HoloAssist is significantly larger than ARCTIC in terms of contact sequences leading to benefits in zero-shot generalization & transferring models trained on HoloAssist to ARCTIC. + +lab datasets. ARCTIC has accurate 3D labels but is limited in scale ( $10 \times$ smaller than HoloAssist for contact sequences). When trained only on ARCTIC, we find the trends to transfer from HoloAssist, i.e., LatentAct is better than HCTFormer. Using codebook & indexer trained on HoloAssist further benefits models trained on ARCTIC. Zero-shot transfer from HoloAssist is competitive to models trained on ARCTIC. + +Benefits of Interaction Codebook: We verify if the 2-stage training procedure is beneficial by comparing with single-stage training methods HCTFormer & HCTDiff in Tab. 4 and observe consistent benefits. Another option is to retrieve an interaction trajectory from the training set, conditioned on the input. However, this does not scale well with the training data & is computationally expensive, e.g. the learned codebook in InterCode has 512 entries whereas the training dataset has $\sim 15\mathrm{K}$ sequences, which is significantly larger. + +Contact maps helps hand predictions: Here, we train the InterCode & InterPred modules with only the hand pose loss, i.e., removing the loss on contact maps. In Tab. 5, we observe that loss on contact maps helps with hand predictions. + +Trends with LatentAct-Diff.: Training InterPred with diffusion losses leads to better hand poses but worse contact maps compared to LatentAct, across all settings (Tab. 3). + +
MethodTask-levelObject-levelAction-levelScene-level
M-PE(cm)↓M-PA(cm)↓F1↑M-PE(cm)↓M-PA(cm)↓F1↑M-PE(cm)↓M-PA(cm)↓F1↑M-PE(cm)↓M-PA(cm)↓F1↑
Hand visibleForecasting
LatentAct7.612.990.758.093.060.758.143.320.777.873.040.74
LatentAct-Diff7.903.030.718.493.200.728.293.340.728.323.100.72
Interpolation
LatentAct6.722.870.807.483.020.787.253.190.807.272.930.78
LatentAct-Diff6.532.620.787.232.810.787.113.100.796.702.840.78
No handsForecasting
LatentAct7.932.970.768.263.010.768.193.280.777.903.000.75
LatentAct-Diff7.822.890.718.633.060.718.193.150.727.932.960.73
Interpolation
LatentAct7.132.890.797.242.970.797.443.210.807.122.890.78
LatentAct-Diff6.602.650.767.732.780.767.423.040.786.892.710.77
+ +Table 3. LatentAct-Diff trends. Training InterPred with diffusion loss leads to better hand poses but worse contact maps than LatentAct. + +
MethodMPJPE↓MPJPE-PA↓F1↑
HCTFormer7.523.020.76
HCTFormer + InterCode6.722.870.80
HCTDiff8.302.720.77
HCTDiff + InterCode6.532.620.78
+ +![](images/1a55bf12358673256d2af4439faa8ca1c6ea5c68fbb8f295467f405f0dccfb73.jpg) +Figure 5. More training data helps both LatentAct & HCTFormer. + +Dataset scale: In Fig. 5, we see that both LatentAct & HCTFormer improve with larger training dataset size. + +# 4.3. Visualizations + +We compare the predictions of LatentAct & the best baseline on a few examples from task level generalization setting (Fig. 6) for both forecasting & interpolation for 3 timesteps $(t = 5, 15, 25)$ . The left column shows the input image with the contact point (projected in the image in cyan blob), goal image (for interpolation task) & text describing the action. The other columns show: (a) Camera View: predicted hand in the camera view with the contact point, this captures the placement of the hand around the contact point, (b) Another View: this better visualizes the hand articulation from a different camera viewpoint, (c) Contact Maps are shown as red parts of the hand mesh. LatentAct leads to better orientation of hand & sharper contact maps than the baseline. Note that these tasks are not seen during training. More visualizations & analysis are provided in the supplementary. Limitations: We do not predict the object state change. While 3D object models are not available, alternatives like + +Table 4. Two stage training with InterCode improves over single-stage training methods HCTFormer & HCTDiff. Note that HCT-Former + InterCode is same as LatentAct. + +
MethodMPJPE↓MPJPE-PA↓F1↑
HCTFormer7.523.020.76
- no contact map7.652.89-
LatentAct6.722.870.80
- no contact map7.762.91-
HCTDiff8.302.720.77
- no contact map8.432.78-
LatentAct-Diff6.532.620.78
- no contact map7.092.90-
+ +Table 5. Loss on contact maps helps hand predictions for all models. + +predicting 2D object masks could be explored. While we assume a 3D contact point to be available, it can also be estimated using off-the-shelf depth models & ground truth intrinsics to get the 3D location of a 2D point that can be prompted by a user (e.g. by clicking a point on the image). Our experiments measure whether the predicted interaction trajectory is accurate, but it does not account for the multimodal nature of future prediction. The annotations in our work are programatically generated and may not be accurate in all cases. While we filter out the erroneous cases, using a better annotation tool [47] would be beneficial. + +# 5. Conclusion + +Our model predicts interaction trajectories, i.e. 3D hand motion & contact maps, from single RGB view, action text & 3D contact point as input. It consists of: (1) Interaction Codebook: a VQVAE to learn a latent codebook of 3D hand poses & contact points, i.e., tokenizing interactions. (2) Interaction Predictor: a transformer-decoder module to predict the interaction trajectory from test time inputs, by using an indexer module to retrieve a latent interaction from the learned codebook. Our large-scale experiments on the diverse HoloAssist dataset show benefits across 4 generalization settings. Our work is a first step towards learning a motion prior on interaction sequences and can be augmented + +
Input Image (t=0)Camera ViewAnother ViewContact Map
mix/stir coffeeBaselinet=5t=15t=25t=5t=15t=25t=5t=15t=25
GT.........
.........
rotate lensBaseline.........
GT.........
.........
pull batteryBaseline.........
GT.........
.........
slide controllerBaseline.........
GT.........
+ +Figure 6. Visualizations. We compare the predictions of LatentAct & the best baseline on a few examples from the task level generalization setting for both forecasting & interpolation for 3 timesteps (t=5, 15, 25). The left column shows the input image with the contact point (projected in the image in cyan blob), goal image (for interpolation task) & text describing the action. The other columns show: (a) Camera View: predicted hand in the camera view with the contact point, this captures the placement of the hand around the contact point, (b) Another View: this better visualizes the hand pose from a different camera viewpoint, (c) Contact Maps are shown as red parts of the hand mesh. LatentAct leads to better orientation of hand & sharper contact maps than the baseline. Note that these tasks are not seen during training. + +with object trajectories with better 3D models in the future. + +Our data pipeline is modular and can also incorporate better + +hand & segmentation models in future. + +Acknowledgments: This material is based upon work supported by an NSF CAREER Award (IIS2143873), USDA-NIFA AIFARMS National AI Institute (USDA#2020-67021-32799) and NSF (IIS-2007035). Part of the work was done when Aditya Prakash was an intern at Microsoft. We acknowledge compute support by NVIDIA Academic Hardware Grant and a DURIP grant. We thank Shaowei Liu and Arjun Gupta for feedback on the draft. + +# References + +[1] Shikhar Bahl, Russell Mendonca, Lili Chen, Unnat Jain, and Deepak Pathak. Affordances from human videos as a versatile representation for robotics. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3 +[2] Prithviraj Banerjee, Sindi Shkodrani, Pierre Moulon, Shreyas Hampali, Fan Zhang, Jade Fountain, Edward Miller, Selen Basol, Richard Newcombe, Robert Wang, Jakob Julian Engel, and Tomas Hodan. Introducing hot3d: An egocentric dataset for 3d hand and object tracking. arXiv, 2406.09598, 2024. 2, 5 +[3] Jona Braun, Sammy Joe Christen, Muhammed Kocabas, Emre Aksan, and Otmar Hilliges. Physically plausible full-body hand-object interaction synthesis. arXiv, 2309.07907, 2023. 2 +[4] Junuk Cha, Jihyeon Kim, Jae Shin Yoon, and Seungryul Baek. Text2hoi: Text-guided 3d motion generation for handobject interaction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 5 +[5] Matthew Chang, Aditya Prakash, and Saurabh Gupta. Look ma, no hands! agent-environment factorization of egocentric videos. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 3 +[6] Tianyi Cheng, Dandan Shan, Ayda Sultan Hassen, Richard Ely Locke Higgins, and David Fouhey. Towards a richer 2d understanding of hands at scale. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 2, 5 +[7] Sammy Christen, Shreyas Hampali, Fadime Sener, Edoardo Remelli, Tomas Hodan, Eric Sauser, Shugao Ma, and Bugra Tekin. Diffh2o: Diffusion-based synthesis of hand-object interactions from textual descriptions. arXiv, 2403.17827, 2024. 2, 3, 5 +[8] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Sanja Fidler, Antonino Furnari, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, et al. Scaling egocentric vision: The epic-kitchens dataset. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 2 +[9] Ahmad Darkhalil, Dandan Shan, Bin Zhu, Jian Ma, Amlan Kar, Richard Higgins, Sanja Fidler, David Fouhey, and Dima Damen. Epic-kitchens visor benchmark: Video segmentations and object relations. In NeurIPS Track on Datasets and Benchmarks, 2022. 2 +[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image + +is worth 16x16 words: Transformers for image recognition at scale. In Proceedings of the International Conference on Learning Representations (ICLR), 2021. 5 +[11] Zicong Fan, Omid Taheri, Dimitrios Tzionas, Muhammed Kocabas, Manuel Kaufmann, Michael J. Black, and Otmar Hilliges. ARCTIC: A dataset for dexterous bimanual handobject manipulation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3, 5, 6 +[12] Zicong Fan, Takehiko Ohkawa, Linlin Yang, Nie Lin, Zhishan Zhou, Shihao Zhou, Jiajun Liang, Zhong Gao, Xuanyang Zhang, Xue Zhang, Fei Li, Zheng Liu, Feng Lu, Karim Abou Zeid, Bastian Leibe, Jeongwan On, Seungryul Baek, Aditya Prakash, Saurabh Gupta, Kun He, Yoichi Sato, Otmar Hilliges, Hyung Jin Chang, and Angela Yao. Benchmarks and challenges in pose estimation for egocentric hand interactions with objects. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 3 +[13] Anindita Ghosh, Rishabh Dabral, Vladislav Golyanik, Christian Theobalt, and Philipp Slusallek. Imos: Intent-driven full-body motion synthesis for human-object interactions. EuroGraphics, 42(2):1-12, 2023. 2 +[14] J.J. Gibson. The ecological approach to visual perception, 1979. 3 +[15] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[16] Chuan Guo, Yuxuan Mu, Muhammad Gohar Javed, Sen Wang, and Li Cheng. Momask: Generative masked modeling of 3d human motions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3 +[17] Shreyas Hampali, Mahdi Rad, Markus Oberweger, and Vincent Lepetit. Honnotate: A method for 3d annotation of hand and object poses. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2 +[18] Shreyas Hampali, Sayan Deb Sarkar, Mahdi Rad, and Vincent Lepetit. Keypoint transformer: Solving joint identification in challenging hands and object interactions for accurate 3d pose estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[19] Mohamed Hassan, Partha Ghosh, Joachim Tesch, Dimitrios Tzionas, and Michael J. Black. Populating 3d scenes by learning human-scene interaction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3 +[20] Di Huang, Xiaopeng Ji, Xingyi He, Jiaming Sun, Tong He, Qing Shuai, Wanli Ouyang, and Xiaowei Zhou. Reconstructing hand-held objects from monocular video. In ACM Transactions on Graphics, 2022. 2 +[21] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. In Proceedings of the International Conference on Learning Representations (ICLR), 2017. 3 +[22] Jaewoo Jeong, Daehee Park, and Kuk-Jin Yoon. Multi-agent long-term 3d human pose forecasting via interaction-aware + +trajectory conditioning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 5, 6 +[23] Hanwen Jiang, Shaowei Liu, Jiashun Wang, and Xiaolong Wang. Hand-object contact consistency reasoning for human grasps generation. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. 3 +[24] Taein Kwon, Bugra Tekin, Jan Stühmer, Federica Bogo, and Marc Pollefeys. H2o: Two hands manipulating objects for first person interaction recognition. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. 2 +[25] Gen Li, Varun Jampani, Deqing Sun, and Laura Sevilla-Lara. LOCATE: localize and transfer object parts for weakly supervised affordance grounding. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3 +[26] Quanzhou Li, Jingbo Wang, Chen Change Loy, and Bo Dai. Task-oriented human-object interactions generation with implicit neural representations. In Proceedings of the IEEE Winter Conference on Applications of Computer Vision (WACV), 2024. 2 +[27] Shaowei Liu, Yang Zhou, Jimei Yang, Saurabh Gupta, and Shenlong Wang. Contactgen: Generative contact modeling for grasp generation. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2023. 3 +[28] Yunze Liu, Yun Liu, Che Jiang, Kangbo Lyu, Weikang Wan, Hao Shen, Boqiang Liang, Zhoujie Fu, He Wang, and Li Yi. HOI4D: A 4d egocentric dataset for category-level human-object interaction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 5 +[29] Takehiko Ohkawa, Kun He, Fadime Sener, Tomas Hodan, Luan Tran, and Cem Keskin. Assemblyhands: Towards egocentric activity understanding via 3d hand pose estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 12999-13008, 2023. 2 +[30] Georgios Pavlakos, Dandan Shan, Iija Radosavovic, Angjoo Kanazawa, David Fouhey, and Jitendra Malik. Reconstructing hands in 3d with transformers. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 5 +[31] Aditya Prakash, Matthew Chang, Matthew Jin, and Saurabh Gupta. 3d reconstruction of objects in hands without real world 3d supervision. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 2 +[32] Aditya Prakash, Arjun Gupta, and Saurabh Gupta. Mitigating perspective distortion-induced shape ambiguity in image crops. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 3 +[33] Aditya Prakash, Ruisen Tu, Matthew Chang, and Saurabh Gupta. 3d hand pose estimation in everyday egocentric images. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 2 +[34] Sergey Prokudin, Christoph Lassner, and Javier Romero. Efficient learning on point clouds with basis point sets. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2019. 2 + +[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the International Conference on Machine Learning (ICML), 2021. 3, 4 +[36] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dollár, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos. arXiv, 2408.00714, 2024. 2, 5 +[37] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: modeling and capturing hands and bodies together. ACM Transactions on Graphics, 36(6):245:1-245:17, 2017. 2, 3 +[38] Javier Romero, Dimitrios Tzionas, and Michael J Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Transactions on Graphics (ToG), 2017. 3, 5 +[39] Rómer Rosales and Stan Sclaroff. Combining generative and discriminative models in a framework for articulated pose estimation. International Journal of Computer Vision, 67: 251-276, 2006. 2 +[40] Dandan Shan, Jiaqi Geng, Michelle Shu, and David F Fouhey. Understanding human hands in contact at internet scale. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2 +[41] Rajinder S Sodhi, Brett R Jones, David Forsyth, Brian P Bailey, and Giuliano Maciocci. Bethere: 3d mobile collaboration with spatial input. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, pages 179-188, 2013. 2 +[42] Mohan Kumar Srirama, Sudeep Dasari, Shikhhar Bahl, and Abhinav Gupta. HRP: human affordances for robotic pretraining. arxiv, 2407.18911, 2024. 3 +[43] Omid Taheri, Nima Ghorbani, Michael J Black, and Dimitrios Tzionas. GRAB: A dataset of whole-body human grasping of objects. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2, 3, 5 +[44] Omid Taheri, Yi Zhou, Dimitrios Tzionas, Yang Zhou, Duygu Ceylan, Soren Pirk, and Michael J. Black. GRIP: Generating interaction poses using latent consistency and spatial cues. In Proceedings of the International Conference on 3D Vision (3DV), 2024. 2, 3, 5 +[45] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit Haim Bermano. Human motion diffusion model. In Proceedings of the International Conference on Learning Representations (ICLR), 2023. 5, 6 +[46] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In Proceedings of the International Conference on Machine Learning (ICML), 2021. 3, 4 +[47] Shashank Tripathi, Agniv Chatterjee, Jean-Claude Passy, Hongwei Yi, Dimitrios Tzionas, and Michael J. Black. DECO: + +dense estimation of 3d human-scene contact in the wild. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2023. 3, 7 +[48] Aäron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu. Neural discrete representation learning. In Advances in Neural Information Processing Systems (NeurIPS), 2017. 2 +[49] Sirui Xu, Zhengyuan Li, Yu-Xiong Wang, and Liang-Yan Gui. Interdiff: Generating 3d human-object interactions with physics-informed diffusion. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2023. 3 +[50] Yuhang Yang, Wei Zhai, Hongchen Luo, Yang Cao, Jiebo Luo, and Zheng-Jun Zha. Grounding 3d object affordance from 2d interactions in images. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2023. 3 +[51] Yufei Ye, Abhinav Gupta, and Shubham Tulsiani. What's in your hands? 3D reconstruction of generic objects in hands. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[52] Yufei Ye, Poorvi Hebbar, Abhinav Gupta, and Shubham Tulsiani. Diffusion-guided reconstruction of everyday handobject interaction clips. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2023. 5 +[53] Yufei Ye, Xueting Li, Abhinav Gupta, Shalini De Mello, Stan Birchfield, Jiaming Song, Shubham Tulsiani, and Sifei Liu. Affordance diffusion: Synthesizing hand-object interactions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3 +[54] Yufei Ye, Abhinav Gupta, Kris Kitani, and Shubham Tulsiani. G-hop: Generative hand-object prior for interaction reconstruction and grasp synthesis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 3, 5 +[55] Hui Zhang, Sammy Joe Christen, Zicong Fan, Luocheng Zheng, Jemin Hwangbo, Jie Song, and Otmar Hilliges. Artigrasp: Physically plausible synthesis of bi-manual dexterous grasping and articulation. arXiv, 2309.03891, 2023. 2 \ No newline at end of file diff --git a/2025/How Do I Do That_ Synthesizing 3D Hand Motion and Contacts for Everyday Interactions/images.zip b/2025/How Do I Do That_ Synthesizing 3D Hand Motion and Contacts for Everyday Interactions/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..281e671e65420c6eef7ad52935dcbc164fcd6b15 --- /dev/null +++ b/2025/How Do I Do That_ Synthesizing 3D Hand Motion and Contacts for Everyday Interactions/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3a14450e72b651e47b47b11f4668234d9f055a6aeb1ceccce4b6419b96ab35e0 +size 641237 diff --git a/2025/How Do I Do That_ Synthesizing 3D Hand Motion and Contacts for Everyday Interactions/layout.json b/2025/How Do I Do That_ Synthesizing 3D Hand Motion and Contacts for Everyday Interactions/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0042d008f3a3f9e752fcd39a9083c56e07d6c5a9 --- /dev/null +++ b/2025/How Do I Do That_ Synthesizing 3D Hand Motion and Contacts for Everyday Interactions/layout.json @@ -0,0 +1,6573 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 88, + 103, + 523, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 103, + 523, + 140 + ], + "spans": [ + { + "bbox": [ + 88, + 103, + 523, + 140 + ], + "type": "text", + "content": "How Do I Do That? Synthesizing 3D Hand Motion and Contacts for Everyday Interactions" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 89, + 160, + 520, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 160, + 520, + 217 + ], + "spans": [ + { + "bbox": [ + 89, + 160, + 520, + 217 + ], + "type": "text", + "content": "Aditya Prakash\\*1 Benjamin Lundell\\*2 Dmitry Andreychuk\\* David Forsyth\\*1 Saurabh Gupta\\*1 Harpreet Sawhney\\*2 University of Illinois Urbana-Champaign \\*2Microsoft https://bit.ly/LatentAct" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 57, + 230, + 301, + 467 + ], + "blocks": [ + { + "bbox": [ + 57, + 230, + 301, + 467 + ], + "lines": [ + { + "bbox": [ + 57, + 230, + 301, + 467 + ], + "spans": [ + { + "bbox": [ + 57, + 230, + 301, + 467 + ], + "type": "image", + "image_path": "4fd603c9ecbd617d8a6fbbd53cb2d9e00cc626c4340a90083783b61e102ff2d8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 474, + 555, + 520 + ], + "lines": [ + { + "bbox": [ + 54, + 474, + 555, + 520 + ], + "spans": [ + { + "bbox": [ + 54, + 474, + 555, + 520 + ], + "type": "text", + "content": "Figure 1. Interaction Trajectories. We tackle the novel task of predicting future 3D hand poses & contact maps, i.e., interaction trajectories, from a single image showing the object, text describing the action and a 3D contact point on the object, in everyday activities. We show the trajectory predicted by our method (LatentAct) & ground truth (GT) for 3 future timesteps along with the contact point. We consider 2 settings (top) Forecasting: single RGB view, action text & 3D contact points as input, (bottom) Interpolation: goal image is also provided." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 312, + 232, + 547, + 468 + ], + "blocks": [ + { + "bbox": [ + 312, + 232, + 547, + 468 + ], + "lines": [ + { + "bbox": [ + 312, + 232, + 547, + 468 + ], + "spans": [ + { + "bbox": [ + 312, + 232, + 547, + 468 + ], + "type": "image", + "image_path": "2a87569ab56d8d97375a246d6ccd9a6594d3766d4e9fdc1260255f37ad1a2d8b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 152, + 534, + 200, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 534, + 200, + 547 + ], + "spans": [ + { + "bbox": [ + 152, + 534, + 200, + 547 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 560, + 297, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 560, + 297, + 681 + ], + "spans": [ + { + "bbox": [ + 54, + 560, + 297, + 681 + ], + "type": "text", + "content": "We tackle the novel problem of predicting 3D hand motion and contact maps (or Interaction Trajectories) given a single RGB view, action text, and a 3D contact point on the object as input. Our approach consists of (1) Interaction Codebook: a VQVAE model to learn a latent codebook of hand poses and contact points, effectively tokenizing interaction trajectories, (2) Interaction Predictor: a transformer-decoder module to predict the interaction trajectory from test time inputs by using an indexer module to retrieve a latent affordance from the learned codebook. To train our model," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 535, + 556, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 535, + 556, + 632 + ], + "spans": [ + { + "bbox": [ + 313, + 535, + 556, + 632 + ], + "type": "text", + "content": "we develop a data engine that extracts 3D hand poses and contact trajectories from the diverse HoloAssist dataset. We evaluate our model on a benchmark that is " + }, + { + "bbox": [ + 313, + 535, + 556, + 632 + ], + "type": "inline_equation", + "content": "2.5 - 10 \\times" + }, + { + "bbox": [ + 313, + 535, + 556, + 632 + ], + "type": "text", + "content": " larger than existing works, in terms of diversity of objects and interactions observed, and test for generalization of the model across object categories, action categories, tasks, and scenes. Experimental results show the effectiveness of our approach over transformer & diffusion baselines across all settings." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 657, + 395, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 657, + 395, + 670 + ], + "spans": [ + { + "bbox": [ + 314, + 657, + 395, + 670 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 678, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 678, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 678, + 555, + 715 + ], + "type": "text", + "content": "Hands interact with objects in diverse ways, often requiring varying skill levels to complete different tasks. Consider teaching someone to repair a bike. It is challenging to specify" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 264, + 704 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 264, + 704 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 264, + 704 + ], + "type": "text", + "content": "*Part of the work was done during an internship at Microsoft" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 67, + 704, + 219, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 704, + 219, + 714 + ], + "spans": [ + { + "bbox": [ + 67, + 704, + 219, + 714 + ], + "type": "text", + "content": "† Joint last authors, indicates equal contribution" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7026" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "content": "what part to manipulate [41], what axis to flip a part over such that it inserts well, etc. Instead, it is easier to directly show how the hand should move and what parts of it should make contact with the object over time to complete the task. In this work, we study this problem. Given a single image of an object, a 3D contact point, and a text indicating the high-level action, we predict Interaction Trajectories, the sequence of 3D hand poses and contact maps that specify the initial and evolving hand interactions (Fig. 1), interpretable by a human. We target interaction with everyday objects that are often small, thin, transparent, occluded, and deformable (Fig. 1) for which 3D object models may not be available. Our work is a first step towards learning a motion prior for diverse interactions in everyday activities, that can be combined with 3D object models in the future." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 255, + 297, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 255, + 297, + 481 + ], + "spans": [ + { + "bbox": [ + 56, + 255, + 297, + 481 + ], + "type": "text", + "content": "In the absence of a 3D model, explicitly specifying contact points in the scene grounds the hand-object interaction, especially when there are multiple objects & 3D models are unavailable. This information is provided in the form of a 3D contact point. Moreover, we observe that while objects are diverse, hands interact with them in stylized ways [39]. There are only a few prototypical ways in which hands act in everyday interactions. To exploit this property, we learn a latent codebook of trajectories of 3D hand motions & contact points (Interaction Codebook, Sec. 3.1), effectively tokenizing interaction trajectories using a VQVAE [48]. Intuitively, the codebook learns the prototypical motion involved in (say) screwing in a screw that can be repurposed into motion involved in (say) closing a bottle cap. The codebook is indexed using visual inputs at test time to retrieve the closest codebook entry (Indexer module, Sec. 3.2), and aligned with the scene using the 3D contact point & visual context from the image input (transformer-decoder based Interaction Predictor, Sec. 3.3). We refer to our approach as LatentAct." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 483, + 296, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 296, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 296, + 639 + ], + "type": "text", + "content": "Since existing HOI datasets with 3D annotations have constrained setups, we use egocentric videos from the HoloAssist dataset, showing everyday interactions. We leverage recent advances in 3D hand pose estimation [30] and segmentation in videos [36] to build a semi-automatic data engine (Sec. 3.4) to extract trajectories of 3D hand poses and contact maps from HoloAssist for 800 tasks performed by ordinary users across 120 object categories and 24 action categories. Our setup is " + }, + { + "bbox": [ + 55, + 483, + 296, + 639 + ], + "type": "inline_equation", + "content": "2.5\\mathrm{X} - 10\\mathrm{X}" + }, + { + "bbox": [ + 55, + 483, + 296, + 639 + ], + "type": "text", + "content": " larger than existing works, e.g. HOT3D [2] (33 rigid objects), ARCTIC [11] (10 articulated objects), GRAB [43] (51 objects), HOI4D [28] (16 object categories), in terms of diversity of objects and interactions observed in everyday activities." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "For evaluation, we focus on generalization to 4 aspects: novel object categories, action categories, tasks and scenes. We show results on 2 settings: (1) Forecasting: the model generated predictions conditioned on the textual description of the action, current image showing the object of interaction and the 3D contact point at the current timestep. (2) Inter-" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": "polation: Here, we also provide the goal image, along with the previous inputs, showing the final state of interaction. For each of these settings, we consider 2 variants: (a) hand visible in the image, (b) hand is absent in the image (often the case at start of interactions). Across this large-scale experiment setting, we find our approach to generalize better than transformer & diffusion baselines that directly learn to predict interaction trajectories from images (Sec. 4.2)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 178, + 400, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 178, + 400, + 190 + ], + "spans": [ + { + "bbox": [ + 314, + 178, + 400, + 190 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 198, + 556, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 198, + 556, + 414 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 556, + 414 + ], + "type": "text", + "content": "Generating HOI sequences: Recent works in HOI [3, 7, 13, 26, 44, 54, 55] have studied generating hand-object motions in different settings, e.g. conditioned on textual descriptions [7, 13], body poses [3, 13, 26, 44] for both rigid [3, 7, 13, 26] and articulated objects [4, 55]. While they focus on pick & place tasks involving simple & constrained interactions, we consider everyday activities, often involving small, thin, textureless, occluded, articulated, deformable and dynamic objects in contact for which obtaining 3D models is non-trivial. Existing works in this domain use either object-centric [4, 7] or hand-centric [44, 54] frame to represent trajectories, this becomes a limitation in our setting since 3D objects are unknown and hands may not be present in the image (often the case at the start of the interaction). Instead, we focus on predicting future 3D hand poses & contact maps from a single image showing the object, text describing the action & a 3D contact point to localize the interaction, for diverse interactions in everyday activities." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 415, + 556, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 556, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 556, + 605 + ], + "type": "text", + "content": "HOI datasets: Existing works use datasets [11, 17, 18, 20, 24, 28] collected in controlled settings which provide 3D ground truth using MoCap [11, 43] or multi-camera setups [17, 18, 20, 24, 28, 29]. However, the diversity of objects and interactions in these datasets is limited due to constrained nature of the capture setup. While recent datasets [6, 9, 40] have explored natural interactions in egocentric videos [8, 15], they only provide 2D annotations, e.g. segmentation masks [6, 9], 2D bounding boxes [40] and grasp labels [6]. To mitigate this issue, we design a data engine to extract 3D hand poses and contact maps (Sec. 3.4) from egocentric videos showing everyday interactions. Specifically, we extract 3D trajectories from HoloAssist for 800 tasks performed by ordinary users across 120 object categories and 24 action categories, which is " + }, + { + "bbox": [ + 313, + 415, + 556, + 605 + ], + "type": "inline_equation", + "content": "2.5 \\times 10 \\times" + }, + { + "bbox": [ + 313, + 415, + 556, + 605 + ], + "type": "text", + "content": " larger than existing datasets [2, 11, 28, 43]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 606, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 557, + 714 + ], + "type": "text", + "content": "HOI representation: Hands are often represented using a parametric mesh model, e.g., MANO [37], along with the global translation and rotation of the wrist [7, 11, 13, 30, 31, 33, 51, 53, 54]. Given an object mesh, its 6DoF pose can be denoted using translation & rotation [13, 55], signed distance field [54] or using Basis Point Set (BPS [34]) distances [7, 26]. For hand-object contact, existing approaches typically consider a canonicalized representation with the hand [4, 54] or object [7] at the origin & use relative dis" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7027" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 73, + 289, + 217 + ], + "blocks": [ + { + "bbox": [ + 59, + 73, + 289, + 217 + ], + "lines": [ + { + "bbox": [ + 59, + 73, + 289, + 217 + ], + "spans": [ + { + "bbox": [ + 59, + 73, + 289, + 217 + ], + "type": "image", + "image_path": "603f40e6e920e5ab8c0ad6144b250ab3944488914530f194c78d79c380d2f865.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 230, + 296, + 298 + ], + "lines": [ + { + "bbox": [ + 55, + 230, + 296, + 298 + ], + "spans": [ + { + "bbox": [ + 55, + 230, + 296, + 298 + ], + "type": "text", + "content": "Figure 2. We represent contact points as binary masks on the hand mesh vertices. The hand mesh is represented in the camera coordinate system, consisting of the local hand pose in the MANO [38] coordinate frame and a global transformation from the MANO frame to the camera frame. These hand poses and contact points over several timesteps form interaction trajectories." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 309, + 295, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 309, + 295, + 369 + ], + "spans": [ + { + "bbox": [ + 54, + 309, + 295, + 369 + ], + "type": "text", + "content": "stances [7, 11, 12, 32, 44] or a distance field [54] from hand joints to represent contact. Since 3D hand-object motion annotations are unavailable for diverse objects, e.g., thin, occluded, deformable, we instead focus on sequences of 3D hand poses and contact maps, i.e., interaction trajectories." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 369, + 295, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 369, + 295, + 489 + ], + "spans": [ + { + "bbox": [ + 54, + 369, + 295, + 489 + ], + "type": "text", + "content": "Affordances: Gibson [14] developed the concept of affordances as a set of functionalities that the environment furnishes to an agent, which can be a human, a robot, an animal, or hands. Affordances are typically represented as the contact regions on the objects [1, 5, 23, 25, 27, 28, 43, 50] & the agent [1, 19, 23, 42, 43, 47, 53, 53, 54]. They have been explored in domains like robotics [1, 42], human interactions [19, 43, 47, 49] & hand grasp synthesis [23, 42, 43, 53, 54]. Recent works have also proposed generative models for affordances synthesis [49, 53, 54]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 501, + 111, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 501, + 111, + 513 + ], + "spans": [ + { + "bbox": [ + 55, + 501, + 111, + 513 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 521, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 653 + ], + "type": "text", + "content": "Given a single RGB view, action text, and a 3D contact point as input, we focus on predicting 3D hand poses & contact points, i.e., interaction trajectories (Fig. 1), for future timesteps. Our approach consists of: (1) Interaction Codebook: learns a latent codebook of affordances using a VQVAE (Fig. 3a), (2) Learned Indexer: maps the action text, image (showing an object) & contact point inputs to the codebook indices to obtain the corresponding embeddings (Fig. 3b), (3) Interaction Predictor: takes the queried embeddings along with text, image and contact point inputs to output the 3D interaction trajectory (Fig. 3b)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": "Hand mesh: We use the MANO [37] model to represent the hand. It consists of pose " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": " and shape " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": " parameters, which can be converted to a 3D hand mesh. These mesh vertices lie in the MANO coordinate system, which can be transformed to the camera coordinate system at the first" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": "timestep (reference frame) using a rotation and translation. This gives us the hand trajectory in the reference frame." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 97, + 555, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 97, + 555, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 97, + 555, + 156 + ], + "type": "text", + "content": "Contact map: We represent the contact points as a binary mask over the MANO hand mesh vertices, i.e., a 778 dimensional binary vector for each timestep, referred to as contact map in the paper (Fig. 2). The contact points may also change over time depending on the action, e.g. twisting." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 157, + 556, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 157, + 556, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 556, + 373 + ], + "type": "text", + "content": "Once the 3D hand mesh is transformed into reference frame, we estimate the trajectory of the contact points by computing the centroid of the contact points at each timestep. The contact centroid in the reference frame is used as the 3D contact point input during training. We assume a 3D contact point to be available in the first frame at test time. To input this point, we create a 3D voxel grid of resolution " + }, + { + "bbox": [ + 313, + 157, + 556, + 373 + ], + "type": "inline_equation", + "content": "16 \\times 16 \\times 16" + }, + { + "bbox": [ + 313, + 157, + 556, + 373 + ], + "type": "text", + "content": ". We take minimum and maximum values of 3D hand trajectory (along each axis, in the reference frame) across all training sequences. This span is divided into 16 equal parts per axis, associating each voxel to 3D metric range. It is kept fixed at train & test time. Thus, the center of each voxel represents a 3D location in metric space, i.e., a 3D metric grid, in the camera coordinate system. We create a 3D gaussian heatmap centered at 3D contact point in this grid, referred to as voxelized contact points in the paper. This changes as the trajectory evolves. This is different than the contact map (binary mask over mesh vertices)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 382, + 497, + 394 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 382, + 497, + 394 + ], + "spans": [ + { + "bbox": [ + 313, + 382, + 497, + 394 + ], + "type": "text", + "content": "3.1. Interaction Codebook (InterCode)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 400, + 556, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 400, + 556, + 423 + ], + "spans": [ + { + "bbox": [ + 313, + 400, + 556, + 423 + ], + "type": "text", + "content": "This is implemented as a VQVAE with a transformer encoder-decoder architecture [16]. It consists of 3 modules:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 424, + 556, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 424, + 556, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 424, + 556, + 533 + ], + "type": "text", + "content": "Encoder: The hand and contact map trajectories for a fixed time horizon " + }, + { + "bbox": [ + 313, + 424, + 556, + 533 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 424, + 556, + 533 + ], + "type": "text", + "content": " are passed to a transformer encoder to get the feature embedding. These features are used to query the codebook to obtain the relevant embeddings using L2 distance. Specifically, the input trajectory, consisting of MANO parameters & contact maps, is first passed to a 1-layer MLP to match the feature dimension of the transformer. We use a 1-layer transformer encoder with 1 head & feature dimension = 512. It outputs a " + }, + { + "bbox": [ + 313, + 424, + 556, + 533 + ], + "type": "inline_equation", + "content": "T \\times 512" + }, + { + "bbox": [ + 313, + 424, + 556, + 533 + ], + "type": "text", + "content": " dimensional feature." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 533, + 556, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 556, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 556, + 665 + ], + "type": "text", + "content": "Latent codebook: It is initialized as a " + }, + { + "bbox": [ + 313, + 533, + 556, + 665 + ], + "type": "inline_equation", + "content": "K \\times E" + }, + { + "bbox": [ + 313, + 533, + 556, + 665 + ], + "type": "text", + "content": " dimensional matrix and is updated during training using exponential moving average of the embeddings output from the encoder, following recent works [16]. Concretely, we use a multilayer residual VQVAE architecture [16]. We set the number of quantizers for VQVAE to 6, " + }, + { + "bbox": [ + 313, + 533, + 556, + 665 + ], + "type": "inline_equation", + "content": "K = 512" + }, + { + "bbox": [ + 313, + 533, + 556, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 533, + 556, + 665 + ], + "type": "inline_equation", + "content": "E = 512" + }, + { + "bbox": [ + 313, + 533, + 556, + 665 + ], + "type": "text", + "content": ". When querying the codebook during training, sampling is done using the Gumbel-SoftMax trick [21] to preserve differentiability. We set the temperature of Gumbel-Softmax to 0.5 & inject noise at training for stochastic sampling. At test time, argmax is used to get the closest codebook entry." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": "Decoder: It takes multiple inputs: (1) the latent embeddings queried from the codebook " + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "inline_equation", + "content": "(T \\times 512)" + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": ", (2) CLIP [35] embeddings of the text describing the action (512 dimensional feature extended for Timesteps), (3) DeiT [46] embeddings" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7028" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 69, + 294, + 168 + ], + "blocks": [ + { + "bbox": [ + 60, + 69, + 294, + 168 + ], + "lines": [ + { + "bbox": [ + 60, + 69, + 294, + 168 + ], + "spans": [ + { + "bbox": [ + 60, + 69, + 294, + 168 + ], + "type": "image", + "image_path": "aecf9241ff992cc297975b0a79be9bf9f0f79944f5ad80ef75d575a7213ac360.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 170, + 298, + 227 + ], + "lines": [ + { + "bbox": [ + 55, + 170, + 298, + 227 + ], + "spans": [ + { + "bbox": [ + 55, + 170, + 298, + 227 + ], + "type": "text", + "content": "(a) Interaction Codebook is a VQVAE model that learns a latent codebook of 3D interaction trajectories, consisting of a transformer encoder-decoder architecture. The encoder features are used to sample codebook indices and corresponding embeddings which are passed to a decoder that also takes in the video input and text describing the action and reconstructs the 3D interaction trajectories." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 313, + 70, + 553, + 164 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 553, + 164 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 553, + 164 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 553, + 164 + ], + "type": "image", + "image_path": "a489bae29a12395225aed3b557f007ab2e69ccf0305c4770b44cf9479adcbc53.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 312, + 169, + 556, + 228 + ], + "lines": [ + { + "bbox": [ + 312, + 169, + 556, + 228 + ], + "spans": [ + { + "bbox": [ + 312, + 169, + 556, + 228 + ], + "type": "text", + "content": "(b) We train an Indexer module to map the inputs at test time to the codebook indices to extract the relevant embeddings. These are then passed to a interaction predictor module that outputs the 3D interaction trajectories. We consider 2 settings: (1) Forecasting: text describing the action, single image & 3D contact point, (2) Interpolation: also providing the goal image showing the final state of the interaction (not shown here for clarity)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 236, + 555, + 281 + ], + "lines": [ + { + "bbox": [ + 54, + 236, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 54, + 236, + 555, + 281 + ], + "type": "text", + "content": "Figure 3. Overview. Our framework involves a 2-stage training procedure: (left) Interaction Codebook: to learn a latent codebook of hand poses and contact points, i.e., tokenizing interaction trajectories, (right) a learned Indexer & an Interaction Predictor module to predict the interaction trajectories from single image, action text & 3D contact point. We use pretrained features for images (from DeiT [46]) and text (from CLIP [35]). 3D contact point is input as a 3D gaussian heatmap in a 3D voxel grid (omitted here for clarity)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 293, + 297, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 293, + 297, + 484 + ], + "spans": [ + { + "bbox": [ + 54, + 293, + 297, + 484 + ], + "type": "text", + "content": "of the video showing the actions " + }, + { + "bbox": [ + 54, + 293, + 297, + 484 + ], + "type": "inline_equation", + "content": "(T \\times 768)" + }, + { + "bbox": [ + 54, + 293, + 297, + 484 + ], + "type": "text", + "content": ", (4) voxelized contact points at each timestep, which are passed through four 3D convolution layers. We also pass in a 4D grid with each voxel containing the 3D location of its center in metric space, and process with 3D convolution layers. This gives us a 32 dimensional feature vector from the contact module. Concatenating all these input features, we get a 1824-dimensional joint feature that is passed through a 1-layer MLP to change the dimensionality to 512 and then fed to a transformer decoder (1-layer with 1 head & dropout of 0.2). The query embeddings for the transformer decoder are set as trainable parameters. It outputs a " + }, + { + "bbox": [ + 54, + 293, + 297, + 484 + ], + "type": "inline_equation", + "content": "T \\times 512" + }, + { + "bbox": [ + 54, + 293, + 297, + 484 + ], + "type": "text", + "content": " feature map which is passed to separate decoders for predicting the MANO parameters and the contact map. We use a 2-layer MLP to predict the MANO parameters and a 3-layer MLP to predict the contact maps for the entire time horizon " + }, + { + "bbox": [ + 54, + 293, + 297, + 484 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 54, + 293, + 297, + 484 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 484, + 296, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 484, + 296, + 557 + ], + "spans": [ + { + "bbox": [ + 55, + 484, + 296, + 557 + ], + "type": "text", + "content": "Training: The loss function consists of several terms: (1) Smooth L1 loss on the MANO parameters, (2) L1 loss on the contact centroid, (3) L1 loss on the translation between the future frame and first frame, (4) L1 loss on the 6D rotation representation between the future frame and first frame, (5) binary cross-entropy loss on the contact map predictions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 564, + 157, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 157, + 575 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 157, + 575 + ], + "type": "text", + "content": "3.2. Learned Indexer" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 582, + 296, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 689 + ], + "type": "text", + "content": "At test time, we need to query the learned codebook to get the latent interaction embeddings. This is done using a learned module which takes the test time inputs, i.e. text describing an action, image showing an object & a 3D contact point, and outputs a probability distribution over the codebook indices. The indices and corresponding embeddings can be sampled using the predicted probabilities. Note that the test time inputs only consist of a single image and a 3D contact point, unlike the entire trajectory used for training InterCode." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Specifically, we first compute features from different inputs: CLIP embedding of text (512), DeiT embeddings of" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 293, + 556, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 556, + 424 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 556, + 424 + ], + "type": "text", + "content": "images (768), features from voxelized contact point, in the reference frame, after passing through 4 3D convolutional layers (32) and the 3D contact point. We concatenate these features and process through a 2-layer MLP (1024 & 512 nodes) to match the input dimensions of the transformer classifier module. It is implemented as a 1 layer, 1 head transformer decoder layer that takes in learnable query embeddings and aforementioned features from different modalities to output " + }, + { + "bbox": [ + 313, + 293, + 556, + 424 + ], + "type": "inline_equation", + "content": "T \\times 512" + }, + { + "bbox": [ + 313, + 293, + 556, + 424 + ], + "type": "text", + "content": " dimensional features. These features are then passed to a 3-layer MLP to output a probability distribution over the indices of the codebook." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 425, + 556, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 425, + 556, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 425, + 556, + 460 + ], + "type": "text", + "content": "Training: It consists of a cross-entropy loss with ground truth computed as the closest codebook indices (L2 distance) from the encoder features of the InterCode module." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 468, + 493, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 493, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 493, + 480 + ], + "type": "text", + "content": "3.3. Interaction Predictor (InterPred)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 486, + 556, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 556, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 556, + 640 + ], + "type": "text", + "content": "Since the inputs at test time differ from the decoder inputs of the InterCode module during training, we need a separate learned module to predict the interaction trajectories. It consists of a transformer decoder module with the following inputs: (1) the latent embeddings queried from the codebook using the learned prior, (2) CLIP embeddings of the text describing the action, (3) DeiT embeddings of the image showing the object, and (4) voxelized contact point in the reference frame. These inputs are processed in the same way as for Learned Indexer before passing to the transformer decoder. The architecture is same as the decoder in InterCode. While the inputs are different than InterCode, they are passed to an MLP to match transformer dimension." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "Training: The loss function consists of several terms: (1) Smooth L1 loss on the MANO parameters, (2) L1 loss on the contact centroid, (3) L1 loss on the translation between the future frame and first frame, (4) L1 loss on the 6D rotation representation between the future frame and first frame, (5) binary cross-entropy loss on the contact map predictions." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7029" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 69, + 293, + 133 + ], + "blocks": [ + { + "bbox": [ + 57, + 69, + 293, + 133 + ], + "lines": [ + { + "bbox": [ + 57, + 69, + 293, + 133 + ], + "spans": [ + { + "bbox": [ + 57, + 69, + 293, + 133 + ], + "type": "image", + "image_path": "545cd36b3f1354f3f05c212dd7e401b3661fc3b8615379fc74767c4c5b169365.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 147, + 293, + 199 + ], + "blocks": [ + { + "bbox": [ + 61, + 137, + 217, + 146 + ], + "lines": [ + { + "bbox": [ + 61, + 137, + 217, + 146 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 217, + 146 + ], + "type": "text", + "content": "Generated annotations for the task pull lever" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 147, + 293, + 199 + ], + "lines": [ + { + "bbox": [ + 59, + 147, + 293, + 199 + ], + "spans": [ + { + "bbox": [ + 59, + 147, + 293, + 199 + ], + "type": "image", + "image_path": "83221cfffeb490e8dee0df08fa3c171e73b2485445b4bc94f5743b95ad0e35a2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 206, + 295, + 273 + ], + "lines": [ + { + "bbox": [ + 55, + 206, + 295, + 273 + ], + "spans": [ + { + "bbox": [ + 55, + 206, + 295, + 273 + ], + "type": "text", + "content": "Figure 4. Data engine. (top) Object masks are extracted using SAMv2 [36], 3D hand poses & masks (2D rendering of mesh) are from HaMeR [30], contact points are computed by projecting the 3D hand points into the 2D contact region (intersection of hand & object masks). (bottom) Generated object masks (highlighted in white), 3D hand mesh & contact points for 3 timesteps." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 285, + 137, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 285, + 137, + 298 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 137, + 298 + ], + "type": "text", + "content": "3.4. Data Engine" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 304, + 296, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 304, + 296, + 519 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 296, + 519 + ], + "type": "text", + "content": "Existing lab datasets are limited to simple interactions in constrained setups. Instead, we use egocentric videos from HoloAssist showing everyday interactions with rigid, articulated & deformable objects, e.g. operating machines, tool use, kitchen tasks. It consists of " + }, + { + "bbox": [ + 55, + 304, + 296, + 519 + ], + "type": "inline_equation", + "content": "2\\mathrm{k}+" + }, + { + "bbox": [ + 55, + 304, + 296, + 519 + ], + "type": "text", + "content": " videos with temporally localized atomic actions, e.g. grab, open, screw, mix, rotate, align, slide, etc. with diverse objects. We extract these clips of atomic actions and convert them into the required format for training & evaluation. We consider 24 action categories with 120 object categories spanning 800 tasks. This is 2.5- " + }, + { + "bbox": [ + 55, + 304, + 296, + 519 + ], + "type": "inline_equation", + "content": "10\\times" + }, + { + "bbox": [ + 55, + 304, + 296, + 519 + ], + "type": "text", + "content": " larger than existing datasets, e.g. HOT3D [2] (33 rigid objects), ARCTIC [11] (10 articulated objects), GRAB [43] (51 objects), HOI4D [28] (16 object categories), in terms of objects & diversity of interactions. While HoloAssist contains diverse interactions, it does not provide 3D annotations for hand poses & contact points. To extract these annotations, we design a semi-automatic data pipeline (Fig. 4) using 2D segmentation masks, 3D hand poses & 2D contact region." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 521, + 295, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 580 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 580 + ], + "type": "text", + "content": "Segmentation masks: We adopt a semi-automatic approach for object masks, that involves clicking a point on the target object in the first frame of a video & tracking the object mask across the video using SAMv2 [36]. For 2D hand masks, we render the 3D hand mesh (described next) into the image." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 581, + 296, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 581, + 296, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 581, + 296, + 652 + ], + "type": "text", + "content": "3D hand poses: We use the off-the-shelf HaMeR [30] model to get 3D hand meshes in each video frame. HaMeR takes an image crop around the hand as input (estimated from Hands23 model [6]). The predicted 3D mesh in each frame is transformed to the reference frame using known intrinsics & extrinsics available in the HoloAssist dataset." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": "3D Contact points: We first compute a 2D contact region in an image as the overlap between the hand & object masks. We add gaussian noise at the boundary to obtain a dense 2D contact region. We back project the 2D contact region into 3D hand mesh to obtain 3D contact points on the hand mesh." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 71, + 394, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 394, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 394, + 84 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 91, + 376, + 103 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 91, + 376, + 103 + ], + "spans": [ + { + "bbox": [ + 314, + 91, + 376, + 103 + ], + "type": "text", + "content": "4.1. Protocol" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 110, + 555, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 110, + 555, + 253 + ], + "spans": [ + { + "bbox": [ + 313, + 110, + 555, + 253 + ], + "type": "text", + "content": "Task: We predict trajectories of MANO hand parameters & contact map (a binary mask over 778 vertices of the MANO [38] hand mesh) in 2 settings: (1) Forecasting: input consists of a textual description of the action, a current image showing the object of interaction, and a 3D contact point. (2) Interpolation: We provide the goal image, along with the previous inputs, showing the final state of interaction. For each setting, we 2 two variants: (a) current image with a hand, (b) current image without a hand (removed via in-painting [52]) since the hand may not be visible before the start of the interaction in many practical scenarios. We only consider right hand motion in this work." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 255, + 555, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 255, + 555, + 327 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 555, + 327 + ], + "type": "text", + "content": "Generalization aspects: We study generalization capabilities for 4 different settings: (1) Object categories, (2) Action categories, (3) Novel tasks: a combination of object & action categories, (4) Novel scene: we hold out videos from 1 location (data is collected in 2 locations). For each setting, we create train, validation & test splits in the ratio 80:10:10." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 327, + 555, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 555, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 555, + 411 + ], + "type": "text", + "content": "Metrics: (1) MPJPE: Mean Per Joint Position Error between predicted & ground truth hand joints, averaged over all future timesteps. (2) MPJPE-PA is the procrustus aligned variant, which optimizes for a single rotation, translation & scale to align the entire trajectory with ground truth trajectory, before computing MPJPE. (3) F1 score: harmonic mean of precision & recall to measure accuracy of contact maps." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 412, + 556, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 412, + 556, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 412, + 556, + 592 + ], + "type": "text", + "content": "Baselines: Since no prior works predict interaction trajectories from image input, we adapt 2 recent baselines from human pose literature to work in our setting: (1) HCTFormer: The input image is encoded using a ViT [10] encoder followed by a transformer decoder, similar to the pose decoder used in T2P [22]. TCP decoder takes in trajectory query (for global intent), pose query & pose embeddings. In our case, intent comes from text. We replace pose embeddings with {image, contact, text} features, pass them to an MLP & then feed to {TRM, MLP} modules, as in TCP. (2) HCTDiff: We modify the motion diffusion model (MDM [45]) for hand trajectory and contact map prediction. For this, we change conditioning to use {image, text, contact} features but diffusion architecture is same. Both HCTFormer and HCTDiff are trained in the same setting as ours for fair comparisons." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 593, + 555, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 593, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 593, + 555, + 653 + ], + "type": "text", + "content": "Existing methods that use object-centric representations [4, 7] can not be applied in our setting since we do not have the 3D object models. Hand-centric representations [44, 54] are also not suitable in our setting since the hand is not always visible in the input image at test time." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "text", + "content": "Our approach involves a 2-stage training procedure: InterCode followed by Indexer & InterPred, from different inputs. We train with both transformer & diffusion variants for InterPred, referred to as LatentAct & LatentAct-Diff respectively. LatentAct-Diff uses same InterCode & Indexer as" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7030" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 555, + 264 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 555, + 264 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 555, + 264 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 555, + 264 + ], + "type": "table", + "html": "
MethodTask-levelObject-levelAction-levelScene-level
M-PE(cm)↓M-PA(cm)↓F1↑M-PE(cm)↓M-PA(cm)↓F1↑M-PE(cm)↓M-PA(cm)↓F1↑M-PE(cm)↓M-PA(cm)↓F1↑
Hand visibleForecasting
HCTFormer8.323.060.728.703.190.728.593.250.738.523.090.72
HCTDiff8.422.750.729.012.880.739.113.100.749.492.840.71
LatentAct (Ours)7.612.990.758.093.060.758.143.320.777.873.040.74
Interpolation
HCTFormer7.523.020.768.103.100.767.773.290.787.543.000.76
HCTDiff8.302.720.778.522.910.788.383.160.788.602.810.77
LatentAct (Ours)6.722.870.807.483.020.787.253.190.807.272.930.78
No handsForecasting
HCTFormer8.342.950.728.853.020.738.543.210.748.533.050.72
HCTDiff8.852.650.729.742.810.729.503.110.749.212.740.71
LatentAct (Ours)7.932.970.768.263.010.768.193.280.777.903.000.75
Interpolation
HCTFormer7.302.950.768.423.150.768.013.300.777.292.990.77
HCTDiff8.272.760.768.682.910.769.333.100.7810.022.820.76
LatentAct (Ours)7.132.890.797.242.970.797.443.210.807.122.890.78
", + "image_path": "315f8a4fe9627890556c1ed5b267a30cebe823a9dc7a35a32e2fb59e0669e4f1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 328, + 294, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 328, + 294, + 389 + ], + "spans": [ + { + "bbox": [ + 55, + 328, + 294, + 389 + ], + "type": "text", + "content": "LatentAct but different InterPred, i.e., an iterative denoising model from MDM instead of a single step prediction. The denoising is done in latent space of the codebook followed by a decoder to generate trajectories. It is trained using L2 loss on denoised embedding and loss in Sec. 3.3." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 396, + 113, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 113, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 113, + 407 + ], + "type": "text", + "content": "4.2. Results" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 414, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 414, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 54, + 414, + 295, + 581 + ], + "type": "text", + "content": "Generalization results: First, we study if the model can generate trajectories from the text describing the action, single image & 3D contact point, i.e., Forecasting task. In the absence of a goal, this is quite challenging and inherently multimodal in nature. We compare with the GT trajectories for all 4 generalization aspects: novel objects, actions, tasks & scene. We report MPJPE, MPJPE-PA for hand trajectories & F1-score for contact maps. We observe that our approach leads to consistent gains in absolute hand poses & contact maps. We show results (Tab. 1) on 2 variants of the task, (1) hand visible in the image, (2) hand absent in the image, often the case at the start of interactions. The hand visible in the image makes the tasks slightly easier since the initial pose is visible, as also evident from the results." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 582, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 665 + ], + "type": "text", + "content": "Next, we also provide the goal image, along with the previous inputs, showing the final state of interaction, referred to as goal-conditioned interpolation. This is slightly easier than the Forecasting task since the final state of the interaction is also provided. The results (Tab. 1) on 2 varaints of the task, (1) hand visible in the image, (2) hand absent in the image, show similar benefits as in Forecasting." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 665, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 689 + ], + "type": "text", + "content": "These results are for " + }, + { + "bbox": [ + 55, + 665, + 295, + 689 + ], + "type": "inline_equation", + "content": "T = 30" + }, + { + "bbox": [ + 55, + 665, + 295, + 689 + ], + "type": "text", + "content": " timesteps. We also report results for " + }, + { + "bbox": [ + 55, + 665, + 295, + 689 + ], + "type": "inline_equation", + "content": "T = 16" + }, + { + "bbox": [ + 55, + 665, + 295, + 689 + ], + "type": "text", + "content": " in the supplementary with similar trends." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "Results on ARCTIC: We also show results (Tab. 2) on ARCTIC [11] to check if LatentAct leads to gains on existing" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 315, + 327, + 553, + 386 + ], + "blocks": [ + { + "bbox": [ + 55, + 271, + 555, + 316 + ], + "lines": [ + { + "bbox": [ + 55, + 271, + 555, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 271, + 555, + 316 + ], + "type": "text", + "content": "Table 1. Generalization results. We report MPJPE (M-PE: cm), MPJPE-PA (M-PA: cm) & F1-score in 4 generalization settings: novel tasks, objects, actions & scene, to measure the accuracy of the predicted trajectories. We adapt two recent methods from human pose literature to work with image inputs, i.e., HCTFormer: ViT encoded image features passed to a transformer decoder (similar to pose decoder of T2P [22]), HCTDiff: MDM [45] modified to take image features as well. LatentAct leads to better absolute hand poses & contact maps." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 327, + 553, + 386 + ], + "lines": [ + { + "bbox": [ + 315, + 327, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 315, + 327, + 553, + 386 + ], + "type": "table", + "html": "
MethodMPJPE↓MPJPE-PA↓F1↑
HCTFormer: trained on ARCTIC15.763.610.36
LatentAct: trained on ARCTIC14.773.760.41
LatentAct: zero-shot from Holo15.723.710.19
LatentAct: Codebook & Indexer from Holo15.363.580.45
", + "image_path": "e44fe53372e239988e18729858f9f93d3ae0c2cf4f5cd985012d77ddf000b4d9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 394, + 555, + 428 + ], + "lines": [ + { + "bbox": [ + 313, + 394, + 555, + 428 + ], + "spans": [ + { + "bbox": [ + 313, + 394, + 555, + 428 + ], + "type": "text", + "content": "Table 2. HoloAssist is significantly larger than ARCTIC in terms of contact sequences leading to benefits in zero-shot generalization & transferring models trained on HoloAssist to ARCTIC." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 436, + 555, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 436, + 555, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 436, + 555, + 519 + ], + "type": "text", + "content": "lab datasets. ARCTIC has accurate 3D labels but is limited in scale (" + }, + { + "bbox": [ + 313, + 436, + 555, + 519 + ], + "type": "inline_equation", + "content": "10 \\times" + }, + { + "bbox": [ + 313, + 436, + 555, + 519 + ], + "type": "text", + "content": " smaller than HoloAssist for contact sequences). When trained only on ARCTIC, we find the trends to transfer from HoloAssist, i.e., LatentAct is better than HCTFormer. Using codebook & indexer trained on HoloAssist further benefits models trained on ARCTIC. Zero-shot transfer from HoloAssist is competitive to models trained on ARCTIC." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 520, + 556, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 520, + 556, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 520, + 556, + 628 + ], + "type": "text", + "content": "Benefits of Interaction Codebook: We verify if the 2-stage training procedure is beneficial by comparing with single-stage training methods HCTFormer & HCTDiff in Tab. 4 and observe consistent benefits. Another option is to retrieve an interaction trajectory from the training set, conditioned on the input. However, this does not scale well with the training data & is computationally expensive, e.g. the learned codebook in InterCode has 512 entries whereas the training dataset has " + }, + { + "bbox": [ + 313, + 520, + 556, + 628 + ], + "type": "inline_equation", + "content": "\\sim 15\\mathrm{K}" + }, + { + "bbox": [ + 313, + 520, + 556, + 628 + ], + "type": "text", + "content": " sequences, which is significantly larger." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 629, + 555, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 555, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 555, + 677 + ], + "type": "text", + "content": "Contact maps helps hand predictions: Here, we train the InterCode & InterPred modules with only the hand pose loss, i.e., removing the loss on contact maps. In Tab. 5, we observe that loss on contact maps helps with hand predictions." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "Trends with LatentAct-Diff.: Training InterPred with diffusion losses leads to better hand poses but worse contact maps compared to LatentAct, across all settings (Tab. 3)." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7031" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 553, + 228 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 228 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 228 + ], + "type": "table", + "html": "
MethodTask-levelObject-levelAction-levelScene-level
M-PE(cm)↓M-PA(cm)↓F1↑M-PE(cm)↓M-PA(cm)↓F1↑M-PE(cm)↓M-PA(cm)↓F1↑M-PE(cm)↓M-PA(cm)↓F1↑
Hand visibleForecasting
LatentAct7.612.990.758.093.060.758.143.320.777.873.040.74
LatentAct-Diff7.903.030.718.493.200.728.293.340.728.323.100.72
Interpolation
LatentAct6.722.870.807.483.020.787.253.190.807.272.930.78
LatentAct-Diff6.532.620.787.232.810.787.113.100.796.702.840.78
No handsForecasting
LatentAct7.932.970.768.263.010.768.193.280.777.903.000.75
LatentAct-Diff7.822.890.718.633.060.718.193.150.727.932.960.73
Interpolation
LatentAct7.132.890.797.242.970.797.443.210.807.122.890.78
LatentAct-Diff6.602.650.767.732.780.767.423.040.786.892.710.77
", + "image_path": "99ebcb5415c45814e09171271b1dbf84db2f4f428f4c4e56bc9c0442e4d5ecca.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 65, + 257, + 288, + 322 + ], + "blocks": [ + { + "bbox": [ + 56, + 236, + 553, + 247 + ], + "lines": [ + { + "bbox": [ + 56, + 236, + 553, + 247 + ], + "spans": [ + { + "bbox": [ + 56, + 236, + 553, + 247 + ], + "type": "text", + "content": "Table 3. LatentAct-Diff trends. Training InterPred with diffusion loss leads to better hand poses but worse contact maps than LatentAct." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 65, + 257, + 288, + 322 + ], + "lines": [ + { + "bbox": [ + 65, + 257, + 288, + 322 + ], + "spans": [ + { + "bbox": [ + 65, + 257, + 288, + 322 + ], + "type": "table", + "html": "
MethodMPJPE↓MPJPE-PA↓F1↑
HCTFormer7.523.020.76
HCTFormer + InterCode6.722.870.80
HCTDiff8.302.720.77
HCTDiff + InterCode6.532.620.78
", + "image_path": "4b72fc47b89c6054713915770dafe41844ae10b6b661076e84c571abd9a192dd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 58, + 370, + 293, + 432 + ], + "blocks": [ + { + "bbox": [ + 58, + 370, + 293, + 432 + ], + "lines": [ + { + "bbox": [ + 58, + 370, + 293, + 432 + ], + "spans": [ + { + "bbox": [ + 58, + 370, + 293, + 432 + ], + "type": "image", + "image_path": "1a55bf12358673256d2af4439faa8ca1c6ea5c68fbb8f295467f405f0dccfb73.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 443, + 295, + 455 + ], + "lines": [ + { + "bbox": [ + 55, + 443, + 295, + 455 + ], + "spans": [ + { + "bbox": [ + 55, + 443, + 295, + 455 + ], + "type": "text", + "content": "Figure 5. More training data helps both LatentAct & HCTFormer." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 461, + 296, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 461, + 296, + 486 + ], + "spans": [ + { + "bbox": [ + 55, + 461, + 296, + 486 + ], + "type": "text", + "content": "Dataset scale: In Fig. 5, we see that both LatentAct & HCTFormer improve with larger training dataset size." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 493, + 144, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 493, + 144, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 493, + 144, + 504 + ], + "type": "text", + "content": "4.3. Visualizations" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 510, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 510, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 510, + 297, + 713 + ], + "type": "text", + "content": "We compare the predictions of LatentAct & the best baseline on a few examples from task level generalization setting (Fig. 6) for both forecasting & interpolation for 3 timesteps " + }, + { + "bbox": [ + 54, + 510, + 297, + 713 + ], + "type": "inline_equation", + "content": "(t = 5, 15, 25)" + }, + { + "bbox": [ + 54, + 510, + 297, + 713 + ], + "type": "text", + "content": ". The left column shows the input image with the contact point (projected in the image in cyan blob), goal image (for interpolation task) & text describing the action. The other columns show: (a) Camera View: predicted hand in the camera view with the contact point, this captures the placement of the hand around the contact point, (b) Another View: this better visualizes the hand articulation from a different camera viewpoint, (c) Contact Maps are shown as red parts of the hand mesh. LatentAct leads to better orientation of hand & sharper contact maps than the baseline. Note that these tasks are not seen during training. More visualizations & analysis are provided in the supplementary. Limitations: We do not predict the object state change. While 3D object models are not available, alternatives like" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 323, + 257, + 547, + 372 + ], + "blocks": [ + { + "bbox": [ + 55, + 330, + 297, + 363 + ], + "lines": [ + { + "bbox": [ + 55, + 330, + 297, + 363 + ], + "spans": [ + { + "bbox": [ + 55, + 330, + 297, + 363 + ], + "type": "text", + "content": "Table 4. Two stage training with InterCode improves over single-stage training methods HCTFormer & HCTDiff. Note that HCT-Former + InterCode is same as LatentAct." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 323, + 257, + 547, + 372 + ], + "lines": [ + { + "bbox": [ + 323, + 257, + 547, + 372 + ], + "spans": [ + { + "bbox": [ + 323, + 257, + 547, + 372 + ], + "type": "table", + "html": "
MethodMPJPE↓MPJPE-PA↓F1↑
HCTFormer7.523.020.76
- no contact map7.652.89-
LatentAct6.722.870.80
- no contact map7.762.91-
HCTDiff8.302.720.77
- no contact map8.432.78-
LatentAct-Diff6.532.620.78
- no contact map7.092.90-
", + "image_path": "16081278a5f7f2890cd19a4c15c8a69d26ca25370bf3a80523f1f690e7278801.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 380, + 553, + 392 + ], + "lines": [ + { + "bbox": [ + 313, + 380, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 553, + 392 + ], + "type": "text", + "content": "Table 5. Loss on contact maps helps hand predictions for all models." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 405, + 556, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 405, + 556, + 537 + ], + "spans": [ + { + "bbox": [ + 313, + 405, + 556, + 537 + ], + "type": "text", + "content": "predicting 2D object masks could be explored. While we assume a 3D contact point to be available, it can also be estimated using off-the-shelf depth models & ground truth intrinsics to get the 3D location of a 2D point that can be prompted by a user (e.g. by clicking a point on the image). Our experiments measure whether the predicted interaction trajectory is accurate, but it does not account for the multimodal nature of future prediction. The annotations in our work are programatically generated and may not be accurate in all cases. While we filter out the erroneous cases, using a better annotation tool [47] would be beneficial." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 549, + 388, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 549, + 388, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 549, + 388, + 562 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": "Our model predicts interaction trajectories, i.e. 3D hand motion & contact maps, from single RGB view, action text & 3D contact point as input. It consists of: (1) Interaction Codebook: a VQVAE to learn a latent codebook of 3D hand poses & contact points, i.e., tokenizing interactions. (2) Interaction Predictor: a transformer-decoder module to predict the interaction trajectory from test time inputs, by using an indexer module to retrieve a latent interaction from the learned codebook. Our large-scale experiments on the diverse HoloAssist dataset show benefits across 4 generalization settings. Our work is a first step towards learning a motion prior on interaction sequences and can be augmented" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7032" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 78, + 69, + 541, + 608 + ], + "blocks": [ + { + "bbox": [ + 78, + 69, + 541, + 608 + ], + "lines": [ + { + "bbox": [ + 78, + 69, + 541, + 608 + ], + "spans": [ + { + "bbox": [ + 78, + 69, + 541, + 608 + ], + "type": "table", + "html": "
Input Image (t=0)Camera ViewAnother ViewContact Map
mix/stir coffeeBaselinet=5t=15t=25t=5t=15t=25t=5t=15t=25
GT.........
.........
rotate lensBaseline.........
GT.........
.........
pull batteryBaseline.........
GT.........
.........
slide controllerBaseline.........
GT.........
", + "image_path": "cf32ca31901cabed8b9368942629133b7bba977a85ba0a23923a3d4cdca3f66c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 615, + 555, + 682 + ], + "lines": [ + { + "bbox": [ + 55, + 615, + 555, + 682 + ], + "spans": [ + { + "bbox": [ + 55, + 615, + 555, + 682 + ], + "type": "text", + "content": "Figure 6. Visualizations. We compare the predictions of LatentAct & the best baseline on a few examples from the task level generalization setting for both forecasting & interpolation for 3 timesteps (t=5, 15, 25). The left column shows the input image with the contact point (projected in the image in cyan blob), goal image (for interpolation task) & text describing the action. The other columns show: (a) Camera View: predicted hand in the camera view with the contact point, this captures the placement of the hand around the contact point, (b) Another View: this better visualizes the hand pose from a different camera viewpoint, (c) Contact Maps are shown as red parts of the hand mesh. LatentAct leads to better orientation of hand & sharper contact maps than the baseline. Note that these tasks are not seen during training." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 689, + 297, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 297, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 297, + 700 + ], + "type": "text", + "content": "with object trajectories with better 3D models in the future." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 700, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 700, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 700, + 296, + 712 + ], + "type": "text", + "content": "Our data pipeline is modular and can also incorporate better" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 314, + 689, + 473, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 689, + 473, + 700 + ], + "spans": [ + { + "bbox": [ + 314, + 689, + 473, + 700 + ], + "type": "text", + "content": "hand & segmentation models in future." + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7033" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 168 + ], + "type": "text", + "content": "Acknowledgments: This material is based upon work supported by an NSF CAREER Award (IIS2143873), USDA-NIFA AIFARMS National AI Institute (USDA#2020-67021-32799) and NSF (IIS-2007035). Part of the work was done when Aditya Prakash was an intern at Microsoft. We acknowledge compute support by NVIDIA Academic Hardware Grant and a DURIP grant. We thank Shaowei Liu and Arjun Gupta for feedback on the draft." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 177, + 115, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 177, + 115, + 190 + ], + "spans": [ + { + "bbox": [ + 56, + 177, + 115, + 190 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 196, + 297, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 196, + 297, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 196, + 297, + 251 + ], + "spans": [ + { + "bbox": [ + 61, + 196, + 297, + 251 + ], + "type": "text", + "content": "[1] Shikhar Bahl, Russell Mendonca, Lili Chen, Unnat Jain, and Deepak Pathak. Affordances from human videos as a versatile representation for robotics. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 251, + 296, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 251, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 61, + 251, + 296, + 316 + ], + "type": "text", + "content": "[2] Prithviraj Banerjee, Sindi Shkodrani, Pierre Moulon, Shreyas Hampali, Fan Zhang, Jade Fountain, Edward Miller, Selen Basol, Richard Newcombe, Robert Wang, Jakob Julian Engel, and Tomas Hodan. Introducing hot3d: An egocentric dataset for 3d hand and object tracking. arXiv, 2406.09598, 2024. 2, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 318, + 296, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 318, + 296, + 360 + ], + "spans": [ + { + "bbox": [ + 61, + 318, + 296, + 360 + ], + "type": "text", + "content": "[3] Jona Braun, Sammy Joe Christen, Muhammed Kocabas, Emre Aksan, and Otmar Hilliges. Physically plausible full-body hand-object interaction synthesis. arXiv, 2309.07907, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 361, + 296, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 361, + 296, + 406 + ], + "spans": [ + { + "bbox": [ + 62, + 361, + 296, + 406 + ], + "type": "text", + "content": "[4] Junuk Cha, Jihyeon Kim, Jae Shin Yoon, and Seungryul Baek. Text2hoi: Text-guided 3d motion generation for handobject interaction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 406, + 296, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 406, + 296, + 449 + ], + "spans": [ + { + "bbox": [ + 62, + 406, + 296, + 449 + ], + "type": "text", + "content": "[5] Matthew Chang, Aditya Prakash, and Saurabh Gupta. Look ma, no hands! agent-environment factorization of egocentric videos. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 449, + 296, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 449, + 296, + 494 + ], + "spans": [ + { + "bbox": [ + 62, + 449, + 296, + 494 + ], + "type": "text", + "content": "[6] Tianyi Cheng, Dandan Shan, Ayda Sultan Hassen, Richard Ely Locke Higgins, and David Fouhey. Towards a richer 2d understanding of hands at scale. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 2, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 494, + 296, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 494, + 296, + 548 + ], + "spans": [ + { + "bbox": [ + 62, + 494, + 296, + 548 + ], + "type": "text", + "content": "[7] Sammy Christen, Shreyas Hampali, Fadime Sener, Edoardo Remelli, Tomas Hodan, Eric Sauser, Shugao Ma, and Bugra Tekin. Diffh2o: Diffusion-based synthesis of hand-object interactions from textual descriptions. arXiv, 2403.17827, 2024. 2, 3, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 548, + 296, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 548, + 296, + 613 + ], + "spans": [ + { + "bbox": [ + 62, + 548, + 296, + 613 + ], + "type": "text", + "content": "[8] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Sanja Fidler, Antonino Furnari, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, et al. Scaling egocentric vision: The epic-kitchens dataset. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 614, + 296, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 614, + 296, + 669 + ], + "spans": [ + { + "bbox": [ + 62, + 614, + 296, + 669 + ], + "type": "text", + "content": "[9] Ahmad Darkhalil, Dandan Shan, Bin Zhu, Jian Ma, Amlan Kar, Richard Higgins, Sanja Fidler, David Fouhey, and Dima Damen. Epic-kitchens visor benchmark: Video segmentations and object relations. In NeurIPS Track on Datasets and Benchmarks, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 670, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 296, + 714 + ], + "type": "text", + "content": "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "text", + "content": "is worth 16x16 words: Transformers for image recognition at scale. In Proceedings of the International Conference on Learning Representations (ICLR), 2021. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 106, + 555, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 106, + 555, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 106, + 555, + 171 + ], + "type": "text", + "content": "[11] Zicong Fan, Omid Taheri, Dimitrios Tzionas, Muhammed Kocabas, Manuel Kaufmann, Michael J. Black, and Otmar Hilliges. ARCTIC: A dataset for dexterous bimanual handobject manipulation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3, 5, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 172, + 555, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 172, + 555, + 270 + ], + "spans": [ + { + "bbox": [ + 316, + 172, + 555, + 270 + ], + "type": "text", + "content": "[12] Zicong Fan, Takehiko Ohkawa, Linlin Yang, Nie Lin, Zhishan Zhou, Shihao Zhou, Jiajun Liang, Zhong Gao, Xuanyang Zhang, Xue Zhang, Fei Li, Zheng Liu, Feng Lu, Karim Abou Zeid, Bastian Leibe, Jeongwan On, Seungryul Baek, Aditya Prakash, Saurabh Gupta, Kun He, Yoichi Sato, Otmar Hilliges, Hyung Jin Chang, and Angela Yao. Benchmarks and challenges in pose estimation for egocentric hand interactions with objects. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 271, + 555, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 271, + 555, + 315 + ], + "spans": [ + { + "bbox": [ + 316, + 271, + 555, + 315 + ], + "type": "text", + "content": "[13] Anindita Ghosh, Rishabh Dabral, Vladislav Golyanik, Christian Theobalt, and Philipp Slusallek. Imos: Intent-driven full-body motion synthesis for human-object interactions. EuroGraphics, 42(2):1-12, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 316, + 555, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 555, + 336 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 555, + 336 + ], + "type": "text", + "content": "[14] J.J. Gibson. The ecological approach to visual perception, 1979. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 338, + 554, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 338, + 554, + 403 + ], + "spans": [ + { + "bbox": [ + 317, + 338, + 554, + 403 + ], + "type": "text", + "content": "[15] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 403, + 555, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 403, + 555, + 448 + ], + "spans": [ + { + "bbox": [ + 317, + 403, + 555, + 448 + ], + "type": "text", + "content": "[16] Chuan Guo, Yuxuan Mu, Muhammad Gohar Javed, Sen Wang, and Li Cheng. Momask: Generative masked modeling of 3d human motions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 449, + 555, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 449, + 555, + 491 + ], + "spans": [ + { + "bbox": [ + 317, + 449, + 555, + 491 + ], + "type": "text", + "content": "[17] Shreyas Hampali, Mahdi Rad, Markus Oberweger, and Vincent Lepetit. Honnotate: A method for 3d annotation of hand and object poses. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 492, + 554, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 492, + 554, + 547 + ], + "spans": [ + { + "bbox": [ + 317, + 492, + 554, + 547 + ], + "type": "text", + "content": "[18] Shreyas Hampali, Sayan Deb Sarkar, Mahdi Rad, and Vincent Lepetit. Keypoint transformer: Solving joint identification in challenging hands and object interactions for accurate 3d pose estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 548, + 555, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 548, + 555, + 601 + ], + "spans": [ + { + "bbox": [ + 317, + 548, + 555, + 601 + ], + "type": "text", + "content": "[19] Mohamed Hassan, Partha Ghosh, Joachim Tesch, Dimitrios Tzionas, and Michael J. Black. Populating 3d scenes by learning human-scene interaction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 602, + 555, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 602, + 555, + 646 + ], + "spans": [ + { + "bbox": [ + 316, + 602, + 555, + 646 + ], + "type": "text", + "content": "[20] Di Huang, Xiaopeng Ji, Xingyi He, Jiaming Sun, Tong He, Qing Shuai, Wanli Ouyang, and Xiaowei Zhou. Reconstructing hand-held objects from monocular video. In ACM Transactions on Graphics, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 647, + 555, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 555, + 689 + ], + "type": "text", + "content": "[21] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. In Proceedings of the International Conference on Learning Representations (ICLR), 2017. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 690, + 554, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 690, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 690, + 554, + 714 + ], + "type": "text", + "content": "[22] Jaewoo Jeong, Daehee Park, and Kuk-Jin Yoon. Multi-agent long-term 3d human pose forecasting via interaction-aware" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7034" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 105 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 105 + ], + "type": "text", + "content": "trajectory conditioning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 5, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 296, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 296, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 296, + 150 + ], + "type": "text", + "content": "[23] Hanwen Jiang, Shaowei Liu, Jiashun Wang, and Xiaolong Wang. Hand-object contact consistency reasoning for human grasps generation. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 296, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 296, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 296, + 205 + ], + "type": "text", + "content": "[24] Taein Kwon, Bugra Tekin, Jan Stühmer, Federica Bogo, and Marc Pollefeys. H2o: Two hands manipulating objects for first person interaction recognition. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 205, + 296, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 205, + 296, + 259 + ], + "spans": [ + { + "bbox": [ + 56, + 205, + 296, + 259 + ], + "type": "text", + "content": "[25] Gen Li, Varun Jampani, Deqing Sun, and Laura Sevilla-Lara. LOCATE: localize and transfer object parts for weakly supervised affordance grounding. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 261, + 296, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 296, + 315 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 296, + 315 + ], + "type": "text", + "content": "[26] Quanzhou Li, Jingbo Wang, Chen Change Loy, and Bo Dai. Task-oriented human-object interactions generation with implicit neural representations. In Proceedings of the IEEE Winter Conference on Applications of Computer Vision (WACV), 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 316, + 296, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 296, + 360 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 296, + 360 + ], + "type": "text", + "content": "[27] Shaowei Liu, Yang Zhou, Jimei Yang, Saurabh Gupta, and Shenlong Wang. Contactgen: Generative contact modeling for grasp generation. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 361, + 296, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 296, + 425 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 296, + 425 + ], + "type": "text", + "content": "[28] Yunze Liu, Yun Liu, Che Jiang, Kangbo Lyu, Weikang Wan, Hao Shen, Boqiang Liang, Zhoujie Fu, He Wang, and Li Yi. HOI4D: A 4d egocentric dataset for category-level human-object interaction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 426, + 296, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 296, + 482 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 296, + 482 + ], + "type": "text", + "content": "[29] Takehiko Ohkawa, Kun He, Fadime Sener, Tomas Hodan, Luan Tran, and Cem Keskin. Assemblyhands: Towards egocentric activity understanding via 3d hand pose estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 12999-13008, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 482, + 296, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 482, + 296, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 482, + 296, + 536 + ], + "type": "text", + "content": "[30] Georgios Pavlakos, Dandan Shan, Iija Radosavovic, Angjoo Kanazawa, David Fouhey, and Jitendra Malik. Reconstructing hands in 3d with transformers. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 537, + 296, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 296, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 296, + 581 + ], + "type": "text", + "content": "[31] Aditya Prakash, Matthew Chang, Matthew Jin, and Saurabh Gupta. 3d reconstruction of objects in hands without real world 3d supervision. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 582, + 296, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 296, + 624 + ], + "type": "text", + "content": "[32] Aditya Prakash, Arjun Gupta, and Saurabh Gupta. Mitigating perspective distortion-induced shape ambiguity in image crops. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 625, + 296, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 296, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 296, + 669 + ], + "type": "text", + "content": "[33] Aditya Prakash, Ruisen Tu, Matthew Chang, and Saurabh Gupta. 3d hand pose estimation in everyday egocentric images. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 670, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 296, + 713 + ], + "type": "text", + "content": "[34] Sergey Prokudin, Christoph Lassner, and Javier Romero. Efficient learning on point clouds with basis point sets. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2019. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 72, + 555, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 555, + 148 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 555, + 148 + ], + "type": "text", + "content": "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the International Conference on Machine Learning (ICML), 2021. 3, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 150, + 555, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 150, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 150, + 555, + 228 + ], + "type": "text", + "content": "[36] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dollár, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos. arXiv, 2408.00714, 2024. 2, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 229, + 555, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 229, + 555, + 272 + ], + "spans": [ + { + "bbox": [ + 316, + 229, + 555, + 272 + ], + "type": "text", + "content": "[37] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: modeling and capturing hands and bodies together. ACM Transactions on Graphics, 36(6):245:1-245:17, 2017. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 274, + 555, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 274, + 555, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 274, + 555, + 308 + ], + "type": "text", + "content": "[38] Javier Romero, Dimitrios Tzionas, and Michael J Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Transactions on Graphics (ToG), 2017. 3, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 308, + 555, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 555, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 555, + 352 + ], + "type": "text", + "content": "[39] Rómer Rosales and Stan Sclaroff. Combining generative and discriminative models in a framework for articulated pose estimation. International Journal of Computer Vision, 67: 251-276, 2006. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 353, + 555, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 353, + 555, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 353, + 555, + 398 + ], + "type": "text", + "content": "[40] Dandan Shan, Jiaqi Geng, Michelle Shu, and David F Fouhey. Understanding human hands in contact at internet scale. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 399, + 555, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 555, + 453 + ], + "type": "text", + "content": "[41] Rajinder S Sodhi, Brett R Jones, David Forsyth, Brian P Bailey, and Giuliano Maciocci. Bethere: 3d mobile collaboration with spatial input. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, pages 179-188, 2013. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 455, + 555, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 455, + 555, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 455, + 555, + 487 + ], + "type": "text", + "content": "[42] Mohan Kumar Srirama, Sudeep Dasari, Shikhhar Bahl, and Abhinav Gupta. HRP: human affordances for robotic pretraining. arxiv, 2407.18911, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 488, + 555, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 488, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 488, + 555, + 533 + ], + "type": "text", + "content": "[43] Omid Taheri, Nima Ghorbani, Michael J Black, and Dimitrios Tzionas. GRAB: A dataset of whole-body human grasping of objects. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2, 3, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 534, + 555, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 534, + 555, + 588 + ], + "spans": [ + { + "bbox": [ + 316, + 534, + 555, + 588 + ], + "type": "text", + "content": "[44] Omid Taheri, Yi Zhou, Dimitrios Tzionas, Yang Zhou, Duygu Ceylan, Soren Pirk, and Michael J. Black. GRIP: Generating interaction poses using latent consistency and spatial cues. In Proceedings of the International Conference on 3D Vision (3DV), 2024. 2, 3, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 590, + 555, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 555, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 555, + 634 + ], + "type": "text", + "content": "[45] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit Haim Bermano. Human motion diffusion model. In Proceedings of the International Conference on Learning Representations (ICLR), 2023. 5, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 635, + 555, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 555, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 555, + 690 + ], + "type": "text", + "content": "[46] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In Proceedings of the International Conference on Machine Learning (ICML), 2021. 3, 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 691, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 555, + 713 + ], + "type": "text", + "content": "[47] Shashank Tripathi, Agniv Chatterjee, Jean-Claude Passy, Hongwei Yi, Dimitrios Tzionas, and Michael J. Black. DECO:" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7035" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 498 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "text", + "content": "dense estimation of 3d human-scene contact in the wild. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2023. 3, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 296, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 296, + 140 + ], + "type": "text", + "content": "[48] Aäron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu. Neural discrete representation learning. In Advances in Neural Information Processing Systems (NeurIPS), 2017. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 296, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 296, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 296, + 194 + ], + "type": "text", + "content": "[49] Sirui Xu, Zhengyuan Li, Yu-Xiong Wang, and Liang-Yan Gui. Interdiff: Generating 3d human-object interactions with physics-informed diffusion. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 295, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 295, + 250 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 295, + 250 + ], + "type": "text", + "content": "[50] Yuhang Yang, Wei Zhai, Hongchen Luo, Yang Cao, Jiebo Luo, and Zheng-Jun Zha. Grounding 3d object affordance from 2d interactions in images. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 252, + 296, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 252, + 296, + 296 + ], + "spans": [ + { + "bbox": [ + 56, + 252, + 296, + 296 + ], + "type": "text", + "content": "[51] Yufei Ye, Abhinav Gupta, and Shubham Tulsiani. What's in your hands? 3D reconstruction of generic objects in hands. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 297, + 296, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 296, + 341 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 296, + 341 + ], + "type": "text", + "content": "[52] Yufei Ye, Poorvi Hebbar, Abhinav Gupta, and Shubham Tulsiani. Diffusion-guided reconstruction of everyday handobject interaction clips. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2023. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 342, + 296, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 342, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 56, + 342, + 296, + 396 + ], + "type": "text", + "content": "[53] Yufei Ye, Xueting Li, Abhinav Gupta, Shalini De Mello, Stan Birchfield, Jiaming Song, Shubham Tulsiani, and Sifei Liu. Affordance diffusion: Synthesizing hand-object interactions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 398, + 296, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 398, + 296, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 398, + 296, + 453 + ], + "type": "text", + "content": "[54] Yufei Ye, Abhinav Gupta, Kris Kitani, and Shubham Tulsiani. G-hop: Generative hand-object prior for interaction reconstruction and grasp synthesis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 3, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 454, + 296, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 454, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 454, + 296, + 498 + ], + "type": "text", + "content": "[55] Hui Zhang, Sammy Joe Christen, Zicong Fan, Luocheng Zheng, Jemin Hwangbo, Jie Song, and Otmar Hilliges. Artigrasp: Physically plausible synthesis of bi-manual dexterous grasping and articulation. arXiv, 2309.03891, 2023. 2" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "7036" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/How to Merge Your Multimodal Models Over Time_/76e2370a-2b50-436a-9bdc-9b8af8c8e7ca_content_list.json b/2025/How to Merge Your Multimodal Models Over Time_/76e2370a-2b50-436a-9bdc-9b8af8c8e7ca_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9ca76ed44381c410fdc444a7306690c158873d55 --- /dev/null +++ b/2025/How to Merge Your Multimodal Models Over Time_/76e2370a-2b50-436a-9bdc-9b8af8c8e7ca_content_list.json @@ -0,0 +1,1747 @@ +[ + { + "type": "text", + "text": "How to Merge Your Multimodal Models Over Time?", + "text_level": 1, + "bbox": [ + 212, + 130, + 743, + 152 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d3ca2ac2a62466f29cd4d95337275d8e2a7cb93585bf35347106ea6962047f39.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 746, + 125, + 782, + 150 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sebastian Dziadzio*¹ Vishaal Udandarao*¹,² Karsten Roth*¹,³ Ameya Prabhu°¹ Zeynep Akata³† Samuel Albanie† Matthias Bethge¹†", + "bbox": [ + 156, + 179, + 834, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Tübingen AI Center, University of Tübingen $^{2}$ University of Cambridge $^{3}$ Munich Center for ML, Technical University of Munich", + "bbox": [ + 112, + 218, + 880, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Model merging combines multiple \"expert\" models—finetuned from a base foundation model on diverse tasks and domains—into a single, more capable model. However, most existing model merging approaches assume that all experts are available simultaneously. In reality, new tasks and domains emerge progressively over time, requiring strategies to integrate the knowledge of expert models as they become available: a process we call temporal model merging. The temporal dimension introduces unique challenges not addressed in prior work, raising new questions such as: when training for a new task, should the expert model start from the merged past experts or from the original base model? Should we merge all models at each time step? Which merging techniques are best suited for temporal merging? Should different strategies be used to initialize the training and deploy the model? To answer these questions, we propose a unified framework called TIME—Temporal Integration of Model Expertise—which defines temporal model merging across three axes: (1) Initialization Phase, (2) Deployment Phase, and (3) Merging Technique. Using TIME, we study temporal model merging across model sizes, compute budgets, and learning horizons on the FoMo-in-Flux benchmark. Our comprehensive suite of experiments across TIME allows us to uncover key insights for temporal model merging, offering a better understanding of current challenges and best practices for effective temporal model merging. Our code is available here.", + "bbox": [ + 88, + 300, + 485, + 709 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 753, + 220, + 768 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Foundation models consolidate a wide range of capabilities and knowledge into a single, large model [5]. Consequently, model merging [24, 34, 63, 87] has emerged as a key technique for unifying multiple task-specific fine-tuned models", + "bbox": [ + 89, + 779, + 482, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3c8338fe5b160b2a80269142625cfe5aefaa614fa03a02a07f4f5debd3b41ad5.jpg", + "image_caption": [ + "Figure 1. Temporal Model Merging generalizes standard model merging (yellow), which merges multiple trained experts just once, in a single step. Our systematic study of this realistic multi-step regime reveals that initialization and deployment strategies dominate the importance of the single-step weight merging strategy." + ], + "image_footnote": [], + "bbox": [ + 516, + 267, + 906, + 398 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "derived from a shared base into a single model, effectively preserving the strengths of each specialized expert.", + "bbox": [ + 511, + 498, + 906, + 529 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Current model merging approaches typically assume a fixed base model that is fine-tuned independently on $k$ diverse tasks and domains to produce a set of independent experts [17, 24, 34, 64, 89], which are then merged simultaneously. Research in this field has therefore focused primarily on improving merging techniques for larger or structurally differing $k$ -sets, exploring the impact of the diversity and scale of finetuning domains, tasks and experts [27, 69, 86, 89].", + "bbox": [ + 509, + 531, + 908, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, the world is constantly evolving, leading to continuous shifts over data distributions, domains, and tasks, with new concepts emerging [45, 68] that may have been insufficiently covered during large-scale pretraining [20, 22, 32, 45, 57, 65, 66, 81-83, 98]. This dynamic nature of real-world applications motivates a hitherto missing systematic exploration into temporal model merging (see Fig. 1) to better understand model merging along an additional, overlooked [13, 100] axis: time. Specifically, in this work, we ask: (1) Is temporal model merging significantly influenced by the choice of initialization for expert training? (2) When evaluated over time, which model merging techniques emerge as most suitable? (3) Should merging strategies differ between initialization and model deployment? To answer these questions, we propose a unified framework", + "bbox": [ + 511, + 674, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*equal contribution, random order †equal supervision, random order, o core contributor. Correspondence to: vishaal.udarao@bethgelab.org or sebastian.dziadzio@bethgelab.org", + "bbox": [ + 89, + 851, + 482, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20479", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "for studying temporal model merging—TIME (Temporal Integration of Model Expertise)—structured around three key axes spanning the design space of temporal merging solutions (as shown in Fig. 2):", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Initialization Phase. As expert models are trained and created continuously over time, initialization becomes a crucial design element.", + "2. Deployment Phase. After training an expert on each task, the next step is to deploy a suitable final model. For temporal model merging, this process has to account for a varying number of past expert models and deployed variants, striking a balance between retaining past knowledge and incorporating new task-specific knowledge.", + "3. Model Merging Techniques. Merging literature has explored a range of approaches for simultaneous merging, from simple weight-averaging and interpolation [24, 63, 68, 76] to more complex methods involving weight and candidate selection and re-weighting [11, 42, 43, 89, 95]. To apply these methods temporally, it is crucial to understand how they perform with varying numbers of merge candidates $(k_{t})$ and shifting distributions." + ], + "bbox": [ + 89, + 155, + 482, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Using our TIME framework, we position existing model merging approaches along each key axis and conduct a systematic study of model merging over time. For this, we leverage the multimodal FoMo-in-Flux [68] benchmark which includes 63 datasets with well-documented sequential properties, enabling a thorough investigation of temporal model merging under practical compute constraints, as proposed in Roth et al. [68]. Our experiments systematically explore different merging techniques, initialization, and deployment strategies, providing several key insights:", + "bbox": [ + 89, + 425, + 482, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Key Insights for Temporal Model Merging", + "text_level": 1, + "bbox": [ + 117, + 590, + 401, + 607 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[A] Accounting for time is crucial. Standard \"off-line\" model merging techniques do not generalize well to the temporal merging setting (Sec. 4.2).", + "[B] Complex merging techniques matter little. Choosing sophisticated merging techniques beyond simple weighted averaging provides at best marginal benefits for temporal model merging, especially for long task sequences (Sec. 4.5).", + "[C] Initialization and deployment are critical. Choosing how to select and combine available weights before and after each task $t$ is most important for temporal model merging (Sec. 4.4).", + "[D] Temporal model merging scales well. Larger models or compute budgets allows greater benefits from temporal merging. Scaling enables temporal model merging to even outperform the multitask model, trained on all tasks at once (Sec. 4.6)." + ], + "bbox": [ + 114, + 618, + 455, + 885 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 89, + 663, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Model Merging. We provide a short overview of the model merging literature, detailed in these excellent surveys [90, 92]. While both model aggregation through distillation [10, 67] and averaging checkpoints during training [29, 35, 69] have shown success, the requirement of additional compute limits practicability of these methods [55]. Instead, recent work [18, 21, 24, 25, 41, 47, 49, 50, 52, 59, 69, 71, 72, 75, 77, 85–88, 90, 93, 96, 99] has shown the effectiveness of training-free weight averaging and interpolation of fine-tuned expert models to produce an improved base model, benefiting from (linear) mode connectivity in models fine-tuned from a single pre-trained checkpoint [1, 14, 26, 48, 61]. These insights have been extended into weight-averaged reward models [62], policy models [60] with spherical interpolation, and KL-constrained RLHF [19, 36, 39, 46]. Works such as Fisher-Merge [43], TIES [89], RegMean [28], MATS [78], DELLA [12], DARE [95], Broadcrumbs [11], evolutionary merging [2] and MagMax [42] have explored merging strategies beyond simple interpolation to determine which weights should be merged across expert models. These methods have different benefits for in- and out-of-distribution generalization across domains [79], though recently they have been shown to perform similarly at scale [91]. Additionally, some works have explored the initialization dimension for effectively merging models [9, 13, 42, 100]. In this work, we propose a unifying framework for temporal merging and conduct the most comprehensive study of this topic to date.", + "bbox": [ + 511, + 114, + 906, + 537 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Continual Pretraining extends beyond standard Continual Learning [54, 68], focusing on large-scale model updates starting from pretrained foundation models [16, 20, 23, 56, 68] and addressing more complex and substantial update tasks [6, 7, 16, 37, 38, 68]. There has been limited exploration into using model merging for continual pretraining [3, 42, 68, 76], as most prior works focus on training strategies including regularization objectives and learning-rate schedules [16, 23, 33, 44, 51, 55, 68, 74, 80, 94]. We keep the training strategy fixed, and provide an in-depth exploration beyond simple merging techniques.", + "bbox": [ + 511, + 542, + 908, + 709 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Design Space of Temporal Model Merging", + "text_level": 1, + "bbox": [ + 511, + 723, + 890, + 742 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Notation. Throughout this work, we use $t$ to refer to a given task at time $t$ . Full model parameterization is denoted by $\\theta$ , with the following key instantiations: $\\theta_t$ represents model weights at task $t$ , while $\\theta_t^I$ , $\\theta_t^S$ , and $\\theta_t^O$ denote weights used for initialization, saved weight checkpoints at task $t$ (i.e., the trained expert models), and the output deployed model weights, respectively. Note that while standard model merging considers model weights as elements of a fixed set $\\{\\theta_k\\}_{k=1}^K$ , temporal model merging organizes them along the time axis $\\theta_t$ .", + "bbox": [ + 511, + 750, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20480", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a42037083fab748e2b76f8611cfef4e3d1df23ad74a56b9313e8bf709ebc64ef.jpg", + "image_caption": [ + "Figure 2. Design Space of Temporal Model Merging through TIME. We showcase our framework for the per-task pipeline of temporal model merging over multiple tasks: At each task $t$ , we first initialize the current checkpoint to start training from, $\\theta_t^i$ , by using one or more previously stored checkpoints from previous tasks, either directly or by merging them. We train $\\theta_t^i$ on current task data $\\mathcal{D}_t$ to yield the current task checkpoint $\\theta_t^s$ , which is inserted into the checkpoint buffer. Finally, to produce the output model, $\\theta_t^o$ , we either merge previously stored checkpoints from the buffer or use them directly. The entire framework is depicted in the pseudo-code on the right panel." + ], + "image_footnote": [], + "bbox": [ + 93, + 85, + 903, + 340 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Temporal Model Merging through TIME", + "text_level": 1, + "bbox": [ + 89, + 429, + 437, + 446 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Standard model merging is typically performed offline, after all experts have been trained to convergence [15, 87, 89, 91]. In contrast, model merging in continual pretraining is generally done sequentially, using past checkpoints [68, 76]. Both approaches are specific instances of our more general temporal merging framework, TIME, which defines temporal merging along three key axes: initialization of each expert, merging for deployment at step $t$ , and merging techniques $f_{\\mathrm{merge}}$ applied over time:", + "bbox": [ + 89, + 452, + 483, + 589 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Axis 1: Initialization", + "text_level": 1, + "bbox": [ + 89, + 595, + 240, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As expert models are created continuously over time, initialization becomes a crucial choice. Unlike model merging at a single point in time, the number of potential starting points grows exponentially over time as new experts are created. This raises the question: should starting points for each time step be derived from the base weights (as in traditional merging), from a merged combination of previous experts, or from most recent weights, as commonly done in continual pretraining [68, 76]? In this work, we study the following initialization protocols at time step $t$ for TIME:", + "bbox": [ + 89, + 614, + 483, + 766 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- initZS, which consistently initializes with the base zero-shot model weights $\\theta_0$ at each timestep $t$ .", + "- initFT, which for step $t$ always initializes with the latest available finetuned model weights $\\theta_{t - 1}^{S}$ .", + "- initEMA, which computes an unrolled exponential moving average merge over all previously seen expert models $\\{\\theta_{t'}^S\\}_{1,\\dots,t - 1}$ following the equation:" + ], + "bbox": [ + 91, + 767, + 483, + 873 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {t ^ {\\prime}} ^ {\\mathrm {E M A}} = f _ {\\text {m e r g e}} \\left(\\theta_ {t ^ {\\prime} - 1} ^ {\\mathrm {E M A}}, \\theta_ {t ^ {\\prime} - 1} ^ {S}, \\mathcal {F}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 186, + 883, + 482, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "with merging hyperparameters $\\mathcal{F}$ . Consequently, the initialization weights are given as $\\theta_t^I = \\theta_t^{\\mathrm{EMA}}$ .", + "bbox": [ + 526, + 430, + 906, + 460 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Axis 2: Deployment", + "text_level": 1, + "bbox": [ + 513, + 470, + 658, + 486 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With each update iteration and expert training phase $t$ , a decision must be made on the final model to deploy, determining which weights to present for downstream use. In continual pretraining, the trained model $\\theta_t^S$ is deployed directly. In contrast, standard model merging applies a merging technique $f_{\\text{merge}}$ to a fixed set of $k$ expert models. Temporal model merging, however, must account for both previously deployed models and the growing number of expert models available over time. Unlike standard merging, where $k$ remains constant, the number of experts to merge increases with each step. As a result, temporal merging introduces the idea of weighted combinations, balancing recent updates with retained past knowledge to achieve adaptability and stability—both critical for effective continual learning [31, 54, 55, 68, 97]. In this work, we study three strategies for model deployment:", + "bbox": [ + 511, + 491, + 906, + 733 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- deployyFT, which at step $t$ deploys the latest finetuned expert model, i.e. $\\theta_s^O = \\theta_t^S$ .", + "- deployEMA, which computes an unrolled exponential moving average merge over all expert models, i.e. $\\theta_t^O = \\theta_{t+1}^{EMA}$ following Eq. (1).", + "- deployALL, which applies a merging technique $f_{\\text{merge}}$ over all previously computed expert models $\\{\\theta_{t'}^S\\}_{1}^{t - 1}$ ." + ], + "bbox": [ + 511, + 734, + 903, + 840 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Axis 3: Merging Techniques", + "text_level": 1, + "bbox": [ + 513, + 849, + 714, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "At each point in time, for both initialization and deployment, merging technique $f_{\\text{merge}}$ defines how to combine", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20481", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/0a81180b946c8cb3fe395f9c53e6b9bcfd9897f8b4f7e16b414491ee7ad59a40.jpg", + "table_caption": [ + "Table 1. Comparison of model merging techniques." + ], + "table_footnote": [], + "table_body": "
MethodSparsificationConsensusScaling
Weight averaging [63, 86]XLinear Int.Weight coeff.
SLERP [60]XSpherical Int.Weight coeff.
Task Arithmetic [25]XLinear Int.Scaling factor
MagMax [42]XMax. MagnitudeScaling factor
TIES [89]Top-kSign AgreementScaling factor
DARE-TIES [95]RandomSign AgreementScaling factor
Breadcrumbs-TIES [11]Top/Bottom-kSign AgreementScaling factor
Model Stock [27]XGeometricAdaptive ratio
LiNeS [84]XXLayer weights
", + "bbox": [ + 94, + 114, + 482, + 233 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the available expert models and checkpoints. In this work, we study nine different variants in total, shown in Tab. 1. Denoting the number of models to merge at timestep $t$ as $M_{t}$ (with $t = 0$ and $M_{t} = M$ for standard model merging), we can define these methods as follows:", + "bbox": [ + 89, + 250, + 482, + 325 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Weight Averaging [24, 63, 68, 76, 87] simply employs a uniformly weighted, element-wise average over all models $\\theta_{t,i}$ , resulting in a merge function $f_{\\mathrm{merge}}^{\\mathrm{WA}}$ :", + "bbox": [ + 89, + 339, + 483, + 388 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {t} = \\frac {1}{M _ {t}} \\sum_ {i} \\theta_ {t, i}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 223, + 397, + 482, + 433 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SLERP [60, 73] assumes weights to live on a hypersphere, and consequently conducts interpolation along a curved path connecting weight entries. In particular, for two models $\\theta_{t,1}$ and $\\theta_{t,2}$ deriving from some base weight $\\theta_{t-1}$ and the corresponding task vectors [25] $\\delta_{t,i} = \\theta_{t,i} - \\theta_{t-1}$ , SLERP with interpolation weight $\\lambda$ is defined as", + "bbox": [ + 89, + 443, + 483, + 536 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {t} = \\theta_ {t - 1} + \\frac {\\sin (1 - \\lambda) \\Omega_ {1 , 2}}{\\sin \\Omega_ {1 , 2}} \\cdot \\delta_ {t, 1} + \\frac {\\sin \\lambda \\Omega_ {1 , 2}}{\\sin \\Omega_ {1 , 2}} \\cdot \\delta_ {t, 2} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 545, + 482, + 579 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with $\\Omega_{1,2}$ being the angle between task vectors $\\delta_{t,1}$ and $\\delta_{t,2}$ . We denote the corresponding merge function $f_{\\mathrm{merge}}^{\\mathrm{SLERP}}$", + "bbox": [ + 89, + 589, + 482, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Task Arithmetic [25] defines the merge as a function over task vectors $\\delta_{t,i} = \\theta_{t,i} - \\theta_{t-1}$ for each weight $\\theta_{t,i}$ fine-tuned from $\\theta_{t-1}$ . This introduces a simple merge formalism $f_{\\text{merge}}^{\\text{TA}}$ for weighted parameter averaging with a scale $\\lambda$ :", + "bbox": [ + 89, + 619, + 483, + 679 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {t} = \\theta_ {t - 1} + \\lambda \\frac {1}{M _ {t}} \\sum_ {i} \\delta_ {t, i} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 689, + 482, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "TIES [89] builds on the task arithmetic formalism through controlled pruning of task vector entries with low magnitude. Moreover, the sign for each final merged parameter is set based on the sign of the highest total magnitude across the merge candidates. The final update follows basic task arithmetic, only for entries with matching signs. We refer to the respective merge function as $f_{\\text{merge}}^{\\text{TIES}}$ .", + "bbox": [ + 89, + 744, + 482, + 853 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "DARE [95] is a similar extension of task arithmetic, but instead of targeted pruning, it randomly zeroes out task vector entries using a random mask $Z_{i} \\sim \\mathrm{Bernoulli}(p)$ and", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "masking probability $p$ . Final task vector values for $f_{\\mathrm{merge}}^{\\mathrm{DARE}}$ are then rescaled based on $p$ :", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\delta_ {t, i} ^ {\\mathrm {D A R E}} = \\frac {(1 - Z _ {i}) \\delta_ {t , i}}{1 - p}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 632, + 128, + 903, + 162 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Model Stock [27] provides a geometric extension of simple weight averaging as done in Model Soup [86] by incorporating base weights $\\theta_{t-1}$ into the merging process. Given finetuned weights $\\theta_{t,1}$ and $\\theta_{t,2}$ , the Model Stock merge $f_{\\text{merge}}^{\\text{Stock}}$ is defined as follows:", + "bbox": [ + 511, + 180, + 906, + 255 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {t} = \\frac {2 \\cdot \\cos \\Omega_ {1 , 2}}{1 + \\Omega_ {1 , 2}} \\cdot \\left(\\theta_ {t, 2} - \\theta_ {t, 1}\\right) + \\left(1 - \\frac {2 \\cdot \\cos \\Omega_ {1 , 2}}{1 + \\cos \\Omega_ {1 , 2}}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 261, + 903, + 296 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "utilizing angle $\\Omega_{1,2}$ between task vectors $\\delta_{t,1}$ and $\\delta_{t,2}$ .", + "bbox": [ + 511, + 301, + 872, + 318 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Breadcrumbs [11] deploys another variation on task arithmetic for model merging. In particular, for a given task vector $\\delta_{t,i}$ , extreme left and right tails of the absolute magnitude distribution in $\\delta_{t,i}$ are zeroed out with left and right thresholds $\\beta$ and $\\gamma$ . The modified task vectors $\\delta_{t,i}^{\\mathrm{Bread}}$ are then applied on base weights $\\theta_{t-1}$ following the task arithmetic setup, and giving $f_{\\mathrm{merge}}^{\\mathrm{Bread}}$ .", + "bbox": [ + 511, + 320, + 903, + 428 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "MagMax [42] also uses task vectors—given multiple task vectors $\\delta_{t,i}$ (with increments possible along both time $t$ and count axis $i$ ), the final task vector $\\delta_t$ is yielded through maximum magnitude entry selection; copying the largest magnitude entries across all $\\{\\delta_{t,i}\\}$ into $\\delta_t$ , giving $f_{\\mathrm{merge}}^{\\mathrm{Max}}$ .", + "bbox": [ + 511, + 430, + 905, + 507 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "LiNeS [84], for Layer-increasing Network Scaling, scales weight updates based on their respective layer depth enabling early layers to remain close to original pretraining weights (cf. Neyshabur et al. [48]). Given task vectors $\\delta_{t,i}$ , now broken down across model layers $\\delta_{t,i}^{l}$ with $l\\in [1,\\dots,L]$ and $L$ the number of layers, LiNeS follows the base task arithmetic merging formalism, but updates task vectors as", + "bbox": [ + 511, + 508, + 905, + 614 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\delta_ {t, i} ^ {\\mathrm {L i N e S}} = \\operatorname {c o n c a t} \\left(\\lambda^ {l = 1} \\delta_ {t, i} ^ {l = 1}, \\dots , \\lambda^ {l = L} \\delta_ {t, i} ^ {l = L}\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 622, + 903, + 642 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with layer-scaled interpolation weights $\\lambda^l = \\alpha +\\beta \\frac{l - 1}{L - 1}$ and hyperparameters $\\alpha ,\\beta$ giving $f_{\\mathrm{merge}}^{\\mathrm{Lines}}$", + "bbox": [ + 511, + 650, + 905, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Complete Temporal Model Merging Pipeline", + "text_level": 1, + "bbox": [ + 511, + 695, + 890, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Incorporating all three axes of temporal model merging, we define a five-stage update pipeline for each task $t$ (see Fig. 2), consisting of the following steps:", + "bbox": [ + 511, + 717, + 905, + 762 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Init. Choose one of the aforementioned initialization protocols: initZS, initFT, or initEMA. This produces initialization weights $\\theta_t^I$ at task $t$ .", + "2. Train. Given $\\theta_t^I$ , train on current task data $\\mathcal{D}_t$ within a set compute budget to produce the expert model $\\theta_t^S$ .", + "3. Store. Append $\\theta_t^S$ to storage of expert model weights $S$ .", + "4. Deploy. Choose a deployment protocol: deployFT, deployEMA, or deployALL, and produce the output weights $\\theta_t^O = \\mathrm{deploy}(\\mathcal{S})$ ." + ], + "bbox": [ + 511, + 763, + 903, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20482", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/226214c5b586e741c060c57869bb9e89d99ccf43de286bec220f490540fb3303.jpg", + "image_caption": [ + "Figure 3. Offline merging methods struggle with TIME. All tested merging techniques perform extremely poorly, and are unable to adapt to the temporal setting, underperforming even a simple replay baseline that sequentially trains the base model on task-replayed data." + ], + "image_footnote": [], + "bbox": [ + 94, + 92, + 903, + 289 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Eval. The deployed $\\theta_t^O$ is used for downstream applications and, in our case, extensive evaluation.", + "bbox": [ + 89, + 333, + 480, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Particular choices of init, deploy and $f_{\\text{merge}}$ correspond to existing approaches, for example (initZS, deployALL, $f_{\\text{merge}}^{\\text{WA}}$ ) simply recovers offline merging through weight averaging over experts models derived from original base weights $\\theta_0$ for each task $t$ . Similarly, (initFT, deployEMA, $f_{\\text{merge}}^{\\text{WA}}$ ) recovers exponential moving average approaches as done in [68, 76]. Using our framework, we can explore temporal model merging in a systematic manner across various combinations of all three design choices to understand their impact and recommend best practices.", + "bbox": [ + 88, + 367, + 482, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 549, + 223, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first discuss experimental details in Sec. 4.1. Our initial experiments study the applicability of standard, \"offline\" model merging for the temporal paradigm in Sec. 4.2 and possible extensions in Sec. 4.3. We then transition to a comprehensive exploration of applicable strategies within our TIME framework in Sec. 4.4 and Sec. 4.5, before finally scaling up temporal model merging in Sec. 4.6.", + "bbox": [ + 89, + 575, + 483, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Pipeline Details", + "text_level": 1, + "bbox": [ + 89, + 691, + 356, + 709 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To study temporal model merging in practically relevant scenarios, we use the continual pretraining benchmark Fomo-in-Flux [68]. It includes 41 adaptation datasets and 22 separate evaluation datasets, covering an array of visual and semantic distribution shifts, alongside well-defined, practically motivated compute constraints for each task.", + "bbox": [ + 89, + 714, + 482, + 806 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training at task $t$ . By default, we continuously finetune and merge a ViT-B/16 CLIP [8, 58] model pretrained on LAION-2B [70]. The model is pretrained and finetuned using the standard CLIP objective [58]. Following [68], we fix the training steps for each task based on the DataComp-Small computation budget of $1.8 \\times 10^{9}$", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "GFLOPS, split equally across 20 tasks. We also additionally explore longer sequences with $T \\in \\{50,100\\}$ . At each temporal merging step, we allow unrestricted access to a pretraining data pool $\\mathcal{P}$ , using the same 2M random subset of LAION-400M as in [68]. We use a cosine-decay LR schedule with a linear warmup of $10\\%$ , AdamW optimizer [40], a batchsize of 512 and gradient norm clipping to 1. All experiments use PyTorch [53], and are run on a compute cluster using NVIDIA A100/H100s.", + "bbox": [ + 511, + 333, + 905, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Checkpoint Storage. To enable arbitrary merging strategies at any task $t$ , we store each trained expert $\\theta_t^S$ . We do not store any merged initialization weights $\\theta_t^I$ or deployed models $\\theta_t^O$ , as these can be recovered from stored experts.", + "bbox": [ + 511, + 474, + 905, + 535 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation and Metrics. We focus on two key quantities to evaluate temporal model merging: the level of adaptation, reflecting performance improvement with each merging step, and retention, capturing the preservation of prior knowledge. Specifically, we report two metrics following Roth et al. [68]: Knowledge Accumulation $(\\mathcal{A}_{KA})$ , the average accuracy (or recall@5 for retrieval) across all 41 adaptation datasets, and Zero-Shot Retention $(\\mathcal{A}_{ZS})$ , the zero-shot accuracy or recall@5 on all 22 held-out evaluation datasets. Additional details, including a description of our plotting style, can be found in the supplementary.", + "bbox": [ + 511, + 539, + 908, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Do We Need Model Merging Across Time?", + "text_level": 1, + "bbox": [ + 511, + 720, + 877, + 738 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The simplest approach to tackle temporal merging is to disregard the time axis and follow standard offline merging paradigms. In TIME terms, this corresponds to a configuration of (init_ZS, deploy_ALL, $f_{\\text{merge}}$ ), which always fine-tunes the initial base weights $\\theta_0$ . To study the effectiveness of such an offline strategy, we test the above triplet with various choices of $f_{\\text{merge}}$ , including $f_{\\text{merge}}^{\\text{WA}}$ (averaging in Fig. 3), $f_{\\text{merge}}^{\\text{TA}}$ (task-arithmetic), $f_{\\text{merge}}^{\\text{Max}}$ (magmax), $f_{\\text{merge}}^{\\text{TIES}}$ (ties), $f_{\\text{merge}}^{\\text{DARE}}$ applied over TIES (dare-ties), $f_{\\text{merge}}^{\\text{Bread}}$ (breadcrumbs-ties), and $f_{\\text{merge}}^{\\text{Lines}}$ (lines-ties). To put the results", + "bbox": [ + 511, + 744, + 906, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20483", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ebd093cda73193ac5477e477267bffeaa313b54c949aebe43f0656e36ffbd749.jpg", + "image_caption": [ + "Figure 4. Improving offline merging. We identify two simple methods for adapting offline-merging methods to the temporal setting: (1) replaying data from previous tasks (best-(offline+replay)) and (2) recency-biased weighting of task checkpoints (best-(offline+replay+weighting)). With these method improvements, offline merging methods can match the replay baseline." + ], + "image_footnote": [], + "bbox": [ + 94, + 92, + 356, + 287 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a3bed99fb67c88813c27ba41aefb99d45e29aed54c8d3f2fba5ab1140fd70a8d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 90, + 630, + 287 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2b76313ce61b1817010d8582c8fa3585db7e2df3537da818899c92169cc466d7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 90, + 903, + 287 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "in context, we incorporate (1) a simple continual fine-tuning baseline (replay), following Sec. 4.1, which replays on both pretraining and previous task data, (2) initial zero-shot $(\\theta_0)$ performance lower bound, and (3) multitask training upper bound. We visualize trajectories over time for knowledge accumulation $\\mathcal{A}_{\\mathrm{KA}}$ , zero-shot retention $\\mathcal{A}_{\\mathrm{ZS}}$ , and the geometric mean of both in Fig. 3. Our results clearly highlight that there are only marginal differences between merging techniques $f_{\\mathrm{merge}}$ when deployed in an offline manner for a temporal problem, and they all trace similar trajectories in the $\\mathcal{A}_{\\mathrm{KA}}$ and $\\mathcal{A}_{\\mathrm{ZS}}$ space and achieve similar final performance. We do find that the vanilla $f_{\\mathrm{merge}}^{\\mathrm{TA}}$ (task-arithmetic) starts to deviate after a few update steps, an issue fixed by the various subsequent extensions such as TIES, DARE or LiNeS. Overall, however, unlike straightforward continual fine-tuning (replay), offline merging with any technique fails to address the temporal aspects of the problem, particularly struggling to consistently acquire new knowledge over time (as observed in Fig. 3, left).", + "bbox": [ + 88, + 347, + 485, + 633 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Key takeaway: offline merging with any technique yields similar results, failing to add new knowledge over time.", + "bbox": [ + 89, + 635, + 483, + 666 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Can Replay or Time-Weighting Fix the Issue?", + "text_level": 1, + "bbox": [ + 89, + 679, + 478, + 695 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We next ask: \"What are some simple extensions to offline methods to close the gap to the replay baseline\"? As the continual fine-tuning baseline replays on past data from all previous tasks while training at the current task $t$ , can this task data-mixing also help offline merging methods?", + "bbox": [ + 89, + 702, + 482, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Data replay improves offline merging. Since offline methods operate entirely under a task-independent assumption, they fail to capture any temporal dependencies. Fig. 4 shows that simply applying data-replay on top of standard offline merging leads to significant boosts in the overall performance. For instance, best-(offline+replay) achieves $58.2\\%$ compared to best-offline at $54.6\\%$ , bringing it closer to the replay baseline. However, a notable performance gap", + "bbox": [ + 89, + 780, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "remains, with best-(offline+replay) at $58.2\\%$ falling short of replay at $59.1\\%$ .", + "bbox": [ + 511, + 347, + 906, + 377 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Recency-biased weighting helps. Next, unlike in standard offline averaging, where all task checkpoints are weighted uniformly, we impose temporal ordering via non-uniform weighting for offline merging. We explore several recency-biased, non-uniform weighting schemes, assigning higher weights to more recent tasks to account for the temporal nature of the setting.", + "bbox": [ + 511, + 378, + 908, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We explore various discounting schemes: logarithmic, quadratic, exponential, and cubic, applied to the best offline merge replay method from the previous experiment (please refer to the supplementary for details). As shown in Fig. 4, these schemes improve performance, with best-offline+replay+weighting) reaching $58.9\\%$ yet still falling slightly short of the replay baseline at $59.1\\%$ . These results provide strong evidence that accounting for the new temporal axis is crucial for effective temporal model merging, even when implemented as an extension of offline merging.", + "bbox": [ + 511, + 484, + 908, + 637 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Key Takeaway: Accounting for the time aspect is crucial for effective temporal model merging, even as an extension on top of standard offline merging. Still, a small gap to the simple replay baseline remains.", + "bbox": [ + 511, + 637, + 906, + 698 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. TIME Travel", + "text_level": 1, + "bbox": [ + 511, + 710, + 648, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In our effort to adapt offline merging to the temporal setting, we have explored two axes of our TIME framework—merging technique and deployment. As discussed in Sec. 4.2, the choice of merging technique has minimal impact on final performance, while the deployment strategy plays a much more significant role. Next, we systematically explore the design space for temporal merging by testing all valid combinations of three initialization protocols and three deployment protocols described in Sec. 3.2. After discarding incompatible pairs, such as init_ZS with deployyFT, we evaluated the remaining eight variants us", + "bbox": [ + 511, + 734, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20484", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/639e655301947a8811b4e6988ed89b22f627cbed4850befffae695104730f1dc.jpg", + "image_caption": [ + "Figure 5. A journey through TIME. We explore various initialization and deployment protocols, finding that the EMA initialization-deployment strikes the best balance between knowledge accumulation and zero-shot retention. We refer to this strategy as Best-in-TIME." + ], + "image_footnote": [], + "bbox": [ + 94, + 90, + 356, + 287 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9018a050b489fbfba94241270d6ae30891521d44371f681a3b6be35510469d5f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 90, + 638, + 287 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f7eaf2208b1b1b57e2c95e02a525818266bb39ebd475943f00c1e3a39c631018.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 643, + 90, + 903, + 287 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f9e3177a66ec2ed32b5e3807233b2ce5d05124ee1757d422ac950a2703ffb210.jpg", + "image_caption": [ + "Figure 6. Sweeping Best-in-TIME. All merging techniques perform well with the Best-in-TIME strategy. Indeed, there are no significant differences between techniques, indicating that initialization and deployment matter more for temporal merging." + ], + "image_footnote": [], + "bbox": [ + 94, + 324, + 357, + 521 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1dd6998220954e558705d17b1fb70bccbb96fab789a5e4db0bd6c00a66bacaed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 324, + 638, + 521 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0f3a09fe0ac769d0af23e33f4f49b5b1dacd65324835466a103972dd2328c2f0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 324, + 903, + 521 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ing weight averaging as the merging technique. As shown in Fig. 5, the choice of initialization and deployment strategy largely determines performance, significantly affecting both knowledge accumulation and retention. One combination that stands out consistently is initEMA alongside deployyEMA. This supports the findings of [68, 76] on the applicability of EMA-like solutions to small-scale continual learning and pretraining, while extending the analysis to a much larger, more systematic study of temporal model merging.", + "bbox": [ + 88, + 566, + 482, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Since EMA-style model merging achieves a better balance between knowledge accumulation and retention than other methods, we call this approach Best-in-TIME. In the next section, we will explore the robustness of this strategy across different merging techniques.", + "bbox": [ + 88, + 715, + 482, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Key Takeaway: The EMA strategy for initialization and deployment consistently performs the best.", + "bbox": [ + 89, + 792, + 482, + 823 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. What is the best merge for Best-in-TIME?", + "text_level": 1, + "bbox": [ + 89, + 832, + 447, + 849 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Having identified the optimal initialization and deployment merging strategy in the previous section using weight-averaging as a canonical merging technique, we now invest", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tigate the robustness of our finding by sweeping over other merging techniques, similar to the offline-merging analysis in Sec. 4.2. In particular, we test 7 different merging techniques while keeping the Best-in-TIME initialization and deployment strategy. From Fig. 6, it is immediately evident that all merging techniques perform very similarly—in fact, the knowledge accumulation $(\\mathcal{A}_{KA})$ of all techniques lie within a single standard deviation of performance variation across multiple runs. This indicates that, over a sufficiently long time horizon, all techniques converge to a similar behavior, echoing our results in Sec. 4.2. However, we do notice higher variance in the retention metric $(\\mathcal{A}_{ZS})$ . Overall, our results suggest that for temporal model merging, the specific merging technique used is much less important than selecting the best initialization and deployment strategies.", + "bbox": [ + 511, + 566, + 906, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Key Takeaway: The choice of merging technique is not particularly important—all techniques perform similarly.", + "bbox": [ + 511, + 792, + 905, + 823 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.6. Scaling Up Temporal Model Merging", + "text_level": 1, + "bbox": [ + 511, + 832, + 836, + 849 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We next study the scaling behaviour of temporal model merging across three-axes: model size, compute budget, and number of tasks (results in Fig. 7 and supplen", + "bbox": [ + 511, + 854, + 905, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20485", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ce61e309977d426698779fdbea703da737579f65da74ee43dd42edb9c34de121.jpg", + "image_caption": [ + "Figure 7. Scaling up model merging. (left) With scale, we observe continued improvements of model merging compared to the standard replay baseline. (right) Our Best-in-TIME method continues to improve with scaled total compute budget moving close to the multitask upper-bound. Gray points in the plot visualize suboptimal Best-in-TIME hyperparameter-instantiations." + ], + "image_footnote": [], + "bbox": [ + 99, + 87, + 285, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/996c1b71b013f13effc7c881444e329b4560e24441b189ecfa91c6da5f5592c9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 290, + 88, + 475, + 244 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tary). All our experiments utilize the Best-in-TIME setup described previously, conducting hyperparameter-optimal EMA-weight-averaging at each task.", + "bbox": [ + 89, + 363, + 482, + 409 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Scaling the Model. As we increase the model scale from $S / 16$ (62.3M parameters) to $B / 16$ (149.6M), $L / 14$ (427.6M), and finally $g / 14$ (1.37B) in Fig. 7 (left), we compare sequential fine-tuning (circles) and Best-in-TIME (squares) in terms of knowledge accumulation and retention. We find that: (1) Best-in-TIME scales well with model size, with larger models exhibiting increased affinity to temporal merging. This extends and further corroborates standard offline merging insights by Yadav et al. [89], who showed that model scale facilitates merging. (2) Moreover, while Roth et al. [68] highlight better continual fine-tuning with scale (as also studied by Ibrahim et al. [23]), we show temporal model merging to be substantially more effective across scale. For larger models all the way to the largest ViT-g/14, Best-in-TIME vastly outperforms or matches sequential fine-tuning and the multitask target in knowledge retention and positive backward transfer. Moreover, scale facilitates equivalent degrees of knowledge accumulation between sequential fine-tuning and temporal model merging. Therefore, our model scaling results strongly advocate the use of temporal model merging solutions over standard continual fine-tuning methods [23, 68].", + "bbox": [ + 91, + 414, + 482, + 747 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Scaling the Compute. Keeping the model fixed to ViTB/16, we next scale up the available compute budget by increasing the number of allocated update steps per task. We compare a multitask model, trained on all tasks simultaneously, to a budget-optimal Best-in-TIME. The only hyperparameter for Best-in-TIME is the interpolation weight $w$ . For each compute budget, there is a clear optimal choice of that hyperparameter (suboptimal runs shown as gray dots in Fig. 7 (right)). Higher values of $w$ put greater emphasis on accumulation, allowing optimal accumulation-retention", + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "trade-offs to be reached at lower compute budgets. However, if a larger compute budget is available, less aggressive temporal model merging can achieve higher absolute trade-offs. Note that in Fig. 7 (right), we report the geometric mean between accumulation and retention, corresponding to the right-most panel in previous plots. Our results indicate that Best-in-TIME scales very well across compute budgets, clearly approaching the multitask upper bound in accumulation-retention balance at larger compute budgets. Put together, this gives practitioners the flexibility to adjust Best-in-TIME to their compute budget through the choice of $w$ . Moreover, with more compute to train each respective expert model, the merging process becomes more effective.", + "bbox": [ + 511, + 90, + 903, + 287 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Scaling the Number of Tasks. We next study how Best-in-TIME performs as we increase the number of merging time-steps from $T = 20$ to $T = 50$ and $T = 100$ . As shown in Figs. 10 and 11, Best-in-TIME remains the optimal method across different initialization and deployment strategies. For $T = 50$ and $T = 100$ , we notice that other methods are initially competitive, but face a significant drop-off in performance near the end of the task sequence, while best-in-TIME continues to improve monotonically. This further corroborates that our Best-in-TIME is an extremely scalable method for temporal model merging.", + "bbox": [ + 511, + 289, + 903, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Key Takeaway: Temporal model merging with Best-in-TIME scales efficiently across model sizes and the number of tasks. Moreover, compute scaling improves the effectiveness of temporal model merging.", + "bbox": [ + 511, + 458, + 903, + 520 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 540, + 633, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we study temporal model merging, addressing the challenge of continually merging multimodal models as new tasks and data arrive, and new expert models are trained in succession. To formalize this setting, we propose TIME, a novel unifying framework breaking down temporal model merging into three key axes: (1) initialization phase defining starting weights before each task, (2) deployment phase denoting post-training expert model aggregation, and (3) the choice of merging technique. Using TIME, we conduct a large-scale systematic study uncovering crucial practical guidelines for temporal model merging. Our experiments on the FoMo-in-Flux benchmark spanning 63 datasets, showcase that accounting for the temporal aspect is crucial, with standard offline merging techniques falling short in this dynamic setting. Moreover, we find the particular choice of merging technique matters far less than the merging strategy for initialization and deployment. Finally, we introduce Best-in-TIME, which scales favorably with model size and outperforms existing methods for continual multimodal pretraining. Together, our work provides a systematic entry point into temporal model merging and establishes best practices for this emerging field.", + "bbox": [ + 511, + 568, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20486", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. The authors would like to thank (in random order) Shyamgopal Karthik, Shashwat Goel, Ankit Sonthalia, Olivier Henaff, Alexandre Ramé, and Daniel Marczak for helpful feedback. The plotting style in our work is inspired by figures from Beyer et al. [4]. The style of Fig. 2 is inspired by Figure 1 of Karamcheti et al. [30]. VU, KR, and SD thank the International Max Planck Research School for Intelligent Systems (IMPRS-IS). VU, KR, and SD also thank the European Laboratory for Learning and Intelligent Systems (ELLIS) PhD program for support. VU is supported by a Google PhD Fellowship in Machine Intelligence. SA is supported by a Newton Trust Grant. AP is supported by the Federal Ministry of Education and Research (BMBF), FKZ: 011524085B. MB acknowledges financial support via the Open Philanthropy Foundation funded by the Good Ventures Foundation. MB is a member of the Machine Learning Cluster of Excellence, funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany's Excellence Strategy - EXC number 2064/1 - Project number 390727645. ZA acknowledges the support from the German Research Foundation (DFG): SFB 1233, Robust Vision: Inference Principles and Neural Mechanisms, project number: 276693517 and ERC Grant DEXIM, project number: 853489. This research utilized compute resources at the Tübingen Machine Learning Cloud, DFG FKZ INST 37/1057-1 FUGG.", + "bbox": [ + 89, + 90, + 485, + 467 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 494, + 187, + 510 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Samuel Ainsworth, Jonathan Hayase, and Siddhartha Srinivasa. Git re-basin: Merging models modulo permutation symmetries. In The Eleventh International Conference on Learning Representations, 2023. 2", + "[2] Takuya Akiba, Makoto Shing, Yujin Tang, Qi Sun, and David Ha. Evolutionary optimization of model merging recipes. arXiv preprint arXiv:2403.13187, 2024. 2", + "[3] Anton Alexandrov, Veselin Raychev, Mark Niklas Müller, Ce Zhang, Martin Vechev, and Kristina Toutanova. Mitigating catastrophic forgetting in language transfer via model merging. arXiv preprint arXiv:2407.08699, 2024. 2", + "[4] Lucas Beyer, Xiaohua Zhai, Amélie Royer, Larisa Markeeva, Rohan Anil, and Alexander Kolesnikov. Knowledge distillation: A good teacher is patient and consistent. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10925-10934, 2022. 9, 14", + "[5] Rishi Bommasani, Drew A. Hudson, Ehsan Adeli, Russ B. Altman, Simran Arora, Sydney von Arx, Michael S. Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, Erik Brynjolfsson, Shyamal Buch, Dallas Card, Rodrigo Castellon, Niladri S. Chatterji, Annie S. Chen, Kathleen Creel, Jared Quincy Davis, Dorottya Demszky, Chris Donahue, Moussa Doumbouya, Esin Durmus, Stefano Ermon, John Etchemendy, Kawin Ethayarajh, Li Fei-Fei, Chelsea Finn, Trevor Gale, Lauren Gillespie, Karan Goel, Noah D. Goodman, Shelby Grossman, Neel Guha, Tat" + ], + "bbox": [ + 107, + 520, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "sunori Hashimoto, Peter Henderson, John Hewitt, Daniel E. Ho, Jenny Hong, Kyle Hsu, Jing Huang, Thomas Icard, Saahil Jain, Dan Jurafsky, Pratyusha Kalluri, Siddharth Karamcheti, Geoff Keeling, Fereshte Khani, Omar Khattab, Pang Wei Koh, Mark S. Krass, Ranjay Krishna, Rohith Kuditipudi, and et al. On the opportunities and risks of foundation models. CoRR, abs/2108.07258, 2021. 1", + "[6] Jorg Bornschein, Alexandre Galashov, Ross Hemsley, Amal Rannen-Triki, Yutian Chen, Arslan Chaudhry, Xu Owen He, Arthur Douillard, Massimo Caccia, Qixuan Feng, et al. Nevis' 22: A stream of 100 tasks sampled from 30 years of computer vision research. JMLR, 2023. 2", + "[7] Zhipeng Cai, Ozan Sener, and Vladlen Koltun. Online continual learning with natural distribution shifts: An empirical study with visual data. In ICCV, 2021. 2", + "[8] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. arXiv preprint arXiv:2212.07143, 2022. 5", + "[9] Leshem Choshen, Elad Venezian, Shachar Don-Yehia, Noam Slonim, and Yoav Katz. Where to start? analyzing the potential value of intermediate models. arXiv preprint arXiv:2211.00107, 2022. 2", + "[10] Geoffrey Cideron, Andrea Agostinelli, Johan Ferret, Sertan Girgin, Romuald Elie, Olivier Bachem, Sarah Perrin, and Alexandre Rame. Diversity-rewarded CFG distillation. arXiv preprint arXiv:2410.06084, 2024. 2", + "[11] MohammadReza Davari and Eugene Belilovsky. Model breadcrumbs: Scaling multi-task model merging with sparse masks. In Computer Vision - ECCV 2024, pages 270-287, Cham, 2025. Springer Nature Switzerland. 2, 4", + "[12] Pala Tej Deep, Rishabh Bhardwaj, and Soujanya Poria. Della-merging: Reducing interference in model merging through magnitude-based sampling. arXiv preprint arXiv:2406.11617, 2024. 2", + "[13] Shachar Don-Yehiya, Elad Venezian, Colin Raffel, Noam Slonim, Yoav Katz, and Leshem Choshen. Cold fusion: Collaborative descent for distributed multitask finetuning. arXiv preprint arXiv:2212.01378, 2022. 1, 2", + "[14] Jonathan Frankle, Gintare Karolina Dziugaite, Daniel Roy, and Michael Carbin. Linear mode connectivity and the lottery ticket hypothesis. In Proceedings of the 37th International Conference on Machine Learning, pages 3259-3269. PMLR, 2020. 2", + "[15] João Gama, Indre Žliobaité, Albert Bifet, Mykola Pechenizkiy, and Abdelhamid Bouchachia. A survey on concept drift adaptation. ACM computing surveys, 2014. 3", + "[16] Saurabh Garg, Mehrdad Farajtabar, Hadi Pouransari, Raviteja Vemulapalli, Sachin Mehta, Oncel Tuzel, Vaishaal Shankar, and Fartash Faghri. Tic-clip: Continual training of clip models. In ICLR, 2024. 2", + "[17] Timur Garipov, Pavel Izmailov, Dmitrii Podoprikhin, Dmitry P Vetrov, and Andrew G Wilson. Loss surfaces, mode connectivity, and fast ensembling of dnns. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2018. 1" + ], + "bbox": [ + 522, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20487", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[18] Charles Goddard, Shamane Siriwardhana, Malikeh Ehghaghi, Luke Meyers, Vlad Karpukhin, Brian Benedict, Mark McQuade, and Jacob Solawetz. Arcee's mergekit: A toolkit for merging large language models. arXiv preprint arXiv:2403.13257, 2024. 2", + "[19] Alexey Gorbatovski, Boris Shaposhnikov, Alexey Malakhov, Nikita Surnachev, Yaroslav Aksenov, Ian Maksimov, Nikita Balagansky, and Daniil Gavrilov. Learn your reference model for real good alignment. arXiv preprint arXiv:2404.09656, 2024. 2", + "[20] Zhongrui Gui, Shuyang Sun, Runjia Li, Jianhao Yuan, Zhaochong An, Karsten Roth, Ameya Prabhu, and Philip Torr. knn-clip: Retrieval enables training-free segmentation on continually expanding large vocabularies. arXiv preprint arXiv:2404.09447, 2024. 1, 2", + "[21] Yifei He, Yuzheng Hu, Yong Lin, Tong Zhang, and Han Zhao. Localize-and-stitch: Efficient model merging via sparse task arithmetic. arXiv preprint arXiv:2408.13656, 2024. 2", + "[22] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 1", + "[23] Adam Ibrahim, Benjamin Thérien, Kshitij Gupta, Mats L Richter, Quentin Anthony, Timothée Lesort, Eugene Belilovsky, and Irina Rish. Simple and scalable strategies to continually pre-train large language models. arXiv preprint arXiv:2403.08763, 2024. 2, 8", + "[24] Gabriel Ilharco, Mitchell Wortsman, Samir Yitzhak Gadre, Shuran Song, Hannaneh Hajishirzi, Simon Kornblith, Ali Farhadi, and Ludwig Schmidt. Patching open-vocabulary models by interpolating weights. In NeurIPS, 2022. 1, 2, 4, 14", + "[25] Gabriel Ilharco, Marco Tulio Ribeiro, Mitchell Wortsman, Ludwig Schmidt, Hannaneh Hajishirzi, and Ali Farhadi. Editing models with task arithmetic. In The Eleventh International Conference on Learning Representations, 2023. 2, 4, 14", + "[26] Pavel Izmailov, Dmitrii Podoprikhin, Timur Garipov, Dmitry Vetrov, and Andrew Gordon Wilson. Averaging weights leads to wider optima and better generalization. arXiv preprint arXiv:1803.05407, 2018. 2", + "[27] Dong-Hwan Jang, Sangdoo Yun, and Dongyoon Han. Model stock: All we need is just a few fine-tuned models. In Computer Vision - ECCV 2024: 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part XLIV, page 207-223, Berlin, Heidelberg, 2024. Springer-Verlag. 1, 4", + "[28] Xisen Jin, Xiang Ren, Daniel Preotiac-Pietro, and Pengxiang Cheng. Dataless knowledge fusion by merging weights of language models. In The Eleventh International Conference on Learning Representations, 2023. 2", + "[29] Jean Kaddour. Stop wasting my time! saving days of imageet and bert training with latest weight averaging. arXiv preprint arXiv:2209.14981, 2022. 2", + "[30] Siddharth Karamcheti, Suraj Nair, Ashwin Balakrishna, Percy Liang, Thomas Kollar, and Dorsa Sadigh." + ], + "bbox": [ + 99, + 90, + 485, + 902 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Prismatic vlms: Investigating the design space of visually-conditioned language models. arXiv preprint arXiv:2402.07865, 2024.9", + "[31] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. PNAS, 2017. 3", + "[32] Pang Wei Koh, Shiori Sagawa, Henrik Marklund, Sang Michael Xie, Marvin Zhang, Akshay Balsubramani, Weihua Hu, Michihiro Yasunaga, Richard Lanas Phillips, Irena Gao, et al. Wilds: A benchmark of in-the-wild distribution shifts. In International conference on machine learning, pages 5637-5664. PMLR, 2021. 1", + "[33] Jeffrey Li, Mohammadreza Armandpour, Seyed Iman Mirzadeh, Sachin Mehta, Vaishaal Shankar, Raviteja Vemulapalli, Oncel Tuzel, Mehrdad Farajtabar, Hadi Pouransari, and Fartash Faghri. Tic-lm: A multi-year benchmark for continual pretraining of language models. In NeurIPS 2024 Workshop on Scalable Continual Learning for Lifelong Foundation Models. 2", + "[34] Margaret Li, Suchin Gururangan, Tim Dettmers, Mike Lewis, Tim Althoff, Noah A Smith, and Luke Zettlemoyer. Branch-train-merge: Embarrassingly parallel training of expert language models. arXiv preprint arXiv:2208.03306, 2022. 1", + "[35] Tao Li, Weisen Jiang, Fanghui Liu, Xiaolin Huang, and James T Kwok. Learning scalable model soup on a singlegpu: An efficient subspace training strategy. arXiv preprint arXiv:2407.03641, 2024. 2", + "[36] Yong Lin, Hangyu Lin, Wei Xiong, Shizhe Diao, Jianmeng Liu, Jipeng Zhang, Rui Pan, Haoxiang Wang, Wenbin Hu, Hanning Zhang, Hanze Dong, Renjie Pi, Han Zhao, Nan Jiang, Heng Ji, Yuan Yao, and Tong Zhang. Mitigating the alignment tax of RLHF. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 580-606, Miami, Florida, USA, 2024. Association for Computational Linguistics. 2", + "[37] Zhiqiu Lin, Jia Shi, Deepak Pathak, and Deva Ramanan. The clear benchmark: Continual learning on real-world imagery. In NeurIPS, 2021. 2", + "[38] Adam Liska, Tomas Kocisky, Elena Gribovskaya, Tayfun Terzi, Eren Sezener, Devang Agrawal, D'Autume Cyprien De Masson, Tim Scholtes, Manzil Zaheer, Susannah Young, et al. Streamingqa: A benchmark for adaptation to new knowledge over time in question answering models. In International Conference on Machine Learning, pages 13604-13622. PMLR, 2022. 2", + "[39] Tianlin Liu, Shangmin Guo, Leonardo Bianco, Daniele Calandriello, Quentin Berthet, Felipe Llinares-López, Jessica Hoffmann, Lucas Dixon, Michal Valko, and Mathieu Blondel. Decoding-time realignment of language models. In Proceedings of the 41st International Conference on Machine Learning, pages 31015-31031. PMLR, 2024. 2", + "[40] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5", + "[41] Jinliang Lu, Ziliang Pang, Min Xiao, Yaochen Zhu, Rui Xia, and Jiajun Zhang. Merge, ensemble, and cooperate!" + ], + "bbox": [ + 522, + 92, + 906, + 902 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20488", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "a survey on collaborative strategies in the era of large language models. arXiv preprint arXiv:2407.06089, 2024. 2", + "[42] Daniel Marczak, Bartlomiej Twardowski, Tomasz Trzeciński, and Sebastian Cygert. Magmax: Leveraging model merging for seamless continual learning. arXiv preprint arXiv:2407.06322, 2024. 2, 4", + "[43] Michael S Matena and Colin Raffel. Merging models with fisher-weighted averaging. In Advances in Neural Information Processing Systems, 2022. 2", + "[44] Matías Mendieta, Boran Han, Xingjian Shi, Yi Zhu, and Chen Chen. Towards geospatial foundation models via continual pretraining. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16806-16816, 2023. 2", + "[45] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. In The Eleventh International Conference on Learning Representations, 2023. 1", + "[46] Remi Munos, Michal Valko, Daniele Calandriello, Mohammad Gheshlaghi Azar, Mark Rowland, Zhaohan Daniel Guo, Yunhao Tang, Matthieu Geist, Thomas Mesnard, Come Fiegel, Andrea Michi, Marco Selvi, Sertan Girgin, Nikola Momchev, Olivier Bachem, Daniel J Mankowitz, Doina Precup, and Bilal Piot. Nash learning from human feedback. In Proceedings of the 41st International Conference on Machine Learning, pages 36743-36768. PMLR, 2024. 2", + "[47] Anshul Nasery, Jonathan Hayase, Pang Wei Koh, and Sewoong Oh. Pleas-merging models with permutations and least squares. arXiv preprint arXiv:2407.02447, 2024. 2", + "[48] Behnam Neyshabur, Hanie Sedghi, and Chiyuan Zhang. What is being transferred in transfer learning? In Advances in Neural Information Processing Systems, pages 512-523. Curran Associates, Inc., 2020. 2, 4", + "[49] Kai Nylund, Suchin Gururangan, and Noah A Smith. Time is encoded in the weights of finetuned language models. arXiv preprint arXiv:2312.13401, 2023. 2", + "[50] Changdaee Oh, Yixuan Li, Kyungwoo Song, Sangdoo Yun, and Dongyoon Han. Dawn: Training-free dynamic weight interpolation for robust adaptation. arXiv preprint arXiv:2410.03782, 2024. 2", + "[51] Oleksiy Ostapenko, Timothee Lesort, Pau Rodríguez, Md Rifat Arefin, Arthur Douillard, Irina Rish, and Laurent Charlin. Continual learning with foundation models: An empirical study of latent replay. In Conference on Lifelong Learning Agents (CoLLAs), 2022. 2", + "[52] Jyothish Pari, Samy Jelassi, and Pulkit Agrawal. Collective model intelligence requires compatible specialization. arXiv preprint arXiv:2411.02207, 2024. 2", + "[53] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 5", + "[54] Ameya Prabhu, Zhipeng Cai, Puneet Dokania, Philip Torr, Vladlen Koltun, and Ozan Sener. Online contin" + ], + "bbox": [ + 99, + 92, + 482, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ual learning without the storage constraint. arXiv preprint arXiv:2305.09253, 2023. 2, 3", + "[55] Ameya Prabhu, Hasan Abed Al Kader Hammoud, Puneet Dokania, Philip HS Torr, Ser-Nam Lim, Bernard Ghanem, and Adel Bibi. Computationally budgeted continual learning: What does matter? In CVPR, 2023. 2, 3", + "[56] Ameya Prabhu, Hasan Abed Al Kader Hammoud, Ser-Nam Lim, Bernard Ghanem, Philip HS Torr, and Adel Bibi. From categories to classifier: Name-only continual learning by exploring the web. arXiv preprint arXiv:2311.11293, 2023. 2", + "[57] Sarah Pratt, Ian Covert, Rosanne Liu, and Ali Farhadi. What does a platypus look like? generating customized prompts for zero-shot image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 15691-15701, 2023. 1", + "[58] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5", + "[59] Alexandre Rame, Kartik Ahuja, Jianyu Zhang, Matthieu Cord, Leon Bottou, and David Lopez-Paz. Model rata-touille: Recycling diverse models for out-of-distribution generalization. In Proceedings of the 40th International Conference on Machine Learning, pages 28656-28679. PMLR, 2023. 2", + "[60] Alexandre Ramé, Johan Ferret, Nino Vieillard, Robert Dadashi, Léonard Hussenot, Pierre-Louis Cedoz, Pier Giuseppe Sessa, Sertan Girgin, Arthur Douillard, and Olivier Bachem. Warp: On the benefits of weight averaged rewarded policies. arXiv preprint arXiv:2406.16768, 2024. 2, 4", + "[61] Alexandre Ramé, Matthieu Kirchmeyer, Thibaud Rahier, Alain Rakotomamonjy, Patrick Gallinari, and Matthieu Cord. Diverse weight averaging for out-of-distribution generalization. In Proceedings of the 36th International Conference on Neural Information Processing Systems, 2024. 2", + "[62] Alexandre Rame, Nino Vieillard, Léonard Hussenot, Robert Dadashi, Geoffrey Cideron, Olivier Bachem, and Johan Ferret. Warm: On the benefits of weight averaged reward models. arXiv preprint arXiv:2401.12187, 2024. 2", + "[63] Regent's, ParkLondon, Ukj, and . Utans. Weight averaging for neural networks and local resampling. 1996. 1, 2, 4", + "[64] Mark Rofin, Nikita Balagansky, and Daniil Gavrilov. Linear interpolation in parameter space is good enough for finetuned language models. arXiv preprint arXiv:2211.12092, 2022. 1", + "[65] Karsten Roth, Jae Myung Kim, A. Sophia Koepke, Oriol Vinyals, Cordelia Schmid, and Zeynep Akata. Waffling around for performance: Visual classification with random words and broad concepts. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 15746-15757, 2023. 1", + "[66] Karsten Roth, Zeynep Akata, Dima Damen, Ivana" + ], + "bbox": [ + 522, + 92, + 903, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "20489", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Balažević, and Olivier J Henaff. Context-aware multimodal pretraining. arXiv preprint arXiv:2411.15099, 2024. 1", + "[67] Karsten Roth, Lukas Thede, A. Sophia Koepke, Oriol Vinyals, Olivier J Henaff, and Zeynep Akata. Fantastic gains and where to find them: On the existence and prospect of general knowledge transfer between any pretrained model. In The Twelfth International Conference on Learning Representations, 2024. 2", + "[68] Karsten Roth, Vishaal Udandarao, Sebastian Dziadzio, Ameya Prabhu, Mehdi Cherti, Oriol Vinyals, Olivier Henaff, Samuel Albanie, Matthias Bethge, and Zeynep Akata. A practitioner's guide to continual multimodal pretraining. arXiv preprint arXiv:2408.14471, 2024. 1, 2, 3, 4, 5, 7, 8, 14", + "[69] Sunny Sanyal, Atula Tejaswi Neerkaje, Jean Kaddour, Abhishek Kumar, and sujay sanghavi. Early weight averaging meets high learning rates for LLM pre-training. In First Conference on Language Modeling, 2024. 1, 2", + "[70] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278–25294, 2022. 5", + "[71] Ekansh Sharma, Daniel M Roy, and Gintare Karolina Dziuagaite. The non-local model merging problem: Permutation symmetries and variance collapse. arXiv preprint arXiv:2410.12766, 2024. 2", + "[72] Li Shen, Anke Tang, Enneng Yang, Guibing Guo, Yong Luo, Lefei Zhang, Xiaochun Cao, Bo Du, and Dacheng Tao. Efficient and effective weight-ensembling mixture of experts for multi-task model merging. arXiv preprint arXiv:2410.21804, 2024. 2", + "[73] Ken Shoemake. Animating rotation with quaternion curves. Proceedings of the 12th annual conference on Computer graphics and interactive techniques, 1985. 4", + "[74] Shikhar Srivastava, Md Yousuf Harun, Robik Shrestha, and Christopher Kanan. Improving multimodal large language models using continual learning. arXiv preprint arXiv:2410.19925, 2024. 2", + "[75] George Stoica, Pratik Ramesh, Boglarka Ecsedi, Leshem Choshen, and Judy Hoffman. Model merging with svd to tie the knots. arXiv preprint arXiv:2410.19735, 2024. 2", + "[76] Zafir Stojanovski, Karsten Roth, and Zeynep Akata. Momentum-based weight interpolation of strong zero-shot models for continual learning. arXiv preprint arXiv:2211.03186, 2022. 2, 3, 4, 5, 7", + "[77] Yi-Lin Sung, Linjie Li, Kevin Lin, Zhe Gan, Mohit Bansal, and Lijuan Wang. An empirical study of multimodal model merging. arXiv preprint arXiv:2304.14933, 2023. 2", + "[78] Derek Tam, Mohit Bansal, and Colin Raffel. Merging by matching models in task parameter subspaces. Transactions on Machine Learning Research, 2024. 2", + "[79] Derek Tam, Yash Kant, Brian Lester, Igor Gilitschenski, and Colin Raffel. Realistic evaluation of model merging for compositional generalization. arXiv preprint arXiv:2409.18314, 2024. 2" + ], + "bbox": [ + 99, + 90, + 482, + 898 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[80] Lukas Thede, Karsten Roth, Olivier J Henaff, Matthias Bethge, and Zeynep Akata. Reflecting on the state of rehearsal-free continual learning with pretrained models. arXiv preprint arXiv:2406.09384, 2024. 2", + "[81] Vishaal Udandarao, Ankush Gupta, and Samuel Albanie. Sus-x: Training-free name-only transfer of vision-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2725-2736, 2023. 1", + "[82] Vishaal Udandarao, Nikhil Parthasarathy, Muhammad Ferjad Naeem, Talfan Evans, Samuel Albanie, Federico Tombari, Yongqin Xian, Alessio Tonioni, and Olivier J Henaff. Active data curation effectively distills large-scale multimodal models. arXiv preprint arXiv:2411.18674, 2024.", + "[83] Vishaal Udandarao, Ameya Prabhu, Adhiraj Ghosh, Yash Sharma, Philip HS Torr, Adel Bibi, Samuel Albanie, and Matthias Bethge. No “zero-shot” without exponential data: Pretraining concept frequency determines multimodal model performance. arXiv preprint arXiv:2404.04125, 2024. 1", + "[84] Ke Wang, Nikolaos Dimitriadis, Alessandro Favero, Guillermo Ortiz-Jimenez, Francois Fleuret, and Pascal Frossard. Lines: Post-training layer scaling prevents forgetting and enhances model merging. arXiv preprint arXiv:2410.17146, 2024. 4, 18", + "[85] Ke Wang, Nikolaos Dimitriadis, Guillermo Ortiz-Jimenez, François Fleuret, and Pascal Frossard. Localizing task information for improved model merging and compression. arXiv preprint arXiv:2405.07813, 2024. 2", + "[86] Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, and Ludwig Schmidt. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In Proceedings of the 39th International Conference on Machine Learning, pages 23965-23998. PMLR, 2022. 1, 4", + "[87] Mitchell Wortsman, Gabriel Ilharco, Jong Wook Kim, Mike Li, Simon Kornblith, Rebecca Roelofs, Raphael Gontijo Lopes, Hannaneh Hajishirzi, Ali Farhadi, Hongseok Namkoong, and Ludwig Schmidt. Robust fine-tuning of zero-shot models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7959-7971, 2022. 1, 3, 4", + "[88] Feng Xiong, Runxi Cheng, Wang Chen, Zhanqiu Zhang, Yiwen Guo, Chun Yuan, and Ruifeng Xu. Multi-task model merging via adaptive weight disentanglement. arXiv preprint arXiv:2411.18729, 2024. 2", + "[89] Prateek Yadav, Derek Tam, Leshem Choshen, Colin Raffel, and Mohit Bansal. TIES-merging: Resolving interference when merging models. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 1, 2, 3, 4, 8, 14", + "[90] Prateek Yadav, Colin Raffel, Mohammed Muqeeth, Lucas Caccia, Haokun Liu, Tianlong Chen, Mohit Bansal, Leshem Choshen, and Alessandro Sordoni. A survey on" + ], + "bbox": [ + 522, + 92, + 903, + 900 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "20490", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "model moering: Recycling and routing among specialized experts for collaborative learning. arXiv preprint arXiv:2408.07057, 2024. 2", + "[91] Prateek Yadav, Tu Vu, Jonathan Lai, Alexandra Chronopoulou, Manaal Faruqui, Mohit Bansal, and Tsendsuren Munkhdalai. What matters for model merging at scale? arXiv preprint arXiv:2410.03617, 2024. 2, 3", + "[92] Enneng Yang, Li Shen, Guibing Guo, Xingwei Wang, Xiaochun Cao, Jie Zhang, and Dacheng Tao. Model merging in llms, mllms, and beyond: Methods, theories, applications and opportunities. arXiv preprint arXiv:2408.07666, 2024. 2", + "[93] Enneng Yang, Li Shen, Zhenyi Wang, Guibing Guo, Xiaojun Chen, Xingwei Wang, and Dacheng Tao. Representation surgery for multi-task model merging. arXiv preprint arXiv:2402.02705, 2024. 2", + "[94] Cagatay Yildiz, Nishaanth Kanna Ravichandran, Prishruit Punia, Matthias Bethge, and Beyza Ermis. Investigating continual pretraining in large language models: Insights and implications. arXiv preprint arXiv:2402.17400, 2024. 2", + "[95] Le Yu, Bowen Yu, Haiyang Yu, Fei Huang, and Yongbin Li. Language models are super mario: Absorbing abilities from homologous models as a free lunch. In Proceedings of the 41st International Conference on Machine Learning, pages 57755-57775. PMLR, 2024. 2, 4", + "[96] Kerem Zaman, Leshem Choshen, and Shashank Srivastava. Fuse to forget: Bias reduction and selective memorization through model fusion. arXiv preprint arXiv:2311.07682, 2023. 2", + "[97] Friedemann Zenke, Ben Poole, and Surya Ganguli. Continuual learning through synaptic intelligence. In ICML, 2017. 3", + "[98] Renrui Zhang, Rongyao Fang, Peng Gao, Wei Zhang, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free clip-adapter for better vision-language modeling. arXiv preprint arXiv:2111.03930, 2021. 1", + "[99] Shenghe Zheng and Hongzhi Wang. Free-merging: Fourier transform for model merging with lightweight experts. arXiv preprint arXiv:2411.16815, 2024. 2", + "[100] Luca Zhou, Daniele Solombrino, Donato Crisostomi, Maria Sofia Bucarelli, Fabrizio Silvestri, and Emanuele Rodolà. Atm: Improving model merging by alternating tuning and merging. arXiv preprint arXiv:2411.03055, 2024. 1, 2" + ], + "bbox": [ + 93, + 92, + 482, + 739 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "20491", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 12 + } +] \ No newline at end of file diff --git a/2025/How to Merge Your Multimodal Models Over Time_/76e2370a-2b50-436a-9bdc-9b8af8c8e7ca_model.json b/2025/How to Merge Your Multimodal Models Over Time_/76e2370a-2b50-436a-9bdc-9b8af8c8e7ca_model.json new file mode 100644 index 0000000000000000000000000000000000000000..83fb5f66e750f6016fedd699edf2c8e5814b64e4 --- /dev/null +++ b/2025/How to Merge Your Multimodal Models Over Time_/76e2370a-2b50-436a-9bdc-9b8af8c8e7ca_model.json @@ -0,0 +1,2987 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.131, + 0.744, + 0.153 + ], + "angle": 0, + "content": "How to Merge Your Multimodal Models Over Time?" + }, + { + "type": "image", + "bbox": [ + 0.747, + 0.125, + 0.783, + 0.151 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.158, + 0.18, + 0.836, + 0.218 + ], + "angle": 0, + "content": "Sebastian Dziadzio*¹ Vishaal Udandarao*¹,² Karsten Roth*¹,³ Ameya Prabhu°¹ Zeynep Akata³† Samuel Albanie† Matthias Bethge¹†" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.219, + 0.882, + 0.234 + ], + "angle": 0, + "content": "\\(^{1}\\)Tübingen AI Center, University of Tübingen \\(^{2}\\)University of Cambridge \\(^{3}\\)Munich Center for ML, Technical University of Munich" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.301, + 0.486, + 0.71 + ], + "angle": 0, + "content": "Model merging combines multiple \"expert\" models—finetuned from a base foundation model on diverse tasks and domains—into a single, more capable model. However, most existing model merging approaches assume that all experts are available simultaneously. In reality, new tasks and domains emerge progressively over time, requiring strategies to integrate the knowledge of expert models as they become available: a process we call temporal model merging. The temporal dimension introduces unique challenges not addressed in prior work, raising new questions such as: when training for a new task, should the expert model start from the merged past experts or from the original base model? Should we merge all models at each time step? Which merging techniques are best suited for temporal merging? Should different strategies be used to initialize the training and deploy the model? To answer these questions, we propose a unified framework called TIME—Temporal Integration of Model Expertise—which defines temporal model merging across three axes: (1) Initialization Phase, (2) Deployment Phase, and (3) Merging Technique. Using TIME, we study temporal model merging across model sizes, compute budgets, and learning horizons on the FoMo-in-Flux benchmark. Our comprehensive suite of experiments across TIME allows us to uncover key insights for temporal model merging, offering a better understanding of current challenges and best practices for effective temporal model merging. Our code is available here." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.754, + 0.222, + 0.77 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.841 + ], + "angle": 0, + "content": "Foundation models consolidate a wide range of capabilities and knowledge into a single, large model [5]. Consequently, model merging [24, 34, 63, 87] has emerged as a key technique for unifying multiple task-specific fine-tuned models" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.268, + 0.907, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.409, + 0.907, + 0.48 + ], + "angle": 0, + "content": "Figure 1. Temporal Model Merging generalizes standard model merging (yellow), which merges multiple trained experts just once, in a single step. Our systematic study of this realistic multi-step regime reveals that initialization and deployment strategies dominate the importance of the single-step weight merging strategy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.499, + 0.907, + 0.53 + ], + "angle": 0, + "content": "derived from a shared base into a single model, effectively preserving the strengths of each specialized expert." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.532, + 0.909, + 0.669 + ], + "angle": 0, + "content": "Current model merging approaches typically assume a fixed base model that is fine-tuned independently on \\( k \\) diverse tasks and domains to produce a set of independent experts [17, 24, 34, 64, 89], which are then merged simultaneously. Research in this field has therefore focused primarily on improving merging techniques for larger or structurally differing \\( k \\)-sets, exploring the impact of the diversity and scale of finetuning domains, tasks and experts [27, 69, 86, 89]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.901 + ], + "angle": 0, + "content": "However, the world is constantly evolving, leading to continuous shifts over data distributions, domains, and tasks, with new concepts emerging [45, 68] that may have been insufficiently covered during large-scale pretraining [20, 22, 32, 45, 57, 65, 66, 81-83, 98]. This dynamic nature of real-world applications motivates a hitherto missing systematic exploration into temporal model merging (see Fig. 1) to better understand model merging along an additional, overlooked [13, 100] axis: time. Specifically, in this work, we ask: (1) Is temporal model merging significantly influenced by the choice of initialization for expert training? (2) When evaluated over time, which model merging techniques emerge as most suitable? (3) Should merging strategies differ between initialization and model deployment? To answer these questions, we propose a unified framework" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.852, + 0.483, + 0.901 + ], + "angle": 0, + "content": "*equal contribution, random order †equal supervision, random order, o core contributor. Correspondence to: vishaal.udarao@bethgelab.org or sebastian.dziadzio@bethgelab.org" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20479" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "for studying temporal model merging—TIME (Temporal Integration of Model Expertise)—structured around three key axes spanning the design space of temporal merging solutions (as shown in Fig. 2):" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.156, + 0.483, + 0.202 + ], + "angle": 0, + "content": "1. Initialization Phase. As expert models are trained and created continuously over time, initialization becomes a crucial design element." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.207, + 0.483, + 0.298 + ], + "angle": 0, + "content": "2. Deployment Phase. After training an expert on each task, the next step is to deploy a suitable final model. For temporal model merging, this process has to account for a varying number of past expert models and deployed variants, striking a balance between retaining past knowledge and incorporating new task-specific knowledge." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.303, + 0.483, + 0.424 + ], + "angle": 0, + "content": "3. Model Merging Techniques. Merging literature has explored a range of approaches for simultaneous merging, from simple weight-averaging and interpolation [24, 63, 68, 76] to more complex methods involving weight and candidate selection and re-weighting [11, 42, 43, 89, 95]. To apply these methods temporally, it is crucial to understand how they perform with varying numbers of merge candidates \\((k_{t})\\) and shifting distributions." + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.156, + 0.483, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.426, + 0.483, + 0.577 + ], + "angle": 0, + "content": "Using our TIME framework, we position existing model merging approaches along each key axis and conduct a systematic study of model merging over time. For this, we leverage the multimodal FoMo-in-Flux [68] benchmark which includes 63 datasets with well-documented sequential properties, enabling a thorough investigation of temporal model merging under practical compute constraints, as proposed in Roth et al. [68]. Our experiments systematically explore different merging techniques, initialization, and deployment strategies, providing several key insights:" + }, + { + "type": "title", + "bbox": [ + 0.118, + 0.591, + 0.402, + 0.608 + ], + "angle": 0, + "content": "Key Insights for Temporal Model Merging" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.619, + 0.457, + 0.665 + ], + "angle": 0, + "content": "[A] Accounting for time is crucial. Standard \"off-line\" model merging techniques do not generalize well to the temporal merging setting (Sec. 4.2)." + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.668, + 0.457, + 0.743 + ], + "angle": 0, + "content": "[B] Complex merging techniques matter little. Choosing sophisticated merging techniques beyond simple weighted averaging provides at best marginal benefits for temporal model merging, especially for long task sequences (Sec. 4.5)." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.747, + 0.457, + 0.808 + ], + "angle": 0, + "content": "[C] Initialization and deployment are critical. Choosing how to select and combine available weights before and after each task \\( t \\) is most important for temporal model merging (Sec. 4.4)." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.811, + 0.457, + 0.886 + ], + "angle": 0, + "content": "[D] Temporal model merging scales well. Larger models or compute budgets allows greater benefits from temporal merging. Scaling enables temporal model merging to even outperform the multitask model, trained on all tasks at once (Sec. 4.6)." + }, + { + "type": "list", + "bbox": [ + 0.116, + 0.619, + 0.457, + 0.886 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.664, + 0.106 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.116, + 0.907, + 0.539 + ], + "angle": 0, + "content": "Model Merging. We provide a short overview of the model merging literature, detailed in these excellent surveys [90, 92]. While both model aggregation through distillation [10, 67] and averaging checkpoints during training [29, 35, 69] have shown success, the requirement of additional compute limits practicability of these methods [55]. Instead, recent work [18, 21, 24, 25, 41, 47, 49, 50, 52, 59, 69, 71, 72, 75, 77, 85–88, 90, 93, 96, 99] has shown the effectiveness of training-free weight averaging and interpolation of fine-tuned expert models to produce an improved base model, benefiting from (linear) mode connectivity in models fine-tuned from a single pre-trained checkpoint [1, 14, 26, 48, 61]. These insights have been extended into weight-averaged reward models [62], policy models [60] with spherical interpolation, and KL-constrained RLHF [19, 36, 39, 46]. Works such as Fisher-Merge [43], TIES [89], RegMean [28], MATS [78], DELLA [12], DARE [95], Broadcrumbs [11], evolutionary merging [2] and MagMax [42] have explored merging strategies beyond simple interpolation to determine which weights should be merged across expert models. These methods have different benefits for in- and out-of-distribution generalization across domains [79], though recently they have been shown to perform similarly at scale [91]. Additionally, some works have explored the initialization dimension for effectively merging models [9, 13, 42, 100]. In this work, we propose a unifying framework for temporal merging and conduct the most comprehensive study of this topic to date." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.543, + 0.909, + 0.71 + ], + "angle": 0, + "content": "Continual Pretraining extends beyond standard Continual Learning [54, 68], focusing on large-scale model updates starting from pretrained foundation models [16, 20, 23, 56, 68] and addressing more complex and substantial update tasks [6, 7, 16, 37, 38, 68]. There has been limited exploration into using model merging for continual pretraining [3, 42, 68, 76], as most prior works focus on training strategies including regularization objectives and learning-rate schedules [16, 23, 33, 44, 51, 55, 68, 74, 80, 94]. We keep the training strategy fixed, and provide an in-depth exploration beyond simple merging techniques." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.724, + 0.892, + 0.743 + ], + "angle": 0, + "content": "3. Design Space of Temporal Model Merging" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Notation. Throughout this work, we use \\( t \\) to refer to a given task at time \\( t \\). Full model parameterization is denoted by \\( \\theta \\), with the following key instantiations: \\( \\theta_t \\) represents model weights at task \\( t \\), while \\( \\theta_t^I \\), \\( \\theta_t^S \\), and \\( \\theta_t^O \\) denote weights used for initialization, saved weight checkpoints at task \\( t \\) (i.e., the trained expert models), and the output deployed model weights, respectively. Note that while standard model merging considers model weights as elements of a fixed set \\( \\{\\theta_k\\}_{k=1}^K \\), temporal model merging organizes them along the time axis \\( \\theta_t \\)." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20480" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.087, + 0.905, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.345, + 0.907, + 0.416 + ], + "angle": 0, + "content": "Figure 2. Design Space of Temporal Model Merging through TIME. We showcase our framework for the per-task pipeline of temporal model merging over multiple tasks: At each task \\( t \\), we first initialize the current checkpoint to start training from, \\( \\theta_t^i \\), by using one or more previously stored checkpoints from previous tasks, either directly or by merging them. We train \\( \\theta_t^i \\) on current task data \\( \\mathcal{D}_t \\) to yield the current task checkpoint \\( \\theta_t^s \\), which is inserted into the checkpoint buffer. Finally, to produce the output model, \\( \\theta_t^o \\), we either merge previously stored checkpoints from the buffer or use them directly. The entire framework is depicted in the pseudo-code on the right panel." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.43, + 0.439, + 0.447 + ], + "angle": 0, + "content": "3.1. Temporal Model Merging through TIME" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.453, + 0.484, + 0.59 + ], + "angle": 0, + "content": "Standard model merging is typically performed offline, after all experts have been trained to convergence [15, 87, 89, 91]. In contrast, model merging in continual pretraining is generally done sequentially, using past checkpoints [68, 76]. Both approaches are specific instances of our more general temporal merging framework, TIME, which defines temporal merging along three key axes: initialization of each expert, merging for deployment at step \\( t \\), and merging techniques \\( f_{\\mathrm{merge}} \\) applied over time:" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.597, + 0.241, + 0.61 + ], + "angle": 0, + "content": "Axis 1: Initialization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.616, + 0.484, + 0.767 + ], + "angle": 0, + "content": "As expert models are created continuously over time, initialization becomes a crucial choice. Unlike model merging at a single point in time, the number of potential starting points grows exponentially over time as new experts are created. This raises the question: should starting points for each time step be derived from the base weights (as in traditional merging), from a merged combination of previous experts, or from most recent weights, as commonly done in continual pretraining [68, 76]? In this work, we study the following initialization protocols at time step \\( t \\) for TIME:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.768, + 0.483, + 0.797 + ], + "angle": 0, + "content": "- initZS, which consistently initializes with the base zero-shot model weights \\(\\theta_0\\) at each timestep \\(t\\)." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.798, + 0.484, + 0.828 + ], + "angle": 0, + "content": "- initFT, which for step \\( t \\) always initializes with the latest available finetuned model weights \\( \\theta_{t - 1}^{S} \\)." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.828, + 0.484, + 0.874 + ], + "angle": 0, + "content": "- initEMA, which computes an unrolled exponential moving average merge over all previously seen expert models \\(\\{\\theta_{t'}^S\\}_{1,\\dots,t - 1}\\) following the equation:" + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.768, + 0.484, + 0.874 + ], + "angle": 0, + "content": null + }, + { + "type": "equation", + "bbox": [ + 0.187, + 0.884, + 0.483, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\theta_ {t ^ {\\prime}} ^ {\\mathrm {E M A}} = f _ {\\text {m e r g e}} \\left(\\theta_ {t ^ {\\prime} - 1} ^ {\\mathrm {E M A}}, \\theta_ {t ^ {\\prime} - 1} ^ {S}, \\mathcal {F}\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.431, + 0.907, + 0.462 + ], + "angle": 0, + "content": "with merging hyperparameters \\(\\mathcal{F}\\). Consequently, the initialization weights are given as \\(\\theta_t^I = \\theta_t^{\\mathrm{EMA}}\\)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.472, + 0.659, + 0.487 + ], + "angle": 0, + "content": "Axis 2: Deployment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.492, + 0.907, + 0.734 + ], + "angle": 0, + "content": "With each update iteration and expert training phase \\( t \\), a decision must be made on the final model to deploy, determining which weights to present for downstream use. In continual pretraining, the trained model \\( \\theta_t^S \\) is deployed directly. In contrast, standard model merging applies a merging technique \\( f_{\\text{merge}} \\) to a fixed set of \\( k \\) expert models. Temporal model merging, however, must account for both previously deployed models and the growing number of expert models available over time. Unlike standard merging, where \\( k \\) remains constant, the number of experts to merge increases with each step. As a result, temporal merging introduces the idea of weighted combinations, balancing recent updates with retained past knowledge to achieve adaptability and stability—both critical for effective continual learning [31, 54, 55, 68, 97]. In this work, we study three strategies for model deployment:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.735, + 0.905, + 0.765 + ], + "angle": 0, + "content": "- deployyFT, which at step \\(t\\) deploys the latest finetuned expert model, i.e. \\(\\theta_s^O = \\theta_t^S\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.765, + 0.905, + 0.811 + ], + "angle": 0, + "content": "- deployEMA, which computes an unrolled exponential moving average merge over all expert models, i.e. \\( \\theta_t^O = \\theta_{t+1}^{EMA} \\) following Eq. (1)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.811, + 0.905, + 0.842 + ], + "angle": 0, + "content": "- deployALL, which applies a merging technique \\( f_{\\text{merge}} \\) over all previously computed expert models \\( \\{\\theta_{t'}^S\\}_{1}^{t - 1} \\)." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.735, + 0.905, + 0.842 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.851, + 0.715, + 0.866 + ], + "angle": 0, + "content": "Axis 3: Merging Techniques" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "At each point in time, for both initialization and deployment, merging technique \\( f_{\\text{merge}} \\) defines how to combine" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20481" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.127, + 0.09, + 0.447, + 0.104 + ], + "angle": 0, + "content": "Table 1. Comparison of model merging techniques." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.115, + 0.483, + 0.234 + ], + "angle": 0, + "content": "
MethodSparsificationConsensusScaling
Weight averaging [63, 86]XLinear Int.Weight coeff.
SLERP [60]XSpherical Int.Weight coeff.
Task Arithmetic [25]XLinear Int.Scaling factor
MagMax [42]XMax. MagnitudeScaling factor
TIES [89]Top-kSign AgreementScaling factor
DARE-TIES [95]RandomSign AgreementScaling factor
Breadcrumbs-TIES [11]Top/Bottom-kSign AgreementScaling factor
Model Stock [27]XGeometricAdaptive ratio
LiNeS [84]XXLayer weights
" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.25, + 0.483, + 0.326 + ], + "angle": 0, + "content": "the available expert models and checkpoints. In this work, we study nine different variants in total, shown in Tab. 1. Denoting the number of models to merge at timestep \\( t \\) as \\( M_{t} \\) (with \\( t = 0 \\) and \\( M_{t} = M \\) for standard model merging), we can define these methods as follows:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.34, + 0.484, + 0.389 + ], + "angle": 0, + "content": "Weight Averaging [24, 63, 68, 76, 87] simply employs a uniformly weighted, element-wise average over all models \\(\\theta_{t,i}\\), resulting in a merge function \\(f_{\\mathrm{merge}}^{\\mathrm{WA}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.225, + 0.398, + 0.483, + 0.434 + ], + "angle": 0, + "content": "\\[\n\\theta_ {t} = \\frac {1}{M _ {t}} \\sum_ {i} \\theta_ {t, i}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.444, + 0.484, + 0.537 + ], + "angle": 0, + "content": "SLERP [60, 73] assumes weights to live on a hypersphere, and consequently conducts interpolation along a curved path connecting weight entries. In particular, for two models \\(\\theta_{t,1}\\) and \\(\\theta_{t,2}\\) deriving from some base weight \\(\\theta_{t-1}\\) and the corresponding task vectors [25] \\(\\delta_{t,i} = \\theta_{t,i} - \\theta_{t-1}\\), SLERP with interpolation weight \\(\\lambda\\) is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.546, + 0.483, + 0.58 + ], + "angle": 0, + "content": "\\[\n\\theta_ {t} = \\theta_ {t - 1} + \\frac {\\sin (1 - \\lambda) \\Omega_ {1 , 2}}{\\sin \\Omega_ {1 , 2}} \\cdot \\delta_ {t, 1} + \\frac {\\sin \\lambda \\Omega_ {1 , 2}}{\\sin \\Omega_ {1 , 2}} \\cdot \\delta_ {t, 2} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.59, + 0.483, + 0.621 + ], + "angle": 0, + "content": "with \\(\\Omega_{1,2}\\) being the angle between task vectors \\(\\delta_{t,1}\\) and \\(\\delta_{t,2}\\). We denote the corresponding merge function \\(f_{\\mathrm{merge}}^{\\mathrm{SLERP}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.621, + 0.484, + 0.68 + ], + "angle": 0, + "content": "Task Arithmetic [25] defines the merge as a function over task vectors \\(\\delta_{t,i} = \\theta_{t,i} - \\theta_{t-1}\\) for each weight \\(\\theta_{t,i}\\) fine-tuned from \\(\\theta_{t-1}\\). This introduces a simple merge formalism \\(f_{\\text{merge}}^{\\text{TA}}\\) for weighted parameter averaging with a scale \\(\\lambda\\):" + }, + { + "type": "equation", + "bbox": [ + 0.198, + 0.69, + 0.483, + 0.726 + ], + "angle": 0, + "content": "\\[\n\\theta_ {t} = \\theta_ {t - 1} + \\lambda \\frac {1}{M _ {t}} \\sum_ {i} \\delta_ {t, i} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.746, + 0.483, + 0.854 + ], + "angle": 0, + "content": "TIES [89] builds on the task arithmetic formalism through controlled pruning of task vector entries with low magnitude. Moreover, the sign for each final merged parameter is set based on the sign of the highest total magnitude across the merge candidates. The final update follows basic task arithmetic, only for entries with matching signs. We refer to the respective merge function as \\( f_{\\text{merge}}^{\\text{TIES}} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "DARE [95] is a similar extension of task arithmetic, but instead of targeted pruning, it randomly zeroes out task vector entries using a random mask \\( Z_{i} \\sim \\mathrm{Bernoulli}(p) \\) and" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.091, + 0.905, + 0.122 + ], + "angle": 0, + "content": "masking probability \\( p \\). Final task vector values for \\( f_{\\mathrm{merge}}^{\\mathrm{DARE}} \\) are then rescaled based on \\( p \\):" + }, + { + "type": "equation", + "bbox": [ + 0.633, + 0.129, + 0.905, + 0.163 + ], + "angle": 0, + "content": "\\[\n\\delta_ {t, i} ^ {\\mathrm {D A R E}} = \\frac {(1 - Z _ {i}) \\delta_ {t , i}}{1 - p}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.181, + 0.907, + 0.256 + ], + "angle": 0, + "content": "Model Stock [27] provides a geometric extension of simple weight averaging as done in Model Soup [86] by incorporating base weights \\(\\theta_{t-1}\\) into the merging process. Given finetuned weights \\(\\theta_{t,1}\\) and \\(\\theta_{t,2}\\), the Model Stock merge \\(f_{\\text{merge}}^{\\text{Stock}}\\) is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.262, + 0.905, + 0.297 + ], + "angle": 0, + "content": "\\[\n\\theta_ {t} = \\frac {2 \\cdot \\cos \\Omega_ {1 , 2}}{1 + \\Omega_ {1 , 2}} \\cdot \\left(\\theta_ {t, 2} - \\theta_ {t, 1}\\right) + \\left(1 - \\frac {2 \\cdot \\cos \\Omega_ {1 , 2}}{1 + \\cos \\Omega_ {1 , 2}}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.303, + 0.874, + 0.319 + ], + "angle": 0, + "content": "utilizing angle \\(\\Omega_{1,2}\\) between task vectors \\(\\delta_{t,1}\\) and \\(\\delta_{t,2}\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.321, + 0.905, + 0.429 + ], + "angle": 0, + "content": "Breadcrumbs [11] deploys another variation on task arithmetic for model merging. In particular, for a given task vector \\(\\delta_{t,i}\\), extreme left and right tails of the absolute magnitude distribution in \\(\\delta_{t,i}\\) are zeroed out with left and right thresholds \\(\\beta\\) and \\(\\gamma\\). The modified task vectors \\(\\delta_{t,i}^{\\mathrm{Bread}}\\) are then applied on base weights \\(\\theta_{t-1}\\) following the task arithmetic setup, and giving \\(f_{\\mathrm{merge}}^{\\mathrm{Bread}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.431, + 0.906, + 0.508 + ], + "angle": 0, + "content": "MagMax [42] also uses task vectors—given multiple task vectors \\(\\delta_{t,i}\\) (with increments possible along both time \\(t\\) and count axis \\(i\\)), the final task vector \\(\\delta_t\\) is yielded through maximum magnitude entry selection; copying the largest magnitude entries across all \\(\\{\\delta_{t,i}\\}\\) into \\(\\delta_t\\), giving \\(f_{\\mathrm{merge}}^{\\mathrm{Max}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.51, + 0.906, + 0.616 + ], + "angle": 0, + "content": "LiNeS [84], for Layer-increasing Network Scaling, scales weight updates based on their respective layer depth enabling early layers to remain close to original pretraining weights (cf. Neyshabur et al. [48]). Given task vectors \\(\\delta_{t,i}\\), now broken down across model layers \\(\\delta_{t,i}^{l}\\) with \\(l\\in [1,\\dots,L]\\) and \\(L\\) the number of layers, LiNeS follows the base task arithmetic merging formalism, but updates task vectors as" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.623, + 0.905, + 0.643 + ], + "angle": 0, + "content": "\\[\n\\delta_ {t, i} ^ {\\mathrm {L i N e S}} = \\operatorname {c o n c a t} \\left(\\lambda^ {l = 1} \\delta_ {t, i} ^ {l = 1}, \\dots , \\lambda^ {l = L} \\delta_ {t, i} ^ {l = L}\\right) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.651, + 0.906, + 0.687 + ], + "angle": 0, + "content": "with layer-scaled interpolation weights \\(\\lambda^l = \\alpha +\\beta \\frac{l - 1}{L - 1}\\) and hyperparameters \\(\\alpha ,\\beta\\) giving \\(f_{\\mathrm{merge}}^{\\mathrm{Lines}}\\)" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.696, + 0.892, + 0.712 + ], + "angle": 0, + "content": "3.2. Complete Temporal Model Merging Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.718, + 0.906, + 0.763 + ], + "angle": 0, + "content": "Incorporating all three axes of temporal model merging, we define a five-stage update pipeline for each task \\( t \\) (see Fig. 2), consisting of the following steps:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.765, + 0.905, + 0.81 + ], + "angle": 0, + "content": "1. Init. Choose one of the aforementioned initialization protocols: initZS, initFT, or initEMA. This produces initialization weights \\(\\theta_t^I\\) at task \\(t\\)." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.81, + 0.905, + 0.841 + ], + "angle": 0, + "content": "2. Train. Given \\(\\theta_t^I\\), train on current task data \\(\\mathcal{D}_t\\) within a set compute budget to produce the expert model \\(\\theta_t^S\\)." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.841, + 0.905, + 0.856 + ], + "angle": 0, + "content": "3. Store. Append \\(\\theta_t^S\\) to storage of expert model weights \\(S\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.905, + 0.902 + ], + "angle": 0, + "content": "4. Deploy. Choose a deployment protocol: deployFT, deployEMA, or deployALL, and produce the output weights \\(\\theta_t^O = \\mathrm{deploy}(\\mathcal{S})\\)." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.765, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "20482" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.093, + 0.905, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.292, + 0.908, + 0.322 + ], + "angle": 0, + "content": "Figure 3. Offline merging methods struggle with TIME. All tested merging techniques perform extremely poorly, and are unable to adapt to the temporal setting, underperforming even a simple replay baseline that sequentially trains the base model on task-replayed data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.334, + 0.482, + 0.365 + ], + "angle": 0, + "content": "5. Eval. The deployed \\(\\theta_t^O\\) is used for downstream applications and, in our case, extensive evaluation." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.368, + 0.483, + 0.534 + ], + "angle": 0, + "content": "Particular choices of init, deploy and \\( f_{\\text{merge}} \\) correspond to existing approaches, for example (initZS, deployALL, \\( f_{\\text{merge}}^{\\text{WA}} \\)) simply recovers offline merging through weight averaging over experts models derived from original base weights \\( \\theta_0 \\) for each task \\( t \\). Similarly, (initFT, deployEMA, \\( f_{\\text{merge}}^{\\text{WA}} \\)) recovers exponential moving average approaches as done in [68, 76]. Using our framework, we can explore temporal model merging in a systematic manner across various combinations of all three design choices to understand their impact and recommend best practices." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.55, + 0.224, + 0.567 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.576, + 0.484, + 0.682 + ], + "angle": 0, + "content": "We first discuss experimental details in Sec. 4.1. Our initial experiments study the applicability of standard, \"offline\" model merging for the temporal paradigm in Sec. 4.2 and possible extensions in Sec. 4.3. We then transition to a comprehensive exploration of applicable strategies within our TIME framework in Sec. 4.4 and Sec. 4.5, before finally scaling up temporal model merging in Sec. 4.6." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.693, + 0.357, + 0.71 + ], + "angle": 0, + "content": "4.1. Experimental Pipeline Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.715, + 0.483, + 0.807 + ], + "angle": 0, + "content": "To study temporal model merging in practically relevant scenarios, we use the continual pretraining benchmark Fomo-in-Flux [68]. It includes 41 adaptation datasets and 22 separate evaluation datasets, covering an array of visual and semantic distribution shifts, alongside well-defined, practically motivated compute constraints for each task." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Training at task \\( t \\). By default, we continuously finetune and merge a ViT-B/16 CLIP [8, 58] model pretrained on LAION-2B [70]. The model is pretrained and finetuned using the standard CLIP objective [58]. Following [68], we fix the training steps for each task based on the DataComp-Small computation budget of \\( 1.8 \\times 10^{9} \\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.334, + 0.906, + 0.472 + ], + "angle": 0, + "content": "GFLOPS, split equally across 20 tasks. We also additionally explore longer sequences with \\( T \\in \\{50,100\\} \\). At each temporal merging step, we allow unrestricted access to a pretraining data pool \\( \\mathcal{P} \\), using the same 2M random subset of LAION-400M as in [68]. We use a cosine-decay LR schedule with a linear warmup of \\( 10\\% \\), AdamW optimizer [40], a batchsize of 512 and gradient norm clipping to 1. All experiments use PyTorch [53], and are run on a compute cluster using NVIDIA A100/H100s." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.475, + 0.906, + 0.536 + ], + "angle": 0, + "content": "Checkpoint Storage. To enable arbitrary merging strategies at any task \\( t \\), we store each trained expert \\( \\theta_t^S \\). We do not store any merged initialization weights \\( \\theta_t^I \\) or deployed models \\( \\theta_t^O \\), as these can be recovered from stored experts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.54, + 0.909, + 0.707 + ], + "angle": 0, + "content": "Evaluation and Metrics. We focus on two key quantities to evaluate temporal model merging: the level of adaptation, reflecting performance improvement with each merging step, and retention, capturing the preservation of prior knowledge. Specifically, we report two metrics following Roth et al. [68]: Knowledge Accumulation \\((\\mathcal{A}_{KA})\\), the average accuracy (or recall@5 for retrieval) across all 41 adaptation datasets, and Zero-Shot Retention \\((\\mathcal{A}_{ZS})\\), the zero-shot accuracy or recall@5 on all 22 held-out evaluation datasets. Additional details, including a description of our plotting style, can be found in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.722, + 0.878, + 0.739 + ], + "angle": 0, + "content": "4.2. Do We Need Model Merging Across Time?" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.745, + 0.907, + 0.904 + ], + "angle": 0, + "content": "The simplest approach to tackle temporal merging is to disregard the time axis and follow standard offline merging paradigms. In TIME terms, this corresponds to a configuration of (init_ZS, deploy_ALL, \\(f_{\\text{merge}}\\)), which always fine-tunes the initial base weights \\(\\theta_0\\). To study the effectiveness of such an offline strategy, we test the above triplet with various choices of \\(f_{\\text{merge}}\\), including \\(f_{\\text{merge}}^{\\text{WA}}\\) (averaging in Fig. 3), \\(f_{\\text{merge}}^{\\text{TA}}\\) (task-arithmetic), \\(f_{\\text{merge}}^{\\text{Max}}\\) (magmax), \\(f_{\\text{merge}}^{\\text{TIES}}\\) (ties), \\(f_{\\text{merge}}^{\\text{DARE}}\\) applied over TIES (dare-ties), \\(f_{\\text{merge}}^{\\text{Bread}}\\) (breadcrumbs-ties), and \\(f_{\\text{merge}}^{\\text{Lines}}\\) (lines-ties). To put the results" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20483" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.093, + 0.357, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.092, + 0.631, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.092, + 0.904, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.292, + 0.908, + 0.335 + ], + "angle": 0, + "content": "Figure 4. Improving offline merging. We identify two simple methods for adapting offline-merging methods to the temporal setting: (1) replaying data from previous tasks (best-(offline+replay)) and (2) recency-biased weighting of task checkpoints (best-(offline+replay+weighting)). With these method improvements, offline merging methods can match the replay baseline." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.348, + 0.486, + 0.635 + ], + "angle": 0, + "content": "in context, we incorporate (1) a simple continual fine-tuning baseline (replay), following Sec. 4.1, which replays on both pretraining and previous task data, (2) initial zero-shot \\((\\theta_0)\\) performance lower bound, and (3) multitask training upper bound. We visualize trajectories over time for knowledge accumulation \\(\\mathcal{A}_{\\mathrm{KA}}\\), zero-shot retention \\(\\mathcal{A}_{\\mathrm{ZS}}\\), and the geometric mean of both in Fig. 3. Our results clearly highlight that there are only marginal differences between merging techniques \\(f_{\\mathrm{merge}}\\) when deployed in an offline manner for a temporal problem, and they all trace similar trajectories in the \\(\\mathcal{A}_{\\mathrm{KA}}\\) and \\(\\mathcal{A}_{\\mathrm{ZS}}\\) space and achieve similar final performance. We do find that the vanilla \\(f_{\\mathrm{merge}}^{\\mathrm{TA}}\\) (task-arithmetic) starts to deviate after a few update steps, an issue fixed by the various subsequent extensions such as TIES, DARE or LiNeS. Overall, however, unlike straightforward continual fine-tuning (replay), offline merging with any technique fails to address the temporal aspects of the problem, particularly struggling to consistently acquire new knowledge over time (as observed in Fig. 3, left)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.636, + 0.485, + 0.667 + ], + "angle": 0, + "content": "Key takeaway: offline merging with any technique yields similar results, failing to add new knowledge over time." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.68, + 0.48, + 0.696 + ], + "angle": 0, + "content": "4.3. Can Replay or Time-Weighting Fix the Issue?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.703, + 0.483, + 0.779 + ], + "angle": 0, + "content": "We next ask: \"What are some simple extensions to offline methods to close the gap to the replay baseline\"? As the continual fine-tuning baseline replays on past data from all previous tasks while training at the current task \\( t \\), can this task data-mixing also help offline merging methods?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Data replay improves offline merging. Since offline methods operate entirely under a task-independent assumption, they fail to capture any temporal dependencies. Fig. 4 shows that simply applying data-replay on top of standard offline merging leads to significant boosts in the overall performance. For instance, best-(offline+replay) achieves \\(58.2\\%\\) compared to best-offline at \\(54.6\\%\\), bringing it closer to the replay baseline. However, a notable performance gap" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.348, + 0.907, + 0.378 + ], + "angle": 0, + "content": "remains, with best-(offline+replay) at \\(58.2\\%\\) falling short of replay at \\(59.1\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.379, + 0.909, + 0.485 + ], + "angle": 0, + "content": "Recency-biased weighting helps. Next, unlike in standard offline averaging, where all task checkpoints are weighted uniformly, we impose temporal ordering via non-uniform weighting for offline merging. We explore several recency-biased, non-uniform weighting schemes, assigning higher weights to more recent tasks to account for the temporal nature of the setting." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.486, + 0.909, + 0.638 + ], + "angle": 0, + "content": "We explore various discounting schemes: logarithmic, quadratic, exponential, and cubic, applied to the best offline merge replay method from the previous experiment (please refer to the supplementary for details). As shown in Fig. 4, these schemes improve performance, with best-offline+replay+weighting) reaching \\(58.9\\%\\) yet still falling slightly short of the replay baseline at \\(59.1\\%\\). These results provide strong evidence that accounting for the new temporal axis is crucial for effective temporal model merging, even when implemented as an extension of offline merging." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.638, + 0.907, + 0.699 + ], + "angle": 0, + "content": "Key Takeaway: Accounting for the time aspect is crucial for effective temporal model merging, even as an extension on top of standard offline merging. Still, a small gap to the simple replay baseline remains." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.712, + 0.649, + 0.726 + ], + "angle": 0, + "content": "4.4. TIME Travel" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In our effort to adapt offline merging to the temporal setting, we have explored two axes of our TIME framework—merging technique and deployment. As discussed in Sec. 4.2, the choice of merging technique has minimal impact on final performance, while the deployment strategy plays a much more significant role. Next, we systematically explore the design space for temporal merging by testing all valid combinations of three initialization protocols and three deployment protocols described in Sec. 3.2. After discarding incompatible pairs, such as init_ZS with deployyFT, we evaluated the remaining eight variants us" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20484" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.092, + 0.357, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.092, + 0.639, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.644, + 0.092, + 0.904, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.292, + 0.907, + 0.321 + ], + "angle": 0, + "content": "Figure 5. A journey through TIME. We explore various initialization and deployment protocols, finding that the EMA initialization-deployment strikes the best balance between knowledge accumulation and zero-shot retention. We refer to this strategy as Best-in-TIME." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.325, + 0.359, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.325, + 0.639, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.325, + 0.904, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.524, + 0.907, + 0.553 + ], + "angle": 0, + "content": "Figure 6. Sweeping Best-in-TIME. All merging techniques perform well with the Best-in-TIME strategy. Indeed, there are no significant differences between techniques, indicating that initialization and deployment matter more for temporal merging." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.567, + 0.483, + 0.717 + ], + "angle": 0, + "content": "ing weight averaging as the merging technique. As shown in Fig. 5, the choice of initialization and deployment strategy largely determines performance, significantly affecting both knowledge accumulation and retention. One combination that stands out consistently is initEMA alongside deployyEMA. This supports the findings of [68, 76] on the applicability of EMA-like solutions to small-scale continual learning and pretraining, while extending the analysis to a much larger, more systematic study of temporal model merging." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.717, + 0.483, + 0.793 + ], + "angle": 0, + "content": "Since EMA-style model merging achieves a better balance between knowledge accumulation and retention than other methods, we call this approach Best-in-TIME. In the next section, we will explore the robustness of this strategy across different merging techniques." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.793, + 0.483, + 0.824 + ], + "angle": 0, + "content": "Key Takeaway: The EMA strategy for initialization and deployment consistently performs the best." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.833, + 0.449, + 0.851 + ], + "angle": 0, + "content": "4.5. What is the best merge for Best-in-TIME?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Having identified the optimal initialization and deployment merging strategy in the previous section using weight-averaging as a canonical merging technique, we now invest" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.567, + 0.907, + 0.793 + ], + "angle": 0, + "content": "tigate the robustness of our finding by sweeping over other merging techniques, similar to the offline-merging analysis in Sec. 4.2. In particular, we test 7 different merging techniques while keeping the Best-in-TIME initialization and deployment strategy. From Fig. 6, it is immediately evident that all merging techniques perform very similarly—in fact, the knowledge accumulation \\((\\mathcal{A}_{KA})\\) of all techniques lie within a single standard deviation of performance variation across multiple runs. This indicates that, over a sufficiently long time horizon, all techniques converge to a similar behavior, echoing our results in Sec. 4.2. However, we do notice higher variance in the retention metric \\((\\mathcal{A}_{ZS})\\). Overall, our results suggest that for temporal model merging, the specific merging technique used is much less important than selecting the best initialization and deployment strategies." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.793, + 0.906, + 0.824 + ], + "angle": 0, + "content": "Key Takeaway: The choice of merging technique is not particularly important—all techniques perform similarly." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.833, + 0.837, + 0.851 + ], + "angle": 0, + "content": "4.6. Scaling Up Temporal Model Merging" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.902 + ], + "angle": 0, + "content": "We next study the scaling behaviour of temporal model merging across three-axes: model size, compute budget, and number of tasks (results in Fig. 7 and supplen" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20485" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.088, + 0.287, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.089, + 0.477, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.261, + 0.483, + 0.346 + ], + "angle": 0, + "content": "Figure 7. Scaling up model merging. (left) With scale, we observe continued improvements of model merging compared to the standard replay baseline. (right) Our Best-in-TIME method continues to improve with scaled total compute budget moving close to the multitask upper-bound. Gray points in the plot visualize suboptimal Best-in-TIME hyperparameter-instantiations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.364, + 0.483, + 0.41 + ], + "angle": 0, + "content": "tary). All our experiments utilize the Best-in-TIME setup described previously, conducting hyperparameter-optimal EMA-weight-averaging at each task." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.415, + 0.483, + 0.748 + ], + "angle": 0, + "content": "Scaling the Model. As we increase the model scale from \\( S / 16 \\) (62.3M parameters) to \\( B / 16 \\) (149.6M), \\( L / 14 \\) (427.6M), and finally \\( g / 14 \\) (1.37B) in Fig. 7 (left), we compare sequential fine-tuning (circles) and Best-in-TIME (squares) in terms of knowledge accumulation and retention. We find that: (1) Best-in-TIME scales well with model size, with larger models exhibiting increased affinity to temporal merging. This extends and further corroborates standard offline merging insights by Yadav et al. [89], who showed that model scale facilitates merging. (2) Moreover, while Roth et al. [68] highlight better continual fine-tuning with scale (as also studied by Ibrahim et al. [23]), we show temporal model merging to be substantially more effective across scale. For larger models all the way to the largest ViT-g/14, Best-in-TIME vastly outperforms or matches sequential fine-tuning and the multitask target in knowledge retention and positive backward transfer. Moreover, scale facilitates equivalent degrees of knowledge accumulation between sequential fine-tuning and temporal model merging. Therefore, our model scaling results strongly advocate the use of temporal model merging solutions over standard continual fine-tuning methods [23, 68]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Scaling the Compute. Keeping the model fixed to ViTB/16, we next scale up the available compute budget by increasing the number of allocated update steps per task. We compare a multitask model, trained on all tasks simultaneously, to a budget-optimal Best-in-TIME. The only hyperparameter for Best-in-TIME is the interpolation weight \\( w \\). For each compute budget, there is a clear optimal choice of that hyperparameter (suboptimal runs shown as gray dots in Fig. 7 (right)). Higher values of \\( w \\) put greater emphasis on accumulation, allowing optimal accumulation-retention" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.289 + ], + "angle": 0, + "content": "trade-offs to be reached at lower compute budgets. However, if a larger compute budget is available, less aggressive temporal model merging can achieve higher absolute trade-offs. Note that in Fig. 7 (right), we report the geometric mean between accumulation and retention, corresponding to the right-most panel in previous plots. Our results indicate that Best-in-TIME scales very well across compute budgets, clearly approaching the multitask upper bound in accumulation-retention balance at larger compute budgets. Put together, this gives practitioners the flexibility to adjust Best-in-TIME to their compute budget through the choice of \\( w \\). Moreover, with more compute to train each respective expert model, the merging process becomes more effective." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.29, + 0.905, + 0.457 + ], + "angle": 0, + "content": "Scaling the Number of Tasks. We next study how Best-in-TIME performs as we increase the number of merging time-steps from \\( T = 20 \\) to \\( T = 50 \\) and \\( T = 100 \\). As shown in Figs. 10 and 11, Best-in-TIME remains the optimal method across different initialization and deployment strategies. For \\( T = 50 \\) and \\( T = 100 \\), we notice that other methods are initially competitive, but face a significant drop-off in performance near the end of the task sequence, while best-in-TIME continues to improve monotonically. This further corroborates that our Best-in-TIME is an extremely scalable method for temporal model merging." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.459, + 0.905, + 0.521 + ], + "angle": 0, + "content": "Key Takeaway: Temporal model merging with Best-in-TIME scales efficiently across model sizes and the number of tasks. Moreover, compute scaling improves the effectiveness of temporal model merging." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.541, + 0.634, + 0.557 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.907, + 0.903 + ], + "angle": 0, + "content": "In this work, we study temporal model merging, addressing the challenge of continually merging multimodal models as new tasks and data arrive, and new expert models are trained in succession. To formalize this setting, we propose TIME, a novel unifying framework breaking down temporal model merging into three key axes: (1) initialization phase defining starting weights before each task, (2) deployment phase denoting post-training expert model aggregation, and (3) the choice of merging technique. Using TIME, we conduct a large-scale systematic study uncovering crucial practical guidelines for temporal model merging. Our experiments on the FoMo-in-Flux benchmark spanning 63 datasets, showcase that accounting for the temporal aspect is crucial, with standard offline merging techniques falling short in this dynamic setting. Moreover, we find the particular choice of merging technique matters far less than the merging strategy for initialization and deployment. Finally, we introduce Best-in-TIME, which scales favorably with model size and outperforms existing methods for continual multimodal pretraining. Together, our work provides a systematic entry point into temporal model merging and establishes best practices for this emerging field." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "20486" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.486, + 0.468 + ], + "angle": 0, + "content": "Acknowledgements. The authors would like to thank (in random order) Shyamgopal Karthik, Shashwat Goel, Ankit Sonthalia, Olivier Henaff, Alexandre Ramé, and Daniel Marczak for helpful feedback. The plotting style in our work is inspired by figures from Beyer et al. [4]. The style of Fig. 2 is inspired by Figure 1 of Karamcheti et al. [30]. VU, KR, and SD thank the International Max Planck Research School for Intelligent Systems (IMPRS-IS). VU, KR, and SD also thank the European Laboratory for Learning and Intelligent Systems (ELLIS) PhD program for support. VU is supported by a Google PhD Fellowship in Machine Intelligence. SA is supported by a Newton Trust Grant. AP is supported by the Federal Ministry of Education and Research (BMBF), FKZ: 011524085B. MB acknowledges financial support via the Open Philanthropy Foundation funded by the Good Ventures Foundation. MB is a member of the Machine Learning Cluster of Excellence, funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany's Excellence Strategy - EXC number 2064/1 - Project number 390727645. ZA acknowledges the support from the German Research Foundation (DFG): SFB 1233, Robust Vision: Inference Principles and Neural Mechanisms, project number: 276693517 and ERC Grant DEXIM, project number: 853489. This research utilized compute resources at the Tübingen Machine Learning Cloud, DFG FKZ INST 37/1057-1 FUGG." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.495, + 0.188, + 0.511 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.521, + 0.483, + 0.576 + ], + "angle": 0, + "content": "[1] Samuel Ainsworth, Jonathan Hayase, and Siddhartha Srinivasa. Git re-basin: Merging models modulo permutation symmetries. In The Eleventh International Conference on Learning Representations, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.579, + 0.483, + 0.619 + ], + "angle": 0, + "content": "[2] Takuya Akiba, Makoto Shing, Yujin Tang, Qi Sun, and David Ha. Evolutionary optimization of model merging recipes. arXiv preprint arXiv:2403.13187, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.621, + 0.483, + 0.676 + ], + "angle": 0, + "content": "[3] Anton Alexandrov, Veselin Raychev, Mark Niklas Müller, Ce Zhang, Martin Vechev, and Kristina Toutanova. Mitigating catastrophic forgetting in language transfer via model merging. arXiv preprint arXiv:2407.08699, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.678, + 0.482, + 0.759 + ], + "angle": 0, + "content": "[4] Lucas Beyer, Xiaohua Zhai, Amélie Royer, Larisa Markeeva, Rohan Anil, and Alexander Kolesnikov. Knowledge distillation: A good teacher is patient and consistent. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10925-10934, 2022. 9, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.763, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[5] Rishi Bommasani, Drew A. Hudson, Ehsan Adeli, Russ B. Altman, Simran Arora, Sydney von Arx, Michael S. Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, Erik Brynjolfsson, Shyamal Buch, Dallas Card, Rodrigo Castellon, Niladri S. Chatterji, Annie S. Chen, Kathleen Creel, Jared Quincy Davis, Dorottya Demszky, Chris Donahue, Moussa Doumbouya, Esin Durmus, Stefano Ermon, John Etchemendy, Kawin Ethayarajh, Li Fei-Fei, Chelsea Finn, Trevor Gale, Lauren Gillespie, Karan Goel, Noah D. Goodman, Shelby Grossman, Neel Guha, Tat" + }, + { + "type": "list", + "bbox": [ + 0.108, + 0.521, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.553, + 0.093, + 0.907, + 0.189 + ], + "angle": 0, + "content": "sunori Hashimoto, Peter Henderson, John Hewitt, Daniel E. Ho, Jenny Hong, Kyle Hsu, Jing Huang, Thomas Icard, Saahil Jain, Dan Jurafsky, Pratyusha Kalluri, Siddharth Karamcheti, Geoff Keeling, Fereshte Khani, Omar Khattab, Pang Wei Koh, Mark S. Krass, Ranjay Krishna, Rohith Kuditipudi, and et al. On the opportunities and risks of foundation models. CoRR, abs/2108.07258, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.532, + 0.191, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[6] Jorg Bornschein, Alexandre Galashov, Ross Hemsley, Amal Rannen-Triki, Yutian Chen, Arslan Chaudhry, Xu Owen He, Arthur Douillard, Massimo Caccia, Qixuan Feng, et al. Nevis' 22: A stream of 100 tasks sampled from 30 years of computer vision research. JMLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.532, + 0.261, + 0.905, + 0.301 + ], + "angle": 0, + "content": "[7] Zhipeng Cai, Ozan Sener, and Vladlen Koltun. Online continual learning with natural distribution shifts: An empirical study with visual data. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.532, + 0.302, + 0.905, + 0.37 + ], + "angle": 0, + "content": "[8] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. arXiv preprint arXiv:2212.07143, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.532, + 0.372, + 0.905, + 0.426 + ], + "angle": 0, + "content": "[9] Leshem Choshen, Elad Venezian, Shachar Don-Yehia, Noam Slonim, and Yoav Katz. Where to start? analyzing the potential value of intermediate models. arXiv preprint arXiv:2211.00107, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.428, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[10] Geoffrey Cideron, Andrea Agostinelli, Johan Ferret, Sertan Girgin, Romuald Elie, Olivier Bachem, Sarah Perrin, and Alexandre Rame. Diversity-rewarded CFG distillation. arXiv preprint arXiv:2410.06084, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.484, + 0.905, + 0.551 + ], + "angle": 0, + "content": "[11] MohammadReza Davari and Eugene Belilovsky. Model breadcrumbs: Scaling multi-task model merging with sparse masks. In Computer Vision - ECCV 2024, pages 270-287, Cham, 2025. Springer Nature Switzerland. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.553, + 0.905, + 0.607 + ], + "angle": 0, + "content": "[12] Pala Tej Deep, Rishabh Bhardwaj, and Soujanya Poria. Della-merging: Reducing interference in model merging through magnitude-based sampling. arXiv preprint arXiv:2406.11617, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.609, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[13] Shachar Don-Yehiya, Elad Venezian, Colin Raffel, Noam Slonim, Yoav Katz, and Leshem Choshen. Cold fusion: Collaborative descent for distributed multitask finetuning. arXiv preprint arXiv:2212.01378, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.665, + 0.905, + 0.733 + ], + "angle": 0, + "content": "[14] Jonathan Frankle, Gintare Karolina Dziugaite, Daniel Roy, and Michael Carbin. Linear mode connectivity and the lottery ticket hypothesis. In Proceedings of the 37th International Conference on Machine Learning, pages 3259-3269. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.735, + 0.905, + 0.775 + ], + "angle": 0, + "content": "[15] João Gama, Indre Žliobaité, Albert Bifet, Mykola Pechenizkiy, and Abdelhamid Bouchachia. A survey on concept drift adaptation. ACM computing surveys, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.777, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[16] Saurabh Garg, Mehrdad Farajtabar, Hadi Pouransari, Raviteja Vemulapalli, Sachin Mehta, Oncel Tuzel, Vaishaal Shankar, and Fartash Faghri. Tic-clip: Continual training of clip models. In ICLR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.833, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[17] Timur Garipov, Pavel Izmailov, Dmitrii Podoprikhin, Dmitry P Vetrov, and Andrew G Wilson. Loss surfaces, mode connectivity, and fast ensembling of dnns. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2018. 1" + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20487" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.092, + 0.486, + 0.161 + ], + "angle": 0, + "content": "[18] Charles Goddard, Shamane Siriwardhana, Malikeh Ehghaghi, Luke Meyers, Vlad Karpukhin, Brian Benedict, Mark McQuade, and Jacob Solawetz. Arcee's mergekit: A toolkit for merging large language models. arXiv preprint arXiv:2403.13257, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.162, + 0.484, + 0.232 + ], + "angle": 0, + "content": "[19] Alexey Gorbatovski, Boris Shaposhnikov, Alexey Malakhov, Nikita Surnachev, Yaroslav Aksenov, Ian Maksimov, Nikita Balagansky, and Daniil Gavrilov. Learn your reference model for real good alignment. arXiv preprint arXiv:2404.09656, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.233, + 0.484, + 0.3 + ], + "angle": 0, + "content": "[20] Zhongrui Gui, Shuyang Sun, Runjia Li, Jianhao Yuan, Zhaochong An, Karsten Roth, Ameya Prabhu, and Philip Torr. knn-clip: Retrieval enables training-free segmentation on continually expanding large vocabularies. arXiv preprint arXiv:2404.09447, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.301, + 0.484, + 0.357 + ], + "angle": 0, + "content": "[21] Yifei He, Yuzheng Hu, Yong Lin, Tong Zhang, and Han Zhao. Localize-and-stitch: Efficient model merging via sparse task arithmetic. arXiv preprint arXiv:2408.13656, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.358, + 0.484, + 0.426 + ], + "angle": 0, + "content": "[22] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.427, + 0.484, + 0.496 + ], + "angle": 0, + "content": "[23] Adam Ibrahim, Benjamin Thérien, Kshitij Gupta, Mats L Richter, Quentin Anthony, Timothée Lesort, Eugene Belilovsky, and Irina Rish. Simple and scalable strategies to continually pre-train large language models. arXiv preprint arXiv:2403.08763, 2024. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.497, + 0.484, + 0.564 + ], + "angle": 0, + "content": "[24] Gabriel Ilharco, Mitchell Wortsman, Samir Yitzhak Gadre, Shuran Song, Hannaneh Hajishirzi, Simon Kornblith, Ali Farhadi, and Ludwig Schmidt. Patching open-vocabulary models by interpolating weights. In NeurIPS, 2022. 1, 2, 4, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.566, + 0.484, + 0.635 + ], + "angle": 0, + "content": "[25] Gabriel Ilharco, Marco Tulio Ribeiro, Mitchell Wortsman, Ludwig Schmidt, Hannaneh Hajishirzi, and Ali Farhadi. Editing models with task arithmetic. In The Eleventh International Conference on Learning Representations, 2023. 2, 4, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.636, + 0.484, + 0.692 + ], + "angle": 0, + "content": "[26] Pavel Izmailov, Dmitrii Podoprikhin, Timur Garipov, Dmitry Vetrov, and Andrew Gordon Wilson. Averaging weights leads to wider optima and better generalization. arXiv preprint arXiv:1803.05407, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.692, + 0.484, + 0.775 + ], + "angle": 0, + "content": "[27] Dong-Hwan Jang, Sangdoo Yun, and Dongyoon Han. Model stock: All we need is just a few fine-tuned models. In Computer Vision - ECCV 2024: 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part XLIV, page 207-223, Berlin, Heidelberg, 2024. Springer-Verlag. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.776, + 0.484, + 0.831 + ], + "angle": 0, + "content": "[28] Xisen Jin, Xiang Ren, Daniel Preotiac-Pietro, and Pengxiang Cheng. Dataless knowledge fusion by merging weights of language models. In The Eleventh International Conference on Learning Representations, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.832, + 0.484, + 0.873 + ], + "angle": 0, + "content": "[29] Jean Kaddour. Stop wasting my time! saving days of imageet and bert training with latest weight averaging. arXiv preprint arXiv:2209.14981, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.873, + 0.484, + 0.903 + ], + "angle": 0, + "content": "[30] Siddharth Karamcheti, Suraj Nair, Ashwin Balakrishna, Percy Liang, Thomas Kollar, and Dorsa Sadigh." + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.092, + 0.486, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.553, + 0.093, + 0.907, + 0.134 + ], + "angle": 0, + "content": "Prismatic vlms: Investigating the design space of visually-conditioned language models. arXiv preprint arXiv:2402.07865, 2024.9" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.135, + 0.906, + 0.203 + ], + "angle": 0, + "content": "[31] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. PNAS, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.204, + 0.906, + 0.288 + ], + "angle": 0, + "content": "[32] Pang Wei Koh, Shiori Sagawa, Henrik Marklund, Sang Michael Xie, Marvin Zhang, Akshay Balsubramani, Weihua Hu, Michihiro Yasunaga, Richard Lanas Phillips, Irena Gao, et al. Wilds: A benchmark of in-the-wild distribution shifts. In International conference on machine learning, pages 5637-5664. PMLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.289, + 0.906, + 0.385 + ], + "angle": 0, + "content": "[33] Jeffrey Li, Mohammadreza Armandpour, Seyed Iman Mirzadeh, Sachin Mehta, Vaishaal Shankar, Raviteja Vemulapalli, Oncel Tuzel, Mehrdad Farajtabar, Hadi Pouransari, and Fartash Faghri. Tic-lm: A multi-year benchmark for continual pretraining of language models. In NeurIPS 2024 Workshop on Scalable Continual Learning for Lifelong Foundation Models. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.386, + 0.906, + 0.453 + ], + "angle": 0, + "content": "[34] Margaret Li, Suchin Gururangan, Tim Dettmers, Mike Lewis, Tim Althoff, Noah A Smith, and Luke Zettlemoyer. Branch-train-merge: Embarrassingly parallel training of expert language models. arXiv preprint arXiv:2208.03306, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.455, + 0.906, + 0.51 + ], + "angle": 0, + "content": "[35] Tao Li, Weisen Jiang, Fanghui Liu, Xiaolin Huang, and James T Kwok. Learning scalable model soup on a singlegpu: An efficient subspace training strategy. arXiv preprint arXiv:2407.03641, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.511, + 0.906, + 0.622 + ], + "angle": 0, + "content": "[36] Yong Lin, Hangyu Lin, Wei Xiong, Shizhe Diao, Jianmeng Liu, Jipeng Zhang, Rui Pan, Haoxiang Wang, Wenbin Hu, Hanning Zhang, Hanze Dong, Renjie Pi, Han Zhao, Nan Jiang, Heng Ji, Yuan Yao, and Tong Zhang. Mitigating the alignment tax of RLHF. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 580-606, Miami, Florida, USA, 2024. Association for Computational Linguistics. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.622, + 0.906, + 0.664 + ], + "angle": 0, + "content": "[37] Zhiqiu Lin, Jia Shi, Deepak Pathak, and Deva Ramanan. The clear benchmark: Continual learning on real-world imagery. In NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.664, + 0.906, + 0.76 + ], + "angle": 0, + "content": "[38] Adam Liska, Tomas Kocisky, Elena Gribovskaya, Tayfun Terzi, Eren Sezener, Devang Agrawal, D'Autume Cyprien De Masson, Tim Scholtes, Manzil Zaheer, Susannah Young, et al. Streamingqa: A benchmark for adaptation to new knowledge over time in question answering models. In International Conference on Machine Learning, pages 13604-13622. PMLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.761, + 0.906, + 0.845 + ], + "angle": 0, + "content": "[39] Tianlin Liu, Shangmin Guo, Leonardo Bianco, Daniele Calandriello, Quentin Berthet, Felipe Llinares-López, Jessica Hoffmann, Lucas Dixon, Michal Valko, and Mathieu Blondel. Decoding-time realignment of language models. In Proceedings of the 41st International Conference on Machine Learning, pages 31015-31031. PMLR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.845, + 0.906, + 0.873 + ], + "angle": 0, + "content": "[40] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.873, + 0.906, + 0.903 + ], + "angle": 0, + "content": "[41] Jinliang Lu, Ziliang Pang, Min Xiao, Yaochen Zhu, Rui Xia, and Jiajun Zhang. Merge, ensemble, and cooperate!" + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.093, + 0.907, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20488" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.133, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "a survey on collaborative strategies in the era of large language models. arXiv preprint arXiv:2407.06089, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.122, + 0.483, + 0.177 + ], + "angle": 0, + "content": "[42] Daniel Marczak, Bartlomiej Twardowski, Tomasz Trzeciński, and Sebastian Cygert. Magmax: Leveraging model merging for seamless continual learning. arXiv preprint arXiv:2407.06322, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.179, + 0.482, + 0.22 + ], + "angle": 0, + "content": "[43] Michael S Matena and Colin Raffel. Merging models with fisher-weighted averaging. In Advances in Neural Information Processing Systems, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.222, + 0.483, + 0.29 + ], + "angle": 0, + "content": "[44] Matías Mendieta, Boran Han, Xingjian Shi, Yi Zhu, and Chen Chen. Towards geospatial foundation models via continual pretraining. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16806-16816, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.293, + 0.482, + 0.347 + ], + "angle": 0, + "content": "[45] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. In The Eleventh International Conference on Learning Representations, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.35, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[46] Remi Munos, Michal Valko, Daniele Calandriello, Mohammad Gheshlaghi Azar, Mark Rowland, Zhaohan Daniel Guo, Yunhao Tang, Matthieu Geist, Thomas Mesnard, Come Fiegel, Andrea Michi, Marco Selvi, Sertan Girgin, Nikola Momchev, Olivier Bachem, Daniel J Mankowitz, Doina Precup, and Bilal Piot. Nash learning from human feedback. In Proceedings of the 41st International Conference on Machine Learning, pages 36743-36768. PMLR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.476, + 0.482, + 0.517 + ], + "angle": 0, + "content": "[47] Anshul Nasery, Jonathan Hayase, Pang Wei Koh, and Sewoong Oh. Pleas-merging models with permutations and least squares. arXiv preprint arXiv:2407.02447, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.519, + 0.482, + 0.573 + ], + "angle": 0, + "content": "[48] Behnam Neyshabur, Hanie Sedghi, and Chiyuan Zhang. What is being transferred in transfer learning? In Advances in Neural Information Processing Systems, pages 512-523. Curran Associates, Inc., 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.576, + 0.482, + 0.616 + ], + "angle": 0, + "content": "[49] Kai Nylund, Suchin Gururangan, and Noah A Smith. Time is encoded in the weights of finetuned language models. arXiv preprint arXiv:2312.13401, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.619, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[50] Changdaee Oh, Yixuan Li, Kyungwoo Song, Sangdoo Yun, and Dongyoon Han. Dawn: Training-free dynamic weight interpolation for robust adaptation. arXiv preprint arXiv:2410.03782, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.676, + 0.482, + 0.744 + ], + "angle": 0, + "content": "[51] Oleksiy Ostapenko, Timothee Lesort, Pau Rodríguez, Md Rifat Arefin, Arthur Douillard, Irina Rish, and Laurent Charlin. Continual learning with foundation models: An empirical study of latent replay. In Conference on Lifelong Learning Agents (CoLLAs), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.746, + 0.482, + 0.787 + ], + "angle": 0, + "content": "[52] Jyothish Pari, Samy Jelassi, and Pulkit Agrawal. Collective model intelligence requires compatible specialization. arXiv preprint arXiv:2411.02207, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.789, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[53] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[54] Ameya Prabhu, Zhipeng Cai, Puneet Dokania, Philip Torr, Vladlen Koltun, and Ozan Sener. Online contin" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.554, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "ual learning without the storage constraint. arXiv preprint arXiv:2305.09253, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.122, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[55] Ameya Prabhu, Hasan Abed Al Kader Hammoud, Puneet Dokania, Philip HS Torr, Ser-Nam Lim, Bernard Ghanem, and Adel Bibi. Computationally budgeted continual learning: What does matter? In CVPR, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.179, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[56] Ameya Prabhu, Hasan Abed Al Kader Hammoud, Ser-Nam Lim, Bernard Ghanem, Philip HS Torr, and Adel Bibi. From categories to classifier: Name-only continual learning by exploring the web. arXiv preprint arXiv:2311.11293, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.25, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[57] Sarah Pratt, Ian Covert, Rosanne Liu, and Ali Farhadi. What does a platypus look like? generating customized prompts for zero-shot image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 15691-15701, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.321, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[58] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.406, + 0.905, + 0.487 + ], + "angle": 0, + "content": "[59] Alexandre Rame, Kartik Ahuja, Jianyu Zhang, Matthieu Cord, Leon Bottou, and David Lopez-Paz. Model rata-touille: Recycling diverse models for out-of-distribution generalization. In Proceedings of the 40th International Conference on Machine Learning, pages 28656-28679. PMLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.49, + 0.905, + 0.572 + ], + "angle": 0, + "content": "[60] Alexandre Ramé, Johan Ferret, Nino Vieillard, Robert Dadashi, Léonard Hussenot, Pierre-Louis Cedoz, Pier Giuseppe Sessa, Sertan Girgin, Arthur Douillard, and Olivier Bachem. Warp: On the benefits of weight averaged rewarded policies. arXiv preprint arXiv:2406.16768, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.575, + 0.905, + 0.655 + ], + "angle": 0, + "content": "[61] Alexandre Ramé, Matthieu Kirchmeyer, Thibaud Rahier, Alain Rakotomamonjy, Patrick Gallinari, and Matthieu Cord. Diverse weight averaging for out-of-distribution generalization. In Proceedings of the 36th International Conference on Neural Information Processing Systems, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.659, + 0.905, + 0.714 + ], + "angle": 0, + "content": "[62] Alexandre Rame, Nino Vieillard, Léonard Hussenot, Robert Dadashi, Geoffrey Cideron, Olivier Bachem, and Johan Ferret. Warm: On the benefits of weight averaged reward models. arXiv preprint arXiv:2401.12187, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.717, + 0.905, + 0.744 + ], + "angle": 0, + "content": "[63] Regent's, ParkLondon, Ukj, and . Utans. Weight averaging for neural networks and local resampling. 1996. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.746, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[64] Mark Rofin, Nikita Balagansky, and Daniil Gavrilov. Linear interpolation in parameter space is good enough for finetuned language models. arXiv preprint arXiv:2211.12092, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.803, + 0.905, + 0.885 + ], + "angle": 0, + "content": "[65] Karsten Roth, Jae Myung Kim, A. Sophia Koepke, Oriol Vinyals, Cordelia Schmid, and Zeynep Akata. Waffling around for performance: Visual classification with random words and broad concepts. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 15746-15757, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.888, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[66] Karsten Roth, Zeynep Akata, Dima Damen, Ivana" + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20489" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.133, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "Balažević, and Olivier J Henaff. Context-aware multimodal pretraining. arXiv preprint arXiv:2411.15099, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.122, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[67] Karsten Roth, Lukas Thede, A. Sophia Koepke, Oriol Vinyals, Olivier J Henaff, and Zeynep Akata. Fantastic gains and where to find them: On the existence and prospect of general knowledge transfer between any pretrained model. In The Twelfth International Conference on Learning Representations, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.207, + 0.482, + 0.288 + ], + "angle": 0, + "content": "[68] Karsten Roth, Vishaal Udandarao, Sebastian Dziadzio, Ameya Prabhu, Mehdi Cherti, Oriol Vinyals, Olivier Henaff, Samuel Albanie, Matthias Bethge, and Zeynep Akata. A practitioner's guide to continual multimodal pretraining. arXiv preprint arXiv:2408.14471, 2024. 1, 2, 3, 4, 5, 7, 8, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.291, + 0.483, + 0.346 + ], + "angle": 0, + "content": "[69] Sunny Sanyal, Atula Tejaswi Neerkaje, Jean Kaddour, Abhishek Kumar, and sujay sanghavi. Early weight averaging meets high learning rates for LLM pre-training. In First Conference on Language Modeling, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.348, + 0.483, + 0.431 + ], + "angle": 0, + "content": "[70] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278–25294, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.433, + 0.483, + 0.487 + ], + "angle": 0, + "content": "[71] Ekansh Sharma, Daniel M Roy, and Gintare Karolina Dziuagaite. The non-local model merging problem: Permutation symmetries and variance collapse. arXiv preprint arXiv:2410.12766, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.489, + 0.483, + 0.557 + ], + "angle": 0, + "content": "[72] Li Shen, Anke Tang, Enneng Yang, Guibing Guo, Yong Luo, Lefei Zhang, Xiaochun Cao, Bo Du, and Dacheng Tao. Efficient and effective weight-ensembling mixture of experts for multi-task model merging. arXiv preprint arXiv:2410.21804, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.56, + 0.483, + 0.602 + ], + "angle": 0, + "content": "[73] Ken Shoemake. Animating rotation with quaternion curves. Proceedings of the 12th annual conference on Computer graphics and interactive techniques, 1985. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.604, + 0.483, + 0.657 + ], + "angle": 0, + "content": "[74] Shikhar Srivastava, Md Yousuf Harun, Robik Shrestha, and Christopher Kanan. Improving multimodal large language models using continual learning. arXiv preprint arXiv:2410.19925, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.66, + 0.483, + 0.702 + ], + "angle": 0, + "content": "[75] George Stoica, Pratik Ramesh, Boglarka Ecsedi, Leshem Choshen, and Judy Hoffman. Model merging with svd to tie the knots. arXiv preprint arXiv:2410.19735, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.704, + 0.483, + 0.758 + ], + "angle": 0, + "content": "[76] Zafir Stojanovski, Karsten Roth, and Zeynep Akata. Momentum-based weight interpolation of strong zero-shot models for continual learning. arXiv preprint arXiv:2211.03186, 2022. 2, 3, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.76, + 0.483, + 0.801 + ], + "angle": 0, + "content": "[77] Yi-Lin Sung, Linjie Li, Kevin Lin, Zhe Gan, Mohit Bansal, and Lijuan Wang. An empirical study of multimodal model merging. arXiv preprint arXiv:2304.14933, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.803, + 0.483, + 0.843 + ], + "angle": 0, + "content": "[78] Derek Tam, Mohit Bansal, and Colin Raffel. Merging by matching models in task parameter subspaces. Transactions on Machine Learning Research, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.846, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[79] Derek Tam, Yash Kant, Brian Lester, Igor Gilitschenski, and Colin Raffel. Realistic evaluation of model merging for compositional generalization. arXiv preprint arXiv:2409.18314, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[80] Lukas Thede, Karsten Roth, Olivier J Henaff, Matthias Bethge, and Zeynep Akata. Reflecting on the state of rehearsal-free continual learning with pretrained models. arXiv preprint arXiv:2406.09384, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.151, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[81] Vishaal Udandarao, Ankush Gupta, and Samuel Albanie. Sus-x: Training-free name-only transfer of vision-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2725-2736, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.222, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[82] Vishaal Udandarao, Nikhil Parthasarathy, Muhammad Ferjad Naeem, Talfan Evans, Samuel Albanie, Federico Tombari, Yongqin Xian, Alessio Tonioni, and Olivier J Henaff. Active data curation effectively distills large-scale multimodal models. arXiv preprint arXiv:2411.18674, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.307, + 0.905, + 0.387 + ], + "angle": 0, + "content": "[83] Vishaal Udandarao, Ameya Prabhu, Adhiraj Ghosh, Yash Sharma, Philip HS Torr, Adel Bibi, Samuel Albanie, and Matthias Bethge. No “zero-shot” without exponential data: Pretraining concept frequency determines multimodal model performance. arXiv preprint arXiv:2404.04125, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.392, + 0.905, + 0.46 + ], + "angle": 0, + "content": "[84] Ke Wang, Nikolaos Dimitriadis, Alessandro Favero, Guillermo Ortiz-Jimenez, Francois Fleuret, and Pascal Frossard. Lines: Post-training layer scaling prevents forgetting and enhances model merging. arXiv preprint arXiv:2410.17146, 2024. 4, 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.463, + 0.905, + 0.517 + ], + "angle": 0, + "content": "[85] Ke Wang, Nikolaos Dimitriadis, Guillermo Ortiz-Jimenez, François Fleuret, and Pascal Frossard. Localizing task information for improved model merging and compression. arXiv preprint arXiv:2405.07813, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.52, + 0.905, + 0.63 + ], + "angle": 0, + "content": "[86] Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, and Ludwig Schmidt. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In Proceedings of the 39th International Conference on Machine Learning, pages 23965-23998. PMLR, 2022. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.633, + 0.905, + 0.73 + ], + "angle": 0, + "content": "[87] Mitchell Wortsman, Gabriel Ilharco, Jong Wook Kim, Mike Li, Simon Kornblith, Rebecca Roelofs, Raphael Gontijo Lopes, Hannaneh Hajishirzi, Ali Farhadi, Hongseok Namkoong, and Ludwig Schmidt. Robust fine-tuning of zero-shot models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7959-7971, 2022. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.732, + 0.905, + 0.786 + ], + "angle": 0, + "content": "[88] Feng Xiong, Runxi Cheng, Wang Chen, Zhanqiu Zhang, Yiwen Guo, Chun Yuan, and Ruifeng Xu. Multi-task model merging via adaptive weight disentanglement. arXiv preprint arXiv:2411.18729, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.789, + 0.905, + 0.855 + ], + "angle": 0, + "content": "[89] Prateek Yadav, Derek Tam, Leshem Choshen, Colin Raffel, and Mohit Bansal. TIES-merging: Resolving interference when merging models. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 1, 2, 3, 4, 8, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[90] Prateek Yadav, Colin Raffel, Mohammed Muqeeth, Lucas Caccia, Haokun Liu, Tianlong Chen, Mohit Bansal, Leshem Choshen, and Alessandro Sordoni. A survey on" + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20490" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.131, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "model moering: Recycling and routing among specialized experts for collaborative learning. arXiv preprint arXiv:2408.07057, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.136, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[91] Prateek Yadav, Tu Vu, Jonathan Lai, Alexandra Chronopoulou, Manaal Faruqui, Mohit Bansal, and Tsendsuren Munkhdalai. What matters for model merging at scale? arXiv preprint arXiv:2410.03617, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.192, + 0.482, + 0.26 + ], + "angle": 0, + "content": "[92] Enneng Yang, Li Shen, Guibing Guo, Xingwei Wang, Xiaochun Cao, Jie Zhang, and Dacheng Tao. Model merging in llms, mllms, and beyond: Methods, theories, applications and opportunities. arXiv preprint arXiv:2408.07666, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.263, + 0.482, + 0.317 + ], + "angle": 0, + "content": "[93] Enneng Yang, Li Shen, Zhenyi Wang, Guibing Guo, Xiaojun Chen, Xingwei Wang, and Dacheng Tao. Representation surgery for multi-task model merging. arXiv preprint arXiv:2402.02705, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.32, + 0.483, + 0.386 + ], + "angle": 0, + "content": "[94] Cagatay Yildiz, Nishaanth Kanna Ravichandran, Prishruit Punia, Matthias Bethge, and Beyza Ermis. Investigating continual pretraining in large language models: Insights and implications. arXiv preprint arXiv:2402.17400, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.39, + 0.482, + 0.459 + ], + "angle": 0, + "content": "[95] Le Yu, Bowen Yu, Haiyang Yu, Fei Huang, and Yongbin Li. Language models are super mario: Absorbing abilities from homologous models as a free lunch. In Proceedings of the 41st International Conference on Machine Learning, pages 57755-57775. PMLR, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.461, + 0.482, + 0.514 + ], + "angle": 0, + "content": "[96] Kerem Zaman, Leshem Choshen, and Shashank Srivastava. Fuse to forget: Bias reduction and selective memorization through model fusion. arXiv preprint arXiv:2311.07682, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.517, + 0.482, + 0.556 + ], + "angle": 0, + "content": "[97] Friedemann Zenke, Ben Poole, and Surya Ganguli. Continuual learning through synaptic intelligence. In ICML, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.56, + 0.482, + 0.627 + ], + "angle": 0, + "content": "[98] Renrui Zhang, Rongyao Fang, Peng Gao, Wei Zhang, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free clip-adapter for better vision-language modeling. arXiv preprint arXiv:2111.03930, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.631, + 0.482, + 0.671 + ], + "angle": 0, + "content": "[99] Shenghe Zheng and Hongzhi Wang. Free-merging: Fourier transform for model merging with lightweight experts. arXiv preprint arXiv:2411.16815, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.673, + 0.482, + 0.74 + ], + "angle": 0, + "content": "[100] Luca Zhou, Daniele Solombrino, Donato Crisostomi, Maria Sofia Bucarelli, Fabrizio Silvestri, and Emanuele Rodolà. Atm: Improving model merging by alternating tuning and merging. arXiv preprint arXiv:2411.03055, 2024. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.483, + 0.74 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20491" + } + ] +] \ No newline at end of file diff --git a/2025/How to Merge Your Multimodal Models Over Time_/76e2370a-2b50-436a-9bdc-9b8af8c8e7ca_origin.pdf b/2025/How to Merge Your Multimodal Models Over Time_/76e2370a-2b50-436a-9bdc-9b8af8c8e7ca_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2104476695aefbf89bc5517c9053c14fc50a3a03 --- /dev/null +++ b/2025/How to Merge Your Multimodal Models Over Time_/76e2370a-2b50-436a-9bdc-9b8af8c8e7ca_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8b914c88955cedbb28ac38262161bbc8ba665811963c34098b8b28bae542285b +size 1715248 diff --git a/2025/How to Merge Your Multimodal Models Over Time_/full.md b/2025/How to Merge Your Multimodal Models Over Time_/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f88817f167a019ed940a17d5a8fca0ac1dbe0379 --- /dev/null +++ b/2025/How to Merge Your Multimodal Models Over Time_/full.md @@ -0,0 +1,375 @@ +# How to Merge Your Multimodal Models Over Time? + +![](images/d3ca2ac2a62466f29cd4d95337275d8e2a7cb93585bf35347106ea6962047f39.jpg) + +Sebastian Dziadzio*¹ Vishaal Udandarao*¹,² Karsten Roth*¹,³ Ameya Prabhu°¹ Zeynep Akata³† Samuel Albanie† Matthias Bethge¹† + +$^{1}$ Tübingen AI Center, University of Tübingen $^{2}$ University of Cambridge $^{3}$ Munich Center for ML, Technical University of Munich + +# Abstract + +Model merging combines multiple "expert" models—finetuned from a base foundation model on diverse tasks and domains—into a single, more capable model. However, most existing model merging approaches assume that all experts are available simultaneously. In reality, new tasks and domains emerge progressively over time, requiring strategies to integrate the knowledge of expert models as they become available: a process we call temporal model merging. The temporal dimension introduces unique challenges not addressed in prior work, raising new questions such as: when training for a new task, should the expert model start from the merged past experts or from the original base model? Should we merge all models at each time step? Which merging techniques are best suited for temporal merging? Should different strategies be used to initialize the training and deploy the model? To answer these questions, we propose a unified framework called TIME—Temporal Integration of Model Expertise—which defines temporal model merging across three axes: (1) Initialization Phase, (2) Deployment Phase, and (3) Merging Technique. Using TIME, we study temporal model merging across model sizes, compute budgets, and learning horizons on the FoMo-in-Flux benchmark. Our comprehensive suite of experiments across TIME allows us to uncover key insights for temporal model merging, offering a better understanding of current challenges and best practices for effective temporal model merging. Our code is available here. + +# 1. Introduction + +Foundation models consolidate a wide range of capabilities and knowledge into a single, large model [5]. Consequently, model merging [24, 34, 63, 87] has emerged as a key technique for unifying multiple task-specific fine-tuned models + +![](images/3c8338fe5b160b2a80269142625cfe5aefaa614fa03a02a07f4f5debd3b41ad5.jpg) +Figure 1. Temporal Model Merging generalizes standard model merging (yellow), which merges multiple trained experts just once, in a single step. Our systematic study of this realistic multi-step regime reveals that initialization and deployment strategies dominate the importance of the single-step weight merging strategy. + +derived from a shared base into a single model, effectively preserving the strengths of each specialized expert. + +Current model merging approaches typically assume a fixed base model that is fine-tuned independently on $k$ diverse tasks and domains to produce a set of independent experts [17, 24, 34, 64, 89], which are then merged simultaneously. Research in this field has therefore focused primarily on improving merging techniques for larger or structurally differing $k$ -sets, exploring the impact of the diversity and scale of finetuning domains, tasks and experts [27, 69, 86, 89]. + +However, the world is constantly evolving, leading to continuous shifts over data distributions, domains, and tasks, with new concepts emerging [45, 68] that may have been insufficiently covered during large-scale pretraining [20, 22, 32, 45, 57, 65, 66, 81-83, 98]. This dynamic nature of real-world applications motivates a hitherto missing systematic exploration into temporal model merging (see Fig. 1) to better understand model merging along an additional, overlooked [13, 100] axis: time. Specifically, in this work, we ask: (1) Is temporal model merging significantly influenced by the choice of initialization for expert training? (2) When evaluated over time, which model merging techniques emerge as most suitable? (3) Should merging strategies differ between initialization and model deployment? To answer these questions, we propose a unified framework + +for studying temporal model merging—TIME (Temporal Integration of Model Expertise)—structured around three key axes spanning the design space of temporal merging solutions (as shown in Fig. 2): + +1. Initialization Phase. As expert models are trained and created continuously over time, initialization becomes a crucial design element. +2. Deployment Phase. After training an expert on each task, the next step is to deploy a suitable final model. For temporal model merging, this process has to account for a varying number of past expert models and deployed variants, striking a balance between retaining past knowledge and incorporating new task-specific knowledge. +3. Model Merging Techniques. Merging literature has explored a range of approaches for simultaneous merging, from simple weight-averaging and interpolation [24, 63, 68, 76] to more complex methods involving weight and candidate selection and re-weighting [11, 42, 43, 89, 95]. To apply these methods temporally, it is crucial to understand how they perform with varying numbers of merge candidates $(k_{t})$ and shifting distributions. + +Using our TIME framework, we position existing model merging approaches along each key axis and conduct a systematic study of model merging over time. For this, we leverage the multimodal FoMo-in-Flux [68] benchmark which includes 63 datasets with well-documented sequential properties, enabling a thorough investigation of temporal model merging under practical compute constraints, as proposed in Roth et al. [68]. Our experiments systematically explore different merging techniques, initialization, and deployment strategies, providing several key insights: + +# Key Insights for Temporal Model Merging + +[A] Accounting for time is crucial. Standard "off-line" model merging techniques do not generalize well to the temporal merging setting (Sec. 4.2). +[B] Complex merging techniques matter little. Choosing sophisticated merging techniques beyond simple weighted averaging provides at best marginal benefits for temporal model merging, especially for long task sequences (Sec. 4.5). +[C] Initialization and deployment are critical. Choosing how to select and combine available weights before and after each task $t$ is most important for temporal model merging (Sec. 4.4). +[D] Temporal model merging scales well. Larger models or compute budgets allows greater benefits from temporal merging. Scaling enables temporal model merging to even outperform the multitask model, trained on all tasks at once (Sec. 4.6). + +# 2. Related Works + +Model Merging. We provide a short overview of the model merging literature, detailed in these excellent surveys [90, 92]. While both model aggregation through distillation [10, 67] and averaging checkpoints during training [29, 35, 69] have shown success, the requirement of additional compute limits practicability of these methods [55]. Instead, recent work [18, 21, 24, 25, 41, 47, 49, 50, 52, 59, 69, 71, 72, 75, 77, 85–88, 90, 93, 96, 99] has shown the effectiveness of training-free weight averaging and interpolation of fine-tuned expert models to produce an improved base model, benefiting from (linear) mode connectivity in models fine-tuned from a single pre-trained checkpoint [1, 14, 26, 48, 61]. These insights have been extended into weight-averaged reward models [62], policy models [60] with spherical interpolation, and KL-constrained RLHF [19, 36, 39, 46]. Works such as Fisher-Merge [43], TIES [89], RegMean [28], MATS [78], DELLA [12], DARE [95], Broadcrumbs [11], evolutionary merging [2] and MagMax [42] have explored merging strategies beyond simple interpolation to determine which weights should be merged across expert models. These methods have different benefits for in- and out-of-distribution generalization across domains [79], though recently they have been shown to perform similarly at scale [91]. Additionally, some works have explored the initialization dimension for effectively merging models [9, 13, 42, 100]. In this work, we propose a unifying framework for temporal merging and conduct the most comprehensive study of this topic to date. + +Continual Pretraining extends beyond standard Continual Learning [54, 68], focusing on large-scale model updates starting from pretrained foundation models [16, 20, 23, 56, 68] and addressing more complex and substantial update tasks [6, 7, 16, 37, 38, 68]. There has been limited exploration into using model merging for continual pretraining [3, 42, 68, 76], as most prior works focus on training strategies including regularization objectives and learning-rate schedules [16, 23, 33, 44, 51, 55, 68, 74, 80, 94]. We keep the training strategy fixed, and provide an in-depth exploration beyond simple merging techniques. + +# 3. Design Space of Temporal Model Merging + +Notation. Throughout this work, we use $t$ to refer to a given task at time $t$ . Full model parameterization is denoted by $\theta$ , with the following key instantiations: $\theta_t$ represents model weights at task $t$ , while $\theta_t^I$ , $\theta_t^S$ , and $\theta_t^O$ denote weights used for initialization, saved weight checkpoints at task $t$ (i.e., the trained expert models), and the output deployed model weights, respectively. Note that while standard model merging considers model weights as elements of a fixed set $\{\theta_k\}_{k=1}^K$ , temporal model merging organizes them along the time axis $\theta_t$ . + +![](images/a42037083fab748e2b76f8611cfef4e3d1df23ad74a56b9313e8bf709ebc64ef.jpg) +Figure 2. Design Space of Temporal Model Merging through TIME. We showcase our framework for the per-task pipeline of temporal model merging over multiple tasks: At each task $t$ , we first initialize the current checkpoint to start training from, $\theta_t^i$ , by using one or more previously stored checkpoints from previous tasks, either directly or by merging them. We train $\theta_t^i$ on current task data $\mathcal{D}_t$ to yield the current task checkpoint $\theta_t^s$ , which is inserted into the checkpoint buffer. Finally, to produce the output model, $\theta_t^o$ , we either merge previously stored checkpoints from the buffer or use them directly. The entire framework is depicted in the pseudo-code on the right panel. + +# 3.1. Temporal Model Merging through TIME + +Standard model merging is typically performed offline, after all experts have been trained to convergence [15, 87, 89, 91]. In contrast, model merging in continual pretraining is generally done sequentially, using past checkpoints [68, 76]. Both approaches are specific instances of our more general temporal merging framework, TIME, which defines temporal merging along three key axes: initialization of each expert, merging for deployment at step $t$ , and merging techniques $f_{\mathrm{merge}}$ applied over time: + +# Axis 1: Initialization + +As expert models are created continuously over time, initialization becomes a crucial choice. Unlike model merging at a single point in time, the number of potential starting points grows exponentially over time as new experts are created. This raises the question: should starting points for each time step be derived from the base weights (as in traditional merging), from a merged combination of previous experts, or from most recent weights, as commonly done in continual pretraining [68, 76]? In this work, we study the following initialization protocols at time step $t$ for TIME: + +- initZS, which consistently initializes with the base zero-shot model weights $\theta_0$ at each timestep $t$ . +- initFT, which for step $t$ always initializes with the latest available finetuned model weights $\theta_{t - 1}^{S}$ . +- initEMA, which computes an unrolled exponential moving average merge over all previously seen expert models $\{\theta_{t'}^S\}_{1,\dots,t - 1}$ following the equation: + +$$ +\theta_ {t ^ {\prime}} ^ {\mathrm {E M A}} = f _ {\text {m e r g e}} \left(\theta_ {t ^ {\prime} - 1} ^ {\mathrm {E M A}}, \theta_ {t ^ {\prime} - 1} ^ {S}, \mathcal {F}\right) \tag {1} +$$ + +with merging hyperparameters $\mathcal{F}$ . Consequently, the initialization weights are given as $\theta_t^I = \theta_t^{\mathrm{EMA}}$ . + +# Axis 2: Deployment + +With each update iteration and expert training phase $t$ , a decision must be made on the final model to deploy, determining which weights to present for downstream use. In continual pretraining, the trained model $\theta_t^S$ is deployed directly. In contrast, standard model merging applies a merging technique $f_{\text{merge}}$ to a fixed set of $k$ expert models. Temporal model merging, however, must account for both previously deployed models and the growing number of expert models available over time. Unlike standard merging, where $k$ remains constant, the number of experts to merge increases with each step. As a result, temporal merging introduces the idea of weighted combinations, balancing recent updates with retained past knowledge to achieve adaptability and stability—both critical for effective continual learning [31, 54, 55, 68, 97]. In this work, we study three strategies for model deployment: + +- deployyFT, which at step $t$ deploys the latest finetuned expert model, i.e. $\theta_s^O = \theta_t^S$ . +- deployEMA, which computes an unrolled exponential moving average merge over all expert models, i.e. $\theta_t^O = \theta_{t+1}^{EMA}$ following Eq. (1). +- deployALL, which applies a merging technique $f_{\text{merge}}$ over all previously computed expert models $\{\theta_{t'}^S\}_{1}^{t - 1}$ . + +# Axis 3: Merging Techniques + +At each point in time, for both initialization and deployment, merging technique $f_{\text{merge}}$ defines how to combine + +Table 1. Comparison of model merging techniques. + +
MethodSparsificationConsensusScaling
Weight averaging [63, 86]XLinear Int.Weight coeff.
SLERP [60]XSpherical Int.Weight coeff.
Task Arithmetic [25]XLinear Int.Scaling factor
MagMax [42]XMax. MagnitudeScaling factor
TIES [89]Top-kSign AgreementScaling factor
DARE-TIES [95]RandomSign AgreementScaling factor
Breadcrumbs-TIES [11]Top/Bottom-kSign AgreementScaling factor
Model Stock [27]XGeometricAdaptive ratio
LiNeS [84]XXLayer weights
+ +the available expert models and checkpoints. In this work, we study nine different variants in total, shown in Tab. 1. Denoting the number of models to merge at timestep $t$ as $M_{t}$ (with $t = 0$ and $M_{t} = M$ for standard model merging), we can define these methods as follows: + +Weight Averaging [24, 63, 68, 76, 87] simply employs a uniformly weighted, element-wise average over all models $\theta_{t,i}$ , resulting in a merge function $f_{\mathrm{merge}}^{\mathrm{WA}}$ : + +$$ +\theta_ {t} = \frac {1}{M _ {t}} \sum_ {i} \theta_ {t, i}. \tag {2} +$$ + +SLERP [60, 73] assumes weights to live on a hypersphere, and consequently conducts interpolation along a curved path connecting weight entries. In particular, for two models $\theta_{t,1}$ and $\theta_{t,2}$ deriving from some base weight $\theta_{t-1}$ and the corresponding task vectors [25] $\delta_{t,i} = \theta_{t,i} - \theta_{t-1}$ , SLERP with interpolation weight $\lambda$ is defined as + +$$ +\theta_ {t} = \theta_ {t - 1} + \frac {\sin (1 - \lambda) \Omega_ {1 , 2}}{\sin \Omega_ {1 , 2}} \cdot \delta_ {t, 1} + \frac {\sin \lambda \Omega_ {1 , 2}}{\sin \Omega_ {1 , 2}} \cdot \delta_ {t, 2} \tag {3} +$$ + +with $\Omega_{1,2}$ being the angle between task vectors $\delta_{t,1}$ and $\delta_{t,2}$ . We denote the corresponding merge function $f_{\mathrm{merge}}^{\mathrm{SLERP}}$ + +Task Arithmetic [25] defines the merge as a function over task vectors $\delta_{t,i} = \theta_{t,i} - \theta_{t-1}$ for each weight $\theta_{t,i}$ fine-tuned from $\theta_{t-1}$ . This introduces a simple merge formalism $f_{\text{merge}}^{\text{TA}}$ for weighted parameter averaging with a scale $\lambda$ : + +$$ +\theta_ {t} = \theta_ {t - 1} + \lambda \frac {1}{M _ {t}} \sum_ {i} \delta_ {t, i} \tag {4} +$$ + +TIES [89] builds on the task arithmetic formalism through controlled pruning of task vector entries with low magnitude. Moreover, the sign for each final merged parameter is set based on the sign of the highest total magnitude across the merge candidates. The final update follows basic task arithmetic, only for entries with matching signs. We refer to the respective merge function as $f_{\text{merge}}^{\text{TIES}}$ . + +DARE [95] is a similar extension of task arithmetic, but instead of targeted pruning, it randomly zeroes out task vector entries using a random mask $Z_{i} \sim \mathrm{Bernoulli}(p)$ and + +masking probability $p$ . Final task vector values for $f_{\mathrm{merge}}^{\mathrm{DARE}}$ are then rescaled based on $p$ : + +$$ +\delta_ {t, i} ^ {\mathrm {D A R E}} = \frac {(1 - Z _ {i}) \delta_ {t , i}}{1 - p}. \tag {5} +$$ + +Model Stock [27] provides a geometric extension of simple weight averaging as done in Model Soup [86] by incorporating base weights $\theta_{t-1}$ into the merging process. Given finetuned weights $\theta_{t,1}$ and $\theta_{t,2}$ , the Model Stock merge $f_{\text{merge}}^{\text{Stock}}$ is defined as follows: + +$$ +\theta_ {t} = \frac {2 \cdot \cos \Omega_ {1 , 2}}{1 + \Omega_ {1 , 2}} \cdot \left(\theta_ {t, 2} - \theta_ {t, 1}\right) + \left(1 - \frac {2 \cdot \cos \Omega_ {1 , 2}}{1 + \cos \Omega_ {1 , 2}}\right), \tag {6} +$$ + +utilizing angle $\Omega_{1,2}$ between task vectors $\delta_{t,1}$ and $\delta_{t,2}$ . + +Breadcrumbs [11] deploys another variation on task arithmetic for model merging. In particular, for a given task vector $\delta_{t,i}$ , extreme left and right tails of the absolute magnitude distribution in $\delta_{t,i}$ are zeroed out with left and right thresholds $\beta$ and $\gamma$ . The modified task vectors $\delta_{t,i}^{\mathrm{Bread}}$ are then applied on base weights $\theta_{t-1}$ following the task arithmetic setup, and giving $f_{\mathrm{merge}}^{\mathrm{Bread}}$ . + +MagMax [42] also uses task vectors—given multiple task vectors $\delta_{t,i}$ (with increments possible along both time $t$ and count axis $i$ ), the final task vector $\delta_t$ is yielded through maximum magnitude entry selection; copying the largest magnitude entries across all $\{\delta_{t,i}\}$ into $\delta_t$ , giving $f_{\mathrm{merge}}^{\mathrm{Max}}$ . + +LiNeS [84], for Layer-increasing Network Scaling, scales weight updates based on their respective layer depth enabling early layers to remain close to original pretraining weights (cf. Neyshabur et al. [48]). Given task vectors $\delta_{t,i}$ , now broken down across model layers $\delta_{t,i}^{l}$ with $l\in [1,\dots,L]$ and $L$ the number of layers, LiNeS follows the base task arithmetic merging formalism, but updates task vectors as + +$$ +\delta_ {t, i} ^ {\mathrm {L i N e S}} = \operatorname {c o n c a t} \left(\lambda^ {l = 1} \delta_ {t, i} ^ {l = 1}, \dots , \lambda^ {l = L} \delta_ {t, i} ^ {l = L}\right) \tag {7} +$$ + +with layer-scaled interpolation weights $\lambda^l = \alpha +\beta \frac{l - 1}{L - 1}$ and hyperparameters $\alpha ,\beta$ giving $f_{\mathrm{merge}}^{\mathrm{Lines}}$ + +# 3.2. Complete Temporal Model Merging Pipeline + +Incorporating all three axes of temporal model merging, we define a five-stage update pipeline for each task $t$ (see Fig. 2), consisting of the following steps: + +1. Init. Choose one of the aforementioned initialization protocols: initZS, initFT, or initEMA. This produces initialization weights $\theta_t^I$ at task $t$ . +2. Train. Given $\theta_t^I$ , train on current task data $\mathcal{D}_t$ within a set compute budget to produce the expert model $\theta_t^S$ . +3. Store. Append $\theta_t^S$ to storage of expert model weights $S$ . +4. Deploy. Choose a deployment protocol: deployFT, deployEMA, or deployALL, and produce the output weights $\theta_t^O = \mathrm{deploy}(\mathcal{S})$ . + +![](images/226214c5b586e741c060c57869bb9e89d99ccf43de286bec220f490540fb3303.jpg) +Figure 3. Offline merging methods struggle with TIME. All tested merging techniques perform extremely poorly, and are unable to adapt to the temporal setting, underperforming even a simple replay baseline that sequentially trains the base model on task-replayed data. + +5. Eval. The deployed $\theta_t^O$ is used for downstream applications and, in our case, extensive evaluation. + +Particular choices of init, deploy and $f_{\text{merge}}$ correspond to existing approaches, for example (initZS, deployALL, $f_{\text{merge}}^{\text{WA}}$ ) simply recovers offline merging through weight averaging over experts models derived from original base weights $\theta_0$ for each task $t$ . Similarly, (initFT, deployEMA, $f_{\text{merge}}^{\text{WA}}$ ) recovers exponential moving average approaches as done in [68, 76]. Using our framework, we can explore temporal model merging in a systematic manner across various combinations of all three design choices to understand their impact and recommend best practices. + +# 4. Experiments + +We first discuss experimental details in Sec. 4.1. Our initial experiments study the applicability of standard, "offline" model merging for the temporal paradigm in Sec. 4.2 and possible extensions in Sec. 4.3. We then transition to a comprehensive exploration of applicable strategies within our TIME framework in Sec. 4.4 and Sec. 4.5, before finally scaling up temporal model merging in Sec. 4.6. + +# 4.1. Experimental Pipeline Details + +To study temporal model merging in practically relevant scenarios, we use the continual pretraining benchmark Fomo-in-Flux [68]. It includes 41 adaptation datasets and 22 separate evaluation datasets, covering an array of visual and semantic distribution shifts, alongside well-defined, practically motivated compute constraints for each task. + +Training at task $t$ . By default, we continuously finetune and merge a ViT-B/16 CLIP [8, 58] model pretrained on LAION-2B [70]. The model is pretrained and finetuned using the standard CLIP objective [58]. Following [68], we fix the training steps for each task based on the DataComp-Small computation budget of $1.8 \times 10^{9}$ + +GFLOPS, split equally across 20 tasks. We also additionally explore longer sequences with $T \in \{50,100\}$ . At each temporal merging step, we allow unrestricted access to a pretraining data pool $\mathcal{P}$ , using the same 2M random subset of LAION-400M as in [68]. We use a cosine-decay LR schedule with a linear warmup of $10\%$ , AdamW optimizer [40], a batchsize of 512 and gradient norm clipping to 1. All experiments use PyTorch [53], and are run on a compute cluster using NVIDIA A100/H100s. + +Checkpoint Storage. To enable arbitrary merging strategies at any task $t$ , we store each trained expert $\theta_t^S$ . We do not store any merged initialization weights $\theta_t^I$ or deployed models $\theta_t^O$ , as these can be recovered from stored experts. + +Evaluation and Metrics. We focus on two key quantities to evaluate temporal model merging: the level of adaptation, reflecting performance improvement with each merging step, and retention, capturing the preservation of prior knowledge. Specifically, we report two metrics following Roth et al. [68]: Knowledge Accumulation $(\mathcal{A}_{KA})$ , the average accuracy (or recall@5 for retrieval) across all 41 adaptation datasets, and Zero-Shot Retention $(\mathcal{A}_{ZS})$ , the zero-shot accuracy or recall@5 on all 22 held-out evaluation datasets. Additional details, including a description of our plotting style, can be found in the supplementary. + +# 4.2. Do We Need Model Merging Across Time? + +The simplest approach to tackle temporal merging is to disregard the time axis and follow standard offline merging paradigms. In TIME terms, this corresponds to a configuration of (init_ZS, deploy_ALL, $f_{\text{merge}}$ ), which always fine-tunes the initial base weights $\theta_0$ . To study the effectiveness of such an offline strategy, we test the above triplet with various choices of $f_{\text{merge}}$ , including $f_{\text{merge}}^{\text{WA}}$ (averaging in Fig. 3), $f_{\text{merge}}^{\text{TA}}$ (task-arithmetic), $f_{\text{merge}}^{\text{Max}}$ (magmax), $f_{\text{merge}}^{\text{TIES}}$ (ties), $f_{\text{merge}}^{\text{DARE}}$ applied over TIES (dare-ties), $f_{\text{merge}}^{\text{Bread}}$ (breadcrumbs-ties), and $f_{\text{merge}}^{\text{Lines}}$ (lines-ties). To put the results + +![](images/ebd093cda73193ac5477e477267bffeaa313b54c949aebe43f0656e36ffbd749.jpg) +Figure 4. Improving offline merging. We identify two simple methods for adapting offline-merging methods to the temporal setting: (1) replaying data from previous tasks (best-(offline+replay)) and (2) recency-biased weighting of task checkpoints (best-(offline+replay+weighting)). With these method improvements, offline merging methods can match the replay baseline. + +![](images/a3bed99fb67c88813c27ba41aefb99d45e29aed54c8d3f2fba5ab1140fd70a8d.jpg) + +![](images/2b76313ce61b1817010d8582c8fa3585db7e2df3537da818899c92169cc466d7.jpg) + +in context, we incorporate (1) a simple continual fine-tuning baseline (replay), following Sec. 4.1, which replays on both pretraining and previous task data, (2) initial zero-shot $(\theta_0)$ performance lower bound, and (3) multitask training upper bound. We visualize trajectories over time for knowledge accumulation $\mathcal{A}_{\mathrm{KA}}$ , zero-shot retention $\mathcal{A}_{\mathrm{ZS}}$ , and the geometric mean of both in Fig. 3. Our results clearly highlight that there are only marginal differences between merging techniques $f_{\mathrm{merge}}$ when deployed in an offline manner for a temporal problem, and they all trace similar trajectories in the $\mathcal{A}_{\mathrm{KA}}$ and $\mathcal{A}_{\mathrm{ZS}}$ space and achieve similar final performance. We do find that the vanilla $f_{\mathrm{merge}}^{\mathrm{TA}}$ (task-arithmetic) starts to deviate after a few update steps, an issue fixed by the various subsequent extensions such as TIES, DARE or LiNeS. Overall, however, unlike straightforward continual fine-tuning (replay), offline merging with any technique fails to address the temporal aspects of the problem, particularly struggling to consistently acquire new knowledge over time (as observed in Fig. 3, left). + +Key takeaway: offline merging with any technique yields similar results, failing to add new knowledge over time. + +# 4.3. Can Replay or Time-Weighting Fix the Issue? + +We next ask: "What are some simple extensions to offline methods to close the gap to the replay baseline"? As the continual fine-tuning baseline replays on past data from all previous tasks while training at the current task $t$ , can this task data-mixing also help offline merging methods? + +Data replay improves offline merging. Since offline methods operate entirely under a task-independent assumption, they fail to capture any temporal dependencies. Fig. 4 shows that simply applying data-replay on top of standard offline merging leads to significant boosts in the overall performance. For instance, best-(offline+replay) achieves $58.2\%$ compared to best-offline at $54.6\%$ , bringing it closer to the replay baseline. However, a notable performance gap + +remains, with best-(offline+replay) at $58.2\%$ falling short of replay at $59.1\%$ . + +Recency-biased weighting helps. Next, unlike in standard offline averaging, where all task checkpoints are weighted uniformly, we impose temporal ordering via non-uniform weighting for offline merging. We explore several recency-biased, non-uniform weighting schemes, assigning higher weights to more recent tasks to account for the temporal nature of the setting. + +We explore various discounting schemes: logarithmic, quadratic, exponential, and cubic, applied to the best offline merge replay method from the previous experiment (please refer to the supplementary for details). As shown in Fig. 4, these schemes improve performance, with best-offline+replay+weighting) reaching $58.9\%$ yet still falling slightly short of the replay baseline at $59.1\%$ . These results provide strong evidence that accounting for the new temporal axis is crucial for effective temporal model merging, even when implemented as an extension of offline merging. + +Key Takeaway: Accounting for the time aspect is crucial for effective temporal model merging, even as an extension on top of standard offline merging. Still, a small gap to the simple replay baseline remains. + +# 4.4. TIME Travel + +In our effort to adapt offline merging to the temporal setting, we have explored two axes of our TIME framework—merging technique and deployment. As discussed in Sec. 4.2, the choice of merging technique has minimal impact on final performance, while the deployment strategy plays a much more significant role. Next, we systematically explore the design space for temporal merging by testing all valid combinations of three initialization protocols and three deployment protocols described in Sec. 3.2. After discarding incompatible pairs, such as init_ZS with deployyFT, we evaluated the remaining eight variants us + +![](images/639e655301947a8811b4e6988ed89b22f627cbed4850befffae695104730f1dc.jpg) +Figure 5. A journey through TIME. We explore various initialization and deployment protocols, finding that the EMA initialization-deployment strikes the best balance between knowledge accumulation and zero-shot retention. We refer to this strategy as Best-in-TIME. + +![](images/9018a050b489fbfba94241270d6ae30891521d44371f681a3b6be35510469d5f.jpg) + +![](images/f7eaf2208b1b1b57e2c95e02a525818266bb39ebd475943f00c1e3a39c631018.jpg) + +![](images/f9e3177a66ec2ed32b5e3807233b2ce5d05124ee1757d422ac950a2703ffb210.jpg) +Figure 6. Sweeping Best-in-TIME. All merging techniques perform well with the Best-in-TIME strategy. Indeed, there are no significant differences between techniques, indicating that initialization and deployment matter more for temporal merging. + +![](images/1dd6998220954e558705d17b1fb70bccbb96fab789a5e4db0bd6c00a66bacaed.jpg) + +![](images/0f3a09fe0ac769d0af23e33f4f49b5b1dacd65324835466a103972dd2328c2f0.jpg) + +ing weight averaging as the merging technique. As shown in Fig. 5, the choice of initialization and deployment strategy largely determines performance, significantly affecting both knowledge accumulation and retention. One combination that stands out consistently is initEMA alongside deployyEMA. This supports the findings of [68, 76] on the applicability of EMA-like solutions to small-scale continual learning and pretraining, while extending the analysis to a much larger, more systematic study of temporal model merging. + +Since EMA-style model merging achieves a better balance between knowledge accumulation and retention than other methods, we call this approach Best-in-TIME. In the next section, we will explore the robustness of this strategy across different merging techniques. + +Key Takeaway: The EMA strategy for initialization and deployment consistently performs the best. + +# 4.5. What is the best merge for Best-in-TIME? + +Having identified the optimal initialization and deployment merging strategy in the previous section using weight-averaging as a canonical merging technique, we now invest + +tigate the robustness of our finding by sweeping over other merging techniques, similar to the offline-merging analysis in Sec. 4.2. In particular, we test 7 different merging techniques while keeping the Best-in-TIME initialization and deployment strategy. From Fig. 6, it is immediately evident that all merging techniques perform very similarly—in fact, the knowledge accumulation $(\mathcal{A}_{KA})$ of all techniques lie within a single standard deviation of performance variation across multiple runs. This indicates that, over a sufficiently long time horizon, all techniques converge to a similar behavior, echoing our results in Sec. 4.2. However, we do notice higher variance in the retention metric $(\mathcal{A}_{ZS})$ . Overall, our results suggest that for temporal model merging, the specific merging technique used is much less important than selecting the best initialization and deployment strategies. + +Key Takeaway: The choice of merging technique is not particularly important—all techniques perform similarly. + +# 4.6. Scaling Up Temporal Model Merging + +We next study the scaling behaviour of temporal model merging across three-axes: model size, compute budget, and number of tasks (results in Fig. 7 and supplen + +![](images/ce61e309977d426698779fdbea703da737579f65da74ee43dd42edb9c34de121.jpg) +Figure 7. Scaling up model merging. (left) With scale, we observe continued improvements of model merging compared to the standard replay baseline. (right) Our Best-in-TIME method continues to improve with scaled total compute budget moving close to the multitask upper-bound. Gray points in the plot visualize suboptimal Best-in-TIME hyperparameter-instantiations. + +![](images/996c1b71b013f13effc7c881444e329b4560e24441b189ecfa91c6da5f5592c9.jpg) + +tary). All our experiments utilize the Best-in-TIME setup described previously, conducting hyperparameter-optimal EMA-weight-averaging at each task. + +Scaling the Model. As we increase the model scale from $S / 16$ (62.3M parameters) to $B / 16$ (149.6M), $L / 14$ (427.6M), and finally $g / 14$ (1.37B) in Fig. 7 (left), we compare sequential fine-tuning (circles) and Best-in-TIME (squares) in terms of knowledge accumulation and retention. We find that: (1) Best-in-TIME scales well with model size, with larger models exhibiting increased affinity to temporal merging. This extends and further corroborates standard offline merging insights by Yadav et al. [89], who showed that model scale facilitates merging. (2) Moreover, while Roth et al. [68] highlight better continual fine-tuning with scale (as also studied by Ibrahim et al. [23]), we show temporal model merging to be substantially more effective across scale. For larger models all the way to the largest ViT-g/14, Best-in-TIME vastly outperforms or matches sequential fine-tuning and the multitask target in knowledge retention and positive backward transfer. Moreover, scale facilitates equivalent degrees of knowledge accumulation between sequential fine-tuning and temporal model merging. Therefore, our model scaling results strongly advocate the use of temporal model merging solutions over standard continual fine-tuning methods [23, 68]. + +Scaling the Compute. Keeping the model fixed to ViTB/16, we next scale up the available compute budget by increasing the number of allocated update steps per task. We compare a multitask model, trained on all tasks simultaneously, to a budget-optimal Best-in-TIME. The only hyperparameter for Best-in-TIME is the interpolation weight $w$ . For each compute budget, there is a clear optimal choice of that hyperparameter (suboptimal runs shown as gray dots in Fig. 7 (right)). Higher values of $w$ put greater emphasis on accumulation, allowing optimal accumulation-retention + +trade-offs to be reached at lower compute budgets. However, if a larger compute budget is available, less aggressive temporal model merging can achieve higher absolute trade-offs. Note that in Fig. 7 (right), we report the geometric mean between accumulation and retention, corresponding to the right-most panel in previous plots. Our results indicate that Best-in-TIME scales very well across compute budgets, clearly approaching the multitask upper bound in accumulation-retention balance at larger compute budgets. Put together, this gives practitioners the flexibility to adjust Best-in-TIME to their compute budget through the choice of $w$ . Moreover, with more compute to train each respective expert model, the merging process becomes more effective. + +Scaling the Number of Tasks. We next study how Best-in-TIME performs as we increase the number of merging time-steps from $T = 20$ to $T = 50$ and $T = 100$ . As shown in Figs. 10 and 11, Best-in-TIME remains the optimal method across different initialization and deployment strategies. For $T = 50$ and $T = 100$ , we notice that other methods are initially competitive, but face a significant drop-off in performance near the end of the task sequence, while best-in-TIME continues to improve monotonically. This further corroborates that our Best-in-TIME is an extremely scalable method for temporal model merging. + +Key Takeaway: Temporal model merging with Best-in-TIME scales efficiently across model sizes and the number of tasks. Moreover, compute scaling improves the effectiveness of temporal model merging. + +# 5. Conclusion + +In this work, we study temporal model merging, addressing the challenge of continually merging multimodal models as new tasks and data arrive, and new expert models are trained in succession. To formalize this setting, we propose TIME, a novel unifying framework breaking down temporal model merging into three key axes: (1) initialization phase defining starting weights before each task, (2) deployment phase denoting post-training expert model aggregation, and (3) the choice of merging technique. Using TIME, we conduct a large-scale systematic study uncovering crucial practical guidelines for temporal model merging. Our experiments on the FoMo-in-Flux benchmark spanning 63 datasets, showcase that accounting for the temporal aspect is crucial, with standard offline merging techniques falling short in this dynamic setting. Moreover, we find the particular choice of merging technique matters far less than the merging strategy for initialization and deployment. Finally, we introduce Best-in-TIME, which scales favorably with model size and outperforms existing methods for continual multimodal pretraining. Together, our work provides a systematic entry point into temporal model merging and establishes best practices for this emerging field. + +Acknowledgements. The authors would like to thank (in random order) Shyamgopal Karthik, Shashwat Goel, Ankit Sonthalia, Olivier Henaff, Alexandre Ramé, and Daniel Marczak for helpful feedback. The plotting style in our work is inspired by figures from Beyer et al. [4]. The style of Fig. 2 is inspired by Figure 1 of Karamcheti et al. [30]. VU, KR, and SD thank the International Max Planck Research School for Intelligent Systems (IMPRS-IS). VU, KR, and SD also thank the European Laboratory for Learning and Intelligent Systems (ELLIS) PhD program for support. VU is supported by a Google PhD Fellowship in Machine Intelligence. SA is supported by a Newton Trust Grant. AP is supported by the Federal Ministry of Education and Research (BMBF), FKZ: 011524085B. MB acknowledges financial support via the Open Philanthropy Foundation funded by the Good Ventures Foundation. MB is a member of the Machine Learning Cluster of Excellence, funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany's Excellence Strategy - EXC number 2064/1 - Project number 390727645. ZA acknowledges the support from the German Research Foundation (DFG): SFB 1233, Robust Vision: Inference Principles and Neural Mechanisms, project number: 276693517 and ERC Grant DEXIM, project number: 853489. This research utilized compute resources at the Tübingen Machine Learning Cloud, DFG FKZ INST 37/1057-1 FUGG. + +# References + +[1] Samuel Ainsworth, Jonathan Hayase, and Siddhartha Srinivasa. Git re-basin: Merging models modulo permutation symmetries. In The Eleventh International Conference on Learning Representations, 2023. 2 +[2] Takuya Akiba, Makoto Shing, Yujin Tang, Qi Sun, and David Ha. Evolutionary optimization of model merging recipes. arXiv preprint arXiv:2403.13187, 2024. 2 +[3] Anton Alexandrov, Veselin Raychev, Mark Niklas Müller, Ce Zhang, Martin Vechev, and Kristina Toutanova. Mitigating catastrophic forgetting in language transfer via model merging. arXiv preprint arXiv:2407.08699, 2024. 2 +[4] Lucas Beyer, Xiaohua Zhai, Amélie Royer, Larisa Markeeva, Rohan Anil, and Alexander Kolesnikov. Knowledge distillation: A good teacher is patient and consistent. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10925-10934, 2022. 9, 14 +[5] Rishi Bommasani, Drew A. Hudson, Ehsan Adeli, Russ B. Altman, Simran Arora, Sydney von Arx, Michael S. Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, Erik Brynjolfsson, Shyamal Buch, Dallas Card, Rodrigo Castellon, Niladri S. Chatterji, Annie S. Chen, Kathleen Creel, Jared Quincy Davis, Dorottya Demszky, Chris Donahue, Moussa Doumbouya, Esin Durmus, Stefano Ermon, John Etchemendy, Kawin Ethayarajh, Li Fei-Fei, Chelsea Finn, Trevor Gale, Lauren Gillespie, Karan Goel, Noah D. Goodman, Shelby Grossman, Neel Guha, Tat + +sunori Hashimoto, Peter Henderson, John Hewitt, Daniel E. Ho, Jenny Hong, Kyle Hsu, Jing Huang, Thomas Icard, Saahil Jain, Dan Jurafsky, Pratyusha Kalluri, Siddharth Karamcheti, Geoff Keeling, Fereshte Khani, Omar Khattab, Pang Wei Koh, Mark S. Krass, Ranjay Krishna, Rohith Kuditipudi, and et al. On the opportunities and risks of foundation models. CoRR, abs/2108.07258, 2021. 1 +[6] Jorg Bornschein, Alexandre Galashov, Ross Hemsley, Amal Rannen-Triki, Yutian Chen, Arslan Chaudhry, Xu Owen He, Arthur Douillard, Massimo Caccia, Qixuan Feng, et al. Nevis' 22: A stream of 100 tasks sampled from 30 years of computer vision research. JMLR, 2023. 2 +[7] Zhipeng Cai, Ozan Sener, and Vladlen Koltun. Online continual learning with natural distribution shifts: An empirical study with visual data. In ICCV, 2021. 2 +[8] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. arXiv preprint arXiv:2212.07143, 2022. 5 +[9] Leshem Choshen, Elad Venezian, Shachar Don-Yehia, Noam Slonim, and Yoav Katz. Where to start? analyzing the potential value of intermediate models. arXiv preprint arXiv:2211.00107, 2022. 2 +[10] Geoffrey Cideron, Andrea Agostinelli, Johan Ferret, Sertan Girgin, Romuald Elie, Olivier Bachem, Sarah Perrin, and Alexandre Rame. Diversity-rewarded CFG distillation. arXiv preprint arXiv:2410.06084, 2024. 2 +[11] MohammadReza Davari and Eugene Belilovsky. Model breadcrumbs: Scaling multi-task model merging with sparse masks. In Computer Vision - ECCV 2024, pages 270-287, Cham, 2025. Springer Nature Switzerland. 2, 4 +[12] Pala Tej Deep, Rishabh Bhardwaj, and Soujanya Poria. Della-merging: Reducing interference in model merging through magnitude-based sampling. arXiv preprint arXiv:2406.11617, 2024. 2 +[13] Shachar Don-Yehiya, Elad Venezian, Colin Raffel, Noam Slonim, Yoav Katz, and Leshem Choshen. Cold fusion: Collaborative descent for distributed multitask finetuning. arXiv preprint arXiv:2212.01378, 2022. 1, 2 +[14] Jonathan Frankle, Gintare Karolina Dziugaite, Daniel Roy, and Michael Carbin. Linear mode connectivity and the lottery ticket hypothesis. In Proceedings of the 37th International Conference on Machine Learning, pages 3259-3269. PMLR, 2020. 2 +[15] João Gama, Indre Žliobaité, Albert Bifet, Mykola Pechenizkiy, and Abdelhamid Bouchachia. A survey on concept drift adaptation. ACM computing surveys, 2014. 3 +[16] Saurabh Garg, Mehrdad Farajtabar, Hadi Pouransari, Raviteja Vemulapalli, Sachin Mehta, Oncel Tuzel, Vaishaal Shankar, and Fartash Faghri. Tic-clip: Continual training of clip models. In ICLR, 2024. 2 +[17] Timur Garipov, Pavel Izmailov, Dmitrii Podoprikhin, Dmitry P Vetrov, and Andrew G Wilson. Loss surfaces, mode connectivity, and fast ensembling of dnns. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2018. 1 + +[18] Charles Goddard, Shamane Siriwardhana, Malikeh Ehghaghi, Luke Meyers, Vlad Karpukhin, Brian Benedict, Mark McQuade, and Jacob Solawetz. Arcee's mergekit: A toolkit for merging large language models. arXiv preprint arXiv:2403.13257, 2024. 2 +[19] Alexey Gorbatovski, Boris Shaposhnikov, Alexey Malakhov, Nikita Surnachev, Yaroslav Aksenov, Ian Maksimov, Nikita Balagansky, and Daniil Gavrilov. Learn your reference model for real good alignment. arXiv preprint arXiv:2404.09656, 2024. 2 +[20] Zhongrui Gui, Shuyang Sun, Runjia Li, Jianhao Yuan, Zhaochong An, Karsten Roth, Ameya Prabhu, and Philip Torr. knn-clip: Retrieval enables training-free segmentation on continually expanding large vocabularies. arXiv preprint arXiv:2404.09447, 2024. 1, 2 +[21] Yifei He, Yuzheng Hu, Yong Lin, Tong Zhang, and Han Zhao. Localize-and-stitch: Efficient model merging via sparse task arithmetic. arXiv preprint arXiv:2408.13656, 2024. 2 +[22] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 1 +[23] Adam Ibrahim, Benjamin Thérien, Kshitij Gupta, Mats L Richter, Quentin Anthony, Timothée Lesort, Eugene Belilovsky, and Irina Rish. Simple and scalable strategies to continually pre-train large language models. arXiv preprint arXiv:2403.08763, 2024. 2, 8 +[24] Gabriel Ilharco, Mitchell Wortsman, Samir Yitzhak Gadre, Shuran Song, Hannaneh Hajishirzi, Simon Kornblith, Ali Farhadi, and Ludwig Schmidt. Patching open-vocabulary models by interpolating weights. In NeurIPS, 2022. 1, 2, 4, 14 +[25] Gabriel Ilharco, Marco Tulio Ribeiro, Mitchell Wortsman, Ludwig Schmidt, Hannaneh Hajishirzi, and Ali Farhadi. Editing models with task arithmetic. In The Eleventh International Conference on Learning Representations, 2023. 2, 4, 14 +[26] Pavel Izmailov, Dmitrii Podoprikhin, Timur Garipov, Dmitry Vetrov, and Andrew Gordon Wilson. Averaging weights leads to wider optima and better generalization. arXiv preprint arXiv:1803.05407, 2018. 2 +[27] Dong-Hwan Jang, Sangdoo Yun, and Dongyoon Han. Model stock: All we need is just a few fine-tuned models. In Computer Vision - ECCV 2024: 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part XLIV, page 207-223, Berlin, Heidelberg, 2024. Springer-Verlag. 1, 4 +[28] Xisen Jin, Xiang Ren, Daniel Preotiac-Pietro, and Pengxiang Cheng. Dataless knowledge fusion by merging weights of language models. In The Eleventh International Conference on Learning Representations, 2023. 2 +[29] Jean Kaddour. Stop wasting my time! saving days of imageet and bert training with latest weight averaging. arXiv preprint arXiv:2209.14981, 2022. 2 +[30] Siddharth Karamcheti, Suraj Nair, Ashwin Balakrishna, Percy Liang, Thomas Kollar, and Dorsa Sadigh. + +Prismatic vlms: Investigating the design space of visually-conditioned language models. arXiv preprint arXiv:2402.07865, 2024.9 +[31] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. PNAS, 2017. 3 +[32] Pang Wei Koh, Shiori Sagawa, Henrik Marklund, Sang Michael Xie, Marvin Zhang, Akshay Balsubramani, Weihua Hu, Michihiro Yasunaga, Richard Lanas Phillips, Irena Gao, et al. Wilds: A benchmark of in-the-wild distribution shifts. In International conference on machine learning, pages 5637-5664. PMLR, 2021. 1 +[33] Jeffrey Li, Mohammadreza Armandpour, Seyed Iman Mirzadeh, Sachin Mehta, Vaishaal Shankar, Raviteja Vemulapalli, Oncel Tuzel, Mehrdad Farajtabar, Hadi Pouransari, and Fartash Faghri. Tic-lm: A multi-year benchmark for continual pretraining of language models. In NeurIPS 2024 Workshop on Scalable Continual Learning for Lifelong Foundation Models. 2 +[34] Margaret Li, Suchin Gururangan, Tim Dettmers, Mike Lewis, Tim Althoff, Noah A Smith, and Luke Zettlemoyer. Branch-train-merge: Embarrassingly parallel training of expert language models. arXiv preprint arXiv:2208.03306, 2022. 1 +[35] Tao Li, Weisen Jiang, Fanghui Liu, Xiaolin Huang, and James T Kwok. Learning scalable model soup on a singlegpu: An efficient subspace training strategy. arXiv preprint arXiv:2407.03641, 2024. 2 +[36] Yong Lin, Hangyu Lin, Wei Xiong, Shizhe Diao, Jianmeng Liu, Jipeng Zhang, Rui Pan, Haoxiang Wang, Wenbin Hu, Hanning Zhang, Hanze Dong, Renjie Pi, Han Zhao, Nan Jiang, Heng Ji, Yuan Yao, and Tong Zhang. Mitigating the alignment tax of RLHF. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 580-606, Miami, Florida, USA, 2024. Association for Computational Linguistics. 2 +[37] Zhiqiu Lin, Jia Shi, Deepak Pathak, and Deva Ramanan. The clear benchmark: Continual learning on real-world imagery. In NeurIPS, 2021. 2 +[38] Adam Liska, Tomas Kocisky, Elena Gribovskaya, Tayfun Terzi, Eren Sezener, Devang Agrawal, D'Autume Cyprien De Masson, Tim Scholtes, Manzil Zaheer, Susannah Young, et al. Streamingqa: A benchmark for adaptation to new knowledge over time in question answering models. In International Conference on Machine Learning, pages 13604-13622. PMLR, 2022. 2 +[39] Tianlin Liu, Shangmin Guo, Leonardo Bianco, Daniele Calandriello, Quentin Berthet, Felipe Llinares-López, Jessica Hoffmann, Lucas Dixon, Michal Valko, and Mathieu Blondel. Decoding-time realignment of language models. In Proceedings of the 41st International Conference on Machine Learning, pages 31015-31031. PMLR, 2024. 2 +[40] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5 +[41] Jinliang Lu, Ziliang Pang, Min Xiao, Yaochen Zhu, Rui Xia, and Jiajun Zhang. Merge, ensemble, and cooperate! + +a survey on collaborative strategies in the era of large language models. arXiv preprint arXiv:2407.06089, 2024. 2 +[42] Daniel Marczak, Bartlomiej Twardowski, Tomasz Trzeciński, and Sebastian Cygert. Magmax: Leveraging model merging for seamless continual learning. arXiv preprint arXiv:2407.06322, 2024. 2, 4 +[43] Michael S Matena and Colin Raffel. Merging models with fisher-weighted averaging. In Advances in Neural Information Processing Systems, 2022. 2 +[44] Matías Mendieta, Boran Han, Xingjian Shi, Yi Zhu, and Chen Chen. Towards geospatial foundation models via continual pretraining. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16806-16816, 2023. 2 +[45] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. In The Eleventh International Conference on Learning Representations, 2023. 1 +[46] Remi Munos, Michal Valko, Daniele Calandriello, Mohammad Gheshlaghi Azar, Mark Rowland, Zhaohan Daniel Guo, Yunhao Tang, Matthieu Geist, Thomas Mesnard, Come Fiegel, Andrea Michi, Marco Selvi, Sertan Girgin, Nikola Momchev, Olivier Bachem, Daniel J Mankowitz, Doina Precup, and Bilal Piot. Nash learning from human feedback. In Proceedings of the 41st International Conference on Machine Learning, pages 36743-36768. PMLR, 2024. 2 +[47] Anshul Nasery, Jonathan Hayase, Pang Wei Koh, and Sewoong Oh. Pleas-merging models with permutations and least squares. arXiv preprint arXiv:2407.02447, 2024. 2 +[48] Behnam Neyshabur, Hanie Sedghi, and Chiyuan Zhang. What is being transferred in transfer learning? In Advances in Neural Information Processing Systems, pages 512-523. Curran Associates, Inc., 2020. 2, 4 +[49] Kai Nylund, Suchin Gururangan, and Noah A Smith. Time is encoded in the weights of finetuned language models. arXiv preprint arXiv:2312.13401, 2023. 2 +[50] Changdaee Oh, Yixuan Li, Kyungwoo Song, Sangdoo Yun, and Dongyoon Han. Dawn: Training-free dynamic weight interpolation for robust adaptation. arXiv preprint arXiv:2410.03782, 2024. 2 +[51] Oleksiy Ostapenko, Timothee Lesort, Pau Rodríguez, Md Rifat Arefin, Arthur Douillard, Irina Rish, and Laurent Charlin. Continual learning with foundation models: An empirical study of latent replay. In Conference on Lifelong Learning Agents (CoLLAs), 2022. 2 +[52] Jyothish Pari, Samy Jelassi, and Pulkit Agrawal. Collective model intelligence requires compatible specialization. arXiv preprint arXiv:2411.02207, 2024. 2 +[53] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 5 +[54] Ameya Prabhu, Zhipeng Cai, Puneet Dokania, Philip Torr, Vladlen Koltun, and Ozan Sener. Online contin + +ual learning without the storage constraint. arXiv preprint arXiv:2305.09253, 2023. 2, 3 +[55] Ameya Prabhu, Hasan Abed Al Kader Hammoud, Puneet Dokania, Philip HS Torr, Ser-Nam Lim, Bernard Ghanem, and Adel Bibi. Computationally budgeted continual learning: What does matter? In CVPR, 2023. 2, 3 +[56] Ameya Prabhu, Hasan Abed Al Kader Hammoud, Ser-Nam Lim, Bernard Ghanem, Philip HS Torr, and Adel Bibi. From categories to classifier: Name-only continual learning by exploring the web. arXiv preprint arXiv:2311.11293, 2023. 2 +[57] Sarah Pratt, Ian Covert, Rosanne Liu, and Ali Farhadi. What does a platypus look like? generating customized prompts for zero-shot image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 15691-15701, 2023. 1 +[58] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5 +[59] Alexandre Rame, Kartik Ahuja, Jianyu Zhang, Matthieu Cord, Leon Bottou, and David Lopez-Paz. Model rata-touille: Recycling diverse models for out-of-distribution generalization. In Proceedings of the 40th International Conference on Machine Learning, pages 28656-28679. PMLR, 2023. 2 +[60] Alexandre Ramé, Johan Ferret, Nino Vieillard, Robert Dadashi, Léonard Hussenot, Pierre-Louis Cedoz, Pier Giuseppe Sessa, Sertan Girgin, Arthur Douillard, and Olivier Bachem. Warp: On the benefits of weight averaged rewarded policies. arXiv preprint arXiv:2406.16768, 2024. 2, 4 +[61] Alexandre Ramé, Matthieu Kirchmeyer, Thibaud Rahier, Alain Rakotomamonjy, Patrick Gallinari, and Matthieu Cord. Diverse weight averaging for out-of-distribution generalization. In Proceedings of the 36th International Conference on Neural Information Processing Systems, 2024. 2 +[62] Alexandre Rame, Nino Vieillard, Léonard Hussenot, Robert Dadashi, Geoffrey Cideron, Olivier Bachem, and Johan Ferret. Warm: On the benefits of weight averaged reward models. arXiv preprint arXiv:2401.12187, 2024. 2 +[63] Regent's, ParkLondon, Ukj, and . Utans. Weight averaging for neural networks and local resampling. 1996. 1, 2, 4 +[64] Mark Rofin, Nikita Balagansky, and Daniil Gavrilov. Linear interpolation in parameter space is good enough for finetuned language models. arXiv preprint arXiv:2211.12092, 2022. 1 +[65] Karsten Roth, Jae Myung Kim, A. Sophia Koepke, Oriol Vinyals, Cordelia Schmid, and Zeynep Akata. Waffling around for performance: Visual classification with random words and broad concepts. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 15746-15757, 2023. 1 +[66] Karsten Roth, Zeynep Akata, Dima Damen, Ivana + +Balažević, and Olivier J Henaff. Context-aware multimodal pretraining. arXiv preprint arXiv:2411.15099, 2024. 1 +[67] Karsten Roth, Lukas Thede, A. Sophia Koepke, Oriol Vinyals, Olivier J Henaff, and Zeynep Akata. Fantastic gains and where to find them: On the existence and prospect of general knowledge transfer between any pretrained model. In The Twelfth International Conference on Learning Representations, 2024. 2 +[68] Karsten Roth, Vishaal Udandarao, Sebastian Dziadzio, Ameya Prabhu, Mehdi Cherti, Oriol Vinyals, Olivier Henaff, Samuel Albanie, Matthias Bethge, and Zeynep Akata. A practitioner's guide to continual multimodal pretraining. arXiv preprint arXiv:2408.14471, 2024. 1, 2, 3, 4, 5, 7, 8, 14 +[69] Sunny Sanyal, Atula Tejaswi Neerkaje, Jean Kaddour, Abhishek Kumar, and sujay sanghavi. Early weight averaging meets high learning rates for LLM pre-training. In First Conference on Language Modeling, 2024. 1, 2 +[70] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278–25294, 2022. 5 +[71] Ekansh Sharma, Daniel M Roy, and Gintare Karolina Dziuagaite. The non-local model merging problem: Permutation symmetries and variance collapse. arXiv preprint arXiv:2410.12766, 2024. 2 +[72] Li Shen, Anke Tang, Enneng Yang, Guibing Guo, Yong Luo, Lefei Zhang, Xiaochun Cao, Bo Du, and Dacheng Tao. Efficient and effective weight-ensembling mixture of experts for multi-task model merging. arXiv preprint arXiv:2410.21804, 2024. 2 +[73] Ken Shoemake. Animating rotation with quaternion curves. Proceedings of the 12th annual conference on Computer graphics and interactive techniques, 1985. 4 +[74] Shikhar Srivastava, Md Yousuf Harun, Robik Shrestha, and Christopher Kanan. Improving multimodal large language models using continual learning. arXiv preprint arXiv:2410.19925, 2024. 2 +[75] George Stoica, Pratik Ramesh, Boglarka Ecsedi, Leshem Choshen, and Judy Hoffman. Model merging with svd to tie the knots. arXiv preprint arXiv:2410.19735, 2024. 2 +[76] Zafir Stojanovski, Karsten Roth, and Zeynep Akata. Momentum-based weight interpolation of strong zero-shot models for continual learning. arXiv preprint arXiv:2211.03186, 2022. 2, 3, 4, 5, 7 +[77] Yi-Lin Sung, Linjie Li, Kevin Lin, Zhe Gan, Mohit Bansal, and Lijuan Wang. An empirical study of multimodal model merging. arXiv preprint arXiv:2304.14933, 2023. 2 +[78] Derek Tam, Mohit Bansal, and Colin Raffel. Merging by matching models in task parameter subspaces. Transactions on Machine Learning Research, 2024. 2 +[79] Derek Tam, Yash Kant, Brian Lester, Igor Gilitschenski, and Colin Raffel. Realistic evaluation of model merging for compositional generalization. arXiv preprint arXiv:2409.18314, 2024. 2 + +[80] Lukas Thede, Karsten Roth, Olivier J Henaff, Matthias Bethge, and Zeynep Akata. Reflecting on the state of rehearsal-free continual learning with pretrained models. arXiv preprint arXiv:2406.09384, 2024. 2 +[81] Vishaal Udandarao, Ankush Gupta, and Samuel Albanie. Sus-x: Training-free name-only transfer of vision-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2725-2736, 2023. 1 +[82] Vishaal Udandarao, Nikhil Parthasarathy, Muhammad Ferjad Naeem, Talfan Evans, Samuel Albanie, Federico Tombari, Yongqin Xian, Alessio Tonioni, and Olivier J Henaff. Active data curation effectively distills large-scale multimodal models. arXiv preprint arXiv:2411.18674, 2024. +[83] Vishaal Udandarao, Ameya Prabhu, Adhiraj Ghosh, Yash Sharma, Philip HS Torr, Adel Bibi, Samuel Albanie, and Matthias Bethge. No “zero-shot” without exponential data: Pretraining concept frequency determines multimodal model performance. arXiv preprint arXiv:2404.04125, 2024. 1 +[84] Ke Wang, Nikolaos Dimitriadis, Alessandro Favero, Guillermo Ortiz-Jimenez, Francois Fleuret, and Pascal Frossard. Lines: Post-training layer scaling prevents forgetting and enhances model merging. arXiv preprint arXiv:2410.17146, 2024. 4, 18 +[85] Ke Wang, Nikolaos Dimitriadis, Guillermo Ortiz-Jimenez, François Fleuret, and Pascal Frossard. Localizing task information for improved model merging and compression. arXiv preprint arXiv:2405.07813, 2024. 2 +[86] Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, and Ludwig Schmidt. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In Proceedings of the 39th International Conference on Machine Learning, pages 23965-23998. PMLR, 2022. 1, 4 +[87] Mitchell Wortsman, Gabriel Ilharco, Jong Wook Kim, Mike Li, Simon Kornblith, Rebecca Roelofs, Raphael Gontijo Lopes, Hannaneh Hajishirzi, Ali Farhadi, Hongseok Namkoong, and Ludwig Schmidt. Robust fine-tuning of zero-shot models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7959-7971, 2022. 1, 3, 4 +[88] Feng Xiong, Runxi Cheng, Wang Chen, Zhanqiu Zhang, Yiwen Guo, Chun Yuan, and Ruifeng Xu. Multi-task model merging via adaptive weight disentanglement. arXiv preprint arXiv:2411.18729, 2024. 2 +[89] Prateek Yadav, Derek Tam, Leshem Choshen, Colin Raffel, and Mohit Bansal. TIES-merging: Resolving interference when merging models. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 1, 2, 3, 4, 8, 14 +[90] Prateek Yadav, Colin Raffel, Mohammed Muqeeth, Lucas Caccia, Haokun Liu, Tianlong Chen, Mohit Bansal, Leshem Choshen, and Alessandro Sordoni. A survey on + +model moering: Recycling and routing among specialized experts for collaborative learning. arXiv preprint arXiv:2408.07057, 2024. 2 +[91] Prateek Yadav, Tu Vu, Jonathan Lai, Alexandra Chronopoulou, Manaal Faruqui, Mohit Bansal, and Tsendsuren Munkhdalai. What matters for model merging at scale? arXiv preprint arXiv:2410.03617, 2024. 2, 3 +[92] Enneng Yang, Li Shen, Guibing Guo, Xingwei Wang, Xiaochun Cao, Jie Zhang, and Dacheng Tao. Model merging in llms, mllms, and beyond: Methods, theories, applications and opportunities. arXiv preprint arXiv:2408.07666, 2024. 2 +[93] Enneng Yang, Li Shen, Zhenyi Wang, Guibing Guo, Xiaojun Chen, Xingwei Wang, and Dacheng Tao. Representation surgery for multi-task model merging. arXiv preprint arXiv:2402.02705, 2024. 2 +[94] Cagatay Yildiz, Nishaanth Kanna Ravichandran, Prishruit Punia, Matthias Bethge, and Beyza Ermis. Investigating continual pretraining in large language models: Insights and implications. arXiv preprint arXiv:2402.17400, 2024. 2 +[95] Le Yu, Bowen Yu, Haiyang Yu, Fei Huang, and Yongbin Li. Language models are super mario: Absorbing abilities from homologous models as a free lunch. In Proceedings of the 41st International Conference on Machine Learning, pages 57755-57775. PMLR, 2024. 2, 4 +[96] Kerem Zaman, Leshem Choshen, and Shashank Srivastava. Fuse to forget: Bias reduction and selective memorization through model fusion. arXiv preprint arXiv:2311.07682, 2023. 2 +[97] Friedemann Zenke, Ben Poole, and Surya Ganguli. Continuual learning through synaptic intelligence. In ICML, 2017. 3 +[98] Renrui Zhang, Rongyao Fang, Peng Gao, Wei Zhang, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free clip-adapter for better vision-language modeling. arXiv preprint arXiv:2111.03930, 2021. 1 +[99] Shenghe Zheng and Hongzhi Wang. Free-merging: Fourier transform for model merging with lightweight experts. arXiv preprint arXiv:2411.16815, 2024. 2 +[100] Luca Zhou, Daniele Solombrino, Donato Crisostomi, Maria Sofia Bucarelli, Fabrizio Silvestri, and Emanuele Rodolà. Atm: Improving model merging by alternating tuning and merging. arXiv preprint arXiv:2411.03055, 2024. 1, 2 \ No newline at end of file diff --git a/2025/How to Merge Your Multimodal Models Over Time_/images.zip b/2025/How to Merge Your Multimodal Models Over Time_/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..25761491d501403cb180a41cdc02d0e545c030e8 --- /dev/null +++ b/2025/How to Merge Your Multimodal Models Over Time_/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a48b1da78cd15fb1536fe9936f489bf441a0719b3fff3bc3b9082971188171e7 +size 472633 diff --git a/2025/How to Merge Your Multimodal Models Over Time_/layout.json b/2025/How to Merge Your Multimodal Models Over Time_/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a5ff70630679a4b420cefa918c3cbbc8ae72037f --- /dev/null +++ b/2025/How to Merge Your Multimodal Models Over Time_/layout.json @@ -0,0 +1,12160 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 130, + 103, + 455, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 103, + 455, + 121 + ], + "spans": [ + { + "bbox": [ + 130, + 103, + 455, + 121 + ], + "type": "text", + "content": "How to Merge Your Multimodal Models Over Time?" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 457, + 99, + 479, + 119 + ], + "blocks": [ + { + "bbox": [ + 457, + 99, + 479, + 119 + ], + "lines": [ + { + "bbox": [ + 457, + 99, + 479, + 119 + ], + "spans": [ + { + "bbox": [ + 457, + 99, + 479, + 119 + ], + "type": "image", + "image_path": "d3ca2ac2a62466f29cd4d95337275d8e2a7cb93585bf35347106ea6962047f39.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 96, + 142, + 511, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 142, + 511, + 172 + ], + "spans": [ + { + "bbox": [ + 96, + 142, + 511, + 172 + ], + "type": "text", + "content": "Sebastian Dziadzio*¹ Vishaal Udandarao*¹,² Karsten Roth*¹,³ Ameya Prabhu°¹ Zeynep Akata³† Samuel Albanie† Matthias Bethge¹†" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 173, + 539, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 173, + 539, + 185 + ], + "spans": [ + { + "bbox": [ + 69, + 173, + 539, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 69, + 173, + 539, + 185 + ], + "type": "text", + "content": "Tübingen AI Center, University of Tübingen " + }, + { + "bbox": [ + 69, + 173, + 539, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 69, + 173, + 539, + 185 + ], + "type": "text", + "content": "University of Cambridge " + }, + { + "bbox": [ + 69, + 173, + 539, + 185 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 69, + 173, + 539, + 185 + ], + "type": "text", + "content": "Munich Center for ML, Technical University of Munich" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 238, + 297, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 238, + 297, + 562 + ], + "spans": [ + { + "bbox": [ + 54, + 238, + 297, + 562 + ], + "type": "text", + "content": "Model merging combines multiple \"expert\" models—finetuned from a base foundation model on diverse tasks and domains—into a single, more capable model. However, most existing model merging approaches assume that all experts are available simultaneously. In reality, new tasks and domains emerge progressively over time, requiring strategies to integrate the knowledge of expert models as they become available: a process we call temporal model merging. The temporal dimension introduces unique challenges not addressed in prior work, raising new questions such as: when training for a new task, should the expert model start from the merged past experts or from the original base model? Should we merge all models at each time step? Which merging techniques are best suited for temporal merging? Should different strategies be used to initialize the training and deploy the model? To answer these questions, we propose a unified framework called TIME—Temporal Integration of Model Expertise—which defines temporal model merging across three axes: (1) Initialization Phase, (2) Deployment Phase, and (3) Merging Technique. Using TIME, we study temporal model merging across model sizes, compute budgets, and learning horizons on the FoMo-in-Flux benchmark. Our comprehensive suite of experiments across TIME allows us to uncover key insights for temporal model merging, offering a better understanding of current challenges and best practices for effective temporal model merging. Our code is available here." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 597, + 135, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 135, + 609 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 135, + 609 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 617, + 295, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 666 + ], + "type": "text", + "content": "Foundation models consolidate a wide range of capabilities and knowledge into a single, large model [5]. Consequently, model merging [24, 34, 63, 87] has emerged as a key technique for unifying multiple task-specific fine-tuned models" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 212, + 555, + 316 + ], + "blocks": [ + { + "bbox": [ + 316, + 212, + 555, + 316 + ], + "lines": [ + { + "bbox": [ + 316, + 212, + 555, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 212, + 555, + 316 + ], + "type": "image", + "image_path": "3c8338fe5b160b2a80269142625cfe5aefaa614fa03a02a07f4f5debd3b41ad5.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 323, + 555, + 380 + ], + "lines": [ + { + "bbox": [ + 313, + 323, + 555, + 380 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 555, + 380 + ], + "type": "text", + "content": "Figure 1. Temporal Model Merging generalizes standard model merging (yellow), which merges multiple trained experts just once, in a single step. Our systematic study of this realistic multi-step regime reveals that initialization and deployment strategies dominate the importance of the single-step weight merging strategy." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 395, + 555, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 395, + 555, + 419 + ], + "spans": [ + { + "bbox": [ + 313, + 395, + 555, + 419 + ], + "type": "text", + "content": "derived from a shared base into a single model, effectively preserving the strengths of each specialized expert." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 421, + 556, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 421, + 556, + 529 + ], + "spans": [ + { + "bbox": [ + 312, + 421, + 556, + 529 + ], + "type": "text", + "content": "Current model merging approaches typically assume a fixed base model that is fine-tuned independently on " + }, + { + "bbox": [ + 312, + 421, + 556, + 529 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 312, + 421, + 556, + 529 + ], + "type": "text", + "content": " diverse tasks and domains to produce a set of independent experts [17, 24, 34, 64, 89], which are then merged simultaneously. Research in this field has therefore focused primarily on improving merging techniques for larger or structurally differing " + }, + { + "bbox": [ + 312, + 421, + 556, + 529 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 312, + 421, + 556, + 529 + ], + "type": "text", + "content": "-sets, exploring the impact of the diversity and scale of finetuning domains, tasks and experts [27, 69, 86, 89]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "text", + "content": "However, the world is constantly evolving, leading to continuous shifts over data distributions, domains, and tasks, with new concepts emerging [45, 68] that may have been insufficiently covered during large-scale pretraining [20, 22, 32, 45, 57, 65, 66, 81-83, 98]. This dynamic nature of real-world applications motivates a hitherto missing systematic exploration into temporal model merging (see Fig. 1) to better understand model merging along an additional, overlooked [13, 100] axis: time. Specifically, in this work, we ask: (1) Is temporal model merging significantly influenced by the choice of initialization for expert training? (2) When evaluated over time, which model merging techniques emerge as most suitable? (3) Should merging strategies differ between initialization and model deployment? To answer these questions, we propose a unified framework" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 674, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 674, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 674, + 295, + 713 + ], + "type": "text", + "content": "*equal contribution, random order †equal supervision, random order, o core contributor. Correspondence to: vishaal.udarao@bethgelab.org or sebastian.dziadzio@bethgelab.org" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20479" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "for studying temporal model merging—TIME (Temporal Integration of Model Expertise)—structured around three key axes spanning the design space of temporal merging solutions (as shown in Fig. 2):" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 123, + 295, + 335 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 55, + 123, + 295, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 123, + 295, + 159 + ], + "spans": [ + { + "bbox": [ + 55, + 123, + 295, + 159 + ], + "type": "text", + "content": "1. Initialization Phase. As expert models are trained and created continuously over time, initialization becomes a crucial design element." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 163, + 295, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 163, + 295, + 236 + ], + "spans": [ + { + "bbox": [ + 55, + 163, + 295, + 236 + ], + "type": "text", + "content": "2. Deployment Phase. After training an expert on each task, the next step is to deploy a suitable final model. For temporal model merging, this process has to account for a varying number of past expert models and deployed variants, striking a balance between retaining past knowledge and incorporating new task-specific knowledge." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 239, + 295, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 295, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 295, + 335 + ], + "type": "text", + "content": "3. Model Merging Techniques. Merging literature has explored a range of approaches for simultaneous merging, from simple weight-averaging and interpolation [24, 63, 68, 76] to more complex methods involving weight and candidate selection and re-weighting [11, 42, 43, 89, 95]. To apply these methods temporally, it is crucial to understand how they perform with varying numbers of merge candidates " + }, + { + "bbox": [ + 55, + 239, + 295, + 335 + ], + "type": "inline_equation", + "content": "(k_{t})" + }, + { + "bbox": [ + 55, + 239, + 295, + 335 + ], + "type": "text", + "content": " and shifting distributions." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 337, + 295, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 337, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 337, + 295, + 456 + ], + "type": "text", + "content": "Using our TIME framework, we position existing model merging approaches along each key axis and conduct a systematic study of model merging over time. For this, we leverage the multimodal FoMo-in-Flux [68] benchmark which includes 63 datasets with well-documented sequential properties, enabling a thorough investigation of temporal model merging under practical compute constraints, as proposed in Roth et al. [68]. Our experiments systematically explore different merging techniques, initialization, and deployment strategies, providing several key insights:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 72, + 468, + 246, + 481 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 468, + 246, + 481 + ], + "spans": [ + { + "bbox": [ + 72, + 468, + 246, + 481 + ], + "type": "text", + "content": "Key Insights for Temporal Model Merging" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 70, + 490, + 279, + 701 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 70, + 490, + 279, + 526 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 490, + 279, + 526 + ], + "spans": [ + { + "bbox": [ + 70, + 490, + 279, + 526 + ], + "type": "text", + "content": "[A] Accounting for time is crucial. Standard \"off-line\" model merging techniques do not generalize well to the temporal merging setting (Sec. 4.2)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 70, + 529, + 279, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 529, + 279, + 588 + ], + "spans": [ + { + "bbox": [ + 70, + 529, + 279, + 588 + ], + "type": "text", + "content": "[B] Complex merging techniques matter little. Choosing sophisticated merging techniques beyond simple weighted averaging provides at best marginal benefits for temporal model merging, especially for long task sequences (Sec. 4.5)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 71, + 591, + 279, + 639 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 591, + 279, + 639 + ], + "spans": [ + { + "bbox": [ + 71, + 591, + 279, + 639 + ], + "type": "text", + "content": "[C] Initialization and deployment are critical. Choosing how to select and combine available weights before and after each task " + }, + { + "bbox": [ + 71, + 591, + 279, + 639 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 71, + 591, + 279, + 639 + ], + "type": "text", + "content": " is most important for temporal model merging (Sec. 4.4)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 71, + 642, + 279, + 701 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 642, + 279, + 701 + ], + "spans": [ + { + "bbox": [ + 71, + 642, + 279, + 701 + ], + "type": "text", + "content": "[D] Temporal model merging scales well. Larger models or compute budgets allows greater benefits from temporal merging. Scaling enables temporal model merging to even outperform the multitask model, trained on all tasks at once (Sec. 4.6)." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 314, + 71, + 406, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 406, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 406, + 83 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 91, + 555, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 555, + 426 + ], + "type": "text", + "content": "Model Merging. We provide a short overview of the model merging literature, detailed in these excellent surveys [90, 92]. While both model aggregation through distillation [10, 67] and averaging checkpoints during training [29, 35, 69] have shown success, the requirement of additional compute limits practicability of these methods [55]. Instead, recent work [18, 21, 24, 25, 41, 47, 49, 50, 52, 59, 69, 71, 72, 75, 77, 85–88, 90, 93, 96, 99] has shown the effectiveness of training-free weight averaging and interpolation of fine-tuned expert models to produce an improved base model, benefiting from (linear) mode connectivity in models fine-tuned from a single pre-trained checkpoint [1, 14, 26, 48, 61]. These insights have been extended into weight-averaged reward models [62], policy models [60] with spherical interpolation, and KL-constrained RLHF [19, 36, 39, 46]. Works such as Fisher-Merge [43], TIES [89], RegMean [28], MATS [78], DELLA [12], DARE [95], Broadcrumbs [11], evolutionary merging [2] and MagMax [42] have explored merging strategies beyond simple interpolation to determine which weights should be merged across expert models. These methods have different benefits for in- and out-of-distribution generalization across domains [79], though recently they have been shown to perform similarly at scale [91]. Additionally, some works have explored the initialization dimension for effectively merging models [9, 13, 42, 100]. In this work, we propose a unifying framework for temporal merging and conduct the most comprehensive study of this topic to date." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 430, + 556, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 556, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 556, + 562 + ], + "type": "text", + "content": "Continual Pretraining extends beyond standard Continual Learning [54, 68], focusing on large-scale model updates starting from pretrained foundation models [16, 20, 23, 56, 68] and addressing more complex and substantial update tasks [6, 7, 16, 37, 38, 68]. There has been limited exploration into using model merging for continual pretraining [3, 42, 68, 76], as most prior works focus on training strategies including regularization objectives and learning-rate schedules [16, 23, 33, 44, 51, 55, 68, 74, 80, 94]. We keep the training strategy fixed, and provide an in-depth exploration beyond simple merging techniques." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 573, + 545, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 573, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 573, + 545, + 588 + ], + "type": "text", + "content": "3. Design Space of Temporal Model Merging" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "Notation. Throughout this work, we use " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": " to refer to a given task at time " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": ". Full model parameterization is denoted by " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": ", with the following key instantiations: " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\theta_t" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": " represents model weights at task " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\theta_t^I" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\theta_t^S" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\theta_t^O" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": " denote weights used for initialization, saved weight checkpoints at task " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": " (i.e., the trained expert models), and the output deployed model weights, respectively. Note that while standard model merging considers model weights as elements of a fixed set " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\{\\theta_k\\}_{k=1}^K" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": ", temporal model merging organizes them along the time axis " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\theta_t" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20480" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 68, + 553, + 270 + ], + "blocks": [ + { + "bbox": [ + 57, + 68, + 553, + 270 + ], + "lines": [ + { + "bbox": [ + 57, + 68, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 57, + 68, + 553, + 270 + ], + "type": "image", + "image_path": "a42037083fab748e2b76f8611cfef4e3d1df23ad74a56b9313e8bf709ebc64ef.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 273, + 555, + 329 + ], + "lines": [ + { + "bbox": [ + 55, + 273, + 555, + 329 + ], + "spans": [ + { + "bbox": [ + 55, + 273, + 555, + 329 + ], + "type": "text", + "content": "Figure 2. Design Space of Temporal Model Merging through TIME. We showcase our framework for the per-task pipeline of temporal model merging over multiple tasks: At each task " + }, + { + "bbox": [ + 55, + 273, + 555, + 329 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 273, + 555, + 329 + ], + "type": "text", + "content": ", we first initialize the current checkpoint to start training from, " + }, + { + "bbox": [ + 55, + 273, + 555, + 329 + ], + "type": "inline_equation", + "content": "\\theta_t^i" + }, + { + "bbox": [ + 55, + 273, + 555, + 329 + ], + "type": "text", + "content": ", by using one or more previously stored checkpoints from previous tasks, either directly or by merging them. We train " + }, + { + "bbox": [ + 55, + 273, + 555, + 329 + ], + "type": "inline_equation", + "content": "\\theta_t^i" + }, + { + "bbox": [ + 55, + 273, + 555, + 329 + ], + "type": "text", + "content": " on current task data " + }, + { + "bbox": [ + 55, + 273, + 555, + 329 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_t" + }, + { + "bbox": [ + 55, + 273, + 555, + 329 + ], + "type": "text", + "content": " to yield the current task checkpoint " + }, + { + "bbox": [ + 55, + 273, + 555, + 329 + ], + "type": "inline_equation", + "content": "\\theta_t^s" + }, + { + "bbox": [ + 55, + 273, + 555, + 329 + ], + "type": "text", + "content": ", which is inserted into the checkpoint buffer. Finally, to produce the output model, " + }, + { + "bbox": [ + 55, + 273, + 555, + 329 + ], + "type": "inline_equation", + "content": "\\theta_t^o" + }, + { + "bbox": [ + 55, + 273, + 555, + 329 + ], + "type": "text", + "content": ", we either merge previously stored checkpoints from the buffer or use them directly. The entire framework is depicted in the pseudo-code on the right panel." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 340, + 268, + 354 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 340, + 268, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 340, + 268, + 354 + ], + "type": "text", + "content": "3.1. Temporal Model Merging through TIME" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 358, + 296, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 358, + 296, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 358, + 296, + 467 + ], + "type": "text", + "content": "Standard model merging is typically performed offline, after all experts have been trained to convergence [15, 87, 89, 91]. In contrast, model merging in continual pretraining is generally done sequentially, using past checkpoints [68, 76]. Both approaches are specific instances of our more general temporal merging framework, TIME, which defines temporal merging along three key axes: initialization of each expert, merging for deployment at step " + }, + { + "bbox": [ + 55, + 358, + 296, + 467 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 358, + 296, + 467 + ], + "type": "text", + "content": ", and merging techniques " + }, + { + "bbox": [ + 55, + 358, + 296, + 467 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{merge}}" + }, + { + "bbox": [ + 55, + 358, + 296, + 467 + ], + "type": "text", + "content": " applied over time:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 472, + 147, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 147, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 147, + 483 + ], + "type": "text", + "content": "Axis 1: Initialization" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 487, + 296, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 296, + 607 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 296, + 607 + ], + "type": "text", + "content": "As expert models are created continuously over time, initialization becomes a crucial choice. Unlike model merging at a single point in time, the number of potential starting points grows exponentially over time as new experts are created. This raises the question: should starting points for each time step be derived from the base weights (as in traditional merging), from a merged combination of previous experts, or from most recent weights, as commonly done in continual pretraining [68, 76]? In this work, we study the following initialization protocols at time step " + }, + { + "bbox": [ + 55, + 487, + 296, + 607 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 487, + 296, + 607 + ], + "type": "text", + "content": " for TIME:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 608, + 296, + 692 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 56, + 608, + 295, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 608, + 295, + 631 + ], + "spans": [ + { + "bbox": [ + 56, + 608, + 295, + 631 + ], + "type": "text", + "content": "- initZS, which consistently initializes with the base zero-shot model weights " + }, + { + "bbox": [ + 56, + 608, + 295, + 631 + ], + "type": "inline_equation", + "content": "\\theta_0" + }, + { + "bbox": [ + 56, + 608, + 295, + 631 + ], + "type": "text", + "content": " at each timestep " + }, + { + "bbox": [ + 56, + 608, + 295, + 631 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 56, + 608, + 295, + 631 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 632, + 296, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 632, + 296, + 655 + ], + "spans": [ + { + "bbox": [ + 56, + 632, + 296, + 655 + ], + "type": "text", + "content": "- initFT, which for step " + }, + { + "bbox": [ + 56, + 632, + 296, + 655 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 56, + 632, + 296, + 655 + ], + "type": "text", + "content": " always initializes with the latest available finetuned model weights " + }, + { + "bbox": [ + 56, + 632, + 296, + 655 + ], + "type": "inline_equation", + "content": "\\theta_{t - 1}^{S}" + }, + { + "bbox": [ + 56, + 632, + 296, + 655 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 655, + 296, + 692 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 655, + 296, + 692 + ], + "spans": [ + { + "bbox": [ + 56, + 655, + 296, + 692 + ], + "type": "text", + "content": "- initEMA, which computes an unrolled exponential moving average merge over all previously seen expert models " + }, + { + "bbox": [ + 56, + 655, + 296, + 692 + ], + "type": "inline_equation", + "content": "\\{\\theta_{t'}^S\\}_{1,\\dots,t - 1}" + }, + { + "bbox": [ + 56, + 655, + 296, + 692 + ], + "type": "text", + "content": " following the equation:" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 114, + 700, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 700, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 114, + 700, + 295, + 715 + ], + "type": "interline_equation", + "content": "\\theta_ {t ^ {\\prime}} ^ {\\mathrm {E M A}} = f _ {\\text {m e r g e}} \\left(\\theta_ {t ^ {\\prime} - 1} ^ {\\mathrm {E M A}}, \\theta_ {t ^ {\\prime} - 1} ^ {S}, \\mathcal {F}\\right) \\tag {1}", + "image_path": "0d3b4d1f8010ff82c06da10bab4e091a25900edc6e3759c0f96f17e1d451eacb.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 322, + 341, + 555, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 341, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 322, + 341, + 555, + 365 + ], + "type": "text", + "content": "with merging hyperparameters " + }, + { + "bbox": [ + 322, + 341, + 555, + 365 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 322, + 341, + 555, + 365 + ], + "type": "text", + "content": ". Consequently, the initialization weights are given as " + }, + { + "bbox": [ + 322, + 341, + 555, + 365 + ], + "type": "inline_equation", + "content": "\\theta_t^I = \\theta_t^{\\mathrm{EMA}}" + }, + { + "bbox": [ + 322, + 341, + 555, + 365 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 373, + 403, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 373, + 403, + 385 + ], + "spans": [ + { + "bbox": [ + 314, + 373, + 403, + 385 + ], + "type": "text", + "content": "Axis 2: Deployment" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 389, + 555, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 555, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 555, + 581 + ], + "type": "text", + "content": "With each update iteration and expert training phase " + }, + { + "bbox": [ + 313, + 389, + 555, + 581 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 389, + 555, + 581 + ], + "type": "text", + "content": ", a decision must be made on the final model to deploy, determining which weights to present for downstream use. In continual pretraining, the trained model " + }, + { + "bbox": [ + 313, + 389, + 555, + 581 + ], + "type": "inline_equation", + "content": "\\theta_t^S" + }, + { + "bbox": [ + 313, + 389, + 555, + 581 + ], + "type": "text", + "content": " is deployed directly. In contrast, standard model merging applies a merging technique " + }, + { + "bbox": [ + 313, + 389, + 555, + 581 + ], + "type": "inline_equation", + "content": "f_{\\text{merge}}" + }, + { + "bbox": [ + 313, + 389, + 555, + 581 + ], + "type": "text", + "content": " to a fixed set of " + }, + { + "bbox": [ + 313, + 389, + 555, + 581 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 389, + 555, + 581 + ], + "type": "text", + "content": " expert models. Temporal model merging, however, must account for both previously deployed models and the growing number of expert models available over time. Unlike standard merging, where " + }, + { + "bbox": [ + 313, + 389, + 555, + 581 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 389, + 555, + 581 + ], + "type": "text", + "content": " remains constant, the number of experts to merge increases with each step. As a result, temporal merging introduces the idea of weighted combinations, balancing recent updates with retained past knowledge to achieve adaptability and stability—both critical for effective continual learning [31, 54, 55, 68, 97]. In this work, we study three strategies for model deployment:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 582, + 553, + 666 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 313, + 582, + 553, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 553, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 553, + 605 + ], + "type": "text", + "content": "- deployyFT, which at step " + }, + { + "bbox": [ + 313, + 582, + 553, + 605 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 582, + 553, + 605 + ], + "type": "text", + "content": " deploys the latest finetuned expert model, i.e. " + }, + { + "bbox": [ + 313, + 582, + 553, + 605 + ], + "type": "inline_equation", + "content": "\\theta_s^O = \\theta_t^S" + }, + { + "bbox": [ + 313, + 582, + 553, + 605 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 605, + 553, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 553, + 642 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 553, + 642 + ], + "type": "text", + "content": "- deployEMA, which computes an unrolled exponential moving average merge over all expert models, i.e. " + }, + { + "bbox": [ + 313, + 605, + 553, + 642 + ], + "type": "inline_equation", + "content": "\\theta_t^O = \\theta_{t+1}^{EMA}" + }, + { + "bbox": [ + 313, + 605, + 553, + 642 + ], + "type": "text", + "content": " following Eq. (1)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 642, + 553, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 553, + 666 + ], + "type": "text", + "content": "- deployALL, which applies a merging technique " + }, + { + "bbox": [ + 313, + 642, + 553, + 666 + ], + "type": "inline_equation", + "content": "f_{\\text{merge}}" + }, + { + "bbox": [ + 313, + 642, + 553, + 666 + ], + "type": "text", + "content": " over all previously computed expert models " + }, + { + "bbox": [ + 313, + 642, + 553, + 666 + ], + "type": "inline_equation", + "content": "\\{\\theta_{t'}^S\\}_{1}^{t - 1}" + }, + { + "bbox": [ + 313, + 642, + 553, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 673, + 437, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 673, + 437, + 685 + ], + "spans": [ + { + "bbox": [ + 314, + 673, + 437, + 685 + ], + "type": "text", + "content": "Axis 3: Merging Techniques" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "At each point in time, for both initialization and deployment, merging technique " + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "inline_equation", + "content": "f_{\\text{merge}}" + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": " defines how to combine" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20481" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 91, + 295, + 185 + ], + "blocks": [ + { + "bbox": [ + 77, + 71, + 273, + 82 + ], + "lines": [ + { + "bbox": [ + 77, + 71, + 273, + 82 + ], + "spans": [ + { + "bbox": [ + 77, + 71, + 273, + 82 + ], + "type": "text", + "content": "Table 1. Comparison of model merging techniques." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 91, + 295, + 185 + ], + "lines": [ + { + "bbox": [ + 58, + 91, + 295, + 185 + ], + "spans": [ + { + "bbox": [ + 58, + 91, + 295, + 185 + ], + "type": "table", + "html": "
MethodSparsificationConsensusScaling
Weight averaging [63, 86]XLinear Int.Weight coeff.
SLERP [60]XSpherical Int.Weight coeff.
Task Arithmetic [25]XLinear Int.Scaling factor
MagMax [42]XMax. MagnitudeScaling factor
TIES [89]Top-kSign AgreementScaling factor
DARE-TIES [95]RandomSign AgreementScaling factor
Breadcrumbs-TIES [11]Top/Bottom-kSign AgreementScaling factor
Model Stock [27]XGeometricAdaptive ratio
LiNeS [84]XXLayer weights
", + "image_path": "0a81180b946c8cb3fe395f9c53e6b9bcfd9897f8b4f7e16b414491ee7ad59a40.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 198, + 295, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 198, + 295, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 295, + 258 + ], + "type": "text", + "content": "the available expert models and checkpoints. In this work, we study nine different variants in total, shown in Tab. 1. Denoting the number of models to merge at timestep " + }, + { + "bbox": [ + 55, + 198, + 295, + 258 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 198, + 295, + 258 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 198, + 295, + 258 + ], + "type": "inline_equation", + "content": "M_{t}" + }, + { + "bbox": [ + 55, + 198, + 295, + 258 + ], + "type": "text", + "content": " (with " + }, + { + "bbox": [ + 55, + 198, + 295, + 258 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 55, + 198, + 295, + 258 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 198, + 295, + 258 + ], + "type": "inline_equation", + "content": "M_{t} = M" + }, + { + "bbox": [ + 55, + 198, + 295, + 258 + ], + "type": "text", + "content": " for standard model merging), we can define these methods as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 269, + 296, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 269, + 296, + 308 + ], + "spans": [ + { + "bbox": [ + 55, + 269, + 296, + 308 + ], + "type": "text", + "content": "Weight Averaging [24, 63, 68, 76, 87] simply employs a uniformly weighted, element-wise average over all models " + }, + { + "bbox": [ + 55, + 269, + 296, + 308 + ], + "type": "inline_equation", + "content": "\\theta_{t,i}" + }, + { + "bbox": [ + 55, + 269, + 296, + 308 + ], + "type": "text", + "content": ", resulting in a merge function " + }, + { + "bbox": [ + 55, + 269, + 296, + 308 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{merge}}^{\\mathrm{WA}}" + }, + { + "bbox": [ + 55, + 269, + 296, + 308 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 137, + 315, + 295, + 343 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 315, + 295, + 343 + ], + "spans": [ + { + "bbox": [ + 137, + 315, + 295, + 343 + ], + "type": "interline_equation", + "content": "\\theta_ {t} = \\frac {1}{M _ {t}} \\sum_ {i} \\theta_ {t, i}. \\tag {2}", + "image_path": "e48c7a9682635f84f5ba4ccf93c0304625314a164dda55645baa2e9e8c8555a0.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 351, + 296, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 351, + 296, + 425 + ], + "spans": [ + { + "bbox": [ + 55, + 351, + 296, + 425 + ], + "type": "text", + "content": "SLERP [60, 73] assumes weights to live on a hypersphere, and consequently conducts interpolation along a curved path connecting weight entries. In particular, for two models " + }, + { + "bbox": [ + 55, + 351, + 296, + 425 + ], + "type": "inline_equation", + "content": "\\theta_{t,1}" + }, + { + "bbox": [ + 55, + 351, + 296, + 425 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 351, + 296, + 425 + ], + "type": "inline_equation", + "content": "\\theta_{t,2}" + }, + { + "bbox": [ + 55, + 351, + 296, + 425 + ], + "type": "text", + "content": " deriving from some base weight " + }, + { + "bbox": [ + 55, + 351, + 296, + 425 + ], + "type": "inline_equation", + "content": "\\theta_{t-1}" + }, + { + "bbox": [ + 55, + 351, + 296, + 425 + ], + "type": "text", + "content": " and the corresponding task vectors [25] " + }, + { + "bbox": [ + 55, + 351, + 296, + 425 + ], + "type": "inline_equation", + "content": "\\delta_{t,i} = \\theta_{t,i} - \\theta_{t-1}" + }, + { + "bbox": [ + 55, + 351, + 296, + 425 + ], + "type": "text", + "content": ", SLERP with interpolation weight " + }, + { + "bbox": [ + 55, + 351, + 296, + 425 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 351, + 296, + 425 + ], + "type": "text", + "content": " is defined as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 432, + 295, + 459 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 432, + 295, + 459 + ], + "spans": [ + { + "bbox": [ + 62, + 432, + 295, + 459 + ], + "type": "interline_equation", + "content": "\\theta_ {t} = \\theta_ {t - 1} + \\frac {\\sin (1 - \\lambda) \\Omega_ {1 , 2}}{\\sin \\Omega_ {1 , 2}} \\cdot \\delta_ {t, 1} + \\frac {\\sin \\lambda \\Omega_ {1 , 2}}{\\sin \\Omega_ {1 , 2}} \\cdot \\delta_ {t, 2} \\tag {3}", + "image_path": "9cc2227c889b462568b9de16ee3afdc80ca21a99deae76aeb02cb8574fb980ed.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 467, + 295, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 467, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 467, + 295, + 491 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 55, + 467, + 295, + 491 + ], + "type": "inline_equation", + "content": "\\Omega_{1,2}" + }, + { + "bbox": [ + 55, + 467, + 295, + 491 + ], + "type": "text", + "content": " being the angle between task vectors " + }, + { + "bbox": [ + 55, + 467, + 295, + 491 + ], + "type": "inline_equation", + "content": "\\delta_{t,1}" + }, + { + "bbox": [ + 55, + 467, + 295, + 491 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 467, + 295, + 491 + ], + "type": "inline_equation", + "content": "\\delta_{t,2}" + }, + { + "bbox": [ + 55, + 467, + 295, + 491 + ], + "type": "text", + "content": ". We denote the corresponding merge function " + }, + { + "bbox": [ + 55, + 467, + 295, + 491 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{merge}}^{\\mathrm{SLERP}}" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 491, + 296, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 296, + 538 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 296, + 538 + ], + "type": "text", + "content": "Task Arithmetic [25] defines the merge as a function over task vectors " + }, + { + "bbox": [ + 55, + 491, + 296, + 538 + ], + "type": "inline_equation", + "content": "\\delta_{t,i} = \\theta_{t,i} - \\theta_{t-1}" + }, + { + "bbox": [ + 55, + 491, + 296, + 538 + ], + "type": "text", + "content": " for each weight " + }, + { + "bbox": [ + 55, + 491, + 296, + 538 + ], + "type": "inline_equation", + "content": "\\theta_{t,i}" + }, + { + "bbox": [ + 55, + 491, + 296, + 538 + ], + "type": "text", + "content": " fine-tuned from " + }, + { + "bbox": [ + 55, + 491, + 296, + 538 + ], + "type": "inline_equation", + "content": "\\theta_{t-1}" + }, + { + "bbox": [ + 55, + 491, + 296, + 538 + ], + "type": "text", + "content": ". This introduces a simple merge formalism " + }, + { + "bbox": [ + 55, + 491, + 296, + 538 + ], + "type": "inline_equation", + "content": "f_{\\text{merge}}^{\\text{TA}}" + }, + { + "bbox": [ + 55, + 491, + 296, + 538 + ], + "type": "text", + "content": " for weighted parameter averaging with a scale " + }, + { + "bbox": [ + 55, + 491, + 296, + 538 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 491, + 296, + 538 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 121, + 546, + 295, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 546, + 295, + 574 + ], + "spans": [ + { + "bbox": [ + 121, + 546, + 295, + 574 + ], + "type": "interline_equation", + "content": "\\theta_ {t} = \\theta_ {t - 1} + \\lambda \\frac {1}{M _ {t}} \\sum_ {i} \\delta_ {t, i} \\tag {4}", + "image_path": "858245edc5c8c09884f43c796d5ac4a3d423ebe793a7cfb43b85f1fad28f95e6.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 590, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 295, + 676 + ], + "type": "text", + "content": "TIES [89] builds on the task arithmetic formalism through controlled pruning of task vector entries with low magnitude. Moreover, the sign for each final merged parameter is set based on the sign of the highest total magnitude across the merge candidates. The final update follows basic task arithmetic, only for entries with matching signs. We refer to the respective merge function as " + }, + { + "bbox": [ + 55, + 590, + 295, + 676 + ], + "type": "inline_equation", + "content": "f_{\\text{merge}}^{\\text{TIES}}" + }, + { + "bbox": [ + 55, + 590, + 295, + 676 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "DARE [95] is a similar extension of task arithmetic, but instead of targeted pruning, it randomly zeroes out task vector entries using a random mask " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "Z_{i} \\sim \\mathrm{Bernoulli}(p)" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " and" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "masking probability " + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": ". Final task vector values for " + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{merge}}^{\\mathrm{DARE}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": " are then rescaled based on " + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 387, + 102, + 553, + 129 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 387, + 102, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 387, + 102, + 553, + 129 + ], + "type": "interline_equation", + "content": "\\delta_ {t, i} ^ {\\mathrm {D A R E}} = \\frac {(1 - Z _ {i}) \\delta_ {t , i}}{1 - p}. \\tag {5}", + "image_path": "f690744aaf38767dc2de2b01153c48f51bef2401654482fac24b34896abf5fb8.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 143, + 555, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 143, + 555, + 202 + ], + "spans": [ + { + "bbox": [ + 313, + 143, + 555, + 202 + ], + "type": "text", + "content": "Model Stock [27] provides a geometric extension of simple weight averaging as done in Model Soup [86] by incorporating base weights " + }, + { + "bbox": [ + 313, + 143, + 555, + 202 + ], + "type": "inline_equation", + "content": "\\theta_{t-1}" + }, + { + "bbox": [ + 313, + 143, + 555, + 202 + ], + "type": "text", + "content": " into the merging process. Given finetuned weights " + }, + { + "bbox": [ + 313, + 143, + 555, + 202 + ], + "type": "inline_equation", + "content": "\\theta_{t,1}" + }, + { + "bbox": [ + 313, + 143, + 555, + 202 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 143, + 555, + 202 + ], + "type": "inline_equation", + "content": "\\theta_{t,2}" + }, + { + "bbox": [ + 313, + 143, + 555, + 202 + ], + "type": "text", + "content": ", the Model Stock merge " + }, + { + "bbox": [ + 313, + 143, + 555, + 202 + ], + "type": "inline_equation", + "content": "f_{\\text{merge}}^{\\text{Stock}}" + }, + { + "bbox": [ + 313, + 143, + 555, + 202 + ], + "type": "text", + "content": " is defined as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 318, + 207, + 553, + 235 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 207, + 553, + 235 + ], + "spans": [ + { + "bbox": [ + 318, + 207, + 553, + 235 + ], + "type": "interline_equation", + "content": "\\theta_ {t} = \\frac {2 \\cdot \\cos \\Omega_ {1 , 2}}{1 + \\Omega_ {1 , 2}} \\cdot \\left(\\theta_ {t, 2} - \\theta_ {t, 1}\\right) + \\left(1 - \\frac {2 \\cdot \\cos \\Omega_ {1 , 2}}{1 + \\cos \\Omega_ {1 , 2}}\\right), \\tag {6}", + "image_path": "a613fecb68a588f9675c3abf48ebea7f1a8e1140286bcfc27fe430b538f16eb0.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 239, + 534, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 534, + 252 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 534, + 252 + ], + "type": "text", + "content": "utilizing angle " + }, + { + "bbox": [ + 313, + 239, + 534, + 252 + ], + "type": "inline_equation", + "content": "\\Omega_{1,2}" + }, + { + "bbox": [ + 313, + 239, + 534, + 252 + ], + "type": "text", + "content": " between task vectors " + }, + { + "bbox": [ + 313, + 239, + 534, + 252 + ], + "type": "inline_equation", + "content": "\\delta_{t,1}" + }, + { + "bbox": [ + 313, + 239, + 534, + 252 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 239, + 534, + 252 + ], + "type": "inline_equation", + "content": "\\delta_{t,2}" + }, + { + "bbox": [ + 313, + 239, + 534, + 252 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 254, + 553, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 254, + 553, + 339 + ], + "spans": [ + { + "bbox": [ + 313, + 254, + 553, + 339 + ], + "type": "text", + "content": "Breadcrumbs [11] deploys another variation on task arithmetic for model merging. In particular, for a given task vector " + }, + { + "bbox": [ + 313, + 254, + 553, + 339 + ], + "type": "inline_equation", + "content": "\\delta_{t,i}" + }, + { + "bbox": [ + 313, + 254, + 553, + 339 + ], + "type": "text", + "content": ", extreme left and right tails of the absolute magnitude distribution in " + }, + { + "bbox": [ + 313, + 254, + 553, + 339 + ], + "type": "inline_equation", + "content": "\\delta_{t,i}" + }, + { + "bbox": [ + 313, + 254, + 553, + 339 + ], + "type": "text", + "content": " are zeroed out with left and right thresholds " + }, + { + "bbox": [ + 313, + 254, + 553, + 339 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 254, + 553, + 339 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 254, + 553, + 339 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 254, + 553, + 339 + ], + "type": "text", + "content": ". The modified task vectors " + }, + { + "bbox": [ + 313, + 254, + 553, + 339 + ], + "type": "inline_equation", + "content": "\\delta_{t,i}^{\\mathrm{Bread}}" + }, + { + "bbox": [ + 313, + 254, + 553, + 339 + ], + "type": "text", + "content": " are then applied on base weights " + }, + { + "bbox": [ + 313, + 254, + 553, + 339 + ], + "type": "inline_equation", + "content": "\\theta_{t-1}" + }, + { + "bbox": [ + 313, + 254, + 553, + 339 + ], + "type": "text", + "content": " following the task arithmetic setup, and giving " + }, + { + "bbox": [ + 313, + 254, + 553, + 339 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{merge}}^{\\mathrm{Bread}}" + }, + { + "bbox": [ + 313, + 254, + 553, + 339 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 341, + 554, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 341, + 554, + 402 + ], + "spans": [ + { + "bbox": [ + 313, + 341, + 554, + 402 + ], + "type": "text", + "content": "MagMax [42] also uses task vectors—given multiple task vectors " + }, + { + "bbox": [ + 313, + 341, + 554, + 402 + ], + "type": "inline_equation", + "content": "\\delta_{t,i}" + }, + { + "bbox": [ + 313, + 341, + 554, + 402 + ], + "type": "text", + "content": " (with increments possible along both time " + }, + { + "bbox": [ + 313, + 341, + 554, + 402 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 341, + 554, + 402 + ], + "type": "text", + "content": " and count axis " + }, + { + "bbox": [ + 313, + 341, + 554, + 402 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 341, + 554, + 402 + ], + "type": "text", + "content": "), the final task vector " + }, + { + "bbox": [ + 313, + 341, + 554, + 402 + ], + "type": "inline_equation", + "content": "\\delta_t" + }, + { + "bbox": [ + 313, + 341, + 554, + 402 + ], + "type": "text", + "content": " is yielded through maximum magnitude entry selection; copying the largest magnitude entries across all " + }, + { + "bbox": [ + 313, + 341, + 554, + 402 + ], + "type": "inline_equation", + "content": "\\{\\delta_{t,i}\\}" + }, + { + "bbox": [ + 313, + 341, + 554, + 402 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 313, + 341, + 554, + 402 + ], + "type": "inline_equation", + "content": "\\delta_t" + }, + { + "bbox": [ + 313, + 341, + 554, + 402 + ], + "type": "text", + "content": ", giving " + }, + { + "bbox": [ + 313, + 341, + 554, + 402 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{merge}}^{\\mathrm{Max}}" + }, + { + "bbox": [ + 313, + 341, + 554, + 402 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 403, + 554, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 554, + 487 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 554, + 487 + ], + "type": "text", + "content": "LiNeS [84], for Layer-increasing Network Scaling, scales weight updates based on their respective layer depth enabling early layers to remain close to original pretraining weights (cf. Neyshabur et al. [48]). Given task vectors " + }, + { + "bbox": [ + 313, + 403, + 554, + 487 + ], + "type": "inline_equation", + "content": "\\delta_{t,i}" + }, + { + "bbox": [ + 313, + 403, + 554, + 487 + ], + "type": "text", + "content": ", now broken down across model layers " + }, + { + "bbox": [ + 313, + 403, + 554, + 487 + ], + "type": "inline_equation", + "content": "\\delta_{t,i}^{l}" + }, + { + "bbox": [ + 313, + 403, + 554, + 487 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 403, + 554, + 487 + ], + "type": "inline_equation", + "content": "l\\in [1,\\dots,L]" + }, + { + "bbox": [ + 313, + 403, + 554, + 487 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 403, + 554, + 487 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 403, + 554, + 487 + ], + "type": "text", + "content": " the number of layers, LiNeS follows the base task arithmetic merging formalism, but updates task vectors as" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 343, + 493, + 553, + 509 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 493, + 553, + 509 + ], + "spans": [ + { + "bbox": [ + 343, + 493, + 553, + 509 + ], + "type": "interline_equation", + "content": "\\delta_ {t, i} ^ {\\mathrm {L i N e S}} = \\operatorname {c o n c a t} \\left(\\lambda^ {l = 1} \\delta_ {t, i} ^ {l = 1}, \\dots , \\lambda^ {l = L} \\delta_ {t, i} ^ {l = L}\\right) \\tag {7}", + "image_path": "552c1d83fa269973f4a2ba75b86cea5344a903d8fc284bf9e78c5c5aa3085cc9.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 515, + 554, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 515, + 554, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 515, + 554, + 544 + ], + "type": "text", + "content": "with layer-scaled interpolation weights " + }, + { + "bbox": [ + 313, + 515, + 554, + 544 + ], + "type": "inline_equation", + "content": "\\lambda^l = \\alpha +\\beta \\frac{l - 1}{L - 1}" + }, + { + "bbox": [ + 313, + 515, + 554, + 544 + ], + "type": "text", + "content": " and hyperparameters " + }, + { + "bbox": [ + 313, + 515, + 554, + 544 + ], + "type": "inline_equation", + "content": "\\alpha ,\\beta" + }, + { + "bbox": [ + 313, + 515, + 554, + 544 + ], + "type": "text", + "content": " giving " + }, + { + "bbox": [ + 313, + 515, + 554, + 544 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{merge}}^{\\mathrm{Lines}}" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 551, + 545, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 551, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 551, + 545, + 563 + ], + "type": "text", + "content": "3.2. Complete Temporal Model Merging Pipeline" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 568, + 554, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 568, + 554, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 568, + 554, + 604 + ], + "type": "text", + "content": "Incorporating all three axes of temporal model merging, we define a five-stage update pipeline for each task " + }, + { + "bbox": [ + 313, + 568, + 554, + 604 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 568, + 554, + 604 + ], + "type": "text", + "content": " (see Fig. 2), consisting of the following steps:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 605, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 314, + 605, + 553, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 605, + 553, + 641 + ], + "spans": [ + { + "bbox": [ + 314, + 605, + 553, + 641 + ], + "type": "text", + "content": "1. Init. Choose one of the aforementioned initialization protocols: initZS, initFT, or initEMA. This produces initialization weights " + }, + { + "bbox": [ + 314, + 605, + 553, + 641 + ], + "type": "inline_equation", + "content": "\\theta_t^I" + }, + { + "bbox": [ + 314, + 605, + 553, + 641 + ], + "type": "text", + "content": " at task " + }, + { + "bbox": [ + 314, + 605, + 553, + 641 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 314, + 605, + 553, + 641 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 314, + 641, + 553, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 641, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 314, + 641, + 553, + 666 + ], + "type": "text", + "content": "2. Train. Given " + }, + { + "bbox": [ + 314, + 641, + 553, + 666 + ], + "type": "inline_equation", + "content": "\\theta_t^I" + }, + { + "bbox": [ + 314, + 641, + 553, + 666 + ], + "type": "text", + "content": ", train on current task data " + }, + { + "bbox": [ + 314, + 641, + 553, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_t" + }, + { + "bbox": [ + 314, + 641, + 553, + 666 + ], + "type": "text", + "content": " within a set compute budget to produce the expert model " + }, + { + "bbox": [ + 314, + 641, + 553, + 666 + ], + "type": "inline_equation", + "content": "\\theta_t^S" + }, + { + "bbox": [ + 314, + 641, + 553, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 314, + 666, + 553, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 666, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 314, + 666, + 553, + 677 + ], + "type": "text", + "content": "3. Store. Append " + }, + { + "bbox": [ + 314, + 666, + 553, + 677 + ], + "type": "inline_equation", + "content": "\\theta_t^S" + }, + { + "bbox": [ + 314, + 666, + 553, + 677 + ], + "type": "text", + "content": " to storage of expert model weights " + }, + { + "bbox": [ + 314, + 666, + 553, + 677 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 314, + 666, + 553, + 677 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "text", + "content": "4. Deploy. Choose a deployment protocol: deployFT, deployEMA, or deployALL, and produce the output weights " + }, + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "inline_equation", + "content": "\\theta_t^O = \\mathrm{deploy}(\\mathcal{S})" + }, + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20482" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 73, + 553, + 229 + ], + "blocks": [ + { + "bbox": [ + 58, + 73, + 553, + 229 + ], + "lines": [ + { + "bbox": [ + 58, + 73, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 58, + 73, + 553, + 229 + ], + "type": "image", + "image_path": "226214c5b586e741c060c57869bb9e89d99ccf43de286bec220f490540fb3303.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 231, + 555, + 255 + ], + "lines": [ + { + "bbox": [ + 55, + 231, + 555, + 255 + ], + "spans": [ + { + "bbox": [ + 55, + 231, + 555, + 255 + ], + "type": "text", + "content": "Figure 3. Offline merging methods struggle with TIME. All tested merging techniques perform extremely poorly, and are unable to adapt to the temporal setting, underperforming even a simple replay baseline that sequentially trains the base model on task-replayed data." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 264, + 294, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 264, + 294, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 264, + 294, + 289 + ], + "type": "text", + "content": "5. Eval. The deployed " + }, + { + "bbox": [ + 55, + 264, + 294, + 289 + ], + "type": "inline_equation", + "content": "\\theta_t^O" + }, + { + "bbox": [ + 55, + 264, + 294, + 289 + ], + "type": "text", + "content": " is used for downstream applications and, in our case, extensive evaluation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 291, + 295, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 291, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 54, + 291, + 295, + 422 + ], + "type": "text", + "content": "Particular choices of init, deploy and " + }, + { + "bbox": [ + 54, + 291, + 295, + 422 + ], + "type": "inline_equation", + "content": "f_{\\text{merge}}" + }, + { + "bbox": [ + 54, + 291, + 295, + 422 + ], + "type": "text", + "content": " correspond to existing approaches, for example (initZS, deployALL, " + }, + { + "bbox": [ + 54, + 291, + 295, + 422 + ], + "type": "inline_equation", + "content": "f_{\\text{merge}}^{\\text{WA}}" + }, + { + "bbox": [ + 54, + 291, + 295, + 422 + ], + "type": "text", + "content": ") simply recovers offline merging through weight averaging over experts models derived from original base weights " + }, + { + "bbox": [ + 54, + 291, + 295, + 422 + ], + "type": "inline_equation", + "content": "\\theta_0" + }, + { + "bbox": [ + 54, + 291, + 295, + 422 + ], + "type": "text", + "content": " for each task " + }, + { + "bbox": [ + 54, + 291, + 295, + 422 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 54, + 291, + 295, + 422 + ], + "type": "text", + "content": ". Similarly, (initFT, deployEMA, " + }, + { + "bbox": [ + 54, + 291, + 295, + 422 + ], + "type": "inline_equation", + "content": "f_{\\text{merge}}^{\\text{WA}}" + }, + { + "bbox": [ + 54, + 291, + 295, + 422 + ], + "type": "text", + "content": ") recovers exponential moving average approaches as done in [68, 76]. Using our framework, we can explore temporal model merging in a systematic manner across various combinations of all three design choices to understand their impact and recommend best practices." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 435, + 137, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 435, + 137, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 435, + 137, + 449 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 456, + 296, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 296, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 296, + 540 + ], + "type": "text", + "content": "We first discuss experimental details in Sec. 4.1. Our initial experiments study the applicability of standard, \"offline\" model merging for the temporal paradigm in Sec. 4.2 and possible extensions in Sec. 4.3. We then transition to a comprehensive exploration of applicable strategies within our TIME framework in Sec. 4.4 and Sec. 4.5, before finally scaling up temporal model merging in Sec. 4.6." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 548, + 218, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 548, + 218, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 548, + 218, + 562 + ], + "type": "text", + "content": "4.1. Experimental Pipeline Details" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 566, + 295, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 566, + 295, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 566, + 295, + 639 + ], + "type": "text", + "content": "To study temporal model merging in practically relevant scenarios, we use the continual pretraining benchmark Fomo-in-Flux [68]. It includes 41 adaptation datasets and 22 separate evaluation datasets, covering an array of visual and semantic distribution shifts, alongside well-defined, practically motivated compute constraints for each task." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "Training at task " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": ". By default, we continuously finetune and merge a ViT-B/16 CLIP [8, 58] model pretrained on LAION-2B [70]. The model is pretrained and finetuned using the standard CLIP objective [58]. Following [68], we fix the training steps for each task based on the DataComp-Small computation budget of " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "1.8 \\times 10^{9}" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 264, + 554, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 264, + 554, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 264, + 554, + 373 + ], + "type": "text", + "content": "GFLOPS, split equally across 20 tasks. We also additionally explore longer sequences with " + }, + { + "bbox": [ + 313, + 264, + 554, + 373 + ], + "type": "inline_equation", + "content": "T \\in \\{50,100\\}" + }, + { + "bbox": [ + 313, + 264, + 554, + 373 + ], + "type": "text", + "content": ". At each temporal merging step, we allow unrestricted access to a pretraining data pool " + }, + { + "bbox": [ + 313, + 264, + 554, + 373 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 313, + 264, + 554, + 373 + ], + "type": "text", + "content": ", using the same 2M random subset of LAION-400M as in [68]. We use a cosine-decay LR schedule with a linear warmup of " + }, + { + "bbox": [ + 313, + 264, + 554, + 373 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 264, + 554, + 373 + ], + "type": "text", + "content": ", AdamW optimizer [40], a batchsize of 512 and gradient norm clipping to 1. All experiments use PyTorch [53], and are run on a compute cluster using NVIDIA A100/H100s." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 376, + 554, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 554, + 424 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 554, + 424 + ], + "type": "text", + "content": "Checkpoint Storage. To enable arbitrary merging strategies at any task " + }, + { + "bbox": [ + 313, + 376, + 554, + 424 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 376, + 554, + 424 + ], + "type": "text", + "content": ", we store each trained expert " + }, + { + "bbox": [ + 313, + 376, + 554, + 424 + ], + "type": "inline_equation", + "content": "\\theta_t^S" + }, + { + "bbox": [ + 313, + 376, + 554, + 424 + ], + "type": "text", + "content": ". We do not store any merged initialization weights " + }, + { + "bbox": [ + 313, + 376, + 554, + 424 + ], + "type": "inline_equation", + "content": "\\theta_t^I" + }, + { + "bbox": [ + 313, + 376, + 554, + 424 + ], + "type": "text", + "content": " or deployed models " + }, + { + "bbox": [ + 313, + 376, + 554, + 424 + ], + "type": "inline_equation", + "content": "\\theta_t^O" + }, + { + "bbox": [ + 313, + 376, + 554, + 424 + ], + "type": "text", + "content": ", as these can be recovered from stored experts." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 427, + 556, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 427, + 556, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 427, + 556, + 559 + ], + "type": "text", + "content": "Evaluation and Metrics. We focus on two key quantities to evaluate temporal model merging: the level of adaptation, reflecting performance improvement with each merging step, and retention, capturing the preservation of prior knowledge. Specifically, we report two metrics following Roth et al. [68]: Knowledge Accumulation " + }, + { + "bbox": [ + 313, + 427, + 556, + 559 + ], + "type": "inline_equation", + "content": "(\\mathcal{A}_{KA})" + }, + { + "bbox": [ + 313, + 427, + 556, + 559 + ], + "type": "text", + "content": ", the average accuracy (or recall@5 for retrieval) across all 41 adaptation datasets, and Zero-Shot Retention " + }, + { + "bbox": [ + 313, + 427, + 556, + 559 + ], + "type": "inline_equation", + "content": "(\\mathcal{A}_{ZS})" + }, + { + "bbox": [ + 313, + 427, + 556, + 559 + ], + "type": "text", + "content": ", the zero-shot accuracy or recall@5 on all 22 held-out evaluation datasets. Additional details, including a description of our plotting style, can be found in the supplementary." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 571, + 537, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 571, + 537, + 585 + ], + "spans": [ + { + "bbox": [ + 313, + 571, + 537, + 585 + ], + "type": "text", + "content": "4.2. Do We Need Model Merging Across Time?" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 590, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 590, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 590, + 555, + 715 + ], + "type": "text", + "content": "The simplest approach to tackle temporal merging is to disregard the time axis and follow standard offline merging paradigms. In TIME terms, this corresponds to a configuration of (init_ZS, deploy_ALL, " + }, + { + "bbox": [ + 313, + 590, + 555, + 715 + ], + "type": "inline_equation", + "content": "f_{\\text{merge}}" + }, + { + "bbox": [ + 313, + 590, + 555, + 715 + ], + "type": "text", + "content": "), which always fine-tunes the initial base weights " + }, + { + "bbox": [ + 313, + 590, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\theta_0" + }, + { + "bbox": [ + 313, + 590, + 555, + 715 + ], + "type": "text", + "content": ". To study the effectiveness of such an offline strategy, we test the above triplet with various choices of " + }, + { + "bbox": [ + 313, + 590, + 555, + 715 + ], + "type": "inline_equation", + "content": "f_{\\text{merge}}" + }, + { + "bbox": [ + 313, + 590, + 555, + 715 + ], + "type": "text", + "content": ", including " + }, + { + "bbox": [ + 313, + 590, + 555, + 715 + ], + "type": "inline_equation", + "content": "f_{\\text{merge}}^{\\text{WA}}" + }, + { + "bbox": [ + 313, + 590, + 555, + 715 + ], + "type": "text", + "content": " (averaging in Fig. 3), " + }, + { + "bbox": [ + 313, + 590, + 555, + 715 + ], + "type": "inline_equation", + "content": "f_{\\text{merge}}^{\\text{TA}}" + }, + { + "bbox": [ + 313, + 590, + 555, + 715 + ], + "type": "text", + "content": " (task-arithmetic), " + }, + { + "bbox": [ + 313, + 590, + 555, + 715 + ], + "type": "inline_equation", + "content": "f_{\\text{merge}}^{\\text{Max}}" + }, + { + "bbox": [ + 313, + 590, + 555, + 715 + ], + "type": "text", + "content": " (magmax), " + }, + { + "bbox": [ + 313, + 590, + 555, + 715 + ], + "type": "inline_equation", + "content": "f_{\\text{merge}}^{\\text{TIES}}" + }, + { + "bbox": [ + 313, + 590, + 555, + 715 + ], + "type": "text", + "content": " (ties), " + }, + { + "bbox": [ + 313, + 590, + 555, + 715 + ], + "type": "inline_equation", + "content": "f_{\\text{merge}}^{\\text{DARE}}" + }, + { + "bbox": [ + 313, + 590, + 555, + 715 + ], + "type": "text", + "content": " applied over TIES (dare-ties), " + }, + { + "bbox": [ + 313, + 590, + 555, + 715 + ], + "type": "inline_equation", + "content": "f_{\\text{merge}}^{\\text{Bread}}" + }, + { + "bbox": [ + 313, + 590, + 555, + 715 + ], + "type": "text", + "content": " (breadcrumbs-ties), and " + }, + { + "bbox": [ + 313, + 590, + 555, + 715 + ], + "type": "inline_equation", + "content": "f_{\\text{merge}}^{\\text{Lines}}" + }, + { + "bbox": [ + 313, + 590, + 555, + 715 + ], + "type": "text", + "content": " (lines-ties). To put the results" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20483" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 73, + 218, + 228 + ], + "blocks": [ + { + "bbox": [ + 58, + 73, + 218, + 228 + ], + "lines": [ + { + "bbox": [ + 58, + 73, + 218, + 228 + ], + "spans": [ + { + "bbox": [ + 58, + 73, + 218, + 228 + ], + "type": "image", + "image_path": "ebd093cda73193ac5477e477267bffeaa313b54c949aebe43f0656e36ffbd749.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 231, + 555, + 265 + ], + "lines": [ + { + "bbox": [ + 55, + 231, + 555, + 265 + ], + "spans": [ + { + "bbox": [ + 55, + 231, + 555, + 265 + ], + "type": "text", + "content": "Figure 4. Improving offline merging. We identify two simple methods for adapting offline-merging methods to the temporal setting: (1) replaying data from previous tasks (best-(offline+replay)) and (2) recency-biased weighting of task checkpoints (best-(offline+replay+weighting)). With these method improvements, offline merging methods can match the replay baseline." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 225, + 72, + 386, + 228 + ], + "blocks": [ + { + "bbox": [ + 225, + 72, + 386, + 228 + ], + "lines": [ + { + "bbox": [ + 225, + 72, + 386, + 228 + ], + "spans": [ + { + "bbox": [ + 225, + 72, + 386, + 228 + ], + "type": "image", + "image_path": "a3bed99fb67c88813c27ba41aefb99d45e29aed54c8d3f2fba5ab1140fd70a8d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 389, + 72, + 553, + 228 + ], + "blocks": [ + { + "bbox": [ + 389, + 72, + 553, + 228 + ], + "lines": [ + { + "bbox": [ + 389, + 72, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 389, + 72, + 553, + 228 + ], + "type": "image", + "image_path": "2b76313ce61b1817010d8582c8fa3585db7e2df3537da818899c92169cc466d7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 275, + 297, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 275, + 297, + 502 + ], + "spans": [ + { + "bbox": [ + 54, + 275, + 297, + 502 + ], + "type": "text", + "content": "in context, we incorporate (1) a simple continual fine-tuning baseline (replay), following Sec. 4.1, which replays on both pretraining and previous task data, (2) initial zero-shot " + }, + { + "bbox": [ + 54, + 275, + 297, + 502 + ], + "type": "inline_equation", + "content": "(\\theta_0)" + }, + { + "bbox": [ + 54, + 275, + 297, + 502 + ], + "type": "text", + "content": " performance lower bound, and (3) multitask training upper bound. We visualize trajectories over time for knowledge accumulation " + }, + { + "bbox": [ + 54, + 275, + 297, + 502 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\mathrm{KA}}" + }, + { + "bbox": [ + 54, + 275, + 297, + 502 + ], + "type": "text", + "content": ", zero-shot retention " + }, + { + "bbox": [ + 54, + 275, + 297, + 502 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\mathrm{ZS}}" + }, + { + "bbox": [ + 54, + 275, + 297, + 502 + ], + "type": "text", + "content": ", and the geometric mean of both in Fig. 3. Our results clearly highlight that there are only marginal differences between merging techniques " + }, + { + "bbox": [ + 54, + 275, + 297, + 502 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{merge}}" + }, + { + "bbox": [ + 54, + 275, + 297, + 502 + ], + "type": "text", + "content": " when deployed in an offline manner for a temporal problem, and they all trace similar trajectories in the " + }, + { + "bbox": [ + 54, + 275, + 297, + 502 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\mathrm{KA}}" + }, + { + "bbox": [ + 54, + 275, + 297, + 502 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 275, + 297, + 502 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\mathrm{ZS}}" + }, + { + "bbox": [ + 54, + 275, + 297, + 502 + ], + "type": "text", + "content": " space and achieve similar final performance. We do find that the vanilla " + }, + { + "bbox": [ + 54, + 275, + 297, + 502 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{merge}}^{\\mathrm{TA}}" + }, + { + "bbox": [ + 54, + 275, + 297, + 502 + ], + "type": "text", + "content": " (task-arithmetic) starts to deviate after a few update steps, an issue fixed by the various subsequent extensions such as TIES, DARE or LiNeS. Overall, however, unlike straightforward continual fine-tuning (replay), offline merging with any technique fails to address the temporal aspects of the problem, particularly struggling to consistently acquire new knowledge over time (as observed in Fig. 3, left)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 503, + 296, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 503, + 296, + 528 + ], + "spans": [ + { + "bbox": [ + 55, + 503, + 296, + 528 + ], + "type": "text", + "content": "Key takeaway: offline merging with any technique yields similar results, failing to add new knowledge over time." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 538, + 293, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 538, + 293, + 551 + ], + "spans": [ + { + "bbox": [ + 55, + 538, + 293, + 551 + ], + "type": "text", + "content": "4.3. Can Replay or Time-Weighting Fix the Issue?" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 556, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 556, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 295, + 616 + ], + "type": "text", + "content": "We next ask: \"What are some simple extensions to offline methods to close the gap to the replay baseline\"? As the continual fine-tuning baseline replays on past data from all previous tasks while training at the current task " + }, + { + "bbox": [ + 55, + 556, + 295, + 616 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 556, + 295, + 616 + ], + "type": "text", + "content": ", can this task data-mixing also help offline merging methods?" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": "Data replay improves offline merging. Since offline methods operate entirely under a task-independent assumption, they fail to capture any temporal dependencies. Fig. 4 shows that simply applying data-replay on top of standard offline merging leads to significant boosts in the overall performance. For instance, best-(offline+replay) achieves " + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "inline_equation", + "content": "58.2\\%" + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": " compared to best-offline at " + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "inline_equation", + "content": "54.6\\%" + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": ", bringing it closer to the replay baseline. However, a notable performance gap" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 275, + 555, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 275, + 555, + 299 + ], + "spans": [ + { + "bbox": [ + 313, + 275, + 555, + 299 + ], + "type": "text", + "content": "remains, with best-(offline+replay) at " + }, + { + "bbox": [ + 313, + 275, + 555, + 299 + ], + "type": "inline_equation", + "content": "58.2\\%" + }, + { + "bbox": [ + 313, + 275, + 555, + 299 + ], + "type": "text", + "content": " falling short of replay at " + }, + { + "bbox": [ + 313, + 275, + 555, + 299 + ], + "type": "inline_equation", + "content": "59.1\\%" + }, + { + "bbox": [ + 313, + 275, + 555, + 299 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 300, + 556, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 556, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 556, + 384 + ], + "type": "text", + "content": "Recency-biased weighting helps. Next, unlike in standard offline averaging, where all task checkpoints are weighted uniformly, we impose temporal ordering via non-uniform weighting for offline merging. We explore several recency-biased, non-uniform weighting schemes, assigning higher weights to more recent tasks to account for the temporal nature of the setting." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 384, + 556, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 556, + 505 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 556, + 505 + ], + "type": "text", + "content": "We explore various discounting schemes: logarithmic, quadratic, exponential, and cubic, applied to the best offline merge replay method from the previous experiment (please refer to the supplementary for details). As shown in Fig. 4, these schemes improve performance, with best-offline+replay+weighting) reaching " + }, + { + "bbox": [ + 313, + 384, + 556, + 505 + ], + "type": "inline_equation", + "content": "58.9\\%" + }, + { + "bbox": [ + 313, + 384, + 556, + 505 + ], + "type": "text", + "content": " yet still falling slightly short of the replay baseline at " + }, + { + "bbox": [ + 313, + 384, + 556, + 505 + ], + "type": "inline_equation", + "content": "59.1\\%" + }, + { + "bbox": [ + 313, + 384, + 556, + 505 + ], + "type": "text", + "content": ". These results provide strong evidence that accounting for the new temporal axis is crucial for effective temporal model merging, even when implemented as an extension of offline merging." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 505, + 555, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 505, + 555, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 505, + 555, + 553 + ], + "type": "text", + "content": "Key Takeaway: Accounting for the time aspect is crucial for effective temporal model merging, even as an extension on top of standard offline merging. Still, a small gap to the simple replay baseline remains." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 563, + 397, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 397, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 397, + 574 + ], + "type": "text", + "content": "4.4. TIME Travel" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": "In our effort to adapt offline merging to the temporal setting, we have explored two axes of our TIME framework—merging technique and deployment. As discussed in Sec. 4.2, the choice of merging technique has minimal impact on final performance, while the deployment strategy plays a much more significant role. Next, we systematically explore the design space for temporal merging by testing all valid combinations of three initialization protocols and three deployment protocols described in Sec. 3.2. After discarding incompatible pairs, such as init_ZS with deployyFT, we evaluated the remaining eight variants us" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20484" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 72, + 218, + 228 + ], + "blocks": [ + { + "bbox": [ + 58, + 72, + 218, + 228 + ], + "lines": [ + { + "bbox": [ + 58, + 72, + 218, + 228 + ], + "spans": [ + { + "bbox": [ + 58, + 72, + 218, + 228 + ], + "type": "image", + "image_path": "639e655301947a8811b4e6988ed89b22f627cbed4850befffae695104730f1dc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 231, + 555, + 254 + ], + "lines": [ + { + "bbox": [ + 55, + 231, + 555, + 254 + ], + "spans": [ + { + "bbox": [ + 55, + 231, + 555, + 254 + ], + "type": "text", + "content": "Figure 5. A journey through TIME. We explore various initialization and deployment protocols, finding that the EMA initialization-deployment strikes the best balance between knowledge accumulation and zero-shot retention. We refer to this strategy as Best-in-TIME." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 225, + 72, + 391, + 228 + ], + "blocks": [ + { + "bbox": [ + 225, + 72, + 391, + 228 + ], + "lines": [ + { + "bbox": [ + 225, + 72, + 391, + 228 + ], + "spans": [ + { + "bbox": [ + 225, + 72, + 391, + 228 + ], + "type": "image", + "image_path": "9018a050b489fbfba94241270d6ae30891521d44371f681a3b6be35510469d5f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 394, + 72, + 553, + 228 + ], + "blocks": [ + { + "bbox": [ + 394, + 72, + 553, + 228 + ], + "lines": [ + { + "bbox": [ + 394, + 72, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 394, + 72, + 553, + 228 + ], + "type": "image", + "image_path": "f7eaf2208b1b1b57e2c95e02a525818266bb39ebd475943f00c1e3a39c631018.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 58, + 257, + 219, + 413 + ], + "blocks": [ + { + "bbox": [ + 58, + 257, + 219, + 413 + ], + "lines": [ + { + "bbox": [ + 58, + 257, + 219, + 413 + ], + "spans": [ + { + "bbox": [ + 58, + 257, + 219, + 413 + ], + "type": "image", + "image_path": "f9e3177a66ec2ed32b5e3807233b2ce5d05124ee1757d422ac950a2703ffb210.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 415, + 555, + 437 + ], + "lines": [ + { + "bbox": [ + 55, + 415, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 555, + 437 + ], + "type": "text", + "content": "Figure 6. Sweeping Best-in-TIME. All merging techniques perform well with the Best-in-TIME strategy. Indeed, there are no significant differences between techniques, indicating that initialization and deployment matter more for temporal merging." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 225, + 257, + 391, + 413 + ], + "blocks": [ + { + "bbox": [ + 225, + 257, + 391, + 413 + ], + "lines": [ + { + "bbox": [ + 225, + 257, + 391, + 413 + ], + "spans": [ + { + "bbox": [ + 225, + 257, + 391, + 413 + ], + "type": "image", + "image_path": "1dd6998220954e558705d17b1fb70bccbb96fab789a5e4db0bd6c00a66bacaed.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 392, + 257, + 553, + 413 + ], + "blocks": [ + { + "bbox": [ + 392, + 257, + 553, + 413 + ], + "lines": [ + { + "bbox": [ + 392, + 257, + 553, + 413 + ], + "spans": [ + { + "bbox": [ + 392, + 257, + 553, + 413 + ], + "type": "image", + "image_path": "0f3a09fe0ac769d0af23e33f4f49b5b1dacd65324835466a103972dd2328c2f0.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 449, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 449, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 54, + 449, + 295, + 567 + ], + "type": "text", + "content": "ing weight averaging as the merging technique. As shown in Fig. 5, the choice of initialization and deployment strategy largely determines performance, significantly affecting both knowledge accumulation and retention. One combination that stands out consistently is initEMA alongside deployyEMA. This supports the findings of [68, 76] on the applicability of EMA-like solutions to small-scale continual learning and pretraining, while extending the analysis to a much larger, more systematic study of temporal model merging." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 567, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 567, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 54, + 567, + 295, + 628 + ], + "type": "text", + "content": "Since EMA-style model merging achieves a better balance between knowledge accumulation and retention than other methods, we call this approach Best-in-TIME. In the next section, we will explore the robustness of this strategy across different merging techniques." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 628, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 628, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 295, + 652 + ], + "type": "text", + "content": "Key Takeaway: The EMA strategy for initialization and deployment consistently performs the best." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 659, + 274, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 274, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 274, + 673 + ], + "type": "text", + "content": "4.5. What is the best merge for Best-in-TIME?" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "Having identified the optimal initialization and deployment merging strategy in the previous section using weight-averaging as a canonical merging technique, we now invest" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 449, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 555, + 628 + ], + "type": "text", + "content": "tigate the robustness of our finding by sweeping over other merging techniques, similar to the offline-merging analysis in Sec. 4.2. In particular, we test 7 different merging techniques while keeping the Best-in-TIME initialization and deployment strategy. From Fig. 6, it is immediately evident that all merging techniques perform very similarly—in fact, the knowledge accumulation " + }, + { + "bbox": [ + 313, + 449, + 555, + 628 + ], + "type": "inline_equation", + "content": "(\\mathcal{A}_{KA})" + }, + { + "bbox": [ + 313, + 449, + 555, + 628 + ], + "type": "text", + "content": " of all techniques lie within a single standard deviation of performance variation across multiple runs. This indicates that, over a sufficiently long time horizon, all techniques converge to a similar behavior, echoing our results in Sec. 4.2. However, we do notice higher variance in the retention metric " + }, + { + "bbox": [ + 313, + 449, + 555, + 628 + ], + "type": "inline_equation", + "content": "(\\mathcal{A}_{ZS})" + }, + { + "bbox": [ + 313, + 449, + 555, + 628 + ], + "type": "text", + "content": ". Overall, our results suggest that for temporal model merging, the specific merging technique used is much less important than selecting the best initialization and deployment strategies." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 628, + 554, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 628, + 554, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 628, + 554, + 652 + ], + "type": "text", + "content": "Key Takeaway: The choice of merging technique is not particularly important—all techniques perform similarly." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 659, + 512, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 659, + 512, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 659, + 512, + 673 + ], + "type": "text", + "content": "4.6. Scaling Up Temporal Model Merging" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": "We next study the scaling behaviour of temporal model merging across three-axes: model size, compute budget, and number of tasks (results in Fig. 7 and supplen" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20485" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 69, + 175, + 194 + ], + "blocks": [ + { + "bbox": [ + 61, + 69, + 175, + 194 + ], + "lines": [ + { + "bbox": [ + 61, + 69, + 175, + 194 + ], + "spans": [ + { + "bbox": [ + 61, + 69, + 175, + 194 + ], + "type": "image", + "image_path": "ce61e309977d426698779fdbea703da737579f65da74ee43dd42edb9c34de121.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 206, + 295, + 274 + ], + "lines": [ + { + "bbox": [ + 55, + 206, + 295, + 274 + ], + "spans": [ + { + "bbox": [ + 55, + 206, + 295, + 274 + ], + "type": "text", + "content": "Figure 7. Scaling up model merging. (left) With scale, we observe continued improvements of model merging compared to the standard replay baseline. (right) Our Best-in-TIME method continues to improve with scaled total compute budget moving close to the multitask upper-bound. Gray points in the plot visualize suboptimal Best-in-TIME hyperparameter-instantiations." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 178, + 70, + 291, + 194 + ], + "blocks": [ + { + "bbox": [ + 178, + 70, + 291, + 194 + ], + "lines": [ + { + "bbox": [ + 178, + 70, + 291, + 194 + ], + "spans": [ + { + "bbox": [ + 178, + 70, + 291, + 194 + ], + "type": "image", + "image_path": "996c1b71b013f13effc7c881444e329b4560e24441b189ecfa91c6da5f5592c9.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 288, + 295, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 288, + 295, + 324 + ], + "spans": [ + { + "bbox": [ + 55, + 288, + 295, + 324 + ], + "type": "text", + "content": "tary). All our experiments utilize the Best-in-TIME setup described previously, conducting hyperparameter-optimal EMA-weight-averaging at each task." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 328, + 295, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 328, + 295, + 592 + ], + "spans": [ + { + "bbox": [ + 56, + 328, + 295, + 592 + ], + "type": "text", + "content": "Scaling the Model. As we increase the model scale from " + }, + { + "bbox": [ + 56, + 328, + 295, + 592 + ], + "type": "inline_equation", + "content": "S / 16" + }, + { + "bbox": [ + 56, + 328, + 295, + 592 + ], + "type": "text", + "content": " (62.3M parameters) to " + }, + { + "bbox": [ + 56, + 328, + 295, + 592 + ], + "type": "inline_equation", + "content": "B / 16" + }, + { + "bbox": [ + 56, + 328, + 295, + 592 + ], + "type": "text", + "content": " (149.6M), " + }, + { + "bbox": [ + 56, + 328, + 295, + 592 + ], + "type": "inline_equation", + "content": "L / 14" + }, + { + "bbox": [ + 56, + 328, + 295, + 592 + ], + "type": "text", + "content": " (427.6M), and finally " + }, + { + "bbox": [ + 56, + 328, + 295, + 592 + ], + "type": "inline_equation", + "content": "g / 14" + }, + { + "bbox": [ + 56, + 328, + 295, + 592 + ], + "type": "text", + "content": " (1.37B) in Fig. 7 (left), we compare sequential fine-tuning (circles) and Best-in-TIME (squares) in terms of knowledge accumulation and retention. We find that: (1) Best-in-TIME scales well with model size, with larger models exhibiting increased affinity to temporal merging. This extends and further corroborates standard offline merging insights by Yadav et al. [89], who showed that model scale facilitates merging. (2) Moreover, while Roth et al. [68] highlight better continual fine-tuning with scale (as also studied by Ibrahim et al. [23]), we show temporal model merging to be substantially more effective across scale. For larger models all the way to the largest ViT-g/14, Best-in-TIME vastly outperforms or matches sequential fine-tuning and the multitask target in knowledge retention and positive backward transfer. Moreover, scale facilitates equivalent degrees of knowledge accumulation between sequential fine-tuning and temporal model merging. Therefore, our model scaling results strongly advocate the use of temporal model merging solutions over standard continual fine-tuning methods [23, 68]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": "Scaling the Compute. Keeping the model fixed to ViTB/16, we next scale up the available compute budget by increasing the number of allocated update steps per task. We compare a multitask model, trained on all tasks simultaneously, to a budget-optimal Best-in-TIME. The only hyperparameter for Best-in-TIME is the interpolation weight " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": ". For each compute budget, there is a clear optimal choice of that hyperparameter (suboptimal runs shown as gray dots in Fig. 7 (right)). Higher values of " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": " put greater emphasis on accumulation, allowing optimal accumulation-retention" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "content": "trade-offs to be reached at lower compute budgets. However, if a larger compute budget is available, less aggressive temporal model merging can achieve higher absolute trade-offs. Note that in Fig. 7 (right), we report the geometric mean between accumulation and retention, corresponding to the right-most panel in previous plots. Our results indicate that Best-in-TIME scales very well across compute budgets, clearly approaching the multitask upper bound in accumulation-retention balance at larger compute budgets. Put together, this gives practitioners the flexibility to adjust Best-in-TIME to their compute budget through the choice of " + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "content": ". Moreover, with more compute to train each respective expert model, the merging process becomes more effective." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 229, + 553, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 229, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 229, + 553, + 361 + ], + "type": "text", + "content": "Scaling the Number of Tasks. We next study how Best-in-TIME performs as we increase the number of merging time-steps from " + }, + { + "bbox": [ + 313, + 229, + 553, + 361 + ], + "type": "inline_equation", + "content": "T = 20" + }, + { + "bbox": [ + 313, + 229, + 553, + 361 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 229, + 553, + 361 + ], + "type": "inline_equation", + "content": "T = 50" + }, + { + "bbox": [ + 313, + 229, + 553, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 229, + 553, + 361 + ], + "type": "inline_equation", + "content": "T = 100" + }, + { + "bbox": [ + 313, + 229, + 553, + 361 + ], + "type": "text", + "content": ". As shown in Figs. 10 and 11, Best-in-TIME remains the optimal method across different initialization and deployment strategies. For " + }, + { + "bbox": [ + 313, + 229, + 553, + 361 + ], + "type": "inline_equation", + "content": "T = 50" + }, + { + "bbox": [ + 313, + 229, + 553, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 229, + 553, + 361 + ], + "type": "inline_equation", + "content": "T = 100" + }, + { + "bbox": [ + 313, + 229, + 553, + 361 + ], + "type": "text", + "content": ", we notice that other methods are initially competitive, but face a significant drop-off in performance near the end of the task sequence, while best-in-TIME continues to improve monotonically. This further corroborates that our Best-in-TIME is an extremely scalable method for temporal model merging." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 363, + 553, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 363, + 553, + 412 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 553, + 412 + ], + "type": "text", + "content": "Key Takeaway: Temporal model merging with Best-in-TIME scales efficiently across model sizes and the number of tasks. Moreover, compute scaling improves the effectiveness of temporal model merging." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 428, + 388, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 428, + 388, + 441 + ], + "spans": [ + { + "bbox": [ + 314, + 428, + 388, + 441 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "text", + "content": "In this work, we study temporal model merging, addressing the challenge of continually merging multimodal models as new tasks and data arrive, and new expert models are trained in succession. To formalize this setting, we propose TIME, a novel unifying framework breaking down temporal model merging into three key axes: (1) initialization phase defining starting weights before each task, (2) deployment phase denoting post-training expert model aggregation, and (3) the choice of merging technique. Using TIME, we conduct a large-scale systematic study uncovering crucial practical guidelines for temporal model merging. Our experiments on the FoMo-in-Flux benchmark spanning 63 datasets, showcase that accounting for the temporal aspect is crucial, with standard offline merging techniques falling short in this dynamic setting. Moreover, we find the particular choice of merging technique matters far less than the merging strategy for initialization and deployment. Finally, we introduce Best-in-TIME, which scales favorably with model size and outperforms existing methods for continual multimodal pretraining. Together, our work provides a systematic entry point into temporal model merging and establishes best practices for this emerging field." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "20486" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 370 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 370 + ], + "type": "text", + "content": "Acknowledgements. The authors would like to thank (in random order) Shyamgopal Karthik, Shashwat Goel, Ankit Sonthalia, Olivier Henaff, Alexandre Ramé, and Daniel Marczak for helpful feedback. The plotting style in our work is inspired by figures from Beyer et al. [4]. The style of Fig. 2 is inspired by Figure 1 of Karamcheti et al. [30]. VU, KR, and SD thank the International Max Planck Research School for Intelligent Systems (IMPRS-IS). VU, KR, and SD also thank the European Laboratory for Learning and Intelligent Systems (ELLIS) PhD program for support. VU is supported by a Google PhD Fellowship in Machine Intelligence. SA is supported by a Newton Trust Grant. AP is supported by the Federal Ministry of Education and Research (BMBF), FKZ: 011524085B. MB acknowledges financial support via the Open Philanthropy Foundation funded by the Good Ventures Foundation. MB is a member of the Machine Learning Cluster of Excellence, funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany's Excellence Strategy - EXC number 2064/1 - Project number 390727645. ZA acknowledges the support from the German Research Foundation (DFG): SFB 1233, Robust Vision: Inference Principles and Neural Mechanisms, project number: 276693517 and ERC Grant DEXIM, project number: 853489. This research utilized compute resources at the Tübingen Machine Learning Cloud, DFG FKZ INST 37/1057-1 FUGG." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 392, + 115, + 404 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 392, + 115, + 404 + ], + "spans": [ + { + "bbox": [ + 56, + 392, + 115, + 404 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 412, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 66, + 412, + 295, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 412, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 66, + 412, + 295, + 456 + ], + "type": "text", + "content": "[1] Samuel Ainsworth, Jonathan Hayase, and Siddhartha Srinivasa. Git re-basin: Merging models modulo permutation symmetries. In The Eleventh International Conference on Learning Representations, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 458, + 295, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 458, + 295, + 490 + ], + "spans": [ + { + "bbox": [ + 66, + 458, + 295, + 490 + ], + "type": "text", + "content": "[2] Takuya Akiba, Makoto Shing, Yujin Tang, Qi Sun, and David Ha. Evolutionary optimization of model merging recipes. arXiv preprint arXiv:2403.13187, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 491, + 295, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 491, + 295, + 535 + ], + "spans": [ + { + "bbox": [ + 66, + 491, + 295, + 535 + ], + "type": "text", + "content": "[3] Anton Alexandrov, Veselin Raychev, Mark Niklas Müller, Ce Zhang, Martin Vechev, and Kristina Toutanova. Mitigating catastrophic forgetting in language transfer via model merging. arXiv preprint arXiv:2407.08699, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 66, + 536, + 294, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 536, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 66, + 536, + 294, + 601 + ], + "type": "text", + "content": "[4] Lucas Beyer, Xiaohua Zhai, Amélie Royer, Larisa Markeeva, Rohan Anil, and Alexander Kolesnikov. Knowledge distillation: A good teacher is patient and consistent. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10925-10934, 2022. 9, 14" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 66, + 604, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 604, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 604, + 295, + 713 + ], + "type": "text", + "content": "[5] Rishi Bommasani, Drew A. Hudson, Ehsan Adeli, Russ B. Altman, Simran Arora, Sydney von Arx, Michael S. Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, Erik Brynjolfsson, Shyamal Buch, Dallas Card, Rodrigo Castellon, Niladri S. Chatterji, Annie S. Chen, Kathleen Creel, Jared Quincy Davis, Dorottya Demszky, Chris Donahue, Moussa Doumbouya, Esin Durmus, Stefano Ermon, John Etchemendy, Kawin Ethayarajh, Li Fei-Fei, Chelsea Finn, Trevor Gale, Lauren Gillespie, Karan Goel, Noah D. Goodman, Shelby Grossman, Neel Guha, Tat" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 338, + 73, + 555, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 73, + 555, + 149 + ], + "spans": [ + { + "bbox": [ + 338, + 73, + 555, + 149 + ], + "type": "text", + "content": "sunori Hashimoto, Peter Henderson, John Hewitt, Daniel E. Ho, Jenny Hong, Kyle Hsu, Jing Huang, Thomas Icard, Saahil Jain, Dan Jurafsky, Pratyusha Kalluri, Siddharth Karamcheti, Geoff Keeling, Fereshte Khani, Omar Khattab, Pang Wei Koh, Mark S. Krass, Ranjay Krishna, Rohith Kuditipudi, and et al. On the opportunities and risks of foundation models. CoRR, abs/2108.07258, 2021. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 325, + 151, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 151, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 325, + 151, + 553, + 205 + ], + "type": "text", + "content": "[6] Jorg Bornschein, Alexandre Galashov, Ross Hemsley, Amal Rannen-Triki, Yutian Chen, Arslan Chaudhry, Xu Owen He, Arthur Douillard, Massimo Caccia, Qixuan Feng, et al. Nevis' 22: A stream of 100 tasks sampled from 30 years of computer vision research. JMLR, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 325, + 206, + 553, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 206, + 553, + 238 + ], + "spans": [ + { + "bbox": [ + 325, + 206, + 553, + 238 + ], + "type": "text", + "content": "[7] Zhipeng Cai, Ozan Sener, and Vladlen Koltun. Online continual learning with natural distribution shifts: An empirical study with visual data. In ICCV, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 325, + 239, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 239, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 325, + 239, + 553, + 293 + ], + "type": "text", + "content": "[8] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. arXiv preprint arXiv:2212.07143, 2022. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 325, + 294, + 553, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 294, + 553, + 337 + ], + "spans": [ + { + "bbox": [ + 325, + 294, + 553, + 337 + ], + "type": "text", + "content": "[9] Leshem Choshen, Elad Venezian, Shachar Don-Yehia, Noam Slonim, and Yoav Katz. Where to start? analyzing the potential value of intermediate models. arXiv preprint arXiv:2211.00107, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 320, + 338, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 338, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 320, + 338, + 553, + 381 + ], + "type": "text", + "content": "[10] Geoffrey Cideron, Andrea Agostinelli, Johan Ferret, Sertan Girgin, Romuald Elie, Olivier Bachem, Sarah Perrin, and Alexandre Rame. Diversity-rewarded CFG distillation. arXiv preprint arXiv:2410.06084, 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 320, + 383, + 553, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 383, + 553, + 436 + ], + "spans": [ + { + "bbox": [ + 320, + 383, + 553, + 436 + ], + "type": "text", + "content": "[11] MohammadReza Davari and Eugene Belilovsky. Model breadcrumbs: Scaling multi-task model merging with sparse masks. In Computer Vision - ECCV 2024, pages 270-287, Cham, 2025. Springer Nature Switzerland. 2, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 320, + 437, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 437, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 320, + 437, + 553, + 480 + ], + "type": "text", + "content": "[12] Pala Tej Deep, Rishabh Bhardwaj, and Soujanya Poria. Della-merging: Reducing interference in model merging through magnitude-based sampling. arXiv preprint arXiv:2406.11617, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 482, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 482, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 320, + 482, + 553, + 525 + ], + "type": "text", + "content": "[13] Shachar Don-Yehiya, Elad Venezian, Colin Raffel, Noam Slonim, Yoav Katz, and Leshem Choshen. Cold fusion: Collaborative descent for distributed multitask finetuning. arXiv preprint arXiv:2212.01378, 2022. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 526, + 553, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 526, + 553, + 580 + ], + "spans": [ + { + "bbox": [ + 320, + 526, + 553, + 580 + ], + "type": "text", + "content": "[14] Jonathan Frankle, Gintare Karolina Dziugaite, Daniel Roy, and Michael Carbin. Linear mode connectivity and the lottery ticket hypothesis. In Proceedings of the 37th International Conference on Machine Learning, pages 3259-3269. PMLR, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 582, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 582, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 320, + 582, + 553, + 613 + ], + "type": "text", + "content": "[15] João Gama, Indre Žliobaité, Albert Bifet, Mykola Pechenizkiy, and Abdelhamid Bouchachia. A survey on concept drift adaptation. ACM computing surveys, 2014. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 615, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 615, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 320, + 615, + 553, + 658 + ], + "type": "text", + "content": "[16] Saurabh Garg, Mehrdad Farajtabar, Hadi Pouransari, Raviteja Vemulapalli, Sachin Mehta, Oncel Tuzel, Vaishaal Shankar, and Fartash Faghri. Tic-clip: Continual training of clip models. In ICLR, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 659, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 659, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 320, + 659, + 553, + 712 + ], + "type": "text", + "content": "[17] Timur Garipov, Pavel Izmailov, Dmitrii Podoprikhin, Dmitry P Vetrov, and Andrew G Wilson. Loss surfaces, mode connectivity, and fast ensembling of dnns. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2018. 1" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20487" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 72, + 297, + 715 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 72, + 297, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 72, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 61, + 72, + 297, + 127 + ], + "type": "text", + "content": "[18] Charles Goddard, Shamane Siriwardhana, Malikeh Ehghaghi, Luke Meyers, Vlad Karpukhin, Brian Benedict, Mark McQuade, and Jacob Solawetz. Arcee's mergekit: A toolkit for merging large language models. arXiv preprint arXiv:2403.13257, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 128, + 296, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 128, + 296, + 183 + ], + "spans": [ + { + "bbox": [ + 61, + 128, + 296, + 183 + ], + "type": "text", + "content": "[19] Alexey Gorbatovski, Boris Shaposhnikov, Alexey Malakhov, Nikita Surnachev, Yaroslav Aksenov, Ian Maksimov, Nikita Balagansky, and Daniil Gavrilov. Learn your reference model for real good alignment. arXiv preprint arXiv:2404.09656, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 184, + 296, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 184, + 296, + 237 + ], + "spans": [ + { + "bbox": [ + 61, + 184, + 296, + 237 + ], + "type": "text", + "content": "[20] Zhongrui Gui, Shuyang Sun, Runjia Li, Jianhao Yuan, Zhaochong An, Karsten Roth, Ameya Prabhu, and Philip Torr. knn-clip: Retrieval enables training-free segmentation on continually expanding large vocabularies. arXiv preprint arXiv:2404.09447, 2024. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 238, + 296, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 238, + 296, + 282 + ], + "spans": [ + { + "bbox": [ + 61, + 238, + 296, + 282 + ], + "type": "text", + "content": "[21] Yifei He, Yuzheng Hu, Yong Lin, Tong Zhang, and Han Zhao. Localize-and-stitch: Efficient model merging via sparse task arithmetic. arXiv preprint arXiv:2408.13656, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 283, + 296, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 283, + 296, + 337 + ], + "spans": [ + { + "bbox": [ + 61, + 283, + 296, + 337 + ], + "type": "text", + "content": "[22] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 338, + 296, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 338, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 61, + 338, + 296, + 392 + ], + "type": "text", + "content": "[23] Adam Ibrahim, Benjamin Thérien, Kshitij Gupta, Mats L Richter, Quentin Anthony, Timothée Lesort, Eugene Belilovsky, and Irina Rish. Simple and scalable strategies to continually pre-train large language models. arXiv preprint arXiv:2403.08763, 2024. 2, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 393, + 296, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 393, + 296, + 446 + ], + "spans": [ + { + "bbox": [ + 61, + 393, + 296, + 446 + ], + "type": "text", + "content": "[24] Gabriel Ilharco, Mitchell Wortsman, Samir Yitzhak Gadre, Shuran Song, Hannaneh Hajishirzi, Simon Kornblith, Ali Farhadi, and Ludwig Schmidt. Patching open-vocabulary models by interpolating weights. In NeurIPS, 2022. 1, 2, 4, 14" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 448, + 296, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 448, + 296, + 502 + ], + "spans": [ + { + "bbox": [ + 61, + 448, + 296, + 502 + ], + "type": "text", + "content": "[25] Gabriel Ilharco, Marco Tulio Ribeiro, Mitchell Wortsman, Ludwig Schmidt, Hannaneh Hajishirzi, and Ali Farhadi. Editing models with task arithmetic. In The Eleventh International Conference on Learning Representations, 2023. 2, 4, 14" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 503, + 296, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 503, + 296, + 548 + ], + "spans": [ + { + "bbox": [ + 61, + 503, + 296, + 548 + ], + "type": "text", + "content": "[26] Pavel Izmailov, Dmitrii Podoprikhin, Timur Garipov, Dmitry Vetrov, and Andrew Gordon Wilson. Averaging weights leads to wider optima and better generalization. arXiv preprint arXiv:1803.05407, 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 61, + 548, + 296, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 548, + 296, + 613 + ], + "spans": [ + { + "bbox": [ + 61, + 548, + 296, + 613 + ], + "type": "text", + "content": "[27] Dong-Hwan Jang, Sangdoo Yun, and Dongyoon Han. Model stock: All we need is just a few fine-tuned models. In Computer Vision - ECCV 2024: 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part XLIV, page 207-223, Berlin, Heidelberg, 2024. Springer-Verlag. 1, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 61, + 614, + 296, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 614, + 296, + 658 + ], + "spans": [ + { + "bbox": [ + 61, + 614, + 296, + 658 + ], + "type": "text", + "content": "[28] Xisen Jin, Xiang Ren, Daniel Preotiac-Pietro, and Pengxiang Cheng. Dataless knowledge fusion by merging weights of language models. In The Eleventh International Conference on Learning Representations, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 61, + 658, + 296, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 658, + 296, + 691 + ], + "spans": [ + { + "bbox": [ + 61, + 658, + 296, + 691 + ], + "type": "text", + "content": "[29] Jean Kaddour. Stop wasting my time! saving days of imageet and bert training with latest weight averaging. arXiv preprint arXiv:2209.14981, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 61, + 691, + 296, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 691, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 61, + 691, + 296, + 715 + ], + "type": "text", + "content": "[30] Siddharth Karamcheti, Suraj Nair, Ashwin Balakrishna, Percy Liang, Thomas Kollar, and Dorsa Sadigh." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 555, + 715 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 338, + 73, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 73, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 338, + 73, + 555, + 106 + ], + "type": "text", + "content": "Prismatic vlms: Investigating the design space of visually-conditioned language models. arXiv preprint arXiv:2402.07865, 2024.9" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 320, + 106, + 554, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 106, + 554, + 160 + ], + "spans": [ + { + "bbox": [ + 320, + 106, + 554, + 160 + ], + "type": "text", + "content": "[31] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. PNAS, 2017. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 161, + 554, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 161, + 554, + 228 + ], + "spans": [ + { + "bbox": [ + 320, + 161, + 554, + 228 + ], + "type": "text", + "content": "[32] Pang Wei Koh, Shiori Sagawa, Henrik Marklund, Sang Michael Xie, Marvin Zhang, Akshay Balsubramani, Weihua Hu, Michihiro Yasunaga, Richard Lanas Phillips, Irena Gao, et al. Wilds: A benchmark of in-the-wild distribution shifts. In International conference on machine learning, pages 5637-5664. PMLR, 2021. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 228, + 554, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 228, + 554, + 304 + ], + "spans": [ + { + "bbox": [ + 320, + 228, + 554, + 304 + ], + "type": "text", + "content": "[33] Jeffrey Li, Mohammadreza Armandpour, Seyed Iman Mirzadeh, Sachin Mehta, Vaishaal Shankar, Raviteja Vemulapalli, Oncel Tuzel, Mehrdad Farajtabar, Hadi Pouransari, and Fartash Faghri. Tic-lm: A multi-year benchmark for continual pretraining of language models. In NeurIPS 2024 Workshop on Scalable Continual Learning for Lifelong Foundation Models. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 305, + 554, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 305, + 554, + 358 + ], + "spans": [ + { + "bbox": [ + 320, + 305, + 554, + 358 + ], + "type": "text", + "content": "[34] Margaret Li, Suchin Gururangan, Tim Dettmers, Mike Lewis, Tim Althoff, Noah A Smith, and Luke Zettlemoyer. Branch-train-merge: Embarrassingly parallel training of expert language models. arXiv preprint arXiv:2208.03306, 2022. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 360, + 554, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 360, + 554, + 403 + ], + "spans": [ + { + "bbox": [ + 320, + 360, + 554, + 403 + ], + "type": "text", + "content": "[35] Tao Li, Weisen Jiang, Fanghui Liu, Xiaolin Huang, and James T Kwok. Learning scalable model soup on a singlegpu: An efficient subspace training strategy. arXiv preprint arXiv:2407.03641, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 404, + 554, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 404, + 554, + 492 + ], + "spans": [ + { + "bbox": [ + 320, + 404, + 554, + 492 + ], + "type": "text", + "content": "[36] Yong Lin, Hangyu Lin, Wei Xiong, Shizhe Diao, Jianmeng Liu, Jipeng Zhang, Rui Pan, Haoxiang Wang, Wenbin Hu, Hanning Zhang, Hanze Dong, Renjie Pi, Han Zhao, Nan Jiang, Heng Ji, Yuan Yao, and Tong Zhang. Mitigating the alignment tax of RLHF. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 580-606, Miami, Florida, USA, 2024. Association for Computational Linguistics. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 492, + 554, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 492, + 554, + 525 + ], + "spans": [ + { + "bbox": [ + 320, + 492, + 554, + 525 + ], + "type": "text", + "content": "[37] Zhiqiu Lin, Jia Shi, Deepak Pathak, and Deva Ramanan. The clear benchmark: Continual learning on real-world imagery. In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 525, + 554, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 525, + 554, + 601 + ], + "spans": [ + { + "bbox": [ + 320, + 525, + 554, + 601 + ], + "type": "text", + "content": "[38] Adam Liska, Tomas Kocisky, Elena Gribovskaya, Tayfun Terzi, Eren Sezener, Devang Agrawal, D'Autume Cyprien De Masson, Tim Scholtes, Manzil Zaheer, Susannah Young, et al. Streamingqa: A benchmark for adaptation to new knowledge over time in question answering models. In International Conference on Machine Learning, pages 13604-13622. PMLR, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 602, + 554, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 602, + 554, + 669 + ], + "spans": [ + { + "bbox": [ + 320, + 602, + 554, + 669 + ], + "type": "text", + "content": "[39] Tianlin Liu, Shangmin Guo, Leonardo Bianco, Daniele Calandriello, Quentin Berthet, Felipe Llinares-López, Jessica Hoffmann, Lucas Dixon, Michal Valko, and Mathieu Blondel. Decoding-time realignment of language models. In Proceedings of the 41st International Conference on Machine Learning, pages 31015-31031. PMLR, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 669, + 554, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 669, + 554, + 691 + ], + "spans": [ + { + "bbox": [ + 320, + 669, + 554, + 691 + ], + "type": "text", + "content": "[40] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 691, + 554, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 691, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 320, + 691, + 554, + 715 + ], + "type": "text", + "content": "[41] Jinliang Lu, Ziliang Pang, Min Xiao, Yaochen Zhu, Rui Xia, and Jiajun Zhang. Merge, ensemble, and cooperate!" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20488" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 81, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 81, + 73, + 294, + 95 + ], + "type": "text", + "content": "a survey on collaborative strategies in the era of large language models. arXiv preprint arXiv:2407.06089, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 96, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 96, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 61, + 96, + 295, + 140 + ], + "type": "text", + "content": "[42] Daniel Marczak, Bartlomiej Twardowski, Tomasz Trzeciński, and Sebastian Cygert. Magmax: Leveraging model merging for seamless continual learning. arXiv preprint arXiv:2407.06322, 2024. 2, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 141, + 294, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 141, + 294, + 174 + ], + "spans": [ + { + "bbox": [ + 61, + 141, + 294, + 174 + ], + "type": "text", + "content": "[43] Michael S Matena and Colin Raffel. Merging models with fisher-weighted averaging. In Advances in Neural Information Processing Systems, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 175, + 295, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 175, + 295, + 229 + ], + "spans": [ + { + "bbox": [ + 62, + 175, + 295, + 229 + ], + "type": "text", + "content": "[44] Matías Mendieta, Boran Han, Xingjian Shi, Yi Zhu, and Chen Chen. Towards geospatial foundation models via continual pretraining. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16806-16816, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 232, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 232, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 62, + 232, + 294, + 274 + ], + "type": "text", + "content": "[45] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. In The Eleventh International Conference on Learning Representations, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 277, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 277, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 62, + 277, + 294, + 374 + ], + "type": "text", + "content": "[46] Remi Munos, Michal Valko, Daniele Calandriello, Mohammad Gheshlaghi Azar, Mark Rowland, Zhaohan Daniel Guo, Yunhao Tang, Matthieu Geist, Thomas Mesnard, Come Fiegel, Andrea Michi, Marco Selvi, Sertan Girgin, Nikola Momchev, Olivier Bachem, Daniel J Mankowitz, Doina Precup, and Bilal Piot. Nash learning from human feedback. In Proceedings of the 41st International Conference on Machine Learning, pages 36743-36768. PMLR, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 376, + 294, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 376, + 294, + 409 + ], + "spans": [ + { + "bbox": [ + 62, + 376, + 294, + 409 + ], + "type": "text", + "content": "[47] Anshul Nasery, Jonathan Hayase, Pang Wei Koh, and Sewoong Oh. Pleas-merging models with permutations and least squares. arXiv preprint arXiv:2407.02447, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 411, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 411, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 62, + 411, + 294, + 453 + ], + "type": "text", + "content": "[48] Behnam Neyshabur, Hanie Sedghi, and Chiyuan Zhang. What is being transferred in transfer learning? In Advances in Neural Information Processing Systems, pages 512-523. Curran Associates, Inc., 2020. 2, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 456, + 294, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 456, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 62, + 456, + 294, + 487 + ], + "type": "text", + "content": "[49] Kai Nylund, Suchin Gururangan, and Noah A Smith. Time is encoded in the weights of finetuned language models. arXiv preprint arXiv:2312.13401, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 490, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 490, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 62, + 490, + 294, + 533 + ], + "type": "text", + "content": "[50] Changdaee Oh, Yixuan Li, Kyungwoo Song, Sangdoo Yun, and Dongyoon Han. Dawn: Training-free dynamic weight interpolation for robust adaptation. arXiv preprint arXiv:2410.03782, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 535, + 294, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 535, + 294, + 589 + ], + "spans": [ + { + "bbox": [ + 62, + 535, + 294, + 589 + ], + "type": "text", + "content": "[51] Oleksiy Ostapenko, Timothee Lesort, Pau Rodríguez, Md Rifat Arefin, Arthur Douillard, Irina Rish, and Laurent Charlin. Continual learning with foundation models: An empirical study of latent replay. In Conference on Lifelong Learning Agents (CoLLAs), 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 590, + 294, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 590, + 294, + 623 + ], + "spans": [ + { + "bbox": [ + 62, + 590, + 294, + 623 + ], + "type": "text", + "content": "[52] Jyothish Pari, Samy Jelassi, and Pulkit Agrawal. Collective model intelligence requires compatible specialization. arXiv preprint arXiv:2411.02207, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 624, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 624, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 62, + 624, + 294, + 689 + ], + "type": "text", + "content": "[53] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 691, + 294, + 713 + ], + "type": "text", + "content": "[54] Ameya Prabhu, Zhipeng Cai, Puneet Dokania, Philip Torr, Vladlen Koltun, and Ozan Sener. Online contin" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "type": "text", + "content": "ual learning without the storage constraint. arXiv preprint arXiv:2305.09253, 2023. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 321, + 96, + 553, + 140 + ], + "type": "text", + "content": "[55] Ameya Prabhu, Hasan Abed Al Kader Hammoud, Puneet Dokania, Philip HS Torr, Ser-Nam Lim, Bernard Ghanem, and Adel Bibi. Computationally budgeted continual learning: What does matter? In CVPR, 2023. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 141, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 141, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 320, + 141, + 553, + 195 + ], + "type": "text", + "content": "[56] Ameya Prabhu, Hasan Abed Al Kader Hammoud, Ser-Nam Lim, Bernard Ghanem, Philip HS Torr, and Adel Bibi. From categories to classifier: Name-only continual learning by exploring the web. arXiv preprint arXiv:2311.11293, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 198, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 198, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 320, + 198, + 553, + 252 + ], + "type": "text", + "content": "[57] Sarah Pratt, Ian Covert, Rosanne Liu, and Ali Farhadi. What does a platypus look like? generating customized prompts for zero-shot image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 15691-15701, 2023. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 254, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 254, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 320, + 254, + 553, + 319 + ], + "type": "text", + "content": "[58] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 321, + 553, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 321, + 553, + 385 + ], + "spans": [ + { + "bbox": [ + 320, + 321, + 553, + 385 + ], + "type": "text", + "content": "[59] Alexandre Rame, Kartik Ahuja, Jianyu Zhang, Matthieu Cord, Leon Bottou, and David Lopez-Paz. Model rata-touille: Recycling diverse models for out-of-distribution generalization. In Proceedings of the 40th International Conference on Machine Learning, pages 28656-28679. PMLR, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 388, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 388, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 320, + 388, + 553, + 453 + ], + "type": "text", + "content": "[60] Alexandre Ramé, Johan Ferret, Nino Vieillard, Robert Dadashi, Léonard Hussenot, Pierre-Louis Cedoz, Pier Giuseppe Sessa, Sertan Girgin, Arthur Douillard, and Olivier Bachem. Warp: On the benefits of weight averaged rewarded policies. arXiv preprint arXiv:2406.16768, 2024. 2, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 455, + 553, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 455, + 553, + 518 + ], + "spans": [ + { + "bbox": [ + 320, + 455, + 553, + 518 + ], + "type": "text", + "content": "[61] Alexandre Ramé, Matthieu Kirchmeyer, Thibaud Rahier, Alain Rakotomamonjy, Patrick Gallinari, and Matthieu Cord. Diverse weight averaging for out-of-distribution generalization. In Proceedings of the 36th International Conference on Neural Information Processing Systems, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 521, + 553, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 521, + 553, + 565 + ], + "spans": [ + { + "bbox": [ + 320, + 521, + 553, + 565 + ], + "type": "text", + "content": "[62] Alexandre Rame, Nino Vieillard, Léonard Hussenot, Robert Dadashi, Geoffrey Cideron, Olivier Bachem, and Johan Ferret. Warm: On the benefits of weight averaged reward models. arXiv preprint arXiv:2401.12187, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 567, + 553, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 567, + 553, + 589 + ], + "spans": [ + { + "bbox": [ + 320, + 567, + 553, + 589 + ], + "type": "text", + "content": "[63] Regent's, ParkLondon, Ukj, and . Utans. Weight averaging for neural networks and local resampling. 1996. 1, 2, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 590, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 590, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 320, + 590, + 553, + 633 + ], + "type": "text", + "content": "[64] Mark Rofin, Nikita Balagansky, and Daniil Gavrilov. Linear interpolation in parameter space is good enough for finetuned language models. arXiv preprint arXiv:2211.12092, 2022. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 635, + 553, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 635, + 553, + 700 + ], + "spans": [ + { + "bbox": [ + 320, + 635, + 553, + 700 + ], + "type": "text", + "content": "[65] Karsten Roth, Jae Myung Kim, A. Sophia Koepke, Oriol Vinyals, Cordelia Schmid, and Zeynep Akata. Waffling around for performance: Visual classification with random words and broad concepts. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 15746-15757, 2023. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 320, + 703, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 703, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 320, + 703, + 553, + 713 + ], + "type": "text", + "content": "[66] Karsten Roth, Zeynep Akata, Dima Damen, Ivana" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20489" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 81, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 81, + 72, + 294, + 95 + ], + "type": "text", + "content": "Balažević, and Olivier J Henaff. Context-aware multimodal pretraining. arXiv preprint arXiv:2411.15099, 2024. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 96, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 96, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 61, + 96, + 295, + 162 + ], + "type": "text", + "content": "[67] Karsten Roth, Lukas Thede, A. Sophia Koepke, Oriol Vinyals, Olivier J Henaff, and Zeynep Akata. Fantastic gains and where to find them: On the existence and prospect of general knowledge transfer between any pretrained model. In The Twelfth International Conference on Learning Representations, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 163, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 163, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 61, + 163, + 294, + 228 + ], + "type": "text", + "content": "[68] Karsten Roth, Vishaal Udandarao, Sebastian Dziadzio, Ameya Prabhu, Mehdi Cherti, Oriol Vinyals, Olivier Henaff, Samuel Albanie, Matthias Bethge, and Zeynep Akata. A practitioner's guide to continual multimodal pretraining. arXiv preprint arXiv:2408.14471, 2024. 1, 2, 3, 4, 5, 7, 8, 14" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 230, + 295, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 230, + 295, + 274 + ], + "spans": [ + { + "bbox": [ + 62, + 230, + 295, + 274 + ], + "type": "text", + "content": "[69] Sunny Sanyal, Atula Tejaswi Neerkaje, Jean Kaddour, Abhishek Kumar, and sujay sanghavi. Early weight averaging meets high learning rates for LLM pre-training. In First Conference on Language Modeling, 2024. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 275, + 295, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 275, + 295, + 341 + ], + "spans": [ + { + "bbox": [ + 62, + 275, + 295, + 341 + ], + "type": "text", + "content": "[70] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278–25294, 2022. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 342, + 295, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 342, + 295, + 385 + ], + "spans": [ + { + "bbox": [ + 62, + 342, + 295, + 385 + ], + "type": "text", + "content": "[71] Ekansh Sharma, Daniel M Roy, and Gintare Karolina Dziuagaite. The non-local model merging problem: Permutation symmetries and variance collapse. arXiv preprint arXiv:2410.12766, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 387, + 295, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 387, + 295, + 441 + ], + "spans": [ + { + "bbox": [ + 62, + 387, + 295, + 441 + ], + "type": "text", + "content": "[72] Li Shen, Anke Tang, Enneng Yang, Guibing Guo, Yong Luo, Lefei Zhang, Xiaochun Cao, Bo Du, and Dacheng Tao. Efficient and effective weight-ensembling mixture of experts for multi-task model merging. arXiv preprint arXiv:2410.21804, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 443, + 295, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 443, + 295, + 476 + ], + "spans": [ + { + "bbox": [ + 62, + 443, + 295, + 476 + ], + "type": "text", + "content": "[73] Ken Shoemake. Animating rotation with quaternion curves. Proceedings of the 12th annual conference on Computer graphics and interactive techniques, 1985. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 478, + 295, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 478, + 295, + 520 + ], + "spans": [ + { + "bbox": [ + 62, + 478, + 295, + 520 + ], + "type": "text", + "content": "[74] Shikhar Srivastava, Md Yousuf Harun, Robik Shrestha, and Christopher Kanan. Improving multimodal large language models using continual learning. arXiv preprint arXiv:2410.19925, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 522, + 295, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 522, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 62, + 522, + 295, + 555 + ], + "type": "text", + "content": "[75] George Stoica, Pratik Ramesh, Boglarka Ecsedi, Leshem Choshen, and Judy Hoffman. Model merging with svd to tie the knots. arXiv preprint arXiv:2410.19735, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 557, + 295, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 557, + 295, + 600 + ], + "spans": [ + { + "bbox": [ + 62, + 557, + 295, + 600 + ], + "type": "text", + "content": "[76] Zafir Stojanovski, Karsten Roth, and Zeynep Akata. Momentum-based weight interpolation of strong zero-shot models for continual learning. arXiv preprint arXiv:2211.03186, 2022. 2, 3, 4, 5, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 601, + 295, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 601, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 62, + 601, + 295, + 634 + ], + "type": "text", + "content": "[77] Yi-Lin Sung, Linjie Li, Kevin Lin, Zhe Gan, Mohit Bansal, and Lijuan Wang. An empirical study of multimodal model merging. arXiv preprint arXiv:2304.14933, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 635, + 295, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 635, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 62, + 635, + 295, + 667 + ], + "type": "text", + "content": "[78] Derek Tam, Mohit Bansal, and Colin Raffel. Merging by matching models in task parameter subspaces. Transactions on Machine Learning Research, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 670, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 670, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 670, + 295, + 712 + ], + "type": "text", + "content": "[79] Derek Tam, Yash Kant, Brian Lester, Igor Gilitschenski, and Colin Raffel. Realistic evaluation of model merging for compositional generalization. arXiv preprint arXiv:2409.18314, 2024. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 321, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 321, + 73, + 553, + 116 + ], + "type": "text", + "content": "[80] Lukas Thede, Karsten Roth, Olivier J Henaff, Matthias Bethge, and Zeynep Akata. Reflecting on the state of rehearsal-free continual learning with pretrained models. arXiv preprint arXiv:2406.09384, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 119, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 119, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 320, + 119, + 553, + 173 + ], + "type": "text", + "content": "[81] Vishaal Udandarao, Ankush Gupta, and Samuel Albanie. Sus-x: Training-free name-only transfer of vision-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2725-2736, 2023. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 175, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 175, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 320, + 175, + 553, + 239 + ], + "type": "text", + "content": "[82] Vishaal Udandarao, Nikhil Parthasarathy, Muhammad Ferjad Naeem, Talfan Evans, Samuel Albanie, Federico Tombari, Yongqin Xian, Alessio Tonioni, and Olivier J Henaff. Active data curation effectively distills large-scale multimodal models. arXiv preprint arXiv:2411.18674, 2024." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 243, + 553, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 243, + 553, + 306 + ], + "spans": [ + { + "bbox": [ + 320, + 243, + 553, + 306 + ], + "type": "text", + "content": "[83] Vishaal Udandarao, Ameya Prabhu, Adhiraj Ghosh, Yash Sharma, Philip HS Torr, Adel Bibi, Samuel Albanie, and Matthias Bethge. No “zero-shot” without exponential data: Pretraining concept frequency determines multimodal model performance. arXiv preprint arXiv:2404.04125, 2024. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 310, + 553, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 310, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 320, + 310, + 553, + 364 + ], + "type": "text", + "content": "[84] Ke Wang, Nikolaos Dimitriadis, Alessandro Favero, Guillermo Ortiz-Jimenez, Francois Fleuret, and Pascal Frossard. Lines: Post-training layer scaling prevents forgetting and enhances model merging. arXiv preprint arXiv:2410.17146, 2024. 4, 18" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 366, + 553, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 366, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 320, + 366, + 553, + 409 + ], + "type": "text", + "content": "[85] Ke Wang, Nikolaos Dimitriadis, Guillermo Ortiz-Jimenez, François Fleuret, and Pascal Frossard. Localizing task information for improved model merging and compression. arXiv preprint arXiv:2405.07813, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 411, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 411, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 320, + 411, + 553, + 498 + ], + "type": "text", + "content": "[86] Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, and Ludwig Schmidt. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In Proceedings of the 39th International Conference on Machine Learning, pages 23965-23998. PMLR, 2022. 1, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 501, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 501, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 320, + 501, + 553, + 578 + ], + "type": "text", + "content": "[87] Mitchell Wortsman, Gabriel Ilharco, Jong Wook Kim, Mike Li, Simon Kornblith, Rebecca Roelofs, Raphael Gontijo Lopes, Hannaneh Hajishirzi, Ali Farhadi, Hongseok Namkoong, and Ludwig Schmidt. Robust fine-tuning of zero-shot models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7959-7971, 2022. 1, 3, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 579, + 553, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 579, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 320, + 579, + 553, + 622 + ], + "type": "text", + "content": "[88] Feng Xiong, Runxi Cheng, Wang Chen, Zhanqiu Zhang, Yiwen Guo, Chun Yuan, and Ruifeng Xu. Multi-task model merging via adaptive weight disentanglement. arXiv preprint arXiv:2411.18729, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 624, + 553, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 624, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 320, + 624, + 553, + 677 + ], + "type": "text", + "content": "[89] Prateek Yadav, Derek Tam, Leshem Choshen, Colin Raffel, and Mohit Bansal. TIES-merging: Resolving interference when merging models. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 1, 2, 3, 4, 8, 14" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 320, + 681, + 553, + 713 + ], + "type": "text", + "content": "[90] Prateek Yadav, Colin Raffel, Mohammed Muqeeth, Lucas Caccia, Haokun Liu, Tianlong Chen, Mohit Bansal, Leshem Choshen, and Alessandro Sordoni. A survey on" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20490" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 73, + 295, + 586 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 80, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 80, + 73, + 294, + 106 + ], + "type": "text", + "content": "model moering: Recycling and routing among specialized experts for collaborative learning. arXiv preprint arXiv:2408.07057, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 107, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 107, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 61, + 107, + 295, + 150 + ], + "type": "text", + "content": "[91] Prateek Yadav, Tu Vu, Jonathan Lai, Alexandra Chronopoulou, Manaal Faruqui, Mohit Bansal, and Tsendsuren Munkhdalai. What matters for model merging at scale? arXiv preprint arXiv:2410.03617, 2024. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 152, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 152, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 61, + 152, + 294, + 205 + ], + "type": "text", + "content": "[92] Enneng Yang, Li Shen, Guibing Guo, Xingwei Wang, Xiaochun Cao, Jie Zhang, and Dacheng Tao. Model merging in llms, mllms, and beyond: Methods, theories, applications and opportunities. arXiv preprint arXiv:2408.07666, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 208, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 208, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 61, + 208, + 294, + 251 + ], + "type": "text", + "content": "[93] Enneng Yang, Li Shen, Zhenyi Wang, Guibing Guo, Xiaojun Chen, Xingwei Wang, and Dacheng Tao. Representation surgery for multi-task model merging. arXiv preprint arXiv:2402.02705, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 253, + 295, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 253, + 295, + 305 + ], + "spans": [ + { + "bbox": [ + 62, + 253, + 295, + 305 + ], + "type": "text", + "content": "[94] Cagatay Yildiz, Nishaanth Kanna Ravichandran, Prishruit Punia, Matthias Bethge, and Beyza Ermis. Investigating continual pretraining in large language models: Insights and implications. arXiv preprint arXiv:2402.17400, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 308, + 294, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 308, + 294, + 363 + ], + "spans": [ + { + "bbox": [ + 62, + 308, + 294, + 363 + ], + "type": "text", + "content": "[95] Le Yu, Bowen Yu, Haiyang Yu, Fei Huang, and Yongbin Li. Language models are super mario: Absorbing abilities from homologous models as a free lunch. In Proceedings of the 41st International Conference on Machine Learning, pages 57755-57775. PMLR, 2024. 2, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 365, + 294, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 365, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 62, + 365, + 294, + 407 + ], + "type": "text", + "content": "[96] Kerem Zaman, Leshem Choshen, and Shashank Srivastava. Fuse to forget: Bias reduction and selective memorization through model fusion. arXiv preprint arXiv:2311.07682, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 409, + 294, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 409, + 294, + 440 + ], + "spans": [ + { + "bbox": [ + 62, + 409, + 294, + 440 + ], + "type": "text", + "content": "[97] Friedemann Zenke, Ben Poole, and Surya Ganguli. Continuual learning through synaptic intelligence. In ICML, 2017. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 443, + 294, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 443, + 294, + 496 + ], + "spans": [ + { + "bbox": [ + 62, + 443, + 294, + 496 + ], + "type": "text", + "content": "[98] Renrui Zhang, Rongyao Fang, Peng Gao, Wei Zhang, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free clip-adapter for better vision-language modeling. arXiv preprint arXiv:2111.03930, 2021. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 499, + 294, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 499, + 294, + 531 + ], + "spans": [ + { + "bbox": [ + 62, + 499, + 294, + 531 + ], + "type": "text", + "content": "[99] Shenghe Zheng and Hongzhi Wang. Free-merging: Fourier transform for model merging with lightweight experts. arXiv preprint arXiv:2411.16815, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 533, + 294, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 533, + 294, + 586 + ], + "spans": [ + { + "bbox": [ + 57, + 533, + 294, + 586 + ], + "type": "text", + "content": "[100] Luca Zhou, Daniele Solombrino, Donato Crisostomi, Maria Sofia Bucarelli, Fabrizio Silvestri, and Emanuele Rodolà. Atm: Improving model merging by alternating tuning and merging. arXiv preprint arXiv:2411.03055, 2024. 1, 2" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "20491" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/HuMoCon_ Concept Discovery for Human Motion Understanding/b5dd425e-4f3e-4b2d-9c88-36e4f0bafd85_content_list.json b/2025/HuMoCon_ Concept Discovery for Human Motion Understanding/b5dd425e-4f3e-4b2d-9c88-36e4f0bafd85_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..46ce5f465694a46456237a1b7ffd5b6b4feff651 --- /dev/null +++ b/2025/HuMoCon_ Concept Discovery for Human Motion Understanding/b5dd425e-4f3e-4b2d-9c88-36e4f0bafd85_content_list.json @@ -0,0 +1,1710 @@ +[ + { + "type": "text", + "text": "HuMoCon: Concept Discovery for Human Motion Understanding", + "text_level": 1, + "bbox": [ + 165, + 130, + 833, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qihang Fang $^{1,2}$", + "bbox": [ + 150, + 179, + 274, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chengcheng Tang", + "bbox": [ + 295, + 180, + 449, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1The University of Hong Kong", + "bbox": [ + 339, + 198, + 584, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shugao Ma2", + "bbox": [ + 598, + 180, + 699, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yanchao Yang", + "bbox": [ + 722, + 180, + 841, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Meta", + "bbox": [ + 604, + 198, + 656, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "qihfang@gmail.com, {chengcheng.tang,bugratekin,shugao}@meta.com, yanchaoy@hku.hk", + "bbox": [ + 142, + 218, + 846, + 233 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/897542e4c766ca6d2895097d7377ad1855886fce72dc2242c7b2d4bd4dd03f85.jpg", + "image_caption": [ + "Figure 1. We present HuMoCon for concept discovery and human motion understanding with video or motion input. To address the challenge of feature misalignment and high-frequency information loss, we propose a novel feature alignment strategy and an advanced masked auto-encoder reconstructing velocity. HuMoCon empowers effective motion concept discovery and accurate Question Answering tasks, significantly outperforming state-of-the-art methods qualitatively and quantitatively." + ], + "image_footnote": [], + "bbox": [ + 106, + 250, + 558, + 396 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/97026424ca8d3a63ed694ff5fda914904f300cc2f49c155f651d897df1ae9167.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 575, + 251, + 903, + 393 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 481, + 326, + 498 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present HuMoCon, a novel motion-video understanding framework designed for advanced human behavior analysis. The core of our method is a human motion concept discovery framework that efficiently trains multi-modal encoders to extract semantically meaningful and generalizable features. HuMoCon addresses key challenges in motion concept discovery for understanding and reasoning, including the lack of explicit multi-modality feature alignment and the loss of high-frequency information in masked autoencoding frameworks. Our approach integrates a feature alignment strategy that leverages video for contextual understanding and motion for fine-grained interaction modeling, further with a velocity reconstruction mechanism to enhance high-frequency feature expression and mitigate temporal over-smoothing. Comprehensive experiments on standard benchmarks demonstrate that HuMoCon enables effective motion concept discovery and significantly outperforms state-of-the-art methods in training large models for human motion understanding. We will open-source the associated code with our paper.", + "bbox": [ + 88, + 513, + 485, + 816 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 844, + 220, + 859 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Understanding human behavior is crucial for gaining deeper insights into human actions and reasoning, which are foun", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ditional for developing human-centered AI systems [5, 26, 37, 61] that effectively address real-world tasks [14, 68, 72]. Recent advancements in large language models (LLMs) have led to significant progress in video- and motion-based methods [5, 15, 37, 40, 75] for analyzing human behavior. Despite these advancements, achieving a precise and fine-grained understanding of human actions remains a persistent and complex challenge.", + "bbox": [ + 511, + 482, + 906, + 604 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Human behavior is often represented as motion sequences structured based on the kinematics of the human skeleton [32?]. Traditional approaches to analyzing motion sequences have predominantly focused on classification [11, 24, 34, 42, 44, 45, 58, 76], relying on predefined action categories. While effective for specific tasks, these methods often lack the flexibility for deeper or more nuanced analysis. Other methods employ statistical models to assess motion quality [3, 6, 18, 43, 51], but they require significant expert knowledge, limiting scalability.", + "bbox": [ + 511, + 609, + 908, + 760 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advancements in language models [40, 50, 75] have opened new possibilities in motion analysis, enabling tasks such as motion generation and motion captioning based on free-text inputs [4, 20, 52, 59, 61], advancing the field of motion analysis. However, obtaining high-quality motion sequences and annotations remains costly, and motion sequences generally lack information about the surrounding environment, limiting models' ability to understand human-environment interactions.", + "bbox": [ + 511, + 763, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7179", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In contrast, video data offers rich visual context and is easier to collect. It has been widely used for action classification [16, 62, 65] and video Q&A [5, 37, 40]. However, video data often contains irrelevant elements that can hinder accurate content analysis. For example, a model might learn from training data that most running occurs in stadiums, leading it to mistakenly classify other stadium activities as running during testing.", + "bbox": [ + 89, + 90, + 480, + 212 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In recent years, significant progress has been made in developing multi-modality models [5, 87]. Integrating multiple modalities, as opposed to relying on a single modality, enhances the network's ability to learn fine-grained and semantic features. For instance, LART [56] combines video and motion data to ensure that video-based action classification focuses on the human body rather than overemphasizing environmental context. MotionLLM [5] expands large language models (LLMs) to video and motion, building a model that can understand both. MotionLLM demonstrated that learning from video and motion improves the network's ability to analyze human behavior. However, MotionLLM implicitly aligns video and motion using motion-video pairs during LLM training without optimizing the encoder for explicit cross-modal alignment. To address this limitation and further enhance feature representation, we propose explicitly aligning features from different modalities during the encoding process within a human motion concept discovery framework.", + "bbox": [ + 91, + 215, + 486, + 501 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Achieving effective cross-modal alignment and feature representation relies heavily on robust encoder training methods. Many existing studies [70, 71, 87] adopt a masked autoencoder framework [27] when training encoders. This approach involves applying a mask to a certain proportion of the input, extracting features from the masked data using the encoder, and reconstructing the original unmasked input with a decoder. While effective in learning high-quality features, this masking strategy inherently causes the loss of high-frequency information, resulting in temporally over-smoothed reconstructions that hinder the LLM's ability to understand human behavior. [47] To address this issue, we propose reconstructing velocity as a means to preserve high-frequency information. Following InfoCon [46], we design a VQ-VAE-based loss function that helps retain the fine-grained features of the original masked autoencoder while further enhancing high-frequency feature expression. This approach significantly boosts the subsequent LLM's capacity to model and understand human motion effectively.", + "bbox": [ + 89, + 505, + 482, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, we present HuMoCon, an effective human motion concept discovery pertaining framework for advanced human behavior understanding from both video and motion data. Our key contributions include: (i) a novel feature alignment strategy that explicitly leverages video data for contextual understanding and motion data for fine-grained interaction modeling, and (ii) an enhanced masked", + "bbox": [ + 89, + 795, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "autoencoder framework incorporating velocity reconstruction to significantly improve high-frequency feature expression and mitigate temporal over-smoothing. To evaluate the effectiveness of our proposed method, we conduct experiments on the Activity-QA [78] and BABEL-QA [13] benchmarks [54], comparing our approach with the state-of-the-art methods. Comprehensive quantitative and qualitative evaluations demonstrate that HuMoCon achieves state-of-the-art performance on standard benchmarks for motion-video understanding.", + "bbox": [ + 511, + 90, + 903, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 260, + 653, + 275 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Human Motion Understanding", + "text_level": 1, + "bbox": [ + 511, + 286, + 785, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The core of human motion understanding is to learn the semantics of motion sequences or videos. In the early stages, research in this field is often based on predefined action categories. Such work can be divided into two types: action classification [11, 24, 34, 42, 44, 45, 58, 76] and motion generation targeted at specific categories [4, 20, 28, 31, 36, 52, 59-61, 74, 77]. For action recognition, several works [11, 44] utilize Recurrent Neural Networks (RNN) to extract temporal information from human joint sequences, beyond the CNN-based ones [34, 45]. Subsequent works [24, 42, 58, 76] propose representing motion data as a graph structure, enabling the use of spatio-temporal graph convolutional networks (ST-GCNs) to effectively capture the connectivity and interactions between human joints. For motion generation, several studies [28, 31, 36, 60, 74, 77] have concentrated on generating specific, single actions. In contrast, other works [4, 20, 52, 59, 61] aim to generate a broader range of diverse action types, addressing the challenge of variability in motion synthesis.", + "bbox": [ + 511, + 309, + 903, + 597 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In recent years, more research has shifted towards captioning [23, 30], generation [1, 22, 23, 30, 35, 53, 64, 81, 82, 86], and analysis [5, 9, 15, 17] based on free-form text. Leveraging alignment of motion latent features with the CLIP [55] embedding space, MotionCLIP [63] enables the generation of out-of-distribution motions. Transformers [1, 22, 23, 53, 81] and diffusion [35, 64, 82] models are also utilized for text-conditioned motion generation. Moreover, MotionGPT [30] and AvatarGPT [86] utilize large language models (LLMs) to handle multiple motion-related tasks simultaneously. Despite these advancements, research focused specifically on motion analysis remains relatively limited. PoseFix [9] and CigTime [15] are proposed to annotate the corrective instructions from source to target for pose and motion sequences. MotionLLM [5] proposes a method to analyze both the video and motion sequences with detailed spatial-temporal awareness and reasoning abilities. However, MotionLLM relies on implicit alignment through paired data and lacks explicit cross-modal feature alignment.", + "bbox": [ + 511, + 599, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7180", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Video Understanding", + "text_level": 1, + "bbox": [ + 89, + 90, + 292, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In addition to motion understanding, video understanding has emerged as a prominent research area, attracting extensive attention in recent years. Early works on video classification and segmentation relied on handcrafted features [10, 44, 66, 67] or convolutional neural networks [16, 62, 65]. Beyond classification, video analysis [3, 6, 18, 25, 26, 37, 40, 43, 50, 51, 57, 73, 75, 85] has become a highly active field of research. Action quality assessment has been the focus of several works [3, 6, 18, 43, 51], with methods such as Aifit [18] and PoseTrainer [6] proposed to correct user actions. These approaches evaluate actions in conjunction with complex textual instructions [3, 43, 51]. Other studies [25, 26, 57, 73, 85] aim to generate descriptive narratives for videos, broadening the scope of video understanding. Recently, many works [2, 37, 40, 50, 75, 80] have leveraged large language models (LLMs) to understand video content and perform complex reasoning tasks. These methods combine video processing with the advanced reasoning capabilities of LLMs, further advancing the state-of-the-art in video analysis. Video-LLaVA [40] build a model to perform visual reasoning capabilities on images and videos simultaneously. Follow-up works [37, 50, 75] extend Video-LLaVA for faster inference and more accurate reasoning ability. Unlike typical video understanding algorithms, which rely heavily on visual context alone, our approach integrates motion data directly, enhancing the model's ability to understand human-centric actions.", + "bbox": [ + 89, + 112, + 485, + 521 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Multi-modality Pre-training", + "text_level": 1, + "bbox": [ + 89, + 530, + 344, + 546 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multi-modality pre-training has primarily focused on aligning vision and language representations. CLIP [55] pioneered this effort by using contrastive learning to align image and text features on large-scale datasets. Subsequent works [12, 48, 63, 83] extended CLIP's framework to other modalities, including video [48], audio [12] and motion [63].", + "bbox": [ + 89, + 551, + 482, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent studies, such as VALOR [7] and VAST [8], demonstrated that cross-modal alignment not only enhances model robustness but also maintains high performance across modalities. Building on this, follow-up works [19, 87] have aimed to expand the number of aligned modalities, further advancing multi-modality research.", + "bbox": [ + 89, + 657, + 482, + 748 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the context of motion and video alignment, Motion-LLMs [5] is the first work to address this challenge. However, it relies solely on implicit alignment during the fine-tuning stage of large language models.", + "bbox": [ + 89, + 750, + 482, + 810 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 823, + 181, + 838 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 89, + 848, + 200, + 862 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our objective is to develop an LLM-based system capable of understanding and reasoning both human video and mo", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tion inputs. Given a question $Q$ and an input, which can be either a video $U$ or motion sequence $M$ , the model generates the corresponding answer $W$ . Specifically, a video $U$ is represented as a sequence of $T$ keyframes, denoted as $\\{u_0, u_1, \\dots, u_{T-1}\\}$ . Similarly, a motion sequence $M$ is represented as $K$ consecutive pose frames, denoted as $\\{m_0, m_1, \\dots, m_{K-1}\\}$ .", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We present an overview of the pipeline in Fig. 2. To enable effective multi-modal human motion understanding, our framework focuses on: (1) discovering human motion concepts from both video and motion data, (2) preserving high-frequency motion details using velocity reconstruction to reduce temporal smoothing, and (3) aligning video and motion features to enhance multi-modal comprehension. To accomplish this, we train our video and motion encoders with four key learning objectives: Masked Reconstruction, Discriminative Informativeness, Actionable Informativeness, and Feature Alignment. Consequently, we leverage the encoded features of videos and motions to facilitate the fine-tuning of LLMs for video-motion understanding and reasoning. The following sections detail each component of our approach, including multi-modality encoder pre-training, modality translation, and instruction tuning.", + "bbox": [ + 511, + 198, + 906, + 439 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Multi-Modality Encoder Pre-training via Human Motion Concept Discovery", + "text_level": 1, + "bbox": [ + 511, + 448, + 864, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Pre-training encoders are critical for multi-modality large language models (LLMs) [70, 71, 87], as they enable the effective extraction of essential features from large-scale input data. To enhance the encoder's ability to capture key information, we design three distinct learning objectives: masked reconstruction, velocity reconstruction, and feature alignment, all contributing to the quality of the discovered human motion concepts as well as the pre-training efficiency. For both videos and motions, we adopt a similar network structure during encoder pre-training. Consequently, in the following section, we focus on detailing the network structure for video inputs.", + "bbox": [ + 511, + 484, + 906, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Encoding Function. In our approach, we design masked autoencoder architectures tailored for both video and motion data. Specifically, for videos, we utilize a pre-trained encoder $\\mathcal{E}_u$ , to encode the input $U$ into a feature representation $f^u$ :", + "bbox": [ + 511, + 686, + 905, + 761 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nF ^ {u} = \\mathcal {E} _ {u} (U). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 661, + 762, + 903, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Then, we employ a codebook $C^u$ , modeled with VQ-VAE to quantize $f^u$ into a discrete feature representation $d^u$ :", + "bbox": [ + 511, + 785, + 905, + 829 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nD ^ {u} = \\operatorname {Q u a n t i z e} \\left(F ^ {u}, C ^ {u}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 830, + 903, + 848 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where 'Quantize' denotes the quantization operation, and $C^u$ is the codebook used for discretization, also serving as the discovered human motion concepts.", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7181", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e40073539a848eda21d5e8fb455fc0e97d59db9d6253a2473b5c1d203ac79cb9.jpg", + "image_caption": [ + "Figure 2. System overview of our method. (a) The encoder pre-training process for learning and aligning video and motion features and enhancing high-frequency details through velocity reconstruction. We utilize a VQ-VAE-based structure, and we design effective learning objectives to enhance the encoder to extract semantic meaningful and fine-grained features. (b) The fine-tuning of the large language model (LLM) for video and motion reasoning consists of two stages: Modality Translation and Multi-modality Instruction Tuning. In the Modality Translation stage, we train a translation layer for each modality to map the encoding feature to the LLM space. In the Instruction Tuning stage, we fine-tune the LLM to understand human motion and videos for downstream tasks." + ], + "image_footnote": [], + "bbox": [ + 143, + 88, + 854, + 353 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Masked Reconstruction. The masked autoencoder has proven to be highly effective for pre-training [27], as it facilitates learning of meaningful and low-frequency features from input data. Thus, we leverage masked auto-encoding to enhance the discovered concepts and facilitate the multimodal feature pre-training.", + "bbox": [ + 88, + 460, + 482, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "More explicitly, we mask $D^u$ by a probability of $\\sigma$ to obtain the masked feature $D_{\\mathrm{mask}}^u$ :", + "bbox": [ + 89, + 551, + 482, + 585 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nD _ {\\text {m a s k}} ^ {u} = \\operatorname {M a s k} \\left(D ^ {u}, \\sigma\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 595, + 480, + 612 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where 'Mask' is the operation that applies masking to the features, and $\\sigma < 1.0$ controls the masking ratio. A video decoder $D^{u}$ is used to reconstruct the video $U^{\\prime}$ from $\\mathcal{D}_{\\mathrm{mask}}^{u}$ :", + "bbox": [ + 89, + 625, + 483, + 672 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nU ^ {\\prime} = \\mathcal {D} ^ {u} \\left(D _ {\\text {m a s k}} ^ {u}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 225, + 685, + 480, + 702 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We compute the reconstruction loss between the original video $U$ and the reconstructed video $U^{\\prime}$ as follows:", + "bbox": [ + 89, + 710, + 482, + 739 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {v i d e o}} ^ {\\text {r e c}} = \\left| \\left| U ^ {\\prime} - U \\right| \\right| _ {2} ^ {2}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 214, + 752, + 480, + 770 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Similarly, we introduce a reconstruction loss for motion data, $\\mathcal{L}_{\\mathrm{motion}}^{\\mathrm{rec}}$ , following the same process. The total reconstruction loss is defined as:", + "bbox": [ + 89, + 782, + 482, + 827 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {\\text {r e c}} = \\mathcal {L} _ {\\text {v i d e o}} ^ {\\text {r e c}} + \\mathcal {L} _ {\\text {m o t i o n}} ^ {\\text {r e c}}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 205, + 840, + 482, + 858 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The above masked reconstruction loss ensures that the encoded features contain sufficient information about the", + "bbox": [ + 89, + 869, + 483, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "video or motion input, while learning away possible nuisances through masking and making the learning of the concepts (codebook) more focused on critical information. On the other hand, the discretization imposed by the concepts further enhances the meaningfulness and robustness of the pre-trained multi-modal (motion) features.", + "bbox": [ + 511, + 460, + 906, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Velocity Reconstruction. Reconstructing inputs from masked features inevitably results in the loss of high-frequency details, leading to temporally over-smoothed reconstruction outcomes and preventing the discovery of more nuanced yet useful motion concepts.", + "bbox": [ + 511, + 580, + 905, + 655 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address this limitation, we propose an objective that further reconstructs the \"velocity\" of the input sequence. To be more specific, we define \"state\" as the encoding or feature of a frame in the video or motion sequence, while the \"velocity\" means the difference between consecutive \"states\" that captures high-frequency information by focusing on dynamic changes. For video and motion data, we denote the \"states\" as $U$ and $m$ , and the \"velocity\" as the optical flow $O$ (computed from image frames) and delta motion $\\delta M$ , respectively.", + "bbox": [ + 511, + 657, + 905, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, directly reconstructing velocity from the same set of masked features introduces difficulty in figuring out the dynamic information, given the above-mentioned oversmoothing effect in the encoded features. To mitigate this, we draw inspiration from the robotic manipulation literature [46] on action concept discovery and incorporate two", + "bbox": [ + 511, + 810, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7182", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b4033cb12be74bb6bf64101f5319199a9aa4a4861eb8fedd50f3c9f868542bd1.jpg", + "image_caption": [ + "Figure 3. Overview of the velocity reconstruction components. We build similar network structures for both video and motion, and present the video part in this figure. This module is composed of two learning objectives. 1) Discriminative informativeness (left) aims to improve the distinctiveness of encoded features by reducing representational ambiguity. 2) Actionable informativeness (right) focuses on reconstructing the velocity by leveraging gradient information from the discrimination hypernetwork. As for the video data, we employ optical flow as the representation of the velocity." + ], + "image_footnote": [], + "bbox": [ + 173, + 87, + 823, + 229 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tailored learning objectives - discriminative informativeness and actionable informativeness - to enhance the efficiency of representations and improve the quality of temporal dynamics in multi-modality pretaining, as shown in Fig. 3.", + "bbox": [ + 88, + 328, + 480, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Discriminative informativeness. The discriminative informativeness objective aims to improve the distinctiveness of encoded features by increasing the correlation between the discovered concepts and the corresponding state features. Specifically, we introduce a video hyper-network $\\mathcal{H}^u$ , which takes discrete features $c_k^u$ from the codebook $C^u$ as input and generates a classification network. The hyper-network $\\mathcal{H}^u$ (together with the generated classifier) is trained to determine whether an input state is compatible with its corresponding code class, which is represented by a confidence value $s_u$ (classification score) from 0 to 1, thereby enhancing the distinguishability and separation of each feature class in the representation space.", + "bbox": [ + 89, + 405, + 483, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The discriminative loss for video is defined as:", + "bbox": [ + 109, + 602, + 419, + 617 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {v i d e o}} ^ {\\text {d i s}} = \\sum_ {k = 0} ^ {| C ^ {u} | - 1} \\sum_ {i = 0} ^ {T - 1} C E \\left(\\mathcal {H} ^ {u} \\left(c _ {k} ^ {u}, u _ {i}\\right), d _ {i} ^ {u}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 633, + 482, + 678 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $|C^u|$ is the size of the codebook $C^u$ , $CE(\\cdot)$ is the cross-entropy loss for classification, $d_i^u$ represents the \"ground-truth label\" for the input state $u_i$ determined by the VQ-VAE, and $T$ is the number of frames.", + "bbox": [ + 89, + 686, + 482, + 747 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Actionable informativeness. The actionable informativeness objective focuses on reconstructing the velocity by leveraging gradient information from the discriminative hyper-network, $\\mathcal{H}^u$ introduced above. Inspired by the idea that gradients of a discriminator function can inform the state changes, as observed in [46], this objective ensures that the model captures not only static state information, but also the dynamic changes in the sequence. By leveraging gradient fluctuations around a given state, the approach effectively captures dynamic motion details, such as velocity,", + "bbox": [ + 88, + 750, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "resulting in more fine-grained motion concepts, and enriching the motion representation.", + "bbox": [ + 511, + 328, + 903, + 358 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, we initialize a network $\\mathcal{A}^u$ for video data, which takes key image frames $U$ and gradients from the discriminator $\\nabla_U\\mathcal{H}^u (D^u,U)$ as inputs and predicts the velocity $V_{u}^{\\prime}$ . The actionable informativeness loss for video is then computed as:", + "bbox": [ + 511, + 359, + 905, + 434 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {v i d e o}} ^ {\\text {a c t}} = \\left| \\left| V _ {u} - V _ {u} ^ {\\prime} \\right| \\right| _ {2} ^ {2}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 446, + 903, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $V_{u}$ represents the ground-truth velocity of video frames. This process is repeated for motion data to compute the corresponding loss $\\mathcal{L}_{\\mathrm{motion}}^{\\mathrm{act}}$ using $V_{m}$ and $V_{m}^{\\prime}$ .", + "bbox": [ + 511, + 477, + 903, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Combined Objectives. By integrating both video and motion components, we derive the total velocity reconstruction loss for human motion concept discovery and motion feature pre-training as in the following:", + "bbox": [ + 511, + 523, + 905, + 583 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {\\text {d i s}} = \\mathcal {L} _ {\\text {v i d e o}} ^ {\\text {d i s}} + \\mathcal {L} _ {\\text {m o t i o n}} ^ {\\text {d i s}} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 632, + 594, + 903, + 614 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {\\mathrm {a c t}} = \\mathcal {L} _ {\\text {v i d e o}} ^ {\\mathrm {a c t}} + \\mathcal {L} _ {\\text {m o t i o n}} ^ {\\mathrm {a c t}} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 642, + 903, + 660 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With the above objectives, our approach ensures that the model effectively reconstructs high-frequency details, improving the temporal fidelity and overall quality of multimodal representations.", + "bbox": [ + 511, + 667, + 905, + 728 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Feature Alignment. Multi-modality models often leverage the shared information across different modalities to enhance performance [5, 56, 87]. In the case of motion and video, video data provides contextual information that motion data alone may be lacking, which is critical for accurate motion prediction and understanding. Conversely, motion data focuses on human-centric features in the video, reducing over-reliance on environmental details. An efficient utilization or balancing of both is beneficial for the human motion LLMs. However, prior work on human motion", + "bbox": [ + 511, + 750, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7183", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and video understanding [5] only achieves implicit alignment through paired data during LLM fine-tuning, without explicit alignment in encoder training. To address this issue, we propose an alignment loss to improve information sharing and cross-modal coherence.", + "bbox": [ + 89, + 90, + 480, + 165 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To achieve alignment between video and motion features, we collect a set of video-motion pairs from Motion-X [41], which are aligned by frames. We instantiate two projection layers: The video projection layer $\\mathcal{P}^u$ and the motion projection layer $\\mathcal{P}^m$ . These layers are designed to map the respective features from the video and motion domains into a shared aligned space.", + "bbox": [ + 89, + 167, + 480, + 271 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "More explicitly, given paired video and motion data $\\{U,M\\}$ , we use $\\mathcal{P}^u$ and $\\mathcal{P}^m$ to project the discrete video feature $d_i^u$ and motion feature $d_j^m$ into the aligned feature space, resulting in $a_i^u$ and $a_j^m$ , where $a_i^u, a_j^m \\in \\mathbb{R}^H$ . Here, $H$ denotes the dimensionality of the aligned features in the shared space.", + "bbox": [ + 89, + 272, + 480, + 364 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To ensure effective alignment, we define an alignment loss as follows:", + "bbox": [ + 89, + 364, + 480, + 393 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {\\text {a l i g n}} = \\sum_ {0} ^ {T - 1} \\frac {\\operatorname {n o r m} \\left(a _ {i} ^ {u}\\right) \\cdot \\operatorname {n o r m} \\left(a _ {\\text {a l i g n} (i)} ^ {m}\\right) / \\epsilon}{\\sum_ {0} ^ {K - 1} \\operatorname {n o r m} \\left(a _ {i} ^ {u}\\right) \\cdot \\operatorname {n o r m} \\left(a _ {j} ^ {m}\\right) / \\epsilon}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 400, + 480, + 444 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathrm{align}(i)$ indicates the pose index in the motion sequence corresponding to the key video frame $u_{i}$ , norm represents a normalization operation to stabilize feature magnitudes and $\\epsilon$ is a temperature parameter that controls the sharpness of the alignment distribution.", + "bbox": [ + 89, + 450, + 480, + 525 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Note that $\\mathcal{L}^{\\mathrm{align}}$ is computed only when motion and video are paired. For standalone motion or video inputs, this loss term is omitted for applicability.", + "bbox": [ + 89, + 526, + 480, + 571 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The overall human motion concept discovery objective is defined as:", + "bbox": [ + 89, + 571, + 480, + 599 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} ^ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {m o t i o n}} ^ {\\text {r e c}} + \\mathcal {L} _ {\\text {v i d e o}} ^ {\\text {r e d e}} + \\\\ \\lambda^ {\\text {d i s}} \\mathcal {L} ^ {\\text {d i s}} + \\lambda^ {\\text {a c t}} \\mathcal {L} ^ {\\text {a c t}} + \\lambda^ {\\text {a l i g n}} \\mathcal {L} ^ {\\text {a l i g n}}, \\tag {12} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 609, + 480, + 648 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\lambda^{\\mathrm{dis}}$ , $\\lambda^{\\mathrm{act}}$ and $\\lambda^{\\mathrm{align}}$ are weighting coefficients for the respective losses $\\mathcal{L}^{\\mathrm{dis}}$ , $\\mathcal{L}^{\\mathrm{act}}$ and $\\mathcal{L}^{\\mathrm{align}}$ .", + "bbox": [ + 89, + 657, + 480, + 688 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3. Modality Translation", + "text_level": 1, + "bbox": [ + 89, + 696, + 289, + 713 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Although pre-trained encoders capture rich features from the original input, these features are often difficult for LLMs to interpret directly [5, 37, 40]. Therefore, it is essential to introduce an additional projection layer to map these features into a space that LLMs can effectively understand.", + "bbox": [ + 89, + 719, + 480, + 794 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As illustrated in Fig 2, we add tunable translation layers after the encoders $\\mathcal{E}^u$ and $\\mathcal{E}^m$ . These layers are trained using annotated video and motion sequence data to learn task-specific transformations. During this stage, the encoders and the LLM $\\mathcal{F}$ are kept frozen to prevent adverse feature drift, and only the weights of the translation layers are updated. This strategy preserves the pre-trained knowledge of", + "bbox": [ + 89, + 795, + 480, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the encoders and language model while optimizing the feature mapping for downstream tasks.", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Multi-Modality Instruction Tuning", + "text_level": 1, + "bbox": [ + 511, + 131, + 816, + 148 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finally, we employ instruction tuning to train the language model $\\mathcal{F}$ to respond to user queries using the corresponding video or motion sequences as context. During this process, the encoders and projection layers remain frozen, ensuring their pre-trained features are retained, while the weights of the language model $\\mathcal{F}$ are updated to accommodate the task.", + "bbox": [ + 511, + 154, + 903, + 243 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this stage, the model can further enhance cross-modal alignment using paired motion-video data in an implicit manner. To streamline the fine-tuning process, we adopt LoRA [29] with a rank of 8, allowing efficient and lightweight adjustment to the language model.", + "bbox": [ + 511, + 244, + 903, + 321 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 335, + 643, + 351 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experiments Settings", + "text_level": 1, + "bbox": [ + 511, + 359, + 712, + 376 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details For videos and motions, we initialize a codebook with 512 codes, a one-layer projector for feature alignment, and a 2-layer MLP to generate the weights for a 2-layer discriminator separately. For video inputs, we utilize pre-trained SigLIP [79] with an image patch size of 16 pixels and hidden dimension of 256 as the video encoder, and we take a 12-layer transformer as the reconstruction decoder. A 4-layer transformer is utilized as the motion encoder and a 2-layer transformer is used for motion reconstruction. For the velocity reconstruction, we initialize a two-layer hypernet and two-layer velocity decoder. Besides, we utilize one-layer MLPs for feature alignment. We pre-train the motion encoder with only the reconstruction loss for 60K iterations with a learning rate of 1e-4. Then, we train both the video and motion encoders with the total loss for 8K iterations with a learning rate of 1e-5.", + "bbox": [ + 511, + 382, + 903, + 623 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We utilize Llama3.1-8B as our base LLM model. For the modality translation stage, we initialize a one-layer projector for the video and motion encoder separately. We train the projectors on the annotation dataset for one batch with a batch size of 32 and a learning rate of 1e-4. In the Multimodality instruction tuning stage, we train the LLM model with a batch size of 16 and a learning rate of 3e-4.", + "bbox": [ + 511, + 625, + 903, + 729 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training datasets For the encoder pre-training stage, we utilize Motion-X [41] datasets which contain 81K motion sequences and 36K motion-video pairs. In order to facilitate the training process of LLMs, we follow MotionLLM [5] to prompt LLMs to generate 20K annotation for motion-video pairs from the Motion-X dataset, and 10K annotation data from the HumanML3D [21] dataset. This data is utilized for modality translation. We additionally utilize the VATEX [69] dataset which contains 20K annotated videos for video translator training. For multi-modality instruction", + "bbox": [ + 511, + 750, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7184", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/bf1f754f722ccd4b33f8a5870fb4a715d774381e66e38192bb623b6f01a00e33.jpg", + "table_caption": [ + "Table 1. Comparison of different methods on BABEL-QA test set. The pred type represent the prediction type, including classification and generation. For 2s-AGCN and MotionCLIP, 'M' means MLP, 'R' means RNN. Our method outperforms the baselines across all the metrics." + ], + "table_footnote": [], + "table_body": "
ModelPred typeOverallQuery typeRelation Type
ActionDirectionBodyPartBeforeAfterOther
2s-AGCN-Mcls.0.3550.3840.3520.2280.3310.2640.295
2s-AGCN-Rcls.0.3570.3960.3520.1940.3370.3010.285
MotionCLIP-Mcls.0.4300.4850.3610.2720.3720.3210.404
MotionCLIP-Rcls.0.4200.4890.3100.2500.3980.3140.387
MotionLLMgen.0.4360.5170.3540.1540.4270.3680.529
Oursgen.0.7110.8090.6970.6230.7070.6350.797
", + "bbox": [ + 156, + 141, + 836, + 268 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tuning, we follow MotionLLM to generate 200K Q&A pairs for both motions and videos. We also utilize the VideoChatGPT [49] and BABEL-QA [13] datasets to enhance the motion and video understanding ability of our model, which has 100K and 2K Q&A data separately.", + "bbox": [ + 89, + 279, + 480, + 357 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation datasets Although MotionLLM [5] creates a large benchmark for both motion and video understanding, by the time we completed our paper, they had not released the specific details or access methods for the benchmark, nor did they provide a clear timeline for its release. Therefore, we did not perform comparisons on that benchmark but instead selected the test set of ActivityNet-QA [78] and BABEL-QA [13] to separately validate the performance of our algorithm on video and motion data.", + "bbox": [ + 89, + 382, + 483, + 518 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation metrics On the BABEL-QA benchmark, we evaluate the prediction accuracy for evaluation. As for the classification-based baselines, we evaluate the accuracy directly. As for the generative baselines, we adapt the evaluation protocol used in previous works [5, 37] by utilizing LLMs. Specifically, the evaluation LLM is given some examples of questions, answers, and scores (from 0 to 1), and it is required to follow these examples and score the experiment results. On the Activity-QA benchmark, we only compare our methods with our generative baselines. Thus, we directly utilize LLMs to evaluate the methods.", + "bbox": [ + 89, + 545, + 483, + 712 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Quantitative Results", + "text_level": 1, + "bbox": [ + 89, + 726, + 284, + 742 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Experiments on ActivityNet-QA. We evaluate the video understanding ability of our method and other baselines on the ActivityNet-QA benchmark as shown in Tab. 2. We follow previous methods [5, 33, 38] to report the accuracy and scores that are evaluated by LLMs. We show in Tab. 2 that our method outperforms all the baselines on both metrics. Compared to the video-only method, our method aligns the video feature with the motion feature, making the encoder concentrate on human-centric information. Compared to MotionLLM, our encoder pre-training helps the network to", + "bbox": [ + 89, + 750, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "learn more fine-grained information through velocity reconstruction.", + "bbox": [ + 511, + 280, + 903, + 310 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Experiments on BABEL-QA. We evaluate the motion understanding ability of our method as compared to existing studies on the BABEL-QA [13] dataset. Following [5, 13], we compare our method with a diverse set of baselines: 1) 2s-AGCN [58], a method using 2s-GCN to extract motion features and classify motion by MLP or RNN; 2) Motion-CLIP [63], a method that aligns the features of motion and text using CLIP [55] and predicts motion class by MLP or RNN; 3) MotionLLM [5], an LLM-based method which has motion reasoning ability. We follow [5] to fine-tune our method with the training set of BABEL-QA and evaluate our method with LLMs.", + "bbox": [ + 511, + 329, + 906, + 511 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "From Tab. 1, our method outperforms other methods in all metrics by a large margin of nearly $60\\%$ overall. In the \"bodypart\" query task, our algorithm achieves 2.29 times the score of the second-place method, MotionCLIP-M (0.623 vs. 0.272). This advantage stems partly from our velocity-based encoder training, which, compared to traditional motion feature extraction, helps us better capture high-frequency information and ensures the learning of fine-grained features. Additionally, unlike MotionLLM, we perform alignment at the encoder pre-training stage, ensuring that the features learned by the encoder are inherently more semantically rich.", + "bbox": [ + 511, + 511, + 908, + 693 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8cf8d7e77932952657d5f70387a90d0e28a163086d1e0b69ca21499244939bd9.jpg", + "table_caption": [ + "Table 2. Experiments results on ActivityNet-QA benchmark." + ], + "table_footnote": [], + "table_body": "
ModelAcc↑Score ↑
FrozenBiLM [57]24.7-
VideoChat [38]-2.2
LLaMA-Adapter [84]34.22.7
Video-LLaMA [80]12.41.1
Video-ChatGPT [49]35.22.7
Video-LLaVA [40]45.33.3
VideoChat2 [39]49.13.3
MotionLLM [5]53.33.5
Ours54.23.6
", + "bbox": [ + 555, + 741, + 864, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7185", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/bd0306a1cfcc25e3dd319628632e417252bd03f2fc0c7a6c7dbae1a88fa52ed2.jpg", + "table_caption": [ + "Table 3. Ablation study on BABEL-QA test set." + ], + "table_footnote": [], + "table_body": "
ModelPred typeOverallQuery typeRelation Type
ActionDirectionBodyPartBeforeAfterOther
MotionLLMgen.0.4360.5170.3540.1540.4270.3680.529
Ours-w/oLrecgen.0.6960.7410.6450.5770.6000.5970.762
Ours-w/oLdis & Lactgen.0.6370.6930.4780.6060.6670.5260.709
Ours-w/oLaligngen.0.6750.7430.5790.5230.5840.5700.743
Oursgen.0.7110.8090.6970.6230.7070.6350.797
", + "bbox": [ + 140, + 113, + 854, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/592a68e5187fcb85cf5fb369af3df6751909641baebc47d0657f266ac6a8741c.jpg", + "image_caption": [ + "Figure 4. Example of motion understanding. We validate the proposed model's ability to understand motion from multiple aspects. Q1 on the left and Q1-3 on the right evaluate its comprehension of motion sequences, while Q2 and Q3 on the left push it to analyze kinematic, kinesthetic, and physical properties. These results demonstrate our algorithm's ability to recognize, describe in detail, and analyze motions." + ], + "image_footnote": [], + "bbox": [ + 140, + 227, + 854, + 460 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation study. We present the ablation results in Tab. 3. We conducted ablation studies on masked reconstruction, velocity reconstruction, and feature alignment. Motion-LLM serves as an ablation baseline without any of these techniques. As presented, the absence of any single module leads to a decrease in performance. Among these, masked reconstruction has the least impact, as we initially train the motion encoder solely with masked reconstruction loss to ensure stable subsequent training. Removing velocity reconstruction has the greatest impact on pexrformance since velocity provides high-frequency information that other modules lack. Likewise, removing the alignment loss significantly affects results, highlighting the importance of multi-modality alignment in motion understanding.", + "bbox": [ + 89, + 527, + 485, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Qualitative Results", + "text_level": 1, + "bbox": [ + 89, + 755, + 272, + 771 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To validate our algorithm's ability to understand motions and its contextual knowledge, we conducted multiple rounds of Q&A with the motion model. We present two examples in Fig. 4. We discover that, beyond its basic motion recognition capabilities, our algorithm can provide detailed descriptions of motions. In the example on the right, it can precisely describe the actions of the left hand and foot, as well as the right hand and foot. Our algorithm also", + "bbox": [ + 89, + 780, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "possesses extensive knowledge about various movements. For instance, when asked, \"what muscles does the person engage when bending their elbows during a push-up?\" it can accurately respond with \"triceps muscles.\" This demonstrates that our algorithm can understand the input motion sequences and perform corresponding analyses. We present more qualitative results in the supplementary material.", + "bbox": [ + 511, + 527, + 906, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 647, + 633, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we presented HuMoCon, a novel motion-video understanding framework for human action understanding and concept discovery. We demonstrated the effectiveness of our approach through comprehensive experiments, showcasing major improvements over state-of-the-art methods. Our work provides a strong foundation for future research in human action understanding and motion concept discovery. There are still some drawbacks of our work, such as the lack of human-environment contact or understanding niche motions. We plan to address those limitations by expanding the framework to incorporate additional modalities, particularly as environments, objects, and contact, and curate a wider range of specialized datasets.", + "bbox": [ + 511, + 672, + 908, + 869 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7186", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 230, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is supported by the Early Career Scheme of the Research Grants Council (grant # 27207224), the HKU-100 Award, a donation from the Musketeers Foundation, and an Academic Gift from Meta.", + "bbox": [ + 89, + 112, + 485, + 172 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 186, + 187, + 202 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Nikos Athanasiou, Mathis Petrovich, Michael J Black, and Gúl Varol. Teach: Temporal action composition for 3d humans. In 2022 International Conference on 3D Vision (3DV), pages 414-423. IEEE, 2022. 2", + "[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 1(2):3, 2023. 3", + "[3] Yang Bai, Desen Zhou, Songyang Zhang, Jian Wang, Errui Ding, Yu Guan, Yang Long, and Jingdong Wang. Action quality assessment with temporal parsing transformer. In European conference on computer vision, pages 422-438. Springer, 2022. 1, 3", + "[4] Pablo Cervantes, Yusuke Sekikawa, Ikuro Sato, and Koichi Shinoda. Implicit neural representations for variable length human motion generation. In European Conference on Computer Vision, pages 356-372. Springer, 2022. 1, 2", + "[5] Ling-Hao Chen, Shunlin Lu, Ailing Zeng, Hao Zhang, Benyou Wang, Ruimao Zhang, and Lei Zhang. Motionllm: Understanding human behaviors from human motions and videos. arXiv preprint arXiv:2405.20340, 2024. 1, 2, 3, 5, 6, 7", + "[6] Steven Chen and Richard R Yang. Pose trainer: correcting exercise posture using pose estimation. arXiv preprint arXiv:2006.11718, 2020. 1, 3", + "[7] Sihan Chen, Xingjian He, Longteng Guo, Xinxin Zhu, Weining Wang, Jinhui Tang, and Jing Liu. Valor: Vision-audiolanguage omni-perception pretraining model and dataset. arXiv preprint arXiv:2304.08345, 2023. 3", + "[8] Sihan Chen, Handong Li, Qunbo Wang, Zijia Zhao, Mingzhen Sun, Xinxin Zhu, and Jing Liu. Vast: A vision-audio-subtitle-text omni-modality foundation model and dataset. Advances in Neural Information Processing Systems, 36:72842-72866, 2023. 3", + "[9] Ginger Delmas, Philippe Weinzaepfel, Francesc Moreno-Noguer, and Grégory Rogez. Posefix: correcting 3d human poses with natural language. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15018-15028, 2023. 2", + "[10] Piotr Dólár, Vincent Rabaud, Garrison Cottrell, and Serge Belongie. Behavior recognition via sparse spatio-temporal features. In 2005 IEEE international workshop on visual surveillance and performance evaluation of tracking and surveillance, pages 65-72. IEEE, 2005. 3", + "[11] Yong Du, Wei Wang, and Liang Wang. Hierarchical recurrent neural network for skeleton based action recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1110-1118, 2015. 1, 2" + ], + "bbox": [ + 93, + 210, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Benjamin Elizalde, Soham Deshmukh, Mahmoud Al Ismail, and Huaming Wang. Clap learning audio concepts from natural language supervision. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023. 3", + "[13] Mark Endo, Joy Hsu, Jiaman Li, and Jiajun Wu. Motion question answering via modular motion programs. In International Conference on Machine Learning, pages 9312-9328. PMLR, 2023. 2, 7", + "[14] Qihang Fang, Yingda Yin, Qingnan Fan, Fei Xia, Siyan Dong, Sheng Wang, Jue Wang, Leonidas J Guibas, and Baoquan Chen. Towards accurate active camera localization. In European Conference on Computer Vision, pages 122-139. Springer, 2022. 1", + "[15] Qihang Fang, Chengcheng Tang, Bugra Tekin, and Yanchao Yang. Cigtime: Corrective instruction generation through inverse motion editing. In The Thirty-Eighth Annual Conference on Neural Information Processing Systems, 2024. 1, 2", + "[16] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 2, 3", + "[17] Yao Feng, Jing Lin, Sai Kumar Dwivedi, Yu Sun, Priyanka Patel, and Michael J Black. Chatpose: Chatting about 3d human pose. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2093-2103, 2024. 2", + "[18] Mihai Fieraru, Mihai Zanfir, Silviu Cristian Pirlela, Vlad Olaru, and Cristian Sminchisescu. Aifit: Automatic 3d human-interpretable feedback models for fitness training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9919–9928, 2021. 1, 3", + "[19] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15180-15190, 2023. 3", + "[20] Chuan Guo, Xinxin Zuo, Sen Wang, Shihao Zou, Qingyao Sun, Annan Deng, Minglun Gong, and Li Cheng. Action2motion: Conditioned generation of 3d human motions. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2021-2029, 2020. 1, 2", + "[21] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5152-5161, 2022. 6", + "[22] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5152-5161, 2022. 2", + "[23] Chuan Guo, Xinxin Zuo, Sen Wang, and Li Cheng. Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In European Conference on Computer Vision, pages 580-597. Springer, 2022. 2" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7187", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Tianyu Guo, Hong Liu, Zhan Chen, Mengyuan Liu, Tao Wang, and Runwei Ding. Contrastive learning from extremely augmented skeleton sequences for self-supervised action recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 762-770, 2022. 1, 2", + "[25] Tengda Han, Max Bain, Arsha Nagrani, Gul Varol, Weidi Xie, and Andrew Zisserman. Autoad ii: The sequel-who, when, and what in movie audio description. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13645–13655, 2023. 3", + "[26] Tengda Han, Max Bain, Arsha Nagrani, Gül Varol, Weidi Xie, and Andrew Zisserman. Autoad iii: The prequel-back to the pixels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18164-18174, 2024. 1, 3", + "[27] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 2, 4", + "[28] Daniel Holden, Taku Komura, and Jun Saito. Phase-functional neural networks for character control. ACM Transactions on Graphics (TOG), 36(4):1-13, 2017. 2", + "[29] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 6", + "[30] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. Advances in Neural Information Processing Systems, 36:20067-20079, 2023. 2", + "[31] Hanwen Jiang, Shaowei Liu, Jiashun Wang, and Xiaolong Wang. Hand-object contact consistency reasoning for human grasps generation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11107-11116, 2021. 2", + "[32] Yujiao Jiang, Qingmin Liao, Zhaolong Wang, Xiangru Lin, Zongqing Lu, Yuxi Zhao, Hanqing Wei, Jingrui Ye, Yu Zhang, and Zhijing Shao. Smplx-lite: A realistic and drivable avatar benchmark with rich geometry and texture annotations. arXiv preprint arXiv:2405.19609, 2024. 1", + "[33] Peng Jin, Ryuichi Takanobu, Wancai Zhang, Xiaochun Cao, and Li Yuan. Chat-univi: Unified visual representation empowers large language models with image and video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13700-13710, 2024. 7", + "[34] Qiuhong Ke, Mohammed Bennamoun, Senjian An, Ferdous Sohel, and Farid Boussaid. A new representation of skeleton sequences for 3d action recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3288-3297, 2017. 1, 2", + "[35] Jihoon Kim, Jiseob Kim, and Sungjoon Choi. Flame: Freeform language-based motion synthesis & editing. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8255-8263, 2023. 2", + "[36] Kosmas Kritsis, Aggelos Gkiokas, Aggelos Pikrakis, and Vassilis Katsouros. Danceconv: Dance motion genera" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tion with convolutional networks. IEEE Access, 10:44982-45000, 2022. 2", + "[37] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 1, 2, 3, 6, 7", + "[38] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 7", + "[39] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mybench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 7", + "[40] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 1, 2, 3, 6, 7", + "[41] Jing Lin, Ailing Zeng, Shunlin Lu, Yuanhao Cai, Ruimao Zhang, Haoqian Wang, and Lei Zhang. Motion-x: A large-scale 3d expressive whole-body human motion dataset. Advances in Neural Information Processing Systems, 36, 2024. 6", + "[42] Lilang Lin, Jiahang Zhang, and Jiaying Liu. Actionlet-dependent contrastive learning for unsupervised skeleton-based action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2363-2372, 2023. 1, 2", + "[43] Daochang Liu, Qiyue Li, Tingting Jiang, Yizhou Wang, Rulin Miao, Fei Shan, and Ziyu Li. Towards unified surgical skill assessment. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9522-9531, 2021. 1, 3", + "[44] Jun Liu, Amir Shahroudy, Dong Xu, and Gang Wang. Spatio-temporal LSTM with trust gates for 3d human action recognition. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 816-833. Springer, 2016. 1, 2, 3", + "[45] Mengyuan Liu, Hong Liu, and Chen Chen. Enhanced skeleton visualization for view invariant human action recognition. Pattern Recognition, 68:346-362, 2017. 1, 2", + "[46] Ruizhe Liu, Qian Luo, and Yanchao Yang. Infocon: Concept discovery with generative and discriminative informativeness. arXiv preprint arXiv:2404.10606, 2024. 2, 4, 5", + "[47] Thomas Lucas, Fabien Baradel, Philippe Weinzaepfel, and Grégory Rogez. Posegpt: Quantization-based 3d human motion generation and forecasting. In European Conference on Computer Vision, pages 417-435. Springer, 2022. 2", + "[48] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing, 508:293-304, 2022. 3", + "[49] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video" + ], + "bbox": [ + 516, + 90, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7188", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 7", + "[50] Shehan Munasinghe, Rusiru Thushara, Muhammad Maaz, Hanoona Abdul Rasheed, Salman Khan, Mubarak Shah, and Fahad Khan. Pg-video-llava: Pixel grounding large video-language models. arXiv preprint arXiv:2311.13435, 2023. 1, 3", + "[51] Paritosh Parmar and Brendan Tran Morris. What and how well you performed? a multitask learning approach to action quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 304-313, 2019. 1, 3", + "[52] Mathis Petrovich, Michael J Black, and Gül Varol. Action-conditioned 3d human motion synthesis with transformer vae. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10985-10995, 2021. 1, 2", + "[53] Mathis Petrovich, Michael J Black, and Gül Varol. Temos: Generating diverse human motions from textual descriptions. In European Conference on Computer Vision, pages 480-497. Springer, 2022. 2", + "[54] Abhinanda R Punnakkal, Arjun Chandrasekaran, Nikos Athanasiou, Alejandra Quiros-Ramirez, and Michael J Black. Babel: Bodies, action and behavior with english labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 722-731, 2021. 2", + "[55] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3, 7", + "[56] Jathushan Rajasegaran, Georgios Pavlakos, Angjoo Kanazawa, Christoph Feichtenhofer, and Jitendra Malik. On the benefits of 3d pose and tracking for human action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 640-649, 2023. 2, 5", + "[57] Leonard Salewski, Stefan Fauth, A Koepke, and Zeynep Akata. Zero-shot audio captioning with audio-language model guidance and audio context keywords. arXiv preprint arXiv:2311.08396, 2023. 3, 7", + "[58] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Two-stream adaptive graph convolutional networks for skeleton-based action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12026-12035, 2019. 1, 2, 7", + "[59] Sebastian Starke, He Zhang, Taku Komura, and Jun Saito. Neural state machine for character-scene interactions. ACM Transactions on Graphics, 38(6):178, 2019. 1, 2", + "[60] Sebastian Starke, Yiwei Zhao, Taku Komura, and Kazi Zaman. Local motion phases for learning multi-contact character movements. ACM Transactions on Graphics (TOG), 39 (4):54-1, 2020. 2", + "[61] Sebastian Starke, Ian Mason, and Taku Komura. Deeppphase: Periodic autoencoders for learning motion phase manifolds." + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ACM Transactions on Graphics (TOG), 41(4):1-13, 2022. 1, 2", + "[62] Graham W Taylor, Rob Fergus, Yann LeCun, and Christoph Bregler. Convolutional learning of spatio-temporal features. In Computer Vision-ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part VI 11, pages 140-153. Springer, 2010. 2, 3", + "[63] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. In European Conference on Computer Vision, pages 358–374. Springer, 2022. 2, 3, 7", + "[64] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022. 2", + "[65] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 4489-4497, 2015. 2, 3", + "[66] Raviteja Vemulapalli, Felipe Arrate, and Rama Chellappa. Human action recognition by representing 3d skeletons as points in a lie group. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 588-595, 2014. 3", + "[67] Heng Wang and Cordelia Schmid. Action recognition with improved trajectories. In Proceedings of the IEEE international conference on computer vision, pages 3551-3558, 2013. 3", + "[68] Jingbo Wang, Ye Yuan, Zhengyi Luo, Kevin Xie, Dahua Lin, Umar Iqbal, Sanja Fidler, and Sameh Khamis. Learning human dynamics in autonomous driving scenarios. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20796-20806, 2023. 1", + "[69] Xin Wang, Jiawei Wu, Junkun Chen, Lei Li, Yuan-Fang Wang, and William Yang Wang. Vatex: A large-scale, high-quality multilingual dataset for video-and-language research. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4581-4591, 2019. 6", + "[70] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, et al. Internvideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191, 2022. 2, 3", + "[71] Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Guo Chen, Baoqi Pei, Rongkun Zheng, Jilan Xu, Zun Wang, et al. Intervideo2: Scaling video foundation models for multimodal video understanding. arXiv preprint arXiv:2403.15377, 2024. 2, 3", + "[72] Zeqi Xiao, Tai Wang, Jingbo Wang, Jinkun Cao, Wenwei Zhang, Bo Dai, Dahua Lin, and Jiangmiao Pang. Unified human-scene interaction via prompted chain-of-contacts. arXiv preprint arXiv:2309.07918, 2023. 1", + "[73] Junyu Xie, Tengda Han, Max Bain, Arsha Nagrani, Güll Varol, Weidi Xie, and Andrew Zisserman. Autoad-zero: A training-free framework for zero-shot audio description. arXiv preprint arXiv:2407.15850, 2024. 3" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "7189", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[74] Guo-Hao Xu, Yi-Lin Wei, Dian Zheng, Xiao-Ming Wu, and Wei-Shi Zheng. Dexterous grasp transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17933-17942, 2024. 2", + "[75] Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. Pllava: Parameter-free llava extension from images to videos for video dense captioning. arXiv preprint arXiv:2404.16994, 2024. 1, 3", + "[76] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In Proceedings of the AAAI conference on artificial intelligence, 2018. 1, 2", + "[77] Zeshi Yang, Kangkang Yin, and Libin Liu. Learning to use chopsticks in diverse gripping styles. 2022. 2", + "[78] Zhou Yu, Dejing Xu, Jun Yu, Ting Yu, Zhou Zhao, Yueting Zhuang, and Dacheng Tao. Activitynet-qa: A dataset for understanding complex web videos via question answering. In AAAI, pages 9127–9134, 2019. 2, 7", + "[79] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 6", + "[80] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 3, 7", + "[81] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Shaoli Huang, Yong Zhang, Hongwei Zhao, Hongtao Lu, and Xi Shen. T2m-gpt: Generating human motion from textual descriptions with discrete representations. arXiv preprint arXiv:2301.06052, 2023. 2", + "[82] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2", + "[83] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xupeng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. Pointclip: Point cloud understanding by clip. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8552-8562, 2022. 3", + "[84] Renrui Zhang, Jiaming Han, Chris Liu, Peng Gao, Aojun Zhou, Xiangfei Hu, Shilin Yan, Pan Lu, Hongsheng Li, and Yu Qiao. Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199, 2023. 7", + "[85] Xingyi Zhou, Anurag Arnab, Shyamal Buch, Shen Yan, Austin Myers, Xuehan Xiong, Arsha Nagrani, and Cordelia Schmid. Streaming dense video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18243-18252, 2024. 3", + "[86] Zixiang Zhou, Yu Wan, and Baoyuan Wang. Avatargpt: All-in-one framework for motion understanding planning generation and beyond. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1357-1366, 2024. 2", + "[87] Bin Zhu, Bin Lin, Munan Ning, Yang Yan, Jiaxi Cui, HongFa Wang, Yatian Pang, Wenhao Jiang, Junwu Zhang, Zongwei" + ], + "bbox": [ + 91, + 90, + 485, + 901 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Li, et al. Languagebind: Extending video-language pretraining to n-modality by language-based semantic alignment. arXiv preprint arXiv:2310.01852, 2023. 2, 3, 5", + "bbox": [ + 545, + 90, + 903, + 133 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "7190", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/HuMoCon_ Concept Discovery for Human Motion Understanding/b5dd425e-4f3e-4b2d-9c88-36e4f0bafd85_model.json b/2025/HuMoCon_ Concept Discovery for Human Motion Understanding/b5dd425e-4f3e-4b2d-9c88-36e4f0bafd85_model.json new file mode 100644 index 0000000000000000000000000000000000000000..fee45f25d4bda3ce3ed4d6b01fab0de061fe8658 --- /dev/null +++ b/2025/HuMoCon_ Concept Discovery for Human Motion Understanding/b5dd425e-4f3e-4b2d-9c88-36e4f0bafd85_model.json @@ -0,0 +1,2633 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.166, + 0.131, + 0.834, + 0.154 + ], + "angle": 0, + "content": "HuMoCon: Concept Discovery for Human Motion Understanding" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.18, + 0.276, + 0.2 + ], + "angle": 0, + "content": "Qihang Fang\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.297, + 0.181, + 0.45, + 0.199 + ], + "angle": 0, + "content": "Chengcheng Tang" + }, + { + "type": "text", + "bbox": [ + 0.341, + 0.199, + 0.586, + 0.218 + ], + "angle": 0, + "content": "1The University of Hong Kong" + }, + { + "type": "text", + "bbox": [ + 0.599, + 0.181, + 0.7, + 0.199 + ], + "angle": 0, + "content": "Shugao Ma2" + }, + { + "type": "text", + "bbox": [ + 0.723, + 0.181, + 0.843, + 0.199 + ], + "angle": 0, + "content": "Yanchao Yang" + }, + { + "type": "text", + "bbox": [ + 0.606, + 0.199, + 0.658, + 0.217 + ], + "angle": 0, + "content": "\\(^{2}\\)Meta" + }, + { + "type": "text", + "bbox": [ + 0.143, + 0.219, + 0.848, + 0.234 + ], + "angle": 0, + "content": "qihfang@gmail.com, {chengcheng.tang,bugratekin,shugao}@meta.com, yanchaoy@hku.hk" + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.251, + 0.56, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.576, + 0.252, + 0.905, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.413, + 0.907, + 0.471 + ], + "angle": 0, + "content": "Figure 1. We present HuMoCon for concept discovery and human motion understanding with video or motion input. To address the challenge of feature misalignment and high-frequency information loss, we propose a novel feature alignment strategy and an advanced masked auto-encoder reconstructing velocity. HuMoCon empowers effective motion concept discovery and accurate Question Answering tasks, significantly outperforming state-of-the-art methods qualitatively and quantitatively." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.482, + 0.327, + 0.499 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.514, + 0.486, + 0.818 + ], + "angle": 0, + "content": "We present HuMoCon, a novel motion-video understanding framework designed for advanced human behavior analysis. The core of our method is a human motion concept discovery framework that efficiently trains multi-modal encoders to extract semantically meaningful and generalizable features. HuMoCon addresses key challenges in motion concept discovery for understanding and reasoning, including the lack of explicit multi-modality feature alignment and the loss of high-frequency information in masked autoencoding frameworks. Our approach integrates a feature alignment strategy that leverages video for contextual understanding and motion for fine-grained interaction modeling, further with a velocity reconstruction mechanism to enhance high-frequency feature expression and mitigate temporal over-smoothing. Comprehensive experiments on standard benchmarks demonstrate that HuMoCon enables effective motion concept discovery and significantly outperforms state-of-the-art methods in training large models for human motion understanding. We will open-source the associated code with our paper." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.845, + 0.222, + 0.86 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Understanding human behavior is crucial for gaining deeper insights into human actions and reasoning, which are foun" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.483, + 0.907, + 0.606 + ], + "angle": 0, + "content": "ditional for developing human-centered AI systems [5, 26, 37, 61] that effectively address real-world tasks [14, 68, 72]. Recent advancements in large language models (LLMs) have led to significant progress in video- and motion-based methods [5, 15, 37, 40, 75] for analyzing human behavior. Despite these advancements, achieving a precise and fine-grained understanding of human actions remains a persistent and complex challenge." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.61, + 0.909, + 0.761 + ], + "angle": 0, + "content": "Human behavior is often represented as motion sequences structured based on the kinematics of the human skeleton [32?]. Traditional approaches to analyzing motion sequences have predominantly focused on classification [11, 24, 34, 42, 44, 45, 58, 76], relying on predefined action categories. While effective for specific tasks, these methods often lack the flexibility for deeper or more nuanced analysis. Other methods employ statistical models to assess motion quality [3, 6, 18, 43, 51], but they require significant expert knowledge, limiting scalability." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Recent advancements in language models [40, 50, 75] have opened new possibilities in motion analysis, enabling tasks such as motion generation and motion captioning based on free-text inputs [4, 20, 52, 59, 61], advancing the field of motion analysis. However, obtaining high-quality motion sequences and annotations remains costly, and motion sequences generally lack information about the surrounding environment, limiting models' ability to understand human-environment interactions." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7179" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.213 + ], + "angle": 0, + "content": "In contrast, video data offers rich visual context and is easier to collect. It has been widely used for action classification [16, 62, 65] and video Q&A [5, 37, 40]. However, video data often contains irrelevant elements that can hinder accurate content analysis. For example, a model might learn from training data that most running occurs in stadiums, leading it to mistakenly classify other stadium activities as running during testing." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.216, + 0.487, + 0.502 + ], + "angle": 0, + "content": "In recent years, significant progress has been made in developing multi-modality models [5, 87]. Integrating multiple modalities, as opposed to relying on a single modality, enhances the network's ability to learn fine-grained and semantic features. For instance, LART [56] combines video and motion data to ensure that video-based action classification focuses on the human body rather than overemphasizing environmental context. MotionLLM [5] expands large language models (LLMs) to video and motion, building a model that can understand both. MotionLLM demonstrated that learning from video and motion improves the network's ability to analyze human behavior. However, MotionLLM implicitly aligns video and motion using motion-video pairs during LLM training without optimizing the encoder for explicit cross-modal alignment. To address this limitation and further enhance feature representation, we propose explicitly aligning features from different modalities during the encoding process within a human motion concept discovery framework." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.506, + 0.483, + 0.793 + ], + "angle": 0, + "content": "Achieving effective cross-modal alignment and feature representation relies heavily on robust encoder training methods. Many existing studies [70, 71, 87] adopt a masked autoencoder framework [27] when training encoders. This approach involves applying a mask to a certain proportion of the input, extracting features from the masked data using the encoder, and reconstructing the original unmasked input with a decoder. While effective in learning high-quality features, this masking strategy inherently causes the loss of high-frequency information, resulting in temporally over-smoothed reconstructions that hinder the LLM's ability to understand human behavior. [47] To address this issue, we propose reconstructing velocity as a means to preserve high-frequency information. Following InfoCon [46], we design a VQ-VAE-based loss function that helps retain the fine-grained features of the original masked autoencoder while further enhancing high-frequency feature expression. This approach significantly boosts the subsequent LLM's capacity to model and understand human motion effectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In summary, we present HuMoCon, an effective human motion concept discovery pertaining framework for advanced human behavior understanding from both video and motion data. Our key contributions include: (i) a novel feature alignment strategy that explicitly leverages video data for contextual understanding and motion data for fine-grained interaction modeling, and (ii) an enhanced masked" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.243 + ], + "angle": 0, + "content": "autoencoder framework incorporating velocity reconstruction to significantly improve high-frequency feature expression and mitigate temporal over-smoothing. To evaluate the effectiveness of our proposed method, we conduct experiments on the Activity-QA [78] and BABEL-QA [13] benchmarks [54], comparing our approach with the state-of-the-art methods. Comprehensive quantitative and qualitative evaluations demonstrate that HuMoCon achieves state-of-the-art performance on standard benchmarks for motion-video understanding." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.261, + 0.655, + 0.276 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.287, + 0.787, + 0.304 + ], + "angle": 0, + "content": "2.1. Human Motion Understanding" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.31, + 0.905, + 0.598 + ], + "angle": 0, + "content": "The core of human motion understanding is to learn the semantics of motion sequences or videos. In the early stages, research in this field is often based on predefined action categories. Such work can be divided into two types: action classification [11, 24, 34, 42, 44, 45, 58, 76] and motion generation targeted at specific categories [4, 20, 28, 31, 36, 52, 59-61, 74, 77]. For action recognition, several works [11, 44] utilize Recurrent Neural Networks (RNN) to extract temporal information from human joint sequences, beyond the CNN-based ones [34, 45]. Subsequent works [24, 42, 58, 76] propose representing motion data as a graph structure, enabling the use of spatio-temporal graph convolutional networks (ST-GCNs) to effectively capture the connectivity and interactions between human joints. For motion generation, several studies [28, 31, 36, 60, 74, 77] have concentrated on generating specific, single actions. In contrast, other works [4, 20, 52, 59, 61] aim to generate a broader range of diverse action types, addressing the challenge of variability in motion synthesis." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.6, + 0.905, + 0.901 + ], + "angle": 0, + "content": "In recent years, more research has shifted towards captioning [23, 30], generation [1, 22, 23, 30, 35, 53, 64, 81, 82, 86], and analysis [5, 9, 15, 17] based on free-form text. Leveraging alignment of motion latent features with the CLIP [55] embedding space, MotionCLIP [63] enables the generation of out-of-distribution motions. Transformers [1, 22, 23, 53, 81] and diffusion [35, 64, 82] models are also utilized for text-conditioned motion generation. Moreover, MotionGPT [30] and AvatarGPT [86] utilize large language models (LLMs) to handle multiple motion-related tasks simultaneously. Despite these advancements, research focused specifically on motion analysis remains relatively limited. PoseFix [9] and CigTime [15] are proposed to annotate the corrective instructions from source to target for pose and motion sequences. MotionLLM [5] proposes a method to analyze both the video and motion sequences with detailed spatial-temporal awareness and reasoning abilities. However, MotionLLM relies on implicit alignment through paired data and lacks explicit cross-modal feature alignment." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7180" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.293, + 0.108 + ], + "angle": 0, + "content": "2.2. Video Understanding" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.486, + 0.522 + ], + "angle": 0, + "content": "In addition to motion understanding, video understanding has emerged as a prominent research area, attracting extensive attention in recent years. Early works on video classification and segmentation relied on handcrafted features [10, 44, 66, 67] or convolutional neural networks [16, 62, 65]. Beyond classification, video analysis [3, 6, 18, 25, 26, 37, 40, 43, 50, 51, 57, 73, 75, 85] has become a highly active field of research. Action quality assessment has been the focus of several works [3, 6, 18, 43, 51], with methods such as Aifit [18] and PoseTrainer [6] proposed to correct user actions. These approaches evaluate actions in conjunction with complex textual instructions [3, 43, 51]. Other studies [25, 26, 57, 73, 85] aim to generate descriptive narratives for videos, broadening the scope of video understanding. Recently, many works [2, 37, 40, 50, 75, 80] have leveraged large language models (LLMs) to understand video content and perform complex reasoning tasks. These methods combine video processing with the advanced reasoning capabilities of LLMs, further advancing the state-of-the-art in video analysis. Video-LLaVA [40] build a model to perform visual reasoning capabilities on images and videos simultaneously. Follow-up works [37, 50, 75] extend Video-LLaVA for faster inference and more accurate reasoning ability. Unlike typical video understanding algorithms, which rely heavily on visual context alone, our approach integrates motion data directly, enhancing the model's ability to understand human-centric actions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.531, + 0.345, + 0.547 + ], + "angle": 0, + "content": "2.3. Multi-modality Pre-training" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.553, + 0.483, + 0.658 + ], + "angle": 0, + "content": "Multi-modality pre-training has primarily focused on aligning vision and language representations. CLIP [55] pioneered this effort by using contrastive learning to align image and text features on large-scale datasets. Subsequent works [12, 48, 63, 83] extended CLIP's framework to other modalities, including video [48], audio [12] and motion [63]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.483, + 0.749 + ], + "angle": 0, + "content": "Recent studies, such as VALOR [7] and VAST [8], demonstrated that cross-modal alignment not only enhances model robustness but also maintains high performance across modalities. Building on this, follow-up works [19, 87] have aimed to expand the number of aligned modalities, further advancing multi-modality research." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.811 + ], + "angle": 0, + "content": "In the context of motion and video alignment, Motion-LLMs [5] is the first work to address this challenge. However, it relies solely on implicit alignment during the fine-tuning stage of large language models." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.824, + 0.182, + 0.839 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.201, + 0.863 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Our objective is to develop an LLM-based system capable of understanding and reasoning both human video and mo" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "tion inputs. Given a question \\( Q \\) and an input, which can be either a video \\( U \\) or motion sequence \\( M \\), the model generates the corresponding answer \\( W \\). Specifically, a video \\( U \\) is represented as a sequence of \\( T \\) keyframes, denoted as \\( \\{u_0, u_1, \\dots, u_{T-1}\\} \\). Similarly, a motion sequence \\( M \\) is represented as \\( K \\) consecutive pose frames, denoted as \\( \\{m_0, m_1, \\dots, m_{K-1}\\} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.199, + 0.907, + 0.44 + ], + "angle": 0, + "content": "We present an overview of the pipeline in Fig. 2. To enable effective multi-modal human motion understanding, our framework focuses on: (1) discovering human motion concepts from both video and motion data, (2) preserving high-frequency motion details using velocity reconstruction to reduce temporal smoothing, and (3) aligning video and motion features to enhance multi-modal comprehension. To accomplish this, we train our video and motion encoders with four key learning objectives: Masked Reconstruction, Discriminative Informativeness, Actionable Informativeness, and Feature Alignment. Consequently, we leverage the encoded features of videos and motions to facilitate the fine-tuning of LLMs for video-motion understanding and reasoning. The following sections detail each component of our approach, including multi-modality encoder pre-training, modality translation, and instruction tuning." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.449, + 0.866, + 0.481 + ], + "angle": 0, + "content": "3.2. Multi-Modality Encoder Pre-training via Human Motion Concept Discovery" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.486, + 0.907, + 0.669 + ], + "angle": 0, + "content": "Pre-training encoders are critical for multi-modality large language models (LLMs) [70, 71, 87], as they enable the effective extraction of essential features from large-scale input data. To enhance the encoder's ability to capture key information, we design three distinct learning objectives: masked reconstruction, velocity reconstruction, and feature alignment, all contributing to the quality of the discovered human motion concepts as well as the pre-training efficiency. For both videos and motions, we adopt a similar network structure during encoder pre-training. Consequently, in the following section, we focus on detailing the network structure for video inputs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.687, + 0.906, + 0.762 + ], + "angle": 0, + "content": "Encoding Function. In our approach, we design masked autoencoder architectures tailored for both video and motion data. Specifically, for videos, we utilize a pre-trained encoder \\(\\mathcal{E}_u\\), to encode the input \\(U\\) into a feature representation \\(f^u\\):" + }, + { + "type": "equation", + "bbox": [ + 0.663, + 0.763, + 0.905, + 0.78 + ], + "angle": 0, + "content": "\\[\nF ^ {u} = \\mathcal {E} _ {u} (U). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.786, + 0.906, + 0.83 + ], + "angle": 0, + "content": "Then, we employ a codebook \\( C^u \\), modeled with VQ-VAE to quantize \\( f^u \\) into a discrete feature representation \\( d^u \\):" + }, + { + "type": "equation", + "bbox": [ + 0.623, + 0.832, + 0.905, + 0.849 + ], + "angle": 0, + "content": "\\[\nD ^ {u} = \\operatorname {Q u a n t i z e} \\left(F ^ {u}, C ^ {u}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "where 'Quantize' denotes the quantization operation, and \\( C^u \\) is the codebook used for discretization, also serving as the discovered human motion concepts." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7181" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.145, + 0.089, + 0.855, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.364, + 0.907, + 0.448 + ], + "angle": 0, + "content": "Figure 2. System overview of our method. (a) The encoder pre-training process for learning and aligning video and motion features and enhancing high-frequency details through velocity reconstruction. We utilize a VQ-VAE-based structure, and we design effective learning objectives to enhance the encoder to extract semantic meaningful and fine-grained features. (b) The fine-tuning of the large language model (LLM) for video and motion reasoning consists of two stages: Modality Translation and Multi-modality Instruction Tuning. In the Modality Translation stage, we train a translation layer for each modality to map the encoding feature to the LLM space. In the Instruction Tuning stage, we fine-tune the LLM to understand human motion and videos for downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.462, + 0.483, + 0.552 + ], + "angle": 0, + "content": "Masked Reconstruction. The masked autoencoder has proven to be highly effective for pre-training [27], as it facilitates learning of meaningful and low-frequency features from input data. Thus, we leverage masked auto-encoding to enhance the discovered concepts and facilitate the multimodal feature pre-training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.553, + 0.483, + 0.586 + ], + "angle": 0, + "content": "More explicitly, we mask \\(D^u\\) by a probability of \\(\\sigma\\) to obtain the masked feature \\(D_{\\mathrm{mask}}^u\\):" + }, + { + "type": "equation", + "bbox": [ + 0.208, + 0.596, + 0.482, + 0.613 + ], + "angle": 0, + "content": "\\[\nD _ {\\text {m a s k}} ^ {u} = \\operatorname {M a s k} \\left(D ^ {u}, \\sigma\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.625, + 0.484, + 0.673 + ], + "angle": 0, + "content": "where 'Mask' is the operation that applies masking to the features, and \\(\\sigma < 1.0\\) controls the masking ratio. A video decoder \\(D^{u}\\) is used to reconstruct the video \\(U^{\\prime}\\) from \\(\\mathcal{D}_{\\mathrm{mask}}^{u}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.226, + 0.686, + 0.482, + 0.703 + ], + "angle": 0, + "content": "\\[\nU ^ {\\prime} = \\mathcal {D} ^ {u} \\left(D _ {\\text {m a s k}} ^ {u}\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.711, + 0.483, + 0.74 + ], + "angle": 0, + "content": "We compute the reconstruction loss between the original video \\(U\\) and the reconstructed video \\(U^{\\prime}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.215, + 0.753, + 0.482, + 0.771 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {v i d e o}} ^ {\\text {r e c}} = \\left| \\left| U ^ {\\prime} - U \\right| \\right| _ {2} ^ {2}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.784, + 0.483, + 0.828 + ], + "angle": 0, + "content": "Similarly, we introduce a reconstruction loss for motion data, \\(\\mathcal{L}_{\\mathrm{motion}}^{\\mathrm{rec}}\\), following the same process. The total reconstruction loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.207, + 0.841, + 0.483, + 0.859 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {\\text {r e c}} = \\mathcal {L} _ {\\text {v i d e o}} ^ {\\text {r e c}} + \\mathcal {L} _ {\\text {m o t i o n}} ^ {\\text {r e c}}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.9 + ], + "angle": 0, + "content": "The above masked reconstruction loss ensures that the encoded features contain sufficient information about the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.462, + 0.907, + 0.553 + ], + "angle": 0, + "content": "video or motion input, while learning away possible nuisances through masking and making the learning of the concepts (codebook) more focused on critical information. On the other hand, the discretization imposed by the concepts further enhances the meaningfulness and robustness of the pre-trained multi-modal (motion) features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.581, + 0.906, + 0.656 + ], + "angle": 0, + "content": "Velocity Reconstruction. Reconstructing inputs from masked features inevitably results in the loss of high-frequency details, leading to temporally over-smoothed reconstruction outcomes and preventing the discovery of more nuanced yet useful motion concepts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.658, + 0.906, + 0.809 + ], + "angle": 0, + "content": "To address this limitation, we propose an objective that further reconstructs the \"velocity\" of the input sequence. To be more specific, we define \"state\" as the encoding or feature of a frame in the video or motion sequence, while the \"velocity\" means the difference between consecutive \"states\" that captures high-frequency information by focusing on dynamic changes. For video and motion data, we denote the \"states\" as \\( U \\) and \\( m \\), and the \"velocity\" as the optical flow \\( O \\) (computed from image frames) and delta motion \\( \\delta M \\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.901 + ], + "angle": 0, + "content": "However, directly reconstructing velocity from the same set of masked features introduces difficulty in figuring out the dynamic information, given the above-mentioned oversmoothing effect in the encoded features. To mitigate this, we draw inspiration from the robotic manipulation literature [46] on action concept discovery and incorporate two" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7182" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.174, + 0.088, + 0.825, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.245, + 0.908, + 0.318 + ], + "angle": 0, + "content": "Figure 3. Overview of the velocity reconstruction components. We build similar network structures for both video and motion, and present the video part in this figure. This module is composed of two learning objectives. 1) Discriminative informativeness (left) aims to improve the distinctiveness of encoded features by reducing representational ambiguity. 2) Actionable informativeness (right) focuses on reconstructing the velocity by leveraging gradient information from the discrimination hypernetwork. As for the video data, we employ optical flow as the representation of the velocity." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.329, + 0.482, + 0.404 + ], + "angle": 0, + "content": "tailored learning objectives - discriminative informativeness and actionable informativeness - to enhance the efficiency of representations and improve the quality of temporal dynamics in multi-modality pretaining, as shown in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.406, + 0.484, + 0.601 + ], + "angle": 0, + "content": "Discriminative informativeness. The discriminative informativeness objective aims to improve the distinctiveness of encoded features by increasing the correlation between the discovered concepts and the corresponding state features. Specifically, we introduce a video hyper-network \\(\\mathcal{H}^u\\), which takes discrete features \\(c_k^u\\) from the codebook \\(C^u\\) as input and generates a classification network. The hyper-network \\(\\mathcal{H}^u\\) (together with the generated classifier) is trained to determine whether an input state is compatible with its corresponding code class, which is represented by a confidence value \\(s_u\\) (classification score) from 0 to 1, thereby enhancing the distinguishability and separation of each feature class in the representation space." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.603, + 0.42, + 0.618 + ], + "angle": 0, + "content": "The discriminative loss for video is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.147, + 0.635, + 0.483, + 0.679 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {v i d e o}} ^ {\\text {d i s}} = \\sum_ {k = 0} ^ {| C ^ {u} | - 1} \\sum_ {i = 0} ^ {T - 1} C E \\left(\\mathcal {H} ^ {u} \\left(c _ {k} ^ {u}, u _ {i}\\right), d _ {i} ^ {u}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.687, + 0.483, + 0.748 + ], + "angle": 0, + "content": "where \\( |C^u| \\) is the size of the codebook \\( C^u \\), \\( CE(\\cdot) \\) is the cross-entropy loss for classification, \\( d_i^u \\) represents the \"ground-truth label\" for the input state \\( u_i \\) determined by the VQ-VAE, and \\( T \\) is the number of frames." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.75, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Actionable informativeness. The actionable informativeness objective focuses on reconstructing the velocity by leveraging gradient information from the discriminative hyper-network, \\(\\mathcal{H}^u\\) introduced above. Inspired by the idea that gradients of a discriminator function can inform the state changes, as observed in [46], this objective ensures that the model captures not only static state information, but also the dynamic changes in the sequence. By leveraging gradient fluctuations around a given state, the approach effectively captures dynamic motion details, such as velocity," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.329, + 0.905, + 0.359 + ], + "angle": 0, + "content": "resulting in more fine-grained motion concepts, and enriching the motion representation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.36, + 0.906, + 0.435 + ], + "angle": 0, + "content": "Specifically, we initialize a network \\(\\mathcal{A}^u\\) for video data, which takes key image frames \\(U\\) and gradients from the discriminator \\(\\nabla_U\\mathcal{H}^u (D^u,U)\\) as inputs and predicts the velocity \\(V_{u}^{\\prime}\\). The actionable informativeness loss for video is then computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.635, + 0.447, + 0.905, + 0.466 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {v i d e o}} ^ {\\text {a c t}} = \\left| \\left| V _ {u} - V _ {u} ^ {\\prime} \\right| \\right| _ {2} ^ {2}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.478, + 0.905, + 0.524 + ], + "angle": 0, + "content": "where \\(V_{u}\\) represents the ground-truth velocity of video frames. This process is repeated for motion data to compute the corresponding loss \\(\\mathcal{L}_{\\mathrm{motion}}^{\\mathrm{act}}\\) using \\(V_{m}\\) and \\(V_{m}^{\\prime}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.906, + 0.584 + ], + "angle": 0, + "content": "Combined Objectives. By integrating both video and motion components, we derive the total velocity reconstruction loss for human motion concept discovery and motion feature pre-training as in the following:" + }, + { + "type": "equation", + "bbox": [ + 0.633, + 0.595, + 0.905, + 0.615 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {\\text {d i s}} = \\mathcal {L} _ {\\text {v i d e o}} ^ {\\text {d i s}} + \\mathcal {L} _ {\\text {m o t i o n}} ^ {\\text {d i s}} \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.634, + 0.643, + 0.905, + 0.661 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {\\mathrm {a c t}} = \\mathcal {L} _ {\\text {v i d e o}} ^ {\\mathrm {a c t}} + \\mathcal {L} _ {\\text {m o t i o n}} ^ {\\mathrm {a c t}} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.668, + 0.906, + 0.729 + ], + "angle": 0, + "content": "With the above objectives, our approach ensures that the model effectively reconstructs high-frequency details, improving the temporal fidelity and overall quality of multimodal representations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Feature Alignment. Multi-modality models often leverage the shared information across different modalities to enhance performance [5, 56, 87]. In the case of motion and video, video data provides contextual information that motion data alone may be lacking, which is critical for accurate motion prediction and understanding. Conversely, motion data focuses on human-centric features in the video, reducing over-reliance on environmental details. An efficient utilization or balancing of both is beneficial for the human motion LLMs. However, prior work on human motion" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7183" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.166 + ], + "angle": 0, + "content": "and video understanding [5] only achieves implicit alignment through paired data during LLM fine-tuning, without explicit alignment in encoder training. To address this issue, we propose an alignment loss to improve information sharing and cross-modal coherence." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.168, + 0.482, + 0.272 + ], + "angle": 0, + "content": "To achieve alignment between video and motion features, we collect a set of video-motion pairs from Motion-X [41], which are aligned by frames. We instantiate two projection layers: The video projection layer \\(\\mathcal{P}^u\\) and the motion projection layer \\(\\mathcal{P}^m\\). These layers are designed to map the respective features from the video and motion domains into a shared aligned space." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.273, + 0.482, + 0.365 + ], + "angle": 0, + "content": "More explicitly, given paired video and motion data \\(\\{U,M\\}\\), we use \\(\\mathcal{P}^u\\) and \\(\\mathcal{P}^m\\) to project the discrete video feature \\(d_i^u\\) and motion feature \\(d_j^m\\) into the aligned feature space, resulting in \\(a_i^u\\) and \\(a_j^m\\), where \\(a_i^u, a_j^m \\in \\mathbb{R}^H\\). Here, \\(H\\) denotes the dimensionality of the aligned features in the shared space." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.365, + 0.482, + 0.394 + ], + "angle": 0, + "content": "To ensure effective alignment, we define an alignment loss as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.401, + 0.482, + 0.445 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {\\text {a l i g n}} = \\sum_ {0} ^ {T - 1} \\frac {\\operatorname {n o r m} \\left(a _ {i} ^ {u}\\right) \\cdot \\operatorname {n o r m} \\left(a _ {\\text {a l i g n} (i)} ^ {m}\\right) / \\epsilon}{\\sum_ {0} ^ {K - 1} \\operatorname {n o r m} \\left(a _ {i} ^ {u}\\right) \\cdot \\operatorname {n o r m} \\left(a _ {j} ^ {m}\\right) / \\epsilon}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.452, + 0.482, + 0.526 + ], + "angle": 0, + "content": "where \\(\\mathrm{align}(i)\\) indicates the pose index in the motion sequence corresponding to the key video frame \\(u_{i}\\), norm represents a normalization operation to stabilize feature magnitudes and \\(\\epsilon\\) is a temperature parameter that controls the sharpness of the alignment distribution." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.527, + 0.482, + 0.572 + ], + "angle": 0, + "content": "Note that \\(\\mathcal{L}^{\\mathrm{align}}\\) is computed only when motion and video are paired. For standalone motion or video inputs, this loss term is omitted for applicability." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.573, + 0.482, + 0.601 + ], + "angle": 0, + "content": "The overall human motion concept discovery objective is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.611, + 0.482, + 0.649 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} ^ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {m o t i o n}} ^ {\\text {r e c}} + \\mathcal {L} _ {\\text {v i d e o}} ^ {\\text {r e d e}} + \\\\ \\lambda^ {\\text {d i s}} \\mathcal {L} ^ {\\text {d i s}} + \\lambda^ {\\text {a c t}} \\mathcal {L} ^ {\\text {a c t}} + \\lambda^ {\\text {a l i g n}} \\mathcal {L} ^ {\\text {a l i g n}}, \\tag {12} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.658, + 0.482, + 0.689 + ], + "angle": 0, + "content": "where \\(\\lambda^{\\mathrm{dis}}\\), \\(\\lambda^{\\mathrm{act}}\\) and \\(\\lambda^{\\mathrm{align}}\\) are weighting coefficients for the respective losses \\(\\mathcal{L}^{\\mathrm{dis}}\\), \\(\\mathcal{L}^{\\mathrm{act}}\\) and \\(\\mathcal{L}^{\\mathrm{align}}\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.698, + 0.29, + 0.714 + ], + "angle": 0, + "content": "3.3. Modality Translation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.482, + 0.795 + ], + "angle": 0, + "content": "Although pre-trained encoders capture rich features from the original input, these features are often difficult for LLMs to interpret directly [5, 37, 40]. Therefore, it is essential to introduce an additional projection layer to map these features into a space that LLMs can effectively understand." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.482, + 0.901 + ], + "angle": 0, + "content": "As illustrated in Fig 2, we add tunable translation layers after the encoders \\(\\mathcal{E}^u\\) and \\(\\mathcal{E}^m\\). These layers are trained using annotated video and motion sequence data to learn task-specific transformations. During this stage, the encoders and the LLM \\(\\mathcal{F}\\) are kept frozen to prevent adverse feature drift, and only the weights of the translation layers are updated. This strategy preserves the pre-trained knowledge of" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "the encoders and language model while optimizing the feature mapping for downstream tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.132, + 0.818, + 0.149 + ], + "angle": 0, + "content": "3.4. Multi-Modality Instruction Tuning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.155, + 0.905, + 0.244 + ], + "angle": 0, + "content": "Finally, we employ instruction tuning to train the language model \\(\\mathcal{F}\\) to respond to user queries using the corresponding video or motion sequences as context. During this process, the encoders and projection layers remain frozen, ensuring their pre-trained features are retained, while the weights of the language model \\(\\mathcal{F}\\) are updated to accommodate the task." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.246, + 0.905, + 0.322 + ], + "angle": 0, + "content": "In this stage, the model can further enhance cross-modal alignment using paired motion-video data in an implicit manner. To streamline the fine-tuning process, we adopt LoRA [29] with a rank of 8, allowing efficient and lightweight adjustment to the language model." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.336, + 0.645, + 0.352 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.36, + 0.714, + 0.377 + ], + "angle": 0, + "content": "4.1. Experiments Settings" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.383, + 0.905, + 0.624 + ], + "angle": 0, + "content": "Implementation details For videos and motions, we initialize a codebook with 512 codes, a one-layer projector for feature alignment, and a 2-layer MLP to generate the weights for a 2-layer discriminator separately. For video inputs, we utilize pre-trained SigLIP [79] with an image patch size of 16 pixels and hidden dimension of 256 as the video encoder, and we take a 12-layer transformer as the reconstruction decoder. A 4-layer transformer is utilized as the motion encoder and a 2-layer transformer is used for motion reconstruction. For the velocity reconstruction, we initialize a two-layer hypernet and two-layer velocity decoder. Besides, we utilize one-layer MLPs for feature alignment. We pre-train the motion encoder with only the reconstruction loss for 60K iterations with a learning rate of 1e-4. Then, we train both the video and motion encoders with the total loss for 8K iterations with a learning rate of 1e-5." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.625, + 0.905, + 0.73 + ], + "angle": 0, + "content": "We utilize Llama3.1-8B as our base LLM model. For the modality translation stage, we initialize a one-layer projector for the video and motion encoder separately. We train the projectors on the annotation dataset for one batch with a batch size of 32 and a learning rate of 1e-4. In the Multimodality instruction tuning stage, we train the LLM model with a batch size of 16 and a learning rate of 3e-4." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Training datasets For the encoder pre-training stage, we utilize Motion-X [41] datasets which contain 81K motion sequences and 36K motion-video pairs. In order to facilitate the training process of LLMs, we follow MotionLLM [5] to prompt LLMs to generate 20K annotation for motion-video pairs from the Motion-X dataset, and 10K annotation data from the HumanML3D [21] dataset. This data is utilized for modality translation. We additionally utilize the VATEX [69] dataset which contains 20K annotated videos for video translator training. For multi-modality instruction" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7184" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.131 + ], + "angle": 0, + "content": "Table 1. Comparison of different methods on BABEL-QA test set. The pred type represent the prediction type, including classification and generation. For 2s-AGCN and MotionCLIP, 'M' means MLP, 'R' means RNN. Our method outperforms the baselines across all the metrics." + }, + { + "type": "table", + "bbox": [ + 0.157, + 0.142, + 0.838, + 0.27 + ], + "angle": 0, + "content": "
ModelPred typeOverallQuery typeRelation Type
ActionDirectionBodyPartBeforeAfterOther
2s-AGCN-Mcls.0.3550.3840.3520.2280.3310.2640.295
2s-AGCN-Rcls.0.3570.3960.3520.1940.3370.3010.285
MotionCLIP-Mcls.0.4300.4850.3610.2720.3720.3210.404
MotionCLIP-Rcls.0.4200.4890.3100.2500.3980.3140.387
MotionLLMgen.0.4360.5170.3540.1540.4270.3680.529
Oursgen.0.7110.8090.6970.6230.7070.6350.797
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.28, + 0.482, + 0.358 + ], + "angle": 0, + "content": "tuning, we follow MotionLLM to generate 200K Q&A pairs for both motions and videos. We also utilize the VideoChatGPT [49] and BABEL-QA [13] datasets to enhance the motion and video understanding ability of our model, which has 100K and 2K Q&A data separately." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.383, + 0.484, + 0.519 + ], + "angle": 0, + "content": "Evaluation datasets Although MotionLLM [5] creates a large benchmark for both motion and video understanding, by the time we completed our paper, they had not released the specific details or access methods for the benchmark, nor did they provide a clear timeline for its release. Therefore, we did not perform comparisons on that benchmark but instead selected the test set of ActivityNet-QA [78] and BABEL-QA [13] to separately validate the performance of our algorithm on video and motion data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.546, + 0.484, + 0.713 + ], + "angle": 0, + "content": "Evaluation metrics On the BABEL-QA benchmark, we evaluate the prediction accuracy for evaluation. As for the classification-based baselines, we evaluate the accuracy directly. As for the generative baselines, we adapt the evaluation protocol used in previous works [5, 37] by utilizing LLMs. Specifically, the evaluation LLM is given some examples of questions, answers, and scores (from 0 to 1), and it is required to follow these examples and score the experiment results. On the Activity-QA benchmark, we only compare our methods with our generative baselines. Thus, we directly utilize LLMs to evaluate the methods." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.727, + 0.285, + 0.743 + ], + "angle": 0, + "content": "4.2. Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Experiments on ActivityNet-QA. We evaluate the video understanding ability of our method and other baselines on the ActivityNet-QA benchmark as shown in Tab. 2. We follow previous methods [5, 33, 38] to report the accuracy and scores that are evaluated by LLMs. We show in Tab. 2 that our method outperforms all the baselines on both metrics. Compared to the video-only method, our method aligns the video feature with the motion feature, making the encoder concentrate on human-centric information. Compared to MotionLLM, our encoder pre-training helps the network to" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.281, + 0.905, + 0.311 + ], + "angle": 0, + "content": "learn more fine-grained information through velocity reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.33, + 0.907, + 0.512 + ], + "angle": 0, + "content": "Experiments on BABEL-QA. We evaluate the motion understanding ability of our method as compared to existing studies on the BABEL-QA [13] dataset. Following [5, 13], we compare our method with a diverse set of baselines: 1) 2s-AGCN [58], a method using 2s-GCN to extract motion features and classify motion by MLP or RNN; 2) Motion-CLIP [63], a method that aligns the features of motion and text using CLIP [55] and predicts motion class by MLP or RNN; 3) MotionLLM [5], an LLM-based method which has motion reasoning ability. We follow [5] to fine-tune our method with the training set of BABEL-QA and evaluate our method with LLMs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.512, + 0.909, + 0.694 + ], + "angle": 0, + "content": "From Tab. 1, our method outperforms other methods in all metrics by a large margin of nearly \\(60\\%\\) overall. In the \"bodypart\" query task, our algorithm achieves 2.29 times the score of the second-place method, MotionCLIP-M (0.623 vs. 0.272). This advantage stems partly from our velocity-based encoder training, which, compared to traditional motion feature extraction, helps us better capture high-frequency information and ensures the learning of fine-grained features. Additionally, unlike MotionLLM, we perform alignment at the encoder pre-training stage, ensuring that the features learned by the encoder are inherently more semantically rich." + }, + { + "type": "table_caption", + "bbox": [ + 0.52, + 0.717, + 0.9, + 0.731 + ], + "angle": 0, + "content": "Table 2. Experiments results on ActivityNet-QA benchmark." + }, + { + "type": "table", + "bbox": [ + 0.556, + 0.742, + 0.866, + 0.899 + ], + "angle": 0, + "content": "
ModelAcc↑Score ↑
FrozenBiLM [57]24.7-
VideoChat [38]-2.2
LLaMA-Adapter [84]34.22.7
Video-LLaMA [80]12.41.1
Video-ChatGPT [49]35.22.7
Video-LLaVA [40]45.33.3
VideoChat2 [39]49.13.3
MotionLLM [5]53.33.5
Ours54.23.6
" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7185" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.356, + 0.09, + 0.644, + 0.104 + ], + "angle": 0, + "content": "Table 3. Ablation study on BABEL-QA test set." + }, + { + "type": "table", + "bbox": [ + 0.142, + 0.114, + 0.855, + 0.226 + ], + "angle": 0, + "content": "
ModelPred typeOverallQuery typeRelation Type
ActionDirectionBodyPartBeforeAfterOther
MotionLLMgen.0.4360.5170.3540.1540.4270.3680.529
Ours-w/oLrecgen.0.6960.7410.6450.5770.6000.5970.762
Ours-w/oLdis & Lactgen.0.6370.6930.4780.6060.6670.5260.709
Ours-w/oLaligngen.0.6750.7430.5790.5230.5840.5700.743
Oursgen.0.7110.8090.6970.6230.7070.6350.797
" + }, + { + "type": "image", + "bbox": [ + 0.141, + 0.228, + 0.856, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.473, + 0.907, + 0.517 + ], + "angle": 0, + "content": "Figure 4. Example of motion understanding. We validate the proposed model's ability to understand motion from multiple aspects. Q1 on the left and Q1-3 on the right evaluate its comprehension of motion sequences, while Q2 and Q3 on the left push it to analyze kinematic, kinesthetic, and physical properties. These results demonstrate our algorithm's ability to recognize, describe in detail, and analyze motions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.529, + 0.486, + 0.741 + ], + "angle": 0, + "content": "Ablation study. We present the ablation results in Tab. 3. We conducted ablation studies on masked reconstruction, velocity reconstruction, and feature alignment. Motion-LLM serves as an ablation baseline without any of these techniques. As presented, the absence of any single module leads to a decrease in performance. Among these, masked reconstruction has the least impact, as we initially train the motion encoder solely with masked reconstruction loss to ensure stable subsequent training. Removing velocity reconstruction has the greatest impact on pexrformance since velocity provides high-frequency information that other modules lack. Likewise, removing the alignment loss significantly affects results, highlighting the importance of multi-modality alignment in motion understanding." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.756, + 0.274, + 0.772 + ], + "angle": 0, + "content": "4.3. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.485, + 0.903 + ], + "angle": 0, + "content": "To validate our algorithm's ability to understand motions and its contextual knowledge, we conducted multiple rounds of Q&A with the motion model. We present two examples in Fig. 4. We discover that, beyond its basic motion recognition capabilities, our algorithm can provide detailed descriptions of motions. In the example on the right, it can precisely describe the actions of the left hand and foot, as well as the right hand and foot. Our algorithm also" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.529, + 0.907, + 0.635 + ], + "angle": 0, + "content": "possesses extensive knowledge about various movements. For instance, when asked, \"what muscles does the person engage when bending their elbows during a push-up?\" it can accurately respond with \"triceps muscles.\" This demonstrates that our algorithm can understand the input motion sequences and perform corresponding analyses. We present more qualitative results in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.648, + 0.634, + 0.663 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.673, + 0.909, + 0.871 + ], + "angle": 0, + "content": "In this paper, we presented HuMoCon, a novel motion-video understanding framework for human action understanding and concept discovery. We demonstrated the effectiveness of our approach through comprehensive experiments, showcasing major improvements over state-of-the-art methods. Our work provides a strong foundation for future research in human action understanding and motion concept discovery. There are still some drawbacks of our work, such as the lack of human-environment contact or understanding niche motions. We plan to address those limitations by expanding the framework to incorporate additional modalities, particularly as environments, objects, and contact, and curate a wider range of specialized datasets." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7186" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.092, + 0.232, + 0.108 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.113, + 0.486, + 0.173 + ], + "angle": 0, + "content": "This work is supported by the Early Career Scheme of the Research Grants Council (grant # 27207224), the HKU-100 Award, a donation from the Musketeers Foundation, and an Academic Gift from Meta." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.187, + 0.188, + 0.203 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.212, + 0.484, + 0.265 + ], + "angle": 0, + "content": "[1] Nikos Athanasiou, Mathis Petrovich, Michael J Black, and Gúl Varol. Teach: Temporal action composition for 3d humans. In 2022 International Conference on 3D Vision (3DV), pages 414-423. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.268, + 0.484, + 0.337 + ], + "angle": 0, + "content": "[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 1(2):3, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.339, + 0.483, + 0.407 + ], + "angle": 0, + "content": "[3] Yang Bai, Desen Zhou, Songyang Zhang, Jian Wang, Errui Ding, Yu Guan, Yang Long, and Jingdong Wang. Action quality assessment with temporal parsing transformer. In European conference on computer vision, pages 422-438. Springer, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.409, + 0.483, + 0.464 + ], + "angle": 0, + "content": "[4] Pablo Cervantes, Yusuke Sekikawa, Ikuro Sato, and Koichi Shinoda. Implicit neural representations for variable length human motion generation. In European Conference on Computer Vision, pages 356-372. Springer, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.466, + 0.484, + 0.532 + ], + "angle": 0, + "content": "[5] Ling-Hao Chen, Shunlin Lu, Ailing Zeng, Hao Zhang, Benyou Wang, Ruimao Zhang, and Lei Zhang. Motionllm: Understanding human behaviors from human motions and videos. arXiv preprint arXiv:2405.20340, 2024. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.536, + 0.484, + 0.576 + ], + "angle": 0, + "content": "[6] Steven Chen and Richard R Yang. Pose trainer: correcting exercise posture using pose estimation. arXiv preprint arXiv:2006.11718, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.579, + 0.484, + 0.633 + ], + "angle": 0, + "content": "[7] Sihan Chen, Xingjian He, Longteng Guo, Xinxin Zhu, Weining Wang, Jinhui Tang, and Jing Liu. Valor: Vision-audiolanguage omni-perception pretraining model and dataset. arXiv preprint arXiv:2304.08345, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.635, + 0.484, + 0.703 + ], + "angle": 0, + "content": "[8] Sihan Chen, Handong Li, Qunbo Wang, Zijia Zhao, Mingzhen Sun, Xinxin Zhu, and Jing Liu. Vast: A vision-audio-subtitle-text omni-modality foundation model and dataset. Advances in Neural Information Processing Systems, 36:72842-72866, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.705, + 0.484, + 0.773 + ], + "angle": 0, + "content": "[9] Ginger Delmas, Philippe Weinzaepfel, Francesc Moreno-Noguer, and Grégory Rogez. Posefix: correcting 3d human poses with natural language. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15018-15028, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.776, + 0.484, + 0.844 + ], + "angle": 0, + "content": "[10] Piotr Dólár, Vincent Rabaud, Garrison Cottrell, and Serge Belongie. Behavior recognition via sparse spatio-temporal features. In 2005 IEEE international workshop on visual surveillance and performance evaluation of tracking and surveillance, pages 65-72. IEEE, 2005. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[11] Yong Du, Wei Wang, and Liang Wang. Hierarchical recurrent neural network for skeleton based action recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1110-1118, 2015. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.212, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.161 + ], + "angle": 0, + "content": "[12] Benjamin Elizalde, Soham Deshmukh, Mahmoud Al Ismail, and Huaming Wang. Clap learning audio concepts from natural language supervision. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.163, + 0.907, + 0.217 + ], + "angle": 0, + "content": "[13] Mark Endo, Joy Hsu, Jiaman Li, and Jiajun Wu. Motion question answering via modular motion programs. In International Conference on Machine Learning, pages 9312-9328. PMLR, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.219, + 0.906, + 0.288 + ], + "angle": 0, + "content": "[14] Qihang Fang, Yingda Yin, Qingnan Fan, Fei Xia, Siyan Dong, Sheng Wang, Jue Wang, Leonidas J Guibas, and Baoquan Chen. Towards accurate active camera localization. In European Conference on Computer Vision, pages 122-139. Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.289, + 0.906, + 0.354 + ], + "angle": 0, + "content": "[15] Qihang Fang, Chengcheng Tang, Bugra Tekin, and Yanchao Yang. Cigtime: Corrective instruction generation through inverse motion editing. In The Thirty-Eighth Annual Conference on Neural Information Processing Systems, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.357, + 0.906, + 0.413 + ], + "angle": 0, + "content": "[16] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.906, + 0.481 + ], + "angle": 0, + "content": "[17] Yao Feng, Jing Lin, Sai Kumar Dwivedi, Yu Sun, Priyanka Patel, and Michael J Black. Chatpose: Chatting about 3d human pose. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2093-2103, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.484, + 0.906, + 0.552 + ], + "angle": 0, + "content": "[18] Mihai Fieraru, Mihai Zanfir, Silviu Cristian Pirlela, Vlad Olaru, and Cristian Sminchisescu. Aifit: Automatic 3d human-interpretable feedback models for fitness training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9919–9928, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.554, + 0.906, + 0.633 + ], + "angle": 0, + "content": "[19] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15180-15190, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.637, + 0.906, + 0.705 + ], + "angle": 0, + "content": "[20] Chuan Guo, Xinxin Zuo, Sen Wang, Shihao Zou, Qingyao Sun, Annan Deng, Minglun Gong, and Li Cheng. Action2motion: Conditioned generation of 3d human motions. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2021-2029, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.707, + 0.906, + 0.775 + ], + "angle": 0, + "content": "[21] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5152-5161, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.906, + 0.845 + ], + "angle": 0, + "content": "[22] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5152-5161, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[23] Chuan Guo, Xinxin Zuo, Sen Wang, and Li Cheng. Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In European Conference on Computer Vision, pages 580-597. Springer, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7187" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[24] Tianyu Guo, Hong Liu, Zhan Chen, Mengyuan Liu, Tao Wang, and Runwei Ding. Contrastive learning from extremely augmented skeleton sequences for self-supervised action recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 762-770, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.162, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[25] Tengda Han, Max Bain, Arsha Nagrani, Gul Varol, Weidi Xie, and Andrew Zisserman. Autoad ii: The sequel-who, when, and what in movie audio description. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13645–13655, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.482, + 0.3 + ], + "angle": 0, + "content": "[26] Tengda Han, Max Bain, Arsha Nagrani, Gül Varol, Weidi Xie, and Andrew Zisserman. Autoad iii: The prequel-back to the pixels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18164-18174, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.301, + 0.482, + 0.37 + ], + "angle": 0, + "content": "[27] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.371, + 0.482, + 0.412 + ], + "angle": 0, + "content": "[28] Daniel Holden, Taku Komura, and Jun Saito. Phase-functional neural networks for character control. ACM Transactions on Graphics (TOG), 36(4):1-13, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.413, + 0.482, + 0.468 + ], + "angle": 0, + "content": "[29] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.482, + 0.523 + ], + "angle": 0, + "content": "[30] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. Advances in Neural Information Processing Systems, 36:20067-20079, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.525, + 0.482, + 0.593 + ], + "angle": 0, + "content": "[31] Hanwen Jiang, Shaowei Liu, Jiashun Wang, and Xiaolong Wang. Hand-object contact consistency reasoning for human grasps generation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11107-11116, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.482, + 0.663 + ], + "angle": 0, + "content": "[32] Yujiao Jiang, Qingmin Liao, Zhaolong Wang, Xiangru Lin, Zongqing Lu, Yuxi Zhao, Hanqing Wei, Jingrui Ye, Yu Zhang, and Zhijing Shao. Smplx-lite: A realistic and drivable avatar benchmark with rich geometry and texture annotations. arXiv preprint arXiv:2405.19609, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.664, + 0.482, + 0.746 + ], + "angle": 0, + "content": "[33] Peng Jin, Ryuichi Takanobu, Wancai Zhang, Xiaochun Cao, and Li Yuan. Chat-univi: Unified visual representation empowers large language models with image and video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13700-13710, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.482, + 0.816 + ], + "angle": 0, + "content": "[34] Qiuhong Ke, Mohammed Bennamoun, Senjian An, Ferdous Sohel, and Farid Boussaid. A new representation of skeleton sequences for 3d action recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3288-3297, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[35] Jihoon Kim, Jiseob Kim, and Sungjoon Choi. Flame: Freeform language-based motion synthesis & editing. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8255-8263, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[36] Kosmas Kritsis, Aggelos Gkiokas, Aggelos Pikrakis, and Vassilis Katsouros. Danceconv: Dance motion genera" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.092, + 0.905, + 0.12 + ], + "angle": 0, + "content": "tion with convolutional networks. IEEE Access, 10:44982-45000, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[37] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[38] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.905, + 0.317 + ], + "angle": 0, + "content": "[39] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mybench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.319, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[40] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[41] Jing Lin, Ailing Zeng, Shunlin Lu, Yuanhao Cai, Ruimao Zhang, Haoqian Wang, and Lei Zhang. Motion-x: A large-scale 3d expressive whole-body human motion dataset. Advances in Neural Information Processing Systems, 36, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.447, + 0.905, + 0.516 + ], + "angle": 0, + "content": "[42] Lilang Lin, Jiahang Zhang, and Jiaying Liu. Actionlet-dependent contrastive learning for unsupervised skeleton-based action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2363-2372, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.518, + 0.905, + 0.587 + ], + "angle": 0, + "content": "[43] Daochang Liu, Qiyue Li, Tingting Jiang, Yizhou Wang, Rulin Miao, Fei Shan, and Ziyu Li. Towards unified surgical skill assessment. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9522-9531, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.589, + 0.905, + 0.671 + ], + "angle": 0, + "content": "[44] Jun Liu, Amir Shahroudy, Dong Xu, and Gang Wang. Spatio-temporal LSTM with trust gates for 3d human action recognition. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 816-833. Springer, 2016. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.673, + 0.905, + 0.715 + ], + "angle": 0, + "content": "[45] Mengyuan Liu, Hong Liu, and Chen Chen. Enhanced skeleton visualization for view invariant human action recognition. Pattern Recognition, 68:346-362, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.716, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[46] Ruizhe Liu, Qian Luo, and Yanchao Yang. Infocon: Concept discovery with generative and discriminative informativeness. arXiv preprint arXiv:2404.10606, 2024. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.759, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[47] Thomas Lucas, Fabien Baradel, Philippe Weinzaepfel, and Grégory Rogez. Posegpt: Quantization-based 3d human motion generation and forecasting. In European Conference on Computer Vision, pages 417-435. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[48] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing, 508:293-304, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.872, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[49] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7188" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.12 + ], + "angle": 0, + "content": "understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.191 + ], + "angle": 0, + "content": "[50] Shehan Munasinghe, Rusiru Thushara, Muhammad Maaz, Hanoona Abdul Rasheed, Salman Khan, Mubarak Shah, and Fahad Khan. Pg-video-llava: Pixel grounding large video-language models. arXiv preprint arXiv:2311.13435, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.484, + 0.262 + ], + "angle": 0, + "content": "[51] Paritosh Parmar and Brendan Tran Morris. What and how well you performed? a multitask learning approach to action quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 304-313, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.484, + 0.331 + ], + "angle": 0, + "content": "[52] Mathis Petrovich, Michael J Black, and Gül Varol. Action-conditioned 3d human motion synthesis with transformer vae. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10985-10995, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.484, + 0.39 + ], + "angle": 0, + "content": "[53] Mathis Petrovich, Michael J Black, and Gül Varol. Temos: Generating diverse human motions from textual descriptions. In European Conference on Computer Vision, pages 480-497. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.392, + 0.484, + 0.474 + ], + "angle": 0, + "content": "[54] Abhinanda R Punnakkal, Arjun Chandrasekaran, Nikos Athanasiou, Alejandra Quiros-Ramirez, and Michael J Black. Babel: Bodies, action and behavior with english labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 722-731, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.484, + 0.559 + ], + "angle": 0, + "content": "[55] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.484, + 0.643 + ], + "angle": 0, + "content": "[56] Jathushan Rajasegaran, Georgios Pavlakos, Angjoo Kanazawa, Christoph Feichtenhofer, and Jitendra Malik. On the benefits of 3d pose and tracking for human action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 640-649, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.646, + 0.484, + 0.701 + ], + "angle": 0, + "content": "[57] Leonard Salewski, Stefan Fauth, A Koepke, and Zeynep Akata. Zero-shot audio captioning with audio-language model guidance and audio context keywords. arXiv preprint arXiv:2311.08396, 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.703, + 0.484, + 0.771 + ], + "angle": 0, + "content": "[58] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Two-stream adaptive graph convolutional networks for skeleton-based action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12026-12035, 2019. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.773, + 0.484, + 0.814 + ], + "angle": 0, + "content": "[59] Sebastian Starke, He Zhang, Taku Komura, and Jun Saito. Neural state machine for character-scene interactions. ACM Transactions on Graphics, 38(6):178, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.484, + 0.871 + ], + "angle": 0, + "content": "[60] Sebastian Starke, Yiwei Zhao, Taku Komura, and Kazi Zaman. Local motion phases for learning multi-contact character movements. ACM Transactions on Graphics (TOG), 39 (4):54-1, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[61] Sebastian Starke, Ian Mason, and Taku Komura. Deeppphase: Periodic autoencoders for learning motion phase manifolds." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "ACM Transactions on Graphics (TOG), 41(4):1-13, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[62] Graham W Taylor, Rob Fergus, Yann LeCun, and Christoph Bregler. Convolutional learning of spatio-temporal features. In Computer Vision-ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part VI 11, pages 140-153. Springer, 2010. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.263 + ], + "angle": 0, + "content": "[63] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. In European Conference on Computer Vision, pages 358–374. Springer, 2022. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.905, + 0.305 + ], + "angle": 0, + "content": "[64] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.307, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[65] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 4489-4497, 2015. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.905, + 0.446 + ], + "angle": 0, + "content": "[66] Raviteja Vemulapalli, Felipe Arrate, and Rama Chellappa. Human action recognition by representing 3d skeletons as points in a lie group. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 588-595, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.449, + 0.905, + 0.503 + ], + "angle": 0, + "content": "[67] Heng Wang and Cordelia Schmid. Action recognition with improved trajectories. In Proceedings of the IEEE international conference on computer vision, pages 3551-3558, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.505, + 0.905, + 0.575 + ], + "angle": 0, + "content": "[68] Jingbo Wang, Ye Yuan, Zhengyi Luo, Kevin Xie, Dahua Lin, Umar Iqbal, Sanja Fidler, and Sameh Khamis. Learning human dynamics in autonomous driving scenarios. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20796-20806, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.577, + 0.905, + 0.646 + ], + "angle": 0, + "content": "[69] Xin Wang, Jiawei Wu, Junkun Chen, Lei Li, Yuan-Fang Wang, and William Yang Wang. Vatex: A large-scale, high-quality multilingual dataset for video-and-language research. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4581-4591, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.647, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[70] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, et al. Internvideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.905, + 0.786 + ], + "angle": 0, + "content": "[71] Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Guo Chen, Baoqi Pei, Rongkun Zheng, Jilan Xu, Zun Wang, et al. Intervideo2: Scaling video foundation models for multimodal video understanding. arXiv preprint arXiv:2403.15377, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.788, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[72] Zeqi Xiao, Tai Wang, Jingbo Wang, Jinkun Cao, Wenwei Zhang, Bo Dai, Dahua Lin, and Jiangmiao Pang. Unified human-scene interaction via prompted chain-of-contacts. arXiv preprint arXiv:2309.07918, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[73] Junyu Xie, Tengda Han, Max Bain, Arsha Nagrani, Güll Varol, Weidi Xie, and Andrew Zisserman. Autoad-zero: A training-free framework for zero-shot audio description. arXiv preprint arXiv:2407.15850, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7189" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.148 + ], + "angle": 0, + "content": "[74] Guo-Hao Xu, Yi-Lin Wei, Dian Zheng, Xiao-Ming Wu, and Wei-Shi Zheng. Dexterous grasp transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17933-17942, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[75] Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. Pllava: Parameter-free llava extension from images to videos for video dense captioning. arXiv preprint arXiv:2404.16994, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.483, + 0.262 + ], + "angle": 0, + "content": "[76] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In Proceedings of the AAAI conference on artificial intelligence, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.483, + 0.292 + ], + "angle": 0, + "content": "[77] Zeshi Yang, Kangkang Yin, and Libin Liu. Learning to use chopsticks in diverse gripping styles. 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.293, + 0.483, + 0.348 + ], + "angle": 0, + "content": "[78] Zhou Yu, Dejing Xu, Jun Yu, Ting Yu, Zhou Zhao, Yueting Zhuang, and Dacheng Tao. Activitynet-qa: A dataset for understanding complex web videos via question answering. In AAAI, pages 9127–9134, 2019. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.483, + 0.405 + ], + "angle": 0, + "content": "[79] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.483, + 0.448 + ], + "angle": 0, + "content": "[80] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.449, + 0.483, + 0.517 + ], + "angle": 0, + "content": "[81] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Shaoli Huang, Yong Zhang, Hongwei Zhao, Hongtao Lu, and Xi Shen. T2m-gpt: Generating human motion from textual descriptions with discrete representations. arXiv preprint arXiv:2301.06052, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.483, + 0.589 + ], + "angle": 0, + "content": "[82] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.591, + 0.483, + 0.66 + ], + "angle": 0, + "content": "[83] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xupeng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. Pointclip: Point cloud understanding by clip. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8552-8562, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.483, + 0.731 + ], + "angle": 0, + "content": "[84] Renrui Zhang, Jiaming Han, Chris Liu, Peng Gao, Aojun Zhou, Xiangfei Hu, Shilin Yan, Pan Lu, Hongsheng Li, and Yu Qiao. Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.483, + 0.802 + ], + "angle": 0, + "content": "[85] Xingyi Zhou, Anurag Arnab, Shyamal Buch, Shen Yan, Austin Myers, Xuehan Xiong, Arsha Nagrani, and Cordelia Schmid. Streaming dense video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18243-18252, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[86] Zixiang Zhou, Yu Wan, and Baoyuan Wang. Avatargpt: All-in-one framework for motion understanding planning generation and beyond. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1357-1366, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.872, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[87] Bin Zhu, Bin Lin, Munan Ning, Yang Yan, Jiaxi Cui, HongFa Wang, Yatian Pang, Wenhao Jiang, Junwu Zhang, Zongwei" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.547, + 0.092, + 0.905, + 0.135 + ], + "angle": 0, + "content": "Li, et al. Languagebind: Extending video-language pretraining to n-modality by language-based semantic alignment. arXiv preprint arXiv:2310.01852, 2023. 2, 3, 5" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "7190" + } + ] +] \ No newline at end of file diff --git a/2025/HuMoCon_ Concept Discovery for Human Motion Understanding/b5dd425e-4f3e-4b2d-9c88-36e4f0bafd85_origin.pdf b/2025/HuMoCon_ Concept Discovery for Human Motion Understanding/b5dd425e-4f3e-4b2d-9c88-36e4f0bafd85_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0ae1fdbf099cad7185ba984e42e9a4ce867ad76e --- /dev/null +++ b/2025/HuMoCon_ Concept Discovery for Human Motion Understanding/b5dd425e-4f3e-4b2d-9c88-36e4f0bafd85_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:00d517a7e2275734c1dfaf1d95b9918379170e84cafad787aa3793905d084bc7 +size 3430840 diff --git a/2025/HuMoCon_ Concept Discovery for Human Motion Understanding/full.md b/2025/HuMoCon_ Concept Discovery for Human Motion Understanding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a4ec688abcade73e591dff559d50968ddd6761fa --- /dev/null +++ b/2025/HuMoCon_ Concept Discovery for Human Motion Understanding/full.md @@ -0,0 +1,368 @@ +# HuMoCon: Concept Discovery for Human Motion Understanding + +Qihang Fang $^{1,2}$ + +Chengcheng Tang + +1The University of Hong Kong + +Shugao Ma2 + +Yanchao Yang + +$^{2}$ Meta + +qihfang@gmail.com, {chengcheng.tang,bugratekin,shugao}@meta.com, yanchaoy@hku.hk + +![](images/897542e4c766ca6d2895097d7377ad1855886fce72dc2242c7b2d4bd4dd03f85.jpg) +Figure 1. We present HuMoCon for concept discovery and human motion understanding with video or motion input. To address the challenge of feature misalignment and high-frequency information loss, we propose a novel feature alignment strategy and an advanced masked auto-encoder reconstructing velocity. HuMoCon empowers effective motion concept discovery and accurate Question Answering tasks, significantly outperforming state-of-the-art methods qualitatively and quantitatively. + +![](images/97026424ca8d3a63ed694ff5fda914904f300cc2f49c155f651d897df1ae9167.jpg) + +# Abstract + +We present HuMoCon, a novel motion-video understanding framework designed for advanced human behavior analysis. The core of our method is a human motion concept discovery framework that efficiently trains multi-modal encoders to extract semantically meaningful and generalizable features. HuMoCon addresses key challenges in motion concept discovery for understanding and reasoning, including the lack of explicit multi-modality feature alignment and the loss of high-frequency information in masked autoencoding frameworks. Our approach integrates a feature alignment strategy that leverages video for contextual understanding and motion for fine-grained interaction modeling, further with a velocity reconstruction mechanism to enhance high-frequency feature expression and mitigate temporal over-smoothing. Comprehensive experiments on standard benchmarks demonstrate that HuMoCon enables effective motion concept discovery and significantly outperforms state-of-the-art methods in training large models for human motion understanding. We will open-source the associated code with our paper. + +# 1. Introduction + +Understanding human behavior is crucial for gaining deeper insights into human actions and reasoning, which are foun + +ditional for developing human-centered AI systems [5, 26, 37, 61] that effectively address real-world tasks [14, 68, 72]. Recent advancements in large language models (LLMs) have led to significant progress in video- and motion-based methods [5, 15, 37, 40, 75] for analyzing human behavior. Despite these advancements, achieving a precise and fine-grained understanding of human actions remains a persistent and complex challenge. + +Human behavior is often represented as motion sequences structured based on the kinematics of the human skeleton [32?]. Traditional approaches to analyzing motion sequences have predominantly focused on classification [11, 24, 34, 42, 44, 45, 58, 76], relying on predefined action categories. While effective for specific tasks, these methods often lack the flexibility for deeper or more nuanced analysis. Other methods employ statistical models to assess motion quality [3, 6, 18, 43, 51], but they require significant expert knowledge, limiting scalability. + +Recent advancements in language models [40, 50, 75] have opened new possibilities in motion analysis, enabling tasks such as motion generation and motion captioning based on free-text inputs [4, 20, 52, 59, 61], advancing the field of motion analysis. However, obtaining high-quality motion sequences and annotations remains costly, and motion sequences generally lack information about the surrounding environment, limiting models' ability to understand human-environment interactions. + +In contrast, video data offers rich visual context and is easier to collect. It has been widely used for action classification [16, 62, 65] and video Q&A [5, 37, 40]. However, video data often contains irrelevant elements that can hinder accurate content analysis. For example, a model might learn from training data that most running occurs in stadiums, leading it to mistakenly classify other stadium activities as running during testing. + +In recent years, significant progress has been made in developing multi-modality models [5, 87]. Integrating multiple modalities, as opposed to relying on a single modality, enhances the network's ability to learn fine-grained and semantic features. For instance, LART [56] combines video and motion data to ensure that video-based action classification focuses on the human body rather than overemphasizing environmental context. MotionLLM [5] expands large language models (LLMs) to video and motion, building a model that can understand both. MotionLLM demonstrated that learning from video and motion improves the network's ability to analyze human behavior. However, MotionLLM implicitly aligns video and motion using motion-video pairs during LLM training without optimizing the encoder for explicit cross-modal alignment. To address this limitation and further enhance feature representation, we propose explicitly aligning features from different modalities during the encoding process within a human motion concept discovery framework. + +Achieving effective cross-modal alignment and feature representation relies heavily on robust encoder training methods. Many existing studies [70, 71, 87] adopt a masked autoencoder framework [27] when training encoders. This approach involves applying a mask to a certain proportion of the input, extracting features from the masked data using the encoder, and reconstructing the original unmasked input with a decoder. While effective in learning high-quality features, this masking strategy inherently causes the loss of high-frequency information, resulting in temporally over-smoothed reconstructions that hinder the LLM's ability to understand human behavior. [47] To address this issue, we propose reconstructing velocity as a means to preserve high-frequency information. Following InfoCon [46], we design a VQ-VAE-based loss function that helps retain the fine-grained features of the original masked autoencoder while further enhancing high-frequency feature expression. This approach significantly boosts the subsequent LLM's capacity to model and understand human motion effectively. + +In summary, we present HuMoCon, an effective human motion concept discovery pertaining framework for advanced human behavior understanding from both video and motion data. Our key contributions include: (i) a novel feature alignment strategy that explicitly leverages video data for contextual understanding and motion data for fine-grained interaction modeling, and (ii) an enhanced masked + +autoencoder framework incorporating velocity reconstruction to significantly improve high-frequency feature expression and mitigate temporal over-smoothing. To evaluate the effectiveness of our proposed method, we conduct experiments on the Activity-QA [78] and BABEL-QA [13] benchmarks [54], comparing our approach with the state-of-the-art methods. Comprehensive quantitative and qualitative evaluations demonstrate that HuMoCon achieves state-of-the-art performance on standard benchmarks for motion-video understanding. + +# 2. Related Work + +# 2.1. Human Motion Understanding + +The core of human motion understanding is to learn the semantics of motion sequences or videos. In the early stages, research in this field is often based on predefined action categories. Such work can be divided into two types: action classification [11, 24, 34, 42, 44, 45, 58, 76] and motion generation targeted at specific categories [4, 20, 28, 31, 36, 52, 59-61, 74, 77]. For action recognition, several works [11, 44] utilize Recurrent Neural Networks (RNN) to extract temporal information from human joint sequences, beyond the CNN-based ones [34, 45]. Subsequent works [24, 42, 58, 76] propose representing motion data as a graph structure, enabling the use of spatio-temporal graph convolutional networks (ST-GCNs) to effectively capture the connectivity and interactions between human joints. For motion generation, several studies [28, 31, 36, 60, 74, 77] have concentrated on generating specific, single actions. In contrast, other works [4, 20, 52, 59, 61] aim to generate a broader range of diverse action types, addressing the challenge of variability in motion synthesis. + +In recent years, more research has shifted towards captioning [23, 30], generation [1, 22, 23, 30, 35, 53, 64, 81, 82, 86], and analysis [5, 9, 15, 17] based on free-form text. Leveraging alignment of motion latent features with the CLIP [55] embedding space, MotionCLIP [63] enables the generation of out-of-distribution motions. Transformers [1, 22, 23, 53, 81] and diffusion [35, 64, 82] models are also utilized for text-conditioned motion generation. Moreover, MotionGPT [30] and AvatarGPT [86] utilize large language models (LLMs) to handle multiple motion-related tasks simultaneously. Despite these advancements, research focused specifically on motion analysis remains relatively limited. PoseFix [9] and CigTime [15] are proposed to annotate the corrective instructions from source to target for pose and motion sequences. MotionLLM [5] proposes a method to analyze both the video and motion sequences with detailed spatial-temporal awareness and reasoning abilities. However, MotionLLM relies on implicit alignment through paired data and lacks explicit cross-modal feature alignment. + +# 2.2. Video Understanding + +In addition to motion understanding, video understanding has emerged as a prominent research area, attracting extensive attention in recent years. Early works on video classification and segmentation relied on handcrafted features [10, 44, 66, 67] or convolutional neural networks [16, 62, 65]. Beyond classification, video analysis [3, 6, 18, 25, 26, 37, 40, 43, 50, 51, 57, 73, 75, 85] has become a highly active field of research. Action quality assessment has been the focus of several works [3, 6, 18, 43, 51], with methods such as Aifit [18] and PoseTrainer [6] proposed to correct user actions. These approaches evaluate actions in conjunction with complex textual instructions [3, 43, 51]. Other studies [25, 26, 57, 73, 85] aim to generate descriptive narratives for videos, broadening the scope of video understanding. Recently, many works [2, 37, 40, 50, 75, 80] have leveraged large language models (LLMs) to understand video content and perform complex reasoning tasks. These methods combine video processing with the advanced reasoning capabilities of LLMs, further advancing the state-of-the-art in video analysis. Video-LLaVA [40] build a model to perform visual reasoning capabilities on images and videos simultaneously. Follow-up works [37, 50, 75] extend Video-LLaVA for faster inference and more accurate reasoning ability. Unlike typical video understanding algorithms, which rely heavily on visual context alone, our approach integrates motion data directly, enhancing the model's ability to understand human-centric actions. + +# 2.3. Multi-modality Pre-training + +Multi-modality pre-training has primarily focused on aligning vision and language representations. CLIP [55] pioneered this effort by using contrastive learning to align image and text features on large-scale datasets. Subsequent works [12, 48, 63, 83] extended CLIP's framework to other modalities, including video [48], audio [12] and motion [63]. + +Recent studies, such as VALOR [7] and VAST [8], demonstrated that cross-modal alignment not only enhances model robustness but also maintains high performance across modalities. Building on this, follow-up works [19, 87] have aimed to expand the number of aligned modalities, further advancing multi-modality research. + +In the context of motion and video alignment, Motion-LLMs [5] is the first work to address this challenge. However, it relies solely on implicit alignment during the fine-tuning stage of large language models. + +# 3. Method + +# 3.1. Overview + +Our objective is to develop an LLM-based system capable of understanding and reasoning both human video and mo + +tion inputs. Given a question $Q$ and an input, which can be either a video $U$ or motion sequence $M$ , the model generates the corresponding answer $W$ . Specifically, a video $U$ is represented as a sequence of $T$ keyframes, denoted as $\{u_0, u_1, \dots, u_{T-1}\}$ . Similarly, a motion sequence $M$ is represented as $K$ consecutive pose frames, denoted as $\{m_0, m_1, \dots, m_{K-1}\}$ . + +We present an overview of the pipeline in Fig. 2. To enable effective multi-modal human motion understanding, our framework focuses on: (1) discovering human motion concepts from both video and motion data, (2) preserving high-frequency motion details using velocity reconstruction to reduce temporal smoothing, and (3) aligning video and motion features to enhance multi-modal comprehension. To accomplish this, we train our video and motion encoders with four key learning objectives: Masked Reconstruction, Discriminative Informativeness, Actionable Informativeness, and Feature Alignment. Consequently, we leverage the encoded features of videos and motions to facilitate the fine-tuning of LLMs for video-motion understanding and reasoning. The following sections detail each component of our approach, including multi-modality encoder pre-training, modality translation, and instruction tuning. + +# 3.2. Multi-Modality Encoder Pre-training via Human Motion Concept Discovery + +Pre-training encoders are critical for multi-modality large language models (LLMs) [70, 71, 87], as they enable the effective extraction of essential features from large-scale input data. To enhance the encoder's ability to capture key information, we design three distinct learning objectives: masked reconstruction, velocity reconstruction, and feature alignment, all contributing to the quality of the discovered human motion concepts as well as the pre-training efficiency. For both videos and motions, we adopt a similar network structure during encoder pre-training. Consequently, in the following section, we focus on detailing the network structure for video inputs. + +Encoding Function. In our approach, we design masked autoencoder architectures tailored for both video and motion data. Specifically, for videos, we utilize a pre-trained encoder $\mathcal{E}_u$ , to encode the input $U$ into a feature representation $f^u$ : + +$$ +F ^ {u} = \mathcal {E} _ {u} (U). \tag {1} +$$ + +Then, we employ a codebook $C^u$ , modeled with VQ-VAE to quantize $f^u$ into a discrete feature representation $d^u$ : + +$$ +D ^ {u} = \operatorname {Q u a n t i z e} \left(F ^ {u}, C ^ {u}\right), \tag {2} +$$ + +where 'Quantize' denotes the quantization operation, and $C^u$ is the codebook used for discretization, also serving as the discovered human motion concepts. + +![](images/e40073539a848eda21d5e8fb455fc0e97d59db9d6253a2473b5c1d203ac79cb9.jpg) +Figure 2. System overview of our method. (a) The encoder pre-training process for learning and aligning video and motion features and enhancing high-frequency details through velocity reconstruction. We utilize a VQ-VAE-based structure, and we design effective learning objectives to enhance the encoder to extract semantic meaningful and fine-grained features. (b) The fine-tuning of the large language model (LLM) for video and motion reasoning consists of two stages: Modality Translation and Multi-modality Instruction Tuning. In the Modality Translation stage, we train a translation layer for each modality to map the encoding feature to the LLM space. In the Instruction Tuning stage, we fine-tune the LLM to understand human motion and videos for downstream tasks. + +Masked Reconstruction. The masked autoencoder has proven to be highly effective for pre-training [27], as it facilitates learning of meaningful and low-frequency features from input data. Thus, we leverage masked auto-encoding to enhance the discovered concepts and facilitate the multimodal feature pre-training. + +More explicitly, we mask $D^u$ by a probability of $\sigma$ to obtain the masked feature $D_{\mathrm{mask}}^u$ : + +$$ +D _ {\text {m a s k}} ^ {u} = \operatorname {M a s k} \left(D ^ {u}, \sigma\right), \tag {3} +$$ + +where 'Mask' is the operation that applies masking to the features, and $\sigma < 1.0$ controls the masking ratio. A video decoder $D^{u}$ is used to reconstruct the video $U^{\prime}$ from $\mathcal{D}_{\mathrm{mask}}^{u}$ : + +$$ +U ^ {\prime} = \mathcal {D} ^ {u} \left(D _ {\text {m a s k}} ^ {u}\right). \tag {4} +$$ + +We compute the reconstruction loss between the original video $U$ and the reconstructed video $U^{\prime}$ as follows: + +$$ +\mathcal {L} _ {\text {v i d e o}} ^ {\text {r e c}} = \left| \left| U ^ {\prime} - U \right| \right| _ {2} ^ {2}. \tag {5} +$$ + +Similarly, we introduce a reconstruction loss for motion data, $\mathcal{L}_{\mathrm{motion}}^{\mathrm{rec}}$ , following the same process. The total reconstruction loss is defined as: + +$$ +\mathcal {L} ^ {\text {r e c}} = \mathcal {L} _ {\text {v i d e o}} ^ {\text {r e c}} + \mathcal {L} _ {\text {m o t i o n}} ^ {\text {r e c}}. \tag {6} +$$ + +The above masked reconstruction loss ensures that the encoded features contain sufficient information about the + +video or motion input, while learning away possible nuisances through masking and making the learning of the concepts (codebook) more focused on critical information. On the other hand, the discretization imposed by the concepts further enhances the meaningfulness and robustness of the pre-trained multi-modal (motion) features. + +Velocity Reconstruction. Reconstructing inputs from masked features inevitably results in the loss of high-frequency details, leading to temporally over-smoothed reconstruction outcomes and preventing the discovery of more nuanced yet useful motion concepts. + +To address this limitation, we propose an objective that further reconstructs the "velocity" of the input sequence. To be more specific, we define "state" as the encoding or feature of a frame in the video or motion sequence, while the "velocity" means the difference between consecutive "states" that captures high-frequency information by focusing on dynamic changes. For video and motion data, we denote the "states" as $U$ and $m$ , and the "velocity" as the optical flow $O$ (computed from image frames) and delta motion $\delta M$ , respectively. + +However, directly reconstructing velocity from the same set of masked features introduces difficulty in figuring out the dynamic information, given the above-mentioned oversmoothing effect in the encoded features. To mitigate this, we draw inspiration from the robotic manipulation literature [46] on action concept discovery and incorporate two + +![](images/b4033cb12be74bb6bf64101f5319199a9aa4a4861eb8fedd50f3c9f868542bd1.jpg) +Figure 3. Overview of the velocity reconstruction components. We build similar network structures for both video and motion, and present the video part in this figure. This module is composed of two learning objectives. 1) Discriminative informativeness (left) aims to improve the distinctiveness of encoded features by reducing representational ambiguity. 2) Actionable informativeness (right) focuses on reconstructing the velocity by leveraging gradient information from the discrimination hypernetwork. As for the video data, we employ optical flow as the representation of the velocity. + +tailored learning objectives - discriminative informativeness and actionable informativeness - to enhance the efficiency of representations and improve the quality of temporal dynamics in multi-modality pretaining, as shown in Fig. 3. + +Discriminative informativeness. The discriminative informativeness objective aims to improve the distinctiveness of encoded features by increasing the correlation between the discovered concepts and the corresponding state features. Specifically, we introduce a video hyper-network $\mathcal{H}^u$ , which takes discrete features $c_k^u$ from the codebook $C^u$ as input and generates a classification network. The hyper-network $\mathcal{H}^u$ (together with the generated classifier) is trained to determine whether an input state is compatible with its corresponding code class, which is represented by a confidence value $s_u$ (classification score) from 0 to 1, thereby enhancing the distinguishability and separation of each feature class in the representation space. + +The discriminative loss for video is defined as: + +$$ +\mathcal {L} _ {\text {v i d e o}} ^ {\text {d i s}} = \sum_ {k = 0} ^ {| C ^ {u} | - 1} \sum_ {i = 0} ^ {T - 1} C E \left(\mathcal {H} ^ {u} \left(c _ {k} ^ {u}, u _ {i}\right), d _ {i} ^ {u}\right), \tag {7} +$$ + +where $|C^u|$ is the size of the codebook $C^u$ , $CE(\cdot)$ is the cross-entropy loss for classification, $d_i^u$ represents the "ground-truth label" for the input state $u_i$ determined by the VQ-VAE, and $T$ is the number of frames. + +Actionable informativeness. The actionable informativeness objective focuses on reconstructing the velocity by leveraging gradient information from the discriminative hyper-network, $\mathcal{H}^u$ introduced above. Inspired by the idea that gradients of a discriminator function can inform the state changes, as observed in [46], this objective ensures that the model captures not only static state information, but also the dynamic changes in the sequence. By leveraging gradient fluctuations around a given state, the approach effectively captures dynamic motion details, such as velocity, + +resulting in more fine-grained motion concepts, and enriching the motion representation. + +Specifically, we initialize a network $\mathcal{A}^u$ for video data, which takes key image frames $U$ and gradients from the discriminator $\nabla_U\mathcal{H}^u (D^u,U)$ as inputs and predicts the velocity $V_{u}^{\prime}$ . The actionable informativeness loss for video is then computed as: + +$$ +\mathcal {L} _ {\text {v i d e o}} ^ {\text {a c t}} = \left| \left| V _ {u} - V _ {u} ^ {\prime} \right| \right| _ {2} ^ {2}. \tag {8} +$$ + +where $V_{u}$ represents the ground-truth velocity of video frames. This process is repeated for motion data to compute the corresponding loss $\mathcal{L}_{\mathrm{motion}}^{\mathrm{act}}$ using $V_{m}$ and $V_{m}^{\prime}$ . + +Combined Objectives. By integrating both video and motion components, we derive the total velocity reconstruction loss for human motion concept discovery and motion feature pre-training as in the following: + +$$ +\mathcal {L} ^ {\text {d i s}} = \mathcal {L} _ {\text {v i d e o}} ^ {\text {d i s}} + \mathcal {L} _ {\text {m o t i o n}} ^ {\text {d i s}} \tag {9} +$$ + +$$ +\mathcal {L} ^ {\mathrm {a c t}} = \mathcal {L} _ {\text {v i d e o}} ^ {\mathrm {a c t}} + \mathcal {L} _ {\text {m o t i o n}} ^ {\mathrm {a c t}} \tag {10} +$$ + +With the above objectives, our approach ensures that the model effectively reconstructs high-frequency details, improving the temporal fidelity and overall quality of multimodal representations. + +Feature Alignment. Multi-modality models often leverage the shared information across different modalities to enhance performance [5, 56, 87]. In the case of motion and video, video data provides contextual information that motion data alone may be lacking, which is critical for accurate motion prediction and understanding. Conversely, motion data focuses on human-centric features in the video, reducing over-reliance on environmental details. An efficient utilization or balancing of both is beneficial for the human motion LLMs. However, prior work on human motion + +and video understanding [5] only achieves implicit alignment through paired data during LLM fine-tuning, without explicit alignment in encoder training. To address this issue, we propose an alignment loss to improve information sharing and cross-modal coherence. + +To achieve alignment between video and motion features, we collect a set of video-motion pairs from Motion-X [41], which are aligned by frames. We instantiate two projection layers: The video projection layer $\mathcal{P}^u$ and the motion projection layer $\mathcal{P}^m$ . These layers are designed to map the respective features from the video and motion domains into a shared aligned space. + +More explicitly, given paired video and motion data $\{U,M\}$ , we use $\mathcal{P}^u$ and $\mathcal{P}^m$ to project the discrete video feature $d_i^u$ and motion feature $d_j^m$ into the aligned feature space, resulting in $a_i^u$ and $a_j^m$ , where $a_i^u, a_j^m \in \mathbb{R}^H$ . Here, $H$ denotes the dimensionality of the aligned features in the shared space. + +To ensure effective alignment, we define an alignment loss as follows: + +$$ +\mathcal {L} ^ {\text {a l i g n}} = \sum_ {0} ^ {T - 1} \frac {\operatorname {n o r m} \left(a _ {i} ^ {u}\right) \cdot \operatorname {n o r m} \left(a _ {\text {a l i g n} (i)} ^ {m}\right) / \epsilon}{\sum_ {0} ^ {K - 1} \operatorname {n o r m} \left(a _ {i} ^ {u}\right) \cdot \operatorname {n o r m} \left(a _ {j} ^ {m}\right) / \epsilon}, \tag {11} +$$ + +where $\mathrm{align}(i)$ indicates the pose index in the motion sequence corresponding to the key video frame $u_{i}$ , norm represents a normalization operation to stabilize feature magnitudes and $\epsilon$ is a temperature parameter that controls the sharpness of the alignment distribution. + +Note that $\mathcal{L}^{\mathrm{align}}$ is computed only when motion and video are paired. For standalone motion or video inputs, this loss term is omitted for applicability. + +The overall human motion concept discovery objective is defined as: + +$$ +\begin{array}{l} \mathcal {L} ^ {\text {t o t a l}} = \mathcal {L} _ {\text {m o t i o n}} ^ {\text {r e c}} + \mathcal {L} _ {\text {v i d e o}} ^ {\text {r e d e}} + \\ \lambda^ {\text {d i s}} \mathcal {L} ^ {\text {d i s}} + \lambda^ {\text {a c t}} \mathcal {L} ^ {\text {a c t}} + \lambda^ {\text {a l i g n}} \mathcal {L} ^ {\text {a l i g n}}, \tag {12} \\ \end{array} +$$ + +where $\lambda^{\mathrm{dis}}$ , $\lambda^{\mathrm{act}}$ and $\lambda^{\mathrm{align}}$ are weighting coefficients for the respective losses $\mathcal{L}^{\mathrm{dis}}$ , $\mathcal{L}^{\mathrm{act}}$ and $\mathcal{L}^{\mathrm{align}}$ . + +# 3.3. Modality Translation + +Although pre-trained encoders capture rich features from the original input, these features are often difficult for LLMs to interpret directly [5, 37, 40]. Therefore, it is essential to introduce an additional projection layer to map these features into a space that LLMs can effectively understand. + +As illustrated in Fig 2, we add tunable translation layers after the encoders $\mathcal{E}^u$ and $\mathcal{E}^m$ . These layers are trained using annotated video and motion sequence data to learn task-specific transformations. During this stage, the encoders and the LLM $\mathcal{F}$ are kept frozen to prevent adverse feature drift, and only the weights of the translation layers are updated. This strategy preserves the pre-trained knowledge of + +the encoders and language model while optimizing the feature mapping for downstream tasks. + +# 3.4. Multi-Modality Instruction Tuning + +Finally, we employ instruction tuning to train the language model $\mathcal{F}$ to respond to user queries using the corresponding video or motion sequences as context. During this process, the encoders and projection layers remain frozen, ensuring their pre-trained features are retained, while the weights of the language model $\mathcal{F}$ are updated to accommodate the task. + +In this stage, the model can further enhance cross-modal alignment using paired motion-video data in an implicit manner. To streamline the fine-tuning process, we adopt LoRA [29] with a rank of 8, allowing efficient and lightweight adjustment to the language model. + +# 4. Experiments + +# 4.1. Experiments Settings + +Implementation details For videos and motions, we initialize a codebook with 512 codes, a one-layer projector for feature alignment, and a 2-layer MLP to generate the weights for a 2-layer discriminator separately. For video inputs, we utilize pre-trained SigLIP [79] with an image patch size of 16 pixels and hidden dimension of 256 as the video encoder, and we take a 12-layer transformer as the reconstruction decoder. A 4-layer transformer is utilized as the motion encoder and a 2-layer transformer is used for motion reconstruction. For the velocity reconstruction, we initialize a two-layer hypernet and two-layer velocity decoder. Besides, we utilize one-layer MLPs for feature alignment. We pre-train the motion encoder with only the reconstruction loss for 60K iterations with a learning rate of 1e-4. Then, we train both the video and motion encoders with the total loss for 8K iterations with a learning rate of 1e-5. + +We utilize Llama3.1-8B as our base LLM model. For the modality translation stage, we initialize a one-layer projector for the video and motion encoder separately. We train the projectors on the annotation dataset for one batch with a batch size of 32 and a learning rate of 1e-4. In the Multimodality instruction tuning stage, we train the LLM model with a batch size of 16 and a learning rate of 3e-4. + +Training datasets For the encoder pre-training stage, we utilize Motion-X [41] datasets which contain 81K motion sequences and 36K motion-video pairs. In order to facilitate the training process of LLMs, we follow MotionLLM [5] to prompt LLMs to generate 20K annotation for motion-video pairs from the Motion-X dataset, and 10K annotation data from the HumanML3D [21] dataset. This data is utilized for modality translation. We additionally utilize the VATEX [69] dataset which contains 20K annotated videos for video translator training. For multi-modality instruction + +Table 1. Comparison of different methods on BABEL-QA test set. The pred type represent the prediction type, including classification and generation. For 2s-AGCN and MotionCLIP, 'M' means MLP, 'R' means RNN. Our method outperforms the baselines across all the metrics. + +
ModelPred typeOverallQuery typeRelation Type
ActionDirectionBodyPartBeforeAfterOther
2s-AGCN-Mcls.0.3550.3840.3520.2280.3310.2640.295
2s-AGCN-Rcls.0.3570.3960.3520.1940.3370.3010.285
MotionCLIP-Mcls.0.4300.4850.3610.2720.3720.3210.404
MotionCLIP-Rcls.0.4200.4890.3100.2500.3980.3140.387
MotionLLMgen.0.4360.5170.3540.1540.4270.3680.529
Oursgen.0.7110.8090.6970.6230.7070.6350.797
+ +tuning, we follow MotionLLM to generate 200K Q&A pairs for both motions and videos. We also utilize the VideoChatGPT [49] and BABEL-QA [13] datasets to enhance the motion and video understanding ability of our model, which has 100K and 2K Q&A data separately. + +Evaluation datasets Although MotionLLM [5] creates a large benchmark for both motion and video understanding, by the time we completed our paper, they had not released the specific details or access methods for the benchmark, nor did they provide a clear timeline for its release. Therefore, we did not perform comparisons on that benchmark but instead selected the test set of ActivityNet-QA [78] and BABEL-QA [13] to separately validate the performance of our algorithm on video and motion data. + +Evaluation metrics On the BABEL-QA benchmark, we evaluate the prediction accuracy for evaluation. As for the classification-based baselines, we evaluate the accuracy directly. As for the generative baselines, we adapt the evaluation protocol used in previous works [5, 37] by utilizing LLMs. Specifically, the evaluation LLM is given some examples of questions, answers, and scores (from 0 to 1), and it is required to follow these examples and score the experiment results. On the Activity-QA benchmark, we only compare our methods with our generative baselines. Thus, we directly utilize LLMs to evaluate the methods. + +# 4.2. Quantitative Results + +Experiments on ActivityNet-QA. We evaluate the video understanding ability of our method and other baselines on the ActivityNet-QA benchmark as shown in Tab. 2. We follow previous methods [5, 33, 38] to report the accuracy and scores that are evaluated by LLMs. We show in Tab. 2 that our method outperforms all the baselines on both metrics. Compared to the video-only method, our method aligns the video feature with the motion feature, making the encoder concentrate on human-centric information. Compared to MotionLLM, our encoder pre-training helps the network to + +learn more fine-grained information through velocity reconstruction. + +Experiments on BABEL-QA. We evaluate the motion understanding ability of our method as compared to existing studies on the BABEL-QA [13] dataset. Following [5, 13], we compare our method with a diverse set of baselines: 1) 2s-AGCN [58], a method using 2s-GCN to extract motion features and classify motion by MLP or RNN; 2) Motion-CLIP [63], a method that aligns the features of motion and text using CLIP [55] and predicts motion class by MLP or RNN; 3) MotionLLM [5], an LLM-based method which has motion reasoning ability. We follow [5] to fine-tune our method with the training set of BABEL-QA and evaluate our method with LLMs. + +From Tab. 1, our method outperforms other methods in all metrics by a large margin of nearly $60\%$ overall. In the "bodypart" query task, our algorithm achieves 2.29 times the score of the second-place method, MotionCLIP-M (0.623 vs. 0.272). This advantage stems partly from our velocity-based encoder training, which, compared to traditional motion feature extraction, helps us better capture high-frequency information and ensures the learning of fine-grained features. Additionally, unlike MotionLLM, we perform alignment at the encoder pre-training stage, ensuring that the features learned by the encoder are inherently more semantically rich. + +Table 2. Experiments results on ActivityNet-QA benchmark. + +
ModelAcc↑Score ↑
FrozenBiLM [57]24.7-
VideoChat [38]-2.2
LLaMA-Adapter [84]34.22.7
Video-LLaMA [80]12.41.1
Video-ChatGPT [49]35.22.7
Video-LLaVA [40]45.33.3
VideoChat2 [39]49.13.3
MotionLLM [5]53.33.5
Ours54.23.6
+ +Table 3. Ablation study on BABEL-QA test set. + +
ModelPred typeOverallQuery typeRelation Type
ActionDirectionBodyPartBeforeAfterOther
MotionLLMgen.0.4360.5170.3540.1540.4270.3680.529
Ours-w/oLrecgen.0.6960.7410.6450.5770.6000.5970.762
Ours-w/oLdis & Lactgen.0.6370.6930.4780.6060.6670.5260.709
Ours-w/oLaligngen.0.6750.7430.5790.5230.5840.5700.743
Oursgen.0.7110.8090.6970.6230.7070.6350.797
+ +![](images/592a68e5187fcb85cf5fb369af3df6751909641baebc47d0657f266ac6a8741c.jpg) +Figure 4. Example of motion understanding. We validate the proposed model's ability to understand motion from multiple aspects. Q1 on the left and Q1-3 on the right evaluate its comprehension of motion sequences, while Q2 and Q3 on the left push it to analyze kinematic, kinesthetic, and physical properties. These results demonstrate our algorithm's ability to recognize, describe in detail, and analyze motions. + +Ablation study. We present the ablation results in Tab. 3. We conducted ablation studies on masked reconstruction, velocity reconstruction, and feature alignment. Motion-LLM serves as an ablation baseline without any of these techniques. As presented, the absence of any single module leads to a decrease in performance. Among these, masked reconstruction has the least impact, as we initially train the motion encoder solely with masked reconstruction loss to ensure stable subsequent training. Removing velocity reconstruction has the greatest impact on pexrformance since velocity provides high-frequency information that other modules lack. Likewise, removing the alignment loss significantly affects results, highlighting the importance of multi-modality alignment in motion understanding. + +# 4.3. Qualitative Results + +To validate our algorithm's ability to understand motions and its contextual knowledge, we conducted multiple rounds of Q&A with the motion model. We present two examples in Fig. 4. We discover that, beyond its basic motion recognition capabilities, our algorithm can provide detailed descriptions of motions. In the example on the right, it can precisely describe the actions of the left hand and foot, as well as the right hand and foot. Our algorithm also + +possesses extensive knowledge about various movements. For instance, when asked, "what muscles does the person engage when bending their elbows during a push-up?" it can accurately respond with "triceps muscles." This demonstrates that our algorithm can understand the input motion sequences and perform corresponding analyses. We present more qualitative results in the supplementary material. + +# 5. Conclusion + +In this paper, we presented HuMoCon, a novel motion-video understanding framework for human action understanding and concept discovery. We demonstrated the effectiveness of our approach through comprehensive experiments, showcasing major improvements over state-of-the-art methods. Our work provides a strong foundation for future research in human action understanding and motion concept discovery. There are still some drawbacks of our work, such as the lack of human-environment contact or understanding niche motions. We plan to address those limitations by expanding the framework to incorporate additional modalities, particularly as environments, objects, and contact, and curate a wider range of specialized datasets. + +# Acknowledgment + +This work is supported by the Early Career Scheme of the Research Grants Council (grant # 27207224), the HKU-100 Award, a donation from the Musketeers Foundation, and an Academic Gift from Meta. + +# References + +[1] Nikos Athanasiou, Mathis Petrovich, Michael J Black, and Gúl Varol. Teach: Temporal action composition for 3d humans. In 2022 International Conference on 3D Vision (3DV), pages 414-423. IEEE, 2022. 2 +[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 1(2):3, 2023. 3 +[3] Yang Bai, Desen Zhou, Songyang Zhang, Jian Wang, Errui Ding, Yu Guan, Yang Long, and Jingdong Wang. Action quality assessment with temporal parsing transformer. In European conference on computer vision, pages 422-438. Springer, 2022. 1, 3 +[4] Pablo Cervantes, Yusuke Sekikawa, Ikuro Sato, and Koichi Shinoda. Implicit neural representations for variable length human motion generation. In European Conference on Computer Vision, pages 356-372. Springer, 2022. 1, 2 +[5] Ling-Hao Chen, Shunlin Lu, Ailing Zeng, Hao Zhang, Benyou Wang, Ruimao Zhang, and Lei Zhang. Motionllm: Understanding human behaviors from human motions and videos. arXiv preprint arXiv:2405.20340, 2024. 1, 2, 3, 5, 6, 7 +[6] Steven Chen and Richard R Yang. Pose trainer: correcting exercise posture using pose estimation. arXiv preprint arXiv:2006.11718, 2020. 1, 3 +[7] Sihan Chen, Xingjian He, Longteng Guo, Xinxin Zhu, Weining Wang, Jinhui Tang, and Jing Liu. Valor: Vision-audiolanguage omni-perception pretraining model and dataset. arXiv preprint arXiv:2304.08345, 2023. 3 +[8] Sihan Chen, Handong Li, Qunbo Wang, Zijia Zhao, Mingzhen Sun, Xinxin Zhu, and Jing Liu. Vast: A vision-audio-subtitle-text omni-modality foundation model and dataset. Advances in Neural Information Processing Systems, 36:72842-72866, 2023. 3 +[9] Ginger Delmas, Philippe Weinzaepfel, Francesc Moreno-Noguer, and Grégory Rogez. Posefix: correcting 3d human poses with natural language. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15018-15028, 2023. 2 +[10] Piotr Dólár, Vincent Rabaud, Garrison Cottrell, and Serge Belongie. Behavior recognition via sparse spatio-temporal features. In 2005 IEEE international workshop on visual surveillance and performance evaluation of tracking and surveillance, pages 65-72. IEEE, 2005. 3 +[11] Yong Du, Wei Wang, and Liang Wang. Hierarchical recurrent neural network for skeleton based action recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1110-1118, 2015. 1, 2 + +[12] Benjamin Elizalde, Soham Deshmukh, Mahmoud Al Ismail, and Huaming Wang. Clap learning audio concepts from natural language supervision. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023. 3 +[13] Mark Endo, Joy Hsu, Jiaman Li, and Jiajun Wu. Motion question answering via modular motion programs. In International Conference on Machine Learning, pages 9312-9328. PMLR, 2023. 2, 7 +[14] Qihang Fang, Yingda Yin, Qingnan Fan, Fei Xia, Siyan Dong, Sheng Wang, Jue Wang, Leonidas J Guibas, and Baoquan Chen. Towards accurate active camera localization. In European Conference on Computer Vision, pages 122-139. Springer, 2022. 1 +[15] Qihang Fang, Chengcheng Tang, Bugra Tekin, and Yanchao Yang. Cigtime: Corrective instruction generation through inverse motion editing. In The Thirty-Eighth Annual Conference on Neural Information Processing Systems, 2024. 1, 2 +[16] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 2, 3 +[17] Yao Feng, Jing Lin, Sai Kumar Dwivedi, Yu Sun, Priyanka Patel, and Michael J Black. Chatpose: Chatting about 3d human pose. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2093-2103, 2024. 2 +[18] Mihai Fieraru, Mihai Zanfir, Silviu Cristian Pirlela, Vlad Olaru, and Cristian Sminchisescu. Aifit: Automatic 3d human-interpretable feedback models for fitness training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9919–9928, 2021. 1, 3 +[19] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15180-15190, 2023. 3 +[20] Chuan Guo, Xinxin Zuo, Sen Wang, Shihao Zou, Qingyao Sun, Annan Deng, Minglun Gong, and Li Cheng. Action2motion: Conditioned generation of 3d human motions. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2021-2029, 2020. 1, 2 +[21] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5152-5161, 2022. 6 +[22] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5152-5161, 2022. 2 +[23] Chuan Guo, Xinxin Zuo, Sen Wang, and Li Cheng. Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In European Conference on Computer Vision, pages 580-597. Springer, 2022. 2 + +[24] Tianyu Guo, Hong Liu, Zhan Chen, Mengyuan Liu, Tao Wang, and Runwei Ding. Contrastive learning from extremely augmented skeleton sequences for self-supervised action recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 762-770, 2022. 1, 2 +[25] Tengda Han, Max Bain, Arsha Nagrani, Gul Varol, Weidi Xie, and Andrew Zisserman. Autoad ii: The sequel-who, when, and what in movie audio description. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13645–13655, 2023. 3 +[26] Tengda Han, Max Bain, Arsha Nagrani, Gül Varol, Weidi Xie, and Andrew Zisserman. Autoad iii: The prequel-back to the pixels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18164-18174, 2024. 1, 3 +[27] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 2, 4 +[28] Daniel Holden, Taku Komura, and Jun Saito. Phase-functional neural networks for character control. ACM Transactions on Graphics (TOG), 36(4):1-13, 2017. 2 +[29] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 6 +[30] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. Advances in Neural Information Processing Systems, 36:20067-20079, 2023. 2 +[31] Hanwen Jiang, Shaowei Liu, Jiashun Wang, and Xiaolong Wang. Hand-object contact consistency reasoning for human grasps generation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11107-11116, 2021. 2 +[32] Yujiao Jiang, Qingmin Liao, Zhaolong Wang, Xiangru Lin, Zongqing Lu, Yuxi Zhao, Hanqing Wei, Jingrui Ye, Yu Zhang, and Zhijing Shao. Smplx-lite: A realistic and drivable avatar benchmark with rich geometry and texture annotations. arXiv preprint arXiv:2405.19609, 2024. 1 +[33] Peng Jin, Ryuichi Takanobu, Wancai Zhang, Xiaochun Cao, and Li Yuan. Chat-univi: Unified visual representation empowers large language models with image and video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13700-13710, 2024. 7 +[34] Qiuhong Ke, Mohammed Bennamoun, Senjian An, Ferdous Sohel, and Farid Boussaid. A new representation of skeleton sequences for 3d action recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3288-3297, 2017. 1, 2 +[35] Jihoon Kim, Jiseob Kim, and Sungjoon Choi. Flame: Freeform language-based motion synthesis & editing. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8255-8263, 2023. 2 +[36] Kosmas Kritsis, Aggelos Gkiokas, Aggelos Pikrakis, and Vassilis Katsouros. Danceconv: Dance motion genera + +tion with convolutional networks. IEEE Access, 10:44982-45000, 2022. 2 +[37] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 1, 2, 3, 6, 7 +[38] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 7 +[39] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mybench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 7 +[40] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 1, 2, 3, 6, 7 +[41] Jing Lin, Ailing Zeng, Shunlin Lu, Yuanhao Cai, Ruimao Zhang, Haoqian Wang, and Lei Zhang. Motion-x: A large-scale 3d expressive whole-body human motion dataset. Advances in Neural Information Processing Systems, 36, 2024. 6 +[42] Lilang Lin, Jiahang Zhang, and Jiaying Liu. Actionlet-dependent contrastive learning for unsupervised skeleton-based action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2363-2372, 2023. 1, 2 +[43] Daochang Liu, Qiyue Li, Tingting Jiang, Yizhou Wang, Rulin Miao, Fei Shan, and Ziyu Li. Towards unified surgical skill assessment. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9522-9531, 2021. 1, 3 +[44] Jun Liu, Amir Shahroudy, Dong Xu, and Gang Wang. Spatio-temporal LSTM with trust gates for 3d human action recognition. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 816-833. Springer, 2016. 1, 2, 3 +[45] Mengyuan Liu, Hong Liu, and Chen Chen. Enhanced skeleton visualization for view invariant human action recognition. Pattern Recognition, 68:346-362, 2017. 1, 2 +[46] Ruizhe Liu, Qian Luo, and Yanchao Yang. Infocon: Concept discovery with generative and discriminative informativeness. arXiv preprint arXiv:2404.10606, 2024. 2, 4, 5 +[47] Thomas Lucas, Fabien Baradel, Philippe Weinzaepfel, and Grégory Rogez. Posegpt: Quantization-based 3d human motion generation and forecasting. In European Conference on Computer Vision, pages 417-435. Springer, 2022. 2 +[48] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing, 508:293-304, 2022. 3 +[49] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video + +understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 7 +[50] Shehan Munasinghe, Rusiru Thushara, Muhammad Maaz, Hanoona Abdul Rasheed, Salman Khan, Mubarak Shah, and Fahad Khan. Pg-video-llava: Pixel grounding large video-language models. arXiv preprint arXiv:2311.13435, 2023. 1, 3 +[51] Paritosh Parmar and Brendan Tran Morris. What and how well you performed? a multitask learning approach to action quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 304-313, 2019. 1, 3 +[52] Mathis Petrovich, Michael J Black, and Gül Varol. Action-conditioned 3d human motion synthesis with transformer vae. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10985-10995, 2021. 1, 2 +[53] Mathis Petrovich, Michael J Black, and Gül Varol. Temos: Generating diverse human motions from textual descriptions. In European Conference on Computer Vision, pages 480-497. Springer, 2022. 2 +[54] Abhinanda R Punnakkal, Arjun Chandrasekaran, Nikos Athanasiou, Alejandra Quiros-Ramirez, and Michael J Black. Babel: Bodies, action and behavior with english labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 722-731, 2021. 2 +[55] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3, 7 +[56] Jathushan Rajasegaran, Georgios Pavlakos, Angjoo Kanazawa, Christoph Feichtenhofer, and Jitendra Malik. On the benefits of 3d pose and tracking for human action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 640-649, 2023. 2, 5 +[57] Leonard Salewski, Stefan Fauth, A Koepke, and Zeynep Akata. Zero-shot audio captioning with audio-language model guidance and audio context keywords. arXiv preprint arXiv:2311.08396, 2023. 3, 7 +[58] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Two-stream adaptive graph convolutional networks for skeleton-based action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12026-12035, 2019. 1, 2, 7 +[59] Sebastian Starke, He Zhang, Taku Komura, and Jun Saito. Neural state machine for character-scene interactions. ACM Transactions on Graphics, 38(6):178, 2019. 1, 2 +[60] Sebastian Starke, Yiwei Zhao, Taku Komura, and Kazi Zaman. Local motion phases for learning multi-contact character movements. ACM Transactions on Graphics (TOG), 39 (4):54-1, 2020. 2 +[61] Sebastian Starke, Ian Mason, and Taku Komura. Deeppphase: Periodic autoencoders for learning motion phase manifolds. + +ACM Transactions on Graphics (TOG), 41(4):1-13, 2022. 1, 2 +[62] Graham W Taylor, Rob Fergus, Yann LeCun, and Christoph Bregler. Convolutional learning of spatio-temporal features. In Computer Vision-ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part VI 11, pages 140-153. Springer, 2010. 2, 3 +[63] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. In European Conference on Computer Vision, pages 358–374. Springer, 2022. 2, 3, 7 +[64] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022. 2 +[65] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 4489-4497, 2015. 2, 3 +[66] Raviteja Vemulapalli, Felipe Arrate, and Rama Chellappa. Human action recognition by representing 3d skeletons as points in a lie group. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 588-595, 2014. 3 +[67] Heng Wang and Cordelia Schmid. Action recognition with improved trajectories. In Proceedings of the IEEE international conference on computer vision, pages 3551-3558, 2013. 3 +[68] Jingbo Wang, Ye Yuan, Zhengyi Luo, Kevin Xie, Dahua Lin, Umar Iqbal, Sanja Fidler, and Sameh Khamis. Learning human dynamics in autonomous driving scenarios. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20796-20806, 2023. 1 +[69] Xin Wang, Jiawei Wu, Junkun Chen, Lei Li, Yuan-Fang Wang, and William Yang Wang. Vatex: A large-scale, high-quality multilingual dataset for video-and-language research. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4581-4591, 2019. 6 +[70] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, et al. Internvideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191, 2022. 2, 3 +[71] Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Guo Chen, Baoqi Pei, Rongkun Zheng, Jilan Xu, Zun Wang, et al. Intervideo2: Scaling video foundation models for multimodal video understanding. arXiv preprint arXiv:2403.15377, 2024. 2, 3 +[72] Zeqi Xiao, Tai Wang, Jingbo Wang, Jinkun Cao, Wenwei Zhang, Bo Dai, Dahua Lin, and Jiangmiao Pang. Unified human-scene interaction via prompted chain-of-contacts. arXiv preprint arXiv:2309.07918, 2023. 1 +[73] Junyu Xie, Tengda Han, Max Bain, Arsha Nagrani, Güll Varol, Weidi Xie, and Andrew Zisserman. Autoad-zero: A training-free framework for zero-shot audio description. arXiv preprint arXiv:2407.15850, 2024. 3 + +[74] Guo-Hao Xu, Yi-Lin Wei, Dian Zheng, Xiao-Ming Wu, and Wei-Shi Zheng. Dexterous grasp transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17933-17942, 2024. 2 +[75] Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. Pllava: Parameter-free llava extension from images to videos for video dense captioning. arXiv preprint arXiv:2404.16994, 2024. 1, 3 +[76] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In Proceedings of the AAAI conference on artificial intelligence, 2018. 1, 2 +[77] Zeshi Yang, Kangkang Yin, and Libin Liu. Learning to use chopsticks in diverse gripping styles. 2022. 2 +[78] Zhou Yu, Dejing Xu, Jun Yu, Ting Yu, Zhou Zhao, Yueting Zhuang, and Dacheng Tao. Activitynet-qa: A dataset for understanding complex web videos via question answering. In AAAI, pages 9127–9134, 2019. 2, 7 +[79] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 6 +[80] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 3, 7 +[81] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Shaoli Huang, Yong Zhang, Hongwei Zhao, Hongtao Lu, and Xi Shen. T2m-gpt: Generating human motion from textual descriptions with discrete representations. arXiv preprint arXiv:2301.06052, 2023. 2 +[82] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2 +[83] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xupeng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. Pointclip: Point cloud understanding by clip. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8552-8562, 2022. 3 +[84] Renrui Zhang, Jiaming Han, Chris Liu, Peng Gao, Aojun Zhou, Xiangfei Hu, Shilin Yan, Pan Lu, Hongsheng Li, and Yu Qiao. Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199, 2023. 7 +[85] Xingyi Zhou, Anurag Arnab, Shyamal Buch, Shen Yan, Austin Myers, Xuehan Xiong, Arsha Nagrani, and Cordelia Schmid. Streaming dense video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18243-18252, 2024. 3 +[86] Zixiang Zhou, Yu Wan, and Baoyuan Wang. Avatargpt: All-in-one framework for motion understanding planning generation and beyond. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1357-1366, 2024. 2 +[87] Bin Zhu, Bin Lin, Munan Ning, Yang Yan, Jiaxi Cui, HongFa Wang, Yatian Pang, Wenhao Jiang, Junwu Zhang, Zongwei + +Li, et al. Languagebind: Extending video-language pretraining to n-modality by language-based semantic alignment. arXiv preprint arXiv:2310.01852, 2023. 2, 3, 5 \ No newline at end of file diff --git a/2025/HuMoCon_ Concept Discovery for Human Motion Understanding/images.zip b/2025/HuMoCon_ Concept Discovery for Human Motion Understanding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..26e2b897b3e5518afc48ca20d717090b11fbe6c6 --- /dev/null +++ b/2025/HuMoCon_ Concept Discovery for Human Motion Understanding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8eeda36ff8d0ab04d53032351be77260d549a00a663f34fe13f5885d5e7482e0 +size 439631 diff --git a/2025/HuMoCon_ Concept Discovery for Human Motion Understanding/layout.json b/2025/HuMoCon_ Concept Discovery for Human Motion Understanding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..55ebc3a328b4daa2c40908074fd6c7443ef065d6 --- /dev/null +++ b/2025/HuMoCon_ Concept Discovery for Human Motion Understanding/layout.json @@ -0,0 +1,9517 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 101, + 103, + 510, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 103, + 510, + 121 + ], + "spans": [ + { + "bbox": [ + 101, + 103, + 510, + 121 + ], + "type": "text", + "content": "HuMoCon: Concept Discovery for Human Motion Understanding" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 92, + 142, + 168, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 142, + 168, + 158 + ], + "spans": [ + { + "bbox": [ + 92, + 142, + 168, + 158 + ], + "type": "text", + "content": "Qihang Fang" + }, + { + "bbox": [ + 92, + 142, + 168, + 158 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 181, + 143, + 275, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 143, + 275, + 157 + ], + "spans": [ + { + "bbox": [ + 181, + 143, + 275, + 157 + ], + "type": "text", + "content": "Chengcheng Tang" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 208, + 157, + 358, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 157, + 358, + 172 + ], + "spans": [ + { + "bbox": [ + 208, + 157, + 358, + 172 + ], + "type": "text", + "content": "1The University of Hong Kong" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 366, + 143, + 428, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 143, + 428, + 157 + ], + "spans": [ + { + "bbox": [ + 366, + 143, + 428, + 157 + ], + "type": "text", + "content": "Shugao Ma2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 442, + 143, + 515, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 442, + 143, + 515, + 157 + ], + "spans": [ + { + "bbox": [ + 442, + 143, + 515, + 157 + ], + "type": "text", + "content": "Yanchao Yang" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 370, + 157, + 402, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 157, + 402, + 171 + ], + "spans": [ + { + "bbox": [ + 370, + 157, + 402, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 370, + 157, + 402, + 171 + ], + "type": "text", + "content": "Meta" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 87, + 173, + 518, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 173, + 518, + 185 + ], + "spans": [ + { + "bbox": [ + 87, + 173, + 518, + 185 + ], + "type": "text", + "content": "qihfang@gmail.com, {chengcheng.tang,bugratekin,shugao}@meta.com, yanchaoy@hku.hk" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 65, + 198, + 342, + 314 + ], + "blocks": [ + { + "bbox": [ + 65, + 198, + 342, + 314 + ], + "lines": [ + { + "bbox": [ + 65, + 198, + 342, + 314 + ], + "spans": [ + { + "bbox": [ + 65, + 198, + 342, + 314 + ], + "type": "image", + "image_path": "897542e4c766ca6d2895097d7377ad1855886fce72dc2242c7b2d4bd4dd03f85.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 327, + 555, + 373 + ], + "lines": [ + { + "bbox": [ + 54, + 327, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 54, + 327, + 555, + 373 + ], + "type": "text", + "content": "Figure 1. We present HuMoCon for concept discovery and human motion understanding with video or motion input. To address the challenge of feature misalignment and high-frequency information loss, we propose a novel feature alignment strategy and an advanced masked auto-encoder reconstructing velocity. HuMoCon empowers effective motion concept discovery and accurate Question Answering tasks, significantly outperforming state-of-the-art methods qualitatively and quantitatively." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 352, + 199, + 553, + 312 + ], + "blocks": [ + { + "bbox": [ + 352, + 199, + 553, + 312 + ], + "lines": [ + { + "bbox": [ + 352, + 199, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 352, + 199, + 553, + 312 + ], + "type": "image", + "image_path": "97026424ca8d3a63ed694ff5fda914904f300cc2f49c155f651d897df1ae9167.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 151, + 381, + 200, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 381, + 200, + 395 + ], + "spans": [ + { + "bbox": [ + 151, + 381, + 200, + 395 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 54, + 407, + 297, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 407, + 297, + 647 + ], + "spans": [ + { + "bbox": [ + 54, + 407, + 297, + 647 + ], + "type": "text", + "content": "We present HuMoCon, a novel motion-video understanding framework designed for advanced human behavior analysis. The core of our method is a human motion concept discovery framework that efficiently trains multi-modal encoders to extract semantically meaningful and generalizable features. HuMoCon addresses key challenges in motion concept discovery for understanding and reasoning, including the lack of explicit multi-modality feature alignment and the loss of high-frequency information in masked autoencoding frameworks. Our approach integrates a feature alignment strategy that leverages video for contextual understanding and motion for fine-grained interaction modeling, further with a velocity reconstruction mechanism to enhance high-frequency feature expression and mitigate temporal over-smoothing. Comprehensive experiments on standard benchmarks demonstrate that HuMoCon enables effective motion concept discovery and significantly outperforms state-of-the-art methods in training large models for human motion understanding. We will open-source the associated code with our paper." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 669, + 135, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 669, + 135, + 681 + ], + "spans": [ + { + "bbox": [ + 55, + 669, + 135, + 681 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "Understanding human behavior is crucial for gaining deeper insights into human actions and reasoning, which are foun" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 382, + 555, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 382, + 555, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 382, + 555, + 479 + ], + "type": "text", + "content": "ditional for developing human-centered AI systems [5, 26, 37, 61] that effectively address real-world tasks [14, 68, 72]. Recent advancements in large language models (LLMs) have led to significant progress in video- and motion-based methods [5, 15, 37, 40, 75] for analyzing human behavior. Despite these advancements, achieving a precise and fine-grained understanding of human actions remains a persistent and complex challenge." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 483, + 556, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 556, + 602 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 556, + 602 + ], + "type": "text", + "content": "Human behavior is often represented as motion sequences structured based on the kinematics of the human skeleton [32?]. Traditional approaches to analyzing motion sequences have predominantly focused on classification [11, 24, 34, 42, 44, 45, 58, 76], relying on predefined action categories. While effective for specific tasks, these methods often lack the flexibility for deeper or more nuanced analysis. Other methods employ statistical models to assess motion quality [3, 6, 18, 43, 51], but they require significant expert knowledge, limiting scalability." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "content": "Recent advancements in language models [40, 50, 75] have opened new possibilities in motion analysis, enabling tasks such as motion generation and motion captioning based on free-text inputs [4, 20, 52, 59, 61], advancing the field of motion analysis. However, obtaining high-quality motion sequences and annotations remains costly, and motion sequences generally lack information about the surrounding environment, limiting models' ability to understand human-environment interactions." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7179" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "type": "text", + "content": "In contrast, video data offers rich visual context and is easier to collect. It has been widely used for action classification [16, 62, 65] and video Q&A [5, 37, 40]. However, video data often contains irrelevant elements that can hinder accurate content analysis. For example, a model might learn from training data that most running occurs in stadiums, leading it to mistakenly classify other stadium activities as running during testing." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 171, + 298, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 171, + 298, + 397 + ], + "spans": [ + { + "bbox": [ + 56, + 171, + 298, + 397 + ], + "type": "text", + "content": "In recent years, significant progress has been made in developing multi-modality models [5, 87]. Integrating multiple modalities, as opposed to relying on a single modality, enhances the network's ability to learn fine-grained and semantic features. For instance, LART [56] combines video and motion data to ensure that video-based action classification focuses on the human body rather than overemphasizing environmental context. MotionLLM [5] expands large language models (LLMs) to video and motion, building a model that can understand both. MotionLLM demonstrated that learning from video and motion improves the network's ability to analyze human behavior. However, MotionLLM implicitly aligns video and motion using motion-video pairs during LLM training without optimizing the encoder for explicit cross-modal alignment. To address this limitation and further enhance feature representation, we propose explicitly aligning features from different modalities during the encoding process within a human motion concept discovery framework." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 400, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 400, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 400, + 295, + 628 + ], + "type": "text", + "content": "Achieving effective cross-modal alignment and feature representation relies heavily on robust encoder training methods. Many existing studies [70, 71, 87] adopt a masked autoencoder framework [27] when training encoders. This approach involves applying a mask to a certain proportion of the input, extracting features from the masked data using the encoder, and reconstructing the original unmasked input with a decoder. While effective in learning high-quality features, this masking strategy inherently causes the loss of high-frequency information, resulting in temporally over-smoothed reconstructions that hinder the LLM's ability to understand human behavior. [47] To address this issue, we propose reconstructing velocity as a means to preserve high-frequency information. Following InfoCon [46], we design a VQ-VAE-based loss function that helps retain the fine-grained features of the original masked autoencoder while further enhancing high-frequency feature expression. This approach significantly boosts the subsequent LLM's capacity to model and understand human motion effectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "content": "In summary, we present HuMoCon, an effective human motion concept discovery pertaining framework for advanced human behavior understanding from both video and motion data. Our key contributions include: (i) a novel feature alignment strategy that explicitly leverages video data for contextual understanding and motion data for fine-grained interaction modeling, and (ii) an enhanced masked" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "content": "autoencoder framework incorporating velocity reconstruction to significantly improve high-frequency feature expression and mitigate temporal over-smoothing. To evaluate the effectiveness of our proposed method, we conduct experiments on the Activity-QA [78] and BABEL-QA [13] benchmarks [54], comparing our approach with the state-of-the-art methods. Comprehensive quantitative and qualitative evaluations demonstrate that HuMoCon achieves state-of-the-art performance on standard benchmarks for motion-video understanding." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 206, + 400, + 218 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 206, + 400, + 218 + ], + "spans": [ + { + "bbox": [ + 313, + 206, + 400, + 218 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 227, + 481, + 240 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 227, + 481, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 227, + 481, + 240 + ], + "type": "text", + "content": "2.1. Human Motion Understanding" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 245, + 553, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 245, + 553, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 245, + 553, + 473 + ], + "type": "text", + "content": "The core of human motion understanding is to learn the semantics of motion sequences or videos. In the early stages, research in this field is often based on predefined action categories. Such work can be divided into two types: action classification [11, 24, 34, 42, 44, 45, 58, 76] and motion generation targeted at specific categories [4, 20, 28, 31, 36, 52, 59-61, 74, 77]. For action recognition, several works [11, 44] utilize Recurrent Neural Networks (RNN) to extract temporal information from human joint sequences, beyond the CNN-based ones [34, 45]. Subsequent works [24, 42, 58, 76] propose representing motion data as a graph structure, enabling the use of spatio-temporal graph convolutional networks (ST-GCNs) to effectively capture the connectivity and interactions between human joints. For motion generation, several studies [28, 31, 36, 60, 74, 77] have concentrated on generating specific, single actions. In contrast, other works [4, 20, 52, 59, 61] aim to generate a broader range of diverse action types, addressing the challenge of variability in motion synthesis." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 475, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 553, + 713 + ], + "type": "text", + "content": "In recent years, more research has shifted towards captioning [23, 30], generation [1, 22, 23, 30, 35, 53, 64, 81, 82, 86], and analysis [5, 9, 15, 17] based on free-form text. Leveraging alignment of motion latent features with the CLIP [55] embedding space, MotionCLIP [63] enables the generation of out-of-distribution motions. Transformers [1, 22, 23, 53, 81] and diffusion [35, 64, 82] models are also utilized for text-conditioned motion generation. Moreover, MotionGPT [30] and AvatarGPT [86] utilize large language models (LLMs) to handle multiple motion-related tasks simultaneously. Despite these advancements, research focused specifically on motion analysis remains relatively limited. PoseFix [9] and CigTime [15] are proposed to annotate the corrective instructions from source to target for pose and motion sequences. MotionLLM [5] proposes a method to analyze both the video and motion sequences with detailed spatial-temporal awareness and reasoning abilities. However, MotionLLM relies on implicit alignment through paired data and lacks explicit cross-modal feature alignment." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7180" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 179, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 179, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 179, + 85 + ], + "type": "text", + "content": "2.2. Video Understanding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 297, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 297, + 413 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 297, + 413 + ], + "type": "text", + "content": "In addition to motion understanding, video understanding has emerged as a prominent research area, attracting extensive attention in recent years. Early works on video classification and segmentation relied on handcrafted features [10, 44, 66, 67] or convolutional neural networks [16, 62, 65]. Beyond classification, video analysis [3, 6, 18, 25, 26, 37, 40, 43, 50, 51, 57, 73, 75, 85] has become a highly active field of research. Action quality assessment has been the focus of several works [3, 6, 18, 43, 51], with methods such as Aifit [18] and PoseTrainer [6] proposed to correct user actions. These approaches evaluate actions in conjunction with complex textual instructions [3, 43, 51]. Other studies [25, 26, 57, 73, 85] aim to generate descriptive narratives for videos, broadening the scope of video understanding. Recently, many works [2, 37, 40, 50, 75, 80] have leveraged large language models (LLMs) to understand video content and perform complex reasoning tasks. These methods combine video processing with the advanced reasoning capabilities of LLMs, further advancing the state-of-the-art in video analysis. Video-LLaVA [40] build a model to perform visual reasoning capabilities on images and videos simultaneously. Follow-up works [37, 50, 75] extend Video-LLaVA for faster inference and more accurate reasoning ability. Unlike typical video understanding algorithms, which rely heavily on visual context alone, our approach integrates motion data directly, enhancing the model's ability to understand human-centric actions." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 420, + 211, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 420, + 211, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 420, + 211, + 433 + ], + "type": "text", + "content": "2.3. Multi-modality Pre-training" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 437, + 295, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 295, + 521 + ], + "type": "text", + "content": "Multi-modality pre-training has primarily focused on aligning vision and language representations. CLIP [55] pioneered this effort by using contrastive learning to align image and text features on large-scale datasets. Subsequent works [12, 48, 63, 83] extended CLIP's framework to other modalities, including video [48], audio [12] and motion [63]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 521, + 295, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 593 + ], + "type": "text", + "content": "Recent studies, such as VALOR [7] and VAST [8], demonstrated that cross-modal alignment not only enhances model robustness but also maintains high performance across modalities. Building on this, follow-up works [19, 87] have aimed to expand the number of aligned modalities, further advancing multi-modality research." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 594, + 295, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 642 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 642 + ], + "type": "text", + "content": "In the context of motion and video alignment, Motion-LLMs [5] is the first work to address this challenge. However, it relies solely on implicit alignment during the fine-tuning stage of large language models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 652, + 111, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 652, + 111, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 652, + 111, + 664 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 672, + 123, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 123, + 683 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 123, + 683 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "Our objective is to develop an LLM-based system capable of understanding and reasoning both human video and mo" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "tion inputs. Given a question " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": " and an input, which can be either a video " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": " or motion sequence " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": ", the model generates the corresponding answer " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": ". Specifically, a video " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": " is represented as a sequence of " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": " keyframes, denoted as " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "\\{u_0, u_1, \\dots, u_{T-1}\\}" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": ". Similarly, a motion sequence " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": " is represented as " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": " consecutive pose frames, denoted as " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "\\{m_0, m_1, \\dots, m_{K-1}\\}" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 157, + 555, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 157, + 555, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 555, + 348 + ], + "type": "text", + "content": "We present an overview of the pipeline in Fig. 2. To enable effective multi-modal human motion understanding, our framework focuses on: (1) discovering human motion concepts from both video and motion data, (2) preserving high-frequency motion details using velocity reconstruction to reduce temporal smoothing, and (3) aligning video and motion features to enhance multi-modal comprehension. To accomplish this, we train our video and motion encoders with four key learning objectives: Masked Reconstruction, Discriminative Informativeness, Actionable Informativeness, and Feature Alignment. Consequently, we leverage the encoded features of videos and motions to facilitate the fine-tuning of LLMs for video-motion understanding and reasoning. The following sections detail each component of our approach, including multi-modality encoder pre-training, modality translation, and instruction tuning." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 355, + 529, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 355, + 529, + 380 + ], + "spans": [ + { + "bbox": [ + 313, + 355, + 529, + 380 + ], + "type": "text", + "content": "3.2. Multi-Modality Encoder Pre-training via Human Motion Concept Discovery" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 384, + 555, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 555, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 555, + 529 + ], + "type": "text", + "content": "Pre-training encoders are critical for multi-modality large language models (LLMs) [70, 71, 87], as they enable the effective extraction of essential features from large-scale input data. To enhance the encoder's ability to capture key information, we design three distinct learning objectives: masked reconstruction, velocity reconstruction, and feature alignment, all contributing to the quality of the discovered human motion concepts as well as the pre-training efficiency. For both videos and motions, we adopt a similar network structure during encoder pre-training. Consequently, in the following section, we focus on detailing the network structure for video inputs." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 544, + 554, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 554, + 603 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 554, + 603 + ], + "type": "text", + "content": "Encoding Function. In our approach, we design masked autoencoder architectures tailored for both video and motion data. Specifically, for videos, we utilize a pre-trained encoder " + }, + { + "bbox": [ + 313, + 544, + 554, + 603 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_u" + }, + { + "bbox": [ + 313, + 544, + 554, + 603 + ], + "type": "text", + "content": ", to encode the input " + }, + { + "bbox": [ + 313, + 544, + 554, + 603 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 313, + 544, + 554, + 603 + ], + "type": "text", + "content": " into a feature representation " + }, + { + "bbox": [ + 313, + 544, + 554, + 603 + ], + "type": "inline_equation", + "content": "f^u" + }, + { + "bbox": [ + 313, + 544, + 554, + 603 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 405, + 604, + 553, + 617 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 604, + 553, + 617 + ], + "spans": [ + { + "bbox": [ + 405, + 604, + 553, + 617 + ], + "type": "interline_equation", + "content": "F ^ {u} = \\mathcal {E} _ {u} (U). \\tag {1}", + "image_path": "ebdba158b7cb37b1b40a45d37dd4ff9ecfcbb64c729f9d7356f328e632a9e366.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 622, + 554, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 622, + 554, + 657 + ], + "spans": [ + { + "bbox": [ + 313, + 622, + 554, + 657 + ], + "type": "text", + "content": "Then, we employ a codebook " + }, + { + "bbox": [ + 313, + 622, + 554, + 657 + ], + "type": "inline_equation", + "content": "C^u" + }, + { + "bbox": [ + 313, + 622, + 554, + 657 + ], + "type": "text", + "content": ", modeled with VQ-VAE to quantize " + }, + { + "bbox": [ + 313, + 622, + 554, + 657 + ], + "type": "inline_equation", + "content": "f^u" + }, + { + "bbox": [ + 313, + 622, + 554, + 657 + ], + "type": "text", + "content": " into a discrete feature representation " + }, + { + "bbox": [ + 313, + 622, + 554, + 657 + ], + "type": "inline_equation", + "content": "d^u" + }, + { + "bbox": [ + 313, + 622, + 554, + 657 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 381, + 658, + 553, + 672 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 658, + 553, + 672 + ], + "spans": [ + { + "bbox": [ + 381, + 658, + 553, + 672 + ], + "type": "interline_equation", + "content": "D ^ {u} = \\operatorname {Q u a n t i z e} \\left(F ^ {u}, C ^ {u}\\right), \\tag {2}", + "image_path": "71abe5c62dd55dcf10ef6482ec3044b300c10da3d6c9b897be6d901fdd10d02e.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "where 'Quantize' denotes the quantization operation, and " + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "inline_equation", + "content": "C^u" + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": " is the codebook used for discretization, also serving as the discovered human motion concepts." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7181" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 88, + 70, + 523, + 280 + ], + "blocks": [ + { + "bbox": [ + 88, + 70, + 523, + 280 + ], + "lines": [ + { + "bbox": [ + 88, + 70, + 523, + 280 + ], + "spans": [ + { + "bbox": [ + 88, + 70, + 523, + 280 + ], + "type": "image", + "image_path": "e40073539a848eda21d5e8fb455fc0e97d59db9d6253a2473b5c1d203ac79cb9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 288, + 555, + 354 + ], + "lines": [ + { + "bbox": [ + 54, + 288, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 54, + 288, + 555, + 354 + ], + "type": "text", + "content": "Figure 2. System overview of our method. (a) The encoder pre-training process for learning and aligning video and motion features and enhancing high-frequency details through velocity reconstruction. We utilize a VQ-VAE-based structure, and we design effective learning objectives to enhance the encoder to extract semantic meaningful and fine-grained features. (b) The fine-tuning of the large language model (LLM) for video and motion reasoning consists of two stages: Modality Translation and Multi-modality Instruction Tuning. In the Modality Translation stage, we train a translation layer for each modality to map the encoding feature to the LLM space. In the Instruction Tuning stage, we fine-tune the LLM to understand human motion and videos for downstream tasks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 365, + 295, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 365, + 295, + 437 + ], + "spans": [ + { + "bbox": [ + 54, + 365, + 295, + 437 + ], + "type": "text", + "content": "Masked Reconstruction. The masked autoencoder has proven to be highly effective for pre-training [27], as it facilitates learning of meaningful and low-frequency features from input data. Thus, we leverage masked auto-encoding to enhance the discovered concepts and facilitate the multimodal feature pre-training." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 437, + 295, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 295, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 295, + 464 + ], + "type": "text", + "content": "More explicitly, we mask " + }, + { + "bbox": [ + 55, + 437, + 295, + 464 + ], + "type": "inline_equation", + "content": "D^u" + }, + { + "bbox": [ + 55, + 437, + 295, + 464 + ], + "type": "text", + "content": " by a probability of " + }, + { + "bbox": [ + 55, + 437, + 295, + 464 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 55, + 437, + 295, + 464 + ], + "type": "text", + "content": " to obtain the masked feature " + }, + { + "bbox": [ + 55, + 437, + 295, + 464 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{mask}}^u" + }, + { + "bbox": [ + 55, + 437, + 295, + 464 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 127, + 472, + 294, + 485 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 472, + 294, + 485 + ], + "spans": [ + { + "bbox": [ + 127, + 472, + 294, + 485 + ], + "type": "interline_equation", + "content": "D _ {\\text {m a s k}} ^ {u} = \\operatorname {M a s k} \\left(D ^ {u}, \\sigma\\right), \\tag {3}", + "image_path": "2d9d78fec5c168afad01ee7b7a9ca8eb09a1588ad91e8868999aca406d3a7be7.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 495, + 296, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 296, + 533 + ], + "type": "text", + "content": "where 'Mask' is the operation that applies masking to the features, and " + }, + { + "bbox": [ + 55, + 495, + 296, + 533 + ], + "type": "inline_equation", + "content": "\\sigma < 1.0" + }, + { + "bbox": [ + 55, + 495, + 296, + 533 + ], + "type": "text", + "content": " controls the masking ratio. A video decoder " + }, + { + "bbox": [ + 55, + 495, + 296, + 533 + ], + "type": "inline_equation", + "content": "D^{u}" + }, + { + "bbox": [ + 55, + 495, + 296, + 533 + ], + "type": "text", + "content": " is used to reconstruct the video " + }, + { + "bbox": [ + 55, + 495, + 296, + 533 + ], + "type": "inline_equation", + "content": "U^{\\prime}" + }, + { + "bbox": [ + 55, + 495, + 296, + 533 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 495, + 296, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{mask}}^{u}" + }, + { + "bbox": [ + 55, + 495, + 296, + 533 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 543, + 294, + 556 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 543, + 294, + 556 + ], + "spans": [ + { + "bbox": [ + 138, + 543, + 294, + 556 + ], + "type": "interline_equation", + "content": "U ^ {\\prime} = \\mathcal {D} ^ {u} \\left(D _ {\\text {m a s k}} ^ {u}\\right). \\tag {4}", + "image_path": "fe3fa1f83bc2804fb0fc1eda809a5ba9238c8b7dab89ddf9b3368596c7d6db2a.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 563, + 295, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 295, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 295, + 586 + ], + "type": "text", + "content": "We compute the reconstruction loss between the original video " + }, + { + "bbox": [ + 55, + 563, + 295, + 586 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 55, + 563, + 295, + 586 + ], + "type": "text", + "content": " and the reconstructed video " + }, + { + "bbox": [ + 55, + 563, + 295, + 586 + ], + "type": "inline_equation", + "content": "U^{\\prime}" + }, + { + "bbox": [ + 55, + 563, + 295, + 586 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 596, + 294, + 610 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 596, + 294, + 610 + ], + "spans": [ + { + "bbox": [ + 131, + 596, + 294, + 610 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {v i d e o}} ^ {\\text {r e c}} = \\left| \\left| U ^ {\\prime} - U \\right| \\right| _ {2} ^ {2}. \\tag {5}", + "image_path": "f84cd0b873b90403bc1bff3fb89149dd6e8c7a0d177bb412b792dc6c13725866.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 620, + 295, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 620, + 295, + 655 + ], + "spans": [ + { + "bbox": [ + 55, + 620, + 295, + 655 + ], + "type": "text", + "content": "Similarly, we introduce a reconstruction loss for motion data, " + }, + { + "bbox": [ + 55, + 620, + 295, + 655 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{motion}}^{\\mathrm{rec}}" + }, + { + "bbox": [ + 55, + 620, + 295, + 655 + ], + "type": "text", + "content": ", following the same process. The total reconstruction loss is defined as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 126, + 666, + 295, + 680 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 666, + 295, + 680 + ], + "spans": [ + { + "bbox": [ + 126, + 666, + 295, + 680 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {\\text {r e c}} = \\mathcal {L} _ {\\text {v i d e o}} ^ {\\text {r e c}} + \\mathcal {L} _ {\\text {m o t i o n}} ^ {\\text {r e c}}. \\tag {6}", + "image_path": "c723c634a81765a4b41a49adb6bd2be950affaf73e1641786ecb836b39d8e5a3.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 689, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 712 + ], + "type": "text", + "content": "The above masked reconstruction loss ensures that the encoded features contain sufficient information about the" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 365, + 555, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 555, + 437 + ], + "type": "text", + "content": "video or motion input, while learning away possible nuisances through masking and making the learning of the concepts (codebook) more focused on critical information. On the other hand, the discretization imposed by the concepts further enhances the meaningfulness and robustness of the pre-trained multi-modal (motion) features." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 460, + 554, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 554, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 554, + 519 + ], + "type": "text", + "content": "Velocity Reconstruction. Reconstructing inputs from masked features inevitably results in the loss of high-frequency details, leading to temporally over-smoothed reconstruction outcomes and preventing the discovery of more nuanced yet useful motion concepts." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 521, + 554, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 554, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 554, + 640 + ], + "type": "text", + "content": "To address this limitation, we propose an objective that further reconstructs the \"velocity\" of the input sequence. To be more specific, we define \"state\" as the encoding or feature of a frame in the video or motion sequence, while the \"velocity\" means the difference between consecutive \"states\" that captures high-frequency information by focusing on dynamic changes. For video and motion data, we denote the \"states\" as " + }, + { + "bbox": [ + 313, + 521, + 554, + 640 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 313, + 521, + 554, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 521, + 554, + 640 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 521, + 554, + 640 + ], + "type": "text", + "content": ", and the \"velocity\" as the optical flow " + }, + { + "bbox": [ + 313, + 521, + 554, + 640 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 313, + 521, + 554, + 640 + ], + "type": "text", + "content": " (computed from image frames) and delta motion " + }, + { + "bbox": [ + 313, + 521, + 554, + 640 + ], + "type": "inline_equation", + "content": "\\delta M" + }, + { + "bbox": [ + 313, + 521, + 554, + 640 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "content": "However, directly reconstructing velocity from the same set of masked features introduces difficulty in figuring out the dynamic information, given the above-mentioned oversmoothing effect in the encoded features. To mitigate this, we draw inspiration from the robotic manipulation literature [46] on action concept discovery and incorporate two" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7182" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 106, + 69, + 504, + 182 + ], + "blocks": [ + { + "bbox": [ + 106, + 69, + 504, + 182 + ], + "lines": [ + { + "bbox": [ + 106, + 69, + 504, + 182 + ], + "spans": [ + { + "bbox": [ + 106, + 69, + 504, + 182 + ], + "type": "image", + "image_path": "b4033cb12be74bb6bf64101f5319199a9aa4a4861eb8fedd50f3c9f868542bd1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 194, + 555, + 251 + ], + "lines": [ + { + "bbox": [ + 54, + 194, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 54, + 194, + 555, + 251 + ], + "type": "text", + "content": "Figure 3. Overview of the velocity reconstruction components. We build similar network structures for both video and motion, and present the video part in this figure. This module is composed of two learning objectives. 1) Discriminative informativeness (left) aims to improve the distinctiveness of encoded features by reducing representational ambiguity. 2) Actionable informativeness (right) focuses on reconstructing the velocity by leveraging gradient information from the discrimination hypernetwork. As for the video data, we employ optical flow as the representation of the velocity." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 260, + 294, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 260, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 54, + 260, + 294, + 319 + ], + "type": "text", + "content": "tailored learning objectives - discriminative informativeness and actionable informativeness - to enhance the efficiency of representations and improve the quality of temporal dynamics in multi-modality pretaining, as shown in Fig. 3." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 321, + 296, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 321, + 296, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 321, + 296, + 475 + ], + "type": "text", + "content": "Discriminative informativeness. The discriminative informativeness objective aims to improve the distinctiveness of encoded features by increasing the correlation between the discovered concepts and the corresponding state features. Specifically, we introduce a video hyper-network " + }, + { + "bbox": [ + 55, + 321, + 296, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{H}^u" + }, + { + "bbox": [ + 55, + 321, + 296, + 475 + ], + "type": "text", + "content": ", which takes discrete features " + }, + { + "bbox": [ + 55, + 321, + 296, + 475 + ], + "type": "inline_equation", + "content": "c_k^u" + }, + { + "bbox": [ + 55, + 321, + 296, + 475 + ], + "type": "text", + "content": " from the codebook " + }, + { + "bbox": [ + 55, + 321, + 296, + 475 + ], + "type": "inline_equation", + "content": "C^u" + }, + { + "bbox": [ + 55, + 321, + 296, + 475 + ], + "type": "text", + "content": " as input and generates a classification network. The hyper-network " + }, + { + "bbox": [ + 55, + 321, + 296, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{H}^u" + }, + { + "bbox": [ + 55, + 321, + 296, + 475 + ], + "type": "text", + "content": " (together with the generated classifier) is trained to determine whether an input state is compatible with its corresponding code class, which is represented by a confidence value " + }, + { + "bbox": [ + 55, + 321, + 296, + 475 + ], + "type": "inline_equation", + "content": "s_u" + }, + { + "bbox": [ + 55, + 321, + 296, + 475 + ], + "type": "text", + "content": " (classification score) from 0 to 1, thereby enhancing the distinguishability and separation of each feature class in the representation space." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 477, + 257, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 477, + 257, + 489 + ], + "spans": [ + { + "bbox": [ + 67, + 477, + 257, + 489 + ], + "type": "text", + "content": "The discriminative loss for video is defined as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 89, + 502, + 295, + 537 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 502, + 295, + 537 + ], + "spans": [ + { + "bbox": [ + 89, + 502, + 295, + 537 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {v i d e o}} ^ {\\text {d i s}} = \\sum_ {k = 0} ^ {| C ^ {u} | - 1} \\sum_ {i = 0} ^ {T - 1} C E \\left(\\mathcal {H} ^ {u} \\left(c _ {k} ^ {u}, u _ {i}\\right), d _ {i} ^ {u}\\right), \\tag {7}", + "image_path": "e7af1803ee6f9f23f7683b32e3c33c700b441ff678b9961959f7904fd7754227.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 544, + 295, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 592 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 544, + 295, + 592 + ], + "type": "inline_equation", + "content": "|C^u|" + }, + { + "bbox": [ + 55, + 544, + 295, + 592 + ], + "type": "text", + "content": " is the size of the codebook " + }, + { + "bbox": [ + 55, + 544, + 295, + 592 + ], + "type": "inline_equation", + "content": "C^u" + }, + { + "bbox": [ + 55, + 544, + 295, + 592 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 544, + 295, + 592 + ], + "type": "inline_equation", + "content": "CE(\\cdot)" + }, + { + "bbox": [ + 55, + 544, + 295, + 592 + ], + "type": "text", + "content": " is the cross-entropy loss for classification, " + }, + { + "bbox": [ + 55, + 544, + 295, + 592 + ], + "type": "inline_equation", + "content": "d_i^u" + }, + { + "bbox": [ + 55, + 544, + 295, + 592 + ], + "type": "text", + "content": " represents the \"ground-truth label\" for the input state " + }, + { + "bbox": [ + 55, + 544, + 295, + 592 + ], + "type": "inline_equation", + "content": "u_i" + }, + { + "bbox": [ + 55, + 544, + 295, + 592 + ], + "type": "text", + "content": " determined by the VQ-VAE, and " + }, + { + "bbox": [ + 55, + 544, + 295, + 592 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 544, + 295, + 592 + ], + "type": "text", + "content": " is the number of frames." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 594, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 594, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 594, + 296, + 714 + ], + "type": "text", + "content": "Actionable informativeness. The actionable informativeness objective focuses on reconstructing the velocity by leveraging gradient information from the discriminative hyper-network, " + }, + { + "bbox": [ + 54, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{H}^u" + }, + { + "bbox": [ + 54, + 594, + 296, + 714 + ], + "type": "text", + "content": " introduced above. Inspired by the idea that gradients of a discriminator function can inform the state changes, as observed in [46], this objective ensures that the model captures not only static state information, but also the dynamic changes in the sequence. By leveraging gradient fluctuations around a given state, the approach effectively captures dynamic motion details, such as velocity," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 260, + 553, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 260, + 553, + 284 + ], + "spans": [ + { + "bbox": [ + 313, + 260, + 553, + 284 + ], + "type": "text", + "content": "resulting in more fine-grained motion concepts, and enriching the motion representation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 285, + 554, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 554, + 344 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 554, + 344 + ], + "type": "text", + "content": "Specifically, we initialize a network " + }, + { + "bbox": [ + 313, + 285, + 554, + 344 + ], + "type": "inline_equation", + "content": "\\mathcal{A}^u" + }, + { + "bbox": [ + 313, + 285, + 554, + 344 + ], + "type": "text", + "content": " for video data, which takes key image frames " + }, + { + "bbox": [ + 313, + 285, + 554, + 344 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 313, + 285, + 554, + 344 + ], + "type": "text", + "content": " and gradients from the discriminator " + }, + { + "bbox": [ + 313, + 285, + 554, + 344 + ], + "type": "inline_equation", + "content": "\\nabla_U\\mathcal{H}^u (D^u,U)" + }, + { + "bbox": [ + 313, + 285, + 554, + 344 + ], + "type": "text", + "content": " as inputs and predicts the velocity " + }, + { + "bbox": [ + 313, + 285, + 554, + 344 + ], + "type": "inline_equation", + "content": "V_{u}^{\\prime}" + }, + { + "bbox": [ + 313, + 285, + 554, + 344 + ], + "type": "text", + "content": ". The actionable informativeness loss for video is then computed as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 388, + 354, + 553, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 354, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 388, + 354, + 553, + 369 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {v i d e o}} ^ {\\text {a c t}} = \\left| \\left| V _ {u} - V _ {u} ^ {\\prime} \\right| \\right| _ {2} ^ {2}. \\tag {8}", + "image_path": "2ac670911331acb847d6a0b18796e3302bea599f4b722452dbcba09c03873689.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 378, + 553, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 553, + 415 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 378, + 553, + 415 + ], + "type": "inline_equation", + "content": "V_{u}" + }, + { + "bbox": [ + 313, + 378, + 553, + 415 + ], + "type": "text", + "content": " represents the ground-truth velocity of video frames. This process is repeated for motion data to compute the corresponding loss " + }, + { + "bbox": [ + 313, + 378, + 553, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{motion}}^{\\mathrm{act}}" + }, + { + "bbox": [ + 313, + 378, + 553, + 415 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 313, + 378, + 553, + 415 + ], + "type": "inline_equation", + "content": "V_{m}" + }, + { + "bbox": [ + 313, + 378, + 553, + 415 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 378, + 553, + 415 + ], + "type": "inline_equation", + "content": "V_{m}^{\\prime}" + }, + { + "bbox": [ + 313, + 378, + 553, + 415 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 415, + 554, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 554, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 554, + 462 + ], + "type": "text", + "content": "Combined Objectives. By integrating both video and motion components, we derive the total velocity reconstruction loss for human motion concept discovery and motion feature pre-training as in the following:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 387, + 471, + 553, + 487 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 387, + 471, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 387, + 471, + 553, + 487 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {\\text {d i s}} = \\mathcal {L} _ {\\text {v i d e o}} ^ {\\text {d i s}} + \\mathcal {L} _ {\\text {m o t i o n}} ^ {\\text {d i s}} \\tag {9}", + "image_path": "2058a83ad9584b7041fa9ee11586b8f6b52b1aa3bd03be125ddebde288d74cf2.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 388, + 509, + 553, + 523 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 509, + 553, + 523 + ], + "spans": [ + { + "bbox": [ + 388, + 509, + 553, + 523 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {\\mathrm {a c t}} = \\mathcal {L} _ {\\text {v i d e o}} ^ {\\mathrm {a c t}} + \\mathcal {L} _ {\\text {m o t i o n}} ^ {\\mathrm {a c t}} \\tag {10}", + "image_path": "b8f6896960c5a85e1140b6e5d6268f2ad21cec876d8c0aa90050f45a18b0ca22.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 529, + 554, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 554, + 577 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 554, + 577 + ], + "type": "text", + "content": "With the above objectives, our approach ensures that the model effectively reconstructs high-frequency details, improving the temporal fidelity and overall quality of multimodal representations." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "Feature Alignment. Multi-modality models often leverage the shared information across different modalities to enhance performance [5, 56, 87]. In the case of motion and video, video data provides contextual information that motion data alone may be lacking, which is critical for accurate motion prediction and understanding. Conversely, motion data focuses on human-centric features in the video, reducing over-reliance on environmental details. An efficient utilization or balancing of both is beneficial for the human motion LLMs. However, prior work on human motion" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7183" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "type": "text", + "content": "and video understanding [5] only achieves implicit alignment through paired data during LLM fine-tuning, without explicit alignment in encoder training. To address this issue, we propose an alignment loss to improve information sharing and cross-modal coherence." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 133, + 294, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 294, + 215 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 294, + 215 + ], + "type": "text", + "content": "To achieve alignment between video and motion features, we collect a set of video-motion pairs from Motion-X [41], which are aligned by frames. We instantiate two projection layers: The video projection layer " + }, + { + "bbox": [ + 55, + 133, + 294, + 215 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^u" + }, + { + "bbox": [ + 55, + 133, + 294, + 215 + ], + "type": "text", + "content": " and the motion projection layer " + }, + { + "bbox": [ + 55, + 133, + 294, + 215 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^m" + }, + { + "bbox": [ + 55, + 133, + 294, + 215 + ], + "type": "text", + "content": ". These layers are designed to map the respective features from the video and motion domains into a shared aligned space." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 216, + 294, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 216, + 294, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 216, + 294, + 289 + ], + "type": "text", + "content": "More explicitly, given paired video and motion data " + }, + { + "bbox": [ + 55, + 216, + 294, + 289 + ], + "type": "inline_equation", + "content": "\\{U,M\\}" + }, + { + "bbox": [ + 55, + 216, + 294, + 289 + ], + "type": "text", + "content": ", we use " + }, + { + "bbox": [ + 55, + 216, + 294, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^u" + }, + { + "bbox": [ + 55, + 216, + 294, + 289 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 216, + 294, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^m" + }, + { + "bbox": [ + 55, + 216, + 294, + 289 + ], + "type": "text", + "content": " to project the discrete video feature " + }, + { + "bbox": [ + 55, + 216, + 294, + 289 + ], + "type": "inline_equation", + "content": "d_i^u" + }, + { + "bbox": [ + 55, + 216, + 294, + 289 + ], + "type": "text", + "content": " and motion feature " + }, + { + "bbox": [ + 55, + 216, + 294, + 289 + ], + "type": "inline_equation", + "content": "d_j^m" + }, + { + "bbox": [ + 55, + 216, + 294, + 289 + ], + "type": "text", + "content": " into the aligned feature space, resulting in " + }, + { + "bbox": [ + 55, + 216, + 294, + 289 + ], + "type": "inline_equation", + "content": "a_i^u" + }, + { + "bbox": [ + 55, + 216, + 294, + 289 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 216, + 294, + 289 + ], + "type": "inline_equation", + "content": "a_j^m" + }, + { + "bbox": [ + 55, + 216, + 294, + 289 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 216, + 294, + 289 + ], + "type": "inline_equation", + "content": "a_i^u, a_j^m \\in \\mathbb{R}^H" + }, + { + "bbox": [ + 55, + 216, + 294, + 289 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 55, + 216, + 294, + 289 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 216, + 294, + 289 + ], + "type": "text", + "content": " denotes the dimensionality of the aligned features in the shared space." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 289, + 294, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 294, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 294, + 312 + ], + "type": "text", + "content": "To ensure effective alignment, we define an alignment loss as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 73, + 317, + 294, + 352 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 317, + 294, + 352 + ], + "spans": [ + { + "bbox": [ + 73, + 317, + 294, + 352 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {\\text {a l i g n}} = \\sum_ {0} ^ {T - 1} \\frac {\\operatorname {n o r m} \\left(a _ {i} ^ {u}\\right) \\cdot \\operatorname {n o r m} \\left(a _ {\\text {a l i g n} (i)} ^ {m}\\right) / \\epsilon}{\\sum_ {0} ^ {K - 1} \\operatorname {n o r m} \\left(a _ {i} ^ {u}\\right) \\cdot \\operatorname {n o r m} \\left(a _ {j} ^ {m}\\right) / \\epsilon}, \\tag {11}", + "image_path": "f4d47d320900a002ef509193e58282ef2ac19bc1110c3747d43a4dc974a8e1af.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 357, + 294, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 357, + 294, + 416 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 294, + 416 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 357, + 294, + 416 + ], + "type": "inline_equation", + "content": "\\mathrm{align}(i)" + }, + { + "bbox": [ + 55, + 357, + 294, + 416 + ], + "type": "text", + "content": " indicates the pose index in the motion sequence corresponding to the key video frame " + }, + { + "bbox": [ + 55, + 357, + 294, + 416 + ], + "type": "inline_equation", + "content": "u_{i}" + }, + { + "bbox": [ + 55, + 357, + 294, + 416 + ], + "type": "text", + "content": ", norm represents a normalization operation to stabilize feature magnitudes and " + }, + { + "bbox": [ + 55, + 357, + 294, + 416 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 357, + 294, + 416 + ], + "type": "text", + "content": " is a temperature parameter that controls the sharpness of the alignment distribution." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 417, + 294, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 417, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 417, + 294, + 453 + ], + "type": "text", + "content": "Note that " + }, + { + "bbox": [ + 55, + 417, + 294, + 453 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{\\mathrm{align}}" + }, + { + "bbox": [ + 55, + 417, + 294, + 453 + ], + "type": "text", + "content": " is computed only when motion and video are paired. For standalone motion or video inputs, this loss term is omitted for applicability." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 453, + 294, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 453, + 294, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 294, + 475 + ], + "type": "text", + "content": "The overall human motion concept discovery objective is defined as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 94, + 483, + 294, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 483, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 94, + 483, + 294, + 514 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} ^ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {m o t i o n}} ^ {\\text {r e c}} + \\mathcal {L} _ {\\text {v i d e o}} ^ {\\text {r e d e}} + \\\\ \\lambda^ {\\text {d i s}} \\mathcal {L} ^ {\\text {d i s}} + \\lambda^ {\\text {a c t}} \\mathcal {L} ^ {\\text {a c t}} + \\lambda^ {\\text {a l i g n}} \\mathcal {L} ^ {\\text {a l i g n}}, \\tag {12} \\\\ \\end{array}", + "image_path": "72a4d58aef686f48b228fa64ead21d802d00e5d60ed9c1b78def307cca34bb49.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 521, + 294, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 294, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 294, + 545 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 521, + 294, + 545 + ], + "type": "inline_equation", + "content": "\\lambda^{\\mathrm{dis}}" + }, + { + "bbox": [ + 55, + 521, + 294, + 545 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 521, + 294, + 545 + ], + "type": "inline_equation", + "content": "\\lambda^{\\mathrm{act}}" + }, + { + "bbox": [ + 55, + 521, + 294, + 545 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 521, + 294, + 545 + ], + "type": "inline_equation", + "content": "\\lambda^{\\mathrm{align}}" + }, + { + "bbox": [ + 55, + 521, + 294, + 545 + ], + "type": "text", + "content": " are weighting coefficients for the respective losses " + }, + { + "bbox": [ + 55, + 521, + 294, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{\\mathrm{dis}}" + }, + { + "bbox": [ + 55, + 521, + 294, + 545 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 521, + 294, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{\\mathrm{act}}" + }, + { + "bbox": [ + 55, + 521, + 294, + 545 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 521, + 294, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{\\mathrm{align}}" + }, + { + "bbox": [ + 55, + 521, + 294, + 545 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 552, + 177, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 177, + 565 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 177, + 565 + ], + "type": "text", + "content": "3.3. Modality Translation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 570, + 294, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 294, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 294, + 629 + ], + "type": "text", + "content": "Although pre-trained encoders capture rich features from the original input, these features are often difficult for LLMs to interpret directly [5, 37, 40]. Therefore, it is essential to introduce an additional projection layer to map these features into a space that LLMs can effectively understand." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 630, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 294, + 713 + ], + "type": "text", + "content": "As illustrated in Fig 2, we add tunable translation layers after the encoders " + }, + { + "bbox": [ + 55, + 630, + 294, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{E}^u" + }, + { + "bbox": [ + 55, + 630, + 294, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 630, + 294, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{E}^m" + }, + { + "bbox": [ + 55, + 630, + 294, + 713 + ], + "type": "text", + "content": ". These layers are trained using annotated video and motion sequence data to learn task-specific transformations. During this stage, the encoders and the LLM " + }, + { + "bbox": [ + 55, + 630, + 294, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 55, + 630, + 294, + 713 + ], + "type": "text", + "content": " are kept frozen to prevent adverse feature drift, and only the weights of the translation layers are updated. This strategy preserves the pre-trained knowledge of" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "the encoders and language model while optimizing the feature mapping for downstream tasks." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 104, + 500, + 118 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 104, + 500, + 118 + ], + "spans": [ + { + "bbox": [ + 313, + 104, + 500, + 118 + ], + "type": "text", + "content": "3.4. Multi-Modality Instruction Tuning" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 122, + 553, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 122, + 553, + 193 + ], + "spans": [ + { + "bbox": [ + 313, + 122, + 553, + 193 + ], + "type": "text", + "content": "Finally, we employ instruction tuning to train the language model " + }, + { + "bbox": [ + 313, + 122, + 553, + 193 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 313, + 122, + 553, + 193 + ], + "type": "text", + "content": " to respond to user queries using the corresponding video or motion sequences as context. During this process, the encoders and projection layers remain frozen, ensuring their pre-trained features are retained, while the weights of the language model " + }, + { + "bbox": [ + 313, + 122, + 553, + 193 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 313, + 122, + 553, + 193 + ], + "type": "text", + "content": " are updated to accommodate the task." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 194, + 553, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 194, + 553, + 255 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 553, + 255 + ], + "type": "text", + "content": "In this stage, the model can further enhance cross-modal alignment using paired motion-video data in an implicit manner. To streamline the fine-tuning process, we adopt LoRA [29] with a rank of 8, allowing efficient and lightweight adjustment to the language model." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 266, + 394, + 278 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 394, + 278 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 394, + 278 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 285, + 436, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 436, + 298 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 436, + 298 + ], + "type": "text", + "content": "4.1. Experiments Settings" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 303, + 553, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 303, + 553, + 494 + ], + "spans": [ + { + "bbox": [ + 313, + 303, + 553, + 494 + ], + "type": "text", + "content": "Implementation details For videos and motions, we initialize a codebook with 512 codes, a one-layer projector for feature alignment, and a 2-layer MLP to generate the weights for a 2-layer discriminator separately. For video inputs, we utilize pre-trained SigLIP [79] with an image patch size of 16 pixels and hidden dimension of 256 as the video encoder, and we take a 12-layer transformer as the reconstruction decoder. A 4-layer transformer is utilized as the motion encoder and a 2-layer transformer is used for motion reconstruction. For the velocity reconstruction, we initialize a two-layer hypernet and two-layer velocity decoder. Besides, we utilize one-layer MLPs for feature alignment. We pre-train the motion encoder with only the reconstruction loss for 60K iterations with a learning rate of 1e-4. Then, we train both the video and motion encoders with the total loss for 8K iterations with a learning rate of 1e-5." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 495, + 553, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 553, + 578 + ], + "type": "text", + "content": "We utilize Llama3.1-8B as our base LLM model. For the modality translation stage, we initialize a one-layer projector for the video and motion encoder separately. We train the projectors on the annotation dataset for one batch with a batch size of 32 and a learning rate of 1e-4. In the Multimodality instruction tuning stage, we train the LLM model with a batch size of 16 and a learning rate of 3e-4." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "Training datasets For the encoder pre-training stage, we utilize Motion-X [41] datasets which contain 81K motion sequences and 36K motion-video pairs. In order to facilitate the training process of LLMs, we follow MotionLLM [5] to prompt LLMs to generate 20K annotation for motion-video pairs from the Motion-X dataset, and 10K annotation data from the HumanML3D [21] dataset. This data is utilized for modality translation. We additionally utilize the VATEX [69] dataset which contains 20K annotated videos for video translator training. For multi-modality instruction" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7184" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 96, + 112, + 512, + 213 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "type": "text", + "content": "Table 1. Comparison of different methods on BABEL-QA test set. The pred type represent the prediction type, including classification and generation. For 2s-AGCN and MotionCLIP, 'M' means MLP, 'R' means RNN. Our method outperforms the baselines across all the metrics." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 96, + 112, + 512, + 213 + ], + "lines": [ + { + "bbox": [ + 96, + 112, + 512, + 213 + ], + "spans": [ + { + "bbox": [ + 96, + 112, + 512, + 213 + ], + "type": "table", + "html": "
ModelPred typeOverallQuery typeRelation Type
ActionDirectionBodyPartBeforeAfterOther
2s-AGCN-Mcls.0.3550.3840.3520.2280.3310.2640.295
2s-AGCN-Rcls.0.3570.3960.3520.1940.3370.3010.285
MotionCLIP-Mcls.0.4300.4850.3610.2720.3720.3210.404
MotionCLIP-Rcls.0.4200.4890.3100.2500.3980.3140.387
MotionLLMgen.0.4360.5170.3540.1540.4270.3680.529
Oursgen.0.7110.8090.6970.6230.7070.6350.797
", + "image_path": "bf1f754f722ccd4b33f8a5870fb4a715d774381e66e38192bb623b6f01a00e33.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 221, + 294, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 221, + 294, + 283 + ], + "spans": [ + { + "bbox": [ + 55, + 221, + 294, + 283 + ], + "type": "text", + "content": "tuning, we follow MotionLLM to generate 200K Q&A pairs for both motions and videos. We also utilize the VideoChatGPT [49] and BABEL-QA [13] datasets to enhance the motion and video understanding ability of our model, which has 100K and 2K Q&A data separately." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 303, + 296, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 303, + 296, + 411 + ], + "spans": [ + { + "bbox": [ + 55, + 303, + 296, + 411 + ], + "type": "text", + "content": "Evaluation datasets Although MotionLLM [5] creates a large benchmark for both motion and video understanding, by the time we completed our paper, they had not released the specific details or access methods for the benchmark, nor did they provide a clear timeline for its release. Therefore, we did not perform comparisons on that benchmark but instead selected the test set of ActivityNet-QA [78] and BABEL-QA [13] to separately validate the performance of our algorithm on video and motion data." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 432, + 296, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 432, + 296, + 564 + ], + "spans": [ + { + "bbox": [ + 55, + 432, + 296, + 564 + ], + "type": "text", + "content": "Evaluation metrics On the BABEL-QA benchmark, we evaluate the prediction accuracy for evaluation. As for the classification-based baselines, we evaluate the accuracy directly. As for the generative baselines, we adapt the evaluation protocol used in previous works [5, 37] by utilizing LLMs. Specifically, the evaluation LLM is given some examples of questions, answers, and scores (from 0 to 1), and it is required to follow these examples and score the experiment results. On the Activity-QA benchmark, we only compare our methods with our generative baselines. Thus, we directly utilize LLMs to evaluate the methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 575, + 174, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 575, + 174, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 575, + 174, + 588 + ], + "type": "text", + "content": "4.2. Quantitative Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": "Experiments on ActivityNet-QA. We evaluate the video understanding ability of our method and other baselines on the ActivityNet-QA benchmark as shown in Tab. 2. We follow previous methods [5, 33, 38] to report the accuracy and scores that are evaluated by LLMs. We show in Tab. 2 that our method outperforms all the baselines on both metrics. Compared to the video-only method, our method aligns the video feature with the motion feature, making the encoder concentrate on human-centric information. Compared to MotionLLM, our encoder pre-training helps the network to" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 222, + 553, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 222, + 553, + 246 + ], + "spans": [ + { + "bbox": [ + 313, + 222, + 553, + 246 + ], + "type": "text", + "content": "learn more fine-grained information through velocity reconstruction." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 261, + 555, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 261, + 555, + 405 + ], + "spans": [ + { + "bbox": [ + 313, + 261, + 555, + 405 + ], + "type": "text", + "content": "Experiments on BABEL-QA. We evaluate the motion understanding ability of our method as compared to existing studies on the BABEL-QA [13] dataset. Following [5, 13], we compare our method with a diverse set of baselines: 1) 2s-AGCN [58], a method using 2s-GCN to extract motion features and classify motion by MLP or RNN; 2) Motion-CLIP [63], a method that aligns the features of motion and text using CLIP [55] and predicts motion class by MLP or RNN; 3) MotionLLM [5], an LLM-based method which has motion reasoning ability. We follow [5] to fine-tune our method with the training set of BABEL-QA and evaluate our method with LLMs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 405, + 556, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 405, + 556, + 549 + ], + "spans": [ + { + "bbox": [ + 313, + 405, + 556, + 549 + ], + "type": "text", + "content": "From Tab. 1, our method outperforms other methods in all metrics by a large margin of nearly " + }, + { + "bbox": [ + 313, + 405, + 556, + 549 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 313, + 405, + 556, + 549 + ], + "type": "text", + "content": " overall. In the \"bodypart\" query task, our algorithm achieves 2.29 times the score of the second-place method, MotionCLIP-M (0.623 vs. 0.272). This advantage stems partly from our velocity-based encoder training, which, compared to traditional motion feature extraction, helps us better capture high-frequency information and ensures the learning of fine-grained features. Additionally, unlike MotionLLM, we perform alignment at the encoder pre-training stage, ensuring that the features learned by the encoder are inherently more semantically rich." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 340, + 587, + 529, + 712 + ], + "blocks": [ + { + "bbox": [ + 318, + 567, + 550, + 578 + ], + "lines": [ + { + "bbox": [ + 318, + 567, + 550, + 578 + ], + "spans": [ + { + "bbox": [ + 318, + 567, + 550, + 578 + ], + "type": "text", + "content": "Table 2. Experiments results on ActivityNet-QA benchmark." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 340, + 587, + 529, + 712 + ], + "lines": [ + { + "bbox": [ + 340, + 587, + 529, + 712 + ], + "spans": [ + { + "bbox": [ + 340, + 587, + 529, + 712 + ], + "type": "table", + "html": "
ModelAcc↑Score ↑
FrozenBiLM [57]24.7-
VideoChat [38]-2.2
LLaMA-Adapter [84]34.22.7
Video-LLaMA [80]12.41.1
Video-ChatGPT [49]35.22.7
Video-LLaVA [40]45.33.3
VideoChat2 [39]49.13.3
MotionLLM [5]53.33.5
Ours54.23.6
", + "image_path": "8cf8d7e77932952657d5f70387a90d0e28a163086d1e0b69ca21499244939bd9.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7185" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 86, + 90, + 523, + 178 + ], + "blocks": [ + { + "bbox": [ + 217, + 71, + 394, + 82 + ], + "lines": [ + { + "bbox": [ + 217, + 71, + 394, + 82 + ], + "spans": [ + { + "bbox": [ + 217, + 71, + 394, + 82 + ], + "type": "text", + "content": "Table 3. Ablation study on BABEL-QA test set." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 86, + 90, + 523, + 178 + ], + "lines": [ + { + "bbox": [ + 86, + 90, + 523, + 178 + ], + "spans": [ + { + "bbox": [ + 86, + 90, + 523, + 178 + ], + "type": "table", + "html": "
ModelPred typeOverallQuery typeRelation Type
ActionDirectionBodyPartBeforeAfterOther
MotionLLMgen.0.4360.5170.3540.1540.4270.3680.529
Ours-w/oLrecgen.0.6960.7410.6450.5770.6000.5970.762
Ours-w/oLdis & Lactgen.0.6370.6930.4780.6060.6670.5260.709
Ours-w/oLaligngen.0.6750.7430.5790.5230.5840.5700.743
Oursgen.0.7110.8090.6970.6230.7070.6350.797
", + "image_path": "bd0306a1cfcc25e3dd319628632e417252bd03f2fc0c7a6c7dbae1a88fa52ed2.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 86, + 180, + 523, + 365 + ], + "blocks": [ + { + "bbox": [ + 86, + 180, + 523, + 365 + ], + "lines": [ + { + "bbox": [ + 86, + 180, + 523, + 365 + ], + "spans": [ + { + "bbox": [ + 86, + 180, + 523, + 365 + ], + "type": "image", + "image_path": "592a68e5187fcb85cf5fb369af3df6751909641baebc47d0657f266ac6a8741c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 374, + 555, + 409 + ], + "lines": [ + { + "bbox": [ + 55, + 374, + 555, + 409 + ], + "spans": [ + { + "bbox": [ + 55, + 374, + 555, + 409 + ], + "type": "text", + "content": "Figure 4. Example of motion understanding. We validate the proposed model's ability to understand motion from multiple aspects. Q1 on the left and Q1-3 on the right evaluate its comprehension of motion sequences, while Q2 and Q3 on the left push it to analyze kinematic, kinesthetic, and physical properties. These results demonstrate our algorithm's ability to recognize, describe in detail, and analyze motions." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 418, + 297, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 297, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 297, + 586 + ], + "type": "text", + "content": "Ablation study. We present the ablation results in Tab. 3. We conducted ablation studies on masked reconstruction, velocity reconstruction, and feature alignment. Motion-LLM serves as an ablation baseline without any of these techniques. As presented, the absence of any single module leads to a decrease in performance. Among these, masked reconstruction has the least impact, as we initially train the motion encoder solely with masked reconstruction loss to ensure stable subsequent training. Removing velocity reconstruction has the greatest impact on pexrformance since velocity provides high-frequency information that other modules lack. Likewise, removing the alignment loss significantly affects results, highlighting the importance of multi-modality alignment in motion understanding." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 598, + 167, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 598, + 167, + 611 + ], + "spans": [ + { + "bbox": [ + 55, + 598, + 167, + 611 + ], + "type": "text", + "content": "4.3. Qualitative Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 618, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 296, + 715 + ], + "type": "text", + "content": "To validate our algorithm's ability to understand motions and its contextual knowledge, we conducted multiple rounds of Q&A with the motion model. We present two examples in Fig. 4. We discover that, beyond its basic motion recognition capabilities, our algorithm can provide detailed descriptions of motions. In the example on the right, it can precisely describe the actions of the left hand and foot, as well as the right hand and foot. Our algorithm also" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 418, + 555, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 555, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 555, + 502 + ], + "type": "text", + "content": "possesses extensive knowledge about various movements. For instance, when asked, \"what muscles does the person engage when bending their elbows during a push-up?\" it can accurately respond with \"triceps muscles.\" This demonstrates that our algorithm can understand the input motion sequences and perform corresponding analyses. We present more qualitative results in the supplementary material." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 513, + 388, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 513, + 388, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 513, + 388, + 525 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 533, + 556, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 556, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 556, + 689 + ], + "type": "text", + "content": "In this paper, we presented HuMoCon, a novel motion-video understanding framework for human action understanding and concept discovery. We demonstrated the effectiveness of our approach through comprehensive experiments, showcasing major improvements over state-of-the-art methods. Our work provides a strong foundation for future research in human action understanding and motion concept discovery. There are still some drawbacks of our work, such as the lack of human-environment contact or understanding niche motions. We plan to address those limitations by expanding the framework to incorporate additional modalities, particularly as environments, objects, and contact, and curate a wider range of specialized datasets." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7186" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 141, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 141, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 141, + 85 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 297, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 297, + 137 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 297, + 137 + ], + "type": "text", + "content": "This work is supported by the Early Career Scheme of the Research Grants Council (grant # 27207224), the HKU-100 Award, a donation from the Musketeers Foundation, and an Academic Gift from Meta." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 148, + 115, + 160 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 148, + 115, + 160 + ], + "spans": [ + { + "bbox": [ + 56, + 148, + 115, + 160 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 167, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 167, + 296, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 167, + 296, + 209 + ], + "spans": [ + { + "bbox": [ + 61, + 167, + 296, + 209 + ], + "type": "text", + "content": "[1] Nikos Athanasiou, Mathis Petrovich, Michael J Black, and Gúl Varol. Teach: Temporal action composition for 3d humans. In 2022 International Conference on 3D Vision (3DV), pages 414-423. IEEE, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 212, + 296, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 212, + 296, + 266 + ], + "spans": [ + { + "bbox": [ + 61, + 212, + 296, + 266 + ], + "type": "text", + "content": "[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 1(2):3, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 268, + 295, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 268, + 295, + 322 + ], + "spans": [ + { + "bbox": [ + 61, + 268, + 295, + 322 + ], + "type": "text", + "content": "[3] Yang Bai, Desen Zhou, Songyang Zhang, Jian Wang, Errui Ding, Yu Guan, Yang Long, and Jingdong Wang. Action quality assessment with temporal parsing transformer. In European conference on computer vision, pages 422-438. Springer, 2022. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 323, + 295, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 323, + 295, + 367 + ], + "spans": [ + { + "bbox": [ + 61, + 323, + 295, + 367 + ], + "type": "text", + "content": "[4] Pablo Cervantes, Yusuke Sekikawa, Ikuro Sato, and Koichi Shinoda. Implicit neural representations for variable length human motion generation. In European Conference on Computer Vision, pages 356-372. Springer, 2022. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 369, + 296, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 369, + 296, + 421 + ], + "spans": [ + { + "bbox": [ + 62, + 369, + 296, + 421 + ], + "type": "text", + "content": "[5] Ling-Hao Chen, Shunlin Lu, Ailing Zeng, Hao Zhang, Benyou Wang, Ruimao Zhang, and Lei Zhang. Motionllm: Understanding human behaviors from human motions and videos. arXiv preprint arXiv:2405.20340, 2024. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 424, + 296, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 424, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 62, + 424, + 296, + 456 + ], + "type": "text", + "content": "[6] Steven Chen and Richard R Yang. Pose trainer: correcting exercise posture using pose estimation. arXiv preprint arXiv:2006.11718, 2020. 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 458, + 296, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 458, + 296, + 501 + ], + "spans": [ + { + "bbox": [ + 62, + 458, + 296, + 501 + ], + "type": "text", + "content": "[7] Sihan Chen, Xingjian He, Longteng Guo, Xinxin Zhu, Weining Wang, Jinhui Tang, and Jing Liu. Valor: Vision-audiolanguage omni-perception pretraining model and dataset. arXiv preprint arXiv:2304.08345, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 502, + 296, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 502, + 296, + 556 + ], + "spans": [ + { + "bbox": [ + 62, + 502, + 296, + 556 + ], + "type": "text", + "content": "[8] Sihan Chen, Handong Li, Qunbo Wang, Zijia Zhao, Mingzhen Sun, Xinxin Zhu, and Jing Liu. Vast: A vision-audio-subtitle-text omni-modality foundation model and dataset. Advances in Neural Information Processing Systems, 36:72842-72866, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 558, + 296, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 558, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 62, + 558, + 296, + 612 + ], + "type": "text", + "content": "[9] Ginger Delmas, Philippe Weinzaepfel, Francesc Moreno-Noguer, and Grégory Rogez. Posefix: correcting 3d human poses with natural language. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15018-15028, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 614, + 296, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 614, + 296, + 668 + ], + "spans": [ + { + "bbox": [ + 57, + 614, + 296, + 668 + ], + "type": "text", + "content": "[10] Piotr Dólár, Vincent Rabaud, Garrison Cottrell, and Serge Belongie. Behavior recognition via sparse spatio-temporal features. In 2005 IEEE international workshop on visual surveillance and performance evaluation of tracking and surveillance, pages 65-72. IEEE, 2005. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "type": "text", + "content": "[11] Yong Du, Wei Wang, and Liang Wang. Hierarchical recurrent neural network for skeleton based action recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1110-1118, 2015. 1, 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 127 + ], + "type": "text", + "content": "[12] Benjamin Elizalde, Soham Deshmukh, Mahmoud Al Ismail, and Huaming Wang. Clap learning audio concepts from natural language supervision. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 129, + 555, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 555, + 171 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 555, + 171 + ], + "type": "text", + "content": "[13] Mark Endo, Joy Hsu, Jiaman Li, and Jiajun Wu. Motion question answering via modular motion programs. In International Conference on Machine Learning, pages 9312-9328. PMLR, 2023. 2, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 173, + 554, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 173, + 554, + 228 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 554, + 228 + ], + "type": "text", + "content": "[14] Qihang Fang, Yingda Yin, Qingnan Fan, Fei Xia, Siyan Dong, Sheng Wang, Jue Wang, Leonidas J Guibas, and Baoquan Chen. Towards accurate active camera localization. In European Conference on Computer Vision, pages 122-139. Springer, 2022. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 228, + 554, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 554, + 280 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 554, + 280 + ], + "type": "text", + "content": "[15] Qihang Fang, Chengcheng Tang, Bugra Tekin, and Yanchao Yang. Cigtime: Corrective instruction generation through inverse motion editing. In The Thirty-Eighth Annual Conference on Neural Information Processing Systems, 2024. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 282, + 554, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 282, + 554, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 282, + 554, + 327 + ], + "type": "text", + "content": "[16] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 327, + 554, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 554, + 380 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 554, + 380 + ], + "type": "text", + "content": "[17] Yao Feng, Jing Lin, Sai Kumar Dwivedi, Yu Sun, Priyanka Patel, and Michael J Black. Chatpose: Chatting about 3d human pose. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2093-2103, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 383, + 554, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 383, + 554, + 437 + ], + "spans": [ + { + "bbox": [ + 317, + 383, + 554, + 437 + ], + "type": "text", + "content": "[18] Mihai Fieraru, Mihai Zanfir, Silviu Cristian Pirlela, Vlad Olaru, and Cristian Sminchisescu. Aifit: Automatic 3d human-interpretable feedback models for fitness training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9919–9928, 2021. 1, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 438, + 554, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 438, + 554, + 501 + ], + "spans": [ + { + "bbox": [ + 317, + 438, + 554, + 501 + ], + "type": "text", + "content": "[19] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15180-15190, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 504, + 554, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 504, + 554, + 558 + ], + "spans": [ + { + "bbox": [ + 317, + 504, + 554, + 558 + ], + "type": "text", + "content": "[20] Chuan Guo, Xinxin Zuo, Sen Wang, Shihao Zou, Qingyao Sun, Annan Deng, Minglun Gong, and Li Cheng. Action2motion: Conditioned generation of 3d human motions. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2021-2029, 2020. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 559, + 554, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 559, + 554, + 613 + ], + "spans": [ + { + "bbox": [ + 317, + 559, + 554, + 613 + ], + "type": "text", + "content": "[21] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5152-5161, 2022. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 614, + 554, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 554, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 554, + 669 + ], + "type": "text", + "content": "[22] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5152-5161, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "type": "text", + "content": "[23] Chuan Guo, Xinxin Zuo, Sen Wang, and Li Cheng. Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In European Conference on Computer Vision, pages 580-597. Springer, 2022. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7187" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[24] Tianyu Guo, Hong Liu, Zhan Chen, Mengyuan Liu, Tao Wang, and Runwei Ding. Contrastive learning from extremely augmented skeleton sequences for self-supervised action recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 762-770, 2022. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 128, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 294, + 183 + ], + "type": "text", + "content": "[25] Tengda Han, Max Bain, Arsha Nagrani, Gul Varol, Weidi Xie, and Andrew Zisserman. Autoad ii: The sequel-who, when, and what in movie audio description. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13645–13655, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 184, + 294, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 294, + 237 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 294, + 237 + ], + "type": "text", + "content": "[26] Tengda Han, Max Bain, Arsha Nagrani, Gül Varol, Weidi Xie, and Andrew Zisserman. Autoad iii: The prequel-back to the pixels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18164-18174, 2024. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 238, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 238, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 238, + 294, + 293 + ], + "type": "text", + "content": "[27] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 2, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 293, + 294, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 293, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 293, + 294, + 326 + ], + "type": "text", + "content": "[28] Daniel Holden, Taku Komura, and Jun Saito. Phase-functional neural networks for character control. ACM Transactions on Graphics (TOG), 36(4):1-13, 2017. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 327, + 294, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 370 + ], + "type": "text", + "content": "[29] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 371, + 294, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 294, + 414 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 294, + 414 + ], + "type": "text", + "content": "[30] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. Advances in Neural Information Processing Systems, 36:20067-20079, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 415, + 294, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 294, + 469 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 294, + 469 + ], + "type": "text", + "content": "[31] Hanwen Jiang, Shaowei Liu, Jiashun Wang, and Xiaolong Wang. Hand-object contact consistency reasoning for human grasps generation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11107-11116, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 471, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 294, + 525 + ], + "type": "text", + "content": "[32] Yujiao Jiang, Qingmin Liao, Zhaolong Wang, Xiangru Lin, Zongqing Lu, Yuxi Zhao, Hanqing Wei, Jingrui Ye, Yu Zhang, and Zhijing Shao. Smplx-lite: A realistic and drivable avatar benchmark with rich geometry and texture annotations. arXiv preprint arXiv:2405.19609, 2024. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 525, + 294, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 525, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 525, + 294, + 590 + ], + "type": "text", + "content": "[33] Peng Jin, Ryuichi Takanobu, Wancai Zhang, Xiaochun Cao, and Li Yuan. Chat-univi: Unified visual representation empowers large language models with image and video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13700-13710, 2024. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 592, + 294, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 294, + 646 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 294, + 646 + ], + "type": "text", + "content": "[34] Qiuhong Ke, Mohammed Bennamoun, Senjian An, Ferdous Sohel, and Farid Boussaid. A new representation of skeleton sequences for 3d action recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3288-3297, 2017. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 647, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 690 + ], + "type": "text", + "content": "[35] Jihoon Kim, Jiseob Kim, and Sungjoon Choi. Flame: Freeform language-based motion synthesis & editing. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8255-8263, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[36] Kosmas Kritsis, Aggelos Gkiokas, Aggelos Pikrakis, and Vassilis Katsouros. Danceconv: Dance motion genera" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 72, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 72, + 553, + 95 + ], + "type": "text", + "content": "tion with convolutional networks. IEEE Access, 10:44982-45000, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "text", + "content": "[37] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "type": "text", + "content": "[38] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 186, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 251 + ], + "type": "text", + "content": "[39] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mybench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 252, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 252, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 252, + 553, + 297 + ], + "type": "text", + "content": "[40] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 298, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 553, + 352 + ], + "type": "text", + "content": "[41] Jing Lin, Ailing Zeng, Shunlin Lu, Yuanhao Cai, Ruimao Zhang, Haoqian Wang, and Lei Zhang. Motion-x: A large-scale 3d expressive whole-body human motion dataset. Advances in Neural Information Processing Systems, 36, 2024. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 354, + 553, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 408 + ], + "type": "text", + "content": "[42] Lilang Lin, Jiahang Zhang, and Jiaying Liu. Actionlet-dependent contrastive learning for unsupervised skeleton-based action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2363-2372, 2023. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 410, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 410, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 410, + 553, + 464 + ], + "type": "text", + "content": "[43] Daochang Liu, Qiyue Li, Tingting Jiang, Yizhou Wang, Rulin Miao, Fei Shan, and Ziyu Li. Towards unified surgical skill assessment. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9522-9531, 2021. 1, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 466, + 553, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 466, + 553, + 531 + ], + "spans": [ + { + "bbox": [ + 316, + 466, + 553, + 531 + ], + "type": "text", + "content": "[44] Jun Liu, Amir Shahroudy, Dong Xu, and Gang Wang. Spatio-temporal LSTM with trust gates for 3d human action recognition. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 816-833. Springer, 2016. 1, 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 533, + 553, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 533, + 553, + 566 + ], + "spans": [ + { + "bbox": [ + 316, + 533, + 553, + 566 + ], + "type": "text", + "content": "[45] Mengyuan Liu, Hong Liu, and Chen Chen. Enhanced skeleton visualization for view invariant human action recognition. Pattern Recognition, 68:346-362, 2017. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 567, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 567, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 316, + 567, + 553, + 600 + ], + "type": "text", + "content": "[46] Ruizhe Liu, Qian Luo, and Yanchao Yang. Infocon: Concept discovery with generative and discriminative informativeness. arXiv preprint arXiv:2404.10606, 2024. 2, 4, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 601, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 645 + ], + "type": "text", + "content": "[47] Thomas Lucas, Fabien Baradel, Philippe Weinzaepfel, and Grégory Rogez. Posegpt: Quantization-based 3d human motion generation and forecasting. In European Conference on Computer Vision, pages 417-435. Springer, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "text", + "content": "[48] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing, 508:293-304, 2022. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 690, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 690, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 690, + 553, + 713 + ], + "type": "text", + "content": "[49] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7188" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "text", + "content": "understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 151 + ], + "type": "text", + "content": "[50] Shehan Munasinghe, Rusiru Thushara, Muhammad Maaz, Hanoona Abdul Rasheed, Salman Khan, Mubarak Shah, and Fahad Khan. Pg-video-llava: Pixel grounding large video-language models. arXiv preprint arXiv:2311.13435, 2023. 1, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 296, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 296, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 296, + 207 + ], + "type": "text", + "content": "[51] Paritosh Parmar and Brendan Tran Morris. What and how well you performed? a multitask learning approach to action quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 304-313, 2019. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 296, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 296, + 262 + ], + "type": "text", + "content": "[52] Mathis Petrovich, Michael J Black, and Gül Varol. Action-conditioned 3d human motion synthesis with transformer vae. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10985-10995, 2021. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 296, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 296, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 296, + 308 + ], + "type": "text", + "content": "[53] Mathis Petrovich, Michael J Black, and Gül Varol. Temos: Generating diverse human motions from textual descriptions. In European Conference on Computer Vision, pages 480-497. Springer, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 310, + 296, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 310, + 296, + 375 + ], + "spans": [ + { + "bbox": [ + 56, + 310, + 296, + 375 + ], + "type": "text", + "content": "[54] Abhinanda R Punnakkal, Arjun Chandrasekaran, Nikos Athanasiou, Alejandra Quiros-Ramirez, and Michael J Black. Babel: Bodies, action and behavior with english labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 722-731, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 376, + 296, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 296, + 442 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 296, + 442 + ], + "type": "text", + "content": "[55] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 445, + 296, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 296, + 509 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 296, + 509 + ], + "type": "text", + "content": "[56] Jathushan Rajasegaran, Georgios Pavlakos, Angjoo Kanazawa, Christoph Feichtenhofer, and Jitendra Malik. On the benefits of 3d pose and tracking for human action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 640-649, 2023. 2, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 511, + 296, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 511, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 511, + 296, + 555 + ], + "type": "text", + "content": "[57] Leonard Salewski, Stefan Fauth, A Koepke, and Zeynep Akata. Zero-shot audio captioning with audio-language model guidance and audio context keywords. arXiv preprint arXiv:2311.08396, 2023. 3, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 556, + 296, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 556, + 296, + 610 + ], + "spans": [ + { + "bbox": [ + 56, + 556, + 296, + 610 + ], + "type": "text", + "content": "[58] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Two-stream adaptive graph convolutional networks for skeleton-based action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12026-12035, 2019. 1, 2, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 612, + 296, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 612, + 296, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 612, + 296, + 644 + ], + "type": "text", + "content": "[59] Sebastian Starke, He Zhang, Taku Komura, and Jun Saito. Neural state machine for character-scene interactions. ACM Transactions on Graphics, 38(6):178, 2019. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 647, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 296, + 689 + ], + "type": "text", + "content": "[60] Sebastian Starke, Yiwei Zhao, Taku Komura, and Kazi Zaman. Local motion phases for learning multi-contact character movements. ACM Transactions on Graphics (TOG), 39 (4):54-1, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 296, + 713 + ], + "type": "text", + "content": "[61] Sebastian Starke, Ian Mason, and Taku Komura. Deeppphase: Periodic autoencoders for learning motion phase manifolds." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "ACM Transactions on Graphics (TOG), 41(4):1-13, 2022. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 162 + ], + "type": "text", + "content": "[62] Graham W Taylor, Rob Fergus, Yann LeCun, and Christoph Bregler. Convolutional learning of spatio-temporal features. In Computer Vision-ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part VI 11, pages 140-153. Springer, 2010. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 163, + 553, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 208 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 208 + ], + "type": "text", + "content": "[63] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. In European Conference on Computer Vision, pages 358–374. Springer, 2022. 2, 3, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 209, + 553, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 553, + 241 + ], + "type": "text", + "content": "[64] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 243, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 553, + 297 + ], + "type": "text", + "content": "[65] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 4489-4497, 2015. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 298, + 553, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 553, + 353 + ], + "type": "text", + "content": "[66] Raviteja Vemulapalli, Felipe Arrate, and Rama Chellappa. Human action recognition by representing 3d skeletons as points in a lie group. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 588-595, 2014. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 355, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 355, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 355, + 553, + 398 + ], + "type": "text", + "content": "[67] Heng Wang and Cordelia Schmid. Action recognition with improved trajectories. In Proceedings of the IEEE international conference on computer vision, pages 3551-3558, 2013. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 399, + 553, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 455 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 455 + ], + "type": "text", + "content": "[68] Jingbo Wang, Ye Yuan, Zhengyi Luo, Kevin Xie, Dahua Lin, Umar Iqbal, Sanja Fidler, and Sameh Khamis. Learning human dynamics in autonomous driving scenarios. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20796-20806, 2023. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 456, + 553, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 511 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 511 + ], + "type": "text", + "content": "[69] Xin Wang, Jiawei Wu, Junkun Chen, Lei Li, Yuan-Fang Wang, and William Yang Wang. Vatex: A large-scale, high-quality multilingual dataset for video-and-language research. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4581-4591, 2019. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 512, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 512, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 512, + 553, + 567 + ], + "type": "text", + "content": "[70] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, et al. Internvideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191, 2022. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 568, + 553, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 553, + 622 + ], + "type": "text", + "content": "[71] Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Guo Chen, Baoqi Pei, Rongkun Zheng, Jilan Xu, Zun Wang, et al. Intervideo2: Scaling video foundation models for multimodal video understanding. arXiv preprint arXiv:2403.15377, 2024. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 624, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 668 + ], + "type": "text", + "content": "[72] Zeqi Xiao, Tai Wang, Jingbo Wang, Jinkun Cao, Wenwei Zhang, Bo Dai, Dahua Lin, and Jiangmiao Pang. Unified human-scene interaction via prompted chain-of-contacts. arXiv preprint arXiv:2309.07918, 2023. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[73] Junyu Xie, Tengda Han, Max Bain, Arsha Nagrani, Güll Varol, Weidi Xie, and Andrew Zisserman. Autoad-zero: A training-free framework for zero-shot audio description. arXiv preprint arXiv:2407.15850, 2024. 3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7189" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 117 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 117 + ], + "type": "text", + "content": "[74] Guo-Hao Xu, Yi-Lin Wei, Dian Zheng, Xiao-Ming Wu, and Wei-Shi Zheng. Dexterous grasp transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17933-17942, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 162 + ], + "type": "text", + "content": "[75] Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. Pllava: Parameter-free llava extension from images to videos for video dense captioning. arXiv preprint arXiv:2404.16994, 2024. 1, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 295, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 295, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 295, + 207 + ], + "type": "text", + "content": "[76] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In Proceedings of the AAAI conference on artificial intelligence, 2018. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 295, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 295, + 231 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 295, + 231 + ], + "type": "text", + "content": "[77] Zeshi Yang, Kangkang Yin, and Libin Liu. Learning to use chopsticks in diverse gripping styles. 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 232, + 295, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 232, + 295, + 275 + ], + "spans": [ + { + "bbox": [ + 56, + 232, + 295, + 275 + ], + "type": "text", + "content": "[78] Zhou Yu, Dejing Xu, Jun Yu, Ting Yu, Zhou Zhao, Yueting Zhuang, and Dacheng Tao. Activitynet-qa: A dataset for understanding complex web videos via question answering. In AAAI, pages 9127–9134, 2019. 2, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 276, + 295, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 295, + 320 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 295, + 320 + ], + "type": "text", + "content": "[79] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 321, + 295, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 295, + 354 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 295, + 354 + ], + "type": "text", + "content": "[80] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 3, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 355, + 295, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 355, + 295, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 355, + 295, + 409 + ], + "type": "text", + "content": "[81] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Shaoli Huang, Yong Zhang, Hongwei Zhao, Hongtao Lu, and Xi Shen. T2m-gpt: Generating human motion from textual descriptions with discrete representations. arXiv preprint arXiv:2301.06052, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 411, + 295, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 295, + 466 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 295, + 466 + ], + "type": "text", + "content": "[82] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 468, + 295, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 468, + 295, + 522 + ], + "spans": [ + { + "bbox": [ + 56, + 468, + 295, + 522 + ], + "type": "text", + "content": "[83] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xupeng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. Pointclip: Point cloud understanding by clip. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8552-8562, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 523, + 295, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 295, + 578 + ], + "type": "text", + "content": "[84] Renrui Zhang, Jiaming Han, Chris Liu, Peng Gao, Aojun Zhou, Xiangfei Hu, Shilin Yan, Pan Lu, Hongsheng Li, and Yu Qiao. Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199, 2023. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 579, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 295, + 635 + ], + "type": "text", + "content": "[85] Xingyi Zhou, Anurag Arnab, Shyamal Buch, Shen Yan, Austin Myers, Xuehan Xiong, Arsha Nagrani, and Cordelia Schmid. Streaming dense video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18243-18252, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 635, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 295, + 689 + ], + "type": "text", + "content": "[86] Zixiang Zhou, Yu Wan, and Baoyuan Wang. Avatargpt: All-in-one framework for motion understanding planning generation and beyond. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1357-1366, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 690, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 690, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 690, + 295, + 714 + ], + "type": "text", + "content": "[87] Bin Zhu, Bin Lin, Munan Ning, Yang Yan, Jiaxi Cui, HongFa Wang, Yatian Pang, Wenhao Jiang, Junwu Zhang, Zongwei" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 334, + 72, + 553, + 106 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 334, + 72, + 553, + 106 + ], + "type": "text", + "content": "Li, et al. Languagebind: Extending video-language pretraining to n-modality by language-based semantic alignment. arXiv preprint arXiv:2310.01852, 2023. 2, 3, 5" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "7190" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/HuPerFlow_ A Comprehensive Benchmark for Human vs. Machine Motion Estimation Comparison/5a790d2d-155e-42dc-bc07-eb0a483c9c11_content_list.json b/2025/HuPerFlow_ A Comprehensive Benchmark for Human vs. Machine Motion Estimation Comparison/5a790d2d-155e-42dc-bc07-eb0a483c9c11_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..bdef20d29b4a62f108ca9c7b0d95c8e382435cb3 --- /dev/null +++ b/2025/HuPerFlow_ A Comprehensive Benchmark for Human vs. Machine Motion Estimation Comparison/5a790d2d-155e-42dc-bc07-eb0a483c9c11_content_list.json @@ -0,0 +1,1244 @@ +[ + { + "type": "text", + "text": "HuPerFlow: A Comprehensive Benchmark for Human vs. Machine Motion Estimation Comparison", + "text_level": 1, + "bbox": [ + 119, + 128, + 879, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yung-Hao Yang $^{1}$ Zitang Sun $^{1}$ Taiki Fukiage $^{2}$ Shin'ya Nishida $^{1,2*}$", + "bbox": [ + 210, + 202, + 782, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Cognitive Informatics Lab, Graduate School of Informatics, Kyoto University, Japan \n $^{2}$ Communication Science Laboratories, Nippon Telegraph and Telephone Corporation, Japan", + "bbox": [ + 133, + 220, + 862, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://kucognitiveinformaticslab.github.io/Hyperflow-Website/", + "bbox": [ + 218, + 258, + 769, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 308, + 326, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As AI models are increasingly integrated into applications involving human interaction, understanding the alignment between human perception and machine vision has become essential. One example is the estimation of visual motion (optical flow) in dynamic applications such as driving assistance. While there are numerous optical flow datasets and benchmarks with ground truth information, human-perceived flow in natural scenes remains underexplored. We introduce HuPerFlow—a benchmark for human-perceived flow, measured at 2,400 locations across ten optical flow datasets, with $\\sim 38,400$ response vectors collected through online psychophysical experiments. Our data demonstrate that human-perceived flow aligns with ground truth in spatiotemporally smooth locations while also showing systematic errors influenced by various environmental properties. Additionally, we evaluated several optical flow algorithms against human-perceived flow, uncovering both similarities and unique aspects of human perception in complex natural scenes. HuPerFlow is the first large-scale human-perceived flow benchmark for alignment between computer vision models and human perception, as well as for scientific exploration of human motion perception in natural scenes. The HuPerFlow benchmark is publicly available on the HuPerFlow website.", + "bbox": [ + 88, + 342, + 483, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 737, + 220, + 752 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Identifying object motion within the natural environment is crucial for humans, as it supports essential tasks such as navigation, tracking, prediction, and pursuit. In computer vision (CV), optical flow—representing an object's movement in pixel coordinates—serves as a foundational spatiotemporal feature for downstream applications, such as action recognition [36, 39], driver assistance [8, 26], robot navigation [6], video processing [43], etc. With the rise of data-driven approaches, the availability of datasets and", + "bbox": [ + 89, + 763, + 483, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9065b6b6390934b41680e1b09a6fa6d387f9241aba765dc42250afa3be6ddcd4.jpg", + "image_caption": [ + "Figure 1. Example of HuPerFlow. The red arrows indicate human response to perceived flow, and the green arrows indicate ground truth motion vectors. The circles indicate the magnitudes of endpoint errors, the difference between perceived flow and ground truth." + ], + "image_footnote": [], + "bbox": [ + 517, + 309, + 905, + 578 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "benchmarks has consistently fueled advancements in optical flow methods. Numerous studies have leveraged physical engines to generate realistic synthetic data [5] or employed specialized laser and depth cameras [13] to capture ground-truth optical flow in real scenes.", + "bbox": [ + 511, + 684, + 905, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, few studies have examined how humans perceive motion in natural scenes. AI models increasingly interact with humans in various application fields, such as image quality assessment [45, 48], video generation [28], and large language model evaluation [29]). Human responses often guide the development of models, which in turn enhances the models' support of human activities. In the context of visual motion perception, estimating human-perceived flow in addition to the ground truth will be use", + "bbox": [ + 509, + 763, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22799", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ful, for example, in automatic driving [7, 10, 22]. While fully autonomous driving systems can be solely based on the physical ground truth estimation, partially autonomous systems that assist human drivers will be empowered by the ability to predict when, where, and how the drivers are likely to misinterpret motion flow. Another potential use for our dataset is animation synthesis. To control the quality of animation, the animators need to check whether human viewers perceive the movements of the created animation in the way they originally intended. If there are disagreements, the animators should modify the images to minimize the errors. This process can be advanced if human-perceived motion can be automatically predicted. Machine prediction of human-perceived flow is a hard problem for hand-drawn animations since there is no physical ground truth. Even when the animation is rendered from a physically correct dynamic model of the scene, human-perceived motion can be distorted by many factors, including frame rates, image resolution, and the touch of the painting.", + "bbox": [ + 89, + 90, + 480, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Humans possess a sophisticated visual system capable of estimating object movement in complex natural scenes; however, the underlying mechanisms differ significantly from those used in current CV models. Motion inference is inherently an under-constrained problem (such as the aperture problem [30]), and the solution in humans is a perception that unfolds over a spatiotemporal process spanning multiple frames [1]. In contrast, CV methods typically focus on deterministic dense correspondence between consecutive frames. This difference can sometimes result in entirely different motion perceptions between human vision and CV algorithms, as seen in phenomena like the reverse phi [2] and missing fundamental illusions [4] investigated in vision science and psychology.", + "bbox": [ + 89, + 382, + 480, + 594 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite this significant gap, currently, no large-scale benchmark directly evaluates the relationship between CV model responses, physical ground truth, and human perceptual responses to motion in natural scenes, thus hindering further research within this field. To address this problem, we construct the first large-scale comprehensive benchmark for human-perceived optical flow across several types of naturalistic scenes. Our study introduces several unique features compared to prior work: 1): Human-centric labeling: The dataset's labels are directly derived from human perception, gathered through rigorously designed psychophysical experiments. We refer to these human-perceived labels as \"Human Perception\" and use the term \"Ground Truth (GT)\" for the physical-correct labels generated by conventional methods (Fig. 1); 2) Large-scale: our dataset includes around 38,400 data points (trials) derived from 480 participant sessions, spanning over 240 hours of collection time. The benchmark comprises ten representative computer vision motion datasets—such as Sintel [5], KITTI [26], Spring [25], VIPER [35], etc—each", + "bbox": [ + 89, + 598, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "with specifically probed locations and optical flow GT; 3) A novel experimental paradigm: Unlike traditional human assessments (e.g., image quality evaluation), collecting motion perception data is complex and time-intensive. Motion data collection requires a strict display setting, including a constant refresh rate, high resolution, and high-speed hardware to ensure them. Our innovative online experimental paradigm, grounded in a psychophysical method developed for a lab experiment [47], ensures stable and reliable data collection even in online settings, making large-scale human-perceived motion data feasible.", + "bbox": [ + 511, + 90, + 903, + 256 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In addition to constructing this benchmark dataset, we conducted several initial comparative analyses between several representative CV models and human responses, including spatiotemporal gradient-based methods [11], recurrent refinement models [42], multi-frame methods [37], vision transformers [14], human-inspired approaches [38, 40], etc. Our findings reveal several key differences between model predictions and human perception, as well as unique characteristics of human perception in complex natural scenes, including abrupt direction switches (e.g., VIPER [35]), motion blur and non-rigid motion (e.g., MPI Sintel [5]), object translation and rotation (e.g., Flythings3D [24]), camera motion (e.g., Monkaa [24]), and local versus global motion (e.g., MHOF [34]). While the recent CV models correlate closely with human perception, they often exhibit low partial correlation when controlling the ground-truth effects, suggesting their design inherently fails to capture humans' unique perceptual biases.", + "bbox": [ + 511, + 258, + 903, + 529 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions include: (1) the first large-scale human-perceived optical flow benchmark across ten diverse motion scenarios, (2) a new online psychophysics-based experimental paradigm enabling the large-scale collection of precise human-perceived optical flow in natural scenes, and (3) an evaluation of ten different representative models from both computer vision and human vision fields, revealing significant gaps between human motion interpretation and model predictions. These insights advance computer vision research on human-aligned AI and support the development of technologies for better interaction and assistance with humans.", + "bbox": [ + 511, + 532, + 903, + 712 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Previous optical flow datasets", + "text_level": 1, + "bbox": [ + 511, + 729, + 782, + 747 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Computer optical flow datasets", + "text_level": 1, + "bbox": [ + 511, + 756, + 787, + 772 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "There are several optical flow datasets with physical motion ground truth, and we selected some of them based on motion content, image resolution, sequence count, and frames per sequence for online human observers. KITTI flow 2015 [26] is a real-world driving condition, where ground-truth optical flows were captured by stereo camera, laser scanner, and GPS. Virtual KITTI 2 [12] is a photo-realistic synthetic driving dataset featuring different weather conditions", + "bbox": [ + 511, + 779, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22800", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and viewing angles. The Driving subset of the Scene Flow Datasets [24] includes eight synthetic sequences (slow / fast $\\times$ 15mm / 300mm focal length $\\times$ forwards / backward scene). VIPER [35] and TartanAir [44] provide synthetic video sequences with various speed and direction changes, representing diverse navigation scenarios beyond typical driving benchmarks for distinct object motion and camera motion.", + "bbox": [ + 88, + 90, + 480, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In contrast to automotive navigation scenarios, MPI Sintel [5] and SPRING [25] feature long-range, non-rigid motion with discontinue, large jumps, motion blur, and varying illumination, scene, and material properties. Monkaa, also from Scene Flow Datasets [24], includes significant camera translation and rotation, crucial for revealing motion perception influenced by surrounding objects and background. MHOF [34] captures diverse human motion, featuring both local and global motions across various postures and actions. Flythings3D subset of the Scene Flow Datasets [24], includes multiple geometric shapes and objects translating and rotating along complex 3D trajectories. Since the video presentation required several frames for human observers, we excluded well-known benchmarks such as Middlebury [3] (2-8 frames), FlyingChairs2 [16] (2 frames), ChairsSDHom [15] (2 frames), etc.", + "bbox": [ + 93, + 213, + 485, + 454 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Human motion perception datasets", + "text_level": 1, + "bbox": [ + 89, + 468, + 395, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Previous studies on human visual motion perception generally used well-controlled artificial visual stimuli, such as sine-wave gratings [21], random-dot kinematograms [20], and biological motion [19], to investigate specific processing mechanisms. Some studies used more complex natural-scenes stimuli to explore neural responses for them [27, 33]. While these studies provided valuable insights into human motion perception, only one study [47], to our knowledge, has directly reported human-perceived vectors for nature scenes, and compared them with ground-truth motion vectors and computer vision model predictions. This psychophysical study revealed that human-perceived optical flow shows various perceptual deviations from the ground-truth flow. Note that human-perceived vectors in dynamical movies, which are mainly based on motion energy within a $\\sim 100$ ms temporal window [1], could be significantly different from the correspondence of salient feature points between adjacent frames (cf. [9]). However, [47] examined only five movies, all selected from the SlowFlow version (1008 frames per second) of MPI Sintel benchmark [18], in which the jump sizes between adjacent frames are substantially smaller than those of the standard version used in the current study. It is also worth noting that the movies in [47] were presented through a circular aperture centered on the probed location.", + "bbox": [ + 91, + 492, + 482, + 868 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the current study, we adopt an approach similar to [47] with significant modifications: (1) using ten optical flow", + "bbox": [ + 89, + 869, + 480, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "benchmarks that are widely used for training and testing in computer vision fields, (2) choosing a larger number of motion clips with dynamic and diverse spatiotemporal properties, and (3) presenting the entire image on the display, as observers would naturally perceive it, rather than through a central aperture. This approach provides a greater motion context and ecological validity.", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. The HuPerFlow dataset", + "text_level": 1, + "bbox": [ + 511, + 210, + 736, + 226 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Methods", + "text_level": 1, + "bbox": [ + 511, + 234, + 617, + 250 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.1. Ethical statement", + "text_level": 1, + "bbox": [ + 511, + 257, + 681, + 271 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This study was approved by the Research Ethics Committee of the Graduate School of Informatics at Kyoto University. All participants received online informed consent before the experiments.", + "bbox": [ + 511, + 276, + 903, + 335 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.2. Stimuli and Apparatus", + "text_level": 1, + "bbox": [ + 511, + 338, + 723, + 353 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The experiment was conducted using PsychoPy [32], a software for building behavioral experiments, and Pavlovia [31], an online platform for running these experiments in a web browser. To account for diverse hardware environments in online experiments, we controlled the visual presentation in both spatial and temporal domains. Spatially, to standardize viewing angles, observers performed a \"credit card calibration,\" adjusting an on-screen virtual credit card to match a physical one. This allowed us to calculate the physical pixel size after observers adjusted their viewing distance to ensure that 50 pixels corresponded to $1^{\\circ}$ of visual angle, maintaining consistent stimulus pixel sizes across devices. Temporally, visual stimuli were presented at $30\\mathrm{Hz}$ . Since optimal timing control is crucial for online motion perception experiments, we did not use a higher frame rate for which frame drops were more likely to occur under some users' environments. In addition, We recorded frame drops and onset times during online experiments (see Supplementary Appendix 1 for more details). The frame dropping rate was .041%, and the frame onset error was -0.055 ms $(\\mathrm{SD} = 3.664\\mathrm{ms})$ , indicating stable and precise visual stimulus presentation across devices.", + "bbox": [ + 511, + 357, + 906, + 688 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This project contained two phases: a preliminary Random Dot Kinematogram (RDK) for training and selection of the observers, and a main session to collect human-perceived flow data using multiple optical flow datasets. In the RDK phase, 5,000 black and white dots moved uniformly within a 600-pixel circular aperture. In the main experiment, we used ten optical flow datasets including KITTI Flow 2015 [26], Virtual KITTI 2 [12], Driving [24], VIPER [35], TartanAir [44], MPI Sintel Flow [5], SPRING [25], Monkaa [24], MHOF [34], and Flythings3D [24]. As most benchmarks include sub-datasets, we present our selection in detail in Tab. 1.", + "bbox": [ + 511, + 689, + 903, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We applied a standard movie-formating method across datasets with the following exceptions: (1) Most datasets", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22801", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a1c840cf71d7cf5fbc39373e25d995797093501e950a10a26adb35e9a7b09f5c.jpg", + "image_caption": [ + "Figure 2. Experimental procedure: At the start of each trial, a green circle appeared to mark the selected area. Next, a motion sequence and a matching stimulus were presented alternately until a response was made. For most optical flow datasets, the motion sequences consisted of 15 frames (500 ms). It was followed by an inter-stimulus interval (ISI) of 1000 ms, then a matching stimulus (500 ms), and a second ISI (1000 ms). A flash probe was shown in the middle frame (the 8th frame) of the motion sequence to indicate the timing and location of the target motion. The green circle location cue, four-dot placeholders, and matching stimulus were presented sequentially at the same location." + ], + "image_footnote": [], + "bbox": [ + 212, + 59, + 782, + 239 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/27a2f7563a2498246b2357b8b1f41306632acc6371fc6256cce699a432c54675.jpg", + "table_caption": [ + "Table 1. Summary of the selected optical flow datasets. * indicates the image size is resized by cubic interpolation." + ], + "table_footnote": [], + "table_body": "
DatasetsResolution (pixels)SequencesTotal FramesSelection
KITTI 2015 [26]1,242 × 3752004,200Training, left image
Virtual KITTI 2 [12]1,242 × 375521,260Clone, camera_0
Driving [24]960 × 54082,400Finalpass, left, slow
VIPER [35]960 × 540*77134,097Val, jpg
TartanAir [44]640 × 480166,982image_left
MPI Sintel [5]1,024 × 436231,064Training, Final
Spring [25]960 × 540*375,000Train_frame_left
Monkaa [24]960 × 540248,664Finalpass, left
MHOF [34]640 × 6401,23895,868Train, masked with segm_EXR
FlyingThings [24]960 × 5402,18021,818Val, image.clean, left
", + "bbox": [ + 91, + 393, + 498, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "were presented at their original sizes, except for VIPER [35] and SPRING [25], which were resized from $1920 \\times 1080$ to $960 \\times 540$ for online observers, with the optical flow adjusted accordingly. (2) Frames F-7 to F+7 around a selected testing frame were used as movie sequences, except for MHOF [34] and Flythings3D [24], which were limited to F-4 to F+5 due to each sequence having only ten frames. (3) Virtual KITTI2 [12], SPRING [25] and Scene Flow Datasets [24] (i.e., Driving, Monkaa, and Flythings3D) contain both future (forward) and past (backward) flow directions. However, only future-directed flows were reported as video images were presented sequentially in the forward direction, frame-by-frame. (4) We did not modify the original images, except for MHOF [34] dataset, where frames contained irrelevant background scenes. To standardize the", + "bbox": [ + 89, + 674, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "background, we used the \"segm_EXR\" provided by MHOF to render it as a uniform gray background.", + "bbox": [ + 511, + 357, + 906, + 388 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1.3. Design and Procedure", + "text_level": 1, + "bbox": [ + 511, + 393, + 714, + 407 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In both the RDK and main experiments, each session began with six practice trials followed by 80 formal trials. The practice trials familiarized observers with the procedure and helped them understand the differences between the ground truth and their perceived vectors. To support this, we provided visual feedback of both GT and perceived vectors after each response during practice trials. If the angle between perceived vectors and the GT exceeded $30^{\\circ}$ , or if the endpoint error (EPE, the Euclidean distance between the human responses and GT, Equation 1.) was greater than 10 pixels, the same practice trial would be repeated.", + "bbox": [ + 511, + 411, + 906, + 578 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {E P E} = \\sqrt {\\left(u _ {\\text {R e s p}} - u _ {\\mathrm {G T}}\\right) ^ {2} + \\left(v _ {\\text {R e s p}} - v _ {\\mathrm {G T}}\\right) ^ {2}} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 565, + 585, + 903, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the RDK session, the direction $(0^{\\circ} - 360^{\\circ})$ and speed (1-20 pixels per frame, PPF) of the dots were varied randomly across trials but kept constant within a trial. During formal trials in this session, feedback was also shown to reinforce training. In the main experimental session, we used ten datasets as previously described, selecting 24 sequences per dataset. Each sequence included ten target motion locations with speeds under 20 PPF at a specific frame. At each probed location, we collected 16 repeated trials (four repetitions per observer $\\times$ four observers) to obtain a stable measure of perceived flow, totaling 2,400 probed locations across 38,400 trials. In each session, participants viewed two movie sequences with ten target motion locations and four repetitions each (80 trials in total). This yielded 480 sessions across 10 datasets $\\times$ 12 sessions (2 movie sequences each) $\\times$ 4 observers. No feedback was provided during formal trials in the main experiments.", + "bbox": [ + 511, + 613, + 906, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fig. 2 illustrates the display layout and procedure in the main experiment. Both the RDK and main experiments", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22802", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "used a method of adjustment to measure human-perceived flow at the flash-probed spatiotemporal location. Observers reported the perceived flow of a target motion by matching it with a circular Brownian noise pattern $(1/f^2, 120$ pixels in diameter), which has a spectrum similar to natural images and gives the least motion aliasing for a wide range of motion speeds. Using a mouse cursor (a purple dot, 10 pixels in diameter) within a circular control panel (600 pixels in diameter), observers could freely adjust the direction $(0^\\circ - 360^\\circ)$ and speed (1-20 pixels per frame) of the matching Brownian noise. The direction was mapped to angles $(\\theta)$ in a polar coordinate, and speed was adjusted logarithmically based on the cursor's distance from the center of the panel. This method provides more accurate measurements of perceived motion vectors than subjective reports. The control panel for the matching stimulus was always fixed at the display's center. See Supplementary Video S1 for the animated procedure.", + "bbox": [ + 89, + 90, + 480, + 363 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the RDK session, where direction and speed were uniform within each trial, both the target motion and matching stimulus were presented at the center of the aperture. In the main experiment, a green circle (120 pixels in diameter), presented before each trial, served as a spatial cue indicating the target motion area. Additionally, four additional dots (green dots, 10 pixels in diameter) served as placeholders throughout the trial. They were positioned 60 pixels above, below, on the left, and on the right of the probed dot. For most datasets, a repetition cycle (see below) consisted of the target movie sequence (500 ms, 15 frames), an interstimulus interval (ISI, 1000 ms), and the matching stimulus (500 ms), followed by a second ISI (1000 ms) before the next cycle. For the MHOF [34] and FlyingThings3D [24] datasets, which contain only ten frames (333 ms), the second ISI was adjusted to 1167 ms to keep the cycle duration consistent.", + "bbox": [ + 89, + 375, + 482, + 617 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In each trial, a flash-probed dot (a green dot, seven pixels in diameter) was presented at a selected testing frame and location to indicate the timing and position of the target motion. The flash probe appeared in the middle of each movie sequence presentation (typically at the $8^{\\text{th}}$ frame, and at the $5^{\\text{th}}$ frame for MHOF and FlyingThings3D), indicating when observers should respond. The probed target motions were intentionally selected from informative locations, including main objects, the background, and minor objects. In the main experiment, the four-dot placeholders appeared around the probed location, with the matching stimulus displayed directly at that placeholder's location. During each trial, observers could view the movie sequences and match Brownian noise multiple times (repetition cycles), while responses were only accepted after completing three repetition cycles. Since the target motion was shown for just one frame, this adjustment process allowed observers to refine their responses through repeated viewing.", + "bbox": [ + 89, + 628, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Results", + "text_level": 1, + "bbox": [ + 513, + 90, + 607, + 104 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.1. Data quality control", + "text_level": 1, + "bbox": [ + 513, + 113, + 700, + 128 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The study included 55 observers (mean age = 22.18, SD = 2.82, 24 females and 31 males) who participated in a preliminary Random Dot Kinematogram (RDK) experiment. To ensure reliable perceived flow data, we initially selected observers whose correlation coefficient between perceived vectors and ground truth vectors exceeded .90 in the RDK session. Nine observers with low correlation (individual correlation range: .501 to .885) and four who were unable to complete the experiment were excluded. The group-level correlation for the 42 remaining observers was high (Pearson correlation coefficient $r = .955$ , with individual correlations ranging from .917 to .985), demonstrating their ability to reproduce perceived motion accurately. These 42 observers (mean age = 22.19, SD = 2.46; 17 females, 25 males) then proceeded to the main experimental sessions. They participated in 480 sessions since some observers attended multiple different sessions.", + "bbox": [ + 511, + 132, + 903, + 388 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.2. Perceived vectors and \"flow illusions\"", + "text_level": 1, + "bbox": [ + 513, + 398, + 821, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Fig. 3, we show representative examples of averaged human responses for each dataset (see Supplementary Video S2 for target motions and the corresponding averaged human responses across all probed locations within each dataset). Our findings reveal systematic and stable human-specific biases in motion perception under certain scenarios. We termed these systematic discrepancies between human responses and ground truth motions as \"flow illusions\". These deviations were not random errors but provided valuable insights into the underlying processing and mechanism of human motion perception. On the other hand, we also observed individual differences, with an average inter-observers correlation of 0.407 (SD: 0.417; range from -0.329 to 0.990) across all probed locations. These Individual variations would be useful for training models to reproduce the distribution of human motion perception. To facilitate further research, we have made all raw individual data publicly available on the HuPerFlow website [46].", + "bbox": [ + 511, + 417, + 903, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In most driving datasets such as KITTI 2015 [26], Virtual KITTI 2 [12], and Driving [24], the direction and speed of optical flows transition smoothly across frames, with perceived flows aligning closely with the ground truth motion. This consistency supports the reliability of our measurement method in capturing observers' perceptions under spatiotemporally smooth scenes. Notably, as optical flow locations deviated from the scene center, observers often perceived the directions to be slightly shifting toward the central region. This finding may reflect observers' attempts to cancel out the image motion components produced by self-motion (camera motion) during driving. In contrast, other navigation datasets like VIPER [35] and TartanAir [44] show abrupt direction changes, particularly during ac", + "bbox": [ + 511, + 689, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22803", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/eab63f52866073a542b9286741b20c8687dda5751d44bc7d53cd3cd4590f1702.jpg", + "image_caption": [ + "Figure 3. Demonstration of human perceived motion vectors. The results cover ten representative existing optical flow benchmarks." + ], + "image_footnote": [], + "bbox": [ + 212, + 61, + 782, + 344 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tions such as turning, lane changing, or flying backward. These involve interactions between image motion and camera motion, with perceived flows often revealing systematic errors primarily driven by camera motion.", + "bbox": [ + 88, + 396, + 480, + 457 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In datasets like MPI Sintel [5] and SPRING [25], where images include blurred, non-rigid, discontinuous motion and large jump size (cf. [47] using stimuli with small jumps), observers generally failed to discern individual locations, perceiving them as part of a unified grouping movement instead. Unlike the datasets described in the previous paragraph, which were viewed with smoothly moving cameras, Monkaa [24] dataset involved significant camera rotation around relatively stationary objects. Observers often perceived these objects as moving due to the motion of surrounding objects or backgrounds, illustrating the phenomena of relative and induced motion.", + "bbox": [ + 88, + 458, + 482, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the MHOF [34] datasets, human figures generally exhibit global body movement accompanied by local motions from the head and limbs. Although the motion structures are less complex than in MPI Sintel and Spring, observers tended to focus on the direction of the global body movement and ignored the local gestures. In the Flythings3D [24], multiple objects float along 3D random trajectories, requiring observers to perform object-ground and object-object segmentation. They were successful in perceiving motion against simple backgrounds and distinct surrounding objects but made more errors in cluttered, messy conditions.", + "bbox": [ + 88, + 642, + 482, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As different datasets revealed various types of errors, we calculated the correlation between human response and ground truth for each dataset separately. In general, the driving datasets (i.e., KITTI 2015 [26], Virtual KITTI 2 [12], and Driving [24]) showed high correlations, but oth", + "bbox": [ + 89, + 825, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ers showed middle to low correlation (see Supplementary Appendix 2 for the scatter plot of each dataset).", + "bbox": [ + 511, + 396, + 906, + 426 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2.3. EPE in relation to optical flow properties", + "text_level": 1, + "bbox": [ + 511, + 439, + 844, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Fig. 3, the EPE of human response (indicated by the diameter of the green circle) varied across different positions and scenes. Following [5], we evaluated how EPE changed in relation to various optical flow properties. Fig. 4 shows that EPE increased as GT speed increased, which suggests that faster motions imposed more significant computational challenges. To assess temporal flow dynamics, we compared optical flow from the previous frame (t-1) to the selected probed frame (t-0) and analyzed the temporal gradient changes in uv-vectors over two intervals: from t-1 to t-0, and from t-0 to t+1 (ground truth), normalized by GT speed from the same location. Generally, we observed that EPE increased as the normalized temporal flow gradient increased up to the ratio of 2, and then leveled off.", + "bbox": [ + 511, + 460, + 906, + 671 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also considered spatial image properties by using a Sobel kernel to derive the image gradient, calculating the average edge magnitude within a $\\pm 40$ -pixel neighborhood around the probed location. The results indicate that EPE decreased as the image gradient increased. Additionally, using instance boundary from the MPI Sintel dataset, we calculated the shortest distance to the boundary and found that EPE also decreased as the distance increased. These findings suggest that edges and boundaries may create ambiguities in motion flow interpretation and interrupt motion perception. Finally, using camera motion data from the MPI Sintel as a proxy for self-motion within a scene, we observed that EPE increased with self-motion, and this effect persisted even when excluding the effect of GT speed (image motion).", + "bbox": [ + 511, + 674, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22804", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f6cd9664e467d9370b42dc86d615c6c96676048813b0f30e67546cbb352b968b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 171, + 68, + 390, + 169 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5d280e5ffb9e5666da580f64b8e33b934a326fcd4e133f38097804794bd2512f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 392, + 69, + 604, + 164 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1e574383185e22e89a7e2b88dd5fd65e2d45bb8d91af8b06cdb2967637b392e9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 609, + 70, + 821, + 166 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1f5836719aec56efb93590efc195e6442316d2f7757bb8aa004fc5598871bdb8.jpg", + "image_caption": [ + "Figure 4. Endpoint errors (EPEs) of human responses and two machine vision models as functions of optical flow properties. The horizontal axis of each panel is the GT speed, normalized temporal flow gradient, spatial image gradient, instance boundary distance, and self-motion with(out) normalized by the image GT speed. The vertical axis represents EPE with different scales for visualization." + ], + "image_footnote": [], + "bbox": [ + 171, + 170, + 390, + 270 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e1d79cc765d5d1b6e3ee3fda065d0b267e02bacee21d7edcea4e52926cae09ee.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 392, + 170, + 607, + 270 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/46216c18208f5bf43c2632f556bad05b9eada102d7a79ed0e37316bb20676104.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 609, + 170, + 823, + 265 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In addition to the EPE of human response, Fig. 4 shows the EPEs of two machine vision models, VideoFlow and FFV1MT (see Sec. 4). While there were clear differences in the absolute level of performance, the optical flow properties generally affect EPE in similar ways for human and machine vision models, except for the effects of spatial image gradient effect on VideoFlow.", + "bbox": [ + 89, + 338, + 483, + 444 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4. Comparison with optical flow algorithms", + "text_level": 1, + "bbox": [ + 89, + 449, + 457, + 467 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We selected a range of representative optical flow algorithms to assess the similarities and differences among human-perceived flows, ground truth motion, and predictions from computer vision models. These included classical methods based on brightness constancy assumptions, such as Farneback [11]; biologically inspired approaches like FFV1MT [38]; multiscale inference models, such as FlowNet2 [15]; recurrent refinement models like RAFT [42]; transformer-based models like FlowFormer [14]; graph reasoning methods such as AGFlow [23]; multiframe approaches like VideoFlow [37]; Motion energy and self-attention based models like V1Attention [40] and DualModel [41]; as well as general-purpose architectures like Perceiver IO [17]. For a fair comparison, we used each model's default configuration or the best pre-trained model fine-tuned on the Sintel dataset. Since unsupervised learning methods approximate the supervised ground truth, we limited our comparison to supervised methods.", + "bbox": [ + 89, + 476, + 482, + 747 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For evaluation, we first calculated the average EPE and Pearson correlation coefficients between model predictions and both human responses and ground truth. To further understand the models' alignment with human perception independently of ground truth influence—since most models were trained to fit ground truth and the ground truth was inherently correlated with human responses—we also calculated the partial correlation between model predictions and human responses, controlling for ground truth, as shown in Equation 2. Here, $r_{x,y}$ denotes correlation coefficients be", + "bbox": [ + 89, + 750, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tween $x$ and $y$ .", + "bbox": [ + 513, + 339, + 614, + 352 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\rho_ {\\text {m o d e l}} = r _ {\\text {r e s p , m o d e l} \\cdot \\mathrm {G T}} = \\frac {r _ {\\text {r e s p , m o d e l}} - r _ {\\text {r e s p , G T}} \\cdot r _ {\\text {m o d e l , G T}}}{\\sqrt {1 - r _ {\\text {r e s p , G T}} ^ {2}} \\cdot \\sqrt {1 - r _ {\\text {m o d e l , G T}} ^ {2}}}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 364, + 903, + 416 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To complement partial correlation analysis, we also adopted a Response Consistency Index (RCI) from [47]. This index measures the similarity between model predictions and human responses at each probed location, approaching $+1$ when the model's prediction aligns closely with human flow illusions (see Supplementary Appendix 3 for more details).", + "bbox": [ + 511, + 417, + 905, + 521 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 2 summarizes model predictions across all datasets. Overall, VideoFlow's predictions were the closest to the ground truth. RAFT showed the most closely resembling human responses while also having a strong alignment with the ground truth (second-best among models). When controlling for the ground truth effect, FFV1MT and V1Attention-StageI were the most similar to human flow illusions, as indicated by high partial correlations and average RCI values. A detailed analysis comparing the predictions of optical flow algorithms with both the ground truth and human responses for each dataset is provided in Tables S1-S12 of Supplementary Appendix 4.", + "bbox": [ + 511, + 523, + 905, + 703 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 5 provides examples of model predictions, highlighting examples where predictions aligned closely with the ground truth (e.g., VideoFlow and RAFT) versus those having some resemblance to human response errors (e.g., FFV1MT and V1Attention-StageI). For instance, at Point (A-I) in Fig. 5, both flow illusions and FFV1MT exhibited temporal-sluggish responses but VideoFlow aligned well with GT. Conversely, point (A-II) in Fig. 5 both model predictions were similar to flow illusions, likely due to image signal loss, as the GT indicated a small, dark stone falling. At Point (B-I) and other selected body locations, both human perception and FFV1MT reflected grouping motion in the same direction, regardless of local gesture directions.", + "bbox": [ + 511, + 704, + 908, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22805", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/21cee0d09d52486930b0921198cf5b3d0e05b7c9a8764907d8372104fc4b0146.jpg", + "table_caption": [ + "Table 2. Predictions of optical flow algorithms versus Human response or ground truth (GT). $\\rho$ : Partial correlation; $r$ : Pearson correlation coefficient; EPE: vector end-point error; uv, dir, spd represent motion components in Cartesian space, direction, and speed, respectively. RCI is introduced from [47] to represent the model's similarity to human perception (the larger, the more human-aligned)." + ], + "table_footnote": [], + "table_body": "
MethodρRCIv.s. Humanv.s. GT
ρuvρdirρspdruvrdirrspdEPEruvrdirrspdEPE
Ground TruthNaNNaNNaNNaN0.660.430.296.961.001.001.000.00
Farnebäck [11]0.170.140.040.130.540.380.156.910.680.650.395.52
FFV1MT [38]0.310.280.060.150.630.450.206.580.700.590.545.84
RAFT [42]0.110.220.010.050.640.470.237.370.930.850.792.19
FlowNet2 [15]0.100.260.000.070.590.480.187.550.820.800.613.24
FlowFormer [14]0.070.19-0.020.040.640.460.217.430.930.870.792.08
VideoFlow [49]0.080.22-0.020.040.640.470.227.470.940.880.811.91
Perceiver IO [17]0.090.170.030.050.620.440.237.750.890.820.742.87
AGFlow [23]0.160.240.010.110.540.450.157.940.680.660.465.44
V1Attention-StageI [40]0.230.280.020.180.550.450.146.360.630.540.427.00
DualModel [41]0.200.250.060.130.570.450.207.310.700.670.535.66
", + "bbox": [ + 117, + 116, + 875, + 335 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/098a852fcfbaf5c7796842bd6676713224dd1afe568ece9ecb2331b7dcda6e76.jpg", + "image_caption": [ + "Figure 5. Predicted vectors of optical flow algorithms. Examples are FFV1MT [38], V1Attention [40], VideoFlow [37], and RAFT [42]. We marked some points with large RCI differences between the models in the left and right columns. Generally, the model on the left side shows a more human-aligned response over ground truth, while the right side shows the opposite result." + ], + "image_footnote": [], + "bbox": [ + 96, + 359, + 478, + 705 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "At Point (C-I), a strong self-motion effect occurred toward a static motorcycle. Despite the strong physical optical flow in the image, human observers and V1Attention-StageI appeared to compensate for self-motion to perceive the motorcycle as stationary. For more examples, please visit the HuPerFlow website [46]. Note, however, that there remains", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "a large gap between human responses and the most similar model predictions, possibly because the purpose of human visual motion processing is not merely to estimate image flow. The primary purpose of making our dataset is to help future studies develop models that can fill this gap.", + "bbox": [ + 511, + 361, + 906, + 438 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 478, + 633, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "By overcoming technical challenges, we developed a novel online paradigm to collect human motion perception data, introducing the first large-scale benchmark for human-perceived optical flow. Our findings identified distinct perceptual illusions across various scene scenarios, some aligning with and others diverging from the predictions of CV optical flow algorithms. Human motion perception, compared to state-of-the-art CV models, has limited spatial, temporal, and intensity resolutions and struggles with motion estimation in retinal coordinates. However, human motion vision has remarkable capabilities in analyzing hierarchical movement structures and distinguishing object motion from self-motion in world coordinates. Our proposed methods can be extended to collect broader human data from diverse datasets, further bridging gaps between CV, human perception, and ground truth.", + "bbox": [ + 511, + 513, + 906, + 755 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While our dataset may not be sufficient to train models from zero, it offers valuable benchmarks for model evaluation and fine-tuning. Physically accurate CV models are crucial, but aligning with human perception is equally important for human-centered applications. Notably, current optical flow models fail to predict various human-perceived motion illusions, highlighting the need to address this gap both for practical applications and for advancing the scientific understanding of human vision.", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22806", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by JSPS Grants-in-Aid for Scientific Research (KAKENHI), Grant Numbers JP20H00603, JP20H05957, and JP24H00721, and by the Spring Fellowship, Grant Number JPMJFS2123.", + "bbox": [ + 89, + 114, + 483, + 176 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 191, + 187, + 207 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Edward H Adelson and James R Bergen. Spatiotemporal energy models for the perception of motion. Josa a, 2(2): 284-299, 1985. 2, 3", + "[2] Stuart M Anstis and Brian J Rogers. Illusory continuous motion from oscillating positive-negative patterns: Implications for motion perception. Perception, 15(5):627-640, 1986. 2", + "[3] Simon Baker, Daniel Scharstein, J. P. Lewis, Stefan Roth, Michael J. Black, and Richard Szeliski. A Database and Evaluation Methodology for Optical Flow. International Journal of Computer Vision, 92(1):1-31, 2011. 3", + "[4] Richard O Brown and Sheng He. Visual motion of missing-fundamental patterns: motion energy versus feature correspondence. Vision Research, 40(16):2135-2147, 2000. 2", + "[5] Daniel J. Butler, Jonas Wulff, Garrett B. Stanley, and Michael J. Black. A naturalistic open source movie for optical flow evaluation. In European Conf. on Computer Vision (ECCV), pages 611-625. Springer-Verlag, 2012. 1, 2, 3, 4, 6", + "[6] Haiyang Chao, Yu Gu, and Marcello Napolitano. A survey of optical flow techniques for robotics navigation applications. Journal of Intelligent & Robotic Systems, 73:361-372, 2014. 1", + "[7] Chenyi Chen, Ari Seff, Alain Kornhauser, and Jianxiong Xiao. Deepdriving: Learning affordance for direct perception in autonomous driving. In Proceedings of the IEEE international conference on computer vision, pages 2722-2730, 2015. 2", + "[8] Jaechan Cho, Yongchul Jung, Dong-Sun Kim, Seongjoo Lee, and Yunho Jung. Moving object detection based on optical flow estimation and a gaussian mixture model for advanced driver assistance systems. Sensors, 19(14):3217, 2019. 1", + "[9] Carl Doersch, Ankush Gupta, Larisa Markeeeva, Adrià Recasens, Lucas Smaira, Yusuf Aytar, João Carreira, Andrew Zisserman, and Yi Yang. Tap-vid: A benchmark for tracking any point in a video. arXiv preprint arXiv:2211.03726, 2022. Published in NeurIPS Datasets and Benchmarks track, 2022. 3", + "[10] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 2", + "[11] Gunnar Farnebäck. Two-frame motion estimation based on polynomial expansion. In Proceedings of the Scandinavian conference on Image analysis, pages 363-370. Springer, 2003. 2, 7, 8", + "[12] Adrien Gaidon, Qiao Wang, Yohann Cabon, and Eleonora Vig. Virtual worlds as proxy for multi-object tracking analysis. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2, 3, 4, 5, 6" + ], + "bbox": [ + 93, + 215, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 1", + "[14] Tzehoong Huang, Zhichao Zhang, Wenjie Zeng, Ming Liu, Chunjing Zhang, Hongsheng Li, and Shijian Lu. Flowformer: A transformer architecture for optical flow. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2, 7, 8", + "[15] Eddy Ilg, Nikolaus Mayer, Tonmoy Saikia, Margret Keuper, Alexey Dosovitskiy, and Thomas Brox. Flownet 2.0: Evolution of optical flow estimation with deep networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2462-2470, 2017. 3, 7, 8", + "[16] Eddy Ilg, Tonmoy Saikia, Margret Keuper, and Thomas Brox. Occlusions, motion and depth boundaries with a generic network for disparity, optical flow or scene flow estimation. In European Conference on Computer Vision (ECCV), 2018. 3", + "[17] Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, et al. Perceiver io: A general architecture for structured inputs & outputs. arXiv preprint arXiv:2107.14795, 2021. 7, 8", + "[18] Joel Janai, Fatma Güney, Jonas Wulff, Michael J Black, and Andreas Geiger. Slow flow: Exploiting high-speed cameras for accurate and diverse optical flow reference data. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3597-3606, 2017. 3", + "[19] Gunnar Johansson. Visual perception of biological motion and a model for its analysis. Perception & Psychophysics, 14(2):201-211, 1973. 3", + "[20] Bela Julesz. Foundations of Cyclopean Perception. University of Chicago Press, 1971. 3", + "[21] D. H. Kelly. Fourier components of moving gratings. Behavior Research Methods, 14(6):435-437, 1982. 3", + "[22] Monika Lohani, Brennan R Payne, and David L Strayer. A review of psychophysiological measures to assess cognitive states in real-world driving. Frontiers in human neuroscience, 13:57, 2019. 2", + "[23] Ao Luo, Fan Yang, Kunming Luo, Xin Li, Haoqiang Fan, and Shuaicheng Liu. Learning optical flow with adaptive graph reasoning. In Proceedings of the AAAI conference on artificial intelligence, pages 1890-1898, 2022. 7, 8", + "[24] Nikolaus Mayer, Eddy Ilg, Philip Häusser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In IEEE International Conference on Computer Vision and Pattern Recognition (CVPR), 2016. arXiv:1512.02134. 2, 3, 4, 5, 6", + "[25] Lukas Mehl, Jenny Schmalfuss, Azin Jahedi, Yaroslava Naliyayko, and Andres Bruhn. Spring: A high-resolution high-detail dataset and benchmark for scene flow, optical flow and stereo. In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3, 4, 6", + "[26] Moritz Menze and Andreas Geiger. Object scene flow for autonomous vehicles. In Proceedings of the IEEE conference" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22807", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "on computer vision and pattern recognition, pages 3061-3070, 2015. 1, 2, 3, 4, 5, 6", + "[27] Shinji Nishimoto and Jack L. Gallant. A three-dimensional spatiotemporal receptive field model explains responses of area mt neurons to naturalistic movies. Journal of Neuroscience, 31(41):14551-14564, 2011. 3", + "[28] Mayu Otani, Riku Togashi, Yu Sawai, Ryosuke Ishigami, Yuta Nakashima, Esa Rahtu, Janne Heikkilä, and Shin'ichi Satoh. Toward verifiable and reproducible human evaluation for text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14277-14286, 2023. 1", + "[29] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730-27744, 2022. 1", + "[30] Christopher C Pack and Richard T Born. Temporal dynamics of a neural solution to the aperture problem in visual area mt of macaque brain. Nature, 409(6823):1040-1042, 2001. 2", + "[31] Jonathan Peirce and the PsychoPy team. Pavlovia: Online platform for running experiments. Open Science Tools, 2018. Accessed: 2024-10-25. 3", + "[32] Jonathan W. Peirce and Michael MacAskill. Building Experiments in PsychoPy. Sage Publications Ltd, London, United Kingdom, 2018. 3", + "[33] Silvia Pitzalis, Carlo Serra, Valerio Sulpizio, Giovanni d'Avossa, and Gaspare Galati. Neural bases of self- and object-motion in a naturalistic vision. Human Brain Mapping, 41(4):1084–1111, 2020. 3", + "[34] Anurag Ranjan, David T. Hoffmann, Dimitrios Tzionas, Siyu Tang, Javier Romero, and Michael J. Black. Learning multi-human optical flow. International Journal of Computer Vision (IJCV), 2020. 2, 3, 4, 5, 6", + "[35] Stephan R. Richter, Zeeshan Hayder, and Vladlen Koltun. Playing for benchmarks. In IEEE International Conference on Computer Vision, ICCV 2017, Venice, Italy, October 22-29, 2017, pages 2232-2241, 2017. 2, 3, 4, 5", + "[36] Laura Sevilla-Lara, Yiyi Liao, Fatma Güney, Varun Jampani, Andreas Geiger, and Michael J Black. On the integration of optical flow and action recognition. In Pattern Recognition: 40th German Conference, GCPR 2018, Stuttgart, Germany, October 9-12, 2018, Proceedings 40, pages 281-297. Springer, 2019. 1", + "[37] Xiaoyu Shi, Zhaoyang Huang, Weikang Bian, Dasong Li, Manyuan Zhang, Ka Chun Cheung, Simon See, Hongwei Qin, Jifeng Dai, and Hongsheng Li. Videoflow: Exploiting temporal cues for multi-frame optical flow estimation. arXiv preprint arXiv:2303.08340, 2023. 2, 7, 8", + "[38] Fabio Solari, Manuela Chessa, Narasimha K. Medathati, and Pierre Kornprobst. What can we expect from a v1-mt feedforward architecture for optical flow estimation? Signal Processing: Image Communication, 39:342-354, 2015. 2, 7, 8", + "[39] Shuyang Sun, Zhanghui Kuang, Lu Sheng, Wanli Ouyang, and Wei Zhang. Optical flow guided feature: A fast and robust motion representation for video action recognition. In" + ], + "bbox": [ + 91, + 92, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1390-1399, 2018. 1", + "[40] Zitang Sun, Yen-Ju Chen, Yung-Hao Yang, and Shin'ya Nishida. Modeling human visual motion processing with trainable motion energy sensing and a self-attention network. In Advances in Neural Information Processing Systems, 2023. 2, 7, 8", + "[41] Zitang Sun, Yen-Ju Chen, Yung-Hao Yang, and Shin'ya Nishida. Acquisition of second-order motion perception by learning to recognize the motion of objects made by nondiffusive materials. Journal of Vision, Vision Sciences Society Annual Meeting Abstract, 24(9), 2024. 7, 8", + "[42] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Proceedings of the European Conference on Computer Vision (ECCV), pages 402-419, 2020. 2, 7, 8", + "[43] Longguang Wang, Yulan Guo, Li Liu, Zaiping Lin, Xinpu Deng, and Wei An. Deep video super-resolution using hr optical flow estimation. IEEE Transactions on Image Processing, 29:4323-4336, 2020. 1", + "[44] Wenshan Wang, Delong Zhu, Xiangwei Wang, Yaoyu Hu, Yuheng Qiu, Chen Wang, Yafei Hu, Ashish Kapoor, and Sebastian Scherer. Tartanair: A dataset to push the limits of visual slam. arXiv preprint arXiv:2003.14338, 2020. 3, 4, 5", + "[45] H. Wu, Y. Liu, Z. Zhang, L. Du, L. Wang, Y. Wang, D. Lin, and L. Lin. Q-Instruct: Improving Low-Level Visual Abilities for Multi-Modality Foundation Models. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 25490–25500, Seattle, WA, USA, 2024. IEEE. 1", + "[46] Yung-Hao Yang, Zitang Sun, Taiki Fukiage, and Shin'ya Nishida. HuPerFlow: A comprehensive benchmark for human vs. machine motion estimation comparison. https://kucognitiveinformaticslab.github.io/Huperflow-Website/. Accessed: March 2025. 5, 8", + "[47] Yung-Hao Yang, Taiki Fukiage, Zitang Sun, and Shin'ya Nishida. Psychophysical measurement of perceived motion flow of naturalistic scenes. iScience, 26(12), 2023. 2, 3, 6, 7, 8", + "[48] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 1", + "[49] Xizhou Zhu, Yuwen Xiong, Bo Dai, and Dahua Lin. Deep feature flow for video recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2349-2358, 2017. 8" + ], + "bbox": [ + 516, + 92, + 903, + 768 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22808", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/HuPerFlow_ A Comprehensive Benchmark for Human vs. Machine Motion Estimation Comparison/5a790d2d-155e-42dc-bc07-eb0a483c9c11_model.json b/2025/HuPerFlow_ A Comprehensive Benchmark for Human vs. Machine Motion Estimation Comparison/5a790d2d-155e-42dc-bc07-eb0a483c9c11_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1e8c97b7bc4b7d5582c9ab53b1400d770fcac44b --- /dev/null +++ b/2025/HuPerFlow_ A Comprehensive Benchmark for Human vs. Machine Motion Estimation Comparison/5a790d2d-155e-42dc-bc07-eb0a483c9c11_model.json @@ -0,0 +1,1782 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.12, + 0.13, + 0.88, + 0.177 + ], + "angle": 0, + "content": "HuPerFlow: A Comprehensive Benchmark for Human vs. Machine Motion Estimation Comparison" + }, + { + "type": "text", + "bbox": [ + 0.212, + 0.203, + 0.783, + 0.221 + ], + "angle": 0, + "content": "Yung-Hao Yang\\(^{1}\\) Zitang Sun\\(^{1}\\) Taiki Fukiage\\(^{2}\\) Shin'ya Nishida\\(^{1,2*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.134, + 0.222, + 0.864, + 0.257 + ], + "angle": 0, + "content": "\\(^{1}\\)Cognitive Informatics Lab, Graduate School of Informatics, Kyoto University, Japan \n\\(^{2}\\)Communication Science Laboratories, Nippon Telegraph and Telephone Corporation, Japan" + }, + { + "type": "text", + "bbox": [ + 0.22, + 0.26, + 0.771, + 0.275 + ], + "angle": 0, + "content": "https://kucognitiveinformaticslab.github.io/Hyperflow-Website/" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.309, + 0.327, + 0.326 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.343, + 0.485, + 0.705 + ], + "angle": 0, + "content": "As AI models are increasingly integrated into applications involving human interaction, understanding the alignment between human perception and machine vision has become essential. One example is the estimation of visual motion (optical flow) in dynamic applications such as driving assistance. While there are numerous optical flow datasets and benchmarks with ground truth information, human-perceived flow in natural scenes remains underexplored. We introduce HuPerFlow—a benchmark for human-perceived flow, measured at 2,400 locations across ten optical flow datasets, with \\(\\sim 38,400\\) response vectors collected through online psychophysical experiments. Our data demonstrate that human-perceived flow aligns with ground truth in spatiotemporally smooth locations while also showing systematic errors influenced by various environmental properties. Additionally, we evaluated several optical flow algorithms against human-perceived flow, uncovering both similarities and unique aspects of human perception in complex natural scenes. HuPerFlow is the first large-scale human-perceived flow benchmark for alignment between computer vision models and human perception, as well as for scientific exploration of human motion perception in natural scenes. The HuPerFlow benchmark is publicly available on the HuPerFlow website." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.738, + 0.222, + 0.753 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Identifying object motion within the natural environment is crucial for humans, as it supports essential tasks such as navigation, tracking, prediction, and pursuit. In computer vision (CV), optical flow—representing an object's movement in pixel coordinates—serves as a foundational spatiotemporal feature for downstream applications, such as action recognition [36, 39], driver assistance [8, 26], robot navigation [6], video processing [43], etc. With the rise of data-driven approaches, the availability of datasets and" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.31, + 0.906, + 0.579 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.592, + 0.907, + 0.66 + ], + "angle": 0, + "content": "Figure 1. Example of HuPerFlow. The red arrows indicate human response to perceived flow, and the green arrows indicate ground truth motion vectors. The circles indicate the magnitudes of endpoint errors, the difference between perceived flow and ground truth." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.685, + 0.906, + 0.76 + ], + "angle": 0, + "content": "benchmarks has consistently fueled advancements in optical flow methods. Numerous studies have leveraged physical engines to generate realistic synthetic data [5] or employed specialized laser and depth cameras [13] to capture ground-truth optical flow in real scenes." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.765, + 0.907, + 0.903 + ], + "angle": 0, + "content": "However, few studies have examined how humans perceive motion in natural scenes. AI models increasingly interact with humans in various application fields, such as image quality assessment [45, 48], video generation [28], and large language model evaluation [29]). Human responses often guide the development of models, which in turn enhances the models' support of human activities. In the context of visual motion perception, estimating human-perceived flow in addition to the ground truth will be use" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22799" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.482, + 0.38 + ], + "angle": 0, + "content": "ful, for example, in automatic driving [7, 10, 22]. While fully autonomous driving systems can be solely based on the physical ground truth estimation, partially autonomous systems that assist human drivers will be empowered by the ability to predict when, where, and how the drivers are likely to misinterpret motion flow. Another potential use for our dataset is animation synthesis. To control the quality of animation, the animators need to check whether human viewers perceive the movements of the created animation in the way they originally intended. If there are disagreements, the animators should modify the images to minimize the errors. This process can be advanced if human-perceived motion can be automatically predicted. Machine prediction of human-perceived flow is a hard problem for hand-drawn animations since there is no physical ground truth. Even when the animation is rendered from a physically correct dynamic model of the scene, human-perceived motion can be distorted by many factors, including frame rates, image resolution, and the touch of the painting." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.383, + 0.482, + 0.595 + ], + "angle": 0, + "content": "Humans possess a sophisticated visual system capable of estimating object movement in complex natural scenes; however, the underlying mechanisms differ significantly from those used in current CV models. Motion inference is inherently an under-constrained problem (such as the aperture problem [30]), and the solution in humans is a perception that unfolds over a spatiotemporal process spanning multiple frames [1]. In contrast, CV methods typically focus on deterministic dense correspondence between consecutive frames. This difference can sometimes result in entirely different motion perceptions between human vision and CV algorithms, as seen in phenomena like the reverse phi [2] and missing fundamental illusions [4] investigated in vision science and psychology." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Despite this significant gap, currently, no large-scale benchmark directly evaluates the relationship between CV model responses, physical ground truth, and human perceptual responses to motion in natural scenes, thus hindering further research within this field. To address this problem, we construct the first large-scale comprehensive benchmark for human-perceived optical flow across several types of naturalistic scenes. Our study introduces several unique features compared to prior work: 1): Human-centric labeling: The dataset's labels are directly derived from human perception, gathered through rigorously designed psychophysical experiments. We refer to these human-perceived labels as \"Human Perception\" and use the term \"Ground Truth (GT)\" for the physical-correct labels generated by conventional methods (Fig. 1); 2) Large-scale: our dataset includes around 38,400 data points (trials) derived from 480 participant sessions, spanning over 240 hours of collection time. The benchmark comprises ten representative computer vision motion datasets—such as Sintel [5], KITTI [26], Spring [25], VIPER [35], etc—each" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.257 + ], + "angle": 0, + "content": "with specifically probed locations and optical flow GT; 3) A novel experimental paradigm: Unlike traditional human assessments (e.g., image quality evaluation), collecting motion perception data is complex and time-intensive. Motion data collection requires a strict display setting, including a constant refresh rate, high resolution, and high-speed hardware to ensure them. Our innovative online experimental paradigm, grounded in a psychophysical method developed for a lab experiment [47], ensures stable and reliable data collection even in online settings, making large-scale human-perceived motion data feasible." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.259, + 0.905, + 0.53 + ], + "angle": 0, + "content": "In addition to constructing this benchmark dataset, we conducted several initial comparative analyses between several representative CV models and human responses, including spatiotemporal gradient-based methods [11], recurrent refinement models [42], multi-frame methods [37], vision transformers [14], human-inspired approaches [38, 40], etc. Our findings reveal several key differences between model predictions and human perception, as well as unique characteristics of human perception in complex natural scenes, including abrupt direction switches (e.g., VIPER [35]), motion blur and non-rigid motion (e.g., MPI Sintel [5]), object translation and rotation (e.g., Flythings3D [24]), camera motion (e.g., Monkaa [24]), and local versus global motion (e.g., MHOF [34]). While the recent CV models correlate closely with human perception, they often exhibit low partial correlation when controlling the ground-truth effects, suggesting their design inherently fails to capture humans' unique perceptual biases." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.533, + 0.905, + 0.713 + ], + "angle": 0, + "content": "In summary, our contributions include: (1) the first large-scale human-perceived optical flow benchmark across ten diverse motion scenarios, (2) a new online psychophysics-based experimental paradigm enabling the large-scale collection of precise human-perceived optical flow in natural scenes, and (3) an evaluation of ten different representative models from both computer vision and human vision fields, revealing significant gaps between human motion interpretation and model predictions. These insights advance computer vision research on human-aligned AI and support the development of technologies for better interaction and assistance with humans." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.731, + 0.783, + 0.748 + ], + "angle": 0, + "content": "2. Previous optical flow datasets" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.757, + 0.788, + 0.773 + ], + "angle": 0, + "content": "2.1. Computer optical flow datasets" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.905, + 0.901 + ], + "angle": 0, + "content": "There are several optical flow datasets with physical motion ground truth, and we selected some of them based on motion content, image resolution, sequence count, and frames per sequence for online human observers. KITTI flow 2015 [26] is a real-world driving condition, where ground-truth optical flows were captured by stereo camera, laser scanner, and GPS. Virtual KITTI 2 [12] is a photo-realistic synthetic driving dataset featuring different weather conditions" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22800" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.092, + 0.482, + 0.211 + ], + "angle": 0, + "content": "and viewing angles. The Driving subset of the Scene Flow Datasets [24] includes eight synthetic sequences (slow / fast \\(\\times\\) 15mm / 300mm focal length \\(\\times\\) forwards / backward scene). VIPER [35] and TartanAir [44] provide synthetic video sequences with various speed and direction changes, representing diverse navigation scenarios beyond typical driving benchmarks for distinct object motion and camera motion." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.214, + 0.486, + 0.455 + ], + "angle": 0, + "content": "In contrast to automotive navigation scenarios, MPI Sintel [5] and SPRING [25] feature long-range, non-rigid motion with discontinue, large jumps, motion blur, and varying illumination, scene, and material properties. Monkaa, also from Scene Flow Datasets [24], includes significant camera translation and rotation, crucial for revealing motion perception influenced by surrounding objects and background. MHOF [34] captures diverse human motion, featuring both local and global motions across various postures and actions. Flythings3D subset of the Scene Flow Datasets [24], includes multiple geometric shapes and objects translating and rotating along complex 3D trajectories. Since the video presentation required several frames for human observers, we excluded well-known benchmarks such as Middlebury [3] (2-8 frames), FlyingChairs2 [16] (2 frames), ChairsSDHom [15] (2 frames), etc." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.469, + 0.396, + 0.485 + ], + "angle": 0, + "content": "2.2. Human motion perception datasets" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.493, + 0.483, + 0.869 + ], + "angle": 0, + "content": "Previous studies on human visual motion perception generally used well-controlled artificial visual stimuli, such as sine-wave gratings [21], random-dot kinematograms [20], and biological motion [19], to investigate specific processing mechanisms. Some studies used more complex natural-scenes stimuli to explore neural responses for them [27, 33]. While these studies provided valuable insights into human motion perception, only one study [47], to our knowledge, has directly reported human-perceived vectors for nature scenes, and compared them with ground-truth motion vectors and computer vision model predictions. This psychophysical study revealed that human-perceived optical flow shows various perceptual deviations from the ground-truth flow. Note that human-perceived vectors in dynamical movies, which are mainly based on motion energy within a \\(\\sim 100\\) ms temporal window [1], could be significantly different from the correspondence of salient feature points between adjacent frames (cf. [9]). However, [47] examined only five movies, all selected from the SlowFlow version (1008 frames per second) of MPI Sintel benchmark [18], in which the jump sizes between adjacent frames are substantially smaller than those of the standard version used in the current study. It is also worth noting that the movies in [47] were presented through a circular aperture centered on the probed location." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.482, + 0.901 + ], + "angle": 0, + "content": "In the current study, we adopt an approach similar to [47] with significant modifications: (1) using ten optical flow" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "benchmarks that are widely used for training and testing in computer vision fields, (2) choosing a larger number of motion clips with dynamic and diverse spatiotemporal properties, and (3) presenting the entire image on the display, as observers would naturally perceive it, rather than through a central aperture. This approach provides a greater motion context and ecological validity." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.211, + 0.738, + 0.227 + ], + "angle": 0, + "content": "3. The HuPerFlow dataset" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.235, + 0.618, + 0.251 + ], + "angle": 0, + "content": "3.1. Methods" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.258, + 0.682, + 0.272 + ], + "angle": 0, + "content": "3.1.1. Ethical statement" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.277, + 0.905, + 0.337 + ], + "angle": 0, + "content": "This study was approved by the Research Ethics Committee of the Graduate School of Informatics at Kyoto University. All participants received online informed consent before the experiments." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.339, + 0.724, + 0.354 + ], + "angle": 0, + "content": "3.1.2. Stimuli and Apparatus" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.358, + 0.907, + 0.689 + ], + "angle": 0, + "content": "The experiment was conducted using PsychoPy [32], a software for building behavioral experiments, and Pavlovia [31], an online platform for running these experiments in a web browser. To account for diverse hardware environments in online experiments, we controlled the visual presentation in both spatial and temporal domains. Spatially, to standardize viewing angles, observers performed a \"credit card calibration,\" adjusting an on-screen virtual credit card to match a physical one. This allowed us to calculate the physical pixel size after observers adjusted their viewing distance to ensure that 50 pixels corresponded to \\(1^{\\circ}\\) of visual angle, maintaining consistent stimulus pixel sizes across devices. Temporally, visual stimuli were presented at \\(30\\mathrm{Hz}\\). Since optimal timing control is crucial for online motion perception experiments, we did not use a higher frame rate for which frame drops were more likely to occur under some users' environments. In addition, We recorded frame drops and onset times during online experiments (see Supplementary Appendix 1 for more details). The frame dropping rate was .041%, and the frame onset error was -0.055 ms \\((\\mathrm{SD} = 3.664\\mathrm{ms})\\), indicating stable and precise visual stimulus presentation across devices." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.905, + 0.87 + ], + "angle": 0, + "content": "This project contained two phases: a preliminary Random Dot Kinematogram (RDK) for training and selection of the observers, and a main session to collect human-perceived flow data using multiple optical flow datasets. In the RDK phase, 5,000 black and white dots moved uniformly within a 600-pixel circular aperture. In the main experiment, we used ten optical flow datasets including KITTI Flow 2015 [26], Virtual KITTI 2 [12], Driving [24], VIPER [35], TartanAir [44], MPI Sintel Flow [5], SPRING [25], Monkaa [24], MHOF [34], and Flythings3D [24]. As most benchmarks include sub-datasets, we present our selection in detail in Tab. 1." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "We applied a standard movie-formating method across datasets with the following exceptions: (1) Most datasets" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22801" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.214, + 0.06, + 0.784, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.254, + 0.907, + 0.337 + ], + "angle": 0, + "content": "Figure 2. Experimental procedure: At the start of each trial, a green circle appeared to mark the selected area. Next, a motion sequence and a matching stimulus were presented alternately until a response was made. For most optical flow datasets, the motion sequences consisted of 15 frames (500 ms). It was followed by an inter-stimulus interval (ISI) of 1000 ms, then a matching stimulus (500 ms), and a second ISI (1000 ms). A flash probe was shown in the middle frame (the 8th frame) of the motion sequence to indicate the timing and location of the target motion. The green circle location cue, four-dot placeholders, and matching stimulus were presented sequentially at the same location." + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.356, + 0.483, + 0.384 + ], + "angle": 0, + "content": "Table 1. Summary of the selected optical flow datasets. * indicates the image size is resized by cubic interpolation." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.395, + 0.499, + 0.65 + ], + "angle": 0, + "content": "
DatasetsResolution (pixels)SequencesTotal FramesSelection
KITTI 2015 [26]1,242 × 3752004,200Training, left image
Virtual KITTI 2 [12]1,242 × 375521,260Clone, camera_0
Driving [24]960 × 54082,400Finalpass, left, slow
VIPER [35]960 × 540*77134,097Val, jpg
TartanAir [44]640 × 480166,982image_left
MPI Sintel [5]1,024 × 436231,064Training, Final
Spring [25]960 × 540*375,000Train_frame_left
Monkaa [24]960 × 540248,664Finalpass, left
MHOF [34]640 × 6401,23895,868Train, masked with segm_EXR
FlyingThings [24]960 × 5402,18021,818Val, image.clean, left
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.484, + 0.901 + ], + "angle": 0, + "content": "were presented at their original sizes, except for VIPER [35] and SPRING [25], which were resized from \\(1920 \\times 1080\\) to \\(960 \\times 540\\) for online observers, with the optical flow adjusted accordingly. (2) Frames F-7 to F+7 around a selected testing frame were used as movie sequences, except for MHOF [34] and Flythings3D [24], which were limited to F-4 to F+5 due to each sequence having only ten frames. (3) Virtual KITTI2 [12], SPRING [25] and Scene Flow Datasets [24] (i.e., Driving, Monkaa, and Flythings3D) contain both future (forward) and past (backward) flow directions. However, only future-directed flows were reported as video images were presented sequentially in the forward direction, frame-by-frame. (4) We did not modify the original images, except for MHOF [34] dataset, where frames contained irrelevant background scenes. To standardize the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.358, + 0.907, + 0.389 + ], + "angle": 0, + "content": "background, we used the \"segm_EXR\" provided by MHOF to render it as a uniform gray background." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.394, + 0.715, + 0.408 + ], + "angle": 0, + "content": "3.1.3. Design and Procedure" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.412, + 0.907, + 0.579 + ], + "angle": 0, + "content": "In both the RDK and main experiments, each session began with six practice trials followed by 80 formal trials. The practice trials familiarized observers with the procedure and helped them understand the differences between the ground truth and their perceived vectors. To support this, we provided visual feedback of both GT and perceived vectors after each response during practice trials. If the angle between perceived vectors and the GT exceeded \\(30^{\\circ}\\), or if the endpoint error (EPE, the Euclidean distance between the human responses and GT, Equation 1.) was greater than 10 pixels, the same practice trial would be repeated." + }, + { + "type": "equation", + "bbox": [ + 0.566, + 0.586, + 0.905, + 0.61 + ], + "angle": 0, + "content": "\\[\n\\mathrm {E P E} = \\sqrt {\\left(u _ {\\text {R e s p}} - u _ {\\mathrm {G T}}\\right) ^ {2} + \\left(v _ {\\text {R e s p}} - v _ {\\mathrm {G T}}\\right) ^ {2}} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.907, + 0.87 + ], + "angle": 0, + "content": "In the RDK session, the direction \\((0^{\\circ} - 360^{\\circ})\\) and speed (1-20 pixels per frame, PPF) of the dots were varied randomly across trials but kept constant within a trial. During formal trials in this session, feedback was also shown to reinforce training. In the main experimental session, we used ten datasets as previously described, selecting 24 sequences per dataset. Each sequence included ten target motion locations with speeds under 20 PPF at a specific frame. At each probed location, we collected 16 repeated trials (four repetitions per observer \\(\\times\\) four observers) to obtain a stable measure of perceived flow, totaling 2,400 probed locations across 38,400 trials. In each session, participants viewed two movie sequences with ten target motion locations and four repetitions each (80 trials in total). This yielded 480 sessions across 10 datasets \\(\\times\\) 12 sessions (2 movie sequences each) \\(\\times\\) 4 observers. No feedback was provided during formal trials in the main experiments." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Fig. 2 illustrates the display layout and procedure in the main experiment. Both the RDK and main experiments" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22802" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.364 + ], + "angle": 0, + "content": "used a method of adjustment to measure human-perceived flow at the flash-probed spatiotemporal location. Observers reported the perceived flow of a target motion by matching it with a circular Brownian noise pattern \\((1/f^2, 120\\) pixels in diameter), which has a spectrum similar to natural images and gives the least motion aliasing for a wide range of motion speeds. Using a mouse cursor (a purple dot, 10 pixels in diameter) within a circular control panel (600 pixels in diameter), observers could freely adjust the direction \\((0^\\circ - 360^\\circ)\\) and speed (1-20 pixels per frame) of the matching Brownian noise. The direction was mapped to angles \\((\\theta)\\) in a polar coordinate, and speed was adjusted logarithmically based on the cursor's distance from the center of the panel. This method provides more accurate measurements of perceived motion vectors than subjective reports. The control panel for the matching stimulus was always fixed at the display's center. See Supplementary Video S1 for the animated procedure." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.376, + 0.483, + 0.618 + ], + "angle": 0, + "content": "In the RDK session, where direction and speed were uniform within each trial, both the target motion and matching stimulus were presented at the center of the aperture. In the main experiment, a green circle (120 pixels in diameter), presented before each trial, served as a spatial cue indicating the target motion area. Additionally, four additional dots (green dots, 10 pixels in diameter) served as placeholders throughout the trial. They were positioned 60 pixels above, below, on the left, and on the right of the probed dot. For most datasets, a repetition cycle (see below) consisted of the target movie sequence (500 ms, 15 frames), an interstimulus interval (ISI, 1000 ms), and the matching stimulus (500 ms), followed by a second ISI (1000 ms) before the next cycle. For the MHOF [34] and FlyingThings3D [24] datasets, which contain only ten frames (333 ms), the second ISI was adjusted to 1167 ms to keep the cycle duration consistent." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.902 + ], + "angle": 0, + "content": "In each trial, a flash-probed dot (a green dot, seven pixels in diameter) was presented at a selected testing frame and location to indicate the timing and position of the target motion. The flash probe appeared in the middle of each movie sequence presentation (typically at the \\(8^{\\text{th}}\\) frame, and at the \\(5^{\\text{th}}\\) frame for MHOF and FlyingThings3D), indicating when observers should respond. The probed target motions were intentionally selected from informative locations, including main objects, the background, and minor objects. In the main experiment, the four-dot placeholders appeared around the probed location, with the matching stimulus displayed directly at that placeholder's location. During each trial, observers could view the movie sequences and match Brownian noise multiple times (repetition cycles), while responses were only accepted after completing three repetition cycles. Since the target motion was shown for just one frame, this adjustment process allowed observers to refine their responses through repeated viewing." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.092, + 0.608, + 0.106 + ], + "angle": 0, + "content": "3.2. Results" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.114, + 0.702, + 0.129 + ], + "angle": 0, + "content": "3.2.1. Data quality control" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.133, + 0.905, + 0.389 + ], + "angle": 0, + "content": "The study included 55 observers (mean age = 22.18, SD = 2.82, 24 females and 31 males) who participated in a preliminary Random Dot Kinematogram (RDK) experiment. To ensure reliable perceived flow data, we initially selected observers whose correlation coefficient between perceived vectors and ground truth vectors exceeded .90 in the RDK session. Nine observers with low correlation (individual correlation range: .501 to .885) and four who were unable to complete the experiment were excluded. The group-level correlation for the 42 remaining observers was high (Pearson correlation coefficient \\( r = .955 \\), with individual correlations ranging from .917 to .985), demonstrating their ability to reproduce perceived motion accurately. These 42 observers (mean age = 22.19, SD = 2.46; 17 females, 25 males) then proceeded to the main experimental sessions. They participated in 480 sessions since some observers attended multiple different sessions." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.399, + 0.822, + 0.413 + ], + "angle": 0, + "content": "3.2.2. Perceived vectors and \"flow illusions\"" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.418, + 0.905, + 0.689 + ], + "angle": 0, + "content": "In Fig. 3, we show representative examples of averaged human responses for each dataset (see Supplementary Video S2 for target motions and the corresponding averaged human responses across all probed locations within each dataset). Our findings reveal systematic and stable human-specific biases in motion perception under certain scenarios. We termed these systematic discrepancies between human responses and ground truth motions as \"flow illusions\". These deviations were not random errors but provided valuable insights into the underlying processing and mechanism of human motion perception. On the other hand, we also observed individual differences, with an average inter-observers correlation of 0.407 (SD: 0.417; range from -0.329 to 0.990) across all probed locations. These Individual variations would be useful for training models to reproduce the distribution of human motion perception. To facilitate further research, we have made all raw individual data publicly available on the HuPerFlow website [46]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.69, + 0.905, + 0.901 + ], + "angle": 0, + "content": "In most driving datasets such as KITTI 2015 [26], Virtual KITTI 2 [12], and Driving [24], the direction and speed of optical flows transition smoothly across frames, with perceived flows aligning closely with the ground truth motion. This consistency supports the reliability of our measurement method in capturing observers' perceptions under spatiotemporally smooth scenes. Notably, as optical flow locations deviated from the scene center, observers often perceived the directions to be slightly shifting toward the central region. This finding may reflect observers' attempts to cancel out the image motion components produced by self-motion (camera motion) during driving. In contrast, other navigation datasets like VIPER [35] and TartanAir [44] show abrupt direction changes, particularly during ac" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22803" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.214, + 0.063, + 0.784, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.109, + 0.355, + 0.885, + 0.371 + ], + "angle": 0, + "content": "Figure 3. Demonstration of human perceived motion vectors. The results cover ten representative existing optical flow benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.397, + 0.482, + 0.458 + ], + "angle": 0, + "content": "tions such as turning, lane changing, or flying backward. These involve interactions between image motion and camera motion, with perceived flows often revealing systematic errors primarily driven by camera motion." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.459, + 0.483, + 0.639 + ], + "angle": 0, + "content": "In datasets like MPI Sintel [5] and SPRING [25], where images include blurred, non-rigid, discontinuous motion and large jump size (cf. [47] using stimuli with small jumps), observers generally failed to discern individual locations, perceiving them as part of a unified grouping movement instead. Unlike the datasets described in the previous paragraph, which were viewed with smoothly moving cameras, Monkaa [24] dataset involved significant camera rotation around relatively stationary objects. Observers often perceived these objects as moving due to the motion of surrounding objects or backgrounds, illustrating the phenomena of relative and induced motion." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.643, + 0.483, + 0.822 + ], + "angle": 0, + "content": "In the MHOF [34] datasets, human figures generally exhibit global body movement accompanied by local motions from the head and limbs. Although the motion structures are less complex than in MPI Sintel and Spring, observers tended to focus on the direction of the global body movement and ignored the local gestures. In the Flythings3D [24], multiple objects float along 3D random trajectories, requiring observers to perform object-ground and object-object segmentation. They were successful in perceiving motion against simple backgrounds and distinct surrounding objects but made more errors in cluttered, messy conditions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.903 + ], + "angle": 0, + "content": "As different datasets revealed various types of errors, we calculated the correlation between human response and ground truth for each dataset separately. In general, the driving datasets (i.e., KITTI 2015 [26], Virtual KITTI 2 [12], and Driving [24]) showed high correlations, but oth" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.397, + 0.907, + 0.427 + ], + "angle": 0, + "content": "ers showed middle to low correlation (see Supplementary Appendix 2 for the scatter plot of each dataset)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.44, + 0.845, + 0.457 + ], + "angle": 0, + "content": "3.2.3. EPE in relation to optical flow properties" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.461, + 0.907, + 0.672 + ], + "angle": 0, + "content": "As shown in Fig. 3, the EPE of human response (indicated by the diameter of the green circle) varied across different positions and scenes. Following [5], we evaluated how EPE changed in relation to various optical flow properties. Fig. 4 shows that EPE increased as GT speed increased, which suggests that faster motions imposed more significant computational challenges. To assess temporal flow dynamics, we compared optical flow from the previous frame (t-1) to the selected probed frame (t-0) and analyzed the temporal gradient changes in uv-vectors over two intervals: from t-1 to t-0, and from t-0 to t+1 (ground truth), normalized by GT speed from the same location. Generally, we observed that EPE increased as the normalized temporal flow gradient increased up to the ratio of 2, and then leveled off." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.901 + ], + "angle": 0, + "content": "We also considered spatial image properties by using a Sobel kernel to derive the image gradient, calculating the average edge magnitude within a \\(\\pm 40\\)-pixel neighborhood around the probed location. The results indicate that EPE decreased as the image gradient increased. Additionally, using instance boundary from the MPI Sintel dataset, we calculated the shortest distance to the boundary and found that EPE also decreased as the distance increased. These findings suggest that edges and boundaries may create ambiguities in motion flow interpretation and interrupt motion perception. Finally, using camera motion data from the MPI Sintel as a proxy for self-motion within a scene, we observed that EPE increased with self-motion, and this effect persisted even when excluding the effect of GT speed (image motion)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22804" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.172, + 0.069, + 0.391, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.393, + 0.07, + 0.605, + 0.165 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.61, + 0.071, + 0.822, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.172, + 0.171, + 0.391, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.393, + 0.171, + 0.609, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.61, + 0.171, + 0.824, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.282, + 0.908, + 0.325 + ], + "angle": 0, + "content": "Figure 4. Endpoint errors (EPEs) of human responses and two machine vision models as functions of optical flow properties. The horizontal axis of each panel is the GT speed, normalized temporal flow gradient, spatial image gradient, instance boundary distance, and self-motion with(out) normalized by the image GT speed. The vertical axis represents EPE with different scales for visualization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.339, + 0.484, + 0.445 + ], + "angle": 0, + "content": "In addition to the EPE of human response, Fig. 4 shows the EPEs of two machine vision models, VideoFlow and FFV1MT (see Sec. 4). While there were clear differences in the absolute level of performance, the optical flow properties generally affect EPE in similar ways for human and machine vision models, except for the effects of spatial image gradient effect on VideoFlow." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.45, + 0.458, + 0.468 + ], + "angle": 0, + "content": "4. Comparison with optical flow algorithms" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.477, + 0.483, + 0.748 + ], + "angle": 0, + "content": "We selected a range of representative optical flow algorithms to assess the similarities and differences among human-perceived flows, ground truth motion, and predictions from computer vision models. These included classical methods based on brightness constancy assumptions, such as Farneback [11]; biologically inspired approaches like FFV1MT [38]; multiscale inference models, such as FlowNet2 [15]; recurrent refinement models like RAFT [42]; transformer-based models like FlowFormer [14]; graph reasoning methods such as AGFlow [23]; multiframe approaches like VideoFlow [37]; Motion energy and self-attention based models like V1Attention [40] and DualModel [41]; as well as general-purpose architectures like Perceiver IO [17]. For a fair comparison, we used each model's default configuration or the best pre-trained model fine-tuned on the Sintel dataset. Since unsupervised learning methods approximate the supervised ground truth, we limited our comparison to supervised methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.903 + ], + "angle": 0, + "content": "For evaluation, we first calculated the average EPE and Pearson correlation coefficients between model predictions and both human responses and ground truth. To further understand the models' alignment with human perception independently of ground truth influence—since most models were trained to fit ground truth and the ground truth was inherently correlated with human responses—we also calculated the partial correlation between model predictions and human responses, controlling for ground truth, as shown in Equation 2. Here, \\( r_{x,y} \\) denotes correlation coefficients be" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.34, + 0.615, + 0.353 + ], + "angle": 0, + "content": "tween \\(x\\) and \\(y\\)." + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.365, + 0.905, + 0.417 + ], + "angle": 0, + "content": "\\[\n\\rho_ {\\text {m o d e l}} = r _ {\\text {r e s p , m o d e l} \\cdot \\mathrm {G T}} = \\frac {r _ {\\text {r e s p , m o d e l}} - r _ {\\text {r e s p , G T}} \\cdot r _ {\\text {m o d e l , G T}}}{\\sqrt {1 - r _ {\\text {r e s p , G T}} ^ {2}} \\cdot \\sqrt {1 - r _ {\\text {m o d e l , G T}} ^ {2}}}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.418, + 0.906, + 0.522 + ], + "angle": 0, + "content": "To complement partial correlation analysis, we also adopted a Response Consistency Index (RCI) from [47]. This index measures the similarity between model predictions and human responses at each probed location, approaching \\(+1\\) when the model's prediction aligns closely with human flow illusions (see Supplementary Appendix 3 for more details)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.906, + 0.704 + ], + "angle": 0, + "content": "Tab. 2 summarizes model predictions across all datasets. Overall, VideoFlow's predictions were the closest to the ground truth. RAFT showed the most closely resembling human responses while also having a strong alignment with the ground truth (second-best among models). When controlling for the ground truth effect, FFV1MT and V1Attention-StageI were the most similar to human flow illusions, as indicated by high partial correlations and average RCI values. A detailed analysis comparing the predictions of optical flow algorithms with both the ground truth and human responses for each dataset is provided in Tables S1-S12 of Supplementary Appendix 4." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Fig. 5 provides examples of model predictions, highlighting examples where predictions aligned closely with the ground truth (e.g., VideoFlow and RAFT) versus those having some resemblance to human response errors (e.g., FFV1MT and V1Attention-StageI). For instance, at Point (A-I) in Fig. 5, both flow illusions and FFV1MT exhibited temporal-sluggish responses but VideoFlow aligned well with GT. Conversely, point (A-II) in Fig. 5 both model predictions were similar to flow illusions, likely due to image signal loss, as the GT indicated a small, dark stone falling. At Point (B-I) and other selected body locations, both human perception and FFV1MT reflected grouping motion in the same direction, regardless of local gesture directions." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22805" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.064, + 0.907, + 0.108 + ], + "angle": 0, + "content": "Table 2. Predictions of optical flow algorithms versus Human response or ground truth (GT). \\( \\rho \\) : Partial correlation; \\( r \\) : Pearson correlation coefficient; EPE: vector end-point error; uv, dir, spd represent motion components in Cartesian space, direction, and speed, respectively. RCI is introduced from [47] to represent the model's similarity to human perception (the larger, the more human-aligned)." + }, + { + "type": "table", + "bbox": [ + 0.119, + 0.117, + 0.877, + 0.337 + ], + "angle": 0, + "content": "
MethodρRCIv.s. Humanv.s. GT
ρuvρdirρspdruvrdirrspdEPEruvrdirrspdEPE
Ground TruthNaNNaNNaNNaN0.660.430.296.961.001.001.000.00
Farnebäck [11]0.170.140.040.130.540.380.156.910.680.650.395.52
FFV1MT [38]0.310.280.060.150.630.450.206.580.700.590.545.84
RAFT [42]0.110.220.010.050.640.470.237.370.930.850.792.19
FlowNet2 [15]0.100.260.000.070.590.480.187.550.820.800.613.24
FlowFormer [14]0.070.19-0.020.040.640.460.217.430.930.870.792.08
VideoFlow [49]0.080.22-0.020.040.640.470.227.470.940.880.811.91
Perceiver IO [17]0.090.170.030.050.620.440.237.750.890.820.742.87
AGFlow [23]0.160.240.010.110.540.450.157.940.680.660.465.44
V1Attention-StageI [40]0.230.280.020.180.550.450.146.360.630.540.427.00
DualModel [41]0.200.250.060.130.570.450.207.310.700.670.535.66
" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.361, + 0.48, + 0.707 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.804 + ], + "angle": 0, + "content": "Figure 5. Predicted vectors of optical flow algorithms. Examples are FFV1MT [38], V1Attention [40], VideoFlow [37], and RAFT [42]. We marked some points with large RCI differences between the models in the left and right columns. Generally, the model on the left side shows a more human-aligned response over ground truth, while the right side shows the opposite result." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.902 + ], + "angle": 0, + "content": "At Point (C-I), a strong self-motion effect occurred toward a static motorcycle. Despite the strong physical optical flow in the image, human observers and V1Attention-StageI appeared to compensate for self-motion to perceive the motorcycle as stationary. For more examples, please visit the HuPerFlow website [46]. Note, however, that there remains" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.362, + 0.907, + 0.439 + ], + "angle": 0, + "content": "a large gap between human responses and the most similar model predictions, possibly because the purpose of human visual motion processing is not merely to estimate image flow. The primary purpose of making our dataset is to help future studies develop models that can fill this gap." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.479, + 0.634, + 0.496 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.514, + 0.907, + 0.756 + ], + "angle": 0, + "content": "By overcoming technical challenges, we developed a novel online paradigm to collect human motion perception data, introducing the first large-scale benchmark for human-perceived optical flow. Our findings identified distinct perceptual illusions across various scene scenarios, some aligning with and others diverging from the predictions of CV optical flow algorithms. Human motion perception, compared to state-of-the-art CV models, has limited spatial, temporal, and intensity resolutions and struggles with motion estimation in retinal coordinates. However, human motion vision has remarkable capabilities in analyzing hierarchical movement structures and distinguishing object motion from self-motion in world coordinates. Our proposed methods can be extended to collect broader human data from diverse datasets, further bridging gaps between CV, human perception, and ground truth." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "While our dataset may not be sufficient to train models from zero, it offers valuable benchmarks for model evaluation and fine-tuning. Physically accurate CV models are crucial, but aligning with human perception is equally important for human-centered applications. Notably, current optical flow models fail to predict various human-perceived motion illusions, highlighting the need to address this gap both for practical applications and for advancing the scientific understanding of human vision." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "22806" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.116, + 0.484, + 0.178 + ], + "angle": 0, + "content": "This work was supported in part by JSPS Grants-in-Aid for Scientific Research (KAKENHI), Grant Numbers JP20H00603, JP20H05957, and JP24H00721, and by the Spring Fellowship, Grant Number JPMJFS2123." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.192, + 0.188, + 0.208 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.217, + 0.484, + 0.258 + ], + "angle": 0, + "content": "[1] Edward H Adelson and James R Bergen. Spatiotemporal energy models for the perception of motion. Josa a, 2(2): 284-299, 1985. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.261, + 0.484, + 0.302 + ], + "angle": 0, + "content": "[2] Stuart M Anstis and Brian J Rogers. Illusory continuous motion from oscillating positive-negative patterns: Implications for motion perception. Perception, 15(5):627-640, 1986. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.304, + 0.484, + 0.358 + ], + "angle": 0, + "content": "[3] Simon Baker, Daniel Scharstein, J. P. Lewis, Stefan Roth, Michael J. Black, and Richard Szeliski. A Database and Evaluation Methodology for Optical Flow. International Journal of Computer Vision, 92(1):1-31, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.361, + 0.483, + 0.402 + ], + "angle": 0, + "content": "[4] Richard O Brown and Sheng He. Visual motion of missing-fundamental patterns: motion energy versus feature correspondence. Vision Research, 40(16):2135-2147, 2000. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.404, + 0.484, + 0.46 + ], + "angle": 0, + "content": "[5] Daniel J. Butler, Jonas Wulff, Garrett B. Stanley, and Michael J. Black. A naturalistic open source movie for optical flow evaluation. In European Conf. on Computer Vision (ECCV), pages 611-625. Springer-Verlag, 2012. 1, 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.462, + 0.484, + 0.516 + ], + "angle": 0, + "content": "[6] Haiyang Chao, Yu Gu, and Marcello Napolitano. A survey of optical flow techniques for robotics navigation applications. Journal of Intelligent & Robotic Systems, 73:361-372, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.519, + 0.484, + 0.587 + ], + "angle": 0, + "content": "[7] Chenyi Chen, Ari Seff, Alain Kornhauser, and Jianxiong Xiao. Deepdriving: Learning affordance for direct perception in autonomous driving. In Proceedings of the IEEE international conference on computer vision, pages 2722-2730, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.59, + 0.484, + 0.644 + ], + "angle": 0, + "content": "[8] Jaechan Cho, Yongchul Jung, Dong-Sun Kim, Seongjoo Lee, and Yunho Jung. Moving object detection based on optical flow estimation and a gaussian mixture model for advanced driver assistance systems. Sensors, 19(14):3217, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.647, + 0.484, + 0.728 + ], + "angle": 0, + "content": "[9] Carl Doersch, Ankush Gupta, Larisa Markeeeva, Adrià Recasens, Lucas Smaira, Yusuf Aytar, João Carreira, Andrew Zisserman, and Yi Yang. Tap-vid: A benchmark for tracking any point in a video. arXiv preprint arXiv:2211.03726, 2022. Published in NeurIPS Datasets and Benchmarks track, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.732, + 0.483, + 0.786 + ], + "angle": 0, + "content": "[10] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.483, + 0.843 + ], + "angle": 0, + "content": "[11] Gunnar Farnebäck. Two-frame motion estimation based on polynomial expansion. In Proceedings of the Scandinavian conference on Image analysis, pages 363-370. Springer, 2003. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[12] Adrien Gaidon, Qiao Wang, Yohann Cabon, and Eleonora Vig. Virtual worlds as proxy for multi-object tracking analysis. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2, 3, 4, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.217, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.148 + ], + "angle": 0, + "content": "[13] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.906, + 0.218 + ], + "angle": 0, + "content": "[14] Tzehoong Huang, Zhichao Zhang, Wenjie Zeng, Ming Liu, Chunjing Zhang, Hongsheng Li, and Shijian Lu. Flowformer: A transformer architecture for optical flow. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.221, + 0.906, + 0.302 + ], + "angle": 0, + "content": "[15] Eddy Ilg, Nikolaus Mayer, Tonmoy Saikia, Margret Keuper, Alexey Dosovitskiy, and Thomas Brox. Flownet 2.0: Evolution of optical flow estimation with deep networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2462-2470, 2017. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.304, + 0.906, + 0.374 + ], + "angle": 0, + "content": "[16] Eddy Ilg, Tonmoy Saikia, Margret Keuper, and Thomas Brox. Occlusions, motion and depth boundaries with a generic network for disparity, optical flow or scene flow estimation. In European Conference on Computer Vision (ECCV), 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.376, + 0.906, + 0.431 + ], + "angle": 0, + "content": "[17] Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, et al. Perceiver io: A general architecture for structured inputs & outputs. arXiv preprint arXiv:2107.14795, 2021. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.433, + 0.906, + 0.501 + ], + "angle": 0, + "content": "[18] Joel Janai, Fatma Güney, Jonas Wulff, Michael J Black, and Andreas Geiger. Slow flow: Exploiting high-speed cameras for accurate and diverse optical flow reference data. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3597-3606, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.503, + 0.906, + 0.544 + ], + "angle": 0, + "content": "[19] Gunnar Johansson. Visual perception of biological motion and a model for its analysis. Perception & Psychophysics, 14(2):201-211, 1973. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.546, + 0.906, + 0.574 + ], + "angle": 0, + "content": "[20] Bela Julesz. Foundations of Cyclopean Perception. University of Chicago Press, 1971. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.906, + 0.603 + ], + "angle": 0, + "content": "[21] D. H. Kelly. Fourier components of moving gratings. Behavior Research Methods, 14(6):435-437, 1982. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.605, + 0.906, + 0.659 + ], + "angle": 0, + "content": "[22] Monika Lohani, Brennan R Payne, and David L Strayer. A review of psychophysiological measures to assess cognitive states in real-world driving. Frontiers in human neuroscience, 13:57, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.661, + 0.906, + 0.716 + ], + "angle": 0, + "content": "[23] Ao Luo, Fan Yang, Kunming Luo, Xin Li, Haoqiang Fan, and Shuaicheng Liu. Learning optical flow with adaptive graph reasoning. In Proceedings of the AAAI conference on artificial intelligence, pages 1890-1898, 2022. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.718, + 0.906, + 0.8 + ], + "angle": 0, + "content": "[24] Nikolaus Mayer, Eddy Ilg, Philip Häusser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In IEEE International Conference on Computer Vision and Pattern Recognition (CVPR), 2016. arXiv:1512.02134. 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.802, + 0.906, + 0.871 + ], + "angle": 0, + "content": "[25] Lukas Mehl, Jenny Schmalfuss, Azin Jahedi, Yaroslava Naliyayko, and Andres Bruhn. Spring: A high-resolution high-detail dataset and benchmark for scene flow, optical flow and stereo. In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[26] Moritz Menze and Andreas Geiger. Object scene flow for autonomous vehicles. In Proceedings of the IEEE conference" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22807" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.483, + 0.12 + ], + "angle": 0, + "content": "on computer vision and pattern recognition, pages 3061-3070, 2015. 1, 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.177 + ], + "angle": 0, + "content": "[27] Shinji Nishimoto and Jack L. Gallant. A three-dimensional spatiotemporal receptive field model explains responses of area mt neurons to naturalistic movies. Journal of Neuroscience, 31(41):14551-14564, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.262 + ], + "angle": 0, + "content": "[28] Mayu Otani, Riku Togashi, Yu Sawai, Ryosuke Ishigami, Yuta Nakashima, Esa Rahtu, Janne Heikkilä, and Shin'ichi Satoh. Toward verifiable and reproducible human evaluation for text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14277-14286, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.483, + 0.345 + ], + "angle": 0, + "content": "[29] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730-27744, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.348, + 0.483, + 0.389 + ], + "angle": 0, + "content": "[30] Christopher C Pack and Richard T Born. Temporal dynamics of a neural solution to the aperture problem in visual area mt of macaque brain. Nature, 409(6823):1040-1042, 2001. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.483, + 0.432 + ], + "angle": 0, + "content": "[31] Jonathan Peirce and the PsychoPy team. Pavlovia: Online platform for running experiments. Open Science Tools, 2018. Accessed: 2024-10-25. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.483, + 0.475 + ], + "angle": 0, + "content": "[32] Jonathan W. Peirce and Michael MacAskill. Building Experiments in PsychoPy. Sage Publications Ltd, London, United Kingdom, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.477, + 0.483, + 0.532 + ], + "angle": 0, + "content": "[33] Silvia Pitzalis, Carlo Serra, Valerio Sulpizio, Giovanni d'Avossa, and Gaspare Galati. Neural bases of self- and object-motion in a naturalistic vision. Human Brain Mapping, 41(4):1084–1111, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.534, + 0.483, + 0.588 + ], + "angle": 0, + "content": "[34] Anurag Ranjan, David T. Hoffmann, Dimitrios Tzionas, Siyu Tang, Javier Romero, and Michael J. Black. Learning multi-human optical flow. International Journal of Computer Vision (IJCV), 2020. 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.591, + 0.483, + 0.646 + ], + "angle": 0, + "content": "[35] Stephan R. Richter, Zeeshan Hayder, and Vladlen Koltun. Playing for benchmarks. In IEEE International Conference on Computer Vision, ICCV 2017, Venice, Italy, October 22-29, 2017, pages 2232-2241, 2017. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.648, + 0.483, + 0.731 + ], + "angle": 0, + "content": "[36] Laura Sevilla-Lara, Yiyi Liao, Fatma Güney, Varun Jampani, Andreas Geiger, and Michael J Black. On the integration of optical flow and action recognition. In Pattern Recognition: 40th German Conference, GCPR 2018, Stuttgart, Germany, October 9-12, 2018, Proceedings 40, pages 281-297. Springer, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.733, + 0.483, + 0.801 + ], + "angle": 0, + "content": "[37] Xiaoyu Shi, Zhaoyang Huang, Weikang Bian, Dasong Li, Manyuan Zhang, Ka Chun Cheung, Simon See, Hongwei Qin, Jifeng Dai, and Hongsheng Li. Videoflow: Exploiting temporal cues for multi-frame optical flow estimation. arXiv preprint arXiv:2303.08340, 2023. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.483, + 0.857 + ], + "angle": 0, + "content": "[38] Fabio Solari, Manuela Chessa, Narasimha K. Medathati, and Pierre Kornprobst. What can we expect from a v1-mt feedforward architecture for optical flow estimation? Signal Processing: Image Communication, 39:342-354, 2015. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[39] Shuyang Sun, Zhanghui Kuang, Lu Sheng, Wanli Ouyang, and Wei Zhang. Optical flow guided feature: A fast and robust motion representation for video action recognition. In" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1390-1399, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[40] Zitang Sun, Yen-Ju Chen, Yung-Hao Yang, and Shin'ya Nishida. Modeling human visual motion processing with trainable motion energy sensing and a self-attention network. In Advances in Neural Information Processing Systems, 2023. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[41] Zitang Sun, Yen-Ju Chen, Yung-Hao Yang, and Shin'ya Nishida. Acquisition of second-order motion perception by learning to recognize the motion of objects made by nondiffusive materials. Journal of Vision, Vision Sciences Society Annual Meeting Abstract, 24(9), 2024. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.317 + ], + "angle": 0, + "content": "[42] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Proceedings of the European Conference on Computer Vision (ECCV), pages 402-419, 2020. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.319, + 0.905, + 0.374 + ], + "angle": 0, + "content": "[43] Longguang Wang, Yulan Guo, Li Liu, Zaiping Lin, Xinpu Deng, and Wei An. Deep video super-resolution using hr optical flow estimation. IEEE Transactions on Image Processing, 29:4323-4336, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.376, + 0.905, + 0.432 + ], + "angle": 0, + "content": "[44] Wenshan Wang, Delong Zhu, Xiangwei Wang, Yaoyu Hu, Yuheng Qiu, Chen Wang, Yafei Hu, Ashish Kapoor, and Sebastian Scherer. Tartanair: A dataset to push the limits of visual slam. arXiv preprint arXiv:2003.14338, 2020. 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.905, + 0.514 + ], + "angle": 0, + "content": "[45] H. Wu, Y. Liu, Z. Zhang, L. Du, L. Wang, Y. Wang, D. Lin, and L. Lin. Q-Instruct: Improving Low-Level Visual Abilities for Multi-Modality Foundation Models. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 25490–25500, Seattle, WA, USA, 2024. IEEE. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.517, + 0.905, + 0.585 + ], + "angle": 0, + "content": "[46] Yung-Hao Yang, Zitang Sun, Taiki Fukiage, and Shin'ya Nishida. HuPerFlow: A comprehensive benchmark for human vs. machine motion estimation comparison. https://kucognitiveinformaticslab.github.io/Huperflow-Website/. Accessed: March 2025. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.587, + 0.905, + 0.641 + ], + "angle": 0, + "content": "[47] Yung-Hao Yang, Taiki Fukiage, Zitang Sun, and Shin'ya Nishida. Psychophysical measurement of perceived motion flow of naturalistic scenes. iScience, 26(12), 2023. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.643, + 0.905, + 0.713 + ], + "angle": 0, + "content": "[48] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.715, + 0.905, + 0.769 + ], + "angle": 0, + "content": "[49] Xizhou Zhu, Yuwen Xiong, Bo Dai, and Dahua Lin. Deep feature flow for video recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2349-2358, 2017. 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.769 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22808" + } + ] +] \ No newline at end of file diff --git a/2025/HuPerFlow_ A Comprehensive Benchmark for Human vs. Machine Motion Estimation Comparison/5a790d2d-155e-42dc-bc07-eb0a483c9c11_origin.pdf b/2025/HuPerFlow_ A Comprehensive Benchmark for Human vs. Machine Motion Estimation Comparison/5a790d2d-155e-42dc-bc07-eb0a483c9c11_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c03374b9630005aa5c39a1aadb7ee5e68360aa96 --- /dev/null +++ b/2025/HuPerFlow_ A Comprehensive Benchmark for Human vs. Machine Motion Estimation Comparison/5a790d2d-155e-42dc-bc07-eb0a483c9c11_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0bd9b9edd045941c0636b6898d02812e43673421f42da40ce4956996403d04a0 +size 8467343 diff --git a/2025/HuPerFlow_ A Comprehensive Benchmark for Human vs. Machine Motion Estimation Comparison/full.md b/2025/HuPerFlow_ A Comprehensive Benchmark for Human vs. Machine Motion Estimation Comparison/full.md new file mode 100644 index 0000000000000000000000000000000000000000..297ea4acb19ec6b9dcca04f085f25051db7e93b5 --- /dev/null +++ b/2025/HuPerFlow_ A Comprehensive Benchmark for Human vs. Machine Motion Estimation Comparison/full.md @@ -0,0 +1,240 @@ +# HuPerFlow: A Comprehensive Benchmark for Human vs. Machine Motion Estimation Comparison + +Yung-Hao Yang $^{1}$ Zitang Sun $^{1}$ Taiki Fukiage $^{2}$ Shin'ya Nishida $^{1,2*}$ + +$^{1}$ Cognitive Informatics Lab, Graduate School of Informatics, Kyoto University, Japan + $^{2}$ Communication Science Laboratories, Nippon Telegraph and Telephone Corporation, Japan + +https://kucognitiveinformaticslab.github.io/Hyperflow-Website/ + +# Abstract + +As AI models are increasingly integrated into applications involving human interaction, understanding the alignment between human perception and machine vision has become essential. One example is the estimation of visual motion (optical flow) in dynamic applications such as driving assistance. While there are numerous optical flow datasets and benchmarks with ground truth information, human-perceived flow in natural scenes remains underexplored. We introduce HuPerFlow—a benchmark for human-perceived flow, measured at 2,400 locations across ten optical flow datasets, with $\sim 38,400$ response vectors collected through online psychophysical experiments. Our data demonstrate that human-perceived flow aligns with ground truth in spatiotemporally smooth locations while also showing systematic errors influenced by various environmental properties. Additionally, we evaluated several optical flow algorithms against human-perceived flow, uncovering both similarities and unique aspects of human perception in complex natural scenes. HuPerFlow is the first large-scale human-perceived flow benchmark for alignment between computer vision models and human perception, as well as for scientific exploration of human motion perception in natural scenes. The HuPerFlow benchmark is publicly available on the HuPerFlow website. + +# 1. Introduction + +Identifying object motion within the natural environment is crucial for humans, as it supports essential tasks such as navigation, tracking, prediction, and pursuit. In computer vision (CV), optical flow—representing an object's movement in pixel coordinates—serves as a foundational spatiotemporal feature for downstream applications, such as action recognition [36, 39], driver assistance [8, 26], robot navigation [6], video processing [43], etc. With the rise of data-driven approaches, the availability of datasets and + +![](images/9065b6b6390934b41680e1b09a6fa6d387f9241aba765dc42250afa3be6ddcd4.jpg) +Figure 1. Example of HuPerFlow. The red arrows indicate human response to perceived flow, and the green arrows indicate ground truth motion vectors. The circles indicate the magnitudes of endpoint errors, the difference between perceived flow and ground truth. + +benchmarks has consistently fueled advancements in optical flow methods. Numerous studies have leveraged physical engines to generate realistic synthetic data [5] or employed specialized laser and depth cameras [13] to capture ground-truth optical flow in real scenes. + +However, few studies have examined how humans perceive motion in natural scenes. AI models increasingly interact with humans in various application fields, such as image quality assessment [45, 48], video generation [28], and large language model evaluation [29]). Human responses often guide the development of models, which in turn enhances the models' support of human activities. In the context of visual motion perception, estimating human-perceived flow in addition to the ground truth will be use + +ful, for example, in automatic driving [7, 10, 22]. While fully autonomous driving systems can be solely based on the physical ground truth estimation, partially autonomous systems that assist human drivers will be empowered by the ability to predict when, where, and how the drivers are likely to misinterpret motion flow. Another potential use for our dataset is animation synthesis. To control the quality of animation, the animators need to check whether human viewers perceive the movements of the created animation in the way they originally intended. If there are disagreements, the animators should modify the images to minimize the errors. This process can be advanced if human-perceived motion can be automatically predicted. Machine prediction of human-perceived flow is a hard problem for hand-drawn animations since there is no physical ground truth. Even when the animation is rendered from a physically correct dynamic model of the scene, human-perceived motion can be distorted by many factors, including frame rates, image resolution, and the touch of the painting. + +Humans possess a sophisticated visual system capable of estimating object movement in complex natural scenes; however, the underlying mechanisms differ significantly from those used in current CV models. Motion inference is inherently an under-constrained problem (such as the aperture problem [30]), and the solution in humans is a perception that unfolds over a spatiotemporal process spanning multiple frames [1]. In contrast, CV methods typically focus on deterministic dense correspondence between consecutive frames. This difference can sometimes result in entirely different motion perceptions between human vision and CV algorithms, as seen in phenomena like the reverse phi [2] and missing fundamental illusions [4] investigated in vision science and psychology. + +Despite this significant gap, currently, no large-scale benchmark directly evaluates the relationship between CV model responses, physical ground truth, and human perceptual responses to motion in natural scenes, thus hindering further research within this field. To address this problem, we construct the first large-scale comprehensive benchmark for human-perceived optical flow across several types of naturalistic scenes. Our study introduces several unique features compared to prior work: 1): Human-centric labeling: The dataset's labels are directly derived from human perception, gathered through rigorously designed psychophysical experiments. We refer to these human-perceived labels as "Human Perception" and use the term "Ground Truth (GT)" for the physical-correct labels generated by conventional methods (Fig. 1); 2) Large-scale: our dataset includes around 38,400 data points (trials) derived from 480 participant sessions, spanning over 240 hours of collection time. The benchmark comprises ten representative computer vision motion datasets—such as Sintel [5], KITTI [26], Spring [25], VIPER [35], etc—each + +with specifically probed locations and optical flow GT; 3) A novel experimental paradigm: Unlike traditional human assessments (e.g., image quality evaluation), collecting motion perception data is complex and time-intensive. Motion data collection requires a strict display setting, including a constant refresh rate, high resolution, and high-speed hardware to ensure them. Our innovative online experimental paradigm, grounded in a psychophysical method developed for a lab experiment [47], ensures stable and reliable data collection even in online settings, making large-scale human-perceived motion data feasible. + +In addition to constructing this benchmark dataset, we conducted several initial comparative analyses between several representative CV models and human responses, including spatiotemporal gradient-based methods [11], recurrent refinement models [42], multi-frame methods [37], vision transformers [14], human-inspired approaches [38, 40], etc. Our findings reveal several key differences between model predictions and human perception, as well as unique characteristics of human perception in complex natural scenes, including abrupt direction switches (e.g., VIPER [35]), motion blur and non-rigid motion (e.g., MPI Sintel [5]), object translation and rotation (e.g., Flythings3D [24]), camera motion (e.g., Monkaa [24]), and local versus global motion (e.g., MHOF [34]). While the recent CV models correlate closely with human perception, they often exhibit low partial correlation when controlling the ground-truth effects, suggesting their design inherently fails to capture humans' unique perceptual biases. + +In summary, our contributions include: (1) the first large-scale human-perceived optical flow benchmark across ten diverse motion scenarios, (2) a new online psychophysics-based experimental paradigm enabling the large-scale collection of precise human-perceived optical flow in natural scenes, and (3) an evaluation of ten different representative models from both computer vision and human vision fields, revealing significant gaps between human motion interpretation and model predictions. These insights advance computer vision research on human-aligned AI and support the development of technologies for better interaction and assistance with humans. + +# 2. Previous optical flow datasets + +# 2.1. Computer optical flow datasets + +There are several optical flow datasets with physical motion ground truth, and we selected some of them based on motion content, image resolution, sequence count, and frames per sequence for online human observers. KITTI flow 2015 [26] is a real-world driving condition, where ground-truth optical flows were captured by stereo camera, laser scanner, and GPS. Virtual KITTI 2 [12] is a photo-realistic synthetic driving dataset featuring different weather conditions + +and viewing angles. The Driving subset of the Scene Flow Datasets [24] includes eight synthetic sequences (slow / fast $\times$ 15mm / 300mm focal length $\times$ forwards / backward scene). VIPER [35] and TartanAir [44] provide synthetic video sequences with various speed and direction changes, representing diverse navigation scenarios beyond typical driving benchmarks for distinct object motion and camera motion. + +In contrast to automotive navigation scenarios, MPI Sintel [5] and SPRING [25] feature long-range, non-rigid motion with discontinue, large jumps, motion blur, and varying illumination, scene, and material properties. Monkaa, also from Scene Flow Datasets [24], includes significant camera translation and rotation, crucial for revealing motion perception influenced by surrounding objects and background. MHOF [34] captures diverse human motion, featuring both local and global motions across various postures and actions. Flythings3D subset of the Scene Flow Datasets [24], includes multiple geometric shapes and objects translating and rotating along complex 3D trajectories. Since the video presentation required several frames for human observers, we excluded well-known benchmarks such as Middlebury [3] (2-8 frames), FlyingChairs2 [16] (2 frames), ChairsSDHom [15] (2 frames), etc. + +# 2.2. Human motion perception datasets + +Previous studies on human visual motion perception generally used well-controlled artificial visual stimuli, such as sine-wave gratings [21], random-dot kinematograms [20], and biological motion [19], to investigate specific processing mechanisms. Some studies used more complex natural-scenes stimuli to explore neural responses for them [27, 33]. While these studies provided valuable insights into human motion perception, only one study [47], to our knowledge, has directly reported human-perceived vectors for nature scenes, and compared them with ground-truth motion vectors and computer vision model predictions. This psychophysical study revealed that human-perceived optical flow shows various perceptual deviations from the ground-truth flow. Note that human-perceived vectors in dynamical movies, which are mainly based on motion energy within a $\sim 100$ ms temporal window [1], could be significantly different from the correspondence of salient feature points between adjacent frames (cf. [9]). However, [47] examined only five movies, all selected from the SlowFlow version (1008 frames per second) of MPI Sintel benchmark [18], in which the jump sizes between adjacent frames are substantially smaller than those of the standard version used in the current study. It is also worth noting that the movies in [47] were presented through a circular aperture centered on the probed location. + +In the current study, we adopt an approach similar to [47] with significant modifications: (1) using ten optical flow + +benchmarks that are widely used for training and testing in computer vision fields, (2) choosing a larger number of motion clips with dynamic and diverse spatiotemporal properties, and (3) presenting the entire image on the display, as observers would naturally perceive it, rather than through a central aperture. This approach provides a greater motion context and ecological validity. + +# 3. The HuPerFlow dataset + +# 3.1. Methods + +# 3.1.1. Ethical statement + +This study was approved by the Research Ethics Committee of the Graduate School of Informatics at Kyoto University. All participants received online informed consent before the experiments. + +# 3.1.2. Stimuli and Apparatus + +The experiment was conducted using PsychoPy [32], a software for building behavioral experiments, and Pavlovia [31], an online platform for running these experiments in a web browser. To account for diverse hardware environments in online experiments, we controlled the visual presentation in both spatial and temporal domains. Spatially, to standardize viewing angles, observers performed a "credit card calibration," adjusting an on-screen virtual credit card to match a physical one. This allowed us to calculate the physical pixel size after observers adjusted their viewing distance to ensure that 50 pixels corresponded to $1^{\circ}$ of visual angle, maintaining consistent stimulus pixel sizes across devices. Temporally, visual stimuli were presented at $30\mathrm{Hz}$ . Since optimal timing control is crucial for online motion perception experiments, we did not use a higher frame rate for which frame drops were more likely to occur under some users' environments. In addition, We recorded frame drops and onset times during online experiments (see Supplementary Appendix 1 for more details). The frame dropping rate was .041%, and the frame onset error was -0.055 ms $(\mathrm{SD} = 3.664\mathrm{ms})$ , indicating stable and precise visual stimulus presentation across devices. + +This project contained two phases: a preliminary Random Dot Kinematogram (RDK) for training and selection of the observers, and a main session to collect human-perceived flow data using multiple optical flow datasets. In the RDK phase, 5,000 black and white dots moved uniformly within a 600-pixel circular aperture. In the main experiment, we used ten optical flow datasets including KITTI Flow 2015 [26], Virtual KITTI 2 [12], Driving [24], VIPER [35], TartanAir [44], MPI Sintel Flow [5], SPRING [25], Monkaa [24], MHOF [34], and Flythings3D [24]. As most benchmarks include sub-datasets, we present our selection in detail in Tab. 1. + +We applied a standard movie-formating method across datasets with the following exceptions: (1) Most datasets + +![](images/a1c840cf71d7cf5fbc39373e25d995797093501e950a10a26adb35e9a7b09f5c.jpg) +Figure 2. Experimental procedure: At the start of each trial, a green circle appeared to mark the selected area. Next, a motion sequence and a matching stimulus were presented alternately until a response was made. For most optical flow datasets, the motion sequences consisted of 15 frames (500 ms). It was followed by an inter-stimulus interval (ISI) of 1000 ms, then a matching stimulus (500 ms), and a second ISI (1000 ms). A flash probe was shown in the middle frame (the 8th frame) of the motion sequence to indicate the timing and location of the target motion. The green circle location cue, four-dot placeholders, and matching stimulus were presented sequentially at the same location. + +Table 1. Summary of the selected optical flow datasets. * indicates the image size is resized by cubic interpolation. + +
DatasetsResolution (pixels)SequencesTotal FramesSelection
KITTI 2015 [26]1,242 × 3752004,200Training, left image
Virtual KITTI 2 [12]1,242 × 375521,260Clone, camera_0
Driving [24]960 × 54082,400Finalpass, left, slow
VIPER [35]960 × 540*77134,097Val, jpg
TartanAir [44]640 × 480166,982image_left
MPI Sintel [5]1,024 × 436231,064Training, Final
Spring [25]960 × 540*375,000Train_frame_left
Monkaa [24]960 × 540248,664Finalpass, left
MHOF [34]640 × 6401,23895,868Train, masked with segm_EXR
FlyingThings [24]960 × 5402,18021,818Val, image.clean, left
+ +were presented at their original sizes, except for VIPER [35] and SPRING [25], which were resized from $1920 \times 1080$ to $960 \times 540$ for online observers, with the optical flow adjusted accordingly. (2) Frames F-7 to F+7 around a selected testing frame were used as movie sequences, except for MHOF [34] and Flythings3D [24], which were limited to F-4 to F+5 due to each sequence having only ten frames. (3) Virtual KITTI2 [12], SPRING [25] and Scene Flow Datasets [24] (i.e., Driving, Monkaa, and Flythings3D) contain both future (forward) and past (backward) flow directions. However, only future-directed flows were reported as video images were presented sequentially in the forward direction, frame-by-frame. (4) We did not modify the original images, except for MHOF [34] dataset, where frames contained irrelevant background scenes. To standardize the + +background, we used the "segm_EXR" provided by MHOF to render it as a uniform gray background. + +# 3.1.3. Design and Procedure + +In both the RDK and main experiments, each session began with six practice trials followed by 80 formal trials. The practice trials familiarized observers with the procedure and helped them understand the differences between the ground truth and their perceived vectors. To support this, we provided visual feedback of both GT and perceived vectors after each response during practice trials. If the angle between perceived vectors and the GT exceeded $30^{\circ}$ , or if the endpoint error (EPE, the Euclidean distance between the human responses and GT, Equation 1.) was greater than 10 pixels, the same practice trial would be repeated. + +$$ +\mathrm {E P E} = \sqrt {\left(u _ {\text {R e s p}} - u _ {\mathrm {G T}}\right) ^ {2} + \left(v _ {\text {R e s p}} - v _ {\mathrm {G T}}\right) ^ {2}} \tag {1} +$$ + +In the RDK session, the direction $(0^{\circ} - 360^{\circ})$ and speed (1-20 pixels per frame, PPF) of the dots were varied randomly across trials but kept constant within a trial. During formal trials in this session, feedback was also shown to reinforce training. In the main experimental session, we used ten datasets as previously described, selecting 24 sequences per dataset. Each sequence included ten target motion locations with speeds under 20 PPF at a specific frame. At each probed location, we collected 16 repeated trials (four repetitions per observer $\times$ four observers) to obtain a stable measure of perceived flow, totaling 2,400 probed locations across 38,400 trials. In each session, participants viewed two movie sequences with ten target motion locations and four repetitions each (80 trials in total). This yielded 480 sessions across 10 datasets $\times$ 12 sessions (2 movie sequences each) $\times$ 4 observers. No feedback was provided during formal trials in the main experiments. + +Fig. 2 illustrates the display layout and procedure in the main experiment. Both the RDK and main experiments + +used a method of adjustment to measure human-perceived flow at the flash-probed spatiotemporal location. Observers reported the perceived flow of a target motion by matching it with a circular Brownian noise pattern $(1/f^2, 120$ pixels in diameter), which has a spectrum similar to natural images and gives the least motion aliasing for a wide range of motion speeds. Using a mouse cursor (a purple dot, 10 pixels in diameter) within a circular control panel (600 pixels in diameter), observers could freely adjust the direction $(0^\circ - 360^\circ)$ and speed (1-20 pixels per frame) of the matching Brownian noise. The direction was mapped to angles $(\theta)$ in a polar coordinate, and speed was adjusted logarithmically based on the cursor's distance from the center of the panel. This method provides more accurate measurements of perceived motion vectors than subjective reports. The control panel for the matching stimulus was always fixed at the display's center. See Supplementary Video S1 for the animated procedure. + +In the RDK session, where direction and speed were uniform within each trial, both the target motion and matching stimulus were presented at the center of the aperture. In the main experiment, a green circle (120 pixels in diameter), presented before each trial, served as a spatial cue indicating the target motion area. Additionally, four additional dots (green dots, 10 pixels in diameter) served as placeholders throughout the trial. They were positioned 60 pixels above, below, on the left, and on the right of the probed dot. For most datasets, a repetition cycle (see below) consisted of the target movie sequence (500 ms, 15 frames), an interstimulus interval (ISI, 1000 ms), and the matching stimulus (500 ms), followed by a second ISI (1000 ms) before the next cycle. For the MHOF [34] and FlyingThings3D [24] datasets, which contain only ten frames (333 ms), the second ISI was adjusted to 1167 ms to keep the cycle duration consistent. + +In each trial, a flash-probed dot (a green dot, seven pixels in diameter) was presented at a selected testing frame and location to indicate the timing and position of the target motion. The flash probe appeared in the middle of each movie sequence presentation (typically at the $8^{\text{th}}$ frame, and at the $5^{\text{th}}$ frame for MHOF and FlyingThings3D), indicating when observers should respond. The probed target motions were intentionally selected from informative locations, including main objects, the background, and minor objects. In the main experiment, the four-dot placeholders appeared around the probed location, with the matching stimulus displayed directly at that placeholder's location. During each trial, observers could view the movie sequences and match Brownian noise multiple times (repetition cycles), while responses were only accepted after completing three repetition cycles. Since the target motion was shown for just one frame, this adjustment process allowed observers to refine their responses through repeated viewing. + +# 3.2. Results + +# 3.2.1. Data quality control + +The study included 55 observers (mean age = 22.18, SD = 2.82, 24 females and 31 males) who participated in a preliminary Random Dot Kinematogram (RDK) experiment. To ensure reliable perceived flow data, we initially selected observers whose correlation coefficient between perceived vectors and ground truth vectors exceeded .90 in the RDK session. Nine observers with low correlation (individual correlation range: .501 to .885) and four who were unable to complete the experiment were excluded. The group-level correlation for the 42 remaining observers was high (Pearson correlation coefficient $r = .955$ , with individual correlations ranging from .917 to .985), demonstrating their ability to reproduce perceived motion accurately. These 42 observers (mean age = 22.19, SD = 2.46; 17 females, 25 males) then proceeded to the main experimental sessions. They participated in 480 sessions since some observers attended multiple different sessions. + +# 3.2.2. Perceived vectors and "flow illusions" + +In Fig. 3, we show representative examples of averaged human responses for each dataset (see Supplementary Video S2 for target motions and the corresponding averaged human responses across all probed locations within each dataset). Our findings reveal systematic and stable human-specific biases in motion perception under certain scenarios. We termed these systematic discrepancies between human responses and ground truth motions as "flow illusions". These deviations were not random errors but provided valuable insights into the underlying processing and mechanism of human motion perception. On the other hand, we also observed individual differences, with an average inter-observers correlation of 0.407 (SD: 0.417; range from -0.329 to 0.990) across all probed locations. These Individual variations would be useful for training models to reproduce the distribution of human motion perception. To facilitate further research, we have made all raw individual data publicly available on the HuPerFlow website [46]. + +In most driving datasets such as KITTI 2015 [26], Virtual KITTI 2 [12], and Driving [24], the direction and speed of optical flows transition smoothly across frames, with perceived flows aligning closely with the ground truth motion. This consistency supports the reliability of our measurement method in capturing observers' perceptions under spatiotemporally smooth scenes. Notably, as optical flow locations deviated from the scene center, observers often perceived the directions to be slightly shifting toward the central region. This finding may reflect observers' attempts to cancel out the image motion components produced by self-motion (camera motion) during driving. In contrast, other navigation datasets like VIPER [35] and TartanAir [44] show abrupt direction changes, particularly during ac + +![](images/eab63f52866073a542b9286741b20c8687dda5751d44bc7d53cd3cd4590f1702.jpg) +Figure 3. Demonstration of human perceived motion vectors. The results cover ten representative existing optical flow benchmarks. + +tions such as turning, lane changing, or flying backward. These involve interactions between image motion and camera motion, with perceived flows often revealing systematic errors primarily driven by camera motion. + +In datasets like MPI Sintel [5] and SPRING [25], where images include blurred, non-rigid, discontinuous motion and large jump size (cf. [47] using stimuli with small jumps), observers generally failed to discern individual locations, perceiving them as part of a unified grouping movement instead. Unlike the datasets described in the previous paragraph, which were viewed with smoothly moving cameras, Monkaa [24] dataset involved significant camera rotation around relatively stationary objects. Observers often perceived these objects as moving due to the motion of surrounding objects or backgrounds, illustrating the phenomena of relative and induced motion. + +In the MHOF [34] datasets, human figures generally exhibit global body movement accompanied by local motions from the head and limbs. Although the motion structures are less complex than in MPI Sintel and Spring, observers tended to focus on the direction of the global body movement and ignored the local gestures. In the Flythings3D [24], multiple objects float along 3D random trajectories, requiring observers to perform object-ground and object-object segmentation. They were successful in perceiving motion against simple backgrounds and distinct surrounding objects but made more errors in cluttered, messy conditions. + +As different datasets revealed various types of errors, we calculated the correlation between human response and ground truth for each dataset separately. In general, the driving datasets (i.e., KITTI 2015 [26], Virtual KITTI 2 [12], and Driving [24]) showed high correlations, but oth + +ers showed middle to low correlation (see Supplementary Appendix 2 for the scatter plot of each dataset). + +# 3.2.3. EPE in relation to optical flow properties + +As shown in Fig. 3, the EPE of human response (indicated by the diameter of the green circle) varied across different positions and scenes. Following [5], we evaluated how EPE changed in relation to various optical flow properties. Fig. 4 shows that EPE increased as GT speed increased, which suggests that faster motions imposed more significant computational challenges. To assess temporal flow dynamics, we compared optical flow from the previous frame (t-1) to the selected probed frame (t-0) and analyzed the temporal gradient changes in uv-vectors over two intervals: from t-1 to t-0, and from t-0 to t+1 (ground truth), normalized by GT speed from the same location. Generally, we observed that EPE increased as the normalized temporal flow gradient increased up to the ratio of 2, and then leveled off. + +We also considered spatial image properties by using a Sobel kernel to derive the image gradient, calculating the average edge magnitude within a $\pm 40$ -pixel neighborhood around the probed location. The results indicate that EPE decreased as the image gradient increased. Additionally, using instance boundary from the MPI Sintel dataset, we calculated the shortest distance to the boundary and found that EPE also decreased as the distance increased. These findings suggest that edges and boundaries may create ambiguities in motion flow interpretation and interrupt motion perception. Finally, using camera motion data from the MPI Sintel as a proxy for self-motion within a scene, we observed that EPE increased with self-motion, and this effect persisted even when excluding the effect of GT speed (image motion). + +![](images/f6cd9664e467d9370b42dc86d615c6c96676048813b0f30e67546cbb352b968b.jpg) + +![](images/5d280e5ffb9e5666da580f64b8e33b934a326fcd4e133f38097804794bd2512f.jpg) + +![](images/1e574383185e22e89a7e2b88dd5fd65e2d45bb8d91af8b06cdb2967637b392e9.jpg) + +![](images/1f5836719aec56efb93590efc195e6442316d2f7757bb8aa004fc5598871bdb8.jpg) +Figure 4. Endpoint errors (EPEs) of human responses and two machine vision models as functions of optical flow properties. The horizontal axis of each panel is the GT speed, normalized temporal flow gradient, spatial image gradient, instance boundary distance, and self-motion with(out) normalized by the image GT speed. The vertical axis represents EPE with different scales for visualization. + +![](images/e1d79cc765d5d1b6e3ee3fda065d0b267e02bacee21d7edcea4e52926cae09ee.jpg) + +![](images/46216c18208f5bf43c2632f556bad05b9eada102d7a79ed0e37316bb20676104.jpg) + +In addition to the EPE of human response, Fig. 4 shows the EPEs of two machine vision models, VideoFlow and FFV1MT (see Sec. 4). While there were clear differences in the absolute level of performance, the optical flow properties generally affect EPE in similar ways for human and machine vision models, except for the effects of spatial image gradient effect on VideoFlow. + +# 4. Comparison with optical flow algorithms + +We selected a range of representative optical flow algorithms to assess the similarities and differences among human-perceived flows, ground truth motion, and predictions from computer vision models. These included classical methods based on brightness constancy assumptions, such as Farneback [11]; biologically inspired approaches like FFV1MT [38]; multiscale inference models, such as FlowNet2 [15]; recurrent refinement models like RAFT [42]; transformer-based models like FlowFormer [14]; graph reasoning methods such as AGFlow [23]; multiframe approaches like VideoFlow [37]; Motion energy and self-attention based models like V1Attention [40] and DualModel [41]; as well as general-purpose architectures like Perceiver IO [17]. For a fair comparison, we used each model's default configuration or the best pre-trained model fine-tuned on the Sintel dataset. Since unsupervised learning methods approximate the supervised ground truth, we limited our comparison to supervised methods. + +For evaluation, we first calculated the average EPE and Pearson correlation coefficients between model predictions and both human responses and ground truth. To further understand the models' alignment with human perception independently of ground truth influence—since most models were trained to fit ground truth and the ground truth was inherently correlated with human responses—we also calculated the partial correlation between model predictions and human responses, controlling for ground truth, as shown in Equation 2. Here, $r_{x,y}$ denotes correlation coefficients be + +tween $x$ and $y$ . + +$$ +\rho_ {\text {m o d e l}} = r _ {\text {r e s p , m o d e l} \cdot \mathrm {G T}} = \frac {r _ {\text {r e s p , m o d e l}} - r _ {\text {r e s p , G T}} \cdot r _ {\text {m o d e l , G T}}}{\sqrt {1 - r _ {\text {r e s p , G T}} ^ {2}} \cdot \sqrt {1 - r _ {\text {m o d e l , G T}} ^ {2}}}. \tag {2} +$$ + +To complement partial correlation analysis, we also adopted a Response Consistency Index (RCI) from [47]. This index measures the similarity between model predictions and human responses at each probed location, approaching $+1$ when the model's prediction aligns closely with human flow illusions (see Supplementary Appendix 3 for more details). + +Tab. 2 summarizes model predictions across all datasets. Overall, VideoFlow's predictions were the closest to the ground truth. RAFT showed the most closely resembling human responses while also having a strong alignment with the ground truth (second-best among models). When controlling for the ground truth effect, FFV1MT and V1Attention-StageI were the most similar to human flow illusions, as indicated by high partial correlations and average RCI values. A detailed analysis comparing the predictions of optical flow algorithms with both the ground truth and human responses for each dataset is provided in Tables S1-S12 of Supplementary Appendix 4. + +Fig. 5 provides examples of model predictions, highlighting examples where predictions aligned closely with the ground truth (e.g., VideoFlow and RAFT) versus those having some resemblance to human response errors (e.g., FFV1MT and V1Attention-StageI). For instance, at Point (A-I) in Fig. 5, both flow illusions and FFV1MT exhibited temporal-sluggish responses but VideoFlow aligned well with GT. Conversely, point (A-II) in Fig. 5 both model predictions were similar to flow illusions, likely due to image signal loss, as the GT indicated a small, dark stone falling. At Point (B-I) and other selected body locations, both human perception and FFV1MT reflected grouping motion in the same direction, regardless of local gesture directions. + +Table 2. Predictions of optical flow algorithms versus Human response or ground truth (GT). $\rho$ : Partial correlation; $r$ : Pearson correlation coefficient; EPE: vector end-point error; uv, dir, spd represent motion components in Cartesian space, direction, and speed, respectively. RCI is introduced from [47] to represent the model's similarity to human perception (the larger, the more human-aligned). + +
MethodρRCIv.s. Humanv.s. GT
ρuvρdirρspdruvrdirrspdEPEruvrdirrspdEPE
Ground TruthNaNNaNNaNNaN0.660.430.296.961.001.001.000.00
Farnebäck [11]0.170.140.040.130.540.380.156.910.680.650.395.52
FFV1MT [38]0.310.280.060.150.630.450.206.580.700.590.545.84
RAFT [42]0.110.220.010.050.640.470.237.370.930.850.792.19
FlowNet2 [15]0.100.260.000.070.590.480.187.550.820.800.613.24
FlowFormer [14]0.070.19-0.020.040.640.460.217.430.930.870.792.08
VideoFlow [49]0.080.22-0.020.040.640.470.227.470.940.880.811.91
Perceiver IO [17]0.090.170.030.050.620.440.237.750.890.820.742.87
AGFlow [23]0.160.240.010.110.540.450.157.940.680.660.465.44
V1Attention-StageI [40]0.230.280.020.180.550.450.146.360.630.540.427.00
DualModel [41]0.200.250.060.130.570.450.207.310.700.670.535.66
+ +![](images/098a852fcfbaf5c7796842bd6676713224dd1afe568ece9ecb2331b7dcda6e76.jpg) +Figure 5. Predicted vectors of optical flow algorithms. Examples are FFV1MT [38], V1Attention [40], VideoFlow [37], and RAFT [42]. We marked some points with large RCI differences between the models in the left and right columns. Generally, the model on the left side shows a more human-aligned response over ground truth, while the right side shows the opposite result. + +At Point (C-I), a strong self-motion effect occurred toward a static motorcycle. Despite the strong physical optical flow in the image, human observers and V1Attention-StageI appeared to compensate for self-motion to perceive the motorcycle as stationary. For more examples, please visit the HuPerFlow website [46]. Note, however, that there remains + +a large gap between human responses and the most similar model predictions, possibly because the purpose of human visual motion processing is not merely to estimate image flow. The primary purpose of making our dataset is to help future studies develop models that can fill this gap. + +# 5. Conclusion + +By overcoming technical challenges, we developed a novel online paradigm to collect human motion perception data, introducing the first large-scale benchmark for human-perceived optical flow. Our findings identified distinct perceptual illusions across various scene scenarios, some aligning with and others diverging from the predictions of CV optical flow algorithms. Human motion perception, compared to state-of-the-art CV models, has limited spatial, temporal, and intensity resolutions and struggles with motion estimation in retinal coordinates. However, human motion vision has remarkable capabilities in analyzing hierarchical movement structures and distinguishing object motion from self-motion in world coordinates. Our proposed methods can be extended to collect broader human data from diverse datasets, further bridging gaps between CV, human perception, and ground truth. + +While our dataset may not be sufficient to train models from zero, it offers valuable benchmarks for model evaluation and fine-tuning. Physically accurate CV models are crucial, but aligning with human perception is equally important for human-centered applications. Notably, current optical flow models fail to predict various human-perceived motion illusions, highlighting the need to address this gap both for practical applications and for advancing the scientific understanding of human vision. + +# Acknowledgments + +This work was supported in part by JSPS Grants-in-Aid for Scientific Research (KAKENHI), Grant Numbers JP20H00603, JP20H05957, and JP24H00721, and by the Spring Fellowship, Grant Number JPMJFS2123. + +# References + +[1] Edward H Adelson and James R Bergen. Spatiotemporal energy models for the perception of motion. Josa a, 2(2): 284-299, 1985. 2, 3 +[2] Stuart M Anstis and Brian J Rogers. Illusory continuous motion from oscillating positive-negative patterns: Implications for motion perception. Perception, 15(5):627-640, 1986. 2 +[3] Simon Baker, Daniel Scharstein, J. P. Lewis, Stefan Roth, Michael J. Black, and Richard Szeliski. A Database and Evaluation Methodology for Optical Flow. International Journal of Computer Vision, 92(1):1-31, 2011. 3 +[4] Richard O Brown and Sheng He. Visual motion of missing-fundamental patterns: motion energy versus feature correspondence. Vision Research, 40(16):2135-2147, 2000. 2 +[5] Daniel J. Butler, Jonas Wulff, Garrett B. Stanley, and Michael J. Black. A naturalistic open source movie for optical flow evaluation. In European Conf. on Computer Vision (ECCV), pages 611-625. Springer-Verlag, 2012. 1, 2, 3, 4, 6 +[6] Haiyang Chao, Yu Gu, and Marcello Napolitano. A survey of optical flow techniques for robotics navigation applications. Journal of Intelligent & Robotic Systems, 73:361-372, 2014. 1 +[7] Chenyi Chen, Ari Seff, Alain Kornhauser, and Jianxiong Xiao. Deepdriving: Learning affordance for direct perception in autonomous driving. In Proceedings of the IEEE international conference on computer vision, pages 2722-2730, 2015. 2 +[8] Jaechan Cho, Yongchul Jung, Dong-Sun Kim, Seongjoo Lee, and Yunho Jung. Moving object detection based on optical flow estimation and a gaussian mixture model for advanced driver assistance systems. Sensors, 19(14):3217, 2019. 1 +[9] Carl Doersch, Ankush Gupta, Larisa Markeeeva, Adrià Recasens, Lucas Smaira, Yusuf Aytar, João Carreira, Andrew Zisserman, and Yi Yang. Tap-vid: A benchmark for tracking any point in a video. arXiv preprint arXiv:2211.03726, 2022. Published in NeurIPS Datasets and Benchmarks track, 2022. 3 +[10] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 2 +[11] Gunnar Farnebäck. Two-frame motion estimation based on polynomial expansion. In Proceedings of the Scandinavian conference on Image analysis, pages 363-370. Springer, 2003. 2, 7, 8 +[12] Adrien Gaidon, Qiao Wang, Yohann Cabon, and Eleonora Vig. Virtual worlds as proxy for multi-object tracking analysis. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2, 3, 4, 5, 6 + +[13] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 1 +[14] Tzehoong Huang, Zhichao Zhang, Wenjie Zeng, Ming Liu, Chunjing Zhang, Hongsheng Li, and Shijian Lu. Flowformer: A transformer architecture for optical flow. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2, 7, 8 +[15] Eddy Ilg, Nikolaus Mayer, Tonmoy Saikia, Margret Keuper, Alexey Dosovitskiy, and Thomas Brox. Flownet 2.0: Evolution of optical flow estimation with deep networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2462-2470, 2017. 3, 7, 8 +[16] Eddy Ilg, Tonmoy Saikia, Margret Keuper, and Thomas Brox. Occlusions, motion and depth boundaries with a generic network for disparity, optical flow or scene flow estimation. In European Conference on Computer Vision (ECCV), 2018. 3 +[17] Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, et al. Perceiver io: A general architecture for structured inputs & outputs. arXiv preprint arXiv:2107.14795, 2021. 7, 8 +[18] Joel Janai, Fatma Güney, Jonas Wulff, Michael J Black, and Andreas Geiger. Slow flow: Exploiting high-speed cameras for accurate and diverse optical flow reference data. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3597-3606, 2017. 3 +[19] Gunnar Johansson. Visual perception of biological motion and a model for its analysis. Perception & Psychophysics, 14(2):201-211, 1973. 3 +[20] Bela Julesz. Foundations of Cyclopean Perception. University of Chicago Press, 1971. 3 +[21] D. H. Kelly. Fourier components of moving gratings. Behavior Research Methods, 14(6):435-437, 1982. 3 +[22] Monika Lohani, Brennan R Payne, and David L Strayer. A review of psychophysiological measures to assess cognitive states in real-world driving. Frontiers in human neuroscience, 13:57, 2019. 2 +[23] Ao Luo, Fan Yang, Kunming Luo, Xin Li, Haoqiang Fan, and Shuaicheng Liu. Learning optical flow with adaptive graph reasoning. In Proceedings of the AAAI conference on artificial intelligence, pages 1890-1898, 2022. 7, 8 +[24] Nikolaus Mayer, Eddy Ilg, Philip Häusser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In IEEE International Conference on Computer Vision and Pattern Recognition (CVPR), 2016. arXiv:1512.02134. 2, 3, 4, 5, 6 +[25] Lukas Mehl, Jenny Schmalfuss, Azin Jahedi, Yaroslava Naliyayko, and Andres Bruhn. Spring: A high-resolution high-detail dataset and benchmark for scene flow, optical flow and stereo. In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3, 4, 6 +[26] Moritz Menze and Andreas Geiger. Object scene flow for autonomous vehicles. In Proceedings of the IEEE conference + +on computer vision and pattern recognition, pages 3061-3070, 2015. 1, 2, 3, 4, 5, 6 +[27] Shinji Nishimoto and Jack L. Gallant. A three-dimensional spatiotemporal receptive field model explains responses of area mt neurons to naturalistic movies. Journal of Neuroscience, 31(41):14551-14564, 2011. 3 +[28] Mayu Otani, Riku Togashi, Yu Sawai, Ryosuke Ishigami, Yuta Nakashima, Esa Rahtu, Janne Heikkilä, and Shin'ichi Satoh. Toward verifiable and reproducible human evaluation for text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14277-14286, 2023. 1 +[29] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730-27744, 2022. 1 +[30] Christopher C Pack and Richard T Born. Temporal dynamics of a neural solution to the aperture problem in visual area mt of macaque brain. Nature, 409(6823):1040-1042, 2001. 2 +[31] Jonathan Peirce and the PsychoPy team. Pavlovia: Online platform for running experiments. Open Science Tools, 2018. Accessed: 2024-10-25. 3 +[32] Jonathan W. Peirce and Michael MacAskill. Building Experiments in PsychoPy. Sage Publications Ltd, London, United Kingdom, 2018. 3 +[33] Silvia Pitzalis, Carlo Serra, Valerio Sulpizio, Giovanni d'Avossa, and Gaspare Galati. Neural bases of self- and object-motion in a naturalistic vision. Human Brain Mapping, 41(4):1084–1111, 2020. 3 +[34] Anurag Ranjan, David T. Hoffmann, Dimitrios Tzionas, Siyu Tang, Javier Romero, and Michael J. Black. Learning multi-human optical flow. International Journal of Computer Vision (IJCV), 2020. 2, 3, 4, 5, 6 +[35] Stephan R. Richter, Zeeshan Hayder, and Vladlen Koltun. Playing for benchmarks. In IEEE International Conference on Computer Vision, ICCV 2017, Venice, Italy, October 22-29, 2017, pages 2232-2241, 2017. 2, 3, 4, 5 +[36] Laura Sevilla-Lara, Yiyi Liao, Fatma Güney, Varun Jampani, Andreas Geiger, and Michael J Black. On the integration of optical flow and action recognition. In Pattern Recognition: 40th German Conference, GCPR 2018, Stuttgart, Germany, October 9-12, 2018, Proceedings 40, pages 281-297. Springer, 2019. 1 +[37] Xiaoyu Shi, Zhaoyang Huang, Weikang Bian, Dasong Li, Manyuan Zhang, Ka Chun Cheung, Simon See, Hongwei Qin, Jifeng Dai, and Hongsheng Li. Videoflow: Exploiting temporal cues for multi-frame optical flow estimation. arXiv preprint arXiv:2303.08340, 2023. 2, 7, 8 +[38] Fabio Solari, Manuela Chessa, Narasimha K. Medathati, and Pierre Kornprobst. What can we expect from a v1-mt feedforward architecture for optical flow estimation? Signal Processing: Image Communication, 39:342-354, 2015. 2, 7, 8 +[39] Shuyang Sun, Zhanghui Kuang, Lu Sheng, Wanli Ouyang, and Wei Zhang. Optical flow guided feature: A fast and robust motion representation for video action recognition. In + +Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1390-1399, 2018. 1 +[40] Zitang Sun, Yen-Ju Chen, Yung-Hao Yang, and Shin'ya Nishida. Modeling human visual motion processing with trainable motion energy sensing and a self-attention network. In Advances in Neural Information Processing Systems, 2023. 2, 7, 8 +[41] Zitang Sun, Yen-Ju Chen, Yung-Hao Yang, and Shin'ya Nishida. Acquisition of second-order motion perception by learning to recognize the motion of objects made by nondiffusive materials. Journal of Vision, Vision Sciences Society Annual Meeting Abstract, 24(9), 2024. 7, 8 +[42] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Proceedings of the European Conference on Computer Vision (ECCV), pages 402-419, 2020. 2, 7, 8 +[43] Longguang Wang, Yulan Guo, Li Liu, Zaiping Lin, Xinpu Deng, and Wei An. Deep video super-resolution using hr optical flow estimation. IEEE Transactions on Image Processing, 29:4323-4336, 2020. 1 +[44] Wenshan Wang, Delong Zhu, Xiangwei Wang, Yaoyu Hu, Yuheng Qiu, Chen Wang, Yafei Hu, Ashish Kapoor, and Sebastian Scherer. Tartanair: A dataset to push the limits of visual slam. arXiv preprint arXiv:2003.14338, 2020. 3, 4, 5 +[45] H. Wu, Y. Liu, Z. Zhang, L. Du, L. Wang, Y. Wang, D. Lin, and L. Lin. Q-Instruct: Improving Low-Level Visual Abilities for Multi-Modality Foundation Models. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 25490–25500, Seattle, WA, USA, 2024. IEEE. 1 +[46] Yung-Hao Yang, Zitang Sun, Taiki Fukiage, and Shin'ya Nishida. HuPerFlow: A comprehensive benchmark for human vs. machine motion estimation comparison. https://kucognitiveinformaticslab.github.io/Huperflow-Website/. Accessed: March 2025. 5, 8 +[47] Yung-Hao Yang, Taiki Fukiage, Zitang Sun, and Shin'ya Nishida. Psychophysical measurement of perceived motion flow of naturalistic scenes. iScience, 26(12), 2023. 2, 3, 6, 7, 8 +[48] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 1 +[49] Xizhou Zhu, Yuwen Xiong, Bo Dai, and Dahua Lin. Deep feature flow for video recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2349-2358, 2017. 8 \ No newline at end of file diff --git a/2025/HuPerFlow_ A Comprehensive Benchmark for Human vs. Machine Motion Estimation Comparison/images.zip b/2025/HuPerFlow_ A Comprehensive Benchmark for Human vs. Machine Motion Estimation Comparison/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..789ac2006763636e016fe93e766430eaaceb1b9a --- /dev/null +++ b/2025/HuPerFlow_ A Comprehensive Benchmark for Human vs. Machine Motion Estimation Comparison/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:55a79e3c0270f2c84b865d2229ce2f58d03b41b1b105a7cbcae1e14feb59f093 +size 526924 diff --git a/2025/HuPerFlow_ A Comprehensive Benchmark for Human vs. Machine Motion Estimation Comparison/layout.json b/2025/HuPerFlow_ A Comprehensive Benchmark for Human vs. Machine Motion Estimation Comparison/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6976506595ed537d327259c863050ba73b8ba562 --- /dev/null +++ b/2025/HuPerFlow_ A Comprehensive Benchmark for Human vs. Machine Motion Estimation Comparison/layout.json @@ -0,0 +1,6108 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 73, + 102, + 538, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 102, + 538, + 140 + ], + "spans": [ + { + "bbox": [ + 73, + 102, + 538, + 140 + ], + "type": "text", + "content": "HuPerFlow: A Comprehensive Benchmark for Human vs. Machine Motion Estimation Comparison" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 129, + 160, + 479, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 160, + 479, + 175 + ], + "spans": [ + { + "bbox": [ + 129, + 160, + 479, + 175 + ], + "type": "text", + "content": "Yung-Hao Yang" + }, + { + "bbox": [ + 129, + 160, + 479, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 129, + 160, + 479, + 175 + ], + "type": "text", + "content": " Zitang Sun" + }, + { + "bbox": [ + 129, + 160, + 479, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 129, + 160, + 479, + 175 + ], + "type": "text", + "content": " Taiki Fukiage" + }, + { + "bbox": [ + 129, + 160, + 479, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 129, + 160, + 479, + 175 + ], + "type": "text", + "content": " Shin'ya Nishida" + }, + { + "bbox": [ + 129, + 160, + 479, + 175 + ], + "type": "inline_equation", + "content": "^{1,2*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 82, + 175, + 528, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 175, + 528, + 203 + ], + "spans": [ + { + "bbox": [ + 82, + 175, + 528, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 82, + 175, + 528, + 203 + ], + "type": "text", + "content": "Cognitive Informatics Lab, Graduate School of Informatics, Kyoto University, Japan \n" + }, + { + "bbox": [ + 82, + 175, + 528, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 82, + 175, + 528, + 203 + ], + "type": "text", + "content": "Communication Science Laboratories, Nippon Telegraph and Telephone Corporation, Japan" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 205, + 471, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 205, + 471, + 217 + ], + "spans": [ + { + "bbox": [ + 134, + 205, + 471, + 217 + ], + "type": "text", + "content": "https://kucognitiveinformaticslab.github.io/Hyperflow-Website/" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "spans": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 271, + 296, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 271, + 296, + 558 + ], + "spans": [ + { + "bbox": [ + 54, + 271, + 296, + 558 + ], + "type": "text", + "content": "As AI models are increasingly integrated into applications involving human interaction, understanding the alignment between human perception and machine vision has become essential. One example is the estimation of visual motion (optical flow) in dynamic applications such as driving assistance. While there are numerous optical flow datasets and benchmarks with ground truth information, human-perceived flow in natural scenes remains underexplored. We introduce HuPerFlow—a benchmark for human-perceived flow, measured at 2,400 locations across ten optical flow datasets, with " + }, + { + "bbox": [ + 54, + 271, + 296, + 558 + ], + "type": "inline_equation", + "content": "\\sim 38,400" + }, + { + "bbox": [ + 54, + 271, + 296, + 558 + ], + "type": "text", + "content": " response vectors collected through online psychophysical experiments. Our data demonstrate that human-perceived flow aligns with ground truth in spatiotemporally smooth locations while also showing systematic errors influenced by various environmental properties. Additionally, we evaluated several optical flow algorithms against human-perceived flow, uncovering both similarities and unique aspects of human perception in complex natural scenes. HuPerFlow is the first large-scale human-perceived flow benchmark for alignment between computer vision models and human perception, as well as for scientific exploration of human motion perception in natural scenes. The HuPerFlow benchmark is publicly available on the HuPerFlow website." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 584, + 135, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 584, + 135, + 596 + ], + "spans": [ + { + "bbox": [ + 55, + 584, + 135, + 596 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": "Identifying object motion within the natural environment is crucial for humans, as it supports essential tasks such as navigation, tracking, prediction, and pursuit. In computer vision (CV), optical flow—representing an object's movement in pixel coordinates—serves as a foundational spatiotemporal feature for downstream applications, such as action recognition [36, 39], driver assistance [8, 26], robot navigation [6], video processing [43], etc. With the rise of data-driven approaches, the availability of datasets and" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 245, + 554, + 458 + ], + "blocks": [ + { + "bbox": [ + 317, + 245, + 554, + 458 + ], + "lines": [ + { + "bbox": [ + 317, + 245, + 554, + 458 + ], + "spans": [ + { + "bbox": [ + 317, + 245, + 554, + 458 + ], + "type": "image", + "image_path": "9065b6b6390934b41680e1b09a6fa6d387f9241aba765dc42250afa3be6ddcd4.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 468, + 555, + 522 + ], + "lines": [ + { + "bbox": [ + 313, + 468, + 555, + 522 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 555, + 522 + ], + "type": "text", + "content": "Figure 1. Example of HuPerFlow. The red arrows indicate human response to perceived flow, and the green arrows indicate ground truth motion vectors. The circles indicate the magnitudes of endpoint errors, the difference between perceived flow and ground truth." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 542, + 554, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 542, + 554, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 542, + 554, + 601 + ], + "type": "text", + "content": "benchmarks has consistently fueled advancements in optical flow methods. Numerous studies have leveraged physical engines to generate realistic synthetic data [5] or employed specialized laser and depth cameras [13] to capture ground-truth optical flow in real scenes." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 605, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 605, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 605, + 555, + 715 + ], + "type": "text", + "content": "However, few studies have examined how humans perceive motion in natural scenes. AI models increasingly interact with humans in various application fields, such as image quality assessment [45, 48], video generation [28], and large language model evaluation [29]). Human responses often guide the development of models, which in turn enhances the models' support of human activities. In the context of visual motion perception, estimating human-perceived flow in addition to the ground truth will be use" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22799" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 300 + ], + "type": "text", + "content": "ful, for example, in automatic driving [7, 10, 22]. While fully autonomous driving systems can be solely based on the physical ground truth estimation, partially autonomous systems that assist human drivers will be empowered by the ability to predict when, where, and how the drivers are likely to misinterpret motion flow. Another potential use for our dataset is animation synthesis. To control the quality of animation, the animators need to check whether human viewers perceive the movements of the created animation in the way they originally intended. If there are disagreements, the animators should modify the images to minimize the errors. This process can be advanced if human-perceived motion can be automatically predicted. Machine prediction of human-perceived flow is a hard problem for hand-drawn animations since there is no physical ground truth. Even when the animation is rendered from a physically correct dynamic model of the scene, human-perceived motion can be distorted by many factors, including frame rates, image resolution, and the touch of the painting." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 303, + 294, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 303, + 294, + 471 + ], + "spans": [ + { + "bbox": [ + 55, + 303, + 294, + 471 + ], + "type": "text", + "content": "Humans possess a sophisticated visual system capable of estimating object movement in complex natural scenes; however, the underlying mechanisms differ significantly from those used in current CV models. Motion inference is inherently an under-constrained problem (such as the aperture problem [30]), and the solution in humans is a perception that unfolds over a spatiotemporal process spanning multiple frames [1]. In contrast, CV methods typically focus on deterministic dense correspondence between consecutive frames. This difference can sometimes result in entirely different motion perceptions between human vision and CV algorithms, as seen in phenomena like the reverse phi [2] and missing fundamental illusions [4] investigated in vision science and psychology." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 474, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 295, + 713 + ], + "type": "text", + "content": "Despite this significant gap, currently, no large-scale benchmark directly evaluates the relationship between CV model responses, physical ground truth, and human perceptual responses to motion in natural scenes, thus hindering further research within this field. To address this problem, we construct the first large-scale comprehensive benchmark for human-perceived optical flow across several types of naturalistic scenes. Our study introduces several unique features compared to prior work: 1): Human-centric labeling: The dataset's labels are directly derived from human perception, gathered through rigorously designed psychophysical experiments. We refer to these human-perceived labels as \"Human Perception\" and use the term \"Ground Truth (GT)\" for the physical-correct labels generated by conventional methods (Fig. 1); 2) Large-scale: our dataset includes around 38,400 data points (trials) derived from 480 participant sessions, spanning over 240 hours of collection time. The benchmark comprises ten representative computer vision motion datasets—such as Sintel [5], KITTI [26], Spring [25], VIPER [35], etc—each" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 553, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 203 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 203 + ], + "type": "text", + "content": "with specifically probed locations and optical flow GT; 3) A novel experimental paradigm: Unlike traditional human assessments (e.g., image quality evaluation), collecting motion perception data is complex and time-intensive. Motion data collection requires a strict display setting, including a constant refresh rate, high resolution, and high-speed hardware to ensure them. Our innovative online experimental paradigm, grounded in a psychophysical method developed for a lab experiment [47], ensures stable and reliable data collection even in online settings, making large-scale human-perceived motion data feasible." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 205, + 553, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 553, + 419 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 553, + 419 + ], + "type": "text", + "content": "In addition to constructing this benchmark dataset, we conducted several initial comparative analyses between several representative CV models and human responses, including spatiotemporal gradient-based methods [11], recurrent refinement models [42], multi-frame methods [37], vision transformers [14], human-inspired approaches [38, 40], etc. Our findings reveal several key differences between model predictions and human perception, as well as unique characteristics of human perception in complex natural scenes, including abrupt direction switches (e.g., VIPER [35]), motion blur and non-rigid motion (e.g., MPI Sintel [5]), object translation and rotation (e.g., Flythings3D [24]), camera motion (e.g., Monkaa [24]), and local versus global motion (e.g., MHOF [34]). While the recent CV models correlate closely with human perception, they often exhibit low partial correlation when controlling the ground-truth effects, suggesting their design inherently fails to capture humans' unique perceptual biases." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 422, + 553, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 553, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 553, + 564 + ], + "type": "text", + "content": "In summary, our contributions include: (1) the first large-scale human-perceived optical flow benchmark across ten diverse motion scenarios, (2) a new online psychophysics-based experimental paradigm enabling the large-scale collection of precise human-perceived optical flow in natural scenes, and (3) an evaluation of ten different representative models from both computer vision and human vision fields, revealing significant gaps between human motion interpretation and model predictions. These insights advance computer vision research on human-aligned AI and support the development of technologies for better interaction and assistance with humans." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 578, + 479, + 592 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 578, + 479, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 578, + 479, + 592 + ], + "type": "text", + "content": "2. Previous optical flow datasets" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 599, + 482, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 599, + 482, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 599, + 482, + 612 + ], + "type": "text", + "content": "2.1. Computer optical flow datasets" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "type": "text", + "content": "There are several optical flow datasets with physical motion ground truth, and we selected some of them based on motion content, image resolution, sequence count, and frames per sequence for online human observers. KITTI flow 2015 [26] is a real-world driving condition, where ground-truth optical flows were captured by stereo camera, laser scanner, and GPS. Virtual KITTI 2 [12] is a photo-realistic synthetic driving dataset featuring different weather conditions" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22800" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 294, + 167 + ], + "type": "text", + "content": "and viewing angles. The Driving subset of the Scene Flow Datasets [24] includes eight synthetic sequences (slow / fast " + }, + { + "bbox": [ + 54, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 54, + 72, + 294, + 167 + ], + "type": "text", + "content": " 15mm / 300mm focal length " + }, + { + "bbox": [ + 54, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 54, + 72, + 294, + 167 + ], + "type": "text", + "content": " forwards / backward scene). VIPER [35] and TartanAir [44] provide synthetic video sequences with various speed and direction changes, representing diverse navigation scenarios beyond typical driving benchmarks for distinct object motion and camera motion." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 169, + 297, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 169, + 297, + 360 + ], + "spans": [ + { + "bbox": [ + 57, + 169, + 297, + 360 + ], + "type": "text", + "content": "In contrast to automotive navigation scenarios, MPI Sintel [5] and SPRING [25] feature long-range, non-rigid motion with discontinue, large jumps, motion blur, and varying illumination, scene, and material properties. Monkaa, also from Scene Flow Datasets [24], includes significant camera translation and rotation, crucial for revealing motion perception influenced by surrounding objects and background. MHOF [34] captures diverse human motion, featuring both local and global motions across various postures and actions. Flythings3D subset of the Scene Flow Datasets [24], includes multiple geometric shapes and objects translating and rotating along complex 3D trajectories. Since the video presentation required several frames for human observers, we excluded well-known benchmarks such as Middlebury [3] (2-8 frames), FlyingChairs2 [16] (2 frames), ChairsSDHom [15] (2 frames), etc." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 371, + 242, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 371, + 242, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 371, + 242, + 384 + ], + "type": "text", + "content": "2.2. Human motion perception datasets" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 390, + 295, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 390, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 56, + 390, + 295, + 688 + ], + "type": "text", + "content": "Previous studies on human visual motion perception generally used well-controlled artificial visual stimuli, such as sine-wave gratings [21], random-dot kinematograms [20], and biological motion [19], to investigate specific processing mechanisms. Some studies used more complex natural-scenes stimuli to explore neural responses for them [27, 33]. While these studies provided valuable insights into human motion perception, only one study [47], to our knowledge, has directly reported human-perceived vectors for nature scenes, and compared them with ground-truth motion vectors and computer vision model predictions. This psychophysical study revealed that human-perceived optical flow shows various perceptual deviations from the ground-truth flow. Note that human-perceived vectors in dynamical movies, which are mainly based on motion energy within a " + }, + { + "bbox": [ + 56, + 390, + 295, + 688 + ], + "type": "inline_equation", + "content": "\\sim 100" + }, + { + "bbox": [ + 56, + 390, + 295, + 688 + ], + "type": "text", + "content": " ms temporal window [1], could be significantly different from the correspondence of salient feature points between adjacent frames (cf. [9]). However, [47] examined only five movies, all selected from the SlowFlow version (1008 frames per second) of MPI Sintel benchmark [18], in which the jump sizes between adjacent frames are substantially smaller than those of the standard version used in the current study. It is also worth noting that the movies in [47] were presented through a circular aperture centered on the probed location." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 689, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 294, + 713 + ], + "type": "text", + "content": "In the current study, we adopt an approach similar to [47] with significant modifications: (1) using ten optical flow" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "benchmarks that are widely used for training and testing in computer vision fields, (2) choosing a larger number of motion clips with dynamic and diverse spatiotemporal properties, and (3) presenting the entire image on the display, as observers would naturally perceive it, rather than through a central aperture. This approach provides a greater motion context and ecological validity." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 167, + 451, + 179 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 167, + 451, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 167, + 451, + 179 + ], + "type": "text", + "content": "3. The HuPerFlow dataset" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 186, + 378, + 198 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 186, + 378, + 198 + ], + "spans": [ + { + "bbox": [ + 313, + 186, + 378, + 198 + ], + "type": "text", + "content": "3.1. Methods" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 204, + 417, + 215 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 204, + 417, + 215 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 417, + 215 + ], + "type": "text", + "content": "3.1.1. Ethical statement" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 219, + 553, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 219, + 553, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 219, + 553, + 266 + ], + "type": "text", + "content": "This study was approved by the Research Ethics Committee of the Graduate School of Informatics at Kyoto University. All participants received online informed consent before the experiments." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 268, + 443, + 280 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 268, + 443, + 280 + ], + "spans": [ + { + "bbox": [ + 313, + 268, + 443, + 280 + ], + "type": "text", + "content": "3.1.2. Stimuli and Apparatus" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 283, + 555, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 283, + 555, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 283, + 555, + 545 + ], + "type": "text", + "content": "The experiment was conducted using PsychoPy [32], a software for building behavioral experiments, and Pavlovia [31], an online platform for running these experiments in a web browser. To account for diverse hardware environments in online experiments, we controlled the visual presentation in both spatial and temporal domains. Spatially, to standardize viewing angles, observers performed a \"credit card calibration,\" adjusting an on-screen virtual credit card to match a physical one. This allowed us to calculate the physical pixel size after observers adjusted their viewing distance to ensure that 50 pixels corresponded to " + }, + { + "bbox": [ + 313, + 283, + 555, + 545 + ], + "type": "inline_equation", + "content": "1^{\\circ}" + }, + { + "bbox": [ + 313, + 283, + 555, + 545 + ], + "type": "text", + "content": " of visual angle, maintaining consistent stimulus pixel sizes across devices. Temporally, visual stimuli were presented at " + }, + { + "bbox": [ + 313, + 283, + 555, + 545 + ], + "type": "inline_equation", + "content": "30\\mathrm{Hz}" + }, + { + "bbox": [ + 313, + 283, + 555, + 545 + ], + "type": "text", + "content": ". Since optimal timing control is crucial for online motion perception experiments, we did not use a higher frame rate for which frame drops were more likely to occur under some users' environments. In addition, We recorded frame drops and onset times during online experiments (see Supplementary Appendix 1 for more details). The frame dropping rate was .041%, and the frame onset error was -0.055 ms " + }, + { + "bbox": [ + 313, + 283, + 555, + 545 + ], + "type": "inline_equation", + "content": "(\\mathrm{SD} = 3.664\\mathrm{ms})" + }, + { + "bbox": [ + 313, + 283, + 555, + 545 + ], + "type": "text", + "content": ", indicating stable and precise visual stimulus presentation across devices." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 546, + 553, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 553, + 689 + ], + "type": "text", + "content": "This project contained two phases: a preliminary Random Dot Kinematogram (RDK) for training and selection of the observers, and a main session to collect human-perceived flow data using multiple optical flow datasets. In the RDK phase, 5,000 black and white dots moved uniformly within a 600-pixel circular aperture. In the main experiment, we used ten optical flow datasets including KITTI Flow 2015 [26], Virtual KITTI 2 [12], Driving [24], VIPER [35], TartanAir [44], MPI Sintel Flow [5], SPRING [25], Monkaa [24], MHOF [34], and Flythings3D [24]. As most benchmarks include sub-datasets, we present our selection in detail in Tab. 1." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "We applied a standard movie-formating method across datasets with the following exceptions: (1) Most datasets" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22801" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 130, + 47, + 479, + 190 + ], + "blocks": [ + { + "bbox": [ + 130, + 47, + 479, + 190 + ], + "lines": [ + { + "bbox": [ + 130, + 47, + 479, + 190 + ], + "spans": [ + { + "bbox": [ + 130, + 47, + 479, + 190 + ], + "type": "image", + "image_path": "a1c840cf71d7cf5fbc39373e25d995797093501e950a10a26adb35e9a7b09f5c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 201, + 555, + 266 + ], + "lines": [ + { + "bbox": [ + 54, + 201, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 54, + 201, + 555, + 266 + ], + "type": "text", + "content": "Figure 2. Experimental procedure: At the start of each trial, a green circle appeared to mark the selected area. Next, a motion sequence and a matching stimulus were presented alternately until a response was made. For most optical flow datasets, the motion sequences consisted of 15 frames (500 ms). It was followed by an inter-stimulus interval (ISI) of 1000 ms, then a matching stimulus (500 ms), and a second ISI (1000 ms). A flash probe was shown in the middle frame (the 8th frame) of the motion sequence to indicate the timing and location of the target motion. The green circle location cue, four-dot placeholders, and matching stimulus were presented sequentially at the same location." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 312, + 305, + 514 + ], + "blocks": [ + { + "bbox": [ + 55, + 281, + 295, + 304 + ], + "lines": [ + { + "bbox": [ + 55, + 281, + 295, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 295, + 304 + ], + "type": "text", + "content": "Table 1. Summary of the selected optical flow datasets. * indicates the image size is resized by cubic interpolation." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 312, + 305, + 514 + ], + "lines": [ + { + "bbox": [ + 56, + 312, + 305, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 312, + 305, + 514 + ], + "type": "table", + "html": "
DatasetsResolution (pixels)SequencesTotal FramesSelection
KITTI 2015 [26]1,242 × 3752004,200Training, left image
Virtual KITTI 2 [12]1,242 × 375521,260Clone, camera_0
Driving [24]960 × 54082,400Finalpass, left, slow
VIPER [35]960 × 540*77134,097Val, jpg
TartanAir [44]640 × 480166,982image_left
MPI Sintel [5]1,024 × 436231,064Training, Final
Spring [25]960 × 540*375,000Train_frame_left
Monkaa [24]960 × 540248,664Finalpass, left
MHOF [34]640 × 6401,23895,868Train, masked with segm_EXR
FlyingThings [24]960 × 5402,18021,818Val, image.clean, left
", + "image_path": "27a2f7563a2498246b2357b8b1f41306632acc6371fc6256cce699a432c54675.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "content": "were presented at their original sizes, except for VIPER [35] and SPRING [25], which were resized from " + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "inline_equation", + "content": "1920 \\times 1080" + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "inline_equation", + "content": "960 \\times 540" + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "content": " for online observers, with the optical flow adjusted accordingly. (2) Frames F-7 to F+7 around a selected testing frame were used as movie sequences, except for MHOF [34] and Flythings3D [24], which were limited to F-4 to F+5 due to each sequence having only ten frames. (3) Virtual KITTI2 [12], SPRING [25] and Scene Flow Datasets [24] (i.e., Driving, Monkaa, and Flythings3D) contain both future (forward) and past (backward) flow directions. However, only future-directed flows were reported as video images were presented sequentially in the forward direction, frame-by-frame. (4) We did not modify the original images, except for MHOF [34] dataset, where frames contained irrelevant background scenes. To standardize the" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 283, + 555, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 283, + 555, + 308 + ], + "spans": [ + { + "bbox": [ + 313, + 283, + 555, + 308 + ], + "type": "text", + "content": "background, we used the \"segm_EXR\" provided by MHOF to render it as a uniform gray background." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 312, + 437, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 437, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 437, + 323 + ], + "type": "text", + "content": "3.1.3. Design and Procedure" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 326, + 555, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 326, + 555, + 458 + ], + "spans": [ + { + "bbox": [ + 313, + 326, + 555, + 458 + ], + "type": "text", + "content": "In both the RDK and main experiments, each session began with six practice trials followed by 80 formal trials. The practice trials familiarized observers with the procedure and helped them understand the differences between the ground truth and their perceived vectors. To support this, we provided visual feedback of both GT and perceived vectors after each response during practice trials. If the angle between perceived vectors and the GT exceeded " + }, + { + "bbox": [ + 313, + 326, + 555, + 458 + ], + "type": "inline_equation", + "content": "30^{\\circ}" + }, + { + "bbox": [ + 313, + 326, + 555, + 458 + ], + "type": "text", + "content": ", or if the endpoint error (EPE, the Euclidean distance between the human responses and GT, Equation 1.) was greater than 10 pixels, the same practice trial would be repeated." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 346, + 464, + 553, + 483 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 464, + 553, + 483 + ], + "spans": [ + { + "bbox": [ + 346, + 464, + 553, + 483 + ], + "type": "interline_equation", + "content": "\\mathrm {E P E} = \\sqrt {\\left(u _ {\\text {R e s p}} - u _ {\\mathrm {G T}}\\right) ^ {2} + \\left(v _ {\\text {R e s p}} - v _ {\\mathrm {G T}}\\right) ^ {2}} \\tag {1}", + "image_path": "7d3761c493bcb66a9f9ca7c69cd9c1f1d24df21562919f6553ae4e6a9496fc2b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 486, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 555, + 689 + ], + "type": "text", + "content": "In the RDK session, the direction " + }, + { + "bbox": [ + 313, + 486, + 555, + 689 + ], + "type": "inline_equation", + "content": "(0^{\\circ} - 360^{\\circ})" + }, + { + "bbox": [ + 313, + 486, + 555, + 689 + ], + "type": "text", + "content": " and speed (1-20 pixels per frame, PPF) of the dots were varied randomly across trials but kept constant within a trial. During formal trials in this session, feedback was also shown to reinforce training. In the main experimental session, we used ten datasets as previously described, selecting 24 sequences per dataset. Each sequence included ten target motion locations with speeds under 20 PPF at a specific frame. At each probed location, we collected 16 repeated trials (four repetitions per observer " + }, + { + "bbox": [ + 313, + 486, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 313, + 486, + 555, + 689 + ], + "type": "text", + "content": " four observers) to obtain a stable measure of perceived flow, totaling 2,400 probed locations across 38,400 trials. In each session, participants viewed two movie sequences with ten target motion locations and four repetitions each (80 trials in total). This yielded 480 sessions across 10 datasets " + }, + { + "bbox": [ + 313, + 486, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 313, + 486, + 555, + 689 + ], + "type": "text", + "content": " 12 sessions (2 movie sequences each) " + }, + { + "bbox": [ + 313, + 486, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 313, + 486, + 555, + 689 + ], + "type": "text", + "content": " 4 observers. No feedback was provided during formal trials in the main experiments." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "Fig. 2 illustrates the display layout and procedure in the main experiment. Both the RDK and main experiments" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22802" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "type": "text", + "content": "used a method of adjustment to measure human-perceived flow at the flash-probed spatiotemporal location. Observers reported the perceived flow of a target motion by matching it with a circular Brownian noise pattern " + }, + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "type": "inline_equation", + "content": "(1/f^2, 120" + }, + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "type": "text", + "content": " pixels in diameter), which has a spectrum similar to natural images and gives the least motion aliasing for a wide range of motion speeds. Using a mouse cursor (a purple dot, 10 pixels in diameter) within a circular control panel (600 pixels in diameter), observers could freely adjust the direction " + }, + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "type": "inline_equation", + "content": "(0^\\circ - 360^\\circ)" + }, + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "type": "text", + "content": " and speed (1-20 pixels per frame) of the matching Brownian noise. The direction was mapped to angles " + }, + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "type": "inline_equation", + "content": "(\\theta)" + }, + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "type": "text", + "content": " in a polar coordinate, and speed was adjusted logarithmically based on the cursor's distance from the center of the panel. This method provides more accurate measurements of perceived motion vectors than subjective reports. The control panel for the matching stimulus was always fixed at the display's center. See Supplementary Video S1 for the animated procedure." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 297, + 295, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 297, + 295, + 489 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 295, + 489 + ], + "type": "text", + "content": "In the RDK session, where direction and speed were uniform within each trial, both the target motion and matching stimulus were presented at the center of the aperture. In the main experiment, a green circle (120 pixels in diameter), presented before each trial, served as a spatial cue indicating the target motion area. Additionally, four additional dots (green dots, 10 pixels in diameter) served as placeholders throughout the trial. They were positioned 60 pixels above, below, on the left, and on the right of the probed dot. For most datasets, a repetition cycle (see below) consisted of the target movie sequence (500 ms, 15 frames), an interstimulus interval (ISI, 1000 ms), and the matching stimulus (500 ms), followed by a second ISI (1000 ms) before the next cycle. For the MHOF [34] and FlyingThings3D [24] datasets, which contain only ten frames (333 ms), the second ISI was adjusted to 1167 ms to keep the cycle duration consistent." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "type": "text", + "content": "In each trial, a flash-probed dot (a green dot, seven pixels in diameter) was presented at a selected testing frame and location to indicate the timing and position of the target motion. The flash probe appeared in the middle of each movie sequence presentation (typically at the " + }, + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "type": "inline_equation", + "content": "8^{\\text{th}}" + }, + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "type": "text", + "content": " frame, and at the " + }, + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "type": "inline_equation", + "content": "5^{\\text{th}}" + }, + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "type": "text", + "content": " frame for MHOF and FlyingThings3D), indicating when observers should respond. The probed target motions were intentionally selected from informative locations, including main objects, the background, and minor objects. In the main experiment, the four-dot placeholders appeared around the probed location, with the matching stimulus displayed directly at that placeholder's location. During each trial, observers could view the movie sequences and match Brownian noise multiple times (repetition cycles), while responses were only accepted after completing three repetition cycles. Since the target motion was shown for just one frame, this adjustment process allowed observers to refine their responses through repeated viewing." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 314, + 72, + 372, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 372, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 372, + 83 + ], + "type": "text", + "content": "3.2. Results" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 314, + 90, + 429, + 102 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 90, + 429, + 102 + ], + "spans": [ + { + "bbox": [ + 314, + 90, + 429, + 102 + ], + "type": "text", + "content": "3.2.1. Data quality control" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 105, + 553, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 105, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 313, + 105, + 553, + 308 + ], + "type": "text", + "content": "The study included 55 observers (mean age = 22.18, SD = 2.82, 24 females and 31 males) who participated in a preliminary Random Dot Kinematogram (RDK) experiment. To ensure reliable perceived flow data, we initially selected observers whose correlation coefficient between perceived vectors and ground truth vectors exceeded .90 in the RDK session. Nine observers with low correlation (individual correlation range: .501 to .885) and four who were unable to complete the experiment were excluded. The group-level correlation for the 42 remaining observers was high (Pearson correlation coefficient " + }, + { + "bbox": [ + 313, + 105, + 553, + 308 + ], + "type": "inline_equation", + "content": "r = .955" + }, + { + "bbox": [ + 313, + 105, + 553, + 308 + ], + "type": "text", + "content": ", with individual correlations ranging from .917 to .985), demonstrating their ability to reproduce perceived motion accurately. These 42 observers (mean age = 22.19, SD = 2.46; 17 females, 25 males) then proceeded to the main experimental sessions. They participated in 480 sessions since some observers attended multiple different sessions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 316, + 503, + 327 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 316, + 503, + 327 + ], + "spans": [ + { + "bbox": [ + 314, + 316, + 503, + 327 + ], + "type": "text", + "content": "3.2.2. Perceived vectors and \"flow illusions\"" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 331, + 553, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 553, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 553, + 545 + ], + "type": "text", + "content": "In Fig. 3, we show representative examples of averaged human responses for each dataset (see Supplementary Video S2 for target motions and the corresponding averaged human responses across all probed locations within each dataset). Our findings reveal systematic and stable human-specific biases in motion perception under certain scenarios. We termed these systematic discrepancies between human responses and ground truth motions as \"flow illusions\". These deviations were not random errors but provided valuable insights into the underlying processing and mechanism of human motion perception. On the other hand, we also observed individual differences, with an average inter-observers correlation of 0.407 (SD: 0.417; range from -0.329 to 0.990) across all probed locations. These Individual variations would be useful for training models to reproduce the distribution of human motion perception. To facilitate further research, we have made all raw individual data publicly available on the HuPerFlow website [46]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 546, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 553, + 713 + ], + "type": "text", + "content": "In most driving datasets such as KITTI 2015 [26], Virtual KITTI 2 [12], and Driving [24], the direction and speed of optical flows transition smoothly across frames, with perceived flows aligning closely with the ground truth motion. This consistency supports the reliability of our measurement method in capturing observers' perceptions under spatiotemporally smooth scenes. Notably, as optical flow locations deviated from the scene center, observers often perceived the directions to be slightly shifting toward the central region. This finding may reflect observers' attempts to cancel out the image motion components produced by self-motion (camera motion) during driving. In contrast, other navigation datasets like VIPER [35] and TartanAir [44] show abrupt direction changes, particularly during ac" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22803" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 130, + 49, + 479, + 273 + ], + "blocks": [ + { + "bbox": [ + 130, + 49, + 479, + 273 + ], + "lines": [ + { + "bbox": [ + 130, + 49, + 479, + 273 + ], + "spans": [ + { + "bbox": [ + 130, + 49, + 479, + 273 + ], + "type": "image", + "image_path": "eab63f52866073a542b9286741b20c8687dda5751d44bc7d53cd3cd4590f1702.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 66, + 281, + 541, + 293 + ], + "lines": [ + { + "bbox": [ + 66, + 281, + 541, + 293 + ], + "spans": [ + { + "bbox": [ + 66, + 281, + 541, + 293 + ], + "type": "text", + "content": "Figure 3. Demonstration of human perceived motion vectors. The results cover ten representative existing optical flow benchmarks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 314, + 294, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 314, + 294, + 362 + ], + "spans": [ + { + "bbox": [ + 54, + 314, + 294, + 362 + ], + "type": "text", + "content": "tions such as turning, lane changing, or flying backward. These involve interactions between image motion and camera motion, with perceived flows often revealing systematic errors primarily driven by camera motion." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 363, + 295, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 363, + 295, + 506 + ], + "spans": [ + { + "bbox": [ + 54, + 363, + 295, + 506 + ], + "type": "text", + "content": "In datasets like MPI Sintel [5] and SPRING [25], where images include blurred, non-rigid, discontinuous motion and large jump size (cf. [47] using stimuli with small jumps), observers generally failed to discern individual locations, perceiving them as part of a unified grouping movement instead. Unlike the datasets described in the previous paragraph, which were viewed with smoothly moving cameras, Monkaa [24] dataset involved significant camera rotation around relatively stationary objects. Observers often perceived these objects as moving due to the motion of surrounding objects or backgrounds, illustrating the phenomena of relative and induced motion." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 509, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 509, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 54, + 509, + 295, + 651 + ], + "type": "text", + "content": "In the MHOF [34] datasets, human figures generally exhibit global body movement accompanied by local motions from the head and limbs. Although the motion structures are less complex than in MPI Sintel and Spring, observers tended to focus on the direction of the global body movement and ignored the local gestures. In the Flythings3D [24], multiple objects float along 3D random trajectories, requiring observers to perform object-ground and object-object segmentation. They were successful in perceiving motion against simple backgrounds and distinct surrounding objects but made more errors in cluttered, messy conditions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "content": "As different datasets revealed various types of errors, we calculated the correlation between human response and ground truth for each dataset separately. In general, the driving datasets (i.e., KITTI 2015 [26], Virtual KITTI 2 [12], and Driving [24]) showed high correlations, but oth" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 314, + 555, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 314, + 555, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 314, + 555, + 338 + ], + "type": "text", + "content": "ers showed middle to low correlation (see Supplementary Appendix 2 for the scatter plot of each dataset)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 348, + 517, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 348, + 517, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 348, + 517, + 361 + ], + "type": "text", + "content": "3.2.3. EPE in relation to optical flow properties" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 365, + 555, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 555, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 555, + 532 + ], + "type": "text", + "content": "As shown in Fig. 3, the EPE of human response (indicated by the diameter of the green circle) varied across different positions and scenes. Following [5], we evaluated how EPE changed in relation to various optical flow properties. Fig. 4 shows that EPE increased as GT speed increased, which suggests that faster motions imposed more significant computational challenges. To assess temporal flow dynamics, we compared optical flow from the previous frame (t-1) to the selected probed frame (t-0) and analyzed the temporal gradient changes in uv-vectors over two intervals: from t-1 to t-0, and from t-0 to t+1 (ground truth), normalized by GT speed from the same location. Generally, we observed that EPE increased as the normalized temporal flow gradient increased up to the ratio of 2, and then leveled off." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "text", + "content": "We also considered spatial image properties by using a Sobel kernel to derive the image gradient, calculating the average edge magnitude within a " + }, + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\pm 40" + }, + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "text", + "content": "-pixel neighborhood around the probed location. The results indicate that EPE decreased as the image gradient increased. Additionally, using instance boundary from the MPI Sintel dataset, we calculated the shortest distance to the boundary and found that EPE also decreased as the distance increased. These findings suggest that edges and boundaries may create ambiguities in motion flow interpretation and interrupt motion perception. Finally, using camera motion data from the MPI Sintel as a proxy for self-motion within a scene, we observed that EPE increased with self-motion, and this effect persisted even when excluding the effect of GT speed (image motion)." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22804" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 105, + 54, + 239, + 134 + ], + "blocks": [ + { + "bbox": [ + 105, + 54, + 239, + 134 + ], + "lines": [ + { + "bbox": [ + 105, + 54, + 239, + 134 + ], + "spans": [ + { + "bbox": [ + 105, + 54, + 239, + 134 + ], + "type": "image", + "image_path": "f6cd9664e467d9370b42dc86d615c6c96676048813b0f30e67546cbb352b968b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 240, + 55, + 370, + 130 + ], + "blocks": [ + { + "bbox": [ + 240, + 55, + 370, + 130 + ], + "lines": [ + { + "bbox": [ + 240, + 55, + 370, + 130 + ], + "spans": [ + { + "bbox": [ + 240, + 55, + 370, + 130 + ], + "type": "image", + "image_path": "5d280e5ffb9e5666da580f64b8e33b934a326fcd4e133f38097804794bd2512f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 373, + 56, + 503, + 132 + ], + "blocks": [ + { + "bbox": [ + 373, + 56, + 503, + 132 + ], + "lines": [ + { + "bbox": [ + 373, + 56, + 503, + 132 + ], + "spans": [ + { + "bbox": [ + 373, + 56, + 503, + 132 + ], + "type": "image", + "image_path": "1e574383185e22e89a7e2b88dd5fd65e2d45bb8d91af8b06cdb2967637b392e9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 105, + 135, + 239, + 214 + ], + "blocks": [ + { + "bbox": [ + 105, + 135, + 239, + 214 + ], + "lines": [ + { + "bbox": [ + 105, + 135, + 239, + 214 + ], + "spans": [ + { + "bbox": [ + 105, + 135, + 239, + 214 + ], + "type": "image", + "image_path": "1f5836719aec56efb93590efc195e6442316d2f7757bb8aa004fc5598871bdb8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 223, + 555, + 257 + ], + "lines": [ + { + "bbox": [ + 55, + 223, + 555, + 257 + ], + "spans": [ + { + "bbox": [ + 55, + 223, + 555, + 257 + ], + "type": "text", + "content": "Figure 4. Endpoint errors (EPEs) of human responses and two machine vision models as functions of optical flow properties. The horizontal axis of each panel is the GT speed, normalized temporal flow gradient, spatial image gradient, instance boundary distance, and self-motion with(out) normalized by the image GT speed. The vertical axis represents EPE with different scales for visualization." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 240, + 135, + 372, + 214 + ], + "blocks": [ + { + "bbox": [ + 240, + 135, + 372, + 214 + ], + "lines": [ + { + "bbox": [ + 240, + 135, + 372, + 214 + ], + "spans": [ + { + "bbox": [ + 240, + 135, + 372, + 214 + ], + "type": "image", + "image_path": "e1d79cc765d5d1b6e3ee3fda065d0b267e02bacee21d7edcea4e52926cae09ee.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 373, + 135, + 504, + 210 + ], + "blocks": [ + { + "bbox": [ + 373, + 135, + 504, + 210 + ], + "lines": [ + { + "bbox": [ + 373, + 135, + 504, + 210 + ], + "spans": [ + { + "bbox": [ + 373, + 135, + 504, + 210 + ], + "type": "image", + "image_path": "46216c18208f5bf43c2632f556bad05b9eada102d7a79ed0e37316bb20676104.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 268, + 296, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 268, + 296, + 352 + ], + "spans": [ + { + "bbox": [ + 55, + 268, + 296, + 352 + ], + "type": "text", + "content": "In addition to the EPE of human response, Fig. 4 shows the EPEs of two machine vision models, VideoFlow and FFV1MT (see Sec. 4). While there were clear differences in the absolute level of performance, the optical flow properties generally affect EPE in similar ways for human and machine vision models, except for the effects of spatial image gradient effect on VideoFlow." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 356, + 280, + 370 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 356, + 280, + 370 + ], + "spans": [ + { + "bbox": [ + 55, + 356, + 280, + 370 + ], + "type": "text", + "content": "4. Comparison with optical flow algorithms" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 377, + 295, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 377, + 295, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 377, + 295, + 592 + ], + "type": "text", + "content": "We selected a range of representative optical flow algorithms to assess the similarities and differences among human-perceived flows, ground truth motion, and predictions from computer vision models. These included classical methods based on brightness constancy assumptions, such as Farneback [11]; biologically inspired approaches like FFV1MT [38]; multiscale inference models, such as FlowNet2 [15]; recurrent refinement models like RAFT [42]; transformer-based models like FlowFormer [14]; graph reasoning methods such as AGFlow [23]; multiframe approaches like VideoFlow [37]; Motion energy and self-attention based models like V1Attention [40] and DualModel [41]; as well as general-purpose architectures like Perceiver IO [17]. For a fair comparison, we used each model's default configuration or the best pre-trained model fine-tuned on the Sintel dataset. Since unsupervised learning methods approximate the supervised ground truth, we limited our comparison to supervised methods." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "text", + "content": "For evaluation, we first calculated the average EPE and Pearson correlation coefficients between model predictions and both human responses and ground truth. To further understand the models' alignment with human perception independently of ground truth influence—since most models were trained to fit ground truth and the ground truth was inherently correlated with human responses—we also calculated the partial correlation between model predictions and human responses, controlling for ground truth, as shown in Equation 2. Here, " + }, + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "inline_equation", + "content": "r_{x,y}" + }, + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "text", + "content": " denotes correlation coefficients be" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 269, + 376, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 269, + 376, + 279 + ], + "spans": [ + { + "bbox": [ + 314, + 269, + 376, + 279 + ], + "type": "text", + "content": "tween " + }, + { + "bbox": [ + 314, + 269, + 376, + 279 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 314, + 269, + 376, + 279 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 314, + 269, + 376, + 279 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 314, + 269, + 376, + 279 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 317, + 289, + 553, + 330 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 289, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 317, + 289, + 553, + 330 + ], + "type": "interline_equation", + "content": "\\rho_ {\\text {m o d e l}} = r _ {\\text {r e s p , m o d e l} \\cdot \\mathrm {G T}} = \\frac {r _ {\\text {r e s p , m o d e l}} - r _ {\\text {r e s p , G T}} \\cdot r _ {\\text {m o d e l , G T}}}{\\sqrt {1 - r _ {\\text {r e s p , G T}} ^ {2}} \\cdot \\sqrt {1 - r _ {\\text {m o d e l , G T}} ^ {2}}}. \\tag {2}", + "image_path": "d56c00fee504393261fe565042efe2429fe304015d65619920e1835cca609080.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 331, + 554, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 554, + 413 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 554, + 413 + ], + "type": "text", + "content": "To complement partial correlation analysis, we also adopted a Response Consistency Index (RCI) from [47]. This index measures the similarity between model predictions and human responses at each probed location, approaching " + }, + { + "bbox": [ + 313, + 331, + 554, + 413 + ], + "type": "inline_equation", + "content": "+1" + }, + { + "bbox": [ + 313, + 331, + 554, + 413 + ], + "type": "text", + "content": " when the model's prediction aligns closely with human flow illusions (see Supplementary Appendix 3 for more details)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 415, + 554, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 554, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 554, + 557 + ], + "type": "text", + "content": "Tab. 2 summarizes model predictions across all datasets. Overall, VideoFlow's predictions were the closest to the ground truth. RAFT showed the most closely resembling human responses while also having a strong alignment with the ground truth (second-best among models). When controlling for the ground truth effect, FFV1MT and V1Attention-StageI were the most similar to human flow illusions, as indicated by high partial correlations and average RCI values. A detailed analysis comparing the predictions of optical flow algorithms with both the ground truth and human responses for each dataset is provided in Tables S1-S12 of Supplementary Appendix 4." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 558, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 556, + 714 + ], + "type": "text", + "content": "Fig. 5 provides examples of model predictions, highlighting examples where predictions aligned closely with the ground truth (e.g., VideoFlow and RAFT) versus those having some resemblance to human response errors (e.g., FFV1MT and V1Attention-StageI). For instance, at Point (A-I) in Fig. 5, both flow illusions and FFV1MT exhibited temporal-sluggish responses but VideoFlow aligned well with GT. Conversely, point (A-II) in Fig. 5 both model predictions were similar to flow illusions, likely due to image signal loss, as the GT indicated a small, dark stone falling. At Point (B-I) and other selected body locations, both human perception and FFV1MT reflected grouping motion in the same direction, regardless of local gesture directions." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22805" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 72, + 92, + 536, + 266 + ], + "blocks": [ + { + "bbox": [ + 55, + 50, + 555, + 85 + ], + "lines": [ + { + "bbox": [ + 55, + 50, + 555, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 50, + 555, + 85 + ], + "type": "text", + "content": "Table 2. Predictions of optical flow algorithms versus Human response or ground truth (GT). " + }, + { + "bbox": [ + 55, + 50, + 555, + 85 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 55, + 50, + 555, + 85 + ], + "type": "text", + "content": " : Partial correlation; " + }, + { + "bbox": [ + 55, + 50, + 555, + 85 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 50, + 555, + 85 + ], + "type": "text", + "content": " : Pearson correlation coefficient; EPE: vector end-point error; uv, dir, spd represent motion components in Cartesian space, direction, and speed, respectively. RCI is introduced from [47] to represent the model's similarity to human perception (the larger, the more human-aligned)." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 72, + 92, + 536, + 266 + ], + "lines": [ + { + "bbox": [ + 72, + 92, + 536, + 266 + ], + "spans": [ + { + "bbox": [ + 72, + 92, + 536, + 266 + ], + "type": "table", + "html": "
MethodρRCIv.s. Humanv.s. GT
ρuvρdirρspdruvrdirrspdEPEruvrdirrspdEPE
Ground TruthNaNNaNNaNNaN0.660.430.296.961.001.001.000.00
Farnebäck [11]0.170.140.040.130.540.380.156.910.680.650.395.52
FFV1MT [38]0.310.280.060.150.630.450.206.580.700.590.545.84
RAFT [42]0.110.220.010.050.640.470.237.370.930.850.792.19
FlowNet2 [15]0.100.260.000.070.590.480.187.550.820.800.613.24
FlowFormer [14]0.070.19-0.020.040.640.460.217.430.930.870.792.08
VideoFlow [49]0.080.22-0.020.040.640.470.227.470.940.880.811.91
Perceiver IO [17]0.090.170.030.050.620.440.237.750.890.820.742.87
AGFlow [23]0.160.240.010.110.540.450.157.940.680.660.465.44
V1Attention-StageI [40]0.230.280.020.180.550.450.146.360.630.540.427.00
DualModel [41]0.200.250.060.130.570.450.207.310.700.670.535.66
", + "image_path": "21cee0d09d52486930b0921198cf5b3d0e05b7c9a8764907d8372104fc4b0146.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 59, + 285, + 293, + 559 + ], + "blocks": [ + { + "bbox": [ + 59, + 285, + 293, + 559 + ], + "lines": [ + { + "bbox": [ + 59, + 285, + 293, + 559 + ], + "spans": [ + { + "bbox": [ + 59, + 285, + 293, + 559 + ], + "type": "image", + "image_path": "098a852fcfbaf5c7796842bd6676713224dd1afe568ece9ecb2331b7dcda6e76.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 570, + 295, + 636 + ], + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 636 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 636 + ], + "type": "text", + "content": "Figure 5. Predicted vectors of optical flow algorithms. Examples are FFV1MT [38], V1Attention [40], VideoFlow [37], and RAFT [42]. We marked some points with large RCI differences between the models in the left and right columns. Generally, the model on the left side shows a more human-aligned response over ground truth, while the right side shows the opposite result." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "At Point (C-I), a strong self-motion effect occurred toward a static motorcycle. Despite the strong physical optical flow in the image, human observers and V1Attention-StageI appeared to compensate for self-motion to perceive the motorcycle as stationary. For more examples, please visit the HuPerFlow website [46]. Note, however, that there remains" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 286, + 555, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 286, + 555, + 347 + ], + "spans": [ + { + "bbox": [ + 313, + 286, + 555, + 347 + ], + "type": "text", + "content": "a large gap between human responses and the most similar model predictions, possibly because the purpose of human visual motion processing is not merely to estimate image flow. The primary purpose of making our dataset is to help future studies develop models that can fill this gap." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 379, + 388, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 379, + 388, + 392 + ], + "spans": [ + { + "bbox": [ + 314, + 379, + 388, + 392 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 407, + 555, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 555, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 555, + 598 + ], + "type": "text", + "content": "By overcoming technical challenges, we developed a novel online paradigm to collect human motion perception data, introducing the first large-scale benchmark for human-perceived optical flow. Our findings identified distinct perceptual illusions across various scene scenarios, some aligning with and others diverging from the predictions of CV optical flow algorithms. Human motion perception, compared to state-of-the-art CV models, has limited spatial, temporal, and intensity resolutions and struggles with motion estimation in retinal coordinates. However, human motion vision has remarkable capabilities in analyzing hierarchical movement structures and distinguishing object motion from self-motion in world coordinates. Our proposed methods can be extended to collect broader human data from diverse datasets, further bridging gaps between CV, human perception, and ground truth." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "While our dataset may not be sufficient to train models from zero, it offers valuable benchmarks for model evaluation and fine-tuning. Physically accurate CV models are crucial, but aligning with human perception is equally important for human-centered applications. Notably, current optical flow models fail to predict various human-perceived motion illusions, highlighting the need to address this gap both for practical applications and for advancing the scientific understanding of human vision." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "22806" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 140 + ], + "type": "text", + "content": "This work was supported in part by JSPS Grants-in-Aid for Scientific Research (KAKENHI), Grant Numbers JP20H00603, JP20H05957, and JP24H00721, and by the Spring Fellowship, Grant Number JPMJFS2123." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 115, + 164 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 115, + 164 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 115, + 164 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 171, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 171, + 296, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 171, + 296, + 204 + ], + "spans": [ + { + "bbox": [ + 61, + 171, + 296, + 204 + ], + "type": "text", + "content": "[1] Edward H Adelson and James R Bergen. Spatiotemporal energy models for the perception of motion. Josa a, 2(2): 284-299, 1985. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 206, + 296, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 206, + 296, + 239 + ], + "spans": [ + { + "bbox": [ + 61, + 206, + 296, + 239 + ], + "type": "text", + "content": "[2] Stuart M Anstis and Brian J Rogers. Illusory continuous motion from oscillating positive-negative patterns: Implications for motion perception. Perception, 15(5):627-640, 1986. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 240, + 296, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 240, + 296, + 283 + ], + "spans": [ + { + "bbox": [ + 62, + 240, + 296, + 283 + ], + "type": "text", + "content": "[3] Simon Baker, Daniel Scharstein, J. P. Lewis, Stefan Roth, Michael J. Black, and Richard Szeliski. A Database and Evaluation Methodology for Optical Flow. International Journal of Computer Vision, 92(1):1-31, 2011. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 285, + 295, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 285, + 295, + 318 + ], + "spans": [ + { + "bbox": [ + 62, + 285, + 295, + 318 + ], + "type": "text", + "content": "[4] Richard O Brown and Sheng He. Visual motion of missing-fundamental patterns: motion energy versus feature correspondence. Vision Research, 40(16):2135-2147, 2000. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 319, + 296, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 319, + 296, + 364 + ], + "spans": [ + { + "bbox": [ + 62, + 319, + 296, + 364 + ], + "type": "text", + "content": "[5] Daniel J. Butler, Jonas Wulff, Garrett B. Stanley, and Michael J. Black. A naturalistic open source movie for optical flow evaluation. In European Conf. on Computer Vision (ECCV), pages 611-625. Springer-Verlag, 2012. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 365, + 296, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 365, + 296, + 408 + ], + "spans": [ + { + "bbox": [ + 62, + 365, + 296, + 408 + ], + "type": "text", + "content": "[6] Haiyang Chao, Yu Gu, and Marcello Napolitano. A survey of optical flow techniques for robotics navigation applications. Journal of Intelligent & Robotic Systems, 73:361-372, 2014. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 411, + 296, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 411, + 296, + 464 + ], + "spans": [ + { + "bbox": [ + 62, + 411, + 296, + 464 + ], + "type": "text", + "content": "[7] Chenyi Chen, Ari Seff, Alain Kornhauser, and Jianxiong Xiao. Deepdriving: Learning affordance for direct perception in autonomous driving. In Proceedings of the IEEE international conference on computer vision, pages 2722-2730, 2015. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 467, + 296, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 467, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 62, + 467, + 296, + 510 + ], + "type": "text", + "content": "[8] Jaechan Cho, Yongchul Jung, Dong-Sun Kim, Seongjoo Lee, and Yunho Jung. Moving object detection based on optical flow estimation and a gaussian mixture model for advanced driver assistance systems. Sensors, 19(14):3217, 2019. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 512, + 296, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 512, + 296, + 576 + ], + "spans": [ + { + "bbox": [ + 62, + 512, + 296, + 576 + ], + "type": "text", + "content": "[9] Carl Doersch, Ankush Gupta, Larisa Markeeeva, Adrià Recasens, Lucas Smaira, Yusuf Aytar, João Carreira, Andrew Zisserman, and Yi Yang. Tap-vid: A benchmark for tracking any point in a video. arXiv preprint arXiv:2211.03726, 2022. Published in NeurIPS Datasets and Benchmarks track, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 579, + 295, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 579, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 57, + 579, + 295, + 622 + ], + "type": "text", + "content": "[10] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 624, + 295, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 295, + 667 + ], + "type": "text", + "content": "[11] Gunnar Farnebäck. Two-frame motion estimation based on polynomial expansion. In Proceedings of the Scandinavian conference on Image analysis, pages 363-370. Springer, 2003. 2, 7, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "text", + "content": "[12] Adrien Gaidon, Qiao Wang, Yohann Cabon, and Eleonora Vig. Virtual worlds as proxy for multi-object tracking analysis. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 117 + ], + "type": "text", + "content": "[13] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 119, + 554, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 554, + 172 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 554, + 172 + ], + "type": "text", + "content": "[14] Tzehoong Huang, Zhichao Zhang, Wenjie Zeng, Ming Liu, Chunjing Zhang, Hongsheng Li, and Shijian Lu. Flowformer: A transformer architecture for optical flow. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 175, + 554, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 175, + 554, + 239 + ], + "spans": [ + { + "bbox": [ + 317, + 175, + 554, + 239 + ], + "type": "text", + "content": "[15] Eddy Ilg, Nikolaus Mayer, Tonmoy Saikia, Margret Keuper, Alexey Dosovitskiy, and Thomas Brox. Flownet 2.0: Evolution of optical flow estimation with deep networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2462-2470, 2017. 3, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 240, + 554, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 240, + 554, + 296 + ], + "spans": [ + { + "bbox": [ + 316, + 240, + 554, + 296 + ], + "type": "text", + "content": "[16] Eddy Ilg, Tonmoy Saikia, Margret Keuper, and Thomas Brox. Occlusions, motion and depth boundaries with a generic network for disparity, optical flow or scene flow estimation. In European Conference on Computer Vision (ECCV), 2018. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 297, + 554, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 554, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 554, + 341 + ], + "type": "text", + "content": "[17] Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, et al. Perceiver io: A general architecture for structured inputs & outputs. arXiv preprint arXiv:2107.14795, 2021. 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 342, + 554, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 554, + 396 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 554, + 396 + ], + "type": "text", + "content": "[18] Joel Janai, Fatma Güney, Jonas Wulff, Michael J Black, and Andreas Geiger. Slow flow: Exploiting high-speed cameras for accurate and diverse optical flow reference data. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3597-3606, 2017. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 398, + 554, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 398, + 554, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 398, + 554, + 430 + ], + "type": "text", + "content": "[19] Gunnar Johansson. Visual perception of biological motion and a model for its analysis. Perception & Psychophysics, 14(2):201-211, 1973. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 432, + 554, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 432, + 554, + 454 + ], + "spans": [ + { + "bbox": [ + 316, + 432, + 554, + 454 + ], + "type": "text", + "content": "[20] Bela Julesz. Foundations of Cyclopean Perception. University of Chicago Press, 1971. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 456, + 554, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 554, + 477 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 554, + 477 + ], + "type": "text", + "content": "[21] D. H. Kelly. Fourier components of moving gratings. Behavior Research Methods, 14(6):435-437, 1982. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 479, + 554, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 479, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 317, + 479, + 554, + 521 + ], + "type": "text", + "content": "[22] Monika Lohani, Brennan R Payne, and David L Strayer. A review of psychophysiological measures to assess cognitive states in real-world driving. Frontiers in human neuroscience, 13:57, 2019. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 523, + 554, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 523, + 554, + 567 + ], + "spans": [ + { + "bbox": [ + 317, + 523, + 554, + 567 + ], + "type": "text", + "content": "[23] Ao Luo, Fan Yang, Kunming Luo, Xin Li, Haoqiang Fan, and Shuaicheng Liu. Learning optical flow with adaptive graph reasoning. In Proceedings of the AAAI conference on artificial intelligence, pages 1890-1898, 2022. 7, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 568, + 554, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 568, + 554, + 633 + ], + "spans": [ + { + "bbox": [ + 317, + 568, + 554, + 633 + ], + "type": "text", + "content": "[24] Nikolaus Mayer, Eddy Ilg, Philip Häusser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In IEEE International Conference on Computer Vision and Pattern Recognition (CVPR), 2016. arXiv:1512.02134. 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 635, + 554, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 554, + 689 + ], + "type": "text", + "content": "[25] Lukas Mehl, Jenny Schmalfuss, Azin Jahedi, Yaroslava Naliyayko, and Andres Bruhn. Spring: A high-resolution high-detail dataset and benchmark for scene flow, optical flow and stereo. In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3, 4, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 691, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 554, + 713 + ], + "type": "text", + "content": "[26] Moritz Menze and Andreas Geiger. Object scene flow for autonomous vehicles. In Proceedings of the IEEE conference" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22807" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 73, + 295, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 295, + 95 + ], + "type": "text", + "content": "on computer vision and pattern recognition, pages 3061-3070, 2015. 1, 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "type": "text", + "content": "[27] Shinji Nishimoto and Jack L. Gallant. A three-dimensional spatiotemporal receptive field model explains responses of area mt neurons to naturalistic movies. Journal of Neuroscience, 31(41):14551-14564, 2011. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 207 + ], + "type": "text", + "content": "[28] Mayu Otani, Riku Togashi, Yu Sawai, Ryosuke Ishigami, Yuta Nakashima, Esa Rahtu, Janne Heikkilä, and Shin'ichi Satoh. Toward verifiable and reproducible human evaluation for text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14277-14286, 2023. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 295, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 295, + 273 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 295, + 273 + ], + "type": "text", + "content": "[29] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730-27744, 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 275, + 295, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 275, + 295, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 275, + 295, + 308 + ], + "type": "text", + "content": "[30] Christopher C Pack and Richard T Born. Temporal dynamics of a neural solution to the aperture problem in visual area mt of macaque brain. Nature, 409(6823):1040-1042, 2001. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 309, + 295, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 295, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 295, + 342 + ], + "type": "text", + "content": "[31] Jonathan Peirce and the PsychoPy team. Pavlovia: Online platform for running experiments. Open Science Tools, 2018. Accessed: 2024-10-25. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 343, + 295, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 295, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 295, + 376 + ], + "type": "text", + "content": "[32] Jonathan W. Peirce and Michael MacAskill. Building Experiments in PsychoPy. Sage Publications Ltd, London, United Kingdom, 2018. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 377, + 295, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 377, + 295, + 421 + ], + "spans": [ + { + "bbox": [ + 56, + 377, + 295, + 421 + ], + "type": "text", + "content": "[33] Silvia Pitzalis, Carlo Serra, Valerio Sulpizio, Giovanni d'Avossa, and Gaspare Galati. Neural bases of self- and object-motion in a naturalistic vision. Human Brain Mapping, 41(4):1084–1111, 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 422, + 295, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 295, + 465 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 295, + 465 + ], + "type": "text", + "content": "[34] Anurag Ranjan, David T. Hoffmann, Dimitrios Tzionas, Siyu Tang, Javier Romero, and Michael J. Black. Learning multi-human optical flow. International Journal of Computer Vision (IJCV), 2020. 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 468, + 295, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 468, + 295, + 511 + ], + "spans": [ + { + "bbox": [ + 56, + 468, + 295, + 511 + ], + "type": "text", + "content": "[35] Stephan R. Richter, Zeeshan Hayder, and Vladlen Koltun. Playing for benchmarks. In IEEE International Conference on Computer Vision, ICCV 2017, Venice, Italy, October 22-29, 2017, pages 2232-2241, 2017. 2, 3, 4, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 513, + 295, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 513, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 513, + 295, + 578 + ], + "type": "text", + "content": "[36] Laura Sevilla-Lara, Yiyi Liao, Fatma Güney, Varun Jampani, Andreas Geiger, and Michael J Black. On the integration of optical flow and action recognition. In Pattern Recognition: 40th German Conference, GCPR 2018, Stuttgart, Germany, October 9-12, 2018, Proceedings 40, pages 281-297. Springer, 2019. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 580, + 295, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 580, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 580, + 295, + 634 + ], + "type": "text", + "content": "[37] Xiaoyu Shi, Zhaoyang Huang, Weikang Bian, Dasong Li, Manyuan Zhang, Ka Chun Cheung, Simon See, Hongwei Qin, Jifeng Dai, and Hongsheng Li. Videoflow: Exploiting temporal cues for multi-frame optical flow estimation. arXiv preprint arXiv:2303.08340, 2023. 2, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 635, + 295, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 295, + 678 + ], + "type": "text", + "content": "[38] Fabio Solari, Manuela Chessa, Narasimha K. Medathati, and Pierre Kornprobst. What can we expect from a v1-mt feedforward architecture for optical flow estimation? Signal Processing: Image Communication, 39:342-354, 2015. 2, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "type": "text", + "content": "[39] Shuyang Sun, Zhanghui Kuang, Lu Sheng, Wanli Ouyang, and Wei Zhang. Optical flow guided feature: A fast and robust motion representation for video action recognition. In" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 609 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1390-1399, 2018. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "text", + "content": "[40] Zitang Sun, Yen-Ju Chen, Yung-Hao Yang, and Shin'ya Nishida. Modeling human visual motion processing with trainable motion energy sensing and a self-attention network. In Advances in Neural Information Processing Systems, 2023. 2, 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "type": "text", + "content": "[41] Zitang Sun, Yen-Ju Chen, Yung-Hao Yang, and Shin'ya Nishida. Acquisition of second-order motion perception by learning to recognize the motion of objects made by nondiffusive materials. Journal of Vision, Vision Sciences Society Annual Meeting Abstract, 24(9), 2024. 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "type": "text", + "content": "[42] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Proceedings of the European Conference on Computer Vision (ECCV), pages 402-419, 2020. 2, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 252, + 553, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 252, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 316, + 252, + 553, + 296 + ], + "type": "text", + "content": "[43] Longguang Wang, Yulan Guo, Li Liu, Zaiping Lin, Xinpu Deng, and Wei An. Deep video super-resolution using hr optical flow estimation. IEEE Transactions on Image Processing, 29:4323-4336, 2020. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 297, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 342 + ], + "type": "text", + "content": "[44] Wenshan Wang, Delong Zhu, Xiangwei Wang, Yaoyu Hu, Yuheng Qiu, Chen Wang, Yafei Hu, Ashish Kapoor, and Sebastian Scherer. Tartanair: A dataset to push the limits of visual slam. arXiv preprint arXiv:2003.14338, 2020. 3, 4, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 343, + 553, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 553, + 407 + ], + "type": "text", + "content": "[45] H. Wu, Y. Liu, Z. Zhang, L. Du, L. Wang, Y. Wang, D. Lin, and L. Lin. Q-Instruct: Improving Low-Level Visual Abilities for Multi-Modality Foundation Models. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 25490–25500, Seattle, WA, USA, 2024. IEEE. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 409, + 553, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 409, + 553, + 463 + ], + "spans": [ + { + "bbox": [ + 316, + 409, + 553, + 463 + ], + "type": "text", + "content": "[46] Yung-Hao Yang, Zitang Sun, Taiki Fukiage, and Shin'ya Nishida. HuPerFlow: A comprehensive benchmark for human vs. machine motion estimation comparison. https://kucognitiveinformaticslab.github.io/Huperflow-Website/. Accessed: March 2025. 5, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 464, + 553, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 464, + 553, + 507 + ], + "spans": [ + { + "bbox": [ + 316, + 464, + 553, + 507 + ], + "type": "text", + "content": "[47] Yung-Hao Yang, Taiki Fukiage, Zitang Sun, and Shin'ya Nishida. Psychophysical measurement of perceived motion flow of naturalistic scenes. iScience, 26(12), 2023. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 509, + 553, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 509, + 553, + 564 + ], + "spans": [ + { + "bbox": [ + 316, + 509, + 553, + 564 + ], + "type": "text", + "content": "[48] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 566, + 553, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 566, + 553, + 609 + ], + "spans": [ + { + "bbox": [ + 316, + 566, + 553, + 609 + ], + "type": "text", + "content": "[49] Xizhou Zhu, Yuwen Xiong, Bo Dai, and Dahua Lin. Deep feature flow for video recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2349-2358, 2017. 8" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22808" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Human Motion Instruction Tuning/8c08c8b0-9aae-431a-b68e-699a5ddd5c32_content_list.json b/2025/Human Motion Instruction Tuning/8c08c8b0-9aae-431a-b68e-699a5ddd5c32_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..61212672362fb1881297b611d42a7da0b8a6d477 --- /dev/null +++ b/2025/Human Motion Instruction Tuning/8c08c8b0-9aae-431a-b68e-699a5ddd5c32_content_list.json @@ -0,0 +1,1857 @@ +[ + { + "type": "text", + "text": "Human Motion Instruction Tuning", + "text_level": 1, + "bbox": [ + 321, + 130, + 676, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lei Li $^{1,2,\\ast,\\dagger}$ , Sen Jia $^{1,\\ast}$ , Jianhao Wang $^{3}$ , Zhongyu Jiang $^{1}$ , Feng Zhou $^{4}$ , Ju Dai $^{5}$ , Tianfang Zhang $^{6}$ , Zongkai Wu $^{1}$ , Jenq-Neng Hwang $^{1}$", + "bbox": [ + 218, + 191, + 776, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 265, + 326, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper presents LLaMo (Large Language and Human Motion Assistant), a multimodal framework for human motion instruction tuning. In contrast to conventional instruction-tuning approaches that convert non-linguistic inputs, such as video or motion sequences, into language tokens, LLaMo retains motion in its native form for instruction tuning. This method preserves motion-specific details that are often diminished in tokenization, thereby improving the model's ability to interpret complex human behaviors. By processing both video and motion data alongside textual inputs, LLaMo enables a flexible, human-centric analysis. Experimental evaluations across high-complexity domains, including human behaviors and professional activities, indicate that LLaMo effectively captures domain-specific knowledge, enhancing comprehension and prediction in motion-intensive scenarios. We hope LLaMo offers a foundation for future multimodal AI systems with broad applications, from sports analytics to behavioral prediction.", + "bbox": [ + 88, + 297, + 483, + 571 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 602, + 220, + 618 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Understanding human motion is a central challenge in multimodal AI, impacting numerous fields such as digital human, human-computer interaction, sports analytics, healthcare, and virtual human modeling [3-6]. Motion data, which captures skeletal movements and body dynamics, provides a structured, appearance-invariant representation of human actions, concentrating on the essential movement patterns while excluding irrelevant visual details [7]. This data has emerged as a privacy-conscious alternative to visual inputs, offering finer control over action recognition and facilitating analysis in applications ranging from digital human avatars to behavioral monitoring in healthcare [8-10]. As digital human representations and motion-centric", + "bbox": [ + 89, + 628, + 483, + 825 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6929f4853c166dfc7e60a280057512ce73bd37d5595f226416333ac6f813430e.jpg", + "image_caption": [ + "Figure 1. A comparison of MotionLLM [1], MotionGPT [2], and LLaMo highlights LLaMo's motion-specific capabilities. Equipped with a Motion Enhancer and Cross Talker module to align motion and text, LLaMo supports both video and motion inputs, enabling text-aware, fine-grained motion analysis." + ], + "image_footnote": [], + "bbox": [ + 514, + 263, + 908, + 470 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "applications continue to expand, developing models that can leverage the unique characteristics of motion data becomes increasingly crucial for achieving robust, context-aware behavior analysis.", + "bbox": [ + 511, + 580, + 906, + 641 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, large language models (LLMs) have shown promise in motion analysis, leveraging their capacity to model temporal dependencies across multimodal data [1, 11]. State-of-the-art models, such as MotionGPT [12] and MotionLLM [1], treat motion data as a multimodal extension, encoding it as discrete tokens or translating it into language-like representations for processing by LLMs. These models have yielded notable results in motion understanding and recognition tasks, illustrating the adaptability of language models to non-linguistic data. However, their reliance on tokenization or textual conversion can limit the richness of the motion representation, abstracting away critical spatial-temporal details essential for high-fidelity human activity recognition.", + "bbox": [ + 511, + 642, + 908, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite recent advancements, notable challenges remain in current approaches to motion analysis using LLMs. First, encoding motion data into language tokens, as implemented", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* These authors contributed equally to this work.", + "bbox": [ + 112, + 838, + 374, + 851 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding Author. (lilei@di.ku.dk)", + "bbox": [ + 112, + 851, + 372, + 863 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1University of Washington 2University of Copenhagen 3Xi'an Jiaotong University 4North China University of Technology 5Peng Cheng Laboratory 6Tsinghua University", + "bbox": [ + 91, + 863, + 482, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "17582", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "in models such as MotionGPT, relies on a quantization process that may obscure critical fine-grained motion information, potentially limiting the model's capacity to accurately capture detailed spatial-temporal dynamics [13, 14]. The common approach of translating motion data into text tokens further contributes to this issue, often resulting in the loss of essential motion-specific nuances that are critical for in-depth behavior interpretation [11, 15]. This translation process poses considerable challenges in understanding the intricate details of human movement, impeding the model's ability to achieve spatial and physical comprehension of human actions. Moreover, motion data inherently contains 3D structural information and insights into connected human behaviors, which are often difficult to fully capture when relying solely on language tokens, as this representation can overlook the underlying spatial dependencies crucial to modeling human motion accurately.", + "bbox": [ + 89, + 90, + 480, + 348 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In addition, existing methods frequently treat video and motion data as isolated modalities, overlooking the potential benefits of a complementary, integrated approach [16]. Processing these data types independently neglects the rich environmental context and interaction cues provided by video inputs, which are crucial for accurately interpreting complex human behaviors [5, 8]. This separation limits the capacity for a holistic understanding of human motion, particularly in scenarios where context, spatial relations, and dynamic interactions play essential roles in behavior analysis. Furthermore, video-based approaches, while potentially more informative, are computationally intensive, posing significant constraints for large-scale or real-time applications, which limits their feasibility in practical motion analysis systems.", + "bbox": [ + 89, + 351, + 482, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In response to these challenges, we propose LLaMo, a framework designed to integrate motion data as a distinct modality within LLMs without translating it into textual intermediaries. LLaMo incorporates a Motion Estimator and an Enhancer to facilitate the direct input of motion or/and video data, thereby offering a unified, human-centric analysis pipeline. The framework also introduces a Cross Talker module, which dynamically aligns the motion and text features, allowing for text-aware, fine-grained motion representations. This approach retains the original spatial-temporal characteristics of the motion data, providing a more precise understanding of human actions and enabling high-resolution behavior analysis in motion-intensive contexts [17, 18].", + "bbox": [ + 89, + 580, + 482, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Through extensive evaluations on benchmarks such as MoVid-Bench [1] and BABEL-QA [19], we demonstrate that LLaMo achieves state-of-the-art performance in motion-centric human activity recognition [11, 15]. These results underscore LLaMo's potential to advance multimodal AI, providing a robust foundation for human behavior understanding. Our contributions can be summarized as", + "bbox": [ + 89, + 795, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "follows:", + "bbox": [ + 513, + 92, + 571, + 104 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce LLaMo, which treats motion data as an independent modality within LLMs, preserving essential motion-specific details for robust analysis.", + "- We propose Cross Talker, a text-guided mechanism that dynamically focuses on and aggregates key motion frames features, optimizing computational efficiency and enabling the model to pick relevant motion features.", + "- LLaMo's architecture supports the direct input of both raw motion and video data, providing a generalized framework for human-centric analysis across diverse applications, such as sports analytics, healthcare, and behavioral monitoring." + ], + "bbox": [ + 513, + 107, + 903, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 303, + 653, + 318 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Human-Centric Multimodal Representation", + "text_level": 1, + "bbox": [ + 513, + 328, + 885, + 344 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multimodal representation learning is integral to human-centric analysis, especially for applications requiring spatial-temporal reasoning to decode complex behaviors [20-22]. Recent advances, such as Video-LLaVA, have made significant strides by embedding visual data from images and videos into a shared linguistic feature space, enabling sophisticated visual reasoning for behavioral analysis tasks [20, 23]. Despite these achievements, many current models are optimized primarily for static images or discrete video frames, limiting their effectiveness in sequential, dynamic scenarios where understanding progression and continuity in motion is crucial [9, 21, 24]. Privacy concerns further challenge the adoption of video-based methods in sensitive domains; thus, researchers are exploring motion data as a privacy-conscious alternative, enabling analysis focused on human actions without revealing identifiable visual information [25-27]. By combining visual and motion data, emerging multimodal frameworks hold promise for more comprehensive, privacy-aware human behavior analysis, leveraging the strengths of both modalities for adaptability and depth across applications.", + "bbox": [ + 511, + 352, + 906, + 670 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Human Motion Understanding", + "text_level": 1, + "bbox": [ + 513, + 680, + 785, + 696 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Human motion understanding has traditionally relied on skeletal data represented as sequences of joint keypoints to capture movement while preserving user privacy [13, 28-30]. Early models like 2s-AGCN [31] and more recent transformer-based methods, such as MotionCLIP [2], have achieved success in tasks like activity recognition, captioning, and behavior analysis by mapping motion data to language tokens, effectively capturing structural aspects of movement. However, these methods often lack environmental context, which is crucial for nuanced interpretations in real-world applications, as similar movements can convey different meanings across varied scenarios [24, 32, 33]. To address this limitation, recent frameworks inte", + "bbox": [ + 511, + 704, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "17583", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "grate motion and visual data, allowing models to generalize across diverse, dynamic environments [34-36]. This fusion of skeletal data with video-provided context has shown promise in applications like sports and professional activity analysis, where contextual cues are essential. Additionally, skeletal-based data remains valuable in privacy-sensitive applications, such as healthcare and security, where it allows for robust analysis without compromising personal identity.", + "bbox": [ + 89, + 90, + 480, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methods", + "text_level": 1, + "bbox": [ + 89, + 239, + 189, + 255 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The LLaMo framework is composed of three primary modules to effectively process and integrate video, motion data and text. An overview of this pipeline is provided in Figure 2. The architecture comprises the following components:", + "bbox": [ + 89, + 263, + 482, + 339 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Multimodal Feature Extraction Module: This module independently encodes the video or motion inputs, which makes LLaMo is a general human motion assistant with video input or motion input. Then the Motion features are enhanced by distilling valuable representations in video.", + "2. Cross Talker: Here, the enhanced motion features are aggregated in a text-guided manner and aligned with text within a shared semantic space, allowing the natural feature fusion between motion and text, which provides fine-grained inputs for LLM.", + "3. Behavior Generation Module: Using the aggregated and text-aligned motion features, along with the text representations, this module generates a contextually aware textual description of the observed human behavior." + ], + "bbox": [ + 89, + 340, + 482, + 569 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Enhanced motion Feature Extraction Module", + "text_level": 1, + "bbox": [ + 89, + 577, + 475, + 592 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The enhanced motion feature extraction module consists of four primary components: the motion estimator, followed by the motion and video encoders, and the feature enhancement module. The motion estimator addresses scenarios where motion data may be unavailable by estimating motion information directly from video frames. The motion and video encoders then independently process the motion and video data, encoding them into their respective feature spaces. This approach allows the system to leverage complementary information from both modalities, creating a richer representation of human behavior.", + "bbox": [ + 89, + 599, + 482, + 765 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a sequence of video frames $V = \\{v_{1}, v_{2}, \\ldots, v_{T}\\}$ , the video encoder $f_{v}(\\cdot)$ extracts a set of visual features:", + "bbox": [ + 89, + 766, + 482, + 797 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nF _ {V} = f _ {v} (V) = \\left\\{f _ {v} \\left(v _ {1}\\right), f _ {v} \\left(v _ {2}\\right), \\dots , f _ {v} \\left(v _ {T}\\right) \\right\\}.\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 805, + 441, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Simultaneously, the motion encoder $f_{m}(\\cdot)$ processes the motion data $M = \\{m_1, m_2, \\ldots, m_T\\}$ to generate motion-specific features:", + "bbox": [ + 89, + 830, + 482, + 876 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nF _ {M} = f _ {m} (M) = \\left\\{f _ {m} \\left(m _ {1}\\right), f _ {m} \\left(m _ {2}\\right), \\dots , f _ {m} \\left(m _ {T}\\right) \\right\\}.\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 883, + 465, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, $F_{V}\\in \\mathbb{R}^{T\\times H}$ and $F_{M}\\in \\mathbb{R}^{T\\times H}$ represent the extracted feature sets for the video and motion data, respectively, where $T$ is the sequence length and $H$ denotes the feature dimension.", + "bbox": [ + 511, + 89, + 903, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The feature enhancement module then aligns and fuses these modality-specific features. By using the motion features $F_{M}$ as queries, this module selectively extracts semantically relevant information from the video features $F_{V}$ . This alignment and fusion process allows the motion features to be enriched with contextual information from the video, producing refined, contextually enriched representations that enhance the model's understanding of nuanced human behaviors. For the detailed design refer to the Appendix.", + "bbox": [ + 511, + 154, + 906, + 306 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "First, the video features and motion features will conduct self-attention operations respectively to get their augmented features $F_V'$ and $F_M'$ . This allows the video and motion to initially capture the dependencies on their respective time steps. Then, in order to enrich the motion data features with semantic information from the video, we subtly design a cross-attention block where augmented motion features as the query to extract valid semantic information in the video features. Finally, the selected video semantic representation and motion features are connected by residuals followed by a feed-forward neural network to obtain the enhanced motion features $\\tilde{F_M}$", + "bbox": [ + 511, + 308, + 908, + 491 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Cross Talker Module", + "text_level": 1, + "bbox": [ + 513, + 508, + 712, + 523 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given the enhanced motion features $\\tilde{F}_M$ from the feature enhancement module, directly feeding the entire sequence of motion frames into the Large Language Model (LLM) incurs significant computational costs and may lead to suboptimal performance. This limitation arises from the self-attention mechanism in transformers, which has a computational complexity of $\\mathcal{O}(L^2)$ , where $L$ denotes the sequence length. When incorporating both motion and text, the sequence length becomes $L = L_{T} + T$ , where $L_{T}$ is the number of text tokens and $T$ is the number of motion frames. As $T$ increases, the quadratic growth in computational cost becomes prohibitive, leading to complexity proportional to $(L_{T} + T)^{2}$ .", + "bbox": [ + 511, + 534, + 906, + 733 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Furthermore, computing attention over long sequences can dilute attention weights, making it challenging for the model to focus on essential parts of the sequence and effectively capture long-range dependencies. Furthermore, the motion features are not well aligned with text either. To address these issues, we propose the Cross Talker Module, which selectively distills key motion frames—referred to as viewpoint frames—based on their relevance to the text context, as well as aligning the motion features with the text. An overview of the Cross Talker Module is shown in Figure 3.", + "bbox": [ + 511, + 734, + 908, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "17584", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/48347167cdb846970fabb12d7a345502e93e48aaca805da0b7e77a0124f7d964.jpg", + "image_caption": [ + "Figure 2. Overview of the LLaMo framework. It includes three main modules: (1) Multimodal Feature Extraction for encoding video and motion data; (2) Cross Talker for aligning and fusing motion and text features; and (3) Behavior Generation Module to produce text descriptions of human behavior based on integrated features." + ], + "image_footnote": [], + "bbox": [ + 93, + 85, + 869, + 367 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f9a416491ad6f8c391feccd39022f6b68db29765534857a3c9a8afc333aea68e.jpg", + "image_caption": [ + "Figure 3. Overview of the Cross Talker Module, which selects key frames based on text guidance and fuses them with text features for enhanced analysis." + ], + "image_footnote": [], + "bbox": [ + 94, + 448, + 480, + 741 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Language-Guided Frame Selection Given the enhanced motion features $\\tilde{F}_M\\in \\mathbb{R}^{T\\times H}$ and text embeddings $F_{T}\\in$ $\\mathbb{R}^{L_T\\times H}$ , we aim to identify $K$ viewpoint frames that are most relevant to the textual context. This selection is achieved through a cross-attention operation, where text", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "features act as queries and motion features as keys and values:", + "bbox": [ + 511, + 448, + 906, + 474 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nA = \\operatorname {S o f t m a x} \\left(\\frac {F _ {T} W _ {Q} \\left(\\tilde {F} _ {M} W _ {K}\\right) ^ {\\top}}{\\sqrt {d}}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 474, + 906, + 516 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $W_{Q}, W_{K} \\in \\mathbb{R}^{H \\times d}$ are learnable projection matrices, and $d$ is the dimensionality of the queries and keys. The resulting attention matrix $A \\in \\mathbb{R}^{L_{T} \\times T}$ contains attention weights, with each element $A_{i,j}$ indicating the relevance between the $i$ -th text token and the $j$ -th motion frame.", + "bbox": [ + 511, + 520, + 906, + 595 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To determine the importance of each motion frame, we aggregate attention weights across all text tokens using max pooling:", + "bbox": [ + 511, + 597, + 906, + 642 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns _ {j} = \\max _ {i = 1, \\dots , L _ {T}} A _ {i, j}, \\quad j = 1, \\dots , T. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 652, + 906, + 676 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The resulting scores $s_j$ represent the significance of each frame with respect to the text context. We then select the top $K$ frames with the highest relevance scores as viewpoint frames, reducing the effective sequence length from $T$ to $K$ . This reduction in sequence length lowers the self-attention complexity from $\\mathcal{O}((L_T + T)^2)$ to $\\mathcal{O}((L_T + K)^2)$ , which is computationally efficient when $K \\ll T$ .", + "bbox": [ + 511, + 686, + 908, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Adaptive Contextual Feature Aggregation For each selected viewpoint frame, we enrich its representation by aggregating both local and global contexts. An adaptive receptive field size $r_k$ is predicted for each viewpoint frame $k$ using a receptive field regression module. The module performs a cross-attention operation where each viewpoint", + "bbox": [ + 511, + 809, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "17585", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "frame serves as a query, and the unselected motion features serve as keys and values, followed by a sigmoid activation to estimate $r_k$ . Based on $r_k$ , we define a local window $W_k$ around frame $k$ :", + "bbox": [ + 89, + 90, + 483, + 151 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nW _ {k} = \\{j \\mid | j - k | \\leq r _ {k} \\times T \\}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 183, + 166, + 482, + 184 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Within this window, we apply local attention to capture fine-grained details:", + "bbox": [ + 89, + 199, + 483, + 231 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {\\text {l o c a l}} (k) = \\tilde {F} _ {M} (k) + \\text {A t t e n t i o n} (\\tilde {F} _ {M} (k), \\tilde {F} _ {M} \\left(W _ {k}\\right), \\tilde {F} _ {M} \\left(W _ {k}\\right)), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 243, + 491, + 275 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\tilde{F}_M(W_k)$ denotes the motion features within the local window around frame $k$ .", + "bbox": [ + 89, + 275, + 483, + 306 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To incorporate global context, we partition the full motion sequence into $N$ segments and compute segment-level features via average pooling:", + "bbox": [ + 89, + 308, + 483, + 354 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {M} ^ {\\mathrm {s e g}} = \\left\\{\\bar {F} _ {M} (n) \\right\\} _ {n = 1} ^ {N}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 212, + 367, + 482, + 386 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the segment size is controlled by the hyperparameter $S_{n}$ . We then apply global attention:", + "bbox": [ + 89, + 400, + 483, + 431 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {\\text {g l o b a l}} (k) = F _ {\\text {l o c a l}} (k) + \\text {A t t e n t i o n} \\left(F _ {\\text {l o c a l}} (k), F _ {M} ^ {\\text {s e g}}, F _ {M} ^ {\\text {s e g}}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 444, + 482, + 476 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The final representation for each viewpoint frame is obtained by concatenating local and global features:", + "bbox": [ + 89, + 477, + 483, + 506 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {M} (k) = \\left[ F _ {\\text {l o c a l}} (k); F _ {\\text {g l o b a l}} (k) \\right] \\in \\mathbb {R} ^ {2 H}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 155, + 520, + 482, + 539 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "allowing the model to capture both fine-grained motion details and broader contextual information.", + "bbox": [ + 89, + 553, + 483, + 582 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Bidirectional Cross-Modal Fusion To integrate the enhanced motion features with the text embeddings, we employ a bidirectional cross-attention mechanism. First, we update the motion features by performing a cross-attention operation, where the motion features act as queries and the text features serve as keys and values, and vice versa for the text features. The updated features are then passed through feed-forward networks, and we concatenate the motion and text representations to form the input for the next stage:", + "bbox": [ + 89, + 609, + 483, + 746 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {\\text {f u s i o n}} = \\left[ F _ {T}; \\{F _ {M} (k) \\} _ {k \\in K} \\right]. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 761, + 482, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This bidirectional fusion enhances the interaction between motion and text, enabling a coherent, contextually grounded understanding of human behavior. By selectively focusing on key motion frames and integrating enriched motion and text features, the Cross Talker Module reduces computational costs and improves the model's ability to capture complex human actions.", + "bbox": [ + 89, + 795, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Behavior Generation Module", + "text_level": 1, + "bbox": [ + 513, + 90, + 774, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Behavior Generation Module leverages the fused features $F_{\\mathrm{fusion}}$ to generate a textual description $Y$ that encapsulates the observed human behavior. This module utilizes a language model $h(\\cdot)$ to transform $F_{\\mathrm{fusion}}$ into a semantically coherent and contextually relevant output sequence:", + "bbox": [ + 511, + 113, + 906, + 189 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nY = h \\left(F _ {\\text {f u s i o n}}\\right) = \\left\\{y _ {1}, y _ {2}, \\dots , y _ {L} \\right\\}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 199, + 906, + 217 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $y_{t}$ denotes the token generated at the $t$ -th time step, and $L$ is the length of the generated description. The probability of generating each token $y_{t}$ is given by:", + "bbox": [ + 511, + 226, + 905, + 272 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np \\left(y _ {t} \\mid y _ {1: t - 1}, F _ {\\text {f u s i o n}}\\right) = \\operatorname {S o f t m a x} \\left(W _ {o} h _ {t}\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 281, + 906, + 297 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $h_t$ represents the hidden state of the language model at time $t$ , and $W_o$ is an output projection matrix. The hidden state $h_t$ is derived from the model's internal mechanisms, which integrate attention over the input features and previously generated tokens. This autoregressive approach enables LLaMo to produce detailed and contextually aligned descriptions that incorporate nuanced information from both motion and text modalities.", + "bbox": [ + 511, + 306, + 905, + 428 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By iteratively generating tokens based on the fused features and preceding context, the Behavior Generation Module constructs a meaningful and comprehensive narrative of human activity, grounded in the multimodal input data.", + "bbox": [ + 511, + 429, + 906, + 489 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Training Objective", + "text_level": 1, + "bbox": [ + 513, + 497, + 694, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To train LLaMo, we finetune a pretrained language model using supervised learning, guiding it to generate accurate and contextually relevant descriptions based on the input motion, video, and text features. Both the motion and video encoders remain frozen during training, focusing the optimization on the language model's alignment with multimodal inputs. The training objective minimizes the discrepancy between the generated descriptions and the groundtruth annotations in the dataset.", + "bbox": [ + 511, + 520, + 905, + 655 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given a training set of $N$ samples, each with video frames $V^{(i)}$ , motion data $M^{(i)}$ , and corresponding ground-truth textual descriptions $\\hat{Y}^{(i)} = \\{\\hat{y}_1^{(i)},\\hat{y}_2^{(i)},\\dots ,\\hat{y}_L^{(i)}\\}$ , the overall training objective is to minimize the negative log-likelihood of the ground-truth tokens across all samples:", + "bbox": [ + 511, + 657, + 905, + 736 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\sum_ {t = 1} ^ {L ^ {(i)}} \\log p \\left(\\hat {y} _ {t} ^ {(i)} \\mid \\hat {y} _ {1: t - 1} ^ {(i)}, F _ {\\text {f u s i o n}} ^ {(i)}\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 744, + 906, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $F_{\\mathrm{fusion}}^{(i)}$ represents the fused input features for the $i$ -th sample. The term $p(\\hat{y}_t^{(i)} \\mid \\hat{y}_{1:t-1}^{(i)}, F_{\\mathrm{fusion}}^{(i)})$ denotes the probability of the ground-truth token $\\hat{y}_t^{(i)}$ given the previous tokens $\\hat{y}_{1:t-1}^{(i)}$ and the fused features.", + "bbox": [ + 511, + 797, + 905, + 871 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This objective ensures that the generated descriptions are syntactically and semantically aligned with the target", + "bbox": [ + 511, + 871, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "17586", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "outputs, capturing the detailed aspects of human behavior present in the video and motion inputs.", + "bbox": [ + 89, + 90, + 485, + 122 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 135, + 225, + 152 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To evaluate LLaMo, we performed experiments across datasets representing complex human behaviors and tasks, including professional sports and cyclic action repetition counts.", + "bbox": [ + 89, + 161, + 483, + 220 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 231, + 308, + 247 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training Datasets: LLaMo was trained using a combination of bimodal video-motion datasets and unimodal video or motion datasets to enable a robust understanding of both modalities. For video-motion data, we utilized MoVid from MotionLLM [1] and our custom Swing dataset containing 20,000 annotated videos of baseball and golf swings with corresponding motion capture data and expert Q&A instructions. Motion-only datasets included HumanML3D [37] and KIT-ML [3], while Mo-RepCount, a filtered subset of RepCount [38], provided high-quality repetition count videos. This multimodal data configuration ensures that LLaMo effectively learns from both single and combined modalities.", + "bbox": [ + 89, + 255, + 483, + 450 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Datasets: LLaMo was evaluated on MoVid-Bench [1] to assess both video and motion understanding capabilities, with additional tests on BABEL-QA [19] to evaluate motion-based question answering. For specific tasks, we used Mo-RepCount to gauge motion details and Swing to evaluate professional sports guidance.", + "bbox": [ + 89, + 470, + 483, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics: For MoVid-Bench, we follow previous LLM evaluation metrics [39] [22] [40] on accuracy and scores, which contains body-part awareness, sequentially, direction analysis, reasoning ability, and hallucination, respectively. BABEL-QA used the metric followed by [19], Mo-RepCount applied metrics such as OBO, MAE, OBZ, and RMSE, and Swing performance was assessed on reasonableness, coherence, pertinence, and adaptability using GPT-4 evaluations, scoring from 0 to 5.", + "bbox": [ + 89, + 580, + 483, + 717 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Results", + "text_level": 1, + "bbox": [ + 89, + 726, + 184, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation on Motion Understanding in MoVid-Bench. We evaluated LLaMo's motion understanding capabilities on the MoVid-Bench, focusing on five critical aspects. Both accuracy and score metrics were used to comprehensively assess LLaMo's performance. Table 1 provides a comparison with several baselines, including GPT-3.5 [41], MotionGPT [2], and MotionLLM [1]. The GPT-3.5 baseline, limited to textual data processing, struggles significantly with motion interpretation, yielding low performance. MotionGPT, while capable of incorporating motion data, is pri", + "bbox": [ + 89, + 750, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "marily code motion into a codebook, which limits its reasoning ability on coiled motion data and leaves it vulnerable to hallucinations. Although MotionLLM provides improvements in understanding motion sequences, it still conducts translation on motion data, resulting in low accuracy on seq. and Hall.", + "bbox": [ + 511, + 90, + 903, + 180 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "LLaMo, in contrast, achieves superior results across nearly all metrics on MoVid-Bench-Motion, attaining the highest accuracy in most categories. This performance is primarily attributed to its preservation of critical motion-specific nuances, as LLaMo processes motion data as a distinct modality without requiring any conversion. LLaMo effectively captures the intricate details of 3D sequences associated with human behaviors. Its Feature Enhancer, combined with the Cross Talker module, enables simultaneous input of video and motion data, allowing LLaMo to selectively focus on motion frames most relevant to the text context. This design enhances the model's ability to capture complex patterns of human motion and improves its reasoning on motion-intensive tasks. Overall, LLaMo's results on MoVid-Bench demonstrate its effectiveness in advancing human behavior analysis, highlighting its potential for broader applications in general motion understanding.", + "bbox": [ + 511, + 181, + 906, + 439 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation on Video Understanding in MoVid-Bench.", + "text_level": 1, + "bbox": [ + 511, + 458, + 906, + 472 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluated LLaMo on the video component of the MoVid-Bench dataset to assess its capability for comprehensive video understanding. As presented in Table 1, LLaMo achieves strong performance across most metrics, demonstrating LLaMo's extraordinary comprehension ability on videos even without motion input.", + "bbox": [ + 511, + 473, + 903, + 563 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Compared to other models like GPT-3.5 [41], VideoLLAVA [39], and MotionLLM [1], while LLaMo slightly trails behind MotionLLM in body-part awareness, LLaMo shows significant improvements in reasoning ability and overall scores. This indicates that our model excels in understanding complex video content and making accurate inferences about human behaviors.", + "bbox": [ + 511, + 564, + 905, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The rewarding results can be attributed to our model's powerful ability to estimate motion data accurately in video and align it seamlessly with video inputs. This approach retains the motion capture of motion data and the auxiliary information provided by video, enriching the understanding of dynamic scenes.", + "bbox": [ + 511, + 669, + 906, + 760 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation on BABEL-QA. We evaluated LLaMo on the BABEL-QA dataset to assess its capability to handle complex motion-based queries. As shown in Table 2, LLaMo achieves the highest overall score of 0.458, outperforming all baselines. It excels particularly in the Action category with a score of 0.525, demonstrating superior action recognition capabilities. Additionally, it achieves top scores in temporal reasoning tasks (Before and After), indicating a", + "bbox": [ + 511, + 780, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "17587", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/21e8b7bec3a137a35036556a5f4d6ca28740a73bdf86b7affb8eef808260b72a.jpg", + "image_caption": [ + "e e" + ], + "image_footnote": [], + "bbox": [ + 117, + 92, + 176, + 162 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/308b04019fbdaaf485a6f47cea714dc2fa91b12e073a31bfdfe18fbb56dfcf21.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 187, + 90, + 238, + 162 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e8387f7b304d4d04ab54c316bf28e4953df442d47b1c1d40616b51882572d02d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 250, + 92, + 297, + 162 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/772529eca92b7d5d18040a0ad6aa554a2d80c39bb16f4357cb90dd7ea3133dd3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 310, + 92, + 356, + 162 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/68edf55cc48e16ae54ce913a517b4a80c72688b34871bce62471906dd2472e0a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 92, + 408, + 161 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0ae195acc9f7b15ff2081792485f2d8dd6c7729044f9d78c2d78f423d8806a97.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 421, + 92, + 468, + 161 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1f3f8bf582d6a4f5651e5036f0451e9f4f97f94e301ad6b8f30ea25f7d38c92d.jpg", + "image_caption": [ + "#" + ], + "image_footnote": [], + "bbox": [ + 478, + 92, + 527, + 161 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Describe the following human motion video concisely, including but not limited to what actions are being performed, the standards followed, and the characteristics.", + "bbox": [ + 558, + 95, + 872, + 128 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The person is practicing a Tai Chi movement called \"Golden Rooster Stands on One Leg.\" The movement starts from a wide stance with extended arms and gradually transitions into a single-leg balance, with the other knee lifted and the hands drawn close to the body. The entire process is slow and controlled, with smooth and continuous motions, emphasizing the integration of mind and body, showcasing the characteristics of Tai Chi.", + "bbox": [ + 558, + 137, + 875, + 227 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0d97bb1db2bca601221d1e2abf22ae3355e5c8eb1d3821a0c3a3039283d7e8e6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 101, + 164, + 339, + 234 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a816d961af460a53aacb302c19591bfd13830226e517b2bf8e269ec831b3c6ce.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 339, + 167, + 527, + 232 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a2dec7880660cc958862edae3a8d3bef710de30643fcf191b216786537369b73.jpg", + "image_caption": [ + "Figure 4. Example outputs from LLaMo across human activities and professional sports, showcasing its reasoning capabilities and domain-specific knowledge in motion-intensive scenarios." + ], + "image_footnote": [], + "bbox": [ + 117, + 236, + 339, + 297 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/56370119188f0947e59157efbc6062889103e27afd4e5f7da97d58831cfa36bf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 237, + 633, + 325 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2ca5dbd12e367ff59fa0c4951037e02d07f6db1f0315be09a70929c3c78e2794.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 234, + 862, + 295 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "What is the person in the video doing? The person in the video is making a phone", + "Can you analyze her current emotions? Sure, she seems a little impatient and angry", + "Why do you think she's angry? Could it be that she's feeling surprised instead?" + ], + "bbox": [ + 96, + 300, + 352, + 386 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Determining whether a person is angry can be based on their body language. She is likely feeling angry because she stopped and stomped her foot while walking, and her body movements became somewhat sluggish after hanging up the phone.", + "bbox": [ + 96, + 390, + 361, + 446 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "How many push-ups did the person do in the video?", + "bbox": [ + 364, + 327, + 614, + 351 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The person in the video did three push-ups. Is it definitely 3 times? Why do I remember it as 5 times?", + "bbox": [ + 366, + 359, + 627, + 395 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Let me observe carefully again. It's likely that the smaller range of motion led to your misjudgment. Indeed, he did three push-ups, with three sets of repeated movements.", + "bbox": [ + 364, + 401, + 630, + 446 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "What should be the position of the knees in golf technique?", + "bbox": [ + 635, + 297, + 900, + 320 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "GT: You want to keep your knees a little wider and more stable in golf technique.", + "bbox": [ + 635, + 325, + 764, + 369 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "LLaMo: Keep knees slightly bent and stable, with the left knee moving forward during the downswing.", + "bbox": [ + 771, + 316, + 900, + 372 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "What joint angles are important at the top of the swing for elite golfers?", + "bbox": [ + 635, + 375, + 900, + 397 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "GT: Most elite golfers have joint angles ranging from 35 to 45 degrees at the top of the swing.", + "bbox": [ + 635, + 401, + 764, + 445 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "LLaMo: At the top of the swing, elite golfers have hips at 45 degrees and wrists at 90 degrees.", + "bbox": [ + 771, + 401, + 900, + 445 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b199ec79aa272c4e1f61227f8e1c3059d63b7ba695aa62b4aabca667acb136c2.jpg", + "table_caption": [ + "Table 1. Expected Comparison on the MoVid-Bench. The top part of the table presents motion-related results, and the bottom part presents video-related results. Higher accuracy and score values indicate better performance." + ], + "table_footnote": [], + "table_body": "
MoVid-Bench-MotionBody.Seq.Dir.Rea.Hall.All
Acc.ScoreAcc.ScoreAcc.ScoreAcc.ScoreAcc.ScoreAcc.Score
GT100.005.00100.005.00100.005.00100.005.00100.005.00100.005.00
GPT-3.5 [41]24.512.0430.412.2527.142.1939.192.6458.333.2231.332.31
MotionGPT [2]31.223.9842.693.1644.293.5035.813.0616.662.2536.863.11
MotionLLM [1]50.493.5536.843.1458.573.7652.703.5855.563.3949.503.49
LLaMo59.304.0144.013.1260.913.9958.213.6461.173.5355.323.67
MoVid-Bench-VideoBody.Seq.Dir.Rea.Hull.All
Acc.ScoreAcc.ScoreAcc.ScoreAcc.ScoreAcc.ScoreAcc.Score
GT100.005.00100.005.00100.005.00100.005.00100.005,00100.005.00
GPT-3.5 [41]2.401.231.391.004.651.095.411.650.000.943.031.26
Video-LLAVA [39]33.532.7625.462.7241.862.8452.973.2858.831.8942.532.70
MotionLLM [1]34.132.9332.872.9244.183.1463.203.5570.592.3049.002.97
LLaMo33.832.8536.013.1145.503.3267.593.7372.812.2552.333.10
", + "bbox": [ + 94, + 542, + 901, + 718 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "strong understanding of temporal relationships in motion sequences. This capability benefits from the direct integration of motion data as an independent modality and the effective text-aware motion frames extract mechanism in LLaMo, preserving crucial motion-specific nuances, as well as leading to more accurate and contextually rich interpretations.", + "bbox": [ + 89, + 743, + 482, + 847 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Professional Sports Analysis. We further evaluate LLaMo on the challenging swing-dataset, termed as", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "profession-swing, details refer to Appendix, where the ground truth (GT) consists of answers provided by professional coaches, serving as expert benchmarks. The evaluation focuses on four key indicators: Reasonableness: Logical soundness and plausibility of the responses. Coherence: Consistency and logical flow within the responses. Pertinence: Relevance of the answers to the specific questions. Adaptability: Ability to adjust responses based on the stage on the athlete.", + "bbox": [ + 511, + 743, + 906, + 878 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As presented in Table 3, LLaMo significantly outper", + "bbox": [ + 532, + 885, + 903, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "17588", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/28869079c851c28fe262b722eba7d2e8274659d3839d8aab34ad9d147233fc6e.jpg", + "table_caption": [ + "Table 2. Comparison on BABEL-QA dataset. Higher scores indicate better performance." + ], + "table_footnote": [], + "table_body": "
ModelPred. typeOverall ↑Action ↑Direction ↑Body Part ↑Before ↑After ↑Other ↑
MotionCLIP-M [41]cls.0.4300.4850.3610.2720.3720.3210.404
MotionCLIP-R [41]cls.0.4200.4890.3100.2500.3980.3140.387
MotionLLM [41]gen.0.4360.5170.3540.1540.4270.3680.529
LLaMogen.0.4580.5250.3980.2240.4430.3920.518
", + "bbox": [ + 98, + 114, + 897, + 193 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2cf2ce3eedc5510a29a3572ed6c723a272e76fa6beaacb5fa13a1fb5864efb04.jpg", + "table_caption": [ + "Table 3. Performance on the profession-swing dataset across four indicators. Higher accuracy and score values indicate better performance." + ], + "table_footnote": [], + "table_body": "
ModelReasonablenessCoherencePertinenceAdaptabilityAll
AccScoreAccScoreAccScoreAccScoreAccScore
GT10051005100510051005
GPT-3.5 [41]4.510.834.200.743.650.661.710.223.420.61
MotionGPT [2]10.531.2515.421.6612.641.114.791.3514.351.4
MotionLLM [1]12.331.5119.201.8717.981.7910.201.2216.531.57
LLaMo21.102.1127.102.7131.813.1220.221.9824.802.48
", + "bbox": [ + 98, + 232, + 898, + 333 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "forms both GPT-3.5 [41] and MotionLLM [1] across all metrics. While MotionLLM shows improvements over GPT-3.5, with higher accuracy and scores, it still lags behind LLaMo. Specifically, LLaMo achieves an overall accuracy of 24.80", + "bbox": [ + 88, + 357, + 480, + 434 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The superior performance of LLaMo can be attributed to the design of the Cross Talker module, which enables language-guided frame selection to capture complex motion relationships inherent in professional swings. This allows LLaMo to effectively model intricate motion patterns and generate responses that are more reasonable, coherent, pertinent, and adaptable, closely aligning with the expert-level insights provided by professional coaches.", + "bbox": [ + 88, + 434, + 482, + 554 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In contrast, while MotionLLM demonstrates certain strengths over MotionGPT—such as better pertinence and coherence. It lacks the advanced motion understanding capabilities of LLaMo, lagging behind LLaMo distinctly.", + "bbox": [ + 89, + 555, + 482, + 617 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/142cd9266401a1132524d38c1689d810fd9f41fb7cb2dc60371874d504f65865.jpg", + "table_caption": [ + "Table 4. Motion and video details capture evaluation on MoRepCount" + ], + "table_footnote": [], + "table_body": "
ModelOBOMAEOBZRMSE
EScounts [42]0.3970.2910.1985.58
PoseRAC [43]0.3820.3120.2045.95
TransRAC [38]0.2760.4440.1058.56
RepNet [44]0.009///
LLaMo0.3890.3240.2226.15
", + "bbox": [ + 112, + 667, + 457, + 776 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Evaluation on Mo-RepCount. We evaluated LLaMo on the Mo-RepCount dataset to assess its ability to capture fine-grained motion details and complex temporal features required for accurate repetition counting. For a fair comparison, we trained all state-of-the-art models, including RepNet [44], TransRAC [38], PoseRAC [43] and ES", + "bbox": [ + 89, + 809, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "counts [42], on the Mo-RepCount dataset without using any additional data. As presented in Table 4, LLaMo outperforms the RepNet and TransRAC across almost all metrics. Since both of them are models built for loop counting tasks, the effect of our model on this task is convincing, which shows that LLaMo is also competent in capturing the details of the motion and videos.", + "bbox": [ + 511, + 357, + 903, + 464 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While LLaMo did not outperform the specialized counting models EScounts and poseRAC on many metrics, it excelled on the OBZ metric, demonstrating an unmatched ability to capture fine-grained motion details and spatiotemporal relationships. These results confirm LLaMo's strong potential in motion and video analysis, underscoring its suitability for detail-oriented tasks and applications.", + "bbox": [ + 511, + 467, + 906, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 599, + 633, + 614 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce LLaMo, a novel framework that integrates motion data directly into large language models, thereby enhancing multimodal understanding. By unifying text, video, and motion data in an architecture fine-tuned for human motion instruction, LLaMo achieves notable progress in behavior comprehension, motion-focused captioning, and interactive Q&A, capturing the fine-grained motion details vital for precise, context-aware insights. Experimental results, especially in professional sports analysis, demonstrate LLaMo's capacity to interpret complex human motion with accuracy. Future work will extend LLaMo to a broad range of domains where detailed motion analysis and multimodal understanding are essential. The main challenges lie in refining cross-modal integration techniques and improving efficiency for real-time applications, thus broadening LLaMo's potential in human-centered AI. We hope our work will spur further exploration into more advanced human-centric multimodal models.", + "bbox": [ + 511, + 628, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "17589", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by the Pioneer Centre for AI, DNRF grant number P1.", + "bbox": [ + 91, + 114, + 483, + 145 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 157, + 187, + 172 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Ling-Hao Chen, Shunlin Lu, Ailing Zeng, Hao Zhang, Benyou Wang, Ruimao Zhang, and Lei Zhang. Motionllm: Understanding human behaviors from human motions and videos. arXiv preprint arXiv:2405.20340, 2024. 1, 2, 6, 7, 8", + "[2] Ling-Hao Chen, Jiawei Zhang, Wen Liu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. In Proceedings of the Conference on Neural Information Processing Systems (NeurIPS), 2024. 1, 2, 6, 7, 8", + "[3] Matthias Plappert, Christian Mandery, and Tamim Asfour. The kit motion-language dataset. Big data, 4(4):236-252, 2016. 1, 6", + "[4] Yan Zhang, Michael J Black, and Siyu Tang. We are more than our joints: Predicting how 3d bodies move. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021.", + "[5] Fangzhou Hong, Liang Pan, Zhongang Cai, and Ziwei Liu. Versatile multi-modal pre-training for human-centric perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16156-16166, 2022. 2", + "[6] Haoxuan Qu, Yujun Cai, and Jun Liu. Llms are good action recognizers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18395-18406, 2024. 1", + "[7] Shan Chen, Jiale Zhou, and Lei Li. Dense point clouds matter: Dust-gs for scene reconstruction from sparse viewpoints. arXiv preprint arXiv:2409.08613, 2024. 1", + "[8] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 1, 2", + "[9] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 2", + "[10] Li Lei, Jia Sen, Wang Jianhao, An Zhaochong, Li Jiaang, Hwang Jenq-Neng, and Belongie Serge. Chatmotion: A multimodal multi-agent for human motion analysis. arXiv preprint arXiv:2502.18180, 2025. 1", + "[11] Mark Endo, Joy Hsu, Jiaman Li, and Jiajun Wu. Motion question answering via modular motion programs. In Proceedings of the International Conference on Machine Learning (ICML), 2023. 1, 2", + "[12] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 1", + "[13] Yaya Shi, Haiyang Xu, Chunfeng Yuan, Bing Li, Weiming Hu, and Zheng-Jun Zha. Learning video-text aligned repre" + ], + "bbox": [ + 93, + 183, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "sentations for video captioning. IEEE Transactions on Multimedia (TMM), 2023. 2", + "[14] Chuan Guo, Xinxin Zuo, Sen Wang, and Li Cheng. Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In Proceedings of the European Conference on Computer Vision (ECCV), pages 580-597, 2022. 2", + "[15] Kunchang Li, Yali Wang, Yinan He, et al. Mvbench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2", + "[16] Zhaochong An, Guolei Sun, Yun Liu, Runjia Li, Min Wu, Ming-Ming Cheng, Ender Konukoglu, and Serge Belongie. Multimodality helps few-shot 3d point cloud semantic segmentation. arXiv preprint arXiv:2410.22489, 2024. 2", + "[17] Bobby Bodenheimer, Chuck Rose, Seth Rosenthal, and John Pella. The process of motion capture: Dealing with the data. In Proceedings of the Eurographics Workshop on Computer Animation and Simulation (EG Workshop), pages 3-18, 1997. 2", + "[18] Xiangyu Peng, Yuxuan Zhang, Xiaolong Wang, and Ziwei Liu. Text-driven human motion generation with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2", + "[19] Mark Endo, Joy Hsu, Jiaman Li, and Jiajun Wu. Motion question answering via modular motion programs. In International Conference on Machine Learning, pages 9312-9328. PMLR, 2023. 2, 6", + "[20] Jie Lin, Wei Zhang, and Zhe Chen. Videollm: Language models for video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2405-2415, 2023. 2", + "[21] Xu Ning, Hongwei Li, and Yu Zhao. Videobench: A benchmark for large-scale video understanding. arXiv preprint arXiv:2304.12345, 2023. 2", + "[22] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 2, 6", + "[23] Lei Li. Segment any building. In Computer Graphics International Conference, pages 155-166. Springer, 2023. 2", + "[24] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 2", + "[25] Enxin Song, Guanhong Zhang, and Haoyang Zhou. Adaptive multimodal learning for behavior analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2", + "[26] Yunhua Yang, Ziwang Zhao, and Yiming Xie. Recognizing human behaviors with skeletal data in llm-based frameworks. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023.", + "[27] Lei Li, Tianfang Zhang, Zhongfeng Kang, and Xikun Jiang. Mask-fpan: Semi-supervised face parsing in the wild with de-occlusion and uv gan. Computers & Graphics, 116:185–193, 2023. 2" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "17590", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Matthias Plappert, Christian Mandery, and Tamim Asfour. Learning a bidirectional mapping between human whole-body motion and natural language using deep recurrent neural networks. Robotics and Autonomous Systems (RAS), 109: 13-26, 2018. 2", + "[29] Lei Li. Cpseg: Finer-grained image semantic segmentation via chain-of-thought language prompting. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 513-522, 2024.", + "[30] Yunhua Yang, Liang Zhang, and Hui Li. Understanding human behaviors from skeletal data: A review of datasets and methods. arXiv preprint arXiv:2310.12998, 2023. 2", + "[31] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Two-stream adaptive graph convolutional networks for skeleton-based action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12026-12035, 2019. 2", + "[32] Enxin Song, Wenhao Chai, and Yucheng Zhang. Fine-grained spatial-temporal motion understanding in complex video environments. arXiv preprint arXiv:2310.08639, 2023. 2", + "[33] Zhaochong An, Guolei Sun, Zongwei Wu, Hao Tang, and Luc Van Gool. Temporal-aware hierarchical mask classification for video semantic segmentation. arXiv preprint arXiv:2309.08020, 2023. 2", + "[34] Xianghong Liu, Haoxuan Wang, Zhiwei Zhang, Huan Wu, Baoquan Chen, and Xiaoguang Han. Category-agnostic pose estimation for point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3", + "[35] Xiaolong He, Yi Zhang, Chen Chen, and Kyoung Mu Lee. Activitynet++: A large-scale benchmark for video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023.", + "[36] Lei Li. Hierarchical edge aware learning for 3d point cloud. In Computer Graphics International Conference, pages 81-92. Springer, 2023. 3", + "[37] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5152-5161, 2022. 6", + "[38] Huazhang Hu, Sixun Dong, Yiqun Zhao, Dongze Lian, Zhengxin Li, and Shenghua Gao. Transrac: Encoding multiscale temporal correlation with transformers for repetitive action counting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19013-19022, 2022. 6, 8", + "[39] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 6, 7", + "[40] Peng Jin, Ryuichi Takanobu, Wancai Zhang, Xiaochun Cao, and Li Yuan. Chat-univi: Unified visual representation empowers large language models with image and video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13700-13710, 2024. 6" + ], + "bbox": [ + 91, + 92, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] OpenAI. Gpt-3.5: Generative pre-trained transformer 3.5, 2023. https://platform.openai.com/docs/models/gpt-3-5.6,7,8", + "[42] Saptarshi Sinha, Alexandros Stergiou, and Dima Damen. Every shot counts: Using exemplars for repetition counting in videos. arXiv preprint arXiv:2403.18074, 2024. 8", + "[43] Ziyu Yao, Xuxin Cheng, and Yuexian Zou. Poserac: Pose saliency transformer for repetitive action counting. arXiv preprint arXiv:2303.08450, 2023. 8", + "[44] Bastian Wandt and Bodo Rosenhahn. Repnet: Weakly supervised training of an adversarial reprojection network for 3d human pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7782-7791, 2019. 8" + ], + "bbox": [ + 514, + 92, + 903, + 290 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "17591", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Human Motion Instruction Tuning/8c08c8b0-9aae-431a-b68e-699a5ddd5c32_model.json b/2025/Human Motion Instruction Tuning/8c08c8b0-9aae-431a-b68e-699a5ddd5c32_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8499bd3a52950115bfe242345d060a317ed2bd38 --- /dev/null +++ b/2025/Human Motion Instruction Tuning/8c08c8b0-9aae-431a-b68e-699a5ddd5c32_model.json @@ -0,0 +1,2431 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.322, + 0.131, + 0.677, + 0.154 + ], + "angle": 0, + "content": "Human Motion Instruction Tuning" + }, + { + "type": "text", + "bbox": [ + 0.22, + 0.192, + 0.777, + 0.245 + ], + "angle": 0, + "content": "Lei Li\\(^{1,2,\\ast,\\dagger}\\), Sen Jia\\(^{1,\\ast}\\), Jianhao Wang\\(^{3}\\), Zhongyu Jiang\\(^{1}\\), Feng Zhou\\(^{4}\\), Ju Dai\\(^{5}\\), Tianfang Zhang\\(^{6}\\), Zongkai Wu\\(^{1}\\), Jenq-Neng Hwang\\(^{1}\\)" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.266, + 0.327, + 0.283 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.299, + 0.485, + 0.572 + ], + "angle": 0, + "content": "This paper presents LLaMo (Large Language and Human Motion Assistant), a multimodal framework for human motion instruction tuning. In contrast to conventional instruction-tuning approaches that convert non-linguistic inputs, such as video or motion sequences, into language tokens, LLaMo retains motion in its native form for instruction tuning. This method preserves motion-specific details that are often diminished in tokenization, thereby improving the model's ability to interpret complex human behaviors. By processing both video and motion data alongside textual inputs, LLaMo enables a flexible, human-centric analysis. Experimental evaluations across high-complexity domains, including human behaviors and professional activities, indicate that LLaMo effectively captures domain-specific knowledge, enhancing comprehension and prediction in motion-intensive scenarios. We hope LLaMo offers a foundation for future multimodal AI systems with broad applications, from sports analytics to behavioral prediction." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.603, + 0.222, + 0.619 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.63, + 0.484, + 0.827 + ], + "angle": 0, + "content": "Understanding human motion is a central challenge in multimodal AI, impacting numerous fields such as digital human, human-computer interaction, sports analytics, healthcare, and virtual human modeling [3-6]. Motion data, which captures skeletal movements and body dynamics, provides a structured, appearance-invariant representation of human actions, concentrating on the essential movement patterns while excluding irrelevant visual details [7]. This data has emerged as a privacy-conscious alternative to visual inputs, offering finer control over action recognition and facilitating analysis in applications ranging from digital human avatars to behavioral monitoring in healthcare [8-10]. As digital human representations and motion-centric" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.264, + 0.909, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.481, + 0.907, + 0.553 + ], + "angle": 0, + "content": "Figure 1. A comparison of MotionLLM [1], MotionGPT [2], and LLaMo highlights LLaMo's motion-specific capabilities. Equipped with a Motion Enhancer and Cross Talker module to align motion and text, LLaMo supports both video and motion inputs, enabling text-aware, fine-grained motion analysis." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.581, + 0.907, + 0.642 + ], + "angle": 0, + "content": "applications continue to expand, developing models that can leverage the unique characteristics of motion data becomes increasingly crucial for achieving robust, context-aware behavior analysis." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.909, + 0.854 + ], + "angle": 0, + "content": "Recently, large language models (LLMs) have shown promise in motion analysis, leveraging their capacity to model temporal dependencies across multimodal data [1, 11]. State-of-the-art models, such as MotionGPT [12] and MotionLLM [1], treat motion data as a multimodal extension, encoding it as discrete tokens or translating it into language-like representations for processing by LLMs. These models have yielded notable results in motion understanding and recognition tasks, illustrating the adaptability of language models to non-linguistic data. However, their reliance on tokenization or textual conversion can limit the richness of the motion representation, abstracting away critical spatial-temporal details essential for high-fidelity human activity recognition." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Despite recent advancements, notable challenges remain in current approaches to motion analysis using LLMs. First, encoding motion data into language tokens, as implemented" + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.839, + 0.375, + 0.852 + ], + "angle": 0, + "content": "* These authors contributed equally to this work." + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.852, + 0.373, + 0.864 + ], + "angle": 0, + "content": "† Corresponding Author. (lilei@di.ku.dk)" + }, + { + "type": "page_footnote", + "bbox": [ + 0.093, + 0.864, + 0.483, + 0.901 + ], + "angle": 0, + "content": "1University of Washington 2University of Copenhagen 3Xi'an Jiaotong University 4North China University of Technology 5Peng Cheng Laboratory 6Tsinghua University" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.839, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17582" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.349 + ], + "angle": 0, + "content": "in models such as MotionGPT, relies on a quantization process that may obscure critical fine-grained motion information, potentially limiting the model's capacity to accurately capture detailed spatial-temporal dynamics [13, 14]. The common approach of translating motion data into text tokens further contributes to this issue, often resulting in the loss of essential motion-specific nuances that are critical for in-depth behavior interpretation [11, 15]. This translation process poses considerable challenges in understanding the intricate details of human movement, impeding the model's ability to achieve spatial and physical comprehension of human actions. Moreover, motion data inherently contains 3D structural information and insights into connected human behaviors, which are often difficult to fully capture when relying solely on language tokens, as this representation can overlook the underlying spatial dependencies crucial to modeling human motion accurately." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.352, + 0.483, + 0.577 + ], + "angle": 0, + "content": "In addition, existing methods frequently treat video and motion data as isolated modalities, overlooking the potential benefits of a complementary, integrated approach [16]. Processing these data types independently neglects the rich environmental context and interaction cues provided by video inputs, which are crucial for accurately interpreting complex human behaviors [5, 8]. This separation limits the capacity for a holistic understanding of human motion, particularly in scenarios where context, spatial relations, and dynamic interactions play essential roles in behavior analysis. Furthermore, video-based approaches, while potentially more informative, are computationally intensive, posing significant constraints for large-scale or real-time applications, which limits their feasibility in practical motion analysis systems." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.582, + 0.483, + 0.793 + ], + "angle": 0, + "content": "In response to these challenges, we propose LLaMo, a framework designed to integrate motion data as a distinct modality within LLMs without translating it into textual intermediaries. LLaMo incorporates a Motion Estimator and an Enhancer to facilitate the direct input of motion or/and video data, thereby offering a unified, human-centric analysis pipeline. The framework also introduces a Cross Talker module, which dynamically aligns the motion and text features, allowing for text-aware, fine-grained motion representations. This approach retains the original spatial-temporal characteristics of the motion data, providing a more precise understanding of human actions and enabling high-resolution behavior analysis in motion-intensive contexts [17, 18]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Through extensive evaluations on benchmarks such as MoVid-Bench [1] and BABEL-QA [19], we demonstrate that LLaMo achieves state-of-the-art performance in motion-centric human activity recognition [11, 15]. These results underscore LLaMo's potential to advance multimodal AI, providing a robust foundation for human behavior understanding. Our contributions can be summarized as" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.093, + 0.572, + 0.105 + ], + "angle": 0, + "content": "follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.108, + 0.905, + 0.152 + ], + "angle": 0, + "content": "- We introduce LLaMo, which treats motion data as an independent modality within LLMs, preserving essential motion-specific details for robust analysis." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.153, + 0.905, + 0.213 + ], + "angle": 0, + "content": "- We propose Cross Talker, a text-guided mechanism that dynamically focuses on and aggregates key motion frames features, optimizing computational efficiency and enabling the model to pick relevant motion features." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.213, + 0.905, + 0.288 + ], + "angle": 0, + "content": "- LLaMo's architecture supports the direct input of both raw motion and video data, providing a generalized framework for human-centric analysis across diverse applications, such as sports analytics, healthcare, and behavioral monitoring." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.108, + 0.905, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.304, + 0.655, + 0.319 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.329, + 0.887, + 0.345 + ], + "angle": 0, + "content": "2.1. Human-Centric Multimodal Representation" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.353, + 0.907, + 0.671 + ], + "angle": 0, + "content": "Multimodal representation learning is integral to human-centric analysis, especially for applications requiring spatial-temporal reasoning to decode complex behaviors [20-22]. Recent advances, such as Video-LLaVA, have made significant strides by embedding visual data from images and videos into a shared linguistic feature space, enabling sophisticated visual reasoning for behavioral analysis tasks [20, 23]. Despite these achievements, many current models are optimized primarily for static images or discrete video frames, limiting their effectiveness in sequential, dynamic scenarios where understanding progression and continuity in motion is crucial [9, 21, 24]. Privacy concerns further challenge the adoption of video-based methods in sensitive domains; thus, researchers are exploring motion data as a privacy-conscious alternative, enabling analysis focused on human actions without revealing identifiable visual information [25-27]. By combining visual and motion data, emerging multimodal frameworks hold promise for more comprehensive, privacy-aware human behavior analysis, leveraging the strengths of both modalities for adaptability and depth across applications." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.681, + 0.787, + 0.697 + ], + "angle": 0, + "content": "2.2. Human Motion Understanding" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.705, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Human motion understanding has traditionally relied on skeletal data represented as sequences of joint keypoints to capture movement while preserving user privacy [13, 28-30]. Early models like 2s-AGCN [31] and more recent transformer-based methods, such as MotionCLIP [2], have achieved success in tasks like activity recognition, captioning, and behavior analysis by mapping motion data to language tokens, effectively capturing structural aspects of movement. However, these methods often lack environmental context, which is crucial for nuanced interpretations in real-world applications, as similar movements can convey different meanings across varied scenarios [24, 32, 33]. To address this limitation, recent frameworks inte" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "17583" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.228 + ], + "angle": 0, + "content": "grate motion and visual data, allowing models to generalize across diverse, dynamic environments [34-36]. This fusion of skeletal data with video-provided context has shown promise in applications like sports and professional activity analysis, where contextual cues are essential. Additionally, skeletal-based data remains valuable in privacy-sensitive applications, such as healthcare and security, where it allows for robust analysis without compromising personal identity." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.24, + 0.19, + 0.256 + ], + "angle": 0, + "content": "3. Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.265, + 0.483, + 0.34 + ], + "angle": 0, + "content": "The LLaMo framework is composed of three primary modules to effectively process and integrate video, motion data and text. An overview of this pipeline is provided in Figure 2. The architecture comprises the following components:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.342, + 0.483, + 0.432 + ], + "angle": 0, + "content": "1. Multimodal Feature Extraction Module: This module independently encodes the video or motion inputs, which makes LLaMo is a general human motion assistant with video input or motion input. Then the Motion features are enhanced by distilling valuable representations in video." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.433, + 0.483, + 0.509 + ], + "angle": 0, + "content": "2. Cross Talker: Here, the enhanced motion features are aggregated in a text-guided manner and aligned with text within a shared semantic space, allowing the natural feature fusion between motion and text, which provides fine-grained inputs for LLM." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.509, + 0.483, + 0.57 + ], + "angle": 0, + "content": "3. Behavior Generation Module: Using the aggregated and text-aligned motion features, along with the text representations, this module generates a contextually aware textual description of the observed human behavior." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.342, + 0.483, + 0.57 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.578, + 0.476, + 0.593 + ], + "angle": 0, + "content": "3.1. Enhanced motion Feature Extraction Module" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.6, + 0.483, + 0.766 + ], + "angle": 0, + "content": "The enhanced motion feature extraction module consists of four primary components: the motion estimator, followed by the motion and video encoders, and the feature enhancement module. The motion estimator addresses scenarios where motion data may be unavailable by estimating motion information directly from video frames. The motion and video encoders then independently process the motion and video data, encoding them into their respective feature spaces. This approach allows the system to leverage complementary information from both modalities, creating a richer representation of human behavior." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.767, + 0.483, + 0.798 + ], + "angle": 0, + "content": "Given a sequence of video frames \\(V = \\{v_{1}, v_{2}, \\ldots, v_{T}\\}\\), the video encoder \\(f_{v}(\\cdot)\\) extracts a set of visual features:" + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.806, + 0.442, + 0.823 + ], + "angle": 0, + "content": "\\[\nF _ {V} = f _ {v} (V) = \\left\\{f _ {v} \\left(v _ {1}\\right), f _ {v} \\left(v _ {2}\\right), \\dots , f _ {v} \\left(v _ {T}\\right) \\right\\}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.831, + 0.483, + 0.877 + ], + "angle": 0, + "content": "Simultaneously, the motion encoder \\( f_{m}(\\cdot) \\) processes the motion data \\( M = \\{m_1, m_2, \\ldots, m_T\\} \\) to generate motion-specific features:" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.885, + 0.466, + 0.903 + ], + "angle": 0, + "content": "\\[\nF _ {M} = f _ {m} (M) = \\left\\{f _ {m} \\left(m _ {1}\\right), f _ {m} \\left(m _ {2}\\right), \\dots , f _ {m} \\left(m _ {T}\\right) \\right\\}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.09, + 0.905, + 0.152 + ], + "angle": 0, + "content": "Here, \\(F_{V}\\in \\mathbb{R}^{T\\times H}\\) and \\(F_{M}\\in \\mathbb{R}^{T\\times H}\\) represent the extracted feature sets for the video and motion data, respectively, where \\(T\\) is the sequence length and \\(H\\) denotes the feature dimension." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.155, + 0.907, + 0.307 + ], + "angle": 0, + "content": "The feature enhancement module then aligns and fuses these modality-specific features. By using the motion features \\( F_{M} \\) as queries, this module selectively extracts semantically relevant information from the video features \\( F_{V} \\). This alignment and fusion process allows the motion features to be enriched with contextual information from the video, producing refined, contextually enriched representations that enhance the model's understanding of nuanced human behaviors. For the detailed design refer to the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.309, + 0.909, + 0.492 + ], + "angle": 0, + "content": "First, the video features and motion features will conduct self-attention operations respectively to get their augmented features \\( F_V' \\) and \\( F_M' \\). This allows the video and motion to initially capture the dependencies on their respective time steps. Then, in order to enrich the motion data features with semantic information from the video, we subtly design a cross-attention block where augmented motion features as the query to extract valid semantic information in the video features. Finally, the selected video semantic representation and motion features are connected by residuals followed by a feed-forward neural network to obtain the enhanced motion features \\( \\tilde{F_M} \\)" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.509, + 0.714, + 0.525 + ], + "angle": 0, + "content": "3.2. Cross Talker Module" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.535, + 0.907, + 0.734 + ], + "angle": 0, + "content": "Given the enhanced motion features \\(\\tilde{F}_M\\) from the feature enhancement module, directly feeding the entire sequence of motion frames into the Large Language Model (LLM) incurs significant computational costs and may lead to suboptimal performance. This limitation arises from the self-attention mechanism in transformers, which has a computational complexity of \\(\\mathcal{O}(L^2)\\), where \\(L\\) denotes the sequence length. When incorporating both motion and text, the sequence length becomes \\(L = L_{T} + T\\), where \\(L_{T}\\) is the number of text tokens and \\(T\\) is the number of motion frames. As \\(T\\) increases, the quadratic growth in computational cost becomes prohibitive, leading to complexity proportional to \\((L_{T} + T)^{2}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Furthermore, computing attention over long sequences can dilute attention weights, making it challenging for the model to focus on essential parts of the sequence and effectively capture long-range dependencies. Furthermore, the motion features are not well aligned with text either. To address these issues, we propose the Cross Talker Module, which selectively distills key motion frames—referred to as viewpoint frames—based on their relevance to the text context, as well as aligning the motion features with the text. An overview of the Cross Talker Module is shown in Figure 3." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17584" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.087, + 0.87, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.379, + 0.908, + 0.424 + ], + "angle": 0, + "content": "Figure 2. Overview of the LLaMo framework. It includes three main modules: (1) Multimodal Feature Extraction for encoding video and motion data; (2) Cross Talker for aligning and fusing motion and text features; and (3) Behavior Generation Module to produce text descriptions of human behavior based on integrated features." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.449, + 0.481, + 0.742 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.754, + 0.483, + 0.797 + ], + "angle": 0, + "content": "Figure 3. Overview of the Cross Talker Module, which selects key frames based on text guidance and fuses them with text features for enhanced analysis." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Language-Guided Frame Selection Given the enhanced motion features \\(\\tilde{F}_M\\in \\mathbb{R}^{T\\times H}\\) and text embeddings \\(F_{T}\\in\\) \\(\\mathbb{R}^{L_T\\times H}\\), we aim to identify \\(K\\) viewpoint frames that are most relevant to the textual context. This selection is achieved through a cross-attention operation, where text" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.449, + 0.907, + 0.476 + ], + "angle": 0, + "content": "features act as queries and motion features as keys and values:" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.476, + 0.907, + 0.517 + ], + "angle": 0, + "content": "\\[\nA = \\operatorname {S o f t m a x} \\left(\\frac {F _ {T} W _ {Q} \\left(\\tilde {F} _ {M} W _ {K}\\right) ^ {\\top}}{\\sqrt {d}}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.521, + 0.907, + 0.597 + ], + "angle": 0, + "content": "where \\( W_{Q}, W_{K} \\in \\mathbb{R}^{H \\times d} \\) are learnable projection matrices, and \\( d \\) is the dimensionality of the queries and keys. The resulting attention matrix \\( A \\in \\mathbb{R}^{L_{T} \\times T} \\) contains attention weights, with each element \\( A_{i,j} \\) indicating the relevance between the \\( i \\)-th text token and the \\( j \\)-th motion frame." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.907, + 0.643 + ], + "angle": 0, + "content": "To determine the importance of each motion frame, we aggregate attention weights across all text tokens using max pooling:" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.654, + 0.907, + 0.677 + ], + "angle": 0, + "content": "\\[\ns _ {j} = \\max _ {i = 1, \\dots , L _ {T}} A _ {i, j}, \\quad j = 1, \\dots , T. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.687, + 0.909, + 0.793 + ], + "angle": 0, + "content": "The resulting scores \\( s_j \\) represent the significance of each frame with respect to the text context. We then select the top \\( K \\) frames with the highest relevance scores as viewpoint frames, reducing the effective sequence length from \\( T \\) to \\( K \\). This reduction in sequence length lowers the self-attention complexity from \\( \\mathcal{O}((L_T + T)^2) \\) to \\( \\mathcal{O}((L_T + K)^2) \\), which is computationally efficient when \\( K \\ll T \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Adaptive Contextual Feature Aggregation For each selected viewpoint frame, we enrich its representation by aggregating both local and global contexts. An adaptive receptive field size \\( r_k \\) is predicted for each viewpoint frame \\( k \\) using a receptive field regression module. The module performs a cross-attention operation where each viewpoint" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17585" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.152 + ], + "angle": 0, + "content": "frame serves as a query, and the unselected motion features serve as keys and values, followed by a sigmoid activation to estimate \\( r_k \\). Based on \\( r_k \\), we define a local window \\( W_k \\) around frame \\( k \\):" + }, + { + "type": "equation", + "bbox": [ + 0.184, + 0.167, + 0.483, + 0.185 + ], + "angle": 0, + "content": "\\[\nW _ {k} = \\{j \\mid | j - k | \\leq r _ {k} \\times T \\}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.2, + 0.484, + 0.232 + ], + "angle": 0, + "content": "Within this window, we apply local attention to capture fine-grained details:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.244, + 0.493, + 0.276 + ], + "angle": 0, + "content": "\\[\nF _ {\\text {l o c a l}} (k) = \\tilde {F} _ {M} (k) + \\text {A t t e n t i o n} (\\tilde {F} _ {M} (k), \\tilde {F} _ {M} \\left(W _ {k}\\right), \\tilde {F} _ {M} \\left(W _ {k}\\right)), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.276, + 0.484, + 0.307 + ], + "angle": 0, + "content": "where \\(\\tilde{F}_M(W_k)\\) denotes the motion features within the local window around frame \\(k\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.309, + 0.484, + 0.355 + ], + "angle": 0, + "content": "To incorporate global context, we partition the full motion sequence into \\(N\\) segments and compute segment-level features via average pooling:" + }, + { + "type": "equation", + "bbox": [ + 0.213, + 0.368, + 0.483, + 0.387 + ], + "angle": 0, + "content": "\\[\nF _ {M} ^ {\\mathrm {s e g}} = \\left\\{\\bar {F} _ {M} (n) \\right\\} _ {n = 1} ^ {N}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.401, + 0.484, + 0.432 + ], + "angle": 0, + "content": "where the segment size is controlled by the hyperparameter \\( S_{n} \\). We then apply global attention:" + }, + { + "type": "equation", + "bbox": [ + 0.103, + 0.445, + 0.483, + 0.477 + ], + "angle": 0, + "content": "\\[\nF _ {\\text {g l o b a l}} (k) = F _ {\\text {l o c a l}} (k) + \\text {A t t e n t i o n} \\left(F _ {\\text {l o c a l}} (k), F _ {M} ^ {\\text {s e g}}, F _ {M} ^ {\\text {s e g}}\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.478, + 0.484, + 0.507 + ], + "angle": 0, + "content": "The final representation for each viewpoint frame is obtained by concatenating local and global features:" + }, + { + "type": "equation", + "bbox": [ + 0.156, + 0.521, + 0.483, + 0.54 + ], + "angle": 0, + "content": "\\[\nF _ {M} (k) = \\left[ F _ {\\text {l o c a l}} (k); F _ {\\text {g l o b a l}} (k) \\right] \\in \\mathbb {R} ^ {2 H}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.554, + 0.484, + 0.583 + ], + "angle": 0, + "content": "allowing the model to capture both fine-grained motion details and broader contextual information." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.611, + 0.484, + 0.747 + ], + "angle": 0, + "content": "Bidirectional Cross-Modal Fusion To integrate the enhanced motion features with the text embeddings, we employ a bidirectional cross-attention mechanism. First, we update the motion features by performing a cross-attention operation, where the motion features act as queries and the text features serve as keys and values, and vice versa for the text features. The updated features are then passed through feed-forward networks, and we concatenate the motion and text representations to form the input for the next stage:" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.762, + 0.483, + 0.78 + ], + "angle": 0, + "content": "\\[\nF _ {\\text {f u s i o n}} = \\left[ F _ {T}; \\{F _ {M} (k) \\} _ {k \\in K} \\right]. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.903 + ], + "angle": 0, + "content": "This bidirectional fusion enhances the interaction between motion and text, enabling a coherent, contextually grounded understanding of human behavior. By selectively focusing on key motion frames and integrating enriched motion and text features, the Cross Talker Module reduces computational costs and improves the model's ability to capture complex human actions." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.776, + 0.107 + ], + "angle": 0, + "content": "3.3. Behavior Generation Module" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.114, + 0.907, + 0.19 + ], + "angle": 0, + "content": "The Behavior Generation Module leverages the fused features \\( F_{\\mathrm{fusion}} \\) to generate a textual description \\( Y \\) that encapsulates the observed human behavior. This module utilizes a language model \\( h(\\cdot) \\) to transform \\( F_{\\mathrm{fusion}} \\) into a semantically coherent and contextually relevant output sequence:" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.2, + 0.907, + 0.218 + ], + "angle": 0, + "content": "\\[\nY = h \\left(F _ {\\text {f u s i o n}}\\right) = \\left\\{y _ {1}, y _ {2}, \\dots , y _ {L} \\right\\}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.227, + 0.906, + 0.273 + ], + "angle": 0, + "content": "where \\( y_{t} \\) denotes the token generated at the \\( t \\)-th time step, and \\( L \\) is the length of the generated description. The probability of generating each token \\( y_{t} \\) is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.572, + 0.282, + 0.907, + 0.299 + ], + "angle": 0, + "content": "\\[\np \\left(y _ {t} \\mid y _ {1: t - 1}, F _ {\\text {f u s i o n}}\\right) = \\operatorname {S o f t m a x} \\left(W _ {o} h _ {t}\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.308, + 0.906, + 0.429 + ], + "angle": 0, + "content": "where \\( h_t \\) represents the hidden state of the language model at time \\( t \\), and \\( W_o \\) is an output projection matrix. The hidden state \\( h_t \\) is derived from the model's internal mechanisms, which integrate attention over the input features and previously generated tokens. This autoregressive approach enables LLaMo to produce detailed and contextually aligned descriptions that incorporate nuanced information from both motion and text modalities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.43, + 0.907, + 0.491 + ], + "angle": 0, + "content": "By iteratively generating tokens based on the fused features and preceding context, the Behavior Generation Module constructs a meaningful and comprehensive narrative of human activity, grounded in the multimodal input data." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.498, + 0.696, + 0.515 + ], + "angle": 0, + "content": "3.4. Training Objective" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.521, + 0.906, + 0.656 + ], + "angle": 0, + "content": "To train LLaMo, we finetune a pretrained language model using supervised learning, guiding it to generate accurate and contextually relevant descriptions based on the input motion, video, and text features. Both the motion and video encoders remain frozen during training, focusing the optimization on the language model's alignment with multimodal inputs. The training objective minimizes the discrepancy between the generated descriptions and the groundtruth annotations in the dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.658, + 0.906, + 0.737 + ], + "angle": 0, + "content": "Given a training set of \\(N\\) samples, each with video frames \\(V^{(i)}\\), motion data \\(M^{(i)}\\), and corresponding ground-truth textual descriptions \\(\\hat{Y}^{(i)} = \\{\\hat{y}_1^{(i)},\\hat{y}_2^{(i)},\\dots ,\\hat{y}_L^{(i)}\\}\\), the overall training objective is to minimize the negative log-likelihood of the ground-truth tokens across all samples:" + }, + { + "type": "equation", + "bbox": [ + 0.547, + 0.745, + 0.907, + 0.789 + ], + "angle": 0, + "content": "\\[\nL = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\sum_ {t = 1} ^ {L ^ {(i)}} \\log p \\left(\\hat {y} _ {t} ^ {(i)} \\mid \\hat {y} _ {1: t - 1} ^ {(i)}, F _ {\\text {f u s i o n}} ^ {(i)}\\right), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.799, + 0.906, + 0.872 + ], + "angle": 0, + "content": "where \\(F_{\\mathrm{fusion}}^{(i)}\\) represents the fused input features for the \\(i\\)-th sample. The term \\(p(\\hat{y}_t^{(i)} \\mid \\hat{y}_{1:t-1}^{(i)}, F_{\\mathrm{fusion}}^{(i)})\\) denotes the probability of the ground-truth token \\(\\hat{y}_t^{(i)}\\) given the previous tokens \\(\\hat{y}_{1:t-1}^{(i)}\\) and the fused features." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.872, + 0.907, + 0.902 + ], + "angle": 0, + "content": "This objective ensures that the generated descriptions are syntactically and semantically aligned with the target" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17586" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.123 + ], + "angle": 0, + "content": "outputs, capturing the detailed aspects of human behavior present in the video and motion inputs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.136, + 0.226, + 0.154 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.162, + 0.485, + 0.222 + ], + "angle": 0, + "content": "To evaluate LLaMo, we performed experiments across datasets representing complex human behaviors and tasks, including professional sports and cyclic action repetition counts." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.232, + 0.31, + 0.248 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.256, + 0.485, + 0.451 + ], + "angle": 0, + "content": "Training Datasets: LLaMo was trained using a combination of bimodal video-motion datasets and unimodal video or motion datasets to enable a robust understanding of both modalities. For video-motion data, we utilized MoVid from MotionLLM [1] and our custom Swing dataset containing 20,000 annotated videos of baseball and golf swings with corresponding motion capture data and expert Q&A instructions. Motion-only datasets included HumanML3D [37] and KIT-ML [3], while Mo-RepCount, a filtered subset of RepCount [38], provided high-quality repetition count videos. This multimodal data configuration ensures that LLaMo effectively learns from both single and combined modalities." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.471, + 0.484, + 0.562 + ], + "angle": 0, + "content": "Evaluation Datasets: LLaMo was evaluated on MoVid-Bench [1] to assess both video and motion understanding capabilities, with additional tests on BABEL-QA [19] to evaluate motion-based question answering. For specific tasks, we used Mo-RepCount to gauge motion details and Swing to evaluate professional sports guidance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.581, + 0.484, + 0.718 + ], + "angle": 0, + "content": "Evaluation Metrics: For MoVid-Bench, we follow previous LLM evaluation metrics [39] [22] [40] on accuracy and scores, which contains body-part awareness, sequentially, direction analysis, reasoning ability, and hallucination, respectively. BABEL-QA used the metric followed by [19], Mo-RepCount applied metrics such as OBO, MAE, OBZ, and RMSE, and Swing performance was assessed on reasonableness, coherence, pertinence, and adaptability using GPT-4 evaluations, scoring from 0 to 5." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.727, + 0.185, + 0.743 + ], + "angle": 0, + "content": "4.2. Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Evaluation on Motion Understanding in MoVid-Bench. We evaluated LLaMo's motion understanding capabilities on the MoVid-Bench, focusing on five critical aspects. Both accuracy and score metrics were used to comprehensively assess LLaMo's performance. Table 1 provides a comparison with several baselines, including GPT-3.5 [41], MotionGPT [2], and MotionLLM [1]. The GPT-3.5 baseline, limited to textual data processing, struggles significantly with motion interpretation, yielding low performance. MotionGPT, while capable of incorporating motion data, is pri" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.181 + ], + "angle": 0, + "content": "marily code motion into a codebook, which limits its reasoning ability on coiled motion data and leaves it vulnerable to hallucinations. Although MotionLLM provides improvements in understanding motion sequences, it still conducts translation on motion data, resulting in low accuracy on seq. and Hall." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.183, + 0.908, + 0.44 + ], + "angle": 0, + "content": "LLaMo, in contrast, achieves superior results across nearly all metrics on MoVid-Bench-Motion, attaining the highest accuracy in most categories. This performance is primarily attributed to its preservation of critical motion-specific nuances, as LLaMo processes motion data as a distinct modality without requiring any conversion. LLaMo effectively captures the intricate details of 3D sequences associated with human behaviors. Its Feature Enhancer, combined with the Cross Talker module, enables simultaneous input of video and motion data, allowing LLaMo to selectively focus on motion frames most relevant to the text context. This design enhances the model's ability to capture complex patterns of human motion and improves its reasoning on motion-intensive tasks. Overall, LLaMo's results on MoVid-Bench demonstrate its effectiveness in advancing human behavior analysis, highlighting its potential for broader applications in general motion understanding." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.459, + 0.907, + 0.473 + ], + "angle": 0, + "content": "Evaluation on Video Understanding in MoVid-Bench." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.474, + 0.905, + 0.564 + ], + "angle": 0, + "content": "We evaluated LLaMo on the video component of the MoVid-Bench dataset to assess its capability for comprehensive video understanding. As presented in Table 1, LLaMo achieves strong performance across most metrics, demonstrating LLaMo's extraordinary comprehension ability on videos even without motion input." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.565, + 0.906, + 0.669 + ], + "angle": 0, + "content": "Compared to other models like GPT-3.5 [41], VideoLLAVA [39], and MotionLLM [1], while LLaMo slightly trails behind MotionLLM in body-part awareness, LLaMo shows significant improvements in reasoning ability and overall scores. This indicates that our model excels in understanding complex video content and making accurate inferences about human behaviors." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.67, + 0.907, + 0.761 + ], + "angle": 0, + "content": "The rewarding results can be attributed to our model's powerful ability to estimate motion data accurately in video and align it seamlessly with video inputs. This approach retains the motion capture of motion data and the auxiliary information provided by video, enriching the understanding of dynamic scenes." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.781, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Evaluation on BABEL-QA. We evaluated LLaMo on the BABEL-QA dataset to assess its capability to handle complex motion-based queries. As shown in Table 2, LLaMo achieves the highest overall score of 0.458, outperforming all baselines. It excels particularly in the Action category with a score of 0.525, demonstrating superior action recognition capabilities. Additionally, it achieves top scores in temporal reasoning tasks (Before and After), indicating a" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17587" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.101, + 0.115, + 0.113, + 0.137 + ], + "angle": 0, + "content": "e e" + }, + { + "type": "image", + "bbox": [ + 0.119, + 0.093, + 0.178, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.188, + 0.092, + 0.24, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.251, + 0.093, + 0.299, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.311, + 0.093, + 0.357, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.093, + 0.409, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.422, + 0.093, + 0.469, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.48, + 0.093, + 0.528, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.537, + 0.1, + 0.555, + 0.115 + ], + "angle": 0, + "content": "#" + }, + { + "type": "text", + "bbox": [ + 0.56, + 0.097, + 0.873, + 0.129 + ], + "angle": 0, + "content": "Describe the following human motion video concisely, including but not limited to what actions are being performed, the standards followed, and the characteristics." + }, + { + "type": "text", + "bbox": [ + 0.559, + 0.138, + 0.876, + 0.228 + ], + "angle": 0, + "content": "The person is practicing a Tai Chi movement called \"Golden Rooster Stands on One Leg.\" The movement starts from a wide stance with extended arms and gradually transitions into a single-leg balance, with the other knee lifted and the hands drawn close to the body. The entire process is slow and controlled, with smooth and continuous motions, emphasizing the integration of mind and body, showcasing the characteristics of Tai Chi." + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.165, + 0.341, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.341, + 0.169, + 0.528, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.119, + 0.237, + 0.341, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.238, + 0.634, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.236, + 0.864, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.301, + 0.312, + 0.328 + ], + "angle": 0, + "content": "What is the person in the video doing? The person in the video is making a phone" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.331, + 0.317, + 0.359 + ], + "angle": 0, + "content": "Can you analyze her current emotions? Sure, she seems a little impatient and angry" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.363, + 0.353, + 0.387 + ], + "angle": 0, + "content": "Why do you think she's angry? Could it be that she's feeling surprised instead?" + }, + { + "type": "list", + "bbox": [ + 0.097, + 0.301, + 0.353, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.391, + 0.362, + 0.447 + ], + "angle": 0, + "content": "Determining whether a person is angry can be based on their body language. She is likely feeling angry because she stopped and stomped her foot while walking, and her body movements became somewhat sluggish after hanging up the phone." + }, + { + "type": "text", + "bbox": [ + 0.366, + 0.328, + 0.615, + 0.352 + ], + "angle": 0, + "content": "How many push-ups did the person do in the video?" + }, + { + "type": "text", + "bbox": [ + 0.367, + 0.361, + 0.628, + 0.396 + ], + "angle": 0, + "content": "The person in the video did three push-ups. Is it definitely 3 times? Why do I remember it as 5 times?" + }, + { + "type": "text", + "bbox": [ + 0.366, + 0.402, + 0.632, + 0.447 + ], + "angle": 0, + "content": "Let me observe carefully again. It's likely that the smaller range of motion led to your misjudgment. Indeed, he did three push-ups, with three sets of repeated movements." + }, + { + "type": "text", + "bbox": [ + 0.636, + 0.298, + 0.901, + 0.321 + ], + "angle": 0, + "content": "What should be the position of the knees in golf technique?" + }, + { + "type": "text", + "bbox": [ + 0.637, + 0.326, + 0.766, + 0.37 + ], + "angle": 0, + "content": "GT: You want to keep your knees a little wider and more stable in golf technique." + }, + { + "type": "text", + "bbox": [ + 0.772, + 0.318, + 0.901, + 0.373 + ], + "angle": 0, + "content": "LLaMo: Keep knees slightly bent and stable, with the left knee moving forward during the downswing." + }, + { + "type": "text", + "bbox": [ + 0.637, + 0.376, + 0.901, + 0.398 + ], + "angle": 0, + "content": "What joint angles are important at the top of the swing for elite golfers?" + }, + { + "type": "text", + "bbox": [ + 0.637, + 0.402, + 0.766, + 0.446 + ], + "angle": 0, + "content": "GT: Most elite golfers have joint angles ranging from 35 to 45 degrees at the top of the swing." + }, + { + "type": "text", + "bbox": [ + 0.772, + 0.402, + 0.901, + 0.446 + ], + "angle": 0, + "content": "LLaMo: At the top of the swing, elite golfers have hips at 45 degrees and wrists at 90 degrees." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.461, + 0.905, + 0.487 + ], + "angle": 0, + "content": "Figure 4. Example outputs from LLaMo across human activities and professional sports, showcasing its reasoning capabilities and domain-specific knowledge in motion-intensive scenarios." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.503, + 0.906, + 0.531 + ], + "angle": 0, + "content": "Table 1. Expected Comparison on the MoVid-Bench. The top part of the table presents motion-related results, and the bottom part presents video-related results. Higher accuracy and score values indicate better performance." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.543, + 0.903, + 0.719 + ], + "angle": 0, + "content": "
MoVid-Bench-MotionBody.Seq.Dir.Rea.Hall.All
Acc.ScoreAcc.ScoreAcc.ScoreAcc.ScoreAcc.ScoreAcc.Score
GT100.005.00100.005.00100.005.00100.005.00100.005.00100.005.00
GPT-3.5 [41]24.512.0430.412.2527.142.1939.192.6458.333.2231.332.31
MotionGPT [2]31.223.9842.693.1644.293.5035.813.0616.662.2536.863.11
MotionLLM [1]50.493.5536.843.1458.573.7652.703.5855.563.3949.503.49
LLaMo59.304.0144.013.1260.913.9958.213.6461.173.5355.323.67
MoVid-Bench-VideoBody.Seq.Dir.Rea.Hull.All
Acc.ScoreAcc.ScoreAcc.ScoreAcc.ScoreAcc.ScoreAcc.Score
GT100.005.00100.005.00100.005.00100.005.00100.005,00100.005.00
GPT-3.5 [41]2.401.231.391.004.651.095.411.650.000.943.031.26
Video-LLAVA [39]33.532.7625.462.7241.862.8452.973.2858.831.8942.532.70
MotionLLM [1]34.132.9332.872.9244.183.1463.203.5570.592.3049.002.97
LLaMo33.832.8536.013.1145.503.3267.593.7372.812.2552.333.10
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.744, + 0.483, + 0.848 + ], + "angle": 0, + "content": "strong understanding of temporal relationships in motion sequences. This capability benefits from the direct integration of motion data as an independent modality and the effective text-aware motion frames extract mechanism in LLaMo, preserving crucial motion-specific nuances, as well as leading to more accurate and contextually rich interpretations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Professional Sports Analysis. We further evaluate LLaMo on the challenging swing-dataset, termed as" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.744, + 0.907, + 0.879 + ], + "angle": 0, + "content": "profession-swing, details refer to Appendix, where the ground truth (GT) consists of answers provided by professional coaches, serving as expert benchmarks. The evaluation focuses on four key indicators: Reasonableness: Logical soundness and plausibility of the responses. Coherence: Consistency and logical flow within the responses. Pertinence: Relevance of the answers to the specific questions. Adaptability: Ability to adjust responses based on the stage on the athlete." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "As presented in Table 3, LLaMo significantly outper" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17588" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.236, + 0.089, + 0.761, + 0.104 + ], + "angle": 0, + "content": "Table 2. Comparison on BABEL-QA dataset. Higher scores indicate better performance." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.115, + 0.898, + 0.194 + ], + "angle": 0, + "content": "
ModelPred. typeOverall ↑Action ↑Direction ↑Body Part ↑Before ↑After ↑Other ↑
MotionCLIP-M [41]cls.0.4300.4850.3610.2720.3720.3210.404
MotionCLIP-R [41]cls.0.4200.4890.3100.2500.3980.3140.387
MotionLLM [41]gen.0.4360.5170.3540.1540.4270.3680.529
LLaMogen.0.4580.5250.3980.2240.4430.3920.518
" + }, + { + "type": "table_caption", + "bbox": [ + 0.093, + 0.207, + 0.904, + 0.222 + ], + "angle": 0, + "content": "Table 3. Performance on the profession-swing dataset across four indicators. Higher accuracy and score values indicate better performance." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.233, + 0.9, + 0.334 + ], + "angle": 0, + "content": "
ModelReasonablenessCoherencePertinenceAdaptabilityAll
AccScoreAccScoreAccScoreAccScoreAccScore
GT10051005100510051005
GPT-3.5 [41]4.510.834.200.743.650.661.710.223.420.61
MotionGPT [2]10.531.2515.421.6612.641.114.791.3514.351.4
MotionLLM [1]12.331.5119.201.8717.981.7910.201.2216.531.57
LLaMo21.102.1127.102.7131.813.1220.221.9824.802.48
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.358, + 0.482, + 0.435 + ], + "angle": 0, + "content": "forms both GPT-3.5 [41] and MotionLLM [1] across all metrics. While MotionLLM shows improvements over GPT-3.5, with higher accuracy and scores, it still lags behind LLaMo. Specifically, LLaMo achieves an overall accuracy of 24.80" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.435, + 0.483, + 0.555 + ], + "angle": 0, + "content": "The superior performance of LLaMo can be attributed to the design of the Cross Talker module, which enables language-guided frame selection to capture complex motion relationships inherent in professional swings. This allows LLaMo to effectively model intricate motion patterns and generate responses that are more reasonable, coherent, pertinent, and adaptable, closely aligning with the expert-level insights provided by professional coaches." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.556, + 0.483, + 0.618 + ], + "angle": 0, + "content": "In contrast, while MotionLLM demonstrates certain strengths over MotionGPT—such as better pertinence and coherence. It lacks the advanced motion understanding capabilities of LLaMo, lagging behind LLaMo distinctly." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.658 + ], + "angle": 0, + "content": "Table 4. Motion and video details capture evaluation on MoRepCount" + }, + { + "type": "table", + "bbox": [ + 0.114, + 0.669, + 0.458, + 0.777 + ], + "angle": 0, + "content": "
ModelOBOMAEOBZRMSE
EScounts [42]0.3970.2910.1985.58
PoseRAC [43]0.3820.3120.2045.95
TransRAC [38]0.2760.4440.1058.56
RepNet [44]0.009///
LLaMo0.3890.3240.2226.15
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Evaluation on Mo-RepCount. We evaluated LLaMo on the Mo-RepCount dataset to assess its ability to capture fine-grained motion details and complex temporal features required for accurate repetition counting. For a fair comparison, we trained all state-of-the-art models, including RepNet [44], TransRAC [38], PoseRAC [43] and ES" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.358, + 0.905, + 0.465 + ], + "angle": 0, + "content": "counts [42], on the Mo-RepCount dataset without using any additional data. As presented in Table 4, LLaMo outperforms the RepNet and TransRAC across almost all metrics. Since both of them are models built for loop counting tasks, the effect of our model on this task is convincing, which shows that LLaMo is also competent in capturing the details of the motion and videos." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.468, + 0.907, + 0.577 + ], + "angle": 0, + "content": "While LLaMo did not outperform the specialized counting models EScounts and poseRAC on many metrics, it excelled on the OBZ metric, demonstrating an unmatched ability to capture fine-grained motion details and spatiotemporal relationships. These results confirm LLaMo's strong potential in motion and video analysis, underscoring its suitability for detail-oriented tasks and applications." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.6, + 0.634, + 0.616 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We introduce LLaMo, a novel framework that integrates motion data directly into large language models, thereby enhancing multimodal understanding. By unifying text, video, and motion data in an architecture fine-tuned for human motion instruction, LLaMo achieves notable progress in behavior comprehension, motion-focused captioning, and interactive Q&A, capturing the fine-grained motion details vital for precise, context-aware insights. Experimental results, especially in professional sports analysis, demonstrate LLaMo's capacity to interpret complex human motion with accuracy. Future work will extend LLaMo to a broad range of domains where detailed motion analysis and multimodal understanding are essential. The main challenges lie in refining cross-modal integration techniques and improving efficiency for real-time applications, thus broadening LLaMo's potential in human-centered AI. We hope our work will spur further exploration into more advanced human-centric multimodal models." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17589" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.115, + 0.484, + 0.146 + ], + "angle": 0, + "content": "This work was supported in part by the Pioneer Centre for AI, DNRF grant number P1." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.159, + 0.188, + 0.174 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.184, + 0.484, + 0.24 + ], + "angle": 0, + "content": "[1] Ling-Hao Chen, Shunlin Lu, Ailing Zeng, Hao Zhang, Benyou Wang, Ruimao Zhang, and Lei Zhang. Motionllm: Understanding human behaviors from human motions and videos. arXiv preprint arXiv:2405.20340, 2024. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.241, + 0.484, + 0.295 + ], + "angle": 0, + "content": "[2] Ling-Hao Chen, Jiawei Zhang, Wen Liu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. In Proceedings of the Conference on Neural Information Processing Systems (NeurIPS), 2024. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.297, + 0.482, + 0.337 + ], + "angle": 0, + "content": "[3] Matthias Plappert, Christian Mandery, and Tamim Asfour. The kit motion-language dataset. Big data, 4(4):236-252, 2016. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.339, + 0.484, + 0.394 + ], + "angle": 0, + "content": "[4] Yan Zhang, Michael J Black, and Siyu Tang. We are more than our joints: Predicting how 3d bodies move. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.396, + 0.484, + 0.464 + ], + "angle": 0, + "content": "[5] Fangzhou Hong, Liang Pan, Zhongang Cai, and Ziwei Liu. Versatile multi-modal pre-training for human-centric perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16156-16166, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.466, + 0.484, + 0.521 + ], + "angle": 0, + "content": "[6] Haoxuan Qu, Yujun Cai, and Jun Liu. Llms are good action recognizers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18395-18406, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.522, + 0.482, + 0.563 + ], + "angle": 0, + "content": "[7] Shan Chen, Jiale Zhou, and Lei Li. Dense point clouds matter: Dust-gs for scene reconstruction from sparse viewpoints. arXiv preprint arXiv:2409.08613, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.565, + 0.484, + 0.633 + ], + "angle": 0, + "content": "[8] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.635, + 0.484, + 0.704 + ], + "angle": 0, + "content": "[9] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.705, + 0.484, + 0.76 + ], + "angle": 0, + "content": "[10] Li Lei, Jia Sen, Wang Jianhao, An Zhaochong, Li Jiaang, Hwang Jenq-Neng, and Belongie Serge. Chatmotion: A multimodal multi-agent for human motion analysis. arXiv preprint arXiv:2502.18180, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.761, + 0.484, + 0.816 + ], + "angle": 0, + "content": "[11] Mark Endo, Joy Hsu, Jiaman Li, and Jiajun Wu. Motion question answering via modular motion programs. In Proceedings of the International Conference on Machine Learning (ICML), 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.484, + 0.872 + ], + "angle": 0, + "content": "[12] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[13] Yaya Shi, Haiyang Xu, Chunfeng Yuan, Bing Li, Weiming Hu, and Zheng-Jun Zha. Learning video-text aligned repre" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.184, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "sentations for video captioning. IEEE Transactions on Multimedia (TMM), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[14] Chuan Guo, Xinxin Zuo, Sen Wang, and Li Cheng. Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In Proceedings of the European Conference on Computer Vision (ECCV), pages 580-597, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[15] Kunchang Li, Yali Wang, Yinan He, et al. Mvbench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.305 + ], + "angle": 0, + "content": "[16] Zhaochong An, Guolei Sun, Yun Liu, Runjia Li, Min Wu, Ming-Ming Cheng, Ender Konukoglu, and Serge Belongie. Multimodality helps few-shot 3d point cloud semantic segmentation. arXiv preprint arXiv:2410.22489, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.306, + 0.905, + 0.374 + ], + "angle": 0, + "content": "[17] Bobby Bodenheimer, Chuck Rose, Seth Rosenthal, and John Pella. The process of motion capture: Dealing with the data. In Proceedings of the Eurographics Workshop on Computer Animation and Simulation (EG Workshop), pages 3-18, 1997. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.376, + 0.905, + 0.433 + ], + "angle": 0, + "content": "[18] Xiangyu Peng, Yuxuan Zhang, Xiaolong Wang, and Ziwei Liu. Text-driven human motion generation with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.905, + 0.489 + ], + "angle": 0, + "content": "[19] Mark Endo, Joy Hsu, Jiaman Li, and Jiajun Wu. Motion question answering via modular motion programs. In International Conference on Machine Learning, pages 9312-9328. PMLR, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.907, + 0.546 + ], + "angle": 0, + "content": "[20] Jie Lin, Wei Zhang, and Zhe Chen. Videollm: Language models for video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2405-2415, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.547, + 0.905, + 0.587 + ], + "angle": 0, + "content": "[21] Xu Ning, Hongwei Li, and Yu Zhao. Videobench: A benchmark for large-scale video understanding. arXiv preprint arXiv:2304.12345, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.59, + 0.905, + 0.645 + ], + "angle": 0, + "content": "[22] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.647, + 0.905, + 0.675 + ], + "angle": 0, + "content": "[23] Lei Li. Segment any building. In Computer Graphics International Conference, pages 155-166. Springer, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.676, + 0.905, + 0.732 + ], + "angle": 0, + "content": "[24] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.733, + 0.907, + 0.788 + ], + "angle": 0, + "content": "[25] Enxin Song, Guanhong Zhang, and Haoyang Zhou. Adaptive multimodal learning for behavior analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[26] Yunhua Yang, Ziwang Zhao, and Yiming Xie. Recognizing human behaviors with skeletal data in llm-based frameworks. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[27] Lei Li, Tianfang Zhang, Zhongfeng Kang, and Xikun Jiang. Mask-fpan: Semi-supervised face parsing in the wild with de-occlusion and uv gan. Computers & Graphics, 116:185–193, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17590" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.16 + ], + "angle": 0, + "content": "[28] Matthias Plappert, Christian Mandery, and Tamim Asfour. Learning a bidirectional mapping between human whole-body motion and natural language using deep recurrent neural networks. Robotics and Autonomous Systems (RAS), 109: 13-26, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.483, + 0.218 + ], + "angle": 0, + "content": "[29] Lei Li. Cpseg: Finer-grained image semantic segmentation via chain-of-thought language prompting. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 513-522, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.219, + 0.482, + 0.26 + ], + "angle": 0, + "content": "[30] Yunhua Yang, Liang Zhang, and Hui Li. Understanding human behaviors from skeletal data: A review of datasets and methods. arXiv preprint arXiv:2310.12998, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.261, + 0.482, + 0.328 + ], + "angle": 0, + "content": "[31] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Two-stream adaptive graph convolutional networks for skeleton-based action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12026-12035, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.482, + 0.383 + ], + "angle": 0, + "content": "[32] Enxin Song, Wenhao Chai, and Yucheng Zhang. Fine-grained spatial-temporal motion understanding in complex video environments. arXiv preprint arXiv:2310.08639, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.385, + 0.482, + 0.439 + ], + "angle": 0, + "content": "[33] Zhaochong An, Guolei Sun, Zongwei Wu, Hao Tang, and Luc Van Gool. Temporal-aware hierarchical mask classification for video semantic segmentation. arXiv preprint arXiv:2309.08020, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.482, + 0.51 + ], + "angle": 0, + "content": "[34] Xianghong Liu, Haoxuan Wang, Zhiwei Zhang, Huan Wu, Baoquan Chen, and Xiaoguang Han. Category-agnostic pose estimation for point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.482, + 0.566 + ], + "angle": 0, + "content": "[35] Xiaolong He, Yi Zhang, Chen Chen, and Kyoung Mu Lee. Activitynet++: A large-scale benchmark for video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.482, + 0.608 + ], + "angle": 0, + "content": "[36] Lei Li. Hierarchical edge aware learning for 3d point cloud. In Computer Graphics International Conference, pages 81-92. Springer, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.609, + 0.482, + 0.678 + ], + "angle": 0, + "content": "[37] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5152-5161, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.482, + 0.761 + ], + "angle": 0, + "content": "[38] Huazhang Hu, Sixun Dong, Yiqun Zhao, Dongze Lian, Zhengxin Li, and Shenghua Gao. Transrac: Encoding multiscale temporal correlation with transformers for repetitive action counting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19013-19022, 2022. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.482, + 0.816 + ], + "angle": 0, + "content": "[39] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[40] Peng Jin, Ryuichi Takanobu, Wancai Zhang, Xiaochun Cao, and Li Yuan. Chat-univi: Unified visual representation empowers large language models with image and video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13700-13710, 2024. 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[41] OpenAI. Gpt-3.5: Generative pre-trained transformer 3.5, 2023. https://platform.openai.com/docs/models/gpt-3-5.6,7,8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[42] Saptarshi Sinha, Alexandros Stergiou, and Dima Damen. Every shot counts: Using exemplars for repetition counting in videos. arXiv preprint arXiv:2403.18074, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.22 + ], + "angle": 0, + "content": "[43] Ziyu Yao, Xuxin Cheng, and Yuexian Zou. Poserac: Pose saliency transformer for repetitive action counting. arXiv preprint arXiv:2303.08450, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.222, + 0.905, + 0.291 + ], + "angle": 0, + "content": "[44] Bastian Wandt and Bodo Rosenhahn. Repnet: Weakly supervised training of an adversarial reprojection network for 3d human pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7782-7791, 2019. 8" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.093, + 0.905, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17591" + } + ] +] \ No newline at end of file diff --git a/2025/Human Motion Instruction Tuning/8c08c8b0-9aae-431a-b68e-699a5ddd5c32_origin.pdf b/2025/Human Motion Instruction Tuning/8c08c8b0-9aae-431a-b68e-699a5ddd5c32_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1827a7a44f0c9160e07e255b9b00f47987e68add --- /dev/null +++ b/2025/Human Motion Instruction Tuning/8c08c8b0-9aae-431a-b68e-699a5ddd5c32_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:19bf0887ddd97a29bd5fbde468d75f498ad8eca3fb0a36a7354e833935a8a717 +size 2251575 diff --git a/2025/Human Motion Instruction Tuning/full.md b/2025/Human Motion Instruction Tuning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a8e605b56f7c37ece7fd3fccb3c67873a153722f --- /dev/null +++ b/2025/Human Motion Instruction Tuning/full.md @@ -0,0 +1,364 @@ +# Human Motion Instruction Tuning + +Lei Li $^{1,2,\ast,\dagger}$ , Sen Jia $^{1,\ast}$ , Jianhao Wang $^{3}$ , Zhongyu Jiang $^{1}$ , Feng Zhou $^{4}$ , Ju Dai $^{5}$ , Tianfang Zhang $^{6}$ , Zongkai Wu $^{1}$ , Jenq-Neng Hwang $^{1}$ + +# Abstract + +This paper presents LLaMo (Large Language and Human Motion Assistant), a multimodal framework for human motion instruction tuning. In contrast to conventional instruction-tuning approaches that convert non-linguistic inputs, such as video or motion sequences, into language tokens, LLaMo retains motion in its native form for instruction tuning. This method preserves motion-specific details that are often diminished in tokenization, thereby improving the model's ability to interpret complex human behaviors. By processing both video and motion data alongside textual inputs, LLaMo enables a flexible, human-centric analysis. Experimental evaluations across high-complexity domains, including human behaviors and professional activities, indicate that LLaMo effectively captures domain-specific knowledge, enhancing comprehension and prediction in motion-intensive scenarios. We hope LLaMo offers a foundation for future multimodal AI systems with broad applications, from sports analytics to behavioral prediction. + +# 1. Introduction + +Understanding human motion is a central challenge in multimodal AI, impacting numerous fields such as digital human, human-computer interaction, sports analytics, healthcare, and virtual human modeling [3-6]. Motion data, which captures skeletal movements and body dynamics, provides a structured, appearance-invariant representation of human actions, concentrating on the essential movement patterns while excluding irrelevant visual details [7]. This data has emerged as a privacy-conscious alternative to visual inputs, offering finer control over action recognition and facilitating analysis in applications ranging from digital human avatars to behavioral monitoring in healthcare [8-10]. As digital human representations and motion-centric + +![](images/6929f4853c166dfc7e60a280057512ce73bd37d5595f226416333ac6f813430e.jpg) +Figure 1. A comparison of MotionLLM [1], MotionGPT [2], and LLaMo highlights LLaMo's motion-specific capabilities. Equipped with a Motion Enhancer and Cross Talker module to align motion and text, LLaMo supports both video and motion inputs, enabling text-aware, fine-grained motion analysis. + +applications continue to expand, developing models that can leverage the unique characteristics of motion data becomes increasingly crucial for achieving robust, context-aware behavior analysis. + +Recently, large language models (LLMs) have shown promise in motion analysis, leveraging their capacity to model temporal dependencies across multimodal data [1, 11]. State-of-the-art models, such as MotionGPT [12] and MotionLLM [1], treat motion data as a multimodal extension, encoding it as discrete tokens or translating it into language-like representations for processing by LLMs. These models have yielded notable results in motion understanding and recognition tasks, illustrating the adaptability of language models to non-linguistic data. However, their reliance on tokenization or textual conversion can limit the richness of the motion representation, abstracting away critical spatial-temporal details essential for high-fidelity human activity recognition. + +Despite recent advancements, notable challenges remain in current approaches to motion analysis using LLMs. First, encoding motion data into language tokens, as implemented + +in models such as MotionGPT, relies on a quantization process that may obscure critical fine-grained motion information, potentially limiting the model's capacity to accurately capture detailed spatial-temporal dynamics [13, 14]. The common approach of translating motion data into text tokens further contributes to this issue, often resulting in the loss of essential motion-specific nuances that are critical for in-depth behavior interpretation [11, 15]. This translation process poses considerable challenges in understanding the intricate details of human movement, impeding the model's ability to achieve spatial and physical comprehension of human actions. Moreover, motion data inherently contains 3D structural information and insights into connected human behaviors, which are often difficult to fully capture when relying solely on language tokens, as this representation can overlook the underlying spatial dependencies crucial to modeling human motion accurately. + +In addition, existing methods frequently treat video and motion data as isolated modalities, overlooking the potential benefits of a complementary, integrated approach [16]. Processing these data types independently neglects the rich environmental context and interaction cues provided by video inputs, which are crucial for accurately interpreting complex human behaviors [5, 8]. This separation limits the capacity for a holistic understanding of human motion, particularly in scenarios where context, spatial relations, and dynamic interactions play essential roles in behavior analysis. Furthermore, video-based approaches, while potentially more informative, are computationally intensive, posing significant constraints for large-scale or real-time applications, which limits their feasibility in practical motion analysis systems. + +In response to these challenges, we propose LLaMo, a framework designed to integrate motion data as a distinct modality within LLMs without translating it into textual intermediaries. LLaMo incorporates a Motion Estimator and an Enhancer to facilitate the direct input of motion or/and video data, thereby offering a unified, human-centric analysis pipeline. The framework also introduces a Cross Talker module, which dynamically aligns the motion and text features, allowing for text-aware, fine-grained motion representations. This approach retains the original spatial-temporal characteristics of the motion data, providing a more precise understanding of human actions and enabling high-resolution behavior analysis in motion-intensive contexts [17, 18]. + +Through extensive evaluations on benchmarks such as MoVid-Bench [1] and BABEL-QA [19], we demonstrate that LLaMo achieves state-of-the-art performance in motion-centric human activity recognition [11, 15]. These results underscore LLaMo's potential to advance multimodal AI, providing a robust foundation for human behavior understanding. Our contributions can be summarized as + +follows: + +- We introduce LLaMo, which treats motion data as an independent modality within LLMs, preserving essential motion-specific details for robust analysis. +- We propose Cross Talker, a text-guided mechanism that dynamically focuses on and aggregates key motion frames features, optimizing computational efficiency and enabling the model to pick relevant motion features. +- LLaMo's architecture supports the direct input of both raw motion and video data, providing a generalized framework for human-centric analysis across diverse applications, such as sports analytics, healthcare, and behavioral monitoring. + +# 2. Related Work + +# 2.1. Human-Centric Multimodal Representation + +Multimodal representation learning is integral to human-centric analysis, especially for applications requiring spatial-temporal reasoning to decode complex behaviors [20-22]. Recent advances, such as Video-LLaVA, have made significant strides by embedding visual data from images and videos into a shared linguistic feature space, enabling sophisticated visual reasoning for behavioral analysis tasks [20, 23]. Despite these achievements, many current models are optimized primarily for static images or discrete video frames, limiting their effectiveness in sequential, dynamic scenarios where understanding progression and continuity in motion is crucial [9, 21, 24]. Privacy concerns further challenge the adoption of video-based methods in sensitive domains; thus, researchers are exploring motion data as a privacy-conscious alternative, enabling analysis focused on human actions without revealing identifiable visual information [25-27]. By combining visual and motion data, emerging multimodal frameworks hold promise for more comprehensive, privacy-aware human behavior analysis, leveraging the strengths of both modalities for adaptability and depth across applications. + +# 2.2. Human Motion Understanding + +Human motion understanding has traditionally relied on skeletal data represented as sequences of joint keypoints to capture movement while preserving user privacy [13, 28-30]. Early models like 2s-AGCN [31] and more recent transformer-based methods, such as MotionCLIP [2], have achieved success in tasks like activity recognition, captioning, and behavior analysis by mapping motion data to language tokens, effectively capturing structural aspects of movement. However, these methods often lack environmental context, which is crucial for nuanced interpretations in real-world applications, as similar movements can convey different meanings across varied scenarios [24, 32, 33]. To address this limitation, recent frameworks inte + +grate motion and visual data, allowing models to generalize across diverse, dynamic environments [34-36]. This fusion of skeletal data with video-provided context has shown promise in applications like sports and professional activity analysis, where contextual cues are essential. Additionally, skeletal-based data remains valuable in privacy-sensitive applications, such as healthcare and security, where it allows for robust analysis without compromising personal identity. + +# 3. Methods + +The LLaMo framework is composed of three primary modules to effectively process and integrate video, motion data and text. An overview of this pipeline is provided in Figure 2. The architecture comprises the following components: + +1. Multimodal Feature Extraction Module: This module independently encodes the video or motion inputs, which makes LLaMo is a general human motion assistant with video input or motion input. Then the Motion features are enhanced by distilling valuable representations in video. +2. Cross Talker: Here, the enhanced motion features are aggregated in a text-guided manner and aligned with text within a shared semantic space, allowing the natural feature fusion between motion and text, which provides fine-grained inputs for LLM. +3. Behavior Generation Module: Using the aggregated and text-aligned motion features, along with the text representations, this module generates a contextually aware textual description of the observed human behavior. + +# 3.1. Enhanced motion Feature Extraction Module + +The enhanced motion feature extraction module consists of four primary components: the motion estimator, followed by the motion and video encoders, and the feature enhancement module. The motion estimator addresses scenarios where motion data may be unavailable by estimating motion information directly from video frames. The motion and video encoders then independently process the motion and video data, encoding them into their respective feature spaces. This approach allows the system to leverage complementary information from both modalities, creating a richer representation of human behavior. + +Given a sequence of video frames $V = \{v_{1}, v_{2}, \ldots, v_{T}\}$ , the video encoder $f_{v}(\cdot)$ extracts a set of visual features: + +$$ +F _ {V} = f _ {v} (V) = \left\{f _ {v} \left(v _ {1}\right), f _ {v} \left(v _ {2}\right), \dots , f _ {v} \left(v _ {T}\right) \right\}. +$$ + +Simultaneously, the motion encoder $f_{m}(\cdot)$ processes the motion data $M = \{m_1, m_2, \ldots, m_T\}$ to generate motion-specific features: + +$$ +F _ {M} = f _ {m} (M) = \left\{f _ {m} \left(m _ {1}\right), f _ {m} \left(m _ {2}\right), \dots , f _ {m} \left(m _ {T}\right) \right\}. +$$ + +Here, $F_{V}\in \mathbb{R}^{T\times H}$ and $F_{M}\in \mathbb{R}^{T\times H}$ represent the extracted feature sets for the video and motion data, respectively, where $T$ is the sequence length and $H$ denotes the feature dimension. + +The feature enhancement module then aligns and fuses these modality-specific features. By using the motion features $F_{M}$ as queries, this module selectively extracts semantically relevant information from the video features $F_{V}$ . This alignment and fusion process allows the motion features to be enriched with contextual information from the video, producing refined, contextually enriched representations that enhance the model's understanding of nuanced human behaviors. For the detailed design refer to the Appendix. + +First, the video features and motion features will conduct self-attention operations respectively to get their augmented features $F_V'$ and $F_M'$ . This allows the video and motion to initially capture the dependencies on their respective time steps. Then, in order to enrich the motion data features with semantic information from the video, we subtly design a cross-attention block where augmented motion features as the query to extract valid semantic information in the video features. Finally, the selected video semantic representation and motion features are connected by residuals followed by a feed-forward neural network to obtain the enhanced motion features $\tilde{F_M}$ + +# 3.2. Cross Talker Module + +Given the enhanced motion features $\tilde{F}_M$ from the feature enhancement module, directly feeding the entire sequence of motion frames into the Large Language Model (LLM) incurs significant computational costs and may lead to suboptimal performance. This limitation arises from the self-attention mechanism in transformers, which has a computational complexity of $\mathcal{O}(L^2)$ , where $L$ denotes the sequence length. When incorporating both motion and text, the sequence length becomes $L = L_{T} + T$ , where $L_{T}$ is the number of text tokens and $T$ is the number of motion frames. As $T$ increases, the quadratic growth in computational cost becomes prohibitive, leading to complexity proportional to $(L_{T} + T)^{2}$ . + +Furthermore, computing attention over long sequences can dilute attention weights, making it challenging for the model to focus on essential parts of the sequence and effectively capture long-range dependencies. Furthermore, the motion features are not well aligned with text either. To address these issues, we propose the Cross Talker Module, which selectively distills key motion frames—referred to as viewpoint frames—based on their relevance to the text context, as well as aligning the motion features with the text. An overview of the Cross Talker Module is shown in Figure 3. + +![](images/48347167cdb846970fabb12d7a345502e93e48aaca805da0b7e77a0124f7d964.jpg) +Figure 2. Overview of the LLaMo framework. It includes three main modules: (1) Multimodal Feature Extraction for encoding video and motion data; (2) Cross Talker for aligning and fusing motion and text features; and (3) Behavior Generation Module to produce text descriptions of human behavior based on integrated features. + +![](images/f9a416491ad6f8c391feccd39022f6b68db29765534857a3c9a8afc333aea68e.jpg) +Figure 3. Overview of the Cross Talker Module, which selects key frames based on text guidance and fuses them with text features for enhanced analysis. + +Language-Guided Frame Selection Given the enhanced motion features $\tilde{F}_M\in \mathbb{R}^{T\times H}$ and text embeddings $F_{T}\in$ $\mathbb{R}^{L_T\times H}$ , we aim to identify $K$ viewpoint frames that are most relevant to the textual context. This selection is achieved through a cross-attention operation, where text + +features act as queries and motion features as keys and values: + +$$ +A = \operatorname {S o f t m a x} \left(\frac {F _ {T} W _ {Q} \left(\tilde {F} _ {M} W _ {K}\right) ^ {\top}}{\sqrt {d}}\right), \tag {1} +$$ + +where $W_{Q}, W_{K} \in \mathbb{R}^{H \times d}$ are learnable projection matrices, and $d$ is the dimensionality of the queries and keys. The resulting attention matrix $A \in \mathbb{R}^{L_{T} \times T}$ contains attention weights, with each element $A_{i,j}$ indicating the relevance between the $i$ -th text token and the $j$ -th motion frame. + +To determine the importance of each motion frame, we aggregate attention weights across all text tokens using max pooling: + +$$ +s _ {j} = \max _ {i = 1, \dots , L _ {T}} A _ {i, j}, \quad j = 1, \dots , T. \tag {2} +$$ + +The resulting scores $s_j$ represent the significance of each frame with respect to the text context. We then select the top $K$ frames with the highest relevance scores as viewpoint frames, reducing the effective sequence length from $T$ to $K$ . This reduction in sequence length lowers the self-attention complexity from $\mathcal{O}((L_T + T)^2)$ to $\mathcal{O}((L_T + K)^2)$ , which is computationally efficient when $K \ll T$ . + +Adaptive Contextual Feature Aggregation For each selected viewpoint frame, we enrich its representation by aggregating both local and global contexts. An adaptive receptive field size $r_k$ is predicted for each viewpoint frame $k$ using a receptive field regression module. The module performs a cross-attention operation where each viewpoint + +frame serves as a query, and the unselected motion features serve as keys and values, followed by a sigmoid activation to estimate $r_k$ . Based on $r_k$ , we define a local window $W_k$ around frame $k$ : + +$$ +W _ {k} = \{j \mid | j - k | \leq r _ {k} \times T \}. \tag {3} +$$ + +Within this window, we apply local attention to capture fine-grained details: + +$$ +F _ {\text {l o c a l}} (k) = \tilde {F} _ {M} (k) + \text {A t t e n t i o n} (\tilde {F} _ {M} (k), \tilde {F} _ {M} \left(W _ {k}\right), \tilde {F} _ {M} \left(W _ {k}\right)), \tag {4} +$$ + +where $\tilde{F}_M(W_k)$ denotes the motion features within the local window around frame $k$ . + +To incorporate global context, we partition the full motion sequence into $N$ segments and compute segment-level features via average pooling: + +$$ +F _ {M} ^ {\mathrm {s e g}} = \left\{\bar {F} _ {M} (n) \right\} _ {n = 1} ^ {N}, \tag {5} +$$ + +where the segment size is controlled by the hyperparameter $S_{n}$ . We then apply global attention: + +$$ +F _ {\text {g l o b a l}} (k) = F _ {\text {l o c a l}} (k) + \text {A t t e n t i o n} \left(F _ {\text {l o c a l}} (k), F _ {M} ^ {\text {s e g}}, F _ {M} ^ {\text {s e g}}\right). \tag {6} +$$ + +The final representation for each viewpoint frame is obtained by concatenating local and global features: + +$$ +F _ {M} (k) = \left[ F _ {\text {l o c a l}} (k); F _ {\text {g l o b a l}} (k) \right] \in \mathbb {R} ^ {2 H}, \tag {7} +$$ + +allowing the model to capture both fine-grained motion details and broader contextual information. + +Bidirectional Cross-Modal Fusion To integrate the enhanced motion features with the text embeddings, we employ a bidirectional cross-attention mechanism. First, we update the motion features by performing a cross-attention operation, where the motion features act as queries and the text features serve as keys and values, and vice versa for the text features. The updated features are then passed through feed-forward networks, and we concatenate the motion and text representations to form the input for the next stage: + +$$ +F _ {\text {f u s i o n}} = \left[ F _ {T}; \{F _ {M} (k) \} _ {k \in K} \right]. \tag {8} +$$ + +This bidirectional fusion enhances the interaction between motion and text, enabling a coherent, contextually grounded understanding of human behavior. By selectively focusing on key motion frames and integrating enriched motion and text features, the Cross Talker Module reduces computational costs and improves the model's ability to capture complex human actions. + +# 3.3. Behavior Generation Module + +The Behavior Generation Module leverages the fused features $F_{\mathrm{fusion}}$ to generate a textual description $Y$ that encapsulates the observed human behavior. This module utilizes a language model $h(\cdot)$ to transform $F_{\mathrm{fusion}}$ into a semantically coherent and contextually relevant output sequence: + +$$ +Y = h \left(F _ {\text {f u s i o n}}\right) = \left\{y _ {1}, y _ {2}, \dots , y _ {L} \right\}, \tag {9} +$$ + +where $y_{t}$ denotes the token generated at the $t$ -th time step, and $L$ is the length of the generated description. The probability of generating each token $y_{t}$ is given by: + +$$ +p \left(y _ {t} \mid y _ {1: t - 1}, F _ {\text {f u s i o n}}\right) = \operatorname {S o f t m a x} \left(W _ {o} h _ {t}\right), \tag {10} +$$ + +where $h_t$ represents the hidden state of the language model at time $t$ , and $W_o$ is an output projection matrix. The hidden state $h_t$ is derived from the model's internal mechanisms, which integrate attention over the input features and previously generated tokens. This autoregressive approach enables LLaMo to produce detailed and contextually aligned descriptions that incorporate nuanced information from both motion and text modalities. + +By iteratively generating tokens based on the fused features and preceding context, the Behavior Generation Module constructs a meaningful and comprehensive narrative of human activity, grounded in the multimodal input data. + +# 3.4. Training Objective + +To train LLaMo, we finetune a pretrained language model using supervised learning, guiding it to generate accurate and contextually relevant descriptions based on the input motion, video, and text features. Both the motion and video encoders remain frozen during training, focusing the optimization on the language model's alignment with multimodal inputs. The training objective minimizes the discrepancy between the generated descriptions and the groundtruth annotations in the dataset. + +Given a training set of $N$ samples, each with video frames $V^{(i)}$ , motion data $M^{(i)}$ , and corresponding ground-truth textual descriptions $\hat{Y}^{(i)} = \{\hat{y}_1^{(i)},\hat{y}_2^{(i)},\dots ,\hat{y}_L^{(i)}\}$ , the overall training objective is to minimize the negative log-likelihood of the ground-truth tokens across all samples: + +$$ +L = - \frac {1}{N} \sum_ {i = 1} ^ {N} \sum_ {t = 1} ^ {L ^ {(i)}} \log p \left(\hat {y} _ {t} ^ {(i)} \mid \hat {y} _ {1: t - 1} ^ {(i)}, F _ {\text {f u s i o n}} ^ {(i)}\right), \tag {11} +$$ + +where $F_{\mathrm{fusion}}^{(i)}$ represents the fused input features for the $i$ -th sample. The term $p(\hat{y}_t^{(i)} \mid \hat{y}_{1:t-1}^{(i)}, F_{\mathrm{fusion}}^{(i)})$ denotes the probability of the ground-truth token $\hat{y}_t^{(i)}$ given the previous tokens $\hat{y}_{1:t-1}^{(i)}$ and the fused features. + +This objective ensures that the generated descriptions are syntactically and semantically aligned with the target + +outputs, capturing the detailed aspects of human behavior present in the video and motion inputs. + +# 4. Experiments + +To evaluate LLaMo, we performed experiments across datasets representing complex human behaviors and tasks, including professional sports and cyclic action repetition counts. + +# 4.1. Implementation Details + +Training Datasets: LLaMo was trained using a combination of bimodal video-motion datasets and unimodal video or motion datasets to enable a robust understanding of both modalities. For video-motion data, we utilized MoVid from MotionLLM [1] and our custom Swing dataset containing 20,000 annotated videos of baseball and golf swings with corresponding motion capture data and expert Q&A instructions. Motion-only datasets included HumanML3D [37] and KIT-ML [3], while Mo-RepCount, a filtered subset of RepCount [38], provided high-quality repetition count videos. This multimodal data configuration ensures that LLaMo effectively learns from both single and combined modalities. + +Evaluation Datasets: LLaMo was evaluated on MoVid-Bench [1] to assess both video and motion understanding capabilities, with additional tests on BABEL-QA [19] to evaluate motion-based question answering. For specific tasks, we used Mo-RepCount to gauge motion details and Swing to evaluate professional sports guidance. + +Evaluation Metrics: For MoVid-Bench, we follow previous LLM evaluation metrics [39] [22] [40] on accuracy and scores, which contains body-part awareness, sequentially, direction analysis, reasoning ability, and hallucination, respectively. BABEL-QA used the metric followed by [19], Mo-RepCount applied metrics such as OBO, MAE, OBZ, and RMSE, and Swing performance was assessed on reasonableness, coherence, pertinence, and adaptability using GPT-4 evaluations, scoring from 0 to 5. + +# 4.2. Results + +Evaluation on Motion Understanding in MoVid-Bench. We evaluated LLaMo's motion understanding capabilities on the MoVid-Bench, focusing on five critical aspects. Both accuracy and score metrics were used to comprehensively assess LLaMo's performance. Table 1 provides a comparison with several baselines, including GPT-3.5 [41], MotionGPT [2], and MotionLLM [1]. The GPT-3.5 baseline, limited to textual data processing, struggles significantly with motion interpretation, yielding low performance. MotionGPT, while capable of incorporating motion data, is pri + +marily code motion into a codebook, which limits its reasoning ability on coiled motion data and leaves it vulnerable to hallucinations. Although MotionLLM provides improvements in understanding motion sequences, it still conducts translation on motion data, resulting in low accuracy on seq. and Hall. + +LLaMo, in contrast, achieves superior results across nearly all metrics on MoVid-Bench-Motion, attaining the highest accuracy in most categories. This performance is primarily attributed to its preservation of critical motion-specific nuances, as LLaMo processes motion data as a distinct modality without requiring any conversion. LLaMo effectively captures the intricate details of 3D sequences associated with human behaviors. Its Feature Enhancer, combined with the Cross Talker module, enables simultaneous input of video and motion data, allowing LLaMo to selectively focus on motion frames most relevant to the text context. This design enhances the model's ability to capture complex patterns of human motion and improves its reasoning on motion-intensive tasks. Overall, LLaMo's results on MoVid-Bench demonstrate its effectiveness in advancing human behavior analysis, highlighting its potential for broader applications in general motion understanding. + +# Evaluation on Video Understanding in MoVid-Bench. + +We evaluated LLaMo on the video component of the MoVid-Bench dataset to assess its capability for comprehensive video understanding. As presented in Table 1, LLaMo achieves strong performance across most metrics, demonstrating LLaMo's extraordinary comprehension ability on videos even without motion input. + +Compared to other models like GPT-3.5 [41], VideoLLAVA [39], and MotionLLM [1], while LLaMo slightly trails behind MotionLLM in body-part awareness, LLaMo shows significant improvements in reasoning ability and overall scores. This indicates that our model excels in understanding complex video content and making accurate inferences about human behaviors. + +The rewarding results can be attributed to our model's powerful ability to estimate motion data accurately in video and align it seamlessly with video inputs. This approach retains the motion capture of motion data and the auxiliary information provided by video, enriching the understanding of dynamic scenes. + +Evaluation on BABEL-QA. We evaluated LLaMo on the BABEL-QA dataset to assess its capability to handle complex motion-based queries. As shown in Table 2, LLaMo achieves the highest overall score of 0.458, outperforming all baselines. It excels particularly in the Action category with a score of 0.525, demonstrating superior action recognition capabilities. Additionally, it achieves top scores in temporal reasoning tasks (Before and After), indicating a + +![](images/21e8b7bec3a137a35036556a5f4d6ca28740a73bdf86b7affb8eef808260b72a.jpg) +e e + +![](images/308b04019fbdaaf485a6f47cea714dc2fa91b12e073a31bfdfe18fbb56dfcf21.jpg) + +![](images/e8387f7b304d4d04ab54c316bf28e4953df442d47b1c1d40616b51882572d02d.jpg) + +![](images/772529eca92b7d5d18040a0ad6aa554a2d80c39bb16f4357cb90dd7ea3133dd3.jpg) + +![](images/68edf55cc48e16ae54ce913a517b4a80c72688b34871bce62471906dd2472e0a.jpg) + +![](images/0ae195acc9f7b15ff2081792485f2d8dd6c7729044f9d78c2d78f423d8806a97.jpg) + +![](images/1f3f8bf582d6a4f5651e5036f0451e9f4f97f94e301ad6b8f30ea25f7d38c92d.jpg) +# + +Describe the following human motion video concisely, including but not limited to what actions are being performed, the standards followed, and the characteristics. + +The person is practicing a Tai Chi movement called "Golden Rooster Stands on One Leg." The movement starts from a wide stance with extended arms and gradually transitions into a single-leg balance, with the other knee lifted and the hands drawn close to the body. The entire process is slow and controlled, with smooth and continuous motions, emphasizing the integration of mind and body, showcasing the characteristics of Tai Chi. + +![](images/0d97bb1db2bca601221d1e2abf22ae3355e5c8eb1d3821a0c3a3039283d7e8e6.jpg) + +![](images/a816d961af460a53aacb302c19591bfd13830226e517b2bf8e269ec831b3c6ce.jpg) + +![](images/a2dec7880660cc958862edae3a8d3bef710de30643fcf191b216786537369b73.jpg) +Figure 4. Example outputs from LLaMo across human activities and professional sports, showcasing its reasoning capabilities and domain-specific knowledge in motion-intensive scenarios. + +![](images/56370119188f0947e59157efbc6062889103e27afd4e5f7da97d58831cfa36bf.jpg) + +![](images/2ca5dbd12e367ff59fa0c4951037e02d07f6db1f0315be09a70929c3c78e2794.jpg) + +What is the person in the video doing? The person in the video is making a phone +Can you analyze her current emotions? Sure, she seems a little impatient and angry +Why do you think she's angry? Could it be that she's feeling surprised instead? + +Determining whether a person is angry can be based on their body language. She is likely feeling angry because she stopped and stomped her foot while walking, and her body movements became somewhat sluggish after hanging up the phone. + +How many push-ups did the person do in the video? + +The person in the video did three push-ups. Is it definitely 3 times? Why do I remember it as 5 times? + +Let me observe carefully again. It's likely that the smaller range of motion led to your misjudgment. Indeed, he did three push-ups, with three sets of repeated movements. + +What should be the position of the knees in golf technique? + +GT: You want to keep your knees a little wider and more stable in golf technique. + +LLaMo: Keep knees slightly bent and stable, with the left knee moving forward during the downswing. + +What joint angles are important at the top of the swing for elite golfers? + +GT: Most elite golfers have joint angles ranging from 35 to 45 degrees at the top of the swing. + +LLaMo: At the top of the swing, elite golfers have hips at 45 degrees and wrists at 90 degrees. + +Table 1. Expected Comparison on the MoVid-Bench. The top part of the table presents motion-related results, and the bottom part presents video-related results. Higher accuracy and score values indicate better performance. + +
MoVid-Bench-MotionBody.Seq.Dir.Rea.Hall.All
Acc.ScoreAcc.ScoreAcc.ScoreAcc.ScoreAcc.ScoreAcc.Score
GT100.005.00100.005.00100.005.00100.005.00100.005.00100.005.00
GPT-3.5 [41]24.512.0430.412.2527.142.1939.192.6458.333.2231.332.31
MotionGPT [2]31.223.9842.693.1644.293.5035.813.0616.662.2536.863.11
MotionLLM [1]50.493.5536.843.1458.573.7652.703.5855.563.3949.503.49
LLaMo59.304.0144.013.1260.913.9958.213.6461.173.5355.323.67
MoVid-Bench-VideoBody.Seq.Dir.Rea.Hull.All
Acc.ScoreAcc.ScoreAcc.ScoreAcc.ScoreAcc.ScoreAcc.Score
GT100.005.00100.005.00100.005.00100.005.00100.005,00100.005.00
GPT-3.5 [41]2.401.231.391.004.651.095.411.650.000.943.031.26
Video-LLAVA [39]33.532.7625.462.7241.862.8452.973.2858.831.8942.532.70
MotionLLM [1]34.132.9332.872.9244.183.1463.203.5570.592.3049.002.97
LLaMo33.832.8536.013.1145.503.3267.593.7372.812.2552.333.10
+ +strong understanding of temporal relationships in motion sequences. This capability benefits from the direct integration of motion data as an independent modality and the effective text-aware motion frames extract mechanism in LLaMo, preserving crucial motion-specific nuances, as well as leading to more accurate and contextually rich interpretations. + +Professional Sports Analysis. We further evaluate LLaMo on the challenging swing-dataset, termed as + +profession-swing, details refer to Appendix, where the ground truth (GT) consists of answers provided by professional coaches, serving as expert benchmarks. The evaluation focuses on four key indicators: Reasonableness: Logical soundness and plausibility of the responses. Coherence: Consistency and logical flow within the responses. Pertinence: Relevance of the answers to the specific questions. Adaptability: Ability to adjust responses based on the stage on the athlete. + +As presented in Table 3, LLaMo significantly outper + +Table 2. Comparison on BABEL-QA dataset. Higher scores indicate better performance. + +
ModelPred. typeOverall ↑Action ↑Direction ↑Body Part ↑Before ↑After ↑Other ↑
MotionCLIP-M [41]cls.0.4300.4850.3610.2720.3720.3210.404
MotionCLIP-R [41]cls.0.4200.4890.3100.2500.3980.3140.387
MotionLLM [41]gen.0.4360.5170.3540.1540.4270.3680.529
LLaMogen.0.4580.5250.3980.2240.4430.3920.518
+ +Table 3. Performance on the profession-swing dataset across four indicators. Higher accuracy and score values indicate better performance. + +
ModelReasonablenessCoherencePertinenceAdaptabilityAll
AccScoreAccScoreAccScoreAccScoreAccScore
GT10051005100510051005
GPT-3.5 [41]4.510.834.200.743.650.661.710.223.420.61
MotionGPT [2]10.531.2515.421.6612.641.114.791.3514.351.4
MotionLLM [1]12.331.5119.201.8717.981.7910.201.2216.531.57
LLaMo21.102.1127.102.7131.813.1220.221.9824.802.48
+ +forms both GPT-3.5 [41] and MotionLLM [1] across all metrics. While MotionLLM shows improvements over GPT-3.5, with higher accuracy and scores, it still lags behind LLaMo. Specifically, LLaMo achieves an overall accuracy of 24.80 + +The superior performance of LLaMo can be attributed to the design of the Cross Talker module, which enables language-guided frame selection to capture complex motion relationships inherent in professional swings. This allows LLaMo to effectively model intricate motion patterns and generate responses that are more reasonable, coherent, pertinent, and adaptable, closely aligning with the expert-level insights provided by professional coaches. + +In contrast, while MotionLLM demonstrates certain strengths over MotionGPT—such as better pertinence and coherence. It lacks the advanced motion understanding capabilities of LLaMo, lagging behind LLaMo distinctly. + +Table 4. Motion and video details capture evaluation on MoRepCount + +
ModelOBOMAEOBZRMSE
EScounts [42]0.3970.2910.1985.58
PoseRAC [43]0.3820.3120.2045.95
TransRAC [38]0.2760.4440.1058.56
RepNet [44]0.009///
LLaMo0.3890.3240.2226.15
+ +Evaluation on Mo-RepCount. We evaluated LLaMo on the Mo-RepCount dataset to assess its ability to capture fine-grained motion details and complex temporal features required for accurate repetition counting. For a fair comparison, we trained all state-of-the-art models, including RepNet [44], TransRAC [38], PoseRAC [43] and ES + +counts [42], on the Mo-RepCount dataset without using any additional data. As presented in Table 4, LLaMo outperforms the RepNet and TransRAC across almost all metrics. Since both of them are models built for loop counting tasks, the effect of our model on this task is convincing, which shows that LLaMo is also competent in capturing the details of the motion and videos. + +While LLaMo did not outperform the specialized counting models EScounts and poseRAC on many metrics, it excelled on the OBZ metric, demonstrating an unmatched ability to capture fine-grained motion details and spatiotemporal relationships. These results confirm LLaMo's strong potential in motion and video analysis, underscoring its suitability for detail-oriented tasks and applications. + +# 5. Conclusion + +We introduce LLaMo, a novel framework that integrates motion data directly into large language models, thereby enhancing multimodal understanding. By unifying text, video, and motion data in an architecture fine-tuned for human motion instruction, LLaMo achieves notable progress in behavior comprehension, motion-focused captioning, and interactive Q&A, capturing the fine-grained motion details vital for precise, context-aware insights. Experimental results, especially in professional sports analysis, demonstrate LLaMo's capacity to interpret complex human motion with accuracy. Future work will extend LLaMo to a broad range of domains where detailed motion analysis and multimodal understanding are essential. The main challenges lie in refining cross-modal integration techniques and improving efficiency for real-time applications, thus broadening LLaMo's potential in human-centered AI. We hope our work will spur further exploration into more advanced human-centric multimodal models. + +# Acknowledgement + +This work was supported in part by the Pioneer Centre for AI, DNRF grant number P1. + +# References + +[1] Ling-Hao Chen, Shunlin Lu, Ailing Zeng, Hao Zhang, Benyou Wang, Ruimao Zhang, and Lei Zhang. Motionllm: Understanding human behaviors from human motions and videos. arXiv preprint arXiv:2405.20340, 2024. 1, 2, 6, 7, 8 +[2] Ling-Hao Chen, Jiawei Zhang, Wen Liu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. In Proceedings of the Conference on Neural Information Processing Systems (NeurIPS), 2024. 1, 2, 6, 7, 8 +[3] Matthias Plappert, Christian Mandery, and Tamim Asfour. The kit motion-language dataset. Big data, 4(4):236-252, 2016. 1, 6 +[4] Yan Zhang, Michael J Black, and Siyu Tang. We are more than our joints: Predicting how 3d bodies move. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. +[5] Fangzhou Hong, Liang Pan, Zhongang Cai, and Ziwei Liu. Versatile multi-modal pre-training for human-centric perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16156-16166, 2022. 2 +[6] Haoxuan Qu, Yujun Cai, and Jun Liu. Llms are good action recognizers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18395-18406, 2024. 1 +[7] Shan Chen, Jiale Zhou, and Lei Li. Dense point clouds matter: Dust-gs for scene reconstruction from sparse viewpoints. arXiv preprint arXiv:2409.08613, 2024. 1 +[8] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 1, 2 +[9] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 2 +[10] Li Lei, Jia Sen, Wang Jianhao, An Zhaochong, Li Jiaang, Hwang Jenq-Neng, and Belongie Serge. Chatmotion: A multimodal multi-agent for human motion analysis. arXiv preprint arXiv:2502.18180, 2025. 1 +[11] Mark Endo, Joy Hsu, Jiaman Li, and Jiajun Wu. Motion question answering via modular motion programs. In Proceedings of the International Conference on Machine Learning (ICML), 2023. 1, 2 +[12] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 1 +[13] Yaya Shi, Haiyang Xu, Chunfeng Yuan, Bing Li, Weiming Hu, and Zheng-Jun Zha. Learning video-text aligned repre + +sentations for video captioning. IEEE Transactions on Multimedia (TMM), 2023. 2 +[14] Chuan Guo, Xinxin Zuo, Sen Wang, and Li Cheng. Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In Proceedings of the European Conference on Computer Vision (ECCV), pages 580-597, 2022. 2 +[15] Kunchang Li, Yali Wang, Yinan He, et al. Mvbench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2 +[16] Zhaochong An, Guolei Sun, Yun Liu, Runjia Li, Min Wu, Ming-Ming Cheng, Ender Konukoglu, and Serge Belongie. Multimodality helps few-shot 3d point cloud semantic segmentation. arXiv preprint arXiv:2410.22489, 2024. 2 +[17] Bobby Bodenheimer, Chuck Rose, Seth Rosenthal, and John Pella. The process of motion capture: Dealing with the data. In Proceedings of the Eurographics Workshop on Computer Animation and Simulation (EG Workshop), pages 3-18, 1997. 2 +[18] Xiangyu Peng, Yuxuan Zhang, Xiaolong Wang, and Ziwei Liu. Text-driven human motion generation with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2 +[19] Mark Endo, Joy Hsu, Jiaman Li, and Jiajun Wu. Motion question answering via modular motion programs. In International Conference on Machine Learning, pages 9312-9328. PMLR, 2023. 2, 6 +[20] Jie Lin, Wei Zhang, and Zhe Chen. Videollm: Language models for video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2405-2415, 2023. 2 +[21] Xu Ning, Hongwei Li, and Yu Zhao. Videobench: A benchmark for large-scale video understanding. arXiv preprint arXiv:2304.12345, 2023. 2 +[22] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 2, 6 +[23] Lei Li. Segment any building. In Computer Graphics International Conference, pages 155-166. Springer, 2023. 2 +[24] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 2 +[25] Enxin Song, Guanhong Zhang, and Haoyang Zhou. Adaptive multimodal learning for behavior analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2 +[26] Yunhua Yang, Ziwang Zhao, and Yiming Xie. Recognizing human behaviors with skeletal data in llm-based frameworks. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. +[27] Lei Li, Tianfang Zhang, Zhongfeng Kang, and Xikun Jiang. Mask-fpan: Semi-supervised face parsing in the wild with de-occlusion and uv gan. Computers & Graphics, 116:185–193, 2023. 2 + +[28] Matthias Plappert, Christian Mandery, and Tamim Asfour. Learning a bidirectional mapping between human whole-body motion and natural language using deep recurrent neural networks. Robotics and Autonomous Systems (RAS), 109: 13-26, 2018. 2 +[29] Lei Li. Cpseg: Finer-grained image semantic segmentation via chain-of-thought language prompting. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 513-522, 2024. +[30] Yunhua Yang, Liang Zhang, and Hui Li. Understanding human behaviors from skeletal data: A review of datasets and methods. arXiv preprint arXiv:2310.12998, 2023. 2 +[31] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Two-stream adaptive graph convolutional networks for skeleton-based action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12026-12035, 2019. 2 +[32] Enxin Song, Wenhao Chai, and Yucheng Zhang. Fine-grained spatial-temporal motion understanding in complex video environments. arXiv preprint arXiv:2310.08639, 2023. 2 +[33] Zhaochong An, Guolei Sun, Zongwei Wu, Hao Tang, and Luc Van Gool. Temporal-aware hierarchical mask classification for video semantic segmentation. arXiv preprint arXiv:2309.08020, 2023. 2 +[34] Xianghong Liu, Haoxuan Wang, Zhiwei Zhang, Huan Wu, Baoquan Chen, and Xiaoguang Han. Category-agnostic pose estimation for point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3 +[35] Xiaolong He, Yi Zhang, Chen Chen, and Kyoung Mu Lee. Activitynet++: A large-scale benchmark for video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. +[36] Lei Li. Hierarchical edge aware learning for 3d point cloud. In Computer Graphics International Conference, pages 81-92. Springer, 2023. 3 +[37] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5152-5161, 2022. 6 +[38] Huazhang Hu, Sixun Dong, Yiqun Zhao, Dongze Lian, Zhengxin Li, and Shenghua Gao. Transrac: Encoding multiscale temporal correlation with transformers for repetitive action counting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19013-19022, 2022. 6, 8 +[39] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 6, 7 +[40] Peng Jin, Ryuichi Takanobu, Wancai Zhang, Xiaochun Cao, and Li Yuan. Chat-univi: Unified visual representation empowers large language models with image and video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13700-13710, 2024. 6 + +[41] OpenAI. Gpt-3.5: Generative pre-trained transformer 3.5, 2023. https://platform.openai.com/docs/models/gpt-3-5.6,7,8 +[42] Saptarshi Sinha, Alexandros Stergiou, and Dima Damen. Every shot counts: Using exemplars for repetition counting in videos. arXiv preprint arXiv:2403.18074, 2024. 8 +[43] Ziyu Yao, Xuxin Cheng, and Yuexian Zou. Poserac: Pose saliency transformer for repetitive action counting. arXiv preprint arXiv:2303.08450, 2023. 8 +[44] Bastian Wandt and Bodo Rosenhahn. Repnet: Weakly supervised training of an adversarial reprojection network for 3d human pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7782-7791, 2019. 8 \ No newline at end of file diff --git a/2025/Human Motion Instruction Tuning/images.zip b/2025/Human Motion Instruction Tuning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e90536f743ba5151bd45ca51ad319009658476b7 --- /dev/null +++ b/2025/Human Motion Instruction Tuning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1d0b7711db27d73499eaeb27d3689d19f814c59598c9ba04ed78c22f4c1d8137 +size 491460 diff --git a/2025/Human Motion Instruction Tuning/layout.json b/2025/Human Motion Instruction Tuning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6494b2e01b8b8feea2d4c550a10665e9b7257793 --- /dev/null +++ b/2025/Human Motion Instruction Tuning/layout.json @@ -0,0 +1,8952 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 197, + 103, + 414, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 103, + 414, + 121 + ], + "spans": [ + { + "bbox": [ + 197, + 103, + 414, + 121 + ], + "type": "text", + "content": "Human Motion Instruction Tuning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 152, + 475, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 152, + 475, + 194 + ], + "spans": [ + { + "bbox": [ + 134, + 152, + 475, + 194 + ], + "type": "text", + "content": "Lei Li" + }, + { + "bbox": [ + 134, + 152, + 475, + 194 + ], + "type": "inline_equation", + "content": "^{1,2,\\ast,\\dagger}" + }, + { + "bbox": [ + 134, + 152, + 475, + 194 + ], + "type": "text", + "content": ", Sen Jia" + }, + { + "bbox": [ + 134, + 152, + 475, + 194 + ], + "type": "inline_equation", + "content": "^{1,\\ast}" + }, + { + "bbox": [ + 134, + 152, + 475, + 194 + ], + "type": "text", + "content": ", Jianhao Wang" + }, + { + "bbox": [ + 134, + 152, + 475, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 134, + 152, + 475, + 194 + ], + "type": "text", + "content": ", Zhongyu Jiang" + }, + { + "bbox": [ + 134, + 152, + 475, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 134, + 152, + 475, + 194 + ], + "type": "text", + "content": ", Feng Zhou" + }, + { + "bbox": [ + 134, + 152, + 475, + 194 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 134, + 152, + 475, + 194 + ], + "type": "text", + "content": ", Ju Dai" + }, + { + "bbox": [ + 134, + 152, + 475, + 194 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 134, + 152, + 475, + 194 + ], + "type": "text", + "content": ", Tianfang Zhang" + }, + { + "bbox": [ + 134, + 152, + 475, + 194 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 134, + 152, + 475, + 194 + ], + "type": "text", + "content": ", Zongkai Wu" + }, + { + "bbox": [ + 134, + 152, + 475, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 134, + 152, + 475, + 194 + ], + "type": "text", + "content": ", Jenq-Neng Hwang" + }, + { + "bbox": [ + 134, + 152, + 475, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 210, + 200, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 210, + 200, + 224 + ], + "spans": [ + { + "bbox": [ + 151, + 210, + 200, + 224 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 236, + 296, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 236, + 296, + 453 + ], + "spans": [ + { + "bbox": [ + 54, + 236, + 296, + 453 + ], + "type": "text", + "content": "This paper presents LLaMo (Large Language and Human Motion Assistant), a multimodal framework for human motion instruction tuning. In contrast to conventional instruction-tuning approaches that convert non-linguistic inputs, such as video or motion sequences, into language tokens, LLaMo retains motion in its native form for instruction tuning. This method preserves motion-specific details that are often diminished in tokenization, thereby improving the model's ability to interpret complex human behaviors. By processing both video and motion data alongside textual inputs, LLaMo enables a flexible, human-centric analysis. Experimental evaluations across high-complexity domains, including human behaviors and professional activities, indicate that LLaMo effectively captures domain-specific knowledge, enhancing comprehension and prediction in motion-intensive scenarios. We hope LLaMo offers a foundation for future multimodal AI systems with broad applications, from sports analytics to behavioral prediction." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 477, + 135, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 477, + 135, + 490 + ], + "spans": [ + { + "bbox": [ + 55, + 477, + 135, + 490 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 498, + 296, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 654 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 654 + ], + "type": "text", + "content": "Understanding human motion is a central challenge in multimodal AI, impacting numerous fields such as digital human, human-computer interaction, sports analytics, healthcare, and virtual human modeling [3-6]. Motion data, which captures skeletal movements and body dynamics, provides a structured, appearance-invariant representation of human actions, concentrating on the essential movement patterns while excluding irrelevant visual details [7]. This data has emerged as a privacy-conscious alternative to visual inputs, offering finer control over action recognition and facilitating analysis in applications ranging from digital human avatars to behavioral monitoring in healthcare [8-10]. As digital human representations and motion-centric" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 315, + 209, + 556, + 373 + ], + "blocks": [ + { + "bbox": [ + 315, + 209, + 556, + 373 + ], + "lines": [ + { + "bbox": [ + 315, + 209, + 556, + 373 + ], + "spans": [ + { + "bbox": [ + 315, + 209, + 556, + 373 + ], + "type": "image", + "image_path": "6929f4853c166dfc7e60a280057512ce73bd37d5595f226416333ac6f813430e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 380, + 555, + 437 + ], + "lines": [ + { + "bbox": [ + 313, + 380, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 555, + 437 + ], + "type": "text", + "content": "Figure 1. A comparison of MotionLLM [1], MotionGPT [2], and LLaMo highlights LLaMo's motion-specific capabilities. Equipped with a Motion Enhancer and Cross Talker module to align motion and text, LLaMo supports both video and motion inputs, enabling text-aware, fine-grained motion analysis." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 460, + 555, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 555, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 555, + 508 + ], + "type": "text", + "content": "applications continue to expand, developing models that can leverage the unique characteristics of motion data becomes increasingly crucial for achieving robust, context-aware behavior analysis." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 509, + 556, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 556, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 556, + 676 + ], + "type": "text", + "content": "Recently, large language models (LLMs) have shown promise in motion analysis, leveraging their capacity to model temporal dependencies across multimodal data [1, 11]. State-of-the-art models, such as MotionGPT [12] and MotionLLM [1], treat motion data as a multimodal extension, encoding it as discrete tokens or translating it into language-like representations for processing by LLMs. These models have yielded notable results in motion understanding and recognition tasks, illustrating the adaptability of language models to non-linguistic data. However, their reliance on tokenization or textual conversion can limit the richness of the motion representation, abstracting away critical spatial-temporal details essential for high-fidelity human activity recognition." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "Despite recent advancements, notable challenges remain in current approaches to motion analysis using LLMs. First, encoding motion data into language tokens, as implemented" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 664, + 229, + 674 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 664, + 229, + 674 + ], + "spans": [ + { + "bbox": [ + 69, + 664, + 229, + 674 + ], + "type": "text", + "content": "* These authors contributed equally to this work." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 69, + 674, + 228, + 684 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 674, + 228, + 684 + ], + "spans": [ + { + "bbox": [ + 69, + 674, + 228, + 684 + ], + "type": "text", + "content": "† Corresponding Author. (lilei@di.ku.dk)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 684, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 684, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 684, + 295, + 713 + ], + "type": "text", + "content": "1University of Washington 2University of Copenhagen 3Xi'an Jiaotong University 4North China University of Technology 5Peng Cheng Laboratory 6Tsinghua University" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "17582" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 276 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 276 + ], + "type": "text", + "content": "in models such as MotionGPT, relies on a quantization process that may obscure critical fine-grained motion information, potentially limiting the model's capacity to accurately capture detailed spatial-temporal dynamics [13, 14]. The common approach of translating motion data into text tokens further contributes to this issue, often resulting in the loss of essential motion-specific nuances that are critical for in-depth behavior interpretation [11, 15]. This translation process poses considerable challenges in understanding the intricate details of human movement, impeding the model's ability to achieve spatial and physical comprehension of human actions. Moreover, motion data inherently contains 3D structural information and insights into connected human behaviors, which are often difficult to fully capture when relying solely on language tokens, as this representation can overlook the underlying spatial dependencies crucial to modeling human motion accurately." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 278, + 295, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 278, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 278, + 295, + 456 + ], + "type": "text", + "content": "In addition, existing methods frequently treat video and motion data as isolated modalities, overlooking the potential benefits of a complementary, integrated approach [16]. Processing these data types independently neglects the rich environmental context and interaction cues provided by video inputs, which are crucial for accurately interpreting complex human behaviors [5, 8]. This separation limits the capacity for a holistic understanding of human motion, particularly in scenarios where context, spatial relations, and dynamic interactions play essential roles in behavior analysis. Furthermore, video-based approaches, while potentially more informative, are computationally intensive, posing significant constraints for large-scale or real-time applications, which limits their feasibility in practical motion analysis systems." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 460, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 628 + ], + "type": "text", + "content": "In response to these challenges, we propose LLaMo, a framework designed to integrate motion data as a distinct modality within LLMs without translating it into textual intermediaries. LLaMo incorporates a Motion Estimator and an Enhancer to facilitate the direct input of motion or/and video data, thereby offering a unified, human-centric analysis pipeline. The framework also introduces a Cross Talker module, which dynamically aligns the motion and text features, allowing for text-aware, fine-grained motion representations. This approach retains the original spatial-temporal characteristics of the motion data, providing a more precise understanding of human actions and enabling high-resolution behavior analysis in motion-intensive contexts [17, 18]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "content": "Through extensive evaluations on benchmarks such as MoVid-Bench [1] and BABEL-QA [19], we demonstrate that LLaMo achieves state-of-the-art performance in motion-centric human activity recognition [11, 15]. These results underscore LLaMo's potential to advance multimodal AI, providing a robust foundation for human behavior understanding. Our contributions can be summarized as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 314, + 73, + 350, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 73, + 350, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 73, + 350, + 83 + ], + "type": "text", + "content": "follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 85, + 553, + 228 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 314, + 85, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 85, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 314, + 85, + 553, + 120 + ], + "type": "text", + "content": "- We introduce LLaMo, which treats motion data as an independent modality within LLMs, preserving essential motion-specific details for robust analysis." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 121, + 553, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 121, + 553, + 168 + ], + "spans": [ + { + "bbox": [ + 314, + 121, + 553, + 168 + ], + "type": "text", + "content": "- We propose Cross Talker, a text-guided mechanism that dynamically focuses on and aggregates key motion frames features, optimizing computational efficiency and enabling the model to pick relevant motion features." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 168, + 553, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 168, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 314, + 168, + 553, + 228 + ], + "type": "text", + "content": "- LLaMo's architecture supports the direct input of both raw motion and video data, providing a generalized framework for human-centric analysis across diverse applications, such as sports analytics, healthcare, and behavioral monitoring." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 240, + 400, + 252 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 240, + 400, + 252 + ], + "spans": [ + { + "bbox": [ + 314, + 240, + 400, + 252 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 260, + 542, + 273 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 260, + 542, + 273 + ], + "spans": [ + { + "bbox": [ + 314, + 260, + 542, + 273 + ], + "type": "text", + "content": "2.1. Human-Centric Multimodal Representation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 279, + 555, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 279, + 555, + 531 + ], + "spans": [ + { + "bbox": [ + 313, + 279, + 555, + 531 + ], + "type": "text", + "content": "Multimodal representation learning is integral to human-centric analysis, especially for applications requiring spatial-temporal reasoning to decode complex behaviors [20-22]. Recent advances, such as Video-LLaVA, have made significant strides by embedding visual data from images and videos into a shared linguistic feature space, enabling sophisticated visual reasoning for behavioral analysis tasks [20, 23]. Despite these achievements, many current models are optimized primarily for static images or discrete video frames, limiting their effectiveness in sequential, dynamic scenarios where understanding progression and continuity in motion is crucial [9, 21, 24]. Privacy concerns further challenge the adoption of video-based methods in sensitive domains; thus, researchers are exploring motion data as a privacy-conscious alternative, enabling analysis focused on human actions without revealing identifiable visual information [25-27]. By combining visual and motion data, emerging multimodal frameworks hold promise for more comprehensive, privacy-aware human behavior analysis, leveraging the strengths of both modalities for adaptability and depth across applications." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 539, + 481, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 539, + 481, + 552 + ], + "spans": [ + { + "bbox": [ + 314, + 539, + 481, + 552 + ], + "type": "text", + "content": "2.2. Human Motion Understanding" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": "Human motion understanding has traditionally relied on skeletal data represented as sequences of joint keypoints to capture movement while preserving user privacy [13, 28-30]. Early models like 2s-AGCN [31] and more recent transformer-based methods, such as MotionCLIP [2], have achieved success in tasks like activity recognition, captioning, and behavior analysis by mapping motion data to language tokens, effectively capturing structural aspects of movement. However, these methods often lack environmental context, which is crucial for nuanced interpretations in real-world applications, as similar movements can convey different meanings across varied scenarios [24, 32, 33]. To address this limitation, recent frameworks inte" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17583" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "content": "grate motion and visual data, allowing models to generalize across diverse, dynamic environments [34-36]. This fusion of skeletal data with video-provided context has shown promise in applications like sports and professional activity analysis, where contextual cues are essential. Additionally, skeletal-based data remains valuable in privacy-sensitive applications, such as healthcare and security, where it allows for robust analysis without compromising personal identity." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 190, + 116, + 202 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 190, + 116, + 202 + ], + "spans": [ + { + "bbox": [ + 55, + 190, + 116, + 202 + ], + "type": "text", + "content": "3. Methods" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 209, + 295, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 209, + 295, + 269 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 295, + 269 + ], + "type": "text", + "content": "The LLaMo framework is composed of three primary modules to effectively process and integrate video, motion data and text. An overview of this pipeline is provided in Figure 2. The architecture comprises the following components:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 270, + 295, + 451 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 56, + 270, + 295, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 270, + 295, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 270, + 295, + 342 + ], + "type": "text", + "content": "1. Multimodal Feature Extraction Module: This module independently encodes the video or motion inputs, which makes LLaMo is a general human motion assistant with video input or motion input. Then the Motion features are enhanced by distilling valuable representations in video." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 342, + 295, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 295, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 295, + 403 + ], + "type": "text", + "content": "2. Cross Talker: Here, the enhanced motion features are aggregated in a text-guided manner and aligned with text within a shared semantic space, allowing the natural feature fusion between motion and text, which provides fine-grained inputs for LLM." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 403, + 295, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 295, + 451 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 295, + 451 + ], + "type": "text", + "content": "3. Behavior Generation Module: Using the aggregated and text-aligned motion features, along with the text representations, this module generates a contextually aware textual description of the observed human behavior." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 457, + 291, + 469 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 457, + 291, + 469 + ], + "spans": [ + { + "bbox": [ + 55, + 457, + 291, + 469 + ], + "type": "text", + "content": "3.1. Enhanced motion Feature Extraction Module" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 475, + 295, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 295, + 606 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 295, + 606 + ], + "type": "text", + "content": "The enhanced motion feature extraction module consists of four primary components: the motion estimator, followed by the motion and video encoders, and the feature enhancement module. The motion estimator addresses scenarios where motion data may be unavailable by estimating motion information directly from video frames. The motion and video encoders then independently process the motion and video data, encoding them into their respective feature spaces. This approach allows the system to leverage complementary information from both modalities, creating a richer representation of human behavior." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 607, + 295, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 607, + 295, + 632 + ], + "spans": [ + { + "bbox": [ + 55, + 607, + 295, + 632 + ], + "type": "text", + "content": "Given a sequence of video frames " + }, + { + "bbox": [ + 55, + 607, + 295, + 632 + ], + "type": "inline_equation", + "content": "V = \\{v_{1}, v_{2}, \\ldots, v_{T}\\}" + }, + { + "bbox": [ + 55, + 607, + 295, + 632 + ], + "type": "text", + "content": ", the video encoder " + }, + { + "bbox": [ + 55, + 607, + 295, + 632 + ], + "type": "inline_equation", + "content": "f_{v}(\\cdot)" + }, + { + "bbox": [ + 55, + 607, + 295, + 632 + ], + "type": "text", + "content": " extracts a set of visual features:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 80, + 638, + 270, + 651 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 638, + 270, + 651 + ], + "spans": [ + { + "bbox": [ + 80, + 638, + 270, + 651 + ], + "type": "interline_equation", + "content": "F _ {V} = f _ {v} (V) = \\left\\{f _ {v} \\left(v _ {1}\\right), f _ {v} \\left(v _ {2}\\right), \\dots , f _ {v} \\left(v _ {T}\\right) \\right\\}.", + "image_path": "055c05d77d67838b59fa6b557bc170f5b9bb4b7f0385668fb3549362341aeabd.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 658, + 295, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 658, + 295, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 658, + 295, + 694 + ], + "type": "text", + "content": "Simultaneously, the motion encoder " + }, + { + "bbox": [ + 55, + 658, + 295, + 694 + ], + "type": "inline_equation", + "content": "f_{m}(\\cdot)" + }, + { + "bbox": [ + 55, + 658, + 295, + 694 + ], + "type": "text", + "content": " processes the motion data " + }, + { + "bbox": [ + 55, + 658, + 295, + 694 + ], + "type": "inline_equation", + "content": "M = \\{m_1, m_2, \\ldots, m_T\\}" + }, + { + "bbox": [ + 55, + 658, + 295, + 694 + ], + "type": "text", + "content": " to generate motion-specific features:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 67, + 700, + 285, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 700, + 285, + 715 + ], + "spans": [ + { + "bbox": [ + 67, + 700, + 285, + 715 + ], + "type": "interline_equation", + "content": "F _ {M} = f _ {m} (M) = \\left\\{f _ {m} \\left(m _ {1}\\right), f _ {m} \\left(m _ {2}\\right), \\dots , f _ {m} \\left(m _ {T}\\right) \\right\\}.", + "image_path": "e49ba7eb44ac91491ac6a2417c9361fc9493deadf96456d4d8ef84dac9da1829.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 71, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 71, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 553, + 120 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 71, + 553, + 120 + ], + "type": "inline_equation", + "content": "F_{V}\\in \\mathbb{R}^{T\\times H}" + }, + { + "bbox": [ + 313, + 71, + 553, + 120 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 71, + 553, + 120 + ], + "type": "inline_equation", + "content": "F_{M}\\in \\mathbb{R}^{T\\times H}" + }, + { + "bbox": [ + 313, + 71, + 553, + 120 + ], + "type": "text", + "content": " represent the extracted feature sets for the video and motion data, respectively, where " + }, + { + "bbox": [ + 313, + 71, + 553, + 120 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 71, + 553, + 120 + ], + "type": "text", + "content": " is the sequence length and " + }, + { + "bbox": [ + 313, + 71, + 553, + 120 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 71, + 553, + 120 + ], + "type": "text", + "content": " denotes the feature dimension." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 122, + 555, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 122, + 555, + 243 + ], + "spans": [ + { + "bbox": [ + 313, + 122, + 555, + 243 + ], + "type": "text", + "content": "The feature enhancement module then aligns and fuses these modality-specific features. By using the motion features " + }, + { + "bbox": [ + 313, + 122, + 555, + 243 + ], + "type": "inline_equation", + "content": "F_{M}" + }, + { + "bbox": [ + 313, + 122, + 555, + 243 + ], + "type": "text", + "content": " as queries, this module selectively extracts semantically relevant information from the video features " + }, + { + "bbox": [ + 313, + 122, + 555, + 243 + ], + "type": "inline_equation", + "content": "F_{V}" + }, + { + "bbox": [ + 313, + 122, + 555, + 243 + ], + "type": "text", + "content": ". This alignment and fusion process allows the motion features to be enriched with contextual information from the video, producing refined, contextually enriched representations that enhance the model's understanding of nuanced human behaviors. For the detailed design refer to the Appendix." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 244, + 556, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 244, + 556, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 244, + 556, + 389 + ], + "type": "text", + "content": "First, the video features and motion features will conduct self-attention operations respectively to get their augmented features " + }, + { + "bbox": [ + 313, + 244, + 556, + 389 + ], + "type": "inline_equation", + "content": "F_V'" + }, + { + "bbox": [ + 313, + 244, + 556, + 389 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 244, + 556, + 389 + ], + "type": "inline_equation", + "content": "F_M'" + }, + { + "bbox": [ + 313, + 244, + 556, + 389 + ], + "type": "text", + "content": ". This allows the video and motion to initially capture the dependencies on their respective time steps. Then, in order to enrich the motion data features with semantic information from the video, we subtly design a cross-attention block where augmented motion features as the query to extract valid semantic information in the video features. Finally, the selected video semantic representation and motion features are connected by residuals followed by a feed-forward neural network to obtain the enhanced motion features " + }, + { + "bbox": [ + 313, + 244, + 556, + 389 + ], + "type": "inline_equation", + "content": "\\tilde{F_M}" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 403, + 436, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 403, + 436, + 415 + ], + "spans": [ + { + "bbox": [ + 314, + 403, + 436, + 415 + ], + "type": "text", + "content": "3.2. Cross Talker Module" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 423, + 555, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 555, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 555, + 581 + ], + "type": "text", + "content": "Given the enhanced motion features " + }, + { + "bbox": [ + 313, + 423, + 555, + 581 + ], + "type": "inline_equation", + "content": "\\tilde{F}_M" + }, + { + "bbox": [ + 313, + 423, + 555, + 581 + ], + "type": "text", + "content": " from the feature enhancement module, directly feeding the entire sequence of motion frames into the Large Language Model (LLM) incurs significant computational costs and may lead to suboptimal performance. This limitation arises from the self-attention mechanism in transformers, which has a computational complexity of " + }, + { + "bbox": [ + 313, + 423, + 555, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(L^2)" + }, + { + "bbox": [ + 313, + 423, + 555, + 581 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 423, + 555, + 581 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 423, + 555, + 581 + ], + "type": "text", + "content": " denotes the sequence length. When incorporating both motion and text, the sequence length becomes " + }, + { + "bbox": [ + 313, + 423, + 555, + 581 + ], + "type": "inline_equation", + "content": "L = L_{T} + T" + }, + { + "bbox": [ + 313, + 423, + 555, + 581 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 423, + 555, + 581 + ], + "type": "inline_equation", + "content": "L_{T}" + }, + { + "bbox": [ + 313, + 423, + 555, + 581 + ], + "type": "text", + "content": " is the number of text tokens and " + }, + { + "bbox": [ + 313, + 423, + 555, + 581 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 423, + 555, + 581 + ], + "type": "text", + "content": " is the number of motion frames. As " + }, + { + "bbox": [ + 313, + 423, + 555, + 581 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 423, + 555, + 581 + ], + "type": "text", + "content": " increases, the quadratic growth in computational cost becomes prohibitive, leading to complexity proportional to " + }, + { + "bbox": [ + 313, + 423, + 555, + 581 + ], + "type": "inline_equation", + "content": "(L_{T} + T)^{2}" + }, + { + "bbox": [ + 313, + 423, + 555, + 581 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "content": "Furthermore, computing attention over long sequences can dilute attention weights, making it challenging for the model to focus on essential parts of the sequence and effectively capture long-range dependencies. Furthermore, the motion features are not well aligned with text either. To address these issues, we propose the Cross Talker Module, which selectively distills key motion frames—referred to as viewpoint frames—based on their relevance to the text context, as well as aligning the motion features with the text. An overview of the Cross Talker Module is shown in Figure 3." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17584" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 68, + 532, + 291 + ], + "blocks": [ + { + "bbox": [ + 57, + 68, + 532, + 291 + ], + "lines": [ + { + "bbox": [ + 57, + 68, + 532, + 291 + ], + "spans": [ + { + "bbox": [ + 57, + 68, + 532, + 291 + ], + "type": "image", + "image_path": "48347167cdb846970fabb12d7a345502e93e48aaca805da0b7e77a0124f7d964.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 300, + 555, + 335 + ], + "lines": [ + { + "bbox": [ + 54, + 300, + 555, + 335 + ], + "spans": [ + { + "bbox": [ + 54, + 300, + 555, + 335 + ], + "type": "text", + "content": "Figure 2. Overview of the LLaMo framework. It includes three main modules: (1) Multimodal Feature Extraction for encoding video and motion data; (2) Cross Talker for aligning and fusing motion and text features; and (3) Behavior Generation Module to produce text descriptions of human behavior based on integrated features." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 355, + 294, + 587 + ], + "blocks": [ + { + "bbox": [ + 58, + 355, + 294, + 587 + ], + "lines": [ + { + "bbox": [ + 58, + 355, + 294, + 587 + ], + "spans": [ + { + "bbox": [ + 58, + 355, + 294, + 587 + ], + "type": "image", + "image_path": "f9a416491ad6f8c391feccd39022f6b68db29765534857a3c9a8afc333aea68e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 597, + 295, + 631 + ], + "lines": [ + { + "bbox": [ + 55, + 597, + 295, + 631 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 295, + 631 + ], + "type": "text", + "content": "Figure 3. Overview of the Cross Talker Module, which selects key frames based on text guidance and fuses them with text features for enhanced analysis." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "Language-Guided Frame Selection Given the enhanced motion features " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\tilde{F}_M\\in \\mathbb{R}^{T\\times H}" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": " and text embeddings " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "F_{T}\\in" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{L_T\\times H}" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": ", we aim to identify " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": " viewpoint frames that are most relevant to the textual context. This selection is achieved through a cross-attention operation, where text" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 355, + 555, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 355, + 555, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 355, + 555, + 376 + ], + "type": "text", + "content": "features act as queries and motion features as keys and values:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 354, + 376, + 555, + 409 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 376, + 555, + 409 + ], + "spans": [ + { + "bbox": [ + 354, + 376, + 555, + 409 + ], + "type": "interline_equation", + "content": "A = \\operatorname {S o f t m a x} \\left(\\frac {F _ {T} W _ {Q} \\left(\\tilde {F} _ {M} W _ {K}\\right) ^ {\\top}}{\\sqrt {d}}\\right), \\tag {1}", + "image_path": "8b3e2054eb3fa7a58652931940b0b5dbc7fa1d1abe5a88a3416edd5f89821c36.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 412, + 555, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 412, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 412, + 555, + 472 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 412, + 555, + 472 + ], + "type": "inline_equation", + "content": "W_{Q}, W_{K} \\in \\mathbb{R}^{H \\times d}" + }, + { + "bbox": [ + 313, + 412, + 555, + 472 + ], + "type": "text", + "content": " are learnable projection matrices, and " + }, + { + "bbox": [ + 313, + 412, + 555, + 472 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 412, + 555, + 472 + ], + "type": "text", + "content": " is the dimensionality of the queries and keys. The resulting attention matrix " + }, + { + "bbox": [ + 313, + 412, + 555, + 472 + ], + "type": "inline_equation", + "content": "A \\in \\mathbb{R}^{L_{T} \\times T}" + }, + { + "bbox": [ + 313, + 412, + 555, + 472 + ], + "type": "text", + "content": " contains attention weights, with each element " + }, + { + "bbox": [ + 313, + 412, + 555, + 472 + ], + "type": "inline_equation", + "content": "A_{i,j}" + }, + { + "bbox": [ + 313, + 412, + 555, + 472 + ], + "type": "text", + "content": " indicating the relevance between the " + }, + { + "bbox": [ + 313, + 412, + 555, + 472 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 412, + 555, + 472 + ], + "type": "text", + "content": "-th text token and the " + }, + { + "bbox": [ + 313, + 412, + 555, + 472 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 412, + 555, + 472 + ], + "type": "text", + "content": "-th motion frame." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 473, + 555, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 555, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 555, + 509 + ], + "type": "text", + "content": "To determine the importance of each motion frame, we aggregate attention weights across all text tokens using max pooling:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 359, + 517, + 555, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 517, + 555, + 536 + ], + "spans": [ + { + "bbox": [ + 359, + 517, + 555, + 536 + ], + "type": "interline_equation", + "content": "s _ {j} = \\max _ {i = 1, \\dots , L _ {T}} A _ {i, j}, \\quad j = 1, \\dots , T. \\tag {2}", + "image_path": "1aaf834da0739bff146409d42fe61dd6c15d1b58d9abfecdaef751bfe5668d8a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 544, + 556, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 556, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 556, + 628 + ], + "type": "text", + "content": "The resulting scores " + }, + { + "bbox": [ + 313, + 544, + 556, + 628 + ], + "type": "inline_equation", + "content": "s_j" + }, + { + "bbox": [ + 313, + 544, + 556, + 628 + ], + "type": "text", + "content": " represent the significance of each frame with respect to the text context. We then select the top " + }, + { + "bbox": [ + 313, + 544, + 556, + 628 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 544, + 556, + 628 + ], + "type": "text", + "content": " frames with the highest relevance scores as viewpoint frames, reducing the effective sequence length from " + }, + { + "bbox": [ + 313, + 544, + 556, + 628 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 544, + 556, + 628 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 544, + 556, + 628 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 544, + 556, + 628 + ], + "type": "text", + "content": ". This reduction in sequence length lowers the self-attention complexity from " + }, + { + "bbox": [ + 313, + 544, + 556, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{O}((L_T + T)^2)" + }, + { + "bbox": [ + 313, + 544, + 556, + 628 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 544, + 556, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{O}((L_T + K)^2)" + }, + { + "bbox": [ + 313, + 544, + 556, + 628 + ], + "type": "text", + "content": ", which is computationally efficient when " + }, + { + "bbox": [ + 313, + 544, + 556, + 628 + ], + "type": "inline_equation", + "content": "K \\ll T" + }, + { + "bbox": [ + 313, + 544, + 556, + 628 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": "Adaptive Contextual Feature Aggregation For each selected viewpoint frame, we enrich its representation by aggregating both local and global contexts. An adaptive receptive field size " + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "inline_equation", + "content": "r_k" + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": " is predicted for each viewpoint frame " + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": " using a receptive field regression module. The module performs a cross-attention operation where each viewpoint" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "17585" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": "frame serves as a query, and the unselected motion features serve as keys and values, followed by a sigmoid activation to estimate " + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "inline_equation", + "content": "r_k" + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": ". Based on " + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "inline_equation", + "content": "r_k" + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": ", we define a local window " + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "inline_equation", + "content": "W_k" + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": " around frame " + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 112, + 132, + 295, + 146 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 132, + 295, + 146 + ], + "spans": [ + { + "bbox": [ + 112, + 132, + 295, + 146 + ], + "type": "interline_equation", + "content": "W _ {k} = \\{j \\mid | j - k | \\leq r _ {k} \\times T \\}. \\tag {3}", + "image_path": "6fe0b0e934c1f864966481dc8177973205e84ba3d2cfb702e1d8f68ba03aff33.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 158, + 296, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 158, + 296, + 183 + ], + "spans": [ + { + "bbox": [ + 55, + 158, + 296, + 183 + ], + "type": "text", + "content": "Within this window, we apply local attention to capture fine-grained details:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 193, + 301, + 218 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 193, + 301, + 218 + ], + "spans": [ + { + "bbox": [ + 55, + 193, + 301, + 218 + ], + "type": "interline_equation", + "content": "F _ {\\text {l o c a l}} (k) = \\tilde {F} _ {M} (k) + \\text {A t t e n t i o n} (\\tilde {F} _ {M} (k), \\tilde {F} _ {M} \\left(W _ {k}\\right), \\tilde {F} _ {M} \\left(W _ {k}\\right)), \\tag {4}", + "image_path": "cdd4c933d24238fe2820ec9173ce2f622afab6229a31fcee1f03e6d6e22c6e12.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 218, + 296, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 218, + 296, + 243 + ], + "spans": [ + { + "bbox": [ + 55, + 218, + 296, + 243 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 218, + 296, + 243 + ], + "type": "inline_equation", + "content": "\\tilde{F}_M(W_k)" + }, + { + "bbox": [ + 55, + 218, + 296, + 243 + ], + "type": "text", + "content": " denotes the motion features within the local window around frame " + }, + { + "bbox": [ + 55, + 218, + 296, + 243 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 218, + 296, + 243 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 244, + 296, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 244, + 296, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 244, + 296, + 281 + ], + "type": "text", + "content": "To incorporate global context, we partition the full motion sequence into " + }, + { + "bbox": [ + 55, + 244, + 296, + 281 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 244, + 296, + 281 + ], + "type": "text", + "content": " segments and compute segment-level features via average pooling:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 291, + 295, + 306 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 291, + 295, + 306 + ], + "spans": [ + { + "bbox": [ + 130, + 291, + 295, + 306 + ], + "type": "interline_equation", + "content": "F _ {M} ^ {\\mathrm {s e g}} = \\left\\{\\bar {F} _ {M} (n) \\right\\} _ {n = 1} ^ {N}, \\tag {5}", + "image_path": "8ba1641c5a64aa71e6dc48574c2897ce5b608331ebdabf93e3c74a986a6a36c3.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 317, + 296, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 317, + 296, + 342 + ], + "spans": [ + { + "bbox": [ + 55, + 317, + 296, + 342 + ], + "type": "text", + "content": "where the segment size is controlled by the hyperparameter " + }, + { + "bbox": [ + 55, + 317, + 296, + 342 + ], + "type": "inline_equation", + "content": "S_{n}" + }, + { + "bbox": [ + 55, + 317, + 296, + 342 + ], + "type": "text", + "content": ". We then apply global attention:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 63, + 352, + 295, + 377 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 352, + 295, + 377 + ], + "spans": [ + { + "bbox": [ + 63, + 352, + 295, + 377 + ], + "type": "interline_equation", + "content": "F _ {\\text {g l o b a l}} (k) = F _ {\\text {l o c a l}} (k) + \\text {A t t e n t i o n} \\left(F _ {\\text {l o c a l}} (k), F _ {M} ^ {\\text {s e g}}, F _ {M} ^ {\\text {s e g}}\\right). \\tag {6}", + "image_path": "68d5bb33aeaa49f2edbf753946765a01540938ac0cf8d92011be9c2b873bfdf3.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 378, + 296, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 378, + 296, + 401 + ], + "spans": [ + { + "bbox": [ + 55, + 378, + 296, + 401 + ], + "type": "text", + "content": "The final representation for each viewpoint frame is obtained by concatenating local and global features:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 95, + 412, + 295, + 427 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 412, + 295, + 427 + ], + "spans": [ + { + "bbox": [ + 95, + 412, + 295, + 427 + ], + "type": "interline_equation", + "content": "F _ {M} (k) = \\left[ F _ {\\text {l o c a l}} (k); F _ {\\text {g l o b a l}} (k) \\right] \\in \\mathbb {R} ^ {2 H}, \\tag {7}", + "image_path": "166cb6e51326e66a9393d35aad1f0cc3c8f65eb7dec6bd9b334720ff93699377.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 438, + 296, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 438, + 296, + 461 + ], + "spans": [ + { + "bbox": [ + 55, + 438, + 296, + 461 + ], + "type": "text", + "content": "allowing the model to capture both fine-grained motion details and broader contextual information." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 483, + 296, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 296, + 591 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 296, + 591 + ], + "type": "text", + "content": "Bidirectional Cross-Modal Fusion To integrate the enhanced motion features with the text embeddings, we employ a bidirectional cross-attention mechanism. First, we update the motion features by performing a cross-attention operation, where the motion features act as queries and the text features serve as keys and values, and vice versa for the text features. The updated features are then passed through feed-forward networks, and we concatenate the motion and text representations to form the input for the next stage:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 603, + 295, + 617 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 603, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 115, + 603, + 295, + 617 + ], + "type": "interline_equation", + "content": "F _ {\\text {f u s i o n}} = \\left[ F _ {T}; \\{F _ {M} (k) \\} _ {k \\in K} \\right]. \\tag {8}", + "image_path": "9df0c262482e897023dfa01c382f5355d82cd4ad19b5495615b5987147f75a82.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "content": "This bidirectional fusion enhances the interaction between motion and text, enabling a coherent, contextually grounded understanding of human behavior. By selectively focusing on key motion frames and integrating enriched motion and text features, the Cross Talker Module reduces computational costs and improves the model's ability to capture complex human actions." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 72, + 474, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 474, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 474, + 84 + ], + "type": "text", + "content": "3.3. Behavior Generation Module" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 90, + 555, + 150 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 90, + 555, + 150 + ], + "spans": [ + { + "bbox": [ + 313, + 90, + 555, + 150 + ], + "type": "text", + "content": "The Behavior Generation Module leverages the fused features " + }, + { + "bbox": [ + 313, + 90, + 555, + 150 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{fusion}}" + }, + { + "bbox": [ + 313, + 90, + 555, + 150 + ], + "type": "text", + "content": " to generate a textual description " + }, + { + "bbox": [ + 313, + 90, + 555, + 150 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 313, + 90, + 555, + 150 + ], + "type": "text", + "content": " that encapsulates the observed human behavior. This module utilizes a language model " + }, + { + "bbox": [ + 313, + 90, + 555, + 150 + ], + "type": "inline_equation", + "content": "h(\\cdot)" + }, + { + "bbox": [ + 313, + 90, + 555, + 150 + ], + "type": "text", + "content": " to transform " + }, + { + "bbox": [ + 313, + 90, + 555, + 150 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{fusion}}" + }, + { + "bbox": [ + 313, + 90, + 555, + 150 + ], + "type": "text", + "content": " into a semantically coherent and contextually relevant output sequence:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 361, + 158, + 555, + 172 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 158, + 555, + 172 + ], + "spans": [ + { + "bbox": [ + 361, + 158, + 555, + 172 + ], + "type": "interline_equation", + "content": "Y = h \\left(F _ {\\text {f u s i o n}}\\right) = \\left\\{y _ {1}, y _ {2}, \\dots , y _ {L} \\right\\}, \\tag {9}", + "image_path": "3d86899e70c95f98ab0a7dff6ac64755f06c8c27c296603f33d103e17baad5d6.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 179, + 554, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 179, + 554, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 179, + 554, + 216 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 179, + 554, + 216 + ], + "type": "inline_equation", + "content": "y_{t}" + }, + { + "bbox": [ + 313, + 179, + 554, + 216 + ], + "type": "text", + "content": " denotes the token generated at the " + }, + { + "bbox": [ + 313, + 179, + 554, + 216 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 179, + 554, + 216 + ], + "type": "text", + "content": "-th time step, and " + }, + { + "bbox": [ + 313, + 179, + 554, + 216 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 179, + 554, + 216 + ], + "type": "text", + "content": " is the length of the generated description. The probability of generating each token " + }, + { + "bbox": [ + 313, + 179, + 554, + 216 + ], + "type": "inline_equation", + "content": "y_{t}" + }, + { + "bbox": [ + 313, + 179, + 554, + 216 + ], + "type": "text", + "content": " is given by:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 350, + 223, + 555, + 236 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 223, + 555, + 236 + ], + "spans": [ + { + "bbox": [ + 350, + 223, + 555, + 236 + ], + "type": "interline_equation", + "content": "p \\left(y _ {t} \\mid y _ {1: t - 1}, F _ {\\text {f u s i o n}}\\right) = \\operatorname {S o f t m a x} \\left(W _ {o} h _ {t}\\right), \\tag {10}", + "image_path": "22391ac7c0c2bc930cb3c0259cb79b7fb6068cd703b4bbf52158fe3757903a31.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 243, + 554, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 554, + 339 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 554, + 339 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 243, + 554, + 339 + ], + "type": "inline_equation", + "content": "h_t" + }, + { + "bbox": [ + 313, + 243, + 554, + 339 + ], + "type": "text", + "content": " represents the hidden state of the language model at time " + }, + { + "bbox": [ + 313, + 243, + 554, + 339 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 243, + 554, + 339 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 243, + 554, + 339 + ], + "type": "inline_equation", + "content": "W_o" + }, + { + "bbox": [ + 313, + 243, + 554, + 339 + ], + "type": "text", + "content": " is an output projection matrix. The hidden state " + }, + { + "bbox": [ + 313, + 243, + 554, + 339 + ], + "type": "inline_equation", + "content": "h_t" + }, + { + "bbox": [ + 313, + 243, + 554, + 339 + ], + "type": "text", + "content": " is derived from the model's internal mechanisms, which integrate attention over the input features and previously generated tokens. This autoregressive approach enables LLaMo to produce detailed and contextually aligned descriptions that incorporate nuanced information from both motion and text modalities." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 340, + 555, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 340, + 555, + 388 + ], + "spans": [ + { + "bbox": [ + 313, + 340, + 555, + 388 + ], + "type": "text", + "content": "By iteratively generating tokens based on the fused features and preceding context, the Behavior Generation Module constructs a meaningful and comprehensive narrative of human activity, grounded in the multimodal input data." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 314, + 394, + 425, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 394, + 425, + 407 + ], + "spans": [ + { + "bbox": [ + 314, + 394, + 425, + 407 + ], + "type": "text", + "content": "3.4. Training Objective" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 412, + 554, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 412, + 554, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 412, + 554, + 519 + ], + "type": "text", + "content": "To train LLaMo, we finetune a pretrained language model using supervised learning, guiding it to generate accurate and contextually relevant descriptions based on the input motion, video, and text features. Both the motion and video encoders remain frozen during training, focusing the optimization on the language model's alignment with multimodal inputs. The training objective minimizes the discrepancy between the generated descriptions and the groundtruth annotations in the dataset." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 521, + 554, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 554, + 583 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 554, + 583 + ], + "type": "text", + "content": "Given a training set of " + }, + { + "bbox": [ + 313, + 521, + 554, + 583 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 521, + 554, + 583 + ], + "type": "text", + "content": " samples, each with video frames " + }, + { + "bbox": [ + 313, + 521, + 554, + 583 + ], + "type": "inline_equation", + "content": "V^{(i)}" + }, + { + "bbox": [ + 313, + 521, + 554, + 583 + ], + "type": "text", + "content": ", motion data " + }, + { + "bbox": [ + 313, + 521, + 554, + 583 + ], + "type": "inline_equation", + "content": "M^{(i)}" + }, + { + "bbox": [ + 313, + 521, + 554, + 583 + ], + "type": "text", + "content": ", and corresponding ground-truth textual descriptions " + }, + { + "bbox": [ + 313, + 521, + 554, + 583 + ], + "type": "inline_equation", + "content": "\\hat{Y}^{(i)} = \\{\\hat{y}_1^{(i)},\\hat{y}_2^{(i)},\\dots ,\\hat{y}_L^{(i)}\\}" + }, + { + "bbox": [ + 313, + 521, + 554, + 583 + ], + "type": "text", + "content": ", the overall training objective is to minimize the negative log-likelihood of the ground-truth tokens across all samples:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 334, + 590, + 555, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 590, + 555, + 624 + ], + "spans": [ + { + "bbox": [ + 334, + 590, + 555, + 624 + ], + "type": "interline_equation", + "content": "L = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\sum_ {t = 1} ^ {L ^ {(i)}} \\log p \\left(\\hat {y} _ {t} ^ {(i)} \\mid \\hat {y} _ {1: t - 1} ^ {(i)}, F _ {\\text {f u s i o n}} ^ {(i)}\\right), \\tag {11}", + "image_path": "f09328e0960d4a38c347e35d9e1ff4a86009a2b168060bd437910fa0f7fe3405.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 632, + 554, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 632, + 554, + 690 + ], + "spans": [ + { + "bbox": [ + 313, + 632, + 554, + 690 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 632, + 554, + 690 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{fusion}}^{(i)}" + }, + { + "bbox": [ + 313, + 632, + 554, + 690 + ], + "type": "text", + "content": " represents the fused input features for the " + }, + { + "bbox": [ + 313, + 632, + 554, + 690 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 632, + 554, + 690 + ], + "type": "text", + "content": "-th sample. The term " + }, + { + "bbox": [ + 313, + 632, + 554, + 690 + ], + "type": "inline_equation", + "content": "p(\\hat{y}_t^{(i)} \\mid \\hat{y}_{1:t-1}^{(i)}, F_{\\mathrm{fusion}}^{(i)})" + }, + { + "bbox": [ + 313, + 632, + 554, + 690 + ], + "type": "text", + "content": " denotes the probability of the ground-truth token " + }, + { + "bbox": [ + 313, + 632, + 554, + 690 + ], + "type": "inline_equation", + "content": "\\hat{y}_t^{(i)}" + }, + { + "bbox": [ + 313, + 632, + 554, + 690 + ], + "type": "text", + "content": " given the previous tokens " + }, + { + "bbox": [ + 313, + 632, + 554, + 690 + ], + "type": "inline_equation", + "content": "\\hat{y}_{1:t-1}^{(i)}" + }, + { + "bbox": [ + 313, + 632, + 554, + 690 + ], + "type": "text", + "content": " and the fused features." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 690, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 690, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 690, + 555, + 714 + ], + "type": "text", + "content": "This objective ensures that the generated descriptions are syntactically and semantically aligned with the target" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17586" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "content": "outputs, capturing the detailed aspects of human behavior present in the video and motion inputs." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 107, + 138, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 107, + 138, + 121 + ], + "spans": [ + { + "bbox": [ + 55, + 107, + 138, + 121 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 128, + 296, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 128, + 296, + 175 + ], + "spans": [ + { + "bbox": [ + 55, + 128, + 296, + 175 + ], + "type": "text", + "content": "To evaluate LLaMo, we performed experiments across datasets representing complex human behaviors and tasks, including professional sports and cyclic action repetition counts." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 183, + 189, + 196 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 183, + 189, + 196 + ], + "spans": [ + { + "bbox": [ + 55, + 183, + 189, + 196 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 202, + 296, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 202, + 296, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 202, + 296, + 357 + ], + "type": "text", + "content": "Training Datasets: LLaMo was trained using a combination of bimodal video-motion datasets and unimodal video or motion datasets to enable a robust understanding of both modalities. For video-motion data, we utilized MoVid from MotionLLM [1] and our custom Swing dataset containing 20,000 annotated videos of baseball and golf swings with corresponding motion capture data and expert Q&A instructions. Motion-only datasets included HumanML3D [37] and KIT-ML [3], while Mo-RepCount, a filtered subset of RepCount [38], provided high-quality repetition count videos. This multimodal data configuration ensures that LLaMo effectively learns from both single and combined modalities." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 373, + 296, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 296, + 445 + ], + "type": "text", + "content": "Evaluation Datasets: LLaMo was evaluated on MoVid-Bench [1] to assess both video and motion understanding capabilities, with additional tests on BABEL-QA [19] to evaluate motion-based question answering. For specific tasks, we used Mo-RepCount to gauge motion details and Swing to evaluate professional sports guidance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 460, + 296, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 296, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 296, + 568 + ], + "type": "text", + "content": "Evaluation Metrics: For MoVid-Bench, we follow previous LLM evaluation metrics [39] [22] [40] on accuracy and scores, which contains body-part awareness, sequentially, direction analysis, reasoning ability, and hallucination, respectively. BABEL-QA used the metric followed by [19], Mo-RepCount applied metrics such as OBO, MAE, OBZ, and RMSE, and Swing performance was assessed on reasonableness, coherence, pertinence, and adaptability using GPT-4 evaluations, scoring from 0 to 5." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 575, + 113, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 575, + 113, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 575, + 113, + 588 + ], + "type": "text", + "content": "4.2. Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "text", + "content": "Evaluation on Motion Understanding in MoVid-Bench. We evaluated LLaMo's motion understanding capabilities on the MoVid-Bench, focusing on five critical aspects. Both accuracy and score metrics were used to comprehensively assess LLaMo's performance. Table 1 provides a comparison with several baselines, including GPT-3.5 [41], MotionGPT [2], and MotionLLM [1]. The GPT-3.5 baseline, limited to textual data processing, struggles significantly with motion interpretation, yielding low performance. MotionGPT, while capable of incorporating motion data, is pri" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 143 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 143 + ], + "type": "text", + "content": "marily code motion into a codebook, which limits its reasoning ability on coiled motion data and leaves it vulnerable to hallucinations. Although MotionLLM provides improvements in understanding motion sequences, it still conducts translation on motion data, resulting in low accuracy on seq. and Hall." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 144, + 555, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 555, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 555, + 348 + ], + "type": "text", + "content": "LLaMo, in contrast, achieves superior results across nearly all metrics on MoVid-Bench-Motion, attaining the highest accuracy in most categories. This performance is primarily attributed to its preservation of critical motion-specific nuances, as LLaMo processes motion data as a distinct modality without requiring any conversion. LLaMo effectively captures the intricate details of 3D sequences associated with human behaviors. Its Feature Enhancer, combined with the Cross Talker module, enables simultaneous input of video and motion data, allowing LLaMo to selectively focus on motion frames most relevant to the text context. This design enhances the model's ability to capture complex patterns of human motion and improves its reasoning on motion-intensive tasks. Overall, LLaMo's results on MoVid-Bench demonstrate its effectiveness in advancing human behavior analysis, highlighting its potential for broader applications in general motion understanding." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 363, + 555, + 374 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 363, + 555, + 374 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 555, + 374 + ], + "type": "text", + "content": "Evaluation on Video Understanding in MoVid-Bench." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 375, + 553, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 375, + 553, + 446 + ], + "spans": [ + { + "bbox": [ + 313, + 375, + 553, + 446 + ], + "type": "text", + "content": "We evaluated LLaMo on the video component of the MoVid-Bench dataset to assess its capability for comprehensive video understanding. As presented in Table 1, LLaMo achieves strong performance across most metrics, demonstrating LLaMo's extraordinary comprehension ability on videos even without motion input." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 447, + 554, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 554, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 554, + 529 + ], + "type": "text", + "content": "Compared to other models like GPT-3.5 [41], VideoLLAVA [39], and MotionLLM [1], while LLaMo slightly trails behind MotionLLM in body-part awareness, LLaMo shows significant improvements in reasoning ability and overall scores. This indicates that our model excels in understanding complex video content and making accurate inferences about human behaviors." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 530, + 555, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 530, + 555, + 602 + ], + "spans": [ + { + "bbox": [ + 313, + 530, + 555, + 602 + ], + "type": "text", + "content": "The rewarding results can be attributed to our model's powerful ability to estimate motion data accurately in video and align it seamlessly with video inputs. This approach retains the motion capture of motion data and the auxiliary information provided by video, enriching the understanding of dynamic scenes." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "type": "text", + "content": "Evaluation on BABEL-QA. We evaluated LLaMo on the BABEL-QA dataset to assess its capability to handle complex motion-based queries. As shown in Table 2, LLaMo achieves the highest overall score of 0.458, outperforming all baselines. It excels particularly in the Action category with a score of 0.525, demonstrating superior action recognition capabilities. Additionally, it achieves top scores in temporal reasoning tasks (Before and After), indicating a" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17587" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 73, + 108, + 129 + ], + "blocks": [ + { + "bbox": [ + 61, + 91, + 69, + 108 + ], + "lines": [ + { + "bbox": [ + 61, + 91, + 69, + 108 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 69, + 108 + ], + "type": "text", + "content": "e e" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 72, + 73, + 108, + 129 + ], + "lines": [ + { + "bbox": [ + 72, + 73, + 108, + 129 + ], + "spans": [ + { + "bbox": [ + 72, + 73, + 108, + 129 + ], + "type": "image", + "image_path": "21e8b7bec3a137a35036556a5f4d6ca28740a73bdf86b7affb8eef808260b72a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 115, + 72, + 146, + 129 + ], + "blocks": [ + { + "bbox": [ + 115, + 72, + 146, + 129 + ], + "lines": [ + { + "bbox": [ + 115, + 72, + 146, + 129 + ], + "spans": [ + { + "bbox": [ + 115, + 72, + 146, + 129 + ], + "type": "image", + "image_path": "308b04019fbdaaf485a6f47cea714dc2fa91b12e073a31bfdfe18fbb56dfcf21.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 153, + 73, + 182, + 129 + ], + "blocks": [ + { + "bbox": [ + 153, + 73, + 182, + 129 + ], + "lines": [ + { + "bbox": [ + 153, + 73, + 182, + 129 + ], + "spans": [ + { + "bbox": [ + 153, + 73, + 182, + 129 + ], + "type": "image", + "image_path": "e8387f7b304d4d04ab54c316bf28e4953df442d47b1c1d40616b51882572d02d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 190, + 73, + 218, + 129 + ], + "blocks": [ + { + "bbox": [ + 190, + 73, + 218, + 129 + ], + "lines": [ + { + "bbox": [ + 190, + 73, + 218, + 129 + ], + "spans": [ + { + "bbox": [ + 190, + 73, + 218, + 129 + ], + "type": "image", + "image_path": "772529eca92b7d5d18040a0ad6aa554a2d80c39bb16f4357cb90dd7ea3133dd3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 225, + 73, + 250, + 128 + ], + "blocks": [ + { + "bbox": [ + 225, + 73, + 250, + 128 + ], + "lines": [ + { + "bbox": [ + 225, + 73, + 250, + 128 + ], + "spans": [ + { + "bbox": [ + 225, + 73, + 250, + 128 + ], + "type": "image", + "image_path": "68edf55cc48e16ae54ce913a517b4a80c72688b34871bce62471906dd2472e0a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 258, + 73, + 287, + 128 + ], + "blocks": [ + { + "bbox": [ + 258, + 73, + 287, + 128 + ], + "lines": [ + { + "bbox": [ + 258, + 73, + 287, + 128 + ], + "spans": [ + { + "bbox": [ + 258, + 73, + 287, + 128 + ], + "type": "image", + "image_path": "0ae195acc9f7b15ff2081792485f2d8dd6c7729044f9d78c2d78f423d8806a97.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 293, + 73, + 323, + 128 + ], + "blocks": [ + { + "bbox": [ + 293, + 73, + 323, + 128 + ], + "lines": [ + { + "bbox": [ + 293, + 73, + 323, + 128 + ], + "spans": [ + { + "bbox": [ + 293, + 73, + 323, + 128 + ], + "type": "image", + "image_path": "1f3f8bf582d6a4f5651e5036f0451e9f4f97f94e301ad6b8f30ea25f7d38c92d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 328, + 79, + 339, + 91 + ], + "lines": [ + { + "bbox": [ + 328, + 79, + 339, + 91 + ], + "spans": [ + { + "bbox": [ + 328, + 79, + 339, + 91 + ], + "type": "text", + "content": "#" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 342, + 76, + 534, + 102 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 76, + 534, + 102 + ], + "spans": [ + { + "bbox": [ + 342, + 76, + 534, + 102 + ], + "type": "text", + "content": "Describe the following human motion video concisely, including but not limited to what actions are being performed, the standards followed, and the characteristics." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 342, + 109, + 536, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 109, + 536, + 180 + ], + "spans": [ + { + "bbox": [ + 342, + 109, + 536, + 180 + ], + "type": "text", + "content": "The person is practicing a Tai Chi movement called \"Golden Rooster Stands on One Leg.\" The movement starts from a wide stance with extended arms and gradually transitions into a single-leg balance, with the other knee lifted and the hands drawn close to the body. The entire process is slow and controlled, with smooth and continuous motions, emphasizing the integration of mind and body, showcasing the characteristics of Tai Chi." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 62, + 130, + 208, + 186 + ], + "blocks": [ + { + "bbox": [ + 62, + 130, + 208, + 186 + ], + "lines": [ + { + "bbox": [ + 62, + 130, + 208, + 186 + ], + "spans": [ + { + "bbox": [ + 62, + 130, + 208, + 186 + ], + "type": "image", + "image_path": "0d97bb1db2bca601221d1e2abf22ae3355e5c8eb1d3821a0c3a3039283d7e8e6.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 208, + 133, + 323, + 184 + ], + "blocks": [ + { + "bbox": [ + 208, + 133, + 323, + 184 + ], + "lines": [ + { + "bbox": [ + 208, + 133, + 323, + 184 + ], + "spans": [ + { + "bbox": [ + 208, + 133, + 323, + 184 + ], + "type": "image", + "image_path": "a816d961af460a53aacb302c19591bfd13830226e517b2bf8e269ec831b3c6ce.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 72, + 187, + 208, + 236 + ], + "blocks": [ + { + "bbox": [ + 72, + 187, + 208, + 236 + ], + "lines": [ + { + "bbox": [ + 72, + 187, + 208, + 236 + ], + "spans": [ + { + "bbox": [ + 72, + 187, + 208, + 236 + ], + "type": "image", + "image_path": "a2dec7880660cc958862edae3a8d3bef710de30643fcf191b216786537369b73.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 365, + 553, + 385 + ], + "lines": [ + { + "bbox": [ + 55, + 365, + 553, + 385 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 553, + 385 + ], + "type": "text", + "content": "Figure 4. Example outputs from LLaMo across human activities and professional sports, showcasing its reasoning capabilities and domain-specific knowledge in motion-intensive scenarios." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 223, + 188, + 388, + 258 + ], + "blocks": [ + { + "bbox": [ + 223, + 188, + 388, + 258 + ], + "lines": [ + { + "bbox": [ + 223, + 188, + 388, + 258 + ], + "spans": [ + { + "bbox": [ + 223, + 188, + 388, + 258 + ], + "type": "image", + "image_path": "56370119188f0947e59157efbc6062889103e27afd4e5f7da97d58831cfa36bf.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 410, + 186, + 528, + 234 + ], + "blocks": [ + { + "bbox": [ + 410, + 186, + 528, + 234 + ], + "lines": [ + { + "bbox": [ + 410, + 186, + 528, + 234 + ], + "spans": [ + { + "bbox": [ + 410, + 186, + 528, + 234 + ], + "type": "image", + "image_path": "2ca5dbd12e367ff59fa0c4951037e02d07f6db1f0315be09a70929c3c78e2794.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 59, + 238, + 216, + 306 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 59, + 238, + 190, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 238, + 190, + 259 + ], + "spans": [ + { + "bbox": [ + 59, + 238, + 190, + 259 + ], + "type": "text", + "content": "What is the person in the video doing? The person in the video is making a phone" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 59, + 262, + 194, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 262, + 194, + 284 + ], + "spans": [ + { + "bbox": [ + 59, + 262, + 194, + 284 + ], + "type": "text", + "content": "Can you analyze her current emotions? Sure, she seems a little impatient and angry" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 59, + 287, + 216, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 287, + 216, + 306 + ], + "spans": [ + { + "bbox": [ + 59, + 287, + 216, + 306 + ], + "type": "text", + "content": "Why do you think she's angry? Could it be that she's feeling surprised instead?" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 59, + 309, + 221, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 309, + 221, + 354 + ], + "spans": [ + { + "bbox": [ + 59, + 309, + 221, + 354 + ], + "type": "text", + "content": "Determining whether a person is angry can be based on their body language. She is likely feeling angry because she stopped and stomped her foot while walking, and her body movements became somewhat sluggish after hanging up the phone." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 223, + 259, + 376, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 259, + 376, + 278 + ], + "spans": [ + { + "bbox": [ + 223, + 259, + 376, + 278 + ], + "type": "text", + "content": "How many push-ups did the person do in the video?" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 224, + 285, + 384, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 285, + 384, + 313 + ], + "spans": [ + { + "bbox": [ + 224, + 285, + 384, + 313 + ], + "type": "text", + "content": "The person in the video did three push-ups. Is it definitely 3 times? Why do I remember it as 5 times?" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 223, + 318, + 386, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 318, + 386, + 354 + ], + "spans": [ + { + "bbox": [ + 223, + 318, + 386, + 354 + ], + "type": "text", + "content": "Let me observe carefully again. It's likely that the smaller range of motion led to your misjudgment. Indeed, he did three push-ups, with three sets of repeated movements." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 389, + 236, + 551, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 236, + 551, + 254 + ], + "spans": [ + { + "bbox": [ + 389, + 236, + 551, + 254 + ], + "type": "text", + "content": "What should be the position of the knees in golf technique?" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 389, + 258, + 468, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 258, + 468, + 293 + ], + "spans": [ + { + "bbox": [ + 389, + 258, + 468, + 293 + ], + "type": "text", + "content": "GT: You want to keep your knees a little wider and more stable in golf technique." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 472, + 251, + 551, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 472, + 251, + 551, + 295 + ], + "spans": [ + { + "bbox": [ + 472, + 251, + 551, + 295 + ], + "type": "text", + "content": "LLaMo: Keep knees slightly bent and stable, with the left knee moving forward during the downswing." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 389, + 297, + 551, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 297, + 551, + 315 + ], + "spans": [ + { + "bbox": [ + 389, + 297, + 551, + 315 + ], + "type": "text", + "content": "What joint angles are important at the top of the swing for elite golfers?" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 389, + 318, + 468, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 318, + 468, + 353 + ], + "spans": [ + { + "bbox": [ + 389, + 318, + 468, + 353 + ], + "type": "text", + "content": "GT: Most elite golfers have joint angles ranging from 35 to 45 degrees at the top of the swing." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 472, + 318, + 551, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 472, + 318, + 551, + 353 + ], + "spans": [ + { + "bbox": [ + 472, + 318, + 551, + 353 + ], + "type": "text", + "content": "LLaMo: At the top of the swing, elite golfers have hips at 45 degrees and wrists at 90 degrees." + } + ] + } + ], + "index": 29 + }, + { + "type": "table", + "bbox": [ + 58, + 430, + 552, + 569 + ], + "blocks": [ + { + "bbox": [ + 55, + 398, + 554, + 420 + ], + "lines": [ + { + "bbox": [ + 55, + 398, + 554, + 420 + ], + "spans": [ + { + "bbox": [ + 55, + 398, + 554, + 420 + ], + "type": "text", + "content": "Table 1. Expected Comparison on the MoVid-Bench. The top part of the table presents motion-related results, and the bottom part presents video-related results. Higher accuracy and score values indicate better performance." + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 430, + 552, + 569 + ], + "lines": [ + { + "bbox": [ + 58, + 430, + 552, + 569 + ], + "spans": [ + { + "bbox": [ + 58, + 430, + 552, + 569 + ], + "type": "table", + "html": "
MoVid-Bench-MotionBody.Seq.Dir.Rea.Hall.All
Acc.ScoreAcc.ScoreAcc.ScoreAcc.ScoreAcc.ScoreAcc.Score
GT100.005.00100.005.00100.005.00100.005.00100.005.00100.005.00
GPT-3.5 [41]24.512.0430.412.2527.142.1939.192.6458.333.2231.332.31
MotionGPT [2]31.223.9842.693.1644.293.5035.813.0616.662.2536.863.11
MotionLLM [1]50.493.5536.843.1458.573.7652.703.5855.563.3949.503.49
LLaMo59.304.0144.013.1260.913.9958.213.6461.173.5355.323.67
MoVid-Bench-VideoBody.Seq.Dir.Rea.Hull.All
Acc.ScoreAcc.ScoreAcc.ScoreAcc.ScoreAcc.ScoreAcc.Score
GT100.005.00100.005.00100.005.00100.005.00100.005,00100.005.00
GPT-3.5 [41]2.401.231.391.004.651.095.411.650.000.943.031.26
Video-LLAVA [39]33.532.7625.462.7241.862.8452.973.2858.831.8942.532.70
MotionLLM [1]34.132.9332.872.9244.183.1463.203.5570.592.3049.002.97
LLaMo33.832.8536.013.1145.503.3267.593.7372.812.2552.333.10
", + "image_path": "b199ec79aa272c4e1f61227f8e1c3059d63b7ba695aa62b4aabca667acb136c2.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "table_body" + } + ], + "index": 32 + }, + { + "bbox": [ + 55, + 589, + 295, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 589, + 295, + 671 + ], + "spans": [ + { + "bbox": [ + 55, + 589, + 295, + 671 + ], + "type": "text", + "content": "strong understanding of temporal relationships in motion sequences. This capability benefits from the direct integration of motion data as an independent modality and the effective text-aware motion frames extract mechanism in LLaMo, preserving crucial motion-specific nuances, as well as leading to more accurate and contextually rich interpretations." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "Professional Sports Analysis. We further evaluate LLaMo on the challenging swing-dataset, termed as" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 313, + 589, + 555, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 589, + 555, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 589, + 555, + 696 + ], + "type": "text", + "content": "profession-swing, details refer to Appendix, where the ground truth (GT) consists of answers provided by professional coaches, serving as expert benchmarks. The evaluation focuses on four key indicators: Reasonableness: Logical soundness and plausibility of the responses. Coherence: Consistency and logical flow within the responses. Pertinence: Relevance of the answers to the specific questions. Adaptability: Ability to adjust responses based on the stage on the athlete." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 326, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 553, + 714 + ], + "type": "text", + "content": "As presented in Table 3, LLaMo significantly outper" + } + ] + } + ], + "index": 36 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17588" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 60, + 91, + 549, + 153 + ], + "blocks": [ + { + "bbox": [ + 144, + 70, + 465, + 82 + ], + "lines": [ + { + "bbox": [ + 144, + 70, + 465, + 82 + ], + "spans": [ + { + "bbox": [ + 144, + 70, + 465, + 82 + ], + "type": "text", + "content": "Table 2. Comparison on BABEL-QA dataset. Higher scores indicate better performance." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 91, + 549, + 153 + ], + "lines": [ + { + "bbox": [ + 60, + 91, + 549, + 153 + ], + "spans": [ + { + "bbox": [ + 60, + 91, + 549, + 153 + ], + "type": "table", + "html": "
ModelPred. typeOverall ↑Action ↑Direction ↑Body Part ↑Before ↑After ↑Other ↑
MotionCLIP-M [41]cls.0.4300.4850.3610.2720.3720.3210.404
MotionCLIP-R [41]cls.0.4200.4890.3100.2500.3980.3140.387
MotionLLM [41]gen.0.4360.5170.3540.1540.4270.3680.529
LLaMogen.0.4580.5250.3980.2240.4430.3920.518
", + "image_path": "28869079c851c28fe262b722eba7d2e8274659d3839d8aab34ad9d147233fc6e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 60, + 184, + 550, + 264 + ], + "blocks": [ + { + "bbox": [ + 56, + 163, + 553, + 175 + ], + "lines": [ + { + "bbox": [ + 56, + 163, + 553, + 175 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 553, + 175 + ], + "type": "text", + "content": "Table 3. Performance on the profession-swing dataset across four indicators. Higher accuracy and score values indicate better performance." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 184, + 550, + 264 + ], + "lines": [ + { + "bbox": [ + 60, + 184, + 550, + 264 + ], + "spans": [ + { + "bbox": [ + 60, + 184, + 550, + 264 + ], + "type": "table", + "html": "
ModelReasonablenessCoherencePertinenceAdaptabilityAll
AccScoreAccScoreAccScoreAccScoreAccScore
GT10051005100510051005
GPT-3.5 [41]4.510.834.200.743.650.661.710.223.420.61
MotionGPT [2]10.531.2515.421.6612.641.114.791.3514.351.4
MotionLLM [1]12.331.5119.201.8717.981.7910.201.2216.531.57
LLaMo21.102.1127.102.7131.813.1220.221.9824.802.48
", + "image_path": "2cf2ce3eedc5510a29a3572ed6c723a272e76fa6beaacb5fa13a1fb5864efb04.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 283, + 294, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 283, + 294, + 344 + ], + "spans": [ + { + "bbox": [ + 54, + 283, + 294, + 344 + ], + "type": "text", + "content": "forms both GPT-3.5 [41] and MotionLLM [1] across all metrics. While MotionLLM shows improvements over GPT-3.5, with higher accuracy and scores, it still lags behind LLaMo. Specifically, LLaMo achieves an overall accuracy of 24.80" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 344, + 295, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 344, + 295, + 439 + ], + "spans": [ + { + "bbox": [ + 54, + 344, + 295, + 439 + ], + "type": "text", + "content": "The superior performance of LLaMo can be attributed to the design of the Cross Talker module, which enables language-guided frame selection to capture complex motion relationships inherent in professional swings. This allows LLaMo to effectively model intricate motion patterns and generate responses that are more reasonable, coherent, pertinent, and adaptable, closely aligning with the expert-level insights provided by professional coaches." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 440, + 295, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 440, + 295, + 489 + ], + "spans": [ + { + "bbox": [ + 55, + 440, + 295, + 489 + ], + "type": "text", + "content": "In contrast, while MotionLLM demonstrates certain strengths over MotionGPT—such as better pertinence and coherence. It lacks the advanced motion understanding capabilities of LLaMo, lagging behind LLaMo distinctly." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 69, + 529, + 280, + 615 + ], + "blocks": [ + { + "bbox": [ + 55, + 498, + 295, + 521 + ], + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 521 + ], + "type": "text", + "content": "Table 4. Motion and video details capture evaluation on MoRepCount" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 529, + 280, + 615 + ], + "lines": [ + { + "bbox": [ + 69, + 529, + 280, + 615 + ], + "spans": [ + { + "bbox": [ + 69, + 529, + 280, + 615 + ], + "type": "table", + "html": "
ModelOBOMAEOBZRMSE
EScounts [42]0.3970.2910.1985.58
PoseRAC [43]0.3820.3120.2045.95
TransRAC [38]0.2760.4440.1058.56
RepNet [44]0.009///
LLaMo0.3890.3240.2226.15
", + "image_path": "142cd9266401a1132524d38c1689d810fd9f41fb7cb2dc60371874d504f65865.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "text", + "content": "Evaluation on Mo-RepCount. We evaluated LLaMo on the Mo-RepCount dataset to assess its ability to capture fine-grained motion details and complex temporal features required for accurate repetition counting. For a fair comparison, we trained all state-of-the-art models, including RepNet [44], TransRAC [38], PoseRAC [43] and ES" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 283, + 553, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 283, + 553, + 368 + ], + "spans": [ + { + "bbox": [ + 313, + 283, + 553, + 368 + ], + "type": "text", + "content": "counts [42], on the Mo-RepCount dataset without using any additional data. As presented in Table 4, LLaMo outperforms the RepNet and TransRAC across almost all metrics. Since both of them are models built for loop counting tasks, the effect of our model on this task is convincing, which shows that LLaMo is also competent in capturing the details of the motion and videos." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 370, + 555, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 370, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 370, + 555, + 456 + ], + "type": "text", + "content": "While LLaMo did not outperform the specialized counting models EScounts and poseRAC on many metrics, it excelled on the OBZ metric, demonstrating an unmatched ability to capture fine-grained motion details and spatiotemporal relationships. These results confirm LLaMo's strong potential in motion and video analysis, underscoring its suitability for detail-oriented tasks and applications." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 475, + 388, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 388, + 487 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 388, + 487 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "content": "We introduce LLaMo, a novel framework that integrates motion data directly into large language models, thereby enhancing multimodal understanding. By unifying text, video, and motion data in an architecture fine-tuned for human motion instruction, LLaMo achieves notable progress in behavior comprehension, motion-focused captioning, and interactive Q&A, capturing the fine-grained motion details vital for precise, context-aware insights. Experimental results, especially in professional sports analysis, demonstrate LLaMo's capacity to interpret complex human motion with accuracy. Future work will extend LLaMo to a broad range of domains where detailed motion analysis and multimodal understanding are essential. The main challenges lie in refining cross-modal integration techniques and improving efficiency for real-time applications, thus broadening LLaMo's potential in human-centered AI. We hope our work will spur further exploration into more advanced human-centric multimodal models." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17589" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 91, + 296, + 115 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 91, + 296, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 91, + 296, + 115 + ], + "type": "text", + "content": "This work was supported in part by the Pioneer Centre for AI, DNRF grant number P1." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 125, + 115, + 137 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 125, + 115, + 137 + ], + "spans": [ + { + "bbox": [ + 56, + 125, + 115, + 137 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 145, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 145, + 296, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 145, + 296, + 190 + ], + "spans": [ + { + "bbox": [ + 61, + 145, + 296, + 190 + ], + "type": "text", + "content": "[1] Ling-Hao Chen, Shunlin Lu, Ailing Zeng, Hao Zhang, Benyou Wang, Ruimao Zhang, and Lei Zhang. Motionllm: Understanding human behaviors from human motions and videos. arXiv preprint arXiv:2405.20340, 2024. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 190, + 296, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 190, + 296, + 233 + ], + "spans": [ + { + "bbox": [ + 62, + 190, + 296, + 233 + ], + "type": "text", + "content": "[2] Ling-Hao Chen, Jiawei Zhang, Wen Liu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. In Proceedings of the Conference on Neural Information Processing Systems (NeurIPS), 2024. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 235, + 294, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 235, + 294, + 266 + ], + "spans": [ + { + "bbox": [ + 62, + 235, + 294, + 266 + ], + "type": "text", + "content": "[3] Matthias Plappert, Christian Mandery, and Tamim Asfour. The kit motion-language dataset. Big data, 4(4):236-252, 2016. 1, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 268, + 296, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 268, + 296, + 312 + ], + "spans": [ + { + "bbox": [ + 62, + 268, + 296, + 312 + ], + "type": "text", + "content": "[4] Yan Zhang, Michael J Black, and Siyu Tang. We are more than our joints: Predicting how 3d bodies move. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 313, + 296, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 313, + 296, + 367 + ], + "spans": [ + { + "bbox": [ + 62, + 313, + 296, + 367 + ], + "type": "text", + "content": "[5] Fangzhou Hong, Liang Pan, Zhongang Cai, and Ziwei Liu. Versatile multi-modal pre-training for human-centric perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16156-16166, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 369, + 296, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 369, + 296, + 412 + ], + "spans": [ + { + "bbox": [ + 62, + 369, + 296, + 412 + ], + "type": "text", + "content": "[6] Haoxuan Qu, Yujun Cai, and Jun Liu. Llms are good action recognizers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18395-18406, 2024. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 413, + 294, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 413, + 294, + 445 + ], + "spans": [ + { + "bbox": [ + 62, + 413, + 294, + 445 + ], + "type": "text", + "content": "[7] Shan Chen, Jiale Zhou, and Lei Li. Dense point clouds matter: Dust-gs for scene reconstruction from sparse viewpoints. arXiv preprint arXiv:2409.08613, 2024. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 447, + 296, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 447, + 296, + 501 + ], + "spans": [ + { + "bbox": [ + 62, + 447, + 296, + 501 + ], + "type": "text", + "content": "[8] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 502, + 296, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 502, + 296, + 557 + ], + "spans": [ + { + "bbox": [ + 62, + 502, + 296, + 557 + ], + "type": "text", + "content": "[9] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 558, + 296, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 558, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 57, + 558, + 296, + 601 + ], + "type": "text", + "content": "[10] Li Lei, Jia Sen, Wang Jianhao, An Zhaochong, Li Jiaang, Hwang Jenq-Neng, and Belongie Serge. Chatmotion: A multimodal multi-agent for human motion analysis. arXiv preprint arXiv:2502.18180, 2025. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 602, + 296, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 602, + 296, + 646 + ], + "spans": [ + { + "bbox": [ + 57, + 602, + 296, + 646 + ], + "type": "text", + "content": "[11] Mark Endo, Joy Hsu, Jiaman Li, and Jiajun Wu. Motion question answering via modular motion programs. In Proceedings of the International Conference on Machine Learning (ICML), 2023. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 647, + 296, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 296, + 690 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 296, + 690 + ], + "type": "text", + "content": "[12] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "type": "text", + "content": "[13] Yaya Shi, Haiyang Xu, Chunfeng Yuan, Bing Li, Weiming Hu, and Zheng-Jun Zha. Learning video-text aligned repre" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "sentations for video captioning. IEEE Transactions on Multimedia (TMM), 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "text", + "content": "[14] Chuan Guo, Xinxin Zuo, Sen Wang, and Li Cheng. Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In Proceedings of the European Conference on Computer Vision (ECCV), pages 580-597, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "text", + "content": "[15] Kunchang Li, Yali Wang, Yinan He, et al. Mvbench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 197, + 553, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 241 + ], + "type": "text", + "content": "[16] Zhaochong An, Guolei Sun, Yun Liu, Runjia Li, Min Wu, Ming-Ming Cheng, Ender Konukoglu, and Serge Belongie. Multimodality helps few-shot 3d point cloud semantic segmentation. arXiv preprint arXiv:2410.22489, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 242, + 553, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 242, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 316, + 242, + 553, + 296 + ], + "type": "text", + "content": "[17] Bobby Bodenheimer, Chuck Rose, Seth Rosenthal, and John Pella. The process of motion capture: Dealing with the data. In Proceedings of the Eurographics Workshop on Computer Animation and Simulation (EG Workshop), pages 3-18, 1997. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 297, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 342 + ], + "type": "text", + "content": "[18] Xiangyu Peng, Yuxuan Zhang, Xiaolong Wang, and Ziwei Liu. Text-driven human motion generation with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 343, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 553, + 387 + ], + "type": "text", + "content": "[19] Mark Endo, Joy Hsu, Jiaman Li, and Jiajun Wu. Motion question answering via modular motion programs. In International Conference on Machine Learning, pages 9312-9328. PMLR, 2023. 2, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 388, + 555, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 555, + 432 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 555, + 432 + ], + "type": "text", + "content": "[20] Jie Lin, Wei Zhang, and Zhe Chen. Videollm: Language models for video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2405-2415, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 433, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 433, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 433, + 553, + 464 + ], + "type": "text", + "content": "[21] Xu Ning, Hongwei Li, and Yu Zhao. Videobench: A benchmark for large-scale video understanding. arXiv preprint arXiv:2304.12345, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 467, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 467, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 317, + 467, + 553, + 510 + ], + "type": "text", + "content": "[22] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 2, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 512, + 553, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 512, + 553, + 534 + ], + "spans": [ + { + "bbox": [ + 316, + 512, + 553, + 534 + ], + "type": "text", + "content": "[23] Lei Li. Segment any building. In Computer Graphics International Conference, pages 155-166. Springer, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 535, + 553, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 535, + 553, + 579 + ], + "spans": [ + { + "bbox": [ + 317, + 535, + 553, + 579 + ], + "type": "text", + "content": "[24] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 580, + 555, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 580, + 555, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 580, + 555, + 624 + ], + "type": "text", + "content": "[25] Enxin Song, Guanhong Zhang, and Haoyang Zhou. Adaptive multimodal learning for behavior analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 624, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 669 + ], + "type": "text", + "content": "[26] Yunhua Yang, Ziwang Zhao, and Yiming Xie. Recognizing human behaviors with skeletal data in llm-based frameworks. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[27] Lei Li, Tianfang Zhang, Zhongfeng Kang, and Xikun Jiang. Mask-fpan: Semi-supervised face parsing in the wild with de-occlusion and uv gan. Computers & Graphics, 116:185–193, 2023. 2" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17590" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "type": "text", + "content": "[28] Matthias Plappert, Christian Mandery, and Tamim Asfour. Learning a bidirectional mapping between human whole-body motion and natural language using deep recurrent neural networks. Robotics and Autonomous Systems (RAS), 109: 13-26, 2018. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 295, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 295, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 295, + 172 + ], + "type": "text", + "content": "[29] Lei Li. Cpseg: Finer-grained image semantic segmentation via chain-of-thought language prompting. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 513-522, 2024." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 173, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 173, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 173, + 294, + 205 + ], + "type": "text", + "content": "[30] Yunhua Yang, Liang Zhang, and Hui Li. Understanding human behaviors from skeletal data: A review of datasets and methods. arXiv preprint arXiv:2310.12998, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 206, + 294, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 294, + 259 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 294, + 259 + ], + "type": "text", + "content": "[31] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Two-stream adaptive graph convolutional networks for skeleton-based action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12026-12035, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 261, + 294, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 294, + 303 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 294, + 303 + ], + "type": "text", + "content": "[32] Enxin Song, Wenhao Chai, and Yucheng Zhang. Fine-grained spatial-temporal motion understanding in complex video environments. arXiv preprint arXiv:2310.08639, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 304, + 294, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 304, + 294, + 347 + ], + "spans": [ + { + "bbox": [ + 56, + 304, + 294, + 347 + ], + "type": "text", + "content": "[33] Zhaochong An, Guolei Sun, Zongwei Wu, Hao Tang, and Luc Van Gool. Temporal-aware hierarchical mask classification for video semantic segmentation. arXiv preprint arXiv:2309.08020, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 350, + 294, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 294, + 403 + ], + "type": "text", + "content": "[34] Xianghong Liu, Haoxuan Wang, Zhiwei Zhang, Huan Wu, Baoquan Chen, and Xiaoguang Han. Category-agnostic pose estimation for point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 404, + 294, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 294, + 448 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 294, + 448 + ], + "type": "text", + "content": "[35] Xiaolong He, Yi Zhang, Chen Chen, and Kyoung Mu Lee. Activitynet++: A large-scale benchmark for video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 449, + 294, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 294, + 481 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 294, + 481 + ], + "type": "text", + "content": "[36] Lei Li. Hierarchical edge aware learning for 3d point cloud. In Computer Graphics International Conference, pages 81-92. Springer, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 482, + 294, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 482, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 482, + 294, + 536 + ], + "type": "text", + "content": "[37] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5152-5161, 2022. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 537, + 294, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 294, + 602 + ], + "type": "text", + "content": "[38] Huazhang Hu, Sixun Dong, Yiqun Zhao, Dongze Lian, Zhengxin Li, and Shenghua Gao. Transrac: Encoding multiscale temporal correlation with transformers for repetitive action counting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19013-19022, 2022. 6, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 603, + 294, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 294, + 646 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 294, + 646 + ], + "type": "text", + "content": "[39] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "type": "text", + "content": "[40] Peng Jin, Ryuichi Takanobu, Wancai Zhang, Xiaochun Cao, and Li Yuan. Chat-univi: Unified visual representation empowers large language models with image and video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13700-13710, 2024. 6" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 553, + 230 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 315, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 315, + 73, + 553, + 106 + ], + "type": "text", + "content": "[41] OpenAI. Gpt-3.5: Generative pre-trained transformer 3.5, 2023. https://platform.openai.com/docs/models/gpt-3-5.6,7,8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "type": "text", + "content": "[42] Saptarshi Sinha, Alexandros Stergiou, and Dima Damen. Every shot counts: Using exemplars for repetition counting in videos. arXiv preprint arXiv:2403.18074, 2024. 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 141, + 553, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 174 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 174 + ], + "type": "text", + "content": "[43] Ziyu Yao, Xuxin Cheng, and Yuexian Zou. Poserac: Pose saliency transformer for repetitive action counting. arXiv preprint arXiv:2303.08450, 2023. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 175, + 553, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 230 + ], + "type": "text", + "content": "[44] Bastian Wandt and Bodo Rosenhahn. Repnet: Weakly supervised training of an adversarial reprojection network for 3d human pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7782-7791, 2019. 8" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "17591" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Human-centered Interactive Learning via MLLMs for Text-to-Image Person Re-identification/5cc75be5-e122-4a4f-ba40-7450ab436d6a_content_list.json b/2025/Human-centered Interactive Learning via MLLMs for Text-to-Image Person Re-identification/5cc75be5-e122-4a4f-ba40-7450ab436d6a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..bdb489eea7eabafd83a6c1f6beb34d1a5d45f86c --- /dev/null +++ b/2025/Human-centered Interactive Learning via MLLMs for Text-to-Image Person Re-identification/5cc75be5-e122-4a4f-ba40-7450ab436d6a_content_list.json @@ -0,0 +1,1489 @@ +[ + { + "type": "text", + "text": "Human-centered Interactive Learning via MLLMs for Text-to-Image Person Re-identification", + "text_level": 1, + "bbox": [ + 151, + 130, + 846, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yang Qin $^{1}$ Chao Chen $^{2}$ Zhihang Fu $^{2}$ Dezhong Peng $^{1,4,5}$ Xi Peng $^{1,3}$ Peng Hu $^{1*}$", + "bbox": [ + 156, + 202, + 844, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1College of Computer Science, Sichuan University 2Independent Researcher", + "bbox": [ + 240, + 222, + 756, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ National Key Laboratory of Fundamental Algorithms and Models for Engineering Simulation, Sichuan University", + "bbox": [ + 119, + 239, + 875, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sichuan National Innovation New Vision UHD Video Technology Co., Ltd Tianfu Jincheng Laboratory", + "bbox": [ + 140, + 257, + 849, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 308, + 326, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite remarkable advancements in text-to-image person re-identification (TIREID) facilitated by the breakthrough of cross-modal embedding models, existing methods often struggle to distinguish challenging candidate images due to intrinsic limitations, such as network architecture and data quality. To address these issues, we propose an Interactive Cross-modal Learning framework (ICL), which leverages human-centered interaction to enhance the discriminability of text queries through external multimodal knowledge. To achieve this, we propose a plug-and-play Test-time Humane-centered Interaction (THI) module, which performs visual question answering focused on human characteristics, facilitating multi-round interactions with a multimodal large language model (MLLM) to align query intent with latent target images. Specifically, THI refines user queries based on the MLLM responses to reduce the gap to the best-matching images, thereby boosting ranking accuracy. Additionally, to address the limitation of low-quality training texts, we introduce a novel Reorganization Data Augmentation (RDA) strategy based on information enrichment and diversity enhancement to enhance query discriminability by enriching, decomposing, and reorganizing person descriptions. Extensive experiments on four TIREID benchmarks, i.e., CUHK-PEDES, ICFG-PEDES RSTPReid, RSTPReid, and UFine6926, demonstrate that our method achieves remarkable performance with substantial improvement. Code is available at https://github.com/QinYang79/ICL.", + "bbox": [ + 89, + 340, + 483, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 791, + 220, + 806 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, text-to-image person re-identification (TIREID) [3-5, 7, 33, 34, 41] has made great progress in aligning text descriptions and person images, enabling high-accuracy person search and identification. Different from traditional", + "bbox": [ + 89, + 816, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fba52da64d005a0539e6ec922a99984925182eb012c93d057d25364f0a94c25b.jpg", + "image_caption": [ + "(a) Existing methods" + ], + "image_footnote": [], + "bbox": [ + 516, + 306, + 906, + 386 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0f3b0b8518086d3904eb882445c010dfce5f4e8b3ce5da480db96badb7397d1b.jpg", + "image_caption": [ + "(b) Interactive TIREID", + "Figure 1. The illustration of our motivation. When performing text-based person re-identification, (a) existing methods commonly exploit cross-modal models to calculate similarity (such as IRRA [12] and RDE [21]) and then obtain the candidate person images by ranking. However, due to the intrinsic limitations of models and training data, they may not be able to distinguish challenging candidate images well enough to obtain satisfactory results. (b) Our motivation is to interact with the system for external guidance like a human and gradually refine the user's query from the candidate items shown in (a) by multiple rounds of question-answering, ultimately improving the overall ranking." + ], + "image_footnote": [], + "bbox": [ + 517, + 400, + 906, + 510 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "image-to-image re-identification [32, 39, 42, 43], TIREID, as a rising task in the multimodal community [19, 20, 22, 23, 27], retrieves person images using user-customized text queries, which is more practical in many scenarios where an image query is unavailable, such as in video surveillance systems [2] or crowd management [6]. A primary challenge in TIREID is learning to associate text queries with the corresponding person images at a fine-grained level, bridging the modality gap for accurate similarity measurement.", + "bbox": [ + 511, + 686, + 906, + 823 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To overcome this challenge, recent efforts have developed various strategies to improve cross-modal associations, including local attribute modeling [37, 44], loss function designing [12, 21], ReID-domain pre-training [28, 38], etc. Although these methods achieve promising perfor", + "bbox": [ + 511, + 825, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author: Peng Hu (penghu.ml@gmail.com).", + "bbox": [ + 107, + 886, + 423, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14390", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "mance, they are limited by the inherent defect of the offline models and training data, making it hard to handle dynamic user query inputs and leading to poor generalization. In practice, however, different users tend to input concise, vague, and diverse text queries based on their memories to the target person, thereby raising challenges for the offline models to understand fine-grained human characteristics from the queries. For example, as illustrated in Figure 1 (a), the user text query \"A woman walking visible from the back is wearing a white shirt, and black pants.\" fails to capture some important contextual details like handheld objects or background, resulting in inaccurate candidate images. In brief, it is hard for offline models to handle dynamic and challenging queries solely relying on their learned internal knowledge. To break through this bottleneck, we could seek to align queries with image candidates with the help of external knowledge, thereby reranking the retrieved images and boosting the identification accuracy.", + "bbox": [ + 89, + 90, + 480, + 363 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent studies [8, 13, 17] have explored integrating external knowledge from pre-trained models or (multimodal) large language models to guide various downstream tasks. For instance, Li et al. [17] leveraged multimodal knowledge of pre-trained vision-language models [24] to enhance unimodal image representations for external-guided clustering. However, this approach cannot achieve post-hoc improvement for offline models, leading to infeasibility in practice due to the prohibitive cost of fine-tuning or even retraining. In cross-modal retrieval, Han et al. [8] exploited the excellent comprehension and generation capabilities of multimodal large language models (MLLMs) to design an interactive re-ranking pipeline for text-to-video retrieval. However, Han et al. [8] introduces label leakage by using the ground truth video as the video in mind for the MLLM answerer agent, which is not applicable in actual testing.", + "bbox": [ + 89, + 367, + 482, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we present a novel Interactive Cross-modal Learning framework (ICL) that exploits the multimodal knowledge implicit in MLLMs to enhance the alignment between text queries and target person images. Specifically, we first propose a plug-and-play Test-time Human-centered Interaction (THI) module that refines text queries through interactions with an MLLM, enhancing the post-hoc ability of trained models to distinguish challenging candidates. THI identifies the latent target images through multi-round interactions with MLLMs, asking human-centered questions to the MLLM for fine-grained answers about the images, which refine the query texts to strengthen the alignment with the images, ultimately address the inherent limitations of the input queries and enhancing ranking accuracy. To address the intrinsic limitations in the trained models, we propose a Reorganization Data Augmentation strategy (RDA) to enrich and diversify pedestrian texts for training enhancement, thereby transferring external knowledge in MLLMs into the model. To enrich the texts, RDA applies", + "bbox": [ + 89, + 613, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "visual question answering via the MLLM on each image-text pair to supplement fine-grained person characteristics. Moreover, to enhance the diversity of texts, RDA presents a decomposition-reorganization strategy to decompose person descriptions into attribute-specific sub-sentences (e.g., clothes, pants, shoes, etc.), and rewrite them with MLLM into multiple sentences with the same meaning. Subsequently, the rewritten sentences are randomly reordered and recombined to generate varied augmented texts, thus boosting data diversity and enhancing the generalization of models. Our main contributions are as follows:", + "bbox": [ + 511, + 90, + 903, + 256 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce human-centered interaction to TIREID, proposing a novel MLLM-driven Interactive Cross-modal Learning framework (ICL), which leverages external knowledge to overcome the inherent limitations of existing offline methods in handling dynamic queries.", + "- A plug-and-play Test-time Human-centered Interaction module (THI) is presented to align query intent with latent target images through multi-round interactions with MLLMs, improving the ranking quality.", + "- An effective Reorganization Data Augmentation strategy (RDA), applying MLLMs for text decomposition and recombination, is developed to generate discriminative and diverse training texts, improving cross-modal learning.", + "- Extensive experiments on four text-to-image person re-identification benchmarks verify the effectiveness and superiority of our method, which achieves promising performance and superior generalization." + ], + "bbox": [ + 513, + 258, + 903, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 532, + 653, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Text-to-Image Person Re-identification", + "text_level": 1, + "bbox": [ + 513, + 559, + 846, + 574 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As a challenging topic in multimodal learning, text-to-image person re-identification (TIREID) aims to search the image of the target person with a given natural language query. Existing methods [1, 12, 21, 28, 41] can be roughly divided into three categories according to the used backbone type: unimodal backbones, general multimodal pre-trained backbones, and ReID-domain multimodal pre-trained backbones. The early TIREID methods [33, 35] commonly use modality-specific unimodal backbones to encode text or images for cross-modal alignment, e.g., ResNet [10], BERT [30], etc. Recently, benefiting from the rapid development of visual-language pre-training models [14, 24], more and more researchers [3, 12, 21] have begun to use general pre-trained models as the backbone networks for solving TIREID, hoping that the pre-trained alignment knowledge can improve modality representation and cross-modal alignment. However, the performance improvement is still limited due to the domain gap between the general pre-training data and TIREID tasks. To this end, some attempts [28, 38] exploit image-caption models or MLLMs to annotate images, thus obtaining a large number of image", + "bbox": [ + 511, + 583, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14391", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5486563401a74a428c6b4e71d4a148c3038f7d2f3506da315bc0cd95a87f5da7.jpg", + "image_caption": [ + "Figure 2. The illustration of our Test-time Human-centered Interaction (THI) module. THI includes $K$ rounds of interactions to align query intention with the latent target image by external guidance, where in each round, we perform human-centered visual question answering around fine-grained person attributes to enhance the semantic consistency between the query and the intended person image, and then improve the final ReID performance on the large-scale evaluation through efficient re-ranking. Besides, we perform supervised fine-tuning via LoRA to inspire the discriminate ability of MLLM for ReID domain images and better align queries with latent target images." + ], + "image_footnote": [], + "bbox": [ + 117, + 89, + 883, + 318 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "text pairs for ReID-domain pre-training. Although promising performance has been achieved, they still cannot get rid of the inherent defect of offline models during handling dynamic queries. In this paper, we propose a novel interactive cross-modal learning framework based on MLLMs, which exploits dynamic interactions for external guidance to improve the generalization ability.", + "bbox": [ + 89, + 406, + 483, + 512 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Visual Interactive Learning", + "text_level": 1, + "bbox": [ + 89, + 527, + 339, + 544 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With the booming development of multimodal large language models (MLLMs) [15, 18, 29], exploiting visual interaction to improve downstream tasks has attracted increasing attention from researchers. We collectively refer to these as visual interaction learning and divide them into offline interaction [28, 31] and online interaction [8, 9, 13] groups by engaging in specific task periods. The former aims to use interactions to improve the training/pre-training data, such as information richness and diversity. It is usually static and separated from users, which is the mainstream of existing TIREID methods, e.g., Tan et al. [28] apply MLLMs to conduct interactions with multiple prompt templates for better diversity. Unlike them, online interactions are designed to be able to use real-time feedback to improve training or testing. For example, Levy et al. [13] propose to gradually clarify user intention through dialogue to improve image retrieval performance. Likewise, Han et al. [8] develop a rerank pipeline based on LLM-based iterative navigation. However, these methods cannot be directly used for TIREID effectively due to the differences in tasks and data domains. In this paper, we develop a test-time module based on MLLMs to conduct interactions for external guidance, overcoming the limitations of the intra-", + "bbox": [ + 89, + 553, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "model knowledge in TIReID offline models.", + "bbox": [ + 511, + 406, + 805, + 420 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 511, + 435, + 648, + 453 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce an Interactive Cross-modal Learning framework (ICL), which consists of two core comments to address the inherent challenges in offline models and training data, i.e., Test-time Human-centered Interaction module (THI) and Reorganization Data Augmentation (RDA). In Section 3.1, we provide the necessary definitions to facilitate the study. Then, we outline the details of our THI and RDA in Sections 3.2 and 3.3, respectively.", + "bbox": [ + 511, + 460, + 906, + 583 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem and Symbol Statement", + "text_level": 1, + "bbox": [ + 511, + 592, + 794, + 608 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Suppose that we have the query text $q \\in \\mathcal{Q}$ , a pedestrian image $v \\in \\mathcal{V}$ , where $\\mathcal{Q}$ and $\\mathcal{V}$ are the text query set and image gallery. The purpose of TIReID is to utilize the text query to retrieve the ideal images from the gallery, thus achieving person searching by identities of retrieved images. To achieve this, existing methods usually train an offline cross-modal model $f_{\\mathrm{cross}} = \\{f,g\\}$ to measure the similarities between text queries and pedestrian images. For a text $q$ and an image $v$ , the similarity to measure the matching degree can be represented as $S_{q,v} \\equiv \\operatorname{Sim}(f(q), g(v))$ , where $\\operatorname{Sim}(*)$ is the similarity function, $f$ and $g$ are the text and image encoders, respectively. Then, we can utilize $q$ to search relevant candidate images $\\hat{v}$ from $\\mathcal{V}$ as follows:", + "bbox": [ + 511, + 614, + 906, + 811 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {V}} (q) = \\left\\{\\hat {v} _ {k} \\right\\} _ {k = 1} ^ {K} = \\operatorname {T o p} - K _ {v \\in \\mathcal {V}} \\left(S _ {q, v}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 823, + 903, + 842 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\hat{\\mathcal{V}}(q)$ is the candidate set for $q$ and $K$ is the number of candidate images. In addition, our ICL also involves MLLMs to conduct interactions and training augmentation,", + "bbox": [ + 511, + 854, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14392", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "which we denote as $\\mathcal{M}$ , and the prompt template function for $\\mathcal{M}$ is denoted by the symbol $\\mathcal{T}$ . Based on the above definitions, we will elaborate on ICL in the following sections.", + "bbox": [ + 89, + 90, + 482, + 137 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Test-time Humane-center Interaction", + "text_level": 1, + "bbox": [ + 89, + 146, + 413, + 161 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Anchor Localization. Due to the model bottleneck, it is hard for existing methods to distinguish challenging candidate images relying solely on internal knowledge. Thus, we exploit the fine-grained image understanding capability of MLLMs to conduct external guidance. We first propose asking questions such as \"Can this text accurately describe the image?\" to let MLLMs explicitly tell us the latent target (anchor) image. As shown in Figure 2, given a text query $q$ , we first get the corresponding candidate set $\\hat{\\mathcal{V}}(q)$ and then perform multiple $(K)$ rounds of interactions according to the ranking until the ideal image is determined. For the $k$ -th round, the answer for interaction can be represented as:", + "bbox": [ + 89, + 169, + 483, + 351 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\na _ {\\hat {v} _ {k}} ^ {q} = \\mathcal {M} \\left(\\mathcal {T} _ {\\text {l o c}} (q, \\hat {v} _ {k})\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 362, + 482, + 381 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $k \\in \\{1, \\dots, K\\}$ , $a_{\\hat{v}_k}^q$ is the answer of 'Yes' or 'No', $\\hat{v}_k$ is the Top- $k$ candidate image for query $q$ , and $T_{\\mathrm{loc}}$ is the prompt template function for anchor localization. However, since there is a domain gap between ReID images and the generic images used for pre-training or instruction-tuning MLLMs, the answers of MLLMs are often unreliable. To solve this, we exploit LoRA [11] to perform supervised fine-tuning (SFT) on MLLMs to inspire the fine-grained ability to identify the person image and the SFT loss is:", + "bbox": [ + 89, + 391, + 483, + 527 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {l o r a} = - \\sum_ {(x, y) \\in \\mathcal {Z}} \\sum_ {t = 1} ^ {| y |} \\log \\left(p _ {\\mathcal {M}} \\left(y _ {t} \\mid x, y < t\\right)\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 539, + 482, + 584 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $x$ is the input Prompt and $y$ is the output Response, $\\mathcal{Z} = \\{\\mathcal{Z}^{+},\\mathcal{Z}^{-}\\}$ is the SFT dataset. For $\\mathcal{Z}^{+}$ , we select part of training texts $(\\{q_i\\}_{i = 1}^{N_l})$ and the corresponding ground-truth images $(\\{v_{i}^{+}\\}_{i = 1}^{N_{l}})$ from the training set to construct the input Pormpts, i.e., $\\{\\mathcal{T}_{\\mathrm{loc}}(q_i,v_i^+)\\}_{i = 1}^{N_l}$ expecting MLLMs to output the Response of 'Yes'. Similarly, for $\\mathcal{Z}^{-}$ , we also use the text queries and the corresponding negative images from the training set to construct the input Pormpts, expecting MLLMs to output the Response of 'No'. To make $\\mathcal{Z}^{-}$ more discriminative, we use a TIREID pre-trained cross-modal model to obtain the Top-10 image with different person ID as input negative sample image for each text query by similarity ranking.", + "bbox": [ + 89, + 595, + 482, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Human-centered VQA. For each round, once we confirm the response of 'Yes', we will ask the anchor image $(\\bar{v})$ a series of questions to learn more details about pedestrian characteristics. These details can help us alleviate the discrepancy between the user query and the information within the anchor image, thus improving overall ranking by refining the query. We call this process human-centered visual", + "bbox": [ + 89, + 795, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c824de588e1e6a60c0dfc19054f26889e11793083914afd1bc9fc1d7ff33ef1f.jpg", + "image_caption": [ + "(a) CUHK-PEDES", + "Figure 3. The similarity statistics of the Top-1 retrieved items in the test sets of the three benchmarks. It is obvious that the items with higher similarity are more likely to be correct retrievals." + ], + "image_footnote": [], + "bbox": [ + 517, + 87, + 643, + 162 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f5e32f4be4afd3a7078b77138f1c5a39b731211571d68d7a7759261c686935bd.jpg", + "image_caption": [ + "(b) ICFG-PEDES" + ], + "image_footnote": [], + "bbox": [ + 643, + 88, + 774, + 161 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6a7d7a9a7a350b03bdf20481d1c0a06d7f9d22c6199507201d610b14c37eee61.jpg", + "image_caption": [ + "(c) RSTPReid" + ], + "image_footnote": [], + "bbox": [ + 774, + 88, + 905, + 162 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "question answering (VQA), which is expressed as:", + "bbox": [ + 511, + 233, + 849, + 250 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nr _ {\\bar {v}} = \\mathcal {M} \\left(\\mathcal {T} _ {\\mathrm {v q a}} \\left(\\left\\{c _ {i} \\right\\} _ {i = 1} ^ {N _ {q}}, \\bar {v}\\right)\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 255, + 903, + 273 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\{c_i\\}_{i=1}^{N_q}$ are $N_q$ questions directed at the detail characteristics (e.g., gender, hair, upper body, lower body, shoes, etc.) of the pedestrian image, $r_{\\bar{v}}$ is the answer to $\\{c_i\\}_{i=1}^{N_q}$ one by one, and $\\mathcal{T}_{\\mathrm{vqa}}$ is the prompt template function for human-centered VQA. To improve the consistency between the image and the user query, a natural strategy is to concatenate these one-by-one answers after the original text. However, due to the limitation of the maximum text length that can be processed by the text model (e.g., it is usually 77 for CLIP), this will destroy the sequence structure. To this end, we recommend merging these answers and the original text by MLLMs to obtain a fluent and concise text, i.e.,", + "bbox": [ + 511, + 282, + 906, + 467 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {q} = \\mathcal {M} \\left(\\mathcal {T} _ {\\text {a g g r}} \\left(r _ {\\bar {v}}, q\\right)\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 635, + 474, + 903, + 491 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{q}$ is the merged text query and $\\mathcal{T}_{\\mathrm{aggr}}$ is the prompt template function for text aggregation.", + "bbox": [ + 511, + 498, + 905, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Efficient Re-ranking. To apply the above process to large-scale evaluation, efficiency is a factor that needs to be considered in the retrieval task, since introducing MLLMs to interactions for external guidance inevitably brings additional inference costs. To improve the interaction efficiency, we recommended adopting different strategies for different rounds of interactions. We observe that the retrieval quality is positively correlated with the similarity between the query and the Top-1 candidate image, as shown in Figure 3. To this end, in the first round, we only interact with images whose cross-modal similarity is greater than a threshold $\\xi$ and the answer of anchor localization is 'Yes'. This is because the retrievals with low similarities are more likely to be wrong items. After the first round, we only interact with the image that is all considered 'No' in the previous rounds of anchor localization and whose cross-modal similarity is less than the threshold $\\xi$ . This allows us to reduce unnecessary interactions and find text queries that really require interactions, making our method more efficient on large-scale evaluation. Given a query $q$ , if the above constraints are met, the interaction is carried out and we can get merged text $\\hat{q}$ after human-centered VQA, the re-ranking similarity $\\hat{S}_{q,v}$ of any image $v$ in $\\mathcal{V}$ is:", + "bbox": [ + 511, + 529, + 906, + 876 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {S} _ {q, v} = \\lambda S _ {q, v} + (1 - \\lambda) \\bar {S} _ {\\hat {q}, v}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 611, + 883, + 903, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14393", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\bar{S}_{\\hat{q},v} \\equiv 1$ if $v$ is $\\hat{v}_1$ , otherwise, $\\bar{S}_{\\hat{q},v} \\equiv S_{\\hat{q},v}$ , and $\\lambda \\in [0,1]$ is a hyperparameter to balance the contribution of the raw query and the refined text. To make our approach clearer, we describe the detailed algorithm process of our THI in Algorithm 1. Due to space limitations, all prompt templates can be found in the supplementary material.", + "bbox": [ + 89, + 89, + 483, + 183 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 The interaction process of our THI" + ], + "code_body": "Input: The query set $\\mathcal{Q}$ , the image gallery $\\mathcal{V}$ , the offline model $f_{\\mathrm{cross}}$ , the MLLM $\\mathcal{M}$ , the similarity threshold $\\xi$ , the number of interaction rounds $K$ ; \n1: Obtain candidate sets $\\{\\hat{\\mathcal{V}}(q_i)\\}_{i=1}^{|\\mathcal{Q}|}$ for all queries in $\\mathcal{Q}$ via Equation (1); \n2: for $k = 1,2,\\dots,K$ do \n3: for $i = 1,2,\\dots,|\\mathcal{Q}|$ do \n4: Conduct anchor localization via Equation (2) and output the answer of $a_{\\hat{v}_k^i}^{q_i}$ based on the $k$ -th candidate image $\\hat{v}_k^i$ in $\\hat{\\mathcal{V}}(q_i)$ ; \n5: if $a_{\\hat{v}_k^i}^{q_i}$ shows 'Yes', $k = 1$ , and $S_{q_i,\\hat{v}_1^i} > \\xi$ then \n6: Conduct human-centered VQA via Equations (4) and (5) to get the refined query $\\hat{q}_i$ ; \n7: Compute the re-ranking similarities between query $q_i$ and all images via Equation (6); \n8: end if \n9: if $\\{a_{\\hat{v}_j^i}\\}_{j=1}^{k-1}$ all show 'No', $a_{\\hat{v}_k^i}^{q_i}$ shows 'Yes', $k > 1$ , and $S_{q_i,\\hat{v}_1^i} \\leq \\xi$ then \n10: Conduct human-centered VQA via Equations (4) and (5) to get the refined query $\\hat{q}_i$ ; \n11: Compute the re-ranking similarities between query $q_i$ and all images via Equation (6); \n12: end if \n13: end for \n14: end for \n15: Re-ranking based on similarities; \nOutput: The new candidate images.", + "bbox": [ + 91, + 214, + 483, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Reorganization Data Augmentation", + "text_level": 1, + "bbox": [ + 89, + 664, + 398, + 680 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although THI can exploit interactions to provide external knowledge during test time to improve retrieval quality, the cross-modal embedding model is still a bottleneck that limits its further performance improvement. To this end, as shown in Figure 4, we introduce a new Reorganization Data Augmentation (RDA) strategy by enriching, decomposing, and reorganizing person descriptions to improve the discriminability and diversity of training data. We first apply Human-centered VQA in THI to obtain relevant characteristic descriptions and merge them with the original text, thus obtaining a text with richer information, i.e., $\\hat{q} = \\mathcal{M}(\\mathcal{T}_{\\mathrm{aggr}}(\\mathcal{M}(\\mathcal{T}_{\\mathrm{vqa}}(\\{c_i\\}_{i=1}^{N_q}, v)), q))$ . Then, we apply MLLMs to decompose enriched text $\\hat{q}$ into multiple independent sub-sentences that describe individual at-", + "bbox": [ + 89, + 686, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f12d5a579f973bb0872b8ae0e4528bf3b9a1fe6a3b6fd2b5de03f9b475a34af2.jpg", + "image_caption": [ + "Figure 4. The illustration of our RDA. The purpose of RDA is to supplement more details to the original training texts through human-centered VQA, improving the discriminability of texts. In addition, to enhance diversity, RDA maximizes diversity through the Decomposition-Rewriting-Reorganization strategy." + ], + "image_footnote": [], + "bbox": [ + 517, + 88, + 906, + 244 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tributes without interfering with each other, i.e., $\\{\\tilde{q}_i\\}_{i=1}^n = \\mathcal{M}(\\mathcal{T}_{\\mathrm{dec}}(\\hat{q}))$ , where $\\{\\tilde{q}_i\\}_{i=1}^n$ are the $n$ sub-sentences and $\\mathcal{T}_{\\mathrm{dec}}$ is the prompt template function for text decomposition. The purpose of decomposition is to be able to reorganize the text in a different order later. To increase diversity, we rewrite each sub-sentence into multiple sentences with different styles but the same meaning. For sub-sentence $\\tilde{q}$ , we conduct rewriting by $\\mathcal{R} = \\{\\check{q}_j\\}_{j=1}^m = \\mathcal{M}(\\mathcal{T}_{\\mathrm{rwt}}(\\tilde{q}))$ , where $\\mathcal{R}$ is the set with $m$ rewritten sentences and $\\mathcal{T}_{\\mathrm{rwt}}$ is the prompt template function for text rewriting.", + "bbox": [ + 511, + 344, + 906, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, for a text $q$ , we can obtain the set $\\{\\mathcal{R}_i\\}_{i=1}^n$ that contains a large number of style sub-sentences. So far, we can obtain augmented texts with different sub-sentence orders and different style combinations as shown in Figure 4, which we represent as $\\check{q}$ . During training, we can mix the augmentation texts $(\\check{q})$ with the original texts $(q)$ for cross-modal learning, thereby improving model-intra knowledge and generalization. Due to space limitations, more training details can be found in the supplementary material.", + "bbox": [ + 511, + 496, + 908, + 632 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 645, + 645, + 662 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we conduct extensive experiments to verify the effectiveness, superiority, and generalization of the proposed ICL on four public benchmark datasets.", + "bbox": [ + 511, + 671, + 906, + 717 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets and Evaluation Protocols", + "text_level": 1, + "bbox": [ + 511, + 726, + 812, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. In our experiments, we use three coarse-grained benchmarks, CHUK-PEDES [16], ICFG-PEDES [5], and RSTPReid [41], and one fine-grained benchmark, UFine6926 [44], to evaluate our ICL. Compared with the coarse-grained benchmark, the fine-grained benchmark has richer texts with fine-grained information. For all datasets, we follow their official settings for data partitioning. More details are provided in the supplementary materials.", + "bbox": [ + 511, + 750, + 906, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Protocols. To measure performance, we utilize the widely accepted Rank-K (1,5,10) metrics to mea", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14394", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/ad537603959926ce02e4e5749bd0543607315691337d3b32b9aa66e6e03118dd.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsImage Enc.Text Enc.CUHK-PEDESICFG-PEDESRSTPReid
Rank-1Rank-5Rank-10mAPmINPRank-1Rank-5Rank-10mAPmINPRank-1Rank-5Rank-10mAPmINP
① VL-Backbones w/o ReID-domain pre-training
IVT [26]ViT-BaseBERT65.6985.9391.1560.66-56.0473.6080.22--46.7070.0078.80--
\\( LCR^2S \\) [36]RN50BERT67.3684.1989.6259.20-57.9376.0882.4038.21-54.9576.6584.7040.92-
CFine [37]CLIP-ViTBERT69.5785.9391.15--60.8376.5582.42--50.5572.5081.60--
RaSa [1]Swin-BBERT76.5190.2994.2569.38-65.2880.4085.1241.29-66.9086.5091.3552.31-
IRRA [12]CLIP-ViTCLIP-X.73.3889.9393.7166.1350.2463.4680.2585.8238.067.9360.2081.3088.2047.1725.28
TBPS [3]CLIP-ViTCLIP-X.73.5488.1992.3565.3849.2565.0580.3485.4739.837.8762.1081.9087.7548.0025.86
CFAM [44]CLIP-ViTCLIP-X.75.6090.5394.3667.27-65.3881.1786.3539.42-62.4583.5591.1049.50-
RDE [21]CLIP-ViTCLIP-X.75.9490.1494.1267.5651.4467.6882.4787.3640.067.8765.3583.9589.9050.8828.08
Our ICLCLIP-ViTCLIP-X.76.4190.4894.3368.0451.9968.1182.5987.5240.818.1867.7086.0591.7552.6229.36
Our ICL*CLIP-ViTCLIP-X.77.9190.2794.1469.1353.4069.0282.4587.3641.218.3070.5585.9591.6553.6830.13
② VL-Backbones with ReID-domain pre-training
\\( IRRA^b \\) [12]CLIP-ViTCLIP-X.74.0589.4893.6466.57-64.3780.7586.1238.85-61.9080.6089.3048.08-
APTM [38]Swin-BBERT76.5390.0494.1566.91-68.5182.9987.5641.22-67.5085.7091.4552.56-
\\( NAM^d \\) [28]CLIP-ViTCLIP-X.77.4790.8494.6769.4354.0866.7682.0287.1741.459.5367.1586.5591.9052.0028.46
Our ICLCLIP-ViTCLIP-X.78.1891.6394.8369.5853.4869.2283.4988.0642.349.0170.0086.6091.7054.1630.93
Our ICL*CLIP-ViTCLIP-X.79.0691.2694.7270.4454.7070.0583.3587.9142.709.1372.5586.6091.3055.1931.72
", + "bbox": [ + 96, + 88, + 903, + 329 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b4fe73b5e24de3e9197a2e76a0b52305bb50c2aed128791fdcec14e62d8763f7.jpg", + "table_caption": [ + "Table 1. Performance on the three coarse-grained benchmarks. The results with THI are marked with $\\star$ . Note that IRRA means using the pre-trained Backbones with MALS [38] and the results of NAM are reproduced by us." + ], + "table_footnote": [], + "table_body": "
MethodsRank-1Rank-5Rank-10mAPmINP
LGUR [25]70.6984.5789.9168.93-
SSAN [5]75.0988.6392.8473.14-
IRRA [12]85.0294.3196.7583.9177.30
RDE [21]87.6095.6597.4686.1079.54
CFAM(B/16) [44]85.5594.5197.0284.23-
CFAM(L/14) [44]88.5195.5897.4987.09-
1 Our ICL89.1796.1397.8887.4981.50
1 Our ICL★90.6795.9897.8688.2982.60
2 Our ICL91.0296.9898.1789.7684.70
2 Our ICL★91.7896.8398.1690.3385.62
", + "bbox": [ + 98, + 375, + 478, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Performance comparison on the UFine6926 dataset. The results of IRRA and RDE are reproduced by us.", + "bbox": [ + 89, + 516, + 482, + 544 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "sure the TIREID performance. Like [12, 21], we also report the mean Average Precision (mAP) and mean Inverse Negative Penalty (mINP) as auxiliary metrics.", + "bbox": [ + 89, + 554, + 483, + 602 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 617, + 308, + 633 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To achieve interactive TIReID, we choose RDE [21] and Qwen2-VL-7B-Instruct [29] as the investigated TIReID method and MLLM. For the sake of fairness, we do not modify any settings of RDE, including model architecture (CLIP-ViTB/16 [24]), hyperparameters, and training parameters. To be compatible with the fine-grained benchmark, following [44], the maximum length of the textual tokens of CLIP is set to 168 by interpolating the positional embedding layer with an initial learning rate of $5e - 5$ . As for fine-tuning MLLMs, we use the Llama-Factory framework [40] to conduct SFT with LoRA [11] for 2 epochs. The train batch size per device is set to 4, the gradient accumulation steps are set to 16, the initial learning rate is $5e - 5$ and the hyperparameters $\\alpha$ and $r$ of LoRA are set to 16 and 8, respectively. During inference, the temperature is set to 0.01 to keep reproducibility. Note that all experiments can be completed on two GeForce RTX 24GB 3090 GPUs.", + "bbox": [ + 88, + 643, + 485, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Comparison with State-of-the-Arts", + "text_level": 1, + "bbox": [ + 511, + 377, + 816, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, to verify the superiority of our ICL, we compare our method with more than 15 baselines including recent advanced methods (e.g., RDE (CVPR'24) [21] and NAM (CVPR'24) [28]). Based on the backbone type, we divide the baselines into two groups (1 and 2) as shown Table 1, i.e., the baselines using VL-Backbones w/o and with ReID-domain pre-training. For the group 2, we use the pretrained weights released by [28] to initialize CLIP for fair comparison. The results are reported in Tables 1 and 2.", + "bbox": [ + 511, + 401, + 906, + 537 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results on coarse-grained datasets. Table 1 report the results evaluated on the coarse-grained datasets. We can see that our method can achieve competitive performance even without THI. By performing THI, the Rank-1 score of our method is greatly improved, e.g., in the group ①, the Rank-1 scores on the three datasets are improved by $1.50\\%$ , $0.91\\%$ , and $2.85\\%$ , respectively. In addition, mAP and mINP scores have also improved greatly, which indicates that the overall ranking has improved. In group ②, our method achieves the best scores on most metrics, especially Rank-1 reached $72.55\\%$ on RSTPReid, which is sufficient to verify the superiority. However, we found that THI slightly degraded Rank-5 and Rank-10, which is because we only conducted 5 rounds of interaction. As long as the locating anchor image is incorrect, the errors will accumulate, which can be relieved by increasing the rounds of interaction. But in general, THI can significantly improve Rank-1 and the overall ranking (mAP and mINP).", + "bbox": [ + 511, + 537, + 908, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results on fine-grained dataset. Table 2 shows the results on the fine-grained dataset whose average text length is over 80. Such kind of fine-grained description often has a clear query intention. For better comparison, we provide the performance of IRRA and RDE on the UFine6926 dataset. From the results, our method can still achieve excellent per", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14395", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0b79a12b6ffe274b85c93fea257ab4a3ef30dcfc5f0e716faffa6089522fbcb6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsTraining SetsCUHK-PEDESICFG-PEDESRSTPReid
Rank-1Rank-5Rank-10mAPmINPRank-1Rank-5Rank-10mAPmINPRank-1Rank-5Rank-10mAPmINP
IRRA [12]CUHK-PEDES73.3889.9393.7166.1350.2442.4162.1169.6221.771.9553.2577.1585.3539.6316.60
ICFG-PEDES33.4856.2966.3331.5619.2063.4680.2585.8238.067.9345.3069.2578.8036.8218.38
RSTPReid32.8055.2665.8130.2917.6132.3049.6757.8020.543.8460.2081.3088.2047.1725.28
RDE [21]CUHK-PEDES75.9490.1494.1267.5651.4448.1866.3073.7025.002.3354.9077.5086.5041.2717.84
ICFG-PEDES38.1159.2468.4434.1620.4467.6882.4787.3640.067.8749.2572.1080.2038.4618.33
RSTPReid36.9458.2267.5833.6520.4242.1758.3265.4926.374.9465.3583.9589.9050.8828.08
Our ICLCUHK-PEDES76.4190.4894.3368.0451.9948.5766.6673.7525.302.4055.8079.6087.6542.0917.41
ICFG-PEDES42.8764.2073.4438.1923.5868.1182.5987.5240.818.1852.5075.0583.0041.8221.14
RSTPReid41.3161.8670.3136.7822.3745.9362.7068.8028.895.6367.7086.0591.7552.6229.36
Our ICL*CUHK-PEDES77.9190.2794.1469.1353.4052.8066.4973.4925.602.4461.3079.2587.4043.4218.01
ICFG-PEDES49.2964.3473.5540.8225.3869.0282.4587.3641.218.3060.1575.3083.1543.7222.04
RSTPReid47.3561.4570.3438.9123.6850.5261.5668.5729.265.7370.5585.9591.6553.6830.13
", + "bbox": [ + 99, + 88, + 903, + 271 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e353bcf20aa123499b0fd678eb9c40ffdf77a8dc77a0d386749cfafb9959544e.jpg", + "table_caption": [ + "Table 3. Comparison of mutual generalization capabilities between coarse-grained datasets." + ], + "table_footnote": [ + "Table 4. Generalization capabilities between coarse-grained and fine-grained datasets. The best scores in each task are in bold." + ], + "table_body": "
Source → TargetMethodsRank-1Rank-5Rank-10mAPmINP
CUHK. → UFine.IRRA [12]37.5154.9264.2940.7634.33
RDE [21]40.3757.4966.0542.6835.78
Our ICL46.4063.5572.0848.6841.56
Our ICL☆57.7664.1372.8153.9745.64
ICFG. → UFine.IRRA [12]15.0226.7933.9017.1012.75
RDE [21]17.8631.0138.5619.8214.74
Our ICL27.9544.2052.2029.8523.20
Our ICL☆36.8144.6552.7334.1226.61
RSTP. → UFine.IRRA [12]13.2125.6733.9315.6011.09
RDE [21]14.0025.2332.6416.2211.90
Our ICL23.8938.3046.7025.5419.20
Our ICL☆31.2338.5647.0228.9021.80
UFine. → CUHKIRRA [12]37.7460.1270.1335.9423.21
RDE [21]39.4161.1470.1136.4923.32
Our ICL49.0470.2778.6444.5429.58
Our ICL☆56.8770.1978.5347.3131.20
UFine. → ICFG.IRRA [12]34.5255.4164.4417.961.95
RDE [21]40.3760.1468.4120.542.19
Our ICL43.1062.9270.7322.732.56
Our ICL☆47.8362.6770.4823.162.62
UFine. → RSTP.IRRA [12]37.6563.7073.0029.0011.80
RDE [21]39.9063.5074.7529.9212.43
Our ICL48.8572.6581.8036.9116.39
Our ICL☆55.3572.4081.5038.6417.23
", + "bbox": [ + 98, + 311, + 482, + 628 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "formance, with Rank-1 exceeding $91\\%$ . This shows that interaction is also applicable to the fine-grained scenario.", + "bbox": [ + 89, + 686, + 482, + 717 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Generalization Study", + "text_level": 1, + "bbox": [ + 89, + 726, + 290, + 742 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To evaluate the generalization, Tables 3 and 4 report the cross-domain performance on four datasets, including the coarse-to-fine, coarse-to-fine, and fine-to-coarse generalization experiments. From the results, our method achieves better generalization than RDE even without THI thanks to the training enhancement of RDA. When THI is performed, the cross-domain performance is dramatically improved, for example, from CUHK-PEDES to RSTPReid, THI brings an improvement of more than $4\\%$ on Rank-1. This shows that THI is a potential solution to generaliza", + "bbox": [ + 89, + 750, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tion challenges in the future. From the generalization experiments between fine-grained and coarse-grained datasets shown in Table 4, ICL can also achieve the best cross-domain performance, e.g., compared with the best baseline RDE, from UFine6926 domain to CUHK-PEDES domain, our method improves Rank-1 and mAP by $17.49\\%$ and $10.86\\%$ , respectively, which further verifies the cross-domain generalization of our method.", + "bbox": [ + 511, + 315, + 906, + 436 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/02a38d9086de556b5f604e246c96e45289c867c683aa4175f322f0a959b640c4.jpg", + "table_caption": [], + "table_footnote": [ + "Table 5. Transferability results on three coarse-grained benchmarks. $\\Delta$ Avg represents the average improvement." + ], + "table_body": "
MethodsTHICUHK-PEDESICFG-PEDESRSTPReidΔAvg
Rank-1mAPRank-1mAPRank-1mAP
CLIP [24]X71.6463.9260.1134.5256.5544.52+2.57
✓73.7765.6663.5734.9561.4546.25
IRRA [12]X73.3866.1363.4638.0660.2047.17+1.86
✓76.0667.4265.2638.5863.7548.47
RDE [21]X75.9467.5667.6840.0665.3550.88+1.41
✓77.4768.6268.7240.6368.4552.01
", + "bbox": [ + 519, + 448, + 903, + 547 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Transferability Study", + "text_level": 1, + "bbox": [ + 511, + 602, + 714, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our interactive module is separate and independent from the training of existing TIReID methods, so it can be plug-and-play with existing methods to improve ReID performance. To verify the transferability of our THI, we conduct experiments on multiple baselines, and the results are shown in Table 5. Except for the CLIP results we reproduced, all other experiments used public pre-trained model weights. From the results of Table 5, the interactive strategy application can significantly improve Rank-1 and mAP, which shows that the external guidance by interactions via MLLMs can further clarify the text-image alignments and improve the overall ranking. The average improvements on the three datasets are $2.57\\%$ , $1.86\\%$ , and $1.41\\%$ , respectively, which seriously proves the transferability of our THI.", + "bbox": [ + 511, + 626, + 906, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.6. Ablation and Parameter Study", + "text_level": 1, + "bbox": [ + 511, + 847, + 785, + 864 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To explore the effects of each proposed component, i.e., THI and RDA, we first conduct the ablation study on three", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14396", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "coarse-grained datasets as reported in Table 6. From the results, each component can bring performance gains on Rank-1 and mAP, which verifies the reliability and rationality of the method. Especially the introduction of THI has greatly improved the Rank-1 accuracy from $68.95\\%$ to $70.55\\%$ on the RSTPreid dataset. In addition, our RDA also brings considerable performance gains, especially on the RSTPreid dataset. Also, we conduct the parameter analysis on the CHUK-PEDES dataset on two free hyperparameters, i.e., the similarity threshold $\\xi$ and the balance factor $\\lambda$ . The former filters unnecessary retrievals to improve the interaction efficiency, while the latter controls the contribution of refined texts. Based on Figure 5, in our experiments, we set $\\xi$ in the range of $0.5 \\sim 0.6$ and $\\lambda$ to 0.8, thus mitigating the risk of introducing noisy external knowledge.", + "bbox": [ + 89, + 90, + 483, + 318 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2fcadfc6a9a88b84ef08d3161ccfe833f7724ab84e154b95bfa4dfaa6b400281.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
No.THIRDALoRACUHK-PEDESICFG-PEDESRSTPReid
Rank-1mAPRank-1mAPRank-1mAP
#1✓✓✓77.9169.3269.0241.2170.5553.68
#2✓✓×76.3868.5967.9241.1369.0053.11
#3×✓×76.4168.0468.1140.8167.7052.62
#4×××75.9467.5667.6840.0665.3550.88
", + "bbox": [ + 98, + 327, + 478, + 406 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Ablation studies on CHUK-PEDES, ICFG-PEDES, and RSTPReid datasets. The best scores are in bold.", + "bbox": [ + 89, + 407, + 482, + 436 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3d7772f515054dae074f670307524539aa3f362bf5097f1182732169c1447a2f.jpg", + "image_caption": [ + "(a) The similarity threshold $\\xi$", + "Figure 5. Variation of performance with different $\\xi$ and $\\lambda$ ." + ], + "image_footnote": [], + "bbox": [ + 91, + 449, + 287, + 547 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c584e6d8acd84e0846325cc373435b3a25138bd51eba753542aedb69fc859d7b.jpg", + "image_caption": [ + "(b) The balance factor $\\lambda$" + ], + "image_footnote": [], + "bbox": [ + 285, + 449, + 483, + 547 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.7. Interactive Study", + "text_level": 1, + "bbox": [ + 89, + 590, + 261, + 606 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This section further explores the interactive module, i.e. THI. We report the performance (mAP) changes after multiple rounds of interactions in Figure 6. The mAP score can reflect the overall retrieval quality. From the results, the performance improvement is obvious in the first few rounds $(\\leq 2)$ since the items that meet the query semantics are mostly concentrated at the top of the ranking. However, after $>2$ rounds, the performance gain is gradually not obvious as the number of queries requiring interaction decreases. But generally, as the number of rounds increases, the overall performance gradually improves. In all our experiments, we performed 5 rounds of interaction. In addition, we visualize the top-10 retrieved results before and after applying THI in Figure 7. Due to the inherent limitations of the intra-model knowledge, it is difficult to obtain satisfactory results without the help of THI. In contrast, our THI can dynamically enrich queries by interacting with MLLMs to achieve a more reliable ranking. More example results are given in the supplementary material.", + "bbox": [ + 89, + 613, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a468f43e12e93e9db9044752c1b7bab4faeafca3d2345fae931fe51558325a3b.jpg", + "image_caption": [ + "(a) CUHK-PEDES" + ], + "image_footnote": [], + "bbox": [ + 514, + 89, + 647, + 180 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/934c89310c3ab053d56bdfe58e5389df97e51d022f40a831122983b2620aa41d.jpg", + "image_caption": [ + "(b) ICFG-PEDES", + "Figure 6. Performance (mAP) versus rounds on three datasets. Round 0 indicates the setting without using THI." + ], + "image_footnote": [], + "bbox": [ + 647, + 88, + 774, + 180 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/543546290431ffbbd3f2bc3c372f4df8c65819deec1bed410650b8973891e94e.jpg", + "image_caption": [ + "(c) RSTPReid" + ], + "image_footnote": [], + "bbox": [ + 774, + 89, + 903, + 180 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d81673e82a8d80de989625b02f3c922b1d4291443a750f08dde17504ee196e20.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 246, + 660, + 321 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ad0c671cfb7951ea6984a1adb15a69a27340c9ac356219ca45d2459bec9af3f6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 660, + 246, + 903, + 320 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fb579c33600b81633192f7179c87f2f24f84bb02991a511821ce96e48471a5bb.jpg", + "image_caption": [ + "Figure 7. Top-10 retrieved results on CUHK-PEDES dataset between ICL (the first row) and ICL with THI (the second row)." + ], + "image_footnote": [], + "bbox": [ + 517, + 325, + 660, + 392 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7042d8b8e506934f997341f636bbb76dfbd27533654a8eaaf277610a97d429f5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 660, + 324, + 903, + 396 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 459, + 633, + 474 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we explore interactive text-to-image person re-identification, which aims to improve the alignment between dynamic queries and challenging candidate images by leveraging external guidance from MLLMs. To achieve this, we develop an Interactive Cross-modal Learning (ICL) framework to alleviate the inherent challenges of offline models and training data by, including a plug-and-play Test-time Human-centered Interaction (THI) module and Reorganization Data Augmentation (RDA). Extensive experiments and analysis show that our framework can effectively transfer external knowledge in MLLMs into offline models for guiding re-identification, showing excellent performance and generalization.", + "bbox": [ + 511, + 487, + 906, + 684 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 513, + 705, + 671, + 722 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported in part by the National Key R&D Program of China under Grant 2024YFB4710604; in part by NSFC under Grants 62472295, 62372315, 62176171, and U21B2040; in part by Sichuan Science and Technology Planning Projects (2024YFHZ0089, 2024NSFTD0049, 2024YFHZ0144, 2024NSFTD0047, 2024NSFTD0038); in part by System of Systems and Artificial Intelligence Laboratory pioneer fund grant; in part by the Fundamental Research Funds for the Central Universities under Grant CJ202403 and CJ202303; in part by Chengdu Science and Technology Project under Grant 2023-XT00-00004-GX.", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14397", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yang Bai, Min Cao, Daming Gao, Ziqiang Cao, Chen Chen, Zhenfeng Fan, Liqiang Nie, and Min Zhang. Rasa: relation and sensitivity aware representation learning for text-based person search. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, pages 555–563, 2023. 2, 6", + "[2] Maryam Bukhari, Sadaf Yasmin, Sheneela Naz, Muazzam Maqsood, Jehyeok Rew, and Seungmin Rho. Language and vision based person re-identification for surveillance systems using deep learning with lip layers. Image and Vision Computing, 132:104658, 2023. 1", + "[3] Min Cao, Yang Bai, Ziyin Zeng, Mang Ye, and Min Zhang. An empirical study of clip for text-based person search. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 465-473, 2024. 1, 2, 6", + "[4] Yuhao Chen, Guoqing Zhang, Yujiang Lu, Zhenxing Wang, and Yuhui Zheng. Tipcb: A simple but effective part-based convolutional baseline for text-based person search. Neurocomputing, 494:171-181, 2022.", + "[5] Zefeng Ding, Changxing Ding, Zhiyin Shao, and Dacheng Tao. Semantically self-aligned network for text-to-image part-aware person re-identification. arXiv preprint arXiv:2107.12666, 2021. 1, 5, 6", + "[6] Hiren Galiyawala and Mehul S Raval. Person retrieval in surveillance using textual query: a review. Multimedia Tools and Applications, 80(18):27343-27383, 2021. 1", + "[7] Daming Gao, Yang Bai, Min Cao, Hao Dou, Mang Ye, and Min Zhang. Semi-supervised text-based person search. arXiv preprint arXiv:2404.18106, 2024. 1", + "[8] Donghoon Han, Eunhwan Park, Gisang Lee, Adam Lee, and Nojun Kwak. Merlin: Multimodal embedding refinement via llm-based iterative navigation for text-video retrieval-erank pipeline. arXiv preprint arXiv:2407.12508, 2024. 2, 3", + "[9] Chen He, Shenshen Li, Zheng Wang, Hua Chen, Fumin Shen, and Xing Xu. Chatting with interactive memory for text-based person retrieval. Multimedia Systems, 31(1):31, 2025. 3", + "[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 2", + "[11] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 4, 6", + "[12] Ding Jiang and Mang Ye. Cross-modal implicit relation reasoning and aligning for text-to-image person retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2787-2797, 2023. 1, 2, 6, 7", + "[13] Matan Levy, Rami Ben-Ari, Nir Darshan, and Dani Lischinski. Chatting makes perfect: Chat-based image retrieval. Advances in Neural Information Processing Systems, 36, 2024. 2, 3", + "[14] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022. 2", + "[15] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 3", + "[16] Shuang Li, Tong Xiao, Hongsheng Li, Bolei Zhou, Dayu Yue, and Xiaogang Wang. Person search with natural language description. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1970-1979, 2017. 5", + "[17] Yunfan Li, Peng Hu, Dezhong Peng, Jiancheng Lv, Jianping Fan, and Xi Peng. Image clustering with external guidance. In *Forty-first International Conference on Machine Learning*. 2", + "[18] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 3", + "[19] Yang Qin, Dezhong Peng, Xi Peng, Xu Wang, and Peng Hu. Deep evidential learning with noisy correspondence for cross-modal retrieval. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4948-4956, 2022. 1", + "[20] Yang Qin, Yuan Sun, Dezhong Peng, Joey Tianyi Zhou, Xi Peng, and Peng Hu. Cross-modal active complementary learning with self-refining correspondence. Advances in neural information processing systems, 36:24829-24840, 2023. 1", + "[21] Yang Qin, Yingke Chen, Dezhong Peng, Xi Peng, Joey Tianyi Zhou, and Peng Hu. Noisy-correspondence learning for text-to-image person re-identification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27197-27206, 2024. 1, 2, 6, 7", + "[22] Yalan Qin, Chuan Qin, Xinpeng Zhang, and Guorui Feng. Dual consensus anchor learning for fast multi-view clustering. IEEE Transactions on Image Processing, 2024. 1", + "[23] Yalan Qin, Nan Pu, Nicu Sebe, and Guorui Feng. Latent space learning based ensemble clustering. IEEE Transactions on Image Processing, 2025. 1", + "[24] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 6, 7", + "[25] Zhiyin Shao, Xinyu Zhang, Meng Fang, Zhifeng Lin, Jian Wang, and Changxing Ding. Learning granularity-unified representations for text-to-image person re-identification. In Proceedings of the 30th ACM International Conference on Multimedia, 2022. 6", + "[26] Xiujun Shu, Wei Wen, Haoqian Wu, Keyu Chen, Yiran Song, Ruizhi Qiao, Bo Ren, and Xiao Wang. See finer, see more: Implicit modality alignment for text-based person retrieval. In European Conference on Computer Vision, pages 624-641. Springer, 2022. 6" + ], + "bbox": [ + 516, + 93, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14398", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Yuan Sun, Yang Qin, Yongxiang Li, Dezhong Peng, Xi Peng, and Peng Hu. Robust multi-view clustering with noisy correspondence. IEEE Transactions on Knowledge and Data Engineering, 2024. 1", + "[28] Wentan Tan, Changxing Ding, Jiayu Jiang, Fei Wang, Yibing Zhan, and Dapeng Tao. Harnessing the power of mllms for transferable text-to-image person reid. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17127-17137, 2024. 1, 2, 3, 6", + "[29] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 3, 6", + "[30] Yansong Wang, Yundong Sun, Yansheng Fu, Dongjie Zhu, and Zhaoshuo Tian. Spectrum-bert: pre-training of deep bidirectional transformers for spectral classification of Chinese liquors. IEEE Transactions on Instrumentation and Measurement, 2024. 2", + "[31] Yabing Wang, Le Wang, Qiang Zhou, Zhibin Wang, Hao Li, Gang Hua, and Wei Tang. Multimodal llm enhanced cross-lingual cross-modal retrieval. arXiv preprint arXiv:2409.19961, 2024. 3", + "[32] Yuhao Wang, Yang Liu, Aihua Zheng, and Pingping Zhang. Demo: Decoupled feature-based mixture of experts for multi-modal object re-identification. In Proceedings of the AAAI Conference on Artificial Intelligence, 2025. 1", + "[33] Zhe Wang, Zhiyuan Fang, Jun Wang, and Yezhou Yang. Vistaa: Visual-textual attributes alignment in person search by natural language. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XII 16, pages 402-420. Springer, 2020. 1, 2", + "[34] Zijie Wang, Aichun Zhu, Jingyi Xue, Xili Wan, Chao Liu, Tian Wang, and Yifeng Li. Caibc: Capturing all-round information beyond color for text-based person retrieval. In Proceedings of the 30th ACM International Conference on Multimedia, pages 5314-5322, 2022. 1", + "[35] Yushuang Wu, Zizheng Yan, Xiaoguang Han, Guanbin Li, Changqing Zou, and Shuguang Cui. Lipscore: language-guided person search via color reasoning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1624-1633, 2021. 2", + "[36] Shuanglin Yan, Neng Dong, Jun Liu, Liyan Zhang, and Jinhui Tang. Learning comprehensive representations with richer self for text-to-image person re-identification. In Proceedings of the 31st ACM international conference on multimedia, pages 6202–6211, 2023. 6", + "[37] Shuanglin Yan, Neng Dong, Liyan Zhang, and Jinhui Tang. Clip-driven fine-grained text-image person re-identification. IEEE Transactions on Image Processing, 2023. 1, 6", + "[38] Shuyu Yang, Yinan Zhou, Zhedong Zheng, Yaxiong Wang, Li Zhu, and Yujiao Wu. Towards unified text-based person retrieval: A large-scale multi-attribute and language search benchmark. In Proceedings of the 31st ACM International Conference on Multimedia, pages 4492–4501, 2023. 1, 2, 6" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Pingping Zhang, Yuhao Wang, Yang Liu, Zhengzheng Tu, and Huchuan Lu. Magic tokens: Select diverse tokens for multi-modal object re-identification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17117-17126, 2024. 1", + "[40] Yaowei Zheng, Richong Zhang, Junhao Zhang, Yanhan Ye, Zheyan Luo, Zhangchi Feng, and Yongqiang Ma. Llamafactory: Unified efficient fine-tuning of $100+$ language models. arXiv preprint arXiv:2403.13372, 2024. 6", + "[41] Aichun Zhu, Zijie Wang, Yifeng Li, Xili Wan, Jing Jin, Tian Wang, Fangqiang Hu, and Gang Hua. Dssl: Deep surroundings-person separation learning for text-based person retrieval. In Proceedings of the 29th ACM International Conference on Multimedia, pages 209-217, 2021. 1, 2, 5", + "[42] Jialong Zuo, Changqian Yu, Nong Sang, and Changxin Gao. Plip: Language-image pre-training for person representation learning, 2023. 1", + "[43] Jialong Zuo, Ying Nie, Hanyu Zhou, Huaxin Zhang, Haoyu Wang, Tianyu Guo, Nong Sang, and Changxin Gao. Cross-video identity correlating for person re-identification pretraining. arXiv preprint arXiv:2409.18569, 2024. 1", + "[44] Jialong Zuo, Hanyu Zhou, Ying Nie, Feng Zhang, Tianyu Guo, Nong Sang, Yunhe Wang, and Changxin Gao. Ufinebench: Towards text-based person retrieval with ultra-fine granularity. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22010-22019, 2024. 1, 5, 6" + ], + "bbox": [ + 516, + 92, + 905, + 470 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14399", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Human-centered Interactive Learning via MLLMs for Text-to-Image Person Re-identification/5cc75be5-e122-4a4f-ba40-7450ab436d6a_model.json b/2025/Human-centered Interactive Learning via MLLMs for Text-to-Image Person Re-identification/5cc75be5-e122-4a4f-ba40-7450ab436d6a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a7e8a9c0105eb41e0963aba1e42b6f2205f1129d --- /dev/null +++ b/2025/Human-centered Interactive Learning via MLLMs for Text-to-Image Person Re-identification/5cc75be5-e122-4a4f-ba40-7450ab436d6a_model.json @@ -0,0 +1,2112 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.152, + 0.131, + 0.848, + 0.175 + ], + "angle": 0, + "content": "Human-centered Interactive Learning via MLLMs for Text-to-Image Person Re-identification" + }, + { + "type": "text", + "bbox": [ + 0.157, + 0.203, + 0.846, + 0.223 + ], + "angle": 0, + "content": "Yang Qin\\(^{1}\\) Chao Chen\\(^{2}\\) Zhihang Fu\\(^{2}\\) Dezhong Peng\\(^{1,4,5}\\) Xi Peng\\(^{1,3}\\) Peng Hu\\(^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.223, + 0.758, + 0.24 + ], + "angle": 0, + "content": "1College of Computer Science, Sichuan University 2Independent Researcher" + }, + { + "type": "text", + "bbox": [ + 0.12, + 0.24, + 0.877, + 0.258 + ], + "angle": 0, + "content": "\\(^{3}\\)National Key Laboratory of Fundamental Algorithms and Models for Engineering Simulation, Sichuan University" + }, + { + "type": "text", + "bbox": [ + 0.142, + 0.258, + 0.851, + 0.275 + ], + "angle": 0, + "content": "Sichuan National Innovation New Vision UHD Video Technology Co., Ltd Tianfu Jincheng Laboratory" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.309, + 0.327, + 0.325 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.342, + 0.485, + 0.765 + ], + "angle": 0, + "content": "Despite remarkable advancements in text-to-image person re-identification (TIREID) facilitated by the breakthrough of cross-modal embedding models, existing methods often struggle to distinguish challenging candidate images due to intrinsic limitations, such as network architecture and data quality. To address these issues, we propose an Interactive Cross-modal Learning framework (ICL), which leverages human-centered interaction to enhance the discriminability of text queries through external multimodal knowledge. To achieve this, we propose a plug-and-play Test-time Humane-centered Interaction (THI) module, which performs visual question answering focused on human characteristics, facilitating multi-round interactions with a multimodal large language model (MLLM) to align query intent with latent target images. Specifically, THI refines user queries based on the MLLM responses to reduce the gap to the best-matching images, thereby boosting ranking accuracy. Additionally, to address the limitation of low-quality training texts, we introduce a novel Reorganization Data Augmentation (RDA) strategy based on information enrichment and diversity enhancement to enhance query discriminability by enriching, decomposing, and reorganizing person descriptions. Extensive experiments on four TIREID benchmarks, i.e., CUHK-PEDES, ICFG-PEDES RSTPReid, RSTPReid, and UFine6926, demonstrate that our method achieves remarkable performance with substantial improvement. Code is available at https://github.com/QinYang79/ICL." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.792, + 0.222, + 0.808 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.818, + 0.484, + 0.879 + ], + "angle": 0, + "content": "Recently, text-to-image person re-identification (TIREID) [3-5, 7, 33, 34, 41] has made great progress in aligning text descriptions and person images, enabling high-accuracy person search and identification. Different from traditional" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.308, + 0.907, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.663, + 0.387, + 0.764, + 0.397 + ], + "angle": 0, + "content": "(a) Existing methods" + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.401, + 0.907, + 0.511 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.658, + 0.513, + 0.768, + 0.524 + ], + "angle": 0, + "content": "(b) Interactive TIREID" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.527, + 0.907, + 0.68 + ], + "angle": 0, + "content": "Figure 1. The illustration of our motivation. When performing text-based person re-identification, (a) existing methods commonly exploit cross-modal models to calculate similarity (such as IRRA [12] and RDE [21]) and then obtain the candidate person images by ranking. However, due to the intrinsic limitations of models and training data, they may not be able to distinguish challenging candidate images well enough to obtain satisfactory results. (b) Our motivation is to interact with the system for external guidance like a human and gradually refine the user's query from the candidate items shown in (a) by multiple rounds of question-answering, ultimately improving the overall ranking." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.688, + 0.907, + 0.824 + ], + "angle": 0, + "content": "image-to-image re-identification [32, 39, 42, 43], TIREID, as a rising task in the multimodal community [19, 20, 22, 23, 27], retrieves person images using user-customized text queries, which is more practical in many scenarios where an image query is unavailable, such as in video surveillance systems [2] or crowd management [6]. A primary challenge in TIREID is learning to associate text queries with the corresponding person images at a fine-grained level, bridging the modality gap for accurate similarity measurement." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.903 + ], + "angle": 0, + "content": "To overcome this challenge, recent efforts have developed various strategies to improve cross-modal associations, including local attribute modeling [37, 44], loss function designing [12, 21], ReID-domain pre-training [28, 38], etc. Although these methods achieve promising perfor" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.424, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author: Peng Hu (penghu.ml@gmail.com)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14390" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.364 + ], + "angle": 0, + "content": "mance, they are limited by the inherent defect of the offline models and training data, making it hard to handle dynamic user query inputs and leading to poor generalization. In practice, however, different users tend to input concise, vague, and diverse text queries based on their memories to the target person, thereby raising challenges for the offline models to understand fine-grained human characteristics from the queries. For example, as illustrated in Figure 1 (a), the user text query \"A woman walking visible from the back is wearing a white shirt, and black pants.\" fails to capture some important contextual details like handheld objects or background, resulting in inaccurate candidate images. In brief, it is hard for offline models to handle dynamic and challenging queries solely relying on their learned internal knowledge. To break through this bottleneck, we could seek to align queries with image candidates with the help of external knowledge, thereby reranking the retrieved images and boosting the identification accuracy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.368, + 0.483, + 0.61 + ], + "angle": 0, + "content": "Recent studies [8, 13, 17] have explored integrating external knowledge from pre-trained models or (multimodal) large language models to guide various downstream tasks. For instance, Li et al. [17] leveraged multimodal knowledge of pre-trained vision-language models [24] to enhance unimodal image representations for external-guided clustering. However, this approach cannot achieve post-hoc improvement for offline models, leading to infeasibility in practice due to the prohibitive cost of fine-tuning or even retraining. In cross-modal retrieval, Han et al. [8] exploited the excellent comprehension and generation capabilities of multimodal large language models (MLLMs) to design an interactive re-ranking pipeline for text-to-video retrieval. However, Han et al. [8] introduces label leakage by using the ground truth video as the video in mind for the MLLM answerer agent, which is not applicable in actual testing." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In this paper, we present a novel Interactive Cross-modal Learning framework (ICL) that exploits the multimodal knowledge implicit in MLLMs to enhance the alignment between text queries and target person images. Specifically, we first propose a plug-and-play Test-time Human-centered Interaction (THI) module that refines text queries through interactions with an MLLM, enhancing the post-hoc ability of trained models to distinguish challenging candidates. THI identifies the latent target images through multi-round interactions with MLLMs, asking human-centered questions to the MLLM for fine-grained answers about the images, which refine the query texts to strengthen the alignment with the images, ultimately address the inherent limitations of the input queries and enhancing ranking accuracy. To address the intrinsic limitations in the trained models, we propose a Reorganization Data Augmentation strategy (RDA) to enrich and diversify pedestrian texts for training enhancement, thereby transferring external knowledge in MLLMs into the model. To enrich the texts, RDA applies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.257 + ], + "angle": 0, + "content": "visual question answering via the MLLM on each image-text pair to supplement fine-grained person characteristics. Moreover, to enhance the diversity of texts, RDA presents a decomposition-reorganization strategy to decompose person descriptions into attribute-specific sub-sentences (e.g., clothes, pants, shoes, etc.), and rewrite them with MLLM into multiple sentences with the same meaning. Subsequently, the rewritten sentences are randomly reordered and recombined to generate varied augmented texts, thus boosting data diversity and enhancing the generalization of models. Our main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.259, + 0.905, + 0.335 + ], + "angle": 0, + "content": "- We introduce human-centered interaction to TIREID, proposing a novel MLLM-driven Interactive Cross-modal Learning framework (ICL), which leverages external knowledge to overcome the inherent limitations of existing offline methods in handling dynamic queries." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.335, + 0.905, + 0.395 + ], + "angle": 0, + "content": "- A plug-and-play Test-time Human-centered Interaction module (THI) is presented to align query intent with latent target images through multi-round interactions with MLLMs, improving the ranking quality." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.396, + 0.905, + 0.455 + ], + "angle": 0, + "content": "- An effective Reorganization Data Augmentation strategy (RDA), applying MLLMs for text decomposition and recombination, is developed to generate discriminative and diverse training texts, improving cross-modal learning." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.456, + 0.905, + 0.515 + ], + "angle": 0, + "content": "- Extensive experiments on four text-to-image person re-identification benchmarks verify the effectiveness and superiority of our method, which achieves promising performance and superior generalization." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.259, + 0.905, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.534, + 0.655, + 0.549 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.56, + 0.847, + 0.575 + ], + "angle": 0, + "content": "2.1. Text-to-Image Person Re-identification" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.907, + 0.903 + ], + "angle": 0, + "content": "As a challenging topic in multimodal learning, text-to-image person re-identification (TIREID) aims to search the image of the target person with a given natural language query. Existing methods [1, 12, 21, 28, 41] can be roughly divided into three categories according to the used backbone type: unimodal backbones, general multimodal pre-trained backbones, and ReID-domain multimodal pre-trained backbones. The early TIREID methods [33, 35] commonly use modality-specific unimodal backbones to encode text or images for cross-modal alignment, e.g., ResNet [10], BERT [30], etc. Recently, benefiting from the rapid development of visual-language pre-training models [14, 24], more and more researchers [3, 12, 21] have begun to use general pre-trained models as the backbone networks for solving TIREID, hoping that the pre-trained alignment knowledge can improve modality representation and cross-modal alignment. However, the performance improvement is still limited due to the domain gap between the general pre-training data and TIREID tasks. To this end, some attempts [28, 38] exploit image-caption models or MLLMs to annotate images, thus obtaining a large number of image" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14391" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.09, + 0.885, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.322, + 0.908, + 0.392 + ], + "angle": 0, + "content": "Figure 2. The illustration of our Test-time Human-centered Interaction (THI) module. THI includes \\(K\\) rounds of interactions to align query intention with the latent target image by external guidance, where in each round, we perform human-centered visual question answering around fine-grained person attributes to enhance the semantic consistency between the query and the intended person image, and then improve the final ReID performance on the large-scale evaluation through efficient re-ranking. Besides, we perform supervised fine-tuning via LoRA to inspire the discriminate ability of MLLM for ReID domain images and better align queries with latent target images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.407, + 0.484, + 0.513 + ], + "angle": 0, + "content": "text pairs for ReID-domain pre-training. Although promising performance has been achieved, they still cannot get rid of the inherent defect of offline models during handling dynamic queries. In this paper, we propose a novel interactive cross-modal learning framework based on MLLMs, which exploits dynamic interactions for external guidance to improve the generalization ability." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.528, + 0.34, + 0.545 + ], + "angle": 0, + "content": "2.2. Visual Interactive Learning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.554, + 0.484, + 0.901 + ], + "angle": 0, + "content": "With the booming development of multimodal large language models (MLLMs) [15, 18, 29], exploiting visual interaction to improve downstream tasks has attracted increasing attention from researchers. We collectively refer to these as visual interaction learning and divide them into offline interaction [28, 31] and online interaction [8, 9, 13] groups by engaging in specific task periods. The former aims to use interactions to improve the training/pre-training data, such as information richness and diversity. It is usually static and separated from users, which is the mainstream of existing TIREID methods, e.g., Tan et al. [28] apply MLLMs to conduct interactions with multiple prompt templates for better diversity. Unlike them, online interactions are designed to be able to use real-time feedback to improve training or testing. For example, Levy et al. [13] propose to gradually clarify user intention through dialogue to improve image retrieval performance. Likewise, Han et al. [8] develop a rerank pipeline based on LLM-based iterative navigation. However, these methods cannot be directly used for TIREID effectively due to the differences in tasks and data domains. In this paper, we develop a test-time module based on MLLMs to conduct interactions for external guidance, overcoming the limitations of the intra-" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.407, + 0.807, + 0.421 + ], + "angle": 0, + "content": "model knowledge in TIReID offline models." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.436, + 0.649, + 0.454 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.462, + 0.907, + 0.584 + ], + "angle": 0, + "content": "In this section, we introduce an Interactive Cross-modal Learning framework (ICL), which consists of two core comments to address the inherent challenges in offline models and training data, i.e., Test-time Human-centered Interaction module (THI) and Reorganization Data Augmentation (RDA). In Section 3.1, we provide the necessary definitions to facilitate the study. Then, we outline the details of our THI and RDA in Sections 3.2 and 3.3, respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.593, + 0.795, + 0.609 + ], + "angle": 0, + "content": "3.1. Problem and Symbol Statement" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.616, + 0.907, + 0.813 + ], + "angle": 0, + "content": "Suppose that we have the query text \\( q \\in \\mathcal{Q} \\), a pedestrian image \\( v \\in \\mathcal{V} \\), where \\( \\mathcal{Q} \\) and \\( \\mathcal{V} \\) are the text query set and image gallery. The purpose of TIReID is to utilize the text query to retrieve the ideal images from the gallery, thus achieving person searching by identities of retrieved images. To achieve this, existing methods usually train an offline cross-modal model \\( f_{\\mathrm{cross}} = \\{f,g\\} \\) to measure the similarities between text queries and pedestrian images. For a text \\( q \\) and an image \\( v \\), the similarity to measure the matching degree can be represented as \\( S_{q,v} \\equiv \\operatorname{Sim}(f(q), g(v)) \\), where \\( \\operatorname{Sim}(*) \\) is the similarity function, \\( f \\) and \\( g \\) are the text and image encoders, respectively. Then, we can utilize \\( q \\) to search relevant candidate images \\( \\hat{v} \\) from \\( \\mathcal{V} \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.583, + 0.824, + 0.905, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {V}} (q) = \\left\\{\\hat {v} _ {k} \\right\\} _ {k = 1} ^ {K} = \\operatorname {T o p} - K _ {v \\in \\mathcal {V}} \\left(S _ {q, v}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.855, + 0.907, + 0.902 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathcal{V}}(q)\\) is the candidate set for \\(q\\) and \\(K\\) is the number of candidate images. In addition, our ICL also involves MLLMs to conduct interactions and training augmentation," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14392" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.138 + ], + "angle": 0, + "content": "which we denote as \\(\\mathcal{M}\\), and the prompt template function for \\(\\mathcal{M}\\) is denoted by the symbol \\(\\mathcal{T}\\). Based on the above definitions, we will elaborate on ICL in the following sections." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.147, + 0.414, + 0.162 + ], + "angle": 0, + "content": "3.2. Test-time Humane-center Interaction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.17, + 0.484, + 0.352 + ], + "angle": 0, + "content": "Anchor Localization. Due to the model bottleneck, it is hard for existing methods to distinguish challenging candidate images relying solely on internal knowledge. Thus, we exploit the fine-grained image understanding capability of MLLMs to conduct external guidance. We first propose asking questions such as \"Can this text accurately describe the image?\" to let MLLMs explicitly tell us the latent target (anchor) image. As shown in Figure 2, given a text query \\( q \\), we first get the corresponding candidate set \\( \\hat{\\mathcal{V}}(q) \\) and then perform multiple \\( (K) \\) rounds of interactions according to the ranking until the ideal image is determined. For the \\( k \\)-th round, the answer for interaction can be represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.209, + 0.363, + 0.483, + 0.382 + ], + "angle": 0, + "content": "\\[\na _ {\\hat {v} _ {k}} ^ {q} = \\mathcal {M} \\left(\\mathcal {T} _ {\\text {l o c}} (q, \\hat {v} _ {k})\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.392, + 0.484, + 0.528 + ], + "angle": 0, + "content": "where \\( k \\in \\{1, \\dots, K\\} \\), \\( a_{\\hat{v}_k}^q \\) is the answer of 'Yes' or 'No', \\( \\hat{v}_k \\) is the Top-\\( k \\) candidate image for query \\( q \\), and \\( T_{\\mathrm{loc}} \\) is the prompt template function for anchor localization. However, since there is a domain gap between ReID images and the generic images used for pre-training or instruction-tuning MLLMs, the answers of MLLMs are often unreliable. To solve this, we exploit LoRA [11] to perform supervised fine-tuning (SFT) on MLLMs to inspire the fine-grained ability to identify the person image and the SFT loss is:" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.54, + 0.483, + 0.585 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {l o r a} = - \\sum_ {(x, y) \\in \\mathcal {Z}} \\sum_ {t = 1} ^ {| y |} \\log \\left(p _ {\\mathcal {M}} \\left(y _ {t} \\mid x, y < t\\right)\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.597, + 0.483, + 0.795 + ], + "angle": 0, + "content": "where \\( x \\) is the input Prompt and \\( y \\) is the output Response, \\( \\mathcal{Z} = \\{\\mathcal{Z}^{+},\\mathcal{Z}^{-}\\} \\) is the SFT dataset. For \\( \\mathcal{Z}^{+} \\), we select part of training texts \\( (\\{q_i\\}_{i = 1}^{N_l}) \\) and the corresponding ground-truth images \\( (\\{v_{i}^{+}\\}_{i = 1}^{N_{l}}) \\) from the training set to construct the input Pormpts, i.e., \\( \\{\\mathcal{T}_{\\mathrm{loc}}(q_i,v_i^+)\\}_{i = 1}^{N_l} \\) expecting MLLMs to output the Response of 'Yes'. Similarly, for \\( \\mathcal{Z}^{-} \\), we also use the text queries and the corresponding negative images from the training set to construct the input Pormpts, expecting MLLMs to output the Response of 'No'. To make \\( \\mathcal{Z}^{-} \\) more discriminative, we use a TIREID pre-trained cross-modal model to obtain the Top-10 image with different person ID as input negative sample image for each text query by similarity ranking." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Human-centered VQA. For each round, once we confirm the response of 'Yes', we will ask the anchor image \\((\\bar{v})\\) a series of questions to learn more details about pedestrian characteristics. These details can help us alleviate the discrepancy between the user query and the information within the anchor image, thus improving overall ranking by refining the query. We call this process human-centered visual" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.088, + 0.644, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.531, + 0.164, + 0.632, + 0.175 + ], + "angle": 0, + "content": "(a) CUHK-PEDES" + }, + { + "type": "image", + "bbox": [ + 0.645, + 0.089, + 0.776, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.663, + 0.164, + 0.758, + 0.175 + ], + "angle": 0, + "content": "(b) ICFG-PEDES" + }, + { + "type": "image", + "bbox": [ + 0.776, + 0.089, + 0.906, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.803, + 0.164, + 0.878, + 0.175 + ], + "angle": 0, + "content": "(c) RSTPReid" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.18, + 0.907, + 0.222 + ], + "angle": 0, + "content": "Figure 3. The similarity statistics of the Top-1 retrieved items in the test sets of the three benchmarks. It is obvious that the items with higher similarity are more likely to be correct retrievals." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.234, + 0.85, + 0.25 + ], + "angle": 0, + "content": "question answering (VQA), which is expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.256, + 0.905, + 0.275 + ], + "angle": 0, + "content": "\\[\nr _ {\\bar {v}} = \\mathcal {M} \\left(\\mathcal {T} _ {\\mathrm {v q a}} \\left(\\left\\{c _ {i} \\right\\} _ {i = 1} ^ {N _ {q}}, \\bar {v}\\right)\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.283, + 0.907, + 0.468 + ], + "angle": 0, + "content": "where \\(\\{c_i\\}_{i=1}^{N_q}\\) are \\(N_q\\) questions directed at the detail characteristics (e.g., gender, hair, upper body, lower body, shoes, etc.) of the pedestrian image, \\(r_{\\bar{v}}\\) is the answer to \\(\\{c_i\\}_{i=1}^{N_q}\\) one by one, and \\(\\mathcal{T}_{\\mathrm{vqa}}\\) is the prompt template function for human-centered VQA. To improve the consistency between the image and the user query, a natural strategy is to concatenate these one-by-one answers after the original text. However, due to the limitation of the maximum text length that can be processed by the text model (e.g., it is usually 77 for CLIP), this will destroy the sequence structure. To this end, we recommend merging these answers and the original text by MLLMs to obtain a fluent and concise text, i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.637, + 0.476, + 0.905, + 0.492 + ], + "angle": 0, + "content": "\\[\n\\hat {q} = \\mathcal {M} \\left(\\mathcal {T} _ {\\text {a g g r}} \\left(r _ {\\bar {v}}, q\\right)\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.499, + 0.906, + 0.529 + ], + "angle": 0, + "content": "where \\(\\hat{q}\\) is the merged text query and \\(\\mathcal{T}_{\\mathrm{aggr}}\\) is the prompt template function for text aggregation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.53, + 0.907, + 0.877 + ], + "angle": 0, + "content": "Efficient Re-ranking. To apply the above process to large-scale evaluation, efficiency is a factor that needs to be considered in the retrieval task, since introducing MLLMs to interactions for external guidance inevitably brings additional inference costs. To improve the interaction efficiency, we recommended adopting different strategies for different rounds of interactions. We observe that the retrieval quality is positively correlated with the similarity between the query and the Top-1 candidate image, as shown in Figure 3. To this end, in the first round, we only interact with images whose cross-modal similarity is greater than a threshold \\(\\xi\\) and the answer of anchor localization is 'Yes'. This is because the retrievals with low similarities are more likely to be wrong items. After the first round, we only interact with the image that is all considered 'No' in the previous rounds of anchor localization and whose cross-modal similarity is less than the threshold \\(\\xi\\). This allows us to reduce unnecessary interactions and find text queries that really require interactions, making our method more efficient on large-scale evaluation. Given a query \\(q\\), if the above constraints are met, the interaction is carried out and we can get merged text \\(\\hat{q}\\) after human-centered VQA, the re-ranking similarity \\(\\hat{S}_{q,v}\\) of any image \\(v\\) in \\(\\mathcal{V}\\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.612, + 0.884, + 0.905, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\hat {S} _ {q, v} = \\lambda S _ {q, v} + (1 - \\lambda) \\bar {S} _ {\\hat {q}, v}, \\tag {6}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14393" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.09, + 0.485, + 0.184 + ], + "angle": 0, + "content": "where \\(\\bar{S}_{\\hat{q},v} \\equiv 1\\) if \\(v\\) is \\(\\hat{v}_1\\), otherwise, \\(\\bar{S}_{\\hat{q},v} \\equiv S_{\\hat{q},v}\\), and \\(\\lambda \\in [0,1]\\) is a hyperparameter to balance the contribution of the raw query and the refined text. To make our approach clearer, we describe the detailed algorithm process of our THI in Algorithm 1. Due to space limitations, all prompt templates can be found in the supplementary material." + }, + { + "type": "code_caption", + "bbox": [ + 0.093, + 0.196, + 0.411, + 0.212 + ], + "angle": 0, + "content": "Algorithm 1 The interaction process of our THI" + }, + { + "type": "algorithm", + "bbox": [ + 0.093, + 0.215, + 0.484, + 0.64 + ], + "angle": 0, + "content": "Input: The query set \\(\\mathcal{Q}\\), the image gallery \\(\\mathcal{V}\\), the offline model \\(f_{\\mathrm{cross}}\\), the MLLM \\(\\mathcal{M}\\), the similarity threshold \\(\\xi\\), the number of interaction rounds \\(K\\); \n1: Obtain candidate sets \\(\\{\\hat{\\mathcal{V}}(q_i)\\}_{i=1}^{|\\mathcal{Q}|}\\) for all queries in \\(\\mathcal{Q}\\) via Equation (1); \n2: for \\(k = 1,2,\\dots,K\\) do \n3: for \\(i = 1,2,\\dots,|\\mathcal{Q}|\\) do \n4: Conduct anchor localization via Equation (2) and output the answer of \\(a_{\\hat{v}_k^i}^{q_i}\\) based on the \\(k\\)-th candidate image \\(\\hat{v}_k^i\\) in \\(\\hat{\\mathcal{V}}(q_i)\\); \n5: if \\(a_{\\hat{v}_k^i}^{q_i}\\) shows 'Yes', \\(k = 1\\), and \\(S_{q_i,\\hat{v}_1^i} > \\xi\\) then \n6: Conduct human-centered VQA via Equations (4) and (5) to get the refined query \\(\\hat{q}_i\\); \n7: Compute the re-ranking similarities between query \\(q_i\\) and all images via Equation (6); \n8: end if \n9: if \\(\\{a_{\\hat{v}_j^i}\\}_{j=1}^{k-1}\\) all show 'No', \\(a_{\\hat{v}_k^i}^{q_i}\\) shows 'Yes', \\(k > 1\\), and \\(S_{q_i,\\hat{v}_1^i} \\leq \\xi\\) then \n10: Conduct human-centered VQA via Equations (4) and (5) to get the refined query \\(\\hat{q}_i\\); \n11: Compute the re-ranking similarities between query \\(q_i\\) and all images via Equation (6); \n12: end if \n13: end for \n14: end for \n15: Re-ranking based on similarities; \nOutput: The new candidate images." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.665, + 0.4, + 0.681 + ], + "angle": 0, + "content": "3.3. Reorganization Data Augmentation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.688, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Although THI can exploit interactions to provide external knowledge during test time to improve retrieval quality, the cross-modal embedding model is still a bottleneck that limits its further performance improvement. To this end, as shown in Figure 4, we introduce a new Reorganization Data Augmentation (RDA) strategy by enriching, decomposing, and reorganizing person descriptions to improve the discriminability and diversity of training data. We first apply Human-centered VQA in THI to obtain relevant characteristic descriptions and merge them with the original text, thus obtaining a text with richer information, i.e., \\(\\hat{q} = \\mathcal{M}(\\mathcal{T}_{\\mathrm{aggr}}(\\mathcal{M}(\\mathcal{T}_{\\mathrm{vqa}}(\\{c_i\\}_{i=1}^{N_q}, v)), q))\\). Then, we apply MLLMs to decompose enriched text \\(\\hat{q}\\) into multiple independent sub-sentences that describe individual at-" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.089, + 0.907, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.249, + 0.907, + 0.32 + ], + "angle": 0, + "content": "Figure 4. The illustration of our RDA. The purpose of RDA is to supplement more details to the original training texts through human-centered VQA, improving the discriminability of texts. In addition, to enhance diversity, RDA maximizes diversity through the Decomposition-Rewriting-Reorganization strategy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.345, + 0.907, + 0.496 + ], + "angle": 0, + "content": "tributes without interfering with each other, i.e., \\(\\{\\tilde{q}_i\\}_{i=1}^n = \\mathcal{M}(\\mathcal{T}_{\\mathrm{dec}}(\\hat{q}))\\), where \\(\\{\\tilde{q}_i\\}_{i=1}^n\\) are the \\(n\\) sub-sentences and \\(\\mathcal{T}_{\\mathrm{dec}}\\) is the prompt template function for text decomposition. The purpose of decomposition is to be able to reorganize the text in a different order later. To increase diversity, we rewrite each sub-sentence into multiple sentences with different styles but the same meaning. For sub-sentence \\(\\tilde{q}\\), we conduct rewriting by \\(\\mathcal{R} = \\{\\check{q}_j\\}_{j=1}^m = \\mathcal{M}(\\mathcal{T}_{\\mathrm{rwt}}(\\tilde{q}))\\), where \\(\\mathcal{R}\\) is the set with \\(m\\) rewritten sentences and \\(\\mathcal{T}_{\\mathrm{rwt}}\\) is the prompt template function for text rewriting." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.497, + 0.909, + 0.633 + ], + "angle": 0, + "content": "Finally, for a text \\( q \\), we can obtain the set \\( \\{\\mathcal{R}_i\\}_{i=1}^n \\) that contains a large number of style sub-sentences. So far, we can obtain augmented texts with different sub-sentence orders and different style combinations as shown in Figure 4, which we represent as \\( \\check{q} \\). During training, we can mix the augmentation texts \\( (\\check{q}) \\) with the original texts \\( (q) \\) for cross-modal learning, thereby improving model-intra knowledge and generalization. Due to space limitations, more training details can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.646, + 0.646, + 0.663 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.672, + 0.907, + 0.718 + ], + "angle": 0, + "content": "In this section, we conduct extensive experiments to verify the effectiveness, superiority, and generalization of the proposed ICL on four public benchmark datasets." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.727, + 0.813, + 0.741 + ], + "angle": 0, + "content": "4.1. Datasets and Evaluation Protocols" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.871 + ], + "angle": 0, + "content": "Datasets. In our experiments, we use three coarse-grained benchmarks, CHUK-PEDES [16], ICFG-PEDES [5], and RSTPReid [41], and one fine-grained benchmark, UFine6926 [44], to evaluate our ICL. Compared with the coarse-grained benchmark, the fine-grained benchmark has richer texts with fine-grained information. For all datasets, we follow their official settings for data partitioning. More details are provided in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Evaluation Protocols. To measure performance, we utilize the widely accepted Rank-K (1,5,10) metrics to mea" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14394" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.098, + 0.089, + 0.904, + 0.33 + ], + "angle": 0, + "content": "
MethodsImage Enc.Text Enc.CUHK-PEDESICFG-PEDESRSTPReid
Rank-1Rank-5Rank-10mAPmINPRank-1Rank-5Rank-10mAPmINPRank-1Rank-5Rank-10mAPmINP
① VL-Backbones w/o ReID-domain pre-training
IVT [26]ViT-BaseBERT65.6985.9391.1560.66-56.0473.6080.22--46.7070.0078.80--
\\( LCR^2S \\) [36]RN50BERT67.3684.1989.6259.20-57.9376.0882.4038.21-54.9576.6584.7040.92-
CFine [37]CLIP-ViTBERT69.5785.9391.15--60.8376.5582.42--50.5572.5081.60--
RaSa [1]Swin-BBERT76.5190.2994.2569.38-65.2880.4085.1241.29-66.9086.5091.3552.31-
IRRA [12]CLIP-ViTCLIP-X.73.3889.9393.7166.1350.2463.4680.2585.8238.067.9360.2081.3088.2047.1725.28
TBPS [3]CLIP-ViTCLIP-X.73.5488.1992.3565.3849.2565.0580.3485.4739.837.8762.1081.9087.7548.0025.86
CFAM [44]CLIP-ViTCLIP-X.75.6090.5394.3667.27-65.3881.1786.3539.42-62.4583.5591.1049.50-
RDE [21]CLIP-ViTCLIP-X.75.9490.1494.1267.5651.4467.6882.4787.3640.067.8765.3583.9589.9050.8828.08
Our ICLCLIP-ViTCLIP-X.76.4190.4894.3368.0451.9968.1182.5987.5240.818.1867.7086.0591.7552.6229.36
Our ICL*CLIP-ViTCLIP-X.77.9190.2794.1469.1353.4069.0282.4587.3641.218.3070.5585.9591.6553.6830.13
② VL-Backbones with ReID-domain pre-training
\\( IRRA^b \\) [12]CLIP-ViTCLIP-X.74.0589.4893.6466.57-64.3780.7586.1238.85-61.9080.6089.3048.08-
APTM [38]Swin-BBERT76.5390.0494.1566.91-68.5182.9987.5641.22-67.5085.7091.4552.56-
\\( NAM^d \\) [28]CLIP-ViTCLIP-X.77.4790.8494.6769.4354.0866.7682.0287.1741.459.5367.1586.5591.9052.0028.46
Our ICLCLIP-ViTCLIP-X.78.1891.6394.8369.5853.4869.2283.4988.0642.349.0170.0086.6091.7054.1630.93
Our ICL*CLIP-ViTCLIP-X.79.0691.2694.7270.4454.7070.0583.3587.9142.709.1372.5586.6091.3055.1931.72
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.334, + 0.907, + 0.364 + ], + "angle": 0, + "content": "Table 1. Performance on the three coarse-grained benchmarks. The results with THI are marked with \\(\\star\\). Note that IRRA means using the pre-trained Backbones with MALS [38] and the results of NAM are reproduced by us." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.376, + 0.48, + 0.514 + ], + "angle": 0, + "content": "
MethodsRank-1Rank-5Rank-10mAPmINP
LGUR [25]70.6984.5789.9168.93-
SSAN [5]75.0988.6392.8473.14-
IRRA [12]85.0294.3196.7583.9177.30
RDE [21]87.6095.6597.4686.1079.54
CFAM(B/16) [44]85.5594.5197.0284.23-
CFAM(L/14) [44]88.5195.5897.4987.09-
1 Our ICL89.1796.1397.8887.4981.50
1 Our ICL★90.6795.9897.8688.2982.60
2 Our ICL91.0296.9898.1789.7684.70
2 Our ICL★91.7896.8398.1690.3385.62
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.517, + 0.483, + 0.545 + ], + "angle": 0, + "content": "Table 2. Performance comparison on the UFine6926 dataset. The results of IRRA and RDE are reproduced by us." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.555, + 0.484, + 0.603 + ], + "angle": 0, + "content": "sure the TIREID performance. Like [12, 21], we also report the mean Average Precision (mAP) and mean Inverse Negative Penalty (mINP) as auxiliary metrics." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.618, + 0.31, + 0.635 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.644, + 0.486, + 0.901 + ], + "angle": 0, + "content": "To achieve interactive TIReID, we choose RDE [21] and Qwen2-VL-7B-Instruct [29] as the investigated TIReID method and MLLM. For the sake of fairness, we do not modify any settings of RDE, including model architecture (CLIP-ViTB/16 [24]), hyperparameters, and training parameters. To be compatible with the fine-grained benchmark, following [44], the maximum length of the textual tokens of CLIP is set to 168 by interpolating the positional embedding layer with an initial learning rate of \\(5e - 5\\). As for fine-tuning MLLMs, we use the Llama-Factory framework [40] to conduct SFT with LoRA [11] for 2 epochs. The train batch size per device is set to 4, the gradient accumulation steps are set to 16, the initial learning rate is \\(5e - 5\\) and the hyperparameters \\(\\alpha\\) and \\(r\\) of LoRA are set to 16 and 8, respectively. During inference, the temperature is set to 0.01 to keep reproducibility. Note that all experiments can be completed on two GeForce RTX 24GB 3090 GPUs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.378, + 0.818, + 0.394 + ], + "angle": 0, + "content": "4.3. Comparison with State-of-the-Arts" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.402, + 0.907, + 0.538 + ], + "angle": 0, + "content": "In this section, to verify the superiority of our ICL, we compare our method with more than 15 baselines including recent advanced methods (e.g., RDE (CVPR'24) [21] and NAM (CVPR'24) [28]). Based on the backbone type, we divide the baselines into two groups (1 and 2) as shown Table 1, i.e., the baselines using VL-Backbones w/o and with ReID-domain pre-training. For the group 2, we use the pretrained weights released by [28] to initialize CLIP for fair comparison. The results are reported in Tables 1 and 2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.909, + 0.81 + ], + "angle": 0, + "content": "Results on coarse-grained datasets. Table 1 report the results evaluated on the coarse-grained datasets. We can see that our method can achieve competitive performance even without THI. By performing THI, the Rank-1 score of our method is greatly improved, e.g., in the group ①, the Rank-1 scores on the three datasets are improved by \\(1.50\\%\\), \\(0.91\\%\\), and \\(2.85\\%\\), respectively. In addition, mAP and mINP scores have also improved greatly, which indicates that the overall ranking has improved. In group ②, our method achieves the best scores on most metrics, especially Rank-1 reached \\(72.55\\%\\) on RSTPReid, which is sufficient to verify the superiority. However, we found that THI slightly degraded Rank-5 and Rank-10, which is because we only conducted 5 rounds of interaction. As long as the locating anchor image is incorrect, the errors will accumulate, which can be relieved by increasing the rounds of interaction. But in general, THI can significantly improve Rank-1 and the overall ranking (mAP and mINP)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Results on fine-grained dataset. Table 2 shows the results on the fine-grained dataset whose average text length is over 80. Such kind of fine-grained description often has a clear query intention. For better comparison, we provide the performance of IRRA and RDE on the UFine6926 dataset. From the results, our method can still achieve excellent per" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14395" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.1, + 0.089, + 0.904, + 0.272 + ], + "angle": 0, + "content": "
MethodsTraining SetsCUHK-PEDESICFG-PEDESRSTPReid
Rank-1Rank-5Rank-10mAPmINPRank-1Rank-5Rank-10mAPmINPRank-1Rank-5Rank-10mAPmINP
IRRA [12]CUHK-PEDES73.3889.9393.7166.1350.2442.4162.1169.6221.771.9553.2577.1585.3539.6316.60
ICFG-PEDES33.4856.2966.3331.5619.2063.4680.2585.8238.067.9345.3069.2578.8036.8218.38
RSTPReid32.8055.2665.8130.2917.6132.3049.6757.8020.543.8460.2081.3088.2047.1725.28
RDE [21]CUHK-PEDES75.9490.1494.1267.5651.4448.1866.3073.7025.002.3354.9077.5086.5041.2717.84
ICFG-PEDES38.1159.2468.4434.1620.4467.6882.4787.3640.067.8749.2572.1080.2038.4618.33
RSTPReid36.9458.2267.5833.6520.4242.1758.3265.4926.374.9465.3583.9589.9050.8828.08
Our ICLCUHK-PEDES76.4190.4894.3368.0451.9948.5766.6673.7525.302.4055.8079.6087.6542.0917.41
ICFG-PEDES42.8764.2073.4438.1923.5868.1182.5987.5240.818.1852.5075.0583.0041.8221.14
RSTPReid41.3161.8670.3136.7822.3745.9362.7068.8028.895.6367.7086.0591.7552.6229.36
Our ICL*CUHK-PEDES77.9190.2794.1469.1353.4052.8066.4973.4925.602.4461.3079.2587.4043.4218.01
ICFG-PEDES49.2964.3473.5540.8225.3869.0282.4587.3641.218.3060.1575.3083.1543.7222.04
RSTPReid47.3561.4570.3438.9123.6850.5261.5668.5729.265.7370.5585.9591.6553.6830.13
" + }, + { + "type": "table_caption", + "bbox": [ + 0.227, + 0.276, + 0.768, + 0.29 + ], + "angle": 0, + "content": "Table 3. Comparison of mutual generalization capabilities between coarse-grained datasets." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.313, + 0.483, + 0.63 + ], + "angle": 0, + "content": "
Source → TargetMethodsRank-1Rank-5Rank-10mAPmINP
CUHK. → UFine.IRRA [12]37.5154.9264.2940.7634.33
RDE [21]40.3757.4966.0542.6835.78
Our ICL46.4063.5572.0848.6841.56
Our ICL☆57.7664.1372.8153.9745.64
ICFG. → UFine.IRRA [12]15.0226.7933.9017.1012.75
RDE [21]17.8631.0138.5619.8214.74
Our ICL27.9544.2052.2029.8523.20
Our ICL☆36.8144.6552.7334.1226.61
RSTP. → UFine.IRRA [12]13.2125.6733.9315.6011.09
RDE [21]14.0025.2332.6416.2211.90
Our ICL23.8938.3046.7025.5419.20
Our ICL☆31.2338.5647.0228.9021.80
UFine. → CUHKIRRA [12]37.7460.1270.1335.9423.21
RDE [21]39.4161.1470.1136.4923.32
Our ICL49.0470.2778.6444.5429.58
Our ICL☆56.8770.1978.5347.3131.20
UFine. → ICFG.IRRA [12]34.5255.4164.4417.961.95
RDE [21]40.3760.1468.4120.542.19
Our ICL43.1062.9270.7322.732.56
Our ICL☆47.8362.6770.4823.162.62
UFine. → RSTP.IRRA [12]37.6563.7073.0029.0011.80
RDE [21]39.9063.5074.7529.9212.43
Our ICL48.8572.6581.8036.9116.39
Our ICL☆55.3572.4081.5038.6417.23
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.632, + 0.483, + 0.661 + ], + "angle": 0, + "content": "Table 4. Generalization capabilities between coarse-grained and fine-grained datasets. The best scores in each task are in bold." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.687, + 0.483, + 0.718 + ], + "angle": 0, + "content": "formance, with Rank-1 exceeding \\(91\\%\\). This shows that interaction is also applicable to the fine-grained scenario." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.727, + 0.292, + 0.743 + ], + "angle": 0, + "content": "4.4. Generalization Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To evaluate the generalization, Tables 3 and 4 report the cross-domain performance on four datasets, including the coarse-to-fine, coarse-to-fine, and fine-to-coarse generalization experiments. From the results, our method achieves better generalization than RDE even without THI thanks to the training enhancement of RDA. When THI is performed, the cross-domain performance is dramatically improved, for example, from CUHK-PEDES to RSTPReid, THI brings an improvement of more than \\(4\\%\\) on Rank-1. This shows that THI is a potential solution to generaliza" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.316, + 0.907, + 0.437 + ], + "angle": 0, + "content": "tion challenges in the future. From the generalization experiments between fine-grained and coarse-grained datasets shown in Table 4, ICL can also achieve the best cross-domain performance, e.g., compared with the best baseline RDE, from UFine6926 domain to CUHK-PEDES domain, our method improves Rank-1 and mAP by \\(17.49\\%\\) and \\(10.86\\%\\), respectively, which further verifies the cross-domain generalization of our method." + }, + { + "type": "table", + "bbox": [ + 0.52, + 0.449, + 0.905, + 0.548 + ], + "angle": 0, + "content": "
MethodsTHICUHK-PEDESICFG-PEDESRSTPReidΔAvg
Rank-1mAPRank-1mAPRank-1mAP
CLIP [24]X71.6463.9260.1134.5256.5544.52+2.57
✓73.7765.6663.5734.9561.4546.25
IRRA [12]X73.3866.1363.4638.0660.2047.17+1.86
✓76.0667.4265.2638.5863.7548.47
RDE [21]X75.9467.5667.6840.0665.3550.88+1.41
✓77.4768.6268.7240.6368.4552.01
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.513, + 0.55, + 0.905, + 0.577 + ], + "angle": 0, + "content": "Table 5. Transferability results on three coarse-grained benchmarks. \\( \\Delta \\) Avg represents the average improvement." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.603, + 0.715, + 0.62 + ], + "angle": 0, + "content": "4.5. Transferability Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.627, + 0.907, + 0.839 + ], + "angle": 0, + "content": "Our interactive module is separate and independent from the training of existing TIReID methods, so it can be plug-and-play with existing methods to improve ReID performance. To verify the transferability of our THI, we conduct experiments on multiple baselines, and the results are shown in Table 5. Except for the CLIP results we reproduced, all other experiments used public pre-trained model weights. From the results of Table 5, the interactive strategy application can significantly improve Rank-1 and mAP, which shows that the external guidance by interactions via MLLMs can further clarify the text-image alignments and improve the overall ranking. The average improvements on the three datasets are \\(2.57\\%\\), \\(1.86\\%\\), and \\(1.41\\%\\), respectively, which seriously proves the transferability of our THI." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.787, + 0.865 + ], + "angle": 0, + "content": "4.6. Ablation and Parameter Study" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.903 + ], + "angle": 0, + "content": "To explore the effects of each proposed component, i.e., THI and RDA, we first conduct the ablation study on three" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14396" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.319 + ], + "angle": 0, + "content": "coarse-grained datasets as reported in Table 6. From the results, each component can bring performance gains on Rank-1 and mAP, which verifies the reliability and rationality of the method. Especially the introduction of THI has greatly improved the Rank-1 accuracy from \\(68.95\\%\\) to \\(70.55\\%\\) on the RSTPreid dataset. In addition, our RDA also brings considerable performance gains, especially on the RSTPreid dataset. Also, we conduct the parameter analysis on the CHUK-PEDES dataset on two free hyperparameters, i.e., the similarity threshold \\(\\xi\\) and the balance factor \\(\\lambda\\). The former filters unnecessary retrievals to improve the interaction efficiency, while the latter controls the contribution of refined texts. Based on Figure 5, in our experiments, we set \\(\\xi\\) in the range of \\(0.5 \\sim 0.6\\) and \\(\\lambda\\) to 0.8, thus mitigating the risk of introducing noisy external knowledge." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.328, + 0.48, + 0.407 + ], + "angle": 0, + "content": "
No.THIRDALoRACUHK-PEDESICFG-PEDESRSTPReid
Rank-1mAPRank-1mAPRank-1mAP
#1✓✓✓77.9169.3269.0241.2170.5553.68
#2✓✓×76.3868.5967.9241.1369.0053.11
#3×✓×76.4168.0468.1140.8167.7052.62
#4×××75.9467.5667.6840.0665.3550.88
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.409, + 0.483, + 0.437 + ], + "angle": 0, + "content": "Table 6. Ablation studies on CHUK-PEDES, ICFG-PEDES, and RSTPReid datasets. The best scores are in bold." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.45, + 0.288, + 0.548 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.551, + 0.268, + 0.563 + ], + "angle": 0, + "content": "(a) The similarity threshold \\(\\xi\\)" + }, + { + "type": "image", + "bbox": [ + 0.287, + 0.45, + 0.484, + 0.548 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.321, + 0.551, + 0.45, + 0.562 + ], + "angle": 0, + "content": "(b) The balance factor \\(\\lambda\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.566, + 0.46, + 0.58 + ], + "angle": 0, + "content": "Figure 5. Variation of performance with different \\(\\xi\\) and \\(\\lambda\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.591, + 0.262, + 0.607 + ], + "angle": 0, + "content": "4.7. Interactive Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.484, + 0.901 + ], + "angle": 0, + "content": "This section further explores the interactive module, i.e. THI. We report the performance (mAP) changes after multiple rounds of interactions in Figure 6. The mAP score can reflect the overall retrieval quality. From the results, the performance improvement is obvious in the first few rounds \\((\\leq 2)\\) since the items that meet the query semantics are mostly concentrated at the top of the ranking. However, after \\(>2\\) rounds, the performance gain is gradually not obvious as the number of queries requiring interaction decreases. But generally, as the number of rounds increases, the overall performance gradually improves. In all our experiments, we performed 5 rounds of interaction. In addition, we visualize the top-10 retrieved results before and after applying THI in Figure 7. Due to the inherent limitations of the intra-model knowledge, it is difficult to obtain satisfactory results without the help of THI. In contrast, our THI can dynamically enrich queries by interacting with MLLMs to achieve a more reliable ranking. More example results are given in the supplementary material." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.09, + 0.648, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.531, + 0.184, + 0.632, + 0.195 + ], + "angle": 0, + "content": "(a) CUHK-PEDES" + }, + { + "type": "image", + "bbox": [ + 0.648, + 0.089, + 0.775, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.664, + 0.184, + 0.758, + 0.195 + ], + "angle": 0, + "content": "(b) ICFG-PEDES" + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.09, + 0.905, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.803, + 0.184, + 0.878, + 0.195 + ], + "angle": 0, + "content": "(c) RSTPReid" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.199, + 0.905, + 0.227 + ], + "angle": 0, + "content": "Figure 6. Performance (mAP) versus rounds on three datasets. Round 0 indicates the setting without using THI." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.247, + 0.661, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.661, + 0.247, + 0.905, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.326, + 0.661, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.661, + 0.325, + 0.904, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.4, + 0.905, + 0.428 + ], + "angle": 0, + "content": "Figure 7. Top-10 retrieved results on CUHK-PEDES dataset between ICL (the first row) and ICL with THI (the second row)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.46, + 0.634, + 0.475 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.488, + 0.907, + 0.685 + ], + "angle": 0, + "content": "In this paper, we explore interactive text-to-image person re-identification, which aims to improve the alignment between dynamic queries and challenging candidate images by leveraging external guidance from MLLMs. To achieve this, we develop an Interactive Cross-modal Learning (ICL) framework to alleviate the inherent challenges of offline models and training data by, including a plug-and-play Test-time Human-centered Interaction (THI) module and Reorganization Data Augmentation (RDA). Extensive experiments and analysis show that our framework can effectively transfer external knowledge in MLLMs into offline models for guiding re-identification, showing excellent performance and generalization." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.707, + 0.673, + 0.723 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "This work was supported in part by the National Key R&D Program of China under Grant 2024YFB4710604; in part by NSFC under Grants 62472295, 62372315, 62176171, and U21B2040; in part by Sichuan Science and Technology Planning Projects (2024YFHZ0089, 2024NSFTD0049, 2024YFHZ0144, 2024NSFTD0047, 2024NSFTD0038); in part by System of Systems and Artificial Intelligence Laboratory pioneer fund grant; in part by the Fundamental Research Funds for the Central Universities under Grant CJ202403 and CJ202303; in part by Chengdu Science and Technology Project under Grant 2023-XT00-00004-GX." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14397" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.197 + ], + "angle": 0, + "content": "[1] Yang Bai, Min Cao, Daming Gao, Ziqiang Cao, Chen Chen, Zhenfeng Fan, Liqiang Nie, and Min Zhang. Rasa: relation and sensitivity aware representation learning for text-based person search. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, pages 555–563, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.2, + 0.484, + 0.268 + ], + "angle": 0, + "content": "[2] Maryam Bukhari, Sadaf Yasmin, Sheneela Naz, Muazzam Maqsood, Jehyeok Rew, and Seungmin Rho. Language and vision based person re-identification for surveillance systems using deep learning with lip layers. Image and Vision Computing, 132:104658, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.27, + 0.483, + 0.324 + ], + "angle": 0, + "content": "[3] Min Cao, Yang Bai, Ziyin Zeng, Mang Ye, and Min Zhang. An empirical study of clip for text-based person search. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 465-473, 2024. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.326, + 0.483, + 0.38 + ], + "angle": 0, + "content": "[4] Yuhao Chen, Guoqing Zhang, Yujiang Lu, Zhenxing Wang, and Yuhui Zheng. Tipcb: A simple but effective part-based convolutional baseline for text-based person search. Neurocomputing, 494:171-181, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.382, + 0.483, + 0.436 + ], + "angle": 0, + "content": "[5] Zefeng Ding, Changxing Ding, Zhiyin Shao, and Dacheng Tao. Semantically self-aligned network for text-to-image part-aware person re-identification. arXiv preprint arXiv:2107.12666, 2021. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.438, + 0.483, + 0.478 + ], + "angle": 0, + "content": "[6] Hiren Galiyawala and Mehul S Raval. Person retrieval in surveillance using textual query: a review. Multimedia Tools and Applications, 80(18):27343-27383, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.481, + 0.482, + 0.521 + ], + "angle": 0, + "content": "[7] Daming Gao, Yang Bai, Min Cao, Hao Dou, Mang Ye, and Min Zhang. Semi-supervised text-based person search. arXiv preprint arXiv:2404.18106, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.524, + 0.483, + 0.577 + ], + "angle": 0, + "content": "[8] Donghoon Han, Eunhwan Park, Gisang Lee, Adam Lee, and Nojun Kwak. Merlin: Multimodal embedding refinement via llm-based iterative navigation for text-video retrieval-erank pipeline. arXiv preprint arXiv:2407.12508, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.579, + 0.483, + 0.632 + ], + "angle": 0, + "content": "[9] Chen He, Shenshen Li, Zheng Wang, Hua Chen, Fumin Shen, and Xing Xu. Chatting with interactive memory for text-based person retrieval. Multimedia Systems, 31(1):31, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.635, + 0.483, + 0.689 + ], + "angle": 0, + "content": "[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.691, + 0.483, + 0.746 + ], + "angle": 0, + "content": "[11] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.748, + 0.483, + 0.815 + ], + "angle": 0, + "content": "[12] Ding Jiang and Mang Ye. Cross-modal implicit relation reasoning and aligning for text-to-image person retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2787-2797, 2023. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.483, + 0.87 + ], + "angle": 0, + "content": "[13] Matan Levy, Rami Ben-Ari, Nir Darshan, and Dani Lischinski. Chatting makes perfect: Chat-based image retrieval. Advances in Neural Information Processing Systems, 36, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[14] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.094, + 0.905, + 0.133 + ], + "angle": 0, + "content": "vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[15] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.207, + 0.905, + 0.274 + ], + "angle": 0, + "content": "[16] Shuang Li, Tong Xiao, Hongsheng Li, Bolei Zhou, Dayu Yue, and Xiaogang Wang. Person search with natural language description. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1970-1979, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.278, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[17] Yunfan Li, Peng Hu, Dezhong Peng, Jiancheng Lv, Jianping Fan, and Xi Peng. Image clustering with external guidance. In *Forty-first International Conference on Machine Learning*. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.904, + 0.362 + ], + "angle": 0, + "content": "[18] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.364, + 0.905, + 0.432 + ], + "angle": 0, + "content": "[19] Yang Qin, Dezhong Peng, Xi Peng, Xu Wang, and Peng Hu. Deep evidential learning with noisy correspondence for cross-modal retrieval. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4948-4956, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.435, + 0.905, + 0.503 + ], + "angle": 0, + "content": "[20] Yang Qin, Yuan Sun, Dezhong Peng, Joey Tianyi Zhou, Xi Peng, and Peng Hu. Cross-modal active complementary learning with self-refining correspondence. Advances in neural information processing systems, 36:24829-24840, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.506, + 0.905, + 0.585 + ], + "angle": 0, + "content": "[21] Yang Qin, Yingke Chen, Dezhong Peng, Xi Peng, Joey Tianyi Zhou, and Peng Hu. Noisy-correspondence learning for text-to-image person re-identification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27197-27206, 2024. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.59, + 0.905, + 0.631 + ], + "angle": 0, + "content": "[22] Yalan Qin, Chuan Qin, Xinpeng Zhang, and Guorui Feng. Dual consensus anchor learning for fast multi-view clustering. IEEE Transactions on Image Processing, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.633, + 0.905, + 0.674 + ], + "angle": 0, + "content": "[23] Yalan Qin, Nan Pu, Nicu Sebe, and Guorui Feng. Latent space learning based ensemble clustering. IEEE Transactions on Image Processing, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.676, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[24] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.761, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[25] Zhiyin Shao, Xinyu Zhang, Meng Fang, Zhifeng Lin, Jian Wang, and Changxing Ding. Learning granularity-unified representations for text-to-image person re-identification. In Proceedings of the 30th ACM International Conference on Multimedia, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[26] Xiujun Shu, Wei Wen, Haoqian Wu, Keyu Chen, Yiran Song, Ruizhi Qiao, Bo Ren, and Xiao Wang. See finer, see more: Implicit modality alignment for text-based person retrieval. In European Conference on Computer Vision, pages 624-641. Springer, 2022. 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.094, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14398" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[27] Yuan Sun, Yang Qin, Yongxiang Li, Dezhong Peng, Xi Peng, and Peng Hu. Robust multi-view clustering with noisy correspondence. IEEE Transactions on Knowledge and Data Engineering, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.218 + ], + "angle": 0, + "content": "[28] Wentan Tan, Changxing Ding, Jiayu Jiang, Fei Wang, Yibing Zhan, and Dapeng Tao. Harnessing the power of mllms for transferable text-to-image person reid. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17127-17137, 2024. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.219, + 0.482, + 0.315 + ], + "angle": 0, + "content": "[29] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.482, + 0.383 + ], + "angle": 0, + "content": "[30] Yansong Wang, Yundong Sun, Yansheng Fu, Dongjie Zhu, and Zhaoshuo Tian. Spectrum-bert: pre-training of deep bidirectional transformers for spectral classification of Chinese liquors. IEEE Transactions on Instrumentation and Measurement, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.385, + 0.482, + 0.44 + ], + "angle": 0, + "content": "[31] Yabing Wang, Le Wang, Qiang Zhou, Zhibin Wang, Hao Li, Gang Hua, and Wei Tang. Multimodal llm enhanced cross-lingual cross-modal retrieval. arXiv preprint arXiv:2409.19961, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.482, + 0.496 + ], + "angle": 0, + "content": "[32] Yuhao Wang, Yang Liu, Aihua Zheng, and Pingping Zhang. Demo: Decoupled feature-based mixture of experts for multi-modal object re-identification. In Proceedings of the AAAI Conference on Artificial Intelligence, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.498, + 0.482, + 0.579 + ], + "angle": 0, + "content": "[33] Zhe Wang, Zhiyuan Fang, Jun Wang, and Yezhou Yang. Vistaa: Visual-textual attributes alignment in person search by natural language. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XII 16, pages 402-420. Springer, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.482, + 0.649 + ], + "angle": 0, + "content": "[34] Zijie Wang, Aichun Zhu, Jingyi Xue, Xili Wan, Chao Liu, Tian Wang, and Yifeng Li. Caibc: Capturing all-round information beyond color for text-based person retrieval. In Proceedings of the 30th ACM International Conference on Multimedia, pages 5314-5322, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.651, + 0.482, + 0.718 + ], + "angle": 0, + "content": "[35] Yushuang Wu, Zizheng Yan, Xiaoguang Han, Guanbin Li, Changqing Zou, and Shuguang Cui. Lipscore: language-guided person search via color reasoning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1624-1633, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.482, + 0.788 + ], + "angle": 0, + "content": "[36] Shuanglin Yan, Neng Dong, Jun Liu, Liyan Zhang, and Jinhui Tang. Learning comprehensive representations with richer self for text-to-image person re-identification. In Proceedings of the 31st ACM international conference on multimedia, pages 6202–6211, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.482, + 0.831 + ], + "angle": 0, + "content": "[37] Shuanglin Yan, Neng Dong, Liyan Zhang, and Jinhui Tang. Clip-driven fine-grained text-image person re-identification. IEEE Transactions on Image Processing, 2023. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[38] Shuyu Yang, Yinan Zhou, Zhedong Zheng, Yaxiong Wang, Li Zhu, and Yujiao Wu. Towards unified text-based person retrieval: A large-scale multi-attribute and language search benchmark. In Proceedings of the 31st ACM International Conference on Multimedia, pages 4492–4501, 2023. 1, 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.162 + ], + "angle": 0, + "content": "[39] Pingping Zhang, Yuhao Wang, Yang Liu, Zhengzheng Tu, and Huchuan Lu. Magic tokens: Select diverse tokens for multi-modal object re-identification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17117-17126, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[40] Yaowei Zheng, Richong Zhang, Junhao Zhang, Yanhan Ye, Zheyan Luo, Zhangchi Feng, and Yongqiang Ma. Llamafactory: Unified efficient fine-tuning of \\(100+\\) language models. arXiv preprint arXiv:2403.13372, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[41] Aichun Zhu, Zijie Wang, Yifeng Li, Xili Wan, Jing Jin, Tian Wang, Fangqiang Hu, and Gang Hua. Dssl: Deep surroundings-person separation learning for text-based person retrieval. In Proceedings of the 29th ACM International Conference on Multimedia, pages 209-217, 2021. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.905, + 0.331 + ], + "angle": 0, + "content": "[42] Jialong Zuo, Changqian Yu, Nong Sang, and Changxin Gao. Plip: Language-image pre-training for person representation learning, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.333, + 0.905, + 0.388 + ], + "angle": 0, + "content": "[43] Jialong Zuo, Ying Nie, Hanyu Zhou, Huaxin Zhang, Haoyu Wang, Tianyu Guo, Nong Sang, and Changxin Gao. Cross-video identity correlating for person re-identification pretraining. arXiv preprint arXiv:2409.18569, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.905, + 0.471 + ], + "angle": 0, + "content": "[44] Jialong Zuo, Hanyu Zhou, Ying Nie, Feng Zhang, Tianyu Guo, Nong Sang, Yunhe Wang, and Changxin Gao. Ufinebench: Towards text-based person retrieval with ultra-fine granularity. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22010-22019, 2024. 1, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.471 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "14399" + } + ] +] \ No newline at end of file diff --git a/2025/Human-centered Interactive Learning via MLLMs for Text-to-Image Person Re-identification/5cc75be5-e122-4a4f-ba40-7450ab436d6a_origin.pdf b/2025/Human-centered Interactive Learning via MLLMs for Text-to-Image Person Re-identification/5cc75be5-e122-4a4f-ba40-7450ab436d6a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..25984f373534f60f8f1fb840f7fcd40e0ff0f19b --- /dev/null +++ b/2025/Human-centered Interactive Learning via MLLMs for Text-to-Image Person Re-identification/5cc75be5-e122-4a4f-ba40-7450ab436d6a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1094b6214e56cf797007dce56bbbcc8e5b4c0e4026b6a7b312edca650af8252b +size 1649034 diff --git a/2025/Human-centered Interactive Learning via MLLMs for Text-to-Image Person Re-identification/full.md b/2025/Human-centered Interactive Learning via MLLMs for Text-to-Image Person Re-identification/full.md new file mode 100644 index 0000000000000000000000000000000000000000..dce1e826b3c356586282d77e4d820f54092c45fb --- /dev/null +++ b/2025/Human-centered Interactive Learning via MLLMs for Text-to-Image Person Re-identification/full.md @@ -0,0 +1,309 @@ +# Human-centered Interactive Learning via MLLMs for Text-to-Image Person Re-identification + +Yang Qin $^{1}$ Chao Chen $^{2}$ Zhihang Fu $^{2}$ Dezhong Peng $^{1,4,5}$ Xi Peng $^{1,3}$ Peng Hu $^{1*}$ + +1College of Computer Science, Sichuan University 2Independent Researcher + +$^{3}$ National Key Laboratory of Fundamental Algorithms and Models for Engineering Simulation, Sichuan University + +Sichuan National Innovation New Vision UHD Video Technology Co., Ltd Tianfu Jincheng Laboratory + +# Abstract + +Despite remarkable advancements in text-to-image person re-identification (TIREID) facilitated by the breakthrough of cross-modal embedding models, existing methods often struggle to distinguish challenging candidate images due to intrinsic limitations, such as network architecture and data quality. To address these issues, we propose an Interactive Cross-modal Learning framework (ICL), which leverages human-centered interaction to enhance the discriminability of text queries through external multimodal knowledge. To achieve this, we propose a plug-and-play Test-time Humane-centered Interaction (THI) module, which performs visual question answering focused on human characteristics, facilitating multi-round interactions with a multimodal large language model (MLLM) to align query intent with latent target images. Specifically, THI refines user queries based on the MLLM responses to reduce the gap to the best-matching images, thereby boosting ranking accuracy. Additionally, to address the limitation of low-quality training texts, we introduce a novel Reorganization Data Augmentation (RDA) strategy based on information enrichment and diversity enhancement to enhance query discriminability by enriching, decomposing, and reorganizing person descriptions. Extensive experiments on four TIREID benchmarks, i.e., CUHK-PEDES, ICFG-PEDES RSTPReid, RSTPReid, and UFine6926, demonstrate that our method achieves remarkable performance with substantial improvement. Code is available at https://github.com/QinYang79/ICL. + +# 1. Introduction + +Recently, text-to-image person re-identification (TIREID) [3-5, 7, 33, 34, 41] has made great progress in aligning text descriptions and person images, enabling high-accuracy person search and identification. Different from traditional + +![](images/fba52da64d005a0539e6ec922a99984925182eb012c93d057d25364f0a94c25b.jpg) +(a) Existing methods + +![](images/0f3b0b8518086d3904eb882445c010dfce5f4e8b3ce5da480db96badb7397d1b.jpg) +(b) Interactive TIREID +Figure 1. The illustration of our motivation. When performing text-based person re-identification, (a) existing methods commonly exploit cross-modal models to calculate similarity (such as IRRA [12] and RDE [21]) and then obtain the candidate person images by ranking. However, due to the intrinsic limitations of models and training data, they may not be able to distinguish challenging candidate images well enough to obtain satisfactory results. (b) Our motivation is to interact with the system for external guidance like a human and gradually refine the user's query from the candidate items shown in (a) by multiple rounds of question-answering, ultimately improving the overall ranking. + +image-to-image re-identification [32, 39, 42, 43], TIREID, as a rising task in the multimodal community [19, 20, 22, 23, 27], retrieves person images using user-customized text queries, which is more practical in many scenarios where an image query is unavailable, such as in video surveillance systems [2] or crowd management [6]. A primary challenge in TIREID is learning to associate text queries with the corresponding person images at a fine-grained level, bridging the modality gap for accurate similarity measurement. + +To overcome this challenge, recent efforts have developed various strategies to improve cross-modal associations, including local attribute modeling [37, 44], loss function designing [12, 21], ReID-domain pre-training [28, 38], etc. Although these methods achieve promising perfor + +mance, they are limited by the inherent defect of the offline models and training data, making it hard to handle dynamic user query inputs and leading to poor generalization. In practice, however, different users tend to input concise, vague, and diverse text queries based on their memories to the target person, thereby raising challenges for the offline models to understand fine-grained human characteristics from the queries. For example, as illustrated in Figure 1 (a), the user text query "A woman walking visible from the back is wearing a white shirt, and black pants." fails to capture some important contextual details like handheld objects or background, resulting in inaccurate candidate images. In brief, it is hard for offline models to handle dynamic and challenging queries solely relying on their learned internal knowledge. To break through this bottleneck, we could seek to align queries with image candidates with the help of external knowledge, thereby reranking the retrieved images and boosting the identification accuracy. + +Recent studies [8, 13, 17] have explored integrating external knowledge from pre-trained models or (multimodal) large language models to guide various downstream tasks. For instance, Li et al. [17] leveraged multimodal knowledge of pre-trained vision-language models [24] to enhance unimodal image representations for external-guided clustering. However, this approach cannot achieve post-hoc improvement for offline models, leading to infeasibility in practice due to the prohibitive cost of fine-tuning or even retraining. In cross-modal retrieval, Han et al. [8] exploited the excellent comprehension and generation capabilities of multimodal large language models (MLLMs) to design an interactive re-ranking pipeline for text-to-video retrieval. However, Han et al. [8] introduces label leakage by using the ground truth video as the video in mind for the MLLM answerer agent, which is not applicable in actual testing. + +In this paper, we present a novel Interactive Cross-modal Learning framework (ICL) that exploits the multimodal knowledge implicit in MLLMs to enhance the alignment between text queries and target person images. Specifically, we first propose a plug-and-play Test-time Human-centered Interaction (THI) module that refines text queries through interactions with an MLLM, enhancing the post-hoc ability of trained models to distinguish challenging candidates. THI identifies the latent target images through multi-round interactions with MLLMs, asking human-centered questions to the MLLM for fine-grained answers about the images, which refine the query texts to strengthen the alignment with the images, ultimately address the inherent limitations of the input queries and enhancing ranking accuracy. To address the intrinsic limitations in the trained models, we propose a Reorganization Data Augmentation strategy (RDA) to enrich and diversify pedestrian texts for training enhancement, thereby transferring external knowledge in MLLMs into the model. To enrich the texts, RDA applies + +visual question answering via the MLLM on each image-text pair to supplement fine-grained person characteristics. Moreover, to enhance the diversity of texts, RDA presents a decomposition-reorganization strategy to decompose person descriptions into attribute-specific sub-sentences (e.g., clothes, pants, shoes, etc.), and rewrite them with MLLM into multiple sentences with the same meaning. Subsequently, the rewritten sentences are randomly reordered and recombined to generate varied augmented texts, thus boosting data diversity and enhancing the generalization of models. Our main contributions are as follows: + +- We introduce human-centered interaction to TIREID, proposing a novel MLLM-driven Interactive Cross-modal Learning framework (ICL), which leverages external knowledge to overcome the inherent limitations of existing offline methods in handling dynamic queries. +- A plug-and-play Test-time Human-centered Interaction module (THI) is presented to align query intent with latent target images through multi-round interactions with MLLMs, improving the ranking quality. +- An effective Reorganization Data Augmentation strategy (RDA), applying MLLMs for text decomposition and recombination, is developed to generate discriminative and diverse training texts, improving cross-modal learning. +- Extensive experiments on four text-to-image person re-identification benchmarks verify the effectiveness and superiority of our method, which achieves promising performance and superior generalization. + +# 2. Related Work + +# 2.1. Text-to-Image Person Re-identification + +As a challenging topic in multimodal learning, text-to-image person re-identification (TIREID) aims to search the image of the target person with a given natural language query. Existing methods [1, 12, 21, 28, 41] can be roughly divided into three categories according to the used backbone type: unimodal backbones, general multimodal pre-trained backbones, and ReID-domain multimodal pre-trained backbones. The early TIREID methods [33, 35] commonly use modality-specific unimodal backbones to encode text or images for cross-modal alignment, e.g., ResNet [10], BERT [30], etc. Recently, benefiting from the rapid development of visual-language pre-training models [14, 24], more and more researchers [3, 12, 21] have begun to use general pre-trained models as the backbone networks for solving TIREID, hoping that the pre-trained alignment knowledge can improve modality representation and cross-modal alignment. However, the performance improvement is still limited due to the domain gap between the general pre-training data and TIREID tasks. To this end, some attempts [28, 38] exploit image-caption models or MLLMs to annotate images, thus obtaining a large number of image + +![](images/5486563401a74a428c6b4e71d4a148c3038f7d2f3506da315bc0cd95a87f5da7.jpg) +Figure 2. The illustration of our Test-time Human-centered Interaction (THI) module. THI includes $K$ rounds of interactions to align query intention with the latent target image by external guidance, where in each round, we perform human-centered visual question answering around fine-grained person attributes to enhance the semantic consistency between the query and the intended person image, and then improve the final ReID performance on the large-scale evaluation through efficient re-ranking. Besides, we perform supervised fine-tuning via LoRA to inspire the discriminate ability of MLLM for ReID domain images and better align queries with latent target images. + +text pairs for ReID-domain pre-training. Although promising performance has been achieved, they still cannot get rid of the inherent defect of offline models during handling dynamic queries. In this paper, we propose a novel interactive cross-modal learning framework based on MLLMs, which exploits dynamic interactions for external guidance to improve the generalization ability. + +# 2.2. Visual Interactive Learning + +With the booming development of multimodal large language models (MLLMs) [15, 18, 29], exploiting visual interaction to improve downstream tasks has attracted increasing attention from researchers. We collectively refer to these as visual interaction learning and divide them into offline interaction [28, 31] and online interaction [8, 9, 13] groups by engaging in specific task periods. The former aims to use interactions to improve the training/pre-training data, such as information richness and diversity. It is usually static and separated from users, which is the mainstream of existing TIREID methods, e.g., Tan et al. [28] apply MLLMs to conduct interactions with multiple prompt templates for better diversity. Unlike them, online interactions are designed to be able to use real-time feedback to improve training or testing. For example, Levy et al. [13] propose to gradually clarify user intention through dialogue to improve image retrieval performance. Likewise, Han et al. [8] develop a rerank pipeline based on LLM-based iterative navigation. However, these methods cannot be directly used for TIREID effectively due to the differences in tasks and data domains. In this paper, we develop a test-time module based on MLLMs to conduct interactions for external guidance, overcoming the limitations of the intra- + +model knowledge in TIReID offline models. + +# 3. Methodology + +In this section, we introduce an Interactive Cross-modal Learning framework (ICL), which consists of two core comments to address the inherent challenges in offline models and training data, i.e., Test-time Human-centered Interaction module (THI) and Reorganization Data Augmentation (RDA). In Section 3.1, we provide the necessary definitions to facilitate the study. Then, we outline the details of our THI and RDA in Sections 3.2 and 3.3, respectively. + +# 3.1. Problem and Symbol Statement + +Suppose that we have the query text $q \in \mathcal{Q}$ , a pedestrian image $v \in \mathcal{V}$ , where $\mathcal{Q}$ and $\mathcal{V}$ are the text query set and image gallery. The purpose of TIReID is to utilize the text query to retrieve the ideal images from the gallery, thus achieving person searching by identities of retrieved images. To achieve this, existing methods usually train an offline cross-modal model $f_{\mathrm{cross}} = \{f,g\}$ to measure the similarities between text queries and pedestrian images. For a text $q$ and an image $v$ , the similarity to measure the matching degree can be represented as $S_{q,v} \equiv \operatorname{Sim}(f(q), g(v))$ , where $\operatorname{Sim}(*)$ is the similarity function, $f$ and $g$ are the text and image encoders, respectively. Then, we can utilize $q$ to search relevant candidate images $\hat{v}$ from $\mathcal{V}$ as follows: + +$$ +\hat {\mathcal {V}} (q) = \left\{\hat {v} _ {k} \right\} _ {k = 1} ^ {K} = \operatorname {T o p} - K _ {v \in \mathcal {V}} \left(S _ {q, v}\right), \tag {1} +$$ + +where $\hat{\mathcal{V}}(q)$ is the candidate set for $q$ and $K$ is the number of candidate images. In addition, our ICL also involves MLLMs to conduct interactions and training augmentation, + +which we denote as $\mathcal{M}$ , and the prompt template function for $\mathcal{M}$ is denoted by the symbol $\mathcal{T}$ . Based on the above definitions, we will elaborate on ICL in the following sections. + +# 3.2. Test-time Humane-center Interaction + +Anchor Localization. Due to the model bottleneck, it is hard for existing methods to distinguish challenging candidate images relying solely on internal knowledge. Thus, we exploit the fine-grained image understanding capability of MLLMs to conduct external guidance. We first propose asking questions such as "Can this text accurately describe the image?" to let MLLMs explicitly tell us the latent target (anchor) image. As shown in Figure 2, given a text query $q$ , we first get the corresponding candidate set $\hat{\mathcal{V}}(q)$ and then perform multiple $(K)$ rounds of interactions according to the ranking until the ideal image is determined. For the $k$ -th round, the answer for interaction can be represented as: + +$$ +a _ {\hat {v} _ {k}} ^ {q} = \mathcal {M} \left(\mathcal {T} _ {\text {l o c}} (q, \hat {v} _ {k})\right), \tag {2} +$$ + +where $k \in \{1, \dots, K\}$ , $a_{\hat{v}_k}^q$ is the answer of 'Yes' or 'No', $\hat{v}_k$ is the Top- $k$ candidate image for query $q$ , and $T_{\mathrm{loc}}$ is the prompt template function for anchor localization. However, since there is a domain gap between ReID images and the generic images used for pre-training or instruction-tuning MLLMs, the answers of MLLMs are often unreliable. To solve this, we exploit LoRA [11] to perform supervised fine-tuning (SFT) on MLLMs to inspire the fine-grained ability to identify the person image and the SFT loss is: + +$$ +\mathcal {L} _ {l o r a} = - \sum_ {(x, y) \in \mathcal {Z}} \sum_ {t = 1} ^ {| y |} \log \left(p _ {\mathcal {M}} \left(y _ {t} \mid x, y < t\right)\right), \tag {3} +$$ + +where $x$ is the input Prompt and $y$ is the output Response, $\mathcal{Z} = \{\mathcal{Z}^{+},\mathcal{Z}^{-}\}$ is the SFT dataset. For $\mathcal{Z}^{+}$ , we select part of training texts $(\{q_i\}_{i = 1}^{N_l})$ and the corresponding ground-truth images $(\{v_{i}^{+}\}_{i = 1}^{N_{l}})$ from the training set to construct the input Pormpts, i.e., $\{\mathcal{T}_{\mathrm{loc}}(q_i,v_i^+)\}_{i = 1}^{N_l}$ expecting MLLMs to output the Response of 'Yes'. Similarly, for $\mathcal{Z}^{-}$ , we also use the text queries and the corresponding negative images from the training set to construct the input Pormpts, expecting MLLMs to output the Response of 'No'. To make $\mathcal{Z}^{-}$ more discriminative, we use a TIREID pre-trained cross-modal model to obtain the Top-10 image with different person ID as input negative sample image for each text query by similarity ranking. + +Human-centered VQA. For each round, once we confirm the response of 'Yes', we will ask the anchor image $(\bar{v})$ a series of questions to learn more details about pedestrian characteristics. These details can help us alleviate the discrepancy between the user query and the information within the anchor image, thus improving overall ranking by refining the query. We call this process human-centered visual + +![](images/c824de588e1e6a60c0dfc19054f26889e11793083914afd1bc9fc1d7ff33ef1f.jpg) +(a) CUHK-PEDES +Figure 3. The similarity statistics of the Top-1 retrieved items in the test sets of the three benchmarks. It is obvious that the items with higher similarity are more likely to be correct retrievals. + +![](images/f5e32f4be4afd3a7078b77138f1c5a39b731211571d68d7a7759261c686935bd.jpg) +(b) ICFG-PEDES + +![](images/6a7d7a9a7a350b03bdf20481d1c0a06d7f9d22c6199507201d610b14c37eee61.jpg) +(c) RSTPReid + +question answering (VQA), which is expressed as: + +$$ +r _ {\bar {v}} = \mathcal {M} \left(\mathcal {T} _ {\mathrm {v q a}} \left(\left\{c _ {i} \right\} _ {i = 1} ^ {N _ {q}}, \bar {v}\right)\right), \tag {4} +$$ + +where $\{c_i\}_{i=1}^{N_q}$ are $N_q$ questions directed at the detail characteristics (e.g., gender, hair, upper body, lower body, shoes, etc.) of the pedestrian image, $r_{\bar{v}}$ is the answer to $\{c_i\}_{i=1}^{N_q}$ one by one, and $\mathcal{T}_{\mathrm{vqa}}$ is the prompt template function for human-centered VQA. To improve the consistency between the image and the user query, a natural strategy is to concatenate these one-by-one answers after the original text. However, due to the limitation of the maximum text length that can be processed by the text model (e.g., it is usually 77 for CLIP), this will destroy the sequence structure. To this end, we recommend merging these answers and the original text by MLLMs to obtain a fluent and concise text, i.e., + +$$ +\hat {q} = \mathcal {M} \left(\mathcal {T} _ {\text {a g g r}} \left(r _ {\bar {v}}, q\right)\right), \tag {5} +$$ + +where $\hat{q}$ is the merged text query and $\mathcal{T}_{\mathrm{aggr}}$ is the prompt template function for text aggregation. + +Efficient Re-ranking. To apply the above process to large-scale evaluation, efficiency is a factor that needs to be considered in the retrieval task, since introducing MLLMs to interactions for external guidance inevitably brings additional inference costs. To improve the interaction efficiency, we recommended adopting different strategies for different rounds of interactions. We observe that the retrieval quality is positively correlated with the similarity between the query and the Top-1 candidate image, as shown in Figure 3. To this end, in the first round, we only interact with images whose cross-modal similarity is greater than a threshold $\xi$ and the answer of anchor localization is 'Yes'. This is because the retrievals with low similarities are more likely to be wrong items. After the first round, we only interact with the image that is all considered 'No' in the previous rounds of anchor localization and whose cross-modal similarity is less than the threshold $\xi$ . This allows us to reduce unnecessary interactions and find text queries that really require interactions, making our method more efficient on large-scale evaluation. Given a query $q$ , if the above constraints are met, the interaction is carried out and we can get merged text $\hat{q}$ after human-centered VQA, the re-ranking similarity $\hat{S}_{q,v}$ of any image $v$ in $\mathcal{V}$ is: + +$$ +\hat {S} _ {q, v} = \lambda S _ {q, v} + (1 - \lambda) \bar {S} _ {\hat {q}, v}, \tag {6} +$$ + +where $\bar{S}_{\hat{q},v} \equiv 1$ if $v$ is $\hat{v}_1$ , otherwise, $\bar{S}_{\hat{q},v} \equiv S_{\hat{q},v}$ , and $\lambda \in [0,1]$ is a hyperparameter to balance the contribution of the raw query and the refined text. To make our approach clearer, we describe the detailed algorithm process of our THI in Algorithm 1. Due to space limitations, all prompt templates can be found in the supplementary material. + +Algorithm 1 The interaction process of our THI +Input: The query set $\mathcal{Q}$ , the image gallery $\mathcal{V}$ , the offline model $f_{\mathrm{cross}}$ , the MLLM $\mathcal{M}$ , the similarity threshold $\xi$ , the number of interaction rounds $K$ ; +1: Obtain candidate sets $\{\hat{\mathcal{V}}(q_i)\}_{i=1}^{|\mathcal{Q}|}$ for all queries in $\mathcal{Q}$ via Equation (1); +2: for $k = 1,2,\dots,K$ do +3: for $i = 1,2,\dots,|\mathcal{Q}|$ do +4: Conduct anchor localization via Equation (2) and output the answer of $a_{\hat{v}_k^i}^{q_i}$ based on the $k$ -th candidate image $\hat{v}_k^i$ in $\hat{\mathcal{V}}(q_i)$ ; +5: if $a_{\hat{v}_k^i}^{q_i}$ shows 'Yes', $k = 1$ , and $S_{q_i,\hat{v}_1^i} > \xi$ then +6: Conduct human-centered VQA via Equations (4) and (5) to get the refined query $\hat{q}_i$ ; +7: Compute the re-ranking similarities between query $q_i$ and all images via Equation (6); +8: end if +9: if $\{a_{\hat{v}_j^i}\}_{j=1}^{k-1}$ all show 'No', $a_{\hat{v}_k^i}^{q_i}$ shows 'Yes', $k > 1$ , and $S_{q_i,\hat{v}_1^i} \leq \xi$ then +10: Conduct human-centered VQA via Equations (4) and (5) to get the refined query $\hat{q}_i$ ; +11: Compute the re-ranking similarities between query $q_i$ and all images via Equation (6); +12: end if +13: end for +14: end for +15: Re-ranking based on similarities; +Output: The new candidate images. + +# 3.3. Reorganization Data Augmentation + +Although THI can exploit interactions to provide external knowledge during test time to improve retrieval quality, the cross-modal embedding model is still a bottleneck that limits its further performance improvement. To this end, as shown in Figure 4, we introduce a new Reorganization Data Augmentation (RDA) strategy by enriching, decomposing, and reorganizing person descriptions to improve the discriminability and diversity of training data. We first apply Human-centered VQA in THI to obtain relevant characteristic descriptions and merge them with the original text, thus obtaining a text with richer information, i.e., $\hat{q} = \mathcal{M}(\mathcal{T}_{\mathrm{aggr}}(\mathcal{M}(\mathcal{T}_{\mathrm{vqa}}(\{c_i\}_{i=1}^{N_q}, v)), q))$ . Then, we apply MLLMs to decompose enriched text $\hat{q}$ into multiple independent sub-sentences that describe individual at- + +![](images/f12d5a579f973bb0872b8ae0e4528bf3b9a1fe6a3b6fd2b5de03f9b475a34af2.jpg) +Figure 4. The illustration of our RDA. The purpose of RDA is to supplement more details to the original training texts through human-centered VQA, improving the discriminability of texts. In addition, to enhance diversity, RDA maximizes diversity through the Decomposition-Rewriting-Reorganization strategy. + +tributes without interfering with each other, i.e., $\{\tilde{q}_i\}_{i=1}^n = \mathcal{M}(\mathcal{T}_{\mathrm{dec}}(\hat{q}))$ , where $\{\tilde{q}_i\}_{i=1}^n$ are the $n$ sub-sentences and $\mathcal{T}_{\mathrm{dec}}$ is the prompt template function for text decomposition. The purpose of decomposition is to be able to reorganize the text in a different order later. To increase diversity, we rewrite each sub-sentence into multiple sentences with different styles but the same meaning. For sub-sentence $\tilde{q}$ , we conduct rewriting by $\mathcal{R} = \{\check{q}_j\}_{j=1}^m = \mathcal{M}(\mathcal{T}_{\mathrm{rwt}}(\tilde{q}))$ , where $\mathcal{R}$ is the set with $m$ rewritten sentences and $\mathcal{T}_{\mathrm{rwt}}$ is the prompt template function for text rewriting. + +Finally, for a text $q$ , we can obtain the set $\{\mathcal{R}_i\}_{i=1}^n$ that contains a large number of style sub-sentences. So far, we can obtain augmented texts with different sub-sentence orders and different style combinations as shown in Figure 4, which we represent as $\check{q}$ . During training, we can mix the augmentation texts $(\check{q})$ with the original texts $(q)$ for cross-modal learning, thereby improving model-intra knowledge and generalization. Due to space limitations, more training details can be found in the supplementary material. + +# 4. Experiments + +In this section, we conduct extensive experiments to verify the effectiveness, superiority, and generalization of the proposed ICL on four public benchmark datasets. + +# 4.1. Datasets and Evaluation Protocols + +Datasets. In our experiments, we use three coarse-grained benchmarks, CHUK-PEDES [16], ICFG-PEDES [5], and RSTPReid [41], and one fine-grained benchmark, UFine6926 [44], to evaluate our ICL. Compared with the coarse-grained benchmark, the fine-grained benchmark has richer texts with fine-grained information. For all datasets, we follow their official settings for data partitioning. More details are provided in the supplementary materials. + +Evaluation Protocols. To measure performance, we utilize the widely accepted Rank-K (1,5,10) metrics to mea + +
MethodsImage Enc.Text Enc.CUHK-PEDESICFG-PEDESRSTPReid
Rank-1Rank-5Rank-10mAPmINPRank-1Rank-5Rank-10mAPmINPRank-1Rank-5Rank-10mAPmINP
① VL-Backbones w/o ReID-domain pre-training
IVT [26]ViT-BaseBERT65.6985.9391.1560.66-56.0473.6080.22--46.7070.0078.80--
\( LCR^2S \) [36]RN50BERT67.3684.1989.6259.20-57.9376.0882.4038.21-54.9576.6584.7040.92-
CFine [37]CLIP-ViTBERT69.5785.9391.15--60.8376.5582.42--50.5572.5081.60--
RaSa [1]Swin-BBERT76.5190.2994.2569.38-65.2880.4085.1241.29-66.9086.5091.3552.31-
IRRA [12]CLIP-ViTCLIP-X.73.3889.9393.7166.1350.2463.4680.2585.8238.067.9360.2081.3088.2047.1725.28
TBPS [3]CLIP-ViTCLIP-X.73.5488.1992.3565.3849.2565.0580.3485.4739.837.8762.1081.9087.7548.0025.86
CFAM [44]CLIP-ViTCLIP-X.75.6090.5394.3667.27-65.3881.1786.3539.42-62.4583.5591.1049.50-
RDE [21]CLIP-ViTCLIP-X.75.9490.1494.1267.5651.4467.6882.4787.3640.067.8765.3583.9589.9050.8828.08
Our ICLCLIP-ViTCLIP-X.76.4190.4894.3368.0451.9968.1182.5987.5240.818.1867.7086.0591.7552.6229.36
Our ICL*CLIP-ViTCLIP-X.77.9190.2794.1469.1353.4069.0282.4587.3641.218.3070.5585.9591.6553.6830.13
② VL-Backbones with ReID-domain pre-training
\( IRRA^b \) [12]CLIP-ViTCLIP-X.74.0589.4893.6466.57-64.3780.7586.1238.85-61.9080.6089.3048.08-
APTM [38]Swin-BBERT76.5390.0494.1566.91-68.5182.9987.5641.22-67.5085.7091.4552.56-
\( NAM^d \) [28]CLIP-ViTCLIP-X.77.4790.8494.6769.4354.0866.7682.0287.1741.459.5367.1586.5591.9052.0028.46
Our ICLCLIP-ViTCLIP-X.78.1891.6394.8369.5853.4869.2283.4988.0642.349.0170.0086.6091.7054.1630.93
Our ICL*CLIP-ViTCLIP-X.79.0691.2694.7270.4454.7070.0583.3587.9142.709.1372.5586.6091.3055.1931.72
+ +Table 1. Performance on the three coarse-grained benchmarks. The results with THI are marked with $\star$ . Note that IRRA means using the pre-trained Backbones with MALS [38] and the results of NAM are reproduced by us. + +
MethodsRank-1Rank-5Rank-10mAPmINP
LGUR [25]70.6984.5789.9168.93-
SSAN [5]75.0988.6392.8473.14-
IRRA [12]85.0294.3196.7583.9177.30
RDE [21]87.6095.6597.4686.1079.54
CFAM(B/16) [44]85.5594.5197.0284.23-
CFAM(L/14) [44]88.5195.5897.4987.09-
1 Our ICL89.1796.1397.8887.4981.50
1 Our ICL★90.6795.9897.8688.2982.60
2 Our ICL91.0296.9898.1789.7684.70
2 Our ICL★91.7896.8398.1690.3385.62
+ +Table 2. Performance comparison on the UFine6926 dataset. The results of IRRA and RDE are reproduced by us. + +sure the TIREID performance. Like [12, 21], we also report the mean Average Precision (mAP) and mean Inverse Negative Penalty (mINP) as auxiliary metrics. + +# 4.2. Implementation Details + +To achieve interactive TIReID, we choose RDE [21] and Qwen2-VL-7B-Instruct [29] as the investigated TIReID method and MLLM. For the sake of fairness, we do not modify any settings of RDE, including model architecture (CLIP-ViTB/16 [24]), hyperparameters, and training parameters. To be compatible with the fine-grained benchmark, following [44], the maximum length of the textual tokens of CLIP is set to 168 by interpolating the positional embedding layer with an initial learning rate of $5e - 5$ . As for fine-tuning MLLMs, we use the Llama-Factory framework [40] to conduct SFT with LoRA [11] for 2 epochs. The train batch size per device is set to 4, the gradient accumulation steps are set to 16, the initial learning rate is $5e - 5$ and the hyperparameters $\alpha$ and $r$ of LoRA are set to 16 and 8, respectively. During inference, the temperature is set to 0.01 to keep reproducibility. Note that all experiments can be completed on two GeForce RTX 24GB 3090 GPUs. + +# 4.3. Comparison with State-of-the-Arts + +In this section, to verify the superiority of our ICL, we compare our method with more than 15 baselines including recent advanced methods (e.g., RDE (CVPR'24) [21] and NAM (CVPR'24) [28]). Based on the backbone type, we divide the baselines into two groups (1 and 2) as shown Table 1, i.e., the baselines using VL-Backbones w/o and with ReID-domain pre-training. For the group 2, we use the pretrained weights released by [28] to initialize CLIP for fair comparison. The results are reported in Tables 1 and 2. + +Results on coarse-grained datasets. Table 1 report the results evaluated on the coarse-grained datasets. We can see that our method can achieve competitive performance even without THI. By performing THI, the Rank-1 score of our method is greatly improved, e.g., in the group ①, the Rank-1 scores on the three datasets are improved by $1.50\%$ , $0.91\%$ , and $2.85\%$ , respectively. In addition, mAP and mINP scores have also improved greatly, which indicates that the overall ranking has improved. In group ②, our method achieves the best scores on most metrics, especially Rank-1 reached $72.55\%$ on RSTPReid, which is sufficient to verify the superiority. However, we found that THI slightly degraded Rank-5 and Rank-10, which is because we only conducted 5 rounds of interaction. As long as the locating anchor image is incorrect, the errors will accumulate, which can be relieved by increasing the rounds of interaction. But in general, THI can significantly improve Rank-1 and the overall ranking (mAP and mINP). + +Results on fine-grained dataset. Table 2 shows the results on the fine-grained dataset whose average text length is over 80. Such kind of fine-grained description often has a clear query intention. For better comparison, we provide the performance of IRRA and RDE on the UFine6926 dataset. From the results, our method can still achieve excellent per + +
MethodsTraining SetsCUHK-PEDESICFG-PEDESRSTPReid
Rank-1Rank-5Rank-10mAPmINPRank-1Rank-5Rank-10mAPmINPRank-1Rank-5Rank-10mAPmINP
IRRA [12]CUHK-PEDES73.3889.9393.7166.1350.2442.4162.1169.6221.771.9553.2577.1585.3539.6316.60
ICFG-PEDES33.4856.2966.3331.5619.2063.4680.2585.8238.067.9345.3069.2578.8036.8218.38
RSTPReid32.8055.2665.8130.2917.6132.3049.6757.8020.543.8460.2081.3088.2047.1725.28
RDE [21]CUHK-PEDES75.9490.1494.1267.5651.4448.1866.3073.7025.002.3354.9077.5086.5041.2717.84
ICFG-PEDES38.1159.2468.4434.1620.4467.6882.4787.3640.067.8749.2572.1080.2038.4618.33
RSTPReid36.9458.2267.5833.6520.4242.1758.3265.4926.374.9465.3583.9589.9050.8828.08
Our ICLCUHK-PEDES76.4190.4894.3368.0451.9948.5766.6673.7525.302.4055.8079.6087.6542.0917.41
ICFG-PEDES42.8764.2073.4438.1923.5868.1182.5987.5240.818.1852.5075.0583.0041.8221.14
RSTPReid41.3161.8670.3136.7822.3745.9362.7068.8028.895.6367.7086.0591.7552.6229.36
Our ICL*CUHK-PEDES77.9190.2794.1469.1353.4052.8066.4973.4925.602.4461.3079.2587.4043.4218.01
ICFG-PEDES49.2964.3473.5540.8225.3869.0282.4587.3641.218.3060.1575.3083.1543.7222.04
RSTPReid47.3561.4570.3438.9123.6850.5261.5668.5729.265.7370.5585.9591.6553.6830.13
+ +Table 3. Comparison of mutual generalization capabilities between coarse-grained datasets. + +
Source → TargetMethodsRank-1Rank-5Rank-10mAPmINP
CUHK. → UFine.IRRA [12]37.5154.9264.2940.7634.33
RDE [21]40.3757.4966.0542.6835.78
Our ICL46.4063.5572.0848.6841.56
Our ICL☆57.7664.1372.8153.9745.64
ICFG. → UFine.IRRA [12]15.0226.7933.9017.1012.75
RDE [21]17.8631.0138.5619.8214.74
Our ICL27.9544.2052.2029.8523.20
Our ICL☆36.8144.6552.7334.1226.61
RSTP. → UFine.IRRA [12]13.2125.6733.9315.6011.09
RDE [21]14.0025.2332.6416.2211.90
Our ICL23.8938.3046.7025.5419.20
Our ICL☆31.2338.5647.0228.9021.80
UFine. → CUHKIRRA [12]37.7460.1270.1335.9423.21
RDE [21]39.4161.1470.1136.4923.32
Our ICL49.0470.2778.6444.5429.58
Our ICL☆56.8770.1978.5347.3131.20
UFine. → ICFG.IRRA [12]34.5255.4164.4417.961.95
RDE [21]40.3760.1468.4120.542.19
Our ICL43.1062.9270.7322.732.56
Our ICL☆47.8362.6770.4823.162.62
UFine. → RSTP.IRRA [12]37.6563.7073.0029.0011.80
RDE [21]39.9063.5074.7529.9212.43
Our ICL48.8572.6581.8036.9116.39
Our ICL☆55.3572.4081.5038.6417.23
+ +Table 4. Generalization capabilities between coarse-grained and fine-grained datasets. The best scores in each task are in bold. + +formance, with Rank-1 exceeding $91\%$ . This shows that interaction is also applicable to the fine-grained scenario. + +# 4.4. Generalization Study + +To evaluate the generalization, Tables 3 and 4 report the cross-domain performance on four datasets, including the coarse-to-fine, coarse-to-fine, and fine-to-coarse generalization experiments. From the results, our method achieves better generalization than RDE even without THI thanks to the training enhancement of RDA. When THI is performed, the cross-domain performance is dramatically improved, for example, from CUHK-PEDES to RSTPReid, THI brings an improvement of more than $4\%$ on Rank-1. This shows that THI is a potential solution to generaliza + +tion challenges in the future. From the generalization experiments between fine-grained and coarse-grained datasets shown in Table 4, ICL can also achieve the best cross-domain performance, e.g., compared with the best baseline RDE, from UFine6926 domain to CUHK-PEDES domain, our method improves Rank-1 and mAP by $17.49\%$ and $10.86\%$ , respectively, which further verifies the cross-domain generalization of our method. + +
MethodsTHICUHK-PEDESICFG-PEDESRSTPReidΔAvg
Rank-1mAPRank-1mAPRank-1mAP
CLIP [24]X71.6463.9260.1134.5256.5544.52+2.57
✓73.7765.6663.5734.9561.4546.25
IRRA [12]X73.3866.1363.4638.0660.2047.17+1.86
✓76.0667.4265.2638.5863.7548.47
RDE [21]X75.9467.5667.6840.0665.3550.88+1.41
✓77.4768.6268.7240.6368.4552.01
+ +Table 5. Transferability results on three coarse-grained benchmarks. $\Delta$ Avg represents the average improvement. + +# 4.5. Transferability Study + +Our interactive module is separate and independent from the training of existing TIReID methods, so it can be plug-and-play with existing methods to improve ReID performance. To verify the transferability of our THI, we conduct experiments on multiple baselines, and the results are shown in Table 5. Except for the CLIP results we reproduced, all other experiments used public pre-trained model weights. From the results of Table 5, the interactive strategy application can significantly improve Rank-1 and mAP, which shows that the external guidance by interactions via MLLMs can further clarify the text-image alignments and improve the overall ranking. The average improvements on the three datasets are $2.57\%$ , $1.86\%$ , and $1.41\%$ , respectively, which seriously proves the transferability of our THI. + +# 4.6. Ablation and Parameter Study + +To explore the effects of each proposed component, i.e., THI and RDA, we first conduct the ablation study on three + +coarse-grained datasets as reported in Table 6. From the results, each component can bring performance gains on Rank-1 and mAP, which verifies the reliability and rationality of the method. Especially the introduction of THI has greatly improved the Rank-1 accuracy from $68.95\%$ to $70.55\%$ on the RSTPreid dataset. In addition, our RDA also brings considerable performance gains, especially on the RSTPreid dataset. Also, we conduct the parameter analysis on the CHUK-PEDES dataset on two free hyperparameters, i.e., the similarity threshold $\xi$ and the balance factor $\lambda$ . The former filters unnecessary retrievals to improve the interaction efficiency, while the latter controls the contribution of refined texts. Based on Figure 5, in our experiments, we set $\xi$ in the range of $0.5 \sim 0.6$ and $\lambda$ to 0.8, thus mitigating the risk of introducing noisy external knowledge. + +
No.THIRDALoRACUHK-PEDESICFG-PEDESRSTPReid
Rank-1mAPRank-1mAPRank-1mAP
#1✓✓✓77.9169.3269.0241.2170.5553.68
#2✓✓×76.3868.5967.9241.1369.0053.11
#3×✓×76.4168.0468.1140.8167.7052.62
#4×××75.9467.5667.6840.0665.3550.88
+ +Table 6. Ablation studies on CHUK-PEDES, ICFG-PEDES, and RSTPReid datasets. The best scores are in bold. + +![](images/3d7772f515054dae074f670307524539aa3f362bf5097f1182732169c1447a2f.jpg) +(a) The similarity threshold $\xi$ +Figure 5. Variation of performance with different $\xi$ and $\lambda$ . + +![](images/c584e6d8acd84e0846325cc373435b3a25138bd51eba753542aedb69fc859d7b.jpg) +(b) The balance factor $\lambda$ + +# 4.7. Interactive Study + +This section further explores the interactive module, i.e. THI. We report the performance (mAP) changes after multiple rounds of interactions in Figure 6. The mAP score can reflect the overall retrieval quality. From the results, the performance improvement is obvious in the first few rounds $(\leq 2)$ since the items that meet the query semantics are mostly concentrated at the top of the ranking. However, after $>2$ rounds, the performance gain is gradually not obvious as the number of queries requiring interaction decreases. But generally, as the number of rounds increases, the overall performance gradually improves. In all our experiments, we performed 5 rounds of interaction. In addition, we visualize the top-10 retrieved results before and after applying THI in Figure 7. Due to the inherent limitations of the intra-model knowledge, it is difficult to obtain satisfactory results without the help of THI. In contrast, our THI can dynamically enrich queries by interacting with MLLMs to achieve a more reliable ranking. More example results are given in the supplementary material. + +![](images/a468f43e12e93e9db9044752c1b7bab4faeafca3d2345fae931fe51558325a3b.jpg) +(a) CUHK-PEDES + +![](images/934c89310c3ab053d56bdfe58e5389df97e51d022f40a831122983b2620aa41d.jpg) +(b) ICFG-PEDES +Figure 6. Performance (mAP) versus rounds on three datasets. Round 0 indicates the setting without using THI. + +![](images/543546290431ffbbd3f2bc3c372f4df8c65819deec1bed410650b8973891e94e.jpg) +(c) RSTPReid + +![](images/d81673e82a8d80de989625b02f3c922b1d4291443a750f08dde17504ee196e20.jpg) + +![](images/ad0c671cfb7951ea6984a1adb15a69a27340c9ac356219ca45d2459bec9af3f6.jpg) + +![](images/fb579c33600b81633192f7179c87f2f24f84bb02991a511821ce96e48471a5bb.jpg) +Figure 7. Top-10 retrieved results on CUHK-PEDES dataset between ICL (the first row) and ICL with THI (the second row). + +![](images/7042d8b8e506934f997341f636bbb76dfbd27533654a8eaaf277610a97d429f5.jpg) + +# 5. Conclusion + +In this paper, we explore interactive text-to-image person re-identification, which aims to improve the alignment between dynamic queries and challenging candidate images by leveraging external guidance from MLLMs. To achieve this, we develop an Interactive Cross-modal Learning (ICL) framework to alleviate the inherent challenges of offline models and training data by, including a plug-and-play Test-time Human-centered Interaction (THI) module and Reorganization Data Augmentation (RDA). Extensive experiments and analysis show that our framework can effectively transfer external knowledge in MLLMs into offline models for guiding re-identification, showing excellent performance and generalization. + +# Acknowledgments + +This work was supported in part by the National Key R&D Program of China under Grant 2024YFB4710604; in part by NSFC under Grants 62472295, 62372315, 62176171, and U21B2040; in part by Sichuan Science and Technology Planning Projects (2024YFHZ0089, 2024NSFTD0049, 2024YFHZ0144, 2024NSFTD0047, 2024NSFTD0038); in part by System of Systems and Artificial Intelligence Laboratory pioneer fund grant; in part by the Fundamental Research Funds for the Central Universities under Grant CJ202403 and CJ202303; in part by Chengdu Science and Technology Project under Grant 2023-XT00-00004-GX. + +# References + +[1] Yang Bai, Min Cao, Daming Gao, Ziqiang Cao, Chen Chen, Zhenfeng Fan, Liqiang Nie, and Min Zhang. Rasa: relation and sensitivity aware representation learning for text-based person search. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, pages 555–563, 2023. 2, 6 +[2] Maryam Bukhari, Sadaf Yasmin, Sheneela Naz, Muazzam Maqsood, Jehyeok Rew, and Seungmin Rho. Language and vision based person re-identification for surveillance systems using deep learning with lip layers. Image and Vision Computing, 132:104658, 2023. 1 +[3] Min Cao, Yang Bai, Ziyin Zeng, Mang Ye, and Min Zhang. An empirical study of clip for text-based person search. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 465-473, 2024. 1, 2, 6 +[4] Yuhao Chen, Guoqing Zhang, Yujiang Lu, Zhenxing Wang, and Yuhui Zheng. Tipcb: A simple but effective part-based convolutional baseline for text-based person search. Neurocomputing, 494:171-181, 2022. +[5] Zefeng Ding, Changxing Ding, Zhiyin Shao, and Dacheng Tao. Semantically self-aligned network for text-to-image part-aware person re-identification. arXiv preprint arXiv:2107.12666, 2021. 1, 5, 6 +[6] Hiren Galiyawala and Mehul S Raval. Person retrieval in surveillance using textual query: a review. Multimedia Tools and Applications, 80(18):27343-27383, 2021. 1 +[7] Daming Gao, Yang Bai, Min Cao, Hao Dou, Mang Ye, and Min Zhang. Semi-supervised text-based person search. arXiv preprint arXiv:2404.18106, 2024. 1 +[8] Donghoon Han, Eunhwan Park, Gisang Lee, Adam Lee, and Nojun Kwak. Merlin: Multimodal embedding refinement via llm-based iterative navigation for text-video retrieval-erank pipeline. arXiv preprint arXiv:2407.12508, 2024. 2, 3 +[9] Chen He, Shenshen Li, Zheng Wang, Hua Chen, Fumin Shen, and Xing Xu. Chatting with interactive memory for text-based person retrieval. Multimedia Systems, 31(1):31, 2025. 3 +[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 2 +[11] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 4, 6 +[12] Ding Jiang and Mang Ye. Cross-modal implicit relation reasoning and aligning for text-to-image person retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2787-2797, 2023. 1, 2, 6, 7 +[13] Matan Levy, Rami Ben-Ari, Nir Darshan, and Dani Lischinski. Chatting makes perfect: Chat-based image retrieval. Advances in Neural Information Processing Systems, 36, 2024. 2, 3 +[14] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified + +vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022. 2 +[15] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 3 +[16] Shuang Li, Tong Xiao, Hongsheng Li, Bolei Zhou, Dayu Yue, and Xiaogang Wang. Person search with natural language description. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1970-1979, 2017. 5 +[17] Yunfan Li, Peng Hu, Dezhong Peng, Jiancheng Lv, Jianping Fan, and Xi Peng. Image clustering with external guidance. In *Forty-first International Conference on Machine Learning*. 2 +[18] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 3 +[19] Yang Qin, Dezhong Peng, Xi Peng, Xu Wang, and Peng Hu. Deep evidential learning with noisy correspondence for cross-modal retrieval. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4948-4956, 2022. 1 +[20] Yang Qin, Yuan Sun, Dezhong Peng, Joey Tianyi Zhou, Xi Peng, and Peng Hu. Cross-modal active complementary learning with self-refining correspondence. Advances in neural information processing systems, 36:24829-24840, 2023. 1 +[21] Yang Qin, Yingke Chen, Dezhong Peng, Xi Peng, Joey Tianyi Zhou, and Peng Hu. Noisy-correspondence learning for text-to-image person re-identification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27197-27206, 2024. 1, 2, 6, 7 +[22] Yalan Qin, Chuan Qin, Xinpeng Zhang, and Guorui Feng. Dual consensus anchor learning for fast multi-view clustering. IEEE Transactions on Image Processing, 2024. 1 +[23] Yalan Qin, Nan Pu, Nicu Sebe, and Guorui Feng. Latent space learning based ensemble clustering. IEEE Transactions on Image Processing, 2025. 1 +[24] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 6, 7 +[25] Zhiyin Shao, Xinyu Zhang, Meng Fang, Zhifeng Lin, Jian Wang, and Changxing Ding. Learning granularity-unified representations for text-to-image person re-identification. In Proceedings of the 30th ACM International Conference on Multimedia, 2022. 6 +[26] Xiujun Shu, Wei Wen, Haoqian Wu, Keyu Chen, Yiran Song, Ruizhi Qiao, Bo Ren, and Xiao Wang. See finer, see more: Implicit modality alignment for text-based person retrieval. In European Conference on Computer Vision, pages 624-641. Springer, 2022. 6 + +[27] Yuan Sun, Yang Qin, Yongxiang Li, Dezhong Peng, Xi Peng, and Peng Hu. Robust multi-view clustering with noisy correspondence. IEEE Transactions on Knowledge and Data Engineering, 2024. 1 +[28] Wentan Tan, Changxing Ding, Jiayu Jiang, Fei Wang, Yibing Zhan, and Dapeng Tao. Harnessing the power of mllms for transferable text-to-image person reid. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17127-17137, 2024. 1, 2, 3, 6 +[29] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 3, 6 +[30] Yansong Wang, Yundong Sun, Yansheng Fu, Dongjie Zhu, and Zhaoshuo Tian. Spectrum-bert: pre-training of deep bidirectional transformers for spectral classification of Chinese liquors. IEEE Transactions on Instrumentation and Measurement, 2024. 2 +[31] Yabing Wang, Le Wang, Qiang Zhou, Zhibin Wang, Hao Li, Gang Hua, and Wei Tang. Multimodal llm enhanced cross-lingual cross-modal retrieval. arXiv preprint arXiv:2409.19961, 2024. 3 +[32] Yuhao Wang, Yang Liu, Aihua Zheng, and Pingping Zhang. Demo: Decoupled feature-based mixture of experts for multi-modal object re-identification. In Proceedings of the AAAI Conference on Artificial Intelligence, 2025. 1 +[33] Zhe Wang, Zhiyuan Fang, Jun Wang, and Yezhou Yang. Vistaa: Visual-textual attributes alignment in person search by natural language. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XII 16, pages 402-420. Springer, 2020. 1, 2 +[34] Zijie Wang, Aichun Zhu, Jingyi Xue, Xili Wan, Chao Liu, Tian Wang, and Yifeng Li. Caibc: Capturing all-round information beyond color for text-based person retrieval. In Proceedings of the 30th ACM International Conference on Multimedia, pages 5314-5322, 2022. 1 +[35] Yushuang Wu, Zizheng Yan, Xiaoguang Han, Guanbin Li, Changqing Zou, and Shuguang Cui. Lipscore: language-guided person search via color reasoning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1624-1633, 2021. 2 +[36] Shuanglin Yan, Neng Dong, Jun Liu, Liyan Zhang, and Jinhui Tang. Learning comprehensive representations with richer self for text-to-image person re-identification. In Proceedings of the 31st ACM international conference on multimedia, pages 6202–6211, 2023. 6 +[37] Shuanglin Yan, Neng Dong, Liyan Zhang, and Jinhui Tang. Clip-driven fine-grained text-image person re-identification. IEEE Transactions on Image Processing, 2023. 1, 6 +[38] Shuyu Yang, Yinan Zhou, Zhedong Zheng, Yaxiong Wang, Li Zhu, and Yujiao Wu. Towards unified text-based person retrieval: A large-scale multi-attribute and language search benchmark. In Proceedings of the 31st ACM International Conference on Multimedia, pages 4492–4501, 2023. 1, 2, 6 + +[39] Pingping Zhang, Yuhao Wang, Yang Liu, Zhengzheng Tu, and Huchuan Lu. Magic tokens: Select diverse tokens for multi-modal object re-identification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17117-17126, 2024. 1 +[40] Yaowei Zheng, Richong Zhang, Junhao Zhang, Yanhan Ye, Zheyan Luo, Zhangchi Feng, and Yongqiang Ma. Llamafactory: Unified efficient fine-tuning of $100+$ language models. arXiv preprint arXiv:2403.13372, 2024. 6 +[41] Aichun Zhu, Zijie Wang, Yifeng Li, Xili Wan, Jing Jin, Tian Wang, Fangqiang Hu, and Gang Hua. Dssl: Deep surroundings-person separation learning for text-based person retrieval. In Proceedings of the 29th ACM International Conference on Multimedia, pages 209-217, 2021. 1, 2, 5 +[42] Jialong Zuo, Changqian Yu, Nong Sang, and Changxin Gao. Plip: Language-image pre-training for person representation learning, 2023. 1 +[43] Jialong Zuo, Ying Nie, Hanyu Zhou, Huaxin Zhang, Haoyu Wang, Tianyu Guo, Nong Sang, and Changxin Gao. Cross-video identity correlating for person re-identification pretraining. arXiv preprint arXiv:2409.18569, 2024. 1 +[44] Jialong Zuo, Hanyu Zhou, Ying Nie, Feng Zhang, Tianyu Guo, Nong Sang, Yunhe Wang, and Changxin Gao. Ufinebench: Towards text-based person retrieval with ultra-fine granularity. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22010-22019, 2024. 1, 5, 6 \ No newline at end of file diff --git a/2025/Human-centered Interactive Learning via MLLMs for Text-to-Image Person Re-identification/images.zip b/2025/Human-centered Interactive Learning via MLLMs for Text-to-Image Person Re-identification/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..da815a18577f8922a429a551a9ad007f2fd93a32 --- /dev/null +++ b/2025/Human-centered Interactive Learning via MLLMs for Text-to-Image Person Re-identification/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9e6909357cc875d03c1ed74f2abaff1cadbc47bde84533a3aa11883cb347df2f +size 903387 diff --git a/2025/Human-centered Interactive Learning via MLLMs for Text-to-Image Person Re-identification/layout.json b/2025/Human-centered Interactive Learning via MLLMs for Text-to-Image Person Re-identification/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9ee6547d0725769ca41762c478ecd0af7de9b6d9 --- /dev/null +++ b/2025/Human-centered Interactive Learning via MLLMs for Text-to-Image Person Re-identification/layout.json @@ -0,0 +1,9320 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 93, + 103, + 518, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 103, + 518, + 138 + ], + "spans": [ + { + "bbox": [ + 93, + 103, + 518, + 138 + ], + "type": "text", + "content": "Human-centered Interactive Learning via MLLMs for Text-to-Image Person Re-identification" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 96, + 160, + 517, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 160, + 517, + 176 + ], + "spans": [ + { + "bbox": [ + 96, + 160, + 517, + 176 + ], + "type": "text", + "content": "Yang Qin" + }, + { + "bbox": [ + 96, + 160, + 517, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 96, + 160, + 517, + 176 + ], + "type": "text", + "content": " Chao Chen" + }, + { + "bbox": [ + 96, + 160, + 517, + 176 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 96, + 160, + 517, + 176 + ], + "type": "text", + "content": " Zhihang Fu" + }, + { + "bbox": [ + 96, + 160, + 517, + 176 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 96, + 160, + 517, + 176 + ], + "type": "text", + "content": " Dezhong Peng" + }, + { + "bbox": [ + 96, + 160, + 517, + 176 + ], + "type": "inline_equation", + "content": "^{1,4,5}" + }, + { + "bbox": [ + 96, + 160, + 517, + 176 + ], + "type": "text", + "content": " Xi Peng" + }, + { + "bbox": [ + 96, + 160, + 517, + 176 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 96, + 160, + 517, + 176 + ], + "type": "text", + "content": " Peng Hu" + }, + { + "bbox": [ + 96, + 160, + 517, + 176 + ], + "type": "inline_equation", + "content": "^{1*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 147, + 176, + 463, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 176, + 463, + 190 + ], + "spans": [ + { + "bbox": [ + 147, + 176, + 463, + 190 + ], + "type": "text", + "content": "1College of Computer Science, Sichuan University 2Independent Researcher" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 73, + 190, + 536, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 190, + 536, + 204 + ], + "spans": [ + { + "bbox": [ + 73, + 190, + 536, + 204 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 73, + 190, + 536, + 204 + ], + "type": "text", + "content": "National Key Laboratory of Fundamental Algorithms and Models for Engineering Simulation, Sichuan University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 86, + 204, + 520, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 204, + 520, + 217 + ], + "spans": [ + { + "bbox": [ + 86, + 204, + 520, + 217 + ], + "type": "text", + "content": "Sichuan National Innovation New Vision UHD Video Technology Co., Ltd Tianfu Jincheng Laboratory" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 244, + 200, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 244, + 200, + 257 + ], + "spans": [ + { + "bbox": [ + 151, + 244, + 200, + 257 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 270, + 296, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 296, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 296, + 605 + ], + "type": "text", + "content": "Despite remarkable advancements in text-to-image person re-identification (TIREID) facilitated by the breakthrough of cross-modal embedding models, existing methods often struggle to distinguish challenging candidate images due to intrinsic limitations, such as network architecture and data quality. To address these issues, we propose an Interactive Cross-modal Learning framework (ICL), which leverages human-centered interaction to enhance the discriminability of text queries through external multimodal knowledge. To achieve this, we propose a plug-and-play Test-time Humane-centered Interaction (THI) module, which performs visual question answering focused on human characteristics, facilitating multi-round interactions with a multimodal large language model (MLLM) to align query intent with latent target images. Specifically, THI refines user queries based on the MLLM responses to reduce the gap to the best-matching images, thereby boosting ranking accuracy. Additionally, to address the limitation of low-quality training texts, we introduce a novel Reorganization Data Augmentation (RDA) strategy based on information enrichment and diversity enhancement to enhance query discriminability by enriching, decomposing, and reorganizing person descriptions. Extensive experiments on four TIREID benchmarks, i.e., CUHK-PEDES, ICFG-PEDES RSTPReid, RSTPReid, and UFine6926, demonstrate that our method achieves remarkable performance with substantial improvement. Code is available at https://github.com/QinYang79/ICL." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 627, + 135, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 627, + 135, + 639 + ], + "spans": [ + { + "bbox": [ + 56, + 627, + 135, + 639 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 647, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 296, + 696 + ], + "type": "text", + "content": "Recently, text-to-image person re-identification (TIREID) [3-5, 7, 33, 34, 41] has made great progress in aligning text descriptions and person images, enabling high-accuracy person search and identification. Different from traditional" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 316, + 243, + 555, + 306 + ], + "blocks": [ + { + "bbox": [ + 316, + 243, + 555, + 306 + ], + "lines": [ + { + "bbox": [ + 316, + 243, + 555, + 306 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 555, + 306 + ], + "type": "image", + "image_path": "fba52da64d005a0539e6ec922a99984925182eb012c93d057d25364f0a94c25b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 405, + 306, + 467, + 314 + ], + "lines": [ + { + "bbox": [ + 405, + 306, + 467, + 314 + ], + "spans": [ + { + "bbox": [ + 405, + 306, + 467, + 314 + ], + "type": "text", + "content": "(a) Existing methods" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 317, + 317, + 555, + 404 + ], + "blocks": [ + { + "bbox": [ + 317, + 317, + 555, + 404 + ], + "lines": [ + { + "bbox": [ + 317, + 317, + 555, + 404 + ], + "spans": [ + { + "bbox": [ + 317, + 317, + 555, + 404 + ], + "type": "image", + "image_path": "0f3b0b8518086d3904eb882445c010dfce5f4e8b3ce5da480db96badb7397d1b.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 402, + 406, + 470, + 415 + ], + "lines": [ + { + "bbox": [ + 402, + 406, + 470, + 415 + ], + "spans": [ + { + "bbox": [ + 402, + 406, + 470, + 415 + ], + "type": "text", + "content": "(b) Interactive TIREID" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 417, + 555, + 538 + ], + "lines": [ + { + "bbox": [ + 313, + 417, + 555, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 417, + 555, + 538 + ], + "type": "text", + "content": "Figure 1. The illustration of our motivation. When performing text-based person re-identification, (a) existing methods commonly exploit cross-modal models to calculate similarity (such as IRRA [12] and RDE [21]) and then obtain the candidate person images by ranking. However, due to the intrinsic limitations of models and training data, they may not be able to distinguish challenging candidate images well enough to obtain satisfactory results. (b) Our motivation is to interact with the system for external guidance like a human and gradually refine the user's query from the candidate items shown in (a) by multiple rounds of question-answering, ultimately improving the overall ranking." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 544, + 555, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 555, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 555, + 652 + ], + "type": "text", + "content": "image-to-image re-identification [32, 39, 42, 43], TIREID, as a rising task in the multimodal community [19, 20, 22, 23, 27], retrieves person images using user-customized text queries, which is more practical in many scenarios where an image query is unavailable, such as in video surveillance systems [2] or crowd management [6]. A primary challenge in TIREID is learning to associate text queries with the corresponding person images at a fine-grained level, bridging the modality gap for accurate similarity measurement." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": "To overcome this challenge, recent efforts have developed various strategies to improve cross-modal associations, including local attribute modeling [37, 44], loss function designing [12, 21], ReID-domain pre-training [28, 38], etc. Although these methods achieve promising perfor" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 259, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 259, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 259, + 713 + ], + "type": "text", + "content": "*Corresponding author: Peng Hu (penghu.ml@gmail.com)." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14390" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "type": "text", + "content": "mance, they are limited by the inherent defect of the offline models and training data, making it hard to handle dynamic user query inputs and leading to poor generalization. In practice, however, different users tend to input concise, vague, and diverse text queries based on their memories to the target person, thereby raising challenges for the offline models to understand fine-grained human characteristics from the queries. For example, as illustrated in Figure 1 (a), the user text query \"A woman walking visible from the back is wearing a white shirt, and black pants.\" fails to capture some important contextual details like handheld objects or background, resulting in inaccurate candidate images. In brief, it is hard for offline models to handle dynamic and challenging queries solely relying on their learned internal knowledge. To break through this bottleneck, we could seek to align queries with image candidates with the help of external knowledge, thereby reranking the retrieved images and boosting the identification accuracy." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 291, + 295, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 291, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 291, + 295, + 483 + ], + "type": "text", + "content": "Recent studies [8, 13, 17] have explored integrating external knowledge from pre-trained models or (multimodal) large language models to guide various downstream tasks. For instance, Li et al. [17] leveraged multimodal knowledge of pre-trained vision-language models [24] to enhance unimodal image representations for external-guided clustering. However, this approach cannot achieve post-hoc improvement for offline models, leading to infeasibility in practice due to the prohibitive cost of fine-tuning or even retraining. In cross-modal retrieval, Han et al. [8] exploited the excellent comprehension and generation capabilities of multimodal large language models (MLLMs) to design an interactive re-ranking pipeline for text-to-video retrieval. However, Han et al. [8] introduces label leakage by using the ground truth video as the video in mind for the MLLM answerer agent, which is not applicable in actual testing." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "type": "text", + "content": "In this paper, we present a novel Interactive Cross-modal Learning framework (ICL) that exploits the multimodal knowledge implicit in MLLMs to enhance the alignment between text queries and target person images. Specifically, we first propose a plug-and-play Test-time Human-centered Interaction (THI) module that refines text queries through interactions with an MLLM, enhancing the post-hoc ability of trained models to distinguish challenging candidates. THI identifies the latent target images through multi-round interactions with MLLMs, asking human-centered questions to the MLLM for fine-grained answers about the images, which refine the query texts to strengthen the alignment with the images, ultimately address the inherent limitations of the input queries and enhancing ranking accuracy. To address the intrinsic limitations in the trained models, we propose a Reorganization Data Augmentation strategy (RDA) to enrich and diversify pedestrian texts for training enhancement, thereby transferring external knowledge in MLLMs into the model. To enrich the texts, RDA applies" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 553, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 203 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 203 + ], + "type": "text", + "content": "visual question answering via the MLLM on each image-text pair to supplement fine-grained person characteristics. Moreover, to enhance the diversity of texts, RDA presents a decomposition-reorganization strategy to decompose person descriptions into attribute-specific sub-sentences (e.g., clothes, pants, shoes, etc.), and rewrite them with MLLM into multiple sentences with the same meaning. Subsequently, the rewritten sentences are randomly reordered and recombined to generate varied augmented texts, thus boosting data diversity and enhancing the generalization of models. Our main contributions are as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 314, + 205, + 553, + 407 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 314, + 205, + 553, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 205, + 553, + 265 + ], + "spans": [ + { + "bbox": [ + 314, + 205, + 553, + 265 + ], + "type": "text", + "content": "- We introduce human-centered interaction to TIREID, proposing a novel MLLM-driven Interactive Cross-modal Learning framework (ICL), which leverages external knowledge to overcome the inherent limitations of existing offline methods in handling dynamic queries." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 265, + 553, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 265, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 314, + 265, + 553, + 312 + ], + "type": "text", + "content": "- A plug-and-play Test-time Human-centered Interaction module (THI) is presented to align query intent with latent target images through multi-round interactions with MLLMs, improving the ranking quality." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 313, + 553, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 313, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 314, + 313, + 553, + 360 + ], + "type": "text", + "content": "- An effective Reorganization Data Augmentation strategy (RDA), applying MLLMs for text decomposition and recombination, is developed to generate discriminative and diverse training texts, improving cross-modal learning." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 361, + 553, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 361, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 314, + 361, + 553, + 407 + ], + "type": "text", + "content": "- Extensive experiments on four text-to-image person re-identification benchmarks verify the effectiveness and superiority of our method, which achieves promising performance and superior generalization." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 422, + 400, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 422, + 400, + 434 + ], + "spans": [ + { + "bbox": [ + 314, + 422, + 400, + 434 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 443, + 518, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 443, + 518, + 455 + ], + "spans": [ + { + "bbox": [ + 314, + 443, + 518, + 455 + ], + "type": "text", + "content": "2.1. Text-to-Image Person Re-identification" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "text", + "content": "As a challenging topic in multimodal learning, text-to-image person re-identification (TIREID) aims to search the image of the target person with a given natural language query. Existing methods [1, 12, 21, 28, 41] can be roughly divided into three categories according to the used backbone type: unimodal backbones, general multimodal pre-trained backbones, and ReID-domain multimodal pre-trained backbones. The early TIREID methods [33, 35] commonly use modality-specific unimodal backbones to encode text or images for cross-modal alignment, e.g., ResNet [10], BERT [30], etc. Recently, benefiting from the rapid development of visual-language pre-training models [14, 24], more and more researchers [3, 12, 21] have begun to use general pre-trained models as the backbone networks for solving TIREID, hoping that the pre-trained alignment knowledge can improve modality representation and cross-modal alignment. However, the performance improvement is still limited due to the domain gap between the general pre-training data and TIREID tasks. To this end, some attempts [28, 38] exploit image-caption models or MLLMs to annotate images, thus obtaining a large number of image" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14391" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 71, + 541, + 252 + ], + "blocks": [ + { + "bbox": [ + 72, + 71, + 541, + 252 + ], + "lines": [ + { + "bbox": [ + 72, + 71, + 541, + 252 + ], + "spans": [ + { + "bbox": [ + 72, + 71, + 541, + 252 + ], + "type": "image", + "image_path": "5486563401a74a428c6b4e71d4a148c3038f7d2f3506da315bc0cd95a87f5da7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 255, + 555, + 310 + ], + "lines": [ + { + "bbox": [ + 55, + 255, + 555, + 310 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 555, + 310 + ], + "type": "text", + "content": "Figure 2. The illustration of our Test-time Human-centered Interaction (THI) module. THI includes " + }, + { + "bbox": [ + 55, + 255, + 555, + 310 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 255, + 555, + 310 + ], + "type": "text", + "content": " rounds of interactions to align query intention with the latent target image by external guidance, where in each round, we perform human-centered visual question answering around fine-grained person attributes to enhance the semantic consistency between the query and the intended person image, and then improve the final ReID performance on the large-scale evaluation through efficient re-ranking. Besides, we perform supervised fine-tuning via LoRA to inspire the discriminate ability of MLLM for ReID domain images and better align queries with latent target images." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 322, + 296, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 322, + 296, + 406 + ], + "spans": [ + { + "bbox": [ + 55, + 322, + 296, + 406 + ], + "type": "text", + "content": "text pairs for ReID-domain pre-training. Although promising performance has been achieved, they still cannot get rid of the inherent defect of offline models during handling dynamic queries. In this paper, we propose a novel interactive cross-modal learning framework based on MLLMs, which exploits dynamic interactions for external guidance to improve the generalization ability." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 418, + 208, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 208, + 431 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 208, + 431 + ], + "type": "text", + "content": "2.2. Visual Interactive Learning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 438, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 438, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 438, + 296, + 713 + ], + "type": "text", + "content": "With the booming development of multimodal large language models (MLLMs) [15, 18, 29], exploiting visual interaction to improve downstream tasks has attracted increasing attention from researchers. We collectively refer to these as visual interaction learning and divide them into offline interaction [28, 31] and online interaction [8, 9, 13] groups by engaging in specific task periods. The former aims to use interactions to improve the training/pre-training data, such as information richness and diversity. It is usually static and separated from users, which is the mainstream of existing TIREID methods, e.g., Tan et al. [28] apply MLLMs to conduct interactions with multiple prompt templates for better diversity. Unlike them, online interactions are designed to be able to use real-time feedback to improve training or testing. For example, Levy et al. [13] propose to gradually clarify user intention through dialogue to improve image retrieval performance. Likewise, Han et al. [8] develop a rerank pipeline based on LLM-based iterative navigation. However, these methods cannot be directly used for TIREID effectively due to the differences in tasks and data domains. In this paper, we develop a test-time module based on MLLMs to conduct interactions for external guidance, overcoming the limitations of the intra-" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 322, + 493, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 322, + 493, + 333 + ], + "spans": [ + { + "bbox": [ + 313, + 322, + 493, + 333 + ], + "type": "text", + "content": "model knowledge in TIReID offline models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 345, + 397, + 359 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 345, + 397, + 359 + ], + "spans": [ + { + "bbox": [ + 313, + 345, + 397, + 359 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 365, + 555, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 555, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 555, + 462 + ], + "type": "text", + "content": "In this section, we introduce an Interactive Cross-modal Learning framework (ICL), which consists of two core comments to address the inherent challenges in offline models and training data, i.e., Test-time Human-centered Interaction module (THI) and Reorganization Data Augmentation (RDA). In Section 3.1, we provide the necessary definitions to facilitate the study. Then, we outline the details of our THI and RDA in Sections 3.2 and 3.3, respectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 469, + 486, + 482 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 469, + 486, + 482 + ], + "spans": [ + { + "bbox": [ + 313, + 469, + 486, + 482 + ], + "type": "text", + "content": "3.1. Problem and Symbol Statement" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 487, + 555, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 555, + 643 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 555, + 643 + ], + "type": "text", + "content": "Suppose that we have the query text " + }, + { + "bbox": [ + 313, + 487, + 555, + 643 + ], + "type": "inline_equation", + "content": "q \\in \\mathcal{Q}" + }, + { + "bbox": [ + 313, + 487, + 555, + 643 + ], + "type": "text", + "content": ", a pedestrian image " + }, + { + "bbox": [ + 313, + 487, + 555, + 643 + ], + "type": "inline_equation", + "content": "v \\in \\mathcal{V}" + }, + { + "bbox": [ + 313, + 487, + 555, + 643 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 487, + 555, + 643 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 313, + 487, + 555, + 643 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 487, + 555, + 643 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 313, + 487, + 555, + 643 + ], + "type": "text", + "content": " are the text query set and image gallery. The purpose of TIReID is to utilize the text query to retrieve the ideal images from the gallery, thus achieving person searching by identities of retrieved images. To achieve this, existing methods usually train an offline cross-modal model " + }, + { + "bbox": [ + 313, + 487, + 555, + 643 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{cross}} = \\{f,g\\}" + }, + { + "bbox": [ + 313, + 487, + 555, + 643 + ], + "type": "text", + "content": " to measure the similarities between text queries and pedestrian images. For a text " + }, + { + "bbox": [ + 313, + 487, + 555, + 643 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 313, + 487, + 555, + 643 + ], + "type": "text", + "content": " and an image " + }, + { + "bbox": [ + 313, + 487, + 555, + 643 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 313, + 487, + 555, + 643 + ], + "type": "text", + "content": ", the similarity to measure the matching degree can be represented as " + }, + { + "bbox": [ + 313, + 487, + 555, + 643 + ], + "type": "inline_equation", + "content": "S_{q,v} \\equiv \\operatorname{Sim}(f(q), g(v))" + }, + { + "bbox": [ + 313, + 487, + 555, + 643 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 487, + 555, + 643 + ], + "type": "inline_equation", + "content": "\\operatorname{Sim}(*)" + }, + { + "bbox": [ + 313, + 487, + 555, + 643 + ], + "type": "text", + "content": " is the similarity function, " + }, + { + "bbox": [ + 313, + 487, + 555, + 643 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 487, + 555, + 643 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 487, + 555, + 643 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 313, + 487, + 555, + 643 + ], + "type": "text", + "content": " are the text and image encoders, respectively. Then, we can utilize " + }, + { + "bbox": [ + 313, + 487, + 555, + 643 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 313, + 487, + 555, + 643 + ], + "type": "text", + "content": " to search relevant candidate images " + }, + { + "bbox": [ + 313, + 487, + 555, + 643 + ], + "type": "inline_equation", + "content": "\\hat{v}" + }, + { + "bbox": [ + 313, + 487, + 555, + 643 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 487, + 555, + 643 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 313, + 487, + 555, + 643 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 356, + 652, + 553, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 652, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 356, + 652, + 553, + 667 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {V}} (q) = \\left\\{\\hat {v} _ {k} \\right\\} _ {k = 1} ^ {K} = \\operatorname {T o p} - K _ {v \\in \\mathcal {V}} \\left(S _ {q, v}\\right), \\tag {1}", + "image_path": "c6653eff2df8bd98a5d1850171f75bb0bbc77dd30d091093116594da23cd2769.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{V}}(q)" + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": " is the candidate set for " + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": " is the number of candidate images. In addition, our ICL also involves MLLMs to conduct interactions and training augmentation," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14392" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "content": "which we denote as " + }, + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "content": ", and the prompt template function for " + }, + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "content": " is denoted by the symbol " + }, + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "content": ". Based on the above definitions, we will elaborate on ICL in the following sections." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 116, + 253, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 116, + 253, + 128 + ], + "spans": [ + { + "bbox": [ + 55, + 116, + 253, + 128 + ], + "type": "text", + "content": "3.2. Test-time Humane-center Interaction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 134, + 296, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 134, + 296, + 278 + ], + "spans": [ + { + "bbox": [ + 55, + 134, + 296, + 278 + ], + "type": "text", + "content": "Anchor Localization. Due to the model bottleneck, it is hard for existing methods to distinguish challenging candidate images relying solely on internal knowledge. Thus, we exploit the fine-grained image understanding capability of MLLMs to conduct external guidance. We first propose asking questions such as \"Can this text accurately describe the image?\" to let MLLMs explicitly tell us the latent target (anchor) image. As shown in Figure 2, given a text query " + }, + { + "bbox": [ + 55, + 134, + 296, + 278 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 55, + 134, + 296, + 278 + ], + "type": "text", + "content": ", we first get the corresponding candidate set " + }, + { + "bbox": [ + 55, + 134, + 296, + 278 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{V}}(q)" + }, + { + "bbox": [ + 55, + 134, + 296, + 278 + ], + "type": "text", + "content": " and then perform multiple " + }, + { + "bbox": [ + 55, + 134, + 296, + 278 + ], + "type": "inline_equation", + "content": "(K)" + }, + { + "bbox": [ + 55, + 134, + 296, + 278 + ], + "type": "text", + "content": " rounds of interactions according to the ranking until the ideal image is determined. For the " + }, + { + "bbox": [ + 55, + 134, + 296, + 278 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 134, + 296, + 278 + ], + "type": "text", + "content": "-th round, the answer for interaction can be represented as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 127, + 287, + 295, + 302 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 287, + 295, + 302 + ], + "spans": [ + { + "bbox": [ + 127, + 287, + 295, + 302 + ], + "type": "interline_equation", + "content": "a _ {\\hat {v} _ {k}} ^ {q} = \\mathcal {M} \\left(\\mathcal {T} _ {\\text {l o c}} (q, \\hat {v} _ {k})\\right), \\tag {2}", + "image_path": "abb889d61020ccc38085c81509537c64503e013ed9d30160fb7a9ab113bed648.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 310, + 296, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 310, + 296, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 310, + 296, + 418 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 310, + 296, + 418 + ], + "type": "inline_equation", + "content": "k \\in \\{1, \\dots, K\\}" + }, + { + "bbox": [ + 55, + 310, + 296, + 418 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 310, + 296, + 418 + ], + "type": "inline_equation", + "content": "a_{\\hat{v}_k}^q" + }, + { + "bbox": [ + 55, + 310, + 296, + 418 + ], + "type": "text", + "content": " is the answer of 'Yes' or 'No', " + }, + { + "bbox": [ + 55, + 310, + 296, + 418 + ], + "type": "inline_equation", + "content": "\\hat{v}_k" + }, + { + "bbox": [ + 55, + 310, + 296, + 418 + ], + "type": "text", + "content": " is the Top-" + }, + { + "bbox": [ + 55, + 310, + 296, + 418 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 310, + 296, + 418 + ], + "type": "text", + "content": " candidate image for query " + }, + { + "bbox": [ + 55, + 310, + 296, + 418 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 55, + 310, + 296, + 418 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 310, + 296, + 418 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{loc}}" + }, + { + "bbox": [ + 55, + 310, + 296, + 418 + ], + "type": "text", + "content": " is the prompt template function for anchor localization. However, since there is a domain gap between ReID images and the generic images used for pre-training or instruction-tuning MLLMs, the answers of MLLMs are often unreliable. To solve this, we exploit LoRA [11] to perform supervised fine-tuning (SFT) on MLLMs to inspire the fine-grained ability to identify the person image and the SFT loss is:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 83, + 427, + 295, + 463 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 427, + 295, + 463 + ], + "spans": [ + { + "bbox": [ + 83, + 427, + 295, + 463 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {l o r a} = - \\sum_ {(x, y) \\in \\mathcal {Z}} \\sum_ {t = 1} ^ {| y |} \\log \\left(p _ {\\mathcal {M}} \\left(y _ {t} \\mid x, y < t\\right)\\right), \\tag {3}", + "image_path": "19ed2f87cab60d0fab84158bb1e2d05a0155eece6130ae460bee12d34e76b7c5.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 472, + 295, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 295, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 295, + 629 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 472, + 295, + 629 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 472, + 295, + 629 + ], + "type": "text", + "content": " is the input Prompt and " + }, + { + "bbox": [ + 55, + 472, + 295, + 629 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 472, + 295, + 629 + ], + "type": "text", + "content": " is the output Response, " + }, + { + "bbox": [ + 55, + 472, + 295, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{Z} = \\{\\mathcal{Z}^{+},\\mathcal{Z}^{-}\\}" + }, + { + "bbox": [ + 55, + 472, + 295, + 629 + ], + "type": "text", + "content": " is the SFT dataset. For " + }, + { + "bbox": [ + 55, + 472, + 295, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}^{+}" + }, + { + "bbox": [ + 55, + 472, + 295, + 629 + ], + "type": "text", + "content": ", we select part of training texts " + }, + { + "bbox": [ + 55, + 472, + 295, + 629 + ], + "type": "inline_equation", + "content": "(\\{q_i\\}_{i = 1}^{N_l})" + }, + { + "bbox": [ + 55, + 472, + 295, + 629 + ], + "type": "text", + "content": " and the corresponding ground-truth images " + }, + { + "bbox": [ + 55, + 472, + 295, + 629 + ], + "type": "inline_equation", + "content": "(\\{v_{i}^{+}\\}_{i = 1}^{N_{l}})" + }, + { + "bbox": [ + 55, + 472, + 295, + 629 + ], + "type": "text", + "content": " from the training set to construct the input Pormpts, i.e., " + }, + { + "bbox": [ + 55, + 472, + 295, + 629 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{T}_{\\mathrm{loc}}(q_i,v_i^+)\\}_{i = 1}^{N_l}" + }, + { + "bbox": [ + 55, + 472, + 295, + 629 + ], + "type": "text", + "content": " expecting MLLMs to output the Response of 'Yes'. Similarly, for " + }, + { + "bbox": [ + 55, + 472, + 295, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}^{-}" + }, + { + "bbox": [ + 55, + 472, + 295, + 629 + ], + "type": "text", + "content": ", we also use the text queries and the corresponding negative images from the training set to construct the input Pormpts, expecting MLLMs to output the Response of 'No'. To make " + }, + { + "bbox": [ + 55, + 472, + 295, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}^{-}" + }, + { + "bbox": [ + 55, + 472, + 295, + 629 + ], + "type": "text", + "content": " more discriminative, we use a TIREID pre-trained cross-modal model to obtain the Top-10 image with different person ID as input negative sample image for each text query by similarity ranking." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": "Human-centered VQA. For each round, once we confirm the response of 'Yes', we will ask the anchor image " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "(\\bar{v})" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": " a series of questions to learn more details about pedestrian characteristics. These details can help us alleviate the discrepancy between the user query and the information within the anchor image, thus improving overall ranking by refining the query. We call this process human-centered visual" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 317, + 69, + 394, + 129 + ], + "blocks": [ + { + "bbox": [ + 317, + 69, + 394, + 129 + ], + "lines": [ + { + "bbox": [ + 317, + 69, + 394, + 129 + ], + "spans": [ + { + "bbox": [ + 317, + 69, + 394, + 129 + ], + "type": "image", + "image_path": "c824de588e1e6a60c0dfc19054f26889e11793083914afd1bc9fc1d7ff33ef1f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 324, + 129, + 386, + 138 + ], + "lines": [ + { + "bbox": [ + 324, + 129, + 386, + 138 + ], + "spans": [ + { + "bbox": [ + 324, + 129, + 386, + 138 + ], + "type": "text", + "content": "(a) CUHK-PEDES" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 142, + 555, + 175 + ], + "lines": [ + { + "bbox": [ + 313, + 142, + 555, + 175 + ], + "spans": [ + { + "bbox": [ + 313, + 142, + 555, + 175 + ], + "type": "text", + "content": "Figure 3. The similarity statistics of the Top-1 retrieved items in the test sets of the three benchmarks. It is obvious that the items with higher similarity are more likely to be correct retrievals." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 394, + 70, + 474, + 128 + ], + "blocks": [ + { + "bbox": [ + 394, + 70, + 474, + 128 + ], + "lines": [ + { + "bbox": [ + 394, + 70, + 474, + 128 + ], + "spans": [ + { + "bbox": [ + 394, + 70, + 474, + 128 + ], + "type": "image", + "image_path": "f5e32f4be4afd3a7078b77138f1c5a39b731211571d68d7a7759261c686935bd.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 405, + 129, + 463, + 138 + ], + "lines": [ + { + "bbox": [ + 405, + 129, + 463, + 138 + ], + "spans": [ + { + "bbox": [ + 405, + 129, + 463, + 138 + ], + "type": "text", + "content": "(b) ICFG-PEDES" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 474, + 70, + 554, + 129 + ], + "blocks": [ + { + "bbox": [ + 474, + 70, + 554, + 129 + ], + "lines": [ + { + "bbox": [ + 474, + 70, + 554, + 129 + ], + "spans": [ + { + "bbox": [ + 474, + 70, + 554, + 129 + ], + "type": "image", + "image_path": "6a7d7a9a7a350b03bdf20481d1c0a06d7f9d22c6199507201d610b14c37eee61.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 491, + 129, + 537, + 138 + ], + "lines": [ + { + "bbox": [ + 491, + 129, + 537, + 138 + ], + "spans": [ + { + "bbox": [ + 491, + 129, + 537, + 138 + ], + "type": "text", + "content": "(c) RSTPReid" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 185, + 520, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 185, + 520, + 198 + ], + "spans": [ + { + "bbox": [ + 313, + 185, + 520, + 198 + ], + "type": "text", + "content": "question answering (VQA), which is expressed as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 377, + 202, + 553, + 217 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 202, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 377, + 202, + 553, + 217 + ], + "type": "interline_equation", + "content": "r _ {\\bar {v}} = \\mathcal {M} \\left(\\mathcal {T} _ {\\mathrm {v q a}} \\left(\\left\\{c _ {i} \\right\\} _ {i = 1} ^ {N _ {q}}, \\bar {v}\\right)\\right), \\tag {4}", + "image_path": "770fdb3c6d8d4d524d91b5e23ebd1d32b324ac86d7d8d2259c6bea50f673d0fd.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 224, + 555, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 224, + 555, + 370 + ], + "spans": [ + { + "bbox": [ + 313, + 224, + 555, + 370 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 224, + 555, + 370 + ], + "type": "inline_equation", + "content": "\\{c_i\\}_{i=1}^{N_q}" + }, + { + "bbox": [ + 313, + 224, + 555, + 370 + ], + "type": "text", + "content": " are " + }, + { + "bbox": [ + 313, + 224, + 555, + 370 + ], + "type": "inline_equation", + "content": "N_q" + }, + { + "bbox": [ + 313, + 224, + 555, + 370 + ], + "type": "text", + "content": " questions directed at the detail characteristics (e.g., gender, hair, upper body, lower body, shoes, etc.) of the pedestrian image, " + }, + { + "bbox": [ + 313, + 224, + 555, + 370 + ], + "type": "inline_equation", + "content": "r_{\\bar{v}}" + }, + { + "bbox": [ + 313, + 224, + 555, + 370 + ], + "type": "text", + "content": " is the answer to " + }, + { + "bbox": [ + 313, + 224, + 555, + 370 + ], + "type": "inline_equation", + "content": "\\{c_i\\}_{i=1}^{N_q}" + }, + { + "bbox": [ + 313, + 224, + 555, + 370 + ], + "type": "text", + "content": " one by one, and " + }, + { + "bbox": [ + 313, + 224, + 555, + 370 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{vqa}}" + }, + { + "bbox": [ + 313, + 224, + 555, + 370 + ], + "type": "text", + "content": " is the prompt template function for human-centered VQA. To improve the consistency between the image and the user query, a natural strategy is to concatenate these one-by-one answers after the original text. However, due to the limitation of the maximum text length that can be processed by the text model (e.g., it is usually 77 for CLIP), this will destroy the sequence structure. To this end, we recommend merging these answers and the original text by MLLMs to obtain a fluent and concise text, i.e.," + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 389, + 376, + 553, + 389 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 376, + 553, + 389 + ], + "spans": [ + { + "bbox": [ + 389, + 376, + 553, + 389 + ], + "type": "interline_equation", + "content": "\\hat {q} = \\mathcal {M} \\left(\\mathcal {T} _ {\\text {a g g r}} \\left(r _ {\\bar {v}}, q\\right)\\right), \\tag {5}", + "image_path": "8c3137305115068bff413cbdba78889ff17b2f5311514042a3f668405a253c9c.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 395, + 554, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 395, + 554, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 395, + 554, + 418 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 395, + 554, + 418 + ], + "type": "inline_equation", + "content": "\\hat{q}" + }, + { + "bbox": [ + 313, + 395, + 554, + 418 + ], + "type": "text", + "content": " is the merged text query and " + }, + { + "bbox": [ + 313, + 395, + 554, + 418 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{aggr}}" + }, + { + "bbox": [ + 313, + 395, + 554, + 418 + ], + "type": "text", + "content": " is the prompt template function for text aggregation." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 419, + 555, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 419, + 555, + 694 + ], + "spans": [ + { + "bbox": [ + 313, + 419, + 555, + 694 + ], + "type": "text", + "content": "Efficient Re-ranking. To apply the above process to large-scale evaluation, efficiency is a factor that needs to be considered in the retrieval task, since introducing MLLMs to interactions for external guidance inevitably brings additional inference costs. To improve the interaction efficiency, we recommended adopting different strategies for different rounds of interactions. We observe that the retrieval quality is positively correlated with the similarity between the query and the Top-1 candidate image, as shown in Figure 3. To this end, in the first round, we only interact with images whose cross-modal similarity is greater than a threshold " + }, + { + "bbox": [ + 313, + 419, + 555, + 694 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 313, + 419, + 555, + 694 + ], + "type": "text", + "content": " and the answer of anchor localization is 'Yes'. This is because the retrievals with low similarities are more likely to be wrong items. After the first round, we only interact with the image that is all considered 'No' in the previous rounds of anchor localization and whose cross-modal similarity is less than the threshold " + }, + { + "bbox": [ + 313, + 419, + 555, + 694 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 313, + 419, + 555, + 694 + ], + "type": "text", + "content": ". This allows us to reduce unnecessary interactions and find text queries that really require interactions, making our method more efficient on large-scale evaluation. Given a query " + }, + { + "bbox": [ + 313, + 419, + 555, + 694 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 313, + 419, + 555, + 694 + ], + "type": "text", + "content": ", if the above constraints are met, the interaction is carried out and we can get merged text " + }, + { + "bbox": [ + 313, + 419, + 555, + 694 + ], + "type": "inline_equation", + "content": "\\hat{q}" + }, + { + "bbox": [ + 313, + 419, + 555, + 694 + ], + "type": "text", + "content": " after human-centered VQA, the re-ranking similarity " + }, + { + "bbox": [ + 313, + 419, + 555, + 694 + ], + "type": "inline_equation", + "content": "\\hat{S}_{q,v}" + }, + { + "bbox": [ + 313, + 419, + 555, + 694 + ], + "type": "text", + "content": " of any image " + }, + { + "bbox": [ + 313, + 419, + 555, + 694 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 313, + 419, + 555, + 694 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 313, + 419, + 555, + 694 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 313, + 419, + 555, + 694 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 374, + 700, + 553, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 374, + 700, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 374, + 700, + 553, + 715 + ], + "type": "interline_equation", + "content": "\\hat {S} _ {q, v} = \\lambda S _ {q, v} + (1 - \\lambda) \\bar {S} _ {\\hat {q}, v}, \\tag {6}", + "image_path": "875eb56aa48c231d6513fdc98fb2283ee86307a658fafd9e2bd499aec2932a32.jpg" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14393" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 296, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 296, + 145 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 71, + 296, + 145 + ], + "type": "inline_equation", + "content": "\\bar{S}_{\\hat{q},v} \\equiv 1" + }, + { + "bbox": [ + 55, + 71, + 296, + 145 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 55, + 71, + 296, + 145 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 55, + 71, + 296, + 145 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 55, + 71, + 296, + 145 + ], + "type": "inline_equation", + "content": "\\hat{v}_1" + }, + { + "bbox": [ + 55, + 71, + 296, + 145 + ], + "type": "text", + "content": ", otherwise, " + }, + { + "bbox": [ + 55, + 71, + 296, + 145 + ], + "type": "inline_equation", + "content": "\\bar{S}_{\\hat{q},v} \\equiv S_{\\hat{q},v}" + }, + { + "bbox": [ + 55, + 71, + 296, + 145 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 71, + 296, + 145 + ], + "type": "inline_equation", + "content": "\\lambda \\in [0,1]" + }, + { + "bbox": [ + 55, + 71, + 296, + 145 + ], + "type": "text", + "content": " is a hyperparameter to balance the contribution of the raw query and the refined text. To make our approach clearer, we describe the detailed algorithm process of our THI in Algorithm 1. Due to space limitations, all prompt templates can be found in the supplementary material." + } + ] + } + ], + "index": 0 + }, + { + "type": "code", + "bbox": [ + 56, + 170, + 296, + 506 + ], + "blocks": [ + { + "bbox": [ + 56, + 155, + 251, + 167 + ], + "lines": [ + { + "bbox": [ + 56, + 155, + 251, + 167 + ], + "spans": [ + { + "bbox": [ + 56, + 155, + 251, + 167 + ], + "type": "text", + "content": "Algorithm 1 The interaction process of our THI" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 56, + 170, + 296, + 506 + ], + "lines": [ + { + "bbox": [ + 56, + 170, + 296, + 506 + ], + "spans": [ + { + "bbox": [ + 56, + 170, + 296, + 506 + ], + "type": "text", + "content": "Input: The query set " + }, + { + "bbox": [ + 56, + 170, + 296, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 56, + 170, + 296, + 506 + ], + "type": "text", + "content": ", the image gallery " + }, + { + "bbox": [ + 56, + 170, + 296, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 56, + 170, + 296, + 506 + ], + "type": "text", + "content": ", the offline model " + }, + { + "bbox": [ + 56, + 170, + 296, + 506 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{cross}}" + }, + { + "bbox": [ + 56, + 170, + 296, + 506 + ], + "type": "text", + "content": ", the MLLM " + }, + { + "bbox": [ + 56, + 170, + 296, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 56, + 170, + 296, + 506 + ], + "type": "text", + "content": ", the similarity threshold " + }, + { + "bbox": [ + 56, + 170, + 296, + 506 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 56, + 170, + 296, + 506 + ], + "type": "text", + "content": ", the number of interaction rounds " + }, + { + "bbox": [ + 56, + 170, + 296, + 506 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 56, + 170, + 296, + 506 + ], + "type": "text", + "content": "; \n1: Obtain candidate sets " + }, + { + "bbox": [ + 56, + 170, + 296, + 506 + ], + "type": "inline_equation", + "content": "\\{\\hat{\\mathcal{V}}(q_i)\\}_{i=1}^{|\\mathcal{Q}|}" + }, + { + "bbox": [ + 56, + 170, + 296, + 506 + ], + "type": "text", + "content": " for all queries in " + }, + { + "bbox": [ + 56, + 170, + 296, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 56, + 170, + 296, + 506 + ], + "type": "text", + "content": " via Equation (1); \n2: for " + }, + { + "bbox": [ + 56, + 170, + 296, + 506 + ], + "type": "inline_equation", + "content": "k = 1,2,\\dots,K" + }, + { + "bbox": [ + 56, + 170, + 296, + 506 + ], + "type": "text", + "content": " do \n3: for " + }, + { + "bbox": [ + 56, + 170, + 296, + 506 + ], + "type": "inline_equation", + "content": "i = 1,2,\\dots,|\\mathcal{Q}|" + }, + { + "bbox": [ + 56, + 170, + 296, + 506 + ], + "type": "text", + "content": " do \n4: Conduct anchor localization via Equation (2) and output the answer of " + }, + { + "bbox": [ + 56, + 170, + 296, + 506 + ], + "type": "inline_equation", + "content": "a_{\\hat{v}_k^i}^{q_i}" + }, + { + "bbox": [ + 56, + 170, + 296, + 506 + ], + "type": "text", + "content": " based on the " + }, + { + "bbox": [ + 56, + 170, + 296, + 506 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 56, + 170, + 296, + 506 + ], + "type": "text", + "content": "-th candidate image " + }, + { + "bbox": [ + 56, + 170, + 296, + 506 + ], + "type": "inline_equation", + "content": "\\hat{v}_k^i" + }, + { + "bbox": [ + 56, + 170, + 296, + 506 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 56, + 170, + 296, + 506 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{V}}(q_i)" + }, + { + "bbox": [ + 56, + 170, + 296, + 506 + ], + "type": "text", + "content": "; \n5: if " + }, + { + "bbox": [ + 56, + 170, + 296, + 506 + ], + "type": "inline_equation", + "content": "a_{\\hat{v}_k^i}^{q_i}" + }, + { + "bbox": [ + 56, + 170, + 296, + 506 + ], + "type": "text", + "content": " shows 'Yes', " + }, + { + "bbox": [ + 56, + 170, + 296, + 506 + ], + "type": "inline_equation", + "content": "k = 1" + }, + { + "bbox": [ + 56, + 170, + 296, + 506 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 56, + 170, + 296, + 506 + ], + "type": "inline_equation", + "content": "S_{q_i,\\hat{v}_1^i} > \\xi" + }, + { + "bbox": [ + 56, + 170, + 296, + 506 + ], + "type": "text", + "content": " then \n6: Conduct human-centered VQA via Equations (4) and (5) to get the refined query " + }, + { + "bbox": [ + 56, + 170, + 296, + 506 + ], + "type": "inline_equation", + "content": "\\hat{q}_i" + }, + { + "bbox": [ + 56, + 170, + 296, + 506 + ], + "type": "text", + "content": "; \n7: Compute the re-ranking similarities between query " + }, + { + "bbox": [ + 56, + 170, + 296, + 506 + ], + "type": "inline_equation", + "content": "q_i" + }, + { + "bbox": [ + 56, + 170, + 296, + 506 + ], + "type": "text", + "content": " and all images via Equation (6); \n8: end if \n9: if " + }, + { + "bbox": [ + 56, + 170, + 296, + 506 + ], + "type": "inline_equation", + "content": "\\{a_{\\hat{v}_j^i}\\}_{j=1}^{k-1}" + }, + { + "bbox": [ + 56, + 170, + 296, + 506 + ], + "type": "text", + "content": " all show 'No', " + }, + { + "bbox": [ + 56, + 170, + 296, + 506 + ], + "type": "inline_equation", + "content": "a_{\\hat{v}_k^i}^{q_i}" + }, + { + "bbox": [ + 56, + 170, + 296, + 506 + ], + "type": "text", + "content": " shows 'Yes', " + }, + { + "bbox": [ + 56, + 170, + 296, + 506 + ], + "type": "inline_equation", + "content": "k > 1" + }, + { + "bbox": [ + 56, + 170, + 296, + 506 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 56, + 170, + 296, + 506 + ], + "type": "inline_equation", + "content": "S_{q_i,\\hat{v}_1^i} \\leq \\xi" + }, + { + "bbox": [ + 56, + 170, + 296, + 506 + ], + "type": "text", + "content": " then \n10: Conduct human-centered VQA via Equations (4) and (5) to get the refined query " + }, + { + "bbox": [ + 56, + 170, + 296, + 506 + ], + "type": "inline_equation", + "content": "\\hat{q}_i" + }, + { + "bbox": [ + 56, + 170, + 296, + 506 + ], + "type": "text", + "content": "; \n11: Compute the re-ranking similarities between query " + }, + { + "bbox": [ + 56, + 170, + 296, + 506 + ], + "type": "inline_equation", + "content": "q_i" + }, + { + "bbox": [ + 56, + 170, + 296, + 506 + ], + "type": "text", + "content": " and all images via Equation (6); \n12: end if \n13: end for \n14: end for \n15: Re-ranking based on similarities; \nOutput: The new candidate images." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_body" + } + ], + "index": 2, + "sub_type": "algorithm" + }, + { + "bbox": [ + 55, + 526, + 244, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 526, + 244, + 539 + ], + "spans": [ + { + "bbox": [ + 55, + 526, + 244, + 539 + ], + "type": "text", + "content": "3.3. Reorganization Data Augmentation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 544, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 296, + 713 + ], + "type": "text", + "content": "Although THI can exploit interactions to provide external knowledge during test time to improve retrieval quality, the cross-modal embedding model is still a bottleneck that limits its further performance improvement. To this end, as shown in Figure 4, we introduce a new Reorganization Data Augmentation (RDA) strategy by enriching, decomposing, and reorganizing person descriptions to improve the discriminability and diversity of training data. We first apply Human-centered VQA in THI to obtain relevant characteristic descriptions and merge them with the original text, thus obtaining a text with richer information, i.e., " + }, + { + "bbox": [ + 55, + 544, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\hat{q} = \\mathcal{M}(\\mathcal{T}_{\\mathrm{aggr}}(\\mathcal{M}(\\mathcal{T}_{\\mathrm{vqa}}(\\{c_i\\}_{i=1}^{N_q}, v)), q))" + }, + { + "bbox": [ + 55, + 544, + 296, + 713 + ], + "type": "text", + "content": ". Then, we apply MLLMs to decompose enriched text " + }, + { + "bbox": [ + 55, + 544, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\hat{q}" + }, + { + "bbox": [ + 55, + 544, + 296, + 713 + ], + "type": "text", + "content": " into multiple independent sub-sentences that describe individual at-" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 317, + 70, + 555, + 194 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 555, + 194 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 555, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 555, + 194 + ], + "type": "image", + "image_path": "f12d5a579f973bb0872b8ae0e4528bf3b9a1fe6a3b6fd2b5de03f9b475a34af2.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 197, + 555, + 253 + ], + "lines": [ + { + "bbox": [ + 313, + 197, + 555, + 253 + ], + "spans": [ + { + "bbox": [ + 313, + 197, + 555, + 253 + ], + "type": "text", + "content": "Figure 4. The illustration of our RDA. The purpose of RDA is to supplement more details to the original training texts through human-centered VQA, improving the discriminability of texts. In addition, to enhance diversity, RDA maximizes diversity through the Decomposition-Rewriting-Reorganization strategy." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 273, + 555, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 273, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 273, + 555, + 392 + ], + "type": "text", + "content": "tributes without interfering with each other, i.e., " + }, + { + "bbox": [ + 313, + 273, + 555, + 392 + ], + "type": "inline_equation", + "content": "\\{\\tilde{q}_i\\}_{i=1}^n = \\mathcal{M}(\\mathcal{T}_{\\mathrm{dec}}(\\hat{q}))" + }, + { + "bbox": [ + 313, + 273, + 555, + 392 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 273, + 555, + 392 + ], + "type": "inline_equation", + "content": "\\{\\tilde{q}_i\\}_{i=1}^n" + }, + { + "bbox": [ + 313, + 273, + 555, + 392 + ], + "type": "text", + "content": " are the " + }, + { + "bbox": [ + 313, + 273, + 555, + 392 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 273, + 555, + 392 + ], + "type": "text", + "content": " sub-sentences and " + }, + { + "bbox": [ + 313, + 273, + 555, + 392 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{dec}}" + }, + { + "bbox": [ + 313, + 273, + 555, + 392 + ], + "type": "text", + "content": " is the prompt template function for text decomposition. The purpose of decomposition is to be able to reorganize the text in a different order later. To increase diversity, we rewrite each sub-sentence into multiple sentences with different styles but the same meaning. For sub-sentence " + }, + { + "bbox": [ + 313, + 273, + 555, + 392 + ], + "type": "inline_equation", + "content": "\\tilde{q}" + }, + { + "bbox": [ + 313, + 273, + 555, + 392 + ], + "type": "text", + "content": ", we conduct rewriting by " + }, + { + "bbox": [ + 313, + 273, + 555, + 392 + ], + "type": "inline_equation", + "content": "\\mathcal{R} = \\{\\check{q}_j\\}_{j=1}^m = \\mathcal{M}(\\mathcal{T}_{\\mathrm{rwt}}(\\tilde{q}))" + }, + { + "bbox": [ + 313, + 273, + 555, + 392 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 273, + 555, + 392 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 313, + 273, + 555, + 392 + ], + "type": "text", + "content": " is the set with " + }, + { + "bbox": [ + 313, + 273, + 555, + 392 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 273, + 555, + 392 + ], + "type": "text", + "content": " rewritten sentences and " + }, + { + "bbox": [ + 313, + 273, + 555, + 392 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{rwt}}" + }, + { + "bbox": [ + 313, + 273, + 555, + 392 + ], + "type": "text", + "content": " is the prompt template function for text rewriting." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 393, + 556, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 393, + 556, + 501 + ], + "spans": [ + { + "bbox": [ + 313, + 393, + 556, + 501 + ], + "type": "text", + "content": "Finally, for a text " + }, + { + "bbox": [ + 313, + 393, + 556, + 501 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 313, + 393, + 556, + 501 + ], + "type": "text", + "content": ", we can obtain the set " + }, + { + "bbox": [ + 313, + 393, + 556, + 501 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{R}_i\\}_{i=1}^n" + }, + { + "bbox": [ + 313, + 393, + 556, + 501 + ], + "type": "text", + "content": " that contains a large number of style sub-sentences. So far, we can obtain augmented texts with different sub-sentence orders and different style combinations as shown in Figure 4, which we represent as " + }, + { + "bbox": [ + 313, + 393, + 556, + 501 + ], + "type": "inline_equation", + "content": "\\check{q}" + }, + { + "bbox": [ + 313, + 393, + 556, + 501 + ], + "type": "text", + "content": ". During training, we can mix the augmentation texts " + }, + { + "bbox": [ + 313, + 393, + 556, + 501 + ], + "type": "inline_equation", + "content": "(\\check{q})" + }, + { + "bbox": [ + 313, + 393, + 556, + 501 + ], + "type": "text", + "content": " with the original texts " + }, + { + "bbox": [ + 313, + 393, + 556, + 501 + ], + "type": "inline_equation", + "content": "(q)" + }, + { + "bbox": [ + 313, + 393, + 556, + 501 + ], + "type": "text", + "content": " for cross-modal learning, thereby improving model-intra knowledge and generalization. Due to space limitations, more training details can be found in the supplementary material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 511, + 395, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 511, + 395, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 511, + 395, + 525 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 532, + 555, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 532, + 555, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 532, + 555, + 568 + ], + "type": "text", + "content": "In this section, we conduct extensive experiments to verify the effectiveness, superiority, and generalization of the proposed ICL on four public benchmark datasets." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 575, + 497, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 575, + 497, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 575, + 497, + 586 + ], + "type": "text", + "content": "4.1. Datasets and Evaluation Protocols" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 594, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 689 + ], + "type": "text", + "content": "Datasets. In our experiments, we use three coarse-grained benchmarks, CHUK-PEDES [16], ICFG-PEDES [5], and RSTPReid [41], and one fine-grained benchmark, UFine6926 [44], to evaluate our ICL. Compared with the coarse-grained benchmark, the fine-grained benchmark has richer texts with fine-grained information. For all datasets, we follow their official settings for data partitioning. More details are provided in the supplementary materials." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "Evaluation Protocols. To measure performance, we utilize the widely accepted Rank-K (1,5,10) metrics to mea" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14394" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 70, + 553, + 261 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 553, + 261 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 553, + 261 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 553, + 261 + ], + "type": "table", + "html": "
MethodsImage Enc.Text Enc.CUHK-PEDESICFG-PEDESRSTPReid
Rank-1Rank-5Rank-10mAPmINPRank-1Rank-5Rank-10mAPmINPRank-1Rank-5Rank-10mAPmINP
① VL-Backbones w/o ReID-domain pre-training
IVT [26]ViT-BaseBERT65.6985.9391.1560.66-56.0473.6080.22--46.7070.0078.80--
\\( LCR^2S \\) [36]RN50BERT67.3684.1989.6259.20-57.9376.0882.4038.21-54.9576.6584.7040.92-
CFine [37]CLIP-ViTBERT69.5785.9391.15--60.8376.5582.42--50.5572.5081.60--
RaSa [1]Swin-BBERT76.5190.2994.2569.38-65.2880.4085.1241.29-66.9086.5091.3552.31-
IRRA [12]CLIP-ViTCLIP-X.73.3889.9393.7166.1350.2463.4680.2585.8238.067.9360.2081.3088.2047.1725.28
TBPS [3]CLIP-ViTCLIP-X.73.5488.1992.3565.3849.2565.0580.3485.4739.837.8762.1081.9087.7548.0025.86
CFAM [44]CLIP-ViTCLIP-X.75.6090.5394.3667.27-65.3881.1786.3539.42-62.4583.5591.1049.50-
RDE [21]CLIP-ViTCLIP-X.75.9490.1494.1267.5651.4467.6882.4787.3640.067.8765.3583.9589.9050.8828.08
Our ICLCLIP-ViTCLIP-X.76.4190.4894.3368.0451.9968.1182.5987.5240.818.1867.7086.0591.7552.6229.36
Our ICL*CLIP-ViTCLIP-X.77.9190.2794.1469.1353.4069.0282.4587.3641.218.3070.5585.9591.6553.6830.13
② VL-Backbones with ReID-domain pre-training
\\( IRRA^b \\) [12]CLIP-ViTCLIP-X.74.0589.4893.6466.57-64.3780.7586.1238.85-61.9080.6089.3048.08-
APTM [38]Swin-BBERT76.5390.0494.1566.91-68.5182.9987.5641.22-67.5085.7091.4552.56-
\\( NAM^d \\) [28]CLIP-ViTCLIP-X.77.4790.8494.6769.4354.0866.7682.0287.1741.459.5367.1586.5591.9052.0028.46
Our ICLCLIP-ViTCLIP-X.78.1891.6394.8369.5853.4869.2283.4988.0642.349.0170.0086.6091.7054.1630.93
Our ICL*CLIP-ViTCLIP-X.79.0691.2694.7270.4454.7070.0583.3587.9142.709.1372.5586.6091.3055.1931.72
", + "image_path": "ad537603959926ce02e4e5749bd0543607315691337d3b32b9aa66e6e03118dd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 60, + 297, + 293, + 407 + ], + "blocks": [ + { + "bbox": [ + 55, + 264, + 555, + 288 + ], + "lines": [ + { + "bbox": [ + 55, + 264, + 555, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 264, + 555, + 288 + ], + "type": "text", + "content": "Table 1. Performance on the three coarse-grained benchmarks. The results with THI are marked with " + }, + { + "bbox": [ + 55, + 264, + 555, + 288 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 55, + 264, + 555, + 288 + ], + "type": "text", + "content": ". Note that IRRA means using the pre-trained Backbones with MALS [38] and the results of NAM are reproduced by us." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 297, + 293, + 407 + ], + "lines": [ + { + "bbox": [ + 60, + 297, + 293, + 407 + ], + "spans": [ + { + "bbox": [ + 60, + 297, + 293, + 407 + ], + "type": "table", + "html": "
MethodsRank-1Rank-5Rank-10mAPmINP
LGUR [25]70.6984.5789.9168.93-
SSAN [5]75.0988.6392.8473.14-
IRRA [12]85.0294.3196.7583.9177.30
RDE [21]87.6095.6597.4686.1079.54
CFAM(B/16) [44]85.5594.5197.0284.23-
CFAM(L/14) [44]88.5195.5897.4987.09-
1 Our ICL89.1796.1397.8887.4981.50
1 Our ICL★90.6795.9897.8688.2982.60
2 Our ICL91.0296.9898.1789.7684.70
2 Our ICL★91.7896.8398.1690.3385.62
", + "image_path": "b4fe73b5e24de3e9197a2e76a0b52305bb50c2aed128791fdcec14e62d8763f7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 409, + 295, + 431 + ], + "lines": [ + { + "bbox": [ + 55, + 409, + 295, + 431 + ], + "spans": [ + { + "bbox": [ + 55, + 409, + 295, + 431 + ], + "type": "text", + "content": "Table 2. Performance comparison on the UFine6926 dataset. The results of IRRA and RDE are reproduced by us." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 439, + 296, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 439, + 296, + 477 + ], + "spans": [ + { + "bbox": [ + 55, + 439, + 296, + 477 + ], + "type": "text", + "content": "sure the TIREID performance. Like [12, 21], we also report the mean Average Precision (mAP) and mean Inverse Negative Penalty (mINP) as auxiliary metrics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 489, + 189, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 489, + 189, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 489, + 189, + 502 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 510, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 510, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 510, + 297, + 713 + ], + "type": "text", + "content": "To achieve interactive TIReID, we choose RDE [21] and Qwen2-VL-7B-Instruct [29] as the investigated TIReID method and MLLM. For the sake of fairness, we do not modify any settings of RDE, including model architecture (CLIP-ViTB/16 [24]), hyperparameters, and training parameters. To be compatible with the fine-grained benchmark, following [44], the maximum length of the textual tokens of CLIP is set to 168 by interpolating the positional embedding layer with an initial learning rate of " + }, + { + "bbox": [ + 54, + 510, + 297, + 713 + ], + "type": "inline_equation", + "content": "5e - 5" + }, + { + "bbox": [ + 54, + 510, + 297, + 713 + ], + "type": "text", + "content": ". As for fine-tuning MLLMs, we use the Llama-Factory framework [40] to conduct SFT with LoRA [11] for 2 epochs. The train batch size per device is set to 4, the gradient accumulation steps are set to 16, the initial learning rate is " + }, + { + "bbox": [ + 54, + 510, + 297, + 713 + ], + "type": "inline_equation", + "content": "5e - 5" + }, + { + "bbox": [ + 54, + 510, + 297, + 713 + ], + "type": "text", + "content": " and the hyperparameters " + }, + { + "bbox": [ + 54, + 510, + 297, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 54, + 510, + 297, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 510, + 297, + 713 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 54, + 510, + 297, + 713 + ], + "type": "text", + "content": " of LoRA are set to 16 and 8, respectively. During inference, the temperature is set to 0.01 to keep reproducibility. Note that all experiments can be completed on two GeForce RTX 24GB 3090 GPUs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 299, + 500, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 299, + 500, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 299, + 500, + 312 + ], + "type": "text", + "content": "4.3. Comparison with State-of-the-Arts" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 318, + 555, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 318, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 318, + 555, + 426 + ], + "type": "text", + "content": "In this section, to verify the superiority of our ICL, we compare our method with more than 15 baselines including recent advanced methods (e.g., RDE (CVPR'24) [21] and NAM (CVPR'24) [28]). Based on the backbone type, we divide the baselines into two groups (1 and 2) as shown Table 1, i.e., the baselines using VL-Backbones w/o and with ReID-domain pre-training. For the group 2, we use the pretrained weights released by [28] to initialize CLIP for fair comparison. The results are reported in Tables 1 and 2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 426, + 556, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 556, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 556, + 641 + ], + "type": "text", + "content": "Results on coarse-grained datasets. Table 1 report the results evaluated on the coarse-grained datasets. We can see that our method can achieve competitive performance even without THI. By performing THI, the Rank-1 score of our method is greatly improved, e.g., in the group ①, the Rank-1 scores on the three datasets are improved by " + }, + { + "bbox": [ + 313, + 426, + 556, + 641 + ], + "type": "inline_equation", + "content": "1.50\\%" + }, + { + "bbox": [ + 313, + 426, + 556, + 641 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 426, + 556, + 641 + ], + "type": "inline_equation", + "content": "0.91\\%" + }, + { + "bbox": [ + 313, + 426, + 556, + 641 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 426, + 556, + 641 + ], + "type": "inline_equation", + "content": "2.85\\%" + }, + { + "bbox": [ + 313, + 426, + 556, + 641 + ], + "type": "text", + "content": ", respectively. In addition, mAP and mINP scores have also improved greatly, which indicates that the overall ranking has improved. In group ②, our method achieves the best scores on most metrics, especially Rank-1 reached " + }, + { + "bbox": [ + 313, + 426, + 556, + 641 + ], + "type": "inline_equation", + "content": "72.55\\%" + }, + { + "bbox": [ + 313, + 426, + 556, + 641 + ], + "type": "text", + "content": " on RSTPReid, which is sufficient to verify the superiority. However, we found that THI slightly degraded Rank-5 and Rank-10, which is because we only conducted 5 rounds of interaction. As long as the locating anchor image is incorrect, the errors will accumulate, which can be relieved by increasing the rounds of interaction. But in general, THI can significantly improve Rank-1 and the overall ranking (mAP and mINP)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "Results on fine-grained dataset. Table 2 shows the results on the fine-grained dataset whose average text length is over 80. Such kind of fine-grained description often has a clear query intention. For better comparison, we provide the performance of IRRA and RDE on the UFine6926 dataset. From the results, our method can still achieve excellent per" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14395" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 70, + 553, + 215 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 553, + 215 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 553, + 215 + ], + "type": "table", + "html": "
MethodsTraining SetsCUHK-PEDESICFG-PEDESRSTPReid
Rank-1Rank-5Rank-10mAPmINPRank-1Rank-5Rank-10mAPmINPRank-1Rank-5Rank-10mAPmINP
IRRA [12]CUHK-PEDES73.3889.9393.7166.1350.2442.4162.1169.6221.771.9553.2577.1585.3539.6316.60
ICFG-PEDES33.4856.2966.3331.5619.2063.4680.2585.8238.067.9345.3069.2578.8036.8218.38
RSTPReid32.8055.2665.8130.2917.6132.3049.6757.8020.543.8460.2081.3088.2047.1725.28
RDE [21]CUHK-PEDES75.9490.1494.1267.5651.4448.1866.3073.7025.002.3354.9077.5086.5041.2717.84
ICFG-PEDES38.1159.2468.4434.1620.4467.6882.4787.3640.067.8749.2572.1080.2038.4618.33
RSTPReid36.9458.2267.5833.6520.4242.1758.3265.4926.374.9465.3583.9589.9050.8828.08
Our ICLCUHK-PEDES76.4190.4894.3368.0451.9948.5766.6673.7525.302.4055.8079.6087.6542.0917.41
ICFG-PEDES42.8764.2073.4438.1923.5868.1182.5987.5240.818.1852.5075.0583.0041.8221.14
RSTPReid41.3161.8670.3136.7822.3745.9362.7068.8028.895.6367.7086.0591.7552.6229.36
Our ICL*CUHK-PEDES77.9190.2794.1469.1353.4052.8066.4973.4925.602.4461.3079.2587.4043.4218.01
ICFG-PEDES49.2964.3473.5540.8225.3869.0282.4587.3641.218.3060.1575.3083.1543.7222.04
RSTPReid47.3561.4570.3438.9123.6850.5261.5668.5729.265.7370.5585.9591.6553.6830.13
", + "image_path": "0b79a12b6ffe274b85c93fea257ab4a3ef30dcfc5f0e716faffa6089522fbcb6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 60, + 247, + 295, + 498 + ], + "blocks": [ + { + "bbox": [ + 138, + 218, + 470, + 229 + ], + "lines": [ + { + "bbox": [ + 138, + 218, + 470, + 229 + ], + "spans": [ + { + "bbox": [ + 138, + 218, + 470, + 229 + ], + "type": "text", + "content": "Table 3. Comparison of mutual generalization capabilities between coarse-grained datasets." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 247, + 295, + 498 + ], + "lines": [ + { + "bbox": [ + 60, + 247, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 60, + 247, + 295, + 498 + ], + "type": "table", + "html": "
Source → TargetMethodsRank-1Rank-5Rank-10mAPmINP
CUHK. → UFine.IRRA [12]37.5154.9264.2940.7634.33
RDE [21]40.3757.4966.0542.6835.78
Our ICL46.4063.5572.0848.6841.56
Our ICL☆57.7664.1372.8153.9745.64
ICFG. → UFine.IRRA [12]15.0226.7933.9017.1012.75
RDE [21]17.8631.0138.5619.8214.74
Our ICL27.9544.2052.2029.8523.20
Our ICL☆36.8144.6552.7334.1226.61
RSTP. → UFine.IRRA [12]13.2125.6733.9315.6011.09
RDE [21]14.0025.2332.6416.2211.90
Our ICL23.8938.3046.7025.5419.20
Our ICL☆31.2338.5647.0228.9021.80
UFine. → CUHKIRRA [12]37.7460.1270.1335.9423.21
RDE [21]39.4161.1470.1136.4923.32
Our ICL49.0470.2778.6444.5429.58
Our ICL☆56.8770.1978.5347.3131.20
UFine. → ICFG.IRRA [12]34.5255.4164.4417.961.95
RDE [21]40.3760.1468.4120.542.19
Our ICL43.1062.9270.7322.732.56
Our ICL☆47.8362.6770.4823.162.62
UFine. → RSTP.IRRA [12]37.6563.7073.0029.0011.80
RDE [21]39.9063.5074.7529.9212.43
Our ICL48.8572.6581.8036.9116.39
Our ICL☆55.3572.4081.5038.6417.23
", + "image_path": "e353bcf20aa123499b0fd678eb9c40ffdf77a8dc77a0d386749cfafb9959544e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 500, + 295, + 523 + ], + "lines": [ + { + "bbox": [ + 55, + 500, + 295, + 523 + ], + "spans": [ + { + "bbox": [ + 55, + 500, + 295, + 523 + ], + "type": "text", + "content": "Table 4. Generalization capabilities between coarse-grained and fine-grained datasets. The best scores in each task are in bold." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 544, + 295, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 568 + ], + "type": "text", + "content": "formance, with Rank-1 exceeding " + }, + { + "bbox": [ + 55, + 544, + 295, + 568 + ], + "type": "inline_equation", + "content": "91\\%" + }, + { + "bbox": [ + 55, + 544, + 295, + 568 + ], + "type": "text", + "content": ". This shows that interaction is also applicable to the fine-grained scenario." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 575, + 178, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 575, + 178, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 575, + 178, + 588 + ], + "type": "text", + "content": "4.4. Generalization Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": "To evaluate the generalization, Tables 3 and 4 report the cross-domain performance on four datasets, including the coarse-to-fine, coarse-to-fine, and fine-to-coarse generalization experiments. From the results, our method achieves better generalization than RDE even without THI thanks to the training enhancement of RDA. When THI is performed, the cross-domain performance is dramatically improved, for example, from CUHK-PEDES to RSTPReid, THI brings an improvement of more than " + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": " on Rank-1. This shows that THI is a potential solution to generaliza" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 250, + 555, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 250, + 555, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 250, + 555, + 346 + ], + "type": "text", + "content": "tion challenges in the future. From the generalization experiments between fine-grained and coarse-grained datasets shown in Table 4, ICL can also achieve the best cross-domain performance, e.g., compared with the best baseline RDE, from UFine6926 domain to CUHK-PEDES domain, our method improves Rank-1 and mAP by " + }, + { + "bbox": [ + 313, + 250, + 555, + 346 + ], + "type": "inline_equation", + "content": "17.49\\%" + }, + { + "bbox": [ + 313, + 250, + 555, + 346 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 250, + 555, + 346 + ], + "type": "inline_equation", + "content": "10.86\\%" + }, + { + "bbox": [ + 313, + 250, + 555, + 346 + ], + "type": "text", + "content": ", respectively, which further verifies the cross-domain generalization of our method." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 318, + 355, + 553, + 434 + ], + "blocks": [ + { + "bbox": [ + 318, + 355, + 553, + 434 + ], + "lines": [ + { + "bbox": [ + 318, + 355, + 553, + 434 + ], + "spans": [ + { + "bbox": [ + 318, + 355, + 553, + 434 + ], + "type": "table", + "html": "
MethodsTHICUHK-PEDESICFG-PEDESRSTPReidΔAvg
Rank-1mAPRank-1mAPRank-1mAP
CLIP [24]X71.6463.9260.1134.5256.5544.52+2.57
✓73.7765.6663.5734.9561.4546.25
IRRA [12]X73.3866.1363.4638.0660.2047.17+1.86
✓76.0667.4265.2638.5863.7548.47
RDE [21]X75.9467.5667.6840.0665.3550.88+1.41
✓77.4768.6268.7240.6368.4552.01
", + "image_path": "02a38d9086de556b5f604e246c96e45289c867c683aa4175f322f0a959b640c4.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 313, + 435, + 553, + 456 + ], + "lines": [ + { + "bbox": [ + 313, + 435, + 553, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 435, + 553, + 456 + ], + "type": "text", + "content": "Table 5. Transferability results on three coarse-grained benchmarks. " + }, + { + "bbox": [ + 313, + 435, + 553, + 456 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 313, + 435, + 553, + 456 + ], + "type": "text", + "content": " Avg represents the average improvement." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 477, + 437, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 477, + 437, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 477, + 437, + 491 + ], + "type": "text", + "content": "4.5. Transferability Study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 496, + 555, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 496, + 555, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 496, + 555, + 664 + ], + "type": "text", + "content": "Our interactive module is separate and independent from the training of existing TIReID methods, so it can be plug-and-play with existing methods to improve ReID performance. To verify the transferability of our THI, we conduct experiments on multiple baselines, and the results are shown in Table 5. Except for the CLIP results we reproduced, all other experiments used public pre-trained model weights. From the results of Table 5, the interactive strategy application can significantly improve Rank-1 and mAP, which shows that the external guidance by interactions via MLLMs can further clarify the text-image alignments and improve the overall ranking. The average improvements on the three datasets are " + }, + { + "bbox": [ + 313, + 496, + 555, + 664 + ], + "type": "inline_equation", + "content": "2.57\\%" + }, + { + "bbox": [ + 313, + 496, + 555, + 664 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 496, + 555, + 664 + ], + "type": "inline_equation", + "content": "1.86\\%" + }, + { + "bbox": [ + 313, + 496, + 555, + 664 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 496, + 555, + 664 + ], + "type": "inline_equation", + "content": "1.41\\%" + }, + { + "bbox": [ + 313, + 496, + 555, + 664 + ], + "type": "text", + "content": ", respectively, which seriously proves the transferability of our THI." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 671, + 481, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 481, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 481, + 685 + ], + "type": "text", + "content": "4.6. Ablation and Parameter Study" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "To explore the effects of each proposed component, i.e., THI and RDA, we first conduct the ablation study on three" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "14396" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "type": "text", + "content": "coarse-grained datasets as reported in Table 6. From the results, each component can bring performance gains on Rank-1 and mAP, which verifies the reliability and rationality of the method. Especially the introduction of THI has greatly improved the Rank-1 accuracy from " + }, + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "type": "inline_equation", + "content": "68.95\\%" + }, + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "type": "inline_equation", + "content": "70.55\\%" + }, + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "type": "text", + "content": " on the RSTPreid dataset. In addition, our RDA also brings considerable performance gains, especially on the RSTPreid dataset. Also, we conduct the parameter analysis on the CHUK-PEDES dataset on two free hyperparameters, i.e., the similarity threshold " + }, + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "type": "text", + "content": " and the balance factor " + }, + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "type": "text", + "content": ". The former filters unnecessary retrievals to improve the interaction efficiency, while the latter controls the contribution of refined texts. Based on Figure 5, in our experiments, we set " + }, + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "type": "text", + "content": " in the range of " + }, + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "type": "inline_equation", + "content": "0.5 \\sim 0.6" + }, + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "type": "text", + "content": " to 0.8, thus mitigating the risk of introducing noisy external knowledge." + } + ] + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 60, + 259, + 293, + 322 + ], + "blocks": [ + { + "bbox": [ + 60, + 259, + 293, + 322 + ], + "lines": [ + { + "bbox": [ + 60, + 259, + 293, + 322 + ], + "spans": [ + { + "bbox": [ + 60, + 259, + 293, + 322 + ], + "type": "table", + "html": "
No.THIRDALoRACUHK-PEDESICFG-PEDESRSTPReid
Rank-1mAPRank-1mAPRank-1mAP
#1✓✓✓77.9169.3269.0241.2170.5553.68
#2✓✓×76.3868.5967.9241.1369.0053.11
#3×✓×76.4168.0468.1140.8167.7052.62
#4×××75.9467.5667.6840.0665.3550.88
", + "image_path": "2fcadfc6a9a88b84ef08d3161ccfe833f7724ab84e154b95bfa4dfaa6b400281.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 323, + 295, + 346 + ], + "lines": [ + { + "bbox": [ + 55, + 323, + 295, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 295, + 346 + ], + "type": "text", + "content": "Table 6. Ablation studies on CHUK-PEDES, ICFG-PEDES, and RSTPReid datasets. The best scores are in bold." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 56, + 356, + 176, + 434 + ], + "blocks": [ + { + "bbox": [ + 56, + 356, + 176, + 434 + ], + "lines": [ + { + "bbox": [ + 56, + 356, + 176, + 434 + ], + "spans": [ + { + "bbox": [ + 56, + 356, + 176, + 434 + ], + "type": "image", + "image_path": "3d7772f515054dae074f670307524539aa3f362bf5097f1182732169c1447a2f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 436, + 164, + 445 + ], + "lines": [ + { + "bbox": [ + 69, + 436, + 164, + 445 + ], + "spans": [ + { + "bbox": [ + 69, + 436, + 164, + 445 + ], + "type": "text", + "content": "(a) The similarity threshold " + }, + { + "bbox": [ + 69, + 436, + 164, + 445 + ], + "type": "inline_equation", + "content": "\\xi" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 69, + 448, + 281, + 459 + ], + "lines": [ + { + "bbox": [ + 69, + 448, + 281, + 459 + ], + "spans": [ + { + "bbox": [ + 69, + 448, + 281, + 459 + ], + "type": "text", + "content": "Figure 5. Variation of performance with different " + }, + { + "bbox": [ + 69, + 448, + 281, + 459 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 69, + 448, + 281, + 459 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 69, + 448, + 281, + 459 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 69, + 448, + 281, + 459 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 175, + 356, + 296, + 434 + ], + "blocks": [ + { + "bbox": [ + 175, + 356, + 296, + 434 + ], + "lines": [ + { + "bbox": [ + 175, + 356, + 296, + 434 + ], + "spans": [ + { + "bbox": [ + 175, + 356, + 296, + 434 + ], + "type": "image", + "image_path": "c584e6d8acd84e0846325cc373435b3a25138bd51eba753542aedb69fc859d7b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 196, + 436, + 275, + 445 + ], + "lines": [ + { + "bbox": [ + 196, + 436, + 275, + 445 + ], + "spans": [ + { + "bbox": [ + 196, + 436, + 275, + 445 + ], + "type": "text", + "content": "(b) The balance factor " + }, + { + "bbox": [ + 196, + 436, + 275, + 445 + ], + "type": "inline_equation", + "content": "\\lambda" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 468, + 160, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 160, + 480 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 160, + 480 + ], + "type": "text", + "content": "4.7. Interactive Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "type": "text", + "content": "This section further explores the interactive module, i.e. THI. We report the performance (mAP) changes after multiple rounds of interactions in Figure 6. The mAP score can reflect the overall retrieval quality. From the results, the performance improvement is obvious in the first few rounds " + }, + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "type": "inline_equation", + "content": "(\\leq 2)" + }, + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "type": "text", + "content": " since the items that meet the query semantics are mostly concentrated at the top of the ranking. However, after " + }, + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "type": "inline_equation", + "content": ">2" + }, + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "type": "text", + "content": " rounds, the performance gain is gradually not obvious as the number of queries requiring interaction decreases. But generally, as the number of rounds increases, the overall performance gradually improves. In all our experiments, we performed 5 rounds of interaction. In addition, we visualize the top-10 retrieved results before and after applying THI in Figure 7. Due to the inherent limitations of the intra-model knowledge, it is difficult to obtain satisfactory results without the help of THI. In contrast, our THI can dynamically enrich queries by interacting with MLLMs to achieve a more reliable ranking. More example results are given in the supplementary material." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 315, + 71, + 396, + 143 + ], + "blocks": [ + { + "bbox": [ + 315, + 71, + 396, + 143 + ], + "lines": [ + { + "bbox": [ + 315, + 71, + 396, + 143 + ], + "spans": [ + { + "bbox": [ + 315, + 71, + 396, + 143 + ], + "type": "image", + "image_path": "a468f43e12e93e9db9044752c1b7bab4faeafca3d2345fae931fe51558325a3b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 324, + 145, + 386, + 154 + ], + "lines": [ + { + "bbox": [ + 324, + 145, + 386, + 154 + ], + "spans": [ + { + "bbox": [ + 324, + 145, + 386, + 154 + ], + "type": "text", + "content": "(a) CUHK-PEDES" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 396, + 70, + 474, + 143 + ], + "blocks": [ + { + "bbox": [ + 396, + 70, + 474, + 143 + ], + "lines": [ + { + "bbox": [ + 396, + 70, + 474, + 143 + ], + "spans": [ + { + "bbox": [ + 396, + 70, + 474, + 143 + ], + "type": "image", + "image_path": "934c89310c3ab053d56bdfe58e5389df97e51d022f40a831122983b2620aa41d.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 145, + 463, + 154 + ], + "lines": [ + { + "bbox": [ + 406, + 145, + 463, + 154 + ], + "spans": [ + { + "bbox": [ + 406, + 145, + 463, + 154 + ], + "type": "text", + "content": "(b) ICFG-PEDES" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 157, + 553, + 179 + ], + "lines": [ + { + "bbox": [ + 313, + 157, + 553, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 553, + 179 + ], + "type": "text", + "content": "Figure 6. Performance (mAP) versus rounds on three datasets. Round 0 indicates the setting without using THI." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 474, + 71, + 553, + 143 + ], + "blocks": [ + { + "bbox": [ + 474, + 71, + 553, + 143 + ], + "lines": [ + { + "bbox": [ + 474, + 71, + 553, + 143 + ], + "spans": [ + { + "bbox": [ + 474, + 71, + 553, + 143 + ], + "type": "image", + "image_path": "543546290431ffbbd3f2bc3c372f4df8c65819deec1bed410650b8973891e94e.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 491, + 145, + 537, + 154 + ], + "lines": [ + { + "bbox": [ + 491, + 145, + 537, + 154 + ], + "spans": [ + { + "bbox": [ + 491, + 145, + 537, + 154 + ], + "type": "text", + "content": "(c) RSTPReid" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 317, + 195, + 404, + 255 + ], + "blocks": [ + { + "bbox": [ + 317, + 195, + 404, + 255 + ], + "lines": [ + { + "bbox": [ + 317, + 195, + 404, + 255 + ], + "spans": [ + { + "bbox": [ + 317, + 195, + 404, + 255 + ], + "type": "image", + "image_path": "d81673e82a8d80de989625b02f3c922b1d4291443a750f08dde17504ee196e20.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 404, + 195, + 553, + 254 + ], + "blocks": [ + { + "bbox": [ + 404, + 195, + 553, + 254 + ], + "lines": [ + { + "bbox": [ + 404, + 195, + 553, + 254 + ], + "spans": [ + { + "bbox": [ + 404, + 195, + 553, + 254 + ], + "type": "image", + "image_path": "ad0c671cfb7951ea6984a1adb15a69a27340c9ac356219ca45d2459bec9af3f6.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 317, + 258, + 404, + 311 + ], + "blocks": [ + { + "bbox": [ + 317, + 258, + 404, + 311 + ], + "lines": [ + { + "bbox": [ + 317, + 258, + 404, + 311 + ], + "spans": [ + { + "bbox": [ + 317, + 258, + 404, + 311 + ], + "type": "image", + "image_path": "fb579c33600b81633192f7179c87f2f24f84bb02991a511821ce96e48471a5bb.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 316, + 553, + 338 + ], + "lines": [ + { + "bbox": [ + 313, + 316, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 553, + 338 + ], + "type": "text", + "content": "Figure 7. Top-10 retrieved results on CUHK-PEDES dataset between ICL (the first row) and ICL with THI (the second row)." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 404, + 257, + 553, + 314 + ], + "blocks": [ + { + "bbox": [ + 404, + 257, + 553, + 314 + ], + "lines": [ + { + "bbox": [ + 404, + 257, + 553, + 314 + ], + "spans": [ + { + "bbox": [ + 404, + 257, + 553, + 314 + ], + "type": "image", + "image_path": "7042d8b8e506934f997341f636bbb76dfbd27533654a8eaaf277610a97d429f5.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 314, + 364, + 388, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 364, + 388, + 376 + ], + "spans": [ + { + "bbox": [ + 314, + 364, + 388, + 376 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 386, + 555, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 386, + 555, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 386, + 555, + 542 + ], + "type": "text", + "content": "In this paper, we explore interactive text-to-image person re-identification, which aims to improve the alignment between dynamic queries and challenging candidate images by leveraging external guidance from MLLMs. To achieve this, we develop an Interactive Cross-modal Learning (ICL) framework to alleviate the inherent challenges of offline models and training data by, including a plug-and-play Test-time Human-centered Interaction (THI) module and Reorganization Data Augmentation (RDA). Extensive experiments and analysis show that our framework can effectively transfer external knowledge in MLLMs into offline models for guiding re-identification, showing excellent performance and generalization." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 314, + 559, + 411, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 559, + 411, + 572 + ], + "spans": [ + { + "bbox": [ + 314, + 559, + 411, + 572 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "This work was supported in part by the National Key R&D Program of China under Grant 2024YFB4710604; in part by NSFC under Grants 62472295, 62372315, 62176171, and U21B2040; in part by Sichuan Science and Technology Planning Projects (2024YFHZ0089, 2024NSFTD0049, 2024YFHZ0144, 2024NSFTD0047, 2024NSFTD0038); in part by System of Systems and Artificial Intelligence Laboratory pioneer fund grant; in part by the Fundamental Research Funds for the Central Universities under Grant CJ202403 and CJ202303; in part by Chengdu Science and Technology Project under Grant 2023-XT00-00004-GX." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14397" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "type": "text", + "content": "[1] Yang Bai, Min Cao, Daming Gao, Ziqiang Cao, Chen Chen, Zhenfeng Fan, Liqiang Nie, and Min Zhang. Rasa: relation and sensitivity aware representation learning for text-based person search. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, pages 555–563, 2023. 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 158, + 296, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 158, + 296, + 212 + ], + "spans": [ + { + "bbox": [ + 61, + 158, + 296, + 212 + ], + "type": "text", + "content": "[2] Maryam Bukhari, Sadaf Yasmin, Sheneela Naz, Muazzam Maqsood, Jehyeok Rew, and Seungmin Rho. Language and vision based person re-identification for surveillance systems using deep learning with lip layers. Image and Vision Computing, 132:104658, 2023. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 213, + 295, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 213, + 295, + 256 + ], + "spans": [ + { + "bbox": [ + 62, + 213, + 295, + 256 + ], + "type": "text", + "content": "[3] Min Cao, Yang Bai, Ziyin Zeng, Mang Ye, and Min Zhang. An empirical study of clip for text-based person search. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 465-473, 2024. 1, 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 258, + 295, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 258, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 62, + 258, + 295, + 300 + ], + "type": "text", + "content": "[4] Yuhao Chen, Guoqing Zhang, Yujiang Lu, Zhenxing Wang, and Yuhui Zheng. Tipcb: A simple but effective part-based convolutional baseline for text-based person search. Neurocomputing, 494:171-181, 2022." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 302, + 295, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 302, + 295, + 345 + ], + "spans": [ + { + "bbox": [ + 62, + 302, + 295, + 345 + ], + "type": "text", + "content": "[5] Zefeng Ding, Changxing Ding, Zhiyin Shao, and Dacheng Tao. Semantically self-aligned network for text-to-image part-aware person re-identification. arXiv preprint arXiv:2107.12666, 2021. 1, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 346, + 295, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 346, + 295, + 378 + ], + "spans": [ + { + "bbox": [ + 62, + 346, + 295, + 378 + ], + "type": "text", + "content": "[6] Hiren Galiyawala and Mehul S Raval. Person retrieval in surveillance using textual query: a review. Multimedia Tools and Applications, 80(18):27343-27383, 2021. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 380, + 294, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 294, + 412 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 294, + 412 + ], + "type": "text", + "content": "[7] Daming Gao, Yang Bai, Min Cao, Hao Dou, Mang Ye, and Min Zhang. Semi-supervised text-based person search. arXiv preprint arXiv:2404.18106, 2024. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 415, + 295, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 415, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 62, + 415, + 295, + 456 + ], + "type": "text", + "content": "[8] Donghoon Han, Eunhwan Park, Gisang Lee, Adam Lee, and Nojun Kwak. Merlin: Multimodal embedding refinement via llm-based iterative navigation for text-video retrieval-erank pipeline. arXiv preprint arXiv:2407.12508, 2024. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 458, + 295, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 458, + 295, + 500 + ], + "spans": [ + { + "bbox": [ + 62, + 458, + 295, + 500 + ], + "type": "text", + "content": "[9] Chen He, Shenshen Li, Zheng Wang, Hua Chen, Fumin Shen, and Xing Xu. Chatting with interactive memory for text-based person retrieval. Multimedia Systems, 31(1):31, 2025. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 502, + 295, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 502, + 295, + 545 + ], + "spans": [ + { + "bbox": [ + 57, + 502, + 295, + 545 + ], + "type": "text", + "content": "[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 547, + 295, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 547, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 57, + 547, + 295, + 590 + ], + "type": "text", + "content": "[11] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 4, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 592, + 295, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 592, + 295, + 645 + ], + "spans": [ + { + "bbox": [ + 57, + 592, + 295, + 645 + ], + "type": "text", + "content": "[12] Ding Jiang and Mang Ye. Cross-modal implicit relation reasoning and aligning for text-to-image person retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2787-2797, 2023. 1, 2, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "type": "text", + "content": "[13] Matan Levy, Rami Ben-Ari, Nir Darshan, and Dani Lischinski. Chatting makes perfect: Chat-based image retrieval. Advances in Neural Information Processing Systems, 36, 2024. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "text", + "content": "[14] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 74, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 335, + 74, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 553, + 105 + ], + "type": "text", + "content": "vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 107, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 161 + ], + "type": "text", + "content": "[15] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 163, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 553, + 217 + ], + "type": "text", + "content": "[16] Shuang Li, Tong Xiao, Hongsheng Li, Bolei Zhou, Dayu Yue, and Xiaogang Wang. Person search with natural language description. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1970-1979, 2017. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 220, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 220, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 317, + 220, + 553, + 262 + ], + "type": "text", + "content": "[17] Yunfan Li, Peng Hu, Dezhong Peng, Jiancheng Lv, Jianping Fan, and Xi Peng. Image clustering with external guidance. In *Forty-first International Conference on Machine Learning*. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 264, + 553, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 286 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 286 + ], + "type": "text", + "content": "[18] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 288, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 288, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 317, + 288, + 553, + 342 + ], + "type": "text", + "content": "[19] Yang Qin, Dezhong Peng, Xi Peng, Xu Wang, and Peng Hu. Deep evidential learning with noisy correspondence for cross-modal retrieval. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4948-4956, 2022. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 344, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 344, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 317, + 344, + 553, + 398 + ], + "type": "text", + "content": "[20] Yang Qin, Yuan Sun, Dezhong Peng, Joey Tianyi Zhou, Xi Peng, and Peng Hu. Cross-modal active complementary learning with self-refining correspondence. Advances in neural information processing systems, 36:24829-24840, 2023. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 400, + 553, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 400, + 553, + 463 + ], + "spans": [ + { + "bbox": [ + 317, + 400, + 553, + 463 + ], + "type": "text", + "content": "[21] Yang Qin, Yingke Chen, Dezhong Peng, Xi Peng, Joey Tianyi Zhou, and Peng Hu. Noisy-correspondence learning for text-to-image person re-identification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27197-27206, 2024. 1, 2, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 467, + 553, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 467, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 317, + 467, + 553, + 499 + ], + "type": "text", + "content": "[22] Yalan Qin, Chuan Qin, Xinpeng Zhang, and Guorui Feng. Dual consensus anchor learning for fast multi-view clustering. IEEE Transactions on Image Processing, 2024. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 501, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 501, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 317, + 501, + 553, + 533 + ], + "type": "text", + "content": "[23] Yalan Qin, Nan Pu, Nicu Sebe, and Guorui Feng. Latent space learning based ensemble clustering. IEEE Transactions on Image Processing, 2025. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 535, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 535, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 317, + 535, + 553, + 600 + ], + "type": "text", + "content": "[24] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 602, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 602, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 317, + 602, + 553, + 656 + ], + "type": "text", + "content": "[25] Zhiyin Shao, Xinyu Zhang, Meng Fang, Zhifeng Lin, Jian Wang, and Changxing Ding. Learning granularity-unified representations for text-to-image person re-identification. In Proceedings of the 30th ACM International Conference on Multimedia, 2022. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 553, + 713 + ], + "type": "text", + "content": "[26] Xiujun Shu, Wei Wen, Haoqian Wu, Keyu Chen, Yiran Song, Ruizhi Qiao, Bo Ren, and Xiao Wang. See finer, see more: Implicit modality alignment for text-based person retrieval. In European Conference on Computer Vision, pages 624-641. Springer, 2022. 6" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14398" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[27] Yuan Sun, Yang Qin, Yongxiang Li, Dezhong Peng, Xi Peng, and Peng Hu. Robust multi-view clustering with noisy correspondence. IEEE Transactions on Knowledge and Data Engineering, 2024. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 172 + ], + "type": "text", + "content": "[28] Wentan Tan, Changxing Ding, Jiayu Jiang, Fei Wang, Yibing Zhan, and Dapeng Tao. Harnessing the power of mllms for transferable text-to-image person reid. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17127-17137, 2024. 1, 2, 3, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 173, + 294, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 173, + 294, + 249 + ], + "spans": [ + { + "bbox": [ + 56, + 173, + 294, + 249 + ], + "type": "text", + "content": "[29] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 250, + 294, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 294, + 303 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 294, + 303 + ], + "type": "text", + "content": "[30] Yansong Wang, Yundong Sun, Yansheng Fu, Dongjie Zhu, and Zhaoshuo Tian. Spectrum-bert: pre-training of deep bidirectional transformers for spectral classification of Chinese liquors. IEEE Transactions on Instrumentation and Measurement, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 304, + 294, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 304, + 294, + 348 + ], + "spans": [ + { + "bbox": [ + 56, + 304, + 294, + 348 + ], + "type": "text", + "content": "[31] Yabing Wang, Le Wang, Qiang Zhou, Zhibin Wang, Hao Li, Gang Hua, and Wei Tang. Multimodal llm enhanced cross-lingual cross-modal retrieval. arXiv preprint arXiv:2409.19961, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 350, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 294, + 392 + ], + "type": "text", + "content": "[32] Yuhao Wang, Yang Liu, Aihua Zheng, and Pingping Zhang. Demo: Decoupled feature-based mixture of experts for multi-modal object re-identification. In Proceedings of the AAAI Conference on Artificial Intelligence, 2025. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 394, + 294, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 394, + 294, + 458 + ], + "spans": [ + { + "bbox": [ + 56, + 394, + 294, + 458 + ], + "type": "text", + "content": "[33] Zhe Wang, Zhiyuan Fang, Jun Wang, and Yezhou Yang. Vistaa: Visual-textual attributes alignment in person search by natural language. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XII 16, pages 402-420. Springer, 2020. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 460, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 294, + 514 + ], + "type": "text", + "content": "[34] Zijie Wang, Aichun Zhu, Jingyi Xue, Xili Wan, Chao Liu, Tian Wang, and Yifeng Li. Caibc: Capturing all-round information beyond color for text-based person retrieval. In Proceedings of the 30th ACM International Conference on Multimedia, pages 5314-5322, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 515, + 294, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 515, + 294, + 568 + ], + "spans": [ + { + "bbox": [ + 56, + 515, + 294, + 568 + ], + "type": "text", + "content": "[35] Yushuang Wu, Zizheng Yan, Xiaoguang Han, Guanbin Li, Changqing Zou, and Shuguang Cui. Lipscore: language-guided person search via color reasoning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1624-1633, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 570, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 294, + 624 + ], + "type": "text", + "content": "[36] Shuanglin Yan, Neng Dong, Jun Liu, Liyan Zhang, and Jinhui Tang. Learning comprehensive representations with richer self for text-to-image person re-identification. In Proceedings of the 31st ACM international conference on multimedia, pages 6202–6211, 2023. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 625, + 294, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 294, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 294, + 658 + ], + "type": "text", + "content": "[37] Shuanglin Yan, Neng Dong, Liyan Zhang, and Jinhui Tang. Clip-driven fine-grained text-image person re-identification. IEEE Transactions on Image Processing, 2023. 1, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "text", + "content": "[38] Shuyu Yang, Yinan Zhou, Zhedong Zheng, Yaxiong Wang, Li Zhu, and Yujiao Wu. Towards unified text-based person retrieval: A large-scale multi-attribute and language search benchmark. In Proceedings of the 31st ACM International Conference on Multimedia, pages 4492–4501, 2023. 1, 2, 6" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 373 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "type": "text", + "content": "[39] Pingping Zhang, Yuhao Wang, Yang Liu, Zhengzheng Tu, and Huchuan Lu. Magic tokens: Select diverse tokens for multi-modal object re-identification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17117-17126, 2024. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "type": "text", + "content": "[40] Yaowei Zheng, Richong Zhang, Junhao Zhang, Yanhan Ye, Zheyan Luo, Zhangchi Feng, and Yongqiang Ma. Llamafactory: Unified efficient fine-tuning of " + }, + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "type": "inline_equation", + "content": "100+" + }, + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "type": "text", + "content": " language models. arXiv preprint arXiv:2403.13372, 2024. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 174, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 553, + 228 + ], + "type": "text", + "content": "[41] Aichun Zhu, Zijie Wang, Yifeng Li, Xili Wan, Jing Jin, Tian Wang, Fangqiang Hu, and Gang Hua. Dssl: Deep surroundings-person separation learning for text-based person retrieval. In Proceedings of the 29th ACM International Conference on Multimedia, pages 209-217, 2021. 1, 2, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 230, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 553, + 262 + ], + "type": "text", + "content": "[42] Jialong Zuo, Changqian Yu, Nong Sang, and Changxin Gao. Plip: Language-image pre-training for person representation learning, 2023. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 263, + 553, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 263, + 553, + 307 + ], + "spans": [ + { + "bbox": [ + 316, + 263, + 553, + 307 + ], + "type": "text", + "content": "[43] Jialong Zuo, Ying Nie, Hanyu Zhou, Huaxin Zhang, Haoyu Wang, Tianyu Guo, Nong Sang, and Changxin Gao. Cross-video identity correlating for person re-identification pretraining. arXiv preprint arXiv:2409.18569, 2024. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 308, + 553, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 553, + 373 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 553, + 373 + ], + "type": "text", + "content": "[44] Jialong Zuo, Hanyu Zhou, Ying Nie, Feng Zhang, Tianyu Guo, Nong Sang, Yunhe Wang, and Changxin Gao. Ufinebench: Towards text-based person retrieval with ultra-fine granularity. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22010-22019, 2024. 1, 5, 6" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14399" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/HumanDreamer_ Generating Controllable Human-Motion Videos via Decoupled Generation/8f919b44-e98e-44f2-a774-268fc39345bc_content_list.json b/2025/HumanDreamer_ Generating Controllable Human-Motion Videos via Decoupled Generation/8f919b44-e98e-44f2-a774-268fc39345bc_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8ecc7655b5608fb445b116ff977b8304311739cc --- /dev/null +++ b/2025/HumanDreamer_ Generating Controllable Human-Motion Videos via Decoupled Generation/8f919b44-e98e-44f2-a774-268fc39345bc_content_list.json @@ -0,0 +1,1922 @@ +[ + { + "type": "text", + "text": "HumanDreamer: Generating Controllable Human-Motion Videos via Decoupled Generation", + "text_level": 1, + "bbox": [ + 166, + 130, + 831, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Boyuan Wang $^{1,2,3*}$ , Xiaofeng Wang $^{1,2,3*}$ , Chaojun Ni $^{4,5}$ , Guosheng Zhao $^{1,2,3}$ , Zhiqin Yang $^{6}$ , Zheng Zhu $^{4\\dagger}$ , Muyang Zhang $^{1,2}$ , Yukun Zhou $^{4}$ , Xinze Chen $^{4}$ , Guan Huang $^{4}$ , Lihong Liu $^{1}$ , Xingang Wang $^{1,3\\dagger}$", + "bbox": [ + 99, + 202, + 947, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Institute of Automation, Chinese Academy of Sciences, China", + "bbox": [ + 274, + 238, + 774, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ School of Artificial Intelligence, University of Chinese Academy of Sciences, China", + "bbox": [ + 186, + 256, + 859, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Luoyang Institute for Robot and Intelligent Equipment, China", + "bbox": [ + 274, + 273, + 772, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ GigaAI, China $^{5}$ Peking University, China $^{6}$ The Chinese University of Hong Kong", + "bbox": [ + 184, + 292, + 854, + 311 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4d9fb75c2e6b73f5f1aeca537ac865c23d69eefa743b4ccb77d156339745e9de.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 158, + 362, + 181, + 381 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text", + "bbox": [ + 187, + 364, + 225, + 378 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A man is performing a yoga pose on a mat, and he is seen moving his legs and arms in different positions.", + "bbox": [ + 261, + 349, + 560, + 372 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A man is speaking to the camera and then looking off into the distance. Later, he is speaking to the camera again.", + "bbox": [ + 261, + 372, + 560, + 398 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1854e0342367af45949e612f9934120c5176b22a34f9ef0c1da950452cd83400.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 575, + 363, + 637, + 383 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "HumanDreamer", + "bbox": [ + 643, + 366, + 779, + 381 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4dbab8c84cbcdf186b3e3fdcf022ab2cfdbab02333e726ef38f2707e48ad6a4d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 148, + 416, + 263, + 484 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e8f827329573956ee15bab64424ad7bc7c6f34f1f11dfe4b663ce163e37243e9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 264, + 416, + 379, + 484 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3f8ca9494a8d2e8e435af44a0f2df93b6b51b43364b09e682e3072412ab75e3a.jpg", + "image_caption": [ + "Text-to-Pose Generation" + ], + "image_footnote": [], + "bbox": [ + 380, + 416, + 495, + 484 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1e94a5cf8ef8bd75b2ad9c2661ba34fdc8921647fa25899131cf3cf80e56de86.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 416, + 612, + 484 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/77a51d7e752bd3ae932752d0e934f9c1e271d1eeecffea74a16638c168a7cdb1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 612, + 416, + 725, + 484 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9c7b5e002ac917dc1cfc9102ea658206347663dee52a7e2f6edc83ee7451d53b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 727, + 416, + 843, + 484 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bbfc19f914967ff7e30916898b4b3ee811b9cfd6da0732ca59febdb7bea83fd7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 147, + 491, + 246, + 580 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b81e79a4b9f71ee5e97d63f32f3c11fc55203e3111f583c3e9351eb2c185e3e4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 248, + 491, + 344, + 580 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7dd132b0bb811a968d662444727540635d190c1d7e21b54d0d37c1f7a80c3ffd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 346, + 491, + 444, + 580 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1e973707081fd3653105c4b1e44468297d29da928eaac73f3a74a2dcd102bdeb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 446, + 491, + 544, + 580 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a34c8b76e8c9131652bcc889daf840037bc469e98f83e00b679696eb609ab49e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 545, + 491, + 643, + 580 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0533623de33a056f233d3a583df12fc7c5c10ef586237acddae55e348e69b62d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 643, + 491, + 743, + 580 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/daa34c4e486e3ea4a975e5efae0a064395f7aed52bd0cf482497f8933fb9ae6d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 743, + 491, + 843, + 580 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e2f2f9cd04c43aae3b98fd31ee51a77c167696d04fa73e9db9102a72558d0f33.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 148, + 604, + 263, + 672 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e27aa5ad574bd389e80ad374abcec69fb9322e4291872c0dadc62657f668ddf2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 264, + 604, + 379, + 672 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1bab931afc1bde3e962f5d768f56ef3f5a2d7adbbb4c109d05c082412097e3a9.jpg", + "image_caption": [ + "Pose-to-Video Generation" + ], + "image_footnote": [], + "bbox": [ + 380, + 604, + 495, + 672 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fea104d67a494498303501440134ab397c9eb1f9675169296cbbd13c4ef46712.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 604, + 612, + 672 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e70b871a72fb3f17152b178008bc2601392060f6a4328678ac72df32b81cdc61.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 612, + 604, + 725, + 672 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/235df14e4cc9e27bf1fd654c6e21af1c1b1d7fa5afdb3e00cc6a1437e8843926.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 727, + 604, + 843, + 672 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d7ba48354bf5fcd85a86e32ad98c15ff25e7dc99e929dd3ef6aa607dec419475.jpg", + "image_caption": [ + "Figure 1. Illustration of HumanDreamer. The human-motion video generation is decoupled into two steps: Text-to-Pose generation and Pose-to-Video generation. The decoupled process integrates the flexibility of text control and the controllability of pose guidance." + ], + "image_footnote": [], + "bbox": [ + 148, + 676, + 250, + 768 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7ad538ec01880978193cbdb436ba3dba39d3f443523d990691fedf6052e754fa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 251, + 676, + 348, + 768 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/46ff04edc0bab07374dea19d2fdfd72686944221f216e398e5dcaf146389d08a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 349, + 676, + 447, + 768 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/07f5ce86b40c3105d2cc62afb95a08c7c7f214aba3271a3c7cea434c4315b78c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 449, + 676, + 547, + 768 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7815a721cd409d9f706e6362a7f174275509cd0e45c9b9b5272119058240b1bd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 547, + 676, + 645, + 768 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/41ec9997937ddcb56dbb189f33ab8f6938b27efc9a3201cd7fa3983a5543a29f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 647, + 676, + 745, + 768 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4e999f3cd01a9303c07fe61f8c505da6d7a4b18c3278fa87d11620f5a9d3ab2e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 746, + 676, + 844, + 768 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 669, + 830, + 750, + 847 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation.", + "bbox": [ + 236, + 0, + 810, + 20 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "Except for this watermark, it is identical to the accepted version;", + "bbox": [ + 323, + 16, + 723, + 31 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 292, + 31, + 754, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*These authors contributed equally to this work.", + "bbox": [ + 107, + 862, + 364, + 875 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding authors. zhengzhu@ieee.org, xingang.wang@ia.ac.cn.", + "bbox": [ + 109, + 875, + 482, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\ddagger$ Project Page: https://humandreamer.github.io/", + "bbox": [ + 109, + 887, + 433, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Human-motion video generation has been a challenging task, primarily due to the difficulty inherent in learning hu", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12391", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "man body movements. While some approaches have attempted to drive human-centric video generation explicitly through pose control, these methods typically rely on poses derived from existing videos, thereby lacking flexibility. To address this, we propose HumanDreamer, a decoupled human video generation framework that first generates diverse poses from text prompts and then leverages these poses to generate human-motion videos. Specifically, we propose MotionVid, the largest dataset for human-motion pose generation. Based on the dataset, we present MotionDiT, which is trained to generate structured human-motion poses from text prompts. Besides, a novel LAMA loss is introduced, which together contribute to a significant improvement in FID by $62.4\\%$ , along with respective enhancements in R-precision for top1, top2, and top3 by $41.8\\%$ , $26.3\\%$ , and $18.3\\%$ , thereby advancing both the Text-to-Pose control accuracy and FID metrics. Our experiments across various Pose-to-Video baselines demonstrate that the poses generated by our method can produce diverse and high-quality human-motion videos. Furthermore, our model can facilitate other downstream tasks, such as pose sequence prediction and 2D-3D motion lifting.", + "bbox": [ + 93, + 90, + 480, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 94, + 452, + 220, + 467 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generating human-motion videos remains a particularly challenging task due to the inherent complexity of modeling human body movements. Despite current advancements in generative modeling [9, 13, 31, 48, 71], state-of-the-art video generation models [3, 16, 19, 36, 59, 61], equipped with billions of parameters and trained on millions of video and image data, still struggle to capture human body movements, frequently resulting in fragmented or unrealistic portrayals. This limitation is exacerbated when controlling human videos via textual conditions, highlighting the fundamental challenge of directly mapping text prompts to human visual data.", + "bbox": [ + 93, + 478, + 480, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To enhance the generation quality, approaches such asAnimate Anyone [20], UniAnimate [53], Mimic Motion [66], Champ [69] andAnimate-X [49] explicitly generate human-motion videos through pose control. By utilizing the pose-guided generation, these methods effectively reduce the complexities associated with human-motion videos. However, a notable limitation of this approach is its reliance on human-motion poses derived from existing videos, which restricts the flexibility.", + "bbox": [ + 93, + 659, + 480, + 794 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Therefore, we propose HumanDreamer, a decoupled human-motion video generation framework that first generates human-motion poses from text prompts and subsequently produces human-motion videos based on the generated poses. The motivation behind this decoupled approach lies in that Text-to-Pose presents a more manageable search space compared to the direct learning of text-to-pixel repre", + "bbox": [ + 93, + 795, + 480, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "sensation. As shown in Fig. 1, this decoupled framework facilitates a more effective generation of human movements. Additionally, the proposed HumanDreamer utilizes text as the input, offering greater flexibility than [20, 34, 49, 53, 66] that directly relies on pre-defined poses. Specifically, we construct a 1.2 million text-pose pairs dataset MotionVid, which is the largest dataset for human-motion pose generation. To ensure dataset quality, a comprehensive data cleaning pipeline is introduced, where the cleaning factors involve body movement amplitude, human presence duration, facial visibility, and the proportion of human figures. This rigorous cleaning guarantees that the dataset is reliable for training Text-to-Pose tasks. Based on the dataset, we propose the MotionDiT to generate structured human-motion poses from text prompts. The MotionDiT integrates a global attention block designed to extract the global characteristics of the entire pose sequence, alongside a local feature aggregation mechanism that captures information from adjacent pose points, effectively combining global and local perspectives to enhance the quality and coherence of the generated poses. Additionally, the LAMA loss is utilized in the latent space to align the semantic features of the ground truth motion with those produced by MotionDiT, which not only improves the interpretability of the model but also boosts its overall performance. The introduced techniques collectively result in a notable $62.4\\%$ enhancement in FID, accompanied by respective increases in R-precision for the top1, top2, and top3 categories by $41.8\\%$ , $26.3\\%$ , and $18.3\\%$ .", + "bbox": [ + 516, + 90, + 903, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The primary contributions of this work are as follows: (1) We present HumanDreamer, the first decoupled framework for human-motion video generation, which integrates the flexibility of text control with the controllability of pose guidance. (2) We propose the MotionVid, the largest dataset for human-motion pose generation. In MotionVid, we conduct a comprehensive data annotation and data cleaning to ensure data reliability. (3) The MotionDiT is introduced to generate diverse human-motion poses under text control. To enhance the generation process, we propose the LAMA loss, which significantly improves pose fidelity and diversity. (4) Through extensive experiments, we show that MotionDiT and the LAMA loss improve control accuracy and FID by $62.4\\%$ , with top-k R-precision gains of $41.8\\%$ , $26.3\\%$ , and $18.3\\%$ for top-1, top-2, and top-3. The generated poses also support downstream tasks like pose sequence prediction and 2D-to-3D motion enhancement.", + "bbox": [ + 516, + 530, + 903, + 786 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 516, + 805, + 653, + 820 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Human-motion Pose Datasets", + "text_level": 1, + "bbox": [ + 516, + 830, + 774, + 845 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Human motion datasets are essential for the development of human-centric perception [7, 8, 21, 30, 44, 57] or generation [6, 29, 32, 51, 52, 54, 55, 63, 68] tasks. Specifically,", + "bbox": [ + 516, + 854, + 903, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12392", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/22a1e483c3189286dfed3e85906f8d2d67c0438971256fb3eb066be9130f0993.jpg", + "image_caption": [ + "Figure 2. The data cleaning and annotation pipeline for MotionVid begins with raw data sourced from public datasets and the internet, which is then segmented into video clips. To ensure high-quality data, we apply video quality filter, data annotation, human quality filter, and caption quality filter." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 898, + 209 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the KIT Motion-Language Dataset [39] provides sequence-level descriptions to support multi-modal motion tasks. In contrast, the HumanML3D [15] dataset, which is built upon the AMASS [35] and HumanAct12 [14] datasets, offers richer textual annotations and a wider variety of activities. Additionally, Motion-X [29] establishes a robust annotation protocol and is the first to compile a comprehensive fine-grained 3D whole-body motion dataset derived from extensive scene captures. As a result, existing text-driven 3D motion datasets often lack sufficient volume and diversity, which limits their scalability. A dataset that shares similarities with ours is Holistic-Motion2D [55], which features 1M text-pose pairs but only 16K of these are publicly accessible. However, the proposed MotionVid significantly surpasses this with a larger and more diverse set of 1.2M text-pose data, coupled with a more thorough data cleaning process that ensures the dataset's reliability for training Text-to-Pose tasks.", + "bbox": [ + 91, + 287, + 483, + 559 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Human-motion Pose Generation Methods", + "text_level": 1, + "bbox": [ + 89, + 585, + 446, + 601 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text-driven human motion generation, which converts textual descriptions into motion sequences, has gained significant attention in recent years. Early approaches focus on aligning motion and text through shared latent spaces. For instance, MotionCLIP [51] enhances autoencoder generalization by aligning the latent space with the expressive CLIP [40] embedding. T2M-GPT [63] frames the task as predicting discrete motion indices. More recently, Diffusion-based models, including MDM [52], MotionDiffuse [65], and DiffGesture [68], have set new performance standards. MLD [6] introduces a latent Diffusion model that generates diverse, realistic motions, while Humantomato [32] expands the task to whole-body motion generation, advancing GPT-like models for finer motion sequences. However, limited 3D data hampers model diversity and detail. Tender [55] overcomes this by training on extensive 2D pose data. In contrast, the proposed MotionDiT employs a Diffusion Transformer to better capture both local and global pose relationships, improving human-motion pose generation.", + "bbox": [ + 89, + 613, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. HumanDreamer", + "text_level": 1, + "bbox": [ + 513, + 286, + 676, + 301 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first introduce the proposed MotionVid dataset that facilitates the training of human-motion pose generation. Next, we elaborate on the details of the decoupled Text-to-Pose generation and Pose-to-Video generation.", + "bbox": [ + 511, + 311, + 906, + 372 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. MotionVid", + "text_level": 1, + "bbox": [ + 513, + 380, + 633, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Existing datasets face limitations in training Text-to-Pose tasks. For instance, current 3D datasets [14, 35] primarily capture torso movements, lacking essential full-body keypoints, particularly for the face and hands, which makes them unsuitable for diverse pose generation. Although the Motion-X dataset [29] provides full-body 3D data, its scale is limited to only 81.1K samples, as the expensive annotation of 3D data. For the 2D full-body dataset, the Holistic-Motion2D dataset [55] includes approximately 1M samples, though only 16K of these are publicly accessible. Therefore, there is a pressing need for a comprehensive and publicly available human-motion pose dataset that facilitates the training of Text-to-Pose tasks. To address these limitations, we introduce the largest human-motion pose dataset MotionVid, which contains approximately 1.2M data pairs, with videos of varying resolutions and durations exceeding 64 frames. Besides, we provide a robust data collection pipeline to obtain high-quality textpose pairs from diverse and complex data sources.", + "bbox": [ + 511, + 402, + 906, + 688 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Data Source. We construct a diverse set of text-pose pairs by curating data from various publicly available video datasets, including action recognition datasets Kinetics-400 [23], Kinetics-700 [4], ActivityNet-200 [11], and Something-Something V2 [12], as well as human-related datasets such as CAER [26], DFEW [22], UBody [28], Charades [45], Charades-Ego [46], HAA500 [10], and HMDB51 [25]. This selection totals $\\sim 6.6\\mathrm{M}$ samples. To further enhance scene and subject diversity, we add $3.4\\mathrm{M}$ videos sourced from the internet. Given the complexity of these videos, we perform extensive annotation and cleaning to ensure high-quality data. As shown in Fig. 2, we first use a shot transition detection model [47] to segment the video clips, followed by the video quality filter, data annotation,", + "bbox": [ + 511, + 689, + 908, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12393", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f8b8a27cb1d50360cbea2f8058c48f9ffcd609658e0fc405136e29b822d1abdc.jpg", + "image_caption": [ + "Figure 3. Training pipeline of the proposed Text-to-Pose generation. Pose data are encoded in latent space via the Pose VAE, which are then processed by the proposed MotionDiT, where local feature aggregation and global attention are utilized to capture information from the entire pose sequence. Finally, the LAMA loss is calculated via the proposed CLoP, which enhances the training of MotionDiT." + ], + "image_footnote": [], + "bbox": [ + 94, + 90, + 898, + 287 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "human quality filter, and caption quality filter.", + "bbox": [ + 89, + 369, + 393, + 383 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Video Quality Filter. After acquiring a large number of videos, it is necessary to verify the video quality. Drawing inspiration from SVD's data cleaning strategy [3], we propose a Video Quality Filter to select high-quality video clips. Specifically, we use the GMFlow method [56] to estimate optical flow and filter out videos with insufficient movement intensity. Subsequently, we employ the method described in [1] to detect text regions and remove videos where text occupies a significant portion of the frame. We also utilize LAION-AI's aesthetic predictor [41] to compute aesthetic scores and eliminate videos with low aesthetic quality. Finally, the Laplacian operator [2] is applied to assess blur intensity, and videos deemed excessively blurry are discarded. As a result, approximately $50\\%$ of the data is filtered out (see supplementary materials for more details).", + "bbox": [ + 88, + 385, + 482, + 611 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Data Annotation. We first prompt the Vision Language Model (VLM) [5] to generate action-oriented captions that describe human activities. Besides, for each frame, we extract 2D poses using the DWPose model [58]. This process yields a large number of text-pose pairs. However, the annotated data pairs are unsuitable for training Text-to-Pose tasks. This is due to several factors: the VLM's potential inaccuracy in describing multi-person actions, difficulty in recognizing poses occupying smaller areas in the frame, bias toward static poses from individuals with minimal motion, and limited detail in poses where faces are not visible.", + "bbox": [ + 89, + 612, + 482, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Human Quality Filter. To further refine the dataset, we implement a Human Quality Filter, focusing on selecting data based on human-related characteristics. We calculate the difference between two consecutive frames' 2D poses to filter out sequences with insufficient motion magnitude. Additionally, we compute the ratio of the human detection bounding box to the entire frame to remove clips where the human coverage is too small. At this stage, we focus on", + "bbox": [ + 89, + 779, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "simpler scenarios by selecting scenes with a human count of one and ensuring face visibility for training. Based on the Human Quality Filter, approximately $75\\%$ of the data is filtered out (see supplementary materials for more details).", + "bbox": [ + 511, + 369, + 903, + 429 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Caption Quality Filter. To enhance the quality of caption annotations, we apply a Caption Quality Filter to refine alignment accuracy. Specifically, using the proposed CLoP (see Sec. 3.2 for details), we compute the semantic similarity between text and poses, filtering data based on the similarity score. This ensures that each text annotation semantically aligns with the corresponding 2D pose data, thereby improving overall data quality.", + "bbox": [ + 511, + 430, + 905, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Text-to-Pose", + "text_level": 1, + "bbox": [ + 511, + 561, + 643, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fig. 3 illustrates the training pipeline of the proposed Text-to-Pose generation. In this pipeline, pose data are first encoded into a latent space via the Pose VAE and subsequently processed by the proposed MotionDiT model. Within MotionDiT, local feature aggregation and global attention mechanisms are employed to capture comprehensive information across the entire pose sequence. Finally, LAMA loss is computed using the proposed CLoP, which further refines the training of MotionDiT. In the next, we introduce the Pose VAE, MotionDiT and CLoP.", + "bbox": [ + 511, + 582, + 905, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pose VAE. Variational Autoencoders (VAEs) [17, 24, 59, 67] have been widely applied in 2D image and 3D video processing, but their use for pose data remains underexplored, presenting a rich area for research. We represent a pose sequence as $\\mathbf{p} \\in \\mathbb{R}^{f \\times N \\times 3}$ , where $f$ is the frame count, $N$ the keypoints, and dimension 3 includes the $x$ , $y$ coordinates and confidence scores. Given the sequential structure and frame-to-frame correlations, a 1D convolutional encoder with multi-layer downsampling effectively extracts temporal features, while symmetric decoder-encoder architecture with skip connections aids reconstruction. Confi", + "bbox": [ + 511, + 734, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12394", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "dence scores of each keypoint help mitigate occlusion issues, optimizing the model via KL divergence and reconstruction loss. Further details are available in the supplementary materials.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "MotionDiT. The proposed MotionDiT extends the Diffusion Transformer (DiT) [37] architecture specifically tailored to construct associations between human-motion poses and text control. In the proposed MotionDiT, we incorporate a global attention block to capture global spatial-temporal patterns inherent in human poses. Besides, the MotionDiT involves a local feature aggregation module to strengthen correlations between adjacent joints. Specifically, we first obtain the latent representation from the Pose VAE, which is then processed via a patch-embedding:", + "bbox": [ + 89, + 151, + 482, + 303 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nl _ {p} = \\mathcal {F} _ {\\text {P a t c h}} (\\mathcal {E} (\\mathbf {p})), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 220, + 313, + 482, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{F}_{\\mathrm{Patch}}(\\cdot)$ is the patch embedding operation [33], and $\\mathcal{E}(\\cdot)$ is the Pose VAE encoder. Then we employ Diffusion blocks to process $l_{p}$ . As shown in Fig. 3, in each block, the local feature aggregation module is first utilized to strengthen local feature correlation:", + "bbox": [ + 89, + 340, + 483, + 416 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nl _ {p} = \\mathcal {F} _ {\\mathrm {s a}} \\left(\\mathcal {F} _ {\\mathrm {r e s}} \\left(l _ {p}\\right) + l _ {p}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 426, + 482, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{F}_{\\mathrm{res}}$ is the 1D ResNet Block with kernel size 3, and $\\mathcal{F}_{\\mathrm{sa}}$ is the spatial self attention. For text-based control, textual features obtained from CLIP [40] are then integrated into the network through cross-attention. Besides, following conventional video Diffusion architecture [3, 33], temporal attention is used to ensure continuity of pose features. Notably, it is essential for the model to capture global pose feature information. Although the model already employs spatiotemporal attention mechanisms, certain pose characteristics at one location and time may influence other locations at different times. Thus, global attention is required to extract these global features among internal pose keypoints. Instead of incorporating global attention in every attention module, we decide to apply it to the output of the central layer of the network to reduce computational complexity. Specifically, we reshape the latent output of the middle layer to $z \\in \\mathbb{R}^{(f \\times n) \\times c}$ , where $n$ is the number of points $N$ divided by the down-sampling factor $r$ and $c$ is the number of latent channels. We then perform self-attention across all frames $f$ and all points $n$ . Finally, we reshape the output back to its original dimensions.", + "bbox": [ + 89, + 454, + 482, + 770 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following [33], we employ the noise prediction loss to optimize MotionDiT, thus the model output is:", + "bbox": [ + 89, + 771, + 482, + 801 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\epsilon_ {\\text {p r e d}} = g _ {\\theta} \\left(z _ {t}, t, s\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 220, + 814, + 482, + 830 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $g_{\\theta}$ represents the MotionDiT parameterized by $\\theta$ , $z_{t}$ is the noisy latent variable at time step $t$ , and $s$ denotes the conditional input (e.g., textual features), we compute the noise prediction at time step $t$ . The model learns to predict", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the noise $\\epsilon_{\\mathrm{pred}}$ conditioned on the input. We calculate the mean squared error (MSE) loss between the predicted noise $\\epsilon_{\\mathrm{pred}}$ and the true noise $\\epsilon$ as:", + "bbox": [ + 511, + 90, + 903, + 136 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {d} = \\mathbb {E} _ {t, z _ {0}, \\epsilon} \\left[ \\| \\epsilon - \\epsilon_ {\\text {p r e d}} \\| ^ {2} \\right], \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 142, + 903, + 162 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_d$ represents the loss over all time steps $t$ , true noise $\\epsilon \\sim \\mathcal{N}(0, I)$ , and ground truth latent variable $z_0$ .", + "bbox": [ + 511, + 167, + 903, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "CLoP. Additionally, prior works [27, 60] emphasize the significance of intermediate feature representations, which inspire us to propose a LAtent seMantic Alignment (LAMA) loss aimed at enhancing text-based motion control. We expect this term to provide benefits in two key ways: (1) By aligning the latent space of MotionDiT with a largescale pre-trained motion encoder, we can improve both fidelity and diversity of the generated 2D pose; (2) Aligning with the pre-trained pose encoder that is already aligned with text inherently strengthens the influence of text control on motion generation. Unlike CLIP [40], there are currently no readily available large-scale pre-trained models for aligning motion and text. Leveraging the MotionVid dataset, we introduce Contrastive Language-Motion Pretraining (CLoP), which aligns text with 2D pose. This improves the evaluation of 2D pose and latent semantic alignment. Specifically, inspired by CLIP, we employ a similar mechanism to extract textual features. The extracted text features and 2D pose data are then processed through dedicated text encoder $\\mathcal{F}_e(\\cdot)$ and pose encoder $\\mathcal{F}_p(\\cdot)$ , respectively. The alignment of the two modalities is facilitated by optimizing a contrastive loss function [40]. Our architecture represents a 2D extension of the approach in [38]. Unlike the original model, we utilize CLIP for poses representation as $\\mathbf{p} \\in \\mathbb{R}^{f \\times N \\times 3}$ with its corresponding text e, capturing frame-wise keypoint coordinates and confidence scores instead of 3D features, the CLoP training objective as follows:", + "bbox": [ + 511, + 198, + 906, + 619 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c} = \\ell_ {c e} \\left(\\ell_ {2} \\left(\\mathbf {h} _ {e} \\mathbf {h} _ {p} ^ {T}\\right); y\\right) + \\ell_ {c e} \\left(\\ell_ {2} \\left(\\mathbf {h} _ {p} \\mathbf {h} _ {e} ^ {T}\\right); y\\right) / 2, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 625, + 903, + 645 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\ell_2(\\cdot)$ denote the $\\ell_2$ normalization and $\\ell_{ce}(\\cdot ;y)$ represent the cross-entropy loss with ground truth label $y$ . $\\mathbf{h}_e = \\mathcal{F}_e(\\mathbf{e})\\mathbf{W}_e$ is the projected text embedding, calculated by projection matrix $\\mathbf{W}_e$ . $\\mathbf{h}_p = \\mathcal{F}_p(\\mathbf{p})\\mathbf{W}_p$ is the projected pose embedding.", + "bbox": [ + 511, + 650, + 906, + 726 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on the CLoP, we evaluate the dissimilarity within the feature space (e.g., MSE or cosine similarity metrics), achieving the benefits of the previously mentioned LAMA loss. Rather than using prior methods to extract structural features, we focus on the semantic features of motion, as motion's structural information is simpler while textual information is richer. We also employ a feature projection using a two-layer MLP $g_{\\omega}(\\cdot)$ parameterized by $\\omega$ to project the latent feature space learned by MotionDiT. As a result, the LAMA loss $\\mathcal{L}_f$ can be formulated:", + "bbox": [ + 511, + 726, + 906, + 878 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {f} = d \\left(g _ {\\omega} \\left(\\mathbf {h} _ {d} ^ {l}\\right), \\mathbf {h} _ {p}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 627, + 883, + 903, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12395", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/0a9ae254838681a4081d276bc5a3eabd4ecd2ca240cba8de092d125812f72501.jpg", + "table_caption": [ + "Table 1. Comparison to Other State-of-the-Art Methods on the MotionVid Subset. The metrics demonstrate that our method outperforms others in terms of pose-text alignment and diversity. Bold indicates the best result." + ], + "table_footnote": [], + "table_body": "
MethodFID ↓Rp-top1 ↑Rp-top2 ↑Rp-top3 ↑Diversity ↑MM Dist ↓MM ↑
T2M-GPT [63]496.9070.1760.2990.39564.00042.95014.480
PriorMDM [43]538.9490.1560.2700.36662.33852.55044.135
MLD [6]396.9490.3180.5050.62864.44240.19657.862
Ours149.0070.4510.6380.74368.22032.76160.460
", + "bbox": [ + 145, + 128, + 854, + 222 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{h}_d^l$ is the $l$ -th layer latent representation of MotionDiT, $d(\\cdot ,\\cdot)$ is the CLoP function. Consequently, the overall objective $\\mathcal{L}$ of MotionDiT can be formulated as:", + "bbox": [ + 89, + 244, + 482, + 292 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {d} + \\lambda_ {f} \\mathcal {L} _ {f}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 225, + 301, + 482, + 318 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\lambda_{f}$ is a hyperparameter to trade off the alignment and denoising process.", + "bbox": [ + 89, + 327, + 483, + 358 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3. Pose-to-Video", + "text_level": 1, + "bbox": [ + 89, + 366, + 233, + 380 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We propose a model for Pose-to-Video generation, capable of producing human-motion videos based on an initial frame image and a pose sequence. The model leverages [59] as its backbone, maintaining consistent parameters with the original model, incorporates conditional control inspired by [64], and employs data augmentation techniques from [49] to enhance performance. Notably, all training was conducted on the MotionVid dataset, demonstrating its versatility for both Text-to-Pose and Pose-to-Video tasks. It is worth noting that other Pose-to-Video models, such as [20, 49, 66], are also capable of handling this task. For further details on the model architecture, please refer to the supplementary materials.", + "bbox": [ + 89, + 388, + 483, + 583 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 597, + 223, + 614 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental setup", + "text_level": 1, + "bbox": [ + 89, + 622, + 277, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset. For computational efficiency, we select a representative subset of 50K samples from MotionVid, covering diverse actions like dancing, squatting, and lifting to ensure diversity and expedite experimentation. This subset maintains result representativeness while enabling efficient testing. We also conducted scaling experiments to assess model performance across various dataset sizes. Qualitative visualization experiments utilize a dataset of 1.2M samples.", + "bbox": [ + 89, + 643, + 482, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. For Text-to-Pose, our motion data consists of 128 points per frame, covering the body, face, and hands. The input is structured as $f \\times 128 \\times 3$ , representing $f$ frames, with 128 points in each frame and 3 channels that encode image coordinates (x, y) along with confidence scores. To maintain uniformity across samples, all sequences are cropped to 64 frames. For Pose VAE, we employ a downsampling factor $r = 8$ , yielding a latent representation $z \\in \\mathbb{R}^{f \\times 16 \\times 3}$ . In MotionDiT, we apply a two-", + "bbox": [ + 89, + 763, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "fold downsampling in the patch embedding. The network consists of 13 layers, with text features extracted from the SD2.1 CLIP model [42] as our text encoder, resulting in a conditional embedding $t_{emb} \\in \\mathbb{R}^{1 \\times 1024}$ . The AdamW optimizer is used with a learning rate of $1 \\times 10^{-5}$ . The training is conducted on 8 H20 GPUs, with inference performed on a single H20.", + "bbox": [ + 511, + 244, + 906, + 351 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. To evaluate generated motion data quality, several metrics are employed, consistent with previous works [6, 15, 43, 52, 55, 63]. Frechet Inception Distance (FID) [18] assesses distributional similarity between ground truth and generated motions, using features from a custom-trained CLOP model. MultiModality evaluates diversity for a single text input, while Diversity [15] calculates variance across all motions to quantify output diversity. R-precision measures top-1, top-2, and top-3 recall rates in motion retrieval, evaluating alignment accuracy. MultiModality Distance gauges the distance between textual inputs and generated motions, indicating adherence to prompts. Further details are available in the supplementary materials.", + "bbox": [ + 511, + 354, + 908, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Results on Text-to-Pose", + "text_level": 1, + "bbox": [ + 511, + 585, + 728, + 599 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Currently, Tender [55] is the only method for generating 2D poses based on text but its code and dataset are unavailable. Thus, we select three SOTA 3D motion generation methods for comparison: T2M-GPT [62], PriorMDM [43], and MLD [6], converting their inputs to 2D poses. All methods are trained on the MotionVid subset with proposed settings to standardize evaluation metrics. Comparison results (Tab. 1) show our model outperforms others, achieving $62.4\\%$ improvement in FID, $41.8\\%$ in Rp-top1, $26.3\\%$ in Rp-top2, and $18.3\\%$ in Rp-top3. These results indicate our model generates semantically aligned outputs with high diversity.", + "bbox": [ + 511, + 609, + 906, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The visual comparison with other methods is presented in Fig. 4, and the results demonstrate a significant improvement over existing models. Our method produces poses that better align with the textual constraints, with keypoints remaining intact and exhibiting minimal motion jitter. A more detailed visual comparison can be found in the supplementary materials.", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12396", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0093a61e54227de98980f961005121e3c477e6411f6e7dcb35073869fc63688f.jpg", + "table_caption": [ + "Table 2. The ablation study presents four configurations, progressively adding components of Local, Global, and LAMA loss to the original model. As we move from the initial configuration to the fully enhanced model, performance metrics consistently improve, highlighting the positive impact of each additional component." + ], + "table_footnote": [], + "table_body": "
MethodFID ↓Rp-top1 ↑Rp-top2 ↑Rp-top3 ↑Diversity ↑MM Dist ↓MM ↑
Vanilla DiT283.0910.3670.5610.68266.58737.04621.461
DiT+Local183.2130.4150.6060.72167.59134.29141.890
DiT+Local+Global162.0220.4330.6200.73067.35733.26957.848
DiT+Local+Global+LAMA149.0070.4510.6380.74368.22032.76160.460
", + "bbox": [ + 91, + 142, + 903, + 238 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A man is holding his head in his hands and continues to do so while looking down at them.", + "bbox": [ + 116, + 253, + 415, + 273 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A woman is dancing, moving her arms and legs around.", + "bbox": [ + 450, + 253, + 720, + 265 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b15f958cd9d769ce18e89f1e707a24f5f3dbc1f38caddf77df9fbf9862682219.jpg", + "image_caption": [ + "Figure 4. Visualization results compared to SOTA Text-to-Pose methods. The results demonstrate that our model significantly outperforms other models. Our method generates poses that are more consistent with the text constraints, with keypoints maintaining their integrity and minimal motion jitter. For a better visual comparison, please refer to the supplementary materials." + ], + "image_footnote": [], + "bbox": [ + 93, + 276, + 439, + 500 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7f254fa475fbd039c70fbdacc559df5196b80fc5204c1f5c278fff1447034b64.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 450, + 275, + 890, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 588, + 254, + 603 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To validate the efficacy of the three modules proposed in the MotionDiT (i.e., the local feature aggregation, the global attention, and the LAMA loss), we conduct a series of ablation studies. The outcomes are summarized as follows:", + "bbox": [ + 89, + 611, + 482, + 670 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Local Feature Aggregation. Removing the local attention block and integrating a DiT structure similar to Latte [33] in each attention mechanism, we showcase our method's efficacy in capturing local details. The performance improvements, shown in Tab. 2, confirm the robustness of our local feature extraction approach.", + "bbox": [ + 89, + 672, + 482, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Global Attention Block. To assess the impact of our global attention mechanism, we perform an ablation test where this component is removed from the model. The findings, detailed in Tab. 2, indicate a notable decline in performance, which underscores the importance of the global attention mechanism in enhancing the model's ability to capture long-range dependencies and global context.", + "bbox": [ + 89, + 763, + 482, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "LAMA Loss. We investigate the role of the semantic alignment loss by training MotionDiT without it. The analysis in", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 2 shows a significant performance drop, validating the LAMA loss's critical role in guiding optimization towards accurate and stable solutions. These ablation studies provide strong evidence for the proposed algorithms' effectiveness and necessity in enhancing model performance across various aspects.", + "bbox": [ + 511, + 588, + 906, + 679 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Scaling Law Results", + "text_level": 1, + "bbox": [ + 511, + 691, + 705, + 707 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We expand the training set size, and as shown in Tab. 3, our metrics improved with the increase in data volume. This demonstrates the validity of using large-scale 2D pose data and suggests that model performance can be further enhanced through an efficient and cost-effective data collection pipeline.", + "bbox": [ + 511, + 714, + 905, + 805 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Qualitative Results on Pose-to-Video", + "text_level": 1, + "bbox": [ + 511, + 816, + 828, + 832 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In our qualitative experiments, we use the generated motion sequences as templates and provide them to the Pose-to-Video model. The specific video results are shown in Fig. 5. Compared to state-of-the-art text-to-video models,", + "bbox": [ + 511, + 840, + 906, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12397", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9794ba2c9d729f7714a123d1c105c760f5e22d64f6e735046839a90e5c5d2e2d.jpg", + "image_caption": [ + "Figure 5. Visualization results compared to SOTA Text-to-Video methods. Mochi1 [50] and CogVideoX [59] exhibit issues such as body distortion, weak motion continuity, and neglecting facial generation. In contrast, HumanDreamer is able to generate more coherent and consistent videos with smoother transitions and better attention to details such as facial expressions. For a better visual comparison, please refer to the supplementary materials." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 449, + 243 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cf83992ebd1916a8a86ed0092e94b113f5009d74d67720e875a104f1b611ece1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 455, + 89, + 893, + 243 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b328feb5f7d525edc1034586214cdbd01dd1bc850d580c32e9dce797e25a3148.jpg", + "table_caption": [ + "Table 3. By increasing the amount of training data, we observe an improvement in model performance, which validates the potential for rapid scalability using large-scale 2D motion data." + ], + "table_footnote": [], + "table_body": "
DatasizeRp-top1 ↑Rp-top2 ↑Rp-top3 ↑MM Dist ↓
50K(4%)0.4510.6380.74332.761
250K(20%)0.4520.6410.75132.666
500K(40%)0.4640.6570.76632.399
1M(80%)0.4920.6710.76931.116
1.25M(100%)0.5130.6940.79130.139
", + "bbox": [ + 94, + 388, + 480, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "our model exhibits larger and more dynamic movements, with the characters' actions more accurately reflecting the textual descriptions.", + "bbox": [ + 89, + 502, + 483, + 547 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4be80cf722b8d40a1101efcbb61b9c1f64e314890ec4d99a43bdff41a57edb16.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 93, + 559, + 478, + 626 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e22d0e599981c539142a464e74693c402e4052585c33e9767107d6cb4cf2fd25.jpg", + "image_caption": [ + "Figure 6. The visualization results demonstrate that using the initial frame pose, different prompt texts can guide the generation of distinct pose sequences." + ], + "image_footnote": [], + "bbox": [ + 94, + 627, + 477, + 693 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Other Downstream Tasks", + "text_level": 1, + "bbox": [ + 89, + 772, + 323, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In addition to the Pose-to-Video task, the proposed Text-to-Pose component can also be utilized for pose sequence prediction, 2D-3D motion lifting tasks.", + "bbox": [ + 89, + 794, + 482, + 840 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Pose Sequence Prediction. In scenarios with partial pose sequences, the Text-to-Pose model can infer and generate missing parts by conditioning on existing poses and textual movement descriptions. This enables the synthesis of co-", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "herent pose continuations. Given initial and final frames, the model predicts intermediate pose states, completing the sequence (Fig. 6). This is especially useful in animation for supplementing incomplete pose data.", + "bbox": [ + 511, + 338, + 906, + 398 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "2D-3D Motion Lifting. Models like [70] can enhance motion data dimensionality, converting 2D motions into realistic 3D motions (Fig. 7). With textual descriptions specifying depth or spatial aspects, the Text-to-Pose system generates richer 3D motion representations. This is crucial for virtual reality (VR) and augmented reality (AR) environments, where realistic 3D motion significantly improves user experience.", + "bbox": [ + 511, + 400, + 908, + 521 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8704c8d46a9d0423f47135731a6dc66a7d570932a0b85f220fc02a42437719f4.jpg", + "image_caption": [ + "Figure 7. Visualization results on 2D-3D Motion Lifting." + ], + "image_footnote": [], + "bbox": [ + 516, + 537, + 903, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Discussion and Conclusion", + "text_level": 1, + "bbox": [ + 511, + 689, + 763, + 705 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this study, we present HumanDreamer, a pioneering decoupled framework for generating human-motion videos that merges text control flexibility with pose guidance controllability. Utilizing MotionVid, the largest dataset for human-motion pose generation, we train MotionDiT for producing structured poses. We introduce LAMA loss to improve semantic alignment, ensuring coherent outputs.", + "bbox": [ + 511, + 715, + 906, + 823 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Experimental results indicate that using generated poses in Pose-to-Video yields high-quality, diverse human-motion videos, surpassing current benchmarks. These findings confirm the effectiveness and adaptability of our decoupled framework, facilitating versatile video generation.", + "bbox": [ + 511, + 825, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12398", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Youngmin Baek, Bado Lee, Dongyoon Han, Sangdoo Yun, and Hwalsuk Lee. Character region awareness for text detection. In CVPR, 2019. 4", + "[2] Raghav Bansal, Gaurav Raj, and Tanupriya Choudhury. Blur image detection using laplacian operator and open-cv. In SMART, 2016. 4", + "[3] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2, 4, 5", + "[4] Joao Carreira, Eric Noland, Chloe Hillier, and Andrew Zisserman. A short note on the kinetics-700 human action dataset, 2022. 3", + "[5] Lin Chen, Xilin Wei, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Bin Lin, Zhenyu Tang, et al. Sharegpt4video: Improving video understanding and generation with better captions. arXiv preprint arXiv:2406.04325, 2024. 4", + "[6] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, and Gang Yu. Executing your commands via motion diffusion in latent space. In CVPR, 2023. 2, 3, 6", + "[7] Yuxin Chen, Ziqi Zhang, Chunfeng Yuan, Bing Li, Ying Deng, and Weiming Hu. Channel-wise topology refinement graph convolution for skeleton-based action recognition. In ICCV, 2021. 2", + "[8] Ziwei Chen, Qiang Li, Xiaofeng Wang, and Wankou Yang. Liftedcl: Lifting contrastive learning for human-centric perception. In ICLR, 2023. 2", + "[9] Joseph Cho, Fachrina Dewi Puspitasari, Sheng Zheng, Jingyao Zheng, Lik-Hang Lee, Tae-Ho Kim, Choong Seon Hong, and Chaoning Zhang. Sora as an agi world model? a complete survey on text-to-video generation. arXiv preprint arXiv:2403.05131, 2024. 2", + "[10] Jihoon Chung, Cheng hsin Wu, Hsuan ru Yang, Yu-Wing Tai, and Chi-Keung Tang. Haa500: Human-centric atomic action dataset with curated videos. In ICCV, 2011. 3", + "[11] Bernard Ghanem Fabian Caba Heilbron, Victor Escorcia and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In CVPR, 2015. 3", + "[12] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The\" something something\" video database for learning and evaluating visual common sense. In ICCV, 2017. 3", + "[13] Yanchen Guan, Haicheng Liao, Zhenning Li, Guohui Zhang, and Chengzhong Xu. World models for autonomous driving: An initial survey. arXiv preprint arXiv:2403.02622, 2024. 2", + "[14] Chuan Guo, Xinxin Zuo, Sen Wang, Shihao Zou, Qingyao Sun, Annan Deng, Minglun Gong, and Li Cheng. Action2motion: Conditioned generation of 3d human motions. In ACM MM, 2020. 3", + "[15] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In CVPR, 2022. 3, 6" + ], + "bbox": [ + 93, + 114, + 482, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Agrim Gupta, Lijun Yu, Kihyuk Sohn, Xiuye Gu, Meera Hahn, Li Fei-Fei, Irfan Essa, Lu Jiang, and José Lezama. Photorealistic video generation with diffusion models. arXiv preprint arXiv:2312.06662, 2023. 2", + "[17] Louay Hazami, Rayhane Mama, and Ragavan Thurairatnam. Efficient-vdvae: Less is more. arXiv preprint arXiv:2203.13751, 2022. 4", + "[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS, 2017. 6", + "[19] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 2", + "[20] Li Hu. Animate anyone: Consistent and controllable image-to-video synthesis for character animation. In CVPR, 2024. 2, 6", + "[21] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. TPAMI, 2013. 2", + "[22] Xingxun Jiang, Yuan Zong, Wenming Zheng, Chuangao Tang, Wanchuang Xia, Cheng Lu, and Jiateng Liu. Dfew: A large-scale database for recognizing dynamic facial expressions in the wild. In ACM MM, 2020. 3", + "[23] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, Mustafa Suleyman, and Andrew Zisserman. The kinetics human action video dataset, 2017. 3", + "[24] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 4", + "[25] H. Kuehne, H. Jhuang, E. Garrote, T. Poggio, and T. Serre. HMDB: a large video database for human motion recognition. In ICCV, 2011. 3", + "[26] Jiyoung Lee, Seungryong Kim, Sunok Kim, Jungin Park, and Kwanghoonn Sohn. Context-aware emotion recognition networks. In ICCV, 2019. 3", + "[27] Alexander C Li, Mihir Prabhudesai, Shivam Duggal, Ellis Brown, and Deepak Pathak. Your diffusion model is secretly a zero-shot classifier. In ICCV, 2023. 5", + "[28] Jing Lin, Ailing Zeng, Haoqian Wang, Lei Zhang, and Yu Li. One-stage 3d whole-body mesh recovery with component aware transformer. CVPR, 2023. 3", + "[29] Jing Lin, Ailing Zeng, Shunlin Lu, Yuanhao Cai, Ruimao Zhang, Haoqian Wang, and Lei Zhang. Motion-x: A large-scale 3d expressive whole-body human motion dataset. NeurIPS, 2024. 2, 3", + "[30] Jun Liu, Amir Shahroudy, Mauricio Perez, Gang Wang, Ling-Yu Duan, and Alex C Kot. Ntu rgb+ d 120: A large-scale benchmark for 3d human activity understanding. TPAMI, 2019. 2", + "[31] Yixin Liu, Kai Zhang, Yuan Li, Zhiling Yan, Chujie Gao, Ruoxi Chen, Zhengqing Yuan, Yue Huang, Hanchi Sun, Jianfeng Gao, et al. Sora: A review on background, technology, limitations, and opportunities of large vision models. arXiv preprint arXiv:2402.17177, 2024. 2" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12399", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[32] Shunlin Lu, Ling-Hao Chen, Ailing Zeng, Jing Lin, Ruimao Zhang, Lei Zhang, and Heung-Yeung Shum. Humantomato: Text-aligned whole-body motion generation. arXiv preprint arXiv:2310.12978, 2023. 2, 3", + "[33] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 5, 7", + "[34] Yue Ma, Yingqing He, Xiaodong Cun, Xintao Wang, Siran Chen, Xiu Li, and Qifeng Chen. Follow your pose: Pose-guided text-to-video generation using pose-free videos. In AAAI, 2024. 2", + "[35] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In ICCV, 2019. 3", + "[36] Haomiao Ni, Changhao Shi, Kai Li, Sharon X Huang, and Martin Renqiang Min. Conditional image-to-video generation with latent flow diffusion models. In CVPR, 2023. 2", + "[37] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, 2023. 5", + "[38] Mathis Petrovich, Michael J. Black, and Gül Varol. TMR: Text-to-motion retrieval using contrastive 3D human motion synthesis. In ICCV, 2023. 5", + "[39] Matthias Plappert, Christian Mandery, and Tamim Asfour. The kit motion-language dataset. Big data, 2016. 3", + "[40] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 3, 5", + "[41] Beaumont Romain and Schuhmann Christoph. Laion aesthetics predictor v1. 2022. 4", + "[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 6", + "[43] Yoni Shafir, Guy Tevet, Roy Kapon, and Amit Haim Bermano. Human motion diffusion as a generative prior. In ICLR, 2024. 6", + "[44] Amir Shahroudy, Jun Liu, Tian-Tsong Ng, and Gang Wang. Ntu $\\mathrm{rgb + }$ d: A large scale dataset for 3d human activity analysis. In CVPR, 2016. 2", + "[45] Gunnar A. Sigurdsson, Gül Varol, Xiaolong Wang, Ali Farhadi, Ivan Laptev, and Abhinav Gupta. Hollywood in homes: Crowdsourcing data collection for activity understanding, 2016. 3", + "[46] Gunnar A. Sigurdsson, Abhinav Gupta, Cordelia Schmid, Ali Farhadi, and Karteek Alahari. Charades-ego: A large-scale dataset of paired third and first person videos. In ArXiv, 2018. 3", + "[47] Tomáš Souček and Jakub Lokoc. Transnet v2: An effective deep network architecture for fast shot transition detection. arXiv preprint arXiv:2008.04838, 2020. 3", + "[48] Rui Sun, Yumin Zhang, Tejal Shah, Jiahao Sun, Shuoying Zhang, Wenqi Li, Haoran Duan, Bo Wei, and Rajiv Ranjan. From sora what we can see: A survey of text-to-video generation. arXiv preprint arXiv:2405.10674, 2024. 2" + ], + "bbox": [ + 91, + 90, + 485, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Shuai Tan, Biao Gong, Xiang Wang, Shiwei Zhang, Dandan Zheng, Ruobing Zheng, Kecheng Zheng, Jingdong Chen, and Ming Yang. Animate-x: Universal character image animation with enhanced motion representation. arXiv preprint arXiv:2410.10306, 2024. 2, 6", + "[50] Genmo Team. Mochi 1. https://github.com/genmoai/models, 2024.8", + "[51] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. In ECCV, 2022. 2, 3", + "[52] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In ICLR, 2023. 2, 3, 6", + "[53] Xiang Wang, Shiwei Zhang, Changxin Gao, Jiayu Wang, Xiaoqiang Zhou, Yingya Zhang, Luxin Yan, and Nong Sang. Unianimate: Taming unified video diffusion models for consistent human image animation. arXiv preprint arXiv:2406.01188, 2024. 2", + "[54] Xiaofeng Wang, Kang Zhao, Feng Liu, Jiayu Wang, Guosheng Zhao, Xiaoyi Bao, Zheng Zhu, Yingya Zhang, and Xingang Wang. Egovid-5m: A large-scale video-action dataset for egocentric video generation. arXiv preprint arXiv:2411.08380, 2024. 2", + "[55] Yuan Wang, Zhao Wang, Junhao Gong, Di Huang, Tong He, Wanli Ouyang, Jile Jiao, Xuetao Feng, Qi Dou, Shixiang Tang, et al. Holistic-motion2d: Scalable whole-body human motion generation in 2d space. arXiv preprint arXiv:2406.11253, 2024. 2, 3, 6", + "[56] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, and Dacheng Tao. Gmflow: Learning optical flow via global matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8121-8130, 2022. 4", + "[57] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In AAAI, 2018. 2", + "[58] Zhendong Yang, Ailing Zeng, Chun Yuan, and Yu Li. Effective whole-body pose estimation with two-stages distillation. In ICCV, 2023. 4", + "[59] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 2, 4, 6, 8", + "[60] Sihyun Yu, Sangkyung Kwak, Huiwon Jang, Jongheon Jeong, Jonathan Huang, Jinwoo Shin, and Saining Xie. Representation alignment for generation: Training diffusion transformers is easier than you think, 2024. 5", + "[61] Yan Zeng, Guoqiang Wei, Jiani Zheng, Jiaxin Zou, Yang Wei, Yuchen Zhang, and Hang Li. Make pixels dance: High-dynamic video generation. arXiv preprint arXiv:2311.10982, 2023. 2", + "[62] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Shaoli Huang, Yong Zhang, Hongwei Zhao, Hongtao Lu, and Xi Shen. T2m-gpt: Generating human motion from textual descriptions with discrete representations. In CVPR, 2023. 6", + "[63] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Yong Zhang, Hongwei Zhao, Hongtao Lu, Xi Shen, and Ying" + ], + "bbox": [ + 514, + 90, + 905, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12400", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Shan. Generating human motion from textual descriptions with discrete representations. In CVPR, 2023. 2, 3, 6", + "[64] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 6", + "[65] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001, 2022.3", + "[66] Yang Zhang, Jiaxi Gu, Li-Wen Wang, Han Wang, Junqi Cheng, Yuefeng Zhu, and Fangyuan Zou. Mimic motion: High-quality human motion video generation with confidence-aware pose guidance. arXiv preprint arXiv:2406.19680, 2024. 2, 6", + "[67] Sijie Zhao, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Muyao Niu, Xiaoyu Li, Wenbo Hu, and Ying Shan. Cv-vae: A compatible video vae for latent generative video models. https://arxiv.org/abs/2405.20279, 2024. 4", + "[68] Weiyu Zhao, Liangxiao Hu, and Shengping Zhang. Diffugesture: Generating human gesture from two-person dialogue with diffusion models. In ICMI, 2023. 2, 3", + "[69] Shenhao Zhu, Junming Leo Chen, Zuozhuo Dai, Qingkun Su, Yinghui Xu, Xun Cao, Yao Yao, Hao Zhu, and Siyu Zhu. Champ: Controllable and consistent human image animation with 3d parametric guidance. arXiv preprint arXiv:2403.14781, 2024. 2", + "[70] Wentao Zhu, Xiaoxuan Ma, Zhaoyang Liu, Libin Liu, Wayne Wu, and Yizhou Wang. Motionbert: A unified perspective on learning human motion representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 8", + "[71] Zheng Zhu, Xiaofeng Wang, Wangbo Zhao, Chen Min, Ni-anchen Deng, Min Dou, Yuqi Wang, Botian Shi, Kai Wang, Chi Zhang, et al. Is sora a world simulator? a comprehensive survey on general world models and beyond. arXiv preprint arXiv:2405.03520, 2024. 2" + ], + "bbox": [ + 91, + 90, + 482, + 599 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "12401", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/HumanDreamer_ Generating Controllable Human-Motion Videos via Decoupled Generation/8f919b44-e98e-44f2-a774-268fc39345bc_model.json b/2025/HumanDreamer_ Generating Controllable Human-Motion Videos via Decoupled Generation/8f919b44-e98e-44f2-a774-268fc39345bc_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b3207685b6aa5ad9dfc977e040acc873ef70a465 --- /dev/null +++ b/2025/HumanDreamer_ Generating Controllable Human-Motion Videos via Decoupled Generation/8f919b44-e98e-44f2-a774-268fc39345bc_model.json @@ -0,0 +1,2664 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.021 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + }, + { + "type": "header", + "bbox": [ + 0.325, + 0.017, + 0.725, + 0.032 + ], + "angle": 0, + "content": "Except for this watermark, it is identical to the accepted version;" + }, + { + "type": "header", + "bbox": [ + 0.294, + 0.032, + 0.756, + 0.047 + ], + "angle": 0, + "content": "the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.167, + 0.131, + 0.833, + 0.177 + ], + "angle": 0, + "content": "HumanDreamer: Generating Controllable Human-Motion Videos via Decoupled Generation" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.203, + 0.949, + 0.24 + ], + "angle": 0, + "content": "Boyuan Wang\\(^{1,2,3*}\\), Xiaofeng Wang\\(^{1,2,3*}\\), Chaojun Ni\\(^{4,5}\\), Guosheng Zhao\\(^{1,2,3}\\), Zhiqin Yang\\(^{6}\\), Zheng Zhu\\(^{4\\dagger}\\), Muyang Zhang\\(^{1,2}\\), Yukun Zhou\\(^{4}\\), Xinze Chen\\(^{4}\\), Guan Huang\\(^{4}\\), Lihong Liu\\(^{1}\\), Xingang Wang\\(^{1,3\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.276, + 0.239, + 0.775, + 0.257 + ], + "angle": 0, + "content": "\\(^{1}\\)Institute of Automation, Chinese Academy of Sciences, China" + }, + { + "type": "text", + "bbox": [ + 0.187, + 0.257, + 0.86, + 0.275 + ], + "angle": 0, + "content": "\\(^{2}\\)School of Artificial Intelligence, University of Chinese Academy of Sciences, China" + }, + { + "type": "text", + "bbox": [ + 0.276, + 0.275, + 0.774, + 0.293 + ], + "angle": 0, + "content": "\\(^{3}\\)Luoyang Institute for Robot and Intelligent Equipment, China" + }, + { + "type": "text", + "bbox": [ + 0.185, + 0.293, + 0.855, + 0.312 + ], + "angle": 0, + "content": "\\(^{4}\\)GigaAI, China \\(^{5}\\)Peking University, China \\(^{6}\\)The Chinese University of Hong Kong" + }, + { + "type": "image", + "bbox": [ + 0.159, + 0.363, + 0.182, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.188, + 0.365, + 0.227, + 0.38 + ], + "angle": 0, + "content": "Text" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.35, + 0.562, + 0.373 + ], + "angle": 0, + "content": "A man is performing a yoga pose on a mat, and he is seen moving his legs and arms in different positions." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.373, + 0.561, + 0.4 + ], + "angle": 0, + "content": "A man is speaking to the camera and then looking off into the distance. Later, he is speaking to the camera again." + }, + { + "type": "image", + "bbox": [ + 0.576, + 0.364, + 0.638, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.645, + 0.367, + 0.78, + 0.382 + ], + "angle": 0, + "content": "HumanDreamer" + }, + { + "type": "image_caption", + "bbox": [ + 0.397, + 0.403, + 0.593, + 0.417 + ], + "angle": 0, + "content": "Text-to-Pose Generation" + }, + { + "type": "image", + "bbox": [ + 0.149, + 0.417, + 0.264, + 0.486 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.265, + 0.417, + 0.38, + 0.486 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.381, + 0.417, + 0.496, + 0.486 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.497, + 0.417, + 0.613, + 0.486 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.614, + 0.417, + 0.727, + 0.486 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.728, + 0.417, + 0.844, + 0.486 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.148, + 0.492, + 0.248, + 0.581 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.249, + 0.492, + 0.346, + 0.581 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.347, + 0.492, + 0.446, + 0.581 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.447, + 0.492, + 0.545, + 0.581 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.546, + 0.492, + 0.644, + 0.581 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.645, + 0.492, + 0.744, + 0.581 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.745, + 0.492, + 0.844, + 0.581 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.391, + 0.591, + 0.6, + 0.605 + ], + "angle": 0, + "content": "Pose-to-Video Generation" + }, + { + "type": "image", + "bbox": [ + 0.149, + 0.606, + 0.264, + 0.673 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.265, + 0.606, + 0.38, + 0.673 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.381, + 0.606, + 0.496, + 0.673 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.497, + 0.606, + 0.613, + 0.673 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.614, + 0.606, + 0.727, + 0.673 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.728, + 0.606, + 0.844, + 0.673 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.149, + 0.678, + 0.251, + 0.77 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.252, + 0.678, + 0.349, + 0.77 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.35, + 0.678, + 0.448, + 0.77 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.45, + 0.678, + 0.548, + 0.77 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.549, + 0.678, + 0.647, + 0.77 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.648, + 0.678, + 0.746, + 0.77 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.747, + 0.678, + 0.846, + 0.77 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.79, + 0.908, + 0.82 + ], + "angle": 0, + "content": "Figure 1. Illustration of HumanDreamer. The human-motion video generation is decoupled into two steps: Text-to-Pose generation and Pose-to-Video generation. The decoupled process integrates the flexibility of text control and the controllability of pose guidance." + }, + { + "type": "title", + "bbox": [ + 0.671, + 0.831, + 0.751, + 0.848 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.863, + 0.366, + 0.876 + ], + "angle": 0, + "content": "*These authors contributed equally to this work." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.876, + 0.483, + 0.888 + ], + "angle": 0, + "content": "† Corresponding authors. zhengzhu@ieee.org, xingang.wang@ia.ac.cn." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.434, + 0.9 + ], + "angle": 0, + "content": "\\(\\ddagger\\)Project Page: https://humandreamer.github.io/" + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.863, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Human-motion video generation has been a challenging task, primarily due to the difficulty inherent in learning hu" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12391" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.094, + 0.092, + 0.482, + 0.424 + ], + "angle": 0, + "content": "man body movements. While some approaches have attempted to drive human-centric video generation explicitly through pose control, these methods typically rely on poses derived from existing videos, thereby lacking flexibility. To address this, we propose HumanDreamer, a decoupled human video generation framework that first generates diverse poses from text prompts and then leverages these poses to generate human-motion videos. Specifically, we propose MotionVid, the largest dataset for human-motion pose generation. Based on the dataset, we present MotionDiT, which is trained to generate structured human-motion poses from text prompts. Besides, a novel LAMA loss is introduced, which together contribute to a significant improvement in FID by \\(62.4\\%\\), along with respective enhancements in R-precision for top1, top2, and top3 by \\(41.8\\%\\), \\(26.3\\%\\), and \\(18.3\\%\\), thereby advancing both the Text-to-Pose control accuracy and FID metrics. Our experiments across various Pose-to-Video baselines demonstrate that the poses generated by our method can produce diverse and high-quality human-motion videos. Furthermore, our model can facilitate other downstream tasks, such as pose sequence prediction and 2D-3D motion lifting." + }, + { + "type": "title", + "bbox": [ + 0.095, + 0.453, + 0.221, + 0.468 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.479, + 0.482, + 0.658 + ], + "angle": 0, + "content": "Generating human-motion videos remains a particularly challenging task due to the inherent complexity of modeling human body movements. Despite current advancements in generative modeling [9, 13, 31, 48, 71], state-of-the-art video generation models [3, 16, 19, 36, 59, 61], equipped with billions of parameters and trained on millions of video and image data, still struggle to capture human body movements, frequently resulting in fragmented or unrealistic portrayals. This limitation is exacerbated when controlling human videos via textual conditions, highlighting the fundamental challenge of directly mapping text prompts to human visual data." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.66, + 0.482, + 0.795 + ], + "angle": 0, + "content": "To enhance the generation quality, approaches such asAnimate Anyone [20], UniAnimate [53], Mimic Motion [66], Champ [69] andAnimate-X [49] explicitly generate human-motion videos through pose control. By utilizing the pose-guided generation, these methods effectively reduce the complexities associated with human-motion videos. However, a notable limitation of this approach is its reliance on human-motion poses derived from existing videos, which restricts the flexibility." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.796, + 0.482, + 0.9 + ], + "angle": 0, + "content": "Therefore, we propose HumanDreamer, a decoupled human-motion video generation framework that first generates human-motion poses from text prompts and subsequently produces human-motion videos based on the generated poses. The motivation behind this decoupled approach lies in that Text-to-Pose presents a more manageable search space compared to the direct learning of text-to-pixel repre" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.529 + ], + "angle": 0, + "content": "sensation. As shown in Fig. 1, this decoupled framework facilitates a more effective generation of human movements. Additionally, the proposed HumanDreamer utilizes text as the input, offering greater flexibility than [20, 34, 49, 53, 66] that directly relies on pre-defined poses. Specifically, we construct a 1.2 million text-pose pairs dataset MotionVid, which is the largest dataset for human-motion pose generation. To ensure dataset quality, a comprehensive data cleaning pipeline is introduced, where the cleaning factors involve body movement amplitude, human presence duration, facial visibility, and the proportion of human figures. This rigorous cleaning guarantees that the dataset is reliable for training Text-to-Pose tasks. Based on the dataset, we propose the MotionDiT to generate structured human-motion poses from text prompts. The MotionDiT integrates a global attention block designed to extract the global characteristics of the entire pose sequence, alongside a local feature aggregation mechanism that captures information from adjacent pose points, effectively combining global and local perspectives to enhance the quality and coherence of the generated poses. Additionally, the LAMA loss is utilized in the latent space to align the semantic features of the ground truth motion with those produced by MotionDiT, which not only improves the interpretability of the model but also boosts its overall performance. The introduced techniques collectively result in a notable \\(62.4\\%\\) enhancement in FID, accompanied by respective increases in R-precision for the top1, top2, and top3 categories by \\(41.8\\%\\), \\(26.3\\%\\), and \\(18.3\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.531, + 0.905, + 0.787 + ], + "angle": 0, + "content": "The primary contributions of this work are as follows: (1) We present HumanDreamer, the first decoupled framework for human-motion video generation, which integrates the flexibility of text control with the controllability of pose guidance. (2) We propose the MotionVid, the largest dataset for human-motion pose generation. In MotionVid, we conduct a comprehensive data annotation and data cleaning to ensure data reliability. (3) The MotionDiT is introduced to generate diverse human-motion poses under text control. To enhance the generation process, we propose the LAMA loss, which significantly improves pose fidelity and diversity. (4) Through extensive experiments, we show that MotionDiT and the LAMA loss improve control accuracy and FID by \\(62.4\\%\\), with top-k R-precision gains of \\(41.8\\%\\), \\(26.3\\%\\), and \\(18.3\\%\\) for top-1, top-2, and top-3. The generated poses also support downstream tasks like pose sequence prediction and 2D-to-3D motion enhancement." + }, + { + "type": "title", + "bbox": [ + 0.517, + 0.806, + 0.654, + 0.821 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.517, + 0.832, + 0.775, + 0.847 + ], + "angle": 0, + "content": "2.1. Human-motion Pose Datasets" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.856, + 0.905, + 0.9 + ], + "angle": 0, + "content": "Human motion datasets are essential for the development of human-centric perception [7, 8, 21, 30, 44, 57] or generation [6, 29, 32, 51, 52, 54, 55, 63, 68] tasks. Specifically," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12392" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.089, + 0.9, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.219, + 0.908, + 0.263 + ], + "angle": 0, + "content": "Figure 2. The data cleaning and annotation pipeline for MotionVid begins with raw data sourced from public datasets and the internet, which is then segmented into video clips. To ensure high-quality data, we apply video quality filter, data annotation, human quality filter, and caption quality filter." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.288, + 0.485, + 0.56 + ], + "angle": 0, + "content": "the KIT Motion-Language Dataset [39] provides sequence-level descriptions to support multi-modal motion tasks. In contrast, the HumanML3D [15] dataset, which is built upon the AMASS [35] and HumanAct12 [14] datasets, offers richer textual annotations and a wider variety of activities. Additionally, Motion-X [29] establishes a robust annotation protocol and is the first to compile a comprehensive fine-grained 3D whole-body motion dataset derived from extensive scene captures. As a result, existing text-driven 3D motion datasets often lack sufficient volume and diversity, which limits their scalability. A dataset that shares similarities with ours is Holistic-Motion2D [55], which features 1M text-pose pairs but only 16K of these are publicly accessible. However, the proposed MotionVid significantly surpasses this with a larger and more diverse set of 1.2M text-pose data, coupled with a more thorough data cleaning process that ensures the dataset's reliability for training Text-to-Pose tasks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.587, + 0.447, + 0.602 + ], + "angle": 0, + "content": "2.2. Human-motion Pose Generation Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Text-driven human motion generation, which converts textual descriptions into motion sequences, has gained significant attention in recent years. Early approaches focus on aligning motion and text through shared latent spaces. For instance, MotionCLIP [51] enhances autoencoder generalization by aligning the latent space with the expressive CLIP [40] embedding. T2M-GPT [63] frames the task as predicting discrete motion indices. More recently, Diffusion-based models, including MDM [52], MotionDiffuse [65], and DiffGesture [68], have set new performance standards. MLD [6] introduces a latent Diffusion model that generates diverse, realistic motions, while Humantomato [32] expands the task to whole-body motion generation, advancing GPT-like models for finer motion sequences. However, limited 3D data hampers model diversity and detail. Tender [55] overcomes this by training on extensive 2D pose data. In contrast, the proposed MotionDiT employs a Diffusion Transformer to better capture both local and global pose relationships, improving human-motion pose generation." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.287, + 0.677, + 0.303 + ], + "angle": 0, + "content": "3. HumanDreamer" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.312, + 0.907, + 0.373 + ], + "angle": 0, + "content": "In this section, we first introduce the proposed MotionVid dataset that facilitates the training of human-motion pose generation. Next, we elaborate on the details of the decoupled Text-to-Pose generation and Pose-to-Video generation." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.381, + 0.634, + 0.395 + ], + "angle": 0, + "content": "3.1. MotionVid" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.403, + 0.907, + 0.689 + ], + "angle": 0, + "content": "Existing datasets face limitations in training Text-to-Pose tasks. For instance, current 3D datasets [14, 35] primarily capture torso movements, lacking essential full-body keypoints, particularly for the face and hands, which makes them unsuitable for diverse pose generation. Although the Motion-X dataset [29] provides full-body 3D data, its scale is limited to only 81.1K samples, as the expensive annotation of 3D data. For the 2D full-body dataset, the Holistic-Motion2D dataset [55] includes approximately 1M samples, though only 16K of these are publicly accessible. Therefore, there is a pressing need for a comprehensive and publicly available human-motion pose dataset that facilitates the training of Text-to-Pose tasks. To address these limitations, we introduce the largest human-motion pose dataset MotionVid, which contains approximately 1.2M data pairs, with videos of varying resolutions and durations exceeding 64 frames. Besides, we provide a robust data collection pipeline to obtain high-quality textpose pairs from diverse and complex data sources." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Data Source. We construct a diverse set of text-pose pairs by curating data from various publicly available video datasets, including action recognition datasets Kinetics-400 [23], Kinetics-700 [4], ActivityNet-200 [11], and Something-Something V2 [12], as well as human-related datasets such as CAER [26], DFEW [22], UBody [28], Charades [45], Charades-Ego [46], HAA500 [10], and HMDB51 [25]. This selection totals \\(\\sim 6.6\\mathrm{M}\\) samples. To further enhance scene and subject diversity, we add \\(3.4\\mathrm{M}\\) videos sourced from the internet. Given the complexity of these videos, we perform extensive annotation and cleaning to ensure high-quality data. As shown in Fig. 2, we first use a shot transition detection model [47] to segment the video clips, followed by the video quality filter, data annotation," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12393" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.096, + 0.092, + 0.9, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.301, + 0.908, + 0.345 + ], + "angle": 0, + "content": "Figure 3. Training pipeline of the proposed Text-to-Pose generation. Pose data are encoded in latent space via the Pose VAE, which are then processed by the proposed MotionDiT, where local feature aggregation and global attention are utilized to capture information from the entire pose sequence. Finally, the LAMA loss is calculated via the proposed CLoP, which enhances the training of MotionDiT." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.371, + 0.395, + 0.385 + ], + "angle": 0, + "content": "human quality filter, and caption quality filter." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.386, + 0.483, + 0.612 + ], + "angle": 0, + "content": "Video Quality Filter. After acquiring a large number of videos, it is necessary to verify the video quality. Drawing inspiration from SVD's data cleaning strategy [3], we propose a Video Quality Filter to select high-quality video clips. Specifically, we use the GMFlow method [56] to estimate optical flow and filter out videos with insufficient movement intensity. Subsequently, we employ the method described in [1] to detect text regions and remove videos where text occupies a significant portion of the frame. We also utilize LAION-AI's aesthetic predictor [41] to compute aesthetic scores and eliminate videos with low aesthetic quality. Finally, the Laplacian operator [2] is applied to assess blur intensity, and videos deemed excessively blurry are discarded. As a result, approximately \\(50\\%\\) of the data is filtered out (see supplementary materials for more details)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.613, + 0.483, + 0.779 + ], + "angle": 0, + "content": "Data Annotation. We first prompt the Vision Language Model (VLM) [5] to generate action-oriented captions that describe human activities. Besides, for each frame, we extract 2D poses using the DWPose model [58]. This process yields a large number of text-pose pairs. However, the annotated data pairs are unsuitable for training Text-to-Pose tasks. This is due to several factors: the VLM's potential inaccuracy in describing multi-person actions, difficulty in recognizing poses occupying smaller areas in the frame, bias toward static poses from individuals with minimal motion, and limited detail in poses where faces are not visible." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Human Quality Filter. To further refine the dataset, we implement a Human Quality Filter, focusing on selecting data based on human-related characteristics. We calculate the difference between two consecutive frames' 2D poses to filter out sequences with insufficient motion magnitude. Additionally, we compute the ratio of the human detection bounding box to the entire frame to remove clips where the human coverage is too small. At this stage, we focus on" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.371, + 0.905, + 0.43 + ], + "angle": 0, + "content": "simpler scenarios by selecting scenes with a human count of one and ensuring face visibility for training. Based on the Human Quality Filter, approximately \\(75\\%\\) of the data is filtered out (see supplementary materials for more details)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.431, + 0.906, + 0.552 + ], + "angle": 0, + "content": "Caption Quality Filter. To enhance the quality of caption annotations, we apply a Caption Quality Filter to refine alignment accuracy. Specifically, using the proposed CLoP (see Sec. 3.2 for details), we compute the semantic similarity between text and poses, filtering data based on the similarity score. This ensures that each text annotation semantically aligns with the corresponding 2D pose data, thereby improving overall data quality." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.562, + 0.645, + 0.576 + ], + "angle": 0, + "content": "3.2. Text-to-Pose" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.906, + 0.734 + ], + "angle": 0, + "content": "Fig. 3 illustrates the training pipeline of the proposed Text-to-Pose generation. In this pipeline, pose data are first encoded into a latent space via the Pose VAE and subsequently processed by the proposed MotionDiT model. Within MotionDiT, local feature aggregation and global attention mechanisms are employed to capture comprehensive information across the entire pose sequence. Finally, LAMA loss is computed using the proposed CLoP, which further refines the training of MotionDiT. In the next, we introduce the Pose VAE, MotionDiT and CLoP." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Pose VAE. Variational Autoencoders (VAEs) [17, 24, 59, 67] have been widely applied in 2D image and 3D video processing, but their use for pose data remains underexplored, presenting a rich area for research. We represent a pose sequence as \\(\\mathbf{p} \\in \\mathbb{R}^{f \\times N \\times 3}\\), where \\(f\\) is the frame count, \\(N\\) the keypoints, and dimension 3 includes the \\(x\\), \\(y\\) coordinates and confidence scores. Given the sequential structure and frame-to-frame correlations, a 1D convolutional encoder with multi-layer downsampling effectively extracts temporal features, while symmetric decoder-encoder architecture with skip connections aids reconstruction. Confi" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12394" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "dence scores of each keypoint help mitigate occlusion issues, optimizing the model via KL divergence and reconstruction loss. Further details are available in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.152, + 0.483, + 0.304 + ], + "angle": 0, + "content": "MotionDiT. The proposed MotionDiT extends the Diffusion Transformer (DiT) [37] architecture specifically tailored to construct associations between human-motion poses and text control. In the proposed MotionDiT, we incorporate a global attention block to capture global spatial-temporal patterns inherent in human poses. Besides, the MotionDiT involves a local feature aggregation module to strengthen correlations between adjacent joints. Specifically, we first obtain the latent representation from the Pose VAE, which is then processed via a patch-embedding:" + }, + { + "type": "equation", + "bbox": [ + 0.222, + 0.314, + 0.483, + 0.332 + ], + "angle": 0, + "content": "\\[\nl _ {p} = \\mathcal {F} _ {\\text {P a t c h}} (\\mathcal {E} (\\mathbf {p})), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.341, + 0.484, + 0.417 + ], + "angle": 0, + "content": "where \\(\\mathcal{F}_{\\mathrm{Patch}}(\\cdot)\\) is the patch embedding operation [33], and \\(\\mathcal{E}(\\cdot)\\) is the Pose VAE encoder. Then we employ Diffusion blocks to process \\(l_{p}\\). As shown in Fig. 3, in each block, the local feature aggregation module is first utilized to strengthen local feature correlation:" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.428, + 0.483, + 0.448 + ], + "angle": 0, + "content": "\\[\nl _ {p} = \\mathcal {F} _ {\\mathrm {s a}} \\left(\\mathcal {F} _ {\\mathrm {r e s}} \\left(l _ {p}\\right) + l _ {p}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.455, + 0.483, + 0.771 + ], + "angle": 0, + "content": "where \\(\\mathcal{F}_{\\mathrm{res}}\\) is the 1D ResNet Block with kernel size 3, and \\(\\mathcal{F}_{\\mathrm{sa}}\\) is the spatial self attention. For text-based control, textual features obtained from CLIP [40] are then integrated into the network through cross-attention. Besides, following conventional video Diffusion architecture [3, 33], temporal attention is used to ensure continuity of pose features. Notably, it is essential for the model to capture global pose feature information. Although the model already employs spatiotemporal attention mechanisms, certain pose characteristics at one location and time may influence other locations at different times. Thus, global attention is required to extract these global features among internal pose keypoints. Instead of incorporating global attention in every attention module, we decide to apply it to the output of the central layer of the network to reduce computational complexity. Specifically, we reshape the latent output of the middle layer to \\(z \\in \\mathbb{R}^{(f \\times n) \\times c}\\), where \\(n\\) is the number of points \\(N\\) divided by the down-sampling factor \\(r\\) and \\(c\\) is the number of latent channels. We then perform self-attention across all frames \\(f\\) and all points \\(n\\). Finally, we reshape the output back to its original dimensions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.772, + 0.483, + 0.803 + ], + "angle": 0, + "content": "Following [33], we employ the noise prediction loss to optimize MotionDiT, thus the model output is:" + }, + { + "type": "equation", + "bbox": [ + 0.222, + 0.815, + 0.483, + 0.831 + ], + "angle": 0, + "content": "\\[\n\\epsilon_ {\\text {p r e d}} = g _ {\\theta} \\left(z _ {t}, t, s\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(g_{\\theta}\\) represents the MotionDiT parameterized by \\(\\theta\\), \\(z_{t}\\) is the noisy latent variable at time step \\(t\\), and \\(s\\) denotes the conditional input (e.g., textual features), we compute the noise prediction at time step \\(t\\). The model learns to predict" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "the noise \\(\\epsilon_{\\mathrm{pred}}\\) conditioned on the input. We calculate the mean squared error (MSE) loss between the predicted noise \\(\\epsilon_{\\mathrm{pred}}\\) and the true noise \\(\\epsilon\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.615, + 0.143, + 0.905, + 0.163 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {d} = \\mathbb {E} _ {t, z _ {0}, \\epsilon} \\left[ \\| \\epsilon - \\epsilon_ {\\text {p r e d}} \\| ^ {2} \\right], \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.168, + 0.905, + 0.198 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_d\\) represents the loss over all time steps \\(t\\), true noise \\(\\epsilon \\sim \\mathcal{N}(0, I)\\), and ground truth latent variable \\(z_0\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.199, + 0.907, + 0.62 + ], + "angle": 0, + "content": "CLoP. Additionally, prior works [27, 60] emphasize the significance of intermediate feature representations, which inspire us to propose a LAtent seMantic Alignment (LAMA) loss aimed at enhancing text-based motion control. We expect this term to provide benefits in two key ways: (1) By aligning the latent space of MotionDiT with a largescale pre-trained motion encoder, we can improve both fidelity and diversity of the generated 2D pose; (2) Aligning with the pre-trained pose encoder that is already aligned with text inherently strengthens the influence of text control on motion generation. Unlike CLIP [40], there are currently no readily available large-scale pre-trained models for aligning motion and text. Leveraging the MotionVid dataset, we introduce Contrastive Language-Motion Pretraining (CLoP), which aligns text with 2D pose. This improves the evaluation of 2D pose and latent semantic alignment. Specifically, inspired by CLIP, we employ a similar mechanism to extract textual features. The extracted text features and 2D pose data are then processed through dedicated text encoder \\(\\mathcal{F}_e(\\cdot)\\) and pose encoder \\(\\mathcal{F}_p(\\cdot)\\), respectively. The alignment of the two modalities is facilitated by optimizing a contrastive loss function [40]. Our architecture represents a 2D extension of the approach in [38]. Unlike the original model, we utilize CLIP for poses representation as \\(\\mathbf{p} \\in \\mathbb{R}^{f \\times N \\times 3}\\) with its corresponding text e, capturing frame-wise keypoint coordinates and confidence scores instead of 3D features, the CLoP training objective as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.626, + 0.905, + 0.646 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c} = \\ell_ {c e} \\left(\\ell_ {2} \\left(\\mathbf {h} _ {e} \\mathbf {h} _ {p} ^ {T}\\right); y\\right) + \\ell_ {c e} \\left(\\ell_ {2} \\left(\\mathbf {h} _ {p} \\mathbf {h} _ {e} ^ {T}\\right); y\\right) / 2, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.651, + 0.907, + 0.727 + ], + "angle": 0, + "content": "where \\(\\ell_2(\\cdot)\\) denote the \\(\\ell_2\\) normalization and \\(\\ell_{ce}(\\cdot ;y)\\) represent the cross-entropy loss with ground truth label \\(y\\). \\(\\mathbf{h}_e = \\mathcal{F}_e(\\mathbf{e})\\mathbf{W}_e\\) is the projected text embedding, calculated by projection matrix \\(\\mathbf{W}_e\\). \\(\\mathbf{h}_p = \\mathcal{F}_p(\\mathbf{p})\\mathbf{W}_p\\) is the projected pose embedding." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.727, + 0.907, + 0.879 + ], + "angle": 0, + "content": "Based on the CLoP, we evaluate the dissimilarity within the feature space (e.g., MSE or cosine similarity metrics), achieving the benefits of the previously mentioned LAMA loss. Rather than using prior methods to extract structural features, we focus on the semantic features of motion, as motion's structural information is simpler while textual information is richer. We also employ a feature projection using a two-layer MLP \\( g_{\\omega}(\\cdot) \\) parameterized by \\( \\omega \\) to project the latent feature space learned by MotionDiT. As a result, the LAMA loss \\( \\mathcal{L}_f \\) can be formulated:" + }, + { + "type": "equation", + "bbox": [ + 0.629, + 0.884, + 0.905, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {f} = d \\left(g _ {\\omega} \\left(\\mathbf {h} _ {d} ^ {l}\\right), \\mathbf {h} _ {p}\\right), \\tag {6}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12395" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 1. Comparison to Other State-of-the-Art Methods on the MotionVid Subset. The metrics demonstrate that our method outperforms others in terms of pose-text alignment and diversity. Bold indicates the best result." + }, + { + "type": "table", + "bbox": [ + 0.147, + 0.129, + 0.855, + 0.223 + ], + "angle": 0, + "content": "
MethodFID ↓Rp-top1 ↑Rp-top2 ↑Rp-top3 ↑Diversity ↑MM Dist ↓MM ↑
T2M-GPT [63]496.9070.1760.2990.39564.00042.95014.480
PriorMDM [43]538.9490.1560.2700.36662.33852.55044.135
MLD [6]396.9490.3180.5050.62864.44240.19657.862
Ours149.0070.4510.6380.74368.22032.76160.460
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.246, + 0.483, + 0.293 + ], + "angle": 0, + "content": "where \\(\\mathbf{h}_d^l\\) is the \\(l\\)-th layer latent representation of MotionDiT, \\(d(\\cdot ,\\cdot)\\) is the CLoP function. Consequently, the overall objective \\(\\mathcal{L}\\) of MotionDiT can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.227, + 0.302, + 0.483, + 0.319 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {d} + \\lambda_ {f} \\mathcal {L} _ {f}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.328, + 0.484, + 0.359 + ], + "angle": 0, + "content": "where \\(\\lambda_{f}\\) is a hyperparameter to trade off the alignment and denoising process." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.367, + 0.235, + 0.381 + ], + "angle": 0, + "content": "3.3. Pose-to-Video" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.389, + 0.484, + 0.584 + ], + "angle": 0, + "content": "We propose a model for Pose-to-Video generation, capable of producing human-motion videos based on an initial frame image and a pose sequence. The model leverages [59] as its backbone, maintaining consistent parameters with the original model, incorporates conditional control inspired by [64], and employs data augmentation techniques from [49] to enhance performance. Notably, all training was conducted on the MotionVid dataset, demonstrating its versatility for both Text-to-Pose and Pose-to-Video tasks. It is worth noting that other Pose-to-Video models, such as [20, 49, 66], are also capable of handling this task. For further details on the model architecture, please refer to the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.598, + 0.224, + 0.615 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.623, + 0.279, + 0.64 + ], + "angle": 0, + "content": "4.1. Experimental setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.483, + 0.765 + ], + "angle": 0, + "content": "Dataset. For computational efficiency, we select a representative subset of 50K samples from MotionVid, covering diverse actions like dancing, squatting, and lifting to ensure diversity and expedite experimentation. This subset maintains result representativeness while enabling efficient testing. We also conducted scaling experiments to assess model performance across various dataset sizes. Qualitative visualization experiments utilize a dataset of 1.2M samples." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Implementation Details. For Text-to-Pose, our motion data consists of 128 points per frame, covering the body, face, and hands. The input is structured as \\( f \\times 128 \\times 3 \\), representing \\( f \\) frames, with 128 points in each frame and 3 channels that encode image coordinates (x, y) along with confidence scores. To maintain uniformity across samples, all sequences are cropped to 64 frames. For Pose VAE, we employ a downsampling factor \\( r = 8 \\), yielding a latent representation \\( z \\in \\mathbb{R}^{f \\times 16 \\times 3} \\). In MotionDiT, we apply a two-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.246, + 0.907, + 0.352 + ], + "angle": 0, + "content": "fold downsampling in the patch embedding. The network consists of 13 layers, with text features extracted from the SD2.1 CLIP model [42] as our text encoder, resulting in a conditional embedding \\( t_{emb} \\in \\mathbb{R}^{1 \\times 1024} \\). The AdamW optimizer is used with a learning rate of \\( 1 \\times 10^{-5} \\). The training is conducted on 8 H20 GPUs, with inference performed on a single H20." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.356, + 0.909, + 0.568 + ], + "angle": 0, + "content": "Evaluation Metrics. To evaluate generated motion data quality, several metrics are employed, consistent with previous works [6, 15, 43, 52, 55, 63]. Frechet Inception Distance (FID) [18] assesses distributional similarity between ground truth and generated motions, using features from a custom-trained CLOP model. MultiModality evaluates diversity for a single text input, while Diversity [15] calculates variance across all motions to quantify output diversity. R-precision measures top-1, top-2, and top-3 recall rates in motion retrieval, evaluating alignment accuracy. MultiModality Distance gauges the distance between textual inputs and generated motions, indicating adherence to prompts. Further details are available in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.586, + 0.73, + 0.601 + ], + "angle": 0, + "content": "4.2. Results on Text-to-Pose" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.611, + 0.907, + 0.793 + ], + "angle": 0, + "content": "Currently, Tender [55] is the only method for generating 2D poses based on text but its code and dataset are unavailable. Thus, we select three SOTA 3D motion generation methods for comparison: T2M-GPT [62], PriorMDM [43], and MLD [6], converting their inputs to 2D poses. All methods are trained on the MotionVid subset with proposed settings to standardize evaluation metrics. Comparison results (Tab. 1) show our model outperforms others, achieving \\(62.4\\%\\) improvement in FID, \\(41.8\\%\\) in Rp-top1, \\(26.3\\%\\) in Rp-top2, and \\(18.3\\%\\) in Rp-top3. These results indicate our model generates semantically aligned outputs with high diversity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "The visual comparison with other methods is presented in Fig. 4, and the results demonstrate a significant improvement over existing models. Our method produces poses that better align with the textual constraints, with keypoints remaining intact and exhibiting minimal motion jitter. A more detailed visual comparison can be found in the supplementary materials." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12396" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.133 + ], + "angle": 0, + "content": "Table 2. The ablation study presents four configurations, progressively adding components of Local, Global, and LAMA loss to the original model. As we move from the initial configuration to the fully enhanced model, performance metrics consistently improve, highlighting the positive impact of each additional component." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.143, + 0.905, + 0.239 + ], + "angle": 0, + "content": "
MethodFID ↓Rp-top1 ↑Rp-top2 ↑Rp-top3 ↑Diversity ↑MM Dist ↓MM ↑
Vanilla DiT283.0910.3670.5610.68266.58737.04621.461
DiT+Local183.2130.4150.6060.72167.59134.29141.890
DiT+Local+Global162.0220.4330.6200.73067.35733.26957.848
DiT+Local+Global+LAMA149.0070.4510.6380.74368.22032.76160.460
" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.254, + 0.416, + 0.275 + ], + "angle": 0, + "content": "A man is holding his head in his hands and continues to do so while looking down at them." + }, + { + "type": "text", + "bbox": [ + 0.451, + 0.254, + 0.722, + 0.266 + ], + "angle": 0, + "content": "A woman is dancing, moving her arms and legs around." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.277, + 0.441, + 0.501 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.452, + 0.276, + 0.892, + 0.501 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.521, + 0.908, + 0.566 + ], + "angle": 0, + "content": "Figure 4. Visualization results compared to SOTA Text-to-Pose methods. The results demonstrate that our model significantly outperforms other models. Our method generates poses that are more consistent with the text constraints, with keypoints maintaining their integrity and minimal motion jitter. For a better visual comparison, please refer to the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.589, + 0.255, + 0.604 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.612, + 0.483, + 0.671 + ], + "angle": 0, + "content": "To validate the efficacy of the three modules proposed in the MotionDiT (i.e., the local feature aggregation, the global attention, and the LAMA loss), we conduct a series of ablation studies. The outcomes are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.673, + 0.483, + 0.763 + ], + "angle": 0, + "content": "Local Feature Aggregation. Removing the local attention block and integrating a DiT structure similar to Latte [33] in each attention mechanism, we showcase our method's efficacy in capturing local details. The performance improvements, shown in Tab. 2, confirm the robustness of our local feature extraction approach." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.764, + 0.483, + 0.87 + ], + "angle": 0, + "content": "Global Attention Block. To assess the impact of our global attention mechanism, we perform an ablation test where this component is removed from the model. The findings, detailed in Tab. 2, indicate a notable decline in performance, which underscores the importance of the global attention mechanism in enhancing the model's ability to capture long-range dependencies and global context." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "LAMA Loss. We investigate the role of the semantic alignment loss by training MotionDiT without it. The analysis in" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.589, + 0.907, + 0.68 + ], + "angle": 0, + "content": "Tab. 2 shows a significant performance drop, validating the LAMA loss's critical role in guiding optimization towards accurate and stable solutions. These ablation studies provide strong evidence for the proposed algorithms' effectiveness and necessity in enhancing model performance across various aspects." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.692, + 0.707, + 0.708 + ], + "angle": 0, + "content": "4.4. Scaling Law Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.715, + 0.906, + 0.806 + ], + "angle": 0, + "content": "We expand the training set size, and as shown in Tab. 3, our metrics improved with the increase in data volume. This demonstrates the validity of using large-scale 2D pose data and suggests that model performance can be further enhanced through an efficient and cost-effective data collection pipeline." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.818, + 0.83, + 0.833 + ], + "angle": 0, + "content": "4.5. Qualitative Results on Pose-to-Video" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.907, + 0.902 + ], + "angle": 0, + "content": "In our qualitative experiments, we use the generated motion sequences as templates and provide them to the Pose-to-Video model. The specific video results are shown in Fig. 5. Compared to state-of-the-art text-to-video models," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12397" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.45, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.456, + 0.09, + 0.895, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.256, + 0.908, + 0.314 + ], + "angle": 0, + "content": "Figure 5. Visualization results compared to SOTA Text-to-Video methods. Mochi1 [50] and CogVideoX [59] exhibit issues such as body distortion, weak motion continuity, and neglecting facial generation. In contrast, HumanDreamer is able to generate more coherent and consistent videos with smoother transitions and better attention to details such as facial expressions. For a better visual comparison, please refer to the supplementary materials." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.336, + 0.483, + 0.38 + ], + "angle": 0, + "content": "Table 3. By increasing the amount of training data, we observe an improvement in model performance, which validates the potential for rapid scalability using large-scale 2D motion data." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.389, + 0.481, + 0.48 + ], + "angle": 0, + "content": "
DatasizeRp-top1 ↑Rp-top2 ↑Rp-top3 ↑MM Dist ↓
50K(4%)0.4510.6380.74332.761
250K(20%)0.4520.6410.75132.666
500K(40%)0.4640.6570.76632.399
1M(80%)0.4920.6710.76931.116
1.25M(100%)0.5130.6940.79130.139
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.503, + 0.484, + 0.548 + ], + "angle": 0, + "content": "our model exhibits larger and more dynamic movements, with the characters' actions more accurately reflecting the textual descriptions." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.56, + 0.48, + 0.627 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.628, + 0.478, + 0.694 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.707, + 0.484, + 0.749 + ], + "angle": 0, + "content": "Figure 6. The visualization results demonstrate that using the initial frame pose, different prompt texts can guide the generation of distinct pose sequences." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.773, + 0.325, + 0.789 + ], + "angle": 0, + "content": "4.6. Other Downstream Tasks" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.841 + ], + "angle": 0, + "content": "In addition to the Pose-to-Video task, the proposed Text-to-Pose component can also be utilized for pose sequence prediction, 2D-3D motion lifting tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Pose Sequence Prediction. In scenarios with partial pose sequences, the Text-to-Pose model can infer and generate missing parts by conditioning on existing poses and textual movement descriptions. This enables the synthesis of co-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.339, + 0.907, + 0.399 + ], + "angle": 0, + "content": "herent pose continuations. Given initial and final frames, the model predicts intermediate pose states, completing the sequence (Fig. 6). This is especially useful in animation for supplementing incomplete pose data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.401, + 0.909, + 0.522 + ], + "angle": 0, + "content": "2D-3D Motion Lifting. Models like [70] can enhance motion data dimensionality, converting 2D motions into realistic 3D motions (Fig. 7). With textual descriptions specifying depth or spatial aspects, the Text-to-Pose system generates richer 3D motion representations. This is crucial for virtual reality (VR) and augmented reality (AR) environments, where realistic 3D motion significantly improves user experience." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.538, + 0.904, + 0.626 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.539, + 0.636, + 0.88, + 0.651 + ], + "angle": 0, + "content": "Figure 7. Visualization results on 2D-3D Motion Lifting." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.69, + 0.764, + 0.706 + ], + "angle": 0, + "content": "5. Discussion and Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.717, + 0.907, + 0.824 + ], + "angle": 0, + "content": "In this study, we present HumanDreamer, a pioneering decoupled framework for generating human-motion videos that merges text control flexibility with pose guidance controllability. Utilizing MotionVid, the largest dataset for human-motion pose generation, we train MotionDiT for producing structured poses. We introduce LAMA loss to improve semantic alignment, ensuring coherent outputs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Experimental results indicate that using generated poses in Pose-to-Video yields high-quality, diverse human-motion videos, surpassing current benchmarks. These findings confirm the effectiveness and adaptability of our decoupled framework, facilitating versatile video generation." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12398" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.482, + 0.156 + ], + "angle": 0, + "content": "[1] Youngmin Baek, Bado Lee, Dongyoon Han, Sangdoo Yun, and Hwalsuk Lee. Character region awareness for text detection. In CVPR, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.157, + 0.483, + 0.198 + ], + "angle": 0, + "content": "[2] Raghav Bansal, Gaurav Raj, and Tanupriya Choudhury. Blur image detection using laplacian operator and open-cv. In SMART, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.2, + 0.483, + 0.269 + ], + "angle": 0, + "content": "[3] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.27, + 0.482, + 0.311 + ], + "angle": 0, + "content": "[4] Joao Carreira, Eric Noland, Chloe Hillier, and Andrew Zisserman. A short note on the kinetics-700 human action dataset, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.312, + 0.482, + 0.38 + ], + "angle": 0, + "content": "[5] Lin Chen, Xilin Wei, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Bin Lin, Zhenyu Tang, et al. Sharegpt4video: Improving video understanding and generation with better captions. arXiv preprint arXiv:2406.04325, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.382, + 0.482, + 0.422 + ], + "angle": 0, + "content": "[6] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, and Gang Yu. Executing your commands via motion diffusion in latent space. In CVPR, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.424, + 0.482, + 0.478 + ], + "angle": 0, + "content": "[7] Yuxin Chen, Ziqi Zhang, Chunfeng Yuan, Bing Li, Ying Deng, and Weiming Hu. Channel-wise topology refinement graph convolution for skeleton-based action recognition. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.481, + 0.482, + 0.521 + ], + "angle": 0, + "content": "[8] Ziwei Chen, Qiang Li, Xiaofeng Wang, and Wankou Yang. Liftedcl: Lifting contrastive learning for human-centric perception. In ICLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.523, + 0.482, + 0.59 + ], + "angle": 0, + "content": "[9] Joseph Cho, Fachrina Dewi Puspitasari, Sheng Zheng, Jingyao Zheng, Lik-Hang Lee, Tae-Ho Kim, Choong Seon Hong, and Chaoning Zhang. Sora as an agi world model? a complete survey on text-to-video generation. arXiv preprint arXiv:2403.05131, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.593, + 0.482, + 0.634 + ], + "angle": 0, + "content": "[10] Jihoon Chung, Cheng hsin Wu, Hsuan ru Yang, Yu-Wing Tai, and Chi-Keung Tang. Haa500: Human-centric atomic action dataset with curated videos. In ICCV, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.635, + 0.482, + 0.676 + ], + "angle": 0, + "content": "[11] Bernard Ghanem Fabian Caba Heilbron, Victor Escorcia and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In CVPR, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.677, + 0.482, + 0.759 + ], + "angle": 0, + "content": "[12] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The\" something something\" video database for learning and evaluating visual common sense. In ICCV, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.761, + 0.482, + 0.802 + ], + "angle": 0, + "content": "[13] Yanchen Guan, Haicheng Liao, Zhenning Li, Guohui Zhang, and Chengzhong Xu. World models for autonomous driving: An initial survey. arXiv preprint arXiv:2403.02622, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[14] Chuan Guo, Xinxin Zuo, Sen Wang, Shihao Zou, Qingyao Sun, Annan Deng, Minglun Gong, and Li Cheng. Action2motion: Conditioned generation of 3d human motions. In ACM MM, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[15] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In CVPR, 2022. 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[16] Agrim Gupta, Lijun Yu, Kihyuk Sohn, Xiuye Gu, Meera Hahn, Li Fei-Fei, Irfan Essa, Lu Jiang, and José Lezama. Photorealistic video generation with diffusion models. arXiv preprint arXiv:2312.06662, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.905, + 0.188 + ], + "angle": 0, + "content": "[17] Louay Hazami, Rayhane Mama, and Ragavan Thurairatnam. Efficient-vdvae: Less is more. arXiv preprint arXiv:2203.13751, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.19, + 0.905, + 0.244 + ], + "angle": 0, + "content": "[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.246, + 0.905, + 0.3 + ], + "angle": 0, + "content": "[19] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.904, + 0.342 + ], + "angle": 0, + "content": "[20] Li Hu. Animate anyone: Consistent and controllable image-to-video synthesis for character animation. In CVPR, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.904, + 0.397 + ], + "angle": 0, + "content": "[21] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. TPAMI, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.904, + 0.453 + ], + "angle": 0, + "content": "[22] Xingxun Jiang, Yuan Zong, Wenming Zheng, Chuangao Tang, Wanchuang Xia, Cheng Lu, and Jiateng Liu. Dfew: A large-scale database for recognizing dynamic facial expressions in the wild. In ACM MM, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.456, + 0.904, + 0.523 + ], + "angle": 0, + "content": "[23] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, Mustafa Suleyman, and Andrew Zisserman. The kinetics human action video dataset, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.904, + 0.551 + ], + "angle": 0, + "content": "[24] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.553, + 0.904, + 0.592 + ], + "angle": 0, + "content": "[25] H. Kuehne, H. Jhuang, E. Garrote, T. Poggio, and T. Serre. HMDB: a large video database for human motion recognition. In ICCV, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.595, + 0.904, + 0.635 + ], + "angle": 0, + "content": "[26] Jiyoung Lee, Seungryong Kim, Sunok Kim, Jungin Park, and Kwanghoonn Sohn. Context-aware emotion recognition networks. In ICCV, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.637, + 0.904, + 0.677 + ], + "angle": 0, + "content": "[27] Alexander C Li, Mihir Prabhudesai, Shivam Duggal, Ellis Brown, and Deepak Pathak. Your diffusion model is secretly a zero-shot classifier. In ICCV, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.904, + 0.718 + ], + "angle": 0, + "content": "[28] Jing Lin, Ailing Zeng, Haoqian Wang, Lei Zhang, and Yu Li. One-stage 3d whole-body mesh recovery with component aware transformer. CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.72, + 0.905, + 0.774 + ], + "angle": 0, + "content": "[29] Jing Lin, Ailing Zeng, Shunlin Lu, Yuanhao Cai, Ruimao Zhang, Haoqian Wang, and Lei Zhang. Motion-x: A large-scale 3d expressive whole-body human motion dataset. NeurIPS, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.776, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[30] Jun Liu, Amir Shahroudy, Mauricio Perez, Gang Wang, Ling-Yu Duan, and Alex C Kot. Ntu rgb+ d 120: A large-scale benchmark for 3d human activity understanding. TPAMI, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[31] Yixin Liu, Kai Zhang, Yuan Li, Zhiling Yan, Chujie Gao, Ruoxi Chen, Zhengqing Yuan, Yue Huang, Hanchi Sun, Jianfeng Gao, et al. Sora: A review on background, technology, limitations, and opportunities of large vision models. arXiv preprint arXiv:2402.17177, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12399" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.147 + ], + "angle": 0, + "content": "[32] Shunlin Lu, Ling-Hao Chen, Ailing Zeng, Jing Lin, Ruimao Zhang, Lei Zhang, and Heung-Yeung Shum. Humantomato: Text-aligned whole-body motion generation. arXiv preprint arXiv:2310.12978, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.15, + 0.483, + 0.207 + ], + "angle": 0, + "content": "[33] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.208, + 0.483, + 0.263 + ], + "angle": 0, + "content": "[34] Yue Ma, Yingqing He, Xiaodong Cun, Xintao Wang, Siran Chen, Xiu Li, and Qifeng Chen. Follow your pose: Pose-guided text-to-video generation using pose-free videos. In AAAI, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.265, + 0.483, + 0.307 + ], + "angle": 0, + "content": "[35] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In ICCV, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.309, + 0.483, + 0.35 + ], + "angle": 0, + "content": "[36] Haomiao Ni, Changhao Shi, Kai Li, Sharon X Huang, and Martin Renqiang Min. Conditional image-to-video generation with latent flow diffusion models. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.352, + 0.483, + 0.38 + ], + "angle": 0, + "content": "[37] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.382, + 0.483, + 0.422 + ], + "angle": 0, + "content": "[38] Mathis Petrovich, Michael J. Black, and Gül Varol. TMR: Text-to-motion retrieval using contrastive 3D human motion synthesis. In ICCV, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.425, + 0.483, + 0.453 + ], + "angle": 0, + "content": "[39] Matthias Plappert, Christian Mandery, and Tamim Asfour. The kit motion-language dataset. Big data, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.455, + 0.483, + 0.525 + ], + "angle": 0, + "content": "[40] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.527, + 0.483, + 0.554 + ], + "angle": 0, + "content": "[41] Beaumont Romain and Schuhmann Christoph. Laion aesthetics predictor v1. 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.556, + 0.483, + 0.597 + ], + "angle": 0, + "content": "[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.6, + 0.483, + 0.64 + ], + "angle": 0, + "content": "[43] Yoni Shafir, Guy Tevet, Roy Kapon, and Amit Haim Bermano. Human motion diffusion as a generative prior. In ICLR, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.643, + 0.483, + 0.685 + ], + "angle": 0, + "content": "[44] Amir Shahroudy, Jun Liu, Tian-Tsong Ng, and Gang Wang. Ntu \\(\\mathrm{rgb + }\\) d: A large scale dataset for 3d human activity analysis. In CVPR, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.687, + 0.483, + 0.743 + ], + "angle": 0, + "content": "[45] Gunnar A. Sigurdsson, Gül Varol, Xiaolong Wang, Ali Farhadi, Ivan Laptev, and Abhinav Gupta. Hollywood in homes: Crowdsourcing data collection for activity understanding, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.745, + 0.483, + 0.799 + ], + "angle": 0, + "content": "[46] Gunnar A. Sigurdsson, Abhinav Gupta, Cordelia Schmid, Ali Farhadi, and Karteek Alahari. Charades-ego: A large-scale dataset of paired third and first person videos. In ArXiv, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.802, + 0.483, + 0.843 + ], + "angle": 0, + "content": "[47] Tomáš Souček and Jakub Lokoc. Transnet v2: An effective deep network architecture for fast shot transition detection. arXiv preprint arXiv:2008.04838, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[48] Rui Sun, Yumin Zhang, Tejal Shah, Jiahao Sun, Shuoying Zhang, Wenqi Li, Haoran Duan, Bo Wei, and Rajiv Ranjan. From sora what we can see: A survey of text-to-video generation. arXiv preprint arXiv:2405.10674, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.906, + 0.161 + ], + "angle": 0, + "content": "[49] Shuai Tan, Biao Gong, Xiang Wang, Shiwei Zhang, Dandan Zheng, Ruobing Zheng, Kecheng Zheng, Jingdong Chen, and Ming Yang. Animate-x: Universal character image animation with enhanced motion representation. arXiv preprint arXiv:2410.10306, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.163, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[50] Genmo Team. Mochi 1. https://github.com/genmoai/models, 2024.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.191, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[51] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. In ECCV, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.234, + 0.905, + 0.273 + ], + "angle": 0, + "content": "[52] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In ICLR, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.275, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[53] Xiang Wang, Shiwei Zhang, Changxin Gao, Jiayu Wang, Xiaoqiang Zhou, Yingya Zhang, Luxin Yan, and Nong Sang. Unianimate: Taming unified video diffusion models for consistent human image animation. arXiv preprint arXiv:2406.01188, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.345, + 0.905, + 0.412 + ], + "angle": 0, + "content": "[54] Xiaofeng Wang, Kang Zhao, Feng Liu, Jiayu Wang, Guosheng Zhao, Xiaoyi Bao, Zheng Zhu, Yingya Zhang, and Xingang Wang. Egovid-5m: A large-scale video-action dataset for egocentric video generation. arXiv preprint arXiv:2411.08380, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.414, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[55] Yuan Wang, Zhao Wang, Junhao Gong, Di Huang, Tong He, Wanli Ouyang, Jile Jiao, Xuetao Feng, Qi Dou, Shixiang Tang, et al. Holistic-motion2d: Scalable whole-body human motion generation in 2d space. arXiv preprint arXiv:2406.11253, 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.484, + 0.905, + 0.551 + ], + "angle": 0, + "content": "[56] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, and Dacheng Tao. Gmflow: Learning optical flow via global matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8121-8130, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.553, + 0.905, + 0.593 + ], + "angle": 0, + "content": "[57] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In AAAI, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.595, + 0.905, + 0.635 + ], + "angle": 0, + "content": "[58] Zhendong Yang, Ailing Zeng, Chun Yuan, and Yu Li. Effective whole-body pose estimation with two-stages distillation. In ICCV, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.637, + 0.905, + 0.704 + ], + "angle": 0, + "content": "[59] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 2, 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.707, + 0.905, + 0.761 + ], + "angle": 0, + "content": "[60] Sihyun Yu, Sangkyung Kwak, Huiwon Jang, Jongheon Jeong, Jonathan Huang, Jinwoo Shin, and Saining Xie. Representation alignment for generation: Training diffusion transformers is easier than you think, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.763, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[61] Yan Zeng, Guoqiang Wei, Jiani Zheng, Jiaxin Zou, Yang Wei, Yuchen Zhang, and Hang Li. Make pixels dance: High-dynamic video generation. arXiv preprint arXiv:2311.10982, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.818, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[62] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Shaoli Huang, Yong Zhang, Hongwei Zhao, Hongtao Lu, and Xi Shen. T2m-gpt: Generating human motion from textual descriptions with discrete representations. In CVPR, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[63] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Yong Zhang, Hongwei Zhao, Hongtao Lu, Xi Shen, and Ying" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12400" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "Shan. Generating human motion from textual descriptions with discrete representations. In CVPR, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.162 + ], + "angle": 0, + "content": "[64] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.22 + ], + "angle": 0, + "content": "[65] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.222, + 0.482, + 0.29 + ], + "angle": 0, + "content": "[66] Yang Zhang, Jiaxi Gu, Li-Wen Wang, Han Wang, Junqi Cheng, Yuefeng Zhu, and Fangyuan Zou. Mimic motion: High-quality human motion video generation with confidence-aware pose guidance. arXiv preprint arXiv:2406.19680, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.292, + 0.482, + 0.347 + ], + "angle": 0, + "content": "[67] Sijie Zhao, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Muyao Niu, Xiaoyu Li, Wenbo Hu, and Ying Shan. Cv-vae: A compatible video vae for latent generative video models. https://arxiv.org/abs/2405.20279, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.482, + 0.389 + ], + "angle": 0, + "content": "[68] Weiyu Zhao, Liangxiao Hu, and Shengping Zhang. Diffugesture: Generating human gesture from two-person dialogue with diffusion models. In ICMI, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.482, + 0.459 + ], + "angle": 0, + "content": "[69] Shenhao Zhu, Junming Leo Chen, Zuozhuo Dai, Qingkun Su, Yinghui Xu, Xun Cao, Yao Yao, Hao Zhu, and Siyu Zhu. Champ: Controllable and consistent human image animation with 3d parametric guidance. arXiv preprint arXiv:2403.14781, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.461, + 0.482, + 0.53 + ], + "angle": 0, + "content": "[70] Wentao Zhu, Xiaoxuan Ma, Zhaoyang Liu, Libin Liu, Wayne Wu, and Yizhou Wang. Motionbert: A unified perspective on learning human motion representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.532, + 0.482, + 0.6 + ], + "angle": 0, + "content": "[71] Zheng Zhu, Xiaofeng Wang, Wangbo Zhao, Chen Min, Ni-anchen Deng, Min Dou, Yuqi Wang, Botian Shi, Kai Wang, Chi Zhang, et al. Is sora a world simulator? a comprehensive survey on general world models and beyond. arXiv preprint arXiv:2405.03520, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.6 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "12401" + } + ] +] \ No newline at end of file diff --git a/2025/HumanDreamer_ Generating Controllable Human-Motion Videos via Decoupled Generation/8f919b44-e98e-44f2-a774-268fc39345bc_origin.pdf b/2025/HumanDreamer_ Generating Controllable Human-Motion Videos via Decoupled Generation/8f919b44-e98e-44f2-a774-268fc39345bc_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..752ae82da5ea0f0ef1d0aba030a02e2397257f27 --- /dev/null +++ b/2025/HumanDreamer_ Generating Controllable Human-Motion Videos via Decoupled Generation/8f919b44-e98e-44f2-a774-268fc39345bc_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b906394dfa7d6629ff87c6936f9ac45e26a92cbea81cd5d2ea77ef6a3b60bba2 +size 9201172 diff --git a/2025/HumanDreamer_ Generating Controllable Human-Motion Videos via Decoupled Generation/full.md b/2025/HumanDreamer_ Generating Controllable Human-Motion Videos via Decoupled Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..31a97b519061997938cd18458a7a106faf9d08bd --- /dev/null +++ b/2025/HumanDreamer_ Generating Controllable Human-Motion Videos via Decoupled Generation/full.md @@ -0,0 +1,367 @@ +# HumanDreamer: Generating Controllable Human-Motion Videos via Decoupled Generation + +Boyuan Wang $^{1,2,3*}$ , Xiaofeng Wang $^{1,2,3*}$ , Chaojun Ni $^{4,5}$ , Guosheng Zhao $^{1,2,3}$ , Zhiqin Yang $^{6}$ , Zheng Zhu $^{4\dagger}$ , Muyang Zhang $^{1,2}$ , Yukun Zhou $^{4}$ , Xinze Chen $^{4}$ , Guan Huang $^{4}$ , Lihong Liu $^{1}$ , Xingang Wang $^{1,3\dagger}$ + +$^{1}$ Institute of Automation, Chinese Academy of Sciences, China + +$^{2}$ School of Artificial Intelligence, University of Chinese Academy of Sciences, China + +$^{3}$ Luoyang Institute for Robot and Intelligent Equipment, China + +$^{4}$ GigaAI, China $^{5}$ Peking University, China $^{6}$ The Chinese University of Hong Kong + +![](images/4d9fb75c2e6b73f5f1aeca537ac865c23d69eefa743b4ccb77d156339745e9de.jpg) + +Text + +A man is performing a yoga pose on a mat, and he is seen moving his legs and arms in different positions. + +A man is speaking to the camera and then looking off into the distance. Later, he is speaking to the camera again. + +![](images/1854e0342367af45949e612f9934120c5176b22a34f9ef0c1da950452cd83400.jpg) + +HumanDreamer + +![](images/4dbab8c84cbcdf186b3e3fdcf022ab2cfdbab02333e726ef38f2707e48ad6a4d.jpg) + +![](images/e8f827329573956ee15bab64424ad7bc7c6f34f1f11dfe4b663ce163e37243e9.jpg) + +![](images/3f8ca9494a8d2e8e435af44a0f2df93b6b51b43364b09e682e3072412ab75e3a.jpg) +Text-to-Pose Generation + +![](images/1e94a5cf8ef8bd75b2ad9c2661ba34fdc8921647fa25899131cf3cf80e56de86.jpg) + +![](images/77a51d7e752bd3ae932752d0e934f9c1e271d1eeecffea74a16638c168a7cdb1.jpg) + +![](images/9c7b5e002ac917dc1cfc9102ea658206347663dee52a7e2f6edc83ee7451d53b.jpg) + +![](images/bbfc19f914967ff7e30916898b4b3ee811b9cfd6da0732ca59febdb7bea83fd7.jpg) + +![](images/b81e79a4b9f71ee5e97d63f32f3c11fc55203e3111f583c3e9351eb2c185e3e4.jpg) + +![](images/7dd132b0bb811a968d662444727540635d190c1d7e21b54d0d37c1f7a80c3ffd.jpg) + +![](images/1e973707081fd3653105c4b1e44468297d29da928eaac73f3a74a2dcd102bdeb.jpg) + +![](images/a34c8b76e8c9131652bcc889daf840037bc469e98f83e00b679696eb609ab49e.jpg) + +![](images/0533623de33a056f233d3a583df12fc7c5c10ef586237acddae55e348e69b62d.jpg) + +![](images/daa34c4e486e3ea4a975e5efae0a064395f7aed52bd0cf482497f8933fb9ae6d.jpg) + +![](images/e2f2f9cd04c43aae3b98fd31ee51a77c167696d04fa73e9db9102a72558d0f33.jpg) + +![](images/e27aa5ad574bd389e80ad374abcec69fb9322e4291872c0dadc62657f668ddf2.jpg) + +![](images/1bab931afc1bde3e962f5d768f56ef3f5a2d7adbbb4c109d05c082412097e3a9.jpg) +Pose-to-Video Generation + +![](images/fea104d67a494498303501440134ab397c9eb1f9675169296cbbd13c4ef46712.jpg) + +![](images/e70b871a72fb3f17152b178008bc2601392060f6a4328678ac72df32b81cdc61.jpg) + +![](images/235df14e4cc9e27bf1fd654c6e21af1c1b1d7fa5afdb3e00cc6a1437e8843926.jpg) + +![](images/d7ba48354bf5fcd85a86e32ad98c15ff25e7dc99e929dd3ef6aa607dec419475.jpg) +Figure 1. Illustration of HumanDreamer. The human-motion video generation is decoupled into two steps: Text-to-Pose generation and Pose-to-Video generation. The decoupled process integrates the flexibility of text control and the controllability of pose guidance. + +![](images/7ad538ec01880978193cbdb436ba3dba39d3f443523d990691fedf6052e754fa.jpg) + +![](images/46ff04edc0bab07374dea19d2fdfd72686944221f216e398e5dcaf146389d08a.jpg) + +![](images/07f5ce86b40c3105d2cc62afb95a08c7c7f214aba3271a3c7cea434c4315b78c.jpg) + +![](images/7815a721cd409d9f706e6362a7f174275509cd0e45c9b9b5272119058240b1bd.jpg) + +![](images/41ec9997937ddcb56dbb189f33ab8f6938b27efc9a3201cd7fa3983a5543a29f.jpg) + +![](images/4e999f3cd01a9303c07fe61f8c505da6d7a4b18c3278fa87d11620f5a9d3ab2e.jpg) + +# Abstract + +man body movements. While some approaches have attempted to drive human-centric video generation explicitly through pose control, these methods typically rely on poses derived from existing videos, thereby lacking flexibility. To address this, we propose HumanDreamer, a decoupled human video generation framework that first generates diverse poses from text prompts and then leverages these poses to generate human-motion videos. Specifically, we propose MotionVid, the largest dataset for human-motion pose generation. Based on the dataset, we present MotionDiT, which is trained to generate structured human-motion poses from text prompts. Besides, a novel LAMA loss is introduced, which together contribute to a significant improvement in FID by $62.4\%$ , along with respective enhancements in R-precision for top1, top2, and top3 by $41.8\%$ , $26.3\%$ , and $18.3\%$ , thereby advancing both the Text-to-Pose control accuracy and FID metrics. Our experiments across various Pose-to-Video baselines demonstrate that the poses generated by our method can produce diverse and high-quality human-motion videos. Furthermore, our model can facilitate other downstream tasks, such as pose sequence prediction and 2D-3D motion lifting. + +# 1. Introduction + +Generating human-motion videos remains a particularly challenging task due to the inherent complexity of modeling human body movements. Despite current advancements in generative modeling [9, 13, 31, 48, 71], state-of-the-art video generation models [3, 16, 19, 36, 59, 61], equipped with billions of parameters and trained on millions of video and image data, still struggle to capture human body movements, frequently resulting in fragmented or unrealistic portrayals. This limitation is exacerbated when controlling human videos via textual conditions, highlighting the fundamental challenge of directly mapping text prompts to human visual data. + +To enhance the generation quality, approaches such asAnimate Anyone [20], UniAnimate [53], Mimic Motion [66], Champ [69] andAnimate-X [49] explicitly generate human-motion videos through pose control. By utilizing the pose-guided generation, these methods effectively reduce the complexities associated with human-motion videos. However, a notable limitation of this approach is its reliance on human-motion poses derived from existing videos, which restricts the flexibility. + +Therefore, we propose HumanDreamer, a decoupled human-motion video generation framework that first generates human-motion poses from text prompts and subsequently produces human-motion videos based on the generated poses. The motivation behind this decoupled approach lies in that Text-to-Pose presents a more manageable search space compared to the direct learning of text-to-pixel repre + +sensation. As shown in Fig. 1, this decoupled framework facilitates a more effective generation of human movements. Additionally, the proposed HumanDreamer utilizes text as the input, offering greater flexibility than [20, 34, 49, 53, 66] that directly relies on pre-defined poses. Specifically, we construct a 1.2 million text-pose pairs dataset MotionVid, which is the largest dataset for human-motion pose generation. To ensure dataset quality, a comprehensive data cleaning pipeline is introduced, where the cleaning factors involve body movement amplitude, human presence duration, facial visibility, and the proportion of human figures. This rigorous cleaning guarantees that the dataset is reliable for training Text-to-Pose tasks. Based on the dataset, we propose the MotionDiT to generate structured human-motion poses from text prompts. The MotionDiT integrates a global attention block designed to extract the global characteristics of the entire pose sequence, alongside a local feature aggregation mechanism that captures information from adjacent pose points, effectively combining global and local perspectives to enhance the quality and coherence of the generated poses. Additionally, the LAMA loss is utilized in the latent space to align the semantic features of the ground truth motion with those produced by MotionDiT, which not only improves the interpretability of the model but also boosts its overall performance. The introduced techniques collectively result in a notable $62.4\%$ enhancement in FID, accompanied by respective increases in R-precision for the top1, top2, and top3 categories by $41.8\%$ , $26.3\%$ , and $18.3\%$ . + +The primary contributions of this work are as follows: (1) We present HumanDreamer, the first decoupled framework for human-motion video generation, which integrates the flexibility of text control with the controllability of pose guidance. (2) We propose the MotionVid, the largest dataset for human-motion pose generation. In MotionVid, we conduct a comprehensive data annotation and data cleaning to ensure data reliability. (3) The MotionDiT is introduced to generate diverse human-motion poses under text control. To enhance the generation process, we propose the LAMA loss, which significantly improves pose fidelity and diversity. (4) Through extensive experiments, we show that MotionDiT and the LAMA loss improve control accuracy and FID by $62.4\%$ , with top-k R-precision gains of $41.8\%$ , $26.3\%$ , and $18.3\%$ for top-1, top-2, and top-3. The generated poses also support downstream tasks like pose sequence prediction and 2D-to-3D motion enhancement. + +# 2. Related Work + +# 2.1. Human-motion Pose Datasets + +Human motion datasets are essential for the development of human-centric perception [7, 8, 21, 30, 44, 57] or generation [6, 29, 32, 51, 52, 54, 55, 63, 68] tasks. Specifically, + +![](images/22a1e483c3189286dfed3e85906f8d2d67c0438971256fb3eb066be9130f0993.jpg) +Figure 2. The data cleaning and annotation pipeline for MotionVid begins with raw data sourced from public datasets and the internet, which is then segmented into video clips. To ensure high-quality data, we apply video quality filter, data annotation, human quality filter, and caption quality filter. + +the KIT Motion-Language Dataset [39] provides sequence-level descriptions to support multi-modal motion tasks. In contrast, the HumanML3D [15] dataset, which is built upon the AMASS [35] and HumanAct12 [14] datasets, offers richer textual annotations and a wider variety of activities. Additionally, Motion-X [29] establishes a robust annotation protocol and is the first to compile a comprehensive fine-grained 3D whole-body motion dataset derived from extensive scene captures. As a result, existing text-driven 3D motion datasets often lack sufficient volume and diversity, which limits their scalability. A dataset that shares similarities with ours is Holistic-Motion2D [55], which features 1M text-pose pairs but only 16K of these are publicly accessible. However, the proposed MotionVid significantly surpasses this with a larger and more diverse set of 1.2M text-pose data, coupled with a more thorough data cleaning process that ensures the dataset's reliability for training Text-to-Pose tasks. + +# 2.2. Human-motion Pose Generation Methods + +Text-driven human motion generation, which converts textual descriptions into motion sequences, has gained significant attention in recent years. Early approaches focus on aligning motion and text through shared latent spaces. For instance, MotionCLIP [51] enhances autoencoder generalization by aligning the latent space with the expressive CLIP [40] embedding. T2M-GPT [63] frames the task as predicting discrete motion indices. More recently, Diffusion-based models, including MDM [52], MotionDiffuse [65], and DiffGesture [68], have set new performance standards. MLD [6] introduces a latent Diffusion model that generates diverse, realistic motions, while Humantomato [32] expands the task to whole-body motion generation, advancing GPT-like models for finer motion sequences. However, limited 3D data hampers model diversity and detail. Tender [55] overcomes this by training on extensive 2D pose data. In contrast, the proposed MotionDiT employs a Diffusion Transformer to better capture both local and global pose relationships, improving human-motion pose generation. + +# 3. HumanDreamer + +In this section, we first introduce the proposed MotionVid dataset that facilitates the training of human-motion pose generation. Next, we elaborate on the details of the decoupled Text-to-Pose generation and Pose-to-Video generation. + +# 3.1. MotionVid + +Existing datasets face limitations in training Text-to-Pose tasks. For instance, current 3D datasets [14, 35] primarily capture torso movements, lacking essential full-body keypoints, particularly for the face and hands, which makes them unsuitable for diverse pose generation. Although the Motion-X dataset [29] provides full-body 3D data, its scale is limited to only 81.1K samples, as the expensive annotation of 3D data. For the 2D full-body dataset, the Holistic-Motion2D dataset [55] includes approximately 1M samples, though only 16K of these are publicly accessible. Therefore, there is a pressing need for a comprehensive and publicly available human-motion pose dataset that facilitates the training of Text-to-Pose tasks. To address these limitations, we introduce the largest human-motion pose dataset MotionVid, which contains approximately 1.2M data pairs, with videos of varying resolutions and durations exceeding 64 frames. Besides, we provide a robust data collection pipeline to obtain high-quality textpose pairs from diverse and complex data sources. + +Data Source. We construct a diverse set of text-pose pairs by curating data from various publicly available video datasets, including action recognition datasets Kinetics-400 [23], Kinetics-700 [4], ActivityNet-200 [11], and Something-Something V2 [12], as well as human-related datasets such as CAER [26], DFEW [22], UBody [28], Charades [45], Charades-Ego [46], HAA500 [10], and HMDB51 [25]. This selection totals $\sim 6.6\mathrm{M}$ samples. To further enhance scene and subject diversity, we add $3.4\mathrm{M}$ videos sourced from the internet. Given the complexity of these videos, we perform extensive annotation and cleaning to ensure high-quality data. As shown in Fig. 2, we first use a shot transition detection model [47] to segment the video clips, followed by the video quality filter, data annotation, + +![](images/f8b8a27cb1d50360cbea2f8058c48f9ffcd609658e0fc405136e29b822d1abdc.jpg) +Figure 3. Training pipeline of the proposed Text-to-Pose generation. Pose data are encoded in latent space via the Pose VAE, which are then processed by the proposed MotionDiT, where local feature aggregation and global attention are utilized to capture information from the entire pose sequence. Finally, the LAMA loss is calculated via the proposed CLoP, which enhances the training of MotionDiT. + +human quality filter, and caption quality filter. + +Video Quality Filter. After acquiring a large number of videos, it is necessary to verify the video quality. Drawing inspiration from SVD's data cleaning strategy [3], we propose a Video Quality Filter to select high-quality video clips. Specifically, we use the GMFlow method [56] to estimate optical flow and filter out videos with insufficient movement intensity. Subsequently, we employ the method described in [1] to detect text regions and remove videos where text occupies a significant portion of the frame. We also utilize LAION-AI's aesthetic predictor [41] to compute aesthetic scores and eliminate videos with low aesthetic quality. Finally, the Laplacian operator [2] is applied to assess blur intensity, and videos deemed excessively blurry are discarded. As a result, approximately $50\%$ of the data is filtered out (see supplementary materials for more details). + +Data Annotation. We first prompt the Vision Language Model (VLM) [5] to generate action-oriented captions that describe human activities. Besides, for each frame, we extract 2D poses using the DWPose model [58]. This process yields a large number of text-pose pairs. However, the annotated data pairs are unsuitable for training Text-to-Pose tasks. This is due to several factors: the VLM's potential inaccuracy in describing multi-person actions, difficulty in recognizing poses occupying smaller areas in the frame, bias toward static poses from individuals with minimal motion, and limited detail in poses where faces are not visible. + +Human Quality Filter. To further refine the dataset, we implement a Human Quality Filter, focusing on selecting data based on human-related characteristics. We calculate the difference between two consecutive frames' 2D poses to filter out sequences with insufficient motion magnitude. Additionally, we compute the ratio of the human detection bounding box to the entire frame to remove clips where the human coverage is too small. At this stage, we focus on + +simpler scenarios by selecting scenes with a human count of one and ensuring face visibility for training. Based on the Human Quality Filter, approximately $75\%$ of the data is filtered out (see supplementary materials for more details). + +Caption Quality Filter. To enhance the quality of caption annotations, we apply a Caption Quality Filter to refine alignment accuracy. Specifically, using the proposed CLoP (see Sec. 3.2 for details), we compute the semantic similarity between text and poses, filtering data based on the similarity score. This ensures that each text annotation semantically aligns with the corresponding 2D pose data, thereby improving overall data quality. + +# 3.2. Text-to-Pose + +Fig. 3 illustrates the training pipeline of the proposed Text-to-Pose generation. In this pipeline, pose data are first encoded into a latent space via the Pose VAE and subsequently processed by the proposed MotionDiT model. Within MotionDiT, local feature aggregation and global attention mechanisms are employed to capture comprehensive information across the entire pose sequence. Finally, LAMA loss is computed using the proposed CLoP, which further refines the training of MotionDiT. In the next, we introduce the Pose VAE, MotionDiT and CLoP. + +Pose VAE. Variational Autoencoders (VAEs) [17, 24, 59, 67] have been widely applied in 2D image and 3D video processing, but their use for pose data remains underexplored, presenting a rich area for research. We represent a pose sequence as $\mathbf{p} \in \mathbb{R}^{f \times N \times 3}$ , where $f$ is the frame count, $N$ the keypoints, and dimension 3 includes the $x$ , $y$ coordinates and confidence scores. Given the sequential structure and frame-to-frame correlations, a 1D convolutional encoder with multi-layer downsampling effectively extracts temporal features, while symmetric decoder-encoder architecture with skip connections aids reconstruction. Confi + +dence scores of each keypoint help mitigate occlusion issues, optimizing the model via KL divergence and reconstruction loss. Further details are available in the supplementary materials. + +MotionDiT. The proposed MotionDiT extends the Diffusion Transformer (DiT) [37] architecture specifically tailored to construct associations between human-motion poses and text control. In the proposed MotionDiT, we incorporate a global attention block to capture global spatial-temporal patterns inherent in human poses. Besides, the MotionDiT involves a local feature aggregation module to strengthen correlations between adjacent joints. Specifically, we first obtain the latent representation from the Pose VAE, which is then processed via a patch-embedding: + +$$ +l _ {p} = \mathcal {F} _ {\text {P a t c h}} (\mathcal {E} (\mathbf {p})), \tag {1} +$$ + +where $\mathcal{F}_{\mathrm{Patch}}(\cdot)$ is the patch embedding operation [33], and $\mathcal{E}(\cdot)$ is the Pose VAE encoder. Then we employ Diffusion blocks to process $l_{p}$ . As shown in Fig. 3, in each block, the local feature aggregation module is first utilized to strengthen local feature correlation: + +$$ +l _ {p} = \mathcal {F} _ {\mathrm {s a}} \left(\mathcal {F} _ {\mathrm {r e s}} \left(l _ {p}\right) + l _ {p}\right), \tag {2} +$$ + +where $\mathcal{F}_{\mathrm{res}}$ is the 1D ResNet Block with kernel size 3, and $\mathcal{F}_{\mathrm{sa}}$ is the spatial self attention. For text-based control, textual features obtained from CLIP [40] are then integrated into the network through cross-attention. Besides, following conventional video Diffusion architecture [3, 33], temporal attention is used to ensure continuity of pose features. Notably, it is essential for the model to capture global pose feature information. Although the model already employs spatiotemporal attention mechanisms, certain pose characteristics at one location and time may influence other locations at different times. Thus, global attention is required to extract these global features among internal pose keypoints. Instead of incorporating global attention in every attention module, we decide to apply it to the output of the central layer of the network to reduce computational complexity. Specifically, we reshape the latent output of the middle layer to $z \in \mathbb{R}^{(f \times n) \times c}$ , where $n$ is the number of points $N$ divided by the down-sampling factor $r$ and $c$ is the number of latent channels. We then perform self-attention across all frames $f$ and all points $n$ . Finally, we reshape the output back to its original dimensions. + +Following [33], we employ the noise prediction loss to optimize MotionDiT, thus the model output is: + +$$ +\epsilon_ {\text {p r e d}} = g _ {\theta} \left(z _ {t}, t, s\right), \tag {3} +$$ + +where $g_{\theta}$ represents the MotionDiT parameterized by $\theta$ , $z_{t}$ is the noisy latent variable at time step $t$ , and $s$ denotes the conditional input (e.g., textual features), we compute the noise prediction at time step $t$ . The model learns to predict + +the noise $\epsilon_{\mathrm{pred}}$ conditioned on the input. We calculate the mean squared error (MSE) loss between the predicted noise $\epsilon_{\mathrm{pred}}$ and the true noise $\epsilon$ as: + +$$ +\mathcal {L} _ {d} = \mathbb {E} _ {t, z _ {0}, \epsilon} \left[ \| \epsilon - \epsilon_ {\text {p r e d}} \| ^ {2} \right], \tag {4} +$$ + +where $\mathcal{L}_d$ represents the loss over all time steps $t$ , true noise $\epsilon \sim \mathcal{N}(0, I)$ , and ground truth latent variable $z_0$ . + +CLoP. Additionally, prior works [27, 60] emphasize the significance of intermediate feature representations, which inspire us to propose a LAtent seMantic Alignment (LAMA) loss aimed at enhancing text-based motion control. We expect this term to provide benefits in two key ways: (1) By aligning the latent space of MotionDiT with a largescale pre-trained motion encoder, we can improve both fidelity and diversity of the generated 2D pose; (2) Aligning with the pre-trained pose encoder that is already aligned with text inherently strengthens the influence of text control on motion generation. Unlike CLIP [40], there are currently no readily available large-scale pre-trained models for aligning motion and text. Leveraging the MotionVid dataset, we introduce Contrastive Language-Motion Pretraining (CLoP), which aligns text with 2D pose. This improves the evaluation of 2D pose and latent semantic alignment. Specifically, inspired by CLIP, we employ a similar mechanism to extract textual features. The extracted text features and 2D pose data are then processed through dedicated text encoder $\mathcal{F}_e(\cdot)$ and pose encoder $\mathcal{F}_p(\cdot)$ , respectively. The alignment of the two modalities is facilitated by optimizing a contrastive loss function [40]. Our architecture represents a 2D extension of the approach in [38]. Unlike the original model, we utilize CLIP for poses representation as $\mathbf{p} \in \mathbb{R}^{f \times N \times 3}$ with its corresponding text e, capturing frame-wise keypoint coordinates and confidence scores instead of 3D features, the CLoP training objective as follows: + +$$ +\mathcal {L} _ {c} = \ell_ {c e} \left(\ell_ {2} \left(\mathbf {h} _ {e} \mathbf {h} _ {p} ^ {T}\right); y\right) + \ell_ {c e} \left(\ell_ {2} \left(\mathbf {h} _ {p} \mathbf {h} _ {e} ^ {T}\right); y\right) / 2, \tag {5} +$$ + +where $\ell_2(\cdot)$ denote the $\ell_2$ normalization and $\ell_{ce}(\cdot ;y)$ represent the cross-entropy loss with ground truth label $y$ . $\mathbf{h}_e = \mathcal{F}_e(\mathbf{e})\mathbf{W}_e$ is the projected text embedding, calculated by projection matrix $\mathbf{W}_e$ . $\mathbf{h}_p = \mathcal{F}_p(\mathbf{p})\mathbf{W}_p$ is the projected pose embedding. + +Based on the CLoP, we evaluate the dissimilarity within the feature space (e.g., MSE or cosine similarity metrics), achieving the benefits of the previously mentioned LAMA loss. Rather than using prior methods to extract structural features, we focus on the semantic features of motion, as motion's structural information is simpler while textual information is richer. We also employ a feature projection using a two-layer MLP $g_{\omega}(\cdot)$ parameterized by $\omega$ to project the latent feature space learned by MotionDiT. As a result, the LAMA loss $\mathcal{L}_f$ can be formulated: + +$$ +\mathcal {L} _ {f} = d \left(g _ {\omega} \left(\mathbf {h} _ {d} ^ {l}\right), \mathbf {h} _ {p}\right), \tag {6} +$$ + +Table 1. Comparison to Other State-of-the-Art Methods on the MotionVid Subset. The metrics demonstrate that our method outperforms others in terms of pose-text alignment and diversity. Bold indicates the best result. + +
MethodFID ↓Rp-top1 ↑Rp-top2 ↑Rp-top3 ↑Diversity ↑MM Dist ↓MM ↑
T2M-GPT [63]496.9070.1760.2990.39564.00042.95014.480
PriorMDM [43]538.9490.1560.2700.36662.33852.55044.135
MLD [6]396.9490.3180.5050.62864.44240.19657.862
Ours149.0070.4510.6380.74368.22032.76160.460
+ +where $\mathbf{h}_d^l$ is the $l$ -th layer latent representation of MotionDiT, $d(\cdot ,\cdot)$ is the CLoP function. Consequently, the overall objective $\mathcal{L}$ of MotionDiT can be formulated as: + +$$ +\mathcal {L} = \mathcal {L} _ {d} + \lambda_ {f} \mathcal {L} _ {f}, \tag {7} +$$ + +where $\lambda_{f}$ is a hyperparameter to trade off the alignment and denoising process. + +# 3.3. Pose-to-Video + +We propose a model for Pose-to-Video generation, capable of producing human-motion videos based on an initial frame image and a pose sequence. The model leverages [59] as its backbone, maintaining consistent parameters with the original model, incorporates conditional control inspired by [64], and employs data augmentation techniques from [49] to enhance performance. Notably, all training was conducted on the MotionVid dataset, demonstrating its versatility for both Text-to-Pose and Pose-to-Video tasks. It is worth noting that other Pose-to-Video models, such as [20, 49, 66], are also capable of handling this task. For further details on the model architecture, please refer to the supplementary materials. + +# 4. Experiments + +# 4.1. Experimental setup + +Dataset. For computational efficiency, we select a representative subset of 50K samples from MotionVid, covering diverse actions like dancing, squatting, and lifting to ensure diversity and expedite experimentation. This subset maintains result representativeness while enabling efficient testing. We also conducted scaling experiments to assess model performance across various dataset sizes. Qualitative visualization experiments utilize a dataset of 1.2M samples. + +Implementation Details. For Text-to-Pose, our motion data consists of 128 points per frame, covering the body, face, and hands. The input is structured as $f \times 128 \times 3$ , representing $f$ frames, with 128 points in each frame and 3 channels that encode image coordinates (x, y) along with confidence scores. To maintain uniformity across samples, all sequences are cropped to 64 frames. For Pose VAE, we employ a downsampling factor $r = 8$ , yielding a latent representation $z \in \mathbb{R}^{f \times 16 \times 3}$ . In MotionDiT, we apply a two- + +fold downsampling in the patch embedding. The network consists of 13 layers, with text features extracted from the SD2.1 CLIP model [42] as our text encoder, resulting in a conditional embedding $t_{emb} \in \mathbb{R}^{1 \times 1024}$ . The AdamW optimizer is used with a learning rate of $1 \times 10^{-5}$ . The training is conducted on 8 H20 GPUs, with inference performed on a single H20. + +Evaluation Metrics. To evaluate generated motion data quality, several metrics are employed, consistent with previous works [6, 15, 43, 52, 55, 63]. Frechet Inception Distance (FID) [18] assesses distributional similarity between ground truth and generated motions, using features from a custom-trained CLOP model. MultiModality evaluates diversity for a single text input, while Diversity [15] calculates variance across all motions to quantify output diversity. R-precision measures top-1, top-2, and top-3 recall rates in motion retrieval, evaluating alignment accuracy. MultiModality Distance gauges the distance between textual inputs and generated motions, indicating adherence to prompts. Further details are available in the supplementary materials. + +# 4.2. Results on Text-to-Pose + +Currently, Tender [55] is the only method for generating 2D poses based on text but its code and dataset are unavailable. Thus, we select three SOTA 3D motion generation methods for comparison: T2M-GPT [62], PriorMDM [43], and MLD [6], converting their inputs to 2D poses. All methods are trained on the MotionVid subset with proposed settings to standardize evaluation metrics. Comparison results (Tab. 1) show our model outperforms others, achieving $62.4\%$ improvement in FID, $41.8\%$ in Rp-top1, $26.3\%$ in Rp-top2, and $18.3\%$ in Rp-top3. These results indicate our model generates semantically aligned outputs with high diversity. + +The visual comparison with other methods is presented in Fig. 4, and the results demonstrate a significant improvement over existing models. Our method produces poses that better align with the textual constraints, with keypoints remaining intact and exhibiting minimal motion jitter. A more detailed visual comparison can be found in the supplementary materials. + +Table 2. The ablation study presents four configurations, progressively adding components of Local, Global, and LAMA loss to the original model. As we move from the initial configuration to the fully enhanced model, performance metrics consistently improve, highlighting the positive impact of each additional component. + +
MethodFID ↓Rp-top1 ↑Rp-top2 ↑Rp-top3 ↑Diversity ↑MM Dist ↓MM ↑
Vanilla DiT283.0910.3670.5610.68266.58737.04621.461
DiT+Local183.2130.4150.6060.72167.59134.29141.890
DiT+Local+Global162.0220.4330.6200.73067.35733.26957.848
DiT+Local+Global+LAMA149.0070.4510.6380.74368.22032.76160.460
+ +A man is holding his head in his hands and continues to do so while looking down at them. + +A woman is dancing, moving her arms and legs around. + +![](images/b15f958cd9d769ce18e89f1e707a24f5f3dbc1f38caddf77df9fbf9862682219.jpg) +Figure 4. Visualization results compared to SOTA Text-to-Pose methods. The results demonstrate that our model significantly outperforms other models. Our method generates poses that are more consistent with the text constraints, with keypoints maintaining their integrity and minimal motion jitter. For a better visual comparison, please refer to the supplementary materials. + +![](images/7f254fa475fbd039c70fbdacc559df5196b80fc5204c1f5c278fff1447034b64.jpg) + +# 4.3. Ablation Studies + +To validate the efficacy of the three modules proposed in the MotionDiT (i.e., the local feature aggregation, the global attention, and the LAMA loss), we conduct a series of ablation studies. The outcomes are summarized as follows: + +Local Feature Aggregation. Removing the local attention block and integrating a DiT structure similar to Latte [33] in each attention mechanism, we showcase our method's efficacy in capturing local details. The performance improvements, shown in Tab. 2, confirm the robustness of our local feature extraction approach. + +Global Attention Block. To assess the impact of our global attention mechanism, we perform an ablation test where this component is removed from the model. The findings, detailed in Tab. 2, indicate a notable decline in performance, which underscores the importance of the global attention mechanism in enhancing the model's ability to capture long-range dependencies and global context. + +LAMA Loss. We investigate the role of the semantic alignment loss by training MotionDiT without it. The analysis in + +Tab. 2 shows a significant performance drop, validating the LAMA loss's critical role in guiding optimization towards accurate and stable solutions. These ablation studies provide strong evidence for the proposed algorithms' effectiveness and necessity in enhancing model performance across various aspects. + +# 4.4. Scaling Law Results + +We expand the training set size, and as shown in Tab. 3, our metrics improved with the increase in data volume. This demonstrates the validity of using large-scale 2D pose data and suggests that model performance can be further enhanced through an efficient and cost-effective data collection pipeline. + +# 4.5. Qualitative Results on Pose-to-Video + +In our qualitative experiments, we use the generated motion sequences as templates and provide them to the Pose-to-Video model. The specific video results are shown in Fig. 5. Compared to state-of-the-art text-to-video models, + +![](images/9794ba2c9d729f7714a123d1c105c760f5e22d64f6e735046839a90e5c5d2e2d.jpg) +Figure 5. Visualization results compared to SOTA Text-to-Video methods. Mochi1 [50] and CogVideoX [59] exhibit issues such as body distortion, weak motion continuity, and neglecting facial generation. In contrast, HumanDreamer is able to generate more coherent and consistent videos with smoother transitions and better attention to details such as facial expressions. For a better visual comparison, please refer to the supplementary materials. + +![](images/cf83992ebd1916a8a86ed0092e94b113f5009d74d67720e875a104f1b611ece1.jpg) + +Table 3. By increasing the amount of training data, we observe an improvement in model performance, which validates the potential for rapid scalability using large-scale 2D motion data. + +
DatasizeRp-top1 ↑Rp-top2 ↑Rp-top3 ↑MM Dist ↓
50K(4%)0.4510.6380.74332.761
250K(20%)0.4520.6410.75132.666
500K(40%)0.4640.6570.76632.399
1M(80%)0.4920.6710.76931.116
1.25M(100%)0.5130.6940.79130.139
+ +our model exhibits larger and more dynamic movements, with the characters' actions more accurately reflecting the textual descriptions. + +![](images/4be80cf722b8d40a1101efcbb61b9c1f64e314890ec4d99a43bdff41a57edb16.jpg) + +![](images/e22d0e599981c539142a464e74693c402e4052585c33e9767107d6cb4cf2fd25.jpg) +Figure 6. The visualization results demonstrate that using the initial frame pose, different prompt texts can guide the generation of distinct pose sequences. + +# 4.6. Other Downstream Tasks + +In addition to the Pose-to-Video task, the proposed Text-to-Pose component can also be utilized for pose sequence prediction, 2D-3D motion lifting tasks. + +Pose Sequence Prediction. In scenarios with partial pose sequences, the Text-to-Pose model can infer and generate missing parts by conditioning on existing poses and textual movement descriptions. This enables the synthesis of co- + +herent pose continuations. Given initial and final frames, the model predicts intermediate pose states, completing the sequence (Fig. 6). This is especially useful in animation for supplementing incomplete pose data. + +2D-3D Motion Lifting. Models like [70] can enhance motion data dimensionality, converting 2D motions into realistic 3D motions (Fig. 7). With textual descriptions specifying depth or spatial aspects, the Text-to-Pose system generates richer 3D motion representations. This is crucial for virtual reality (VR) and augmented reality (AR) environments, where realistic 3D motion significantly improves user experience. + +![](images/8704c8d46a9d0423f47135731a6dc66a7d570932a0b85f220fc02a42437719f4.jpg) +Figure 7. Visualization results on 2D-3D Motion Lifting. + +# 5. Discussion and Conclusion + +In this study, we present HumanDreamer, a pioneering decoupled framework for generating human-motion videos that merges text control flexibility with pose guidance controllability. Utilizing MotionVid, the largest dataset for human-motion pose generation, we train MotionDiT for producing structured poses. We introduce LAMA loss to improve semantic alignment, ensuring coherent outputs. + +Experimental results indicate that using generated poses in Pose-to-Video yields high-quality, diverse human-motion videos, surpassing current benchmarks. These findings confirm the effectiveness and adaptability of our decoupled framework, facilitating versatile video generation. + +# References + +[1] Youngmin Baek, Bado Lee, Dongyoon Han, Sangdoo Yun, and Hwalsuk Lee. Character region awareness for text detection. In CVPR, 2019. 4 +[2] Raghav Bansal, Gaurav Raj, and Tanupriya Choudhury. Blur image detection using laplacian operator and open-cv. In SMART, 2016. 4 +[3] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2, 4, 5 +[4] Joao Carreira, Eric Noland, Chloe Hillier, and Andrew Zisserman. A short note on the kinetics-700 human action dataset, 2022. 3 +[5] Lin Chen, Xilin Wei, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Bin Lin, Zhenyu Tang, et al. Sharegpt4video: Improving video understanding and generation with better captions. arXiv preprint arXiv:2406.04325, 2024. 4 +[6] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, and Gang Yu. Executing your commands via motion diffusion in latent space. In CVPR, 2023. 2, 3, 6 +[7] Yuxin Chen, Ziqi Zhang, Chunfeng Yuan, Bing Li, Ying Deng, and Weiming Hu. Channel-wise topology refinement graph convolution for skeleton-based action recognition. In ICCV, 2021. 2 +[8] Ziwei Chen, Qiang Li, Xiaofeng Wang, and Wankou Yang. Liftedcl: Lifting contrastive learning for human-centric perception. In ICLR, 2023. 2 +[9] Joseph Cho, Fachrina Dewi Puspitasari, Sheng Zheng, Jingyao Zheng, Lik-Hang Lee, Tae-Ho Kim, Choong Seon Hong, and Chaoning Zhang. Sora as an agi world model? a complete survey on text-to-video generation. arXiv preprint arXiv:2403.05131, 2024. 2 +[10] Jihoon Chung, Cheng hsin Wu, Hsuan ru Yang, Yu-Wing Tai, and Chi-Keung Tang. Haa500: Human-centric atomic action dataset with curated videos. In ICCV, 2011. 3 +[11] Bernard Ghanem Fabian Caba Heilbron, Victor Escorcia and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In CVPR, 2015. 3 +[12] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The" something something" video database for learning and evaluating visual common sense. In ICCV, 2017. 3 +[13] Yanchen Guan, Haicheng Liao, Zhenning Li, Guohui Zhang, and Chengzhong Xu. World models for autonomous driving: An initial survey. arXiv preprint arXiv:2403.02622, 2024. 2 +[14] Chuan Guo, Xinxin Zuo, Sen Wang, Shihao Zou, Qingyao Sun, Annan Deng, Minglun Gong, and Li Cheng. Action2motion: Conditioned generation of 3d human motions. In ACM MM, 2020. 3 +[15] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In CVPR, 2022. 3, 6 + +[16] Agrim Gupta, Lijun Yu, Kihyuk Sohn, Xiuye Gu, Meera Hahn, Li Fei-Fei, Irfan Essa, Lu Jiang, and José Lezama. Photorealistic video generation with diffusion models. arXiv preprint arXiv:2312.06662, 2023. 2 +[17] Louay Hazami, Rayhane Mama, and Ragavan Thurairatnam. Efficient-vdvae: Less is more. arXiv preprint arXiv:2203.13751, 2022. 4 +[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS, 2017. 6 +[19] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 2 +[20] Li Hu. Animate anyone: Consistent and controllable image-to-video synthesis for character animation. In CVPR, 2024. 2, 6 +[21] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. TPAMI, 2013. 2 +[22] Xingxun Jiang, Yuan Zong, Wenming Zheng, Chuangao Tang, Wanchuang Xia, Cheng Lu, and Jiateng Liu. Dfew: A large-scale database for recognizing dynamic facial expressions in the wild. In ACM MM, 2020. 3 +[23] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, Mustafa Suleyman, and Andrew Zisserman. The kinetics human action video dataset, 2017. 3 +[24] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 4 +[25] H. Kuehne, H. Jhuang, E. Garrote, T. Poggio, and T. Serre. HMDB: a large video database for human motion recognition. In ICCV, 2011. 3 +[26] Jiyoung Lee, Seungryong Kim, Sunok Kim, Jungin Park, and Kwanghoonn Sohn. Context-aware emotion recognition networks. In ICCV, 2019. 3 +[27] Alexander C Li, Mihir Prabhudesai, Shivam Duggal, Ellis Brown, and Deepak Pathak. Your diffusion model is secretly a zero-shot classifier. In ICCV, 2023. 5 +[28] Jing Lin, Ailing Zeng, Haoqian Wang, Lei Zhang, and Yu Li. One-stage 3d whole-body mesh recovery with component aware transformer. CVPR, 2023. 3 +[29] Jing Lin, Ailing Zeng, Shunlin Lu, Yuanhao Cai, Ruimao Zhang, Haoqian Wang, and Lei Zhang. Motion-x: A large-scale 3d expressive whole-body human motion dataset. NeurIPS, 2024. 2, 3 +[30] Jun Liu, Amir Shahroudy, Mauricio Perez, Gang Wang, Ling-Yu Duan, and Alex C Kot. Ntu rgb+ d 120: A large-scale benchmark for 3d human activity understanding. TPAMI, 2019. 2 +[31] Yixin Liu, Kai Zhang, Yuan Li, Zhiling Yan, Chujie Gao, Ruoxi Chen, Zhengqing Yuan, Yue Huang, Hanchi Sun, Jianfeng Gao, et al. Sora: A review on background, technology, limitations, and opportunities of large vision models. arXiv preprint arXiv:2402.17177, 2024. 2 + +[32] Shunlin Lu, Ling-Hao Chen, Ailing Zeng, Jing Lin, Ruimao Zhang, Lei Zhang, and Heung-Yeung Shum. Humantomato: Text-aligned whole-body motion generation. arXiv preprint arXiv:2310.12978, 2023. 2, 3 +[33] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 5, 7 +[34] Yue Ma, Yingqing He, Xiaodong Cun, Xintao Wang, Siran Chen, Xiu Li, and Qifeng Chen. Follow your pose: Pose-guided text-to-video generation using pose-free videos. In AAAI, 2024. 2 +[35] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In ICCV, 2019. 3 +[36] Haomiao Ni, Changhao Shi, Kai Li, Sharon X Huang, and Martin Renqiang Min. Conditional image-to-video generation with latent flow diffusion models. In CVPR, 2023. 2 +[37] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, 2023. 5 +[38] Mathis Petrovich, Michael J. Black, and Gül Varol. TMR: Text-to-motion retrieval using contrastive 3D human motion synthesis. In ICCV, 2023. 5 +[39] Matthias Plappert, Christian Mandery, and Tamim Asfour. The kit motion-language dataset. Big data, 2016. 3 +[40] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 3, 5 +[41] Beaumont Romain and Schuhmann Christoph. Laion aesthetics predictor v1. 2022. 4 +[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 6 +[43] Yoni Shafir, Guy Tevet, Roy Kapon, and Amit Haim Bermano. Human motion diffusion as a generative prior. In ICLR, 2024. 6 +[44] Amir Shahroudy, Jun Liu, Tian-Tsong Ng, and Gang Wang. Ntu $\mathrm{rgb + }$ d: A large scale dataset for 3d human activity analysis. In CVPR, 2016. 2 +[45] Gunnar A. Sigurdsson, Gül Varol, Xiaolong Wang, Ali Farhadi, Ivan Laptev, and Abhinav Gupta. Hollywood in homes: Crowdsourcing data collection for activity understanding, 2016. 3 +[46] Gunnar A. Sigurdsson, Abhinav Gupta, Cordelia Schmid, Ali Farhadi, and Karteek Alahari. Charades-ego: A large-scale dataset of paired third and first person videos. In ArXiv, 2018. 3 +[47] Tomáš Souček and Jakub Lokoc. Transnet v2: An effective deep network architecture for fast shot transition detection. arXiv preprint arXiv:2008.04838, 2020. 3 +[48] Rui Sun, Yumin Zhang, Tejal Shah, Jiahao Sun, Shuoying Zhang, Wenqi Li, Haoran Duan, Bo Wei, and Rajiv Ranjan. From sora what we can see: A survey of text-to-video generation. arXiv preprint arXiv:2405.10674, 2024. 2 + +[49] Shuai Tan, Biao Gong, Xiang Wang, Shiwei Zhang, Dandan Zheng, Ruobing Zheng, Kecheng Zheng, Jingdong Chen, and Ming Yang. Animate-x: Universal character image animation with enhanced motion representation. arXiv preprint arXiv:2410.10306, 2024. 2, 6 +[50] Genmo Team. Mochi 1. https://github.com/genmoai/models, 2024.8 +[51] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. In ECCV, 2022. 2, 3 +[52] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In ICLR, 2023. 2, 3, 6 +[53] Xiang Wang, Shiwei Zhang, Changxin Gao, Jiayu Wang, Xiaoqiang Zhou, Yingya Zhang, Luxin Yan, and Nong Sang. Unianimate: Taming unified video diffusion models for consistent human image animation. arXiv preprint arXiv:2406.01188, 2024. 2 +[54] Xiaofeng Wang, Kang Zhao, Feng Liu, Jiayu Wang, Guosheng Zhao, Xiaoyi Bao, Zheng Zhu, Yingya Zhang, and Xingang Wang. Egovid-5m: A large-scale video-action dataset for egocentric video generation. arXiv preprint arXiv:2411.08380, 2024. 2 +[55] Yuan Wang, Zhao Wang, Junhao Gong, Di Huang, Tong He, Wanli Ouyang, Jile Jiao, Xuetao Feng, Qi Dou, Shixiang Tang, et al. Holistic-motion2d: Scalable whole-body human motion generation in 2d space. arXiv preprint arXiv:2406.11253, 2024. 2, 3, 6 +[56] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, and Dacheng Tao. Gmflow: Learning optical flow via global matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8121-8130, 2022. 4 +[57] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In AAAI, 2018. 2 +[58] Zhendong Yang, Ailing Zeng, Chun Yuan, and Yu Li. Effective whole-body pose estimation with two-stages distillation. In ICCV, 2023. 4 +[59] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 2, 4, 6, 8 +[60] Sihyun Yu, Sangkyung Kwak, Huiwon Jang, Jongheon Jeong, Jonathan Huang, Jinwoo Shin, and Saining Xie. Representation alignment for generation: Training diffusion transformers is easier than you think, 2024. 5 +[61] Yan Zeng, Guoqiang Wei, Jiani Zheng, Jiaxin Zou, Yang Wei, Yuchen Zhang, and Hang Li. Make pixels dance: High-dynamic video generation. arXiv preprint arXiv:2311.10982, 2023. 2 +[62] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Shaoli Huang, Yong Zhang, Hongwei Zhao, Hongtao Lu, and Xi Shen. T2m-gpt: Generating human motion from textual descriptions with discrete representations. In CVPR, 2023. 6 +[63] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Yong Zhang, Hongwei Zhao, Hongtao Lu, Xi Shen, and Ying + +Shan. Generating human motion from textual descriptions with discrete representations. In CVPR, 2023. 2, 3, 6 +[64] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 6 +[65] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001, 2022.3 +[66] Yang Zhang, Jiaxi Gu, Li-Wen Wang, Han Wang, Junqi Cheng, Yuefeng Zhu, and Fangyuan Zou. Mimic motion: High-quality human motion video generation with confidence-aware pose guidance. arXiv preprint arXiv:2406.19680, 2024. 2, 6 +[67] Sijie Zhao, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Muyao Niu, Xiaoyu Li, Wenbo Hu, and Ying Shan. Cv-vae: A compatible video vae for latent generative video models. https://arxiv.org/abs/2405.20279, 2024. 4 +[68] Weiyu Zhao, Liangxiao Hu, and Shengping Zhang. Diffugesture: Generating human gesture from two-person dialogue with diffusion models. In ICMI, 2023. 2, 3 +[69] Shenhao Zhu, Junming Leo Chen, Zuozhuo Dai, Qingkun Su, Yinghui Xu, Xun Cao, Yao Yao, Hao Zhu, and Siyu Zhu. Champ: Controllable and consistent human image animation with 3d parametric guidance. arXiv preprint arXiv:2403.14781, 2024. 2 +[70] Wentao Zhu, Xiaoxuan Ma, Zhaoyang Liu, Libin Liu, Wayne Wu, and Yizhou Wang. Motionbert: A unified perspective on learning human motion representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 8 +[71] Zheng Zhu, Xiaofeng Wang, Wangbo Zhao, Chen Min, Ni-anchen Deng, Min Dou, Yuqi Wang, Botian Shi, Kai Wang, Chi Zhang, et al. Is sora a world simulator? a comprehensive survey on general world models and beyond. arXiv preprint arXiv:2405.03520, 2024. 2 \ No newline at end of file diff --git a/2025/HumanDreamer_ Generating Controllable Human-Motion Videos via Decoupled Generation/images.zip b/2025/HumanDreamer_ Generating Controllable Human-Motion Videos via Decoupled Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9137cfef52206828fbb7bbacdab69ee48f0629df --- /dev/null +++ b/2025/HumanDreamer_ Generating Controllable Human-Motion Videos via Decoupled Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:59d5d90103427502b842460718dc89f16135a7b8323482b5acaff6f7cb9d6630 +size 610273 diff --git a/2025/HumanDreamer_ Generating Controllable Human-Motion Videos via Decoupled Generation/layout.json b/2025/HumanDreamer_ Generating Controllable Human-Motion Videos via Decoupled Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b937eb7e91fad242372b54096dbaddb132031b6f --- /dev/null +++ b/2025/HumanDreamer_ Generating Controllable Human-Motion Videos via Decoupled Generation/layout.json @@ -0,0 +1,10229 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 102, + 103, + 509, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 103, + 509, + 140 + ], + "spans": [ + { + "bbox": [ + 102, + 103, + 509, + 140 + ], + "type": "text", + "content": "HumanDreamer: Generating Controllable Human-Motion Videos via Decoupled Generation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 160, + 580, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 160, + 580, + 190 + ], + "spans": [ + { + "bbox": [ + 61, + 160, + 580, + 190 + ], + "type": "text", + "content": "Boyuan Wang" + }, + { + "bbox": [ + 61, + 160, + 580, + 190 + ], + "type": "inline_equation", + "content": "^{1,2,3*}" + }, + { + "bbox": [ + 61, + 160, + 580, + 190 + ], + "type": "text", + "content": ", Xiaofeng Wang" + }, + { + "bbox": [ + 61, + 160, + 580, + 190 + ], + "type": "inline_equation", + "content": "^{1,2,3*}" + }, + { + "bbox": [ + 61, + 160, + 580, + 190 + ], + "type": "text", + "content": ", Chaojun Ni" + }, + { + "bbox": [ + 61, + 160, + 580, + 190 + ], + "type": "inline_equation", + "content": "^{4,5}" + }, + { + "bbox": [ + 61, + 160, + 580, + 190 + ], + "type": "text", + "content": ", Guosheng Zhao" + }, + { + "bbox": [ + 61, + 160, + 580, + 190 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + }, + { + "bbox": [ + 61, + 160, + 580, + 190 + ], + "type": "text", + "content": ", Zhiqin Yang" + }, + { + "bbox": [ + 61, + 160, + 580, + 190 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 61, + 160, + 580, + 190 + ], + "type": "text", + "content": ", Zheng Zhu" + }, + { + "bbox": [ + 61, + 160, + 580, + 190 + ], + "type": "inline_equation", + "content": "^{4\\dagger}" + }, + { + "bbox": [ + 61, + 160, + 580, + 190 + ], + "type": "text", + "content": ", Muyang Zhang" + }, + { + "bbox": [ + 61, + 160, + 580, + 190 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 61, + 160, + 580, + 190 + ], + "type": "text", + "content": ", Yukun Zhou" + }, + { + "bbox": [ + 61, + 160, + 580, + 190 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 61, + 160, + 580, + 190 + ], + "type": "text", + "content": ", Xinze Chen" + }, + { + "bbox": [ + 61, + 160, + 580, + 190 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 61, + 160, + 580, + 190 + ], + "type": "text", + "content": ", Guan Huang" + }, + { + "bbox": [ + 61, + 160, + 580, + 190 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 61, + 160, + 580, + 190 + ], + "type": "text", + "content": ", Lihong Liu" + }, + { + "bbox": [ + 61, + 160, + 580, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 160, + 580, + 190 + ], + "type": "text", + "content": ", Xingang Wang" + }, + { + "bbox": [ + 61, + 160, + 580, + 190 + ], + "type": "inline_equation", + "content": "^{1,3\\dagger}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 168, + 189, + 474, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 189, + 474, + 203 + ], + "spans": [ + { + "bbox": [ + 168, + 189, + 474, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 168, + 189, + 474, + 203 + ], + "type": "text", + "content": "Institute of Automation, Chinese Academy of Sciences, China" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 114, + 203, + 526, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 203, + 526, + 217 + ], + "spans": [ + { + "bbox": [ + 114, + 203, + 526, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 114, + 203, + 526, + 217 + ], + "type": "text", + "content": "School of Artificial Intelligence, University of Chinese Academy of Sciences, China" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 168, + 217, + 473, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 217, + 473, + 232 + ], + "spans": [ + { + "bbox": [ + 168, + 217, + 473, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 168, + 217, + 473, + 232 + ], + "type": "text", + "content": "Luoyang Institute for Robot and Intelligent Equipment, China" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 113, + 232, + 523, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 232, + 523, + 247 + ], + "spans": [ + { + "bbox": [ + 113, + 232, + 523, + 247 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 113, + 232, + 523, + 247 + ], + "type": "text", + "content": "GigaAI, China " + }, + { + "bbox": [ + 113, + 232, + 523, + 247 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 113, + 232, + 523, + 247 + ], + "type": "text", + "content": "Peking University, China " + }, + { + "bbox": [ + 113, + 232, + 523, + 247 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 113, + 232, + 523, + 247 + ], + "type": "text", + "content": "The Chinese University of Hong Kong" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 97, + 287, + 111, + 302 + ], + "blocks": [ + { + "bbox": [ + 97, + 287, + 111, + 302 + ], + "lines": [ + { + "bbox": [ + 97, + 287, + 111, + 302 + ], + "spans": [ + { + "bbox": [ + 97, + 287, + 111, + 302 + ], + "type": "image", + "image_path": "4d9fb75c2e6b73f5f1aeca537ac865c23d69eefa743b4ccb77d156339745e9de.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 289, + 138, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 289, + 138, + 300 + ], + "spans": [ + { + "bbox": [ + 115, + 289, + 138, + 300 + ], + "type": "text", + "content": "Text" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 160, + 277, + 343, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 277, + 343, + 295 + ], + "spans": [ + { + "bbox": [ + 160, + 277, + 343, + 295 + ], + "type": "text", + "content": "A man is performing a yoga pose on a mat, and he is seen moving his legs and arms in different positions." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 160, + 295, + 343, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 295, + 343, + 316 + ], + "spans": [ + { + "bbox": [ + 160, + 295, + 343, + 316 + ], + "type": "text", + "content": "A man is speaking to the camera and then looking off into the distance. Later, he is speaking to the camera again." + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 352, + 288, + 390, + 304 + ], + "blocks": [ + { + "bbox": [ + 352, + 288, + 390, + 304 + ], + "lines": [ + { + "bbox": [ + 352, + 288, + 390, + 304 + ], + "spans": [ + { + "bbox": [ + 352, + 288, + 390, + 304 + ], + "type": "image", + "image_path": "1854e0342367af45949e612f9934120c5176b22a34f9ef0c1da950452cd83400.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 394, + 290, + 477, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 290, + 477, + 302 + ], + "spans": [ + { + "bbox": [ + 394, + 290, + 477, + 302 + ], + "type": "text", + "content": "HumanDreamer" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 91, + 330, + 161, + 384 + ], + "blocks": [ + { + "bbox": [ + 91, + 330, + 161, + 384 + ], + "lines": [ + { + "bbox": [ + 91, + 330, + 161, + 384 + ], + "spans": [ + { + "bbox": [ + 91, + 330, + 161, + 384 + ], + "type": "image", + "image_path": "4dbab8c84cbcdf186b3e3fdcf022ab2cfdbab02333e726ef38f2707e48ad6a4d.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 162, + 330, + 232, + 384 + ], + "blocks": [ + { + "bbox": [ + 162, + 330, + 232, + 384 + ], + "lines": [ + { + "bbox": [ + 162, + 330, + 232, + 384 + ], + "spans": [ + { + "bbox": [ + 162, + 330, + 232, + 384 + ], + "type": "image", + "image_path": "e8f827329573956ee15bab64424ad7bc7c6f34f1f11dfe4b663ce163e37243e9.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 233, + 330, + 303, + 384 + ], + "blocks": [ + { + "bbox": [ + 242, + 319, + 362, + 330 + ], + "lines": [ + { + "bbox": [ + 242, + 319, + 362, + 330 + ], + "spans": [ + { + "bbox": [ + 242, + 319, + 362, + 330 + ], + "type": "text", + "content": "Text-to-Pose Generation" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 233, + 330, + 303, + 384 + ], + "lines": [ + { + "bbox": [ + 233, + 330, + 303, + 384 + ], + "spans": [ + { + "bbox": [ + 233, + 330, + 303, + 384 + ], + "type": "image", + "image_path": "3f8ca9494a8d2e8e435af44a0f2df93b6b51b43364b09e682e3072412ab75e3a.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 304, + 330, + 375, + 384 + ], + "blocks": [ + { + "bbox": [ + 304, + 330, + 375, + 384 + ], + "lines": [ + { + "bbox": [ + 304, + 330, + 375, + 384 + ], + "spans": [ + { + "bbox": [ + 304, + 330, + 375, + 384 + ], + "type": "image", + "image_path": "1e94a5cf8ef8bd75b2ad9c2661ba34fdc8921647fa25899131cf3cf80e56de86.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 375, + 330, + 444, + 384 + ], + "blocks": [ + { + "bbox": [ + 375, + 330, + 444, + 384 + ], + "lines": [ + { + "bbox": [ + 375, + 330, + 444, + 384 + ], + "spans": [ + { + "bbox": [ + 375, + 330, + 444, + 384 + ], + "type": "image", + "image_path": "77a51d7e752bd3ae932752d0e934f9c1e271d1eeecffea74a16638c168a7cdb1.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 445, + 330, + 516, + 384 + ], + "blocks": [ + { + "bbox": [ + 445, + 330, + 516, + 384 + ], + "lines": [ + { + "bbox": [ + 445, + 330, + 516, + 384 + ], + "spans": [ + { + "bbox": [ + 445, + 330, + 516, + 384 + ], + "type": "image", + "image_path": "9c7b5e002ac917dc1cfc9102ea658206347663dee52a7e2f6edc83ee7451d53b.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 90, + 389, + 151, + 460 + ], + "blocks": [ + { + "bbox": [ + 90, + 389, + 151, + 460 + ], + "lines": [ + { + "bbox": [ + 90, + 389, + 151, + 460 + ], + "spans": [ + { + "bbox": [ + 90, + 389, + 151, + 460 + ], + "type": "image", + "image_path": "bbfc19f914967ff7e30916898b4b3ee811b9cfd6da0732ca59febdb7bea83fd7.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 152, + 389, + 211, + 460 + ], + "blocks": [ + { + "bbox": [ + 152, + 389, + 211, + 460 + ], + "lines": [ + { + "bbox": [ + 152, + 389, + 211, + 460 + ], + "spans": [ + { + "bbox": [ + 152, + 389, + 211, + 460 + ], + "type": "image", + "image_path": "b81e79a4b9f71ee5e97d63f32f3c11fc55203e3111f583c3e9351eb2c185e3e4.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 212, + 389, + 272, + 460 + ], + "blocks": [ + { + "bbox": [ + 212, + 389, + 272, + 460 + ], + "lines": [ + { + "bbox": [ + 212, + 389, + 272, + 460 + ], + "spans": [ + { + "bbox": [ + 212, + 389, + 272, + 460 + ], + "type": "image", + "image_path": "7dd132b0bb811a968d662444727540635d190c1d7e21b54d0d37c1f7a80c3ffd.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 273, + 389, + 333, + 460 + ], + "blocks": [ + { + "bbox": [ + 273, + 389, + 333, + 460 + ], + "lines": [ + { + "bbox": [ + 273, + 389, + 333, + 460 + ], + "spans": [ + { + "bbox": [ + 273, + 389, + 333, + 460 + ], + "type": "image", + "image_path": "1e973707081fd3653105c4b1e44468297d29da928eaac73f3a74a2dcd102bdeb.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 334, + 389, + 394, + 460 + ], + "blocks": [ + { + "bbox": [ + 334, + 389, + 394, + 460 + ], + "lines": [ + { + "bbox": [ + 334, + 389, + 394, + 460 + ], + "spans": [ + { + "bbox": [ + 334, + 389, + 394, + 460 + ], + "type": "image", + "image_path": "a34c8b76e8c9131652bcc889daf840037bc469e98f83e00b679696eb609ab49e.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 394, + 389, + 455, + 460 + ], + "blocks": [ + { + "bbox": [ + 394, + 389, + 455, + 460 + ], + "lines": [ + { + "bbox": [ + 394, + 389, + 455, + 460 + ], + "spans": [ + { + "bbox": [ + 394, + 389, + 455, + 460 + ], + "type": "image", + "image_path": "0533623de33a056f233d3a583df12fc7c5c10ef586237acddae55e348e69b62d.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 455, + 389, + 516, + 460 + ], + "blocks": [ + { + "bbox": [ + 455, + 389, + 516, + 460 + ], + "lines": [ + { + "bbox": [ + 455, + 389, + 516, + 460 + ], + "spans": [ + { + "bbox": [ + 455, + 389, + 516, + 460 + ], + "type": "image", + "image_path": "daa34c4e486e3ea4a975e5efae0a064395f7aed52bd0cf482497f8933fb9ae6d.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 91, + 479, + 161, + 533 + ], + "blocks": [ + { + "bbox": [ + 91, + 479, + 161, + 533 + ], + "lines": [ + { + "bbox": [ + 91, + 479, + 161, + 533 + ], + "spans": [ + { + "bbox": [ + 91, + 479, + 161, + 533 + ], + "type": "image", + "image_path": "e2f2f9cd04c43aae3b98fd31ee51a77c167696d04fa73e9db9102a72558d0f33.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 162, + 479, + 232, + 533 + ], + "blocks": [ + { + "bbox": [ + 162, + 479, + 232, + 533 + ], + "lines": [ + { + "bbox": [ + 162, + 479, + 232, + 533 + ], + "spans": [ + { + "bbox": [ + 162, + 479, + 232, + 533 + ], + "type": "image", + "image_path": "e27aa5ad574bd389e80ad374abcec69fb9322e4291872c0dadc62657f668ddf2.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 233, + 479, + 303, + 533 + ], + "blocks": [ + { + "bbox": [ + 239, + 468, + 367, + 479 + ], + "lines": [ + { + "bbox": [ + 239, + 468, + 367, + 479 + ], + "spans": [ + { + "bbox": [ + 239, + 468, + 367, + 479 + ], + "type": "text", + "content": "Pose-to-Video Generation" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 233, + 479, + 303, + 533 + ], + "lines": [ + { + "bbox": [ + 233, + 479, + 303, + 533 + ], + "spans": [ + { + "bbox": [ + 233, + 479, + 303, + 533 + ], + "type": "image", + "image_path": "1bab931afc1bde3e962f5d768f56ef3f5a2d7adbbb4c109d05c082412097e3a9.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 304, + 479, + 375, + 533 + ], + "blocks": [ + { + "bbox": [ + 304, + 479, + 375, + 533 + ], + "lines": [ + { + "bbox": [ + 304, + 479, + 375, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 479, + 375, + 533 + ], + "type": "image", + "image_path": "fea104d67a494498303501440134ab397c9eb1f9675169296cbbd13c4ef46712.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 375, + 479, + 444, + 533 + ], + "blocks": [ + { + "bbox": [ + 375, + 479, + 444, + 533 + ], + "lines": [ + { + "bbox": [ + 375, + 479, + 444, + 533 + ], + "spans": [ + { + "bbox": [ + 375, + 479, + 444, + 533 + ], + "type": "image", + "image_path": "e70b871a72fb3f17152b178008bc2601392060f6a4328678ac72df32b81cdc61.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 445, + 479, + 516, + 533 + ], + "blocks": [ + { + "bbox": [ + 445, + 479, + 516, + 533 + ], + "lines": [ + { + "bbox": [ + 445, + 479, + 516, + 533 + ], + "spans": [ + { + "bbox": [ + 445, + 479, + 516, + 533 + ], + "type": "image", + "image_path": "235df14e4cc9e27bf1fd654c6e21af1c1b1d7fa5afdb3e00cc6a1437e8843926.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 91, + 536, + 153, + 609 + ], + "blocks": [ + { + "bbox": [ + 91, + 536, + 153, + 609 + ], + "lines": [ + { + "bbox": [ + 91, + 536, + 153, + 609 + ], + "spans": [ + { + "bbox": [ + 91, + 536, + 153, + 609 + ], + "type": "image", + "image_path": "d7ba48354bf5fcd85a86e32ad98c15ff25e7dc99e929dd3ef6aa607dec419475.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 625, + 555, + 649 + ], + "lines": [ + { + "bbox": [ + 55, + 625, + 555, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 625, + 555, + 649 + ], + "type": "text", + "content": "Figure 1. Illustration of HumanDreamer. The human-motion video generation is decoupled into two steps: Text-to-Pose generation and Pose-to-Video generation. The decoupled process integrates the flexibility of text control and the controllability of pose guidance." + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_caption" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 154, + 536, + 213, + 609 + ], + "blocks": [ + { + "bbox": [ + 154, + 536, + 213, + 609 + ], + "lines": [ + { + "bbox": [ + 154, + 536, + 213, + 609 + ], + "spans": [ + { + "bbox": [ + 154, + 536, + 213, + 609 + ], + "type": "image", + "image_path": "7ad538ec01880978193cbdb436ba3dba39d3f443523d990691fedf6052e754fa.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 214, + 536, + 274, + 609 + ], + "blocks": [ + { + "bbox": [ + 214, + 536, + 274, + 609 + ], + "lines": [ + { + "bbox": [ + 214, + 536, + 274, + 609 + ], + "spans": [ + { + "bbox": [ + 214, + 536, + 274, + 609 + ], + "type": "image", + "image_path": "46ff04edc0bab07374dea19d2fdfd72686944221f216e398e5dcaf146389d08a.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 275, + 536, + 335, + 609 + ], + "blocks": [ + { + "bbox": [ + 275, + 536, + 335, + 609 + ], + "lines": [ + { + "bbox": [ + 275, + 536, + 335, + 609 + ], + "spans": [ + { + "bbox": [ + 275, + 536, + 335, + 609 + ], + "type": "image", + "image_path": "07f5ce86b40c3105d2cc62afb95a08c7c7f214aba3271a3c7cea434c4315b78c.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 335, + 536, + 395, + 609 + ], + "blocks": [ + { + "bbox": [ + 335, + 536, + 395, + 609 + ], + "lines": [ + { + "bbox": [ + 335, + 536, + 395, + 609 + ], + "spans": [ + { + "bbox": [ + 335, + 536, + 395, + 609 + ], + "type": "image", + "image_path": "7815a721cd409d9f706e6362a7f174275509cd0e45c9b9b5272119058240b1bd.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 396, + 536, + 456, + 609 + ], + "blocks": [ + { + "bbox": [ + 396, + 536, + 456, + 609 + ], + "lines": [ + { + "bbox": [ + 396, + 536, + 456, + 609 + ], + "spans": [ + { + "bbox": [ + 396, + 536, + 456, + 609 + ], + "type": "image", + "image_path": "41ec9997937ddcb56dbb189f33ab8f6938b27efc9a3201cd7fa3983a5543a29f.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 457, + 536, + 517, + 609 + ], + "blocks": [ + { + "bbox": [ + 457, + 536, + 517, + 609 + ], + "lines": [ + { + "bbox": [ + 457, + 536, + 517, + 609 + ], + "spans": [ + { + "bbox": [ + 457, + 536, + 517, + 609 + ], + "type": "image", + "image_path": "4e999f3cd01a9303c07fe61f8c505da6d7a4b18c3278fa87d11620f5a9d3ab2e.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + } + ], + "index": 43 + }, + { + "bbox": [ + 410, + 658, + 459, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 410, + 658, + 459, + 671 + ], + "spans": [ + { + "bbox": [ + 410, + 658, + 459, + 671 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 45 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 16 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 16 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 16 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 198, + 13, + 443, + 25 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 13, + 443, + 25 + ], + "spans": [ + { + "bbox": [ + 198, + 13, + 443, + 25 + ], + "type": "text", + "content": "Except for this watermark, it is identical to the accepted version;" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "spans": [ + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "type": "text", + "content": "the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 683, + 223, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 683, + 223, + 693 + ], + "spans": [ + { + "bbox": [ + 66, + 683, + 223, + 693 + ], + "type": "text", + "content": "*These authors contributed equally to this work." + } + ] + } + ], + "index": 46 + }, + { + "bbox": [ + 67, + 693, + 295, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 693, + 295, + 703 + ], + "spans": [ + { + "bbox": [ + 67, + 693, + 295, + 703 + ], + "type": "text", + "content": "† Corresponding authors. zhengzhu@ieee.org, xingang.wang@ia.ac.cn." + } + ] + } + ], + "index": 47 + }, + { + "bbox": [ + 67, + 703, + 265, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 265, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 265, + 712 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 67, + 703, + 265, + 712 + ], + "type": "text", + "content": "Project Page: https://humandreamer.github.io/" + } + ] + } + ], + "index": 48 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Human-motion video generation has been a challenging task, primarily due to the difficulty inherent in learning hu" + } + ] + } + ], + "index": 50 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12391" + } + ] + } + ], + "index": 51 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 72, + 294, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 72, + 294, + 335 + ], + "spans": [ + { + "bbox": [ + 57, + 72, + 294, + 335 + ], + "type": "text", + "content": "man body movements. While some approaches have attempted to drive human-centric video generation explicitly through pose control, these methods typically rely on poses derived from existing videos, thereby lacking flexibility. To address this, we propose HumanDreamer, a decoupled human video generation framework that first generates diverse poses from text prompts and then leverages these poses to generate human-motion videos. Specifically, we propose MotionVid, the largest dataset for human-motion pose generation. Based on the dataset, we present MotionDiT, which is trained to generate structured human-motion poses from text prompts. Besides, a novel LAMA loss is introduced, which together contribute to a significant improvement in FID by " + }, + { + "bbox": [ + 57, + 72, + 294, + 335 + ], + "type": "inline_equation", + "content": "62.4\\%" + }, + { + "bbox": [ + 57, + 72, + 294, + 335 + ], + "type": "text", + "content": ", along with respective enhancements in R-precision for top1, top2, and top3 by " + }, + { + "bbox": [ + 57, + 72, + 294, + 335 + ], + "type": "inline_equation", + "content": "41.8\\%" + }, + { + "bbox": [ + 57, + 72, + 294, + 335 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 57, + 72, + 294, + 335 + ], + "type": "inline_equation", + "content": "26.3\\%" + }, + { + "bbox": [ + 57, + 72, + 294, + 335 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 57, + 72, + 294, + 335 + ], + "type": "inline_equation", + "content": "18.3\\%" + }, + { + "bbox": [ + 57, + 72, + 294, + 335 + ], + "type": "text", + "content": ", thereby advancing both the Text-to-Pose control accuracy and FID metrics. Our experiments across various Pose-to-Video baselines demonstrate that the poses generated by our method can produce diverse and high-quality human-motion videos. Furthermore, our model can facilitate other downstream tasks, such as pose sequence prediction and 2D-3D motion lifting." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 58, + 358, + 135, + 370 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 358, + 135, + 370 + ], + "spans": [ + { + "bbox": [ + 58, + 358, + 135, + 370 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 379, + 294, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 379, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 57, + 379, + 294, + 521 + ], + "type": "text", + "content": "Generating human-motion videos remains a particularly challenging task due to the inherent complexity of modeling human body movements. Despite current advancements in generative modeling [9, 13, 31, 48, 71], state-of-the-art video generation models [3, 16, 19, 36, 59, 61], equipped with billions of parameters and trained on millions of video and image data, still struggle to capture human body movements, frequently resulting in fragmented or unrealistic portrayals. This limitation is exacerbated when controlling human videos via textual conditions, highlighting the fundamental challenge of directly mapping text prompts to human visual data." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 522, + 294, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 522, + 294, + 629 + ], + "spans": [ + { + "bbox": [ + 57, + 522, + 294, + 629 + ], + "type": "text", + "content": "To enhance the generation quality, approaches such asAnimate Anyone [20], UniAnimate [53], Mimic Motion [66], Champ [69] andAnimate-X [49] explicitly generate human-motion videos through pose control. By utilizing the pose-guided generation, these methods effectively reduce the complexities associated with human-motion videos. However, a notable limitation of this approach is its reliance on human-motion poses derived from existing videos, which restricts the flexibility." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 630, + 294, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 630, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 630, + 294, + 712 + ], + "type": "text", + "content": "Therefore, we propose HumanDreamer, a decoupled human-motion video generation framework that first generates human-motion poses from text prompts and subsequently produces human-motion videos based on the generated poses. The motivation behind this decoupled approach lies in that Text-to-Pose presents a more manageable search space compared to the direct learning of text-to-pixel repre" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 316, + 72, + 553, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 418 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 418 + ], + "type": "text", + "content": "sensation. As shown in Fig. 1, this decoupled framework facilitates a more effective generation of human movements. Additionally, the proposed HumanDreamer utilizes text as the input, offering greater flexibility than [20, 34, 49, 53, 66] that directly relies on pre-defined poses. Specifically, we construct a 1.2 million text-pose pairs dataset MotionVid, which is the largest dataset for human-motion pose generation. To ensure dataset quality, a comprehensive data cleaning pipeline is introduced, where the cleaning factors involve body movement amplitude, human presence duration, facial visibility, and the proportion of human figures. This rigorous cleaning guarantees that the dataset is reliable for training Text-to-Pose tasks. Based on the dataset, we propose the MotionDiT to generate structured human-motion poses from text prompts. The MotionDiT integrates a global attention block designed to extract the global characteristics of the entire pose sequence, alongside a local feature aggregation mechanism that captures information from adjacent pose points, effectively combining global and local perspectives to enhance the quality and coherence of the generated poses. Additionally, the LAMA loss is utilized in the latent space to align the semantic features of the ground truth motion with those produced by MotionDiT, which not only improves the interpretability of the model but also boosts its overall performance. The introduced techniques collectively result in a notable " + }, + { + "bbox": [ + 316, + 72, + 553, + 418 + ], + "type": "inline_equation", + "content": "62.4\\%" + }, + { + "bbox": [ + 316, + 72, + 553, + 418 + ], + "type": "text", + "content": " enhancement in FID, accompanied by respective increases in R-precision for the top1, top2, and top3 categories by " + }, + { + "bbox": [ + 316, + 72, + 553, + 418 + ], + "type": "inline_equation", + "content": "41.8\\%" + }, + { + "bbox": [ + 316, + 72, + 553, + 418 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 316, + 72, + 553, + 418 + ], + "type": "inline_equation", + "content": "26.3\\%" + }, + { + "bbox": [ + 316, + 72, + 553, + 418 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 316, + 72, + 553, + 418 + ], + "type": "inline_equation", + "content": "18.3\\%" + }, + { + "bbox": [ + 316, + 72, + 553, + 418 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 316, + 420, + 553, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 420, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 316, + 420, + 553, + 623 + ], + "type": "text", + "content": "The primary contributions of this work are as follows: (1) We present HumanDreamer, the first decoupled framework for human-motion video generation, which integrates the flexibility of text control with the controllability of pose guidance. (2) We propose the MotionVid, the largest dataset for human-motion pose generation. In MotionVid, we conduct a comprehensive data annotation and data cleaning to ensure data reliability. (3) The MotionDiT is introduced to generate diverse human-motion poses under text control. To enhance the generation process, we propose the LAMA loss, which significantly improves pose fidelity and diversity. (4) Through extensive experiments, we show that MotionDiT and the LAMA loss improve control accuracy and FID by " + }, + { + "bbox": [ + 316, + 420, + 553, + 623 + ], + "type": "inline_equation", + "content": "62.4\\%" + }, + { + "bbox": [ + 316, + 420, + 553, + 623 + ], + "type": "text", + "content": ", with top-k R-precision gains of " + }, + { + "bbox": [ + 316, + 420, + 553, + 623 + ], + "type": "inline_equation", + "content": "41.8\\%" + }, + { + "bbox": [ + 316, + 420, + 553, + 623 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 316, + 420, + 553, + 623 + ], + "type": "inline_equation", + "content": "26.3\\%" + }, + { + "bbox": [ + 316, + 420, + 553, + 623 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 316, + 420, + 553, + 623 + ], + "type": "inline_equation", + "content": "18.3\\%" + }, + { + "bbox": [ + 316, + 420, + 553, + 623 + ], + "type": "text", + "content": " for top-1, top-2, and top-3. The generated poses also support downstream tasks like pose sequence prediction and 2D-to-3D motion enhancement." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 316, + 638, + 400, + 650 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 638, + 400, + 650 + ], + "spans": [ + { + "bbox": [ + 316, + 638, + 400, + 650 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 316, + 658, + 474, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 474, + 670 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 474, + 670 + ], + "type": "text", + "content": "2.1. Human-motion Pose Datasets" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 316, + 677, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 677, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 677, + 553, + 712 + ], + "type": "text", + "content": "Human motion datasets are essential for the development of human-centric perception [7, 8, 21, 30, 44, 57] or generation [6, 29, 32, 51, 52, 54, 55, 63, 68] tasks. Specifically," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12392" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 550, + 166 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 550, + 166 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 550, + 166 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 550, + 166 + ], + "type": "image", + "image_path": "22a1e483c3189286dfed3e85906f8d2d67c0438971256fb3eb066be9130f0993.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 173, + 555, + 208 + ], + "lines": [ + { + "bbox": [ + 55, + 173, + 555, + 208 + ], + "spans": [ + { + "bbox": [ + 55, + 173, + 555, + 208 + ], + "type": "text", + "content": "Figure 2. The data cleaning and annotation pipeline for MotionVid begins with raw data sourced from public datasets and the internet, which is then segmented into video clips. To ensure high-quality data, we apply video quality filter, data annotation, human quality filter, and caption quality filter." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 228, + 296, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 296, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 296, + 443 + ], + "type": "text", + "content": "the KIT Motion-Language Dataset [39] provides sequence-level descriptions to support multi-modal motion tasks. In contrast, the HumanML3D [15] dataset, which is built upon the AMASS [35] and HumanAct12 [14] datasets, offers richer textual annotations and a wider variety of activities. Additionally, Motion-X [29] establishes a robust annotation protocol and is the first to compile a comprehensive fine-grained 3D whole-body motion dataset derived from extensive scene captures. As a result, existing text-driven 3D motion datasets often lack sufficient volume and diversity, which limits their scalability. A dataset that shares similarities with ours is Holistic-Motion2D [55], which features 1M text-pose pairs but only 16K of these are publicly accessible. However, the proposed MotionVid significantly surpasses this with a larger and more diverse set of 1.2M text-pose data, coupled with a more thorough data cleaning process that ensures the dataset's reliability for training Text-to-Pose tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 464, + 273, + 476 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 464, + 273, + 476 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 273, + 476 + ], + "type": "text", + "content": "2.2. Human-motion Pose Generation Methods" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 486, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 296, + 714 + ], + "type": "text", + "content": "Text-driven human motion generation, which converts textual descriptions into motion sequences, has gained significant attention in recent years. Early approaches focus on aligning motion and text through shared latent spaces. For instance, MotionCLIP [51] enhances autoencoder generalization by aligning the latent space with the expressive CLIP [40] embedding. T2M-GPT [63] frames the task as predicting discrete motion indices. More recently, Diffusion-based models, including MDM [52], MotionDiffuse [65], and DiffGesture [68], have set new performance standards. MLD [6] introduces a latent Diffusion model that generates diverse, realistic motions, while Humantomato [32] expands the task to whole-body motion generation, advancing GPT-like models for finer motion sequences. However, limited 3D data hampers model diversity and detail. Tender [55] overcomes this by training on extensive 2D pose data. In contrast, the proposed MotionDiT employs a Diffusion Transformer to better capture both local and global pose relationships, improving human-motion pose generation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 227, + 414, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 227, + 414, + 239 + ], + "spans": [ + { + "bbox": [ + 314, + 227, + 414, + 239 + ], + "type": "text", + "content": "3. HumanDreamer" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 247, + 555, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 555, + 295 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 555, + 295 + ], + "type": "text", + "content": "In this section, we first introduce the proposed MotionVid dataset that facilitates the training of human-motion pose generation. Next, we elaborate on the details of the decoupled Text-to-Pose generation and Pose-to-Video generation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 301, + 388, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 301, + 388, + 312 + ], + "spans": [ + { + "bbox": [ + 314, + 301, + 388, + 312 + ], + "type": "text", + "content": "3.1. MotionVid" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 319, + 555, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 555, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 555, + 545 + ], + "type": "text", + "content": "Existing datasets face limitations in training Text-to-Pose tasks. For instance, current 3D datasets [14, 35] primarily capture torso movements, lacking essential full-body keypoints, particularly for the face and hands, which makes them unsuitable for diverse pose generation. Although the Motion-X dataset [29] provides full-body 3D data, its scale is limited to only 81.1K samples, as the expensive annotation of 3D data. For the 2D full-body dataset, the Holistic-Motion2D dataset [55] includes approximately 1M samples, though only 16K of these are publicly accessible. Therefore, there is a pressing need for a comprehensive and publicly available human-motion pose dataset that facilitates the training of Text-to-Pose tasks. To address these limitations, we introduce the largest human-motion pose dataset MotionVid, which contains approximately 1.2M data pairs, with videos of varying resolutions and durations exceeding 64 frames. Besides, we provide a robust data collection pipeline to obtain high-quality textpose pairs from diverse and complex data sources." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 546, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 556, + 714 + ], + "type": "text", + "content": "Data Source. We construct a diverse set of text-pose pairs by curating data from various publicly available video datasets, including action recognition datasets Kinetics-400 [23], Kinetics-700 [4], ActivityNet-200 [11], and Something-Something V2 [12], as well as human-related datasets such as CAER [26], DFEW [22], UBody [28], Charades [45], Charades-Ego [46], HAA500 [10], and HMDB51 [25]. This selection totals " + }, + { + "bbox": [ + 313, + 546, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\sim 6.6\\mathrm{M}" + }, + { + "bbox": [ + 313, + 546, + 556, + 714 + ], + "type": "text", + "content": " samples. To further enhance scene and subject diversity, we add " + }, + { + "bbox": [ + 313, + 546, + 556, + 714 + ], + "type": "inline_equation", + "content": "3.4\\mathrm{M}" + }, + { + "bbox": [ + 313, + 546, + 556, + 714 + ], + "type": "text", + "content": " videos sourced from the internet. Given the complexity of these videos, we perform extensive annotation and cleaning to ensure high-quality data. As shown in Fig. 2, we first use a shot transition detection model [47] to segment the video clips, followed by the video quality filter, data annotation," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12393" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 72, + 550, + 228 + ], + "blocks": [ + { + "bbox": [ + 58, + 72, + 550, + 228 + ], + "lines": [ + { + "bbox": [ + 58, + 72, + 550, + 228 + ], + "spans": [ + { + "bbox": [ + 58, + 72, + 550, + 228 + ], + "type": "image", + "image_path": "f8b8a27cb1d50360cbea2f8058c48f9ffcd609658e0fc405136e29b822d1abdc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 238, + 555, + 273 + ], + "lines": [ + { + "bbox": [ + 55, + 238, + 555, + 273 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 555, + 273 + ], + "type": "text", + "content": "Figure 3. Training pipeline of the proposed Text-to-Pose generation. Pose data are encoded in latent space via the Pose VAE, which are then processed by the proposed MotionDiT, where local feature aggregation and global attention are utilized to capture information from the entire pose sequence. Finally, the LAMA loss is calculated via the proposed CLoP, which enhances the training of MotionDiT." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 293, + 241, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 293, + 241, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 241, + 304 + ], + "type": "text", + "content": "human quality filter, and caption quality filter." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 305, + 295, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 305, + 295, + 484 + ], + "spans": [ + { + "bbox": [ + 54, + 305, + 295, + 484 + ], + "type": "text", + "content": "Video Quality Filter. After acquiring a large number of videos, it is necessary to verify the video quality. Drawing inspiration from SVD's data cleaning strategy [3], we propose a Video Quality Filter to select high-quality video clips. Specifically, we use the GMFlow method [56] to estimate optical flow and filter out videos with insufficient movement intensity. Subsequently, we employ the method described in [1] to detect text regions and remove videos where text occupies a significant portion of the frame. We also utilize LAION-AI's aesthetic predictor [41] to compute aesthetic scores and eliminate videos with low aesthetic quality. Finally, the Laplacian operator [2] is applied to assess blur intensity, and videos deemed excessively blurry are discarded. As a result, approximately " + }, + { + "bbox": [ + 54, + 305, + 295, + 484 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 54, + 305, + 295, + 484 + ], + "type": "text", + "content": " of the data is filtered out (see supplementary materials for more details)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 485, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 485, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 485, + 295, + 616 + ], + "type": "text", + "content": "Data Annotation. We first prompt the Vision Language Model (VLM) [5] to generate action-oriented captions that describe human activities. Besides, for each frame, we extract 2D poses using the DWPose model [58]. This process yields a large number of text-pose pairs. However, the annotated data pairs are unsuitable for training Text-to-Pose tasks. This is due to several factors: the VLM's potential inaccuracy in describing multi-person actions, difficulty in recognizing poses occupying smaller areas in the frame, bias toward static poses from individuals with minimal motion, and limited detail in poses where faces are not visible." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": "Human Quality Filter. To further refine the dataset, we implement a Human Quality Filter, focusing on selecting data based on human-related characteristics. We calculate the difference between two consecutive frames' 2D poses to filter out sequences with insufficient motion magnitude. Additionally, we compute the ratio of the human detection bounding box to the entire frame to remove clips where the human coverage is too small. At this stage, we focus on" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 293, + 553, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 553, + 340 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 553, + 340 + ], + "type": "text", + "content": "simpler scenarios by selecting scenes with a human count of one and ensuring face visibility for training. Based on the Human Quality Filter, approximately " + }, + { + "bbox": [ + 313, + 293, + 553, + 340 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 313, + 293, + 553, + 340 + ], + "type": "text", + "content": " of the data is filtered out (see supplementary materials for more details)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 341, + 554, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 341, + 554, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 341, + 554, + 437 + ], + "type": "text", + "content": "Caption Quality Filter. To enhance the quality of caption annotations, we apply a Caption Quality Filter to refine alignment accuracy. Specifically, using the proposed CLoP (see Sec. 3.2 for details), we compute the semantic similarity between text and poses, filtering data based on the similarity score. This ensures that each text annotation semantically aligns with the corresponding 2D pose data, thereby improving overall data quality." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 445, + 394, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 394, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 394, + 456 + ], + "type": "text", + "content": "3.2. Text-to-Pose" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 461, + 554, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 554, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 554, + 581 + ], + "type": "text", + "content": "Fig. 3 illustrates the training pipeline of the proposed Text-to-Pose generation. In this pipeline, pose data are first encoded into a latent space via the Pose VAE and subsequently processed by the proposed MotionDiT model. Within MotionDiT, local feature aggregation and global attention mechanisms are employed to capture comprehensive information across the entire pose sequence. Finally, LAMA loss is computed using the proposed CLoP, which further refines the training of MotionDiT. In the next, we introduce the Pose VAE, MotionDiT and CLoP." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "text", + "content": "Pose VAE. Variational Autoencoders (VAEs) [17, 24, 59, 67] have been widely applied in 2D image and 3D video processing, but their use for pose data remains underexplored, presenting a rich area for research. We represent a pose sequence as " + }, + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{p} \\in \\mathbb{R}^{f \\times N \\times 3}" + }, + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "text", + "content": " is the frame count, " + }, + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "text", + "content": " the keypoints, and dimension 3 includes the " + }, + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "text", + "content": " coordinates and confidence scores. Given the sequential structure and frame-to-frame correlations, a 1D convolutional encoder with multi-layer downsampling effectively extracts temporal features, while symmetric decoder-encoder architecture with skip connections aids reconstruction. Confi" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12394" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "dence scores of each keypoint help mitigate occlusion issues, optimizing the model via KL divergence and reconstruction loss. Further details are available in the supplementary materials." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 120, + 295, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 120, + 295, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 120, + 295, + 240 + ], + "type": "text", + "content": "MotionDiT. The proposed MotionDiT extends the Diffusion Transformer (DiT) [37] architecture specifically tailored to construct associations between human-motion poses and text control. In the proposed MotionDiT, we incorporate a global attention block to capture global spatial-temporal patterns inherent in human poses. Besides, the MotionDiT involves a local feature aggregation module to strengthen correlations between adjacent joints. Specifically, we first obtain the latent representation from the Pose VAE, which is then processed via a patch-embedding:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 135, + 248, + 295, + 262 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 248, + 295, + 262 + ], + "spans": [ + { + "bbox": [ + 135, + 248, + 295, + 262 + ], + "type": "interline_equation", + "content": "l _ {p} = \\mathcal {F} _ {\\text {P a t c h}} (\\mathcal {E} (\\mathbf {p})), \\tag {1}", + "image_path": "1d409c385b53c41f140be6c811376357ec04fc891d07441049da60f49ea5dc51.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 270, + 296, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 296, + 330 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 296, + 330 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 270, + 296, + 330 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{Patch}}(\\cdot)" + }, + { + "bbox": [ + 55, + 270, + 296, + 330 + ], + "type": "text", + "content": " is the patch embedding operation [33], and " + }, + { + "bbox": [ + 55, + 270, + 296, + 330 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(\\cdot)" + }, + { + "bbox": [ + 55, + 270, + 296, + 330 + ], + "type": "text", + "content": " is the Pose VAE encoder. Then we employ Diffusion blocks to process " + }, + { + "bbox": [ + 55, + 270, + 296, + 330 + ], + "type": "inline_equation", + "content": "l_{p}" + }, + { + "bbox": [ + 55, + 270, + 296, + 330 + ], + "type": "text", + "content": ". As shown in Fig. 3, in each block, the local feature aggregation module is first utilized to strengthen local feature correlation:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 125, + 338, + 295, + 354 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 338, + 295, + 354 + ], + "spans": [ + { + "bbox": [ + 125, + 338, + 295, + 354 + ], + "type": "interline_equation", + "content": "l _ {p} = \\mathcal {F} _ {\\mathrm {s a}} \\left(\\mathcal {F} _ {\\mathrm {r e s}} \\left(l _ {p}\\right) + l _ {p}\\right), \\tag {2}", + "image_path": "7126574ae8490137f59e29d6c471de5192142033304f4217dbbc5ed6958f57c7.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 360, + 295, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 360, + 295, + 610 + ], + "spans": [ + { + "bbox": [ + 55, + 360, + 295, + 610 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 360, + 295, + 610 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{res}}" + }, + { + "bbox": [ + 55, + 360, + 295, + 610 + ], + "type": "text", + "content": " is the 1D ResNet Block with kernel size 3, and " + }, + { + "bbox": [ + 55, + 360, + 295, + 610 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{sa}}" + }, + { + "bbox": [ + 55, + 360, + 295, + 610 + ], + "type": "text", + "content": " is the spatial self attention. For text-based control, textual features obtained from CLIP [40] are then integrated into the network through cross-attention. Besides, following conventional video Diffusion architecture [3, 33], temporal attention is used to ensure continuity of pose features. Notably, it is essential for the model to capture global pose feature information. Although the model already employs spatiotemporal attention mechanisms, certain pose characteristics at one location and time may influence other locations at different times. Thus, global attention is required to extract these global features among internal pose keypoints. Instead of incorporating global attention in every attention module, we decide to apply it to the output of the central layer of the network to reduce computational complexity. Specifically, we reshape the latent output of the middle layer to " + }, + { + "bbox": [ + 55, + 360, + 295, + 610 + ], + "type": "inline_equation", + "content": "z \\in \\mathbb{R}^{(f \\times n) \\times c}" + }, + { + "bbox": [ + 55, + 360, + 295, + 610 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 360, + 295, + 610 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 360, + 295, + 610 + ], + "type": "text", + "content": " is the number of points " + }, + { + "bbox": [ + 55, + 360, + 295, + 610 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 360, + 295, + 610 + ], + "type": "text", + "content": " divided by the down-sampling factor " + }, + { + "bbox": [ + 55, + 360, + 295, + 610 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 360, + 295, + 610 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 360, + 295, + 610 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 360, + 295, + 610 + ], + "type": "text", + "content": " is the number of latent channels. We then perform self-attention across all frames " + }, + { + "bbox": [ + 55, + 360, + 295, + 610 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 360, + 295, + 610 + ], + "type": "text", + "content": " and all points " + }, + { + "bbox": [ + 55, + 360, + 295, + 610 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 360, + 295, + 610 + ], + "type": "text", + "content": ". Finally, we reshape the output back to its original dimensions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 611, + 295, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 295, + 635 + ], + "type": "text", + "content": "Following [33], we employ the noise prediction loss to optimize MotionDiT, thus the model output is:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 135, + 645, + 295, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 645, + 295, + 658 + ], + "spans": [ + { + "bbox": [ + 135, + 645, + 295, + 658 + ], + "type": "interline_equation", + "content": "\\epsilon_ {\\text {p r e d}} = g _ {\\theta} \\left(z _ {t}, t, s\\right), \\tag {3}", + "image_path": "8fe4f3d2ce96bf69ed0ef8bba5a0e49caab0cc5bfd7b3a42d3c20f8bc4fdd6d0.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "g_{\\theta}" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " represents the MotionDiT parameterized by " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " is the noisy latent variable at time step " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " denotes the conditional input (e.g., textual features), we compute the noise prediction at time step " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": ". The model learns to predict" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": "the noise " + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\mathrm{pred}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": " conditioned on the input. We calculate the mean squared error (MSE) loss between the predicted noise " + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\mathrm{pred}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": " and the true noise " + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 376, + 113, + 553, + 129 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 113, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 376, + 113, + 553, + 129 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {d} = \\mathbb {E} _ {t, z _ {0}, \\epsilon} \\left[ \\| \\epsilon - \\epsilon_ {\\text {p r e d}} \\| ^ {2} \\right], \\tag {4}", + "image_path": "0886c9955b93e05375207ea7ba5445dd255f4199e4b867e18f7732f91a054567.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 133, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 553, + 156 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 133, + 553, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_d" + }, + { + "bbox": [ + 313, + 133, + 553, + 156 + ], + "type": "text", + "content": " represents the loss over all time steps " + }, + { + "bbox": [ + 313, + 133, + 553, + 156 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 133, + 553, + 156 + ], + "type": "text", + "content": ", true noise " + }, + { + "bbox": [ + 313, + 133, + 553, + 156 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0, I)" + }, + { + "bbox": [ + 313, + 133, + 553, + 156 + ], + "type": "text", + "content": ", and ground truth latent variable " + }, + { + "bbox": [ + 313, + 133, + 553, + 156 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 313, + 133, + 553, + 156 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 157, + 555, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 157, + 555, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 555, + 491 + ], + "type": "text", + "content": "CLoP. Additionally, prior works [27, 60] emphasize the significance of intermediate feature representations, which inspire us to propose a LAtent seMantic Alignment (LAMA) loss aimed at enhancing text-based motion control. We expect this term to provide benefits in two key ways: (1) By aligning the latent space of MotionDiT with a largescale pre-trained motion encoder, we can improve both fidelity and diversity of the generated 2D pose; (2) Aligning with the pre-trained pose encoder that is already aligned with text inherently strengthens the influence of text control on motion generation. Unlike CLIP [40], there are currently no readily available large-scale pre-trained models for aligning motion and text. Leveraging the MotionVid dataset, we introduce Contrastive Language-Motion Pretraining (CLoP), which aligns text with 2D pose. This improves the evaluation of 2D pose and latent semantic alignment. Specifically, inspired by CLIP, we employ a similar mechanism to extract textual features. The extracted text features and 2D pose data are then processed through dedicated text encoder " + }, + { + "bbox": [ + 313, + 157, + 555, + 491 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_e(\\cdot)" + }, + { + "bbox": [ + 313, + 157, + 555, + 491 + ], + "type": "text", + "content": " and pose encoder " + }, + { + "bbox": [ + 313, + 157, + 555, + 491 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_p(\\cdot)" + }, + { + "bbox": [ + 313, + 157, + 555, + 491 + ], + "type": "text", + "content": ", respectively. The alignment of the two modalities is facilitated by optimizing a contrastive loss function [40]. Our architecture represents a 2D extension of the approach in [38]. Unlike the original model, we utilize CLIP for poses representation as " + }, + { + "bbox": [ + 313, + 157, + 555, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{p} \\in \\mathbb{R}^{f \\times N \\times 3}" + }, + { + "bbox": [ + 313, + 157, + 555, + 491 + ], + "type": "text", + "content": " with its corresponding text e, capturing frame-wise keypoint coordinates and confidence scores instead of 3D features, the CLoP training objective as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 324, + 495, + 553, + 511 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 495, + 553, + 511 + ], + "spans": [ + { + "bbox": [ + 324, + 495, + 553, + 511 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c} = \\ell_ {c e} \\left(\\ell_ {2} \\left(\\mathbf {h} _ {e} \\mathbf {h} _ {p} ^ {T}\\right); y\\right) + \\ell_ {c e} \\left(\\ell_ {2} \\left(\\mathbf {h} _ {p} \\mathbf {h} _ {e} ^ {T}\\right); y\\right) / 2, \\tag {5}", + "image_path": "4e9c4d6659e8016307a04589b83f00ca8a5842dd5ea42fca33f887e69defbd5f.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 515, + 555, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 515, + 555, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 515, + 555, + 575 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 515, + 555, + 575 + ], + "type": "inline_equation", + "content": "\\ell_2(\\cdot)" + }, + { + "bbox": [ + 313, + 515, + 555, + 575 + ], + "type": "text", + "content": " denote the " + }, + { + "bbox": [ + 313, + 515, + 555, + 575 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 313, + 515, + 555, + 575 + ], + "type": "text", + "content": " normalization and " + }, + { + "bbox": [ + 313, + 515, + 555, + 575 + ], + "type": "inline_equation", + "content": "\\ell_{ce}(\\cdot ;y)" + }, + { + "bbox": [ + 313, + 515, + 555, + 575 + ], + "type": "text", + "content": " represent the cross-entropy loss with ground truth label " + }, + { + "bbox": [ + 313, + 515, + 555, + 575 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 515, + 555, + 575 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 515, + 555, + 575 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_e = \\mathcal{F}_e(\\mathbf{e})\\mathbf{W}_e" + }, + { + "bbox": [ + 313, + 515, + 555, + 575 + ], + "type": "text", + "content": " is the projected text embedding, calculated by projection matrix " + }, + { + "bbox": [ + 313, + 515, + 555, + 575 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_e" + }, + { + "bbox": [ + 313, + 515, + 555, + 575 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 515, + 555, + 575 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_p = \\mathcal{F}_p(\\mathbf{p})\\mathbf{W}_p" + }, + { + "bbox": [ + 313, + 515, + 555, + 575 + ], + "type": "text", + "content": " is the projected pose embedding." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 575, + 555, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 575, + 555, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 575, + 555, + 696 + ], + "type": "text", + "content": "Based on the CLoP, we evaluate the dissimilarity within the feature space (e.g., MSE or cosine similarity metrics), achieving the benefits of the previously mentioned LAMA loss. Rather than using prior methods to extract structural features, we focus on the semantic features of motion, as motion's structural information is simpler while textual information is richer. We also employ a feature projection using a two-layer MLP " + }, + { + "bbox": [ + 313, + 575, + 555, + 696 + ], + "type": "inline_equation", + "content": "g_{\\omega}(\\cdot)" + }, + { + "bbox": [ + 313, + 575, + 555, + 696 + ], + "type": "text", + "content": " parameterized by " + }, + { + "bbox": [ + 313, + 575, + 555, + 696 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 313, + 575, + 555, + 696 + ], + "type": "text", + "content": " to project the latent feature space learned by MotionDiT. As a result, the LAMA loss " + }, + { + "bbox": [ + 313, + 575, + 555, + 696 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_f" + }, + { + "bbox": [ + 313, + 575, + 555, + 696 + ], + "type": "text", + "content": " can be formulated:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 384, + 700, + 553, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 700, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 384, + 700, + 553, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {f} = d \\left(g _ {\\omega} \\left(\\mathbf {h} _ {d} ^ {l}\\right), \\mathbf {h} _ {p}\\right), \\tag {6}", + "image_path": "c0956d4a828a19af9c2ced52ff0e27525340b561f9162595c7e79202c0569ec0.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12395" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 89, + 102, + 523, + 176 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 1. Comparison to Other State-of-the-Art Methods on the MotionVid Subset. The metrics demonstrate that our method outperforms others in terms of pose-text alignment and diversity. Bold indicates the best result." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 89, + 102, + 523, + 176 + ], + "lines": [ + { + "bbox": [ + 89, + 102, + 523, + 176 + ], + "spans": [ + { + "bbox": [ + 89, + 102, + 523, + 176 + ], + "type": "table", + "html": "
MethodFID ↓Rp-top1 ↑Rp-top2 ↑Rp-top3 ↑Diversity ↑MM Dist ↓MM ↑
T2M-GPT [63]496.9070.1760.2990.39564.00042.95014.480
PriorMDM [43]538.9490.1560.2700.36662.33852.55044.135
MLD [6]396.9490.3180.5050.62864.44240.19657.862
Ours149.0070.4510.6380.74368.22032.76160.460
", + "image_path": "0a9ae254838681a4081d276bc5a3eabd4ecd2ca240cba8de092d125812f72501.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 194, + 295, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 194, + 295, + 232 + ], + "spans": [ + { + "bbox": [ + 55, + 194, + 295, + 232 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 194, + 295, + 232 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_d^l" + }, + { + "bbox": [ + 55, + 194, + 295, + 232 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 55, + 194, + 295, + 232 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 194, + 295, + 232 + ], + "type": "text", + "content": "-th layer latent representation of MotionDiT, " + }, + { + "bbox": [ + 55, + 194, + 295, + 232 + ], + "type": "inline_equation", + "content": "d(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 55, + 194, + 295, + 232 + ], + "type": "text", + "content": " is the CLoP function. Consequently, the overall objective " + }, + { + "bbox": [ + 55, + 194, + 295, + 232 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 55, + 194, + 295, + 232 + ], + "type": "text", + "content": " of MotionDiT can be formulated as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 239, + 295, + 252 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 239, + 295, + 252 + ], + "spans": [ + { + "bbox": [ + 138, + 239, + 295, + 252 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {d} + \\lambda_ {f} \\mathcal {L} _ {f}, \\tag {7}", + "image_path": "7097af81369593ea6fa3880409b99defa5d690e7a4875ad785e55c49e4b10a9c.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 259, + 296, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 259, + 296, + 284 + ], + "spans": [ + { + "bbox": [ + 55, + 259, + 296, + 284 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 259, + 296, + 284 + ], + "type": "inline_equation", + "content": "\\lambda_{f}" + }, + { + "bbox": [ + 55, + 259, + 296, + 284 + ], + "type": "text", + "content": " is a hyperparameter to trade off the alignment and denoising process." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 290, + 143, + 301 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 290, + 143, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 290, + 143, + 301 + ], + "type": "text", + "content": "3.3. Pose-to-Video" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 308, + 296, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 308, + 296, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 296, + 462 + ], + "type": "text", + "content": "We propose a model for Pose-to-Video generation, capable of producing human-motion videos based on an initial frame image and a pose sequence. The model leverages [59] as its backbone, maintaining consistent parameters with the original model, incorporates conditional control inspired by [64], and employs data augmentation techniques from [49] to enhance performance. Notably, all training was conducted on the MotionVid dataset, demonstrating its versatility for both Text-to-Pose and Pose-to-Video tasks. It is worth noting that other Pose-to-Video models, such as [20, 49, 66], are also capable of handling this task. For further details on the model architecture, please refer to the supplementary materials." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 473, + 137, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 137, + 487 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 137, + 487 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 493, + 170, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 493, + 170, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 493, + 170, + 506 + ], + "type": "text", + "content": "4.1. Experimental setup" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 510, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 605 + ], + "type": "text", + "content": "Dataset. For computational efficiency, we select a representative subset of 50K samples from MotionVid, covering diverse actions like dancing, squatting, and lifting to ensure diversity and expedite experimentation. This subset maintains result representativeness while enabling efficient testing. We also conducted scaling experiments to assess model performance across various dataset sizes. Qualitative visualization experiments utilize a dataset of 1.2M samples." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": "Implementation Details. For Text-to-Pose, our motion data consists of 128 points per frame, covering the body, face, and hands. The input is structured as " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "f \\times 128 \\times 3" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": ", representing " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": " frames, with 128 points in each frame and 3 channels that encode image coordinates (x, y) along with confidence scores. To maintain uniformity across samples, all sequences are cropped to 64 frames. For Pose VAE, we employ a downsampling factor " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "r = 8" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": ", yielding a latent representation " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "z \\in \\mathbb{R}^{f \\times 16 \\times 3}" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": ". In MotionDiT, we apply a two-" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 194, + 555, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 194, + 555, + 278 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 555, + 278 + ], + "type": "text", + "content": "fold downsampling in the patch embedding. The network consists of 13 layers, with text features extracted from the SD2.1 CLIP model [42] as our text encoder, resulting in a conditional embedding " + }, + { + "bbox": [ + 313, + 194, + 555, + 278 + ], + "type": "inline_equation", + "content": "t_{emb} \\in \\mathbb{R}^{1 \\times 1024}" + }, + { + "bbox": [ + 313, + 194, + 555, + 278 + ], + "type": "text", + "content": ". The AdamW optimizer is used with a learning rate of " + }, + { + "bbox": [ + 313, + 194, + 555, + 278 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-5}" + }, + { + "bbox": [ + 313, + 194, + 555, + 278 + ], + "type": "text", + "content": ". The training is conducted on 8 H20 GPUs, with inference performed on a single H20." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 281, + 556, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 556, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 556, + 449 + ], + "type": "text", + "content": "Evaluation Metrics. To evaluate generated motion data quality, several metrics are employed, consistent with previous works [6, 15, 43, 52, 55, 63]. Frechet Inception Distance (FID) [18] assesses distributional similarity between ground truth and generated motions, using features from a custom-trained CLOP model. MultiModality evaluates diversity for a single text input, while Diversity [15] calculates variance across all motions to quantify output diversity. R-precision measures top-1, top-2, and top-3 recall rates in motion retrieval, evaluating alignment accuracy. MultiModality Distance gauges the distance between textual inputs and generated motions, indicating adherence to prompts. Further details are available in the supplementary materials." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 464, + 446, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 446, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 446, + 475 + ], + "type": "text", + "content": "4.2. Results on Text-to-Pose" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 483, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 555, + 628 + ], + "type": "text", + "content": "Currently, Tender [55] is the only method for generating 2D poses based on text but its code and dataset are unavailable. Thus, we select three SOTA 3D motion generation methods for comparison: T2M-GPT [62], PriorMDM [43], and MLD [6], converting their inputs to 2D poses. All methods are trained on the MotionVid subset with proposed settings to standardize evaluation metrics. Comparison results (Tab. 1) show our model outperforms others, achieving " + }, + { + "bbox": [ + 313, + 483, + 555, + 628 + ], + "type": "inline_equation", + "content": "62.4\\%" + }, + { + "bbox": [ + 313, + 483, + 555, + 628 + ], + "type": "text", + "content": " improvement in FID, " + }, + { + "bbox": [ + 313, + 483, + 555, + 628 + ], + "type": "inline_equation", + "content": "41.8\\%" + }, + { + "bbox": [ + 313, + 483, + 555, + 628 + ], + "type": "text", + "content": " in Rp-top1, " + }, + { + "bbox": [ + 313, + 483, + 555, + 628 + ], + "type": "inline_equation", + "content": "26.3\\%" + }, + { + "bbox": [ + 313, + 483, + 555, + 628 + ], + "type": "text", + "content": " in Rp-top2, and " + }, + { + "bbox": [ + 313, + 483, + 555, + 628 + ], + "type": "inline_equation", + "content": "18.3\\%" + }, + { + "bbox": [ + 313, + 483, + 555, + 628 + ], + "type": "text", + "content": " in Rp-top3. These results indicate our model generates semantically aligned outputs with high diversity." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "The visual comparison with other methods is presented in Fig. 4, and the results demonstrate a significant improvement over existing models. Our method produces poses that better align with the textual constraints, with keypoints remaining intact and exhibiting minimal motion jitter. A more detailed visual comparison can be found in the supplementary materials." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12396" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 113, + 553, + 189 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "text", + "content": "Table 2. The ablation study presents four configurations, progressively adding components of Local, Global, and LAMA loss to the original model. As we move from the initial configuration to the fully enhanced model, performance metrics consistently improve, highlighting the positive impact of each additional component." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 113, + 553, + 189 + ], + "lines": [ + { + "bbox": [ + 56, + 113, + 553, + 189 + ], + "spans": [ + { + "bbox": [ + 56, + 113, + 553, + 189 + ], + "type": "table", + "html": "
MethodFID ↓Rp-top1 ↑Rp-top2 ↑Rp-top3 ↑Diversity ↑MM Dist ↓MM ↑
Vanilla DiT283.0910.3670.5610.68266.58737.04621.461
DiT+Local183.2130.4150.6060.72167.59134.29141.890
DiT+Local+Global162.0220.4330.6200.73067.35733.26957.848
DiT+Local+Global+LAMA149.0070.4510.6380.74368.22032.76160.460
", + "image_path": "0093a61e54227de98980f961005121e3c477e6411f6e7dcb35073869fc63688f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 71, + 201, + 254, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 201, + 254, + 217 + ], + "spans": [ + { + "bbox": [ + 71, + 201, + 254, + 217 + ], + "type": "text", + "content": "A man is holding his head in his hands and continues to do so while looking down at them." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 276, + 201, + 441, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 201, + 441, + 210 + ], + "spans": [ + { + "bbox": [ + 276, + 201, + 441, + 210 + ], + "type": "text", + "content": "A woman is dancing, moving her arms and legs around." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 57, + 219, + 269, + 396 + ], + "blocks": [ + { + "bbox": [ + 57, + 219, + 269, + 396 + ], + "lines": [ + { + "bbox": [ + 57, + 219, + 269, + 396 + ], + "spans": [ + { + "bbox": [ + 57, + 219, + 269, + 396 + ], + "type": "image", + "image_path": "b15f958cd9d769ce18e89f1e707a24f5f3dbc1f38caddf77df9fbf9862682219.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 412, + 555, + 448 + ], + "lines": [ + { + "bbox": [ + 55, + 412, + 555, + 448 + ], + "spans": [ + { + "bbox": [ + 55, + 412, + 555, + 448 + ], + "type": "text", + "content": "Figure 4. Visualization results compared to SOTA Text-to-Pose methods. The results demonstrate that our model significantly outperforms other models. Our method generates poses that are more consistent with the text constraints, with keypoints maintaining their integrity and minimal motion jitter. For a better visual comparison, please refer to the supplementary materials." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 276, + 218, + 545, + 396 + ], + "blocks": [ + { + "bbox": [ + 276, + 218, + 545, + 396 + ], + "lines": [ + { + "bbox": [ + 276, + 218, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 276, + 218, + 545, + 396 + ], + "type": "image", + "image_path": "7f254fa475fbd039c70fbdacc559df5196b80fc5204c1f5c278fff1447034b64.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 466, + 156, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 466, + 156, + 478 + ], + "spans": [ + { + "bbox": [ + 55, + 466, + 156, + 478 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 484, + 295, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 484, + 295, + 531 + ], + "spans": [ + { + "bbox": [ + 55, + 484, + 295, + 531 + ], + "type": "text", + "content": "To validate the efficacy of the three modules proposed in the MotionDiT (i.e., the local feature aggregation, the global attention, and the LAMA loss), we conduct a series of ablation studies. The outcomes are summarized as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 533, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 604 + ], + "type": "text", + "content": "Local Feature Aggregation. Removing the local attention block and integrating a DiT structure similar to Latte [33] in each attention mechanism, we showcase our method's efficacy in capturing local details. The performance improvements, shown in Tab. 2, confirm the robustness of our local feature extraction approach." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "type": "text", + "content": "Global Attention Block. To assess the impact of our global attention mechanism, we perform an ablation test where this component is removed from the model. The findings, detailed in Tab. 2, indicate a notable decline in performance, which underscores the importance of the global attention mechanism in enhancing the model's ability to capture long-range dependencies and global context." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "LAMA Loss. We investigate the role of the semantic alignment loss by training MotionDiT without it. The analysis in" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 466, + 555, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 466, + 555, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 466, + 555, + 538 + ], + "type": "text", + "content": "Tab. 2 shows a significant performance drop, validating the LAMA loss's critical role in guiding optimization towards accurate and stable solutions. These ablation studies provide strong evidence for the proposed algorithms' effectiveness and necessity in enhancing model performance across various aspects." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 548, + 432, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 548, + 432, + 560 + ], + "spans": [ + { + "bbox": [ + 313, + 548, + 432, + 560 + ], + "type": "text", + "content": "4.4. Scaling Law Results" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 566, + 554, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 566, + 554, + 638 + ], + "spans": [ + { + "bbox": [ + 313, + 566, + 554, + 638 + ], + "type": "text", + "content": "We expand the training set size, and as shown in Tab. 3, our metrics improved with the increase in data volume. This demonstrates the validity of using large-scale 2D pose data and suggests that model performance can be further enhanced through an efficient and cost-effective data collection pipeline." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 647, + 507, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 507, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 507, + 659 + ], + "type": "text", + "content": "4.5. Qualitative Results on Pose-to-Video" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "text", + "content": "In our qualitative experiments, we use the generated motion sequences as templates and provide them to the Pose-to-Video model. The specific video results are shown in Fig. 5. Compared to state-of-the-art text-to-video models," + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12397" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 275, + 193 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 275, + 193 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 275, + 193 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 275, + 193 + ], + "type": "image", + "image_path": "9794ba2c9d729f7714a123d1c105c760f5e22d64f6e735046839a90e5c5d2e2d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 202, + 555, + 248 + ], + "lines": [ + { + "bbox": [ + 54, + 202, + 555, + 248 + ], + "spans": [ + { + "bbox": [ + 54, + 202, + 555, + 248 + ], + "type": "text", + "content": "Figure 5. Visualization results compared to SOTA Text-to-Video methods. Mochi1 [50] and CogVideoX [59] exhibit issues such as body distortion, weak motion continuity, and neglecting facial generation. In contrast, HumanDreamer is able to generate more coherent and consistent videos with smoother transitions and better attention to details such as facial expressions. For a better visual comparison, please refer to the supplementary materials." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 279, + 71, + 547, + 193 + ], + "blocks": [ + { + "bbox": [ + 279, + 71, + 547, + 193 + ], + "lines": [ + { + "bbox": [ + 279, + 71, + 547, + 193 + ], + "spans": [ + { + "bbox": [ + 279, + 71, + 547, + 193 + ], + "type": "image", + "image_path": "cf83992ebd1916a8a86ed0092e94b113f5009d74d67720e875a104f1b611ece1.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 58, + 308, + 294, + 380 + ], + "blocks": [ + { + "bbox": [ + 55, + 266, + 295, + 300 + ], + "lines": [ + { + "bbox": [ + 55, + 266, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 295, + 300 + ], + "type": "text", + "content": "Table 3. By increasing the amount of training data, we observe an improvement in model performance, which validates the potential for rapid scalability using large-scale 2D motion data." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 308, + 294, + 380 + ], + "lines": [ + { + "bbox": [ + 58, + 308, + 294, + 380 + ], + "spans": [ + { + "bbox": [ + 58, + 308, + 294, + 380 + ], + "type": "table", + "html": "
DatasizeRp-top1 ↑Rp-top2 ↑Rp-top3 ↑MM Dist ↓
50K(4%)0.4510.6380.74332.761
250K(20%)0.4520.6410.75132.666
500K(40%)0.4640.6570.76632.399
1M(80%)0.4920.6710.76931.116
1.25M(100%)0.5130.6940.79130.139
", + "image_path": "b328feb5f7d525edc1034586214cdbd01dd1bc850d580c32e9dce797e25a3148.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 398, + 296, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 398, + 296, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 398, + 296, + 434 + ], + "type": "text", + "content": "our model exhibits larger and more dynamic movements, with the characters' actions more accurately reflecting the textual descriptions." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 57, + 443, + 293, + 496 + ], + "blocks": [ + { + "bbox": [ + 57, + 443, + 293, + 496 + ], + "lines": [ + { + "bbox": [ + 57, + 443, + 293, + 496 + ], + "spans": [ + { + "bbox": [ + 57, + 443, + 293, + 496 + ], + "type": "image", + "image_path": "4be80cf722b8d40a1101efcbb61b9c1f64e314890ec4d99a43bdff41a57edb16.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 58, + 497, + 292, + 549 + ], + "blocks": [ + { + "bbox": [ + 58, + 497, + 292, + 549 + ], + "lines": [ + { + "bbox": [ + 58, + 497, + 292, + 549 + ], + "spans": [ + { + "bbox": [ + 58, + 497, + 292, + 549 + ], + "type": "image", + "image_path": "e22d0e599981c539142a464e74693c402e4052585c33e9767107d6cb4cf2fd25.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 559, + 296, + 593 + ], + "lines": [ + { + "bbox": [ + 55, + 559, + 296, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 296, + 593 + ], + "type": "text", + "content": "Figure 6. The visualization results demonstrate that using the initial frame pose, different prompt texts can guide the generation of distinct pose sequences." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 612, + 198, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 198, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 198, + 624 + ], + "type": "text", + "content": "4.6. Other Downstream Tasks" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 629, + 295, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 666 + ], + "type": "text", + "content": "In addition to the Pose-to-Video task, the proposed Text-to-Pose component can also be utilized for pose sequence prediction, 2D-3D motion lifting tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "Pose Sequence Prediction. In scenarios with partial pose sequences, the Text-to-Pose model can infer and generate missing parts by conditioning on existing poses and textual movement descriptions. This enables the synthesis of co-" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 268, + 555, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 268, + 555, + 316 + ], + "spans": [ + { + "bbox": [ + 313, + 268, + 555, + 316 + ], + "type": "text", + "content": "herent pose continuations. Given initial and final frames, the model predicts intermediate pose states, completing the sequence (Fig. 6). This is especially useful in animation for supplementing incomplete pose data." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 317, + 556, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 317, + 556, + 413 + ], + "spans": [ + { + "bbox": [ + 313, + 317, + 556, + 413 + ], + "type": "text", + "content": "2D-3D Motion Lifting. Models like [70] can enhance motion data dimensionality, converting 2D motions into realistic 3D motions (Fig. 7). With textual descriptions specifying depth or spatial aspects, the Text-to-Pose system generates richer 3D motion representations. This is crucial for virtual reality (VR) and augmented reality (AR) environments, where realistic 3D motion significantly improves user experience." + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 316, + 426, + 553, + 495 + ], + "blocks": [ + { + "bbox": [ + 316, + 426, + 553, + 495 + ], + "lines": [ + { + "bbox": [ + 316, + 426, + 553, + 495 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 553, + 495 + ], + "type": "image", + "image_path": "8704c8d46a9d0423f47135731a6dc66a7d570932a0b85f220fc02a42437719f4.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 503, + 538, + 515 + ], + "lines": [ + { + "bbox": [ + 329, + 503, + 538, + 515 + ], + "spans": [ + { + "bbox": [ + 329, + 503, + 538, + 515 + ], + "type": "text", + "content": "Figure 7. Visualization results on 2D-3D Motion Lifting." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 546, + 467, + 559 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 467, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 467, + 559 + ], + "type": "text", + "content": "5. Discussion and Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 567, + 555, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 555, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 555, + 652 + ], + "type": "text", + "content": "In this study, we present HumanDreamer, a pioneering decoupled framework for generating human-motion videos that merges text control flexibility with pose guidance controllability. Utilizing MotionVid, the largest dataset for human-motion pose generation, we train MotionDiT for producing structured poses. We introduce LAMA loss to improve semantic alignment, ensuring coherent outputs." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": "Experimental results indicate that using generated poses in Pose-to-Video yields high-quality, diverse human-motion videos, surpassing current benchmarks. These findings confirm the effectiveness and adaptability of our decoupled framework, facilitating versatile video generation." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12398" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 91, + 294, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 294, + 123 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 294, + 123 + ], + "type": "text", + "content": "[1] Youngmin Baek, Bado Lee, Dongyoon Han, Sangdoo Yun, and Hwalsuk Lee. Character region awareness for text detection. In CVPR, 2019. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 124, + 295, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 124, + 295, + 156 + ], + "spans": [ + { + "bbox": [ + 61, + 124, + 295, + 156 + ], + "type": "text", + "content": "[2] Raghav Bansal, Gaurav Raj, and Tanupriya Choudhury. Blur image detection using laplacian operator and open-cv. In SMART, 2016. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 158, + 295, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 158, + 295, + 213 + ], + "spans": [ + { + "bbox": [ + 62, + 158, + 295, + 213 + ], + "type": "text", + "content": "[3] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2, 4, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 213, + 294, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 213, + 294, + 246 + ], + "spans": [ + { + "bbox": [ + 62, + 213, + 294, + 246 + ], + "type": "text", + "content": "[4] Joao Carreira, Eric Noland, Chloe Hillier, and Andrew Zisserman. A short note on the kinetics-700 human action dataset, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 247, + 294, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 247, + 294, + 300 + ], + "spans": [ + { + "bbox": [ + 62, + 247, + 294, + 300 + ], + "type": "text", + "content": "[5] Lin Chen, Xilin Wei, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Bin Lin, Zhenyu Tang, et al. Sharegpt4video: Improving video understanding and generation with better captions. arXiv preprint arXiv:2406.04325, 2024. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 302, + 294, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 302, + 294, + 334 + ], + "spans": [ + { + "bbox": [ + 62, + 302, + 294, + 334 + ], + "type": "text", + "content": "[6] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, and Gang Yu. Executing your commands via motion diffusion in latent space. In CVPR, 2023. 2, 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 335, + 294, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 335, + 294, + 378 + ], + "spans": [ + { + "bbox": [ + 62, + 335, + 294, + 378 + ], + "type": "text", + "content": "[7] Yuxin Chen, Ziqi Zhang, Chunfeng Yuan, Bing Li, Ying Deng, and Weiming Hu. Channel-wise topology refinement graph convolution for skeleton-based action recognition. In ICCV, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 380, + 294, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 294, + 412 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 294, + 412 + ], + "type": "text", + "content": "[8] Ziwei Chen, Qiang Li, Xiaofeng Wang, and Wankou Yang. Liftedcl: Lifting contrastive learning for human-centric perception. In ICLR, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 414, + 294, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 414, + 294, + 467 + ], + "spans": [ + { + "bbox": [ + 62, + 414, + 294, + 467 + ], + "type": "text", + "content": "[9] Joseph Cho, Fachrina Dewi Puspitasari, Sheng Zheng, Jingyao Zheng, Lik-Hang Lee, Tae-Ho Kim, Choong Seon Hong, and Chaoning Zhang. Sora as an agi world model? a complete survey on text-to-video generation. arXiv preprint arXiv:2403.05131, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 469, + 294, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 469, + 294, + 502 + ], + "spans": [ + { + "bbox": [ + 57, + 469, + 294, + 502 + ], + "type": "text", + "content": "[10] Jihoon Chung, Cheng hsin Wu, Hsuan ru Yang, Yu-Wing Tai, and Chi-Keung Tang. Haa500: Human-centric atomic action dataset with curated videos. In ICCV, 2011. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 502, + 294, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 502, + 294, + 535 + ], + "spans": [ + { + "bbox": [ + 57, + 502, + 294, + 535 + ], + "type": "text", + "content": "[11] Bernard Ghanem Fabian Caba Heilbron, Victor Escorcia and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In CVPR, 2015. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 536, + 294, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 536, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 57, + 536, + 294, + 601 + ], + "type": "text", + "content": "[12] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The\" something something\" video database for learning and evaluating visual common sense. In ICCV, 2017. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 602, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 602, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 602, + 294, + 635 + ], + "type": "text", + "content": "[13] Yanchen Guan, Haicheng Liao, Zhenning Li, Guohui Zhang, and Chengzhong Xu. World models for autonomous driving: An initial survey. arXiv preprint arXiv:2403.02622, 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 636, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 294, + 678 + ], + "type": "text", + "content": "[14] Chuan Guo, Xinxin Zuo, Sen Wang, Shihao Zou, Qingyao Sun, Annan Deng, Minglun Gong, and Li Cheng. Action2motion: Conditioned generation of 3d human motions. In ACM MM, 2020. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 680, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 294, + 712 + ], + "type": "text", + "content": "[15] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In CVPR, 2022. 3, 6" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[16] Agrim Gupta, Lijun Yu, Kihyuk Sohn, Xiuye Gu, Meera Hahn, Li Fei-Fei, Irfan Essa, Lu Jiang, and José Lezama. Photorealistic video generation with diffusion models. arXiv preprint arXiv:2312.06662, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 118, + 553, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 148 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 148 + ], + "type": "text", + "content": "[17] Louay Hazami, Rayhane Mama, and Ragavan Thurairatnam. Efficient-vdvae: Less is more. arXiv preprint arXiv:2203.13751, 2022. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 150, + 553, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 150, + 553, + 193 + ], + "spans": [ + { + "bbox": [ + 317, + 150, + 553, + 193 + ], + "type": "text", + "content": "[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS, 2017. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 194, + 553, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 194, + 553, + 237 + ], + "spans": [ + { + "bbox": [ + 317, + 194, + 553, + 237 + ], + "type": "text", + "content": "[19] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 239, + 553, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 270 + ], + "type": "text", + "content": "[20] Li Hu. Animate anyone: Consistent and controllable image-to-video synthesis for character animation. In CVPR, 2024. 2, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 272, + 553, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 553, + 314 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 553, + 314 + ], + "type": "text", + "content": "[21] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. TPAMI, 2013. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 316, + 553, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 553, + 358 + ], + "type": "text", + "content": "[22] Xingxun Jiang, Yuan Zong, Wenming Zheng, Chuangao Tang, Wanchuang Xia, Cheng Lu, and Jiateng Liu. Dfew: A large-scale database for recognizing dynamic facial expressions in the wild. In ACM MM, 2020. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 361, + 553, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 361, + 553, + 414 + ], + "spans": [ + { + "bbox": [ + 317, + 361, + 553, + 414 + ], + "type": "text", + "content": "[23] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, Mustafa Suleyman, and Andrew Zisserman. The kinetics human action video dataset, 2017. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 415, + 553, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 553, + 436 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 553, + 436 + ], + "type": "text", + "content": "[24] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 437, + 553, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 437, + 553, + 468 + ], + "spans": [ + { + "bbox": [ + 316, + 437, + 553, + 468 + ], + "type": "text", + "content": "[25] H. Kuehne, H. Jhuang, E. Garrote, T. Poggio, and T. Serre. HMDB: a large video database for human motion recognition. In ICCV, 2011. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 471, + 553, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 471, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 316, + 471, + 553, + 502 + ], + "type": "text", + "content": "[26] Jiyoung Lee, Seungryong Kim, Sunok Kim, Jungin Park, and Kwanghoonn Sohn. Context-aware emotion recognition networks. In ICCV, 2019. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 504, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 504, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 316, + 504, + 553, + 536 + ], + "type": "text", + "content": "[27] Alexander C Li, Mihir Prabhudesai, Shivam Duggal, Ellis Brown, and Deepak Pathak. Your diffusion model is secretly a zero-shot classifier. In ICCV, 2023. 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 537, + 553, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 553, + 568 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 553, + 568 + ], + "type": "text", + "content": "[28] Jing Lin, Ailing Zeng, Haoqian Wang, Lei Zhang, and Yu Li. One-stage 3d whole-body mesh recovery with component aware transformer. CVPR, 2023. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 570, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 570, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 317, + 570, + 553, + 613 + ], + "type": "text", + "content": "[29] Jing Lin, Ailing Zeng, Shunlin Lu, Yuanhao Cai, Ruimao Zhang, Haoqian Wang, and Lei Zhang. Motion-x: A large-scale 3d expressive whole-body human motion dataset. NeurIPS, 2024. 2, 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 614, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 614, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 317, + 614, + 553, + 656 + ], + "type": "text", + "content": "[30] Jun Liu, Amir Shahroudy, Mauricio Perez, Gang Wang, Ling-Yu Duan, and Alex C Kot. Ntu rgb+ d 120: A large-scale benchmark for 3d human activity understanding. TPAMI, 2019. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "type": "text", + "content": "[31] Yixin Liu, Kai Zhang, Yuan Li, Zhiling Yan, Chujie Gao, Ruoxi Chen, Zhengqing Yuan, Yue Huang, Hanchi Sun, Jianfeng Gao, et al. Sora: A review on background, technology, limitations, and opportunities of large vision models. arXiv preprint arXiv:2402.17177, 2024. 2" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12399" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 116 + ], + "type": "text", + "content": "[32] Shunlin Lu, Ling-Hao Chen, Ailing Zeng, Jing Lin, Ruimao Zhang, Lei Zhang, and Heung-Yeung Shum. Humantomato: Text-aligned whole-body motion generation. arXiv preprint arXiv:2310.12978, 2023. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 163 + ], + "type": "text", + "content": "[33] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 5, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 164, + 295, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 164, + 295, + 208 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 295, + 208 + ], + "type": "text", + "content": "[34] Yue Ma, Yingqing He, Xiaodong Cun, Xintao Wang, Siran Chen, Xiu Li, and Qifeng Chen. Follow your pose: Pose-guided text-to-video generation using pose-free videos. In AAAI, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 295, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 295, + 243 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 295, + 243 + ], + "type": "text", + "content": "[35] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In ICCV, 2019. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 244, + 295, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 244, + 295, + 277 + ], + "spans": [ + { + "bbox": [ + 56, + 244, + 295, + 277 + ], + "type": "text", + "content": "[36] Haomiao Ni, Changhao Shi, Kai Li, Sharon X Huang, and Martin Renqiang Min. Conditional image-to-video generation with latent flow diffusion models. In CVPR, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 278, + 295, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 278, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 56, + 278, + 295, + 300 + ], + "type": "text", + "content": "[37] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, 2023. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 302, + 295, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 302, + 295, + 334 + ], + "spans": [ + { + "bbox": [ + 56, + 302, + 295, + 334 + ], + "type": "text", + "content": "[38] Mathis Petrovich, Michael J. Black, and Gül Varol. TMR: Text-to-motion retrieval using contrastive 3D human motion synthesis. In ICCV, 2023. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 336, + 295, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 336, + 295, + 358 + ], + "spans": [ + { + "bbox": [ + 56, + 336, + 295, + 358 + ], + "type": "text", + "content": "[39] Matthias Plappert, Christian Mandery, and Tamim Asfour. The kit motion-language dataset. Big data, 2016. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 360, + 295, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 360, + 295, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 360, + 295, + 415 + ], + "type": "text", + "content": "[40] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 3, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 417, + 295, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 417, + 295, + 438 + ], + "spans": [ + { + "bbox": [ + 56, + 417, + 295, + 438 + ], + "type": "text", + "content": "[41] Beaumont Romain and Schuhmann Christoph. Laion aesthetics predictor v1. 2022. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 440, + 295, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 440, + 295, + 472 + ], + "spans": [ + { + "bbox": [ + 56, + 440, + 295, + 472 + ], + "type": "text", + "content": "[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 475, + 295, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 475, + 295, + 506 + ], + "spans": [ + { + "bbox": [ + 56, + 475, + 295, + 506 + ], + "type": "text", + "content": "[43] Yoni Shafir, Guy Tevet, Roy Kapon, and Amit Haim Bermano. Human motion diffusion as a generative prior. In ICLR, 2024. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 509, + 295, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 509, + 295, + 542 + ], + "spans": [ + { + "bbox": [ + 56, + 509, + 295, + 542 + ], + "type": "text", + "content": "[44] Amir Shahroudy, Jun Liu, Tian-Tsong Ng, and Gang Wang. Ntu " + }, + { + "bbox": [ + 56, + 509, + 295, + 542 + ], + "type": "inline_equation", + "content": "\\mathrm{rgb + }" + }, + { + "bbox": [ + 56, + 509, + 295, + 542 + ], + "type": "text", + "content": " d: A large scale dataset for 3d human activity analysis. In CVPR, 2016. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 544, + 295, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 544, + 295, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 295, + 588 + ], + "type": "text", + "content": "[45] Gunnar A. Sigurdsson, Gül Varol, Xiaolong Wang, Ali Farhadi, Ivan Laptev, and Abhinav Gupta. Hollywood in homes: Crowdsourcing data collection for activity understanding, 2016. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 590, + 295, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 295, + 632 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 295, + 632 + ], + "type": "text", + "content": "[46] Gunnar A. Sigurdsson, Abhinav Gupta, Cordelia Schmid, Ali Farhadi, and Karteek Alahari. Charades-ego: A large-scale dataset of paired third and first person videos. In ArXiv, 2018. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 635, + 295, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 295, + 667 + ], + "type": "text", + "content": "[47] Tomáš Souček and Jakub Lokoc. Transnet v2: An effective deep network architecture for fast shot transition detection. arXiv preprint arXiv:2008.04838, 2020. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 295, + 713 + ], + "type": "text", + "content": "[48] Rui Sun, Yumin Zhang, Tejal Shah, Jiahao Sun, Shuoying Zhang, Wenqi Li, Haoran Duan, Bo Wei, and Rajiv Ranjan. From sora what we can see: A survey of text-to-video generation. arXiv preprint arXiv:2405.10674, 2024. 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 315, + 72, + 554, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 554, + 127 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 554, + 127 + ], + "type": "text", + "content": "[49] Shuai Tan, Biao Gong, Xiang Wang, Shiwei Zhang, Dandan Zheng, Ruobing Zheng, Kecheng Zheng, Jingdong Chen, and Ming Yang. Animate-x: Universal character image animation with enhanced motion representation. arXiv preprint arXiv:2410.10306, 2024. 2, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 129, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 129, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 315, + 129, + 553, + 149 + ], + "type": "text", + "content": "[50] Genmo Team. Mochi 1. https://github.com/genmoai/models, 2024.8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 151, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 151, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 315, + 151, + 553, + 183 + ], + "type": "text", + "content": "[51] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. In ECCV, 2022. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 185, + 553, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 185, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 315, + 185, + 553, + 216 + ], + "type": "text", + "content": "[52] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In ICLR, 2023. 2, 3, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 315, + 217, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 217, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 315, + 217, + 553, + 271 + ], + "type": "text", + "content": "[53] Xiang Wang, Shiwei Zhang, Changxin Gao, Jiayu Wang, Xiaoqiang Zhou, Yingya Zhang, Luxin Yan, and Nong Sang. Unianimate: Taming unified video diffusion models for consistent human image animation. arXiv preprint arXiv:2406.01188, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 315, + 273, + 553, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 273, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 315, + 273, + 553, + 326 + ], + "type": "text", + "content": "[54] Xiaofeng Wang, Kang Zhao, Feng Liu, Jiayu Wang, Guosheng Zhao, Xiaoyi Bao, Zheng Zhu, Yingya Zhang, and Xingang Wang. Egovid-5m: A large-scale video-action dataset for egocentric video generation. arXiv preprint arXiv:2411.08380, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 315, + 327, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 327, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 315, + 327, + 553, + 381 + ], + "type": "text", + "content": "[55] Yuan Wang, Zhao Wang, Junhao Gong, Di Huang, Tong He, Wanli Ouyang, Jile Jiao, Xuetao Feng, Qi Dou, Shixiang Tang, et al. Holistic-motion2d: Scalable whole-body human motion generation in 2d space. arXiv preprint arXiv:2406.11253, 2024. 2, 3, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 315, + 383, + 553, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 383, + 553, + 436 + ], + "spans": [ + { + "bbox": [ + 315, + 383, + 553, + 436 + ], + "type": "text", + "content": "[56] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, and Dacheng Tao. Gmflow: Learning optical flow via global matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8121-8130, 2022. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 315, + 437, + 553, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 437, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 315, + 437, + 553, + 469 + ], + "type": "text", + "content": "[57] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In AAAI, 2018. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 315, + 471, + 553, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 471, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 315, + 471, + 553, + 502 + ], + "type": "text", + "content": "[58] Zhendong Yang, Ailing Zeng, Chun Yuan, and Yu Li. Effective whole-body pose estimation with two-stages distillation. In ICCV, 2023. 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 315, + 504, + 553, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 504, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 315, + 504, + 553, + 557 + ], + "type": "text", + "content": "[59] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 2, 4, 6, 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 315, + 559, + 553, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 559, + 553, + 602 + ], + "spans": [ + { + "bbox": [ + 315, + 559, + 553, + 602 + ], + "type": "text", + "content": "[60] Sihyun Yu, Sangkyung Kwak, Huiwon Jang, Jongheon Jeong, Jonathan Huang, Jinwoo Shin, and Saining Xie. Representation alignment for generation: Training diffusion transformers is easier than you think, 2024. 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 315, + 604, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 604, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 315, + 604, + 553, + 645 + ], + "type": "text", + "content": "[61] Yan Zeng, Guoqiang Wei, Jiani Zheng, Jiaxin Zou, Yang Wei, Yuchen Zhang, and Hang Li. Make pixels dance: High-dynamic video generation. arXiv preprint arXiv:2311.10982, 2023. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 315, + 647, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 647, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 315, + 647, + 553, + 690 + ], + "type": "text", + "content": "[62] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Shaoli Huang, Yong Zhang, Hongwei Zhao, Hongtao Lu, and Xi Shen. T2m-gpt: Generating human motion from textual descriptions with discrete representations. In CVPR, 2023. 6" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 315, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 315, + 692, + 553, + 713 + ], + "type": "text", + "content": "[63] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Yong Zhang, Hongwei Zhao, Hongtao Lu, Xi Shen, and Ying" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12400" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 475 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "Shan. Generating human motion from textual descriptions with discrete representations. In CVPR, 2023. 2, 3, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 128 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 128 + ], + "type": "text", + "content": "[64] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 129, + 294, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 174 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 174 + ], + "type": "text", + "content": "[65] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001, 2022.3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 175, + 294, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 229 + ], + "type": "text", + "content": "[66] Yang Zhang, Jiaxi Gu, Li-Wen Wang, Han Wang, Junqi Cheng, Yuefeng Zhu, and Fangyuan Zou. Mimic motion: High-quality human motion video generation with confidence-aware pose guidance. arXiv preprint arXiv:2406.19680, 2024. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 231, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 294, + 274 + ], + "type": "text", + "content": "[67] Sijie Zhao, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Muyao Niu, Xiaoyu Li, Wenbo Hu, and Ying Shan. Cv-vae: A compatible video vae for latent generative video models. https://arxiv.org/abs/2405.20279, 2024. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 276, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 308 + ], + "type": "text", + "content": "[68] Weiyu Zhao, Liangxiao Hu, and Shengping Zhang. Diffugesture: Generating human gesture from two-person dialogue with diffusion models. In ICMI, 2023. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 309, + 294, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 294, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 294, + 363 + ], + "type": "text", + "content": "[69] Shenhao Zhu, Junming Leo Chen, Zuozhuo Dai, Qingkun Su, Yinghui Xu, Xun Cao, Yao Yao, Hao Zhu, and Siyu Zhu. Champ: Controllable and consistent human image animation with 3d parametric guidance. arXiv preprint arXiv:2403.14781, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 365, + 294, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 419 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 419 + ], + "type": "text", + "content": "[70] Wentao Zhu, Xiaoxuan Ma, Zhaoyang Liu, Libin Liu, Wayne Wu, and Yizhou Wang. Motionbert: A unified perspective on learning human motion representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 421, + 294, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 421, + 294, + 475 + ], + "spans": [ + { + "bbox": [ + 56, + 421, + 294, + 475 + ], + "type": "text", + "content": "[71] Zheng Zhu, Xiaofeng Wang, Wangbo Zhao, Chen Min, Ni-anchen Deng, Min Dou, Yuqi Wang, Botian Shi, Kai Wang, Chi Zhang, et al. Is sora a world simulator? a comprehensive survey on general world models and beyond. arXiv preprint arXiv:2405.03520, 2024. 2" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12401" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/HumanMM_ Global Human Motion Recovery from Multi-shot Videos/8ae05aea-0260-4aa4-bbb0-352a7f188bbe_content_list.json b/2025/HumanMM_ Global Human Motion Recovery from Multi-shot Videos/8ae05aea-0260-4aa4-bbb0-352a7f188bbe_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3f4e91ab624e0edcd09543b734cb3e05d8d3c81b --- /dev/null +++ b/2025/HumanMM_ Global Human Motion Recovery from Multi-shot Videos/8ae05aea-0260-4aa4-bbb0-352a7f188bbe_content_list.json @@ -0,0 +1,1724 @@ +[ + { + "type": "text", + "text": "HumanMM: Global Human Motion Recovery from Multi-shot Videos", + "text_level": 1, + "bbox": [ + 143, + 108, + 851, + 128 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuhong Zhang $^{1,2,\\dagger*}$ Guanlin Wu $^{2,3,\\ddagger*}$ Ling-Hao Chen $^{1,2,\\ddagger}$ Zhuokai Zhao $^{4}$ Jing Lin $^{1,2}$ Xiaoke Jiang $^{2}$ Jiamin Wu $^{2,5}$ Zhuoheng Li $^{6}$ Hao Frank Yang $^{3}$ Haoqian Wang $^{1\\dagger}$ Lei Zhang $^{2\\dagger}$", + "bbox": [ + 127, + 157, + 866, + 195 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Tsinghua University $^{2}$ IDEA Research $^{3}$ Johns Hopkins University $^{4}$ University of Chicago $^{5}$ HKUST $^{6}$ HKU", + "bbox": [ + 138, + 196, + 859, + 212 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{dsyuhong, guanlinwu0930, thu.lhchen}@gmail.com", + "bbox": [ + 292, + 215, + 707, + 229 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Project page: https://zhangyuhong01.github.io/HumanMM", + "bbox": [ + 210, + 229, + 781, + 248 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5cd68b86bc48e672d8354ecd0927f8ed06c9b3b9c195f60338892f3cf801c111.jpg", + "image_caption": [ + "Figure 1. Recovering a human motion from multi-shot videos. Top: We take two multi-shot table tennis game videos with shot transitions as input. We aim to recover two motions of two athletes (Long MA and Zhendong FAN) from two videos, respectively. The first video is recorded by three shots (“①”, “②”, and “③”), and the second one is recovered by two shots (“④” and “⑤”). Bottom: We recover two motions (Long MA in green and Zhendong FAN in pink), different shots, and camera poses for each multi-shot video. The recovered motion is aligned with the motion in the videos." + ], + "image_footnote": [], + "bbox": [ + 99, + 258, + 897, + 619 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 699, + 326, + 714 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we present a novel framework designed to reconstruct long-sequence 3D human motion in the world coordinates from in-the-wild videos with multiple shot transitions. Such long-sequence in-the-wild motions are highly valuable to applications such as motion generation and motion understanding, but are of great challenge to be recovered due to abrupt shot transitions, partial occlusions, and dynamic backgrounds presented in such videos. Existing", + "bbox": [ + 89, + 731, + 485, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "methods primarily focus on single-shot videos, where continuity is maintained within a single camera view, or simplify multi-shot alignment in camera space only. In this work, we tackle the challenges by integrating an enhanced camera pose estimation with Human Motion Recovery (HMR) by incorporating a shot transition detector and a robust alignment module for accurate pose and orientation continuity across shots. By leveraging a custom motion integrator, we effectively mitigate the problem of foot sliding and ensure temporal consistency in human pose. Extensive evaluations on our created multi-shot dataset from public 3D human datasets demonstrate the robustness of our method in reconstructing realistic human motion in world coordinates.", + "bbox": [ + 511, + 700, + 908, + 897 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution, ‡Core contributor, †Corresponding author. Work done by Yuhong Zhang, Guanlin Wu, Ling-Hao Chen, Jing Lin and Jiamin Wu during the internship at IDEA Research.", + "bbox": [ + 89, + 862, + 483, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1973", + "bbox": [ + 483, + 944, + 514, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 111, + 222, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In recent years, significant advances have been made in 3D human pose estimation, particularly in enhancing the accuracy of human motion recovery (HMR)1 from monocular video sequences. HMR has demonstrated extensive applications in human-AI interaction [1, 2], human motion understanding [3-6], and motion generation [3, 7-24]. While existing methods [25] achieved relatively high performance in recovering human mesh in camera coordinates, estimating human motion in world coordinates remains challenging [26-29] due to inaccurate camera pose estimation and the complexity of reconstructing human motion spatially.", + "bbox": [ + 88, + 136, + 482, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Most current progress in 3D human motion community mainly benefits from large scale data [25, 27-31], and longsequence videos. These resources enhance estimation accuracy for HMR methods and improve the understanding and generation of longer motion sequences for tasks such as motion understanding [3, 32, 33] and generation [3, 7-20, 34-47], even when annotations are derived from markerless capturing methods like pseudo labels [48-51].", + "bbox": [ + 89, + 304, + 483, + 425 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A promising approach to enlarge the scale of the motion databases is to estimate human motions from unlimited online videos in a markerless manner. However, many longsequence online videos are recorded with multiple shots, referred to as multi-shot videos $^2$ , especially prevalent in domains such as sports broadcasting, talk shows, and concerts. In filmmaking and television live show, a \"shot\" denotes an individual camera view capturing a specific moment or action from a particular vantage point [52].", + "bbox": [ + 89, + 425, + 483, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Segmenting multi-shot videos into separate shots inevitably reduces the length of the video sequences, which can be detrimental to tasks that benefit from longer sequences, such as long motion generation [47, 53]. This limitation is highlighted in the existing datasets [54, 55], where the longest clip is less than 20 seconds after segmentation, as shown in Fig. 2. Moreover, focusing exclusively on online single-shot videos diminishes the utilization ratio of available online videos and may negatively impact the diversity of scenarios represented in the created datasets.", + "bbox": [ + 89, + 561, + 482, + 712 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Therefore, how to address the issue of discontinuities caused by shot transitions is notoriously difficult in the community. To resolve this problem, previous works [56-59] have proposed algorithms to address human mesh recovery in a camera space from movies containing shot change between long shots and close-ups.", + "bbox": [ + 89, + 712, + 482, + 805 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/11082310bcbfb4fec8c0288df95839fd66f45dc6e327ff64fc02a4f0b28c9d25.jpg", + "image_caption": [ + "Figure 2. The comparison between the distribution of sequence lengths in different existing large-scale markerless motion datasets with ours. The $x$ -axis and $y$ -axis denote the duration time (s) and percentage of video number, respectively. Our dataset (in green) contains more portion of long-sequence videos in general." + ], + "image_footnote": [], + "bbox": [ + 521, + 89, + 893, + 237 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, recovering human motions in world coordinates from multi-shot videos presents two fundamental challenges that remain underexplored. 1) How to align the human motion and orientation in the world coordinates during shot transitions? Ensuring continuity of human orientation and pose across shots is complicated by factors such as partial visibility of human body (e.g. transitioning from long shot to close-up) and changes in human orientation (e.g. two long shots from different viewpoints). These issues, caused by abrupt changes in camera viewpoints, necessitate robust alignment mechanisms. 2) How to reconstruct accurate human motion in world coordinates? Existing approaches employ Simultaneous Localization and Mapping (SLAM) methods to estimate camera parameters, which are then used to project recovered human meshes from camera to world coordinates [26-29]. This process requires highly accurate camera estimation and must address motion consistency and foot sliding in the recovered human motion within the world space.", + "bbox": [ + 511, + 310, + 906, + 598 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite these challenges, human motion in multi-shot videos often remain continuous across shots, even as camera viewpoints change. This observation suggests that with appropriate handling of shot transitions and camera motion, it is possible to reconstruct consistent and complete 3D human motions throughout multi-shot videos.", + "bbox": [ + 511, + 598, + 905, + 688 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a novel framework HumanMM, Human Motion recovery from Multi-shot videos, to address these challenges. It integrates human pose estimation across shots with robust camera estimation in the world space. Firstly, we develop a shot transition detector to identify frames with shot transitions. To ensure a more robust camera pose estimation, we introduce an enhanced SLAM method incorporating long-term tracking of feature points and exclusion of moving human from bundle adjustment process. We utilize existing HMR method integrated with our enhanced camera estimation to get the initial human parameters for each separated shot. Subsequently, we implement an alignment module to align human orientation based on stereo calibration and smooth human poses through a", + "bbox": [ + 511, + 689, + 908, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1In this paper, the \"human mesh recovery\" refers to recovery in the camera coordinates and the \"human motion recovery\" denotes recovery in the world coordinates. Unless specified otherwise, HMR refers to human motion recovery.", + "bbox": [ + 89, + 814, + 482, + 863 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "2In this paper, a multi-shot video refers to a long-sequence video containing multiple shot transitions. We assume that the camera intrinsics remain consistent across different shots within a multi-shot video.", + "bbox": [ + 89, + 863, + 482, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1974", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "trained multi-shot HMR encoder, which effectively captures the temporal context of human movements across different shots. Finally, after aligning human and camera parameters between shot transitions, we train a motion decoder and a trajectory refiner to smooth the human pose and mitigate issues such as foot sliding, thereby enhancing the overall motion consistency in the reconstructed 3D human motions.", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows.", + "bbox": [ + 109, + 198, + 434, + 210 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present the first approach to reconstruct human motion from multi-shot videos in world coordinates.", + "- We introduce HumanMM, a HMR framework for multi-shot videos. It includes an enhanced camera trajectory estimation method, a human motion alignment module and a motion integrator to ensure accurate and consistent recovery of human pose and orientation in world coordinates across different shots in the whole video.", + "- We develop a multi-shot video dataset ms-Motion to evaluate the performance of HMR from multi-shot videos, based on existing public datasets such as AIST [60] and Human3.6M [61]. Extensive experiments on related benchmarks verify the effectiveness of our method." + ], + "bbox": [ + 89, + 214, + 483, + 409 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 426, + 232, + 441 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. HMR from One-shot Video", + "text_level": 1, + "bbox": [ + 89, + 452, + 334, + 465 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "One-shot videos, captured with a single camera without shot transitions, has been extensively studied within the community for human mesh and motion recovery.", + "bbox": [ + 89, + 476, + 482, + 521 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Human mesh recovery in camera coordinates can be broadly categorized into two approaches: optimization-based methods [62-66] and regression-based methods [30, 67-70]. With the significant advancements of transformer [71], HMR2.0 [25] has surpassed previous methods and benefits several downstream tasks related to HMR.", + "bbox": [ + 89, + 522, + 482, + 611 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Although there are several previous works tried to recover motions in world coordinates with multi-camera capture system [60, 72] and IMU-based methods [73, 74] and enjoy relatively satisfying results, this setup limits their use for applications of infinite in-the-wild monocular videos. To address this limitation, several attempts [26-29] integrate SLAM into the HMR pipeline by first estimating the camera pose using SLAM methods, e.g. DROID-SLAM [75] or DPVO [76], and then project the recovered human motion from camera to world coordinates. To exclude the inconsistencies caused by dynamic objects, such as moving humans, TRAM [27] modifies DROID-SLAM by incorporating human masking and depth-based distance rescaling. However, DROID-SLAM performs dense bundle adjustment (DBA) on feature maps from downsampled images and selects features based only on two consecutive frames rather than long-term video sequences [75-77]. Consequently, masking significantly reduces the number of informative and consistent features, especially when humans occupy large", + "bbox": [ + 89, + 613, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "portions of the image, leading to inaccuracies. Therefore, developing a SLAM method that retains sufficient and representative features for DBA after masking is important.", + "bbox": [ + 511, + 90, + 903, + 137 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. HMR from Multi-shot Video", + "text_level": 1, + "bbox": [ + 511, + 146, + 769, + 161 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multiple shots are fundamental elements of cinematic storytelling and live performances, utilizing various camera positions and focal lengths to create immersive and detailed viewing experiences for audiences [52]. However, most marker-based motion capture (MoCap) datasets [60, 72, 73, 78, 79] consist single-shot videos only, resulting in limited research on HMR from multi-shot videos.", + "bbox": [ + 511, + 169, + 903, + 273 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recovering human motion from multi-shot videos in camera coordinates is already challenging. This is because treating each pose estimation result of each shot separately leads to inconsistencies when combining all estimations, caused by partially or fully invisible human bodies across shot transitions. Pavlakos et al. [56] addresses this issue by focusing on shot changes from long shots to close-ups, which are common in film. They develop smoothness constraints within a temporal Human Mesh and Motion Recovery (t-HMMR) model to infer motions during occlusions caused by shot transitions.", + "bbox": [ + 511, + 273, + 905, + 439 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Advancements in HMR methods [29] for single-shot videos in world coordinates have paved the way for extending HMR to multi-shot videos with varying camera viewpoints. However, aligning human orientation, body pose, and translation continuously across multi-shot videos in world coordinates underexplored. Effective alignment is crucial to maintain motion continuity and coherence, especially when dealing with diverse camera perspectives and abrupt transitions between shots.", + "bbox": [ + 511, + 440, + 905, + 574 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, while substantial progress has been made in HMR from single-shot videos, extending these techniques to multi-shot videos requires addressing additional complexities related to camera pose alignment and motion consistency across shot transitions. We address this challenge by proposing a novel pipeline that ensures accurate and continuous 3D HMR from multi-shot monocular videos.", + "bbox": [ + 511, + 575, + 905, + 679 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 511, + 694, + 604, + 709 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we propose HumanMM to recover human motion from multi-shot videos. The system overview is shown in Fig. 3. Given an input video sequence $\\mathbf{V} = \\{I_t\\}_{t=1}^T$ of length $T$ , where $I_t$ denotes the $t$ -th frame, our objective is to recover human motion in world coordinates. We begin by detecting shot transition frames based on human bounding box (a.k.a. bbox) and 2D keypoints (a.k.a. KPTs) through a shot transition detector (Sec. 3.2). For each clipped shot, we initialize the camera pose (camera rotation and camera translation) and recover initial human motion in world coordinates (Sec. 3.3). The initialized SMPL parameters and camera poses are then fed into a hu", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1975", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ca2ca7aec029cc764656585c8a8e1546a71db368b30708f80823d8e6d5e2ade0.jpg", + "image_caption": [ + "Figure 3. The overview of HumanMM. HumanMM processes multi-shot video sequences by first extracting motion feature such as keypoints and bounding boxes, using ViTPose [80] and image feature using ViT [81]. These features are then segmented into single-shot clips via Shot Transition Detection (Sec. 3.2). Initialized camera (camera rotation $\\mathbf{R}$ and camera translation $\\mathbf{T}$ ) and human (SMPL) parameters for each shot are estimated using Masked LEAP-VO (Sec. 3.3) and GVHMR [29]. Human orientation is aligned across shots through camera calibration (3.4.1), and ms-HMR (Sec. 3.4.2) ensures consistent pose alignment. Finally, a bi-directional LSTM-based trajectory predictor with trajectory refiner predicts trajectory based on aligned motion and mitigates foot sliding throughout the video." + ], + "image_footnote": [], + "bbox": [ + 94, + 90, + 898, + 281 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "man motion alignment module (Sec. 3.4), which aligns human orientations via camera calibration based on human 2D KPTs and smooth the human pose by incorporating pose information across different shots. Additionally, it refines the entire motion sequence through whole video using a temporal motion encoder ms-HMR. Finally, we introduce a post-processing module for motion integration (Sec. 3.5).", + "bbox": [ + 89, + 369, + 483, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Preliminary: 3D Human Model", + "text_level": 1, + "bbox": [ + 89, + 483, + 372, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our method aims to recover motions in world coordinates in the SMPL [82] format, whose pose at frame $t$ can be represented as $\\mathcal{M}_t(\\theta_t,\\beta_t,\\Gamma_t,\\tau_t)\\in \\mathbb{R}^{6890\\times 3}$ . Here, the body pose, body shape, root orientation, and translation are $\\theta_{t}\\in \\mathbb{R}^{23\\times 3}$ , $\\beta_{t}\\in \\mathbb{R}^{10}$ , $\\Gamma_t\\in \\mathbb{R}^3$ , and $\\tau_{t}\\in \\mathbb{R}^{3}$ , respectively. We use $\\mathbf{K}_t^{2D}$ to denote human 2D KPTs at each frame $t$ .", + "bbox": [ + 89, + 506, + 483, + 598 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Shot Transition Detector For Multi-shot Video", + "text_level": 1, + "bbox": [ + 89, + 604, + 483, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our algorithm begins with shot transition detection in one video. As shown in Fig. 3, the shot transition detector has three key components, scene transition detector, bounding box (a.k.a. bounding box) tracking, and human keypoints tracking. (1) Scene change transition detector: Initially, we employ the SceneDetect [83] algorithm to identify scene changes based on significant variations in the background. However, the SceneDetect fails to detect shot transitions when background changes are unnoticeable, illustrated in Fig. 4. Subsequently, we leverage the following modules to bridge the gap. (2) Bbox tracking for shot transition. As a shot change often accompanies with a sudden change of human subject size, we track humans in a video via mmtracking [84]. Consequently, we compute the Intersection over Union (IoU) between neighbor bboxes and identify a shot transition when the IoU falls smaller than a manually tuned threshold. (3) Human pose tracking for shot transition detection. To achieve a finer granularity, we additionally intro", + "bbox": [ + 89, + 628, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c26a99eaab97d284f84c633ae3c919750b0f716898a770539707640218814788.jpg", + "image_caption": [ + "(a) Scene Change" + ], + "image_footnote": [], + "bbox": [ + 516, + 369, + 651, + 477 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/96aeb8efd260b8fa739c9d6857c6f8eeda93fc99f2d21f9e1cea00ef87964fbb.jpg", + "image_caption": [ + "(b) Position Change" + ], + "image_footnote": [], + "bbox": [ + 653, + 369, + 777, + 477 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/35edaa31a1ff346fdad3638262003b6d7ff3eb8e5945f481aa50a7a61b87ec6e.jpg", + "image_caption": [ + "(c) Pose Change" + ], + "image_footnote": [], + "bbox": [ + 779, + 369, + 900, + 477 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Figure 4. Shot transition detection examples. Examples (a), (b), and (c) illustrate multi-shot scenarios in online videos. (a) shows scene transitions detectable by SceneDetect. (b) illustrates significant position changes undetectable by SceneDetect but resolvable withbbox tracking-based method. (c) shows pose or orientation transition, requiring pose tracking-based methods as they cannot be addressed by either SceneDetect or_bbox tracking.", + "bbox": [ + 511, + 489, + 905, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "duce human 2D KPTs to detect extreme corner shot changes in a video. By thresholding the IoU of corresponding keypoints between neighbor frames, we can accurately identify shot transitions even with subtle human movements.", + "bbox": [ + 511, + 592, + 905, + 651 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As each separate module cannot identify all kinds of shot transitions, the three modules are jointly used to clip a video into several sub-sequences serially.", + "bbox": [ + 511, + 652, + 905, + 699 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Human Motion and Camera Pose Estimation For Each Shot", + "text_level": 1, + "bbox": [ + 511, + 709, + 905, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After obtaining the clipped videos, our next goal is to estimate the camera pose and SMPL parameters in the world coordinates for each clipped video. The estimated camera pose and motions for each shot will be used to construct the whole motion sequence in the next stage (Sec. 3.4).", + "bbox": [ + 511, + 750, + 905, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "How to estimate the camera parameters accurately? Our approach for camera parameter calculation is based on a visual odometry (VO) estimation method, LEAP-VO [77]. Utilizing the CoTracker method [85], LEAP-VO estimates the visibility and trajectories of $N$ selected points by ana", + "bbox": [ + 511, + 825, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1976", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "lyzing image gradients across the video sequence. LEAPVO subsequently computes confidence scores for each trajectory, retaining only those with high confidence while discarding trajectories shorter than a predefined threshold. The remaining trajectories undergo bundle adjustment (BA) within a fixed window size to estimate the camera poses.", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, simply applying LEAP-VO in the camera estimation process is still unsatisfactory in most human-centric scenarios. The primary limitation stems from the dynamic movements of human subjects, which typically occupy a substantial portion of each image in human-centric videos. This dynamic presence introduces noise into the camera pose estimation in world coordinates, as the estimation process relies heavily on the relationship between the camera and the static environment. To address this issue, we propose a Masked LEAP-VO algorithm. Our approach involves inputting the image $I_{t}$ and the human bbox at frame $t$ into SAM [86] to generate a human mask. We then assign a visibility value of zero to points within the human mask, effectively excluding these trajectories from the BA process. For clarity, we denote $S_{BA}$ as the window size of BA, $\\hat{n}$ denotes the number of filtered point trajectories, and $w_{ij,\\hat{n}}$ as the normalized weight based on confidence score and visibility. For estimating the camera poses $\\mathbf{G} = \\{\\mathbf{R},\\mathbf{T}\\}$ of orientation and translation, the reprojection loss function for BA can then be formulated as follows,", + "bbox": [ + 91, + 181, + 483, + 483 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {G} = \\underset {\\mathbf {G}, d _ {i}, \\hat {n}} {\\arg \\min } \\sum_ {i} \\sum_ {j \\in | i - j | \\leq S} \\sum_ {B A} w _ {i j, \\hat {n}} | | \\mathcal {F} (\\mathbf {G} _ {i}, \\mathbf {G} _ {j}, d _ {i, \\hat {n}}) - \\Pi_ {i j} (\\mathbf {p} _ {i, \\hat {n}}) | |,\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 489, + 496, + 516 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{F}(\\mathbf{G}_i,\\mathbf{G}_j,d_{i,\\hat{n}})$ denotes the point positions calculated by camera pose $\\mathbf{G}$ at frame $i$ and $j$ with depth $d_{i,\\hat{n}}$ . $\\Pi_{ij}(\\mathbf{p}_{i,\\hat{n}})$ denotes the position for project position of $\\mathbf{p}_{i,\\hat{n}}$ from frame $i$ to $j$ . Consequently, we obtain the camera rotation $\\mathbf{R}_t$ and translation $\\mathbf{T}_t$ from camera pose $\\mathbf{G}_t$ at $t$ .", + "bbox": [ + 89, + 523, + 482, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Recovering human motion in world coordinates with estimated camera parameters. Given an input video, we feed the estimated camera parameters $(\\mathbf{R}_t$ and $\\mathbf{T}_t)$ into the state-of-the-art motion recovering model, GVHMR [29],", + "bbox": [ + 89, + 599, + 482, + 660 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {t} ^ {w}, \\beta_ {t} ^ {w}, \\Gamma_ {t} ^ {w}, \\tau_ {t} ^ {w} = \\mathrm {G V H M R} \\left(I _ {t}, \\mathbf {R} _ {t}, \\mathbf {T} _ {t}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 666, + 480, + 683 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": " Initialized human parameters $\\theta_t^w,\\beta_t^w,\\Gamma_t^w,\\tau_t^w$ and camera parameters $\\mathbf{R}_t,\\mathbf{T}_t$ will input to human motion alignment.", + "bbox": [ + 89, + 685, + 482, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Aligning Human Motion Between Shots", + "text_level": 1, + "bbox": [ + 89, + 726, + 431, + 742 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on initialized world motion for each individual shot, the subsequent question is how to merge discontinuous motions from different shots into a continuous motion sequence as a whole in world coordinates. A straightforward solution is to align all motion sequences to the world coordinate system of the first shot. However, finding the correspondence between different shots is still under-explored and challenging. To resolve this issue, we decompose the motion parameters into camera-dependent and camera-independent ones. The former (Sec. 3.4.1) achieves alignment between shots", + "bbox": [ + 89, + 750, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c18d8cf03b5852f2d79f6a998a219a15655cf336e96ac1be4db29ebd159ef0b2.jpg", + "image_caption": [ + "Figure 5. Human orientation alignment module. Following a shot transition after the foremost purple human mesh (shot $①$ captured by camera $C_0$ ), the unaligned (blue) and aligned (green) motions are captured as shot $②$ and shot “ $③$ ” by camera $C_0'$ and $C_1$ , respectively. $C_0' = C_0$ . To achieve human orientation alignment from shot “ $①$ ” to “ $③$ ”, the camera rotation matrix from $C_0'$ to $C_1$ is computed and applied as the offset of human orientation." + ], + "image_footnote": [], + "bbox": [ + 522, + 90, + 895, + 227 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "via human orientation alignment based on camera calibration, whereas the latter (Sec. 3.4.2) is a trainable module to enhance the continuity of human motion sequence. These two key designs ensure a consistent motion sequence between frames when encountering shot transitions.", + "bbox": [ + 511, + 333, + 905, + 409 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4.1 Aligning Human Orientations Between Shots", + "text_level": 1, + "bbox": [ + 511, + 419, + 880, + 434 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After obtaining the initial SMPL and camera parameters $\\{\\theta_t^i,\\beta_t^i,\\Gamma_t^i,\\tau_t^i,\\mathbf{R}_t^i,\\mathbf{T}_t^i\\}$ for each shot, directly concatenating motions between shots result abrupt changes of human poses and orientations. To address this issue, we introduce the Orientation Alignment Module (OAM), as shown in Fig. 5, to align human orientations. As the whole motion sequence is continuous, we have the following assumption. Assumption 1 Human orientations and translations during the shot transition in world coordinates are continuous. To align the orientations between two frames with shot transition under Assumption 1, we decompose the human orientation with shot transitions in world coordinates as,", + "bbox": [ + 511, + 441, + 906, + 625 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {R} \\left(\\Gamma_ {\\text {w o r l d}}\\right) = \\mathrm {R} _ {\\delta_ {\\text {c a m}}} \\mathrm {R} \\left(\\Gamma_ {\\text {v i e w}}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 619, + 627, + 903, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{R}_{\\delta_{\\mathrm{cam}}}$ represents the camera rotation on the Y-axis between current $t$ -th and previous $t - 1$ -th frame, $\\Gamma_{\\mathrm{view}}$ denotes the human orientation estimated by the current shot, and $\\mathbb{R}(\\cdot):\\mathbb{R}^3\\to \\mathbb{R}^9$ is the mapping from axis angle to rotation matrix. As $\\Gamma_{\\mathrm{view}}$ in current shot can be estimated independently, mentioned in Sec. 3.3, obtaining accurate $\\Gamma_{\\mathrm{world}}$ in Eq. (2) remains a key challenge to estimate the relative camera rotation $\\mathbf{R}_{\\delta_{\\mathrm{cam}}}$ between frames in shot transitions.", + "bbox": [ + 511, + 643, + 906, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Estimating the relative camera pose $\\mathbf{R}_{\\delta_{\\mathrm{cam}}}$ between transition frames. Different from our approach of estimating camera pose in each shot (Sec. 3.3), we do not mask the human subject when estimating camera rotation $\\mathbf{R}_{\\delta_{\\mathrm{cam}}}$ . Instead, we use human 2D KPTs as explicit feature matching. Specifically, we filter out unmatched keypoints based on their visibility and unaligned direction using RANSAC [87], effectively addressing camera pose estimation during shot transitions. This procedure is referred", + "bbox": [ + 511, + 765, + 908, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1977", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "to as Camera Calibration (a.k.a. epipolar-geometry-based camera extrinsic estimation), and is detailed below.", + "bbox": [ + 89, + 90, + 482, + 119 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Camera Calibration, we assume that the human translations remain unchanged across the shot transition, implying that only the camera's orientation changes (i.e. Assumption 1). Consequently, we calculate the orientation offset by determining the change in camera orientation using camera calibration. We begin by extracting human 2D KPTs from two consecutive frames during the shot transition. Due to the shot transition, the visibility of 2D KPTs may vary, e.g. occlusion in some shots. Therefore, we employ ED-Pose [88] to filter out invisible 2D KPTs between shot transition frames. Subsequently, RANSAC identifies matching 2D KPTs corresponding to the most possible camera rotation direction. These matched 2D KPTs facilitate the estimation of the aligned camera rotation $\\mathbf{R}_{\\delta_{\\mathrm{cam}}}$ . The detailed estimation process is as follows.", + "bbox": [ + 89, + 121, + 483, + 347 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We denote the detected 2D KPTs of two frames in the shot transition as $\\mathbf{S}_1 = [(x_1^{(1)},y_1^{(1)}),(x_1^{(2)},y_1^{(2)}),\\dots ,(x_1^{(N)},y_1^{(N)})]^{\\top}\\in \\mathbb{R}^{2\\times N}$ and $\\mathbf{S}_2 = [(x_2^{(1)},y_2^{(1)}),(x_2^{(2)},y_2^{(2)}),\\dots ,(x_2^{(N)},y_2^{(N)})]^{\\top}\\in$ $\\mathbb{R}^{2\\times N}$ . The essential matrix $\\mathbf{E} = [\\mathbf{T}]_{\\times}\\mathbf{R}$ should satisfy the following orthogonal property such that,", + "bbox": [ + 89, + 348, + 483, + 441 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {S} _ {1} ^ {\\top} \\mathbf {E} \\mathbf {S} _ {2} = \\mathbf {0}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 238, + 441, + 480, + 460 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Once $\\mathbf{E}$ is obtained by solving Eq. (3), we enforce the rank-2 constraint on $\\mathbf{E}$ through SVD decomposition and subsequently derive the aligned camera rotation $\\mathbf{R}_{\\delta_{\\mathrm{cam}}}$ between two frames (cf. Hartley et al. [89] for more details).", + "bbox": [ + 89, + 470, + 482, + 531 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In summary, we reformulate the alignment problem of human orientation in shot transitions as estimating the relative camera rotation $\\mathbf{R}_{\\delta_{\\mathrm{cam}}}$ between frames. Accordingly, we obtain the camera rotation $\\mathbf{R}_{\\delta_{\\mathrm{cam}}}$ via camera calibration.", + "bbox": [ + 89, + 532, + 483, + 594 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4.2 Aligning Human Poses Between Shots", + "text_level": 1, + "bbox": [ + 89, + 604, + 408, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In shot transition, video sequences recorded by two shots are often with various occlusions. However, unoccluded body parts in two shots can be complementary to each other for motion alignment. Thus, we introduce the multi-shot HMR (ms-HMR, i.e. $\\mathsf{E}_M(\\cdot)$ ) module to refine the whole motion sequence. As shown in Fig. 6, the ms-HMR is a Transformer encoder-like architecture, whose input and output are the estimated global motion and the refined global motion, respectively. The process can be formulated as,", + "bbox": [ + 89, + 628, + 483, + 765 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\phi_ {1}, \\phi_ {2}, \\dots , \\phi_ {T} = \\mathrm {E} _ {M} \\left(\\theta_ {1}, \\theta_ {2}, \\dots , \\theta_ {T}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 155, + 770, + 482, + 786 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\phi_*$ denotes the refined motion of each frame. With this design, our method can adapt to diverse occlusions of human body brought by shot transitions.", + "bbox": [ + 89, + 791, + 483, + 838 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.5. Post-processing Module for Motion Integration", + "text_level": 1, + "bbox": [ + 89, + 847, + 482, + 864 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Trajectory Predictor and Trajectory Refiner. Based on the aligned human pose and orientation, we introduce a", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1cb1151ba9da0460afacd94ca6f5dc2be9bccac51fe297050c8a9acf8ea47910.jpg", + "image_caption": [ + "Figure 6. ms-HMR Structure. The initial human pose parameters $\\theta$ across multiple video shots are input into a transformer with shot-index-based positional encoding. This enables ms-HMR to generate consistent human poses across all shots in the video. bi-directional LSTM trajectory predictor to recover foot-ground contact probabilities $p_t^c$ , and root velocity $v_t$ as," + ], + "image_footnote": [], + "bbox": [ + 519, + 90, + 913, + 330 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\np _ {t} ^ {c}, v _ {t} = \\operatorname {L S T M} \\left(\\phi_ {1} ^ {m}, \\Gamma_ {1}, \\mathbb {F} (I _ {1}), \\phi_ {2} ^ {m}, \\Gamma_ {2}, \\mathbb {F} (I _ {2}), \\dots , \\right. \\\\ \\left. \\quad \\left(m, \\mathbb {F} - \\mathbb {F} (I _ {m})\\right)\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 417, + 903, + 443 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\phi_ {T} ^ {m}, \\Gamma_ {T}, F (I _ {T})),\n$$\n", + "text_format": "latex", + "bbox": [ + 751, + 436, + 861, + 453 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathsf{F}(\\cdot)$ denotes the image feature of each frame extracted by ViT [81]. Accordingly, the contact probabilities $p_t^c$ , and velocity $v_t$ are supervised by the ground-truth labels with MSE loss and are used to reconstruct the trajectory. Besides, we extend the trajectory refiner in WHAM [28] to alleviate foot sliding problem in our estimated trajectory.", + "bbox": [ + 511, + 453, + 905, + 544 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Benchmarking Multi-shot Motion Recovery", + "text_level": 1, + "bbox": [ + 511, + 549, + 903, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset Construction. To create a multi-shot 3D human motion dataset, we introduce ms-Motion by processing existing public 3D human datasets with multiple camera settings and ground truth human and camera parameters, specifically AIST [60] and Human3.6M (H3.6M) [61]. In our construction pipeline, we randomly separate each original one-shot video into several clips. Then, we choose each clip from different shots and concatenate them together as one video recorded by multiple shots. For example, AIST provides each video with eight cameras C0, C1, ..., C7 from different view point and we choose a video and split it into 5 clips at t0, t1, ..., t4. For frames in these separated clips, we choose frames shot by a random camera for each clip and combine five clips as one multi-shot video. Therefore, we construct a multi-shot version of AIST and H3.6M, which are named ms-AIST and ms-H3.6M subsets. Then we combine them and name this new dataset ms-Motion. The detailed statistics of ms-Motion are shown in Tab. 2. We do not compare with other existing 3D human datasets as they contain limited number of multi-shot videos.", + "bbox": [ + 511, + 568, + 906, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Benchmark Evaluation Protocol. To evaluate the performance of our proposed methods on multi-shot videos, our", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1978", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/69b37aa15e1d1f5cb6eb89130ebbcbec7d4b5f6906ffe6992fc9d17bc22a67e4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetModels2-Shot3-Shot4-Shot
PA.↓WA.↓RTE↓ROE↓Jitter↓F.S.↓PA.↓WA.↓RTE↓ROE↓Jitter↓F.S.↓PA↓WA↓RTE↓ROE↓Jitter↓F.S.↓
ms-AISTSLAHMR [2023]72.34341.759.6296.2662.593.2680.35510.7710.33101.3672.394.4390.32803.6912.11104.0780.3716.52
WHAM [2024]65.34336.824.3984.4825.242.7578.68451.325.1489.8424.062.99102.88603.935.5790.0726.293.62
GVHMR [2024]60.72231.366.2096.5834.877.6570.33357.167.5599.6934.469.4283.77563.178.96104.5335.679.78
Ours36.82121.352.5669.2333.272.6638.52141.383.6467.7135.073.5539.63161.524.5570.3139.494.09
ms-H3.6MSLAHMR [2023]80.67352.6116.67111.9737.807.9397.15562.1016.91118.4652.239.96107.90748.5817.85116.7265.1511.58
WHAM [2024]71.32313.5811.4182.4218.405.0979.51423.9812.3684.8518.875.0390.50512.6612.9190.3418.405.69
GVHMR [2024]64.63254.306.9481.9318.458.8080.79296.7485.2558.2618.3610.6285.19471.539.1291.6319.4710.65
Ours40.52132.133.6553.3919.054.1745.35145.365.3358.2617.354.6250.59147.626.2061.2219.775.12
", + "bbox": [ + 94, + 88, + 903, + 214 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/871d27dcb28520e36167e8c52ba5dff7bd85d5247cd9f6b5a7f968b8b43caf38.jpg", + "table_caption": [ + "Table 1. Quantitative comparison of different HMR methods on ms-Motion dataset. We record the results for ms-AIST and ms-H3.6M separately. PA and WA means PA-MPJPE and WA-MPJPE respectively, while F.S. is Foot Sliding. Our proposed method has achieved the best performance in PA-MPJPE, WA-MPJPE, RTE and ROE across ms-Motion among these methods." + ], + "table_footnote": [], + "table_body": "
DatasetDuration(s)VideosFPSMax LengthMin LengthShots
ms-Motion23.76003014783142, 3, 4
", + "bbox": [ + 91, + 258, + 480, + 296 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "target is to evaluate metrics for accurately reflecting the performance on videos with shot transitions. To this end, we use Root Orientation Error (a.k.a. ROE in deg) to measure the performance of the proposed method on human orientation alignment across different shots. Besides, we use Root Translation Error (a.k.a. RTE in $m$ ) to assess the performance of the proposed method on global trajectory recovery. Jitter $\\left( \\frac{10m}{fps^3} \\right)$ is also used to evaluate the stability of recovered human pose from multi-shot videos. We also include foot sliding $(cm)$ , the averaged displacement of foot vertices during contact with the ground, to assess the precision of recovered motion in the world coordinates [28].", + "bbox": [ + 88, + 332, + 483, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Experiment", + "text_level": 1, + "bbox": [ + 89, + 531, + 215, + 547 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Datasets and Metrics", + "text_level": 1, + "bbox": [ + 89, + 556, + 289, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation Datasets. To evaluate the performance of our proposed pipeline for multi-shot videos, we use ms-Motion dataset and EMDB-1 dataset [73] with self-added noise for the evaluation of ablation study. For camera trajectory estimation, we use EMDB-1 and EMDB-2 split [73] as they contain the GT moving camera trajectory. Our self-created dataset contains 600 multi-shot videos, 42.7K frames, totaling 237 minutes. EMDB-1 split contains 17 video sequences totaling 13.5 minutes and EMDB-2 split contains 25 sequences in a total of 24.0 minutes.", + "bbox": [ + 88, + 580, + 482, + 731 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation Metrics. For shot detection we use Recall, Precision and F1 Score as evaluation metrics. For 3D human pose estimation-related tasks, we use ROE, RTE, jitter, and foot-sliding for evaluating the human motion recovery results on multi-shot videos. For the ablation study of our proposed pipeline, we evaluate the Procrustes-aligned Mean Per Joint Position Error (a.k.a. PA-MPJPE) and Per Vertex Error (a.k.a. PVE) as additional metrics besides previous mentioned ones. For camera pose estimation, we use absolute trajectory error (a.k.a. ATE) $(m)$ , Relative Pose Error (a.k.a. RPE) rotation $(deg)$ , and RPE translation $(m)$ .", + "bbox": [ + 88, + 734, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/53927265636e81df53b21d52febe7d146dc8b88399c9d22d92586f89535d5d6e.jpg", + "table_caption": [ + "Table 2. Statistics of the ms-Motion dataset. By shots, we mean the number of shot transitions in a single video." + ], + "table_footnote": [], + "table_body": "
Methodsms-Motion
Recall↑Precision↑F1 Score↑
Scenes Detect (SD) [83]0.740.720.70
SD+Bbox Tracking (Bbox)0.880.850.86
SD+Bbox+Pose Tracking0.960.880.92
", + "bbox": [ + 514, + 258, + 702, + 306 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2259ce20b8487ba6de83159f464280004aafe73f2abdafbd5694f00b23c20c6f.jpg", + "table_caption": [ + "Table 3. Comparison between difference shot detection algorithms. We evaluate our shot transition detector on our proposed multi-shot video human motion dataset ms-Motion." + ], + "table_footnote": [], + "table_body": "
MethodsATE↓RPE Trans.↓RPE Rot.↓
DPVO (w/o mask)0.481.851.06
Masked DPVO0.481.570.97
LEAP-VO (w/o mask)0.500.930.97
Masked LEAP-VO0.510.920.95
", + "bbox": [ + 712, + 261, + 897, + 304 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Camera tracking results on EMDB-1 [73]. Our method has achieved $\\sim 50\\%$ ↓ on RPE trans. than that of the original DPVO and perform the best in RPE rot. metrics.", + "bbox": [ + 709, + 305, + 898, + 358 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Implementation Details", + "text_level": 1, + "bbox": [ + 513, + 364, + 730, + 380 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The ms-HMR, the trajectory, and foot sliding refiner are trained on the AMASS [78], 3DPW [79], Human3.6M [61], and BEDLAM [90] datasets, and evaluated on EMDB and our ms-Motion. During training, we introduce random rotational noise (ranging from 0 to 1 radian) along the y-axis to the root pose $\\Gamma$ and random noise to the body pose $\\theta$ at random positions to simulate the inaccuracies of pre-estimated human motions caused by shot transitions in multi-shot videos. This strategy enables the network to robustly recover smooth and consistent human motion from noisy initial parameters. The benchmark test results were obtained after training for 80 epochs on one NVIDIA-A100 GPU.", + "bbox": [ + 509, + 385, + 906, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Main Results: Comparison of Global Human Motion Recovery Results on the Benchmark", + "text_level": 1, + "bbox": [ + 511, + 575, + 903, + 607 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare our proposed method HumanMM with several state-of-the-art HMR methods (SLAHMR [26], WHAM [28] and GVHMR [29]) on our proposed benchmark ms-Motion. As illustrated in Tab. 1, our proposed method has achieved the best performance for PA&WAMPJPE, RTE and ROE through videos with all numbers of shots across ms-AIST and ms-H3.6M, indicating that our method reconstructs both the global human motion and orientations in the world coordinates more accurately and robustly. For the foot sliding metric, our method also performs as the best on ms-H3.6M across all numbers of shots.", + "bbox": [ + 509, + 612, + 906, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 789, + 679, + 804 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Human-centric Scene Shot Boundary Detection Evaluation. To evaluate the performance of our proposed Shot Transition Detector, we test the algorithm on our proposed multi-shot human motion recovery benchmark and compare the output frame list of shot transitions with the ground truth (GT) of our dataset. As shown in Tab. 3, by apply", + "bbox": [ + 511, + 810, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1979", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a8e7df8b005299e7c51581efca7768e5e9d04f98d298dc31fcbcc2a35bfae541.jpg", + "image_caption": [ + "Figure 7. Qualitative comparison of different HMR methods on ms-Motion dataset. The side view of the rendered mesh for input mutli-shot video is shown in (a), while the top view is shown in (c). We also draw the comparison of the human trajectory as shown in (b). Our method is the most similar as GT in both rendered motion and trajectories among these methods." + ], + "image_footnote": [], + "bbox": [ + 96, + 89, + 898, + 219 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b1e52d08de99c1313e86f3c01c5a2dafd33d011a4955eadce8b70b6df28b453b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Methods|PA-MPJE↓|MPJPE↓WA-MPJE↓W-MPJPE↓PVE↓ACCEL↓RTE↓ROE↓F.S.↓
Baseline (Concat)106.48141.67273.15553.67122.156.1410.8691.5514.91
w/o ms-HMR78.24101.52246.42436.5785.775.873.8950.633.54
w/o OAM73.5692.13243.65425.1879.645.676.6176.744.45
w/o traj. pred.50.4983.68231.75432.1775.775.755.5247.684.96
w/o traj. ref.50.4983.68198.58397.6575.775.234.0647.687.84
HumanMM (Ours)50.4983.68194.77393.2175.775.163.5447.683.28
", + "bbox": [ + 91, + 262, + 486, + 328 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ing the proposed finer granularity shot detection methods, the number of recall, precision, and F1 score all increases consistently. The combination of three steps (ScenesDetect, bbox tracking, and pose tracking) has achieved 0.96, 0.88, and 0.92 on the recall, precision, and F1 score, respectively, which indicates a comparable performance in shot boundary detection. Besides, as can be seen in the results, the latter two steps of shot detection contribute to the fine-grained final results significantly and jointly.", + "bbox": [ + 88, + 356, + 482, + 491 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Key modules in the Proposed Method. We compare our methods with four variants on EMDB with noise dataset, as shown in Tab. 5, ms-HMR is the key component for the improvement in PA-MPJPE and PVE, which indicates a more accurate modeling of the whole motion sequence. This design serves as a recovery module to estimate some invisible body parts in some shots. Additionally, the orientation alignment module (OAM, in Sec. 3.4) is also a critical block for accurate human orientation estimation, indicated by the metric ROE. This module helps to model the global human motion between shots. For foot sliding, the results in Tab. 5 also show that the trajectory refiner (Sec. 3.5) in our method helps mitigate the foot sliding issue.", + "bbox": [ + 88, + 492, + 482, + 688 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison on Camera Trajectory Estimation. To evaluate the performance of our proposed camera trajectory estimation method Masked LEAP-VO, we evaluate the camera trajectory accuracy on EMDB-1 and EMDB-2. For more convenient comparison, we introduce two baselines, DPVO [76], which has been widely used in HMR methods such as WHAM [28] and GVHMR [29], and LEAP-VO [77]. To provide more intuition about the insights of masking dynamic humans in the video, we also implement a variant, Masked DPVO, by applying SAM at the patchify stage of DPVO to exclude patches containing human pixels. As shown in Tab. 4 and Tab. 6, compared with baseline methods, our key design of masking dynamic human subjects improves the result in both RPE Translation and", + "bbox": [ + 89, + 688, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7ba6046069786c2d9a31500c5cc711655eb9d5f0a6198989c42cfedb9e81f834.jpg", + "table_caption": [ + "Table 5. Ablation studies on different combinations of HumanMM modules. We evaluate the contributions of each key module on EMDB-1." + ], + "table_footnote": [], + "table_body": "
Methods\\( \\left| \\mathrm{{ATE}}\\right| \\downarrow \\)RPE Trans. \\( \\downarrow \\)RPE Rot. \\( \\downarrow \\)
DPVO (w/o mask)0.481.071.26
Masked DPVO0.500.861.21
LEAP-VO (w/o mask)0.500.831.21
Masked LEAP-VO0.490.831.19
", + "bbox": [ + 496, + 262, + 671, + 304 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1897d43a75b8be4a6316816eef6742912cea9d95717f9a8b2852d051779766b1.jpg", + "table_caption": [ + "Table 6. Camera tracking results on EMDB-2 [73]. Masked LEAP-VO performs better in both RPE Trans. and RPE Rot. metrics." + ], + "table_footnote": [], + "table_body": "
MethodsWA-MPJE↓W-MPJE↓RTE↓Jitter↓F.S.↓
DPVO305.40117.105.1017.903.50
Masked DPVO303.90116.404.1017.403.50
LEAP-VO284.10112.803.1016.303.50
Masked LEAP-VO283.70112.703.1016.303.50
", + "bbox": [ + 681, + 262, + 901, + 303 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7. Global motion recovery results on EMDB-2 [73]. We input estimated camera parameters from different methods into GVHMR for comparison on HMR.", + "bbox": [ + 679, + 304, + 903, + 351 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "RPE Rotation while maintaining competitive ATE. This result indicates the effectiveness of the design of masking dynamic human subjects in the process of camera trajectory estimation. Compared with the DPVO baseline, our method achieves $\\sim 50\\%$ $\\downarrow$ RPE translation on EMDB-1. In addition, we run GVHMR [29] on EMDB-2 with different estimated camera trajectories. The results is shown in Tab. 7, which further illustrates the effectiveness of our method.", + "bbox": [ + 511, + 356, + 906, + 476 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion and Discussion", + "text_level": 1, + "bbox": [ + 511, + 487, + 761, + 502 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Conclusion. In this paper, we introduce HumanMM, the first framework designed for human motion recovery from multi-shot videos in world coordinates. HumanMM addresses the challenges inherent in multi-shot videos by incorporating three key components: an enhanced camera trajectory estimation method called masked LEAP-VO, a human motion alignment module that ensures consistency across different shots, and a post-processing module for seamless motion integration. Extensive experiments demonstrate that HumanMM outperforms existing human motion recovery methods across various benchmarks, achieving state-of-the-art accuracy on our newly created multi-shot human motion dataset, ms-Motion.", + "bbox": [ + 511, + 507, + 906, + 702 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and Future Work. While HumanMM represents an dvancement in human motion recovery from multishot videos in world coordinates, its performance may decline when faced with an excessive number of shot transitions. Despite these challenges, HumanMM provides a solid baseline for human motion recovery from multi-shot videos and can be employed in annotating markerless human motion datasets. Our newly introduced dataset, msMotion, offers a valuable benchmark for evaluating general human motion recovery methods in world coordinates, especially regarding their performance on multi-shot videos. Based on the proposed method, our future work aims to enlarge the related datasets for larger-scale motion databases.", + "bbox": [ + 511, + 704, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1980", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This research was funded through the National Key Research and Development Program of China Project No. 2022YFB36066), in part by the Shenzhen Science and Technology Project under Grant (KJZD20240903103210014). The author team would also like to convey sincere thanks to Ms. Yaxin Chen from IDEA Research for the expressive dance motion used in the demo presentation.", + "bbox": [ + 89, + 114, + 483, + 227 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 256, + 187, + 272 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jingbo Wang, Yu Rong, Jingyuan Liu, Sijie Yan, Dahua Lin, and Bo Dai. Towards diverse and natural scene-aware 3d human motion synthesis. In CVPR, pages 20428-20437, 2022. 2", + "[2] Zeqi Xiao, Tai Wang, Jingbo Wang, Jinkun Cao, Wenwei Zhang, Bo Dai, Dahua Lin, and Jiangmiao Pang. Unified human-scene interaction via prompted chain-of-contacts. In ICLR, 2024. 2", + "[3] Chuan Guo, Xinxin Zuo, Sen Wang, and Li Cheng. Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In ECCV, pages 580-597, 2022. 2", + "[4] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. NeurIPS, 2024.", + "[5] Liang Pan, Jingbo Wang, Buzhen Huang, Junyu Zhang, Haofan Wang, Xu Tang, and Yangang Wang. Synthesizing physically plausible human motions in 3d scenes. In 3DV, 2024.", + "[6] Jingbo Wang, Ye Yuan, Zhengyi Luo, Kevin Xie, Dahua Lin, Umar Iqbal, Sanja Fidler, and Sameh Khamis. Learning human dynamics in autonomous driving scenarios. In ICCV, pages 20739-20749, 2023. 2", + "[7] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. In ECCV, pages 358–374, 2022. 2", + "[8] Mathis Petrovich, Michael J Black, and Gül Varol. Temos: Generating diverse human motions from textual descriptions. In ECCV, pages 480-497, 2022.", + "[9] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. IEEE TPAMI, 2024.", + "[10] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model. In ICLR, 2022.", + "[11] Zan Wang, Yixin Chen, Tengyu Liu, Yixin Zhu, Wei Liang, and Siyuan Huang. Humanise: Language-conditioned human motion generation in 3d scenes. NeurIPS, pages 14959-14971, 2022.", + "[12] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, and Gang Yu. Executing your commands via motion diffusion in latent space. In CVPR, pages 18000-18010, 2023." + ], + "bbox": [ + 93, + 282, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Ye Yuan, Jiaming Song, Umar Iqbal, Arash Vahdat, and Jan Kautz. Physdiff: Physics-guided human motion diffusion model. In ICCV, pages 16010-16021, 2023.", + "[14] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Yong Zhang, Hongwei Zhao, Hongtao Lu, Xi Shen, and Ying Shan. Generating human motion from textual descriptions with discrete representations. In CVPR, pages 14730-14740, 2023.", + "[15] Yonatan Shafir, Guy Tevet, Roy Kapon, and Amit H Bermano. Human motion diffusion as a generative prior. In ICLR, 2024.", + "[16] Korrawe Karunratanakul, Konpat Preechakul, Supasorn Suwajanakorn, and Siyu Tang. Guided motion diffusion for controllable human motion synthesis. In CVPR, pages 2151-2162, 2023.", + "[17] Ling-Hao Chen, Shunlin Lu, Wenxun Dai, Zhiyang Dou, Xuan Ju, Jingbo Wang, Taku Komura, and Lei Zhang. Pay attention and move better: Harnessing attention for interactive motion generation and training-free editing, 2024.", + "[18] Zeqi Xiao, Tai Wang, Jingbo Wang, Jinkun Cao, Wenwei Zhang, Bo Dai, Dahua Lin, and Jiangmiao Pang. Unified human-scene interaction via prompted chain-of-contacts. In ICLR, 2024.", + "[19] Yiming Xie, Varun Jampani, Lei Zhong, Deqing Sun, and Huaizu Jiang. Omnicontrol: Control any joint at any time for human motion generation. In ICLR, 2024.", + "[20] Shunlin Lu, Ling-Hao Chen, Ailing Zeng, Jing Lin, Ruimao Zhang, Lei Zhang, and Heung-Yeung Shum. Humantomato: Text-aligned whole-body motion generation. ICML, 2024. 2", + "[21] Wenxun Dai, Ling-Hao Chen, Jingbo Wang, Jinpeng Liu, Bo Dai, and Yansong Tang. Motionlcm: Real-time controllable motion generation via latent consistency model. ECCV, 2024.", + "[22] Ling-Hao Chen, Jiawei Zhang, Yewen Li, Yiren Pang, Xiaobo Xia, and Tongliang Liu. Humanmac: Masked motion completion for human motion prediction. In ICCV, pages 9544-9555, 2023.", + "[23] Shunlin Lu, Jingbo Wang, Zeyu Lu, Ling-Hao Chen, Wenxun Dai, Junting Dong, Zhiyang Dou, Bo Dai, and Ruimao Zhang. Scamo: Exploring the scaling law in autoregressive motion generation model. arXiv preprint arXiv:2412.14559, 2024.", + "[24] Wenxun Dai, Ling-Hao Chen, Yufei Huo, Jingbo Wang, Jinpeng Liu, Bo Dai, and Yansong Tang. Real-time controllable motion generation via latent consistency model. 2025. 2", + "[25] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa*, and Jitendra Malik*. Humans in 4D: Reconstructing and tracking humans with transformers. In ICCV, 2023. 2, 3", + "[26] Vickie Ye, Georgios Pavlakos, Jitendra Malik, and Angjoo Kanazawa. Decoupling human and camera motion from videos in the wild. In CVPR, 2023. 2, 3, 7", + "[27] Yufu Wang, Ziyun Wang, Lingjie Liu, and Kostas Daniilidis. Tram: Global trajectory and motion of 3d humans from inthe-wild videos. ECCV, 2024. 2, 3", + "[28] Soyong Shin, Juyong Kim, Eni Halilaj, and Michael J. Black. WHAM: Reconstructing world-grounded humans with accurate 3D motion. In CVPR, 2024. 6, 7, 8" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1981", + "bbox": [ + 483, + 945, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Zehong Shen, Huaijin Pi, Yan Xia, Zhi Cen, Sida Peng, Zechen Hu, Hujun Bao, Ruizhen Hu, and Xiaowei Zhou. World-grounded human motion recovery via gravity-view coordinates. In ACM SIGGRAPH Asia, 2024. 2, 3, 4, 5, 7, 8", + "[30] Angjoo Kanazawa, Jason Y. Zhang, Panna Felsen, and Jitendra Malik. Learning 3d human dynamics from video. In CVPR, 2019. 3", + "[31] Muhammed Kocabas, Ye Yuan, Pavlo Molchanov, Yunrong Guo, Michael J. Black, Otmar Hilliges, Jan Kautz, and Umar Iqbal. Pace: Human and motion estimation from in-the-wild videos. In 3DV, 2024. 2", + "[32] Ling-Hao Chen, Shunlin Lu, Ailing Zeng, Hao Zhang, Benyou Wang, Ruimao Zhang, and Lei Zhang. Motionllm: Understanding human behaviors from human motions and videos. arXiv preprint arXiv:2405.20340, 2024. 2", + "[33] Matthias Plappert, Christian Mandery, and Tamim Asfour. Learning a bidirectional mapping between human whole-body motion and natural language using deep recurrent neural networks. RAS, 109:13-26, 2018. 2", + "[34] Fangzhou Hong, Mingyuan Zhang, Liang Pan, Zhongang Cai, Lei Yang, and Ziwei Liu. Avatarclip: Zero-shot text-driven generation and animation of 3d avatars. ACM SIGGRAPH, 2022. 2", + "[35] Nikos Athanasiou, Mathis Petrovich, Michael J Black, and Gül Varol. Teach: Temporal action composition for 3d humans. In 3DV, pages 414-423, 2022.", + "[36] Rishabh Dabral, Muhammad Hamza Mughal, Vladislav Golyanik, and Christian Theobalt. Mofusion: A framework for denoising-diffusion-based motion synthesis. In CVPR, pages 9760-9770, 2023.", + "[37] Mingyuan Zhang, Xinying Guo, Liang Pan, Zhongang Cai, Fangzhou Hong, Huirong Li, Lei Yang, and Ziwei Liu. Remodiffuse: Retrieval-augmented motion diffusion model. In ICCV, 2023.", + "[38] Weilin Wan, Zhiyang Dou, Taku Komura, Wenping Wang, Dinesh Jayaraman, and Lingjie Liu. Tlcontrol: Trajectory and language control for human motion synthesis. ECCV, 2024.", + "[39] Chuan Guo, Yuxuan Mu, Muhammad Gohar Javed, Sen Wang, and Li Cheng. Momask: Generative masked modeling of 3d human motions. In CVPR, pages 1900-1910, 2024.", + "[40] Jinpeng Liu, Wenxun Dai, Chunyu Wang, Yiji Cheng, Yansong Tang, and Xin Tong. Plan, posture and go: Towards open-world text-to-motion generation. ECCV, 2024.", + "[41] Bo Han, Hao Peng, Minjing Dong, Yi Ren, Yixuan Shen, and Chang Xu. Amd: Autoregressive motion diffusion. In AAAI, pages 2022-2030, 2024.", + "[42] Zhenyu Xie, Yang Wu, Xuehao Gao, Zhongqian Sun, Wei Yang, and Xiaodan Liang. Towards detailed text-to-motion synthesis via basic-to-advanced hierarchical diffusion model. In AAAI, pages 6252–6260, 2024.", + "[43] Wenyang Zhou, Zhiyang Dou, Zeyu Cao, Zhouyingcheng Liao, Jingbo Wang, Wenjia Wang, Yuan Liu, Taku Komura, Wenping Wang, and Lingjie Liu. Emdm: Efficient motion diffusion model for fast, high-quality motion generation. ECCV, 2024." + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[44] Mathis Petrovich, Or Litany, Umar Iqbal, Michael J Black, Gul Varol, Xue Bin Peng, and Davis Rempe. Multi-track timeline control for text-driven 3d human motion generation. In CVPRW, pages 1911-1921, 2024.", + "[45] Boxun Hu, Mingze Xia, Ding Zhao, and Guanlin Wu. Mona: Moving object detection from videos shot by dynamic camera, 2025.", + "[46] Zan Wang, Yixin Chen, Baoxiong Jia, Puhao Li, Jinlu Zhang, Jingze Zhang, Tengyu Liu, Yixin Zhu, Wei Liang, and Siyuan Huang. Move as you say interact as you can: Language-guided human motion generation with scene affordance. In CVPR, pages 433-444, 2024.", + "[47] Zhongfei Qing, Zhongang Cai, Zhitao Yang, and Lei Yang. Story-to-motion: Synthesizing infinite and controllable character animation from long text, 2023. 2", + "[48] Hui En Pang, Zhongang Cai, Lei Yang, Tianwei Zhang, and Ziwei Liu. Benchmarking and analyzing 3d human pose and shape estimation beyond algorithms. In NeurIPS, 2022. 2", + "[49] Gyeongsik Moon, Hongsuk Choi, and Kyoung Mu Lee. Neuralannnot: Neural annotator for 3d human mesh training sets. In CVPRW, 2022.", + "[50] Gyeongsik Moon, Hongsuk Choi, Sanghyuk Chun, Jiyoung Lee, and Sangdoo Yun. Three recipes for better 3d pseudogots of 3d human mesh estimation in the wild. In CVPRW, 2023.", + "[51] Hongwei Yi, Hualin Liang, Yifei Liu, Qiong Cao, Yandong Wen, Timo Bolkart, Dacheng Tao, and Michael J Black. Generating holistic 3d human motion from speech. In CVPR, pages 469-480, 2023. 2", + "[52] C.J. Bowen and R. Thompson. Grammar of the Edit. Focal Press, 2013. 2, 3", + "[53] Mathis Petrovich, Or Litany, Umar Iqbal, Michael J. Black, Gül Varol, Xue Bin Peng, and Davis Rempe. Multi-track timeline control for text-driven 3d human motion generation. In CVPRW, 2024. 2", + "[54] Yuhong Zhang, Jing Lin, Ailing Zeng, Guanlin Wu, Shunlin Lu, Yurong Fu, Yuanhao Cai, Ruimao Zhang, Haoqian Wang, and Lei Zhang. Motion- $\\mathbf{x} + +$ : A large-scale multimodal 3d whole-body human motion dataset. arXiv preprint arXiv:2501.05098, 2025. 2", + "[55] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In CVPR, pages 5152-5161, 2022. 2", + "[56] Georgios Pavlakos, Jitendra Malik, and Angjoo Kanazawa. Human mesh recovery from multiple shots. In CVPR, 2022. 2, 3", + "[57] Peng Wu, Xiankai Lu, Jianbing Shen, and Yilong Yin. Clip fusion with bi-level optimization for human mesh reconstruction from monocular videos. In ACM MM, page 105-115, New York, NY, USA, 2023. Association for Computing Machinery.", + "[58] Kuan-Chieh Wang, Zhenzhen Weng, Maria Xenochristou, Joao Pedro Araujo, Jeffrey Gu, C Karen Liu, and Serena Yeung. Nemo: 3d neural motion fields from multiple video instances of the same action. In CVPR, 2023." + ], + "bbox": [ + 514, + 90, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1982", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[59] Fabien Baradel, Thibault Groueix, Philippe Weinzaepfel, Romain Brégier, Yannis Kalantidis, and Grégory Rogez. Leveraging mocap data for human mesh recovery. In 3DV, pages 586-595, 2021. 2", + "[60] Ruilong Li, Shan Yang, David A. Ross, and Angjoo Kanazawa. Ai choreographer: Music conditioned 3d dance generation with aist++, 2021. 3, 6", + "[61] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE TPAMI, 36(7):1325-1339, 2014. 3, 6, 7", + "[62] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. ACM TOG, 36(6), 2017. 3", + "[63] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J. Black. Keep it SMPL: Automatic estimation of 3D human pose and shape from a single image. In Computer Vision - ECCV 2016. Springer International Publishing, 2016.", + "[64] Anurag* Arnab, Carl* Doersch, and Andrew Zisserman. Exploiting temporal context for 3d human pose estimation in the wild. In CVPR, 2019.", + "[65] Ahmed A. A. Osman, Timo Bolkart, and Michael J. Black. STAR: Sparse Trained Articulated Human Body Regressor, page 598-613. Springer International Publishing, 2020.", + "[66] Yinghao Huang, Federica Bogo, Christoph Lassner, Angjoo Kanazawa, Peter V. Gehler, Javier Romero, Ijaz Akhter, and Michael J. Black. Towards accurate marker-less human shape and pose estimation over time. In 3DV, 2017. 3", + "[67] Angjoo Kanazawa, Michael J. Black, David W. Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In CVPR, 2018. 3", + "[68] Muhammed Kocabas, Nikos Athanasiou, and Michael J. Black. Vibe: Video inference for human body pose and shape estimation. In CVPR, 2020.", + "[69] Nikos Kolotouros, Georgios Pavlakos, and Kostas Dani-ilidis. Convolutional mesh regression for single-image human shape reconstruction. In CVPR, 2019.", + "[70] Zhengyi Luo, S. Alireza Golestaneh, and Kris M. Kitani. 3d human motion estimation via motion compression and refinement. In ACCV, 2020. 3", + "[71] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS. Curran Associates, Inc., 2017. 3", + "[72] Chun-Hao P. Huang, Hongwei Yi, Markus Höschle, Matvey Safroshkin, Tsvetelina Alexiadis, Senya Polikovsky, Daniel Scharstein, and Michael J. Black. Capturing and inferring dense full-body human-scene contact. In CVPR, pages 13274-13285, 2022. 3", + "[73] Manuel Kaufmann, Jie Song, Chen Guo, Kaiyue Shen, Tianjian Jiang, Chengcheng Tang, Juan Jose Zárate, and Otmar Hilliges. EMDB: The Electromagnetic Database of Global 3D Human Pose and Shape in the Wild. In ICCV, 2023. 3, 7, 8", + "[74] Xinyu Yi, Yuxiao Zhou, and Feng Xu. Transpose: Real-time 3d human translation and pose estimation with six inertial sensors. ACM TOG, 40(4), 2021. 3" + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[75] Zachary Teed and Jia Deng. Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. In NeurIPS, pages 16558-16569. Curran Associates, Inc., 2021. 3", + "[76] Zachary Teed, Lahav Lipson, and Jia Deng. Deep patch visual odometry. NeurIPS, 2023. 3, 8", + "[77] Weirong Chen, Le Chen, Rui Wang, and Marc Pollefeys. Leap-vo: Long-term effective any point tracking for visual odometry. In CVPR, 2024. 3, 4, 8", + "[78] Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, and Michael J. Black. AMASS: Archive of motion capture as surface shapes. In ICCV, pages 5442-5451, 2019. 3, 7", + "[79] Timo von Marcard, Roberto Henschel, Michael Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In ECCV, 2018. 3, 7", + "[80] Yufei Xu, Jing Zhang, Qiming Zhang, and Dacheng Tao. ViTPose: Simple vision transformer baselines for human pose estimation. In NeurIPS, 2022. 4", + "[81] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 4, 6", + "[82] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multiperson linear model. ACM TOG, 34(6):248:1-248:16, 2015. 4", + "[83] Qingqiu Huang, Yu Xiong, Anyi Rao, Jiaze Wang, and Dahua Lin. Movienet: A holistic dataset for movie understanding. In ECCV, 2020. 4, 7", + "[84] MMTracking Contributors. MMTracking: OpenMMLab video perception toolbox and benchmark. https://github.com/open-mmlab/mmtracking, 2020.4", + "[85] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. CoTracker: It is better to track together. In ECCV, 2024. 4", + "[86] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. In ICCV, 2023. 5", + "[87] Martin A. Fischler and Robert C. Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Commun. ACM, 24(6):381-395, 1981. 5", + "[88] Jie Yang, Ailing Zeng, Shilong Liu, Feng Li, Ruimao Zhang, and Lei Zhang. Explicit box detection unifies end-to-end multi-person pose estimation. In ICLR, 2023. 6", + "[89] Richard Hartley and Andrew Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, USA, 2 edition, 2003. 6", + "[90] Michael J. Black, Priyanka Patel, Joachim Tesch, and Jinlong Yang. BEDLAM: A synthetic dataset of bodies exhibiting detailed lifelike animated motion. In CVPR, pages 8726-8737, 2023. 7" + ], + "bbox": [ + 516, + 92, + 903, + 883 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "1983", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/HumanMM_ Global Human Motion Recovery from Multi-shot Videos/8ae05aea-0260-4aa4-bbb0-352a7f188bbe_model.json b/2025/HumanMM_ Global Human Motion Recovery from Multi-shot Videos/8ae05aea-0260-4aa4-bbb0-352a7f188bbe_model.json new file mode 100644 index 0000000000000000000000000000000000000000..64cf9c5538db58d26af311c0a893a893e6079b71 --- /dev/null +++ b/2025/HumanMM_ Global Human Motion Recovery from Multi-shot Videos/8ae05aea-0260-4aa4-bbb0-352a7f188bbe_model.json @@ -0,0 +1,2730 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.145, + 0.109, + 0.852, + 0.13 + ], + "angle": 0, + "content": "HumanMM: Global Human Motion Recovery from Multi-shot Videos" + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.159, + 0.867, + 0.196 + ], + "angle": 0, + "content": "Yuhong Zhang\\(^{1,2,\\dagger*}\\) Guanlin Wu\\(^{2,3,\\ddagger*}\\) Ling-Hao Chen\\(^{1,2,\\ddagger}\\) Zhuokai Zhao\\(^{4}\\) Jing Lin\\(^{1,2}\\) Xiaoke Jiang\\(^{2}\\) Jiamin Wu\\(^{2,5}\\) Zhuoheng Li\\(^{6}\\) Hao Frank Yang\\(^{3}\\) Haoqian Wang\\(^{1\\dagger}\\) Lei Zhang\\(^{2\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.139, + 0.197, + 0.86, + 0.213 + ], + "angle": 0, + "content": "\\(^{1}\\)Tsinghua University \\(^{2}\\)IDEA Research \\(^{3}\\)Johns Hopkins University \\(^{4}\\)University of Chicago \\(^{5}\\)HKUST \\(^{6}\\)HKU" + }, + { + "type": "text", + "bbox": [ + 0.293, + 0.216, + 0.708, + 0.23 + ], + "angle": 0, + "content": "{dsyuhong, guanlinwu0930, thu.lhchen}@gmail.com" + }, + { + "type": "text", + "bbox": [ + 0.212, + 0.231, + 0.782, + 0.249 + ], + "angle": 0, + "content": "Project page: https://zhangyuhong01.github.io/HumanMM" + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.259, + 0.898, + 0.621 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.623, + 0.908, + 0.695 + ], + "angle": 0, + "content": "Figure 1. Recovering a human motion from multi-shot videos. Top: We take two multi-shot table tennis game videos with shot transitions as input. We aim to recover two motions of two athletes (Long MA and Zhendong FAN) from two videos, respectively. The first video is recorded by three shots (“①”, “②”, and “③”), and the second one is recovered by two shots (“④” and “⑤”). Bottom: We recover two motions (Long MA in green and Zhendong FAN in pink), different shots, and camera poses for each multi-shot video. The recovered motion is aligned with the motion in the videos." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.7, + 0.327, + 0.715 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.732, + 0.486, + 0.854 + ], + "angle": 0, + "content": "In this paper, we present a novel framework designed to reconstruct long-sequence 3D human motion in the world coordinates from in-the-wild videos with multiple shot transitions. Such long-sequence in-the-wild motions are highly valuable to applications such as motion generation and motion understanding, but are of great challenge to be recovered due to abrupt shot transitions, partial occlusions, and dynamic backgrounds presented in such videos. Existing" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.701, + 0.909, + 0.898 + ], + "angle": 0, + "content": "methods primarily focus on single-shot videos, where continuity is maintained within a single camera view, or simplify multi-shot alignment in camera space only. In this work, we tackle the challenges by integrating an enhanced camera pose estimation with Human Motion Recovery (HMR) by incorporating a shot transition detector and a robust alignment module for accurate pose and orientation continuity across shots. By leveraging a custom motion integrator, we effectively mitigate the problem of foot sliding and ensure temporal consistency in human pose. Extensive evaluations on our created multi-shot dataset from public 3D human datasets demonstrate the robustness of our method in reconstructing realistic human motion in world coordinates." + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.863, + 0.485, + 0.901 + ], + "angle": 0, + "content": "*Equal contribution, ‡Core contributor, †Corresponding author. Work done by Yuhong Zhang, Guanlin Wu, Ling-Hao Chen, Jing Lin and Jiamin Wu during the internship at IDEA Research." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "1973" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.112, + 0.223, + 0.128 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.137, + 0.483, + 0.304 + ], + "angle": 0, + "content": "In recent years, significant advances have been made in 3D human pose estimation, particularly in enhancing the accuracy of human motion recovery (HMR)1 from monocular video sequences. HMR has demonstrated extensive applications in human-AI interaction [1, 2], human motion understanding [3-6], and motion generation [3, 7-24]. While existing methods [25] achieved relatively high performance in recovering human mesh in camera coordinates, estimating human motion in world coordinates remains challenging [26-29] due to inaccurate camera pose estimation and the complexity of reconstructing human motion spatially." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.305, + 0.484, + 0.426 + ], + "angle": 0, + "content": "Most current progress in 3D human motion community mainly benefits from large scale data [25, 27-31], and longsequence videos. These resources enhance estimation accuracy for HMR methods and improve the understanding and generation of longer motion sequences for tasks such as motion understanding [3, 32, 33] and generation [3, 7-20, 34-47], even when annotations are derived from markerless capturing methods like pseudo labels [48-51]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.426, + 0.484, + 0.562 + ], + "angle": 0, + "content": "A promising approach to enlarge the scale of the motion databases is to estimate human motions from unlimited online videos in a markerless manner. However, many longsequence online videos are recorded with multiple shots, referred to as multi-shot videos\\(^2\\), especially prevalent in domains such as sports broadcasting, talk shows, and concerts. In filmmaking and television live show, a \"shot\" denotes an individual camera view capturing a specific moment or action from a particular vantage point [52]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.563, + 0.483, + 0.713 + ], + "angle": 0, + "content": "Segmenting multi-shot videos into separate shots inevitably reduces the length of the video sequences, which can be detrimental to tasks that benefit from longer sequences, such as long motion generation [47, 53]. This limitation is highlighted in the existing datasets [54, 55], where the longest clip is less than 20 seconds after segmentation, as shown in Fig. 2. Moreover, focusing exclusively on online single-shot videos diminishes the utilization ratio of available online videos and may negatively impact the diversity of scenarios represented in the created datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.713, + 0.483, + 0.806 + ], + "angle": 0, + "content": "Therefore, how to address the issue of discontinuities caused by shot transitions is notoriously difficult in the community. To resolve this problem, previous works [56-59] have proposed algorithms to address human mesh recovery in a camera space from movies containing shot change between long shots and close-ups." + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.09, + 0.895, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.238, + 0.907, + 0.307 + ], + "angle": 0, + "content": "Figure 2. The comparison between the distribution of sequence lengths in different existing large-scale markerless motion datasets with ours. The \\( x \\)-axis and \\( y \\)-axis denote the duration time (s) and percentage of video number, respectively. Our dataset (in green) contains more portion of long-sequence videos in general." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.311, + 0.907, + 0.599 + ], + "angle": 0, + "content": "However, recovering human motions in world coordinates from multi-shot videos presents two fundamental challenges that remain underexplored. 1) How to align the human motion and orientation in the world coordinates during shot transitions? Ensuring continuity of human orientation and pose across shots is complicated by factors such as partial visibility of human body (e.g. transitioning from long shot to close-up) and changes in human orientation (e.g. two long shots from different viewpoints). These issues, caused by abrupt changes in camera viewpoints, necessitate robust alignment mechanisms. 2) How to reconstruct accurate human motion in world coordinates? Existing approaches employ Simultaneous Localization and Mapping (SLAM) methods to estimate camera parameters, which are then used to project recovered human meshes from camera to world coordinates [26-29]. This process requires highly accurate camera estimation and must address motion consistency and foot sliding in the recovered human motion within the world space." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.906, + 0.689 + ], + "angle": 0, + "content": "Despite these challenges, human motion in multi-shot videos often remain continuous across shots, even as camera viewpoints change. This observation suggests that with appropriate handling of shot transitions and camera motion, it is possible to reconstruct consistent and complete 3D human motions throughout multi-shot videos." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.909, + 0.902 + ], + "angle": 0, + "content": "In this paper, we propose a novel framework HumanMM, Human Motion recovery from Multi-shot videos, to address these challenges. It integrates human pose estimation across shots with robust camera estimation in the world space. Firstly, we develop a shot transition detector to identify frames with shot transitions. To ensure a more robust camera pose estimation, we introduce an enhanced SLAM method incorporating long-term tracking of feature points and exclusion of moving human from bundle adjustment process. We utilize existing HMR method integrated with our enhanced camera estimation to get the initial human parameters for each separated shot. Subsequently, we implement an alignment module to align human orientation based on stereo calibration and smooth human poses through a" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.815, + 0.483, + 0.864 + ], + "angle": 0, + "content": "1In this paper, the \"human mesh recovery\" refers to recovery in the camera coordinates and the \"human motion recovery\" denotes recovery in the world coordinates. Unless specified otherwise, HMR refers to human motion recovery." + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.864, + 0.483, + 0.9 + ], + "angle": 0, + "content": "2In this paper, a multi-shot video refers to a long-sequence video containing multiple shot transitions. We assume that the camera intrinsics remain consistent across different shots within a multi-shot video." + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.815, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1974" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.197 + ], + "angle": 0, + "content": "trained multi-shot HMR encoder, which effectively captures the temporal context of human movements across different shots. Finally, after aligning human and camera parameters between shot transitions, we train a motion decoder and a trajectory refiner to smooth the human pose and mitigate issues such as foot sliding, thereby enhancing the overall motion consistency in the reconstructed 3D human motions." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.199, + 0.435, + 0.212 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.215, + 0.483, + 0.243 + ], + "angle": 0, + "content": "- We present the first approach to reconstruct human motion from multi-shot videos in world coordinates." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.245, + 0.484, + 0.333 + ], + "angle": 0, + "content": "- We introduce HumanMM, a HMR framework for multi-shot videos. It includes an enhanced camera trajectory estimation method, a human motion alignment module and a motion integrator to ensure accurate and consistent recovery of human pose and orientation in world coordinates across different shots in the whole video." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.335, + 0.484, + 0.41 + ], + "angle": 0, + "content": "- We develop a multi-shot video dataset ms-Motion to evaluate the performance of HMR from multi-shot videos, based on existing public datasets such as AIST [60] and Human3.6M [61]. Extensive experiments on related benchmarks verify the effectiveness of our method." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.215, + 0.484, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.427, + 0.233, + 0.442 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.453, + 0.336, + 0.467 + ], + "angle": 0, + "content": "2.1. HMR from One-shot Video" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.477, + 0.483, + 0.522 + ], + "angle": 0, + "content": "One-shot videos, captured with a single camera without shot transitions, has been extensively studied within the community for human mesh and motion recovery." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.523, + 0.483, + 0.612 + ], + "angle": 0, + "content": "Human mesh recovery in camera coordinates can be broadly categorized into two approaches: optimization-based methods [62-66] and regression-based methods [30, 67-70]. With the significant advancements of transformer [71], HMR2.0 [25] has surpassed previous methods and benefits several downstream tasks related to HMR." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Although there are several previous works tried to recover motions in world coordinates with multi-camera capture system [60, 72] and IMU-based methods [73, 74] and enjoy relatively satisfying results, this setup limits their use for applications of infinite in-the-wild monocular videos. To address this limitation, several attempts [26-29] integrate SLAM into the HMR pipeline by first estimating the camera pose using SLAM methods, e.g. DROID-SLAM [75] or DPVO [76], and then project the recovered human motion from camera to world coordinates. To exclude the inconsistencies caused by dynamic objects, such as moving humans, TRAM [27] modifies DROID-SLAM by incorporating human masking and depth-based distance rescaling. However, DROID-SLAM performs dense bundle adjustment (DBA) on feature maps from downsampled images and selects features based only on two consecutive frames rather than long-term video sequences [75-77]. Consequently, masking significantly reduces the number of informative and consistent features, especially when humans occupy large" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.138 + ], + "angle": 0, + "content": "portions of the image, leading to inaccuracies. Therefore, developing a SLAM method that retains sufficient and representative features for DBA after masking is important." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.147, + 0.771, + 0.162 + ], + "angle": 0, + "content": "2.2. HMR from Multi-shot Video" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.17, + 0.905, + 0.274 + ], + "angle": 0, + "content": "Multiple shots are fundamental elements of cinematic storytelling and live performances, utilizing various camera positions and focal lengths to create immersive and detailed viewing experiences for audiences [52]. However, most marker-based motion capture (MoCap) datasets [60, 72, 73, 78, 79] consist single-shot videos only, resulting in limited research on HMR from multi-shot videos." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.275, + 0.906, + 0.44 + ], + "angle": 0, + "content": "Recovering human motion from multi-shot videos in camera coordinates is already challenging. This is because treating each pose estimation result of each shot separately leads to inconsistencies when combining all estimations, caused by partially or fully invisible human bodies across shot transitions. Pavlakos et al. [56] addresses this issue by focusing on shot changes from long shots to close-ups, which are common in film. They develop smoothness constraints within a temporal Human Mesh and Motion Recovery (t-HMMR) model to infer motions during occlusions caused by shot transitions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.441, + 0.906, + 0.575 + ], + "angle": 0, + "content": "Advancements in HMR methods [29] for single-shot videos in world coordinates have paved the way for extending HMR to multi-shot videos with varying camera viewpoints. However, aligning human orientation, body pose, and translation continuously across multi-shot videos in world coordinates underexplored. Effective alignment is crucial to maintain motion continuity and coherence, especially when dealing with diverse camera perspectives and abrupt transitions between shots." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.576, + 0.906, + 0.68 + ], + "angle": 0, + "content": "In summary, while substantial progress has been made in HMR from single-shot videos, extending these techniques to multi-shot videos requires addressing additional complexities related to camera pose alignment and motion consistency across shot transitions. We address this challenge by proposing a novel pipeline that ensures accurate and continuous 3D HMR from multi-shot monocular videos." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.695, + 0.605, + 0.71 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this section, we propose HumanMM to recover human motion from multi-shot videos. The system overview is shown in Fig. 3. Given an input video sequence \\(\\mathbf{V} = \\{I_t\\}_{t=1}^T\\) of length \\(T\\), where \\(I_t\\) denotes the \\(t\\)-th frame, our objective is to recover human motion in world coordinates. We begin by detecting shot transition frames based on human bounding box (a.k.a. bbox) and 2D keypoints (a.k.a. KPTs) through a shot transition detector (Sec. 3.2). For each clipped shot, we initialize the camera pose (camera rotation and camera translation) and recover initial human motion in world coordinates (Sec. 3.3). The initialized SMPL parameters and camera poses are then fed into a hu" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1975" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.092, + 0.9, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.283, + 0.907, + 0.367 + ], + "angle": 0, + "content": "Figure 3. The overview of HumanMM. HumanMM processes multi-shot video sequences by first extracting motion feature such as keypoints and bounding boxes, using ViTPose [80] and image feature using ViT [81]. These features are then segmented into single-shot clips via Shot Transition Detection (Sec. 3.2). Initialized camera (camera rotation \\(\\mathbf{R}\\) and camera translation \\(\\mathbf{T}\\)) and human (SMPL) parameters for each shot are estimated using Masked LEAP-VO (Sec. 3.3) and GVHMR [29]. Human orientation is aligned across shots through camera calibration (3.4.1), and ms-HMR (Sec. 3.4.2) ensures consistent pose alignment. Finally, a bi-directional LSTM-based trajectory predictor with trajectory refiner predicts trajectory based on aligned motion and mitigates foot sliding throughout the video." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.37, + 0.484, + 0.476 + ], + "angle": 0, + "content": "man motion alignment module (Sec. 3.4), which aligns human orientations via camera calibration based on human 2D KPTs and smooth the human pose by incorporating pose information across different shots. Additionally, it refines the entire motion sequence through whole video using a temporal motion encoder ms-HMR. Finally, we introduce a post-processing module for motion integration (Sec. 3.5)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.484, + 0.373, + 0.5 + ], + "angle": 0, + "content": "3.1. Preliminary: 3D Human Model" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.507, + 0.484, + 0.599 + ], + "angle": 0, + "content": "Our method aims to recover motions in world coordinates in the SMPL [82] format, whose pose at frame \\(t\\) can be represented as \\(\\mathcal{M}_t(\\theta_t,\\beta_t,\\Gamma_t,\\tau_t)\\in \\mathbb{R}^{6890\\times 3}\\). Here, the body pose, body shape, root orientation, and translation are \\(\\theta_{t}\\in \\mathbb{R}^{23\\times 3}\\), \\(\\beta_{t}\\in \\mathbb{R}^{10}\\), \\(\\Gamma_t\\in \\mathbb{R}^3\\), and \\(\\tau_{t}\\in \\mathbb{R}^{3}\\), respectively. We use \\(\\mathbf{K}_t^{2D}\\) to denote human 2D KPTs at each frame \\(t\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.606, + 0.484, + 0.621 + ], + "angle": 0, + "content": "3.2. Shot Transition Detector For Multi-shot Video" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Our algorithm begins with shot transition detection in one video. As shown in Fig. 3, the shot transition detector has three key components, scene transition detector, bounding box (a.k.a. bounding box) tracking, and human keypoints tracking. (1) Scene change transition detector: Initially, we employ the SceneDetect [83] algorithm to identify scene changes based on significant variations in the background. However, the SceneDetect fails to detect shot transitions when background changes are unnoticeable, illustrated in Fig. 4. Subsequently, we leverage the following modules to bridge the gap. (2) Bbox tracking for shot transition. As a shot change often accompanies with a sudden change of human subject size, we track humans in a video via mmtracking [84]. Consequently, we compute the Intersection over Union (IoU) between neighbor bboxes and identify a shot transition when the IoU falls smaller than a manually tuned threshold. (3) Human pose tracking for shot transition detection. To achieve a finer granularity, we additionally intro" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.37, + 0.653, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.544, + 0.48, + 0.63, + 0.491 + ], + "angle": 0, + "content": "(a) Scene Change" + }, + { + "type": "image", + "bbox": [ + 0.654, + 0.37, + 0.778, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.665, + 0.48, + 0.762, + 0.491 + ], + "angle": 0, + "content": "(b) Position Change" + }, + { + "type": "image", + "bbox": [ + 0.781, + 0.37, + 0.901, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.797, + 0.48, + 0.876, + 0.491 + ], + "angle": 0, + "content": "(c) Pose Change" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.491, + 0.906, + 0.587 + ], + "angle": 0, + "content": "Figure 4. Shot transition detection examples. Examples (a), (b), and (c) illustrate multi-shot scenarios in online videos. (a) shows scene transitions detectable by SceneDetect. (b) illustrates significant position changes undetectable by SceneDetect but resolvable withbbox tracking-based method. (c) shows pose or orientation transition, requiring pose tracking-based methods as they cannot be addressed by either SceneDetect or_bbox tracking." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.593, + 0.906, + 0.652 + ], + "angle": 0, + "content": "duce human 2D KPTs to detect extreme corner shot changes in a video. By thresholding the IoU of corresponding keypoints between neighbor frames, we can accurately identify shot transitions even with subtle human movements." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.654, + 0.906, + 0.7 + ], + "angle": 0, + "content": "As each separate module cannot identify all kinds of shot transitions, the three modules are jointly used to clip a video into several sub-sequences serially." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.71, + 0.906, + 0.74 + ], + "angle": 0, + "content": "3.3. Human Motion and Camera Pose Estimation For Each Shot" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.906, + 0.825 + ], + "angle": 0, + "content": "After obtaining the clipped videos, our next goal is to estimate the camera pose and SMPL parameters in the world coordinates for each clipped video. The estimated camera pose and motions for each shot will be used to construct the whole motion sequence in the next stage (Sec. 3.4)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.901 + ], + "angle": 0, + "content": "How to estimate the camera parameters accurately? Our approach for camera parameter calculation is based on a visual odometry (VO) estimation method, LEAP-VO [77]. Utilizing the CoTracker method [85], LEAP-VO estimates the visibility and trajectories of \\(N\\) selected points by ana" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1976" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "lyzing image gradients across the video sequence. LEAPVO subsequently computes confidence scores for each trajectory, retaining only those with high confidence while discarding trajectories shorter than a predefined threshold. The remaining trajectories undergo bundle adjustment (BA) within a fixed window size to estimate the camera poses." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.183, + 0.485, + 0.484 + ], + "angle": 0, + "content": "However, simply applying LEAP-VO in the camera estimation process is still unsatisfactory in most human-centric scenarios. The primary limitation stems from the dynamic movements of human subjects, which typically occupy a substantial portion of each image in human-centric videos. This dynamic presence introduces noise into the camera pose estimation in world coordinates, as the estimation process relies heavily on the relationship between the camera and the static environment. To address this issue, we propose a Masked LEAP-VO algorithm. Our approach involves inputting the image \\(I_{t}\\) and the human bbox at frame \\(t\\) into SAM [86] to generate a human mask. We then assign a visibility value of zero to points within the human mask, effectively excluding these trajectories from the BA process. For clarity, we denote \\(S_{BA}\\) as the window size of BA, \\(\\hat{n}\\) denotes the number of filtered point trajectories, and \\(w_{ij,\\hat{n}}\\) as the normalized weight based on confidence score and visibility. For estimating the camera poses \\(\\mathbf{G} = \\{\\mathbf{R},\\mathbf{T}\\}\\) of orientation and translation, the reprojection loss function for BA can then be formulated as follows," + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.491, + 0.497, + 0.517 + ], + "angle": 0, + "content": "\\[\n\\mathbf {G} = \\underset {\\mathbf {G}, d _ {i}, \\hat {n}} {\\arg \\min } \\sum_ {i} \\sum_ {j \\in | i - j | \\leq S} \\sum_ {B A} w _ {i j, \\hat {n}} | | \\mathcal {F} (\\mathbf {G} _ {i}, \\mathbf {G} _ {j}, d _ {i, \\hat {n}}) - \\Pi_ {i j} (\\mathbf {p} _ {i, \\hat {n}}) | |,\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.524, + 0.483, + 0.6 + ], + "angle": 0, + "content": "where \\(\\mathcal{F}(\\mathbf{G}_i,\\mathbf{G}_j,d_{i,\\hat{n}})\\) denotes the point positions calculated by camera pose \\(\\mathbf{G}\\) at frame \\(i\\) and \\(j\\) with depth \\(d_{i,\\hat{n}}\\). \\(\\Pi_{ij}(\\mathbf{p}_{i,\\hat{n}})\\) denotes the position for project position of \\(\\mathbf{p}_{i,\\hat{n}}\\) from frame \\(i\\) to \\(j\\). Consequently, we obtain the camera rotation \\(\\mathbf{R}_t\\) and translation \\(\\mathbf{T}_t\\) from camera pose \\(\\mathbf{G}_t\\) at \\(t\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.601, + 0.483, + 0.661 + ], + "angle": 0, + "content": "Recovering human motion in world coordinates with estimated camera parameters. Given an input video, we feed the estimated camera parameters \\((\\mathbf{R}_t\\) and \\(\\mathbf{T}_t)\\) into the state-of-the-art motion recovering model, GVHMR [29]," + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.667, + 0.482, + 0.684 + ], + "angle": 0, + "content": "\\[\n\\theta_ {t} ^ {w}, \\beta_ {t} ^ {w}, \\Gamma_ {t} ^ {w}, \\tau_ {t} ^ {w} = \\mathrm {G V H M R} \\left(I _ {t}, \\mathbf {R} _ {t}, \\mathbf {T} _ {t}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.686, + 0.483, + 0.717 + ], + "angle": 0, + "content": " Initialized human parameters \\(\\theta_t^w,\\beta_t^w,\\Gamma_t^w,\\tau_t^w\\) and camera parameters \\(\\mathbf{R}_t,\\mathbf{T}_t\\) will input to human motion alignment." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.727, + 0.433, + 0.743 + ], + "angle": 0, + "content": "3.4. Aligning Human Motion Between Shots" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Based on initialized world motion for each individual shot, the subsequent question is how to merge discontinuous motions from different shots into a continuous motion sequence as a whole in world coordinates. A straightforward solution is to align all motion sequences to the world coordinate system of the first shot. However, finding the correspondence between different shots is still under-explored and challenging. To resolve this issue, we decompose the motion parameters into camera-dependent and camera-independent ones. The former (Sec. 3.4.1) achieves alignment between shots" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.091, + 0.897, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.228, + 0.907, + 0.329 + ], + "angle": 0, + "content": "Figure 5. Human orientation alignment module. Following a shot transition after the foremost purple human mesh (shot \\(①\\) captured by camera \\(C_0\\)), the unaligned (blue) and aligned (green) motions are captured as shot \\(②\\) and shot “\\(③\\)” by camera \\(C_0'\\) and \\(C_1\\), respectively. \\(C_0' = C_0\\). To achieve human orientation alignment from shot “\\(①\\)” to “\\(③\\)”, the camera rotation matrix from \\(C_0'\\) to \\(C_1\\) is computed and applied as the offset of human orientation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.334, + 0.906, + 0.41 + ], + "angle": 0, + "content": "via human orientation alignment based on camera calibration, whereas the latter (Sec. 3.4.2) is a trainable module to enhance the continuity of human motion sequence. These two key designs ensure a consistent motion sequence between frames when encountering shot transitions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.42, + 0.882, + 0.435 + ], + "angle": 0, + "content": "3.4.1 Aligning Human Orientations Between Shots" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.443, + 0.907, + 0.626 + ], + "angle": 0, + "content": "After obtaining the initial SMPL and camera parameters \\(\\{\\theta_t^i,\\beta_t^i,\\Gamma_t^i,\\tau_t^i,\\mathbf{R}_t^i,\\mathbf{T}_t^i\\}\\) for each shot, directly concatenating motions between shots result abrupt changes of human poses and orientations. To address this issue, we introduce the Orientation Alignment Module (OAM), as shown in Fig. 5, to align human orientations. As the whole motion sequence is continuous, we have the following assumption. Assumption 1 Human orientations and translations during the shot transition in world coordinates are continuous. To align the orientations between two frames with shot transition under Assumption 1, we decompose the human orientation with shot transitions in world coordinates as," + }, + { + "type": "equation", + "bbox": [ + 0.62, + 0.628, + 0.905, + 0.644 + ], + "angle": 0, + "content": "\\[\n\\mathrm {R} \\left(\\Gamma_ {\\text {w o r l d}}\\right) = \\mathrm {R} _ {\\delta_ {\\text {c a m}}} \\mathrm {R} \\left(\\Gamma_ {\\text {v i e w}}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.907, + 0.765 + ], + "angle": 0, + "content": "where \\(\\mathbf{R}_{\\delta_{\\mathrm{cam}}}\\) represents the camera rotation on the Y-axis between current \\(t\\)-th and previous \\(t - 1\\)-th frame, \\(\\Gamma_{\\mathrm{view}}\\) denotes the human orientation estimated by the current shot, and \\(\\mathbb{R}(\\cdot):\\mathbb{R}^3\\to \\mathbb{R}^9\\) is the mapping from axis angle to rotation matrix. As \\(\\Gamma_{\\mathrm{view}}\\) in current shot can be estimated independently, mentioned in Sec. 3.3, obtaining accurate \\(\\Gamma_{\\mathrm{world}}\\) in Eq. (2) remains a key challenge to estimate the relative camera rotation \\(\\mathbf{R}_{\\delta_{\\mathrm{cam}}}\\) between frames in shot transitions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Estimating the relative camera pose \\(\\mathbf{R}_{\\delta_{\\mathrm{cam}}}\\) between transition frames. Different from our approach of estimating camera pose in each shot (Sec. 3.3), we do not mask the human subject when estimating camera rotation \\(\\mathbf{R}_{\\delta_{\\mathrm{cam}}}\\). Instead, we use human 2D KPTs as explicit feature matching. Specifically, we filter out unmatched keypoints based on their visibility and unaligned direction using RANSAC [87], effectively addressing camera pose estimation during shot transitions. This procedure is referred" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1977" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.121 + ], + "angle": 0, + "content": "to as Camera Calibration (a.k.a. epipolar-geometry-based camera extrinsic estimation), and is detailed below." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.122, + 0.484, + 0.348 + ], + "angle": 0, + "content": "In Camera Calibration, we assume that the human translations remain unchanged across the shot transition, implying that only the camera's orientation changes (i.e. Assumption 1). Consequently, we calculate the orientation offset by determining the change in camera orientation using camera calibration. We begin by extracting human 2D KPTs from two consecutive frames during the shot transition. Due to the shot transition, the visibility of 2D KPTs may vary, e.g. occlusion in some shots. Therefore, we employ ED-Pose [88] to filter out invisible 2D KPTs between shot transition frames. Subsequently, RANSAC identifies matching 2D KPTs corresponding to the most possible camera rotation direction. These matched 2D KPTs facilitate the estimation of the aligned camera rotation \\(\\mathbf{R}_{\\delta_{\\mathrm{cam}}}\\). The detailed estimation process is as follows." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.349, + 0.484, + 0.442 + ], + "angle": 0, + "content": "We denote the detected 2D KPTs of two frames in the shot transition as \\(\\mathbf{S}_1 = [(x_1^{(1)},y_1^{(1)}),(x_1^{(2)},y_1^{(2)}),\\dots ,(x_1^{(N)},y_1^{(N)})]^{\\top}\\in \\mathbb{R}^{2\\times N}\\) and \\(\\mathbf{S}_2 = [(x_2^{(1)},y_2^{(1)}),(x_2^{(2)},y_2^{(2)}),\\dots ,(x_2^{(N)},y_2^{(N)})]^{\\top}\\in\\) \\(\\mathbb{R}^{2\\times N}\\). The essential matrix \\(\\mathbf{E} = [\\mathbf{T}]_{\\times}\\mathbf{R}\\) should satisfy the following orthogonal property such that," + }, + { + "type": "equation", + "bbox": [ + 0.24, + 0.442, + 0.482, + 0.462 + ], + "angle": 0, + "content": "\\[\n\\mathbf {S} _ {1} ^ {\\top} \\mathbf {E} \\mathbf {S} _ {2} = \\mathbf {0}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.472, + 0.483, + 0.532 + ], + "angle": 0, + "content": "Once \\(\\mathbf{E}\\) is obtained by solving Eq. (3), we enforce the rank-2 constraint on \\(\\mathbf{E}\\) through SVD decomposition and subsequently derive the aligned camera rotation \\(\\mathbf{R}_{\\delta_{\\mathrm{cam}}}\\) between two frames (cf. Hartley et al. [89] for more details)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.533, + 0.484, + 0.595 + ], + "angle": 0, + "content": "In summary, we reformulate the alignment problem of human orientation in shot transitions as estimating the relative camera rotation \\(\\mathbf{R}_{\\delta_{\\mathrm{cam}}}\\) between frames. Accordingly, we obtain the camera rotation \\(\\mathbf{R}_{\\delta_{\\mathrm{cam}}}\\) via camera calibration." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.606, + 0.409, + 0.621 + ], + "angle": 0, + "content": "3.4.2 Aligning Human Poses Between Shots" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.63, + 0.484, + 0.766 + ], + "angle": 0, + "content": "In shot transition, video sequences recorded by two shots are often with various occlusions. However, unoccluded body parts in two shots can be complementary to each other for motion alignment. Thus, we introduce the multi-shot HMR (ms-HMR, i.e. \\(\\mathsf{E}_M(\\cdot)\\)) module to refine the whole motion sequence. As shown in Fig. 6, the ms-HMR is a Transformer encoder-like architecture, whose input and output are the estimated global motion and the refined global motion, respectively. The process can be formulated as," + }, + { + "type": "equation", + "bbox": [ + 0.156, + 0.771, + 0.483, + 0.787 + ], + "angle": 0, + "content": "\\[\n\\phi_ {1}, \\phi_ {2}, \\dots , \\phi_ {T} = \\mathrm {E} _ {M} \\left(\\theta_ {1}, \\theta_ {2}, \\dots , \\theta_ {T}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.792, + 0.484, + 0.839 + ], + "angle": 0, + "content": "where \\(\\phi_*\\) denotes the refined motion of each frame. With this design, our method can adapt to diverse occlusions of human body brought by shot transitions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.848, + 0.483, + 0.865 + ], + "angle": 0, + "content": "3.5. Post-processing Module for Motion Integration" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Trajectory Predictor and Trajectory Refiner. Based on the aligned human pose and orientation, we introduce a" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.091, + 0.915, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.333, + 0.907, + 0.418 + ], + "angle": 0, + "content": "Figure 6. ms-HMR Structure. The initial human pose parameters \\(\\theta\\) across multiple video shots are input into a transformer with shot-index-based positional encoding. This enables ms-HMR to generate consistent human poses across all shots in the video. bi-directional LSTM trajectory predictor to recover foot-ground contact probabilities \\(p_t^c\\), and root velocity \\(v_t\\) as," + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.419, + 0.905, + 0.444 + ], + "angle": 0, + "content": "\\[\np _ {t} ^ {c}, v _ {t} = \\operatorname {L S T M} \\left(\\phi_ {1} ^ {m}, \\Gamma_ {1}, \\mathbb {F} (I _ {1}), \\phi_ {2} ^ {m}, \\Gamma_ {2}, \\mathbb {F} (I _ {2}), \\dots , \\right. \\\\ \\left. \\quad \\left(m, \\mathbb {F} - \\mathbb {F} (I _ {m})\\right)\\right) \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.753, + 0.438, + 0.862, + 0.454 + ], + "angle": 0, + "content": "\\[\n\\phi_ {T} ^ {m}, \\Gamma_ {T}, F (I _ {T})),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.454, + 0.906, + 0.545 + ], + "angle": 0, + "content": "where \\( \\mathsf{F}(\\cdot) \\) denotes the image feature of each frame extracted by ViT [81]. Accordingly, the contact probabilities \\( p_t^c \\), and velocity \\( v_t \\) are supervised by the ground-truth labels with MSE loss and are used to reconstruct the trajectory. Besides, we extend the trajectory refiner in WHAM [28] to alleviate foot sliding problem in our estimated trajectory." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.55, + 0.904, + 0.567 + ], + "angle": 0, + "content": "4. Benchmarking Multi-shot Motion Recovery" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.907, + 0.87 + ], + "angle": 0, + "content": "Dataset Construction. To create a multi-shot 3D human motion dataset, we introduce ms-Motion by processing existing public 3D human datasets with multiple camera settings and ground truth human and camera parameters, specifically AIST [60] and Human3.6M (H3.6M) [61]. In our construction pipeline, we randomly separate each original one-shot video into several clips. Then, we choose each clip from different shots and concatenate them together as one video recorded by multiple shots. For example, AIST provides each video with eight cameras C0, C1, ..., C7 from different view point and we choose a video and split it into 5 clips at t0, t1, ..., t4. For frames in these separated clips, we choose frames shot by a random camera for each clip and combine five clips as one multi-shot video. Therefore, we construct a multi-shot version of AIST and H3.6M, which are named ms-AIST and ms-H3.6M subsets. Then we combine them and name this new dataset ms-Motion. The detailed statistics of ms-Motion are shown in Tab. 2. We do not compare with other existing 3D human datasets as they contain limited number of multi-shot videos." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Benchmark Evaluation Protocol. To evaluate the performance of our proposed methods on multi-shot videos, our" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1978" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.905, + 0.215 + ], + "angle": 0, + "content": "
DatasetModels2-Shot3-Shot4-Shot
PA.↓WA.↓RTE↓ROE↓Jitter↓F.S.↓PA.↓WA.↓RTE↓ROE↓Jitter↓F.S.↓PA↓WA↓RTE↓ROE↓Jitter↓F.S.↓
ms-AISTSLAHMR [2023]72.34341.759.6296.2662.593.2680.35510.7710.33101.3672.394.4390.32803.6912.11104.0780.3716.52
WHAM [2024]65.34336.824.3984.4825.242.7578.68451.325.1489.8424.062.99102.88603.935.5790.0726.293.62
GVHMR [2024]60.72231.366.2096.5834.877.6570.33357.167.5599.6934.469.4283.77563.178.96104.5335.679.78
Ours36.82121.352.5669.2333.272.6638.52141.383.6467.7135.073.5539.63161.524.5570.3139.494.09
ms-H3.6MSLAHMR [2023]80.67352.6116.67111.9737.807.9397.15562.1016.91118.4652.239.96107.90748.5817.85116.7265.1511.58
WHAM [2024]71.32313.5811.4182.4218.405.0979.51423.9812.3684.8518.875.0390.50512.6612.9190.3418.405.69
GVHMR [2024]64.63254.306.9481.9318.458.8080.79296.7485.2558.2618.3610.6285.19471.539.1291.6319.4710.65
Ours40.52132.133.6553.3919.054.1745.35145.365.3358.2617.354.6250.59147.626.2061.2219.775.12
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.217, + 0.905, + 0.257 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison of different HMR methods on ms-Motion dataset. We record the results for ms-AIST and ms-H3.6M separately. PA and WA means PA-MPJPE and WA-MPJPE respectively, while F.S. is Foot Sliding. Our proposed method has achieved the best performance in PA-MPJPE, WA-MPJPE, RTE and ROE across ms-Motion among these methods." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.26, + 0.481, + 0.297 + ], + "angle": 0, + "content": "
DatasetDuration(s)VideosFPSMax LengthMin LengthShots
ms-Motion23.76003014783142, 3, 4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.297, + 0.483, + 0.325 + ], + "angle": 0, + "content": "Table 2. Statistics of the ms-Motion dataset. By shots, we mean the number of shot transitions in a single video." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.333, + 0.484, + 0.514 + ], + "angle": 0, + "content": "target is to evaluate metrics for accurately reflecting the performance on videos with shot transitions. To this end, we use Root Orientation Error (a.k.a. ROE in deg) to measure the performance of the proposed method on human orientation alignment across different shots. Besides, we use Root Translation Error (a.k.a. RTE in \\( m \\)) to assess the performance of the proposed method on global trajectory recovery. Jitter \\( \\left( \\frac{10m}{fps^3} \\right) \\) is also used to evaluate the stability of recovered human pose from multi-shot videos. We also include foot sliding \\( (cm) \\), the averaged displacement of foot vertices during contact with the ground, to assess the precision of recovered motion in the world coordinates [28]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.532, + 0.216, + 0.549 + ], + "angle": 0, + "content": "5. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.557, + 0.29, + 0.572 + ], + "angle": 0, + "content": "5.1. Datasets and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.581, + 0.483, + 0.732 + ], + "angle": 0, + "content": "Evaluation Datasets. To evaluate the performance of our proposed pipeline for multi-shot videos, we use ms-Motion dataset and EMDB-1 dataset [73] with self-added noise for the evaluation of ablation study. For camera trajectory estimation, we use EMDB-1 and EMDB-2 split [73] as they contain the GT moving camera trajectory. Our self-created dataset contains 600 multi-shot videos, 42.7K frames, totaling 237 minutes. EMDB-1 split contains 17 video sequences totaling 13.5 minutes and EMDB-2 split contains 25 sequences in a total of 24.0 minutes." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.735, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Evaluation Metrics. For shot detection we use Recall, Precision and F1 Score as evaluation metrics. For 3D human pose estimation-related tasks, we use ROE, RTE, jitter, and foot-sliding for evaluating the human motion recovery results on multi-shot videos. For the ablation study of our proposed pipeline, we evaluate the Procrustes-aligned Mean Per Joint Position Error (a.k.a. PA-MPJPE) and Per Vertex Error (a.k.a. PVE) as additional metrics besides previous mentioned ones. For camera pose estimation, we use absolute trajectory error (a.k.a. ATE) \\((m)\\), Relative Pose Error (a.k.a. RPE) rotation \\((deg)\\), and RPE translation \\((m)\\)." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.26, + 0.704, + 0.307 + ], + "angle": 0, + "content": "
Methodsms-Motion
Recall↑Precision↑F1 Score↑
Scenes Detect (SD) [83]0.740.720.70
SD+Bbox Tracking (Bbox)0.880.850.86
SD+Bbox+Pose Tracking0.960.880.92
" + }, + { + "type": "table_caption", + "bbox": [ + 0.514, + 0.308, + 0.705, + 0.362 + ], + "angle": 0, + "content": "Table 3. Comparison between difference shot detection algorithms. We evaluate our shot transition detector on our proposed multi-shot video human motion dataset ms-Motion." + }, + { + "type": "table", + "bbox": [ + 0.714, + 0.262, + 0.898, + 0.305 + ], + "angle": 0, + "content": "
MethodsATE↓RPE Trans.↓RPE Rot.↓
DPVO (w/o mask)0.481.851.06
Masked DPVO0.481.570.97
LEAP-VO (w/o mask)0.500.930.97
Masked LEAP-VO0.510.920.95
" + }, + { + "type": "table_caption", + "bbox": [ + 0.71, + 0.306, + 0.9, + 0.359 + ], + "angle": 0, + "content": "Table 4. Camera tracking results on EMDB-1 [73]. Our method has achieved \\(\\sim 50\\%\\) ↓ on RPE trans. than that of the original DPVO and perform the best in RPE rot. metrics." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.366, + 0.731, + 0.381 + ], + "angle": 0, + "content": "5.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.386, + 0.907, + 0.568 + ], + "angle": 0, + "content": "The ms-HMR, the trajectory, and foot sliding refiner are trained on the AMASS [78], 3DPW [79], Human3.6M [61], and BEDLAM [90] datasets, and evaluated on EMDB and our ms-Motion. During training, we introduce random rotational noise (ranging from 0 to 1 radian) along the y-axis to the root pose \\(\\Gamma\\) and random noise to the body pose \\(\\theta\\) at random positions to simulate the inaccuracies of pre-estimated human motions caused by shot transitions in multi-shot videos. This strategy enables the network to robustly recover smooth and consistent human motion from noisy initial parameters. The benchmark test results were obtained after training for 80 epochs on one NVIDIA-A100 GPU." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.577, + 0.905, + 0.608 + ], + "angle": 0, + "content": "5.3. Main Results: Comparison of Global Human Motion Recovery Results on the Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.613, + 0.907, + 0.78 + ], + "angle": 0, + "content": "We compare our proposed method HumanMM with several state-of-the-art HMR methods (SLAHMR [26], WHAM [28] and GVHMR [29]) on our proposed benchmark ms-Motion. As illustrated in Tab. 1, our proposed method has achieved the best performance for PA&WAMPJPE, RTE and ROE through videos with all numbers of shots across ms-AIST and ms-H3.6M, indicating that our method reconstructs both the global human motion and orientations in the world coordinates more accurately and robustly. For the foot sliding metric, our method also performs as the best on ms-H3.6M across all numbers of shots." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.79, + 0.68, + 0.805 + ], + "angle": 0, + "content": "5.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Human-centric Scene Shot Boundary Detection Evaluation. To evaluate the performance of our proposed Shot Transition Detector, we test the algorithm on our proposed multi-shot human motion recovery benchmark and compare the output frame list of shot transitions with the ground truth (GT) of our dataset. As shown in Tab. 3, by apply" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1979" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.09, + 0.9, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.22, + 0.907, + 0.261 + ], + "angle": 0, + "content": "Figure 7. Qualitative comparison of different HMR methods on ms-Motion dataset. The side view of the rendered mesh for input mutli-shot video is shown in (a), while the top view is shown in (c). We also draw the comparison of the human trajectory as shown in (b). Our method is the most similar as GT in both rendered motion and trajectories among these methods." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.263, + 0.487, + 0.329 + ], + "angle": 0, + "content": "
Methods|PA-MPJE↓|MPJPE↓WA-MPJE↓W-MPJPE↓PVE↓ACCEL↓RTE↓ROE↓F.S.↓
Baseline (Concat)106.48141.67273.15553.67122.156.1410.8691.5514.91
w/o ms-HMR78.24101.52246.42436.5785.775.873.8950.633.54
w/o OAM73.5692.13243.65425.1879.645.676.6176.744.45
w/o traj. pred.50.4983.68231.75432.1775.775.755.5247.684.96
w/o traj. ref.50.4983.68198.58397.6575.775.234.0647.687.84
HumanMM (Ours)50.4983.68194.77393.2175.775.163.5447.683.28
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.33, + 0.49, + 0.352 + ], + "angle": 0, + "content": "Table 5. Ablation studies on different combinations of HumanMM modules. We evaluate the contributions of each key module on EMDB-1." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.357, + 0.483, + 0.492 + ], + "angle": 0, + "content": "ing the proposed finer granularity shot detection methods, the number of recall, precision, and F1 score all increases consistently. The combination of three steps (ScenesDetect, bbox tracking, and pose tracking) has achieved 0.96, 0.88, and 0.92 on the recall, precision, and F1 score, respectively, which indicates a comparable performance in shot boundary detection. Besides, as can be seen in the results, the latter two steps of shot detection contribute to the fine-grained final results significantly and jointly." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.493, + 0.483, + 0.689 + ], + "angle": 0, + "content": "Key modules in the Proposed Method. We compare our methods with four variants on EMDB with noise dataset, as shown in Tab. 5, ms-HMR is the key component for the improvement in PA-MPJPE and PVE, which indicates a more accurate modeling of the whole motion sequence. This design serves as a recovery module to estimate some invisible body parts in some shots. Additionally, the orientation alignment module (OAM, in Sec. 3.4) is also a critical block for accurate human orientation estimation, indicated by the metric ROE. This module helps to model the global human motion between shots. For foot sliding, the results in Tab. 5 also show that the trajectory refiner (Sec. 3.5) in our method helps mitigate the foot sliding issue." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Comparison on Camera Trajectory Estimation. To evaluate the performance of our proposed camera trajectory estimation method Masked LEAP-VO, we evaluate the camera trajectory accuracy on EMDB-1 and EMDB-2. For more convenient comparison, we introduce two baselines, DPVO [76], which has been widely used in HMR methods such as WHAM [28] and GVHMR [29], and LEAP-VO [77]. To provide more intuition about the insights of masking dynamic humans in the video, we also implement a variant, Masked DPVO, by applying SAM at the patchify stage of DPVO to exclude patches containing human pixels. As shown in Tab. 4 and Tab. 6, compared with baseline methods, our key design of masking dynamic human subjects improves the result in both RPE Translation and" + }, + { + "type": "table", + "bbox": [ + 0.498, + 0.263, + 0.672, + 0.305 + ], + "angle": 0, + "content": "
Methods\\( \\left| \\mathrm{{ATE}}\\right| \\downarrow \\)RPE Trans. \\( \\downarrow \\)RPE Rot. \\( \\downarrow \\)
DPVO (w/o mask)0.481.071.26
Masked DPVO0.500.861.21
LEAP-VO (w/o mask)0.500.831.21
Masked LEAP-VO0.490.831.19
" + }, + { + "type": "table_caption", + "bbox": [ + 0.496, + 0.306, + 0.674, + 0.352 + ], + "angle": 0, + "content": "Table 6. Camera tracking results on EMDB-2 [73]. Masked LEAP-VO performs better in both RPE Trans. and RPE Rot. metrics." + }, + { + "type": "table", + "bbox": [ + 0.683, + 0.263, + 0.903, + 0.304 + ], + "angle": 0, + "content": "
MethodsWA-MPJE↓W-MPJE↓RTE↓Jitter↓F.S.↓
DPVO305.40117.105.1017.903.50
Masked DPVO303.90116.404.1017.403.50
LEAP-VO284.10112.803.1016.303.50
Masked LEAP-VO283.70112.703.1016.303.50
" + }, + { + "type": "table_caption", + "bbox": [ + 0.68, + 0.305, + 0.905, + 0.352 + ], + "angle": 0, + "content": "Table 7. Global motion recovery results on EMDB-2 [73]. We input estimated camera parameters from different methods into GVHMR for comparison on HMR." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.357, + 0.907, + 0.477 + ], + "angle": 0, + "content": "RPE Rotation while maintaining competitive ATE. This result indicates the effectiveness of the design of masking dynamic human subjects in the process of camera trajectory estimation. Compared with the DPVO baseline, our method achieves \\(\\sim 50\\%\\) \\(\\downarrow\\) RPE translation on EMDB-1. In addition, we run GVHMR [29] on EMDB-2 with different estimated camera trajectories. The results is shown in Tab. 7, which further illustrates the effectiveness of our method." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.488, + 0.763, + 0.503 + ], + "angle": 0, + "content": "6. Conclusion and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.907, + 0.703 + ], + "angle": 0, + "content": "Conclusion. In this paper, we introduce HumanMM, the first framework designed for human motion recovery from multi-shot videos in world coordinates. HumanMM addresses the challenges inherent in multi-shot videos by incorporating three key components: an enhanced camera trajectory estimation method called masked LEAP-VO, a human motion alignment module that ensures consistency across different shots, and a post-processing module for seamless motion integration. Extensive experiments demonstrate that HumanMM outperforms existing human motion recovery methods across various benchmarks, achieving state-of-the-art accuracy on our newly created multi-shot human motion dataset, ms-Motion." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Limitations and Future Work. While HumanMM represents an dvancement in human motion recovery from multishot videos in world coordinates, its performance may decline when faced with an excessive number of shot transitions. Despite these challenges, HumanMM provides a solid baseline for human motion recovery from multi-shot videos and can be employed in annotating markerless human motion datasets. Our newly introduced dataset, msMotion, offers a valuable benchmark for evaluating general human motion recovery methods in world coordinates, especially regarding their performance on multi-shot videos. Based on the proposed method, our future work aims to enlarge the related datasets for larger-scale motion databases." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1980" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.228 + ], + "angle": 0, + "content": "This research was funded through the National Key Research and Development Program of China Project No. 2022YFB36066), in part by the Shenzhen Science and Technology Project under Grant (KJZD20240903103210014). The author team would also like to convey sincere thanks to Ms. Yaxin Chen from IDEA Research for the expressive dance motion used in the demo presentation." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.257, + 0.188, + 0.273 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.283, + 0.484, + 0.337 + ], + "angle": 0, + "content": "[1] Jingbo Wang, Yu Rong, Jingyuan Liu, Sijie Yan, Dahua Lin, and Bo Dai. Towards diverse and natural scene-aware 3d human motion synthesis. In CVPR, pages 20428-20437, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.34, + 0.484, + 0.395 + ], + "angle": 0, + "content": "[2] Zeqi Xiao, Tai Wang, Jingbo Wang, Jinkun Cao, Wenwei Zhang, Bo Dai, Dahua Lin, and Jiangmiao Pang. Unified human-scene interaction via prompted chain-of-contacts. In ICLR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.398, + 0.484, + 0.452 + ], + "angle": 0, + "content": "[3] Chuan Guo, Xinxin Zuo, Sen Wang, and Li Cheng. Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In ECCV, pages 580-597, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.455, + 0.483, + 0.496 + ], + "angle": 0, + "content": "[4] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. NeurIPS, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.499, + 0.484, + 0.54 + ], + "angle": 0, + "content": "[5] Liang Pan, Jingbo Wang, Buzhen Huang, Junyu Zhang, Haofan Wang, Xu Tang, and Yangang Wang. Synthesizing physically plausible human motions in 3d scenes. In 3DV, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.543, + 0.484, + 0.598 + ], + "angle": 0, + "content": "[6] Jingbo Wang, Ye Yuan, Zhengyi Luo, Kevin Xie, Dahua Lin, Umar Iqbal, Sanja Fidler, and Sameh Khamis. Learning human dynamics in autonomous driving scenarios. In ICCV, pages 20739-20749, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.601, + 0.484, + 0.642 + ], + "angle": 0, + "content": "[7] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. In ECCV, pages 358–374, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.644, + 0.484, + 0.685 + ], + "angle": 0, + "content": "[8] Mathis Petrovich, Michael J Black, and Gül Varol. Temos: Generating diverse human motions from textual descriptions. In ECCV, pages 480-497, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.688, + 0.484, + 0.741 + ], + "angle": 0, + "content": "[9] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. IEEE TPAMI, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.745, + 0.483, + 0.785 + ], + "angle": 0, + "content": "[10] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model. In ICLR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.788, + 0.484, + 0.843 + ], + "angle": 0, + "content": "[11] Zan Wang, Yixin Chen, Tengyu Liu, Yixin Zhu, Wei Liang, and Siyuan Huang. Humanise: Language-conditioned human motion generation in 3d scenes. NeurIPS, pages 14959-14971, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[12] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, and Gang Yu. Executing your commands via motion diffusion in latent space. In CVPR, pages 18000-18010, 2023." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.283, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[13] Ye Yuan, Jiaming Song, Umar Iqbal, Arash Vahdat, and Jan Kautz. Physdiff: Physics-guided human motion diffusion model. In ICCV, pages 16010-16021, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.203 + ], + "angle": 0, + "content": "[14] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Yong Zhang, Hongwei Zhao, Hongtao Lu, Xi Shen, and Ying Shan. Generating human motion from textual descriptions with discrete representations. In CVPR, pages 14730-14740, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.206, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[15] Yonatan Shafir, Guy Tevet, Roy Kapon, and Amit H Bermano. Human motion diffusion as a generative prior. In ICLR, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.905, + 0.3 + ], + "angle": 0, + "content": "[16] Korrawe Karunratanakul, Konpat Preechakul, Supasorn Suwajanakorn, and Siyu Tang. Guided motion diffusion for controllable human motion synthesis. In CVPR, pages 2151-2162, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.905, + 0.357 + ], + "angle": 0, + "content": "[17] Ling-Hao Chen, Shunlin Lu, Wenxun Dai, Zhiyang Dou, Xuan Ju, Jingbo Wang, Taku Komura, and Lei Zhang. Pay attention and move better: Harnessing attention for interactive motion generation and training-free editing, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.359, + 0.905, + 0.412 + ], + "angle": 0, + "content": "[18] Zeqi Xiao, Tai Wang, Jingbo Wang, Jinkun Cao, Wenwei Zhang, Bo Dai, Dahua Lin, and Jiangmiao Pang. Unified human-scene interaction via prompted chain-of-contacts. In ICLR, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.905, + 0.455 + ], + "angle": 0, + "content": "[19] Yiming Xie, Varun Jampani, Lei Zhong, Deqing Sun, and Huaizu Jiang. Omnicontrol: Control any joint at any time for human motion generation. In ICLR, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.457, + 0.905, + 0.497 + ], + "angle": 0, + "content": "[20] Shunlin Lu, Ling-Hao Chen, Ailing Zeng, Jing Lin, Ruimao Zhang, Lei Zhang, and Heung-Yeung Shum. Humantomato: Text-aligned whole-body motion generation. ICML, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.498, + 0.905, + 0.551 + ], + "angle": 0, + "content": "[21] Wenxun Dai, Ling-Hao Chen, Jingbo Wang, Jinpeng Liu, Bo Dai, and Yansong Tang. Motionlcm: Real-time controllable motion generation via latent consistency model. ECCV, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.554, + 0.905, + 0.607 + ], + "angle": 0, + "content": "[22] Ling-Hao Chen, Jiawei Zhang, Yewen Li, Yiren Pang, Xiaobo Xia, and Tongliang Liu. Humanmac: Masked motion completion for human motion prediction. In ICCV, pages 9544-9555, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.609, + 0.905, + 0.676 + ], + "angle": 0, + "content": "[23] Shunlin Lu, Jingbo Wang, Zeyu Lu, Ling-Hao Chen, Wenxun Dai, Junting Dong, Zhiyang Dou, Bo Dai, and Ruimao Zhang. Scamo: Exploring the scaling law in autoregressive motion generation model. arXiv preprint arXiv:2412.14559, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.905, + 0.719 + ], + "angle": 0, + "content": "[24] Wenxun Dai, Ling-Hao Chen, Yufei Huo, Jingbo Wang, Jinpeng Liu, Bo Dai, and Yansong Tang. Real-time controllable motion generation via latent consistency model. 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.721, + 0.905, + 0.775 + ], + "angle": 0, + "content": "[25] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa*, and Jitendra Malik*. Humans in 4D: Reconstructing and tracking humans with transformers. In ICCV, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.777, + 0.905, + 0.816 + ], + "angle": 0, + "content": "[26] Vickie Ye, Georgios Pavlakos, Jitendra Malik, and Angjoo Kanazawa. Decoupling human and camera motion from videos in the wild. In CVPR, 2023. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[27] Yufu Wang, Ziyun Wang, Lingjie Liu, and Kostas Daniilidis. Tram: Global trajectory and motion of 3d humans from inthe-wild videos. ECCV, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[28] Soyong Shin, Juyong Kim, Eni Halilaj, and Michael J. Black. WHAM: Reconstructing world-grounded humans with accurate 3D motion. In CVPR, 2024. 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1981" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.16 + ], + "angle": 0, + "content": "[29] Zehong Shen, Huaijin Pi, Yan Xia, Zhi Cen, Sida Peng, Zechen Hu, Hujun Bao, Ruizhen Hu, and Xiaowei Zhou. World-grounded human motion recovery via gravity-view coordinates. In ACM SIGGRAPH Asia, 2024. 2, 3, 4, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.161, + 0.482, + 0.203 + ], + "angle": 0, + "content": "[30] Angjoo Kanazawa, Jason Y. Zhang, Panna Felsen, and Jitendra Malik. Learning 3d human dynamics from video. In CVPR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.204, + 0.482, + 0.258 + ], + "angle": 0, + "content": "[31] Muhammed Kocabas, Ye Yuan, Pavlo Molchanov, Yunrong Guo, Michael J. Black, Otmar Hilliges, Jan Kautz, and Umar Iqbal. Pace: Human and motion estimation from in-the-wild videos. In 3DV, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.26, + 0.482, + 0.315 + ], + "angle": 0, + "content": "[32] Ling-Hao Chen, Shunlin Lu, Ailing Zeng, Hao Zhang, Benyou Wang, Ruimao Zhang, and Lei Zhang. Motionllm: Understanding human behaviors from human motions and videos. arXiv preprint arXiv:2405.20340, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.482, + 0.37 + ], + "angle": 0, + "content": "[33] Matthias Plappert, Christian Mandery, and Tamim Asfour. Learning a bidirectional mapping between human whole-body motion and natural language using deep recurrent neural networks. RAS, 109:13-26, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.371, + 0.482, + 0.425 + ], + "angle": 0, + "content": "[34] Fangzhou Hong, Mingyuan Zhang, Liang Pan, Zhongang Cai, Lei Yang, and Ziwei Liu. Avatarclip: Zero-shot text-driven generation and animation of 3d avatars. ACM SIGGRAPH, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.427, + 0.482, + 0.468 + ], + "angle": 0, + "content": "[35] Nikos Athanasiou, Mathis Petrovich, Michael J Black, and Gül Varol. Teach: Temporal action composition for 3d humans. In 3DV, pages 414-423, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.482, + 0.524 + ], + "angle": 0, + "content": "[36] Rishabh Dabral, Muhammad Hamza Mughal, Vladislav Golyanik, and Christian Theobalt. Mofusion: A framework for denoising-diffusion-based motion synthesis. In CVPR, pages 9760-9770, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.525, + 0.482, + 0.579 + ], + "angle": 0, + "content": "[37] Mingyuan Zhang, Xinying Guo, Liang Pan, Zhongang Cai, Fangzhou Hong, Huirong Li, Lei Yang, and Ziwei Liu. Remodiffuse: Retrieval-augmented motion diffusion model. In ICCV, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.482, + 0.634 + ], + "angle": 0, + "content": "[38] Weilin Wan, Zhiyang Dou, Taku Komura, Wenping Wang, Dinesh Jayaraman, and Lingjie Liu. Tlcontrol: Trajectory and language control for human motion synthesis. ECCV, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.636, + 0.482, + 0.69 + ], + "angle": 0, + "content": "[39] Chuan Guo, Yuxuan Mu, Muhammad Gohar Javed, Sen Wang, and Li Cheng. Momask: Generative masked modeling of 3d human motions. In CVPR, pages 1900-1910, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.482, + 0.734 + ], + "angle": 0, + "content": "[40] Jinpeng Liu, Wenxun Dai, Chunyu Wang, Yiji Cheng, Yansong Tang, and Xin Tong. Plan, posture and go: Towards open-world text-to-motion generation. ECCV, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.482, + 0.775 + ], + "angle": 0, + "content": "[41] Bo Han, Hao Peng, Minjing Dong, Yi Ren, Yixuan Shen, and Chang Xu. Amd: Autoregressive motion diffusion. In AAAI, pages 2022-2030, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.482, + 0.831 + ], + "angle": 0, + "content": "[42] Zhenyu Xie, Yang Wu, Xuehao Gao, Zhongqian Sun, Wei Yang, and Xiaodan Liang. Towards detailed text-to-motion synthesis via basic-to-advanced hierarchical diffusion model. In AAAI, pages 6252–6260, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[43] Wenyang Zhou, Zhiyang Dou, Zeyu Cao, Zhouyingcheng Liao, Jingbo Wang, Wenjia Wang, Yuan Liu, Taku Komura, Wenping Wang, and Lingjie Liu. Emdm: Efficient motion diffusion model for fast, high-quality motion generation. ECCV, 2024." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[44] Mathis Petrovich, Or Litany, Umar Iqbal, Michael J Black, Gul Varol, Xue Bin Peng, and Davis Rempe. Multi-track timeline control for text-driven 3d human motion generation. In CVPRW, pages 1911-1921, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.15, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[45] Boxun Hu, Mingze Xia, Ding Zhao, and Guanlin Wu. Mona: Moving object detection from videos shot by dynamic camera, 2025." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.194, + 0.905, + 0.263 + ], + "angle": 0, + "content": "[46] Zan Wang, Yixin Chen, Baoxiong Jia, Puhao Li, Jinlu Zhang, Jingze Zhang, Tengyu Liu, Yixin Zhu, Wei Liang, and Siyuan Huang. Move as you say interact as you can: Language-guided human motion generation with scene affordance. In CVPR, pages 433-444, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.266, + 0.905, + 0.307 + ], + "angle": 0, + "content": "[47] Zhongfei Qing, Zhongang Cai, Zhitao Yang, and Lei Yang. Story-to-motion: Synthesizing infinite and controllable character animation from long text, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.309, + 0.905, + 0.351 + ], + "angle": 0, + "content": "[48] Hui En Pang, Zhongang Cai, Lei Yang, Tianwei Zhang, and Ziwei Liu. Benchmarking and analyzing 3d human pose and shape estimation beyond algorithms. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.354, + 0.905, + 0.394 + ], + "angle": 0, + "content": "[49] Gyeongsik Moon, Hongsuk Choi, and Kyoung Mu Lee. Neuralannnot: Neural annotator for 3d human mesh training sets. In CVPRW, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.397, + 0.905, + 0.451 + ], + "angle": 0, + "content": "[50] Gyeongsik Moon, Hongsuk Choi, Sanghyuk Chun, Jiyoung Lee, and Sangdoo Yun. Three recipes for better 3d pseudogots of 3d human mesh estimation in the wild. In CVPRW, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.455, + 0.905, + 0.511 + ], + "angle": 0, + "content": "[51] Hongwei Yi, Hualin Liang, Yifei Liu, Qiong Cao, Yandong Wen, Timo Bolkart, Dacheng Tao, and Michael J Black. Generating holistic 3d human motion from speech. In CVPR, pages 469-480, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.513, + 0.905, + 0.54 + ], + "angle": 0, + "content": "[52] C.J. Bowen and R. Thompson. Grammar of the Edit. Focal Press, 2013. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.543, + 0.905, + 0.597 + ], + "angle": 0, + "content": "[53] Mathis Petrovich, Or Litany, Umar Iqbal, Michael J. Black, Gül Varol, Xue Bin Peng, and Davis Rempe. Multi-track timeline control for text-driven 3d human motion generation. In CVPRW, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.601, + 0.905, + 0.668 + ], + "angle": 0, + "content": "[54] Yuhong Zhang, Jing Lin, Ailing Zeng, Guanlin Wu, Shunlin Lu, Yurong Fu, Yuanhao Cai, Ruimao Zhang, Haoqian Wang, and Lei Zhang. Motion-\\(\\mathbf{x} + +\\): A large-scale multimodal 3d whole-body human motion dataset. arXiv preprint arXiv:2501.05098, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.672, + 0.905, + 0.727 + ], + "angle": 0, + "content": "[55] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In CVPR, pages 5152-5161, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.73, + 0.905, + 0.771 + ], + "angle": 0, + "content": "[56] Georgios Pavlakos, Jitendra Malik, and Angjoo Kanazawa. Human mesh recovery from multiple shots. In CVPR, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.774, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[57] Peng Wu, Xiankai Lu, Jianbing Shen, and Yilong Yin. Clip fusion with bi-level optimization for human mesh reconstruction from monocular videos. In ACM MM, page 105-115, New York, NY, USA, 2023. Association for Computing Machinery." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.846, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[58] Kuan-Chieh Wang, Zhenzhen Weng, Maria Xenochristou, Joao Pedro Araujo, Jeffrey Gu, C Karen Liu, and Serena Yeung. Nemo: 3d neural motion fields from multiple video instances of the same action. In CVPR, 2023." + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1982" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[59] Fabien Baradel, Thibault Groueix, Philippe Weinzaepfel, Romain Brégier, Yannis Kalantidis, and Grégory Rogez. Leveraging mocap data for human mesh recovery. In 3DV, pages 586-595, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.189 + ], + "angle": 0, + "content": "[60] Ruilong Li, Shan Yang, David A. Ross, and Angjoo Kanazawa. Ai choreographer: Music conditioned 3d dance generation with aist++, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.191, + 0.482, + 0.245 + ], + "angle": 0, + "content": "[61] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE TPAMI, 36(7):1325-1339, 2014. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.247, + 0.482, + 0.287 + ], + "angle": 0, + "content": "[62] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. ACM TOG, 36(6), 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.289, + 0.482, + 0.356 + ], + "angle": 0, + "content": "[63] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J. Black. Keep it SMPL: Automatic estimation of 3D human pose and shape from a single image. In Computer Vision - ECCV 2016. Springer International Publishing, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.358, + 0.482, + 0.398 + ], + "angle": 0, + "content": "[64] Anurag* Arnab, Carl* Doersch, and Andrew Zisserman. Exploiting temporal context for 3d human pose estimation in the wild. In CVPR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.482, + 0.44 + ], + "angle": 0, + "content": "[65] Ahmed A. A. Osman, Timo Bolkart, and Michael J. Black. STAR: Sparse Trained Articulated Human Body Regressor, page 598-613. Springer International Publishing, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.482, + 0.496 + ], + "angle": 0, + "content": "[66] Yinghao Huang, Federica Bogo, Christoph Lassner, Angjoo Kanazawa, Peter V. Gehler, Javier Romero, Ijaz Akhter, and Michael J. Black. Towards accurate marker-less human shape and pose estimation over time. In 3DV, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.498, + 0.482, + 0.538 + ], + "angle": 0, + "content": "[67] Angjoo Kanazawa, Michael J. Black, David W. Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In CVPR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.54, + 0.482, + 0.579 + ], + "angle": 0, + "content": "[68] Muhammed Kocabas, Nikos Athanasiou, and Michael J. Black. Vibe: Video inference for human body pose and shape estimation. In CVPR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.482, + 0.622 + ], + "angle": 0, + "content": "[69] Nikos Kolotouros, Georgios Pavlakos, and Kostas Dani-ilidis. Convolutional mesh regression for single-image human shape reconstruction. In CVPR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.624, + 0.482, + 0.663 + ], + "angle": 0, + "content": "[70] Zhengyi Luo, S. Alireza Golestaneh, and Kris M. Kitani. 3d human motion estimation via motion compression and refinement. In ACCV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.665, + 0.482, + 0.718 + ], + "angle": 0, + "content": "[71] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS. Curran Associates, Inc., 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.721, + 0.482, + 0.788 + ], + "angle": 0, + "content": "[72] Chun-Hao P. Huang, Hongwei Yi, Markus Höschle, Matvey Safroshkin, Tsvetelina Alexiadis, Senya Polikovsky, Daniel Scharstein, and Michael J. Black. Capturing and inferring dense full-body human-scene contact. In CVPR, pages 13274-13285, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[73] Manuel Kaufmann, Jie Song, Chen Guo, Kaiyue Shen, Tianjian Jiang, Chengcheng Tang, Juan Jose Zárate, and Otmar Hilliges. EMDB: The Electromagnetic Database of Global 3D Human Pose and Shape in the Wild. In ICCV, 2023. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[74] Xinyu Yi, Yuxiao Zhou, and Feng Xu. Transpose: Real-time 3d human translation and pose estimation with six inertial sensors. ACM TOG, 40(4), 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[75] Zachary Teed and Jia Deng. Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. In NeurIPS, pages 16558-16569. Curran Associates, Inc., 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.137, + 0.905, + 0.163 + ], + "angle": 0, + "content": "[76] Zachary Teed, Lahav Lipson, and Jia Deng. Deep patch visual odometry. NeurIPS, 2023. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.905, + 0.206 + ], + "angle": 0, + "content": "[77] Weirong Chen, Le Chen, Rui Wang, and Marc Pollefeys. Leap-vo: Long-term effective any point tracking for visual odometry. In CVPR, 2024. 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.208, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[78] Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, and Michael J. Black. AMASS: Archive of motion capture as surface shapes. In ICCV, pages 5442-5451, 2019. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[79] Timo von Marcard, Roberto Henschel, Michael Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In ECCV, 2018. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[80] Yufei Xu, Jing Zhang, Qiming Zhang, and Dacheng Tao. ViTPose: Simple vision transformer baselines for human pose estimation. In NeurIPS, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.364, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[81] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.448, + 0.905, + 0.501 + ], + "angle": 0, + "content": "[82] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multiperson linear model. ACM TOG, 34(6):248:1-248:16, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.905, + 0.545 + ], + "angle": 0, + "content": "[83] Qingqiu Huang, Yu Xiong, Anyi Rao, Jiaze Wang, and Dahua Lin. Movienet: A holistic dataset for movie understanding. In ECCV, 2020. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.905, + 0.588 + ], + "angle": 0, + "content": "[84] MMTracking Contributors. MMTracking: OpenMMLab video perception toolbox and benchmark. https://github.com/open-mmlab/mmtracking, 2020.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.591, + 0.905, + 0.631 + ], + "angle": 0, + "content": "[85] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. CoTracker: It is better to track together. In ECCV, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.633, + 0.905, + 0.687 + ], + "angle": 0, + "content": "[86] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. In ICCV, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.905, + 0.744 + ], + "angle": 0, + "content": "[87] Martin A. Fischler and Robert C. Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Commun. ACM, 24(6):381-395, 1981. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.747, + 0.905, + 0.787 + ], + "angle": 0, + "content": "[88] Jie Yang, Ailing Zeng, Shilong Liu, Feng Li, Ruimao Zhang, and Lei Zhang. Explicit box detection unifies end-to-end multi-person pose estimation. In ICLR, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[89] Richard Hartley and Andrew Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, USA, 2 edition, 2003. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.885 + ], + "angle": 0, + "content": "[90] Michael J. Black, Priyanka Patel, Joachim Tesch, and Jinlong Yang. BEDLAM: A synthetic dataset of bodies exhibiting detailed lifelike animated motion. In CVPR, pages 8726-8737, 2023. 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.885 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1983" + } + ] +] \ No newline at end of file diff --git a/2025/HumanMM_ Global Human Motion Recovery from Multi-shot Videos/8ae05aea-0260-4aa4-bbb0-352a7f188bbe_origin.pdf b/2025/HumanMM_ Global Human Motion Recovery from Multi-shot Videos/8ae05aea-0260-4aa4-bbb0-352a7f188bbe_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..56acac799d5b3cb7715c689336759e5a61cc34a9 --- /dev/null +++ b/2025/HumanMM_ Global Human Motion Recovery from Multi-shot Videos/8ae05aea-0260-4aa4-bbb0-352a7f188bbe_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:27a37fbfa3cc606e73d80faee9299c2488a91f0cbc9696500bdc92eca1078020 +size 1758892 diff --git a/2025/HumanMM_ Global Human Motion Recovery from Multi-shot Videos/full.md b/2025/HumanMM_ Global Human Motion Recovery from Multi-shot Videos/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f1d4f17622dd764256bb26d2ca83ec38aa1453bb --- /dev/null +++ b/2025/HumanMM_ Global Human Motion Recovery from Multi-shot Videos/full.md @@ -0,0 +1,362 @@ +# HumanMM: Global Human Motion Recovery from Multi-shot Videos + +Yuhong Zhang $^{1,2,\dagger*}$ Guanlin Wu $^{2,3,\ddagger*}$ Ling-Hao Chen $^{1,2,\ddagger}$ Zhuokai Zhao $^{4}$ Jing Lin $^{1,2}$ Xiaoke Jiang $^{2}$ Jiamin Wu $^{2,5}$ Zhuoheng Li $^{6}$ Hao Frank Yang $^{3}$ Haoqian Wang $^{1\dagger}$ Lei Zhang $^{2\dagger}$ + +$^{1}$ Tsinghua University $^{2}$ IDEA Research $^{3}$ Johns Hopkins University $^{4}$ University of Chicago $^{5}$ HKUST $^{6}$ HKU + +{dsyuhong, guanlinwu0930, thu.lhchen}@gmail.com + +Project page: https://zhangyuhong01.github.io/HumanMM + +![](images/5cd68b86bc48e672d8354ecd0927f8ed06c9b3b9c195f60338892f3cf801c111.jpg) +Figure 1. Recovering a human motion from multi-shot videos. Top: We take two multi-shot table tennis game videos with shot transitions as input. We aim to recover two motions of two athletes (Long MA and Zhendong FAN) from two videos, respectively. The first video is recorded by three shots (“①”, “②”, and “③”), and the second one is recovered by two shots (“④” and “⑤”). Bottom: We recover two motions (Long MA in green and Zhendong FAN in pink), different shots, and camera poses for each multi-shot video. The recovered motion is aligned with the motion in the videos. + +# Abstract + +In this paper, we present a novel framework designed to reconstruct long-sequence 3D human motion in the world coordinates from in-the-wild videos with multiple shot transitions. Such long-sequence in-the-wild motions are highly valuable to applications such as motion generation and motion understanding, but are of great challenge to be recovered due to abrupt shot transitions, partial occlusions, and dynamic backgrounds presented in such videos. Existing + +methods primarily focus on single-shot videos, where continuity is maintained within a single camera view, or simplify multi-shot alignment in camera space only. In this work, we tackle the challenges by integrating an enhanced camera pose estimation with Human Motion Recovery (HMR) by incorporating a shot transition detector and a robust alignment module for accurate pose and orientation continuity across shots. By leveraging a custom motion integrator, we effectively mitigate the problem of foot sliding and ensure temporal consistency in human pose. Extensive evaluations on our created multi-shot dataset from public 3D human datasets demonstrate the robustness of our method in reconstructing realistic human motion in world coordinates. + +# 1. Introduction + +In recent years, significant advances have been made in 3D human pose estimation, particularly in enhancing the accuracy of human motion recovery (HMR)1 from monocular video sequences. HMR has demonstrated extensive applications in human-AI interaction [1, 2], human motion understanding [3-6], and motion generation [3, 7-24]. While existing methods [25] achieved relatively high performance in recovering human mesh in camera coordinates, estimating human motion in world coordinates remains challenging [26-29] due to inaccurate camera pose estimation and the complexity of reconstructing human motion spatially. + +Most current progress in 3D human motion community mainly benefits from large scale data [25, 27-31], and longsequence videos. These resources enhance estimation accuracy for HMR methods and improve the understanding and generation of longer motion sequences for tasks such as motion understanding [3, 32, 33] and generation [3, 7-20, 34-47], even when annotations are derived from markerless capturing methods like pseudo labels [48-51]. + +A promising approach to enlarge the scale of the motion databases is to estimate human motions from unlimited online videos in a markerless manner. However, many longsequence online videos are recorded with multiple shots, referred to as multi-shot videos $^2$ , especially prevalent in domains such as sports broadcasting, talk shows, and concerts. In filmmaking and television live show, a "shot" denotes an individual camera view capturing a specific moment or action from a particular vantage point [52]. + +Segmenting multi-shot videos into separate shots inevitably reduces the length of the video sequences, which can be detrimental to tasks that benefit from longer sequences, such as long motion generation [47, 53]. This limitation is highlighted in the existing datasets [54, 55], where the longest clip is less than 20 seconds after segmentation, as shown in Fig. 2. Moreover, focusing exclusively on online single-shot videos diminishes the utilization ratio of available online videos and may negatively impact the diversity of scenarios represented in the created datasets. + +Therefore, how to address the issue of discontinuities caused by shot transitions is notoriously difficult in the community. To resolve this problem, previous works [56-59] have proposed algorithms to address human mesh recovery in a camera space from movies containing shot change between long shots and close-ups. + +![](images/11082310bcbfb4fec8c0288df95839fd66f45dc6e327ff64fc02a4f0b28c9d25.jpg) +Figure 2. The comparison between the distribution of sequence lengths in different existing large-scale markerless motion datasets with ours. The $x$ -axis and $y$ -axis denote the duration time (s) and percentage of video number, respectively. Our dataset (in green) contains more portion of long-sequence videos in general. + +However, recovering human motions in world coordinates from multi-shot videos presents two fundamental challenges that remain underexplored. 1) How to align the human motion and orientation in the world coordinates during shot transitions? Ensuring continuity of human orientation and pose across shots is complicated by factors such as partial visibility of human body (e.g. transitioning from long shot to close-up) and changes in human orientation (e.g. two long shots from different viewpoints). These issues, caused by abrupt changes in camera viewpoints, necessitate robust alignment mechanisms. 2) How to reconstruct accurate human motion in world coordinates? Existing approaches employ Simultaneous Localization and Mapping (SLAM) methods to estimate camera parameters, which are then used to project recovered human meshes from camera to world coordinates [26-29]. This process requires highly accurate camera estimation and must address motion consistency and foot sliding in the recovered human motion within the world space. + +Despite these challenges, human motion in multi-shot videos often remain continuous across shots, even as camera viewpoints change. This observation suggests that with appropriate handling of shot transitions and camera motion, it is possible to reconstruct consistent and complete 3D human motions throughout multi-shot videos. + +In this paper, we propose a novel framework HumanMM, Human Motion recovery from Multi-shot videos, to address these challenges. It integrates human pose estimation across shots with robust camera estimation in the world space. Firstly, we develop a shot transition detector to identify frames with shot transitions. To ensure a more robust camera pose estimation, we introduce an enhanced SLAM method incorporating long-term tracking of feature points and exclusion of moving human from bundle adjustment process. We utilize existing HMR method integrated with our enhanced camera estimation to get the initial human parameters for each separated shot. Subsequently, we implement an alignment module to align human orientation based on stereo calibration and smooth human poses through a + +trained multi-shot HMR encoder, which effectively captures the temporal context of human movements across different shots. Finally, after aligning human and camera parameters between shot transitions, we train a motion decoder and a trajectory refiner to smooth the human pose and mitigate issues such as foot sliding, thereby enhancing the overall motion consistency in the reconstructed 3D human motions. + +Our contributions can be summarized as follows. + +- We present the first approach to reconstruct human motion from multi-shot videos in world coordinates. +- We introduce HumanMM, a HMR framework for multi-shot videos. It includes an enhanced camera trajectory estimation method, a human motion alignment module and a motion integrator to ensure accurate and consistent recovery of human pose and orientation in world coordinates across different shots in the whole video. +- We develop a multi-shot video dataset ms-Motion to evaluate the performance of HMR from multi-shot videos, based on existing public datasets such as AIST [60] and Human3.6M [61]. Extensive experiments on related benchmarks verify the effectiveness of our method. + +# 2. Related Work + +# 2.1. HMR from One-shot Video + +One-shot videos, captured with a single camera without shot transitions, has been extensively studied within the community for human mesh and motion recovery. + +Human mesh recovery in camera coordinates can be broadly categorized into two approaches: optimization-based methods [62-66] and regression-based methods [30, 67-70]. With the significant advancements of transformer [71], HMR2.0 [25] has surpassed previous methods and benefits several downstream tasks related to HMR. + +Although there are several previous works tried to recover motions in world coordinates with multi-camera capture system [60, 72] and IMU-based methods [73, 74] and enjoy relatively satisfying results, this setup limits their use for applications of infinite in-the-wild monocular videos. To address this limitation, several attempts [26-29] integrate SLAM into the HMR pipeline by first estimating the camera pose using SLAM methods, e.g. DROID-SLAM [75] or DPVO [76], and then project the recovered human motion from camera to world coordinates. To exclude the inconsistencies caused by dynamic objects, such as moving humans, TRAM [27] modifies DROID-SLAM by incorporating human masking and depth-based distance rescaling. However, DROID-SLAM performs dense bundle adjustment (DBA) on feature maps from downsampled images and selects features based only on two consecutive frames rather than long-term video sequences [75-77]. Consequently, masking significantly reduces the number of informative and consistent features, especially when humans occupy large + +portions of the image, leading to inaccuracies. Therefore, developing a SLAM method that retains sufficient and representative features for DBA after masking is important. + +# 2.2. HMR from Multi-shot Video + +Multiple shots are fundamental elements of cinematic storytelling and live performances, utilizing various camera positions and focal lengths to create immersive and detailed viewing experiences for audiences [52]. However, most marker-based motion capture (MoCap) datasets [60, 72, 73, 78, 79] consist single-shot videos only, resulting in limited research on HMR from multi-shot videos. + +Recovering human motion from multi-shot videos in camera coordinates is already challenging. This is because treating each pose estimation result of each shot separately leads to inconsistencies when combining all estimations, caused by partially or fully invisible human bodies across shot transitions. Pavlakos et al. [56] addresses this issue by focusing on shot changes from long shots to close-ups, which are common in film. They develop smoothness constraints within a temporal Human Mesh and Motion Recovery (t-HMMR) model to infer motions during occlusions caused by shot transitions. + +Advancements in HMR methods [29] for single-shot videos in world coordinates have paved the way for extending HMR to multi-shot videos with varying camera viewpoints. However, aligning human orientation, body pose, and translation continuously across multi-shot videos in world coordinates underexplored. Effective alignment is crucial to maintain motion continuity and coherence, especially when dealing with diverse camera perspectives and abrupt transitions between shots. + +In summary, while substantial progress has been made in HMR from single-shot videos, extending these techniques to multi-shot videos requires addressing additional complexities related to camera pose alignment and motion consistency across shot transitions. We address this challenge by proposing a novel pipeline that ensures accurate and continuous 3D HMR from multi-shot monocular videos. + +# 3. Method + +In this section, we propose HumanMM to recover human motion from multi-shot videos. The system overview is shown in Fig. 3. Given an input video sequence $\mathbf{V} = \{I_t\}_{t=1}^T$ of length $T$ , where $I_t$ denotes the $t$ -th frame, our objective is to recover human motion in world coordinates. We begin by detecting shot transition frames based on human bounding box (a.k.a. bbox) and 2D keypoints (a.k.a. KPTs) through a shot transition detector (Sec. 3.2). For each clipped shot, we initialize the camera pose (camera rotation and camera translation) and recover initial human motion in world coordinates (Sec. 3.3). The initialized SMPL parameters and camera poses are then fed into a hu + +![](images/ca2ca7aec029cc764656585c8a8e1546a71db368b30708f80823d8e6d5e2ade0.jpg) +Figure 3. The overview of HumanMM. HumanMM processes multi-shot video sequences by first extracting motion feature such as keypoints and bounding boxes, using ViTPose [80] and image feature using ViT [81]. These features are then segmented into single-shot clips via Shot Transition Detection (Sec. 3.2). Initialized camera (camera rotation $\mathbf{R}$ and camera translation $\mathbf{T}$ ) and human (SMPL) parameters for each shot are estimated using Masked LEAP-VO (Sec. 3.3) and GVHMR [29]. Human orientation is aligned across shots through camera calibration (3.4.1), and ms-HMR (Sec. 3.4.2) ensures consistent pose alignment. Finally, a bi-directional LSTM-based trajectory predictor with trajectory refiner predicts trajectory based on aligned motion and mitigates foot sliding throughout the video. + +man motion alignment module (Sec. 3.4), which aligns human orientations via camera calibration based on human 2D KPTs and smooth the human pose by incorporating pose information across different shots. Additionally, it refines the entire motion sequence through whole video using a temporal motion encoder ms-HMR. Finally, we introduce a post-processing module for motion integration (Sec. 3.5). + +# 3.1. Preliminary: 3D Human Model + +Our method aims to recover motions in world coordinates in the SMPL [82] format, whose pose at frame $t$ can be represented as $\mathcal{M}_t(\theta_t,\beta_t,\Gamma_t,\tau_t)\in \mathbb{R}^{6890\times 3}$ . Here, the body pose, body shape, root orientation, and translation are $\theta_{t}\in \mathbb{R}^{23\times 3}$ , $\beta_{t}\in \mathbb{R}^{10}$ , $\Gamma_t\in \mathbb{R}^3$ , and $\tau_{t}\in \mathbb{R}^{3}$ , respectively. We use $\mathbf{K}_t^{2D}$ to denote human 2D KPTs at each frame $t$ . + +# 3.2. Shot Transition Detector For Multi-shot Video + +Our algorithm begins with shot transition detection in one video. As shown in Fig. 3, the shot transition detector has three key components, scene transition detector, bounding box (a.k.a. bounding box) tracking, and human keypoints tracking. (1) Scene change transition detector: Initially, we employ the SceneDetect [83] algorithm to identify scene changes based on significant variations in the background. However, the SceneDetect fails to detect shot transitions when background changes are unnoticeable, illustrated in Fig. 4. Subsequently, we leverage the following modules to bridge the gap. (2) Bbox tracking for shot transition. As a shot change often accompanies with a sudden change of human subject size, we track humans in a video via mmtracking [84]. Consequently, we compute the Intersection over Union (IoU) between neighbor bboxes and identify a shot transition when the IoU falls smaller than a manually tuned threshold. (3) Human pose tracking for shot transition detection. To achieve a finer granularity, we additionally intro + +![](images/c26a99eaab97d284f84c633ae3c919750b0f716898a770539707640218814788.jpg) +(a) Scene Change + +![](images/96aeb8efd260b8fa739c9d6857c6f8eeda93fc99f2d21f9e1cea00ef87964fbb.jpg) +(b) Position Change + +![](images/35edaa31a1ff346fdad3638262003b6d7ff3eb8e5945f481aa50a7a61b87ec6e.jpg) +(c) Pose Change + +Figure 4. Shot transition detection examples. Examples (a), (b), and (c) illustrate multi-shot scenarios in online videos. (a) shows scene transitions detectable by SceneDetect. (b) illustrates significant position changes undetectable by SceneDetect but resolvable withbbox tracking-based method. (c) shows pose or orientation transition, requiring pose tracking-based methods as they cannot be addressed by either SceneDetect or_bbox tracking. + +duce human 2D KPTs to detect extreme corner shot changes in a video. By thresholding the IoU of corresponding keypoints between neighbor frames, we can accurately identify shot transitions even with subtle human movements. + +As each separate module cannot identify all kinds of shot transitions, the three modules are jointly used to clip a video into several sub-sequences serially. + +# 3.3. Human Motion and Camera Pose Estimation For Each Shot + +After obtaining the clipped videos, our next goal is to estimate the camera pose and SMPL parameters in the world coordinates for each clipped video. The estimated camera pose and motions for each shot will be used to construct the whole motion sequence in the next stage (Sec. 3.4). + +How to estimate the camera parameters accurately? Our approach for camera parameter calculation is based on a visual odometry (VO) estimation method, LEAP-VO [77]. Utilizing the CoTracker method [85], LEAP-VO estimates the visibility and trajectories of $N$ selected points by ana + +lyzing image gradients across the video sequence. LEAPVO subsequently computes confidence scores for each trajectory, retaining only those with high confidence while discarding trajectories shorter than a predefined threshold. The remaining trajectories undergo bundle adjustment (BA) within a fixed window size to estimate the camera poses. + +However, simply applying LEAP-VO in the camera estimation process is still unsatisfactory in most human-centric scenarios. The primary limitation stems from the dynamic movements of human subjects, which typically occupy a substantial portion of each image in human-centric videos. This dynamic presence introduces noise into the camera pose estimation in world coordinates, as the estimation process relies heavily on the relationship between the camera and the static environment. To address this issue, we propose a Masked LEAP-VO algorithm. Our approach involves inputting the image $I_{t}$ and the human bbox at frame $t$ into SAM [86] to generate a human mask. We then assign a visibility value of zero to points within the human mask, effectively excluding these trajectories from the BA process. For clarity, we denote $S_{BA}$ as the window size of BA, $\hat{n}$ denotes the number of filtered point trajectories, and $w_{ij,\hat{n}}$ as the normalized weight based on confidence score and visibility. For estimating the camera poses $\mathbf{G} = \{\mathbf{R},\mathbf{T}\}$ of orientation and translation, the reprojection loss function for BA can then be formulated as follows, + +$$ +\mathbf {G} = \underset {\mathbf {G}, d _ {i}, \hat {n}} {\arg \min } \sum_ {i} \sum_ {j \in | i - j | \leq S} \sum_ {B A} w _ {i j, \hat {n}} | | \mathcal {F} (\mathbf {G} _ {i}, \mathbf {G} _ {j}, d _ {i, \hat {n}}) - \Pi_ {i j} (\mathbf {p} _ {i, \hat {n}}) | |, +$$ + +where $\mathcal{F}(\mathbf{G}_i,\mathbf{G}_j,d_{i,\hat{n}})$ denotes the point positions calculated by camera pose $\mathbf{G}$ at frame $i$ and $j$ with depth $d_{i,\hat{n}}$ . $\Pi_{ij}(\mathbf{p}_{i,\hat{n}})$ denotes the position for project position of $\mathbf{p}_{i,\hat{n}}$ from frame $i$ to $j$ . Consequently, we obtain the camera rotation $\mathbf{R}_t$ and translation $\mathbf{T}_t$ from camera pose $\mathbf{G}_t$ at $t$ . + +Recovering human motion in world coordinates with estimated camera parameters. Given an input video, we feed the estimated camera parameters $(\mathbf{R}_t$ and $\mathbf{T}_t)$ into the state-of-the-art motion recovering model, GVHMR [29], + +$$ +\theta_ {t} ^ {w}, \beta_ {t} ^ {w}, \Gamma_ {t} ^ {w}, \tau_ {t} ^ {w} = \mathrm {G V H M R} \left(I _ {t}, \mathbf {R} _ {t}, \mathbf {T} _ {t}\right). \tag {1} +$$ + +Initialized human parameters $\theta_t^w,\beta_t^w,\Gamma_t^w,\tau_t^w$ and camera parameters $\mathbf{R}_t,\mathbf{T}_t$ will input to human motion alignment. + +# 3.4. Aligning Human Motion Between Shots + +Based on initialized world motion for each individual shot, the subsequent question is how to merge discontinuous motions from different shots into a continuous motion sequence as a whole in world coordinates. A straightforward solution is to align all motion sequences to the world coordinate system of the first shot. However, finding the correspondence between different shots is still under-explored and challenging. To resolve this issue, we decompose the motion parameters into camera-dependent and camera-independent ones. The former (Sec. 3.4.1) achieves alignment between shots + +![](images/c18d8cf03b5852f2d79f6a998a219a15655cf336e96ac1be4db29ebd159ef0b2.jpg) +Figure 5. Human orientation alignment module. Following a shot transition after the foremost purple human mesh (shot $①$ captured by camera $C_0$ ), the unaligned (blue) and aligned (green) motions are captured as shot $②$ and shot “ $③$ ” by camera $C_0'$ and $C_1$ , respectively. $C_0' = C_0$ . To achieve human orientation alignment from shot “ $①$ ” to “ $③$ ”, the camera rotation matrix from $C_0'$ to $C_1$ is computed and applied as the offset of human orientation. + +via human orientation alignment based on camera calibration, whereas the latter (Sec. 3.4.2) is a trainable module to enhance the continuity of human motion sequence. These two key designs ensure a consistent motion sequence between frames when encountering shot transitions. + +# 3.4.1 Aligning Human Orientations Between Shots + +After obtaining the initial SMPL and camera parameters $\{\theta_t^i,\beta_t^i,\Gamma_t^i,\tau_t^i,\mathbf{R}_t^i,\mathbf{T}_t^i\}$ for each shot, directly concatenating motions between shots result abrupt changes of human poses and orientations. To address this issue, we introduce the Orientation Alignment Module (OAM), as shown in Fig. 5, to align human orientations. As the whole motion sequence is continuous, we have the following assumption. Assumption 1 Human orientations and translations during the shot transition in world coordinates are continuous. To align the orientations between two frames with shot transition under Assumption 1, we decompose the human orientation with shot transitions in world coordinates as, + +$$ +\mathrm {R} \left(\Gamma_ {\text {w o r l d}}\right) = \mathrm {R} _ {\delta_ {\text {c a m}}} \mathrm {R} \left(\Gamma_ {\text {v i e w}}\right), \tag {2} +$$ + +where $\mathbf{R}_{\delta_{\mathrm{cam}}}$ represents the camera rotation on the Y-axis between current $t$ -th and previous $t - 1$ -th frame, $\Gamma_{\mathrm{view}}$ denotes the human orientation estimated by the current shot, and $\mathbb{R}(\cdot):\mathbb{R}^3\to \mathbb{R}^9$ is the mapping from axis angle to rotation matrix. As $\Gamma_{\mathrm{view}}$ in current shot can be estimated independently, mentioned in Sec. 3.3, obtaining accurate $\Gamma_{\mathrm{world}}$ in Eq. (2) remains a key challenge to estimate the relative camera rotation $\mathbf{R}_{\delta_{\mathrm{cam}}}$ between frames in shot transitions. + +Estimating the relative camera pose $\mathbf{R}_{\delta_{\mathrm{cam}}}$ between transition frames. Different from our approach of estimating camera pose in each shot (Sec. 3.3), we do not mask the human subject when estimating camera rotation $\mathbf{R}_{\delta_{\mathrm{cam}}}$ . Instead, we use human 2D KPTs as explicit feature matching. Specifically, we filter out unmatched keypoints based on their visibility and unaligned direction using RANSAC [87], effectively addressing camera pose estimation during shot transitions. This procedure is referred + +to as Camera Calibration (a.k.a. epipolar-geometry-based camera extrinsic estimation), and is detailed below. + +In Camera Calibration, we assume that the human translations remain unchanged across the shot transition, implying that only the camera's orientation changes (i.e. Assumption 1). Consequently, we calculate the orientation offset by determining the change in camera orientation using camera calibration. We begin by extracting human 2D KPTs from two consecutive frames during the shot transition. Due to the shot transition, the visibility of 2D KPTs may vary, e.g. occlusion in some shots. Therefore, we employ ED-Pose [88] to filter out invisible 2D KPTs between shot transition frames. Subsequently, RANSAC identifies matching 2D KPTs corresponding to the most possible camera rotation direction. These matched 2D KPTs facilitate the estimation of the aligned camera rotation $\mathbf{R}_{\delta_{\mathrm{cam}}}$ . The detailed estimation process is as follows. + +We denote the detected 2D KPTs of two frames in the shot transition as $\mathbf{S}_1 = [(x_1^{(1)},y_1^{(1)}),(x_1^{(2)},y_1^{(2)}),\dots ,(x_1^{(N)},y_1^{(N)})]^{\top}\in \mathbb{R}^{2\times N}$ and $\mathbf{S}_2 = [(x_2^{(1)},y_2^{(1)}),(x_2^{(2)},y_2^{(2)}),\dots ,(x_2^{(N)},y_2^{(N)})]^{\top}\in$ $\mathbb{R}^{2\times N}$ . The essential matrix $\mathbf{E} = [\mathbf{T}]_{\times}\mathbf{R}$ should satisfy the following orthogonal property such that, + +$$ +\mathbf {S} _ {1} ^ {\top} \mathbf {E} \mathbf {S} _ {2} = \mathbf {0}. \tag {3} +$$ + +Once $\mathbf{E}$ is obtained by solving Eq. (3), we enforce the rank-2 constraint on $\mathbf{E}$ through SVD decomposition and subsequently derive the aligned camera rotation $\mathbf{R}_{\delta_{\mathrm{cam}}}$ between two frames (cf. Hartley et al. [89] for more details). + +In summary, we reformulate the alignment problem of human orientation in shot transitions as estimating the relative camera rotation $\mathbf{R}_{\delta_{\mathrm{cam}}}$ between frames. Accordingly, we obtain the camera rotation $\mathbf{R}_{\delta_{\mathrm{cam}}}$ via camera calibration. + +# 3.4.2 Aligning Human Poses Between Shots + +In shot transition, video sequences recorded by two shots are often with various occlusions. However, unoccluded body parts in two shots can be complementary to each other for motion alignment. Thus, we introduce the multi-shot HMR (ms-HMR, i.e. $\mathsf{E}_M(\cdot)$ ) module to refine the whole motion sequence. As shown in Fig. 6, the ms-HMR is a Transformer encoder-like architecture, whose input and output are the estimated global motion and the refined global motion, respectively. The process can be formulated as, + +$$ +\phi_ {1}, \phi_ {2}, \dots , \phi_ {T} = \mathrm {E} _ {M} \left(\theta_ {1}, \theta_ {2}, \dots , \theta_ {T}\right), \tag {4} +$$ + +where $\phi_*$ denotes the refined motion of each frame. With this design, our method can adapt to diverse occlusions of human body brought by shot transitions. + +# 3.5. Post-processing Module for Motion Integration + +Trajectory Predictor and Trajectory Refiner. Based on the aligned human pose and orientation, we introduce a + +![](images/1cb1151ba9da0460afacd94ca6f5dc2be9bccac51fe297050c8a9acf8ea47910.jpg) +Figure 6. ms-HMR Structure. The initial human pose parameters $\theta$ across multiple video shots are input into a transformer with shot-index-based positional encoding. This enables ms-HMR to generate consistent human poses across all shots in the video. bi-directional LSTM trajectory predictor to recover foot-ground contact probabilities $p_t^c$ , and root velocity $v_t$ as, + +$$ +p _ {t} ^ {c}, v _ {t} = \operatorname {L S T M} \left(\phi_ {1} ^ {m}, \Gamma_ {1}, \mathbb {F} (I _ {1}), \phi_ {2} ^ {m}, \Gamma_ {2}, \mathbb {F} (I _ {2}), \dots , \right. \\ \left. \quad \left(m, \mathbb {F} - \mathbb {F} (I _ {m})\right)\right) \tag {5} +$$ + +$$ +\phi_ {T} ^ {m}, \Gamma_ {T}, F (I _ {T})), +$$ + +where $\mathsf{F}(\cdot)$ denotes the image feature of each frame extracted by ViT [81]. Accordingly, the contact probabilities $p_t^c$ , and velocity $v_t$ are supervised by the ground-truth labels with MSE loss and are used to reconstruct the trajectory. Besides, we extend the trajectory refiner in WHAM [28] to alleviate foot sliding problem in our estimated trajectory. + +# 4. Benchmarking Multi-shot Motion Recovery + +Dataset Construction. To create a multi-shot 3D human motion dataset, we introduce ms-Motion by processing existing public 3D human datasets with multiple camera settings and ground truth human and camera parameters, specifically AIST [60] and Human3.6M (H3.6M) [61]. In our construction pipeline, we randomly separate each original one-shot video into several clips. Then, we choose each clip from different shots and concatenate them together as one video recorded by multiple shots. For example, AIST provides each video with eight cameras C0, C1, ..., C7 from different view point and we choose a video and split it into 5 clips at t0, t1, ..., t4. For frames in these separated clips, we choose frames shot by a random camera for each clip and combine five clips as one multi-shot video. Therefore, we construct a multi-shot version of AIST and H3.6M, which are named ms-AIST and ms-H3.6M subsets. Then we combine them and name this new dataset ms-Motion. The detailed statistics of ms-Motion are shown in Tab. 2. We do not compare with other existing 3D human datasets as they contain limited number of multi-shot videos. + +Benchmark Evaluation Protocol. To evaluate the performance of our proposed methods on multi-shot videos, our + +
DatasetModels2-Shot3-Shot4-Shot
PA.↓WA.↓RTE↓ROE↓Jitter↓F.S.↓PA.↓WA.↓RTE↓ROE↓Jitter↓F.S.↓PA↓WA↓RTE↓ROE↓Jitter↓F.S.↓
ms-AISTSLAHMR [2023]72.34341.759.6296.2662.593.2680.35510.7710.33101.3672.394.4390.32803.6912.11104.0780.3716.52
WHAM [2024]65.34336.824.3984.4825.242.7578.68451.325.1489.8424.062.99102.88603.935.5790.0726.293.62
GVHMR [2024]60.72231.366.2096.5834.877.6570.33357.167.5599.6934.469.4283.77563.178.96104.5335.679.78
Ours36.82121.352.5669.2333.272.6638.52141.383.6467.7135.073.5539.63161.524.5570.3139.494.09
ms-H3.6MSLAHMR [2023]80.67352.6116.67111.9737.807.9397.15562.1016.91118.4652.239.96107.90748.5817.85116.7265.1511.58
WHAM [2024]71.32313.5811.4182.4218.405.0979.51423.9812.3684.8518.875.0390.50512.6612.9190.3418.405.69
GVHMR [2024]64.63254.306.9481.9318.458.8080.79296.7485.2558.2618.3610.6285.19471.539.1291.6319.4710.65
Ours40.52132.133.6553.3919.054.1745.35145.365.3358.2617.354.6250.59147.626.2061.2219.775.12
+ +Table 1. Quantitative comparison of different HMR methods on ms-Motion dataset. We record the results for ms-AIST and ms-H3.6M separately. PA and WA means PA-MPJPE and WA-MPJPE respectively, while F.S. is Foot Sliding. Our proposed method has achieved the best performance in PA-MPJPE, WA-MPJPE, RTE and ROE across ms-Motion among these methods. + +
DatasetDuration(s)VideosFPSMax LengthMin LengthShots
ms-Motion23.76003014783142, 3, 4
+ +target is to evaluate metrics for accurately reflecting the performance on videos with shot transitions. To this end, we use Root Orientation Error (a.k.a. ROE in deg) to measure the performance of the proposed method on human orientation alignment across different shots. Besides, we use Root Translation Error (a.k.a. RTE in $m$ ) to assess the performance of the proposed method on global trajectory recovery. Jitter $\left( \frac{10m}{fps^3} \right)$ is also used to evaluate the stability of recovered human pose from multi-shot videos. We also include foot sliding $(cm)$ , the averaged displacement of foot vertices during contact with the ground, to assess the precision of recovered motion in the world coordinates [28]. + +# 5. Experiment + +# 5.1. Datasets and Metrics + +Evaluation Datasets. To evaluate the performance of our proposed pipeline for multi-shot videos, we use ms-Motion dataset and EMDB-1 dataset [73] with self-added noise for the evaluation of ablation study. For camera trajectory estimation, we use EMDB-1 and EMDB-2 split [73] as they contain the GT moving camera trajectory. Our self-created dataset contains 600 multi-shot videos, 42.7K frames, totaling 237 minutes. EMDB-1 split contains 17 video sequences totaling 13.5 minutes and EMDB-2 split contains 25 sequences in a total of 24.0 minutes. + +Evaluation Metrics. For shot detection we use Recall, Precision and F1 Score as evaluation metrics. For 3D human pose estimation-related tasks, we use ROE, RTE, jitter, and foot-sliding for evaluating the human motion recovery results on multi-shot videos. For the ablation study of our proposed pipeline, we evaluate the Procrustes-aligned Mean Per Joint Position Error (a.k.a. PA-MPJPE) and Per Vertex Error (a.k.a. PVE) as additional metrics besides previous mentioned ones. For camera pose estimation, we use absolute trajectory error (a.k.a. ATE) $(m)$ , Relative Pose Error (a.k.a. RPE) rotation $(deg)$ , and RPE translation $(m)$ . + +Table 2. Statistics of the ms-Motion dataset. By shots, we mean the number of shot transitions in a single video. + +
Methodsms-Motion
Recall↑Precision↑F1 Score↑
Scenes Detect (SD) [83]0.740.720.70
SD+Bbox Tracking (Bbox)0.880.850.86
SD+Bbox+Pose Tracking0.960.880.92
+ +Table 3. Comparison between difference shot detection algorithms. We evaluate our shot transition detector on our proposed multi-shot video human motion dataset ms-Motion. + +
MethodsATE↓RPE Trans.↓RPE Rot.↓
DPVO (w/o mask)0.481.851.06
Masked DPVO0.481.570.97
LEAP-VO (w/o mask)0.500.930.97
Masked LEAP-VO0.510.920.95
+ +Table 4. Camera tracking results on EMDB-1 [73]. Our method has achieved $\sim 50\%$ ↓ on RPE trans. than that of the original DPVO and perform the best in RPE rot. metrics. + +# 5.2. Implementation Details + +The ms-HMR, the trajectory, and foot sliding refiner are trained on the AMASS [78], 3DPW [79], Human3.6M [61], and BEDLAM [90] datasets, and evaluated on EMDB and our ms-Motion. During training, we introduce random rotational noise (ranging from 0 to 1 radian) along the y-axis to the root pose $\Gamma$ and random noise to the body pose $\theta$ at random positions to simulate the inaccuracies of pre-estimated human motions caused by shot transitions in multi-shot videos. This strategy enables the network to robustly recover smooth and consistent human motion from noisy initial parameters. The benchmark test results were obtained after training for 80 epochs on one NVIDIA-A100 GPU. + +# 5.3. Main Results: Comparison of Global Human Motion Recovery Results on the Benchmark + +We compare our proposed method HumanMM with several state-of-the-art HMR methods (SLAHMR [26], WHAM [28] and GVHMR [29]) on our proposed benchmark ms-Motion. As illustrated in Tab. 1, our proposed method has achieved the best performance for PA&WAMPJPE, RTE and ROE through videos with all numbers of shots across ms-AIST and ms-H3.6M, indicating that our method reconstructs both the global human motion and orientations in the world coordinates more accurately and robustly. For the foot sliding metric, our method also performs as the best on ms-H3.6M across all numbers of shots. + +# 5.4. Ablation Studies + +Human-centric Scene Shot Boundary Detection Evaluation. To evaluate the performance of our proposed Shot Transition Detector, we test the algorithm on our proposed multi-shot human motion recovery benchmark and compare the output frame list of shot transitions with the ground truth (GT) of our dataset. As shown in Tab. 3, by apply + +![](images/a8e7df8b005299e7c51581efca7768e5e9d04f98d298dc31fcbcc2a35bfae541.jpg) +Figure 7. Qualitative comparison of different HMR methods on ms-Motion dataset. The side view of the rendered mesh for input mutli-shot video is shown in (a), while the top view is shown in (c). We also draw the comparison of the human trajectory as shown in (b). Our method is the most similar as GT in both rendered motion and trajectories among these methods. + +
Methods|PA-MPJE↓|MPJPE↓WA-MPJE↓W-MPJPE↓PVE↓ACCEL↓RTE↓ROE↓F.S.↓
Baseline (Concat)106.48141.67273.15553.67122.156.1410.8691.5514.91
w/o ms-HMR78.24101.52246.42436.5785.775.873.8950.633.54
w/o OAM73.5692.13243.65425.1879.645.676.6176.744.45
w/o traj. pred.50.4983.68231.75432.1775.775.755.5247.684.96
w/o traj. ref.50.4983.68198.58397.6575.775.234.0647.687.84
HumanMM (Ours)50.4983.68194.77393.2175.775.163.5447.683.28
+ +ing the proposed finer granularity shot detection methods, the number of recall, precision, and F1 score all increases consistently. The combination of three steps (ScenesDetect, bbox tracking, and pose tracking) has achieved 0.96, 0.88, and 0.92 on the recall, precision, and F1 score, respectively, which indicates a comparable performance in shot boundary detection. Besides, as can be seen in the results, the latter two steps of shot detection contribute to the fine-grained final results significantly and jointly. + +Key modules in the Proposed Method. We compare our methods with four variants on EMDB with noise dataset, as shown in Tab. 5, ms-HMR is the key component for the improvement in PA-MPJPE and PVE, which indicates a more accurate modeling of the whole motion sequence. This design serves as a recovery module to estimate some invisible body parts in some shots. Additionally, the orientation alignment module (OAM, in Sec. 3.4) is also a critical block for accurate human orientation estimation, indicated by the metric ROE. This module helps to model the global human motion between shots. For foot sliding, the results in Tab. 5 also show that the trajectory refiner (Sec. 3.5) in our method helps mitigate the foot sliding issue. + +Comparison on Camera Trajectory Estimation. To evaluate the performance of our proposed camera trajectory estimation method Masked LEAP-VO, we evaluate the camera trajectory accuracy on EMDB-1 and EMDB-2. For more convenient comparison, we introduce two baselines, DPVO [76], which has been widely used in HMR methods such as WHAM [28] and GVHMR [29], and LEAP-VO [77]. To provide more intuition about the insights of masking dynamic humans in the video, we also implement a variant, Masked DPVO, by applying SAM at the patchify stage of DPVO to exclude patches containing human pixels. As shown in Tab. 4 and Tab. 6, compared with baseline methods, our key design of masking dynamic human subjects improves the result in both RPE Translation and + +Table 5. Ablation studies on different combinations of HumanMM modules. We evaluate the contributions of each key module on EMDB-1. + +
Methods\( \left| \mathrm{{ATE}}\right| \downarrow \)RPE Trans. \( \downarrow \)RPE Rot. \( \downarrow \)
DPVO (w/o mask)0.481.071.26
Masked DPVO0.500.861.21
LEAP-VO (w/o mask)0.500.831.21
Masked LEAP-VO0.490.831.19
+ +Table 6. Camera tracking results on EMDB-2 [73]. Masked LEAP-VO performs better in both RPE Trans. and RPE Rot. metrics. + +
MethodsWA-MPJE↓W-MPJE↓RTE↓Jitter↓F.S.↓
DPVO305.40117.105.1017.903.50
Masked DPVO303.90116.404.1017.403.50
LEAP-VO284.10112.803.1016.303.50
Masked LEAP-VO283.70112.703.1016.303.50
+ +Table 7. Global motion recovery results on EMDB-2 [73]. We input estimated camera parameters from different methods into GVHMR for comparison on HMR. + +RPE Rotation while maintaining competitive ATE. This result indicates the effectiveness of the design of masking dynamic human subjects in the process of camera trajectory estimation. Compared with the DPVO baseline, our method achieves $\sim 50\%$ $\downarrow$ RPE translation on EMDB-1. In addition, we run GVHMR [29] on EMDB-2 with different estimated camera trajectories. The results is shown in Tab. 7, which further illustrates the effectiveness of our method. + +# 6. Conclusion and Discussion + +Conclusion. In this paper, we introduce HumanMM, the first framework designed for human motion recovery from multi-shot videos in world coordinates. HumanMM addresses the challenges inherent in multi-shot videos by incorporating three key components: an enhanced camera trajectory estimation method called masked LEAP-VO, a human motion alignment module that ensures consistency across different shots, and a post-processing module for seamless motion integration. Extensive experiments demonstrate that HumanMM outperforms existing human motion recovery methods across various benchmarks, achieving state-of-the-art accuracy on our newly created multi-shot human motion dataset, ms-Motion. + +Limitations and Future Work. While HumanMM represents an dvancement in human motion recovery from multishot videos in world coordinates, its performance may decline when faced with an excessive number of shot transitions. Despite these challenges, HumanMM provides a solid baseline for human motion recovery from multi-shot videos and can be employed in annotating markerless human motion datasets. Our newly introduced dataset, msMotion, offers a valuable benchmark for evaluating general human motion recovery methods in world coordinates, especially regarding their performance on multi-shot videos. Based on the proposed method, our future work aims to enlarge the related datasets for larger-scale motion databases. + +# Acknowledgement + +This research was funded through the National Key Research and Development Program of China Project No. 2022YFB36066), in part by the Shenzhen Science and Technology Project under Grant (KJZD20240903103210014). The author team would also like to convey sincere thanks to Ms. Yaxin Chen from IDEA Research for the expressive dance motion used in the demo presentation. + +# References + +[1] Jingbo Wang, Yu Rong, Jingyuan Liu, Sijie Yan, Dahua Lin, and Bo Dai. Towards diverse and natural scene-aware 3d human motion synthesis. In CVPR, pages 20428-20437, 2022. 2 +[2] Zeqi Xiao, Tai Wang, Jingbo Wang, Jinkun Cao, Wenwei Zhang, Bo Dai, Dahua Lin, and Jiangmiao Pang. Unified human-scene interaction via prompted chain-of-contacts. In ICLR, 2024. 2 +[3] Chuan Guo, Xinxin Zuo, Sen Wang, and Li Cheng. Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In ECCV, pages 580-597, 2022. 2 +[4] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. NeurIPS, 2024. +[5] Liang Pan, Jingbo Wang, Buzhen Huang, Junyu Zhang, Haofan Wang, Xu Tang, and Yangang Wang. Synthesizing physically plausible human motions in 3d scenes. In 3DV, 2024. +[6] Jingbo Wang, Ye Yuan, Zhengyi Luo, Kevin Xie, Dahua Lin, Umar Iqbal, Sanja Fidler, and Sameh Khamis. Learning human dynamics in autonomous driving scenarios. In ICCV, pages 20739-20749, 2023. 2 +[7] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. In ECCV, pages 358–374, 2022. 2 +[8] Mathis Petrovich, Michael J Black, and Gül Varol. Temos: Generating diverse human motions from textual descriptions. In ECCV, pages 480-497, 2022. +[9] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. IEEE TPAMI, 2024. +[10] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model. In ICLR, 2022. +[11] Zan Wang, Yixin Chen, Tengyu Liu, Yixin Zhu, Wei Liang, and Siyuan Huang. Humanise: Language-conditioned human motion generation in 3d scenes. NeurIPS, pages 14959-14971, 2022. +[12] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, and Gang Yu. Executing your commands via motion diffusion in latent space. In CVPR, pages 18000-18010, 2023. + +[13] Ye Yuan, Jiaming Song, Umar Iqbal, Arash Vahdat, and Jan Kautz. Physdiff: Physics-guided human motion diffusion model. In ICCV, pages 16010-16021, 2023. +[14] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Yong Zhang, Hongwei Zhao, Hongtao Lu, Xi Shen, and Ying Shan. Generating human motion from textual descriptions with discrete representations. In CVPR, pages 14730-14740, 2023. +[15] Yonatan Shafir, Guy Tevet, Roy Kapon, and Amit H Bermano. Human motion diffusion as a generative prior. In ICLR, 2024. +[16] Korrawe Karunratanakul, Konpat Preechakul, Supasorn Suwajanakorn, and Siyu Tang. Guided motion diffusion for controllable human motion synthesis. In CVPR, pages 2151-2162, 2023. +[17] Ling-Hao Chen, Shunlin Lu, Wenxun Dai, Zhiyang Dou, Xuan Ju, Jingbo Wang, Taku Komura, and Lei Zhang. Pay attention and move better: Harnessing attention for interactive motion generation and training-free editing, 2024. +[18] Zeqi Xiao, Tai Wang, Jingbo Wang, Jinkun Cao, Wenwei Zhang, Bo Dai, Dahua Lin, and Jiangmiao Pang. Unified human-scene interaction via prompted chain-of-contacts. In ICLR, 2024. +[19] Yiming Xie, Varun Jampani, Lei Zhong, Deqing Sun, and Huaizu Jiang. Omnicontrol: Control any joint at any time for human motion generation. In ICLR, 2024. +[20] Shunlin Lu, Ling-Hao Chen, Ailing Zeng, Jing Lin, Ruimao Zhang, Lei Zhang, and Heung-Yeung Shum. Humantomato: Text-aligned whole-body motion generation. ICML, 2024. 2 +[21] Wenxun Dai, Ling-Hao Chen, Jingbo Wang, Jinpeng Liu, Bo Dai, and Yansong Tang. Motionlcm: Real-time controllable motion generation via latent consistency model. ECCV, 2024. +[22] Ling-Hao Chen, Jiawei Zhang, Yewen Li, Yiren Pang, Xiaobo Xia, and Tongliang Liu. Humanmac: Masked motion completion for human motion prediction. In ICCV, pages 9544-9555, 2023. +[23] Shunlin Lu, Jingbo Wang, Zeyu Lu, Ling-Hao Chen, Wenxun Dai, Junting Dong, Zhiyang Dou, Bo Dai, and Ruimao Zhang. Scamo: Exploring the scaling law in autoregressive motion generation model. arXiv preprint arXiv:2412.14559, 2024. +[24] Wenxun Dai, Ling-Hao Chen, Yufei Huo, Jingbo Wang, Jinpeng Liu, Bo Dai, and Yansong Tang. Real-time controllable motion generation via latent consistency model. 2025. 2 +[25] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa*, and Jitendra Malik*. Humans in 4D: Reconstructing and tracking humans with transformers. In ICCV, 2023. 2, 3 +[26] Vickie Ye, Georgios Pavlakos, Jitendra Malik, and Angjoo Kanazawa. Decoupling human and camera motion from videos in the wild. In CVPR, 2023. 2, 3, 7 +[27] Yufu Wang, Ziyun Wang, Lingjie Liu, and Kostas Daniilidis. Tram: Global trajectory and motion of 3d humans from inthe-wild videos. ECCV, 2024. 2, 3 +[28] Soyong Shin, Juyong Kim, Eni Halilaj, and Michael J. Black. WHAM: Reconstructing world-grounded humans with accurate 3D motion. In CVPR, 2024. 6, 7, 8 + +[29] Zehong Shen, Huaijin Pi, Yan Xia, Zhi Cen, Sida Peng, Zechen Hu, Hujun Bao, Ruizhen Hu, and Xiaowei Zhou. World-grounded human motion recovery via gravity-view coordinates. In ACM SIGGRAPH Asia, 2024. 2, 3, 4, 5, 7, 8 +[30] Angjoo Kanazawa, Jason Y. Zhang, Panna Felsen, and Jitendra Malik. Learning 3d human dynamics from video. In CVPR, 2019. 3 +[31] Muhammed Kocabas, Ye Yuan, Pavlo Molchanov, Yunrong Guo, Michael J. Black, Otmar Hilliges, Jan Kautz, and Umar Iqbal. Pace: Human and motion estimation from in-the-wild videos. In 3DV, 2024. 2 +[32] Ling-Hao Chen, Shunlin Lu, Ailing Zeng, Hao Zhang, Benyou Wang, Ruimao Zhang, and Lei Zhang. Motionllm: Understanding human behaviors from human motions and videos. arXiv preprint arXiv:2405.20340, 2024. 2 +[33] Matthias Plappert, Christian Mandery, and Tamim Asfour. Learning a bidirectional mapping between human whole-body motion and natural language using deep recurrent neural networks. RAS, 109:13-26, 2018. 2 +[34] Fangzhou Hong, Mingyuan Zhang, Liang Pan, Zhongang Cai, Lei Yang, and Ziwei Liu. Avatarclip: Zero-shot text-driven generation and animation of 3d avatars. ACM SIGGRAPH, 2022. 2 +[35] Nikos Athanasiou, Mathis Petrovich, Michael J Black, and Gül Varol. Teach: Temporal action composition for 3d humans. In 3DV, pages 414-423, 2022. +[36] Rishabh Dabral, Muhammad Hamza Mughal, Vladislav Golyanik, and Christian Theobalt. Mofusion: A framework for denoising-diffusion-based motion synthesis. In CVPR, pages 9760-9770, 2023. +[37] Mingyuan Zhang, Xinying Guo, Liang Pan, Zhongang Cai, Fangzhou Hong, Huirong Li, Lei Yang, and Ziwei Liu. Remodiffuse: Retrieval-augmented motion diffusion model. In ICCV, 2023. +[38] Weilin Wan, Zhiyang Dou, Taku Komura, Wenping Wang, Dinesh Jayaraman, and Lingjie Liu. Tlcontrol: Trajectory and language control for human motion synthesis. ECCV, 2024. +[39] Chuan Guo, Yuxuan Mu, Muhammad Gohar Javed, Sen Wang, and Li Cheng. Momask: Generative masked modeling of 3d human motions. In CVPR, pages 1900-1910, 2024. +[40] Jinpeng Liu, Wenxun Dai, Chunyu Wang, Yiji Cheng, Yansong Tang, and Xin Tong. Plan, posture and go: Towards open-world text-to-motion generation. ECCV, 2024. +[41] Bo Han, Hao Peng, Minjing Dong, Yi Ren, Yixuan Shen, and Chang Xu. Amd: Autoregressive motion diffusion. In AAAI, pages 2022-2030, 2024. +[42] Zhenyu Xie, Yang Wu, Xuehao Gao, Zhongqian Sun, Wei Yang, and Xiaodan Liang. Towards detailed text-to-motion synthesis via basic-to-advanced hierarchical diffusion model. In AAAI, pages 6252–6260, 2024. +[43] Wenyang Zhou, Zhiyang Dou, Zeyu Cao, Zhouyingcheng Liao, Jingbo Wang, Wenjia Wang, Yuan Liu, Taku Komura, Wenping Wang, and Lingjie Liu. Emdm: Efficient motion diffusion model for fast, high-quality motion generation. ECCV, 2024. + +[44] Mathis Petrovich, Or Litany, Umar Iqbal, Michael J Black, Gul Varol, Xue Bin Peng, and Davis Rempe. Multi-track timeline control for text-driven 3d human motion generation. In CVPRW, pages 1911-1921, 2024. +[45] Boxun Hu, Mingze Xia, Ding Zhao, and Guanlin Wu. Mona: Moving object detection from videos shot by dynamic camera, 2025. +[46] Zan Wang, Yixin Chen, Baoxiong Jia, Puhao Li, Jinlu Zhang, Jingze Zhang, Tengyu Liu, Yixin Zhu, Wei Liang, and Siyuan Huang. Move as you say interact as you can: Language-guided human motion generation with scene affordance. In CVPR, pages 433-444, 2024. +[47] Zhongfei Qing, Zhongang Cai, Zhitao Yang, and Lei Yang. Story-to-motion: Synthesizing infinite and controllable character animation from long text, 2023. 2 +[48] Hui En Pang, Zhongang Cai, Lei Yang, Tianwei Zhang, and Ziwei Liu. Benchmarking and analyzing 3d human pose and shape estimation beyond algorithms. In NeurIPS, 2022. 2 +[49] Gyeongsik Moon, Hongsuk Choi, and Kyoung Mu Lee. Neuralannnot: Neural annotator for 3d human mesh training sets. In CVPRW, 2022. +[50] Gyeongsik Moon, Hongsuk Choi, Sanghyuk Chun, Jiyoung Lee, and Sangdoo Yun. Three recipes for better 3d pseudogots of 3d human mesh estimation in the wild. In CVPRW, 2023. +[51] Hongwei Yi, Hualin Liang, Yifei Liu, Qiong Cao, Yandong Wen, Timo Bolkart, Dacheng Tao, and Michael J Black. Generating holistic 3d human motion from speech. In CVPR, pages 469-480, 2023. 2 +[52] C.J. Bowen and R. Thompson. Grammar of the Edit. Focal Press, 2013. 2, 3 +[53] Mathis Petrovich, Or Litany, Umar Iqbal, Michael J. Black, Gül Varol, Xue Bin Peng, and Davis Rempe. Multi-track timeline control for text-driven 3d human motion generation. In CVPRW, 2024. 2 +[54] Yuhong Zhang, Jing Lin, Ailing Zeng, Guanlin Wu, Shunlin Lu, Yurong Fu, Yuanhao Cai, Ruimao Zhang, Haoqian Wang, and Lei Zhang. Motion- $\mathbf{x} + +$ : A large-scale multimodal 3d whole-body human motion dataset. arXiv preprint arXiv:2501.05098, 2025. 2 +[55] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In CVPR, pages 5152-5161, 2022. 2 +[56] Georgios Pavlakos, Jitendra Malik, and Angjoo Kanazawa. Human mesh recovery from multiple shots. In CVPR, 2022. 2, 3 +[57] Peng Wu, Xiankai Lu, Jianbing Shen, and Yilong Yin. Clip fusion with bi-level optimization for human mesh reconstruction from monocular videos. In ACM MM, page 105-115, New York, NY, USA, 2023. Association for Computing Machinery. +[58] Kuan-Chieh Wang, Zhenzhen Weng, Maria Xenochristou, Joao Pedro Araujo, Jeffrey Gu, C Karen Liu, and Serena Yeung. Nemo: 3d neural motion fields from multiple video instances of the same action. In CVPR, 2023. + +[59] Fabien Baradel, Thibault Groueix, Philippe Weinzaepfel, Romain Brégier, Yannis Kalantidis, and Grégory Rogez. Leveraging mocap data for human mesh recovery. In 3DV, pages 586-595, 2021. 2 +[60] Ruilong Li, Shan Yang, David A. Ross, and Angjoo Kanazawa. Ai choreographer: Music conditioned 3d dance generation with aist++, 2021. 3, 6 +[61] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE TPAMI, 36(7):1325-1339, 2014. 3, 6, 7 +[62] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. ACM TOG, 36(6), 2017. 3 +[63] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J. Black. Keep it SMPL: Automatic estimation of 3D human pose and shape from a single image. In Computer Vision - ECCV 2016. Springer International Publishing, 2016. +[64] Anurag* Arnab, Carl* Doersch, and Andrew Zisserman. Exploiting temporal context for 3d human pose estimation in the wild. In CVPR, 2019. +[65] Ahmed A. A. Osman, Timo Bolkart, and Michael J. Black. STAR: Sparse Trained Articulated Human Body Regressor, page 598-613. Springer International Publishing, 2020. +[66] Yinghao Huang, Federica Bogo, Christoph Lassner, Angjoo Kanazawa, Peter V. Gehler, Javier Romero, Ijaz Akhter, and Michael J. Black. Towards accurate marker-less human shape and pose estimation over time. In 3DV, 2017. 3 +[67] Angjoo Kanazawa, Michael J. Black, David W. Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In CVPR, 2018. 3 +[68] Muhammed Kocabas, Nikos Athanasiou, and Michael J. Black. Vibe: Video inference for human body pose and shape estimation. In CVPR, 2020. +[69] Nikos Kolotouros, Georgios Pavlakos, and Kostas Dani-ilidis. Convolutional mesh regression for single-image human shape reconstruction. In CVPR, 2019. +[70] Zhengyi Luo, S. Alireza Golestaneh, and Kris M. Kitani. 3d human motion estimation via motion compression and refinement. In ACCV, 2020. 3 +[71] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS. Curran Associates, Inc., 2017. 3 +[72] Chun-Hao P. Huang, Hongwei Yi, Markus Höschle, Matvey Safroshkin, Tsvetelina Alexiadis, Senya Polikovsky, Daniel Scharstein, and Michael J. Black. Capturing and inferring dense full-body human-scene contact. In CVPR, pages 13274-13285, 2022. 3 +[73] Manuel Kaufmann, Jie Song, Chen Guo, Kaiyue Shen, Tianjian Jiang, Chengcheng Tang, Juan Jose Zárate, and Otmar Hilliges. EMDB: The Electromagnetic Database of Global 3D Human Pose and Shape in the Wild. In ICCV, 2023. 3, 7, 8 +[74] Xinyu Yi, Yuxiao Zhou, and Feng Xu. Transpose: Real-time 3d human translation and pose estimation with six inertial sensors. ACM TOG, 40(4), 2021. 3 + +[75] Zachary Teed and Jia Deng. Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. In NeurIPS, pages 16558-16569. Curran Associates, Inc., 2021. 3 +[76] Zachary Teed, Lahav Lipson, and Jia Deng. Deep patch visual odometry. NeurIPS, 2023. 3, 8 +[77] Weirong Chen, Le Chen, Rui Wang, and Marc Pollefeys. Leap-vo: Long-term effective any point tracking for visual odometry. In CVPR, 2024. 3, 4, 8 +[78] Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, and Michael J. Black. AMASS: Archive of motion capture as surface shapes. In ICCV, pages 5442-5451, 2019. 3, 7 +[79] Timo von Marcard, Roberto Henschel, Michael Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In ECCV, 2018. 3, 7 +[80] Yufei Xu, Jing Zhang, Qiming Zhang, and Dacheng Tao. ViTPose: Simple vision transformer baselines for human pose estimation. In NeurIPS, 2022. 4 +[81] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 4, 6 +[82] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multiperson linear model. ACM TOG, 34(6):248:1-248:16, 2015. 4 +[83] Qingqiu Huang, Yu Xiong, Anyi Rao, Jiaze Wang, and Dahua Lin. Movienet: A holistic dataset for movie understanding. In ECCV, 2020. 4, 7 +[84] MMTracking Contributors. MMTracking: OpenMMLab video perception toolbox and benchmark. https://github.com/open-mmlab/mmtracking, 2020.4 +[85] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. CoTracker: It is better to track together. In ECCV, 2024. 4 +[86] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. In ICCV, 2023. 5 +[87] Martin A. Fischler and Robert C. Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Commun. ACM, 24(6):381-395, 1981. 5 +[88] Jie Yang, Ailing Zeng, Shilong Liu, Feng Li, Ruimao Zhang, and Lei Zhang. Explicit box detection unifies end-to-end multi-person pose estimation. In ICLR, 2023. 6 +[89] Richard Hartley and Andrew Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, USA, 2 edition, 2003. 6 +[90] Michael J. Black, Priyanka Patel, Joachim Tesch, and Jinlong Yang. BEDLAM: A synthetic dataset of bodies exhibiting detailed lifelike animated motion. In CVPR, pages 8726-8737, 2023. 7 \ No newline at end of file diff --git a/2025/HumanMM_ Global Human Motion Recovery from Multi-shot Videos/images.zip b/2025/HumanMM_ Global Human Motion Recovery from Multi-shot Videos/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ebfff732122378fd1aa677211a89b30cde97b72c --- /dev/null +++ b/2025/HumanMM_ Global Human Motion Recovery from Multi-shot Videos/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a58ea7d1eb2037bb8c34f3a9e8704250317194288b6f80faf9f25f8dacdeebef +size 585456 diff --git a/2025/HumanMM_ Global Human Motion Recovery from Multi-shot Videos/layout.json b/2025/HumanMM_ Global Human Motion Recovery from Multi-shot Videos/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..df2f96398ecfaf3b71a1a784f043a7f7a98d5296 --- /dev/null +++ b/2025/HumanMM_ Global Human Motion Recovery from Multi-shot Videos/layout.json @@ -0,0 +1,10445 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 88, + 86, + 521, + 102 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 86, + 521, + 102 + ], + "spans": [ + { + "bbox": [ + 88, + 86, + 521, + 102 + ], + "type": "text", + "content": "HumanMM: Global Human Motion Recovery from Multi-shot Videos" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 78, + 125, + 530, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 125, + 530, + 155 + ], + "spans": [ + { + "bbox": [ + 78, + 125, + 530, + 155 + ], + "type": "text", + "content": "Yuhong Zhang" + }, + { + "bbox": [ + 78, + 125, + 530, + 155 + ], + "type": "inline_equation", + "content": "^{1,2,\\dagger*}" + }, + { + "bbox": [ + 78, + 125, + 530, + 155 + ], + "type": "text", + "content": " Guanlin Wu" + }, + { + "bbox": [ + 78, + 125, + 530, + 155 + ], + "type": "inline_equation", + "content": "^{2,3,\\ddagger*}" + }, + { + "bbox": [ + 78, + 125, + 530, + 155 + ], + "type": "text", + "content": " Ling-Hao Chen" + }, + { + "bbox": [ + 78, + 125, + 530, + 155 + ], + "type": "inline_equation", + "content": "^{1,2,\\ddagger}" + }, + { + "bbox": [ + 78, + 125, + 530, + 155 + ], + "type": "text", + "content": " Zhuokai Zhao" + }, + { + "bbox": [ + 78, + 125, + 530, + 155 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 78, + 125, + 530, + 155 + ], + "type": "text", + "content": " Jing Lin" + }, + { + "bbox": [ + 78, + 125, + 530, + 155 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 78, + 125, + 530, + 155 + ], + "type": "text", + "content": " Xiaoke Jiang" + }, + { + "bbox": [ + 78, + 125, + 530, + 155 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 78, + 125, + 530, + 155 + ], + "type": "text", + "content": " Jiamin Wu" + }, + { + "bbox": [ + 78, + 125, + 530, + 155 + ], + "type": "inline_equation", + "content": "^{2,5}" + }, + { + "bbox": [ + 78, + 125, + 530, + 155 + ], + "type": "text", + "content": " Zhuoheng Li" + }, + { + "bbox": [ + 78, + 125, + 530, + 155 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 78, + 125, + 530, + 155 + ], + "type": "text", + "content": " Hao Frank Yang" + }, + { + "bbox": [ + 78, + 125, + 530, + 155 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 78, + 125, + 530, + 155 + ], + "type": "text", + "content": " Haoqian Wang" + }, + { + "bbox": [ + 78, + 125, + 530, + 155 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 78, + 125, + 530, + 155 + ], + "type": "text", + "content": " Lei Zhang" + }, + { + "bbox": [ + 78, + 125, + 530, + 155 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 85, + 156, + 526, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 156, + 526, + 168 + ], + "spans": [ + { + "bbox": [ + 85, + 156, + 526, + 168 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 85, + 156, + 526, + 168 + ], + "type": "text", + "content": "Tsinghua University " + }, + { + "bbox": [ + 85, + 156, + 526, + 168 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 85, + 156, + 526, + 168 + ], + "type": "text", + "content": "IDEA Research " + }, + { + "bbox": [ + 85, + 156, + 526, + 168 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 85, + 156, + 526, + 168 + ], + "type": "text", + "content": "Johns Hopkins University " + }, + { + "bbox": [ + 85, + 156, + 526, + 168 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 85, + 156, + 526, + 168 + ], + "type": "text", + "content": "University of Chicago " + }, + { + "bbox": [ + 85, + 156, + 526, + 168 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 85, + 156, + 526, + 168 + ], + "type": "text", + "content": "HKUST " + }, + { + "bbox": [ + 85, + 156, + 526, + 168 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 85, + 156, + 526, + 168 + ], + "type": "text", + "content": "HKU" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 179, + 171, + 433, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 171, + 433, + 182 + ], + "spans": [ + { + "bbox": [ + 179, + 171, + 433, + 182 + ], + "type": "text", + "content": "{dsyuhong, guanlinwu0930, thu.lhchen}@gmail.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 129, + 182, + 478, + 197 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 182, + 478, + 197 + ], + "spans": [ + { + "bbox": [ + 129, + 182, + 478, + 197 + ], + "type": "text", + "content": "Project page: https://zhangyuhong01.github.io/HumanMM" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 61, + 205, + 549, + 491 + ], + "blocks": [ + { + "bbox": [ + 61, + 205, + 549, + 491 + ], + "lines": [ + { + "bbox": [ + 61, + 205, + 549, + 491 + ], + "spans": [ + { + "bbox": [ + 61, + 205, + 549, + 491 + ], + "type": "image", + "image_path": "5cd68b86bc48e672d8354ecd0927f8ed06c9b3b9c195f60338892f3cf801c111.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 493, + 555, + 550 + ], + "lines": [ + { + "bbox": [ + 54, + 493, + 555, + 550 + ], + "spans": [ + { + "bbox": [ + 54, + 493, + 555, + 550 + ], + "type": "text", + "content": "Figure 1. Recovering a human motion from multi-shot videos. Top: We take two multi-shot table tennis game videos with shot transitions as input. We aim to recover two motions of two athletes (Long MA and Zhendong FAN) from two videos, respectively. The first video is recorded by three shots (“①”, “②”, and “③”), and the second one is recovered by two shots (“④” and “⑤”). Bottom: We recover two motions (Long MA in green and Zhendong FAN in pink), different shots, and camera poses for each multi-shot video. The recovered motion is aligned with the motion in the videos." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 151, + 554, + 200, + 566 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 554, + 200, + 566 + ], + "spans": [ + { + "bbox": [ + 151, + 554, + 200, + 566 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 579, + 297, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 579, + 297, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 579, + 297, + 676 + ], + "type": "text", + "content": "In this paper, we present a novel framework designed to reconstruct long-sequence 3D human motion in the world coordinates from in-the-wild videos with multiple shot transitions. Such long-sequence in-the-wild motions are highly valuable to applications such as motion generation and motion understanding, but are of great challenge to be recovered due to abrupt shot transitions, partial occlusions, and dynamic backgrounds presented in such videos. Existing" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 555, + 556, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 556, + 711 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 556, + 711 + ], + "type": "text", + "content": "methods primarily focus on single-shot videos, where continuity is maintained within a single camera view, or simplify multi-shot alignment in camera space only. In this work, we tackle the challenges by integrating an enhanced camera pose estimation with Human Motion Recovery (HMR) by incorporating a shot transition detector and a robust alignment module for accurate pose and orientation continuity across shots. By leveraging a custom motion integrator, we effectively mitigate the problem of foot sliding and ensure temporal consistency in human pose. Extensive evaluations on our created multi-shot dataset from public 3D human datasets demonstrate the robustness of our method in reconstructing realistic human motion in world coordinates." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 683, + 296, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 683, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 683, + 296, + 713 + ], + "type": "text", + "content": "*Equal contribution, ‡Core contributor, †Corresponding author. Work done by Yuhong Zhang, Guanlin Wu, Ling-Hao Chen, Jing Lin and Jiamin Wu during the internship at IDEA Research." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 296, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 758 + ], + "type": "text", + "content": "1973" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 88, + 136, + 101 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 88, + 136, + 101 + ], + "spans": [ + { + "bbox": [ + 56, + 88, + 136, + 101 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 108, + 295, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 108, + 295, + 240 + ], + "spans": [ + { + "bbox": [ + 54, + 108, + 295, + 240 + ], + "type": "text", + "content": "In recent years, significant advances have been made in 3D human pose estimation, particularly in enhancing the accuracy of human motion recovery (HMR)1 from monocular video sequences. HMR has demonstrated extensive applications in human-AI interaction [1, 2], human motion understanding [3-6], and motion generation [3, 7-24]. While existing methods [25] achieved relatively high performance in recovering human mesh in camera coordinates, estimating human motion in world coordinates remains challenging [26-29] due to inaccurate camera pose estimation and the complexity of reconstructing human motion spatially." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 241, + 296, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 241, + 296, + 337 + ], + "spans": [ + { + "bbox": [ + 55, + 241, + 296, + 337 + ], + "type": "text", + "content": "Most current progress in 3D human motion community mainly benefits from large scale data [25, 27-31], and longsequence videos. These resources enhance estimation accuracy for HMR methods and improve the understanding and generation of longer motion sequences for tasks such as motion understanding [3, 32, 33] and generation [3, 7-20, 34-47], even when annotations are derived from markerless capturing methods like pseudo labels [48-51]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 337, + 296, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 337, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 337, + 296, + 445 + ], + "type": "text", + "content": "A promising approach to enlarge the scale of the motion databases is to estimate human motions from unlimited online videos in a markerless manner. However, many longsequence online videos are recorded with multiple shots, referred to as multi-shot videos" + }, + { + "bbox": [ + 55, + 337, + 296, + 445 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 55, + 337, + 296, + 445 + ], + "type": "text", + "content": ", especially prevalent in domains such as sports broadcasting, talk shows, and concerts. In filmmaking and television live show, a \"shot\" denotes an individual camera view capturing a specific moment or action from a particular vantage point [52]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 445, + 295, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 295, + 564 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 295, + 564 + ], + "type": "text", + "content": "Segmenting multi-shot videos into separate shots inevitably reduces the length of the video sequences, which can be detrimental to tasks that benefit from longer sequences, such as long motion generation [47, 53]. This limitation is highlighted in the existing datasets [54, 55], where the longest clip is less than 20 seconds after segmentation, as shown in Fig. 2. Moreover, focusing exclusively on online single-shot videos diminishes the utilization ratio of available online videos and may negatively impact the diversity of scenarios represented in the created datasets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 564, + 295, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 295, + 638 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 295, + 638 + ], + "type": "text", + "content": "Therefore, how to address the issue of discontinuities caused by shot transitions is notoriously difficult in the community. To resolve this problem, previous works [56-59] have proposed algorithms to address human mesh recovery in a camera space from movies containing shot change between long shots and close-ups." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 319, + 71, + 547, + 188 + ], + "blocks": [ + { + "bbox": [ + 319, + 71, + 547, + 188 + ], + "lines": [ + { + "bbox": [ + 319, + 71, + 547, + 188 + ], + "spans": [ + { + "bbox": [ + 319, + 71, + 547, + 188 + ], + "type": "image", + "image_path": "11082310bcbfb4fec8c0288df95839fd66f45dc6e327ff64fc02a4f0b28c9d25.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 188, + 555, + 243 + ], + "lines": [ + { + "bbox": [ + 313, + 188, + 555, + 243 + ], + "spans": [ + { + "bbox": [ + 313, + 188, + 555, + 243 + ], + "type": "text", + "content": "Figure 2. The comparison between the distribution of sequence lengths in different existing large-scale markerless motion datasets with ours. The " + }, + { + "bbox": [ + 313, + 188, + 555, + 243 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 188, + 555, + 243 + ], + "type": "text", + "content": "-axis and " + }, + { + "bbox": [ + 313, + 188, + 555, + 243 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 188, + 555, + 243 + ], + "type": "text", + "content": "-axis denote the duration time (s) and percentage of video number, respectively. Our dataset (in green) contains more portion of long-sequence videos in general." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 246, + 555, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 246, + 555, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 246, + 555, + 474 + ], + "type": "text", + "content": "However, recovering human motions in world coordinates from multi-shot videos presents two fundamental challenges that remain underexplored. 1) How to align the human motion and orientation in the world coordinates during shot transitions? Ensuring continuity of human orientation and pose across shots is complicated by factors such as partial visibility of human body (e.g. transitioning from long shot to close-up) and changes in human orientation (e.g. two long shots from different viewpoints). These issues, caused by abrupt changes in camera viewpoints, necessitate robust alignment mechanisms. 2) How to reconstruct accurate human motion in world coordinates? Existing approaches employ Simultaneous Localization and Mapping (SLAM) methods to estimate camera parameters, which are then used to project recovered human meshes from camera to world coordinates [26-29]. This process requires highly accurate camera estimation and must address motion consistency and foot sliding in the recovered human motion within the world space." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 474, + 554, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 554, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 554, + 545 + ], + "type": "text", + "content": "Despite these challenges, human motion in multi-shot videos often remain continuous across shots, even as camera viewpoints change. This observation suggests that with appropriate handling of shot transitions and camera motion, it is possible to reconstruct consistent and complete 3D human motions throughout multi-shot videos." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 546, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 556, + 714 + ], + "type": "text", + "content": "In this paper, we propose a novel framework HumanMM, Human Motion recovery from Multi-shot videos, to address these challenges. It integrates human pose estimation across shots with robust camera estimation in the world space. Firstly, we develop a shot transition detector to identify frames with shot transitions. To ensure a more robust camera pose estimation, we introduce an enhanced SLAM method incorporating long-term tracking of feature points and exclusion of moving human from bundle adjustment process. We utilize existing HMR method integrated with our enhanced camera estimation to get the initial human parameters for each separated shot. Subsequently, we implement an alignment module to align human orientation based on stereo calibration and smooth human poses through a" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 645, + 295, + 684 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 645, + 295, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 645, + 295, + 684 + ], + "type": "text", + "content": "1In this paper, the \"human mesh recovery\" refers to recovery in the camera coordinates and the \"human motion recovery\" denotes recovery in the world coordinates. Unless specified otherwise, HMR refers to human motion recovery." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 684, + 295, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 684, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 684, + 295, + 712 + ], + "type": "text", + "content": "2In this paper, a multi-shot video refers to a long-sequence video containing multiple shot transitions. We assume that the camera intrinsics remain consistent across different shots within a multi-shot video." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1974" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "trained multi-shot HMR encoder, which effectively captures the temporal context of human movements across different shots. Finally, after aligning human and camera parameters between shot transitions, we train a motion decoder and a trajectory refiner to smooth the human pose and mitigate issues such as foot sliding, thereby enhancing the overall motion consistency in the reconstructed 3D human motions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 157, + 266, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 157, + 266, + 167 + ], + "spans": [ + { + "bbox": [ + 67, + 157, + 266, + 167 + ], + "type": "text", + "content": "Our contributions can be summarized as follows." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 170, + 296, + 324 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 170, + 295, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 170, + 295, + 192 + ], + "spans": [ + { + "bbox": [ + 55, + 170, + 295, + 192 + ], + "type": "text", + "content": "- We present the first approach to reconstruct human motion from multi-shot videos in world coordinates." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 194, + 296, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 194, + 296, + 263 + ], + "spans": [ + { + "bbox": [ + 55, + 194, + 296, + 263 + ], + "type": "text", + "content": "- We introduce HumanMM, a HMR framework for multi-shot videos. It includes an enhanced camera trajectory estimation method, a human motion alignment module and a motion integrator to ensure accurate and consistent recovery of human pose and orientation in world coordinates across different shots in the whole video." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 265, + 296, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 265, + 296, + 324 + ], + "spans": [ + { + "bbox": [ + 55, + 265, + 296, + 324 + ], + "type": "text", + "content": "- We develop a multi-shot video dataset ms-Motion to evaluate the performance of HMR from multi-shot videos, based on existing public datasets such as AIST [60] and Human3.6M [61]. Extensive experiments on related benchmarks verify the effectiveness of our method." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 338, + 142, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 338, + 142, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 142, + 350 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 358, + 205, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 358, + 205, + 369 + ], + "spans": [ + { + "bbox": [ + 55, + 358, + 205, + 369 + ], + "type": "text", + "content": "2.1. HMR from One-shot Video" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 377, + 295, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 377, + 295, + 413 + ], + "spans": [ + { + "bbox": [ + 55, + 377, + 295, + 413 + ], + "type": "text", + "content": "One-shot videos, captured with a single camera without shot transitions, has been extensively studied within the community for human mesh and motion recovery." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 414, + 295, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 414, + 295, + 484 + ], + "spans": [ + { + "bbox": [ + 55, + 414, + 295, + 484 + ], + "type": "text", + "content": "Human mesh recovery in camera coordinates can be broadly categorized into two approaches: optimization-based methods [62-66] and regression-based methods [30, 67-70]. With the significant advancements of transformer [71], HMR2.0 [25] has surpassed previous methods and benefits several downstream tasks related to HMR." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "type": "text", + "content": "Although there are several previous works tried to recover motions in world coordinates with multi-camera capture system [60, 72] and IMU-based methods [73, 74] and enjoy relatively satisfying results, this setup limits their use for applications of infinite in-the-wild monocular videos. To address this limitation, several attempts [26-29] integrate SLAM into the HMR pipeline by first estimating the camera pose using SLAM methods, e.g. DROID-SLAM [75] or DPVO [76], and then project the recovered human motion from camera to world coordinates. To exclude the inconsistencies caused by dynamic objects, such as moving humans, TRAM [27] modifies DROID-SLAM by incorporating human masking and depth-based distance rescaling. However, DROID-SLAM performs dense bundle adjustment (DBA) on feature maps from downsampled images and selects features based only on two consecutive frames rather than long-term video sequences [75-77]. Consequently, masking significantly reduces the number of informative and consistent features, especially when humans occupy large" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "type": "text", + "content": "portions of the image, leading to inaccuracies. Therefore, developing a SLAM method that retains sufficient and representative features for DBA after masking is important." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 116, + 471, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 116, + 471, + 128 + ], + "spans": [ + { + "bbox": [ + 313, + 116, + 471, + 128 + ], + "type": "text", + "content": "2.2. HMR from Multi-shot Video" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 134, + 553, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 134, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 313, + 134, + 553, + 217 + ], + "type": "text", + "content": "Multiple shots are fundamental elements of cinematic storytelling and live performances, utilizing various camera positions and focal lengths to create immersive and detailed viewing experiences for audiences [52]. However, most marker-based motion capture (MoCap) datasets [60, 72, 73, 78, 79] consist single-shot videos only, resulting in limited research on HMR from multi-shot videos." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 217, + 554, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 554, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 554, + 348 + ], + "type": "text", + "content": "Recovering human motion from multi-shot videos in camera coordinates is already challenging. This is because treating each pose estimation result of each shot separately leads to inconsistencies when combining all estimations, caused by partially or fully invisible human bodies across shot transitions. Pavlakos et al. [56] addresses this issue by focusing on shot changes from long shots to close-ups, which are common in film. They develop smoothness constraints within a temporal Human Mesh and Motion Recovery (t-HMMR) model to infer motions during occlusions caused by shot transitions." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 349, + 554, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 349, + 554, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 349, + 554, + 455 + ], + "type": "text", + "content": "Advancements in HMR methods [29] for single-shot videos in world coordinates have paved the way for extending HMR to multi-shot videos with varying camera viewpoints. However, aligning human orientation, body pose, and translation continuously across multi-shot videos in world coordinates underexplored. Effective alignment is crucial to maintain motion continuity and coherence, especially when dealing with diverse camera perspectives and abrupt transitions between shots." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 456, + 554, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 554, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 554, + 538 + ], + "type": "text", + "content": "In summary, while substantial progress has been made in HMR from single-shot videos, extending these techniques to multi-shot videos requires addressing additional complexities related to camera pose alignment and motion consistency across shot transitions. We address this challenge by proposing a novel pipeline that ensures accurate and continuous 3D HMR from multi-shot monocular videos." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 550, + 370, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 550, + 370, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 550, + 370, + 562 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "In this section, we propose HumanMM to recover human motion from multi-shot videos. The system overview is shown in Fig. 3. Given an input video sequence " + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{V} = \\{I_t\\}_{t=1}^T" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": " of length " + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "inline_equation", + "content": "I_t" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "-th frame, our objective is to recover human motion in world coordinates. We begin by detecting shot transition frames based on human bounding box (a.k.a. bbox) and 2D keypoints (a.k.a. KPTs) through a shot transition detector (Sec. 3.2). For each clipped shot, we initialize the camera pose (camera rotation and camera translation) and recover initial human motion in world coordinates (Sec. 3.3). The initialized SMPL parameters and camera poses are then fed into a hu" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1975" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 72, + 550, + 223 + ], + "blocks": [ + { + "bbox": [ + 58, + 72, + 550, + 223 + ], + "lines": [ + { + "bbox": [ + 58, + 72, + 550, + 223 + ], + "spans": [ + { + "bbox": [ + 58, + 72, + 550, + 223 + ], + "type": "image", + "image_path": "ca2ca7aec029cc764656585c8a8e1546a71db368b30708f80823d8e6d5e2ade0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 224, + 555, + 290 + ], + "lines": [ + { + "bbox": [ + 55, + 224, + 555, + 290 + ], + "spans": [ + { + "bbox": [ + 55, + 224, + 555, + 290 + ], + "type": "text", + "content": "Figure 3. The overview of HumanMM. HumanMM processes multi-shot video sequences by first extracting motion feature such as keypoints and bounding boxes, using ViTPose [80] and image feature using ViT [81]. These features are then segmented into single-shot clips via Shot Transition Detection (Sec. 3.2). Initialized camera (camera rotation " + }, + { + "bbox": [ + 55, + 224, + 555, + 290 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 55, + 224, + 555, + 290 + ], + "type": "text", + "content": " and camera translation " + }, + { + "bbox": [ + 55, + 224, + 555, + 290 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 55, + 224, + 555, + 290 + ], + "type": "text", + "content": ") and human (SMPL) parameters for each shot are estimated using Masked LEAP-VO (Sec. 3.3) and GVHMR [29]. Human orientation is aligned across shots through camera calibration (3.4.1), and ms-HMR (Sec. 3.4.2) ensures consistent pose alignment. Finally, a bi-directional LSTM-based trajectory predictor with trajectory refiner predicts trajectory based on aligned motion and mitigates foot sliding throughout the video." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 293, + 296, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 293, + 296, + 376 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 296, + 376 + ], + "type": "text", + "content": "man motion alignment module (Sec. 3.4), which aligns human orientations via camera calibration based on human 2D KPTs and smooth the human pose by incorporating pose information across different shots. Additionally, it refines the entire motion sequence through whole video using a temporal motion encoder ms-HMR. Finally, we introduce a post-processing module for motion integration (Sec. 3.5)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 383, + 228, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 383, + 228, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 383, + 228, + 396 + ], + "type": "text", + "content": "3.1. Preliminary: 3D Human Model" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 401, + 296, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 401, + 296, + 474 + ], + "spans": [ + { + "bbox": [ + 55, + 401, + 296, + 474 + ], + "type": "text", + "content": "Our method aims to recover motions in world coordinates in the SMPL [82] format, whose pose at frame " + }, + { + "bbox": [ + 55, + 401, + 296, + 474 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 401, + 296, + 474 + ], + "type": "text", + "content": " can be represented as " + }, + { + "bbox": [ + 55, + 401, + 296, + 474 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_t(\\theta_t,\\beta_t,\\Gamma_t,\\tau_t)\\in \\mathbb{R}^{6890\\times 3}" + }, + { + "bbox": [ + 55, + 401, + 296, + 474 + ], + "type": "text", + "content": ". Here, the body pose, body shape, root orientation, and translation are " + }, + { + "bbox": [ + 55, + 401, + 296, + 474 + ], + "type": "inline_equation", + "content": "\\theta_{t}\\in \\mathbb{R}^{23\\times 3}" + }, + { + "bbox": [ + 55, + 401, + 296, + 474 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 401, + 296, + 474 + ], + "type": "inline_equation", + "content": "\\beta_{t}\\in \\mathbb{R}^{10}" + }, + { + "bbox": [ + 55, + 401, + 296, + 474 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 401, + 296, + 474 + ], + "type": "inline_equation", + "content": "\\Gamma_t\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 401, + 296, + 474 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 401, + 296, + 474 + ], + "type": "inline_equation", + "content": "\\tau_{t}\\in \\mathbb{R}^{3}" + }, + { + "bbox": [ + 55, + 401, + 296, + 474 + ], + "type": "text", + "content": ", respectively. We use " + }, + { + "bbox": [ + 55, + 401, + 296, + 474 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_t^{2D}" + }, + { + "bbox": [ + 55, + 401, + 296, + 474 + ], + "type": "text", + "content": " to denote human 2D KPTs at each frame " + }, + { + "bbox": [ + 55, + 401, + 296, + 474 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 401, + 296, + 474 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 479, + 296, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 296, + 491 + ], + "type": "text", + "content": "3.2. Shot Transition Detector For Multi-shot Video" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 498, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 713 + ], + "type": "text", + "content": "Our algorithm begins with shot transition detection in one video. As shown in Fig. 3, the shot transition detector has three key components, scene transition detector, bounding box (a.k.a. bounding box) tracking, and human keypoints tracking. (1) Scene change transition detector: Initially, we employ the SceneDetect [83] algorithm to identify scene changes based on significant variations in the background. However, the SceneDetect fails to detect shot transitions when background changes are unnoticeable, illustrated in Fig. 4. Subsequently, we leverage the following modules to bridge the gap. (2) Bbox tracking for shot transition. As a shot change often accompanies with a sudden change of human subject size, we track humans in a video via mmtracking [84]. Consequently, we compute the Intersection over Union (IoU) between neighbor bboxes and identify a shot transition when the IoU falls smaller than a manually tuned threshold. (3) Human pose tracking for shot transition detection. To achieve a finer granularity, we additionally intro" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 316, + 293, + 399, + 378 + ], + "blocks": [ + { + "bbox": [ + 316, + 293, + 399, + 378 + ], + "lines": [ + { + "bbox": [ + 316, + 293, + 399, + 378 + ], + "spans": [ + { + "bbox": [ + 316, + 293, + 399, + 378 + ], + "type": "image", + "image_path": "c26a99eaab97d284f84c633ae3c919750b0f716898a770539707640218814788.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 332, + 380, + 385, + 388 + ], + "lines": [ + { + "bbox": [ + 332, + 380, + 385, + 388 + ], + "spans": [ + { + "bbox": [ + 332, + 380, + 385, + 388 + ], + "type": "text", + "content": "(a) Scene Change" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 400, + 293, + 476, + 378 + ], + "blocks": [ + { + "bbox": [ + 400, + 293, + 476, + 378 + ], + "lines": [ + { + "bbox": [ + 400, + 293, + 476, + 378 + ], + "spans": [ + { + "bbox": [ + 400, + 293, + 476, + 378 + ], + "type": "image", + "image_path": "96aeb8efd260b8fa739c9d6857c6f8eeda93fc99f2d21f9e1cea00ef87964fbb.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 380, + 466, + 388 + ], + "lines": [ + { + "bbox": [ + 406, + 380, + 466, + 388 + ], + "spans": [ + { + "bbox": [ + 406, + 380, + 466, + 388 + ], + "type": "text", + "content": "(b) Position Change" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 477, + 293, + 551, + 378 + ], + "blocks": [ + { + "bbox": [ + 477, + 293, + 551, + 378 + ], + "lines": [ + { + "bbox": [ + 477, + 293, + 551, + 378 + ], + "spans": [ + { + "bbox": [ + 477, + 293, + 551, + 378 + ], + "type": "image", + "image_path": "35edaa31a1ff346fdad3638262003b6d7ff3eb8e5945f481aa50a7a61b87ec6e.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 487, + 380, + 536, + 388 + ], + "lines": [ + { + "bbox": [ + 487, + 380, + 536, + 388 + ], + "spans": [ + { + "bbox": [ + 487, + 380, + 536, + 388 + ], + "type": "text", + "content": "(c) Pose Change" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 388, + 554, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 554, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 554, + 464 + ], + "type": "text", + "content": "Figure 4. Shot transition detection examples. Examples (a), (b), and (c) illustrate multi-shot scenarios in online videos. (a) shows scene transitions detectable by SceneDetect. (b) illustrates significant position changes undetectable by SceneDetect but resolvable withbbox tracking-based method. (c) shows pose or orientation transition, requiring pose tracking-based methods as they cannot be addressed by either SceneDetect or_bbox tracking." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 469, + 554, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 469, + 554, + 516 + ], + "spans": [ + { + "bbox": [ + 313, + 469, + 554, + 516 + ], + "type": "text", + "content": "duce human 2D KPTs to detect extreme corner shot changes in a video. By thresholding the IoU of corresponding keypoints between neighbor frames, we can accurately identify shot transitions even with subtle human movements." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 517, + 554, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 554, + 554 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 554, + 554 + ], + "type": "text", + "content": "As each separate module cannot identify all kinds of shot transitions, the three modules are jointly used to clip a video into several sub-sequences serially." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 562, + 554, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 562, + 554, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 562, + 554, + 586 + ], + "type": "text", + "content": "3.3. Human Motion and Camera Pose Estimation For Each Shot" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 594, + 554, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 554, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 554, + 653 + ], + "type": "text", + "content": "After obtaining the clipped videos, our next goal is to estimate the camera pose and SMPL parameters in the world coordinates for each clipped video. The estimated camera pose and motions for each shot will be used to construct the whole motion sequence in the next stage (Sec. 3.4)." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": "How to estimate the camera parameters accurately? Our approach for camera parameter calculation is based on a visual odometry (VO) estimation method, LEAP-VO [77]. Utilizing the CoTracker method [85], LEAP-VO estimates the visibility and trajectories of " + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": " selected points by ana" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1976" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "lyzing image gradients across the video sequence. LEAPVO subsequently computes confidence scores for each trajectory, retaining only those with high confidence while discarding trajectories shorter than a predefined threshold. The remaining trajectories undergo bundle adjustment (BA) within a fixed window size to estimate the camera poses." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 144, + 296, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 144, + 296, + 383 + ], + "spans": [ + { + "bbox": [ + 56, + 144, + 296, + 383 + ], + "type": "text", + "content": "However, simply applying LEAP-VO in the camera estimation process is still unsatisfactory in most human-centric scenarios. The primary limitation stems from the dynamic movements of human subjects, which typically occupy a substantial portion of each image in human-centric videos. This dynamic presence introduces noise into the camera pose estimation in world coordinates, as the estimation process relies heavily on the relationship between the camera and the static environment. To address this issue, we propose a Masked LEAP-VO algorithm. Our approach involves inputting the image " + }, + { + "bbox": [ + 56, + 144, + 296, + 383 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 56, + 144, + 296, + 383 + ], + "type": "text", + "content": " and the human bbox at frame " + }, + { + "bbox": [ + 56, + 144, + 296, + 383 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 56, + 144, + 296, + 383 + ], + "type": "text", + "content": " into SAM [86] to generate a human mask. We then assign a visibility value of zero to points within the human mask, effectively excluding these trajectories from the BA process. For clarity, we denote " + }, + { + "bbox": [ + 56, + 144, + 296, + 383 + ], + "type": "inline_equation", + "content": "S_{BA}" + }, + { + "bbox": [ + 56, + 144, + 296, + 383 + ], + "type": "text", + "content": " as the window size of BA, " + }, + { + "bbox": [ + 56, + 144, + 296, + 383 + ], + "type": "inline_equation", + "content": "\\hat{n}" + }, + { + "bbox": [ + 56, + 144, + 296, + 383 + ], + "type": "text", + "content": " denotes the number of filtered point trajectories, and " + }, + { + "bbox": [ + 56, + 144, + 296, + 383 + ], + "type": "inline_equation", + "content": "w_{ij,\\hat{n}}" + }, + { + "bbox": [ + 56, + 144, + 296, + 383 + ], + "type": "text", + "content": " as the normalized weight based on confidence score and visibility. For estimating the camera poses " + }, + { + "bbox": [ + 56, + 144, + 296, + 383 + ], + "type": "inline_equation", + "content": "\\mathbf{G} = \\{\\mathbf{R},\\mathbf{T}\\}" + }, + { + "bbox": [ + 56, + 144, + 296, + 383 + ], + "type": "text", + "content": " of orientation and translation, the reprojection loss function for BA can then be formulated as follows," + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 388, + 304, + 409 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 304, + 409 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 304, + 409 + ], + "type": "interline_equation", + "content": "\\mathbf {G} = \\underset {\\mathbf {G}, d _ {i}, \\hat {n}} {\\arg \\min } \\sum_ {i} \\sum_ {j \\in | i - j | \\leq S} \\sum_ {B A} w _ {i j, \\hat {n}} | | \\mathcal {F} (\\mathbf {G} _ {i}, \\mathbf {G} _ {j}, d _ {i, \\hat {n}}) - \\Pi_ {i j} (\\mathbf {p} _ {i, \\hat {n}}) | |,", + "image_path": "ae4f8c18f3b654e4dfa68d7088487225e56a04b97c8bc32121371208233e6a5b.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 415, + 295, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 295, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 295, + 475 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 415, + 295, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\mathbf{G}_i,\\mathbf{G}_j,d_{i,\\hat{n}})" + }, + { + "bbox": [ + 55, + 415, + 295, + 475 + ], + "type": "text", + "content": " denotes the point positions calculated by camera pose " + }, + { + "bbox": [ + 55, + 415, + 295, + 475 + ], + "type": "inline_equation", + "content": "\\mathbf{G}" + }, + { + "bbox": [ + 55, + 415, + 295, + 475 + ], + "type": "text", + "content": " at frame " + }, + { + "bbox": [ + 55, + 415, + 295, + 475 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 415, + 295, + 475 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 415, + 295, + 475 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 415, + 295, + 475 + ], + "type": "text", + "content": " with depth " + }, + { + "bbox": [ + 55, + 415, + 295, + 475 + ], + "type": "inline_equation", + "content": "d_{i,\\hat{n}}" + }, + { + "bbox": [ + 55, + 415, + 295, + 475 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 415, + 295, + 475 + ], + "type": "inline_equation", + "content": "\\Pi_{ij}(\\mathbf{p}_{i,\\hat{n}})" + }, + { + "bbox": [ + 55, + 415, + 295, + 475 + ], + "type": "text", + "content": " denotes the position for project position of " + }, + { + "bbox": [ + 55, + 415, + 295, + 475 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{i,\\hat{n}}" + }, + { + "bbox": [ + 55, + 415, + 295, + 475 + ], + "type": "text", + "content": " from frame " + }, + { + "bbox": [ + 55, + 415, + 295, + 475 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 415, + 295, + 475 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 415, + 295, + 475 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 415, + 295, + 475 + ], + "type": "text", + "content": ". Consequently, we obtain the camera rotation " + }, + { + "bbox": [ + 55, + 415, + 295, + 475 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_t" + }, + { + "bbox": [ + 55, + 415, + 295, + 475 + ], + "type": "text", + "content": " and translation " + }, + { + "bbox": [ + 55, + 415, + 295, + 475 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_t" + }, + { + "bbox": [ + 55, + 415, + 295, + 475 + ], + "type": "text", + "content": " from camera pose " + }, + { + "bbox": [ + 55, + 415, + 295, + 475 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_t" + }, + { + "bbox": [ + 55, + 415, + 295, + 475 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 55, + 415, + 295, + 475 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 415, + 295, + 475 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 475, + 295, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 295, + 523 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 295, + 523 + ], + "type": "text", + "content": "Recovering human motion in world coordinates with estimated camera parameters. Given an input video, we feed the estimated camera parameters " + }, + { + "bbox": [ + 55, + 475, + 295, + 523 + ], + "type": "inline_equation", + "content": "(\\mathbf{R}_t" + }, + { + "bbox": [ + 55, + 475, + 295, + 523 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 475, + 295, + 523 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_t)" + }, + { + "bbox": [ + 55, + 475, + 295, + 523 + ], + "type": "text", + "content": " into the state-of-the-art motion recovering model, GVHMR [29]," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 97, + 528, + 294, + 541 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 528, + 294, + 541 + ], + "spans": [ + { + "bbox": [ + 97, + 528, + 294, + 541 + ], + "type": "interline_equation", + "content": "\\theta_ {t} ^ {w}, \\beta_ {t} ^ {w}, \\Gamma_ {t} ^ {w}, \\tau_ {t} ^ {w} = \\mathrm {G V H M R} \\left(I _ {t}, \\mathbf {R} _ {t}, \\mathbf {T} _ {t}\\right). \\tag {1}", + "image_path": "8c71337f3325242e3de73b711a1962c09de74a03f34ceac4bde2c6e0042b25e3.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 543, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 543, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 543, + 295, + 567 + ], + "type": "text", + "content": " Initialized human parameters " + }, + { + "bbox": [ + 55, + 543, + 295, + 567 + ], + "type": "inline_equation", + "content": "\\theta_t^w,\\beta_t^w,\\Gamma_t^w,\\tau_t^w" + }, + { + "bbox": [ + 55, + 543, + 295, + 567 + ], + "type": "text", + "content": " and camera parameters " + }, + { + "bbox": [ + 55, + 543, + 295, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_t,\\mathbf{T}_t" + }, + { + "bbox": [ + 55, + 543, + 295, + 567 + ], + "type": "text", + "content": " will input to human motion alignment." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 575, + 264, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 575, + 264, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 575, + 264, + 588 + ], + "type": "text", + "content": "3.4. Aligning Human Motion Between Shots" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": "Based on initialized world motion for each individual shot, the subsequent question is how to merge discontinuous motions from different shots into a continuous motion sequence as a whole in world coordinates. A straightforward solution is to align all motion sequences to the world coordinate system of the first shot. However, finding the correspondence between different shots is still under-explored and challenging. To resolve this issue, we decompose the motion parameters into camera-dependent and camera-independent ones. The former (Sec. 3.4.1) achieves alignment between shots" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 320, + 72, + 548, + 180 + ], + "blocks": [ + { + "bbox": [ + 320, + 72, + 548, + 180 + ], + "lines": [ + { + "bbox": [ + 320, + 72, + 548, + 180 + ], + "spans": [ + { + "bbox": [ + 320, + 72, + 548, + 180 + ], + "type": "image", + "image_path": "c18d8cf03b5852f2d79f6a998a219a15655cf336e96ac1be4db29ebd159ef0b2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 180, + 555, + 260 + ], + "lines": [ + { + "bbox": [ + 313, + 180, + 555, + 260 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 555, + 260 + ], + "type": "text", + "content": "Figure 5. Human orientation alignment module. Following a shot transition after the foremost purple human mesh (shot " + }, + { + "bbox": [ + 313, + 180, + 555, + 260 + ], + "type": "inline_equation", + "content": "①" + }, + { + "bbox": [ + 313, + 180, + 555, + 260 + ], + "type": "text", + "content": " captured by camera " + }, + { + "bbox": [ + 313, + 180, + 555, + 260 + ], + "type": "inline_equation", + "content": "C_0" + }, + { + "bbox": [ + 313, + 180, + 555, + 260 + ], + "type": "text", + "content": "), the unaligned (blue) and aligned (green) motions are captured as shot " + }, + { + "bbox": [ + 313, + 180, + 555, + 260 + ], + "type": "inline_equation", + "content": "②" + }, + { + "bbox": [ + 313, + 180, + 555, + 260 + ], + "type": "text", + "content": " and shot “" + }, + { + "bbox": [ + 313, + 180, + 555, + 260 + ], + "type": "inline_equation", + "content": "③" + }, + { + "bbox": [ + 313, + 180, + 555, + 260 + ], + "type": "text", + "content": "” by camera " + }, + { + "bbox": [ + 313, + 180, + 555, + 260 + ], + "type": "inline_equation", + "content": "C_0'" + }, + { + "bbox": [ + 313, + 180, + 555, + 260 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 180, + 555, + 260 + ], + "type": "inline_equation", + "content": "C_1" + }, + { + "bbox": [ + 313, + 180, + 555, + 260 + ], + "type": "text", + "content": ", respectively. " + }, + { + "bbox": [ + 313, + 180, + 555, + 260 + ], + "type": "inline_equation", + "content": "C_0' = C_0" + }, + { + "bbox": [ + 313, + 180, + 555, + 260 + ], + "type": "text", + "content": ". To achieve human orientation alignment from shot “" + }, + { + "bbox": [ + 313, + 180, + 555, + 260 + ], + "type": "inline_equation", + "content": "①" + }, + { + "bbox": [ + 313, + 180, + 555, + 260 + ], + "type": "text", + "content": "” to “" + }, + { + "bbox": [ + 313, + 180, + 555, + 260 + ], + "type": "inline_equation", + "content": "③" + }, + { + "bbox": [ + 313, + 180, + 555, + 260 + ], + "type": "text", + "content": "”, the camera rotation matrix from " + }, + { + "bbox": [ + 313, + 180, + 555, + 260 + ], + "type": "inline_equation", + "content": "C_0'" + }, + { + "bbox": [ + 313, + 180, + 555, + 260 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 180, + 555, + 260 + ], + "type": "inline_equation", + "content": "C_1" + }, + { + "bbox": [ + 313, + 180, + 555, + 260 + ], + "type": "text", + "content": " is computed and applied as the offset of human orientation." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 264, + 554, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 264, + 554, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 264, + 554, + 324 + ], + "type": "text", + "content": "via human orientation alignment based on camera calibration, whereas the latter (Sec. 3.4.2) is a trainable module to enhance the continuity of human motion sequence. These two key designs ensure a consistent motion sequence between frames when encountering shot transitions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 332, + 539, + 344 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 332, + 539, + 344 + ], + "spans": [ + { + "bbox": [ + 313, + 332, + 539, + 344 + ], + "type": "text", + "content": "3.4.1 Aligning Human Orientations Between Shots" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 350, + 555, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 555, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 555, + 495 + ], + "type": "text", + "content": "After obtaining the initial SMPL and camera parameters " + }, + { + "bbox": [ + 313, + 350, + 555, + 495 + ], + "type": "inline_equation", + "content": "\\{\\theta_t^i,\\beta_t^i,\\Gamma_t^i,\\tau_t^i,\\mathbf{R}_t^i,\\mathbf{T}_t^i\\}" + }, + { + "bbox": [ + 313, + 350, + 555, + 495 + ], + "type": "text", + "content": " for each shot, directly concatenating motions between shots result abrupt changes of human poses and orientations. To address this issue, we introduce the Orientation Alignment Module (OAM), as shown in Fig. 5, to align human orientations. As the whole motion sequence is continuous, we have the following assumption. Assumption 1 Human orientations and translations during the shot transition in world coordinates are continuous. To align the orientations between two frames with shot transition under Assumption 1, we decompose the human orientation with shot transitions in world coordinates as," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 379, + 497, + 553, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 379, + 497, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 379, + 497, + 553, + 510 + ], + "type": "interline_equation", + "content": "\\mathrm {R} \\left(\\Gamma_ {\\text {w o r l d}}\\right) = \\mathrm {R} _ {\\delta_ {\\text {c a m}}} \\mathrm {R} \\left(\\Gamma_ {\\text {v i e w}}\\right), \\tag {2}", + "image_path": "a8229b4c5708cf849c3a31872c3ed216ee0926ca1f22b788fc37e11a7ec0d717.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 510, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 605 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 510, + 555, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{\\delta_{\\mathrm{cam}}}" + }, + { + "bbox": [ + 313, + 510, + 555, + 605 + ], + "type": "text", + "content": " represents the camera rotation on the Y-axis between current " + }, + { + "bbox": [ + 313, + 510, + 555, + 605 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 510, + 555, + 605 + ], + "type": "text", + "content": "-th and previous " + }, + { + "bbox": [ + 313, + 510, + 555, + 605 + ], + "type": "inline_equation", + "content": "t - 1" + }, + { + "bbox": [ + 313, + 510, + 555, + 605 + ], + "type": "text", + "content": "-th frame, " + }, + { + "bbox": [ + 313, + 510, + 555, + 605 + ], + "type": "inline_equation", + "content": "\\Gamma_{\\mathrm{view}}" + }, + { + "bbox": [ + 313, + 510, + 555, + 605 + ], + "type": "text", + "content": " denotes the human orientation estimated by the current shot, and " + }, + { + "bbox": [ + 313, + 510, + 555, + 605 + ], + "type": "inline_equation", + "content": "\\mathbb{R}(\\cdot):\\mathbb{R}^3\\to \\mathbb{R}^9" + }, + { + "bbox": [ + 313, + 510, + 555, + 605 + ], + "type": "text", + "content": " is the mapping from axis angle to rotation matrix. As " + }, + { + "bbox": [ + 313, + 510, + 555, + 605 + ], + "type": "inline_equation", + "content": "\\Gamma_{\\mathrm{view}}" + }, + { + "bbox": [ + 313, + 510, + 555, + 605 + ], + "type": "text", + "content": " in current shot can be estimated independently, mentioned in Sec. 3.3, obtaining accurate " + }, + { + "bbox": [ + 313, + 510, + 555, + 605 + ], + "type": "inline_equation", + "content": "\\Gamma_{\\mathrm{world}}" + }, + { + "bbox": [ + 313, + 510, + 555, + 605 + ], + "type": "text", + "content": " in Eq. (2) remains a key challenge to estimate the relative camera rotation " + }, + { + "bbox": [ + 313, + 510, + 555, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{\\delta_{\\mathrm{cam}}}" + }, + { + "bbox": [ + 313, + 510, + 555, + 605 + ], + "type": "text", + "content": " between frames in shot transitions." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 606, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 556, + 714 + ], + "type": "text", + "content": "Estimating the relative camera pose " + }, + { + "bbox": [ + 313, + 606, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{\\delta_{\\mathrm{cam}}}" + }, + { + "bbox": [ + 313, + 606, + 556, + 714 + ], + "type": "text", + "content": " between transition frames. Different from our approach of estimating camera pose in each shot (Sec. 3.3), we do not mask the human subject when estimating camera rotation " + }, + { + "bbox": [ + 313, + 606, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{\\delta_{\\mathrm{cam}}}" + }, + { + "bbox": [ + 313, + 606, + 556, + 714 + ], + "type": "text", + "content": ". Instead, we use human 2D KPTs as explicit feature matching. Specifically, we filter out unmatched keypoints based on their visibility and unaligned direction using RANSAC [87], effectively addressing camera pose estimation during shot transitions. This procedure is referred" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1977" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": "to as Camera Calibration (a.k.a. epipolar-geometry-based camera extrinsic estimation), and is detailed below." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 96, + 296, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 96, + 296, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 96, + 296, + 275 + ], + "type": "text", + "content": "In Camera Calibration, we assume that the human translations remain unchanged across the shot transition, implying that only the camera's orientation changes (i.e. Assumption 1). Consequently, we calculate the orientation offset by determining the change in camera orientation using camera calibration. We begin by extracting human 2D KPTs from two consecutive frames during the shot transition. Due to the shot transition, the visibility of 2D KPTs may vary, e.g. occlusion in some shots. Therefore, we employ ED-Pose [88] to filter out invisible 2D KPTs between shot transition frames. Subsequently, RANSAC identifies matching 2D KPTs corresponding to the most possible camera rotation direction. These matched 2D KPTs facilitate the estimation of the aligned camera rotation " + }, + { + "bbox": [ + 55, + 96, + 296, + 275 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{\\delta_{\\mathrm{cam}}}" + }, + { + "bbox": [ + 55, + 96, + 296, + 275 + ], + "type": "text", + "content": ". The detailed estimation process is as follows." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 276, + 296, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 276, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 276, + 296, + 350 + ], + "type": "text", + "content": "We denote the detected 2D KPTs of two frames in the shot transition as " + }, + { + "bbox": [ + 55, + 276, + 296, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_1 = [(x_1^{(1)},y_1^{(1)}),(x_1^{(2)},y_1^{(2)}),\\dots ,(x_1^{(N)},y_1^{(N)})]^{\\top}\\in \\mathbb{R}^{2\\times N}" + }, + { + "bbox": [ + 55, + 276, + 296, + 350 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 276, + 296, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_2 = [(x_2^{(1)},y_2^{(1)}),(x_2^{(2)},y_2^{(2)}),\\dots ,(x_2^{(N)},y_2^{(N)})]^{\\top}\\in" + }, + { + "bbox": [ + 55, + 276, + 296, + 350 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{2\\times N}" + }, + { + "bbox": [ + 55, + 276, + 296, + 350 + ], + "type": "text", + "content": ". The essential matrix " + }, + { + "bbox": [ + 55, + 276, + 296, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{E} = [\\mathbf{T}]_{\\times}\\mathbf{R}" + }, + { + "bbox": [ + 55, + 276, + 296, + 350 + ], + "type": "text", + "content": " should satisfy the following orthogonal property such that," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 146, + 350, + 294, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 350, + 294, + 365 + ], + "spans": [ + { + "bbox": [ + 146, + 350, + 294, + 365 + ], + "type": "interline_equation", + "content": "\\mathbf {S} _ {1} ^ {\\top} \\mathbf {E} \\mathbf {S} _ {2} = \\mathbf {0}. \\tag {3}", + "image_path": "eda800b80e865408339fa10191be9c1f4f54af0d98e6da98bfcf3455e7cde5ed.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 373, + 295, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 295, + 421 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 295, + 421 + ], + "type": "text", + "content": "Once " + }, + { + "bbox": [ + 55, + 373, + 295, + 421 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 55, + 373, + 295, + 421 + ], + "type": "text", + "content": " is obtained by solving Eq. (3), we enforce the rank-2 constraint on " + }, + { + "bbox": [ + 55, + 373, + 295, + 421 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 55, + 373, + 295, + 421 + ], + "type": "text", + "content": " through SVD decomposition and subsequently derive the aligned camera rotation " + }, + { + "bbox": [ + 55, + 373, + 295, + 421 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{\\delta_{\\mathrm{cam}}}" + }, + { + "bbox": [ + 55, + 373, + 295, + 421 + ], + "type": "text", + "content": " between two frames (cf. Hartley et al. [89] for more details)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 422, + 296, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 296, + 471 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 296, + 471 + ], + "type": "text", + "content": "In summary, we reformulate the alignment problem of human orientation in shot transitions as estimating the relative camera rotation " + }, + { + "bbox": [ + 55, + 422, + 296, + 471 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{\\delta_{\\mathrm{cam}}}" + }, + { + "bbox": [ + 55, + 422, + 296, + 471 + ], + "type": "text", + "content": " between frames. Accordingly, we obtain the camera rotation " + }, + { + "bbox": [ + 55, + 422, + 296, + 471 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{\\delta_{\\mathrm{cam}}}" + }, + { + "bbox": [ + 55, + 422, + 296, + 471 + ], + "type": "text", + "content": " via camera calibration." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 479, + 250, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 250, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 250, + 491 + ], + "type": "text", + "content": "3.4.2 Aligning Human Poses Between Shots" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 498, + 296, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 606 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 606 + ], + "type": "text", + "content": "In shot transition, video sequences recorded by two shots are often with various occlusions. However, unoccluded body parts in two shots can be complementary to each other for motion alignment. Thus, we introduce the multi-shot HMR (ms-HMR, i.e. " + }, + { + "bbox": [ + 55, + 498, + 296, + 606 + ], + "type": "inline_equation", + "content": "\\mathsf{E}_M(\\cdot)" + }, + { + "bbox": [ + 55, + 498, + 296, + 606 + ], + "type": "text", + "content": ") module to refine the whole motion sequence. As shown in Fig. 6, the ms-HMR is a Transformer encoder-like architecture, whose input and output are the estimated global motion and the refined global motion, respectively. The process can be formulated as," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 95, + 610, + 295, + 623 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 610, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 95, + 610, + 295, + 623 + ], + "type": "interline_equation", + "content": "\\phi_ {1}, \\phi_ {2}, \\dots , \\phi_ {T} = \\mathrm {E} _ {M} \\left(\\theta_ {1}, \\theta_ {2}, \\dots , \\theta_ {T}\\right), \\tag {4}", + "image_path": "ce6f9168d966de34c83c9c828b1d85239a203c91492b2c86117c9691320b2cf8.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 627, + 296, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 627, + 296, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 627, + 296, + 664 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 627, + 296, + 664 + ], + "type": "inline_equation", + "content": "\\phi_*" + }, + { + "bbox": [ + 55, + 627, + 296, + 664 + ], + "type": "text", + "content": " denotes the refined motion of each frame. With this design, our method can adapt to diverse occlusions of human body brought by shot transitions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 671, + 295, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 295, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 295, + 685 + ], + "type": "text", + "content": "3.5. Post-processing Module for Motion Integration" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "Trajectory Predictor and Trajectory Refiner. Based on the aligned human pose and orientation, we introduce a" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 318, + 72, + 559, + 262 + ], + "blocks": [ + { + "bbox": [ + 318, + 72, + 559, + 262 + ], + "lines": [ + { + "bbox": [ + 318, + 72, + 559, + 262 + ], + "spans": [ + { + "bbox": [ + 318, + 72, + 559, + 262 + ], + "type": "image", + "image_path": "1cb1151ba9da0460afacd94ca6f5dc2be9bccac51fe297050c8a9acf8ea47910.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 263, + 555, + 331 + ], + "lines": [ + { + "bbox": [ + 313, + 263, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 263, + 555, + 331 + ], + "type": "text", + "content": "Figure 6. ms-HMR Structure. The initial human pose parameters " + }, + { + "bbox": [ + 313, + 263, + 555, + 331 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 263, + 555, + 331 + ], + "type": "text", + "content": " across multiple video shots are input into a transformer with shot-index-based positional encoding. This enables ms-HMR to generate consistent human poses across all shots in the video. bi-directional LSTM trajectory predictor to recover foot-ground contact probabilities " + }, + { + "bbox": [ + 313, + 263, + 555, + 331 + ], + "type": "inline_equation", + "content": "p_t^c" + }, + { + "bbox": [ + 313, + 263, + 555, + 331 + ], + "type": "text", + "content": ", and root velocity " + }, + { + "bbox": [ + 313, + 263, + 555, + 331 + ], + "type": "inline_equation", + "content": "v_t" + }, + { + "bbox": [ + 313, + 263, + 555, + 331 + ], + "type": "text", + "content": " as," + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 328, + 331, + 553, + 351 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 331, + 553, + 351 + ], + "spans": [ + { + "bbox": [ + 328, + 331, + 553, + 351 + ], + "type": "interline_equation", + "content": "p _ {t} ^ {c}, v _ {t} = \\operatorname {L S T M} \\left(\\phi_ {1} ^ {m}, \\Gamma_ {1}, \\mathbb {F} (I _ {1}), \\phi_ {2} ^ {m}, \\Gamma_ {2}, \\mathbb {F} (I _ {2}), \\dots , \\right. \\\\ \\left. \\quad \\left(m, \\mathbb {F} - \\mathbb {F} (I _ {m})\\right)\\right) \\tag {5}", + "image_path": "e7bf3777a6d9d1008446f68ed225652dba1543e66406702cf487fb615b5474d5.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 460, + 346, + 527, + 359 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 460, + 346, + 527, + 359 + ], + "spans": [ + { + "bbox": [ + 460, + 346, + 527, + 359 + ], + "type": "interline_equation", + "content": "\\phi_ {T} ^ {m}, \\Gamma_ {T}, F (I _ {T})),", + "image_path": "719c8c7b6eebecd18e66fa5cc902cfc881a40b7517a8bb0d73f1975ca8d58d07.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 359, + 554, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 359, + 554, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 359, + 554, + 431 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 359, + 554, + 431 + ], + "type": "inline_equation", + "content": "\\mathsf{F}(\\cdot)" + }, + { + "bbox": [ + 313, + 359, + 554, + 431 + ], + "type": "text", + "content": " denotes the image feature of each frame extracted by ViT [81]. Accordingly, the contact probabilities " + }, + { + "bbox": [ + 313, + 359, + 554, + 431 + ], + "type": "inline_equation", + "content": "p_t^c" + }, + { + "bbox": [ + 313, + 359, + 554, + 431 + ], + "type": "text", + "content": ", and velocity " + }, + { + "bbox": [ + 313, + 359, + 554, + 431 + ], + "type": "inline_equation", + "content": "v_t" + }, + { + "bbox": [ + 313, + 359, + 554, + 431 + ], + "type": "text", + "content": " are supervised by the ground-truth labels with MSE loss and are used to reconstruct the trajectory. Besides, we extend the trajectory refiner in WHAM [28] to alleviate foot sliding problem in our estimated trajectory." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 435, + 553, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 435, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 435, + 553, + 449 + ], + "type": "text", + "content": "4. Benchmarking Multi-shot Motion Recovery" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 450, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 555, + 689 + ], + "type": "text", + "content": "Dataset Construction. To create a multi-shot 3D human motion dataset, we introduce ms-Motion by processing existing public 3D human datasets with multiple camera settings and ground truth human and camera parameters, specifically AIST [60] and Human3.6M (H3.6M) [61]. In our construction pipeline, we randomly separate each original one-shot video into several clips. Then, we choose each clip from different shots and concatenate them together as one video recorded by multiple shots. For example, AIST provides each video with eight cameras C0, C1, ..., C7 from different view point and we choose a video and split it into 5 clips at t0, t1, ..., t4. For frames in these separated clips, we choose frames shot by a random camera for each clip and combine five clips as one multi-shot video. Therefore, we construct a multi-shot version of AIST and H3.6M, which are named ms-AIST and ms-H3.6M subsets. Then we combine them and name this new dataset ms-Motion. The detailed statistics of ms-Motion are shown in Tab. 2. We do not compare with other existing 3D human datasets as they contain limited number of multi-shot videos." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "Benchmark Evaluation Protocol. To evaluate the performance of our proposed methods on multi-shot videos, our" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1978" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 553, + 170 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 170 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 170 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 170 + ], + "type": "table", + "html": "
DatasetModels2-Shot3-Shot4-Shot
PA.↓WA.↓RTE↓ROE↓Jitter↓F.S.↓PA.↓WA.↓RTE↓ROE↓Jitter↓F.S.↓PA↓WA↓RTE↓ROE↓Jitter↓F.S.↓
ms-AISTSLAHMR [2023]72.34341.759.6296.2662.593.2680.35510.7710.33101.3672.394.4390.32803.6912.11104.0780.3716.52
WHAM [2024]65.34336.824.3984.4825.242.7578.68451.325.1489.8424.062.99102.88603.935.5790.0726.293.62
GVHMR [2024]60.72231.366.2096.5834.877.6570.33357.167.5599.6934.469.4283.77563.178.96104.5335.679.78
Ours36.82121.352.5669.2333.272.6638.52141.383.6467.7135.073.5539.63161.524.5570.3139.494.09
ms-H3.6MSLAHMR [2023]80.67352.6116.67111.9737.807.9397.15562.1016.91118.4652.239.96107.90748.5817.85116.7265.1511.58
WHAM [2024]71.32313.5811.4182.4218.405.0979.51423.9812.3684.8518.875.0390.50512.6612.9190.3418.405.69
GVHMR [2024]64.63254.306.9481.9318.458.8080.79296.7485.2558.2618.3610.6285.19471.539.1291.6319.4710.65
Ours40.52132.133.6553.3919.054.1745.35145.365.3358.2617.354.6250.59147.626.2061.2219.775.12
", + "image_path": "69b37aa15e1d1f5cb6eb89130ebbcbec7d4b5f6906ffe6992fc9d17bc22a67e4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 205, + 294, + 235 + ], + "blocks": [ + { + "bbox": [ + 55, + 171, + 553, + 203 + ], + "lines": [ + { + "bbox": [ + 55, + 171, + 553, + 203 + ], + "spans": [ + { + "bbox": [ + 55, + 171, + 553, + 203 + ], + "type": "text", + "content": "Table 1. Quantitative comparison of different HMR methods on ms-Motion dataset. We record the results for ms-AIST and ms-H3.6M separately. PA and WA means PA-MPJPE and WA-MPJPE respectively, while F.S. is Foot Sliding. Our proposed method has achieved the best performance in PA-MPJPE, WA-MPJPE, RTE and ROE across ms-Motion among these methods." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 205, + 294, + 235 + ], + "lines": [ + { + "bbox": [ + 56, + 205, + 294, + 235 + ], + "spans": [ + { + "bbox": [ + 56, + 205, + 294, + 235 + ], + "type": "table", + "html": "
DatasetDuration(s)VideosFPSMax LengthMin LengthShots
ms-Motion23.76003014783142, 3, 4
", + "image_path": "871d27dcb28520e36167e8c52ba5dff7bd85d5247cd9f6b5a7f968b8b43caf38.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 263, + 296, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 263, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 54, + 263, + 296, + 407 + ], + "type": "text", + "content": "target is to evaluate metrics for accurately reflecting the performance on videos with shot transitions. To this end, we use Root Orientation Error (a.k.a. ROE in deg) to measure the performance of the proposed method on human orientation alignment across different shots. Besides, we use Root Translation Error (a.k.a. RTE in " + }, + { + "bbox": [ + 54, + 263, + 296, + 407 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 54, + 263, + 296, + 407 + ], + "type": "text", + "content": ") to assess the performance of the proposed method on global trajectory recovery. Jitter " + }, + { + "bbox": [ + 54, + 263, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\left( \\frac{10m}{fps^3} \\right)" + }, + { + "bbox": [ + 54, + 263, + 296, + 407 + ], + "type": "text", + "content": " is also used to evaluate the stability of recovered human pose from multi-shot videos. We also include foot sliding " + }, + { + "bbox": [ + 54, + 263, + 296, + 407 + ], + "type": "inline_equation", + "content": "(cm)" + }, + { + "bbox": [ + 54, + 263, + 296, + 407 + ], + "type": "text", + "content": ", the averaged displacement of foot vertices during contact with the ground, to assess the precision of recovered motion in the world coordinates [28]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 421, + 132, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 421, + 132, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 421, + 132, + 434 + ], + "type": "text", + "content": "5. Experiment" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 441, + 177, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 177, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 177, + 453 + ], + "type": "text", + "content": "5.1. Datasets and Metrics" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 460, + 295, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 460, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 54, + 460, + 295, + 579 + ], + "type": "text", + "content": "Evaluation Datasets. To evaluate the performance of our proposed pipeline for multi-shot videos, we use ms-Motion dataset and EMDB-1 dataset [73] with self-added noise for the evaluation of ablation study. For camera trajectory estimation, we use EMDB-1 and EMDB-2 split [73] as they contain the GT moving camera trajectory. Our self-created dataset contains 600 multi-shot videos, 42.7K frames, totaling 237 minutes. EMDB-1 split contains 17 video sequences totaling 13.5 minutes and EMDB-2 split contains 25 sequences in a total of 24.0 minutes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "type": "text", + "content": "Evaluation Metrics. For shot detection we use Recall, Precision and F1 Score as evaluation metrics. For 3D human pose estimation-related tasks, we use ROE, RTE, jitter, and foot-sliding for evaluating the human motion recovery results on multi-shot videos. For the ablation study of our proposed pipeline, we evaluate the Procrustes-aligned Mean Per Joint Position Error (a.k.a. PA-MPJPE) and Per Vertex Error (a.k.a. PVE) as additional metrics besides previous mentioned ones. For camera pose estimation, we use absolute trajectory error (a.k.a. ATE) " + }, + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "type": "inline_equation", + "content": "(m)" + }, + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "type": "text", + "content": ", Relative Pose Error (a.k.a. RPE) rotation " + }, + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "type": "inline_equation", + "content": "(deg)" + }, + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "type": "text", + "content": ", and RPE translation " + }, + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "type": "inline_equation", + "content": "(m)" + }, + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 315, + 205, + 430, + 243 + ], + "blocks": [ + { + "bbox": [ + 55, + 235, + 295, + 257 + ], + "lines": [ + { + "bbox": [ + 55, + 235, + 295, + 257 + ], + "spans": [ + { + "bbox": [ + 55, + 235, + 295, + 257 + ], + "type": "text", + "content": "Table 2. Statistics of the ms-Motion dataset. By shots, we mean the number of shot transitions in a single video." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 205, + 430, + 243 + ], + "lines": [ + { + "bbox": [ + 315, + 205, + 430, + 243 + ], + "spans": [ + { + "bbox": [ + 315, + 205, + 430, + 243 + ], + "type": "table", + "html": "
Methodsms-Motion
Recall↑Precision↑F1 Score↑
Scenes Detect (SD) [83]0.740.720.70
SD+Bbox Tracking (Bbox)0.880.850.86
SD+Bbox+Pose Tracking0.960.880.92
", + "image_path": "53927265636e81df53b21d52febe7d146dc8b88399c9d22d92586f89535d5d6e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 436, + 207, + 549, + 241 + ], + "blocks": [ + { + "bbox": [ + 314, + 243, + 431, + 286 + ], + "lines": [ + { + "bbox": [ + 314, + 243, + 431, + 286 + ], + "spans": [ + { + "bbox": [ + 314, + 243, + 431, + 286 + ], + "type": "text", + "content": "Table 3. Comparison between difference shot detection algorithms. We evaluate our shot transition detector on our proposed multi-shot video human motion dataset ms-Motion." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 436, + 207, + 549, + 241 + ], + "lines": [ + { + "bbox": [ + 436, + 207, + 549, + 241 + ], + "spans": [ + { + "bbox": [ + 436, + 207, + 549, + 241 + ], + "type": "table", + "html": "
MethodsATE↓RPE Trans.↓RPE Rot.↓
DPVO (w/o mask)0.481.851.06
Masked DPVO0.481.570.97
LEAP-VO (w/o mask)0.500.930.97
Masked LEAP-VO0.510.920.95
", + "image_path": "2259ce20b8487ba6de83159f464280004aafe73f2abdafbd5694f00b23c20c6f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 434, + 242, + 550, + 284 + ], + "lines": [ + { + "bbox": [ + 434, + 242, + 550, + 284 + ], + "spans": [ + { + "bbox": [ + 434, + 242, + 550, + 284 + ], + "type": "text", + "content": "Table 4. Camera tracking results on EMDB-1 [73]. Our method has achieved " + }, + { + "bbox": [ + 434, + 242, + 550, + 284 + ], + "type": "inline_equation", + "content": "\\sim 50\\%" + }, + { + "bbox": [ + 434, + 242, + 550, + 284 + ], + "type": "text", + "content": " ↓ on RPE trans. than that of the original DPVO and perform the best in RPE rot. metrics." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 314, + 289, + 447, + 301 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 289, + 447, + 301 + ], + "spans": [ + { + "bbox": [ + 314, + 289, + 447, + 301 + ], + "type": "text", + "content": "5.2. Implementation Details" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 305, + 555, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 305, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 312, + 305, + 555, + 449 + ], + "type": "text", + "content": "The ms-HMR, the trajectory, and foot sliding refiner are trained on the AMASS [78], 3DPW [79], Human3.6M [61], and BEDLAM [90] datasets, and evaluated on EMDB and our ms-Motion. During training, we introduce random rotational noise (ranging from 0 to 1 radian) along the y-axis to the root pose " + }, + { + "bbox": [ + 312, + 305, + 555, + 449 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 312, + 305, + 555, + 449 + ], + "type": "text", + "content": " and random noise to the body pose " + }, + { + "bbox": [ + 312, + 305, + 555, + 449 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 312, + 305, + 555, + 449 + ], + "type": "text", + "content": " at random positions to simulate the inaccuracies of pre-estimated human motions caused by shot transitions in multi-shot videos. This strategy enables the network to robustly recover smooth and consistent human motion from noisy initial parameters. The benchmark test results were obtained after training for 80 epochs on one NVIDIA-A100 GPU." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 456, + 553, + 481 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 553, + 481 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 553, + 481 + ], + "type": "text", + "content": "5.3. Main Results: Comparison of Global Human Motion Recovery Results on the Benchmark" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 485, + 555, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 485, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 312, + 485, + 555, + 617 + ], + "type": "text", + "content": "We compare our proposed method HumanMM with several state-of-the-art HMR methods (SLAHMR [26], WHAM [28] and GVHMR [29]) on our proposed benchmark ms-Motion. As illustrated in Tab. 1, our proposed method has achieved the best performance for PA&WAMPJPE, RTE and ROE through videos with all numbers of shots across ms-AIST and ms-H3.6M, indicating that our method reconstructs both the global human motion and orientations in the world coordinates more accurately and robustly. For the foot sliding metric, our method also performs as the best on ms-H3.6M across all numbers of shots." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 625, + 416, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 625, + 416, + 637 + ], + "spans": [ + { + "bbox": [ + 313, + 625, + 416, + 637 + ], + "type": "text", + "content": "5.4. Ablation Studies" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 642, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 555, + 715 + ], + "type": "text", + "content": "Human-centric Scene Shot Boundary Detection Evaluation. To evaluate the performance of our proposed Shot Transition Detector, we test the algorithm on our proposed multi-shot human motion recovery benchmark and compare the output frame list of shot transitions with the ground truth (GT) of our dataset. As shown in Tab. 3, by apply" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1979" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 550, + 174 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 550, + 174 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 550, + 174 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 550, + 174 + ], + "type": "image", + "image_path": "a8e7df8b005299e7c51581efca7768e5e9d04f98d298dc31fcbcc2a35bfae541.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 174, + 555, + 206 + ], + "lines": [ + { + "bbox": [ + 55, + 174, + 555, + 206 + ], + "spans": [ + { + "bbox": [ + 55, + 174, + 555, + 206 + ], + "type": "text", + "content": "Figure 7. Qualitative comparison of different HMR methods on ms-Motion dataset. The side view of the rendered mesh for input mutli-shot video is shown in (a), while the top view is shown in (c). We also draw the comparison of the human trajectory as shown in (b). Our method is the most similar as GT in both rendered motion and trajectories among these methods." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 208, + 298, + 260 + ], + "blocks": [ + { + "bbox": [ + 56, + 208, + 298, + 260 + ], + "lines": [ + { + "bbox": [ + 56, + 208, + 298, + 260 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 298, + 260 + ], + "type": "table", + "html": "
Methods|PA-MPJE↓|MPJPE↓WA-MPJE↓W-MPJPE↓PVE↓ACCEL↓RTE↓ROE↓F.S.↓
Baseline (Concat)106.48141.67273.15553.67122.156.1410.8691.5514.91
w/o ms-HMR78.24101.52246.42436.5785.775.873.8950.633.54
w/o OAM73.5692.13243.65425.1879.645.676.6176.744.45
w/o traj. pred.50.4983.68231.75432.1775.775.755.5247.684.96
w/o traj. ref.50.4983.68198.58397.6575.775.234.0647.687.84
HumanMM (Ours)50.4983.68194.77393.2175.775.163.5447.683.28
", + "image_path": "b1e52d08de99c1313e86f3c01c5a2dafd33d011a4955eadce8b70b6df28b453b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 282, + 295, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 282, + 295, + 389 + ], + "spans": [ + { + "bbox": [ + 54, + 282, + 295, + 389 + ], + "type": "text", + "content": "ing the proposed finer granularity shot detection methods, the number of recall, precision, and F1 score all increases consistently. The combination of three steps (ScenesDetect, bbox tracking, and pose tracking) has achieved 0.96, 0.88, and 0.92 on the recall, precision, and F1 score, respectively, which indicates a comparable performance in shot boundary detection. Besides, as can be seen in the results, the latter two steps of shot detection contribute to the fine-grained final results significantly and jointly." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 390, + 295, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 390, + 295, + 545 + ], + "spans": [ + { + "bbox": [ + 54, + 390, + 295, + 545 + ], + "type": "text", + "content": "Key modules in the Proposed Method. We compare our methods with four variants on EMDB with noise dataset, as shown in Tab. 5, ms-HMR is the key component for the improvement in PA-MPJPE and PVE, which indicates a more accurate modeling of the whole motion sequence. This design serves as a recovery module to estimate some invisible body parts in some shots. Additionally, the orientation alignment module (OAM, in Sec. 3.4) is also a critical block for accurate human orientation estimation, indicated by the metric ROE. This module helps to model the global human motion between shots. For foot sliding, the results in Tab. 5 also show that the trajectory refiner (Sec. 3.5) in our method helps mitigate the foot sliding issue." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": "Comparison on Camera Trajectory Estimation. To evaluate the performance of our proposed camera trajectory estimation method Masked LEAP-VO, we evaluate the camera trajectory accuracy on EMDB-1 and EMDB-2. For more convenient comparison, we introduce two baselines, DPVO [76], which has been widely used in HMR methods such as WHAM [28] and GVHMR [29], and LEAP-VO [77]. To provide more intuition about the insights of masking dynamic humans in the video, we also implement a variant, Masked DPVO, by applying SAM at the patchify stage of DPVO to exclude patches containing human pixels. As shown in Tab. 4 and Tab. 6, compared with baseline methods, our key design of masking dynamic human subjects improves the result in both RPE Translation and" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 304, + 208, + 411, + 241 + ], + "blocks": [ + { + "bbox": [ + 55, + 261, + 299, + 278 + ], + "lines": [ + { + "bbox": [ + 55, + 261, + 299, + 278 + ], + "spans": [ + { + "bbox": [ + 55, + 261, + 299, + 278 + ], + "type": "text", + "content": "Table 5. Ablation studies on different combinations of HumanMM modules. We evaluate the contributions of each key module on EMDB-1." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 304, + 208, + 411, + 241 + ], + "lines": [ + { + "bbox": [ + 304, + 208, + 411, + 241 + ], + "spans": [ + { + "bbox": [ + 304, + 208, + 411, + 241 + ], + "type": "table", + "html": "
Methods\\( \\left| \\mathrm{{ATE}}\\right| \\downarrow \\)RPE Trans. \\( \\downarrow \\)RPE Rot. \\( \\downarrow \\)
DPVO (w/o mask)0.481.071.26
Masked DPVO0.500.861.21
LEAP-VO (w/o mask)0.500.831.21
Masked LEAP-VO0.490.831.19
", + "image_path": "7ba6046069786c2d9a31500c5cc711655eb9d5f0a6198989c42cfedb9e81f834.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 417, + 208, + 552, + 240 + ], + "blocks": [ + { + "bbox": [ + 303, + 242, + 412, + 278 + ], + "lines": [ + { + "bbox": [ + 303, + 242, + 412, + 278 + ], + "spans": [ + { + "bbox": [ + 303, + 242, + 412, + 278 + ], + "type": "text", + "content": "Table 6. Camera tracking results on EMDB-2 [73]. Masked LEAP-VO performs better in both RPE Trans. and RPE Rot. metrics." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 417, + 208, + 552, + 240 + ], + "lines": [ + { + "bbox": [ + 417, + 208, + 552, + 240 + ], + "spans": [ + { + "bbox": [ + 417, + 208, + 552, + 240 + ], + "type": "table", + "html": "
MethodsWA-MPJE↓W-MPJE↓RTE↓Jitter↓F.S.↓
DPVO305.40117.105.1017.903.50
Masked DPVO303.90116.404.1017.403.50
LEAP-VO284.10112.803.1016.303.50
Masked LEAP-VO283.70112.703.1016.303.50
", + "image_path": "1897d43a75b8be4a6316816eef6742912cea9d95717f9a8b2852d051779766b1.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 416, + 241, + 553, + 278 + ], + "lines": [ + { + "bbox": [ + 416, + 241, + 553, + 278 + ], + "spans": [ + { + "bbox": [ + 416, + 241, + 553, + 278 + ], + "type": "text", + "content": "Table 7. Global motion recovery results on EMDB-2 [73]. We input estimated camera parameters from different methods into GVHMR for comparison on HMR." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 282, + 555, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 282, + 555, + 377 + ], + "spans": [ + { + "bbox": [ + 313, + 282, + 555, + 377 + ], + "type": "text", + "content": "RPE Rotation while maintaining competitive ATE. This result indicates the effectiveness of the design of masking dynamic human subjects in the process of camera trajectory estimation. Compared with the DPVO baseline, our method achieves " + }, + { + "bbox": [ + 313, + 282, + 555, + 377 + ], + "type": "inline_equation", + "content": "\\sim 50\\%" + }, + { + "bbox": [ + 313, + 282, + 555, + 377 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 313, + 282, + 555, + 377 + ], + "type": "text", + "content": " RPE translation on EMDB-1. In addition, we run GVHMR [29] on EMDB-2 with different estimated camera trajectories. The results is shown in Tab. 7, which further illustrates the effectiveness of our method." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 386, + 466, + 398 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 386, + 466, + 398 + ], + "spans": [ + { + "bbox": [ + 313, + 386, + 466, + 398 + ], + "type": "text", + "content": "6. Conclusion and Discussion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 402, + 555, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 555, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 555, + 556 + ], + "type": "text", + "content": "Conclusion. In this paper, we introduce HumanMM, the first framework designed for human motion recovery from multi-shot videos in world coordinates. HumanMM addresses the challenges inherent in multi-shot videos by incorporating three key components: an enhanced camera trajectory estimation method called masked LEAP-VO, a human motion alignment module that ensures consistency across different shots, and a post-processing module for seamless motion integration. Extensive experiments demonstrate that HumanMM outperforms existing human motion recovery methods across various benchmarks, achieving state-of-the-art accuracy on our newly created multi-shot human motion dataset, ms-Motion." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "content": "Limitations and Future Work. While HumanMM represents an dvancement in human motion recovery from multishot videos in world coordinates, its performance may decline when faced with an excessive number of shot transitions. Despite these challenges, HumanMM provides a solid baseline for human motion recovery from multi-shot videos and can be employed in annotating markerless human motion datasets. Our newly introduced dataset, msMotion, offers a valuable benchmark for evaluating general human motion recovery methods in world coordinates, especially regarding their performance on multi-shot videos. Based on the proposed method, our future work aims to enlarge the related datasets for larger-scale motion databases." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1980" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 180 + ], + "type": "text", + "content": "This research was funded through the National Key Research and Development Program of China Project No. 2022YFB36066), in part by the Shenzhen Science and Technology Project under Grant (KJZD20240903103210014). The author team would also like to convey sincere thanks to Ms. Yaxin Chen from IDEA Research for the expressive dance motion used in the demo presentation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 203, + 115, + 216 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 203, + 115, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 203, + 115, + 216 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 224, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 224, + 296, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 224, + 296, + 266 + ], + "spans": [ + { + "bbox": [ + 61, + 224, + 296, + 266 + ], + "type": "text", + "content": "[1] Jingbo Wang, Yu Rong, Jingyuan Liu, Sijie Yan, Dahua Lin, and Bo Dai. Towards diverse and natural scene-aware 3d human motion synthesis. In CVPR, pages 20428-20437, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 269, + 296, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 269, + 296, + 312 + ], + "spans": [ + { + "bbox": [ + 61, + 269, + 296, + 312 + ], + "type": "text", + "content": "[2] Zeqi Xiao, Tai Wang, Jingbo Wang, Jinkun Cao, Wenwei Zhang, Bo Dai, Dahua Lin, and Jiangmiao Pang. Unified human-scene interaction via prompted chain-of-contacts. In ICLR, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 315, + 296, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 315, + 296, + 357 + ], + "spans": [ + { + "bbox": [ + 61, + 315, + 296, + 357 + ], + "type": "text", + "content": "[3] Chuan Guo, Xinxin Zuo, Sen Wang, and Li Cheng. Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In ECCV, pages 580-597, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 360, + 295, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 360, + 295, + 392 + ], + "spans": [ + { + "bbox": [ + 62, + 360, + 295, + 392 + ], + "type": "text", + "content": "[4] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. NeurIPS, 2024." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 395, + 296, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 395, + 296, + 427 + ], + "spans": [ + { + "bbox": [ + 62, + 395, + 296, + 427 + ], + "type": "text", + "content": "[5] Liang Pan, Jingbo Wang, Buzhen Huang, Junyu Zhang, Haofan Wang, Xu Tang, and Yangang Wang. Synthesizing physically plausible human motions in 3d scenes. In 3DV, 2024." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 430, + 296, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 430, + 296, + 473 + ], + "spans": [ + { + "bbox": [ + 62, + 430, + 296, + 473 + ], + "type": "text", + "content": "[6] Jingbo Wang, Ye Yuan, Zhengyi Luo, Kevin Xie, Dahua Lin, Umar Iqbal, Sanja Fidler, and Sameh Khamis. Learning human dynamics in autonomous driving scenarios. In ICCV, pages 20739-20749, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 475, + 296, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 475, + 296, + 508 + ], + "spans": [ + { + "bbox": [ + 62, + 475, + 296, + 508 + ], + "type": "text", + "content": "[7] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. In ECCV, pages 358–374, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 510, + 296, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 510, + 296, + 542 + ], + "spans": [ + { + "bbox": [ + 62, + 510, + 296, + 542 + ], + "type": "text", + "content": "[8] Mathis Petrovich, Michael J Black, and Gül Varol. Temos: Generating diverse human motions from textual descriptions. In ECCV, pages 480-497, 2022." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 544, + 296, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 544, + 296, + 586 + ], + "spans": [ + { + "bbox": [ + 62, + 544, + 296, + 586 + ], + "type": "text", + "content": "[9] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. IEEE TPAMI, 2024." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 590, + 295, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 590, + 295, + 621 + ], + "spans": [ + { + "bbox": [ + 57, + 590, + 295, + 621 + ], + "type": "text", + "content": "[10] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model. In ICLR, 2022." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 624, + 296, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 296, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 296, + 667 + ], + "type": "text", + "content": "[11] Zan Wang, Yixin Chen, Tengyu Liu, Yixin Zhu, Wei Liang, and Siyuan Huang. Humanise: Language-conditioned human motion generation in 3d scenes. NeurIPS, pages 14959-14971, 2022." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 670, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 296, + 712 + ], + "type": "text", + "content": "[12] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, and Gang Yu. Executing your commands via motion diffusion in latent space. In CVPR, pages 18000-18010, 2023." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[13] Ye Yuan, Jiaming Song, Umar Iqbal, Arash Vahdat, and Jan Kautz. Physdiff: Physics-guided human motion diffusion model. In ICCV, pages 16010-16021, 2023." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 107, + 553, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 160 + ], + "type": "text", + "content": "[14] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Yong Zhang, Hongwei Zhao, Hongtao Lu, Xi Shen, and Ying Shan. Generating human motion from textual descriptions with discrete representations. In CVPR, pages 14730-14740, 2023." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 163, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 194 + ], + "type": "text", + "content": "[15] Yonatan Shafir, Guy Tevet, Roy Kapon, and Amit H Bermano. Human motion diffusion as a generative prior. In ICLR, 2024." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 195, + 553, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 553, + 237 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 553, + 237 + ], + "type": "text", + "content": "[16] Korrawe Karunratanakul, Konpat Preechakul, Supasorn Suwajanakorn, and Siyu Tang. Guided motion diffusion for controllable human motion synthesis. In CVPR, pages 2151-2162, 2023." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 239, + 553, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 282 + ], + "type": "text", + "content": "[17] Ling-Hao Chen, Shunlin Lu, Wenxun Dai, Zhiyang Dou, Xuan Ju, Jingbo Wang, Taku Komura, and Lei Zhang. Pay attention and move better: Harnessing attention for interactive motion generation and training-free editing, 2024." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 284, + 553, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 284, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 316, + 284, + 553, + 326 + ], + "type": "text", + "content": "[18] Zeqi Xiao, Tai Wang, Jingbo Wang, Jinkun Cao, Wenwei Zhang, Bo Dai, Dahua Lin, and Jiangmiao Pang. Unified human-scene interaction via prompted chain-of-contacts. In ICLR, 2024." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 327, + 553, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 360 + ], + "type": "text", + "content": "[19] Yiming Xie, Varun Jampani, Lei Zhong, Deqing Sun, and Huaizu Jiang. Omnicontrol: Control any joint at any time for human motion generation. In ICLR, 2024." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 361, + 553, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 553, + 393 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 553, + 393 + ], + "type": "text", + "content": "[20] Shunlin Lu, Ling-Hao Chen, Ailing Zeng, Jing Lin, Ruimao Zhang, Lei Zhang, and Heung-Yeung Shum. Humantomato: Text-aligned whole-body motion generation. ICML, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 394, + 553, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 394, + 553, + 436 + ], + "spans": [ + { + "bbox": [ + 316, + 394, + 553, + 436 + ], + "type": "text", + "content": "[21] Wenxun Dai, Ling-Hao Chen, Jingbo Wang, Jinpeng Liu, Bo Dai, and Yansong Tang. Motionlcm: Real-time controllable motion generation via latent consistency model. ECCV, 2024." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 438, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 438, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 316, + 438, + 553, + 480 + ], + "type": "text", + "content": "[22] Ling-Hao Chen, Jiawei Zhang, Yewen Li, Yiren Pang, Xiaobo Xia, and Tongliang Liu. Humanmac: Masked motion completion for human motion prediction. In ICCV, pages 9544-9555, 2023." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 482, + 553, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 482, + 553, + 535 + ], + "spans": [ + { + "bbox": [ + 316, + 482, + 553, + 535 + ], + "type": "text", + "content": "[23] Shunlin Lu, Jingbo Wang, Zeyu Lu, Ling-Hao Chen, Wenxun Dai, Junting Dong, Zhiyang Dou, Bo Dai, and Ruimao Zhang. Scamo: Exploring the scaling law in autoregressive motion generation model. arXiv preprint arXiv:2412.14559, 2024." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 537, + 553, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 553, + 569 + ], + "type": "text", + "content": "[24] Wenxun Dai, Ling-Hao Chen, Yufei Huo, Jingbo Wang, Jinpeng Liu, Bo Dai, and Yansong Tang. Real-time controllable motion generation via latent consistency model. 2025. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 571, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 571, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 571, + 553, + 613 + ], + "type": "text", + "content": "[25] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa*, and Jitendra Malik*. Humans in 4D: Reconstructing and tracking humans with transformers. In ICCV, 2023. 2, 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 615, + 553, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 615, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 316, + 615, + 553, + 646 + ], + "type": "text", + "content": "[26] Vickie Ye, Georgios Pavlakos, Jitendra Malik, and Angjoo Kanazawa. Decoupling human and camera motion from videos in the wild. In CVPR, 2023. 2, 3, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 647, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 679 + ], + "type": "text", + "content": "[27] Yufu Wang, Ziyun Wang, Lingjie Liu, and Kostas Daniilidis. Tram: Global trajectory and motion of 3d humans from inthe-wild videos. ECCV, 2024. 2, 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "type": "text", + "content": "[28] Soyong Shin, Juyong Kim, Eni Halilaj, and Michael J. Black. WHAM: Reconstructing world-grounded humans with accurate 3D motion. In CVPR, 2024. 6, 7, 8" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 314, + 757 + ], + "type": "text", + "content": "1981" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 126 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 126 + ], + "type": "text", + "content": "[29] Zehong Shen, Huaijin Pi, Yan Xia, Zhi Cen, Sida Peng, Zechen Hu, Hujun Bao, Ruizhen Hu, and Xiaowei Zhou. World-grounded human motion recovery via gravity-view coordinates. In ACM SIGGRAPH Asia, 2024. 2, 3, 4, 5, 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 127, + 294, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 127, + 294, + 160 + ], + "spans": [ + { + "bbox": [ + 56, + 127, + 294, + 160 + ], + "type": "text", + "content": "[30] Angjoo Kanazawa, Jason Y. Zhang, Panna Felsen, and Jitendra Malik. Learning 3d human dynamics from video. In CVPR, 2019. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 161, + 294, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 161, + 294, + 204 + ], + "spans": [ + { + "bbox": [ + 56, + 161, + 294, + 204 + ], + "type": "text", + "content": "[31] Muhammed Kocabas, Ye Yuan, Pavlo Molchanov, Yunrong Guo, Michael J. Black, Otmar Hilliges, Jan Kautz, and Umar Iqbal. Pace: Human and motion estimation from in-the-wild videos. In 3DV, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 205, + 294, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 205, + 294, + 249 + ], + "spans": [ + { + "bbox": [ + 56, + 205, + 294, + 249 + ], + "type": "text", + "content": "[32] Ling-Hao Chen, Shunlin Lu, Ailing Zeng, Hao Zhang, Benyou Wang, Ruimao Zhang, and Lei Zhang. Motionllm: Understanding human behaviors from human motions and videos. arXiv preprint arXiv:2405.20340, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 250, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 294, + 293 + ], + "type": "text", + "content": "[33] Matthias Plappert, Christian Mandery, and Tamim Asfour. Learning a bidirectional mapping between human whole-body motion and natural language using deep recurrent neural networks. RAS, 109:13-26, 2018. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 293, + 294, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 293, + 294, + 336 + ], + "spans": [ + { + "bbox": [ + 56, + 293, + 294, + 336 + ], + "type": "text", + "content": "[34] Fangzhou Hong, Mingyuan Zhang, Liang Pan, Zhongang Cai, Lei Yang, and Ziwei Liu. Avatarclip: Zero-shot text-driven generation and animation of 3d avatars. ACM SIGGRAPH, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 338, + 294, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 370 + ], + "type": "text", + "content": "[35] Nikos Athanasiou, Mathis Petrovich, Michael J Black, and Gül Varol. Teach: Temporal action composition for 3d humans. In 3DV, pages 414-423, 2022." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 371, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 294, + 415 + ], + "type": "text", + "content": "[36] Rishabh Dabral, Muhammad Hamza Mughal, Vladislav Golyanik, and Christian Theobalt. Mofusion: A framework for denoising-diffusion-based motion synthesis. In CVPR, pages 9760-9770, 2023." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 415, + 294, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 294, + 458 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 294, + 458 + ], + "type": "text", + "content": "[37] Mingyuan Zhang, Xinying Guo, Liang Pan, Zhongang Cai, Fangzhou Hong, Huirong Li, Lei Yang, and Ziwei Liu. Remodiffuse: Retrieval-augmented motion diffusion model. In ICCV, 2023." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 460, + 294, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 294, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 294, + 502 + ], + "type": "text", + "content": "[38] Weilin Wan, Zhiyang Dou, Taku Komura, Wenping Wang, Dinesh Jayaraman, and Lingjie Liu. Tlcontrol: Trajectory and language control for human motion synthesis. ECCV, 2024." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 503, + 294, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 503, + 294, + 546 + ], + "spans": [ + { + "bbox": [ + 56, + 503, + 294, + 546 + ], + "type": "text", + "content": "[39] Chuan Guo, Yuxuan Mu, Muhammad Gohar Javed, Sen Wang, and Li Cheng. Momask: Generative masked modeling of 3d human motions. In CVPR, pages 1900-1910, 2024." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 548, + 294, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 581 + ], + "type": "text", + "content": "[40] Jinpeng Liu, Wenxun Dai, Chunyu Wang, Yiji Cheng, Yansong Tang, and Xin Tong. Plan, posture and go: Towards open-world text-to-motion generation. ECCV, 2024." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 582, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 294, + 613 + ], + "type": "text", + "content": "[41] Bo Han, Hao Peng, Minjing Dong, Yi Ren, Yixuan Shen, and Chang Xu. Amd: Autoregressive motion diffusion. In AAAI, pages 2022-2030, 2024." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 614, + 294, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 294, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 294, + 658 + ], + "type": "text", + "content": "[42] Zhenyu Xie, Yang Wu, Xuehao Gao, Zhongqian Sun, Wei Yang, and Xiaodan Liang. Towards detailed text-to-motion synthesis via basic-to-advanced hierarchical diffusion model. In AAAI, pages 6252–6260, 2024." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[43] Wenyang Zhou, Zhiyang Dou, Zeyu Cao, Zhouyingcheng Liao, Jingbo Wang, Wenjia Wang, Yuan Liu, Taku Komura, Wenping Wang, and Lingjie Liu. Emdm: Efficient motion diffusion model for fast, high-quality motion generation. ECCV, 2024." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 315, + 72, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 553, + 116 + ], + "type": "text", + "content": "[44] Mathis Petrovich, Or Litany, Umar Iqbal, Michael J Black, Gul Varol, Xue Bin Peng, and Davis Rempe. Multi-track timeline control for text-driven 3d human motion generation. In CVPRW, pages 1911-1921, 2024." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 118, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 118, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 315, + 118, + 553, + 150 + ], + "type": "text", + "content": "[45] Boxun Hu, Mingze Xia, Ding Zhao, and Guanlin Wu. Mona: Moving object detection from videos shot by dynamic camera, 2025." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 153, + 553, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 153, + 553, + 208 + ], + "spans": [ + { + "bbox": [ + 315, + 153, + 553, + 208 + ], + "type": "text", + "content": "[46] Zan Wang, Yixin Chen, Baoxiong Jia, Puhao Li, Jinlu Zhang, Jingze Zhang, Tengyu Liu, Yixin Zhu, Wei Liang, and Siyuan Huang. Move as you say interact as you can: Language-guided human motion generation with scene affordance. In CVPR, pages 433-444, 2024." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 210, + 553, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 210, + 553, + 243 + ], + "spans": [ + { + "bbox": [ + 315, + 210, + 553, + 243 + ], + "type": "text", + "content": "[47] Zhongfei Qing, Zhongang Cai, Zhitao Yang, and Lei Yang. Story-to-motion: Synthesizing infinite and controllable character animation from long text, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 244, + 553, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 244, + 553, + 277 + ], + "spans": [ + { + "bbox": [ + 315, + 244, + 553, + 277 + ], + "type": "text", + "content": "[48] Hui En Pang, Zhongang Cai, Lei Yang, Tianwei Zhang, and Ziwei Liu. Benchmarking and analyzing 3d human pose and shape estimation beyond algorithms. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 280, + 553, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 280, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 315, + 280, + 553, + 312 + ], + "type": "text", + "content": "[49] Gyeongsik Moon, Hongsuk Choi, and Kyoung Mu Lee. Neuralannnot: Neural annotator for 3d human mesh training sets. In CVPRW, 2022." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 315, + 314, + 553, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 314, + 553, + 357 + ], + "spans": [ + { + "bbox": [ + 315, + 314, + 553, + 357 + ], + "type": "text", + "content": "[50] Gyeongsik Moon, Hongsuk Choi, Sanghyuk Chun, Jiyoung Lee, and Sangdoo Yun. Three recipes for better 3d pseudogots of 3d human mesh estimation in the wild. In CVPRW, 2023." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 315, + 360, + 553, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 360, + 553, + 404 + ], + "spans": [ + { + "bbox": [ + 315, + 360, + 553, + 404 + ], + "type": "text", + "content": "[51] Hongwei Yi, Hualin Liang, Yifei Liu, Qiong Cao, Yandong Wen, Timo Bolkart, Dacheng Tao, and Michael J Black. Generating holistic 3d human motion from speech. In CVPR, pages 469-480, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 315, + 406, + 553, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 406, + 553, + 427 + ], + "spans": [ + { + "bbox": [ + 315, + 406, + 553, + 427 + ], + "type": "text", + "content": "[52] C.J. Bowen and R. Thompson. Grammar of the Edit. Focal Press, 2013. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 315, + 430, + 553, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 430, + 553, + 472 + ], + "spans": [ + { + "bbox": [ + 315, + 430, + 553, + 472 + ], + "type": "text", + "content": "[53] Mathis Petrovich, Or Litany, Umar Iqbal, Michael J. Black, Gül Varol, Xue Bin Peng, and Davis Rempe. Multi-track timeline control for text-driven 3d human motion generation. In CVPRW, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 315, + 475, + 553, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 475, + 553, + 529 + ], + "spans": [ + { + "bbox": [ + 315, + 475, + 553, + 529 + ], + "type": "text", + "content": "[54] Yuhong Zhang, Jing Lin, Ailing Zeng, Guanlin Wu, Shunlin Lu, Yurong Fu, Yuanhao Cai, Ruimao Zhang, Haoqian Wang, and Lei Zhang. Motion-" + }, + { + "bbox": [ + 315, + 475, + 553, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{x} + +" + }, + { + "bbox": [ + 315, + 475, + 553, + 529 + ], + "type": "text", + "content": ": A large-scale multimodal 3d whole-body human motion dataset. arXiv preprint arXiv:2501.05098, 2025. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 315, + 532, + 553, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 532, + 553, + 575 + ], + "spans": [ + { + "bbox": [ + 315, + 532, + 553, + 575 + ], + "type": "text", + "content": "[55] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In CVPR, pages 5152-5161, 2022. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 315, + 578, + 553, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 578, + 553, + 610 + ], + "spans": [ + { + "bbox": [ + 315, + 578, + 553, + 610 + ], + "type": "text", + "content": "[56] Georgios Pavlakos, Jitendra Malik, and Angjoo Kanazawa. Human mesh recovery from multiple shots. In CVPR, 2022. 2, 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 315, + 613, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 613, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 315, + 613, + 553, + 667 + ], + "type": "text", + "content": "[57] Peng Wu, Xiankai Lu, Jianbing Shen, and Yilong Yin. Clip fusion with bi-level optimization for human mesh reconstruction from monocular videos. In ACM MM, page 105-115, New York, NY, USA, 2023. Association for Computing Machinery." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 315, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 315, + 670, + 553, + 712 + ], + "type": "text", + "content": "[58] Kuan-Chieh Wang, Zhenzhen Weng, Maria Xenochristou, Joao Pedro Araujo, Jeffrey Gu, C Karen Liu, and Serena Yeung. Nemo: 3d neural motion fields from multiple video instances of the same action. In CVPR, 2023." + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1982" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[59] Fabien Baradel, Thibault Groueix, Philippe Weinzaepfel, Romain Brégier, Yannis Kalantidis, and Grégory Rogez. Leveraging mocap data for human mesh recovery. In 3DV, pages 586-595, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 149 + ], + "type": "text", + "content": "[60] Ruilong Li, Shan Yang, David A. Ross, and Angjoo Kanazawa. Ai choreographer: Music conditioned 3d dance generation with aist++, 2021. 3, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 151, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 151, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 151, + 294, + 194 + ], + "type": "text", + "content": "[61] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE TPAMI, 36(7):1325-1339, 2014. 3, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 195, + 294, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 195, + 294, + 227 + ], + "spans": [ + { + "bbox": [ + 56, + 195, + 294, + 227 + ], + "type": "text", + "content": "[62] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. ACM TOG, 36(6), 2017. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 228, + 294, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 294, + 281 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 294, + 281 + ], + "type": "text", + "content": "[63] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J. Black. Keep it SMPL: Automatic estimation of 3D human pose and shape from a single image. In Computer Vision - ECCV 2016. Springer International Publishing, 2016." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 283, + 294, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 283, + 294, + 315 + ], + "spans": [ + { + "bbox": [ + 56, + 283, + 294, + 315 + ], + "type": "text", + "content": "[64] Anurag* Arnab, Carl* Doersch, and Andrew Zisserman. Exploiting temporal context for 3d human pose estimation in the wild. In CVPR, 2019." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 316, + 294, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 294, + 348 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 294, + 348 + ], + "type": "text", + "content": "[65] Ahmed A. A. Osman, Timo Bolkart, and Michael J. Black. STAR: Sparse Trained Articulated Human Body Regressor, page 598-613. Springer International Publishing, 2020." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 350, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 294, + 392 + ], + "type": "text", + "content": "[66] Yinghao Huang, Federica Bogo, Christoph Lassner, Angjoo Kanazawa, Peter V. Gehler, Javier Romero, Ijaz Akhter, and Michael J. Black. Towards accurate marker-less human shape and pose estimation over time. In 3DV, 2017. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 394, + 294, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 394, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 394, + 294, + 426 + ], + "type": "text", + "content": "[67] Angjoo Kanazawa, Michael J. Black, David W. Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In CVPR, 2018. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 427, + 294, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 427, + 294, + 458 + ], + "spans": [ + { + "bbox": [ + 56, + 427, + 294, + 458 + ], + "type": "text", + "content": "[68] Muhammed Kocabas, Nikos Athanasiou, and Michael J. Black. Vibe: Video inference for human body pose and shape estimation. In CVPR, 2020." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 460, + 294, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 294, + 492 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 294, + 492 + ], + "type": "text", + "content": "[69] Nikos Kolotouros, Georgios Pavlakos, and Kostas Dani-ilidis. Convolutional mesh regression for single-image human shape reconstruction. In CVPR, 2019." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 494, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 494, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 494, + 294, + 525 + ], + "type": "text", + "content": "[70] Zhengyi Luo, S. Alireza Golestaneh, and Kris M. Kitani. 3d human motion estimation via motion compression and refinement. In ACCV, 2020. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 526, + 294, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 526, + 294, + 568 + ], + "spans": [ + { + "bbox": [ + 56, + 526, + 294, + 568 + ], + "type": "text", + "content": "[71] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS. Curran Associates, Inc., 2017. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 571, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 571, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 571, + 294, + 624 + ], + "type": "text", + "content": "[72] Chun-Hao P. Huang, Hongwei Yi, Markus Höschle, Matvey Safroshkin, Tsvetelina Alexiadis, Senya Polikovsky, Daniel Scharstein, and Michael J. Black. Capturing and inferring dense full-body human-scene contact. In CVPR, pages 13274-13285, 2022. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 625, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 294, + 678 + ], + "type": "text", + "content": "[73] Manuel Kaufmann, Jie Song, Chen Guo, Kaiyue Shen, Tianjian Jiang, Chengcheng Tang, Juan Jose Zárate, and Otmar Hilliges. EMDB: The Electromagnetic Database of Global 3D Human Pose and Shape in the Wild. In ICCV, 2023. 3, 7, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "type": "text", + "content": "[74] Xinyu Yi, Yuxiao Zhou, and Feng Xu. Transpose: Real-time 3d human translation and pose estimation with six inertial sensors. ACM TOG, 40(4), 2021. 3" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 700 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[75] Zachary Teed and Jia Deng. Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. In NeurIPS, pages 16558-16569. Curran Associates, Inc., 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 108, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 108, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 108, + 553, + 129 + ], + "type": "text", + "content": "[76] Zachary Teed, Lahav Lipson, and Jia Deng. Deep patch visual odometry. NeurIPS, 2023. 3, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 130, + 553, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 553, + 163 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 553, + 163 + ], + "type": "text", + "content": "[77] Weirong Chen, Le Chen, Rui Wang, and Marc Pollefeys. Leap-vo: Long-term effective any point tracking for visual odometry. In CVPR, 2024. 3, 4, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 164, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 164, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 164, + 553, + 206 + ], + "type": "text", + "content": "[78] Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, and Michael J. Black. AMASS: Archive of motion capture as surface shapes. In ICCV, pages 5442-5451, 2019. 3, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 209, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 553, + 251 + ], + "type": "text", + "content": "[79] Timo von Marcard, Roberto Henschel, Michael Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In ECCV, 2018. 3, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 254, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 553, + 285 + ], + "type": "text", + "content": "[80] Yufei Xu, Jing Zhang, Qiming Zhang, and Dacheng Tao. ViTPose: Simple vision transformer baselines for human pose estimation. In NeurIPS, 2022. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 288, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 288, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 288, + 553, + 352 + ], + "type": "text", + "content": "[81] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 4, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 354, + 553, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 396 + ], + "type": "text", + "content": "[82] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multiperson linear model. ACM TOG, 34(6):248:1-248:16, 2015. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 399, + 553, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 431 + ], + "type": "text", + "content": "[83] Qingqiu Huang, Yu Xiong, Anyi Rao, Jiaze Wang, and Dahua Lin. Movienet: A holistic dataset for movie understanding. In ECCV, 2020. 4, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 434, + 553, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 553, + 465 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 553, + 465 + ], + "type": "text", + "content": "[84] MMTracking Contributors. MMTracking: OpenMMLab video perception toolbox and benchmark. https://github.com/open-mmlab/mmtracking, 2020.4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 468, + 553, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 468, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 316, + 468, + 553, + 499 + ], + "type": "text", + "content": "[85] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. CoTracker: It is better to track together. In ECCV, 2024. 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 501, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 501, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 501, + 553, + 544 + ], + "type": "text", + "content": "[86] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. In ICCV, 2023. 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 545, + 553, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 553, + 589 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 553, + 589 + ], + "type": "text", + "content": "[87] Martin A. Fischler and Robert C. Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Commun. ACM, 24(6):381-395, 1981. 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 591, + 553, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 591, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 316, + 591, + 553, + 623 + ], + "type": "text", + "content": "[88] Jie Yang, Ailing Zeng, Shilong Liu, Feng Li, Ruimao Zhang, and Lei Zhang. Explicit box detection unifies end-to-end multi-person pose estimation. In ICLR, 2023. 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 624, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 656 + ], + "type": "text", + "content": "[89] Richard Hartley and Andrew Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, USA, 2 edition, 2003. 6" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 658, + 553, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 700 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 700 + ], + "type": "text", + "content": "[90] Michael J. Black, Priyanka Patel, Joachim Tesch, and Jinlong Yang. BEDLAM: A synthetic dataset of bodies exhibiting detailed lifelike animated motion. In CVPR, pages 8726-8737, 2023. 7" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1983" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/HumanRig_ Learning Automatic Rigging for Humanoid Character in a Large Scale Dataset/a5721584-3cc2-4564-a9c0-b5eefdb73fbc_content_list.json b/2025/HumanRig_ Learning Automatic Rigging for Humanoid Character in a Large Scale Dataset/a5721584-3cc2-4564-a9c0-b5eefdb73fbc_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6691a9b67a2f0d0c3f503eeef9732d1fc884c954 --- /dev/null +++ b/2025/HumanRig_ Learning Automatic Rigging for Humanoid Character in a Large Scale Dataset/a5721584-3cc2-4564-a9c0-b5eefdb73fbc_content_list.json @@ -0,0 +1,1400 @@ +[ + { + "type": "text", + "text": "HumanRig: Learning Automatic Rigging for Humanoid Character in a Large Scale Dataset", + "text_level": 1, + "bbox": [ + 107, + 128, + 890, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zedong Chu, Feng Xiong, Meiduo Liu, Jinzhi Zhang, Mingqi Shao, Zhaoxu Sun, Di Wang, Mu Xu Amap, Alibaba Group {chuzedong.czd, xf250971, liumeiduo.l, wushou.zjz, mingqi.smq, szx430629, wd426042, xumu.xm} @alibaba-inc.com", + "bbox": [ + 107, + 203, + 888, + 275 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 308, + 326, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the rapid evolution of 3D generation algorithms, the cost of producing 3D humanoid character models has plummeted, yet the field is impeded by the lack of a comprehensive dataset for automatic rigging—a pivotal step in character animation. Addressing this gap, we present HumanRig, the first large-scale dataset specifically designed for 3D humanoid character rigging, encompassing 11,434 meticulously curated T-posed meshes adhered to a uniform skeleton topology. Capitalizing on this dataset, we introduce an innovative, data-driven automatic rigging framework, which overcomes the limitations of GNN-based methods in handling complex AI-generated meshes. Our approach integrates a Prior-Guided Skeleton Estimator (PGSE) module, which uses 2D skeleton joints to provide a preliminary 3D skeleton, and a Mesh-Skeleton Mutual Attention Network (MSMAN) that fuses skeleton features with 3D mesh features extracted by a U-shaped point transformer. This enables a coarse-to-fine 3D skeleton joint regression and a robust skinning estimation, surpassing previous methods in quality and versatility. This work not only remedies the dataset deficiency in rigging research but also propels the animation industry towards more efficient and automated character rigging pipelines.", + "bbox": [ + 89, + 342, + 485, + 691 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 722, + 220, + 739 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The animation industry is undergoing a revolutionary change with the integration of machine learning and data-driven techniques, particularly in the realms of 3D modeling and character animation. A critical phase in bringing 3D humanoid characters to life is rigging, which entails skeleton construction and skinning to facilitate realistic motion. Traditionally, rigging is a labor-intensive task generally requiring skilled artists to manually construct skeleton joints and paint skinning weights. As content creation pipelines accelerate and grow in complexity, the need for automation", + "bbox": [ + 89, + 750, + 483, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e9243b13090b10c51ecb452479dcc6331d7d79804555d15f7fdd04c261953793.jpg", + "image_caption": [ + "Figure 1. The AI-generated mesh and the artist-created one show distinct face topology distributions. Manual meshes often have varying vertex densities across body parts and special topologies near joints for better deformation, while AI meshes tend to have chaotic topologies which lack semantic information. Previous rigging methods can only handle simple artist-created meshes, while HumanRig deals well with both AI and manual ones, especially for those with complex clothes or accessories and irregular body shapes." + ], + "image_footnote": [], + "bbox": [ + 516, + 321, + 908, + 579 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "in rigging has become more pressing than ever.", + "bbox": [ + 511, + 748, + 823, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Bolstered by extensive 3D datasets, the field of AI-driven 3D model generation [20, 23, 24, 27, 29] has experienced significant advancements. Yet, the advancement in automatic rigging remains constrained due to the lack of a comprehensive and standardized dataset, which is essential for training powerful and robust rigging models. Previous work, like [25], laid the cornerstone for earlier rigging methods but was limited by a small scale of 1,729 T-pose humanoid meshes with inconsistent skeleton topologies and", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "304", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "a lack of meaningful joint labels, hindering their use in animation. Studies such as [11], utilizing the SMPL dataset [13] and its clothed version from [2], were restricted to realistic human body shapes, missing the character diversity needed for varied applications. This has prompted the development of our HumanRig.", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We introduce HumanRig, standing out as the first large-scale dataset tailored for the task of rigging 3D humanoid character models, and offering 11,434 meticulously curated AI-generated humanoid meshes. All models are generated in T-pose and aligned with an industry-standard skeleton topology, which enables direct plug-and-play in standard animation engines. Unparalleled in scale and diversity, HumanRig features an extensive range of body proportions and character types, spanning from real persons to cartoon characters and humanoid animals.", + "bbox": [ + 89, + 183, + 480, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Utilizing the HumanRig dataset, we introduce an innovative automatic rigging framework that diverges from previous approaches [11, 12, 15-17, 25], which primarily rely on Graph Neural Networks (GNNs) [7, 21] to learn rigging directly from 3D mesh geometric features. Such methods struggle with complex 3D models, particularly those generated by AI. The core modules of our method are: a) The Prior-Guided Skeleton Estimator (PGSE), which initializes a coarse skeleton using 2D priors projected into 3D space, significantly reducing the complexity of the rigging task. b) A U-shaped Point Transformer [30] serving as our mesh encoder, disregarding edges to enhance rigging robustness on intricate meshes. c) The Mesh-Skeleton Mutual Attention Network (MSMAN) enriching mesh and skeleton features with mutual information within a high-level semantic space, enabling joint optimization of skeleton construction and skinning.", + "bbox": [ + 89, + 334, + 482, + 590 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our results showcase the method's superior performance, delivering high-quality automatic rigging solutions that eclipse current state-of-the-art techniques. This contribution not only addresses the scarcity of large-scale datasets for humanoid rigging but also presents an automatic rigging solution set to revolutionize the animation industry. By simplifying the rigging process, we pave the way for more efficient and automated character animation, unlocking new frontiers in creative expression and content creation.", + "bbox": [ + 89, + 592, + 480, + 728 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 744, + 232, + 758 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. 3D Rigging Datasets", + "text_level": 1, + "bbox": [ + 89, + 770, + 282, + 786 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The RigNetv1 dataset [25] has been instrumental in recent rigging research. It includes 2,703 rigged models, of which 1,729 are T-posed humanoid. Each character has between 1,000 and 5,000 vertices and an average of 25 skeleton joints. However, these models lack a standard skeleton topology, limiting the applicability of motion retargeting. SMPL [11] has gathered 3D data for real human body", + "bbox": [ + 89, + 794, + 480, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "types using the same skeleton, but it is restricted to hairless bodies without clothes, and all samples share similar topologies and head-to-body ratios. Additionally, meshes in current datasets are manually crafted, which markedly differ from AI-generated meshes (See Fig. 1). This discrepancy limits the generic learning of automatic rigging models for AI-generated humanoid characters. Our HumanRig dataset, consisting of 11,434 AI-generated T-pose meshes with varied head-to-body ratios aligned to a standard Mixamo skeleton, is crafted to boost data-driven methodologies for automatic humanoid rigging.", + "bbox": [ + 511, + 90, + 903, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Automatic Rigging", + "text_level": 1, + "bbox": [ + 511, + 265, + 694, + 281 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Automatic rigging consists of two tasks, skeleton construction and skinning. Skeleton construction is about setting up a hierarchical bone structure and the corresponding joint positions, according to the body anatomy and expected deformation. Skinning involves assigning influence to skeletons over vertices in 3D models, ensuring natural deformation during skeletal animation.", + "bbox": [ + 511, + 287, + 903, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Traditional Method. Pinocchio [1] employs the medial axis to position the skeletons via optimization, but cannot work well when the mesh shape deviates from the skeleton template. Geometric-based skinning methods [1, 8, 10, 22] assume skeleton influence is proportional to proximity, which can cause inaccuracies in meshes with irregular body shapes.", + "bbox": [ + 511, + 393, + 903, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Data-driven Method. RigNet [25] learns mesh offsets, attention, and bone connection probabilities, extracting joints via attention clustering and constructing skeletons with Minimum Spanning Tree. It can not support motion tasks well because the predicted skeleton has uncertain topology and lacks semantic labels. Based on 3D vertex features directly extracted by Graph Neural Network (GNN) backbone, [11, 15] learn template based skeletons and [11, 12, 15-17, 25] learn skinning weights. However, they all struggle with models that have diverse body shapes, intricate clothing and accessories, or complex face topologies. Unlike these approaches, our method effectively utilizes front view rendering to establish robust priors and employs a 2D-3D lifting technique to accurately determine skeleton positions with a coarse-to-fine approach, and overcome the skinning challenge by using a point transformer-based encoder to accurately capture vertex features in 3D space.", + "bbox": [ + 511, + 500, + 903, + 756 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. The Basis - HumanRig Dataset", + "text_level": 1, + "bbox": [ + 511, + 768, + 795, + 784 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As depicted in Fig. 2, the data acquisition pipeline for HumanRig is highly efficient and is detailed as follows.", + "bbox": [ + 511, + 794, + 903, + 823 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. 3D Humanoid Models Preparation", + "text_level": 1, + "bbox": [ + 511, + 832, + 815, + 848 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Constructing a large-scale 3D humanoid model dataset presents significant challenges, primarily due to the labor-intensive nature of requiring extensive input from profes", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "305", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f2581e7396d4c9cc3c6728c7be508d004f9b17b607b66658f688719c3702785d.jpg", + "image_caption": [ + "Figure 2. Data Acquisition Pipeline for our HumanRig dataset." + ], + "image_footnote": [], + "bbox": [ + 133, + 92, + 846, + 239 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/defa07d02366ce2461c9448c2ca561c511e7aa393bfd7203e25c36d4d50a2d10.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Dataset#Uniform SkelClothingHairBind PoseSource
Mixamo [3]108✓✓✓T-poseartist-created
SMPL [13]N/A✓××T-poseartist-created
RigNetv1 [25]1729×✓✓T-poseartist-created
HumanRig (ours)11434✓✓✓T-poseAI-generated
", + "bbox": [ + 91, + 296, + 498, + 357 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1. Humanoid Rigging Dataset Comparisons. For RigNetv1, only T-posed humanoid meshes picked out to compare.", + "bbox": [ + 89, + 368, + 483, + 396 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "sional artists. However, the rapid advancement of AI-based 2D image generation technology has enabled the creation of diverse humanoid character images from textual descriptions. Furthermore, leveraging AI-driven 3D model generation techniques, it has become feasible to efficiently derive high-quality 3D models from a single image.", + "bbox": [ + 88, + 424, + 482, + 515 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "T-pose Image Generation. 3D humanoid characters are commonly modeled in neutral poses like T-pose or A-pose, for enhanced mesh deformation and motion capture (mocap) data compatibility. Our work adopts the T-pose as the bind pose to streamline the rigging process for 3D humanoid models. We introduce a pose-controlled text-to-image framework for generating T-pose character images by integrating stylized text prompts with reference T-pose images. Our method includes the following steps:", + "bbox": [ + 88, + 515, + 482, + 651 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Curating 5,000 detailed text prompts for a range of characters adorned with various clothing accessories, by employing a Large Language Model (LLM) to enhance the prompt template.", + "2. Generating 50 T-pose maps with diverse head-to-body ratios by utilizing DWPose [26] to extract 2D poses from a curated set of humanoid images.", + "3. Synthesizing 20,000 T-posed humanoid images by randomly selecting a text prompt and a T-posed map, and inputting them into Stable Diffusion (SD) [18] and the corresponding Pose ControlNet [28].", + "4. Filtering out low-quality images, ultimately collecting 17,268 high-quality humanoid T-pose images." + ], + "bbox": [ + 89, + 654, + 482, + 852 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This framework guarantees that the generated characters capture the desired style and aesthetics while maintaining accurate body proportions and proper T-pose alignment. As", + "bbox": [ + 88, + 854, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f07dddad3c951309fe89465832b20e6552360a55add1545ee3abf270fa17170a.jpg", + "image_caption": [ + "Figure 3. Head-to-body Ratio Diversity Statistics of HumanRig." + ], + "image_footnote": [], + "bbox": [ + 513, + 292, + 906, + 426 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "a result, we can produce a broad range of diverse and precisely posed T-posed humanoid character images, covering various head-to-body ratios.", + "bbox": [ + 511, + 489, + 903, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "T-posed Mesh Generation. The swift progress in 3D generation has enabled the rapid production of large-scale and high-quality 3D meshes with unprecedented speed. We utilize an Image-to-3D pipeline for generating 3D meshes, which includes open-source methods like InstantMesh [24] and Unique3D [23], as well as commercial tools such as Hyperhuman Rodin [29] and Triposr-v2 [20]. By inputting meticulously prepared T-posed character images, we produce a large batch of T-posed humanoid meshes. We manually filter out low-quality ones, selecting 14,662 high-quality meshes from an initial 17,268.", + "bbox": [ + 511, + 534, + 906, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Humanoid Models Rigging", + "text_level": 1, + "bbox": [ + 511, + 710, + 756, + 727 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To animate a newly created humanoid character using mocap data, it should have a practical skeleton topology that is compatible with mocap data, rather than binding to joints without semantic labels, like [25]. To simplify the mocap-to-deformation process, we utilize the Mixamo skeleton topology, enabling our rigged character to be compatible with existing animation software and game engines. Naturally, we utilize the Mixamo semi-automated rigging tool [3] to speed up the annotation process. Poor-quality rigging data is filtered out, and artists manually refine the remaining models. Ultimately, we produce 11,434 high-quality rigged", + "bbox": [ + 511, + 734, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "306", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a56b1b4c34141ccb33ee030353204f2739e78a52ef6d853f8cc16615e4e93ccd.jpg", + "image_caption": [ + "Figure 4. Method Overview. Given a humanoid mesh, a coarse skeleton is predicted by an Prior-guided skeleton estimator (PGSE) and helps construct the skeleton-aware vertex features. They are fed into Skeleton Encoder and Mesh Encoder, respectively, then fused by a Mesh-Skeleton Mutual Attention Network to predict a refined skeleton and skinning weights with a joint learning strategy. Finally, the skeleton and skinning weights are combined to produce the animation-ready character." + ], + "image_footnote": [], + "bbox": [ + 117, + 125, + 870, + 352 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "humanoid models.", + "bbox": [ + 89, + 450, + 217, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Data Postprocessing", + "text_level": 1, + "bbox": [ + 89, + 478, + 282, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We normalize the prepared rigged humanoid models within a bounding box of $[-0.5, 0.5]$ to ensure proper alignment. From the rigging results, we extract 3D skeleton joint positions of size $[s, 3]$ and the skinning weight matrix of size $[m, s]$ , where $s$ represents the number of skeleton joints, uniformly set to 22 in this work, and $m$ denotes the variable number of vertices across different humanoid models. Additionally, we generate a front-view image using a fixed camera. Utilizing the projection matrix derived from the camera's intrinsic and extrinsic parameters, we project the 3D skeleton joints onto the front-view image, thereby obtaining the corresponding 2D skeleton joint positions.", + "bbox": [ + 89, + 502, + 483, + 684 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Data Summary", + "text_level": 1, + "bbox": [ + 89, + 695, + 243, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As presented in Table 1, we conduct a comparative analysis of our dataset with existing rigged humanoid datasets. Existing datasets are limited in quantity, diversity or skeleton unity. Our dataset bridges these gaps by offering a large-scale collection of 11,434 AI-generated meshes with diverse head-to-body ratios, aligned with a consistent Mixamo skeleton, thereby enhancing both the quantity and diversity of available models for rigging research and applications. Fig. 3 illustrates models within our dataset, showcasing the diverse head-to-body ratios and their corresponding data proportions. Each annotated sample comprises a rigged T-pose humanoid mesh, 3D skeleton joint positions,", + "bbox": [ + 89, + 719, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "a skinning weight matrix, a front-view image, and the associated camera parameters along with the 2D skeleton joint positions.", + "bbox": [ + 511, + 450, + 906, + 497 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Automatic Rigging", + "text_level": 1, + "bbox": [ + 511, + 510, + 694, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Capitalizing on the expansive HumanRig dataset characterized by a uniform skeleton topology, we propose a data-driven automatic rigging framework. As depicted in Fig. 4, our approach begins by initializing a coarse skeleton using a Prior-Guided Skeleton Estimator (PGSE). The skeleton serves a dual purpose: It facilitates the learning of joint positions from coarse to fine, and it enhances vertex attributes to yield skeleton-aware vertex features, thereby simplifying the learning process for both the skeleton and skinning weights. Furthermore, we employ two encoders to extract skeleton and mesh features respectively: a MLP-based Skeleton Encoder for extracting skeleton features and a Point Transformer-based Mesh Encoder for extracting mesh features. These distinct features are then cross-integrated through a Mesh-Skeleton Mutual Attention Network (MS-MAN) to achieve deeper feature integration. We will introduce each of these modules individually.", + "bbox": [ + 511, + 536, + 906, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Prior-guided Skeleton Estimator", + "text_level": 1, + "bbox": [ + 511, + 801, + 799, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Prior-Guided Skeleton Estimator (PGSE) utilizes 2D skeleton priors from the model's front view and projects them into approximate 3D positions using camera parameters. This approach significantly simplifies the task of learning skeleton positions directly from 3D mesh geom", + "bbox": [ + 511, + 824, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "307", + "bbox": [ + 486, + 944, + 509, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "etry by providing an initial estimation, which can then be refined using mesh data. Specifically, we fine-tune RTM-Pose [9] on the front view image to predict 2D template joints $P_{j2D} \\in \\mathbb{R}^{s \\times 2}$ accurately. Then, each 2D joint $\\mathbf{J}_{2D}$ is back-projected into 3D as a ray:", + "bbox": [ + 88, + 90, + 480, + 167 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {X}} _ {r} (\\mu ; \\mathbf {J} _ {2 D}) = \\mathrm {P} _ {c} \\tilde {\\mathbf {J}} _ {2 D} + \\mu \\tilde {\\mathbf {X}} _ {c}\n$$\n", + "text_format": "latex", + "bbox": [ + 183, + 176, + 388, + 194 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{P}_c\\in \\mathbb{R}^{4\\times 3}$ is the pseudo-inverse of the camera projection matrix, $\\mathbf{X}_c$ is the 3D location of the camera center, and the symbol with superscript tilde denotes the corresponding homogeneous coordinate. Finally, we determine the coarse 3D skeleton $P_{skec}\\in \\mathbb{R}^{s\\times 3}$ by calculating the intersections of each ray with the mesh surface and using the midpoints of the first and last intersections for each 3D coarse joint.", + "bbox": [ + 88, + 204, + 482, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Skeleton and Mesh Encoder", + "text_level": 1, + "bbox": [ + 89, + 335, + 341, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We utilize a Skeleton Encoder composed of a simple 3-layer Multi-Layer Perceptron (MLP) to process the coarse 3D skeleton points and extract skeleton features $f_{s} \\in \\mathbb{R}^{s \\times c}$ , where $c$ denotes the number of channels. This encoder is efficient due to the skeleton's fixed topology and limited number of joints.", + "bbox": [ + 89, + 357, + 482, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For mesh feature extraction, we begin by extracting all vertices and constructing skeleton-aware vertex features $f_{v} \\in \\mathbb{R}^{m \\times (3 + s)}$ , where each of the $m$ vertices has 3 positional coordinates and $s$ euclidean distance features relative to $P_{\\text{skec}}$ . We then employ a U-shaped Point Transformer [30] as the Mesh Encoder to extract deep mesh features, with the output of the last up-transition layer serving as our encoder output $f_{m} \\in \\mathbb{R}^{m \\times c}$ . The Point Transformer-based Mesh Encoder effectively integrates local geometric correlation features with global geometric semantic features. Compared to commonly used GNN-based Mesh Encoders, it exhibits superior generalization for AI-generated meshes that often have irregular face topology and incorporate a large number of vertices.", + "bbox": [ + 89, + 449, + 483, + 659 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Mesh-Skeleton Mutual Attention Network", + "text_level": 1, + "bbox": [ + 89, + 669, + 449, + 684 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Mesh-skeleton Mutual Attention Network (MSMAN) is further proposed to integrate Skeleton features $f_{s}$ and Mesh features $f_{m}$ through mutual cross-attention mechanisms. The skeleton features provide mesh features with body part affiliation, which aids in the semantic understanding of mesh vertices and enhances the accuracy of skinning weight predictions. In a similar vein, mesh features bolster the local feature representation of the skeleton, which is beneficial for refining the precise positions of the skeleton. Taking one side as an example, we introduce a multi-head cross-attention mechanism to enhance mesh features, which is formalized as follows:", + "bbox": [ + 89, + 691, + 483, + 872 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nQ = f _ {s}; \\quad K = f _ {m}; \\quad V = f _ {m} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 883, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf _ {m \\rightarrow s} = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d}}\\right) V \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 101, + 903, + 135 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $f_{m\\to s}\\in \\mathbb{R}^{s\\times c}$ is the enhanced skeleton features.", + "bbox": [ + 511, + 138, + 877, + 152 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the other side of the interaction, we adopt the same formulation but set $Q = f_{m}$ ; $K = f_{s}$ ; $V = f_{s}$ to integrate the skeleton features into the mesh features. Finally, we get $f_{s\\rightarrow m}\\in \\mathbb{R}^{m\\times c}$ , which represents skeleton-aware mesh features.", + "bbox": [ + 511, + 154, + 903, + 228 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Rigging Head and Training Loss", + "text_level": 1, + "bbox": [ + 511, + 237, + 799, + 253 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We employ two separate MLP-based heads to independently predict skeleton positions $P_{ske} \\in \\mathbb{R}^{s \\times 3}$ and skinning weights $P_{skin} \\in \\mathbb{R}^{m \\times s}$ . For skeleton joints prediction, we use Mean Squared Error (MSE) Loss:", + "bbox": [ + 511, + 260, + 903, + 320 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s k e l e t o n}} = \\frac {1}{s} \\sum_ {i = 1} ^ {s} \\left(P _ {\\text {s k e} _ {i}} - G _ {\\text {s k e} _ {i}}\\right) ^ {2} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 325, + 903, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Skinning prediction can be conceptualized as optimizing the selection probabilities of various skeleton joints for each mesh vertex. Consequently, we employ the Kullback-Leibler divergence loss to minimize the distance between the predicted skin weight distribution $P_{\\text{skin}}$ and the ground truth distribution $G_{\\text{skin}}$ :", + "bbox": [ + 511, + 371, + 905, + 460 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s k i n n i n g}} = \\frac {1}{m} \\sum_ {i = 1} ^ {m} \\sum_ {j = 1} ^ {s} P _ {\\text {s k i n} _ {i j}} \\left(\\log \\frac {P _ {\\text {s k i n} _ {i j}}}{G _ {\\text {s k i n} _ {i , j}}}\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 465, + 903, + 507 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The total loss is constructed as:", + "bbox": [ + 513, + 512, + 723, + 526 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {s k i n n i n g}} + \\mathcal {L} _ {\\text {s k e l e t o n}} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 535, + 903, + 550 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 511, + 558, + 643, + 574 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details. Our framework is trained in two stages. First, we use the mmpose [4] framework to train a 2D joint predictor based on RTMPose [9]. We then apply PGSE to extract coarse skeletons from input meshes, which are used in the second stage to train other components. Both stages are trained on our HumanRig dataset, with the same $80\\% -10\\% -10\\%$ data split. The training is performed on 2 NVIDIA RTX A6000 GPUs using the AdamW [14] optimizer and a MultiStepScheduler. The training lasts for 300 epochs, with the learning rate halved after 50 epochs, starting from an initial rate of 1e-3 and a batch size of 16.", + "bbox": [ + 511, + 583, + 905, + 748 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation metrics. For skeleton construction evaluation, we employ the metrics CD-J2J, CD-J2B and CD-B2B proposed by [25]. For skinning evaluation, we adopt the same metrics as [12, 16, 17, 25], including skinning precision and L1-norm. Interested readers are referred to the original papers for detailed definitions of these metrics. For deformation quality, we evaluate deformation error by calculating the average Euclidean distance between deformed vertex positions and ground truth ones, using 10 random poses with joint rotations within a range of $\\pm 10$ degrees.", + "bbox": [ + 511, + 750, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "308", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/01e69de8fcabe582de7935a7fed4b08ed568a7a345d9fbbd2ecd49050e35d5e9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Train SetRGNetv1-human(Test)HumanRig(Test)
CD-J2J↓CD-J2B↓CD-B2B↓Prec↑L1↓CD-J2J↓CD-J2B↓CD-B2B↓Prec↑L1↓
RGNetv1-human0.02740.02580.02080.74450.25730.01630.01550.01260.83640.1700
HumanRig-small0.01890.01730.01100.65020.33250.00540.00500.00280.83710.0789
HumanRig0.00710.00640.00330.74520.15640.00270.00250.00120.92710.0610
", + "bbox": [ + 96, + 89, + 906, + 174 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/dbc5e17b2d96f8681222f4cab277bf62e8fdf727477fd7a20288340a57c42889.jpg", + "table_caption": [ + "Table 2. Cross-dataset comparisons of RigNetv1-human and HumanRig." + ], + "table_footnote": [], + "table_body": "
MethodCD-J2J↓CD-J2B↓CD-B2B↓Prec↑L1↓
Ours(w/o PGSE)0.01100.00610.00310.83630.1630
Ours(w/o MSMAN)0.00800.00760.00390.92140.0630
Ours(complete)0.00270.00250.00120.92710.0610
", + "bbox": [ + 94, + 222, + 500, + 281 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0762beff6a4a47ca66afb966996b3b6204165e20fea3bf2fdfa635c6e2ffbe92.jpg", + "table_caption": [ + "Table 3. Architecture Study that shows the influence of PGSE and MSMAN." + ], + "table_footnote": [], + "table_body": "
Mesh EncoderCD-J2J↓CD-J2B↓CD-B2B↓Prec↑L1↓
GraphSAGE [6]0.00310.00300.00160.81530.0764
GraphTransformer [19]0.00300.00290.00160.81000.0760
GMEdgeNet [25]0.00280.00260.00140.82900.0751
Ours(Point Transformer)0.00270.00250.00120.92710.0610
", + "bbox": [ + 94, + 335, + 500, + 402 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4. Mesh Encoder Study that shows our superior performance compared to GNNs.", + "bbox": [ + 89, + 412, + 483, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. HumanRig Dataset Study", + "text_level": 1, + "bbox": [ + 89, + 468, + 323, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To assess the effectiveness of our HumanRig dataset, we perform a comparative study with RigNetv1. RigNetv1, the largest rigging dataset, lacks a uniform skeleton and consistent initial poses. To ensure a fair comparison, we select 1,729 T-posed humanoid meshes, rigged them with the Mixamo skeleton via our annotation pipeline, and labeled them as RigNetv1-human. We train our HumanRig method on the training sets of RigNetv1-human and HumanRig respectively, and subsequently evaluated both models on both test sets. Additionally, to neutralize the effect of dataset size, we created HumanRigsmall, a random subset of HumanRig comprising 1,729 samples. Table 2 demonstrate valuable insights. First, a model trained on HumanRigsmall surpasses one trained on RigNetv1-human on the HumanRig test set, even outperforming the intra-dataset training of RigNetv1-human in skeleton construction task. Second, expanding the dataset size to the full HumanRig dataset significantly improves performance in both tasks. Lastly, the model trained on our dataset excels on both artist-created and AI-generated models.", + "bbox": [ + 88, + 492, + 482, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In addition to the large scale, the superior performance is also attributed to the diverse body proportions (see Fig. 3). We underscore its importance by an experiment that categorizes the HumanRig training and test sets into subsets based on head-body ratios, with the subset containing five heads designated as HumanRig5. We then create a new, balanced subset, HumanRig $u$ , by evenly sampling from all subsets", + "bbox": [ + 89, + 795, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1ffdf905137edcd6f5da773f5271b3a5df5375204a72a79c4d25616937fb6f95.jpg", + "image_caption": [ + "Figure 5. Study on the Importance of Diversity in Body Shapes." + ], + "image_footnote": [], + "bbox": [ + 516, + 223, + 709, + 372 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/73b3546570a4b6e893a3d49eeef57e3e6c8dc0c5819ff7fa40aad8c929680513.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 223, + 903, + 372 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "to match the size of HumanRig5. We train two HumanRig models separately on HumanRig5 and HumanRigu, and then evaluate them on test subsets ranging from two-heads to nine-heads. The experimental results in Fig. 5 indicate that for the model trained on HumanRig5, the performance declines as the head-body ratio deviates from the five heads in both tasks. When considering various head-to-body ratios, it exhibits a more balanced performance distribution and overall improvement.", + "bbox": [ + 511, + 426, + 906, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Architecture Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 573, + 767, + 589 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "PGSE Design. After eliminating PGSE, the network directly regresses joint positions from mesh features without the coarse skeleton's guidance. This also limits the mesh encoder's input to vertex coordinates. Table 3 demonstrates that PGSE significantly reduces the difficulty of skeleton construction (CD-J2J drops from 0.0110 to 0.0027). Additionally, skeleton-aware vertex features substantially improves skinning performance.", + "bbox": [ + 511, + 597, + 905, + 718 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MSMAN Design. The MSMAN is integral to our method as it enables the joint optimization of skeleton construction and skinning. To isolate the effects of MSMAN, we train an expert network that estimates skinning weights directly from the mesh encoder's output, and use the coarse skeleton as the final skeleton output. Table 3 illustrates that the coarse skeleton itself achieves commendable accuracy, but its refinement by MSMAN still results in a notable improvement. Furthermore, skinning performance is improved by the joint optimization. These findings underscore the significance of MSMAN's mutual attention mechanism in effectively integrating mesh features and skeletal repre", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "309", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d7ce88cfd4f0ed8429030f1305a40bcfdb9d42306a401554d2a7443950c0fe86.jpg", + "image_caption": [ + "Figure 6. Skeleton construction comparisons. RigNet outputs irregular skeletons, which presents challenges to motion applications. NBS performs poorly in joint position prediction. Our method achieves the most plausible humanoid skeletons according to different human shapes from character (a) to (e). The skeleton topology we utilized is illustrated in character (f)." + ], + "image_footnote": [], + "bbox": [ + 99, + 92, + 679, + 349 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/014e5cd5e2607b404378e0c9bffff381439d0a6a77b651228700a187015255c9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 686, + 93, + 893, + 349 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sentations within a high-level semantic space.", + "bbox": [ + 89, + 435, + 393, + 450 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Mesh Encoder Design. In this work, we deviate from the traditional approach of utilizing GNNs for mesh feature extraction by introducing a U-shaped Point Transformer. To investigate the impact of this design choice, we replace our mesh encoder with GMEdgeNet proposed by RigNet, and substitute its EdgeConv [21] operator with other widely-used GNNs such as GraphSAGE [6] and GraphTransformer [19]. The results presented in Table 4 demonstrate that our mesh encoder achieves superior performance. This improvement is due to the more comprehensive and efficient aggregation of mesh vertex features based on 3D spatial distances, compared to aggregation along surface edges in GNNs. Additionally, edges in AI-generated 3D models are often irregular and lack semantic information, thus, a mesh encoder that ignores edge information exhibits improved generalization and robustness.", + "bbox": [ + 91, + 452, + 483, + 694 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Comparisons", + "text_level": 1, + "bbox": [ + 89, + 707, + 230, + 724 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Due to the lack of effective datasets, there are few data-driven methods [11, 12, 15-17, 25] for studying humanoid neural rigging. Unfortunately, as far as we know, only RigNet [25] and NBS [11] have been open-sourced. We conduct the following experiments to demonstrate the superiority of our method.", + "bbox": [ + 89, + 732, + 482, + 823 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Skeleton Construction. For skeleton prediction, we compare our method with RigNet [25] and NBS [11], as illustrated in Fig. 6. RigNet results in a lack of different key joints in all samples because it does not incorporate a humanoid skeleton prior, leading to uncertain joint placements", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/360820de8f86e6ad31cd999094d4dfb4623993d33890b1d3ba3bec40c0a2f946.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Test setHeat Diffusion [1]GVB [5]RigNet [25]Ours
RigNetv1-human0.004050.003060.002520.00198
HumanRig0.002360.00224N/A0.00084
", + "bbox": [ + 517, + 431, + 924, + 482 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Deformation Error Study. RigNet is not evaluated on HumanRig as it only handles 1k-5k vertex meshes. Our method achieves the best deformation quality on both test sets.", + "bbox": [ + 511, + 491, + 906, + 532 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and bone connections without semantic context. Although NBS utilizes a fixed skeleton template of SMPL [13], it produces inaccurate joint positions, especially in the lower limbs. This inaccuracy stems from the challenge of directly regressing joint positions from mesh features. Moreover, as it exclusively trained on the SMPL dataset, it struggles to generalize well to diverse meshes that exhibit varying head-to-body ratios. In contrast, by leveraging a robust 2D prior, we predict joints based on a standard skeleton template, resulting in more plausible skeletons.", + "bbox": [ + 511, + 566, + 906, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "**Skinning and Deformation.** We conducted a comparative analysis with Heat Diffusion [1], Geodesic Voxel Binding (GVB) [5], and RigNet [25]. These methods can predict skinning weights given an input mesh and its corresponding skeleton. Heat Diffusion and GVB are traditional geometry-based methods, whereas RigNet employs a data-driven approach by training an independent skinning network. For fair comparison, we input them the same skeletons predicted by our HumanRig method. Table 5 presents the quantitative assessment of deformation errors. RigNet can not handle complex meshes with over 5K vertices, while our approach outperforms on both artist-created and AI-", + "bbox": [ + 511, + 719, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "310", + "bbox": [ + 485, + 944, + 511, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b2a7100ac64b371e6b46aa6a3059c24f8cec891bf54e388a4467a1fd33bfc601.jpg", + "image_caption": [ + "Figure 7. Skinning and Deformation Comparisons. Different colors are assigned to each vertex according to the most influential joint, determined by skinning weights. Previous methods struggle with meshes that have special body shapes or intricate accessories (red box), while our method achieves superior deformation quality. More intuitive results are visualized in the supplementary video." + ], + "image_footnote": [], + "bbox": [ + 176, + 90, + 830, + 564 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "generated meshes. Additionally, Fig. 7 provides a qualitative visualization, which indicates that our method excels in maintaining more lifelike and fluid animations across varying head-to-body ratios, intricate clothes or accessories. This benefits from the strong point-transformer-based mesh encoder, which competently differentiates between distinct body parts to generate more precise skinning weights.", + "bbox": [ + 89, + 651, + 482, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusions", + "text_level": 1, + "bbox": [ + 89, + 768, + 218, + 784 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we have addressed the critical need for a comprehensive dataset and a robust framework for automatic rigging of 3D humanoid character models. We introduce HumanRig, a large-scale dataset of AI-generated T-pose humanoid models, all rigged with a consistent skeleton topology, significantly surpassing the previous datasets in terms of size, diversity, complexity and practical motion", + "bbox": [ + 89, + 794, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "applications. Leveraging the HumanRig dataset, we propose a novel automatic rigging framework that integrates a Prior-Guided Skeleton Estimator (PGSE), a U-shaped Point Transformer-based Mesh Encoder and a Mesh-Skeleton Mutual Attention Network (MSMAN). Our approach excels in skeleton construction and skinning prediction, effectively managing challenges from irregular shapes, detailed clothing or accessories, and differing head-to-body ratios. Compared to GNN-based approaches, our method provides superior generalization to complex, AI-generated meshes.", + "bbox": [ + 511, + 651, + 906, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our method still have limitations. First, the skeleton template has not considered small body parts such as fingers, which requires generating finer meshes in datasets. Second, it is exciting if our method could be extended to the quadruped or any objects.", + "bbox": [ + 511, + 801, + 908, + 878 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "311", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Ilya Baran and Jovan Popovic. Automatic rigging and animation of 3d characters. ACM Transactions on graphics (TOG), 26(3):72-es, 2007. 2, 7", + "[2] Bharat Lal Bhatnagar, Garvita Tiwari, Christian Theobalt, and Gerard Pons-Moll. Multi-garment net: Learning to dress 3d people from images. In IEEE International Conference on Computer Vision (ICCV). IEEE, 2019. 2", + "[3] Sue Blackman and Sue Blackman. Rigging with mixamo. Unity for Absolute Beginners, pages 565-573, 2014. 3", + "[4] MMPose Contributors. Openmmlab pose estimation toolbox and benchmark. https://github.com/open-mmlab/mmpose, 2020.5", + "[5] Olivier Dionne and Martin De Lasa. Geodesic binding for degenerate character geometry using sparse voxelization. IEEE transactions on visualization and computer graphics, 20(10):1367-1378, 2014. 7", + "[6] Will Hamilton, Zhitao Ying, and Jure Leskovec. Inductive representation learning on large graphs. Advances in neural information processing systems, 30, 2017. 6, 7", + "[7] Rana Hanocka, Amir Hertz, Noa Fish, Raja Giryes, Shachar Fleishman, and Daniel Cohen-Or. Meshconn: a network with an edge. ACM Transactions on Graphics (ToG), 38(4):1-12, 2019. 2", + "[8] Alec Jacobson, Ilya Baran, Jovan Popovic, and Olga Sorkine. Bounded biharmonic weights for real-time deformation. ACM Trans. Graph., 30(4):78, 2011. 2", + "[9] Tao Jiang, Peng Lu, Li Zhang, Ningsheng Ma, Rui Han, Chengqi Lyu, Yining Li, and Kai Chen. Rtmpose: Realtime multi-person pose estimation based on mmpose. arXiv preprint arXiv:2303.07399, 2023. 5", + "[10] Ladislav Kavan and Olga Sorkine. Elasticity-inspired deformers for character articulation. ACM Transactions on Graphics (TOG), 31(6):1-8, 2012. 2", + "[11] Peizhuo Li, Kfir Aberman, Rana Hanocka, Libin Liu, Olga Sorkine-Hornung, and Baoquan Chen. Learning skeletal articulations with neural blend shapes. ACM Transactions on Graphics (TOG), 40(4):1-15, 2021. 2, 7", + "[12] Lijuan Liu, Youyi Zheng, Di Tang, Yi Yuan, Changjie Fan, and Kun Zhou. Neuroskinning: Automatic skin binding for production characters with deep graph networks. ACM Transactions on Graphics (ToG), 38(4):1-12, 2019. 2, 5, 7", + "[13] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. Acm Transactions on Graphics, 34 (Article 248), 2015. 2, 3, 7", + "[14] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5", + "[15] Jing Ma and Dongliang Zhang. Tarig: Adaptive template-aware neural rigging for humanoid characters. Computers & Graphics, 114:158-167, 2023. 2, 7", + "[16] Albert Mosella-Montoro and Javier Ruiz-Hidalgo. Skinningnet: Two-stream graph convolutional neural network for skinning prediction of synthetic characters. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18593-18602, 2022. 5" + ], + "bbox": [ + 93, + 114, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Xiaoyu Pan, Jiancong Huang, Jiaming Mai, He Wang, Honglin Li, Tongkui Su, Wenjun Wang, and Xiaogang Jin. Heterskinnet: A heterogeneous network for skin weights prediction. In Proceedings of the ACM on Computer Graphics and Interactive Techniques. Association for Computing Machinery, 2021. 2, 5, 7", + "[18] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3", + "[19] Yunsheng Shi, Zhengjie Huang, Shikun Feng, Hui Zhong, Wenjin Wang, and Yu Sun. Masked label prediction: Unified message passing model for semi-supervised classification. arXiv preprint arXiv:2009.03509, 2020. 6, 7", + "[20] Dmitry Tochilkin, David Pankratz, Zexiang Liu, Zixuan Huang, Adam Letts, Yangguang Li, Ding Liang, Christian Laforte, Varun Jampani, and Yan-Pei Cao. Triposr: Fast 3d object reconstruction from a single image. arXiv preprint arXiv:2403.02151, 2024. 1, 3", + "[21] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (tog), 38(5):1-12, 2019. 2, 7", + "[22] Rich Wareham and Joan Lasenby. Bone glow: An improved method for the assignment of weights for mesh deformation. In Articulated Motion and Deformable Objects: 5th International Conference, AMDO 2008, Port d'Andratx, Mallorca, Spain, July 9-11, 2008. Proceedings 5, pages 63-71. Springer, 2008. 2", + "[23] Kailu Wu, Fangfu Liu, Zhihan Cai, Runjie Yan, Hanyang Wang, Yating Hu, Yueqi Duan, and Kaisheng Ma. Unique3d: High-quality and efficient 3d mesh generation from a single image. arXiv preprint arXiv:2405.20343, 2024. 1, 3", + "[24] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191, 2024.1,3", + "[25] Zhan Xu, Yang Zhou, Evangelos Kalogerakis, Chris Landreth, and Karan Singh. Rignet: neural rigging for articulated characters. ACM Transactions on Graphics (TOG), 39 (4):58-1, 2020. 1, 2, 3, 5, 6, 7", + "[26] Zhendong Yang, Ailing Zeng, Chun Yuan, and Yu Li. Effective whole-body pose estimation with two-stages distillation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4210-4220, 2023. 3", + "[27] Jinzhi Zhang, Feng Xiong, and Mu Xu. G3PT: unleash the power of autoregressive modeling in 3d generation via cross-scale querying transformer. CoRR, abs/2409.06322, 2024. 1", + "[28] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 3", + "[29] Longwen Zhang, Ziyu Wang, Qixuan Zhang, Qiwei Qiu, Anqi Pang, Haoran Jiang, Wei Yang, Lan Xu, and Jingyi Yu. Clay: A controllable large-scale generative model for creat-" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "312", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ing high-quality 3d assets. ACM Transactions on Graphics (TOG), 43(4):1-20, 2024. 1, 3", + "[30] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16259-16268, 2021. 2, 5" + ], + "bbox": [ + 91, + 90, + 483, + 176 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "313", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/HumanRig_ Learning Automatic Rigging for Humanoid Character in a Large Scale Dataset/a5721584-3cc2-4564-a9c0-b5eefdb73fbc_model.json b/2025/HumanRig_ Learning Automatic Rigging for Humanoid Character in a Large Scale Dataset/a5721584-3cc2-4564-a9c0-b5eefdb73fbc_model.json new file mode 100644 index 0000000000000000000000000000000000000000..fd926732ad0cd13921d4f1bda77b5fb8ca3d0cfd --- /dev/null +++ b/2025/HumanRig_ Learning Automatic Rigging for Humanoid Character in a Large Scale Dataset/a5721584-3cc2-4564-a9c0-b5eefdb73fbc_model.json @@ -0,0 +1,1782 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.108, + 0.13, + 0.892, + 0.175 + ], + "angle": 0, + "content": "HumanRig: Learning Automatic Rigging for Humanoid Character in a Large Scale Dataset" + }, + { + "type": "text", + "bbox": [ + 0.108, + 0.204, + 0.89, + 0.276 + ], + "angle": 0, + "content": "Zedong Chu, Feng Xiong, Meiduo Liu, Jinzhi Zhang, Mingqi Shao, Zhaoxu Sun, Di Wang, Mu Xu Amap, Alibaba Group {chuzedong.czd, xf250971, liumeiduo.l, wushou.zjz, mingqi.smq, szx430629, wd426042, xumu.xm} @alibaba-inc.com" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.309, + 0.327, + 0.327 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.343, + 0.486, + 0.692 + ], + "angle": 0, + "content": "With the rapid evolution of 3D generation algorithms, the cost of producing 3D humanoid character models has plummeted, yet the field is impeded by the lack of a comprehensive dataset for automatic rigging—a pivotal step in character animation. Addressing this gap, we present HumanRig, the first large-scale dataset specifically designed for 3D humanoid character rigging, encompassing 11,434 meticulously curated T-posed meshes adhered to a uniform skeleton topology. Capitalizing on this dataset, we introduce an innovative, data-driven automatic rigging framework, which overcomes the limitations of GNN-based methods in handling complex AI-generated meshes. Our approach integrates a Prior-Guided Skeleton Estimator (PGSE) module, which uses 2D skeleton joints to provide a preliminary 3D skeleton, and a Mesh-Skeleton Mutual Attention Network (MSMAN) that fuses skeleton features with 3D mesh features extracted by a U-shaped point transformer. This enables a coarse-to-fine 3D skeleton joint regression and a robust skinning estimation, surpassing previous methods in quality and versatility. This work not only remedies the dataset deficiency in rigging research but also propels the animation industry towards more efficient and automated character rigging pipelines." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.723, + 0.222, + 0.74 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.485, + 0.903 + ], + "angle": 0, + "content": "The animation industry is undergoing a revolutionary change with the integration of machine learning and data-driven techniques, particularly in the realms of 3D modeling and character animation. A critical phase in bringing 3D humanoid characters to life is rigging, which entails skeleton construction and skinning to facilitate realistic motion. Traditionally, rigging is a labor-intensive task generally requiring skilled artists to manually construct skeleton joints and paint skinning weights. As content creation pipelines accelerate and grow in complexity, the need for automation" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.322, + 0.909, + 0.58 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.592, + 0.909, + 0.718 + ], + "angle": 0, + "content": "Figure 1. The AI-generated mesh and the artist-created one show distinct face topology distributions. Manual meshes often have varying vertex densities across body parts and special topologies near joints for better deformation, while AI meshes tend to have chaotic topologies which lack semantic information. Previous rigging methods can only handle simple artist-created meshes, while HumanRig deals well with both AI and manual ones, especially for those with complex clothes or accessories and irregular body shapes." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.749, + 0.825, + 0.764 + ], + "angle": 0, + "content": "in rigging has become more pressing than ever." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Bolstered by extensive 3D datasets, the field of AI-driven 3D model generation [20, 23, 24, 27, 29] has experienced significant advancements. Yet, the advancement in automatic rigging remains constrained due to the lack of a comprehensive and standardized dataset, which is essential for training powerful and robust rigging models. Previous work, like [25], laid the cornerstone for earlier rigging methods but was limited by a small scale of 1,729 T-pose humanoid meshes with inconsistent skeleton topologies and" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "304" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "a lack of meaningful joint labels, hindering their use in animation. Studies such as [11], utilizing the SMPL dataset [13] and its clothed version from [2], were restricted to realistic human body shapes, missing the character diversity needed for varied applications. This has prompted the development of our HumanRig." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.184, + 0.482, + 0.333 + ], + "angle": 0, + "content": "We introduce HumanRig, standing out as the first large-scale dataset tailored for the task of rigging 3D humanoid character models, and offering 11,434 meticulously curated AI-generated humanoid meshes. All models are generated in T-pose and aligned with an industry-standard skeleton topology, which enables direct plug-and-play in standard animation engines. Unparalleled in scale and diversity, HumanRig features an extensive range of body proportions and character types, spanning from real persons to cartoon characters and humanoid animals." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.335, + 0.483, + 0.591 + ], + "angle": 0, + "content": "Utilizing the HumanRig dataset, we introduce an innovative automatic rigging framework that diverges from previous approaches [11, 12, 15-17, 25], which primarily rely on Graph Neural Networks (GNNs) [7, 21] to learn rigging directly from 3D mesh geometric features. Such methods struggle with complex 3D models, particularly those generated by AI. The core modules of our method are: a) The Prior-Guided Skeleton Estimator (PGSE), which initializes a coarse skeleton using 2D priors projected into 3D space, significantly reducing the complexity of the rigging task. b) A U-shaped Point Transformer [30] serving as our mesh encoder, disregarding edges to enhance rigging robustness on intricate meshes. c) The Mesh-Skeleton Mutual Attention Network (MSMAN) enriching mesh and skeleton features with mutual information within a high-level semantic space, enabling joint optimization of skeleton construction and skinning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.593, + 0.482, + 0.729 + ], + "angle": 0, + "content": "Our results showcase the method's superior performance, delivering high-quality automatic rigging solutions that eclipse current state-of-the-art techniques. This contribution not only addresses the scarcity of large-scale datasets for humanoid rigging but also presents an automatic rigging solution set to revolutionize the animation industry. By simplifying the rigging process, we pave the way for more efficient and automated character animation, unlocking new frontiers in creative expression and content creation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.745, + 0.233, + 0.76 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.771, + 0.283, + 0.787 + ], + "angle": 0, + "content": "2.1. 3D Rigging Datasets" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.482, + 0.901 + ], + "angle": 0, + "content": "The RigNetv1 dataset [25] has been instrumental in recent rigging research. It includes 2,703 rigged models, of which 1,729 are T-posed humanoid. Each character has between 1,000 and 5,000 vertices and an average of 25 skeleton joints. However, these models lack a standard skeleton topology, limiting the applicability of motion retargeting. SMPL [11] has gathered 3D data for real human body" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.258 + ], + "angle": 0, + "content": "types using the same skeleton, but it is restricted to hairless bodies without clothes, and all samples share similar topologies and head-to-body ratios. Additionally, meshes in current datasets are manually crafted, which markedly differ from AI-generated meshes (See Fig. 1). This discrepancy limits the generic learning of automatic rigging models for AI-generated humanoid characters. Our HumanRig dataset, consisting of 11,434 AI-generated T-pose meshes with varied head-to-body ratios aligned to a standard Mixamo skeleton, is crafted to boost data-driven methodologies for automatic humanoid rigging." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.266, + 0.695, + 0.282 + ], + "angle": 0, + "content": "2.2. Automatic Rigging" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.289, + 0.905, + 0.394 + ], + "angle": 0, + "content": "Automatic rigging consists of two tasks, skeleton construction and skinning. Skeleton construction is about setting up a hierarchical bone structure and the corresponding joint positions, according to the body anatomy and expected deformation. Skinning involves assigning influence to skeletons over vertices in 3D models, ensuring natural deformation during skeletal animation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.395, + 0.905, + 0.5 + ], + "angle": 0, + "content": "Traditional Method. Pinocchio [1] employs the medial axis to position the skeletons via optimization, but cannot work well when the mesh shape deviates from the skeleton template. Geometric-based skinning methods [1, 8, 10, 22] assume skeleton influence is proportional to proximity, which can cause inaccuracies in meshes with irregular body shapes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.501, + 0.905, + 0.757 + ], + "angle": 0, + "content": "Data-driven Method. RigNet [25] learns mesh offsets, attention, and bone connection probabilities, extracting joints via attention clustering and constructing skeletons with Minimum Spanning Tree. It can not support motion tasks well because the predicted skeleton has uncertain topology and lacks semantic labels. Based on 3D vertex features directly extracted by Graph Neural Network (GNN) backbone, [11, 15] learn template based skeletons and [11, 12, 15-17, 25] learn skinning weights. However, they all struggle with models that have diverse body shapes, intricate clothing and accessories, or complex face topologies. Unlike these approaches, our method effectively utilizes front view rendering to establish robust priors and employs a 2D-3D lifting technique to accurately determine skeleton positions with a coarse-to-fine approach, and overcome the skinning challenge by using a point transformer-based encoder to accurately capture vertex features in 3D space." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.769, + 0.796, + 0.785 + ], + "angle": 0, + "content": "3. The Basis - HumanRig Dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.904, + 0.824 + ], + "angle": 0, + "content": "As depicted in Fig. 2, the data acquisition pipeline for HumanRig is highly efficient and is detailed as follows." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.833, + 0.816, + 0.849 + ], + "angle": 0, + "content": "3.1. 3D Humanoid Models Preparation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Constructing a large-scale 3D humanoid model dataset presents significant challenges, primarily due to the labor-intensive nature of requiring extensive input from profes" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "305" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.093, + 0.848, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.306, + 0.254, + 0.692, + 0.269 + ], + "angle": 0, + "content": "Figure 2. Data Acquisition Pipeline for our HumanRig dataset." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.297, + 0.499, + 0.358 + ], + "angle": 0, + "content": "
Dataset#Uniform SkelClothingHairBind PoseSource
Mixamo [3]108✓✓✓T-poseartist-created
SMPL [13]N/A✓××T-poseartist-created
RigNetv1 [25]1729×✓✓T-poseartist-created
HumanRig (ours)11434✓✓✓T-poseAI-generated
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.369, + 0.484, + 0.397 + ], + "angle": 0, + "content": "Table 1. Humanoid Rigging Dataset Comparisons. For RigNetv1, only T-posed humanoid meshes picked out to compare." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.425, + 0.483, + 0.516 + ], + "angle": 0, + "content": "sional artists. However, the rapid advancement of AI-based 2D image generation technology has enabled the creation of diverse humanoid character images from textual descriptions. Furthermore, leveraging AI-driven 3D model generation techniques, it has become feasible to efficiently derive high-quality 3D models from a single image." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.516, + 0.483, + 0.652 + ], + "angle": 0, + "content": "T-pose Image Generation. 3D humanoid characters are commonly modeled in neutral poses like T-pose or A-pose, for enhanced mesh deformation and motion capture (mocap) data compatibility. Our work adopts the T-pose as the bind pose to streamline the rigging process for 3D humanoid models. We introduce a pose-controlled text-to-image framework for generating T-pose character images by integrating stylized text prompts with reference T-pose images. Our method includes the following steps:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.655, + 0.483, + 0.716 + ], + "angle": 0, + "content": "1. Curating 5,000 detailed text prompts for a range of characters adorned with various clothing accessories, by employing a Large Language Model (LLM) to enhance the prompt template." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.716, + 0.483, + 0.761 + ], + "angle": 0, + "content": "2. Generating 50 T-pose maps with diverse head-to-body ratios by utilizing DWPose [26] to extract 2D poses from a curated set of humanoid images." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.761, + 0.483, + 0.821 + ], + "angle": 0, + "content": "3. Synthesizing 20,000 T-posed humanoid images by randomly selecting a text prompt and a T-posed map, and inputting them into Stable Diffusion (SD) [18] and the corresponding Pose ControlNet [28]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.822, + 0.483, + 0.853 + ], + "angle": 0, + "content": "4. Filtering out low-quality images, ultimately collecting 17,268 high-quality humanoid T-pose images." + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.655, + 0.483, + 0.853 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "This framework guarantees that the generated characters capture the desired style and aesthetics while maintaining accurate body proportions and proper T-pose alignment. As" + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.293, + 0.907, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.448, + 0.904, + 0.463 + ], + "angle": 0, + "content": "Figure 3. Head-to-body Ratio Diversity Statistics of HumanRig." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.49, + 0.905, + 0.534 + ], + "angle": 0, + "content": "a result, we can produce a broad range of diverse and precisely posed T-posed humanoid character images, covering various head-to-body ratios." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.535, + 0.907, + 0.701 + ], + "angle": 0, + "content": "T-posed Mesh Generation. The swift progress in 3D generation has enabled the rapid production of large-scale and high-quality 3D meshes with unprecedented speed. We utilize an Image-to-3D pipeline for generating 3D meshes, which includes open-source methods like InstantMesh [24] and Unique3D [23], as well as commercial tools such as Hyperhuman Rodin [29] and Triposr-v2 [20]. By inputting meticulously prepared T-posed character images, we produce a large batch of T-posed humanoid meshes. We manually filter out low-quality ones, selecting 14,662 high-quality meshes from an initial 17,268." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.712, + 0.757, + 0.728 + ], + "angle": 0, + "content": "3.2. Humanoid Models Rigging" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.903 + ], + "angle": 0, + "content": "To animate a newly created humanoid character using mocap data, it should have a practical skeleton topology that is compatible with mocap data, rather than binding to joints without semantic labels, like [25]. To simplify the mocap-to-deformation process, we utilize the Mixamo skeleton topology, enabling our rigged character to be compatible with existing animation software and game engines. Naturally, we utilize the Mixamo semi-automated rigging tool [3] to speed up the annotation process. Poor-quality rigging data is filtered out, and artists manually refine the remaining models. Ultimately, we produce 11,434 high-quality rigged" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "306" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.126, + 0.872, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.369, + 0.908, + 0.426 + ], + "angle": 0, + "content": "Figure 4. Method Overview. Given a humanoid mesh, a coarse skeleton is predicted by an Prior-guided skeleton estimator (PGSE) and helps construct the skeleton-aware vertex features. They are fed into Skeleton Encoder and Mesh Encoder, respectively, then fused by a Mesh-Skeleton Mutual Attention Network to predict a refined skeleton and skinning weights with a joint learning strategy. Finally, the skeleton and skinning weights are combined to produce the animation-ready character." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.452, + 0.218, + 0.466 + ], + "angle": 0, + "content": "humanoid models." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.479, + 0.284, + 0.496 + ], + "angle": 0, + "content": "3.3. Data Postprocessing" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.503, + 0.484, + 0.685 + ], + "angle": 0, + "content": "We normalize the prepared rigged humanoid models within a bounding box of \\([-0.5, 0.5]\\) to ensure proper alignment. From the rigging results, we extract 3D skeleton joint positions of size \\([s, 3]\\) and the skinning weight matrix of size \\([m, s]\\), where \\(s\\) represents the number of skeleton joints, uniformly set to 22 in this work, and \\(m\\) denotes the variable number of vertices across different humanoid models. Additionally, we generate a front-view image using a fixed camera. Utilizing the projection matrix derived from the camera's intrinsic and extrinsic parameters, we project the 3D skeleton joints onto the front-view image, thereby obtaining the corresponding 2D skeleton joint positions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.696, + 0.245, + 0.712 + ], + "angle": 0, + "content": "3.4. Data Summary" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.902 + ], + "angle": 0, + "content": "As presented in Table 1, we conduct a comparative analysis of our dataset with existing rigged humanoid datasets. Existing datasets are limited in quantity, diversity or skeleton unity. Our dataset bridges these gaps by offering a large-scale collection of 11,434 AI-generated meshes with diverse head-to-body ratios, aligned with a consistent Mixamo skeleton, thereby enhancing both the quantity and diversity of available models for rigging research and applications. Fig. 3 illustrates models within our dataset, showcasing the diverse head-to-body ratios and their corresponding data proportions. Each annotated sample comprises a rigged T-pose humanoid mesh, 3D skeleton joint positions," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.452, + 0.907, + 0.498 + ], + "angle": 0, + "content": "a skinning weight matrix, a front-view image, and the associated camera parameters along with the 2D skeleton joint positions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.511, + 0.696, + 0.529 + ], + "angle": 0, + "content": "4. Automatic Rigging" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.537, + 0.907, + 0.793 + ], + "angle": 0, + "content": "Capitalizing on the expansive HumanRig dataset characterized by a uniform skeleton topology, we propose a data-driven automatic rigging framework. As depicted in Fig. 4, our approach begins by initializing a coarse skeleton using a Prior-Guided Skeleton Estimator (PGSE). The skeleton serves a dual purpose: It facilitates the learning of joint positions from coarse to fine, and it enhances vertex attributes to yield skeleton-aware vertex features, thereby simplifying the learning process for both the skeleton and skinning weights. Furthermore, we employ two encoders to extract skeleton and mesh features respectively: a MLP-based Skeleton Encoder for extracting skeleton features and a Point Transformer-based Mesh Encoder for extracting mesh features. These distinct features are then cross-integrated through a Mesh-Skeleton Mutual Attention Network (MS-MAN) to achieve deeper feature integration. We will introduce each of these modules individually." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.802, + 0.8, + 0.818 + ], + "angle": 0, + "content": "4.1. Prior-guided Skeleton Estimator" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.906, + 0.902 + ], + "angle": 0, + "content": "The Prior-Guided Skeleton Estimator (PGSE) utilizes 2D skeleton priors from the model's front view and projects them into approximate 3D positions using camera parameters. This approach significantly simplifies the task of learning skeleton positions directly from 3D mesh geom" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.945, + 0.511, + 0.956 + ], + "angle": 0, + "content": "307" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.092, + 0.482, + 0.168 + ], + "angle": 0, + "content": "etry by providing an initial estimation, which can then be refined using mesh data. Specifically, we fine-tune RTM-Pose [9] on the front view image to predict 2D template joints \\(P_{j2D} \\in \\mathbb{R}^{s \\times 2}\\) accurately. Then, each 2D joint \\(\\mathbf{J}_{2D}\\) is back-projected into 3D as a ray:" + }, + { + "type": "equation", + "bbox": [ + 0.184, + 0.177, + 0.39, + 0.195 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {X}} _ {r} (\\mu ; \\mathbf {J} _ {2 D}) = \\mathrm {P} _ {c} \\tilde {\\mathbf {J}} _ {2 D} + \\mu \\tilde {\\mathbf {X}} _ {c}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.205, + 0.483, + 0.326 + ], + "angle": 0, + "content": "where \\(\\mathrm{P}_c\\in \\mathbb{R}^{4\\times 3}\\) is the pseudo-inverse of the camera projection matrix, \\(\\mathbf{X}_c\\) is the 3D location of the camera center, and the symbol with superscript tilde denotes the corresponding homogeneous coordinate. Finally, we determine the coarse 3D skeleton \\(P_{skec}\\in \\mathbb{R}^{s\\times 3}\\) by calculating the intersections of each ray with the mesh surface and using the midpoints of the first and last intersections for each 3D coarse joint." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.336, + 0.342, + 0.35 + ], + "angle": 0, + "content": "4.2. Skeleton and Mesh Encoder" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.358, + 0.483, + 0.448 + ], + "angle": 0, + "content": "We utilize a Skeleton Encoder composed of a simple 3-layer Multi-Layer Perceptron (MLP) to process the coarse 3D skeleton points and extract skeleton features \\( f_{s} \\in \\mathbb{R}^{s \\times c} \\), where \\( c \\) denotes the number of channels. This encoder is efficient due to the skeleton's fixed topology and limited number of joints." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.45, + 0.484, + 0.66 + ], + "angle": 0, + "content": "For mesh feature extraction, we begin by extracting all vertices and constructing skeleton-aware vertex features \\( f_{v} \\in \\mathbb{R}^{m \\times (3 + s)} \\), where each of the \\( m \\) vertices has 3 positional coordinates and \\( s \\) euclidean distance features relative to \\( P_{\\text{skec}} \\). We then employ a U-shaped Point Transformer [30] as the Mesh Encoder to extract deep mesh features, with the output of the last up-transition layer serving as our encoder output \\( f_{m} \\in \\mathbb{R}^{m \\times c} \\). The Point Transformer-based Mesh Encoder effectively integrates local geometric correlation features with global geometric semantic features. Compared to commonly used GNN-based Mesh Encoders, it exhibits superior generalization for AI-generated meshes that often have irregular face topology and incorporate a large number of vertices." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.67, + 0.45, + 0.685 + ], + "angle": 0, + "content": "4.3. Mesh-Skeleton Mutual Attention Network" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.693, + 0.484, + 0.873 + ], + "angle": 0, + "content": "Mesh-skeleton Mutual Attention Network (MSMAN) is further proposed to integrate Skeleton features \\( f_{s} \\) and Mesh features \\( f_{m} \\) through mutual cross-attention mechanisms. The skeleton features provide mesh features with body part affiliation, which aids in the semantic understanding of mesh vertices and enhances the accuracy of skinning weight predictions. In a similar vein, mesh features bolster the local feature representation of the skeleton, which is beneficial for refining the precise positions of the skeleton. Taking one side as an example, we introduce a multi-head cross-attention mechanism to enhance mesh features, which is formalized as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.885, + 0.483, + 0.902 + ], + "angle": 0, + "content": "\\[\nQ = f _ {s}; \\quad K = f _ {m}; \\quad V = f _ {m} \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.102, + 0.905, + 0.136 + ], + "angle": 0, + "content": "\\[\nf _ {m \\rightarrow s} = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d}}\\right) V \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.139, + 0.879, + 0.154 + ], + "angle": 0, + "content": "where \\(f_{m\\to s}\\in \\mathbb{R}^{s\\times c}\\) is the enhanced skeleton features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.155, + 0.905, + 0.229 + ], + "angle": 0, + "content": "For the other side of the interaction, we adopt the same formulation but set \\( Q = f_{m} \\); \\( K = f_{s} \\); \\( V = f_{s} \\) to integrate the skeleton features into the mesh features. Finally, we get \\( f_{s\\rightarrow m}\\in \\mathbb{R}^{m\\times c} \\), which represents skeleton-aware mesh features." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.238, + 0.8, + 0.254 + ], + "angle": 0, + "content": "4.4. Rigging Head and Training Loss" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.261, + 0.905, + 0.321 + ], + "angle": 0, + "content": "We employ two separate MLP-based heads to independently predict skeleton positions \\( P_{ske} \\in \\mathbb{R}^{s \\times 3} \\) and skinning weights \\( P_{skin} \\in \\mathbb{R}^{m \\times s} \\). For skeleton joints prediction, we use Mean Squared Error (MSE) Loss:" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.327, + 0.905, + 0.366 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s k e l e t o n}} = \\frac {1}{s} \\sum_ {i = 1} ^ {s} \\left(P _ {\\text {s k e} _ {i}} - G _ {\\text {s k e} _ {i}}\\right) ^ {2} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.372, + 0.906, + 0.462 + ], + "angle": 0, + "content": "Skinning prediction can be conceptualized as optimizing the selection probabilities of various skeleton joints for each mesh vertex. Consequently, we employ the Kullback-Leibler divergence loss to minimize the distance between the predicted skin weight distribution \\( P_{\\text{skin}} \\) and the ground truth distribution \\( G_{\\text{skin}} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.467, + 0.905, + 0.508 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s k i n n i n g}} = \\frac {1}{m} \\sum_ {i = 1} ^ {m} \\sum_ {j = 1} ^ {s} P _ {\\text {s k i n} _ {i j}} \\left(\\log \\frac {P _ {\\text {s k i n} _ {i j}}}{G _ {\\text {s k i n} _ {i , j}}}\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.513, + 0.724, + 0.527 + ], + "angle": 0, + "content": "The total loss is constructed as:" + }, + { + "type": "equation", + "bbox": [ + 0.603, + 0.536, + 0.905, + 0.551 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {s k i n n i n g}} + \\mathcal {L} _ {\\text {s k e l e t o n}} \\tag {5}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.559, + 0.645, + 0.575 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.906, + 0.749 + ], + "angle": 0, + "content": "Implementation details. Our framework is trained in two stages. First, we use the mmpose [4] framework to train a 2D joint predictor based on RTMPose [9]. We then apply PGSE to extract coarse skeletons from input meshes, which are used in the second stage to train other components. Both stages are trained on our HumanRig dataset, with the same \\(80\\% -10\\% -10\\%\\) data split. The training is performed on 2 NVIDIA RTX A6000 GPUs using the AdamW [14] optimizer and a MultiStepScheduler. The training lasts for 300 epochs, with the learning rate halved after 50 epochs, starting from an initial rate of 1e-3 and a batch size of 16." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Evaluation metrics. For skeleton construction evaluation, we employ the metrics CD-J2J, CD-J2B and CD-B2B proposed by [25]. For skinning evaluation, we adopt the same metrics as [12, 16, 17, 25], including skinning precision and L1-norm. Interested readers are referred to the original papers for detailed definitions of these metrics. For deformation quality, we evaluate deformation error by calculating the average Euclidean distance between deformed vertex positions and ground truth ones, using 10 random poses with joint rotations within a range of \\(\\pm 10\\) degrees." + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "308" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.097, + 0.09, + 0.908, + 0.175 + ], + "angle": 0, + "content": "
Train SetRGNetv1-human(Test)HumanRig(Test)
CD-J2J↓CD-J2B↓CD-B2B↓Prec↑L1↓CD-J2J↓CD-J2B↓CD-B2B↓Prec↑L1↓
RGNetv1-human0.02740.02580.02080.74450.25730.01630.01550.01260.83640.1700
HumanRig-small0.01890.01730.01100.65020.33250.00540.00500.00280.83710.0789
HumanRig0.00710.00640.00330.74520.15640.00270.00250.00120.92710.0610
" + }, + { + "type": "table_caption", + "bbox": [ + 0.277, + 0.184, + 0.72, + 0.199 + ], + "angle": 0, + "content": "Table 2. Cross-dataset comparisons of RigNetv1-human and HumanRig." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.223, + 0.501, + 0.282 + ], + "angle": 0, + "content": "
MethodCD-J2J↓CD-J2B↓CD-B2B↓Prec↑L1↓
Ours(w/o PGSE)0.01100.00610.00310.83630.1630
Ours(w/o MSMAN)0.00800.00760.00390.92140.0630
Ours(complete)0.00270.00250.00120.92710.0610
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.293, + 0.484, + 0.32 + ], + "angle": 0, + "content": "Table 3. Architecture Study that shows the influence of PGSE and MSMAN." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.336, + 0.501, + 0.404 + ], + "angle": 0, + "content": "
Mesh EncoderCD-J2J↓CD-J2B↓CD-B2B↓Prec↑L1↓
GraphSAGE [6]0.00310.00300.00160.81530.0764
GraphTransformer [19]0.00300.00290.00160.81000.0760
GMEdgeNet [25]0.00280.00260.00140.82900.0751
Ours(Point Transformer)0.00270.00250.00120.92710.0610
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.414, + 0.484, + 0.443 + ], + "angle": 0, + "content": "Table 4. Mesh Encoder Study that shows our superior performance compared to GNNs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.469, + 0.325, + 0.485 + ], + "angle": 0, + "content": "5.1. HumanRig Dataset Study" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.493, + 0.483, + 0.794 + ], + "angle": 0, + "content": "To assess the effectiveness of our HumanRig dataset, we perform a comparative study with RigNetv1. RigNetv1, the largest rigging dataset, lacks a uniform skeleton and consistent initial poses. To ensure a fair comparison, we select 1,729 T-posed humanoid meshes, rigged them with the Mixamo skeleton via our annotation pipeline, and labeled them as RigNetv1-human. We train our HumanRig method on the training sets of RigNetv1-human and HumanRig respectively, and subsequently evaluated both models on both test sets. Additionally, to neutralize the effect of dataset size, we created HumanRigsmall, a random subset of HumanRig comprising 1,729 samples. Table 2 demonstrate valuable insights. First, a model trained on HumanRigsmall surpasses one trained on RigNetv1-human on the HumanRig test set, even outperforming the intra-dataset training of RigNetv1-human in skeleton construction task. Second, expanding the dataset size to the full HumanRig dataset significantly improves performance in both tasks. Lastly, the model trained on our dataset excels on both artist-created and AI-generated models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.902 + ], + "angle": 0, + "content": "In addition to the large scale, the superior performance is also attributed to the diverse body proportions (see Fig. 3). We underscore its importance by an experiment that categorizes the HumanRig training and test sets into subsets based on head-body ratios, with the subset containing five heads designated as HumanRig5. We then create a new, balanced subset, HumanRig \\( u \\), by evenly sampling from all subsets" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.224, + 0.71, + 0.373 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.224, + 0.904, + 0.373 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.385, + 0.907, + 0.401 + ], + "angle": 0, + "content": "Figure 5. Study on the Importance of Diversity in Body Shapes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.427, + 0.907, + 0.563 + ], + "angle": 0, + "content": "to match the size of HumanRig5. We train two HumanRig models separately on HumanRig5 and HumanRigu, and then evaluate them on test subsets ranging from two-heads to nine-heads. The experimental results in Fig. 5 indicate that for the model trained on HumanRig5, the performance declines as the head-body ratio deviates from the five heads in both tasks. When considering various head-to-body ratios, it exhibits a more balanced performance distribution and overall improvement." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.574, + 0.769, + 0.59 + ], + "angle": 0, + "content": "5.2. Architecture Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.906, + 0.719 + ], + "angle": 0, + "content": "PGSE Design. After eliminating PGSE, the network directly regresses joint positions from mesh features without the coarse skeleton's guidance. This also limits the mesh encoder's input to vertex coordinates. Table 3 demonstrates that PGSE significantly reduces the difficulty of skeleton construction (CD-J2J drops from 0.0110 to 0.0027). Additionally, skeleton-aware vertex features substantially improves skinning performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "MSMAN Design. The MSMAN is integral to our method as it enables the joint optimization of skeleton construction and skinning. To isolate the effects of MSMAN, we train an expert network that estimates skinning weights directly from the mesh encoder's output, and use the coarse skeleton as the final skeleton output. Table 3 illustrates that the coarse skeleton itself achieves commendable accuracy, but its refinement by MSMAN still results in a notable improvement. Furthermore, skinning performance is improved by the joint optimization. These findings underscore the significance of MSMAN's mutual attention mechanism in effectively integrating mesh features and skeletal repre" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "309" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.093, + 0.681, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.687, + 0.094, + 0.895, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.367, + 0.908, + 0.41 + ], + "angle": 0, + "content": "Figure 6. Skeleton construction comparisons. RigNet outputs irregular skeletons, which presents challenges to motion applications. NBS performs poorly in joint position prediction. Our method achieves the most plausible humanoid skeletons according to different human shapes from character (a) to (e). The skeleton topology we utilized is illustrated in character (f)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.436, + 0.395, + 0.451 + ], + "angle": 0, + "content": "sentations within a high-level semantic space." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.453, + 0.485, + 0.695 + ], + "angle": 0, + "content": "Mesh Encoder Design. In this work, we deviate from the traditional approach of utilizing GNNs for mesh feature extraction by introducing a U-shaped Point Transformer. To investigate the impact of this design choice, we replace our mesh encoder with GMEdgeNet proposed by RigNet, and substitute its EdgeConv [21] operator with other widely-used GNNs such as GraphSAGE [6] and GraphTransformer [19]. The results presented in Table 4 demonstrate that our mesh encoder achieves superior performance. This improvement is due to the more comprehensive and efficient aggregation of mesh vertex features based on 3D spatial distances, compared to aggregation along surface edges in GNNs. Additionally, edges in AI-generated 3D models are often irregular and lack semantic information, thus, a mesh encoder that ignores edge information exhibits improved generalization and robustness." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.708, + 0.231, + 0.725 + ], + "angle": 0, + "content": "5.3. Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.733, + 0.483, + 0.824 + ], + "angle": 0, + "content": "Due to the lack of effective datasets, there are few data-driven methods [11, 12, 15-17, 25] for studying humanoid neural rigging. Unfortunately, as far as we know, only RigNet [25] and NBS [11] have been open-sourced. We conduct the following experiments to demonstrate the superiority of our method." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Skeleton Construction. For skeleton prediction, we compare our method with RigNet [25] and NBS [11], as illustrated in Fig. 6. RigNet results in a lack of different key joints in all samples because it does not incorporate a humanoid skeleton prior, leading to uncertain joint placements" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.432, + 0.925, + 0.483 + ], + "angle": 0, + "content": "
Test setHeat Diffusion [1]GVB [5]RigNet [25]Ours
RigNetv1-human0.004050.003060.002520.00198
HumanRig0.002360.00224N/A0.00084
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.492, + 0.907, + 0.534 + ], + "angle": 0, + "content": "Table 5. Deformation Error Study. RigNet is not evaluated on HumanRig as it only handles 1k-5k vertex meshes. Our method achieves the best deformation quality on both test sets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.567, + 0.907, + 0.717 + ], + "angle": 0, + "content": "and bone connections without semantic context. Although NBS utilizes a fixed skeleton template of SMPL [13], it produces inaccurate joint positions, especially in the lower limbs. This inaccuracy stems from the challenge of directly regressing joint positions from mesh features. Moreover, as it exclusively trained on the SMPL dataset, it struggles to generalize well to diverse meshes that exhibit varying head-to-body ratios. In contrast, by leveraging a robust 2D prior, we predict joints based on a standard skeleton template, resulting in more plausible skeletons." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.901 + ], + "angle": 0, + "content": "**Skinning and Deformation.** We conducted a comparative analysis with Heat Diffusion [1], Geodesic Voxel Binding (GVB) [5], and RigNet [25]. These methods can predict skinning weights given an input mesh and its corresponding skeleton. Heat Diffusion and GVB are traditional geometry-based methods, whereas RigNet employs a data-driven approach by training an independent skinning network. For fair comparison, we input them the same skeletons predicted by our HumanRig method. Table 5 presents the quantitative assessment of deformation errors. RigNet can not handle complex meshes with over 5K vertices, while our approach outperforms on both artist-created and AI-" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.512, + 0.957 + ], + "angle": 0, + "content": "310" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.177, + 0.092, + 0.831, + 0.565 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.583, + 0.908, + 0.627 + ], + "angle": 0, + "content": "Figure 7. Skinning and Deformation Comparisons. Different colors are assigned to each vertex according to the most influential joint, determined by skinning weights. Previous methods struggle with meshes that have special body shapes or intricate accessories (red box), while our method achieves superior deformation quality. More intuitive results are visualized in the supplementary video." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.652, + 0.483, + 0.759 + ], + "angle": 0, + "content": "generated meshes. Additionally, Fig. 7 provides a qualitative visualization, which indicates that our method excels in maintaining more lifelike and fluid animations across varying head-to-body ratios, intricate clothes or accessories. This benefits from the strong point-transformer-based mesh encoder, which competently differentiates between distinct body parts to generate more precise skinning weights." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.77, + 0.219, + 0.785 + ], + "angle": 0, + "content": "6. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In this work, we have addressed the critical need for a comprehensive dataset and a robust framework for automatic rigging of 3D humanoid character models. We introduce HumanRig, a large-scale dataset of AI-generated T-pose humanoid models, all rigged with a consistent skeleton topology, significantly surpassing the previous datasets in terms of size, diversity, complexity and practical motion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.652, + 0.907, + 0.803 + ], + "angle": 0, + "content": "applications. Leveraging the HumanRig dataset, we propose a novel automatic rigging framework that integrates a Prior-Guided Skeleton Estimator (PGSE), a U-shaped Point Transformer-based Mesh Encoder and a Mesh-Skeleton Mutual Attention Network (MSMAN). Our approach excels in skeleton construction and skinning prediction, effectively managing challenges from irregular shapes, detailed clothing or accessories, and differing head-to-body ratios. Compared to GNN-based approaches, our method provides superior generalization to complex, AI-generated meshes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.803, + 0.909, + 0.88 + ], + "angle": 0, + "content": "Our method still have limitations. First, the skeleton template has not considered small body parts such as fingers, which requires generating finer meshes in datasets. Second, it is exciting if our method could be extended to the quadruped or any objects." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "311" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.482, + 0.156 + ], + "angle": 0, + "content": "[1] Ilya Baran and Jovan Popovic. Automatic rigging and animation of 3d characters. ACM Transactions on graphics (TOG), 26(3):72-es, 2007. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.159, + 0.483, + 0.213 + ], + "angle": 0, + "content": "[2] Bharat Lal Bhatnagar, Garvita Tiwari, Christian Theobalt, and Gerard Pons-Moll. Multi-garment net: Learning to dress 3d people from images. In IEEE International Conference on Computer Vision (ICCV). IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.216, + 0.482, + 0.243 + ], + "angle": 0, + "content": "[3] Sue Blackman and Sue Blackman. Rigging with mixamo. Unity for Absolute Beginners, pages 565-573, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.246, + 0.482, + 0.285 + ], + "angle": 0, + "content": "[4] MMPose Contributors. Openmmlab pose estimation toolbox and benchmark. https://github.com/open-mmlab/mmpose, 2020.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.288, + 0.482, + 0.342 + ], + "angle": 0, + "content": "[5] Olivier Dionne and Martin De Lasa. Geodesic binding for degenerate character geometry using sparse voxelization. IEEE transactions on visualization and computer graphics, 20(10):1367-1378, 2014. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.345, + 0.483, + 0.386 + ], + "angle": 0, + "content": "[6] Will Hamilton, Zhitao Ying, and Jure Leskovec. Inductive representation learning on large graphs. Advances in neural information processing systems, 30, 2017. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.388, + 0.482, + 0.442 + ], + "angle": 0, + "content": "[7] Rana Hanocka, Amir Hertz, Noa Fish, Raja Giryes, Shachar Fleishman, and Daniel Cohen-Or. Meshconn: a network with an edge. ACM Transactions on Graphics (ToG), 38(4):1-12, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.446, + 0.482, + 0.487 + ], + "angle": 0, + "content": "[8] Alec Jacobson, Ilya Baran, Jovan Popovic, and Olga Sorkine. Bounded biharmonic weights for real-time deformation. ACM Trans. Graph., 30(4):78, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.489, + 0.482, + 0.543 + ], + "angle": 0, + "content": "[9] Tao Jiang, Peng Lu, Li Zhang, Ningsheng Ma, Rui Han, Chengqi Lyu, Yining Li, and Kai Chen. Rtmpose: Realtime multi-person pose estimation based on mmpose. arXiv preprint arXiv:2303.07399, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.546, + 0.482, + 0.586 + ], + "angle": 0, + "content": "[10] Ladislav Kavan and Olga Sorkine. Elasticity-inspired deformers for character articulation. ACM Transactions on Graphics (TOG), 31(6):1-8, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.589, + 0.482, + 0.643 + ], + "angle": 0, + "content": "[11] Peizhuo Li, Kfir Aberman, Rana Hanocka, Libin Liu, Olga Sorkine-Hornung, and Baoquan Chen. Learning skeletal articulations with neural blend shapes. ACM Transactions on Graphics (TOG), 40(4):1-15, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.646, + 0.482, + 0.7 + ], + "angle": 0, + "content": "[12] Lijuan Liu, Youyi Zheng, Di Tang, Yi Yuan, Changjie Fan, and Kun Zhou. Neuroskinning: Automatic skin binding for production characters with deep graph networks. ACM Transactions on Graphics (ToG), 38(4):1-12, 2019. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.703, + 0.482, + 0.757 + ], + "angle": 0, + "content": "[13] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. Acm Transactions on Graphics, 34 (Article 248), 2015. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.76, + 0.482, + 0.787 + ], + "angle": 0, + "content": "[14] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[15] Jing Ma and Dongliang Zhang. Tarig: Adaptive template-aware neural rigging for humanoid characters. Computers & Graphics, 114:158-167, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[16] Albert Mosella-Montoro and Javier Ruiz-Hidalgo. Skinningnet: Two-stream graph convolutional neural network for skinning prediction of synthetic characters. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18593-18602, 2022. 5" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[17] Xiaoyu Pan, Jiancong Huang, Jiaming Mai, He Wang, Honglin Li, Tongkui Su, Wenjun Wang, and Xiaogang Jin. Heterskinnet: A heterogeneous network for skin weights prediction. In Proceedings of the ACM on Computer Graphics and Interactive Techniques. Association for Computing Machinery, 2021. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.179, + 0.907, + 0.247 + ], + "angle": 0, + "content": "[18] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.249, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[19] Yunsheng Shi, Zhengjie Huang, Shikun Feng, Hui Zhong, Wenjin Wang, and Yu Sun. Masked label prediction: Unified message passing model for semi-supervised classification. arXiv preprint arXiv:2009.03509, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.305, + 0.905, + 0.374 + ], + "angle": 0, + "content": "[20] Dmitry Tochilkin, David Pankratz, Zexiang Liu, Zixuan Huang, Adam Letts, Yangguang Li, Ding Liang, Christian Laforte, Varun Jampani, and Yan-Pei Cao. Triposr: Fast 3d object reconstruction from a single image. arXiv preprint arXiv:2403.02151, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.376, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[21] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (tog), 38(5):1-12, 2019. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.433, + 0.905, + 0.516 + ], + "angle": 0, + "content": "[22] Rich Wareham and Joan Lasenby. Bone glow: An improved method for the assignment of weights for mesh deformation. In Articulated Motion and Deformable Objects: 5th International Conference, AMDO 2008, Port d'Andratx, Mallorca, Spain, July 9-11, 2008. Proceedings 5, pages 63-71. Springer, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.518, + 0.905, + 0.572 + ], + "angle": 0, + "content": "[23] Kailu Wu, Fangfu Liu, Zhihan Cai, Runjie Yan, Hanyang Wang, Yating Hu, Yueqi Duan, and Kaisheng Ma. Unique3d: High-quality and efficient 3d mesh generation from a single image. arXiv preprint arXiv:2405.20343, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.575, + 0.905, + 0.642 + ], + "angle": 0, + "content": "[24] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191, 2024.1,3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.646, + 0.905, + 0.7 + ], + "angle": 0, + "content": "[25] Zhan Xu, Yang Zhou, Evangelos Kalogerakis, Chris Landreth, and Karan Singh. Rignet: neural rigging for articulated characters. ACM Transactions on Graphics (TOG), 39 (4):58-1, 2020. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.703, + 0.905, + 0.757 + ], + "angle": 0, + "content": "[26] Zhendong Yang, Ailing Zeng, Chun Yuan, and Yu Li. Effective whole-body pose estimation with two-stages distillation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4210-4220, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.76, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[27] Jinzhi Zhang, Feng Xiong, and Mu Xu. G3PT: unleash the power of autoregressive modeling in 3d generation via cross-scale querying transformer. CoRR, abs/2409.06322, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.803, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[28] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[29] Longwen Zhang, Ziyu Wang, Qixuan Zhang, Qiwei Qiu, Anqi Pang, Haoran Jiang, Wei Yang, Lan Xu, and Jingyi Yu. Clay: A controllable large-scale generative model for creat-" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "312" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.124, + 0.092, + 0.484, + 0.12 + ], + "angle": 0, + "content": "ing high-quality 3d assets. ACM Transactions on Graphics (TOG), 43(4):1-20, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.178 + ], + "angle": 0, + "content": "[30] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16259-16268, 2021. 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "313" + } + ] +] \ No newline at end of file diff --git a/2025/HumanRig_ Learning Automatic Rigging for Humanoid Character in a Large Scale Dataset/a5721584-3cc2-4564-a9c0-b5eefdb73fbc_origin.pdf b/2025/HumanRig_ Learning Automatic Rigging for Humanoid Character in a Large Scale Dataset/a5721584-3cc2-4564-a9c0-b5eefdb73fbc_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9784cd2b137e36ad03fefdcc3adcf4c4b82e10cd --- /dev/null +++ b/2025/HumanRig_ Learning Automatic Rigging for Humanoid Character in a Large Scale Dataset/a5721584-3cc2-4564-a9c0-b5eefdb73fbc_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bc15be9bab3549eb2c8c2cb624eb92f72c8db8697bc56a066c0316d1797b19fe +size 8112036 diff --git a/2025/HumanRig_ Learning Automatic Rigging for Humanoid Character in a Large Scale Dataset/full.md b/2025/HumanRig_ Learning Automatic Rigging for Humanoid Character in a Large Scale Dataset/full.md new file mode 100644 index 0000000000000000000000000000000000000000..faf2039850f107c671a2a3759466298e1d066c19 --- /dev/null +++ b/2025/HumanRig_ Learning Automatic Rigging for Humanoid Character in a Large Scale Dataset/full.md @@ -0,0 +1,261 @@ +# HumanRig: Learning Automatic Rigging for Humanoid Character in a Large Scale Dataset + +Zedong Chu, Feng Xiong, Meiduo Liu, Jinzhi Zhang, Mingqi Shao, Zhaoxu Sun, Di Wang, Mu Xu Amap, Alibaba Group {chuzedong.czd, xf250971, liumeiduo.l, wushou.zjz, mingqi.smq, szx430629, wd426042, xumu.xm} @alibaba-inc.com + +# Abstract + +With the rapid evolution of 3D generation algorithms, the cost of producing 3D humanoid character models has plummeted, yet the field is impeded by the lack of a comprehensive dataset for automatic rigging—a pivotal step in character animation. Addressing this gap, we present HumanRig, the first large-scale dataset specifically designed for 3D humanoid character rigging, encompassing 11,434 meticulously curated T-posed meshes adhered to a uniform skeleton topology. Capitalizing on this dataset, we introduce an innovative, data-driven automatic rigging framework, which overcomes the limitations of GNN-based methods in handling complex AI-generated meshes. Our approach integrates a Prior-Guided Skeleton Estimator (PGSE) module, which uses 2D skeleton joints to provide a preliminary 3D skeleton, and a Mesh-Skeleton Mutual Attention Network (MSMAN) that fuses skeleton features with 3D mesh features extracted by a U-shaped point transformer. This enables a coarse-to-fine 3D skeleton joint regression and a robust skinning estimation, surpassing previous methods in quality and versatility. This work not only remedies the dataset deficiency in rigging research but also propels the animation industry towards more efficient and automated character rigging pipelines. + +# 1. Introduction + +The animation industry is undergoing a revolutionary change with the integration of machine learning and data-driven techniques, particularly in the realms of 3D modeling and character animation. A critical phase in bringing 3D humanoid characters to life is rigging, which entails skeleton construction and skinning to facilitate realistic motion. Traditionally, rigging is a labor-intensive task generally requiring skilled artists to manually construct skeleton joints and paint skinning weights. As content creation pipelines accelerate and grow in complexity, the need for automation + +![](images/e9243b13090b10c51ecb452479dcc6331d7d79804555d15f7fdd04c261953793.jpg) +Figure 1. The AI-generated mesh and the artist-created one show distinct face topology distributions. Manual meshes often have varying vertex densities across body parts and special topologies near joints for better deformation, while AI meshes tend to have chaotic topologies which lack semantic information. Previous rigging methods can only handle simple artist-created meshes, while HumanRig deals well with both AI and manual ones, especially for those with complex clothes or accessories and irregular body shapes. + +in rigging has become more pressing than ever. + +Bolstered by extensive 3D datasets, the field of AI-driven 3D model generation [20, 23, 24, 27, 29] has experienced significant advancements. Yet, the advancement in automatic rigging remains constrained due to the lack of a comprehensive and standardized dataset, which is essential for training powerful and robust rigging models. Previous work, like [25], laid the cornerstone for earlier rigging methods but was limited by a small scale of 1,729 T-pose humanoid meshes with inconsistent skeleton topologies and + +a lack of meaningful joint labels, hindering their use in animation. Studies such as [11], utilizing the SMPL dataset [13] and its clothed version from [2], were restricted to realistic human body shapes, missing the character diversity needed for varied applications. This has prompted the development of our HumanRig. + +We introduce HumanRig, standing out as the first large-scale dataset tailored for the task of rigging 3D humanoid character models, and offering 11,434 meticulously curated AI-generated humanoid meshes. All models are generated in T-pose and aligned with an industry-standard skeleton topology, which enables direct plug-and-play in standard animation engines. Unparalleled in scale and diversity, HumanRig features an extensive range of body proportions and character types, spanning from real persons to cartoon characters and humanoid animals. + +Utilizing the HumanRig dataset, we introduce an innovative automatic rigging framework that diverges from previous approaches [11, 12, 15-17, 25], which primarily rely on Graph Neural Networks (GNNs) [7, 21] to learn rigging directly from 3D mesh geometric features. Such methods struggle with complex 3D models, particularly those generated by AI. The core modules of our method are: a) The Prior-Guided Skeleton Estimator (PGSE), which initializes a coarse skeleton using 2D priors projected into 3D space, significantly reducing the complexity of the rigging task. b) A U-shaped Point Transformer [30] serving as our mesh encoder, disregarding edges to enhance rigging robustness on intricate meshes. c) The Mesh-Skeleton Mutual Attention Network (MSMAN) enriching mesh and skeleton features with mutual information within a high-level semantic space, enabling joint optimization of skeleton construction and skinning. + +Our results showcase the method's superior performance, delivering high-quality automatic rigging solutions that eclipse current state-of-the-art techniques. This contribution not only addresses the scarcity of large-scale datasets for humanoid rigging but also presents an automatic rigging solution set to revolutionize the animation industry. By simplifying the rigging process, we pave the way for more efficient and automated character animation, unlocking new frontiers in creative expression and content creation. + +# 2. Related Work + +# 2.1. 3D Rigging Datasets + +The RigNetv1 dataset [25] has been instrumental in recent rigging research. It includes 2,703 rigged models, of which 1,729 are T-posed humanoid. Each character has between 1,000 and 5,000 vertices and an average of 25 skeleton joints. However, these models lack a standard skeleton topology, limiting the applicability of motion retargeting. SMPL [11] has gathered 3D data for real human body + +types using the same skeleton, but it is restricted to hairless bodies without clothes, and all samples share similar topologies and head-to-body ratios. Additionally, meshes in current datasets are manually crafted, which markedly differ from AI-generated meshes (See Fig. 1). This discrepancy limits the generic learning of automatic rigging models for AI-generated humanoid characters. Our HumanRig dataset, consisting of 11,434 AI-generated T-pose meshes with varied head-to-body ratios aligned to a standard Mixamo skeleton, is crafted to boost data-driven methodologies for automatic humanoid rigging. + +# 2.2. Automatic Rigging + +Automatic rigging consists of two tasks, skeleton construction and skinning. Skeleton construction is about setting up a hierarchical bone structure and the corresponding joint positions, according to the body anatomy and expected deformation. Skinning involves assigning influence to skeletons over vertices in 3D models, ensuring natural deformation during skeletal animation. + +Traditional Method. Pinocchio [1] employs the medial axis to position the skeletons via optimization, but cannot work well when the mesh shape deviates from the skeleton template. Geometric-based skinning methods [1, 8, 10, 22] assume skeleton influence is proportional to proximity, which can cause inaccuracies in meshes with irregular body shapes. + +Data-driven Method. RigNet [25] learns mesh offsets, attention, and bone connection probabilities, extracting joints via attention clustering and constructing skeletons with Minimum Spanning Tree. It can not support motion tasks well because the predicted skeleton has uncertain topology and lacks semantic labels. Based on 3D vertex features directly extracted by Graph Neural Network (GNN) backbone, [11, 15] learn template based skeletons and [11, 12, 15-17, 25] learn skinning weights. However, they all struggle with models that have diverse body shapes, intricate clothing and accessories, or complex face topologies. Unlike these approaches, our method effectively utilizes front view rendering to establish robust priors and employs a 2D-3D lifting technique to accurately determine skeleton positions with a coarse-to-fine approach, and overcome the skinning challenge by using a point transformer-based encoder to accurately capture vertex features in 3D space. + +# 3. The Basis - HumanRig Dataset + +As depicted in Fig. 2, the data acquisition pipeline for HumanRig is highly efficient and is detailed as follows. + +# 3.1. 3D Humanoid Models Preparation + +Constructing a large-scale 3D humanoid model dataset presents significant challenges, primarily due to the labor-intensive nature of requiring extensive input from profes + +![](images/f2581e7396d4c9cc3c6728c7be508d004f9b17b607b66658f688719c3702785d.jpg) +Figure 2. Data Acquisition Pipeline for our HumanRig dataset. + +
Dataset#Uniform SkelClothingHairBind PoseSource
Mixamo [3]108✓✓✓T-poseartist-created
SMPL [13]N/A✓××T-poseartist-created
RigNetv1 [25]1729×✓✓T-poseartist-created
HumanRig (ours)11434✓✓✓T-poseAI-generated
+ +Table 1. Humanoid Rigging Dataset Comparisons. For RigNetv1, only T-posed humanoid meshes picked out to compare. + +sional artists. However, the rapid advancement of AI-based 2D image generation technology has enabled the creation of diverse humanoid character images from textual descriptions. Furthermore, leveraging AI-driven 3D model generation techniques, it has become feasible to efficiently derive high-quality 3D models from a single image. + +T-pose Image Generation. 3D humanoid characters are commonly modeled in neutral poses like T-pose or A-pose, for enhanced mesh deformation and motion capture (mocap) data compatibility. Our work adopts the T-pose as the bind pose to streamline the rigging process for 3D humanoid models. We introduce a pose-controlled text-to-image framework for generating T-pose character images by integrating stylized text prompts with reference T-pose images. Our method includes the following steps: + +1. Curating 5,000 detailed text prompts for a range of characters adorned with various clothing accessories, by employing a Large Language Model (LLM) to enhance the prompt template. +2. Generating 50 T-pose maps with diverse head-to-body ratios by utilizing DWPose [26] to extract 2D poses from a curated set of humanoid images. +3. Synthesizing 20,000 T-posed humanoid images by randomly selecting a text prompt and a T-posed map, and inputting them into Stable Diffusion (SD) [18] and the corresponding Pose ControlNet [28]. +4. Filtering out low-quality images, ultimately collecting 17,268 high-quality humanoid T-pose images. + +This framework guarantees that the generated characters capture the desired style and aesthetics while maintaining accurate body proportions and proper T-pose alignment. As + +![](images/f07dddad3c951309fe89465832b20e6552360a55add1545ee3abf270fa17170a.jpg) +Figure 3. Head-to-body Ratio Diversity Statistics of HumanRig. + +a result, we can produce a broad range of diverse and precisely posed T-posed humanoid character images, covering various head-to-body ratios. + +T-posed Mesh Generation. The swift progress in 3D generation has enabled the rapid production of large-scale and high-quality 3D meshes with unprecedented speed. We utilize an Image-to-3D pipeline for generating 3D meshes, which includes open-source methods like InstantMesh [24] and Unique3D [23], as well as commercial tools such as Hyperhuman Rodin [29] and Triposr-v2 [20]. By inputting meticulously prepared T-posed character images, we produce a large batch of T-posed humanoid meshes. We manually filter out low-quality ones, selecting 14,662 high-quality meshes from an initial 17,268. + +# 3.2. Humanoid Models Rigging + +To animate a newly created humanoid character using mocap data, it should have a practical skeleton topology that is compatible with mocap data, rather than binding to joints without semantic labels, like [25]. To simplify the mocap-to-deformation process, we utilize the Mixamo skeleton topology, enabling our rigged character to be compatible with existing animation software and game engines. Naturally, we utilize the Mixamo semi-automated rigging tool [3] to speed up the annotation process. Poor-quality rigging data is filtered out, and artists manually refine the remaining models. Ultimately, we produce 11,434 high-quality rigged + +![](images/a56b1b4c34141ccb33ee030353204f2739e78a52ef6d853f8cc16615e4e93ccd.jpg) +Figure 4. Method Overview. Given a humanoid mesh, a coarse skeleton is predicted by an Prior-guided skeleton estimator (PGSE) and helps construct the skeleton-aware vertex features. They are fed into Skeleton Encoder and Mesh Encoder, respectively, then fused by a Mesh-Skeleton Mutual Attention Network to predict a refined skeleton and skinning weights with a joint learning strategy. Finally, the skeleton and skinning weights are combined to produce the animation-ready character. + +humanoid models. + +# 3.3. Data Postprocessing + +We normalize the prepared rigged humanoid models within a bounding box of $[-0.5, 0.5]$ to ensure proper alignment. From the rigging results, we extract 3D skeleton joint positions of size $[s, 3]$ and the skinning weight matrix of size $[m, s]$ , where $s$ represents the number of skeleton joints, uniformly set to 22 in this work, and $m$ denotes the variable number of vertices across different humanoid models. Additionally, we generate a front-view image using a fixed camera. Utilizing the projection matrix derived from the camera's intrinsic and extrinsic parameters, we project the 3D skeleton joints onto the front-view image, thereby obtaining the corresponding 2D skeleton joint positions. + +# 3.4. Data Summary + +As presented in Table 1, we conduct a comparative analysis of our dataset with existing rigged humanoid datasets. Existing datasets are limited in quantity, diversity or skeleton unity. Our dataset bridges these gaps by offering a large-scale collection of 11,434 AI-generated meshes with diverse head-to-body ratios, aligned with a consistent Mixamo skeleton, thereby enhancing both the quantity and diversity of available models for rigging research and applications. Fig. 3 illustrates models within our dataset, showcasing the diverse head-to-body ratios and their corresponding data proportions. Each annotated sample comprises a rigged T-pose humanoid mesh, 3D skeleton joint positions, + +a skinning weight matrix, a front-view image, and the associated camera parameters along with the 2D skeleton joint positions. + +# 4. Automatic Rigging + +Capitalizing on the expansive HumanRig dataset characterized by a uniform skeleton topology, we propose a data-driven automatic rigging framework. As depicted in Fig. 4, our approach begins by initializing a coarse skeleton using a Prior-Guided Skeleton Estimator (PGSE). The skeleton serves a dual purpose: It facilitates the learning of joint positions from coarse to fine, and it enhances vertex attributes to yield skeleton-aware vertex features, thereby simplifying the learning process for both the skeleton and skinning weights. Furthermore, we employ two encoders to extract skeleton and mesh features respectively: a MLP-based Skeleton Encoder for extracting skeleton features and a Point Transformer-based Mesh Encoder for extracting mesh features. These distinct features are then cross-integrated through a Mesh-Skeleton Mutual Attention Network (MS-MAN) to achieve deeper feature integration. We will introduce each of these modules individually. + +# 4.1. Prior-guided Skeleton Estimator + +The Prior-Guided Skeleton Estimator (PGSE) utilizes 2D skeleton priors from the model's front view and projects them into approximate 3D positions using camera parameters. This approach significantly simplifies the task of learning skeleton positions directly from 3D mesh geom + +etry by providing an initial estimation, which can then be refined using mesh data. Specifically, we fine-tune RTM-Pose [9] on the front view image to predict 2D template joints $P_{j2D} \in \mathbb{R}^{s \times 2}$ accurately. Then, each 2D joint $\mathbf{J}_{2D}$ is back-projected into 3D as a ray: + +$$ +\tilde {\mathbf {X}} _ {r} (\mu ; \mathbf {J} _ {2 D}) = \mathrm {P} _ {c} \tilde {\mathbf {J}} _ {2 D} + \mu \tilde {\mathbf {X}} _ {c} +$$ + +where $\mathrm{P}_c\in \mathbb{R}^{4\times 3}$ is the pseudo-inverse of the camera projection matrix, $\mathbf{X}_c$ is the 3D location of the camera center, and the symbol with superscript tilde denotes the corresponding homogeneous coordinate. Finally, we determine the coarse 3D skeleton $P_{skec}\in \mathbb{R}^{s\times 3}$ by calculating the intersections of each ray with the mesh surface and using the midpoints of the first and last intersections for each 3D coarse joint. + +# 4.2. Skeleton and Mesh Encoder + +We utilize a Skeleton Encoder composed of a simple 3-layer Multi-Layer Perceptron (MLP) to process the coarse 3D skeleton points and extract skeleton features $f_{s} \in \mathbb{R}^{s \times c}$ , where $c$ denotes the number of channels. This encoder is efficient due to the skeleton's fixed topology and limited number of joints. + +For mesh feature extraction, we begin by extracting all vertices and constructing skeleton-aware vertex features $f_{v} \in \mathbb{R}^{m \times (3 + s)}$ , where each of the $m$ vertices has 3 positional coordinates and $s$ euclidean distance features relative to $P_{\text{skec}}$ . We then employ a U-shaped Point Transformer [30] as the Mesh Encoder to extract deep mesh features, with the output of the last up-transition layer serving as our encoder output $f_{m} \in \mathbb{R}^{m \times c}$ . The Point Transformer-based Mesh Encoder effectively integrates local geometric correlation features with global geometric semantic features. Compared to commonly used GNN-based Mesh Encoders, it exhibits superior generalization for AI-generated meshes that often have irregular face topology and incorporate a large number of vertices. + +# 4.3. Mesh-Skeleton Mutual Attention Network + +Mesh-skeleton Mutual Attention Network (MSMAN) is further proposed to integrate Skeleton features $f_{s}$ and Mesh features $f_{m}$ through mutual cross-attention mechanisms. The skeleton features provide mesh features with body part affiliation, which aids in the semantic understanding of mesh vertices and enhances the accuracy of skinning weight predictions. In a similar vein, mesh features bolster the local feature representation of the skeleton, which is beneficial for refining the precise positions of the skeleton. Taking one side as an example, we introduce a multi-head cross-attention mechanism to enhance mesh features, which is formalized as follows: + +$$ +Q = f _ {s}; \quad K = f _ {m}; \quad V = f _ {m} \tag {1} +$$ + +$$ +f _ {m \rightarrow s} = \operatorname {s o f t m a x} \left(\frac {Q K ^ {T}}{\sqrt {d}}\right) V \tag {2} +$$ + +where $f_{m\to s}\in \mathbb{R}^{s\times c}$ is the enhanced skeleton features. + +For the other side of the interaction, we adopt the same formulation but set $Q = f_{m}$ ; $K = f_{s}$ ; $V = f_{s}$ to integrate the skeleton features into the mesh features. Finally, we get $f_{s\rightarrow m}\in \mathbb{R}^{m\times c}$ , which represents skeleton-aware mesh features. + +# 4.4. Rigging Head and Training Loss + +We employ two separate MLP-based heads to independently predict skeleton positions $P_{ske} \in \mathbb{R}^{s \times 3}$ and skinning weights $P_{skin} \in \mathbb{R}^{m \times s}$ . For skeleton joints prediction, we use Mean Squared Error (MSE) Loss: + +$$ +\mathcal {L} _ {\text {s k e l e t o n}} = \frac {1}{s} \sum_ {i = 1} ^ {s} \left(P _ {\text {s k e} _ {i}} - G _ {\text {s k e} _ {i}}\right) ^ {2} \tag {3} +$$ + +Skinning prediction can be conceptualized as optimizing the selection probabilities of various skeleton joints for each mesh vertex. Consequently, we employ the Kullback-Leibler divergence loss to minimize the distance between the predicted skin weight distribution $P_{\text{skin}}$ and the ground truth distribution $G_{\text{skin}}$ : + +$$ +\mathcal {L} _ {\text {s k i n n i n g}} = \frac {1}{m} \sum_ {i = 1} ^ {m} \sum_ {j = 1} ^ {s} P _ {\text {s k i n} _ {i j}} \left(\log \frac {P _ {\text {s k i n} _ {i j}}}{G _ {\text {s k i n} _ {i , j}}}\right) \tag {4} +$$ + +The total loss is constructed as: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {\text {s k i n n i n g}} + \mathcal {L} _ {\text {s k e l e t o n}} \tag {5} +$$ + +# 5. Experiments + +Implementation details. Our framework is trained in two stages. First, we use the mmpose [4] framework to train a 2D joint predictor based on RTMPose [9]. We then apply PGSE to extract coarse skeletons from input meshes, which are used in the second stage to train other components. Both stages are trained on our HumanRig dataset, with the same $80\% -10\% -10\%$ data split. The training is performed on 2 NVIDIA RTX A6000 GPUs using the AdamW [14] optimizer and a MultiStepScheduler. The training lasts for 300 epochs, with the learning rate halved after 50 epochs, starting from an initial rate of 1e-3 and a batch size of 16. + +Evaluation metrics. For skeleton construction evaluation, we employ the metrics CD-J2J, CD-J2B and CD-B2B proposed by [25]. For skinning evaluation, we adopt the same metrics as [12, 16, 17, 25], including skinning precision and L1-norm. Interested readers are referred to the original papers for detailed definitions of these metrics. For deformation quality, we evaluate deformation error by calculating the average Euclidean distance between deformed vertex positions and ground truth ones, using 10 random poses with joint rotations within a range of $\pm 10$ degrees. + +
Train SetRGNetv1-human(Test)HumanRig(Test)
CD-J2J↓CD-J2B↓CD-B2B↓Prec↑L1↓CD-J2J↓CD-J2B↓CD-B2B↓Prec↑L1↓
RGNetv1-human0.02740.02580.02080.74450.25730.01630.01550.01260.83640.1700
HumanRig-small0.01890.01730.01100.65020.33250.00540.00500.00280.83710.0789
HumanRig0.00710.00640.00330.74520.15640.00270.00250.00120.92710.0610
+ +Table 2. Cross-dataset comparisons of RigNetv1-human and HumanRig. + +
MethodCD-J2J↓CD-J2B↓CD-B2B↓Prec↑L1↓
Ours(w/o PGSE)0.01100.00610.00310.83630.1630
Ours(w/o MSMAN)0.00800.00760.00390.92140.0630
Ours(complete)0.00270.00250.00120.92710.0610
+ +Table 3. Architecture Study that shows the influence of PGSE and MSMAN. + +
Mesh EncoderCD-J2J↓CD-J2B↓CD-B2B↓Prec↑L1↓
GraphSAGE [6]0.00310.00300.00160.81530.0764
GraphTransformer [19]0.00300.00290.00160.81000.0760
GMEdgeNet [25]0.00280.00260.00140.82900.0751
Ours(Point Transformer)0.00270.00250.00120.92710.0610
+ +Table 4. Mesh Encoder Study that shows our superior performance compared to GNNs. + +# 5.1. HumanRig Dataset Study + +To assess the effectiveness of our HumanRig dataset, we perform a comparative study with RigNetv1. RigNetv1, the largest rigging dataset, lacks a uniform skeleton and consistent initial poses. To ensure a fair comparison, we select 1,729 T-posed humanoid meshes, rigged them with the Mixamo skeleton via our annotation pipeline, and labeled them as RigNetv1-human. We train our HumanRig method on the training sets of RigNetv1-human and HumanRig respectively, and subsequently evaluated both models on both test sets. Additionally, to neutralize the effect of dataset size, we created HumanRigsmall, a random subset of HumanRig comprising 1,729 samples. Table 2 demonstrate valuable insights. First, a model trained on HumanRigsmall surpasses one trained on RigNetv1-human on the HumanRig test set, even outperforming the intra-dataset training of RigNetv1-human in skeleton construction task. Second, expanding the dataset size to the full HumanRig dataset significantly improves performance in both tasks. Lastly, the model trained on our dataset excels on both artist-created and AI-generated models. + +In addition to the large scale, the superior performance is also attributed to the diverse body proportions (see Fig. 3). We underscore its importance by an experiment that categorizes the HumanRig training and test sets into subsets based on head-body ratios, with the subset containing five heads designated as HumanRig5. We then create a new, balanced subset, HumanRig $u$ , by evenly sampling from all subsets + +![](images/1ffdf905137edcd6f5da773f5271b3a5df5375204a72a79c4d25616937fb6f95.jpg) +Figure 5. Study on the Importance of Diversity in Body Shapes. + +![](images/73b3546570a4b6e893a3d49eeef57e3e6c8dc0c5819ff7fa40aad8c929680513.jpg) + +to match the size of HumanRig5. We train two HumanRig models separately on HumanRig5 and HumanRigu, and then evaluate them on test subsets ranging from two-heads to nine-heads. The experimental results in Fig. 5 indicate that for the model trained on HumanRig5, the performance declines as the head-body ratio deviates from the five heads in both tasks. When considering various head-to-body ratios, it exhibits a more balanced performance distribution and overall improvement. + +# 5.2. Architecture Ablation Study + +PGSE Design. After eliminating PGSE, the network directly regresses joint positions from mesh features without the coarse skeleton's guidance. This also limits the mesh encoder's input to vertex coordinates. Table 3 demonstrates that PGSE significantly reduces the difficulty of skeleton construction (CD-J2J drops from 0.0110 to 0.0027). Additionally, skeleton-aware vertex features substantially improves skinning performance. + +MSMAN Design. The MSMAN is integral to our method as it enables the joint optimization of skeleton construction and skinning. To isolate the effects of MSMAN, we train an expert network that estimates skinning weights directly from the mesh encoder's output, and use the coarse skeleton as the final skeleton output. Table 3 illustrates that the coarse skeleton itself achieves commendable accuracy, but its refinement by MSMAN still results in a notable improvement. Furthermore, skinning performance is improved by the joint optimization. These findings underscore the significance of MSMAN's mutual attention mechanism in effectively integrating mesh features and skeletal repre + +![](images/d7ce88cfd4f0ed8429030f1305a40bcfdb9d42306a401554d2a7443950c0fe86.jpg) +Figure 6. Skeleton construction comparisons. RigNet outputs irregular skeletons, which presents challenges to motion applications. NBS performs poorly in joint position prediction. Our method achieves the most plausible humanoid skeletons according to different human shapes from character (a) to (e). The skeleton topology we utilized is illustrated in character (f). + +![](images/014e5cd5e2607b404378e0c9bffff381439d0a6a77b651228700a187015255c9.jpg) + +sentations within a high-level semantic space. + +Mesh Encoder Design. In this work, we deviate from the traditional approach of utilizing GNNs for mesh feature extraction by introducing a U-shaped Point Transformer. To investigate the impact of this design choice, we replace our mesh encoder with GMEdgeNet proposed by RigNet, and substitute its EdgeConv [21] operator with other widely-used GNNs such as GraphSAGE [6] and GraphTransformer [19]. The results presented in Table 4 demonstrate that our mesh encoder achieves superior performance. This improvement is due to the more comprehensive and efficient aggregation of mesh vertex features based on 3D spatial distances, compared to aggregation along surface edges in GNNs. Additionally, edges in AI-generated 3D models are often irregular and lack semantic information, thus, a mesh encoder that ignores edge information exhibits improved generalization and robustness. + +# 5.3. Comparisons + +Due to the lack of effective datasets, there are few data-driven methods [11, 12, 15-17, 25] for studying humanoid neural rigging. Unfortunately, as far as we know, only RigNet [25] and NBS [11] have been open-sourced. We conduct the following experiments to demonstrate the superiority of our method. + +Skeleton Construction. For skeleton prediction, we compare our method with RigNet [25] and NBS [11], as illustrated in Fig. 6. RigNet results in a lack of different key joints in all samples because it does not incorporate a humanoid skeleton prior, leading to uncertain joint placements + +
Test setHeat Diffusion [1]GVB [5]RigNet [25]Ours
RigNetv1-human0.004050.003060.002520.00198
HumanRig0.002360.00224N/A0.00084
+ +Table 5. Deformation Error Study. RigNet is not evaluated on HumanRig as it only handles 1k-5k vertex meshes. Our method achieves the best deformation quality on both test sets. + +and bone connections without semantic context. Although NBS utilizes a fixed skeleton template of SMPL [13], it produces inaccurate joint positions, especially in the lower limbs. This inaccuracy stems from the challenge of directly regressing joint positions from mesh features. Moreover, as it exclusively trained on the SMPL dataset, it struggles to generalize well to diverse meshes that exhibit varying head-to-body ratios. In contrast, by leveraging a robust 2D prior, we predict joints based on a standard skeleton template, resulting in more plausible skeletons. + +**Skinning and Deformation.** We conducted a comparative analysis with Heat Diffusion [1], Geodesic Voxel Binding (GVB) [5], and RigNet [25]. These methods can predict skinning weights given an input mesh and its corresponding skeleton. Heat Diffusion and GVB are traditional geometry-based methods, whereas RigNet employs a data-driven approach by training an independent skinning network. For fair comparison, we input them the same skeletons predicted by our HumanRig method. Table 5 presents the quantitative assessment of deformation errors. RigNet can not handle complex meshes with over 5K vertices, while our approach outperforms on both artist-created and AI- + +![](images/b2a7100ac64b371e6b46aa6a3059c24f8cec891bf54e388a4467a1fd33bfc601.jpg) +Figure 7. Skinning and Deformation Comparisons. Different colors are assigned to each vertex according to the most influential joint, determined by skinning weights. Previous methods struggle with meshes that have special body shapes or intricate accessories (red box), while our method achieves superior deformation quality. More intuitive results are visualized in the supplementary video. + +generated meshes. Additionally, Fig. 7 provides a qualitative visualization, which indicates that our method excels in maintaining more lifelike and fluid animations across varying head-to-body ratios, intricate clothes or accessories. This benefits from the strong point-transformer-based mesh encoder, which competently differentiates between distinct body parts to generate more precise skinning weights. + +# 6. Conclusions + +In this work, we have addressed the critical need for a comprehensive dataset and a robust framework for automatic rigging of 3D humanoid character models. We introduce HumanRig, a large-scale dataset of AI-generated T-pose humanoid models, all rigged with a consistent skeleton topology, significantly surpassing the previous datasets in terms of size, diversity, complexity and practical motion + +applications. Leveraging the HumanRig dataset, we propose a novel automatic rigging framework that integrates a Prior-Guided Skeleton Estimator (PGSE), a U-shaped Point Transformer-based Mesh Encoder and a Mesh-Skeleton Mutual Attention Network (MSMAN). Our approach excels in skeleton construction and skinning prediction, effectively managing challenges from irregular shapes, detailed clothing or accessories, and differing head-to-body ratios. Compared to GNN-based approaches, our method provides superior generalization to complex, AI-generated meshes. + +Our method still have limitations. First, the skeleton template has not considered small body parts such as fingers, which requires generating finer meshes in datasets. Second, it is exciting if our method could be extended to the quadruped or any objects. + +# References + +[1] Ilya Baran and Jovan Popovic. Automatic rigging and animation of 3d characters. ACM Transactions on graphics (TOG), 26(3):72-es, 2007. 2, 7 +[2] Bharat Lal Bhatnagar, Garvita Tiwari, Christian Theobalt, and Gerard Pons-Moll. Multi-garment net: Learning to dress 3d people from images. In IEEE International Conference on Computer Vision (ICCV). IEEE, 2019. 2 +[3] Sue Blackman and Sue Blackman. Rigging with mixamo. Unity for Absolute Beginners, pages 565-573, 2014. 3 +[4] MMPose Contributors. Openmmlab pose estimation toolbox and benchmark. https://github.com/open-mmlab/mmpose, 2020.5 +[5] Olivier Dionne and Martin De Lasa. Geodesic binding for degenerate character geometry using sparse voxelization. IEEE transactions on visualization and computer graphics, 20(10):1367-1378, 2014. 7 +[6] Will Hamilton, Zhitao Ying, and Jure Leskovec. Inductive representation learning on large graphs. Advances in neural information processing systems, 30, 2017. 6, 7 +[7] Rana Hanocka, Amir Hertz, Noa Fish, Raja Giryes, Shachar Fleishman, and Daniel Cohen-Or. Meshconn: a network with an edge. ACM Transactions on Graphics (ToG), 38(4):1-12, 2019. 2 +[8] Alec Jacobson, Ilya Baran, Jovan Popovic, and Olga Sorkine. Bounded biharmonic weights for real-time deformation. ACM Trans. Graph., 30(4):78, 2011. 2 +[9] Tao Jiang, Peng Lu, Li Zhang, Ningsheng Ma, Rui Han, Chengqi Lyu, Yining Li, and Kai Chen. Rtmpose: Realtime multi-person pose estimation based on mmpose. arXiv preprint arXiv:2303.07399, 2023. 5 +[10] Ladislav Kavan and Olga Sorkine. Elasticity-inspired deformers for character articulation. ACM Transactions on Graphics (TOG), 31(6):1-8, 2012. 2 +[11] Peizhuo Li, Kfir Aberman, Rana Hanocka, Libin Liu, Olga Sorkine-Hornung, and Baoquan Chen. Learning skeletal articulations with neural blend shapes. ACM Transactions on Graphics (TOG), 40(4):1-15, 2021. 2, 7 +[12] Lijuan Liu, Youyi Zheng, Di Tang, Yi Yuan, Changjie Fan, and Kun Zhou. Neuroskinning: Automatic skin binding for production characters with deep graph networks. ACM Transactions on Graphics (ToG), 38(4):1-12, 2019. 2, 5, 7 +[13] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. Acm Transactions on Graphics, 34 (Article 248), 2015. 2, 3, 7 +[14] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5 +[15] Jing Ma and Dongliang Zhang. Tarig: Adaptive template-aware neural rigging for humanoid characters. Computers & Graphics, 114:158-167, 2023. 2, 7 +[16] Albert Mosella-Montoro and Javier Ruiz-Hidalgo. Skinningnet: Two-stream graph convolutional neural network for skinning prediction of synthetic characters. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18593-18602, 2022. 5 + +[17] Xiaoyu Pan, Jiancong Huang, Jiaming Mai, He Wang, Honglin Li, Tongkui Su, Wenjun Wang, and Xiaogang Jin. Heterskinnet: A heterogeneous network for skin weights prediction. In Proceedings of the ACM on Computer Graphics and Interactive Techniques. Association for Computing Machinery, 2021. 2, 5, 7 +[18] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3 +[19] Yunsheng Shi, Zhengjie Huang, Shikun Feng, Hui Zhong, Wenjin Wang, and Yu Sun. Masked label prediction: Unified message passing model for semi-supervised classification. arXiv preprint arXiv:2009.03509, 2020. 6, 7 +[20] Dmitry Tochilkin, David Pankratz, Zexiang Liu, Zixuan Huang, Adam Letts, Yangguang Li, Ding Liang, Christian Laforte, Varun Jampani, and Yan-Pei Cao. Triposr: Fast 3d object reconstruction from a single image. arXiv preprint arXiv:2403.02151, 2024. 1, 3 +[21] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (tog), 38(5):1-12, 2019. 2, 7 +[22] Rich Wareham and Joan Lasenby. Bone glow: An improved method for the assignment of weights for mesh deformation. In Articulated Motion and Deformable Objects: 5th International Conference, AMDO 2008, Port d'Andratx, Mallorca, Spain, July 9-11, 2008. Proceedings 5, pages 63-71. Springer, 2008. 2 +[23] Kailu Wu, Fangfu Liu, Zhihan Cai, Runjie Yan, Hanyang Wang, Yating Hu, Yueqi Duan, and Kaisheng Ma. Unique3d: High-quality and efficient 3d mesh generation from a single image. arXiv preprint arXiv:2405.20343, 2024. 1, 3 +[24] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191, 2024.1,3 +[25] Zhan Xu, Yang Zhou, Evangelos Kalogerakis, Chris Landreth, and Karan Singh. Rignet: neural rigging for articulated characters. ACM Transactions on Graphics (TOG), 39 (4):58-1, 2020. 1, 2, 3, 5, 6, 7 +[26] Zhendong Yang, Ailing Zeng, Chun Yuan, and Yu Li. Effective whole-body pose estimation with two-stages distillation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4210-4220, 2023. 3 +[27] Jinzhi Zhang, Feng Xiong, and Mu Xu. G3PT: unleash the power of autoregressive modeling in 3d generation via cross-scale querying transformer. CoRR, abs/2409.06322, 2024. 1 +[28] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 3 +[29] Longwen Zhang, Ziyu Wang, Qixuan Zhang, Qiwei Qiu, Anqi Pang, Haoran Jiang, Wei Yang, Lan Xu, and Jingyi Yu. Clay: A controllable large-scale generative model for creat- + +ing high-quality 3d assets. ACM Transactions on Graphics (TOG), 43(4):1-20, 2024. 1, 3 +[30] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16259-16268, 2021. 2, 5 \ No newline at end of file diff --git a/2025/HumanRig_ Learning Automatic Rigging for Humanoid Character in a Large Scale Dataset/images.zip b/2025/HumanRig_ Learning Automatic Rigging for Humanoid Character in a Large Scale Dataset/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..412b21eba2352eeee789153f54171ff5eea71053 --- /dev/null +++ b/2025/HumanRig_ Learning Automatic Rigging for Humanoid Character in a Large Scale Dataset/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:635e92e3b75cf303a08426234055726fd25fa1a64225bead0168ef43312704ba +size 457707 diff --git a/2025/HumanRig_ Learning Automatic Rigging for Humanoid Character in a Large Scale Dataset/layout.json b/2025/HumanRig_ Learning Automatic Rigging for Humanoid Character in a Large Scale Dataset/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e38ecba0fc457dc6d45ee3ef258c72502b4996e7 --- /dev/null +++ b/2025/HumanRig_ Learning Automatic Rigging for Humanoid Character in a Large Scale Dataset/layout.json @@ -0,0 +1,6119 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 66, + 102, + 545, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 102, + 545, + 138 + ], + "spans": [ + { + "bbox": [ + 66, + 102, + 545, + 138 + ], + "type": "text", + "content": "HumanRig: Learning Automatic Rigging for Humanoid Character in a Large Scale Dataset" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 161, + 544, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 161, + 544, + 218 + ], + "spans": [ + { + "bbox": [ + 66, + 161, + 544, + 218 + ], + "type": "text", + "content": "Zedong Chu, Feng Xiong, Meiduo Liu, Jinzhi Zhang, Mingqi Shao, Zhaoxu Sun, Di Wang, Mu Xu Amap, Alibaba Group {chuzedong.czd, xf250971, liumeiduo.l, wushou.zjz, mingqi.smq, szx430629, wd426042, xumu.xm} @alibaba-inc.com" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "spans": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 271, + 297, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 271, + 297, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 271, + 297, + 548 + ], + "type": "text", + "content": "With the rapid evolution of 3D generation algorithms, the cost of producing 3D humanoid character models has plummeted, yet the field is impeded by the lack of a comprehensive dataset for automatic rigging—a pivotal step in character animation. Addressing this gap, we present HumanRig, the first large-scale dataset specifically designed for 3D humanoid character rigging, encompassing 11,434 meticulously curated T-posed meshes adhered to a uniform skeleton topology. Capitalizing on this dataset, we introduce an innovative, data-driven automatic rigging framework, which overcomes the limitations of GNN-based methods in handling complex AI-generated meshes. Our approach integrates a Prior-Guided Skeleton Estimator (PGSE) module, which uses 2D skeleton joints to provide a preliminary 3D skeleton, and a Mesh-Skeleton Mutual Attention Network (MSMAN) that fuses skeleton features with 3D mesh features extracted by a U-shaped point transformer. This enables a coarse-to-fine 3D skeleton joint regression and a robust skinning estimation, surpassing previous methods in quality and versatility. This work not only remedies the dataset deficiency in rigging research but also propels the animation industry towards more efficient and automated character rigging pipelines." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 572, + 135, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 572, + 135, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 572, + 135, + 586 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "text", + "content": "The animation industry is undergoing a revolutionary change with the integration of machine learning and data-driven techniques, particularly in the realms of 3D modeling and character animation. A critical phase in bringing 3D humanoid characters to life is rigging, which entails skeleton construction and skinning to facilitate realistic motion. Traditionally, rigging is a labor-intensive task generally requiring skilled artists to manually construct skeleton joints and paint skinning weights. As content creation pipelines accelerate and grow in complexity, the need for automation" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 316, + 255, + 556, + 459 + ], + "blocks": [ + { + "bbox": [ + 316, + 255, + 556, + 459 + ], + "lines": [ + { + "bbox": [ + 316, + 255, + 556, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 255, + 556, + 459 + ], + "type": "image", + "image_path": "e9243b13090b10c51ecb452479dcc6331d7d79804555d15f7fdd04c261953793.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 468, + 556, + 568 + ], + "lines": [ + { + "bbox": [ + 313, + 468, + 556, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 556, + 568 + ], + "type": "text", + "content": "Figure 1. The AI-generated mesh and the artist-created one show distinct face topology distributions. Manual meshes often have varying vertex densities across body parts and special topologies near joints for better deformation, while AI meshes tend to have chaotic topologies which lack semantic information. Previous rigging methods can only handle simple artist-created meshes, while HumanRig deals well with both AI and manual ones, especially for those with complex clothes or accessories and irregular body shapes." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 593, + 504, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 593, + 504, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 593, + 504, + 605 + ], + "type": "text", + "content": "in rigging has become more pressing than ever." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "Bolstered by extensive 3D datasets, the field of AI-driven 3D model generation [20, 23, 24, 27, 29] has experienced significant advancements. Yet, the advancement in automatic rigging remains constrained due to the lack of a comprehensive and standardized dataset, which is essential for training powerful and robust rigging models. Previous work, like [25], laid the cornerstone for earlier rigging methods but was limited by a small scale of 1,729 T-pose humanoid meshes with inconsistent skeleton topologies and" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "304" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "a lack of meaningful joint labels, hindering their use in animation. Studies such as [11], utilizing the SMPL dataset [13] and its clothed version from [2], were restricted to realistic human body shapes, missing the character diversity needed for varied applications. This has prompted the development of our HumanRig." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 145, + 294, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 145, + 294, + 263 + ], + "spans": [ + { + "bbox": [ + 55, + 145, + 294, + 263 + ], + "type": "text", + "content": "We introduce HumanRig, standing out as the first large-scale dataset tailored for the task of rigging 3D humanoid character models, and offering 11,434 meticulously curated AI-generated humanoid meshes. All models are generated in T-pose and aligned with an industry-standard skeleton topology, which enables direct plug-and-play in standard animation engines. Unparalleled in scale and diversity, HumanRig features an extensive range of body proportions and character types, spanning from real persons to cartoon characters and humanoid animals." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 265, + 295, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 265, + 295, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 265, + 295, + 468 + ], + "type": "text", + "content": "Utilizing the HumanRig dataset, we introduce an innovative automatic rigging framework that diverges from previous approaches [11, 12, 15-17, 25], which primarily rely on Graph Neural Networks (GNNs) [7, 21] to learn rigging directly from 3D mesh geometric features. Such methods struggle with complex 3D models, particularly those generated by AI. The core modules of our method are: a) The Prior-Guided Skeleton Estimator (PGSE), which initializes a coarse skeleton using 2D priors projected into 3D space, significantly reducing the complexity of the rigging task. b) A U-shaped Point Transformer [30] serving as our mesh encoder, disregarding edges to enhance rigging robustness on intricate meshes. c) The Mesh-Skeleton Mutual Attention Network (MSMAN) enriching mesh and skeleton features with mutual information within a high-level semantic space, enabling joint optimization of skeleton construction and skinning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 469, + 294, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 469, + 294, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 469, + 294, + 577 + ], + "type": "text", + "content": "Our results showcase the method's superior performance, delivering high-quality automatic rigging solutions that eclipse current state-of-the-art techniques. This contribution not only addresses the scarcity of large-scale datasets for humanoid rigging but also presents an automatic rigging solution set to revolutionize the animation industry. By simplifying the rigging process, we pave the way for more efficient and automated character animation, unlocking new frontiers in creative expression and content creation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 590, + 142, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 142, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 142, + 601 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 610, + 173, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 610, + 173, + 623 + ], + "spans": [ + { + "bbox": [ + 55, + 610, + 173, + 623 + ], + "type": "text", + "content": "2.1. 3D Rigging Datasets" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 629, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 294, + 713 + ], + "type": "text", + "content": "The RigNetv1 dataset [25] has been instrumental in recent rigging research. It includes 2,703 rigged models, of which 1,729 are T-posed humanoid. Each character has between 1,000 and 5,000 vertices and an average of 25 skeleton joints. However, these models lack a standard skeleton topology, limiting the applicability of motion retargeting. SMPL [11] has gathered 3D data for real human body" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "content": "types using the same skeleton, but it is restricted to hairless bodies without clothes, and all samples share similar topologies and head-to-body ratios. Additionally, meshes in current datasets are manually crafted, which markedly differ from AI-generated meshes (See Fig. 1). This discrepancy limits the generic learning of automatic rigging models for AI-generated humanoid characters. Our HumanRig dataset, consisting of 11,434 AI-generated T-pose meshes with varied head-to-body ratios aligned to a standard Mixamo skeleton, is crafted to boost data-driven methodologies for automatic humanoid rigging." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 210, + 425, + 223 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 210, + 425, + 223 + ], + "spans": [ + { + "bbox": [ + 313, + 210, + 425, + 223 + ], + "type": "text", + "content": "2.2. Automatic Rigging" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 228, + 553, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 553, + 312 + ], + "type": "text", + "content": "Automatic rigging consists of two tasks, skeleton construction and skinning. Skeleton construction is about setting up a hierarchical bone structure and the corresponding joint positions, according to the body anatomy and expected deformation. Skinning involves assigning influence to skeletons over vertices in 3D models, ensuring natural deformation during skeletal animation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 312, + 553, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 553, + 396 + ], + "type": "text", + "content": "Traditional Method. Pinocchio [1] employs the medial axis to position the skeletons via optimization, but cannot work well when the mesh shape deviates from the skeleton template. Geometric-based skinning methods [1, 8, 10, 22] assume skeleton influence is proportional to proximity, which can cause inaccuracies in meshes with irregular body shapes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 396, + 553, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 553, + 599 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 553, + 599 + ], + "type": "text", + "content": "Data-driven Method. RigNet [25] learns mesh offsets, attention, and bone connection probabilities, extracting joints via attention clustering and constructing skeletons with Minimum Spanning Tree. It can not support motion tasks well because the predicted skeleton has uncertain topology and lacks semantic labels. Based on 3D vertex features directly extracted by Graph Neural Network (GNN) backbone, [11, 15] learn template based skeletons and [11, 12, 15-17, 25] learn skinning weights. However, they all struggle with models that have diverse body shapes, intricate clothing and accessories, or complex face topologies. Unlike these approaches, our method effectively utilizes front view rendering to establish robust priors and employs a 2D-3D lifting technique to accurately determine skeleton positions with a coarse-to-fine approach, and overcome the skinning challenge by using a point transformer-based encoder to accurately capture vertex features in 3D space." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 609, + 487, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 609, + 487, + 621 + ], + "spans": [ + { + "bbox": [ + 313, + 609, + 487, + 621 + ], + "type": "text", + "content": "3. The Basis - HumanRig Dataset" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 629, + 553, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 553, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 553, + 652 + ], + "type": "text", + "content": "As depicted in Fig. 2, the data acquisition pipeline for HumanRig is highly efficient and is detailed as follows." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 659, + 499, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 659, + 499, + 672 + ], + "spans": [ + { + "bbox": [ + 313, + 659, + 499, + 672 + ], + "type": "text", + "content": "3.1. 3D Humanoid Models Preparation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "Constructing a large-scale 3D humanoid model dataset presents significant challenges, primarily due to the labor-intensive nature of requiring extensive input from profes" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "305" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 73, + 518, + 190 + ], + "blocks": [ + { + "bbox": [ + 82, + 73, + 518, + 190 + ], + "lines": [ + { + "bbox": [ + 82, + 73, + 518, + 190 + ], + "spans": [ + { + "bbox": [ + 82, + 73, + 518, + 190 + ], + "type": "image", + "image_path": "f2581e7396d4c9cc3c6728c7be508d004f9b17b607b66658f688719c3702785d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 187, + 201, + 423, + 213 + ], + "lines": [ + { + "bbox": [ + 187, + 201, + 423, + 213 + ], + "spans": [ + { + "bbox": [ + 187, + 201, + 423, + 213 + ], + "type": "text", + "content": "Figure 2. Data Acquisition Pipeline for our HumanRig dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 235, + 305, + 283 + ], + "blocks": [ + { + "bbox": [ + 56, + 235, + 305, + 283 + ], + "lines": [ + { + "bbox": [ + 56, + 235, + 305, + 283 + ], + "spans": [ + { + "bbox": [ + 56, + 235, + 305, + 283 + ], + "type": "table", + "html": "
Dataset#Uniform SkelClothingHairBind PoseSource
Mixamo [3]108✓✓✓T-poseartist-created
SMPL [13]N/A✓××T-poseartist-created
RigNetv1 [25]1729×✓✓T-poseartist-created
HumanRig (ours)11434✓✓✓T-poseAI-generated
", + "image_path": "defa07d02366ce2461c9448c2ca561c511e7aa393bfd7203e25c36d4d50a2d10.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 292, + 296, + 314 + ], + "lines": [ + { + "bbox": [ + 55, + 292, + 296, + 314 + ], + "spans": [ + { + "bbox": [ + 55, + 292, + 296, + 314 + ], + "type": "text", + "content": "Table 1. Humanoid Rigging Dataset Comparisons. For RigNetv1, only T-posed humanoid meshes picked out to compare." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 336, + 295, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 336, + 295, + 408 + ], + "spans": [ + { + "bbox": [ + 54, + 336, + 295, + 408 + ], + "type": "text", + "content": "sional artists. However, the rapid advancement of AI-based 2D image generation technology has enabled the creation of diverse humanoid character images from textual descriptions. Furthermore, leveraging AI-driven 3D model generation techniques, it has become feasible to efficiently derive high-quality 3D models from a single image." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 408, + 295, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 408, + 295, + 516 + ], + "spans": [ + { + "bbox": [ + 54, + 408, + 295, + 516 + ], + "type": "text", + "content": "T-pose Image Generation. 3D humanoid characters are commonly modeled in neutral poses like T-pose or A-pose, for enhanced mesh deformation and motion capture (mocap) data compatibility. Our work adopts the T-pose as the bind pose to streamline the rigging process for 3D humanoid models. We introduce a pose-controlled text-to-image framework for generating T-pose character images by integrating stylized text prompts with reference T-pose images. Our method includes the following steps:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 518, + 295, + 675 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 55, + 518, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 518, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 518, + 295, + 567 + ], + "type": "text", + "content": "1. Curating 5,000 detailed text prompts for a range of characters adorned with various clothing accessories, by employing a Large Language Model (LLM) to enhance the prompt template." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 567, + 295, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 567, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 295, + 602 + ], + "type": "text", + "content": "2. Generating 50 T-pose maps with diverse head-to-body ratios by utilizing DWPose [26] to extract 2D poses from a curated set of humanoid images." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 602, + 295, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 602, + 295, + 650 + ], + "spans": [ + { + "bbox": [ + 55, + 602, + 295, + 650 + ], + "type": "text", + "content": "3. Synthesizing 20,000 T-posed humanoid images by randomly selecting a text prompt and a T-posed map, and inputting them into Stable Diffusion (SD) [18] and the corresponding Pose ControlNet [28]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 651, + 295, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 651, + 295, + 675 + ], + "spans": [ + { + "bbox": [ + 55, + 651, + 295, + 675 + ], + "type": "text", + "content": "4. Filtering out low-quality images, ultimately collecting 17,268 high-quality humanoid T-pose images." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 54, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 677, + 296, + 714 + ], + "type": "text", + "content": "This framework guarantees that the generated characters capture the desired style and aesthetics while maintaining accurate body proportions and proper T-pose alignment. As" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 314, + 232, + 555, + 338 + ], + "blocks": [ + { + "bbox": [ + 314, + 232, + 555, + 338 + ], + "lines": [ + { + "bbox": [ + 314, + 232, + 555, + 338 + ], + "spans": [ + { + "bbox": [ + 314, + 232, + 555, + 338 + ], + "type": "image", + "image_path": "f07dddad3c951309fe89465832b20e6552360a55add1545ee3abf270fa17170a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 354, + 553, + 366 + ], + "lines": [ + { + "bbox": [ + 313, + 354, + 553, + 366 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 553, + 366 + ], + "type": "text", + "content": "Figure 3. Head-to-body Ratio Diversity Statistics of HumanRig." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 388, + 553, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 553, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 553, + 422 + ], + "type": "text", + "content": "a result, we can produce a broad range of diverse and precisely posed T-posed humanoid character images, covering various head-to-body ratios." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 423, + 555, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 555, + 555 + ], + "type": "text", + "content": "T-posed Mesh Generation. The swift progress in 3D generation has enabled the rapid production of large-scale and high-quality 3D meshes with unprecedented speed. We utilize an Image-to-3D pipeline for generating 3D meshes, which includes open-source methods like InstantMesh [24] and Unique3D [23], as well as commercial tools such as Hyperhuman Rodin [29] and Triposr-v2 [20]. By inputting meticulously prepared T-posed character images, we produce a large batch of T-posed humanoid meshes. We manually filter out low-quality ones, selecting 14,662 high-quality meshes from an initial 17,268." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 563, + 463, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 463, + 576 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 463, + 576 + ], + "type": "text", + "content": "3.2. Humanoid Models Rigging" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "content": "To animate a newly created humanoid character using mocap data, it should have a practical skeleton topology that is compatible with mocap data, rather than binding to joints without semantic labels, like [25]. To simplify the mocap-to-deformation process, we utilize the Mixamo skeleton topology, enabling our rigged character to be compatible with existing animation software and game engines. Naturally, we utilize the Mixamo semi-automated rigging tool [3] to speed up the annotation process. Poor-quality rigging data is filtered out, and artists manually refine the remaining models. Ultimately, we produce 11,434 high-quality rigged" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "306" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 99, + 533, + 279 + ], + "blocks": [ + { + "bbox": [ + 72, + 99, + 533, + 279 + ], + "lines": [ + { + "bbox": [ + 72, + 99, + 533, + 279 + ], + "spans": [ + { + "bbox": [ + 72, + 99, + 533, + 279 + ], + "type": "image", + "image_path": "a56b1b4c34141ccb33ee030353204f2739e78a52ef6d853f8cc16615e4e93ccd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 292, + 555, + 337 + ], + "lines": [ + { + "bbox": [ + 54, + 292, + 555, + 337 + ], + "spans": [ + { + "bbox": [ + 54, + 292, + 555, + 337 + ], + "type": "text", + "content": "Figure 4. Method Overview. Given a humanoid mesh, a coarse skeleton is predicted by an Prior-guided skeleton estimator (PGSE) and helps construct the skeleton-aware vertex features. They are fed into Skeleton Encoder and Mesh Encoder, respectively, then fused by a Mesh-Skeleton Mutual Attention Network to predict a refined skeleton and skinning weights with a joint learning strategy. Finally, the skeleton and skinning weights are combined to produce the animation-ready character." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 357, + 133, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 357, + 133, + 369 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 133, + 369 + ], + "type": "text", + "content": "humanoid models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 379, + 173, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 379, + 173, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 379, + 173, + 392 + ], + "type": "text", + "content": "3.3. Data Postprocessing" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 398, + 296, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 398, + 296, + 542 + ], + "spans": [ + { + "bbox": [ + 55, + 398, + 296, + 542 + ], + "type": "text", + "content": "We normalize the prepared rigged humanoid models within a bounding box of " + }, + { + "bbox": [ + 55, + 398, + 296, + 542 + ], + "type": "inline_equation", + "content": "[-0.5, 0.5]" + }, + { + "bbox": [ + 55, + 398, + 296, + 542 + ], + "type": "text", + "content": " to ensure proper alignment. From the rigging results, we extract 3D skeleton joint positions of size " + }, + { + "bbox": [ + 55, + 398, + 296, + 542 + ], + "type": "inline_equation", + "content": "[s, 3]" + }, + { + "bbox": [ + 55, + 398, + 296, + 542 + ], + "type": "text", + "content": " and the skinning weight matrix of size " + }, + { + "bbox": [ + 55, + 398, + 296, + 542 + ], + "type": "inline_equation", + "content": "[m, s]" + }, + { + "bbox": [ + 55, + 398, + 296, + 542 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 398, + 296, + 542 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 398, + 296, + 542 + ], + "type": "text", + "content": " represents the number of skeleton joints, uniformly set to 22 in this work, and " + }, + { + "bbox": [ + 55, + 398, + 296, + 542 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 398, + 296, + 542 + ], + "type": "text", + "content": " denotes the variable number of vertices across different humanoid models. Additionally, we generate a front-view image using a fixed camera. Utilizing the projection matrix derived from the camera's intrinsic and extrinsic parameters, we project the 3D skeleton joints onto the front-view image, thereby obtaining the corresponding 2D skeleton joint positions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 551, + 149, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 551, + 149, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 551, + 149, + 563 + ], + "type": "text", + "content": "3.4. Data Summary" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": "As presented in Table 1, we conduct a comparative analysis of our dataset with existing rigged humanoid datasets. Existing datasets are limited in quantity, diversity or skeleton unity. Our dataset bridges these gaps by offering a large-scale collection of 11,434 AI-generated meshes with diverse head-to-body ratios, aligned with a consistent Mixamo skeleton, thereby enhancing both the quantity and diversity of available models for rigging research and applications. Fig. 3 illustrates models within our dataset, showcasing the diverse head-to-body ratios and their corresponding data proportions. Each annotated sample comprises a rigged T-pose humanoid mesh, 3D skeleton joint positions," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 357, + 555, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 555, + 394 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 555, + 394 + ], + "type": "text", + "content": "a skinning weight matrix, a front-view image, and the associated camera parameters along with the 2D skeleton joint positions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 404, + 425, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 404, + 425, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 404, + 425, + 418 + ], + "type": "text", + "content": "4. Automatic Rigging" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 425, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 425, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 425, + 555, + 628 + ], + "type": "text", + "content": "Capitalizing on the expansive HumanRig dataset characterized by a uniform skeleton topology, we propose a data-driven automatic rigging framework. As depicted in Fig. 4, our approach begins by initializing a coarse skeleton using a Prior-Guided Skeleton Estimator (PGSE). The skeleton serves a dual purpose: It facilitates the learning of joint positions from coarse to fine, and it enhances vertex attributes to yield skeleton-aware vertex features, thereby simplifying the learning process for both the skeleton and skinning weights. Furthermore, we employ two encoders to extract skeleton and mesh features respectively: a MLP-based Skeleton Encoder for extracting skeleton features and a Point Transformer-based Mesh Encoder for extracting mesh features. These distinct features are then cross-integrated through a Mesh-Skeleton Mutual Attention Network (MS-MAN) to achieve deeper feature integration. We will introduce each of these modules individually." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 635, + 489, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 489, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 489, + 647 + ], + "type": "text", + "content": "4.1. Prior-guided Skeleton Estimator" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 653, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 554, + 714 + ], + "type": "text", + "content": "The Prior-Guided Skeleton Estimator (PGSE) utilizes 2D skeleton priors from the model's front view and projects them into approximate 3D positions using camera parameters. This approach significantly simplifies the task of learning skeleton positions directly from 3D mesh geom" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 748, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 748, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 748, + 312, + 757 + ], + "type": "text", + "content": "307" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 294, + 133 + ], + "type": "text", + "content": "etry by providing an initial estimation, which can then be refined using mesh data. Specifically, we fine-tune RTM-Pose [9] on the front view image to predict 2D template joints " + }, + { + "bbox": [ + 54, + 72, + 294, + 133 + ], + "type": "inline_equation", + "content": "P_{j2D} \\in \\mathbb{R}^{s \\times 2}" + }, + { + "bbox": [ + 54, + 72, + 294, + 133 + ], + "type": "text", + "content": " accurately. Then, each 2D joint " + }, + { + "bbox": [ + 54, + 72, + 294, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{J}_{2D}" + }, + { + "bbox": [ + 54, + 72, + 294, + 133 + ], + "type": "text", + "content": " is back-projected into 3D as a ray:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 112, + 140, + 238, + 154 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 140, + 238, + 154 + ], + "spans": [ + { + "bbox": [ + 112, + 140, + 238, + 154 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {X}} _ {r} (\\mu ; \\mathbf {J} _ {2 D}) = \\mathrm {P} _ {c} \\tilde {\\mathbf {J}} _ {2 D} + \\mu \\tilde {\\mathbf {X}} _ {c}", + "image_path": "8d277a9a90148db8c1e283b447a241f0b96520da1507109a009408015f1b6de3.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 162, + 295, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 162, + 295, + 258 + ], + "spans": [ + { + "bbox": [ + 54, + 162, + 295, + 258 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 54, + 162, + 295, + 258 + ], + "type": "inline_equation", + "content": "\\mathrm{P}_c\\in \\mathbb{R}^{4\\times 3}" + }, + { + "bbox": [ + 54, + 162, + 295, + 258 + ], + "type": "text", + "content": " is the pseudo-inverse of the camera projection matrix, " + }, + { + "bbox": [ + 54, + 162, + 295, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_c" + }, + { + "bbox": [ + 54, + 162, + 295, + 258 + ], + "type": "text", + "content": " is the 3D location of the camera center, and the symbol with superscript tilde denotes the corresponding homogeneous coordinate. Finally, we determine the coarse 3D skeleton " + }, + { + "bbox": [ + 54, + 162, + 295, + 258 + ], + "type": "inline_equation", + "content": "P_{skec}\\in \\mathbb{R}^{s\\times 3}" + }, + { + "bbox": [ + 54, + 162, + 295, + 258 + ], + "type": "text", + "content": " by calculating the intersections of each ray with the mesh surface and using the midpoints of the first and last intersections for each 3D coarse joint." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 266, + 209, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 209, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 209, + 277 + ], + "type": "text", + "content": "4.2. Skeleton and Mesh Encoder" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 283, + 295, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 283, + 295, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 283, + 295, + 354 + ], + "type": "text", + "content": "We utilize a Skeleton Encoder composed of a simple 3-layer Multi-Layer Perceptron (MLP) to process the coarse 3D skeleton points and extract skeleton features " + }, + { + "bbox": [ + 55, + 283, + 295, + 354 + ], + "type": "inline_equation", + "content": "f_{s} \\in \\mathbb{R}^{s \\times c}" + }, + { + "bbox": [ + 55, + 283, + 295, + 354 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 283, + 295, + 354 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 283, + 295, + 354 + ], + "type": "text", + "content": " denotes the number of channels. This encoder is efficient due to the skeleton's fixed topology and limited number of joints." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 356, + 296, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 356, + 296, + 522 + ], + "spans": [ + { + "bbox": [ + 55, + 356, + 296, + 522 + ], + "type": "text", + "content": "For mesh feature extraction, we begin by extracting all vertices and constructing skeleton-aware vertex features " + }, + { + "bbox": [ + 55, + 356, + 296, + 522 + ], + "type": "inline_equation", + "content": "f_{v} \\in \\mathbb{R}^{m \\times (3 + s)}" + }, + { + "bbox": [ + 55, + 356, + 296, + 522 + ], + "type": "text", + "content": ", where each of the " + }, + { + "bbox": [ + 55, + 356, + 296, + 522 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 356, + 296, + 522 + ], + "type": "text", + "content": " vertices has 3 positional coordinates and " + }, + { + "bbox": [ + 55, + 356, + 296, + 522 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 356, + 296, + 522 + ], + "type": "text", + "content": " euclidean distance features relative to " + }, + { + "bbox": [ + 55, + 356, + 296, + 522 + ], + "type": "inline_equation", + "content": "P_{\\text{skec}}" + }, + { + "bbox": [ + 55, + 356, + 296, + 522 + ], + "type": "text", + "content": ". We then employ a U-shaped Point Transformer [30] as the Mesh Encoder to extract deep mesh features, with the output of the last up-transition layer serving as our encoder output " + }, + { + "bbox": [ + 55, + 356, + 296, + 522 + ], + "type": "inline_equation", + "content": "f_{m} \\in \\mathbb{R}^{m \\times c}" + }, + { + "bbox": [ + 55, + 356, + 296, + 522 + ], + "type": "text", + "content": ". The Point Transformer-based Mesh Encoder effectively integrates local geometric correlation features with global geometric semantic features. Compared to commonly used GNN-based Mesh Encoders, it exhibits superior generalization for AI-generated meshes that often have irregular face topology and incorporate a large number of vertices." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 530, + 275, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 530, + 275, + 542 + ], + "spans": [ + { + "bbox": [ + 55, + 530, + 275, + 542 + ], + "type": "text", + "content": "4.3. Mesh-Skeleton Mutual Attention Network" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 548, + 296, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 548, + 296, + 691 + ], + "spans": [ + { + "bbox": [ + 55, + 548, + 296, + 691 + ], + "type": "text", + "content": "Mesh-skeleton Mutual Attention Network (MSMAN) is further proposed to integrate Skeleton features " + }, + { + "bbox": [ + 55, + 548, + 296, + 691 + ], + "type": "inline_equation", + "content": "f_{s}" + }, + { + "bbox": [ + 55, + 548, + 296, + 691 + ], + "type": "text", + "content": " and Mesh features " + }, + { + "bbox": [ + 55, + 548, + 296, + 691 + ], + "type": "inline_equation", + "content": "f_{m}" + }, + { + "bbox": [ + 55, + 548, + 296, + 691 + ], + "type": "text", + "content": " through mutual cross-attention mechanisms. The skeleton features provide mesh features with body part affiliation, which aids in the semantic understanding of mesh vertices and enhances the accuracy of skinning weight predictions. In a similar vein, mesh features bolster the local feature representation of the skeleton, which is beneficial for refining the precise positions of the skeleton. Taking one side as an example, we introduce a multi-head cross-attention mechanism to enhance mesh features, which is formalized as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 110, + 700, + 295, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 700, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 110, + 700, + 295, + 714 + ], + "type": "interline_equation", + "content": "Q = f _ {s}; \\quad K = f _ {m}; \\quad V = f _ {m} \\tag {1}", + "image_path": "431237208ebef1a8edc9af90b63d569bd803d9c46f6091a144e0111071c0ba7e.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 372, + 80, + 553, + 107 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 80, + 553, + 107 + ], + "spans": [ + { + "bbox": [ + 372, + 80, + 553, + 107 + ], + "type": "interline_equation", + "content": "f _ {m \\rightarrow s} = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d}}\\right) V \\tag {2}", + "image_path": "4daa5bafe69414418b58538fd55bd2ae1b6859f899959adbf8ade95b478429b8.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 110, + 537, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 110, + 537, + 121 + ], + "spans": [ + { + "bbox": [ + 313, + 110, + 537, + 121 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 110, + 537, + 121 + ], + "type": "inline_equation", + "content": "f_{m\\to s}\\in \\mathbb{R}^{s\\times c}" + }, + { + "bbox": [ + 313, + 110, + 537, + 121 + ], + "type": "text", + "content": " is the enhanced skeleton features." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 122, + 553, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 122, + 553, + 181 + ], + "spans": [ + { + "bbox": [ + 313, + 122, + 553, + 181 + ], + "type": "text", + "content": "For the other side of the interaction, we adopt the same formulation but set " + }, + { + "bbox": [ + 313, + 122, + 553, + 181 + ], + "type": "inline_equation", + "content": "Q = f_{m}" + }, + { + "bbox": [ + 313, + 122, + 553, + 181 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 313, + 122, + 553, + 181 + ], + "type": "inline_equation", + "content": "K = f_{s}" + }, + { + "bbox": [ + 313, + 122, + 553, + 181 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 313, + 122, + 553, + 181 + ], + "type": "inline_equation", + "content": "V = f_{s}" + }, + { + "bbox": [ + 313, + 122, + 553, + 181 + ], + "type": "text", + "content": " to integrate the skeleton features into the mesh features. Finally, we get " + }, + { + "bbox": [ + 313, + 122, + 553, + 181 + ], + "type": "inline_equation", + "content": "f_{s\\rightarrow m}\\in \\mathbb{R}^{m\\times c}" + }, + { + "bbox": [ + 313, + 122, + 553, + 181 + ], + "type": "text", + "content": ", which represents skeleton-aware mesh features." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 188, + 489, + 201 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 188, + 489, + 201 + ], + "spans": [ + { + "bbox": [ + 313, + 188, + 489, + 201 + ], + "type": "text", + "content": "4.4. Rigging Head and Training Loss" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 206, + 553, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 206, + 553, + 254 + ], + "spans": [ + { + "bbox": [ + 313, + 206, + 553, + 254 + ], + "type": "text", + "content": "We employ two separate MLP-based heads to independently predict skeleton positions " + }, + { + "bbox": [ + 313, + 206, + 553, + 254 + ], + "type": "inline_equation", + "content": "P_{ske} \\in \\mathbb{R}^{s \\times 3}" + }, + { + "bbox": [ + 313, + 206, + 553, + 254 + ], + "type": "text", + "content": " and skinning weights " + }, + { + "bbox": [ + 313, + 206, + 553, + 254 + ], + "type": "inline_equation", + "content": "P_{skin} \\in \\mathbb{R}^{m \\times s}" + }, + { + "bbox": [ + 313, + 206, + 553, + 254 + ], + "type": "text", + "content": ". For skeleton joints prediction, we use Mean Squared Error (MSE) Loss:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 359, + 258, + 553, + 289 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 258, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 359, + 258, + 553, + 289 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s k e l e t o n}} = \\frac {1}{s} \\sum_ {i = 1} ^ {s} \\left(P _ {\\text {s k e} _ {i}} - G _ {\\text {s k e} _ {i}}\\right) ^ {2} \\tag {3}", + "image_path": "d681d30e71b73846d8e4d2d1db95cb95891eca6e62c65e431052d688c2480fea.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 294, + 554, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 294, + 554, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 294, + 554, + 365 + ], + "type": "text", + "content": "Skinning prediction can be conceptualized as optimizing the selection probabilities of various skeleton joints for each mesh vertex. Consequently, we employ the Kullback-Leibler divergence loss to minimize the distance between the predicted skin weight distribution " + }, + { + "bbox": [ + 313, + 294, + 554, + 365 + ], + "type": "inline_equation", + "content": "P_{\\text{skin}}" + }, + { + "bbox": [ + 313, + 294, + 554, + 365 + ], + "type": "text", + "content": " and the ground truth distribution " + }, + { + "bbox": [ + 313, + 294, + 554, + 365 + ], + "type": "inline_equation", + "content": "G_{\\text{skin}}" + }, + { + "bbox": [ + 313, + 294, + 554, + 365 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 330, + 369, + 553, + 402 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 369, + 553, + 402 + ], + "spans": [ + { + "bbox": [ + 330, + 369, + 553, + 402 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s k i n n i n g}} = \\frac {1}{m} \\sum_ {i = 1} ^ {m} \\sum_ {j = 1} ^ {s} P _ {\\text {s k i n} _ {i j}} \\left(\\log \\frac {P _ {\\text {s k i n} _ {i j}}}{G _ {\\text {s k i n} _ {i , j}}}\\right) \\tag {4}", + "image_path": "a1c527c7222f9ed87f1c1d1e993f2f2160c36b810b68292084a4902dfdbd9a28.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 406, + 443, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 406, + 443, + 417 + ], + "spans": [ + { + "bbox": [ + 314, + 406, + 443, + 417 + ], + "type": "text", + "content": "The total loss is constructed as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 369, + 424, + 553, + 436 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 424, + 553, + 436 + ], + "spans": [ + { + "bbox": [ + 369, + 424, + 553, + 436 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {s k i n n i n g}} + \\mathcal {L} _ {\\text {s k e l e t o n}} \\tag {5}", + "image_path": "66792b44ae350fffcab19fb316cd883bcb7e95c9a5d6cb9cc3fc0c894815e99a.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 442, + 394, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 442, + 394, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 442, + 394, + 455 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 462, + 554, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 554, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 554, + 593 + ], + "type": "text", + "content": "Implementation details. Our framework is trained in two stages. First, we use the mmpose [4] framework to train a 2D joint predictor based on RTMPose [9]. We then apply PGSE to extract coarse skeletons from input meshes, which are used in the second stage to train other components. Both stages are trained on our HumanRig dataset, with the same " + }, + { + "bbox": [ + 313, + 462, + 554, + 593 + ], + "type": "inline_equation", + "content": "80\\% -10\\% -10\\%" + }, + { + "bbox": [ + 313, + 462, + 554, + 593 + ], + "type": "text", + "content": " data split. The training is performed on 2 NVIDIA RTX A6000 GPUs using the AdamW [14] optimizer and a MultiStepScheduler. The training lasts for 300 epochs, with the learning rate halved after 50 epochs, starting from an initial rate of 1e-3 and a batch size of 16." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": "Evaluation metrics. For skeleton construction evaluation, we employ the metrics CD-J2J, CD-J2B and CD-B2B proposed by [25]. For skinning evaluation, we adopt the same metrics as [12, 16, 17, 25], including skinning precision and L1-norm. Interested readers are referred to the original papers for detailed definitions of these metrics. For deformation quality, we evaluate deformation error by calculating the average Euclidean distance between deformed vertex positions and ground truth ones, using 10 random poses with joint rotations within a range of " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\pm 10" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " degrees." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "308" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 71, + 555, + 138 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 555, + 138 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 555, + 138 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 555, + 138 + ], + "type": "table", + "html": "
Train SetRGNetv1-human(Test)HumanRig(Test)
CD-J2J↓CD-J2B↓CD-B2B↓Prec↑L1↓CD-J2J↓CD-J2B↓CD-B2B↓Prec↑L1↓
RGNetv1-human0.02740.02580.02080.74450.25730.01630.01550.01260.83640.1700
HumanRig-small0.01890.01730.01100.65020.33250.00540.00500.00280.83710.0789
HumanRig0.00710.00640.00330.74520.15640.00270.00250.00120.92710.0610
", + "image_path": "01e69de8fcabe582de7935a7fed4b08ed568a7a345d9fbbd2ecd49050e35d5e9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 176, + 306, + 223 + ], + "blocks": [ + { + "bbox": [ + 169, + 145, + 440, + 157 + ], + "lines": [ + { + "bbox": [ + 169, + 145, + 440, + 157 + ], + "spans": [ + { + "bbox": [ + 169, + 145, + 440, + 157 + ], + "type": "text", + "content": "Table 2. Cross-dataset comparisons of RigNetv1-human and HumanRig." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 176, + 306, + 223 + ], + "lines": [ + { + "bbox": [ + 58, + 176, + 306, + 223 + ], + "spans": [ + { + "bbox": [ + 58, + 176, + 306, + 223 + ], + "type": "table", + "html": "
MethodCD-J2J↓CD-J2B↓CD-B2B↓Prec↑L1↓
Ours(w/o PGSE)0.01100.00610.00310.83630.1630
Ours(w/o MSMAN)0.00800.00760.00390.92140.0630
Ours(complete)0.00270.00250.00120.92710.0610
", + "image_path": "dbc5e17b2d96f8681222f4cab277bf62e8fdf727477fd7a20288340a57c42889.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 58, + 266, + 306, + 319 + ], + "blocks": [ + { + "bbox": [ + 55, + 232, + 296, + 253 + ], + "lines": [ + { + "bbox": [ + 55, + 232, + 296, + 253 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 296, + 253 + ], + "type": "text", + "content": "Table 3. Architecture Study that shows the influence of PGSE and MSMAN." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 266, + 306, + 319 + ], + "lines": [ + { + "bbox": [ + 58, + 266, + 306, + 319 + ], + "spans": [ + { + "bbox": [ + 58, + 266, + 306, + 319 + ], + "type": "table", + "html": "
Mesh EncoderCD-J2J↓CD-J2B↓CD-B2B↓Prec↑L1↓
GraphSAGE [6]0.00310.00300.00160.81530.0764
GraphTransformer [19]0.00300.00290.00160.81000.0760
GMEdgeNet [25]0.00280.00260.00140.82900.0751
Ours(Point Transformer)0.00270.00250.00120.92710.0610
", + "image_path": "0762beff6a4a47ca66afb966996b3b6204165e20fea3bf2fdfa635c6e2ffbe92.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 327, + 296, + 350 + ], + "lines": [ + { + "bbox": [ + 55, + 327, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 296, + 350 + ], + "type": "text", + "content": "Table 4. Mesh Encoder Study that shows our superior performance compared to GNNs." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 371, + 198, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 371, + 198, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 371, + 198, + 384 + ], + "type": "text", + "content": "5.1. HumanRig Dataset Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 390, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 390, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 54, + 390, + 295, + 628 + ], + "type": "text", + "content": "To assess the effectiveness of our HumanRig dataset, we perform a comparative study with RigNetv1. RigNetv1, the largest rigging dataset, lacks a uniform skeleton and consistent initial poses. To ensure a fair comparison, we select 1,729 T-posed humanoid meshes, rigged them with the Mixamo skeleton via our annotation pipeline, and labeled them as RigNetv1-human. We train our HumanRig method on the training sets of RigNetv1-human and HumanRig respectively, and subsequently evaluated both models on both test sets. Additionally, to neutralize the effect of dataset size, we created HumanRigsmall, a random subset of HumanRig comprising 1,729 samples. Table 2 demonstrate valuable insights. First, a model trained on HumanRigsmall surpasses one trained on RigNetv1-human on the HumanRig test set, even outperforming the intra-dataset training of RigNetv1-human in skeleton construction task. Second, expanding the dataset size to the full HumanRig dataset significantly improves performance in both tasks. Lastly, the model trained on our dataset excels on both artist-created and AI-generated models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "content": "In addition to the large scale, the superior performance is also attributed to the diverse body proportions (see Fig. 3). We underscore its importance by an experiment that categorizes the HumanRig training and test sets into subsets based on head-body ratios, with the subset containing five heads designated as HumanRig5. We then create a new, balanced subset, HumanRig " + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "content": ", by evenly sampling from all subsets" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 177, + 434, + 295 + ], + "blocks": [ + { + "bbox": [ + 316, + 177, + 434, + 295 + ], + "lines": [ + { + "bbox": [ + 316, + 177, + 434, + 295 + ], + "spans": [ + { + "bbox": [ + 316, + 177, + 434, + 295 + ], + "type": "image", + "image_path": "1ffdf905137edcd6f5da773f5271b3a5df5375204a72a79c4d25616937fb6f95.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 304, + 555, + 317 + ], + "lines": [ + { + "bbox": [ + 313, + 304, + 555, + 317 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 555, + 317 + ], + "type": "text", + "content": "Figure 5. Study on the Importance of Diversity in Body Shapes." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 435, + 177, + 553, + 295 + ], + "blocks": [ + { + "bbox": [ + 435, + 177, + 553, + 295 + ], + "lines": [ + { + "bbox": [ + 435, + 177, + 553, + 295 + ], + "spans": [ + { + "bbox": [ + 435, + 177, + 553, + 295 + ], + "type": "image", + "image_path": "73b3546570a4b6e893a3d49eeef57e3e6c8dc0c5819ff7fa40aad8c929680513.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 338, + 555, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 555, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 555, + 445 + ], + "type": "text", + "content": "to match the size of HumanRig5. We train two HumanRig models separately on HumanRig5 and HumanRigu, and then evaluate them on test subsets ranging from two-heads to nine-heads. The experimental results in Fig. 5 indicate that for the model trained on HumanRig5, the performance declines as the head-body ratio deviates from the five heads in both tasks. When considering various head-to-body ratios, it exhibits a more balanced performance distribution and overall improvement." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 454, + 470, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 454, + 470, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 470, + 467 + ], + "type": "text", + "content": "5.2. Architecture Ablation Study" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 473, + 554, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 554, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 554, + 569 + ], + "type": "text", + "content": "PGSE Design. After eliminating PGSE, the network directly regresses joint positions from mesh features without the coarse skeleton's guidance. This also limits the mesh encoder's input to vertex coordinates. Table 3 demonstrates that PGSE significantly reduces the difficulty of skeleton construction (CD-J2J drops from 0.0110 to 0.0027). Additionally, skeleton-aware vertex features substantially improves skinning performance." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "MSMAN Design. The MSMAN is integral to our method as it enables the joint optimization of skeleton construction and skinning. To isolate the effects of MSMAN, we train an expert network that estimates skinning weights directly from the mesh encoder's output, and use the coarse skeleton as the final skeleton output. Table 3 illustrates that the coarse skeleton itself achieves commendable accuracy, but its refinement by MSMAN still results in a notable improvement. Furthermore, skinning performance is improved by the joint optimization. These findings underscore the significance of MSMAN's mutual attention mechanism in effectively integrating mesh features and skeletal repre" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "309" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 73, + 416, + 277 + ], + "blocks": [ + { + "bbox": [ + 61, + 73, + 416, + 277 + ], + "lines": [ + { + "bbox": [ + 61, + 73, + 416, + 277 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 416, + 277 + ], + "type": "image", + "image_path": "d7ce88cfd4f0ed8429030f1305a40bcfdb9d42306a401554d2a7443950c0fe86.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 290, + 555, + 324 + ], + "lines": [ + { + "bbox": [ + 54, + 290, + 555, + 324 + ], + "spans": [ + { + "bbox": [ + 54, + 290, + 555, + 324 + ], + "type": "text", + "content": "Figure 6. Skeleton construction comparisons. RigNet outputs irregular skeletons, which presents challenges to motion applications. NBS performs poorly in joint position prediction. Our method achieves the most plausible humanoid skeletons according to different human shapes from character (a) to (e). The skeleton topology we utilized is illustrated in character (f)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 420, + 74, + 547, + 277 + ], + "blocks": [ + { + "bbox": [ + 420, + 74, + 547, + 277 + ], + "lines": [ + { + "bbox": [ + 420, + 74, + 547, + 277 + ], + "spans": [ + { + "bbox": [ + 420, + 74, + 547, + 277 + ], + "type": "image", + "image_path": "014e5cd5e2607b404378e0c9bffff381439d0a6a77b651228700a187015255c9.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 345, + 241, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 345, + 241, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 345, + 241, + 357 + ], + "type": "text", + "content": "sentations within a high-level semantic space." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 358, + 296, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 358, + 296, + 550 + ], + "spans": [ + { + "bbox": [ + 56, + 358, + 296, + 550 + ], + "type": "text", + "content": "Mesh Encoder Design. In this work, we deviate from the traditional approach of utilizing GNNs for mesh feature extraction by introducing a U-shaped Point Transformer. To investigate the impact of this design choice, we replace our mesh encoder with GMEdgeNet proposed by RigNet, and substitute its EdgeConv [21] operator with other widely-used GNNs such as GraphSAGE [6] and GraphTransformer [19]. The results presented in Table 4 demonstrate that our mesh encoder achieves superior performance. This improvement is due to the more comprehensive and efficient aggregation of mesh vertex features based on 3D spatial distances, compared to aggregation along surface edges in GNNs. Additionally, edges in AI-generated 3D models are often irregular and lack semantic information, thus, a mesh encoder that ignores edge information exhibits improved generalization and robustness." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 560, + 141, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 560, + 141, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 560, + 141, + 574 + ], + "type": "text", + "content": "5.3. Comparisons" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 580, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 580, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 580, + 295, + 652 + ], + "type": "text", + "content": "Due to the lack of effective datasets, there are few data-driven methods [11, 12, 15-17, 25] for studying humanoid neural rigging. Unfortunately, as far as we know, only RigNet [25] and NBS [11] have been open-sourced. We conduct the following experiments to demonstrate the superiority of our method." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "Skeleton Construction. For skeleton prediction, we compare our method with RigNet [25] and NBS [11], as illustrated in Fig. 6. RigNet results in a lack of different key joints in all samples because it does not incorporate a humanoid skeleton prior, leading to uncertain joint placements" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 317, + 342, + 566, + 382 + ], + "blocks": [ + { + "bbox": [ + 317, + 342, + 566, + 382 + ], + "lines": [ + { + "bbox": [ + 317, + 342, + 566, + 382 + ], + "spans": [ + { + "bbox": [ + 317, + 342, + 566, + 382 + ], + "type": "table", + "html": "
Test setHeat Diffusion [1]GVB [5]RigNet [25]Ours
RigNetv1-human0.004050.003060.002520.00198
HumanRig0.002360.00224N/A0.00084
", + "image_path": "360820de8f86e6ad31cd999094d4dfb4623993d33890b1d3ba3bec40c0a2f946.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 389, + 555, + 422 + ], + "lines": [ + { + "bbox": [ + 313, + 389, + 555, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 555, + 422 + ], + "type": "text", + "content": "Table 5. Deformation Error Study. RigNet is not evaluated on HumanRig as it only handles 1k-5k vertex meshes. Our method achieves the best deformation quality on both test sets." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 449, + 555, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 555, + 567 + ], + "type": "text", + "content": "and bone connections without semantic context. Although NBS utilizes a fixed skeleton template of SMPL [13], it produces inaccurate joint positions, especially in the lower limbs. This inaccuracy stems from the challenge of directly regressing joint positions from mesh features. Moreover, as it exclusively trained on the SMPL dataset, it struggles to generalize well to diverse meshes that exhibit varying head-to-body ratios. In contrast, by leveraging a robust 2D prior, we predict joints based on a standard skeleton template, resulting in more plausible skeletons." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": "**Skinning and Deformation.** We conducted a comparative analysis with Heat Diffusion [1], Geodesic Voxel Binding (GVB) [5], and RigNet [25]. These methods can predict skinning weights given an input mesh and its corresponding skeleton. Heat Diffusion and GVB are traditional geometry-based methods, whereas RigNet employs a data-driven approach by training an independent skinning network. For fair comparison, we input them the same skeletons predicted by our HumanRig method. Table 5 presents the quantitative assessment of deformation errors. RigNet can not handle complex meshes with over 5K vertices, while our approach outperforms on both artist-created and AI-" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "type": "text", + "content": "310" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 72, + 508, + 447 + ], + "blocks": [ + { + "bbox": [ + 108, + 72, + 508, + 447 + ], + "lines": [ + { + "bbox": [ + 108, + 72, + 508, + 447 + ], + "spans": [ + { + "bbox": [ + 108, + 72, + 508, + 447 + ], + "type": "image", + "image_path": "b2a7100ac64b371e6b46aa6a3059c24f8cec891bf54e388a4467a1fd33bfc601.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 461, + 555, + 496 + ], + "lines": [ + { + "bbox": [ + 55, + 461, + 555, + 496 + ], + "spans": [ + { + "bbox": [ + 55, + 461, + 555, + 496 + ], + "type": "text", + "content": "Figure 7. Skinning and Deformation Comparisons. Different colors are assigned to each vertex according to the most influential joint, determined by skinning weights. Previous methods struggle with meshes that have special body shapes or intricate accessories (red box), while our method achieves superior deformation quality. More intuitive results are visualized in the supplementary video." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 516, + 295, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 516, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 516, + 295, + 601 + ], + "type": "text", + "content": "generated meshes. Additionally, Fig. 7 provides a qualitative visualization, which indicates that our method excels in maintaining more lifelike and fluid animations across varying head-to-body ratios, intricate clothes or accessories. This benefits from the strong point-transformer-based mesh encoder, which competently differentiates between distinct body parts to generate more precise skinning weights." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 609, + 134, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 134, + 621 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 134, + 621 + ], + "type": "text", + "content": "6. Conclusions" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": "In this work, we have addressed the critical need for a comprehensive dataset and a robust framework for automatic rigging of 3D humanoid character models. We introduce HumanRig, a large-scale dataset of AI-generated T-pose humanoid models, all rigged with a consistent skeleton topology, significantly surpassing the previous datasets in terms of size, diversity, complexity and practical motion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 516, + 555, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 516, + 555, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 516, + 555, + 635 + ], + "type": "text", + "content": "applications. Leveraging the HumanRig dataset, we propose a novel automatic rigging framework that integrates a Prior-Guided Skeleton Estimator (PGSE), a U-shaped Point Transformer-based Mesh Encoder and a Mesh-Skeleton Mutual Attention Network (MSMAN). Our approach excels in skeleton construction and skinning prediction, effectively managing challenges from irregular shapes, detailed clothing or accessories, and differing head-to-body ratios. Compared to GNN-based approaches, our method provides superior generalization to complex, AI-generated meshes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 635, + 556, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 556, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 556, + 696 + ], + "type": "text", + "content": "Our method still have limitations. First, the skeleton template has not considered small body parts such as fingers, which requires generating finer meshes in datasets. Second, it is exciting if our method could be extended to the quadruped or any objects." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "311" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 61, + 91, + 294, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 294, + 123 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 294, + 123 + ], + "type": "text", + "content": "[1] Ilya Baran and Jovan Popovic. Automatic rigging and animation of 3d characters. ACM Transactions on graphics (TOG), 26(3):72-es, 2007. 2, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 295, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 295, + 168 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 295, + 168 + ], + "type": "text", + "content": "[2] Bharat Lal Bhatnagar, Garvita Tiwari, Christian Theobalt, and Gerard Pons-Moll. Multi-garment net: Learning to dress 3d people from images. In IEEE International Conference on Computer Vision (ICCV). IEEE, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 171, + 294, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 171, + 294, + 192 + ], + "spans": [ + { + "bbox": [ + 62, + 171, + 294, + 192 + ], + "type": "text", + "content": "[3] Sue Blackman and Sue Blackman. Rigging with mixamo. Unity for Absolute Beginners, pages 565-573, 2014. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 194, + 294, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 194, + 294, + 225 + ], + "spans": [ + { + "bbox": [ + 62, + 194, + 294, + 225 + ], + "type": "text", + "content": "[4] MMPose Contributors. Openmmlab pose estimation toolbox and benchmark. https://github.com/open-mmlab/mmpose, 2020.5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 228, + 294, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 228, + 294, + 270 + ], + "spans": [ + { + "bbox": [ + 62, + 228, + 294, + 270 + ], + "type": "text", + "content": "[5] Olivier Dionne and Martin De Lasa. Geodesic binding for degenerate character geometry using sparse voxelization. IEEE transactions on visualization and computer graphics, 20(10):1367-1378, 2014. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 273, + 295, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 273, + 295, + 305 + ], + "spans": [ + { + "bbox": [ + 62, + 273, + 295, + 305 + ], + "type": "text", + "content": "[6] Will Hamilton, Zhitao Ying, and Jure Leskovec. Inductive representation learning on large graphs. Advances in neural information processing systems, 30, 2017. 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 307, + 294, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 307, + 294, + 350 + ], + "spans": [ + { + "bbox": [ + 62, + 307, + 294, + 350 + ], + "type": "text", + "content": "[7] Rana Hanocka, Amir Hertz, Noa Fish, Raja Giryes, Shachar Fleishman, and Daniel Cohen-Or. Meshconn: a network with an edge. ACM Transactions on Graphics (ToG), 38(4):1-12, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 353, + 294, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 353, + 294, + 385 + ], + "spans": [ + { + "bbox": [ + 62, + 353, + 294, + 385 + ], + "type": "text", + "content": "[8] Alec Jacobson, Ilya Baran, Jovan Popovic, and Olga Sorkine. Bounded biharmonic weights for real-time deformation. ACM Trans. Graph., 30(4):78, 2011. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 387, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 387, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 62, + 387, + 294, + 430 + ], + "type": "text", + "content": "[9] Tao Jiang, Peng Lu, Li Zhang, Ningsheng Ma, Rui Han, Chengqi Lyu, Yining Li, and Kai Chen. Rtmpose: Realtime multi-person pose estimation based on mmpose. arXiv preprint arXiv:2303.07399, 2023. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 432, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 432, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 57, + 432, + 294, + 464 + ], + "type": "text", + "content": "[10] Ladislav Kavan and Olga Sorkine. Elasticity-inspired deformers for character articulation. ACM Transactions on Graphics (TOG), 31(6):1-8, 2012. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 466, + 294, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 466, + 294, + 509 + ], + "spans": [ + { + "bbox": [ + 57, + 466, + 294, + 509 + ], + "type": "text", + "content": "[11] Peizhuo Li, Kfir Aberman, Rana Hanocka, Libin Liu, Olga Sorkine-Hornung, and Baoquan Chen. Learning skeletal articulations with neural blend shapes. ACM Transactions on Graphics (TOG), 40(4):1-15, 2021. 2, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 511, + 294, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 511, + 294, + 554 + ], + "spans": [ + { + "bbox": [ + 57, + 511, + 294, + 554 + ], + "type": "text", + "content": "[12] Lijuan Liu, Youyi Zheng, Di Tang, Yi Yuan, Changjie Fan, and Kun Zhou. Neuroskinning: Automatic skin binding for production characters with deep graph networks. ACM Transactions on Graphics (ToG), 38(4):1-12, 2019. 2, 5, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 556, + 294, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 556, + 294, + 599 + ], + "spans": [ + { + "bbox": [ + 57, + 556, + 294, + 599 + ], + "type": "text", + "content": "[13] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. Acm Transactions on Graphics, 34 (Article 248), 2015. 2, 3, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 601, + 294, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 601, + 294, + 623 + ], + "spans": [ + { + "bbox": [ + 57, + 601, + 294, + 623 + ], + "type": "text", + "content": "[14] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 624, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 294, + 656 + ], + "type": "text", + "content": "[15] Jing Ma and Dongliang Zhang. Tarig: Adaptive template-aware neural rigging for humanoid characters. Computers & Graphics, 114:158-167, 2023. 2, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 658, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 294, + 713 + ], + "type": "text", + "content": "[16] Albert Mosella-Montoro and Javier Ruiz-Hidalgo. Skinningnet: Two-stream graph convolutional neural network for skinning prediction of synthetic characters. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18593-18602, 2022. 5" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 139 + ], + "type": "text", + "content": "[17] Xiaoyu Pan, Jiancong Huang, Jiaming Mai, He Wang, Honglin Li, Tongkui Su, Wenjun Wang, and Xiaogang Jin. Heterskinnet: A heterogeneous network for skin weights prediction. In Proceedings of the ACM on Computer Graphics and Interactive Techniques. Association for Computing Machinery, 2021. 2, 5, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 141, + 555, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 141, + 555, + 195 + ], + "spans": [ + { + "bbox": [ + 317, + 141, + 555, + 195 + ], + "type": "text", + "content": "[18] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 197, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 197, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 317, + 197, + 553, + 239 + ], + "type": "text", + "content": "[19] Yunsheng Shi, Zhengjie Huang, Shikun Feng, Hui Zhong, Wenjin Wang, and Yu Sun. Masked label prediction: Unified message passing model for semi-supervised classification. arXiv preprint arXiv:2009.03509, 2020. 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 241, + 553, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 241, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 317, + 241, + 553, + 296 + ], + "type": "text", + "content": "[20] Dmitry Tochilkin, David Pankratz, Zexiang Liu, Zixuan Huang, Adam Letts, Yangguang Li, Ding Liang, Christian Laforte, Varun Jampani, and Yan-Pei Cao. Triposr: Fast 3d object reconstruction from a single image. arXiv preprint arXiv:2403.02151, 2024. 1, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 297, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 297, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 317, + 297, + 553, + 341 + ], + "type": "text", + "content": "[21] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (tog), 38(5):1-12, 2019. 2, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 342, + 553, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 342, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 317, + 342, + 553, + 408 + ], + "type": "text", + "content": "[22] Rich Wareham and Joan Lasenby. Bone glow: An improved method for the assignment of weights for mesh deformation. In Articulated Motion and Deformable Objects: 5th International Conference, AMDO 2008, Port d'Andratx, Mallorca, Spain, July 9-11, 2008. Proceedings 5, pages 63-71. Springer, 2008. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 410, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 410, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 317, + 410, + 553, + 453 + ], + "type": "text", + "content": "[23] Kailu Wu, Fangfu Liu, Zhihan Cai, Runjie Yan, Hanyang Wang, Yating Hu, Yueqi Duan, and Kaisheng Ma. Unique3d: High-quality and efficient 3d mesh generation from a single image. arXiv preprint arXiv:2405.20343, 2024. 1, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 455, + 553, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 455, + 553, + 508 + ], + "spans": [ + { + "bbox": [ + 317, + 455, + 553, + 508 + ], + "type": "text", + "content": "[24] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191, 2024.1,3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 511, + 553, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 511, + 553, + 554 + ], + "spans": [ + { + "bbox": [ + 317, + 511, + 553, + 554 + ], + "type": "text", + "content": "[25] Zhan Xu, Yang Zhou, Evangelos Kalogerakis, Chris Landreth, and Karan Singh. Rignet: neural rigging for articulated characters. ACM Transactions on Graphics (TOG), 39 (4):58-1, 2020. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 556, + 553, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 556, + 553, + 599 + ], + "spans": [ + { + "bbox": [ + 317, + 556, + 553, + 599 + ], + "type": "text", + "content": "[26] Zhendong Yang, Ailing Zeng, Chun Yuan, and Yu Li. Effective whole-body pose estimation with two-stages distillation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4210-4220, 2023. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 601, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 601, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 317, + 601, + 553, + 633 + ], + "type": "text", + "content": "[27] Jinzhi Zhang, Feng Xiong, and Mu Xu. G3PT: unleash the power of autoregressive modeling in 3d generation via cross-scale querying transformer. CoRR, abs/2409.06322, 2024. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 635, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 553, + 678 + ], + "type": "text", + "content": "[28] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "type": "text", + "content": "[29] Longwen Zhang, Ziyu Wang, Qixuan Zhang, Qiwei Qiu, Anqi Pang, Haoran Jiang, Wei Yang, Lan Xu, and Jingyi Yu. Clay: A controllable large-scale generative model for creat-" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "312" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 140 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 75, + 72, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 72, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 296, + 95 + ], + "type": "text", + "content": "ing high-quality 3d assets. ACM Transactions on Graphics (TOG), 43(4):1-20, 2024. 1, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "type": "text", + "content": "[30] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16259-16268, 2021. 2, 5" + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "313" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/HunyuanPortrait_ Implicit Condition Control for Enhanced Portrait Animation/a6cc1e88-c49b-4877-897d-6a6632aadecd_content_list.json b/2025/HunyuanPortrait_ Implicit Condition Control for Enhanced Portrait Animation/a6cc1e88-c49b-4877-897d-6a6632aadecd_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..04b07d2881ab8264423a38880510a200bef3faf9 --- /dev/null +++ b/2025/HunyuanPortrait_ Implicit Condition Control for Enhanced Portrait Animation/a6cc1e88-c49b-4877-897d-6a6632aadecd_content_list.json @@ -0,0 +1,1182 @@ +[ + { + "type": "text", + "text": "HunyuanPortrait: Implicit Condition Control for Enhanced Portrait Animation", + "text_level": 1, + "bbox": [ + 96, + 130, + 898, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zunnan Xu $^{1,2*}$ , Zhentao Yu $^{2}$ , Zixiang Zhou $^{2}$ , Jun Zhou $^{3*}$ , Xiaoyu Jin $^{1*}$ , Fa-ting Hong $^{4*}$ , Xiaozhong Ji $^{2}$ , Junwei Zhu $^{2}$ , Chengfei Cai $^{2}$ , Shiyu Tang $^{2}$ , Qin Lin $^{2}$ , Xiu Li $^{1\\dagger}$ , Qinglin Lu $^{2\\dagger}$ , Tsinghua University ${}^{2}$ Hunyuan, Tencent ${}^{3}$ Sun Yat-Sen Univeristy ${}^{4}$ HKUST", + "bbox": [ + 129, + 179, + 870, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We introduce HunyuanPortrait, a diffusion-based condition control method that employs implicit representations for highly controllable and lifelike portrait animation. Given a single portrait image as an appearance reference and video clips as driving templates, HunyuanPortrait can animate the character in the reference image by the facial expression and head pose of the driving videos. In our framework, we utilize pre-trained encoders to achieve the decoupling of portrait motion information and identity in videos. To do so, implicit representation is adopted to encode motion information and is employed as control signals in the animation phase. By leveraging the power of stable video diffusion as the main building block, we carefully design adapter layers to inject control signals into the denoising unet through attention mechanisms. These bring spatial richness of details and temporal consistency. HunyuanPortrait also exhibits strong generalization performance, which can effectively disentangle appearance and motion under different image styles. Our framework outperforms existing methods, demonstrating superior temporal consistency and controllability. Our project is available at HunyuanPortrait.", + "bbox": [ + 89, + 300, + 483, + 617 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 645, + 220, + 659 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we aim to animate a single static image into a lifelike video using control signals in implicit representations derived from driving videos. As shown in Figure 1, our primary focus is not only on the high controllability of digital portraits but also on the strong consistency of the characters and background in the reference image.", + "bbox": [ + 89, + 670, + 482, + 760 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Portrait animation has been widely applied in fields such as human-computer interaction [11, 56], virtual reality, and the crafting of digital avatars for the gaming and multimedia industries [28-30]. Previous works [4, 7, 62] develop usable portrait animation models utilizing Generative Adversarial Networks (GANs) [10] based on raw video inputs. However, these approaches exhibit limitations in terms of gener", + "bbox": [ + 89, + 761, + 482, + 867 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a1ad754f12f89d77499d7c3e7dddd63a2b74498d5679a4605742fef0238e910a.jpg", + "image_caption": [ + "Figure 1. Our framework employs implicit condition control to generate portrait animations, demonstrating robust generalization performance with high-fidelity facial dynamics and vivid head poses during cross-reenactment. The animated portraits remain unaffected by variations in facial shape and the spatial position of the driving videos, demonstrating strong identity consistency." + ], + "image_footnote": [], + "bbox": [ + 516, + 267, + 903, + 612 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "alization. Given the variability of facial structures and the complexities of facial movements, these methods often face challenges such as artifacts and insufficient control over facial expressions. This is particularly evident when there is a substantial difference in facial shape between the driving video and the source image, resulting in significant issues with facial artifacts and motion distortion. These early methods often produce results with background jitter and blurred portraits during cross-reenactment [36, 41].", + "bbox": [ + 511, + 733, + 906, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The advancement of diffusion models has led to a significant focus in recent methodologies [3, 48] on the fine-tuning", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work done during the internship at Tencent.", + "bbox": [ + 107, + 875, + 346, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15909", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "of image-level diffusion models [39] through the incorporation of motion modules [14], thereby facilitating the generation of video content. These methods effectively address the issues associated with GAN-based approaches in managing complex backgrounds and enhancing generalization. However, even with the incorporation of motion modules, these methods are susceptible to defects during the generation stage (e.g., temporal smoothness and flexibility to different frame rates). This vulnerability arises from the absence of motion pre-training in the underlying models during the pre-training phase. Additionally, these methods remain constrained to the explicit control of facial expressions. The effectiveness of these methods, which depend on explicit keypoints to control facial expressions, not only results in the loss of intricate details but also significantly relies on the accuracy of the landmark extraction techniques employed. Furthermore, as a result of the variability in facial shapes, these methods are limited by post-processing alignment strategies and often face challenges in adapting to diverse facial geometries, which can lead to inaccurate facial control and insufficient preservation of identity in the visual effects of videos. Due to the human eye's sensitivity to subtle facial changes [45], it is crucial to preserve the fidelity of facial details and the smoothness of the videos.", + "bbox": [ + 93, + 90, + 480, + 452 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Therefore, we propose an implicit conditional control model to tackle the challenges posed by diverse facial geometries and intricate expression details. We utilize stable video diffusion as the foundational model and integrate identity and motion information through appearance and motion attention. This approach eliminates the need to fine-tune the image diffusion model and to train the motion module separately. Firstly, we coarsely decouple the identity information from the motion information in the original video by using a pre-trained motion encoder [9]. Considering that this pre-trained encoder has not completely decoupled identity from motion information, we further implement enhanced training strategies and improved network architectures to strengthen the control of motion features and the separation of portrait identities. For the adaptation of temporal modeling tasks in video generation using diffusion models, along with the complementary contextual information present between different frames, we propose the incorporation of a motion memory bank to enhance the implicit representation of motion features. Due to the substantial variations in the degree of blur and pixel distortion in video frames at different levels of motion intensity, we propose an intensity-aware motion encoder. This encoder aims to improve the capacity of decoupled motion features to capture intricate details. Additionally, in order to achieve consistent modeling of portrait identity and background areas in videos, we combined ArcFace [6] with the DiNOv2 [37] backbone to develop an enhanced appearance encoder. To preserve the pre-training knowledge of both models, we", + "bbox": [ + 93, + 463, + 480, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "froze their original parameters and proposed an ID-aware Multi-scale adapter (IMAdapter). This approach achieves high-fidelity modeling of the intricate details in portraits. Our main contributions can be summarized as follows:", + "bbox": [ + 516, + 92, + 903, + 150 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose the first implicit condition animation framework based on stable video diffusion, which significantly improves temporal consistency and achieves high fidelity and precise control for portrait animation.", + "- We develop a series of strategies to enhance identity consistency and improve the accuracy of portrait motion, including data augmentation, enhanced training and inference strategies, and improved network architecture.", + "- Experiments demonstrate the effectiveness of our proposed method, which achieves state-of-the-art performance in both subjective and objective metrics." + ], + "bbox": [ + 511, + 152, + 903, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 516, + 332, + 653, + 345 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Human portrait animation aims to bring a still image to life by utilizing a driving source, such as a sequence of facial landmarks or frames that contain the head of the portrait. Recent advancements in the field of neural avatars can be divided into two subcategories: non-diffusion-based and diffusion-based methods. Non-diffusion-based methods [13, 17, 18, 26, 42, 49, 55, 62, 66] are renowned for their ability to achieve realistic facial animations and fidelity to motion. Many works employed predetermined motion representations that are popular in the literature, such as blendshapes from 3D Morphable Models (3DMM) [7, 20, 22, 35, 44, 60, 61]. However, these works often display significant artifacts when there is a large disparity between the driving portrait and the source image. Due to the use of a warping-based strategy, these methods are also ineffective at capturing videos with significant head and body movements. With the development of diffusion models [25], recent methods [12, 51, 53, 57] achieve human portrait animation by fine-tuning the stable diffusion model [39]. Several studies have investigated the generation of facial expressions through explicit keypoint control [3, 48, 67]. This method shows great promise for enhancing the quality of generated videos and enables precise control of facial expressions through the extraction of facial landmarks. However, given the diversity of facial geometries, there are notable differences in the distribution of facial keypoints among individuals. The alignment processes used in these methods often struggle to manage boundary cases effectively, resulting in identification offsets due to inaccurate alignment. In addition, the reliance on image-based generative models and separately trained motion modules to achieve temporal consistency results in shortcomings related to temporal smoothness. In the pursuit of improved temporal continuity, pioneering work [21, 27, 38, 64] has employed stable video diffusion models to construct end-to-end diffusion models for human video generation. However, these methods are pri", + "bbox": [ + 516, + 356, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15910", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/de615433cfc9cb796a91735626fead6552017e81571f987ea3475810ddc42e85.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 98, + 88, + 903, + 308 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e95f8eea86262ba3c83d889629e2fe3c0687b973cf8f82c8ba08b023aa2dc88e.jpg", + "image_caption": [ + "Figure 2. Our framework utilizes implicit representation to encode motion information as control signals. By harnessing the capabilities of stable video diffusion as the primary building block, we have meticulously designed a fine-grained appearance extractor to maintain the identity of the portrait, along with an intensity-aware motion extractor to capture intricate facial dynamics." + ], + "image_footnote": [], + "bbox": [ + 93, + 318, + 478, + 467 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c8aef9ebb1feb928dbfbf9f27ab00ef311674bcfa8f72237008d8d3065f82bda.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 480, + 321, + 903, + 465 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "marily designed for full-body movements and do not take into account the preservation of facial identity or the diversity of human facial geometries. As a result, they often lead to significant alterations in the identity of the generated face. Additionally, these methods lack the ability to control subtle facial dynamics, which limits their application in generating portraits that incorporate speech content and facial expressions. In this study, we investigate the strategy of adapting to various face shapes through implicit representations of expression descriptors and achieving capture of facial dynamics. Our method surpasses previous approaches in terms of temporal consistency and generalization ability.", + "bbox": [ + 88, + 546, + 482, + 728 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 744, + 225, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 771, + 230, + 786 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion Model. Our framework is built upon diffusion model, as introduced in [39]. The model is designed to facilitate efficient and stable training by conducting the diffusion and denoising processes within a latent space, rather than directly in the image space. The Variational Autoencoder (VAE), as proposed in [23], plays a pivotal role in this process by mapping images from the RGB color space into", + "bbox": [ + 89, + 794, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "a latent space. This transformation enables the diffusion process to be directed by textual embeddings.", + "bbox": [ + 511, + 546, + 906, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The process begins with the VAE model, which projects images into latent embeddings. This model comprises two key components: an encoder $\\mathcal{E}$ and a decoder $\\mathcal{D}$ . The encoder $\\mathcal{E}$ compresses the pixel space image $x$ into a latent representation $z = \\mathcal{E}(x)$ , while the decoder $\\mathcal{D}$ reconstructs it back into an image space, aiming for $\\mathcal{D}(z) \\approx x$ . Subsequently, a UNet-based architecture [40] is employed for learning the reverse denoising process within the latent space. This network integrates attention mechanisms through Transformer Blocks. The cross-attention mechanism ensures that the textual prompt is effectively integrated throughout the process. The overall training objective for the UNet can be articulated as follows:", + "bbox": [ + 511, + 577, + 908, + 772 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\underset {t, z, \\epsilon} {\\mathbb {E}} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(\\sqrt {\\bar {\\alpha} _ {t}} z + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , c, t\\right) \\| ^ {2} \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 779, + 906, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $z$ denotes the latent embedding of the training sample. $\\epsilon_{\\theta}$ and $\\epsilon$ represent the predicted noise by the diffusion model and the ground truth noise at the corresponding timestep $t$ , respectively. $c$ is the condition embedding, and the coefficient $\\bar{\\alpha}_t$ remains consistent with that employed in diffusion models.", + "bbox": [ + 511, + 810, + 908, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15911", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Overall framework", + "text_level": 1, + "bbox": [ + 91, + 90, + 274, + 104 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As illustrated in Figure 2, building upon stable video diffusion [1], our method is designed with two core components for generating portrait videos: the appearance extractor and the motion extractor. For the detailed appearance modeling of the portrait, we integrate the appearance encoder to manage the subject's identity and background in the generated video. We carefully design an ID-aware multi-scale adapter based on the Arcface encoder [6] and DiNOv2 [37] backbone to enhance visual appearance representation. For the motion extractor, we emphasize the enhancement of the implicit representation of facial dynamics. By utilizing the pre-trained motion encoder [8], we further optimize it by taking into account the intensity and the temporal continuity of motions. This involved the introduction of motion memory banks and an intensity-aware motion encoder to enhance the perception of motion intensity and refine features. These representations are conditionally integrated into the denoising U-Net through attention mechanisms. Our framework also integrates a spatial conditioning signal to ensure the stability of regions beyond the face. By utilizing these features, we generate latent features using a denoising UNet and subsequently produce video through VAEs.", + "bbox": [ + 89, + 112, + 483, + 446 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Intensity-Aware Motion Extractor", + "text_level": 1, + "bbox": [ + 89, + 455, + 390, + 470 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For portrait motion modeling, we concentrate on identity-agnostic portrait animation by utilizing an implicit representation of expression descriptors derived from raw videos, which capture a diverse array of facial dynamics (e.g., lip-sync, micro-expressions, eye gaze, and blinking). This approach differs from existing diffusion-based methods that typically employ explicit landmarks for facial expression representation [33, 51]. Considering that the fidelity of facial motion capture is often compromised by extraneous factors—such as variations in facial shape and background noise that can distort essential motion signals—we crop the central area of the face to focus on the regions most indicative of motion dynamics. The boundary is defined as the area between the eyebrows and the bottom of the mouth. This approach allows us to isolate facial movements and enhance the signal-to-noise ratio. The cropped pixels from driving videos are subsequently utilized as inputs for the pre-trained motion encoder [8] to extract a coarse implicit conditioning signal. Since motion intensity varies and exerts distinct influences on the generated pixels, we introduce an Intensity-Aware Motion Encoder into our framework. This addition enhances the coarse motion features from the motion encoder by adapting to the varying intensities of motion. The intensity is estimated by two dimensions: the degree of distortion in facial expressions and the overall amplitude of head movement. Considering that the expression and head pose intensity rely on landmark positioning and their relative differences, we propose a calcu", + "bbox": [ + 89, + 478, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "lation method for these variances using landmarks and normalization, which can be formulated as below:", + "bbox": [ + 511, + 90, + 903, + 119 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} e _ {k, j} = l _ {k, j} - l _ {k, c}, k \\in [ 1, n ], j \\in [ 1, m ] \\\\ s = \\sqrt {\\left(l _ {1 , \\max } - l _ {1 , \\min }\\right) ^ {2}}, \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 133, + 903, + 179 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $l_{k,j}$ denotes the j-th landmark point in the k-th frame and $l_{k,c}$ denotes the center of the landmark point (e.g., nose tip). $n$ and $m$ denote the numbers of frames and points. $s$ represents the face scale derived from the first frame of the landmarks. We further calculate the intensity of expressions $I_e$ and the head poses $I_h$ using the formulation below:", + "bbox": [ + 511, + 191, + 906, + 282 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI _ {e} = \\frac {1}{n \\cdot s} \\sum_ {k = 1} ^ {n} \\sqrt {\\frac {1}{m} \\sum_ {j = 1} ^ {m} \\left(e _ {k , j} - \\frac {1}{n} \\sum_ {k = 1} ^ {n} e _ {k , j}\\right) ^ {2}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 295, + 903, + 351 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI _ {h} = \\frac {1}{s} \\sqrt {\\frac {1}{n} \\sum_ {k = 1} ^ {n} (l _ {k , c} - \\frac {1}{n} \\sum_ {k = 1} ^ {n} l _ {k , c}) ^ {2}}.\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 347, + 800, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Considering the issue of accuracy in landmark detection, we discretize the calculated continuous values by dividing the intensity into $d$ discrete levels based on the range of values. We set $d$ to 64 and then map these discrete values using embeddings $\\mathcal{D}$ to their corresponding intensity feature vectors $E_{s} = \\mathrm{Concat}[\\mathcal{D}(I_{e}),\\mathcal{D}(I_{h})]$ , and concatenate them along the channel dimension to adjust the feature space in response to the dynamic range of motion intensity. In light of the fact that the sequence of motion features is derived from pixel extraction and exhibits a lack of continuity across different contexts, we propose the motion memory bank to improve the context-awareness and temporal modeling capabilities of motion features, which can be formulated as below:", + "bbox": [ + 511, + 407, + 906, + 617 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\hat {f} _ {m} = \\operatorname {A d a L N} \\left(F _ {m} + E _ {s}, E _ {s}\\right), \\\\ \\bar {f} _ {m} = \\operatorname {M H C A} (\\hat {m}, \\hat {f} _ {m}, \\hat {f} _ {m}), \\tag {4} \\\\ \\bar {f} _ {m} = \\operatorname {A d a L N} \\left(\\bar {f} _ {m}, E _ {s}\\right), \\\\ \\bar {f} _ {m} = \\operatorname {M L P} (\\bar {f} _ {m}), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 628, + 903, + 705 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $F_{m}$ denotes the features extracted by the pre-trained motion encoder, and $\\hat{m}$ signifies the query features from the motion memory bank, with $\\hat{f}_{m}$ being utilized as the key and value features. The MHCA denotes multi-head cross-attention [47] and the AdaLN denotes adaptive layer normalization [54]. The learnable memories are designed to supplement contextual information, enriching motion features rather than generating new motions. This helps the model capture pixel correlation between different frames, alleviating blurring when some facial parts undergo deformation. The refined motion features are incorporated into the denoising U-Net using cross-attention mechanisms [59].", + "bbox": [ + 511, + 719, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15912", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a8788756699950c2dd20dd00aa70d062cff2beff2fab1d8f8f07e3245c9e4a48.jpg", + "image_caption": [ + "Figure 3. The illustration of our ID-aware Multi-scale Adapter (IMAdapter). Here, $\\mathbb{C}$ represents the operation of concatenating features along the channel dimension." + ], + "image_footnote": [], + "bbox": [ + 102, + 88, + 473, + 161 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Fine-grained Appearance Extractor", + "text_level": 1, + "bbox": [ + 89, + 242, + 401, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To achieve appearance control, which encompasses both portrait identity information and background details, we propose an appearance extractor that creates a fine-grained representation of the animated portrait and background. Existing video diffusion models often struggle with identity consistency [38, 64], which results in deficiencies in the method's capacity to generate portraits. Due to the human eye's sensitivity to subtle facial changes, visual artifacts can lead to inaccurate facial details and insufficient preservation of identity in the generated videos. Therefore, we adopt a patch-level image encoder [37] as the backbone of the appearance encoder and introduce an ID-aware multi-scale adapter (IMAdapter) to enhance the model's ID consistency capabilities. To process the entire image, we first extract information related to clothing and background elements of the portrait. For the enhancement of consistency, as illustrated in Figure 2, we utilize the ID encoder to extract identity-related features. We further propose extracting multi-scale, fine-grained identity features using adapters.", + "bbox": [ + 88, + 265, + 482, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As illustrated in Figure 3, we first use the linear projection to downsample the visual features. The low-rank features are subsequently fed into convolutional modules at multiple scales, along with a cross-attention mechanism, which captures ID-aware fine-grained details for the preservation of visual identity. The cross-modal features are then linearly projected back to the original dimensions to be merged with patch-level features:", + "bbox": [ + 88, + 553, + 482, + 674 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\hat {f} _ {a} = f _ {a} \\mathbf {W} _ {d o w n} ^ {v}, \\\\ \\hat {f} _ {c} = \\operatorname {M C o n v} (\\hat {f} _ {a}), \\tag {5} \\\\ \\bar {f} _ {a} = \\mathrm {M H C A} (f _ {i d}, \\hat {f} _ {c}, \\hat {f} _ {c}), \\\\ f _ {a} = f _ {a} + \\bar {f} _ {a} \\mathbf {W} _ {u p} ^ {v}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 686, + 482, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $f_{id}$ represents the ID features, which are utilized as the query features, while $\\hat{f}_c$ serves as the key and value features. MConv signifies the multiple convolutions that process appearance features $f_a$ in parallel and concatenate them along the channel dimension. During the training phase, the sampling of ID information comes from a random frame in the video sequence; in the inference stage, it comes from the reference image.", + "bbox": [ + 89, + 779, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Training and Inference Strategies", + "text_level": 1, + "bbox": [ + 511, + 90, + 808, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For adaptation to various anime styles, we utilize AnimeGANv3 [31] to perform style conversion on the dataset. This allows the model to adjust to different styles of driving videos and source images. Instead of directly training models with AnimeGANv3, we employ its style transfer capability to convert facial portrait datasets into anime-stylized representations. These transformed data samples are then integrated into our training dataset for enhanced feature learning. To mitigate the impact of skin color and other color-related information on facial movements, we employ color jittering to augment the data for the facial cropping input of the motion encoder. For maintaining the stability of the regions beyond the face in the generated video, we utilize DWPose [58] as an additional spatial conditioning signal. Specifically, we employ the pose-guided network [19] to achieve spatial control condition mapping. We further incorporate spatial conditional control into the latent features of the first spatial convolutional layer in U-Net through addition. To simulate scenarios in which certain skeletons from the driving video are absent compared to those in the source image during the inference stage, we implement an augmentation strategy [48] for the pose sequence during training. This strategy randomly removes certain edges throughout the training process to improve the stability of spatial control. This mitigates the effects of detector accuracy limitations on detector-based control conditions and offers flexible control capability. During inference, considering that there are differences in the body shape and spatial position of bones between the source image and the driving video, we first translate the position based on the offset of the center point of the face (e.g., the position of the nose tip) and scale it according to the length ratio of the skeletons. We find that the key points of the eyes can lead to subtle variations in face shape. Therefore, in the setting of cross reenactment, we do not utilize the key points and corresponding edges of the eyes.", + "bbox": [ + 511, + 112, + 906, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 671, + 645, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experiments Setting", + "text_level": 1, + "bbox": [ + 511, + 696, + 705, + 713 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We train our model using HDTF [65], VFHQ [52], VoxCeleb2 [5], CelebV-HQ [68], and our own collected dataset. This collected dataset comprises monocular camera recordings showcasing 20 hours of authentic human video footage obtained from 200 subjects. The recordings were captured in both indoor and outdoor environments. During the training process, all of the images and videos are resized to $512 \\times 512$ . For the optimization of our framework, we employ the AdamW optimizer with a learning rate of $1 \\times 10^{-5}$ . The weights of the ID Encoder [6], Variational Autoencoder [24] from Stable Video Diffusion [2] and the blocks of DiNOv2 [37] are fixed. In the motion memory bank, we", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15913", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/d5f9b4d3828ed8733b33b71a7d91d52d1ed177c3d1d6772b67fd39ea6e04e02f.jpg", + "table_caption": [ + "Table 1. Quantitative results on $512 \\times 512$ test images. LMD multiplied by $10^{-3}$ , AED multiplied by $10^{-2}$ , APD multiplied by $10^{-3}$ and Identity Similarity multiplied by $10^{-1}$ . $\\uparrow$ indicates higher is better. $\\downarrow$ indicates lower is better. The best results are in bold." + ], + "table_footnote": [], + "table_body": "
MethodSelf ReenactmentCross Reenactment
LMD ↓FID-VID ↓FVD ↓PSNR ↑SSIM ↑LPIPS ↓AED ↓APD ↓Identity ↑ SimilarityFacial ↑ MovementsVideo ↑ QualityTemporal ↑ Smoothness
LivePortrait [13]9.1482.71483.3831.410.720.2226.8326.978.713.004.114.06
AntiPortrait [51]6.6781.90430.2430.540.670.2722.4221.327.953.834.003.81
FollowYE [34]5.6377.15417.5330.840.680.2520.0620.958.184.114.233.33
X-Potrait [53]6.2382.93416.4130.810.710.1920.6620.398.033.883.733.49
HunyuanPortrait (Ours)2.0275.81333.4832.980.810.1119.4519.208.874.554.694.61
", + "bbox": [ + 94, + 127, + 903, + 232 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "incorporate 64 learnable motion memories, each with a dimension of 768, which share parameters across the various blocks. They are initialized using a scaled normal distribution, where each element is sampled from $\\mathcal{N}(0,\\frac{1}{\\sqrt{\\mathrm{hidden\\_dim}}})$ . To ensure stability, we apply gradient norm clipping at a value of 0.99. During inference, we employ the DDIM sampler [43] and set the scale of classifier-free guidance [16] to 2.0 for our experiment. The framework is trained on 128 NVIDIA A100 GPUs for three days.", + "bbox": [ + 88, + 258, + 480, + 396 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Metrics", + "text_level": 1, + "bbox": [ + 89, + 407, + 187, + 422 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We employ both qualitative and quantitative analyses to assess the generated video quality and motion accuracy of our portrait animation results. For self-reenactment, we utilize the Fréchet Inception Distance (FID) [15], the Fréchet Video Distance (FVD) [46], Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM) [50], Learned Perceptual Image Patch Similarity (LPIPS) [63] and the Landmark Mean Distances (LMD) [32]. For cross-reenactment, we utilize Average Expression Distance (AED) [42], Average Pose Distance (APD) [42] and the ArcFace Score [6] as the identity (ID) similarity. Detailed descriptions of these metrics are provided in supplementary materials. Considering the absence of ground truth and the significant error between current evaluation methods and human perception, we combine the results of user studies that evaluates cross-reenactment from three additional dimensions: facial movements, video quality, and temporal smoothness. The facial movements encompass not only facial expressions but also the overall orientation of the generated portrait head. The quality of the video and its temporal smoothness are primarily assessed by the clarity of the video and the smoothness of the transitions between frames. For evaluation, we utilize source images and driving videos from two prominent datasets [5, 52]. This experiment aims to evaluate the model's ability to transfer expressions and movements from one subject to another while preserving the identity of the source portrait.", + "bbox": [ + 91, + 429, + 483, + 838 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Quantitative and Qualitative Analysis", + "text_level": 1, + "bbox": [ + 89, + 848, + 416, + 864 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Table 1, our method attains the lowest FIDVID (average FID of frames in videos) and FVD, with a", + "bbox": [ + 89, + 869, + 482, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "significant advantage when compared to the previously top-performing method, which underscores the superior capability of HunyuanPortrait. We also find that non-diffusion-based methods have much lower generation quality compared to diffusion-based methods. Although the video resolution is the same, the generated portrait details have significant differences. In more challenging cross-reenactment scenarios, our approach demonstrates strong generalization capabilities, leading to substantial improvements across all metrics. This fully demonstrates the advantage of our method, as cross-reenactment is much more difficult than self-reenactment and is closer to real-life application scenarios. The warping-based method LivePortrait struggles to accurately model overall head movement, resulting in generated portraits that exhibit minimal variation in head pose relative to the reference image. Compared with other methods, our method has stronger ID similarity, facial dynamic generation ability, video quality, and temporal smoothness. Our method employs implicit motion representation rather than explicit facial keypoints to control the generation of facial expressions. Our method further takes into account the intensity of motion, the ability to perceive identity, and complementary information based on temporal context. This approach mitigates identity distortion caused by variations in facial geometries and effectively captures dynamic facial details that explicit keypoints cannot represent. Meanwhile, since our method is based on stable video diffusion, the generated results are smoother compared to previous image-based diffusion methods and have higher video quality. These enhancements highlight the superior accuracy and fidelity of our method in capturing fine-grained details, affirming its efficacy in synthesizing realistic portrait motions. Furthermore, due to the lack of open-source training code, combined with the challenges related to their motion module and the frame rate of the videos, FollowYourEmoji and X-Portrait shows suboptimal performance in terms of temporal smoothness. Our method remains unaffected by variations in frame rate.", + "bbox": [ + 511, + 258, + 906, + 833 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Visualization", + "text_level": 1, + "bbox": [ + 513, + 845, + 651, + 861 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As illustrated in Figure 4, our method can capture more detailed and consistent portrait movements with the driv", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15914", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/81cdda35b14df4f6649f2d0cc102d583a41bad022873b8cd1eb3888319e71d7b.jpg", + "image_caption": [ + "Figure 4. Qualitative comparisons of self-reenactment and cross-reenactment with state-of-the-art methods." + ], + "image_footnote": [], + "bbox": [ + 93, + 87, + 906, + 636 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ing video. For self-reenactment, our method generates results that enable more detailed capture of facial dynamics, including eye gaze direction, eye rotation, and lip synchronization. In addition, our method can also achieve tracking of the overall rotation of the head, which cannot be achieved by previous top-performing warping-based method LivePortrait, such as LivePortrait. Current state-of-the-art diffusion-based methods have significant defects in maintaining the portrait's identity and facial shape due to the deviation of predicted keypoints during head rotation. In contrast, our method effectively preserves the identity of the portrait and captures intricate facial dynamics. Previous methods are influenced by the differences in facial geometries between the driving video and the reference image,", + "bbox": [ + 89, + 688, + 485, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "leading to substantial challenges in preserving the identity of the generated video. This is a result of inadequate disentanglement of their appearance features and facial dynamics. These issues result in problems such as blurry videos and distorted portraits. Our framework effectively addresses the disentanglement of the two key elements involved in portrait animation, thereby demonstrating a strong generalization ability, as shown in Figure 4. The diffusion-based method exhibits superior generation quality; however, discrepancies in facial keypoints due to variations in facial shape, combined with the limited ability of these keypoints to accurately capture dynamic facial details, make the previous state-of-the-art method less effective than ours in generating features such as facial geometry, eye gaze direc", + "bbox": [ + 511, + 688, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15915", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/30ccdfece604e8ff32ea2087997d1481740afb4bb8b3c82ab89f37b2ad4d2e2b.jpg", + "image_caption": [ + "Figure 5. Qualitative ablation studies of different components." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 480, + 349 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tion, and lip synchronization in portraits. Due to variations in facial shapes, methods that rely on explicit keypoints are influenced by the facial characteristics of individuals in the driving video. This leads to insufficient separation between motion and identity. Furthermore, since their approach utilizes the image generation model Stable Diffusion as its foundation, it also exhibits inferior performance compared to ours in terms of temporal smoothness. Building upon stable video diffusion and utilizing our meticulously designed fine-grained feature extractors, our method exhibits exceptional performance in identity preservation, video quality enhancement, and the capture of subtle facial movements, resulting in more vivid and smoother generated videos.", + "bbox": [ + 89, + 383, + 483, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 590, + 243, + 606 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of Motion Memory Bank. We validate the effectiveness of the motion memory bank. As shown in Table 2, the incorporation of the motion memory bank results in higher-quality videos, demonstrating improvements in FID-VID and FVD, as well as enhanced facial movements. We also observe an improvement in smoothness. Through the utilization of motion memories, our methodology demonstrates context-based perception capabilities, thereby enhancing motion features in the presence of ambiguous information. By integrating memories with information from preceding and subsequent frames, the memory bank can provide more detailed information into facial dynamics. As illustrated in Figure 5, the forehead wrinkles associated with the elevation of the eyes become more pronounced, when the character's eyebrows are raised.", + "bbox": [ + 89, + 613, + 482, + 839 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of Intensity-Aware Motion Encoder (IAME). We confirm the significant role of the intensity-aware motion encoder. As illustrated in Table 2, the incorporation of this component, in contrast to the direct utilization of coarse", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5e9895874bda8d589aca8a1a0d6f8e3f7083c3ef7f2588167bf69d932412f232.jpg", + "table_caption": [ + "Table 2. Ablation study to validate the effectiveness of different components. LMD (Landmark Mean Distance) multiplied by $10^{-3}$ and Identity Similarity multiplied by $10^{-1}$ ." + ], + "table_footnote": [], + "table_body": "
SettingIdentity ↑ SimilarityFID-VID ↓FVD ↓LMD ↓
HunyuanPortrait8.8775.81333.482.02
- Memory Bank8.7578.43361.942.78
- IAME8.6380.79385.434.01
- ID Features8.2175.03330.121.98
- IMAdapter8.0977.14352.672.54
", + "bbox": [ + 517, + 142, + 903, + 247 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "motion features derived from the pre-trained motion encoder [8], offers critical guidance and intricate details regarding the facial dynamics associated with portrait motions. This enhancement significantly elevates the overall quality of the generated video, evidenced by an improvement in FID-VID, FVD, and ID similarity. As illustrated in Figure 5, the absence of feature refinement in the IAME may result in inaccuracies in the micro-expressions of the characters, including the direction of their gaze. Furthermore, the intricate details of the face are neglected.", + "bbox": [ + 511, + 273, + 906, + 422 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of ID features and IMAdapter. To validate the effectiveness of the IMAdapter, we first remove the ID features from the IMAdapter, leaving only multiple convolution kernels for low-rank feature processing. As shown in Table 2 and Figure 5, the ablation results demonstrate that incorporating ID features is critical in enhancing the model's ID consistency ability. We further remove IMAdapter from the appearance extractor's blocks. As demonstrated in Figure 5, the incorporation of IMAdapter provides more fine-grained features related to facial texture, geometry, and other facial characteristics.", + "bbox": [ + 511, + 424, + 908, + 590 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 603, + 633, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present an enhanced portrait animation framework that addresses the limitations of previous approaches. Our model leverages implicit representations with stable video diffusion to achieve temporal consistency and precise control over facial dynamics. The proposed model demonstrates significant improvements in the overall quality and temporal consistency of generated videos, allowing for more natural and lifelike animations. We offer a robust framework for generating high-quality, controllable animations, which achieve high disentanglement. Our method not only enhances the visual experience but also opens up possibilities for immersive content creation. With continued research and development, this approach holds great potential for applications in virtual reality, gaming, and human-computer interaction. However, as with any technology that generates realistic human likenesses, it is essential to consider the ethical implications. Future work should also develop safeguards to ensure responsible use.", + "bbox": [ + 511, + 628, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15916", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 4", + "[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 5", + "[3] Zhiyuan Chen, Jiajiong Cao, Zhiquan Chen, Yuming Li, and Chenguang Ma. Echomimic: Lifelike audio-driven portrait animations through editable landmark conditions. arXiv preprint arXiv:2407.08136, 2024. 1, 2", + "[4] Kun Cheng, Xiaodong Cun, Yong Zhang, Menghan Xia, Fei Yin, Mingrui Zhu, Xuan Wang, Jue Wang, and Nannan Wang. Videoretalking: Audio-based lip synchronization for talking head video editing in the wild. In SIGGRAPH Asia 2022, pages 1-9, 2022. 1", + "[5] Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. Voxceleb2: Deep speaker recognition. arXiv preprint arXiv:1806.05622, 2018. 5, 6", + "[6] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 2, 4, 5, 6", + "[7] Michail Christos Doukas, Stefanos Zafeiriou, and Viktoriia Sharmanska. Headgan: One-shot neural head synthesis and editing, 2021. 1, 2", + "[8] Nikita Drobyshev, Jenya Chelishev, Taras Khakhulin, Aleksei Ivakhnenko, Victor Lempitsky, and Egor Zakharov. Megaportraits: One-shot megapixel neural head avatars. In Proceedings of the 30th ACM International Conference on Multimedia, pages 2663-2671, 2022. 4, 8", + "[9] Nikita Drobyshev, Jenya Chelishev, Taras Khakhulin, Aleksei Ivakhnenko, Victor Lempitsky, and Egor Zakharov. Megaportraits: One-shot megapixel neural head avatars, 2023. 2", + "[10] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 1", + "[11] Yuming Gu, Hongyi Xu, You Xie, Guoxian Song, Yichun Shi, Di Chang, Jing Yang, and Linjie Luo. Diffportrait3d: Controllable diffusion for zero-shot portrait view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10456-10465, 2024. 1", + "[12] Hanzhong Guo, Hongwei Yi, Daquan Zhou, Alexander William Bergman, Michael Lingelbach, and Yizhou Yu. Real-time one-step diffusion-based expressive portrait videos generation. arXiv preprint arXiv:2412.13479, 2024. 2", + "[13] Jianzhu Guo, Dingyun Zhang, Xiaogiang Liu, Zhizhou Zhong, Yuan Zhang, Pengfei Wan, and Di Zhang. Livepor" + ], + "bbox": [ + 93, + 114, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "trait: Efficient portrait animation with stitching and retargeting control. arXiv preprint arXiv:2407.03168, 2024. 2, 6", + "[14] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. In The Twelfth International Conference on Learning Representations, 2024. 2", + "[15] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6", + "[16] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 6", + "[17] Fa-Ting Hong and Dan Xu. Implicit identity representation conditioned memory compensation network for talking head video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23062-23072, 2023. 2", + "[18] Fa-Ting Hong, Li Shen, and Dan Xu. Dagan++: Depth-aware generative adversarial network for talking head video generation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2", + "[19] Li Hu. Animate anyone: Consistent and controllable image-to-video synthesis for character animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8153-8163, 2024. 5", + "[20] Xianliang Huang, Jiajie Gou, Shuhang Chen, Zhizhou Zhong, Jihong Guan, and Shuigeng Zhou. Iddr-ngp: Incorporating detectors for distractors removal with instant neural radiance field. In Proceedings of the 31st ACM International Conference on Multimedia, pages 1343-1351, 2023. 2", + "[21] Xiaoyu Jin, Zunnan Xu, Mingwen Ou, and Wenming Yang. Alignment is all you need: A training-free augmentation strategy for pose-guided video generation. arXiv preprint arXiv:2408.16506, 2024. 2", + "[22] Taras Khakhulin, Vanessa Sklyarova, Victor Lempitsky, and Egor Zakharov. Realistic one-shot mesh-based head avatars. In European Conference on Computer Vision, pages 345-362. Springer, 2022. 2", + "[23] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3", + "[24] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013.5", + "[25] Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, et al. Hunyuanvideo: A systematic framework for large video generative models. arXiv preprint arXiv:2412.03603, 2024. 2", + "[26] Hongxiang Li, Meng Cao, Xuxin Cheng, Yaowei Li, Zhihong Zhu, and Yuexian Zou. G2l: Semantically aligned and uniform video grounding via geodesic and game theory. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12032-12042, 2023. 2", + "[27] Hongxiang Li, Yaowei Li, Yuhang Yang, Junjie Cao, Zhihong Zhu, Xuxin Cheng, and Long Chen. Dispose: Disen" + ], + "bbox": [ + 516, + 92, + 906, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15917", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tangling pose guidance for controllable human image animation. arXiv preprint arXiv:2412.09349, 2024. 2", + "[28] Ronghui Li, Junfan Zhao, Yachao Zhang, Mingyang Su, Zeping Ren, Han Zhang, Yansong Tang, and Xiu Li. Finedance: A fine-grained choreography dataset for 3d full body dance generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10234-10243, 2023. 1", + "[29] Ronghui Li, Hongwen Zhang, Yachao Zhang, Yuxiang Zhang, Youliang Zhang, Jie Guo, Yan Zhang, Xiu Li, and Yebin Liu. Lodge++: High-quality and long dance generation with vivid choreography patterns. arXiv preprint arXiv:2410.20389, 2024.", + "[30] Ronghui Li, YuXiang Zhang, Yachao Zhang, Hongwen Zhang, Jie Guo, Yan Zhang, Yebin Liu, and Xiu Li. Lodge: A coarse to fine diffusion network for long dance generation guided by the characteristic dance primitives. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1524-1534, 2024. 1", + "[31] Gang LIU, Xin CHEN, and Zhixiang GAO. A novel double-tail generative adversarial network for fast photo animation. IEICE Transactions on Information and Systems, E107.D(1): 72-82, 2024. 5", + "[32] Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuoling Chang, Ming Guang Yong, Juhyun Lee, et al. Mediapipe: A framework for building perception pipelines. arXiv preprint arXiv:1906.08172, 2019. 6", + "[33] Yue Ma, Yingqing He, Xiaodong Cun, Xintao Wang, Siran Chen, Xiu Li, and Qifeng Chen. Follow your pose: Pose-guided text-to-video generation using pose-free videos. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4117-4125, 2024. 4", + "[34] Yue Ma, Hongyu Liu, Hongfa Wang, Heng Pan, Yingqing He, Junkun Yuan, Ailing Zeng, Chengfei Cai, Heung-Yeung Shum, Wei Liu, et al. Follow-your-emoji: Fine-controllable and expressive freestyle portrait animation. arXiv preprint arXiv:2406.01900, 2024. 6", + "[35] Yuxi Mi, Zhizhou Zhong, Yuge Huang, Jiazhen Ji, Jianqing Xu, Jun Wang, Shaoming Wang, Shouhong Ding, and Shuigeng Zhou. Privacy-preserving face recognition using trainable feature subtraction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 297-307, 2024. 2", + "[36] Yuval Nirkin, Yosi Keller, and Tal Hassner. Fsgan: Subject agnostic face swapping and reenactment. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7184-7193, 2019. 1", + "[37] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 4, 5", + "[38] Bohao Peng, Jian Wang, Yuechen Zhang, Wenbo Li, MingChang Yang, and Jiaya Jia. Controlnext: Powerful and efficient control for image and video generation. arXiv preprint arXiv:2408.06070, 2024. 2, 5" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 2, 3", + "[40] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 3", + "[41] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. Advances in neural information processing systems, 32, 2019. 1", + "[42] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. In Advances in Neural Information Processing Systems, 2019. 2, 6", + "[43] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 6", + "[44] Jiale Tao, Shuhang Gu, Wen Li, and Lixin Duan. Learning motion refinement for unsupervised face animation. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[45] Jiayuan Tian, Yanling LONG, Hong YANG, Huifang WU, Peng XUE, and Zhongqing JIANG. Cues of eye region and their effects on face-personality perception. Advances in Psychological Science, 30(12):2735, 2022. 2", + "[46] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphaël Marinier, Marcin Michalski, and Sylvain Gelly. Fvd: A new metric for video generation. 2019. 6", + "[47] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems, 30, 2017. 4", + "[48] Cong Wang, Kuan Tian, Jun Zhang, Yonghang Guan, Feng Luo, Fei Shen, Zhiwei Jiang, Qing Gu, Xiao Han, and Wei Yang. V-express: Conditional dropout for progressive training of portrait video generation. arXiv preprint arXiv:2406.02511, 2024. 1, 2, 5", + "[49] Suzhen Wang, Lincheng Li, Yu Ding, and Xin Yu. One-shot talking face generation from single-speaker audio-visual correlation learning. In AAAI Conference on Artificial Intelligence, pages 2531–2539, 2022. 2", + "[50] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6", + "[51] Huawei Wei, Zejun Yang, and Zhisheng Wang. Aniportrait: Audio-driven synthesis of photorealistic portrait animations, 2024. 2, 4, 6", + "[52] Liangbin Xie, Xintao Wang, Honglun Zhang, Chao Dong, and Ying Shan. Vfhq: A high-quality dataset and benchmark for video face super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 657-666, 2022. 5, 6" + ], + "bbox": [ + 516, + 90, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15918", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] You Xie, Hongyi Xu, Guoxian Song, Chao Wang, Yichun Shi, and Linjie Luo. X-portrait: Expressive portrait animation with hierarchical motion attention. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2, 6", + "[54] Jingjing Xu, Xu Sun, Zhiyuan Zhang, Guangxiang Zhao, and Junyang Lin. Understanding and improving layer normalization. Advances in neural information processing systems, 32, 2019. 4", + "[55] Sicheng Xu, Guojun Chen, Yu-Xiao Guo, Jiaolong Yang, Chong Li, Zhenyu Zang, Yizhong Zhang, Xin Tong, and Baining Guo. Vasa-1: Lifelike audio-driven talking faces generated in real time. arXiv preprint arXiv:2404.10667, 2024. 2", + "[56] Zunnan Xu, Yukang Lin, Haonan Han, Sicheng Yang, Ronghui Li, Yachao Zhang, and Xiu Li. Mambatak: Efficient holistic gesture synthesis with selective state space models. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 1", + "[57] Shurong Yang, Huadong Li, Juhao Wu, Minhao Jing, Linze Li, Renhe Ji, Jiajun Liang, and Haoqiang Fan. Megactor: Harness the power of raw video for vivid portrait animation. arXiv preprint arXiv:2405.20851, 2024. 2", + "[58] Zhendong Yang, Ailing Zeng, Chun Yuan, and Yu Li. Effective whole-body pose estimation with two-stages distillation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4210-4220, 2023. 5", + "[59] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023.4", + "[60] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. Styleheat: One-shot high-resolution editable talking face generation via pre-trained stylegan. In European Conference on Computer Vision, pages 85–101, 2022. 2", + "[61] Wangbo Yu, Yanbo Fan, Yong Zhang, Xuan Wang, Fei Yin, Yunpeng Bai, Yan-Pei Cao, Ying Shan, Yang Wu, Zhongqian Sun, and Baoyuan Wu. Nofa: Nerf-based one-shot facial avatar reconstruction, 2023. 2", + "[62] Bowen Zhang, Chenyang Qi, Pan Zhang, Bo Zhang, Hsiang-Tao Wu, Dong Chen, Qifeng Chen, Yong Wang, and Fang Wen. Metaportrait: Identity-preserving talking head generation with fast personalized adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22096-22105, 2023. 1, 2", + "[63] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6", + "[64] Yang Zhang, Jiaxi Gu, Li-Wen Wang, Han Wang, Junqi Cheng, Yuefeng Zhu, and Fangyuan Zou. Mimic motion: High-quality human motion video generation with confidence-aware pose guidance. arXiv preprint arXiv:2406.19680, 2024. 2, 5", + "[65] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with" + ], + "bbox": [ + 91, + 90, + 480, + 901 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661-3670, 2021. 5", + "[66] Jian Zhao and Hui Zhang. Thin-plate spline motion model for image animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3657–3666, 2022. 2", + "[67] Longtao Zheng, Yifan Zhang, Hanzhong Guo, Jiachun Pan, Zhenxiong Tan, Jiahao Lu, Chuanxin Tang, Bo An, and Shuicheng Yan. Memo: Memory-guided diffusion for expressive talking video generation. arXiv preprint arXiv:2412.04448, 2024. 2", + "[68] Hao Zhu, Wayne Wu, Wentao Zhu, Liming Jiang, Siwei Tang, Li Zhang, Ziwei Liu, and Chen Change Loy. Celebvehq: A large-scale video facial attributes dataset. In European conference on computer vision, pages 650-667. Springer, 2022. 5" + ], + "bbox": [ + 516, + 90, + 906, + 330 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "15919", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/HunyuanPortrait_ Implicit Condition Control for Enhanced Portrait Animation/a6cc1e88-c49b-4877-897d-6a6632aadecd_model.json b/2025/HunyuanPortrait_ Implicit Condition Control for Enhanced Portrait Animation/a6cc1e88-c49b-4877-897d-6a6632aadecd_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8ad574f9be8346d78154a103ced7d36e9817a304 --- /dev/null +++ b/2025/HunyuanPortrait_ Implicit Condition Control for Enhanced Portrait Animation/a6cc1e88-c49b-4877-897d-6a6632aadecd_model.json @@ -0,0 +1,1960 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.098, + 0.131, + 0.9, + 0.152 + ], + "angle": 0, + "content": "HunyuanPortrait: Implicit Condition Control for Enhanced Portrait Animation" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.18, + 0.871, + 0.235 + ], + "angle": 0, + "content": "Zunnan Xu\\(^{1,2*}\\), Zhentao Yu\\(^{2}\\), Zixiang Zhou\\(^{2}\\), Jun Zhou\\(^{3*}\\), Xiaoyu Jin\\(^{1*}\\), Fa-ting Hong\\(^{4*}\\), Xiaozhong Ji\\(^{2}\\), Junwei Zhu\\(^{2}\\), Chengfei Cai\\(^{2}\\), Shiyu Tang\\(^{2}\\), Qin Lin\\(^{2}\\), Xiu Li\\(^{1\\dagger}\\), Qinglin Lu\\(^{2\\dagger}\\), Tsinghua University \\({}^{2}\\)Hunyuan, Tencent \\({}^{3}\\)Sun Yat-Sen Univeristy \\({}^{4}\\)HKUST" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.301, + 0.484, + 0.618 + ], + "angle": 0, + "content": "We introduce HunyuanPortrait, a diffusion-based condition control method that employs implicit representations for highly controllable and lifelike portrait animation. Given a single portrait image as an appearance reference and video clips as driving templates, HunyuanPortrait can animate the character in the reference image by the facial expression and head pose of the driving videos. In our framework, we utilize pre-trained encoders to achieve the decoupling of portrait motion information and identity in videos. To do so, implicit representation is adopted to encode motion information and is employed as control signals in the animation phase. By leveraging the power of stable video diffusion as the main building block, we carefully design adapter layers to inject control signals into the denoising unet through attention mechanisms. These bring spatial richness of details and temporal consistency. HunyuanPortrait also exhibits strong generalization performance, which can effectively disentangle appearance and motion under different image styles. Our framework outperforms existing methods, demonstrating superior temporal consistency and controllability. Our project is available at HunyuanPortrait." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.646, + 0.222, + 0.66 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.671, + 0.483, + 0.761 + ], + "angle": 0, + "content": "In this paper, we aim to animate a single static image into a lifelike video using control signals in implicit representations derived from driving videos. As shown in Figure 1, our primary focus is not only on the high controllability of digital portraits but also on the strong consistency of the characters and background in the reference image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.762, + 0.483, + 0.868 + ], + "angle": 0, + "content": "Portrait animation has been widely applied in fields such as human-computer interaction [11, 56], virtual reality, and the crafting of digital avatars for the gaming and multimedia industries [28-30]. Previous works [4, 7, 62] develop usable portrait animation models utilizing Generative Adversarial Networks (GANs) [10] based on raw video inputs. However, these approaches exhibit limitations in terms of gener" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.268, + 0.905, + 0.613 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.624, + 0.908, + 0.709 + ], + "angle": 0, + "content": "Figure 1. Our framework employs implicit condition control to generate portrait animations, demonstrating robust generalization performance with high-fidelity facial dynamics and vivid head poses during cross-reenactment. The animated portraits remain unaffected by variations in facial shape and the spatial position of the driving videos, demonstrating strong identity consistency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.734, + 0.907, + 0.87 + ], + "angle": 0, + "content": "alization. Given the variability of facial structures and the complexities of facial movements, these methods often face challenges such as artifacts and insufficient control over facial expressions. This is particularly evident when there is a substantial difference in facial shape between the driving video and the source image, resulting in significant issues with facial artifacts and motion distortion. These early methods often produce results with background jitter and blurred portraits during cross-reenactment [36, 41]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "The advancement of diffusion models has led to a significant focus in recent methodologies [3, 48] on the fine-tuning" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.348, + 0.888 + ], + "angle": 0, + "content": "*Work done during the internship at Tencent." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.236, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.348, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15909" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.094, + 0.092, + 0.482, + 0.453 + ], + "angle": 0, + "content": "of image-level diffusion models [39] through the incorporation of motion modules [14], thereby facilitating the generation of video content. These methods effectively address the issues associated with GAN-based approaches in managing complex backgrounds and enhancing generalization. However, even with the incorporation of motion modules, these methods are susceptible to defects during the generation stage (e.g., temporal smoothness and flexibility to different frame rates). This vulnerability arises from the absence of motion pre-training in the underlying models during the pre-training phase. Additionally, these methods remain constrained to the explicit control of facial expressions. The effectiveness of these methods, which depend on explicit keypoints to control facial expressions, not only results in the loss of intricate details but also significantly relies on the accuracy of the landmark extraction techniques employed. Furthermore, as a result of the variability in facial shapes, these methods are limited by post-processing alignment strategies and often face challenges in adapting to diverse facial geometries, which can lead to inaccurate facial control and insufficient preservation of identity in the visual effects of videos. Due to the human eye's sensitivity to subtle facial changes [45], it is crucial to preserve the fidelity of facial details and the smoothness of the videos." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.464, + 0.482, + 0.9 + ], + "angle": 0, + "content": "Therefore, we propose an implicit conditional control model to tackle the challenges posed by diverse facial geometries and intricate expression details. We utilize stable video diffusion as the foundational model and integrate identity and motion information through appearance and motion attention. This approach eliminates the need to fine-tune the image diffusion model and to train the motion module separately. Firstly, we coarsely decouple the identity information from the motion information in the original video by using a pre-trained motion encoder [9]. Considering that this pre-trained encoder has not completely decoupled identity from motion information, we further implement enhanced training strategies and improved network architectures to strengthen the control of motion features and the separation of portrait identities. For the adaptation of temporal modeling tasks in video generation using diffusion models, along with the complementary contextual information present between different frames, we propose the incorporation of a motion memory bank to enhance the implicit representation of motion features. Due to the substantial variations in the degree of blur and pixel distortion in video frames at different levels of motion intensity, we propose an intensity-aware motion encoder. This encoder aims to improve the capacity of decoupled motion features to capture intricate details. Additionally, in order to achieve consistent modeling of portrait identity and background areas in videos, we combined ArcFace [6] with the DiNOv2 [37] backbone to develop an enhanced appearance encoder. To preserve the pre-training knowledge of both models, we" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.093, + 0.904, + 0.151 + ], + "angle": 0, + "content": "froze their original parameters and proposed an ID-aware Multi-scale adapter (IMAdapter). This approach achieves high-fidelity modeling of the intricate details in portraits. Our main contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.153, + 0.904, + 0.211 + ], + "angle": 0, + "content": "- We propose the first implicit condition animation framework based on stable video diffusion, which significantly improves temporal consistency and achieves high fidelity and precise control for portrait animation." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.213, + 0.904, + 0.271 + ], + "angle": 0, + "content": "- We develop a series of strategies to enhance identity consistency and improve the accuracy of portrait motion, including data augmentation, enhanced training and inference strategies, and improved network architecture." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.273, + 0.904, + 0.318 + ], + "angle": 0, + "content": "- Experiments demonstrate the effectiveness of our proposed method, which achieves state-of-the-art performance in both subjective and objective metrics." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.153, + 0.904, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.517, + 0.333, + 0.654, + 0.347 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.357, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Human portrait animation aims to bring a still image to life by utilizing a driving source, such as a sequence of facial landmarks or frames that contain the head of the portrait. Recent advancements in the field of neural avatars can be divided into two subcategories: non-diffusion-based and diffusion-based methods. Non-diffusion-based methods [13, 17, 18, 26, 42, 49, 55, 62, 66] are renowned for their ability to achieve realistic facial animations and fidelity to motion. Many works employed predetermined motion representations that are popular in the literature, such as blendshapes from 3D Morphable Models (3DMM) [7, 20, 22, 35, 44, 60, 61]. However, these works often display significant artifacts when there is a large disparity between the driving portrait and the source image. Due to the use of a warping-based strategy, these methods are also ineffective at capturing videos with significant head and body movements. With the development of diffusion models [25], recent methods [12, 51, 53, 57] achieve human portrait animation by fine-tuning the stable diffusion model [39]. Several studies have investigated the generation of facial expressions through explicit keypoint control [3, 48, 67]. This method shows great promise for enhancing the quality of generated videos and enables precise control of facial expressions through the extraction of facial landmarks. However, given the diversity of facial geometries, there are notable differences in the distribution of facial keypoints among individuals. The alignment processes used in these methods often struggle to manage boundary cases effectively, resulting in identification offsets due to inaccurate alignment. In addition, the reliance on image-based generative models and separately trained motion modules to achieve temporal consistency results in shortcomings related to temporal smoothness. In the pursuit of improved temporal continuity, pioneering work [21, 27, 38, 64] has employed stable video diffusion models to construct end-to-end diffusion models for human video generation. However, these methods are pri" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15910" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.089, + 0.904, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.319, + 0.48, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.482, + 0.322, + 0.904, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.478, + 0.908, + 0.522 + ], + "angle": 0, + "content": "Figure 2. Our framework utilizes implicit representation to encode motion information as control signals. By harnessing the capabilities of stable video diffusion as the primary building block, we have meticulously designed a fine-grained appearance extractor to maintain the identity of the portrait, along with an intensity-aware motion extractor to capture intricate facial dynamics." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.547, + 0.483, + 0.729 + ], + "angle": 0, + "content": "marily designed for full-body movements and do not take into account the preservation of facial identity or the diversity of human facial geometries. As a result, they often lead to significant alterations in the identity of the generated face. Additionally, these methods lack the ability to control subtle facial dynamics, which limits their application in generating portraits that incorporate speech content and facial expressions. In this study, we investigate the strategy of adapting to various face shapes through implicit representations of expression descriptors and achieving capture of facial dynamics. Our method surpasses previous approaches in terms of temporal consistency and generalization ability." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.746, + 0.226, + 0.763 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.772, + 0.232, + 0.787 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Diffusion Model. Our framework is built upon diffusion model, as introduced in [39]. The model is designed to facilitate efficient and stable training by conducting the diffusion and denoising processes within a latent space, rather than directly in the image space. The Variational Autoencoder (VAE), as proposed in [23], plays a pivotal role in this process by mapping images from the RGB color space into" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.547, + 0.907, + 0.577 + ], + "angle": 0, + "content": "a latent space. This transformation enables the diffusion process to be directed by textual embeddings." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.578, + 0.909, + 0.773 + ], + "angle": 0, + "content": "The process begins with the VAE model, which projects images into latent embeddings. This model comprises two key components: an encoder \\(\\mathcal{E}\\) and a decoder \\(\\mathcal{D}\\). The encoder \\(\\mathcal{E}\\) compresses the pixel space image \\(x\\) into a latent representation \\(z = \\mathcal{E}(x)\\), while the decoder \\(\\mathcal{D}\\) reconstructs it back into an image space, aiming for \\(\\mathcal{D}(z) \\approx x\\). Subsequently, a UNet-based architecture [40] is employed for learning the reverse denoising process within the latent space. This network integrates attention mechanisms through Transformer Blocks. The cross-attention mechanism ensures that the textual prompt is effectively integrated throughout the process. The overall training objective for the UNet can be articulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.78, + 0.907, + 0.803 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\underset {t, z, \\epsilon} {\\mathbb {E}} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(\\sqrt {\\bar {\\alpha} _ {t}} z + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , c, t\\right) \\| ^ {2} \\right], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.9 + ], + "angle": 0, + "content": "where \\(z\\) denotes the latent embedding of the training sample. \\(\\epsilon_{\\theta}\\) and \\(\\epsilon\\) represent the predicted noise by the diffusion model and the ground truth noise at the corresponding timestep \\(t\\), respectively. \\(c\\) is the condition embedding, and the coefficient \\(\\bar{\\alpha}_t\\) remains consistent with that employed in diffusion models." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15911" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.276, + 0.106 + ], + "angle": 0, + "content": "3.2. Overall framework" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.485, + 0.447 + ], + "angle": 0, + "content": "As illustrated in Figure 2, building upon stable video diffusion [1], our method is designed with two core components for generating portrait videos: the appearance extractor and the motion extractor. For the detailed appearance modeling of the portrait, we integrate the appearance encoder to manage the subject's identity and background in the generated video. We carefully design an ID-aware multi-scale adapter based on the Arcface encoder [6] and DiNOv2 [37] backbone to enhance visual appearance representation. For the motion extractor, we emphasize the enhancement of the implicit representation of facial dynamics. By utilizing the pre-trained motion encoder [8], we further optimize it by taking into account the intensity and the temporal continuity of motions. This involved the introduction of motion memory banks and an intensity-aware motion encoder to enhance the perception of motion intensity and refine features. These representations are conditionally integrated into the denoising U-Net through attention mechanisms. Our framework also integrates a spatial conditioning signal to ensure the stability of regions beyond the face. By utilizing these features, we generate latent features using a denoising UNet and subsequently produce video through VAEs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.456, + 0.391, + 0.472 + ], + "angle": 0, + "content": "3.3. Intensity-Aware Motion Extractor" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.479, + 0.484, + 0.902 + ], + "angle": 0, + "content": "For portrait motion modeling, we concentrate on identity-agnostic portrait animation by utilizing an implicit representation of expression descriptors derived from raw videos, which capture a diverse array of facial dynamics (e.g., lip-sync, micro-expressions, eye gaze, and blinking). This approach differs from existing diffusion-based methods that typically employ explicit landmarks for facial expression representation [33, 51]. Considering that the fidelity of facial motion capture is often compromised by extraneous factors—such as variations in facial shape and background noise that can distort essential motion signals—we crop the central area of the face to focus on the regions most indicative of motion dynamics. The boundary is defined as the area between the eyebrows and the bottom of the mouth. This approach allows us to isolate facial movements and enhance the signal-to-noise ratio. The cropped pixels from driving videos are subsequently utilized as inputs for the pre-trained motion encoder [8] to extract a coarse implicit conditioning signal. Since motion intensity varies and exerts distinct influences on the generated pixels, we introduce an Intensity-Aware Motion Encoder into our framework. This addition enhances the coarse motion features from the motion encoder by adapting to the varying intensities of motion. The intensity is estimated by two dimensions: the degree of distortion in facial expressions and the overall amplitude of head movement. Considering that the expression and head pose intensity rely on landmark positioning and their relative differences, we propose a calcu" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "lation method for these variances using landmarks and normalization, which can be formulated as below:" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.135, + 0.905, + 0.18 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} e _ {k, j} = l _ {k, j} - l _ {k, c}, k \\in [ 1, n ], j \\in [ 1, m ] \\\\ s = \\sqrt {\\left(l _ {1 , \\max } - l _ {1 , \\min }\\right) ^ {2}}, \\tag {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.192, + 0.907, + 0.283 + ], + "angle": 0, + "content": "where \\( l_{k,j} \\) denotes the j-th landmark point in the k-th frame and \\( l_{k,c} \\) denotes the center of the landmark point (e.g., nose tip). \\( n \\) and \\( m \\) denote the numbers of frames and points. \\( s \\) represents the face scale derived from the first frame of the landmarks. We further calculate the intensity of expressions \\( I_e \\) and the head poses \\( I_h \\) using the formulation below:" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.296, + 0.905, + 0.352 + ], + "angle": 0, + "content": "\\[\nI _ {e} = \\frac {1}{n \\cdot s} \\sum_ {k = 1} ^ {n} \\sqrt {\\frac {1}{m} \\sum_ {j = 1} ^ {m} \\left(e _ {k , j} - \\frac {1}{n} \\sum_ {k = 1} ^ {n} e _ {k , j}\\right) ^ {2}}, \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.56, + 0.348, + 0.802, + 0.394 + ], + "angle": 0, + "content": "\\[\nI _ {h} = \\frac {1}{s} \\sqrt {\\frac {1}{n} \\sum_ {k = 1} ^ {n} (l _ {k , c} - \\frac {1}{n} \\sum_ {k = 1} ^ {n} l _ {k , c}) ^ {2}}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.408, + 0.907, + 0.618 + ], + "angle": 0, + "content": "Considering the issue of accuracy in landmark detection, we discretize the calculated continuous values by dividing the intensity into \\(d\\) discrete levels based on the range of values. We set \\(d\\) to 64 and then map these discrete values using embeddings \\(\\mathcal{D}\\) to their corresponding intensity feature vectors \\(E_{s} = \\mathrm{Concat}[\\mathcal{D}(I_{e}),\\mathcal{D}(I_{h})]\\), and concatenate them along the channel dimension to adjust the feature space in response to the dynamic range of motion intensity. In light of the fact that the sequence of motion features is derived from pixel extraction and exhibits a lack of continuity across different contexts, we propose the motion memory bank to improve the context-awareness and temporal modeling capabilities of motion features, which can be formulated as below:" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.629, + 0.905, + 0.706 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\hat {f} _ {m} = \\operatorname {A d a L N} \\left(F _ {m} + E _ {s}, E _ {s}\\right), \\\\ \\bar {f} _ {m} = \\operatorname {M H C A} (\\hat {m}, \\hat {f} _ {m}, \\hat {f} _ {m}), \\tag {4} \\\\ \\bar {f} _ {m} = \\operatorname {A d a L N} \\left(\\bar {f} _ {m}, E _ {s}\\right), \\\\ \\bar {f} _ {m} = \\operatorname {M L P} (\\bar {f} _ {m}), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.902 + ], + "angle": 0, + "content": "where \\(F_{m}\\) denotes the features extracted by the pre-trained motion encoder, and \\(\\hat{m}\\) signifies the query features from the motion memory bank, with \\(\\hat{f}_{m}\\) being utilized as the key and value features. The MHCA denotes multi-head cross-attention [47] and the AdaLN denotes adaptive layer normalization [54]. The learnable memories are designed to supplement contextual information, enriching motion features rather than generating new motions. This helps the model capture pixel correlation between different frames, alleviating blurring when some facial parts undergo deformation. The refined motion features are incorporated into the denoising U-Net using cross-attention mechanisms [59]." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15912" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.089, + 0.475, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.174, + 0.483, + 0.217 + ], + "angle": 0, + "content": "Figure 3. The illustration of our ID-aware Multi-scale Adapter (IMAdapter). Here, \\(\\mathbb{C}\\) represents the operation of concatenating features along the channel dimension." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.243, + 0.403, + 0.26 + ], + "angle": 0, + "content": "3.4. Fine-grained Appearance Extractor" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.266, + 0.483, + 0.553 + ], + "angle": 0, + "content": "To achieve appearance control, which encompasses both portrait identity information and background details, we propose an appearance extractor that creates a fine-grained representation of the animated portrait and background. Existing video diffusion models often struggle with identity consistency [38, 64], which results in deficiencies in the method's capacity to generate portraits. Due to the human eye's sensitivity to subtle facial changes, visual artifacts can lead to inaccurate facial details and insufficient preservation of identity in the generated videos. Therefore, we adopt a patch-level image encoder [37] as the backbone of the appearance encoder and introduce an ID-aware multi-scale adapter (IMAdapter) to enhance the model's ID consistency capabilities. To process the entire image, we first extract information related to clothing and background elements of the portrait. For the enhancement of consistency, as illustrated in Figure 2, we utilize the ID encoder to extract identity-related features. We further propose extracting multi-scale, fine-grained identity features using adapters." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.554, + 0.483, + 0.675 + ], + "angle": 0, + "content": "As illustrated in Figure 3, we first use the linear projection to downsample the visual features. The low-rank features are subsequently fed into convolutional modules at multiple scales, along with a cross-attention mechanism, which captures ID-aware fine-grained details for the preservation of visual identity. The cross-modal features are then linearly projected back to the original dimensions to be merged with patch-level features:" + }, + { + "type": "equation", + "bbox": [ + 0.201, + 0.687, + 0.483, + 0.769 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\hat {f} _ {a} = f _ {a} \\mathbf {W} _ {d o w n} ^ {v}, \\\\ \\hat {f} _ {c} = \\operatorname {M C o n v} (\\hat {f} _ {a}), \\tag {5} \\\\ \\bar {f} _ {a} = \\mathrm {M H C A} (f _ {i d}, \\hat {f} _ {c}, \\hat {f} _ {c}), \\\\ f _ {a} = f _ {a} + \\bar {f} _ {a} \\mathbf {W} _ {u p} ^ {v}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.901 + ], + "angle": 0, + "content": "where \\( f_{id} \\) represents the ID features, which are utilized as the query features, while \\( \\hat{f}_c \\) serves as the key and value features. MConv signifies the multiple convolutions that process appearance features \\( f_a \\) in parallel and concatenate them along the channel dimension. During the training phase, the sampling of ID information comes from a random frame in the video sequence; in the inference stage, it comes from the reference image." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.091, + 0.809, + 0.108 + ], + "angle": 0, + "content": "3.5. Training and Inference Strategies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.113, + 0.908, + 0.659 + ], + "angle": 0, + "content": "For adaptation to various anime styles, we utilize AnimeGANv3 [31] to perform style conversion on the dataset. This allows the model to adjust to different styles of driving videos and source images. Instead of directly training models with AnimeGANv3, we employ its style transfer capability to convert facial portrait datasets into anime-stylized representations. These transformed data samples are then integrated into our training dataset for enhanced feature learning. To mitigate the impact of skin color and other color-related information on facial movements, we employ color jittering to augment the data for the facial cropping input of the motion encoder. For maintaining the stability of the regions beyond the face in the generated video, we utilize DWPose [58] as an additional spatial conditioning signal. Specifically, we employ the pose-guided network [19] to achieve spatial control condition mapping. We further incorporate spatial conditional control into the latent features of the first spatial convolutional layer in U-Net through addition. To simulate scenarios in which certain skeletons from the driving video are absent compared to those in the source image during the inference stage, we implement an augmentation strategy [48] for the pose sequence during training. This strategy randomly removes certain edges throughout the training process to improve the stability of spatial control. This mitigates the effects of detector accuracy limitations on detector-based control conditions and offers flexible control capability. During inference, considering that there are differences in the body shape and spatial position of bones between the source image and the driving video, we first translate the position based on the offset of the center point of the face (e.g., the position of the nose tip) and scale it according to the length ratio of the skeletons. We find that the key points of the eyes can lead to subtle variations in face shape. Therefore, in the setting of cross reenactment, we do not utilize the key points and corresponding edges of the eyes." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.672, + 0.646, + 0.689 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.697, + 0.707, + 0.714 + ], + "angle": 0, + "content": "4.1. Experiments Setting" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We train our model using HDTF [65], VFHQ [52], VoxCeleb2 [5], CelebV-HQ [68], and our own collected dataset. This collected dataset comprises monocular camera recordings showcasing 20 hours of authentic human video footage obtained from 200 subjects. The recordings were captured in both indoor and outdoor environments. During the training process, all of the images and videos are resized to \\(512 \\times 512\\). For the optimization of our framework, we employ the AdamW optimizer with a learning rate of \\(1 \\times 10^{-5}\\). The weights of the ID Encoder [6], Variational Autoencoder [24] from Stable Video Diffusion [2] and the blocks of DiNOv2 [37] are fixed. In the motion memory bank, we" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15913" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.119 + ], + "angle": 0, + "content": "Table 1. Quantitative results on \\(512 \\times 512\\) test images. LMD multiplied by \\(10^{-3}\\), AED multiplied by \\(10^{-2}\\), APD multiplied by \\(10^{-3}\\) and Identity Similarity multiplied by \\(10^{-1}\\). \\(\\uparrow\\) indicates higher is better. \\(\\downarrow\\) indicates lower is better. The best results are in bold." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.128, + 0.905, + 0.233 + ], + "angle": 0, + "content": "
MethodSelf ReenactmentCross Reenactment
LMD ↓FID-VID ↓FVD ↓PSNR ↑SSIM ↑LPIPS ↓AED ↓APD ↓Identity ↑ SimilarityFacial ↑ MovementsVideo ↑ QualityTemporal ↑ Smoothness
LivePortrait [13]9.1482.71483.3831.410.720.2226.8326.978.713.004.114.06
AntiPortrait [51]6.6781.90430.2430.540.670.2722.4221.327.953.834.003.81
FollowYE [34]5.6377.15417.5330.840.680.2520.0620.958.184.114.233.33
X-Potrait [53]6.2382.93416.4130.810.710.1920.6620.398.033.883.733.49
HunyuanPortrait (Ours)2.0275.81333.4832.980.810.1119.4519.208.874.554.694.61
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.259, + 0.482, + 0.397 + ], + "angle": 0, + "content": "incorporate 64 learnable motion memories, each with a dimension of 768, which share parameters across the various blocks. They are initialized using a scaled normal distribution, where each element is sampled from \\(\\mathcal{N}(0,\\frac{1}{\\sqrt{\\mathrm{hidden\\_dim}}})\\). To ensure stability, we apply gradient norm clipping at a value of 0.99. During inference, we employ the DDIM sampler [43] and set the scale of classifier-free guidance [16] to 2.0 for our experiment. The framework is trained on 128 NVIDIA A100 GPUs for three days." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.408, + 0.188, + 0.423 + ], + "angle": 0, + "content": "4.2. Metrics" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.43, + 0.485, + 0.839 + ], + "angle": 0, + "content": "We employ both qualitative and quantitative analyses to assess the generated video quality and motion accuracy of our portrait animation results. For self-reenactment, we utilize the Fréchet Inception Distance (FID) [15], the Fréchet Video Distance (FVD) [46], Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM) [50], Learned Perceptual Image Patch Similarity (LPIPS) [63] and the Landmark Mean Distances (LMD) [32]. For cross-reenactment, we utilize Average Expression Distance (AED) [42], Average Pose Distance (APD) [42] and the ArcFace Score [6] as the identity (ID) similarity. Detailed descriptions of these metrics are provided in supplementary materials. Considering the absence of ground truth and the significant error between current evaluation methods and human perception, we combine the results of user studies that evaluates cross-reenactment from three additional dimensions: facial movements, video quality, and temporal smoothness. The facial movements encompass not only facial expressions but also the overall orientation of the generated portrait head. The quality of the video and its temporal smoothness are primarily assessed by the clarity of the video and the smoothness of the transitions between frames. For evaluation, we utilize source images and driving videos from two prominent datasets [5, 52]. This experiment aims to evaluate the model's ability to transfer expressions and movements from one subject to another while preserving the identity of the source portrait." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.849, + 0.418, + 0.866 + ], + "angle": 0, + "content": "4.3. Quantitative and Qualitative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.903 + ], + "angle": 0, + "content": "As shown in Table 1, our method attains the lowest FIDVID (average FID of frames in videos) and FVD, with a" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.259, + 0.908, + 0.834 + ], + "angle": 0, + "content": "significant advantage when compared to the previously top-performing method, which underscores the superior capability of HunyuanPortrait. We also find that non-diffusion-based methods have much lower generation quality compared to diffusion-based methods. Although the video resolution is the same, the generated portrait details have significant differences. In more challenging cross-reenactment scenarios, our approach demonstrates strong generalization capabilities, leading to substantial improvements across all metrics. This fully demonstrates the advantage of our method, as cross-reenactment is much more difficult than self-reenactment and is closer to real-life application scenarios. The warping-based method LivePortrait struggles to accurately model overall head movement, resulting in generated portraits that exhibit minimal variation in head pose relative to the reference image. Compared with other methods, our method has stronger ID similarity, facial dynamic generation ability, video quality, and temporal smoothness. Our method employs implicit motion representation rather than explicit facial keypoints to control the generation of facial expressions. Our method further takes into account the intensity of motion, the ability to perceive identity, and complementary information based on temporal context. This approach mitigates identity distortion caused by variations in facial geometries and effectively captures dynamic facial details that explicit keypoints cannot represent. Meanwhile, since our method is based on stable video diffusion, the generated results are smoother compared to previous image-based diffusion methods and have higher video quality. These enhancements highlight the superior accuracy and fidelity of our method in capturing fine-grained details, affirming its efficacy in synthesizing realistic portrait motions. Furthermore, due to the lack of open-source training code, combined with the challenges related to their motion module and the frame rate of the videos, FollowYourEmoji and X-Portrait shows suboptimal performance in terms of temporal smoothness. Our method remains unaffected by variations in frame rate." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.847, + 0.652, + 0.862 + ], + "angle": 0, + "content": "4.4. Visualization" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.903 + ], + "angle": 0, + "content": "As illustrated in Figure 4, our method can capture more detailed and consistent portrait movements with the driv" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15914" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.088, + 0.907, + 0.637 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.182, + 0.648, + 0.816, + 0.661 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparisons of self-reenactment and cross-reenactment with state-of-the-art methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.486, + 0.901 + ], + "angle": 0, + "content": "ing video. For self-reenactment, our method generates results that enable more detailed capture of facial dynamics, including eye gaze direction, eye rotation, and lip synchronization. In addition, our method can also achieve tracking of the overall rotation of the head, which cannot be achieved by previous top-performing warping-based method LivePortrait, such as LivePortrait. Current state-of-the-art diffusion-based methods have significant defects in maintaining the portrait's identity and facial shape due to the deviation of predicted keypoints during head rotation. In contrast, our method effectively preserves the identity of the portrait and captures intricate facial dynamics. Previous methods are influenced by the differences in facial geometries between the driving video and the reference image," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.909, + 0.901 + ], + "angle": 0, + "content": "leading to substantial challenges in preserving the identity of the generated video. This is a result of inadequate disentanglement of their appearance features and facial dynamics. These issues result in problems such as blurry videos and distorted portraits. Our framework effectively addresses the disentanglement of the two key elements involved in portrait animation, thereby demonstrating a strong generalization ability, as shown in Figure 4. The diffusion-based method exhibits superior generation quality; however, discrepancies in facial keypoints due to variations in facial shape, combined with the limited ability of these keypoints to accurately capture dynamic facial details, make the previous state-of-the-art method less effective than ours in generating features such as facial geometry, eye gaze direc" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15915" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.482, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.101, + 0.362, + 0.472, + 0.377 + ], + "angle": 0, + "content": "Figure 5. Qualitative ablation studies of different components." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.385, + 0.484, + 0.581 + ], + "angle": 0, + "content": "tion, and lip synchronization in portraits. Due to variations in facial shapes, methods that rely on explicit keypoints are influenced by the facial characteristics of individuals in the driving video. This leads to insufficient separation between motion and identity. Furthermore, since their approach utilizes the image generation model Stable Diffusion as its foundation, it also exhibits inferior performance compared to ours in terms of temporal smoothness. Building upon stable video diffusion and utilizing our meticulously designed fine-grained feature extractors, our method exhibits exceptional performance in identity preservation, video quality enhancement, and the capture of subtle facial movements, resulting in more vivid and smoother generated videos." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.592, + 0.245, + 0.607 + ], + "angle": 0, + "content": "4.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.483, + 0.84 + ], + "angle": 0, + "content": "Effect of Motion Memory Bank. We validate the effectiveness of the motion memory bank. As shown in Table 2, the incorporation of the motion memory bank results in higher-quality videos, demonstrating improvements in FID-VID and FVD, as well as enhanced facial movements. We also observe an improvement in smoothness. Through the utilization of motion memories, our methodology demonstrates context-based perception capabilities, thereby enhancing motion features in the presence of ambiguous information. By integrating memories with information from preceding and subsequent frames, the memory bank can provide more detailed information into facial dynamics. As illustrated in Figure 5, the forehead wrinkles associated with the elevation of the eyes become more pronounced, when the character's eyebrows are raised." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Effect of Intensity-Aware Motion Encoder (IAME). We confirm the significant role of the intensity-aware motion encoder. As illustrated in Table 2, the incorporation of this component, in contrast to the direct utilization of coarse" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.089, + 0.907, + 0.133 + ], + "angle": 0, + "content": "Table 2. Ablation study to validate the effectiveness of different components. LMD (Landmark Mean Distance) multiplied by \\(10^{-3}\\) and Identity Similarity multiplied by \\(10^{-1}\\)." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.143, + 0.905, + 0.248 + ], + "angle": 0, + "content": "
SettingIdentity ↑ SimilarityFID-VID ↓FVD ↓LMD ↓
HunyuanPortrait8.8775.81333.482.02
- Memory Bank8.7578.43361.942.78
- IAME8.6380.79385.434.01
- ID Features8.2175.03330.121.98
- IMAdapter8.0977.14352.672.54
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.274, + 0.907, + 0.424 + ], + "angle": 0, + "content": "motion features derived from the pre-trained motion encoder [8], offers critical guidance and intricate details regarding the facial dynamics associated with portrait motions. This enhancement significantly elevates the overall quality of the generated video, evidenced by an improvement in FID-VID, FVD, and ID similarity. As illustrated in Figure 5, the absence of feature refinement in the IAME may result in inaccuracies in the micro-expressions of the characters, including the direction of their gaze. Furthermore, the intricate details of the face are neglected." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.425, + 0.909, + 0.591 + ], + "angle": 0, + "content": "Effect of ID features and IMAdapter. To validate the effectiveness of the IMAdapter, we first remove the ID features from the IMAdapter, leaving only multiple convolution kernels for low-rank feature processing. As shown in Table 2 and Figure 5, the ablation results demonstrate that incorporating ID features is critical in enhancing the model's ID consistency ability. We further remove IMAdapter from the appearance extractor's blocks. As demonstrated in Figure 5, the incorporation of IMAdapter provides more fine-grained features related to facial texture, geometry, and other facial characteristics." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.604, + 0.634, + 0.62 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In this paper, we present an enhanced portrait animation framework that addresses the limitations of previous approaches. Our model leverages implicit representations with stable video diffusion to achieve temporal consistency and precise control over facial dynamics. The proposed model demonstrates significant improvements in the overall quality and temporal consistency of generated videos, allowing for more natural and lifelike animations. We offer a robust framework for generating high-quality, controllable animations, which achieve high disentanglement. Our method not only enhances the visual experience but also opens up possibilities for immersive content creation. With continued research and development, this approach holds great potential for applications in virtual reality, gaming, and human-computer interaction. However, as with any technology that generates realistic human likenesses, it is essential to consider the ethical implications. Future work should also develop safeguards to ensure responsible use." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15916" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.186, + 0.484, + 0.255 + ], + "angle": 0, + "content": "[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.256, + 0.483, + 0.311 + ], + "angle": 0, + "content": "[3] Zhiyuan Chen, Jiajiong Cao, Zhiquan Chen, Yuming Li, and Chenguang Ma. Echomimic: Lifelike audio-driven portrait animations through editable landmark conditions. arXiv preprint arXiv:2407.08136, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.312, + 0.483, + 0.382 + ], + "angle": 0, + "content": "[4] Kun Cheng, Xiaodong Cun, Yong Zhang, Menghan Xia, Fei Yin, Mingrui Zhu, Xuan Wang, Jue Wang, and Nannan Wang. Videoretalking: Audio-based lip synchronization for talking head video editing in the wild. In SIGGRAPH Asia 2022, pages 1-9, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.383, + 0.483, + 0.424 + ], + "angle": 0, + "content": "[5] Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. Voxceleb2: Deep speaker recognition. arXiv preprint arXiv:1806.05622, 2018. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.425, + 0.483, + 0.493 + ], + "angle": 0, + "content": "[6] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.494, + 0.483, + 0.536 + ], + "angle": 0, + "content": "[7] Michail Christos Doukas, Stefanos Zafeiriou, and Viktoriia Sharmanska. Headgan: One-shot neural head synthesis and editing, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.537, + 0.483, + 0.606 + ], + "angle": 0, + "content": "[8] Nikita Drobyshev, Jenya Chelishev, Taras Khakhulin, Aleksei Ivakhnenko, Victor Lempitsky, and Egor Zakharov. Megaportraits: One-shot megapixel neural head avatars. In Proceedings of the 30th ACM International Conference on Multimedia, pages 2663-2671, 2022. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.607, + 0.483, + 0.661 + ], + "angle": 0, + "content": "[9] Nikita Drobyshev, Jenya Chelishev, Taras Khakhulin, Aleksei Ivakhnenko, Victor Lempitsky, and Egor Zakharov. Megaportraits: One-shot megapixel neural head avatars, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.663, + 0.483, + 0.718 + ], + "angle": 0, + "content": "[10] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.719, + 0.483, + 0.801 + ], + "angle": 0, + "content": "[11] Yuming Gu, Hongyi Xu, You Xie, Guoxian Song, Yichun Shi, Di Chang, Jing Yang, and Linjie Luo. Diffportrait3d: Controllable diffusion for zero-shot portrait view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10456-10465, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.483, + 0.87 + ], + "angle": 0, + "content": "[12] Hanzhong Guo, Hongwei Yi, Daquan Zhou, Alexander William Bergman, Michael Lingelbach, and Yizhou Yu. Real-time one-step diffusion-based expressive portrait videos generation. arXiv preprint arXiv:2412.13479, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.872, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[13] Jianzhu Guo, Dingyun Zhang, Xiaogiang Liu, Zhizhou Zhong, Yuan Zhang, Pengfei Wan, and Di Zhang. Livepor" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "trait: Efficient portrait animation with stitching and retargeting control. arXiv preprint arXiv:2407.03168, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[14] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. In The Twelfth International Conference on Learning Representations, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[15] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.905, + 0.305 + ], + "angle": 0, + "content": "[16] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.306, + 0.905, + 0.374 + ], + "angle": 0, + "content": "[17] Fa-Ting Hong and Dan Xu. Implicit identity representation conditioned memory compensation network for talking head video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23062-23072, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.376, + 0.905, + 0.432 + ], + "angle": 0, + "content": "[18] Fa-Ting Hong, Li Shen, and Dan Xu. Dagan++: Depth-aware generative adversarial network for talking head video generation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.907, + 0.489 + ], + "angle": 0, + "content": "[19] Li Hu. Animate anyone: Consistent and controllable image-to-video synthesis for character animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8153-8163, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.905, + 0.559 + ], + "angle": 0, + "content": "[20] Xianliang Huang, Jiajie Gou, Shuhang Chen, Zhizhou Zhong, Jihong Guan, and Shuigeng Zhou. Iddr-ngp: Incorporating detectors for distractors removal with instant neural radiance field. In Proceedings of the 31st ACM International Conference on Multimedia, pages 1343-1351, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.561, + 0.905, + 0.615 + ], + "angle": 0, + "content": "[21] Xiaoyu Jin, Zunnan Xu, Mingwen Ou, and Wenming Yang. Alignment is all you need: A training-free augmentation strategy for pose-guided video generation. arXiv preprint arXiv:2408.16506, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.617, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[22] Taras Khakhulin, Vanessa Sklyarova, Victor Lempitsky, and Egor Zakharov. Realistic one-shot mesh-based head avatars. In European Conference on Computer Vision, pages 345-362. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.674, + 0.905, + 0.702 + ], + "angle": 0, + "content": "[23] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.703, + 0.905, + 0.731 + ], + "angle": 0, + "content": "[24] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.732, + 0.905, + 0.799 + ], + "angle": 0, + "content": "[25] Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, et al. Hunyuanvideo: A systematic framework for large video generative models. arXiv preprint arXiv:2412.03603, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[26] Hongxiang Li, Meng Cao, Xuxin Cheng, Yaowei Li, Zhihong Zhu, and Yuexian Zou. G2l: Semantically aligned and uniform video grounding via geodesic and game theory. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12032-12042, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[27] Hongxiang Li, Yaowei Li, Yuhang Yang, Junjie Cao, Zhihong Zhu, Xuxin Cheng, and Long Chen. Dispose: Disen" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15917" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "tangling pose guidance for controllable human image animation. arXiv preprint arXiv:2412.09349, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[28] Ronghui Li, Junfan Zhao, Yachao Zhang, Mingyang Su, Zeping Ren, Han Zhang, Yansong Tang, and Xiu Li. Finedance: A fine-grained choreography dataset for 3d full body dance generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10234-10243, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.483, + 0.276 + ], + "angle": 0, + "content": "[29] Ronghui Li, Hongwen Zhang, Yachao Zhang, Yuxiang Zhang, Youliang Zhang, Jie Guo, Yan Zhang, Xiu Li, and Yebin Liu. Lodge++: High-quality and long dance generation with vivid choreography patterns. arXiv preprint arXiv:2410.20389, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.361 + ], + "angle": 0, + "content": "[30] Ronghui Li, YuXiang Zhang, Yachao Zhang, Hongwen Zhang, Jie Guo, Yan Zhang, Yebin Liu, and Xiu Li. Lodge: A coarse to fine diffusion network for long dance generation guided by the characteristic dance primitives. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1524-1534, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[31] Gang LIU, Xin CHEN, and Zhixiang GAO. A novel double-tail generative adversarial network for fast photo animation. IEICE Transactions on Information and Systems, E107.D(1): 72-82, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.419, + 0.482, + 0.489 + ], + "angle": 0, + "content": "[32] Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuoling Chang, Ming Guang Yong, Juhyun Lee, et al. Mediapipe: A framework for building perception pipelines. arXiv preprint arXiv:1906.08172, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.482, + 0.56 + ], + "angle": 0, + "content": "[33] Yue Ma, Yingqing He, Xiaodong Cun, Xintao Wang, Siran Chen, Xiu Li, and Qifeng Chen. Follow your pose: Pose-guided text-to-video generation using pose-free videos. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4117-4125, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.482, + 0.63 + ], + "angle": 0, + "content": "[34] Yue Ma, Hongyu Liu, Hongfa Wang, Heng Pan, Yingqing He, Junkun Yuan, Ailing Zeng, Chengfei Cai, Heung-Yeung Shum, Wei Liu, et al. Follow-your-emoji: Fine-controllable and expressive freestyle portrait animation. arXiv preprint arXiv:2406.01900, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.632, + 0.482, + 0.716 + ], + "angle": 0, + "content": "[35] Yuxi Mi, Zhizhou Zhong, Yuge Huang, Jiazhen Ji, Jianqing Xu, Jun Wang, Shaoming Wang, Shouhong Ding, and Shuigeng Zhou. Privacy-preserving face recognition using trainable feature subtraction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 297-307, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.483, + 0.773 + ], + "angle": 0, + "content": "[36] Yuval Nirkin, Yosi Keller, and Tal Hassner. Fsgan: Subject agnostic face swapping and reenactment. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7184-7193, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.482, + 0.844 + ], + "angle": 0, + "content": "[37] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[38] Bohao Peng, Jian Wang, Yuechen Zhang, Wenbo Li, MingChang Yang, and Jiaya Jia. Controlnext: Powerful and efficient control for image and video generation. arXiv preprint arXiv:2408.06070, 2024. 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.137, + 0.907, + 0.22 + ], + "angle": 0, + "content": "[40] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.222, + 0.905, + 0.277 + ], + "angle": 0, + "content": "[41] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. Advances in neural information processing systems, 32, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.279, + 0.905, + 0.336 + ], + "angle": 0, + "content": "[42] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. In Advances in Neural Information Processing Systems, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.338, + 0.905, + 0.379 + ], + "angle": 0, + "content": "[43] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.382, + 0.905, + 0.436 + ], + "angle": 0, + "content": "[44] Jiale Tao, Shuhang Gu, Wen Li, and Lixin Duan. Learning motion refinement for unsupervised face animation. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.439, + 0.905, + 0.495 + ], + "angle": 0, + "content": "[45] Jiayuan Tian, Yanling LONG, Hong YANG, Huifang WU, Peng XUE, and Zhongqing JIANG. Cues of eye region and their effects on face-personality perception. Advances in Psychological Science, 30(12):2735, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.497, + 0.905, + 0.539 + ], + "angle": 0, + "content": "[46] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphaël Marinier, Marcin Michalski, and Sylvain Gelly. Fvd: A new metric for video generation. 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.541, + 0.905, + 0.598 + ], + "angle": 0, + "content": "[47] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems, 30, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.6, + 0.905, + 0.668 + ], + "angle": 0, + "content": "[48] Cong Wang, Kuan Tian, Jun Zhang, Yonghang Guan, Feng Luo, Fei Shen, Zhiwei Jiang, Qing Gu, Xiao Han, and Wei Yang. V-express: Conditional dropout for progressive training of portrait video generation. arXiv preprint arXiv:2406.02511, 2024. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.671, + 0.905, + 0.728 + ], + "angle": 0, + "content": "[49] Suzhen Wang, Lincheng Li, Yu Ding, and Xin Yu. One-shot talking face generation from single-speaker audio-visual correlation learning. In AAAI Conference on Artificial Intelligence, pages 2531–2539, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.729, + 0.905, + 0.784 + ], + "angle": 0, + "content": "[50] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.787, + 0.905, + 0.828 + ], + "angle": 0, + "content": "[51] Huawei Wei, Zejun Yang, and Zhisheng Wang. Aniportrait: Audio-driven synthesis of photorealistic portrait animations, 2024. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[52] Liangbin Xie, Xintao Wang, Honglun Zhang, Chao Dong, and Ying Shan. Vfhq: A high-quality dataset and benchmark for video face super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 657-666, 2022. 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "15918" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.148 + ], + "angle": 0, + "content": "[53] You Xie, Hongyi Xu, Guoxian Song, Chao Wang, Yichun Shi, and Linjie Luo. X-portrait: Expressive portrait animation with hierarchical motion attention. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.204 + ], + "angle": 0, + "content": "[54] Jingjing Xu, Xu Sun, Zhiyuan Zhang, Guangxiang Zhao, and Junyang Lin. Understanding and improving layer normalization. Advances in neural information processing systems, 32, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.206, + 0.482, + 0.275 + ], + "angle": 0, + "content": "[55] Sicheng Xu, Guojun Chen, Yu-Xiao Guo, Jiaolong Yang, Chong Li, Zhenyu Zang, Yizhong Zhang, Xin Tong, and Baining Guo. Vasa-1: Lifelike audio-driven talking faces generated in real time. arXiv preprint arXiv:2404.10667, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.482, + 0.348 + ], + "angle": 0, + "content": "[56] Zunnan Xu, Yukang Lin, Haonan Han, Sicheng Yang, Ronghui Li, Yachao Zhang, and Xiu Li. Mambatak: Efficient holistic gesture synthesis with selective state space models. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.348, + 0.482, + 0.404 + ], + "angle": 0, + "content": "[57] Shurong Yang, Huadong Li, Juhao Wu, Minhao Jing, Linze Li, Renhe Ji, Jiajun Liang, and Haoqiang Fan. Megactor: Harness the power of raw video for vivid portrait animation. arXiv preprint arXiv:2405.20851, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.482, + 0.462 + ], + "angle": 0, + "content": "[58] Zhendong Yang, Ailing Zeng, Chun Yuan, and Yu Li. Effective whole-body pose estimation with two-stages distillation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4210-4220, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.482, + 0.517 + ], + "angle": 0, + "content": "[59] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.482, + 0.589 + ], + "angle": 0, + "content": "[60] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. Styleheat: One-shot high-resolution editable talking face generation via pre-trained stylegan. In European Conference on Computer Vision, pages 85–101, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.482, + 0.645 + ], + "angle": 0, + "content": "[61] Wangbo Yu, Yanbo Fan, Yong Zhang, Xuan Wang, Fei Yin, Yunpeng Bai, Yan-Pei Cao, Ying Shan, Yang Wu, Zhongqian Sun, and Baoyuan Wu. Nofa: Nerf-based one-shot facial avatar reconstruction, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.482, + 0.731 + ], + "angle": 0, + "content": "[62] Bowen Zhang, Chenyang Qi, Pan Zhang, Bo Zhang, Hsiang-Tao Wu, Dong Chen, Qifeng Chen, Yong Wang, and Fang Wen. Metaportrait: Identity-preserving talking head generation with fast personalized adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22096-22105, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.731, + 0.482, + 0.801 + ], + "angle": 0, + "content": "[63] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.802, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[64] Yang Zhang, Jiaxi Gu, Li-Wen Wang, Han Wang, Junqi Cheng, Yuefeng Zhu, and Fangyuan Zou. Mimic motion: High-quality human motion video generation with confidence-aware pose guidance. arXiv preprint arXiv:2406.19680, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.872, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[65] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.092, + 0.908, + 0.135 + ], + "angle": 0, + "content": "a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661-3670, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.907, + 0.19 + ], + "angle": 0, + "content": "[66] Jian Zhao and Hui Zhang. Thin-plate spline motion model for image animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3657–3666, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.907, + 0.261 + ], + "angle": 0, + "content": "[67] Longtao Zheng, Yifan Zhang, Hanzhong Guo, Jiachun Pan, Zhenxiong Tan, Jiahao Lu, Chuanxin Tang, Bo An, and Shuicheng Yan. Memo: Memory-guided diffusion for expressive talking video generation. arXiv preprint arXiv:2412.04448, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.262, + 0.907, + 0.331 + ], + "angle": 0, + "content": "[68] Hao Zhu, Wayne Wu, Wentao Zhu, Liming Jiang, Siwei Tang, Li Zhang, Ziwei Liu, and Chen Change Loy. Celebvehq: A large-scale video facial attributes dataset. In European conference on computer vision, pages 650-667. Springer, 2022. 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15919" + } + ] +] \ No newline at end of file diff --git a/2025/HunyuanPortrait_ Implicit Condition Control for Enhanced Portrait Animation/a6cc1e88-c49b-4877-897d-6a6632aadecd_origin.pdf b/2025/HunyuanPortrait_ Implicit Condition Control for Enhanced Portrait Animation/a6cc1e88-c49b-4877-897d-6a6632aadecd_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5ce692b89afdac03abbaf102fab5379922f8ca69 --- /dev/null +++ b/2025/HunyuanPortrait_ Implicit Condition Control for Enhanced Portrait Animation/a6cc1e88-c49b-4877-897d-6a6632aadecd_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:91e10a1765ca518264654ad38e9ada3c009c70737b72241660eddd27e3133577 +size 6053021 diff --git a/2025/HunyuanPortrait_ Implicit Condition Control for Enhanced Portrait Animation/full.md b/2025/HunyuanPortrait_ Implicit Condition Control for Enhanced Portrait Animation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..580ea030f5c62b66b955542ddd0498e184275466 --- /dev/null +++ b/2025/HunyuanPortrait_ Implicit Condition Control for Enhanced Portrait Animation/full.md @@ -0,0 +1,245 @@ +# HunyuanPortrait: Implicit Condition Control for Enhanced Portrait Animation + +Zunnan Xu $^{1,2*}$ , Zhentao Yu $^{2}$ , Zixiang Zhou $^{2}$ , Jun Zhou $^{3*}$ , Xiaoyu Jin $^{1*}$ , Fa-ting Hong $^{4*}$ , Xiaozhong Ji $^{2}$ , Junwei Zhu $^{2}$ , Chengfei Cai $^{2}$ , Shiyu Tang $^{2}$ , Qin Lin $^{2}$ , Xiu Li $^{1\dagger}$ , Qinglin Lu $^{2\dagger}$ , Tsinghua University ${}^{2}$ Hunyuan, Tencent ${}^{3}$ Sun Yat-Sen Univeristy ${}^{4}$ HKUST + +# Abstract + +We introduce HunyuanPortrait, a diffusion-based condition control method that employs implicit representations for highly controllable and lifelike portrait animation. Given a single portrait image as an appearance reference and video clips as driving templates, HunyuanPortrait can animate the character in the reference image by the facial expression and head pose of the driving videos. In our framework, we utilize pre-trained encoders to achieve the decoupling of portrait motion information and identity in videos. To do so, implicit representation is adopted to encode motion information and is employed as control signals in the animation phase. By leveraging the power of stable video diffusion as the main building block, we carefully design adapter layers to inject control signals into the denoising unet through attention mechanisms. These bring spatial richness of details and temporal consistency. HunyuanPortrait also exhibits strong generalization performance, which can effectively disentangle appearance and motion under different image styles. Our framework outperforms existing methods, demonstrating superior temporal consistency and controllability. Our project is available at HunyuanPortrait. + +# 1. Introduction + +In this paper, we aim to animate a single static image into a lifelike video using control signals in implicit representations derived from driving videos. As shown in Figure 1, our primary focus is not only on the high controllability of digital portraits but also on the strong consistency of the characters and background in the reference image. + +Portrait animation has been widely applied in fields such as human-computer interaction [11, 56], virtual reality, and the crafting of digital avatars for the gaming and multimedia industries [28-30]. Previous works [4, 7, 62] develop usable portrait animation models utilizing Generative Adversarial Networks (GANs) [10] based on raw video inputs. However, these approaches exhibit limitations in terms of gener + +![](images/a1ad754f12f89d77499d7c3e7dddd63a2b74498d5679a4605742fef0238e910a.jpg) +Figure 1. Our framework employs implicit condition control to generate portrait animations, demonstrating robust generalization performance with high-fidelity facial dynamics and vivid head poses during cross-reenactment. The animated portraits remain unaffected by variations in facial shape and the spatial position of the driving videos, demonstrating strong identity consistency. + +alization. Given the variability of facial structures and the complexities of facial movements, these methods often face challenges such as artifacts and insufficient control over facial expressions. This is particularly evident when there is a substantial difference in facial shape between the driving video and the source image, resulting in significant issues with facial artifacts and motion distortion. These early methods often produce results with background jitter and blurred portraits during cross-reenactment [36, 41]. + +The advancement of diffusion models has led to a significant focus in recent methodologies [3, 48] on the fine-tuning + +of image-level diffusion models [39] through the incorporation of motion modules [14], thereby facilitating the generation of video content. These methods effectively address the issues associated with GAN-based approaches in managing complex backgrounds and enhancing generalization. However, even with the incorporation of motion modules, these methods are susceptible to defects during the generation stage (e.g., temporal smoothness and flexibility to different frame rates). This vulnerability arises from the absence of motion pre-training in the underlying models during the pre-training phase. Additionally, these methods remain constrained to the explicit control of facial expressions. The effectiveness of these methods, which depend on explicit keypoints to control facial expressions, not only results in the loss of intricate details but also significantly relies on the accuracy of the landmark extraction techniques employed. Furthermore, as a result of the variability in facial shapes, these methods are limited by post-processing alignment strategies and often face challenges in adapting to diverse facial geometries, which can lead to inaccurate facial control and insufficient preservation of identity in the visual effects of videos. Due to the human eye's sensitivity to subtle facial changes [45], it is crucial to preserve the fidelity of facial details and the smoothness of the videos. + +Therefore, we propose an implicit conditional control model to tackle the challenges posed by diverse facial geometries and intricate expression details. We utilize stable video diffusion as the foundational model and integrate identity and motion information through appearance and motion attention. This approach eliminates the need to fine-tune the image diffusion model and to train the motion module separately. Firstly, we coarsely decouple the identity information from the motion information in the original video by using a pre-trained motion encoder [9]. Considering that this pre-trained encoder has not completely decoupled identity from motion information, we further implement enhanced training strategies and improved network architectures to strengthen the control of motion features and the separation of portrait identities. For the adaptation of temporal modeling tasks in video generation using diffusion models, along with the complementary contextual information present between different frames, we propose the incorporation of a motion memory bank to enhance the implicit representation of motion features. Due to the substantial variations in the degree of blur and pixel distortion in video frames at different levels of motion intensity, we propose an intensity-aware motion encoder. This encoder aims to improve the capacity of decoupled motion features to capture intricate details. Additionally, in order to achieve consistent modeling of portrait identity and background areas in videos, we combined ArcFace [6] with the DiNOv2 [37] backbone to develop an enhanced appearance encoder. To preserve the pre-training knowledge of both models, we + +froze their original parameters and proposed an ID-aware Multi-scale adapter (IMAdapter). This approach achieves high-fidelity modeling of the intricate details in portraits. Our main contributions can be summarized as follows: + +- We propose the first implicit condition animation framework based on stable video diffusion, which significantly improves temporal consistency and achieves high fidelity and precise control for portrait animation. +- We develop a series of strategies to enhance identity consistency and improve the accuracy of portrait motion, including data augmentation, enhanced training and inference strategies, and improved network architecture. +- Experiments demonstrate the effectiveness of our proposed method, which achieves state-of-the-art performance in both subjective and objective metrics. + +# 2. Related Work + +Human portrait animation aims to bring a still image to life by utilizing a driving source, such as a sequence of facial landmarks or frames that contain the head of the portrait. Recent advancements in the field of neural avatars can be divided into two subcategories: non-diffusion-based and diffusion-based methods. Non-diffusion-based methods [13, 17, 18, 26, 42, 49, 55, 62, 66] are renowned for their ability to achieve realistic facial animations and fidelity to motion. Many works employed predetermined motion representations that are popular in the literature, such as blendshapes from 3D Morphable Models (3DMM) [7, 20, 22, 35, 44, 60, 61]. However, these works often display significant artifacts when there is a large disparity between the driving portrait and the source image. Due to the use of a warping-based strategy, these methods are also ineffective at capturing videos with significant head and body movements. With the development of diffusion models [25], recent methods [12, 51, 53, 57] achieve human portrait animation by fine-tuning the stable diffusion model [39]. Several studies have investigated the generation of facial expressions through explicit keypoint control [3, 48, 67]. This method shows great promise for enhancing the quality of generated videos and enables precise control of facial expressions through the extraction of facial landmarks. However, given the diversity of facial geometries, there are notable differences in the distribution of facial keypoints among individuals. The alignment processes used in these methods often struggle to manage boundary cases effectively, resulting in identification offsets due to inaccurate alignment. In addition, the reliance on image-based generative models and separately trained motion modules to achieve temporal consistency results in shortcomings related to temporal smoothness. In the pursuit of improved temporal continuity, pioneering work [21, 27, 38, 64] has employed stable video diffusion models to construct end-to-end diffusion models for human video generation. However, these methods are pri + +![](images/de615433cfc9cb796a91735626fead6552017e81571f987ea3475810ddc42e85.jpg) + +![](images/e95f8eea86262ba3c83d889629e2fe3c0687b973cf8f82c8ba08b023aa2dc88e.jpg) +Figure 2. Our framework utilizes implicit representation to encode motion information as control signals. By harnessing the capabilities of stable video diffusion as the primary building block, we have meticulously designed a fine-grained appearance extractor to maintain the identity of the portrait, along with an intensity-aware motion extractor to capture intricate facial dynamics. + +![](images/c8aef9ebb1feb928dbfbf9f27ab00ef311674bcfa8f72237008d8d3065f82bda.jpg) + +marily designed for full-body movements and do not take into account the preservation of facial identity or the diversity of human facial geometries. As a result, they often lead to significant alterations in the identity of the generated face. Additionally, these methods lack the ability to control subtle facial dynamics, which limits their application in generating portraits that incorporate speech content and facial expressions. In this study, we investigate the strategy of adapting to various face shapes through implicit representations of expression descriptors and achieving capture of facial dynamics. Our method surpasses previous approaches in terms of temporal consistency and generalization ability. + +# 3. Methodology + +# 3.1. Preliminaries + +Diffusion Model. Our framework is built upon diffusion model, as introduced in [39]. The model is designed to facilitate efficient and stable training by conducting the diffusion and denoising processes within a latent space, rather than directly in the image space. The Variational Autoencoder (VAE), as proposed in [23], plays a pivotal role in this process by mapping images from the RGB color space into + +a latent space. This transformation enables the diffusion process to be directed by textual embeddings. + +The process begins with the VAE model, which projects images into latent embeddings. This model comprises two key components: an encoder $\mathcal{E}$ and a decoder $\mathcal{D}$ . The encoder $\mathcal{E}$ compresses the pixel space image $x$ into a latent representation $z = \mathcal{E}(x)$ , while the decoder $\mathcal{D}$ reconstructs it back into an image space, aiming for $\mathcal{D}(z) \approx x$ . Subsequently, a UNet-based architecture [40] is employed for learning the reverse denoising process within the latent space. This network integrates attention mechanisms through Transformer Blocks. The cross-attention mechanism ensures that the textual prompt is effectively integrated throughout the process. The overall training objective for the UNet can be articulated as follows: + +$$ +\mathcal {L} = \underset {t, z, \epsilon} {\mathbb {E}} \left[ \| \epsilon - \epsilon_ {\theta} \left(\sqrt {\bar {\alpha} _ {t}} z + \sqrt {1 - \bar {\alpha} _ {t}} \epsilon , c, t\right) \| ^ {2} \right], \tag {1} +$$ + +where $z$ denotes the latent embedding of the training sample. $\epsilon_{\theta}$ and $\epsilon$ represent the predicted noise by the diffusion model and the ground truth noise at the corresponding timestep $t$ , respectively. $c$ is the condition embedding, and the coefficient $\bar{\alpha}_t$ remains consistent with that employed in diffusion models. + +# 3.2. Overall framework + +As illustrated in Figure 2, building upon stable video diffusion [1], our method is designed with two core components for generating portrait videos: the appearance extractor and the motion extractor. For the detailed appearance modeling of the portrait, we integrate the appearance encoder to manage the subject's identity and background in the generated video. We carefully design an ID-aware multi-scale adapter based on the Arcface encoder [6] and DiNOv2 [37] backbone to enhance visual appearance representation. For the motion extractor, we emphasize the enhancement of the implicit representation of facial dynamics. By utilizing the pre-trained motion encoder [8], we further optimize it by taking into account the intensity and the temporal continuity of motions. This involved the introduction of motion memory banks and an intensity-aware motion encoder to enhance the perception of motion intensity and refine features. These representations are conditionally integrated into the denoising U-Net through attention mechanisms. Our framework also integrates a spatial conditioning signal to ensure the stability of regions beyond the face. By utilizing these features, we generate latent features using a denoising UNet and subsequently produce video through VAEs. + +# 3.3. Intensity-Aware Motion Extractor + +For portrait motion modeling, we concentrate on identity-agnostic portrait animation by utilizing an implicit representation of expression descriptors derived from raw videos, which capture a diverse array of facial dynamics (e.g., lip-sync, micro-expressions, eye gaze, and blinking). This approach differs from existing diffusion-based methods that typically employ explicit landmarks for facial expression representation [33, 51]. Considering that the fidelity of facial motion capture is often compromised by extraneous factors—such as variations in facial shape and background noise that can distort essential motion signals—we crop the central area of the face to focus on the regions most indicative of motion dynamics. The boundary is defined as the area between the eyebrows and the bottom of the mouth. This approach allows us to isolate facial movements and enhance the signal-to-noise ratio. The cropped pixels from driving videos are subsequently utilized as inputs for the pre-trained motion encoder [8] to extract a coarse implicit conditioning signal. Since motion intensity varies and exerts distinct influences on the generated pixels, we introduce an Intensity-Aware Motion Encoder into our framework. This addition enhances the coarse motion features from the motion encoder by adapting to the varying intensities of motion. The intensity is estimated by two dimensions: the degree of distortion in facial expressions and the overall amplitude of head movement. Considering that the expression and head pose intensity rely on landmark positioning and their relative differences, we propose a calcu + +lation method for these variances using landmarks and normalization, which can be formulated as below: + +$$ +\begin{array}{l} e _ {k, j} = l _ {k, j} - l _ {k, c}, k \in [ 1, n ], j \in [ 1, m ] \\ s = \sqrt {\left(l _ {1 , \max } - l _ {1 , \min }\right) ^ {2}}, \tag {2} \\ \end{array} +$$ + +where $l_{k,j}$ denotes the j-th landmark point in the k-th frame and $l_{k,c}$ denotes the center of the landmark point (e.g., nose tip). $n$ and $m$ denote the numbers of frames and points. $s$ represents the face scale derived from the first frame of the landmarks. We further calculate the intensity of expressions $I_e$ and the head poses $I_h$ using the formulation below: + +$$ +I _ {e} = \frac {1}{n \cdot s} \sum_ {k = 1} ^ {n} \sqrt {\frac {1}{m} \sum_ {j = 1} ^ {m} \left(e _ {k , j} - \frac {1}{n} \sum_ {k = 1} ^ {n} e _ {k , j}\right) ^ {2}}, \tag {3} +$$ + +$$ +I _ {h} = \frac {1}{s} \sqrt {\frac {1}{n} \sum_ {k = 1} ^ {n} (l _ {k , c} - \frac {1}{n} \sum_ {k = 1} ^ {n} l _ {k , c}) ^ {2}}. +$$ + +Considering the issue of accuracy in landmark detection, we discretize the calculated continuous values by dividing the intensity into $d$ discrete levels based on the range of values. We set $d$ to 64 and then map these discrete values using embeddings $\mathcal{D}$ to their corresponding intensity feature vectors $E_{s} = \mathrm{Concat}[\mathcal{D}(I_{e}),\mathcal{D}(I_{h})]$ , and concatenate them along the channel dimension to adjust the feature space in response to the dynamic range of motion intensity. In light of the fact that the sequence of motion features is derived from pixel extraction and exhibits a lack of continuity across different contexts, we propose the motion memory bank to improve the context-awareness and temporal modeling capabilities of motion features, which can be formulated as below: + +$$ +\begin{array}{l} \hat {f} _ {m} = \operatorname {A d a L N} \left(F _ {m} + E _ {s}, E _ {s}\right), \\ \bar {f} _ {m} = \operatorname {M H C A} (\hat {m}, \hat {f} _ {m}, \hat {f} _ {m}), \tag {4} \\ \bar {f} _ {m} = \operatorname {A d a L N} \left(\bar {f} _ {m}, E _ {s}\right), \\ \bar {f} _ {m} = \operatorname {M L P} (\bar {f} _ {m}), \\ \end{array} +$$ + +where $F_{m}$ denotes the features extracted by the pre-trained motion encoder, and $\hat{m}$ signifies the query features from the motion memory bank, with $\hat{f}_{m}$ being utilized as the key and value features. The MHCA denotes multi-head cross-attention [47] and the AdaLN denotes adaptive layer normalization [54]. The learnable memories are designed to supplement contextual information, enriching motion features rather than generating new motions. This helps the model capture pixel correlation between different frames, alleviating blurring when some facial parts undergo deformation. The refined motion features are incorporated into the denoising U-Net using cross-attention mechanisms [59]. + +![](images/a8788756699950c2dd20dd00aa70d062cff2beff2fab1d8f8f07e3245c9e4a48.jpg) +Figure 3. The illustration of our ID-aware Multi-scale Adapter (IMAdapter). Here, $\mathbb{C}$ represents the operation of concatenating features along the channel dimension. + +# 3.4. Fine-grained Appearance Extractor + +To achieve appearance control, which encompasses both portrait identity information and background details, we propose an appearance extractor that creates a fine-grained representation of the animated portrait and background. Existing video diffusion models often struggle with identity consistency [38, 64], which results in deficiencies in the method's capacity to generate portraits. Due to the human eye's sensitivity to subtle facial changes, visual artifacts can lead to inaccurate facial details and insufficient preservation of identity in the generated videos. Therefore, we adopt a patch-level image encoder [37] as the backbone of the appearance encoder and introduce an ID-aware multi-scale adapter (IMAdapter) to enhance the model's ID consistency capabilities. To process the entire image, we first extract information related to clothing and background elements of the portrait. For the enhancement of consistency, as illustrated in Figure 2, we utilize the ID encoder to extract identity-related features. We further propose extracting multi-scale, fine-grained identity features using adapters. + +As illustrated in Figure 3, we first use the linear projection to downsample the visual features. The low-rank features are subsequently fed into convolutional modules at multiple scales, along with a cross-attention mechanism, which captures ID-aware fine-grained details for the preservation of visual identity. The cross-modal features are then linearly projected back to the original dimensions to be merged with patch-level features: + +$$ +\begin{array}{l} \hat {f} _ {a} = f _ {a} \mathbf {W} _ {d o w n} ^ {v}, \\ \hat {f} _ {c} = \operatorname {M C o n v} (\hat {f} _ {a}), \tag {5} \\ \bar {f} _ {a} = \mathrm {M H C A} (f _ {i d}, \hat {f} _ {c}, \hat {f} _ {c}), \\ f _ {a} = f _ {a} + \bar {f} _ {a} \mathbf {W} _ {u p} ^ {v}, \\ \end{array} +$$ + +where $f_{id}$ represents the ID features, which are utilized as the query features, while $\hat{f}_c$ serves as the key and value features. MConv signifies the multiple convolutions that process appearance features $f_a$ in parallel and concatenate them along the channel dimension. During the training phase, the sampling of ID information comes from a random frame in the video sequence; in the inference stage, it comes from the reference image. + +# 3.5. Training and Inference Strategies + +For adaptation to various anime styles, we utilize AnimeGANv3 [31] to perform style conversion on the dataset. This allows the model to adjust to different styles of driving videos and source images. Instead of directly training models with AnimeGANv3, we employ its style transfer capability to convert facial portrait datasets into anime-stylized representations. These transformed data samples are then integrated into our training dataset for enhanced feature learning. To mitigate the impact of skin color and other color-related information on facial movements, we employ color jittering to augment the data for the facial cropping input of the motion encoder. For maintaining the stability of the regions beyond the face in the generated video, we utilize DWPose [58] as an additional spatial conditioning signal. Specifically, we employ the pose-guided network [19] to achieve spatial control condition mapping. We further incorporate spatial conditional control into the latent features of the first spatial convolutional layer in U-Net through addition. To simulate scenarios in which certain skeletons from the driving video are absent compared to those in the source image during the inference stage, we implement an augmentation strategy [48] for the pose sequence during training. This strategy randomly removes certain edges throughout the training process to improve the stability of spatial control. This mitigates the effects of detector accuracy limitations on detector-based control conditions and offers flexible control capability. During inference, considering that there are differences in the body shape and spatial position of bones between the source image and the driving video, we first translate the position based on the offset of the center point of the face (e.g., the position of the nose tip) and scale it according to the length ratio of the skeletons. We find that the key points of the eyes can lead to subtle variations in face shape. Therefore, in the setting of cross reenactment, we do not utilize the key points and corresponding edges of the eyes. + +# 4. Experiments + +# 4.1. Experiments Setting + +We train our model using HDTF [65], VFHQ [52], VoxCeleb2 [5], CelebV-HQ [68], and our own collected dataset. This collected dataset comprises monocular camera recordings showcasing 20 hours of authentic human video footage obtained from 200 subjects. The recordings were captured in both indoor and outdoor environments. During the training process, all of the images and videos are resized to $512 \times 512$ . For the optimization of our framework, we employ the AdamW optimizer with a learning rate of $1 \times 10^{-5}$ . The weights of the ID Encoder [6], Variational Autoencoder [24] from Stable Video Diffusion [2] and the blocks of DiNOv2 [37] are fixed. In the motion memory bank, we + +Table 1. Quantitative results on $512 \times 512$ test images. LMD multiplied by $10^{-3}$ , AED multiplied by $10^{-2}$ , APD multiplied by $10^{-3}$ and Identity Similarity multiplied by $10^{-1}$ . $\uparrow$ indicates higher is better. $\downarrow$ indicates lower is better. The best results are in bold. + +
MethodSelf ReenactmentCross Reenactment
LMD ↓FID-VID ↓FVD ↓PSNR ↑SSIM ↑LPIPS ↓AED ↓APD ↓Identity ↑ SimilarityFacial ↑ MovementsVideo ↑ QualityTemporal ↑ Smoothness
LivePortrait [13]9.1482.71483.3831.410.720.2226.8326.978.713.004.114.06
AntiPortrait [51]6.6781.90430.2430.540.670.2722.4221.327.953.834.003.81
FollowYE [34]5.6377.15417.5330.840.680.2520.0620.958.184.114.233.33
X-Potrait [53]6.2382.93416.4130.810.710.1920.6620.398.033.883.733.49
HunyuanPortrait (Ours)2.0275.81333.4832.980.810.1119.4519.208.874.554.694.61
+ +incorporate 64 learnable motion memories, each with a dimension of 768, which share parameters across the various blocks. They are initialized using a scaled normal distribution, where each element is sampled from $\mathcal{N}(0,\frac{1}{\sqrt{\mathrm{hidden\_dim}}})$ . To ensure stability, we apply gradient norm clipping at a value of 0.99. During inference, we employ the DDIM sampler [43] and set the scale of classifier-free guidance [16] to 2.0 for our experiment. The framework is trained on 128 NVIDIA A100 GPUs for three days. + +# 4.2. Metrics + +We employ both qualitative and quantitative analyses to assess the generated video quality and motion accuracy of our portrait animation results. For self-reenactment, we utilize the Fréchet Inception Distance (FID) [15], the Fréchet Video Distance (FVD) [46], Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM) [50], Learned Perceptual Image Patch Similarity (LPIPS) [63] and the Landmark Mean Distances (LMD) [32]. For cross-reenactment, we utilize Average Expression Distance (AED) [42], Average Pose Distance (APD) [42] and the ArcFace Score [6] as the identity (ID) similarity. Detailed descriptions of these metrics are provided in supplementary materials. Considering the absence of ground truth and the significant error between current evaluation methods and human perception, we combine the results of user studies that evaluates cross-reenactment from three additional dimensions: facial movements, video quality, and temporal smoothness. The facial movements encompass not only facial expressions but also the overall orientation of the generated portrait head. The quality of the video and its temporal smoothness are primarily assessed by the clarity of the video and the smoothness of the transitions between frames. For evaluation, we utilize source images and driving videos from two prominent datasets [5, 52]. This experiment aims to evaluate the model's ability to transfer expressions and movements from one subject to another while preserving the identity of the source portrait. + +# 4.3. Quantitative and Qualitative Analysis + +As shown in Table 1, our method attains the lowest FIDVID (average FID of frames in videos) and FVD, with a + +significant advantage when compared to the previously top-performing method, which underscores the superior capability of HunyuanPortrait. We also find that non-diffusion-based methods have much lower generation quality compared to diffusion-based methods. Although the video resolution is the same, the generated portrait details have significant differences. In more challenging cross-reenactment scenarios, our approach demonstrates strong generalization capabilities, leading to substantial improvements across all metrics. This fully demonstrates the advantage of our method, as cross-reenactment is much more difficult than self-reenactment and is closer to real-life application scenarios. The warping-based method LivePortrait struggles to accurately model overall head movement, resulting in generated portraits that exhibit minimal variation in head pose relative to the reference image. Compared with other methods, our method has stronger ID similarity, facial dynamic generation ability, video quality, and temporal smoothness. Our method employs implicit motion representation rather than explicit facial keypoints to control the generation of facial expressions. Our method further takes into account the intensity of motion, the ability to perceive identity, and complementary information based on temporal context. This approach mitigates identity distortion caused by variations in facial geometries and effectively captures dynamic facial details that explicit keypoints cannot represent. Meanwhile, since our method is based on stable video diffusion, the generated results are smoother compared to previous image-based diffusion methods and have higher video quality. These enhancements highlight the superior accuracy and fidelity of our method in capturing fine-grained details, affirming its efficacy in synthesizing realistic portrait motions. Furthermore, due to the lack of open-source training code, combined with the challenges related to their motion module and the frame rate of the videos, FollowYourEmoji and X-Portrait shows suboptimal performance in terms of temporal smoothness. Our method remains unaffected by variations in frame rate. + +# 4.4. Visualization + +As illustrated in Figure 4, our method can capture more detailed and consistent portrait movements with the driv + +![](images/81cdda35b14df4f6649f2d0cc102d583a41bad022873b8cd1eb3888319e71d7b.jpg) +Figure 4. Qualitative comparisons of self-reenactment and cross-reenactment with state-of-the-art methods. + +ing video. For self-reenactment, our method generates results that enable more detailed capture of facial dynamics, including eye gaze direction, eye rotation, and lip synchronization. In addition, our method can also achieve tracking of the overall rotation of the head, which cannot be achieved by previous top-performing warping-based method LivePortrait, such as LivePortrait. Current state-of-the-art diffusion-based methods have significant defects in maintaining the portrait's identity and facial shape due to the deviation of predicted keypoints during head rotation. In contrast, our method effectively preserves the identity of the portrait and captures intricate facial dynamics. Previous methods are influenced by the differences in facial geometries between the driving video and the reference image, + +leading to substantial challenges in preserving the identity of the generated video. This is a result of inadequate disentanglement of their appearance features and facial dynamics. These issues result in problems such as blurry videos and distorted portraits. Our framework effectively addresses the disentanglement of the two key elements involved in portrait animation, thereby demonstrating a strong generalization ability, as shown in Figure 4. The diffusion-based method exhibits superior generation quality; however, discrepancies in facial keypoints due to variations in facial shape, combined with the limited ability of these keypoints to accurately capture dynamic facial details, make the previous state-of-the-art method less effective than ours in generating features such as facial geometry, eye gaze direc + +![](images/30ccdfece604e8ff32ea2087997d1481740afb4bb8b3c82ab89f37b2ad4d2e2b.jpg) +Figure 5. Qualitative ablation studies of different components. + +tion, and lip synchronization in portraits. Due to variations in facial shapes, methods that rely on explicit keypoints are influenced by the facial characteristics of individuals in the driving video. This leads to insufficient separation between motion and identity. Furthermore, since their approach utilizes the image generation model Stable Diffusion as its foundation, it also exhibits inferior performance compared to ours in terms of temporal smoothness. Building upon stable video diffusion and utilizing our meticulously designed fine-grained feature extractors, our method exhibits exceptional performance in identity preservation, video quality enhancement, and the capture of subtle facial movements, resulting in more vivid and smoother generated videos. + +# 4.5. Ablation Study + +Effect of Motion Memory Bank. We validate the effectiveness of the motion memory bank. As shown in Table 2, the incorporation of the motion memory bank results in higher-quality videos, demonstrating improvements in FID-VID and FVD, as well as enhanced facial movements. We also observe an improvement in smoothness. Through the utilization of motion memories, our methodology demonstrates context-based perception capabilities, thereby enhancing motion features in the presence of ambiguous information. By integrating memories with information from preceding and subsequent frames, the memory bank can provide more detailed information into facial dynamics. As illustrated in Figure 5, the forehead wrinkles associated with the elevation of the eyes become more pronounced, when the character's eyebrows are raised. + +Effect of Intensity-Aware Motion Encoder (IAME). We confirm the significant role of the intensity-aware motion encoder. As illustrated in Table 2, the incorporation of this component, in contrast to the direct utilization of coarse + +Table 2. Ablation study to validate the effectiveness of different components. LMD (Landmark Mean Distance) multiplied by $10^{-3}$ and Identity Similarity multiplied by $10^{-1}$ . + +
SettingIdentity ↑ SimilarityFID-VID ↓FVD ↓LMD ↓
HunyuanPortrait8.8775.81333.482.02
- Memory Bank8.7578.43361.942.78
- IAME8.6380.79385.434.01
- ID Features8.2175.03330.121.98
- IMAdapter8.0977.14352.672.54
+ +motion features derived from the pre-trained motion encoder [8], offers critical guidance and intricate details regarding the facial dynamics associated with portrait motions. This enhancement significantly elevates the overall quality of the generated video, evidenced by an improvement in FID-VID, FVD, and ID similarity. As illustrated in Figure 5, the absence of feature refinement in the IAME may result in inaccuracies in the micro-expressions of the characters, including the direction of their gaze. Furthermore, the intricate details of the face are neglected. + +Effect of ID features and IMAdapter. To validate the effectiveness of the IMAdapter, we first remove the ID features from the IMAdapter, leaving only multiple convolution kernels for low-rank feature processing. As shown in Table 2 and Figure 5, the ablation results demonstrate that incorporating ID features is critical in enhancing the model's ID consistency ability. We further remove IMAdapter from the appearance extractor's blocks. As demonstrated in Figure 5, the incorporation of IMAdapter provides more fine-grained features related to facial texture, geometry, and other facial characteristics. + +# 5. Conclusion + +In this paper, we present an enhanced portrait animation framework that addresses the limitations of previous approaches. Our model leverages implicit representations with stable video diffusion to achieve temporal consistency and precise control over facial dynamics. The proposed model demonstrates significant improvements in the overall quality and temporal consistency of generated videos, allowing for more natural and lifelike animations. We offer a robust framework for generating high-quality, controllable animations, which achieve high disentanglement. Our method not only enhances the visual experience but also opens up possibilities for immersive content creation. With continued research and development, this approach holds great potential for applications in virtual reality, gaming, and human-computer interaction. However, as with any technology that generates realistic human likenesses, it is essential to consider the ethical implications. Future work should also develop safeguards to ensure responsible use. + +# References + +[1] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 4 +[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 5 +[3] Zhiyuan Chen, Jiajiong Cao, Zhiquan Chen, Yuming Li, and Chenguang Ma. Echomimic: Lifelike audio-driven portrait animations through editable landmark conditions. arXiv preprint arXiv:2407.08136, 2024. 1, 2 +[4] Kun Cheng, Xiaodong Cun, Yong Zhang, Menghan Xia, Fei Yin, Mingrui Zhu, Xuan Wang, Jue Wang, and Nannan Wang. Videoretalking: Audio-based lip synchronization for talking head video editing in the wild. In SIGGRAPH Asia 2022, pages 1-9, 2022. 1 +[5] Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. Voxceleb2: Deep speaker recognition. arXiv preprint arXiv:1806.05622, 2018. 5, 6 +[6] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 2, 4, 5, 6 +[7] Michail Christos Doukas, Stefanos Zafeiriou, and Viktoriia Sharmanska. Headgan: One-shot neural head synthesis and editing, 2021. 1, 2 +[8] Nikita Drobyshev, Jenya Chelishev, Taras Khakhulin, Aleksei Ivakhnenko, Victor Lempitsky, and Egor Zakharov. Megaportraits: One-shot megapixel neural head avatars. In Proceedings of the 30th ACM International Conference on Multimedia, pages 2663-2671, 2022. 4, 8 +[9] Nikita Drobyshev, Jenya Chelishev, Taras Khakhulin, Aleksei Ivakhnenko, Victor Lempitsky, and Egor Zakharov. Megaportraits: One-shot megapixel neural head avatars, 2023. 2 +[10] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 1 +[11] Yuming Gu, Hongyi Xu, You Xie, Guoxian Song, Yichun Shi, Di Chang, Jing Yang, and Linjie Luo. Diffportrait3d: Controllable diffusion for zero-shot portrait view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10456-10465, 2024. 1 +[12] Hanzhong Guo, Hongwei Yi, Daquan Zhou, Alexander William Bergman, Michael Lingelbach, and Yizhou Yu. Real-time one-step diffusion-based expressive portrait videos generation. arXiv preprint arXiv:2412.13479, 2024. 2 +[13] Jianzhu Guo, Dingyun Zhang, Xiaogiang Liu, Zhizhou Zhong, Yuan Zhang, Pengfei Wan, and Di Zhang. Livepor + +trait: Efficient portrait animation with stitching and retargeting control. arXiv preprint arXiv:2407.03168, 2024. 2, 6 +[14] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. In The Twelfth International Conference on Learning Representations, 2024. 2 +[15] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6 +[16] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 6 +[17] Fa-Ting Hong and Dan Xu. Implicit identity representation conditioned memory compensation network for talking head video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23062-23072, 2023. 2 +[18] Fa-Ting Hong, Li Shen, and Dan Xu. Dagan++: Depth-aware generative adversarial network for talking head video generation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2 +[19] Li Hu. Animate anyone: Consistent and controllable image-to-video synthesis for character animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8153-8163, 2024. 5 +[20] Xianliang Huang, Jiajie Gou, Shuhang Chen, Zhizhou Zhong, Jihong Guan, and Shuigeng Zhou. Iddr-ngp: Incorporating detectors for distractors removal with instant neural radiance field. In Proceedings of the 31st ACM International Conference on Multimedia, pages 1343-1351, 2023. 2 +[21] Xiaoyu Jin, Zunnan Xu, Mingwen Ou, and Wenming Yang. Alignment is all you need: A training-free augmentation strategy for pose-guided video generation. arXiv preprint arXiv:2408.16506, 2024. 2 +[22] Taras Khakhulin, Vanessa Sklyarova, Victor Lempitsky, and Egor Zakharov. Realistic one-shot mesh-based head avatars. In European Conference on Computer Vision, pages 345-362. Springer, 2022. 2 +[23] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3 +[24] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013.5 +[25] Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, et al. Hunyuanvideo: A systematic framework for large video generative models. arXiv preprint arXiv:2412.03603, 2024. 2 +[26] Hongxiang Li, Meng Cao, Xuxin Cheng, Yaowei Li, Zhihong Zhu, and Yuexian Zou. G2l: Semantically aligned and uniform video grounding via geodesic and game theory. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12032-12042, 2023. 2 +[27] Hongxiang Li, Yaowei Li, Yuhang Yang, Junjie Cao, Zhihong Zhu, Xuxin Cheng, and Long Chen. Dispose: Disen + +tangling pose guidance for controllable human image animation. arXiv preprint arXiv:2412.09349, 2024. 2 +[28] Ronghui Li, Junfan Zhao, Yachao Zhang, Mingyang Su, Zeping Ren, Han Zhang, Yansong Tang, and Xiu Li. Finedance: A fine-grained choreography dataset for 3d full body dance generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10234-10243, 2023. 1 +[29] Ronghui Li, Hongwen Zhang, Yachao Zhang, Yuxiang Zhang, Youliang Zhang, Jie Guo, Yan Zhang, Xiu Li, and Yebin Liu. Lodge++: High-quality and long dance generation with vivid choreography patterns. arXiv preprint arXiv:2410.20389, 2024. +[30] Ronghui Li, YuXiang Zhang, Yachao Zhang, Hongwen Zhang, Jie Guo, Yan Zhang, Yebin Liu, and Xiu Li. Lodge: A coarse to fine diffusion network for long dance generation guided by the characteristic dance primitives. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1524-1534, 2024. 1 +[31] Gang LIU, Xin CHEN, and Zhixiang GAO. A novel double-tail generative adversarial network for fast photo animation. IEICE Transactions on Information and Systems, E107.D(1): 72-82, 2024. 5 +[32] Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuoling Chang, Ming Guang Yong, Juhyun Lee, et al. Mediapipe: A framework for building perception pipelines. arXiv preprint arXiv:1906.08172, 2019. 6 +[33] Yue Ma, Yingqing He, Xiaodong Cun, Xintao Wang, Siran Chen, Xiu Li, and Qifeng Chen. Follow your pose: Pose-guided text-to-video generation using pose-free videos. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4117-4125, 2024. 4 +[34] Yue Ma, Hongyu Liu, Hongfa Wang, Heng Pan, Yingqing He, Junkun Yuan, Ailing Zeng, Chengfei Cai, Heung-Yeung Shum, Wei Liu, et al. Follow-your-emoji: Fine-controllable and expressive freestyle portrait animation. arXiv preprint arXiv:2406.01900, 2024. 6 +[35] Yuxi Mi, Zhizhou Zhong, Yuge Huang, Jiazhen Ji, Jianqing Xu, Jun Wang, Shaoming Wang, Shouhong Ding, and Shuigeng Zhou. Privacy-preserving face recognition using trainable feature subtraction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 297-307, 2024. 2 +[36] Yuval Nirkin, Yosi Keller, and Tal Hassner. Fsgan: Subject agnostic face swapping and reenactment. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7184-7193, 2019. 1 +[37] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 4, 5 +[38] Bohao Peng, Jian Wang, Yuechen Zhang, Wenbo Li, MingChang Yang, and Jiaya Jia. Controlnext: Powerful and efficient control for image and video generation. arXiv preprint arXiv:2408.06070, 2024. 2, 5 + +[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 2, 3 +[40] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 3 +[41] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. Advances in neural information processing systems, 32, 2019. 1 +[42] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. In Advances in Neural Information Processing Systems, 2019. 2, 6 +[43] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 6 +[44] Jiale Tao, Shuhang Gu, Wen Li, and Lixin Duan. Learning motion refinement for unsupervised face animation. Advances in Neural Information Processing Systems, 36, 2024. 2 +[45] Jiayuan Tian, Yanling LONG, Hong YANG, Huifang WU, Peng XUE, and Zhongqing JIANG. Cues of eye region and their effects on face-personality perception. Advances in Psychological Science, 30(12):2735, 2022. 2 +[46] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphaël Marinier, Marcin Michalski, and Sylvain Gelly. Fvd: A new metric for video generation. 2019. 6 +[47] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems, 30, 2017. 4 +[48] Cong Wang, Kuan Tian, Jun Zhang, Yonghang Guan, Feng Luo, Fei Shen, Zhiwei Jiang, Qing Gu, Xiao Han, and Wei Yang. V-express: Conditional dropout for progressive training of portrait video generation. arXiv preprint arXiv:2406.02511, 2024. 1, 2, 5 +[49] Suzhen Wang, Lincheng Li, Yu Ding, and Xin Yu. One-shot talking face generation from single-speaker audio-visual correlation learning. In AAAI Conference on Artificial Intelligence, pages 2531–2539, 2022. 2 +[50] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6 +[51] Huawei Wei, Zejun Yang, and Zhisheng Wang. Aniportrait: Audio-driven synthesis of photorealistic portrait animations, 2024. 2, 4, 6 +[52] Liangbin Xie, Xintao Wang, Honglun Zhang, Chao Dong, and Ying Shan. Vfhq: A high-quality dataset and benchmark for video face super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 657-666, 2022. 5, 6 + +[53] You Xie, Hongyi Xu, Guoxian Song, Chao Wang, Yichun Shi, and Linjie Luo. X-portrait: Expressive portrait animation with hierarchical motion attention. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2, 6 +[54] Jingjing Xu, Xu Sun, Zhiyuan Zhang, Guangxiang Zhao, and Junyang Lin. Understanding and improving layer normalization. Advances in neural information processing systems, 32, 2019. 4 +[55] Sicheng Xu, Guojun Chen, Yu-Xiao Guo, Jiaolong Yang, Chong Li, Zhenyu Zang, Yizhong Zhang, Xin Tong, and Baining Guo. Vasa-1: Lifelike audio-driven talking faces generated in real time. arXiv preprint arXiv:2404.10667, 2024. 2 +[56] Zunnan Xu, Yukang Lin, Haonan Han, Sicheng Yang, Ronghui Li, Yachao Zhang, and Xiu Li. Mambatak: Efficient holistic gesture synthesis with selective state space models. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 1 +[57] Shurong Yang, Huadong Li, Juhao Wu, Minhao Jing, Linze Li, Renhe Ji, Jiajun Liang, and Haoqiang Fan. Megactor: Harness the power of raw video for vivid portrait animation. arXiv preprint arXiv:2405.20851, 2024. 2 +[58] Zhendong Yang, Ailing Zeng, Chun Yuan, and Yu Li. Effective whole-body pose estimation with two-stages distillation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4210-4220, 2023. 5 +[59] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023.4 +[60] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. Styleheat: One-shot high-resolution editable talking face generation via pre-trained stylegan. In European Conference on Computer Vision, pages 85–101, 2022. 2 +[61] Wangbo Yu, Yanbo Fan, Yong Zhang, Xuan Wang, Fei Yin, Yunpeng Bai, Yan-Pei Cao, Ying Shan, Yang Wu, Zhongqian Sun, and Baoyuan Wu. Nofa: Nerf-based one-shot facial avatar reconstruction, 2023. 2 +[62] Bowen Zhang, Chenyang Qi, Pan Zhang, Bo Zhang, Hsiang-Tao Wu, Dong Chen, Qifeng Chen, Yong Wang, and Fang Wen. Metaportrait: Identity-preserving talking head generation with fast personalized adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22096-22105, 2023. 1, 2 +[63] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6 +[64] Yang Zhang, Jiaxi Gu, Li-Wen Wang, Han Wang, Junqi Cheng, Yuefeng Zhu, and Fangyuan Zou. Mimic motion: High-quality human motion video generation with confidence-aware pose guidance. arXiv preprint arXiv:2406.19680, 2024. 2, 5 +[65] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with + +a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661-3670, 2021. 5 +[66] Jian Zhao and Hui Zhang. Thin-plate spline motion model for image animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3657–3666, 2022. 2 +[67] Longtao Zheng, Yifan Zhang, Hanzhong Guo, Jiachun Pan, Zhenxiong Tan, Jiahao Lu, Chuanxin Tang, Bo An, and Shuicheng Yan. Memo: Memory-guided diffusion for expressive talking video generation. arXiv preprint arXiv:2412.04448, 2024. 2 +[68] Hao Zhu, Wayne Wu, Wentao Zhu, Liming Jiang, Siwei Tang, Li Zhang, Ziwei Liu, and Chen Change Loy. Celebvehq: A large-scale video facial attributes dataset. In European conference on computer vision, pages 650-667. Springer, 2022. 5 \ No newline at end of file diff --git a/2025/HunyuanPortrait_ Implicit Condition Control for Enhanced Portrait Animation/images.zip b/2025/HunyuanPortrait_ Implicit Condition Control for Enhanced Portrait Animation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f8747494632a73890c6c5e907a3f3317393e9af2 --- /dev/null +++ b/2025/HunyuanPortrait_ Implicit Condition Control for Enhanced Portrait Animation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:27ed9b0be8c763c1c2b9c36ee1f34655020c946256e71314347f954c7801ee10 +size 666125 diff --git a/2025/HunyuanPortrait_ Implicit Condition Control for Enhanced Portrait Animation/layout.json b/2025/HunyuanPortrait_ Implicit Condition Control for Enhanced Portrait Animation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c8c6c1895124430e0bb9da3ef79b6f0db7c2ed63 --- /dev/null +++ b/2025/HunyuanPortrait_ Implicit Condition Control for Enhanced Portrait Animation/layout.json @@ -0,0 +1,7065 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 59, + 103, + 550, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 103, + 550, + 120 + ], + "spans": [ + { + "bbox": [ + 59, + 103, + 550, + 120 + ], + "type": "text", + "content": "HunyuanPortrait: Implicit Condition Control for Enhanced Portrait Animation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 79, + 142, + 533, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 142, + 533, + 186 + ], + "spans": [ + { + "bbox": [ + 79, + 142, + 533, + 186 + ], + "type": "text", + "content": "Zunnan Xu" + }, + { + "bbox": [ + 79, + 142, + 533, + 186 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 79, + 142, + 533, + 186 + ], + "type": "text", + "content": ", Zhentao Yu" + }, + { + "bbox": [ + 79, + 142, + 533, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 79, + 142, + 533, + 186 + ], + "type": "text", + "content": ", Zixiang Zhou" + }, + { + "bbox": [ + 79, + 142, + 533, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 79, + 142, + 533, + 186 + ], + "type": "text", + "content": ", Jun Zhou" + }, + { + "bbox": [ + 79, + 142, + 533, + 186 + ], + "type": "inline_equation", + "content": "^{3*}" + }, + { + "bbox": [ + 79, + 142, + 533, + 186 + ], + "type": "text", + "content": ", Xiaoyu Jin" + }, + { + "bbox": [ + 79, + 142, + 533, + 186 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 79, + 142, + 533, + 186 + ], + "type": "text", + "content": ", Fa-ting Hong" + }, + { + "bbox": [ + 79, + 142, + 533, + 186 + ], + "type": "inline_equation", + "content": "^{4*}" + }, + { + "bbox": [ + 79, + 142, + 533, + 186 + ], + "type": "text", + "content": ", Xiaozhong Ji" + }, + { + "bbox": [ + 79, + 142, + 533, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 79, + 142, + 533, + 186 + ], + "type": "text", + "content": ", Junwei Zhu" + }, + { + "bbox": [ + 79, + 142, + 533, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 79, + 142, + 533, + 186 + ], + "type": "text", + "content": ", Chengfei Cai" + }, + { + "bbox": [ + 79, + 142, + 533, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 79, + 142, + 533, + 186 + ], + "type": "text", + "content": ", Shiyu Tang" + }, + { + "bbox": [ + 79, + 142, + 533, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 79, + 142, + 533, + 186 + ], + "type": "text", + "content": ", Qin Lin" + }, + { + "bbox": [ + 79, + 142, + 533, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 79, + 142, + 533, + 186 + ], + "type": "text", + "content": ", Xiu Li" + }, + { + "bbox": [ + 79, + 142, + 533, + 186 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 79, + 142, + 533, + 186 + ], + "type": "text", + "content": ", Qinglin Lu" + }, + { + "bbox": [ + 79, + 142, + 533, + 186 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 79, + 142, + 533, + 186 + ], + "type": "text", + "content": ", Tsinghua University " + }, + { + "bbox": [ + 79, + 142, + 533, + 186 + ], + "type": "inline_equation", + "content": "{}^{2}" + }, + { + "bbox": [ + 79, + 142, + 533, + 186 + ], + "type": "text", + "content": "Hunyuan, Tencent " + }, + { + "bbox": [ + 79, + 142, + 533, + 186 + ], + "type": "inline_equation", + "content": "{}^{3}" + }, + { + "bbox": [ + 79, + 142, + 533, + 186 + ], + "type": "text", + "content": "Sun Yat-Sen Univeristy " + }, + { + "bbox": [ + 79, + 142, + 533, + 186 + ], + "type": "inline_equation", + "content": "{}^{4}" + }, + { + "bbox": [ + 79, + 142, + 533, + 186 + ], + "type": "text", + "content": "HKUST" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 238, + 296, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 238, + 296, + 489 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 296, + 489 + ], + "type": "text", + "content": "We introduce HunyuanPortrait, a diffusion-based condition control method that employs implicit representations for highly controllable and lifelike portrait animation. Given a single portrait image as an appearance reference and video clips as driving templates, HunyuanPortrait can animate the character in the reference image by the facial expression and head pose of the driving videos. In our framework, we utilize pre-trained encoders to achieve the decoupling of portrait motion information and identity in videos. To do so, implicit representation is adopted to encode motion information and is employed as control signals in the animation phase. By leveraging the power of stable video diffusion as the main building block, we carefully design adapter layers to inject control signals into the denoising unet through attention mechanisms. These bring spatial richness of details and temporal consistency. HunyuanPortrait also exhibits strong generalization performance, which can effectively disentangle appearance and motion under different image styles. Our framework outperforms existing methods, demonstrating superior temporal consistency and controllability. Our project is available at HunyuanPortrait." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 511, + 135, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 511, + 135, + 522 + ], + "spans": [ + { + "bbox": [ + 55, + 511, + 135, + 522 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 531, + 295, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 531, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 531, + 295, + 602 + ], + "type": "text", + "content": "In this paper, we aim to animate a single static image into a lifelike video using control signals in implicit representations derived from driving videos. As shown in Figure 1, our primary focus is not only on the high controllability of digital portraits but also on the strong consistency of the characters and background in the reference image." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 603, + 295, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 603, + 295, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 603, + 295, + 687 + ], + "type": "text", + "content": "Portrait animation has been widely applied in fields such as human-computer interaction [11, 56], virtual reality, and the crafting of digital avatars for the gaming and multimedia industries [28-30]. Previous works [4, 7, 62] develop usable portrait animation models utilizing Generative Adversarial Networks (GANs) [10] based on raw video inputs. However, these approaches exhibit limitations in terms of gener" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 212, + 553, + 485 + ], + "blocks": [ + { + "bbox": [ + 316, + 212, + 553, + 485 + ], + "lines": [ + { + "bbox": [ + 316, + 212, + 553, + 485 + ], + "spans": [ + { + "bbox": [ + 316, + 212, + 553, + 485 + ], + "type": "image", + "image_path": "a1ad754f12f89d77499d7c3e7dddd63a2b74498d5679a4605742fef0238e910a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 494, + 555, + 561 + ], + "lines": [ + { + "bbox": [ + 313, + 494, + 555, + 561 + ], + "spans": [ + { + "bbox": [ + 313, + 494, + 555, + 561 + ], + "type": "text", + "content": "Figure 1. Our framework employs implicit condition control to generate portrait animations, demonstrating robust generalization performance with high-fidelity facial dynamics and vivid head poses during cross-reenactment. The animated portraits remain unaffected by variations in facial shape and the spatial position of the driving videos, demonstrating strong identity consistency." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 581, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 581, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 581, + 555, + 689 + ], + "type": "text", + "content": "alization. Given the variability of facial structures and the complexities of facial movements, these methods often face challenges such as artifacts and insufficient control over facial expressions. This is particularly evident when there is a substantial difference in facial shape between the driving video and the source image, resulting in significant issues with facial artifacts and motion distortion. These early methods often produce results with background jitter and blurred portraits during cross-reenactment [36, 41]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "The advancement of diffusion models has led to a significant focus in recent methodologies [3, 48] on the fine-tuning" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 212, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 212, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 212, + 703 + ], + "type": "text", + "content": "*Work done during the internship at Tencent." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15909" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 72, + 294, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 72, + 294, + 358 + ], + "spans": [ + { + "bbox": [ + 57, + 72, + 294, + 358 + ], + "type": "text", + "content": "of image-level diffusion models [39] through the incorporation of motion modules [14], thereby facilitating the generation of video content. These methods effectively address the issues associated with GAN-based approaches in managing complex backgrounds and enhancing generalization. However, even with the incorporation of motion modules, these methods are susceptible to defects during the generation stage (e.g., temporal smoothness and flexibility to different frame rates). This vulnerability arises from the absence of motion pre-training in the underlying models during the pre-training phase. Additionally, these methods remain constrained to the explicit control of facial expressions. The effectiveness of these methods, which depend on explicit keypoints to control facial expressions, not only results in the loss of intricate details but also significantly relies on the accuracy of the landmark extraction techniques employed. Furthermore, as a result of the variability in facial shapes, these methods are limited by post-processing alignment strategies and often face challenges in adapting to diverse facial geometries, which can lead to inaccurate facial control and insufficient preservation of identity in the visual effects of videos. Due to the human eye's sensitivity to subtle facial changes [45], it is crucial to preserve the fidelity of facial details and the smoothness of the videos." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 367, + 294, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 367, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 367, + 294, + 712 + ], + "type": "text", + "content": "Therefore, we propose an implicit conditional control model to tackle the challenges posed by diverse facial geometries and intricate expression details. We utilize stable video diffusion as the foundational model and integrate identity and motion information through appearance and motion attention. This approach eliminates the need to fine-tune the image diffusion model and to train the motion module separately. Firstly, we coarsely decouple the identity information from the motion information in the original video by using a pre-trained motion encoder [9]. Considering that this pre-trained encoder has not completely decoupled identity from motion information, we further implement enhanced training strategies and improved network architectures to strengthen the control of motion features and the separation of portrait identities. For the adaptation of temporal modeling tasks in video generation using diffusion models, along with the complementary contextual information present between different frames, we propose the incorporation of a motion memory bank to enhance the implicit representation of motion features. Due to the substantial variations in the degree of blur and pixel distortion in video frames at different levels of motion intensity, we propose an intensity-aware motion encoder. This encoder aims to improve the capacity of decoupled motion features to capture intricate details. Additionally, in order to achieve consistent modeling of portrait identity and background areas in videos, we combined ArcFace [6] with the DiNOv2 [37] backbone to develop an enhanced appearance encoder. To preserve the pre-training knowledge of both models, we" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 316, + 73, + 553, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 119 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 119 + ], + "type": "text", + "content": "froze their original parameters and proposed an ID-aware Multi-scale adapter (IMAdapter). This approach achieves high-fidelity modeling of the intricate details in portraits. Our main contributions can be summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 121, + 553, + 251 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 314, + 121, + 553, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 121, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 314, + 121, + 553, + 167 + ], + "type": "text", + "content": "- We propose the first implicit condition animation framework based on stable video diffusion, which significantly improves temporal consistency and achieves high fidelity and precise control for portrait animation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 168, + 553, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 168, + 553, + 214 + ], + "spans": [ + { + "bbox": [ + 313, + 168, + 553, + 214 + ], + "type": "text", + "content": "- We develop a series of strategies to enhance identity consistency and improve the accuracy of portrait motion, including data augmentation, enhanced training and inference strategies, and improved network architecture." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 216, + 553, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 216, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 553, + 251 + ], + "type": "text", + "content": "- Experiments demonstrate the effectiveness of our proposed method, which achieves state-of-the-art performance in both subjective and objective metrics." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 316, + 263, + 400, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 263, + 400, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 263, + 400, + 274 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 316, + 282, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 282, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 282, + 553, + 713 + ], + "type": "text", + "content": "Human portrait animation aims to bring a still image to life by utilizing a driving source, such as a sequence of facial landmarks or frames that contain the head of the portrait. Recent advancements in the field of neural avatars can be divided into two subcategories: non-diffusion-based and diffusion-based methods. Non-diffusion-based methods [13, 17, 18, 26, 42, 49, 55, 62, 66] are renowned for their ability to achieve realistic facial animations and fidelity to motion. Many works employed predetermined motion representations that are popular in the literature, such as blendshapes from 3D Morphable Models (3DMM) [7, 20, 22, 35, 44, 60, 61]. However, these works often display significant artifacts when there is a large disparity between the driving portrait and the source image. Due to the use of a warping-based strategy, these methods are also ineffective at capturing videos with significant head and body movements. With the development of diffusion models [25], recent methods [12, 51, 53, 57] achieve human portrait animation by fine-tuning the stable diffusion model [39]. Several studies have investigated the generation of facial expressions through explicit keypoint control [3, 48, 67]. This method shows great promise for enhancing the quality of generated videos and enables precise control of facial expressions through the extraction of facial landmarks. However, given the diversity of facial geometries, there are notable differences in the distribution of facial keypoints among individuals. The alignment processes used in these methods often struggle to manage boundary cases effectively, resulting in identification offsets due to inaccurate alignment. In addition, the reliance on image-based generative models and separately trained motion modules to achieve temporal consistency results in shortcomings related to temporal smoothness. In the pursuit of improved temporal continuity, pioneering work [21, 27, 38, 64] has employed stable video diffusion models to construct end-to-end diffusion models for human video generation. However, these methods are pri" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "15910" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 70, + 553, + 244 + ], + "blocks": [ + { + "bbox": [ + 60, + 70, + 553, + 244 + ], + "lines": [ + { + "bbox": [ + 60, + 70, + 553, + 244 + ], + "spans": [ + { + "bbox": [ + 60, + 70, + 553, + 244 + ], + "type": "image", + "image_path": "de615433cfc9cb796a91735626fead6552017e81571f987ea3475810ddc42e85.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 57, + 252, + 293, + 370 + ], + "blocks": [ + { + "bbox": [ + 57, + 252, + 293, + 370 + ], + "lines": [ + { + "bbox": [ + 57, + 252, + 293, + 370 + ], + "spans": [ + { + "bbox": [ + 57, + 252, + 293, + 370 + ], + "type": "image", + "image_path": "e95f8eea86262ba3c83d889629e2fe3c0687b973cf8f82c8ba08b023aa2dc88e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 378, + 555, + 413 + ], + "lines": [ + { + "bbox": [ + 54, + 378, + 555, + 413 + ], + "spans": [ + { + "bbox": [ + 54, + 378, + 555, + 413 + ], + "type": "text", + "content": "Figure 2. Our framework utilizes implicit representation to encode motion information as control signals. By harnessing the capabilities of stable video diffusion as the primary building block, we have meticulously designed a fine-grained appearance extractor to maintain the identity of the portrait, along with an intensity-aware motion extractor to capture intricate facial dynamics." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 294, + 255, + 553, + 369 + ], + "blocks": [ + { + "bbox": [ + 294, + 255, + 553, + 369 + ], + "lines": [ + { + "bbox": [ + 294, + 255, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 294, + 255, + 553, + 369 + ], + "type": "image", + "image_path": "c8aef9ebb1feb928dbfbf9f27ab00ef311674bcfa8f72237008d8d3065f82bda.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 433, + 295, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 433, + 295, + 577 + ], + "spans": [ + { + "bbox": [ + 54, + 433, + 295, + 577 + ], + "type": "text", + "content": "marily designed for full-body movements and do not take into account the preservation of facial identity or the diversity of human facial geometries. As a result, they often lead to significant alterations in the identity of the generated face. Additionally, these methods lack the ability to control subtle facial dynamics, which limits their application in generating portraits that incorporate speech content and facial expressions. In this study, we investigate the strategy of adapting to various face shapes through implicit representations of expression descriptors and achieving capture of facial dynamics. Our method surpasses previous approaches in terms of temporal consistency and generalization ability." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 590, + 138, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 138, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 138, + 604 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 611, + 141, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 141, + 623 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 141, + 623 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": "Diffusion Model. Our framework is built upon diffusion model, as introduced in [39]. The model is designed to facilitate efficient and stable training by conducting the diffusion and denoising processes within a latent space, rather than directly in the image space. The Variational Autoencoder (VAE), as proposed in [23], plays a pivotal role in this process by mapping images from the RGB color space into" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 433, + 555, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 433, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 433, + 555, + 456 + ], + "type": "text", + "content": "a latent space. This transformation enables the diffusion process to be directed by textual embeddings." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 457, + 556, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 457, + 556, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 457, + 556, + 612 + ], + "type": "text", + "content": "The process begins with the VAE model, which projects images into latent embeddings. This model comprises two key components: an encoder " + }, + { + "bbox": [ + 313, + 457, + 556, + 612 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 313, + 457, + 556, + 612 + ], + "type": "text", + "content": " and a decoder " + }, + { + "bbox": [ + 313, + 457, + 556, + 612 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 457, + 556, + 612 + ], + "type": "text", + "content": ". The encoder " + }, + { + "bbox": [ + 313, + 457, + 556, + 612 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 313, + 457, + 556, + 612 + ], + "type": "text", + "content": " compresses the pixel space image " + }, + { + "bbox": [ + 313, + 457, + 556, + 612 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 457, + 556, + 612 + ], + "type": "text", + "content": " into a latent representation " + }, + { + "bbox": [ + 313, + 457, + 556, + 612 + ], + "type": "inline_equation", + "content": "z = \\mathcal{E}(x)" + }, + { + "bbox": [ + 313, + 457, + 556, + 612 + ], + "type": "text", + "content": ", while the decoder " + }, + { + "bbox": [ + 313, + 457, + 556, + 612 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 457, + 556, + 612 + ], + "type": "text", + "content": " reconstructs it back into an image space, aiming for " + }, + { + "bbox": [ + 313, + 457, + 556, + 612 + ], + "type": "inline_equation", + "content": "\\mathcal{D}(z) \\approx x" + }, + { + "bbox": [ + 313, + 457, + 556, + 612 + ], + "type": "text", + "content": ". Subsequently, a UNet-based architecture [40] is employed for learning the reverse denoising process within the latent space. This network integrates attention mechanisms through Transformer Blocks. The cross-attention mechanism ensures that the textual prompt is effectively integrated throughout the process. The overall training objective for the UNet can be articulated as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 341, + 617, + 555, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 617, + 555, + 635 + ], + "spans": [ + { + "bbox": [ + 341, + 617, + 555, + 635 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\underset {t, z, \\epsilon} {\\mathbb {E}} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(\\sqrt {\\bar {\\alpha} _ {t}} z + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , c, t\\right) \\| ^ {2} \\right], \\tag {1}", + "image_path": "10455b8e4777c4fcc4547efc29c460ae759e4e5e6a9a110e87ce0805730a39fb.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 642, + 556, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 712 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 642, + 556, + 712 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 642, + 556, + 712 + ], + "type": "text", + "content": " denotes the latent embedding of the training sample. " + }, + { + "bbox": [ + 313, + 642, + 556, + 712 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 313, + 642, + 556, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 642, + 556, + 712 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 642, + 556, + 712 + ], + "type": "text", + "content": " represent the predicted noise by the diffusion model and the ground truth noise at the corresponding timestep " + }, + { + "bbox": [ + 313, + 642, + 556, + 712 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 642, + 556, + 712 + ], + "type": "text", + "content": ", respectively. " + }, + { + "bbox": [ + 313, + 642, + 556, + 712 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 642, + 556, + 712 + ], + "type": "text", + "content": " is the condition embedding, and the coefficient " + }, + { + "bbox": [ + 313, + 642, + 556, + 712 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t" + }, + { + "bbox": [ + 313, + 642, + 556, + 712 + ], + "type": "text", + "content": " remains consistent with that employed in diffusion models." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15911" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 168, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 168, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 168, + 83 + ], + "type": "text", + "content": "3.2. Overall framework" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 296, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 296, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 296, + 354 + ], + "type": "text", + "content": "As illustrated in Figure 2, building upon stable video diffusion [1], our method is designed with two core components for generating portrait videos: the appearance extractor and the motion extractor. For the detailed appearance modeling of the portrait, we integrate the appearance encoder to manage the subject's identity and background in the generated video. We carefully design an ID-aware multi-scale adapter based on the Arcface encoder [6] and DiNOv2 [37] backbone to enhance visual appearance representation. For the motion extractor, we emphasize the enhancement of the implicit representation of facial dynamics. By utilizing the pre-trained motion encoder [8], we further optimize it by taking into account the intensity and the temporal continuity of motions. This involved the introduction of motion memory banks and an intensity-aware motion encoder to enhance the perception of motion intensity and refine features. These representations are conditionally integrated into the denoising U-Net through attention mechanisms. Our framework also integrates a spatial conditioning signal to ensure the stability of regions beyond the face. By utilizing these features, we generate latent features using a denoising UNet and subsequently produce video through VAEs." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 361, + 239, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 239, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 239, + 373 + ], + "type": "text", + "content": "3.3. Intensity-Aware Motion Extractor" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 379, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 379, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 379, + 296, + 714 + ], + "type": "text", + "content": "For portrait motion modeling, we concentrate on identity-agnostic portrait animation by utilizing an implicit representation of expression descriptors derived from raw videos, which capture a diverse array of facial dynamics (e.g., lip-sync, micro-expressions, eye gaze, and blinking). This approach differs from existing diffusion-based methods that typically employ explicit landmarks for facial expression representation [33, 51]. Considering that the fidelity of facial motion capture is often compromised by extraneous factors—such as variations in facial shape and background noise that can distort essential motion signals—we crop the central area of the face to focus on the regions most indicative of motion dynamics. The boundary is defined as the area between the eyebrows and the bottom of the mouth. This approach allows us to isolate facial movements and enhance the signal-to-noise ratio. The cropped pixels from driving videos are subsequently utilized as inputs for the pre-trained motion encoder [8] to extract a coarse implicit conditioning signal. Since motion intensity varies and exerts distinct influences on the generated pixels, we introduce an Intensity-Aware Motion Encoder into our framework. This addition enhances the coarse motion features from the motion encoder by adapting to the varying intensities of motion. The intensity is estimated by two dimensions: the degree of distortion in facial expressions and the overall amplitude of head movement. Considering that the expression and head pose intensity rely on landmark positioning and their relative differences, we propose a calcu" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "content": "lation method for these variances using landmarks and normalization, which can be formulated as below:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 353, + 106, + 553, + 142 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 106, + 553, + 142 + ], + "spans": [ + { + "bbox": [ + 353, + 106, + 553, + 142 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} e _ {k, j} = l _ {k, j} - l _ {k, c}, k \\in [ 1, n ], j \\in [ 1, m ] \\\\ s = \\sqrt {\\left(l _ {1 , \\max } - l _ {1 , \\min }\\right) ^ {2}}, \\tag {2} \\\\ \\end{array}", + "image_path": "e3e3f05b3cccb4051f55f9513c55e8eac749545c4c222b8474a1c55d1133146e.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 152, + 555, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 152, + 555, + 224 + ], + "spans": [ + { + "bbox": [ + 313, + 152, + 555, + 224 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 152, + 555, + 224 + ], + "type": "inline_equation", + "content": "l_{k,j}" + }, + { + "bbox": [ + 313, + 152, + 555, + 224 + ], + "type": "text", + "content": " denotes the j-th landmark point in the k-th frame and " + }, + { + "bbox": [ + 313, + 152, + 555, + 224 + ], + "type": "inline_equation", + "content": "l_{k,c}" + }, + { + "bbox": [ + 313, + 152, + 555, + 224 + ], + "type": "text", + "content": " denotes the center of the landmark point (e.g., nose tip). " + }, + { + "bbox": [ + 313, + 152, + 555, + 224 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 152, + 555, + 224 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 152, + 555, + 224 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 152, + 555, + 224 + ], + "type": "text", + "content": " denote the numbers of frames and points. " + }, + { + "bbox": [ + 313, + 152, + 555, + 224 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 152, + 555, + 224 + ], + "type": "text", + "content": " represents the face scale derived from the first frame of the landmarks. We further calculate the intensity of expressions " + }, + { + "bbox": [ + 313, + 152, + 555, + 224 + ], + "type": "inline_equation", + "content": "I_e" + }, + { + "bbox": [ + 313, + 152, + 555, + 224 + ], + "type": "text", + "content": " and the head poses " + }, + { + "bbox": [ + 313, + 152, + 555, + 224 + ], + "type": "inline_equation", + "content": "I_h" + }, + { + "bbox": [ + 313, + 152, + 555, + 224 + ], + "type": "text", + "content": " using the formulation below:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 341, + 234, + 553, + 278 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 234, + 553, + 278 + ], + "spans": [ + { + "bbox": [ + 341, + 234, + 553, + 278 + ], + "type": "interline_equation", + "content": "I _ {e} = \\frac {1}{n \\cdot s} \\sum_ {k = 1} ^ {n} \\sqrt {\\frac {1}{m} \\sum_ {j = 1} ^ {m} \\left(e _ {k , j} - \\frac {1}{n} \\sum_ {k = 1} ^ {n} e _ {k , j}\\right) ^ {2}}, \\tag {3}", + "image_path": "b9dc970aaf78d935c418d2968b6851cc3a415d4267755095895a7ddbc023c283.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 342, + 275, + 490, + 312 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 275, + 490, + 312 + ], + "spans": [ + { + "bbox": [ + 342, + 275, + 490, + 312 + ], + "type": "interline_equation", + "content": "I _ {h} = \\frac {1}{s} \\sqrt {\\frac {1}{n} \\sum_ {k = 1} ^ {n} (l _ {k , c} - \\frac {1}{n} \\sum_ {k = 1} ^ {n} l _ {k , c}) ^ {2}}.", + "image_path": "96e6306c73764a3371a20c6c471c28a814f1cffc8b1ea591d2b80f849ea71593.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 323, + 555, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 555, + 489 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 555, + 489 + ], + "type": "text", + "content": "Considering the issue of accuracy in landmark detection, we discretize the calculated continuous values by dividing the intensity into " + }, + { + "bbox": [ + 313, + 323, + 555, + 489 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 323, + 555, + 489 + ], + "type": "text", + "content": " discrete levels based on the range of values. We set " + }, + { + "bbox": [ + 313, + 323, + 555, + 489 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 323, + 555, + 489 + ], + "type": "text", + "content": " to 64 and then map these discrete values using embeddings " + }, + { + "bbox": [ + 313, + 323, + 555, + 489 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 323, + 555, + 489 + ], + "type": "text", + "content": " to their corresponding intensity feature vectors " + }, + { + "bbox": [ + 313, + 323, + 555, + 489 + ], + "type": "inline_equation", + "content": "E_{s} = \\mathrm{Concat}[\\mathcal{D}(I_{e}),\\mathcal{D}(I_{h})]" + }, + { + "bbox": [ + 313, + 323, + 555, + 489 + ], + "type": "text", + "content": ", and concatenate them along the channel dimension to adjust the feature space in response to the dynamic range of motion intensity. In light of the fact that the sequence of motion features is derived from pixel extraction and exhibits a lack of continuity across different contexts, we propose the motion memory bank to improve the context-awareness and temporal modeling capabilities of motion features, which can be formulated as below:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 372, + 498, + 553, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 498, + 553, + 559 + ], + "spans": [ + { + "bbox": [ + 372, + 498, + 553, + 559 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\hat {f} _ {m} = \\operatorname {A d a L N} \\left(F _ {m} + E _ {s}, E _ {s}\\right), \\\\ \\bar {f} _ {m} = \\operatorname {M H C A} (\\hat {m}, \\hat {f} _ {m}, \\hat {f} _ {m}), \\tag {4} \\\\ \\bar {f} _ {m} = \\operatorname {A d a L N} \\left(\\bar {f} _ {m}, E _ {s}\\right), \\\\ \\bar {f} _ {m} = \\operatorname {M L P} (\\bar {f} _ {m}), \\\\ \\end{array}", + "image_path": "a34f2a30e93039e81344093a9360e24f1dcbdf9f9396c6e99ebb62a75b7625e9.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "inline_equation", + "content": "F_{m}" + }, + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "text", + "content": " denotes the features extracted by the pre-trained motion encoder, and " + }, + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\hat{m}" + }, + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "text", + "content": " signifies the query features from the motion memory bank, with " + }, + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\hat{f}_{m}" + }, + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "text", + "content": " being utilized as the key and value features. The MHCA denotes multi-head cross-attention [47] and the AdaLN denotes adaptive layer normalization [54]. The learnable memories are designed to supplement contextual information, enriching motion features rather than generating new motions. This helps the model capture pixel correlation between different frames, alleviating blurring when some facial parts undergo deformation. The refined motion features are incorporated into the denoising U-Net using cross-attention mechanisms [59]." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15912" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 70, + 290, + 128 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 290, + 128 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 290, + 128 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 290, + 128 + ], + "type": "image", + "image_path": "a8788756699950c2dd20dd00aa70d062cff2beff2fab1d8f8f07e3245c9e4a48.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 137, + 295, + 171 + ], + "lines": [ + { + "bbox": [ + 55, + 137, + 295, + 171 + ], + "spans": [ + { + "bbox": [ + 55, + 137, + 295, + 171 + ], + "type": "text", + "content": "Figure 3. The illustration of our ID-aware Multi-scale Adapter (IMAdapter). Here, " + }, + { + "bbox": [ + 55, + 137, + 295, + 171 + ], + "type": "inline_equation", + "content": "\\mathbb{C}" + }, + { + "bbox": [ + 55, + 137, + 295, + 171 + ], + "type": "text", + "content": " represents the operation of concatenating features along the channel dimension." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 192, + 246, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 192, + 246, + 205 + ], + "spans": [ + { + "bbox": [ + 55, + 192, + 246, + 205 + ], + "type": "text", + "content": "3.4. Fine-grained Appearance Extractor" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 210, + 295, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 210, + 295, + 437 + ], + "spans": [ + { + "bbox": [ + 54, + 210, + 295, + 437 + ], + "type": "text", + "content": "To achieve appearance control, which encompasses both portrait identity information and background details, we propose an appearance extractor that creates a fine-grained representation of the animated portrait and background. Existing video diffusion models often struggle with identity consistency [38, 64], which results in deficiencies in the method's capacity to generate portraits. Due to the human eye's sensitivity to subtle facial changes, visual artifacts can lead to inaccurate facial details and insufficient preservation of identity in the generated videos. Therefore, we adopt a patch-level image encoder [37] as the backbone of the appearance encoder and introduce an ID-aware multi-scale adapter (IMAdapter) to enhance the model's ID consistency capabilities. To process the entire image, we first extract information related to clothing and background elements of the portrait. For the enhancement of consistency, as illustrated in Figure 2, we utilize the ID encoder to extract identity-related features. We further propose extracting multi-scale, fine-grained identity features using adapters." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 438, + 295, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 438, + 295, + 534 + ], + "spans": [ + { + "bbox": [ + 54, + 438, + 295, + 534 + ], + "type": "text", + "content": "As illustrated in Figure 3, we first use the linear projection to downsample the visual features. The low-rank features are subsequently fed into convolutional modules at multiple scales, along with a cross-attention mechanism, which captures ID-aware fine-grained details for the preservation of visual identity. The cross-modal features are then linearly projected back to the original dimensions to be merged with patch-level features:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 123, + 544, + 295, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 544, + 295, + 609 + ], + "spans": [ + { + "bbox": [ + 123, + 544, + 295, + 609 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\hat {f} _ {a} = f _ {a} \\mathbf {W} _ {d o w n} ^ {v}, \\\\ \\hat {f} _ {c} = \\operatorname {M C o n v} (\\hat {f} _ {a}), \\tag {5} \\\\ \\bar {f} _ {a} = \\mathrm {M H C A} (f _ {i d}, \\hat {f} _ {c}, \\hat {f} _ {c}), \\\\ f _ {a} = f _ {a} + \\bar {f} _ {a} \\mathbf {W} _ {u p} ^ {v}, \\\\ \\end{array}", + "image_path": "59244c0ecd5e44bcde7327fd4e9a84739e185c43215b38b8a0a7107fdf4b8d7f.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "inline_equation", + "content": "f_{id}" + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": " represents the ID features, which are utilized as the query features, while " + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\hat{f}_c" + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": " serves as the key and value features. MConv signifies the multiple convolutions that process appearance features " + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "inline_equation", + "content": "f_a" + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": " in parallel and concatenate them along the channel dimension. During the training phase, the sampling of ID information comes from a random frame in the video sequence; in the inference stage, it comes from the reference image." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 495, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 495, + 85 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 495, + 85 + ], + "type": "text", + "content": "3.5. Training and Inference Strategies" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 89, + 555, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 555, + 521 + ], + "type": "text", + "content": "For adaptation to various anime styles, we utilize AnimeGANv3 [31] to perform style conversion on the dataset. This allows the model to adjust to different styles of driving videos and source images. Instead of directly training models with AnimeGANv3, we employ its style transfer capability to convert facial portrait datasets into anime-stylized representations. These transformed data samples are then integrated into our training dataset for enhanced feature learning. To mitigate the impact of skin color and other color-related information on facial movements, we employ color jittering to augment the data for the facial cropping input of the motion encoder. For maintaining the stability of the regions beyond the face in the generated video, we utilize DWPose [58] as an additional spatial conditioning signal. Specifically, we employ the pose-guided network [19] to achieve spatial control condition mapping. We further incorporate spatial conditional control into the latent features of the first spatial convolutional layer in U-Net through addition. To simulate scenarios in which certain skeletons from the driving video are absent compared to those in the source image during the inference stage, we implement an augmentation strategy [48] for the pose sequence during training. This strategy randomly removes certain edges throughout the training process to improve the stability of spatial control. This mitigates the effects of detector accuracy limitations on detector-based control conditions and offers flexible control capability. During inference, considering that there are differences in the body shape and spatial position of bones between the source image and the driving video, we first translate the position based on the offset of the center point of the face (e.g., the position of the nose tip) and scale it according to the length ratio of the skeletons. We find that the key points of the eyes can lead to subtle variations in face shape. Therefore, in the setting of cross reenactment, we do not utilize the key points and corresponding edges of the eyes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 532, + 395, + 545 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 532, + 395, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 532, + 395, + 545 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 552, + 432, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 432, + 565 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 432, + 565 + ], + "type": "text", + "content": "4.1. Experiments Setting" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "We train our model using HDTF [65], VFHQ [52], VoxCeleb2 [5], CelebV-HQ [68], and our own collected dataset. This collected dataset comprises monocular camera recordings showcasing 20 hours of authentic human video footage obtained from 200 subjects. The recordings were captured in both indoor and outdoor environments. During the training process, all of the images and videos are resized to " + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": ". For the optimization of our framework, we employ the AdamW optimizer with a learning rate of " + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-5}" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": ". The weights of the ID Encoder [6], Variational Autoencoder [24] from Stable Video Diffusion [2] and the blocks of DiNOv2 [37] are fixed. In the motion memory bank, we" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15913" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 101, + 553, + 184 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 1. Quantitative results on " + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": " test images. LMD multiplied by " + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "inline_equation", + "content": "10^{-3}" + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": ", AED multiplied by " + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "inline_equation", + "content": "10^{-2}" + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": ", APD multiplied by " + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "inline_equation", + "content": "10^{-3}" + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": " and Identity Similarity multiplied by " + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "inline_equation", + "content": "10^{-1}" + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": " indicates higher is better. " + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": " indicates lower is better. The best results are in bold." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 101, + 553, + 184 + ], + "lines": [ + { + "bbox": [ + 58, + 101, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 58, + 101, + 553, + 184 + ], + "type": "table", + "html": "
MethodSelf ReenactmentCross Reenactment
LMD ↓FID-VID ↓FVD ↓PSNR ↑SSIM ↑LPIPS ↓AED ↓APD ↓Identity ↑ SimilarityFacial ↑ MovementsVideo ↑ QualityTemporal ↑ Smoothness
LivePortrait [13]9.1482.71483.3831.410.720.2226.8326.978.713.004.114.06
AntiPortrait [51]6.6781.90430.2430.540.670.2722.4221.327.953.834.003.81
FollowYE [34]5.6377.15417.5330.840.680.2520.0620.958.184.114.233.33
X-Potrait [53]6.2382.93416.4130.810.710.1920.6620.398.033.883.733.49
HunyuanPortrait (Ours)2.0275.81333.4832.980.810.1119.4519.208.874.554.694.61
", + "image_path": "d5f9b4d3828ed8733b33b71a7d91d52d1ed177c3d1d6772b67fd39ea6e04e02f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 205, + 294, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 205, + 294, + 314 + ], + "spans": [ + { + "bbox": [ + 54, + 205, + 294, + 314 + ], + "type": "text", + "content": "incorporate 64 learnable motion memories, each with a dimension of 768, which share parameters across the various blocks. They are initialized using a scaled normal distribution, where each element is sampled from " + }, + { + "bbox": [ + 54, + 205, + 294, + 314 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0,\\frac{1}{\\sqrt{\\mathrm{hidden\\_dim}}})" + }, + { + "bbox": [ + 54, + 205, + 294, + 314 + ], + "type": "text", + "content": ". To ensure stability, we apply gradient norm clipping at a value of 0.99. During inference, we employ the DDIM sampler [43] and set the scale of classifier-free guidance [16] to 2.0 for our experiment. The framework is trained on 128 NVIDIA A100 GPUs for three days." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 323, + 115, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 323, + 115, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 115, + 335 + ], + "type": "text", + "content": "4.2. Metrics" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 340, + 296, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 340, + 296, + 664 + ], + "spans": [ + { + "bbox": [ + 56, + 340, + 296, + 664 + ], + "type": "text", + "content": "We employ both qualitative and quantitative analyses to assess the generated video quality and motion accuracy of our portrait animation results. For self-reenactment, we utilize the Fréchet Inception Distance (FID) [15], the Fréchet Video Distance (FVD) [46], Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM) [50], Learned Perceptual Image Patch Similarity (LPIPS) [63] and the Landmark Mean Distances (LMD) [32]. For cross-reenactment, we utilize Average Expression Distance (AED) [42], Average Pose Distance (APD) [42] and the ArcFace Score [6] as the identity (ID) similarity. Detailed descriptions of these metrics are provided in supplementary materials. Considering the absence of ground truth and the significant error between current evaluation methods and human perception, we combine the results of user studies that evaluates cross-reenactment from three additional dimensions: facial movements, video quality, and temporal smoothness. The facial movements encompass not only facial expressions but also the overall orientation of the generated portrait head. The quality of the video and its temporal smoothness are primarily assessed by the clarity of the video and the smoothness of the transitions between frames. For evaluation, we utilize source images and driving videos from two prominent datasets [5, 52]. This experiment aims to evaluate the model's ability to transfer expressions and movements from one subject to another while preserving the identity of the source portrait." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 672, + 255, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 255, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 255, + 685 + ], + "type": "text", + "content": "4.3. Quantitative and Qualitative Analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": "As shown in Table 1, our method attains the lowest FIDVID (average FID of frames in videos) and FVD, with a" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 205, + 555, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 555, + 660 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 555, + 660 + ], + "type": "text", + "content": "significant advantage when compared to the previously top-performing method, which underscores the superior capability of HunyuanPortrait. We also find that non-diffusion-based methods have much lower generation quality compared to diffusion-based methods. Although the video resolution is the same, the generated portrait details have significant differences. In more challenging cross-reenactment scenarios, our approach demonstrates strong generalization capabilities, leading to substantial improvements across all metrics. This fully demonstrates the advantage of our method, as cross-reenactment is much more difficult than self-reenactment and is closer to real-life application scenarios. The warping-based method LivePortrait struggles to accurately model overall head movement, resulting in generated portraits that exhibit minimal variation in head pose relative to the reference image. Compared with other methods, our method has stronger ID similarity, facial dynamic generation ability, video quality, and temporal smoothness. Our method employs implicit motion representation rather than explicit facial keypoints to control the generation of facial expressions. Our method further takes into account the intensity of motion, the ability to perceive identity, and complementary information based on temporal context. This approach mitigates identity distortion caused by variations in facial geometries and effectively captures dynamic facial details that explicit keypoints cannot represent. Meanwhile, since our method is based on stable video diffusion, the generated results are smoother compared to previous image-based diffusion methods and have higher video quality. These enhancements highlight the superior accuracy and fidelity of our method in capturing fine-grained details, affirming its efficacy in synthesizing realistic portrait motions. Furthermore, due to the lack of open-source training code, combined with the challenges related to their motion module and the frame rate of the videos, FollowYourEmoji and X-Portrait shows suboptimal performance in terms of temporal smoothness. Our method remains unaffected by variations in frame rate." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 670, + 399, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 670, + 399, + 682 + ], + "spans": [ + { + "bbox": [ + 314, + 670, + 399, + 682 + ], + "type": "text", + "content": "4.4. Visualization" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "As illustrated in Figure 4, our method can capture more detailed and consistent portrait movements with the driv" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15914" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 69, + 555, + 504 + ], + "blocks": [ + { + "bbox": [ + 57, + 69, + 555, + 504 + ], + "lines": [ + { + "bbox": [ + 57, + 69, + 555, + 504 + ], + "spans": [ + { + "bbox": [ + 57, + 69, + 555, + 504 + ], + "type": "image", + "image_path": "81cdda35b14df4f6649f2d0cc102d583a41bad022873b8cd1eb3888319e71d7b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 111, + 513, + 499, + 523 + ], + "lines": [ + { + "bbox": [ + 111, + 513, + 499, + 523 + ], + "spans": [ + { + "bbox": [ + 111, + 513, + 499, + 523 + ], + "type": "text", + "content": "Figure 4. Qualitative comparisons of self-reenactment and cross-reenactment with state-of-the-art methods." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 545, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 297, + 713 + ], + "type": "text", + "content": "ing video. For self-reenactment, our method generates results that enable more detailed capture of facial dynamics, including eye gaze direction, eye rotation, and lip synchronization. In addition, our method can also achieve tracking of the overall rotation of the head, which cannot be achieved by previous top-performing warping-based method LivePortrait, such as LivePortrait. Current state-of-the-art diffusion-based methods have significant defects in maintaining the portrait's identity and facial shape due to the deviation of predicted keypoints during head rotation. In contrast, our method effectively preserves the identity of the portrait and captures intricate facial dynamics. Previous methods are influenced by the differences in facial geometries between the driving video and the reference image," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "text", + "content": "leading to substantial challenges in preserving the identity of the generated video. This is a result of inadequate disentanglement of their appearance features and facial dynamics. These issues result in problems such as blurry videos and distorted portraits. Our framework effectively addresses the disentanglement of the two key elements involved in portrait animation, thereby demonstrating a strong generalization ability, as shown in Figure 4. The diffusion-based method exhibits superior generation quality; however, discrepancies in facial keypoints due to variations in facial shape, combined with the limited ability of these keypoints to accurately capture dynamic facial details, make the previous state-of-the-art method less effective than ours in generating features such as facial geometry, eye gaze direc" + } + ] + } + ], + "index": 3 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "15915" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 294, + 277 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 294, + 277 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 294, + 277 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 294, + 277 + ], + "type": "image", + "image_path": "30ccdfece604e8ff32ea2087997d1481740afb4bb8b3c82ab89f37b2ad4d2e2b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 286, + 288, + 298 + ], + "lines": [ + { + "bbox": [ + 61, + 286, + 288, + 298 + ], + "spans": [ + { + "bbox": [ + 61, + 286, + 288, + 298 + ], + "type": "text", + "content": "Figure 5. Qualitative ablation studies of different components." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 304, + 296, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 304, + 296, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 296, + 460 + ], + "type": "text", + "content": "tion, and lip synchronization in portraits. Due to variations in facial shapes, methods that rely on explicit keypoints are influenced by the facial characteristics of individuals in the driving video. This leads to insufficient separation between motion and identity. Furthermore, since their approach utilizes the image generation model Stable Diffusion as its foundation, it also exhibits inferior performance compared to ours in terms of temporal smoothness. Building upon stable video diffusion and utilizing our meticulously designed fine-grained feature extractors, our method exhibits exceptional performance in identity preservation, video quality enhancement, and the capture of subtle facial movements, resulting in more vivid and smoother generated videos." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 468, + 149, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 149, + 480 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 149, + 480 + ], + "type": "text", + "content": "4.5. Ablation Study" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 486, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 295, + 665 + ], + "type": "text", + "content": "Effect of Motion Memory Bank. We validate the effectiveness of the motion memory bank. As shown in Table 2, the incorporation of the motion memory bank results in higher-quality videos, demonstrating improvements in FID-VID and FVD, as well as enhanced facial movements. We also observe an improvement in smoothness. Through the utilization of motion memories, our methodology demonstrates context-based perception capabilities, thereby enhancing motion features in the presence of ambiguous information. By integrating memories with information from preceding and subsequent frames, the memory bank can provide more detailed information into facial dynamics. As illustrated in Figure 5, the forehead wrinkles associated with the elevation of the eyes become more pronounced, when the character's eyebrows are raised." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "Effect of Intensity-Aware Motion Encoder (IAME). We confirm the significant role of the intensity-aware motion encoder. As illustrated in Table 2, the incorporation of this component, in contrast to the direct utilization of coarse" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 317, + 113, + 553, + 196 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 555, + 105 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 555, + 105 + ], + "type": "text", + "content": "Table 2. Ablation study to validate the effectiveness of different components. LMD (Landmark Mean Distance) multiplied by " + }, + { + "bbox": [ + 313, + 70, + 555, + 105 + ], + "type": "inline_equation", + "content": "10^{-3}" + }, + { + "bbox": [ + 313, + 70, + 555, + 105 + ], + "type": "text", + "content": " and Identity Similarity multiplied by " + }, + { + "bbox": [ + 313, + 70, + 555, + 105 + ], + "type": "inline_equation", + "content": "10^{-1}" + }, + { + "bbox": [ + 313, + 70, + 555, + 105 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 113, + 553, + 196 + ], + "lines": [ + { + "bbox": [ + 317, + 113, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 317, + 113, + 553, + 196 + ], + "type": "table", + "html": "
SettingIdentity ↑ SimilarityFID-VID ↓FVD ↓LMD ↓
HunyuanPortrait8.8775.81333.482.02
- Memory Bank8.7578.43361.942.78
- IAME8.6380.79385.434.01
- ID Features8.2175.03330.121.98
- IMAdapter8.0977.14352.672.54
", + "image_path": "5e9895874bda8d589aca8a1a0d6f8e3f7083c3ef7f2588167bf69d932412f232.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 217, + 555, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 555, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 555, + 335 + ], + "type": "text", + "content": "motion features derived from the pre-trained motion encoder [8], offers critical guidance and intricate details regarding the facial dynamics associated with portrait motions. This enhancement significantly elevates the overall quality of the generated video, evidenced by an improvement in FID-VID, FVD, and ID similarity. As illustrated in Figure 5, the absence of feature refinement in the IAME may result in inaccuracies in the micro-expressions of the characters, including the direction of their gaze. Furthermore, the intricate details of the face are neglected." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 336, + 556, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 336, + 556, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 336, + 556, + 468 + ], + "type": "text", + "content": "Effect of ID features and IMAdapter. To validate the effectiveness of the IMAdapter, we first remove the ID features from the IMAdapter, leaving only multiple convolution kernels for low-rank feature processing. As shown in Table 2 and Figure 5, the ablation results demonstrate that incorporating ID features is critical in enhancing the model's ID consistency ability. We further remove IMAdapter from the appearance extractor's blocks. As demonstrated in Figure 5, the incorporation of IMAdapter provides more fine-grained features related to facial texture, geometry, and other facial characteristics." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 478, + 388, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 478, + 388, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 478, + 388, + 491 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "text", + "content": "In this paper, we present an enhanced portrait animation framework that addresses the limitations of previous approaches. Our model leverages implicit representations with stable video diffusion to achieve temporal consistency and precise control over facial dynamics. The proposed model demonstrates significant improvements in the overall quality and temporal consistency of generated videos, allowing for more natural and lifelike animations. We offer a robust framework for generating high-quality, controllable animations, which achieve high disentanglement. Our method not only enhances the visual experience but also opens up possibilities for immersive content creation. With continued research and development, this approach holds great potential for applications in virtual reality, gaming, and human-computer interaction. However, as with any technology that generates realistic human likenesses, it is essential to consider the ethical implications. Future work should also develop safeguards to ensure responsible use." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15916" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 147, + 296, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 147, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 61, + 147, + 296, + 201 + ], + "type": "text", + "content": "[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 202, + 295, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 202, + 295, + 246 + ], + "spans": [ + { + "bbox": [ + 62, + 202, + 295, + 246 + ], + "type": "text", + "content": "[3] Zhiyuan Chen, Jiajiong Cao, Zhiquan Chen, Yuming Li, and Chenguang Ma. Echomimic: Lifelike audio-driven portrait animations through editable landmark conditions. arXiv preprint arXiv:2407.08136, 2024. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 247, + 295, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 247, + 295, + 302 + ], + "spans": [ + { + "bbox": [ + 62, + 247, + 295, + 302 + ], + "type": "text", + "content": "[4] Kun Cheng, Xiaodong Cun, Yong Zhang, Menghan Xia, Fei Yin, Mingrui Zhu, Xuan Wang, Jue Wang, and Nannan Wang. Videoretalking: Audio-based lip synchronization for talking head video editing in the wild. In SIGGRAPH Asia 2022, pages 1-9, 2022. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 303, + 295, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 303, + 295, + 335 + ], + "spans": [ + { + "bbox": [ + 62, + 303, + 295, + 335 + ], + "type": "text", + "content": "[5] Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. Voxceleb2: Deep speaker recognition. arXiv preprint arXiv:1806.05622, 2018. 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 336, + 295, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 336, + 295, + 390 + ], + "spans": [ + { + "bbox": [ + 62, + 336, + 295, + 390 + ], + "type": "text", + "content": "[6] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 2, 4, 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 391, + 295, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 391, + 295, + 424 + ], + "spans": [ + { + "bbox": [ + 62, + 391, + 295, + 424 + ], + "type": "text", + "content": "[7] Michail Christos Doukas, Stefanos Zafeiriou, and Viktoriia Sharmanska. Headgan: One-shot neural head synthesis and editing, 2021. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 425, + 295, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 425, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 62, + 425, + 295, + 479 + ], + "type": "text", + "content": "[8] Nikita Drobyshev, Jenya Chelishev, Taras Khakhulin, Aleksei Ivakhnenko, Victor Lempitsky, and Egor Zakharov. Megaportraits: One-shot megapixel neural head avatars. In Proceedings of the 30th ACM International Conference on Multimedia, pages 2663-2671, 2022. 4, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 480, + 295, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 480, + 295, + 523 + ], + "spans": [ + { + "bbox": [ + 62, + 480, + 295, + 523 + ], + "type": "text", + "content": "[9] Nikita Drobyshev, Jenya Chelishev, Taras Khakhulin, Aleksei Ivakhnenko, Victor Lempitsky, and Egor Zakharov. Megaportraits: One-shot megapixel neural head avatars, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 525, + 295, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 525, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 57, + 525, + 295, + 568 + ], + "type": "text", + "content": "[10] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 569, + 295, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 569, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 57, + 569, + 295, + 634 + ], + "type": "text", + "content": "[11] Yuming Gu, Hongyi Xu, You Xie, Guoxian Song, Yichun Shi, Di Chang, Jing Yang, and Linjie Luo. Diffportrait3d: Controllable diffusion for zero-shot portrait view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10456-10465, 2024. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "type": "text", + "content": "[12] Hanzhong Guo, Hongwei Yi, Daquan Zhou, Alexander William Bergman, Michael Lingelbach, and Yizhou Yu. Real-time one-step diffusion-based expressive portrait videos generation. arXiv preprint arXiv:2412.13479, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 690, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 690, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 690, + 295, + 714 + ], + "type": "text", + "content": "[13] Jianzhu Guo, Dingyun Zhang, Xiaogiang Liu, Zhizhou Zhong, Yuan Zhang, Pengfei Wan, and Di Zhang. Livepor" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "trait: Efficient portrait animation with stitching and retargeting control. arXiv preprint arXiv:2407.03168, 2024. 2, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 162 + ], + "type": "text", + "content": "[14] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. In The Twelfth International Conference on Learning Representations, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "type": "text", + "content": "[15] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 219, + 553, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 241 + ], + "type": "text", + "content": "[16] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 242, + 553, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 242, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 316, + 242, + 553, + 296 + ], + "type": "text", + "content": "[17] Fa-Ting Hong and Dan Xu. Implicit identity representation conditioned memory compensation network for talking head video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23062-23072, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 297, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 342 + ], + "type": "text", + "content": "[18] Fa-Ting Hong, Li Shen, and Dan Xu. Dagan++: Depth-aware generative adversarial network for talking head video generation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 343, + 555, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 555, + 387 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 555, + 387 + ], + "type": "text", + "content": "[19] Li Hu. Animate anyone: Consistent and controllable image-to-video synthesis for character animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8153-8163, 2024. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 388, + 553, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 442 + ], + "type": "text", + "content": "[20] Xianliang Huang, Jiajie Gou, Shuhang Chen, Zhizhou Zhong, Jihong Guan, and Shuigeng Zhou. Iddr-ngp: Incorporating detectors for distractors removal with instant neural radiance field. In Proceedings of the 31st ACM International Conference on Multimedia, pages 1343-1351, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 444, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 444, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 444, + 553, + 487 + ], + "type": "text", + "content": "[21] Xiaoyu Jin, Zunnan Xu, Mingwen Ou, and Wenming Yang. Alignment is all you need: A training-free augmentation strategy for pose-guided video generation. arXiv preprint arXiv:2408.16506, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 488, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 488, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 488, + 553, + 533 + ], + "type": "text", + "content": "[22] Taras Khakhulin, Vanessa Sklyarova, Victor Lempitsky, and Egor Zakharov. Realistic one-shot mesh-based head avatars. In European Conference on Computer Vision, pages 345-362. Springer, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 533, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 533, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 533, + 553, + 555 + ], + "type": "text", + "content": "[23] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 556, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 556, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 556, + 553, + 578 + ], + "type": "text", + "content": "[24] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013.5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 579, + 553, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 579, + 553, + 632 + ], + "spans": [ + { + "bbox": [ + 316, + 579, + 553, + 632 + ], + "type": "text", + "content": "[25] Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, et al. Hunyuanvideo: A systematic framework for large video generative models. arXiv preprint arXiv:2412.03603, 2024. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 635, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 690 + ], + "type": "text", + "content": "[26] Hongxiang Li, Meng Cao, Xuxin Cheng, Yaowei Li, Zhihong Zhu, and Yuexian Zou. G2l: Semantically aligned and uniform video grounding via geodesic and game theory. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12032-12042, 2023. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "text", + "content": "[27] Hongxiang Li, Yaowei Li, Yuhang Yang, Junjie Cao, Zhihong Zhu, Xuxin Cheng, and Long Chen. Dispose: Disen" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15917" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "tangling pose guidance for controllable human image animation. arXiv preprint arXiv:2412.09349, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 162 + ], + "type": "text", + "content": "[28] Ronghui Li, Junfan Zhao, Yachao Zhang, Mingyang Su, Zeping Ren, Han Zhang, Yansong Tang, and Xiu Li. Finedance: A fine-grained choreography dataset for 3d full body dance generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10234-10243, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 295, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 295, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 295, + 218 + ], + "type": "text", + "content": "[29] Ronghui Li, Hongwen Zhang, Yachao Zhang, Yuxiang Zhang, Youliang Zhang, Jie Guo, Yan Zhang, Xiu Li, and Yebin Liu. Lodge++: High-quality and long dance generation with vivid choreography patterns. arXiv preprint arXiv:2410.20389, 2024." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 285 + ], + "type": "text", + "content": "[30] Ronghui Li, YuXiang Zhang, Yachao Zhang, Hongwen Zhang, Jie Guo, Yan Zhang, Yebin Liu, and Xiu Li. Lodge: A coarse to fine diffusion network for long dance generation guided by the characteristic dance primitives. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1524-1534, 2024. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 286, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 294, + 330 + ], + "type": "text", + "content": "[31] Gang LIU, Xin CHEN, and Zhixiang GAO. A novel double-tail generative adversarial network for fast photo animation. IEICE Transactions on Information and Systems, E107.D(1): 72-82, 2024. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 331, + 294, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 294, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 294, + 387 + ], + "type": "text", + "content": "[32] Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuoling Chang, Ming Guang Yong, Juhyun Lee, et al. Mediapipe: A framework for building perception pipelines. arXiv preprint arXiv:1906.08172, 2019. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 388, + 294, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 443 + ], + "type": "text", + "content": "[33] Yue Ma, Yingqing He, Xiaodong Cun, Xintao Wang, Siran Chen, Xiu Li, and Qifeng Chen. Follow your pose: Pose-guided text-to-video generation using pose-free videos. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4117-4125, 2024. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 445, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 498 + ], + "type": "text", + "content": "[34] Yue Ma, Hongyu Liu, Hongfa Wang, Heng Pan, Yingqing He, Junkun Yuan, Ailing Zeng, Chengfei Cai, Heung-Yeung Shum, Wei Liu, et al. Follow-your-emoji: Fine-controllable and expressive freestyle portrait animation. arXiv preprint arXiv:2406.01900, 2024. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 500, + 294, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 500, + 294, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 500, + 294, + 567 + ], + "type": "text", + "content": "[35] Yuxi Mi, Zhizhou Zhong, Yuge Huang, Jiazhen Ji, Jianqing Xu, Jun Wang, Shaoming Wang, Shouhong Ding, and Shuigeng Zhou. Privacy-preserving face recognition using trainable feature subtraction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 297-307, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 568, + 295, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 295, + 612 + ], + "type": "text", + "content": "[36] Yuval Nirkin, Yosi Keller, and Tal Hassner. Fsgan: Subject agnostic face swapping and reenactment. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7184-7193, 2019. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 613, + 294, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 294, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 294, + 668 + ], + "type": "text", + "content": "[37] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 4, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 669, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 294, + 712 + ], + "type": "text", + "content": "[38] Bohao Peng, Jian Wang, Yuechen Zhang, Wenbo Li, MingChang Yang, and Jiaya Jia. Controlnext: Powerful and efficient control for image and video generation. arXiv preprint arXiv:2408.06070, 2024. 2, 5" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "type": "text", + "content": "[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 108, + 555, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 108, + 555, + 174 + ], + "spans": [ + { + "bbox": [ + 316, + 108, + 555, + 174 + ], + "type": "text", + "content": "[40] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 175, + 553, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 219 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 219 + ], + "type": "text", + "content": "[41] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. Advances in neural information processing systems, 32, 2019. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 220, + 553, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 266 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 266 + ], + "type": "text", + "content": "[42] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. In Advances in Neural Information Processing Systems, 2019. 2, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 267, + 553, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 267, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 316, + 267, + 553, + 300 + ], + "type": "text", + "content": "[43] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 302, + 553, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 302, + 553, + 345 + ], + "spans": [ + { + "bbox": [ + 316, + 302, + 553, + 345 + ], + "type": "text", + "content": "[44] Jiale Tao, Shuhang Gu, Wen Li, and Lixin Duan. Learning motion refinement for unsupervised face animation. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 347, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 347, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 347, + 553, + 392 + ], + "type": "text", + "content": "[45] Jiayuan Tian, Yanling LONG, Hong YANG, Huifang WU, Peng XUE, and Zhongqing JIANG. Cues of eye region and their effects on face-personality perception. Advances in Psychological Science, 30(12):2735, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 393, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 393, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 393, + 553, + 426 + ], + "type": "text", + "content": "[46] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphaël Marinier, Marcin Michalski, and Sylvain Gelly. Fvd: A new metric for video generation. 2019. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 428, + 553, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 428, + 553, + 473 + ], + "spans": [ + { + "bbox": [ + 316, + 428, + 553, + 473 + ], + "type": "text", + "content": "[47] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems, 30, 2017. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 475, + 553, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 475, + 553, + 529 + ], + "spans": [ + { + "bbox": [ + 316, + 475, + 553, + 529 + ], + "type": "text", + "content": "[48] Cong Wang, Kuan Tian, Jun Zhang, Yonghang Guan, Feng Luo, Fei Shen, Zhiwei Jiang, Qing Gu, Xiao Han, and Wei Yang. V-express: Conditional dropout for progressive training of portrait video generation. arXiv preprint arXiv:2406.02511, 2024. 1, 2, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 531, + 553, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 531, + 553, + 576 + ], + "spans": [ + { + "bbox": [ + 316, + 531, + 553, + 576 + ], + "type": "text", + "content": "[49] Suzhen Wang, Lincheng Li, Yu Ding, and Xin Yu. One-shot talking face generation from single-speaker audio-visual correlation learning. In AAAI Conference on Artificial Intelligence, pages 2531–2539, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 577, + 553, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 577, + 553, + 620 + ], + "spans": [ + { + "bbox": [ + 316, + 577, + 553, + 620 + ], + "type": "text", + "content": "[50] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 623, + 553, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 623, + 553, + 655 + ], + "spans": [ + { + "bbox": [ + 316, + 623, + 553, + 655 + ], + "type": "text", + "content": "[51] Huawei Wei, Zejun Yang, and Zhisheng Wang. Aniportrait: Audio-driven synthesis of photorealistic portrait animations, 2024. 2, 4, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 658, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 555, + 713 + ], + "type": "text", + "content": "[52] Liangbin Xie, Xintao Wang, Honglun Zhang, Chao Dong, and Ying Shan. Vfhq: A high-quality dataset and benchmark for video face super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 657-666, 2022. 5, 6" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "15918" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 117 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 117 + ], + "type": "text", + "content": "[53] You Xie, Hongyi Xu, Guoxian Song, Chao Wang, Yichun Shi, and Linjie Luo. X-portrait: Expressive portrait animation with hierarchical motion attention. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "type": "text", + "content": "[54] Jingjing Xu, Xu Sun, Zhiyuan Zhang, Guangxiang Zhao, and Junyang Lin. Understanding and improving layer normalization. Advances in neural information processing systems, 32, 2019. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "type": "text", + "content": "[55] Sicheng Xu, Guojun Chen, Yu-Xiao Guo, Jiaolong Yang, Chong Li, Zhenyu Zang, Yizhong Zhang, Xin Tong, and Baining Guo. Vasa-1: Lifelike audio-driven talking faces generated in real time. arXiv preprint arXiv:2404.10667, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 294, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 294, + 275 + ], + "type": "text", + "content": "[56] Zunnan Xu, Yukang Lin, Haonan Han, Sicheng Yang, Ronghui Li, Yachao Zhang, and Xiu Li. Mambatak: Efficient holistic gesture synthesis with selective state space models. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 275, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 275, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 275, + 294, + 319 + ], + "type": "text", + "content": "[57] Shurong Yang, Huadong Li, Juhao Wu, Minhao Jing, Linze Li, Renhe Ji, Jiajun Liang, and Haoqiang Fan. Megactor: Harness the power of raw video for vivid portrait animation. arXiv preprint arXiv:2405.20851, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 320, + 294, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 294, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 294, + 365 + ], + "type": "text", + "content": "[58] Zhendong Yang, Ailing Zeng, Chun Yuan, and Yu Li. Effective whole-body pose estimation with two-stages distillation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4210-4220, 2023. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 294, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 409 + ], + "type": "text", + "content": "[59] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023.4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 411, + 294, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 466 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 466 + ], + "type": "text", + "content": "[60] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. Styleheat: One-shot high-resolution editable talking face generation via pre-trained stylegan. In European Conference on Computer Vision, pages 85–101, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 467, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 294, + 510 + ], + "type": "text", + "content": "[61] Wangbo Yu, Yanbo Fan, Yong Zhang, Xuan Wang, Fei Yin, Yunpeng Bai, Yan-Pei Cao, Ying Shan, Yang Wu, Zhongqian Sun, and Baoyuan Wu. Nofa: Nerf-based one-shot facial avatar reconstruction, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 512, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 294, + 578 + ], + "type": "text", + "content": "[62] Bowen Zhang, Chenyang Qi, Pan Zhang, Bo Zhang, Hsiang-Tao Wu, Dong Chen, Qifeng Chen, Yong Wang, and Fang Wen. Metaportrait: Identity-preserving talking head generation with fast personalized adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22096-22105, 2023. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 578, + 294, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 578, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 578, + 294, + 634 + ], + "type": "text", + "content": "[63] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "text", + "content": "[64] Yang Zhang, Jiaxi Gu, Li-Wen Wang, Han Wang, Junqi Cheng, Yuefeng Zhu, and Fangyuan Zou. Mimic motion: High-quality human motion video generation with confidence-aware pose guidance. arXiv preprint arXiv:2406.19680, 2024. 2, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 690, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 690, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 690, + 294, + 714 + ], + "type": "text", + "content": "[65] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 262 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 335, + 72, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 72, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 335, + 72, + 555, + 106 + ], + "type": "text", + "content": "a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661-3670, 2021. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 555, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 555, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 555, + 150 + ], + "type": "text", + "content": "[66] Jian Zhao and Hui Zhang. Thin-plate spline motion model for image animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3657–3666, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 152, + 555, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 555, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 555, + 206 + ], + "type": "text", + "content": "[67] Longtao Zheng, Yifan Zhang, Hanzhong Guo, Jiachun Pan, Zhenxiong Tan, Jiahao Lu, Chuanxin Tang, Bo An, and Shuicheng Yan. Memo: Memory-guided diffusion for expressive talking video generation. arXiv preprint arXiv:2412.04448, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 207, + 555, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 207, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 207, + 555, + 262 + ], + "type": "text", + "content": "[68] Hao Zhu, Wayne Wu, Wentao Zhu, Liming Jiang, Siwei Tang, Li Zhang, Ziwei Liu, and Chen Change Loy. Celebvehq: A large-scale video facial attributes dataset. In European conference on computer vision, pages 650-667. Springer, 2022. 5" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "15919" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Hybrid Concept Bottleneck Models/7a729be8-7e97-4b64-9b6a-2c339ff64815_content_list.json b/2025/Hybrid Concept Bottleneck Models/7a729be8-7e97-4b64-9b6a-2c339ff64815_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..094d56650a827cc18e950443b7fc52a003878547 --- /dev/null +++ b/2025/Hybrid Concept Bottleneck Models/7a729be8-7e97-4b64-9b6a-2c339ff64815_content_list.json @@ -0,0 +1,1608 @@ +[ + { + "type": "text", + "text": "Hybrid Concept Bottleneck Models", + "text_level": 1, + "bbox": [ + 318, + 130, + 679, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yang Liu, Tianwei Zhang, Shi Gu \nUniversity of Electronic Science and Technology of China \nShenzhen Institute for Advanced Study, UESTC", + "bbox": [ + 267, + 181, + 728, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "liu.yang.minéstd.uestc.edu.cn, gsbyztw@gmail.com, gus@uestc.edu.cn", + "bbox": [ + 197, + 237, + 792, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 286, + 326, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Concept Bottleneck Models (CBMs) provide an interpretable framework for neural networks by mapping visual features to predefined, human-understandable concepts. However, the application of CBMs is often constrained by insufficient concept annotations. Recently, multi-modal pretrained models have shown promise in reducing annotation costs by aligning visual representations with textual concept embeddings. Nevertheless, the quality and completeness of the predefined concepts significantly affect the performance of CBMs. In this work, we propose Hybrid Concept Bottleneck Model (HybridCBM), a novel CBM framework to address the challenge of incomplete predefined concepts. Our method consists of two main components: a Static Concept Bank and a Dynamic Concept Bank. The Static Concept Bank directly leverages large language models (LLMs) for concept construction, while the Dynamic Concept Bank employs learnable vectors to capture complementary and valuable concepts continuously during training. After training, a pre-trained translator converts these vectors into human-understandable concepts, further enhancing model interpretability. HybridCBM is highly flexible and can be easily integrated with existing CBMs to improve both interpretability and performance. Experimental results1 on multiple datasets demonstrate that HybridCBM outperforms current state-of-the-art CBMs and achieves comparable results to black-box models. Additionally, we propose novel metrics to assess the quality of learned concepts, showing that they perform comparably to predefined concepts.", + "bbox": [ + 86, + 318, + 485, + 742 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 772, + 223, + 787 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep Neural Networks (DNNs) have become increasingly dominant in various fields, including computer vision, natural language processing, and speech recognition. However their complex and deep structures present a significant challenge to interpretability, often earning them the label", + "bbox": [ + 89, + 797, + 483, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4e6be5eabccdd36d34ca15099725a4981584a8936aa2bf2b1af6b458872f484c.jpg", + "image_caption": [ + "Figure 1. Our proposed concept-trainable CBM reduces the need for meticulously designed concepts by incorporating learned concepts (e.g., 'yellowish throat'), thereby enhancing interpretability and performance while minimizing manual design effort." + ], + "image_footnote": [], + "bbox": [ + 516, + 284, + 906, + 406 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "of 'blackbox' [23]. Addressing this challenge is crucial not only for ensuring accurate predictions, but also for providing scientists and engineers with more accessible tools for designing, developing, and debugging models. Interpretability is particularly critical in domains such as healthcare [10], medicine [33], education [7], and finance [28], where high-stakes decision-making requires transparency and trustworthiness to ensure reliability and accountability. Therefore, there has been a growing focus on research aimed at developing interpretable models. While post-hoc explanation methods [34, 37, 55] have gained popularity, they often fall short by providing unfaithful representations of the model's underlying computations [35].", + "bbox": [ + 509, + 477, + 906, + 674 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A natural approach to improving interpretability in DNNs is to design inherently interpretable models that leverage high-level concepts to describe features extracted by black-box models [36], imitating the way humans encode and understand the world[20]. One notable approach following such methodology is the Concept Bottleneck Models (CBM) [17]. CBMs operate by first predicting concepts from image representations mid-way through the decision-making process. These models pair each image with its predefined concepts and are trained in an end-to-end manner—initially predicting concepts and subsequently using these predictions to make final decisions. While this method of concept prediction enhances model interpretability, it often faces two fundamental challenges: (1) it is perceived to underperform compared to black-box models [11],", + "bbox": [ + 511, + 674, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{1}$ Code available at https://github.com/ly1998117/HybridCBM", + "bbox": [ + 107, + 886, + 437, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20179", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and (2) the requirement for precise annotations for each concept significantly limits their broader application.", + "bbox": [ + 89, + 90, + 480, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To mitigate these shortcomings, recent research has explored multimodal pretrained model-based CBMs [15, 52, 54], leveraging CLIP [32] to establish correspondences between visual and textual representations, thereby reducing the reliance on costly concept annotations. Specifically, CLIP encodes images and human-understandable concepts into a shared feature space, enabling the creation of a concept bottleneck by directly mapping visual representations to concept embeddings instead of through manual annotations. Regarding the trade-off between interpretability and performance, some studies [15, 38, 52] have emphasized the construction of high-quality concept banks through efficient concept selection. However, establishing a comprehensive concept bank remains a significant challenge [53], often constrained by the inherent limitations in language expressiveness and precision. In response, other research efforts have explored the integration of a residual linear layer to address the issue of incomplete concept extraction [39, 54], albeit at the cost of interpretability. For example, ResCBM [39] introduces estimizable vectors to capture missing concepts, utilizing a candidate concept bank in the residual component to regularizing these vectors toward specific candidates during training. However, the model is limited by the scope of the candidate concept bank, restricting its ability to learn concepts outside the predefined set. Additionally, the effectiveness of the residual structure remains an open question.", + "bbox": [ + 91, + 126, + 483, + 534 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose the Hybrid Concept Bottleneck Model (HybridCBM), specifically designed to address the challenges of incomplete concept representation, predefined concept bank dependence, and scalability issue by dynamically discovering new concepts directly from visual representations. Different from prior methods that rely on residual structure or external concept banks, our approach incorporates a hybrid concept bank comprising both a static and a dynamic concept bank. Initially, we establish the static concept bank by leveraging a large language model (LLM) like GPT-3.5 [27, 31], which is celebrated for its broad world knowledge [12, 30, 45]. For instance, in Figure 1, when prompted about a palm warbler, GPT-3.5 provides information such as 'feeds on the ground'. This textual descriptions is then encoded into embeddings by the text encoder of CLIP. Next, we initialize a set of estimizable vectors that form the dynamic concept bank. These vectors are refined to identify concepts present in the input image but absent from the static concept bank. To further enhance the concept discovery, we pre-train a concept translator, such as GPT-2 [31], to translate these newly discovered unknown concepts into textual descriptions, a process we term as 'concept labeling'. To evaluate the understandability, visual relevance and factual accuracy of the dynamic concepts, we", + "bbox": [ + 89, + 537, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "further design a set of evaluation metrics, some of which are based on Vision-Language Model (VLM) GPT-4o $^2$ . Our main contributions can be summarized as follows:", + "bbox": [ + 511, + 90, + 906, + 136 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Innovative Hybrid Concept Bank: Our model introduces a hybrid concept bank that combines static and dynamic concepts, allowing it to adapt and refine its interpretative capabilities dynamically. This structure leverages both predefined expert knowledge while enabling the discovery of new relevant concepts.", + "2. Seamless Integration of LLM Technologies: By employing advanced language models (LLM for defining static concepts, translating new concepts and VLM for evaluation) alongside the visual-textual bridging capabilities of CLIP, our model bridges the gap between visual representation and textual interpretation.", + "3. Enhanced Interpretability and Performance: Our approach maintains high interpretability through transparent concept labeling, while improving performance by discovering new concepts from visual information." + ], + "bbox": [ + 511, + 137, + 906, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 395, + 653, + 410 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Interpretable Neural Networks. One way to build an interpretable neural networks is the use of a concept-based explanation [3, 13, 15, 21, 29, 38, 43, 49]. Concept Bottleneck Models (CBMs) [17] are among the most popular approaches for making predictions based on human-interpretable concepts but require labor-intensive concept annotations for each image. Similarly, Concept Activation Vectors (CAVs) [14] represent concepts as normal vectors to the decision boundaries that separate positive and negative samples of a concept, though they also need extra datasets to train SVMs for each concept. Owing to the simple structure based on human-defined concepts, there have been studies aiming to alleviate these drawbacks, e.g., costly concept annotation and trade-off between interpretability and performance. Post-hoc Concept Bottleneck (PCBM) [54] utilizes ConceptNet [42] to obtain concepts and Leveraging CLIP to generate bottlenecks by projecting visual representations into the concept subspace. However, PCBM-h [54] incorporates a residual linear predictor that compensates for missing concepts by directly adding logits to PCBM's predictor. While this modification helps to recover accuracy, it does so at the expense of interpretability. The label-free CBM [26] uses CLIP for concept annotation, allowing arbitrary visual backbone to be transformed into an CBM without the need for labeled concept data. Labo [52] concentrates on concepts selection after generating candidate concepts through LLM. ResCBM [39] builds on PCBM-h's residual structure by using a candidate concept bank to guide and refine estimizable vectors. However, the model's expressive potential is limited by the extent of the candidate concept bank.", + "bbox": [ + 511, + 421, + 906, + 875 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "$^{2}$ https://openai.com/index/hello-gpt-4o/", + "bbox": [ + 529, + 886, + 743, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20180", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4c63391b7a4a249706c14a7a45c35f7af6190a369095e00c5ccf7c4b8a37d558.jpg", + "image_caption": [ + "Figure 2. Overview of the Hybrid Concept Bottleneck Model (HybridCBM), which is designed for interpretable concept learning and labeling. First, a Concept Translator is pre-trained to do concept-labeling. Next, a pretrained alignment model (CLIP) is used to obtain embeddings of predefined concepts, forming the Static Concept Bank. The Dynamic Concept Bank learns missing concepts and aligns its distribution with the Static Concept Bank to create the final concept bottleneck. Finally, an interpretable concept-based predictor classifies targets based on the concept scores." + ], + "image_footnote": [], + "bbox": [ + 114, + 87, + 883, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "CLIP in Image Captioning. Multimodal pretrained models like CLIP [32], trained on over 400 million image-text pairs using unsupervised contrastive loss, exhibit impressive capabilities in image captioning tasks [1, 19, 46]. Previous studies, such as [24] and [40], have primarily used CLIP as a visual encoder for captioning tasks, while [19] projects visual embeddings into CLIP's text embedding space to preserve visual information. However, these approaches often underutilize CLIP's aligned multi-modal latent space. Similar to [19, 24], we collect image-caption pairs and a vast corpus to train a GPT-2 from scratch as our concept translator, fully leveraging CLIP's latent space and bridging the gap between image and concept embeddings and their corresponding textual descriptions.", + "bbox": [ + 88, + 462, + 485, + 674 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 691, + 183, + 708 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We provide an overview of our methodology in Figure 2. Our model incorporates a hybrid concept bank and a pretrained concept translator (Section 3.2). The concept bank consists of a set of estimizable vectors, which are refined to discover new concepts through a concept-interpretability training scheme (Section 3.3). As a result, the learned concepts are not only discriminative and orthogonal but also aligned with human-understandable semantics. The pre-trained concept translator performs concept labeling by mapping these learned, initially unknown vectors to meaningful human concepts. Finally, we apply a sparse linear layer over the similarity scores between hybrid concepts", + "bbox": [ + 89, + 719, + 485, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and images, enabling us to identify the specific concepts the model relies on for its decisions.", + "bbox": [ + 511, + 462, + 906, + 492 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 511, + 500, + 718, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Consider a dataset $\\mathcal{D} = (x,y)$ , where each image $x$ is paired with a label $y\\in \\mathcal{V}$ . We use the image encoder $\\mathcal{F}_{img}$ and text encoder $\\mathcal{F}_{txt}$ of CLIP [32] to map images and text into a shared $d$ -dimensional feature space, respectively. A pre-trained concept translator $\\mathcal{T}$ then maps these shared $d$ -dimensional features back to textual concepts. In order to store the interpretable information of the model, we construct a hybrid concept bank $E_{C}\\in \\mathbb{R}^{N_{C}\\times d}$ , comprising:", + "bbox": [ + 511, + 523, + 906, + 643 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Static Concept Bank $E_{S} \\in \\mathbb{R}^{N_{S} \\times d}$ . Contains text features $\\mathcal{F}_{txt}(c) \\in \\mathbb{R}^{d}$ extracted by the text encoder $\\mathcal{F}_{txt}$ . Each row corresponds to a predefined concept $C_{S} = \\{c_{1}, c_{2}, \\ldots, c_{N_{S}}\\}$ generated from a large language model.", + "- Dynamic Concept Bank $E_{D} \\in \\mathbb{R}^{N_{D} \\times d}$ . Consists of a set of randomly initialized learnable vectors $e_{i} \\in \\mathbb{R}^{d}$ that are optimized to capture missing visual-specific features. After training, These vectors could be mapped to the textual concept space $C_{D} = \\{c_{1}, c_{2}, \\ldots, c_{N_{D}}\\}$ by the translator $\\mathcal{T}$ , where each concept $c = \\mathcal{T}(e)$ ." + ], + "bbox": [ + 513, + 643, + 906, + 795 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overall hybrid concept bank is defined as $E_{C} = E_{S}\\oplus$ $E_{D}$ , where $\\oplus$ denotes concatenation. The total number of concepts in the bank is indicated by $N_{C} = N_{S} + N_{D}$ . To compute the concept scores, we utilize the CLIP image encoder to extract the image feature $f = \\mathcal{F}_{img}(x)\\in \\mathbb{R}^{d}$ Therein, the similarity between $f$ and $E_{C}$ is computed using projection length, resulting in the concept score $S\\in \\mathbb{R}^{N_C}$", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20181", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "which reflects the presence of particular concepts in the image $x\\colon S_x = \\| f\\| _2\\cos (f,E_C) = f\\cdot \\hat{E_C}^T$ , where $\\hat{\\cdot}$ denotes $L_{2}$ normalization and cos is the cosine similarity. Since $S$ provides a unique representation for each image, it serves as a concept bottleneck. These per-image representations can be naturally used to support an interpretable classification. Following the existing CLIP-based CBMs [26, 52, 54], we train a linear layer as the concept classifier $\\phi :\\hat{y} = \\phi (S),\\mathbb{R}^{N_C}\\to \\mathcal{V}$ to make final predictions in the label space based on the concept scores. In our perspective, we intend to improve the performance and learn abundant concepts from the visual information at the same time. This can be solved by the following optimization problem:", + "bbox": [ + 89, + 90, + 483, + 292 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\min _ {\\phi , E _ {D}} \\underset {(x, y) \\sim \\mathcal {D}} {\\mathbb {E}} \\left[ \\mathcal {L} _ {c l s} \\left(\\phi \\left(f \\cdot \\hat {E} _ {S} ^ {T} \\oplus f \\cdot \\hat {E} _ {D} ^ {T}\\right), y\\right) \\right] \\tag {1} \\\\ + \\lambda \\Omega (\\phi) + \\Gamma_ {E _ {D}} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 299, + 482, + 351 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{cls}(\\hat{y}, y)$ is the cross-entropy loss function, $\\Omega$ is a complexity measure used to promote sparsity in the linear layer, $\\lambda$ is the regularization strength and $\\Gamma$ represents the optimization objective for the Dynamic Concept Bank $E_{D}$ , introduced in Section 3.3.", + "bbox": [ + 89, + 359, + 483, + 436 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Hybrid Concept Bank Establishment", + "text_level": 1, + "bbox": [ + 89, + 444, + 411, + 460 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Previous methods, such as Labo [52], generate a set of relevant concepts and then form a concept bank through the submodular selection method. However, the predefined concepts often suffer from inherent flaws due to their limited scope, and the selection process may inadvertently omit important information. To address these issues, we propose a Hybrid Concept Bank that dynamically discovers concepts directly from the visual representation, combining both a Static Concept Bank and a Dynamic Concept Bank.", + "bbox": [ + 89, + 467, + 483, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Static Concept Bank. Following Labo, we utilize GPT-3.5 to generate a set of candidate concepts for each class name $y \\in \\mathcal{V}$ with prompts such as \"describe what the looks like\". After initially generating 500 concepts per class, we perform concept selection to ensure a balanced representation. Specifically, we select $N$ concepts for each class using both submodular and random selection methods for comparison. Assume $K = |\\mathcal{V}|$ , this process yields a total of $N_S = N \\times K$ concepts, which are then encoded by the text encoder $\\mathcal{F}_{txt}$ of CLIP and fixed during training.", + "bbox": [ + 89, + 606, + 483, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Dynamic Concept Bank. Unlike the static counterpart, the Dynamic Concept Bank is initialized with a set of concept vectors $e_d \\in \\mathbb{R}^d$ , which are optimized through the proposed Concept-Interpretability Training Scheme (Section 3.3). Thus, it can adaptively capture the unique and rich visual-specific features during the training process.", + "bbox": [ + 89, + 761, + 483, + 853 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Concept Labeling. To enhance the interpretability of learned concepts, we pre-train a concept translator using GPT-2 [31] architecture to invert the CLIP text encoder.", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following recent work [19, 24, 50], we use prefix language modeling for training. Details of the pre-training process are provided in Appendix B.2. The translator maps each concept embedding $e$ to its textual description $c$ through $c = \\mathcal{T}(e)$ , assigning semantic meaning to the learned concepts and improving interpretability.", + "bbox": [ + 511, + 90, + 905, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Concept-Interpretability Training Scheme", + "text_level": 1, + "bbox": [ + 511, + 191, + 872, + 208 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The previous CLIP-based CBMs rely solely on predefined concepts, often overlooking the potential of learning concepts directly from visual representation. Since CLIP maps both images and texts into a shared feature space, we could then leverage this capability to optimize the dynamic concept bank $E_{D}$ based on the given image embedding $f$ . To achieve this, we introduce a concept-interpretability training scheme that regularizes $E_{D}$ through multiple loss functions. The goal is to obtain a set of learnable concepts that are discriminative, diverse, and semantically aligned with human-understandable meanings, thereby enhancing both model interpretability and performance.", + "bbox": [ + 511, + 214, + 906, + 397 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Discriminability Loss. The discriminability loss ensures that learned concepts are highly activated for images of their own class and less activated for images of other classes. We first divide the dynamic concept vectors into several subsets corresponding to the classes, such that $E_D = \\{E_D^1, E_D^2, \\ldots, E_D^K\\}$ , where $K$ is the number of classes. Each subset $E_D^k$ contains $N_D / K$ dynamic concept vectors associated with class $k$ . To encourage vectors $\\mathbf{c}_k \\in E_D^k$ to learn specific information of images within class $k$ and incorporate semantic information from the class name, we define the intra-class loss:", + "bbox": [ + 511, + 398, + 908, + 566 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {i n t r a} ^ {i} = y _ {k} \\left\\{- \\hat {f} \\cdot \\hat {c} _ {k} + \\alpha \\left| m - \\hat {\\mathcal {F}} _ {t x t} (k) \\cdot \\hat {c} _ {k} \\right| \\right\\} \\quad (2)\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 573, + 906, + 599 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f = \\mathcal{F}_{img}(x)$ is the image feature, $\\mathcal{F}_{txt}(k)$ is the class embedding, $\\mathbf{c}_k$ is the sampled concept vector for class $k$ from the distribution of vectors $E_D^k$ via reparameterization $^3$ , $\\alpha$ is a hyperparameter and $m$ is a margin used to prevent excessive similarity to class $k$ , with a value of 0.85 as defined in Label-Free [26]. To ensure that concepts are distinct between different classes, we further define the inter-class loss:", + "bbox": [ + 511, + 609, + 905, + 731 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {i n t e r} ^ {i} = (y _ {k} - 1) \\left\\{- \\hat {f} \\cdot \\hat {c} _ {k} + \\alpha \\left| m - \\hat {\\mathcal {F}} _ {t x t} (k) \\cdot \\hat {c} _ {k} \\right| \\right\\} (3)\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 738, + 906, + 766 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The total discriminability loss is defined as $\\mathcal{L}_{dis} = \\mathcal{L}_{intra} + \\beta \\mathcal{L}_{inter}$ , where $\\beta$ is a balancing hyperparameter.", + "bbox": [ + 511, + 773, + 905, + 805 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Orthogonality Loss. To enhance the diversity of dynamic concept embeddings and reduce redundancy, we introduce an orthogonality loss, which comprises two components:", + "bbox": [ + 511, + 806, + 905, + 854 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "${}^{3}\\mathbf{c}_{k} = \\mu_{k} + \\sigma_{k}\\cdot \\epsilon$ , with $\\mu_{k}$ and $\\sigma_{k}$ being the mean and standard deviation of $E_D^k$ , respectively, and $\\epsilon \\sim \\mathcal{N}(0,I)$ is sampled from the standard normal distribution.", + "bbox": [ + 511, + 862, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20182", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "intra-class diversity and inter-bank diversity. First, we promote the dynamic concept embeddings within the same class to be orthogonal to each other. Specifically, we define the intra-class diversity loss as follows:", + "bbox": [ + 89, + 90, + 483, + 151 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {o r t - i n t r a}} = \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\frac {1}{(N _ {D} / K) ^ {2}} \\sum_ {i \\neq j} \\left| \\hat {\\mathbf {e}} _ {D, i} ^ {k} \\cdot \\hat {\\mathbf {e}} _ {D, j} ^ {k} \\right| \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 161, + 482, + 204 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{e}_{D,i}^{k}$ and $\\mathbf{e}_{D,j}^{k}$ are dynamic concept embeddings associated with class $k$ , $K$ is the number of classes, and $N_{D}$ is the total number of dynamic concepts. Second, to minimize redundancy between static and dynamic concepts, we define the inter-bank diversity loss as follows:", + "bbox": [ + 89, + 213, + 483, + 291 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {o r t - i n t e r}} = \\frac {1}{N _ {S} N _ {D}} \\sum_ {i = 1} ^ {N _ {S}} \\sum_ {j = 1} ^ {N _ {D}} | \\hat {\\mathbf {e}} _ {S, i} \\cdot \\hat {\\mathbf {e}} _ {D, j} | \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 297, + 482, + 340 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{e}_{S,i}$ and $\\mathbf{e}_{D,j}$ are static and dynamic concept embeddings, respectively, and $N_{S}$ is the number of static concepts. The total orthogonality loss is then defined as $\\mathcal{L}_{ort} = \\mathcal{L}_{ort - intra} + \\mathcal{L}_{ort - inter}$ .", + "bbox": [ + 89, + 349, + 483, + 411 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Distribution Alignment Loss. To enhance consistency between the distributions of dynamic and static concepts, we employ the Sinkhorn divergence [8], an entropy-regularized version of the Sinkhorn distance. We define the distribution alignment loss as: $\\mathcal{L}_{\\mathrm{align}} = \\mathfrak{S}_{div,\\epsilon}(E_D,E_S)$ , where $\\mathfrak{S}_{div,\\epsilon}$ denotes the Sinkhorn divergence, $E_{D}\\in \\mathbb{R}^{N_{D}\\times d}$ and $E_{S}\\in \\mathbb{R}^{N_{S}\\times d}$ are the dynamic and static concept embeddings, respectively. This loss encourages the distribution of dynamic concepts to align with that of static concepts, thus promoting semantic consistency and interpretability.", + "bbox": [ + 89, + 412, + 483, + 565 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Overall Loss Function. The total loss function combines the classification loss with the proposed regularization terms $\\mathcal{L} = \\mathcal{L}_{\\mathrm{cls}} + \\lambda_{\\mathrm{dis}}\\mathcal{L}_{\\mathrm{dis}} + \\lambda_{\\mathrm{ort}}\\mathcal{L}_{\\mathrm{ort}} + \\lambda_{\\mathrm{align}}\\mathcal{L}_{\\mathrm{align}}$ , where $\\lambda_{\\mathrm{dis}}$ , $\\lambda_{\\mathrm{ort}}$ , and $\\lambda_{\\mathrm{align}}$ are hyperparameters to balance the losses.", + "bbox": [ + 89, + 568, + 483, + 630 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 641, + 223, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experiments Setup", + "text_level": 1, + "bbox": [ + 89, + 665, + 272, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset. We collect image-caption pairs from MSCOCO [4] and then compile a large corpus from ConceptNet [42] and MSCOCO, which also includes concepts generated by GPT-3.5. After filtering out concepts that contain fewer than 15 words, the dataset includes 1,738,985 concepts and 566,747 image-concept pairs, used to pre-train our concept translator for translating both image and text embeddings. We also perform comprehensive experiments on 11 classification datasets, covering a broad spectrum of domains, including: ImageNet [9], CIFAR-10 and CIFAR-100 [18], Food-101 [2], FGVC-Aircraft [22], Flower-102 [25], CUB200-2011 [48], UCF-101 [41], DTD [6], HAM10000 [47] and RESISC45 [5]. Each dataset is split into training, validation, and test sets, as per standard practices.", + "bbox": [ + 89, + 688, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. To evaluate our model, we compare it with a linear probe and several interpretable methods.", + "bbox": [ + 511, + 90, + 906, + 119 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Linear Probe. Following the implementation of CLIP [32], we train a logistic regression model using cuML's L-BFGS solver with an L2 penalty term, directly from the visual representations encoded by CLIP.", + "- PCBM [54]. PCBM conceptualizes each class as a node in ConceptNet [42] and aggregates neighboring nodes as concepts. PCBM-h [54] introduces an additional classifier that maps image embeddings into the label space, serving as residual shortcuts to the original classifier.", + "- Labo [52]. Labo employs a submodular function to select concepts for each class from candidate concepts generated by LLM. Unlike other models, it uses dot-product, without any normalization, to determine the presence of particular concepts in an image.", + "- ResCBM[39]. Building on PCBM-h, ResCBM utilizes ConceptNet [42] to construct a candidate concept bank. It optimizes a set of vectors to align with some concepts from this bank, thereby enhancing the interpretability." + ], + "bbox": [ + 511, + 121, + 906, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. The hybrid concept bank is constructed with a default ratio of 0.5 between the dynamic and static concept bank, ensuring an equal number of concepts for each class. Our model is trained using the Adam [16] optimizer within the PyTorch Lightning framework. We save checkpoints that achieve the highest validation accuracy and perform evaluations on the test set. Further details can be found in the supplementary material.", + "bbox": [ + 511, + 396, + 905, + 517 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Classification Performance", + "text_level": 1, + "bbox": [ + 511, + 526, + 756, + 541 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Few-shot Comparison. To assess the performance gap between HybridCBM and 'black-box' models, we compare our method with the linear probe (LP) in a few-shot setting. We follow the few-shot evaluation protocol introduced by CLIP [32]. Specifically, we randomly sample 1, 2, 4, 8, and 16 images per class from the training set and use all available images in the fully-supervised setting. For all experiments, we use CLIP-ViT-L/14 as the backbone and compare the performance with submodular [52] and random static concept selection methods. As shown in Table 1, the mean test accuracy across all datasets indicates that HybridCBM performs similarly to LP in the fully-supervised setting and significantly surpasses LP in the few-shot scenario. Figure 3 presents detailed comparisons across eight datasets, with the last three provided in the appendix.", + "bbox": [ + 511, + 549, + 906, + 776 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "CBMs Comparison. We further compare HybridCBM with other CBMs and employ CLIP-RN50 as the backbone to ensure fairness. Additionally, since our dynamic concepts are derived from visual representation, we create a variant called CaptionCBM, where $N_{D} / K$ images for each class are sampled and translated into captions to replace the dynamic concepts. As summarized in Table 2, despite strong baselines like LaBo, and ResCBM, Hybrid-", + "bbox": [ + 511, + 779, + 908, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20183", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2b961cb45c7c0cc6dfe309408f3510cd68c230e42949ed4fe74b5ec9c066b917.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 112, + 88, + 310, + 200 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c2c6bb52da85a431e7fe3f67d30187d34b5d5ce25fd8614c68ad88b4fafcca68.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 313, + 88, + 496, + 200 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ac88102dacec52339c538b12d5b9f7db9a52e6b45550bc66c52a1182351a5f21.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 88, + 687, + 200 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fdefcbf83f30ab8b3c95a247b27f4029bc5de62a6f01a39cafea0ff01dc254ff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 89, + 875, + 200 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9606fc4fbc3e3dd0a3e03accc5cc302e432097630772505e412b70b34bbb2c48.jpg", + "image_caption": [ + "Figure 3. Comparison of test accuracy between HybridCBM with submodular and random selection methods, and Linear Probe across 8 datasets. The x-axis denotes the number of labeled images." + ], + "image_footnote": [], + "bbox": [ + 112, + 202, + 310, + 325 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/68649b7494c5ab1d544044bcc7bd73f3ae08f9c7033eccaa81fa7c4d678edeb6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 313, + 202, + 500, + 325 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/33f30f0554c71afaa22cb318a8feb446b4f955eea1f476d1a7d15405fb5d54d2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 202, + 687, + 325 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a63d6407ddb1a80a8960a1a0ef3aaeed2a76f6aa2700bbb29de5a9c5f14b97ca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 202, + 875, + 325 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/61081f84d119dc2b013b88fe13c4eac07afecdb6cda2b4f0b9f92e1cf02d6557.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method12AircraftHAM10000RESISC45
\\( concepts per class: ^{50} \\)16Full\\( concepts per class: ^{50} \\)1.0\\( concepts per class: ^{50} \\)
\\( 4 \\)87816
Linear Probe52.30\\( 64.60^6 \\)71.9478.3381.5586.840.72.59
LaboN/A\\( N/A_{0.5} \\)N/AN/AN/A85.720.6N/A
HybridCBM(submodular)58.64\\( 69.52_{0.4} \\)75.0878.7181.9386.830.75.12
HybridCBM (random)57.58\\( 66.69_{3} \\)74.6278.02Line 0r98 HybridCBM(submodular)86.820.44.16\\( Linear Probe \\)\\( Linear Probe \\)\\( Linear Probe \\)
0.3\\( Non distinct \\)\\( Non distinct \\)\\( Non distinct \\)
\\( HybridCBM (random) \\)\\( HybridCBM (random) \\)\\( HybridCBM (random) \\)
", + "bbox": [ + 94, + 359, + 906, + 460 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f5e7051f2e969b0f68cd5cb777577cfe62b7f13f7d7c6504230fee816c1630b8.jpg", + "table_caption": [ + "Table 1. Mean accuracy across all datasets at different shots." + ], + "table_footnote": [], + "table_body": "
MethodInterpretabilityCIFAR-10CIFAR-100
Linear Probe✗88.870.1
PCBM-h [54]✗87.669.9
PCBM [54]✓84.556.0
Label-Free [26]†✓86.365.2
Labo [52]✓87.665.2
ResCBM [39]†✓88.067.9
CaptionCBM✓81.860.7
HybridCBM✓88.0368.38
", + "bbox": [ + 96, + 476, + 470, + 595 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Test accuracy comparison between HybridCBM and on CIFAR-10 and CIFAR-100. † indicates the reported performance.", + "bbox": [ + 89, + 604, + 482, + 632 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CBM achieves state-of-the-art performance on CIFAR-10 and CIFAR-100, and delivers results comparable to LP.", + "bbox": [ + 89, + 646, + 482, + 676 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Interpretability of the dynamic concepts", + "text_level": 1, + "bbox": [ + 89, + 686, + 434, + 705 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In addition to evaluating classification performance, we assess the interpretability of the learned concepts in this section using the following metrics:", + "bbox": [ + 89, + 712, + 482, + 757 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- Concept Purity: This metric measures how well each concept aligns with its respective class, reflecting the ability of the concepts to capture class-specific features. For each class $k$ , we compute the average cosine similarity between the normalized embedding of the concept $\\hat{e}_k$ and the class name $\\hat{\\mathcal{F}}_{txt}(k)$ .", + "bbox": [ + 89, + 758, + 482, + 849 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nP u r i t y = \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\left(\\bar {\\hat {e}} _ {k} \\cdot \\hat {\\mathcal {F}} _ {t x t} (k)\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 186, + 862, + 482, + 904 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nS e p a r a t i o n = 1 - \\frac {2}{K (K - 1)} \\sum_ {i = 1} ^ {K} \\sum_ {j = i + 1} ^ {K} (\\bar {\\hat {e}} _ {i} \\cdot \\bar {\\hat {e}} _ {j}) \\quad (7)\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 516, + 906, + 560 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\bar{\\hat{e}}_i$ denotes the mean embedding of concepts for class $i$ . Higher separation indicates greater independence between concepts of different classes in semantic space.", + "bbox": [ + 524, + 569, + 906, + 614 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- Semantic Alignment: To ensure semantic alignment between translated concepts and intended dataset classes, we use GPT-3.5 to evaluate each translated concept's accuracy in representing a specific class with a binary \"yes\" or \"no\" response. The validation metric, \"Semantics\", is calculated as the proportion of concepts that GPT-3.5 correctly associates with the intended classes, reflecting the semantic interpretability. The detailed information is provided in Appendix B.3.", + "bbox": [ + 514, + 614, + 906, + 750 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nS e m a n t i c s = \\frac {1}{N _ {D}} \\sum_ {i = 1} ^ {N _ {D}} \\delta \\left(L L M \\left(c _ {i}, \\text {c l a s s e s}\\right), 1\\right) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 758, + 905, + 800 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $LLM(c_i, classes)$ represents the response to whether concept $c_i$ belongs to one of the classes, and $\\delta$ is a binary function that returns 1 for \"yes\" and 0 otherwise. To mitigate potential limitations of the LLM, we emphasize comparing the relative size of static and dynamic concept groups over absolute absolute values.", + "bbox": [ + 524, + 809, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20184", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e621690fc5de3503e3e90060d3600ff1861c3733160d938fdaec639b89bf9795.jpg", + "image_caption": [ + "(a) Dynamic Concept Correlation" + ], + "image_footnote": [], + "bbox": [ + 112, + 88, + 486, + 344 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b3f6ae79e26dd994a3af4e780a38a152a8fb9347cf21025c1ba293aaa8c5677e.jpg", + "image_caption": [ + "(b) Static Concept Correlation", + "Figure 4. Comparison of dynamic and static concept correlation structures on the CUB dataset. (a) Dynamic concepts: Correlation maps and classifier weights illustrate how dynamic concepts adjust their relationships with classes and among themselves. (b) Static concepts: Selected and random concepts show their contributions to class relationships and inter-concept dependencies." + ], + "image_footnote": [], + "bbox": [ + 506, + 88, + 880, + 344 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/393d7f9a2c9afec66250adfc29f2b982f32251106ad27dea16061d945e2d31ee.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
LossConcept BankPurity (%)Separation (%)Semantics (%)Precision@t (%)
LclsStatic (submodular)19.386.238.348.5
Static (random)19.282.430.947.3
Dynamic0.1100.00.010.05
Lcls+disDynamic71.976.930.034.8
Lcls+dis +ortDynamic72.077.128.934.1
Lcls+dis +ort+alignDynamic39.880.032.846.2
", + "bbox": [ + 96, + 417, + 470, + 537 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Evaluation of HybridCBM's static and dynamic concepts across various metrics.", + "bbox": [ + 89, + 542, + 482, + 570 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "- Concept-Image Relevance: As shown on the right in Figure 5, to evaluate the accuracy of the concepts aligning with the actual content of each image, we use the ability of GPT-4o in image understanding to calculate the retrieval Precision@t. For a sampled image of each class, we measure the average proportion of $t$ concepts that GPT-4o confirms as relevant.", + "bbox": [ + 89, + 585, + 482, + 689 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nP r e c i s i o n @ t = \\frac {1}{K} \\sum_ {i = 1} ^ {K} \\frac {\\delta (V L M \\left(c _ {i , t} , i m g\\right) , 1)}{t} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 696, + 480, + 738 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $VLM(c_{i,t},img)$ is the response of GPT-4o and $K$ is the total number of classes. $c_{i,t}$ is $t$ concepts of class $i$ .", + "bbox": [ + 102, + 746, + 482, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Concept Analysis. We perform concept analysis on the CUB dataset. Figure 4 compares the correlation structures of dynamic and static concepts. In Figure 4a, the ConceptClass Correlation Map shows that our dynamic concepts align closely with their corresponding classes while effectively capturing diversity across classes. The ConceptConcept Correlation Map further highlights the diversity among concepts. Notably, regions with higher coefficients", + "bbox": [ + 89, + 779, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "reflect natural semantic relationships between certain class names, such as \"red-eyed vireo\" and \"white-eyed vireo\". The pattern also seen in the Class-Class Correlation Map. The weight matrix of the classifier in Figure 4a reveals that the dynamic concept part has a more pronounced diagonal compared to the static concept, indicating that dynamic concepts are more discriminative and align better with their corresponding classes. In Figure 4b, we present similar relation maps of concepts selected by a submodular function and randomly selected concepts from each class. The selected concepts are more cohesive with their classes, whereas the randomly selected concepts, despite belonging to each class, do not exhibit clear patterns. Table 3 details the evaluation metrics proposed. For static concepts, submodular selection achieves a higher semantics score than random selection (38.3% vs 30.9%). For dynamic concepts, after introducing the loss functions $\\mathcal{L}_{\\text{dis}}$ , $\\mathcal{L}_{\\text{ort}}$ , and $\\mathcal{L}_{\\text{align}}$ , the semantic score of dynamic concepts improved to 32.8%, approaching static concepts. This indicates that our proposed loss functions effectively enhance the alignment and discriminative power of dynamic concepts.", + "bbox": [ + 511, + 421, + 906, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Interpretability. Interpretability can be divided into pattern and case interpretability. Pattern interpretability focuses on general patterns the model has learned, providing insight into its global behavior and how features influence classes. For example, a classifier's weight matrix reflecting bias toward each class. Case interpretability explains the model's decisions for individual instances, offering a local understanding of why a specific prediction was made for an image—for instance, concept scores indicating how strongly certain concepts are present. Following MMCBM [51], We", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20185", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ccb72a03ec3d450c1f70274923e0d913a747efc21559b128f3b78132f9849f36.jpg", + "image_caption": [ + "Concept Retrieval" + ], + "image_footnote": [], + "bbox": [ + 101, + 89, + 560, + 277 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "You are an expert binary concept classifier, capable of determining whether a given concept has any form of relationship with the provided image. If it has any relationship with the image, respond with \"yes\". Otherwise, respond with \"no\".", + "bbox": [ + 580, + 117, + 834, + 175 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2da2dbb27c6ff480001b88d582b14e5fe3e7d178bb1a1a39424320072032ce6d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 576, + 178, + 645, + 227 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- sometimes called the \"black albatross\"", + "- known as the black-browed albatross\n- most abundant albatross species" + ], + "bbox": [ + 645, + 186, + 843, + 220 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3caa0a1b10c2e8d0a607d6952259239cda7a8251ce05c24574bc4620f8285371.jpg", + "image_caption": [ + "Figure 5. On the left, the top-2 static and dynamic concepts for randomly selected classes across three datasets are presented, focusing on case interpretability. The concept retrieval process with vision-language model GPT-4o is illustrated on the right." + ], + "image_footnote": [], + "bbox": [ + 580, + 238, + 614, + 263 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(no, yes, yes)", + "bbox": [ + 686, + 253, + 761, + 266 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/47ddbc17d283d48679d133cd40fa5ea1b85e11758f20c8c69bc05b28626558c3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Methodconcepts per classRatio of Dynamic Concepts
00.20.40.60.81
LPN/A87.09N/AN/AN/AN/AN/A
Hybrid (submodular)583.2184.4685.4586.0886.3486.78
1084.2785.4786.2886.5186.8787.13
1584.7585.8486.4586.7887.0887.25
2084.9386.0986.4486.8486.8887.24
Hybrid (random)580.9584.0485.2885.9686.0786.53
1082.8885.2886.3086.6286.8287.13
1583.6585.8186.5486.8086.8987.25
2083.8586.0986.6686.8687.0687.24
", + "bbox": [ + 96, + 316, + 470, + 441 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/69f558f3b960723c5062798f19f1ae6940a32f103e8bd98cbdf3b3e740c260ca.jpg", + "table_caption": [ + "Table 4. Ablation average results on 10 datasets for varying the ratio of dynamic concepts under different methods." + ], + "table_footnote": [], + "table_body": "
MethodMetric\\( \\mathcal{L}_{cls} \\)\\( \\mathcal{L}_{cls+dis} \\)\\( \\mathcal{L}_{cls+dis}+ort \\)\\( \\mathcal{L}_{cls+dis}+ort+align \\)
Hybrid (submodular)accuracy84.6084.3684.2984.54
precision@t0.04734.834.146.2
Hybrid (random)accuracy84.5084.1683.9184.17
precision@t0.04434.835.547.2
", + "bbox": [ + 99, + 479, + 470, + 546 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Ablation results for varying training loss with a dynamic concept ratio of 0.5 and 10 concepts per class.", + "bbox": [ + 89, + 555, + 482, + 583 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "use the attention matrix, computed as $S_{i} \\times W$ , where $W$ is the classifier's weight, to show that the prediction is based on specific concepts within the image. As illustrated in Figure 5, we show the case interpretability by selecting the top 2 concepts for each image. We evaluate case interpretability by computing retrieval precision@t, measuring how accurately the top 5 identified concepts align with the image content. The results are shown in Table 3, where the static concepts serve as a baseline for comparison with the dynamic concepts; we focus on relative differences due to the potential limitations of GPT-4o. The dynamic concepts, across all loss functions, achieve $46.2\\%$ , closely approaching the static concepts' performance of $47.3\\%$ . This demonstrates that dynamic concepts offer high interpretability and can effectively serve as an alternative to static concepts.", + "bbox": [ + 89, + 595, + 482, + 823 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 832, + 243, + 848 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Dynamic Concept Bank Ratio. To examine the impact of dynamic concept ratios on performance, we test various ratios on 10 datasets except ImageNet, as shown in Table 4.", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As the dynamic concept ratio increases, performance will also improve. Notably, using random concepts per class results in an performance drop compared to the submodular selection method. Furthermore, employing learned concepts, even at a ratio of 0.2, significantly improves performance.", + "bbox": [ + 511, + 330, + 906, + 419 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Concept-Interpretability Loss. Table 5 reports the accuracy and precision@t (t=5) of learned concepts across different combinations of loss functions, comparing the Hybrid (submodular) and Hybrid (random) methods. A slight decrease in accuracy is observed as additional terms are added to the $\\mathcal{L}_{cls}$ but significantly improves precision@t. This suggests a trade-off, where additional constraints, while slightly reducing accuracy (e.g., accuracy drops from 84.60 to 84.29 for the Hybrid (submodular) method), enhance concept interpretability.", + "bbox": [ + 511, + 422, + 908, + 574 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 588, + 633, + 603 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduced HybridCBM, a novel hybrid concept bottleneck model designed to enhance interpretability and performance by combining both static and dynamic concept banks. Unlike traditional CBMs, our method allows the model to dynamically discover new concepts during training, thus expanding its ability to capture unique and relevant features. Our approach also incorporates a concept translator to provide human-readable interpretations for the learned dynamic concepts, making the model's decision process more accessible and understandable. Through extensive experiments on various datasets, our method achieves comparable accuracy to 'black-box' models while retaining transparency, bridging the gap between interpretability and performance. We believe that HybridCBM represents a significant step forward in the development of interpretable machine learning models. Future work could further explore methods for optimizing the accuracy of concept labeling and exploring more sophisticated strategies for dynamic concept discovery.", + "bbox": [ + 511, + 613, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20186", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgements.", + "text_level": 1, + "bbox": [ + 89, + 90, + 282, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This project is supported by NSFC Key Program 62236009, Shenzhen Fundamental Research Program (General Program) JCYJ 20210324140807019, NSFC General Program 61876032, and Key Laboratory of Data Intelligence and Cognitive Computing, Longhua District, Shenzhen.", + "bbox": [ + 89, + 113, + 485, + 185 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 209, + 187, + 224 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 3", + "[2] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 – mining discriminative components with random forests. In European Conference on Computer Vision, 2014. 5", + "[3] Chaofan Chen, Oscar Li, Daniel Tao, Alina Barnett, Cynthia Rudin, and Jonathan K Su. This looks like that: Deep learning for interpretable image recognition. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 2", + "[4] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dollar, and C. Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server, 2015. 5, 1, 2", + "[5] Gong Cheng, Junwei Han, and Xiaoqiang Lu. Remote sensing image scene classification: Benchmark and state of the art. Proceedings of the IEEE, 105(10):1865-1883, 2017. 5", + "[6] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3606-3613, 2014. 5", + "[7] Cristina Conati, Kaska Porayska-Pomsta, and Manolis Mavrikis. Ai in education needs interpretable machine learning: Lessons from open learner modelling, 2018. 1", + "[8] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2013. 5", + "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5", + "[10] Andre Esteva, Alexandre Robicquet, Bharath Ramsundar, Volodymyr Kuleshov, Mark DePristo, Katherine Chou, Claire Cui, Greg Corrado, Sebastian Thrun, and Jeff Dean. A guide to deep learning in healthcare. Nature medicine, 25 (1):24-29, 2019. 1", + "[11] David Gunning and David Aha. Darpa's explainable artificial intelligence (xai) program. AI magazine, 40(2):44-58, 2019. 1", + "[12] Zhengbao Jiang, Frank F. Xu, Jun Araki, and Graham Neubig. How can we know what language models know? Transactions of the Association for Computational Linguistics, 8: 423-438, 2020. 2" + ], + "bbox": [ + 93, + 234, + 480, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Dmitry Kazhdan, Botty Dimanov, Mateja Jamnik, Pietro Lio, and Adrian Weller. Now you see me (cme): Concept-based model extraction, 2020. 2", + "[14] Been Kim, Martin Wattenberg, Justin Gilmer, Carrie Cai, James Wexler, Fernanda Viegas, and Rory sayres. Interpretability beyond feature attribution: Quantitative testing with concept activation vectors (TCAV). In Proceedings of the 35th International Conference on Machine Learning, pages 2668-2677. PMLR, 2018. 2", + "[15] Injae Kim, Jongha Kim, Joonmyung Choi, and Hyunwoo J. Kim. Concept bottleneck with visual concept filtering for explainable medical image classification, 2023. 2", + "[16] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 5, 4", + "[17] Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. Concept bottleneck models. In International Conference on Machine Learning, pages 5338-5348. PMLR, 2020. 1, 2", + "[18] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 5", + "[19] Wei Li, Linchao Zhu, Longyin Wen, and Yi Yang. Decap: Decoding clip latents for zero-shot captioning via text-only training. In The Eleventh International Conference on Learning Representations, 2023. 3, 4, 2", + "[20] Yuanning Li, Huzheng Yang, and Shi Gu. Enhancing neural encoding models for naturalistic perception with a multilevel integration of deep neural networks and cortical networks. Science Bulletin, 2024. 1", + "[21] Max Losch, Mario Fritz, and Bernt Schiele. Interpretability beyond classification output: Semantic bottleneck networks, 2019. 2", + "[22] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 5", + "[23] Tim Miller. Explanation in artificial intelligence: Insights from the social sciences. Artificial Intelligence, 267:1-38, 2019. 1", + "[24] Ron Mokady, Amir Hertz, and Amit H Bermano. Clipcap: Clip prefix for image captioning. arXiv preprint arXiv:2111.09734, 2021. 3, 4, 2", + "[25] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In Indian Conference on Computer Vision, Graphics and Image Processing, 2008. 5", + "[26] Tuomas Oikarinen, Subhro Das, Lam M. Nguyen, and Tsui-Wei Weng. Label-free concept bottleneck models. In The Eleventh International Conference on Learning Representations, 2023. 2, 4, 6", + "[27] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul F Christiano, Jan Leike, and Ryan Lowe. Training language models to follow instructions with human feedback. In Advances in Neural Information" + ], + "bbox": [ + 516, + 92, + 905, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20187", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Processing Systems, pages 27730-27744. Curran Associates, Inc., 2022. 2", + "[28] Ahmet Murat Ozbayoglu, Mehmet Ugur Gudelek, and Omer Berat Sezer. Deep learning for financial applications: A survey. Applied Soft Computing, 93:106384, 2020. 1", + "[29] Konstantinos Panagiotis Panousis, Dino Ienc, and Diego Marcos. Sparse linear concept discovery models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2767-2771, 2023. 2", + "[30] Fabio Petroni, Tim Roktaschel, Sebastian Riedel, Patrick Lewis, Anton Bakhtin, Yuxiang Wu, and Alexander Miller. Language models as knowledge bases? In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 2463-2473, Hong Kong, China, 2019. Association for Computational Linguistics. 2", + "[31] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 2, 4", + "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2, 3, 5", + "[33] Muhammad Imran Razzak, Saeeda Naz, and Ahmad Zaib. Deep Learning for Medical Image Processing: Overview, Challenges and the Future, page 323-350. Springer International Publishing, 2017. 1", + "[34] Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. \"why should i trust you?\" explaining the predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining, pages 1135-1144, 2016. 1", + "[35] Cynthia Rudin. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature Machine Intelligence, 1(5):206-215, 2019. 1", + "[36] Gesina Schwalbe. Concept embedding analysis: A review. arXiv preprint arXiv:2203.13909, 2022. 1", + "[37] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 1", + "[38] Chenming Shang, Hengyuan Zhang, Hao Wen, and Yujiu Yang. Understanding multimodal deep neural networks: A concept selection view, 2024. 2", + "[39] Chenming Shang, Shiji Zhou, Hengyuan Zhang, Xinzhe Ni, Yujiu Yang, and Yuwang Wang. Incremental residual concept bottleneck models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11030-11040, 2024. 2, 5, 6", + "[40] Sheng Shen, Liunian Harold Li, Hao Tan, Mohit Bansal, Anna Rohrbach, Kai-Wei Chang, Zhewei Yao, and Kurt" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Keutzer. How much can CLIP benefit vision-and-language tasks? In International Conference on Learning Representations, 2022. 3", + "[41] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 5", + "[42] Robyn Speer, Joshua Chin, and Catherine Havasi. Conceptnet 5.5: An open multilingual graph of general knowledge. In Thirty-first AAAI conference on artificial intelligence, 2017. 2, 5, 1", + "[43] Divyansh Srivastava, Ge Yan, and Tsui-Wei Weng. Vlg-cbm: Training concept bottleneck models with vision-language guidance, 2024. 2", + "[44] Sandeep Subramanian, Adam Trischler, Yoshua Bengio, and Christopher J Pal. Learning general purpose distributed sentence representations via large scale multi-task learning. arXiv preprint arXiv:1804.00079, 2018. 2", + "[45] Mujeen Sung, Jinhyuk Lee, Sean Yi, Minji Jeon, Sungdong Kim, and Jaewoo Kang. Can language models be biomedical knowledge bases? In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 4723-4734, Online and Punta Cana, Dominican Republic, 2021. Association for Computational Linguistics. 2", + "[46] Yoad Tewel, Yoav Shalev, Idan Schwartz, and Lior Wolf. Zero-shot image-to-text generation for visual-semantic arithmetic. arXiv preprint arXiv:2111.14447, 2021. 3", + "[47] Philipp Tschandl, Cliff Rosendahl, and Harald Kittler. The ham10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions. Scientific data, 5(1):1-9, 2018. 5", + "[48] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 5", + "[49] Bowen Wang, Liangzhi Li, Yuta Nakashima, and Hajime Nagahara. Learning bottleneck concepts in image classification. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2", + "[50] Zirui Wang, Jiahui Yu, Adams Wei Yu, Zihang Dai, Yulia Tsvetkov, and Yuan Cao. SimVLM: Simple visual language model pretraining with weak supervision. In International Conference on Learning Representations, 2022. 4", + "[51] Yifan Wu, Yang Liu, Yue Yang, Michael S. Yao, Wenli Yang, Xuehui Shi, Lihong Yang, Dongjun Li, Yueming Liu, James C. Gee, Xuan Yang, Wenbin Wei, and Shi Gu. A concept-based interpretable model for the diagnosis of choroid neoplasias using multimodal data, 2024. 7", + "[52] Yue Yang, Artemis Panagopoulou, Shenghao Zhou, Daniel Jin, Chris Callison-Burch, and Mark Yatskar. Language in a bottle: Language model guided concept bottlenecks for interpretable image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19187-19197, 2023. 2, 4, 5, 6", + "[53] Chih-Kuan Yeh, Been Kim, Sercan Arik, Chun-Liang Li, Tomas Pfister, and Pradeep Ravikumar. On completeness-aware concept-based explanations in deep neural networks. In Advances in Neural Information Processing Systems, pages 20554-20565. Curran Associates, Inc., 2020. 2" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20188", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[54] Mert Yuksekgonul, Maggie Wang, and James Zou. Posthoc concept bottleneck models. In ICLR 2022 Workshop on PAIR²Struct, 2022. 2, 4, 5, 6", + "[55] Yu Zhang, Peter Tiño, Alex Leonardis, and Ke Tang. A survey on neural network interpretability. IEEE Transactions on Emerging Topics in Computational Intelligence, 2021. 1", + "[56] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 1" + ], + "bbox": [ + 89, + 92, + 482, + 231 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "20189", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Hybrid Concept Bottleneck Models/7a729be8-7e97-4b64-9b6a-2c339ff64815_model.json b/2025/Hybrid Concept Bottleneck Models/7a729be8-7e97-4b64-9b6a-2c339ff64815_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3acd83fd35224423fe2813b6eaddaadb013a02ee --- /dev/null +++ b/2025/Hybrid Concept Bottleneck Models/7a729be8-7e97-4b64-9b6a-2c339ff64815_model.json @@ -0,0 +1,2323 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.32, + 0.131, + 0.68, + 0.154 + ], + "angle": 0, + "content": "Hybrid Concept Bottleneck Models" + }, + { + "type": "text", + "bbox": [ + 0.268, + 0.182, + 0.73, + 0.234 + ], + "angle": 0, + "content": "Yang Liu, Tianwei Zhang, Shi Gu \nUniversity of Electronic Science and Technology of China \nShenzhen Institute for Advanced Study, UESTC" + }, + { + "type": "text", + "bbox": [ + 0.199, + 0.238, + 0.794, + 0.253 + ], + "angle": 0, + "content": "liu.yang.minéstd.uestc.edu.cn, gsbyztw@gmail.com, gus@uestc.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.287, + 0.327, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.319, + 0.486, + 0.743 + ], + "angle": 0, + "content": "Concept Bottleneck Models (CBMs) provide an interpretable framework for neural networks by mapping visual features to predefined, human-understandable concepts. However, the application of CBMs is often constrained by insufficient concept annotations. Recently, multi-modal pretrained models have shown promise in reducing annotation costs by aligning visual representations with textual concept embeddings. Nevertheless, the quality and completeness of the predefined concepts significantly affect the performance of CBMs. In this work, we propose Hybrid Concept Bottleneck Model (HybridCBM), a novel CBM framework to address the challenge of incomplete predefined concepts. Our method consists of two main components: a Static Concept Bank and a Dynamic Concept Bank. The Static Concept Bank directly leverages large language models (LLMs) for concept construction, while the Dynamic Concept Bank employs learnable vectors to capture complementary and valuable concepts continuously during training. After training, a pre-trained translator converts these vectors into human-understandable concepts, further enhancing model interpretability. HybridCBM is highly flexible and can be easily integrated with existing CBMs to improve both interpretability and performance. Experimental results1 on multiple datasets demonstrate that HybridCBM outperforms current state-of-the-art CBMs and achieves comparable results to black-box models. Additionally, we propose novel metrics to assess the quality of learned concepts, showing that they perform comparably to predefined concepts." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.773, + 0.224, + 0.789 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.799, + 0.485, + 0.877 + ], + "angle": 0, + "content": "Deep Neural Networks (DNNs) have become increasingly dominant in various fields, including computer vision, natural language processing, and speech recognition. However their complex and deep structures present a significant challenge to interpretability, often earning them the label" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.285, + 0.907, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.414, + 0.907, + 0.47 + ], + "angle": 0, + "content": "Figure 1. Our proposed concept-trainable CBM reduces the need for meticulously designed concepts by incorporating learned concepts (e.g., 'yellowish throat'), thereby enhancing interpretability and performance while minimizing manual design effort." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.478, + 0.907, + 0.675 + ], + "angle": 0, + "content": "of 'blackbox' [23]. Addressing this challenge is crucial not only for ensuring accurate predictions, but also for providing scientists and engineers with more accessible tools for designing, developing, and debugging models. Interpretability is particularly critical in domains such as healthcare [10], medicine [33], education [7], and finance [28], where high-stakes decision-making requires transparency and trustworthiness to ensure reliability and accountability. Therefore, there has been a growing focus on research aimed at developing interpretable models. While post-hoc explanation methods [34, 37, 55] have gained popularity, they often fall short by providing unfaithful representations of the model's underlying computations [35]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.902 + ], + "angle": 0, + "content": "A natural approach to improving interpretability in DNNs is to design inherently interpretable models that leverage high-level concepts to describe features extracted by black-box models [36], imitating the way humans encode and understand the world[20]. One notable approach following such methodology is the Concept Bottleneck Models (CBM) [17]. CBMs operate by first predicting concepts from image representations mid-way through the decision-making process. These models pair each image with its predefined concepts and are trained in an end-to-end manner—initially predicting concepts and subsequently using these predictions to make final decisions. While this method of concept prediction enhances model interpretability, it often faces two fundamental challenges: (1) it is perceived to underperform compared to black-box models [11]," + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.438, + 0.901 + ], + "angle": 0, + "content": "\\(^{1}\\)Code available at https://github.com/ly1998117/HybridCBM" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20179" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.122 + ], + "angle": 0, + "content": "and (2) the requirement for precise annotations for each concept significantly limits their broader application." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.127, + 0.485, + 0.535 + ], + "angle": 0, + "content": "To mitigate these shortcomings, recent research has explored multimodal pretrained model-based CBMs [15, 52, 54], leveraging CLIP [32] to establish correspondences between visual and textual representations, thereby reducing the reliance on costly concept annotations. Specifically, CLIP encodes images and human-understandable concepts into a shared feature space, enabling the creation of a concept bottleneck by directly mapping visual representations to concept embeddings instead of through manual annotations. Regarding the trade-off between interpretability and performance, some studies [15, 38, 52] have emphasized the construction of high-quality concept banks through efficient concept selection. However, establishing a comprehensive concept bank remains a significant challenge [53], often constrained by the inherent limitations in language expressiveness and precision. In response, other research efforts have explored the integration of a residual linear layer to address the issue of incomplete concept extraction [39, 54], albeit at the cost of interpretability. For example, ResCBM [39] introduces estimizable vectors to capture missing concepts, utilizing a candidate concept bank in the residual component to regularizing these vectors toward specific candidates during training. However, the model is limited by the scope of the candidate concept bank, restricting its ability to learn concepts outside the predefined set. Additionally, the effectiveness of the residual structure remains an open question." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.485, + 0.902 + ], + "angle": 0, + "content": "In this work, we propose the Hybrid Concept Bottleneck Model (HybridCBM), specifically designed to address the challenges of incomplete concept representation, predefined concept bank dependence, and scalability issue by dynamically discovering new concepts directly from visual representations. Different from prior methods that rely on residual structure or external concept banks, our approach incorporates a hybrid concept bank comprising both a static and a dynamic concept bank. Initially, we establish the static concept bank by leveraging a large language model (LLM) like GPT-3.5 [27, 31], which is celebrated for its broad world knowledge [12, 30, 45]. For instance, in Figure 1, when prompted about a palm warbler, GPT-3.5 provides information such as 'feeds on the ground'. This textual descriptions is then encoded into embeddings by the text encoder of CLIP. Next, we initialize a set of estimizable vectors that form the dynamic concept bank. These vectors are refined to identify concepts present in the input image but absent from the static concept bank. To further enhance the concept discovery, we pre-train a concept translator, such as GPT-2 [31], to translate these newly discovered unknown concepts into textual descriptions, a process we term as 'concept labeling'. To evaluate the understandability, visual relevance and factual accuracy of the dynamic concepts, we" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.137 + ], + "angle": 0, + "content": "further design a set of evaluation metrics, some of which are based on Vision-Language Model (VLM) GPT-4o\\(^2\\). Our main contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.138, + 0.907, + 0.228 + ], + "angle": 0, + "content": "1. Innovative Hybrid Concept Bank: Our model introduces a hybrid concept bank that combines static and dynamic concepts, allowing it to adapt and refine its interpretative capabilities dynamically. This structure leverages both predefined expert knowledge while enabling the discovery of new relevant concepts." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.229, + 0.905, + 0.319 + ], + "angle": 0, + "content": "2. Seamless Integration of LLM Technologies: By employing advanced language models (LLM for defining static concepts, translating new concepts and VLM for evaluation) alongside the visual-textual bridging capabilities of CLIP, our model bridges the gap between visual representation and textual interpretation." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.32, + 0.905, + 0.38 + ], + "angle": 0, + "content": "3. Enhanced Interpretability and Performance: Our approach maintains high interpretability through transparent concept labeling, while improving performance by discovering new concepts from visual information." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.138, + 0.907, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.396, + 0.655, + 0.411 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.422, + 0.908, + 0.875 + ], + "angle": 0, + "content": "Interpretable Neural Networks. One way to build an interpretable neural networks is the use of a concept-based explanation [3, 13, 15, 21, 29, 38, 43, 49]. Concept Bottleneck Models (CBMs) [17] are among the most popular approaches for making predictions based on human-interpretable concepts but require labor-intensive concept annotations for each image. Similarly, Concept Activation Vectors (CAVs) [14] represent concepts as normal vectors to the decision boundaries that separate positive and negative samples of a concept, though they also need extra datasets to train SVMs for each concept. Owing to the simple structure based on human-defined concepts, there have been studies aiming to alleviate these drawbacks, e.g., costly concept annotation and trade-off between interpretability and performance. Post-hoc Concept Bottleneck (PCBM) [54] utilizes ConceptNet [42] to obtain concepts and Leveraging CLIP to generate bottlenecks by projecting visual representations into the concept subspace. However, PCBM-h [54] incorporates a residual linear predictor that compensates for missing concepts by directly adding logits to PCBM's predictor. While this modification helps to recover accuracy, it does so at the expense of interpretability. The label-free CBM [26] uses CLIP for concept annotation, allowing arbitrary visual backbone to be transformed into an CBM without the need for labeled concept data. Labo [52] concentrates on concepts selection after generating candidate concepts through LLM. ResCBM [39] builds on PCBM-h's residual structure by using a candidate concept bank to guide and refine estimizable vectors. However, the model's expressive potential is limited by the extent of the candidate concept bank." + }, + { + "type": "page_footnote", + "bbox": [ + 0.531, + 0.887, + 0.744, + 0.901 + ], + "angle": 0, + "content": "\\(^{2}\\)https://openai.com/index/hello-gpt-4o/" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20180" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.115, + 0.088, + 0.885, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.384, + 0.908, + 0.457 + ], + "angle": 0, + "content": "Figure 2. Overview of the Hybrid Concept Bottleneck Model (HybridCBM), which is designed for interpretable concept learning and labeling. First, a Concept Translator is pre-trained to do concept-labeling. Next, a pretrained alignment model (CLIP) is used to obtain embeddings of predefined concepts, forming the Static Concept Bank. The Dynamic Concept Bank learns missing concepts and aligns its distribution with the Static Concept Bank to create the final concept bottleneck. Finally, an interpretable concept-based predictor classifies targets based on the concept scores." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.463, + 0.486, + 0.675 + ], + "angle": 0, + "content": "CLIP in Image Captioning. Multimodal pretrained models like CLIP [32], trained on over 400 million image-text pairs using unsupervised contrastive loss, exhibit impressive capabilities in image captioning tasks [1, 19, 46]. Previous studies, such as [24] and [40], have primarily used CLIP as a visual encoder for captioning tasks, while [19] projects visual embeddings into CLIP's text embedding space to preserve visual information. However, these approaches often underutilize CLIP's aligned multi-modal latent space. Similar to [19, 24], we collect image-caption pairs and a vast corpus to train a GPT-2 from scratch as our concept translator, fully leveraging CLIP's latent space and bridging the gap between image and concept embeddings and their corresponding textual descriptions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.693, + 0.184, + 0.709 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.486, + 0.902 + ], + "angle": 0, + "content": "We provide an overview of our methodology in Figure 2. Our model incorporates a hybrid concept bank and a pretrained concept translator (Section 3.2). The concept bank consists of a set of estimizable vectors, which are refined to discover new concepts through a concept-interpretability training scheme (Section 3.3). As a result, the learned concepts are not only discriminative and orthogonal but also aligned with human-understandable semantics. The pre-trained concept translator performs concept labeling by mapping these learned, initially unknown vectors to meaningful human concepts. Finally, we apply a sparse linear layer over the similarity scores between hybrid concepts" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.463, + 0.907, + 0.493 + ], + "angle": 0, + "content": "and images, enabling us to identify the specific concepts the model relies on for its decisions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.5, + 0.719, + 0.515 + ], + "angle": 0, + "content": "3.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.907, + 0.645 + ], + "angle": 0, + "content": "Consider a dataset \\(\\mathcal{D} = (x,y)\\), where each image \\(x\\) is paired with a label \\(y\\in \\mathcal{V}\\). We use the image encoder \\(\\mathcal{F}_{img}\\) and text encoder \\(\\mathcal{F}_{txt}\\) of CLIP [32] to map images and text into a shared \\(d\\)-dimensional feature space, respectively. A pre-trained concept translator \\(\\mathcal{T}\\) then maps these shared \\(d\\)-dimensional features back to textual concepts. In order to store the interpretable information of the model, we construct a hybrid concept bank \\(E_{C}\\in \\mathbb{R}^{N_{C}\\times d}\\), comprising:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.644, + 0.907, + 0.705 + ], + "angle": 0, + "content": "- Static Concept Bank \\(E_{S} \\in \\mathbb{R}^{N_{S} \\times d}\\). Contains text features \\(\\mathcal{F}_{txt}(c) \\in \\mathbb{R}^{d}\\) extracted by the text encoder \\(\\mathcal{F}_{txt}\\). Each row corresponds to a predefined concept \\(C_{S} = \\{c_{1}, c_{2}, \\ldots, c_{N_{S}}\\}\\) generated from a large language model." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.705, + 0.907, + 0.796 + ], + "angle": 0, + "content": "- Dynamic Concept Bank \\( E_{D} \\in \\mathbb{R}^{N_{D} \\times d} \\). Consists of a set of randomly initialized learnable vectors \\( e_{i} \\in \\mathbb{R}^{d} \\) that are optimized to capture missing visual-specific features. After training, These vectors could be mapped to the textual concept space \\( C_{D} = \\{c_{1}, c_{2}, \\ldots, c_{N_{D}}\\} \\) by the translator \\( \\mathcal{T} \\), where each concept \\( c = \\mathcal{T}(e) \\)." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.644, + 0.907, + 0.796 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "The overall hybrid concept bank is defined as \\(E_{C} = E_{S}\\oplus\\) \\(E_{D}\\) , where \\(\\oplus\\) denotes concatenation. The total number of concepts in the bank is indicated by \\(N_{C} = N_{S} + N_{D}\\) . To compute the concept scores, we utilize the CLIP image encoder to extract the image feature \\(f = \\mathcal{F}_{img}(x)\\in \\mathbb{R}^{d}\\) Therein, the similarity between \\(f\\) and \\(E_{C}\\) is computed using projection length, resulting in the concept score \\(S\\in \\mathbb{R}^{N_C}\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20181" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.293 + ], + "angle": 0, + "content": "which reflects the presence of particular concepts in the image \\(x\\colon S_x = \\| f\\| _2\\cos (f,E_C) = f\\cdot \\hat{E_C}^T\\) , where \\(\\hat{\\cdot}\\) denotes \\(L_{2}\\) normalization and cos is the cosine similarity. Since \\(S\\) provides a unique representation for each image, it serves as a concept bottleneck. These per-image representations can be naturally used to support an interpretable classification. Following the existing CLIP-based CBMs [26, 52, 54], we train a linear layer as the concept classifier \\(\\phi :\\hat{y} = \\phi (S),\\mathbb{R}^{N_C}\\to \\mathcal{V}\\) to make final predictions in the label space based on the concept scores. In our perspective, we intend to improve the performance and learn abundant concepts from the visual information at the same time. This can be solved by the following optimization problem:" + }, + { + "type": "equation", + "bbox": [ + 0.114, + 0.3, + 0.483, + 0.352 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\min _ {\\phi , E _ {D}} \\underset {(x, y) \\sim \\mathcal {D}} {\\mathbb {E}} \\left[ \\mathcal {L} _ {c l s} \\left(\\phi \\left(f \\cdot \\hat {E} _ {S} ^ {T} \\oplus f \\cdot \\hat {E} _ {D} ^ {T}\\right), y\\right) \\right] \\tag {1} \\\\ + \\lambda \\Omega (\\phi) + \\Gamma_ {E _ {D}} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.361, + 0.484, + 0.437 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{cls}(\\hat{y}, y)\\) is the cross-entropy loss function, \\(\\Omega\\) is a complexity measure used to promote sparsity in the linear layer, \\(\\lambda\\) is the regularization strength and \\(\\Gamma\\) represents the optimization objective for the Dynamic Concept Bank \\(E_{D}\\), introduced in Section 3.3." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.445, + 0.413, + 0.461 + ], + "angle": 0, + "content": "3.2. Hybrid Concept Bank Establishment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.468, + 0.484, + 0.605 + ], + "angle": 0, + "content": "Previous methods, such as Labo [52], generate a set of relevant concepts and then form a concept bank through the submodular selection method. However, the predefined concepts often suffer from inherent flaws due to their limited scope, and the selection process may inadvertently omit important information. To address these issues, we propose a Hybrid Concept Bank that dynamically discovers concepts directly from the visual representation, combining both a Static Concept Bank and a Dynamic Concept Bank." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.607, + 0.484, + 0.76 + ], + "angle": 0, + "content": "Static Concept Bank. Following Labo, we utilize GPT-3.5 to generate a set of candidate concepts for each class name \\( y \\in \\mathcal{V} \\) with prompts such as \"describe what the looks like\". After initially generating 500 concepts per class, we perform concept selection to ensure a balanced representation. Specifically, we select \\( N \\) concepts for each class using both submodular and random selection methods for comparison. Assume \\( K = |\\mathcal{V}| \\), this process yields a total of \\( N_S = N \\times K \\) concepts, which are then encoded by the text encoder \\( \\mathcal{F}_{txt} \\) of CLIP and fixed during training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.762, + 0.484, + 0.854 + ], + "angle": 0, + "content": "Dynamic Concept Bank. Unlike the static counterpart, the Dynamic Concept Bank is initialized with a set of concept vectors \\( e_d \\in \\mathbb{R}^d \\), which are optimized through the proposed Concept-Interpretability Training Scheme (Section 3.3). Thus, it can adaptively capture the unique and rich visual-specific features during the training process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Concept Labeling. To enhance the interpretability of learned concepts, we pre-train a concept translator using GPT-2 [31] architecture to invert the CLIP text encoder." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.183 + ], + "angle": 0, + "content": "Following recent work [19, 24, 50], we use prefix language modeling for training. Details of the pre-training process are provided in Appendix B.2. The translator maps each concept embedding \\( e \\) to its textual description \\( c \\) through \\( c = \\mathcal{T}(e) \\), assigning semantic meaning to the learned concepts and improving interpretability." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.192, + 0.874, + 0.209 + ], + "angle": 0, + "content": "3.3. Concept-Interpretability Training Scheme" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.215, + 0.907, + 0.398 + ], + "angle": 0, + "content": "The previous CLIP-based CBMs rely solely on predefined concepts, often overlooking the potential of learning concepts directly from visual representation. Since CLIP maps both images and texts into a shared feature space, we could then leverage this capability to optimize the dynamic concept bank \\( E_{D} \\) based on the given image embedding \\( f \\). To achieve this, we introduce a concept-interpretability training scheme that regularizes \\( E_{D} \\) through multiple loss functions. The goal is to obtain a set of learnable concepts that are discriminative, diverse, and semantically aligned with human-understandable meanings, thereby enhancing both model interpretability and performance." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.4, + 0.909, + 0.567 + ], + "angle": 0, + "content": "Discriminability Loss. The discriminability loss ensures that learned concepts are highly activated for images of their own class and less activated for images of other classes. We first divide the dynamic concept vectors into several subsets corresponding to the classes, such that \\( E_D = \\{E_D^1, E_D^2, \\ldots, E_D^K\\} \\), where \\( K \\) is the number of classes. Each subset \\( E_D^k \\) contains \\( N_D / K \\) dynamic concept vectors associated with class \\( k \\). To encourage vectors \\( \\mathbf{c}_k \\in E_D^k \\) to learn specific information of images within class \\( k \\) and incorporate semantic information from the class name, we define the intra-class loss:" + }, + { + "type": "equation", + "bbox": [ + 0.541, + 0.574, + 0.907, + 0.601 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {i n t r a} ^ {i} = y _ {k} \\left\\{- \\hat {f} \\cdot \\hat {c} _ {k} + \\alpha \\left| m - \\hat {\\mathcal {F}} _ {t x t} (k) \\cdot \\hat {c} _ {k} \\right| \\right\\} \\quad (2)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.61, + 0.906, + 0.732 + ], + "angle": 0, + "content": "where \\( f = \\mathcal{F}_{img}(x) \\) is the image feature, \\( \\mathcal{F}_{txt}(k) \\) is the class embedding, \\( \\mathbf{c}_k \\) is the sampled concept vector for class \\( k \\) from the distribution of vectors \\( E_D^k \\) via reparameterization\\(^3\\), \\( \\alpha \\) is a hyperparameter and \\( m \\) is a margin used to prevent excessive similarity to class \\( k \\), with a value of 0.85 as defined in Label-Free [26]. To ensure that concepts are distinct between different classes, we further define the inter-class loss:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.739, + 0.907, + 0.767 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {i n t e r} ^ {i} = (y _ {k} - 1) \\left\\{- \\hat {f} \\cdot \\hat {c} _ {k} + \\alpha \\left| m - \\hat {\\mathcal {F}} _ {t x t} (k) \\cdot \\hat {c} _ {k} \\right| \\right\\} (3)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.775, + 0.906, + 0.806 + ], + "angle": 0, + "content": "The total discriminability loss is defined as \\(\\mathcal{L}_{dis} = \\mathcal{L}_{intra} + \\beta \\mathcal{L}_{inter}\\), where \\(\\beta\\) is a balancing hyperparameter." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.808, + 0.906, + 0.856 + ], + "angle": 0, + "content": "Orthogonality Loss. To enhance the diversity of dynamic concept embeddings and reduce redundancy, we introduce an orthogonality loss, which comprises two components:" + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.863, + 0.907, + 0.901 + ], + "angle": 0, + "content": "\\({}^{3}\\mathbf{c}_{k} = \\mu_{k} + \\sigma_{k}\\cdot \\epsilon\\), with \\(\\mu_{k}\\) and \\(\\sigma_{k}\\) being the mean and standard deviation of \\(E_D^k\\), respectively, and \\(\\epsilon \\sim \\mathcal{N}(0,I)\\) is sampled from the standard normal distribution." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20182" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.152 + ], + "angle": 0, + "content": "intra-class diversity and inter-bank diversity. First, we promote the dynamic concept embeddings within the same class to be orthogonal to each other. Specifically, we define the intra-class diversity loss as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.113, + 0.162, + 0.483, + 0.205 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {o r t - i n t r a}} = \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\frac {1}{(N _ {D} / K) ^ {2}} \\sum_ {i \\neq j} \\left| \\hat {\\mathbf {e}} _ {D, i} ^ {k} \\cdot \\hat {\\mathbf {e}} _ {D, j} ^ {k} \\right| \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.214, + 0.485, + 0.292 + ], + "angle": 0, + "content": "where \\(\\mathbf{e}_{D,i}^{k}\\) and \\(\\mathbf{e}_{D,j}^{k}\\) are dynamic concept embeddings associated with class \\(k\\), \\(K\\) is the number of classes, and \\(N_{D}\\) is the total number of dynamic concepts. Second, to minimize redundancy between static and dynamic concepts, we define the inter-bank diversity loss as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.299, + 0.483, + 0.342 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {o r t - i n t e r}} = \\frac {1}{N _ {S} N _ {D}} \\sum_ {i = 1} ^ {N _ {S}} \\sum_ {j = 1} ^ {N _ {D}} | \\hat {\\mathbf {e}} _ {S, i} \\cdot \\hat {\\mathbf {e}} _ {D, j} | \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.351, + 0.484, + 0.412 + ], + "angle": 0, + "content": "where \\(\\mathbf{e}_{S,i}\\) and \\(\\mathbf{e}_{D,j}\\) are static and dynamic concept embeddings, respectively, and \\(N_{S}\\) is the number of static concepts. The total orthogonality loss is then defined as \\(\\mathcal{L}_{ort} = \\mathcal{L}_{ort - intra} + \\mathcal{L}_{ort - inter}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.414, + 0.484, + 0.566 + ], + "angle": 0, + "content": "Distribution Alignment Loss. To enhance consistency between the distributions of dynamic and static concepts, we employ the Sinkhorn divergence [8], an entropy-regularized version of the Sinkhorn distance. We define the distribution alignment loss as: \\(\\mathcal{L}_{\\mathrm{align}} = \\mathfrak{S}_{div,\\epsilon}(E_D,E_S)\\), where \\(\\mathfrak{S}_{div,\\epsilon}\\) denotes the Sinkhorn divergence, \\(E_{D}\\in \\mathbb{R}^{N_{D}\\times d}\\) and \\(E_{S}\\in \\mathbb{R}^{N_{S}\\times d}\\) are the dynamic and static concept embeddings, respectively. This loss encourages the distribution of dynamic concepts to align with that of static concepts, thus promoting semantic consistency and interpretability." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.484, + 0.631 + ], + "angle": 0, + "content": "Overall Loss Function. The total loss function combines the classification loss with the proposed regularization terms \\(\\mathcal{L} = \\mathcal{L}_{\\mathrm{cls}} + \\lambda_{\\mathrm{dis}}\\mathcal{L}_{\\mathrm{dis}} + \\lambda_{\\mathrm{ort}}\\mathcal{L}_{\\mathrm{ort}} + \\lambda_{\\mathrm{align}}\\mathcal{L}_{\\mathrm{align}}\\), where \\(\\lambda_{\\mathrm{dis}}\\), \\(\\lambda_{\\mathrm{ort}}\\), and \\(\\lambda_{\\mathrm{align}}\\) are hyperparameters to balance the losses." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.642, + 0.224, + 0.659 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.666, + 0.274, + 0.683 + ], + "angle": 0, + "content": "4.1. Experiments Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Dataset. We collect image-caption pairs from MSCOCO [4] and then compile a large corpus from ConceptNet [42] and MSCOCO, which also includes concepts generated by GPT-3.5. After filtering out concepts that contain fewer than 15 words, the dataset includes 1,738,985 concepts and 566,747 image-concept pairs, used to pre-train our concept translator for translating both image and text embeddings. We also perform comprehensive experiments on 11 classification datasets, covering a broad spectrum of domains, including: ImageNet [9], CIFAR-10 and CIFAR-100 [18], Food-101 [2], FGVC-Aircraft [22], Flower-102 [25], CUB200-2011 [48], UCF-101 [41], DTD [6], HAM10000 [47] and RESISC45 [5]. Each dataset is split into training, validation, and test sets, as per standard practices." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.121 + ], + "angle": 0, + "content": "Baselines. To evaluate our model, we compare it with a linear probe and several interpretable methods." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.122, + 0.907, + 0.182 + ], + "angle": 0, + "content": "- Linear Probe. Following the implementation of CLIP [32], we train a logistic regression model using cuML's L-BFGS solver with an L2 penalty term, directly from the visual representations encoded by CLIP." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.182, + 0.906, + 0.258 + ], + "angle": 0, + "content": "- PCBM [54]. PCBM conceptualizes each class as a node in ConceptNet [42] and aggregates neighboring nodes as concepts. PCBM-h [54] introduces an additional classifier that maps image embeddings into the label space, serving as residual shortcuts to the original classifier." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.258, + 0.907, + 0.333 + ], + "angle": 0, + "content": "- Labo [52]. Labo employs a submodular function to select concepts for each class from candidate concepts generated by LLM. Unlike other models, it uses dot-product, without any normalization, to determine the presence of particular concepts in an image." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.333, + 0.906, + 0.395 + ], + "angle": 0, + "content": "- ResCBM[39]. Building on PCBM-h, ResCBM utilizes ConceptNet [42] to construct a candidate concept bank. It optimizes a set of vectors to align with some concepts from this bank, thereby enhancing the interpretability." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.122, + 0.907, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.397, + 0.906, + 0.518 + ], + "angle": 0, + "content": "Implementation Details. The hybrid concept bank is constructed with a default ratio of 0.5 between the dynamic and static concept bank, ensuring an equal number of concepts for each class. Our model is trained using the Adam [16] optimizer within the PyTorch Lightning framework. We save checkpoints that achieve the highest validation accuracy and perform evaluations on the test set. Further details can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.527, + 0.757, + 0.542 + ], + "angle": 0, + "content": "4.2. Classification Performance" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.55, + 0.907, + 0.777 + ], + "angle": 0, + "content": "Few-shot Comparison. To assess the performance gap between HybridCBM and 'black-box' models, we compare our method with the linear probe (LP) in a few-shot setting. We follow the few-shot evaluation protocol introduced by CLIP [32]. Specifically, we randomly sample 1, 2, 4, 8, and 16 images per class from the training set and use all available images in the fully-supervised setting. For all experiments, we use CLIP-ViT-L/14 as the backbone and compare the performance with submodular [52] and random static concept selection methods. As shown in Table 1, the mean test accuracy across all datasets indicates that HybridCBM performs similarly to LP in the fully-supervised setting and significantly surpasses LP in the few-shot scenario. Figure 3 presents detailed comparisons across eight datasets, with the last three provided in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.902 + ], + "angle": 0, + "content": "CBMs Comparison. We further compare HybridCBM with other CBMs and employ CLIP-RN50 as the backbone to ensure fairness. Additionally, since our dynamic concepts are derived from visual representation, we create a variant called CaptionCBM, where \\( N_{D} / K \\) images for each class are sampled and translated into captions to replace the dynamic concepts. As summarized in Table 2, despite strong baselines like LaBo, and ResCBM, Hybrid-" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20183" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.114, + 0.089, + 0.312, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.089, + 0.498, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.089, + 0.688, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.09, + 0.877, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.203, + 0.312, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.203, + 0.5, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.203, + 0.689, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.203, + 0.877, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.331, + 0.907, + 0.358 + ], + "angle": 0, + "content": "Figure 3. Comparison of test accuracy between HybridCBM with submodular and random selection methods, and Linear Probe across 8 datasets. The x-axis denotes the number of labeled images." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.36, + 0.907, + 0.461 + ], + "angle": 0, + "content": "
Method12AircraftHAM10000RESISC45
\\( concepts per class: ^{50} \\)16Full\\( concepts per class: ^{50} \\)1.0\\( concepts per class: ^{50} \\)
\\( 4 \\)87816
Linear Probe52.30\\( 64.60^6 \\)71.9478.3381.5586.840.72.59
LaboN/A\\( N/A_{0.5} \\)N/AN/AN/A85.720.6N/A
HybridCBM(submodular)58.64\\( 69.52_{0.4} \\)75.0878.7181.9386.830.75.12
HybridCBM (random)57.58\\( 66.69_{3} \\)74.6278.02Line 0r98 HybridCBM(submodular)86.820.44.16\\( Linear Probe \\)\\( Linear Probe \\)\\( Linear Probe \\)
0.3\\( Non distinct \\)\\( Non distinct \\)\\( Non distinct \\)
\\( HybridCBM (random) \\)\\( HybridCBM (random) \\)\\( HybridCBM (random) \\)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.105, + 0.461, + 0.472, + 0.473 + ], + "angle": 0, + "content": "Table 1. Mean accuracy across all datasets at different shots." + }, + { + "type": "table", + "bbox": [ + 0.098, + 0.477, + 0.472, + 0.597 + ], + "angle": 0, + "content": "
MethodInterpretabilityCIFAR-10CIFAR-100
Linear Probe✗88.870.1
PCBM-h [54]✗87.669.9
PCBM [54]✓84.556.0
Label-Free [26]†✓86.365.2
Labo [52]✓87.665.2
ResCBM [39]†✓88.067.9
CaptionCBM✓81.860.7
HybridCBM✓88.0368.38
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.605, + 0.483, + 0.633 + ], + "angle": 0, + "content": "Table 2. Test accuracy comparison between HybridCBM and on CIFAR-10 and CIFAR-100. † indicates the reported performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.647, + 0.483, + 0.678 + ], + "angle": 0, + "content": "CBM achieves state-of-the-art performance on CIFAR-10 and CIFAR-100, and delivers results comparable to LP." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.688, + 0.436, + 0.706 + ], + "angle": 0, + "content": "4.3. Interpretability of the dynamic concepts" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.713, + 0.483, + 0.758 + ], + "angle": 0, + "content": "In addition to evaluating classification performance, we assess the interpretability of the learned concepts in this section using the following metrics:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.759, + 0.483, + 0.85 + ], + "angle": 0, + "content": "- Concept Purity: This metric measures how well each concept aligns with its respective class, reflecting the ability of the concepts to capture class-specific features. For each class \\( k \\), we compute the average cosine similarity between the normalized embedding of the concept \\( \\hat{e}_k \\) and the class name \\( \\hat{\\mathcal{F}}_{txt}(k) \\)." + }, + { + "type": "equation", + "bbox": [ + 0.187, + 0.863, + 0.483, + 0.905 + ], + "angle": 0, + "content": "\\[\nP u r i t y = \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\left(\\bar {\\hat {e}} _ {k} \\cdot \\hat {\\mathcal {F}} _ {t x t} (k)\\right) \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.543, + 0.517, + 0.907, + 0.561 + ], + "angle": 0, + "content": "\\[\nS e p a r a t i o n = 1 - \\frac {2}{K (K - 1)} \\sum_ {i = 1} ^ {K} \\sum_ {j = i + 1} ^ {K} (\\bar {\\hat {e}} _ {i} \\cdot \\bar {\\hat {e}} _ {j}) \\quad (7)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.57, + 0.907, + 0.615 + ], + "angle": 0, + "content": "where \\(\\bar{\\hat{e}}_i\\) denotes the mean embedding of concepts for class \\(i\\). Higher separation indicates greater independence between concepts of different classes in semantic space." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.616, + 0.907, + 0.751 + ], + "angle": 0, + "content": "- Semantic Alignment: To ensure semantic alignment between translated concepts and intended dataset classes, we use GPT-3.5 to evaluate each translated concept's accuracy in representing a specific class with a binary \"yes\" or \"no\" response. The validation metric, \"Semantics\", is calculated as the proportion of concepts that GPT-3.5 correctly associates with the intended classes, reflecting the semantic interpretability. The detailed information is provided in Appendix B.3." + }, + { + "type": "equation", + "bbox": [ + 0.546, + 0.76, + 0.906, + 0.801 + ], + "angle": 0, + "content": "\\[\nS e m a n t i c s = \\frac {1}{N _ {D}} \\sum_ {i = 1} ^ {N _ {D}} \\delta \\left(L L M \\left(c _ {i}, \\text {c l a s s e s}\\right), 1\\right) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.81, + 0.907, + 0.901 + ], + "angle": 0, + "content": "where \\( LLM(c_i, classes) \\) represents the response to whether concept \\( c_i \\) belongs to one of the classes, and \\( \\delta \\) is a binary function that returns 1 for \"yes\" and 0 otherwise. To mitigate potential limitations of the LLM, we emphasize comparing the relative size of static and dynamic concept groups over absolute absolute values." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20184" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.089, + 0.487, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.223, + 0.346, + 0.385, + 0.356 + ], + "angle": 0, + "content": "(a) Dynamic Concept Correlation" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.089, + 0.882, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.607, + 0.346, + 0.755, + 0.356 + ], + "angle": 0, + "content": "(b) Static Concept Correlation" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.364, + 0.907, + 0.407 + ], + "angle": 0, + "content": "Figure 4. Comparison of dynamic and static concept correlation structures on the CUB dataset. (a) Dynamic concepts: Correlation maps and classifier weights illustrate how dynamic concepts adjust their relationships with classes and among themselves. (b) Static concepts: Selected and random concepts show their contributions to class relationships and inter-concept dependencies." + }, + { + "type": "table", + "bbox": [ + 0.098, + 0.419, + 0.472, + 0.538 + ], + "angle": 0, + "content": "
LossConcept BankPurity (%)Separation (%)Semantics (%)Precision@t (%)
LclsStatic (submodular)19.386.238.348.5
Static (random)19.282.430.947.3
Dynamic0.1100.00.010.05
Lcls+disDynamic71.976.930.034.8
Lcls+dis +ortDynamic72.077.128.934.1
Lcls+dis +ort+alignDynamic39.880.032.846.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.544, + 0.483, + 0.571 + ], + "angle": 0, + "content": "Table 3. Evaluation of HybridCBM's static and dynamic concepts across various metrics." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.586, + 0.483, + 0.69 + ], + "angle": 0, + "content": "- Concept-Image Relevance: As shown on the right in Figure 5, to evaluate the accuracy of the concepts aligning with the actual content of each image, we use the ability of GPT-4o in image understanding to calculate the retrieval Precision@t. For a sampled image of each class, we measure the average proportion of \\( t \\) concepts that GPT-4o confirms as relevant." + }, + { + "type": "equation", + "bbox": [ + 0.127, + 0.698, + 0.482, + 0.739 + ], + "angle": 0, + "content": "\\[\nP r e c i s i o n @ t = \\frac {1}{K} \\sum_ {i = 1} ^ {K} \\frac {\\delta (V L M \\left(c _ {i , t} , i m g\\right) , 1)}{t} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.747, + 0.483, + 0.778 + ], + "angle": 0, + "content": "where \\( VLM(c_{i,t},img) \\) is the response of GPT-4o and \\( K \\) is the total number of classes. \\( c_{i,t} \\) is \\( t \\) concepts of class \\( i \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Concept Analysis. We perform concept analysis on the CUB dataset. Figure 4 compares the correlation structures of dynamic and static concepts. In Figure 4a, the ConceptClass Correlation Map shows that our dynamic concepts align closely with their corresponding classes while effectively capturing diversity across classes. The ConceptConcept Correlation Map further highlights the diversity among concepts. Notably, regions with higher coefficients" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.422, + 0.907, + 0.74 + ], + "angle": 0, + "content": "reflect natural semantic relationships between certain class names, such as \"red-eyed vireo\" and \"white-eyed vireo\". The pattern also seen in the Class-Class Correlation Map. The weight matrix of the classifier in Figure 4a reveals that the dynamic concept part has a more pronounced diagonal compared to the static concept, indicating that dynamic concepts are more discriminative and align better with their corresponding classes. In Figure 4b, we present similar relation maps of concepts selected by a submodular function and randomly selected concepts from each class. The selected concepts are more cohesive with their classes, whereas the randomly selected concepts, despite belonging to each class, do not exhibit clear patterns. Table 3 details the evaluation metrics proposed. For static concepts, submodular selection achieves a higher semantics score than random selection (38.3% vs 30.9%). For dynamic concepts, after introducing the loss functions \\(\\mathcal{L}_{\\text{dis}}\\), \\(\\mathcal{L}_{\\text{ort}}\\), and \\(\\mathcal{L}_{\\text{align}}\\), the semantic score of dynamic concepts improved to 32.8%, approaching static concepts. This indicates that our proposed loss functions effectively enhance the alignment and discriminative power of dynamic concepts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Interpretability. Interpretability can be divided into pattern and case interpretability. Pattern interpretability focuses on general patterns the model has learned, providing insight into its global behavior and how features influence classes. For example, a classifier's weight matrix reflecting bias toward each class. Case interpretability explains the model's decisions for individual instances, offering a local understanding of why a specific prediction was made for an image—for instance, concept scores indicating how strongly certain concepts are present. Following MMCBM [51], We" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20185" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.09, + 0.561, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.565, + 0.091, + 0.705, + 0.108 + ], + "angle": 0, + "content": "Concept Retrieval" + }, + { + "type": "text", + "bbox": [ + 0.581, + 0.118, + 0.835, + 0.176 + ], + "angle": 0, + "content": "You are an expert binary concept classifier, capable of determining whether a given concept has any form of relationship with the provided image. If it has any relationship with the image, respond with \"yes\". Otherwise, respond with \"no\"." + }, + { + "type": "image", + "bbox": [ + 0.578, + 0.179, + 0.646, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.647, + 0.188, + 0.844, + 0.198 + ], + "angle": 0, + "content": "- sometimes called the \"black albatross\"" + }, + { + "type": "text", + "bbox": [ + 0.647, + 0.199, + 0.836, + 0.222 + ], + "angle": 0, + "content": "- known as the black-browed albatross\n- most abundant albatross species" + }, + { + "type": "list", + "bbox": [ + 0.647, + 0.188, + 0.844, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.581, + 0.239, + 0.615, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.687, + 0.254, + 0.763, + 0.267 + ], + "angle": 0, + "content": "(no, yes, yes)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.287, + 0.907, + 0.317 + ], + "angle": 0, + "content": "Figure 5. On the left, the top-2 static and dynamic concepts for randomly selected classes across three datasets are presented, focusing on case interpretability. The concept retrieval process with vision-language model GPT-4o is illustrated on the right." + }, + { + "type": "table", + "bbox": [ + 0.098, + 0.318, + 0.472, + 0.442 + ], + "angle": 0, + "content": "
Methodconcepts per classRatio of Dynamic Concepts
00.20.40.60.81
LPN/A87.09N/AN/AN/AN/AN/A
Hybrid (submodular)583.2184.4685.4586.0886.3486.78
1084.2785.4786.2886.5186.8787.13
1584.7585.8486.4586.7887.0887.25
2084.9386.0986.4486.8486.8887.24
Hybrid (random)580.9584.0485.2885.9686.0786.53
1082.8885.2886.3086.6286.8287.13
1583.6585.8186.5486.8086.8987.25
2083.8586.0986.6686.8687.0687.24
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.449, + 0.483, + 0.476 + ], + "angle": 0, + "content": "Table 4. Ablation average results on 10 datasets for varying the ratio of dynamic concepts under different methods." + }, + { + "type": "table", + "bbox": [ + 0.1, + 0.48, + 0.472, + 0.547 + ], + "angle": 0, + "content": "
MethodMetric\\( \\mathcal{L}_{cls} \\)\\( \\mathcal{L}_{cls+dis} \\)\\( \\mathcal{L}_{cls+dis}+ort \\)\\( \\mathcal{L}_{cls+dis}+ort+align \\)
Hybrid (submodular)accuracy84.6084.3684.2984.54
precision@t0.04734.834.146.2
Hybrid (random)accuracy84.5084.1683.9184.17
precision@t0.04434.835.547.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.556, + 0.483, + 0.584 + ], + "angle": 0, + "content": "Table 5. Ablation results for varying training loss with a dynamic concept ratio of 0.5 and 10 concepts per class." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.597, + 0.483, + 0.824 + ], + "angle": 0, + "content": "use the attention matrix, computed as \\( S_{i} \\times W \\), where \\( W \\) is the classifier's weight, to show that the prediction is based on specific concepts within the image. As illustrated in Figure 5, we show the case interpretability by selecting the top 2 concepts for each image. We evaluate case interpretability by computing retrieval precision@t, measuring how accurately the top 5 identified concepts align with the image content. The results are shown in Table 3, where the static concepts serve as a baseline for comparison with the dynamic concepts; we focus on relative differences due to the potential limitations of GPT-4o. The dynamic concepts, across all loss functions, achieve \\( 46.2\\% \\), closely approaching the static concepts' performance of \\( 47.3\\% \\). This demonstrates that dynamic concepts offer high interpretability and can effectively serve as an alternative to static concepts." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.245, + 0.849 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Dynamic Concept Bank Ratio. To examine the impact of dynamic concept ratios on performance, we test various ratios on 10 datasets except ImageNet, as shown in Table 4." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.332, + 0.907, + 0.42 + ], + "angle": 0, + "content": "As the dynamic concept ratio increases, performance will also improve. Notably, using random concepts per class results in an performance drop compared to the submodular selection method. Furthermore, employing learned concepts, even at a ratio of 0.2, significantly improves performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.423, + 0.909, + 0.575 + ], + "angle": 0, + "content": "Concept-Interpretability Loss. Table 5 reports the accuracy and precision@t (t=5) of learned concepts across different combinations of loss functions, comparing the Hybrid (submodular) and Hybrid (random) methods. A slight decrease in accuracy is observed as additional terms are added to the \\(\\mathcal{L}_{cls}\\) but significantly improves precision@t. This suggests a trade-off, where additional constraints, while slightly reducing accuracy (e.g., accuracy drops from 84.60 to 84.29 for the Hybrid (submodular) method), enhance concept interpretability." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.589, + 0.634, + 0.604 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we introduced HybridCBM, a novel hybrid concept bottleneck model designed to enhance interpretability and performance by combining both static and dynamic concept banks. Unlike traditional CBMs, our method allows the model to dynamically discover new concepts during training, thus expanding its ability to capture unique and relevant features. Our approach also incorporates a concept translator to provide human-readable interpretations for the learned dynamic concepts, making the model's decision process more accessible and understandable. Through extensive experiments on various datasets, our method achieves comparable accuracy to 'black-box' models while retaining transparency, bridging the gap between interpretability and performance. We believe that HybridCBM represents a significant step forward in the development of interpretable machine learning models. Future work could further explore methods for optimizing the accuracy of concept labeling and exploring more sophisticated strategies for dynamic concept discovery." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20186" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.283, + 0.108 + ], + "angle": 0, + "content": "6. Acknowledgements." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.486, + 0.186 + ], + "angle": 0, + "content": "This project is supported by NSFC Key Program 62236009, Shenzhen Fundamental Research Program (General Program) JCYJ 20210324140807019, NSFC General Program 61876032, and Key Laboratory of Data Intelligence and Cognitive Computing, Longhua District, Shenzhen." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.21, + 0.188, + 0.226 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.235, + 0.482, + 0.317 + ], + "angle": 0, + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.318, + 0.482, + 0.373 + ], + "angle": 0, + "content": "[2] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 – mining discriminative components with random forests. In European Conference on Computer Vision, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.374, + 0.482, + 0.442 + ], + "angle": 0, + "content": "[3] Chaofan Chen, Oscar Li, Daniel Tao, Alina Barnett, Cynthia Rudin, and Jonathan K Su. This looks like that: Deep learning for interpretable image recognition. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.444, + 0.482, + 0.498 + ], + "angle": 0, + "content": "[4] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dollar, and C. Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server, 2015. 5, 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.499, + 0.482, + 0.54 + ], + "angle": 0, + "content": "[5] Gong Cheng, Junwei Han, and Xiaoqiang Lu. Remote sensing image scene classification: Benchmark and state of the art. Proceedings of the IEEE, 105(10):1865-1883, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.541, + 0.482, + 0.596 + ], + "angle": 0, + "content": "[6] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3606-3613, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.597, + 0.482, + 0.638 + ], + "angle": 0, + "content": "[7] Cristina Conati, Kaska Porayska-Pomsta, and Manolis Mavrikis. Ai in education needs interpretable machine learning: Lessons from open learner modelling, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.638, + 0.482, + 0.679 + ], + "angle": 0, + "content": "[8] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.68, + 0.482, + 0.735 + ], + "angle": 0, + "content": "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.736, + 0.482, + 0.803 + ], + "angle": 0, + "content": "[10] Andre Esteva, Alexandre Robicquet, Bharath Ramsundar, Volodymyr Kuleshov, Mark DePristo, Katherine Chou, Claire Cui, Greg Corrado, Sebastian Thrun, and Jeff Dean. A guide to deep learning in healthcare. Nature medicine, 25 (1):24-29, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.805, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[11] David Gunning and David Aha. Darpa's explainable artificial intelligence (xai) program. AI magazine, 40(2):44-58, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.845, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[12] Zhengbao Jiang, Frank F. Xu, Jun Araki, and Graham Neubig. How can we know what language models know? Transactions of the Association for Computational Linguistics, 8: 423-438, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.235, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.134 + ], + "angle": 0, + "content": "[13] Dmitry Kazhdan, Botty Dimanov, Mateja Jamnik, Pietro Lio, and Adrian Weller. Now you see me (cme): Concept-based model extraction, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.906, + 0.219 + ], + "angle": 0, + "content": "[14] Been Kim, Martin Wattenberg, Justin Gilmer, Carrie Cai, James Wexler, Fernanda Viegas, and Rory sayres. Interpretability beyond feature attribution: Quantitative testing with concept activation vectors (TCAV). In Proceedings of the 35th International Conference on Machine Learning, pages 2668-2677. PMLR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.906, + 0.262 + ], + "angle": 0, + "content": "[15] Injae Kim, Jongha Kim, Joonmyung Choi, and Hyunwoo J. Kim. Concept bottleneck with visual concept filtering for explainable medical image classification, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.906, + 0.303 + ], + "angle": 0, + "content": "[16] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 5, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.906, + 0.361 + ], + "angle": 0, + "content": "[17] Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. Concept bottleneck models. In International Conference on Machine Learning, pages 5338-5348. PMLR, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.906, + 0.391 + ], + "angle": 0, + "content": "[18] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.392, + 0.906, + 0.447 + ], + "angle": 0, + "content": "[19] Wei Li, Linchao Zhu, Longyin Wen, and Yi Yang. Decap: Decoding clip latents for zero-shot captioning via text-only training. In The Eleventh International Conference on Learning Representations, 2023. 3, 4, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.449, + 0.906, + 0.503 + ], + "angle": 0, + "content": "[20] Yuanning Li, Huzheng Yang, and Shi Gu. Enhancing neural encoding models for naturalistic perception with a multilevel integration of deep neural networks and cortical networks. Science Bulletin, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.505, + 0.906, + 0.545 + ], + "angle": 0, + "content": "[21] Max Losch, Mario Fritz, and Bernt Schiele. Interpretability beyond classification output: Semantic bottleneck networks, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.906, + 0.602 + ], + "angle": 0, + "content": "[22] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.906, + 0.645 + ], + "angle": 0, + "content": "[23] Tim Miller. Explanation in artificial intelligence: Insights from the social sciences. Artificial Intelligence, 267:1-38, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.647, + 0.906, + 0.688 + ], + "angle": 0, + "content": "[24] Ron Mokady, Amir Hertz, and Amit H Bermano. Clipcap: Clip prefix for image captioning. arXiv preprint arXiv:2111.09734, 2021. 3, 4, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.69, + 0.906, + 0.746 + ], + "angle": 0, + "content": "[25] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In Indian Conference on Computer Vision, Graphics and Image Processing, 2008. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.747, + 0.906, + 0.802 + ], + "angle": 0, + "content": "[26] Tuomas Oikarinen, Subhro Das, Lam M. Nguyen, and Tsui-Wei Weng. Label-free concept bottleneck models. In The Eleventh International Conference on Learning Representations, 2023. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.906, + 0.9 + ], + "angle": 0, + "content": "[27] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul F Christiano, Jan Leike, and Ryan Lowe. Training language models to follow instructions with human feedback. In Advances in Neural Information" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20187" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "Processing Systems, pages 27730-27744. Curran Associates, Inc., 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.122, + 0.483, + 0.162 + ], + "angle": 0, + "content": "[28] Ahmet Murat Ozbayoglu, Mehmet Ugur Gudelek, and Omer Berat Sezer. Deep learning for financial applications: A survey. Applied Soft Computing, 93:106384, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[29] Konstantinos Panagiotis Panousis, Dino Ienc, and Diego Marcos. Sparse linear concept discovery models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2767-2771, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.219, + 0.483, + 0.329 + ], + "angle": 0, + "content": "[30] Fabio Petroni, Tim Roktaschel, Sebastian Riedel, Patrick Lewis, Anton Bakhtin, Yuxiang Wu, and Alexander Miller. Language models as knowledge bases? In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 2463-2473, Hong Kong, China, 2019. Association for Computational Linguistics. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.331, + 0.482, + 0.382 + ], + "angle": 0, + "content": "[31] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.386, + 0.482, + 0.468 + ], + "angle": 0, + "content": "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.47, + 0.482, + 0.524 + ], + "angle": 0, + "content": "[33] Muhammad Imran Razzak, Saeeda Naz, and Ahmad Zaib. Deep Learning for Medical Image Processing: Overview, Challenges and the Future, page 323-350. Springer International Publishing, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.526, + 0.482, + 0.594 + ], + "angle": 0, + "content": "[34] Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. \"why should i trust you?\" explaining the predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining, pages 1135-1144, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.596, + 0.482, + 0.648 + ], + "angle": 0, + "content": "[35] Cynthia Rudin. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature Machine Intelligence, 1(5):206-215, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.651, + 0.481, + 0.677 + ], + "angle": 0, + "content": "[36] Gesina Schwalbe. Concept embedding analysis: A review. arXiv preprint arXiv:2203.13909, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.482, + 0.76 + ], + "angle": 0, + "content": "[37] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.763, + 0.482, + 0.803 + ], + "angle": 0, + "content": "[38] Chenming Shang, Hengyuan Zhang, Hao Wen, and Yujiu Yang. Understanding multimodal deep neural networks: A concept selection view, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.805, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[39] Chenming Shang, Shiji Zhou, Hengyuan Zhang, Xinzhe Ni, Yujiu Yang, and Yuwang Wang. Incremental residual concept bottleneck models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11030-11040, 2024. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[40] Sheng Shen, Liunian Harold Li, Hao Tan, Mohit Bansal, Anna Rohrbach, Kai-Wei Chang, Zhewei Yao, and Kurt" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "Keutzer. How much can CLIP benefit vision-and-language tasks? In International Conference on Learning Representations, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[41] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[42] Robyn Speer, Joshua Chin, and Catherine Havasi. Conceptnet 5.5: An open multilingual graph of general knowledge. In Thirty-first AAAI conference on artificial intelligence, 2017. 2, 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.236, + 0.905, + 0.277 + ], + "angle": 0, + "content": "[43] Divyansh Srivastava, Ge Yan, and Tsui-Wei Weng. Vlg-cbm: Training concept bottleneck models with vision-language guidance, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.279, + 0.905, + 0.333 + ], + "angle": 0, + "content": "[44] Sandeep Subramanian, Adam Trischler, Yoshua Bengio, and Christopher J Pal. Learning general purpose distributed sentence representations via large scale multi-task learning. arXiv preprint arXiv:1804.00079, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.336, + 0.905, + 0.419 + ], + "angle": 0, + "content": "[45] Mujeen Sung, Jinhyuk Lee, Sean Yi, Minji Jeon, Sungdong Kim, and Jaewoo Kang. Can language models be biomedical knowledge bases? In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 4723-4734, Online and Punta Cana, Dominican Republic, 2021. Association for Computational Linguistics. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.421, + 0.905, + 0.461 + ], + "angle": 0, + "content": "[46] Yoad Tewel, Yoav Shalev, Idan Schwartz, and Lior Wolf. Zero-shot image-to-text generation for visual-semantic arithmetic. arXiv preprint arXiv:2111.14447, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.463, + 0.905, + 0.518 + ], + "angle": 0, + "content": "[47] Philipp Tschandl, Cliff Rosendahl, and Harald Kittler. The ham10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions. Scientific data, 5(1):1-9, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.521, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[48] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.563, + 0.905, + 0.617 + ], + "angle": 0, + "content": "[49] Bowen Wang, Liangzhi Li, Yuta Nakashima, and Hajime Nagahara. Learning bottleneck concepts in image classification. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.62, + 0.905, + 0.675 + ], + "angle": 0, + "content": "[50] Zirui Wang, Jiahui Yu, Adams Wei Yu, Zihang Dai, Yulia Tsvetkov, and Yuan Cao. SimVLM: Simple visual language model pretraining with weak supervision. In International Conference on Learning Representations, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.677, + 0.905, + 0.746 + ], + "angle": 0, + "content": "[51] Yifan Wu, Yang Liu, Yue Yang, Michael S. Yao, Wenli Yang, Xuehui Shi, Lihong Yang, Dongjun Li, Yueming Liu, James C. Gee, Xuan Yang, Wenbin Wei, and Shi Gu. A concept-based interpretable model for the diagnosis of choroid neoplasias using multimodal data, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.748, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[52] Yue Yang, Artemis Panagopoulou, Shenghao Zhou, Daniel Jin, Chris Callison-Burch, and Mark Yatskar. Language in a bottle: Language model guided concept bottlenecks for interpretable image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19187-19197, 2023. 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[53] Chih-Kuan Yeh, Been Kim, Sercan Arik, Chun-Liang Li, Tomas Pfister, and Pradeep Ravikumar. On completeness-aware concept-based explanations in deep neural networks. In Advances in Neural Information Processing Systems, pages 20554-20565. Curran Associates, Inc., 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20188" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.093, + 0.482, + 0.133 + ], + "angle": 0, + "content": "[54] Mert Yuksekgonul, Maggie Wang, and James Zou. Posthoc concept bottleneck models. In ICLR 2022 Workshop on PAIR²Struct, 2022. 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.136, + 0.483, + 0.177 + ], + "angle": 0, + "content": "[55] Yu Zhang, Peter Tiño, Alex Leonardis, and Ke Tang. A survey on neural network interpretability. IEEE Transactions on Emerging Topics in Computational Intelligence, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.179, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[56] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.093, + 0.483, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "20189" + } + ] +] \ No newline at end of file diff --git a/2025/Hybrid Concept Bottleneck Models/7a729be8-7e97-4b64-9b6a-2c339ff64815_origin.pdf b/2025/Hybrid Concept Bottleneck Models/7a729be8-7e97-4b64-9b6a-2c339ff64815_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..647b695450e7e48edc8132da13b06c7fb4054d5a --- /dev/null +++ b/2025/Hybrid Concept Bottleneck Models/7a729be8-7e97-4b64-9b6a-2c339ff64815_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0b681fb1a15377b259ff26c2a2210fbcdc182e25c2d9572228858864a5b11244 +size 3761207 diff --git a/2025/Hybrid Concept Bottleneck Models/full.md b/2025/Hybrid Concept Bottleneck Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a8c1d227fef0f84022e3503c3fb1fb04908c9b06 --- /dev/null +++ b/2025/Hybrid Concept Bottleneck Models/full.md @@ -0,0 +1,320 @@ +# Hybrid Concept Bottleneck Models + +Yang Liu, Tianwei Zhang, Shi Gu +University of Electronic Science and Technology of China +Shenzhen Institute for Advanced Study, UESTC + +liu.yang.minéstd.uestc.edu.cn, gsbyztw@gmail.com, gus@uestc.edu.cn + +# Abstract + +Concept Bottleneck Models (CBMs) provide an interpretable framework for neural networks by mapping visual features to predefined, human-understandable concepts. However, the application of CBMs is often constrained by insufficient concept annotations. Recently, multi-modal pretrained models have shown promise in reducing annotation costs by aligning visual representations with textual concept embeddings. Nevertheless, the quality and completeness of the predefined concepts significantly affect the performance of CBMs. In this work, we propose Hybrid Concept Bottleneck Model (HybridCBM), a novel CBM framework to address the challenge of incomplete predefined concepts. Our method consists of two main components: a Static Concept Bank and a Dynamic Concept Bank. The Static Concept Bank directly leverages large language models (LLMs) for concept construction, while the Dynamic Concept Bank employs learnable vectors to capture complementary and valuable concepts continuously during training. After training, a pre-trained translator converts these vectors into human-understandable concepts, further enhancing model interpretability. HybridCBM is highly flexible and can be easily integrated with existing CBMs to improve both interpretability and performance. Experimental results1 on multiple datasets demonstrate that HybridCBM outperforms current state-of-the-art CBMs and achieves comparable results to black-box models. Additionally, we propose novel metrics to assess the quality of learned concepts, showing that they perform comparably to predefined concepts. + +# 1. Introduction + +Deep Neural Networks (DNNs) have become increasingly dominant in various fields, including computer vision, natural language processing, and speech recognition. However their complex and deep structures present a significant challenge to interpretability, often earning them the label + +![](images/4e6be5eabccdd36d34ca15099725a4981584a8936aa2bf2b1af6b458872f484c.jpg) +Figure 1. Our proposed concept-trainable CBM reduces the need for meticulously designed concepts by incorporating learned concepts (e.g., 'yellowish throat'), thereby enhancing interpretability and performance while minimizing manual design effort. + +of 'blackbox' [23]. Addressing this challenge is crucial not only for ensuring accurate predictions, but also for providing scientists and engineers with more accessible tools for designing, developing, and debugging models. Interpretability is particularly critical in domains such as healthcare [10], medicine [33], education [7], and finance [28], where high-stakes decision-making requires transparency and trustworthiness to ensure reliability and accountability. Therefore, there has been a growing focus on research aimed at developing interpretable models. While post-hoc explanation methods [34, 37, 55] have gained popularity, they often fall short by providing unfaithful representations of the model's underlying computations [35]. + +A natural approach to improving interpretability in DNNs is to design inherently interpretable models that leverage high-level concepts to describe features extracted by black-box models [36], imitating the way humans encode and understand the world[20]. One notable approach following such methodology is the Concept Bottleneck Models (CBM) [17]. CBMs operate by first predicting concepts from image representations mid-way through the decision-making process. These models pair each image with its predefined concepts and are trained in an end-to-end manner—initially predicting concepts and subsequently using these predictions to make final decisions. While this method of concept prediction enhances model interpretability, it often faces two fundamental challenges: (1) it is perceived to underperform compared to black-box models [11], + +and (2) the requirement for precise annotations for each concept significantly limits their broader application. + +To mitigate these shortcomings, recent research has explored multimodal pretrained model-based CBMs [15, 52, 54], leveraging CLIP [32] to establish correspondences between visual and textual representations, thereby reducing the reliance on costly concept annotations. Specifically, CLIP encodes images and human-understandable concepts into a shared feature space, enabling the creation of a concept bottleneck by directly mapping visual representations to concept embeddings instead of through manual annotations. Regarding the trade-off between interpretability and performance, some studies [15, 38, 52] have emphasized the construction of high-quality concept banks through efficient concept selection. However, establishing a comprehensive concept bank remains a significant challenge [53], often constrained by the inherent limitations in language expressiveness and precision. In response, other research efforts have explored the integration of a residual linear layer to address the issue of incomplete concept extraction [39, 54], albeit at the cost of interpretability. For example, ResCBM [39] introduces estimizable vectors to capture missing concepts, utilizing a candidate concept bank in the residual component to regularizing these vectors toward specific candidates during training. However, the model is limited by the scope of the candidate concept bank, restricting its ability to learn concepts outside the predefined set. Additionally, the effectiveness of the residual structure remains an open question. + +In this work, we propose the Hybrid Concept Bottleneck Model (HybridCBM), specifically designed to address the challenges of incomplete concept representation, predefined concept bank dependence, and scalability issue by dynamically discovering new concepts directly from visual representations. Different from prior methods that rely on residual structure or external concept banks, our approach incorporates a hybrid concept bank comprising both a static and a dynamic concept bank. Initially, we establish the static concept bank by leveraging a large language model (LLM) like GPT-3.5 [27, 31], which is celebrated for its broad world knowledge [12, 30, 45]. For instance, in Figure 1, when prompted about a palm warbler, GPT-3.5 provides information such as 'feeds on the ground'. This textual descriptions is then encoded into embeddings by the text encoder of CLIP. Next, we initialize a set of estimizable vectors that form the dynamic concept bank. These vectors are refined to identify concepts present in the input image but absent from the static concept bank. To further enhance the concept discovery, we pre-train a concept translator, such as GPT-2 [31], to translate these newly discovered unknown concepts into textual descriptions, a process we term as 'concept labeling'. To evaluate the understandability, visual relevance and factual accuracy of the dynamic concepts, we + +further design a set of evaluation metrics, some of which are based on Vision-Language Model (VLM) GPT-4o $^2$ . Our main contributions can be summarized as follows: + +1. Innovative Hybrid Concept Bank: Our model introduces a hybrid concept bank that combines static and dynamic concepts, allowing it to adapt and refine its interpretative capabilities dynamically. This structure leverages both predefined expert knowledge while enabling the discovery of new relevant concepts. +2. Seamless Integration of LLM Technologies: By employing advanced language models (LLM for defining static concepts, translating new concepts and VLM for evaluation) alongside the visual-textual bridging capabilities of CLIP, our model bridges the gap between visual representation and textual interpretation. +3. Enhanced Interpretability and Performance: Our approach maintains high interpretability through transparent concept labeling, while improving performance by discovering new concepts from visual information. + +# 2. Related Work + +Interpretable Neural Networks. One way to build an interpretable neural networks is the use of a concept-based explanation [3, 13, 15, 21, 29, 38, 43, 49]. Concept Bottleneck Models (CBMs) [17] are among the most popular approaches for making predictions based on human-interpretable concepts but require labor-intensive concept annotations for each image. Similarly, Concept Activation Vectors (CAVs) [14] represent concepts as normal vectors to the decision boundaries that separate positive and negative samples of a concept, though they also need extra datasets to train SVMs for each concept. Owing to the simple structure based on human-defined concepts, there have been studies aiming to alleviate these drawbacks, e.g., costly concept annotation and trade-off between interpretability and performance. Post-hoc Concept Bottleneck (PCBM) [54] utilizes ConceptNet [42] to obtain concepts and Leveraging CLIP to generate bottlenecks by projecting visual representations into the concept subspace. However, PCBM-h [54] incorporates a residual linear predictor that compensates for missing concepts by directly adding logits to PCBM's predictor. While this modification helps to recover accuracy, it does so at the expense of interpretability. The label-free CBM [26] uses CLIP for concept annotation, allowing arbitrary visual backbone to be transformed into an CBM without the need for labeled concept data. Labo [52] concentrates on concepts selection after generating candidate concepts through LLM. ResCBM [39] builds on PCBM-h's residual structure by using a candidate concept bank to guide and refine estimizable vectors. However, the model's expressive potential is limited by the extent of the candidate concept bank. + +![](images/4c63391b7a4a249706c14a7a45c35f7af6190a369095e00c5ccf7c4b8a37d558.jpg) +Figure 2. Overview of the Hybrid Concept Bottleneck Model (HybridCBM), which is designed for interpretable concept learning and labeling. First, a Concept Translator is pre-trained to do concept-labeling. Next, a pretrained alignment model (CLIP) is used to obtain embeddings of predefined concepts, forming the Static Concept Bank. The Dynamic Concept Bank learns missing concepts and aligns its distribution with the Static Concept Bank to create the final concept bottleneck. Finally, an interpretable concept-based predictor classifies targets based on the concept scores. + +CLIP in Image Captioning. Multimodal pretrained models like CLIP [32], trained on over 400 million image-text pairs using unsupervised contrastive loss, exhibit impressive capabilities in image captioning tasks [1, 19, 46]. Previous studies, such as [24] and [40], have primarily used CLIP as a visual encoder for captioning tasks, while [19] projects visual embeddings into CLIP's text embedding space to preserve visual information. However, these approaches often underutilize CLIP's aligned multi-modal latent space. Similar to [19, 24], we collect image-caption pairs and a vast corpus to train a GPT-2 from scratch as our concept translator, fully leveraging CLIP's latent space and bridging the gap between image and concept embeddings and their corresponding textual descriptions. + +# 3. Method + +We provide an overview of our methodology in Figure 2. Our model incorporates a hybrid concept bank and a pretrained concept translator (Section 3.2). The concept bank consists of a set of estimizable vectors, which are refined to discover new concepts through a concept-interpretability training scheme (Section 3.3). As a result, the learned concepts are not only discriminative and orthogonal but also aligned with human-understandable semantics. The pre-trained concept translator performs concept labeling by mapping these learned, initially unknown vectors to meaningful human concepts. Finally, we apply a sparse linear layer over the similarity scores between hybrid concepts + +and images, enabling us to identify the specific concepts the model relies on for its decisions. + +# 3.1. Problem Formulation + +Consider a dataset $\mathcal{D} = (x,y)$ , where each image $x$ is paired with a label $y\in \mathcal{V}$ . We use the image encoder $\mathcal{F}_{img}$ and text encoder $\mathcal{F}_{txt}$ of CLIP [32] to map images and text into a shared $d$ -dimensional feature space, respectively. A pre-trained concept translator $\mathcal{T}$ then maps these shared $d$ -dimensional features back to textual concepts. In order to store the interpretable information of the model, we construct a hybrid concept bank $E_{C}\in \mathbb{R}^{N_{C}\times d}$ , comprising: + +- Static Concept Bank $E_{S} \in \mathbb{R}^{N_{S} \times d}$ . Contains text features $\mathcal{F}_{txt}(c) \in \mathbb{R}^{d}$ extracted by the text encoder $\mathcal{F}_{txt}$ . Each row corresponds to a predefined concept $C_{S} = \{c_{1}, c_{2}, \ldots, c_{N_{S}}\}$ generated from a large language model. +- Dynamic Concept Bank $E_{D} \in \mathbb{R}^{N_{D} \times d}$ . Consists of a set of randomly initialized learnable vectors $e_{i} \in \mathbb{R}^{d}$ that are optimized to capture missing visual-specific features. After training, These vectors could be mapped to the textual concept space $C_{D} = \{c_{1}, c_{2}, \ldots, c_{N_{D}}\}$ by the translator $\mathcal{T}$ , where each concept $c = \mathcal{T}(e)$ . + +The overall hybrid concept bank is defined as $E_{C} = E_{S}\oplus$ $E_{D}$ , where $\oplus$ denotes concatenation. The total number of concepts in the bank is indicated by $N_{C} = N_{S} + N_{D}$ . To compute the concept scores, we utilize the CLIP image encoder to extract the image feature $f = \mathcal{F}_{img}(x)\in \mathbb{R}^{d}$ Therein, the similarity between $f$ and $E_{C}$ is computed using projection length, resulting in the concept score $S\in \mathbb{R}^{N_C}$ + +which reflects the presence of particular concepts in the image $x\colon S_x = \| f\| _2\cos (f,E_C) = f\cdot \hat{E_C}^T$ , where $\hat{\cdot}$ denotes $L_{2}$ normalization and cos is the cosine similarity. Since $S$ provides a unique representation for each image, it serves as a concept bottleneck. These per-image representations can be naturally used to support an interpretable classification. Following the existing CLIP-based CBMs [26, 52, 54], we train a linear layer as the concept classifier $\phi :\hat{y} = \phi (S),\mathbb{R}^{N_C}\to \mathcal{V}$ to make final predictions in the label space based on the concept scores. In our perspective, we intend to improve the performance and learn abundant concepts from the visual information at the same time. This can be solved by the following optimization problem: + +$$ +\begin{array}{l} \min _ {\phi , E _ {D}} \underset {(x, y) \sim \mathcal {D}} {\mathbb {E}} \left[ \mathcal {L} _ {c l s} \left(\phi \left(f \cdot \hat {E} _ {S} ^ {T} \oplus f \cdot \hat {E} _ {D} ^ {T}\right), y\right) \right] \tag {1} \\ + \lambda \Omega (\phi) + \Gamma_ {E _ {D}} \\ \end{array} +$$ + +where $\mathcal{L}_{cls}(\hat{y}, y)$ is the cross-entropy loss function, $\Omega$ is a complexity measure used to promote sparsity in the linear layer, $\lambda$ is the regularization strength and $\Gamma$ represents the optimization objective for the Dynamic Concept Bank $E_{D}$ , introduced in Section 3.3. + +# 3.2. Hybrid Concept Bank Establishment + +Previous methods, such as Labo [52], generate a set of relevant concepts and then form a concept bank through the submodular selection method. However, the predefined concepts often suffer from inherent flaws due to their limited scope, and the selection process may inadvertently omit important information. To address these issues, we propose a Hybrid Concept Bank that dynamically discovers concepts directly from the visual representation, combining both a Static Concept Bank and a Dynamic Concept Bank. + +Static Concept Bank. Following Labo, we utilize GPT-3.5 to generate a set of candidate concepts for each class name $y \in \mathcal{V}$ with prompts such as "describe what the looks like". After initially generating 500 concepts per class, we perform concept selection to ensure a balanced representation. Specifically, we select $N$ concepts for each class using both submodular and random selection methods for comparison. Assume $K = |\mathcal{V}|$ , this process yields a total of $N_S = N \times K$ concepts, which are then encoded by the text encoder $\mathcal{F}_{txt}$ of CLIP and fixed during training. + +Dynamic Concept Bank. Unlike the static counterpart, the Dynamic Concept Bank is initialized with a set of concept vectors $e_d \in \mathbb{R}^d$ , which are optimized through the proposed Concept-Interpretability Training Scheme (Section 3.3). Thus, it can adaptively capture the unique and rich visual-specific features during the training process. + +Concept Labeling. To enhance the interpretability of learned concepts, we pre-train a concept translator using GPT-2 [31] architecture to invert the CLIP text encoder. + +Following recent work [19, 24, 50], we use prefix language modeling for training. Details of the pre-training process are provided in Appendix B.2. The translator maps each concept embedding $e$ to its textual description $c$ through $c = \mathcal{T}(e)$ , assigning semantic meaning to the learned concepts and improving interpretability. + +# 3.3. Concept-Interpretability Training Scheme + +The previous CLIP-based CBMs rely solely on predefined concepts, often overlooking the potential of learning concepts directly from visual representation. Since CLIP maps both images and texts into a shared feature space, we could then leverage this capability to optimize the dynamic concept bank $E_{D}$ based on the given image embedding $f$ . To achieve this, we introduce a concept-interpretability training scheme that regularizes $E_{D}$ through multiple loss functions. The goal is to obtain a set of learnable concepts that are discriminative, diverse, and semantically aligned with human-understandable meanings, thereby enhancing both model interpretability and performance. + +Discriminability Loss. The discriminability loss ensures that learned concepts are highly activated for images of their own class and less activated for images of other classes. We first divide the dynamic concept vectors into several subsets corresponding to the classes, such that $E_D = \{E_D^1, E_D^2, \ldots, E_D^K\}$ , where $K$ is the number of classes. Each subset $E_D^k$ contains $N_D / K$ dynamic concept vectors associated with class $k$ . To encourage vectors $\mathbf{c}_k \in E_D^k$ to learn specific information of images within class $k$ and incorporate semantic information from the class name, we define the intra-class loss: + +$$ +\mathcal {L} _ {i n t r a} ^ {i} = y _ {k} \left\{- \hat {f} \cdot \hat {c} _ {k} + \alpha \left| m - \hat {\mathcal {F}} _ {t x t} (k) \cdot \hat {c} _ {k} \right| \right\} \quad (2) +$$ + +where $f = \mathcal{F}_{img}(x)$ is the image feature, $\mathcal{F}_{txt}(k)$ is the class embedding, $\mathbf{c}_k$ is the sampled concept vector for class $k$ from the distribution of vectors $E_D^k$ via reparameterization $^3$ , $\alpha$ is a hyperparameter and $m$ is a margin used to prevent excessive similarity to class $k$ , with a value of 0.85 as defined in Label-Free [26]. To ensure that concepts are distinct between different classes, we further define the inter-class loss: + +$$ +\mathcal {L} _ {i n t e r} ^ {i} = (y _ {k} - 1) \left\{- \hat {f} \cdot \hat {c} _ {k} + \alpha \left| m - \hat {\mathcal {F}} _ {t x t} (k) \cdot \hat {c} _ {k} \right| \right\} (3) +$$ + +The total discriminability loss is defined as $\mathcal{L}_{dis} = \mathcal{L}_{intra} + \beta \mathcal{L}_{inter}$ , where $\beta$ is a balancing hyperparameter. + +Orthogonality Loss. To enhance the diversity of dynamic concept embeddings and reduce redundancy, we introduce an orthogonality loss, which comprises two components: + +intra-class diversity and inter-bank diversity. First, we promote the dynamic concept embeddings within the same class to be orthogonal to each other. Specifically, we define the intra-class diversity loss as follows: + +$$ +\mathcal {L} _ {\text {o r t - i n t r a}} = \frac {1}{K} \sum_ {k = 1} ^ {K} \frac {1}{(N _ {D} / K) ^ {2}} \sum_ {i \neq j} \left| \hat {\mathbf {e}} _ {D, i} ^ {k} \cdot \hat {\mathbf {e}} _ {D, j} ^ {k} \right| \tag {4} +$$ + +where $\mathbf{e}_{D,i}^{k}$ and $\mathbf{e}_{D,j}^{k}$ are dynamic concept embeddings associated with class $k$ , $K$ is the number of classes, and $N_{D}$ is the total number of dynamic concepts. Second, to minimize redundancy between static and dynamic concepts, we define the inter-bank diversity loss as follows: + +$$ +\mathcal {L} _ {\text {o r t - i n t e r}} = \frac {1}{N _ {S} N _ {D}} \sum_ {i = 1} ^ {N _ {S}} \sum_ {j = 1} ^ {N _ {D}} | \hat {\mathbf {e}} _ {S, i} \cdot \hat {\mathbf {e}} _ {D, j} | \tag {5} +$$ + +where $\mathbf{e}_{S,i}$ and $\mathbf{e}_{D,j}$ are static and dynamic concept embeddings, respectively, and $N_{S}$ is the number of static concepts. The total orthogonality loss is then defined as $\mathcal{L}_{ort} = \mathcal{L}_{ort - intra} + \mathcal{L}_{ort - inter}$ . + +Distribution Alignment Loss. To enhance consistency between the distributions of dynamic and static concepts, we employ the Sinkhorn divergence [8], an entropy-regularized version of the Sinkhorn distance. We define the distribution alignment loss as: $\mathcal{L}_{\mathrm{align}} = \mathfrak{S}_{div,\epsilon}(E_D,E_S)$ , where $\mathfrak{S}_{div,\epsilon}$ denotes the Sinkhorn divergence, $E_{D}\in \mathbb{R}^{N_{D}\times d}$ and $E_{S}\in \mathbb{R}^{N_{S}\times d}$ are the dynamic and static concept embeddings, respectively. This loss encourages the distribution of dynamic concepts to align with that of static concepts, thus promoting semantic consistency and interpretability. + +Overall Loss Function. The total loss function combines the classification loss with the proposed regularization terms $\mathcal{L} = \mathcal{L}_{\mathrm{cls}} + \lambda_{\mathrm{dis}}\mathcal{L}_{\mathrm{dis}} + \lambda_{\mathrm{ort}}\mathcal{L}_{\mathrm{ort}} + \lambda_{\mathrm{align}}\mathcal{L}_{\mathrm{align}}$ , where $\lambda_{\mathrm{dis}}$ , $\lambda_{\mathrm{ort}}$ , and $\lambda_{\mathrm{align}}$ are hyperparameters to balance the losses. + +# 4. Experiments + +# 4.1. Experiments Setup + +Dataset. We collect image-caption pairs from MSCOCO [4] and then compile a large corpus from ConceptNet [42] and MSCOCO, which also includes concepts generated by GPT-3.5. After filtering out concepts that contain fewer than 15 words, the dataset includes 1,738,985 concepts and 566,747 image-concept pairs, used to pre-train our concept translator for translating both image and text embeddings. We also perform comprehensive experiments on 11 classification datasets, covering a broad spectrum of domains, including: ImageNet [9], CIFAR-10 and CIFAR-100 [18], Food-101 [2], FGVC-Aircraft [22], Flower-102 [25], CUB200-2011 [48], UCF-101 [41], DTD [6], HAM10000 [47] and RESISC45 [5]. Each dataset is split into training, validation, and test sets, as per standard practices. + +Baselines. To evaluate our model, we compare it with a linear probe and several interpretable methods. + +- Linear Probe. Following the implementation of CLIP [32], we train a logistic regression model using cuML's L-BFGS solver with an L2 penalty term, directly from the visual representations encoded by CLIP. +- PCBM [54]. PCBM conceptualizes each class as a node in ConceptNet [42] and aggregates neighboring nodes as concepts. PCBM-h [54] introduces an additional classifier that maps image embeddings into the label space, serving as residual shortcuts to the original classifier. +- Labo [52]. Labo employs a submodular function to select concepts for each class from candidate concepts generated by LLM. Unlike other models, it uses dot-product, without any normalization, to determine the presence of particular concepts in an image. +- ResCBM[39]. Building on PCBM-h, ResCBM utilizes ConceptNet [42] to construct a candidate concept bank. It optimizes a set of vectors to align with some concepts from this bank, thereby enhancing the interpretability. + +Implementation Details. The hybrid concept bank is constructed with a default ratio of 0.5 between the dynamic and static concept bank, ensuring an equal number of concepts for each class. Our model is trained using the Adam [16] optimizer within the PyTorch Lightning framework. We save checkpoints that achieve the highest validation accuracy and perform evaluations on the test set. Further details can be found in the supplementary material. + +# 4.2. Classification Performance + +Few-shot Comparison. To assess the performance gap between HybridCBM and 'black-box' models, we compare our method with the linear probe (LP) in a few-shot setting. We follow the few-shot evaluation protocol introduced by CLIP [32]. Specifically, we randomly sample 1, 2, 4, 8, and 16 images per class from the training set and use all available images in the fully-supervised setting. For all experiments, we use CLIP-ViT-L/14 as the backbone and compare the performance with submodular [52] and random static concept selection methods. As shown in Table 1, the mean test accuracy across all datasets indicates that HybridCBM performs similarly to LP in the fully-supervised setting and significantly surpasses LP in the few-shot scenario. Figure 3 presents detailed comparisons across eight datasets, with the last three provided in the appendix. + +CBMs Comparison. We further compare HybridCBM with other CBMs and employ CLIP-RN50 as the backbone to ensure fairness. Additionally, since our dynamic concepts are derived from visual representation, we create a variant called CaptionCBM, where $N_{D} / K$ images for each class are sampled and translated into captions to replace the dynamic concepts. As summarized in Table 2, despite strong baselines like LaBo, and ResCBM, Hybrid- + +![](images/2b961cb45c7c0cc6dfe309408f3510cd68c230e42949ed4fe74b5ec9c066b917.jpg) + +![](images/c2c6bb52da85a431e7fe3f67d30187d34b5d5ce25fd8614c68ad88b4fafcca68.jpg) + +![](images/ac88102dacec52339c538b12d5b9f7db9a52e6b45550bc66c52a1182351a5f21.jpg) + +![](images/fdefcbf83f30ab8b3c95a247b27f4029bc5de62a6f01a39cafea0ff01dc254ff.jpg) + +![](images/9606fc4fbc3e3dd0a3e03accc5cc302e432097630772505e412b70b34bbb2c48.jpg) +Figure 3. Comparison of test accuracy between HybridCBM with submodular and random selection methods, and Linear Probe across 8 datasets. The x-axis denotes the number of labeled images. + +![](images/68649b7494c5ab1d544044bcc7bd73f3ae08f9c7033eccaa81fa7c4d678edeb6.jpg) + +![](images/33f30f0554c71afaa22cb318a8feb446b4f955eea1f476d1a7d15405fb5d54d2.jpg) + +![](images/a63d6407ddb1a80a8960a1a0ef3aaeed2a76f6aa2700bbb29de5a9c5f14b97ca.jpg) + +
Method12AircraftHAM10000RESISC45
\( concepts per class: ^{50} \)16Full\( concepts per class: ^{50} \)1.0\( concepts per class: ^{50} \)
\( 4 \)87816
Linear Probe52.30\( 64.60^6 \)71.9478.3381.5586.840.72.59
LaboN/A\( N/A_{0.5} \)N/AN/AN/A85.720.6N/A
HybridCBM(submodular)58.64\( 69.52_{0.4} \)75.0878.7181.9386.830.75.12
HybridCBM (random)57.58\( 66.69_{3} \)74.6278.02Line 0r98 HybridCBM(submodular)86.820.44.16\( Linear Probe \)\( Linear Probe \)\( Linear Probe \)
0.3\( Non distinct \)\( Non distinct \)\( Non distinct \)
\( HybridCBM (random) \)\( HybridCBM (random) \)\( HybridCBM (random) \)
+ +Table 1. Mean accuracy across all datasets at different shots. + +
MethodInterpretabilityCIFAR-10CIFAR-100
Linear Probe✗88.870.1
PCBM-h [54]✗87.669.9
PCBM [54]✓84.556.0
Label-Free [26]†✓86.365.2
Labo [52]✓87.665.2
ResCBM [39]†✓88.067.9
CaptionCBM✓81.860.7
HybridCBM✓88.0368.38
+ +Table 2. Test accuracy comparison between HybridCBM and on CIFAR-10 and CIFAR-100. † indicates the reported performance. + +CBM achieves state-of-the-art performance on CIFAR-10 and CIFAR-100, and delivers results comparable to LP. + +# 4.3. Interpretability of the dynamic concepts + +In addition to evaluating classification performance, we assess the interpretability of the learned concepts in this section using the following metrics: + +- Concept Purity: This metric measures how well each concept aligns with its respective class, reflecting the ability of the concepts to capture class-specific features. For each class $k$ , we compute the average cosine similarity between the normalized embedding of the concept $\hat{e}_k$ and the class name $\hat{\mathcal{F}}_{txt}(k)$ . + +$$ +P u r i t y = \frac {1}{K} \sum_ {k = 1} ^ {K} \left(\bar {\hat {e}} _ {k} \cdot \hat {\mathcal {F}} _ {t x t} (k)\right) \tag {6} +$$ + +$$ +S e p a r a t i o n = 1 - \frac {2}{K (K - 1)} \sum_ {i = 1} ^ {K} \sum_ {j = i + 1} ^ {K} (\bar {\hat {e}} _ {i} \cdot \bar {\hat {e}} _ {j}) \quad (7) +$$ + +where $\bar{\hat{e}}_i$ denotes the mean embedding of concepts for class $i$ . Higher separation indicates greater independence between concepts of different classes in semantic space. + +- Semantic Alignment: To ensure semantic alignment between translated concepts and intended dataset classes, we use GPT-3.5 to evaluate each translated concept's accuracy in representing a specific class with a binary "yes" or "no" response. The validation metric, "Semantics", is calculated as the proportion of concepts that GPT-3.5 correctly associates with the intended classes, reflecting the semantic interpretability. The detailed information is provided in Appendix B.3. + +$$ +S e m a n t i c s = \frac {1}{N _ {D}} \sum_ {i = 1} ^ {N _ {D}} \delta \left(L L M \left(c _ {i}, \text {c l a s s e s}\right), 1\right) \tag {8} +$$ + +where $LLM(c_i, classes)$ represents the response to whether concept $c_i$ belongs to one of the classes, and $\delta$ is a binary function that returns 1 for "yes" and 0 otherwise. To mitigate potential limitations of the LLM, we emphasize comparing the relative size of static and dynamic concept groups over absolute absolute values. + +![](images/e621690fc5de3503e3e90060d3600ff1861c3733160d938fdaec639b89bf9795.jpg) +(a) Dynamic Concept Correlation + +![](images/b3f6ae79e26dd994a3af4e780a38a152a8fb9347cf21025c1ba293aaa8c5677e.jpg) +(b) Static Concept Correlation +Figure 4. Comparison of dynamic and static concept correlation structures on the CUB dataset. (a) Dynamic concepts: Correlation maps and classifier weights illustrate how dynamic concepts adjust their relationships with classes and among themselves. (b) Static concepts: Selected and random concepts show their contributions to class relationships and inter-concept dependencies. + +
LossConcept BankPurity (%)Separation (%)Semantics (%)Precision@t (%)
LclsStatic (submodular)19.386.238.348.5
Static (random)19.282.430.947.3
Dynamic0.1100.00.010.05
Lcls+disDynamic71.976.930.034.8
Lcls+dis +ortDynamic72.077.128.934.1
Lcls+dis +ort+alignDynamic39.880.032.846.2
+ +Table 3. Evaluation of HybridCBM's static and dynamic concepts across various metrics. + +- Concept-Image Relevance: As shown on the right in Figure 5, to evaluate the accuracy of the concepts aligning with the actual content of each image, we use the ability of GPT-4o in image understanding to calculate the retrieval Precision@t. For a sampled image of each class, we measure the average proportion of $t$ concepts that GPT-4o confirms as relevant. + +$$ +P r e c i s i o n @ t = \frac {1}{K} \sum_ {i = 1} ^ {K} \frac {\delta (V L M \left(c _ {i , t} , i m g\right) , 1)}{t} \tag {9} +$$ + +where $VLM(c_{i,t},img)$ is the response of GPT-4o and $K$ is the total number of classes. $c_{i,t}$ is $t$ concepts of class $i$ . + +Concept Analysis. We perform concept analysis on the CUB dataset. Figure 4 compares the correlation structures of dynamic and static concepts. In Figure 4a, the ConceptClass Correlation Map shows that our dynamic concepts align closely with their corresponding classes while effectively capturing diversity across classes. The ConceptConcept Correlation Map further highlights the diversity among concepts. Notably, regions with higher coefficients + +reflect natural semantic relationships between certain class names, such as "red-eyed vireo" and "white-eyed vireo". The pattern also seen in the Class-Class Correlation Map. The weight matrix of the classifier in Figure 4a reveals that the dynamic concept part has a more pronounced diagonal compared to the static concept, indicating that dynamic concepts are more discriminative and align better with their corresponding classes. In Figure 4b, we present similar relation maps of concepts selected by a submodular function and randomly selected concepts from each class. The selected concepts are more cohesive with their classes, whereas the randomly selected concepts, despite belonging to each class, do not exhibit clear patterns. Table 3 details the evaluation metrics proposed. For static concepts, submodular selection achieves a higher semantics score than random selection (38.3% vs 30.9%). For dynamic concepts, after introducing the loss functions $\mathcal{L}_{\text{dis}}$ , $\mathcal{L}_{\text{ort}}$ , and $\mathcal{L}_{\text{align}}$ , the semantic score of dynamic concepts improved to 32.8%, approaching static concepts. This indicates that our proposed loss functions effectively enhance the alignment and discriminative power of dynamic concepts. + +Interpretability. Interpretability can be divided into pattern and case interpretability. Pattern interpretability focuses on general patterns the model has learned, providing insight into its global behavior and how features influence classes. For example, a classifier's weight matrix reflecting bias toward each class. Case interpretability explains the model's decisions for individual instances, offering a local understanding of why a specific prediction was made for an image—for instance, concept scores indicating how strongly certain concepts are present. Following MMCBM [51], We + +![](images/ccb72a03ec3d450c1f70274923e0d913a747efc21559b128f3b78132f9849f36.jpg) +Concept Retrieval + +You are an expert binary concept classifier, capable of determining whether a given concept has any form of relationship with the provided image. If it has any relationship with the image, respond with "yes". Otherwise, respond with "no". + +![](images/2da2dbb27c6ff480001b88d582b14e5fe3e7d178bb1a1a39424320072032ce6d.jpg) + +- sometimes called the "black albatross" +- known as the black-browed albatross +- most abundant albatross species + +![](images/3caa0a1b10c2e8d0a607d6952259239cda7a8251ce05c24574bc4620f8285371.jpg) +Figure 5. On the left, the top-2 static and dynamic concepts for randomly selected classes across three datasets are presented, focusing on case interpretability. The concept retrieval process with vision-language model GPT-4o is illustrated on the right. + +(no, yes, yes) + +
Methodconcepts per classRatio of Dynamic Concepts
00.20.40.60.81
LPN/A87.09N/AN/AN/AN/AN/A
Hybrid (submodular)583.2184.4685.4586.0886.3486.78
1084.2785.4786.2886.5186.8787.13
1584.7585.8486.4586.7887.0887.25
2084.9386.0986.4486.8486.8887.24
Hybrid (random)580.9584.0485.2885.9686.0786.53
1082.8885.2886.3086.6286.8287.13
1583.6585.8186.5486.8086.8987.25
2083.8586.0986.6686.8687.0687.24
+ +Table 4. Ablation average results on 10 datasets for varying the ratio of dynamic concepts under different methods. + +
MethodMetric\( \mathcal{L}_{cls} \)\( \mathcal{L}_{cls+dis} \)\( \mathcal{L}_{cls+dis}+ort \)\( \mathcal{L}_{cls+dis}+ort+align \)
Hybrid (submodular)accuracy84.6084.3684.2984.54
precision@t0.04734.834.146.2
Hybrid (random)accuracy84.5084.1683.9184.17
precision@t0.04434.835.547.2
+ +Table 5. Ablation results for varying training loss with a dynamic concept ratio of 0.5 and 10 concepts per class. + +use the attention matrix, computed as $S_{i} \times W$ , where $W$ is the classifier's weight, to show that the prediction is based on specific concepts within the image. As illustrated in Figure 5, we show the case interpretability by selecting the top 2 concepts for each image. We evaluate case interpretability by computing retrieval precision@t, measuring how accurately the top 5 identified concepts align with the image content. The results are shown in Table 3, where the static concepts serve as a baseline for comparison with the dynamic concepts; we focus on relative differences due to the potential limitations of GPT-4o. The dynamic concepts, across all loss functions, achieve $46.2\%$ , closely approaching the static concepts' performance of $47.3\%$ . This demonstrates that dynamic concepts offer high interpretability and can effectively serve as an alternative to static concepts. + +# 4.4. Ablation Study + +Dynamic Concept Bank Ratio. To examine the impact of dynamic concept ratios on performance, we test various ratios on 10 datasets except ImageNet, as shown in Table 4. + +As the dynamic concept ratio increases, performance will also improve. Notably, using random concepts per class results in an performance drop compared to the submodular selection method. Furthermore, employing learned concepts, even at a ratio of 0.2, significantly improves performance. + +Concept-Interpretability Loss. Table 5 reports the accuracy and precision@t (t=5) of learned concepts across different combinations of loss functions, comparing the Hybrid (submodular) and Hybrid (random) methods. A slight decrease in accuracy is observed as additional terms are added to the $\mathcal{L}_{cls}$ but significantly improves precision@t. This suggests a trade-off, where additional constraints, while slightly reducing accuracy (e.g., accuracy drops from 84.60 to 84.29 for the Hybrid (submodular) method), enhance concept interpretability. + +# 5. Conclusion + +In this paper, we introduced HybridCBM, a novel hybrid concept bottleneck model designed to enhance interpretability and performance by combining both static and dynamic concept banks. Unlike traditional CBMs, our method allows the model to dynamically discover new concepts during training, thus expanding its ability to capture unique and relevant features. Our approach also incorporates a concept translator to provide human-readable interpretations for the learned dynamic concepts, making the model's decision process more accessible and understandable. Through extensive experiments on various datasets, our method achieves comparable accuracy to 'black-box' models while retaining transparency, bridging the gap between interpretability and performance. We believe that HybridCBM represents a significant step forward in the development of interpretable machine learning models. Future work could further explore methods for optimizing the accuracy of concept labeling and exploring more sophisticated strategies for dynamic concept discovery. + +# 6. Acknowledgements. + +This project is supported by NSFC Key Program 62236009, Shenzhen Fundamental Research Program (General Program) JCYJ 20210324140807019, NSFC General Program 61876032, and Key Laboratory of Data Intelligence and Cognitive Computing, Longhua District, Shenzhen. + +# References + +[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 3 +[2] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 – mining discriminative components with random forests. In European Conference on Computer Vision, 2014. 5 +[3] Chaofan Chen, Oscar Li, Daniel Tao, Alina Barnett, Cynthia Rudin, and Jonathan K Su. This looks like that: Deep learning for interpretable image recognition. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 2 +[4] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dollar, and C. Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server, 2015. 5, 1, 2 +[5] Gong Cheng, Junwei Han, and Xiaoqiang Lu. Remote sensing image scene classification: Benchmark and state of the art. Proceedings of the IEEE, 105(10):1865-1883, 2017. 5 +[6] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3606-3613, 2014. 5 +[7] Cristina Conati, Kaska Porayska-Pomsta, and Manolis Mavrikis. Ai in education needs interpretable machine learning: Lessons from open learner modelling, 2018. 1 +[8] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2013. 5 +[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5 +[10] Andre Esteva, Alexandre Robicquet, Bharath Ramsundar, Volodymyr Kuleshov, Mark DePristo, Katherine Chou, Claire Cui, Greg Corrado, Sebastian Thrun, and Jeff Dean. A guide to deep learning in healthcare. Nature medicine, 25 (1):24-29, 2019. 1 +[11] David Gunning and David Aha. Darpa's explainable artificial intelligence (xai) program. AI magazine, 40(2):44-58, 2019. 1 +[12] Zhengbao Jiang, Frank F. Xu, Jun Araki, and Graham Neubig. How can we know what language models know? Transactions of the Association for Computational Linguistics, 8: 423-438, 2020. 2 + +[13] Dmitry Kazhdan, Botty Dimanov, Mateja Jamnik, Pietro Lio, and Adrian Weller. Now you see me (cme): Concept-based model extraction, 2020. 2 +[14] Been Kim, Martin Wattenberg, Justin Gilmer, Carrie Cai, James Wexler, Fernanda Viegas, and Rory sayres. Interpretability beyond feature attribution: Quantitative testing with concept activation vectors (TCAV). In Proceedings of the 35th International Conference on Machine Learning, pages 2668-2677. PMLR, 2018. 2 +[15] Injae Kim, Jongha Kim, Joonmyung Choi, and Hyunwoo J. Kim. Concept bottleneck with visual concept filtering for explainable medical image classification, 2023. 2 +[16] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 5, 4 +[17] Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. Concept bottleneck models. In International Conference on Machine Learning, pages 5338-5348. PMLR, 2020. 1, 2 +[18] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 5 +[19] Wei Li, Linchao Zhu, Longyin Wen, and Yi Yang. Decap: Decoding clip latents for zero-shot captioning via text-only training. In The Eleventh International Conference on Learning Representations, 2023. 3, 4, 2 +[20] Yuanning Li, Huzheng Yang, and Shi Gu. Enhancing neural encoding models for naturalistic perception with a multilevel integration of deep neural networks and cortical networks. Science Bulletin, 2024. 1 +[21] Max Losch, Mario Fritz, and Bernt Schiele. Interpretability beyond classification output: Semantic bottleneck networks, 2019. 2 +[22] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 5 +[23] Tim Miller. Explanation in artificial intelligence: Insights from the social sciences. Artificial Intelligence, 267:1-38, 2019. 1 +[24] Ron Mokady, Amir Hertz, and Amit H Bermano. Clipcap: Clip prefix for image captioning. arXiv preprint arXiv:2111.09734, 2021. 3, 4, 2 +[25] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In Indian Conference on Computer Vision, Graphics and Image Processing, 2008. 5 +[26] Tuomas Oikarinen, Subhro Das, Lam M. Nguyen, and Tsui-Wei Weng. Label-free concept bottleneck models. In The Eleventh International Conference on Learning Representations, 2023. 2, 4, 6 +[27] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul F Christiano, Jan Leike, and Ryan Lowe. Training language models to follow instructions with human feedback. In Advances in Neural Information + +Processing Systems, pages 27730-27744. Curran Associates, Inc., 2022. 2 +[28] Ahmet Murat Ozbayoglu, Mehmet Ugur Gudelek, and Omer Berat Sezer. Deep learning for financial applications: A survey. Applied Soft Computing, 93:106384, 2020. 1 +[29] Konstantinos Panagiotis Panousis, Dino Ienc, and Diego Marcos. Sparse linear concept discovery models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2767-2771, 2023. 2 +[30] Fabio Petroni, Tim Roktaschel, Sebastian Riedel, Patrick Lewis, Anton Bakhtin, Yuxiang Wu, and Alexander Miller. Language models as knowledge bases? In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 2463-2473, Hong Kong, China, 2019. Association for Computational Linguistics. 2 +[31] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 2, 4 +[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2, 3, 5 +[33] Muhammad Imran Razzak, Saeeda Naz, and Ahmad Zaib. Deep Learning for Medical Image Processing: Overview, Challenges and the Future, page 323-350. Springer International Publishing, 2017. 1 +[34] Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. "why should i trust you?" explaining the predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining, pages 1135-1144, 2016. 1 +[35] Cynthia Rudin. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature Machine Intelligence, 1(5):206-215, 2019. 1 +[36] Gesina Schwalbe. Concept embedding analysis: A review. arXiv preprint arXiv:2203.13909, 2022. 1 +[37] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 1 +[38] Chenming Shang, Hengyuan Zhang, Hao Wen, and Yujiu Yang. Understanding multimodal deep neural networks: A concept selection view, 2024. 2 +[39] Chenming Shang, Shiji Zhou, Hengyuan Zhang, Xinzhe Ni, Yujiu Yang, and Yuwang Wang. Incremental residual concept bottleneck models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11030-11040, 2024. 2, 5, 6 +[40] Sheng Shen, Liunian Harold Li, Hao Tan, Mohit Bansal, Anna Rohrbach, Kai-Wei Chang, Zhewei Yao, and Kurt + +Keutzer. How much can CLIP benefit vision-and-language tasks? In International Conference on Learning Representations, 2022. 3 +[41] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 5 +[42] Robyn Speer, Joshua Chin, and Catherine Havasi. Conceptnet 5.5: An open multilingual graph of general knowledge. In Thirty-first AAAI conference on artificial intelligence, 2017. 2, 5, 1 +[43] Divyansh Srivastava, Ge Yan, and Tsui-Wei Weng. Vlg-cbm: Training concept bottleneck models with vision-language guidance, 2024. 2 +[44] Sandeep Subramanian, Adam Trischler, Yoshua Bengio, and Christopher J Pal. Learning general purpose distributed sentence representations via large scale multi-task learning. arXiv preprint arXiv:1804.00079, 2018. 2 +[45] Mujeen Sung, Jinhyuk Lee, Sean Yi, Minji Jeon, Sungdong Kim, and Jaewoo Kang. Can language models be biomedical knowledge bases? In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 4723-4734, Online and Punta Cana, Dominican Republic, 2021. Association for Computational Linguistics. 2 +[46] Yoad Tewel, Yoav Shalev, Idan Schwartz, and Lior Wolf. Zero-shot image-to-text generation for visual-semantic arithmetic. arXiv preprint arXiv:2111.14447, 2021. 3 +[47] Philipp Tschandl, Cliff Rosendahl, and Harald Kittler. The ham10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions. Scientific data, 5(1):1-9, 2018. 5 +[48] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 5 +[49] Bowen Wang, Liangzhi Li, Yuta Nakashima, and Hajime Nagahara. Learning bottleneck concepts in image classification. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2 +[50] Zirui Wang, Jiahui Yu, Adams Wei Yu, Zihang Dai, Yulia Tsvetkov, and Yuan Cao. SimVLM: Simple visual language model pretraining with weak supervision. In International Conference on Learning Representations, 2022. 4 +[51] Yifan Wu, Yang Liu, Yue Yang, Michael S. Yao, Wenli Yang, Xuehui Shi, Lihong Yang, Dongjun Li, Yueming Liu, James C. Gee, Xuan Yang, Wenbin Wei, and Shi Gu. A concept-based interpretable model for the diagnosis of choroid neoplasias using multimodal data, 2024. 7 +[52] Yue Yang, Artemis Panagopoulou, Shenghao Zhou, Daniel Jin, Chris Callison-Burch, and Mark Yatskar. Language in a bottle: Language model guided concept bottlenecks for interpretable image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19187-19197, 2023. 2, 4, 5, 6 +[53] Chih-Kuan Yeh, Been Kim, Sercan Arik, Chun-Liang Li, Tomas Pfister, and Pradeep Ravikumar. On completeness-aware concept-based explanations in deep neural networks. In Advances in Neural Information Processing Systems, pages 20554-20565. Curran Associates, Inc., 2020. 2 + +[54] Mert Yuksekgonul, Maggie Wang, and James Zou. Posthoc concept bottleneck models. In ICLR 2022 Workshop on PAIR²Struct, 2022. 2, 4, 5, 6 +[55] Yu Zhang, Peter Tiño, Alex Leonardis, and Ke Tang. A survey on neural network interpretability. IEEE Transactions on Emerging Topics in Computational Intelligence, 2021. 1 +[56] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 1 \ No newline at end of file diff --git a/2025/Hybrid Concept Bottleneck Models/images.zip b/2025/Hybrid Concept Bottleneck Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f1dcf7faa3f4446f44f05e0b354136253ab531b5 --- /dev/null +++ b/2025/Hybrid Concept Bottleneck Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:657585552d195734ccc9fad83a8bcd7f81f1bddae077afe87ada8d9c7ee18c00 +size 667648 diff --git a/2025/Hybrid Concept Bottleneck Models/layout.json b/2025/Hybrid Concept Bottleneck Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..42ae30b2f5598b10180ad5c5cd9a002dd4e25f74 --- /dev/null +++ b/2025/Hybrid Concept Bottleneck Models/layout.json @@ -0,0 +1,9337 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 195, + 103, + 416, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 103, + 416, + 121 + ], + "spans": [ + { + "bbox": [ + 195, + 103, + 416, + 121 + ], + "type": "text", + "content": "Hybrid Concept Bottleneck Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 164, + 144, + 446, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 144, + 446, + 185 + ], + "spans": [ + { + "bbox": [ + 164, + 144, + 446, + 185 + ], + "type": "text", + "content": "Yang Liu, Tianwei Zhang, Shi Gu \nUniversity of Electronic Science and Technology of China \nShenzhen Institute for Advanced Study, UESTC" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 121, + 188, + 485, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 188, + 485, + 200 + ], + "spans": [ + { + "bbox": [ + 121, + 188, + 485, + 200 + ], + "type": "text", + "content": "liu.yang.minéstd.uestc.edu.cn, gsbyztw@gmail.com, gus@uestc.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "spans": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 252, + 297, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 252, + 297, + 588 + ], + "spans": [ + { + "bbox": [ + 53, + 252, + 297, + 588 + ], + "type": "text", + "content": "Concept Bottleneck Models (CBMs) provide an interpretable framework for neural networks by mapping visual features to predefined, human-understandable concepts. However, the application of CBMs is often constrained by insufficient concept annotations. Recently, multi-modal pretrained models have shown promise in reducing annotation costs by aligning visual representations with textual concept embeddings. Nevertheless, the quality and completeness of the predefined concepts significantly affect the performance of CBMs. In this work, we propose Hybrid Concept Bottleneck Model (HybridCBM), a novel CBM framework to address the challenge of incomplete predefined concepts. Our method consists of two main components: a Static Concept Bank and a Dynamic Concept Bank. The Static Concept Bank directly leverages large language models (LLMs) for concept construction, while the Dynamic Concept Bank employs learnable vectors to capture complementary and valuable concepts continuously during training. After training, a pre-trained translator converts these vectors into human-understandable concepts, further enhancing model interpretability. HybridCBM is highly flexible and can be easily integrated with existing CBMs to improve both interpretability and performance. Experimental results1 on multiple datasets demonstrate that HybridCBM outperforms current state-of-the-art CBMs and achieves comparable results to black-box models. Additionally, we propose novel metrics to assess the quality of learned concepts, showing that they perform comparably to predefined concepts." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 612, + 137, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 612, + 137, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 612, + 137, + 624 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 632, + 296, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 632, + 296, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 632, + 296, + 694 + ], + "type": "text", + "content": "Deep Neural Networks (DNNs) have become increasingly dominant in various fields, including computer vision, natural language processing, and speech recognition. However their complex and deep structures present a significant challenge to interpretability, often earning them the label" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 225, + 555, + 322 + ], + "blocks": [ + { + "bbox": [ + 316, + 225, + 555, + 322 + ], + "lines": [ + { + "bbox": [ + 316, + 225, + 555, + 322 + ], + "spans": [ + { + "bbox": [ + 316, + 225, + 555, + 322 + ], + "type": "image", + "image_path": "4e6be5eabccdd36d34ca15099725a4981584a8936aa2bf2b1af6b458872f484c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 327, + 555, + 372 + ], + "lines": [ + { + "bbox": [ + 313, + 327, + 555, + 372 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 555, + 372 + ], + "type": "text", + "content": "Figure 1. Our proposed concept-trainable CBM reduces the need for meticulously designed concepts by incorporating learned concepts (e.g., 'yellowish throat'), thereby enhancing interpretability and performance while minimizing manual design effort." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 378, + 555, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 378, + 555, + 534 + ], + "spans": [ + { + "bbox": [ + 312, + 378, + 555, + 534 + ], + "type": "text", + "content": "of 'blackbox' [23]. Addressing this challenge is crucial not only for ensuring accurate predictions, but also for providing scientists and engineers with more accessible tools for designing, developing, and debugging models. Interpretability is particularly critical in domains such as healthcare [10], medicine [33], education [7], and finance [28], where high-stakes decision-making requires transparency and trustworthiness to ensure reliability and accountability. Therefore, there has been a growing focus on research aimed at developing interpretable models. While post-hoc explanation methods [34, 37, 55] have gained popularity, they often fall short by providing unfaithful representations of the model's underlying computations [35]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "text", + "content": "A natural approach to improving interpretability in DNNs is to design inherently interpretable models that leverage high-level concepts to describe features extracted by black-box models [36], imitating the way humans encode and understand the world[20]. One notable approach following such methodology is the Concept Bottleneck Models (CBM) [17]. CBMs operate by first predicting concepts from image representations mid-way through the decision-making process. These models pair each image with its predefined concepts and are trained in an end-to-end manner—initially predicting concepts and subsequently using these predictions to make final decisions. While this method of concept prediction enhances model interpretability, it often faces two fundamental challenges: (1) it is perceived to underperform compared to black-box models [11]," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 268, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 268, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 268, + 713 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 66, + 702, + 268, + 713 + ], + "type": "text", + "content": "Code available at https://github.com/ly1998117/HybridCBM" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20179" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 96 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 96 + ], + "type": "text", + "content": "and (2) the requirement for precise annotations for each concept significantly limits their broader application." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 100, + 296, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 100, + 296, + 423 + ], + "spans": [ + { + "bbox": [ + 56, + 100, + 296, + 423 + ], + "type": "text", + "content": "To mitigate these shortcomings, recent research has explored multimodal pretrained model-based CBMs [15, 52, 54], leveraging CLIP [32] to establish correspondences between visual and textual representations, thereby reducing the reliance on costly concept annotations. Specifically, CLIP encodes images and human-understandable concepts into a shared feature space, enabling the creation of a concept bottleneck by directly mapping visual representations to concept embeddings instead of through manual annotations. Regarding the trade-off between interpretability and performance, some studies [15, 38, 52] have emphasized the construction of high-quality concept banks through efficient concept selection. However, establishing a comprehensive concept bank remains a significant challenge [53], often constrained by the inherent limitations in language expressiveness and precision. In response, other research efforts have explored the integration of a residual linear layer to address the issue of incomplete concept extraction [39, 54], albeit at the cost of interpretability. For example, ResCBM [39] introduces estimizable vectors to capture missing concepts, utilizing a candidate concept bank in the residual component to regularizing these vectors toward specific candidates during training. However, the model is limited by the scope of the candidate concept bank, restricting its ability to learn concepts outside the predefined set. Additionally, the effectiveness of the residual structure remains an open question." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 426, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 296, + 714 + ], + "type": "text", + "content": "In this work, we propose the Hybrid Concept Bottleneck Model (HybridCBM), specifically designed to address the challenges of incomplete concept representation, predefined concept bank dependence, and scalability issue by dynamically discovering new concepts directly from visual representations. Different from prior methods that rely on residual structure or external concept banks, our approach incorporates a hybrid concept bank comprising both a static and a dynamic concept bank. Initially, we establish the static concept bank by leveraging a large language model (LLM) like GPT-3.5 [27, 31], which is celebrated for its broad world knowledge [12, 30, 45]. For instance, in Figure 1, when prompted about a palm warbler, GPT-3.5 provides information such as 'feeds on the ground'. This textual descriptions is then encoded into embeddings by the text encoder of CLIP. Next, we initialize a set of estimizable vectors that form the dynamic concept bank. These vectors are refined to identify concepts present in the input image but absent from the static concept bank. To further enhance the concept discovery, we pre-train a concept translator, such as GPT-2 [31], to translate these newly discovered unknown concepts into textual descriptions, a process we term as 'concept labeling'. To evaluate the understandability, visual relevance and factual accuracy of the dynamic concepts, we" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "type": "text", + "content": "further design a set of evaluation metrics, some of which are based on Vision-Language Model (VLM) GPT-4o" + }, + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "type": "text", + "content": ". Our main contributions can be summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 109, + 555, + 300 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 315, + 109, + 555, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 109, + 555, + 180 + ], + "spans": [ + { + "bbox": [ + 315, + 109, + 555, + 180 + ], + "type": "text", + "content": "1. Innovative Hybrid Concept Bank: Our model introduces a hybrid concept bank that combines static and dynamic concepts, allowing it to adapt and refine its interpretative capabilities dynamically. This structure leverages both predefined expert knowledge while enabling the discovery of new relevant concepts." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 181, + 553, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 181, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 313, + 181, + 553, + 252 + ], + "type": "text", + "content": "2. Seamless Integration of LLM Technologies: By employing advanced language models (LLM for defining static concepts, translating new concepts and VLM for evaluation) alongside the visual-textual bridging capabilities of CLIP, our model bridges the gap between visual representation and textual interpretation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 253, + 553, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 253, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 253, + 553, + 300 + ], + "type": "text", + "content": "3. Enhanced Interpretability and Performance: Our approach maintains high interpretability through transparent concept labeling, while improving performance by discovering new concepts from visual information." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 313, + 400, + 325 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 313, + 400, + 325 + ], + "spans": [ + { + "bbox": [ + 314, + 313, + 400, + 325 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 334, + 555, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 334, + 555, + 693 + ], + "spans": [ + { + "bbox": [ + 313, + 334, + 555, + 693 + ], + "type": "text", + "content": "Interpretable Neural Networks. One way to build an interpretable neural networks is the use of a concept-based explanation [3, 13, 15, 21, 29, 38, 43, 49]. Concept Bottleneck Models (CBMs) [17] are among the most popular approaches for making predictions based on human-interpretable concepts but require labor-intensive concept annotations for each image. Similarly, Concept Activation Vectors (CAVs) [14] represent concepts as normal vectors to the decision boundaries that separate positive and negative samples of a concept, though they also need extra datasets to train SVMs for each concept. Owing to the simple structure based on human-defined concepts, there have been studies aiming to alleviate these drawbacks, e.g., costly concept annotation and trade-off between interpretability and performance. Post-hoc Concept Bottleneck (PCBM) [54] utilizes ConceptNet [42] to obtain concepts and Leveraging CLIP to generate bottlenecks by projecting visual representations into the concept subspace. However, PCBM-h [54] incorporates a residual linear predictor that compensates for missing concepts by directly adding logits to PCBM's predictor. While this modification helps to recover accuracy, it does so at the expense of interpretability. The label-free CBM [26] uses CLIP for concept annotation, allowing arbitrary visual backbone to be transformed into an CBM without the need for labeled concept data. Labo [52] concentrates on concepts selection after generating candidate concepts through LLM. ResCBM [39] builds on PCBM-h's residual structure by using a candidate concept bank to guide and refine estimizable vectors. However, the model's expressive potential is limited by the extent of the candidate concept bank." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 324, + 702, + 455, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 702, + 455, + 713 + ], + "spans": [ + { + "bbox": [ + 324, + 702, + 455, + 713 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 324, + 702, + 455, + 713 + ], + "type": "text", + "content": "https://openai.com/index/hello-gpt-4o/" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20180" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 69, + 541, + 300 + ], + "blocks": [ + { + "bbox": [ + 70, + 69, + 541, + 300 + ], + "lines": [ + { + "bbox": [ + 70, + 69, + 541, + 300 + ], + "spans": [ + { + "bbox": [ + 70, + 69, + 541, + 300 + ], + "type": "image", + "image_path": "4c63391b7a4a249706c14a7a45c35f7af6190a369095e00c5ccf7c4b8a37d558.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 304, + 555, + 361 + ], + "lines": [ + { + "bbox": [ + 54, + 304, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 54, + 304, + 555, + 361 + ], + "type": "text", + "content": "Figure 2. Overview of the Hybrid Concept Bottleneck Model (HybridCBM), which is designed for interpretable concept learning and labeling. First, a Concept Translator is pre-trained to do concept-labeling. Next, a pretrained alignment model (CLIP) is used to obtain embeddings of predefined concepts, forming the Static Concept Bank. The Dynamic Concept Bank learns missing concepts and aligns its distribution with the Static Concept Bank to create the final concept bottleneck. Finally, an interpretable concept-based predictor classifies targets based on the concept scores." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 366, + 297, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 366, + 297, + 534 + ], + "spans": [ + { + "bbox": [ + 54, + 366, + 297, + 534 + ], + "type": "text", + "content": "CLIP in Image Captioning. Multimodal pretrained models like CLIP [32], trained on over 400 million image-text pairs using unsupervised contrastive loss, exhibit impressive capabilities in image captioning tasks [1, 19, 46]. Previous studies, such as [24] and [40], have primarily used CLIP as a visual encoder for captioning tasks, while [19] projects visual embeddings into CLIP's text embedding space to preserve visual information. However, these approaches often underutilize CLIP's aligned multi-modal latent space. Similar to [19, 24], we collect image-caption pairs and a vast corpus to train a GPT-2 from scratch as our concept translator, fully leveraging CLIP's latent space and bridging the gap between image and concept embeddings and their corresponding textual descriptions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 548, + 112, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 548, + 112, + 561 + ], + "spans": [ + { + "bbox": [ + 55, + 548, + 112, + 561 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "text", + "content": "We provide an overview of our methodology in Figure 2. Our model incorporates a hybrid concept bank and a pretrained concept translator (Section 3.2). The concept bank consists of a set of estimizable vectors, which are refined to discover new concepts through a concept-interpretability training scheme (Section 3.3). As a result, the learned concepts are not only discriminative and orthogonal but also aligned with human-understandable semantics. The pre-trained concept translator performs concept labeling by mapping these learned, initially unknown vectors to meaningful human concepts. Finally, we apply a sparse linear layer over the similarity scores between hybrid concepts" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 366, + 555, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 366, + 555, + 390 + ], + "spans": [ + { + "bbox": [ + 313, + 366, + 555, + 390 + ], + "type": "text", + "content": "and images, enabling us to identify the specific concepts the model relies on for its decisions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 396, + 440, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 440, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 440, + 407 + ], + "type": "text", + "content": "3.1. Problem Formulation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 415, + 555, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 555, + 510 + ], + "type": "text", + "content": "Consider a dataset " + }, + { + "bbox": [ + 313, + 415, + 555, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = (x,y)" + }, + { + "bbox": [ + 313, + 415, + 555, + 510 + ], + "type": "text", + "content": ", where each image " + }, + { + "bbox": [ + 313, + 415, + 555, + 510 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 415, + 555, + 510 + ], + "type": "text", + "content": " is paired with a label " + }, + { + "bbox": [ + 313, + 415, + 555, + 510 + ], + "type": "inline_equation", + "content": "y\\in \\mathcal{V}" + }, + { + "bbox": [ + 313, + 415, + 555, + 510 + ], + "type": "text", + "content": ". We use the image encoder " + }, + { + "bbox": [ + 313, + 415, + 555, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{img}" + }, + { + "bbox": [ + 313, + 415, + 555, + 510 + ], + "type": "text", + "content": " and text encoder " + }, + { + "bbox": [ + 313, + 415, + 555, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{txt}" + }, + { + "bbox": [ + 313, + 415, + 555, + 510 + ], + "type": "text", + "content": " of CLIP [32] to map images and text into a shared " + }, + { + "bbox": [ + 313, + 415, + 555, + 510 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 415, + 555, + 510 + ], + "type": "text", + "content": "-dimensional feature space, respectively. A pre-trained concept translator " + }, + { + "bbox": [ + 313, + 415, + 555, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 313, + 415, + 555, + 510 + ], + "type": "text", + "content": " then maps these shared " + }, + { + "bbox": [ + 313, + 415, + 555, + 510 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 415, + 555, + 510 + ], + "type": "text", + "content": "-dimensional features back to textual concepts. In order to store the interpretable information of the model, we construct a hybrid concept bank " + }, + { + "bbox": [ + 313, + 415, + 555, + 510 + ], + "type": "inline_equation", + "content": "E_{C}\\in \\mathbb{R}^{N_{C}\\times d}" + }, + { + "bbox": [ + 313, + 415, + 555, + 510 + ], + "type": "text", + "content": ", comprising:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 510, + 555, + 630 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 314, + 510, + 555, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 510, + 555, + 558 + ], + "spans": [ + { + "bbox": [ + 314, + 510, + 555, + 558 + ], + "type": "text", + "content": "- Static Concept Bank " + }, + { + "bbox": [ + 314, + 510, + 555, + 558 + ], + "type": "inline_equation", + "content": "E_{S} \\in \\mathbb{R}^{N_{S} \\times d}" + }, + { + "bbox": [ + 314, + 510, + 555, + 558 + ], + "type": "text", + "content": ". Contains text features " + }, + { + "bbox": [ + 314, + 510, + 555, + 558 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{txt}(c) \\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 314, + 510, + 555, + 558 + ], + "type": "text", + "content": " extracted by the text encoder " + }, + { + "bbox": [ + 314, + 510, + 555, + 558 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{txt}" + }, + { + "bbox": [ + 314, + 510, + 555, + 558 + ], + "type": "text", + "content": ". Each row corresponds to a predefined concept " + }, + { + "bbox": [ + 314, + 510, + 555, + 558 + ], + "type": "inline_equation", + "content": "C_{S} = \\{c_{1}, c_{2}, \\ldots, c_{N_{S}}\\}" + }, + { + "bbox": [ + 314, + 510, + 555, + 558 + ], + "type": "text", + "content": " generated from a large language model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 558, + 555, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 558, + 555, + 630 + ], + "spans": [ + { + "bbox": [ + 314, + 558, + 555, + 630 + ], + "type": "text", + "content": "- Dynamic Concept Bank " + }, + { + "bbox": [ + 314, + 558, + 555, + 630 + ], + "type": "inline_equation", + "content": "E_{D} \\in \\mathbb{R}^{N_{D} \\times d}" + }, + { + "bbox": [ + 314, + 558, + 555, + 630 + ], + "type": "text", + "content": ". Consists of a set of randomly initialized learnable vectors " + }, + { + "bbox": [ + 314, + 558, + 555, + 630 + ], + "type": "inline_equation", + "content": "e_{i} \\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 314, + 558, + 555, + 630 + ], + "type": "text", + "content": " that are optimized to capture missing visual-specific features. After training, These vectors could be mapped to the textual concept space " + }, + { + "bbox": [ + 314, + 558, + 555, + 630 + ], + "type": "inline_equation", + "content": "C_{D} = \\{c_{1}, c_{2}, \\ldots, c_{N_{D}}\\}" + }, + { + "bbox": [ + 314, + 558, + 555, + 630 + ], + "type": "text", + "content": " by the translator " + }, + { + "bbox": [ + 314, + 558, + 555, + 630 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 314, + 558, + 555, + 630 + ], + "type": "text", + "content": ", where each concept " + }, + { + "bbox": [ + 314, + 558, + 555, + 630 + ], + "type": "inline_equation", + "content": "c = \\mathcal{T}(e)" + }, + { + "bbox": [ + 314, + 558, + 555, + 630 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "The overall hybrid concept bank is defined as " + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "inline_equation", + "content": "E_{C} = E_{S}\\oplus" + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "inline_equation", + "content": "E_{D}" + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": " , where " + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": " denotes concatenation. The total number of concepts in the bank is indicated by " + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "inline_equation", + "content": "N_{C} = N_{S} + N_{D}" + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": " . To compute the concept scores, we utilize the CLIP image encoder to extract the image feature " + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "inline_equation", + "content": "f = \\mathcal{F}_{img}(x)\\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": " Therein, the similarity between " + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "inline_equation", + "content": "E_{C}" + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": " is computed using projection length, resulting in the concept score " + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "inline_equation", + "content": "S\\in \\mathbb{R}^{N_C}" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20181" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 232 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 232 + ], + "type": "text", + "content": "which reflects the presence of particular concepts in the image " + }, + { + "bbox": [ + 55, + 72, + 296, + 232 + ], + "type": "inline_equation", + "content": "x\\colon S_x = \\| f\\| _2\\cos (f,E_C) = f\\cdot \\hat{E_C}^T" + }, + { + "bbox": [ + 55, + 72, + 296, + 232 + ], + "type": "text", + "content": " , where " + }, + { + "bbox": [ + 55, + 72, + 296, + 232 + ], + "type": "inline_equation", + "content": "\\hat{\\cdot}" + }, + { + "bbox": [ + 55, + 72, + 296, + 232 + ], + "type": "text", + "content": " denotes " + }, + { + "bbox": [ + 55, + 72, + 296, + 232 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 55, + 72, + 296, + 232 + ], + "type": "text", + "content": " normalization and cos is the cosine similarity. Since " + }, + { + "bbox": [ + 55, + 72, + 296, + 232 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 72, + 296, + 232 + ], + "type": "text", + "content": " provides a unique representation for each image, it serves as a concept bottleneck. These per-image representations can be naturally used to support an interpretable classification. Following the existing CLIP-based CBMs [26, 52, 54], we train a linear layer as the concept classifier " + }, + { + "bbox": [ + 55, + 72, + 296, + 232 + ], + "type": "inline_equation", + "content": "\\phi :\\hat{y} = \\phi (S),\\mathbb{R}^{N_C}\\to \\mathcal{V}" + }, + { + "bbox": [ + 55, + 72, + 296, + 232 + ], + "type": "text", + "content": " to make final predictions in the label space based on the concept scores. In our perspective, we intend to improve the performance and learn abundant concepts from the visual information at the same time. This can be solved by the following optimization problem:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 237, + 295, + 278 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 237, + 295, + 278 + ], + "spans": [ + { + "bbox": [ + 69, + 237, + 295, + 278 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\min _ {\\phi , E _ {D}} \\underset {(x, y) \\sim \\mathcal {D}} {\\mathbb {E}} \\left[ \\mathcal {L} _ {c l s} \\left(\\phi \\left(f \\cdot \\hat {E} _ {S} ^ {T} \\oplus f \\cdot \\hat {E} _ {D} ^ {T}\\right), y\\right) \\right] \\tag {1} \\\\ + \\lambda \\Omega (\\phi) + \\Gamma_ {E _ {D}} \\\\ \\end{array}", + "image_path": "e797d9c8e677d6f59568f998ca433771551fcdf34dd7d3ea6f8526dccab1d4e2.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 285, + 296, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 285, + 296, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 296, + 346 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 285, + 296, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cls}(\\hat{y}, y)" + }, + { + "bbox": [ + 55, + 285, + 296, + 346 + ], + "type": "text", + "content": " is the cross-entropy loss function, " + }, + { + "bbox": [ + 55, + 285, + 296, + 346 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 55, + 285, + 296, + 346 + ], + "type": "text", + "content": " is a complexity measure used to promote sparsity in the linear layer, " + }, + { + "bbox": [ + 55, + 285, + 296, + 346 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 285, + 296, + 346 + ], + "type": "text", + "content": " is the regularization strength and " + }, + { + "bbox": [ + 55, + 285, + 296, + 346 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 55, + 285, + 296, + 346 + ], + "type": "text", + "content": " represents the optimization objective for the Dynamic Concept Bank " + }, + { + "bbox": [ + 55, + 285, + 296, + 346 + ], + "type": "inline_equation", + "content": "E_{D}" + }, + { + "bbox": [ + 55, + 285, + 296, + 346 + ], + "type": "text", + "content": ", introduced in Section 3.3." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 352, + 252, + 365 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 352, + 252, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 352, + 252, + 365 + ], + "type": "text", + "content": "3.2. Hybrid Concept Bank Establishment" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 370, + 296, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 370, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 370, + 296, + 479 + ], + "type": "text", + "content": "Previous methods, such as Labo [52], generate a set of relevant concepts and then form a concept bank through the submodular selection method. However, the predefined concepts often suffer from inherent flaws due to their limited scope, and the selection process may inadvertently omit important information. To address these issues, we propose a Hybrid Concept Bank that dynamically discovers concepts directly from the visual representation, combining both a Static Concept Bank and a Dynamic Concept Bank." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 480, + 296, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 480, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 480, + 296, + 601 + ], + "type": "text", + "content": "Static Concept Bank. Following Labo, we utilize GPT-3.5 to generate a set of candidate concepts for each class name " + }, + { + "bbox": [ + 55, + 480, + 296, + 601 + ], + "type": "inline_equation", + "content": "y \\in \\mathcal{V}" + }, + { + "bbox": [ + 55, + 480, + 296, + 601 + ], + "type": "text", + "content": " with prompts such as \"describe what the looks like\". After initially generating 500 concepts per class, we perform concept selection to ensure a balanced representation. Specifically, we select " + }, + { + "bbox": [ + 55, + 480, + 296, + 601 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 480, + 296, + 601 + ], + "type": "text", + "content": " concepts for each class using both submodular and random selection methods for comparison. Assume " + }, + { + "bbox": [ + 55, + 480, + 296, + 601 + ], + "type": "inline_equation", + "content": "K = |\\mathcal{V}|" + }, + { + "bbox": [ + 55, + 480, + 296, + 601 + ], + "type": "text", + "content": ", this process yields a total of " + }, + { + "bbox": [ + 55, + 480, + 296, + 601 + ], + "type": "inline_equation", + "content": "N_S = N \\times K" + }, + { + "bbox": [ + 55, + 480, + 296, + 601 + ], + "type": "text", + "content": " concepts, which are then encoded by the text encoder " + }, + { + "bbox": [ + 55, + 480, + 296, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{txt}" + }, + { + "bbox": [ + 55, + 480, + 296, + 601 + ], + "type": "text", + "content": " of CLIP and fixed during training." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 603, + 296, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 603, + 296, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 603, + 296, + 676 + ], + "type": "text", + "content": "Dynamic Concept Bank. Unlike the static counterpart, the Dynamic Concept Bank is initialized with a set of concept vectors " + }, + { + "bbox": [ + 55, + 603, + 296, + 676 + ], + "type": "inline_equation", + "content": "e_d \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 55, + 603, + 296, + 676 + ], + "type": "text", + "content": ", which are optimized through the proposed Concept-Interpretability Training Scheme (Section 3.3). Thus, it can adaptively capture the unique and rich visual-specific features during the training process." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Concept Labeling. To enhance the interpretability of learned concepts, we pre-train a concept translator using GPT-2 [31] architecture to invert the CLIP text encoder." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": "Following recent work [19, 24, 50], we use prefix language modeling for training. Details of the pre-training process are provided in Appendix B.2. The translator maps each concept embedding " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": " to its textual description " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": " through " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "c = \\mathcal{T}(e)" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": ", assigning semantic meaning to the learned concepts and improving interpretability." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 152, + 534, + 165 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 152, + 534, + 165 + ], + "spans": [ + { + "bbox": [ + 313, + 152, + 534, + 165 + ], + "type": "text", + "content": "3.3. Concept-Interpretability Training Scheme" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 170, + 555, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 170, + 555, + 315 + ], + "spans": [ + { + "bbox": [ + 313, + 170, + 555, + 315 + ], + "type": "text", + "content": "The previous CLIP-based CBMs rely solely on predefined concepts, often overlooking the potential of learning concepts directly from visual representation. Since CLIP maps both images and texts into a shared feature space, we could then leverage this capability to optimize the dynamic concept bank " + }, + { + "bbox": [ + 313, + 170, + 555, + 315 + ], + "type": "inline_equation", + "content": "E_{D}" + }, + { + "bbox": [ + 313, + 170, + 555, + 315 + ], + "type": "text", + "content": " based on the given image embedding " + }, + { + "bbox": [ + 313, + 170, + 555, + 315 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 170, + 555, + 315 + ], + "type": "text", + "content": ". To achieve this, we introduce a concept-interpretability training scheme that regularizes " + }, + { + "bbox": [ + 313, + 170, + 555, + 315 + ], + "type": "inline_equation", + "content": "E_{D}" + }, + { + "bbox": [ + 313, + 170, + 555, + 315 + ], + "type": "text", + "content": " through multiple loss functions. The goal is to obtain a set of learnable concepts that are discriminative, diverse, and semantically aligned with human-understandable meanings, thereby enhancing both model interpretability and performance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 316, + 556, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 556, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 556, + 449 + ], + "type": "text", + "content": "Discriminability Loss. The discriminability loss ensures that learned concepts are highly activated for images of their own class and less activated for images of other classes. We first divide the dynamic concept vectors into several subsets corresponding to the classes, such that " + }, + { + "bbox": [ + 313, + 316, + 556, + 449 + ], + "type": "inline_equation", + "content": "E_D = \\{E_D^1, E_D^2, \\ldots, E_D^K\\}" + }, + { + "bbox": [ + 313, + 316, + 556, + 449 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 316, + 556, + 449 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 316, + 556, + 449 + ], + "type": "text", + "content": " is the number of classes. Each subset " + }, + { + "bbox": [ + 313, + 316, + 556, + 449 + ], + "type": "inline_equation", + "content": "E_D^k" + }, + { + "bbox": [ + 313, + 316, + 556, + 449 + ], + "type": "text", + "content": " contains " + }, + { + "bbox": [ + 313, + 316, + 556, + 449 + ], + "type": "inline_equation", + "content": "N_D / K" + }, + { + "bbox": [ + 313, + 316, + 556, + 449 + ], + "type": "text", + "content": " dynamic concept vectors associated with class " + }, + { + "bbox": [ + 313, + 316, + 556, + 449 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 316, + 556, + 449 + ], + "type": "text", + "content": ". To encourage vectors " + }, + { + "bbox": [ + 313, + 316, + 556, + 449 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_k \\in E_D^k" + }, + { + "bbox": [ + 313, + 316, + 556, + 449 + ], + "type": "text", + "content": " to learn specific information of images within class " + }, + { + "bbox": [ + 313, + 316, + 556, + 449 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 316, + 556, + 449 + ], + "type": "text", + "content": " and incorporate semantic information from the class name, we define the intra-class loss:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 331, + 454, + 555, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 454, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 331, + 454, + 555, + 475 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {i n t r a} ^ {i} = y _ {k} \\left\\{- \\hat {f} \\cdot \\hat {c} _ {k} + \\alpha \\left| m - \\hat {\\mathcal {F}} _ {t x t} (k) \\cdot \\hat {c} _ {k} \\right| \\right\\} \\quad (2)", + "image_path": "8da81b9494066d788d567107b5f6daa025d37d47d947d26b27f1bb369aff397d.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 483, + 554, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 554, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 554, + 579 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 483, + 554, + 579 + ], + "type": "inline_equation", + "content": "f = \\mathcal{F}_{img}(x)" + }, + { + "bbox": [ + 313, + 483, + 554, + 579 + ], + "type": "text", + "content": " is the image feature, " + }, + { + "bbox": [ + 313, + 483, + 554, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{txt}(k)" + }, + { + "bbox": [ + 313, + 483, + 554, + 579 + ], + "type": "text", + "content": " is the class embedding, " + }, + { + "bbox": [ + 313, + 483, + 554, + 579 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_k" + }, + { + "bbox": [ + 313, + 483, + 554, + 579 + ], + "type": "text", + "content": " is the sampled concept vector for class " + }, + { + "bbox": [ + 313, + 483, + 554, + 579 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 483, + 554, + 579 + ], + "type": "text", + "content": " from the distribution of vectors " + }, + { + "bbox": [ + 313, + 483, + 554, + 579 + ], + "type": "inline_equation", + "content": "E_D^k" + }, + { + "bbox": [ + 313, + 483, + 554, + 579 + ], + "type": "text", + "content": " via reparameterization" + }, + { + "bbox": [ + 313, + 483, + 554, + 579 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 313, + 483, + 554, + 579 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 483, + 554, + 579 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 483, + 554, + 579 + ], + "type": "text", + "content": " is a hyperparameter and " + }, + { + "bbox": [ + 313, + 483, + 554, + 579 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 483, + 554, + 579 + ], + "type": "text", + "content": " is a margin used to prevent excessive similarity to class " + }, + { + "bbox": [ + 313, + 483, + 554, + 579 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 483, + 554, + 579 + ], + "type": "text", + "content": ", with a value of 0.85 as defined in Label-Free [26]. To ensure that concepts are distinct between different classes, we further define the inter-class loss:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 319, + 585, + 555, + 607 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 585, + 555, + 607 + ], + "spans": [ + { + "bbox": [ + 319, + 585, + 555, + 607 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {i n t e r} ^ {i} = (y _ {k} - 1) \\left\\{- \\hat {f} \\cdot \\hat {c} _ {k} + \\alpha \\left| m - \\hat {\\mathcal {F}} _ {t x t} (k) \\cdot \\hat {c} _ {k} \\right| \\right\\} (3)", + "image_path": "dfb6df593f515c3b7505ee0f069876cc1870461277bfad008a97d2a1523f1d33.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 613, + 554, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 613, + 554, + 638 + ], + "spans": [ + { + "bbox": [ + 313, + 613, + 554, + 638 + ], + "type": "text", + "content": "The total discriminability loss is defined as " + }, + { + "bbox": [ + 313, + 613, + 554, + 638 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{dis} = \\mathcal{L}_{intra} + \\beta \\mathcal{L}_{inter}" + }, + { + "bbox": [ + 313, + 613, + 554, + 638 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 613, + 554, + 638 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 613, + 554, + 638 + ], + "type": "text", + "content": " is a balancing hyperparameter." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 639, + 554, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 639, + 554, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 639, + 554, + 677 + ], + "type": "text", + "content": "Orthogonality Loss. To enhance the diversity of dynamic concept embeddings and reduce redundancy, we introduce an orthogonality loss, which comprises two components:" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 683, + 555, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 683, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 683, + 555, + 713 + ], + "type": "inline_equation", + "content": "{}^{3}\\mathbf{c}_{k} = \\mu_{k} + \\sigma_{k}\\cdot \\epsilon" + }, + { + "bbox": [ + 313, + 683, + 555, + 713 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 313, + 683, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mu_{k}" + }, + { + "bbox": [ + 313, + 683, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 683, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\sigma_{k}" + }, + { + "bbox": [ + 313, + 683, + 555, + 713 + ], + "type": "text", + "content": " being the mean and standard deviation of " + }, + { + "bbox": [ + 313, + 683, + 555, + 713 + ], + "type": "inline_equation", + "content": "E_D^k" + }, + { + "bbox": [ + 313, + 683, + 555, + 713 + ], + "type": "text", + "content": ", respectively, and " + }, + { + "bbox": [ + 313, + 683, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0,I)" + }, + { + "bbox": [ + 313, + 683, + 555, + 713 + ], + "type": "text", + "content": " is sampled from the standard normal distribution." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20182" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": "intra-class diversity and inter-bank diversity. First, we promote the dynamic concept embeddings within the same class to be orthogonal to each other. Specifically, we define the intra-class diversity loss as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 128, + 295, + 162 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 128, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 69, + 128, + 295, + 162 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {o r t - i n t r a}} = \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\frac {1}{(N _ {D} / K) ^ {2}} \\sum_ {i \\neq j} \\left| \\hat {\\mathbf {e}} _ {D, i} ^ {k} \\cdot \\hat {\\mathbf {e}} _ {D, j} ^ {k} \\right| \\tag {4}", + "image_path": "6253ea352af191197419d0d68c997fa84811b678c31f4cd95e98688a4a7a35a8.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 169, + 296, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 169, + 296, + 231 + ], + "spans": [ + { + "bbox": [ + 55, + 169, + 296, + 231 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 169, + 296, + 231 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_{D,i}^{k}" + }, + { + "bbox": [ + 55, + 169, + 296, + 231 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 169, + 296, + 231 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_{D,j}^{k}" + }, + { + "bbox": [ + 55, + 169, + 296, + 231 + ], + "type": "text", + "content": " are dynamic concept embeddings associated with class " + }, + { + "bbox": [ + 55, + 169, + 296, + 231 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 169, + 296, + 231 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 169, + 296, + 231 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 169, + 296, + 231 + ], + "type": "text", + "content": " is the number of classes, and " + }, + { + "bbox": [ + 55, + 169, + 296, + 231 + ], + "type": "inline_equation", + "content": "N_{D}" + }, + { + "bbox": [ + 55, + 169, + 296, + 231 + ], + "type": "text", + "content": " is the total number of dynamic concepts. Second, to minimize redundancy between static and dynamic concepts, we define the inter-bank diversity loss as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 89, + 236, + 295, + 270 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 236, + 295, + 270 + ], + "spans": [ + { + "bbox": [ + 89, + 236, + 295, + 270 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {o r t - i n t e r}} = \\frac {1}{N _ {S} N _ {D}} \\sum_ {i = 1} ^ {N _ {S}} \\sum_ {j = 1} ^ {N _ {D}} | \\hat {\\mathbf {e}} _ {S, i} \\cdot \\hat {\\mathbf {e}} _ {D, j} | \\tag {5}", + "image_path": "4f5ddf4abdafc30d73ef2afd8353f7f7982d7f3b02ce3a0c6895f8848df28f4b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 277, + 296, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 296, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 296, + 326 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 277, + 296, + 326 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_{S,i}" + }, + { + "bbox": [ + 55, + 277, + 296, + 326 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 277, + 296, + 326 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_{D,j}" + }, + { + "bbox": [ + 55, + 277, + 296, + 326 + ], + "type": "text", + "content": " are static and dynamic concept embeddings, respectively, and " + }, + { + "bbox": [ + 55, + 277, + 296, + 326 + ], + "type": "inline_equation", + "content": "N_{S}" + }, + { + "bbox": [ + 55, + 277, + 296, + 326 + ], + "type": "text", + "content": " is the number of static concepts. The total orthogonality loss is then defined as " + }, + { + "bbox": [ + 55, + 277, + 296, + 326 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ort} = \\mathcal{L}_{ort - intra} + \\mathcal{L}_{ort - inter}" + }, + { + "bbox": [ + 55, + 277, + 296, + 326 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 327, + 296, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 296, + 448 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 296, + 448 + ], + "type": "text", + "content": "Distribution Alignment Loss. To enhance consistency between the distributions of dynamic and static concepts, we employ the Sinkhorn divergence [8], an entropy-regularized version of the Sinkhorn distance. We define the distribution alignment loss as: " + }, + { + "bbox": [ + 55, + 327, + 296, + 448 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{align}} = \\mathfrak{S}_{div,\\epsilon}(E_D,E_S)" + }, + { + "bbox": [ + 55, + 327, + 296, + 448 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 327, + 296, + 448 + ], + "type": "inline_equation", + "content": "\\mathfrak{S}_{div,\\epsilon}" + }, + { + "bbox": [ + 55, + 327, + 296, + 448 + ], + "type": "text", + "content": " denotes the Sinkhorn divergence, " + }, + { + "bbox": [ + 55, + 327, + 296, + 448 + ], + "type": "inline_equation", + "content": "E_{D}\\in \\mathbb{R}^{N_{D}\\times d}" + }, + { + "bbox": [ + 55, + 327, + 296, + 448 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 327, + 296, + 448 + ], + "type": "inline_equation", + "content": "E_{S}\\in \\mathbb{R}^{N_{S}\\times d}" + }, + { + "bbox": [ + 55, + 327, + 296, + 448 + ], + "type": "text", + "content": " are the dynamic and static concept embeddings, respectively. This loss encourages the distribution of dynamic concepts to align with that of static concepts, thus promoting semantic consistency and interpretability." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 450, + 296, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 296, + 499 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 296, + 499 + ], + "type": "text", + "content": "Overall Loss Function. The total loss function combines the classification loss with the proposed regularization terms " + }, + { + "bbox": [ + 55, + 450, + 296, + 499 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\mathcal{L}_{\\mathrm{cls}} + \\lambda_{\\mathrm{dis}}\\mathcal{L}_{\\mathrm{dis}} + \\lambda_{\\mathrm{ort}}\\mathcal{L}_{\\mathrm{ort}} + \\lambda_{\\mathrm{align}}\\mathcal{L}_{\\mathrm{align}}" + }, + { + "bbox": [ + 55, + 450, + 296, + 499 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 450, + 296, + 499 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{dis}}" + }, + { + "bbox": [ + 55, + 450, + 296, + 499 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 450, + 296, + 499 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{ort}}" + }, + { + "bbox": [ + 55, + 450, + 296, + 499 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 450, + 296, + 499 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{align}}" + }, + { + "bbox": [ + 55, + 450, + 296, + 499 + ], + "type": "text", + "content": " are hyperparameters to balance the losses." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 508, + 137, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 508, + 137, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 508, + 137, + 521 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 527, + 167, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 527, + 167, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 527, + 167, + 540 + ], + "type": "text", + "content": "4.1. Experiments Setup" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": "Dataset. We collect image-caption pairs from MSCOCO [4] and then compile a large corpus from ConceptNet [42] and MSCOCO, which also includes concepts generated by GPT-3.5. After filtering out concepts that contain fewer than 15 words, the dataset includes 1,738,985 concepts and 566,747 image-concept pairs, used to pre-train our concept translator for translating both image and text embeddings. We also perform comprehensive experiments on 11 classification datasets, covering a broad spectrum of domains, including: ImageNet [9], CIFAR-10 and CIFAR-100 [18], Food-101 [2], FGVC-Aircraft [22], Flower-102 [25], CUB200-2011 [48], UCF-101 [41], DTD [6], HAM10000 [47] and RESISC45 [5]. Each dataset is split into training, validation, and test sets, as per standard practices." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 555, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 95 + ], + "type": "text", + "content": "Baselines. To evaluate our model, we compare it with a linear probe and several interpretable methods." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 96, + 555, + 312 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 314, + 96, + 555, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 96, + 555, + 144 + ], + "spans": [ + { + "bbox": [ + 314, + 96, + 555, + 144 + ], + "type": "text", + "content": "- Linear Probe. Following the implementation of CLIP [32], we train a logistic regression model using cuML's L-BFGS solver with an L2 penalty term, directly from the visual representations encoded by CLIP." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 144, + 554, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 554, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 554, + 204 + ], + "type": "text", + "content": "- PCBM [54]. PCBM conceptualizes each class as a node in ConceptNet [42] and aggregates neighboring nodes as concepts. PCBM-h [54] introduces an additional classifier that maps image embeddings into the label space, serving as residual shortcuts to the original classifier." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 204, + 555, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 204, + 555, + 263 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 555, + 263 + ], + "type": "text", + "content": "- Labo [52]. Labo employs a submodular function to select concepts for each class from candidate concepts generated by LLM. Unlike other models, it uses dot-product, without any normalization, to determine the presence of particular concepts in an image." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 263, + 554, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 263, + 554, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 263, + 554, + 312 + ], + "type": "text", + "content": "- ResCBM[39]. Building on PCBM-h, ResCBM utilizes ConceptNet [42] to construct a candidate concept bank. It optimizes a set of vectors to align with some concepts from this bank, thereby enhancing the interpretability." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 314, + 554, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 314, + 554, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 314, + 554, + 410 + ], + "type": "text", + "content": "Implementation Details. The hybrid concept bank is constructed with a default ratio of 0.5 between the dynamic and static concept bank, ensuring an equal number of concepts for each class. Our model is trained using the Adam [16] optimizer within the PyTorch Lightning framework. We save checkpoints that achieve the highest validation accuracy and perform evaluations on the test set. Further details can be found in the supplementary material." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 417, + 463, + 429 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 417, + 463, + 429 + ], + "spans": [ + { + "bbox": [ + 313, + 417, + 463, + 429 + ], + "type": "text", + "content": "4.2. Classification Performance" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 435, + 555, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 435, + 555, + 615 + ], + "spans": [ + { + "bbox": [ + 313, + 435, + 555, + 615 + ], + "type": "text", + "content": "Few-shot Comparison. To assess the performance gap between HybridCBM and 'black-box' models, we compare our method with the linear probe (LP) in a few-shot setting. We follow the few-shot evaluation protocol introduced by CLIP [32]. Specifically, we randomly sample 1, 2, 4, 8, and 16 images per class from the training set and use all available images in the fully-supervised setting. For all experiments, we use CLIP-ViT-L/14 as the backbone and compare the performance with submodular [52] and random static concept selection methods. As shown in Table 1, the mean test accuracy across all datasets indicates that HybridCBM performs similarly to LP in the fully-supervised setting and significantly surpasses LP in the few-shot scenario. Figure 3 presents detailed comparisons across eight datasets, with the last three provided in the appendix." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": "CBMs Comparison. We further compare HybridCBM with other CBMs and employ CLIP-RN50 as the backbone to ensure fairness. Additionally, since our dynamic concepts are derived from visual representation, we create a variant called CaptionCBM, where " + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "inline_equation", + "content": "N_{D} / K" + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": " images for each class are sampled and translated into captions to replace the dynamic concepts. As summarized in Table 2, despite strong baselines like LaBo, and ResCBM, Hybrid-" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20183" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 70, + 190, + 159 + ], + "blocks": [ + { + "bbox": [ + 69, + 70, + 190, + 159 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 190, + 159 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 190, + 159 + ], + "type": "image", + "image_path": "2b961cb45c7c0cc6dfe309408f3510cd68c230e42949ed4fe74b5ec9c066b917.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 192, + 70, + 304, + 159 + ], + "blocks": [ + { + "bbox": [ + 192, + 70, + 304, + 159 + ], + "lines": [ + { + "bbox": [ + 192, + 70, + 304, + 159 + ], + "spans": [ + { + "bbox": [ + 192, + 70, + 304, + 159 + ], + "type": "image", + "image_path": "c2c6bb52da85a431e7fe3f67d30187d34b5d5ce25fd8614c68ad88b4fafcca68.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 304, + 70, + 421, + 159 + ], + "blocks": [ + { + "bbox": [ + 304, + 70, + 421, + 159 + ], + "lines": [ + { + "bbox": [ + 304, + 70, + 421, + 159 + ], + "spans": [ + { + "bbox": [ + 304, + 70, + 421, + 159 + ], + "type": "image", + "image_path": "ac88102dacec52339c538b12d5b9f7db9a52e6b45550bc66c52a1182351a5f21.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 425, + 71, + 536, + 159 + ], + "blocks": [ + { + "bbox": [ + 425, + 71, + 536, + 159 + ], + "lines": [ + { + "bbox": [ + 425, + 71, + 536, + 159 + ], + "spans": [ + { + "bbox": [ + 425, + 71, + 536, + 159 + ], + "type": "image", + "image_path": "fdefcbf83f30ab8b3c95a247b27f4029bc5de62a6f01a39cafea0ff01dc254ff.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 69, + 160, + 190, + 258 + ], + "blocks": [ + { + "bbox": [ + 69, + 160, + 190, + 258 + ], + "lines": [ + { + "bbox": [ + 69, + 160, + 190, + 258 + ], + "spans": [ + { + "bbox": [ + 69, + 160, + 190, + 258 + ], + "type": "image", + "image_path": "9606fc4fbc3e3dd0a3e03accc5cc302e432097630772505e412b70b34bbb2c48.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 262, + 555, + 283 + ], + "lines": [ + { + "bbox": [ + 55, + 262, + 555, + 283 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 555, + 283 + ], + "type": "text", + "content": "Figure 3. Comparison of test accuracy between HybridCBM with submodular and random selection methods, and Linear Probe across 8 datasets. The x-axis denotes the number of labeled images." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 192, + 160, + 306, + 258 + ], + "blocks": [ + { + "bbox": [ + 192, + 160, + 306, + 258 + ], + "lines": [ + { + "bbox": [ + 192, + 160, + 306, + 258 + ], + "spans": [ + { + "bbox": [ + 192, + 160, + 306, + 258 + ], + "type": "image", + "image_path": "68649b7494c5ab1d544044bcc7bd73f3ae08f9c7033eccaa81fa7c4d678edeb6.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 309, + 160, + 421, + 258 + ], + "blocks": [ + { + "bbox": [ + 309, + 160, + 421, + 258 + ], + "lines": [ + { + "bbox": [ + 309, + 160, + 421, + 258 + ], + "spans": [ + { + "bbox": [ + 309, + 160, + 421, + 258 + ], + "type": "image", + "image_path": "33f30f0554c71afaa22cb318a8feb446b4f955eea1f476d1a7d15405fb5d54d2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 425, + 160, + 536, + 258 + ], + "blocks": [ + { + "bbox": [ + 425, + 160, + 536, + 258 + ], + "lines": [ + { + "bbox": [ + 425, + 160, + 536, + 258 + ], + "spans": [ + { + "bbox": [ + 425, + 160, + 536, + 258 + ], + "type": "image", + "image_path": "a63d6407ddb1a80a8960a1a0ef3aaeed2a76f6aa2700bbb29de5a9c5f14b97ca.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 58, + 285, + 555, + 365 + ], + "blocks": [ + { + "bbox": [ + 58, + 285, + 555, + 365 + ], + "lines": [ + { + "bbox": [ + 58, + 285, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 58, + 285, + 555, + 365 + ], + "type": "table", + "html": "
Method12AircraftHAM10000RESISC45
\\( concepts per class: ^{50} \\)16Full\\( concepts per class: ^{50} \\)1.0\\( concepts per class: ^{50} \\)
\\( 4 \\)87816
Linear Probe52.30\\( 64.60^6 \\)71.9478.3381.5586.840.72.59
LaboN/A\\( N/A_{0.5} \\)N/AN/AN/A85.720.6N/A
HybridCBM(submodular)58.64\\( 69.52_{0.4} \\)75.0878.7181.9386.830.75.12
HybridCBM (random)57.58\\( 66.69_{3} \\)74.6278.02Line 0r98 HybridCBM(submodular)86.820.44.16\\( Linear Probe \\)\\( Linear Probe \\)\\( Linear Probe \\)
0.3\\( Non distinct \\)\\( Non distinct \\)\\( Non distinct \\)
\\( HybridCBM (random) \\)\\( HybridCBM (random) \\)\\( HybridCBM (random) \\)
", + "image_path": "61081f84d119dc2b013b88fe13c4eac07afecdb6cda2b4f0b9f92e1cf02d6557.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 59, + 377, + 288, + 472 + ], + "blocks": [ + { + "bbox": [ + 64, + 365, + 288, + 374 + ], + "lines": [ + { + "bbox": [ + 64, + 365, + 288, + 374 + ], + "spans": [ + { + "bbox": [ + 64, + 365, + 288, + 374 + ], + "type": "text", + "content": "Table 1. Mean accuracy across all datasets at different shots." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 377, + 288, + 472 + ], + "lines": [ + { + "bbox": [ + 59, + 377, + 288, + 472 + ], + "spans": [ + { + "bbox": [ + 59, + 377, + 288, + 472 + ], + "type": "table", + "html": "
MethodInterpretabilityCIFAR-10CIFAR-100
Linear Probe✗88.870.1
PCBM-h [54]✗87.669.9
PCBM [54]✓84.556.0
Label-Free [26]†✓86.365.2
Labo [52]✓87.665.2
ResCBM [39]†✓88.067.9
CaptionCBM✓81.860.7
HybridCBM✓88.0368.38
", + "image_path": "f5e7051f2e969b0f68cd5cb777577cfe62b7f13f7d7c6504230fee816c1630b8.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 479, + 295, + 501 + ], + "lines": [ + { + "bbox": [ + 55, + 479, + 295, + 501 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 295, + 501 + ], + "type": "text", + "content": "Table 2. Test accuracy comparison between HybridCBM and on CIFAR-10 and CIFAR-100. † indicates the reported performance." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 512, + 295, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 512, + 295, + 536 + ], + "spans": [ + { + "bbox": [ + 55, + 512, + 295, + 536 + ], + "type": "text", + "content": "CBM achieves state-of-the-art performance on CIFAR-10 and CIFAR-100, and delivers results comparable to LP." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 544, + 266, + 559 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 266, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 266, + 559 + ], + "type": "text", + "content": "4.3. Interpretability of the dynamic concepts" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 564, + 295, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 295, + 600 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 295, + 600 + ], + "type": "text", + "content": "In addition to evaluating classification performance, we assess the interpretability of the learned concepts in this section using the following metrics:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 601, + 295, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 601, + 295, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 601, + 295, + 673 + ], + "type": "text", + "content": "- Concept Purity: This metric measures how well each concept aligns with its respective class, reflecting the ability of the concepts to capture class-specific features. For each class " + }, + { + "bbox": [ + 55, + 601, + 295, + 673 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 601, + 295, + 673 + ], + "type": "text", + "content": ", we compute the average cosine similarity between the normalized embedding of the concept " + }, + { + "bbox": [ + 55, + 601, + 295, + 673 + ], + "type": "inline_equation", + "content": "\\hat{e}_k" + }, + { + "bbox": [ + 55, + 601, + 295, + 673 + ], + "type": "text", + "content": " and the class name " + }, + { + "bbox": [ + 55, + 601, + 295, + 673 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{F}}_{txt}(k)" + }, + { + "bbox": [ + 55, + 601, + 295, + 673 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 114, + 683, + 295, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 683, + 295, + 716 + ], + "spans": [ + { + "bbox": [ + 114, + 683, + 295, + 716 + ], + "type": "interline_equation", + "content": "P u r i t y = \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\left(\\bar {\\hat {e}} _ {k} \\cdot \\hat {\\mathcal {F}} _ {t x t} (k)\\right) \\tag {6}", + "image_path": "cffd3ae6322f973306b01e1e6f41282f06b2b592c8a64bbe0775d4dbbeca4418.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 332, + 409, + 555, + 444 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 409, + 555, + 444 + ], + "spans": [ + { + "bbox": [ + 332, + 409, + 555, + 444 + ], + "type": "interline_equation", + "content": "S e p a r a t i o n = 1 - \\frac {2}{K (K - 1)} \\sum_ {i = 1} ^ {K} \\sum_ {j = i + 1} ^ {K} (\\bar {\\hat {e}} _ {i} \\cdot \\bar {\\hat {e}} _ {j}) \\quad (7)", + "image_path": "30fbd59fdfb31351bd1660c8ecb2787e4798739d95013bb479902eb1efd4b2e8.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 451, + 555, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 451, + 555, + 487 + ], + "spans": [ + { + "bbox": [ + 321, + 451, + 555, + 487 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 321, + 451, + 555, + 487 + ], + "type": "inline_equation", + "content": "\\bar{\\hat{e}}_i" + }, + { + "bbox": [ + 321, + 451, + 555, + 487 + ], + "type": "text", + "content": " denotes the mean embedding of concepts for class " + }, + { + "bbox": [ + 321, + 451, + 555, + 487 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 321, + 451, + 555, + 487 + ], + "type": "text", + "content": ". Higher separation indicates greater independence between concepts of different classes in semantic space." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 487, + 555, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 487, + 555, + 594 + ], + "spans": [ + { + "bbox": [ + 315, + 487, + 555, + 594 + ], + "type": "text", + "content": "- Semantic Alignment: To ensure semantic alignment between translated concepts and intended dataset classes, we use GPT-3.5 to evaluate each translated concept's accuracy in representing a specific class with a binary \"yes\" or \"no\" response. The validation metric, \"Semantics\", is calculated as the proportion of concepts that GPT-3.5 correctly associates with the intended classes, reflecting the semantic interpretability. The detailed information is provided in Appendix B.3." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 334, + 601, + 554, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 601, + 554, + 634 + ], + "spans": [ + { + "bbox": [ + 334, + 601, + 554, + 634 + ], + "type": "interline_equation", + "content": "S e m a n t i c s = \\frac {1}{N _ {D}} \\sum_ {i = 1} ^ {N _ {D}} \\delta \\left(L L M \\left(c _ {i}, \\text {c l a s s e s}\\right), 1\\right) \\tag {8}", + "image_path": "80915c7700d10d73fe49bd6bfb92e66e9d8e56d7ed493916c2499708e4c1580c.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 641, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 641, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 321, + 641, + 555, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 321, + 641, + 555, + 713 + ], + "type": "inline_equation", + "content": "LLM(c_i, classes)" + }, + { + "bbox": [ + 321, + 641, + 555, + 713 + ], + "type": "text", + "content": " represents the response to whether concept " + }, + { + "bbox": [ + 321, + 641, + 555, + 713 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 321, + 641, + 555, + 713 + ], + "type": "text", + "content": " belongs to one of the classes, and " + }, + { + "bbox": [ + 321, + 641, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 321, + 641, + 555, + 713 + ], + "type": "text", + "content": " is a binary function that returns 1 for \"yes\" and 0 otherwise. To mitigate potential limitations of the LLM, we emphasize comparing the relative size of static and dynamic concept groups over absolute absolute values." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20184" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 70, + 298, + 273 + ], + "blocks": [ + { + "bbox": [ + 69, + 70, + 298, + 273 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 298, + 273 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 298, + 273 + ], + "type": "image", + "image_path": "e621690fc5de3503e3e90060d3600ff1861c3733160d938fdaec639b89bf9795.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 136, + 274, + 235, + 281 + ], + "lines": [ + { + "bbox": [ + 136, + 274, + 235, + 281 + ], + "spans": [ + { + "bbox": [ + 136, + 274, + 235, + 281 + ], + "type": "text", + "content": "(a) Dynamic Concept Correlation" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 310, + 70, + 539, + 273 + ], + "blocks": [ + { + "bbox": [ + 310, + 70, + 539, + 273 + ], + "lines": [ + { + "bbox": [ + 310, + 70, + 539, + 273 + ], + "spans": [ + { + "bbox": [ + 310, + 70, + 539, + 273 + ], + "type": "image", + "image_path": "b3f6ae79e26dd994a3af4e780a38a152a8fb9347cf21025c1ba293aaa8c5677e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 371, + 274, + 462, + 281 + ], + "lines": [ + { + "bbox": [ + 371, + 274, + 462, + 281 + ], + "spans": [ + { + "bbox": [ + 371, + 274, + 462, + 281 + ], + "type": "text", + "content": "(b) Static Concept Correlation" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 288, + 555, + 322 + ], + "lines": [ + { + "bbox": [ + 55, + 288, + 555, + 322 + ], + "spans": [ + { + "bbox": [ + 55, + 288, + 555, + 322 + ], + "type": "text", + "content": "Figure 4. Comparison of dynamic and static concept correlation structures on the CUB dataset. (a) Dynamic concepts: Correlation maps and classifier weights illustrate how dynamic concepts adjust their relationships with classes and among themselves. (b) Static concepts: Selected and random concepts show their contributions to class relationships and inter-concept dependencies." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 59, + 331, + 288, + 426 + ], + "blocks": [ + { + "bbox": [ + 59, + 331, + 288, + 426 + ], + "lines": [ + { + "bbox": [ + 59, + 331, + 288, + 426 + ], + "spans": [ + { + "bbox": [ + 59, + 331, + 288, + 426 + ], + "type": "table", + "html": "
LossConcept BankPurity (%)Separation (%)Semantics (%)Precision@t (%)
LclsStatic (submodular)19.386.238.348.5
Static (random)19.282.430.947.3
Dynamic0.1100.00.010.05
Lcls+disDynamic71.976.930.034.8
Lcls+dis +ortDynamic72.077.128.934.1
Lcls+dis +ort+alignDynamic39.880.032.846.2
", + "image_path": "393d7f9a2c9afec66250adfc29f2b982f32251106ad27dea16061d945e2d31ee.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 430, + 295, + 452 + ], + "lines": [ + { + "bbox": [ + 55, + 430, + 295, + 452 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 295, + 452 + ], + "type": "text", + "content": "Table 3. Evaluation of HybridCBM's static and dynamic concepts across various metrics." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 464, + 295, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 464, + 295, + 546 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 295, + 546 + ], + "type": "text", + "content": "- Concept-Image Relevance: As shown on the right in Figure 5, to evaluate the accuracy of the concepts aligning with the actual content of each image, we use the ability of GPT-4o in image understanding to calculate the retrieval Precision@t. For a sampled image of each class, we measure the average proportion of " + }, + { + "bbox": [ + 55, + 464, + 295, + 546 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 464, + 295, + 546 + ], + "type": "text", + "content": " concepts that GPT-4o confirms as relevant." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 77, + 552, + 294, + 585 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 552, + 294, + 585 + ], + "spans": [ + { + "bbox": [ + 77, + 552, + 294, + 585 + ], + "type": "interline_equation", + "content": "P r e c i s i o n @ t = \\frac {1}{K} \\sum_ {i = 1} ^ {K} \\frac {\\delta (V L M \\left(c _ {i , t} , i m g\\right) , 1)}{t} \\tag {9}", + "image_path": "5b98ee177f5f2cf0db14442c495ce7b1e187a1ca888f92ad6b59e434d8b83e1e.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 63, + 591, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 591, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 63, + 591, + 295, + 616 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 63, + 591, + 295, + 616 + ], + "type": "inline_equation", + "content": "VLM(c_{i,t},img)" + }, + { + "bbox": [ + 63, + 591, + 295, + 616 + ], + "type": "text", + "content": " is the response of GPT-4o and " + }, + { + "bbox": [ + 63, + 591, + 295, + 616 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 63, + 591, + 295, + 616 + ], + "type": "text", + "content": " is the total number of classes. " + }, + { + "bbox": [ + 63, + 591, + 295, + 616 + ], + "type": "inline_equation", + "content": "c_{i,t}" + }, + { + "bbox": [ + 63, + 591, + 295, + 616 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 63, + 591, + 295, + 616 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 63, + 591, + 295, + 616 + ], + "type": "text", + "content": " concepts of class " + }, + { + "bbox": [ + 63, + 591, + 295, + 616 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 63, + 591, + 295, + 616 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "content": "Concept Analysis. We perform concept analysis on the CUB dataset. Figure 4 compares the correlation structures of dynamic and static concepts. In Figure 4a, the ConceptClass Correlation Map shows that our dynamic concepts align closely with their corresponding classes while effectively capturing diversity across classes. The ConceptConcept Correlation Map further highlights the diversity among concepts. Notably, regions with higher coefficients" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 334, + 555, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 334, + 555, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 334, + 555, + 586 + ], + "type": "text", + "content": "reflect natural semantic relationships between certain class names, such as \"red-eyed vireo\" and \"white-eyed vireo\". The pattern also seen in the Class-Class Correlation Map. The weight matrix of the classifier in Figure 4a reveals that the dynamic concept part has a more pronounced diagonal compared to the static concept, indicating that dynamic concepts are more discriminative and align better with their corresponding classes. In Figure 4b, we present similar relation maps of concepts selected by a submodular function and randomly selected concepts from each class. The selected concepts are more cohesive with their classes, whereas the randomly selected concepts, despite belonging to each class, do not exhibit clear patterns. Table 3 details the evaluation metrics proposed. For static concepts, submodular selection achieves a higher semantics score than random selection (38.3% vs 30.9%). For dynamic concepts, after introducing the loss functions " + }, + { + "bbox": [ + 313, + 334, + 555, + 586 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{dis}}" + }, + { + "bbox": [ + 313, + 334, + 555, + 586 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 334, + 555, + 586 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{ort}}" + }, + { + "bbox": [ + 313, + 334, + 555, + 586 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 334, + 555, + 586 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{align}}" + }, + { + "bbox": [ + 313, + 334, + 555, + 586 + ], + "type": "text", + "content": ", the semantic score of dynamic concepts improved to 32.8%, approaching static concepts. This indicates that our proposed loss functions effectively enhance the alignment and discriminative power of dynamic concepts." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "Interpretability. Interpretability can be divided into pattern and case interpretability. Pattern interpretability focuses on general patterns the model has learned, providing insight into its global behavior and how features influence classes. For example, a classifier's weight matrix reflecting bias toward each class. Case interpretability explains the model's decisions for individual instances, offering a local understanding of why a specific prediction was made for an image—for instance, concept scores indicating how strongly certain concepts are present. Following MMCBM [51], We" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20185" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 71, + 343, + 220 + ], + "blocks": [ + { + "bbox": [ + 62, + 71, + 343, + 220 + ], + "lines": [ + { + "bbox": [ + 62, + 71, + 343, + 220 + ], + "spans": [ + { + "bbox": [ + 62, + 71, + 343, + 220 + ], + "type": "image", + "image_path": "ccb72a03ec3d450c1f70274923e0d913a747efc21559b128f3b78132f9849f36.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 72, + 431, + 85 + ], + "lines": [ + { + "bbox": [ + 345, + 72, + 431, + 85 + ], + "spans": [ + { + "bbox": [ + 345, + 72, + 431, + 85 + ], + "type": "text", + "content": "Concept Retrieval" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 355, + 93, + 511, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 93, + 511, + 139 + ], + "spans": [ + { + "bbox": [ + 355, + 93, + 511, + 139 + ], + "type": "text", + "content": "You are an expert binary concept classifier, capable of determining whether a given concept has any form of relationship with the provided image. If it has any relationship with the image, respond with \"yes\". Otherwise, respond with \"no\"." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 353, + 141, + 395, + 180 + ], + "blocks": [ + { + "bbox": [ + 353, + 141, + 395, + 180 + ], + "lines": [ + { + "bbox": [ + 353, + 141, + 395, + 180 + ], + "spans": [ + { + "bbox": [ + 353, + 141, + 395, + 180 + ], + "type": "image", + "image_path": "2da2dbb27c6ff480001b88d582b14e5fe3e7d178bb1a1a39424320072032ce6d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 395, + 148, + 516, + 175 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 395, + 148, + 516, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 148, + 516, + 156 + ], + "spans": [ + { + "bbox": [ + 395, + 148, + 516, + 156 + ], + "type": "text", + "content": "- sometimes called the \"black albatross\"" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 395, + 157, + 511, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 157, + 511, + 175 + ], + "spans": [ + { + "bbox": [ + 395, + 157, + 511, + 175 + ], + "type": "text", + "content": "- known as the black-browed albatross\n- most abundant albatross species" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 355, + 189, + 376, + 209 + ], + "blocks": [ + { + "bbox": [ + 355, + 189, + 376, + 209 + ], + "lines": [ + { + "bbox": [ + 355, + 189, + 376, + 209 + ], + "spans": [ + { + "bbox": [ + 355, + 189, + 376, + 209 + ], + "type": "image", + "image_path": "3caa0a1b10c2e8d0a607d6952259239cda7a8251ce05c24574bc4620f8285371.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 227, + 555, + 251 + ], + "lines": [ + { + "bbox": [ + 55, + 227, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 227, + 555, + 251 + ], + "type": "text", + "content": "Figure 5. On the left, the top-2 static and dynamic concepts for randomly selected classes across three datasets are presented, focusing on case interpretability. The concept retrieval process with vision-language model GPT-4o is illustrated on the right." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 420, + 201, + 466, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 420, + 201, + 466, + 211 + ], + "spans": [ + { + "bbox": [ + 420, + 201, + 466, + 211 + ], + "type": "text", + "content": "(no, yes, yes)" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 59, + 251, + 288, + 350 + ], + "blocks": [ + { + "bbox": [ + 59, + 251, + 288, + 350 + ], + "lines": [ + { + "bbox": [ + 59, + 251, + 288, + 350 + ], + "spans": [ + { + "bbox": [ + 59, + 251, + 288, + 350 + ], + "type": "table", + "html": "
Methodconcepts per classRatio of Dynamic Concepts
00.20.40.60.81
LPN/A87.09N/AN/AN/AN/AN/A
Hybrid (submodular)583.2184.4685.4586.0886.3486.78
1084.2785.4786.2886.5186.8787.13
1584.7585.8486.4586.7887.0887.25
2084.9386.0986.4486.8486.8887.24
Hybrid (random)580.9584.0485.2885.9686.0786.53
1082.8885.2886.3086.6286.8287.13
1583.6585.8186.5486.8086.8987.25
2083.8586.0986.6686.8687.0687.24
", + "image_path": "47ddbc17d283d48679d133cd40fa5ea1b85e11758f20c8c69bc05b28626558c3.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 61, + 380, + 288, + 433 + ], + "blocks": [ + { + "bbox": [ + 55, + 355, + 295, + 376 + ], + "lines": [ + { + "bbox": [ + 55, + 355, + 295, + 376 + ], + "spans": [ + { + "bbox": [ + 55, + 355, + 295, + 376 + ], + "type": "text", + "content": "Table 4. Ablation average results on 10 datasets for varying the ratio of dynamic concepts under different methods." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 380, + 288, + 433 + ], + "lines": [ + { + "bbox": [ + 61, + 380, + 288, + 433 + ], + "spans": [ + { + "bbox": [ + 61, + 380, + 288, + 433 + ], + "type": "table", + "html": "
MethodMetric\\( \\mathcal{L}_{cls} \\)\\( \\mathcal{L}_{cls+dis} \\)\\( \\mathcal{L}_{cls+dis}+ort \\)\\( \\mathcal{L}_{cls+dis}+ort+align \\)
Hybrid (submodular)accuracy84.6084.3684.2984.54
precision@t0.04734.834.146.2
Hybrid (random)accuracy84.5084.1683.9184.17
precision@t0.04434.835.547.2
", + "image_path": "69f558f3b960723c5062798f19f1ae6940a32f103e8bd98cbdf3b3e740c260ca.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 440, + 295, + 462 + ], + "lines": [ + { + "bbox": [ + 55, + 440, + 295, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 440, + 295, + 462 + ], + "type": "text", + "content": "Table 5. Ablation results for varying training loss with a dynamic concept ratio of 0.5 and 10 concepts per class." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 472, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 295, + 652 + ], + "type": "text", + "content": "use the attention matrix, computed as " + }, + { + "bbox": [ + 55, + 472, + 295, + 652 + ], + "type": "inline_equation", + "content": "S_{i} \\times W" + }, + { + "bbox": [ + 55, + 472, + 295, + 652 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 472, + 295, + 652 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 55, + 472, + 295, + 652 + ], + "type": "text", + "content": " is the classifier's weight, to show that the prediction is based on specific concepts within the image. As illustrated in Figure 5, we show the case interpretability by selecting the top 2 concepts for each image. We evaluate case interpretability by computing retrieval precision@t, measuring how accurately the top 5 identified concepts align with the image content. The results are shown in Table 3, where the static concepts serve as a baseline for comparison with the dynamic concepts; we focus on relative differences due to the potential limitations of GPT-4o. The dynamic concepts, across all loss functions, achieve " + }, + { + "bbox": [ + 55, + 472, + 295, + 652 + ], + "type": "inline_equation", + "content": "46.2\\%" + }, + { + "bbox": [ + 55, + 472, + 295, + 652 + ], + "type": "text", + "content": ", closely approaching the static concepts' performance of " + }, + { + "bbox": [ + 55, + 472, + 295, + 652 + ], + "type": "inline_equation", + "content": "47.3\\%" + }, + { + "bbox": [ + 55, + 472, + 295, + 652 + ], + "type": "text", + "content": ". This demonstrates that dynamic concepts offer high interpretability and can effectively serve as an alternative to static concepts." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 659, + 149, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 149, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 149, + 672 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Dynamic Concept Bank Ratio. To examine the impact of dynamic concept ratios on performance, we test various ratios on 10 datasets except ImageNet, as shown in Table 4." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 262, + 555, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 262, + 555, + 332 + ], + "spans": [ + { + "bbox": [ + 313, + 262, + 555, + 332 + ], + "type": "text", + "content": "As the dynamic concept ratio increases, performance will also improve. Notably, using random concepts per class results in an performance drop compared to the submodular selection method. Furthermore, employing learned concepts, even at a ratio of 0.2, significantly improves performance." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 335, + 556, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 556, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 556, + 455 + ], + "type": "text", + "content": "Concept-Interpretability Loss. Table 5 reports the accuracy and precision@t (t=5) of learned concepts across different combinations of loss functions, comparing the Hybrid (submodular) and Hybrid (random) methods. A slight decrease in accuracy is observed as additional terms are added to the " + }, + { + "bbox": [ + 313, + 335, + 556, + 455 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cls}" + }, + { + "bbox": [ + 313, + 335, + 556, + 455 + ], + "type": "text", + "content": " but significantly improves precision@t. This suggests a trade-off, where additional constraints, while slightly reducing accuracy (e.g., accuracy drops from 84.60 to 84.29 for the Hybrid (submodular) method), enhance concept interpretability." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 466, + 388, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 466, + 388, + 478 + ], + "spans": [ + { + "bbox": [ + 313, + 466, + 388, + 478 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we introduced HybridCBM, a novel hybrid concept bottleneck model designed to enhance interpretability and performance by combining both static and dynamic concept banks. Unlike traditional CBMs, our method allows the model to dynamically discover new concepts during training, thus expanding its ability to capture unique and relevant features. Our approach also incorporates a concept translator to provide human-readable interpretations for the learned dynamic concepts, making the model's decision process more accessible and understandable. Through extensive experiments on various datasets, our method achieves comparable accuracy to 'black-box' models while retaining transparency, bridging the gap between interpretability and performance. We believe that HybridCBM represents a significant step forward in the development of interpretable machine learning models. Future work could further explore methods for optimizing the accuracy of concept labeling and exploring more sophisticated strategies for dynamic concept discovery." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20186" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 173, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 173, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 173, + 85 + ], + "type": "text", + "content": "6. Acknowledgements." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 297, + 147 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 297, + 147 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 297, + 147 + ], + "type": "text", + "content": "This project is supported by NSFC Key Program 62236009, Shenzhen Fundamental Research Program (General Program) JCYJ 20210324140807019, NSFC General Program 61876032, and Key Laboratory of Data Intelligence and Cognitive Computing, Longhua District, Shenzhen." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 166, + 115, + 178 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 166, + 115, + 178 + ], + "spans": [ + { + "bbox": [ + 56, + 166, + 115, + 178 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 186, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 186, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 186, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 61, + 186, + 294, + 251 + ], + "type": "text", + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 251, + 294, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 251, + 294, + 295 + ], + "spans": [ + { + "bbox": [ + 61, + 251, + 294, + 295 + ], + "type": "text", + "content": "[2] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 – mining discriminative components with random forests. In European Conference on Computer Vision, 2014. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 296, + 294, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 296, + 294, + 350 + ], + "spans": [ + { + "bbox": [ + 62, + 296, + 294, + 350 + ], + "type": "text", + "content": "[3] Chaofan Chen, Oscar Li, Daniel Tao, Alina Barnett, Cynthia Rudin, and Jonathan K Su. This looks like that: Deep learning for interpretable image recognition. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 351, + 294, + 394 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 351, + 294, + 394 + ], + "spans": [ + { + "bbox": [ + 62, + 351, + 294, + 394 + ], + "type": "text", + "content": "[4] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dollar, and C. Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server, 2015. 5, 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 395, + 294, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 395, + 294, + 427 + ], + "spans": [ + { + "bbox": [ + 62, + 395, + 294, + 427 + ], + "type": "text", + "content": "[5] Gong Cheng, Junwei Han, and Xiaoqiang Lu. Remote sensing image scene classification: Benchmark and state of the art. Proceedings of the IEEE, 105(10):1865-1883, 2017. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 428, + 294, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 428, + 294, + 472 + ], + "spans": [ + { + "bbox": [ + 62, + 428, + 294, + 472 + ], + "type": "text", + "content": "[6] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3606-3613, 2014. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 472, + 294, + 505 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 472, + 294, + 505 + ], + "spans": [ + { + "bbox": [ + 62, + 472, + 294, + 505 + ], + "type": "text", + "content": "[7] Cristina Conati, Kaska Porayska-Pomsta, and Manolis Mavrikis. Ai in education needs interpretable machine learning: Lessons from open learner modelling, 2018. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 505, + 294, + 537 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 505, + 294, + 537 + ], + "spans": [ + { + "bbox": [ + 62, + 505, + 294, + 537 + ], + "type": "text", + "content": "[8] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2013. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 538, + 294, + 582 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 538, + 294, + 582 + ], + "spans": [ + { + "bbox": [ + 62, + 538, + 294, + 582 + ], + "type": "text", + "content": "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 582, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 582, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 582, + 294, + 635 + ], + "type": "text", + "content": "[10] Andre Esteva, Alexandre Robicquet, Bharath Ramsundar, Volodymyr Kuleshov, Mark DePristo, Katherine Chou, Claire Cui, Greg Corrado, Sebastian Thrun, and Jeff Dean. A guide to deep learning in healthcare. Nature medicine, 25 (1):24-29, 2019. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 637, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 637, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 637, + 294, + 667 + ], + "type": "text", + "content": "[11] David Gunning and David Aha. Darpa's explainable artificial intelligence (xai) program. AI magazine, 40(2):44-58, 2019. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 669, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 669, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 669, + 294, + 712 + ], + "type": "text", + "content": "[12] Zhengbao Jiang, Frank F. Xu, Jun Araki, and Graham Neubig. How can we know what language models know? Transactions of the Association for Computational Linguistics, 8: 423-438, 2020. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 106 + ], + "type": "text", + "content": "[13] Dmitry Kazhdan, Botty Dimanov, Mateja Jamnik, Pietro Lio, and Adrian Weller. Now you see me (cme): Concept-based model extraction, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 107, + 554, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 554, + 173 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 554, + 173 + ], + "type": "text", + "content": "[14] Been Kim, Martin Wattenberg, Justin Gilmer, Carrie Cai, James Wexler, Fernanda Viegas, and Rory sayres. Interpretability beyond feature attribution: Quantitative testing with concept activation vectors (TCAV). In Proceedings of the 35th International Conference on Machine Learning, pages 2668-2677. PMLR, 2018. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 175, + 554, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 554, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 554, + 207 + ], + "type": "text", + "content": "[15] Injae Kim, Jongha Kim, Joonmyung Choi, and Hyunwoo J. Kim. Concept bottleneck with visual concept filtering for explainable medical image classification, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 208, + 554, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 554, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 554, + 239 + ], + "type": "text", + "content": "[16] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 5, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 241, + 554, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 554, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 554, + 285 + ], + "type": "text", + "content": "[17] Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. Concept bottleneck models. In International Conference on Machine Learning, pages 5338-5348. PMLR, 2020. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 286, + 554, + 309 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 554, + 309 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 554, + 309 + ], + "type": "text", + "content": "[18] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 310, + 554, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 310, + 554, + 354 + ], + "spans": [ + { + "bbox": [ + 317, + 310, + 554, + 354 + ], + "type": "text", + "content": "[19] Wei Li, Linchao Zhu, Longyin Wen, and Yi Yang. Decap: Decoding clip latents for zero-shot captioning via text-only training. In The Eleventh International Conference on Learning Representations, 2023. 3, 4, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 355, + 554, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 355, + 554, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 355, + 554, + 398 + ], + "type": "text", + "content": "[20] Yuanning Li, Huzheng Yang, and Shi Gu. Enhancing neural encoding models for naturalistic perception with a multilevel integration of deep neural networks and cortical networks. Science Bulletin, 2024. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 399, + 554, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 554, + 431 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 554, + 431 + ], + "type": "text", + "content": "[21] Max Losch, Mario Fritz, and Bernt Schiele. Interpretability beyond classification output: Semantic bottleneck networks, 2019. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 434, + 554, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 554, + 476 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 554, + 476 + ], + "type": "text", + "content": "[22] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 478, + 554, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 554, + 510 + ], + "type": "text", + "content": "[23] Tim Miller. Explanation in artificial intelligence: Insights from the social sciences. Artificial Intelligence, 267:1-38, 2019. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 512, + 554, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 512, + 554, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 512, + 554, + 544 + ], + "type": "text", + "content": "[24] Ron Mokady, Amir Hertz, and Amit H Bermano. Clipcap: Clip prefix for image captioning. arXiv preprint arXiv:2111.09734, 2021. 3, 4, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 546, + 554, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 546, + 554, + 590 + ], + "spans": [ + { + "bbox": [ + 316, + 546, + 554, + 590 + ], + "type": "text", + "content": "[25] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In Indian Conference on Computer Vision, Graphics and Image Processing, 2008. 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 591, + 554, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 591, + 554, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 591, + 554, + 635 + ], + "type": "text", + "content": "[26] Tuomas Oikarinen, Subhro Das, Lam M. Nguyen, and Tsui-Wei Weng. Label-free concept bottleneck models. In The Eleventh International Conference on Learning Representations, 2023. 2, 4, 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 636, + 554, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 554, + 712 + ], + "type": "text", + "content": "[27] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul F Christiano, Jan Leike, and Ryan Lowe. Training language models to follow instructions with human feedback. In Advances in Neural Information" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20187" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "text", + "content": "Processing Systems, pages 27730-27744. Curran Associates, Inc., 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 96, + 295, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 96, + 295, + 128 + ], + "spans": [ + { + "bbox": [ + 57, + 96, + 295, + 128 + ], + "type": "text", + "content": "[28] Ahmet Murat Ozbayoglu, Mehmet Ugur Gudelek, and Omer Berat Sezer. Deep learning for financial applications: A survey. Applied Soft Computing, 93:106384, 2020. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 129, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 172 + ], + "type": "text", + "content": "[29] Konstantinos Panagiotis Panousis, Dino Ienc, and Diego Marcos. Sparse linear concept discovery models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2767-2771, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 173, + 295, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 173, + 295, + 260 + ], + "spans": [ + { + "bbox": [ + 56, + 173, + 295, + 260 + ], + "type": "text", + "content": "[30] Fabio Petroni, Tim Roktaschel, Sebastian Riedel, Patrick Lewis, Anton Bakhtin, Yuxiang Wu, and Alexander Miller. Language models as knowledge bases? In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 2463-2473, Hong Kong, China, 2019. Association for Computational Linguistics. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 262, + 294, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 262, + 294, + 302 + ], + "spans": [ + { + "bbox": [ + 57, + 262, + 294, + 302 + ], + "type": "text", + "content": "[31] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 2, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 305, + 294, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 305, + 294, + 370 + ], + "spans": [ + { + "bbox": [ + 57, + 305, + 294, + 370 + ], + "type": "text", + "content": "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2, 3, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 372, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 372, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 57, + 372, + 294, + 415 + ], + "type": "text", + "content": "[33] Muhammad Imran Razzak, Saeeda Naz, and Ahmad Zaib. Deep Learning for Medical Image Processing: Overview, Challenges and the Future, page 323-350. Springer International Publishing, 2017. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 416, + 294, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 416, + 294, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 416, + 294, + 470 + ], + "type": "text", + "content": "[34] Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. \"why should i trust you?\" explaining the predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining, pages 1135-1144, 2016. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 472, + 294, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 472, + 294, + 513 + ], + "spans": [ + { + "bbox": [ + 56, + 472, + 294, + 513 + ], + "type": "text", + "content": "[35] Cynthia Rudin. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature Machine Intelligence, 1(5):206-215, 2019. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 515, + 294, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 515, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 57, + 515, + 294, + 536 + ], + "type": "text", + "content": "[36] Gesina Schwalbe. Concept embedding analysis: A review. arXiv preprint arXiv:2203.13909, 2022. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 537, + 294, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 294, + 601 + ], + "type": "text", + "content": "[37] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 604, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 604, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 604, + 294, + 635 + ], + "type": "text", + "content": "[38] Chenming Shang, Hengyuan Zhang, Hao Wen, and Yujiu Yang. Understanding multimodal deep neural networks: A concept selection view, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 637, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 637, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 637, + 294, + 690 + ], + "type": "text", + "content": "[39] Chenming Shang, Shiji Zhou, Hengyuan Zhang, Xinzhe Ni, Yujiu Yang, and Yuwang Wang. Incremental residual concept bottleneck models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11030-11040, 2024. 2, 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "type": "text", + "content": "[40] Sheng Shen, Liunian Harold Li, Hao Tan, Mohit Bansal, Anna Rohrbach, Kai-Wei Chang, Zhewei Yao, and Kurt" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "Keutzer. How much can CLIP benefit vision-and-language tasks? In International Conference on Learning Representations, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "type": "text", + "content": "[41] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "type": "text", + "content": "[42] Robyn Speer, Joshua Chin, and Catherine Havasi. Conceptnet 5.5: An open multilingual graph of general knowledge. In Thirty-first AAAI conference on artificial intelligence, 2017. 2, 5, 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 186, + 553, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 219 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 219 + ], + "type": "text", + "content": "[43] Divyansh Srivastava, Ge Yan, and Tsui-Wei Weng. Vlg-cbm: Training concept bottleneck models with vision-language guidance, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 220, + 553, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 263 + ], + "type": "text", + "content": "[44] Sandeep Subramanian, Adam Trischler, Yoshua Bengio, and Christopher J Pal. Learning general purpose distributed sentence representations via large scale multi-task learning. arXiv preprint arXiv:1804.00079, 2018. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 266, + 553, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 266, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 266, + 553, + 331 + ], + "type": "text", + "content": "[45] Mujeen Sung, Jinhyuk Lee, Sean Yi, Minji Jeon, Sungdong Kim, and Jaewoo Kang. Can language models be biomedical knowledge bases? In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 4723-4734, Online and Punta Cana, Dominican Republic, 2021. Association for Computational Linguistics. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 333, + 553, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 333, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 316, + 333, + 553, + 365 + ], + "type": "text", + "content": "[46] Yoad Tewel, Yoav Shalev, Idan Schwartz, and Lior Wolf. Zero-shot image-to-text generation for visual-semantic arithmetic. arXiv preprint arXiv:2111.14447, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 366, + 553, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 366, + 553, + 410 + ], + "spans": [ + { + "bbox": [ + 316, + 366, + 553, + 410 + ], + "type": "text", + "content": "[47] Philipp Tschandl, Cliff Rosendahl, and Harald Kittler. The ham10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions. Scientific data, 5(1):1-9, 2018. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 412, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 412, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 412, + 553, + 443 + ], + "type": "text", + "content": "[48] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 445, + 553, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 445, + 553, + 488 + ], + "spans": [ + { + "bbox": [ + 316, + 445, + 553, + 488 + ], + "type": "text", + "content": "[49] Bowen Wang, Liangzhi Li, Yuta Nakashima, and Hajime Nagahara. Learning bottleneck concepts in image classification. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 491, + 553, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 491, + 553, + 534 + ], + "spans": [ + { + "bbox": [ + 316, + 491, + 553, + 534 + ], + "type": "text", + "content": "[50] Zirui Wang, Jiahui Yu, Adams Wei Yu, Zihang Dai, Yulia Tsvetkov, and Yuan Cao. SimVLM: Simple visual language model pretraining with weak supervision. In International Conference on Learning Representations, 2022. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 536, + 553, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 536, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 316, + 536, + 553, + 590 + ], + "type": "text", + "content": "[51] Yifan Wu, Yang Liu, Yue Yang, Michael S. Yao, Wenli Yang, Xuehui Shi, Lihong Yang, Dongjun Li, Yueming Liu, James C. Gee, Xuan Yang, Wenbin Wei, and Shi Gu. A concept-based interpretable model for the diagnosis of choroid neoplasias using multimodal data, 2024. 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 592, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 592, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 592, + 553, + 657 + ], + "type": "text", + "content": "[52] Yue Yang, Artemis Panagopoulou, Shenghao Zhou, Daniel Jin, Chris Callison-Burch, and Mark Yatskar. Language in a bottle: Language model guided concept bottlenecks for interpretable image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19187-19197, 2023. 2, 4, 5, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "text", + "content": "[53] Chih-Kuan Yeh, Been Kim, Sercan Arik, Chun-Liang Li, Tomas Pfister, and Pradeep Ravikumar. On completeness-aware concept-based explanations in deep neural networks. In Advances in Neural Information Processing Systems, pages 20554-20565. Curran Associates, Inc., 2020. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20188" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 73, + 295, + 183 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "type": "text", + "content": "[54] Mert Yuksekgonul, Maggie Wang, and James Zou. Posthoc concept bottleneck models. In ICLR 2022 Workshop on PAIR²Struct, 2022. 2, 4, 5, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 107, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 107, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 55, + 107, + 295, + 140 + ], + "type": "text", + "content": "[55] Yu Zhang, Peter Tiño, Alex Leonardis, and Ke Tang. A survey on neural network interpretability. IEEE Transactions on Emerging Topics in Computational Intelligence, 2021. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 141, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 141, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 57, + 141, + 294, + 183 + ], + "type": "text", + "content": "[56] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 1" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "20189" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Hybrid Global-Local Representation with Augmented Spatial Guidance for Zero-Shot Referring Image Segmentation/9c88ce4f-9647-4320-bdd2-f6fc8a88e98e_content_list.json b/2025/Hybrid Global-Local Representation with Augmented Spatial Guidance for Zero-Shot Referring Image Segmentation/9c88ce4f-9647-4320-bdd2-f6fc8a88e98e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9e526b7ce6773b1a11a84c29959c9549a92537fb --- /dev/null +++ b/2025/Hybrid Global-Local Representation with Augmented Spatial Guidance for Zero-Shot Referring Image Segmentation/9c88ce4f-9647-4320-bdd2-f6fc8a88e98e_content_list.json @@ -0,0 +1,1720 @@ +[ + { + "type": "text", + "text": "Hybrid Global-Local Representation with Augmented Spatial Guidance for Zero-Shot Referring Image Segmentation", + "text_level": 1, + "bbox": [ + 120, + 128, + 879, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ting Liu $^{1,2*}$ Siyuan Li $^{1}$", + "bbox": [ + 395, + 202, + 596, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1ASGO, School of Computer Science, Northwestern Polytechnical University \n2Shenzhen Research Institute of Northwestern Polytechnical University", + "bbox": [ + 191, + 220, + 803, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "liuting@nwpu.edu.cn, lisiyuan@mail.nwpu.edu.cn", + "bbox": [ + 285, + 258, + 696, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 309, + 326, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advances in zero-shot referring image segmentation (RIS), driven by models such as the Segment Anything Model (SAM) and CLIP, have made substantial progress in aligning visual and textual information. Despite these successes, the extraction of precise and high-quality mask region representations remains a critical challenge, limiting the full potential of RIS tasks. In this paper, we introduce a training-free, hybrid global-local feature extraction approach that integrates detailed mask-specific features with contextual information from the surrounding area, enhancing mask region representation. To further strengthen alignment between mask regions and referring expressions, we propose a spatial guidance augmentation strategy that improves spatial coherence, which is essential for accurately localizing described areas. By incorporating multiple spatial cues, this approach facilitates more robust and precise referring segmentation. Extensive experiments on standard RIS benchmarks demonstrate that our method significantly outperforms existing zero-shot RIS models, achieving substantial performance gains. We believe our approach advances RIS tasks and establishes a versatile framework for region-text alignment, offering broader implications for cross-modal understanding and interaction. Code is available at https://github.com/fhgyuanshen/HybridGL.", + "bbox": [ + 89, + 340, + 483, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 731, + 220, + 747 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Referring image segmentation (RIS) is a critical task in computer vision, where the goal is to segment a specific object or region in an image based on a natural language expression. This task is essential for applications such as visual search, robot perception, and human-computer interaction. Recent advancements, particularly models like the Segment Anything Model (SAM) [9] and CLIP [23], have significantly advanced zero-shot RIS, enabling object seg", + "bbox": [ + 89, + 756, + 482, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/663573138a87d23a5298fa7f333a3728dd2f00267c942f211ca9a554f60b2d70.jpg", + "image_caption": [ + "Figure 1. Common issues in existing methods: 1) Inaccurate mask feature extraction; 2) Incorrect spatial localization; 3) Incomplete segmentation." + ], + "image_footnote": [], + "bbox": [ + 517, + 309, + 905, + 390 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "mentation without the need for labeled data or task-specific training. In typical RIS pipelines, SAM generates a set of mask proposals, and CLIP extracts visual features for each mask region. For each mask, a similarity score is computed with the referring text, and the mask with the highest score is selected as the final prediction.", + "bbox": [ + 511, + 457, + 906, + 547 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite the promising performance of existing methods, the accurate extraction of mask representations that align well with referring text remains underexplored. Most existing approaches either mask out areas outside the mask region or crop the image to focus solely on the mask region before feeding it into CLIP, which primarily concentrates on local features while often neglecting the essential surrounding context. Some works [22, 32, 41] have sought to introduce global surrounding context features. However, such methods remain simplistic and fail to fully capture the complex interplay between the mask and its context, limiting their ability to accurately match referring expressions with the correct mask regions, as shown in Fig. 1.", + "bbox": [ + 511, + 550, + 908, + 747 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we propose a novel hybrid global-local feature extraction approach to enhance mask region feature extraction, achieving a more precise and contextually rich mask representation without any additional training. Our approach seamlessly integrates local and global features, capturing both region-specific and context-aware information for each mask. Specifically, we design two complementary branches within CLIP to extract local region-specific visual features and broader context-aware visual features for each mask. Features from the global branch", + "bbox": [ + 511, + 750, + 910, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Ting Liu is the corresponding author.", + "bbox": [ + 109, + 887, + 313, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "29634", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "are progressively fused into the local branch to generate a hybrid feature representation. This hybrid fusion allows the visual encoder to automatically capture and interact with the complementary information from both branches, yielding a more contextually enriched mask feature.", + "bbox": [ + 89, + 90, + 480, + 165 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another challenge in referring semantic segmentation lies in the use of spatial relationships within referring expressions (e.g., \"left of\", \"bottom of\") to describe objects, as shown in Fig. 1. These spatial cues introduce complexity, as they require both the recognition of object locations and the relationship between them. Capturing and aligning these spatial descriptions with visual features is inherently difficult. While referring text offers rich contextual information about spatial relationships, effectively integrating this context with visual data remains a significant challenge. Without a mechanism to explicitly model and align this spatial information, accurately matching the textual description to the correct visual region becomes problematic. Moreover, directly computing the similarity between the extracted visual mask features and text features can lead to ambiguity in the segmentation, as the mask features may capture a mixture of information from multiple objects or regions. It could mistakenly segment only part of the target region, rather than the full object. To alleviate these limitations, we introduce an augmentation approach by introducing several spatial guidance including spatial relationships, coherence, and positional cues.", + "bbox": [ + 91, + 167, + 482, + 496 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our experiments show that the proposed method significantly outperforms several zero-shot baselines and weakly-supervised referring segmentation methods, achieving significant accuracy improvements. The proposed framework offers a powerful, efficient approach to zero-shot referring image segmentation, with strong potential for practical deployment in real-world scenarios. The contributions of this paper can be summarized as follows:", + "bbox": [ + 89, + 500, + 480, + 618 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose an innovative hybrid global-local feature extraction approach for RIS, enhancing mask region representation without additional training requirements.", + "- We introduce a spatial guidance augmentation strategy that leverages spatial relationships, coherence, and positional cues to mitigate segmentation ambiguity.", + "- Extensive experiments on the four public datasets, demonstrate that our method significantly outperforms existing state-of-the-art zero-shot semantic segmentation approaches." + ], + "bbox": [ + 91, + 619, + 482, + 770 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 784, + 232, + 800 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Referring Image Segmentation. Referring image segmentation is a visual grounding task that requires the model to understand a natural language expression describing a specific object within an image and accurately segment that object from the rest of the scene [7]. The goal of this task is to bridge the gap between visual and textual modalities, en", + "bbox": [ + 89, + 810, + 480, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "abling more sophisticated interactions with visual content. Fully supervised methods [3, 8, 12, 13, 16, 27, 28, 38] have achieved impressive performance in this area by effectively integrating information from both images and text descriptions. These methods rely on large datasets with detailed annotations, where each object mentioned in the text is precisely segmented in the corresponding image [33, 34]. Currently, some weakly supervised methods [4, 10, 20, 29] can learn and perform segmentation using a smaller amount of labeled data. However, the requirement for such detailed and expensive human-annotated data limits the scalability and applicability of supervised approaches.", + "bbox": [ + 511, + 90, + 903, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Foundation Model in Image Segmentation. Recent research has shown that segmentation knowledge can emerge from pre-trained foundation models (FMs) [1] such as CLIP and Stable Diffusion [25]. Though the standard CLIP model excels at recognizing object appearances, it falls short in grasping their exact locations. However, MaskCLIP [43] demonstrates that it is possible to adapt CLIP for segmentation purposes by making minor adjustments to its attention-pooling mechanism. Other works delve into various strategies for refining attention mechanisms [2, 6, 14], enhancing the model's localization capability. Specifically, SAM has shown promising capabilities in object segmentation in many works [5, 17, 18, 24, 37, 39, 44]. Overall, these findings suggest that FMs can serve as a valuable resource for zero-shot segmentation and related tasks.", + "bbox": [ + 511, + 272, + 903, + 496 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Zero-shot Referring Image Segmentation. To address the limitations of fully supervised methods, zero-shot referring image segmentation has gained significant attention. Zero-shot approaches aim to perform segmentation without requiring any labeled data for the target objects, making them highly flexible and scalable. One notable method in this domain is Global-Local [41], which leverages pretrained models like FreeSOLO [35] and CLIP to achieve zero-shot segmentation. Another approach, CaR [31], further enhances this process by recurrently applying CLIP to refine the segmentation mask iteratively. The introduction of Segment Anything Model (SAM) has marked a significant milestone in zero-shot referring image segmentation. For example, Ref-Diff [22] utilizes diffusion models to better understand the relationship between image and text pairs, leading to more accurate and context-aware segmentations. TAS [32] uses a captioner, BLIP2 [11], to provide additional context and enhance the segmentation results by generating descriptive captions for the images. PseudorIS [42] modifies the Global-Local's pipeline and incorporates a captioner like CoCa [40] for unsupervised training, leveraging the rich knowledge of current foundation models to improve segmentation accuracy.", + "bbox": [ + 511, + 500, + 903, + 845 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Note that although we also adopt both local and global features, our approach differs significantly from Global-Local [41] in terms of both the methods for local and global", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "29635", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a2e22ab99cccd241570d1b57fdb16c3831b7f84f204a5aeb6f89c95b638e9913.jpg", + "image_caption": [ + "Figure 2. The proposed framework combines hybrid global-local feature extraction with multiple spatial guidance mechanisms to improve zero-shot referring image segmentation, using mask proposals generated by SAM. By leveraging both broad context and local details, and enhancing segmentation with spatial guidance, the framework effectively augments the segmentation of target based on textual descriptions." + ], + "image_footnote": [], + "bbox": [ + 135, + 89, + 867, + 359 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "feature extraction and the fusion strategy of these features. Distinct from previous methods, our work introduces a novel, training-free approach to extracting mask representations using a hybrid feature extraction scheme. Furthermore, we incorporate multiple spatial guidance mechanisms to enhance semantic coherence and spatial alignment, improving the referring segmentation process.", + "bbox": [ + 89, + 411, + 483, + 516 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 532, + 181, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 89, + 559, + 200, + 573 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given an input image $I \\in \\mathbb{R}^{H \\times W \\times 3}$ , where $H$ and $W$ are the height and width of the image, respectively, and a referring text $T$ , which describes the target object or region, the goal is to predict the segmentation mask $m^* \\in \\{0,1\\}^{H \\times W}$ that corresponds to the region in $I$ described by $t$ . We first adopt SAM (Segment Anything Model) [9] to generate a set of mask proposals $M = \\{m_1, m_2, \\ldots, m_n\\}$ , where each $m_i$ is a binary mask highlighting a potential segment region in $I$ .", + "bbox": [ + 89, + 580, + 483, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For each mask $m_{i}\\in M$ , we use the CLIP visual encoder to extract our proposed hybrid global-local features $\\mathbf{x}_i\\in \\mathbb{R}^d$ . To match the referring expression with the visual features, we encode the referring expression $t$ using the CLIP text encoder $\\phi_{\\mathrm{text}}(\\cdot)$ to obtain the textual feature $\\mathbf{f}_t\\in \\mathbb{R}^d$ , adopting same strategy with [41]. Subsequently, we can compute the cosine similarity between $\\mathbf{f}_t$ and each feature $\\mathbf{x}_i$ , obtaining the semantic alignment score $\\mathbf{S}_{m_i}^s$ for each mask $m_{i}$ . Furthermore, we introduce a spatial guidance augmentation approach. By leveraging several spatial cues, we augment the scoring mask process to prioritize semantically and spatially aligned masks. Our method in", + "bbox": [ + 89, + 719, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tegrates several spatial cues, specifically spatial coherence guidance, spatial position guidance, and spatial relationships guidance, to enhance the mask scoring process. The overall framework is shown in Fig. 2.", + "bbox": [ + 511, + 411, + 906, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Hybrid Global-Local Feature Extraction", + "text_level": 1, + "bbox": [ + 511, + 483, + 861, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To extract a more precise mask feature for mask $m_{i}$ , we propose the Hybrid Global-Local Feature Extraction method. This approach captures both region-specific and context-aware features, significantly enhancing mask feature extraction. The method consists of two branches: a local branch for region-specific feature extraction and a global branch for context-aware feature extraction. Features from the global branch are progressively fused into the local branch to generate a hybrid feature representation. This hybrid fusion allows the model to automatically capture the complementary information from the two branches.", + "bbox": [ + 511, + 506, + 906, + 671 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Local Feature Extraction. For the local region-specific feature extraction, we apply the binary mask $m$ to the input image $I$ , creating a masked image $I_{\\mathrm{local}} = I \\cdot m_i$ . This ensures that only the relevant region contributes to feature extraction. The masked image $I_{\\mathrm{local}}$ is then processed through the CLIP image encoder to obtain the region-specific features. Unlike most existing methods, we do not crop the mask region, ensuring that the input image remains well-aligned with the input of the global branch. This allows the naturally aligned features to be hybridized and deeply fused for enhanced feature extraction.", + "bbox": [ + 511, + 672, + 908, + 838 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Global Feature Extraction. For the context-aware global feature extraction, we use the original image $I$ and apply a Gaussian blur to the non-mask regions, thereby emphasizing the relevant areas while preserving essential contextual", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "29636", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "information. To further direct the model's focus onto the mask region, we introduce an attention mask $\\mathbf{A}_{m_i}$ during the self-attention operation in the last $l$ layers of the CLIP transformer-based architecture as shown in Fig. 2. Considering the CLS token is embedded into the shared visual-textual space as the visual feature, we nullify the attention scores between the CLS token and image tokens outside the masked region, effectively preventing non-relevant areas from contributing to the attention on the CLS token.", + "bbox": [ + 89, + 90, + 480, + 224 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Hybrid Feature Extraction. To enhance the fusion of local and global features, we propose a hybrid fusion approach that enables the attention mechanism in the network to effectively and automatically capture the complementary information between these two features.", + "bbox": [ + 89, + 226, + 480, + 301 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given that local features primarily focus on the mask region, which typically contains the most relevant information for the task, the global branch is intended to complement the local branch by providing a broader context. While the global branch captures a wider spatial range and provides essential contextual information beyond the mask, it may also introduce irrelevant information from regions outside the target. To maximize the benefits of both local and global features, it is crucial to selectively integrate them, ensuring that the global context enhances the local representation without diluting its focus on the target region.", + "bbox": [ + 89, + 301, + 480, + 468 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To achieve this, we apply a token mask to the global features during the fusion. Specifically, we mask out image tokens outside the mask region to ensure that only the relevant tokens contribute to the feature extraction process. The fusion of local and global features at layer $l$ is performed as follows:", + "bbox": [ + 89, + 468, + 480, + 556 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {\\text {h y b r i d}} ^ {(l)} = \\mathbf {x} _ {\\text {l o c a l}} ^ {(l - 1)} + \\beta \\cdot \\left(\\mathbf {x} _ {\\text {g l o b a l}} ^ {(l - 1)} \\cdot \\mathbf {B} _ {m _ {i}}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 561, + 480, + 588 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{x}_{\\mathrm{local}}^{(l - 1)}$ and $\\mathbf{x}_{\\mathrm{global}}^{(l - 1)}$ represent the local and global features derived from the $l - 1$ -th layer, respectively. The hyper-parameter $\\beta$ is the relative contribution of the global feature to the fusion process. To ensure that only relevant tokens contribute to the feature fusion, we apply a mask $\\mathbf{B}_{m_i} \\in \\{0,1\\}^K$ on the global feature, where each feature contains $K$ image token features, effectively excluding tokens corresponding to regions outside the mask. The resulting hybrid feature $\\mathbf{x}_{\\mathrm{hybrid}}^{(l - 1)}$ is then passed as input to the $l$ -th layer of the CLIP image encoder, denoted $\\phi_{\\mathrm{visual}}^l$ , to produce a feature representation that integrates both global and local information:", + "bbox": [ + 89, + 595, + 482, + 781 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {\\text {l o c a l}} ^ {(l)} = \\phi_ {\\text {v i s u a l}} ^ {l} \\left(\\mathbf {x} _ {\\text {h y b r i d}} ^ {(l)}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 781, + 480, + 803 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The local feature $\\mathbf{x}_{\\mathrm{local}}^{(L)}$ obtained from the last encoder layer $L$ , which effectively combines localized and global contextual information, serves as the final hybrid feature for the mask region.", + "bbox": [ + 89, + 808, + 480, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Semantic Alignment Score. We then extract hybrid visual features $\\mathbf{X}_{\\mathrm{hybrid}} = \\{\\mathbf{x}_1,\\mathbf{x}_2,\\dots ,\\mathbf{x}_n\\}$ for each mask, where", + "bbox": [ + 89, + 869, + 480, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "each $\\mathbf{x}_i$ corresponds to a mask proposal $m_{i}$ .After extracting a referring text feature $\\mathbf{f}_t$ from the CLIP text encoder, we calculate the cosine semantic alignment score between the referring text and each $\\mathbf{x}_i$ ..", + "bbox": [ + 511, + 90, + 903, + 151 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {S} _ {m _ {i}} ^ {s} = \\cos \\left(\\phi_ {\\text {t e x t}} (t), x _ {i}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 627, + 162, + 903, + 180 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Spatial Guidance Augmentation", + "text_level": 1, + "bbox": [ + 511, + 188, + 795, + 204 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We combine multiple spatial guidance mechanisms, including relationships, coherence, and position, which are detailed in the following.", + "bbox": [ + 511, + 210, + 903, + 255 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Spatial Relationship Guidance. To utilize the spatial relationships described in the referring text, we first parse the text into objects and spatial relations, following a method similar to [30]. For instance, objects might include phrases like \"the pizza\" or \"man\" while spatial relations describe the positioning between objects, such as \"right\".", + "bbox": [ + 511, + 256, + 903, + 345 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With the extracted hybrid mask features $x_{i}$ for each mask $m_{i}$ , we compute the semantic alignment score $\\mathcal{P}(p, m_{i}) = \\cos(\\phi_{\\text{text}}(p), x_{i})$ , which measures how well the mask satisfies a given object $p$ (e.g., \"the pizza\"). We then select the top $k$ masks based on this score and apply the softmax function to normalize these selected scores, ensuring that they sum to one. This strategy prioritizes the most relevant and semantically aligned mask proposals while also enhancing computational efficiency for subsequent operations.", + "bbox": [ + 511, + 347, + 903, + 482 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To model the spatial relationships, a spatial relation function $\\mathcal{R}(p,q)$ quantified the relationship between two parsed objects $p$ and $q$ is defined as:", + "bbox": [ + 511, + 483, + 903, + 527 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {R} \\left(\\left(p, m _ {i}\\right), \\left(q, m _ {j}\\right)\\right) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} p \\text {s a t i s f i e s t h e r e l a t i o n w i t h} q, \\\\ 0 & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 539, + 911, + 589 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the spatial relation between $p$ and $q$ can be any of the defined relations like \"left\", \"right\", \"top\", \"bottom\", \"within\", \"smaller\", or \"bigger\". The satisfaction of the spatial relation is computed based on the position and size of the corresponding masks $m_{i}$ and $m_{j}$ for $p$ and $q$ .", + "bbox": [ + 511, + 590, + 903, + 666 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, we combine the probability of each object with the spatial relation probabilities to identify the target object. The overall likelihood of mask $m_{i}$ being the target $p$ is given by:", + "bbox": [ + 511, + 666, + 903, + 726 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {S} _ {m _ {i}} ^ {s} = \\sum_ {m _ {j}} \\mathcal {P} (p, m _ {i}) \\cdot \\mathcal {R} ((p, m _ {i}), (q, m _ {j})) \\cdot \\mathcal {P} (q, m _ {j}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 736, + 903, + 768 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{S}_{m_i}^s$ is the final probability of $m_{i}$ being the correct target, combining both its predicate satisfaction and its spatial relationships with other objects. If the referring text contains those spatial relations, the score of each mask is computed accordingly. Otherwise, $\\mathbf{S}_{m_i}^s$ is directly computed using Equation 3.", + "bbox": [ + 511, + 779, + 903, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Spatial Coherence Guidance. To enhance spatial coherence, we generate a spatial localization guidance $\\mathbf{G}_{\\mathrm{co}} \\in$", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "29637", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$[0,1]^{H\\times W}$ by using the referring text feature $(t)$ to identify the target region. Specifically, we apply the algorithm proposed in [2], which leverages self-attention mechanisms for expression localization and segmentation. This guidance map is constructed by calculating the similarity between $\\phi_{\\mathrm{text}}(t)$ and each visual token embedding derived from the CLIP image encoder with the self-attention mechanisms, yielding a localization map that broadly highlights areas corresponding to the target object or region. Consequently, positions with values closer to 1 represent higher similarity, thereby indicating regions more likely aligned with the target described by the text feature.", + "bbox": [ + 89, + 89, + 480, + 271 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Spatial Position Guidance. To incorporate spatial positions that align with the referring expression context, similar with [22], we introduce a set of positional guidance matrices $\\mathbf{G}_{\\mathrm{pos}} \\in [0,1]^{H \\times W}$ , each representing a position within the image. Here, $pos \\in \\{top, bottom, left, right, middle\\}$ denotes specific spatial attributes, such as \"top\" for upper regions. $\\mathbf{G}_{\\mathrm{pos}}$ is computed by:", + "bbox": [ + 89, + 272, + 480, + 380 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {G} _ {\\text {p o s}} = \\left\\{ \\begin{array}{l l} \\mathcal {E} (\\text {p o s}) & \\text {i f p o s i t i o n p o s a p p e r a s i n} t, \\\\ \\mathbf {1} _ {H \\times W} & \\text {i f n o p o s i t i o n p o s a p p e r a s i n} t, \\end{array} \\right. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 388, + 482, + 429 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{E}(\\mathrm{pos})$ represents the emphasis function that returns a value between 0 and 1 based on the position's emphasis. To generate these matrices, we compute the distance of each pixel to the corresponding region defined by pos, normalize the distances to the range [0, 1], and directly assign these values to the positional guidance matrix. This structure reflects a gradual transition in spatial emphasis across the image, enabling the model to focus on regions by the referring expression's context.", + "bbox": [ + 89, + 436, + 482, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this manner, we define $\\mathbf{G}_{\\mathrm{pos}}$ for each desired direction. Once the pos appears in the referring text, we incorporate this spatial guidance into our spatial coherence guidance $\\mathbf{G}_{\\mathrm{co}}$ by performing an element-wise multiplication. For each direction dir, the spatial guidance matrix $\\mathbf{G}$ is defined as:", + "bbox": [ + 89, + 573, + 482, + 662 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {G} = \\mathbf {G} _ {\\mathrm {c o}} \\odot \\mathbf {G} _ {\\mathrm {p o s}}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 223, + 662, + 480, + 680 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Spatial Guidance Score. The spatial guidance score $\\mathbf{S}_{m_i}^g$ for each mask proposal $m_{i}$ is computed as the difference between the mean spatial guidance values within the mask. Specifically, we define the spatial guidance score as:", + "bbox": [ + 89, + 685, + 482, + 746 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {S} _ {m _ {i}} ^ {g} = \\frac {\\operatorname {S u m} (\\mathbf {G} \\odot m _ {i})}{\\operatorname {S u m} (m _ {i})} - \\lambda \\cdot \\frac {\\operatorname {S u m} (\\mathbf {G} \\odot (1 - m _ {i}))}{\\operatorname {S u m} (1 - m _ {i})}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 753, + 482, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda$ is a hyperparameter that controls the importance of the negative score. This score $\\mathbf{S}_{m_i}^g$ serves as a spatial coherence measure, with larger values indicating better spatial alignment between the mask proposal and the referring expression.", + "bbox": [ + 89, + 794, + 482, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, we apply softmax normalization to the semantic alignment score $\\mathbf{S}_{m_i}^s$ and spatial guidance score $\\mathbf{S}_{m_i}^g$ for all", + "bbox": [ + 89, + 869, + 482, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "masks, and fuse them together. The final score for each mask $m_{i}$ is obtained by:", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {S} _ {m _ {i}} = (1 - \\alpha) \\mathbf {S} _ {m _ {i}} ^ {s} + \\alpha \\mathbf {S} _ {m _ {i}} ^ {g}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 132, + 903, + 150 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha$ controls the trade-off between semantic alignment and spatial guidance. The final referring semantic segmentation result $m^{*}$ is generated by selecting the mask with the highest mask score.", + "bbox": [ + 511, + 159, + 903, + 219 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This comprehensive approach allows us to prioritize mask proposals that are not only semantically aligned with the text feature but also spatially consistent with the specified direction. Thus, masks are refined based on both semantic and spatial coherence, enhancing the accuracy of the selected mask with respect to the referring expression's spatial context.", + "bbox": [ + 511, + 220, + 905, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 339, + 645, + 356 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets and Metrics", + "text_level": 1, + "bbox": [ + 511, + 364, + 710, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To evaluate the proposed method, we use several main RIS datasets: RefCOCO [21], RefCOCO+ [21], and RefCOCOg [19]. These datasets are derived from the MSCOCO [15] dataset and feature images annotated with detailed referring expressions that pinpoint specific objects or regions. Each dataset brings unique aspects to the referring expressions it contains. RefCOCO often includes positional information like \"left\" and \"right,\" which is banned in RefCOCO+, and RefCOCOg has more elaborate sentence structures. We also evaluate it on PhraseCut [36] dataset, which introduces structured textual descriptions that detail attributes, categories, and relationships among objects. We employ two primary metrics to evaluate the performance: overall Intersection over Union (oIoU) and mean Intersection over Union (mIoU). oIoU assesses the total overlap between predicted and ground truth regions relative to their combined area, making it particularly stringent for inaccuracies in larger segments. On the other hand, mIoU calculates the average overlap for each individual instance, ensuring a balanced consideration of performance across objects of varying sizes. Together, these metrics provide a robust framework for assessing the effectiveness of the proposed method in RIS tasks.", + "bbox": [ + 511, + 386, + 906, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 742, + 730, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The experiments were run on a single NVIDIA RTX 3090 GPU. Following previous work [22, 32], we use the default ViT-H SAM model, and the hyperparameters \"predicted iou threshold\" and \"stability score threshold\" were set to 0.7, the \"points per side\" was set to 8. We use CLIP with ViTB/16 backbone in both hybrid feature extraction and spatial coherence guidance. We set $\\alpha$ and $\\beta$ to 0.6 and 2, respectively, for all datasets. The hyperparameter $\\lambda$ is empirically set to 9 to balance the propensity of both excessively large", + "bbox": [ + 511, + 763, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "29638", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/c007b91f6f3be48f01f8c8e6cc35e2cb37059f73fdc37a5fd54170fe11984253.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MetricMethodVision BackbonePre-trained ModelRefCOCORefCOCO+RefCOCOg
valtestAtestBvaltestAtestBvaltest
oloUzero-shot methods w/ additional training
Pseudo-RIS [42]ViT-BSAM, CoCa, CLIP37.3343.4331.9040.1946.4333.6341.6343.52
VLM-VG [33]R101COCO*, VLM-VG*45.4048.0041.4037.0040.7030.5042.8044.10
zero-shot methods w/o additional training
Grad-CAM [26]R50SAM, CLIP23.4423.9121.6026.6727.2024.8423.0023.91
MaskCLIP [43]R50SAM, CLIP20.1820.5221.3022.0622.4324.6123.0523.41
Global-Local [41]R50FreeSOLO, CLIP24.5823.3824.3525.8724.6125.6130.0729.83
Global-Local [41]R50SAM, CLIP24.5526.0021.0326.6229.9922.2328.9230.48
Global-Local [41]ViT-BSAM, CLIP21.7124.4820.5123.7028.1221.8626.5728.21
Ref-Diff [22]ViT-BSAM, SD, CLIP35.1637.4434.5035.5638.6631.4038.6237.50
TAS [32]ViT-BSAM, BLIP2, CLIP29.5330.2628.2433.2138.7728.0135.8436.16
OursViT-BSAM,CLIP41.8144.5238.535.7441.4330.942.4742.97
mIoUweakly-supervised methods
CLRL [10]ViT-B-31.0632.3030.1131.2832.1130.1332.88-
PPT [4]ViT-BSAM46.7645.3346.2845.3445.8444.7742.97-
zero-shot methods w/ additional training
Pseudo-RIS [42]ViT-BSAM, CoCa, CLIP41.0548.1933.4844.3351.4235.0845.9946.67
VLM-VG [33]R101COCO*, VLM-VG*49.9053.1046.7042.7047.3036.2048.0048.50
zero-shot methods w/o additional training
Grad-CAM [26]R50SAM, CLIP30.2231.9027.1733.9625.6632.2933.0532.50
MaskCLIP [43]R50SAM, CLIP25.6226.6625.1727.4928.4930.4730.1330.15
Global-Local [41]R50FreeSOLO, CLIP26.7024.9926.4828.2226.5427.8633.0233.12
Global-Local [41]R50SAM, CLIP31.8332.9328.6434.9737.1130.6140.6640.94
Global-Local [41]ViT-BSAM, CLIP33.1236.5229.5835.2939.5831.8940.0840.74
CaR [31]ViT-B and ViT-LCLIP33.5735.3630.5134.2236.0331.0236.6736.57
Ref-Diff [22]ViT-BSAM, SD, CLIP37.2138.4037.1937.2940.5133.0144.0244.51
TAS [32]ViT-BSAM, BLIP2, CLIP39.8441.0836.2443.6349.1336.5446.6246.80
OursViT-BSAM, CLIP49.4853.3745.1943.4049.1337.1751.2551.59
", + "bbox": [ + 91, + 90, + 906, + 489 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and small mask regions receiving disproportionately high scores. When the referring text includes \"big\" or \"small\", $\\lambda$ is adjusted to 3 or 14, respectively.", + "bbox": [ + 89, + 547, + 482, + 594 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Results", + "text_level": 1, + "bbox": [ + 89, + 604, + 184, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Tab. 1, we evaluate our model on RefCOCO, RefCOCO+, and RefCOCOg and compare it with other state-of-the-art (SOTA) zero-shot models. Here, we evaluate the MaskCLIP, Grad-CAM methods by computing the similarity between their feature maps and SAM's mask proposals. Since the original Global-Local method used FreeSOLO as the mask extractor, we have re-evaluated the performance of Global-Local using SAM as the mask extractor, across different backbones. We also provide results of Ref-Diff, and TAS using different backbones. Additionally, we include weakly-supervised methods and zero-shot methods with extra training or additional datasets. Our method achieves excellent results on all three datasets. In terms of oIoU, our method improves by $4\\% -7\\%$ over SOTA methods on RefCOCO and RefCOCOg, and although it does not achieve the best performance on the testB set of RefCOCO+, it is only $0.5\\%$ lower than the SOTA method. For mIoU, our method outperforms SOTA methods by $4\\% -10\\%$ on", + "bbox": [ + 89, + 628, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "RefCOCO and RefCOCOg, and it also achieves comparable performance on RefCOCO+. Importantly, our method achieves comparable or even higher mIoU on all three datasets compared to methods that use additional training. This indicates that our method has a more balanced consideration across varying objects. In addition, we present the oIoU and mIoU results on the PhraseCut dataset's test set in Tab. 2. Our method demonstrates superior performance compared to previous methods, achieving the highest average score.", + "bbox": [ + 511, + 547, + 906, + 700 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e923b5cc7d91bfe6d85b62ccd30612ae7d581ac47011dcdf847e5ce968d735ef.jpg", + "table_caption": [ + "Table 1. Comparisons with the SOTA zero-shot approaches on RefCOCO, RefCOCO+, and RefCOCOg datasets. The best two results under the same setting, w/o additional training, are highlighted in bold and underlined, respectively. * indicates the extra dataset used to train the model." + ], + "table_footnote": [], + "table_body": "
MethodsoIoUmIoUavg.
Global-Local[41]23.64-23.64
TAS[32]25.6424.6625.15
Ref-Diff[22]29.4241.7535.59
Ours38.3936.9837.69
", + "bbox": [ + 519, + 710, + 898, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Comparison with existing methods on the PhraseCut dataset.", + "bbox": [ + 511, + 814, + 905, + 840 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Notably, our approach relies solely on SAM and CLIP, while TAS additionally incorporates the large BLIP2 model, and Ref-Diff leverages the Stable-Diffusion model; yet, our method still outperforms these more complex models. Be", + "bbox": [ + 511, + 840, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "29639", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1277ce1a277c109ba3b566aa1966d2be186089acf32cf30c8b45bbf86bdeb108.jpg", + "image_caption": [ + "Referring text : The woman in the middle sitting on the couch" + ], + "image_footnote": [], + "bbox": [ + 116, + 111, + 238, + 183 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f5ace7e5698fa3a39b60d8b8840f8ff975943e43f44e74062d902d212d502530.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 243, + 111, + 367, + 184 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/830633bea87253211cbe328d63df5cd3ea1cc8e9ab3a9858a5be510752d49064.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 372, + 111, + 496, + 184 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7b5f8fafda932b3da87cfd324016d04292534821c997d78c85db06ac7bd20dac.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 111, + 624, + 184 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3e19d8aff5ba5f5b18b9dae1849f7acaa6ecae15510770c12a9639641eeb1ed1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 629, + 111, + 753, + 184 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b77d261737c50977f865da24abc425a948dc663ff9f6afce6451a18c65056558.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 759, + 111, + 882, + 184 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7a38693728058a60cbceb8f5fcac0d3023b7b1cc6e1dbe9cb5f024880d7d138c.jpg", + "image_caption": [ + "Referring text : A beige surfboard being carried by a man in a wetsuit" + ], + "image_footnote": [], + "bbox": [ + 116, + 202, + 238, + 267 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/eee53ef79b8e3305441cfa9b64afec0e4f90098c407804bbdc048062d74aa4a3.jpg", + "image_caption": [ + "Referring text : The chair behind the guy wearing the stripes" + ], + "image_footnote": [], + "bbox": [ + 243, + 202, + 367, + 267 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4cee302decba7ab9189780f2885f63a5b63fd102baee62f37b72aa65d57161ff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 372, + 202, + 496, + 267 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/28e0d10764d9ab28c68cb10b705b6febd7f963662663d65390b4681aa14914c8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 202, + 624, + 267 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/469f846a77783f3c31ebfc0e8aae034d0af00f3f8e76ffd73d4206a3e59e0273.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 629, + 202, + 753, + 267 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/31a66926f07aad2c85bdcd0168a9f3044e9c84f1efcb3ffcddc02a8876c6d48e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 759, + 202, + 882, + 267 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1f5914738e6fbe1fccd79f0d00656339a0b5d56e38428bfba5346cac8d372001.jpg", + "image_caption": [ + "Image" + ], + "image_footnote": [], + "bbox": [ + 116, + 286, + 238, + 356 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9e63f670fb8d9503fafe1aad3a8352c603d5ba34e707c57c9f068c2abdf96138.jpg", + "image_caption": [ + "SAM-CLIP" + ], + "image_footnote": [], + "bbox": [ + 243, + 286, + 367, + 357 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0c338692d9ddcd6c0cdbfb4d2219da8964095f9724342ad53dc73a2e37424a1b.jpg", + "image_caption": [ + "Global-Local", + "Figure 3. Visual comparisons with existing methods. Our approach achieves more accurate localization and a complete segmentation of the target object." + ], + "image_footnote": [], + "bbox": [ + 372, + 286, + 496, + 356 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/428996b93833e65d31ceb1825e798a462f8302daa5b8b2819d8aefd20b646b83.jpg", + "image_caption": [ + "TAS" + ], + "image_footnote": [], + "bbox": [ + 501, + 286, + 624, + 356 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e691375c1beddf71d678894fc5e3d1362717e0231db25f3a4a187724500d1f2f.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 629, + 286, + 753, + 356 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a450e5f0eeb9c824b197f6ba298cd2b0e409d72a3be190172c8882b87947ee17.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 759, + 286, + 882, + 356 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sides, TAS and Ref-Diff show better performance in either mIoU or oIoU, whereas our method consistently outperforms both metrics. We present visual comparisons with existing methods in Fig. 3, highlighting the improvements in segmentation accuracy and the ability to capture spatial relationships.", + "bbox": [ + 89, + 407, + 483, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 506, + 243, + 522 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To evaluate the effectiveness of different strategies, we conduct extensive ablation studies on the val dataset of RefCOCO, RefCOCO+, and RefCOCOg datasets. We compare a range of feature extraction methods and spatial guidance strategies, examining their impact on performance.", + "bbox": [ + 89, + 527, + 483, + 604 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4.1. Ablation on Hybrid Feature Extraction", + "text_level": 1, + "bbox": [ + 89, + 609, + 410, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Local and Global Features. We first evaluate the performance of local and global features, including different strategies for global feature extraction. These strategies involve blurring images to reduce the influence of irrelevant regions on the global representation, applying a token mask to exclude non-mask regions, and using an attention mask to minimize the impact of non-mask regions on the CLS token. All these different mask strategies are applied to the last few layers of the CLIP image encoder. As shown in Tab. 3, the local features (L) consistently outperform the global methods across all datasets in terms of both oIoU and mIoU. Notably, combining a blur operation with the attention mask (G(blur+att_mask)) results in a significant improvement, approaching the performance of local features on the RefCOCO+ and RefCOCOg datasets. Hence, we adopt this approach to extract global features.", + "bbox": [ + 89, + 628, + 482, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Hybrid Global-Local Features. The intuitive way to fuse the global and local features is to compute a weighted sum", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a17ee82c9411b971f9040bb08647e07f4feaaa7103c64c9f6b2adda4cca8f007.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodRefCOCO oIoU mIoURefCOCO+ oIoU mIoURefCOCOg oIoU mIoU
L24.6532.6428.1636.4433.6342.35
G(blur)16.1722.0118.6125.1422.0031.24
G(tok_mask)20.5531.7121.8433.1819.8232.55
G(att_mask)20.0530.0721.4431.5124.1136.24
G(blur+att_mask)27.6337.1229.7539.9433.0344.45
G+L [41]21.7133.1223.7035.2926.5740.08
G+L28.9037.6432.2741.1238.0847.51
L2G29.9339.1733.1742.5537.7147.20
G2L31.7140.2734.4443.4039.1248.64
G2L + L2G31.3240.2334.6343.7338.4348.22
", + "bbox": [ + 514, + 405, + 905, + 603 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Results of different mask feature extraction methods on the val split of three datasets.", + "bbox": [ + 511, + 607, + 905, + 635 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "of the two. In Tab. 3, we present experimental results of this fusion strategy, employing the local and global feature extraction method from “G + L[41]”, which denotes the weighted sum of G(tok_mask) and cropped L, while “G + L” refers to the weighted sum of G(blur+att_mask) with L. We report both the results with the optimal weight that yielded the highest performance. Our results demonstrate that, due to the effectiveness of our global feature extraction strategy in generating a more accurate global representation, this weighted sum fusion significantly outperforms the previous method.", + "bbox": [ + 511, + 642, + 906, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To fully exploit the attention mechanism in the model and facilitate a comprehensive interaction between the global and local features, we propose generating a hybrid feature by fusing features from the later layers of one branch's image encoder into the other branch. In Tab. 3, we present the results of different hybrid fusion strategies.", + "bbox": [ + 511, + 810, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "29640", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d4e7d85b32056161d923d33d20ce421f7cb53e56d477d3cfc621278e49f95477.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Start LayerG2LL2G
oIoUmIoUoIoUmIoU
722.6134.7829.1239.33
838.1647.8234.6743.62
939.1248.6437.6147.17
1036.5046.4737.3846.81
1134.1242.2633.2844.06
", + "bbox": [ + 107, + 88, + 464, + 209 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The \"L2G\" strategy, where local branch features are incorporated into the global branch, shows a significant improvement over \"G + L\". This highlights the effectiveness of our hybrid feature extraction method. On the other hand, the \"G2L\" strategy, where global features are fused into the local branch, provides the best results overall. This indicates that local features, which are crucial for fine-grained object identification, benefit most from the additional global context, allowing the model to achieve more accurate and contextually informed segmentations. The key difference between \"G2L\" and \"L2G\" is that the global branch applies an attention mask, while the local branch does not. Global features require a token mask for integration into the local branch, whereas local features can be directly fused into the global branch, resulting in distinct \"G2L\" and \"L2G\" outputs. We also explore the fusion of both \"L2G\" and \"G2L\" strategies using a weighted sum operation, but no further improvements are observed. Consequently, we adopt the \"G2L\" strategy for hybrid feature extraction in our implementation.", + "bbox": [ + 91, + 260, + 482, + 561 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of Starting Layer for Hybrid Fusion We further investigate the effect of starting the hybrid fusion at different layers of the image encoder on the RefCOCOg dataset. In Tab. 4, we present the results of applying the fusion at various layers to determine which starting layer yields the best performance. Our findings show that initiating fusion at the 9-th layer yields the best performance, suggesting that representations at this layer provide the most meaningful and contextually relevant information for effective interaction between global and local features.", + "bbox": [ + 89, + 564, + 482, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4.2. Ablation on Spatial Guidance", + "text_level": 1, + "bbox": [ + 89, + 727, + 344, + 742 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Building upon the proposed hybrid global-local feature extraction method, we further conduct ablation studies to assess the impact of different components in the spatial guidance augmentation approach. Our experiments primarily focus on two datasets: RefCOCO and RefCOCOg.", + "bbox": [ + 89, + 747, + 482, + 823 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of Spatial Relationship Guidance From Tab. 5, we can see that incorporating spatial relationship guidance (Rel) leads to consistent performance improvements across all datasets. However, the improvements are more limited on RefCOCOg, where referring expressions typically con", + "bbox": [ + 89, + 824, + 482, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e77cf6ebfbe95dd480843f4f845f90b571b094e72e1e92fc36af8bf353a3b9d9.jpg", + "table_caption": [ + "Table 4. Results of the starting layer $l$ for hybrid fusion of global and local branches on the val split of RefCOCOg dataset." + ], + "table_footnote": [], + "table_body": "
G2LRelGcoGposRefCOCORefCOCOg
oIoUmIoUoIoUmIoU
✓31.7140.2739.1248.64
✓✓35.6844.2939.6848.99
✓✓✓35.6844.2939.7349.02
✓✓✓✓41.8149.4842.4751.25
", + "bbox": [ + 516, + 89, + 898, + 195 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Ablation study on the different spatial guidance.", + "bbox": [ + 537, + 205, + 879, + 220 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tain fewer spatial relationships. The most significant improvement is seen on RefCOCO, where both mIoU and oIoU increase by approximately $4\\%$ . This is attributed to the richer spatial relationship descriptions in the referring expressions, which are better captured by the introduced guidance. Overall, these results highlight the effectiveness of spatial relationship guidance in enhancing referring expression comprehension, especially in datasets with complex and detailed spatial descriptions.", + "bbox": [ + 511, + 229, + 903, + 366 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of Spatial Coherence & Position Guidance Spatial coherence guidance $\\mathbf{G}_{\\mathrm{co}}$ is introduced to enhance segmentation coherence and mitigate the issue of partial masks being selected for the target object. As a result, improvements in oIoU and mIoU are modest, as only a small subset of data face this issue, which may not be fully reflected in the overall scores. However, when combined with spatial position guidance $\\mathbf{G}_{\\mathrm{pos}}$ together, as shown in Tab. 5, we observe a significant improvement in spatial position awareness, leading to more accurate target localization and enhanced segmentation performance. In referring expressions, spatial positional cues such as \"the left [object]\" often specify the target object's position without additional information. By introducing $\\mathbf{G}_{\\mathrm{pos}}$ , which explicitly encodes spatial positioning information, we improve the alignment between the referring text and the mask prediction.", + "bbox": [ + 511, + 367, + 906, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 625, + 633, + 640 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper presents a novel, training-free approach to zero-shot referring image segmentation (RIS), addressing challenges in aligning visual masks with referring expressions. Leveraging CLIP and SAM, our hybrid global-local feature extraction method combines mask-specific detail with contextual information to improve mask representation. Further, a spatial guidance augmentation strategy enhances spatial coherence and reduces ambiguities, effectively aligning masks with referring text. Experiments on RefCOCO, RefCOCO+, and RefCOCOg demonstrate that our method achieves substantial gains over zero-shot RIS models.", + "bbox": [ + 511, + 650, + 903, + 815 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work was supported in part by the National Natural Science Foundation of China (No.62106201, No.62376217), and in part by the Guangdong Basic and Applied Basic Research Foundation (No.2025A1515011501). The authors thank the anonymous reviewers for their helpful feedback.", + "bbox": [ + 511, + 816, + 903, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "29641", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 121, + 187, + 136 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021. 2", + "[2] Walid Bousselham, Felix Petersen, Vittorio Ferrari, and Hilde Kuehne. Grounding everything: Emerging localization properties in vision-language transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3828-3837, 2024. 2, 5", + "[3] Yong Xien Chng, Henry Zheng, Yizeng Han, Xuchong Qiu, and Gao Huang. Mask grounding for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26573-26583, 2024. 2", + "[4] Qiyuan Dai and Sibei Yang. Curriculum point prompting for weakly-supervised referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13711-13722, 2024. 2, 6", + "[5] Ruining Deng, Can Cui, Quan Liu, Tianyuan Yao, Lucas W Remedios, Shunxing Bao, Bennett A Landman, Lee E Weiless, Lori A Coburn, Keith T Wilson, et al. Segment anything model (sam) for digital pathology: Assess zero-shot segmentation on whole slide imaging. arXiv preprint arXiv:2304.04155, 2023. 2", + "[6] Sina Hajimiri, Ismail Ben Ayed, and Jose Dolz. Pay attention to your neighbours: Training-free open-vocabulary semantic segmentation. arXiv preprint arXiv:2404.08181, 2024. 2", + "[7] Ronghang Hu, Marcus Rohrbach, and Trevor Darrell. Segmentation from natural language expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 108-124. Springer, 2016. 2", + "[8] Ziling Huang and Shin'ichi Satoh. Referring image segmentation via joint mask contextual embedding learning and progressive alignment network. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 7753-7762, 2023. 2", + "[9] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4015-4026, 2023. 1, 3", + "[10] Jungbeom Lee, Sungjin Lee, Jinseok Nam, Seunghak Yu, Jaeyoung Do, and Tara Taghavi. Weakly supervised referring image segmentation with intra-chunk and inter-chunk consistency. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21870–21881, 2023. 2, 6", + "[11] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with" + ], + "bbox": [ + 93, + 147, + 482, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 2", + "[12] Jiachen Li, Qing Xie, Xiaojun Chang, Jinyu Xu, and Yongjian Liu. Mutually-guided hierarchical multi-modal feature learning for referring image segmentation. ACM Trans. Multimedia Comput. Commun. Appl., 2024. Just Accepted. 2", + "[13] Tianxiao Li, Junhong Chen, Yiheng Huang, Kesi Huang, Qiqiang Xia, Muhammad Asim, and Wenyin Liu. Smvt: Spectrum-driven multi-scale vision transformer for referring image segmentation. In Advanced Intelligent Computing Technology and Applications, pages 193-206, Singapore, 2024. Springer Nature Singapore. 2", + "[14] Yi Li, Hualiang Wang, Yiqun Duan, and Xiaomeng Li. Clip surgery for better explainability with enhancement in open-vocabulary tasks. arXiv preprint arXiv:2304.05653, 2023. 2", + "[15] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5", + "[16] Sihan Liu, Yiwei Ma, Xiaqing Zhang, Haowei Wang, Jiayi Ji, Xiaoshuai Sun, and Rongrong Ji. Rotated multi-scale interaction network for referring remote sensing image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26658-26668, 2024. 2", + "[17] Yang Liu, Muzhi Zhu, Hengtao Li, Hao Chen, Xinlong Wang, and Chunhua Shen. Matcher: Segment anything with one shot using all-purpose feature matching. arXiv preprint arXiv:2305.13310, 2023. 2", + "[18] Jun Ma, Yuting He, Feifei Li, Lin Han, Chenyu You, and Bo Wang. Segment anything in medical images. Nature Communications, 15(1):654, 2024. 2", + "[19] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20, 2016. 5", + "[20] Sayan Nag, Koustava Goswami, and Srikrishna Karanam. Safari: Adaptive sequence transformer for weakly supervised referring expression segmentation. In European Conference on Computer Vision, pages 485-503. Springer, 2025. 2", + "[21] Varun K Nagaraja, Vlad I Morariu, and Larry S Davis. Modeling context between objects for referring expression understanding. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 792-807. Springer, 2016. 5", + "[22] Minheng Ni, Yabo Zhang, Kailai Feng, Xiaoming Li, Yiwen Guo, and Wangmeng Zuo. Ref-diff: Zero-shot referring image segmentation with generative models. arXiv preprint arXiv:2308.16777, 2023. 1, 2, 5, 6" + ], + "bbox": [ + 517, + 93, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "29642", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[23] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1", + "[24] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 2", + "[25] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2", + "[26] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 618-626, 2017. 6", + "[27] Nisarg A Shah, Vibashan VS, and Vishal M Patel. Lqmformer: Language-aware query mask transformer for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12903-12913, 2024. 2", + "[28] Chao Shang, Zichen Song, Heqian Qiu, Lanxiao Wang, Fanman Meng, and Hongliang Li. Prompt-driven referring image segmentation with instance contrasting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4124-4134, 2024. 2", + "[29] Robin Strudel, Ivan Laptev, and Cordelia Schmid. Weakly-supervised segmentation of referring expressions. arXiv preprint arXiv:2205.04725, 2022. 2", + "[30] Sanjay Subramanian, William Merrill, Trevor Darrell, Matt Gardner, Sameer Singh, and Anna Rohrbach. Reclip: A strong zero-shot baseline for referring expression comprehension. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 5198-5215, 2022. 4", + "[31] Shuyang Sun, Runjia Li, Philip Torr, Xiuye Gu, and Siyang Li. Clip as rnn: Segment countless visual concepts without training endeavor. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13171-13182, 2024. 2, 6", + "[32] Yucheng Suo, Linchao Zhu, and Yi Yang. Text augmented spatial aware zero-shot referring image segmentation. In The 2023 Conference on Empirical Methods in Natural Language Processing, 2023. 1, 2, 5, 6", + "[33] Shijie Wang, Dahun Kim, Ali Taalimi, Chen Sun, and Weicheng Kuo. Learning visual grounding from generative vision and language model. arXiv preprint arXiv:2407.14563, 2024. 2, 6", + "[34] Wenxuan Wang, Tongtian Yue, Yisi Zhang, Longteng Guo, Xingjian He, Xinlong Wang, and Jing Liu. Unveiling parts beyond objects: Towards finer-granularity referring expres" + ], + "bbox": [ + 91, + 90, + 480, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "sion segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12998-13008, 2024. 2", + "[35] Xinlong Wang, Zhiding Yu, Shalini De Mello, Jan Kautz, Anima Anandkumar, Chunhua Shen, and Jose M Alvarez. Freesolo: Learning to segment objects without annotations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14176-14186, 2022. 2", + "[36] Chenyun Wu, Zhe Lin, Scott Cohen, Trung Bui, and Subhransu Maji. Phrasescut: Language-based image segmentation in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10216-10225, 2020. 5", + "[37] Defeng Xie, Ruichen Wang, Jian Ma, Chen Chen, Haonan Lu, Dong Yang, Fobo Shi, and Xiaodong Lin. Edit everything: A text-guided generative system for images editing. arXiv preprint arXiv:2304.14006, 2023. 2", + "[38] Li Xu, Mark He Huang, Xindi Shang, Zehuan Yuan, Ying Sun, and Jun Liu. Meta compositional referring expression segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19478-19487, 2023. 2", + "[39] Jinyu Yang, Mingqi Gao, Zhe Li, Shang Gao, Fangjing Wang, and Feng Zheng. Track anything: Segment anything meets videos. arXiv preprint arXiv:2304.11968, 2023. 2", + "[40] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022. 2", + "[41] Seonghoon Yu, Paul Hongsuck Seo, and Jeany Son. Zero-shot referring image segmentation with global-local context features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19456-19465, 2023. 1, 2, 3, 6, 7", + "[42] Seonghoon Yu, Paul Hongsuck Seo, and Jeany Son. Pseudoris: Distinctive pseudo-supervision generation for referring image segmentation. In Proceedings of the European Conference on Computer Vision, 2024. 2, 6", + "[43] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In European Conference on Computer Vision, pages 696-712. Springer, 2022. 2, 6", + "[44] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Wang, Lijuan Wang, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. Advances in Neural Information Processing Systems, 36, 2024. 2" + ], + "bbox": [ + 516, + 92, + 903, + 727 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "29643", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Hybrid Global-Local Representation with Augmented Spatial Guidance for Zero-Shot Referring Image Segmentation/9c88ce4f-9647-4320-bdd2-f6fc8a88e98e_model.json b/2025/Hybrid Global-Local Representation with Augmented Spatial Guidance for Zero-Shot Referring Image Segmentation/9c88ce4f-9647-4320-bdd2-f6fc8a88e98e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..40900622f266dc53f8104c221f8be37502ab72f7 --- /dev/null +++ b/2025/Hybrid Global-Local Representation with Augmented Spatial Guidance for Zero-Shot Referring Image Segmentation/9c88ce4f-9647-4320-bdd2-f6fc8a88e98e_model.json @@ -0,0 +1,2266 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.121, + 0.13, + 0.88, + 0.177 + ], + "angle": 0, + "content": "Hybrid Global-Local Representation with Augmented Spatial Guidance for Zero-Shot Referring Image Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.397, + 0.203, + 0.597, + 0.221 + ], + "angle": 0, + "content": "Ting Liu\\(^{1,2*}\\) Siyuan Li\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.192, + 0.221, + 0.805, + 0.257 + ], + "angle": 0, + "content": "1ASGO, School of Computer Science, Northwestern Polytechnical University \n2Shenzhen Research Institute of Northwestern Polytechnical University" + }, + { + "type": "text", + "bbox": [ + 0.287, + 0.26, + 0.697, + 0.275 + ], + "angle": 0, + "content": "liuting@nwpu.edu.cn, lisiyuan@mail.nwpu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.31, + 0.327, + 0.325 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.342, + 0.485, + 0.705 + ], + "angle": 0, + "content": "Recent advances in zero-shot referring image segmentation (RIS), driven by models such as the Segment Anything Model (SAM) and CLIP, have made substantial progress in aligning visual and textual information. Despite these successes, the extraction of precise and high-quality mask region representations remains a critical challenge, limiting the full potential of RIS tasks. In this paper, we introduce a training-free, hybrid global-local feature extraction approach that integrates detailed mask-specific features with contextual information from the surrounding area, enhancing mask region representation. To further strengthen alignment between mask regions and referring expressions, we propose a spatial guidance augmentation strategy that improves spatial coherence, which is essential for accurately localizing described areas. By incorporating multiple spatial cues, this approach facilitates more robust and precise referring segmentation. Extensive experiments on standard RIS benchmarks demonstrate that our method significantly outperforms existing zero-shot RIS models, achieving substantial performance gains. We believe our approach advances RIS tasks and establishes a versatile framework for region-text alignment, offering broader implications for cross-modal understanding and interaction. Code is available at https://github.com/fhgyuanshen/HybridGL." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.732, + 0.222, + 0.748 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.757, + 0.483, + 0.879 + ], + "angle": 0, + "content": "Referring image segmentation (RIS) is a critical task in computer vision, where the goal is to segment a specific object or region in an image based on a natural language expression. This task is essential for applications such as visual search, robot perception, and human-computer interaction. Recent advancements, particularly models like the Segment Anything Model (SAM) [9] and CLIP [23], have significantly advanced zero-shot RIS, enabling object seg" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.31, + 0.906, + 0.391 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.406, + 0.907, + 0.448 + ], + "angle": 0, + "content": "Figure 1. Common issues in existing methods: 1) Inaccurate mask feature extraction; 2) Incorrect spatial localization; 3) Incomplete segmentation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.458, + 0.907, + 0.548 + ], + "angle": 0, + "content": "mentation without the need for labeled data or task-specific training. In typical RIS pipelines, SAM generates a set of mask proposals, and CLIP extracts visual features for each mask region. For each mask, a similarity score is computed with the referring text, and the mask with the highest score is selected as the final prediction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.551, + 0.909, + 0.748 + ], + "angle": 0, + "content": "Despite the promising performance of existing methods, the accurate extraction of mask representations that align well with referring text remains underexplored. Most existing approaches either mask out areas outside the mask region or crop the image to focus solely on the mask region before feeding it into CLIP, which primarily concentrates on local features while often neglecting the essential surrounding context. Some works [22, 32, 41] have sought to introduce global surrounding context features. However, such methods remain simplistic and fail to fully capture the complex interplay between the mask and its context, limiting their ability to accurately match referring expressions with the correct mask regions, as shown in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.911, + 0.902 + ], + "angle": 0, + "content": "In this paper, we propose a novel hybrid global-local feature extraction approach to enhance mask region feature extraction, achieving a more precise and contextually rich mask representation without any additional training. Our approach seamlessly integrates local and global features, capturing both region-specific and context-aware information for each mask. Specifically, we design two complementary branches within CLIP to extract local region-specific visual features and broader context-aware visual features for each mask. Features from the global branch" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.314, + 0.901 + ], + "angle": 0, + "content": "*Ting Liu is the corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29634" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.166 + ], + "angle": 0, + "content": "are progressively fused into the local branch to generate a hybrid feature representation. This hybrid fusion allows the visual encoder to automatically capture and interact with the complementary information from both branches, yielding a more contextually enriched mask feature." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.168, + 0.483, + 0.497 + ], + "angle": 0, + "content": "Another challenge in referring semantic segmentation lies in the use of spatial relationships within referring expressions (e.g., \"left of\", \"bottom of\") to describe objects, as shown in Fig. 1. These spatial cues introduce complexity, as they require both the recognition of object locations and the relationship between them. Capturing and aligning these spatial descriptions with visual features is inherently difficult. While referring text offers rich contextual information about spatial relationships, effectively integrating this context with visual data remains a significant challenge. Without a mechanism to explicitly model and align this spatial information, accurately matching the textual description to the correct visual region becomes problematic. Moreover, directly computing the similarity between the extracted visual mask features and text features can lead to ambiguity in the segmentation, as the mask features may capture a mixture of information from multiple objects or regions. It could mistakenly segment only part of the target region, rather than the full object. To alleviate these limitations, we introduce an augmentation approach by introducing several spatial guidance including spatial relationships, coherence, and positional cues." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.5, + 0.482, + 0.619 + ], + "angle": 0, + "content": "Our experiments show that the proposed method significantly outperforms several zero-shot baselines and weakly-supervised referring segmentation methods, achieving significant accuracy improvements. The proposed framework offers a powerful, efficient approach to zero-shot referring image segmentation, with strong potential for practical deployment in real-world scenarios. The contributions of this paper can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.62, + 0.483, + 0.665 + ], + "angle": 0, + "content": "- We propose an innovative hybrid global-local feature extraction approach for RIS, enhancing mask region representation without additional training requirements." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.666, + 0.483, + 0.71 + ], + "angle": 0, + "content": "- We introduce a spatial guidance augmentation strategy that leverages spatial relationships, coherence, and positional cues to mitigate segmentation ambiguity." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.712, + 0.483, + 0.771 + ], + "angle": 0, + "content": "- Extensive experiments on the four public datasets, demonstrate that our method significantly outperforms existing state-of-the-art zero-shot semantic segmentation approaches." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.62, + 0.483, + 0.771 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.785, + 0.233, + 0.801 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.482, + 0.901 + ], + "angle": 0, + "content": "Referring Image Segmentation. Referring image segmentation is a visual grounding task that requires the model to understand a natural language expression describing a specific object within an image and accurately segment that object from the rest of the scene [7]. The goal of this task is to bridge the gap between visual and textual modalities, en" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.272 + ], + "angle": 0, + "content": "abling more sophisticated interactions with visual content. Fully supervised methods [3, 8, 12, 13, 16, 27, 28, 38] have achieved impressive performance in this area by effectively integrating information from both images and text descriptions. These methods rely on large datasets with detailed annotations, where each object mentioned in the text is precisely segmented in the corresponding image [33, 34]. Currently, some weakly supervised methods [4, 10, 20, 29] can learn and perform segmentation using a smaller amount of labeled data. However, the requirement for such detailed and expensive human-annotated data limits the scalability and applicability of supervised approaches." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.273, + 0.905, + 0.497 + ], + "angle": 0, + "content": "Foundation Model in Image Segmentation. Recent research has shown that segmentation knowledge can emerge from pre-trained foundation models (FMs) [1] such as CLIP and Stable Diffusion [25]. Though the standard CLIP model excels at recognizing object appearances, it falls short in grasping their exact locations. However, MaskCLIP [43] demonstrates that it is possible to adapt CLIP for segmentation purposes by making minor adjustments to its attention-pooling mechanism. Other works delve into various strategies for refining attention mechanisms [2, 6, 14], enhancing the model's localization capability. Specifically, SAM has shown promising capabilities in object segmentation in many works [5, 17, 18, 24, 37, 39, 44]. Overall, these findings suggest that FMs can serve as a valuable resource for zero-shot segmentation and related tasks." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.5, + 0.905, + 0.846 + ], + "angle": 0, + "content": "Zero-shot Referring Image Segmentation. To address the limitations of fully supervised methods, zero-shot referring image segmentation has gained significant attention. Zero-shot approaches aim to perform segmentation without requiring any labeled data for the target objects, making them highly flexible and scalable. One notable method in this domain is Global-Local [41], which leverages pretrained models like FreeSOLO [35] and CLIP to achieve zero-shot segmentation. Another approach, CaR [31], further enhances this process by recurrently applying CLIP to refine the segmentation mask iteratively. The introduction of Segment Anything Model (SAM) has marked a significant milestone in zero-shot referring image segmentation. For example, Ref-Diff [22] utilizes diffusion models to better understand the relationship between image and text pairs, leading to more accurate and context-aware segmentations. TAS [32] uses a captioner, BLIP2 [11], to provide additional context and enhance the segmentation results by generating descriptive captions for the images. PseudorIS [42] modifies the Global-Local's pipeline and incorporates a captioner like CoCa [40] for unsupervised training, leveraging the rich knowledge of current foundation models to improve segmentation accuracy." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Note that although we also adopt both local and global features, our approach differs significantly from Global-Local [41] in terms of both the methods for local and global" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "29635" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.137, + 0.09, + 0.868, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.362, + 0.908, + 0.404 + ], + "angle": 0, + "content": "Figure 2. The proposed framework combines hybrid global-local feature extraction with multiple spatial guidance mechanisms to improve zero-shot referring image segmentation, using mask proposals generated by SAM. By leveraging both broad context and local details, and enhancing segmentation with spatial guidance, the framework effectively augments the segmentation of target based on textual descriptions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.412, + 0.484, + 0.517 + ], + "angle": 0, + "content": "feature extraction and the fusion strategy of these features. Distinct from previous methods, our work introduces a novel, training-free approach to extracting mask representations using a hybrid feature extraction scheme. Furthermore, we incorporate multiple spatial guidance mechanisms to enhance semantic coherence and spatial alignment, improving the referring segmentation process." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.534, + 0.182, + 0.549 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.56, + 0.201, + 0.574 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.582, + 0.484, + 0.718 + ], + "angle": 0, + "content": "Given an input image \\( I \\in \\mathbb{R}^{H \\times W \\times 3} \\), where \\( H \\) and \\( W \\) are the height and width of the image, respectively, and a referring text \\( T \\), which describes the target object or region, the goal is to predict the segmentation mask \\( m^* \\in \\{0,1\\}^{H \\times W} \\) that corresponds to the region in \\( I \\) described by \\( t \\). We first adopt SAM (Segment Anything Model) [9] to generate a set of mask proposals \\( M = \\{m_1, m_2, \\ldots, m_n\\} \\), where each \\( m_i \\) is a binary mask highlighting a potential segment region in \\( I \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.901 + ], + "angle": 0, + "content": "For each mask \\(m_{i}\\in M\\), we use the CLIP visual encoder to extract our proposed hybrid global-local features \\(\\mathbf{x}_i\\in \\mathbb{R}^d\\). To match the referring expression with the visual features, we encode the referring expression \\(t\\) using the CLIP text encoder \\(\\phi_{\\mathrm{text}}(\\cdot)\\) to obtain the textual feature \\(\\mathbf{f}_t\\in \\mathbb{R}^d\\), adopting same strategy with [41]. Subsequently, we can compute the cosine similarity between \\(\\mathbf{f}_t\\) and each feature \\(\\mathbf{x}_i\\), obtaining the semantic alignment score \\(\\mathbf{S}_{m_i}^s\\) for each mask \\(m_{i}\\). Furthermore, we introduce a spatial guidance augmentation approach. By leveraging several spatial cues, we augment the scoring mask process to prioritize semantically and spatially aligned masks. Our method in" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.412, + 0.907, + 0.472 + ], + "angle": 0, + "content": "tegrates several spatial cues, specifically spatial coherence guidance, spatial position guidance, and spatial relationships guidance, to enhance the mask scoring process. The overall framework is shown in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.484, + 0.862, + 0.5 + ], + "angle": 0, + "content": "3.2. Hybrid Global-Local Feature Extraction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.507, + 0.907, + 0.672 + ], + "angle": 0, + "content": "To extract a more precise mask feature for mask \\( m_{i} \\), we propose the Hybrid Global-Local Feature Extraction method. This approach captures both region-specific and context-aware features, significantly enhancing mask feature extraction. The method consists of two branches: a local branch for region-specific feature extraction and a global branch for context-aware feature extraction. Features from the global branch are progressively fused into the local branch to generate a hybrid feature representation. This hybrid fusion allows the model to automatically capture the complementary information from the two branches." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.673, + 0.909, + 0.839 + ], + "angle": 0, + "content": "Local Feature Extraction. For the local region-specific feature extraction, we apply the binary mask \\( m \\) to the input image \\( I \\), creating a masked image \\( I_{\\mathrm{local}} = I \\cdot m_i \\). This ensures that only the relevant region contributes to feature extraction. The masked image \\( I_{\\mathrm{local}} \\) is then processed through the CLIP image encoder to obtain the region-specific features. Unlike most existing methods, we do not crop the mask region, ensuring that the input image remains well-aligned with the input of the global branch. This allows the naturally aligned features to be hybridized and deeply fused for enhanced feature extraction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Global Feature Extraction. For the context-aware global feature extraction, we use the original image \\(I\\) and apply a Gaussian blur to the non-mask regions, thereby emphasizing the relevant areas while preserving essential contextual" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "29636" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.226 + ], + "angle": 0, + "content": "information. To further direct the model's focus onto the mask region, we introduce an attention mask \\(\\mathbf{A}_{m_i}\\) during the self-attention operation in the last \\(l\\) layers of the CLIP transformer-based architecture as shown in Fig. 2. Considering the CLS token is embedded into the shared visual-textual space as the visual feature, we nullify the attention scores between the CLS token and image tokens outside the masked region, effectively preventing non-relevant areas from contributing to the attention on the CLS token." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.227, + 0.482, + 0.302 + ], + "angle": 0, + "content": "Hybrid Feature Extraction. To enhance the fusion of local and global features, we propose a hybrid fusion approach that enables the attention mechanism in the network to effectively and automatically capture the complementary information between these two features." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.303, + 0.482, + 0.469 + ], + "angle": 0, + "content": "Given that local features primarily focus on the mask region, which typically contains the most relevant information for the task, the global branch is intended to complement the local branch by providing a broader context. While the global branch captures a wider spatial range and provides essential contextual information beyond the mask, it may also introduce irrelevant information from regions outside the target. To maximize the benefits of both local and global features, it is crucial to selectively integrate them, ensuring that the global context enhances the local representation without diluting its focus on the target region." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.469, + 0.482, + 0.558 + ], + "angle": 0, + "content": "To achieve this, we apply a token mask to the global features during the fusion. Specifically, we mask out image tokens outside the mask region to ensure that only the relevant tokens contribute to the feature extraction process. The fusion of local and global features at layer \\(l\\) is performed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.155, + 0.563, + 0.482, + 0.589 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {\\text {h y b r i d}} ^ {(l)} = \\mathbf {x} _ {\\text {l o c a l}} ^ {(l - 1)} + \\beta \\cdot \\left(\\mathbf {x} _ {\\text {g l o b a l}} ^ {(l - 1)} \\cdot \\mathbf {B} _ {m _ {i}}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.596, + 0.483, + 0.782 + ], + "angle": 0, + "content": "where \\(\\mathbf{x}_{\\mathrm{local}}^{(l - 1)}\\) and \\(\\mathbf{x}_{\\mathrm{global}}^{(l - 1)}\\) represent the local and global features derived from the \\(l - 1\\)-th layer, respectively. The hyper-parameter \\(\\beta\\) is the relative contribution of the global feature to the fusion process. To ensure that only relevant tokens contribute to the feature fusion, we apply a mask \\(\\mathbf{B}_{m_i} \\in \\{0,1\\}^K\\) on the global feature, where each feature contains \\(K\\) image token features, effectively excluding tokens corresponding to regions outside the mask. The resulting hybrid feature \\(\\mathbf{x}_{\\mathrm{hybrid}}^{(l - 1)}\\) is then passed as input to the \\(l\\)-th layer of the CLIP image encoder, denoted \\(\\phi_{\\mathrm{visual}}^l\\), to produce a feature representation that integrates both global and local information:" + }, + { + "type": "equation", + "bbox": [ + 0.209, + 0.782, + 0.482, + 0.804 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {\\text {l o c a l}} ^ {(l)} = \\phi_ {\\text {v i s u a l}} ^ {l} \\left(\\mathbf {x} _ {\\text {h y b r i d}} ^ {(l)}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.809, + 0.482, + 0.87 + ], + "angle": 0, + "content": "The local feature \\(\\mathbf{x}_{\\mathrm{local}}^{(L)}\\) obtained from the last encoder layer \\(L\\), which effectively combines localized and global contextual information, serves as the final hybrid feature for the mask region." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.482, + 0.902 + ], + "angle": 0, + "content": "Semantic Alignment Score. We then extract hybrid visual features \\(\\mathbf{X}_{\\mathrm{hybrid}} = \\{\\mathbf{x}_1,\\mathbf{x}_2,\\dots ,\\mathbf{x}_n\\}\\) for each mask, where" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "each \\(\\mathbf{x}_i\\) corresponds to a mask proposal \\(m_{i}\\) .After extracting a referring text feature \\(\\mathbf{f}_t\\) from the CLIP text encoder, we calculate the cosine semantic alignment score between the referring text and each \\(\\mathbf{x}_i\\) .." + }, + { + "type": "equation", + "bbox": [ + 0.628, + 0.163, + 0.905, + 0.181 + ], + "angle": 0, + "content": "\\[\n\\mathbf {S} _ {m _ {i}} ^ {s} = \\cos \\left(\\phi_ {\\text {t e x t}} (t), x _ {i}\\right). \\tag {3}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.189, + 0.796, + 0.205 + ], + "angle": 0, + "content": "3.3. Spatial Guidance Augmentation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.211, + 0.905, + 0.256 + ], + "angle": 0, + "content": "We combine multiple spatial guidance mechanisms, including relationships, coherence, and position, which are detailed in the following." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.257, + 0.905, + 0.347 + ], + "angle": 0, + "content": "Spatial Relationship Guidance. To utilize the spatial relationships described in the referring text, we first parse the text into objects and spatial relations, following a method similar to [30]. For instance, objects might include phrases like \"the pizza\" or \"man\" while spatial relations describe the positioning between objects, such as \"right\"." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.348, + 0.905, + 0.483 + ], + "angle": 0, + "content": "With the extracted hybrid mask features \\( x_{i} \\) for each mask \\( m_{i} \\), we compute the semantic alignment score \\( \\mathcal{P}(p, m_{i}) = \\cos(\\phi_{\\text{text}}(p), x_{i}) \\), which measures how well the mask satisfies a given object \\( p \\) (e.g., \"the pizza\"). We then select the top \\( k \\) masks based on this score and apply the softmax function to normalize these selected scores, ensuring that they sum to one. This strategy prioritizes the most relevant and semantically aligned mask proposals while also enhancing computational efficiency for subsequent operations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.484, + 0.905, + 0.528 + ], + "angle": 0, + "content": "To model the spatial relationships, a spatial relation function \\(\\mathcal{R}(p,q)\\) quantified the relationship between two parsed objects \\(p\\) and \\(q\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.54, + 0.912, + 0.59 + ], + "angle": 0, + "content": "\\[\n\\mathcal {R} \\left(\\left(p, m _ {i}\\right), \\left(q, m _ {j}\\right)\\right) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} p \\text {s a t i s f i e s t h e r e l a t i o n w i t h} q, \\\\ 0 & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.592, + 0.905, + 0.667 + ], + "angle": 0, + "content": "where the spatial relation between \\( p \\) and \\( q \\) can be any of the defined relations like \"left\", \"right\", \"top\", \"bottom\", \"within\", \"smaller\", or \"bigger\". The satisfaction of the spatial relation is computed based on the position and size of the corresponding masks \\( m_{i} \\) and \\( m_{j} \\) for \\( p \\) and \\( q \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.667, + 0.905, + 0.727 + ], + "angle": 0, + "content": "Finally, we combine the probability of each object with the spatial relation probabilities to identify the target object. The overall likelihood of mask \\( m_{i} \\) being the target \\( p \\) is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.737, + 0.905, + 0.77 + ], + "angle": 0, + "content": "\\[\n\\mathbf {S} _ {m _ {i}} ^ {s} = \\sum_ {m _ {j}} \\mathcal {P} (p, m _ {i}) \\cdot \\mathcal {R} ((p, m _ {i}), (q, m _ {j})) \\cdot \\mathcal {P} (q, m _ {j}), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.905, + 0.87 + ], + "angle": 0, + "content": "where \\(\\mathbf{S}_{m_i}^s\\) is the final probability of \\(m_{i}\\) being the correct target, combining both its predicate satisfaction and its spatial relationships with other objects. If the referring text contains those spatial relations, the score of each mask is computed accordingly. Otherwise, \\(\\mathbf{S}_{m_i}^s\\) is directly computed using Equation 3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Spatial Coherence Guidance. To enhance spatial coherence, we generate a spatial localization guidance \\(\\mathbf{G}_{\\mathrm{co}} \\in\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "29637" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.09, + 0.482, + 0.272 + ], + "angle": 0, + "content": "\\([0,1]^{H\\times W}\\) by using the referring text feature \\((t)\\) to identify the target region. Specifically, we apply the algorithm proposed in [2], which leverages self-attention mechanisms for expression localization and segmentation. This guidance map is constructed by calculating the similarity between \\(\\phi_{\\mathrm{text}}(t)\\) and each visual token embedding derived from the CLIP image encoder with the self-attention mechanisms, yielding a localization map that broadly highlights areas corresponding to the target object or region. Consequently, positions with values closer to 1 represent higher similarity, thereby indicating regions more likely aligned with the target described by the text feature." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.273, + 0.482, + 0.381 + ], + "angle": 0, + "content": "Spatial Position Guidance. To incorporate spatial positions that align with the referring expression context, similar with [22], we introduce a set of positional guidance matrices \\(\\mathbf{G}_{\\mathrm{pos}} \\in [0,1]^{H \\times W}\\), each representing a position within the image. Here, \\(pos \\in \\{top, bottom, left, right, middle\\}\\) denotes specific spatial attributes, such as \"top\" for upper regions. \\(\\mathbf{G}_{\\mathrm{pos}}\\) is computed by:" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.389, + 0.483, + 0.43 + ], + "angle": 0, + "content": "\\[\n\\mathbf {G} _ {\\text {p o s}} = \\left\\{ \\begin{array}{l l} \\mathcal {E} (\\text {p o s}) & \\text {i f p o s i t i o n p o s a p p e r a s i n} t, \\\\ \\mathbf {1} _ {H \\times W} & \\text {i f n o p o s i t i o n p o s a p p e r a s i n} t, \\end{array} \\right. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.438, + 0.483, + 0.573 + ], + "angle": 0, + "content": "where \\(\\mathcal{E}(\\mathrm{pos})\\) represents the emphasis function that returns a value between 0 and 1 based on the position's emphasis. To generate these matrices, we compute the distance of each pixel to the corresponding region defined by pos, normalize the distances to the range [0, 1], and directly assign these values to the positional guidance matrix. This structure reflects a gradual transition in spatial emphasis across the image, enabling the model to focus on regions by the referring expression's context." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.574, + 0.483, + 0.663 + ], + "angle": 0, + "content": "In this manner, we define \\(\\mathbf{G}_{\\mathrm{pos}}\\) for each desired direction. Once the pos appears in the referring text, we incorporate this spatial guidance into our spatial coherence guidance \\(\\mathbf{G}_{\\mathrm{co}}\\) by performing an element-wise multiplication. For each direction dir, the spatial guidance matrix \\(\\mathbf{G}\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.225, + 0.664, + 0.482, + 0.681 + ], + "angle": 0, + "content": "\\[\n\\mathbf {G} = \\mathbf {G} _ {\\mathrm {c o}} \\odot \\mathbf {G} _ {\\mathrm {p o s}}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.686, + 0.483, + 0.747 + ], + "angle": 0, + "content": "Spatial Guidance Score. The spatial guidance score \\(\\mathbf{S}_{m_i}^g\\) for each mask proposal \\(m_{i}\\) is computed as the difference between the mean spatial guidance values within the mask. Specifically, we define the spatial guidance score as:" + }, + { + "type": "equation", + "bbox": [ + 0.105, + 0.754, + 0.483, + 0.789 + ], + "angle": 0, + "content": "\\[\n\\mathbf {S} _ {m _ {i}} ^ {g} = \\frac {\\operatorname {S u m} (\\mathbf {G} \\odot m _ {i})}{\\operatorname {S u m} (m _ {i})} - \\lambda \\cdot \\frac {\\operatorname {S u m} (\\mathbf {G} \\odot (1 - m _ {i}))}{\\operatorname {S u m} (1 - m _ {i})}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.87 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is a hyperparameter that controls the importance of the negative score. This score \\(\\mathbf{S}_{m_i}^g\\) serves as a spatial coherence measure, with larger values indicating better spatial alignment between the mask proposal and the referring expression." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.904 + ], + "angle": 0, + "content": "Finally, we apply softmax normalization to the semantic alignment score \\(\\mathbf{S}_{m_i}^s\\) and spatial guidance score \\(\\mathbf{S}_{m_i}^g\\) for all" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "masks, and fuse them together. The final score for each mask \\( m_{i} \\) is obtained by:" + }, + { + "type": "equation", + "bbox": [ + 0.611, + 0.133, + 0.905, + 0.151 + ], + "angle": 0, + "content": "\\[\n\\mathbf {S} _ {m _ {i}} = (1 - \\alpha) \\mathbf {S} _ {m _ {i}} ^ {s} + \\alpha \\mathbf {S} _ {m _ {i}} ^ {g}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.16, + 0.905, + 0.22 + ], + "angle": 0, + "content": "where \\(\\alpha\\) controls the trade-off between semantic alignment and spatial guidance. The final referring semantic segmentation result \\(m^{*}\\) is generated by selecting the mask with the highest mask score." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.222, + 0.906, + 0.326 + ], + "angle": 0, + "content": "This comprehensive approach allows us to prioritize mask proposals that are not only semantically aligned with the text feature but also spatially consistent with the specified direction. Thus, masks are refined based on both semantic and spatial coherence, enhancing the accuracy of the selected mask with respect to the referring expression's spatial context." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.34, + 0.646, + 0.357 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.365, + 0.712, + 0.38 + ], + "angle": 0, + "content": "4.1. Datasets and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.387, + 0.907, + 0.735 + ], + "angle": 0, + "content": "To evaluate the proposed method, we use several main RIS datasets: RefCOCO [21], RefCOCO+ [21], and RefCOCOg [19]. These datasets are derived from the MSCOCO [15] dataset and feature images annotated with detailed referring expressions that pinpoint specific objects or regions. Each dataset brings unique aspects to the referring expressions it contains. RefCOCO often includes positional information like \"left\" and \"right,\" which is banned in RefCOCO+, and RefCOCOg has more elaborate sentence structures. We also evaluate it on PhraseCut [36] dataset, which introduces structured textual descriptions that detail attributes, categories, and relationships among objects. We employ two primary metrics to evaluate the performance: overall Intersection over Union (oIoU) and mean Intersection over Union (mIoU). oIoU assesses the total overlap between predicted and ground truth regions relative to their combined area, making it particularly stringent for inaccuracies in larger segments. On the other hand, mIoU calculates the average overlap for each individual instance, ensuring a balanced consideration of performance across objects of varying sizes. Together, these metrics provide a robust framework for assessing the effectiveness of the proposed method in RIS tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.743, + 0.731, + 0.759 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.902 + ], + "angle": 0, + "content": "The experiments were run on a single NVIDIA RTX 3090 GPU. Following previous work [22, 32], we use the default ViT-H SAM model, and the hyperparameters \"predicted iou threshold\" and \"stability score threshold\" were set to 0.7, the \"points per side\" was set to 8. We use CLIP with ViTB/16 backbone in both hybrid feature extraction and spatial coherence guidance. We set \\(\\alpha\\) and \\(\\beta\\) to 0.6 and 2, respectively, for all datasets. The hyperparameter \\(\\lambda\\) is empirically set to 9 to balance the propensity of both excessively large" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29638" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.091, + 0.907, + 0.49 + ], + "angle": 0, + "content": "
MetricMethodVision BackbonePre-trained ModelRefCOCORefCOCO+RefCOCOg
valtestAtestBvaltestAtestBvaltest
oloUzero-shot methods w/ additional training
Pseudo-RIS [42]ViT-BSAM, CoCa, CLIP37.3343.4331.9040.1946.4333.6341.6343.52
VLM-VG [33]R101COCO*, VLM-VG*45.4048.0041.4037.0040.7030.5042.8044.10
zero-shot methods w/o additional training
Grad-CAM [26]R50SAM, CLIP23.4423.9121.6026.6727.2024.8423.0023.91
MaskCLIP [43]R50SAM, CLIP20.1820.5221.3022.0622.4324.6123.0523.41
Global-Local [41]R50FreeSOLO, CLIP24.5823.3824.3525.8724.6125.6130.0729.83
Global-Local [41]R50SAM, CLIP24.5526.0021.0326.6229.9922.2328.9230.48
Global-Local [41]ViT-BSAM, CLIP21.7124.4820.5123.7028.1221.8626.5728.21
Ref-Diff [22]ViT-BSAM, SD, CLIP35.1637.4434.5035.5638.6631.4038.6237.50
TAS [32]ViT-BSAM, BLIP2, CLIP29.5330.2628.2433.2138.7728.0135.8436.16
OursViT-BSAM,CLIP41.8144.5238.535.7441.4330.942.4742.97
mIoUweakly-supervised methods
CLRL [10]ViT-B-31.0632.3030.1131.2832.1130.1332.88-
PPT [4]ViT-BSAM46.7645.3346.2845.3445.8444.7742.97-
zero-shot methods w/ additional training
Pseudo-RIS [42]ViT-BSAM, CoCa, CLIP41.0548.1933.4844.3351.4235.0845.9946.67
VLM-VG [33]R101COCO*, VLM-VG*49.9053.1046.7042.7047.3036.2048.0048.50
zero-shot methods w/o additional training
Grad-CAM [26]R50SAM, CLIP30.2231.9027.1733.9625.6632.2933.0532.50
MaskCLIP [43]R50SAM, CLIP25.6226.6625.1727.4928.4930.4730.1330.15
Global-Local [41]R50FreeSOLO, CLIP26.7024.9926.4828.2226.5427.8633.0233.12
Global-Local [41]R50SAM, CLIP31.8332.9328.6434.9737.1130.6140.6640.94
Global-Local [41]ViT-BSAM, CLIP33.1236.5229.5835.2939.5831.8940.0840.74
CaR [31]ViT-B and ViT-LCLIP33.5735.3630.5134.2236.0331.0236.6736.57
Ref-Diff [22]ViT-BSAM, SD, CLIP37.2138.4037.1937.2940.5133.0144.0244.51
TAS [32]ViT-BSAM, BLIP2, CLIP39.8441.0836.2443.6349.1336.5446.6246.80
OursViT-BSAM, CLIP49.4853.3745.1943.4049.1337.1751.2551.59
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.495, + 0.907, + 0.54 + ], + "angle": 0, + "content": "Table 1. Comparisons with the SOTA zero-shot approaches on RefCOCO, RefCOCO+, and RefCOCOg datasets. The best two results under the same setting, w/o additional training, are highlighted in bold and underlined, respectively. * indicates the extra dataset used to train the model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.549, + 0.483, + 0.595 + ], + "angle": 0, + "content": "and small mask regions receiving disproportionately high scores. When the referring text includes \"big\" or \"small\", \\(\\lambda\\) is adjusted to 3 or 14, respectively." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.606, + 0.185, + 0.621 + ], + "angle": 0, + "content": "4.3. Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In Tab. 1, we evaluate our model on RefCOCO, RefCOCO+, and RefCOCOg and compare it with other state-of-the-art (SOTA) zero-shot models. Here, we evaluate the MaskCLIP, Grad-CAM methods by computing the similarity between their feature maps and SAM's mask proposals. Since the original Global-Local method used FreeSOLO as the mask extractor, we have re-evaluated the performance of Global-Local using SAM as the mask extractor, across different backbones. We also provide results of Ref-Diff, and TAS using different backbones. Additionally, we include weakly-supervised methods and zero-shot methods with extra training or additional datasets. Our method achieves excellent results on all three datasets. In terms of oIoU, our method improves by \\(4\\% -7\\%\\) over SOTA methods on RefCOCO and RefCOCOg, and although it does not achieve the best performance on the testB set of RefCOCO+, it is only \\(0.5\\%\\) lower than the SOTA method. For mIoU, our method outperforms SOTA methods by \\(4\\% -10\\%\\) on" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.549, + 0.907, + 0.701 + ], + "angle": 0, + "content": "RefCOCO and RefCOCOg, and it also achieves comparable performance on RefCOCO+. Importantly, our method achieves comparable or even higher mIoU on all three datasets compared to methods that use additional training. This indicates that our method has a more balanced consideration across varying objects. In addition, we present the oIoU and mIoU results on the PhraseCut dataset's test set in Tab. 2. Our method demonstrates superior performance compared to previous methods, achieving the highest average score." + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.712, + 0.9, + 0.803 + ], + "angle": 0, + "content": "
MethodsoIoUmIoUavg.
Global-Local[41]23.64-23.64
TAS[32]25.6424.6625.15
Ref-Diff[22]29.4241.7535.59
Ours38.3936.9837.69
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.815, + 0.906, + 0.841 + ], + "angle": 0, + "content": "Table 2. Comparison with existing methods on the PhraseCut dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Notably, our approach relies solely on SAM and CLIP, while TAS additionally incorporates the large BLIP2 model, and Ref-Diff leverages the Stable-Diffusion model; yet, our method still outperforms these more complex models. Be" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29639" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.121, + 0.096, + 0.544, + 0.111 + ], + "angle": 0, + "content": "Referring text : The woman in the middle sitting on the couch" + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.112, + 0.24, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.112, + 0.368, + 0.185 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.112, + 0.497, + 0.185 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.112, + 0.625, + 0.185 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.63, + 0.112, + 0.754, + 0.185 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.112, + 0.883, + 0.185 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.12, + 0.187, + 0.598, + 0.202 + ], + "angle": 0, + "content": "Referring text : A beige surfboard being carried by a man in a wetsuit" + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.203, + 0.24, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.203, + 0.368, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.203, + 0.497, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.203, + 0.625, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.63, + 0.203, + 0.754, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.203, + 0.883, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.12, + 0.27, + 0.537, + 0.286 + ], + "angle": 0, + "content": "Referring text : The chair behind the guy wearing the stripes" + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.287, + 0.24, + 0.357 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.162, + 0.358, + 0.203, + 0.37 + ], + "angle": 0, + "content": "Image" + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.287, + 0.368, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.273, + 0.358, + 0.346, + 0.369 + ], + "angle": 0, + "content": "SAM-CLIP" + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.287, + 0.497, + 0.357 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.395, + 0.358, + 0.477, + 0.369 + ], + "angle": 0, + "content": "Global-Local" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.287, + 0.625, + 0.357 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.549, + 0.358, + 0.577, + 0.368 + ], + "angle": 0, + "content": "TAS" + }, + { + "type": "image", + "bbox": [ + 0.63, + 0.287, + 0.754, + 0.357 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.675, + 0.358, + 0.706, + 0.368 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.287, + 0.883, + 0.357 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.806, + 0.358, + 0.828, + 0.369 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.371, + 0.908, + 0.4 + ], + "angle": 0, + "content": "Figure 3. Visual comparisons with existing methods. Our approach achieves more accurate localization and a complete segmentation of the target object." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.408, + 0.484, + 0.5 + ], + "angle": 0, + "content": "sides, TAS and Ref-Diff show better performance in either mIoU or oIoU, whereas our method consistently outperforms both metrics. We present visual comparisons with existing methods in Fig. 3, highlighting the improvements in segmentation accuracy and the ability to capture spatial relationships." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.507, + 0.245, + 0.523 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.529, + 0.484, + 0.605 + ], + "angle": 0, + "content": "To evaluate the effectiveness of different strategies, we conduct extensive ablation studies on the val dataset of RefCOCO, RefCOCO+, and RefCOCOg datasets. We compare a range of feature extraction methods and spatial guidance strategies, examining their impact on performance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.611, + 0.411, + 0.625 + ], + "angle": 0, + "content": "4.4.1. Ablation on Hybrid Feature Extraction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.87 + ], + "angle": 0, + "content": "Local and Global Features. We first evaluate the performance of local and global features, including different strategies for global feature extraction. These strategies involve blurring images to reduce the influence of irrelevant regions on the global representation, applying a token mask to exclude non-mask regions, and using an attention mask to minimize the impact of non-mask regions on the CLS token. All these different mask strategies are applied to the last few layers of the CLIP image encoder. As shown in Tab. 3, the local features (L) consistently outperform the global methods across all datasets in terms of both oIoU and mIoU. Notably, combining a blur operation with the attention mask (G(blur+att_mask)) results in a significant improvement, approaching the performance of local features on the RefCOCO+ and RefCOCOg datasets. Hence, we adopt this approach to extract global features." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Hybrid Global-Local Features. The intuitive way to fuse the global and local features is to compute a weighted sum" + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.406, + 0.906, + 0.604 + ], + "angle": 0, + "content": "
MethodRefCOCO oIoU mIoURefCOCO+ oIoU mIoURefCOCOg oIoU mIoU
L24.6532.6428.1636.4433.6342.35
G(blur)16.1722.0118.6125.1422.0031.24
G(tok_mask)20.5531.7121.8433.1819.8232.55
G(att_mask)20.0530.0721.4431.5124.1136.24
G(blur+att_mask)27.6337.1229.7539.9433.0344.45
G+L [41]21.7133.1223.7035.2926.5740.08
G+L28.9037.6432.2741.1238.0847.51
L2G29.9339.1733.1742.5537.7147.20
G2L31.7140.2734.4443.4039.1248.64
G2L + L2G31.3240.2334.6343.7338.4348.22
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.608, + 0.906, + 0.636 + ], + "angle": 0, + "content": "Table 3. Results of different mask feature extraction methods on the val split of three datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.907, + 0.809 + ], + "angle": 0, + "content": "of the two. In Tab. 3, we present experimental results of this fusion strategy, employing the local and global feature extraction method from “G + L[41]”, which denotes the weighted sum of G(tok_mask) and cropped L, while “G + L” refers to the weighted sum of G(blur+att_mask) with L. We report both the results with the optimal weight that yielded the highest performance. Our results demonstrate that, due to the effectiveness of our global feature extraction strategy in generating a more accurate global representation, this weighted sum fusion significantly outperforms the previous method." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.907, + 0.901 + ], + "angle": 0, + "content": "To fully exploit the attention mechanism in the model and facilitate a comprehensive interaction between the global and local features, we propose generating a hybrid feature by fusing features from the later layers of one branch's image encoder into the other branch. In Tab. 3, we present the results of different hybrid fusion strategies." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "29640" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.108, + 0.089, + 0.465, + 0.21 + ], + "angle": 0, + "content": "
Start LayerG2LL2G
oIoUmIoUoIoUmIoU
722.6134.7829.1239.33
838.1647.8234.6743.62
939.1248.6437.6147.17
1036.5046.4737.3846.81
1134.1242.2633.2844.06
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.221, + 0.483, + 0.251 + ], + "angle": 0, + "content": "Table 4. Results of the starting layer \\( l \\) for hybrid fusion of global and local branches on the val split of RefCOCOg dataset." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.261, + 0.483, + 0.562 + ], + "angle": 0, + "content": "The \"L2G\" strategy, where local branch features are incorporated into the global branch, shows a significant improvement over \"G + L\". This highlights the effectiveness of our hybrid feature extraction method. On the other hand, the \"G2L\" strategy, where global features are fused into the local branch, provides the best results overall. This indicates that local features, which are crucial for fine-grained object identification, benefit most from the additional global context, allowing the model to achieve more accurate and contextually informed segmentations. The key difference between \"G2L\" and \"L2G\" is that the global branch applies an attention mask, while the local branch does not. Global features require a token mask for integration into the local branch, whereas local features can be directly fused into the global branch, resulting in distinct \"G2L\" and \"L2G\" outputs. We also explore the fusion of both \"L2G\" and \"G2L\" strategies using a weighted sum operation, but no further improvements are observed. Consequently, we adopt the \"G2L\" strategy for hybrid feature extraction in our implementation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.565, + 0.483, + 0.716 + ], + "angle": 0, + "content": "Impact of Starting Layer for Hybrid Fusion We further investigate the effect of starting the hybrid fusion at different layers of the image encoder on the RefCOCOg dataset. In Tab. 4, we present the results of applying the fusion at various layers to determine which starting layer yields the best performance. Our findings show that initiating fusion at the 9-th layer yields the best performance, suggesting that representations at this layer provide the most meaningful and contextually relevant information for effective interaction between global and local features." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.728, + 0.345, + 0.743 + ], + "angle": 0, + "content": "4.4.2. Ablation on Spatial Guidance" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.748, + 0.483, + 0.824 + ], + "angle": 0, + "content": "Building upon the proposed hybrid global-local feature extraction method, we further conduct ablation studies to assess the impact of different components in the spatial guidance augmentation approach. Our experiments primarily focus on two datasets: RefCOCO and RefCOCOg." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Impact of Spatial Relationship Guidance From Tab. 5, we can see that incorporating spatial relationship guidance (Rel) leads to consistent performance improvements across all datasets. However, the improvements are more limited on RefCOCOg, where referring expressions typically con" + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.09, + 0.9, + 0.196 + ], + "angle": 0, + "content": "
G2LRelGcoGposRefCOCORefCOCOg
oIoUmIoUoIoUmIoU
✓31.7140.2739.1248.64
✓✓35.6844.2939.6848.99
✓✓✓35.6844.2939.7349.02
✓✓✓✓41.8149.4842.4751.25
" + }, + { + "type": "table_caption", + "bbox": [ + 0.539, + 0.207, + 0.88, + 0.221 + ], + "angle": 0, + "content": "Table 5. Ablation study on the different spatial guidance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.231, + 0.905, + 0.367 + ], + "angle": 0, + "content": "tain fewer spatial relationships. The most significant improvement is seen on RefCOCO, where both mIoU and oIoU increase by approximately \\(4\\%\\). This is attributed to the richer spatial relationship descriptions in the referring expressions, which are better captured by the introduced guidance. Overall, these results highlight the effectiveness of spatial relationship guidance in enhancing referring expression comprehension, especially in datasets with complex and detailed spatial descriptions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.368, + 0.907, + 0.61 + ], + "angle": 0, + "content": "Impact of Spatial Coherence & Position Guidance Spatial coherence guidance \\(\\mathbf{G}_{\\mathrm{co}}\\) is introduced to enhance segmentation coherence and mitigate the issue of partial masks being selected for the target object. As a result, improvements in oIoU and mIoU are modest, as only a small subset of data face this issue, which may not be fully reflected in the overall scores. However, when combined with spatial position guidance \\(\\mathbf{G}_{\\mathrm{pos}}\\) together, as shown in Tab. 5, we observe a significant improvement in spatial position awareness, leading to more accurate target localization and enhanced segmentation performance. In referring expressions, spatial positional cues such as \"the left [object]\" often specify the target object's position without additional information. By introducing \\(\\mathbf{G}_{\\mathrm{pos}}\\), which explicitly encodes spatial positioning information, we improve the alignment between the referring text and the mask prediction." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.625, + 0.634, + 0.641 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.651, + 0.905, + 0.816 + ], + "angle": 0, + "content": "This paper presents a novel, training-free approach to zero-shot referring image segmentation (RIS), addressing challenges in aligning visual masks with referring expressions. Leveraging CLIP and SAM, our hybrid global-local feature extraction method combines mask-specific detail with contextual information to improve mask representation. Further, a spatial guidance augmentation strategy enhances spatial coherence and reduces ambiguities, effectively aligning masks with referring text. Experiments on RefCOCO, RefCOCO+, and RefCOCOg demonstrate that our method achieves substantial gains over zero-shot RIS models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.817, + 0.905, + 0.9 + ], + "angle": 0, + "content": "Acknowledgements. This work was supported in part by the National Natural Science Foundation of China (No.62106201, No.62376217), and in part by the Guangdong Basic and Applied Basic Research Foundation (No.2025A1515011501). The authors thank the anonymous reviewers for their helpful feedback." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "29641" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.122, + 0.188, + 0.137 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.148, + 0.483, + 0.217 + ], + "angle": 0, + "content": "[1] Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.219, + 0.483, + 0.288 + ], + "angle": 0, + "content": "[2] Walid Bousselham, Felix Petersen, Vittorio Ferrari, and Hilde Kuehne. Grounding everything: Emerging localization properties in vision-language transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3828-3837, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.291, + 0.483, + 0.358 + ], + "angle": 0, + "content": "[3] Yong Xien Chng, Henry Zheng, Yizeng Han, Xuchong Qiu, and Gao Huang. Mask grounding for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26573-26583, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.362, + 0.482, + 0.429 + ], + "angle": 0, + "content": "[4] Qiyuan Dai and Sibei Yang. Curriculum point prompting for weakly-supervised referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13711-13722, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.433, + 0.483, + 0.514 + ], + "angle": 0, + "content": "[5] Ruining Deng, Can Cui, Quan Liu, Tianyuan Yao, Lucas W Remedios, Shunxing Bao, Bennett A Landman, Lee E Weiless, Lori A Coburn, Keith T Wilson, et al. Segment anything model (sam) for digital pathology: Assess zero-shot segmentation on whole slide imaging. arXiv preprint arXiv:2304.04155, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.518, + 0.482, + 0.559 + ], + "angle": 0, + "content": "[6] Sina Hajimiri, Ismail Ben Ayed, and Jose Dolz. Pay attention to your neighbours: Training-free open-vocabulary semantic segmentation. arXiv preprint arXiv:2404.08181, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.562, + 0.482, + 0.63 + ], + "angle": 0, + "content": "[7] Ronghang Hu, Marcus Rohrbach, and Trevor Darrell. Segmentation from natural language expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 108-124. Springer, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.633, + 0.482, + 0.701 + ], + "angle": 0, + "content": "[8] Ziling Huang and Shin'ichi Satoh. Referring image segmentation via joint mask contextual embedding learning and progressive alignment network. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 7753-7762, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.704, + 0.482, + 0.786 + ], + "angle": 0, + "content": "[9] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4015-4026, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[10] Jungbeom Lee, Sungjin Lee, Jinseok Nam, Seunghak Yu, Jaeyoung Do, and Tara Taghavi. Weakly supervised referring image segmentation with intra-chunk and inter-chunk consistency. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21870–21881, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[11] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.148, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.094, + 0.905, + 0.134 + ], + "angle": 0, + "content": "frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[12] Jiachen Li, Qing Xie, Xiaojun Chang, Jinyu Xu, and Yongjian Liu. Mutually-guided hierarchical multi-modal feature learning for referring image segmentation. ACM Trans. Multimedia Comput. Commun. Appl., 2024. Just Accepted. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.208, + 0.905, + 0.29 + ], + "angle": 0, + "content": "[13] Tianxiao Li, Junhong Chen, Yiheng Huang, Kesi Huang, Qiqiang Xia, Muhammad Asim, and Wenyin Liu. Smvt: Spectrum-driven multi-scale vision transformer for referring image segmentation. In Advanced Intelligent Computing Technology and Applications, pages 193-206, Singapore, 2024. Springer Nature Singapore. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.292, + 0.905, + 0.345 + ], + "angle": 0, + "content": "[14] Yi Li, Hualiang Wang, Yiqun Duan, and Xiaomeng Li. Clip surgery for better explainability with enhancement in open-vocabulary tasks. arXiv preprint arXiv:2304.05653, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.349, + 0.905, + 0.432 + ], + "angle": 0, + "content": "[15] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.434, + 0.905, + 0.515 + ], + "angle": 0, + "content": "[16] Sihan Liu, Yiwei Ma, Xiaqing Zhang, Haowei Wang, Jiayi Ji, Xiaoshuai Sun, and Rongrong Ji. Rotated multi-scale interaction network for referring remote sensing image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26658-26668, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.519, + 0.905, + 0.572 + ], + "angle": 0, + "content": "[17] Yang Liu, Muzhi Zhu, Hengtao Li, Hao Chen, Xinlong Wang, and Chunhua Shen. Matcher: Segment anything with one shot using all-purpose feature matching. arXiv preprint arXiv:2305.13310, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.576, + 0.905, + 0.617 + ], + "angle": 0, + "content": "[18] Jun Ma, Yuting He, Feifei Li, Lin Han, Chenyu You, and Bo Wang. Segment anything in medical images. Nature Communications, 15(1):654, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.62, + 0.905, + 0.688 + ], + "angle": 0, + "content": "[19] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.69, + 0.905, + 0.757 + ], + "angle": 0, + "content": "[20] Sayan Nag, Koustava Goswami, and Srikrishna Karanam. Safari: Adaptive sequence transformer for weakly supervised referring expression segmentation. In European Conference on Computer Vision, pages 485-503. Springer, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.761, + 0.905, + 0.842 + ], + "angle": 0, + "content": "[21] Varun K Nagaraja, Vlad I Morariu, and Larry S Davis. Modeling context between objects for referring expression understanding. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 792-807. Springer, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[22] Minheng Ni, Yabo Zhang, Kailai Feng, Xiaoming Li, Yiwen Guo, and Wangmeng Zuo. Ref-diff: Zero-shot referring image segmentation with generative models. arXiv preprint arXiv:2308.16777, 2023. 1, 2, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.094, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29642" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.175 + ], + "angle": 0, + "content": "[23] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.178, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[24] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.318 + ], + "angle": 0, + "content": "[25] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.482, + 0.401 + ], + "angle": 0, + "content": "[26] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 618-626, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.404, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[27] Nisarg A Shah, Vibashan VS, and Vishal M Patel. Lqmformer: Language-aware query mask transformer for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12903-12913, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.482, + 0.545 + ], + "angle": 0, + "content": "[28] Chao Shang, Zichen Song, Heqian Qiu, Lanxiao Wang, Fanman Meng, and Hongliang Li. Prompt-driven referring image segmentation with instance contrasting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4124-4134, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.547, + 0.482, + 0.587 + ], + "angle": 0, + "content": "[29] Robin Strudel, Ivan Laptev, and Cordelia Schmid. Weakly-supervised segmentation of referring expressions. arXiv preprint arXiv:2205.04725, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.589, + 0.482, + 0.672 + ], + "angle": 0, + "content": "[30] Sanjay Subramanian, William Merrill, Trevor Darrell, Matt Gardner, Sameer Singh, and Anna Rohrbach. Reclip: A strong zero-shot baseline for referring expression comprehension. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 5198-5215, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.675, + 0.482, + 0.743 + ], + "angle": 0, + "content": "[31] Shuyang Sun, Runjia Li, Philip Torr, Xiuye Gu, and Siyang Li. Clip as rnn: Segment countless visual concepts without training endeavor. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13171-13182, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.482, + 0.801 + ], + "angle": 0, + "content": "[32] Yucheng Suo, Linchao Zhu, and Yi Yang. Text augmented spatial aware zero-shot referring image segmentation. In The 2023 Conference on Empirical Methods in Natural Language Processing, 2023. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[33] Shijie Wang, Dahun Kim, Ali Taalimi, Chen Sun, and Weicheng Kuo. Learning visual grounding from generative vision and language model. arXiv preprint arXiv:2407.14563, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[34] Wenxuan Wang, Tongtian Yue, Yisi Zhang, Longteng Guo, Xingjian He, Xinlong Wang, and Jing Liu. Unveiling parts beyond objects: Towards finer-granularity referring expres" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "sion segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12998-13008, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[35] Xinlong Wang, Zhiding Yu, Shalini De Mello, Jan Kautz, Anima Anandkumar, Chunhua Shen, and Jose M Alvarez. Freesolo: Learning to segment objects without annotations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14176-14186, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.274 + ], + "angle": 0, + "content": "[36] Chenyun Wu, Zhe Lin, Scott Cohen, Trung Bui, and Subhransu Maji. Phrasescut: Language-based image segmentation in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10216-10225, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.905, + 0.331 + ], + "angle": 0, + "content": "[37] Defeng Xie, Ruichen Wang, Jian Ma, Chen Chen, Haonan Lu, Dong Yang, Fobo Shi, and Xiaodong Lin. Edit everything: A text-guided generative system for images editing. arXiv preprint arXiv:2304.14006, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.333, + 0.905, + 0.401 + ], + "angle": 0, + "content": "[38] Li Xu, Mark He Huang, Xindi Shang, Zehuan Yuan, Ying Sun, and Jun Liu. Meta compositional referring expression segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19478-19487, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.404, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[39] Jinyu Yang, Mingqi Gao, Zhe Li, Shang Gao, Fangjing Wang, and Feng Zheng. Track anything: Segment anything meets videos. arXiv preprint arXiv:2304.11968, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.447, + 0.905, + 0.501 + ], + "angle": 0, + "content": "[40] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.503, + 0.905, + 0.572 + ], + "angle": 0, + "content": "[41] Seonghoon Yu, Paul Hongsuck Seo, and Jeany Son. Zero-shot referring image segmentation with global-local context features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19456-19465, 2023. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.574, + 0.905, + 0.628 + ], + "angle": 0, + "content": "[42] Seonghoon Yu, Paul Hongsuck Seo, and Jeany Son. Pseudoris: Distinctive pseudo-supervision generation for referring image segmentation. In Proceedings of the European Conference on Computer Vision, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.63, + 0.905, + 0.672 + ], + "angle": 0, + "content": "[43] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In European Conference on Computer Vision, pages 696-712. Springer, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.673, + 0.905, + 0.728 + ], + "angle": 0, + "content": "[44] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Wang, Lijuan Wang, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.728 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29643" + } + ] +] \ No newline at end of file diff --git a/2025/Hybrid Global-Local Representation with Augmented Spatial Guidance for Zero-Shot Referring Image Segmentation/9c88ce4f-9647-4320-bdd2-f6fc8a88e98e_origin.pdf b/2025/Hybrid Global-Local Representation with Augmented Spatial Guidance for Zero-Shot Referring Image Segmentation/9c88ce4f-9647-4320-bdd2-f6fc8a88e98e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6f0cd0e32d1f14c4b208a4af2c9bcd5f3f85cfa2 --- /dev/null +++ b/2025/Hybrid Global-Local Representation with Augmented Spatial Guidance for Zero-Shot Referring Image Segmentation/9c88ce4f-9647-4320-bdd2-f6fc8a88e98e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e6bd44bb0fc0575b3d9f357d71f5fc18f91ea57ed146aa3b310392f937ea3538 +size 3844214 diff --git a/2025/Hybrid Global-Local Representation with Augmented Spatial Guidance for Zero-Shot Referring Image Segmentation/full.md b/2025/Hybrid Global-Local Representation with Augmented Spatial Guidance for Zero-Shot Referring Image Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b6125e89f670257dbecf38a2ce1daaf00245e0e7 --- /dev/null +++ b/2025/Hybrid Global-Local Representation with Augmented Spatial Guidance for Zero-Shot Referring Image Segmentation/full.md @@ -0,0 +1,334 @@ +# Hybrid Global-Local Representation with Augmented Spatial Guidance for Zero-Shot Referring Image Segmentation + +Ting Liu $^{1,2*}$ Siyuan Li $^{1}$ + +1ASGO, School of Computer Science, Northwestern Polytechnical University +2Shenzhen Research Institute of Northwestern Polytechnical University + +liuting@nwpu.edu.cn, lisiyuan@mail.nwpu.edu.cn + +# Abstract + +Recent advances in zero-shot referring image segmentation (RIS), driven by models such as the Segment Anything Model (SAM) and CLIP, have made substantial progress in aligning visual and textual information. Despite these successes, the extraction of precise and high-quality mask region representations remains a critical challenge, limiting the full potential of RIS tasks. In this paper, we introduce a training-free, hybrid global-local feature extraction approach that integrates detailed mask-specific features with contextual information from the surrounding area, enhancing mask region representation. To further strengthen alignment between mask regions and referring expressions, we propose a spatial guidance augmentation strategy that improves spatial coherence, which is essential for accurately localizing described areas. By incorporating multiple spatial cues, this approach facilitates more robust and precise referring segmentation. Extensive experiments on standard RIS benchmarks demonstrate that our method significantly outperforms existing zero-shot RIS models, achieving substantial performance gains. We believe our approach advances RIS tasks and establishes a versatile framework for region-text alignment, offering broader implications for cross-modal understanding and interaction. Code is available at https://github.com/fhgyuanshen/HybridGL. + +# 1. Introduction + +Referring image segmentation (RIS) is a critical task in computer vision, where the goal is to segment a specific object or region in an image based on a natural language expression. This task is essential for applications such as visual search, robot perception, and human-computer interaction. Recent advancements, particularly models like the Segment Anything Model (SAM) [9] and CLIP [23], have significantly advanced zero-shot RIS, enabling object seg + +![](images/663573138a87d23a5298fa7f333a3728dd2f00267c942f211ca9a554f60b2d70.jpg) +Figure 1. Common issues in existing methods: 1) Inaccurate mask feature extraction; 2) Incorrect spatial localization; 3) Incomplete segmentation. + +mentation without the need for labeled data or task-specific training. In typical RIS pipelines, SAM generates a set of mask proposals, and CLIP extracts visual features for each mask region. For each mask, a similarity score is computed with the referring text, and the mask with the highest score is selected as the final prediction. + +Despite the promising performance of existing methods, the accurate extraction of mask representations that align well with referring text remains underexplored. Most existing approaches either mask out areas outside the mask region or crop the image to focus solely on the mask region before feeding it into CLIP, which primarily concentrates on local features while often neglecting the essential surrounding context. Some works [22, 32, 41] have sought to introduce global surrounding context features. However, such methods remain simplistic and fail to fully capture the complex interplay between the mask and its context, limiting their ability to accurately match referring expressions with the correct mask regions, as shown in Fig. 1. + +In this paper, we propose a novel hybrid global-local feature extraction approach to enhance mask region feature extraction, achieving a more precise and contextually rich mask representation without any additional training. Our approach seamlessly integrates local and global features, capturing both region-specific and context-aware information for each mask. Specifically, we design two complementary branches within CLIP to extract local region-specific visual features and broader context-aware visual features for each mask. Features from the global branch + +are progressively fused into the local branch to generate a hybrid feature representation. This hybrid fusion allows the visual encoder to automatically capture and interact with the complementary information from both branches, yielding a more contextually enriched mask feature. + +Another challenge in referring semantic segmentation lies in the use of spatial relationships within referring expressions (e.g., "left of", "bottom of") to describe objects, as shown in Fig. 1. These spatial cues introduce complexity, as they require both the recognition of object locations and the relationship between them. Capturing and aligning these spatial descriptions with visual features is inherently difficult. While referring text offers rich contextual information about spatial relationships, effectively integrating this context with visual data remains a significant challenge. Without a mechanism to explicitly model and align this spatial information, accurately matching the textual description to the correct visual region becomes problematic. Moreover, directly computing the similarity between the extracted visual mask features and text features can lead to ambiguity in the segmentation, as the mask features may capture a mixture of information from multiple objects or regions. It could mistakenly segment only part of the target region, rather than the full object. To alleviate these limitations, we introduce an augmentation approach by introducing several spatial guidance including spatial relationships, coherence, and positional cues. + +Our experiments show that the proposed method significantly outperforms several zero-shot baselines and weakly-supervised referring segmentation methods, achieving significant accuracy improvements. The proposed framework offers a powerful, efficient approach to zero-shot referring image segmentation, with strong potential for practical deployment in real-world scenarios. The contributions of this paper can be summarized as follows: + +- We propose an innovative hybrid global-local feature extraction approach for RIS, enhancing mask region representation without additional training requirements. +- We introduce a spatial guidance augmentation strategy that leverages spatial relationships, coherence, and positional cues to mitigate segmentation ambiguity. +- Extensive experiments on the four public datasets, demonstrate that our method significantly outperforms existing state-of-the-art zero-shot semantic segmentation approaches. + +# 2. Related Work + +Referring Image Segmentation. Referring image segmentation is a visual grounding task that requires the model to understand a natural language expression describing a specific object within an image and accurately segment that object from the rest of the scene [7]. The goal of this task is to bridge the gap between visual and textual modalities, en + +abling more sophisticated interactions with visual content. Fully supervised methods [3, 8, 12, 13, 16, 27, 28, 38] have achieved impressive performance in this area by effectively integrating information from both images and text descriptions. These methods rely on large datasets with detailed annotations, where each object mentioned in the text is precisely segmented in the corresponding image [33, 34]. Currently, some weakly supervised methods [4, 10, 20, 29] can learn and perform segmentation using a smaller amount of labeled data. However, the requirement for such detailed and expensive human-annotated data limits the scalability and applicability of supervised approaches. + +Foundation Model in Image Segmentation. Recent research has shown that segmentation knowledge can emerge from pre-trained foundation models (FMs) [1] such as CLIP and Stable Diffusion [25]. Though the standard CLIP model excels at recognizing object appearances, it falls short in grasping their exact locations. However, MaskCLIP [43] demonstrates that it is possible to adapt CLIP for segmentation purposes by making minor adjustments to its attention-pooling mechanism. Other works delve into various strategies for refining attention mechanisms [2, 6, 14], enhancing the model's localization capability. Specifically, SAM has shown promising capabilities in object segmentation in many works [5, 17, 18, 24, 37, 39, 44]. Overall, these findings suggest that FMs can serve as a valuable resource for zero-shot segmentation and related tasks. + +Zero-shot Referring Image Segmentation. To address the limitations of fully supervised methods, zero-shot referring image segmentation has gained significant attention. Zero-shot approaches aim to perform segmentation without requiring any labeled data for the target objects, making them highly flexible and scalable. One notable method in this domain is Global-Local [41], which leverages pretrained models like FreeSOLO [35] and CLIP to achieve zero-shot segmentation. Another approach, CaR [31], further enhances this process by recurrently applying CLIP to refine the segmentation mask iteratively. The introduction of Segment Anything Model (SAM) has marked a significant milestone in zero-shot referring image segmentation. For example, Ref-Diff [22] utilizes diffusion models to better understand the relationship between image and text pairs, leading to more accurate and context-aware segmentations. TAS [32] uses a captioner, BLIP2 [11], to provide additional context and enhance the segmentation results by generating descriptive captions for the images. PseudorIS [42] modifies the Global-Local's pipeline and incorporates a captioner like CoCa [40] for unsupervised training, leveraging the rich knowledge of current foundation models to improve segmentation accuracy. + +Note that although we also adopt both local and global features, our approach differs significantly from Global-Local [41] in terms of both the methods for local and global + +![](images/a2e22ab99cccd241570d1b57fdb16c3831b7f84f204a5aeb6f89c95b638e9913.jpg) +Figure 2. The proposed framework combines hybrid global-local feature extraction with multiple spatial guidance mechanisms to improve zero-shot referring image segmentation, using mask proposals generated by SAM. By leveraging both broad context and local details, and enhancing segmentation with spatial guidance, the framework effectively augments the segmentation of target based on textual descriptions. + +feature extraction and the fusion strategy of these features. Distinct from previous methods, our work introduces a novel, training-free approach to extracting mask representations using a hybrid feature extraction scheme. Furthermore, we incorporate multiple spatial guidance mechanisms to enhance semantic coherence and spatial alignment, improving the referring segmentation process. + +# 3. Method + +# 3.1. Overview + +Given an input image $I \in \mathbb{R}^{H \times W \times 3}$ , where $H$ and $W$ are the height and width of the image, respectively, and a referring text $T$ , which describes the target object or region, the goal is to predict the segmentation mask $m^* \in \{0,1\}^{H \times W}$ that corresponds to the region in $I$ described by $t$ . We first adopt SAM (Segment Anything Model) [9] to generate a set of mask proposals $M = \{m_1, m_2, \ldots, m_n\}$ , where each $m_i$ is a binary mask highlighting a potential segment region in $I$ . + +For each mask $m_{i}\in M$ , we use the CLIP visual encoder to extract our proposed hybrid global-local features $\mathbf{x}_i\in \mathbb{R}^d$ . To match the referring expression with the visual features, we encode the referring expression $t$ using the CLIP text encoder $\phi_{\mathrm{text}}(\cdot)$ to obtain the textual feature $\mathbf{f}_t\in \mathbb{R}^d$ , adopting same strategy with [41]. Subsequently, we can compute the cosine similarity between $\mathbf{f}_t$ and each feature $\mathbf{x}_i$ , obtaining the semantic alignment score $\mathbf{S}_{m_i}^s$ for each mask $m_{i}$ . Furthermore, we introduce a spatial guidance augmentation approach. By leveraging several spatial cues, we augment the scoring mask process to prioritize semantically and spatially aligned masks. Our method in + +tegrates several spatial cues, specifically spatial coherence guidance, spatial position guidance, and spatial relationships guidance, to enhance the mask scoring process. The overall framework is shown in Fig. 2. + +# 3.2. Hybrid Global-Local Feature Extraction + +To extract a more precise mask feature for mask $m_{i}$ , we propose the Hybrid Global-Local Feature Extraction method. This approach captures both region-specific and context-aware features, significantly enhancing mask feature extraction. The method consists of two branches: a local branch for region-specific feature extraction and a global branch for context-aware feature extraction. Features from the global branch are progressively fused into the local branch to generate a hybrid feature representation. This hybrid fusion allows the model to automatically capture the complementary information from the two branches. + +Local Feature Extraction. For the local region-specific feature extraction, we apply the binary mask $m$ to the input image $I$ , creating a masked image $I_{\mathrm{local}} = I \cdot m_i$ . This ensures that only the relevant region contributes to feature extraction. The masked image $I_{\mathrm{local}}$ is then processed through the CLIP image encoder to obtain the region-specific features. Unlike most existing methods, we do not crop the mask region, ensuring that the input image remains well-aligned with the input of the global branch. This allows the naturally aligned features to be hybridized and deeply fused for enhanced feature extraction. + +Global Feature Extraction. For the context-aware global feature extraction, we use the original image $I$ and apply a Gaussian blur to the non-mask regions, thereby emphasizing the relevant areas while preserving essential contextual + +information. To further direct the model's focus onto the mask region, we introduce an attention mask $\mathbf{A}_{m_i}$ during the self-attention operation in the last $l$ layers of the CLIP transformer-based architecture as shown in Fig. 2. Considering the CLS token is embedded into the shared visual-textual space as the visual feature, we nullify the attention scores between the CLS token and image tokens outside the masked region, effectively preventing non-relevant areas from contributing to the attention on the CLS token. + +Hybrid Feature Extraction. To enhance the fusion of local and global features, we propose a hybrid fusion approach that enables the attention mechanism in the network to effectively and automatically capture the complementary information between these two features. + +Given that local features primarily focus on the mask region, which typically contains the most relevant information for the task, the global branch is intended to complement the local branch by providing a broader context. While the global branch captures a wider spatial range and provides essential contextual information beyond the mask, it may also introduce irrelevant information from regions outside the target. To maximize the benefits of both local and global features, it is crucial to selectively integrate them, ensuring that the global context enhances the local representation without diluting its focus on the target region. + +To achieve this, we apply a token mask to the global features during the fusion. Specifically, we mask out image tokens outside the mask region to ensure that only the relevant tokens contribute to the feature extraction process. The fusion of local and global features at layer $l$ is performed as follows: + +$$ +\mathbf {x} _ {\text {h y b r i d}} ^ {(l)} = \mathbf {x} _ {\text {l o c a l}} ^ {(l - 1)} + \beta \cdot \left(\mathbf {x} _ {\text {g l o b a l}} ^ {(l - 1)} \cdot \mathbf {B} _ {m _ {i}}\right), \tag {1} +$$ + +where $\mathbf{x}_{\mathrm{local}}^{(l - 1)}$ and $\mathbf{x}_{\mathrm{global}}^{(l - 1)}$ represent the local and global features derived from the $l - 1$ -th layer, respectively. The hyper-parameter $\beta$ is the relative contribution of the global feature to the fusion process. To ensure that only relevant tokens contribute to the feature fusion, we apply a mask $\mathbf{B}_{m_i} \in \{0,1\}^K$ on the global feature, where each feature contains $K$ image token features, effectively excluding tokens corresponding to regions outside the mask. The resulting hybrid feature $\mathbf{x}_{\mathrm{hybrid}}^{(l - 1)}$ is then passed as input to the $l$ -th layer of the CLIP image encoder, denoted $\phi_{\mathrm{visual}}^l$ , to produce a feature representation that integrates both global and local information: + +$$ +\mathbf {x} _ {\text {l o c a l}} ^ {(l)} = \phi_ {\text {v i s u a l}} ^ {l} \left(\mathbf {x} _ {\text {h y b r i d}} ^ {(l)}\right). \tag {2} +$$ + +The local feature $\mathbf{x}_{\mathrm{local}}^{(L)}$ obtained from the last encoder layer $L$ , which effectively combines localized and global contextual information, serves as the final hybrid feature for the mask region. + +Semantic Alignment Score. We then extract hybrid visual features $\mathbf{X}_{\mathrm{hybrid}} = \{\mathbf{x}_1,\mathbf{x}_2,\dots ,\mathbf{x}_n\}$ for each mask, where + +each $\mathbf{x}_i$ corresponds to a mask proposal $m_{i}$ .After extracting a referring text feature $\mathbf{f}_t$ from the CLIP text encoder, we calculate the cosine semantic alignment score between the referring text and each $\mathbf{x}_i$ .. + +$$ +\mathbf {S} _ {m _ {i}} ^ {s} = \cos \left(\phi_ {\text {t e x t}} (t), x _ {i}\right). \tag {3} +$$ + +# 3.3. Spatial Guidance Augmentation + +We combine multiple spatial guidance mechanisms, including relationships, coherence, and position, which are detailed in the following. + +Spatial Relationship Guidance. To utilize the spatial relationships described in the referring text, we first parse the text into objects and spatial relations, following a method similar to [30]. For instance, objects might include phrases like "the pizza" or "man" while spatial relations describe the positioning between objects, such as "right". + +With the extracted hybrid mask features $x_{i}$ for each mask $m_{i}$ , we compute the semantic alignment score $\mathcal{P}(p, m_{i}) = \cos(\phi_{\text{text}}(p), x_{i})$ , which measures how well the mask satisfies a given object $p$ (e.g., "the pizza"). We then select the top $k$ masks based on this score and apply the softmax function to normalize these selected scores, ensuring that they sum to one. This strategy prioritizes the most relevant and semantically aligned mask proposals while also enhancing computational efficiency for subsequent operations. + +To model the spatial relationships, a spatial relation function $\mathcal{R}(p,q)$ quantified the relationship between two parsed objects $p$ and $q$ is defined as: + +$$ +\mathcal {R} \left(\left(p, m _ {i}\right), \left(q, m _ {j}\right)\right) = \left\{ \begin{array}{l l} 1 & \text {i f} p \text {s a t i s f i e s t h e r e l a t i o n w i t h} q, \\ 0 & \text {o t h e r w i s e}, \end{array} \right. \tag {4} +$$ + +where the spatial relation between $p$ and $q$ can be any of the defined relations like "left", "right", "top", "bottom", "within", "smaller", or "bigger". The satisfaction of the spatial relation is computed based on the position and size of the corresponding masks $m_{i}$ and $m_{j}$ for $p$ and $q$ . + +Finally, we combine the probability of each object with the spatial relation probabilities to identify the target object. The overall likelihood of mask $m_{i}$ being the target $p$ is given by: + +$$ +\mathbf {S} _ {m _ {i}} ^ {s} = \sum_ {m _ {j}} \mathcal {P} (p, m _ {i}) \cdot \mathcal {R} ((p, m _ {i}), (q, m _ {j})) \cdot \mathcal {P} (q, m _ {j}), \tag {5} +$$ + +where $\mathbf{S}_{m_i}^s$ is the final probability of $m_{i}$ being the correct target, combining both its predicate satisfaction and its spatial relationships with other objects. If the referring text contains those spatial relations, the score of each mask is computed accordingly. Otherwise, $\mathbf{S}_{m_i}^s$ is directly computed using Equation 3. + +Spatial Coherence Guidance. To enhance spatial coherence, we generate a spatial localization guidance $\mathbf{G}_{\mathrm{co}} \in$ + +$[0,1]^{H\times W}$ by using the referring text feature $(t)$ to identify the target region. Specifically, we apply the algorithm proposed in [2], which leverages self-attention mechanisms for expression localization and segmentation. This guidance map is constructed by calculating the similarity between $\phi_{\mathrm{text}}(t)$ and each visual token embedding derived from the CLIP image encoder with the self-attention mechanisms, yielding a localization map that broadly highlights areas corresponding to the target object or region. Consequently, positions with values closer to 1 represent higher similarity, thereby indicating regions more likely aligned with the target described by the text feature. + +Spatial Position Guidance. To incorporate spatial positions that align with the referring expression context, similar with [22], we introduce a set of positional guidance matrices $\mathbf{G}_{\mathrm{pos}} \in [0,1]^{H \times W}$ , each representing a position within the image. Here, $pos \in \{top, bottom, left, right, middle\}$ denotes specific spatial attributes, such as "top" for upper regions. $\mathbf{G}_{\mathrm{pos}}$ is computed by: + +$$ +\mathbf {G} _ {\text {p o s}} = \left\{ \begin{array}{l l} \mathcal {E} (\text {p o s}) & \text {i f p o s i t i o n p o s a p p e r a s i n} t, \\ \mathbf {1} _ {H \times W} & \text {i f n o p o s i t i o n p o s a p p e r a s i n} t, \end{array} \right. \tag {6} +$$ + +where $\mathcal{E}(\mathrm{pos})$ represents the emphasis function that returns a value between 0 and 1 based on the position's emphasis. To generate these matrices, we compute the distance of each pixel to the corresponding region defined by pos, normalize the distances to the range [0, 1], and directly assign these values to the positional guidance matrix. This structure reflects a gradual transition in spatial emphasis across the image, enabling the model to focus on regions by the referring expression's context. + +In this manner, we define $\mathbf{G}_{\mathrm{pos}}$ for each desired direction. Once the pos appears in the referring text, we incorporate this spatial guidance into our spatial coherence guidance $\mathbf{G}_{\mathrm{co}}$ by performing an element-wise multiplication. For each direction dir, the spatial guidance matrix $\mathbf{G}$ is defined as: + +$$ +\mathbf {G} = \mathbf {G} _ {\mathrm {c o}} \odot \mathbf {G} _ {\mathrm {p o s}}. \tag {7} +$$ + +Spatial Guidance Score. The spatial guidance score $\mathbf{S}_{m_i}^g$ for each mask proposal $m_{i}$ is computed as the difference between the mean spatial guidance values within the mask. Specifically, we define the spatial guidance score as: + +$$ +\mathbf {S} _ {m _ {i}} ^ {g} = \frac {\operatorname {S u m} (\mathbf {G} \odot m _ {i})}{\operatorname {S u m} (m _ {i})} - \lambda \cdot \frac {\operatorname {S u m} (\mathbf {G} \odot (1 - m _ {i}))}{\operatorname {S u m} (1 - m _ {i})}, \tag {8} +$$ + +where $\lambda$ is a hyperparameter that controls the importance of the negative score. This score $\mathbf{S}_{m_i}^g$ serves as a spatial coherence measure, with larger values indicating better spatial alignment between the mask proposal and the referring expression. + +Finally, we apply softmax normalization to the semantic alignment score $\mathbf{S}_{m_i}^s$ and spatial guidance score $\mathbf{S}_{m_i}^g$ for all + +masks, and fuse them together. The final score for each mask $m_{i}$ is obtained by: + +$$ +\mathbf {S} _ {m _ {i}} = (1 - \alpha) \mathbf {S} _ {m _ {i}} ^ {s} + \alpha \mathbf {S} _ {m _ {i}} ^ {g}, \tag {9} +$$ + +where $\alpha$ controls the trade-off between semantic alignment and spatial guidance. The final referring semantic segmentation result $m^{*}$ is generated by selecting the mask with the highest mask score. + +This comprehensive approach allows us to prioritize mask proposals that are not only semantically aligned with the text feature but also spatially consistent with the specified direction. Thus, masks are refined based on both semantic and spatial coherence, enhancing the accuracy of the selected mask with respect to the referring expression's spatial context. + +# 4. Experiments + +# 4.1. Datasets and Metrics + +To evaluate the proposed method, we use several main RIS datasets: RefCOCO [21], RefCOCO+ [21], and RefCOCOg [19]. These datasets are derived from the MSCOCO [15] dataset and feature images annotated with detailed referring expressions that pinpoint specific objects or regions. Each dataset brings unique aspects to the referring expressions it contains. RefCOCO often includes positional information like "left" and "right," which is banned in RefCOCO+, and RefCOCOg has more elaborate sentence structures. We also evaluate it on PhraseCut [36] dataset, which introduces structured textual descriptions that detail attributes, categories, and relationships among objects. We employ two primary metrics to evaluate the performance: overall Intersection over Union (oIoU) and mean Intersection over Union (mIoU). oIoU assesses the total overlap between predicted and ground truth regions relative to their combined area, making it particularly stringent for inaccuracies in larger segments. On the other hand, mIoU calculates the average overlap for each individual instance, ensuring a balanced consideration of performance across objects of varying sizes. Together, these metrics provide a robust framework for assessing the effectiveness of the proposed method in RIS tasks. + +# 4.2. Implementation Details + +The experiments were run on a single NVIDIA RTX 3090 GPU. Following previous work [22, 32], we use the default ViT-H SAM model, and the hyperparameters "predicted iou threshold" and "stability score threshold" were set to 0.7, the "points per side" was set to 8. We use CLIP with ViTB/16 backbone in both hybrid feature extraction and spatial coherence guidance. We set $\alpha$ and $\beta$ to 0.6 and 2, respectively, for all datasets. The hyperparameter $\lambda$ is empirically set to 9 to balance the propensity of both excessively large + +
MetricMethodVision BackbonePre-trained ModelRefCOCORefCOCO+RefCOCOg
valtestAtestBvaltestAtestBvaltest
oloUzero-shot methods w/ additional training
Pseudo-RIS [42]ViT-BSAM, CoCa, CLIP37.3343.4331.9040.1946.4333.6341.6343.52
VLM-VG [33]R101COCO*, VLM-VG*45.4048.0041.4037.0040.7030.5042.8044.10
zero-shot methods w/o additional training
Grad-CAM [26]R50SAM, CLIP23.4423.9121.6026.6727.2024.8423.0023.91
MaskCLIP [43]R50SAM, CLIP20.1820.5221.3022.0622.4324.6123.0523.41
Global-Local [41]R50FreeSOLO, CLIP24.5823.3824.3525.8724.6125.6130.0729.83
Global-Local [41]R50SAM, CLIP24.5526.0021.0326.6229.9922.2328.9230.48
Global-Local [41]ViT-BSAM, CLIP21.7124.4820.5123.7028.1221.8626.5728.21
Ref-Diff [22]ViT-BSAM, SD, CLIP35.1637.4434.5035.5638.6631.4038.6237.50
TAS [32]ViT-BSAM, BLIP2, CLIP29.5330.2628.2433.2138.7728.0135.8436.16
OursViT-BSAM,CLIP41.8144.5238.535.7441.4330.942.4742.97
mIoUweakly-supervised methods
CLRL [10]ViT-B-31.0632.3030.1131.2832.1130.1332.88-
PPT [4]ViT-BSAM46.7645.3346.2845.3445.8444.7742.97-
zero-shot methods w/ additional training
Pseudo-RIS [42]ViT-BSAM, CoCa, CLIP41.0548.1933.4844.3351.4235.0845.9946.67
VLM-VG [33]R101COCO*, VLM-VG*49.9053.1046.7042.7047.3036.2048.0048.50
zero-shot methods w/o additional training
Grad-CAM [26]R50SAM, CLIP30.2231.9027.1733.9625.6632.2933.0532.50
MaskCLIP [43]R50SAM, CLIP25.6226.6625.1727.4928.4930.4730.1330.15
Global-Local [41]R50FreeSOLO, CLIP26.7024.9926.4828.2226.5427.8633.0233.12
Global-Local [41]R50SAM, CLIP31.8332.9328.6434.9737.1130.6140.6640.94
Global-Local [41]ViT-BSAM, CLIP33.1236.5229.5835.2939.5831.8940.0840.74
CaR [31]ViT-B and ViT-LCLIP33.5735.3630.5134.2236.0331.0236.6736.57
Ref-Diff [22]ViT-BSAM, SD, CLIP37.2138.4037.1937.2940.5133.0144.0244.51
TAS [32]ViT-BSAM, BLIP2, CLIP39.8441.0836.2443.6349.1336.5446.6246.80
OursViT-BSAM, CLIP49.4853.3745.1943.4049.1337.1751.2551.59
+ +and small mask regions receiving disproportionately high scores. When the referring text includes "big" or "small", $\lambda$ is adjusted to 3 or 14, respectively. + +# 4.3. Results + +In Tab. 1, we evaluate our model on RefCOCO, RefCOCO+, and RefCOCOg and compare it with other state-of-the-art (SOTA) zero-shot models. Here, we evaluate the MaskCLIP, Grad-CAM methods by computing the similarity between their feature maps and SAM's mask proposals. Since the original Global-Local method used FreeSOLO as the mask extractor, we have re-evaluated the performance of Global-Local using SAM as the mask extractor, across different backbones. We also provide results of Ref-Diff, and TAS using different backbones. Additionally, we include weakly-supervised methods and zero-shot methods with extra training or additional datasets. Our method achieves excellent results on all three datasets. In terms of oIoU, our method improves by $4\% -7\%$ over SOTA methods on RefCOCO and RefCOCOg, and although it does not achieve the best performance on the testB set of RefCOCO+, it is only $0.5\%$ lower than the SOTA method. For mIoU, our method outperforms SOTA methods by $4\% -10\%$ on + +RefCOCO and RefCOCOg, and it also achieves comparable performance on RefCOCO+. Importantly, our method achieves comparable or even higher mIoU on all three datasets compared to methods that use additional training. This indicates that our method has a more balanced consideration across varying objects. In addition, we present the oIoU and mIoU results on the PhraseCut dataset's test set in Tab. 2. Our method demonstrates superior performance compared to previous methods, achieving the highest average score. + +Table 1. Comparisons with the SOTA zero-shot approaches on RefCOCO, RefCOCO+, and RefCOCOg datasets. The best two results under the same setting, w/o additional training, are highlighted in bold and underlined, respectively. * indicates the extra dataset used to train the model. + +
MethodsoIoUmIoUavg.
Global-Local[41]23.64-23.64
TAS[32]25.6424.6625.15
Ref-Diff[22]29.4241.7535.59
Ours38.3936.9837.69
+ +Table 2. Comparison with existing methods on the PhraseCut dataset. + +Notably, our approach relies solely on SAM and CLIP, while TAS additionally incorporates the large BLIP2 model, and Ref-Diff leverages the Stable-Diffusion model; yet, our method still outperforms these more complex models. Be + +![](images/1277ce1a277c109ba3b566aa1966d2be186089acf32cf30c8b45bbf86bdeb108.jpg) +Referring text : The woman in the middle sitting on the couch + +![](images/f5ace7e5698fa3a39b60d8b8840f8ff975943e43f44e74062d902d212d502530.jpg) + +![](images/830633bea87253211cbe328d63df5cd3ea1cc8e9ab3a9858a5be510752d49064.jpg) + +![](images/7b5f8fafda932b3da87cfd324016d04292534821c997d78c85db06ac7bd20dac.jpg) + +![](images/3e19d8aff5ba5f5b18b9dae1849f7acaa6ecae15510770c12a9639641eeb1ed1.jpg) + +![](images/b77d261737c50977f865da24abc425a948dc663ff9f6afce6451a18c65056558.jpg) + +![](images/7a38693728058a60cbceb8f5fcac0d3023b7b1cc6e1dbe9cb5f024880d7d138c.jpg) +Referring text : A beige surfboard being carried by a man in a wetsuit + +![](images/eee53ef79b8e3305441cfa9b64afec0e4f90098c407804bbdc048062d74aa4a3.jpg) +Referring text : The chair behind the guy wearing the stripes + +![](images/4cee302decba7ab9189780f2885f63a5b63fd102baee62f37b72aa65d57161ff.jpg) + +![](images/28e0d10764d9ab28c68cb10b705b6febd7f963662663d65390b4681aa14914c8.jpg) + +![](images/469f846a77783f3c31ebfc0e8aae034d0af00f3f8e76ffd73d4206a3e59e0273.jpg) + +![](images/31a66926f07aad2c85bdcd0168a9f3044e9c84f1efcb3ffcddc02a8876c6d48e.jpg) + +![](images/1f5914738e6fbe1fccd79f0d00656339a0b5d56e38428bfba5346cac8d372001.jpg) +Image + +![](images/9e63f670fb8d9503fafe1aad3a8352c603d5ba34e707c57c9f068c2abdf96138.jpg) +SAM-CLIP + +![](images/0c338692d9ddcd6c0cdbfb4d2219da8964095f9724342ad53dc73a2e37424a1b.jpg) +Global-Local +Figure 3. Visual comparisons with existing methods. Our approach achieves more accurate localization and a complete segmentation of the target object. + +![](images/428996b93833e65d31ceb1825e798a462f8302daa5b8b2819d8aefd20b646b83.jpg) +TAS + +![](images/e691375c1beddf71d678894fc5e3d1362717e0231db25f3a4a187724500d1f2f.jpg) +Ours + +![](images/a450e5f0eeb9c824b197f6ba298cd2b0e409d72a3be190172c8882b87947ee17.jpg) +GT + +sides, TAS and Ref-Diff show better performance in either mIoU or oIoU, whereas our method consistently outperforms both metrics. We present visual comparisons with existing methods in Fig. 3, highlighting the improvements in segmentation accuracy and the ability to capture spatial relationships. + +# 4.4. Ablation Study + +To evaluate the effectiveness of different strategies, we conduct extensive ablation studies on the val dataset of RefCOCO, RefCOCO+, and RefCOCOg datasets. We compare a range of feature extraction methods and spatial guidance strategies, examining their impact on performance. + +# 4.4.1. Ablation on Hybrid Feature Extraction + +Local and Global Features. We first evaluate the performance of local and global features, including different strategies for global feature extraction. These strategies involve blurring images to reduce the influence of irrelevant regions on the global representation, applying a token mask to exclude non-mask regions, and using an attention mask to minimize the impact of non-mask regions on the CLS token. All these different mask strategies are applied to the last few layers of the CLIP image encoder. As shown in Tab. 3, the local features (L) consistently outperform the global methods across all datasets in terms of both oIoU and mIoU. Notably, combining a blur operation with the attention mask (G(blur+att_mask)) results in a significant improvement, approaching the performance of local features on the RefCOCO+ and RefCOCOg datasets. Hence, we adopt this approach to extract global features. + +Hybrid Global-Local Features. The intuitive way to fuse the global and local features is to compute a weighted sum + +
MethodRefCOCO oIoU mIoURefCOCO+ oIoU mIoURefCOCOg oIoU mIoU
L24.6532.6428.1636.4433.6342.35
G(blur)16.1722.0118.6125.1422.0031.24
G(tok_mask)20.5531.7121.8433.1819.8232.55
G(att_mask)20.0530.0721.4431.5124.1136.24
G(blur+att_mask)27.6337.1229.7539.9433.0344.45
G+L [41]21.7133.1223.7035.2926.5740.08
G+L28.9037.6432.2741.1238.0847.51
L2G29.9339.1733.1742.5537.7147.20
G2L31.7140.2734.4443.4039.1248.64
G2L + L2G31.3240.2334.6343.7338.4348.22
+ +Table 3. Results of different mask feature extraction methods on the val split of three datasets. + +of the two. In Tab. 3, we present experimental results of this fusion strategy, employing the local and global feature extraction method from “G + L[41]”, which denotes the weighted sum of G(tok_mask) and cropped L, while “G + L” refers to the weighted sum of G(blur+att_mask) with L. We report both the results with the optimal weight that yielded the highest performance. Our results demonstrate that, due to the effectiveness of our global feature extraction strategy in generating a more accurate global representation, this weighted sum fusion significantly outperforms the previous method. + +To fully exploit the attention mechanism in the model and facilitate a comprehensive interaction between the global and local features, we propose generating a hybrid feature by fusing features from the later layers of one branch's image encoder into the other branch. In Tab. 3, we present the results of different hybrid fusion strategies. + +
Start LayerG2LL2G
oIoUmIoUoIoUmIoU
722.6134.7829.1239.33
838.1647.8234.6743.62
939.1248.6437.6147.17
1036.5046.4737.3846.81
1134.1242.2633.2844.06
+ +The "L2G" strategy, where local branch features are incorporated into the global branch, shows a significant improvement over "G + L". This highlights the effectiveness of our hybrid feature extraction method. On the other hand, the "G2L" strategy, where global features are fused into the local branch, provides the best results overall. This indicates that local features, which are crucial for fine-grained object identification, benefit most from the additional global context, allowing the model to achieve more accurate and contextually informed segmentations. The key difference between "G2L" and "L2G" is that the global branch applies an attention mask, while the local branch does not. Global features require a token mask for integration into the local branch, whereas local features can be directly fused into the global branch, resulting in distinct "G2L" and "L2G" outputs. We also explore the fusion of both "L2G" and "G2L" strategies using a weighted sum operation, but no further improvements are observed. Consequently, we adopt the "G2L" strategy for hybrid feature extraction in our implementation. + +Impact of Starting Layer for Hybrid Fusion We further investigate the effect of starting the hybrid fusion at different layers of the image encoder on the RefCOCOg dataset. In Tab. 4, we present the results of applying the fusion at various layers to determine which starting layer yields the best performance. Our findings show that initiating fusion at the 9-th layer yields the best performance, suggesting that representations at this layer provide the most meaningful and contextually relevant information for effective interaction between global and local features. + +# 4.4.2. Ablation on Spatial Guidance + +Building upon the proposed hybrid global-local feature extraction method, we further conduct ablation studies to assess the impact of different components in the spatial guidance augmentation approach. Our experiments primarily focus on two datasets: RefCOCO and RefCOCOg. + +Impact of Spatial Relationship Guidance From Tab. 5, we can see that incorporating spatial relationship guidance (Rel) leads to consistent performance improvements across all datasets. However, the improvements are more limited on RefCOCOg, where referring expressions typically con + +Table 4. Results of the starting layer $l$ for hybrid fusion of global and local branches on the val split of RefCOCOg dataset. + +
G2LRelGcoGposRefCOCORefCOCOg
oIoUmIoUoIoUmIoU
✓31.7140.2739.1248.64
✓✓35.6844.2939.6848.99
✓✓✓35.6844.2939.7349.02
✓✓✓✓41.8149.4842.4751.25
+ +Table 5. Ablation study on the different spatial guidance. + +tain fewer spatial relationships. The most significant improvement is seen on RefCOCO, where both mIoU and oIoU increase by approximately $4\%$ . This is attributed to the richer spatial relationship descriptions in the referring expressions, which are better captured by the introduced guidance. Overall, these results highlight the effectiveness of spatial relationship guidance in enhancing referring expression comprehension, especially in datasets with complex and detailed spatial descriptions. + +Impact of Spatial Coherence & Position Guidance Spatial coherence guidance $\mathbf{G}_{\mathrm{co}}$ is introduced to enhance segmentation coherence and mitigate the issue of partial masks being selected for the target object. As a result, improvements in oIoU and mIoU are modest, as only a small subset of data face this issue, which may not be fully reflected in the overall scores. However, when combined with spatial position guidance $\mathbf{G}_{\mathrm{pos}}$ together, as shown in Tab. 5, we observe a significant improvement in spatial position awareness, leading to more accurate target localization and enhanced segmentation performance. In referring expressions, spatial positional cues such as "the left [object]" often specify the target object's position without additional information. By introducing $\mathbf{G}_{\mathrm{pos}}$ , which explicitly encodes spatial positioning information, we improve the alignment between the referring text and the mask prediction. + +# 5. Conclusion + +This paper presents a novel, training-free approach to zero-shot referring image segmentation (RIS), addressing challenges in aligning visual masks with referring expressions. Leveraging CLIP and SAM, our hybrid global-local feature extraction method combines mask-specific detail with contextual information to improve mask representation. Further, a spatial guidance augmentation strategy enhances spatial coherence and reduces ambiguities, effectively aligning masks with referring text. Experiments on RefCOCO, RefCOCO+, and RefCOCOg demonstrate that our method achieves substantial gains over zero-shot RIS models. + +Acknowledgements. This work was supported in part by the National Natural Science Foundation of China (No.62106201, No.62376217), and in part by the Guangdong Basic and Applied Basic Research Foundation (No.2025A1515011501). The authors thank the anonymous reviewers for their helpful feedback. + +# References + +[1] Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021. 2 +[2] Walid Bousselham, Felix Petersen, Vittorio Ferrari, and Hilde Kuehne. Grounding everything: Emerging localization properties in vision-language transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3828-3837, 2024. 2, 5 +[3] Yong Xien Chng, Henry Zheng, Yizeng Han, Xuchong Qiu, and Gao Huang. Mask grounding for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26573-26583, 2024. 2 +[4] Qiyuan Dai and Sibei Yang. Curriculum point prompting for weakly-supervised referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13711-13722, 2024. 2, 6 +[5] Ruining Deng, Can Cui, Quan Liu, Tianyuan Yao, Lucas W Remedios, Shunxing Bao, Bennett A Landman, Lee E Weiless, Lori A Coburn, Keith T Wilson, et al. Segment anything model (sam) for digital pathology: Assess zero-shot segmentation on whole slide imaging. arXiv preprint arXiv:2304.04155, 2023. 2 +[6] Sina Hajimiri, Ismail Ben Ayed, and Jose Dolz. Pay attention to your neighbours: Training-free open-vocabulary semantic segmentation. arXiv preprint arXiv:2404.08181, 2024. 2 +[7] Ronghang Hu, Marcus Rohrbach, and Trevor Darrell. Segmentation from natural language expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 108-124. Springer, 2016. 2 +[8] Ziling Huang and Shin'ichi Satoh. Referring image segmentation via joint mask contextual embedding learning and progressive alignment network. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 7753-7762, 2023. 2 +[9] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4015-4026, 2023. 1, 3 +[10] Jungbeom Lee, Sungjin Lee, Jinseok Nam, Seunghak Yu, Jaeyoung Do, and Tara Taghavi. Weakly supervised referring image segmentation with intra-chunk and inter-chunk consistency. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21870–21881, 2023. 2, 6 +[11] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with + +frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 2 +[12] Jiachen Li, Qing Xie, Xiaojun Chang, Jinyu Xu, and Yongjian Liu. Mutually-guided hierarchical multi-modal feature learning for referring image segmentation. ACM Trans. Multimedia Comput. Commun. Appl., 2024. Just Accepted. 2 +[13] Tianxiao Li, Junhong Chen, Yiheng Huang, Kesi Huang, Qiqiang Xia, Muhammad Asim, and Wenyin Liu. Smvt: Spectrum-driven multi-scale vision transformer for referring image segmentation. In Advanced Intelligent Computing Technology and Applications, pages 193-206, Singapore, 2024. Springer Nature Singapore. 2 +[14] Yi Li, Hualiang Wang, Yiqun Duan, and Xiaomeng Li. Clip surgery for better explainability with enhancement in open-vocabulary tasks. arXiv preprint arXiv:2304.05653, 2023. 2 +[15] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5 +[16] Sihan Liu, Yiwei Ma, Xiaqing Zhang, Haowei Wang, Jiayi Ji, Xiaoshuai Sun, and Rongrong Ji. Rotated multi-scale interaction network for referring remote sensing image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26658-26668, 2024. 2 +[17] Yang Liu, Muzhi Zhu, Hengtao Li, Hao Chen, Xinlong Wang, and Chunhua Shen. Matcher: Segment anything with one shot using all-purpose feature matching. arXiv preprint arXiv:2305.13310, 2023. 2 +[18] Jun Ma, Yuting He, Feifei Li, Lin Han, Chenyu You, and Bo Wang. Segment anything in medical images. Nature Communications, 15(1):654, 2024. 2 +[19] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20, 2016. 5 +[20] Sayan Nag, Koustava Goswami, and Srikrishna Karanam. Safari: Adaptive sequence transformer for weakly supervised referring expression segmentation. In European Conference on Computer Vision, pages 485-503. Springer, 2025. 2 +[21] Varun K Nagaraja, Vlad I Morariu, and Larry S Davis. Modeling context between objects for referring expression understanding. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 792-807. Springer, 2016. 5 +[22] Minheng Ni, Yabo Zhang, Kailai Feng, Xiaoming Li, Yiwen Guo, and Wangmeng Zuo. Ref-diff: Zero-shot referring image segmentation with generative models. arXiv preprint arXiv:2308.16777, 2023. 1, 2, 5, 6 + +[23] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1 +[24] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 2 +[25] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2 +[26] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 618-626, 2017. 6 +[27] Nisarg A Shah, Vibashan VS, and Vishal M Patel. Lqmformer: Language-aware query mask transformer for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12903-12913, 2024. 2 +[28] Chao Shang, Zichen Song, Heqian Qiu, Lanxiao Wang, Fanman Meng, and Hongliang Li. Prompt-driven referring image segmentation with instance contrasting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4124-4134, 2024. 2 +[29] Robin Strudel, Ivan Laptev, and Cordelia Schmid. Weakly-supervised segmentation of referring expressions. arXiv preprint arXiv:2205.04725, 2022. 2 +[30] Sanjay Subramanian, William Merrill, Trevor Darrell, Matt Gardner, Sameer Singh, and Anna Rohrbach. Reclip: A strong zero-shot baseline for referring expression comprehension. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 5198-5215, 2022. 4 +[31] Shuyang Sun, Runjia Li, Philip Torr, Xiuye Gu, and Siyang Li. Clip as rnn: Segment countless visual concepts without training endeavor. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13171-13182, 2024. 2, 6 +[32] Yucheng Suo, Linchao Zhu, and Yi Yang. Text augmented spatial aware zero-shot referring image segmentation. In The 2023 Conference on Empirical Methods in Natural Language Processing, 2023. 1, 2, 5, 6 +[33] Shijie Wang, Dahun Kim, Ali Taalimi, Chen Sun, and Weicheng Kuo. Learning visual grounding from generative vision and language model. arXiv preprint arXiv:2407.14563, 2024. 2, 6 +[34] Wenxuan Wang, Tongtian Yue, Yisi Zhang, Longteng Guo, Xingjian He, Xinlong Wang, and Jing Liu. Unveiling parts beyond objects: Towards finer-granularity referring expres + +sion segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12998-13008, 2024. 2 +[35] Xinlong Wang, Zhiding Yu, Shalini De Mello, Jan Kautz, Anima Anandkumar, Chunhua Shen, and Jose M Alvarez. Freesolo: Learning to segment objects without annotations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14176-14186, 2022. 2 +[36] Chenyun Wu, Zhe Lin, Scott Cohen, Trung Bui, and Subhransu Maji. Phrasescut: Language-based image segmentation in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10216-10225, 2020. 5 +[37] Defeng Xie, Ruichen Wang, Jian Ma, Chen Chen, Haonan Lu, Dong Yang, Fobo Shi, and Xiaodong Lin. Edit everything: A text-guided generative system for images editing. arXiv preprint arXiv:2304.14006, 2023. 2 +[38] Li Xu, Mark He Huang, Xindi Shang, Zehuan Yuan, Ying Sun, and Jun Liu. Meta compositional referring expression segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19478-19487, 2023. 2 +[39] Jinyu Yang, Mingqi Gao, Zhe Li, Shang Gao, Fangjing Wang, and Feng Zheng. Track anything: Segment anything meets videos. arXiv preprint arXiv:2304.11968, 2023. 2 +[40] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022. 2 +[41] Seonghoon Yu, Paul Hongsuck Seo, and Jeany Son. Zero-shot referring image segmentation with global-local context features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19456-19465, 2023. 1, 2, 3, 6, 7 +[42] Seonghoon Yu, Paul Hongsuck Seo, and Jeany Son. Pseudoris: Distinctive pseudo-supervision generation for referring image segmentation. In Proceedings of the European Conference on Computer Vision, 2024. 2, 6 +[43] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In European Conference on Computer Vision, pages 696-712. Springer, 2022. 2, 6 +[44] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Wang, Lijuan Wang, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. Advances in Neural Information Processing Systems, 36, 2024. 2 \ No newline at end of file diff --git a/2025/Hybrid Global-Local Representation with Augmented Spatial Guidance for Zero-Shot Referring Image Segmentation/images.zip b/2025/Hybrid Global-Local Representation with Augmented Spatial Guidance for Zero-Shot Referring Image Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8acb8300318682f66ce4adedacb0797905a242f6 --- /dev/null +++ b/2025/Hybrid Global-Local Representation with Augmented Spatial Guidance for Zero-Shot Referring Image Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:29b9767c29580659ed20b4efcbd41cd7e344d0641dd1391eae4c11fc4e3bd3be +size 709609 diff --git a/2025/Hybrid Global-Local Representation with Augmented Spatial Guidance for Zero-Shot Referring Image Segmentation/layout.json b/2025/Hybrid Global-Local Representation with Augmented Spatial Guidance for Zero-Shot Referring Image Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..cde3f132c1fad4fc05a651f5b473ea708fa2e4a7 --- /dev/null +++ b/2025/Hybrid Global-Local Representation with Augmented Spatial Guidance for Zero-Shot Referring Image Segmentation/layout.json @@ -0,0 +1,8977 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 74, + 102, + 538, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 102, + 538, + 140 + ], + "spans": [ + { + "bbox": [ + 74, + 102, + 538, + 140 + ], + "type": "text", + "content": "Hybrid Global-Local Representation with Augmented Spatial Guidance for Zero-Shot Referring Image Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 242, + 160, + 365, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 160, + 365, + 175 + ], + "spans": [ + { + "bbox": [ + 242, + 160, + 365, + 175 + ], + "type": "text", + "content": "Ting Liu" + }, + { + "bbox": [ + 242, + 160, + 365, + 175 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 242, + 160, + 365, + 175 + ], + "type": "text", + "content": " Siyuan Li" + }, + { + "bbox": [ + 242, + 160, + 365, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 117, + 175, + 492, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 175, + 492, + 203 + ], + "spans": [ + { + "bbox": [ + 117, + 175, + 492, + 203 + ], + "type": "text", + "content": "1ASGO, School of Computer Science, Northwestern Polytechnical University \n2Shenzhen Research Institute of Northwestern Polytechnical University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 175, + 205, + 426, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 205, + 426, + 217 + ], + "spans": [ + { + "bbox": [ + 175, + 205, + 426, + 217 + ], + "type": "text", + "content": "liuting@nwpu.edu.cn, lisiyuan@mail.nwpu.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 152, + 245, + 200, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 245, + 200, + 257 + ], + "spans": [ + { + "bbox": [ + 152, + 245, + 200, + 257 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 270, + 296, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 296, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 296, + 558 + ], + "type": "text", + "content": "Recent advances in zero-shot referring image segmentation (RIS), driven by models such as the Segment Anything Model (SAM) and CLIP, have made substantial progress in aligning visual and textual information. Despite these successes, the extraction of precise and high-quality mask region representations remains a critical challenge, limiting the full potential of RIS tasks. In this paper, we introduce a training-free, hybrid global-local feature extraction approach that integrates detailed mask-specific features with contextual information from the surrounding area, enhancing mask region representation. To further strengthen alignment between mask regions and referring expressions, we propose a spatial guidance augmentation strategy that improves spatial coherence, which is essential for accurately localizing described areas. By incorporating multiple spatial cues, this approach facilitates more robust and precise referring segmentation. Extensive experiments on standard RIS benchmarks demonstrate that our method significantly outperforms existing zero-shot RIS models, achieving substantial performance gains. We believe our approach advances RIS tasks and establishes a versatile framework for region-text alignment, offering broader implications for cross-modal understanding and interaction. Code is available at https://github.com/fhgyuanshen/HybridGL." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 579, + 135, + 592 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 135, + 592 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 135, + 592 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 599, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 599, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 599, + 295, + 696 + ], + "type": "text", + "content": "Referring image segmentation (RIS) is a critical task in computer vision, where the goal is to segment a specific object or region in an image based on a natural language expression. This task is essential for applications such as visual search, robot perception, and human-computer interaction. Recent advancements, particularly models like the Segment Anything Model (SAM) [9] and CLIP [23], have significantly advanced zero-shot RIS, enabling object seg" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 245, + 554, + 309 + ], + "blocks": [ + { + "bbox": [ + 317, + 245, + 554, + 309 + ], + "lines": [ + { + "bbox": [ + 317, + 245, + 554, + 309 + ], + "spans": [ + { + "bbox": [ + 317, + 245, + 554, + 309 + ], + "type": "image", + "image_path": "663573138a87d23a5298fa7f333a3728dd2f00267c942f211ca9a554f60b2d70.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 321, + 555, + 354 + ], + "lines": [ + { + "bbox": [ + 313, + 321, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 321, + 555, + 354 + ], + "type": "text", + "content": "Figure 1. Common issues in existing methods: 1) Inaccurate mask feature extraction; 2) Incorrect spatial localization; 3) Incomplete segmentation." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 362, + 555, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 362, + 555, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 362, + 555, + 434 + ], + "type": "text", + "content": "mentation without the need for labeled data or task-specific training. In typical RIS pipelines, SAM generates a set of mask proposals, and CLIP extracts visual features for each mask region. For each mask, a similarity score is computed with the referring text, and the mask with the highest score is selected as the final prediction." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 436, + 556, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 436, + 556, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 436, + 556, + 592 + ], + "type": "text", + "content": "Despite the promising performance of existing methods, the accurate extraction of mask representations that align well with referring text remains underexplored. Most existing approaches either mask out areas outside the mask region or crop the image to focus solely on the mask region before feeding it into CLIP, which primarily concentrates on local features while often neglecting the essential surrounding context. Some works [22, 32, 41] have sought to introduce global surrounding context features. However, such methods remain simplistic and fail to fully capture the complex interplay between the mask and its context, limiting their ability to accurately match referring expressions with the correct mask regions, as shown in Fig. 1." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 594, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 557, + 714 + ], + "type": "text", + "content": "In this paper, we propose a novel hybrid global-local feature extraction approach to enhance mask region feature extraction, achieving a more precise and contextually rich mask representation without any additional training. Our approach seamlessly integrates local and global features, capturing both region-specific and context-aware information for each mask. Specifically, we design two complementary branches within CLIP to extract local region-specific visual features and broader context-aware visual features for each mask. Features from the global branch" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 192, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 192, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 192, + 713 + ], + "type": "text", + "content": "*Ting Liu is the corresponding author." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29634" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "type": "text", + "content": "are progressively fused into the local branch to generate a hybrid feature representation. This hybrid fusion allows the visual encoder to automatically capture and interact with the complementary information from both branches, yielding a more contextually enriched mask feature." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 133, + 295, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 133, + 295, + 393 + ], + "spans": [ + { + "bbox": [ + 56, + 133, + 295, + 393 + ], + "type": "text", + "content": "Another challenge in referring semantic segmentation lies in the use of spatial relationships within referring expressions (e.g., \"left of\", \"bottom of\") to describe objects, as shown in Fig. 1. These spatial cues introduce complexity, as they require both the recognition of object locations and the relationship between them. Capturing and aligning these spatial descriptions with visual features is inherently difficult. While referring text offers rich contextual information about spatial relationships, effectively integrating this context with visual data remains a significant challenge. Without a mechanism to explicitly model and align this spatial information, accurately matching the textual description to the correct visual region becomes problematic. Moreover, directly computing the similarity between the extracted visual mask features and text features can lead to ambiguity in the segmentation, as the mask features may capture a mixture of information from multiple objects or regions. It could mistakenly segment only part of the target region, rather than the full object. To alleviate these limitations, we introduce an augmentation approach by introducing several spatial guidance including spatial relationships, coherence, and positional cues." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 396, + 294, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 294, + 490 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 294, + 490 + ], + "type": "text", + "content": "Our experiments show that the proposed method significantly outperforms several zero-shot baselines and weakly-supervised referring segmentation methods, achieving significant accuracy improvements. The proposed framework offers a powerful, efficient approach to zero-shot referring image segmentation, with strong potential for practical deployment in real-world scenarios. The contributions of this paper can be summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 491, + 295, + 610 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 56, + 491, + 295, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 491, + 295, + 526 + ], + "spans": [ + { + "bbox": [ + 56, + 491, + 295, + 526 + ], + "type": "text", + "content": "- We propose an innovative hybrid global-local feature extraction approach for RIS, enhancing mask region representation without additional training requirements." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 527, + 295, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 527, + 295, + 562 + ], + "spans": [ + { + "bbox": [ + 56, + 527, + 295, + 562 + ], + "type": "text", + "content": "- We introduce a spatial guidance augmentation strategy that leverages spatial relationships, coherence, and positional cues to mitigate segmentation ambiguity." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 563, + 295, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 563, + 295, + 610 + ], + "spans": [ + { + "bbox": [ + 56, + 563, + 295, + 610 + ], + "type": "text", + "content": "- Extensive experiments on the four public datasets, demonstrate that our method significantly outperforms existing state-of-the-art zero-shot semantic segmentation approaches." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 621, + 142, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 621, + 142, + 634 + ], + "spans": [ + { + "bbox": [ + 55, + 621, + 142, + 634 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 642, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 294, + 713 + ], + "type": "text", + "content": "Referring Image Segmentation. Referring image segmentation is a visual grounding task that requires the model to understand a natural language expression describing a specific object within an image and accurately segment that object from the rest of the scene [7]. The goal of this task is to bridge the gap between visual and textual modalities, en" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "content": "abling more sophisticated interactions with visual content. Fully supervised methods [3, 8, 12, 13, 16, 27, 28, 38] have achieved impressive performance in this area by effectively integrating information from both images and text descriptions. These methods rely on large datasets with detailed annotations, where each object mentioned in the text is precisely segmented in the corresponding image [33, 34]. Currently, some weakly supervised methods [4, 10, 20, 29] can learn and perform segmentation using a smaller amount of labeled data. However, the requirement for such detailed and expensive human-annotated data limits the scalability and applicability of supervised approaches." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 216, + 553, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 216, + 553, + 393 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 553, + 393 + ], + "type": "text", + "content": "Foundation Model in Image Segmentation. Recent research has shown that segmentation knowledge can emerge from pre-trained foundation models (FMs) [1] such as CLIP and Stable Diffusion [25]. Though the standard CLIP model excels at recognizing object appearances, it falls short in grasping their exact locations. However, MaskCLIP [43] demonstrates that it is possible to adapt CLIP for segmentation purposes by making minor adjustments to its attention-pooling mechanism. Other works delve into various strategies for refining attention mechanisms [2, 6, 14], enhancing the model's localization capability. Specifically, SAM has shown promising capabilities in object segmentation in many works [5, 17, 18, 24, 37, 39, 44]. Overall, these findings suggest that FMs can serve as a valuable resource for zero-shot segmentation and related tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 396, + 553, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 553, + 670 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 553, + 670 + ], + "type": "text", + "content": "Zero-shot Referring Image Segmentation. To address the limitations of fully supervised methods, zero-shot referring image segmentation has gained significant attention. Zero-shot approaches aim to perform segmentation without requiring any labeled data for the target objects, making them highly flexible and scalable. One notable method in this domain is Global-Local [41], which leverages pretrained models like FreeSOLO [35] and CLIP to achieve zero-shot segmentation. Another approach, CaR [31], further enhances this process by recurrently applying CLIP to refine the segmentation mask iteratively. The introduction of Segment Anything Model (SAM) has marked a significant milestone in zero-shot referring image segmentation. For example, Ref-Diff [22] utilizes diffusion models to better understand the relationship between image and text pairs, leading to more accurate and context-aware segmentations. TAS [32] uses a captioner, BLIP2 [11], to provide additional context and enhance the segmentation results by generating descriptive captions for the images. PseudorIS [42] modifies the Global-Local's pipeline and incorporates a captioner like CoCa [40] for unsupervised training, leveraging the rich knowledge of current foundation models to improve segmentation accuracy." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "Note that although we also adopt both local and global features, our approach differs significantly from Global-Local [41] in terms of both the methods for local and global" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "29635" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 83, + 71, + 531, + 285 + ], + "blocks": [ + { + "bbox": [ + 83, + 71, + 531, + 285 + ], + "lines": [ + { + "bbox": [ + 83, + 71, + 531, + 285 + ], + "spans": [ + { + "bbox": [ + 83, + 71, + 531, + 285 + ], + "type": "image", + "image_path": "a2e22ab99cccd241570d1b57fdb16c3831b7f84f204a5aeb6f89c95b638e9913.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 286, + 555, + 319 + ], + "lines": [ + { + "bbox": [ + 54, + 286, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 54, + 286, + 555, + 319 + ], + "type": "text", + "content": "Figure 2. The proposed framework combines hybrid global-local feature extraction with multiple spatial guidance mechanisms to improve zero-shot referring image segmentation, using mask proposals generated by SAM. By leveraging both broad context and local details, and enhancing segmentation with spatial guidance, the framework effectively augments the segmentation of target based on textual descriptions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 326, + 296, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 326, + 296, + 409 + ], + "spans": [ + { + "bbox": [ + 55, + 326, + 296, + 409 + ], + "type": "text", + "content": "feature extraction and the fusion strategy of these features. Distinct from previous methods, our work introduces a novel, training-free approach to extracting mask representations using a hybrid feature extraction scheme. Furthermore, we incorporate multiple spatial guidance mechanisms to enhance semantic coherence and spatial alignment, improving the referring segmentation process." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 422, + 111, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 111, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 111, + 434 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 443, + 123, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 443, + 123, + 454 + ], + "spans": [ + { + "bbox": [ + 55, + 443, + 123, + 454 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 460, + 296, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 296, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 296, + 568 + ], + "type": "text", + "content": "Given an input image " + }, + { + "bbox": [ + 55, + 460, + 296, + 568 + ], + "type": "inline_equation", + "content": "I \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 55, + 460, + 296, + 568 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 460, + 296, + 568 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 460, + 296, + 568 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 460, + 296, + 568 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 55, + 460, + 296, + 568 + ], + "type": "text", + "content": " are the height and width of the image, respectively, and a referring text " + }, + { + "bbox": [ + 55, + 460, + 296, + 568 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 460, + 296, + 568 + ], + "type": "text", + "content": ", which describes the target object or region, the goal is to predict the segmentation mask " + }, + { + "bbox": [ + 55, + 460, + 296, + 568 + ], + "type": "inline_equation", + "content": "m^* \\in \\{0,1\\}^{H \\times W}" + }, + { + "bbox": [ + 55, + 460, + 296, + 568 + ], + "type": "text", + "content": " that corresponds to the region in " + }, + { + "bbox": [ + 55, + 460, + 296, + 568 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 460, + 296, + 568 + ], + "type": "text", + "content": " described by " + }, + { + "bbox": [ + 55, + 460, + 296, + 568 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 460, + 296, + 568 + ], + "type": "text", + "content": ". We first adopt SAM (Segment Anything Model) [9] to generate a set of mask proposals " + }, + { + "bbox": [ + 55, + 460, + 296, + 568 + ], + "type": "inline_equation", + "content": "M = \\{m_1, m_2, \\ldots, m_n\\}" + }, + { + "bbox": [ + 55, + 460, + 296, + 568 + ], + "type": "text", + "content": ", where each " + }, + { + "bbox": [ + 55, + 460, + 296, + 568 + ], + "type": "inline_equation", + "content": "m_i" + }, + { + "bbox": [ + 55, + 460, + 296, + 568 + ], + "type": "text", + "content": " is a binary mask highlighting a potential segment region in " + }, + { + "bbox": [ + 55, + 460, + 296, + 568 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 460, + 296, + 568 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": "For each mask " + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "m_{i}\\in M" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": ", we use the CLIP visual encoder to extract our proposed hybrid global-local features " + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i\\in \\mathbb{R}^d" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": ". To match the referring expression with the visual features, we encode the referring expression " + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": " using the CLIP text encoder " + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\phi_{\\mathrm{text}}(\\cdot)" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": " to obtain the textual feature " + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_t\\in \\mathbb{R}^d" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": ", adopting same strategy with [41]. Subsequently, we can compute the cosine similarity between " + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_t" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": " and each feature " + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": ", obtaining the semantic alignment score " + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{m_i}^s" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": " for each mask " + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "m_{i}" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": ". Furthermore, we introduce a spatial guidance augmentation approach. By leveraging several spatial cues, we augment the scoring mask process to prioritize semantically and spatially aligned masks. Our method in" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 326, + 555, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 326, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 326, + 555, + 373 + ], + "type": "text", + "content": "tegrates several spatial cues, specifically spatial coherence guidance, spatial position guidance, and spatial relationships guidance, to enhance the mask scoring process. The overall framework is shown in Fig. 2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 383, + 527, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 383, + 527, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 383, + 527, + 396 + ], + "type": "text", + "content": "3.2. Hybrid Global-Local Feature Extraction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 401, + 555, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 555, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 555, + 532 + ], + "type": "text", + "content": "To extract a more precise mask feature for mask " + }, + { + "bbox": [ + 313, + 401, + 555, + 532 + ], + "type": "inline_equation", + "content": "m_{i}" + }, + { + "bbox": [ + 313, + 401, + 555, + 532 + ], + "type": "text", + "content": ", we propose the Hybrid Global-Local Feature Extraction method. This approach captures both region-specific and context-aware features, significantly enhancing mask feature extraction. The method consists of two branches: a local branch for region-specific feature extraction and a global branch for context-aware feature extraction. Features from the global branch are progressively fused into the local branch to generate a hybrid feature representation. This hybrid fusion allows the model to automatically capture the complementary information from the two branches." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 533, + 556, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 556, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 556, + 664 + ], + "type": "text", + "content": "Local Feature Extraction. For the local region-specific feature extraction, we apply the binary mask " + }, + { + "bbox": [ + 313, + 533, + 556, + 664 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 533, + 556, + 664 + ], + "type": "text", + "content": " to the input image " + }, + { + "bbox": [ + 313, + 533, + 556, + 664 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 533, + 556, + 664 + ], + "type": "text", + "content": ", creating a masked image " + }, + { + "bbox": [ + 313, + 533, + 556, + 664 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{local}} = I \\cdot m_i" + }, + { + "bbox": [ + 313, + 533, + 556, + 664 + ], + "type": "text", + "content": ". This ensures that only the relevant region contributes to feature extraction. The masked image " + }, + { + "bbox": [ + 313, + 533, + 556, + 664 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{local}}" + }, + { + "bbox": [ + 313, + 533, + 556, + 664 + ], + "type": "text", + "content": " is then processed through the CLIP image encoder to obtain the region-specific features. Unlike most existing methods, we do not crop the mask region, ensuring that the input image remains well-aligned with the input of the global branch. This allows the naturally aligned features to be hybridized and deeply fused for enhanced feature extraction." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "Global Feature Extraction. For the context-aware global feature extraction, we use the original image " + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": " and apply a Gaussian blur to the non-mask regions, thereby emphasizing the relevant areas while preserving essential contextual" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "29636" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 178 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 178 + ], + "type": "text", + "content": "information. To further direct the model's focus onto the mask region, we introduce an attention mask " + }, + { + "bbox": [ + 55, + 72, + 294, + 178 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{m_i}" + }, + { + "bbox": [ + 55, + 72, + 294, + 178 + ], + "type": "text", + "content": " during the self-attention operation in the last " + }, + { + "bbox": [ + 55, + 72, + 294, + 178 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 72, + 294, + 178 + ], + "type": "text", + "content": " layers of the CLIP transformer-based architecture as shown in Fig. 2. Considering the CLS token is embedded into the shared visual-textual space as the visual feature, we nullify the attention scores between the CLS token and image tokens outside the masked region, effectively preventing non-relevant areas from contributing to the attention on the CLS token." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 179, + 294, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 179, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 179, + 294, + 239 + ], + "type": "text", + "content": "Hybrid Feature Extraction. To enhance the fusion of local and global features, we propose a hybrid fusion approach that enables the attention mechanism in the network to effectively and automatically capture the complementary information between these two features." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 239, + 294, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 294, + 371 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 294, + 371 + ], + "type": "text", + "content": "Given that local features primarily focus on the mask region, which typically contains the most relevant information for the task, the global branch is intended to complement the local branch by providing a broader context. While the global branch captures a wider spatial range and provides essential contextual information beyond the mask, it may also introduce irrelevant information from regions outside the target. To maximize the benefits of both local and global features, it is crucial to selectively integrate them, ensuring that the global context enhances the local representation without diluting its focus on the target region." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 371, + 294, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 371, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 55, + 371, + 294, + 441 + ], + "type": "text", + "content": "To achieve this, we apply a token mask to the global features during the fusion. Specifically, we mask out image tokens outside the mask region to ensure that only the relevant tokens contribute to the feature extraction process. The fusion of local and global features at layer " + }, + { + "bbox": [ + 55, + 371, + 294, + 441 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 371, + 294, + 441 + ], + "type": "text", + "content": " is performed as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 94, + 445, + 294, + 466 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 445, + 294, + 466 + ], + "spans": [ + { + "bbox": [ + 94, + 445, + 294, + 466 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {\\text {h y b r i d}} ^ {(l)} = \\mathbf {x} _ {\\text {l o c a l}} ^ {(l - 1)} + \\beta \\cdot \\left(\\mathbf {x} _ {\\text {g l o b a l}} ^ {(l - 1)} \\cdot \\mathbf {B} _ {m _ {i}}\\right), \\tag {1}", + "image_path": "a8230a504a8763e352847415c8493f28f877cf34130e11d26e6a0761267db062.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 472, + 295, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 295, + 619 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 295, + 619 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 472, + 295, + 619 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{local}}^{(l - 1)}" + }, + { + "bbox": [ + 55, + 472, + 295, + 619 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 472, + 295, + 619 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{global}}^{(l - 1)}" + }, + { + "bbox": [ + 55, + 472, + 295, + 619 + ], + "type": "text", + "content": " represent the local and global features derived from the " + }, + { + "bbox": [ + 55, + 472, + 295, + 619 + ], + "type": "inline_equation", + "content": "l - 1" + }, + { + "bbox": [ + 55, + 472, + 295, + 619 + ], + "type": "text", + "content": "-th layer, respectively. The hyper-parameter " + }, + { + "bbox": [ + 55, + 472, + 295, + 619 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 472, + 295, + 619 + ], + "type": "text", + "content": " is the relative contribution of the global feature to the fusion process. To ensure that only relevant tokens contribute to the feature fusion, we apply a mask " + }, + { + "bbox": [ + 55, + 472, + 295, + 619 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{m_i} \\in \\{0,1\\}^K" + }, + { + "bbox": [ + 55, + 472, + 295, + 619 + ], + "type": "text", + "content": " on the global feature, where each feature contains " + }, + { + "bbox": [ + 55, + 472, + 295, + 619 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 472, + 295, + 619 + ], + "type": "text", + "content": " image token features, effectively excluding tokens corresponding to regions outside the mask. The resulting hybrid feature " + }, + { + "bbox": [ + 55, + 472, + 295, + 619 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{hybrid}}^{(l - 1)}" + }, + { + "bbox": [ + 55, + 472, + 295, + 619 + ], + "type": "text", + "content": " is then passed as input to the " + }, + { + "bbox": [ + 55, + 472, + 295, + 619 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 472, + 295, + 619 + ], + "type": "text", + "content": "-th layer of the CLIP image encoder, denoted " + }, + { + "bbox": [ + 55, + 472, + 295, + 619 + ], + "type": "inline_equation", + "content": "\\phi_{\\mathrm{visual}}^l" + }, + { + "bbox": [ + 55, + 472, + 295, + 619 + ], + "type": "text", + "content": ", to produce a feature representation that integrates both global and local information:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 127, + 619, + 294, + 636 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 619, + 294, + 636 + ], + "spans": [ + { + "bbox": [ + 127, + 619, + 294, + 636 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {\\text {l o c a l}} ^ {(l)} = \\phi_ {\\text {v i s u a l}} ^ {l} \\left(\\mathbf {x} _ {\\text {h y b r i d}} ^ {(l)}\\right). \\tag {2}", + "image_path": "56a6193d6025271304a48a59780513f0e615082fd624f1da0b71206a2757c999.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 640, + 294, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 640, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 640, + 294, + 689 + ], + "type": "text", + "content": "The local feature " + }, + { + "bbox": [ + 55, + 640, + 294, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{local}}^{(L)}" + }, + { + "bbox": [ + 55, + 640, + 294, + 689 + ], + "type": "text", + "content": " obtained from the last encoder layer " + }, + { + "bbox": [ + 55, + 640, + 294, + 689 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 640, + 294, + 689 + ], + "type": "text", + "content": ", which effectively combines localized and global contextual information, serves as the final hybrid feature for the mask region." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 294, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 294, + 714 + ], + "type": "text", + "content": "Semantic Alignment Score. We then extract hybrid visual features " + }, + { + "bbox": [ + 55, + 689, + 294, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\mathrm{hybrid}} = \\{\\mathbf{x}_1,\\mathbf{x}_2,\\dots ,\\mathbf{x}_n\\}" + }, + { + "bbox": [ + 55, + 689, + 294, + 714 + ], + "type": "text", + "content": " for each mask, where" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": "each " + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": " corresponds to a mask proposal " + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "inline_equation", + "content": "m_{i}" + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": " .After extracting a referring text feature " + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_t" + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": " from the CLIP text encoder, we calculate the cosine semantic alignment score between the referring text and each " + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": " .." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 384, + 129, + 553, + 143 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 129, + 553, + 143 + ], + "spans": [ + { + "bbox": [ + 384, + 129, + 553, + 143 + ], + "type": "interline_equation", + "content": "\\mathbf {S} _ {m _ {i}} ^ {s} = \\cos \\left(\\phi_ {\\text {t e x t}} (t), x _ {i}\\right). \\tag {3}", + "image_path": "470877d1bd00ef6e5e5d260827c996f8f37093422329b515d8b9b1b79897669a.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 149, + 487, + 162 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 149, + 487, + 162 + ], + "spans": [ + { + "bbox": [ + 313, + 149, + 487, + 162 + ], + "type": "text", + "content": "3.3. Spatial Guidance Augmentation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 167, + 553, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 167, + 553, + 202 + ], + "spans": [ + { + "bbox": [ + 313, + 167, + 553, + 202 + ], + "type": "text", + "content": "We combine multiple spatial guidance mechanisms, including relationships, coherence, and position, which are detailed in the following." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 203, + 553, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 203, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 313, + 203, + 553, + 274 + ], + "type": "text", + "content": "Spatial Relationship Guidance. To utilize the spatial relationships described in the referring text, we first parse the text into objects and spatial relations, following a method similar to [30]. For instance, objects might include phrases like \"the pizza\" or \"man\" while spatial relations describe the positioning between objects, such as \"right\"." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 275, + 553, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 275, + 553, + 382 + ], + "spans": [ + { + "bbox": [ + 313, + 275, + 553, + 382 + ], + "type": "text", + "content": "With the extracted hybrid mask features " + }, + { + "bbox": [ + 313, + 275, + 553, + 382 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 313, + 275, + 553, + 382 + ], + "type": "text", + "content": " for each mask " + }, + { + "bbox": [ + 313, + 275, + 553, + 382 + ], + "type": "inline_equation", + "content": "m_{i}" + }, + { + "bbox": [ + 313, + 275, + 553, + 382 + ], + "type": "text", + "content": ", we compute the semantic alignment score " + }, + { + "bbox": [ + 313, + 275, + 553, + 382 + ], + "type": "inline_equation", + "content": "\\mathcal{P}(p, m_{i}) = \\cos(\\phi_{\\text{text}}(p), x_{i})" + }, + { + "bbox": [ + 313, + 275, + 553, + 382 + ], + "type": "text", + "content": ", which measures how well the mask satisfies a given object " + }, + { + "bbox": [ + 313, + 275, + 553, + 382 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 275, + 553, + 382 + ], + "type": "text", + "content": " (e.g., \"the pizza\"). We then select the top " + }, + { + "bbox": [ + 313, + 275, + 553, + 382 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 275, + 553, + 382 + ], + "type": "text", + "content": " masks based on this score and apply the softmax function to normalize these selected scores, ensuring that they sum to one. This strategy prioritizes the most relevant and semantically aligned mask proposals while also enhancing computational efficiency for subsequent operations." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 383, + 553, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 383, + 553, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 383, + 553, + 418 + ], + "type": "text", + "content": "To model the spatial relationships, a spatial relation function " + }, + { + "bbox": [ + 313, + 383, + 553, + 418 + ], + "type": "inline_equation", + "content": "\\mathcal{R}(p,q)" + }, + { + "bbox": [ + 313, + 383, + 553, + 418 + ], + "type": "text", + "content": " quantified the relationship between two parsed objects " + }, + { + "bbox": [ + 313, + 383, + 553, + 418 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 383, + 553, + 418 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 383, + 553, + 418 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 313, + 383, + 553, + 418 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 427, + 558, + 467 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 427, + 558, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 427, + 558, + 467 + ], + "type": "interline_equation", + "content": "\\mathcal {R} \\left(\\left(p, m _ {i}\\right), \\left(q, m _ {j}\\right)\\right) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} p \\text {s a t i s f i e s t h e r e l a t i o n w i t h} q, \\\\ 0 & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {4}", + "image_path": "d704cf04164aa3ccc9aace8d2dcea8ce35a9daf91054a59548cde323a4828469.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 468, + 553, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 553, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 553, + 528 + ], + "type": "text", + "content": "where the spatial relation between " + }, + { + "bbox": [ + 313, + 468, + 553, + 528 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 468, + 553, + 528 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 468, + 553, + 528 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 313, + 468, + 553, + 528 + ], + "type": "text", + "content": " can be any of the defined relations like \"left\", \"right\", \"top\", \"bottom\", \"within\", \"smaller\", or \"bigger\". The satisfaction of the spatial relation is computed based on the position and size of the corresponding masks " + }, + { + "bbox": [ + 313, + 468, + 553, + 528 + ], + "type": "inline_equation", + "content": "m_{i}" + }, + { + "bbox": [ + 313, + 468, + 553, + 528 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 468, + 553, + 528 + ], + "type": "inline_equation", + "content": "m_{j}" + }, + { + "bbox": [ + 313, + 468, + 553, + 528 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 313, + 468, + 553, + 528 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 468, + 553, + 528 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 468, + 553, + 528 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 313, + 468, + 553, + 528 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 528, + 553, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 528, + 553, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 528, + 553, + 575 + ], + "type": "text", + "content": "Finally, we combine the probability of each object with the spatial relation probabilities to identify the target object. The overall likelihood of mask " + }, + { + "bbox": [ + 313, + 528, + 553, + 575 + ], + "type": "inline_equation", + "content": "m_{i}" + }, + { + "bbox": [ + 313, + 528, + 553, + 575 + ], + "type": "text", + "content": " being the target " + }, + { + "bbox": [ + 313, + 528, + 553, + 575 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 528, + 553, + 575 + ], + "type": "text", + "content": " is given by:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 318, + 583, + 553, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 583, + 553, + 609 + ], + "spans": [ + { + "bbox": [ + 318, + 583, + 553, + 609 + ], + "type": "interline_equation", + "content": "\\mathbf {S} _ {m _ {i}} ^ {s} = \\sum_ {m _ {j}} \\mathcal {P} (p, m _ {i}) \\cdot \\mathcal {R} ((p, m _ {i}), (q, m _ {j})) \\cdot \\mathcal {P} (q, m _ {j}), \\tag {5}", + "image_path": "365c402622064f5c2b7ec352cd25f49a4b20a8418ae87608dcabf38509cc665b.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 617, + 553, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 553, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 617, + 553, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{m_i}^s" + }, + { + "bbox": [ + 313, + 617, + 553, + 689 + ], + "type": "text", + "content": " is the final probability of " + }, + { + "bbox": [ + 313, + 617, + 553, + 689 + ], + "type": "inline_equation", + "content": "m_{i}" + }, + { + "bbox": [ + 313, + 617, + 553, + 689 + ], + "type": "text", + "content": " being the correct target, combining both its predicate satisfaction and its spatial relationships with other objects. If the referring text contains those spatial relations, the score of each mask is computed accordingly. Otherwise, " + }, + { + "bbox": [ + 313, + 617, + 553, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{m_i}^s" + }, + { + "bbox": [ + 313, + 617, + 553, + 689 + ], + "type": "text", + "content": " is directly computed using Equation 3." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "Spatial Coherence Guidance. To enhance spatial coherence, we generate a spatial localization guidance " + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{\\mathrm{co}} \\in" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29637" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 294, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 294, + 215 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 294, + 215 + ], + "type": "inline_equation", + "content": "[0,1]^{H\\times W}" + }, + { + "bbox": [ + 55, + 71, + 294, + 215 + ], + "type": "text", + "content": " by using the referring text feature " + }, + { + "bbox": [ + 55, + 71, + 294, + 215 + ], + "type": "inline_equation", + "content": "(t)" + }, + { + "bbox": [ + 55, + 71, + 294, + 215 + ], + "type": "text", + "content": " to identify the target region. Specifically, we apply the algorithm proposed in [2], which leverages self-attention mechanisms for expression localization and segmentation. This guidance map is constructed by calculating the similarity between " + }, + { + "bbox": [ + 55, + 71, + 294, + 215 + ], + "type": "inline_equation", + "content": "\\phi_{\\mathrm{text}}(t)" + }, + { + "bbox": [ + 55, + 71, + 294, + 215 + ], + "type": "text", + "content": " and each visual token embedding derived from the CLIP image encoder with the self-attention mechanisms, yielding a localization map that broadly highlights areas corresponding to the target object or region. Consequently, positions with values closer to 1 represent higher similarity, thereby indicating regions more likely aligned with the target described by the text feature." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 216, + 294, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 216, + 294, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 216, + 294, + 301 + ], + "type": "text", + "content": "Spatial Position Guidance. To incorporate spatial positions that align with the referring expression context, similar with [22], we introduce a set of positional guidance matrices " + }, + { + "bbox": [ + 55, + 216, + 294, + 301 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{\\mathrm{pos}} \\in [0,1]^{H \\times W}" + }, + { + "bbox": [ + 55, + 216, + 294, + 301 + ], + "type": "text", + "content": ", each representing a position within the image. Here, " + }, + { + "bbox": [ + 55, + 216, + 294, + 301 + ], + "type": "inline_equation", + "content": "pos \\in \\{top, bottom, left, right, middle\\}" + }, + { + "bbox": [ + 55, + 216, + 294, + 301 + ], + "type": "text", + "content": " denotes specific spatial attributes, such as \"top\" for upper regions. " + }, + { + "bbox": [ + 55, + 216, + 294, + 301 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{\\mathrm{pos}}" + }, + { + "bbox": [ + 55, + 216, + 294, + 301 + ], + "type": "text", + "content": " is computed by:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 308, + 295, + 340 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 308, + 295, + 340 + ], + "spans": [ + { + "bbox": [ + 67, + 308, + 295, + 340 + ], + "type": "interline_equation", + "content": "\\mathbf {G} _ {\\text {p o s}} = \\left\\{ \\begin{array}{l l} \\mathcal {E} (\\text {p o s}) & \\text {i f p o s i t i o n p o s a p p e r a s i n} t, \\\\ \\mathbf {1} _ {H \\times W} & \\text {i f n o p o s i t i o n p o s a p p e r a s i n} t, \\end{array} \\right. \\tag {6}", + "image_path": "d80229b7d8eacc3d482b36fb5579b8a1c7a2cc99ddfbfde2aa6f24a506a7fbd9.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 346, + 295, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 346, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 295, + 453 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 346, + 295, + 453 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(\\mathrm{pos})" + }, + { + "bbox": [ + 55, + 346, + 295, + 453 + ], + "type": "text", + "content": " represents the emphasis function that returns a value between 0 and 1 based on the position's emphasis. To generate these matrices, we compute the distance of each pixel to the corresponding region defined by pos, normalize the distances to the range [0, 1], and directly assign these values to the positional guidance matrix. This structure reflects a gradual transition in spatial emphasis across the image, enabling the model to focus on regions by the referring expression's context." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 454, + 295, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 454, + 295, + 525 + ], + "spans": [ + { + "bbox": [ + 55, + 454, + 295, + 525 + ], + "type": "text", + "content": "In this manner, we define " + }, + { + "bbox": [ + 55, + 454, + 295, + 525 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{\\mathrm{pos}}" + }, + { + "bbox": [ + 55, + 454, + 295, + 525 + ], + "type": "text", + "content": " for each desired direction. Once the pos appears in the referring text, we incorporate this spatial guidance into our spatial coherence guidance " + }, + { + "bbox": [ + 55, + 454, + 295, + 525 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{\\mathrm{co}}" + }, + { + "bbox": [ + 55, + 454, + 295, + 525 + ], + "type": "text", + "content": " by performing an element-wise multiplication. For each direction dir, the spatial guidance matrix " + }, + { + "bbox": [ + 55, + 454, + 295, + 525 + ], + "type": "inline_equation", + "content": "\\mathbf{G}" + }, + { + "bbox": [ + 55, + 454, + 295, + 525 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 525, + 294, + 539 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 525, + 294, + 539 + ], + "spans": [ + { + "bbox": [ + 137, + 525, + 294, + 539 + ], + "type": "interline_equation", + "content": "\\mathbf {G} = \\mathbf {G} _ {\\mathrm {c o}} \\odot \\mathbf {G} _ {\\mathrm {p o s}}. \\tag {7}", + "image_path": "aa5fdd272aac8a06b6a120adf4abd3a6f0ea01936c7ccecde9ad8d755c9dfcfd.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 543, + 295, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 543, + 295, + 591 + ], + "spans": [ + { + "bbox": [ + 55, + 543, + 295, + 591 + ], + "type": "text", + "content": "Spatial Guidance Score. The spatial guidance score " + }, + { + "bbox": [ + 55, + 543, + 295, + 591 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{m_i}^g" + }, + { + "bbox": [ + 55, + 543, + 295, + 591 + ], + "type": "text", + "content": " for each mask proposal " + }, + { + "bbox": [ + 55, + 543, + 295, + 591 + ], + "type": "inline_equation", + "content": "m_{i}" + }, + { + "bbox": [ + 55, + 543, + 295, + 591 + ], + "type": "text", + "content": " is computed as the difference between the mean spatial guidance values within the mask. Specifically, we define the spatial guidance score as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 64, + 597, + 295, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 597, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 64, + 597, + 295, + 624 + ], + "type": "interline_equation", + "content": "\\mathbf {S} _ {m _ {i}} ^ {g} = \\frac {\\operatorname {S u m} (\\mathbf {G} \\odot m _ {i})}{\\operatorname {S u m} (m _ {i})} - \\lambda \\cdot \\frac {\\operatorname {S u m} (\\mathbf {G} \\odot (1 - m _ {i}))}{\\operatorname {S u m} (1 - m _ {i})}, \\tag {8}", + "image_path": "f1fe52f97cf88d8b3b431d9d655238632c46a96b769727f40847c61798da6b81.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 629, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 629, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 629, + 295, + 689 + ], + "type": "text", + "content": " is a hyperparameter that controls the importance of the negative score. This score " + }, + { + "bbox": [ + 55, + 629, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{m_i}^g" + }, + { + "bbox": [ + 55, + 629, + 295, + 689 + ], + "type": "text", + "content": " serves as a spatial coherence measure, with larger values indicating better spatial alignment between the mask proposal and the referring expression." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": "Finally, we apply softmax normalization to the semantic alignment score " + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{m_i}^s" + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": " and spatial guidance score " + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{m_i}^g" + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": " for all" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "masks, and fuse them together. The final score for each mask " + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "inline_equation", + "content": "m_{i}" + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": " is obtained by:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 373, + 105, + 553, + 119 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 105, + 553, + 119 + ], + "spans": [ + { + "bbox": [ + 373, + 105, + 553, + 119 + ], + "type": "interline_equation", + "content": "\\mathbf {S} _ {m _ {i}} = (1 - \\alpha) \\mathbf {S} _ {m _ {i}} ^ {s} + \\alpha \\mathbf {S} _ {m _ {i}} ^ {g}, \\tag {9}", + "image_path": "caa638249ee3ac047d58a3512bae71d6da4702f51b644c0f854793eb3e4cdd47.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 126, + 553, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 126, + 553, + 174 + ], + "spans": [ + { + "bbox": [ + 313, + 126, + 553, + 174 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 126, + 553, + 174 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 126, + 553, + 174 + ], + "type": "text", + "content": " controls the trade-off between semantic alignment and spatial guidance. The final referring semantic segmentation result " + }, + { + "bbox": [ + 313, + 126, + 553, + 174 + ], + "type": "inline_equation", + "content": "m^{*}" + }, + { + "bbox": [ + 313, + 126, + 553, + 174 + ], + "type": "text", + "content": " is generated by selecting the mask with the highest mask score." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 175, + 554, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 175, + 554, + 258 + ], + "spans": [ + { + "bbox": [ + 313, + 175, + 554, + 258 + ], + "type": "text", + "content": "This comprehensive approach allows us to prioritize mask proposals that are not only semantically aligned with the text feature but also spatially consistent with the specified direction. Thus, masks are refined based on both semantic and spatial coherence, enhancing the accuracy of the selected mask with respect to the referring expression's spatial context." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 269, + 395, + 282 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 269, + 395, + 282 + ], + "spans": [ + { + "bbox": [ + 313, + 269, + 395, + 282 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 289, + 435, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 435, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 435, + 300 + ], + "type": "text", + "content": "4.1. Datasets and Metrics" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 306, + 555, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 306, + 555, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 306, + 555, + 582 + ], + "type": "text", + "content": "To evaluate the proposed method, we use several main RIS datasets: RefCOCO [21], RefCOCO+ [21], and RefCOCOg [19]. These datasets are derived from the MSCOCO [15] dataset and feature images annotated with detailed referring expressions that pinpoint specific objects or regions. Each dataset brings unique aspects to the referring expressions it contains. RefCOCO often includes positional information like \"left\" and \"right,\" which is banned in RefCOCO+, and RefCOCOg has more elaborate sentence structures. We also evaluate it on PhraseCut [36] dataset, which introduces structured textual descriptions that detail attributes, categories, and relationships among objects. We employ two primary metrics to evaluate the performance: overall Intersection over Union (oIoU) and mean Intersection over Union (mIoU). oIoU assesses the total overlap between predicted and ground truth regions relative to their combined area, making it particularly stringent for inaccuracies in larger segments. On the other hand, mIoU calculates the average overlap for each individual instance, ensuring a balanced consideration of performance across objects of varying sizes. Together, these metrics provide a robust framework for assessing the effectiveness of the proposed method in RIS tasks." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 588, + 447, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 588, + 447, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 447, + 601 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 605, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 714 + ], + "type": "text", + "content": "The experiments were run on a single NVIDIA RTX 3090 GPU. Following previous work [22, 32], we use the default ViT-H SAM model, and the hyperparameters \"predicted iou threshold\" and \"stability score threshold\" were set to 0.7, the \"points per side\" was set to 8. We use CLIP with ViTB/16 backbone in both hybrid feature extraction and spatial coherence guidance. We set " + }, + { + "bbox": [ + 313, + 605, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 605, + 554, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 605, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 605, + 554, + 714 + ], + "type": "text", + "content": " to 0.6 and 2, respectively, for all datasets. The hyperparameter " + }, + { + "bbox": [ + 313, + 605, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 605, + 554, + 714 + ], + "type": "text", + "content": " is empirically set to 9 to balance the propensity of both excessively large" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "29638" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 72, + 555, + 388 + ], + "blocks": [ + { + "bbox": [ + 56, + 72, + 555, + 388 + ], + "lines": [ + { + "bbox": [ + 56, + 72, + 555, + 388 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 555, + 388 + ], + "type": "table", + "html": "
MetricMethodVision BackbonePre-trained ModelRefCOCORefCOCO+RefCOCOg
valtestAtestBvaltestAtestBvaltest
oloUzero-shot methods w/ additional training
Pseudo-RIS [42]ViT-BSAM, CoCa, CLIP37.3343.4331.9040.1946.4333.6341.6343.52
VLM-VG [33]R101COCO*, VLM-VG*45.4048.0041.4037.0040.7030.5042.8044.10
zero-shot methods w/o additional training
Grad-CAM [26]R50SAM, CLIP23.4423.9121.6026.6727.2024.8423.0023.91
MaskCLIP [43]R50SAM, CLIP20.1820.5221.3022.0622.4324.6123.0523.41
Global-Local [41]R50FreeSOLO, CLIP24.5823.3824.3525.8724.6125.6130.0729.83
Global-Local [41]R50SAM, CLIP24.5526.0021.0326.6229.9922.2328.9230.48
Global-Local [41]ViT-BSAM, CLIP21.7124.4820.5123.7028.1221.8626.5728.21
Ref-Diff [22]ViT-BSAM, SD, CLIP35.1637.4434.5035.5638.6631.4038.6237.50
TAS [32]ViT-BSAM, BLIP2, CLIP29.5330.2628.2433.2138.7728.0135.8436.16
OursViT-BSAM,CLIP41.8144.5238.535.7441.4330.942.4742.97
mIoUweakly-supervised methods
CLRL [10]ViT-B-31.0632.3030.1131.2832.1130.1332.88-
PPT [4]ViT-BSAM46.7645.3346.2845.3445.8444.7742.97-
zero-shot methods w/ additional training
Pseudo-RIS [42]ViT-BSAM, CoCa, CLIP41.0548.1933.4844.3351.4235.0845.9946.67
VLM-VG [33]R101COCO*, VLM-VG*49.9053.1046.7042.7047.3036.2048.0048.50
zero-shot methods w/o additional training
Grad-CAM [26]R50SAM, CLIP30.2231.9027.1733.9625.6632.2933.0532.50
MaskCLIP [43]R50SAM, CLIP25.6226.6625.1727.4928.4930.4730.1330.15
Global-Local [41]R50FreeSOLO, CLIP26.7024.9926.4828.2226.5427.8633.0233.12
Global-Local [41]R50SAM, CLIP31.8332.9328.6434.9737.1130.6140.6640.94
Global-Local [41]ViT-BSAM, CLIP33.1236.5229.5835.2939.5831.8940.0840.74
CaR [31]ViT-B and ViT-LCLIP33.5735.3630.5134.2236.0331.0236.6736.57
Ref-Diff [22]ViT-BSAM, SD, CLIP37.2138.4037.1937.2940.5133.0144.0244.51
TAS [32]ViT-BSAM, BLIP2, CLIP39.8441.0836.2443.6349.1336.5446.6246.80
OursViT-BSAM, CLIP49.4853.3745.1943.4049.1337.1751.2551.59
", + "image_path": "c007b91f6f3be48f01f8c8e6cc35e2cb37059f73fdc37a5fd54170fe11984253.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 434, + 295, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 295, + 471 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 295, + 471 + ], + "type": "text", + "content": "and small mask regions receiving disproportionately high scores. When the referring text includes \"big\" or \"small\", " + }, + { + "bbox": [ + 55, + 434, + 295, + 471 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 434, + 295, + 471 + ], + "type": "text", + "content": " is adjusted to 3 or 14, respectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 479, + 113, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 113, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 113, + 491 + ], + "type": "text", + "content": "4.3. Results" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 498, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 713 + ], + "type": "text", + "content": "In Tab. 1, we evaluate our model on RefCOCO, RefCOCO+, and RefCOCOg and compare it with other state-of-the-art (SOTA) zero-shot models. Here, we evaluate the MaskCLIP, Grad-CAM methods by computing the similarity between their feature maps and SAM's mask proposals. Since the original Global-Local method used FreeSOLO as the mask extractor, we have re-evaluated the performance of Global-Local using SAM as the mask extractor, across different backbones. We also provide results of Ref-Diff, and TAS using different backbones. Additionally, we include weakly-supervised methods and zero-shot methods with extra training or additional datasets. Our method achieves excellent results on all three datasets. In terms of oIoU, our method improves by " + }, + { + "bbox": [ + 55, + 498, + 296, + 713 + ], + "type": "inline_equation", + "content": "4\\% -7\\%" + }, + { + "bbox": [ + 55, + 498, + 296, + 713 + ], + "type": "text", + "content": " over SOTA methods on RefCOCO and RefCOCOg, and although it does not achieve the best performance on the testB set of RefCOCO+, it is only " + }, + { + "bbox": [ + 55, + 498, + 296, + 713 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 55, + 498, + 296, + 713 + ], + "type": "text", + "content": " lower than the SOTA method. For mIoU, our method outperforms SOTA methods by " + }, + { + "bbox": [ + 55, + 498, + 296, + 713 + ], + "type": "inline_equation", + "content": "4\\% -10\\%" + }, + { + "bbox": [ + 55, + 498, + 296, + 713 + ], + "type": "text", + "content": " on" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 434, + 555, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 555, + 555 + ], + "type": "text", + "content": "RefCOCO and RefCOCOg, and it also achieves comparable performance on RefCOCO+. Importantly, our method achieves comparable or even higher mIoU on all three datasets compared to methods that use additional training. This indicates that our method has a more balanced consideration across varying objects. In addition, we present the oIoU and mIoU results on the PhraseCut dataset's test set in Tab. 2. Our method demonstrates superior performance compared to previous methods, achieving the highest average score." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 318, + 563, + 550, + 635 + ], + "blocks": [ + { + "bbox": [ + 55, + 392, + 555, + 427 + ], + "lines": [ + { + "bbox": [ + 55, + 392, + 555, + 427 + ], + "spans": [ + { + "bbox": [ + 55, + 392, + 555, + 427 + ], + "type": "text", + "content": "Table 1. Comparisons with the SOTA zero-shot approaches on RefCOCO, RefCOCO+, and RefCOCOg datasets. The best two results under the same setting, w/o additional training, are highlighted in bold and underlined, respectively. * indicates the extra dataset used to train the model." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 563, + 550, + 635 + ], + "lines": [ + { + "bbox": [ + 318, + 563, + 550, + 635 + ], + "spans": [ + { + "bbox": [ + 318, + 563, + 550, + 635 + ], + "type": "table", + "html": "
MethodsoIoUmIoUavg.
Global-Local[41]23.64-23.64
TAS[32]25.6424.6625.15
Ref-Diff[22]29.4241.7535.59
Ours38.3936.9837.69
", + "image_path": "e923b5cc7d91bfe6d85b62ccd30612ae7d581ac47011dcdf847e5ce968d735ef.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 645, + 554, + 666 + ], + "lines": [ + { + "bbox": [ + 313, + 645, + 554, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 645, + 554, + 666 + ], + "type": "text", + "content": "Table 2. Comparison with existing methods on the PhraseCut dataset." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "text", + "content": "Notably, our approach relies solely on SAM and CLIP, while TAS additionally incorporates the large BLIP2 model, and Ref-Diff leverages the Stable-Diffusion model; yet, our method still outperforms these more complex models. Be" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "29639" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 71, + 88, + 146, + 145 + ], + "blocks": [ + { + "bbox": [ + 74, + 76, + 332, + 87 + ], + "lines": [ + { + "bbox": [ + 74, + 76, + 332, + 87 + ], + "spans": [ + { + "bbox": [ + 74, + 76, + 332, + 87 + ], + "type": "text", + "content": "Referring text : The woman in the middle sitting on the couch" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 71, + 88, + 146, + 145 + ], + "lines": [ + { + "bbox": [ + 71, + 88, + 146, + 145 + ], + "spans": [ + { + "bbox": [ + 71, + 88, + 146, + 145 + ], + "type": "image", + "image_path": "1277ce1a277c109ba3b566aa1966d2be186089acf32cf30c8b45bbf86bdeb108.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 149, + 88, + 225, + 146 + ], + "blocks": [ + { + "bbox": [ + 149, + 88, + 225, + 146 + ], + "lines": [ + { + "bbox": [ + 149, + 88, + 225, + 146 + ], + "spans": [ + { + "bbox": [ + 149, + 88, + 225, + 146 + ], + "type": "image", + "image_path": "f5ace7e5698fa3a39b60d8b8840f8ff975943e43f44e74062d902d212d502530.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 228, + 88, + 304, + 146 + ], + "blocks": [ + { + "bbox": [ + 228, + 88, + 304, + 146 + ], + "lines": [ + { + "bbox": [ + 228, + 88, + 304, + 146 + ], + "spans": [ + { + "bbox": [ + 228, + 88, + 304, + 146 + ], + "type": "image", + "image_path": "830633bea87253211cbe328d63df5cd3ea1cc8e9ab3a9858a5be510752d49064.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 307, + 88, + 382, + 146 + ], + "blocks": [ + { + "bbox": [ + 307, + 88, + 382, + 146 + ], + "lines": [ + { + "bbox": [ + 307, + 88, + 382, + 146 + ], + "spans": [ + { + "bbox": [ + 307, + 88, + 382, + 146 + ], + "type": "image", + "image_path": "7b5f8fafda932b3da87cfd324016d04292534821c997d78c85db06ac7bd20dac.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 385, + 88, + 461, + 146 + ], + "blocks": [ + { + "bbox": [ + 385, + 88, + 461, + 146 + ], + "lines": [ + { + "bbox": [ + 385, + 88, + 461, + 146 + ], + "spans": [ + { + "bbox": [ + 385, + 88, + 461, + 146 + ], + "type": "image", + "image_path": "3e19d8aff5ba5f5b18b9dae1849f7acaa6ecae15510770c12a9639641eeb1ed1.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 465, + 88, + 540, + 146 + ], + "blocks": [ + { + "bbox": [ + 465, + 88, + 540, + 146 + ], + "lines": [ + { + "bbox": [ + 465, + 88, + 540, + 146 + ], + "spans": [ + { + "bbox": [ + 465, + 88, + 540, + 146 + ], + "type": "image", + "image_path": "b77d261737c50977f865da24abc425a948dc663ff9f6afce6451a18c65056558.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 71, + 160, + 146, + 212 + ], + "blocks": [ + { + "bbox": [ + 73, + 148, + 365, + 159 + ], + "lines": [ + { + "bbox": [ + 73, + 148, + 365, + 159 + ], + "spans": [ + { + "bbox": [ + 73, + 148, + 365, + 159 + ], + "type": "text", + "content": "Referring text : A beige surfboard being carried by a man in a wetsuit" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 71, + 160, + 146, + 212 + ], + "lines": [ + { + "bbox": [ + 71, + 160, + 146, + 212 + ], + "spans": [ + { + "bbox": [ + 71, + 160, + 146, + 212 + ], + "type": "image", + "image_path": "7a38693728058a60cbceb8f5fcac0d3023b7b1cc6e1dbe9cb5f024880d7d138c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 149, + 160, + 225, + 212 + ], + "blocks": [ + { + "bbox": [ + 149, + 160, + 225, + 212 + ], + "lines": [ + { + "bbox": [ + 149, + 160, + 225, + 212 + ], + "spans": [ + { + "bbox": [ + 149, + 160, + 225, + 212 + ], + "type": "image", + "image_path": "eee53ef79b8e3305441cfa9b64afec0e4f90098c407804bbdc048062d74aa4a3.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 73, + 213, + 328, + 226 + ], + "lines": [ + { + "bbox": [ + 73, + 213, + 328, + 226 + ], + "spans": [ + { + "bbox": [ + 73, + 213, + 328, + 226 + ], + "type": "text", + "content": "Referring text : The chair behind the guy wearing the stripes" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 228, + 160, + 304, + 212 + ], + "blocks": [ + { + "bbox": [ + 228, + 160, + 304, + 212 + ], + "lines": [ + { + "bbox": [ + 228, + 160, + 304, + 212 + ], + "spans": [ + { + "bbox": [ + 228, + 160, + 304, + 212 + ], + "type": "image", + "image_path": "4cee302decba7ab9189780f2885f63a5b63fd102baee62f37b72aa65d57161ff.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 307, + 160, + 382, + 212 + ], + "blocks": [ + { + "bbox": [ + 307, + 160, + 382, + 212 + ], + "lines": [ + { + "bbox": [ + 307, + 160, + 382, + 212 + ], + "spans": [ + { + "bbox": [ + 307, + 160, + 382, + 212 + ], + "type": "image", + "image_path": "28e0d10764d9ab28c68cb10b705b6febd7f963662663d65390b4681aa14914c8.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 385, + 160, + 461, + 212 + ], + "blocks": [ + { + "bbox": [ + 385, + 160, + 461, + 212 + ], + "lines": [ + { + "bbox": [ + 385, + 160, + 461, + 212 + ], + "spans": [ + { + "bbox": [ + 385, + 160, + 461, + 212 + ], + "type": "image", + "image_path": "469f846a77783f3c31ebfc0e8aae034d0af00f3f8e76ffd73d4206a3e59e0273.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 465, + 160, + 540, + 212 + ], + "blocks": [ + { + "bbox": [ + 465, + 160, + 540, + 212 + ], + "lines": [ + { + "bbox": [ + 465, + 160, + 540, + 212 + ], + "spans": [ + { + "bbox": [ + 465, + 160, + 540, + 212 + ], + "type": "image", + "image_path": "31a66926f07aad2c85bdcd0168a9f3044e9c84f1efcb3ffcddc02a8876c6d48e.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 71, + 227, + 146, + 282 + ], + "blocks": [ + { + "bbox": [ + 71, + 227, + 146, + 282 + ], + "lines": [ + { + "bbox": [ + 71, + 227, + 146, + 282 + ], + "spans": [ + { + "bbox": [ + 71, + 227, + 146, + 282 + ], + "type": "image", + "image_path": "1f5914738e6fbe1fccd79f0d00656339a0b5d56e38428bfba5346cac8d372001.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 99, + 283, + 124, + 293 + ], + "lines": [ + { + "bbox": [ + 99, + 283, + 124, + 293 + ], + "spans": [ + { + "bbox": [ + 99, + 283, + 124, + 293 + ], + "type": "text", + "content": "Image" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 149, + 227, + 225, + 283 + ], + "blocks": [ + { + "bbox": [ + 149, + 227, + 225, + 283 + ], + "lines": [ + { + "bbox": [ + 149, + 227, + 225, + 283 + ], + "spans": [ + { + "bbox": [ + 149, + 227, + 225, + 283 + ], + "type": "image", + "image_path": "9e63f670fb8d9503fafe1aad3a8352c603d5ba34e707c57c9f068c2abdf96138.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 167, + 283, + 211, + 292 + ], + "lines": [ + { + "bbox": [ + 167, + 283, + 211, + 292 + ], + "spans": [ + { + "bbox": [ + 167, + 283, + 211, + 292 + ], + "type": "text", + "content": "SAM-CLIP" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 228, + 227, + 304, + 282 + ], + "blocks": [ + { + "bbox": [ + 228, + 227, + 304, + 282 + ], + "lines": [ + { + "bbox": [ + 228, + 227, + 304, + 282 + ], + "spans": [ + { + "bbox": [ + 228, + 227, + 304, + 282 + ], + "type": "image", + "image_path": "0c338692d9ddcd6c0cdbfb4d2219da8964095f9724342ad53dc73a2e37424a1b.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 241, + 283, + 291, + 292 + ], + "lines": [ + { + "bbox": [ + 241, + 283, + 291, + 292 + ], + "spans": [ + { + "bbox": [ + 241, + 283, + 291, + 292 + ], + "type": "text", + "content": "Global-Local" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 293, + 555, + 316 + ], + "lines": [ + { + "bbox": [ + 55, + 293, + 555, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 555, + 316 + ], + "type": "text", + "content": "Figure 3. Visual comparisons with existing methods. Our approach achieves more accurate localization and a complete segmentation of the target object." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 307, + 227, + 382, + 282 + ], + "blocks": [ + { + "bbox": [ + 307, + 227, + 382, + 282 + ], + "lines": [ + { + "bbox": [ + 307, + 227, + 382, + 282 + ], + "spans": [ + { + "bbox": [ + 307, + 227, + 382, + 282 + ], + "type": "image", + "image_path": "428996b93833e65d31ceb1825e798a462f8302daa5b8b2819d8aefd20b646b83.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 283, + 353, + 291 + ], + "lines": [ + { + "bbox": [ + 335, + 283, + 353, + 291 + ], + "spans": [ + { + "bbox": [ + 335, + 283, + 353, + 291 + ], + "type": "text", + "content": "TAS" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 385, + 227, + 461, + 282 + ], + "blocks": [ + { + "bbox": [ + 385, + 227, + 461, + 282 + ], + "lines": [ + { + "bbox": [ + 385, + 227, + 461, + 282 + ], + "spans": [ + { + "bbox": [ + 385, + 227, + 461, + 282 + ], + "type": "image", + "image_path": "e691375c1beddf71d678894fc5e3d1362717e0231db25f3a4a187724500d1f2f.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 413, + 283, + 432, + 291 + ], + "lines": [ + { + "bbox": [ + 413, + 283, + 432, + 291 + ], + "spans": [ + { + "bbox": [ + 413, + 283, + 432, + 291 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 465, + 227, + 540, + 282 + ], + "blocks": [ + { + "bbox": [ + 465, + 227, + 540, + 282 + ], + "lines": [ + { + "bbox": [ + 465, + 227, + 540, + 282 + ], + "spans": [ + { + "bbox": [ + 465, + 227, + 540, + 282 + ], + "type": "image", + "image_path": "a450e5f0eeb9c824b197f6ba298cd2b0e409d72a3be190172c8882b87947ee17.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 493, + 283, + 506, + 292 + ], + "lines": [ + { + "bbox": [ + 493, + 283, + 506, + 292 + ], + "spans": [ + { + "bbox": [ + 493, + 283, + 506, + 292 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "bbox": [ + 55, + 323, + 296, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 323, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 296, + 396 + ], + "type": "text", + "content": "sides, TAS and Ref-Diff show better performance in either mIoU or oIoU, whereas our method consistently outperforms both metrics. We present visual comparisons with existing methods in Fig. 3, highlighting the improvements in segmentation accuracy and the ability to capture spatial relationships." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 55, + 401, + 149, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 401, + 149, + 414 + ], + "spans": [ + { + "bbox": [ + 55, + 401, + 149, + 414 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 55, + 418, + 296, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 296, + 479 + ], + "type": "text", + "content": "To evaluate the effectiveness of different strategies, we conduct extensive ablation studies on the val dataset of RefCOCO, RefCOCO+, and RefCOCOg datasets. We compare a range of feature extraction methods and spatial guidance strategies, examining their impact on performance." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 55, + 483, + 251, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 251, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 251, + 495 + ], + "type": "text", + "content": "4.4.1. Ablation on Hybrid Feature Extraction" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 55, + 498, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 689 + ], + "type": "text", + "content": "Local and Global Features. We first evaluate the performance of local and global features, including different strategies for global feature extraction. These strategies involve blurring images to reduce the influence of irrelevant regions on the global representation, applying a token mask to exclude non-mask regions, and using an attention mask to minimize the impact of non-mask regions on the CLS token. All these different mask strategies are applied to the last few layers of the CLIP image encoder. As shown in Tab. 3, the local features (L) consistently outperform the global methods across all datasets in terms of both oIoU and mIoU. Notably, combining a blur operation with the attention mask (G(blur+att_mask)) results in a significant improvement, approaching the performance of local features on the RefCOCO+ and RefCOCOg datasets. Hence, we adopt this approach to extract global features." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "Hybrid Global-Local Features. The intuitive way to fuse the global and local features is to compute a weighted sum" + } + ] + } + ], + "index": 33 + }, + { + "type": "table", + "bbox": [ + 315, + 321, + 554, + 478 + ], + "blocks": [ + { + "bbox": [ + 315, + 321, + 554, + 478 + ], + "lines": [ + { + "bbox": [ + 315, + 321, + 554, + 478 + ], + "spans": [ + { + "bbox": [ + 315, + 321, + 554, + 478 + ], + "type": "table", + "html": "
MethodRefCOCO oIoU mIoURefCOCO+ oIoU mIoURefCOCOg oIoU mIoU
L24.6532.6428.1636.4433.6342.35
G(blur)16.1722.0118.6125.1422.0031.24
G(tok_mask)20.5531.7121.8433.1819.8232.55
G(att_mask)20.0530.0721.4431.5124.1136.24
G(blur+att_mask)27.6337.1229.7539.9433.0344.45
G+L [41]21.7133.1223.7035.2926.5740.08
G+L28.9037.6432.2741.1238.0847.51
L2G29.9339.1733.1742.5537.7147.20
G2L31.7140.2734.4443.4039.1248.64
G2L + L2G31.3240.2334.6343.7338.4348.22
", + "image_path": "a17ee82c9411b971f9040bb08647e07f4feaaa7103c64c9f6b2adda4cca8f007.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "table_body" + } + ], + "index": 34 + }, + { + "bbox": [ + 313, + 481, + 554, + 503 + ], + "lines": [ + { + "bbox": [ + 313, + 481, + 554, + 503 + ], + "spans": [ + { + "bbox": [ + 313, + 481, + 554, + 503 + ], + "type": "text", + "content": "Table 3. Results of different mask feature extraction methods on the val split of three datasets." + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 509, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 555, + 640 + ], + "type": "text", + "content": "of the two. In Tab. 3, we present experimental results of this fusion strategy, employing the local and global feature extraction method from “G + L[41]”, which denotes the weighted sum of G(tok_mask) and cropped L, while “G + L” refers to the weighted sum of G(blur+att_mask) with L. We report both the results with the optimal weight that yielded the highest performance. Our results demonstrate that, due to the effectiveness of our global feature extraction strategy in generating a more accurate global representation, this weighted sum fusion significantly outperforms the previous method." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": "To fully exploit the attention mechanism in the model and facilitate a comprehensive interaction between the global and local features, we propose generating a hybrid feature by fusing features from the later layers of one branch's image encoder into the other branch. In Tab. 3, we present the results of different hybrid fusion strategies." + } + ] + } + ], + "index": 37 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "29640" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 66, + 70, + 284, + 166 + ], + "blocks": [ + { + "bbox": [ + 66, + 70, + 284, + 166 + ], + "lines": [ + { + "bbox": [ + 66, + 70, + 284, + 166 + ], + "spans": [ + { + "bbox": [ + 66, + 70, + 284, + 166 + ], + "type": "table", + "html": "
Start LayerG2LL2G
oIoUmIoUoIoUmIoU
722.6134.7829.1239.33
838.1647.8234.6743.62
939.1248.6437.6147.17
1036.5046.4737.3846.81
1134.1242.2633.2844.06
", + "image_path": "d4e7d85b32056161d923d33d20ce421f7cb53e56d477d3cfc621278e49f95477.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 206, + 295, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 295, + 445 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 295, + 445 + ], + "type": "text", + "content": "The \"L2G\" strategy, where local branch features are incorporated into the global branch, shows a significant improvement over \"G + L\". This highlights the effectiveness of our hybrid feature extraction method. On the other hand, the \"G2L\" strategy, where global features are fused into the local branch, provides the best results overall. This indicates that local features, which are crucial for fine-grained object identification, benefit most from the additional global context, allowing the model to achieve more accurate and contextually informed segmentations. The key difference between \"G2L\" and \"L2G\" is that the global branch applies an attention mask, while the local branch does not. Global features require a token mask for integration into the local branch, whereas local features can be directly fused into the global branch, resulting in distinct \"G2L\" and \"L2G\" outputs. We also explore the fusion of both \"L2G\" and \"G2L\" strategies using a weighted sum operation, but no further improvements are observed. Consequently, we adopt the \"G2L\" strategy for hybrid feature extraction in our implementation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 447, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 447, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 447, + 295, + 567 + ], + "type": "text", + "content": "Impact of Starting Layer for Hybrid Fusion We further investigate the effect of starting the hybrid fusion at different layers of the image encoder on the RefCOCOg dataset. In Tab. 4, we present the results of applying the fusion at various layers to determine which starting layer yields the best performance. Our findings show that initiating fusion at the 9-th layer yields the best performance, suggesting that representations at this layer provide the most meaningful and contextually relevant information for effective interaction between global and local features." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 576, + 211, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 211, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 211, + 588 + ], + "type": "text", + "content": "4.4.2. Ablation on Spatial Guidance" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 592, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 592, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 592, + 295, + 652 + ], + "type": "text", + "content": "Building upon the proposed hybrid global-local feature extraction method, we further conduct ablation studies to assess the impact of different components in the spatial guidance augmentation approach. Our experiments primarily focus on two datasets: RefCOCO and RefCOCOg." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 653, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 715 + ], + "type": "text", + "content": "Impact of Spatial Relationship Guidance From Tab. 5, we can see that incorporating spatial relationship guidance (Rel) leads to consistent performance improvements across all datasets. However, the improvements are more limited on RefCOCOg, where referring expressions typically con" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 316, + 71, + 550, + 155 + ], + "blocks": [ + { + "bbox": [ + 55, + 175, + 295, + 198 + ], + "lines": [ + { + "bbox": [ + 55, + 175, + 295, + 198 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 295, + 198 + ], + "type": "text", + "content": "Table 4. Results of the starting layer " + }, + { + "bbox": [ + 55, + 175, + 295, + 198 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 175, + 295, + 198 + ], + "type": "text", + "content": " for hybrid fusion of global and local branches on the val split of RefCOCOg dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 71, + 550, + 155 + ], + "lines": [ + { + "bbox": [ + 316, + 71, + 550, + 155 + ], + "spans": [ + { + "bbox": [ + 316, + 71, + 550, + 155 + ], + "type": "table", + "html": "
G2LRelGcoGposRefCOCORefCOCOg
oIoUmIoUoIoUmIoU
✓31.7140.2739.1248.64
✓✓35.6844.2939.6848.99
✓✓✓35.6844.2939.7349.02
✓✓✓✓41.8149.4842.4751.25
", + "image_path": "e77cf6ebfbe95dd480843f4f845f90b571b094e72e1e92fc36af8bf353a3b9d9.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 329, + 163, + 538, + 175 + ], + "lines": [ + { + "bbox": [ + 329, + 163, + 538, + 175 + ], + "spans": [ + { + "bbox": [ + 329, + 163, + 538, + 175 + ], + "type": "text", + "content": "Table 5. Ablation study on the different spatial guidance." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 182, + 553, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 182, + 553, + 290 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 553, + 290 + ], + "type": "text", + "content": "tain fewer spatial relationships. The most significant improvement is seen on RefCOCO, where both mIoU and oIoU increase by approximately " + }, + { + "bbox": [ + 313, + 182, + 553, + 290 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 313, + 182, + 553, + 290 + ], + "type": "text", + "content": ". This is attributed to the richer spatial relationship descriptions in the referring expressions, which are better captured by the introduced guidance. Overall, these results highlight the effectiveness of spatial relationship guidance in enhancing referring expression comprehension, especially in datasets with complex and detailed spatial descriptions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 291, + 555, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 291, + 555, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 291, + 555, + 483 + ], + "type": "text", + "content": "Impact of Spatial Coherence & Position Guidance Spatial coherence guidance " + }, + { + "bbox": [ + 313, + 291, + 555, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{\\mathrm{co}}" + }, + { + "bbox": [ + 313, + 291, + 555, + 483 + ], + "type": "text", + "content": " is introduced to enhance segmentation coherence and mitigate the issue of partial masks being selected for the target object. As a result, improvements in oIoU and mIoU are modest, as only a small subset of data face this issue, which may not be fully reflected in the overall scores. However, when combined with spatial position guidance " + }, + { + "bbox": [ + 313, + 291, + 555, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{\\mathrm{pos}}" + }, + { + "bbox": [ + 313, + 291, + 555, + 483 + ], + "type": "text", + "content": " together, as shown in Tab. 5, we observe a significant improvement in spatial position awareness, leading to more accurate target localization and enhanced segmentation performance. In referring expressions, spatial positional cues such as \"the left [object]\" often specify the target object's position without additional information. By introducing " + }, + { + "bbox": [ + 313, + 291, + 555, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{\\mathrm{pos}}" + }, + { + "bbox": [ + 313, + 291, + 555, + 483 + ], + "type": "text", + "content": ", which explicitly encodes spatial positioning information, we improve the alignment between the referring text and the mask prediction." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 495, + 388, + 507 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 388, + 507 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 388, + 507 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 515, + 553, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 515, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 313, + 515, + 553, + 646 + ], + "type": "text", + "content": "This paper presents a novel, training-free approach to zero-shot referring image segmentation (RIS), addressing challenges in aligning visual masks with referring expressions. Leveraging CLIP and SAM, our hybrid global-local feature extraction method combines mask-specific detail with contextual information to improve mask representation. Further, a spatial guidance augmentation strategy enhances spatial coherence and reduces ambiguities, effectively aligning masks with referring text. Experiments on RefCOCO, RefCOCO+, and RefCOCOg demonstrate that our method achieves substantial gains over zero-shot RIS models." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 647, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 553, + 712 + ], + "type": "text", + "content": "Acknowledgements. This work was supported in part by the National Natural Science Foundation of China (No.62106201, No.62376217), and in part by the Guangdong Basic and Applied Basic Research Foundation (No.2025A1515011501). The authors thank the anonymous reviewers for their helpful feedback." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29641" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 96, + 115, + 108 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 115, + 108 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 115, + 108 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 117, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 117, + 295, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 117, + 295, + 171 + ], + "spans": [ + { + "bbox": [ + 61, + 117, + 295, + 171 + ], + "type": "text", + "content": "[1] Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 173, + 295, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 173, + 295, + 228 + ], + "spans": [ + { + "bbox": [ + 61, + 173, + 295, + 228 + ], + "type": "text", + "content": "[2] Walid Bousselham, Felix Petersen, Vittorio Ferrari, and Hilde Kuehne. Grounding everything: Emerging localization properties in vision-language transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3828-3837, 2024. 2, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 230, + 295, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 230, + 295, + 283 + ], + "spans": [ + { + "bbox": [ + 62, + 230, + 295, + 283 + ], + "type": "text", + "content": "[3] Yong Xien Chng, Henry Zheng, Yizeng Han, Xuchong Qiu, and Gao Huang. Mask grounding for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26573-26583, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 286, + 294, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 286, + 294, + 339 + ], + "spans": [ + { + "bbox": [ + 62, + 286, + 294, + 339 + ], + "type": "text", + "content": "[4] Qiyuan Dai and Sibei Yang. Curriculum point prompting for weakly-supervised referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13711-13722, 2024. 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 342, + 295, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 342, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 62, + 342, + 295, + 407 + ], + "type": "text", + "content": "[5] Ruining Deng, Can Cui, Quan Liu, Tianyuan Yao, Lucas W Remedios, Shunxing Bao, Bennett A Landman, Lee E Weiless, Lori A Coburn, Keith T Wilson, et al. Segment anything model (sam) for digital pathology: Assess zero-shot segmentation on whole slide imaging. arXiv preprint arXiv:2304.04155, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 410, + 294, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 410, + 294, + 442 + ], + "spans": [ + { + "bbox": [ + 62, + 410, + 294, + 442 + ], + "type": "text", + "content": "[6] Sina Hajimiri, Ismail Ben Ayed, and Jose Dolz. Pay attention to your neighbours: Training-free open-vocabulary semantic segmentation. arXiv preprint arXiv:2404.08181, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 445, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 445, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 62, + 445, + 294, + 498 + ], + "type": "text", + "content": "[7] Ronghang Hu, Marcus Rohrbach, and Trevor Darrell. Segmentation from natural language expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 108-124. Springer, 2016. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 501, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 501, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 62, + 501, + 294, + 555 + ], + "type": "text", + "content": "[8] Ziling Huang and Shin'ichi Satoh. Referring image segmentation via joint mask contextual embedding learning and progressive alignment network. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 7753-7762, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 557, + 294, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 557, + 294, + 622 + ], + "spans": [ + { + "bbox": [ + 62, + 557, + 294, + 622 + ], + "type": "text", + "content": "[9] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4015-4026, 2023. 1, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 624, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 294, + 689 + ], + "type": "text", + "content": "[10] Jungbeom Lee, Sungjin Lee, Jinseok Nam, Seunghak Yu, Jaeyoung Do, and Tara Taghavi. Weakly supervised referring image segmentation with intra-chunk and inter-chunk consistency. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21870–21881, 2023. 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 691, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 294, + 714 + ], + "type": "text", + "content": "[11] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 74, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 335, + 74, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 553, + 106 + ], + "type": "text", + "content": "frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 108, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 553, + 162 + ], + "type": "text", + "content": "[12] Jiachen Li, Qing Xie, Xiaojun Chang, Jinyu Xu, and Yongjian Liu. Mutually-guided hierarchical multi-modal feature learning for referring image segmentation. ACM Trans. Multimedia Comput. Commun. Appl., 2024. Just Accepted. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 164, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 164, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 317, + 164, + 553, + 229 + ], + "type": "text", + "content": "[13] Tianxiao Li, Junhong Chen, Yiheng Huang, Kesi Huang, Qiqiang Xia, Muhammad Asim, and Wenyin Liu. Smvt: Spectrum-driven multi-scale vision transformer for referring image segmentation. In Advanced Intelligent Computing Technology and Applications, pages 193-206, Singapore, 2024. Springer Nature Singapore. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 231, + 553, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 231, + 553, + 273 + ], + "spans": [ + { + "bbox": [ + 317, + 231, + 553, + 273 + ], + "type": "text", + "content": "[14] Yi Li, Hualiang Wang, Yiqun Duan, and Xiaomeng Li. Clip surgery for better explainability with enhancement in open-vocabulary tasks. arXiv preprint arXiv:2304.05653, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 276, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 276, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 317, + 276, + 553, + 342 + ], + "type": "text", + "content": "[15] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 343, + 553, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 343, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 317, + 343, + 553, + 407 + ], + "type": "text", + "content": "[16] Sihan Liu, Yiwei Ma, Xiaqing Zhang, Haowei Wang, Jiayi Ji, Xiaoshuai Sun, and Rongrong Ji. Rotated multi-scale interaction network for referring remote sensing image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26658-26668, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 411, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 411, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 317, + 411, + 553, + 453 + ], + "type": "text", + "content": "[17] Yang Liu, Muzhi Zhu, Hengtao Li, Hao Chen, Xinlong Wang, and Chunhua Shen. Matcher: Segment anything with one shot using all-purpose feature matching. arXiv preprint arXiv:2305.13310, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 456, + 553, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 456, + 553, + 488 + ], + "spans": [ + { + "bbox": [ + 317, + 456, + 553, + 488 + ], + "type": "text", + "content": "[18] Jun Ma, Yuting He, Feifei Li, Lin Han, Chenyu You, and Bo Wang. Segment anything in medical images. Nature Communications, 15(1):654, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 491, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 491, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 317, + 491, + 553, + 544 + ], + "type": "text", + "content": "[19] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20, 2016. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 546, + 553, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 546, + 553, + 599 + ], + "spans": [ + { + "bbox": [ + 317, + 546, + 553, + 599 + ], + "type": "text", + "content": "[20] Sayan Nag, Koustava Goswami, and Srikrishna Karanam. Safari: Adaptive sequence transformer for weakly supervised referring expression segmentation. In European Conference on Computer Vision, pages 485-503. Springer, 2025. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 602, + 553, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 602, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 317, + 602, + 553, + 666 + ], + "type": "text", + "content": "[21] Varun K Nagaraja, Vlad I Morariu, and Larry S Davis. Modeling context between objects for referring expression understanding. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 792-807. Springer, 2016. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 670, + 553, + 712 + ], + "type": "text", + "content": "[22] Minheng Ni, Yabo Zhang, Kailai Feng, Xiaoming Li, Yiwen Guo, and Wangmeng Zuo. Ref-diff: Zero-shot referring image segmentation with generative models. arXiv preprint arXiv:2308.16777, 2023. 1, 2, 5, 6" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "29642" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 138 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 138 + ], + "type": "text", + "content": "[23] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 195 + ], + "type": "text", + "content": "[24] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "type": "text", + "content": "[25] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 253, + 294, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 294, + 317 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 294, + 317 + ], + "type": "text", + "content": "[26] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 618-626, 2017. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 319, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 319, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 319, + 294, + 374 + ], + "type": "text", + "content": "[27] Nisarg A Shah, Vibashan VS, and Vishal M Patel. Lqmformer: Language-aware query mask transformer for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12903-12913, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 376, + 294, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 294, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 294, + 431 + ], + "type": "text", + "content": "[28] Chao Shang, Zichen Song, Heqian Qiu, Lanxiao Wang, Fanman Meng, and Hongliang Li. Prompt-driven referring image segmentation with instance contrasting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4124-4134, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 433, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 433, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 433, + 294, + 464 + ], + "type": "text", + "content": "[29] Robin Strudel, Ivan Laptev, and Cordelia Schmid. Weakly-supervised segmentation of referring expressions. arXiv preprint arXiv:2205.04725, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 466, + 294, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 466, + 294, + 532 + ], + "spans": [ + { + "bbox": [ + 56, + 466, + 294, + 532 + ], + "type": "text", + "content": "[30] Sanjay Subramanian, William Merrill, Trevor Darrell, Matt Gardner, Sameer Singh, and Anna Rohrbach. Reclip: A strong zero-shot baseline for referring expression comprehension. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 5198-5215, 2022. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 534, + 294, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 534, + 294, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 534, + 294, + 588 + ], + "type": "text", + "content": "[31] Shuyang Sun, Runjia Li, Philip Torr, Xiuye Gu, and Siyang Li. Clip as rnn: Segment countless visual concepts without training endeavor. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13171-13182, 2024. 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 590, + 294, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 634 + ], + "type": "text", + "content": "[32] Yucheng Suo, Linchao Zhu, and Yi Yang. Text augmented spatial aware zero-shot referring image segmentation. In The 2023 Conference on Empirical Methods in Natural Language Processing, 2023. 1, 2, 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 635, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 678 + ], + "type": "text", + "content": "[33] Shijie Wang, Dahun Kim, Ali Taalimi, Chen Sun, and Weicheng Kuo. Learning visual grounding from generative vision and language model. arXiv preprint arXiv:2407.14563, 2024. 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 680, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 714 + ], + "type": "text", + "content": "[34] Wenxuan Wang, Tongtian Yue, Yisi Zhang, Longteng Guo, Xingjian He, Xinlong Wang, and Jing Liu. Unveiling parts beyond objects: Towards finer-granularity referring expres" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 576 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "sion segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12998-13008, 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "type": "text", + "content": "[35] Xinlong Wang, Zhiding Yu, Shalini De Mello, Jan Kautz, Anima Anandkumar, Chunhua Shen, and Jose M Alvarez. Freesolo: Learning to segment objects without annotations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14176-14186, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "type": "text", + "content": "[36] Chenyun Wu, Zhe Lin, Scott Cohen, Trung Bui, and Subhransu Maji. Phrasescut: Language-based image segmentation in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10216-10225, 2020. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 219, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 262 + ], + "type": "text", + "content": "[37] Defeng Xie, Ruichen Wang, Jian Ma, Chen Chen, Haonan Lu, Dong Yang, Fobo Shi, and Xiaodong Lin. Edit everything: A text-guided generative system for images editing. arXiv preprint arXiv:2304.14006, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 263, + 553, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 263, + 553, + 317 + ], + "spans": [ + { + "bbox": [ + 316, + 263, + 553, + 317 + ], + "type": "text", + "content": "[38] Li Xu, Mark He Huang, Xindi Shang, Zehuan Yuan, Ying Sun, and Jun Liu. Meta compositional referring expression segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19478-19487, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 319, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 319, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 319, + 553, + 352 + ], + "type": "text", + "content": "[39] Jinyu Yang, Mingqi Gao, Zhe Li, Shang Gao, Fangjing Wang, and Feng Zheng. Track anything: Segment anything meets videos. arXiv preprint arXiv:2304.11968, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 354, + 553, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 396 + ], + "type": "text", + "content": "[40] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 398, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 398, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 398, + 553, + 453 + ], + "type": "text", + "content": "[41] Seonghoon Yu, Paul Hongsuck Seo, and Jeany Son. Zero-shot referring image segmentation with global-local context features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19456-19465, 2023. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 454, + 553, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 454, + 553, + 497 + ], + "spans": [ + { + "bbox": [ + 316, + 454, + 553, + 497 + ], + "type": "text", + "content": "[42] Seonghoon Yu, Paul Hongsuck Seo, and Jeany Son. Pseudoris: Distinctive pseudo-supervision generation for referring image segmentation. In Proceedings of the European Conference on Computer Vision, 2024. 2, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 498, + 553, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 498, + 553, + 532 + ], + "spans": [ + { + "bbox": [ + 316, + 498, + 553, + 532 + ], + "type": "text", + "content": "[43] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In European Conference on Computer Vision, pages 696-712. Springer, 2022. 2, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 533, + 553, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 533, + 553, + 576 + ], + "spans": [ + { + "bbox": [ + 316, + 533, + 553, + 576 + ], + "type": "text", + "content": "[44] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Wang, Lijuan Wang, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29643" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Hybrid Reciprocal Transformer with Triplet Feature Alignment for Scene Graph Generation/11e7a635-4306-4d78-8190-6be9b1f3acac_content_list.json b/2025/Hybrid Reciprocal Transformer with Triplet Feature Alignment for Scene Graph Generation/11e7a635-4306-4d78-8190-6be9b1f3acac_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..659af776208b837ac38830a2e2c616a9311e8ca2 --- /dev/null +++ b/2025/Hybrid Reciprocal Transformer with Triplet Feature Alignment for Scene Graph Generation/11e7a635-4306-4d78-8190-6be9b1f3acac_content_list.json @@ -0,0 +1,1482 @@ +[ + { + "type": "text", + "text": "Hybrid Reciprocal Transformer with Triplet Feature Alignment for Scene Graph Generation", + "text_level": 1, + "bbox": [ + 233, + 130, + 767, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiawei Fu, Tiantian Zhang, Kai Chen, and Qi Dou \nDepartment of Computer Science and Engineering, The Chinese University of Hong Kong \njwfu@cse.cuhk.edu.hk", + "bbox": [ + 142, + 203, + 854, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Scene graph generation is a pivotal task in computer vision, focusing on comprehensive identification of visual relation tuples embedded within images. The advancement of methods involving triplets has sought to enhance task performance by integrating triplets as contextual features for more precise predicate identification from component level. However, challenges remain due to interference from multi-role objects in overlapping tuples within complex environments, which impairs the model's ability to distinguish and align specific triplet features for reasoning diverse semantics of multi-role objects. To address these issues, we introduce a novel framework that incorporates a triplet alignment model into a hybrid reciprocal transformer architecture, starting from using triplet mask features to guide the learning of component-level relation graphs. To effectively distinguish multi-role objects characterized by overlapping visual relation tuples, we introduce a triplet alignment loss, which provides multi-role objects with aligned features from triplet and helps customize them. Additionally, we explore the inherent connectivity between hybrid aligned triplet and component features through a bidirectional refinement module, which enhances feature interaction and reciprocal reinforcement. Experimental results demonstrate that our model achieves state-of-the-art performance on the Visual Genome and Action Genome datasets, underscoring its effectiveness and adaptability. Project page: hg-sg.github.io.", + "bbox": [ + 88, + 325, + 485, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 767, + 220, + 782 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Scene graph generation [2, 4, 18] aims to find all possible visual relation tuples $\\{\\text{subject-predicate-object}\\}$ given an image. It has wide applications in various fields, including image retrieval [37, 40, 49], visual grounding [13, 32, 47], and visual question answering [31, 33, 55]. Typically, diverse two-stage [20, 22, 41, 56, 57] and one-stage [4, 16, 18, 25, 46] methods have been developed to solve this prob", + "bbox": [ + 89, + 794, + 483, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/34955259c84ebc6373acc83bd9bc11c3dffdbcc6c8b7f09e262c52e43188ef1e.jpg", + "image_caption": [ + "Figure 1. Complex scenarios often involve overlapping visual relationships, where a single object can assume multiple roles across different triplets. Identifying these multi-role objects is challenging without knowing their aligned triplet features. Moreover, aligned triplets (top) and component (bottom) representations correspond to the same visual relation tuple, offering an opportunity to leverage their inherent relevance for reciprocal refinement." + ], + "image_footnote": [], + "bbox": [ + 516, + 292, + 645, + 434 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d74d360bf4b0558d3918ef15d96b17aedce03e12efd035371689ff2b184cea30.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 292, + 774, + 434 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bdfa7616dd081249b8131875bc92132dd428c31269a14711fc389ae9048e90cf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 292, + 905, + 434 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "lem by organizing regressed component-level subject, object, and predicate with post-processing. However, accurate predicate recognition remains challenging in scene graph generation, as many images incorporate extraneous visual elements. Directly regressing the predicate often introduces noise from features of unaligned visual relation tuples, owing to the lack of definite corresponding subject-object pair features for the visual relation. Accordingly, recent researches [4, 14, 22, 57] have utilized triplet features as prototypes for contextual semantics, combining detailed cues with broader triplet features to enhance predicate recognition accuracy from the triplet level.", + "bbox": [ + 511, + 550, + 908, + 733 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite advancements, accurately matching predicates to corresponding visual tuples and identifying all potential visual relations in complex environments remains challenging, even with triplet features. As illustrated in Figure 1, a single scene can contain multiple visual relations, with one object assuming multiple roles (either as subject or object) across different tuples. The analogous component-level features of multi-role objects increase overlap and misalignment among visual relation triplets, thereby hindering the model's ability to differentiate among triplets and to further reason diverse semantics of multi-role objects with", + "bbox": [ + 511, + 734, + 910, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8953", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "out precisely aligned triplet features. Furthermore, previous methodologies [4, 14, 22, 57] often overlook the potential of leveraging connectivity between aligned triplet and component features for mutual enhancement. This oversight can be attributed to the misalignment of triplet and component features [4, 14, 22, 57] or to a two-stage design [22, 57] that is not conducive to joint optimization of hybrid features.", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, it is essential to integrate feature alignment at triplet level, which is key in differentiating complex overlapping visuals. This alignment provides objects performing multiple roles with distinct aligned triplet features, facilitating their distinction within overlapped visual relational tuples. Moreover, to maximize the utility of aligned triplet and component features for identical visual relation tuples, and to investigate the consistency and connectivity between hybrid features to enhance mutual performance, implementing an end-to-end, one-stage scene graph generation framework is essential. This framework should enable bidirectional feature refinement for predicate recognition and object detection, thereby achieving joint optimization of both hybrid triplet and component features.", + "bbox": [ + 89, + 199, + 480, + 411 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this study, we integrate a triplet alignment model into a hybrid reciprocal transformer architecture for the generation of scene graphs. This integration leverages triplet-level representations as contextual features to enhance component-level scene generation and improve the accuracy of predicate recognition. To effectively differentiate complex, overlapping visual relation tuples with multirole objects at the feature level, we introduce a triplet alignment loss. This loss function promotes similarity between matched triplet and component features, while penalizing similarity among unmatched components from different tuples. It provides triplet-aligned features to multi-role objects to distinguish them in component level. Moreover, we investigate the inherent connectivity between aligned triplet and component features through a bidirectional refinement module. This module amalgamates aligned hybrid features from both triplets and components to facilitate mutual enhancement and reciprocal refinement.", + "bbox": [ + 89, + 414, + 482, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, we apply an image backbone and transformer encoder to embed the image information across the image. Then the hybrid reciprocal transformer embeds hybrid representations for triplet and component learning, where the triplet-level decoder is utilized to embed the proposal segmentation mask of {subject-predicate-object}, while the component-level decoders are applied to produce the corresponding detected subject, object, and predicate component information. To achieve triplet-guided component-level learning, we follow the iterative learning process [16], the query position embedding of the component-level learning transformer decoder is generated by the processed results of the triplet-level query embedding to assign the feature level between them. To real-", + "bbox": [ + 89, + 688, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ize bidirectional refinement, we additionally design a representation cross-conditioned mechanism among layer dimensions of the hybrid reciprocal transformer using attention [45] mechanism. Furthermore, we build a triplet alignment loss function to match triplet-level feature to be consistent with matched component-level feature, which aims to distinguish overlapped triplets with multi-role objects.", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The comprehensive experimental evaluations conducted on Visual Genome dataset [19] and Action Genome dataset [15] substantiate the effectiveness of our framework. In Visual Genome dataset, we demonstrate superior performance by achieving increments of $+1.2$ and $+2.3$ in Recall at 50 and 100, respectively, thereby establishing a new state-of-the-art. Additionally, we achieve highest recorded metrics at both Recall 20 and 50 in Action Genome dataset, confirming its scalability across diverse datasets.", + "bbox": [ + 511, + 198, + 903, + 333 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The contribution of this paper is summarized as follows:", + "bbox": [ + 532, + 334, + 903, + 349 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We integrate triplet-level and component-level representations within a hybrid reciprocal transformer framework, which uses the triplet mask feature as the proposal to improve predicate recognition performance.", + "- We introduce a triplet feature alignment loss to provide multi-role objects in overlapped visual relation with triplet-aligned features to distinguish them and improve predicate recognition efficacy.", + "- We leverage the aligned triplet and component feature with reciprocal refinement to explore inherent connectivity between them and optimize scene graphs." + ], + "bbox": [ + 513, + 351, + 903, + 516 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 532, + 653, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Two-Stage Scene Graph Generation", + "text_level": 1, + "bbox": [ + 513, + 558, + 823, + 574 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Scene graph generation has become a prominent research area within the vision community, as evidenced by a wealth of recent studies [2, 4, 16-18, 25]. Classic research [29] design two-stage methods for scene graph generation and use the detected object to organize the relations. Two-stage methods commence with the pre-training of a robust object detector for all classes present in the dataset, typically employing detectors such as Faster R-CNN [35]. Upon establishing the object detector, the visual relation detection module is then constructed utilizing the object data (bound- ing boxes and associated features) extracted from the detector. In the subsequent stage, the object detector's parameters are fixed, and the focus shifts solely to training the relation detection module to construct the scene graph. With the development of artificial intelligence model, various architectures have been implemented for the relation detection module [34, 43, 50-52], such as message passing [50], LSTM [43, 48, 52], transformer [7, 20, 30, 38, 41].", + "bbox": [ + 511, + 582, + 903, + 853 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent two-stage methods [22, 57] have incorporated triplet features to enhance predicate recognition. However, their effectiveness remains constrained in complex scenes", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8954", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "involving overlapping visual relation tuples with multi-role objects. Additionally, the inherent separate design of object and predicate architectures in these methods limits their ability to effectively utilize the connections between aligned triplet-level and component-level elements. To this end, we leverage a reciprocal one-stage scene graph generation framework which leverages hybrid triplet and component features to enhance each other. It surpasses the inherent constraints of two-stage systems and solves the obfuscation from multi-role objects with a triplet alignment module.", + "bbox": [ + 89, + 90, + 480, + 242 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. One-Stage Scene Graph Generation", + "text_level": 1, + "bbox": [ + 89, + 253, + 401, + 268 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Contrary to two-stage methods, which primarily suffer from the limited ability of the relation detector to directly utilize features extracted by the detector, one-stage approaches [4, 9, 16-18, 25] overcome this limitation. They achieve this by integrating object detector and graph generation architectures, enabling end-to-end training that facilitates direct access and utilization of extracted features. One-stage methods mainly consists of fully convolutional networks [28] and transformer-based models [4, 9, 16-18, 25]. This configuration facilitates direct interactions between the detection and graph generation modules, significantly enhancing the quality of the resulting scene graphs. The current approaches in one-stage methods primarily concentrate on component-level learning, analyzing subjects, objects, and predicates individually given image feature. These methods [16-18] typically compile these components into tuples $\\{\\text{subject-predicate-object}\\}$ during post-processing.", + "bbox": [ + 89, + 277, + 482, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Detecting accurate predicates and overcoming obfuscation in overlapped visual relation tuples pose significant challenges for current one-stage methods, primarily due to the lack of subject-object contextual features [22] and varying semantics of multi-role objects. We introduce triplet-level mask features as preliminary proposals to facilitate component-level learning in a one-stage architecture. Additionally, our method incorporates a triplet alignment module that aligns triplet and component features, enhancing the recognition of multi-role objects' diverse semantics.", + "bbox": [ + 89, + 535, + 482, + 685 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 700, + 181, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Formulation and Model Overview", + "text_level": 1, + "bbox": [ + 89, + 724, + 387, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Scene Graph Generation Formulation. Classic scene graph generation task aims to organize a relation graph $\\mathbf{G}'$ as $\\mathbf{G}' = \\{\\mathbf{s}_k, \\mathbf{o}_k, \\mathbf{p}_k\\}_{k=1}^K$ as a group of visual relation tuples organized from component-level subject $\\mathbf{s}_k$ , object $\\mathbf{o}_k$ , and predicate $\\mathbf{p}_k$ . In our proposed architecture for hybrid representation tailored for scene graph generation, we refine the conventional task formulation by incorporating a triplet mask for each visual relation tuple. Given an image $\\mathbf{I}$ , our scene graph $\\mathbf{G}$ is represented as a collection of quadruples that simultaneously encapsulates both triplet-level and tra", + "bbox": [ + 89, + 750, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ditional component-level representations:", + "bbox": [ + 513, + 90, + 787, + 106 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {G} = \\left\\{\\mathbf {t} _ {k}, \\mathbf {s} _ {k}, \\mathbf {o} _ {k}, \\mathbf {p} _ {k} \\right\\} _ {k = 1} ^ {K}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 114, + 903, + 133 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{t}_k$ is the triplet-level representation denoting the triplet mask constructed from the classic component-level representation subject $\\mathbf{s}_k$ , object $\\mathbf{o}_k$ , and predicate $\\mathbf{p}_k$ . Each component-level representation of the quadruple, i.e., $\\mathbf{s}_k, \\mathbf{o}_k, \\mathbf{p}_k$ , consists of a class label and bounding box. Note the bounding box corresponding to $\\mathbf{p}_k$ is calculated from the center point of its corresponding $\\mathbf{s}_k$ and $\\mathbf{o}_k$ . While $\\mathbf{t}_k$ represents the binary mask of the triplet of $\\mathbf{s}_k, \\mathbf{o}_k, \\mathbf{p}_k$ with a class label represent predicate class. This augmentation of triplet-level representation, as compared to traditional methodologies [4, 16-18], significantly provides subject-object pair contextual feature and limits explore space for accurate predicate recognition. Furthermore, it fosters exploration of the connectivity among hybrid triplet-level and component-level representations of same visual relations, thereby facilitating mutual enhancement.", + "bbox": [ + 511, + 143, + 906, + 383 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Model Overview. Figure 2 describes our overall architecture, which mainly consists of multi-scale image encoder and hybrid reciprocal transformer. We begin by utilizing an image encoder equipped with a ResNet backbone [12] and a transformer encoder [45] to extract multi-scale image visual features $\\mathcal{I}$ . Based on $\\mathcal{I}$ , our hybrid reciprocal transformer incorporates dual-level representations together, triplet mask queries $Q_{\\mathrm{t}}$ for triplet-level representation and subject queries $Q_{\\mathrm{s}}$ , object queries $Q_{\\mathrm{o}}$ , predicate queries $Q_{\\mathrm{p}}$ , for component-level representation, $Q$ means a sequence of queries as $Q = \\{q^{i}\\}_{i=1}^{N}$ . To leverage triplet-level representation facilitates guidance of component-level learning for accurate predicate recognition, hybrid reciprocal transformer uses iterative conditional chains for condition the process of $Q_{x}; x \\in \\{\\mathrm{s}, \\mathrm{o}, \\mathrm{p}\\}$ on $Q_{\\mathrm{t}}$ (Sec. 3.2). Furthermore, to enable bidirectional refinement in hybrid reciprocal transformer, we develop a cross-conditioning mechanism, which leverages an attention module that operates across layer dimension of hybrid reciprocal transformer to reason aligned hybrid representations (Sec. 3.2). Finally, we propose alignment loss that leverages triplet-component matched relationship, to provide multi-role objects with distinct aligned triplet feature and recognize overlapped visual relation tuples accurately (Sec. 3.3).", + "bbox": [ + 511, + 385, + 906, + 748 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Hybrid Reciprocal Transformer", + "text_level": 1, + "bbox": [ + 511, + 756, + 795, + 772 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Hybrid Reciprocal Transformer Architecture. Hybrid reciprocal transformer integrates triplet mask, subject, object, and predicate decoders to process input queries $Q_{x}$ ; $x \\in \\{\\mathrm{t},\\mathrm{s},\\mathrm{o},\\mathrm{p}\\}$ via cross-attention with cropping image features $\\mathcal{I}$ . $\\mathcal{I}$ is cropped through a multi-scale image encoding process, where the final layer of a ResNet-based convolutional network provides a position-aware, flattened image feature representation for a transformer encoder [45, 59].", + "bbox": [ + 511, + 779, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8955", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f99079d95f29ce1a68b4fc91d1f54a30862c60133d1364aac964f30135fb3caf.jpg", + "image_caption": [ + "Figure 2. Model overview of our method. Hybrid reciprocal transformer deploys four multi-layer decoder structures to embed triplet mask and visual relation tuple components learning using hybrid representations. We illustrate the detailed feature flow of each decoder to visualize two interaction paradigms within this transformer. The first paradigm facilitates triplet iterative guidance and enhances contextual feature integration for predicate recognition at the component level. The second paradigm employs aligned triplet and component features to enable bidirectional refinement of the hybrid triplet-level and component-level representations." + ], + "image_footnote": [], + "bbox": [ + 102, + 90, + 893, + 351 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, we model the elements of the scene graph—triplet mask, subject, object, and predicate, using multi-layer transformer decoders [1] with iterative processing across its $M$ layers step by step. The decoders are represented as $\\mathcal{T}_{\\mathrm{t}}$ , $\\mathcal{T}_{\\mathrm{s}}$ , $\\mathcal{T}_{\\mathrm{o}}$ , and $\\mathcal{T}_{\\mathrm{p}}$ , respectively. In each step $j$ , the $j$ -th layer of the triplet-level decoder, $\\mathcal{T}_{\\mathrm{t}}^{j}$ , generates masks feature for triplet-wise visual relations in the image. While the $j$ -th layer of each component-level decoder, $\\mathcal{T}_{x}^{j}$ where $x \\in \\{\\mathrm{s}, \\mathrm{o}, \\mathrm{p}\\}$ , produces component-level part feature of the scene graph $\\mathbf{G}^{j}$ . To realize triplet-level representation as proposals to boost component-level learning and allow hybrid representations to mutually reason and refine each other, our hybrid reciprocal transformer incorporates a triplet-guided relation generation mechanism with bidirectional learning refinement.", + "bbox": [ + 91, + 433, + 483, + 661 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Triplet-Guided Relation Learning with Bidirectional Refinement. To implement triplet-guided relation generation and bidirectional triplet-component enhancement, we develop two interaction paradigms between triplet-level and component-level features. The first paradigm involves using mask feature at the triplet level to guide the process of generating visual relations at the component level. This paradigm leverages triplet-level features as proposals to enhance predicate recognition from component-level data by incorporating contextual information from triplet-level representations. The second paradigm is to enable reciprocal sharing of triplet-level and component-level learning representations across the multiple layers of the hybrid reciprocal transformer through the use of an attention mechanism. This allows for streamlined integration of hybrid represen", + "bbox": [ + 89, + 672, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tations and mutual enhancement, as they represent different aspects of aligned identical aligned visual relation tuples.", + "bbox": [ + 511, + 433, + 906, + 463 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, for a decoder $\\mathcal{T}_x; x \\in \\{\\mathrm{t}, \\mathrm{s}, \\mathrm{o}, \\mathrm{p}\\}$ , we derive its output query representation by integrating self-attention on the input queries with positional embeddings and encoder-decoder attention using the image feature $\\mathcal{I}$ . In $j$ -th layers, the outputs of decoders are shown as:", + "bbox": [ + 511, + 464, + 906, + 540 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {Q} _ {x} ^ {j} = \\left\\{\\widehat {q} _ {x} ^ {i, j} \\right\\} _ {i = 1} ^ {N}; x \\in \\{t, s, o, p \\}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 547, + 906, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\widehat{Q}_x^j$ is produced from two types of conditional inputs from the aforementioned two interaction paradigms for hybrid representations respectively. The first is triplet-guided positional embedding $\\bar{P}_x^j = \\{\\bar{p}_x^{i,j}\\}_{i=1}^N$ ; $x \\in \\{\\mathrm{t}, \\mathrm{s}, \\mathrm{o}, \\mathrm{p}\\}$ based original position embedding $P_x^j = \\{p_x^{i,j}\\}_{i=1}^N$ ; $x \\in \\{\\mathrm{t}, \\mathrm{s}, \\mathrm{o}, \\mathrm{p}\\}$ respectively. The second is hybrid reciprocal query $\\bar{Q}_x^{j-1}$ ; $x \\in \\{\\mathrm{t}, \\mathrm{s}, \\mathrm{o}, \\mathrm{p}\\}$ from $j-1$ -th layer.", + "bbox": [ + 511, + 578, + 905, + 685 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Explicitly, the first involves conditioning component-level decoders $\\mathcal{T}_x; x \\in \\{\\mathrm{s,o,p}\\}$ on the step $j$ outputs of triplet-level decoder $\\mathcal{T}_{\\mathrm{t}}$ using guided positional embedding $\\bar{P}_x^j; x \\in \\{\\mathrm{t,s,o,p}\\}$ and following iterative process [16]. Specifically, $\\hat{Q}_x^j = \\mathcal{T}_x^j (\\bar{Q}_x^{j - 1}|\\bar{P}_x^j,\\mathcal{I}); x \\in \\{\\mathrm{t,s,o,p}\\}$ , where triplet decoder using original learnable position embedding $\\bar{P}_{\\mathrm{t}}^{j} = P_{\\mathrm{t}}^{j}$ , and triplet-guided mechanism is described as:", + "bbox": [ + 511, + 686, + 905, + 792 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {P} _ {\\mathrm {s}} ^ {j} = P _ {\\mathrm {s}} ^ {j} + \\text {F F N} \\left(\\mathrm {M H A} \\left(P _ {\\mathrm {s}} ^ {j}, \\widehat {Q} _ {\\mathrm {t}} ^ {j}, \\widehat {Q} _ {\\mathrm {t}} ^ {j}\\right)\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 797, + 903, + 816 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {P} _ {\\mathrm {o}} ^ {j} = P _ {\\mathrm {o}} ^ {j} + \\text {F F N} \\left(\\text {M H A} \\left(P _ {\\mathrm {o}} ^ {j}, \\left[ \\widehat {Q} _ {\\mathrm {t}} ^ {j}, \\widehat {Q} _ {\\mathrm {s}} ^ {j} \\right], \\left[ \\widehat {Q} _ {\\mathrm {t}} ^ {j}, \\widehat {Q} _ {\\mathrm {s}} ^ {j} \\right]\\right)\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 819, + 903, + 840 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {P} _ {\\mathrm {p}} ^ {j} = P _ {\\mathrm {p}} ^ {j} + \\text {F F N} \\left(\\mathrm {M H A} \\left(P _ {\\mathrm {p}} ^ {j}, \\left[ \\widehat {Q} _ {\\mathrm {t}} ^ {j}, \\widehat {Q} _ {\\mathrm {s}} ^ {j}, \\widehat {Q} _ {\\mathrm {o}} ^ {j} \\right], \\left[ \\widehat {Q} _ {\\mathrm {t}} ^ {j}, \\widehat {Q} _ {\\mathrm {s}} ^ {j}, \\widehat {Q} _ {\\mathrm {o}} ^ {j} \\right]\\right)\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 840, + 903, + 862 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathsf{FFN}(\\cdot)$ is the feed-forward neural network, and $\\mathsf{MHA}(\\cdot,\\cdot,\\cdot)$ is the multi-head attention mechanism for", + "bbox": [ + 511, + 869, + 908, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8956", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4521cf47e76fb4f11b729eceebd8cc56f4b6ec67ffe7cf5e41cd5639fcd1d021.jpg", + "image_caption": [ + "Figure 3. Triplet feature alignment loss measures consistency between triplet feature with tripled component-level feature. It encourages similarity between triplet and component feature from identical visual relation tuples to distinguish multi-role object by aligning different triplet feature into component level." + ], + "image_footnote": [], + "bbox": [ + 102, + 89, + 470, + 198 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\mathsf{Q},\\mathsf{K},\\mathsf{V}$ respectively, $[\\cdot ,\\cdot ]$ means concatenation of the element in it. While the second paradigm requires $\\mathcal{T}_{\\mathrm{t}}^{j}$ conditioned on $\\mathcal{T}_x^{j - 1};x\\in \\{\\mathrm{s,o,p}\\}$ and $\\mathcal{T}_x^j;x\\in \\{\\mathrm{s,o,p}\\}$ conditioned on $\\mathcal{T}_{\\mathrm{t}}^{j - 1}$ to allow bidirectional refinement. This cross-conditioned allows for feature flow enhancement among hybrid learning representations and is achieved through conditional hybrid reciprocal query $\\bar{Q}_x^{j - 1};x\\in$ $\\{\\mathrm{t,s,o,p}\\}$ as:", + "bbox": [ + 89, + 276, + 483, + 398 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {Q} _ {\\mathrm {t}} ^ {j} = \\widehat {Q} _ {\\mathrm {t}} ^ {j} + \\mathsf {F F N} (\\mathsf {M H A} (\\widehat {Q} _ {t} ^ {j}, [ \\widehat {Q} _ {x} ^ {j} ], [ \\widehat {Q} _ {x} ^ {j} ])); x \\in \\{\\mathrm {s}, \\mathrm {o}, \\mathrm {p} \\}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 407, + 482, + 428 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {Q} _ {x} ^ {j} = \\widehat {Q} _ {x} ^ {j} + \\mathsf {F F N} (\\mathsf {M H A} (\\widehat {Q} _ {x} ^ {j}, \\widehat {Q} _ {\\mathrm {t}} ^ {j}, \\widehat {Q} _ {\\mathrm {t}} ^ {j})); x \\in \\{\\mathrm {s}, 0, \\mathrm {p} \\}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 429, + 482, + 450 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By implementing the two specified conditional input patterns, we enable the generation of triplet-guided relation graph features within the same layer across various segments of hybrid reciprocal transformer, as well as bidirectional enhancement of hybrid representations across multiple layers of hybrid reciprocal transformer. Subsequently, these representations undergo processing by their respective prediction heads to produce desired hybrid outputs. Then alignment loss is applied for supervision, aiming to align hybrid features from the triplet dimension and enhance the differentiation of overlapping visual relation tuples.", + "bbox": [ + 89, + 460, + 482, + 626 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Prediction Heads. We build prediction heads for hybrid triplet-level and component-level features respectively. For triplet-level feature, we follow [6, 58] to refine with masked attention [5] with the multi-scale image feature. Specifically, triplet masks are generated through an iterative process wherein a dot product is applied across multiple scale dimensions along the channel axis between multi-scale image features $\\mathcal{I}$ and $\\widehat{Q}_{\\mathrm{t}}$ . Then the sigmoid function is applied to normalize the segmentation masks. For the component-level detection bounding box representation (subject, object, and predicate), we use MLP to output their corresponding box position and class information directly.", + "bbox": [ + 89, + 626, + 482, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Triplet Feature Alignment Loss", + "text_level": 1, + "bbox": [ + 89, + 816, + 369, + 834 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Alignment Loss. To enhance the differentiation of similar component features in multi-role objects that belong to overlapping visual relation tuples, we propose a triplet-component alignment loss. This loss function refines the", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "representation of similar component features by motivating multiple queries to delineate distinct roles within a single multi-role object, correlating each query specifically to specific triplet configuration. This mechanism is realized by aligning the computed matched similarity matrix with the ground truth (GT) matching matrix to accurately map between triplet configurations and component representations.", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the example illustrated in Figure 3, we observe a scenario featuring several overlapping visual relationships due to the presence of multi-role objects, such as a $\\underline{bag}$ and $\\underline{dog}$ . Here, the alignment loss assigns distinct features for different roles to each multi-role object based on their corresponding triplet-level features. Through the iterative matching and optimization process enabled by multiple layers of hybrid reciprocal transformer, the model promotes distinct representations for the multiple roles of a single object in overlapping triplets, using unique query-level features derived from their respective aligned triplet features. Furthermore, it promotes distinctiveness by encouraging both triplet-level and component-level features of various visual tuples to maintain separation, which aids in accurate recognition of complex and overlapping visual relation tuples.", + "bbox": [ + 511, + 198, + 903, + 425 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, in $j$ -th layer of hybrid reciprocal transformer, we calculate predicated similarity matrix given the processed triplet-level query $\\widehat{Q}_{\\mathrm{t}}^{j}$ and component-level query $\\widehat{Q}_{x}^{j}; x \\in \\{\\mathrm{s}, \\mathrm{o}, \\mathrm{p}\\}$ . We first produce aligned query representation as $\\widetilde{Q}_{x}^{j}; x \\in \\{\\mathrm{t}, \\mathrm{s}, \\mathrm{o}, \\mathrm{p}\\}$ using their respective matching results to identity GT results to maintain bitwise corresponding among query level. Based on the aligned $\\widetilde{Q}_{x}^{j}; x \\in \\{\\mathrm{t}, \\mathrm{s}, \\mathrm{o}, \\mathrm{p}\\}$ , we generate a pseudo triplet-level representation $\\mathbb{Q}_{\\mathrm{t}}^{j}$ as the weighted sum of the output component-level representations $\\widetilde{Q}_{x}^{j}; x \\in \\{\\mathrm{s}, \\mathrm{o}, \\mathrm{p}\\}$ , and calculate the predicted triplet-level similarities matrix $S^{j}$ of $j$ -th layer:", + "bbox": [ + 511, + 426, + 906, + 597 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} ^ {j} = \\mathbb {Q} _ {\\mathrm {t}} ^ {j} \\left(\\widetilde {Q} _ {\\mathrm {t}} ^ {j}\\right) ^ {T}; \\mathbb {Q} _ {\\mathrm {t}} ^ {j} = \\widetilde {Q} _ {\\mathrm {s}} ^ {j} + \\widetilde {Q} _ {\\mathrm {o}} ^ {j} + \\widetilde {Q} _ {\\mathrm {p}} ^ {j}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 608, + 903, + 630 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While the GT similarity matrix $S_{\\mathrm{GT}}$ is derived from the original matching relationship between the GT triplet-level masks and component-level subjects, objects, and predicates bounding boxes as one-hot formatting.", + "bbox": [ + 511, + 641, + 903, + 702 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Once the GT matrix $S_{\\mathrm{GT}}$ and predicted similarity matrix $S^j$ are obtained, binary cross-entropy loss is applied between them to update model parameters. Incorporate with the iterative optimization and matching process within a hybrid reciprocal transformer framework, the alignment loss enhances the model's capability to differentiate overlapping visual relation tuples by promoting a high degree of similarity between matched component-level and triplet-level representations with GT matching results.", + "bbox": [ + 511, + 703, + 905, + 838 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "End-to-End Training. In the end-to-end training process for our method, the overall loss function is the sum of the loss from all $M$ layers of hybrid reciprocal transformer, the loss $\\mathcal{L}$ can be written as:", + "bbox": [ + 511, + 839, + 905, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8957", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/f567b8675fb0ca2f06dbfcf9591bf7a05d81d7d9650dfa0aec6c349e33972fc9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodR@50R@100mR@50mR@100AvgR@50AvgR@100F@50F@100
Two-Stage, X101-FPN Backbone
Motifs [52] [CVPR2018]32.136.95.56.818.821.99.411.5
VCTree [43] [CVPR202]31.836.16.67.719.221.910.912.7
TDE [44] [CVPR2020]19.423.29.311.114.417.212.615.0
EBM [42] [CVPR2021]20.524.79.711.615.118.213.215.8
BPLSA [11] [ICCV2021]21.725.513.515.717.620.616.619.4
DT2-ACBS [8] [ICCV2021]22.024.415.016.318.520.417.819.5
BGNN [24] [CVPR2021]31.035.810.712.720.924.215.918.7
PE-Net [57] [CVPR2023]30.735.212.414.521.624.917.720.5
VETO [41] [CVPR2023]27.531.58.19.517.820.512.514.6
DRM [22] [CVPR2024]34.038.99.011.221.525.014.217.4
One-Stage, ResNet-50 Backbone
RelTR [7] [TPAMI2023]27.530.710.812.319.221.515.517.6
EGTR [14] [CVPR2024]30.234.37.910.119.122.212.515.6
Hydra-SGG [4] [Arxiv2024]28.433.116.019.722.226.420.524.7
One-Stage, ResNet-101 Backbone
RelDN [53] [CVPR2019]30.334.84.45.417.420.17.79.3
AS-Net [3] [CVPR2021]18.721.16.17.212.414.29.210.7
SGTR [25] [CVPR2022]25.126.612.014.618.620.616.218.9
HOTR [17] [CVPR2021]22.427.16.99.714.718.410.614.3
SpeaQHotR [18] [CVPR2024]24.729.19.612.717.220.913.817.7
ISG† [16] [NIPS2022]29.532.17.48.418.520.311.813.3
SpeaQISG† [18] [CVPR2024]32.936.011.814.122.425.117.420.3
ISG [16] [NIPS2022]27.230.115.016.621.123.419.321.4
SpeaQISG [18] [CVPR2024]32.135.515.117.623.626.620.523.5
Ours34.1(+1.2)38.3(+2.3)16.0(+0.9)20.5(+2.9)25.1(+1.5)29.4(+2.8)21.8(+1.3)26.7(+3.2)
", + "bbox": [ + 96, + 88, + 906, + 390 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Scene graph generation performance on Visual Genome. The best results among models with ResNet-101 backbone are marked in bold. $\\dagger$ denotes the performance without loss re-weighting proposed in [16].", + "bbox": [ + 89, + 392, + 906, + 421 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\sum_ {j = 1} ^ {M} \\left(\\mathcal {L} _ {\\mathrm {t}} ^ {j} + \\mathcal {L} _ {\\mathrm {c}} ^ {j} + \\mathcal {L} _ {\\text {a l i g n}} ^ {j}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 426, + 480, + 474 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {t}} = \\mathcal {L} _ {\\mathrm {t}} ^ {\\text {m a s k}} + \\mathcal {L} _ {\\mathrm {t}} ^ {\\text {d i c e}} + \\mathcal {L} _ {\\mathrm {t}} ^ {\\text {c l s}},\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 474, + 367, + 489 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {c}} = \\mathcal {L} _ {\\mathrm {c}} ^ {\\mathrm {s}} + \\mathcal {L} _ {\\mathrm {c}} ^ {\\mathrm {o}} + \\mathcal {L} _ {\\mathrm {c}} ^ {\\mathrm {p}},\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 492, + 328, + 508 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{t}}$ denotes triplet mask loss through Hungarian matching, which containing mask loss $\\mathcal{L}_{\\mathrm{t}}^{\\mathrm{mask}}$ , dice loss $\\mathcal{L}_{\\mathrm{t}}^{\\mathrm{dice}}$ , and mask classification loss $\\mathcal{L}_{\\mathrm{t}}^{\\mathrm{cls}}$ , $\\mathcal{L}_{\\mathrm{c}}$ follows the design of [18] and consists of subject loss $\\mathcal{L}_{\\mathrm{c}}^{\\mathrm{s}}$ , object loss $\\mathcal{L}_{\\mathrm{c}}^{\\mathrm{o}}$ , and predicate loss $\\mathcal{L}_{\\mathrm{c}}^{\\mathrm{p}}$ , each of $\\mathcal{L}_{\\mathrm{c}}^{\\mathrm{s}}$ , $\\mathcal{L}_{\\mathrm{c}}^{\\mathrm{o}}$ , and $\\mathcal{L}_{\\mathrm{c}}^{\\mathrm{p}}$ is constructed from L1 box loss, GIoU loss [36], and classification loss through groupwise query multi-assignment matching [18], $\\mathcal{L}_{\\mathrm{align}}^{j}$ is alignment loss calculated using binary cross-entropy loss between $S^{j}$ and $S_{\\mathrm{GT}}$ .", + "bbox": [ + 89, + 520, + 483, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 89, + 671, + 215, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experiment Setup", + "text_level": 1, + "bbox": [ + 89, + 696, + 266, + 713 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We demonstrate the effectiveness of our proposed approach on two benchmark datasets. Visual Genome (VG) [19] is utilized as a benchmark for scene graph generation. We leverage the commonly processed subset from [50], which comprises 108k images across 150 object categories and 50 predicate categories. Action Genome (AG) [15] provides scene graph annotations across 9,848 videos within the Charades dataset [39] for studying human-object interactions. It covers 35 object categories and 25 relation categories, deliberately excluding the class person. Adhering to protocols from previous studies [10], we only utilize frames that have been annotated according to [15]. Frames", + "bbox": [ + 89, + 719, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "lacking annotations for person or other objects are excluded to ensure the creation of usable scene graphs.", + "bbox": [ + 511, + 430, + 906, + 459 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. Consistent with established methodologies in the literature [50, 52], we evaluate our model using the standard metrics Recall@K (R@K) and Mean Recall@K (mR@K). We further quantify performance by calculating both the arithmetic mean (AvgR@K) [23] and the harmonic mean (F@K) [16]. These additional measures provide a more nuanced assessment of model performance across varying predicate frequencies, ensuring a comprehensive evaluation of the balanced performance.", + "bbox": [ + 511, + 460, + 905, + 597 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. We utilize a ResNet-101 backbone [12] paired with a six-layer Transformer encoder. For each part in $\\mathcal{T}_x$ ; $x \\in \\{\\mathrm{t}, \\mathrm{s}, \\mathrm{o}, \\mathrm{p}\\}$ of the hybrid reciprocal transformer, we deploy a six-layer transformer decoder [45, 59] architecture. The query number for component-level learning is 600 for $\\mathcal{T}_x^j$ ; $x \\in \\{\\mathrm{s}, \\mathrm{o}, \\mathrm{p}\\}$ , and 300 for $\\mathcal{T}_{\\mathrm{tn}}^j$ . The triplet matching strategy employs a naive Hungarian algorithm, optimizing for mask loss, dice loss, and classification loss. For component-level matching, we adopt a grouping matching strategy as proposed in [18], ensuring consistency in matched indices across components. The model undergoes pretraining on VG and AG datasets, focusing on object detection and semantic segmentation, respectively. Experiments are conducted on four A40 GPUs with batch size 16.", + "bbox": [ + 511, + 598, + 906, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison Results with State-of-the-art", + "text_level": 1, + "bbox": [ + 511, + 816, + 867, + 830 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Visual Genome Dataset. In Table 1, we report the performance of our method on VG dataset, the results show that our method achieves new state-of-the-art performance on those one-stage methods based on ResNet-101 backbone.", + "bbox": [ + 511, + 839, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8958", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/72123fc504572586026605f8510bd6cfff348334b017efa127fff1b29d0df2fa.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodR@20/50mR@20/50F@20/50
Two-Stage, X101-FPN Backbone
VRD [29]10.3 / 10.9-
Freq Prior [52]24.0 / 24.9--
IMP [50]23.9 / 25.5--
MSDN [26]24.0 / 25.6--
Graph R-CNN [51]24.1 / 25.8--
RelDN [53]25.0 / 26.2--
SimpleBase [10]27.9 / 30.48.3 / 9.112.8 / 14.0
One-Stage, ResNet-101 Backbone
ISG [16]31.0 / 37.510.9 / 13.216.1 / 19.5
\\( ISG_{\\dagger} \\) [16]30.1 / 36.519.2 / 22.323.4 / 27.7
Spea\\( Q_{\\text{ISG}} \\) [18]30.8 / 38.011.7 / 13.817.0 / 20.2
Spea\\( Q_{\\text{ISG}_{\\dagger}} \\) [18]31.8 / 39.420.9 / 23.125.2 / 29.1
Ours32.0 / 40.221.3 / 23.625.5 / 29.7
", + "bbox": [ + 102, + 88, + 472, + 271 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our model demonstrates significant enhancements with increases of $+1.2$ , $+2.3$ , $+0.9$ , and $+2.9$ in the R@50/100 and mR@50/100 metrics. These enhancements can be credited to our pioneering hybrid reciprocal architecture, which integrates reasoning across aligned triplets and component features. Furthermore, the synergistic interplay between the two-tiered representation of visual relation tuple significantly augments the overall performance. When compared with two-stage methods employing the X101-FPN backbone, our method achieves state-of-the-art results in R@K, as evidenced by its comparison with the DRM model [22]. Moreover, owing to our one-stage design to jointly optimize predicate recognition and object detection, we substantially outperform two-stage approaches in mR@K.", + "bbox": [ + 89, + 309, + 483, + 520 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Action Genome Dataset. We additionally report R@20/50, mR20/50, and F@20/50 values on the Action Genome dataset [15] in Table 2. Consistent with the findings observed on the Visual Genome dataset, we establish a new benchmark for state-of-the-art performance across all utilized metrics, including R@20/50, mR@20/50, and F@20/50. This advancement underscores the efficacy of our hybrid reciprocal architecture with triplet alignment in enhancing scene graph generation capabilities. Moreover, the improvement in performance across various datasets demonstrates the robust scalability and adaptability of the hybrid learning framework, which does not necessitate specific modifications to be effective across different contexts.", + "bbox": [ + 89, + 521, + 483, + 717 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study and Analysis", + "text_level": 1, + "bbox": [ + 89, + 726, + 346, + 742 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Alignment Loss. We apply an alignment loss between hybrid triplet and component features to enhance the distinction of multi-role objects in overlapping triplets, by aligning triplet features into the component level. We investigate the effect of the triplet feature alignment loss by setting different weights in the end-to-end training overall loss function. The results metrics in the VG dataset using different weights (0.0, 0.1, 0.2, 0.3, 0.4, 0.5, 1.0) are reported at Figure 4, the weight used in our vanilla model is 0.2. Our findings suggest that increasing the weight of alignment loss", + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/00063d6094f90b08deb0255954de53456795003940482df8ea136d2065b86a9a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 522, + 93, + 897, + 188 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e6300f9fb6ca522991e9505e354d86fd817060e7983780da842f9dfdef500a87.jpg", + "image_caption": [ + "Figure 4. Scene graph generation performance comparison with Visual Genome dataset using different parameters of weight of alignment loss in overall training loss." + ], + "image_footnote": [], + "bbox": [ + 522, + 190, + 897, + 292 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1ee878197dac3029ab57e207726a5d7fb859188178bb25921c28352e50251bf0.jpg", + "table_caption": [ + "Table 2. Scene graph generation performance on Action Genome. Best performance with bold." + ], + "table_footnote": [], + "table_body": "
αc:tR@50R@100mR@50mR@100
0.532.135.316.522.3
1.033.235.716.121.2
2.033.736.816.422.3
", + "bbox": [ + 588, + 330, + 831, + 388 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Effect of matching triplet-level and component-level together with different ratio ${\\alpha }_{\\mathrm{c} : \\mathrm{t}}$ between component and triplet cost.", + "bbox": [ + 511, + 388, + 906, + 419 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "initially enhances performance as desired, but subsequently deteriorates. This pattern of performance degradation at higher weights of alignment loss indicates that excessive emphasis on alignment loss may hinder learning efficacy.", + "bbox": [ + 511, + 429, + 903, + 491 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Hybrid Representation Guidance Order and Reciprocal Refinement. In our framework, we traditionally utilize the processed query for triplet-level segmentation masks to condition the generation of queries for component-level representations. To explore the impact of reversing this sequence, we instead employ the component-level query as a condition for generating triplet-level representations. The results for this variant, denoted as R@50/100 and mR@50/100, are 33.2, 36.9, 15.3, and 18.9, respectively, when compared with our vanilla model, as detailed in Table 1, it is evident that utilizing the triplet-level query as a condition for component-level learning yields superior performance. This enhancement can be attributed to the global information embedded in the triplet-level query feature, which effectively enhances the learning of local information in the component-level elements—subject, object, and predicate. Additionally, we explore the impact of reciprocal refinement by omitting this feature from our vanilla model. The performance metrics on VG dataset for R@50/100 and mR@50/100 are 32.4, 36.3, 14.4, and 18.5, respectively. The observed decrease in performance underscores the significance of reciprocal refinement in integrating hybrid triplet and component features effectively.", + "bbox": [ + 511, + 491, + 906, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Grouping Hybrid Representations in Uniform Matcher. Inspired by [18], we explore integrating hybrid representations using a single Hungarian matching mechanism to align them directly. Table 3 presents a comparative anal", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8959", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4d36f69c497476372929a2ebb076c3021165d744d5f628e9475735601b159455.jpg", + "image_caption": [ + "Figure 5. Qualitative analysis of our method compared with the baseline models, SpeaQ [18] with ISG [16], indicates that our method yields significantly enhanced accuracy in the prediction of predicates and object detection. Red marks error." + ], + "image_footnote": [], + "bbox": [ + 93, + 95, + 903, + 257 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/930721ed36421146b51a5922b45cf4651bb84265777b8320068bf123a1a044c4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
NQtNQcR@50R@100mR@50mR@100
30030033.837.315.218.3
40033.837.415.318.9
50033.937.615.719.4
60030033.937.815.720.1
40034.037.915.920.5
50034.238.216.221.2
", + "bbox": [ + 156, + 291, + 415, + 378 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ysis of the outcomes, employing varied matching ratios between component and triplet levels. The findings indicate that merging the hybrid representations via a Hungarian matching mechanism yields inferior results compared to our vanilla model. Specifically, reducing component-level matching negatively impacts performance. This decline is attributed to the stringent constraints imposed by grouping two distinct hybrid representations, which restricts their internal matching and adversely affects overall performance.", + "bbox": [ + 88, + 417, + 482, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Query Number. We examine the impact of query number for hybrid representations ( $N_{Q\\mathrm{t}}$ for triplet and $N_{Q\\mathrm{c}}$ for component) in Table 4. The comparative analysis indicates that incremental increases in query numbers consistently enhance outcome performance. This improvement adheres to fundamental principles observed in transformer-based approaches for visual tasks [1, 21, 27, 59] by providing richer representational capacity and coverage [54].", + "bbox": [ + 88, + 551, + 482, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Matching of Hybrid Representations in Inference Stage. In the inference stage, our vanilla model only uses the bitwise multiply results from the component-level subject, object, and predicate scores. We compare two deformations in Table 5: the first directly calculates multiply of bitwise triplet-level mask scores together, and the second uses Hungarian matching to align two-level output and then multiply them together. Their corresponding results are shown in Table 5. The comparison results show that directly multiply component-level with triplet-level score together will definitely cause negative performance, since the two-level output representation cannot be directly assigned to the specific result subject, object, and predicate tuple. We also find that after the Hungarian matching, the multiplied scores of the hybrid two-level representation will improve performance", + "bbox": [ + 88, + 674, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/be7a41d7c05c793554b8298130ebc820540dd24fd47f789b29c2b6d819f4792c.jpg", + "table_caption": [ + "Table 4. Scene graph performance on VG dataset with different number of triplet-level and component-level queries." + ], + "table_footnote": [], + "table_body": "
Inference TypeR@50R@100mR@50mR@100
Component Only34.138.316.020.5
Component × Triplet32.535.614.616.5
Component × Triplet34.238.316.120.6
", + "bbox": [ + 552, + 291, + 862, + 347 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Comparison results among inference matching between hybrid representations. $\\times$ means bitwise multiply and $\\star$ denotes multiply after matching.", + "bbox": [ + 511, + 349, + 903, + 390 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "a bit. We suppose this improvement came from the final aligned hybrid representation by the alignment loss will boost the final scene graph generation performance.", + "bbox": [ + 511, + 398, + 903, + 443 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Qualitative Results", + "text_level": 1, + "bbox": [ + 511, + 448, + 694, + 463 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 5 presents qualitative examples comparing prediction results from the SpeaQ [18] and ours. Ours successfully detects challenging samples that contain overlapped complicated visual relation tuples and provides accurate predicate recognition results. Furthermore, we provide more accurate object and predicate detection results compared to previous baselines. These enhancements are attributed to a hybrid reciprocal architecture that incorporates an alignment module to differentiate triplet features and predict the component relation graph, while receiving supervision at both triplet and component levels.", + "bbox": [ + 511, + 472, + 903, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 651, + 632, + 666 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We integrate a triplet feature alignment module into a hybrid reciprocal transformer to innovate scene graph generation, starting with triplet mask features guide that enhances predicate recognition. We introduce a triplet alignment loss to differentiate multi-role object features in intersecting visual relation tuples, aligning them with triplet features for clearer distinction. Furthermore, our architecture features a reciprocal refinement module in multiple layers for effective fusion and interaction of hybrid aligned features. Experimental results across multiple datasets confirm our superiority over existing methods, marking a substantial advancement in the field of scene graph generation.", + "bbox": [ + 511, + 672, + 903, + 853 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This work is supported by the InnoHK Clusters of the Hong Kong SAR Government via the Hong Kong Centre for Logistics Robotics.", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8960", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 4, 8", + "[2] Guikun Chen, Jin Li, and Wenguan Wang. Scene graph generation with role-playing large language models. arXiv preprint arXiv:2410.15364, 2024. 1, 2", + "[3] Mingfei Chen, Yue Liao, Si Liu, Zhiyuan Chen, Fei Wang, and Chen Qian. Reformulating hoi detection as adaptive set prediction. In CVPR, 2021. 6", + "[4] Minghan Chen, Guikun Chen, Wenguan Wang, and Yi Yang. Hydra-sgg: Hybrid relation assignment for one-stage scene graph generation. arXiv preprint arXiv:2409.10262, 2024. 1, 2, 3, 6", + "[5] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 5", + "[6] Sehwan Choi, Jungho Kim, Hongjae Shin, and Jun Won Choi. Mask2map: Vectorized hd map construction using bird's eye view segmentation masks. arXiv preprint arXiv:2407.13517, 2024. 5", + "[7] Yuren Cong, Michael Ying Yang, and Bodo Rosenhahn. Reltr: Relation transformer for scene graph generation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(9):11169-11183, 2023. 2, 6", + "[8] Alakh Desai, Tz-Ying Wu, Subarna Tripathi, and Nuno Vasconcelos. Learning of visual relations: The devil is in the tails. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15404-15413, 2021. 6", + "[9] Qi Dong, Zhuowen Tu, Haofu Liao, Yuting Zhang, Vijay Mahadevan, and Stefano Soatto. Visual relationship detection using part-and-sum transformers with composite queries. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3550-3559, 2021. 3", + "[10] Raghav Goyal12 and Leonid Sigal123. A simple baseline for weakly-supervised human-centric relation detection. 2021. 6, 7", + "[11] Yuyu Guo, Lianli Gao, Xuanhan Wang, Yuxuan Hu, Xing Xu, Xu Lu, Heng Tao Shen, and Jingkuan Song. From general to specific: Informative scene graph generation via balance adjustment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16383-16392, 2021. 6", + "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3, 6", + "[13] Ruozhen He, Paola Cascante-Bonilla, Ziyan Yang, Alexander C Berg, and Vicente Ordonez. Improved visual grounding through self-consistent explanations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13095-13105, 2024. 1" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Jinbae Im, JeongYeon Nam, Nokyung Park, Hyungmin Lee, and Seunghyun Park. Egtr: Extracting graph from transformer for scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24229-24238, 2024. 1, 2, 6", + "[15] Jingwei Ji, Ranjay Krishna, Li Fei-Fei, and Juan Carlos Niebles. Action genome: Actions as compositions of spatiotemporal scene graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10236-10247, 2020. 2, 6, 7", + "[16] Siddhesh Khandelwal and Leonid Signal. Iterative scene graph generation. Advances in Neural Information Processing Systems, 35:24295-24308, 2022. 1, 2, 3, 4, 6, 7, 8", + "[17] Bumsoo Kim, Junhyun Lee, Jaewoo Kang, Eun-Sol Kim, and Hyunwoo J Kim. Hotr: End-to-end human-object interaction detection with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 74-83, 2021. 6", + "[18] Jongha Kim, Jihwan Park, Jinyoung Park, Jinyoung Kim, Sehyung Kim, and Hyunwoo J Kim. Groupwise query specialization and quality-aware multi-assignment for transformer-based visual relationship detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28160–28169, 2024. 1, 2, 3, 6, 7, 8", + "[19] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 2, 6", + "[20] Sanjoy Kundu and Sathyanarayanan N Aakur. Is-ggt: Iterative scene graph generation with generative transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6292-6301, 2023. 1, 2", + "[21] Feng Li, Hao Zhang, Shilong Liu, Jian Guo, Lionel M Ni, and Lei Zhang. Dn-detr: Accelerate detr training by introducing query denoising. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13619-13627, 2022. 8", + "[22] Jiankai Li, Yunhong Wang, Xiefan Guo, Ruijie Yang, and Weixin Li. Leveraging predicate and triplet learning for scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28369-28379, 2024. 1, 2, 3, 6, 7", + "[23] Lin Li, Long Chen, Yifeng Huang, Zhimeng Zhang, Songyang Zhang, and Jun Xiao. The devil is in the labels: Noisy label correction for robust scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18869-18878, 2022. 6", + "[24] Rongjie Li, Songyang Zhang, Bo Wan, and Xuming He. Bipartite graph network with adaptive message passing for unbiased scene graph generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11109-11119, 2021. 6", + "[25] Rongjie Li, Songyang Zhang, and Xuming He. Sgtr: End-to-end scene graph generation with transformer. In proceedings" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8961", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "of the IEEE/CVF conference on computer vision and pattern recognition, pages 19486-19496, 2022. 1, 2, 3, 6", + "[26] Yikang Li, Wanli Ouyang, Bolei Zhou, Kun Wang, and Xiaogang Wang. Scene graph generation from objects, phrases and region captions. In Proceedings of the IEEE international conference on computer vision, pages 1261-1270, 2017. 7", + "[27] Fanfan Liu, Haoran Wei, Wenzhe Zhao, Guozhen Li, Jingquan Peng, and Zihao Li. Wb-detr: Transformer-based detector without backbone. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2979-2987, 2021. 8", + "[28] Hengyue Liu, Ning Yan, Masood Mortazavi, and Bir Bhanu. Fully convolutional scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11546-11556, 2021. 3", + "[29] Cewu Lu, Ranjay Krishna, Michael Bernstein, and Li Fei-Fei. Visual relationship detection with language priors. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 852-869. Springer, 2016. 2, 7", + "[30] Yichao Lu, Himanshu Rai, Jason Chang, Boris Knyazev, Guangwei Yu, Shashank Shekhar, Graham W Taylor, and Maksims Volkovs. Context-aware scene graph generation with seq2seq transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15931-15941, 2021. 2", + "[31] Jie Ma, Pinghui Wang, Dechen Kong, Zewei Wang, Jun Liu, Hongbin Pei, and Junzhou Zhao. Robust visual question answering: Datasets, methods, and future challenges. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1", + "[32] Tao Ma, Bing Bai, Haozhe Lin, Heyuan Wang, Yu Wang, Lin Luo, and Lu Fang. When visual grounding meets gigapixel-level large-scale scenes: Benchmark and approach. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22119-22128, 2024. 1", + "[33] Nandita Naik, Christopher Potts, and Elisa Kreiss. Context-vqa: Towards context-aware and purposeful visual question answering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2821-2825, 2023. 1", + "[34] Mengshi Qi, Weijian Li, Zhengyuan Yang, Yunhong Wang, and Jiebo Luo. Attentive relational networks for mapping images to scene graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3957-3966, 2019. 2", + "[35] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE transactions on pattern analysis and machine intelligence, 39(6):1137-1149, 2016. 2", + "[36] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 658-666, 2019. 6", + "[37] Shihao Shao, Kaifeng Chen, Arjun Karpur, Qinghua Cui, Andre Araujo, and Bingyi Cao. Global features are all you" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "need for image retrieval and reranking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11036-11046, 2023. 1", + "[38] Suprosanna Shit, Rajat Koner, Bastian Wittmann, Johannes Paetzold, Ivan Ezhov, Hongwei Li, Jiazhen Pan, Sahand Sharifzadeh, Georgios Kaissis, Volker Tresp, et al. Relationformer: A unified framework for image-to-graph generation. In European Conference on Computer Vision, pages 422-439. Springer, 2022. 2", + "[39] Gunnar A Sigurdsson, Gül Varol, Xiaolong Wang, Ali Farhadi, Ivan Laptev, and Abhinav Gupta. Hollywood in homes: Crowdsourcing data collection for activity understanding. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 510-526. Springer, 2016. 6", + "[40] Chull Hwan Song, Jooyoung Yoon, Shunghyun Choi, and Yannis Avrithis. Boosting vision transformers for image retrieval. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 107-117, 2023. 1", + "[41] Gopika Sudhakaran, Devendra Singh Dhami, Kristian Kersting, and Stefan Roth. Vision relation transformer for unbiased scene graph generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21882-21893, 2023. 1, 2, 6", + "[42] Mohammed Suhail, Abhay Mittal, Behjat Siddiquie, Chris Broaddus, Jayan Eledath, Gerard Medioni, and Leonid Sigal. Energy-based learning for scene graph generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13936-13945, 2021. 6", + "[43] Kaihua Tang, Hanwang Zhang, Baoyuan Wu, Wenhan Luo, and Wei Liu. Learning to compose dynamic tree structures for visual contexts. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6619-6628, 2019. 2, 6", + "[44] Kaihua Tang, Yulei Niu, Jianqiang Huang, Jiaxin Shi, and Hanwang Zhang. Unbiased scene graph generation from biased training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3716-3725, 2020. 6", + "[45] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 2, 3, 6", + "[46] Guan Wang, Zhimin Li, Qingchao Chen, and Yang Liu. Oed: Towards one-stage end-to-end dynamic scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27938-27947, 2024. 1", + "[47] Sai Wang, Yutian Lin, and Yu Wu. Omni-q: Omnidirectional scene understanding for unsupervised visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14261-14270, 2024. 1", + "[48] Wenbin Wang, Ruiping Wang, Shiguang Shan, and Xilin Chen. Sketching image gist: Human-mimetic hierarchical scene graph generation. In European conference on computer vision, pages 222-239. Springer, 2020. 2" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8962", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Hui Wu, Min Wang, Wengang Zhou, Zhenbo Lu, and Houqiang Li. Asymmetric feature fusion for image retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11082-11092, 2023. 1", + "[50] Danfei Xu, Yuke Zhu, Christopher B Choy, and Li Fei-Fei. Scene graph generation by iterative message passing. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5410-5419, 2017. 2, 6, 7", + "[51] Jianwei Yang, Jiasen Lu, Stefan Lee, Dhruv Batra, and Devi Parikh. Graph r-cnn for scene graph generation. In Proceedings of the European conference on computer vision (ECCV), pages 670-685, 2018. 7", + "[52] Rowan Zellers, Mark Yatskar, Sam Thomson, and Yejin Choi. Neural motifs: Scene graph parsing with global context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5831-5840, 2018. 2, 6, 7", + "[53] Ji Zhang, Kevin J Shih, Ahmed Elgammal, Andrew Tao, and Bryan Catanzaro. Graphical contrastive losses for scene graph parsing. In CVPR, 2019. 6, 7", + "[54] Jichuan Zhang, Wei Li, Shuang Cheng, Ya-Li Li, and Shengjin Wang. Dynamic object queries for transformer-based incremental object detection. arXiv preprint arXiv:2407.21687, 2024. 8", + "[55] Xi Zhang, Feifei Zhang, and Changsheng Xu. Vqacl: A novel visual question answering continual learning setting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19102-19112, 2023. 1", + "[56] Long Zhao, Liangzhe Yuan, Boqing Gong, Yin Cui, Florian Schroff, Ming-Hsuan Yang, Hartwig Adam, and Ting Liu. Unified visual relationship detection with vision and language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6962-6973, 2023. 1", + "[57] Chaofan Zheng, Xinyu Lyu, Lianli Gao, Bo Dai, and Jingkuan Song. Prototype-based embedding network for scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22783-22792, 2023. 1, 2, 6", + "[58] Yi Zhou, Hui Zhang, Jiaqian Yu, Yifan Yang, Sangil Jung, Seung-In Park, and ByungIn Yoo. Himap: Hybrid representation learning for end-to-end vectorized hd map construction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15396-15406, 2024. 5", + "[59] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable $\\{\\mathrm{detr}\\}$ : Deformable transformers for end-to-end object detection. In International Conference on Learning Representations, 2021. 3, 6, 8" + ], + "bbox": [ + 91, + 90, + 485, + 810 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "8963", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Hybrid Reciprocal Transformer with Triplet Feature Alignment for Scene Graph Generation/11e7a635-4306-4d78-8190-6be9b1f3acac_model.json b/2025/Hybrid Reciprocal Transformer with Triplet Feature Alignment for Scene Graph Generation/11e7a635-4306-4d78-8190-6be9b1f3acac_model.json new file mode 100644 index 0000000000000000000000000000000000000000..798b7e3b1cfcb3703451c9ed2b4a0454fc2c9cc9 --- /dev/null +++ b/2025/Hybrid Reciprocal Transformer with Triplet Feature Alignment for Scene Graph Generation/11e7a635-4306-4d78-8190-6be9b1f3acac_model.json @@ -0,0 +1,2136 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.234, + 0.131, + 0.768, + 0.177 + ], + "angle": 0, + "content": "Hybrid Reciprocal Transformer with Triplet Feature Alignment for Scene Graph Generation" + }, + { + "type": "text", + "bbox": [ + 0.143, + 0.204, + 0.856, + 0.257 + ], + "angle": 0, + "content": "Jiawei Fu, Tiantian Zhang, Kai Chen, and Qi Dou \nDepartment of Computer Science and Engineering, The Chinese University of Hong Kong \njwfu@cse.cuhk.edu.hk" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.326, + 0.486, + 0.735 + ], + "angle": 0, + "content": "Scene graph generation is a pivotal task in computer vision, focusing on comprehensive identification of visual relation tuples embedded within images. The advancement of methods involving triplets has sought to enhance task performance by integrating triplets as contextual features for more precise predicate identification from component level. However, challenges remain due to interference from multi-role objects in overlapping tuples within complex environments, which impairs the model's ability to distinguish and align specific triplet features for reasoning diverse semantics of multi-role objects. To address these issues, we introduce a novel framework that incorporates a triplet alignment model into a hybrid reciprocal transformer architecture, starting from using triplet mask features to guide the learning of component-level relation graphs. To effectively distinguish multi-role objects characterized by overlapping visual relation tuples, we introduce a triplet alignment loss, which provides multi-role objects with aligned features from triplet and helps customize them. Additionally, we explore the inherent connectivity between hybrid aligned triplet and component features through a bidirectional refinement module, which enhances feature interaction and reciprocal reinforcement. Experimental results demonstrate that our model achieves state-of-the-art performance on the Visual Genome and Action Genome datasets, underscoring its effectiveness and adaptability. Project page: hg-sg.github.io." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.768, + 0.222, + 0.784 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Scene graph generation [2, 4, 18] aims to find all possible visual relation tuples \\(\\{\\text{subject-predicate-object}\\}\\) given an image. It has wide applications in various fields, including image retrieval [37, 40, 49], visual grounding [13, 32, 47], and visual question answering [31, 33, 55]. Typically, diverse two-stage [20, 22, 41, 56, 57] and one-stage [4, 16, 18, 25, 46] methods have been developed to solve this prob" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.293, + 0.646, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.646, + 0.293, + 0.776, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.776, + 0.293, + 0.906, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.44, + 0.907, + 0.538 + ], + "angle": 0, + "content": "Figure 1. Complex scenarios often involve overlapping visual relationships, where a single object can assume multiple roles across different triplets. Identifying these multi-role objects is challenging without knowing their aligned triplet features. Moreover, aligned triplets (top) and component (bottom) representations correspond to the same visual relation tuple, offering an opportunity to leverage their inherent relevance for reciprocal refinement." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.551, + 0.909, + 0.734 + ], + "angle": 0, + "content": "lem by organizing regressed component-level subject, object, and predicate with post-processing. However, accurate predicate recognition remains challenging in scene graph generation, as many images incorporate extraneous visual elements. Directly regressing the predicate often introduces noise from features of unaligned visual relation tuples, owing to the lack of definite corresponding subject-object pair features for the visual relation. Accordingly, recent researches [4, 14, 22, 57] have utilized triplet features as prototypes for contextual semantics, combining detailed cues with broader triplet features to enhance predicate recognition accuracy from the triplet level." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Despite advancements, accurately matching predicates to corresponding visual tuples and identifying all potential visual relations in complex environments remains challenging, even with triplet features. As illustrated in Figure 1, a single scene can contain multiple visual relations, with one object assuming multiple roles (either as subject or object) across different tuples. The analogous component-level features of multi-role objects increase overlap and misalignment among visual relation triplets, thereby hindering the model's ability to differentiate among triplets and to further reason diverse semantics of multi-role objects with" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8953" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "out precisely aligned triplet features. Furthermore, previous methodologies [4, 14, 22, 57] often overlook the potential of leveraging connectivity between aligned triplet and component features for mutual enhancement. This oversight can be attributed to the misalignment of triplet and component features [4, 14, 22, 57] or to a two-stage design [22, 57] that is not conducive to joint optimization of hybrid features." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.2, + 0.482, + 0.412 + ], + "angle": 0, + "content": "To this end, it is essential to integrate feature alignment at triplet level, which is key in differentiating complex overlapping visuals. This alignment provides objects performing multiple roles with distinct aligned triplet features, facilitating their distinction within overlapped visual relational tuples. Moreover, to maximize the utility of aligned triplet and component features for identical visual relation tuples, and to investigate the consistency and connectivity between hybrid features to enhance mutual performance, implementing an end-to-end, one-stage scene graph generation framework is essential. This framework should enable bidirectional feature refinement for predicate recognition and object detection, thereby achieving joint optimization of both hybrid triplet and component features." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.415, + 0.483, + 0.687 + ], + "angle": 0, + "content": "In this study, we integrate a triplet alignment model into a hybrid reciprocal transformer architecture for the generation of scene graphs. This integration leverages triplet-level representations as contextual features to enhance component-level scene generation and improve the accuracy of predicate recognition. To effectively differentiate complex, overlapping visual relation tuples with multirole objects at the feature level, we introduce a triplet alignment loss. This loss function promotes similarity between matched triplet and component features, while penalizing similarity among unmatched components from different tuples. It provides triplet-aligned features to multi-role objects to distinguish them in component level. Moreover, we investigate the inherent connectivity between aligned triplet and component features through a bidirectional refinement module. This module amalgamates aligned hybrid features from both triplets and components to facilitate mutual enhancement and reciprocal refinement." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Specifically, we apply an image backbone and transformer encoder to embed the image information across the image. Then the hybrid reciprocal transformer embeds hybrid representations for triplet and component learning, where the triplet-level decoder is utilized to embed the proposal segmentation mask of {subject-predicate-object}, while the component-level decoders are applied to produce the corresponding detected subject, object, and predicate component information. To achieve triplet-guided component-level learning, we follow the iterative learning process [16], the query position embedding of the component-level learning transformer decoder is generated by the processed results of the triplet-level query embedding to assign the feature level between them. To real-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "ize bidirectional refinement, we additionally design a representation cross-conditioned mechanism among layer dimensions of the hybrid reciprocal transformer using attention [45] mechanism. Furthermore, we build a triplet alignment loss function to match triplet-level feature to be consistent with matched component-level feature, which aims to distinguish overlapped triplets with multi-role objects." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.199, + 0.905, + 0.334 + ], + "angle": 0, + "content": "The comprehensive experimental evaluations conducted on Visual Genome dataset [19] and Action Genome dataset [15] substantiate the effectiveness of our framework. In Visual Genome dataset, we demonstrate superior performance by achieving increments of \\(+1.2\\) and \\(+2.3\\) in Recall at 50 and 100, respectively, thereby establishing a new state-of-the-art. Additionally, we achieve highest recorded metrics at both Recall 20 and 50 in Action Genome dataset, confirming its scalability across diverse datasets." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.335, + 0.904, + 0.35 + ], + "angle": 0, + "content": "The contribution of this paper is summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.352, + 0.904, + 0.412 + ], + "angle": 0, + "content": "- We integrate triplet-level and component-level representations within a hybrid reciprocal transformer framework, which uses the triplet mask feature as the proposal to improve predicate recognition performance." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.412, + 0.904, + 0.472 + ], + "angle": 0, + "content": "- We introduce a triplet feature alignment loss to provide multi-role objects in overlapped visual relation with triplet-aligned features to distinguish them and improve predicate recognition efficacy." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.472, + 0.904, + 0.517 + ], + "angle": 0, + "content": "- We leverage the aligned triplet and component feature with reciprocal refinement to explore inherent connectivity between them and optimize scene graphs." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.352, + 0.904, + 0.517 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.534, + 0.655, + 0.549 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.559, + 0.825, + 0.575 + ], + "angle": 0, + "content": "2.1. Two-Stage Scene Graph Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.905, + 0.854 + ], + "angle": 0, + "content": "Scene graph generation has become a prominent research area within the vision community, as evidenced by a wealth of recent studies [2, 4, 16-18, 25]. Classic research [29] design two-stage methods for scene graph generation and use the detected object to organize the relations. Two-stage methods commence with the pre-training of a robust object detector for all classes present in the dataset, typically employing detectors such as Faster R-CNN [35]. Upon establishing the object detector, the visual relation detection module is then constructed utilizing the object data (bound- ing boxes and associated features) extracted from the detector. In the subsequent stage, the object detector's parameters are fixed, and the focus shifts solely to training the relation detection module to construct the scene graph. With the development of artificial intelligence model, various architectures have been implemented for the relation detection module [34, 43, 50-52], such as message passing [50], LSTM [43, 48, 52], transformer [7, 20, 30, 38, 41]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Recent two-stage methods [22, 57] have incorporated triplet features to enhance predicate recognition. However, their effectiveness remains constrained in complex scenes" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8954" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.243 + ], + "angle": 0, + "content": "involving overlapping visual relation tuples with multi-role objects. Additionally, the inherent separate design of object and predicate architectures in these methods limits their ability to effectively utilize the connections between aligned triplet-level and component-level elements. To this end, we leverage a reciprocal one-stage scene graph generation framework which leverages hybrid triplet and component features to enhance each other. It surpasses the inherent constraints of two-stage systems and solves the obfuscation from multi-role objects with a triplet alignment module." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.254, + 0.403, + 0.27 + ], + "angle": 0, + "content": "2.2. One-Stage Scene Graph Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.278, + 0.483, + 0.534 + ], + "angle": 0, + "content": "Contrary to two-stage methods, which primarily suffer from the limited ability of the relation detector to directly utilize features extracted by the detector, one-stage approaches [4, 9, 16-18, 25] overcome this limitation. They achieve this by integrating object detector and graph generation architectures, enabling end-to-end training that facilitates direct access and utilization of extracted features. One-stage methods mainly consists of fully convolutional networks [28] and transformer-based models [4, 9, 16-18, 25]. This configuration facilitates direct interactions between the detection and graph generation modules, significantly enhancing the quality of the resulting scene graphs. The current approaches in one-stage methods primarily concentrate on component-level learning, analyzing subjects, objects, and predicates individually given image feature. These methods [16-18] typically compile these components into tuples \\(\\{\\text{subject-predicate-object}\\}\\) during post-processing." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.536, + 0.483, + 0.686 + ], + "angle": 0, + "content": "Detecting accurate predicates and overcoming obfuscation in overlapped visual relation tuples pose significant challenges for current one-stage methods, primarily due to the lack of subject-object contextual features [22] and varying semantics of multi-role objects. We introduce triplet-level mask features as preliminary proposals to facilitate component-level learning in a one-stage architecture. Additionally, our method incorporates a triplet alignment module that aligns triplet and component features, enhancing the recognition of multi-role objects' diverse semantics." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.701, + 0.182, + 0.716 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.726, + 0.388, + 0.74 + ], + "angle": 0, + "content": "3.1. Formulation and Model Overview" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Scene Graph Generation Formulation. Classic scene graph generation task aims to organize a relation graph \\(\\mathbf{G}'\\) as \\(\\mathbf{G}' = \\{\\mathbf{s}_k, \\mathbf{o}_k, \\mathbf{p}_k\\}_{k=1}^K\\) as a group of visual relation tuples organized from component-level subject \\(\\mathbf{s}_k\\), object \\(\\mathbf{o}_k\\), and predicate \\(\\mathbf{p}_k\\). In our proposed architecture for hybrid representation tailored for scene graph generation, we refine the conventional task formulation by incorporating a triplet mask for each visual relation tuple. Given an image \\(\\mathbf{I}\\), our scene graph \\(\\mathbf{G}\\) is represented as a collection of quadruples that simultaneously encapsulates both triplet-level and tra" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.789, + 0.107 + ], + "angle": 0, + "content": "ditional component-level representations:" + }, + { + "type": "equation", + "bbox": [ + 0.622, + 0.116, + 0.905, + 0.134 + ], + "angle": 0, + "content": "\\[\n\\mathbf {G} = \\left\\{\\mathbf {t} _ {k}, \\mathbf {s} _ {k}, \\mathbf {o} _ {k}, \\mathbf {p} _ {k} \\right\\} _ {k = 1} ^ {K}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.144, + 0.907, + 0.385 + ], + "angle": 0, + "content": "where \\(\\mathbf{t}_k\\) is the triplet-level representation denoting the triplet mask constructed from the classic component-level representation subject \\(\\mathbf{s}_k\\), object \\(\\mathbf{o}_k\\), and predicate \\(\\mathbf{p}_k\\). Each component-level representation of the quadruple, i.e., \\(\\mathbf{s}_k, \\mathbf{o}_k, \\mathbf{p}_k\\), consists of a class label and bounding box. Note the bounding box corresponding to \\(\\mathbf{p}_k\\) is calculated from the center point of its corresponding \\(\\mathbf{s}_k\\) and \\(\\mathbf{o}_k\\). While \\(\\mathbf{t}_k\\) represents the binary mask of the triplet of \\(\\mathbf{s}_k, \\mathbf{o}_k, \\mathbf{p}_k\\) with a class label represent predicate class. This augmentation of triplet-level representation, as compared to traditional methodologies [4, 16-18], significantly provides subject-object pair contextual feature and limits explore space for accurate predicate recognition. Furthermore, it fosters exploration of the connectivity among hybrid triplet-level and component-level representations of same visual relations, thereby facilitating mutual enhancement." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.386, + 0.908, + 0.749 + ], + "angle": 0, + "content": "Model Overview. Figure 2 describes our overall architecture, which mainly consists of multi-scale image encoder and hybrid reciprocal transformer. We begin by utilizing an image encoder equipped with a ResNet backbone [12] and a transformer encoder [45] to extract multi-scale image visual features \\(\\mathcal{I}\\). Based on \\(\\mathcal{I}\\), our hybrid reciprocal transformer incorporates dual-level representations together, triplet mask queries \\(Q_{\\mathrm{t}}\\) for triplet-level representation and subject queries \\(Q_{\\mathrm{s}}\\), object queries \\(Q_{\\mathrm{o}}\\), predicate queries \\(Q_{\\mathrm{p}}\\), for component-level representation, \\(Q\\) means a sequence of queries as \\(Q = \\{q^{i}\\}_{i=1}^{N}\\). To leverage triplet-level representation facilitates guidance of component-level learning for accurate predicate recognition, hybrid reciprocal transformer uses iterative conditional chains for condition the process of \\(Q_{x}; x \\in \\{\\mathrm{s}, \\mathrm{o}, \\mathrm{p}\\}\\) on \\(Q_{\\mathrm{t}}\\) (Sec. 3.2). Furthermore, to enable bidirectional refinement in hybrid reciprocal transformer, we develop a cross-conditioning mechanism, which leverages an attention module that operates across layer dimension of hybrid reciprocal transformer to reason aligned hybrid representations (Sec. 3.2). Finally, we propose alignment loss that leverages triplet-component matched relationship, to provide multi-role objects with distinct aligned triplet feature and recognize overlapped visual relation tuples accurately (Sec. 3.3)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.757, + 0.796, + 0.773 + ], + "angle": 0, + "content": "3.2. Hybrid Reciprocal Transformer" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Hybrid Reciprocal Transformer Architecture. Hybrid reciprocal transformer integrates triplet mask, subject, object, and predicate decoders to process input queries \\( Q_{x} \\); \\( x \\in \\{\\mathrm{t},\\mathrm{s},\\mathrm{o},\\mathrm{p}\\} \\) via cross-attention with cropping image features \\( \\mathcal{I} \\). \\( \\mathcal{I} \\) is cropped through a multi-scale image encoding process, where the final layer of a ResNet-based convolutional network provides a position-aware, flattened image feature representation for a transformer encoder [45, 59]." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8955" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.092, + 0.895, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.356, + 0.908, + 0.427 + ], + "angle": 0, + "content": "Figure 2. Model overview of our method. Hybrid reciprocal transformer deploys four multi-layer decoder structures to embed triplet mask and visual relation tuple components learning using hybrid representations. We illustrate the detailed feature flow of each decoder to visualize two interaction paradigms within this transformer. The first paradigm facilitates triplet iterative guidance and enhances contextual feature integration for predicate recognition at the component level. The second paradigm employs aligned triplet and component features to enable bidirectional refinement of the hybrid triplet-level and component-level representations." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.434, + 0.485, + 0.662 + ], + "angle": 0, + "content": "Specifically, we model the elements of the scene graph—triplet mask, subject, object, and predicate, using multi-layer transformer decoders [1] with iterative processing across its \\(M\\) layers step by step. The decoders are represented as \\(\\mathcal{T}_{\\mathrm{t}}\\), \\(\\mathcal{T}_{\\mathrm{s}}\\), \\(\\mathcal{T}_{\\mathrm{o}}\\), and \\(\\mathcal{T}_{\\mathrm{p}}\\), respectively. In each step \\(j\\), the \\(j\\)-th layer of the triplet-level decoder, \\(\\mathcal{T}_{\\mathrm{t}}^{j}\\), generates masks feature for triplet-wise visual relations in the image. While the \\(j\\)-th layer of each component-level decoder, \\(\\mathcal{T}_{x}^{j}\\) where \\(x \\in \\{\\mathrm{s}, \\mathrm{o}, \\mathrm{p}\\}\\), produces component-level part feature of the scene graph \\(\\mathbf{G}^{j}\\). To realize triplet-level representation as proposals to boost component-level learning and allow hybrid representations to mutually reason and refine each other, our hybrid reciprocal transformer incorporates a triplet-guided relation generation mechanism with bidirectional learning refinement." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Triplet-Guided Relation Learning with Bidirectional Refinement. To implement triplet-guided relation generation and bidirectional triplet-component enhancement, we develop two interaction paradigms between triplet-level and component-level features. The first paradigm involves using mask feature at the triplet level to guide the process of generating visual relations at the component level. This paradigm leverages triplet-level features as proposals to enhance predicate recognition from component-level data by incorporating contextual information from triplet-level representations. The second paradigm is to enable reciprocal sharing of triplet-level and component-level learning representations across the multiple layers of the hybrid reciprocal transformer through the use of an attention mechanism. This allows for streamlined integration of hybrid represen" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.434, + 0.907, + 0.464 + ], + "angle": 0, + "content": "tations and mutual enhancement, as they represent different aspects of aligned identical aligned visual relation tuples." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.465, + 0.907, + 0.541 + ], + "angle": 0, + "content": "Specifically, for a decoder \\(\\mathcal{T}_x; x \\in \\{\\mathrm{t}, \\mathrm{s}, \\mathrm{o}, \\mathrm{p}\\}\\), we derive its output query representation by integrating self-attention on the input queries with positional embeddings and encoder-decoder attention using the image feature \\(\\mathcal{I}\\). In \\(j\\)-th layers, the outputs of decoders are shown as:" + }, + { + "type": "equation", + "bbox": [ + 0.606, + 0.548, + 0.907, + 0.568 + ], + "angle": 0, + "content": "\\[\n\\widehat {Q} _ {x} ^ {j} = \\left\\{\\widehat {q} _ {x} ^ {i, j} \\right\\} _ {i = 1} ^ {N}; x \\in \\{t, s, o, p \\}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.579, + 0.906, + 0.686 + ], + "angle": 0, + "content": "\\(\\widehat{Q}_x^j\\) is produced from two types of conditional inputs from the aforementioned two interaction paradigms for hybrid representations respectively. The first is triplet-guided positional embedding \\(\\bar{P}_x^j = \\{\\bar{p}_x^{i,j}\\}_{i=1}^N\\); \\(x \\in \\{\\mathrm{t}, \\mathrm{s}, \\mathrm{o}, \\mathrm{p}\\}\\) based original position embedding \\(P_x^j = \\{p_x^{i,j}\\}_{i=1}^N\\); \\(x \\in \\{\\mathrm{t}, \\mathrm{s}, \\mathrm{o}, \\mathrm{p}\\}\\) respectively. The second is hybrid reciprocal query \\(\\bar{Q}_x^{j-1}\\); \\(x \\in \\{\\mathrm{t}, \\mathrm{s}, \\mathrm{o}, \\mathrm{p}\\}\\) from \\(j-1\\)-th layer." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.687, + 0.906, + 0.794 + ], + "angle": 0, + "content": "Explicitly, the first involves conditioning component-level decoders \\(\\mathcal{T}_x; x \\in \\{\\mathrm{s,o,p}\\}\\) on the step \\(j\\) outputs of triplet-level decoder \\(\\mathcal{T}_{\\mathrm{t}}\\) using guided positional embedding \\(\\bar{P}_x^j; x \\in \\{\\mathrm{t,s,o,p}\\}\\) and following iterative process [16]. Specifically, \\(\\hat{Q}_x^j = \\mathcal{T}_x^j (\\bar{Q}_x^{j - 1}|\\bar{P}_x^j,\\mathcal{I}); x \\in \\{\\mathrm{t,s,o,p}\\}\\), where triplet decoder using original learnable position embedding \\(\\bar{P}_{\\mathrm{t}}^{j} = P_{\\mathrm{t}}^{j}\\), and triplet-guided mechanism is described as:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.798, + 0.905, + 0.818 + ], + "angle": 0, + "content": "\\[\n\\bar {P} _ {\\mathrm {s}} ^ {j} = P _ {\\mathrm {s}} ^ {j} + \\text {F F N} \\left(\\mathrm {M H A} \\left(P _ {\\mathrm {s}} ^ {j}, \\widehat {Q} _ {\\mathrm {t}} ^ {j}, \\widehat {Q} _ {\\mathrm {t}} ^ {j}\\right)\\right), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.82, + 0.905, + 0.841 + ], + "angle": 0, + "content": "\\[\n\\bar {P} _ {\\mathrm {o}} ^ {j} = P _ {\\mathrm {o}} ^ {j} + \\text {F F N} \\left(\\text {M H A} \\left(P _ {\\mathrm {o}} ^ {j}, \\left[ \\widehat {Q} _ {\\mathrm {t}} ^ {j}, \\widehat {Q} _ {\\mathrm {s}} ^ {j} \\right], \\left[ \\widehat {Q} _ {\\mathrm {t}} ^ {j}, \\widehat {Q} _ {\\mathrm {s}} ^ {j} \\right]\\right)\\right), \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.524, + 0.842, + 0.905, + 0.863 + ], + "angle": 0, + "content": "\\[\n\\bar {P} _ {\\mathrm {p}} ^ {j} = P _ {\\mathrm {p}} ^ {j} + \\text {F F N} \\left(\\mathrm {M H A} \\left(P _ {\\mathrm {p}} ^ {j}, \\left[ \\widehat {Q} _ {\\mathrm {t}} ^ {j}, \\widehat {Q} _ {\\mathrm {s}} ^ {j}, \\widehat {Q} _ {\\mathrm {o}} ^ {j} \\right], \\left[ \\widehat {Q} _ {\\mathrm {t}} ^ {j}, \\widehat {Q} _ {\\mathrm {s}} ^ {j}, \\widehat {Q} _ {\\mathrm {o}} ^ {j} \\right]\\right)\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.87, + 0.909, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathsf{FFN}(\\cdot)\\) is the feed-forward neural network, and \\(\\mathsf{MHA}(\\cdot,\\cdot,\\cdot)\\) is the multi-head attention mechanism for" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8956" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.09, + 0.472, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.2, + 0.483, + 0.271 + ], + "angle": 0, + "content": "Figure 3. Triplet feature alignment loss measures consistency between triplet feature with tripled component-level feature. It encourages similarity between triplet and component feature from identical visual relation tuples to distinguish multi-role object by aligning different triplet feature into component level." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.277, + 0.484, + 0.4 + ], + "angle": 0, + "content": "\\(\\mathsf{Q},\\mathsf{K},\\mathsf{V}\\) respectively, \\([\\cdot ,\\cdot ]\\) means concatenation of the element in it. While the second paradigm requires \\(\\mathcal{T}_{\\mathrm{t}}^{j}\\) conditioned on \\(\\mathcal{T}_x^{j - 1};x\\in \\{\\mathrm{s,o,p}\\}\\) and \\(\\mathcal{T}_x^j;x\\in \\{\\mathrm{s,o,p}\\}\\) conditioned on \\(\\mathcal{T}_{\\mathrm{t}}^{j - 1}\\) to allow bidirectional refinement. This cross-conditioned allows for feature flow enhancement among hybrid learning representations and is achieved through conditional hybrid reciprocal query \\(\\bar{Q}_x^{j - 1};x\\in\\) \\(\\{\\mathrm{t,s,o,p}\\}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.409, + 0.483, + 0.429 + ], + "angle": 0, + "content": "\\[\n\\bar {Q} _ {\\mathrm {t}} ^ {j} = \\widehat {Q} _ {\\mathrm {t}} ^ {j} + \\mathsf {F F N} (\\mathsf {M H A} (\\widehat {Q} _ {t} ^ {j}, [ \\widehat {Q} _ {x} ^ {j} ], [ \\widehat {Q} _ {x} ^ {j} ])); x \\in \\{\\mathrm {s}, \\mathrm {o}, \\mathrm {p} \\}, \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.43, + 0.483, + 0.451 + ], + "angle": 0, + "content": "\\[\n\\bar {Q} _ {x} ^ {j} = \\widehat {Q} _ {x} ^ {j} + \\mathsf {F F N} (\\mathsf {M H A} (\\widehat {Q} _ {x} ^ {j}, \\widehat {Q} _ {\\mathrm {t}} ^ {j}, \\widehat {Q} _ {\\mathrm {t}} ^ {j})); x \\in \\{\\mathrm {s}, 0, \\mathrm {p} \\}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.461, + 0.483, + 0.627 + ], + "angle": 0, + "content": "By implementing the two specified conditional input patterns, we enable the generation of triplet-guided relation graph features within the same layer across various segments of hybrid reciprocal transformer, as well as bidirectional enhancement of hybrid representations across multiple layers of hybrid reciprocal transformer. Subsequently, these representations undergo processing by their respective prediction heads to produce desired hybrid outputs. Then alignment loss is applied for supervision, aiming to align hybrid features from the triplet dimension and enhance the differentiation of overlapping visual relation tuples." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.627, + 0.483, + 0.809 + ], + "angle": 0, + "content": "Prediction Heads. We build prediction heads for hybrid triplet-level and component-level features respectively. For triplet-level feature, we follow [6, 58] to refine with masked attention [5] with the multi-scale image feature. Specifically, triplet masks are generated through an iterative process wherein a dot product is applied across multiple scale dimensions along the channel axis between multi-scale image features \\(\\mathcal{I}\\) and \\(\\widehat{Q}_{\\mathrm{t}}\\). Then the sigmoid function is applied to normalize the segmentation masks. For the component-level detection bounding box representation (subject, object, and predicate), we use MLP to output their corresponding box position and class information directly." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.818, + 0.37, + 0.835 + ], + "angle": 0, + "content": "3.3. Triplet Feature Alignment Loss" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Alignment Loss. To enhance the differentiation of similar component features in multi-role objects that belong to overlapping visual relation tuples, we propose a triplet-component alignment loss. This loss function refines the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "representation of similar component features by motivating multiple queries to delineate distinct roles within a single multi-role object, correlating each query specifically to specific triplet configuration. This mechanism is realized by aligning the computed matched similarity matrix with the ground truth (GT) matching matrix to accurately map between triplet configurations and component representations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.199, + 0.905, + 0.426 + ], + "angle": 0, + "content": "In the example illustrated in Figure 3, we observe a scenario featuring several overlapping visual relationships due to the presence of multi-role objects, such as a \\( \\underline{bag} \\) and \\( \\underline{dog} \\). Here, the alignment loss assigns distinct features for different roles to each multi-role object based on their corresponding triplet-level features. Through the iterative matching and optimization process enabled by multiple layers of hybrid reciprocal transformer, the model promotes distinct representations for the multiple roles of a single object in overlapping triplets, using unique query-level features derived from their respective aligned triplet features. Furthermore, it promotes distinctiveness by encouraging both triplet-level and component-level features of various visual tuples to maintain separation, which aids in accurate recognition of complex and overlapping visual relation tuples." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.427, + 0.907, + 0.598 + ], + "angle": 0, + "content": "Specifically, in \\(j\\)-th layer of hybrid reciprocal transformer, we calculate predicated similarity matrix given the processed triplet-level query \\(\\widehat{Q}_{\\mathrm{t}}^{j}\\) and component-level query \\(\\widehat{Q}_{x}^{j}; x \\in \\{\\mathrm{s}, \\mathrm{o}, \\mathrm{p}\\}\\). We first produce aligned query representation as \\(\\widetilde{Q}_{x}^{j}; x \\in \\{\\mathrm{t}, \\mathrm{s}, \\mathrm{o}, \\mathrm{p}\\}\\) using their respective matching results to identity GT results to maintain bitwise corresponding among query level. Based on the aligned \\(\\widetilde{Q}_{x}^{j}; x \\in \\{\\mathrm{t}, \\mathrm{s}, \\mathrm{o}, \\mathrm{p}\\}\\), we generate a pseudo triplet-level representation \\(\\mathbb{Q}_{\\mathrm{t}}^{j}\\) as the weighted sum of the output component-level representations \\(\\widetilde{Q}_{x}^{j}; x \\in \\{\\mathrm{s}, \\mathrm{o}, \\mathrm{p}\\}\\), and calculate the predicted triplet-level similarities matrix \\(S^{j}\\) of \\(j\\)-th layer:" + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.609, + 0.905, + 0.631 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} ^ {j} = \\mathbb {Q} _ {\\mathrm {t}} ^ {j} \\left(\\widetilde {Q} _ {\\mathrm {t}} ^ {j}\\right) ^ {T}; \\mathbb {Q} _ {\\mathrm {t}} ^ {j} = \\widetilde {Q} _ {\\mathrm {s}} ^ {j} + \\widetilde {Q} _ {\\mathrm {o}} ^ {j} + \\widetilde {Q} _ {\\mathrm {p}} ^ {j}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.642, + 0.905, + 0.703 + ], + "angle": 0, + "content": "While the GT similarity matrix \\( S_{\\mathrm{GT}} \\) is derived from the original matching relationship between the GT triplet-level masks and component-level subjects, objects, and predicates bounding boxes as one-hot formatting." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.906, + 0.839 + ], + "angle": 0, + "content": "Once the GT matrix \\( S_{\\mathrm{GT}} \\) and predicted similarity matrix \\( S^j \\) are obtained, binary cross-entropy loss is applied between them to update model parameters. Incorporate with the iterative optimization and matching process within a hybrid reciprocal transformer framework, the alignment loss enhances the model's capability to differentiate overlapping visual relation tuples by promoting a high degree of similarity between matched component-level and triplet-level representations with GT matching results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.9 + ], + "angle": 0, + "content": "End-to-End Training. In the end-to-end training process for our method, the overall loss function is the sum of the loss from all \\(M\\) layers of hybrid reciprocal transformer, the loss \\(\\mathcal{L}\\) can be written as:" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8957" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.097, + 0.089, + 0.907, + 0.391 + ], + "angle": 0, + "content": "
MethodR@50R@100mR@50mR@100AvgR@50AvgR@100F@50F@100
Two-Stage, X101-FPN Backbone
Motifs [52] [CVPR2018]32.136.95.56.818.821.99.411.5
VCTree [43] [CVPR202]31.836.16.67.719.221.910.912.7
TDE [44] [CVPR2020]19.423.29.311.114.417.212.615.0
EBM [42] [CVPR2021]20.524.79.711.615.118.213.215.8
BPLSA [11] [ICCV2021]21.725.513.515.717.620.616.619.4
DT2-ACBS [8] [ICCV2021]22.024.415.016.318.520.417.819.5
BGNN [24] [CVPR2021]31.035.810.712.720.924.215.918.7
PE-Net [57] [CVPR2023]30.735.212.414.521.624.917.720.5
VETO [41] [CVPR2023]27.531.58.19.517.820.512.514.6
DRM [22] [CVPR2024]34.038.99.011.221.525.014.217.4
One-Stage, ResNet-50 Backbone
RelTR [7] [TPAMI2023]27.530.710.812.319.221.515.517.6
EGTR [14] [CVPR2024]30.234.37.910.119.122.212.515.6
Hydra-SGG [4] [Arxiv2024]28.433.116.019.722.226.420.524.7
One-Stage, ResNet-101 Backbone
RelDN [53] [CVPR2019]30.334.84.45.417.420.17.79.3
AS-Net [3] [CVPR2021]18.721.16.17.212.414.29.210.7
SGTR [25] [CVPR2022]25.126.612.014.618.620.616.218.9
HOTR [17] [CVPR2021]22.427.16.99.714.718.410.614.3
SpeaQHotR [18] [CVPR2024]24.729.19.612.717.220.913.817.7
ISG† [16] [NIPS2022]29.532.17.48.418.520.311.813.3
SpeaQISG† [18] [CVPR2024]32.936.011.814.122.425.117.420.3
ISG [16] [NIPS2022]27.230.115.016.621.123.419.321.4
SpeaQISG [18] [CVPR2024]32.135.515.117.623.626.620.523.5
Ours34.1(+1.2)38.3(+2.3)16.0(+0.9)20.5(+2.9)25.1(+1.5)29.4(+2.8)21.8(+1.3)26.7(+3.2)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.393, + 0.907, + 0.422 + ], + "angle": 0, + "content": "Table 1. Scene graph generation performance on Visual Genome. The best results among models with ResNet-101 backbone are marked in bold. \\(\\dagger\\) denotes the performance without loss re-weighting proposed in [16]." + }, + { + "type": "equation", + "bbox": [ + 0.194, + 0.428, + 0.482, + 0.475 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\sum_ {j = 1} ^ {M} \\left(\\mathcal {L} _ {\\mathrm {t}} ^ {j} + \\mathcal {L} _ {\\mathrm {c}} ^ {j} + \\mathcal {L} _ {\\text {a l i g n}} ^ {j}\\right), \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.194, + 0.475, + 0.369, + 0.49 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {t}} = \\mathcal {L} _ {\\mathrm {t}} ^ {\\text {m a s k}} + \\mathcal {L} _ {\\mathrm {t}} ^ {\\text {d i c e}} + \\mathcal {L} _ {\\mathrm {t}} ^ {\\text {c l s}},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.193, + 0.493, + 0.329, + 0.51 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {c}} = \\mathcal {L} _ {\\mathrm {c}} ^ {\\mathrm {s}} + \\mathcal {L} _ {\\mathrm {c}} ^ {\\mathrm {o}} + \\mathcal {L} _ {\\mathrm {c}} ^ {\\mathrm {p}},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.521, + 0.484, + 0.659 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{t}}\\) denotes triplet mask loss through Hungarian matching, which containing mask loss \\(\\mathcal{L}_{\\mathrm{t}}^{\\mathrm{mask}}\\), dice loss \\(\\mathcal{L}_{\\mathrm{t}}^{\\mathrm{dice}}\\), and mask classification loss \\(\\mathcal{L}_{\\mathrm{t}}^{\\mathrm{cls}}\\), \\(\\mathcal{L}_{\\mathrm{c}}\\) follows the design of [18] and consists of subject loss \\(\\mathcal{L}_{\\mathrm{c}}^{\\mathrm{s}}\\), object loss \\(\\mathcal{L}_{\\mathrm{c}}^{\\mathrm{o}}\\), and predicate loss \\(\\mathcal{L}_{\\mathrm{c}}^{\\mathrm{p}}\\), each of \\(\\mathcal{L}_{\\mathrm{c}}^{\\mathrm{s}}\\), \\(\\mathcal{L}_{\\mathrm{c}}^{\\mathrm{o}}\\), and \\(\\mathcal{L}_{\\mathrm{c}}^{\\mathrm{p}}\\) is constructed from L1 box loss, GIoU loss [36], and classification loss through groupwise query multi-assignment matching [18], \\(\\mathcal{L}_{\\mathrm{align}}^{j}\\) is alignment loss calculated using binary cross-entropy loss between \\(S^{j}\\) and \\(S_{\\mathrm{GT}}\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.672, + 0.216, + 0.688 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.697, + 0.267, + 0.714 + ], + "angle": 0, + "content": "4.1. Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Datasets. We demonstrate the effectiveness of our proposed approach on two benchmark datasets. Visual Genome (VG) [19] is utilized as a benchmark for scene graph generation. We leverage the commonly processed subset from [50], which comprises 108k images across 150 object categories and 50 predicate categories. Action Genome (AG) [15] provides scene graph annotations across 9,848 videos within the Charades dataset [39] for studying human-object interactions. It covers 35 object categories and 25 relation categories, deliberately excluding the class person. Adhering to protocols from previous studies [10], we only utilize frames that have been annotated according to [15]. Frames" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.431, + 0.907, + 0.46 + ], + "angle": 0, + "content": "lacking annotations for person or other objects are excluded to ensure the creation of usable scene graphs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.462, + 0.906, + 0.598 + ], + "angle": 0, + "content": "Evaluation Metrics. Consistent with established methodologies in the literature [50, 52], we evaluate our model using the standard metrics Recall@K (R@K) and Mean Recall@K (mR@K). We further quantify performance by calculating both the arithmetic mean (AvgR@K) [23] and the harmonic mean (F@K) [16]. These additional measures provide a more nuanced assessment of model performance across varying predicate frequencies, ensuring a comprehensive evaluation of the balanced performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.907, + 0.81 + ], + "angle": 0, + "content": "Implementation Details. We utilize a ResNet-101 backbone [12] paired with a six-layer Transformer encoder. For each part in \\(\\mathcal{T}_x\\); \\(x \\in \\{\\mathrm{t}, \\mathrm{s}, \\mathrm{o}, \\mathrm{p}\\}\\) of the hybrid reciprocal transformer, we deploy a six-layer transformer decoder [45, 59] architecture. The query number for component-level learning is 600 for \\(\\mathcal{T}_x^j\\); \\(x \\in \\{\\mathrm{s}, \\mathrm{o}, \\mathrm{p}\\}\\), and 300 for \\(\\mathcal{T}_{\\mathrm{tn}}^j\\). The triplet matching strategy employs a naive Hungarian algorithm, optimizing for mask loss, dice loss, and classification loss. For component-level matching, we adopt a grouping matching strategy as proposed in [18], ensuring consistency in matched indices across components. The model undergoes pretraining on VG and AG datasets, focusing on object detection and semantic segmentation, respectively. Experiments are conducted on four A40 GPUs with batch size 16." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.817, + 0.868, + 0.832 + ], + "angle": 0, + "content": "4.2. Comparison Results with State-of-the-art" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Visual Genome Dataset. In Table 1, we report the performance of our method on VG dataset, the results show that our method achieves new state-of-the-art performance on those one-stage methods based on ResNet-101 backbone." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8958" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.104, + 0.089, + 0.473, + 0.272 + ], + "angle": 0, + "content": "
MethodR@20/50mR@20/50F@20/50
Two-Stage, X101-FPN Backbone
VRD [29]10.3 / 10.9-
Freq Prior [52]24.0 / 24.9--
IMP [50]23.9 / 25.5--
MSDN [26]24.0 / 25.6--
Graph R-CNN [51]24.1 / 25.8--
RelDN [53]25.0 / 26.2--
SimpleBase [10]27.9 / 30.48.3 / 9.112.8 / 14.0
One-Stage, ResNet-101 Backbone
ISG [16]31.0 / 37.510.9 / 13.216.1 / 19.5
\\( ISG_{\\dagger} \\) [16]30.1 / 36.519.2 / 22.323.4 / 27.7
Spea\\( Q_{\\text{ISG}} \\) [18]30.8 / 38.011.7 / 13.817.0 / 20.2
Spea\\( Q_{\\text{ISG}_{\\dagger}} \\) [18]31.8 / 39.420.9 / 23.125.2 / 29.1
Ours32.0 / 40.221.3 / 23.625.5 / 29.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.274, + 0.483, + 0.303 + ], + "angle": 0, + "content": "Table 2. Scene graph generation performance on Action Genome. Best performance with bold." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.31, + 0.484, + 0.521 + ], + "angle": 0, + "content": "Our model demonstrates significant enhancements with increases of \\(+1.2\\), \\(+2.3\\), \\(+0.9\\), and \\(+2.9\\) in the R@50/100 and mR@50/100 metrics. These enhancements can be credited to our pioneering hybrid reciprocal architecture, which integrates reasoning across aligned triplets and component features. Furthermore, the synergistic interplay between the two-tiered representation of visual relation tuple significantly augments the overall performance. When compared with two-stage methods employing the X101-FPN backbone, our method achieves state-of-the-art results in R@K, as evidenced by its comparison with the DRM model [22]. Moreover, owing to our one-stage design to jointly optimize predicate recognition and object detection, we substantially outperform two-stage approaches in mR@K." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.522, + 0.484, + 0.718 + ], + "angle": 0, + "content": "Action Genome Dataset. We additionally report R@20/50, mR20/50, and F@20/50 values on the Action Genome dataset [15] in Table 2. Consistent with the findings observed on the Visual Genome dataset, we establish a new benchmark for state-of-the-art performance across all utilized metrics, including R@20/50, mR@20/50, and F@20/50. This advancement underscores the efficacy of our hybrid reciprocal architecture with triplet alignment in enhancing scene graph generation capabilities. Moreover, the improvement in performance across various datasets demonstrates the robust scalability and adaptability of the hybrid learning framework, which does not necessitate specific modifications to be effective across different contexts." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.727, + 0.347, + 0.743 + ], + "angle": 0, + "content": "4.3. Ablation Study and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Alignment Loss. We apply an alignment loss between hybrid triplet and component features to enhance the distinction of multi-role objects in overlapping triplets, by aligning triplet features into the component level. We investigate the effect of the triplet feature alignment loss by setting different weights in the end-to-end training overall loss function. The results metrics in the VG dataset using different weights (0.0, 0.1, 0.2, 0.3, 0.4, 0.5, 1.0) are reported at Figure 4, the weight used in our vanilla model is 0.2. Our findings suggest that increasing the weight of alignment loss" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.094, + 0.898, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.191, + 0.898, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.291, + 0.907, + 0.332 + ], + "angle": 0, + "content": "Figure 4. Scene graph generation performance comparison with Visual Genome dataset using different parameters of weight of alignment loss in overall training loss." + }, + { + "type": "table", + "bbox": [ + 0.589, + 0.332, + 0.832, + 0.389 + ], + "angle": 0, + "content": "
αc:tR@50R@100mR@50mR@100
0.532.135.316.522.3
1.033.235.716.121.2
2.033.736.816.422.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.39, + 0.907, + 0.42 + ], + "angle": 0, + "content": "Table 3. Effect of matching triplet-level and component-level together with different ratio \\( {\\alpha }_{\\mathrm{c} : \\mathrm{t}} \\) between component and triplet cost." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.43, + 0.905, + 0.492 + ], + "angle": 0, + "content": "initially enhances performance as desired, but subsequently deteriorates. This pattern of performance degradation at higher weights of alignment loss indicates that excessive emphasis on alignment loss may hinder learning efficacy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.492, + 0.907, + 0.84 + ], + "angle": 0, + "content": "Hybrid Representation Guidance Order and Reciprocal Refinement. In our framework, we traditionally utilize the processed query for triplet-level segmentation masks to condition the generation of queries for component-level representations. To explore the impact of reversing this sequence, we instead employ the component-level query as a condition for generating triplet-level representations. The results for this variant, denoted as R@50/100 and mR@50/100, are 33.2, 36.9, 15.3, and 18.9, respectively, when compared with our vanilla model, as detailed in Table 1, it is evident that utilizing the triplet-level query as a condition for component-level learning yields superior performance. This enhancement can be attributed to the global information embedded in the triplet-level query feature, which effectively enhances the learning of local information in the component-level elements—subject, object, and predicate. Additionally, we explore the impact of reciprocal refinement by omitting this feature from our vanilla model. The performance metrics on VG dataset for R@50/100 and mR@50/100 are 32.4, 36.3, 14.4, and 18.5, respectively. The observed decrease in performance underscores the significance of reciprocal refinement in integrating hybrid triplet and component features effectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Grouping Hybrid Representations in Uniform Matcher. Inspired by [18], we explore integrating hybrid representations using a single Hungarian matching mechanism to align them directly. Table 3 presents a comparative anal" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8959" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.096, + 0.904, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.26, + 0.905, + 0.285 + ], + "angle": 0, + "content": "Figure 5. Qualitative analysis of our method compared with the baseline models, SpeaQ [18] with ISG [16], indicates that our method yields significantly enhanced accuracy in the prediction of predicates and object detection. Red marks error." + }, + { + "type": "table", + "bbox": [ + 0.157, + 0.292, + 0.416, + 0.379 + ], + "angle": 0, + "content": "
NQtNQcR@50R@100mR@50mR@100
30030033.837.315.218.3
40033.837.415.318.9
50033.937.615.719.4
60030033.937.815.720.1
40034.037.915.920.5
50034.238.216.221.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.381, + 0.483, + 0.409 + ], + "angle": 0, + "content": "Table 4. Scene graph performance on VG dataset with different number of triplet-level and component-level queries." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.418, + 0.483, + 0.553 + ], + "angle": 0, + "content": "ysis of the outcomes, employing varied matching ratios between component and triplet levels. The findings indicate that merging the hybrid representations via a Hungarian matching mechanism yields inferior results compared to our vanilla model. Specifically, reducing component-level matching negatively impacts performance. This decline is attributed to the stringent constraints imposed by grouping two distinct hybrid representations, which restricts their internal matching and adversely affects overall performance." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.553, + 0.483, + 0.674 + ], + "angle": 0, + "content": "Query Number. We examine the impact of query number for hybrid representations (\\(N_{Q\\mathrm{t}}\\) for triplet and \\(N_{Q\\mathrm{c}}\\) for component) in Table 4. The comparative analysis indicates that incremental increases in query numbers consistently enhance outcome performance. This improvement adheres to fundamental principles observed in transformer-based approaches for visual tasks [1, 21, 27, 59] by providing richer representational capacity and coverage [54]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.675, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Matching of Hybrid Representations in Inference Stage. In the inference stage, our vanilla model only uses the bitwise multiply results from the component-level subject, object, and predicate scores. We compare two deformations in Table 5: the first directly calculates multiply of bitwise triplet-level mask scores together, and the second uses Hungarian matching to align two-level output and then multiply them together. Their corresponding results are shown in Table 5. The comparison results show that directly multiply component-level with triplet-level score together will definitely cause negative performance, since the two-level output representation cannot be directly assigned to the specific result subject, object, and predicate tuple. We also find that after the Hungarian matching, the multiplied scores of the hybrid two-level representation will improve performance" + }, + { + "type": "table", + "bbox": [ + 0.553, + 0.292, + 0.864, + 0.348 + ], + "angle": 0, + "content": "
Inference TypeR@50R@100mR@50mR@100
Component Only34.138.316.020.5
Component × Triplet32.535.614.616.5
Component × Triplet34.238.316.120.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.35, + 0.905, + 0.391 + ], + "angle": 0, + "content": "Table 5. Comparison results among inference matching between hybrid representations. \\(\\times\\) means bitwise multiply and \\(\\star\\) denotes multiply after matching." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.399, + 0.905, + 0.444 + ], + "angle": 0, + "content": "a bit. We suppose this improvement came from the final aligned hybrid representation by the alignment loss will boost the final scene graph generation performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.449, + 0.696, + 0.464 + ], + "angle": 0, + "content": "4.4. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.473, + 0.905, + 0.639 + ], + "angle": 0, + "content": "Figure 5 presents qualitative examples comparing prediction results from the SpeaQ [18] and ours. Ours successfully detects challenging samples that contain overlapped complicated visual relation tuples and provides accurate predicate recognition results. Furthermore, we provide more accurate object and predicate detection results compared to previous baselines. These enhancements are attributed to a hybrid reciprocal architecture that incorporates an alignment module to differentiate triplet features and predict the component relation graph, while receiving supervision at both triplet and component levels." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.652, + 0.633, + 0.667 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.905, + 0.854 + ], + "angle": 0, + "content": "We integrate a triplet feature alignment module into a hybrid reciprocal transformer to innovate scene graph generation, starting with triplet mask features guide that enhances predicate recognition. We introduce a triplet alignment loss to differentiate multi-role object features in intersecting visual relation tuples, aligning them with triplet features for clearer distinction. Furthermore, our architecture features a reciprocal refinement module in multiple layers for effective fusion and interaction of hybrid aligned features. Experimental results across multiple datasets confirm our superiority over existing methods, marking a substantial advancement in the field of scene graph generation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Acknowledgement. This work is supported by the InnoHK Clusters of the Hong Kong SAR Government via the Hong Kong Centre for Logistics Robotics." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8960" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.183 + ], + "angle": 0, + "content": "[1] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.186, + 0.484, + 0.228 + ], + "angle": 0, + "content": "[2] Guikun Chen, Jin Li, and Wenguan Wang. Scene graph generation with role-playing large language models. arXiv preprint arXiv:2410.15364, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.231, + 0.483, + 0.272 + ], + "angle": 0, + "content": "[3] Mingfei Chen, Yue Liao, Si Liu, Zhiyuan Chen, Fei Wang, and Chen Qian. Reformulating hoi detection as adaptive set prediction. In CVPR, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.274, + 0.483, + 0.328 + ], + "angle": 0, + "content": "[4] Minghan Chen, Guikun Chen, Wenguan Wang, and Yi Yang. Hydra-sgg: Hybrid relation assignment for one-stage scene graph generation. arXiv preprint arXiv:2409.10262, 2024. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.331, + 0.483, + 0.401 + ], + "angle": 0, + "content": "[5] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.403, + 0.483, + 0.457 + ], + "angle": 0, + "content": "[6] Sehwan Choi, Jungho Kim, Hongjae Shin, and Jun Won Choi. Mask2map: Vectorized hd map construction using bird's eye view segmentation masks. arXiv preprint arXiv:2407.13517, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.46, + 0.483, + 0.514 + ], + "angle": 0, + "content": "[7] Yuren Cong, Michael Ying Yang, and Bodo Rosenhahn. Reltr: Relation transformer for scene graph generation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(9):11169-11183, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.517, + 0.483, + 0.572 + ], + "angle": 0, + "content": "[8] Alakh Desai, Tz-Ying Wu, Subarna Tripathi, and Nuno Vasconcelos. Learning of visual relations: The devil is in the tails. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15404-15413, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.575, + 0.483, + 0.644 + ], + "angle": 0, + "content": "[9] Qi Dong, Zhuowen Tu, Haofu Liao, Yuting Zhang, Vijay Mahadevan, and Stefano Soatto. Visual relationship detection using part-and-sum transformers with composite queries. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3550-3559, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.647, + 0.483, + 0.686 + ], + "angle": 0, + "content": "[10] Raghav Goyal12 and Leonid Sigal123. A simple baseline for weakly-supervised human-centric relation detection. 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.689, + 0.483, + 0.771 + ], + "angle": 0, + "content": "[11] Yuyu Guo, Lianli Gao, Xuanhan Wang, Yuxuan Hu, Xing Xu, Xu Lu, Heng Tao Shen, and Jingkuan Song. From general to specific: Informative scene graph generation via balance adjustment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16383-16392, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.775, + 0.483, + 0.83 + ], + "angle": 0, + "content": "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[13] Ruozhen He, Paola Cascante-Bonilla, Ziyan Yang, Alexander C Berg, and Vicente Ordonez. Improved visual grounding through self-consistent explanations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13095-13105, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[14] Jinbae Im, JeongYeon Nam, Nokyung Park, Hyungmin Lee, and Seunghyun Park. Egtr: Extracting graph from transformer for scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24229-24238, 2024. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[15] Jingwei Ji, Ranjay Krishna, Li Fei-Fei, and Juan Carlos Niebles. Action genome: Actions as compositions of spatiotemporal scene graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10236-10247, 2020. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.235, + 0.905, + 0.276 + ], + "angle": 0, + "content": "[16] Siddhesh Khandelwal and Leonid Signal. Iterative scene graph generation. Advances in Neural Information Processing Systems, 35:24295-24308, 2022. 1, 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.278, + 0.907, + 0.348 + ], + "angle": 0, + "content": "[17] Bumsoo Kim, Junhyun Lee, Jaewoo Kang, Eun-Sol Kim, and Hyunwoo J Kim. Hotr: End-to-end human-object interaction detection with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 74-83, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.35, + 0.905, + 0.432 + ], + "angle": 0, + "content": "[18] Jongha Kim, Jihwan Park, Jinyoung Park, Jinyoung Kim, Sehyung Kim, and Hyunwoo J Kim. Groupwise query specialization and quality-aware multi-assignment for transformer-based visual relationship detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28160–28169, 2024. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.434, + 0.905, + 0.516 + ], + "angle": 0, + "content": "[19] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.519, + 0.905, + 0.585 + ], + "angle": 0, + "content": "[20] Sanjoy Kundu and Sathyanarayanan N Aakur. Is-ggt: Iterative scene graph generation with generative transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6292-6301, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.589, + 0.905, + 0.658 + ], + "angle": 0, + "content": "[21] Feng Li, Hao Zhang, Shilong Liu, Jian Guo, Lionel M Ni, and Lei Zhang. Dn-detr: Accelerate detr training by introducing query denoising. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13619-13627, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.661, + 0.905, + 0.73 + ], + "angle": 0, + "content": "[22] Jiankai Li, Yunhong Wang, Xiefan Guo, Ruijie Yang, and Weixin Li. Leveraging predicate and triplet learning for scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28369-28379, 2024. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.732, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[23] Lin Li, Long Chen, Yifeng Huang, Zhimeng Zhang, Songyang Zhang, and Jun Xiao. The devil is in the labels: Noisy label correction for robust scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18869-18878, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.802, + 0.907, + 0.871 + ], + "angle": 0, + "content": "[24] Rongjie Li, Songyang Zhang, Bo Wan, and Xuming He. Bipartite graph network with adaptive message passing for unbiased scene graph generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11109-11119, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[25] Rongjie Li, Songyang Zhang, and Xuming He. Sgtr: End-to-end scene graph generation with transformer. In proceedings" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8961" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "of the IEEE/CVF conference on computer vision and pattern recognition, pages 19486-19496, 2022. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.122, + 0.483, + 0.188 + ], + "angle": 0, + "content": "[26] Yikang Li, Wanli Ouyang, Bolei Zhou, Kun Wang, and Xiaogang Wang. Scene graph generation from objects, phrases and region captions. In Proceedings of the IEEE international conference on computer vision, pages 1261-1270, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.483, + 0.259 + ], + "angle": 0, + "content": "[27] Fanfan Liu, Haoran Wei, Wenzhe Zhao, Guozhen Li, Jingquan Peng, and Zihao Li. Wb-detr: Transformer-based detector without backbone. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2979-2987, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.261, + 0.482, + 0.316 + ], + "angle": 0, + "content": "[28] Hengyue Liu, Ning Yan, Masood Mortazavi, and Bir Bhanu. Fully convolutional scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11546-11556, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.317, + 0.483, + 0.385 + ], + "angle": 0, + "content": "[29] Cewu Lu, Ranjay Krishna, Michael Bernstein, and Li Fei-Fei. Visual relationship detection with language priors. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 852-869. Springer, 2016. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.386, + 0.483, + 0.467 + ], + "angle": 0, + "content": "[30] Yichao Lu, Himanshu Rai, Jason Chang, Boris Knyazev, Guangwei Yu, Shashank Shekhar, Graham W Taylor, and Maksims Volkovs. Context-aware scene graph generation with seq2seq transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15931-15941, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.483, + 0.537 + ], + "angle": 0, + "content": "[31] Jie Ma, Pinghui Wang, Dechen Kong, Zewei Wang, Jun Liu, Hongbin Pei, and Junzhou Zhao. Robust visual question answering: Datasets, methods, and future challenges. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.483, + 0.607 + ], + "angle": 0, + "content": "[32] Tao Ma, Bing Bai, Haozhe Lin, Heyuan Wang, Yu Wang, Lin Luo, and Lu Fang. When visual grounding meets gigapixel-level large-scale scenes: Benchmark and approach. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22119-22128, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.609, + 0.483, + 0.663 + ], + "angle": 0, + "content": "[33] Nandita Naik, Christopher Potts, and Elisa Kreiss. Context-vqa: Towards context-aware and purposeful visual question answering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2821-2825, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.665, + 0.483, + 0.734 + ], + "angle": 0, + "content": "[34] Mengshi Qi, Weijian Li, Zhengyuan Yang, Yunhong Wang, and Jiebo Luo. Attentive relational networks for mapping images to scene graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3957-3966, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.483, + 0.789 + ], + "angle": 0, + "content": "[35] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE transactions on pattern analysis and machine intelligence, 39(6):1137-1149, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[36] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 658-666, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[37] Shihao Shao, Kaifeng Chen, Arjun Karpur, Qinghua Cui, Andre Araujo, and Bingyi Cao. Global features are all you" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.908, + 0.135 + ], + "angle": 0, + "content": "need for image retrieval and reranking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11036-11046, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.137, + 0.906, + 0.219 + ], + "angle": 0, + "content": "[38] Suprosanna Shit, Rajat Koner, Bastian Wittmann, Johannes Paetzold, Ivan Ezhov, Hongwei Li, Jiazhen Pan, Sahand Sharifzadeh, Georgios Kaissis, Volker Tresp, et al. Relationformer: A unified framework for image-to-graph generation. In European Conference on Computer Vision, pages 422-439. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.906, + 0.316 + ], + "angle": 0, + "content": "[39] Gunnar A Sigurdsson, Gül Varol, Xiaolong Wang, Ali Farhadi, Ivan Laptev, and Abhinav Gupta. Hollywood in homes: Crowdsourcing data collection for activity understanding. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 510-526. Springer, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.319, + 0.906, + 0.387 + ], + "angle": 0, + "content": "[40] Chull Hwan Song, Jooyoung Yoon, Shunghyun Choi, and Yannis Avrithis. Boosting vision transformers for image retrieval. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 107-117, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.906, + 0.46 + ], + "angle": 0, + "content": "[41] Gopika Sudhakaran, Devendra Singh Dhami, Kristian Kersting, and Stefan Roth. Vision relation transformer for unbiased scene graph generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21882-21893, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.462, + 0.906, + 0.531 + ], + "angle": 0, + "content": "[42] Mohammed Suhail, Abhay Mittal, Behjat Siddiquie, Chris Broaddus, Jayan Eledath, Gerard Medioni, and Leonid Sigal. Energy-based learning for scene graph generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13936-13945, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.533, + 0.906, + 0.601 + ], + "angle": 0, + "content": "[43] Kaihua Tang, Hanwang Zhang, Baoyuan Wu, Wenhan Luo, and Wei Liu. Learning to compose dynamic tree structures for visual contexts. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6619-6628, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.603, + 0.906, + 0.671 + ], + "angle": 0, + "content": "[44] Kaihua Tang, Yulei Niu, Jianqiang Huang, Jiaxin Shi, and Hanwang Zhang. Unbiased scene graph generation from biased training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3716-3725, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.674, + 0.906, + 0.702 + ], + "angle": 0, + "content": "[45] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.704, + 0.906, + 0.771 + ], + "angle": 0, + "content": "[46] Guan Wang, Zhimin Li, Qingchao Chen, and Yang Liu. Oed: Towards one-stage end-to-end dynamic scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27938-27947, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.775, + 0.906, + 0.843 + ], + "angle": 0, + "content": "[47] Sai Wang, Yutian Lin, and Yu Wu. Omni-q: Omnidirectional scene understanding for unsupervised visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14261-14270, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[48] Wenbin Wang, Ruiping Wang, Shiguang Shan, and Xilin Chen. Sketching image gist: Human-mimetic hierarchical scene graph generation. In European conference on computer vision, pages 222-239. Springer, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8962" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.161 + ], + "angle": 0, + "content": "[49] Hui Wu, Min Wang, Wengang Zhou, Zhenbo Lu, and Houqiang Li. Asymmetric feature fusion for image retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11082-11092, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[50] Danfei Xu, Yuke Zhu, Christopher B Choy, and Li Fei-Fei. Scene graph generation by iterative message passing. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5410-5419, 2017. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.275 + ], + "angle": 0, + "content": "[51] Jianwei Yang, Jiasen Lu, Stefan Lee, Dhruv Batra, and Devi Parikh. Graph r-cnn for scene graph generation. In Proceedings of the European conference on computer vision (ECCV), pages 670-685, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.482, + 0.344 + ], + "angle": 0, + "content": "[52] Rowan Zellers, Mark Yatskar, Sam Thomson, and Yejin Choi. Neural motifs: Scene graph parsing with global context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5831-5840, 2018. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.346, + 0.482, + 0.389 + ], + "angle": 0, + "content": "[53] Ji Zhang, Kevin J Shih, Ahmed Elgammal, Andrew Tao, and Bryan Catanzaro. Graphical contrastive losses for scene graph parsing. In CVPR, 2019. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.39, + 0.482, + 0.444 + ], + "angle": 0, + "content": "[54] Jichuan Zhang, Wei Li, Shuang Cheng, Ya-Li Li, and Shengjin Wang. Dynamic object queries for transformer-based incremental object detection. arXiv preprint arXiv:2407.21687, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.446, + 0.482, + 0.514 + ], + "angle": 0, + "content": "[55] Xi Zhang, Feifei Zhang, and Changsheng Xu. Vqacl: A novel visual question answering continual learning setting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19102-19112, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.517, + 0.482, + 0.599 + ], + "angle": 0, + "content": "[56] Long Zhao, Liangzhe Yuan, Boqing Gong, Yin Cui, Florian Schroff, Ming-Hsuan Yang, Hartwig Adam, and Ting Liu. Unified visual relationship detection with vision and language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6962-6973, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.601, + 0.482, + 0.669 + ], + "angle": 0, + "content": "[57] Chaofan Zheng, Xinyu Lyu, Lianli Gao, Bo Dai, and Jingkuan Song. Prototype-based embedding network for scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22783-22792, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.671, + 0.482, + 0.753 + ], + "angle": 0, + "content": "[58] Yi Zhou, Hui Zhang, Jiaqian Yu, Yifan Yang, Sangil Jung, Seung-In Park, and ByungIn Yoo. Himap: Hybrid representation learning for end-to-end vectorized hd map construction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15396-15406, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.756, + 0.482, + 0.811 + ], + "angle": 0, + "content": "[59] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable \\(\\{\\mathrm{detr}\\}\\): Deformable transformers for end-to-end object detection. In International Conference on Learning Representations, 2021. 3, 6, 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.811 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8963" + } + ] +] \ No newline at end of file diff --git a/2025/Hybrid Reciprocal Transformer with Triplet Feature Alignment for Scene Graph Generation/11e7a635-4306-4d78-8190-6be9b1f3acac_origin.pdf b/2025/Hybrid Reciprocal Transformer with Triplet Feature Alignment for Scene Graph Generation/11e7a635-4306-4d78-8190-6be9b1f3acac_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..59482dfcd0d6c6ee944b21bf99d04fe9c90c342a --- /dev/null +++ b/2025/Hybrid Reciprocal Transformer with Triplet Feature Alignment for Scene Graph Generation/11e7a635-4306-4d78-8190-6be9b1f3acac_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8e1e07e56abd521db50714543158066c099c5823dd8bab2e3ecd7b24a402b1f0 +size 725960 diff --git a/2025/Hybrid Reciprocal Transformer with Triplet Feature Alignment for Scene Graph Generation/full.md b/2025/Hybrid Reciprocal Transformer with Triplet Feature Alignment for Scene Graph Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..420e7314e13afdd5875b34acce61313990b83977 --- /dev/null +++ b/2025/Hybrid Reciprocal Transformer with Triplet Feature Alignment for Scene Graph Generation/full.md @@ -0,0 +1,306 @@ +# Hybrid Reciprocal Transformer with Triplet Feature Alignment for Scene Graph Generation + +Jiawei Fu, Tiantian Zhang, Kai Chen, and Qi Dou +Department of Computer Science and Engineering, The Chinese University of Hong Kong +jwfu@cse.cuhk.edu.hk + +# Abstract + +Scene graph generation is a pivotal task in computer vision, focusing on comprehensive identification of visual relation tuples embedded within images. The advancement of methods involving triplets has sought to enhance task performance by integrating triplets as contextual features for more precise predicate identification from component level. However, challenges remain due to interference from multi-role objects in overlapping tuples within complex environments, which impairs the model's ability to distinguish and align specific triplet features for reasoning diverse semantics of multi-role objects. To address these issues, we introduce a novel framework that incorporates a triplet alignment model into a hybrid reciprocal transformer architecture, starting from using triplet mask features to guide the learning of component-level relation graphs. To effectively distinguish multi-role objects characterized by overlapping visual relation tuples, we introduce a triplet alignment loss, which provides multi-role objects with aligned features from triplet and helps customize them. Additionally, we explore the inherent connectivity between hybrid aligned triplet and component features through a bidirectional refinement module, which enhances feature interaction and reciprocal reinforcement. Experimental results demonstrate that our model achieves state-of-the-art performance on the Visual Genome and Action Genome datasets, underscoring its effectiveness and adaptability. Project page: hg-sg.github.io. + +# 1. Introduction + +Scene graph generation [2, 4, 18] aims to find all possible visual relation tuples $\{\text{subject-predicate-object}\}$ given an image. It has wide applications in various fields, including image retrieval [37, 40, 49], visual grounding [13, 32, 47], and visual question answering [31, 33, 55]. Typically, diverse two-stage [20, 22, 41, 56, 57] and one-stage [4, 16, 18, 25, 46] methods have been developed to solve this prob + +![](images/34955259c84ebc6373acc83bd9bc11c3dffdbcc6c8b7f09e262c52e43188ef1e.jpg) +Figure 1. Complex scenarios often involve overlapping visual relationships, where a single object can assume multiple roles across different triplets. Identifying these multi-role objects is challenging without knowing their aligned triplet features. Moreover, aligned triplets (top) and component (bottom) representations correspond to the same visual relation tuple, offering an opportunity to leverage their inherent relevance for reciprocal refinement. + +![](images/d74d360bf4b0558d3918ef15d96b17aedce03e12efd035371689ff2b184cea30.jpg) + +![](images/bdfa7616dd081249b8131875bc92132dd428c31269a14711fc389ae9048e90cf.jpg) + +lem by organizing regressed component-level subject, object, and predicate with post-processing. However, accurate predicate recognition remains challenging in scene graph generation, as many images incorporate extraneous visual elements. Directly regressing the predicate often introduces noise from features of unaligned visual relation tuples, owing to the lack of definite corresponding subject-object pair features for the visual relation. Accordingly, recent researches [4, 14, 22, 57] have utilized triplet features as prototypes for contextual semantics, combining detailed cues with broader triplet features to enhance predicate recognition accuracy from the triplet level. + +Despite advancements, accurately matching predicates to corresponding visual tuples and identifying all potential visual relations in complex environments remains challenging, even with triplet features. As illustrated in Figure 1, a single scene can contain multiple visual relations, with one object assuming multiple roles (either as subject or object) across different tuples. The analogous component-level features of multi-role objects increase overlap and misalignment among visual relation triplets, thereby hindering the model's ability to differentiate among triplets and to further reason diverse semantics of multi-role objects with + +out precisely aligned triplet features. Furthermore, previous methodologies [4, 14, 22, 57] often overlook the potential of leveraging connectivity between aligned triplet and component features for mutual enhancement. This oversight can be attributed to the misalignment of triplet and component features [4, 14, 22, 57] or to a two-stage design [22, 57] that is not conducive to joint optimization of hybrid features. + +To this end, it is essential to integrate feature alignment at triplet level, which is key in differentiating complex overlapping visuals. This alignment provides objects performing multiple roles with distinct aligned triplet features, facilitating their distinction within overlapped visual relational tuples. Moreover, to maximize the utility of aligned triplet and component features for identical visual relation tuples, and to investigate the consistency and connectivity between hybrid features to enhance mutual performance, implementing an end-to-end, one-stage scene graph generation framework is essential. This framework should enable bidirectional feature refinement for predicate recognition and object detection, thereby achieving joint optimization of both hybrid triplet and component features. + +In this study, we integrate a triplet alignment model into a hybrid reciprocal transformer architecture for the generation of scene graphs. This integration leverages triplet-level representations as contextual features to enhance component-level scene generation and improve the accuracy of predicate recognition. To effectively differentiate complex, overlapping visual relation tuples with multirole objects at the feature level, we introduce a triplet alignment loss. This loss function promotes similarity between matched triplet and component features, while penalizing similarity among unmatched components from different tuples. It provides triplet-aligned features to multi-role objects to distinguish them in component level. Moreover, we investigate the inherent connectivity between aligned triplet and component features through a bidirectional refinement module. This module amalgamates aligned hybrid features from both triplets and components to facilitate mutual enhancement and reciprocal refinement. + +Specifically, we apply an image backbone and transformer encoder to embed the image information across the image. Then the hybrid reciprocal transformer embeds hybrid representations for triplet and component learning, where the triplet-level decoder is utilized to embed the proposal segmentation mask of {subject-predicate-object}, while the component-level decoders are applied to produce the corresponding detected subject, object, and predicate component information. To achieve triplet-guided component-level learning, we follow the iterative learning process [16], the query position embedding of the component-level learning transformer decoder is generated by the processed results of the triplet-level query embedding to assign the feature level between them. To real- + +ize bidirectional refinement, we additionally design a representation cross-conditioned mechanism among layer dimensions of the hybrid reciprocal transformer using attention [45] mechanism. Furthermore, we build a triplet alignment loss function to match triplet-level feature to be consistent with matched component-level feature, which aims to distinguish overlapped triplets with multi-role objects. + +The comprehensive experimental evaluations conducted on Visual Genome dataset [19] and Action Genome dataset [15] substantiate the effectiveness of our framework. In Visual Genome dataset, we demonstrate superior performance by achieving increments of $+1.2$ and $+2.3$ in Recall at 50 and 100, respectively, thereby establishing a new state-of-the-art. Additionally, we achieve highest recorded metrics at both Recall 20 and 50 in Action Genome dataset, confirming its scalability across diverse datasets. + +The contribution of this paper is summarized as follows: + +- We integrate triplet-level and component-level representations within a hybrid reciprocal transformer framework, which uses the triplet mask feature as the proposal to improve predicate recognition performance. +- We introduce a triplet feature alignment loss to provide multi-role objects in overlapped visual relation with triplet-aligned features to distinguish them and improve predicate recognition efficacy. +- We leverage the aligned triplet and component feature with reciprocal refinement to explore inherent connectivity between them and optimize scene graphs. + +# 2. Related Work + +# 2.1. Two-Stage Scene Graph Generation + +Scene graph generation has become a prominent research area within the vision community, as evidenced by a wealth of recent studies [2, 4, 16-18, 25]. Classic research [29] design two-stage methods for scene graph generation and use the detected object to organize the relations. Two-stage methods commence with the pre-training of a robust object detector for all classes present in the dataset, typically employing detectors such as Faster R-CNN [35]. Upon establishing the object detector, the visual relation detection module is then constructed utilizing the object data (bound- ing boxes and associated features) extracted from the detector. In the subsequent stage, the object detector's parameters are fixed, and the focus shifts solely to training the relation detection module to construct the scene graph. With the development of artificial intelligence model, various architectures have been implemented for the relation detection module [34, 43, 50-52], such as message passing [50], LSTM [43, 48, 52], transformer [7, 20, 30, 38, 41]. + +Recent two-stage methods [22, 57] have incorporated triplet features to enhance predicate recognition. However, their effectiveness remains constrained in complex scenes + +involving overlapping visual relation tuples with multi-role objects. Additionally, the inherent separate design of object and predicate architectures in these methods limits their ability to effectively utilize the connections between aligned triplet-level and component-level elements. To this end, we leverage a reciprocal one-stage scene graph generation framework which leverages hybrid triplet and component features to enhance each other. It surpasses the inherent constraints of two-stage systems and solves the obfuscation from multi-role objects with a triplet alignment module. + +# 2.2. One-Stage Scene Graph Generation + +Contrary to two-stage methods, which primarily suffer from the limited ability of the relation detector to directly utilize features extracted by the detector, one-stage approaches [4, 9, 16-18, 25] overcome this limitation. They achieve this by integrating object detector and graph generation architectures, enabling end-to-end training that facilitates direct access and utilization of extracted features. One-stage methods mainly consists of fully convolutional networks [28] and transformer-based models [4, 9, 16-18, 25]. This configuration facilitates direct interactions between the detection and graph generation modules, significantly enhancing the quality of the resulting scene graphs. The current approaches in one-stage methods primarily concentrate on component-level learning, analyzing subjects, objects, and predicates individually given image feature. These methods [16-18] typically compile these components into tuples $\{\text{subject-predicate-object}\}$ during post-processing. + +Detecting accurate predicates and overcoming obfuscation in overlapped visual relation tuples pose significant challenges for current one-stage methods, primarily due to the lack of subject-object contextual features [22] and varying semantics of multi-role objects. We introduce triplet-level mask features as preliminary proposals to facilitate component-level learning in a one-stage architecture. Additionally, our method incorporates a triplet alignment module that aligns triplet and component features, enhancing the recognition of multi-role objects' diverse semantics. + +# 3. Method + +# 3.1. Formulation and Model Overview + +Scene Graph Generation Formulation. Classic scene graph generation task aims to organize a relation graph $\mathbf{G}'$ as $\mathbf{G}' = \{\mathbf{s}_k, \mathbf{o}_k, \mathbf{p}_k\}_{k=1}^K$ as a group of visual relation tuples organized from component-level subject $\mathbf{s}_k$ , object $\mathbf{o}_k$ , and predicate $\mathbf{p}_k$ . In our proposed architecture for hybrid representation tailored for scene graph generation, we refine the conventional task formulation by incorporating a triplet mask for each visual relation tuple. Given an image $\mathbf{I}$ , our scene graph $\mathbf{G}$ is represented as a collection of quadruples that simultaneously encapsulates both triplet-level and tra + +ditional component-level representations: + +$$ +\mathbf {G} = \left\{\mathbf {t} _ {k}, \mathbf {s} _ {k}, \mathbf {o} _ {k}, \mathbf {p} _ {k} \right\} _ {k = 1} ^ {K}, \tag {1} +$$ + +where $\mathbf{t}_k$ is the triplet-level representation denoting the triplet mask constructed from the classic component-level representation subject $\mathbf{s}_k$ , object $\mathbf{o}_k$ , and predicate $\mathbf{p}_k$ . Each component-level representation of the quadruple, i.e., $\mathbf{s}_k, \mathbf{o}_k, \mathbf{p}_k$ , consists of a class label and bounding box. Note the bounding box corresponding to $\mathbf{p}_k$ is calculated from the center point of its corresponding $\mathbf{s}_k$ and $\mathbf{o}_k$ . While $\mathbf{t}_k$ represents the binary mask of the triplet of $\mathbf{s}_k, \mathbf{o}_k, \mathbf{p}_k$ with a class label represent predicate class. This augmentation of triplet-level representation, as compared to traditional methodologies [4, 16-18], significantly provides subject-object pair contextual feature and limits explore space for accurate predicate recognition. Furthermore, it fosters exploration of the connectivity among hybrid triplet-level and component-level representations of same visual relations, thereby facilitating mutual enhancement. + +Model Overview. Figure 2 describes our overall architecture, which mainly consists of multi-scale image encoder and hybrid reciprocal transformer. We begin by utilizing an image encoder equipped with a ResNet backbone [12] and a transformer encoder [45] to extract multi-scale image visual features $\mathcal{I}$ . Based on $\mathcal{I}$ , our hybrid reciprocal transformer incorporates dual-level representations together, triplet mask queries $Q_{\mathrm{t}}$ for triplet-level representation and subject queries $Q_{\mathrm{s}}$ , object queries $Q_{\mathrm{o}}$ , predicate queries $Q_{\mathrm{p}}$ , for component-level representation, $Q$ means a sequence of queries as $Q = \{q^{i}\}_{i=1}^{N}$ . To leverage triplet-level representation facilitates guidance of component-level learning for accurate predicate recognition, hybrid reciprocal transformer uses iterative conditional chains for condition the process of $Q_{x}; x \in \{\mathrm{s}, \mathrm{o}, \mathrm{p}\}$ on $Q_{\mathrm{t}}$ (Sec. 3.2). Furthermore, to enable bidirectional refinement in hybrid reciprocal transformer, we develop a cross-conditioning mechanism, which leverages an attention module that operates across layer dimension of hybrid reciprocal transformer to reason aligned hybrid representations (Sec. 3.2). Finally, we propose alignment loss that leverages triplet-component matched relationship, to provide multi-role objects with distinct aligned triplet feature and recognize overlapped visual relation tuples accurately (Sec. 3.3). + +# 3.2. Hybrid Reciprocal Transformer + +Hybrid Reciprocal Transformer Architecture. Hybrid reciprocal transformer integrates triplet mask, subject, object, and predicate decoders to process input queries $Q_{x}$ ; $x \in \{\mathrm{t},\mathrm{s},\mathrm{o},\mathrm{p}\}$ via cross-attention with cropping image features $\mathcal{I}$ . $\mathcal{I}$ is cropped through a multi-scale image encoding process, where the final layer of a ResNet-based convolutional network provides a position-aware, flattened image feature representation for a transformer encoder [45, 59]. + +![](images/f99079d95f29ce1a68b4fc91d1f54a30862c60133d1364aac964f30135fb3caf.jpg) +Figure 2. Model overview of our method. Hybrid reciprocal transformer deploys four multi-layer decoder structures to embed triplet mask and visual relation tuple components learning using hybrid representations. We illustrate the detailed feature flow of each decoder to visualize two interaction paradigms within this transformer. The first paradigm facilitates triplet iterative guidance and enhances contextual feature integration for predicate recognition at the component level. The second paradigm employs aligned triplet and component features to enable bidirectional refinement of the hybrid triplet-level and component-level representations. + +Specifically, we model the elements of the scene graph—triplet mask, subject, object, and predicate, using multi-layer transformer decoders [1] with iterative processing across its $M$ layers step by step. The decoders are represented as $\mathcal{T}_{\mathrm{t}}$ , $\mathcal{T}_{\mathrm{s}}$ , $\mathcal{T}_{\mathrm{o}}$ , and $\mathcal{T}_{\mathrm{p}}$ , respectively. In each step $j$ , the $j$ -th layer of the triplet-level decoder, $\mathcal{T}_{\mathrm{t}}^{j}$ , generates masks feature for triplet-wise visual relations in the image. While the $j$ -th layer of each component-level decoder, $\mathcal{T}_{x}^{j}$ where $x \in \{\mathrm{s}, \mathrm{o}, \mathrm{p}\}$ , produces component-level part feature of the scene graph $\mathbf{G}^{j}$ . To realize triplet-level representation as proposals to boost component-level learning and allow hybrid representations to mutually reason and refine each other, our hybrid reciprocal transformer incorporates a triplet-guided relation generation mechanism with bidirectional learning refinement. + +Triplet-Guided Relation Learning with Bidirectional Refinement. To implement triplet-guided relation generation and bidirectional triplet-component enhancement, we develop two interaction paradigms between triplet-level and component-level features. The first paradigm involves using mask feature at the triplet level to guide the process of generating visual relations at the component level. This paradigm leverages triplet-level features as proposals to enhance predicate recognition from component-level data by incorporating contextual information from triplet-level representations. The second paradigm is to enable reciprocal sharing of triplet-level and component-level learning representations across the multiple layers of the hybrid reciprocal transformer through the use of an attention mechanism. This allows for streamlined integration of hybrid represen + +tations and mutual enhancement, as they represent different aspects of aligned identical aligned visual relation tuples. + +Specifically, for a decoder $\mathcal{T}_x; x \in \{\mathrm{t}, \mathrm{s}, \mathrm{o}, \mathrm{p}\}$ , we derive its output query representation by integrating self-attention on the input queries with positional embeddings and encoder-decoder attention using the image feature $\mathcal{I}$ . In $j$ -th layers, the outputs of decoders are shown as: + +$$ +\widehat {Q} _ {x} ^ {j} = \left\{\widehat {q} _ {x} ^ {i, j} \right\} _ {i = 1} ^ {N}; x \in \{t, s, o, p \}. \tag {2} +$$ + +$\widehat{Q}_x^j$ is produced from two types of conditional inputs from the aforementioned two interaction paradigms for hybrid representations respectively. The first is triplet-guided positional embedding $\bar{P}_x^j = \{\bar{p}_x^{i,j}\}_{i=1}^N$ ; $x \in \{\mathrm{t}, \mathrm{s}, \mathrm{o}, \mathrm{p}\}$ based original position embedding $P_x^j = \{p_x^{i,j}\}_{i=1}^N$ ; $x \in \{\mathrm{t}, \mathrm{s}, \mathrm{o}, \mathrm{p}\}$ respectively. The second is hybrid reciprocal query $\bar{Q}_x^{j-1}$ ; $x \in \{\mathrm{t}, \mathrm{s}, \mathrm{o}, \mathrm{p}\}$ from $j-1$ -th layer. + +Explicitly, the first involves conditioning component-level decoders $\mathcal{T}_x; x \in \{\mathrm{s,o,p}\}$ on the step $j$ outputs of triplet-level decoder $\mathcal{T}_{\mathrm{t}}$ using guided positional embedding $\bar{P}_x^j; x \in \{\mathrm{t,s,o,p}\}$ and following iterative process [16]. Specifically, $\hat{Q}_x^j = \mathcal{T}_x^j (\bar{Q}_x^{j - 1}|\bar{P}_x^j,\mathcal{I}); x \in \{\mathrm{t,s,o,p}\}$ , where triplet decoder using original learnable position embedding $\bar{P}_{\mathrm{t}}^{j} = P_{\mathrm{t}}^{j}$ , and triplet-guided mechanism is described as: + +$$ +\bar {P} _ {\mathrm {s}} ^ {j} = P _ {\mathrm {s}} ^ {j} + \text {F F N} \left(\mathrm {M H A} \left(P _ {\mathrm {s}} ^ {j}, \widehat {Q} _ {\mathrm {t}} ^ {j}, \widehat {Q} _ {\mathrm {t}} ^ {j}\right)\right), \tag {3} +$$ + +$$ +\bar {P} _ {\mathrm {o}} ^ {j} = P _ {\mathrm {o}} ^ {j} + \text {F F N} \left(\text {M H A} \left(P _ {\mathrm {o}} ^ {j}, \left[ \widehat {Q} _ {\mathrm {t}} ^ {j}, \widehat {Q} _ {\mathrm {s}} ^ {j} \right], \left[ \widehat {Q} _ {\mathrm {t}} ^ {j}, \widehat {Q} _ {\mathrm {s}} ^ {j} \right]\right)\right), \tag {4} +$$ + +$$ +\bar {P} _ {\mathrm {p}} ^ {j} = P _ {\mathrm {p}} ^ {j} + \text {F F N} \left(\mathrm {M H A} \left(P _ {\mathrm {p}} ^ {j}, \left[ \widehat {Q} _ {\mathrm {t}} ^ {j}, \widehat {Q} _ {\mathrm {s}} ^ {j}, \widehat {Q} _ {\mathrm {o}} ^ {j} \right], \left[ \widehat {Q} _ {\mathrm {t}} ^ {j}, \widehat {Q} _ {\mathrm {s}} ^ {j}, \widehat {Q} _ {\mathrm {o}} ^ {j} \right]\right)\right), \tag {5} +$$ + +where $\mathsf{FFN}(\cdot)$ is the feed-forward neural network, and $\mathsf{MHA}(\cdot,\cdot,\cdot)$ is the multi-head attention mechanism for + +![](images/4521cf47e76fb4f11b729eceebd8cc56f4b6ec67ffe7cf5e41cd5639fcd1d021.jpg) +Figure 3. Triplet feature alignment loss measures consistency between triplet feature with tripled component-level feature. It encourages similarity between triplet and component feature from identical visual relation tuples to distinguish multi-role object by aligning different triplet feature into component level. + +$\mathsf{Q},\mathsf{K},\mathsf{V}$ respectively, $[\cdot ,\cdot ]$ means concatenation of the element in it. While the second paradigm requires $\mathcal{T}_{\mathrm{t}}^{j}$ conditioned on $\mathcal{T}_x^{j - 1};x\in \{\mathrm{s,o,p}\}$ and $\mathcal{T}_x^j;x\in \{\mathrm{s,o,p}\}$ conditioned on $\mathcal{T}_{\mathrm{t}}^{j - 1}$ to allow bidirectional refinement. This cross-conditioned allows for feature flow enhancement among hybrid learning representations and is achieved through conditional hybrid reciprocal query $\bar{Q}_x^{j - 1};x\in$ $\{\mathrm{t,s,o,p}\}$ as: + +$$ +\bar {Q} _ {\mathrm {t}} ^ {j} = \widehat {Q} _ {\mathrm {t}} ^ {j} + \mathsf {F F N} (\mathsf {M H A} (\widehat {Q} _ {t} ^ {j}, [ \widehat {Q} _ {x} ^ {j} ], [ \widehat {Q} _ {x} ^ {j} ])); x \in \{\mathrm {s}, \mathrm {o}, \mathrm {p} \}, \tag {6} +$$ + +$$ +\bar {Q} _ {x} ^ {j} = \widehat {Q} _ {x} ^ {j} + \mathsf {F F N} (\mathsf {M H A} (\widehat {Q} _ {x} ^ {j}, \widehat {Q} _ {\mathrm {t}} ^ {j}, \widehat {Q} _ {\mathrm {t}} ^ {j})); x \in \{\mathrm {s}, 0, \mathrm {p} \}. \tag {7} +$$ + +By implementing the two specified conditional input patterns, we enable the generation of triplet-guided relation graph features within the same layer across various segments of hybrid reciprocal transformer, as well as bidirectional enhancement of hybrid representations across multiple layers of hybrid reciprocal transformer. Subsequently, these representations undergo processing by their respective prediction heads to produce desired hybrid outputs. Then alignment loss is applied for supervision, aiming to align hybrid features from the triplet dimension and enhance the differentiation of overlapping visual relation tuples. + +Prediction Heads. We build prediction heads for hybrid triplet-level and component-level features respectively. For triplet-level feature, we follow [6, 58] to refine with masked attention [5] with the multi-scale image feature. Specifically, triplet masks are generated through an iterative process wherein a dot product is applied across multiple scale dimensions along the channel axis between multi-scale image features $\mathcal{I}$ and $\widehat{Q}_{\mathrm{t}}$ . Then the sigmoid function is applied to normalize the segmentation masks. For the component-level detection bounding box representation (subject, object, and predicate), we use MLP to output their corresponding box position and class information directly. + +# 3.3. Triplet Feature Alignment Loss + +Alignment Loss. To enhance the differentiation of similar component features in multi-role objects that belong to overlapping visual relation tuples, we propose a triplet-component alignment loss. This loss function refines the + +representation of similar component features by motivating multiple queries to delineate distinct roles within a single multi-role object, correlating each query specifically to specific triplet configuration. This mechanism is realized by aligning the computed matched similarity matrix with the ground truth (GT) matching matrix to accurately map between triplet configurations and component representations. + +In the example illustrated in Figure 3, we observe a scenario featuring several overlapping visual relationships due to the presence of multi-role objects, such as a $\underline{bag}$ and $\underline{dog}$ . Here, the alignment loss assigns distinct features for different roles to each multi-role object based on their corresponding triplet-level features. Through the iterative matching and optimization process enabled by multiple layers of hybrid reciprocal transformer, the model promotes distinct representations for the multiple roles of a single object in overlapping triplets, using unique query-level features derived from their respective aligned triplet features. Furthermore, it promotes distinctiveness by encouraging both triplet-level and component-level features of various visual tuples to maintain separation, which aids in accurate recognition of complex and overlapping visual relation tuples. + +Specifically, in $j$ -th layer of hybrid reciprocal transformer, we calculate predicated similarity matrix given the processed triplet-level query $\widehat{Q}_{\mathrm{t}}^{j}$ and component-level query $\widehat{Q}_{x}^{j}; x \in \{\mathrm{s}, \mathrm{o}, \mathrm{p}\}$ . We first produce aligned query representation as $\widetilde{Q}_{x}^{j}; x \in \{\mathrm{t}, \mathrm{s}, \mathrm{o}, \mathrm{p}\}$ using their respective matching results to identity GT results to maintain bitwise corresponding among query level. Based on the aligned $\widetilde{Q}_{x}^{j}; x \in \{\mathrm{t}, \mathrm{s}, \mathrm{o}, \mathrm{p}\}$ , we generate a pseudo triplet-level representation $\mathbb{Q}_{\mathrm{t}}^{j}$ as the weighted sum of the output component-level representations $\widetilde{Q}_{x}^{j}; x \in \{\mathrm{s}, \mathrm{o}, \mathrm{p}\}$ , and calculate the predicted triplet-level similarities matrix $S^{j}$ of $j$ -th layer: + +$$ +\mathcal {S} ^ {j} = \mathbb {Q} _ {\mathrm {t}} ^ {j} \left(\widetilde {Q} _ {\mathrm {t}} ^ {j}\right) ^ {T}; \mathbb {Q} _ {\mathrm {t}} ^ {j} = \widetilde {Q} _ {\mathrm {s}} ^ {j} + \widetilde {Q} _ {\mathrm {o}} ^ {j} + \widetilde {Q} _ {\mathrm {p}} ^ {j}. \tag {8} +$$ + +While the GT similarity matrix $S_{\mathrm{GT}}$ is derived from the original matching relationship between the GT triplet-level masks and component-level subjects, objects, and predicates bounding boxes as one-hot formatting. + +Once the GT matrix $S_{\mathrm{GT}}$ and predicted similarity matrix $S^j$ are obtained, binary cross-entropy loss is applied between them to update model parameters. Incorporate with the iterative optimization and matching process within a hybrid reciprocal transformer framework, the alignment loss enhances the model's capability to differentiate overlapping visual relation tuples by promoting a high degree of similarity between matched component-level and triplet-level representations with GT matching results. + +End-to-End Training. In the end-to-end training process for our method, the overall loss function is the sum of the loss from all $M$ layers of hybrid reciprocal transformer, the loss $\mathcal{L}$ can be written as: + +
MethodR@50R@100mR@50mR@100AvgR@50AvgR@100F@50F@100
Two-Stage, X101-FPN Backbone
Motifs [52] [CVPR2018]32.136.95.56.818.821.99.411.5
VCTree [43] [CVPR202]31.836.16.67.719.221.910.912.7
TDE [44] [CVPR2020]19.423.29.311.114.417.212.615.0
EBM [42] [CVPR2021]20.524.79.711.615.118.213.215.8
BPLSA [11] [ICCV2021]21.725.513.515.717.620.616.619.4
DT2-ACBS [8] [ICCV2021]22.024.415.016.318.520.417.819.5
BGNN [24] [CVPR2021]31.035.810.712.720.924.215.918.7
PE-Net [57] [CVPR2023]30.735.212.414.521.624.917.720.5
VETO [41] [CVPR2023]27.531.58.19.517.820.512.514.6
DRM [22] [CVPR2024]34.038.99.011.221.525.014.217.4
One-Stage, ResNet-50 Backbone
RelTR [7] [TPAMI2023]27.530.710.812.319.221.515.517.6
EGTR [14] [CVPR2024]30.234.37.910.119.122.212.515.6
Hydra-SGG [4] [Arxiv2024]28.433.116.019.722.226.420.524.7
One-Stage, ResNet-101 Backbone
RelDN [53] [CVPR2019]30.334.84.45.417.420.17.79.3
AS-Net [3] [CVPR2021]18.721.16.17.212.414.29.210.7
SGTR [25] [CVPR2022]25.126.612.014.618.620.616.218.9
HOTR [17] [CVPR2021]22.427.16.99.714.718.410.614.3
SpeaQHotR [18] [CVPR2024]24.729.19.612.717.220.913.817.7
ISG† [16] [NIPS2022]29.532.17.48.418.520.311.813.3
SpeaQISG† [18] [CVPR2024]32.936.011.814.122.425.117.420.3
ISG [16] [NIPS2022]27.230.115.016.621.123.419.321.4
SpeaQISG [18] [CVPR2024]32.135.515.117.623.626.620.523.5
Ours34.1(+1.2)38.3(+2.3)16.0(+0.9)20.5(+2.9)25.1(+1.5)29.4(+2.8)21.8(+1.3)26.7(+3.2)
+ +Table 1. Scene graph generation performance on Visual Genome. The best results among models with ResNet-101 backbone are marked in bold. $\dagger$ denotes the performance without loss re-weighting proposed in [16]. + +$$ +\mathcal {L} = \sum_ {j = 1} ^ {M} \left(\mathcal {L} _ {\mathrm {t}} ^ {j} + \mathcal {L} _ {\mathrm {c}} ^ {j} + \mathcal {L} _ {\text {a l i g n}} ^ {j}\right), \tag {9} +$$ + +$$ +\mathcal {L} _ {\mathrm {t}} = \mathcal {L} _ {\mathrm {t}} ^ {\text {m a s k}} + \mathcal {L} _ {\mathrm {t}} ^ {\text {d i c e}} + \mathcal {L} _ {\mathrm {t}} ^ {\text {c l s}}, +$$ + +$$ +\mathcal {L} _ {\mathrm {c}} = \mathcal {L} _ {\mathrm {c}} ^ {\mathrm {s}} + \mathcal {L} _ {\mathrm {c}} ^ {\mathrm {o}} + \mathcal {L} _ {\mathrm {c}} ^ {\mathrm {p}}, +$$ + +where $\mathcal{L}_{\mathrm{t}}$ denotes triplet mask loss through Hungarian matching, which containing mask loss $\mathcal{L}_{\mathrm{t}}^{\mathrm{mask}}$ , dice loss $\mathcal{L}_{\mathrm{t}}^{\mathrm{dice}}$ , and mask classification loss $\mathcal{L}_{\mathrm{t}}^{\mathrm{cls}}$ , $\mathcal{L}_{\mathrm{c}}$ follows the design of [18] and consists of subject loss $\mathcal{L}_{\mathrm{c}}^{\mathrm{s}}$ , object loss $\mathcal{L}_{\mathrm{c}}^{\mathrm{o}}$ , and predicate loss $\mathcal{L}_{\mathrm{c}}^{\mathrm{p}}$ , each of $\mathcal{L}_{\mathrm{c}}^{\mathrm{s}}$ , $\mathcal{L}_{\mathrm{c}}^{\mathrm{o}}$ , and $\mathcal{L}_{\mathrm{c}}^{\mathrm{p}}$ is constructed from L1 box loss, GIoU loss [36], and classification loss through groupwise query multi-assignment matching [18], $\mathcal{L}_{\mathrm{align}}^{j}$ is alignment loss calculated using binary cross-entropy loss between $S^{j}$ and $S_{\mathrm{GT}}$ . + +# 4. Experiment + +# 4.1. Experiment Setup + +Datasets. We demonstrate the effectiveness of our proposed approach on two benchmark datasets. Visual Genome (VG) [19] is utilized as a benchmark for scene graph generation. We leverage the commonly processed subset from [50], which comprises 108k images across 150 object categories and 50 predicate categories. Action Genome (AG) [15] provides scene graph annotations across 9,848 videos within the Charades dataset [39] for studying human-object interactions. It covers 35 object categories and 25 relation categories, deliberately excluding the class person. Adhering to protocols from previous studies [10], we only utilize frames that have been annotated according to [15]. Frames + +lacking annotations for person or other objects are excluded to ensure the creation of usable scene graphs. + +Evaluation Metrics. Consistent with established methodologies in the literature [50, 52], we evaluate our model using the standard metrics Recall@K (R@K) and Mean Recall@K (mR@K). We further quantify performance by calculating both the arithmetic mean (AvgR@K) [23] and the harmonic mean (F@K) [16]. These additional measures provide a more nuanced assessment of model performance across varying predicate frequencies, ensuring a comprehensive evaluation of the balanced performance. + +Implementation Details. We utilize a ResNet-101 backbone [12] paired with a six-layer Transformer encoder. For each part in $\mathcal{T}_x$ ; $x \in \{\mathrm{t}, \mathrm{s}, \mathrm{o}, \mathrm{p}\}$ of the hybrid reciprocal transformer, we deploy a six-layer transformer decoder [45, 59] architecture. The query number for component-level learning is 600 for $\mathcal{T}_x^j$ ; $x \in \{\mathrm{s}, \mathrm{o}, \mathrm{p}\}$ , and 300 for $\mathcal{T}_{\mathrm{tn}}^j$ . The triplet matching strategy employs a naive Hungarian algorithm, optimizing for mask loss, dice loss, and classification loss. For component-level matching, we adopt a grouping matching strategy as proposed in [18], ensuring consistency in matched indices across components. The model undergoes pretraining on VG and AG datasets, focusing on object detection and semantic segmentation, respectively. Experiments are conducted on four A40 GPUs with batch size 16. + +# 4.2. Comparison Results with State-of-the-art + +Visual Genome Dataset. In Table 1, we report the performance of our method on VG dataset, the results show that our method achieves new state-of-the-art performance on those one-stage methods based on ResNet-101 backbone. + +
MethodR@20/50mR@20/50F@20/50
Two-Stage, X101-FPN Backbone
VRD [29]10.3 / 10.9-
Freq Prior [52]24.0 / 24.9--
IMP [50]23.9 / 25.5--
MSDN [26]24.0 / 25.6--
Graph R-CNN [51]24.1 / 25.8--
RelDN [53]25.0 / 26.2--
SimpleBase [10]27.9 / 30.48.3 / 9.112.8 / 14.0
One-Stage, ResNet-101 Backbone
ISG [16]31.0 / 37.510.9 / 13.216.1 / 19.5
\( ISG_{\dagger} \) [16]30.1 / 36.519.2 / 22.323.4 / 27.7
Spea\( Q_{\text{ISG}} \) [18]30.8 / 38.011.7 / 13.817.0 / 20.2
Spea\( Q_{\text{ISG}_{\dagger}} \) [18]31.8 / 39.420.9 / 23.125.2 / 29.1
Ours32.0 / 40.221.3 / 23.625.5 / 29.7
+ +Our model demonstrates significant enhancements with increases of $+1.2$ , $+2.3$ , $+0.9$ , and $+2.9$ in the R@50/100 and mR@50/100 metrics. These enhancements can be credited to our pioneering hybrid reciprocal architecture, which integrates reasoning across aligned triplets and component features. Furthermore, the synergistic interplay between the two-tiered representation of visual relation tuple significantly augments the overall performance. When compared with two-stage methods employing the X101-FPN backbone, our method achieves state-of-the-art results in R@K, as evidenced by its comparison with the DRM model [22]. Moreover, owing to our one-stage design to jointly optimize predicate recognition and object detection, we substantially outperform two-stage approaches in mR@K. + +Action Genome Dataset. We additionally report R@20/50, mR20/50, and F@20/50 values on the Action Genome dataset [15] in Table 2. Consistent with the findings observed on the Visual Genome dataset, we establish a new benchmark for state-of-the-art performance across all utilized metrics, including R@20/50, mR@20/50, and F@20/50. This advancement underscores the efficacy of our hybrid reciprocal architecture with triplet alignment in enhancing scene graph generation capabilities. Moreover, the improvement in performance across various datasets demonstrates the robust scalability and adaptability of the hybrid learning framework, which does not necessitate specific modifications to be effective across different contexts. + +# 4.3. Ablation Study and Analysis + +Alignment Loss. We apply an alignment loss between hybrid triplet and component features to enhance the distinction of multi-role objects in overlapping triplets, by aligning triplet features into the component level. We investigate the effect of the triplet feature alignment loss by setting different weights in the end-to-end training overall loss function. The results metrics in the VG dataset using different weights (0.0, 0.1, 0.2, 0.3, 0.4, 0.5, 1.0) are reported at Figure 4, the weight used in our vanilla model is 0.2. Our findings suggest that increasing the weight of alignment loss + +![](images/00063d6094f90b08deb0255954de53456795003940482df8ea136d2065b86a9a.jpg) + +![](images/e6300f9fb6ca522991e9505e354d86fd817060e7983780da842f9dfdef500a87.jpg) +Figure 4. Scene graph generation performance comparison with Visual Genome dataset using different parameters of weight of alignment loss in overall training loss. + +Table 2. Scene graph generation performance on Action Genome. Best performance with bold. + +
αc:tR@50R@100mR@50mR@100
0.532.135.316.522.3
1.033.235.716.121.2
2.033.736.816.422.3
+ +Table 3. Effect of matching triplet-level and component-level together with different ratio ${\alpha }_{\mathrm{c} : \mathrm{t}}$ between component and triplet cost. + +initially enhances performance as desired, but subsequently deteriorates. This pattern of performance degradation at higher weights of alignment loss indicates that excessive emphasis on alignment loss may hinder learning efficacy. + +Hybrid Representation Guidance Order and Reciprocal Refinement. In our framework, we traditionally utilize the processed query for triplet-level segmentation masks to condition the generation of queries for component-level representations. To explore the impact of reversing this sequence, we instead employ the component-level query as a condition for generating triplet-level representations. The results for this variant, denoted as R@50/100 and mR@50/100, are 33.2, 36.9, 15.3, and 18.9, respectively, when compared with our vanilla model, as detailed in Table 1, it is evident that utilizing the triplet-level query as a condition for component-level learning yields superior performance. This enhancement can be attributed to the global information embedded in the triplet-level query feature, which effectively enhances the learning of local information in the component-level elements—subject, object, and predicate. Additionally, we explore the impact of reciprocal refinement by omitting this feature from our vanilla model. The performance metrics on VG dataset for R@50/100 and mR@50/100 are 32.4, 36.3, 14.4, and 18.5, respectively. The observed decrease in performance underscores the significance of reciprocal refinement in integrating hybrid triplet and component features effectively. + +Grouping Hybrid Representations in Uniform Matcher. Inspired by [18], we explore integrating hybrid representations using a single Hungarian matching mechanism to align them directly. Table 3 presents a comparative anal + +![](images/4d36f69c497476372929a2ebb076c3021165d744d5f628e9475735601b159455.jpg) +Figure 5. Qualitative analysis of our method compared with the baseline models, SpeaQ [18] with ISG [16], indicates that our method yields significantly enhanced accuracy in the prediction of predicates and object detection. Red marks error. + +
NQtNQcR@50R@100mR@50mR@100
30030033.837.315.218.3
40033.837.415.318.9
50033.937.615.719.4
60030033.937.815.720.1
40034.037.915.920.5
50034.238.216.221.2
+ +ysis of the outcomes, employing varied matching ratios between component and triplet levels. The findings indicate that merging the hybrid representations via a Hungarian matching mechanism yields inferior results compared to our vanilla model. Specifically, reducing component-level matching negatively impacts performance. This decline is attributed to the stringent constraints imposed by grouping two distinct hybrid representations, which restricts their internal matching and adversely affects overall performance. + +Query Number. We examine the impact of query number for hybrid representations ( $N_{Q\mathrm{t}}$ for triplet and $N_{Q\mathrm{c}}$ for component) in Table 4. The comparative analysis indicates that incremental increases in query numbers consistently enhance outcome performance. This improvement adheres to fundamental principles observed in transformer-based approaches for visual tasks [1, 21, 27, 59] by providing richer representational capacity and coverage [54]. + +Matching of Hybrid Representations in Inference Stage. In the inference stage, our vanilla model only uses the bitwise multiply results from the component-level subject, object, and predicate scores. We compare two deformations in Table 5: the first directly calculates multiply of bitwise triplet-level mask scores together, and the second uses Hungarian matching to align two-level output and then multiply them together. Their corresponding results are shown in Table 5. The comparison results show that directly multiply component-level with triplet-level score together will definitely cause negative performance, since the two-level output representation cannot be directly assigned to the specific result subject, object, and predicate tuple. We also find that after the Hungarian matching, the multiplied scores of the hybrid two-level representation will improve performance + +Table 4. Scene graph performance on VG dataset with different number of triplet-level and component-level queries. + +
Inference TypeR@50R@100mR@50mR@100
Component Only34.138.316.020.5
Component × Triplet32.535.614.616.5
Component × Triplet34.238.316.120.6
+ +Table 5. Comparison results among inference matching between hybrid representations. $\times$ means bitwise multiply and $\star$ denotes multiply after matching. + +a bit. We suppose this improvement came from the final aligned hybrid representation by the alignment loss will boost the final scene graph generation performance. + +# 4.4. Qualitative Results + +Figure 5 presents qualitative examples comparing prediction results from the SpeaQ [18] and ours. Ours successfully detects challenging samples that contain overlapped complicated visual relation tuples and provides accurate predicate recognition results. Furthermore, we provide more accurate object and predicate detection results compared to previous baselines. These enhancements are attributed to a hybrid reciprocal architecture that incorporates an alignment module to differentiate triplet features and predict the component relation graph, while receiving supervision at both triplet and component levels. + +# 5. Conclusion + +We integrate a triplet feature alignment module into a hybrid reciprocal transformer to innovate scene graph generation, starting with triplet mask features guide that enhances predicate recognition. We introduce a triplet alignment loss to differentiate multi-role object features in intersecting visual relation tuples, aligning them with triplet features for clearer distinction. Furthermore, our architecture features a reciprocal refinement module in multiple layers for effective fusion and interaction of hybrid aligned features. Experimental results across multiple datasets confirm our superiority over existing methods, marking a substantial advancement in the field of scene graph generation. + +Acknowledgement. This work is supported by the InnoHK Clusters of the Hong Kong SAR Government via the Hong Kong Centre for Logistics Robotics. + +# References + +[1] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 4, 8 +[2] Guikun Chen, Jin Li, and Wenguan Wang. Scene graph generation with role-playing large language models. arXiv preprint arXiv:2410.15364, 2024. 1, 2 +[3] Mingfei Chen, Yue Liao, Si Liu, Zhiyuan Chen, Fei Wang, and Chen Qian. Reformulating hoi detection as adaptive set prediction. In CVPR, 2021. 6 +[4] Minghan Chen, Guikun Chen, Wenguan Wang, and Yi Yang. Hydra-sgg: Hybrid relation assignment for one-stage scene graph generation. arXiv preprint arXiv:2409.10262, 2024. 1, 2, 3, 6 +[5] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 5 +[6] Sehwan Choi, Jungho Kim, Hongjae Shin, and Jun Won Choi. Mask2map: Vectorized hd map construction using bird's eye view segmentation masks. arXiv preprint arXiv:2407.13517, 2024. 5 +[7] Yuren Cong, Michael Ying Yang, and Bodo Rosenhahn. Reltr: Relation transformer for scene graph generation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(9):11169-11183, 2023. 2, 6 +[8] Alakh Desai, Tz-Ying Wu, Subarna Tripathi, and Nuno Vasconcelos. Learning of visual relations: The devil is in the tails. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15404-15413, 2021. 6 +[9] Qi Dong, Zhuowen Tu, Haofu Liao, Yuting Zhang, Vijay Mahadevan, and Stefano Soatto. Visual relationship detection using part-and-sum transformers with composite queries. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3550-3559, 2021. 3 +[10] Raghav Goyal12 and Leonid Sigal123. A simple baseline for weakly-supervised human-centric relation detection. 2021. 6, 7 +[11] Yuyu Guo, Lianli Gao, Xuanhan Wang, Yuxuan Hu, Xing Xu, Xu Lu, Heng Tao Shen, and Jingkuan Song. From general to specific: Informative scene graph generation via balance adjustment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16383-16392, 2021. 6 +[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3, 6 +[13] Ruozhen He, Paola Cascante-Bonilla, Ziyan Yang, Alexander C Berg, and Vicente Ordonez. Improved visual grounding through self-consistent explanations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13095-13105, 2024. 1 + +[14] Jinbae Im, JeongYeon Nam, Nokyung Park, Hyungmin Lee, and Seunghyun Park. Egtr: Extracting graph from transformer for scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24229-24238, 2024. 1, 2, 6 +[15] Jingwei Ji, Ranjay Krishna, Li Fei-Fei, and Juan Carlos Niebles. Action genome: Actions as compositions of spatiotemporal scene graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10236-10247, 2020. 2, 6, 7 +[16] Siddhesh Khandelwal and Leonid Signal. Iterative scene graph generation. Advances in Neural Information Processing Systems, 35:24295-24308, 2022. 1, 2, 3, 4, 6, 7, 8 +[17] Bumsoo Kim, Junhyun Lee, Jaewoo Kang, Eun-Sol Kim, and Hyunwoo J Kim. Hotr: End-to-end human-object interaction detection with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 74-83, 2021. 6 +[18] Jongha Kim, Jihwan Park, Jinyoung Park, Jinyoung Kim, Sehyung Kim, and Hyunwoo J Kim. Groupwise query specialization and quality-aware multi-assignment for transformer-based visual relationship detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28160–28169, 2024. 1, 2, 3, 6, 7, 8 +[19] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 2, 6 +[20] Sanjoy Kundu and Sathyanarayanan N Aakur. Is-ggt: Iterative scene graph generation with generative transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6292-6301, 2023. 1, 2 +[21] Feng Li, Hao Zhang, Shilong Liu, Jian Guo, Lionel M Ni, and Lei Zhang. Dn-detr: Accelerate detr training by introducing query denoising. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13619-13627, 2022. 8 +[22] Jiankai Li, Yunhong Wang, Xiefan Guo, Ruijie Yang, and Weixin Li. Leveraging predicate and triplet learning for scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28369-28379, 2024. 1, 2, 3, 6, 7 +[23] Lin Li, Long Chen, Yifeng Huang, Zhimeng Zhang, Songyang Zhang, and Jun Xiao. The devil is in the labels: Noisy label correction for robust scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18869-18878, 2022. 6 +[24] Rongjie Li, Songyang Zhang, Bo Wan, and Xuming He. Bipartite graph network with adaptive message passing for unbiased scene graph generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11109-11119, 2021. 6 +[25] Rongjie Li, Songyang Zhang, and Xuming He. Sgtr: End-to-end scene graph generation with transformer. In proceedings + +of the IEEE/CVF conference on computer vision and pattern recognition, pages 19486-19496, 2022. 1, 2, 3, 6 +[26] Yikang Li, Wanli Ouyang, Bolei Zhou, Kun Wang, and Xiaogang Wang. Scene graph generation from objects, phrases and region captions. In Proceedings of the IEEE international conference on computer vision, pages 1261-1270, 2017. 7 +[27] Fanfan Liu, Haoran Wei, Wenzhe Zhao, Guozhen Li, Jingquan Peng, and Zihao Li. Wb-detr: Transformer-based detector without backbone. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2979-2987, 2021. 8 +[28] Hengyue Liu, Ning Yan, Masood Mortazavi, and Bir Bhanu. Fully convolutional scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11546-11556, 2021. 3 +[29] Cewu Lu, Ranjay Krishna, Michael Bernstein, and Li Fei-Fei. Visual relationship detection with language priors. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 852-869. Springer, 2016. 2, 7 +[30] Yichao Lu, Himanshu Rai, Jason Chang, Boris Knyazev, Guangwei Yu, Shashank Shekhar, Graham W Taylor, and Maksims Volkovs. Context-aware scene graph generation with seq2seq transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15931-15941, 2021. 2 +[31] Jie Ma, Pinghui Wang, Dechen Kong, Zewei Wang, Jun Liu, Hongbin Pei, and Junzhou Zhao. Robust visual question answering: Datasets, methods, and future challenges. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1 +[32] Tao Ma, Bing Bai, Haozhe Lin, Heyuan Wang, Yu Wang, Lin Luo, and Lu Fang. When visual grounding meets gigapixel-level large-scale scenes: Benchmark and approach. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22119-22128, 2024. 1 +[33] Nandita Naik, Christopher Potts, and Elisa Kreiss. Context-vqa: Towards context-aware and purposeful visual question answering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2821-2825, 2023. 1 +[34] Mengshi Qi, Weijian Li, Zhengyuan Yang, Yunhong Wang, and Jiebo Luo. Attentive relational networks for mapping images to scene graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3957-3966, 2019. 2 +[35] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE transactions on pattern analysis and machine intelligence, 39(6):1137-1149, 2016. 2 +[36] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 658-666, 2019. 6 +[37] Shihao Shao, Kaifeng Chen, Arjun Karpur, Qinghua Cui, Andre Araujo, and Bingyi Cao. Global features are all you + +need for image retrieval and reranking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11036-11046, 2023. 1 +[38] Suprosanna Shit, Rajat Koner, Bastian Wittmann, Johannes Paetzold, Ivan Ezhov, Hongwei Li, Jiazhen Pan, Sahand Sharifzadeh, Georgios Kaissis, Volker Tresp, et al. Relationformer: A unified framework for image-to-graph generation. In European Conference on Computer Vision, pages 422-439. Springer, 2022. 2 +[39] Gunnar A Sigurdsson, Gül Varol, Xiaolong Wang, Ali Farhadi, Ivan Laptev, and Abhinav Gupta. Hollywood in homes: Crowdsourcing data collection for activity understanding. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 510-526. Springer, 2016. 6 +[40] Chull Hwan Song, Jooyoung Yoon, Shunghyun Choi, and Yannis Avrithis. Boosting vision transformers for image retrieval. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 107-117, 2023. 1 +[41] Gopika Sudhakaran, Devendra Singh Dhami, Kristian Kersting, and Stefan Roth. Vision relation transformer for unbiased scene graph generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21882-21893, 2023. 1, 2, 6 +[42] Mohammed Suhail, Abhay Mittal, Behjat Siddiquie, Chris Broaddus, Jayan Eledath, Gerard Medioni, and Leonid Sigal. Energy-based learning for scene graph generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13936-13945, 2021. 6 +[43] Kaihua Tang, Hanwang Zhang, Baoyuan Wu, Wenhan Luo, and Wei Liu. Learning to compose dynamic tree structures for visual contexts. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6619-6628, 2019. 2, 6 +[44] Kaihua Tang, Yulei Niu, Jianqiang Huang, Jiaxin Shi, and Hanwang Zhang. Unbiased scene graph generation from biased training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3716-3725, 2020. 6 +[45] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 2, 3, 6 +[46] Guan Wang, Zhimin Li, Qingchao Chen, and Yang Liu. Oed: Towards one-stage end-to-end dynamic scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27938-27947, 2024. 1 +[47] Sai Wang, Yutian Lin, and Yu Wu. Omni-q: Omnidirectional scene understanding for unsupervised visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14261-14270, 2024. 1 +[48] Wenbin Wang, Ruiping Wang, Shiguang Shan, and Xilin Chen. Sketching image gist: Human-mimetic hierarchical scene graph generation. In European conference on computer vision, pages 222-239. Springer, 2020. 2 + +[49] Hui Wu, Min Wang, Wengang Zhou, Zhenbo Lu, and Houqiang Li. Asymmetric feature fusion for image retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11082-11092, 2023. 1 +[50] Danfei Xu, Yuke Zhu, Christopher B Choy, and Li Fei-Fei. Scene graph generation by iterative message passing. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5410-5419, 2017. 2, 6, 7 +[51] Jianwei Yang, Jiasen Lu, Stefan Lee, Dhruv Batra, and Devi Parikh. Graph r-cnn for scene graph generation. In Proceedings of the European conference on computer vision (ECCV), pages 670-685, 2018. 7 +[52] Rowan Zellers, Mark Yatskar, Sam Thomson, and Yejin Choi. Neural motifs: Scene graph parsing with global context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5831-5840, 2018. 2, 6, 7 +[53] Ji Zhang, Kevin J Shih, Ahmed Elgammal, Andrew Tao, and Bryan Catanzaro. Graphical contrastive losses for scene graph parsing. In CVPR, 2019. 6, 7 +[54] Jichuan Zhang, Wei Li, Shuang Cheng, Ya-Li Li, and Shengjin Wang. Dynamic object queries for transformer-based incremental object detection. arXiv preprint arXiv:2407.21687, 2024. 8 +[55] Xi Zhang, Feifei Zhang, and Changsheng Xu. Vqacl: A novel visual question answering continual learning setting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19102-19112, 2023. 1 +[56] Long Zhao, Liangzhe Yuan, Boqing Gong, Yin Cui, Florian Schroff, Ming-Hsuan Yang, Hartwig Adam, and Ting Liu. Unified visual relationship detection with vision and language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6962-6973, 2023. 1 +[57] Chaofan Zheng, Xinyu Lyu, Lianli Gao, Bo Dai, and Jingkuan Song. Prototype-based embedding network for scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22783-22792, 2023. 1, 2, 6 +[58] Yi Zhou, Hui Zhang, Jiaqian Yu, Yifan Yang, Sangil Jung, Seung-In Park, and ByungIn Yoo. Himap: Hybrid representation learning for end-to-end vectorized hd map construction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15396-15406, 2024. 5 +[59] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable $\{\mathrm{detr}\}$ : Deformable transformers for end-to-end object detection. In International Conference on Learning Representations, 2021. 3, 6, 8 \ No newline at end of file diff --git a/2025/Hybrid Reciprocal Transformer with Triplet Feature Alignment for Scene Graph Generation/images.zip b/2025/Hybrid Reciprocal Transformer with Triplet Feature Alignment for Scene Graph Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9da4c89336721d24402bda00e9429362fc7da5df --- /dev/null +++ b/2025/Hybrid Reciprocal Transformer with Triplet Feature Alignment for Scene Graph Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4599556c8371f882361fb0f2ba495a724c9a47f2922e9e7b6d2a3e31f8ff4f3c +size 645261 diff --git a/2025/Hybrid Reciprocal Transformer with Triplet Feature Alignment for Scene Graph Generation/layout.json b/2025/Hybrid Reciprocal Transformer with Triplet Feature Alignment for Scene Graph Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..88c8fe3a18ee7bf4ab34c6211cb8ee51eb8bc2ea --- /dev/null +++ b/2025/Hybrid Reciprocal Transformer with Triplet Feature Alignment for Scene Graph Generation/layout.json @@ -0,0 +1,8931 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 143, + 103, + 470, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 103, + 470, + 140 + ], + "spans": [ + { + "bbox": [ + 143, + 103, + 470, + 140 + ], + "type": "text", + "content": "Hybrid Reciprocal Transformer with Triplet Feature Alignment for Scene Graph Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 87, + 161, + 523, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 161, + 523, + 203 + ], + "spans": [ + { + "bbox": [ + 87, + 161, + 523, + 203 + ], + "type": "text", + "content": "Jiawei Fu, Tiantian Zhang, Kai Chen, and Qi Dou \nDepartment of Computer Science and Engineering, The Chinese University of Hong Kong \njwfu@cse.cuhk.edu.hk" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 258, + 297, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 258, + 297, + 582 + ], + "spans": [ + { + "bbox": [ + 54, + 258, + 297, + 582 + ], + "type": "text", + "content": "Scene graph generation is a pivotal task in computer vision, focusing on comprehensive identification of visual relation tuples embedded within images. The advancement of methods involving triplets has sought to enhance task performance by integrating triplets as contextual features for more precise predicate identification from component level. However, challenges remain due to interference from multi-role objects in overlapping tuples within complex environments, which impairs the model's ability to distinguish and align specific triplet features for reasoning diverse semantics of multi-role objects. To address these issues, we introduce a novel framework that incorporates a triplet alignment model into a hybrid reciprocal transformer architecture, starting from using triplet mask features to guide the learning of component-level relation graphs. To effectively distinguish multi-role objects characterized by overlapping visual relation tuples, we introduce a triplet alignment loss, which provides multi-role objects with aligned features from triplet and helps customize them. Additionally, we explore the inherent connectivity between hybrid aligned triplet and component features through a bidirectional refinement module, which enhances feature interaction and reciprocal reinforcement. Experimental results demonstrate that our model achieves state-of-the-art performance on the Visual Genome and Action Genome datasets, underscoring its effectiveness and adaptability. Project page: hg-sg.github.io." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 608, + 135, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 608, + 135, + 620 + ], + "spans": [ + { + "bbox": [ + 56, + 608, + 135, + 620 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": "Scene graph generation [2, 4, 18] aims to find all possible visual relation tuples " + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\{\\text{subject-predicate-object}\\}" + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": " given an image. It has wide applications in various fields, including image retrieval [37, 40, 49], visual grounding [13, 32, 47], and visual question answering [31, 33, 55]. Typically, diverse two-stage [20, 22, 41, 56, 57] and one-stage [4, 16, 18, 25, 46] methods have been developed to solve this prob" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 316, + 232, + 395, + 344 + ], + "blocks": [ + { + "bbox": [ + 316, + 232, + 395, + 344 + ], + "lines": [ + { + "bbox": [ + 316, + 232, + 395, + 344 + ], + "spans": [ + { + "bbox": [ + 316, + 232, + 395, + 344 + ], + "type": "image", + "image_path": "34955259c84ebc6373acc83bd9bc11c3dffdbcc6c8b7f09e262c52e43188ef1e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 348, + 555, + 426 + ], + "lines": [ + { + "bbox": [ + 313, + 348, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 348, + 555, + 426 + ], + "type": "text", + "content": "Figure 1. Complex scenarios often involve overlapping visual relationships, where a single object can assume multiple roles across different triplets. Identifying these multi-role objects is challenging without knowing their aligned triplet features. Moreover, aligned triplets (top) and component (bottom) representations correspond to the same visual relation tuple, offering an opportunity to leverage their inherent relevance for reciprocal refinement." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 395, + 232, + 474, + 344 + ], + "blocks": [ + { + "bbox": [ + 395, + 232, + 474, + 344 + ], + "lines": [ + { + "bbox": [ + 395, + 232, + 474, + 344 + ], + "spans": [ + { + "bbox": [ + 395, + 232, + 474, + 344 + ], + "type": "image", + "image_path": "d74d360bf4b0558d3918ef15d96b17aedce03e12efd035371689ff2b184cea30.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 474, + 232, + 554, + 344 + ], + "blocks": [ + { + "bbox": [ + 474, + 232, + 554, + 344 + ], + "lines": [ + { + "bbox": [ + 474, + 232, + 554, + 344 + ], + "spans": [ + { + "bbox": [ + 474, + 232, + 554, + 344 + ], + "type": "image", + "image_path": "bdfa7616dd081249b8131875bc92132dd428c31269a14711fc389ae9048e90cf.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 436, + 556, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 436, + 556, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 436, + 556, + 581 + ], + "type": "text", + "content": "lem by organizing regressed component-level subject, object, and predicate with post-processing. However, accurate predicate recognition remains challenging in scene graph generation, as many images incorporate extraneous visual elements. Directly regressing the predicate often introduces noise from features of unaligned visual relation tuples, owing to the lack of definite corresponding subject-object pair features for the visual relation. Accordingly, recent researches [4, 14, 22, 57] have utilized triplet features as prototypes for contextual semantics, combining detailed cues with broader triplet features to enhance predicate recognition accuracy from the triplet level." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 582, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 557, + 715 + ], + "type": "text", + "content": "Despite advancements, accurately matching predicates to corresponding visual tuples and identifying all potential visual relations in complex environments remains challenging, even with triplet features. As illustrated in Figure 1, a single scene can contain multiple visual relations, with one object assuming multiple roles (either as subject or object) across different tuples. The analogous component-level features of multi-role objects increase overlap and misalignment among visual relation triplets, thereby hindering the model's ability to differentiate among triplets and to further reason diverse semantics of multi-role objects with" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8953" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "out precisely aligned triplet features. Furthermore, previous methodologies [4, 14, 22, 57] often overlook the potential of leveraging connectivity between aligned triplet and component features for mutual enhancement. This oversight can be attributed to the misalignment of triplet and component features [4, 14, 22, 57] or to a two-stage design [22, 57] that is not conducive to joint optimization of hybrid features." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 158, + 294, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 158, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 158, + 294, + 326 + ], + "type": "text", + "content": "To this end, it is essential to integrate feature alignment at triplet level, which is key in differentiating complex overlapping visuals. This alignment provides objects performing multiple roles with distinct aligned triplet features, facilitating their distinction within overlapped visual relational tuples. Moreover, to maximize the utility of aligned triplet and component features for identical visual relation tuples, and to investigate the consistency and connectivity between hybrid features to enhance mutual performance, implementing an end-to-end, one-stage scene graph generation framework is essential. This framework should enable bidirectional feature refinement for predicate recognition and object detection, thereby achieving joint optimization of both hybrid triplet and component features." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 328, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 328, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 328, + 295, + 544 + ], + "type": "text", + "content": "In this study, we integrate a triplet alignment model into a hybrid reciprocal transformer architecture for the generation of scene graphs. This integration leverages triplet-level representations as contextual features to enhance component-level scene generation and improve the accuracy of predicate recognition. To effectively differentiate complex, overlapping visual relation tuples with multirole objects at the feature level, we introduce a triplet alignment loss. This loss function promotes similarity between matched triplet and component features, while penalizing similarity among unmatched components from different tuples. It provides triplet-aligned features to multi-role objects to distinguish them in component level. Moreover, we investigate the inherent connectivity between aligned triplet and component features through a bidirectional refinement module. This module amalgamates aligned hybrid features from both triplets and components to facilitate mutual enhancement and reciprocal refinement." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": "Specifically, we apply an image backbone and transformer encoder to embed the image information across the image. Then the hybrid reciprocal transformer embeds hybrid representations for triplet and component learning, where the triplet-level decoder is utilized to embed the proposal segmentation mask of {subject-predicate-object}, while the component-level decoders are applied to produce the corresponding detected subject, object, and predicate component information. To achieve triplet-guided component-level learning, we follow the iterative learning process [16], the query position embedding of the component-level learning transformer decoder is generated by the processed results of the triplet-level query embedding to assign the feature level between them. To real-" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "ize bidirectional refinement, we additionally design a representation cross-conditioned mechanism among layer dimensions of the hybrid reciprocal transformer using attention [45] mechanism. Furthermore, we build a triplet alignment loss function to match triplet-level feature to be consistent with matched component-level feature, which aims to distinguish overlapped triplets with multi-role objects." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 157, + 553, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 157, + 553, + 264 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 553, + 264 + ], + "type": "text", + "content": "The comprehensive experimental evaluations conducted on Visual Genome dataset [19] and Action Genome dataset [15] substantiate the effectiveness of our framework. In Visual Genome dataset, we demonstrate superior performance by achieving increments of " + }, + { + "bbox": [ + 313, + 157, + 553, + 264 + ], + "type": "inline_equation", + "content": "+1.2" + }, + { + "bbox": [ + 313, + 157, + 553, + 264 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 157, + 553, + 264 + ], + "type": "inline_equation", + "content": "+2.3" + }, + { + "bbox": [ + 313, + 157, + 553, + 264 + ], + "type": "text", + "content": " in Recall at 50 and 100, respectively, thereby establishing a new state-of-the-art. Additionally, we achieve highest recorded metrics at both Recall 20 and 50 in Action Genome dataset, confirming its scalability across diverse datasets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 326, + 265, + 553, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 265, + 553, + 277 + ], + "spans": [ + { + "bbox": [ + 326, + 265, + 553, + 277 + ], + "type": "text", + "content": "The contribution of this paper is summarized as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 278, + 553, + 409 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 314, + 278, + 553, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 278, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 314, + 278, + 553, + 326 + ], + "type": "text", + "content": "- We integrate triplet-level and component-level representations within a hybrid reciprocal transformer framework, which uses the triplet mask feature as the proposal to improve predicate recognition performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 326, + 553, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 326, + 553, + 373 + ], + "spans": [ + { + "bbox": [ + 314, + 326, + 553, + 373 + ], + "type": "text", + "content": "- We introduce a triplet feature alignment loss to provide multi-role objects in overlapped visual relation with triplet-aligned features to distinguish them and improve predicate recognition efficacy." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 373, + 553, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 373, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 314, + 373, + 553, + 409 + ], + "type": "text", + "content": "- We leverage the aligned triplet and component feature with reciprocal refinement to explore inherent connectivity between them and optimize scene graphs." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 422, + 400, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 422, + 400, + 434 + ], + "spans": [ + { + "bbox": [ + 314, + 422, + 400, + 434 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 442, + 504, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 442, + 504, + 455 + ], + "spans": [ + { + "bbox": [ + 314, + 442, + 504, + 455 + ], + "type": "text", + "content": "2.1. Two-Stage Scene Graph Generation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 461, + 553, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 553, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 553, + 676 + ], + "type": "text", + "content": "Scene graph generation has become a prominent research area within the vision community, as evidenced by a wealth of recent studies [2, 4, 16-18, 25]. Classic research [29] design two-stage methods for scene graph generation and use the detected object to organize the relations. Two-stage methods commence with the pre-training of a robust object detector for all classes present in the dataset, typically employing detectors such as Faster R-CNN [35]. Upon establishing the object detector, the visual relation detection module is then constructed utilizing the object data (bound- ing boxes and associated features) extracted from the detector. In the subsequent stage, the object detector's parameters are fixed, and the focus shifts solely to training the relation detection module to construct the scene graph. With the development of artificial intelligence model, various architectures have been implemented for the relation detection module [34, 43, 50-52], such as message passing [50], LSTM [43, 48, 52], transformer [7, 20, 30, 38, 41]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "Recent two-stage methods [22, 57] have incorporated triplet features to enhance predicate recognition. However, their effectiveness remains constrained in complex scenes" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8954" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "type": "text", + "content": "involving overlapping visual relation tuples with multi-role objects. Additionally, the inherent separate design of object and predicate architectures in these methods limits their ability to effectively utilize the connections between aligned triplet-level and component-level elements. To this end, we leverage a reciprocal one-stage scene graph generation framework which leverages hybrid triplet and component features to enhance each other. It surpasses the inherent constraints of two-stage systems and solves the obfuscation from multi-role objects with a triplet alignment module." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 201, + 246, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 201, + 246, + 213 + ], + "spans": [ + { + "bbox": [ + 55, + 201, + 246, + 213 + ], + "type": "text", + "content": "2.2. One-Stage Scene Graph Generation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 220, + 295, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 220, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 220, + 295, + 422 + ], + "type": "text", + "content": "Contrary to two-stage methods, which primarily suffer from the limited ability of the relation detector to directly utilize features extracted by the detector, one-stage approaches [4, 9, 16-18, 25] overcome this limitation. They achieve this by integrating object detector and graph generation architectures, enabling end-to-end training that facilitates direct access and utilization of extracted features. One-stage methods mainly consists of fully convolutional networks [28] and transformer-based models [4, 9, 16-18, 25]. This configuration facilitates direct interactions between the detection and graph generation modules, significantly enhancing the quality of the resulting scene graphs. The current approaches in one-stage methods primarily concentrate on component-level learning, analyzing subjects, objects, and predicates individually given image feature. These methods [16-18] typically compile these components into tuples " + }, + { + "bbox": [ + 55, + 220, + 295, + 422 + ], + "type": "inline_equation", + "content": "\\{\\text{subject-predicate-object}\\}" + }, + { + "bbox": [ + 55, + 220, + 295, + 422 + ], + "type": "text", + "content": " during post-processing." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 424, + 295, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 424, + 295, + 543 + ], + "spans": [ + { + "bbox": [ + 55, + 424, + 295, + 543 + ], + "type": "text", + "content": "Detecting accurate predicates and overcoming obfuscation in overlapped visual relation tuples pose significant challenges for current one-stage methods, primarily due to the lack of subject-object contextual features [22] and varying semantics of multi-role objects. We introduce triplet-level mask features as preliminary proposals to facilitate component-level learning in a one-stage architecture. Additionally, our method incorporates a triplet alignment module that aligns triplet and component features, enhancing the recognition of multi-role objects' diverse semantics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 555, + 111, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 111, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 111, + 567 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 574, + 237, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 574, + 237, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 237, + 586 + ], + "type": "text", + "content": "3.1. Formulation and Model Overview" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": "Scene Graph Generation Formulation. Classic scene graph generation task aims to organize a relation graph " + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{G}'" + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{G}' = \\{\\mathbf{s}_k, \\mathbf{o}_k, \\mathbf{p}_k\\}_{k=1}^K" + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": " as a group of visual relation tuples organized from component-level subject " + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_k" + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": ", object " + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{o}_k" + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": ", and predicate " + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_k" + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": ". In our proposed architecture for hybrid representation tailored for scene graph generation, we refine the conventional task formulation by incorporating a triplet mask for each visual relation tuple. Given an image " + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": ", our scene graph " + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{G}" + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": " is represented as a collection of quadruples that simultaneously encapsulates both triplet-level and tra" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 72, + 482, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 482, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 482, + 84 + ], + "type": "text", + "content": "ditional component-level representations:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 380, + 91, + 553, + 106 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 91, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 380, + 91, + 553, + 106 + ], + "type": "interline_equation", + "content": "\\mathbf {G} = \\left\\{\\mathbf {t} _ {k}, \\mathbf {s} _ {k}, \\mathbf {o} _ {k}, \\mathbf {p} _ {k} \\right\\} _ {k = 1} ^ {K}, \\tag {1}", + "image_path": "3e18448cec4a55fababcb262d8a21e8ed57a7cb085b5739f643537ca809190a6.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 114, + 555, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 114, + 555, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 114, + 555, + 304 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 114, + 555, + 304 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_k" + }, + { + "bbox": [ + 313, + 114, + 555, + 304 + ], + "type": "text", + "content": " is the triplet-level representation denoting the triplet mask constructed from the classic component-level representation subject " + }, + { + "bbox": [ + 313, + 114, + 555, + 304 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_k" + }, + { + "bbox": [ + 313, + 114, + 555, + 304 + ], + "type": "text", + "content": ", object " + }, + { + "bbox": [ + 313, + 114, + 555, + 304 + ], + "type": "inline_equation", + "content": "\\mathbf{o}_k" + }, + { + "bbox": [ + 313, + 114, + 555, + 304 + ], + "type": "text", + "content": ", and predicate " + }, + { + "bbox": [ + 313, + 114, + 555, + 304 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_k" + }, + { + "bbox": [ + 313, + 114, + 555, + 304 + ], + "type": "text", + "content": ". Each component-level representation of the quadruple, i.e., " + }, + { + "bbox": [ + 313, + 114, + 555, + 304 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_k, \\mathbf{o}_k, \\mathbf{p}_k" + }, + { + "bbox": [ + 313, + 114, + 555, + 304 + ], + "type": "text", + "content": ", consists of a class label and bounding box. Note the bounding box corresponding to " + }, + { + "bbox": [ + 313, + 114, + 555, + 304 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_k" + }, + { + "bbox": [ + 313, + 114, + 555, + 304 + ], + "type": "text", + "content": " is calculated from the center point of its corresponding " + }, + { + "bbox": [ + 313, + 114, + 555, + 304 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_k" + }, + { + "bbox": [ + 313, + 114, + 555, + 304 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 114, + 555, + 304 + ], + "type": "inline_equation", + "content": "\\mathbf{o}_k" + }, + { + "bbox": [ + 313, + 114, + 555, + 304 + ], + "type": "text", + "content": ". While " + }, + { + "bbox": [ + 313, + 114, + 555, + 304 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_k" + }, + { + "bbox": [ + 313, + 114, + 555, + 304 + ], + "type": "text", + "content": " represents the binary mask of the triplet of " + }, + { + "bbox": [ + 313, + 114, + 555, + 304 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_k, \\mathbf{o}_k, \\mathbf{p}_k" + }, + { + "bbox": [ + 313, + 114, + 555, + 304 + ], + "type": "text", + "content": " with a class label represent predicate class. This augmentation of triplet-level representation, as compared to traditional methodologies [4, 16-18], significantly provides subject-object pair contextual feature and limits explore space for accurate predicate recognition. Furthermore, it fosters exploration of the connectivity among hybrid triplet-level and component-level representations of same visual relations, thereby facilitating mutual enhancement." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 305, + 555, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 305, + 555, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 305, + 555, + 593 + ], + "type": "text", + "content": "Model Overview. Figure 2 describes our overall architecture, which mainly consists of multi-scale image encoder and hybrid reciprocal transformer. We begin by utilizing an image encoder equipped with a ResNet backbone [12] and a transformer encoder [45] to extract multi-scale image visual features " + }, + { + "bbox": [ + 313, + 305, + 555, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 313, + 305, + 555, + 593 + ], + "type": "text", + "content": ". Based on " + }, + { + "bbox": [ + 313, + 305, + 555, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 313, + 305, + 555, + 593 + ], + "type": "text", + "content": ", our hybrid reciprocal transformer incorporates dual-level representations together, triplet mask queries " + }, + { + "bbox": [ + 313, + 305, + 555, + 593 + ], + "type": "inline_equation", + "content": "Q_{\\mathrm{t}}" + }, + { + "bbox": [ + 313, + 305, + 555, + 593 + ], + "type": "text", + "content": " for triplet-level representation and subject queries " + }, + { + "bbox": [ + 313, + 305, + 555, + 593 + ], + "type": "inline_equation", + "content": "Q_{\\mathrm{s}}" + }, + { + "bbox": [ + 313, + 305, + 555, + 593 + ], + "type": "text", + "content": ", object queries " + }, + { + "bbox": [ + 313, + 305, + 555, + 593 + ], + "type": "inline_equation", + "content": "Q_{\\mathrm{o}}" + }, + { + "bbox": [ + 313, + 305, + 555, + 593 + ], + "type": "text", + "content": ", predicate queries " + }, + { + "bbox": [ + 313, + 305, + 555, + 593 + ], + "type": "inline_equation", + "content": "Q_{\\mathrm{p}}" + }, + { + "bbox": [ + 313, + 305, + 555, + 593 + ], + "type": "text", + "content": ", for component-level representation, " + }, + { + "bbox": [ + 313, + 305, + 555, + 593 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 313, + 305, + 555, + 593 + ], + "type": "text", + "content": " means a sequence of queries as " + }, + { + "bbox": [ + 313, + 305, + 555, + 593 + ], + "type": "inline_equation", + "content": "Q = \\{q^{i}\\}_{i=1}^{N}" + }, + { + "bbox": [ + 313, + 305, + 555, + 593 + ], + "type": "text", + "content": ". To leverage triplet-level representation facilitates guidance of component-level learning for accurate predicate recognition, hybrid reciprocal transformer uses iterative conditional chains for condition the process of " + }, + { + "bbox": [ + 313, + 305, + 555, + 593 + ], + "type": "inline_equation", + "content": "Q_{x}; x \\in \\{\\mathrm{s}, \\mathrm{o}, \\mathrm{p}\\}" + }, + { + "bbox": [ + 313, + 305, + 555, + 593 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 313, + 305, + 555, + 593 + ], + "type": "inline_equation", + "content": "Q_{\\mathrm{t}}" + }, + { + "bbox": [ + 313, + 305, + 555, + 593 + ], + "type": "text", + "content": " (Sec. 3.2). Furthermore, to enable bidirectional refinement in hybrid reciprocal transformer, we develop a cross-conditioning mechanism, which leverages an attention module that operates across layer dimension of hybrid reciprocal transformer to reason aligned hybrid representations (Sec. 3.2). Finally, we propose alignment loss that leverages triplet-component matched relationship, to provide multi-role objects with distinct aligned triplet feature and recognize overlapped visual relation tuples accurately (Sec. 3.3)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 599, + 487, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 599, + 487, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 599, + 487, + 612 + ], + "type": "text", + "content": "3.2. Hybrid Reciprocal Transformer" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": "Hybrid Reciprocal Transformer Architecture. Hybrid reciprocal transformer integrates triplet mask, subject, object, and predicate decoders to process input queries " + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "inline_equation", + "content": "Q_{x}" + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "inline_equation", + "content": "x \\in \\{\\mathrm{t},\\mathrm{s},\\mathrm{o},\\mathrm{p}\\}" + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": " via cross-attention with cropping image features " + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": " is cropped through a multi-scale image encoding process, where the final layer of a ResNet-based convolutional network provides a position-aware, flattened image feature representation for a transformer encoder [45, 59]." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8955" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 72, + 547, + 278 + ], + "blocks": [ + { + "bbox": [ + 63, + 72, + 547, + 278 + ], + "lines": [ + { + "bbox": [ + 63, + 72, + 547, + 278 + ], + "spans": [ + { + "bbox": [ + 63, + 72, + 547, + 278 + ], + "type": "image", + "image_path": "f99079d95f29ce1a68b4fc91d1f54a30862c60133d1364aac964f30135fb3caf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 281, + 555, + 338 + ], + "lines": [ + { + "bbox": [ + 54, + 281, + 555, + 338 + ], + "spans": [ + { + "bbox": [ + 54, + 281, + 555, + 338 + ], + "type": "text", + "content": "Figure 2. Model overview of our method. Hybrid reciprocal transformer deploys four multi-layer decoder structures to embed triplet mask and visual relation tuple components learning using hybrid representations. We illustrate the detailed feature flow of each decoder to visualize two interaction paradigms within this transformer. The first paradigm facilitates triplet iterative guidance and enhances contextual feature integration for predicate recognition at the component level. The second paradigm employs aligned triplet and component features to enable bidirectional refinement of the hybrid triplet-level and component-level representations." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 343, + 296, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 296, + 524 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 296, + 524 + ], + "type": "text", + "content": "Specifically, we model the elements of the scene graph—triplet mask, subject, object, and predicate, using multi-layer transformer decoders [1] with iterative processing across its " + }, + { + "bbox": [ + 56, + 343, + 296, + 524 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 56, + 343, + 296, + 524 + ], + "type": "text", + "content": " layers step by step. The decoders are represented as " + }, + { + "bbox": [ + 56, + 343, + 296, + 524 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{t}}" + }, + { + "bbox": [ + 56, + 343, + 296, + 524 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 56, + 343, + 296, + 524 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{s}}" + }, + { + "bbox": [ + 56, + 343, + 296, + 524 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 56, + 343, + 296, + 524 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{o}}" + }, + { + "bbox": [ + 56, + 343, + 296, + 524 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 56, + 343, + 296, + 524 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{p}}" + }, + { + "bbox": [ + 56, + 343, + 296, + 524 + ], + "type": "text", + "content": ", respectively. In each step " + }, + { + "bbox": [ + 56, + 343, + 296, + 524 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 56, + 343, + 296, + 524 + ], + "type": "text", + "content": ", the " + }, + { + "bbox": [ + 56, + 343, + 296, + 524 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 56, + 343, + 296, + 524 + ], + "type": "text", + "content": "-th layer of the triplet-level decoder, " + }, + { + "bbox": [ + 56, + 343, + 296, + 524 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{t}}^{j}" + }, + { + "bbox": [ + 56, + 343, + 296, + 524 + ], + "type": "text", + "content": ", generates masks feature for triplet-wise visual relations in the image. While the " + }, + { + "bbox": [ + 56, + 343, + 296, + 524 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 56, + 343, + 296, + 524 + ], + "type": "text", + "content": "-th layer of each component-level decoder, " + }, + { + "bbox": [ + 56, + 343, + 296, + 524 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{x}^{j}" + }, + { + "bbox": [ + 56, + 343, + 296, + 524 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 56, + 343, + 296, + 524 + ], + "type": "inline_equation", + "content": "x \\in \\{\\mathrm{s}, \\mathrm{o}, \\mathrm{p}\\}" + }, + { + "bbox": [ + 56, + 343, + 296, + 524 + ], + "type": "text", + "content": ", produces component-level part feature of the scene graph " + }, + { + "bbox": [ + 56, + 343, + 296, + 524 + ], + "type": "inline_equation", + "content": "\\mathbf{G}^{j}" + }, + { + "bbox": [ + 56, + 343, + 296, + 524 + ], + "type": "text", + "content": ". To realize triplet-level representation as proposals to boost component-level learning and allow hybrid representations to mutually reason and refine each other, our hybrid reciprocal transformer incorporates a triplet-guided relation generation mechanism with bidirectional learning refinement." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 533, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 296, + 715 + ], + "type": "text", + "content": "Triplet-Guided Relation Learning with Bidirectional Refinement. To implement triplet-guided relation generation and bidirectional triplet-component enhancement, we develop two interaction paradigms between triplet-level and component-level features. The first paradigm involves using mask feature at the triplet level to guide the process of generating visual relations at the component level. This paradigm leverages triplet-level features as proposals to enhance predicate recognition from component-level data by incorporating contextual information from triplet-level representations. The second paradigm is to enable reciprocal sharing of triplet-level and component-level learning representations across the multiple layers of the hybrid reciprocal transformer through the use of an attention mechanism. This allows for streamlined integration of hybrid represen" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 343, + 555, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 343, + 555, + 367 + ], + "spans": [ + { + "bbox": [ + 313, + 343, + 555, + 367 + ], + "type": "text", + "content": "tations and mutual enhancement, as they represent different aspects of aligned identical aligned visual relation tuples." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 368, + 555, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 368, + 555, + 428 + ], + "spans": [ + { + "bbox": [ + 313, + 368, + 555, + 428 + ], + "type": "text", + "content": "Specifically, for a decoder " + }, + { + "bbox": [ + 313, + 368, + 555, + 428 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_x; x \\in \\{\\mathrm{t}, \\mathrm{s}, \\mathrm{o}, \\mathrm{p}\\}" + }, + { + "bbox": [ + 313, + 368, + 555, + 428 + ], + "type": "text", + "content": ", we derive its output query representation by integrating self-attention on the input queries with positional embeddings and encoder-decoder attention using the image feature " + }, + { + "bbox": [ + 313, + 368, + 555, + 428 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 313, + 368, + 555, + 428 + ], + "type": "text", + "content": ". In " + }, + { + "bbox": [ + 313, + 368, + 555, + 428 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 368, + 555, + 428 + ], + "type": "text", + "content": "-th layers, the outputs of decoders are shown as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 370, + 434, + 555, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 434, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 370, + 434, + 555, + 449 + ], + "type": "interline_equation", + "content": "\\widehat {Q} _ {x} ^ {j} = \\left\\{\\widehat {q} _ {x} ^ {i, j} \\right\\} _ {i = 1} ^ {N}; x \\in \\{t, s, o, p \\}. \\tag {2}", + "image_path": "52c2ea9abf42df14fe963e8564bcd3bf6fd54db79425595700b0712b3906301c.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 458, + 554, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 458, + 554, + 543 + ], + "spans": [ + { + "bbox": [ + 313, + 458, + 554, + 543 + ], + "type": "inline_equation", + "content": "\\widehat{Q}_x^j" + }, + { + "bbox": [ + 313, + 458, + 554, + 543 + ], + "type": "text", + "content": " is produced from two types of conditional inputs from the aforementioned two interaction paradigms for hybrid representations respectively. The first is triplet-guided positional embedding " + }, + { + "bbox": [ + 313, + 458, + 554, + 543 + ], + "type": "inline_equation", + "content": "\\bar{P}_x^j = \\{\\bar{p}_x^{i,j}\\}_{i=1}^N" + }, + { + "bbox": [ + 313, + 458, + 554, + 543 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 313, + 458, + 554, + 543 + ], + "type": "inline_equation", + "content": "x \\in \\{\\mathrm{t}, \\mathrm{s}, \\mathrm{o}, \\mathrm{p}\\}" + }, + { + "bbox": [ + 313, + 458, + 554, + 543 + ], + "type": "text", + "content": " based original position embedding " + }, + { + "bbox": [ + 313, + 458, + 554, + 543 + ], + "type": "inline_equation", + "content": "P_x^j = \\{p_x^{i,j}\\}_{i=1}^N" + }, + { + "bbox": [ + 313, + 458, + 554, + 543 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 313, + 458, + 554, + 543 + ], + "type": "inline_equation", + "content": "x \\in \\{\\mathrm{t}, \\mathrm{s}, \\mathrm{o}, \\mathrm{p}\\}" + }, + { + "bbox": [ + 313, + 458, + 554, + 543 + ], + "type": "text", + "content": " respectively. The second is hybrid reciprocal query " + }, + { + "bbox": [ + 313, + 458, + 554, + 543 + ], + "type": "inline_equation", + "content": "\\bar{Q}_x^{j-1}" + }, + { + "bbox": [ + 313, + 458, + 554, + 543 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 313, + 458, + 554, + 543 + ], + "type": "inline_equation", + "content": "x \\in \\{\\mathrm{t}, \\mathrm{s}, \\mathrm{o}, \\mathrm{p}\\}" + }, + { + "bbox": [ + 313, + 458, + 554, + 543 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 458, + 554, + 543 + ], + "type": "inline_equation", + "content": "j-1" + }, + { + "bbox": [ + 313, + 458, + 554, + 543 + ], + "type": "text", + "content": "-th layer." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 544, + 554, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 554, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 554, + 628 + ], + "type": "text", + "content": "Explicitly, the first involves conditioning component-level decoders " + }, + { + "bbox": [ + 313, + 544, + 554, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_x; x \\in \\{\\mathrm{s,o,p}\\}" + }, + { + "bbox": [ + 313, + 544, + 554, + 628 + ], + "type": "text", + "content": " on the step " + }, + { + "bbox": [ + 313, + 544, + 554, + 628 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 544, + 554, + 628 + ], + "type": "text", + "content": " outputs of triplet-level decoder " + }, + { + "bbox": [ + 313, + 544, + 554, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{t}}" + }, + { + "bbox": [ + 313, + 544, + 554, + 628 + ], + "type": "text", + "content": " using guided positional embedding " + }, + { + "bbox": [ + 313, + 544, + 554, + 628 + ], + "type": "inline_equation", + "content": "\\bar{P}_x^j; x \\in \\{\\mathrm{t,s,o,p}\\}" + }, + { + "bbox": [ + 313, + 544, + 554, + 628 + ], + "type": "text", + "content": " and following iterative process [16]. Specifically, " + }, + { + "bbox": [ + 313, + 544, + 554, + 628 + ], + "type": "inline_equation", + "content": "\\hat{Q}_x^j = \\mathcal{T}_x^j (\\bar{Q}_x^{j - 1}|\\bar{P}_x^j,\\mathcal{I}); x \\in \\{\\mathrm{t,s,o,p}\\}" + }, + { + "bbox": [ + 313, + 544, + 554, + 628 + ], + "type": "text", + "content": ", where triplet decoder using original learnable position embedding " + }, + { + "bbox": [ + 313, + 544, + 554, + 628 + ], + "type": "inline_equation", + "content": "\\bar{P}_{\\mathrm{t}}^{j} = P_{\\mathrm{t}}^{j}" + }, + { + "bbox": [ + 313, + 544, + 554, + 628 + ], + "type": "text", + "content": ", and triplet-guided mechanism is described as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 319, + 632, + 553, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 632, + 553, + 647 + ], + "spans": [ + { + "bbox": [ + 319, + 632, + 553, + 647 + ], + "type": "interline_equation", + "content": "\\bar {P} _ {\\mathrm {s}} ^ {j} = P _ {\\mathrm {s}} ^ {j} + \\text {F F N} \\left(\\mathrm {M H A} \\left(P _ {\\mathrm {s}} ^ {j}, \\widehat {Q} _ {\\mathrm {t}} ^ {j}, \\widehat {Q} _ {\\mathrm {t}} ^ {j}\\right)\\right), \\tag {3}", + "image_path": "099f539b7271f6bfc795599f0ba5947fec231481e68f35db1cef3b840371ea75.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 319, + 649, + 553, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 649, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 319, + 649, + 553, + 666 + ], + "type": "interline_equation", + "content": "\\bar {P} _ {\\mathrm {o}} ^ {j} = P _ {\\mathrm {o}} ^ {j} + \\text {F F N} \\left(\\text {M H A} \\left(P _ {\\mathrm {o}} ^ {j}, \\left[ \\widehat {Q} _ {\\mathrm {t}} ^ {j}, \\widehat {Q} _ {\\mathrm {s}} ^ {j} \\right], \\left[ \\widehat {Q} _ {\\mathrm {t}} ^ {j}, \\widehat {Q} _ {\\mathrm {s}} ^ {j} \\right]\\right)\\right), \\tag {4}", + "image_path": "9f83e3bc1b1b39c6e45d46bbbe680476b879dbf883ce09557c7068ef561de27f.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 320, + 666, + 553, + 683 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 666, + 553, + 683 + ], + "spans": [ + { + "bbox": [ + 320, + 666, + 553, + 683 + ], + "type": "interline_equation", + "content": "\\bar {P} _ {\\mathrm {p}} ^ {j} = P _ {\\mathrm {p}} ^ {j} + \\text {F F N} \\left(\\mathrm {M H A} \\left(P _ {\\mathrm {p}} ^ {j}, \\left[ \\widehat {Q} _ {\\mathrm {t}} ^ {j}, \\widehat {Q} _ {\\mathrm {s}} ^ {j}, \\widehat {Q} _ {\\mathrm {o}} ^ {j} \\right], \\left[ \\widehat {Q} _ {\\mathrm {t}} ^ {j}, \\widehat {Q} _ {\\mathrm {s}} ^ {j}, \\widehat {Q} _ {\\mathrm {o}} ^ {j} \\right]\\right)\\right), \\tag {5}", + "image_path": "4495afb1d769a69be80bc946bc340779342271df38eb53c80d27248a9c62567f.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathsf{FFN}(\\cdot)" + }, + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "text", + "content": " is the feed-forward neural network, and " + }, + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathsf{MHA}(\\cdot,\\cdot,\\cdot)" + }, + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "text", + "content": " is the multi-head attention mechanism for" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8956" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 71, + 288, + 157 + ], + "blocks": [ + { + "bbox": [ + 63, + 71, + 288, + 157 + ], + "lines": [ + { + "bbox": [ + 63, + 71, + 288, + 157 + ], + "spans": [ + { + "bbox": [ + 63, + 71, + 288, + 157 + ], + "type": "image", + "image_path": "4521cf47e76fb4f11b729eceebd8cc56f4b6ec67ffe7cf5e41cd5639fcd1d021.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 158, + 295, + 214 + ], + "lines": [ + { + "bbox": [ + 55, + 158, + 295, + 214 + ], + "spans": [ + { + "bbox": [ + 55, + 158, + 295, + 214 + ], + "type": "text", + "content": "Figure 3. Triplet feature alignment loss measures consistency between triplet feature with tripled component-level feature. It encourages similarity between triplet and component feature from identical visual relation tuples to distinguish multi-role object by aligning different triplet feature into component level." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 219, + 296, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 219, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 219, + 296, + 316 + ], + "type": "inline_equation", + "content": "\\mathsf{Q},\\mathsf{K},\\mathsf{V}" + }, + { + "bbox": [ + 55, + 219, + 296, + 316 + ], + "type": "text", + "content": " respectively, " + }, + { + "bbox": [ + 55, + 219, + 296, + 316 + ], + "type": "inline_equation", + "content": "[\\cdot ,\\cdot ]" + }, + { + "bbox": [ + 55, + 219, + 296, + 316 + ], + "type": "text", + "content": " means concatenation of the element in it. While the second paradigm requires " + }, + { + "bbox": [ + 55, + 219, + 296, + 316 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{t}}^{j}" + }, + { + "bbox": [ + 55, + 219, + 296, + 316 + ], + "type": "text", + "content": " conditioned on " + }, + { + "bbox": [ + 55, + 219, + 296, + 316 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_x^{j - 1};x\\in \\{\\mathrm{s,o,p}\\}" + }, + { + "bbox": [ + 55, + 219, + 296, + 316 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 219, + 296, + 316 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_x^j;x\\in \\{\\mathrm{s,o,p}\\}" + }, + { + "bbox": [ + 55, + 219, + 296, + 316 + ], + "type": "text", + "content": " conditioned on " + }, + { + "bbox": [ + 55, + 219, + 296, + 316 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{t}}^{j - 1}" + }, + { + "bbox": [ + 55, + 219, + 296, + 316 + ], + "type": "text", + "content": " to allow bidirectional refinement. This cross-conditioned allows for feature flow enhancement among hybrid learning representations and is achieved through conditional hybrid reciprocal query " + }, + { + "bbox": [ + 55, + 219, + 296, + 316 + ], + "type": "inline_equation", + "content": "\\bar{Q}_x^{j - 1};x\\in" + }, + { + "bbox": [ + 55, + 219, + 296, + 316 + ], + "type": "inline_equation", + "content": "\\{\\mathrm{t,s,o,p}\\}" + }, + { + "bbox": [ + 55, + 219, + 296, + 316 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 323, + 295, + 339 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 323, + 295, + 339 + ], + "spans": [ + { + "bbox": [ + 62, + 323, + 295, + 339 + ], + "type": "interline_equation", + "content": "\\bar {Q} _ {\\mathrm {t}} ^ {j} = \\widehat {Q} _ {\\mathrm {t}} ^ {j} + \\mathsf {F F N} (\\mathsf {M H A} (\\widehat {Q} _ {t} ^ {j}, [ \\widehat {Q} _ {x} ^ {j} ], [ \\widehat {Q} _ {x} ^ {j} ])); x \\in \\{\\mathrm {s}, \\mathrm {o}, \\mathrm {p} \\}, \\tag {6}", + "image_path": "f2cdfb9486c27a22c2508970b151fdcc14f6ad12a5afee2b39e56fb8c65bf38c.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 340, + 295, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 340, + 295, + 357 + ], + "spans": [ + { + "bbox": [ + 62, + 340, + 295, + 357 + ], + "type": "interline_equation", + "content": "\\bar {Q} _ {x} ^ {j} = \\widehat {Q} _ {x} ^ {j} + \\mathsf {F F N} (\\mathsf {M H A} (\\widehat {Q} _ {x} ^ {j}, \\widehat {Q} _ {\\mathrm {t}} ^ {j}, \\widehat {Q} _ {\\mathrm {t}} ^ {j})); x \\in \\{\\mathrm {s}, 0, \\mathrm {p} \\}. \\tag {7}", + "image_path": "2447f68710dbcf608d4b69c705cdb5a9c33b38a38139cc2e1b23beceaec5a0f5.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 365, + 295, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 295, + 496 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 295, + 496 + ], + "type": "text", + "content": "By implementing the two specified conditional input patterns, we enable the generation of triplet-guided relation graph features within the same layer across various segments of hybrid reciprocal transformer, as well as bidirectional enhancement of hybrid representations across multiple layers of hybrid reciprocal transformer. Subsequently, these representations undergo processing by their respective prediction heads to produce desired hybrid outputs. Then alignment loss is applied for supervision, aiming to align hybrid features from the triplet dimension and enhance the differentiation of overlapping visual relation tuples." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 496, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 496, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 496, + 295, + 640 + ], + "type": "text", + "content": "Prediction Heads. We build prediction heads for hybrid triplet-level and component-level features respectively. For triplet-level feature, we follow [6, 58] to refine with masked attention [5] with the multi-scale image feature. Specifically, triplet masks are generated through an iterative process wherein a dot product is applied across multiple scale dimensions along the channel axis between multi-scale image features " + }, + { + "bbox": [ + 55, + 496, + 295, + 640 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 55, + 496, + 295, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 496, + 295, + 640 + ], + "type": "inline_equation", + "content": "\\widehat{Q}_{\\mathrm{t}}" + }, + { + "bbox": [ + 55, + 496, + 295, + 640 + ], + "type": "text", + "content": ". Then the sigmoid function is applied to normalize the segmentation masks. For the component-level detection bounding box representation (subject, object, and predicate), we use MLP to output their corresponding box position and class information directly." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 647, + 226, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 226, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 226, + 661 + ], + "type": "text", + "content": "3.3. Triplet Feature Alignment Loss" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Alignment Loss. To enhance the differentiation of similar component features in multi-role objects that belong to overlapping visual relation tuples, we propose a triplet-component alignment loss. This loss function refines the" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "representation of similar component features by motivating multiple queries to delineate distinct roles within a single multi-role object, correlating each query specifically to specific triplet configuration. This mechanism is realized by aligning the computed matched similarity matrix with the ground truth (GT) matching matrix to accurately map between triplet configurations and component representations." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 157, + 553, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 157, + 553, + 337 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 553, + 337 + ], + "type": "text", + "content": "In the example illustrated in Figure 3, we observe a scenario featuring several overlapping visual relationships due to the presence of multi-role objects, such as a " + }, + { + "bbox": [ + 313, + 157, + 553, + 337 + ], + "type": "inline_equation", + "content": "\\underline{bag}" + }, + { + "bbox": [ + 313, + 157, + 553, + 337 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 157, + 553, + 337 + ], + "type": "inline_equation", + "content": "\\underline{dog}" + }, + { + "bbox": [ + 313, + 157, + 553, + 337 + ], + "type": "text", + "content": ". Here, the alignment loss assigns distinct features for different roles to each multi-role object based on their corresponding triplet-level features. Through the iterative matching and optimization process enabled by multiple layers of hybrid reciprocal transformer, the model promotes distinct representations for the multiple roles of a single object in overlapping triplets, using unique query-level features derived from their respective aligned triplet features. Furthermore, it promotes distinctiveness by encouraging both triplet-level and component-level features of various visual tuples to maintain separation, which aids in accurate recognition of complex and overlapping visual relation tuples." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 338, + 555, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 555, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 555, + 473 + ], + "type": "text", + "content": "Specifically, in " + }, + { + "bbox": [ + 313, + 338, + 555, + 473 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 338, + 555, + 473 + ], + "type": "text", + "content": "-th layer of hybrid reciprocal transformer, we calculate predicated similarity matrix given the processed triplet-level query " + }, + { + "bbox": [ + 313, + 338, + 555, + 473 + ], + "type": "inline_equation", + "content": "\\widehat{Q}_{\\mathrm{t}}^{j}" + }, + { + "bbox": [ + 313, + 338, + 555, + 473 + ], + "type": "text", + "content": " and component-level query " + }, + { + "bbox": [ + 313, + 338, + 555, + 473 + ], + "type": "inline_equation", + "content": "\\widehat{Q}_{x}^{j}; x \\in \\{\\mathrm{s}, \\mathrm{o}, \\mathrm{p}\\}" + }, + { + "bbox": [ + 313, + 338, + 555, + 473 + ], + "type": "text", + "content": ". We first produce aligned query representation as " + }, + { + "bbox": [ + 313, + 338, + 555, + 473 + ], + "type": "inline_equation", + "content": "\\widetilde{Q}_{x}^{j}; x \\in \\{\\mathrm{t}, \\mathrm{s}, \\mathrm{o}, \\mathrm{p}\\}" + }, + { + "bbox": [ + 313, + 338, + 555, + 473 + ], + "type": "text", + "content": " using their respective matching results to identity GT results to maintain bitwise corresponding among query level. Based on the aligned " + }, + { + "bbox": [ + 313, + 338, + 555, + 473 + ], + "type": "inline_equation", + "content": "\\widetilde{Q}_{x}^{j}; x \\in \\{\\mathrm{t}, \\mathrm{s}, \\mathrm{o}, \\mathrm{p}\\}" + }, + { + "bbox": [ + 313, + 338, + 555, + 473 + ], + "type": "text", + "content": ", we generate a pseudo triplet-level representation " + }, + { + "bbox": [ + 313, + 338, + 555, + 473 + ], + "type": "inline_equation", + "content": "\\mathbb{Q}_{\\mathrm{t}}^{j}" + }, + { + "bbox": [ + 313, + 338, + 555, + 473 + ], + "type": "text", + "content": " as the weighted sum of the output component-level representations " + }, + { + "bbox": [ + 313, + 338, + 555, + 473 + ], + "type": "inline_equation", + "content": "\\widetilde{Q}_{x}^{j}; x \\in \\{\\mathrm{s}, \\mathrm{o}, \\mathrm{p}\\}" + }, + { + "bbox": [ + 313, + 338, + 555, + 473 + ], + "type": "text", + "content": ", and calculate the predicted triplet-level similarities matrix " + }, + { + "bbox": [ + 313, + 338, + 555, + 473 + ], + "type": "inline_equation", + "content": "S^{j}" + }, + { + "bbox": [ + 313, + 338, + 555, + 473 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 338, + 555, + 473 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 338, + 555, + 473 + ], + "type": "text", + "content": "-th layer:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 354, + 482, + 553, + 499 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 482, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 354, + 482, + 553, + 499 + ], + "type": "interline_equation", + "content": "\\mathcal {S} ^ {j} = \\mathbb {Q} _ {\\mathrm {t}} ^ {j} \\left(\\widetilde {Q} _ {\\mathrm {t}} ^ {j}\\right) ^ {T}; \\mathbb {Q} _ {\\mathrm {t}} ^ {j} = \\widetilde {Q} _ {\\mathrm {s}} ^ {j} + \\widetilde {Q} _ {\\mathrm {o}} ^ {j} + \\widetilde {Q} _ {\\mathrm {p}} ^ {j}. \\tag {8}", + "image_path": "a303caef4116245e1b9fae82fd5ca522d69d4b9846b6263d391044889639781b.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 508, + 553, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 508, + 553, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 508, + 553, + 556 + ], + "type": "text", + "content": "While the GT similarity matrix " + }, + { + "bbox": [ + 313, + 508, + 553, + 556 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{GT}}" + }, + { + "bbox": [ + 313, + 508, + 553, + 556 + ], + "type": "text", + "content": " is derived from the original matching relationship between the GT triplet-level masks and component-level subjects, objects, and predicates bounding boxes as one-hot formatting." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 557, + 554, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 554, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 554, + 664 + ], + "type": "text", + "content": "Once the GT matrix " + }, + { + "bbox": [ + 313, + 557, + 554, + 664 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{GT}}" + }, + { + "bbox": [ + 313, + 557, + 554, + 664 + ], + "type": "text", + "content": " and predicted similarity matrix " + }, + { + "bbox": [ + 313, + 557, + 554, + 664 + ], + "type": "inline_equation", + "content": "S^j" + }, + { + "bbox": [ + 313, + 557, + 554, + 664 + ], + "type": "text", + "content": " are obtained, binary cross-entropy loss is applied between them to update model parameters. Incorporate with the iterative optimization and matching process within a hybrid reciprocal transformer framework, the alignment loss enhances the model's capability to differentiate overlapping visual relation tuples by promoting a high degree of similarity between matched component-level and triplet-level representations with GT matching results." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 665, + 554, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 712 + ], + "type": "text", + "content": "End-to-End Training. In the end-to-end training process for our method, the overall loss function is the sum of the loss from all " + }, + { + "bbox": [ + 313, + 665, + 554, + 712 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 665, + 554, + 712 + ], + "type": "text", + "content": " layers of hybrid reciprocal transformer, the loss " + }, + { + "bbox": [ + 313, + 665, + 554, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 313, + 665, + 554, + 712 + ], + "type": "text", + "content": " can be written as:" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8957" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 70, + 555, + 309 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 555, + 309 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 555, + 309 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 555, + 309 + ], + "type": "table", + "html": "
MethodR@50R@100mR@50mR@100AvgR@50AvgR@100F@50F@100
Two-Stage, X101-FPN Backbone
Motifs [52] [CVPR2018]32.136.95.56.818.821.99.411.5
VCTree [43] [CVPR202]31.836.16.67.719.221.910.912.7
TDE [44] [CVPR2020]19.423.29.311.114.417.212.615.0
EBM [42] [CVPR2021]20.524.79.711.615.118.213.215.8
BPLSA [11] [ICCV2021]21.725.513.515.717.620.616.619.4
DT2-ACBS [8] [ICCV2021]22.024.415.016.318.520.417.819.5
BGNN [24] [CVPR2021]31.035.810.712.720.924.215.918.7
PE-Net [57] [CVPR2023]30.735.212.414.521.624.917.720.5
VETO [41] [CVPR2023]27.531.58.19.517.820.512.514.6
DRM [22] [CVPR2024]34.038.99.011.221.525.014.217.4
One-Stage, ResNet-50 Backbone
RelTR [7] [TPAMI2023]27.530.710.812.319.221.515.517.6
EGTR [14] [CVPR2024]30.234.37.910.119.122.212.515.6
Hydra-SGG [4] [Arxiv2024]28.433.116.019.722.226.420.524.7
One-Stage, ResNet-101 Backbone
RelDN [53] [CVPR2019]30.334.84.45.417.420.17.79.3
AS-Net [3] [CVPR2021]18.721.16.17.212.414.29.210.7
SGTR [25] [CVPR2022]25.126.612.014.618.620.616.218.9
HOTR [17] [CVPR2021]22.427.16.99.714.718.410.614.3
SpeaQHotR [18] [CVPR2024]24.729.19.612.717.220.913.817.7
ISG† [16] [NIPS2022]29.532.17.48.418.520.311.813.3
SpeaQISG† [18] [CVPR2024]32.936.011.814.122.425.117.420.3
ISG [16] [NIPS2022]27.230.115.016.621.123.419.321.4
SpeaQISG [18] [CVPR2024]32.135.515.117.623.626.620.523.5
Ours34.1(+1.2)38.3(+2.3)16.0(+0.9)20.5(+2.9)25.1(+1.5)29.4(+2.8)21.8(+1.3)26.7(+3.2)
", + "image_path": "f567b8675fb0ca2f06dbfcf9591bf7a05d81d7d9650dfa0aec6c349e33972fc9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 311, + 555, + 334 + ], + "lines": [ + { + "bbox": [ + 55, + 311, + 555, + 334 + ], + "spans": [ + { + "bbox": [ + 55, + 311, + 555, + 334 + ], + "type": "text", + "content": "Table 1. Scene graph generation performance on Visual Genome. The best results among models with ResNet-101 backbone are marked in bold. " + }, + { + "bbox": [ + 55, + 311, + 555, + 334 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 55, + 311, + 555, + 334 + ], + "type": "text", + "content": " denotes the performance without loss re-weighting proposed in [16]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 118, + 338, + 294, + 376 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 338, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 118, + 338, + 294, + 376 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\sum_ {j = 1} ^ {M} \\left(\\mathcal {L} _ {\\mathrm {t}} ^ {j} + \\mathcal {L} _ {\\mathrm {c}} ^ {j} + \\mathcal {L} _ {\\text {a l i g n}} ^ {j}\\right), \\tag {9}", + "image_path": "43d86f1a7804c8d366d2617fa03fbdd380d79ebfd66873a96c20fcf4a696ca3d.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 118, + 376, + 225, + 388 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 376, + 225, + 388 + ], + "spans": [ + { + "bbox": [ + 118, + 376, + 225, + 388 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {t}} = \\mathcal {L} _ {\\mathrm {t}} ^ {\\text {m a s k}} + \\mathcal {L} _ {\\mathrm {t}} ^ {\\text {d i c e}} + \\mathcal {L} _ {\\mathrm {t}} ^ {\\text {c l s}},", + "image_path": "0608bb7fce6b026c7ab00d819375daafb8134ec4369c255078de7052166afa6c.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 118, + 390, + 201, + 403 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 390, + 201, + 403 + ], + "spans": [ + { + "bbox": [ + 118, + 390, + 201, + 403 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {c}} = \\mathcal {L} _ {\\mathrm {c}} ^ {\\mathrm {s}} + \\mathcal {L} _ {\\mathrm {c}} ^ {\\mathrm {o}} + \\mathcal {L} _ {\\mathrm {c}} ^ {\\mathrm {p}},", + "image_path": "f0fbd13e0a3f2fdbb47c89da66f2986ce9054316a6af92b0c96f5359d8863568.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 412, + 296, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 412, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 412, + 296, + 521 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 412, + 296, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{t}}" + }, + { + "bbox": [ + 55, + 412, + 296, + 521 + ], + "type": "text", + "content": " denotes triplet mask loss through Hungarian matching, which containing mask loss " + }, + { + "bbox": [ + 55, + 412, + 296, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{t}}^{\\mathrm{mask}}" + }, + { + "bbox": [ + 55, + 412, + 296, + 521 + ], + "type": "text", + "content": ", dice loss " + }, + { + "bbox": [ + 55, + 412, + 296, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{t}}^{\\mathrm{dice}}" + }, + { + "bbox": [ + 55, + 412, + 296, + 521 + ], + "type": "text", + "content": ", and mask classification loss " + }, + { + "bbox": [ + 55, + 412, + 296, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{t}}^{\\mathrm{cls}}" + }, + { + "bbox": [ + 55, + 412, + 296, + 521 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 412, + 296, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{c}}" + }, + { + "bbox": [ + 55, + 412, + 296, + 521 + ], + "type": "text", + "content": " follows the design of [18] and consists of subject loss " + }, + { + "bbox": [ + 55, + 412, + 296, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{c}}^{\\mathrm{s}}" + }, + { + "bbox": [ + 55, + 412, + 296, + 521 + ], + "type": "text", + "content": ", object loss " + }, + { + "bbox": [ + 55, + 412, + 296, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{c}}^{\\mathrm{o}}" + }, + { + "bbox": [ + 55, + 412, + 296, + 521 + ], + "type": "text", + "content": ", and predicate loss " + }, + { + "bbox": [ + 55, + 412, + 296, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{c}}^{\\mathrm{p}}" + }, + { + "bbox": [ + 55, + 412, + 296, + 521 + ], + "type": "text", + "content": ", each of " + }, + { + "bbox": [ + 55, + 412, + 296, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{c}}^{\\mathrm{s}}" + }, + { + "bbox": [ + 55, + 412, + 296, + 521 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 412, + 296, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{c}}^{\\mathrm{o}}" + }, + { + "bbox": [ + 55, + 412, + 296, + 521 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 412, + 296, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{c}}^{\\mathrm{p}}" + }, + { + "bbox": [ + 55, + 412, + 296, + 521 + ], + "type": "text", + "content": " is constructed from L1 box loss, GIoU loss [36], and classification loss through groupwise query multi-assignment matching [18], " + }, + { + "bbox": [ + 55, + 412, + 296, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{align}}^{j}" + }, + { + "bbox": [ + 55, + 412, + 296, + 521 + ], + "type": "text", + "content": " is alignment loss calculated using binary cross-entropy loss between " + }, + { + "bbox": [ + 55, + 412, + 296, + 521 + ], + "type": "inline_equation", + "content": "S^{j}" + }, + { + "bbox": [ + 55, + 412, + 296, + 521 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 412, + 296, + 521 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{GT}}" + }, + { + "bbox": [ + 55, + 412, + 296, + 521 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 532, + 132, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 532, + 132, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 532, + 132, + 544 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 552, + 163, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 163, + 565 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 163, + 565 + ], + "type": "text", + "content": "4.1. Experiment Setup" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": "Datasets. We demonstrate the effectiveness of our proposed approach on two benchmark datasets. Visual Genome (VG) [19] is utilized as a benchmark for scene graph generation. We leverage the commonly processed subset from [50], which comprises 108k images across 150 object categories and 50 predicate categories. Action Genome (AG) [15] provides scene graph annotations across 9,848 videos within the Charades dataset [39] for studying human-object interactions. It covers 35 object categories and 25 relation categories, deliberately excluding the class person. Adhering to protocols from previous studies [10], we only utilize frames that have been annotated according to [15]. Frames" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 341, + 555, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 341, + 555, + 364 + ], + "spans": [ + { + "bbox": [ + 313, + 341, + 555, + 364 + ], + "type": "text", + "content": "lacking annotations for person or other objects are excluded to ensure the creation of usable scene graphs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 365, + 554, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 554, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 554, + 473 + ], + "type": "text", + "content": "Evaluation Metrics. Consistent with established methodologies in the literature [50, 52], we evaluate our model using the standard metrics Recall@K (R@K) and Mean Recall@K (mR@K). We further quantify performance by calculating both the arithmetic mean (AvgR@K) [23] and the harmonic mean (F@K) [16]. These additional measures provide a more nuanced assessment of model performance across varying predicate frequencies, ensuring a comprehensive evaluation of the balanced performance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 474, + 555, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 555, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 555, + 641 + ], + "type": "text", + "content": "Implementation Details. We utilize a ResNet-101 backbone [12] paired with a six-layer Transformer encoder. For each part in " + }, + { + "bbox": [ + 313, + 474, + 555, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_x" + }, + { + "bbox": [ + 313, + 474, + 555, + 641 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 313, + 474, + 555, + 641 + ], + "type": "inline_equation", + "content": "x \\in \\{\\mathrm{t}, \\mathrm{s}, \\mathrm{o}, \\mathrm{p}\\}" + }, + { + "bbox": [ + 313, + 474, + 555, + 641 + ], + "type": "text", + "content": " of the hybrid reciprocal transformer, we deploy a six-layer transformer decoder [45, 59] architecture. The query number for component-level learning is 600 for " + }, + { + "bbox": [ + 313, + 474, + 555, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_x^j" + }, + { + "bbox": [ + 313, + 474, + 555, + 641 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 313, + 474, + 555, + 641 + ], + "type": "inline_equation", + "content": "x \\in \\{\\mathrm{s}, \\mathrm{o}, \\mathrm{p}\\}" + }, + { + "bbox": [ + 313, + 474, + 555, + 641 + ], + "type": "text", + "content": ", and 300 for " + }, + { + "bbox": [ + 313, + 474, + 555, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{tn}}^j" + }, + { + "bbox": [ + 313, + 474, + 555, + 641 + ], + "type": "text", + "content": ". The triplet matching strategy employs a naive Hungarian algorithm, optimizing for mask loss, dice loss, and classification loss. For component-level matching, we adopt a grouping matching strategy as proposed in [18], ensuring consistency in matched indices across components. The model undergoes pretraining on VG and AG datasets, focusing on object detection and semantic segmentation, respectively. Experiments are conducted on four A40 GPUs with batch size 16." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 647, + 531, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 531, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 531, + 658 + ], + "type": "text", + "content": "4.2. Comparison Results with State-of-the-art" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": "Visual Genome Dataset. In Table 1, we report the performance of our method on VG dataset, the results show that our method achieves new state-of-the-art performance on those one-stage methods based on ResNet-101 backbone." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8958" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 63, + 70, + 289, + 215 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 289, + 215 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 289, + 215 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 289, + 215 + ], + "type": "table", + "html": "
MethodR@20/50mR@20/50F@20/50
Two-Stage, X101-FPN Backbone
VRD [29]10.3 / 10.9-
Freq Prior [52]24.0 / 24.9--
IMP [50]23.9 / 25.5--
MSDN [26]24.0 / 25.6--
Graph R-CNN [51]24.1 / 25.8--
RelDN [53]25.0 / 26.2--
SimpleBase [10]27.9 / 30.48.3 / 9.112.8 / 14.0
One-Stage, ResNet-101 Backbone
ISG [16]31.0 / 37.510.9 / 13.216.1 / 19.5
\\( ISG_{\\dagger} \\) [16]30.1 / 36.519.2 / 22.323.4 / 27.7
Spea\\( Q_{\\text{ISG}} \\) [18]30.8 / 38.011.7 / 13.817.0 / 20.2
Spea\\( Q_{\\text{ISG}_{\\dagger}} \\) [18]31.8 / 39.420.9 / 23.125.2 / 29.1
Ours32.0 / 40.221.3 / 23.625.5 / 29.7
", + "image_path": "72123fc504572586026605f8510bd6cfff348334b017efa127fff1b29d0df2fa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 245, + 296, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 245, + 296, + 412 + ], + "spans": [ + { + "bbox": [ + 55, + 245, + 296, + 412 + ], + "type": "text", + "content": "Our model demonstrates significant enhancements with increases of " + }, + { + "bbox": [ + 55, + 245, + 296, + 412 + ], + "type": "inline_equation", + "content": "+1.2" + }, + { + "bbox": [ + 55, + 245, + 296, + 412 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 245, + 296, + 412 + ], + "type": "inline_equation", + "content": "+2.3" + }, + { + "bbox": [ + 55, + 245, + 296, + 412 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 245, + 296, + 412 + ], + "type": "inline_equation", + "content": "+0.9" + }, + { + "bbox": [ + 55, + 245, + 296, + 412 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 245, + 296, + 412 + ], + "type": "inline_equation", + "content": "+2.9" + }, + { + "bbox": [ + 55, + 245, + 296, + 412 + ], + "type": "text", + "content": " in the R@50/100 and mR@50/100 metrics. These enhancements can be credited to our pioneering hybrid reciprocal architecture, which integrates reasoning across aligned triplets and component features. Furthermore, the synergistic interplay between the two-tiered representation of visual relation tuple significantly augments the overall performance. When compared with two-stage methods employing the X101-FPN backbone, our method achieves state-of-the-art results in R@K, as evidenced by its comparison with the DRM model [22]. Moreover, owing to our one-stage design to jointly optimize predicate recognition and object detection, we substantially outperform two-stage approaches in mR@K." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 413, + 296, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 413, + 296, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 296, + 568 + ], + "type": "text", + "content": "Action Genome Dataset. We additionally report R@20/50, mR20/50, and F@20/50 values on the Action Genome dataset [15] in Table 2. Consistent with the findings observed on the Visual Genome dataset, we establish a new benchmark for state-of-the-art performance across all utilized metrics, including R@20/50, mR@20/50, and F@20/50. This advancement underscores the efficacy of our hybrid reciprocal architecture with triplet alignment in enhancing scene graph generation capabilities. Moreover, the improvement in performance across various datasets demonstrates the robust scalability and adaptability of the hybrid learning framework, which does not necessitate specific modifications to be effective across different contexts." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 575, + 212, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 575, + 212, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 575, + 212, + 588 + ], + "type": "text", + "content": "4.3. Ablation Study and Analysis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": "Alignment Loss. We apply an alignment loss between hybrid triplet and component features to enhance the distinction of multi-role objects in overlapping triplets, by aligning triplet features into the component level. We investigate the effect of the triplet feature alignment loss by setting different weights in the end-to-end training overall loss function. The results metrics in the VG dataset using different weights (0.0, 0.1, 0.2, 0.3, 0.4, 0.5, 1.0) are reported at Figure 4, the weight used in our vanilla model is 0.2. Our findings suggest that increasing the weight of alignment loss" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 320, + 74, + 549, + 149 + ], + "blocks": [ + { + "bbox": [ + 320, + 74, + 549, + 149 + ], + "lines": [ + { + "bbox": [ + 320, + 74, + 549, + 149 + ], + "spans": [ + { + "bbox": [ + 320, + 74, + 549, + 149 + ], + "type": "image", + "image_path": "00063d6094f90b08deb0255954de53456795003940482df8ea136d2065b86a9a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 320, + 151, + 549, + 232 + ], + "blocks": [ + { + "bbox": [ + 320, + 151, + 549, + 232 + ], + "lines": [ + { + "bbox": [ + 320, + 151, + 549, + 232 + ], + "spans": [ + { + "bbox": [ + 320, + 151, + 549, + 232 + ], + "type": "image", + "image_path": "e6300f9fb6ca522991e9505e354d86fd817060e7983780da842f9dfdef500a87.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 230, + 555, + 262 + ], + "lines": [ + { + "bbox": [ + 313, + 230, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 230, + 555, + 262 + ], + "type": "text", + "content": "Figure 4. Scene graph generation performance comparison with Visual Genome dataset using different parameters of weight of alignment loss in overall training loss." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 360, + 262, + 509, + 308 + ], + "blocks": [ + { + "bbox": [ + 55, + 217, + 295, + 239 + ], + "lines": [ + { + "bbox": [ + 55, + 217, + 295, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 295, + 239 + ], + "type": "text", + "content": "Table 2. Scene graph generation performance on Action Genome. Best performance with bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 360, + 262, + 509, + 308 + ], + "lines": [ + { + "bbox": [ + 360, + 262, + 509, + 308 + ], + "spans": [ + { + "bbox": [ + 360, + 262, + 509, + 308 + ], + "type": "table", + "html": "
αc:tR@50R@100mR@50mR@100
0.532.135.316.522.3
1.033.235.716.121.2
2.033.736.816.422.3
", + "image_path": "1ee878197dac3029ab57e207726a5d7fb859188178bb25921c28352e50251bf0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 308, + 555, + 332 + ], + "lines": [ + { + "bbox": [ + 313, + 308, + 555, + 332 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 555, + 332 + ], + "type": "text", + "content": "Table 3. Effect of matching triplet-level and component-level together with different ratio " + }, + { + "bbox": [ + 313, + 308, + 555, + 332 + ], + "type": "inline_equation", + "content": "{\\alpha }_{\\mathrm{c} : \\mathrm{t}}" + }, + { + "bbox": [ + 313, + 308, + 555, + 332 + ], + "type": "text", + "content": " between component and triplet cost." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 340, + 553, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 340, + 553, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 340, + 553, + 389 + ], + "type": "text", + "content": "initially enhances performance as desired, but subsequently deteriorates. This pattern of performance degradation at higher weights of alignment loss indicates that excessive emphasis on alignment loss may hinder learning efficacy." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 389, + 555, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 555, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 555, + 665 + ], + "type": "text", + "content": "Hybrid Representation Guidance Order and Reciprocal Refinement. In our framework, we traditionally utilize the processed query for triplet-level segmentation masks to condition the generation of queries for component-level representations. To explore the impact of reversing this sequence, we instead employ the component-level query as a condition for generating triplet-level representations. The results for this variant, denoted as R@50/100 and mR@50/100, are 33.2, 36.9, 15.3, and 18.9, respectively, when compared with our vanilla model, as detailed in Table 1, it is evident that utilizing the triplet-level query as a condition for component-level learning yields superior performance. This enhancement can be attributed to the global information embedded in the triplet-level query feature, which effectively enhances the learning of local information in the component-level elements—subject, object, and predicate. Additionally, we explore the impact of reciprocal refinement by omitting this feature from our vanilla model. The performance metrics on VG dataset for R@50/100 and mR@50/100 are 32.4, 36.3, 14.4, and 18.5, respectively. The observed decrease in performance underscores the significance of reciprocal refinement in integrating hybrid triplet and component features effectively." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "Grouping Hybrid Representations in Uniform Matcher. Inspired by [18], we explore integrating hybrid representations using a single Hungarian matching mechanism to align them directly. Table 3 presents a comparative anal" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8959" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 76, + 553, + 204 + ], + "blocks": [ + { + "bbox": [ + 57, + 76, + 553, + 204 + ], + "lines": [ + { + "bbox": [ + 57, + 76, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 57, + 76, + 553, + 204 + ], + "type": "image", + "image_path": "4d36f69c497476372929a2ebb076c3021165d744d5f628e9475735601b159455.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 205, + 553, + 225 + ], + "lines": [ + { + "bbox": [ + 55, + 205, + 553, + 225 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 553, + 225 + ], + "type": "text", + "content": "Figure 5. Qualitative analysis of our method compared with the baseline models, SpeaQ [18] with ISG [16], indicates that our method yields significantly enhanced accuracy in the prediction of predicates and object detection. Red marks error." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 96, + 231, + 254, + 300 + ], + "blocks": [ + { + "bbox": [ + 96, + 231, + 254, + 300 + ], + "lines": [ + { + "bbox": [ + 96, + 231, + 254, + 300 + ], + "spans": [ + { + "bbox": [ + 96, + 231, + 254, + 300 + ], + "type": "table", + "html": "
NQtNQcR@50R@100mR@50mR@100
30030033.837.315.218.3
40033.837.415.318.9
50033.937.615.719.4
60030033.937.815.720.1
40034.037.915.920.5
50034.238.216.221.2
", + "image_path": "930721ed36421146b51a5922b45cf4651bb84265777b8320068bf123a1a044c4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 331, + 295, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 331, + 295, + 437 + ], + "spans": [ + { + "bbox": [ + 54, + 331, + 295, + 437 + ], + "type": "text", + "content": "ysis of the outcomes, employing varied matching ratios between component and triplet levels. The findings indicate that merging the hybrid representations via a Hungarian matching mechanism yields inferior results compared to our vanilla model. Specifically, reducing component-level matching negatively impacts performance. This decline is attributed to the stringent constraints imposed by grouping two distinct hybrid representations, which restricts their internal matching and adversely affects overall performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 437, + 295, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 437, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 54, + 437, + 295, + 533 + ], + "type": "text", + "content": "Query Number. We examine the impact of query number for hybrid representations (" + }, + { + "bbox": [ + 54, + 437, + 295, + 533 + ], + "type": "inline_equation", + "content": "N_{Q\\mathrm{t}}" + }, + { + "bbox": [ + 54, + 437, + 295, + 533 + ], + "type": "text", + "content": " for triplet and " + }, + { + "bbox": [ + 54, + 437, + 295, + 533 + ], + "type": "inline_equation", + "content": "N_{Q\\mathrm{c}}" + }, + { + "bbox": [ + 54, + 437, + 295, + 533 + ], + "type": "text", + "content": " for component) in Table 4. The comparative analysis indicates that incremental increases in query numbers consistently enhance outcome performance. This improvement adheres to fundamental principles observed in transformer-based approaches for visual tasks [1, 21, 27, 59] by providing richer representational capacity and coverage [54]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 534, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 534, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 534, + 295, + 713 + ], + "type": "text", + "content": "Matching of Hybrid Representations in Inference Stage. In the inference stage, our vanilla model only uses the bitwise multiply results from the component-level subject, object, and predicate scores. We compare two deformations in Table 5: the first directly calculates multiply of bitwise triplet-level mask scores together, and the second uses Hungarian matching to align two-level output and then multiply them together. Their corresponding results are shown in Table 5. The comparison results show that directly multiply component-level with triplet-level score together will definitely cause negative performance, since the two-level output representation cannot be directly assigned to the specific result subject, object, and predicate tuple. We also find that after the Hungarian matching, the multiplied scores of the hybrid two-level representation will improve performance" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 338, + 231, + 528, + 275 + ], + "blocks": [ + { + "bbox": [ + 55, + 301, + 295, + 323 + ], + "lines": [ + { + "bbox": [ + 55, + 301, + 295, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 301, + 295, + 323 + ], + "type": "text", + "content": "Table 4. Scene graph performance on VG dataset with different number of triplet-level and component-level queries." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 338, + 231, + 528, + 275 + ], + "lines": [ + { + "bbox": [ + 338, + 231, + 528, + 275 + ], + "spans": [ + { + "bbox": [ + 338, + 231, + 528, + 275 + ], + "type": "table", + "html": "
Inference TypeR@50R@100mR@50mR@100
Component Only34.138.316.020.5
Component × Triplet32.535.614.616.5
Component × Triplet34.238.316.120.6
", + "image_path": "be7a41d7c05c793554b8298130ebc820540dd24fd47f789b29c2b6d819f4792c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 277, + 553, + 309 + ], + "lines": [ + { + "bbox": [ + 313, + 277, + 553, + 309 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 553, + 309 + ], + "type": "text", + "content": "Table 5. Comparison results among inference matching between hybrid representations. " + }, + { + "bbox": [ + 313, + 277, + 553, + 309 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 313, + 277, + 553, + 309 + ], + "type": "text", + "content": " means bitwise multiply and " + }, + { + "bbox": [ + 313, + 277, + 553, + 309 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 313, + 277, + 553, + 309 + ], + "type": "text", + "content": " denotes multiply after matching." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 316, + 553, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 553, + 351 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 553, + 351 + ], + "type": "text", + "content": "a bit. We suppose this improvement came from the final aligned hybrid representation by the alignment loss will boost the final scene graph generation performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 355, + 425, + 367 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 355, + 425, + 367 + ], + "spans": [ + { + "bbox": [ + 313, + 355, + 425, + 367 + ], + "type": "text", + "content": "4.4. Qualitative Results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 374, + 553, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 374, + 553, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 374, + 553, + 506 + ], + "type": "text", + "content": "Figure 5 presents qualitative examples comparing prediction results from the SpeaQ [18] and ours. Ours successfully detects challenging samples that contain overlapped complicated visual relation tuples and provides accurate predicate recognition results. Furthermore, we provide more accurate object and predicate detection results compared to previous baselines. These enhancements are attributed to a hybrid reciprocal architecture that incorporates an alignment module to differentiate triplet features and predict the component relation graph, while receiving supervision at both triplet and component levels." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 516, + 387, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 516, + 387, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 516, + 387, + 528 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 533, + 553, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 553, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 553, + 676 + ], + "type": "text", + "content": "We integrate a triplet feature alignment module into a hybrid reciprocal transformer to innovate scene graph generation, starting with triplet mask features guide that enhances predicate recognition. We introduce a triplet alignment loss to differentiate multi-role object features in intersecting visual relation tuples, aligning them with triplet features for clearer distinction. Furthermore, our architecture features a reciprocal refinement module in multiple layers for effective fusion and interaction of hybrid aligned features. Experimental results across multiple datasets confirm our superiority over existing methods, marking a substantial advancement in the field of scene graph generation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "Acknowledgement. This work is supported by the InnoHK Clusters of the Hong Kong SAR Government via the Hong Kong Centre for Logistics Robotics." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8960" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 144 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 144 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 144 + ], + "type": "text", + "content": "[1] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 4, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 147, + 296, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 147, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 61, + 147, + 296, + 180 + ], + "type": "text", + "content": "[2] Guikun Chen, Jin Li, and Wenguan Wang. Scene graph generation with role-playing large language models. arXiv preprint arXiv:2410.15364, 2024. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 182, + 295, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 182, + 295, + 215 + ], + "spans": [ + { + "bbox": [ + 62, + 182, + 295, + 215 + ], + "type": "text", + "content": "[3] Mingfei Chen, Yue Liao, Si Liu, Zhiyuan Chen, Fei Wang, and Chen Qian. Reformulating hoi detection as adaptive set prediction. In CVPR, 2021. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 217, + 295, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 217, + 295, + 259 + ], + "spans": [ + { + "bbox": [ + 62, + 217, + 295, + 259 + ], + "type": "text", + "content": "[4] Minghan Chen, Guikun Chen, Wenguan Wang, and Yi Yang. Hydra-sgg: Hybrid relation assignment for one-stage scene graph generation. arXiv preprint arXiv:2409.10262, 2024. 1, 2, 3, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 262, + 295, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 262, + 295, + 317 + ], + "spans": [ + { + "bbox": [ + 62, + 262, + 295, + 317 + ], + "type": "text", + "content": "[5] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 319, + 295, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 319, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 62, + 319, + 295, + 361 + ], + "type": "text", + "content": "[6] Sehwan Choi, Jungho Kim, Hongjae Shin, and Jun Won Choi. Mask2map: Vectorized hd map construction using bird's eye view segmentation masks. arXiv preprint arXiv:2407.13517, 2024. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 364, + 295, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 364, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 62, + 364, + 295, + 407 + ], + "type": "text", + "content": "[7] Yuren Cong, Michael Ying Yang, and Bodo Rosenhahn. Reltr: Relation transformer for scene graph generation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(9):11169-11183, 2023. 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 409, + 295, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 409, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 62, + 409, + 295, + 453 + ], + "type": "text", + "content": "[8] Alakh Desai, Tz-Ying Wu, Subarna Tripathi, and Nuno Vasconcelos. Learning of visual relations: The devil is in the tails. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15404-15413, 2021. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 455, + 295, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 455, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 62, + 455, + 295, + 510 + ], + "type": "text", + "content": "[9] Qi Dong, Zhuowen Tu, Haofu Liao, Yuting Zhang, Vijay Mahadevan, and Stefano Soatto. Visual relationship detection using part-and-sum transformers with composite queries. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3550-3559, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 512, + 295, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 512, + 295, + 543 + ], + "spans": [ + { + "bbox": [ + 57, + 512, + 295, + 543 + ], + "type": "text", + "content": "[10] Raghav Goyal12 and Leonid Sigal123. A simple baseline for weakly-supervised human-centric relation detection. 2021. 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 545, + 295, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 545, + 295, + 610 + ], + "spans": [ + { + "bbox": [ + 57, + 545, + 295, + 610 + ], + "type": "text", + "content": "[11] Yuyu Guo, Lianli Gao, Xuanhan Wang, Yuxuan Hu, Xing Xu, Xu Lu, Heng Tao Shen, and Jingkuan Song. From general to specific: Informative scene graph generation via balance adjustment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16383-16392, 2021. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 613, + 295, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 295, + 657 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 295, + 657 + ], + "type": "text", + "content": "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "type": "text", + "content": "[13] Ruozhen He, Paola Cascante-Bonilla, Ziyan Yang, Alexander C Berg, and Vicente Ordonez. Improved visual grounding through self-consistent explanations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13095-13105, 2024. 1" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "text", + "content": "[14] Jinbae Im, JeongYeon Nam, Nokyung Park, Hyungmin Lee, and Seunghyun Park. Egtr: Extracting graph from transformer for scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24229-24238, 2024. 1, 2, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 129, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 553, + 184 + ], + "type": "text", + "content": "[15] Jingwei Ji, Ranjay Krishna, Li Fei-Fei, and Juan Carlos Niebles. Action genome: Actions as compositions of spatiotemporal scene graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10236-10247, 2020. 2, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 186, + 553, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 553, + 218 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 553, + 218 + ], + "type": "text", + "content": "[16] Siddhesh Khandelwal and Leonid Signal. Iterative scene graph generation. Advances in Neural Information Processing Systems, 35:24295-24308, 2022. 1, 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 220, + 555, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 220, + 555, + 275 + ], + "spans": [ + { + "bbox": [ + 317, + 220, + 555, + 275 + ], + "type": "text", + "content": "[17] Bumsoo Kim, Junhyun Lee, Jaewoo Kang, Eun-Sol Kim, and Hyunwoo J Kim. Hotr: End-to-end human-object interaction detection with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 74-83, 2021. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 277, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 277, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 317, + 277, + 553, + 342 + ], + "type": "text", + "content": "[18] Jongha Kim, Jihwan Park, Jinyoung Park, Jinyoung Kim, Sehyung Kim, and Hyunwoo J Kim. Groupwise query specialization and quality-aware multi-assignment for transformer-based visual relationship detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28160–28169, 2024. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 343, + 553, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 343, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 317, + 343, + 553, + 408 + ], + "type": "text", + "content": "[19] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 2, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 411, + 553, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 411, + 553, + 463 + ], + "spans": [ + { + "bbox": [ + 317, + 411, + 553, + 463 + ], + "type": "text", + "content": "[20] Sanjoy Kundu and Sathyanarayanan N Aakur. Is-ggt: Iterative scene graph generation with generative transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6292-6301, 2023. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 466, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 466, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 317, + 466, + 553, + 521 + ], + "type": "text", + "content": "[21] Feng Li, Hao Zhang, Shilong Liu, Jian Guo, Lionel M Ni, and Lei Zhang. Dn-detr: Accelerate detr training by introducing query denoising. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13619-13627, 2022. 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 523, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 523, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 317, + 523, + 553, + 578 + ], + "type": "text", + "content": "[22] Jiankai Li, Yunhong Wang, Xiefan Guo, Ruijie Yang, and Weixin Li. Leveraging predicate and triplet learning for scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28369-28379, 2024. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 579, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 579, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 317, + 579, + 553, + 633 + ], + "type": "text", + "content": "[23] Lin Li, Long Chen, Yifeng Huang, Zhimeng Zhang, Songyang Zhang, and Jun Xiao. The devil is in the labels: Noisy label correction for robust scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18869-18878, 2022. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 635, + 555, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 555, + 689 + ], + "type": "text", + "content": "[24] Rongjie Li, Songyang Zhang, Bo Wan, and Xuming He. Bipartite graph network with adaptive message passing for unbiased scene graph generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11109-11119, 2021. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[25] Rongjie Li, Songyang Zhang, and Xuming He. Sgtr: End-to-end scene graph generation with transformer. In proceedings" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "8961" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "text", + "content": "of the IEEE/CVF conference on computer vision and pattern recognition, pages 19486-19496, 2022. 1, 2, 3, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 96, + 295, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 96, + 295, + 148 + ], + "spans": [ + { + "bbox": [ + 57, + 96, + 295, + 148 + ], + "type": "text", + "content": "[26] Yikang Li, Wanli Ouyang, Bolei Zhou, Kun Wang, and Xiaogang Wang. Scene graph generation from objects, phrases and region captions. In Proceedings of the IEEE international conference on computer vision, pages 1261-1270, 2017. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 295, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 295, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 295, + 205 + ], + "type": "text", + "content": "[27] Fanfan Liu, Haoran Wei, Wenzhe Zhao, Guozhen Li, Jingquan Peng, and Zihao Li. Wb-detr: Transformer-based detector without backbone. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2979-2987, 2021. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 206, + 294, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 206, + 294, + 250 + ], + "spans": [ + { + "bbox": [ + 57, + 206, + 294, + 250 + ], + "type": "text", + "content": "[28] Hengyue Liu, Ning Yan, Masood Mortazavi, and Bir Bhanu. Fully convolutional scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11546-11556, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 251, + 295, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 251, + 295, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 251, + 295, + 304 + ], + "type": "text", + "content": "[29] Cewu Lu, Ranjay Krishna, Michael Bernstein, and Li Fei-Fei. Visual relationship detection with language priors. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 852-869. Springer, 2016. 2, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 305, + 295, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 305, + 295, + 369 + ], + "spans": [ + { + "bbox": [ + 56, + 305, + 295, + 369 + ], + "type": "text", + "content": "[30] Yichao Lu, Himanshu Rai, Jason Chang, Boris Knyazev, Guangwei Yu, Shashank Shekhar, Graham W Taylor, and Maksims Volkovs. Context-aware scene graph generation with seq2seq transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15931-15941, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 371, + 295, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 295, + 425 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 295, + 425 + ], + "type": "text", + "content": "[31] Jie Ma, Pinghui Wang, Dechen Kong, Zewei Wang, Jun Liu, Hongbin Pei, and Junzhou Zhao. Robust visual question answering: Datasets, methods, and future challenges. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 426, + 295, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 295, + 480 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 295, + 480 + ], + "type": "text", + "content": "[32] Tao Ma, Bing Bai, Haozhe Lin, Heyuan Wang, Yu Wang, Lin Luo, and Lu Fang. When visual grounding meets gigapixel-level large-scale scenes: Benchmark and approach. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22119-22128, 2024. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 482, + 295, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 482, + 295, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 482, + 295, + 525 + ], + "type": "text", + "content": "[33] Nandita Naik, Christopher Potts, and Elisa Kreiss. Context-vqa: Towards context-aware and purposeful visual question answering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2821-2825, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 526, + 295, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 526, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 526, + 295, + 581 + ], + "type": "text", + "content": "[34] Mengshi Qi, Weijian Li, Zhengyuan Yang, Yunhong Wang, and Jiebo Luo. Attentive relational networks for mapping images to scene graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3957-3966, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 582, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 295, + 624 + ], + "type": "text", + "content": "[35] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE transactions on pattern analysis and machine intelligence, 39(6):1137-1149, 2016. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 625, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 295, + 689 + ], + "type": "text", + "content": "[36] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 658-666, 2019. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "type": "text", + "content": "[37] Shihao Shao, Kaifeng Chen, Arjun Karpur, Qinghua Cui, Andre Araujo, and Bingyi Cao. Global features are all you" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "text", + "content": "need for image retrieval and reranking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11036-11046, 2023. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 108, + 554, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 108, + 554, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 108, + 554, + 173 + ], + "type": "text", + "content": "[38] Suprosanna Shit, Rajat Koner, Bastian Wittmann, Johannes Paetzold, Ivan Ezhov, Hongwei Li, Jiazhen Pan, Sahand Sharifzadeh, Georgios Kaissis, Volker Tresp, et al. Relationformer: A unified framework for image-to-graph generation. In European Conference on Computer Vision, pages 422-439. Springer, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 175, + 554, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 554, + 250 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 554, + 250 + ], + "type": "text", + "content": "[39] Gunnar A Sigurdsson, Gül Varol, Xiaolong Wang, Ali Farhadi, Ivan Laptev, and Abhinav Gupta. Hollywood in homes: Crowdsourcing data collection for activity understanding. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 510-526. Springer, 2016. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 252, + 554, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 252, + 554, + 306 + ], + "spans": [ + { + "bbox": [ + 316, + 252, + 554, + 306 + ], + "type": "text", + "content": "[40] Chull Hwan Song, Jooyoung Yoon, Shunghyun Choi, and Yannis Avrithis. Boosting vision transformers for image retrieval. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 107-117, 2023. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 308, + 554, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 554, + 364 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 554, + 364 + ], + "type": "text", + "content": "[41] Gopika Sudhakaran, Devendra Singh Dhami, Kristian Kersting, and Stefan Roth. Vision relation transformer for unbiased scene graph generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21882-21893, 2023. 1, 2, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 365, + 554, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 554, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 554, + 420 + ], + "type": "text", + "content": "[42] Mohammed Suhail, Abhay Mittal, Behjat Siddiquie, Chris Broaddus, Jayan Eledath, Gerard Medioni, and Leonid Sigal. Energy-based learning for scene graph generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13936-13945, 2021. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 422, + 554, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 554, + 475 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 554, + 475 + ], + "type": "text", + "content": "[43] Kaihua Tang, Hanwang Zhang, Baoyuan Wu, Wenhan Luo, and Wei Liu. Learning to compose dynamic tree structures for visual contexts. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6619-6628, 2019. 2, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 477, + 554, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 477, + 554, + 531 + ], + "spans": [ + { + "bbox": [ + 316, + 477, + 554, + 531 + ], + "type": "text", + "content": "[44] Kaihua Tang, Yulei Niu, Jianqiang Huang, Jiaxin Shi, and Hanwang Zhang. Unbiased scene graph generation from biased training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3716-3725, 2020. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 533, + 554, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 533, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 533, + 554, + 555 + ], + "type": "text", + "content": "[45] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 2, 3, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 557, + 554, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 557, + 554, + 610 + ], + "spans": [ + { + "bbox": [ + 316, + 557, + 554, + 610 + ], + "type": "text", + "content": "[46] Guan Wang, Zhimin Li, Qingchao Chen, and Yang Liu. Oed: Towards one-stage end-to-end dynamic scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27938-27947, 2024. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 613, + 554, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 554, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 554, + 667 + ], + "type": "text", + "content": "[47] Sai Wang, Yutian Lin, and Yu Wu. Omni-q: Omnidirectional scene understanding for unsupervised visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14261-14270, 2024. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "type": "text", + "content": "[48] Wenbin Wang, Ruiping Wang, Shiguang Shan, and Xilin Chen. Sketching image gist: Human-mimetic hierarchical scene graph generation. In European conference on computer vision, pages 222-239. Springer, 2020. 2" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8962" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 642 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "text", + "content": "[49] Hui Wu, Min Wang, Wengang Zhou, Zhenbo Lu, and Houqiang Li. Asymmetric feature fusion for image retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11082-11092, 2023. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "type": "text", + "content": "[50] Danfei Xu, Yuke Zhu, Christopher B Choy, and Li Fei-Fei. Scene graph generation by iterative message passing. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5410-5419, 2017. 2, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 217 + ], + "type": "text", + "content": "[51] Jianwei Yang, Jiasen Lu, Stefan Lee, Dhruv Batra, and Devi Parikh. Graph r-cnn for scene graph generation. In Proceedings of the European conference on computer vision (ECCV), pages 670-685, 2018. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 294, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 294, + 272 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 294, + 272 + ], + "type": "text", + "content": "[52] Rowan Zellers, Mark Yatskar, Sam Thomson, and Yejin Choi. Neural motifs: Scene graph parsing with global context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5831-5840, 2018. 2, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 274, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 274, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 274, + 294, + 308 + ], + "type": "text", + "content": "[53] Ji Zhang, Kevin J Shih, Ahmed Elgammal, Andrew Tao, and Bryan Catanzaro. Graphical contrastive losses for scene graph parsing. In CVPR, 2019. 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 308, + 294, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 308, + 294, + 351 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 294, + 351 + ], + "type": "text", + "content": "[54] Jichuan Zhang, Wei Li, Shuang Cheng, Ya-Li Li, and Shengjin Wang. Dynamic object queries for transformer-based incremental object detection. arXiv preprint arXiv:2407.21687, 2024. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 353, + 294, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 353, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 353, + 294, + 407 + ], + "type": "text", + "content": "[55] Xi Zhang, Feifei Zhang, and Changsheng Xu. Vqacl: A novel visual question answering continual learning setting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19102-19112, 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 409, + 294, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 409, + 294, + 474 + ], + "spans": [ + { + "bbox": [ + 56, + 409, + 294, + 474 + ], + "type": "text", + "content": "[56] Long Zhao, Liangzhe Yuan, Boqing Gong, Yin Cui, Florian Schroff, Ming-Hsuan Yang, Hartwig Adam, and Ting Liu. Unified visual relationship detection with vision and language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6962-6973, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 475, + 294, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 475, + 294, + 529 + ], + "spans": [ + { + "bbox": [ + 56, + 475, + 294, + 529 + ], + "type": "text", + "content": "[57] Chaofan Zheng, Xinyu Lyu, Lianli Gao, Bo Dai, and Jingkuan Song. Prototype-based embedding network for scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22783-22792, 2023. 1, 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 531, + 294, + 596 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 531, + 294, + 596 + ], + "spans": [ + { + "bbox": [ + 56, + 531, + 294, + 596 + ], + "type": "text", + "content": "[58] Yi Zhou, Hui Zhang, Jiaqian Yu, Yifan Yang, Sangil Jung, Seung-In Park, and ByungIn Yoo. Himap: Hybrid representation learning for end-to-end vectorized hd map construction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15396-15406, 2024. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 598, + 294, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 598, + 294, + 642 + ], + "spans": [ + { + "bbox": [ + 56, + 598, + 294, + 642 + ], + "type": "text", + "content": "[59] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable " + }, + { + "bbox": [ + 56, + 598, + 294, + 642 + ], + "type": "inline_equation", + "content": "\\{\\mathrm{detr}\\}" + }, + { + "bbox": [ + 56, + 598, + 294, + 642 + ], + "type": "text", + "content": ": Deformable transformers for end-to-end object detection. In International Conference on Learning Representations, 2021. 3, 6, 8" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8963" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Hybrid-Level Instruction Injection for Video Token Compression in Multi-modal Large Language Models/a0ce533d-bc87-47ce-aab6-707ed1ae09bc_content_list.json b/2025/Hybrid-Level Instruction Injection for Video Token Compression in Multi-modal Large Language Models/a0ce533d-bc87-47ce-aab6-707ed1ae09bc_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3107397bd0f9636f016e5fbd109175b8c98f1739 --- /dev/null +++ b/2025/Hybrid-Level Instruction Injection for Video Token Compression in Multi-modal Large Language Models/a0ce533d-bc87-47ce-aab6-707ed1ae09bc_content_list.json @@ -0,0 +1,1338 @@ +[ + { + "type": "text", + "text": "Hybrid-Level Instruction Injection for Video Token Compression in Multi-modal Large Language Models", + "text_level": 1, + "bbox": [ + 169, + 128, + 828, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhihang Liu1*, Chen-Wei Xie2, Pandeng Li1,2†, Liming Zhao2, Longxiang Tang3, Yun Zheng2, Chuanbin Liu1, Hongtao Xie1 \n1 University of Science and Technology of China \n2 Tongyi Lab, Alibaba Group 3 Tsinghua University \n{liuzhihang, lpd}@mail.ustc.edu.cn, hxie@ustc.edu.cn", + "bbox": [ + 184, + 202, + 810, + 296 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 329, + 326, + 345 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent Multi-modal Large Language Models (MLLMs) have been challenged by the computational overhead resulting from massive video frames, often alleviated through compression strategies. However, the visual content is not equally contributed to user instructions, existing strategies (e.g., average pool) inevitably lead to the loss of potentially useful information. To tackle this, we propose the Hybrid-level Instruction Injection Strategy for Conditional Token Compression in MLLMs (HICom), utilizing the instruction as a condition to guide the compression from both local and global levels. This encourages the compression to retain the maximum amount of user-focused information while reducing visual tokens to minimize computational burden. Specifically, the instruction condition is injected into the grouped visual tokens at the local level and the learnable tokens at the global level, and we conduct the attention mechanism to complete the conditional compression. From the hybrid-level compression, the instruction-relevant visual parts are highlighted while the temporal-spatial structure is also preserved for easier understanding of LLMs. To further unleash the potential of HICom, we introduce a new conditional pre-training stage with our proposed dataset HICom248K. Experiments show that our HICom can obtain distinguished video understanding ability with fewer tokens, increasing the performance by $2.43\\%$ average on three multiple-choice QA benchmarks and saving $78.8\\%$ tokens compared with the SOTA method. The code is available at https://github.com/1ntzm/HICom.", + "bbox": [ + 89, + 359, + 483, + 785 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 810, + 220, + 825 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multi-modal Large Language Models [2, 14, 22, 31, 32, 74] (MLLMs) have gained significant improvements in multi-", + "bbox": [ + 89, + 835, + 482, + 867 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5ef07a75991de3a13c4c1971b74d4b8fe3231e1bba5beccf9cc7f839b45604ac.jpg", + "image_caption": [ + "Figure 1. An example of the video understanding task, and the comparison between the unconditional compression and our proposed conditional compression with hybrid-level instruction injection. We inject instruction at both local and global levels, guiding the compression to retain the maximum amount of user-focused information and minimize the computational burden." + ], + "image_footnote": [], + "bbox": [ + 521, + 329, + 890, + 566 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "modal understanding by integrating visual information into the LLMs, beating various expert methods on downstream tasks [12, 13, 25, 26, 47, 61, 65, 70]. While initially focused on image understanding, more research [30, 34, 36, 40, 56, 57] has shifted towards challenging video tasks. Most MLLMs handle videos by treating them as sequences of images, sampling frames, and concatenating visual tokens from these frames [6, 20, 42, 52, 66].", + "bbox": [ + 511, + 688, + 906, + 808 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Compared to images, videos comprise multiple frames, resulting in more visual tokens and thus higher computational costs. To make the computation affordable, early methods [23, 30, 34] often employ sparse temporal sampling strategies, leading to significant loss of temporal information. Consequently, current MLLMs focus on compression", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Interns at Alibaba Group", + "bbox": [ + 107, + 875, + 250, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author", + "bbox": [ + 109, + 887, + 233, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8568", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing visual tokens to achieve a trade-off between computational costs and video understanding ability. Mainstream MLLMs [20, 59, 71] use a spatial pooling strategy within each frame, leveraging capabilities gained from image data. Some approaches [19, 43, 67] use Q-Former [22] or Resampler [1] to compress the spatial-temporal information into a fixed number of visual tokens. Other methods employ convolution [7], clustering [17], or memory mechanisms [16, 46, 68] to compress visual tokens in both spatial and temporal dimensions. However, these compression methods are unconditional, lacking explicit guidance, as illustrated in Fig. 1 (b), making it challenging to achieve high compression ratios with minimal information loss. Consequently, user-relevant visual content may be overlooked during such unconditional compression.", + "bbox": [ + 89, + 90, + 480, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Therefore, we focus on introducing the instruction information as a condition to perform targeted compression, as it is rarely explored in previous work. As shown in Fig. 1 (c), conditional compression uses explicit instructional guidance to select user-focused information for retention, thereby enhancing efficiency. To thoroughly explore the conditional compression, we propose Hybrid-level Instruction Injection Strategy for Conditional Token Compression in MLLMs (HICom). Observing how humans process complex visual information, they usually adopt a coarse-to-fine approach, first considering the relative parts in the global coarse impression and then seeking detailed information locally in the video. Motivated by this human perception process, HICom conducts the compression at both local and global levels, retaining the maximum amount of instruction-relevant information while preserving the temporal-spatial structure with fewer tokens, thus achieving a better balance between the computational burden and video comprehension.", + "bbox": [ + 91, + 321, + 483, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, at the local level, we divide the sampled frame features into several groups and compress the tokens of each group into one token based on the attention with the injected instruction condition. Different from Q-Former [22] and Resampler [1], the grouped local attention can explicitly maintain the spatial-temporal structural characteristics of the video while highlighting the relevant parts within each group. At the global level, we inject the instruction condition into a small number of learnable tokens and then conduct the attention with the flattened frame features without grouping. The global compression focuses more on searching instruction-relevant information in the whole video and thus can be expressed by a small number of tokens as an auxiliary. To further unleash the potential of HICom, we introduce a new conditional pre-training stage between the alignment and instruction tuning and construct a new dataset HICom-248K of 248K video clips with high-quality instruction-followed descriptions. The proposed new stage continues to pre-train the parameters of", + "bbox": [ + 89, + 613, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "condition injection, further improving the performance. Experiments show that our HICom can achieve state-of-the-art performance on 5 benchmarks with significantly fewer tokens. (e.g., increasing by $2.43\\%$ average on three multiple-choice QA benchmarks and saving $78.8\\%$ tokens compared with LLaVA-Video-7B [72]).", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions are listed as follows:", + "bbox": [ + 529, + 181, + 828, + 195 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a conditional compression method HICom for MLLMs, achieving effective video understanding while significantly reducing the computational burden.", + "- We conduct hybrid-level instruction injection to achieve the conditional compression, and introduce a new conditional pre-training stage with our constructed HICom248K dataset to further unleash the potential of the conditional compressing.", + "- Experiments on various benchmarks show the proposed HICom can achieve better performance with fewer visual tokens, providing valuable insights for MLLMs." + ], + "bbox": [ + 511, + 196, + 903, + 362 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 376, + 653, + 392 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-modal large language models. Based on the powerful capabilities of LLMs among various linguistic tasks, many works try to extend the understanding abilities to the computer vision area. Flamingo [1] and BLIP series [8, 22] successfully explore the usage of Resampler and Q-Former to bridge the visual tokens to language models. LLaVA series [31-33] use a simple MLP as the visual projector to translate visual tokens to the LLM's embedding space, achieving huge success. Recently, more researchers have focused on extending image tasks to video tasks. The biggest challenge for video tasks is to design an effective method to achieve the trade-off between computational costs and video understanding ability. Early methods [23, 30, 34] sparsely sample frames, losing too much temporal information. Mainstream methods [20, 59, 71] use a simple spatial pooling strategy to reduce the number of visual tokens and get a huge success. [19, 67] employ Q-Former to compress the spatial-temporal tokens into a fixed number. Chat-Univi [17] uses DPC-KNN algorithm to cluster dynamic visual tokens. VideoLLaMA2 [7] utilizes convolution both in temporal and spatial dimensions to down-sample tokens. Flash-VStream [68] introduces a learnable memory mechanism to compress online video streaming. All of them try their best to observe more visual information. However, their unconditional compression without explicit guidance inherently leads to the loss of instruction-relevant information. Different from them, we conduct conditional compression by injecting instruction at hybrid levels, only emphasizing the visual parts related to the instruction and allowing the irrelevant parts to be discarded, thus getting a better representation for each question.", + "bbox": [ + 511, + 402, + 906, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-based visual representation for video LLMs. There are some methods using instruction information during the", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8569", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3cfbab6f830d9d93511879a312410f719e193a7a61311e5cc174e7387ec8058e.jpg", + "image_caption": [ + "Figure 2. The framework of our proposed HICom. We propose the hybrid-level instruction injection to conditionally compress video tokens in MLLMs. We extract instruction-relevant information within each grouped sub-region at the local level, and extract it to a fixed number of tokens at the global level. The instruction condition is injected into the attention process to guide the compression." + ], + "image_footnote": [], + "bbox": [ + 155, + 88, + 614, + 380 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a9bf7acc60fc598ee70d1d5952a758fb3a4ed0b08f9a9405fb91c49a3381d944.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 95, + 843, + 377 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "visual encoding. [53, 63] sample the frames based on the CLIP [41] similarity or a localization model using the instruction, [29, 50] make the frame selection strategy trainable. However, these methods only focus on how to select the correct frames in a rough manner, failing to guide the token compression with the key instruction information. LLaMA-VID [28] follows a similar instruction-guided idea, as it introduces the instruction to interact with visual features. However, it roughly simplifies the interacted but uncompressed result into a single token only as auxiliary information for each frame, destroying the spatial structure and failing to discuss the guiding role of the instruction for conditional compression. Therefore, it is only designed for long videos and the performance is limited. To fully explore conditional compression, we propose to inject the instruction at hybrid levels, exploring a better way for conditional compression to alleviate the information loss while maintaining the performance.", + "bbox": [ + 88, + 460, + 485, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 747, + 183, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Instruction-Guided Conditional Compression", + "text_level": 1, + "bbox": [ + 89, + 771, + 475, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fig. 2 shows the framework of our proposed HICom, which focuses on the design between the vision encoder and the LLM to conduct the conditional compression. We inject the instruction condition into the compression process at both local and global levels for better information extraction with spatial-temporal structure maintained.", + "bbox": [ + 89, + 794, + 483, + 883 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Local-Level Compression. Previous works have demon-", + "bbox": [ + 89, + 885, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "strated the importance of the spatial inductive bias when processing visual representations for image MLMs [27, 49]. Inspired by them, we divide the tokens into several groups and perform conditional compression within each group to preserve the temporal-spatial structure of video features. Specifically, given the encoded video frame features $\\mathbf{V} \\in \\mathbb{R}^{T \\times H \\times W \\times D}$ , and the downsampling ratio $(\\alpha_{T}, \\alpha_{H}, \\alpha_{W})$ , we divide $\\mathbf{V}$ into $N_{T} \\cdot N_{H} \\cdot N_{W}$ groups, where $N_{i} = \\lceil i / \\alpha_{i} \\rceil$ , $i \\in \\{T, H, W\\}$ . Therefore, there are $\\alpha_{T} \\cdot \\alpha_{H} \\cdot \\alpha_{W}$ tokens in each group, which can be formulated as $\\mathbf{V}^{t,h,w} \\in \\mathbb{R}^{\\alpha_{T} \\times \\alpha_{H} \\times \\alpha_{W} \\times D}$ , where $0 \\leq \\{t,h,w\\} \\leq \\{N_{T}, N_{H}, N_{W}\\}$ . We first pool $\\mathbf{V}^{t,h,w}$ into one token $\\mathbf{V}_{p}^{t,h,w} \\in \\mathbb{R}^{1 \\times D}$ and inject the instruction condition $C$ (i.e., the text feature of instructions) into $\\mathbf{V}_{p}^{t,h,w}$ . Then the local attention within each group can be calculated by:", + "bbox": [ + 511, + 460, + 908, + 691 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {Z} _ {l} ^ {t, h, w} = \\operatorname {A t t n} \\left[ \\operatorname {I n j} \\left(\\boldsymbol {V} _ {p} ^ {t, h, w}, \\boldsymbol {C}\\right), \\boldsymbol {V} ^ {t, h, w}, \\boldsymbol {V} ^ {t, h, w} \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 699, + 906, + 727 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathrm{Inj}(\\cdot)$ represents the condition injection, $\\mathrm{Attn}(\\cdot)$ donates the attention mechanism [9] and the inputs of $\\mathrm{Attn}(\\cdot)$ are query, key, and value, respectively. Therefore, the visual features within each group are compressed into only one token, preserving the temporal-spatial structure while conditionally highlighting the instruction-relevant parts. Finally, we concatenate $Z_{l}^{t,h,w}$ as the local compressed results $Z_{l}\\in \\mathbb{R}^{N_{T}\\times N_{H}\\times N_{W}\\times D}$ , and use an MLP layer to project $Z_{l}$ into the LLM's embedding space.", + "bbox": [ + 511, + 734, + 906, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Global-Level Compression. Though the temporal-spatial structure can be maintained at the local level, the attention", + "bbox": [ + 511, + 869, + 906, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8570", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1bcb442224e098a92c0c13676ac3877e3cc4d7fc04eb3f20b2e55852e8c797ca.jpg", + "image_caption": [ + "Figure 3. We introduce a new guidance pre-training stage and implement three-stage training for conditional compression." + ], + "image_footnote": [], + "bbox": [ + 117, + 90, + 460, + 311 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "is forced to focus on small sub-regions. However, each sub-region may contain information that is not equally valid for the question, and some sub-regions are even quite irrelevant. Therefore, we also implement conditional compression at the global level to highlight the most relevant parts within the whole video. Specifically, we initialize a small set of learnable tokens $\\pmb{L} \\in \\mathbb{R}^{N_L \\times D}$ , where $N_{L}$ donates the number of tokens, and then inject the instruction condition into them. Instead of grouping the video frame features, we apply 3D position embedding $\\mathrm{Pos}(\\cdot)$ for them and directly flatten them at global-level compression. Then the global compressed tokens $Z_{g} \\in \\mathbb{R}^{N_{L} \\times D}$ can be calculated by:", + "bbox": [ + 88, + 375, + 482, + 556 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {Z} _ {g} = \\operatorname {A t t n} \\left[ \\operatorname {I n j} (\\boldsymbol {L}, \\boldsymbol {C}), \\boldsymbol {V} + \\operatorname {P o s} (\\boldsymbol {V}), \\boldsymbol {V} \\right]. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 569, + 482, + 587 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "An MLP layer is also utilized to project $Z_{g}$ into the LLM's embedding space. And we finally concatenate $Z_{l}$ and $Z_{g}$ together for further understanding of LLM.", + "bbox": [ + 89, + 599, + 482, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Instruction Condition Injection. In order to fulfill the guiding role of the instruction condition, we explore three different types of injection modules, and define them direct injection, coarse injection, and fine injection, respectively. Note the text encoder can obtain both pooled tokens $C_p \\in \\mathbb{R}^{1 \\times D}$ (i.e., global text embedding) and fine-grained tokens $C_f \\in \\mathbb{R}^{L \\times D}$ (i.e., token-level text embedding). Given the pooled token $C_p$ , the direct injection uses an MLP to translate the single token into the injected output without any interaction with $V_p^{t,h,w}$ or $L$ . We employ coarse injection by the adaptive layer norm [39]. Specifically, we use an MLP to regress the scale and shift from $C_p$ , then add to the visual input $V_p^{t,h,w}$ or learnable tokens $L$ , which we represent them as $A$ , after the layer norm. The process can be formulated as:", + "bbox": [ + 89, + 645, + 482, + 872 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {I n j} (\\boldsymbol {A}, \\boldsymbol {C}) = \\operatorname {L N} (\\boldsymbol {A}) \\cdot \\operatorname {s c a l e} \\left(\\boldsymbol {C} _ {p}\\right) + \\operatorname {s h i f t} \\left(\\boldsymbol {C} _ {p}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 883, + 482, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1. The statistical information of our constructed HICom-248K dataset.", + "bbox": [ + 511, + 89, + 906, + 114 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4a1150e8b5d88adcbbdf6193ef8a0c26196888886a856966727d8ad86691c31e.jpg", + "image_caption": [ + "Figure 4. The visualization of data source (left) and video length (right) of our constructed HICom-248K dataset." + ], + "image_footnote": [], + "bbox": [ + 545, + 128, + 874, + 296 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{LN}(\\cdot)$ is the layer norm. Given the fine-grained tokens $C_f$ as condition $C$ , we conduct the fine injection via the cross attention, which can be formulated as:", + "bbox": [ + 511, + 361, + 906, + 405 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {I n j} (\\boldsymbol {A}, \\boldsymbol {C}) = \\operatorname {A t t n} \\left(\\operatorname {L N} (\\boldsymbol {A}), \\boldsymbol {C} _ {f}, \\boldsymbol {C} _ {f}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 417, + 906, + 435 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As the direct injection directly translates the condition to the query of the attention in Eq. (1) and Eq. (2), the token length limits that it is only suitable for local compression as the local branch compresses each group into only one token. On the contrary, the coarse and fine injections are more flexible to fit various situations as the condition is added into $A$ . We conduct the experiments on the three types of injection and finally choose direct injection for local compression and coarse injection for global compression.", + "bbox": [ + 511, + 446, + 906, + 582 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Conditional Pre-training Stage. The existing mainstream methods follow the pipeline of alignment first and then instruction tuning [7, 31, 32, 68], and previous text-based methods [28, 53] also follow this pipeline. However, the image-caption pairs cannot provide valid instruction information in the alignment stage, direct instruction tuning with modules that are not adequately aligned will confuse the guiding process, thus harming the performance. Therefore, we propose a new conditional pre-training stage between the alignment and the instruction tuning to make the training process easier, as shown in Fig. 3. In this stage, we use our constructed instruction-followed descriptions to pre-train the compression module with condition injection, fulfilling the guiding role of the instruction.", + "bbox": [ + 511, + 582, + 908, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Dataset Construction", + "text_level": 1, + "bbox": [ + 511, + 803, + 714, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To pre-train the condition injection in the compression module, we construct a new instruction-followed description dataset HICom-248K. Different from common instruction tuning datasets, HICom-248K focuses on providing only one data type, i.e., the description of the visual content re", + "bbox": [ + 511, + 825, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8571", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/d70ef979cee86fc9b519a06fcec5ff46e81c7bed7f653c647a5779172d88f445.jpg", + "table_caption": [ + "Table 2. Performance comparison between our HICom and other SOTA methods on multiple-choice QA video benchmarks. † means we use the result reproduced by VideoLLaMA2 [7], * indicates we reproduce the results ourselves using the official checkpoint and inference code provided by authors. § donates we inference with a new length of frames trained by sampling 32 frames." + ], + "table_footnote": [], + "table_body": "
MethodsLLM Size# Frames# TokensVideoMME w/o sub.VideoMME w/ sub.MV-BenchEgo-Schema
shortmidlongallshortmidlongall
Video-LLaVA [30]7B8204845.338.036.239.946.140.738.141.643.0-
VideoChat2-Mistral [24]7B16153648.336.335.039.552.839.439.243.860.4-
LLaMA-VID [28]7B1fps2tps---\\(25.9^{\\dagger}\\)----\\(41.9^{\\dagger}\\)\\(38.5^{\\dagger}\\)
Chat-Univi-1.5 [17]7B6444845.740.335.840.651.244.641.845.945.9-
PLLAVA [59]7B162304--------46.6-
LLaVA-Next-Video [71]7B324608\\(49.9^{*}\\)\\(41.4^{*}\\)\\(37.0^{*}\\)\\(42.8^{*}\\)----\\(46.5^{\\dagger}\\)\\(43.9^{\\dagger}\\)
LongVA [69]7B1281843261.150.446.252.661.653.647.654.3--
VideoLLaMA2 [7]7B16115256.045.442.147.9----54.651.7
Tarsier [51]7B162304--------62.656.0
VITA [11]\\(8\\times 7\\text{B}\\)16409665.952.948.655.870.456.250.959.2--
LLaVA-OneVision [20]7B326272\\(70.1^{*}\\)\\(56.4^{*}\\)\\(48.9^{*}\\)\\(58.5^{*}\\)\\(75.8^{*}\\)\\(58.4^{*}\\)\\(51.6^{*}\\)\\(61.9^{*}\\)56.760.1
LLaVA-Video [72]7B326272\\(73.9^{*}\\)\\(57.3^{*}\\)\\(50.4^{*}\\)\\(60.6^{*}\\)\\(76.6^{*}\\)\\(60.3^{*}\\)\\(51.7^{*}\\)\\(62.9^{*}\\)58.657.3
HICom (Ours)1.5B3268061.751.343.952.365.154.747.355.756.450.1
HICom (Ours)7B3268065.757.151.458.168.458.953.160.164.160.5
HICom (Ours)\\( ^§ \\)7B64132867.658.051.258.971.863.754.563.364.762.2
HICom (Ours)\\( ^§ \\)7B128262469.060.251.560.372.764.657.364.864.062.3
", + "bbox": [ + 138, + 142, + 859, + 375 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "lated to the instruction, and does not include other rich types of instruction data (e.g., reasoning, counting, etc).", + "bbox": [ + 89, + 398, + 482, + 428 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Data Collection. To ensure the diversity and quality of our videos, we collect videos from public datasets Panda70M [5] and Ego4D [15]. Panda-70M contains various open-domain videos from YouTube, and Ego4D collects many high-resolution ego-centric human activity videos. Notably, we use the original untrimmed videos to cut clips ourselves since the provided split clips of both datasets are too short, greatly decreasing the content complexity. To balance the number of different types of videos, we pre-define 29 categories [10, 72] using natural language (e.g. A video about cooking activity.) and use InternVideo2 [54] to extract both the video features and the pre-defined sentence features. Based on the similarities between them, we select 1,500 videos for each category and then randomly select additional 10,000 videos from the others to ensure diversity.", + "bbox": [ + 88, + 430, + 482, + 655 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Video Processing. The untrimmed videos are too long for our pre-training, and also severely affect the quality of captions and annotations by existing SOTA models. Therefore, we use PySceneDetect to split each video into shorter ones with a much looser threshold than Panda-70M. As the split clips can sometimes be redundant with repeat semantics, we further extract keyframes for each video and only select the split clips containing the keyframes as our final results. Specifically, inspired by ShareGPT4Video [4], we calculate the CLIP score for densely sampled frames within a video, and select the less similar frames as keyframes. Finally, we further filter out the video clips shorter than 5 seconds and longer than 120 seconds.", + "bbox": [ + 88, + 657, + 482, + 852 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Instruction-Followed Description Generation. For the processed video clips, we use Qwen2-VL-72B-Instruct [52] to systematically generate the instruction-followed descrip", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tion with the CoT technique [55]. Specifically, we first generate the detailed caption of each input video, then take both the video content and caption as input to generate the three instruction-answer pairs for each video. Different from the normal instruction datasets, we require Qwen2-VL-72B-Instruct to meet the following two rules: 1) The instructions should refer to the specific visual information, such as the people or the objects in the video; 2) The instructions should lead to a descriptive response and the answers should describe the object mentioned in the instruction in detail. Finally, we filter the generated results by discarding answers containing invalid information, such as the phrase \"not provide\", or \"not describe\".", + "bbox": [ + 511, + 398, + 906, + 594 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset Statistics. Tab. 1 and Fig. 4 shows some statistics about our constructed dataset. We collect 248K video clips in total from Panda-70M and Ego4D, the average of the video lengths is 25.67 seconds and the distribution is shown in Fig. 4. With the help of Qwen2-VL-72B-Instruct, we generate 739K instruction-followed descriptions, providing sufficient data for our guidance pre-training.", + "bbox": [ + 511, + 595, + 908, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 715, + 645, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 741, + 702, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. We use SigLIP [64] (so400m-patch14-384) as our vision encoder and text encoder, choose Qwen2.5 series [48] as our LLMs, and randomly initialize the compressor. The vision encoder keeps frozen at all stages, the LLM is frozen at the two pre-train stages, and is fine-tuned at the instruction tuning stage. We follow LLaVA-OneVision [20] to choose our training configurations, the global batch size is set to 512 at the alignment stage, 256 at the conditional pre-train and instruction tun", + "bbox": [ + 511, + 763, + 906, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8572", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/47d9c4ccb18de0fff3a1fe7487d8861d148e241fcbb80bcf2e3dd6ccde83b5c4.jpg", + "table_caption": [ + "Table 3. Performance of our HICom and other SOTA methods on open-ended QA video benchmarks. All the compared models are 7B.†, *, and § keep the same meaning with Tab. 2. #F., #T. are short for # Frames, # Tokens, ANet is short for ActivityNet, LLaVA-NV, LLaVA-OV are short for LLaVA-Next-Video, and LLaVA-OneVision. The format we report ANet is Acc/Score." + ], + "table_footnote": [], + "table_body": "
Methods# F.# T.ANetVideoChatGPT Bench
CIDOCUTUCOAvg
Video-LLaVA [30]8204845.3/3.3------
VideoChat2 [24]16153649.1/3.33.022.883.512.662.812.98
Chat-Univi [17]6444845.8/3.22.892.913.462.892.812.99
LLaMA-VID [28]1fps2tps47.4/3.32.963.003.532.462.512.89
LongVLM [58]10030547.6/3.32.762.863.342.393.112.89
ST-LLM [35]1651250.9/3.33.233.053.742.932.813.15
PLLAVA [59]16230456.3/3.53.212.863.622.332.932.99
LLaVA-NV [71]32460853.5/3.23.393.293.922.63.123.26
LongVA [69]12818432-/2.83.053.093.772.443.643.20
VideoLLaMA2 [7]16115250.2/3.33.163.083.692.563.143.13
Tarsier [51]16230459.5/3.6------
SF-LLaVA [60]50368056.3/3.43.092.703.572.523.353.04
LLaVA-OV [20]32627256.6/3.6*3.45*3.00*3.71*2.68*3.14*3.20*
HICom(Ours)-1.5B3268053.0/3.53.092.673.402.412.982.91
HICom(Ours)-7B3268058.3/3.73.292.853.592.673.223.12
HICom(Ours)-7B$64132859.4/3.73.322.923.652.743.353.20
HICom(Ours)-7B$128262459.5/3.73.332.863.652.743.323.18
", + "bbox": [ + 89, + 181, + 483, + 420 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ing stage. We use the learning rate of 1e-3 at the alignment stage, 1e-5 at the instruction tuning stage. At our proposed conditional pre-training stage, we use 1e-3 for the condition injection sub-module and 1e-4 for other parameters in the compressing module. We train 1 epoch for all stages. More implementation details can be found in Appendix A. Datasets. We use LLaVA-558K [32] image-caption pairs for our alignment stage, the constructed 248K HiComPretrain instruction-followed descriptions for conditional pre-train stage. Inspired by [51, 72], we collect 2.68M video instruction data for instruction tuning, including 1.6M from LLaVA-Video [72], 292K from VideoChat2-IT [23] subset, 255K from M4-Instruct-Data [21], 11K from Charades [45], 114K from NTU RGB+D [44], 122K from TVQA [18], and 290K from MiT [38]. We evaluate our model on 5 video benchmarks, including 3 multiple-choice benchmarks VideoMME [10], MVBench [24], EgoSchema [37], and 2 open-ended benchmarks ActivityNet [3], VideoChatGPT Bench [36].", + "bbox": [ + 88, + 444, + 483, + 734 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Main Results", + "text_level": 1, + "bbox": [ + 89, + 742, + 232, + 757 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Multiple-choice benchmarks. In Tab. 2, we compare the performance of our HICom with different SOTA models on three multiple-choice video benchmarks. We sample 32 frames with 680 tokens to train all our models and sample different frames for inference. Our HICom-7B with 2624 tokens inference achieves the best performance on all three benchmarks. Compared to LLaVA-Video with 6272 tokens, our HICom with only 1328 tokens can obtain better performance, increasing the performance by $2.43\\%$ average and", + "bbox": [ + 89, + 763, + 485, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2514a43b55185e37a9ed690d754913cd404f4f0dfe661338a7d5dafd81cc8c51.jpg", + "table_caption": [ + "Table 4. The ablation study on our proposed HICom, $\\alpha_{(T,H,W)}$ donates the downsampling ratio of each dimension, # Tok. donates the number of tokens input to LLMs. 8f means we sample 8 frames for training. The conditional compression combining both local and global achieves the best." + ], + "table_footnote": [], + "table_body": "
Methodsα(T,H,W) # Tok.VideoMME w/o sub.MV-BenchEgo-Schema
shortmidall
Unconditional Compression
avg pool(1,3,3)259239.334.933.035.744.8
8f, avg pool(1,3,3)64836.334.432.034.343.6
avg pool(4,3,3)64836.333.332.234.043.7
local(4,3,3)64836.734.432.034.443.7
globalNone3230.031.628.930.134.6
local+global(4,3,3)68037.134.832.334.744.1
Conditional Compression
local(4,3,3)64838.836.133.136.044.0
globalNone3230.830.329.830.335.5
local+global(4,3,3)68039.036.734.236.645.0
", + "bbox": [ + 516, + 170, + 903, + 356 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cf33cc9a58344753b90323d9f8ffc36d125af9c1015d01b8aab22a56572af191.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 521, + 376, + 705, + 474 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1d5e2e8debfbb43955815382ba001767b69ba0033218181516a7622fcecffc60.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 714, + 376, + 897, + 473 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8b2f7eb8258a19b8f485249833140701bdc71fc8e7fca194afb600d5eae8fea9.jpg", + "image_caption": [ + "Figure 5. The ablation study on different compressing ratios. The figures show the performance on VideoMME-Short (upper left), MVBench (upper right), EgoSchema (lower left), and the inference time of 7B LLM (lower right)." + ], + "image_footnote": [], + "bbox": [ + 522, + 478, + 704, + 575 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c80cb8ee8b1cd166bc58300308856f173b7bb730762aaf9f3b447734015c2b4a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 478, + 897, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "saving $78.8\\%$ tokens, HICom with 2624 tokens gains $3\\%$ averagely when saving $58.2\\%$ tokens, significantly lowering the computational burden. Notably, LLaVA-Video uses numerous additional image data, and it also unfreezes the vision encoder during training, which can both increase performance. It is also noteworthy that though our HICom performs a little worse than LLaVA-Video on VideoMME short videos (less than 2 minutes), we beat LLaVA-Video on both VideoMME medium (4-15 minutes) and long videos (30-60 minutes). We argue that conditional compression is more helpful for long videos. For short videos, the sampled frames are dense enough for unconditional compression, baselines with enough information would naturally perform powerfully. But for long videos, the sampled frames are more sparse with less redundant, our HICom can preserve", + "bbox": [ + 511, + 672, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8573", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/85c8d27682c47939b602102643ebb8d0b374c1758e73318321849c7c5dc35088.jpg", + "table_caption": [ + "Table 5. The ablation study on the conditional pre-training stage with the constructed HICom-248K dataset." + ], + "table_footnote": [], + "table_body": "
MethodsConditional \nPre-trainedVideoMMEw/o subs.MV-BenchEgo-Schema
shortmidlongall
avg poolX36.632.433.234.142.740.6
✓36.333.332.234.043.741.9
w/o inj.X39.433.733.135.442.841.64
✓37.134.832.334.744.142.4
w/ inj.X38.835.134.136.044.041.6
✓39.036.734.236.645.043.5
", + "bbox": [ + 112, + 128, + 459, + 233 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the maximum amount of instruct-relevant information while it is easily suppressed in LLaVA-Video.", + "bbox": [ + 89, + 255, + 482, + 286 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Open-ended benchmarks. We also compare the performance of different models on two open-ended video benchmarks, as shown in Tab. 3. We evaluate all our results via GPT-3.5-Turbo-0125. Our HICom also achieves comparable results on both benchmarks, as arriving at SOTA on ActivityNet and getting the second best on the VideoChatGPT Bench. Our HICom with 1328 tokens is on par with Tarsier with much more training data, and performs similarly to LLaVA-OneVision with much more training data and much more visual tokens, demonstrating the effectiveness of our compression.", + "bbox": [ + 89, + 287, + 482, + 454 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Generalization Ability on Video Length. Due to the design of the local-level compression, we can easily extend the lengths of sampled frames at the inference stage for our trained model with 32 frames, rather than re-training it. As we increase the number of sampled frames, the metrics of our HICom also improve across all these benchmarks. For example, the performance on short and medium videos of VideoMME grows fast from 32 frames to 128 frames at our high compressing ratio, as more useful frames are sampled. The performance on long videos of VideoMME without subtitles changes slightly, and we argue this may caused by the extremely long videos (30-60 minutes). Both 32 frames and 128 frames are too sparse for them. Our HICom can pay attention to the most relevant parts based on the sparsely sampled frames, thus changing slightly.", + "bbox": [ + 89, + 455, + 482, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 695, + 254, + 709 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To save time costs, we choose VideoChat2-IT [23] 896K video data as our instruction tuning data and Qwen2.5-0.5B [48] as our LLM to conduct ablation studies, and extract 32 frames for training unless otherwise specified. This requires less than 28 hours training on 8 A800 GPUs.", + "bbox": [ + 89, + 717, + 482, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Component Analysis. The key components of our HICom are local- and global- level compression. Tab. 4 shows the results of them in the status of both unconditional (i.e., without condition injection) and conditional compression (i.e., with condition injection). Note we use $V_{p}^{t,h,w}$ as query of Eq. (1) in unconditional compression for direct injection. With the same number of input tokens, the un", + "bbox": [ + 89, + 795, + 482, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7b9771fb764f2a29e950823193f9e15dd1ed6ea6cc65c048779000c2b0f36e7f.jpg", + "table_caption": [ + "Table 6. The ablation study on different types of injection." + ], + "table_footnote": [], + "table_body": "
LocalGlobalVideoMME w/o sub.MV-BenchEgo-Schema
shortmidlongall
directcoarse39.036.734.236.645.043.5
directfine39.335.334.336.344.143.5
coarsecoarse39.834.835.036.544.142.7
coarsefine38.436.734.236.444.343.5
finecoarse39.134.734.236.044.242.7
finefine38.834.934.236.043.642.0
", + "bbox": [ + 535, + 114, + 880, + 229 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "conditional local compression can achieve comparable performance with both spatial pooling (line 2) and temporal-spatial pooling (line 3). Due to the lack of temporal-spatial inductive bias and too few tokens, the global compression significantly harms the performance, and local+global performs only slightly better than local as the additional information of the global branch is limited without explicit guidance. Compared with unconditional compression, the local and global conditional compression increase by $0.8\\%$ and $0.5\\%$ , respectively. The local+global achieves the best, increasing by $1.9\\%$ on VideoMME and $1.3\\%$ average on three benchmarks, also increasing by $0.63\\%$ compared with local conditional compression. As the global branch can provide instruction-relevant information from a global sight, combining both is better. It is noteworthy that our conditional compression with 680 tokens can beat the average pooling with 2592 tokens on all benchmarks, reducing $73.8\\%$ visual tokens and increasing by $0.47\\%$ on the performance.", + "bbox": [ + 511, + 250, + 906, + 521 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Compressing Ratio. We explore the effectiveness of the compressing ratio on both unconditional and conditional compression in Fig. 5. Our unconditional compression performs similarly with average pooling on three benchmarks at different compressing ratios, and the conditional compression performs much better than them. It is noteworthy that the speed of performance decreasing of the conditional compression is also lower than the unconditional compression as the compressing ratio becomes higher. This shows instruction-relevant information remains with the guidance, demonstrating the superiority. In the lower right sub-figure of Fig. 5, the compressing ratio of (4,3,3) saves $57.68\\%$ inference time on Qwen2.5-7B than (1,3,3) with only $2.06\\%$ performance loss on Qwen2.5-0.5B. Higher compressing ratios (4,9,9) and (8,9,9) can only save a little more time with much higher performance loss. Therefore, we choose (4,3,3) as our final compressing ratio.", + "bbox": [ + 511, + 522, + 908, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Conditional Pre-training Stage. To evaluate the effectiveness of our proposed conditional pre-training stage, we conduct experiments on whether to use this stage or not on both unconditional and conditional compression in Tab. 5. For average pooling and our unconditional compression, the conditional pre-training stage is also effective on MVBench and EgoSchema since more data is trained, but is hard to be effective on VideoMME. For the conditional compression,", + "bbox": [ + 511, + 780, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8574", + "bbox": [ + 482, + 944, + 516, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ca936bda48ff260a0914c7ee461ef9a2d5ed2daa242042b0d9ee43b4059701de.jpg", + "image_caption": [ + "Figure 6. The visualization of the attention map at both the local and global level in the situation of unconditional (w/o inj.) and conditional (w/ inj.) compression. We indicate the instruction-relevant parts with the red bounding box for easier reading." + ], + "image_footnote": [], + "bbox": [ + 122, + 88, + 877, + 352 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the conditional pre-training stage is effective on all benchmarks, increasing by $1.17\\%$ averagely, demonstrating the effectiveness of not only the constructed data, but also the newly proposed stage.", + "bbox": [ + 88, + 417, + 480, + 477 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Types of Injection. We propose three types of instruction condition injection in Sec. 3.1, direct injection is designed for local compression only, while coarse and fine injection fit both. We conduct the ablation study on them in Tab. 6. Coarse injection performs better than fine injection on all three benchmarks, and both of them perform better than unconditional compression. We argue the reason may be that the pooled instruction feature is more in line with the CLIP training, and the attention of fine injection is harder to converge than the MLP of coarse injection.", + "bbox": [ + 88, + 478, + 482, + 630 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Qualitative Analysis", + "text_level": 1, + "bbox": [ + 89, + 637, + 282, + 652 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To qualitatively figure out how our HICom realizes the conditional compression, we visualize the attention map of the compression module in Fig. 6. It clearly shows where the model pays attention to based on the instructions. Since we conduct the local-level compression within the subregion of each group, the attention map of the local level is more evenly distributed throughout the video than the global level. Compared to unconditional compression, the attention map of conditional compression is more related to the instruction-relevant visual parts. As shown in Fig. 6, the attention maps at both levels successfully focus on the candle holders in the first example, and successfully on the Christmas tree, the apples, candles, and berries in the second example, which means the model will give more weights of these parts during compressing. We also notice that the global-level compression tends to forget some de", + "bbox": [ + 88, + 657, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tailed visual information (e.g., the candle holders in the first and third frame of the first example, the Christmas tree in the second frame of the second example), while the local-level compression is better to capture these details due to group-limited attention. This shows that hybrid-level compression provides different sights, thus helping the model better understand the videos.", + "bbox": [ + 511, + 417, + 906, + 522 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 541, + 633, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we address the challenge of achieving efficient video understanding in multi-modal large language models (MLLMs) through our proposed HICom. Our approach effectively balances the trade-off between computational costs and information retention by integrating user instructions into the compression process. This method not only preserves the temporal-spatial structure of video data but also emphasizes instruction-relevant content through a dual-level compression strategy. Furthermore, by introducing a conditional pre-training stage with proposed HICom-248K dataset, targeted pre-training enhances the efficacy of instruction-injected schema. Our experiments demonstrate the effectiveness, as HICom can maintain distinguished video understanding ability with much fewer tokens. HICom shows a good generalization ability to extend the number of frames during inference, but the fixed frame sampling strategy during training still limits the ability to understand long videos. In the future, we will try to sample frames based on fps and add more long videos for training to further increase the ability. We will also explore the potential of conditional compression on images, especially high-resolution images, making our HICom more powerful.", + "bbox": [ + 511, + 568, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8575", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 240, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is supported by the National Key Research and Development Program of China (2022YFB3104700), the National Nature Science Foundation of China (62425114, 62121002, U23B2028, 62232006). We acknowledge the support of Alibaba Group, the GPU cluster built by MCC Lab of Information Science and Technology Institution, USTC, and USTC super-computing center for providing computational resources for this project.", + "bbox": [ + 89, + 114, + 483, + 237 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 251, + 187, + 266 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. NeurIPS, 35: 23716-23736, 2022. 2", + "[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 1", + "[3] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In CVPR, pages 961-970, 2015. 6", + "[4] Lin Chen, Xilin Wei, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Bin Lin, Zhenyu Tang, et al. Sharegpt4video: Improving video understanding and generation with better captions. arXiv preprint arXiv:2406.04325, 2024. 5", + "[5] Tsai-Shien Chen, Aliaksandr Siarohin, Willi Menapace, Ekaterina Deyneka, Hsiang-wei Chao, Byung Eun Jeon, Yuwei Fang, Hsin-Ying Lee, Jian Ren, Ming-Hsuan Yang, et al. Panda-70m: Captioning 70m videos with multiple cross-modality teachers. In CVPR, pages 13320-13331, 2024. 5, 3", + "[6] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhang-wei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 1", + "[7] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024. 2, 4, 5, 6", + "[8] Wenliang Dai, Junnan Li, DONGXU LI, Anthony Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale N Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. In NeurIPS, pages 49250-49267, 2023. 2", + "[9] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3", + "[10] Chaoyou Fu, Yuhan Dai, Yondong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen," + ], + "bbox": [ + 93, + 276, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 5, 6, 3", + "[11] Chaoyou Fu, Haojia Lin, Zuwei Long, Yunhang Shen, Meng Zhao, Yifan Zhang, Xiong Wang, Di Yin, Long Ma, Xiawu Zheng, et al. Vita: Towards open-source interactive omni multimodal llm. arXiv preprint arXiv:2408.05211, 2024. 5", + "[12] Zuan Gao, Yuxin Wang, Yadong Qu, Boqiang Zhang, Zixiao Wang, Jianjun Xu, and Hongtao Xie. Self-supervised pretraining with symmetric superimposition modeling for scene text recognition. arXiv preprint arXiv:2405.05841, 2024. 1", + "[13] Jiannan Ge, Hongtao Xie, Pandeng Li, Lingxi Xie, Shaobo Min, and Yongdong Zhang. Towards discriminative feature generation for generalized zero-shot learning. IEEE Transactions on Multimedia, 2024. 1", + "[14] Jiannan Ge, Lingxi Xie, Hongtao Xie, Pandeng Li, Xiaopeng Zhang, Yongdong Zhang, and Qi Tian. Alignzeg: Mitigating objective misalignment for zero-shot semantic segmentation. In ECCV, pages 142-161. Springer, 2024. 1", + "[15] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In CVPR, pages 18995-19012, 2022. 5, 3", + "[16] Bo He, Hengduo Li, Young Kyun Jang, Menglin Jia, Xuefei Cao, Ashish Shah, Abhinav Shrivastava, and Ser-Nam Lim. Ma-lmm: Memory-augmented large multimodal model for long-term video understanding. In CVPR, pages 13504-13514, 2024. 2", + "[17] Peng Jin, Ryuichi Takanobu, Wancai Zhang, Xiaochun Cao, and Li Yuan. Chat-univi: Unified visual representation empowers large language models with image and video understanding. In CVPR, pages 13700-13710, 2024. 2, 5, 6", + "[18] Jie Lei, Licheng Yu, Mohit Bansal, and Tamara L Berg. Tvqa: Localized, compositional video question answering. arXiv preprint arXiv:1809.01696, 2018. 6", + "[19] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726, 2023. 2", + "[20] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 1, 2, 5, 6", + "[21] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava next-interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895, 2024. 6", + "[22] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, pages 19730–19742. PMLR, 2023. 1, 2", + "[23] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 1, 2, 6, 7", + "[24] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al." + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8576", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Mvbench: A comprehensive multi-modal video understanding benchmark. In CVPR, pages 22195-22206, 2024. 5, 6", + "[25] Pandeng Li, Chen-Wei Xie, Hongtao Xie, Liming Zhao, Lei Zhang, Yun Zheng, Deli Zhao, and Yongdong Zhang. Momentdiff: Generative video moment retrieval from random to real. NeurIPS, 36:65948–65966, 2023. 1", + "[26] Pandeng Li, Chen-Wei Xie, Liming Zhao, Hongtao Xie, Jiannan Ge, Yun Zheng, Deli Zhao, and Yongdong Zhang. Progressive spatio-temporal prototype matching for text-video retrieval. In ICCV, pages 4100–4110, 2023. 1", + "[27] Wentong Li, Yuqian Yuan, Jian Liu, Dongqi Tang, Song Wang, Jianke Zhu, and Lei Zhang. Tokenpacker: Efficient visual projector for multimodal llm. arXiv preprint arXiv:2407.02392, 2024. 3", + "[28] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In ECCV, pages 323–340. Springer, 2024. 3, 4, 5, 6", + "[29] Jianxin Liang, Xiaojun Meng, Yueqian Wang, Chang Liu, Qun Liu, and Dongyan Zhao. End-to-end video question answering with frame scoring mechanisms and adaptive sampling. arXiv preprint arXiv:2407.15047, 2024. 3", + "[30] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 1, 2, 5, 6", + "[31] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 36, 2023. 1, 2, 4", + "[32] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In CVPR, pages 26296-26306, 2024. 1, 4, 6", + "[33] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024. 2, 1", + "[34] Ruyang Liu, Chen Li, Yixiao Ge, Thomas H Li, Ying Shan, and Ge Li. Bt-adapter: Video conversation is feasible without video instruction tuning. In CVPR, pages 13658-13667, 2024. 1, 2", + "[35] Ruyang Liu, Chen Li, Haoran Tang, Yixiao Ge, Ying Shan, and Ge Li. St-llm: Large language models are effective temporal learners. In European Conference on Computer Vision, pages 1–18. Springer, 2024. 6", + "[36] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. In ACL, 2024. 1, 6", + "[37] Karttikeya Mangalam, Raiymbek Akshulakov, and Jitendra Malik. Egoschema: A diagnostic benchmark for very long-form video language understanding. NeurIPS, 36:46212-46244, 2023. 6", + "[38] Mathew Monfort, Alex Andonian, Bolei Zhou, Kandan Ramakrishnan, Sarah Adel Bargal, Tom Yan, Lisa Brown, Quanfu Fan, Dan Gutfreund, Carl Vondrick, et al. Moments in time dataset: one million videos for event understanding. IEEE TPAMI, 42(2):502-508, 2019. 6", + "[39] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, pages 4195-4205, 2023. 4" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Tianyuan Qu, Longxiang Tang, Bohao Peng, Senqiao Yang, Bei Yu, and Jiaya Jia. Does your vision-language model get lost in the long video sampling dilemma?, 2025. 1", + "[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 3", + "[42] Machel Reid, Nikolay Savinov, Denis Teplyashin, Dmitry Lepikhin, Timothy Lillicrap, Jean-baptiste Alayrac, Radu Soricut, Angeliki Lazaridou, Orhan First, Julian Schrittwieser, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 1", + "[43] Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. In CVPR, pages 14313-14323, 2024. 2", + "[44] Amir Shahroudy, Jun Liu, Tian-Tsong Ng, and Gang Wang. Ntu rgb+d: A large scale dataset for 3d human activity analysis. In CVPR, pages 1010-1019, 2016. 6", + "[45] Gunnar A. Sigurdsson, Gúl Varol, Xiaolong Wang, Ivan Laptev, Ali Farhadi, and Abhinav Gupta. Hollywood in homes: Crowdsourcing data collection for activity understanding. ArXiv e-prints, 2016. 6", + "[46] Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Haozhe Chi, Xun Guo, Tian Ye, Yanting Zhang, et al. Moviechat: From dense token to sparse memory for long video understanding. In CVPR, pages 18221-18232, 2024. 2", + "[47] Longxiang Tang, Zhuotao Tian, Kai Li, Chunming He, Hantao Zhou, Hengshuang Zhao, Xiu Li, and Jiaya Jia. Mind the interference: Retaining pre-trained knowledge in parameter efficient continual learning of vision-language models. In ECCV, pages 346-365. Springer, 2024. 1", + "[48] Qwen Team. Qwen2.5: A party of foundation models, 2024. 5, 7, 1", + "[49] Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, Sai Charitha Akula, Jihan Yang, Shusheng Yang, Adithya Iyer, Xichen Pan, et al. Cambrian1: A fully open, vision-centric exploration of multimodal llms. arXiv preprint arXiv:2406.16860, 2024. 3", + "[50] Haibo Wang, Chenghang Lai, Yixuan Sun, and Weifeng Ge. Weakly supervised gaussian contrastive grounding with large multimodal models for video question answering. arXiv preprint arXiv:2401.10711, 2024. 3", + "[51] Jiawei Wang, Liping Yuan, Yuchen Zhang, and Haomiao Sun. Tarsier: Recipes for training and evaluating large video description models. arXiv preprint arXiv:2407.00634, 2024. 5, 6", + "[52] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1, 5, 3", + "[53] Yizhou Wang, Ruiyi Zhang, Haoliang Wang, Uttaran Bhattacharya, Yun Fu, and Gang Wu. Vaquita: Enhancing align" + ], + "bbox": [ + 516, + 90, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8577", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ment in llm-assisted video understanding. arXiv preprint arXiv:2312.02310, 2023. 3, 4", + "[54] Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Guo Chen, Baoqi Pei, Rongkun Zheng, Jilan Xu, Zun Wang, et al. Intervideo2: Scaling video foundation models for multimodal video understanding. arXiv preprint arXiv:2403.15377, 2024. 5, 3", + "[55] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. NeurIPS, 35:24824-24837, 2022. 5", + "[56] Yujie Wei, Shiwei Zhang, Zhiwu Qing, Hangjie Yuan, Zhiheng Liu, Yu Liu, Yingya Zhang, Jingren Zhou, and Hongming Shan. Dreamvideo: Composing your dream videos with customized subject and motion. In CVPR, pages 6537-6549, 2024. 1", + "[57] Yujie Wei, Shiwei Zhang, Hangjie Yuan, Xiang Wang, Haonan Qiu, Rui Zhao, Yutong Feng, Feng Liu, Zhizhong Huang, Jiaxin Ye, et al. Dreamvideo-2: Zero-shot subject-driven video customization with precise motion control. arXiv preprint arXiv:2410.13830, 2024. 1", + "[58] Yuetian Weng, Mingfei Han, Haoyu He, Xiaojun Chang, and Bohan Zhuang. Longvlm: Efficient long video understanding via large language models. In ECCV, pages 453-470. Springer, 2024. 6", + "[59] Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. Pllava: Parameter-free llava extension from images to videos for video dense captioning. arXiv preprint arXiv:2404.16994, 2024. 2, 5, 6", + "[60] Mingze Xu, Mingfei Gao, Zhe Gan, Hong-You Chen, Zhengfeng Lai, Haiming Gang, Kai Kang, and Afshin Dehghan. Slowfast-llava: A strong training-free baseline for video large language models. arXiv preprint arXiv:2407.15841, 2024. 6", + "[61] Zhipei Xu, Xuanyu Zhang, Runyi Li, Zecheng Tang, Qing Huang, and Jian Zhang. Fakeshield: Explainable image forgery detection and localization via multi-modal large language models. arXiv preprint arXiv:2410.02761, 2024. 1", + "[62] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 1", + "[63] Shoubin Yu, Jaemin Cho, Prateek Yadav, and Mohit Bansal. Self-chained image-language model for video localization and question answering. NeurIPS, 36, 2023. 3", + "[64] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In ICCV, pages 11975-11986, 2023. 5, 1", + "[65] Boqiang Zhang, Hongtao Xie, Zuan Gao, and Yuxin Wang. Choose what you need: Disentangled representation learning for scene text recognition removal and editing. In CVPR, pages 28358-28368, 2024. 1", + "[66] Boqiang Zhang, Kehan Li, Zesen Cheng, Zhiqiang Hu, Yuqian Yuan, Guanzheng Chen, Sicong Leng, Yuming Jiang, Hang Zhang, Xin Li, et al. Videollama 3: Frontier multimodal foundation models for image and video understanding. arXiv preprint arXiv:2501.13106, 2025. 1" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[67] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 2", + "[68] Haoji Zhang, Yiqin Wang, Yansong Tang, Yong Liu, Jiashi Feng, Jifeng Dai, and Xiaojie Jin. Flash-vstream: Memory-based real-time understanding for long video streams. arXiv preprint arXiv:2406.08085, 2024. 2, 4", + "[69] Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision. arXiv preprint arXiv:2406.16852, 2024. 5, 6", + "[70] Xuanyu Zhang, Runyi Li, Jiwen Yu, Youmin Xu, Weiqi Li, and Jian Zhang. Editguard: Versatile image watermarking for tamper localization and copyright protection. In CVPR, pages 11964-11974, 2024. 1", + "[71] Yuanhan Zhang, Bo Li, haotian Liu, Yong jae Lee, Liangke Gui, Di Fu, Jiashi Feng, Ziwei Liu, and Chunyuan Li. Llavanext: A strong zero-shot video understanding model, 2024. 2, 5, 6", + "[72] Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, and Chunyuan Li. Video instruction tuning with synthetic data. arXiv preprint arXiv:2410.02713, 2024. 2, 5, 6, 3", + "[73] Junjie Zhou, Yan Shu, Bo Zhao, Boya Wu, Shitao Xiao, Xi Yang, Yongping Xiong, Bo Zhang, Tiejun Huang, and Zheng Liu. Mlvu: A comprehensive benchmark for multi-task long video understanding. arXiv preprint arXiv:2406.04264, 2024. 6", + "[74] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 1" + ], + "bbox": [ + 516, + 90, + 903, + 556 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "8578", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Hybrid-Level Instruction Injection for Video Token Compression in Multi-modal Large Language Models/a0ce533d-bc87-47ce-aab6-707ed1ae09bc_model.json b/2025/Hybrid-Level Instruction Injection for Video Token Compression in Multi-modal Large Language Models/a0ce533d-bc87-47ce-aab6-707ed1ae09bc_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d97618c2d91750ab84c893326cc6ac6169f96907 --- /dev/null +++ b/2025/Hybrid-Level Instruction Injection for Video Token Compression in Multi-modal Large Language Models/a0ce533d-bc87-47ce-aab6-707ed1ae09bc_model.json @@ -0,0 +1,2202 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.171, + 0.13, + 0.83, + 0.177 + ], + "angle": 0, + "content": "Hybrid-Level Instruction Injection for Video Token Compression in Multi-modal Large Language Models" + }, + { + "type": "text", + "bbox": [ + 0.186, + 0.203, + 0.812, + 0.297 + ], + "angle": 0, + "content": "Zhihang Liu1*, Chen-Wei Xie2, Pandeng Li1,2†, Liming Zhao2, Longxiang Tang3, Yun Zheng2, Chuanbin Liu1, Hongtao Xie1 \n1 University of Science and Technology of China \n2 Tongyi Lab, Alibaba Group 3 Tsinghua University \n{liuzhihang, lpd}@mail.ustc.edu.cn, hxie@ustc.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.33, + 0.327, + 0.346 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.361, + 0.485, + 0.786 + ], + "angle": 0, + "content": "Recent Multi-modal Large Language Models (MLLMs) have been challenged by the computational overhead resulting from massive video frames, often alleviated through compression strategies. However, the visual content is not equally contributed to user instructions, existing strategies (e.g., average pool) inevitably lead to the loss of potentially useful information. To tackle this, we propose the Hybrid-level Instruction Injection Strategy for Conditional Token Compression in MLLMs (HICom), utilizing the instruction as a condition to guide the compression from both local and global levels. This encourages the compression to retain the maximum amount of user-focused information while reducing visual tokens to minimize computational burden. Specifically, the instruction condition is injected into the grouped visual tokens at the local level and the learnable tokens at the global level, and we conduct the attention mechanism to complete the conditional compression. From the hybrid-level compression, the instruction-relevant visual parts are highlighted while the temporal-spatial structure is also preserved for easier understanding of LLMs. To further unleash the potential of HICom, we introduce a new conditional pre-training stage with our proposed dataset HICom248K. Experiments show that our HICom can obtain distinguished video understanding ability with fewer tokens, increasing the performance by \\(2.43\\%\\) average on three multiple-choice QA benchmarks and saving \\(78.8\\%\\) tokens compared with the SOTA method. The code is available at https://github.com/1ntzm/HICom." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.811, + 0.222, + 0.827 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.837, + 0.483, + 0.868 + ], + "angle": 0, + "content": "Multi-modal Large Language Models [2, 14, 22, 31, 32, 74] (MLLMs) have gained significant improvements in multi-" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.33, + 0.892, + 0.567 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.576, + 0.908, + 0.66 + ], + "angle": 0, + "content": "Figure 1. An example of the video understanding task, and the comparison between the unconditional compression and our proposed conditional compression with hybrid-level instruction injection. We inject instruction at both local and global levels, guiding the compression to retain the maximum amount of user-focused information and minimize the computational burden." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.809 + ], + "angle": 0, + "content": "modal understanding by integrating visual information into the LLMs, beating various expert methods on downstream tasks [12, 13, 25, 26, 47, 61, 65, 70]. While initially focused on image understanding, more research [30, 34, 36, 40, 56, 57] has shifted towards challenging video tasks. Most MLLMs handle videos by treating them as sequences of images, sampling frames, and concatenating visual tokens from these frames [6, 20, 42, 52, 66]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Compared to images, videos comprise multiple frames, resulting in more visual tokens and thus higher computational costs. To make the computation affordable, early methods [23, 30, 34] often employ sparse temporal sampling strategies, leading to significant loss of temporal information. Consequently, current MLLMs focus on compression" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.251, + 0.888 + ], + "angle": 0, + "content": "*Interns at Alibaba Group" + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.235, + 0.901 + ], + "angle": 0, + "content": "† Corresponding author" + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.251, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8568" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.318 + ], + "angle": 0, + "content": "ing visual tokens to achieve a trade-off between computational costs and video understanding ability. Mainstream MLLMs [20, 59, 71] use a spatial pooling strategy within each frame, leveraging capabilities gained from image data. Some approaches [19, 43, 67] use Q-Former [22] or Resampler [1] to compress the spatial-temporal information into a fixed number of visual tokens. Other methods employ convolution [7], clustering [17], or memory mechanisms [16, 46, 68] to compress visual tokens in both spatial and temporal dimensions. However, these compression methods are unconditional, lacking explicit guidance, as illustrated in Fig. 1 (b), making it challenging to achieve high compression ratios with minimal information loss. Consequently, user-relevant visual content may be overlooked during such unconditional compression." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.323, + 0.485, + 0.61 + ], + "angle": 0, + "content": "Therefore, we focus on introducing the instruction information as a condition to perform targeted compression, as it is rarely explored in previous work. As shown in Fig. 1 (c), conditional compression uses explicit instructional guidance to select user-focused information for retention, thereby enhancing efficiency. To thoroughly explore the conditional compression, we propose Hybrid-level Instruction Injection Strategy for Conditional Token Compression in MLLMs (HICom). Observing how humans process complex visual information, they usually adopt a coarse-to-fine approach, first considering the relative parts in the global coarse impression and then seeking detailed information locally in the video. Motivated by this human perception process, HICom conducts the compression at both local and global levels, retaining the maximum amount of instruction-relevant information while preserving the temporal-spatial structure with fewer tokens, thus achieving a better balance between the computational burden and video comprehension." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Specifically, at the local level, we divide the sampled frame features into several groups and compress the tokens of each group into one token based on the attention with the injected instruction condition. Different from Q-Former [22] and Resampler [1], the grouped local attention can explicitly maintain the spatial-temporal structural characteristics of the video while highlighting the relevant parts within each group. At the global level, we inject the instruction condition into a small number of learnable tokens and then conduct the attention with the flattened frame features without grouping. The global compression focuses more on searching instruction-relevant information in the whole video and thus can be expressed by a small number of tokens as an auxiliary. To further unleash the potential of HICom, we introduce a new conditional pre-training stage between the alignment and instruction tuning and construct a new dataset HICom-248K of 248K video clips with high-quality instruction-followed descriptions. The proposed new stage continues to pre-train the parameters of" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.182 + ], + "angle": 0, + "content": "condition injection, further improving the performance. Experiments show that our HICom can achieve state-of-the-art performance on 5 benchmarks with significantly fewer tokens. (e.g., increasing by \\(2.43\\%\\) average on three multiple-choice QA benchmarks and saving \\(78.8\\%\\) tokens compared with LLaVA-Video-7B [72])." + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.183, + 0.829, + 0.196 + ], + "angle": 0, + "content": "Our main contributions are listed as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.198, + 0.905, + 0.242 + ], + "angle": 0, + "content": "- We propose a conditional compression method HICom for MLLMs, achieving effective video understanding while significantly reducing the computational burden." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.243, + 0.905, + 0.318 + ], + "angle": 0, + "content": "- We conduct hybrid-level instruction injection to achieve the conditional compression, and introduce a new conditional pre-training stage with our constructed HICom248K dataset to further unleash the potential of the conditional compressing." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.318, + 0.905, + 0.363 + ], + "angle": 0, + "content": "- Experiments on various benchmarks show the proposed HICom can achieve better performance with fewer visual tokens, providing valuable insights for MLLMs." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.198, + 0.905, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.377, + 0.655, + 0.393 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.403, + 0.907, + 0.87 + ], + "angle": 0, + "content": "Multi-modal large language models. Based on the powerful capabilities of LLMs among various linguistic tasks, many works try to extend the understanding abilities to the computer vision area. Flamingo [1] and BLIP series [8, 22] successfully explore the usage of Resampler and Q-Former to bridge the visual tokens to language models. LLaVA series [31-33] use a simple MLP as the visual projector to translate visual tokens to the LLM's embedding space, achieving huge success. Recently, more researchers have focused on extending image tasks to video tasks. The biggest challenge for video tasks is to design an effective method to achieve the trade-off between computational costs and video understanding ability. Early methods [23, 30, 34] sparsely sample frames, losing too much temporal information. Mainstream methods [20, 59, 71] use a simple spatial pooling strategy to reduce the number of visual tokens and get a huge success. [19, 67] employ Q-Former to compress the spatial-temporal tokens into a fixed number. Chat-Univi [17] uses DPC-KNN algorithm to cluster dynamic visual tokens. VideoLLaMA2 [7] utilizes convolution both in temporal and spatial dimensions to down-sample tokens. Flash-VStream [68] introduces a learnable memory mechanism to compress online video streaming. All of them try their best to observe more visual information. However, their unconditional compression without explicit guidance inherently leads to the loss of instruction-relevant information. Different from them, we conduct conditional compression by injecting instruction at hybrid levels, only emphasizing the visual parts related to the instruction and allowing the irrelevant parts to be discarded, thus getting a better representation for each question." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Text-based visual representation for video LLMs. There are some methods using instruction information during the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8569" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.156, + 0.089, + 0.615, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.097, + 0.844, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.392, + 0.908, + 0.437 + ], + "angle": 0, + "content": "Figure 2. The framework of our proposed HICom. We propose the hybrid-level instruction injection to conditionally compress video tokens in MLLMs. We extract instruction-relevant information within each grouped sub-region at the local level, and extract it to a fixed number of tokens at the global level. The instruction condition is injected into the attention process to guide the compression." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.461, + 0.486, + 0.735 + ], + "angle": 0, + "content": "visual encoding. [53, 63] sample the frames based on the CLIP [41] similarity or a localization model using the instruction, [29, 50] make the frame selection strategy trainable. However, these methods only focus on how to select the correct frames in a rough manner, failing to guide the token compression with the key instruction information. LLaMA-VID [28] follows a similar instruction-guided idea, as it introduces the instruction to interact with visual features. However, it roughly simplifies the interacted but uncompressed result into a single token only as auxiliary information for each frame, destroying the spatial structure and failing to discuss the guiding role of the instruction for conditional compression. Therefore, it is only designed for long videos and the performance is limited. To fully explore conditional compression, we propose to inject the instruction at hybrid levels, exploring a better way for conditional compression to alleviate the information loss while maintaining the performance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.748, + 0.184, + 0.763 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.772, + 0.476, + 0.789 + ], + "angle": 0, + "content": "3.1. Instruction-Guided Conditional Compression" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.884 + ], + "angle": 0, + "content": "Fig. 2 shows the framework of our proposed HICom, which focuses on the design between the vision encoder and the LLM to conduct the conditional compression. We inject the instruction condition into the compression process at both local and global levels for better information extraction with spatial-temporal structure maintained." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Local-Level Compression. Previous works have demon-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.461, + 0.909, + 0.692 + ], + "angle": 0, + "content": "strated the importance of the spatial inductive bias when processing visual representations for image MLMs [27, 49]. Inspired by them, we divide the tokens into several groups and perform conditional compression within each group to preserve the temporal-spatial structure of video features. Specifically, given the encoded video frame features \\(\\mathbf{V} \\in \\mathbb{R}^{T \\times H \\times W \\times D}\\), and the downsampling ratio \\((\\alpha_{T}, \\alpha_{H}, \\alpha_{W})\\), we divide \\(\\mathbf{V}\\) into \\(N_{T} \\cdot N_{H} \\cdot N_{W}\\) groups, where \\(N_{i} = \\lceil i / \\alpha_{i} \\rceil\\), \\(i \\in \\{T, H, W\\}\\). Therefore, there are \\(\\alpha_{T} \\cdot \\alpha_{H} \\cdot \\alpha_{W}\\) tokens in each group, which can be formulated as \\(\\mathbf{V}^{t,h,w} \\in \\mathbb{R}^{\\alpha_{T} \\times \\alpha_{H} \\times \\alpha_{W} \\times D}\\), where \\(0 \\leq \\{t,h,w\\} \\leq \\{N_{T}, N_{H}, N_{W}\\}\\). We first pool \\(\\mathbf{V}^{t,h,w}\\) into one token \\(\\mathbf{V}_{p}^{t,h,w} \\in \\mathbb{R}^{1 \\times D}\\) and inject the instruction condition \\(C\\) (i.e., the text feature of instructions) into \\(\\mathbf{V}_{p}^{t,h,w}\\). Then the local attention within each group can be calculated by:" + }, + { + "type": "equation", + "bbox": [ + 0.534, + 0.7, + 0.907, + 0.728 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {Z} _ {l} ^ {t, h, w} = \\operatorname {A t t n} \\left[ \\operatorname {I n j} \\left(\\boldsymbol {V} _ {p} ^ {t, h, w}, \\boldsymbol {C}\\right), \\boldsymbol {V} ^ {t, h, w}, \\boldsymbol {V} ^ {t, h, w} \\right], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.871 + ], + "angle": 0, + "content": "where \\(\\mathrm{Inj}(\\cdot)\\) represents the condition injection, \\(\\mathrm{Attn}(\\cdot)\\) donates the attention mechanism [9] and the inputs of \\(\\mathrm{Attn}(\\cdot)\\) are query, key, and value, respectively. Therefore, the visual features within each group are compressed into only one token, preserving the temporal-spatial structure while conditionally highlighting the instruction-relevant parts. Finally, we concatenate \\(Z_{l}^{t,h,w}\\) as the local compressed results \\(Z_{l}\\in \\mathbb{R}^{N_{T}\\times N_{H}\\times N_{W}\\times D}\\), and use an MLP layer to project \\(Z_{l}\\) into the LLM's embedding space." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.9 + ], + "angle": 0, + "content": "Global-Level Compression. Though the temporal-spatial structure can be maintained at the local level, the attention" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8570" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.091, + 0.462, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.319, + 0.483, + 0.348 + ], + "angle": 0, + "content": "Figure 3. We introduce a new guidance pre-training stage and implement three-stage training for conditional compression." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.375, + 0.483, + 0.557 + ], + "angle": 0, + "content": "is forced to focus on small sub-regions. However, each sub-region may contain information that is not equally valid for the question, and some sub-regions are even quite irrelevant. Therefore, we also implement conditional compression at the global level to highlight the most relevant parts within the whole video. Specifically, we initialize a small set of learnable tokens \\( \\pmb{L} \\in \\mathbb{R}^{N_L \\times D} \\), where \\( N_{L} \\) donates the number of tokens, and then inject the instruction condition into them. Instead of grouping the video frame features, we apply 3D position embedding \\( \\mathrm{Pos}(\\cdot) \\) for them and directly flatten them at global-level compression. Then the global compressed tokens \\( Z_{g} \\in \\mathbb{R}^{N_{L} \\times D} \\) can be calculated by:" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.57, + 0.483, + 0.588 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {Z} _ {g} = \\operatorname {A t t n} \\left[ \\operatorname {I n j} (\\boldsymbol {L}, \\boldsymbol {C}), \\boldsymbol {V} + \\operatorname {P o s} (\\boldsymbol {V}), \\boldsymbol {V} \\right]. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.6, + 0.483, + 0.645 + ], + "angle": 0, + "content": "An MLP layer is also utilized to project \\( Z_{g} \\) into the LLM's embedding space. And we finally concatenate \\( Z_{l} \\) and \\( Z_{g} \\) together for further understanding of LLM." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.646, + 0.483, + 0.873 + ], + "angle": 0, + "content": "Instruction Condition Injection. In order to fulfill the guiding role of the instruction condition, we explore three different types of injection modules, and define them direct injection, coarse injection, and fine injection, respectively. Note the text encoder can obtain both pooled tokens \\( C_p \\in \\mathbb{R}^{1 \\times D} \\) (i.e., global text embedding) and fine-grained tokens \\( C_f \\in \\mathbb{R}^{L \\times D} \\) (i.e., token-level text embedding). Given the pooled token \\( C_p \\), the direct injection uses an MLP to translate the single token into the injected output without any interaction with \\( V_p^{t,h,w} \\) or \\( L \\). We employ coarse injection by the adaptive layer norm [39]. Specifically, we use an MLP to regress the scale and shift from \\( C_p \\), then add to the visual input \\( V_p^{t,h,w} \\) or learnable tokens \\( L \\), which we represent them as \\( A \\), after the layer norm. The process can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.131, + 0.885, + 0.483, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\operatorname {I n j} (\\boldsymbol {A}, \\boldsymbol {C}) = \\operatorname {L N} (\\boldsymbol {A}) \\cdot \\operatorname {s c a l e} \\left(\\boldsymbol {C} _ {p}\\right) + \\operatorname {s h i f t} \\left(\\boldsymbol {C} _ {p}\\right), \\tag {3}\n\\]" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.907, + 0.116 + ], + "angle": 0, + "content": "Table 1. The statistical information of our constructed HICom-248K dataset." + }, + { + "type": "image", + "bbox": [ + 0.547, + 0.129, + 0.875, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.307, + 0.907, + 0.334 + ], + "angle": 0, + "content": "Figure 4. The visualization of data source (left) and video length (right) of our constructed HICom-248K dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.362, + 0.907, + 0.406 + ], + "angle": 0, + "content": "where \\(\\mathrm{LN}(\\cdot)\\) is the layer norm. Given the fine-grained tokens \\(C_f\\) as condition \\(C\\), we conduct the fine injection via the cross attention, which can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.582, + 0.419, + 0.907, + 0.436 + ], + "angle": 0, + "content": "\\[\n\\operatorname {I n j} (\\boldsymbol {A}, \\boldsymbol {C}) = \\operatorname {A t t n} \\left(\\operatorname {L N} (\\boldsymbol {A}), \\boldsymbol {C} _ {f}, \\boldsymbol {C} _ {f}\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.447, + 0.907, + 0.583 + ], + "angle": 0, + "content": "As the direct injection directly translates the condition to the query of the attention in Eq. (1) and Eq. (2), the token length limits that it is only suitable for local compression as the local branch compresses each group into only one token. On the contrary, the coarse and fine injections are more flexible to fit various situations as the condition is added into \\( A \\). We conduct the experiments on the three types of injection and finally choose direct injection for local compression and coarse injection for global compression." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.909, + 0.795 + ], + "angle": 0, + "content": "Conditional Pre-training Stage. The existing mainstream methods follow the pipeline of alignment first and then instruction tuning [7, 31, 32, 68], and previous text-based methods [28, 53] also follow this pipeline. However, the image-caption pairs cannot provide valid instruction information in the alignment stage, direct instruction tuning with modules that are not adequately aligned will confuse the guiding process, thus harming the performance. Therefore, we propose a new conditional pre-training stage between the alignment and the instruction tuning to make the training process easier, as shown in Fig. 3. In this stage, we use our constructed instruction-followed descriptions to pre-train the compression module with condition injection, fulfilling the guiding role of the instruction." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.804, + 0.715, + 0.818 + ], + "angle": 0, + "content": "3.2. Dataset Construction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.901 + ], + "angle": 0, + "content": "To pre-train the condition injection in the compression module, we construct a new instruction-followed description dataset HICom-248K. Different from common instruction tuning datasets, HICom-248K focuses on providing only one data type, i.e., the description of the visual content re" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8571" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.133 + ], + "angle": 0, + "content": "Table 2. Performance comparison between our HICom and other SOTA methods on multiple-choice QA video benchmarks. † means we use the result reproduced by VideoLLaMA2 [7], * indicates we reproduce the results ourselves using the official checkpoint and inference code provided by authors. § donates we inference with a new length of frames trained by sampling 32 frames." + }, + { + "type": "table", + "bbox": [ + 0.139, + 0.143, + 0.86, + 0.375 + ], + "angle": 0, + "content": "
MethodsLLM Size# Frames# TokensVideoMME w/o sub.VideoMME w/ sub.MV-BenchEgo-Schema
shortmidlongallshortmidlongall
Video-LLaVA [30]7B8204845.338.036.239.946.140.738.141.643.0-
VideoChat2-Mistral [24]7B16153648.336.335.039.552.839.439.243.860.4-
LLaMA-VID [28]7B1fps2tps---\\(25.9^{\\dagger}\\)----\\(41.9^{\\dagger}\\)\\(38.5^{\\dagger}\\)
Chat-Univi-1.5 [17]7B6444845.740.335.840.651.244.641.845.945.9-
PLLAVA [59]7B162304--------46.6-
LLaVA-Next-Video [71]7B324608\\(49.9^{*}\\)\\(41.4^{*}\\)\\(37.0^{*}\\)\\(42.8^{*}\\)----\\(46.5^{\\dagger}\\)\\(43.9^{\\dagger}\\)
LongVA [69]7B1281843261.150.446.252.661.653.647.654.3--
VideoLLaMA2 [7]7B16115256.045.442.147.9----54.651.7
Tarsier [51]7B162304--------62.656.0
VITA [11]\\(8\\times 7\\text{B}\\)16409665.952.948.655.870.456.250.959.2--
LLaVA-OneVision [20]7B326272\\(70.1^{*}\\)\\(56.4^{*}\\)\\(48.9^{*}\\)\\(58.5^{*}\\)\\(75.8^{*}\\)\\(58.4^{*}\\)\\(51.6^{*}\\)\\(61.9^{*}\\)56.760.1
LLaVA-Video [72]7B326272\\(73.9^{*}\\)\\(57.3^{*}\\)\\(50.4^{*}\\)\\(60.6^{*}\\)\\(76.6^{*}\\)\\(60.3^{*}\\)\\(51.7^{*}\\)\\(62.9^{*}\\)58.657.3
HICom (Ours)1.5B3268061.751.343.952.365.154.747.355.756.450.1
HICom (Ours)7B3268065.757.151.458.168.458.953.160.164.160.5
HICom (Ours)\\( ^§ \\)7B64132867.658.051.258.971.863.754.563.364.762.2
HICom (Ours)\\( ^§ \\)7B128262469.060.251.560.372.764.657.364.864.062.3
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.399, + 0.483, + 0.429 + ], + "angle": 0, + "content": "lated to the instruction, and does not include other rich types of instruction data (e.g., reasoning, counting, etc)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.431, + 0.483, + 0.656 + ], + "angle": 0, + "content": "Data Collection. To ensure the diversity and quality of our videos, we collect videos from public datasets Panda70M [5] and Ego4D [15]. Panda-70M contains various open-domain videos from YouTube, and Ego4D collects many high-resolution ego-centric human activity videos. Notably, we use the original untrimmed videos to cut clips ourselves since the provided split clips of both datasets are too short, greatly decreasing the content complexity. To balance the number of different types of videos, we pre-define 29 categories [10, 72] using natural language (e.g. A video about cooking activity.) and use InternVideo2 [54] to extract both the video features and the pre-defined sentence features. Based on the similarities between them, we select 1,500 videos for each category and then randomly select additional 10,000 videos from the others to ensure diversity." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.658, + 0.483, + 0.853 + ], + "angle": 0, + "content": "Video Processing. The untrimmed videos are too long for our pre-training, and also severely affect the quality of captions and annotations by existing SOTA models. Therefore, we use PySceneDetect to split each video into shorter ones with a much looser threshold than Panda-70M. As the split clips can sometimes be redundant with repeat semantics, we further extract keyframes for each video and only select the split clips containing the keyframes as our final results. Specifically, inspired by ShareGPT4Video [4], we calculate the CLIP score for densely sampled frames within a video, and select the less similar frames as keyframes. Finally, we further filter out the video clips shorter than 5 seconds and longer than 120 seconds." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Instruction-Followed Description Generation. For the processed video clips, we use Qwen2-VL-72B-Instruct [52] to systematically generate the instruction-followed descrip" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.399, + 0.907, + 0.595 + ], + "angle": 0, + "content": "tion with the CoT technique [55]. Specifically, we first generate the detailed caption of each input video, then take both the video content and caption as input to generate the three instruction-answer pairs for each video. Different from the normal instruction datasets, we require Qwen2-VL-72B-Instruct to meet the following two rules: 1) The instructions should refer to the specific visual information, such as the people or the objects in the video; 2) The instructions should lead to a descriptive response and the answers should describe the object mentioned in the instruction in detail. Finally, we filter the generated results by discarding answers containing invalid information, such as the phrase \"not provide\", or \"not describe\"." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.596, + 0.909, + 0.704 + ], + "angle": 0, + "content": "Dataset Statistics. Tab. 1 and Fig. 4 shows some statistics about our constructed dataset. We collect 248K video clips in total from Panda-70M and Ego4D, the average of the video lengths is 25.67 seconds and the distribution is shown in Fig. 4. With the help of Qwen2-VL-72B-Instruct, we generate 739K instruction-followed descriptions, providing sufficient data for our guidance pre-training." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.717, + 0.646, + 0.734 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.742, + 0.704, + 0.759 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Implementation Details. We use SigLIP [64] (so400m-patch14-384) as our vision encoder and text encoder, choose Qwen2.5 series [48] as our LLMs, and randomly initialize the compressor. The vision encoder keeps frozen at all stages, the LLM is frozen at the two pre-train stages, and is fine-tuned at the instruction tuning stage. We follow LLaVA-OneVision [20] to choose our training configurations, the global batch size is set to 512 at the alignment stage, 256 at the conditional pre-train and instruction tun" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8572" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.486, + 0.175 + ], + "angle": 0, + "content": "Table 3. Performance of our HICom and other SOTA methods on open-ended QA video benchmarks. All the compared models are 7B.†, *, and § keep the same meaning with Tab. 2. #F., #T. are short for # Frames, # Tokens, ANet is short for ActivityNet, LLaVA-NV, LLaVA-OV are short for LLaVA-Next-Video, and LLaVA-OneVision. The format we report ANet is Acc/Score." + }, + { + "type": "table", + "bbox": [ + 0.09, + 0.182, + 0.485, + 0.421 + ], + "angle": 0, + "content": "
Methods# F.# T.ANetVideoChatGPT Bench
CIDOCUTUCOAvg
Video-LLaVA [30]8204845.3/3.3------
VideoChat2 [24]16153649.1/3.33.022.883.512.662.812.98
Chat-Univi [17]6444845.8/3.22.892.913.462.892.812.99
LLaMA-VID [28]1fps2tps47.4/3.32.963.003.532.462.512.89
LongVLM [58]10030547.6/3.32.762.863.342.393.112.89
ST-LLM [35]1651250.9/3.33.233.053.742.932.813.15
PLLAVA [59]16230456.3/3.53.212.863.622.332.932.99
LLaVA-NV [71]32460853.5/3.23.393.293.922.63.123.26
LongVA [69]12818432-/2.83.053.093.772.443.643.20
VideoLLaMA2 [7]16115250.2/3.33.163.083.692.563.143.13
Tarsier [51]16230459.5/3.6------
SF-LLaVA [60]50368056.3/3.43.092.703.572.523.353.04
LLaVA-OV [20]32627256.6/3.6*3.45*3.00*3.71*2.68*3.14*3.20*
HICom(Ours)-1.5B3268053.0/3.53.092.673.402.412.982.91
HICom(Ours)-7B3268058.3/3.73.292.853.592.673.223.12
HICom(Ours)-7B$64132859.4/3.73.322.923.652.743.353.20
HICom(Ours)-7B$128262459.5/3.73.332.863.652.743.323.18
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.445, + 0.485, + 0.735 + ], + "angle": 0, + "content": "ing stage. We use the learning rate of 1e-3 at the alignment stage, 1e-5 at the instruction tuning stage. At our proposed conditional pre-training stage, we use 1e-3 for the condition injection sub-module and 1e-4 for other parameters in the compressing module. We train 1 epoch for all stages. More implementation details can be found in Appendix A. Datasets. We use LLaVA-558K [32] image-caption pairs for our alignment stage, the constructed 248K HiComPretrain instruction-followed descriptions for conditional pre-train stage. Inspired by [51, 72], we collect 2.68M video instruction data for instruction tuning, including 1.6M from LLaVA-Video [72], 292K from VideoChat2-IT [23] subset, 255K from M4-Instruct-Data [21], 11K from Charades [45], 114K from NTU RGB+D [44], 122K from TVQA [18], and 290K from MiT [38]. We evaluate our model on 5 video benchmarks, including 3 multiple-choice benchmarks VideoMME [10], MVBench [24], EgoSchema [37], and 2 open-ended benchmarks ActivityNet [3], VideoChatGPT Bench [36]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.743, + 0.233, + 0.758 + ], + "angle": 0, + "content": "4.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Multiple-choice benchmarks. In Tab. 2, we compare the performance of our HICom with different SOTA models on three multiple-choice video benchmarks. We sample 32 frames with 680 tokens to train all our models and sample different frames for inference. Our HICom-7B with 2624 tokens inference achieves the best performance on all three benchmarks. Compared to LLaVA-Video with 6272 tokens, our HICom with only 1328 tokens can obtain better performance, increasing the performance by \\(2.43\\%\\) average and" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.089, + 0.909, + 0.161 + ], + "angle": 0, + "content": "Table 4. The ablation study on our proposed HICom, \\(\\alpha_{(T,H,W)}\\) donates the downsampling ratio of each dimension, # Tok. donates the number of tokens input to LLMs. 8f means we sample 8 frames for training. The conditional compression combining both local and global achieves the best." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.171, + 0.905, + 0.357 + ], + "angle": 0, + "content": "
Methodsα(T,H,W) # Tok.VideoMME w/o sub.MV-BenchEgo-Schema
shortmidall
Unconditional Compression
avg pool(1,3,3)259239.334.933.035.744.8
8f, avg pool(1,3,3)64836.334.432.034.343.6
avg pool(4,3,3)64836.333.332.234.043.7
local(4,3,3)64836.734.432.034.443.7
globalNone3230.031.628.930.134.6
local+global(4,3,3)68037.134.832.334.744.1
Conditional Compression
local(4,3,3)64838.836.133.136.044.0
globalNone3230.830.329.830.335.5
local+global(4,3,3)68039.036.734.236.645.0
" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.377, + 0.707, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.377, + 0.898, + 0.474 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.524, + 0.479, + 0.705, + 0.577 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.479, + 0.898, + 0.577 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.586, + 0.909, + 0.645 + ], + "angle": 0, + "content": "Figure 5. The ablation study on different compressing ratios. The figures show the performance on VideoMME-Short (upper left), MVBench (upper right), EgoSchema (lower left), and the inference time of 7B LLM (lower right)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.909, + 0.903 + ], + "angle": 0, + "content": "saving \\(78.8\\%\\) tokens, HICom with 2624 tokens gains \\(3\\%\\) averagely when saving \\(58.2\\%\\) tokens, significantly lowering the computational burden. Notably, LLaVA-Video uses numerous additional image data, and it also unfreezes the vision encoder during training, which can both increase performance. It is also noteworthy that though our HICom performs a little worse than LLaVA-Video on VideoMME short videos (less than 2 minutes), we beat LLaVA-Video on both VideoMME medium (4-15 minutes) and long videos (30-60 minutes). We argue that conditional compression is more helpful for long videos. For short videos, the sampled frames are dense enough for unconditional compression, baselines with enough information would naturally perform powerfully. But for long videos, the sampled frames are more sparse with less redundant, our HICom can preserve" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "8573" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.09, + 0.486, + 0.119 + ], + "angle": 0, + "content": "Table 5. The ablation study on the conditional pre-training stage with the constructed HICom-248K dataset." + }, + { + "type": "table", + "bbox": [ + 0.113, + 0.129, + 0.46, + 0.234 + ], + "angle": 0, + "content": "
MethodsConditional \nPre-trainedVideoMMEw/o subs.MV-BenchEgo-Schema
shortmidlongall
avg poolX36.632.433.234.142.740.6
✓36.333.332.234.043.741.9
w/o inj.X39.433.733.135.442.841.64
✓37.134.832.334.744.142.4
w/ inj.X38.835.134.136.044.041.6
✓39.036.734.236.645.043.5
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.256, + 0.483, + 0.287 + ], + "angle": 0, + "content": "the maximum amount of instruct-relevant information while it is easily suppressed in LLaVA-Video." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.288, + 0.483, + 0.455 + ], + "angle": 0, + "content": "Open-ended benchmarks. We also compare the performance of different models on two open-ended video benchmarks, as shown in Tab. 3. We evaluate all our results via GPT-3.5-Turbo-0125. Our HICom also achieves comparable results on both benchmarks, as arriving at SOTA on ActivityNet and getting the second best on the VideoChatGPT Bench. Our HICom with 1328 tokens is on par with Tarsier with much more training data, and performs similarly to LLaVA-OneVision with much more training data and much more visual tokens, demonstrating the effectiveness of our compression." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.456, + 0.483, + 0.683 + ], + "angle": 0, + "content": "Generalization Ability on Video Length. Due to the design of the local-level compression, we can easily extend the lengths of sampled frames at the inference stage for our trained model with 32 frames, rather than re-training it. As we increase the number of sampled frames, the metrics of our HICom also improve across all these benchmarks. For example, the performance on short and medium videos of VideoMME grows fast from 32 frames to 128 frames at our high compressing ratio, as more useful frames are sampled. The performance on long videos of VideoMME without subtitles changes slightly, and we argue this may caused by the extremely long videos (30-60 minutes). Both 32 frames and 128 frames are too sparse for them. Our HICom can pay attention to the most relevant parts based on the sparsely sampled frames, thus changing slightly." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.696, + 0.255, + 0.71 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.718, + 0.483, + 0.795 + ], + "angle": 0, + "content": "To save time costs, we choose VideoChat2-IT [23] 896K video data as our instruction tuning data and Qwen2.5-0.5B [48] as our LLM to conduct ablation studies, and extract 32 frames for training unless otherwise specified. This requires less than 28 hours training on 8 A800 GPUs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Component Analysis. The key components of our HICom are local- and global- level compression. Tab. 4 shows the results of them in the status of both unconditional (i.e., without condition injection) and conditional compression (i.e., with condition injection). Note we use \\( V_{p}^{t,h,w} \\) as query of Eq. (1) in unconditional compression for direct injection. With the same number of input tokens, the un" + }, + { + "type": "table_caption", + "bbox": [ + 0.537, + 0.09, + 0.882, + 0.105 + ], + "angle": 0, + "content": "Table 6. The ablation study on different types of injection." + }, + { + "type": "table", + "bbox": [ + 0.537, + 0.116, + 0.882, + 0.23 + ], + "angle": 0, + "content": "
LocalGlobalVideoMME w/o sub.MV-BenchEgo-Schema
shortmidlongall
directcoarse39.036.734.236.645.043.5
directfine39.335.334.336.344.143.5
coarsecoarse39.834.835.036.544.142.7
coarsefine38.436.734.236.444.343.5
finecoarse39.134.734.236.044.242.7
finefine38.834.934.236.043.642.0
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.251, + 0.907, + 0.522 + ], + "angle": 0, + "content": "conditional local compression can achieve comparable performance with both spatial pooling (line 2) and temporal-spatial pooling (line 3). Due to the lack of temporal-spatial inductive bias and too few tokens, the global compression significantly harms the performance, and local+global performs only slightly better than local as the additional information of the global branch is limited without explicit guidance. Compared with unconditional compression, the local and global conditional compression increase by \\(0.8\\%\\) and \\(0.5\\%\\), respectively. The local+global achieves the best, increasing by \\(1.9\\%\\) on VideoMME and \\(1.3\\%\\) average on three benchmarks, also increasing by \\(0.63\\%\\) compared with local conditional compression. As the global branch can provide instruction-relevant information from a global sight, combining both is better. It is noteworthy that our conditional compression with 680 tokens can beat the average pooling with 2592 tokens on all benchmarks, reducing \\(73.8\\%\\) visual tokens and increasing by \\(0.47\\%\\) on the performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.523, + 0.909, + 0.78 + ], + "angle": 0, + "content": "Compressing Ratio. We explore the effectiveness of the compressing ratio on both unconditional and conditional compression in Fig. 5. Our unconditional compression performs similarly with average pooling on three benchmarks at different compressing ratios, and the conditional compression performs much better than them. It is noteworthy that the speed of performance decreasing of the conditional compression is also lower than the unconditional compression as the compressing ratio becomes higher. This shows instruction-relevant information remains with the guidance, demonstrating the superiority. In the lower right sub-figure of Fig. 5, the compressing ratio of (4,3,3) saves \\(57.68\\%\\) inference time on Qwen2.5-7B than (1,3,3) with only \\(2.06\\%\\) performance loss on Qwen2.5-0.5B. Higher compressing ratios (4,9,9) and (8,9,9) can only save a little more time with much higher performance loss. Therefore, we choose (4,3,3) as our final compressing ratio." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Conditional Pre-training Stage. To evaluate the effectiveness of our proposed conditional pre-training stage, we conduct experiments on whether to use this stage or not on both unconditional and conditional compression in Tab. 5. For average pooling and our unconditional compression, the conditional pre-training stage is also effective on MVBench and EgoSchema since more data is trained, but is hard to be effective on VideoMME. For the conditional compression," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.958 + ], + "angle": 0, + "content": "8574" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.123, + 0.089, + 0.878, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.364, + 0.908, + 0.395 + ], + "angle": 0, + "content": "Figure 6. The visualization of the attention map at both the local and global level in the situation of unconditional (w/o inj.) and conditional (w/ inj.) compression. We indicate the instruction-relevant parts with the red bounding box for easier reading." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.418, + 0.482, + 0.478 + ], + "angle": 0, + "content": "the conditional pre-training stage is effective on all benchmarks, increasing by \\(1.17\\%\\) averagely, demonstrating the effectiveness of not only the constructed data, but also the newly proposed stage." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.479, + 0.483, + 0.631 + ], + "angle": 0, + "content": "Types of Injection. We propose three types of instruction condition injection in Sec. 3.1, direct injection is designed for local compression only, while coarse and fine injection fit both. We conduct the ablation study on them in Tab. 6. Coarse injection performs better than fine injection on all three benchmarks, and both of them perform better than unconditional compression. We argue the reason may be that the pooled instruction feature is more in line with the CLIP training, and the attention of fine injection is harder to converge than the MLP of coarse injection." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.638, + 0.283, + 0.654 + ], + "angle": 0, + "content": "4.4. Qualitative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.659, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To qualitatively figure out how our HICom realizes the conditional compression, we visualize the attention map of the compression module in Fig. 6. It clearly shows where the model pays attention to based on the instructions. Since we conduct the local-level compression within the subregion of each group, the attention map of the local level is more evenly distributed throughout the video than the global level. Compared to unconditional compression, the attention map of conditional compression is more related to the instruction-relevant visual parts. As shown in Fig. 6, the attention maps at both levels successfully focus on the candle holders in the first example, and successfully on the Christmas tree, the apples, candles, and berries in the second example, which means the model will give more weights of these parts during compressing. We also notice that the global-level compression tends to forget some de" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.418, + 0.907, + 0.523 + ], + "angle": 0, + "content": "tailed visual information (e.g., the candle holders in the first and third frame of the first example, the Christmas tree in the second frame of the second example), while the local-level compression is better to capture these details due to group-limited attention. This shows that hybrid-level compression provides different sights, thus helping the model better understand the videos." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.542, + 0.634, + 0.558 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.907, + 0.903 + ], + "angle": 0, + "content": "In this paper, we address the challenge of achieving efficient video understanding in multi-modal large language models (MLLMs) through our proposed HICom. Our approach effectively balances the trade-off between computational costs and information retention by integrating user instructions into the compression process. This method not only preserves the temporal-spatial structure of video data but also emphasizes instruction-relevant content through a dual-level compression strategy. Furthermore, by introducing a conditional pre-training stage with proposed HICom-248K dataset, targeted pre-training enhances the efficacy of instruction-injected schema. Our experiments demonstrate the effectiveness, as HICom can maintain distinguished video understanding ability with much fewer tokens. HICom shows a good generalization ability to extend the number of frames during inference, but the fixed frame sampling strategy during training still limits the ability to understand long videos. In the future, we will try to sample frames based on fps and add more long videos for training to further increase the ability. We will also explore the potential of conditional compression on images, especially high-resolution images, making our HICom more powerful." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "8575" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.241, + 0.108 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.484, + 0.238 + ], + "angle": 0, + "content": "This work is supported by the National Key Research and Development Program of China (2022YFB3104700), the National Nature Science Foundation of China (62425114, 62121002, U23B2028, 62232006). We acknowledge the support of Alibaba Group, the GPU cluster built by MCC Lab of Information Science and Technology Institution, USTC, and USTC super-computing center for providing computational resources for this project." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.252, + 0.188, + 0.267 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.277, + 0.484, + 0.345 + ], + "angle": 0, + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. NeurIPS, 35: 23716-23736, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.347, + 0.484, + 0.403 + ], + "angle": 0, + "content": "[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.406, + 0.483, + 0.46 + ], + "angle": 0, + "content": "[3] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In CVPR, pages 961-970, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.462, + 0.484, + 0.53 + ], + "angle": 0, + "content": "[4] Lin Chen, Xilin Wei, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Bin Lin, Zhenyu Tang, et al. Sharegpt4video: Improving video understanding and generation with better captions. arXiv preprint arXiv:2406.04325, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.533, + 0.484, + 0.615 + ], + "angle": 0, + "content": "[5] Tsai-Shien Chen, Aliaksandr Siarohin, Willi Menapace, Ekaterina Deyneka, Hsiang-wei Chao, Byung Eun Jeon, Yuwei Fang, Hsin-Ying Lee, Jian Ren, Ming-Hsuan Yang, et al. Panda-70m: Captioning 70m videos with multiple cross-modality teachers. In CVPR, pages 13320-13331, 2024. 5, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.618, + 0.484, + 0.686 + ], + "angle": 0, + "content": "[6] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhang-wei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.688, + 0.484, + 0.757 + ], + "angle": 0, + "content": "[7] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024. 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.76, + 0.484, + 0.828 + ], + "angle": 0, + "content": "[8] Wenliang Dai, Junnan Li, DONGXU LI, Anthony Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale N Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. In NeurIPS, pages 49250-49267, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.831, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[9] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[10] Chaoyou Fu, Yuhan Dai, Yondong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen," + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.277, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 5, 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[11] Chaoyou Fu, Haojia Lin, Zuwei Long, Yunhang Shen, Meng Zhao, Yifan Zhang, Xiong Wang, Di Yin, Long Ma, Xiawu Zheng, et al. Vita: Towards open-source interactive omni multimodal llm. arXiv preprint arXiv:2408.05211, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.191, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[12] Zuan Gao, Yuxin Wang, Yadong Qu, Boqiang Zhang, Zixiao Wang, Jianjun Xu, and Hongtao Xie. Self-supervised pretraining with symmetric superimposition modeling for scene text recognition. arXiv preprint arXiv:2405.05841, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.905, + 0.3 + ], + "angle": 0, + "content": "[13] Jiannan Ge, Hongtao Xie, Pandeng Li, Lingxi Xie, Shaobo Min, and Yongdong Zhang. Towards discriminative feature generation for generalized zero-shot learning. IEEE Transactions on Multimedia, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.905, + 0.357 + ], + "angle": 0, + "content": "[14] Jiannan Ge, Lingxi Xie, Hongtao Xie, Pandeng Li, Xiaopeng Zhang, Yongdong Zhang, and Qi Tian. Alignzeg: Mitigating objective misalignment for zero-shot semantic segmentation. In ECCV, pages 142-161. Springer, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.359, + 0.905, + 0.426 + ], + "angle": 0, + "content": "[15] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In CVPR, pages 18995-19012, 2022. 5, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.905, + 0.495 + ], + "angle": 0, + "content": "[16] Bo He, Hengduo Li, Young Kyun Jang, Menglin Jia, Xuefei Cao, Ashish Shah, Abhinav Shrivastava, and Ser-Nam Lim. Ma-lmm: Memory-augmented large multimodal model for long-term video understanding. In CVPR, pages 13504-13514, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.497, + 0.905, + 0.552 + ], + "angle": 0, + "content": "[17] Peng Jin, Ryuichi Takanobu, Wancai Zhang, Xiaochun Cao, and Li Yuan. Chat-univi: Unified visual representation empowers large language models with image and video understanding. In CVPR, pages 13700-13710, 2024. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.554, + 0.904, + 0.593 + ], + "angle": 0, + "content": "[18] Jie Lei, Licheng Yu, Mohit Bansal, and Tamara L Berg. Tvqa: Localized, compositional video question answering. arXiv preprint arXiv:1809.01696, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.595, + 0.905, + 0.649 + ], + "angle": 0, + "content": "[19] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.651, + 0.905, + 0.705 + ], + "angle": 0, + "content": "[20] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.707, + 0.905, + 0.761 + ], + "angle": 0, + "content": "[21] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava next-interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.762, + 0.905, + 0.817 + ], + "angle": 0, + "content": "[22] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, pages 19730–19742. PMLR, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[23] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[24] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8576" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "Mvbench: A comprehensive multi-modal video understanding benchmark. In CVPR, pages 22195-22206, 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.178 + ], + "angle": 0, + "content": "[25] Pandeng Li, Chen-Wei Xie, Hongtao Xie, Liming Zhao, Lei Zhang, Yun Zheng, Deli Zhao, and Yongdong Zhang. Momentdiff: Generative video moment retrieval from random to real. NeurIPS, 36:65948–65966, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.18, + 0.482, + 0.235 + ], + "angle": 0, + "content": "[26] Pandeng Li, Chen-Wei Xie, Liming Zhao, Hongtao Xie, Jiannan Ge, Yun Zheng, Deli Zhao, and Yongdong Zhang. Progressive spatio-temporal prototype matching for text-video retrieval. In ICCV, pages 4100–4110, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.238, + 0.483, + 0.292 + ], + "angle": 0, + "content": "[27] Wentong Li, Yuqian Yuan, Jian Liu, Dongqi Tang, Song Wang, Jianke Zhu, and Lei Zhang. Tokenpacker: Efficient visual projector for multimodal llm. arXiv preprint arXiv:2407.02392, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.295, + 0.482, + 0.337 + ], + "angle": 0, + "content": "[28] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In ECCV, pages 323–340. Springer, 2024. 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.339, + 0.482, + 0.395 + ], + "angle": 0, + "content": "[29] Jianxin Liang, Xiaojun Meng, Yueqian Wang, Chang Liu, Qun Liu, and Dongyan Zhao. End-to-end video question answering with frame scoring mechanisms and adaptive sampling. arXiv preprint arXiv:2407.15047, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.397, + 0.482, + 0.451 + ], + "angle": 0, + "content": "[30] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.454, + 0.482, + 0.481 + ], + "angle": 0, + "content": "[31] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 36, 2023. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.482, + 0.525 + ], + "angle": 0, + "content": "[32] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In CVPR, pages 26296-26306, 2024. 1, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.528, + 0.482, + 0.57 + ], + "angle": 0, + "content": "[33] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024. 2, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.572, + 0.482, + 0.625 + ], + "angle": 0, + "content": "[34] Ruyang Liu, Chen Li, Yixiao Ge, Thomas H Li, Ying Shan, and Ge Li. Bt-adapter: Video conversation is feasible without video instruction tuning. In CVPR, pages 13658-13667, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.629, + 0.482, + 0.684 + ], + "angle": 0, + "content": "[35] Ruyang Liu, Chen Li, Haoran Tang, Yixiao Ge, Ying Shan, and Ge Li. St-llm: Large language models are effective temporal learners. In European Conference on Computer Vision, pages 1–18. Springer, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.687, + 0.482, + 0.741 + ], + "angle": 0, + "content": "[36] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. In ACL, 2024. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.744, + 0.482, + 0.799 + ], + "angle": 0, + "content": "[37] Karttikeya Mangalam, Raiymbek Akshulakov, and Jitendra Malik. Egoschema: A diagnostic benchmark for very long-form video language understanding. NeurIPS, 36:46212-46244, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.802, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[38] Mathew Monfort, Alex Andonian, Bolei Zhou, Kandan Ramakrishnan, Sarah Adel Bargal, Tom Yan, Lisa Brown, Quanfu Fan, Dan Gutfreund, Carl Vondrick, et al. Moments in time dataset: one million videos for event understanding. IEEE TPAMI, 42(2):502-508, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[39] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, pages 4195-4205, 2023. 4" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[40] Tianyuan Qu, Longxiang Tang, Bohao Peng, Senqiao Yang, Bei Yu, and Jiaya Jia. Does your vision-language model get lost in the long video sampling dilemma?, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.288 + ], + "angle": 0, + "content": "[42] Machel Reid, Nikolay Savinov, Denis Teplyashin, Dmitry Lepikhin, Timothy Lillicrap, Jean-baptiste Alayrac, Radu Soricut, Angeliki Lazaridou, Orhan First, Julian Schrittwieser, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.905, + 0.345 + ], + "angle": 0, + "content": "[43] Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. In CVPR, pages 14313-14323, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.348, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[44] Amir Shahroudy, Jun Liu, Tian-Tsong Ng, and Gang Wang. Ntu rgb+d: A large scale dataset for 3d human activity analysis. In CVPR, pages 1010-1019, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.446 + ], + "angle": 0, + "content": "[45] Gunnar A. Sigurdsson, Gúl Varol, Xiaolong Wang, Ivan Laptev, Ali Farhadi, and Abhinav Gupta. Hollywood in homes: Crowdsourcing data collection for activity understanding. ArXiv e-prints, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.448, + 0.905, + 0.516 + ], + "angle": 0, + "content": "[46] Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Haozhe Chi, Xun Guo, Tian Ye, Yanting Zhang, et al. Moviechat: From dense token to sparse memory for long video understanding. In CVPR, pages 18221-18232, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.587 + ], + "angle": 0, + "content": "[47] Longxiang Tang, Zhuotao Tian, Kai Li, Chunming He, Hantao Zhou, Hengshuang Zhao, Xiu Li, and Jiaya Jia. Mind the interference: Retaining pre-trained knowledge in parameter efficient continual learning of vision-language models. In ECCV, pages 346-365. Springer, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.589, + 0.904, + 0.616 + ], + "angle": 0, + "content": "[48] Qwen Team. Qwen2.5: A party of foundation models, 2024. 5, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.619, + 0.905, + 0.686 + ], + "angle": 0, + "content": "[49] Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, Sai Charitha Akula, Jihan Yang, Shusheng Yang, Adithya Iyer, Xichen Pan, et al. Cambrian1: A fully open, vision-centric exploration of multimodal llms. arXiv preprint arXiv:2406.16860, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.905, + 0.744 + ], + "angle": 0, + "content": "[50] Haibo Wang, Chenghang Lai, Yixuan Sun, and Weifeng Ge. Weakly supervised gaussian contrastive grounding with large multimodal models for video question answering. arXiv preprint arXiv:2401.10711, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[51] Jiawei Wang, Liping Yuan, Yuchen Zhang, and Haomiao Sun. Tarsier: Recipes for training and evaluating large video description models. arXiv preprint arXiv:2407.00634, 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[52] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1, 5, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[53] Yizhou Wang, Ruiyi Zhang, Haoliang Wang, Uttaran Bhattacharya, Yun Fu, and Gang Wu. Vaquita: Enhancing align" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8577" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.119 + ], + "angle": 0, + "content": "ment in llm-assisted video understanding. arXiv preprint arXiv:2312.02310, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.484, + 0.189 + ], + "angle": 0, + "content": "[54] Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Guo Chen, Baoqi Pei, Rongkun Zheng, Jilan Xu, Zun Wang, et al. Intervideo2: Scaling video foundation models for multimodal video understanding. arXiv preprint arXiv:2403.15377, 2024. 5, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.191, + 0.483, + 0.245 + ], + "angle": 0, + "content": "[55] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. NeurIPS, 35:24824-24837, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.247, + 0.483, + 0.314 + ], + "angle": 0, + "content": "[56] Yujie Wei, Shiwei Zhang, Zhiwu Qing, Hangjie Yuan, Zhiheng Liu, Yu Liu, Yingya Zhang, Jingren Zhou, and Hongming Shan. Dreamvideo: Composing your dream videos with customized subject and motion. In CVPR, pages 6537-6549, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.483, + 0.384 + ], + "angle": 0, + "content": "[57] Yujie Wei, Shiwei Zhang, Hangjie Yuan, Xiang Wang, Haonan Qiu, Rui Zhao, Yutong Feng, Feng Liu, Zhizhong Huang, Jiaxin Ye, et al. Dreamvideo-2: Zero-shot subject-driven video customization with precise motion control. arXiv preprint arXiv:2410.13830, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.385, + 0.483, + 0.44 + ], + "angle": 0, + "content": "[58] Yuetian Weng, Mingfei Han, Haoyu He, Xiaojun Chang, and Bohan Zhuang. Longvlm: Efficient long video understanding via large language models. In ECCV, pages 453-470. Springer, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.483, + 0.496 + ], + "angle": 0, + "content": "[59] Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. Pllava: Parameter-free llava extension from images to videos for video dense captioning. arXiv preprint arXiv:2404.16994, 2024. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.498, + 0.483, + 0.565 + ], + "angle": 0, + "content": "[60] Mingze Xu, Mingfei Gao, Zhe Gan, Hong-You Chen, Zhengfeng Lai, Haiming Gang, Kai Kang, and Afshin Dehghan. Slowfast-llava: A strong training-free baseline for video large language models. arXiv preprint arXiv:2407.15841, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.483, + 0.622 + ], + "angle": 0, + "content": "[61] Zhipei Xu, Xuanyu Zhang, Runyi Li, Zecheng Tang, Qing Huang, and Jian Zhang. Fakeshield: Explainable image forgery detection and localization via multi-modal large language models. arXiv preprint arXiv:2410.02761, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.623, + 0.483, + 0.69 + ], + "angle": 0, + "content": "[62] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.483, + 0.734 + ], + "angle": 0, + "content": "[63] Shoubin Yu, Jaemin Cho, Prateek Yadav, and Mohit Bansal. Self-chained image-language model for video localization and question answering. NeurIPS, 36, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.483, + 0.775 + ], + "angle": 0, + "content": "[64] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In ICCV, pages 11975-11986, 2023. 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.777, + 0.483, + 0.831 + ], + "angle": 0, + "content": "[65] Boqiang Zhang, Hongtao Xie, Zuan Gao, and Yuxin Wang. Choose what you need: Disentangled representation learning for scene text recognition removal and editing. In CVPR, pages 28358-28368, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[66] Boqiang Zhang, Kehan Li, Zesen Cheng, Zhiqiang Hu, Yuqian Yuan, Guanzheng Chen, Sicong Leng, Yuming Jiang, Hang Zhang, Xin Li, et al. Videollama 3: Frontier multimodal foundation models for image and video understanding. arXiv preprint arXiv:2501.13106, 2025. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[67] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[68] Haoji Zhang, Yiqin Wang, Yansong Tang, Yong Liu, Jiashi Feng, Jifeng Dai, and Xiaojie Jin. Flash-vstream: Memory-based real-time understanding for long video streams. arXiv preprint arXiv:2406.08085, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[69] Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision. arXiv preprint arXiv:2406.16852, 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[70] Xuanyu Zhang, Runyi Li, Jiwen Yu, Youmin Xu, Weiqi Li, and Jian Zhang. Editguard: Versatile image watermarking for tamper localization and copyright protection. In CVPR, pages 11964-11974, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.905, + 0.374 + ], + "angle": 0, + "content": "[71] Yuanhan Zhang, Bo Li, haotian Liu, Yong jae Lee, Liangke Gui, Di Fu, Jiashi Feng, Ziwei Liu, and Chunyuan Li. Llavanext: A strong zero-shot video understanding model, 2024. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.376, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[72] Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, and Chunyuan Li. Video instruction tuning with synthetic data. arXiv preprint arXiv:2410.02713, 2024. 2, 5, 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.433, + 0.905, + 0.5 + ], + "angle": 0, + "content": "[73] Junjie Zhou, Yan Shu, Bo Zhao, Boya Wu, Shitao Xiao, Xi Yang, Yongping Xiong, Bo Zhang, Tiejun Huang, and Zheng Liu. Mlvu: A comprehensive benchmark for multi-task long video understanding. arXiv preprint arXiv:2406.04264, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.503, + 0.905, + 0.558 + ], + "angle": 0, + "content": "[74] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.558 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "8578" + } + ] +] \ No newline at end of file diff --git a/2025/Hybrid-Level Instruction Injection for Video Token Compression in Multi-modal Large Language Models/a0ce533d-bc87-47ce-aab6-707ed1ae09bc_origin.pdf b/2025/Hybrid-Level Instruction Injection for Video Token Compression in Multi-modal Large Language Models/a0ce533d-bc87-47ce-aab6-707ed1ae09bc_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..30e98c54a17642deb22086752e382e9c8fab0d04 --- /dev/null +++ b/2025/Hybrid-Level Instruction Injection for Video Token Compression in Multi-modal Large Language Models/a0ce533d-bc87-47ce-aab6-707ed1ae09bc_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d97329b374dcd9e389cecb975e142aca07a90dd201af1f0b6bb76676ed350e6e +size 1522667 diff --git a/2025/Hybrid-Level Instruction Injection for Video Token Compression in Multi-modal Large Language Models/full.md b/2025/Hybrid-Level Instruction Injection for Video Token Compression in Multi-modal Large Language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..240cdaa1cf2f3f6e57050308d78857b6fac6c52e --- /dev/null +++ b/2025/Hybrid-Level Instruction Injection for Video Token Compression in Multi-modal Large Language Models/full.md @@ -0,0 +1,282 @@ +# Hybrid-Level Instruction Injection for Video Token Compression in Multi-modal Large Language Models + +Zhihang Liu1*, Chen-Wei Xie2, Pandeng Li1,2†, Liming Zhao2, Longxiang Tang3, Yun Zheng2, Chuanbin Liu1, Hongtao Xie1 +1 University of Science and Technology of China +2 Tongyi Lab, Alibaba Group 3 Tsinghua University +{liuzhihang, lpd}@mail.ustc.edu.cn, hxie@ustc.edu.cn + +# Abstract + +Recent Multi-modal Large Language Models (MLLMs) have been challenged by the computational overhead resulting from massive video frames, often alleviated through compression strategies. However, the visual content is not equally contributed to user instructions, existing strategies (e.g., average pool) inevitably lead to the loss of potentially useful information. To tackle this, we propose the Hybrid-level Instruction Injection Strategy for Conditional Token Compression in MLLMs (HICom), utilizing the instruction as a condition to guide the compression from both local and global levels. This encourages the compression to retain the maximum amount of user-focused information while reducing visual tokens to minimize computational burden. Specifically, the instruction condition is injected into the grouped visual tokens at the local level and the learnable tokens at the global level, and we conduct the attention mechanism to complete the conditional compression. From the hybrid-level compression, the instruction-relevant visual parts are highlighted while the temporal-spatial structure is also preserved for easier understanding of LLMs. To further unleash the potential of HICom, we introduce a new conditional pre-training stage with our proposed dataset HICom248K. Experiments show that our HICom can obtain distinguished video understanding ability with fewer tokens, increasing the performance by $2.43\%$ average on three multiple-choice QA benchmarks and saving $78.8\%$ tokens compared with the SOTA method. The code is available at https://github.com/1ntzm/HICom. + +# 1. Introduction + +Multi-modal Large Language Models [2, 14, 22, 31, 32, 74] (MLLMs) have gained significant improvements in multi- + +![](images/5ef07a75991de3a13c4c1971b74d4b8fe3231e1bba5beccf9cc7f839b45604ac.jpg) +Figure 1. An example of the video understanding task, and the comparison between the unconditional compression and our proposed conditional compression with hybrid-level instruction injection. We inject instruction at both local and global levels, guiding the compression to retain the maximum amount of user-focused information and minimize the computational burden. + +modal understanding by integrating visual information into the LLMs, beating various expert methods on downstream tasks [12, 13, 25, 26, 47, 61, 65, 70]. While initially focused on image understanding, more research [30, 34, 36, 40, 56, 57] has shifted towards challenging video tasks. Most MLLMs handle videos by treating them as sequences of images, sampling frames, and concatenating visual tokens from these frames [6, 20, 42, 52, 66]. + +Compared to images, videos comprise multiple frames, resulting in more visual tokens and thus higher computational costs. To make the computation affordable, early methods [23, 30, 34] often employ sparse temporal sampling strategies, leading to significant loss of temporal information. Consequently, current MLLMs focus on compression + +ing visual tokens to achieve a trade-off between computational costs and video understanding ability. Mainstream MLLMs [20, 59, 71] use a spatial pooling strategy within each frame, leveraging capabilities gained from image data. Some approaches [19, 43, 67] use Q-Former [22] or Resampler [1] to compress the spatial-temporal information into a fixed number of visual tokens. Other methods employ convolution [7], clustering [17], or memory mechanisms [16, 46, 68] to compress visual tokens in both spatial and temporal dimensions. However, these compression methods are unconditional, lacking explicit guidance, as illustrated in Fig. 1 (b), making it challenging to achieve high compression ratios with minimal information loss. Consequently, user-relevant visual content may be overlooked during such unconditional compression. + +Therefore, we focus on introducing the instruction information as a condition to perform targeted compression, as it is rarely explored in previous work. As shown in Fig. 1 (c), conditional compression uses explicit instructional guidance to select user-focused information for retention, thereby enhancing efficiency. To thoroughly explore the conditional compression, we propose Hybrid-level Instruction Injection Strategy for Conditional Token Compression in MLLMs (HICom). Observing how humans process complex visual information, they usually adopt a coarse-to-fine approach, first considering the relative parts in the global coarse impression and then seeking detailed information locally in the video. Motivated by this human perception process, HICom conducts the compression at both local and global levels, retaining the maximum amount of instruction-relevant information while preserving the temporal-spatial structure with fewer tokens, thus achieving a better balance between the computational burden and video comprehension. + +Specifically, at the local level, we divide the sampled frame features into several groups and compress the tokens of each group into one token based on the attention with the injected instruction condition. Different from Q-Former [22] and Resampler [1], the grouped local attention can explicitly maintain the spatial-temporal structural characteristics of the video while highlighting the relevant parts within each group. At the global level, we inject the instruction condition into a small number of learnable tokens and then conduct the attention with the flattened frame features without grouping. The global compression focuses more on searching instruction-relevant information in the whole video and thus can be expressed by a small number of tokens as an auxiliary. To further unleash the potential of HICom, we introduce a new conditional pre-training stage between the alignment and instruction tuning and construct a new dataset HICom-248K of 248K video clips with high-quality instruction-followed descriptions. The proposed new stage continues to pre-train the parameters of + +condition injection, further improving the performance. Experiments show that our HICom can achieve state-of-the-art performance on 5 benchmarks with significantly fewer tokens. (e.g., increasing by $2.43\%$ average on three multiple-choice QA benchmarks and saving $78.8\%$ tokens compared with LLaVA-Video-7B [72]). + +Our main contributions are listed as follows: + +- We propose a conditional compression method HICom for MLLMs, achieving effective video understanding while significantly reducing the computational burden. +- We conduct hybrid-level instruction injection to achieve the conditional compression, and introduce a new conditional pre-training stage with our constructed HICom248K dataset to further unleash the potential of the conditional compressing. +- Experiments on various benchmarks show the proposed HICom can achieve better performance with fewer visual tokens, providing valuable insights for MLLMs. + +# 2. Related Work + +Multi-modal large language models. Based on the powerful capabilities of LLMs among various linguistic tasks, many works try to extend the understanding abilities to the computer vision area. Flamingo [1] and BLIP series [8, 22] successfully explore the usage of Resampler and Q-Former to bridge the visual tokens to language models. LLaVA series [31-33] use a simple MLP as the visual projector to translate visual tokens to the LLM's embedding space, achieving huge success. Recently, more researchers have focused on extending image tasks to video tasks. The biggest challenge for video tasks is to design an effective method to achieve the trade-off between computational costs and video understanding ability. Early methods [23, 30, 34] sparsely sample frames, losing too much temporal information. Mainstream methods [20, 59, 71] use a simple spatial pooling strategy to reduce the number of visual tokens and get a huge success. [19, 67] employ Q-Former to compress the spatial-temporal tokens into a fixed number. Chat-Univi [17] uses DPC-KNN algorithm to cluster dynamic visual tokens. VideoLLaMA2 [7] utilizes convolution both in temporal and spatial dimensions to down-sample tokens. Flash-VStream [68] introduces a learnable memory mechanism to compress online video streaming. All of them try their best to observe more visual information. However, their unconditional compression without explicit guidance inherently leads to the loss of instruction-relevant information. Different from them, we conduct conditional compression by injecting instruction at hybrid levels, only emphasizing the visual parts related to the instruction and allowing the irrelevant parts to be discarded, thus getting a better representation for each question. + +Text-based visual representation for video LLMs. There are some methods using instruction information during the + +![](images/3cfbab6f830d9d93511879a312410f719e193a7a61311e5cc174e7387ec8058e.jpg) +Figure 2. The framework of our proposed HICom. We propose the hybrid-level instruction injection to conditionally compress video tokens in MLLMs. We extract instruction-relevant information within each grouped sub-region at the local level, and extract it to a fixed number of tokens at the global level. The instruction condition is injected into the attention process to guide the compression. + +![](images/a9bf7acc60fc598ee70d1d5952a758fb3a4ed0b08f9a9405fb91c49a3381d944.jpg) + +visual encoding. [53, 63] sample the frames based on the CLIP [41] similarity or a localization model using the instruction, [29, 50] make the frame selection strategy trainable. However, these methods only focus on how to select the correct frames in a rough manner, failing to guide the token compression with the key instruction information. LLaMA-VID [28] follows a similar instruction-guided idea, as it introduces the instruction to interact with visual features. However, it roughly simplifies the interacted but uncompressed result into a single token only as auxiliary information for each frame, destroying the spatial structure and failing to discuss the guiding role of the instruction for conditional compression. Therefore, it is only designed for long videos and the performance is limited. To fully explore conditional compression, we propose to inject the instruction at hybrid levels, exploring a better way for conditional compression to alleviate the information loss while maintaining the performance. + +# 3. Method + +# 3.1. Instruction-Guided Conditional Compression + +Fig. 2 shows the framework of our proposed HICom, which focuses on the design between the vision encoder and the LLM to conduct the conditional compression. We inject the instruction condition into the compression process at both local and global levels for better information extraction with spatial-temporal structure maintained. + +Local-Level Compression. Previous works have demon- + +strated the importance of the spatial inductive bias when processing visual representations for image MLMs [27, 49]. Inspired by them, we divide the tokens into several groups and perform conditional compression within each group to preserve the temporal-spatial structure of video features. Specifically, given the encoded video frame features $\mathbf{V} \in \mathbb{R}^{T \times H \times W \times D}$ , and the downsampling ratio $(\alpha_{T}, \alpha_{H}, \alpha_{W})$ , we divide $\mathbf{V}$ into $N_{T} \cdot N_{H} \cdot N_{W}$ groups, where $N_{i} = \lceil i / \alpha_{i} \rceil$ , $i \in \{T, H, W\}$ . Therefore, there are $\alpha_{T} \cdot \alpha_{H} \cdot \alpha_{W}$ tokens in each group, which can be formulated as $\mathbf{V}^{t,h,w} \in \mathbb{R}^{\alpha_{T} \times \alpha_{H} \times \alpha_{W} \times D}$ , where $0 \leq \{t,h,w\} \leq \{N_{T}, N_{H}, N_{W}\}$ . We first pool $\mathbf{V}^{t,h,w}$ into one token $\mathbf{V}_{p}^{t,h,w} \in \mathbb{R}^{1 \times D}$ and inject the instruction condition $C$ (i.e., the text feature of instructions) into $\mathbf{V}_{p}^{t,h,w}$ . Then the local attention within each group can be calculated by: + +$$ +\boldsymbol {Z} _ {l} ^ {t, h, w} = \operatorname {A t t n} \left[ \operatorname {I n j} \left(\boldsymbol {V} _ {p} ^ {t, h, w}, \boldsymbol {C}\right), \boldsymbol {V} ^ {t, h, w}, \boldsymbol {V} ^ {t, h, w} \right], \tag {1} +$$ + +where $\mathrm{Inj}(\cdot)$ represents the condition injection, $\mathrm{Attn}(\cdot)$ donates the attention mechanism [9] and the inputs of $\mathrm{Attn}(\cdot)$ are query, key, and value, respectively. Therefore, the visual features within each group are compressed into only one token, preserving the temporal-spatial structure while conditionally highlighting the instruction-relevant parts. Finally, we concatenate $Z_{l}^{t,h,w}$ as the local compressed results $Z_{l}\in \mathbb{R}^{N_{T}\times N_{H}\times N_{W}\times D}$ , and use an MLP layer to project $Z_{l}$ into the LLM's embedding space. + +Global-Level Compression. Though the temporal-spatial structure can be maintained at the local level, the attention + +![](images/1bcb442224e098a92c0c13676ac3877e3cc4d7fc04eb3f20b2e55852e8c797ca.jpg) +Figure 3. We introduce a new guidance pre-training stage and implement three-stage training for conditional compression. + +is forced to focus on small sub-regions. However, each sub-region may contain information that is not equally valid for the question, and some sub-regions are even quite irrelevant. Therefore, we also implement conditional compression at the global level to highlight the most relevant parts within the whole video. Specifically, we initialize a small set of learnable tokens $\pmb{L} \in \mathbb{R}^{N_L \times D}$ , where $N_{L}$ donates the number of tokens, and then inject the instruction condition into them. Instead of grouping the video frame features, we apply 3D position embedding $\mathrm{Pos}(\cdot)$ for them and directly flatten them at global-level compression. Then the global compressed tokens $Z_{g} \in \mathbb{R}^{N_{L} \times D}$ can be calculated by: + +$$ +\boldsymbol {Z} _ {g} = \operatorname {A t t n} \left[ \operatorname {I n j} (\boldsymbol {L}, \boldsymbol {C}), \boldsymbol {V} + \operatorname {P o s} (\boldsymbol {V}), \boldsymbol {V} \right]. \tag {2} +$$ + +An MLP layer is also utilized to project $Z_{g}$ into the LLM's embedding space. And we finally concatenate $Z_{l}$ and $Z_{g}$ together for further understanding of LLM. + +Instruction Condition Injection. In order to fulfill the guiding role of the instruction condition, we explore three different types of injection modules, and define them direct injection, coarse injection, and fine injection, respectively. Note the text encoder can obtain both pooled tokens $C_p \in \mathbb{R}^{1 \times D}$ (i.e., global text embedding) and fine-grained tokens $C_f \in \mathbb{R}^{L \times D}$ (i.e., token-level text embedding). Given the pooled token $C_p$ , the direct injection uses an MLP to translate the single token into the injected output without any interaction with $V_p^{t,h,w}$ or $L$ . We employ coarse injection by the adaptive layer norm [39]. Specifically, we use an MLP to regress the scale and shift from $C_p$ , then add to the visual input $V_p^{t,h,w}$ or learnable tokens $L$ , which we represent them as $A$ , after the layer norm. The process can be formulated as: + +$$ +\operatorname {I n j} (\boldsymbol {A}, \boldsymbol {C}) = \operatorname {L N} (\boldsymbol {A}) \cdot \operatorname {s c a l e} \left(\boldsymbol {C} _ {p}\right) + \operatorname {s h i f t} \left(\boldsymbol {C} _ {p}\right), \tag {3} +$$ + +Table 1. The statistical information of our constructed HICom-248K dataset. + +![](images/4a1150e8b5d88adcbbdf6193ef8a0c26196888886a856966727d8ad86691c31e.jpg) +Figure 4. The visualization of data source (left) and video length (right) of our constructed HICom-248K dataset. + +where $\mathrm{LN}(\cdot)$ is the layer norm. Given the fine-grained tokens $C_f$ as condition $C$ , we conduct the fine injection via the cross attention, which can be formulated as: + +$$ +\operatorname {I n j} (\boldsymbol {A}, \boldsymbol {C}) = \operatorname {A t t n} \left(\operatorname {L N} (\boldsymbol {A}), \boldsymbol {C} _ {f}, \boldsymbol {C} _ {f}\right). \tag {4} +$$ + +As the direct injection directly translates the condition to the query of the attention in Eq. (1) and Eq. (2), the token length limits that it is only suitable for local compression as the local branch compresses each group into only one token. On the contrary, the coarse and fine injections are more flexible to fit various situations as the condition is added into $A$ . We conduct the experiments on the three types of injection and finally choose direct injection for local compression and coarse injection for global compression. + +Conditional Pre-training Stage. The existing mainstream methods follow the pipeline of alignment first and then instruction tuning [7, 31, 32, 68], and previous text-based methods [28, 53] also follow this pipeline. However, the image-caption pairs cannot provide valid instruction information in the alignment stage, direct instruction tuning with modules that are not adequately aligned will confuse the guiding process, thus harming the performance. Therefore, we propose a new conditional pre-training stage between the alignment and the instruction tuning to make the training process easier, as shown in Fig. 3. In this stage, we use our constructed instruction-followed descriptions to pre-train the compression module with condition injection, fulfilling the guiding role of the instruction. + +# 3.2. Dataset Construction + +To pre-train the condition injection in the compression module, we construct a new instruction-followed description dataset HICom-248K. Different from common instruction tuning datasets, HICom-248K focuses on providing only one data type, i.e., the description of the visual content re + +Table 2. Performance comparison between our HICom and other SOTA methods on multiple-choice QA video benchmarks. † means we use the result reproduced by VideoLLaMA2 [7], * indicates we reproduce the results ourselves using the official checkpoint and inference code provided by authors. § donates we inference with a new length of frames trained by sampling 32 frames. + +
MethodsLLM Size# Frames# TokensVideoMME w/o sub.VideoMME w/ sub.MV-BenchEgo-Schema
shortmidlongallshortmidlongall
Video-LLaVA [30]7B8204845.338.036.239.946.140.738.141.643.0-
VideoChat2-Mistral [24]7B16153648.336.335.039.552.839.439.243.860.4-
LLaMA-VID [28]7B1fps2tps---\(25.9^{\dagger}\)----\(41.9^{\dagger}\)\(38.5^{\dagger}\)
Chat-Univi-1.5 [17]7B6444845.740.335.840.651.244.641.845.945.9-
PLLAVA [59]7B162304--------46.6-
LLaVA-Next-Video [71]7B324608\(49.9^{*}\)\(41.4^{*}\)\(37.0^{*}\)\(42.8^{*}\)----\(46.5^{\dagger}\)\(43.9^{\dagger}\)
LongVA [69]7B1281843261.150.446.252.661.653.647.654.3--
VideoLLaMA2 [7]7B16115256.045.442.147.9----54.651.7
Tarsier [51]7B162304--------62.656.0
VITA [11]\(8\times 7\text{B}\)16409665.952.948.655.870.456.250.959.2--
LLaVA-OneVision [20]7B326272\(70.1^{*}\)\(56.4^{*}\)\(48.9^{*}\)\(58.5^{*}\)\(75.8^{*}\)\(58.4^{*}\)\(51.6^{*}\)\(61.9^{*}\)56.760.1
LLaVA-Video [72]7B326272\(73.9^{*}\)\(57.3^{*}\)\(50.4^{*}\)\(60.6^{*}\)\(76.6^{*}\)\(60.3^{*}\)\(51.7^{*}\)\(62.9^{*}\)58.657.3
HICom (Ours)1.5B3268061.751.343.952.365.154.747.355.756.450.1
HICom (Ours)7B3268065.757.151.458.168.458.953.160.164.160.5
HICom (Ours)\( ^§ \)7B64132867.658.051.258.971.863.754.563.364.762.2
HICom (Ours)\( ^§ \)7B128262469.060.251.560.372.764.657.364.864.062.3
+ +lated to the instruction, and does not include other rich types of instruction data (e.g., reasoning, counting, etc). + +Data Collection. To ensure the diversity and quality of our videos, we collect videos from public datasets Panda70M [5] and Ego4D [15]. Panda-70M contains various open-domain videos from YouTube, and Ego4D collects many high-resolution ego-centric human activity videos. Notably, we use the original untrimmed videos to cut clips ourselves since the provided split clips of both datasets are too short, greatly decreasing the content complexity. To balance the number of different types of videos, we pre-define 29 categories [10, 72] using natural language (e.g. A video about cooking activity.) and use InternVideo2 [54] to extract both the video features and the pre-defined sentence features. Based on the similarities between them, we select 1,500 videos for each category and then randomly select additional 10,000 videos from the others to ensure diversity. + +Video Processing. The untrimmed videos are too long for our pre-training, and also severely affect the quality of captions and annotations by existing SOTA models. Therefore, we use PySceneDetect to split each video into shorter ones with a much looser threshold than Panda-70M. As the split clips can sometimes be redundant with repeat semantics, we further extract keyframes for each video and only select the split clips containing the keyframes as our final results. Specifically, inspired by ShareGPT4Video [4], we calculate the CLIP score for densely sampled frames within a video, and select the less similar frames as keyframes. Finally, we further filter out the video clips shorter than 5 seconds and longer than 120 seconds. + +Instruction-Followed Description Generation. For the processed video clips, we use Qwen2-VL-72B-Instruct [52] to systematically generate the instruction-followed descrip + +tion with the CoT technique [55]. Specifically, we first generate the detailed caption of each input video, then take both the video content and caption as input to generate the three instruction-answer pairs for each video. Different from the normal instruction datasets, we require Qwen2-VL-72B-Instruct to meet the following two rules: 1) The instructions should refer to the specific visual information, such as the people or the objects in the video; 2) The instructions should lead to a descriptive response and the answers should describe the object mentioned in the instruction in detail. Finally, we filter the generated results by discarding answers containing invalid information, such as the phrase "not provide", or "not describe". + +Dataset Statistics. Tab. 1 and Fig. 4 shows some statistics about our constructed dataset. We collect 248K video clips in total from Panda-70M and Ego4D, the average of the video lengths is 25.67 seconds and the distribution is shown in Fig. 4. With the help of Qwen2-VL-72B-Instruct, we generate 739K instruction-followed descriptions, providing sufficient data for our guidance pre-training. + +# 4. Experiments + +# 4.1. Experimental Setup + +Implementation Details. We use SigLIP [64] (so400m-patch14-384) as our vision encoder and text encoder, choose Qwen2.5 series [48] as our LLMs, and randomly initialize the compressor. The vision encoder keeps frozen at all stages, the LLM is frozen at the two pre-train stages, and is fine-tuned at the instruction tuning stage. We follow LLaVA-OneVision [20] to choose our training configurations, the global batch size is set to 512 at the alignment stage, 256 at the conditional pre-train and instruction tun + +Table 3. Performance of our HICom and other SOTA methods on open-ended QA video benchmarks. All the compared models are 7B.†, *, and § keep the same meaning with Tab. 2. #F., #T. are short for # Frames, # Tokens, ANet is short for ActivityNet, LLaVA-NV, LLaVA-OV are short for LLaVA-Next-Video, and LLaVA-OneVision. The format we report ANet is Acc/Score. + +
Methods# F.# T.ANetVideoChatGPT Bench
CIDOCUTUCOAvg
Video-LLaVA [30]8204845.3/3.3------
VideoChat2 [24]16153649.1/3.33.022.883.512.662.812.98
Chat-Univi [17]6444845.8/3.22.892.913.462.892.812.99
LLaMA-VID [28]1fps2tps47.4/3.32.963.003.532.462.512.89
LongVLM [58]10030547.6/3.32.762.863.342.393.112.89
ST-LLM [35]1651250.9/3.33.233.053.742.932.813.15
PLLAVA [59]16230456.3/3.53.212.863.622.332.932.99
LLaVA-NV [71]32460853.5/3.23.393.293.922.63.123.26
LongVA [69]12818432-/2.83.053.093.772.443.643.20
VideoLLaMA2 [7]16115250.2/3.33.163.083.692.563.143.13
Tarsier [51]16230459.5/3.6------
SF-LLaVA [60]50368056.3/3.43.092.703.572.523.353.04
LLaVA-OV [20]32627256.6/3.6*3.45*3.00*3.71*2.68*3.14*3.20*
HICom(Ours)-1.5B3268053.0/3.53.092.673.402.412.982.91
HICom(Ours)-7B3268058.3/3.73.292.853.592.673.223.12
HICom(Ours)-7B$64132859.4/3.73.322.923.652.743.353.20
HICom(Ours)-7B$128262459.5/3.73.332.863.652.743.323.18
+ +ing stage. We use the learning rate of 1e-3 at the alignment stage, 1e-5 at the instruction tuning stage. At our proposed conditional pre-training stage, we use 1e-3 for the condition injection sub-module and 1e-4 for other parameters in the compressing module. We train 1 epoch for all stages. More implementation details can be found in Appendix A. Datasets. We use LLaVA-558K [32] image-caption pairs for our alignment stage, the constructed 248K HiComPretrain instruction-followed descriptions for conditional pre-train stage. Inspired by [51, 72], we collect 2.68M video instruction data for instruction tuning, including 1.6M from LLaVA-Video [72], 292K from VideoChat2-IT [23] subset, 255K from M4-Instruct-Data [21], 11K from Charades [45], 114K from NTU RGB+D [44], 122K from TVQA [18], and 290K from MiT [38]. We evaluate our model on 5 video benchmarks, including 3 multiple-choice benchmarks VideoMME [10], MVBench [24], EgoSchema [37], and 2 open-ended benchmarks ActivityNet [3], VideoChatGPT Bench [36]. + +# 4.2. Main Results + +Multiple-choice benchmarks. In Tab. 2, we compare the performance of our HICom with different SOTA models on three multiple-choice video benchmarks. We sample 32 frames with 680 tokens to train all our models and sample different frames for inference. Our HICom-7B with 2624 tokens inference achieves the best performance on all three benchmarks. Compared to LLaVA-Video with 6272 tokens, our HICom with only 1328 tokens can obtain better performance, increasing the performance by $2.43\%$ average and + +Table 4. The ablation study on our proposed HICom, $\alpha_{(T,H,W)}$ donates the downsampling ratio of each dimension, # Tok. donates the number of tokens input to LLMs. 8f means we sample 8 frames for training. The conditional compression combining both local and global achieves the best. + +
Methodsα(T,H,W) # Tok.VideoMME w/o sub.MV-BenchEgo-Schema
shortmidall
Unconditional Compression
avg pool(1,3,3)259239.334.933.035.744.8
8f, avg pool(1,3,3)64836.334.432.034.343.6
avg pool(4,3,3)64836.333.332.234.043.7
local(4,3,3)64836.734.432.034.443.7
globalNone3230.031.628.930.134.6
local+global(4,3,3)68037.134.832.334.744.1
Conditional Compression
local(4,3,3)64838.836.133.136.044.0
globalNone3230.830.329.830.335.5
local+global(4,3,3)68039.036.734.236.645.0
+ +![](images/cf33cc9a58344753b90323d9f8ffc36d125af9c1015d01b8aab22a56572af191.jpg) + +![](images/1d5e2e8debfbb43955815382ba001767b69ba0033218181516a7622fcecffc60.jpg) + +![](images/8b2f7eb8258a19b8f485249833140701bdc71fc8e7fca194afb600d5eae8fea9.jpg) +Figure 5. The ablation study on different compressing ratios. The figures show the performance on VideoMME-Short (upper left), MVBench (upper right), EgoSchema (lower left), and the inference time of 7B LLM (lower right). + +![](images/c80cb8ee8b1cd166bc58300308856f173b7bb730762aaf9f3b447734015c2b4a.jpg) + +saving $78.8\%$ tokens, HICom with 2624 tokens gains $3\%$ averagely when saving $58.2\%$ tokens, significantly lowering the computational burden. Notably, LLaVA-Video uses numerous additional image data, and it also unfreezes the vision encoder during training, which can both increase performance. It is also noteworthy that though our HICom performs a little worse than LLaVA-Video on VideoMME short videos (less than 2 minutes), we beat LLaVA-Video on both VideoMME medium (4-15 minutes) and long videos (30-60 minutes). We argue that conditional compression is more helpful for long videos. For short videos, the sampled frames are dense enough for unconditional compression, baselines with enough information would naturally perform powerfully. But for long videos, the sampled frames are more sparse with less redundant, our HICom can preserve + +Table 5. The ablation study on the conditional pre-training stage with the constructed HICom-248K dataset. + +
MethodsConditional +Pre-trainedVideoMMEw/o subs.MV-BenchEgo-Schema
shortmidlongall
avg poolX36.632.433.234.142.740.6
✓36.333.332.234.043.741.9
w/o inj.X39.433.733.135.442.841.64
✓37.134.832.334.744.142.4
w/ inj.X38.835.134.136.044.041.6
✓39.036.734.236.645.043.5
+ +the maximum amount of instruct-relevant information while it is easily suppressed in LLaVA-Video. + +Open-ended benchmarks. We also compare the performance of different models on two open-ended video benchmarks, as shown in Tab. 3. We evaluate all our results via GPT-3.5-Turbo-0125. Our HICom also achieves comparable results on both benchmarks, as arriving at SOTA on ActivityNet and getting the second best on the VideoChatGPT Bench. Our HICom with 1328 tokens is on par with Tarsier with much more training data, and performs similarly to LLaVA-OneVision with much more training data and much more visual tokens, demonstrating the effectiveness of our compression. + +Generalization Ability on Video Length. Due to the design of the local-level compression, we can easily extend the lengths of sampled frames at the inference stage for our trained model with 32 frames, rather than re-training it. As we increase the number of sampled frames, the metrics of our HICom also improve across all these benchmarks. For example, the performance on short and medium videos of VideoMME grows fast from 32 frames to 128 frames at our high compressing ratio, as more useful frames are sampled. The performance on long videos of VideoMME without subtitles changes slightly, and we argue this may caused by the extremely long videos (30-60 minutes). Both 32 frames and 128 frames are too sparse for them. Our HICom can pay attention to the most relevant parts based on the sparsely sampled frames, thus changing slightly. + +# 4.3. Ablation Studies + +To save time costs, we choose VideoChat2-IT [23] 896K video data as our instruction tuning data and Qwen2.5-0.5B [48] as our LLM to conduct ablation studies, and extract 32 frames for training unless otherwise specified. This requires less than 28 hours training on 8 A800 GPUs. + +Component Analysis. The key components of our HICom are local- and global- level compression. Tab. 4 shows the results of them in the status of both unconditional (i.e., without condition injection) and conditional compression (i.e., with condition injection). Note we use $V_{p}^{t,h,w}$ as query of Eq. (1) in unconditional compression for direct injection. With the same number of input tokens, the un + +Table 6. The ablation study on different types of injection. + +
LocalGlobalVideoMME w/o sub.MV-BenchEgo-Schema
shortmidlongall
directcoarse39.036.734.236.645.043.5
directfine39.335.334.336.344.143.5
coarsecoarse39.834.835.036.544.142.7
coarsefine38.436.734.236.444.343.5
finecoarse39.134.734.236.044.242.7
finefine38.834.934.236.043.642.0
+ +conditional local compression can achieve comparable performance with both spatial pooling (line 2) and temporal-spatial pooling (line 3). Due to the lack of temporal-spatial inductive bias and too few tokens, the global compression significantly harms the performance, and local+global performs only slightly better than local as the additional information of the global branch is limited without explicit guidance. Compared with unconditional compression, the local and global conditional compression increase by $0.8\%$ and $0.5\%$ , respectively. The local+global achieves the best, increasing by $1.9\%$ on VideoMME and $1.3\%$ average on three benchmarks, also increasing by $0.63\%$ compared with local conditional compression. As the global branch can provide instruction-relevant information from a global sight, combining both is better. It is noteworthy that our conditional compression with 680 tokens can beat the average pooling with 2592 tokens on all benchmarks, reducing $73.8\%$ visual tokens and increasing by $0.47\%$ on the performance. + +Compressing Ratio. We explore the effectiveness of the compressing ratio on both unconditional and conditional compression in Fig. 5. Our unconditional compression performs similarly with average pooling on three benchmarks at different compressing ratios, and the conditional compression performs much better than them. It is noteworthy that the speed of performance decreasing of the conditional compression is also lower than the unconditional compression as the compressing ratio becomes higher. This shows instruction-relevant information remains with the guidance, demonstrating the superiority. In the lower right sub-figure of Fig. 5, the compressing ratio of (4,3,3) saves $57.68\%$ inference time on Qwen2.5-7B than (1,3,3) with only $2.06\%$ performance loss on Qwen2.5-0.5B. Higher compressing ratios (4,9,9) and (8,9,9) can only save a little more time with much higher performance loss. Therefore, we choose (4,3,3) as our final compressing ratio. + +Conditional Pre-training Stage. To evaluate the effectiveness of our proposed conditional pre-training stage, we conduct experiments on whether to use this stage or not on both unconditional and conditional compression in Tab. 5. For average pooling and our unconditional compression, the conditional pre-training stage is also effective on MVBench and EgoSchema since more data is trained, but is hard to be effective on VideoMME. For the conditional compression, + +![](images/ca936bda48ff260a0914c7ee461ef9a2d5ed2daa242042b0d9ee43b4059701de.jpg) +Figure 6. The visualization of the attention map at both the local and global level in the situation of unconditional (w/o inj.) and conditional (w/ inj.) compression. We indicate the instruction-relevant parts with the red bounding box for easier reading. + +the conditional pre-training stage is effective on all benchmarks, increasing by $1.17\%$ averagely, demonstrating the effectiveness of not only the constructed data, but also the newly proposed stage. + +Types of Injection. We propose three types of instruction condition injection in Sec. 3.1, direct injection is designed for local compression only, while coarse and fine injection fit both. We conduct the ablation study on them in Tab. 6. Coarse injection performs better than fine injection on all three benchmarks, and both of them perform better than unconditional compression. We argue the reason may be that the pooled instruction feature is more in line with the CLIP training, and the attention of fine injection is harder to converge than the MLP of coarse injection. + +# 4.4. Qualitative Analysis + +To qualitatively figure out how our HICom realizes the conditional compression, we visualize the attention map of the compression module in Fig. 6. It clearly shows where the model pays attention to based on the instructions. Since we conduct the local-level compression within the subregion of each group, the attention map of the local level is more evenly distributed throughout the video than the global level. Compared to unconditional compression, the attention map of conditional compression is more related to the instruction-relevant visual parts. As shown in Fig. 6, the attention maps at both levels successfully focus on the candle holders in the first example, and successfully on the Christmas tree, the apples, candles, and berries in the second example, which means the model will give more weights of these parts during compressing. We also notice that the global-level compression tends to forget some de + +tailed visual information (e.g., the candle holders in the first and third frame of the first example, the Christmas tree in the second frame of the second example), while the local-level compression is better to capture these details due to group-limited attention. This shows that hybrid-level compression provides different sights, thus helping the model better understand the videos. + +# 5. Conclusion + +In this paper, we address the challenge of achieving efficient video understanding in multi-modal large language models (MLLMs) through our proposed HICom. Our approach effectively balances the trade-off between computational costs and information retention by integrating user instructions into the compression process. This method not only preserves the temporal-spatial structure of video data but also emphasizes instruction-relevant content through a dual-level compression strategy. Furthermore, by introducing a conditional pre-training stage with proposed HICom-248K dataset, targeted pre-training enhances the efficacy of instruction-injected schema. Our experiments demonstrate the effectiveness, as HICom can maintain distinguished video understanding ability with much fewer tokens. HICom shows a good generalization ability to extend the number of frames during inference, but the fixed frame sampling strategy during training still limits the ability to understand long videos. In the future, we will try to sample frames based on fps and add more long videos for training to further increase the ability. We will also explore the potential of conditional compression on images, especially high-resolution images, making our HICom more powerful. + +# Acknowledgment + +This work is supported by the National Key Research and Development Program of China (2022YFB3104700), the National Nature Science Foundation of China (62425114, 62121002, U23B2028, 62232006). We acknowledge the support of Alibaba Group, the GPU cluster built by MCC Lab of Information Science and Technology Institution, USTC, and USTC super-computing center for providing computational resources for this project. + +# References + +[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. NeurIPS, 35: 23716-23736, 2022. 2 +[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 1 +[3] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In CVPR, pages 961-970, 2015. 6 +[4] Lin Chen, Xilin Wei, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Bin Lin, Zhenyu Tang, et al. Sharegpt4video: Improving video understanding and generation with better captions. arXiv preprint arXiv:2406.04325, 2024. 5 +[5] Tsai-Shien Chen, Aliaksandr Siarohin, Willi Menapace, Ekaterina Deyneka, Hsiang-wei Chao, Byung Eun Jeon, Yuwei Fang, Hsin-Ying Lee, Jian Ren, Ming-Hsuan Yang, et al. Panda-70m: Captioning 70m videos with multiple cross-modality teachers. In CVPR, pages 13320-13331, 2024. 5, 3 +[6] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhang-wei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 1 +[7] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024. 2, 4, 5, 6 +[8] Wenliang Dai, Junnan Li, DONGXU LI, Anthony Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale N Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. In NeurIPS, pages 49250-49267, 2023. 2 +[9] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3 +[10] Chaoyou Fu, Yuhan Dai, Yondong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, + +Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 5, 6, 3 +[11] Chaoyou Fu, Haojia Lin, Zuwei Long, Yunhang Shen, Meng Zhao, Yifan Zhang, Xiong Wang, Di Yin, Long Ma, Xiawu Zheng, et al. Vita: Towards open-source interactive omni multimodal llm. arXiv preprint arXiv:2408.05211, 2024. 5 +[12] Zuan Gao, Yuxin Wang, Yadong Qu, Boqiang Zhang, Zixiao Wang, Jianjun Xu, and Hongtao Xie. Self-supervised pretraining with symmetric superimposition modeling for scene text recognition. arXiv preprint arXiv:2405.05841, 2024. 1 +[13] Jiannan Ge, Hongtao Xie, Pandeng Li, Lingxi Xie, Shaobo Min, and Yongdong Zhang. Towards discriminative feature generation for generalized zero-shot learning. IEEE Transactions on Multimedia, 2024. 1 +[14] Jiannan Ge, Lingxi Xie, Hongtao Xie, Pandeng Li, Xiaopeng Zhang, Yongdong Zhang, and Qi Tian. Alignzeg: Mitigating objective misalignment for zero-shot semantic segmentation. In ECCV, pages 142-161. Springer, 2024. 1 +[15] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In CVPR, pages 18995-19012, 2022. 5, 3 +[16] Bo He, Hengduo Li, Young Kyun Jang, Menglin Jia, Xuefei Cao, Ashish Shah, Abhinav Shrivastava, and Ser-Nam Lim. Ma-lmm: Memory-augmented large multimodal model for long-term video understanding. In CVPR, pages 13504-13514, 2024. 2 +[17] Peng Jin, Ryuichi Takanobu, Wancai Zhang, Xiaochun Cao, and Li Yuan. Chat-univi: Unified visual representation empowers large language models with image and video understanding. In CVPR, pages 13700-13710, 2024. 2, 5, 6 +[18] Jie Lei, Licheng Yu, Mohit Bansal, and Tamara L Berg. Tvqa: Localized, compositional video question answering. arXiv preprint arXiv:1809.01696, 2018. 6 +[19] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726, 2023. 2 +[20] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 1, 2, 5, 6 +[21] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava next-interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895, 2024. 6 +[22] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, pages 19730–19742. PMLR, 2023. 1, 2 +[23] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 1, 2, 6, 7 +[24] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. + +Mvbench: A comprehensive multi-modal video understanding benchmark. In CVPR, pages 22195-22206, 2024. 5, 6 +[25] Pandeng Li, Chen-Wei Xie, Hongtao Xie, Liming Zhao, Lei Zhang, Yun Zheng, Deli Zhao, and Yongdong Zhang. Momentdiff: Generative video moment retrieval from random to real. NeurIPS, 36:65948–65966, 2023. 1 +[26] Pandeng Li, Chen-Wei Xie, Liming Zhao, Hongtao Xie, Jiannan Ge, Yun Zheng, Deli Zhao, and Yongdong Zhang. Progressive spatio-temporal prototype matching for text-video retrieval. In ICCV, pages 4100–4110, 2023. 1 +[27] Wentong Li, Yuqian Yuan, Jian Liu, Dongqi Tang, Song Wang, Jianke Zhu, and Lei Zhang. Tokenpacker: Efficient visual projector for multimodal llm. arXiv preprint arXiv:2407.02392, 2024. 3 +[28] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In ECCV, pages 323–340. Springer, 2024. 3, 4, 5, 6 +[29] Jianxin Liang, Xiaojun Meng, Yueqian Wang, Chang Liu, Qun Liu, and Dongyan Zhao. End-to-end video question answering with frame scoring mechanisms and adaptive sampling. arXiv preprint arXiv:2407.15047, 2024. 3 +[30] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 1, 2, 5, 6 +[31] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 36, 2023. 1, 2, 4 +[32] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In CVPR, pages 26296-26306, 2024. 1, 4, 6 +[33] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024. 2, 1 +[34] Ruyang Liu, Chen Li, Yixiao Ge, Thomas H Li, Ying Shan, and Ge Li. Bt-adapter: Video conversation is feasible without video instruction tuning. In CVPR, pages 13658-13667, 2024. 1, 2 +[35] Ruyang Liu, Chen Li, Haoran Tang, Yixiao Ge, Ying Shan, and Ge Li. St-llm: Large language models are effective temporal learners. In European Conference on Computer Vision, pages 1–18. Springer, 2024. 6 +[36] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. In ACL, 2024. 1, 6 +[37] Karttikeya Mangalam, Raiymbek Akshulakov, and Jitendra Malik. Egoschema: A diagnostic benchmark for very long-form video language understanding. NeurIPS, 36:46212-46244, 2023. 6 +[38] Mathew Monfort, Alex Andonian, Bolei Zhou, Kandan Ramakrishnan, Sarah Adel Bargal, Tom Yan, Lisa Brown, Quanfu Fan, Dan Gutfreund, Carl Vondrick, et al. Moments in time dataset: one million videos for event understanding. IEEE TPAMI, 42(2):502-508, 2019. 6 +[39] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, pages 4195-4205, 2023. 4 + +[40] Tianyuan Qu, Longxiang Tang, Bohao Peng, Senqiao Yang, Bei Yu, and Jiaya Jia. Does your vision-language model get lost in the long video sampling dilemma?, 2025. 1 +[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 3 +[42] Machel Reid, Nikolay Savinov, Denis Teplyashin, Dmitry Lepikhin, Timothy Lillicrap, Jean-baptiste Alayrac, Radu Soricut, Angeliki Lazaridou, Orhan First, Julian Schrittwieser, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 1 +[43] Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. In CVPR, pages 14313-14323, 2024. 2 +[44] Amir Shahroudy, Jun Liu, Tian-Tsong Ng, and Gang Wang. Ntu rgb+d: A large scale dataset for 3d human activity analysis. In CVPR, pages 1010-1019, 2016. 6 +[45] Gunnar A. Sigurdsson, Gúl Varol, Xiaolong Wang, Ivan Laptev, Ali Farhadi, and Abhinav Gupta. Hollywood in homes: Crowdsourcing data collection for activity understanding. ArXiv e-prints, 2016. 6 +[46] Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Haozhe Chi, Xun Guo, Tian Ye, Yanting Zhang, et al. Moviechat: From dense token to sparse memory for long video understanding. In CVPR, pages 18221-18232, 2024. 2 +[47] Longxiang Tang, Zhuotao Tian, Kai Li, Chunming He, Hantao Zhou, Hengshuang Zhao, Xiu Li, and Jiaya Jia. Mind the interference: Retaining pre-trained knowledge in parameter efficient continual learning of vision-language models. In ECCV, pages 346-365. Springer, 2024. 1 +[48] Qwen Team. Qwen2.5: A party of foundation models, 2024. 5, 7, 1 +[49] Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, Sai Charitha Akula, Jihan Yang, Shusheng Yang, Adithya Iyer, Xichen Pan, et al. Cambrian1: A fully open, vision-centric exploration of multimodal llms. arXiv preprint arXiv:2406.16860, 2024. 3 +[50] Haibo Wang, Chenghang Lai, Yixuan Sun, and Weifeng Ge. Weakly supervised gaussian contrastive grounding with large multimodal models for video question answering. arXiv preprint arXiv:2401.10711, 2024. 3 +[51] Jiawei Wang, Liping Yuan, Yuchen Zhang, and Haomiao Sun. Tarsier: Recipes for training and evaluating large video description models. arXiv preprint arXiv:2407.00634, 2024. 5, 6 +[52] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1, 5, 3 +[53] Yizhou Wang, Ruiyi Zhang, Haoliang Wang, Uttaran Bhattacharya, Yun Fu, and Gang Wu. Vaquita: Enhancing align + +ment in llm-assisted video understanding. arXiv preprint arXiv:2312.02310, 2023. 3, 4 +[54] Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Guo Chen, Baoqi Pei, Rongkun Zheng, Jilan Xu, Zun Wang, et al. Intervideo2: Scaling video foundation models for multimodal video understanding. arXiv preprint arXiv:2403.15377, 2024. 5, 3 +[55] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. NeurIPS, 35:24824-24837, 2022. 5 +[56] Yujie Wei, Shiwei Zhang, Zhiwu Qing, Hangjie Yuan, Zhiheng Liu, Yu Liu, Yingya Zhang, Jingren Zhou, and Hongming Shan. Dreamvideo: Composing your dream videos with customized subject and motion. In CVPR, pages 6537-6549, 2024. 1 +[57] Yujie Wei, Shiwei Zhang, Hangjie Yuan, Xiang Wang, Haonan Qiu, Rui Zhao, Yutong Feng, Feng Liu, Zhizhong Huang, Jiaxin Ye, et al. Dreamvideo-2: Zero-shot subject-driven video customization with precise motion control. arXiv preprint arXiv:2410.13830, 2024. 1 +[58] Yuetian Weng, Mingfei Han, Haoyu He, Xiaojun Chang, and Bohan Zhuang. Longvlm: Efficient long video understanding via large language models. In ECCV, pages 453-470. Springer, 2024. 6 +[59] Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. Pllava: Parameter-free llava extension from images to videos for video dense captioning. arXiv preprint arXiv:2404.16994, 2024. 2, 5, 6 +[60] Mingze Xu, Mingfei Gao, Zhe Gan, Hong-You Chen, Zhengfeng Lai, Haiming Gang, Kai Kang, and Afshin Dehghan. Slowfast-llava: A strong training-free baseline for video large language models. arXiv preprint arXiv:2407.15841, 2024. 6 +[61] Zhipei Xu, Xuanyu Zhang, Runyi Li, Zecheng Tang, Qing Huang, and Jian Zhang. Fakeshield: Explainable image forgery detection and localization via multi-modal large language models. arXiv preprint arXiv:2410.02761, 2024. 1 +[62] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 1 +[63] Shoubin Yu, Jaemin Cho, Prateek Yadav, and Mohit Bansal. Self-chained image-language model for video localization and question answering. NeurIPS, 36, 2023. 3 +[64] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In ICCV, pages 11975-11986, 2023. 5, 1 +[65] Boqiang Zhang, Hongtao Xie, Zuan Gao, and Yuxin Wang. Choose what you need: Disentangled representation learning for scene text recognition removal and editing. In CVPR, pages 28358-28368, 2024. 1 +[66] Boqiang Zhang, Kehan Li, Zesen Cheng, Zhiqiang Hu, Yuqian Yuan, Guanzheng Chen, Sicong Leng, Yuming Jiang, Hang Zhang, Xin Li, et al. Videollama 3: Frontier multimodal foundation models for image and video understanding. arXiv preprint arXiv:2501.13106, 2025. 1 + +[67] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 2 +[68] Haoji Zhang, Yiqin Wang, Yansong Tang, Yong Liu, Jiashi Feng, Jifeng Dai, and Xiaojie Jin. Flash-vstream: Memory-based real-time understanding for long video streams. arXiv preprint arXiv:2406.08085, 2024. 2, 4 +[69] Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision. arXiv preprint arXiv:2406.16852, 2024. 5, 6 +[70] Xuanyu Zhang, Runyi Li, Jiwen Yu, Youmin Xu, Weiqi Li, and Jian Zhang. Editguard: Versatile image watermarking for tamper localization and copyright protection. In CVPR, pages 11964-11974, 2024. 1 +[71] Yuanhan Zhang, Bo Li, haotian Liu, Yong jae Lee, Liangke Gui, Di Fu, Jiashi Feng, Ziwei Liu, and Chunyuan Li. Llavanext: A strong zero-shot video understanding model, 2024. 2, 5, 6 +[72] Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, and Chunyuan Li. Video instruction tuning with synthetic data. arXiv preprint arXiv:2410.02713, 2024. 2, 5, 6, 3 +[73] Junjie Zhou, Yan Shu, Bo Zhao, Boya Wu, Shitao Xiao, Xi Yang, Yongping Xiong, Bo Zhang, Tiejun Huang, and Zheng Liu. Mlvu: A comprehensive benchmark for multi-task long video understanding. arXiv preprint arXiv:2406.04264, 2024. 6 +[74] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 1 \ No newline at end of file diff --git a/2025/Hybrid-Level Instruction Injection for Video Token Compression in Multi-modal Large Language Models/images.zip b/2025/Hybrid-Level Instruction Injection for Video Token Compression in Multi-modal Large Language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..94683812f255c5b730dbc7152a91e331681e4c75 --- /dev/null +++ b/2025/Hybrid-Level Instruction Injection for Video Token Compression in Multi-modal Large Language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6db5a32e4d33e204628dbf81ed9f3d3117043c88427df8c1e9760033eb5d8a1a +size 704930 diff --git a/2025/Hybrid-Level Instruction Injection for Video Token Compression in Multi-modal Large Language Models/layout.json b/2025/Hybrid-Level Instruction Injection for Video Token Compression in Multi-modal Large Language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f99c262c55d41d7e87ef03491739573d7eba97ec --- /dev/null +++ b/2025/Hybrid-Level Instruction Injection for Video Token Compression in Multi-modal Large Language Models/layout.json @@ -0,0 +1,7864 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 104, + 102, + 507, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 102, + 507, + 140 + ], + "spans": [ + { + "bbox": [ + 104, + 102, + 507, + 140 + ], + "type": "text", + "content": "Hybrid-Level Instruction Injection for Video Token Compression in Multi-modal Large Language Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 113, + 160, + 496, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 160, + 496, + 235 + ], + "spans": [ + { + "bbox": [ + 113, + 160, + 496, + 235 + ], + "type": "text", + "content": "Zhihang Liu1*, Chen-Wei Xie2, Pandeng Li1,2†, Liming Zhao2, Longxiang Tang3, Yun Zheng2, Chuanbin Liu1, Hongtao Xie1 \n1 University of Science and Technology of China \n2 Tongyi Lab, Alibaba Group 3 Tsinghua University \n{liuzhihang, lpd}@mail.ustc.edu.cn, hxie@ustc.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 152, + 261, + 200, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 261, + 200, + 274 + ], + "spans": [ + { + "bbox": [ + 152, + 261, + 200, + 274 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 285, + 296, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 285, + 296, + 622 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 296, + 622 + ], + "type": "text", + "content": "Recent Multi-modal Large Language Models (MLLMs) have been challenged by the computational overhead resulting from massive video frames, often alleviated through compression strategies. However, the visual content is not equally contributed to user instructions, existing strategies (e.g., average pool) inevitably lead to the loss of potentially useful information. To tackle this, we propose the Hybrid-level Instruction Injection Strategy for Conditional Token Compression in MLLMs (HICom), utilizing the instruction as a condition to guide the compression from both local and global levels. This encourages the compression to retain the maximum amount of user-focused information while reducing visual tokens to minimize computational burden. Specifically, the instruction condition is injected into the grouped visual tokens at the local level and the learnable tokens at the global level, and we conduct the attention mechanism to complete the conditional compression. From the hybrid-level compression, the instruction-relevant visual parts are highlighted while the temporal-spatial structure is also preserved for easier understanding of LLMs. To further unleash the potential of HICom, we introduce a new conditional pre-training stage with our proposed dataset HICom248K. Experiments show that our HICom can obtain distinguished video understanding ability with fewer tokens, increasing the performance by " + }, + { + "bbox": [ + 55, + 285, + 296, + 622 + ], + "type": "inline_equation", + "content": "2.43\\%" + }, + { + "bbox": [ + 55, + 285, + 296, + 622 + ], + "type": "text", + "content": " average on three multiple-choice QA benchmarks and saving " + }, + { + "bbox": [ + 55, + 285, + 296, + 622 + ], + "type": "inline_equation", + "content": "78.8\\%" + }, + { + "bbox": [ + 55, + 285, + 296, + 622 + ], + "type": "text", + "content": " tokens compared with the SOTA method. The code is available at https://github.com/1ntzm/HICom." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 642, + 135, + 654 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 642, + 135, + 654 + ], + "spans": [ + { + "bbox": [ + 56, + 642, + 135, + 654 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 662, + 295, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 662, + 295, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 662, + 295, + 687 + ], + "type": "text", + "content": "Multi-modal Large Language Models [2, 14, 22, 31, 32, 74] (MLLMs) have gained significant improvements in multi-" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 319, + 261, + 545, + 449 + ], + "blocks": [ + { + "bbox": [ + 319, + 261, + 545, + 449 + ], + "lines": [ + { + "bbox": [ + 319, + 261, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 319, + 261, + 545, + 449 + ], + "type": "image", + "image_path": "5ef07a75991de3a13c4c1971b74d4b8fe3231e1bba5beccf9cc7f839b45604ac.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 456, + 555, + 522 + ], + "lines": [ + { + "bbox": [ + 313, + 456, + 555, + 522 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 555, + 522 + ], + "type": "text", + "content": "Figure 1. An example of the video understanding task, and the comparison between the unconditional compression and our proposed conditional compression with hybrid-level instruction injection. We inject instruction at both local and global levels, guiding the compression to retain the maximum amount of user-focused information and minimize the computational burden." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 545, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 640 + ], + "type": "text", + "content": "modal understanding by integrating visual information into the LLMs, beating various expert methods on downstream tasks [12, 13, 25, 26, 47, 61, 65, 70]. While initially focused on image understanding, more research [30, 34, 36, 40, 56, 57] has shifted towards challenging video tasks. Most MLLMs handle videos by treating them as sequences of images, sampling frames, and concatenating visual tokens from these frames [6, 20, 42, 52, 66]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "Compared to images, videos comprise multiple frames, resulting in more visual tokens and thus higher computational costs. To make the computation affordable, early methods [23, 30, 34] often employ sparse temporal sampling strategies, leading to significant loss of temporal information. Consequently, current MLLMs focus on compression" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 153, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 153, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 153, + 703 + ], + "type": "text", + "content": "*Interns at Alibaba Group" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 67, + 703, + 143, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 143, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 143, + 713 + ], + "type": "text", + "content": "† Corresponding author" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8568" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "content": "ing visual tokens to achieve a trade-off between computational costs and video understanding ability. Mainstream MLLMs [20, 59, 71] use a spatial pooling strategy within each frame, leveraging capabilities gained from image data. Some approaches [19, 43, 67] use Q-Former [22] or Resampler [1] to compress the spatial-temporal information into a fixed number of visual tokens. Other methods employ convolution [7], clustering [17], or memory mechanisms [16, 46, 68] to compress visual tokens in both spatial and temporal dimensions. However, these compression methods are unconditional, lacking explicit guidance, as illustrated in Fig. 1 (b), making it challenging to achieve high compression ratios with minimal information loss. Consequently, user-relevant visual content may be overlooked during such unconditional compression." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 255, + 296, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 255, + 296, + 483 + ], + "spans": [ + { + "bbox": [ + 56, + 255, + 296, + 483 + ], + "type": "text", + "content": "Therefore, we focus on introducing the instruction information as a condition to perform targeted compression, as it is rarely explored in previous work. As shown in Fig. 1 (c), conditional compression uses explicit instructional guidance to select user-focused information for retention, thereby enhancing efficiency. To thoroughly explore the conditional compression, we propose Hybrid-level Instruction Injection Strategy for Conditional Token Compression in MLLMs (HICom). Observing how humans process complex visual information, they usually adopt a coarse-to-fine approach, first considering the relative parts in the global coarse impression and then seeking detailed information locally in the video. Motivated by this human perception process, HICom conducts the compression at both local and global levels, retaining the maximum amount of instruction-relevant information while preserving the temporal-spatial structure with fewer tokens, thus achieving a better balance between the computational burden and video comprehension." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "type": "text", + "content": "Specifically, at the local level, we divide the sampled frame features into several groups and compress the tokens of each group into one token based on the attention with the injected instruction condition. Different from Q-Former [22] and Resampler [1], the grouped local attention can explicitly maintain the spatial-temporal structural characteristics of the video while highlighting the relevant parts within each group. At the global level, we inject the instruction condition into a small number of learnable tokens and then conduct the attention with the flattened frame features without grouping. The global compression focuses more on searching instruction-relevant information in the whole video and thus can be expressed by a small number of tokens as an auxiliary. To further unleash the potential of HICom, we introduce a new conditional pre-training stage between the alignment and instruction tuning and construct a new dataset HICom-248K of 248K video clips with high-quality instruction-followed descriptions. The proposed new stage continues to pre-train the parameters of" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "condition injection, further improving the performance. Experiments show that our HICom can achieve state-of-the-art performance on 5 benchmarks with significantly fewer tokens. (e.g., increasing by " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "2.43\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " average on three multiple-choice QA benchmarks and saving " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "78.8\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " tokens compared with LLaVA-Video-7B [72])." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 324, + 144, + 507, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 144, + 507, + 155 + ], + "spans": [ + { + "bbox": [ + 324, + 144, + 507, + 155 + ], + "type": "text", + "content": "Our main contributions are listed as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 156, + 553, + 287 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 314, + 156, + 553, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 156, + 553, + 191 + ], + "spans": [ + { + "bbox": [ + 314, + 156, + 553, + 191 + ], + "type": "text", + "content": "- We propose a conditional compression method HICom for MLLMs, achieving effective video understanding while significantly reducing the computational burden." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 192, + 553, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 192, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 192, + 553, + 251 + ], + "type": "text", + "content": "- We conduct hybrid-level instruction injection to achieve the conditional compression, and introduce a new conditional pre-training stage with our constructed HICom248K dataset to further unleash the potential of the conditional compressing." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 251, + 553, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 553, + 287 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 553, + 287 + ], + "type": "text", + "content": "- Experiments on various benchmarks show the proposed HICom can achieve better performance with fewer visual tokens, providing valuable insights for MLLMs." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 298, + 400, + 311 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 298, + 400, + 311 + ], + "spans": [ + { + "bbox": [ + 314, + 298, + 400, + 311 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 319, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 555, + 689 + ], + "type": "text", + "content": "Multi-modal large language models. Based on the powerful capabilities of LLMs among various linguistic tasks, many works try to extend the understanding abilities to the computer vision area. Flamingo [1] and BLIP series [8, 22] successfully explore the usage of Resampler and Q-Former to bridge the visual tokens to language models. LLaVA series [31-33] use a simple MLP as the visual projector to translate visual tokens to the LLM's embedding space, achieving huge success. Recently, more researchers have focused on extending image tasks to video tasks. The biggest challenge for video tasks is to design an effective method to achieve the trade-off between computational costs and video understanding ability. Early methods [23, 30, 34] sparsely sample frames, losing too much temporal information. Mainstream methods [20, 59, 71] use a simple spatial pooling strategy to reduce the number of visual tokens and get a huge success. [19, 67] employ Q-Former to compress the spatial-temporal tokens into a fixed number. Chat-Univi [17] uses DPC-KNN algorithm to cluster dynamic visual tokens. VideoLLaMA2 [7] utilizes convolution both in temporal and spatial dimensions to down-sample tokens. Flash-VStream [68] introduces a learnable memory mechanism to compress online video streaming. All of them try their best to observe more visual information. However, their unconditional compression without explicit guidance inherently leads to the loss of instruction-relevant information. Different from them, we conduct conditional compression by injecting instruction at hybrid levels, only emphasizing the visual parts related to the instruction and allowing the irrelevant parts to be discarded, thus getting a better representation for each question." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "Text-based visual representation for video LLMs. There are some methods using instruction information during the" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8569" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 95, + 70, + 376, + 301 + ], + "blocks": [ + { + "bbox": [ + 95, + 70, + 376, + 301 + ], + "lines": [ + { + "bbox": [ + 95, + 70, + 376, + 301 + ], + "spans": [ + { + "bbox": [ + 95, + 70, + 376, + 301 + ], + "type": "image", + "image_path": "3cfbab6f830d9d93511879a312410f719e193a7a61311e5cc174e7387ec8058e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 310, + 555, + 346 + ], + "lines": [ + { + "bbox": [ + 54, + 310, + 555, + 346 + ], + "spans": [ + { + "bbox": [ + 54, + 310, + 555, + 346 + ], + "type": "text", + "content": "Figure 2. The framework of our proposed HICom. We propose the hybrid-level instruction injection to conditionally compress video tokens in MLLMs. We extract instruction-relevant information within each grouped sub-region at the local level, and extract it to a fixed number of tokens at the global level. The instruction condition is injected into the attention process to guide the compression." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 391, + 76, + 516, + 299 + ], + "blocks": [ + { + "bbox": [ + 391, + 76, + 516, + 299 + ], + "lines": [ + { + "bbox": [ + 391, + 76, + 516, + 299 + ], + "spans": [ + { + "bbox": [ + 391, + 76, + 516, + 299 + ], + "type": "image", + "image_path": "a9bf7acc60fc598ee70d1d5952a758fb3a4ed0b08f9a9405fb91c49a3381d944.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 365, + 297, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 365, + 297, + 582 + ], + "spans": [ + { + "bbox": [ + 54, + 365, + 297, + 582 + ], + "type": "text", + "content": "visual encoding. [53, 63] sample the frames based on the CLIP [41] similarity or a localization model using the instruction, [29, 50] make the frame selection strategy trainable. However, these methods only focus on how to select the correct frames in a rough manner, failing to guide the token compression with the key instruction information. LLaMA-VID [28] follows a similar instruction-guided idea, as it introduces the instruction to interact with visual features. However, it roughly simplifies the interacted but uncompressed result into a single token only as auxiliary information for each frame, destroying the spatial structure and failing to discuss the guiding role of the instruction for conditional compression. Therefore, it is only designed for long videos and the performance is limited. To fully explore conditional compression, we propose to inject the instruction at hybrid levels, exploring a better way for conditional compression to alleviate the information loss while maintaining the performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 592, + 112, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 592, + 112, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 592, + 112, + 604 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 611, + 291, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 291, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 291, + 624 + ], + "type": "text", + "content": "3.1. Instruction-Guided Conditional Compression" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 629, + 296, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 700 + ], + "type": "text", + "content": "Fig. 2 shows the framework of our proposed HICom, which focuses on the design between the vision encoder and the LLM to conduct the conditional compression. We inject the instruction condition into the compression process at both local and global levels for better information extraction with spatial-temporal structure maintained." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 701, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 296, + 713 + ], + "type": "text", + "content": "Local-Level Compression. Previous works have demon-" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 365, + 556, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 556, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 556, + 548 + ], + "type": "text", + "content": "strated the importance of the spatial inductive bias when processing visual representations for image MLMs [27, 49]. Inspired by them, we divide the tokens into several groups and perform conditional compression within each group to preserve the temporal-spatial structure of video features. Specifically, given the encoded video frame features " + }, + { + "bbox": [ + 313, + 365, + 556, + 548 + ], + "type": "inline_equation", + "content": "\\mathbf{V} \\in \\mathbb{R}^{T \\times H \\times W \\times D}" + }, + { + "bbox": [ + 313, + 365, + 556, + 548 + ], + "type": "text", + "content": ", and the downsampling ratio " + }, + { + "bbox": [ + 313, + 365, + 556, + 548 + ], + "type": "inline_equation", + "content": "(\\alpha_{T}, \\alpha_{H}, \\alpha_{W})" + }, + { + "bbox": [ + 313, + 365, + 556, + 548 + ], + "type": "text", + "content": ", we divide " + }, + { + "bbox": [ + 313, + 365, + 556, + 548 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 313, + 365, + 556, + 548 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 313, + 365, + 556, + 548 + ], + "type": "inline_equation", + "content": "N_{T} \\cdot N_{H} \\cdot N_{W}" + }, + { + "bbox": [ + 313, + 365, + 556, + 548 + ], + "type": "text", + "content": " groups, where " + }, + { + "bbox": [ + 313, + 365, + 556, + 548 + ], + "type": "inline_equation", + "content": "N_{i} = \\lceil i / \\alpha_{i} \\rceil" + }, + { + "bbox": [ + 313, + 365, + 556, + 548 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 365, + 556, + 548 + ], + "type": "inline_equation", + "content": "i \\in \\{T, H, W\\}" + }, + { + "bbox": [ + 313, + 365, + 556, + 548 + ], + "type": "text", + "content": ". Therefore, there are " + }, + { + "bbox": [ + 313, + 365, + 556, + 548 + ], + "type": "inline_equation", + "content": "\\alpha_{T} \\cdot \\alpha_{H} \\cdot \\alpha_{W}" + }, + { + "bbox": [ + 313, + 365, + 556, + 548 + ], + "type": "text", + "content": " tokens in each group, which can be formulated as " + }, + { + "bbox": [ + 313, + 365, + 556, + 548 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^{t,h,w} \\in \\mathbb{R}^{\\alpha_{T} \\times \\alpha_{H} \\times \\alpha_{W} \\times D}" + }, + { + "bbox": [ + 313, + 365, + 556, + 548 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 365, + 556, + 548 + ], + "type": "inline_equation", + "content": "0 \\leq \\{t,h,w\\} \\leq \\{N_{T}, N_{H}, N_{W}\\}" + }, + { + "bbox": [ + 313, + 365, + 556, + 548 + ], + "type": "text", + "content": ". We first pool " + }, + { + "bbox": [ + 313, + 365, + 556, + 548 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^{t,h,w}" + }, + { + "bbox": [ + 313, + 365, + 556, + 548 + ], + "type": "text", + "content": " into one token " + }, + { + "bbox": [ + 313, + 365, + 556, + 548 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{p}^{t,h,w} \\in \\mathbb{R}^{1 \\times D}" + }, + { + "bbox": [ + 313, + 365, + 556, + 548 + ], + "type": "text", + "content": " and inject the instruction condition " + }, + { + "bbox": [ + 313, + 365, + 556, + 548 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 365, + 556, + 548 + ], + "type": "text", + "content": " (i.e., the text feature of instructions) into " + }, + { + "bbox": [ + 313, + 365, + 556, + 548 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{p}^{t,h,w}" + }, + { + "bbox": [ + 313, + 365, + 556, + 548 + ], + "type": "text", + "content": ". Then the local attention within each group can be calculated by:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 326, + 554, + 555, + 576 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 554, + 555, + 576 + ], + "spans": [ + { + "bbox": [ + 326, + 554, + 555, + 576 + ], + "type": "interline_equation", + "content": "\\boldsymbol {Z} _ {l} ^ {t, h, w} = \\operatorname {A t t n} \\left[ \\operatorname {I n j} \\left(\\boldsymbol {V} _ {p} ^ {t, h, w}, \\boldsymbol {C}\\right), \\boldsymbol {V} ^ {t, h, w}, \\boldsymbol {V} ^ {t, h, w} \\right], \\tag {1}", + "image_path": "698335d6af27f213c73cc0fa006c33464945d5250e06a45bc221a01410b452de.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 582, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 582, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\mathrm{Inj}(\\cdot)" + }, + { + "bbox": [ + 313, + 582, + 555, + 689 + ], + "type": "text", + "content": " represents the condition injection, " + }, + { + "bbox": [ + 313, + 582, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\mathrm{Attn}(\\cdot)" + }, + { + "bbox": [ + 313, + 582, + 555, + 689 + ], + "type": "text", + "content": " donates the attention mechanism [9] and the inputs of " + }, + { + "bbox": [ + 313, + 582, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\mathrm{Attn}(\\cdot)" + }, + { + "bbox": [ + 313, + 582, + 555, + 689 + ], + "type": "text", + "content": " are query, key, and value, respectively. Therefore, the visual features within each group are compressed into only one token, preserving the temporal-spatial structure while conditionally highlighting the instruction-relevant parts. Finally, we concatenate " + }, + { + "bbox": [ + 313, + 582, + 555, + 689 + ], + "type": "inline_equation", + "content": "Z_{l}^{t,h,w}" + }, + { + "bbox": [ + 313, + 582, + 555, + 689 + ], + "type": "text", + "content": " as the local compressed results " + }, + { + "bbox": [ + 313, + 582, + 555, + 689 + ], + "type": "inline_equation", + "content": "Z_{l}\\in \\mathbb{R}^{N_{T}\\times N_{H}\\times N_{W}\\times D}" + }, + { + "bbox": [ + 313, + 582, + 555, + 689 + ], + "type": "text", + "content": ", and use an MLP layer to project " + }, + { + "bbox": [ + 313, + 582, + 555, + 689 + ], + "type": "inline_equation", + "content": "Z_{l}" + }, + { + "bbox": [ + 313, + 582, + 555, + 689 + ], + "type": "text", + "content": " into the LLM's embedding space." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 689, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 712 + ], + "type": "text", + "content": "Global-Level Compression. Though the temporal-spatial structure can be maintained at the local level, the attention" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8570" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 72, + 282, + 247 + ], + "blocks": [ + { + "bbox": [ + 72, + 72, + 282, + 247 + ], + "lines": [ + { + "bbox": [ + 72, + 72, + 282, + 247 + ], + "spans": [ + { + "bbox": [ + 72, + 72, + 282, + 247 + ], + "type": "image", + "image_path": "1bcb442224e098a92c0c13676ac3877e3cc4d7fc04eb3f20b2e55852e8c797ca.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 252, + 295, + 275 + ], + "lines": [ + { + "bbox": [ + 55, + 252, + 295, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 252, + 295, + 275 + ], + "type": "text", + "content": "Figure 3. We introduce a new guidance pre-training stage and implement three-stage training for conditional compression." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 297, + 295, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 297, + 295, + 441 + ], + "spans": [ + { + "bbox": [ + 54, + 297, + 295, + 441 + ], + "type": "text", + "content": "is forced to focus on small sub-regions. However, each sub-region may contain information that is not equally valid for the question, and some sub-regions are even quite irrelevant. Therefore, we also implement conditional compression at the global level to highlight the most relevant parts within the whole video. Specifically, we initialize a small set of learnable tokens " + }, + { + "bbox": [ + 54, + 297, + 295, + 441 + ], + "type": "inline_equation", + "content": "\\pmb{L} \\in \\mathbb{R}^{N_L \\times D}" + }, + { + "bbox": [ + 54, + 297, + 295, + 441 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 54, + 297, + 295, + 441 + ], + "type": "inline_equation", + "content": "N_{L}" + }, + { + "bbox": [ + 54, + 297, + 295, + 441 + ], + "type": "text", + "content": " donates the number of tokens, and then inject the instruction condition into them. Instead of grouping the video frame features, we apply 3D position embedding " + }, + { + "bbox": [ + 54, + 297, + 295, + 441 + ], + "type": "inline_equation", + "content": "\\mathrm{Pos}(\\cdot)" + }, + { + "bbox": [ + 54, + 297, + 295, + 441 + ], + "type": "text", + "content": " for them and directly flatten them at global-level compression. Then the global compressed tokens " + }, + { + "bbox": [ + 54, + 297, + 295, + 441 + ], + "type": "inline_equation", + "content": "Z_{g} \\in \\mathbb{R}^{N_{L} \\times D}" + }, + { + "bbox": [ + 54, + 297, + 295, + 441 + ], + "type": "text", + "content": " can be calculated by:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 91, + 451, + 295, + 465 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 451, + 295, + 465 + ], + "spans": [ + { + "bbox": [ + 91, + 451, + 295, + 465 + ], + "type": "interline_equation", + "content": "\\boldsymbol {Z} _ {g} = \\operatorname {A t t n} \\left[ \\operatorname {I n j} (\\boldsymbol {L}, \\boldsymbol {C}), \\boldsymbol {V} + \\operatorname {P o s} (\\boldsymbol {V}), \\boldsymbol {V} \\right]. \\tag {2}", + "image_path": "a99ae4478bfcb2522bd1917671ebc82f86ac5d7db96691e0202f5a9c5396db2b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 475, + 295, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 295, + 510 + ], + "type": "text", + "content": "An MLP layer is also utilized to project " + }, + { + "bbox": [ + 55, + 475, + 295, + 510 + ], + "type": "inline_equation", + "content": "Z_{g}" + }, + { + "bbox": [ + 55, + 475, + 295, + 510 + ], + "type": "text", + "content": " into the LLM's embedding space. And we finally concatenate " + }, + { + "bbox": [ + 55, + 475, + 295, + 510 + ], + "type": "inline_equation", + "content": "Z_{l}" + }, + { + "bbox": [ + 55, + 475, + 295, + 510 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 475, + 295, + 510 + ], + "type": "inline_equation", + "content": "Z_{g}" + }, + { + "bbox": [ + 55, + 475, + 295, + 510 + ], + "type": "text", + "content": " together for further understanding of LLM." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 511, + 295, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 511, + 295, + 691 + ], + "spans": [ + { + "bbox": [ + 55, + 511, + 295, + 691 + ], + "type": "text", + "content": "Instruction Condition Injection. In order to fulfill the guiding role of the instruction condition, we explore three different types of injection modules, and define them direct injection, coarse injection, and fine injection, respectively. Note the text encoder can obtain both pooled tokens " + }, + { + "bbox": [ + 55, + 511, + 295, + 691 + ], + "type": "inline_equation", + "content": "C_p \\in \\mathbb{R}^{1 \\times D}" + }, + { + "bbox": [ + 55, + 511, + 295, + 691 + ], + "type": "text", + "content": " (i.e., global text embedding) and fine-grained tokens " + }, + { + "bbox": [ + 55, + 511, + 295, + 691 + ], + "type": "inline_equation", + "content": "C_f \\in \\mathbb{R}^{L \\times D}" + }, + { + "bbox": [ + 55, + 511, + 295, + 691 + ], + "type": "text", + "content": " (i.e., token-level text embedding). Given the pooled token " + }, + { + "bbox": [ + 55, + 511, + 295, + 691 + ], + "type": "inline_equation", + "content": "C_p" + }, + { + "bbox": [ + 55, + 511, + 295, + 691 + ], + "type": "text", + "content": ", the direct injection uses an MLP to translate the single token into the injected output without any interaction with " + }, + { + "bbox": [ + 55, + 511, + 295, + 691 + ], + "type": "inline_equation", + "content": "V_p^{t,h,w}" + }, + { + "bbox": [ + 55, + 511, + 295, + 691 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 55, + 511, + 295, + 691 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 511, + 295, + 691 + ], + "type": "text", + "content": ". We employ coarse injection by the adaptive layer norm [39]. Specifically, we use an MLP to regress the scale and shift from " + }, + { + "bbox": [ + 55, + 511, + 295, + 691 + ], + "type": "inline_equation", + "content": "C_p" + }, + { + "bbox": [ + 55, + 511, + 295, + 691 + ], + "type": "text", + "content": ", then add to the visual input " + }, + { + "bbox": [ + 55, + 511, + 295, + 691 + ], + "type": "inline_equation", + "content": "V_p^{t,h,w}" + }, + { + "bbox": [ + 55, + 511, + 295, + 691 + ], + "type": "text", + "content": " or learnable tokens " + }, + { + "bbox": [ + 55, + 511, + 295, + 691 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 511, + 295, + 691 + ], + "type": "text", + "content": ", which we represent them as " + }, + { + "bbox": [ + 55, + 511, + 295, + 691 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 55, + 511, + 295, + 691 + ], + "type": "text", + "content": ", after the layer norm. The process can be formulated as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 80, + 700, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 700, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 80, + 700, + 295, + 715 + ], + "type": "interline_equation", + "content": "\\operatorname {I n j} (\\boldsymbol {A}, \\boldsymbol {C}) = \\operatorname {L N} (\\boldsymbol {A}) \\cdot \\operatorname {s c a l e} \\left(\\boldsymbol {C} _ {p}\\right) + \\operatorname {s h i f t} \\left(\\boldsymbol {C} _ {p}\\right), \\tag {3}", + "image_path": "ba9e280e3637970d25a6a7e6432973a2b1ae091b242d80c96952373c9765e354.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 71, + 555, + 91 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 71, + 555, + 91 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 555, + 91 + ], + "type": "text", + "content": "Table 1. The statistical information of our constructed HICom-248K dataset." + } + ] + } + ], + "index": 7, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 334, + 102, + 535, + 235 + ], + "blocks": [ + { + "bbox": [ + 334, + 102, + 535, + 235 + ], + "lines": [ + { + "bbox": [ + 334, + 102, + 535, + 235 + ], + "spans": [ + { + "bbox": [ + 334, + 102, + 535, + 235 + ], + "type": "image", + "image_path": "4a1150e8b5d88adcbbdf6193ef8a0c26196888886a856966727d8ad86691c31e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 243, + 555, + 264 + ], + "lines": [ + { + "bbox": [ + 313, + 243, + 555, + 264 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 555, + 264 + ], + "type": "text", + "content": "Figure 4. The visualization of data source (left) and video length (right) of our constructed HICom-248K dataset." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 286, + 555, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 286, + 555, + 321 + ], + "spans": [ + { + "bbox": [ + 313, + 286, + 555, + 321 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 286, + 555, + 321 + ], + "type": "inline_equation", + "content": "\\mathrm{LN}(\\cdot)" + }, + { + "bbox": [ + 313, + 286, + 555, + 321 + ], + "type": "text", + "content": " is the layer norm. Given the fine-grained tokens " + }, + { + "bbox": [ + 313, + 286, + 555, + 321 + ], + "type": "inline_equation", + "content": "C_f" + }, + { + "bbox": [ + 313, + 286, + 555, + 321 + ], + "type": "text", + "content": " as condition " + }, + { + "bbox": [ + 313, + 286, + 555, + 321 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 286, + 555, + 321 + ], + "type": "text", + "content": ", we conduct the fine injection via the cross attention, which can be formulated as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 356, + 331, + 555, + 345 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 331, + 555, + 345 + ], + "spans": [ + { + "bbox": [ + 356, + 331, + 555, + 345 + ], + "type": "interline_equation", + "content": "\\operatorname {I n j} (\\boldsymbol {A}, \\boldsymbol {C}) = \\operatorname {A t t n} \\left(\\operatorname {L N} (\\boldsymbol {A}), \\boldsymbol {C} _ {f}, \\boldsymbol {C} _ {f}\\right). \\tag {4}", + "image_path": "894740cf9df888e3ca8b5bcfa6c3eb60a3531a4bf9a6c5d662f31b9f5bc1e834.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 354, + 555, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 555, + 461 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 555, + 461 + ], + "type": "text", + "content": "As the direct injection directly translates the condition to the query of the attention in Eq. (1) and Eq. (2), the token length limits that it is only suitable for local compression as the local branch compresses each group into only one token. On the contrary, the coarse and fine injections are more flexible to fit various situations as the condition is added into " + }, + { + "bbox": [ + 313, + 354, + 555, + 461 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 354, + 555, + 461 + ], + "type": "text", + "content": ". We conduct the experiments on the three types of injection and finally choose direct injection for local compression and coarse injection for global compression." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 461, + 556, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 556, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 556, + 629 + ], + "type": "text", + "content": "Conditional Pre-training Stage. The existing mainstream methods follow the pipeline of alignment first and then instruction tuning [7, 31, 32, 68], and previous text-based methods [28, 53] also follow this pipeline. However, the image-caption pairs cannot provide valid instruction information in the alignment stage, direct instruction tuning with modules that are not adequately aligned will confuse the guiding process, thus harming the performance. Therefore, we propose a new conditional pre-training stage between the alignment and the instruction tuning to make the training process easier, as shown in Fig. 3. In this stage, we use our constructed instruction-followed descriptions to pre-train the compression module with condition injection, fulfilling the guiding role of the instruction." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 636, + 437, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 636, + 437, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 636, + 437, + 647 + ], + "type": "text", + "content": "3.2. Dataset Construction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": "To pre-train the condition injection in the compression module, we construct a new instruction-followed description dataset HICom-248K. Different from common instruction tuning datasets, HICom-248K focuses on providing only one data type, i.e., the description of the visual content re" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "8571" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 85, + 113, + 526, + 297 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "text", + "content": "Table 2. Performance comparison between our HICom and other SOTA methods on multiple-choice QA video benchmarks. † means we use the result reproduced by VideoLLaMA2 [7], * indicates we reproduce the results ourselves using the official checkpoint and inference code provided by authors. § donates we inference with a new length of frames trained by sampling 32 frames." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 85, + 113, + 526, + 297 + ], + "lines": [ + { + "bbox": [ + 85, + 113, + 526, + 297 + ], + "spans": [ + { + "bbox": [ + 85, + 113, + 526, + 297 + ], + "type": "table", + "html": "
MethodsLLM Size# Frames# TokensVideoMME w/o sub.VideoMME w/ sub.MV-BenchEgo-Schema
shortmidlongallshortmidlongall
Video-LLaVA [30]7B8204845.338.036.239.946.140.738.141.643.0-
VideoChat2-Mistral [24]7B16153648.336.335.039.552.839.439.243.860.4-
LLaMA-VID [28]7B1fps2tps---\\(25.9^{\\dagger}\\)----\\(41.9^{\\dagger}\\)\\(38.5^{\\dagger}\\)
Chat-Univi-1.5 [17]7B6444845.740.335.840.651.244.641.845.945.9-
PLLAVA [59]7B162304--------46.6-
LLaVA-Next-Video [71]7B324608\\(49.9^{*}\\)\\(41.4^{*}\\)\\(37.0^{*}\\)\\(42.8^{*}\\)----\\(46.5^{\\dagger}\\)\\(43.9^{\\dagger}\\)
LongVA [69]7B1281843261.150.446.252.661.653.647.654.3--
VideoLLaMA2 [7]7B16115256.045.442.147.9----54.651.7
Tarsier [51]7B162304--------62.656.0
VITA [11]\\(8\\times 7\\text{B}\\)16409665.952.948.655.870.456.250.959.2--
LLaVA-OneVision [20]7B326272\\(70.1^{*}\\)\\(56.4^{*}\\)\\(48.9^{*}\\)\\(58.5^{*}\\)\\(75.8^{*}\\)\\(58.4^{*}\\)\\(51.6^{*}\\)\\(61.9^{*}\\)56.760.1
LLaVA-Video [72]7B326272\\(73.9^{*}\\)\\(57.3^{*}\\)\\(50.4^{*}\\)\\(60.6^{*}\\)\\(76.6^{*}\\)\\(60.3^{*}\\)\\(51.7^{*}\\)\\(62.9^{*}\\)58.657.3
HICom (Ours)1.5B3268061.751.343.952.365.154.747.355.756.450.1
HICom (Ours)7B3268065.757.151.458.168.458.953.160.164.160.5
HICom (Ours)\\( ^§ \\)7B64132867.658.051.258.971.863.754.563.364.762.2
HICom (Ours)\\( ^§ \\)7B128262469.060.251.560.372.764.657.364.864.062.3
", + "image_path": "d70ef979cee86fc9b519a06fcec5ff46e81c7bed7f653c647a5779172d88f445.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 316, + 295, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 295, + 339 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 295, + 339 + ], + "type": "text", + "content": "lated to the instruction, and does not include other rich types of instruction data (e.g., reasoning, counting, etc)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 341, + 295, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 341, + 295, + 519 + ], + "spans": [ + { + "bbox": [ + 54, + 341, + 295, + 519 + ], + "type": "text", + "content": "Data Collection. To ensure the diversity and quality of our videos, we collect videos from public datasets Panda70M [5] and Ego4D [15]. Panda-70M contains various open-domain videos from YouTube, and Ego4D collects many high-resolution ego-centric human activity videos. Notably, we use the original untrimmed videos to cut clips ourselves since the provided split clips of both datasets are too short, greatly decreasing the content complexity. To balance the number of different types of videos, we pre-define 29 categories [10, 72] using natural language (e.g. A video about cooking activity.) and use InternVideo2 [54] to extract both the video features and the pre-defined sentence features. Based on the similarities between them, we select 1,500 videos for each category and then randomly select additional 10,000 videos from the others to ensure diversity." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 521, + 295, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 521, + 295, + 675 + ], + "spans": [ + { + "bbox": [ + 54, + 521, + 295, + 675 + ], + "type": "text", + "content": "Video Processing. The untrimmed videos are too long for our pre-training, and also severely affect the quality of captions and annotations by existing SOTA models. Therefore, we use PySceneDetect to split each video into shorter ones with a much looser threshold than Panda-70M. As the split clips can sometimes be redundant with repeat semantics, we further extract keyframes for each video and only select the split clips containing the keyframes as our final results. Specifically, inspired by ShareGPT4Video [4], we calculate the CLIP score for densely sampled frames within a video, and select the less similar frames as keyframes. Finally, we further filter out the video clips shorter than 5 seconds and longer than 120 seconds." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "Instruction-Followed Description Generation. For the processed video clips, we use Qwen2-VL-72B-Instruct [52] to systematically generate the instruction-followed descrip" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 316, + 555, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 555, + 471 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 555, + 471 + ], + "type": "text", + "content": "tion with the CoT technique [55]. Specifically, we first generate the detailed caption of each input video, then take both the video content and caption as input to generate the three instruction-answer pairs for each video. Different from the normal instruction datasets, we require Qwen2-VL-72B-Instruct to meet the following two rules: 1) The instructions should refer to the specific visual information, such as the people or the objects in the video; 2) The instructions should lead to a descriptive response and the answers should describe the object mentioned in the instruction in detail. Finally, we filter the generated results by discarding answers containing invalid information, such as the phrase \"not provide\", or \"not describe\"." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 472, + 556, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 556, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 556, + 557 + ], + "type": "text", + "content": "Dataset Statistics. Tab. 1 and Fig. 4 shows some statistics about our constructed dataset. We collect 248K video clips in total from Panda-70M and Ego4D, the average of the video lengths is 25.67 seconds and the distribution is shown in Fig. 4. With the help of Qwen2-VL-72B-Instruct, we generate 739K instruction-followed descriptions, providing sufficient data for our guidance pre-training." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 567, + 395, + 581 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 395, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 395, + 581 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 587, + 430, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 587, + 430, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 587, + 430, + 601 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": "Implementation Details. We use SigLIP [64] (so400m-patch14-384) as our vision encoder and text encoder, choose Qwen2.5 series [48] as our LLMs, and randomly initialize the compressor. The vision encoder keeps frozen at all stages, the LLM is frozen at the two pre-train stages, and is fine-tuned at the instruction tuning stage. We follow LLaVA-OneVision [20] to choose our training configurations, the global batch size is set to 512 at the alignment stage, 256 at the conditional pre-train and instruction tun" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8572" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 55, + 144, + 296, + 333 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 138 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 138 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 138 + ], + "type": "text", + "content": "Table 3. Performance of our HICom and other SOTA methods on open-ended QA video benchmarks. All the compared models are 7B.†, *, and § keep the same meaning with Tab. 2. #F., #T. are short for # Frames, # Tokens, ANet is short for ActivityNet, LLaVA-NV, LLaVA-OV are short for LLaVA-Next-Video, and LLaVA-OneVision. The format we report ANet is Acc/Score." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 55, + 144, + 296, + 333 + ], + "lines": [ + { + "bbox": [ + 55, + 144, + 296, + 333 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 296, + 333 + ], + "type": "table", + "html": "
Methods# F.# T.ANetVideoChatGPT Bench
CIDOCUTUCOAvg
Video-LLaVA [30]8204845.3/3.3------
VideoChat2 [24]16153649.1/3.33.022.883.512.662.812.98
Chat-Univi [17]6444845.8/3.22.892.913.462.892.812.99
LLaMA-VID [28]1fps2tps47.4/3.32.963.003.532.462.512.89
LongVLM [58]10030547.6/3.32.762.863.342.393.112.89
ST-LLM [35]1651250.9/3.33.233.053.742.932.813.15
PLLAVA [59]16230456.3/3.53.212.863.622.332.932.99
LLaVA-NV [71]32460853.5/3.23.393.293.922.63.123.26
LongVA [69]12818432-/2.83.053.093.772.443.643.20
VideoLLaMA2 [7]16115250.2/3.33.163.083.692.563.143.13
Tarsier [51]16230459.5/3.6------
SF-LLaVA [60]50368056.3/3.43.092.703.572.523.353.04
LLaVA-OV [20]32627256.6/3.6*3.45*3.00*3.71*2.68*3.14*3.20*
HICom(Ours)-1.5B3268053.0/3.53.092.673.402.412.982.91
HICom(Ours)-7B3268058.3/3.73.292.853.592.673.223.12
HICom(Ours)-7B$64132859.4/3.73.322.923.652.743.353.20
HICom(Ours)-7B$128262459.5/3.73.332.863.652.743.323.18
", + "image_path": "47d9c4ccb18de0fff3a1fe7487d8861d148e241fcbb80bcf2e3dd6ccde83b5c4.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 352, + 296, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 352, + 296, + 582 + ], + "spans": [ + { + "bbox": [ + 54, + 352, + 296, + 582 + ], + "type": "text", + "content": "ing stage. We use the learning rate of 1e-3 at the alignment stage, 1e-5 at the instruction tuning stage. At our proposed conditional pre-training stage, we use 1e-3 for the condition injection sub-module and 1e-4 for other parameters in the compressing module. We train 1 epoch for all stages. More implementation details can be found in Appendix A. Datasets. We use LLaVA-558K [32] image-caption pairs for our alignment stage, the constructed 248K HiComPretrain instruction-followed descriptions for conditional pre-train stage. Inspired by [51, 72], we collect 2.68M video instruction data for instruction tuning, including 1.6M from LLaVA-Video [72], 292K from VideoChat2-IT [23] subset, 255K from M4-Instruct-Data [21], 11K from Charades [45], 114K from NTU RGB+D [44], 122K from TVQA [18], and 290K from MiT [38]. We evaluate our model on 5 video benchmarks, including 3 multiple-choice benchmarks VideoMME [10], MVBench [24], EgoSchema [37], and 2 open-ended benchmarks ActivityNet [3], VideoChatGPT Bench [36]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 588, + 142, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 142, + 600 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 142, + 600 + ], + "type": "text", + "content": "4.2. Main Results" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 605, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 297, + 715 + ], + "type": "text", + "content": "Multiple-choice benchmarks. In Tab. 2, we compare the performance of our HICom with different SOTA models on three multiple-choice video benchmarks. We sample 32 frames with 680 tokens to train all our models and sample different frames for inference. Our HICom-7B with 2624 tokens inference achieves the best performance on all three benchmarks. Compared to LLaVA-Video with 6272 tokens, our HICom with only 1328 tokens can obtain better performance, increasing the performance by " + }, + { + "bbox": [ + 55, + 605, + 297, + 715 + ], + "type": "inline_equation", + "content": "2.43\\%" + }, + { + "bbox": [ + 55, + 605, + 297, + 715 + ], + "type": "text", + "content": " average and" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 316, + 135, + 553, + 282 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 556, + 127 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 556, + 127 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 556, + 127 + ], + "type": "text", + "content": "Table 4. The ablation study on our proposed HICom, " + }, + { + "bbox": [ + 313, + 70, + 556, + 127 + ], + "type": "inline_equation", + "content": "\\alpha_{(T,H,W)}" + }, + { + "bbox": [ + 313, + 70, + 556, + 127 + ], + "type": "text", + "content": " donates the downsampling ratio of each dimension, # Tok. donates the number of tokens input to LLMs. 8f means we sample 8 frames for training. The conditional compression combining both local and global achieves the best." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 135, + 553, + 282 + ], + "lines": [ + { + "bbox": [ + 316, + 135, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 135, + 553, + 282 + ], + "type": "table", + "html": "
Methodsα(T,H,W) # Tok.VideoMME w/o sub.MV-BenchEgo-Schema
shortmidall
Unconditional Compression
avg pool(1,3,3)259239.334.933.035.744.8
8f, avg pool(1,3,3)64836.334.432.034.343.6
avg pool(4,3,3)64836.333.332.234.043.7
local(4,3,3)64836.734.432.034.443.7
globalNone3230.031.628.930.134.6
local+global(4,3,3)68037.134.832.334.744.1
Conditional Compression
local(4,3,3)64838.836.133.136.044.0
globalNone3230.830.329.830.335.5
local+global(4,3,3)68039.036.734.236.645.0
", + "image_path": "2514a43b55185e37a9ed690d754913cd404f4f0dfe661338a7d5dafd81cc8c51.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 319, + 298, + 432, + 376 + ], + "blocks": [ + { + "bbox": [ + 319, + 298, + 432, + 376 + ], + "lines": [ + { + "bbox": [ + 319, + 298, + 432, + 376 + ], + "spans": [ + { + "bbox": [ + 319, + 298, + 432, + 376 + ], + "type": "image", + "image_path": "cf33cc9a58344753b90323d9f8ffc36d125af9c1015d01b8aab22a56572af191.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 437, + 298, + 549, + 375 + ], + "blocks": [ + { + "bbox": [ + 437, + 298, + 549, + 375 + ], + "lines": [ + { + "bbox": [ + 437, + 298, + 549, + 375 + ], + "spans": [ + { + "bbox": [ + 437, + 298, + 549, + 375 + ], + "type": "image", + "image_path": "1d5e2e8debfbb43955815382ba001767b69ba0033218181516a7622fcecffc60.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 320, + 379, + 431, + 456 + ], + "blocks": [ + { + "bbox": [ + 320, + 379, + 431, + 456 + ], + "lines": [ + { + "bbox": [ + 320, + 379, + 431, + 456 + ], + "spans": [ + { + "bbox": [ + 320, + 379, + 431, + 456 + ], + "type": "image", + "image_path": "8b2f7eb8258a19b8f485249833140701bdc71fc8e7fca194afb600d5eae8fea9.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 464, + 556, + 510 + ], + "lines": [ + { + "bbox": [ + 313, + 464, + 556, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 556, + 510 + ], + "type": "text", + "content": "Figure 5. The ablation study on different compressing ratios. The figures show the performance on VideoMME-Short (upper left), MVBench (upper right), EgoSchema (lower left), and the inference time of 7B LLM (lower right)." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 434, + 379, + 549, + 456 + ], + "blocks": [ + { + "bbox": [ + 434, + 379, + 549, + 456 + ], + "lines": [ + { + "bbox": [ + 434, + 379, + 549, + 456 + ], + "spans": [ + { + "bbox": [ + 434, + 379, + 549, + 456 + ], + "type": "image", + "image_path": "c80cb8ee8b1cd166bc58300308856f173b7bb730762aaf9f3b447734015c2b4a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 533, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 556, + 715 + ], + "type": "text", + "content": "saving " + }, + { + "bbox": [ + 313, + 533, + 556, + 715 + ], + "type": "inline_equation", + "content": "78.8\\%" + }, + { + "bbox": [ + 313, + 533, + 556, + 715 + ], + "type": "text", + "content": " tokens, HICom with 2624 tokens gains " + }, + { + "bbox": [ + 313, + 533, + 556, + 715 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 313, + 533, + 556, + 715 + ], + "type": "text", + "content": " averagely when saving " + }, + { + "bbox": [ + 313, + 533, + 556, + 715 + ], + "type": "inline_equation", + "content": "58.2\\%" + }, + { + "bbox": [ + 313, + 533, + 556, + 715 + ], + "type": "text", + "content": " tokens, significantly lowering the computational burden. Notably, LLaVA-Video uses numerous additional image data, and it also unfreezes the vision encoder during training, which can both increase performance. It is also noteworthy that though our HICom performs a little worse than LLaVA-Video on VideoMME short videos (less than 2 minutes), we beat LLaVA-Video on both VideoMME medium (4-15 minutes) and long videos (30-60 minutes). We argue that conditional compression is more helpful for long videos. For short videos, the sampled frames are dense enough for unconditional compression, baselines with enough information would naturally perform powerfully. But for long videos, the sampled frames are more sparse with less redundant, our HICom can preserve" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "8573" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 102, + 281, + 185 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 297, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 297, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 297, + 94 + ], + "type": "text", + "content": "Table 5. The ablation study on the conditional pre-training stage with the constructed HICom-248K dataset." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 102, + 281, + 185 + ], + "lines": [ + { + "bbox": [ + 69, + 102, + 281, + 185 + ], + "spans": [ + { + "bbox": [ + 69, + 102, + 281, + 185 + ], + "type": "table", + "html": "
MethodsConditional \nPre-trainedVideoMMEw/o subs.MV-BenchEgo-Schema
shortmidlongall
avg poolX36.632.433.234.142.740.6
✓36.333.332.234.043.741.9
w/o inj.X39.433.733.135.442.841.64
✓37.134.832.334.744.142.4
w/ inj.X38.835.134.136.044.041.6
✓39.036.734.236.645.043.5
", + "image_path": "85c8d27682c47939b602102643ebb8d0b374c1758e73318321849c7c5dc35088.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 202, + 295, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 202, + 295, + 227 + ], + "spans": [ + { + "bbox": [ + 55, + 202, + 295, + 227 + ], + "type": "text", + "content": "the maximum amount of instruct-relevant information while it is easily suppressed in LLaVA-Video." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 228, + 295, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 228, + 295, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 295, + 360 + ], + "type": "text", + "content": "Open-ended benchmarks. We also compare the performance of different models on two open-ended video benchmarks, as shown in Tab. 3. We evaluate all our results via GPT-3.5-Turbo-0125. Our HICom also achieves comparable results on both benchmarks, as arriving at SOTA on ActivityNet and getting the second best on the VideoChatGPT Bench. Our HICom with 1328 tokens is on par with Tarsier with much more training data, and performs similarly to LLaVA-OneVision with much more training data and much more visual tokens, demonstrating the effectiveness of our compression." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 361, + 295, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 295, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 295, + 540 + ], + "type": "text", + "content": "Generalization Ability on Video Length. Due to the design of the local-level compression, we can easily extend the lengths of sampled frames at the inference stage for our trained model with 32 frames, rather than re-training it. As we increase the number of sampled frames, the metrics of our HICom also improve across all these benchmarks. For example, the performance on short and medium videos of VideoMME grows fast from 32 frames to 128 frames at our high compressing ratio, as more useful frames are sampled. The performance on long videos of VideoMME without subtitles changes slightly, and we argue this may caused by the extremely long videos (30-60 minutes). Both 32 frames and 128 frames are too sparse for them. Our HICom can pay attention to the most relevant parts based on the sparsely sampled frames, thus changing slightly." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 551, + 156, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 551, + 156, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 551, + 156, + 562 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 568, + 295, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 568, + 295, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 568, + 295, + 629 + ], + "type": "text", + "content": "To save time costs, we choose VideoChat2-IT [23] 896K video data as our instruction tuning data and Qwen2.5-0.5B [48] as our LLM to conduct ablation studies, and extract 32 frames for training unless otherwise specified. This requires less than 28 hours training on 8 A800 GPUs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "type": "text", + "content": "Component Analysis. The key components of our HICom are local- and global- level compression. Tab. 4 shows the results of them in the status of both unconditional (i.e., without condition injection) and conditional compression (i.e., with condition injection). Note we use " + }, + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "type": "inline_equation", + "content": "V_{p}^{t,h,w}" + }, + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "type": "text", + "content": " as query of Eq. (1) in unconditional compression for direct injection. With the same number of input tokens, the un" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 328, + 91, + 539, + 182 + ], + "blocks": [ + { + "bbox": [ + 328, + 71, + 539, + 83 + ], + "lines": [ + { + "bbox": [ + 328, + 71, + 539, + 83 + ], + "spans": [ + { + "bbox": [ + 328, + 71, + 539, + 83 + ], + "type": "text", + "content": "Table 6. The ablation study on different types of injection." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 328, + 91, + 539, + 182 + ], + "lines": [ + { + "bbox": [ + 328, + 91, + 539, + 182 + ], + "spans": [ + { + "bbox": [ + 328, + 91, + 539, + 182 + ], + "type": "table", + "html": "
LocalGlobalVideoMME w/o sub.MV-BenchEgo-Schema
shortmidlongall
directcoarse39.036.734.236.645.043.5
directfine39.335.334.336.344.143.5
coarsecoarse39.834.835.036.544.142.7
coarsefine38.436.734.236.444.343.5
finecoarse39.134.734.236.044.242.7
finefine38.834.934.236.043.642.0
", + "image_path": "7b9771fb764f2a29e950823193f9e15dd1ed6ea6cc65c048779000c2b0f36e7f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 198, + 555, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 198, + 555, + 413 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 555, + 413 + ], + "type": "text", + "content": "conditional local compression can achieve comparable performance with both spatial pooling (line 2) and temporal-spatial pooling (line 3). Due to the lack of temporal-spatial inductive bias and too few tokens, the global compression significantly harms the performance, and local+global performs only slightly better than local as the additional information of the global branch is limited without explicit guidance. Compared with unconditional compression, the local and global conditional compression increase by " + }, + { + "bbox": [ + 313, + 198, + 555, + 413 + ], + "type": "inline_equation", + "content": "0.8\\%" + }, + { + "bbox": [ + 313, + 198, + 555, + 413 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 198, + 555, + 413 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 313, + 198, + 555, + 413 + ], + "type": "text", + "content": ", respectively. The local+global achieves the best, increasing by " + }, + { + "bbox": [ + 313, + 198, + 555, + 413 + ], + "type": "inline_equation", + "content": "1.9\\%" + }, + { + "bbox": [ + 313, + 198, + 555, + 413 + ], + "type": "text", + "content": " on VideoMME and " + }, + { + "bbox": [ + 313, + 198, + 555, + 413 + ], + "type": "inline_equation", + "content": "1.3\\%" + }, + { + "bbox": [ + 313, + 198, + 555, + 413 + ], + "type": "text", + "content": " average on three benchmarks, also increasing by " + }, + { + "bbox": [ + 313, + 198, + 555, + 413 + ], + "type": "inline_equation", + "content": "0.63\\%" + }, + { + "bbox": [ + 313, + 198, + 555, + 413 + ], + "type": "text", + "content": " compared with local conditional compression. As the global branch can provide instruction-relevant information from a global sight, combining both is better. It is noteworthy that our conditional compression with 680 tokens can beat the average pooling with 2592 tokens on all benchmarks, reducing " + }, + { + "bbox": [ + 313, + 198, + 555, + 413 + ], + "type": "inline_equation", + "content": "73.8\\%" + }, + { + "bbox": [ + 313, + 198, + 555, + 413 + ], + "type": "text", + "content": " visual tokens and increasing by " + }, + { + "bbox": [ + 313, + 198, + 555, + 413 + ], + "type": "inline_equation", + "content": "0.47\\%" + }, + { + "bbox": [ + 313, + 198, + 555, + 413 + ], + "type": "text", + "content": " on the performance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 414, + 556, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 414, + 556, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 414, + 556, + 617 + ], + "type": "text", + "content": "Compressing Ratio. We explore the effectiveness of the compressing ratio on both unconditional and conditional compression in Fig. 5. Our unconditional compression performs similarly with average pooling on three benchmarks at different compressing ratios, and the conditional compression performs much better than them. It is noteworthy that the speed of performance decreasing of the conditional compression is also lower than the unconditional compression as the compressing ratio becomes higher. This shows instruction-relevant information remains with the guidance, demonstrating the superiority. In the lower right sub-figure of Fig. 5, the compressing ratio of (4,3,3) saves " + }, + { + "bbox": [ + 313, + 414, + 556, + 617 + ], + "type": "inline_equation", + "content": "57.68\\%" + }, + { + "bbox": [ + 313, + 414, + 556, + 617 + ], + "type": "text", + "content": " inference time on Qwen2.5-7B than (1,3,3) with only " + }, + { + "bbox": [ + 313, + 414, + 556, + 617 + ], + "type": "inline_equation", + "content": "2.06\\%" + }, + { + "bbox": [ + 313, + 414, + 556, + 617 + ], + "type": "text", + "content": " performance loss on Qwen2.5-0.5B. Higher compressing ratios (4,9,9) and (8,9,9) can only save a little more time with much higher performance loss. Therefore, we choose (4,3,3) as our final compressing ratio." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "type": "text", + "content": "Conditional Pre-training Stage. To evaluate the effectiveness of our proposed conditional pre-training stage, we conduct experiments on whether to use this stage or not on both unconditional and conditional compression in Tab. 5. For average pooling and our unconditional compression, the conditional pre-training stage is also effective on MVBench and EgoSchema since more data is trained, but is hard to be effective on VideoMME. For the conditional compression," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 758 + ], + "type": "text", + "content": "8574" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 75, + 70, + 537, + 279 + ], + "blocks": [ + { + "bbox": [ + 75, + 70, + 537, + 279 + ], + "lines": [ + { + "bbox": [ + 75, + 70, + 537, + 279 + ], + "spans": [ + { + "bbox": [ + 75, + 70, + 537, + 279 + ], + "type": "image", + "image_path": "ca936bda48ff260a0914c7ee461ef9a2d5ed2daa242042b0d9ee43b4059701de.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 288, + 555, + 312 + ], + "lines": [ + { + "bbox": [ + 55, + 288, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 288, + 555, + 312 + ], + "type": "text", + "content": "Figure 6. The visualization of the attention map at both the local and global level in the situation of unconditional (w/o inj.) and conditional (w/ inj.) compression. We indicate the instruction-relevant parts with the red bounding box for easier reading." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 331, + 294, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 331, + 294, + 378 + ], + "spans": [ + { + "bbox": [ + 54, + 331, + 294, + 378 + ], + "type": "text", + "content": "the conditional pre-training stage is effective on all benchmarks, increasing by " + }, + { + "bbox": [ + 54, + 331, + 294, + 378 + ], + "type": "inline_equation", + "content": "1.17\\%" + }, + { + "bbox": [ + 54, + 331, + 294, + 378 + ], + "type": "text", + "content": " averagely, demonstrating the effectiveness of not only the constructed data, but also the newly proposed stage." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 379, + 295, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 379, + 295, + 499 + ], + "spans": [ + { + "bbox": [ + 54, + 379, + 295, + 499 + ], + "type": "text", + "content": "Types of Injection. We propose three types of instruction condition injection in Sec. 3.1, direct injection is designed for local compression only, while coarse and fine injection fit both. We conduct the ablation study on them in Tab. 6. Coarse injection performs better than fine injection on all three benchmarks, and both of them perform better than unconditional compression. We argue the reason may be that the pooled instruction feature is more in line with the CLIP training, and the attention of fine injection is harder to converge than the MLP of coarse injection." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 505, + 173, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 505, + 173, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 505, + 173, + 517 + ], + "type": "text", + "content": "4.4. Qualitative Analysis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 521, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 521, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 521, + 295, + 713 + ], + "type": "text", + "content": "To qualitatively figure out how our HICom realizes the conditional compression, we visualize the attention map of the compression module in Fig. 6. It clearly shows where the model pays attention to based on the instructions. Since we conduct the local-level compression within the subregion of each group, the attention map of the local level is more evenly distributed throughout the video than the global level. Compared to unconditional compression, the attention map of conditional compression is more related to the instruction-relevant visual parts. As shown in Fig. 6, the attention maps at both levels successfully focus on the candle holders in the first example, and successfully on the Christmas tree, the apples, candles, and berries in the second example, which means the model will give more weights of these parts during compressing. We also notice that the global-level compression tends to forget some de" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 331, + 555, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 555, + 414 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 555, + 414 + ], + "type": "text", + "content": "tailed visual information (e.g., the candle holders in the first and third frame of the first example, the Christmas tree in the second frame of the second example), while the local-level compression is better to capture these details due to group-limited attention. This shows that hybrid-level compression provides different sights, thus helping the model better understand the videos." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 429, + 388, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 429, + 388, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 429, + 388, + 441 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "text", + "content": "In this paper, we address the challenge of achieving efficient video understanding in multi-modal large language models (MLLMs) through our proposed HICom. Our approach effectively balances the trade-off between computational costs and information retention by integrating user instructions into the compression process. This method not only preserves the temporal-spatial structure of video data but also emphasizes instruction-relevant content through a dual-level compression strategy. Furthermore, by introducing a conditional pre-training stage with proposed HICom-248K dataset, targeted pre-training enhances the efficacy of instruction-injected schema. Our experiments demonstrate the effectiveness, as HICom can maintain distinguished video understanding ability with much fewer tokens. HICom shows a good generalization ability to extend the number of frames during inference, but the fixed frame sampling strategy during training still limits the ability to understand long videos. In the future, we will try to sample frames based on fps and add more long videos for training to further increase the ability. We will also explore the potential of conditional compression on images, especially high-resolution images, making our HICom more powerful." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "8575" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 188 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 188 + ], + "type": "text", + "content": "This work is supported by the National Key Research and Development Program of China (2022YFB3104700), the National Nature Science Foundation of China (62425114, 62121002, U23B2028, 62232006). We acknowledge the support of Alibaba Group, the GPU cluster built by MCC Lab of Information Science and Technology Institution, USTC, and USTC super-computing center for providing computational resources for this project." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 199, + 115, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 199, + 115, + 211 + ], + "spans": [ + { + "bbox": [ + 56, + 199, + 115, + 211 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 219, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 219, + 296, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 219, + 296, + 273 + ], + "spans": [ + { + "bbox": [ + 61, + 219, + 296, + 273 + ], + "type": "text", + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. NeurIPS, 35: 23716-23736, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 274, + 296, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 274, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 61, + 274, + 296, + 319 + ], + "type": "text", + "content": "[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 321, + 295, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 321, + 295, + 364 + ], + "spans": [ + { + "bbox": [ + 61, + 321, + 295, + 364 + ], + "type": "text", + "content": "[3] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In CVPR, pages 961-970, 2015. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 365, + 296, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 365, + 296, + 419 + ], + "spans": [ + { + "bbox": [ + 61, + 365, + 296, + 419 + ], + "type": "text", + "content": "[4] Lin Chen, Xilin Wei, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Bin Lin, Zhenyu Tang, et al. Sharegpt4video: Improving video understanding and generation with better captions. arXiv preprint arXiv:2406.04325, 2024. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 422, + 296, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 422, + 296, + 487 + ], + "spans": [ + { + "bbox": [ + 62, + 422, + 296, + 487 + ], + "type": "text", + "content": "[5] Tsai-Shien Chen, Aliaksandr Siarohin, Willi Menapace, Ekaterina Deyneka, Hsiang-wei Chao, Byung Eun Jeon, Yuwei Fang, Hsin-Ying Lee, Jian Ren, Ming-Hsuan Yang, et al. Panda-70m: Captioning 70m videos with multiple cross-modality teachers. In CVPR, pages 13320-13331, 2024. 5, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 489, + 296, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 489, + 296, + 543 + ], + "spans": [ + { + "bbox": [ + 62, + 489, + 296, + 543 + ], + "type": "text", + "content": "[6] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhang-wei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 544, + 296, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 544, + 296, + 599 + ], + "spans": [ + { + "bbox": [ + 62, + 544, + 296, + 599 + ], + "type": "text", + "content": "[7] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024. 2, 4, 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 601, + 296, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 601, + 296, + 655 + ], + "spans": [ + { + "bbox": [ + 62, + 601, + 296, + 655 + ], + "type": "text", + "content": "[8] Wenliang Dai, Junnan Li, DONGXU LI, Anthony Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale N Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. In NeurIPS, pages 49250-49267, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 658, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 658, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 62, + 658, + 295, + 689 + ], + "type": "text", + "content": "[9] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "text", + "content": "[10] Chaoyou Fu, Yuhan Dai, Yondong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen," + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 5, 6, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "text", + "content": "[11] Chaoyou Fu, Haojia Lin, Zuwei Long, Yunhang Shen, Meng Zhao, Yifan Zhang, Xiong Wang, Di Yin, Long Ma, Xiawu Zheng, et al. Vita: Towards open-source interactive omni multimodal llm. arXiv preprint arXiv:2408.05211, 2024. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 151, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 151, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 151, + 553, + 194 + ], + "type": "text", + "content": "[12] Zuan Gao, Yuxin Wang, Yadong Qu, Boqiang Zhang, Zixiao Wang, Jianjun Xu, and Hongtao Xie. Self-supervised pretraining with symmetric superimposition modeling for scene text recognition. arXiv preprint arXiv:2405.05841, 2024. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 195, + 553, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 553, + 237 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 553, + 237 + ], + "type": "text", + "content": "[13] Jiannan Ge, Hongtao Xie, Pandeng Li, Lingxi Xie, Shaobo Min, and Yongdong Zhang. Towards discriminative feature generation for generalized zero-shot learning. IEEE Transactions on Multimedia, 2024. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 239, + 553, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 282 + ], + "type": "text", + "content": "[14] Jiannan Ge, Lingxi Xie, Hongtao Xie, Pandeng Li, Xiaopeng Zhang, Yongdong Zhang, and Qi Tian. Alignzeg: Mitigating objective misalignment for zero-shot semantic segmentation. In ECCV, pages 142-161. Springer, 2024. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 284, + 553, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 284, + 553, + 337 + ], + "spans": [ + { + "bbox": [ + 316, + 284, + 553, + 337 + ], + "type": "text", + "content": "[15] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In CVPR, pages 18995-19012, 2022. 5, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 338, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 553, + 392 + ], + "type": "text", + "content": "[16] Bo He, Hengduo Li, Young Kyun Jang, Menglin Jia, Xuefei Cao, Ashish Shah, Abhinav Shrivastava, and Ser-Nam Lim. Ma-lmm: Memory-augmented large multimodal model for long-term video understanding. In CVPR, pages 13504-13514, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 393, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 393, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 393, + 553, + 437 + ], + "type": "text", + "content": "[17] Peng Jin, Ryuichi Takanobu, Wancai Zhang, Xiaochun Cao, and Li Yuan. Chat-univi: Unified visual representation empowers large language models with image and video understanding. In CVPR, pages 13700-13710, 2024. 2, 5, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 438, + 553, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 438, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 316, + 438, + 553, + 469 + ], + "type": "text", + "content": "[18] Jie Lei, Licheng Yu, Mohit Bansal, and Tamara L Berg. Tvqa: Localized, compositional video question answering. arXiv preprint arXiv:1809.01696, 2018. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 471, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 471, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 317, + 471, + 553, + 514 + ], + "type": "text", + "content": "[19] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 515, + 553, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 515, + 553, + 558 + ], + "spans": [ + { + "bbox": [ + 316, + 515, + 553, + 558 + ], + "type": "text", + "content": "[20] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 1, 2, 5, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 559, + 553, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 553, + 602 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 553, + 602 + ], + "type": "text", + "content": "[21] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava next-interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895, 2024. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 603, + 553, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 603, + 553, + 647 + ], + "spans": [ + { + "bbox": [ + 316, + 603, + 553, + 647 + ], + "type": "text", + "content": "[22] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, pages 19730–19742. PMLR, 2023. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "text", + "content": "[23] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 1, 2, 6, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[24] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al." + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8576" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "Mvbench: A comprehensive multi-modal video understanding benchmark. In CVPR, pages 22195-22206, 2024. 5, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "text", + "content": "[25] Pandeng Li, Chen-Wei Xie, Hongtao Xie, Liming Zhao, Lei Zhang, Yun Zheng, Deli Zhao, and Yongdong Zhang. Momentdiff: Generative video moment retrieval from random to real. NeurIPS, 36:65948–65966, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 142, + 294, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 142, + 294, + 186 + ], + "spans": [ + { + "bbox": [ + 56, + 142, + 294, + 186 + ], + "type": "text", + "content": "[26] Pandeng Li, Chen-Wei Xie, Liming Zhao, Hongtao Xie, Jiannan Ge, Yun Zheng, Deli Zhao, and Yongdong Zhang. Progressive spatio-temporal prototype matching for text-video retrieval. In ICCV, pages 4100–4110, 2023. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 188, + 295, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 188, + 295, + 231 + ], + "spans": [ + { + "bbox": [ + 56, + 188, + 295, + 231 + ], + "type": "text", + "content": "[27] Wentong Li, Yuqian Yuan, Jian Liu, Dongqi Tang, Song Wang, Jianke Zhu, and Lei Zhang. Tokenpacker: Efficient visual projector for multimodal llm. arXiv preprint arXiv:2407.02392, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 233, + 294, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 233, + 294, + 266 + ], + "spans": [ + { + "bbox": [ + 56, + 233, + 294, + 266 + ], + "type": "text", + "content": "[28] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In ECCV, pages 323–340. Springer, 2024. 3, 4, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 268, + 294, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 268, + 294, + 312 + ], + "spans": [ + { + "bbox": [ + 56, + 268, + 294, + 312 + ], + "type": "text", + "content": "[29] Jianxin Liang, Xiaojun Meng, Yueqian Wang, Chang Liu, Qun Liu, and Dongyan Zhao. End-to-end video question answering with frame scoring mechanisms and adaptive sampling. arXiv preprint arXiv:2407.15047, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 314, + 294, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 314, + 294, + 357 + ], + "spans": [ + { + "bbox": [ + 56, + 314, + 294, + 357 + ], + "type": "text", + "content": "[30] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 1, 2, 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 359, + 294, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 359, + 294, + 380 + ], + "spans": [ + { + "bbox": [ + 56, + 359, + 294, + 380 + ], + "type": "text", + "content": "[31] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 36, 2023. 1, 2, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 383, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 294, + 415 + ], + "type": "text", + "content": "[32] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In CVPR, pages 26296-26306, 2024. 1, 4, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 418, + 294, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 418, + 294, + 451 + ], + "spans": [ + { + "bbox": [ + 56, + 418, + 294, + 451 + ], + "type": "text", + "content": "[33] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024. 2, 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 453, + 294, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 453, + 294, + 495 + ], + "spans": [ + { + "bbox": [ + 56, + 453, + 294, + 495 + ], + "type": "text", + "content": "[34] Ruyang Liu, Chen Li, Yixiao Ge, Thomas H Li, Ying Shan, and Ge Li. Bt-adapter: Video conversation is feasible without video instruction tuning. In CVPR, pages 13658-13667, 2024. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 498, + 294, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 498, + 294, + 541 + ], + "spans": [ + { + "bbox": [ + 56, + 498, + 294, + 541 + ], + "type": "text", + "content": "[35] Ruyang Liu, Chen Li, Haoran Tang, Yixiao Ge, Ying Shan, and Ge Li. St-llm: Large language models are effective temporal learners. In European Conference on Computer Vision, pages 1–18. Springer, 2024. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 544, + 294, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 544, + 294, + 586 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 294, + 586 + ], + "type": "text", + "content": "[36] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. In ACL, 2024. 1, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 589, + 294, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 589, + 294, + 632 + ], + "spans": [ + { + "bbox": [ + 56, + 589, + 294, + 632 + ], + "type": "text", + "content": "[37] Karttikeya Mangalam, Raiymbek Akshulakov, and Jitendra Malik. Egoschema: A diagnostic benchmark for very long-form video language understanding. NeurIPS, 36:46212-46244, 2023. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "text", + "content": "[38] Mathew Monfort, Alex Andonian, Bolei Zhou, Kandan Ramakrishnan, Sarah Adel Bargal, Tom Yan, Lisa Brown, Quanfu Fan, Dan Gutfreund, Carl Vondrick, et al. Moments in time dataset: one million videos for event understanding. IEEE TPAMI, 42(2):502-508, 2019. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[39] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, pages 4195-4205, 2023. 4" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "type": "text", + "content": "[40] Tianyuan Qu, Longxiang Tang, Bohao Peng, Senqiao Yang, Bei Yu, and Jiaya Jia. Does your vision-language model get lost in the long video sampling dilemma?, 2025. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "type": "text", + "content": "[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 163, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 228 + ], + "type": "text", + "content": "[42] Machel Reid, Nikolay Savinov, Denis Teplyashin, Dmitry Lepikhin, Timothy Lillicrap, Jean-baptiste Alayrac, Radu Soricut, Angeliki Lazaridou, Orhan First, Julian Schrittwieser, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 230, + 553, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 553, + 273 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 553, + 273 + ], + "type": "text", + "content": "[43] Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. In CVPR, pages 14313-14323, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 275, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 275, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 275, + 553, + 308 + ], + "type": "text", + "content": "[44] Amir Shahroudy, Jun Liu, Tian-Tsong Ng, and Gang Wang. Ntu rgb+d: A large scale dataset for 3d human activity analysis. In CVPR, pages 1010-1019, 2016. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 309, + 553, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 353 + ], + "type": "text", + "content": "[45] Gunnar A. Sigurdsson, Gúl Varol, Xiaolong Wang, Ivan Laptev, Ali Farhadi, and Abhinav Gupta. Hollywood in homes: Crowdsourcing data collection for activity understanding. ArXiv e-prints, 2016. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 354, + 553, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 408 + ], + "type": "text", + "content": "[46] Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Haozhe Chi, Xun Guo, Tian Ye, Yanting Zhang, et al. Moviechat: From dense token to sparse memory for long video understanding. In CVPR, pages 18221-18232, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 411, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 464 + ], + "type": "text", + "content": "[47] Longxiang Tang, Zhuotao Tian, Kai Li, Chunming He, Hantao Zhou, Hengshuang Zhao, Xiu Li, and Jiaya Jia. Mind the interference: Retaining pre-trained knowledge in parameter efficient continual learning of vision-language models. In ECCV, pages 346-365. Springer, 2024. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 466, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 466, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 466, + 553, + 487 + ], + "type": "text", + "content": "[48] Qwen Team. Qwen2.5: A party of foundation models, 2024. 5, 7, 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 490, + 553, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 490, + 553, + 543 + ], + "spans": [ + { + "bbox": [ + 316, + 490, + 553, + 543 + ], + "type": "text", + "content": "[49] Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, Sai Charitha Akula, Jihan Yang, Shusheng Yang, Adithya Iyer, Xichen Pan, et al. Cambrian1: A fully open, vision-centric exploration of multimodal llms. arXiv preprint arXiv:2406.16860, 2024. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 545, + 553, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 553, + 589 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 553, + 589 + ], + "type": "text", + "content": "[50] Haibo Wang, Chenghang Lai, Yixuan Sun, and Weifeng Ge. Weakly supervised gaussian contrastive grounding with large multimodal models for video question answering. arXiv preprint arXiv:2401.10711, 2024. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 590, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 553, + 633 + ], + "type": "text", + "content": "[51] Jiawei Wang, Liping Yuan, Yuchen Zhang, and Haomiao Sun. Tarsier: Recipes for training and evaluating large video description models. arXiv preprint arXiv:2407.00634, 2024. 5, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "text", + "content": "[52] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1, 5, 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[53] Yizhou Wang, Ruiyi Zhang, Haoliang Wang, Uttaran Bhattacharya, Yun Fu, and Gang Wu. Vaquita: Enhancing align" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8577" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 94 + ], + "type": "text", + "content": "ment in llm-assisted video understanding. arXiv preprint arXiv:2312.02310, 2023. 3, 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 296, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 296, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 296, + 149 + ], + "type": "text", + "content": "[54] Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Guo Chen, Baoqi Pei, Rongkun Zheng, Jilan Xu, Zun Wang, et al. Intervideo2: Scaling video foundation models for multimodal video understanding. arXiv preprint arXiv:2403.15377, 2024. 5, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 151, + 295, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 151, + 295, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 151, + 295, + 194 + ], + "type": "text", + "content": "[55] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. NeurIPS, 35:24824-24837, 2022. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 195, + 295, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 195, + 295, + 248 + ], + "spans": [ + { + "bbox": [ + 56, + 195, + 295, + 248 + ], + "type": "text", + "content": "[56] Yujie Wei, Shiwei Zhang, Zhiwu Qing, Hangjie Yuan, Zhiheng Liu, Yu Liu, Yingya Zhang, Jingren Zhou, and Hongming Shan. Dreamvideo: Composing your dream videos with customized subject and motion. In CVPR, pages 6537-6549, 2024. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 250, + 295, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 295, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 295, + 304 + ], + "type": "text", + "content": "[57] Yujie Wei, Shiwei Zhang, Hangjie Yuan, Xiang Wang, Haonan Qiu, Rui Zhao, Yutong Feng, Feng Liu, Zhizhong Huang, Jiaxin Ye, et al. Dreamvideo-2: Zero-shot subject-driven video customization with precise motion control. arXiv preprint arXiv:2410.13830, 2024. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 304, + 295, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 304, + 295, + 348 + ], + "spans": [ + { + "bbox": [ + 56, + 304, + 295, + 348 + ], + "type": "text", + "content": "[58] Yuetian Weng, Mingfei Han, Haoyu He, Xiaojun Chang, and Bohan Zhuang. Longvlm: Efficient long video understanding via large language models. In ECCV, pages 453-470. Springer, 2024. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 350, + 295, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 295, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 295, + 392 + ], + "type": "text", + "content": "[59] Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. Pllava: Parameter-free llava extension from images to videos for video dense captioning. arXiv preprint arXiv:2404.16994, 2024. 2, 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 394, + 295, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 394, + 295, + 447 + ], + "spans": [ + { + "bbox": [ + 56, + 394, + 295, + 447 + ], + "type": "text", + "content": "[60] Mingze Xu, Mingfei Gao, Zhe Gan, Hong-You Chen, Zhengfeng Lai, Haiming Gang, Kai Kang, and Afshin Dehghan. Slowfast-llava: A strong training-free baseline for video large language models. arXiv preprint arXiv:2407.15841, 2024. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 449, + 295, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 295, + 492 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 295, + 492 + ], + "type": "text", + "content": "[61] Zhipei Xu, Xuanyu Zhang, Runyi Li, Zecheng Tang, Qing Huang, and Jian Zhang. Fakeshield: Explainable image forgery detection and localization via multi-modal large language models. arXiv preprint arXiv:2410.02761, 2024. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 493, + 295, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 493, + 295, + 546 + ], + "spans": [ + { + "bbox": [ + 56, + 493, + 295, + 546 + ], + "type": "text", + "content": "[62] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 548, + 295, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 295, + 581 + ], + "type": "text", + "content": "[63] Shoubin Yu, Jaemin Cho, Prateek Yadav, and Mohit Bansal. Self-chained image-language model for video localization and question answering. NeurIPS, 36, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 582, + 295, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 295, + 613 + ], + "type": "text", + "content": "[64] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In ICCV, pages 11975-11986, 2023. 5, 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 615, + 295, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 615, + 295, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 615, + 295, + 658 + ], + "type": "text", + "content": "[65] Boqiang Zhang, Hongtao Xie, Zuan Gao, and Yuxin Wang. Choose what you need: Disentangled representation learning for scene text recognition removal and editing. In CVPR, pages 28358-28368, 2024. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "type": "text", + "content": "[66] Boqiang Zhang, Kehan Li, Zesen Cheng, Zhiqiang Hu, Yuqian Yuan, Guanzheng Chen, Sicong Leng, Yuming Jiang, Hang Zhang, Xin Li, et al. Videollama 3: Frontier multimodal foundation models for image and video understanding. arXiv preprint arXiv:2501.13106, 2025. 1" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 441 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "type": "text", + "content": "[67] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "text", + "content": "[68] Haoji Zhang, Yiqin Wang, Yansong Tang, Yong Liu, Jiashi Feng, Jifeng Dai, and Xiaojie Jin. Flash-vstream: Memory-based real-time understanding for long video streams. arXiv preprint arXiv:2406.08085, 2024. 2, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "type": "text", + "content": "[69] Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision. arXiv preprint arXiv:2406.16852, 2024. 5, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "type": "text", + "content": "[70] Xuanyu Zhang, Runyi Li, Jiwen Yu, Youmin Xu, Weiqi Li, and Jian Zhang. Editguard: Versatile image watermarking for tamper localization and copyright protection. In CVPR, pages 11964-11974, 2024. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 253, + 553, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 553, + 296 + ], + "type": "text", + "content": "[71] Yuanhan Zhang, Bo Li, haotian Liu, Yong jae Lee, Liangke Gui, Di Fu, Jiashi Feng, Ziwei Liu, and Chunyuan Li. Llavanext: A strong zero-shot video understanding model, 2024. 2, 5, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 297, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 341 + ], + "type": "text", + "content": "[72] Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, and Chunyuan Li. Video instruction tuning with synthetic data. arXiv preprint arXiv:2410.02713, 2024. 2, 5, 6, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 342, + 553, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 553, + 396 + ], + "type": "text", + "content": "[73] Junjie Zhou, Yan Shu, Bo Zhao, Boya Wu, Shitao Xiao, Xi Yang, Yongping Xiong, Bo Zhang, Tiejun Huang, and Zheng Liu. Mlvu: A comprehensive benchmark for multi-task long video understanding. arXiv preprint arXiv:2406.04264, 2024. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 398, + 553, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 398, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 316, + 398, + 553, + 441 + ], + "type": "text", + "content": "[74] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 1" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "8578" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/HybridGS_ Decoupling Transients and Statics with 2D and 3D Gaussian Splatting/347531dd-b02e-491d-be81-0620c6e2bfcd_content_list.json b/2025/HybridGS_ Decoupling Transients and Statics with 2D and 3D Gaussian Splatting/347531dd-b02e-491d-be81-0620c6e2bfcd_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a5342e123e13a97b468e70b336a25e286937122b --- /dev/null +++ b/2025/HybridGS_ Decoupling Transients and Statics with 2D and 3D Gaussian Splatting/347531dd-b02e-491d-be81-0620c6e2bfcd_content_list.json @@ -0,0 +1,1611 @@ +[ + { + "type": "text", + "text": "HybridGS: Decoupling Transients and Statics with 2D and 3D Gaussian Splitting", + "text_level": 1, + "bbox": [ + 267, + 130, + 728, + 175 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jingyu Lin $^{1\\dagger}$ Jiaqi Gu $^{2\\dagger}$ Lubin Fan $^{2\\dagger}$ Bojian Wu $^{2}$ Yujing Lou $^{4}$ Renjie Chen $^{1\\dagger}$ Ligang Liu $^{1}$ Jieping Ye $^{2}$", + "bbox": [ + 212, + 200, + 777, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of Science and Technology of China 2Independent Researcher 4Shanghai Jiao Tong University", + "bbox": [ + 169, + 239, + 826, + 257 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/041433ab006550dde174f0f360819f9503991771e505a0a0cd50cee3009bb315.jpg", + "image_caption": [ + "Figure 1. HybridGS is the first hybrid representation that combines multi-view consistent 3D Gaussians and single-view independent 2D Gaussians, which is used to decouple the transients and statics presented in the scene. Our results demonstrate reasonable decompositions." + ], + "image_footnote": [], + "bbox": [ + 106, + 265, + 313, + 436 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6464bab9e3de9e887ad3450894aea2244a1e6e60c81f408c52babe888a668ef5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 333, + 266, + 473, + 441 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f99fa44d342c780aba4ad3f673914bc8d58a8fc7dd7b5d4cb71239c52c34dd2b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 477, + 267, + 617, + 441 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/39c629258ad6a6120adae51d206764b86889798fabec06f5f7bfa6986e6c8ddd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 620, + 267, + 761, + 441 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ff48fb684431d963da5c2f04ab53208fb965d3dfa5bb43cb6e53dc286dff494a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 763, + 268, + 903, + 441 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 501, + 326, + 516 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generating high-quality novel view renderings of 3D Gaussian Splatting (3DGS) in scenes featuring transient objects is challenging. We propose a novel hybrid representation, termed as HybridGS, using 2D Gaussians for transient objects per image and maintaining traditional 3D Gaussians for the whole static scenes. 3DGS is suited for modeling static scenes that assume multi-view consistency, but the transient objects appear occasionally and do not adhere to the assumption, thus we model them as planar objects from a single view by 2D Gaussians. Our novel representation decomposes the scene from the perspective of fundamental viewpoint consistency. Additionally, we present a multiview supervision method for 3DGS that leverages information from co-visible regions, further enhancing the distinctions between the transients and statics. Then, we propose a straightforward yet effective multi-stage training strategy to ensure robust training and view synthesis. Experiments on benchmarks show our state-of-the-art performance of novel view synthesis in indoor and outdoor scenes, even in the presence of distracting elements. Project page: https://gujiaqivadin.github.io/hybridgs/", + "bbox": [ + 88, + 535, + 485, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 501, + 643, + 516 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D Gaussians Splating (3DGS) [11] has recently gained popularity for novel view synthesis due to its high-quality rendering, efficiency and low-memory cost. Its applications span across virtual reality, augmented reality, and robotics, etc [3, 6, 15, 16, 26, 31, 35, 39, 50]. Typically, most current approaches assume that the input images are posed and free of noise, which are the prerequisite needed to represent a complete scene well. However, this assumption is often not met. For example, images taken casually with a mobile phone usually contain messy dynamic objects, or transient objects. Therefore, it is hard to model the scene because there are always many transient occlusions. Technically, 3DGS is capable of modeling static scenes effectively because the static objects in images satisfy geometric consistency constraints across different viewpoints. However, transient objects do not follow this assumption. If the images containing transient objects are naively used in training 3DGS, such variations can lead to difficulties in achieving coherent blending of individual Gaussian representations, potentially resulting in artifacts or inaccuracies.", + "bbox": [ + 511, + 532, + 906, + 835 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address this issue, numerous methods have been proposed to improve the scene representation ability involving transient objects. In the era of Neural Radiance Fields (NeRF) [21], RobustNeRF [27] introduced a robust loss", + "bbox": [ + 511, + 839, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Equal contributions. ‡Corresponding authors.", + "bbox": [ + 107, + 886, + 357, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "788", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "function that minimizes the influence of photometrically inconsistent observations during training, thereby enhancing the overall quality of the representation. NeRF On-the-go [25] was the first to leverage the uncertainty of transients by incorporating DINOv2 [23] features into a shallow MLP. It focused on loss supervision in the areas of low uncertainty to effectively remove transients from dynamic scenes. These two methods have established widely-used benchmark datasets featuring transients, aiming to evaluate the reconstruction of clear static scenes with improved results of novel view synthesis. For 3DGS, SpotLessSplats (SLS-mlp) [28] integrated the designs of those aforementioned methods, employing adaptive method to detect outliers through clustering in the feature space, achieving the best current results.", + "bbox": [ + 89, + 90, + 480, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, these previous approaches neglect two critical issues: (1) The ambiguity between semantics and transients. Utilizing features from visual models to capture the semantic characteristics of transients can inherently lead to confusion in distinguishing between them. (2) The absence of explicit transient modeling. These methods primarily focus on achieving a robust statics by merely mitigating the impact of transients, rather than comprehensively modeling the whole scene themselves.", + "bbox": [ + 89, + 319, + 480, + 454 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In response to the issues mentioned above, our key observation is that transients lack multi-view consistency and usually only appear in a single view. Therefore, we consider them as planar objects at that view. That is to say, this planar representation is a unique feature of the current view. Based on this, we propose HybridGS, a hybrid representation with 2D Gaussians for transient objects per image and 3D Gaussians for the whole static scenes as illustrated in Fig. 1, which naturally decouples the transients and statics based on their intrinsic properties. With this novel representation, while modeling transients as planar objects, the fundamental consistency of viewpoints is also maintained for statics. To further enhance the performance, a multiview regulated supervision method for 3DGS is adopted to incorporates information from co-visible regions across multiple views, leading to more accurate representation. Additionally, we employ a simple yet effective multi-stage training strategy that ensures robust training and consistent view synthesis across various scenarios, thereby guaranteeing better stability and convergence during the training process. Extensive experiments on benchmark datasets demonstrate our state-of-the-art quality of novel view synthesis in both indoor and outdoor scenes, leading to the accurate decoupling of transients and statics as well as high efficiency in training and inference stage. The contributions are summarized as follows.", + "bbox": [ + 89, + 457, + 482, + 848 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We are the first to introduce a novel hybrid representation that combines image-specific 2D Gaussians with static 3D Gaussians, enabling effective modeling of transient", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "objects within casually captured images.", + "bbox": [ + 526, + 90, + 795, + 106 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We develop a multi-view supervision scheme for 3DGS that utilizes overlapping regions across multiple views, which enhances the model's capability to distinguish between static and transient elements, ultimately improving the overall quality of the novel view synthesis.", + "- Our HybridGS achieves the state-of-the-art performance in benchmark datasets, which outperforms the previous methods and sets a new standard for novel view synthesis in the scenes with transients." + ], + "bbox": [ + 511, + 106, + 903, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 257, + 653, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Novel View Synthesis", + "text_level": 1, + "bbox": [ + 513, + 282, + 712, + 297 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "NeRF (Neural Radiance Fields) [1, 2, 13, 17, 21, 22, 24, 33, 34] revolutionizes 3D scene reconstruction by representing scenes as continuous neural networks that map 3D coordinates to color and density values, which allows for the synthesis of photorealistic views from arbitrary camera positions. Recently, 3D Gaussian Splatting (3DGS) [9, 11, 12, 18, 19, 44, 45] has gained attention as an effective method for novel view synthesis. It starts with an initial point cloud and transforms it into estimizable 3D Gaussian primitives, known for their high-quality rendering and fast processing speeds. Nevertheless, existing methods require the assumption that only static scenes are present in the input images to effectively utilize multi-view geometric consistency to train a coherent scene representation.", + "bbox": [ + 511, + 306, + 903, + 517 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Modeling Transients and Statics in NeRF/3DGS", + "text_level": 1, + "bbox": [ + 513, + 527, + 903, + 544 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In casually captured footage of real-world scenes, there are often numerous dynamic occlusions, such as moving pedestrians and vehicles. Many works have tried to address challenges in scenes featuring both static and transient objects. RobustNeRF [27] was the pioneer work that introduces a robust loss function that reduces the influence of photometrical inconsistency during training, thereby improving the overall reconstruction quality on its proposed dataset. Other methods [4, 20, 40] also applied the similar idea to strengthen the robustness to outliers by down-weighting or discarding inconsistent observations based on the magnitude of color residuals. NeRF On-the-go was the first to introduce DINOV2 [23] features to predict pixel uncertainty, in an attempt to achieve clean static reconstructions without transients. It also introduced a benchmark dataset of casually captured images, aimed at evaluating static reconstruction performance for novel views.", + "bbox": [ + 511, + 551, + 903, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Following these two methods and their datasets, recent methods of 3DGS have also explored techniques for separating transients and statics to reconstruct the statics more effectively. For example, SpotLessSplats [28] further extended previous methods by utilizing semantic features [23] combined with robust optimization strategies to", + "bbox": [ + 511, + 810, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "789", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "effectively eliminate transient distractors. Besides, several approaches [4, 5, 7, 8, 10, 14, 20, 36, 37, 41, 42, 46] have focused on reconstructing statics and illumination in datasets with unconstrained photo collections, that typically employ appearance embedding vectors to model the unique lighting and styles of each image, allowing for more accurate static reconstruction despite varying illumination status.", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our method specifically focuses on distinguishing between transients and statics in casually captured unordered images with minimal illumination changes. Different from previous methods [25, 27, 28], we propose a hybrid Gaussian representation with 2D and 3D Gaussians with multiview information to model the whole 3D scene without any additional semantic features.", + "bbox": [ + 89, + 196, + 480, + 301 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Image Representation with 2D Gaussians", + "text_level": 1, + "bbox": [ + 89, + 311, + 442, + 325 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recently, inspired by 3DGS, some methods [48, 49] in the field of image representation and compression have started to use 2D Gaussians to represent images. GaussianImage [48] leverages 2D Gaussians on the pixel plane to achieve comparable reconstruction quality while offering enhanced compression and accelerated rendering capabilities. Image-GS [49] creates a content-adaptive image representation by fitting a target image by adaptively allocating and progressively optimizing a set of 2D Gaussians.", + "bbox": [ + 89, + 334, + 482, + 469 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Note that the 2D Gaussians used here differ from the 2D Surfels (also referred to as 2D Gaussians) introduced by [9]. While 2D Surfels are designed to enable perspective-correct splatting through 2D surface modeling, ray-splat intersection, and volumetric integration in 3D space, and the 2D Gaussians in GaussianImage [48] is a flexible, compact, and content-adaptive image representation within 2D space.", + "bbox": [ + 89, + 470, + 482, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, we incorporate the trainable 2D Gaussians to learn transient and uncertainty masks during the training process, decomposing the whole scene with 2D and 3D Gaussians.", + "bbox": [ + 89, + 575, + 480, + 621 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 633, + 227, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Gaussian Splatting (3DGS) represent a 3D scene explicitly with a set of 3D anisotropic Gaussians $\\{\\mathcal{G}_{3d}\\}$ . Each Gaussian is parameterized by its centroid $\\mathbf{x_{3d}}\\in \\mathbb{R}^3$ , scale $\\mathbf{S_{3d}}\\in \\mathbb{R}^3$ , rotation matrix $\\mathbf{R}\\in \\mathbb{R}^{3\\times 3}$ , opacity $\\alpha_{3d}\\in \\mathbb{R}$ and color $\\mathbf{c_{3d}}\\in \\mathbb{R}^3$ encoded in spherical harmonic (SH) coefficients. The 3D covariance matrix $\\Sigma_{3d}$ of the 3D Gaussian is obtained from $\\mathbf{R}_{3d}$ and $\\mathbf{S}_{3d}$ :", + "bbox": [ + 89, + 659, + 482, + 765 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\Sigma} _ {3 \\mathbf {d}} = \\mathbf {R} _ {3 \\mathbf {d}} \\mathbf {S} _ {3 \\mathbf {d}} \\mathbf {S} _ {3 \\mathbf {d}} ^ {T} \\mathbf {R} _ {3 \\mathbf {d}} ^ {T}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 773, + 480, + 791 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The 3D Gaussians are defined in world space for a spatial point $\\mathbf{y}_{3d}$ following:", + "bbox": [ + 89, + 801, + 482, + 830 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {G} _ {3 d} \\left(\\mathbf {y} _ {3 \\mathrm {d}}\\right) = e ^ {- \\frac {1}{2} \\left(\\mathbf {y} _ {3 \\mathrm {d}} - \\mathbf {x} _ {3 \\mathrm {d}}\\right) ^ {T} \\boldsymbol {\\Sigma} _ {3 \\mathrm {d}} ^ {- 1} \\left(\\mathbf {y} _ {3 \\mathrm {d}} - \\mathbf {x} _ {3 \\mathrm {d}}\\right)}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 840, + 480, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given $\\mathbf{J}$ , the Jacobian of the affine projective transformation and $\\mathbf{W}$ , a viewing transformation, the covariance ma", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "trix $\\Sigma_{\\mathbf{3d}}^{\\prime}$ in camera coordinates is defined as follows:", + "bbox": [ + 511, + 90, + 862, + 107 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\Sigma} _ {3 \\mathbf {d}} ^ {\\prime} = \\mathbf {J} \\mathbf {W} \\boldsymbol {\\Sigma} _ {3 \\mathbf {d}} \\mathbf {W} ^ {T} \\mathbf {J} ^ {T}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 117, + 903, + 135 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "These 3D Gaussians are projected to 2D imaging plane and blended through a fast, differentiable $\\alpha$ -blending process to render 2D images. The color $\\mathbf{C}_{3d}$ of each pixel is computed using N-ordered 2D splats with the following formula:", + "bbox": [ + 511, + 148, + 903, + 209 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {C} _ {\\mathbf {3 d}} = \\sum_ {i \\in \\mathbf {N}} \\alpha_ {\\mathbf {3 d} _ {i} ^ {\\prime}} \\mathbf {c} _ {\\mathbf {3 d} _ {i}} \\prod_ {j = 1} ^ {i - 1} \\left(1 - \\alpha_ {\\mathbf {3 d} _ {j} ^ {\\prime}}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 583, + 220, + 903, + 263 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\alpha_{\\mathbf{3d}_i'}$ is the final opacity calculated by $\\alpha_{\\mathbf{3d}_i'} = \\alpha_{\\mathbf{3d}_i} * e^{-\\frac{1}{2} (\\mathbf{y}_{\\mathbf{3d}}' - \\mathbf{x}_{\\mathbf{3d}}')^T \\pmb{\\Sigma}_{\\mathbf{3d}}'^{-1} (\\mathbf{y}_{\\mathbf{3d}}' - \\mathbf{x}_{\\mathbf{3d}}')}$ . And $\\mathbf{x}_{\\mathbf{3d}}'$ and $\\mathbf{y}_{\\mathbf{3d}}'$ are coordinates in the projected space. Typically, the centroids of 3D Gaussians are initialized using the sparse SfM point clouds obtained from the set of input images.", + "bbox": [ + 511, + 275, + 903, + 353 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2D Gaussian Splating (2DGS) is initially introduced by [48] in the task of implicit image representation. Since 2D Gaussians $\\{\\mathcal{G}_{2d}\\}$ are no longer oriented to 3D scenes but to 2D space, many bloated operations and redundant parameters are discarded, such as viewing transformation, spherical harmonics, etc. Concretely, a basic 2D Gaussian is described by its centroid $\\mathbf{x_{2d}}\\in \\mathbb{R}^2$ , 2D covariance matrix $\\Sigma_{2\\mathbf{d}}\\in \\mathbb{R}^2$ , color $\\mathbf{c_{2d}}\\in \\mathbb{R}^3$ and opacity $\\alpha_{2\\mathrm{d}}\\in \\mathbb{R}$ . The 2D covariance matrix is also factorized into a rotation matrix $\\mathbf{R}_{2\\mathbf{d}}\\in \\mathbb{R}^{2\\times 2}$ and a scale matrix $\\mathbf{S}_{2\\mathbf{d}}\\in \\mathbb{R}^2$ as:", + "bbox": [ + 511, + 354, + 906, + 505 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\Sigma} _ {2 \\mathrm {d}} = \\mathbf {R} _ {2 \\mathrm {d}} \\mathbf {S} _ {2 \\mathrm {d}} \\mathbf {S} _ {2 \\mathrm {d}} ^ {T} \\mathbf {R} _ {2 \\mathrm {d}} ^ {T}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 516, + 903, + 534 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For rasterization, since the depth and camera parameters are not included in 2D Gaussians, a simplified accumulated summation of N-ordered 2D splats is calculated as follows:", + "bbox": [ + 511, + 547, + 903, + 592 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {C} _ {2 \\mathbf {d}} = \\sum_ {i \\in \\mathbf {N}} \\alpha_ {\\mathbf {2 d} _ {i} ^ {\\prime}} \\mathbf {c} _ {2 \\mathbf {d} _ {i}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 606, + 903, + 637 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\alpha_{\\mathbf{2d}_i'}$ is the opacity calculated by $\\alpha_{\\mathbf{2d}_i'} = \\alpha_{\\mathbf{2d}_i} * e^{-\\frac{1}{2} (\\mathbf{y}_{\\mathbf{2d}} - \\mathbf{x}_{\\mathbf{2d}})^T \\boldsymbol{\\Sigma}_{\\mathbf{2d}}^{-1} (\\mathbf{y}_{\\mathbf{2d}} - \\mathbf{x}_{\\mathbf{2d}})}$ . And $\\mathbf{x}_{\\mathbf{2d}}$ and $\\mathbf{y}_{\\mathbf{2d}}$ are coordinates of centroid and any pixel in the 2D image space.", + "bbox": [ + 511, + 645, + 903, + 693 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, 2D Gaussians can be considered as projections of 3D Gaussians from a certain perspective, illustrating the relationship of Gaussians between 2D and 3D space. In our case, we combine the single-view independent 2D Gaussians and multi-view consistent 3D Gaussians to represent the entire 3D scene, which helps fundamentally decoupling transients and statics, based on the principles of multi-view geometry.", + "bbox": [ + 511, + 694, + 906, + 814 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 511, + 829, + 604, + 844 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We delineate the pipeline of our approach as illustrated in Fig. 2. Essentially, given a set of input images $\\{I_k|k = 1,2,\\dots,N\\}$ with the corresponding camera parameters, for", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "790", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/164f9b07b66db5668985f6fa52c323c8018d9c859a9c11707c268a9e7c6635f2.jpg", + "image_caption": [ + "Figure 2. Overview. Given a casually captured image sequence, we decompose the whole scene into 2D Gaussians for transient objects and 3D Gaussians for static scenes. To warm up, we start by training a basic 3DGS to capture static elements. This is followed by iterative training of 2D and 3D Gaussians, where our transients and statics are combined using an $\\alpha$ -blending strategy with masks to produce the final renderings. The masks provide guidance for 3D Gaussians in the iterative training stage. During the joint-training, both 2D and 3D Gaussians are trained to further optimize the decomposition results." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 903, + 280 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "each view $I$ , the goal of our method is to reasonably decouple the transients $I_{t}$ and statics $I_{s}$ as follows:", + "bbox": [ + 89, + 385, + 482, + 417 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI = M _ {t} \\odot I _ {t} + (1 - M _ {t}) \\odot I _ {s}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 428, + 482, + 446 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the $M_t \\in [0,1]$ represents the transient mask, and $\\odot$ is per pixel multiplication. When a pixel's value in $M_t$ approaches 1, it indicates a higher probability that the location is transient; conversely, a lower value suggests a greater likelihood that area is static.", + "bbox": [ + 89, + 457, + 483, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To achieve the goal, we decompose the whole 3D scene with two components: (1) Multi-view consistent 3D Gaussians is used for rendering $I_{s}$ , which leverages multi-view information from images and models the static scene with a set of unified 3D Gaussians. The multi-view inputs regulate the 3D Gaussians to be consistent and robust across the different views. (2) Single-view independent 2D Gaussians is responsible for modeling $I_{t}$ , which enables our approach to handle the fact that images are casually captured with varying transients. Concretely, we forms a set of view-independent 2D Gaussians to model transients as planar objects from a single view. This novel combination allows for a more precise and reasonable representation of 3D scenes, enabling better performance on novel views.", + "bbox": [ + 88, + 534, + 482, + 744 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This section is organized as follows: we introduce our hybrid representation of scenes using 3D Gaussians in Sec. 4.1 and 2D Gaussians in Sec. 4.2, followed by a comprehensive overview of the training pipeline in Sec. 4.3.", + "bbox": [ + 89, + 744, + 483, + 806 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Modeling Statics with 3D Gaussians", + "text_level": 1, + "bbox": [ + 89, + 816, + 403, + 832 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Generally speaking, the initialized point clouds obtained by COLMAP [29, 30] can only represent areas in the scene that satisfy multi-view consistency, so 3D Gaussians tend to reconstruct static scenes, such as buildings, grounds, etc.", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, in scenes with transients, continuously training the 3D Gaussians may cause the transient objects to be overfitted into the Gaussian kernels because of constraints imposed by the RGB loss. This approach fails to capture the true geometry and instead overemphasizes the RGB information, leading to annoying artifacts from other viewpoints and ultimately impacting quality of novel view synthesis.", + "bbox": [ + 511, + 385, + 903, + 491 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We address this problem by adopting an enhanced multiview regulated supervision scheme. Unlike previous methods that rasterize and supervise only a single image per iteration, our approach introduces two major changes. (1) We increase the number of images per iteration to $K$ , enabling gradient back-propagation to simultaneously consider mutual information from multiple views. This batch-wise inputs allow each optimization step to leverage multi-view insights to distinguish and infer transients and static elements. (2) We adopt a sparse training schedule that focuses optimization only on the 3D Gaussians within the co-visible areas of the cross-view frustums. This method not only sharpens the training focus but also reduces computational costs, as detailed in Alg. 1, and outputs rasterized $\\hat{I}_s$ .", + "bbox": [ + 511, + 491, + 905, + 702 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Modeling Transients with 2D Gaussians", + "text_level": 1, + "bbox": [ + 511, + 710, + 854, + 727 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Different from other methods that utilize semantic features, we introduce 2D Gaussians to model the transients as planar objects from a single view for each image. This approach is motivated by several key considerations: (1) 2DGS has been proven to be an efficient method for representing images, capable of modeling an arbitrary image effectively. (2) The formation and rasterization processes of 2DGS closely resemble those of 3DGS, allowing both to be expressed within a unified framework. (3) Degenerating 3D objects that do not satisfy multi-view geometric consistency into 2D representation can effectively decouple transients from static", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "791", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1: Multi-view Regulated Supervision." + ], + "code_body": "Input: A set of images $\\{I_k, k = 1, 2, \\dots, N\\}$ ; Multi-view size $K$ ; Maximum iterations $T$ ; \nOutput: Optimized 3D Gaussian parameters $\\theta_{3d}$ within the co-visible regions; \nInitialize $\\theta_{3d}$ ; \nfor each iteration $t = 1$ to $T$ do \n\\[ I_{1,\\dots,K} \\leftarrow \\text{RandomSample}(K, N); \\quad F_{\\text{inter}} \\leftarrow \\text{ComputeFrustum}(I_1); \\quad \\text{for } i \\leftarrow 2 \\text{ to } K \\text{ do} \\quad |F_{\\text{inter}} \\leftarrow F_{\\text{inter}} \\cap \\text{ComputeFrustum}(I_i); \\quad \\text{end} \\quad \\mathcal{G}_{3d} \\leftarrow \\text{ExtractGaussians}(I, F_{\\text{inter}}); \\quad \\text{for each Gaussian } G_j \\in \\mathcal{G}_{3d} \\text{ do} \\quad |\\theta \\leftarrow \\theta - \\eta \\cdot \\nabla \\mathcal{L}(G_j(\\theta)); \\quad \\text{end} \\quad \\text{end}", + "bbox": [ + 109, + 112, + 447, + 356 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "scenes on a fundamental level.", + "bbox": [ + 89, + 390, + 294, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, 2D Gaussians are rasterized into an image $\\hat{I}_t$ with a transient mask $\\hat{M}_t$ for modeling uncertainty, each pixel $\\mathbf{y}_{2\\mathbf{d}}$ of which is calculated as:", + "bbox": [ + 89, + 405, + 482, + 450 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {M} _ {t} \\left(\\mathbf {y} _ {2 d}\\right) = \\sum_ {i \\in \\mathbf {N}} \\alpha_ {2 d i} ^ {\\prime} = \\sum_ {i \\in \\mathbf {N}} \\alpha_ {2 d i} \\cdot e ^ {- \\frac {1}{2} \\left(\\mathbf {y} _ {2 d} - \\mathbf {x} _ {2 d}\\right) ^ {T} \\boldsymbol {\\Sigma} _ {2 d} ^ {- 1} \\left(\\mathbf {y} _ {2 d} - \\mathbf {x} _ {2 d}\\right)}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 464, + 482, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "It is noteworthy that we adopt uncertainty of 2D Gaussians to model the mask, rather than using 3D results. Because after modeling a static scene with 3D Gaussians, the 2D Gaussians are responsible for learning the residual part of the image. Therefore, naturally, the results rendered by the 2D Gaussian implicitly include a mask for transients.", + "bbox": [ + 89, + 492, + 483, + 583 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Multi-stage Training Scheme", + "text_level": 1, + "bbox": [ + 89, + 593, + 349, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Because 2D and 3D Gaussians are trained within the same framework, it is crucial to balance their relationship effectively. To that end, we have proposed a multi-stage training strategy that ensures stable convergence and enhances overall performance.", + "bbox": [ + 89, + 617, + 482, + 691 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Warm up Pre-training. The training process begins by using 3DGS to capture the essential structure of the entire static scene. This stage focuses on modeling the scene with a set of unified 3D Gaussians with our multi-view regulated supervision as detailed in Sec. 4.1, producing foundational, albeit low-quality, rendered images $\\hat{I}_s$ using a combination of DSSIM [38] and L1 losses following:", + "bbox": [ + 89, + 693, + 482, + 799 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {w a r m u p} = \\lambda L _ {D S S I M} (\\hat {I} _ {s}, I) + (1 - \\lambda) L _ {1} (\\hat {I} _ {s}, I), \\quad (9)\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 810, + 482, + 829 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the $\\lambda$ is set to 0.2. During the warm-up, the densification of 3D Gaussians is conducted to gradually increase the number, allowing for a more detailed representation of the static scene structure.", + "bbox": [ + 89, + 839, + 483, + 898 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Iterative Training. The second stage employs iterative training between 2DGS and 3DGS to progressively optimize both components. During the alternation, the 3DGS will provide more accurate static scene renderings, which in turn enables the 2DGS to refine its transient representations $\\hat{I}_t$ and mask predictions $\\hat{M}_t$ with Eq. 10. Conversely, the improved masks from 2DGS allow the 3DGS to focus more precisely on rendering the static regions of the scene with Eq. 11. The rendered image $\\hat{I}$ is a composition of $\\hat{I}_s$ and $\\hat{I}_t$ with $\\hat{M}_t$ following Eq. 7. This process iteratively refines the transients and statics, improving the overall quality and accuracy of the masks. The loss functions are as follows:", + "bbox": [ + 511, + 90, + 906, + 272 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {i t e r 2 d}} = \\lambda L _ {\\text {D S S I M}} (\\hat {I}, I) + (1 - \\lambda) L _ {1} (\\hat {I}, I), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 282, + 906, + 299 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {i t e r 3 d}} = \\left(1 - \\hat {M} _ {t}\\right) \\odot \\mathcal {L} _ {\\text {i t e r 2 d}}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 324, + 903, + 342 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that, during training, 2DGS and 3DGS are trained alternately. When training one branch, the gradient backpropagation of the other branch is turned off.", + "bbox": [ + 511, + 348, + 903, + 392 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Joint Fine-tuning. The final stage involves joint fine-tuning, integrating 2DGS and 3DGS into a unified framework. This stage focuses on fine-tuning on the transient mask and reducing the error between the rendered image $\\hat{I}$ and the ground truth image $I$ . The loss $\\mathcal{L}_{joint}$ used during the joint fine-tuning is a combination as follows:", + "bbox": [ + 511, + 393, + 905, + 483 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {j o i n t}} = \\beta L _ {D S S I M} (\\hat {I}, I) + (1 - \\beta) L _ {1} (\\hat {I}, I), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 494, + 903, + 512 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the $\\beta$ is set to 0.2. Here, both the 2DGS and 3DGS branches will enable gradient backpropagation to simultaneously update all parameters.", + "bbox": [ + 511, + 521, + 905, + 568 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 511, + 580, + 645, + 597 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 604, + 702, + 622 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. Similar to previous methods, we evaluate our HybridGS on two challenging datasets: NeRF On-the-go [25] and RobustNeRF [27]. NeRF On-the-go contains multiple casually captured indoor and outdoor sequences with varying ratios of distractors (from $5\\%$ to over $30\\%$ ). We use the version of the dataset where all images were undistorted as [14]. The RobustNeRF dataset includes several scenes exemplifying different types of distractors. To simulate capturing over extended periods, distractor objects are repositioned between frames to interfere with statically placed objects (from 1 (Statue) to 150 (Yoda)). The splits include distractor-filled training set and distractor-free testing set.", + "bbox": [ + 511, + 628, + 906, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. Our HybridGS, developed on top of open-source implementation gsplat [43] and Taming-3DGS [19], incorporates custom CUDA kernels, boosting the rasterization of $\\alpha$ -blending for 3DGS and weighted summation for 2DGS. The number of views $K$ is 4 in our practice. For a scene consisting of 100 images, the process", + "bbox": [ + 511, + 810, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "792", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/8dc6d16ed1da621965b3a32ca680beeb55dce402ccb6bdadba86a4e2a0668eda.jpg", + "table_caption": [ + "Table 1. Comparison on NeRF On-the-go Dataset. For better visualization, the best results are highlighted in bold. Our method shows generally superior performance over state-of-the-art methods. * indicates that the results are reproduced from the official implementation." + ], + "table_footnote": [], + "table_body": "
MethodLow OcclusionMedium OcclusionHigh Occlusion
MountainFountainCornerPatioSpotPatio-High
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
RobustNeRF [27]17.540.4960.38315.650.3180.57623.040.7640.24420.390.7180.25120.650.6250.39120.540.5780.366
NeRF On-the-go [25]20.150.6440.25920.110.6090.31424.220.8060.19020.780.7540.21923.330.7870.18921.410.7180.235
3DGS [11]19.400.6380.21319.960.6590.18520.900.7130.24117.480.7040.19920.770.6930.31617.290.6040.363
SLS-mlp* [28]19.840.5800.31520.190.6120.25924.030.7950.16621.550.8380.32823.520.7560.23520.310.6640.261
HybridGS(Ours)21.730.6930.22521.110.6740.18825.030.8470.07721.980.8120.11024.330.7940.08621.770.7410.142
", + "bbox": [ + 133, + 128, + 867, + 218 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6e1e74455c8458d43b8e8deb203e336ffd3620e81fd980f77ff3f64e3dc57757.jpg", + "table_caption": [ + "Table 2. Comparison on RobustNeRF Dataset. For better visualization, the best results are highlighted in bold. Our method significantly outperforms all baseline methods. Crab (1) scene had a test set with same set of views as those in the train set, which is fixed in Crab (2)." + ], + "table_footnote": [], + "table_body": "
MethodStatueAndroidYodaCrab (1)Crab (2)
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
RobustNeRF [27]20.600.760.1523.280.750.1329.780.820.1532.220.940.06---
NeRF On-the-go [25]21.580.770.2423.500.750.2129.960.830.24------
3DGS [11]21.020.810.1623.110.810.1326.330.910.1431.800.960.0829.74--
SLS-mlp [28]22.540.840.1325.050.850.0933.660.960.1035.850.970.0834.43--
HybridGS(Ours)22.930.870.1025.150.850.0735.320.960.0736.310.970.0535.170.960.08
", + "bbox": [ + 135, + 270, + 867, + 363 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/860441607ade0dc417a0101d00b0f84d0c11598e72e9245213200cb3e6425d1a.jpg", + "table_caption": [ + "Table 3. Ablation on components in datasets with varying occlusions. The best results are highlighted in bold." + ], + "table_footnote": [], + "table_body": "
Dataset(ratio)On-the-go low.(5%)On-the-go medium.(17%)On-the-go high.(26%)
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Ours21.420.6840.20623.500.8270.09223.050.7680.115
w/o iterative Training21.390.6870.20923.110.8150.09422.550.7530.123
w/o Joint Training21.050.6540.24123.180.8150.10723.010.7480.130
w/o Multi-view21.020.6580.23323.060.8120.10322.710.7630.111
", + "bbox": [ + 93, + 426, + 483, + 494 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "starts with a 1k-step warm-up. This is followed by iterative training, which includes 10k steps for 2DGS and 1k steps for 3DGS. The final joint training involves 30k steps. The entire training process takes approximately 0.18 GPU hours. The number of training steps scales proportionally with the size of the dataset. All experiments are performed on a single NVIDIA RTX 4090 GPU using PyTorch.", + "bbox": [ + 88, + 532, + 482, + 637 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. We follow common practice by employing PSNR, SSIM [38], and LPIPS [47] for evaluation.", + "bbox": [ + 89, + 638, + 482, + 670 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Evaluation", + "text_level": 1, + "bbox": [ + 89, + 681, + 210, + 698 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2.1 Comparison on the NeRF On-the-go Dataset", + "text_level": 1, + "bbox": [ + 89, + 707, + 454, + 722 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Tab. 1 and Fig. 3, our method demonstrates superior performance across different scenarios with varying levels of occlusion. In Tab. 1, our method outperforms the previous SOTA method across multiple metrics, with an average improvement of $1.10\\mathrm{dB}$ in PSNR and $5.27\\%$ in SSIM, demonstrating the superiority of our approach.", + "bbox": [ + 89, + 732, + 482, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Furthermore, Fig. 3 provides a qualitative comparison, illustrating the visual fidelity of our results against other methods. Our method produces more clear and detailed images, closely resembling the ground truth, particularly in challenging scenes with complex geometry and occlusions.", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/eed9fe25a34d1f45d5399c477049cd6d252b93132066668e163b000906335e86.jpg", + "table_caption": [ + "Table 4. Ablation on the number of 2D Gaussians of Corner in NeRF On-the-go dataset. For better visualization, the best results are highlighted in bold." + ], + "table_footnote": [], + "table_body": "
NumberPSNR↑SSIM↑LPIPS↓Storage (MB)
2DGS3DGS
1k24.4930.8400.0773.542.1
5k24.7330.8430.08217.341.9
10k25.0340.8470.08034.742.2
15k24.5120.8390.09242.652
25k24.6980.8400.09469.442.2
", + "bbox": [ + 596, + 440, + 826, + 515 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This highlights our model's capability to effectively handle diverse visual conditions both in indoor and outdoor scenes.", + "bbox": [ + 511, + 531, + 906, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2.2 Comparison on the RobustNeRF Dataset", + "text_level": 1, + "bbox": [ + 511, + 584, + 852, + 599 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We present the quantitative results of RobustNeRF in Tab. 2. Our method demonstrates superior performance quantitatively compared to all previous approaches. In scenes with numerous transients, such as the Yoda scene, we've noticed that the performance metrics of the original 3DGS significantly degrade, occasionally even dropping below those of NeRF-related methods. By leveraging multi-view shared information and incorporating 2D Gaussians to effectively represent transients, our approach surpasses semantic-based methods like SLS-mlp [28]. As shown in Fig. 4, our results have clearer overall boundaries and structured fine-grained details, while SLS-mlp [28] may lead to blurriness in static objects and occluded areas. This further confirms that, even when faced with numerous transients, our method consistently maintains stable and reliable performance.", + "bbox": [ + 511, + 608, + 906, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 845, + 679, + 861 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Components. To assess the effectiveness of each component in our framework, we performed an ablation study, as", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "793", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e33afbb0bdf4907ca393059a9b5759b59034735f10b9ffb421385dd0c15538d6.jpg", + "image_caption": [ + "Figure 3. Visualization of novel view synthesis results on the testing set of NeRF On-the-go dataset. Our method demonstrates superior results by effectively reducing artifacts and providing clearer boundaries. This results in a cleaner statics compared to other methods, showcasing enhanced visual quality and precision in novel views." + ], + "image_footnote": [], + "bbox": [ + 98, + 85, + 903, + 329 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9dfe4fd9428dfd8798707caf0a76b15c3988f2de6b6531b144801a2a0e452196.jpg", + "image_caption": [ + "Figure 4. Visualization of novel view synthesis results on RobustNeRF dataset." + ], + "image_footnote": [], + "bbox": [ + 93, + 407, + 480, + 534 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "presented in Tab. 3. We systematically removed iterative training, joint training, and multi-view supervision to evaluate their effectiveness. Each component, as expected, enhances overall performance, demonstrating their individual and combined significance. In scenes with low occlusion, multi-view supervision allows the model to concentrate on static elements by integrating data from various perspectives. Conversely, in highly occluded environments, joint training with 2D and 3D Gaussians markedly improves performance, highlighting the model's capability to manage complex scenes with significant occlusions and transient changes. Iterative training consistently boosts results across different occlusion levels. In summary, the results underscore that integrating diverse training strategies and multiview information significantly enhances the rendering quality of unseen views, especially in challenging conditions.", + "bbox": [ + 88, + 609, + 482, + 851 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The Number of 2D Gaussians. Tab. 4 presents an ablation study on the number of 2D Gaussians used in the corner sequence of NeRF On-the-go dataset. The results indicate that", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d01dbc1fb34ae3d3635214eaf0fd2c6a95d36a5d457c989c044930dd1eb34234.jpg", + "image_caption": [ + "Figure 5. Comparison of transient masks on NeRF On-the-go dataset." + ], + "image_footnote": [], + "bbox": [ + 517, + 407, + 903, + 531 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "increasing the number of 2D Gaussians initially improves the performance. Specifically, 10k 2D Gaussians achieve the highest PSNR (25.034) and SSIM (0.847), while maintaining a competitive LPIPS score of 0.151 and reasonable storage requirements. Based on these findings, we select 10k as the optimal number in our experiments.", + "bbox": [ + 511, + 597, + 906, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Transients and Masks. Benefiting from our method's explicit modeling of transient objects, we can obtain transient masks without introducing any segmentation networks or other pre-trained features. Typically, these masks capture dynamic elements such as pedestrians and vehicles, as well as some structural details ignored by 3D Gaussians. In Fig. 6, our approach effectively learns the RGB and mask of transients even under strong occlusions, while maintaining the robustness of static elements. This capability underscores the strength of our method in discerning and isolating dynamic components within complex scenes.", + "bbox": [ + 511, + 688, + 908, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also compare our transient masks with SLS-mlp [28] semantic masks in Fig. 5. It is observed that the semantic mask does not perfectly align with the transients due to", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "794", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cf5966e71b1ac324ab8b0c069f467e7624bd96a01358cb122fdc01bf4ea115e4.jpg", + "image_caption": [ + "Figure 6. Visualization of scene decomposition into transients and statics. Our method achieves superior transient mask separation in both indoor and outdoor scenes. It effectively separates transients and statics, and the resulting renderings closely resemble the ground truth images, demonstrating its effectiveness. Top two rows are from NeRF On-the-go [25], bottom two rows are from RobustNeRF [27]." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 903, + 450 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d0e6b63f56ad97a30fb02649652e09c61d5c5c8befb4d3741942aae023fde704.jpg", + "image_caption": [ + "Figure 7. Visualization on Photo Tourism dataset." + ], + "image_footnote": [], + "bbox": [ + 94, + 532, + 478, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the semantic features not being sufficiently precise. Additionally, transients often encompass low-level information, such as motion blur or object shadows, which may not explicitly exhibit clear semantic features. By modeling these transients using 2D Gaussians at the pixel level, our method allows for a more fine-grained capture of transients.", + "bbox": [ + 89, + 695, + 482, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Discussions", + "text_level": 1, + "bbox": [ + 89, + 800, + 215, + 814 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While our method has demonstrated strong performance in distinguishing between transients and statics, it has not yet accounted for the illumination variation in unconstrained photo collections. As illustrated in Fig. 7 on Photo Tourism [32], our method renders fine-grained architectural", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "structures, but the overall photometric result tends towards an average photometric effect, akin to overcast conditions. A promising direction is to incorporate an appearance embedding module in our method for the future work.", + "bbox": [ + 511, + 534, + 906, + 594 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 614, + 633, + 631 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce HybridGS, a novel hybrid representation with 2D and 3D Gaussians for decomposing 3D scenes from casually captured images. By introducing the hybrid Gaussian representation, our method addresses the challenges of handling transients as planar objects while preserving the integrity of the static scene. The multi-view supervision mechanism we developed plays a crucial role in enhancing the discrimination between static and transient elements across multiple views, reducing artifacts and inaccuracies in novel views. We have developed a multi-stage training strategy for the joint training of 2D and 3D Gaussians, ensuring the stable and robust training stage. Our experiments demonstrate that HybridGS significantly outperforms existing methods in terms of quality and efficiency. In summary, HybridGS establishes a novel, robust and efficient foundation for 3D scene representation, paving the way for advancements in various vision-based applications.", + "bbox": [ + 511, + 643, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "795", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. ICCV, 2021. 2", + "[2] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. CVPR, 2022. 2", + "[3] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. Pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19457-19467, 2023. 1", + "[4] Jiahao Chen, Yipeng Qin, Lingjie Liu, Jiangbo Lu, and Guanbin Li. Nerf-hugs: Improved neural radiance fields in non-static scenes using heuristics-guided segmentation. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19436-19446, 2024. 2, 3", + "[5] Xingyu Chen, Qi Zhang, Xiaoyu Li, Yue Chen, Feng Ying, Xuan Wang, and Jue Wang. Hallucinated neural radiance fields in the wild. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12933-12942, 2021. 3", + "[6] Zilong Chen, Feng Wang, and Huaping Liu. Text-to-3d using gaussian splatting. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21401–21412, 2023. 1", + "[7] Hiba Dahmani. Swag: Splitting in the wild images with appearance-conditioned gaussians. In ECCV, 2024. 3", + "[8] Sara Fridovich-Keil, Giacomo Meanti, Frederik Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12479-12488, 2023. 3", + "[9] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In SIGGRAPH 2024 Conference Papers. Association for Computing Machinery, 2024. 2, 3", + "[10] Karim Kassab, Antoine Schnepf, Jean-Yves Franceschi, Laurent Caraffa, Jeremie Mary, and Valérie Gouet-Brunet. Refinedfields: Radiance fields refinement for unconstrained scenes. arXiv preprint arXiv:2312.00639, 2023. 3", + "[11] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuehler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (TOG), 42:1-14, 2023. 1, 2, 6", + "[12] Bernhard Kerbl, Andreas Meuleman, Georgios Kopanas, Michael Wimmer, Alexandre Lanvin, and George Drettakis. A hierarchical 3d gaussian representation for real-time rendering of very large datasets. ACM Transactions on Graphics, 43(4), 2024. 2", + "[13] Jonávs. Kulhánek and Torsten Sattler. Tetra-nerf: Representing neural radiance fields using tetrahedra. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 18412–18423, 2023. 2", + "[14] Jonas Kulhanek, Songyou Peng, Zuzana Kukelova, Marc" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Pollefeys, and Torsten Sattler. Wildgaussians: 3d gaussian splatting in the wild. 2024. 3, 5", + "[15] Pengzhi Li, Chengshuai Tang, Qinxuan Huang, and Zhiheng Li. Art3d: 3d gaussian splatting for text-guided artistic scenes generation, 2024. 1", + "[16] Zhihao Liang, Qi Zhang, Yingfa Feng, Ying Shan, and Kui Jia. Gs-ir: 3d gaussian splatting for inverse rendering. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21644–21653, 2023. 1", + "[17] Jiaqi Lin, Zhihao Li, Xiao Tang, Jianzhuang Liu, Shiyong Liu, Jiayue Liu, Yangdi Lu, Xiaofei Wu, Songcen Xu, Youliang Yan, and Wenming Yang. Vastgaussian: Vast 3d gaussians for large scene reconstruction. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5166-5175, 2024. 2", + "[18] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20654-20664, 2024. 2", + "[19] Saswat Subhajyoti Mallick, Rahul Goel, Bernhard Kerbl, Markus Steinberger, Francisco Vicente Carrasco, and Fernando De La Torre. Taming 3dgs: High-quality radiance fields with limited resources. In SIGGRAPH Asia 2024 Conference Papers, New York, NY, USA, 2024. Association for Computing Machinery. 2, 5", + "[20] Ricardo Martin-Brualla, Noha Radwan, Mehdi S. M. Sajjadi, Jonathan T. Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7206-7215, 2020. 2, 3", + "[21] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2", + "[22] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 2", + "[23] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINOV2: Learning robust visual features without supervision. Transactions on Machine Learning Research, 2024. 2", + "[24] Konstantinos Rematas, Andrew Liu, Pratul P. Srinivasan, Jonathan T. Barron, Andrea Tagliasacchi, Tom Funkhouser, and Vittorio Ferrari. Urban radiance fields. CVPR, 2022. 2", + "[25] Weining Ren, Zihan Zhu, Boyang Sun, Jiaqi Chen, Marc Pollefeys, and Songyou Peng. Nerf on-the-go: Exploiting uncertainty for distractor-free nerfs in the wild. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 3, 5, 6, 8" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "796", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Barbara Roessle, Norman Müller, Lorenzo Porzi, Samuel Rota Bulò, Peter Kontschieder, Angela Dai, and Matthias Nießner. L3dg: Latent 3d gaussian diffusion. In SIGGRAPH Asia 2024 Conference Papers, 2024. 1", + "[27] Sara Sabour, Suhani Vora, Daniel Duckworth, Ivan Krasin, David J. Fleet, and Andrea Tagliasacchi. Robustnerf: Ignoring distractors with robust losses. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20626-20636, 2023. 1, 2, 3, 5, 6, 8", + "[28] Sara Sabour, Lily Goli, George Kopanas, Mark Matthews, Dmitry Lagun, Leonidas Guibas, Alec Jacobson, David J. Fleet, and Andrea Tagliasacchi. SpotLessSplats: Ignoring distractors in 3d gaussian splatting. arXiv:2406.20055, 2024. 2, 3, 6, 7", + "[29] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 4", + "[30] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In European Conference on Computer Vision (ECCV), 2016. 4", + "[31] Jin-Chuan Shi, Miao Wang, Hao-Bin Duan, and Shao-Hua Guan. Language embedded 3d gaussians for open-vocabulary scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5333-5343, 2024. 1", + "[32] Noah Snavely, Steven M. Seitz, and Richard Szeliski. Photo tourism: Exploring photo collections in 3d. In SIGGRAPH Conference Proceedings, pages 835-846, New York, NY, USA, 2006. ACM Press. 8", + "[33] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P. Srinivasan, Jonathan T. Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8238–8248, 2022. 2", + "[34] Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Justin Kerr, Terrance Wang, Alexander Kristoffersen, Jake Austin, Kamyar Salahi, Abhik Ahuja, David McAllister, and Angjoo Kanazawa. Nerfstudio: A modular framework for neural radiance field development. ACM SIGGRAPH 2023 Conference Proceedings, 2023. 2", + "[35] Linhan Wang, Kai Cheng, Shuo Lei, Shengkun Wang, Wei Yin, Chenyang Lei, Xiaoxiao Long, and Chang-Tien Lu. Dcgaussian: Improving 3d gaussian splatting for reflective dash cam videos. NeurIPS 2024, 2024. 1", + "[36] Shuaixian Wang, Haoran Xu, Yaokun Li, Jiwei Chen, and Guang Tan. Ie-nerf: Inpainting enhanced neural radiance fields in the wild. ArXiv, abs/2407.10695, 2024. 3", + "[37] Yuze Wang, Junyi Wang, and Yue Qi. We-gs: An in-the-wild efficient 3d gaussian representation for unconstrained photo collections, 2024. 3", + "[38] Zhou Wang, Alan Conrad Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13:600-612, 2004. 5, 6", + "[39] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang." + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20310-20320, 2024. 1", + "[40] Tianhao Walter Wu, Fangcheng Zhong, Andrea Tagliasacchi, Forrester Cole, and Cengiz Oztireli. D $^2$ neRF: Self-supervised decoupling of dynamic and static objects from a monocular video. In Advances in Neural Information Processing Systems, 2022. 2", + "[41] Congrong Xu, Justin Kerr, and Angjoo Kanazawa. Splatfacto-w: A nefstudio implementation of gaussian splatting for unconstrained photo collections. ArXiv, abs/2407.12306, 2024. 3", + "[42] Yifan Yang, Shuhai Zhang, Zixiong Huang, Yubing Zhang, and Mingkui Tan. Cross-ray neural radiance fields for novel-view synthesis from unconstrained image collections. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 15855-15865, 2023. 3", + "[43] Vickie Ye, Ruilong Li, Justin Kerr, Matias Turkulainen, Brent Yi, Zhuoyang Pan, Otto Seiskari, Jianbo Ye, Jeffrey Hu, Matthew Tancik, and Angjoo Kanazawa. gsplat: An open-source library for Gaussian splatting. arXiv preprint arXiv:2409.06765, 2024. 5", + "[44] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19447-19456, 2024. 2", + "[45] Zehao Yu, Torsten Sattler, and Andreas Geiger. Gaussian opacity fields: Efficient adaptive surface reconstruction in unbounded scenes. ACM Transactions on Graphics, 2024. 2", + "[46] Dongbin Zhang, Chuming Wang, Weitao Wang, Peihao Li, Minghan Qin, and Haoqian Wang. Gaussian in the wild: 3d gaussian splatting for unconstrained image collections. ECCV, 2024. 3", + "[47] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 586-595, 2018. 6", + "[48] Xinjie Zhang, Xingtong Ge, Tongda Xu, Dailan He, Yan Wang, Hongwei Qin, Guo Lu, Jing Geng, and Jun Zhang. Gaussianimage: 1000 fps image representation and compression by 2d gaussian splatting. In European Conference on Computer Vision, 2024. 3", + "[49] Yunxiang Zhang, Alexandr Kuznetsov, Akshay Jindal, Kenneth Chen, Anton Sochenov, Anton Kaplanyan, and Qi Sun. Image-gs: Content-adaptive image representation via 2d gaussians, 2024. 3", + "[50] Xiaoyu Zhou, Zhiwei Lin, Xiaojun Shan, Yongtao Wang, Deqing Sun, and Ming-Hsuan Yang. Drivinggaussian: Composite gaussian splatting for surrounding dynamic autonomous driving scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21634-21643, 2024. 1" + ], + "bbox": [ + 516, + 92, + 903, + 867 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "797", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/HybridGS_ Decoupling Transients and Statics with 2D and 3D Gaussian Splatting/347531dd-b02e-491d-be81-0620c6e2bfcd_model.json b/2025/HybridGS_ Decoupling Transients and Statics with 2D and 3D Gaussian Splatting/347531dd-b02e-491d-be81-0620c6e2bfcd_model.json new file mode 100644 index 0000000000000000000000000000000000000000..632df1144c1a4f8c879ae347a62bcc90b65c5d7a --- /dev/null +++ b/2025/HybridGS_ Decoupling Transients and Statics with 2D and 3D Gaussian Splatting/347531dd-b02e-491d-be81-0620c6e2bfcd_model.json @@ -0,0 +1,2200 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.268, + 0.131, + 0.73, + 0.176 + ], + "angle": 0, + "content": "HybridGS: Decoupling Transients and Statics with 2D and 3D Gaussian Splitting" + }, + { + "type": "text", + "bbox": [ + 0.213, + 0.202, + 0.778, + 0.24 + ], + "angle": 0, + "content": "Jingyu Lin\\(^{1\\dagger}\\) Jiaqi Gu\\(^{2\\dagger}\\) Lubin Fan\\(^{2\\dagger}\\) Bojian Wu\\(^{2}\\) Yujing Lou\\(^{4}\\) Renjie Chen\\(^{1\\dagger}\\) Ligang Liu\\(^{1}\\) Jieping Ye\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.171, + 0.241, + 0.828, + 0.258 + ], + "angle": 0, + "content": "1University of Science and Technology of China 2Independent Researcher 4Shanghai Jiao Tong University" + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.266, + 0.315, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.334, + 0.267, + 0.475, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.478, + 0.268, + 0.619, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.621, + 0.268, + 0.762, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.764, + 0.269, + 0.905, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.45, + 0.908, + 0.479 + ], + "angle": 0, + "content": "Figure 1. HybridGS is the first hybrid representation that combines multi-view consistent 3D Gaussians and single-view independent 2D Gaussians, which is used to decouple the transients and statics presented in the scene. Our results demonstrate reasonable decompositions." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.502, + 0.327, + 0.517 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.536, + 0.486, + 0.854 + ], + "angle": 0, + "content": "Generating high-quality novel view renderings of 3D Gaussian Splatting (3DGS) in scenes featuring transient objects is challenging. We propose a novel hybrid representation, termed as HybridGS, using 2D Gaussians for transient objects per image and maintaining traditional 3D Gaussians for the whole static scenes. 3DGS is suited for modeling static scenes that assume multi-view consistency, but the transient objects appear occasionally and do not adhere to the assumption, thus we model them as planar objects from a single view by 2D Gaussians. Our novel representation decomposes the scene from the perspective of fundamental viewpoint consistency. Additionally, we present a multiview supervision method for 3DGS that leverages information from co-visible regions, further enhancing the distinctions between the transients and statics. Then, we propose a straightforward yet effective multi-stage training strategy to ensure robust training and view synthesis. Experiments on benchmarks show our state-of-the-art performance of novel view synthesis in indoor and outdoor scenes, even in the presence of distracting elements. Project page: https://gujiaqivadin.github.io/hybridgs/" + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.502, + 0.645, + 0.517 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.533, + 0.907, + 0.836 + ], + "angle": 0, + "content": "3D Gaussians Splating (3DGS) [11] has recently gained popularity for novel view synthesis due to its high-quality rendering, efficiency and low-memory cost. Its applications span across virtual reality, augmented reality, and robotics, etc [3, 6, 15, 16, 26, 31, 35, 39, 50]. Typically, most current approaches assume that the input images are posed and free of noise, which are the prerequisite needed to represent a complete scene well. However, this assumption is often not met. For example, images taken casually with a mobile phone usually contain messy dynamic objects, or transient objects. Therefore, it is hard to model the scene because there are always many transient occlusions. Technically, 3DGS is capable of modeling static scenes effectively because the static objects in images satisfy geometric consistency constraints across different viewpoints. However, transient objects do not follow this assumption. If the images containing transient objects are naively used in training 3DGS, such variations can lead to difficulties in achieving coherent blending of individual Gaussian representations, potentially resulting in artifacts or inaccuracies." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.901 + ], + "angle": 0, + "content": "To address this issue, numerous methods have been proposed to improve the scene representation ability involving transient objects. In the era of Neural Radiance Fields (NeRF) [21], RobustNeRF [27] introduced a robust loss" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.359, + 0.901 + ], + "angle": 0, + "content": "†Equal contributions. ‡Corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "788" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.317 + ], + "angle": 0, + "content": "function that minimizes the influence of photometrically inconsistent observations during training, thereby enhancing the overall quality of the representation. NeRF On-the-go [25] was the first to leverage the uncertainty of transients by incorporating DINOv2 [23] features into a shallow MLP. It focused on loss supervision in the areas of low uncertainty to effectively remove transients from dynamic scenes. These two methods have established widely-used benchmark datasets featuring transients, aiming to evaluate the reconstruction of clear static scenes with improved results of novel view synthesis. For 3DGS, SpotLessSplats (SLS-mlp) [28] integrated the designs of those aforementioned methods, employing adaptive method to detect outliers through clustering in the feature space, achieving the best current results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.32, + 0.482, + 0.455 + ], + "angle": 0, + "content": "However, these previous approaches neglect two critical issues: (1) The ambiguity between semantics and transients. Utilizing features from visual models to capture the semantic characteristics of transients can inherently lead to confusion in distinguishing between them. (2) The absence of explicit transient modeling. These methods primarily focus on achieving a robust statics by merely mitigating the impact of transients, rather than comprehensively modeling the whole scene themselves." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.458, + 0.483, + 0.849 + ], + "angle": 0, + "content": "In response to the issues mentioned above, our key observation is that transients lack multi-view consistency and usually only appear in a single view. Therefore, we consider them as planar objects at that view. That is to say, this planar representation is a unique feature of the current view. Based on this, we propose HybridGS, a hybrid representation with 2D Gaussians for transient objects per image and 3D Gaussians for the whole static scenes as illustrated in Fig. 1, which naturally decouples the transients and statics based on their intrinsic properties. With this novel representation, while modeling transients as planar objects, the fundamental consistency of viewpoints is also maintained for statics. To further enhance the performance, a multiview regulated supervision method for 3DGS is adopted to incorporates information from co-visible regions across multiple views, leading to more accurate representation. Additionally, we employ a simple yet effective multi-stage training strategy that ensures robust training and consistent view synthesis across various scenarios, thereby guaranteeing better stability and convergence during the training process. Extensive experiments on benchmark datasets demonstrate our state-of-the-art quality of novel view synthesis in both indoor and outdoor scenes, leading to the accurate decoupling of transients and statics as well as high efficiency in training and inference stage. The contributions are summarized as follows." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "- We are the first to introduce a novel hybrid representation that combines image-specific 2D Gaussians with static 3D Gaussians, enabling effective modeling of transient" + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.092, + 0.796, + 0.107 + ], + "angle": 0, + "content": "objects within casually captured images." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.107, + 0.905, + 0.182 + ], + "angle": 0, + "content": "- We develop a multi-view supervision scheme for 3DGS that utilizes overlapping regions across multiple views, which enhances the model's capability to distinguish between static and transient elements, ultimately improving the overall quality of the novel view synthesis." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.183, + 0.905, + 0.242 + ], + "angle": 0, + "content": "- Our HybridGS achieves the state-of-the-art performance in benchmark datasets, which outperforms the previous methods and sets a new standard for novel view synthesis in the scenes with transients." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.107, + 0.905, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.258, + 0.655, + 0.273 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.283, + 0.713, + 0.299 + ], + "angle": 0, + "content": "2.1. Novel View Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.307, + 0.905, + 0.518 + ], + "angle": 0, + "content": "NeRF (Neural Radiance Fields) [1, 2, 13, 17, 21, 22, 24, 33, 34] revolutionizes 3D scene reconstruction by representing scenes as continuous neural networks that map 3D coordinates to color and density values, which allows for the synthesis of photorealistic views from arbitrary camera positions. Recently, 3D Gaussian Splatting (3DGS) [9, 11, 12, 18, 19, 44, 45] has gained attention as an effective method for novel view synthesis. It starts with an initial point cloud and transforms it into estimizable 3D Gaussian primitives, known for their high-quality rendering and fast processing speeds. Nevertheless, existing methods require the assumption that only static scenes are present in the input images to effectively utilize multi-view geometric consistency to train a coherent scene representation." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.529, + 0.905, + 0.545 + ], + "angle": 0, + "content": "2.2. Modeling Transients and Statics in NeRF/3DGS" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.553, + 0.905, + 0.809 + ], + "angle": 0, + "content": "In casually captured footage of real-world scenes, there are often numerous dynamic occlusions, such as moving pedestrians and vehicles. Many works have tried to address challenges in scenes featuring both static and transient objects. RobustNeRF [27] was the pioneer work that introduces a robust loss function that reduces the influence of photometrical inconsistency during training, thereby improving the overall reconstruction quality on its proposed dataset. Other methods [4, 20, 40] also applied the similar idea to strengthen the robustness to outliers by down-weighting or discarding inconsistent observations based on the magnitude of color residuals. NeRF On-the-go was the first to introduce DINOV2 [23] features to predict pixel uncertainty, in an attempt to achieve clean static reconstructions without transients. It also introduced a benchmark dataset of casually captured images, aimed at evaluating static reconstruction performance for novel views." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.811, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Following these two methods and their datasets, recent methods of 3DGS have also explored techniques for separating transients and statics to reconstruct the statics more effectively. For example, SpotLessSplats [28] further extended previous methods by utilizing semantic features [23] combined with robust optimization strategies to" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "789" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "effectively eliminate transient distractors. Besides, several approaches [4, 5, 7, 8, 10, 14, 20, 36, 37, 41, 42, 46] have focused on reconstructing statics and illumination in datasets with unconstrained photo collections, that typically employ appearance embedding vectors to model the unique lighting and styles of each image, allowing for more accurate static reconstruction despite varying illumination status." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.198, + 0.482, + 0.303 + ], + "angle": 0, + "content": "Our method specifically focuses on distinguishing between transients and statics in casually captured unordered images with minimal illumination changes. Different from previous methods [25, 27, 28], we propose a hybrid Gaussian representation with 2D and 3D Gaussians with multiview information to model the whole 3D scene without any additional semantic features." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.312, + 0.443, + 0.327 + ], + "angle": 0, + "content": "2.3. Image Representation with 2D Gaussians" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.335, + 0.483, + 0.47 + ], + "angle": 0, + "content": "Recently, inspired by 3DGS, some methods [48, 49] in the field of image representation and compression have started to use 2D Gaussians to represent images. GaussianImage [48] leverages 2D Gaussians on the pixel plane to achieve comparable reconstruction quality while offering enhanced compression and accelerated rendering capabilities. Image-GS [49] creates a content-adaptive image representation by fitting a target image by adaptively allocating and progressively optimizing a set of 2D Gaussians." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.471, + 0.483, + 0.576 + ], + "angle": 0, + "content": "Note that the 2D Gaussians used here differ from the 2D Surfels (also referred to as 2D Gaussians) introduced by [9]. While 2D Surfels are designed to enable perspective-correct splatting through 2D surface modeling, ray-splat intersection, and volumetric integration in 3D space, and the 2D Gaussians in GaussianImage [48] is a flexible, compact, and content-adaptive image representation within 2D space." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.577, + 0.482, + 0.622 + ], + "angle": 0, + "content": "Here, we incorporate the trainable 2D Gaussians to learn transient and uncertainty masks during the training process, decomposing the whole scene with 2D and 3D Gaussians." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.635, + 0.228, + 0.649 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.766 + ], + "angle": 0, + "content": "3D Gaussian Splatting (3DGS) represent a 3D scene explicitly with a set of 3D anisotropic Gaussians \\(\\{\\mathcal{G}_{3d}\\}\\). Each Gaussian is parameterized by its centroid \\(\\mathbf{x_{3d}}\\in \\mathbb{R}^3\\), scale \\(\\mathbf{S_{3d}}\\in \\mathbb{R}^3\\), rotation matrix \\(\\mathbf{R}\\in \\mathbb{R}^{3\\times 3}\\), opacity \\(\\alpha_{3d}\\in \\mathbb{R}\\) and color \\(\\mathbf{c_{3d}}\\in \\mathbb{R}^3\\) encoded in spherical harmonic (SH) coefficients. The 3D covariance matrix \\(\\Sigma_{3d}\\) of the 3D Gaussian is obtained from \\(\\mathbf{R}_{3d}\\) and \\(\\mathbf{S}_{3d}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.774, + 0.482, + 0.792 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\Sigma} _ {3 \\mathbf {d}} = \\mathbf {R} _ {3 \\mathbf {d}} \\mathbf {S} _ {3 \\mathbf {d}} \\mathbf {S} _ {3 \\mathbf {d}} ^ {T} \\mathbf {R} _ {3 \\mathbf {d}} ^ {T}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.802, + 0.483, + 0.832 + ], + "angle": 0, + "content": "The 3D Gaussians are defined in world space for a spatial point \\(\\mathbf{y}_{3d}\\) following:" + }, + { + "type": "equation", + "bbox": [ + 0.144, + 0.841, + 0.482, + 0.861 + ], + "angle": 0, + "content": "\\[\n\\mathcal {G} _ {3 d} \\left(\\mathbf {y} _ {3 \\mathrm {d}}\\right) = e ^ {- \\frac {1}{2} \\left(\\mathbf {y} _ {3 \\mathrm {d}} - \\mathbf {x} _ {3 \\mathrm {d}}\\right) ^ {T} \\boldsymbol {\\Sigma} _ {3 \\mathrm {d}} ^ {- 1} \\left(\\mathbf {y} _ {3 \\mathrm {d}} - \\mathbf {x} _ {3 \\mathrm {d}}\\right)}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Given \\(\\mathbf{J}\\), the Jacobian of the affine projective transformation and \\(\\mathbf{W}\\), a viewing transformation, the covariance ma" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.864, + 0.108 + ], + "angle": 0, + "content": "trix \\(\\Sigma_{\\mathbf{3d}}^{\\prime}\\) in camera coordinates is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.624, + 0.118, + 0.905, + 0.136 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\Sigma} _ {3 \\mathbf {d}} ^ {\\prime} = \\mathbf {J} \\mathbf {W} \\boldsymbol {\\Sigma} _ {3 \\mathbf {d}} \\mathbf {W} ^ {T} \\mathbf {J} ^ {T}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.149, + 0.905, + 0.21 + ], + "angle": 0, + "content": "These 3D Gaussians are projected to 2D imaging plane and blended through a fast, differentiable \\(\\alpha\\)-blending process to render 2D images. The color \\(\\mathbf{C}_{3d}\\) of each pixel is computed using N-ordered 2D splats with the following formula:" + }, + { + "type": "equation", + "bbox": [ + 0.584, + 0.222, + 0.905, + 0.264 + ], + "angle": 0, + "content": "\\[\n\\mathbf {C} _ {\\mathbf {3 d}} = \\sum_ {i \\in \\mathbf {N}} \\alpha_ {\\mathbf {3 d} _ {i} ^ {\\prime}} \\mathbf {c} _ {\\mathbf {3 d} _ {i}} \\prod_ {j = 1} ^ {i - 1} \\left(1 - \\alpha_ {\\mathbf {3 d} _ {j} ^ {\\prime}}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.276, + 0.905, + 0.354 + ], + "angle": 0, + "content": "where \\(\\alpha_{\\mathbf{3d}_i'}\\) is the final opacity calculated by \\(\\alpha_{\\mathbf{3d}_i'} = \\alpha_{\\mathbf{3d}_i} * e^{-\\frac{1}{2} (\\mathbf{y}_{\\mathbf{3d}}' - \\mathbf{x}_{\\mathbf{3d}}')^T \\pmb{\\Sigma}_{\\mathbf{3d}}'^{-1} (\\mathbf{y}_{\\mathbf{3d}}' - \\mathbf{x}_{\\mathbf{3d}}')}\\). And \\(\\mathbf{x}_{\\mathbf{3d}}'\\) and \\(\\mathbf{y}_{\\mathbf{3d}}'\\) are coordinates in the projected space. Typically, the centroids of 3D Gaussians are initialized using the sparse SfM point clouds obtained from the set of input images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.355, + 0.907, + 0.506 + ], + "angle": 0, + "content": "2D Gaussian Splating (2DGS) is initially introduced by [48] in the task of implicit image representation. Since 2D Gaussians \\(\\{\\mathcal{G}_{2d}\\}\\) are no longer oriented to 3D scenes but to 2D space, many bloated operations and redundant parameters are discarded, such as viewing transformation, spherical harmonics, etc. Concretely, a basic 2D Gaussian is described by its centroid \\(\\mathbf{x_{2d}}\\in \\mathbb{R}^2\\) , 2D covariance matrix \\(\\Sigma_{2\\mathbf{d}}\\in \\mathbb{R}^2\\) , color \\(\\mathbf{c_{2d}}\\in \\mathbb{R}^3\\) and opacity \\(\\alpha_{2\\mathrm{d}}\\in \\mathbb{R}\\) . The 2D covariance matrix is also factorized into a rotation matrix \\(\\mathbf{R}_{2\\mathbf{d}}\\in \\mathbb{R}^{2\\times 2}\\) and a scale matrix \\(\\mathbf{S}_{2\\mathbf{d}}\\in \\mathbb{R}^2\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.611, + 0.517, + 0.905, + 0.535 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\Sigma} _ {2 \\mathrm {d}} = \\mathbf {R} _ {2 \\mathrm {d}} \\mathbf {S} _ {2 \\mathrm {d}} \\mathbf {S} _ {2 \\mathrm {d}} ^ {T} \\mathbf {R} _ {2 \\mathrm {d}} ^ {T}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.548, + 0.905, + 0.593 + ], + "angle": 0, + "content": "For rasterization, since the depth and camera parameters are not included in 2D Gaussians, a simplified accumulated summation of N-ordered 2D splats is calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.634, + 0.607, + 0.905, + 0.638 + ], + "angle": 0, + "content": "\\[\n\\mathbf {C} _ {2 \\mathbf {d}} = \\sum_ {i \\in \\mathbf {N}} \\alpha_ {\\mathbf {2 d} _ {i} ^ {\\prime}} \\mathbf {c} _ {2 \\mathbf {d} _ {i}}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.646, + 0.905, + 0.694 + ], + "angle": 0, + "content": "where \\(\\alpha_{\\mathbf{2d}_i'}\\) is the opacity calculated by \\(\\alpha_{\\mathbf{2d}_i'} = \\alpha_{\\mathbf{2d}_i} * e^{-\\frac{1}{2} (\\mathbf{y}_{\\mathbf{2d}} - \\mathbf{x}_{\\mathbf{2d}})^T \\boldsymbol{\\Sigma}_{\\mathbf{2d}}^{-1} (\\mathbf{y}_{\\mathbf{2d}} - \\mathbf{x}_{\\mathbf{2d}})}\\). And \\(\\mathbf{x}_{\\mathbf{2d}}\\) and \\(\\mathbf{y}_{\\mathbf{2d}}\\) are coordinates of centroid and any pixel in the 2D image space." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.695, + 0.907, + 0.815 + ], + "angle": 0, + "content": "Specifically, 2D Gaussians can be considered as projections of 3D Gaussians from a certain perspective, illustrating the relationship of Gaussians between 2D and 3D space. In our case, we combine the single-view independent 2D Gaussians and multi-view consistent 3D Gaussians to represent the entire 3D scene, which helps fundamentally decoupling transients and statics, based on the principles of multi-view geometry." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.83, + 0.605, + 0.845 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "We delineate the pipeline of our approach as illustrated in Fig. 2. Essentially, given a set of input images \\(\\{I_k|k = 1,2,\\dots,N\\}\\) with the corresponding camera parameters, for" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "790" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.089, + 0.904, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.289, + 0.908, + 0.361 + ], + "angle": 0, + "content": "Figure 2. Overview. Given a casually captured image sequence, we decompose the whole scene into 2D Gaussians for transient objects and 3D Gaussians for static scenes. To warm up, we start by training a basic 3DGS to capture static elements. This is followed by iterative training of 2D and 3D Gaussians, where our transients and statics are combined using an \\(\\alpha\\)-blending strategy with masks to produce the final renderings. The masks provide guidance for 3D Gaussians in the iterative training stage. During the joint-training, both 2D and 3D Gaussians are trained to further optimize the decomposition results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.386, + 0.483, + 0.418 + ], + "angle": 0, + "content": "each view \\(I\\), the goal of our method is to reasonably decouple the transients \\(I_{t}\\) and statics \\(I_{s}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.181, + 0.429, + 0.483, + 0.447 + ], + "angle": 0, + "content": "\\[\nI = M _ {t} \\odot I _ {t} + (1 - M _ {t}) \\odot I _ {s}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.458, + 0.484, + 0.534 + ], + "angle": 0, + "content": "where the \\(M_t \\in [0,1]\\) represents the transient mask, and \\(\\odot\\) is per pixel multiplication. When a pixel's value in \\(M_t\\) approaches 1, it indicates a higher probability that the location is transient; conversely, a lower value suggests a greater likelihood that area is static." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.535, + 0.483, + 0.745 + ], + "angle": 0, + "content": "To achieve the goal, we decompose the whole 3D scene with two components: (1) Multi-view consistent 3D Gaussians is used for rendering \\( I_{s} \\), which leverages multi-view information from images and models the static scene with a set of unified 3D Gaussians. The multi-view inputs regulate the 3D Gaussians to be consistent and robust across the different views. (2) Single-view independent 2D Gaussians is responsible for modeling \\( I_{t} \\), which enables our approach to handle the fact that images are casually captured with varying transients. Concretely, we forms a set of view-independent 2D Gaussians to model transients as planar objects from a single view. This novel combination allows for a more precise and reasonable representation of 3D scenes, enabling better performance on novel views." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.746, + 0.484, + 0.807 + ], + "angle": 0, + "content": "This section is organized as follows: we introduce our hybrid representation of scenes using 3D Gaussians in Sec. 4.1 and 2D Gaussians in Sec. 4.2, followed by a comprehensive overview of the training pipeline in Sec. 4.3." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.817, + 0.405, + 0.833 + ], + "angle": 0, + "content": "4.1. Modeling Statics with 3D Gaussians" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Generally speaking, the initialized point clouds obtained by COLMAP [29, 30] can only represent areas in the scene that satisfy multi-view consistency, so 3D Gaussians tend to reconstruct static scenes, such as buildings, grounds, etc." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.386, + 0.905, + 0.492 + ], + "angle": 0, + "content": "However, in scenes with transients, continuously training the 3D Gaussians may cause the transient objects to be overfitted into the Gaussian kernels because of constraints imposed by the RGB loss. This approach fails to capture the true geometry and instead overemphasizes the RGB information, leading to annoying artifacts from other viewpoints and ultimately impacting quality of novel view synthesis." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.492, + 0.906, + 0.703 + ], + "angle": 0, + "content": "We address this problem by adopting an enhanced multiview regulated supervision scheme. Unlike previous methods that rasterize and supervise only a single image per iteration, our approach introduces two major changes. (1) We increase the number of images per iteration to \\( K \\), enabling gradient back-propagation to simultaneously consider mutual information from multiple views. This batch-wise inputs allow each optimization step to leverage multi-view insights to distinguish and infer transients and static elements. (2) We adopt a sparse training schedule that focuses optimization only on the 3D Gaussians within the co-visible areas of the cross-view frustums. This method not only sharpens the training focus but also reduces computational costs, as detailed in Alg. 1, and outputs rasterized \\( \\hat{I}_s \\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.712, + 0.855, + 0.728 + ], + "angle": 0, + "content": "4.2. Modeling Transients with 2D Gaussians" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Different from other methods that utilize semantic features, we introduce 2D Gaussians to model the transients as planar objects from a single view for each image. This approach is motivated by several key considerations: (1) 2DGS has been proven to be an efficient method for representing images, capable of modeling an arbitrary image effectively. (2) The formation and rasterization processes of 2DGS closely resemble those of 3DGS, allowing both to be expressed within a unified framework. (3) Degenerating 3D objects that do not satisfy multi-view geometric consistency into 2D representation can effectively decouple transients from static" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "791" + } + ], + [ + { + "type": "code_caption", + "bbox": [ + 0.103, + 0.095, + 0.429, + 0.11 + ], + "angle": 0, + "content": "Algorithm 1: Multi-view Regulated Supervision." + }, + { + "type": "algorithm", + "bbox": [ + 0.11, + 0.113, + 0.449, + 0.357 + ], + "angle": 0, + "content": "Input: A set of images \\(\\{I_k, k = 1, 2, \\dots, N\\}\\); Multi-view size \\(K\\); Maximum iterations \\(T\\); \nOutput: Optimized 3D Gaussian parameters \\(\\theta_{3d}\\) within the co-visible regions; \nInitialize \\(\\theta_{3d}\\); \nfor each iteration \\(t = 1\\) to \\(T\\) do \n\\[ I_{1,\\dots,K} \\leftarrow \\text{RandomSample}(K, N); \\quad F_{\\text{inter}} \\leftarrow \\text{ComputeFrustum}(I_1); \\quad \\text{for } i \\leftarrow 2 \\text{ to } K \\text{ do} \\quad |F_{\\text{inter}} \\leftarrow F_{\\text{inter}} \\cap \\text{ComputeFrustum}(I_i); \\quad \\text{end} \\quad \\mathcal{G}_{3d} \\leftarrow \\text{ExtractGaussians}(I, F_{\\text{inter}}); \\quad \\text{for each Gaussian } G_j \\in \\mathcal{G}_{3d} \\text{ do} \\quad |\\theta \\leftarrow \\theta - \\eta \\cdot \\nabla \\mathcal{L}(G_j(\\theta)); \\quad \\text{end} \\quad \\text{end}" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.391, + 0.295, + 0.404 + ], + "angle": 0, + "content": "scenes on a fundamental level." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.406, + 0.483, + 0.452 + ], + "angle": 0, + "content": "Specifically, 2D Gaussians are rasterized into an image \\(\\hat{I}_t\\) with a transient mask \\(\\hat{M}_t\\) for modeling uncertainty, each pixel \\(\\mathbf{y}_{2\\mathbf{d}}\\) of which is calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.465, + 0.483, + 0.481 + ], + "angle": 0, + "content": "\\[\n\\hat {M} _ {t} \\left(\\mathbf {y} _ {2 d}\\right) = \\sum_ {i \\in \\mathbf {N}} \\alpha_ {2 d i} ^ {\\prime} = \\sum_ {i \\in \\mathbf {N}} \\alpha_ {2 d i} \\cdot e ^ {- \\frac {1}{2} \\left(\\mathbf {y} _ {2 d} - \\mathbf {x} _ {2 d}\\right) ^ {T} \\boldsymbol {\\Sigma} _ {2 d} ^ {- 1} \\left(\\mathbf {y} _ {2 d} - \\mathbf {x} _ {2 d}\\right)}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.493, + 0.484, + 0.584 + ], + "angle": 0, + "content": "It is noteworthy that we adopt uncertainty of 2D Gaussians to model the mask, rather than using 3D results. Because after modeling a static scene with 3D Gaussians, the 2D Gaussians are responsible for learning the residual part of the image. Therefore, naturally, the results rendered by the 2D Gaussian implicitly include a mask for transients." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.594, + 0.35, + 0.611 + ], + "angle": 0, + "content": "4.3. Multi-stage Training Scheme" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.618, + 0.483, + 0.692 + ], + "angle": 0, + "content": "Because 2D and 3D Gaussians are trained within the same framework, it is crucial to balance their relationship effectively. To that end, we have proposed a multi-stage training strategy that ensures stable convergence and enhances overall performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.694, + 0.483, + 0.8 + ], + "angle": 0, + "content": "Warm up Pre-training. The training process begins by using 3DGS to capture the essential structure of the entire static scene. This stage focuses on modeling the scene with a set of unified 3D Gaussians with our multi-view regulated supervision as detailed in Sec. 4.1, producing foundational, albeit low-quality, rendered images \\(\\hat{I}_s\\) using a combination of DSSIM [38] and L1 losses following:" + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.811, + 0.483, + 0.83 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {w a r m u p} = \\lambda L _ {D S S I M} (\\hat {I} _ {s}, I) + (1 - \\lambda) L _ {1} (\\hat {I} _ {s}, I), \\quad (9)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.9 + ], + "angle": 0, + "content": "where the \\(\\lambda\\) is set to 0.2. During the warm-up, the densification of 3D Gaussians is conducted to gradually increase the number, allowing for a more detailed representation of the static scene structure." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.273 + ], + "angle": 0, + "content": "Iterative Training. The second stage employs iterative training between 2DGS and 3DGS to progressively optimize both components. During the alternation, the 3DGS will provide more accurate static scene renderings, which in turn enables the 2DGS to refine its transient representations \\(\\hat{I}_t\\) and mask predictions \\(\\hat{M}_t\\) with Eq. 10. Conversely, the improved masks from 2DGS allow the 3DGS to focus more precisely on rendering the static regions of the scene with Eq. 11. The rendered image \\(\\hat{I}\\) is a composition of \\(\\hat{I}_s\\) and \\(\\hat{I}_t\\) with \\(\\hat{M}_t\\) following Eq. 7. This process iteratively refines the transients and statics, improving the overall quality and accuracy of the masks. The loss functions are as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.283, + 0.907, + 0.3 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {i t e r 2 d}} = \\lambda L _ {\\text {D S S I M}} (\\hat {I}, I) + (1 - \\lambda) L _ {1} (\\hat {I}, I), \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.325, + 0.905, + 0.343 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {i t e r 3 d}} = \\left(1 - \\hat {M} _ {t}\\right) \\odot \\mathcal {L} _ {\\text {i t e r 2 d}}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.349, + 0.905, + 0.393 + ], + "angle": 0, + "content": "Note that, during training, 2DGS and 3DGS are trained alternately. When training one branch, the gradient backpropagation of the other branch is turned off." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.394, + 0.906, + 0.484 + ], + "angle": 0, + "content": "Joint Fine-tuning. The final stage involves joint fine-tuning, integrating 2DGS and 3DGS into a unified framework. This stage focuses on fine-tuning on the transient mask and reducing the error between the rendered image \\(\\hat{I}\\) and the ground truth image \\(I\\). The loss \\(\\mathcal{L}_{joint}\\) used during the joint fine-tuning is a combination as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.534, + 0.495, + 0.905, + 0.513 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {j o i n t}} = \\beta L _ {D S S I M} (\\hat {I}, I) + (1 - \\beta) L _ {1} (\\hat {I}, I), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.522, + 0.906, + 0.569 + ], + "angle": 0, + "content": "where the \\(\\beta\\) is set to 0.2. Here, both the 2DGS and 3DGS branches will enable gradient backpropagation to simultaneously update all parameters." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.581, + 0.646, + 0.598 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.606, + 0.704, + 0.623 + ], + "angle": 0, + "content": "5.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.907, + 0.81 + ], + "angle": 0, + "content": "Datasets. Similar to previous methods, we evaluate our HybridGS on two challenging datasets: NeRF On-the-go [25] and RobustNeRF [27]. NeRF On-the-go contains multiple casually captured indoor and outdoor sequences with varying ratios of distractors (from \\(5\\%\\) to over \\(30\\%\\)). We use the version of the dataset where all images were undistorted as [14]. The RobustNeRF dataset includes several scenes exemplifying different types of distractors. To simulate capturing over extended periods, distractor objects are repositioned between frames to interfere with statically placed objects (from 1 (Statue) to 150 (Yoda)). The splits include distractor-filled training set and distractor-free testing set." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Implementation Details. Our HybridGS, developed on top of open-source implementation gsplat [43] and Taming-3DGS [19], incorporates custom CUDA kernels, boosting the rasterization of \\(\\alpha\\)-blending for 3DGS and weighted summation for 2DGS. The number of views \\(K\\) is 4 in our practice. For a scene consisting of 100 images, the process" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "792" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 1. Comparison on NeRF On-the-go Dataset. For better visualization, the best results are highlighted in bold. Our method shows generally superior performance over state-of-the-art methods. * indicates that the results are reproduced from the official implementation." + }, + { + "type": "table", + "bbox": [ + 0.134, + 0.129, + 0.868, + 0.219 + ], + "angle": 0, + "content": "
MethodLow OcclusionMedium OcclusionHigh Occlusion
MountainFountainCornerPatioSpotPatio-High
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
RobustNeRF [27]17.540.4960.38315.650.3180.57623.040.7640.24420.390.7180.25120.650.6250.39120.540.5780.366
NeRF On-the-go [25]20.150.6440.25920.110.6090.31424.220.8060.19020.780.7540.21923.330.7870.18921.410.7180.235
3DGS [11]19.400.6380.21319.960.6590.18520.900.7130.24117.480.7040.19920.770.6930.31617.290.6040.363
SLS-mlp* [28]19.840.5800.31520.190.6120.25924.030.7950.16621.550.8380.32823.520.7560.23520.310.6640.261
HybridGS(Ours)21.730.6930.22521.110.6740.18825.030.8470.07721.980.8120.11024.330.7940.08621.770.7410.142
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.232, + 0.908, + 0.261 + ], + "angle": 0, + "content": "Table 2. Comparison on RobustNeRF Dataset. For better visualization, the best results are highlighted in bold. Our method significantly outperforms all baseline methods. Crab (1) scene had a test set with same set of views as those in the train set, which is fixed in Crab (2)." + }, + { + "type": "table", + "bbox": [ + 0.136, + 0.271, + 0.868, + 0.364 + ], + "angle": 0, + "content": "
MethodStatueAndroidYodaCrab (1)Crab (2)
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
RobustNeRF [27]20.600.760.1523.280.750.1329.780.820.1532.220.940.06---
NeRF On-the-go [25]21.580.770.2423.500.750.2129.960.830.24------
3DGS [11]21.020.810.1623.110.810.1326.330.910.1431.800.960.0829.74--
SLS-mlp [28]22.540.840.1325.050.850.0933.660.960.1035.850.970.0834.43--
HybridGS(Ours)22.930.870.1025.150.850.0735.320.960.0736.310.970.0535.170.960.08
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.388, + 0.483, + 0.418 + ], + "angle": 0, + "content": "Table 3. Ablation on components in datasets with varying occlusions. The best results are highlighted in bold." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.428, + 0.485, + 0.495 + ], + "angle": 0, + "content": "
Dataset(ratio)On-the-go low.(5%)On-the-go medium.(17%)On-the-go high.(26%)
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Ours21.420.6840.20623.500.8270.09223.050.7680.115
w/o iterative Training21.390.6870.20923.110.8150.09422.550.7530.123
w/o Joint Training21.050.6540.24123.180.8150.10723.010.7480.130
w/o Multi-view21.020.6580.23323.060.8120.10322.710.7630.111
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.533, + 0.483, + 0.638 + ], + "angle": 0, + "content": "starts with a 1k-step warm-up. This is followed by iterative training, which includes 10k steps for 2DGS and 1k steps for 3DGS. The final joint training involves 30k steps. The entire training process takes approximately 0.18 GPU hours. The number of training steps scales proportionally with the size of the dataset. All experiments are performed on a single NVIDIA RTX 4090 GPU using PyTorch." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.64, + 0.483, + 0.671 + ], + "angle": 0, + "content": "Metrics. We follow common practice by employing PSNR, SSIM [38], and LPIPS [47] for evaluation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.683, + 0.212, + 0.699 + ], + "angle": 0, + "content": "5.2. Evaluation" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.708, + 0.455, + 0.723 + ], + "angle": 0, + "content": "5.2.1 Comparison on the NeRF On-the-go Dataset" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.733, + 0.483, + 0.825 + ], + "angle": 0, + "content": "As shown in Tab. 1 and Fig. 3, our method demonstrates superior performance across different scenarios with varying levels of occlusion. In Tab. 1, our method outperforms the previous SOTA method across multiple metrics, with an average improvement of \\(1.10\\mathrm{dB}\\) in PSNR and \\(5.27\\%\\) in SSIM, demonstrating the superiority of our approach." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Furthermore, Fig. 3 provides a qualitative comparison, illustrating the visual fidelity of our results against other methods. Our method produces more clear and detailed images, closely resembling the ground truth, particularly in challenging scenes with complex geometry and occlusions." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.388, + 0.907, + 0.43 + ], + "angle": 0, + "content": "Table 4. Ablation on the number of 2D Gaussians of Corner in NeRF On-the-go dataset. For better visualization, the best results are highlighted in bold." + }, + { + "type": "table", + "bbox": [ + 0.597, + 0.441, + 0.828, + 0.516 + ], + "angle": 0, + "content": "
NumberPSNR↑SSIM↑LPIPS↓Storage (MB)
2DGS3DGS
1k24.4930.8400.0773.542.1
5k24.7330.8430.08217.341.9
10k25.0340.8470.08034.742.2
15k24.5120.8390.09242.652
25k24.6980.8400.09469.442.2
" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.532, + 0.907, + 0.563 + ], + "angle": 0, + "content": "This highlights our model's capability to effectively handle diverse visual conditions both in indoor and outdoor scenes." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.585, + 0.854, + 0.6 + ], + "angle": 0, + "content": "5.2.2 Comparison on the RobustNeRF Dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.609, + 0.907, + 0.837 + ], + "angle": 0, + "content": "We present the quantitative results of RobustNeRF in Tab. 2. Our method demonstrates superior performance quantitatively compared to all previous approaches. In scenes with numerous transients, such as the Yoda scene, we've noticed that the performance metrics of the original 3DGS significantly degrade, occasionally even dropping below those of NeRF-related methods. By leveraging multi-view shared information and incorporating 2D Gaussians to effectively represent transients, our approach surpasses semantic-based methods like SLS-mlp [28]. As shown in Fig. 4, our results have clearer overall boundaries and structured fine-grained details, while SLS-mlp [28] may lead to blurriness in static objects and occluded areas. This further confirms that, even when faced with numerous transients, our method consistently maintains stable and reliable performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.847, + 0.68, + 0.862 + ], + "angle": 0, + "content": "5.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Components. To assess the effectiveness of each component in our framework, we performed an ablation study, as" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "793" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.087, + 0.904, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.34, + 0.908, + 0.385 + ], + "angle": 0, + "content": "Figure 3. Visualization of novel view synthesis results on the testing set of NeRF On-the-go dataset. Our method demonstrates superior results by effectively reducing artifacts and providing clearer boundaries. This results in a cleaner statics compared to other methods, showcasing enhanced visual quality and precision in novel views." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.409, + 0.482, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.547, + 0.483, + 0.575 + ], + "angle": 0, + "content": "Figure 4. Visualization of novel view synthesis results on RobustNeRF dataset." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.61, + 0.483, + 0.852 + ], + "angle": 0, + "content": "presented in Tab. 3. We systematically removed iterative training, joint training, and multi-view supervision to evaluate their effectiveness. Each component, as expected, enhances overall performance, demonstrating their individual and combined significance. In scenes with low occlusion, multi-view supervision allows the model to concentrate on static elements by integrating data from various perspectives. Conversely, in highly occluded environments, joint training with 2D and 3D Gaussians markedly improves performance, highlighting the model's capability to manage complex scenes with significant occlusions and transient changes. Iterative training consistently boosts results across different occlusion levels. In summary, the results underscore that integrating diverse training strategies and multiview information significantly enhances the rendering quality of unseen views, especially in challenging conditions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "The Number of 2D Gaussians. Tab. 4 presents an ablation study on the number of 2D Gaussians used in the corner sequence of NeRF On-the-go dataset. The results indicate that" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.408, + 0.905, + 0.532 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.542, + 0.907, + 0.569 + ], + "angle": 0, + "content": "Figure 5. Comparison of transient masks on NeRF On-the-go dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.907, + 0.688 + ], + "angle": 0, + "content": "increasing the number of 2D Gaussians initially improves the performance. Specifically, 10k 2D Gaussians achieve the highest PSNR (25.034) and SSIM (0.847), while maintaining a competitive LPIPS score of 0.151 and reasonable storage requirements. Based on these findings, we select 10k as the optimal number in our experiments." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.909, + 0.854 + ], + "angle": 0, + "content": "Transients and Masks. Benefiting from our method's explicit modeling of transient objects, we can obtain transient masks without introducing any segmentation networks or other pre-trained features. Typically, these masks capture dynamic elements such as pedestrians and vehicles, as well as some structural details ignored by 3D Gaussians. In Fig. 6, our approach effectively learns the RGB and mask of transients even under strong occlusions, while maintaining the robustness of static elements. This capability underscores the strength of our method in discerning and isolating dynamic components within complex scenes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.901 + ], + "angle": 0, + "content": "We also compare our transient masks with SLS-mlp [28] semantic masks in Fig. 5. It is observed that the semantic mask does not perfectly align with the transients due to" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "794" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.096, + 0.09, + 0.904, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.466, + 0.908, + 0.509 + ], + "angle": 0, + "content": "Figure 6. Visualization of scene decomposition into transients and statics. Our method achieves superior transient mask separation in both indoor and outdoor scenes. It effectively separates transients and statics, and the resulting renderings closely resemble the ground truth images, demonstrating its effectiveness. Top two rows are from NeRF On-the-go [25], bottom two rows are from RobustNeRF [27]." + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.533, + 0.48, + 0.64 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.131, + 0.651, + 0.443, + 0.665 + ], + "angle": 0, + "content": "Figure 7. Visualization on Photo Tourism dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.696, + 0.483, + 0.788 + ], + "angle": 0, + "content": "the semantic features not being sufficiently precise. Additionally, transients often encompass low-level information, such as motion blur or object shadows, which may not explicitly exhibit clear semantic features. By modeling these transients using 2D Gaussians at the pixel level, our method allows for a more fine-grained capture of transients." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.801, + 0.217, + 0.815 + ], + "angle": 0, + "content": "5.4. Discussions" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "While our method has demonstrated strong performance in distinguishing between transients and statics, it has not yet accounted for the illumination variation in unconstrained photo collections. As illustrated in Fig. 7 on Photo Tourism [32], our method renders fine-grained architectural" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.535, + 0.907, + 0.595 + ], + "angle": 0, + "content": "structures, but the overall photometric result tends towards an average photometric effect, akin to overcast conditions. A promising direction is to incorporate an appearance embedding module in our method for the future work." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.616, + 0.634, + 0.632 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.907, + 0.903 + ], + "angle": 0, + "content": "In this paper, we introduce HybridGS, a novel hybrid representation with 2D and 3D Gaussians for decomposing 3D scenes from casually captured images. By introducing the hybrid Gaussian representation, our method addresses the challenges of handling transients as planar objects while preserving the integrity of the static scene. The multi-view supervision mechanism we developed plays a crucial role in enhancing the discrimination between static and transient elements across multiple views, reducing artifacts and inaccuracies in novel views. We have developed a multi-stage training strategy for the joint training of 2D and 3D Gaussians, ensuring the stable and robust training stage. Our experiments demonstrate that HybridGS significantly outperforms existing methods in terms of quality and efficiency. In summary, HybridGS establishes a novel, robust and efficient foundation for 3D scene representation, paving the way for advancements in various vision-based applications." + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "795" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.17 + ], + "angle": 0, + "content": "[1] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.172, + 0.484, + 0.213 + ], + "angle": 0, + "content": "[2] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.216, + 0.484, + 0.285 + ], + "angle": 0, + "content": "[3] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. Pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19457-19467, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.287, + 0.484, + 0.356 + ], + "angle": 0, + "content": "[4] Jiahao Chen, Yipeng Qin, Lingjie Liu, Jiangbo Lu, and Guanbin Li. Nerf-hugs: Improved neural radiance fields in non-static scenes using heuristics-guided segmentation. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19436-19446, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.359, + 0.484, + 0.426 + ], + "angle": 0, + "content": "[5] Xingyu Chen, Qi Zhang, Xiaoyu Li, Yue Chen, Feng Ying, Xuan Wang, and Jue Wang. Hallucinated neural radiance fields in the wild. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12933-12942, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.429, + 0.484, + 0.483 + ], + "angle": 0, + "content": "[6] Zilong Chen, Feng Wang, and Huaping Liu. Text-to-3d using gaussian splatting. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21401–21412, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.487, + 0.484, + 0.514 + ], + "angle": 0, + "content": "[7] Hiba Dahmani. Swag: Splitting in the wild images with appearance-conditioned gaussians. In ECCV, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.516, + 0.484, + 0.585 + ], + "angle": 0, + "content": "[8] Sara Fridovich-Keil, Giacomo Meanti, Frederik Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12479-12488, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.587, + 0.484, + 0.642 + ], + "angle": 0, + "content": "[9] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In SIGGRAPH 2024 Conference Papers. Association for Computing Machinery, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.645, + 0.483, + 0.699 + ], + "angle": 0, + "content": "[10] Karim Kassab, Antoine Schnepf, Jean-Yves Franceschi, Laurent Caraffa, Jeremie Mary, and Valérie Gouet-Brunet. Refinedfields: Radiance fields refinement for unconstrained scenes. arXiv preprint arXiv:2312.00639, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.702, + 0.483, + 0.757 + ], + "angle": 0, + "content": "[11] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuehler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (TOG), 42:1-14, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.759, + 0.483, + 0.828 + ], + "angle": 0, + "content": "[12] Bernhard Kerbl, Andreas Meuleman, Georgios Kopanas, Michael Wimmer, Alexandre Lanvin, and George Drettakis. A hierarchical 3d gaussian representation for real-time rendering of very large datasets. ACM Transactions on Graphics, 43(4), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.83, + 0.483, + 0.883 + ], + "angle": 0, + "content": "[13] Jonávs. Kulhánek and Torsten Sattler. Tetra-nerf: Representing neural radiance fields using tetrahedra. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 18412–18423, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.887, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[14] Jonas Kulhanek, Songyou Peng, Zuzana Kukelova, Marc" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Pollefeys, and Torsten Sattler. Wildgaussians: 3d gaussian splatting in the wild. 2024. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[15] Pengzhi Li, Chengshuai Tang, Qinxuan Huang, and Zhiheng Li. Art3d: 3d gaussian splatting for text-guided artistic scenes generation, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[16] Zhihao Liang, Qi Zhang, Yingfa Feng, Ying Shan, and Kui Jia. Gs-ir: 3d gaussian splatting for inverse rendering. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21644–21653, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.905, + 0.301 + ], + "angle": 0, + "content": "[17] Jiaqi Lin, Zhihao Li, Xiao Tang, Jianzhuang Liu, Shiyong Liu, Jiayue Liu, Yangdi Lu, Xiaofei Wu, Songcen Xu, Youliang Yan, and Wenming Yang. Vastgaussian: Vast 3d gaussians for large scene reconstruction. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5166-5175, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.303, + 0.907, + 0.37 + ], + "angle": 0, + "content": "[18] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20654-20664, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.372, + 0.905, + 0.455 + ], + "angle": 0, + "content": "[19] Saswat Subhajyoti Mallick, Rahul Goel, Bernhard Kerbl, Markus Steinberger, Francisco Vicente Carrasco, and Fernando De La Torre. Taming 3dgs: High-quality radiance fields with limited resources. In SIGGRAPH Asia 2024 Conference Papers, New York, NY, USA, 2024. Association for Computing Machinery. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.457, + 0.905, + 0.537 + ], + "angle": 0, + "content": "[20] Ricardo Martin-Brualla, Noha Radwan, Mehdi S. M. Sajjadi, Jonathan T. Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7206-7215, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.54, + 0.905, + 0.593 + ], + "angle": 0, + "content": "[21] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.596, + 0.905, + 0.649 + ], + "angle": 0, + "content": "[22] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.651, + 0.905, + 0.788 + ], + "angle": 0, + "content": "[23] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINOV2: Learning robust visual features without supervision. Transactions on Machine Learning Research, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[24] Konstantinos Rematas, Andrew Liu, Pratul P. Srinivasan, Jonathan T. Barron, Andrea Tagliasacchi, Tom Funkhouser, and Vittorio Ferrari. Urban radiance fields. CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[25] Weining Ren, Zihan Zhu, Boyang Sun, Jiaqi Chen, Marc Pollefeys, and Songyou Peng. Nerf on-the-go: Exploiting uncertainty for distractor-free nerfs in the wild. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 3, 5, 6, 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "796" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[26] Barbara Roessle, Norman Müller, Lorenzo Porzi, Samuel Rota Bulò, Peter Kontschieder, Angela Dai, and Matthias Nießner. L3dg: Latent 3d gaussian diffusion. In SIGGRAPH Asia 2024 Conference Papers, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[27] Sara Sabour, Suhani Vora, Daniel Duckworth, Ivan Krasin, David J. Fleet, and Andrea Tagliasacchi. Robustnerf: Ignoring distractors with robust losses. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20626-20636, 2023. 1, 2, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.219, + 0.482, + 0.286 + ], + "angle": 0, + "content": "[28] Sara Sabour, Lily Goli, George Kopanas, Mark Matthews, Dmitry Lagun, Leonidas Guibas, Alec Jacobson, David J. Fleet, and Andrea Tagliasacchi. SpotLessSplats: Ignoring distractors in 3d gaussian splatting. arXiv:2406.20055, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.288, + 0.482, + 0.329 + ], + "angle": 0, + "content": "[29] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.482, + 0.384 + ], + "angle": 0, + "content": "[30] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In European Conference on Computer Vision (ECCV), 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.386, + 0.482, + 0.454 + ], + "angle": 0, + "content": "[31] Jin-Chuan Shi, Miao Wang, Hao-Bin Duan, and Shao-Hua Guan. Language embedded 3d gaussians for open-vocabulary scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5333-5343, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.456, + 0.482, + 0.509 + ], + "angle": 0, + "content": "[32] Noah Snavely, Steven M. Seitz, and Richard Szeliski. Photo tourism: Exploring photo collections in 3d. In SIGGRAPH Conference Proceedings, pages 835-846, New York, NY, USA, 2006. ACM Press. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.482, + 0.592 + ], + "angle": 0, + "content": "[33] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P. Srinivasan, Jonathan T. Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8238–8248, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.482, + 0.677 + ], + "angle": 0, + "content": "[34] Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Justin Kerr, Terrance Wang, Alexander Kristoffersen, Jake Austin, Kamyar Salahi, Abhik Ahuja, David McAllister, and Angjoo Kanazawa. Nerfstudio: A modular framework for neural radiance field development. ACM SIGGRAPH 2023 Conference Proceedings, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.482, + 0.732 + ], + "angle": 0, + "content": "[35] Linhan Wang, Kai Cheng, Shuo Lei, Shengkun Wang, Wei Yin, Chenyang Lei, Xiaoxiao Long, and Chang-Tien Lu. Dcgaussian: Improving 3d gaussian splatting for reflective dash cam videos. NeurIPS 2024, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.734, + 0.482, + 0.774 + ], + "angle": 0, + "content": "[36] Shuaixian Wang, Haoran Xu, Yaokun Li, Jiwei Chen, and Guang Tan. Ie-nerf: Inpainting enhanced neural radiance fields in the wild. ArXiv, abs/2407.10695, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.482, + 0.816 + ], + "angle": 0, + "content": "[37] Yuze Wang, Junyi Wang, and Yue Qi. We-gs: An in-the-wild efficient 3d gaussian representation for unconstrained photo collections, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[38] Zhou Wang, Alan Conrad Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13:600-612, 2004. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[39] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20310-20320, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[40] Tianhao Walter Wu, Fangcheng Zhong, Andrea Tagliasacchi, Forrester Cole, and Cengiz Oztireli. D\\(^2\\)neRF: Self-supervised decoupling of dynamic and static objects from a monocular video. In Advances in Neural Information Processing Systems, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.905, + 0.274 + ], + "angle": 0, + "content": "[41] Congrong Xu, Justin Kerr, and Angjoo Kanazawa. Splatfacto-w: A nefstudio implementation of gaussian splatting for unconstrained photo collections. ArXiv, abs/2407.12306, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.905, + 0.345 + ], + "angle": 0, + "content": "[42] Yifan Yang, Shuhai Zhang, Zixiong Huang, Yubing Zhang, and Mingkui Tan. Cross-ray neural radiance fields for novel-view synthesis from unconstrained image collections. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 15855-15865, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.347, + 0.905, + 0.415 + ], + "angle": 0, + "content": "[43] Vickie Ye, Ruilong Li, Justin Kerr, Matias Turkulainen, Brent Yi, Zhuoyang Pan, Otto Seiskari, Jianbo Ye, Jeffrey Hu, Matthew Tancik, and Angjoo Kanazawa. gsplat: An open-source library for Gaussian splatting. arXiv preprint arXiv:2409.06765, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.418, + 0.905, + 0.485 + ], + "angle": 0, + "content": "[44] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19447-19456, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.488, + 0.905, + 0.53 + ], + "angle": 0, + "content": "[45] Zehao Yu, Torsten Sattler, and Andreas Geiger. Gaussian opacity fields: Efficient adaptive surface reconstruction in unbounded scenes. ACM Transactions on Graphics, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.531, + 0.905, + 0.585 + ], + "angle": 0, + "content": "[46] Dongbin Zhang, Chuming Wang, Weitao Wang, Peihao Li, Minghan Qin, and Haoqian Wang. Gaussian in the wild: 3d gaussian splatting for unconstrained image collections. ECCV, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.587, + 0.905, + 0.655 + ], + "angle": 0, + "content": "[47] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 586-595, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.657, + 0.905, + 0.727 + ], + "angle": 0, + "content": "[48] Xinjie Zhang, Xingtong Ge, Tongda Xu, Dailan He, Yan Wang, Hongwei Qin, Guo Lu, Jing Geng, and Jun Zhang. Gaussianimage: 1000 fps image representation and compression by 2d gaussian splatting. In European Conference on Computer Vision, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.728, + 0.905, + 0.783 + ], + "angle": 0, + "content": "[49] Yunxiang Zhang, Alexandr Kuznetsov, Akshay Jindal, Kenneth Chen, Anton Sochenov, Anton Kaplanyan, and Qi Sun. Image-gs: Content-adaptive image representation via 2d gaussians, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.785, + 0.905, + 0.868 + ], + "angle": 0, + "content": "[50] Xiaoyu Zhou, Zhiwei Lin, Xiaojun Shan, Yongtao Wang, Deqing Sun, and Ming-Hsuan Yang. Drivinggaussian: Composite gaussian splatting for surrounding dynamic autonomous driving scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21634-21643, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.868 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "797" + } + ] +] \ No newline at end of file diff --git a/2025/HybridGS_ Decoupling Transients and Statics with 2D and 3D Gaussian Splatting/347531dd-b02e-491d-be81-0620c6e2bfcd_origin.pdf b/2025/HybridGS_ Decoupling Transients and Statics with 2D and 3D Gaussian Splatting/347531dd-b02e-491d-be81-0620c6e2bfcd_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..341e765586da4c40b84d1fd511109fa40f582ce9 --- /dev/null +++ b/2025/HybridGS_ Decoupling Transients and Statics with 2D and 3D Gaussian Splatting/347531dd-b02e-491d-be81-0620c6e2bfcd_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4e8017dad034827599907b8cc889856274b15a771b3f02ac3421c04fe6dc9a22 +size 5459021 diff --git a/2025/HybridGS_ Decoupling Transients and Statics with 2D and 3D Gaussian Splatting/full.md b/2025/HybridGS_ Decoupling Transients and Statics with 2D and 3D Gaussian Splatting/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9e91f3610415531d9b8a42bb02ecdc54dc944169 --- /dev/null +++ b/2025/HybridGS_ Decoupling Transients and Statics with 2D and 3D Gaussian Splatting/full.md @@ -0,0 +1,330 @@ +# HybridGS: Decoupling Transients and Statics with 2D and 3D Gaussian Splitting + +Jingyu Lin $^{1\dagger}$ Jiaqi Gu $^{2\dagger}$ Lubin Fan $^{2\dagger}$ Bojian Wu $^{2}$ Yujing Lou $^{4}$ Renjie Chen $^{1\dagger}$ Ligang Liu $^{1}$ Jieping Ye $^{2}$ + +1University of Science and Technology of China 2Independent Researcher 4Shanghai Jiao Tong University + +![](images/041433ab006550dde174f0f360819f9503991771e505a0a0cd50cee3009bb315.jpg) +Figure 1. HybridGS is the first hybrid representation that combines multi-view consistent 3D Gaussians and single-view independent 2D Gaussians, which is used to decouple the transients and statics presented in the scene. Our results demonstrate reasonable decompositions. + +![](images/6464bab9e3de9e887ad3450894aea2244a1e6e60c81f408c52babe888a668ef5.jpg) + +![](images/f99fa44d342c780aba4ad3f673914bc8d58a8fc7dd7b5d4cb71239c52c34dd2b.jpg) + +![](images/39c629258ad6a6120adae51d206764b86889798fabec06f5f7bfa6986e6c8ddd.jpg) + +![](images/ff48fb684431d963da5c2f04ab53208fb965d3dfa5bb43cb6e53dc286dff494a.jpg) + +# Abstract + +Generating high-quality novel view renderings of 3D Gaussian Splatting (3DGS) in scenes featuring transient objects is challenging. We propose a novel hybrid representation, termed as HybridGS, using 2D Gaussians for transient objects per image and maintaining traditional 3D Gaussians for the whole static scenes. 3DGS is suited for modeling static scenes that assume multi-view consistency, but the transient objects appear occasionally and do not adhere to the assumption, thus we model them as planar objects from a single view by 2D Gaussians. Our novel representation decomposes the scene from the perspective of fundamental viewpoint consistency. Additionally, we present a multiview supervision method for 3DGS that leverages information from co-visible regions, further enhancing the distinctions between the transients and statics. Then, we propose a straightforward yet effective multi-stage training strategy to ensure robust training and view synthesis. Experiments on benchmarks show our state-of-the-art performance of novel view synthesis in indoor and outdoor scenes, even in the presence of distracting elements. Project page: https://gujiaqivadin.github.io/hybridgs/ + +# 1. Introduction + +3D Gaussians Splating (3DGS) [11] has recently gained popularity for novel view synthesis due to its high-quality rendering, efficiency and low-memory cost. Its applications span across virtual reality, augmented reality, and robotics, etc [3, 6, 15, 16, 26, 31, 35, 39, 50]. Typically, most current approaches assume that the input images are posed and free of noise, which are the prerequisite needed to represent a complete scene well. However, this assumption is often not met. For example, images taken casually with a mobile phone usually contain messy dynamic objects, or transient objects. Therefore, it is hard to model the scene because there are always many transient occlusions. Technically, 3DGS is capable of modeling static scenes effectively because the static objects in images satisfy geometric consistency constraints across different viewpoints. However, transient objects do not follow this assumption. If the images containing transient objects are naively used in training 3DGS, such variations can lead to difficulties in achieving coherent blending of individual Gaussian representations, potentially resulting in artifacts or inaccuracies. + +To address this issue, numerous methods have been proposed to improve the scene representation ability involving transient objects. In the era of Neural Radiance Fields (NeRF) [21], RobustNeRF [27] introduced a robust loss + +function that minimizes the influence of photometrically inconsistent observations during training, thereby enhancing the overall quality of the representation. NeRF On-the-go [25] was the first to leverage the uncertainty of transients by incorporating DINOv2 [23] features into a shallow MLP. It focused on loss supervision in the areas of low uncertainty to effectively remove transients from dynamic scenes. These two methods have established widely-used benchmark datasets featuring transients, aiming to evaluate the reconstruction of clear static scenes with improved results of novel view synthesis. For 3DGS, SpotLessSplats (SLS-mlp) [28] integrated the designs of those aforementioned methods, employing adaptive method to detect outliers through clustering in the feature space, achieving the best current results. + +However, these previous approaches neglect two critical issues: (1) The ambiguity between semantics and transients. Utilizing features from visual models to capture the semantic characteristics of transients can inherently lead to confusion in distinguishing between them. (2) The absence of explicit transient modeling. These methods primarily focus on achieving a robust statics by merely mitigating the impact of transients, rather than comprehensively modeling the whole scene themselves. + +In response to the issues mentioned above, our key observation is that transients lack multi-view consistency and usually only appear in a single view. Therefore, we consider them as planar objects at that view. That is to say, this planar representation is a unique feature of the current view. Based on this, we propose HybridGS, a hybrid representation with 2D Gaussians for transient objects per image and 3D Gaussians for the whole static scenes as illustrated in Fig. 1, which naturally decouples the transients and statics based on their intrinsic properties. With this novel representation, while modeling transients as planar objects, the fundamental consistency of viewpoints is also maintained for statics. To further enhance the performance, a multiview regulated supervision method for 3DGS is adopted to incorporates information from co-visible regions across multiple views, leading to more accurate representation. Additionally, we employ a simple yet effective multi-stage training strategy that ensures robust training and consistent view synthesis across various scenarios, thereby guaranteeing better stability and convergence during the training process. Extensive experiments on benchmark datasets demonstrate our state-of-the-art quality of novel view synthesis in both indoor and outdoor scenes, leading to the accurate decoupling of transients and statics as well as high efficiency in training and inference stage. The contributions are summarized as follows. + +- We are the first to introduce a novel hybrid representation that combines image-specific 2D Gaussians with static 3D Gaussians, enabling effective modeling of transient + +objects within casually captured images. + +- We develop a multi-view supervision scheme for 3DGS that utilizes overlapping regions across multiple views, which enhances the model's capability to distinguish between static and transient elements, ultimately improving the overall quality of the novel view synthesis. +- Our HybridGS achieves the state-of-the-art performance in benchmark datasets, which outperforms the previous methods and sets a new standard for novel view synthesis in the scenes with transients. + +# 2. Related Work + +# 2.1. Novel View Synthesis + +NeRF (Neural Radiance Fields) [1, 2, 13, 17, 21, 22, 24, 33, 34] revolutionizes 3D scene reconstruction by representing scenes as continuous neural networks that map 3D coordinates to color and density values, which allows for the synthesis of photorealistic views from arbitrary camera positions. Recently, 3D Gaussian Splatting (3DGS) [9, 11, 12, 18, 19, 44, 45] has gained attention as an effective method for novel view synthesis. It starts with an initial point cloud and transforms it into estimizable 3D Gaussian primitives, known for their high-quality rendering and fast processing speeds. Nevertheless, existing methods require the assumption that only static scenes are present in the input images to effectively utilize multi-view geometric consistency to train a coherent scene representation. + +# 2.2. Modeling Transients and Statics in NeRF/3DGS + +In casually captured footage of real-world scenes, there are often numerous dynamic occlusions, such as moving pedestrians and vehicles. Many works have tried to address challenges in scenes featuring both static and transient objects. RobustNeRF [27] was the pioneer work that introduces a robust loss function that reduces the influence of photometrical inconsistency during training, thereby improving the overall reconstruction quality on its proposed dataset. Other methods [4, 20, 40] also applied the similar idea to strengthen the robustness to outliers by down-weighting or discarding inconsistent observations based on the magnitude of color residuals. NeRF On-the-go was the first to introduce DINOV2 [23] features to predict pixel uncertainty, in an attempt to achieve clean static reconstructions without transients. It also introduced a benchmark dataset of casually captured images, aimed at evaluating static reconstruction performance for novel views. + +Following these two methods and their datasets, recent methods of 3DGS have also explored techniques for separating transients and statics to reconstruct the statics more effectively. For example, SpotLessSplats [28] further extended previous methods by utilizing semantic features [23] combined with robust optimization strategies to + +effectively eliminate transient distractors. Besides, several approaches [4, 5, 7, 8, 10, 14, 20, 36, 37, 41, 42, 46] have focused on reconstructing statics and illumination in datasets with unconstrained photo collections, that typically employ appearance embedding vectors to model the unique lighting and styles of each image, allowing for more accurate static reconstruction despite varying illumination status. + +Our method specifically focuses on distinguishing between transients and statics in casually captured unordered images with minimal illumination changes. Different from previous methods [25, 27, 28], we propose a hybrid Gaussian representation with 2D and 3D Gaussians with multiview information to model the whole 3D scene without any additional semantic features. + +# 2.3. Image Representation with 2D Gaussians + +Recently, inspired by 3DGS, some methods [48, 49] in the field of image representation and compression have started to use 2D Gaussians to represent images. GaussianImage [48] leverages 2D Gaussians on the pixel plane to achieve comparable reconstruction quality while offering enhanced compression and accelerated rendering capabilities. Image-GS [49] creates a content-adaptive image representation by fitting a target image by adaptively allocating and progressively optimizing a set of 2D Gaussians. + +Note that the 2D Gaussians used here differ from the 2D Surfels (also referred to as 2D Gaussians) introduced by [9]. While 2D Surfels are designed to enable perspective-correct splatting through 2D surface modeling, ray-splat intersection, and volumetric integration in 3D space, and the 2D Gaussians in GaussianImage [48] is a flexible, compact, and content-adaptive image representation within 2D space. + +Here, we incorporate the trainable 2D Gaussians to learn transient and uncertainty masks during the training process, decomposing the whole scene with 2D and 3D Gaussians. + +# 3. Preliminaries + +3D Gaussian Splatting (3DGS) represent a 3D scene explicitly with a set of 3D anisotropic Gaussians $\{\mathcal{G}_{3d}\}$ . Each Gaussian is parameterized by its centroid $\mathbf{x_{3d}}\in \mathbb{R}^3$ , scale $\mathbf{S_{3d}}\in \mathbb{R}^3$ , rotation matrix $\mathbf{R}\in \mathbb{R}^{3\times 3}$ , opacity $\alpha_{3d}\in \mathbb{R}$ and color $\mathbf{c_{3d}}\in \mathbb{R}^3$ encoded in spherical harmonic (SH) coefficients. The 3D covariance matrix $\Sigma_{3d}$ of the 3D Gaussian is obtained from $\mathbf{R}_{3d}$ and $\mathbf{S}_{3d}$ : + +$$ +\boldsymbol {\Sigma} _ {3 \mathbf {d}} = \mathbf {R} _ {3 \mathbf {d}} \mathbf {S} _ {3 \mathbf {d}} \mathbf {S} _ {3 \mathbf {d}} ^ {T} \mathbf {R} _ {3 \mathbf {d}} ^ {T}. \tag {1} +$$ + +The 3D Gaussians are defined in world space for a spatial point $\mathbf{y}_{3d}$ following: + +$$ +\mathcal {G} _ {3 d} \left(\mathbf {y} _ {3 \mathrm {d}}\right) = e ^ {- \frac {1}{2} \left(\mathbf {y} _ {3 \mathrm {d}} - \mathbf {x} _ {3 \mathrm {d}}\right) ^ {T} \boldsymbol {\Sigma} _ {3 \mathrm {d}} ^ {- 1} \left(\mathbf {y} _ {3 \mathrm {d}} - \mathbf {x} _ {3 \mathrm {d}}\right)}. \tag {2} +$$ + +Given $\mathbf{J}$ , the Jacobian of the affine projective transformation and $\mathbf{W}$ , a viewing transformation, the covariance ma + +trix $\Sigma_{\mathbf{3d}}^{\prime}$ in camera coordinates is defined as follows: + +$$ +\boldsymbol {\Sigma} _ {3 \mathbf {d}} ^ {\prime} = \mathbf {J} \mathbf {W} \boldsymbol {\Sigma} _ {3 \mathbf {d}} \mathbf {W} ^ {T} \mathbf {J} ^ {T}. \tag {3} +$$ + +These 3D Gaussians are projected to 2D imaging plane and blended through a fast, differentiable $\alpha$ -blending process to render 2D images. The color $\mathbf{C}_{3d}$ of each pixel is computed using N-ordered 2D splats with the following formula: + +$$ +\mathbf {C} _ {\mathbf {3 d}} = \sum_ {i \in \mathbf {N}} \alpha_ {\mathbf {3 d} _ {i} ^ {\prime}} \mathbf {c} _ {\mathbf {3 d} _ {i}} \prod_ {j = 1} ^ {i - 1} \left(1 - \alpha_ {\mathbf {3 d} _ {j} ^ {\prime}}\right), \tag {4} +$$ + +where $\alpha_{\mathbf{3d}_i'}$ is the final opacity calculated by $\alpha_{\mathbf{3d}_i'} = \alpha_{\mathbf{3d}_i} * e^{-\frac{1}{2} (\mathbf{y}_{\mathbf{3d}}' - \mathbf{x}_{\mathbf{3d}}')^T \pmb{\Sigma}_{\mathbf{3d}}'^{-1} (\mathbf{y}_{\mathbf{3d}}' - \mathbf{x}_{\mathbf{3d}}')}$ . And $\mathbf{x}_{\mathbf{3d}}'$ and $\mathbf{y}_{\mathbf{3d}}'$ are coordinates in the projected space. Typically, the centroids of 3D Gaussians are initialized using the sparse SfM point clouds obtained from the set of input images. + +2D Gaussian Splating (2DGS) is initially introduced by [48] in the task of implicit image representation. Since 2D Gaussians $\{\mathcal{G}_{2d}\}$ are no longer oriented to 3D scenes but to 2D space, many bloated operations and redundant parameters are discarded, such as viewing transformation, spherical harmonics, etc. Concretely, a basic 2D Gaussian is described by its centroid $\mathbf{x_{2d}}\in \mathbb{R}^2$ , 2D covariance matrix $\Sigma_{2\mathbf{d}}\in \mathbb{R}^2$ , color $\mathbf{c_{2d}}\in \mathbb{R}^3$ and opacity $\alpha_{2\mathrm{d}}\in \mathbb{R}$ . The 2D covariance matrix is also factorized into a rotation matrix $\mathbf{R}_{2\mathbf{d}}\in \mathbb{R}^{2\times 2}$ and a scale matrix $\mathbf{S}_{2\mathbf{d}}\in \mathbb{R}^2$ as: + +$$ +\boldsymbol {\Sigma} _ {2 \mathrm {d}} = \mathbf {R} _ {2 \mathrm {d}} \mathbf {S} _ {2 \mathrm {d}} \mathbf {S} _ {2 \mathrm {d}} ^ {T} \mathbf {R} _ {2 \mathrm {d}} ^ {T}. \tag {5} +$$ + +For rasterization, since the depth and camera parameters are not included in 2D Gaussians, a simplified accumulated summation of N-ordered 2D splats is calculated as follows: + +$$ +\mathbf {C} _ {2 \mathbf {d}} = \sum_ {i \in \mathbf {N}} \alpha_ {\mathbf {2 d} _ {i} ^ {\prime}} \mathbf {c} _ {2 \mathbf {d} _ {i}}, \tag {6} +$$ + +where $\alpha_{\mathbf{2d}_i'}$ is the opacity calculated by $\alpha_{\mathbf{2d}_i'} = \alpha_{\mathbf{2d}_i} * e^{-\frac{1}{2} (\mathbf{y}_{\mathbf{2d}} - \mathbf{x}_{\mathbf{2d}})^T \boldsymbol{\Sigma}_{\mathbf{2d}}^{-1} (\mathbf{y}_{\mathbf{2d}} - \mathbf{x}_{\mathbf{2d}})}$ . And $\mathbf{x}_{\mathbf{2d}}$ and $\mathbf{y}_{\mathbf{2d}}$ are coordinates of centroid and any pixel in the 2D image space. + +Specifically, 2D Gaussians can be considered as projections of 3D Gaussians from a certain perspective, illustrating the relationship of Gaussians between 2D and 3D space. In our case, we combine the single-view independent 2D Gaussians and multi-view consistent 3D Gaussians to represent the entire 3D scene, which helps fundamentally decoupling transients and statics, based on the principles of multi-view geometry. + +# 4. Method + +We delineate the pipeline of our approach as illustrated in Fig. 2. Essentially, given a set of input images $\{I_k|k = 1,2,\dots,N\}$ with the corresponding camera parameters, for + +![](images/164f9b07b66db5668985f6fa52c323c8018d9c859a9c11707c268a9e7c6635f2.jpg) +Figure 2. Overview. Given a casually captured image sequence, we decompose the whole scene into 2D Gaussians for transient objects and 3D Gaussians for static scenes. To warm up, we start by training a basic 3DGS to capture static elements. This is followed by iterative training of 2D and 3D Gaussians, where our transients and statics are combined using an $\alpha$ -blending strategy with masks to produce the final renderings. The masks provide guidance for 3D Gaussians in the iterative training stage. During the joint-training, both 2D and 3D Gaussians are trained to further optimize the decomposition results. + +each view $I$ , the goal of our method is to reasonably decouple the transients $I_{t}$ and statics $I_{s}$ as follows: + +$$ +I = M _ {t} \odot I _ {t} + (1 - M _ {t}) \odot I _ {s}, \tag {7} +$$ + +where the $M_t \in [0,1]$ represents the transient mask, and $\odot$ is per pixel multiplication. When a pixel's value in $M_t$ approaches 1, it indicates a higher probability that the location is transient; conversely, a lower value suggests a greater likelihood that area is static. + +To achieve the goal, we decompose the whole 3D scene with two components: (1) Multi-view consistent 3D Gaussians is used for rendering $I_{s}$ , which leverages multi-view information from images and models the static scene with a set of unified 3D Gaussians. The multi-view inputs regulate the 3D Gaussians to be consistent and robust across the different views. (2) Single-view independent 2D Gaussians is responsible for modeling $I_{t}$ , which enables our approach to handle the fact that images are casually captured with varying transients. Concretely, we forms a set of view-independent 2D Gaussians to model transients as planar objects from a single view. This novel combination allows for a more precise and reasonable representation of 3D scenes, enabling better performance on novel views. + +This section is organized as follows: we introduce our hybrid representation of scenes using 3D Gaussians in Sec. 4.1 and 2D Gaussians in Sec. 4.2, followed by a comprehensive overview of the training pipeline in Sec. 4.3. + +# 4.1. Modeling Statics with 3D Gaussians + +Generally speaking, the initialized point clouds obtained by COLMAP [29, 30] can only represent areas in the scene that satisfy multi-view consistency, so 3D Gaussians tend to reconstruct static scenes, such as buildings, grounds, etc. + +However, in scenes with transients, continuously training the 3D Gaussians may cause the transient objects to be overfitted into the Gaussian kernels because of constraints imposed by the RGB loss. This approach fails to capture the true geometry and instead overemphasizes the RGB information, leading to annoying artifacts from other viewpoints and ultimately impacting quality of novel view synthesis. + +We address this problem by adopting an enhanced multiview regulated supervision scheme. Unlike previous methods that rasterize and supervise only a single image per iteration, our approach introduces two major changes. (1) We increase the number of images per iteration to $K$ , enabling gradient back-propagation to simultaneously consider mutual information from multiple views. This batch-wise inputs allow each optimization step to leverage multi-view insights to distinguish and infer transients and static elements. (2) We adopt a sparse training schedule that focuses optimization only on the 3D Gaussians within the co-visible areas of the cross-view frustums. This method not only sharpens the training focus but also reduces computational costs, as detailed in Alg. 1, and outputs rasterized $\hat{I}_s$ . + +# 4.2. Modeling Transients with 2D Gaussians + +Different from other methods that utilize semantic features, we introduce 2D Gaussians to model the transients as planar objects from a single view for each image. This approach is motivated by several key considerations: (1) 2DGS has been proven to be an efficient method for representing images, capable of modeling an arbitrary image effectively. (2) The formation and rasterization processes of 2DGS closely resemble those of 3DGS, allowing both to be expressed within a unified framework. (3) Degenerating 3D objects that do not satisfy multi-view geometric consistency into 2D representation can effectively decouple transients from static + +Algorithm 1: Multi-view Regulated Supervision. +Input: A set of images $\{I_k, k = 1, 2, \dots, N\}$ ; Multi-view size $K$ ; Maximum iterations $T$ ; +Output: Optimized 3D Gaussian parameters $\theta_{3d}$ within the co-visible regions; +Initialize $\theta_{3d}$ ; +for each iteration $t = 1$ to $T$ do +\[ I_{1,\dots,K} \leftarrow \text{RandomSample}(K, N); \quad F_{\text{inter}} \leftarrow \text{ComputeFrustum}(I_1); \quad \text{for } i \leftarrow 2 \text{ to } K \text{ do} \quad |F_{\text{inter}} \leftarrow F_{\text{inter}} \cap \text{ComputeFrustum}(I_i); \quad \text{end} \quad \mathcal{G}_{3d} \leftarrow \text{ExtractGaussians}(I, F_{\text{inter}}); \quad \text{for each Gaussian } G_j \in \mathcal{G}_{3d} \text{ do} \quad |\theta \leftarrow \theta - \eta \cdot \nabla \mathcal{L}(G_j(\theta)); \quad \text{end} \quad \text{end} + +scenes on a fundamental level. + +Specifically, 2D Gaussians are rasterized into an image $\hat{I}_t$ with a transient mask $\hat{M}_t$ for modeling uncertainty, each pixel $\mathbf{y}_{2\mathbf{d}}$ of which is calculated as: + +$$ +\hat {M} _ {t} \left(\mathbf {y} _ {2 d}\right) = \sum_ {i \in \mathbf {N}} \alpha_ {2 d i} ^ {\prime} = \sum_ {i \in \mathbf {N}} \alpha_ {2 d i} \cdot e ^ {- \frac {1}{2} \left(\mathbf {y} _ {2 d} - \mathbf {x} _ {2 d}\right) ^ {T} \boldsymbol {\Sigma} _ {2 d} ^ {- 1} \left(\mathbf {y} _ {2 d} - \mathbf {x} _ {2 d}\right)}. \tag {8} +$$ + +It is noteworthy that we adopt uncertainty of 2D Gaussians to model the mask, rather than using 3D results. Because after modeling a static scene with 3D Gaussians, the 2D Gaussians are responsible for learning the residual part of the image. Therefore, naturally, the results rendered by the 2D Gaussian implicitly include a mask for transients. + +# 4.3. Multi-stage Training Scheme + +Because 2D and 3D Gaussians are trained within the same framework, it is crucial to balance their relationship effectively. To that end, we have proposed a multi-stage training strategy that ensures stable convergence and enhances overall performance. + +Warm up Pre-training. The training process begins by using 3DGS to capture the essential structure of the entire static scene. This stage focuses on modeling the scene with a set of unified 3D Gaussians with our multi-view regulated supervision as detailed in Sec. 4.1, producing foundational, albeit low-quality, rendered images $\hat{I}_s$ using a combination of DSSIM [38] and L1 losses following: + +$$ +\mathcal {L} _ {w a r m u p} = \lambda L _ {D S S I M} (\hat {I} _ {s}, I) + (1 - \lambda) L _ {1} (\hat {I} _ {s}, I), \quad (9) +$$ + +where the $\lambda$ is set to 0.2. During the warm-up, the densification of 3D Gaussians is conducted to gradually increase the number, allowing for a more detailed representation of the static scene structure. + +Iterative Training. The second stage employs iterative training between 2DGS and 3DGS to progressively optimize both components. During the alternation, the 3DGS will provide more accurate static scene renderings, which in turn enables the 2DGS to refine its transient representations $\hat{I}_t$ and mask predictions $\hat{M}_t$ with Eq. 10. Conversely, the improved masks from 2DGS allow the 3DGS to focus more precisely on rendering the static regions of the scene with Eq. 11. The rendered image $\hat{I}$ is a composition of $\hat{I}_s$ and $\hat{I}_t$ with $\hat{M}_t$ following Eq. 7. This process iteratively refines the transients and statics, improving the overall quality and accuracy of the masks. The loss functions are as follows: + +$$ +\mathcal {L} _ {\text {i t e r 2 d}} = \lambda L _ {\text {D S S I M}} (\hat {I}, I) + (1 - \lambda) L _ {1} (\hat {I}, I), \tag {10} +$$ + +$$ +\mathcal {L} _ {\text {i t e r 3 d}} = \left(1 - \hat {M} _ {t}\right) \odot \mathcal {L} _ {\text {i t e r 2 d}}. \tag {11} +$$ + +Note that, during training, 2DGS and 3DGS are trained alternately. When training one branch, the gradient backpropagation of the other branch is turned off. + +Joint Fine-tuning. The final stage involves joint fine-tuning, integrating 2DGS and 3DGS into a unified framework. This stage focuses on fine-tuning on the transient mask and reducing the error between the rendered image $\hat{I}$ and the ground truth image $I$ . The loss $\mathcal{L}_{joint}$ used during the joint fine-tuning is a combination as follows: + +$$ +\mathcal {L} _ {\text {j o i n t}} = \beta L _ {D S S I M} (\hat {I}, I) + (1 - \beta) L _ {1} (\hat {I}, I), \tag {12} +$$ + +where the $\beta$ is set to 0.2. Here, both the 2DGS and 3DGS branches will enable gradient backpropagation to simultaneously update all parameters. + +# 5. Experiments + +# 5.1. Experimental Setup + +Datasets. Similar to previous methods, we evaluate our HybridGS on two challenging datasets: NeRF On-the-go [25] and RobustNeRF [27]. NeRF On-the-go contains multiple casually captured indoor and outdoor sequences with varying ratios of distractors (from $5\%$ to over $30\%$ ). We use the version of the dataset where all images were undistorted as [14]. The RobustNeRF dataset includes several scenes exemplifying different types of distractors. To simulate capturing over extended periods, distractor objects are repositioned between frames to interfere with statically placed objects (from 1 (Statue) to 150 (Yoda)). The splits include distractor-filled training set and distractor-free testing set. + +Implementation Details. Our HybridGS, developed on top of open-source implementation gsplat [43] and Taming-3DGS [19], incorporates custom CUDA kernels, boosting the rasterization of $\alpha$ -blending for 3DGS and weighted summation for 2DGS. The number of views $K$ is 4 in our practice. For a scene consisting of 100 images, the process + +Table 1. Comparison on NeRF On-the-go Dataset. For better visualization, the best results are highlighted in bold. Our method shows generally superior performance over state-of-the-art methods. * indicates that the results are reproduced from the official implementation. + +
MethodLow OcclusionMedium OcclusionHigh Occlusion
MountainFountainCornerPatioSpotPatio-High
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
RobustNeRF [27]17.540.4960.38315.650.3180.57623.040.7640.24420.390.7180.25120.650.6250.39120.540.5780.366
NeRF On-the-go [25]20.150.6440.25920.110.6090.31424.220.8060.19020.780.7540.21923.330.7870.18921.410.7180.235
3DGS [11]19.400.6380.21319.960.6590.18520.900.7130.24117.480.7040.19920.770.6930.31617.290.6040.363
SLS-mlp* [28]19.840.5800.31520.190.6120.25924.030.7950.16621.550.8380.32823.520.7560.23520.310.6640.261
HybridGS(Ours)21.730.6930.22521.110.6740.18825.030.8470.07721.980.8120.11024.330.7940.08621.770.7410.142
+ +Table 2. Comparison on RobustNeRF Dataset. For better visualization, the best results are highlighted in bold. Our method significantly outperforms all baseline methods. Crab (1) scene had a test set with same set of views as those in the train set, which is fixed in Crab (2). + +
MethodStatueAndroidYodaCrab (1)Crab (2)
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
RobustNeRF [27]20.600.760.1523.280.750.1329.780.820.1532.220.940.06---
NeRF On-the-go [25]21.580.770.2423.500.750.2129.960.830.24------
3DGS [11]21.020.810.1623.110.810.1326.330.910.1431.800.960.0829.74--
SLS-mlp [28]22.540.840.1325.050.850.0933.660.960.1035.850.970.0834.43--
HybridGS(Ours)22.930.870.1025.150.850.0735.320.960.0736.310.970.0535.170.960.08
+ +Table 3. Ablation on components in datasets with varying occlusions. The best results are highlighted in bold. + +
Dataset(ratio)On-the-go low.(5%)On-the-go medium.(17%)On-the-go high.(26%)
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Ours21.420.6840.20623.500.8270.09223.050.7680.115
w/o iterative Training21.390.6870.20923.110.8150.09422.550.7530.123
w/o Joint Training21.050.6540.24123.180.8150.10723.010.7480.130
w/o Multi-view21.020.6580.23323.060.8120.10322.710.7630.111
+ +starts with a 1k-step warm-up. This is followed by iterative training, which includes 10k steps for 2DGS and 1k steps for 3DGS. The final joint training involves 30k steps. The entire training process takes approximately 0.18 GPU hours. The number of training steps scales proportionally with the size of the dataset. All experiments are performed on a single NVIDIA RTX 4090 GPU using PyTorch. + +Metrics. We follow common practice by employing PSNR, SSIM [38], and LPIPS [47] for evaluation. + +# 5.2. Evaluation + +# 5.2.1 Comparison on the NeRF On-the-go Dataset + +As shown in Tab. 1 and Fig. 3, our method demonstrates superior performance across different scenarios with varying levels of occlusion. In Tab. 1, our method outperforms the previous SOTA method across multiple metrics, with an average improvement of $1.10\mathrm{dB}$ in PSNR and $5.27\%$ in SSIM, demonstrating the superiority of our approach. + +Furthermore, Fig. 3 provides a qualitative comparison, illustrating the visual fidelity of our results against other methods. Our method produces more clear and detailed images, closely resembling the ground truth, particularly in challenging scenes with complex geometry and occlusions. + +Table 4. Ablation on the number of 2D Gaussians of Corner in NeRF On-the-go dataset. For better visualization, the best results are highlighted in bold. + +
NumberPSNR↑SSIM↑LPIPS↓Storage (MB)
2DGS3DGS
1k24.4930.8400.0773.542.1
5k24.7330.8430.08217.341.9
10k25.0340.8470.08034.742.2
15k24.5120.8390.09242.652
25k24.6980.8400.09469.442.2
+ +This highlights our model's capability to effectively handle diverse visual conditions both in indoor and outdoor scenes. + +# 5.2.2 Comparison on the RobustNeRF Dataset + +We present the quantitative results of RobustNeRF in Tab. 2. Our method demonstrates superior performance quantitatively compared to all previous approaches. In scenes with numerous transients, such as the Yoda scene, we've noticed that the performance metrics of the original 3DGS significantly degrade, occasionally even dropping below those of NeRF-related methods. By leveraging multi-view shared information and incorporating 2D Gaussians to effectively represent transients, our approach surpasses semantic-based methods like SLS-mlp [28]. As shown in Fig. 4, our results have clearer overall boundaries and structured fine-grained details, while SLS-mlp [28] may lead to blurriness in static objects and occluded areas. This further confirms that, even when faced with numerous transients, our method consistently maintains stable and reliable performance. + +# 5.3. Ablation Studies + +Components. To assess the effectiveness of each component in our framework, we performed an ablation study, as + +![](images/e33afbb0bdf4907ca393059a9b5759b59034735f10b9ffb421385dd0c15538d6.jpg) +Figure 3. Visualization of novel view synthesis results on the testing set of NeRF On-the-go dataset. Our method demonstrates superior results by effectively reducing artifacts and providing clearer boundaries. This results in a cleaner statics compared to other methods, showcasing enhanced visual quality and precision in novel views. + +![](images/9dfe4fd9428dfd8798707caf0a76b15c3988f2de6b6531b144801a2a0e452196.jpg) +Figure 4. Visualization of novel view synthesis results on RobustNeRF dataset. + +presented in Tab. 3. We systematically removed iterative training, joint training, and multi-view supervision to evaluate their effectiveness. Each component, as expected, enhances overall performance, demonstrating their individual and combined significance. In scenes with low occlusion, multi-view supervision allows the model to concentrate on static elements by integrating data from various perspectives. Conversely, in highly occluded environments, joint training with 2D and 3D Gaussians markedly improves performance, highlighting the model's capability to manage complex scenes with significant occlusions and transient changes. Iterative training consistently boosts results across different occlusion levels. In summary, the results underscore that integrating diverse training strategies and multiview information significantly enhances the rendering quality of unseen views, especially in challenging conditions. + +The Number of 2D Gaussians. Tab. 4 presents an ablation study on the number of 2D Gaussians used in the corner sequence of NeRF On-the-go dataset. The results indicate that + +![](images/d01dbc1fb34ae3d3635214eaf0fd2c6a95d36a5d457c989c044930dd1eb34234.jpg) +Figure 5. Comparison of transient masks on NeRF On-the-go dataset. + +increasing the number of 2D Gaussians initially improves the performance. Specifically, 10k 2D Gaussians achieve the highest PSNR (25.034) and SSIM (0.847), while maintaining a competitive LPIPS score of 0.151 and reasonable storage requirements. Based on these findings, we select 10k as the optimal number in our experiments. + +Transients and Masks. Benefiting from our method's explicit modeling of transient objects, we can obtain transient masks without introducing any segmentation networks or other pre-trained features. Typically, these masks capture dynamic elements such as pedestrians and vehicles, as well as some structural details ignored by 3D Gaussians. In Fig. 6, our approach effectively learns the RGB and mask of transients even under strong occlusions, while maintaining the robustness of static elements. This capability underscores the strength of our method in discerning and isolating dynamic components within complex scenes. + +We also compare our transient masks with SLS-mlp [28] semantic masks in Fig. 5. It is observed that the semantic mask does not perfectly align with the transients due to + +![](images/cf5966e71b1ac324ab8b0c069f467e7624bd96a01358cb122fdc01bf4ea115e4.jpg) +Figure 6. Visualization of scene decomposition into transients and statics. Our method achieves superior transient mask separation in both indoor and outdoor scenes. It effectively separates transients and statics, and the resulting renderings closely resemble the ground truth images, demonstrating its effectiveness. Top two rows are from NeRF On-the-go [25], bottom two rows are from RobustNeRF [27]. + +![](images/d0e6b63f56ad97a30fb02649652e09c61d5c5c8befb4d3741942aae023fde704.jpg) +Figure 7. Visualization on Photo Tourism dataset. + +the semantic features not being sufficiently precise. Additionally, transients often encompass low-level information, such as motion blur or object shadows, which may not explicitly exhibit clear semantic features. By modeling these transients using 2D Gaussians at the pixel level, our method allows for a more fine-grained capture of transients. + +# 5.4. Discussions + +While our method has demonstrated strong performance in distinguishing between transients and statics, it has not yet accounted for the illumination variation in unconstrained photo collections. As illustrated in Fig. 7 on Photo Tourism [32], our method renders fine-grained architectural + +structures, but the overall photometric result tends towards an average photometric effect, akin to overcast conditions. A promising direction is to incorporate an appearance embedding module in our method for the future work. + +# 6. Conclusion + +In this paper, we introduce HybridGS, a novel hybrid representation with 2D and 3D Gaussians for decomposing 3D scenes from casually captured images. By introducing the hybrid Gaussian representation, our method addresses the challenges of handling transients as planar objects while preserving the integrity of the static scene. The multi-view supervision mechanism we developed plays a crucial role in enhancing the discrimination between static and transient elements across multiple views, reducing artifacts and inaccuracies in novel views. We have developed a multi-stage training strategy for the joint training of 2D and 3D Gaussians, ensuring the stable and robust training stage. Our experiments demonstrate that HybridGS significantly outperforms existing methods in terms of quality and efficiency. In summary, HybridGS establishes a novel, robust and efficient foundation for 3D scene representation, paving the way for advancements in various vision-based applications. + +# References + +[1] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. ICCV, 2021. 2 +[2] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. CVPR, 2022. 2 +[3] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. Pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19457-19467, 2023. 1 +[4] Jiahao Chen, Yipeng Qin, Lingjie Liu, Jiangbo Lu, and Guanbin Li. Nerf-hugs: Improved neural radiance fields in non-static scenes using heuristics-guided segmentation. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19436-19446, 2024. 2, 3 +[5] Xingyu Chen, Qi Zhang, Xiaoyu Li, Yue Chen, Feng Ying, Xuan Wang, and Jue Wang. Hallucinated neural radiance fields in the wild. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12933-12942, 2021. 3 +[6] Zilong Chen, Feng Wang, and Huaping Liu. Text-to-3d using gaussian splatting. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21401–21412, 2023. 1 +[7] Hiba Dahmani. Swag: Splitting in the wild images with appearance-conditioned gaussians. In ECCV, 2024. 3 +[8] Sara Fridovich-Keil, Giacomo Meanti, Frederik Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12479-12488, 2023. 3 +[9] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In SIGGRAPH 2024 Conference Papers. Association for Computing Machinery, 2024. 2, 3 +[10] Karim Kassab, Antoine Schnepf, Jean-Yves Franceschi, Laurent Caraffa, Jeremie Mary, and Valérie Gouet-Brunet. Refinedfields: Radiance fields refinement for unconstrained scenes. arXiv preprint arXiv:2312.00639, 2023. 3 +[11] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuehler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (TOG), 42:1-14, 2023. 1, 2, 6 +[12] Bernhard Kerbl, Andreas Meuleman, Georgios Kopanas, Michael Wimmer, Alexandre Lanvin, and George Drettakis. A hierarchical 3d gaussian representation for real-time rendering of very large datasets. ACM Transactions on Graphics, 43(4), 2024. 2 +[13] Jonávs. Kulhánek and Torsten Sattler. Tetra-nerf: Representing neural radiance fields using tetrahedra. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 18412–18423, 2023. 2 +[14] Jonas Kulhanek, Songyou Peng, Zuzana Kukelova, Marc + +Pollefeys, and Torsten Sattler. Wildgaussians: 3d gaussian splatting in the wild. 2024. 3, 5 +[15] Pengzhi Li, Chengshuai Tang, Qinxuan Huang, and Zhiheng Li. Art3d: 3d gaussian splatting for text-guided artistic scenes generation, 2024. 1 +[16] Zhihao Liang, Qi Zhang, Yingfa Feng, Ying Shan, and Kui Jia. Gs-ir: 3d gaussian splatting for inverse rendering. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21644–21653, 2023. 1 +[17] Jiaqi Lin, Zhihao Li, Xiao Tang, Jianzhuang Liu, Shiyong Liu, Jiayue Liu, Yangdi Lu, Xiaofei Wu, Songcen Xu, Youliang Yan, and Wenming Yang. Vastgaussian: Vast 3d gaussians for large scene reconstruction. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5166-5175, 2024. 2 +[18] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20654-20664, 2024. 2 +[19] Saswat Subhajyoti Mallick, Rahul Goel, Bernhard Kerbl, Markus Steinberger, Francisco Vicente Carrasco, and Fernando De La Torre. Taming 3dgs: High-quality radiance fields with limited resources. In SIGGRAPH Asia 2024 Conference Papers, New York, NY, USA, 2024. Association for Computing Machinery. 2, 5 +[20] Ricardo Martin-Brualla, Noha Radwan, Mehdi S. M. Sajjadi, Jonathan T. Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7206-7215, 2020. 2, 3 +[21] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2 +[22] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 2 +[23] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINOV2: Learning robust visual features without supervision. Transactions on Machine Learning Research, 2024. 2 +[24] Konstantinos Rematas, Andrew Liu, Pratul P. Srinivasan, Jonathan T. Barron, Andrea Tagliasacchi, Tom Funkhouser, and Vittorio Ferrari. Urban radiance fields. CVPR, 2022. 2 +[25] Weining Ren, Zihan Zhu, Boyang Sun, Jiaqi Chen, Marc Pollefeys, and Songyou Peng. Nerf on-the-go: Exploiting uncertainty for distractor-free nerfs in the wild. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 3, 5, 6, 8 + +[26] Barbara Roessle, Norman Müller, Lorenzo Porzi, Samuel Rota Bulò, Peter Kontschieder, Angela Dai, and Matthias Nießner. L3dg: Latent 3d gaussian diffusion. In SIGGRAPH Asia 2024 Conference Papers, 2024. 1 +[27] Sara Sabour, Suhani Vora, Daniel Duckworth, Ivan Krasin, David J. Fleet, and Andrea Tagliasacchi. Robustnerf: Ignoring distractors with robust losses. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20626-20636, 2023. 1, 2, 3, 5, 6, 8 +[28] Sara Sabour, Lily Goli, George Kopanas, Mark Matthews, Dmitry Lagun, Leonidas Guibas, Alec Jacobson, David J. Fleet, and Andrea Tagliasacchi. SpotLessSplats: Ignoring distractors in 3d gaussian splatting. arXiv:2406.20055, 2024. 2, 3, 6, 7 +[29] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 4 +[30] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In European Conference on Computer Vision (ECCV), 2016. 4 +[31] Jin-Chuan Shi, Miao Wang, Hao-Bin Duan, and Shao-Hua Guan. Language embedded 3d gaussians for open-vocabulary scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5333-5343, 2024. 1 +[32] Noah Snavely, Steven M. Seitz, and Richard Szeliski. Photo tourism: Exploring photo collections in 3d. In SIGGRAPH Conference Proceedings, pages 835-846, New York, NY, USA, 2006. ACM Press. 8 +[33] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P. Srinivasan, Jonathan T. Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8238–8248, 2022. 2 +[34] Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Justin Kerr, Terrance Wang, Alexander Kristoffersen, Jake Austin, Kamyar Salahi, Abhik Ahuja, David McAllister, and Angjoo Kanazawa. Nerfstudio: A modular framework for neural radiance field development. ACM SIGGRAPH 2023 Conference Proceedings, 2023. 2 +[35] Linhan Wang, Kai Cheng, Shuo Lei, Shengkun Wang, Wei Yin, Chenyang Lei, Xiaoxiao Long, and Chang-Tien Lu. Dcgaussian: Improving 3d gaussian splatting for reflective dash cam videos. NeurIPS 2024, 2024. 1 +[36] Shuaixian Wang, Haoran Xu, Yaokun Li, Jiwei Chen, and Guang Tan. Ie-nerf: Inpainting enhanced neural radiance fields in the wild. ArXiv, abs/2407.10695, 2024. 3 +[37] Yuze Wang, Junyi Wang, and Yue Qi. We-gs: An in-the-wild efficient 3d gaussian representation for unconstrained photo collections, 2024. 3 +[38] Zhou Wang, Alan Conrad Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13:600-612, 2004. 5, 6 +[39] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. + +4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20310-20320, 2024. 1 +[40] Tianhao Walter Wu, Fangcheng Zhong, Andrea Tagliasacchi, Forrester Cole, and Cengiz Oztireli. D $^2$ neRF: Self-supervised decoupling of dynamic and static objects from a monocular video. In Advances in Neural Information Processing Systems, 2022. 2 +[41] Congrong Xu, Justin Kerr, and Angjoo Kanazawa. Splatfacto-w: A nefstudio implementation of gaussian splatting for unconstrained photo collections. ArXiv, abs/2407.12306, 2024. 3 +[42] Yifan Yang, Shuhai Zhang, Zixiong Huang, Yubing Zhang, and Mingkui Tan. Cross-ray neural radiance fields for novel-view synthesis from unconstrained image collections. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 15855-15865, 2023. 3 +[43] Vickie Ye, Ruilong Li, Justin Kerr, Matias Turkulainen, Brent Yi, Zhuoyang Pan, Otto Seiskari, Jianbo Ye, Jeffrey Hu, Matthew Tancik, and Angjoo Kanazawa. gsplat: An open-source library for Gaussian splatting. arXiv preprint arXiv:2409.06765, 2024. 5 +[44] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19447-19456, 2024. 2 +[45] Zehao Yu, Torsten Sattler, and Andreas Geiger. Gaussian opacity fields: Efficient adaptive surface reconstruction in unbounded scenes. ACM Transactions on Graphics, 2024. 2 +[46] Dongbin Zhang, Chuming Wang, Weitao Wang, Peihao Li, Minghan Qin, and Haoqian Wang. Gaussian in the wild: 3d gaussian splatting for unconstrained image collections. ECCV, 2024. 3 +[47] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 586-595, 2018. 6 +[48] Xinjie Zhang, Xingtong Ge, Tongda Xu, Dailan He, Yan Wang, Hongwei Qin, Guo Lu, Jing Geng, and Jun Zhang. Gaussianimage: 1000 fps image representation and compression by 2d gaussian splatting. In European Conference on Computer Vision, 2024. 3 +[49] Yunxiang Zhang, Alexandr Kuznetsov, Akshay Jindal, Kenneth Chen, Anton Sochenov, Anton Kaplanyan, and Qi Sun. Image-gs: Content-adaptive image representation via 2d gaussians, 2024. 3 +[50] Xiaoyu Zhou, Zhiwei Lin, Xiaojun Shan, Yongtao Wang, Deqing Sun, and Ming-Hsuan Yang. Drivinggaussian: Composite gaussian splatting for surrounding dynamic autonomous driving scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21634-21643, 2024. 1 \ No newline at end of file diff --git a/2025/HybridGS_ Decoupling Transients and Statics with 2D and 3D Gaussian Splatting/images.zip b/2025/HybridGS_ Decoupling Transients and Statics with 2D and 3D Gaussian Splatting/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4cf99801d74898d87802ba411b32cb6c13d073ba --- /dev/null +++ b/2025/HybridGS_ Decoupling Transients and Statics with 2D and 3D Gaussian Splatting/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:022f1f5ae74f2ad6f1c54a18176cf619639514368f672b55263580d44da95b5f +size 849249 diff --git a/2025/HybridGS_ Decoupling Transients and Statics with 2D and 3D Gaussian Splatting/layout.json b/2025/HybridGS_ Decoupling Transients and Statics with 2D and 3D Gaussian Splatting/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e606e2dde056307467d8ed357679b25a0ef6ff68 --- /dev/null +++ b/2025/HybridGS_ Decoupling Transients and Statics with 2D and 3D Gaussian Splatting/layout.json @@ -0,0 +1,8334 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 164, + 103, + 446, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 103, + 446, + 139 + ], + "spans": [ + { + "bbox": [ + 164, + 103, + 446, + 139 + ], + "type": "text", + "content": "HybridGS: Decoupling Transients and Statics with 2D and 3D Gaussian Splitting" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 159, + 476, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 159, + 476, + 190 + ], + "spans": [ + { + "bbox": [ + 130, + 159, + 476, + 190 + ], + "type": "text", + "content": "Jingyu Lin" + }, + { + "bbox": [ + 130, + 159, + 476, + 190 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 130, + 159, + 476, + 190 + ], + "type": "text", + "content": " Jiaqi Gu" + }, + { + "bbox": [ + 130, + 159, + 476, + 190 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 130, + 159, + 476, + 190 + ], + "type": "text", + "content": " Lubin Fan" + }, + { + "bbox": [ + 130, + 159, + 476, + 190 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 130, + 159, + 476, + 190 + ], + "type": "text", + "content": " Bojian Wu" + }, + { + "bbox": [ + 130, + 159, + 476, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 130, + 159, + 476, + 190 + ], + "type": "text", + "content": " Yujing Lou" + }, + { + "bbox": [ + 130, + 159, + 476, + 190 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 130, + 159, + 476, + 190 + ], + "type": "text", + "content": " Renjie Chen" + }, + { + "bbox": [ + 130, + 159, + 476, + 190 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 130, + 159, + 476, + 190 + ], + "type": "text", + "content": " Ligang Liu" + }, + { + "bbox": [ + 130, + 159, + 476, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 130, + 159, + 476, + 190 + ], + "type": "text", + "content": " Jieping Ye" + }, + { + "bbox": [ + 130, + 159, + 476, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 190, + 506, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 190, + 506, + 204 + ], + "spans": [ + { + "bbox": [ + 104, + 190, + 506, + 204 + ], + "type": "text", + "content": "1University of Science and Technology of China 2Independent Researcher 4Shanghai Jiao Tong University" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 65, + 210, + 192, + 346 + ], + "blocks": [ + { + "bbox": [ + 65, + 210, + 192, + 346 + ], + "lines": [ + { + "bbox": [ + 65, + 210, + 192, + 346 + ], + "spans": [ + { + "bbox": [ + 65, + 210, + 192, + 346 + ], + "type": "image", + "image_path": "041433ab006550dde174f0f360819f9503991771e505a0a0cd50cee3009bb315.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 356, + 555, + 379 + ], + "lines": [ + { + "bbox": [ + 55, + 356, + 555, + 379 + ], + "spans": [ + { + "bbox": [ + 55, + 356, + 555, + 379 + ], + "type": "text", + "content": "Figure 1. HybridGS is the first hybrid representation that combines multi-view consistent 3D Gaussians and single-view independent 2D Gaussians, which is used to decouple the transients and statics presented in the scene. Our results demonstrate reasonable decompositions." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 204, + 211, + 290, + 350 + ], + "blocks": [ + { + "bbox": [ + 204, + 211, + 290, + 350 + ], + "lines": [ + { + "bbox": [ + 204, + 211, + 290, + 350 + ], + "spans": [ + { + "bbox": [ + 204, + 211, + 290, + 350 + ], + "type": "image", + "image_path": "6464bab9e3de9e887ad3450894aea2244a1e6e60c81f408c52babe888a668ef5.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 292, + 212, + 378, + 350 + ], + "blocks": [ + { + "bbox": [ + 292, + 212, + 378, + 350 + ], + "lines": [ + { + "bbox": [ + 292, + 212, + 378, + 350 + ], + "spans": [ + { + "bbox": [ + 292, + 212, + 378, + 350 + ], + "type": "image", + "image_path": "f99fa44d342c780aba4ad3f673914bc8d58a8fc7dd7b5d4cb71239c52c34dd2b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 380, + 212, + 466, + 350 + ], + "blocks": [ + { + "bbox": [ + 380, + 212, + 466, + 350 + ], + "lines": [ + { + "bbox": [ + 380, + 212, + 466, + 350 + ], + "spans": [ + { + "bbox": [ + 380, + 212, + 466, + 350 + ], + "type": "image", + "image_path": "39c629258ad6a6120adae51d206764b86889798fabec06f5f7bfa6986e6c8ddd.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 467, + 213, + 553, + 350 + ], + "blocks": [ + { + "bbox": [ + 467, + 213, + 553, + 350 + ], + "lines": [ + { + "bbox": [ + 467, + 213, + 553, + 350 + ], + "spans": [ + { + "bbox": [ + 467, + 213, + 553, + 350 + ], + "type": "image", + "image_path": "ff48fb684431d963da5c2f04ab53208fb965d3dfa5bb43cb6e53dc286dff494a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 152, + 397, + 200, + 409 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 397, + 200, + 409 + ], + "spans": [ + { + "bbox": [ + 152, + 397, + 200, + 409 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 54, + 424, + 297, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 424, + 297, + 676 + ], + "spans": [ + { + "bbox": [ + 54, + 424, + 297, + 676 + ], + "type": "text", + "content": "Generating high-quality novel view renderings of 3D Gaussian Splatting (3DGS) in scenes featuring transient objects is challenging. We propose a novel hybrid representation, termed as HybridGS, using 2D Gaussians for transient objects per image and maintaining traditional 3D Gaussians for the whole static scenes. 3DGS is suited for modeling static scenes that assume multi-view consistency, but the transient objects appear occasionally and do not adhere to the assumption, thus we model them as planar objects from a single view by 2D Gaussians. Our novel representation decomposes the scene from the perspective of fundamental viewpoint consistency. Additionally, we present a multiview supervision method for 3DGS that leverages information from co-visible regions, further enhancing the distinctions between the transients and statics. Then, we propose a straightforward yet effective multi-stage training strategy to ensure robust training and view synthesis. Experiments on benchmarks show our state-of-the-art performance of novel view synthesis in indoor and outdoor scenes, even in the presence of distracting elements. Project page: https://gujiaqivadin.github.io/hybridgs/" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 315, + 397, + 394, + 409 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 397, + 394, + 409 + ], + "spans": [ + { + "bbox": [ + 315, + 397, + 394, + 409 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 422, + 555, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 555, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 555, + 662 + ], + "type": "text", + "content": "3D Gaussians Splating (3DGS) [11] has recently gained popularity for novel view synthesis due to its high-quality rendering, efficiency and low-memory cost. Its applications span across virtual reality, augmented reality, and robotics, etc [3, 6, 15, 16, 26, 31, 35, 39, 50]. Typically, most current approaches assume that the input images are posed and free of noise, which are the prerequisite needed to represent a complete scene well. However, this assumption is often not met. For example, images taken casually with a mobile phone usually contain messy dynamic objects, or transient objects. Therefore, it is hard to model the scene because there are always many transient occlusions. Technically, 3DGS is capable of modeling static scenes effectively because the static objects in images satisfy geometric consistency constraints across different viewpoints. However, transient objects do not follow this assumption. If the images containing transient objects are naively used in training 3DGS, such variations can lead to difficulties in achieving coherent blending of individual Gaussian representations, potentially resulting in artifacts or inaccuracies." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "content": "To address this issue, numerous methods have been proposed to improve the scene representation ability involving transient objects. In the era of Neural Radiance Fields (NeRF) [21], RobustNeRF [27] introduced a robust loss" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 219, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 219, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 219, + 713 + ], + "type": "text", + "content": "†Equal contributions. ‡Corresponding authors." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "788" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "content": "function that minimizes the influence of photometrically inconsistent observations during training, thereby enhancing the overall quality of the representation. NeRF On-the-go [25] was the first to leverage the uncertainty of transients by incorporating DINOv2 [23] features into a shallow MLP. It focused on loss supervision in the areas of low uncertainty to effectively remove transients from dynamic scenes. These two methods have established widely-used benchmark datasets featuring transients, aiming to evaluate the reconstruction of clear static scenes with improved results of novel view synthesis. For 3DGS, SpotLessSplats (SLS-mlp) [28] integrated the designs of those aforementioned methods, employing adaptive method to detect outliers through clustering in the feature space, achieving the best current results." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 253, + 294, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 253, + 294, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 253, + 294, + 360 + ], + "type": "text", + "content": "However, these previous approaches neglect two critical issues: (1) The ambiguity between semantics and transients. Utilizing features from visual models to capture the semantic characteristics of transients can inherently lead to confusion in distinguishing between them. (2) The absence of explicit transient modeling. These methods primarily focus on achieving a robust statics by merely mitigating the impact of transients, rather than comprehensively modeling the whole scene themselves." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 362, + 295, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 362, + 295, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 362, + 295, + 672 + ], + "type": "text", + "content": "In response to the issues mentioned above, our key observation is that transients lack multi-view consistency and usually only appear in a single view. Therefore, we consider them as planar objects at that view. That is to say, this planar representation is a unique feature of the current view. Based on this, we propose HybridGS, a hybrid representation with 2D Gaussians for transient objects per image and 3D Gaussians for the whole static scenes as illustrated in Fig. 1, which naturally decouples the transients and statics based on their intrinsic properties. With this novel representation, while modeling transients as planar objects, the fundamental consistency of viewpoints is also maintained for statics. To further enhance the performance, a multiview regulated supervision method for 3DGS is adopted to incorporates information from co-visible regions across multiple views, leading to more accurate representation. Additionally, we employ a simple yet effective multi-stage training strategy that ensures robust training and consistent view synthesis across various scenarios, thereby guaranteeing better stability and convergence during the training process. Extensive experiments on benchmark datasets demonstrate our state-of-the-art quality of novel view synthesis in both indoor and outdoor scenes, leading to the accurate decoupling of transients and statics as well as high efficiency in training and inference stage. The contributions are summarized as follows." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "- We are the first to introduce a novel hybrid representation that combines image-specific 2D Gaussians with static 3D Gaussians, enabling effective modeling of transient" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 322, + 72, + 487, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 72, + 487, + 84 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 487, + 84 + ], + "type": "text", + "content": "objects within casually captured images." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 84, + 553, + 191 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 314, + 84, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 84, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 314, + 84, + 553, + 144 + ], + "type": "text", + "content": "- We develop a multi-view supervision scheme for 3DGS that utilizes overlapping regions across multiple views, which enhances the model's capability to distinguish between static and transient elements, ultimately improving the overall quality of the novel view synthesis." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 144, + 553, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 553, + 191 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 553, + 191 + ], + "type": "text", + "content": "- Our HybridGS achieves the state-of-the-art performance in benchmark datasets, which outperforms the previous methods and sets a new standard for novel view synthesis in the scenes with transients." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 204, + 400, + 216 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 204, + 400, + 216 + ], + "spans": [ + { + "bbox": [ + 314, + 204, + 400, + 216 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 224, + 436, + 236 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 224, + 436, + 236 + ], + "spans": [ + { + "bbox": [ + 314, + 224, + 436, + 236 + ], + "type": "text", + "content": "2.1. Novel View Synthesis" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 243, + 553, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 553, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 553, + 410 + ], + "type": "text", + "content": "NeRF (Neural Radiance Fields) [1, 2, 13, 17, 21, 22, 24, 33, 34] revolutionizes 3D scene reconstruction by representing scenes as continuous neural networks that map 3D coordinates to color and density values, which allows for the synthesis of photorealistic views from arbitrary camera positions. Recently, 3D Gaussian Splatting (3DGS) [9, 11, 12, 18, 19, 44, 45] has gained attention as an effective method for novel view synthesis. It starts with an initial point cloud and transforms it into estimizable 3D Gaussian primitives, known for their high-quality rendering and fast processing speeds. Nevertheless, existing methods require the assumption that only static scenes are present in the input images to effectively utilize multi-view geometric consistency to train a coherent scene representation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 418, + 553, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 418, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 314, + 418, + 553, + 431 + ], + "type": "text", + "content": "2.2. Modeling Transients and Statics in NeRF/3DGS" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 437, + 553, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 553, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 553, + 640 + ], + "type": "text", + "content": "In casually captured footage of real-world scenes, there are often numerous dynamic occlusions, such as moving pedestrians and vehicles. Many works have tried to address challenges in scenes featuring both static and transient objects. RobustNeRF [27] was the pioneer work that introduces a robust loss function that reduces the influence of photometrical inconsistency during training, thereby improving the overall reconstruction quality on its proposed dataset. Other methods [4, 20, 40] also applied the similar idea to strengthen the robustness to outliers by down-weighting or discarding inconsistent observations based on the magnitude of color residuals. NeRF On-the-go was the first to introduce DINOV2 [23] features to predict pixel uncertainty, in an attempt to achieve clean static reconstructions without transients. It also introduced a benchmark dataset of casually captured images, aimed at evaluating static reconstruction performance for novel views." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "type": "text", + "content": "Following these two methods and their datasets, recent methods of 3DGS have also explored techniques for separating transients and statics to reconstruct the statics more effectively. For example, SpotLessSplats [28] further extended previous methods by utilizing semantic features [23] combined with robust optimization strategies to" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "789" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "effectively eliminate transient distractors. Besides, several approaches [4, 5, 7, 8, 10, 14, 20, 36, 37, 41, 42, 46] have focused on reconstructing statics and illumination in datasets with unconstrained photo collections, that typically employ appearance embedding vectors to model the unique lighting and styles of each image, allowing for more accurate static reconstruction despite varying illumination status." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 156, + 294, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 156, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 156, + 294, + 239 + ], + "type": "text", + "content": "Our method specifically focuses on distinguishing between transients and statics in casually captured unordered images with minimal illumination changes. Different from previous methods [25, 27, 28], we propose a hybrid Gaussian representation with 2D and 3D Gaussians with multiview information to model the whole 3D scene without any additional semantic features." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 247, + 271, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 247, + 271, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 271, + 258 + ], + "type": "text", + "content": "2.3. Image Representation with 2D Gaussians" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 265, + 295, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 265, + 295, + 372 + ], + "spans": [ + { + "bbox": [ + 55, + 265, + 295, + 372 + ], + "type": "text", + "content": "Recently, inspired by 3DGS, some methods [48, 49] in the field of image representation and compression have started to use 2D Gaussians to represent images. GaussianImage [48] leverages 2D Gaussians on the pixel plane to achieve comparable reconstruction quality while offering enhanced compression and accelerated rendering capabilities. Image-GS [49] creates a content-adaptive image representation by fitting a target image by adaptively allocating and progressively optimizing a set of 2D Gaussians." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 373, + 295, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 295, + 456 + ], + "type": "text", + "content": "Note that the 2D Gaussians used here differ from the 2D Surfels (also referred to as 2D Gaussians) introduced by [9]. While 2D Surfels are designed to enable perspective-correct splatting through 2D surface modeling, ray-splat intersection, and volumetric integration in 3D space, and the 2D Gaussians in GaussianImage [48] is a flexible, compact, and content-adaptive image representation within 2D space." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 456, + 294, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 294, + 492 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 294, + 492 + ], + "type": "text", + "content": "Here, we incorporate the trainable 2D Gaussians to learn transient and uncertainty masks during the training process, decomposing the whole scene with 2D and 3D Gaussians." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 502, + 139, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 139, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 139, + 514 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 522, + 295, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 606 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 606 + ], + "type": "text", + "content": "3D Gaussian Splatting (3DGS) represent a 3D scene explicitly with a set of 3D anisotropic Gaussians " + }, + { + "bbox": [ + 55, + 522, + 295, + 606 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{G}_{3d}\\}" + }, + { + "bbox": [ + 55, + 522, + 295, + 606 + ], + "type": "text", + "content": ". Each Gaussian is parameterized by its centroid " + }, + { + "bbox": [ + 55, + 522, + 295, + 606 + ], + "type": "inline_equation", + "content": "\\mathbf{x_{3d}}\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 522, + 295, + 606 + ], + "type": "text", + "content": ", scale " + }, + { + "bbox": [ + 55, + 522, + 295, + 606 + ], + "type": "inline_equation", + "content": "\\mathbf{S_{3d}}\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 522, + 295, + 606 + ], + "type": "text", + "content": ", rotation matrix " + }, + { + "bbox": [ + 55, + 522, + 295, + 606 + ], + "type": "inline_equation", + "content": "\\mathbf{R}\\in \\mathbb{R}^{3\\times 3}" + }, + { + "bbox": [ + 55, + 522, + 295, + 606 + ], + "type": "text", + "content": ", opacity " + }, + { + "bbox": [ + 55, + 522, + 295, + 606 + ], + "type": "inline_equation", + "content": "\\alpha_{3d}\\in \\mathbb{R}" + }, + { + "bbox": [ + 55, + 522, + 295, + 606 + ], + "type": "text", + "content": " and color " + }, + { + "bbox": [ + 55, + 522, + 295, + 606 + ], + "type": "inline_equation", + "content": "\\mathbf{c_{3d}}\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 522, + 295, + 606 + ], + "type": "text", + "content": " encoded in spherical harmonic (SH) coefficients. The 3D covariance matrix " + }, + { + "bbox": [ + 55, + 522, + 295, + 606 + ], + "type": "inline_equation", + "content": "\\Sigma_{3d}" + }, + { + "bbox": [ + 55, + 522, + 295, + 606 + ], + "type": "text", + "content": " of the 3D Gaussian is obtained from " + }, + { + "bbox": [ + 55, + 522, + 295, + 606 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{3d}" + }, + { + "bbox": [ + 55, + 522, + 295, + 606 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 522, + 295, + 606 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{3d}" + }, + { + "bbox": [ + 55, + 522, + 295, + 606 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 613, + 294, + 627 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 613, + 294, + 627 + ], + "spans": [ + { + "bbox": [ + 115, + 613, + 294, + 627 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\Sigma} _ {3 \\mathbf {d}} = \\mathbf {R} _ {3 \\mathbf {d}} \\mathbf {S} _ {3 \\mathbf {d}} \\mathbf {S} _ {3 \\mathbf {d}} ^ {T} \\mathbf {R} _ {3 \\mathbf {d}} ^ {T}. \\tag {1}", + "image_path": "c702312f28539973bcba1013df4cb4e1da89fc6dfda7bdc9a7d962e9bf98d587.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 635, + 295, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 295, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 295, + 658 + ], + "type": "text", + "content": "The 3D Gaussians are defined in world space for a spatial point " + }, + { + "bbox": [ + 55, + 635, + 295, + 658 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_{3d}" + }, + { + "bbox": [ + 55, + 635, + 295, + 658 + ], + "type": "text", + "content": " following:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 88, + 666, + 294, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 666, + 294, + 681 + ], + "spans": [ + { + "bbox": [ + 88, + 666, + 294, + 681 + ], + "type": "interline_equation", + "content": "\\mathcal {G} _ {3 d} \\left(\\mathbf {y} _ {3 \\mathrm {d}}\\right) = e ^ {- \\frac {1}{2} \\left(\\mathbf {y} _ {3 \\mathrm {d}} - \\mathbf {x} _ {3 \\mathrm {d}}\\right) ^ {T} \\boldsymbol {\\Sigma} _ {3 \\mathrm {d}} ^ {- 1} \\left(\\mathbf {y} _ {3 \\mathrm {d}} - \\mathbf {x} _ {3 \\mathrm {d}}\\right)}. \\tag {2}", + "image_path": "f5127f9abdfd310ff8ba01cdecbf5da50bba16fd074708997ce34a5f383ed1f0.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "Given " + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{J}" + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": ", the Jacobian of the affine projective transformation and " + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": ", a viewing transformation, the covariance ma" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 528, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 528, + 85 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 528, + 85 + ], + "type": "text", + "content": "trix " + }, + { + "bbox": [ + 313, + 72, + 528, + 85 + ], + "type": "inline_equation", + "content": "\\Sigma_{\\mathbf{3d}}^{\\prime}" + }, + { + "bbox": [ + 313, + 72, + 528, + 85 + ], + "type": "text", + "content": " in camera coordinates is defined as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 381, + 93, + 553, + 107 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 93, + 553, + 107 + ], + "spans": [ + { + "bbox": [ + 381, + 93, + 553, + 107 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\Sigma} _ {3 \\mathbf {d}} ^ {\\prime} = \\mathbf {J} \\mathbf {W} \\boldsymbol {\\Sigma} _ {3 \\mathbf {d}} \\mathbf {W} ^ {T} \\mathbf {J} ^ {T}. \\tag {3}", + "image_path": "5d8c22e62b1db418fa876e99a4492e142cc153a5ed8b0f2a07db68095c7f897b.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 118, + 553, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 118, + 553, + 166 + ], + "spans": [ + { + "bbox": [ + 313, + 118, + 553, + 166 + ], + "type": "text", + "content": "These 3D Gaussians are projected to 2D imaging plane and blended through a fast, differentiable " + }, + { + "bbox": [ + 313, + 118, + 553, + 166 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 118, + 553, + 166 + ], + "type": "text", + "content": "-blending process to render 2D images. The color " + }, + { + "bbox": [ + 313, + 118, + 553, + 166 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{3d}" + }, + { + "bbox": [ + 313, + 118, + 553, + 166 + ], + "type": "text", + "content": " of each pixel is computed using N-ordered 2D splats with the following formula:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 357, + 175, + 553, + 209 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 175, + 553, + 209 + ], + "spans": [ + { + "bbox": [ + 357, + 175, + 553, + 209 + ], + "type": "interline_equation", + "content": "\\mathbf {C} _ {\\mathbf {3 d}} = \\sum_ {i \\in \\mathbf {N}} \\alpha_ {\\mathbf {3 d} _ {i} ^ {\\prime}} \\mathbf {c} _ {\\mathbf {3 d} _ {i}} \\prod_ {j = 1} ^ {i - 1} \\left(1 - \\alpha_ {\\mathbf {3 d} _ {j} ^ {\\prime}}\\right), \\tag {4}", + "image_path": "8bf4f2c64f8f74c72e147ad0d46405b5f1c28339d4c275f695086ffdf164f080.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 218, + 553, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 218, + 553, + 280 + ], + "spans": [ + { + "bbox": [ + 313, + 218, + 553, + 280 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 218, + 553, + 280 + ], + "type": "inline_equation", + "content": "\\alpha_{\\mathbf{3d}_i'}" + }, + { + "bbox": [ + 313, + 218, + 553, + 280 + ], + "type": "text", + "content": " is the final opacity calculated by " + }, + { + "bbox": [ + 313, + 218, + 553, + 280 + ], + "type": "inline_equation", + "content": "\\alpha_{\\mathbf{3d}_i'} = \\alpha_{\\mathbf{3d}_i} * e^{-\\frac{1}{2} (\\mathbf{y}_{\\mathbf{3d}}' - \\mathbf{x}_{\\mathbf{3d}}')^T \\pmb{\\Sigma}_{\\mathbf{3d}}'^{-1} (\\mathbf{y}_{\\mathbf{3d}}' - \\mathbf{x}_{\\mathbf{3d}}')}" + }, + { + "bbox": [ + 313, + 218, + 553, + 280 + ], + "type": "text", + "content": ". And " + }, + { + "bbox": [ + 313, + 218, + 553, + 280 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathbf{3d}}'" + }, + { + "bbox": [ + 313, + 218, + 553, + 280 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 218, + 553, + 280 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_{\\mathbf{3d}}'" + }, + { + "bbox": [ + 313, + 218, + 553, + 280 + ], + "type": "text", + "content": " are coordinates in the projected space. Typically, the centroids of 3D Gaussians are initialized using the sparse SfM point clouds obtained from the set of input images." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 281, + 555, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 555, + 400 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 555, + 400 + ], + "type": "text", + "content": "2D Gaussian Splating (2DGS) is initially introduced by [48] in the task of implicit image representation. Since 2D Gaussians " + }, + { + "bbox": [ + 313, + 281, + 555, + 400 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{G}_{2d}\\}" + }, + { + "bbox": [ + 313, + 281, + 555, + 400 + ], + "type": "text", + "content": " are no longer oriented to 3D scenes but to 2D space, many bloated operations and redundant parameters are discarded, such as viewing transformation, spherical harmonics, etc. Concretely, a basic 2D Gaussian is described by its centroid " + }, + { + "bbox": [ + 313, + 281, + 555, + 400 + ], + "type": "inline_equation", + "content": "\\mathbf{x_{2d}}\\in \\mathbb{R}^2" + }, + { + "bbox": [ + 313, + 281, + 555, + 400 + ], + "type": "text", + "content": " , 2D covariance matrix " + }, + { + "bbox": [ + 313, + 281, + 555, + 400 + ], + "type": "inline_equation", + "content": "\\Sigma_{2\\mathbf{d}}\\in \\mathbb{R}^2" + }, + { + "bbox": [ + 313, + 281, + 555, + 400 + ], + "type": "text", + "content": " , color " + }, + { + "bbox": [ + 313, + 281, + 555, + 400 + ], + "type": "inline_equation", + "content": "\\mathbf{c_{2d}}\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 281, + 555, + 400 + ], + "type": "text", + "content": " and opacity " + }, + { + "bbox": [ + 313, + 281, + 555, + 400 + ], + "type": "inline_equation", + "content": "\\alpha_{2\\mathrm{d}}\\in \\mathbb{R}" + }, + { + "bbox": [ + 313, + 281, + 555, + 400 + ], + "type": "text", + "content": " . The 2D covariance matrix is also factorized into a rotation matrix " + }, + { + "bbox": [ + 313, + 281, + 555, + 400 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{2\\mathbf{d}}\\in \\mathbb{R}^{2\\times 2}" + }, + { + "bbox": [ + 313, + 281, + 555, + 400 + ], + "type": "text", + "content": " and a scale matrix " + }, + { + "bbox": [ + 313, + 281, + 555, + 400 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{2\\mathbf{d}}\\in \\mathbb{R}^2" + }, + { + "bbox": [ + 313, + 281, + 555, + 400 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 373, + 409, + 553, + 423 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 409, + 553, + 423 + ], + "spans": [ + { + "bbox": [ + 373, + 409, + 553, + 423 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\Sigma} _ {2 \\mathrm {d}} = \\mathbf {R} _ {2 \\mathrm {d}} \\mathbf {S} _ {2 \\mathrm {d}} \\mathbf {S} _ {2 \\mathrm {d}} ^ {T} \\mathbf {R} _ {2 \\mathrm {d}} ^ {T}. \\tag {5}", + "image_path": "1af9d495c07d7294df791a40b66646f4646749c585f31a4827b2230ad3182d3b.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 434, + 553, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 553, + 469 + ], + "type": "text", + "content": "For rasterization, since the depth and camera parameters are not included in 2D Gaussians, a simplified accumulated summation of N-ordered 2D splats is calculated as follows:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 388, + 480, + 553, + 505 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 480, + 553, + 505 + ], + "spans": [ + { + "bbox": [ + 388, + 480, + 553, + 505 + ], + "type": "interline_equation", + "content": "\\mathbf {C} _ {2 \\mathbf {d}} = \\sum_ {i \\in \\mathbf {N}} \\alpha_ {\\mathbf {2 d} _ {i} ^ {\\prime}} \\mathbf {c} _ {2 \\mathbf {d} _ {i}}, \\tag {6}", + "image_path": "209db5f97cf9f135ed8fd272d323edc0a69a4424b6ee5e7d97c9c8527bc395df.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 511, + 553, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 511, + 553, + 549 + ], + "spans": [ + { + "bbox": [ + 313, + 511, + 553, + 549 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 511, + 553, + 549 + ], + "type": "inline_equation", + "content": "\\alpha_{\\mathbf{2d}_i'}" + }, + { + "bbox": [ + 313, + 511, + 553, + 549 + ], + "type": "text", + "content": " is the opacity calculated by " + }, + { + "bbox": [ + 313, + 511, + 553, + 549 + ], + "type": "inline_equation", + "content": "\\alpha_{\\mathbf{2d}_i'} = \\alpha_{\\mathbf{2d}_i} * e^{-\\frac{1}{2} (\\mathbf{y}_{\\mathbf{2d}} - \\mathbf{x}_{\\mathbf{2d}})^T \\boldsymbol{\\Sigma}_{\\mathbf{2d}}^{-1} (\\mathbf{y}_{\\mathbf{2d}} - \\mathbf{x}_{\\mathbf{2d}})}" + }, + { + "bbox": [ + 313, + 511, + 553, + 549 + ], + "type": "text", + "content": ". And " + }, + { + "bbox": [ + 313, + 511, + 553, + 549 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathbf{2d}}" + }, + { + "bbox": [ + 313, + 511, + 553, + 549 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 511, + 553, + 549 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_{\\mathbf{2d}}" + }, + { + "bbox": [ + 313, + 511, + 553, + 549 + ], + "type": "text", + "content": " are coordinates of centroid and any pixel in the 2D image space." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 550, + 555, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 550, + 555, + 645 + ], + "spans": [ + { + "bbox": [ + 313, + 550, + 555, + 645 + ], + "type": "text", + "content": "Specifically, 2D Gaussians can be considered as projections of 3D Gaussians from a certain perspective, illustrating the relationship of Gaussians between 2D and 3D space. In our case, we combine the single-view independent 2D Gaussians and multi-view consistent 3D Gaussians to represent the entire 3D scene, which helps fundamentally decoupling transients and statics, based on the principles of multi-view geometry." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 657, + 370, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 657, + 370, + 669 + ], + "spans": [ + { + "bbox": [ + 313, + 657, + 370, + 669 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "We delineate the pipeline of our approach as illustrated in Fig. 2. Essentially, given a set of input images " + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\{I_k|k = 1,2,\\dots,N\\}" + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": " with the corresponding camera parameters, for" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "790" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 553, + 222 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 553, + 222 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 553, + 222 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 553, + 222 + ], + "type": "image", + "image_path": "164f9b07b66db5668985f6fa52c323c8018d9c859a9c11707c268a9e7c6635f2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 228, + 555, + 285 + ], + "lines": [ + { + "bbox": [ + 54, + 228, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 54, + 228, + 555, + 285 + ], + "type": "text", + "content": "Figure 2. Overview. Given a casually captured image sequence, we decompose the whole scene into 2D Gaussians for transient objects and 3D Gaussians for static scenes. To warm up, we start by training a basic 3DGS to capture static elements. This is followed by iterative training of 2D and 3D Gaussians, where our transients and statics are combined using an " + }, + { + "bbox": [ + 54, + 228, + 555, + 285 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 54, + 228, + 555, + 285 + ], + "type": "text", + "content": "-blending strategy with masks to produce the final renderings. The masks provide guidance for 3D Gaussians in the iterative training stage. During the joint-training, both 2D and 3D Gaussians are trained to further optimize the decomposition results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 305, + 295, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 305, + 295, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 305, + 295, + 331 + ], + "type": "text", + "content": "each view " + }, + { + "bbox": [ + 55, + 305, + 295, + 331 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 305, + 295, + 331 + ], + "type": "text", + "content": ", the goal of our method is to reasonably decouple the transients " + }, + { + "bbox": [ + 55, + 305, + 295, + 331 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 55, + 305, + 295, + 331 + ], + "type": "text", + "content": " and statics " + }, + { + "bbox": [ + 55, + 305, + 295, + 331 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 55, + 305, + 295, + 331 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 110, + 339, + 295, + 354 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 339, + 295, + 354 + ], + "spans": [ + { + "bbox": [ + 110, + 339, + 295, + 354 + ], + "type": "interline_equation", + "content": "I = M _ {t} \\odot I _ {t} + (1 - M _ {t}) \\odot I _ {s}, \\tag {7}", + "image_path": "840d7a2adb2535fadff88cab5ddd456ea828cbdf5a7d8ac59fbc8745a91f8be0.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 362, + 296, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 362, + 296, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 362, + 296, + 422 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 55, + 362, + 296, + 422 + ], + "type": "inline_equation", + "content": "M_t \\in [0,1]" + }, + { + "bbox": [ + 55, + 362, + 296, + 422 + ], + "type": "text", + "content": " represents the transient mask, and " + }, + { + "bbox": [ + 55, + 362, + 296, + 422 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 55, + 362, + 296, + 422 + ], + "type": "text", + "content": " is per pixel multiplication. When a pixel's value in " + }, + { + "bbox": [ + 55, + 362, + 296, + 422 + ], + "type": "inline_equation", + "content": "M_t" + }, + { + "bbox": [ + 55, + 362, + 296, + 422 + ], + "type": "text", + "content": " approaches 1, it indicates a higher probability that the location is transient; conversely, a lower value suggests a greater likelihood that area is static." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 423, + 295, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 423, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 54, + 423, + 295, + 590 + ], + "type": "text", + "content": "To achieve the goal, we decompose the whole 3D scene with two components: (1) Multi-view consistent 3D Gaussians is used for rendering " + }, + { + "bbox": [ + 54, + 423, + 295, + 590 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 54, + 423, + 295, + 590 + ], + "type": "text", + "content": ", which leverages multi-view information from images and models the static scene with a set of unified 3D Gaussians. The multi-view inputs regulate the 3D Gaussians to be consistent and robust across the different views. (2) Single-view independent 2D Gaussians is responsible for modeling " + }, + { + "bbox": [ + 54, + 423, + 295, + 590 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 54, + 423, + 295, + 590 + ], + "type": "text", + "content": ", which enables our approach to handle the fact that images are casually captured with varying transients. Concretely, we forms a set of view-independent 2D Gaussians to model transients as planar objects from a single view. This novel combination allows for a more precise and reasonable representation of 3D scenes, enabling better performance on novel views." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 590, + 296, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 296, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 296, + 639 + ], + "type": "text", + "content": "This section is organized as follows: we introduce our hybrid representation of scenes using 3D Gaussians in Sec. 4.1 and 2D Gaussians in Sec. 4.2, followed by a comprehensive overview of the training pipeline in Sec. 4.3." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 647, + 247, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 247, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 247, + 659 + ], + "type": "text", + "content": "4.1. Modeling Statics with 3D Gaussians" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Generally speaking, the initialized point clouds obtained by COLMAP [29, 30] can only represent areas in the scene that satisfy multi-view consistency, so 3D Gaussians tend to reconstruct static scenes, such as buildings, grounds, etc." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 305, + 553, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 305, + 553, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 305, + 553, + 389 + ], + "type": "text", + "content": "However, in scenes with transients, continuously training the 3D Gaussians may cause the transient objects to be overfitted into the Gaussian kernels because of constraints imposed by the RGB loss. This approach fails to capture the true geometry and instead overemphasizes the RGB information, leading to annoying artifacts from other viewpoints and ultimately impacting quality of novel view synthesis." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 389, + 554, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 554, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 554, + 556 + ], + "type": "text", + "content": "We address this problem by adopting an enhanced multiview regulated supervision scheme. Unlike previous methods that rasterize and supervise only a single image per iteration, our approach introduces two major changes. (1) We increase the number of images per iteration to " + }, + { + "bbox": [ + 313, + 389, + 554, + 556 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 389, + 554, + 556 + ], + "type": "text", + "content": ", enabling gradient back-propagation to simultaneously consider mutual information from multiple views. This batch-wise inputs allow each optimization step to leverage multi-view insights to distinguish and infer transients and static elements. (2) We adopt a sparse training schedule that focuses optimization only on the 3D Gaussians within the co-visible areas of the cross-view frustums. This method not only sharpens the training focus but also reduces computational costs, as detailed in Alg. 1, and outputs rasterized " + }, + { + "bbox": [ + 313, + 389, + 554, + 556 + ], + "type": "inline_equation", + "content": "\\hat{I}_s" + }, + { + "bbox": [ + 313, + 389, + 554, + 556 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 563, + 523, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 523, + 576 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 523, + 576 + ], + "type": "text", + "content": "4.2. Modeling Transients with 2D Gaussians" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "Different from other methods that utilize semantic features, we introduce 2D Gaussians to model the transients as planar objects from a single view for each image. This approach is motivated by several key considerations: (1) 2DGS has been proven to be an efficient method for representing images, capable of modeling an arbitrary image effectively. (2) The formation and rasterization processes of 2DGS closely resemble those of 3DGS, allowing both to be expressed within a unified framework. (3) Degenerating 3D objects that do not satisfy multi-view geometric consistency into 2D representation can effectively decouple transients from static" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "791" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 67, + 89, + 274, + 282 + ], + "blocks": [ + { + "bbox": [ + 63, + 75, + 262, + 87 + ], + "lines": [ + { + "bbox": [ + 63, + 75, + 262, + 87 + ], + "spans": [ + { + "bbox": [ + 63, + 75, + 262, + 87 + ], + "type": "text", + "content": "Algorithm 1: Multi-view Regulated Supervision." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 67, + 89, + 274, + 282 + ], + "lines": [ + { + "bbox": [ + 67, + 89, + 274, + 282 + ], + "spans": [ + { + "bbox": [ + 67, + 89, + 274, + 282 + ], + "type": "text", + "content": "Input: A set of images " + }, + { + "bbox": [ + 67, + 89, + 274, + 282 + ], + "type": "inline_equation", + "content": "\\{I_k, k = 1, 2, \\dots, N\\}" + }, + { + "bbox": [ + 67, + 89, + 274, + 282 + ], + "type": "text", + "content": "; Multi-view size " + }, + { + "bbox": [ + 67, + 89, + 274, + 282 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 67, + 89, + 274, + 282 + ], + "type": "text", + "content": "; Maximum iterations " + }, + { + "bbox": [ + 67, + 89, + 274, + 282 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 67, + 89, + 274, + 282 + ], + "type": "text", + "content": "; \nOutput: Optimized 3D Gaussian parameters " + }, + { + "bbox": [ + 67, + 89, + 274, + 282 + ], + "type": "inline_equation", + "content": "\\theta_{3d}" + }, + { + "bbox": [ + 67, + 89, + 274, + 282 + ], + "type": "text", + "content": " within the co-visible regions; \nInitialize " + }, + { + "bbox": [ + 67, + 89, + 274, + 282 + ], + "type": "inline_equation", + "content": "\\theta_{3d}" + }, + { + "bbox": [ + 67, + 89, + 274, + 282 + ], + "type": "text", + "content": "; \nfor each iteration " + }, + { + "bbox": [ + 67, + 89, + 274, + 282 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 67, + 89, + 274, + 282 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 67, + 89, + 274, + 282 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 67, + 89, + 274, + 282 + ], + "type": "text", + "content": " do \n\\[ I_{1,\\dots,K} \\leftarrow \\text{RandomSample}(K, N); \\quad F_{\\text{inter}} \\leftarrow \\text{ComputeFrustum}(I_1); \\quad \\text{for } i \\leftarrow 2 \\text{ to } K \\text{ do} \\quad |F_{\\text{inter}} \\leftarrow F_{\\text{inter}} \\cap \\text{ComputeFrustum}(I_i); \\quad \\text{end} \\quad \\mathcal{G}_{3d} \\leftarrow \\text{ExtractGaussians}(I, F_{\\text{inter}}); \\quad \\text{for each Gaussian } G_j \\in \\mathcal{G}_{3d} \\text{ do} \\quad |\\theta \\leftarrow \\theta - \\eta \\cdot \\nabla \\mathcal{L}(G_j(\\theta)); \\quad \\text{end} \\quad \\text{end}" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "code_body" + } + ], + "index": 1, + "sub_type": "algorithm" + }, + { + "bbox": [ + 55, + 309, + 180, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 309, + 180, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 309, + 180, + 319 + ], + "type": "text", + "content": "scenes on a fundamental level." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 321, + 295, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 321, + 295, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 321, + 295, + 357 + ], + "type": "text", + "content": "Specifically, 2D Gaussians are rasterized into an image " + }, + { + "bbox": [ + 55, + 321, + 295, + 357 + ], + "type": "inline_equation", + "content": "\\hat{I}_t" + }, + { + "bbox": [ + 55, + 321, + 295, + 357 + ], + "type": "text", + "content": " with a transient mask " + }, + { + "bbox": [ + 55, + 321, + 295, + 357 + ], + "type": "inline_equation", + "content": "\\hat{M}_t" + }, + { + "bbox": [ + 55, + 321, + 295, + 357 + ], + "type": "text", + "content": " for modeling uncertainty, each pixel " + }, + { + "bbox": [ + 55, + 321, + 295, + 357 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_{2\\mathbf{d}}" + }, + { + "bbox": [ + 55, + 321, + 295, + 357 + ], + "type": "text", + "content": " of which is calculated as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 368, + 295, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 368, + 295, + 380 + ], + "spans": [ + { + "bbox": [ + 62, + 368, + 295, + 380 + ], + "type": "interline_equation", + "content": "\\hat {M} _ {t} \\left(\\mathbf {y} _ {2 d}\\right) = \\sum_ {i \\in \\mathbf {N}} \\alpha_ {2 d i} ^ {\\prime} = \\sum_ {i \\in \\mathbf {N}} \\alpha_ {2 d i} \\cdot e ^ {- \\frac {1}{2} \\left(\\mathbf {y} _ {2 d} - \\mathbf {x} _ {2 d}\\right) ^ {T} \\boldsymbol {\\Sigma} _ {2 d} ^ {- 1} \\left(\\mathbf {y} _ {2 d} - \\mathbf {x} _ {2 d}\\right)}. \\tag {8}", + "image_path": "45f590ce8950fe2a98d1b444d3fac39e7d3076cc9c364c1f6b45f664e1ec4168.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 390, + 296, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 390, + 296, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 390, + 296, + 462 + ], + "type": "text", + "content": "It is noteworthy that we adopt uncertainty of 2D Gaussians to model the mask, rather than using 3D results. Because after modeling a static scene with 3D Gaussians, the 2D Gaussians are responsible for learning the residual part of the image. Therefore, naturally, the results rendered by the 2D Gaussian implicitly include a mask for transients." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 470, + 214, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 470, + 214, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 470, + 214, + 483 + ], + "type": "text", + "content": "4.3. Multi-stage Training Scheme" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 489, + 295, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 489, + 295, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 489, + 295, + 548 + ], + "type": "text", + "content": "Because 2D and 3D Gaussians are trained within the same framework, it is crucial to balance their relationship effectively. To that end, we have proposed a multi-stage training strategy that ensures stable convergence and enhances overall performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 549, + 295, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 549, + 295, + 633 + ], + "spans": [ + { + "bbox": [ + 55, + 549, + 295, + 633 + ], + "type": "text", + "content": "Warm up Pre-training. The training process begins by using 3DGS to capture the essential structure of the entire static scene. This stage focuses on modeling the scene with a set of unified 3D Gaussians with our multi-view regulated supervision as detailed in Sec. 4.1, producing foundational, albeit low-quality, rendered images " + }, + { + "bbox": [ + 55, + 549, + 295, + 633 + ], + "type": "inline_equation", + "content": "\\hat{I}_s" + }, + { + "bbox": [ + 55, + 549, + 295, + 633 + ], + "type": "text", + "content": " using a combination of DSSIM [38] and L1 losses following:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 61, + 642, + 295, + 657 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 642, + 295, + 657 + ], + "spans": [ + { + "bbox": [ + 61, + 642, + 295, + 657 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {w a r m u p} = \\lambda L _ {D S S I M} (\\hat {I} _ {s}, I) + (1 - \\lambda) L _ {1} (\\hat {I} _ {s}, I), \\quad (9)", + "image_path": "1ca1f2d76a98251d64ff384ff50ef1ab566351da758b166caf14f26d85b8c006.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 665, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 712 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 55, + 665, + 296, + 712 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 665, + 296, + 712 + ], + "type": "text", + "content": " is set to 0.2. During the warm-up, the densification of 3D Gaussians is conducted to gradually increase the number, allowing for a more detailed representation of the static scene structure." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": "Iterative Training. The second stage employs iterative training between 2DGS and 3DGS to progressively optimize both components. During the alternation, the 3DGS will provide more accurate static scene renderings, which in turn enables the 2DGS to refine its transient representations " + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "inline_equation", + "content": "\\hat{I}_t" + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": " and mask predictions " + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "inline_equation", + "content": "\\hat{M}_t" + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": " with Eq. 10. Conversely, the improved masks from 2DGS allow the 3DGS to focus more precisely on rendering the static regions of the scene with Eq. 11. The rendered image " + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "inline_equation", + "content": "\\hat{I}" + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": " is a composition of " + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "inline_equation", + "content": "\\hat{I}_s" + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "inline_equation", + "content": "\\hat{I}_t" + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "inline_equation", + "content": "\\hat{M}_t" + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": " following Eq. 7. This process iteratively refines the transients and statics, improving the overall quality and accuracy of the masks. The loss functions are as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 325, + 224, + 555, + 237 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 224, + 555, + 237 + ], + "spans": [ + { + "bbox": [ + 325, + 224, + 555, + 237 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {i t e r 2 d}} = \\lambda L _ {\\text {D S S I M}} (\\hat {I}, I) + (1 - \\lambda) L _ {1} (\\hat {I}, I), \\tag {10}", + "image_path": "35f143c3a6a2048fa37720646a7d1e2e7abb7a77e6bd28442e2a47a43238497d.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 365, + 257, + 553, + 271 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 257, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 365, + 257, + 553, + 271 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {i t e r 3 d}} = \\left(1 - \\hat {M} _ {t}\\right) \\odot \\mathcal {L} _ {\\text {i t e r 2 d}}. \\tag {11}", + "image_path": "ed39ae598420f48cadbc10061b8f0133be8a2d54c51a959dfe7177f09303b448.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 276, + 553, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 276, + 553, + 311 + ], + "spans": [ + { + "bbox": [ + 313, + 276, + 553, + 311 + ], + "type": "text", + "content": "Note that, during training, 2DGS and 3DGS are trained alternately. When training one branch, the gradient backpropagation of the other branch is turned off." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 312, + 554, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 554, + 383 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 554, + 383 + ], + "type": "text", + "content": "Joint Fine-tuning. The final stage involves joint fine-tuning, integrating 2DGS and 3DGS into a unified framework. This stage focuses on fine-tuning on the transient mask and reducing the error between the rendered image " + }, + { + "bbox": [ + 313, + 312, + 554, + 383 + ], + "type": "inline_equation", + "content": "\\hat{I}" + }, + { + "bbox": [ + 313, + 312, + 554, + 383 + ], + "type": "text", + "content": " and the ground truth image " + }, + { + "bbox": [ + 313, + 312, + 554, + 383 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 312, + 554, + 383 + ], + "type": "text", + "content": ". The loss " + }, + { + "bbox": [ + 313, + 312, + 554, + 383 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{joint}" + }, + { + "bbox": [ + 313, + 312, + 554, + 383 + ], + "type": "text", + "content": " used during the joint fine-tuning is a combination as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 326, + 392, + 553, + 406 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 392, + 553, + 406 + ], + "spans": [ + { + "bbox": [ + 326, + 392, + 553, + 406 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {j o i n t}} = \\beta L _ {D S S I M} (\\hat {I}, I) + (1 - \\beta) L _ {1} (\\hat {I}, I), \\tag {12}", + "image_path": "4dcbb5944649c3579b6ea4021529351db6b6a002679620fe227b9b71290bf6bc.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 413, + 554, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 554, + 450 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 554, + 450 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 313, + 413, + 554, + 450 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 413, + 554, + 450 + ], + "type": "text", + "content": " is set to 0.2. Here, both the 2DGS and 3DGS branches will enable gradient backpropagation to simultaneously update all parameters." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 460, + 395, + 473 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 395, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 395, + 473 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 479, + 430, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 430, + 493 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 430, + 493 + ], + "type": "text", + "content": "5.1. Experimental Setup" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 498, + 555, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 641 + ], + "type": "text", + "content": "Datasets. Similar to previous methods, we evaluate our HybridGS on two challenging datasets: NeRF On-the-go [25] and RobustNeRF [27]. NeRF On-the-go contains multiple casually captured indoor and outdoor sequences with varying ratios of distractors (from " + }, + { + "bbox": [ + 313, + 498, + 555, + 641 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 313, + 498, + 555, + 641 + ], + "type": "text", + "content": " to over " + }, + { + "bbox": [ + 313, + 498, + 555, + 641 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 313, + 498, + 555, + 641 + ], + "type": "text", + "content": "). We use the version of the dataset where all images were undistorted as [14]. The RobustNeRF dataset includes several scenes exemplifying different types of distractors. To simulate capturing over extended periods, distractor objects are repositioned between frames to interfere with statically placed objects (from 1 (Statue) to 150 (Yoda)). The splits include distractor-filled training set and distractor-free testing set." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "text", + "content": "Implementation Details. Our HybridGS, developed on top of open-source implementation gsplat [43] and Taming-3DGS [19], incorporates custom CUDA kernels, boosting the rasterization of " + }, + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "text", + "content": "-blending for 3DGS and weighted summation for 2DGS. The number of views " + }, + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "text", + "content": " is 4 in our practice. For a scene consisting of 100 images, the process" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "792" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 82, + 102, + 531, + 173 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 1. Comparison on NeRF On-the-go Dataset. For better visualization, the best results are highlighted in bold. Our method shows generally superior performance over state-of-the-art methods. * indicates that the results are reproduced from the official implementation." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 82, + 102, + 531, + 173 + ], + "lines": [ + { + "bbox": [ + 82, + 102, + 531, + 173 + ], + "spans": [ + { + "bbox": [ + 82, + 102, + 531, + 173 + ], + "type": "table", + "html": "
MethodLow OcclusionMedium OcclusionHigh Occlusion
MountainFountainCornerPatioSpotPatio-High
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
RobustNeRF [27]17.540.4960.38315.650.3180.57623.040.7640.24420.390.7180.25120.650.6250.39120.540.5780.366
NeRF On-the-go [25]20.150.6440.25920.110.6090.31424.220.8060.19020.780.7540.21923.330.7870.18921.410.7180.235
3DGS [11]19.400.6380.21319.960.6590.18520.900.7130.24117.480.7040.19920.770.6930.31617.290.6040.363
SLS-mlp* [28]19.840.5800.31520.190.6120.25924.030.7950.16621.550.8380.32823.520.7560.23520.310.6640.261
HybridGS(Ours)21.730.6930.22521.110.6740.18825.030.8470.07721.980.8120.11024.330.7940.08621.770.7410.142
", + "image_path": "8dc6d16ed1da621965b3a32ca680beeb55dce402ccb6bdadba86a4e2a0668eda.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 83, + 214, + 531, + 288 + ], + "blocks": [ + { + "bbox": [ + 55, + 183, + 555, + 206 + ], + "lines": [ + { + "bbox": [ + 55, + 183, + 555, + 206 + ], + "spans": [ + { + "bbox": [ + 55, + 183, + 555, + 206 + ], + "type": "text", + "content": "Table 2. Comparison on RobustNeRF Dataset. For better visualization, the best results are highlighted in bold. Our method significantly outperforms all baseline methods. Crab (1) scene had a test set with same set of views as those in the train set, which is fixed in Crab (2)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 83, + 214, + 531, + 288 + ], + "lines": [ + { + "bbox": [ + 83, + 214, + 531, + 288 + ], + "spans": [ + { + "bbox": [ + 83, + 214, + 531, + 288 + ], + "type": "table", + "html": "
MethodStatueAndroidYodaCrab (1)Crab (2)
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
RobustNeRF [27]20.600.760.1523.280.750.1329.780.820.1532.220.940.06---
NeRF On-the-go [25]21.580.770.2423.500.750.2129.960.830.24------
3DGS [11]21.020.810.1623.110.810.1326.330.910.1431.800.960.0829.74--
SLS-mlp [28]22.540.840.1325.050.850.0933.660.960.1035.850.970.0834.43--
HybridGS(Ours)22.930.870.1025.150.850.0735.320.960.0736.310.970.0535.170.960.08
", + "image_path": "6e1e74455c8458d43b8e8deb203e336ffd3620e81fd980f77ff3f64e3dc57757.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 57, + 338, + 296, + 392 + ], + "blocks": [ + { + "bbox": [ + 55, + 307, + 295, + 331 + ], + "lines": [ + { + "bbox": [ + 55, + 307, + 295, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 307, + 295, + 331 + ], + "type": "text", + "content": "Table 3. Ablation on components in datasets with varying occlusions. The best results are highlighted in bold." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 338, + 296, + 392 + ], + "lines": [ + { + "bbox": [ + 57, + 338, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 57, + 338, + 296, + 392 + ], + "type": "table", + "html": "
Dataset(ratio)On-the-go low.(5%)On-the-go medium.(17%)On-the-go high.(26%)
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Ours21.420.6840.20623.500.8270.09223.050.7680.115
w/o iterative Training21.390.6870.20923.110.8150.09422.550.7530.123
w/o Joint Training21.050.6540.24123.180.8150.10723.010.7480.130
w/o Multi-view21.020.6580.23323.060.8120.10322.710.7630.111
", + "image_path": "860441607ade0dc417a0101d00b0f84d0c11598e72e9245213200cb3e6425d1a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 422, + 295, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 422, + 295, + 505 + ], + "spans": [ + { + "bbox": [ + 54, + 422, + 295, + 505 + ], + "type": "text", + "content": "starts with a 1k-step warm-up. This is followed by iterative training, which includes 10k steps for 2DGS and 1k steps for 3DGS. The final joint training involves 30k steps. The entire training process takes approximately 0.18 GPU hours. The number of training steps scales proportionally with the size of the dataset. All experiments are performed on a single NVIDIA RTX 4090 GPU using PyTorch." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 506, + 295, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 506, + 295, + 531 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 295, + 531 + ], + "type": "text", + "content": "Metrics. We follow common practice by employing PSNR, SSIM [38], and LPIPS [47] for evaluation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 540, + 129, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 129, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 129, + 553 + ], + "type": "text", + "content": "5.2. Evaluation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 560, + 278, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 560, + 278, + 572 + ], + "spans": [ + { + "bbox": [ + 55, + 560, + 278, + 572 + ], + "type": "text", + "content": "5.2.1 Comparison on the NeRF On-the-go Dataset" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 580, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 580, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 580, + 295, + 653 + ], + "type": "text", + "content": "As shown in Tab. 1 and Fig. 3, our method demonstrates superior performance across different scenarios with varying levels of occlusion. In Tab. 1, our method outperforms the previous SOTA method across multiple metrics, with an average improvement of " + }, + { + "bbox": [ + 55, + 580, + 295, + 653 + ], + "type": "inline_equation", + "content": "1.10\\mathrm{dB}" + }, + { + "bbox": [ + 55, + 580, + 295, + 653 + ], + "type": "text", + "content": " in PSNR and " + }, + { + "bbox": [ + 55, + 580, + 295, + 653 + ], + "type": "inline_equation", + "content": "5.27\\%" + }, + { + "bbox": [ + 55, + 580, + 295, + 653 + ], + "type": "text", + "content": " in SSIM, demonstrating the superiority of our approach." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "Furthermore, Fig. 3 provides a qualitative comparison, illustrating the visual fidelity of our results against other methods. Our method produces more clear and detailed images, closely resembling the ground truth, particularly in challenging scenes with complex geometry and occlusions." + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 365, + 349, + 506, + 408 + ], + "blocks": [ + { + "bbox": [ + 313, + 307, + 555, + 340 + ], + "lines": [ + { + "bbox": [ + 313, + 307, + 555, + 340 + ], + "spans": [ + { + "bbox": [ + 313, + 307, + 555, + 340 + ], + "type": "text", + "content": "Table 4. Ablation on the number of 2D Gaussians of Corner in NeRF On-the-go dataset. For better visualization, the best results are highlighted in bold." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 365, + 349, + 506, + 408 + ], + "lines": [ + { + "bbox": [ + 365, + 349, + 506, + 408 + ], + "spans": [ + { + "bbox": [ + 365, + 349, + 506, + 408 + ], + "type": "table", + "html": "
NumberPSNR↑SSIM↑LPIPS↓Storage (MB)
2DGS3DGS
1k24.4930.8400.0773.542.1
5k24.7330.8430.08217.341.9
10k25.0340.8470.08034.742.2
15k24.5120.8390.09242.652
25k24.6980.8400.09469.442.2
", + "image_path": "eed9fe25a34d1f45d5399c477049cd6d252b93132066668e163b000906335e86.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 421, + 555, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 421, + 555, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 421, + 555, + 445 + ], + "type": "text", + "content": "This highlights our model's capability to effectively handle diverse visual conditions both in indoor and outdoor scenes." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 463, + 522, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 463, + 522, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 522, + 475 + ], + "type": "text", + "content": "5.2.2 Comparison on the RobustNeRF Dataset" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 482, + 555, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 482, + 555, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 482, + 555, + 662 + ], + "type": "text", + "content": "We present the quantitative results of RobustNeRF in Tab. 2. Our method demonstrates superior performance quantitatively compared to all previous approaches. In scenes with numerous transients, such as the Yoda scene, we've noticed that the performance metrics of the original 3DGS significantly degrade, occasionally even dropping below those of NeRF-related methods. By leveraging multi-view shared information and incorporating 2D Gaussians to effectively represent transients, our approach surpasses semantic-based methods like SLS-mlp [28]. As shown in Fig. 4, our results have clearer overall boundaries and structured fine-grained details, while SLS-mlp [28] may lead to blurriness in static objects and occluded areas. This further confirms that, even when faced with numerous transients, our method consistently maintains stable and reliable performance." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 670, + 416, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 670, + 416, + 682 + ], + "spans": [ + { + "bbox": [ + 313, + 670, + 416, + 682 + ], + "type": "text", + "content": "5.3. Ablation Studies" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "Components. To assess the effectiveness of each component in our framework, we performed an ablation study, as" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "793" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 68, + 553, + 261 + ], + "blocks": [ + { + "bbox": [ + 60, + 68, + 553, + 261 + ], + "lines": [ + { + "bbox": [ + 60, + 68, + 553, + 261 + ], + "spans": [ + { + "bbox": [ + 60, + 68, + 553, + 261 + ], + "type": "image", + "image_path": "e33afbb0bdf4907ca393059a9b5759b59034735f10b9ffb421385dd0c15538d6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 269, + 555, + 304 + ], + "lines": [ + { + "bbox": [ + 55, + 269, + 555, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 269, + 555, + 304 + ], + "type": "text", + "content": "Figure 3. Visualization of novel view synthesis results on the testing set of NeRF On-the-go dataset. Our method demonstrates superior results by effectively reducing artifacts and providing clearer boundaries. This results in a cleaner statics compared to other methods, showcasing enhanced visual quality and precision in novel views." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 57, + 323, + 294, + 423 + ], + "blocks": [ + { + "bbox": [ + 57, + 323, + 294, + 423 + ], + "lines": [ + { + "bbox": [ + 57, + 323, + 294, + 423 + ], + "spans": [ + { + "bbox": [ + 57, + 323, + 294, + 423 + ], + "type": "image", + "image_path": "9dfe4fd9428dfd8798707caf0a76b15c3988f2de6b6531b144801a2a0e452196.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 433, + 295, + 455 + ], + "lines": [ + { + "bbox": [ + 55, + 433, + 295, + 455 + ], + "spans": [ + { + "bbox": [ + 55, + 433, + 295, + 455 + ], + "type": "text", + "content": "Figure 4. Visualization of novel view synthesis results on RobustNeRF dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 483, + 295, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 483, + 295, + 674 + ], + "spans": [ + { + "bbox": [ + 54, + 483, + 295, + 674 + ], + "type": "text", + "content": "presented in Tab. 3. We systematically removed iterative training, joint training, and multi-view supervision to evaluate their effectiveness. Each component, as expected, enhances overall performance, demonstrating their individual and combined significance. In scenes with low occlusion, multi-view supervision allows the model to concentrate on static elements by integrating data from various perspectives. Conversely, in highly occluded environments, joint training with 2D and 3D Gaussians markedly improves performance, highlighting the model's capability to manage complex scenes with significant occlusions and transient changes. Iterative training consistently boosts results across different occlusion levels. In summary, the results underscore that integrating diverse training strategies and multiview information significantly enhances the rendering quality of unseen views, especially in challenging conditions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "The Number of 2D Gaussians. Tab. 4 presents an ablation study on the number of 2D Gaussians used in the corner sequence of NeRF On-the-go dataset. The results indicate that" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 317, + 323, + 553, + 421 + ], + "blocks": [ + { + "bbox": [ + 317, + 323, + 553, + 421 + ], + "lines": [ + { + "bbox": [ + 317, + 323, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 317, + 323, + 553, + 421 + ], + "type": "image", + "image_path": "d01dbc1fb34ae3d3635214eaf0fd2c6a95d36a5d457c989c044930dd1eb34234.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 429, + 555, + 450 + ], + "lines": [ + { + "bbox": [ + 313, + 429, + 555, + 450 + ], + "spans": [ + { + "bbox": [ + 313, + 429, + 555, + 450 + ], + "type": "text", + "content": "Figure 5. Comparison of transient masks on NeRF On-the-go dataset." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 473, + 555, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 555, + 544 + ], + "type": "text", + "content": "increasing the number of 2D Gaussians initially improves the performance. Specifically, 10k 2D Gaussians achieve the highest PSNR (25.034) and SSIM (0.847), while maintaining a competitive LPIPS score of 0.151 and reasonable storage requirements. Based on these findings, we select 10k as the optimal number in our experiments." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 545, + 556, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 556, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 556, + 676 + ], + "type": "text", + "content": "Transients and Masks. Benefiting from our method's explicit modeling of transient objects, we can obtain transient masks without introducing any segmentation networks or other pre-trained features. Typically, these masks capture dynamic elements such as pedestrians and vehicles, as well as some structural details ignored by 3D Gaussians. In Fig. 6, our approach effectively learns the RGB and mask of transients even under strong occlusions, while maintaining the robustness of static elements. This capability underscores the strength of our method in discerning and isolating dynamic components within complex scenes." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "We also compare our transient masks with SLS-mlp [28] semantic masks in Fig. 5. It is observed that the semantic mask does not perfectly align with the transients due to" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "794" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 553, + 357 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 553, + 357 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 553, + 357 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 553, + 357 + ], + "type": "image", + "image_path": "cf5966e71b1ac324ab8b0c069f467e7624bd96a01358cb122fdc01bf4ea115e4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 369, + 555, + 403 + ], + "lines": [ + { + "bbox": [ + 55, + 369, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 555, + 403 + ], + "type": "text", + "content": "Figure 6. Visualization of scene decomposition into transients and statics. Our method achieves superior transient mask separation in both indoor and outdoor scenes. It effectively separates transients and statics, and the resulting renderings closely resemble the ground truth images, demonstrating its effectiveness. Top two rows are from NeRF On-the-go [25], bottom two rows are from RobustNeRF [27]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 422, + 293, + 506 + ], + "blocks": [ + { + "bbox": [ + 58, + 422, + 293, + 506 + ], + "lines": [ + { + "bbox": [ + 58, + 422, + 293, + 506 + ], + "spans": [ + { + "bbox": [ + 58, + 422, + 293, + 506 + ], + "type": "image", + "image_path": "d0e6b63f56ad97a30fb02649652e09c61d5c5c8befb4d3741942aae023fde704.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 80, + 515, + 271, + 526 + ], + "lines": [ + { + "bbox": [ + 80, + 515, + 271, + 526 + ], + "spans": [ + { + "bbox": [ + 80, + 515, + 271, + 526 + ], + "type": "text", + "content": "Figure 7. Visualization on Photo Tourism dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 551, + 295, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 551, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 551, + 295, + 624 + ], + "type": "text", + "content": "the semantic features not being sufficiently precise. Additionally, transients often encompass low-level information, such as motion blur or object shadows, which may not explicitly exhibit clear semantic features. By modeling these transients using 2D Gaussians at the pixel level, our method allows for a more fine-grained capture of transients." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 634, + 132, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 634, + 132, + 645 + ], + "spans": [ + { + "bbox": [ + 55, + 634, + 132, + 645 + ], + "type": "text", + "content": "5.4. Discussions" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "While our method has demonstrated strong performance in distinguishing between transients and statics, it has not yet accounted for the illumination variation in unconstrained photo collections. As illustrated in Fig. 7 on Photo Tourism [32], our method renders fine-grained architectural" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 423, + 555, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 555, + 471 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 555, + 471 + ], + "type": "text", + "content": "structures, but the overall photometric result tends towards an average photometric effect, akin to overcast conditions. A promising direction is to incorporate an appearance embedding module in our method for the future work." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 487, + 388, + 500 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 487, + 388, + 500 + ], + "spans": [ + { + "bbox": [ + 314, + 487, + 388, + 500 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 510, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 715 + ], + "type": "text", + "content": "In this paper, we introduce HybridGS, a novel hybrid representation with 2D and 3D Gaussians for decomposing 3D scenes from casually captured images. By introducing the hybrid Gaussian representation, our method addresses the challenges of handling transients as planar objects while preserving the integrity of the static scene. The multi-view supervision mechanism we developed plays a crucial role in enhancing the discrimination between static and transient elements across multiple views, reducing artifacts and inaccuracies in novel views. We have developed a multi-stage training strategy for the joint training of 2D and 3D Gaussians, ensuring the stable and robust training stage. Our experiments demonstrate that HybridGS significantly outperforms existing methods in terms of quality and efficiency. In summary, HybridGS establishes a novel, robust and efficient foundation for 3D scene representation, paving the way for advancements in various vision-based applications." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "795" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "text", + "content": "[1] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. ICCV, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 136, + 296, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 136, + 296, + 168 + ], + "spans": [ + { + "bbox": [ + 61, + 136, + 296, + 168 + ], + "type": "text", + "content": "[2] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. CVPR, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 171, + 296, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 171, + 296, + 225 + ], + "spans": [ + { + "bbox": [ + 61, + 171, + 296, + 225 + ], + "type": "text", + "content": "[3] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. Pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19457-19467, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 227, + 296, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 227, + 296, + 281 + ], + "spans": [ + { + "bbox": [ + 62, + 227, + 296, + 281 + ], + "type": "text", + "content": "[4] Jiahao Chen, Yipeng Qin, Lingjie Liu, Jiangbo Lu, and Guanbin Li. Nerf-hugs: Improved neural radiance fields in non-static scenes using heuristics-guided segmentation. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19436-19446, 2024. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 284, + 296, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 284, + 296, + 337 + ], + "spans": [ + { + "bbox": [ + 62, + 284, + 296, + 337 + ], + "type": "text", + "content": "[5] Xingyu Chen, Qi Zhang, Xiaoyu Li, Yue Chen, Feng Ying, Xuan Wang, and Jue Wang. Hallucinated neural radiance fields in the wild. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12933-12942, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 339, + 296, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 339, + 296, + 382 + ], + "spans": [ + { + "bbox": [ + 62, + 339, + 296, + 382 + ], + "type": "text", + "content": "[6] Zilong Chen, Feng Wang, and Huaping Liu. Text-to-3d using gaussian splatting. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21401–21412, 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 385, + 296, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 385, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 62, + 385, + 296, + 407 + ], + "type": "text", + "content": "[7] Hiba Dahmani. Swag: Splitting in the wild images with appearance-conditioned gaussians. In ECCV, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 408, + 296, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 408, + 296, + 463 + ], + "spans": [ + { + "bbox": [ + 62, + 408, + 296, + 463 + ], + "type": "text", + "content": "[8] Sara Fridovich-Keil, Giacomo Meanti, Frederik Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12479-12488, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 464, + 296, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 464, + 296, + 508 + ], + "spans": [ + { + "bbox": [ + 62, + 464, + 296, + 508 + ], + "type": "text", + "content": "[9] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In SIGGRAPH 2024 Conference Papers. Association for Computing Machinery, 2024. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 510, + 295, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 510, + 295, + 553 + ], + "spans": [ + { + "bbox": [ + 57, + 510, + 295, + 553 + ], + "type": "text", + "content": "[10] Karim Kassab, Antoine Schnepf, Jean-Yves Franceschi, Laurent Caraffa, Jeremie Mary, and Valérie Gouet-Brunet. Refinedfields: Radiance fields refinement for unconstrained scenes. arXiv preprint arXiv:2312.00639, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 555, + 295, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 555, + 295, + 599 + ], + "spans": [ + { + "bbox": [ + 57, + 555, + 295, + 599 + ], + "type": "text", + "content": "[11] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuehler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (TOG), 42:1-14, 2023. 1, 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 601, + 295, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 601, + 295, + 655 + ], + "spans": [ + { + "bbox": [ + 57, + 601, + 295, + 655 + ], + "type": "text", + "content": "[12] Bernhard Kerbl, Andreas Meuleman, Georgios Kopanas, Michael Wimmer, Alexandre Lanvin, and George Drettakis. A hierarchical 3d gaussian representation for real-time rendering of very large datasets. ACM Transactions on Graphics, 43(4), 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 657, + 295, + 699 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 657, + 295, + 699 + ], + "spans": [ + { + "bbox": [ + 57, + 657, + 295, + 699 + ], + "type": "text", + "content": "[13] Jonávs. Kulhánek and Torsten Sattler. Tetra-nerf: Representing neural radiance fields using tetrahedra. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 18412–18423, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 702, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 702, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 702, + 295, + 713 + ], + "type": "text", + "content": "[14] Jonas Kulhanek, Songyou Peng, Zuzana Kukelova, Marc" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "Pollefeys, and Torsten Sattler. Wildgaussians: 3d gaussian splatting in the wild. 2024. 3, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 96, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 128 + ], + "type": "text", + "content": "[15] Pengzhi Li, Chengshuai Tang, Qinxuan Huang, and Zhiheng Li. Art3d: 3d gaussian splatting for text-guided artistic scenes generation, 2024. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "type": "text", + "content": "[16] Zhihao Liang, Qi Zhang, Yingfa Feng, Ying Shan, and Kui Jia. Gs-ir: 3d gaussian splatting for inverse rendering. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21644–21653, 2023. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 174, + 553, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 553, + 238 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 553, + 238 + ], + "type": "text", + "content": "[17] Jiaqi Lin, Zhihao Li, Xiao Tang, Jianzhuang Liu, Shiyong Liu, Jiayue Liu, Yangdi Lu, Xiaofei Wu, Songcen Xu, Youliang Yan, and Wenming Yang. Vastgaussian: Vast 3d gaussians for large scene reconstruction. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5166-5175, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 239, + 555, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 555, + 293 + ], + "type": "text", + "content": "[18] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20654-20664, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 294, + 553, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 294, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 317, + 294, + 553, + 360 + ], + "type": "text", + "content": "[19] Saswat Subhajyoti Mallick, Rahul Goel, Bernhard Kerbl, Markus Steinberger, Francisco Vicente Carrasco, and Fernando De La Torre. Taming 3dgs: High-quality radiance fields with limited resources. In SIGGRAPH Asia 2024 Conference Papers, New York, NY, USA, 2024. Association for Computing Machinery. 2, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 361, + 553, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 361, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 317, + 361, + 553, + 425 + ], + "type": "text", + "content": "[20] Ricardo Martin-Brualla, Noha Radwan, Mehdi S. M. Sajjadi, Jonathan T. Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7206-7215, 2020. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 427, + 553, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 427, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 316, + 427, + 553, + 469 + ], + "type": "text", + "content": "[21] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 472, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 472, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 317, + 472, + 553, + 514 + ], + "type": "text", + "content": "[22] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 515, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 515, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 317, + 515, + 553, + 624 + ], + "type": "text", + "content": "[23] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINOV2: Learning robust visual features without supervision. Transactions on Machine Learning Research, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 625, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 657 + ], + "type": "text", + "content": "[24] Konstantinos Rematas, Andrew Liu, Pratul P. Srinivasan, Jonathan T. Barron, Andrea Tagliasacchi, Tom Funkhouser, and Vittorio Ferrari. Urban radiance fields. CVPR, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "text", + "content": "[25] Weining Ren, Zihan Zhu, Boyang Sun, Jiaqi Chen, Marc Pollefeys, and Songyou Peng. Nerf on-the-go: Exploiting uncertainty for distractor-free nerfs in the wild. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "796" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[26] Barbara Roessle, Norman Müller, Lorenzo Porzi, Samuel Rota Bulò, Peter Kontschieder, Angela Dai, and Matthias Nießner. L3dg: Latent 3d gaussian diffusion. In SIGGRAPH Asia 2024 Conference Papers, 2024. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 172 + ], + "type": "text", + "content": "[27] Sara Sabour, Suhani Vora, Daniel Duckworth, Ivan Krasin, David J. Fleet, and Andrea Tagliasacchi. Robustnerf: Ignoring distractors with robust losses. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20626-20636, 2023. 1, 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 173, + 294, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 173, + 294, + 226 + ], + "spans": [ + { + "bbox": [ + 56, + 173, + 294, + 226 + ], + "type": "text", + "content": "[28] Sara Sabour, Lily Goli, George Kopanas, Mark Matthews, Dmitry Lagun, Leonidas Guibas, Alec Jacobson, David J. Fleet, and Andrea Tagliasacchi. SpotLessSplats: Ignoring distractors in 3d gaussian splatting. arXiv:2406.20055, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 228, + 294, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 294, + 260 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 294, + 260 + ], + "type": "text", + "content": "[29] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 261, + 294, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 294, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 294, + 304 + ], + "type": "text", + "content": "[30] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In European Conference on Computer Vision (ECCV), 2016. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 305, + 294, + 359 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 305, + 294, + 359 + ], + "spans": [ + { + "bbox": [ + 56, + 305, + 294, + 359 + ], + "type": "text", + "content": "[31] Jin-Chuan Shi, Miao Wang, Hao-Bin Duan, and Shao-Hua Guan. Language embedded 3d gaussians for open-vocabulary scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5333-5343, 2024. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 361, + 294, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 294, + 403 + ], + "type": "text", + "content": "[32] Noah Snavely, Steven M. Seitz, and Richard Szeliski. Photo tourism: Exploring photo collections in 3d. In SIGGRAPH Conference Proceedings, pages 835-846, New York, NY, USA, 2006. ACM Press. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 404, + 294, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 294, + 468 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 294, + 468 + ], + "type": "text", + "content": "[33] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P. Srinivasan, Jonathan T. Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8238–8248, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 471, + 294, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 294, + 536 + ], + "type": "text", + "content": "[34] Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Justin Kerr, Terrance Wang, Alexander Kristoffersen, Jake Austin, Kamyar Salahi, Abhik Ahuja, David McAllister, and Angjoo Kanazawa. Nerfstudio: A modular framework for neural radiance field development. ACM SIGGRAPH 2023 Conference Proceedings, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 537, + 294, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 294, + 579 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 294, + 579 + ], + "type": "text", + "content": "[35] Linhan Wang, Kai Cheng, Shuo Lei, Shengkun Wang, Wei Yin, Chenyang Lei, Xiaoxiao Long, and Chang-Tien Lu. Dcgaussian: Improving 3d gaussian splatting for reflective dash cam videos. NeurIPS 2024, 2024. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 581, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 581, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 581, + 294, + 613 + ], + "type": "text", + "content": "[36] Shuaixian Wang, Haoran Xu, Yaokun Li, Jiwei Chen, and Guang Tan. Ie-nerf: Inpainting enhanced neural radiance fields in the wild. ArXiv, abs/2407.10695, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 614, + 294, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 294, + 646 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 294, + 646 + ], + "type": "text", + "content": "[37] Yuze Wang, Junyi Wang, and Yue Qi. We-gs: An in-the-wild efficient 3d gaussian representation for unconstrained photo collections, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 647, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 690 + ], + "type": "text", + "content": "[38] Zhou Wang, Alan Conrad Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13:600-612, 2004. 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "text", + "content": "[39] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 687 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 116 + ], + "type": "text", + "content": "4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20310-20320, 2024. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 172 + ], + "type": "text", + "content": "[40] Tianhao Walter Wu, Fangcheng Zhong, Andrea Tagliasacchi, Forrester Cole, and Cengiz Oztireli. D" + }, + { + "bbox": [ + 316, + 118, + 553, + 172 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 316, + 118, + 553, + 172 + ], + "type": "text", + "content": "neRF: Self-supervised decoupling of dynamic and static objects from a monocular video. In Advances in Neural Information Processing Systems, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 174, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 553, + 217 + ], + "type": "text", + "content": "[41] Congrong Xu, Justin Kerr, and Angjoo Kanazawa. Splatfacto-w: A nefstudio implementation of gaussian splatting for unconstrained photo collections. ArXiv, abs/2407.12306, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 219, + 553, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 273 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 273 + ], + "type": "text", + "content": "[42] Yifan Yang, Shuhai Zhang, Zixiong Huang, Yubing Zhang, and Mingkui Tan. Cross-ray neural radiance fields for novel-view synthesis from unconstrained image collections. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 15855-15865, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 274, + 553, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 274, + 553, + 328 + ], + "spans": [ + { + "bbox": [ + 316, + 274, + 553, + 328 + ], + "type": "text", + "content": "[43] Vickie Ye, Ruilong Li, Justin Kerr, Matias Turkulainen, Brent Yi, Zhuoyang Pan, Otto Seiskari, Jianbo Ye, Jeffrey Hu, Matthew Tancik, and Angjoo Kanazawa. gsplat: An open-source library for Gaussian splatting. arXiv preprint arXiv:2409.06765, 2024. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 331, + 553, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 384 + ], + "type": "text", + "content": "[44] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19447-19456, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 386, + 553, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 386, + 553, + 419 + ], + "spans": [ + { + "bbox": [ + 316, + 386, + 553, + 419 + ], + "type": "text", + "content": "[45] Zehao Yu, Torsten Sattler, and Andreas Geiger. Gaussian opacity fields: Efficient adaptive surface reconstruction in unbounded scenes. ACM Transactions on Graphics, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 420, + 553, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 420, + 553, + 463 + ], + "spans": [ + { + "bbox": [ + 316, + 420, + 553, + 463 + ], + "type": "text", + "content": "[46] Dongbin Zhang, Chuming Wang, Weitao Wang, Peihao Li, Minghan Qin, and Haoqian Wang. Gaussian in the wild: 3d gaussian splatting for unconstrained image collections. ECCV, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 464, + 553, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 464, + 553, + 518 + ], + "spans": [ + { + "bbox": [ + 316, + 464, + 553, + 518 + ], + "type": "text", + "content": "[47] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 586-595, 2018. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 520, + 553, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 520, + 553, + 575 + ], + "spans": [ + { + "bbox": [ + 316, + 520, + 553, + 575 + ], + "type": "text", + "content": "[48] Xinjie Zhang, Xingtong Ge, Tongda Xu, Dailan He, Yan Wang, Hongwei Qin, Guo Lu, Jing Geng, and Jun Zhang. Gaussianimage: 1000 fps image representation and compression by 2d gaussian splatting. In European Conference on Computer Vision, 2024. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 576, + 553, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 576, + 553, + 620 + ], + "spans": [ + { + "bbox": [ + 316, + 576, + 553, + 620 + ], + "type": "text", + "content": "[49] Yunxiang Zhang, Alexandr Kuznetsov, Akshay Jindal, Kenneth Chen, Anton Sochenov, Anton Kaplanyan, and Qi Sun. Image-gs: Content-adaptive image representation via 2d gaussians, 2024. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 621, + 553, + 687 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 621, + 553, + 687 + ], + "spans": [ + { + "bbox": [ + 316, + 621, + 553, + 687 + ], + "type": "text", + "content": "[50] Xiaoyu Zhou, Zhiwei Lin, Xiaojun Shan, Yongtao Wang, Deqing Sun, and Ming-Hsuan Yang. Drivinggaussian: Composite gaussian splatting for surrounding dynamic autonomous driving scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21634-21643, 2024. 1" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "797" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/HybridMQA_ Exploring Geometry-Texture Interactions for Colored Mesh Quality Assessment/a45af5e0-e0d7-4402-9367-2386d583248b_content_list.json b/2025/HybridMQA_ Exploring Geometry-Texture Interactions for Colored Mesh Quality Assessment/a45af5e0-e0d7-4402-9367-2386d583248b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..48e9f51b7026d110e6efca6647bf776ec2bc2c14 --- /dev/null +++ b/2025/HybridMQA_ Exploring Geometry-Texture Interactions for Colored Mesh Quality Assessment/a45af5e0-e0d7-4402-9367-2386d583248b_content_list.json @@ -0,0 +1,1605 @@ +[ + { + "type": "text", + "text": "HybridMQA: Exploring Geometry-Texture Interactions for Colored Mesh Quality Assessment", + "text_level": 1, + "bbox": [ + 125, + 128, + 872, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Armin Shafiee Sarvestani,\\* Sheyang Tang\\*, Zhou Wang \nDepartment of Electrical and Computer Engineering, University of Waterloo \n{a5shafie, sheyang.tang, zhou.wang}@uwaterloo.ca", + "bbox": [ + 199, + 203, + 800, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mesh quality assessment (MQA) models play a critical role in the design, optimization, and evaluation of mesh operation systems in a wide variety of applications. Current MQA models, whether model-based methods using topology-aware features or projection-based approaches working on rendered 2D projections, often fail to capture the intricate interactions between texture and 3D geometry. We introduce HybridMQA, a first-of-its-kind hybrid full-reference colored MQA framework that integrates model-based and projection-based approaches, capturing complex interactions between textural information and 3D structures for enriched quality representations. Our method employs graph learning to extract detailed 3D representations, which are then projected to 2D using a novel feature rendering process that precisely aligns them with colored projections. This enables the exploration of geometry-texture interactions via cross-attention, producing comprehensive mesh quality representations. Extensive experiments demonstrate HybridMQA's superior performance across diverse datasets, highlighting its ability to effectively leverage geometry-texture interactions for a thorough understanding of mesh quality. Our project website is available at https://arshafiee.github.io/hybridmqa/.", + "bbox": [ + 89, + 321, + 483, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 700, + 220, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Advancements in 3D capture and display technologies have sparked a growing interest in immersive media. 3D meshes, a key form of 3D media, are widely used in applications like virtual and augmented reality [20, 49], gaming, animation, medical modeling [37], and generative 3D content creation [21, 35]. A mesh, comprising triangular faces formed by vertices, is colorized by assigning RGB colors to each vertex (vertex-color mesh) or applying a 2D texture map with UV coordinates (textured mesh). The demand for colored meshes calls for high-quality acquisition [3], compress-", + "bbox": [ + 89, + 724, + 482, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/17d7cf6dddfddddcc61e1e1c4dd123aa3c5a4a950c08ffd82e3a1e5ddd0b7c53.jpg", + "image_caption": [ + "Figure 1. Interactions between texture and geometry. top: complex texture makes the geometry distortion imperceptible. bottom: modifying the geometry affects the appearance of texture distortion. Right half (gray) of each object represents geometry." + ], + "image_footnote": [], + "bbox": [ + 521, + 292, + 901, + 465 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sion [11, 24], and transmission [25]. However, these processes often introduce artifacts that degrade visual quality, highlighting the need for robust mesh quality assessment (MQA) methods. Full-reference (FR) approaches, which take distorted meshes and their pristine references as input and generate a quality score by comparing their visual quality, are essential for accurate mesh quality assessment.", + "bbox": [ + 511, + 564, + 906, + 671 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The perceived quality of a 3D colored mesh is affected by its geometry and texture. Different mesh processing operations cause diverse geometrical and texture distortions that degrade the visual quality of meshes by perturbing the interactions between the object's shape and color. Figure 1 illustrates an example of such geometry-texture interactions, where either the geometry or texture can affect the visual appearance of distortions in the other. In the top row, the easily visible geometry distortion becomes imperceptible when we replace the uniform texture with a complex one. In the bottom row, the noisy texture patterns become smooth when we modify the geometry. This highlights the need for MQA methods that capture these complex geometry-texture interaction—a significant factor largely overlooked by existing methods.", + "bbox": [ + 511, + 674, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution", + "bbox": [ + 109, + 887, + 217, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21414", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3087de5fca1eec8884553d404e47aef7d5dccc11af4dbe186a1baafcf47920ad.jpg", + "image_caption": [ + "Figure 2. Reference and distorted meshes under geometry distortions. Although the distorted meshes (hawk and bowl) have distinct visual qualities (different mean opinion scores (MOS)), Graphics-LPIPS [27] assigns similar scores and reverses their ranking. HybridMQA aligns well with human perception as it understands the mesh's geometry properties. Blue boxes denote the same regions across viewpoints." + ], + "image_footnote": [], + "bbox": [ + 101, + 88, + 477, + 252 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "MQA methods may be generally categorized as model-based and projection-based depending on their operation space. Model-based MQA operates directly in the 3D vertex space to extract topology-aware quality features [18, 42]. Their performance is limited due to the lack of access to the object's actual appearance (rendered projections). To compensate, they either (1) apply image quality assessment (IQA) to 2D texture maps [13, 38]; or (2) define color quality descriptors for vertices or faces [28, 48, 53]. However, both fail to capture the final appearance human viewers see. Projection-based MQA, on the other hand, operates on rendered 2D projections [19, 27, 55], hence more effective at assessing texture distortions. Nonetheless, they lack a 3D understanding of the object's topology, which is crucial for accurate quality evaluation. Although underlying geometry distortions may be less noticeable in certain projections (Fig. 1, top row), they become apparent to human viewers when the object is viewed from different angles. Humans intuitively perceive underlying structures: connecting viewpoints, incorporating 3D cues, and detecting distortions that may be imperceptible in isolated 2D projections [6, 31, 57]. This 3D understanding is hard to obtain with projections alone due to the nature of operation space. Fig. 2 illustrates this: the geometry distortion (perturbing vertex positions) is more noticeable along the contours of the hawk and bowl than within their inner regions in 2D projections. However, humans naturally identify viewpoint connections (blue boxes) as objects rotate in 3D, perceiving geometry distortions beneath the texture. Consequently, projection-based MQA metrics (Graphics-LPIPS [27]), which focus on pictorial differences, struggle to detect and penalize these distortions, resulting in wrong rankings (Fig. 2). In contrast, HybridMQA extracts quality-related features from mesh surfaces in 3D, enabling a more accurate MQA.", + "bbox": [ + 91, + 387, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a novel hybrid FR MQA method, namely HybridMQA, that integrates model- and projection-based approaches to explore geometry-texture interactions for comprehensive MQA. Specifically, a base encoder extracts initial 3D features from 2D texture, normal, and vertex maps, which initialize a feature graph based on mesh connectivity. A graph convolutional network (GCN) then learns detailed 3D surface representations, building a 3D understanding of the object's topology. These surface representations are then complemented by textural information from the mesh's renderings, enabling a hybrid integration of model- and projection-based methods across all mesh data modalities: 2D maps, 3D structure, and colored renderings. Additionally, to explore the intricate geometry-texture interactions between the two operational spaces, we propose a novel feature projection technique that renders 2D projections of surface representations from the graph, precisely aligned with colored renderings. This alignment facilitates the exploration of geometry-texture interactions via cross-attention. Below are our contributions:", + "bbox": [ + 511, + 90, + 903, + 392 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose the first hybrid FR MQA method that consolidates the strengths of model- and projection-based approaches across all mesh data modalities, enabling a comprehensive understanding of mesh quality.", + "- We make the first attempt to explore geometry-texture interactions for MQA, drawing meaningful connections between the two domains. Our method shows the importance of leveraging such interactions to build reliable MQA methods.", + "- We propose a novel feature projection technique to render 2D projections of 3D surface representations from graph, aligned with the mesh's colored renderings, establishing pixel-to-pixel correspondence to explore interactions.", + "- Our model outperforms state-of-the-art FR MQA methods, aligning with human perception and generalizing better. The results highlight the effectiveness of capturing geometry-texture interactions to achieve accurate MQA." + ], + "bbox": [ + 513, + 393, + 903, + 650 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 667, + 661, + 684 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Model-based Mesh Quality Assessment", + "text_level": 1, + "bbox": [ + 511, + 695, + 849, + 710 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Model-based MQA methods operate directly in 3D, with most existing methods designed for uncolored meshes [4, 5, 17, 18, 39, 41, 42], resulting in suboptimal performance for colored meshes. Early approaches, such as Hausdorff distance [4] and mean squared error (MSE) [5], use Euclidean distance as a quality measure, while Lavoué et al. [17, 18] employ curvature statistics. Other methods use local curvature and roughness pooling [39, 42] or employ dihedral angles as surface quality indicators [41].", + "bbox": [ + 511, + 717, + 903, + 853 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, the rise of colored meshes drives the need for color integration in MQA methods. Tian and AlRegib [38] and Guo et al. [13] apply MSE and MS-SSIM [44] to", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21415", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2D texture maps, combining these with geometric descriptors to evaluate colored meshes. Nevertheless, 2D texture maps do not represent the post-rendering appearance of 3D meshes and carry little semantic information, leading to suboptimal performance. In contrast, a second group of methods defines color quality descriptors on per-vertex or per-face color values [9, 28, 48, 53]. Nehme et al. [28] propose CMDM, a model for vertex-color meshes that extracts multi-scale color features from vertex colors. Similarly, Zhang et al. [53] apply statistical measures on vertex colors for no-reference MQA. Fu et al. [9] also sample vertex colors and propose an efficient surface sampling approach to convert meshes into point clouds for quality evaluation. Finally, Yang et al. [48] develop GeodesicPSIM, using surface colors to create textured patches for feature extraction. However, these methods fail to capture the rendered appearance of 3D objects as perceived by ultimate human viewers.", + "bbox": [ + 89, + 90, + 480, + 347 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overall, while model-based MQA methods benefit from a profound understanding of the object's topology, their performance is constrained by the lack of access to the object's colored appearance. We address this issue by introducing a novel hybrid MQA method that complements the 3D awareness of model-based methods with textural representations derived from the colored appearance of 3D meshes.", + "bbox": [ + 89, + 348, + 480, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Projection-based Mesh Quality Assessment", + "text_level": 1, + "bbox": [ + 89, + 468, + 457, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In projection-based MQA, quality is assessed on rendered 2D projections of 3D meshes, allowing well-established IQA methods such as PSNR [43], SSIM [45], and VIF [34] to be adapted for MQA [7, 32, 47]. However, IQA methods perform poorly in MQA as they are tailored for natural scenes, while 3D meshes involve different distortions, and their 2D projections differ statistically from natural images.", + "bbox": [ + 89, + 491, + 482, + 598 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A few projection-based methods have been proposed for meshes without color [1, 2], and they expectedly perform poorly on colored meshes. To address this, Nehme et al. [27] propose Graphics-LPIPS, the first projection-based MQA method designed for colored meshes, which builds on LPIPS [52] and uses pre-trained AlexNet [15] to extract quality features from 2D projections. Similarly, Zhang et al. [55] use Swin Transformer [22] with an efficient mini-patch sampling process for no-reference MQA. Lee et al. [19] introduce 3D-PSSIM, which uses a framework similar to Graphics-LPIPS to extract textural information, while complementing it with geometry-aware information derived from the 2D projection space. Nevertheless, these projection-based methods lack 3D understanding of the object's topology. Our approach overcomes this through a well-defined and end-to-end trainable GCN that operates on vertices in the 3D space. Furthermore, these methods fail to account for interactions between the mesh's textural information and the underlying 3D geometry, which limits their performance in detecting geometry-involved distor", + "bbox": [ + 89, + 598, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tions. We address this by proposing a cross-attention framework that draws connections between the two domains.", + "bbox": [ + 511, + 90, + 903, + 119 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Method", + "text_level": 1, + "bbox": [ + 511, + 133, + 687, + 152 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Figure 3a shows the overall framework of HybridMQA, consisting of model and texture branches and a quality encoder. In the texture branch, colored projections of the input mesh (reference or distorted) are rendered from six perpendicular viewpoints. In the model branch, 2D normal, vertex, and texture maps are processed by a base encoder to extract initial 3D features, which initialize a graph based on mesh connectivity. A GCN then learns detailed 3D surface representations. At the same viewpoints as the texture branch, we render 3D feature projections from the graph, ensuring pixel-to-pixel alignment with the colored projections. Finally, the quality encoder processes both sets of projections, exploring geometry-texture interactions to produce a comprehensive quality representation of the mesh. Overall, the proposed HybridMQA can be expressed as follows:", + "bbox": [ + 511, + 160, + 906, + 388 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {f} _ {\\text {m e s h}} = \\text {H y b r i d M Q A} (\\mathcal {M} (\\boldsymbol {V}, \\boldsymbol {E}, \\boldsymbol {U}, \\boldsymbol {T} _ {t}); \\boldsymbol {\\theta}). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 398, + 903, + 416 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here $\\mathcal{M}$ is the input mesh, $\\pmb{\\theta}$ represents model parameters, and $f_{mesh}$ denotes the final quality representation. The mesh $\\mathcal{M}$ consists of vertices $\\pmb{V}$ , vertex connectivity $\\pmb{E}$ , UV coordinates $\\pmb{U}$ , and the 2D texture map $\\pmb{T}_t$ . Vertices and UV coordinates are defined as $\\pmb{V} = \\{\\pmb{v}_i\\}_{i=1}^N$ , $\\pmb{v}_i \\in \\mathbb{R}^3$ and $\\pmb{U} = \\{\\pmb{u}_i\\}_{i=1}^N$ , $\\pmb{u}_i \\in \\mathbb{R}^2$ , where $N$ is the number of vertices, and $\\pmb{v}_i$ and $\\pmb{u}_i$ are the positions of the $i$ -th vertex in 3D space and 2D map, respectively.", + "bbox": [ + 511, + 428, + 905, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Finally, as shown in Fig. 3b, HybridMQA generates quality representations for both reference and distorted meshes. For full-reference quality regression, the absolute differences between these representations are fed into a fully connected network (FCN) to obtain a quality score.", + "bbox": [ + 511, + 549, + 905, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Base Encoder", + "text_level": 1, + "bbox": [ + 511, + 633, + 655, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given an input mesh $\\mathcal{M}$ , we first project its 3D geometrical attributes (normals and vertex positions) into 2D maps aligned with the texture map using barycentric interpolation and UV mapping. The shared UV coordinates $\\pmb{U}$ ensure alignment between the resulting maps (normal and vertex) and the texture map $\\pmb{T}_t$ . These aligned maps are concatenated and processed by a convolutional neural network (CNN) base encoder to jointly capture textural and geometrical information, learning initial quality-aware representations of the mesh's 3D shape and surface (Fig. 3a). This approach of projecting geometrical data into 2D maps has been successfully applied in mesh super-resolution [46].", + "bbox": [ + 511, + 657, + 906, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Graph Convolutional Network", + "text_level": 1, + "bbox": [ + 511, + 847, + 784, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Due to vertex-neighborhood discontinuities in 2D maps, we use graph learning to deepen our model's understanding", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21416", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/779296ae23dab8a424c7f5466b19a516c8295accee3f6b24ed13f197dc53ff24.jpg", + "image_caption": [ + "(a) HybridMQA's framework." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 697, + 279 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/10c17530b52a03b726ee239cb2cf3168dd1aec8eaf71f3b1aa73683167153677.jpg", + "image_caption": [ + "(b) FR quality regression.", + "Figure 3. (a) Overview of HybridMQA. In the model branch, a base encoder extracts 3D features from the mesh's 2D maps, initializing a mesh graph. A GCN extracts 3D surface representations, which are rendered as 2D projections aligned with the colored projections from the texture branch. A quality encoder then captures geometry-texture interactions between the two branches, producing the final mesh quality representation. (b) Full-reference (FR) quality regression, where the absolute difference of HybridMQA's mesh quality representations for reference and distorted meshes is mapped to a quality score via a fully connected network (FCN)." + ], + "image_footnote": [], + "bbox": [ + 709, + 93, + 906, + 279 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "of the mesh's 3D structure via message-passing between neighboring vertices in 3D, enabling them to share useful quality-related information and highlight abnormalities.", + "bbox": [ + 88, + 402, + 480, + 448 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Graph initialization. Given the base encoder's output feature map with $C_1$ channels, we sample a $C_1$ -dimensional feature vector for each mesh vertex using UV mapping:", + "bbox": [ + 88, + 449, + 482, + 496 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {f} _ {i} = \\boldsymbol {T} _ {B E} \\left[ u _ {i} ^ {h}, u _ {i} ^ {w} \\right], i \\in \\{1, \\dots , N \\}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 510, + 482, + 530 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $\\pmb{f}_i \\in \\mathbb{R}^{C_1}$ is the sampled feature for $i$ -th vertex, $u_i^h$ and $u_i^w$ are its UV coordinates, and $T_{BE}$ denotes base encoder's output feature maps. For vertices with multiple UV mappings, we average the sampled features. This process produces a feature graph with vertices initialized by sampled features and edges defined by mesh connectivity $\\pmb{E}$ .", + "bbox": [ + 88, + 542, + 482, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Graph update. We apply a GCN to integrate features from neighboring vertices to learn surface properties. As proposed by Morris et al. [26], we update vertex features by:", + "bbox": [ + 88, + 636, + 482, + 681 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {f} _ {i} ^ {\\prime} = \\boldsymbol {\\theta} _ {1} \\boldsymbol {f} _ {i} + \\boldsymbol {\\theta} _ {2} \\sum_ {j \\in \\Gamma (i)} e _ {j, i} \\boldsymbol {f} _ {j}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 700, + 482, + 736 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $f_{i}^{\\prime} \\in \\mathbb{R}^{C_{2}}$ is the updated feature vector for the $i$ -th vertex, $\\Gamma(i)$ is its 1-ring neighborhood, $\\theta_{1}, \\theta_{2} \\in \\mathbb{R}^{C_{2} \\times C_{1}}$ are learnable weights, and $e_{j,i}$ is the edge weight between vertices $i$ and $j$ , defined as the inverse of their Euclidean distance. Through training, the GCN refines vertex feature embeddings to learn quality-aware 3D surface representations, which are later complemented by textural embeddings to deliver a hybrid and comprehensive quality assessment. This approach of graph learning has been effectively applied to mesh texture downsampling [29].", + "bbox": [ + 88, + 748, + 482, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Feature & RGB Rendering", + "text_level": 1, + "bbox": [ + 511, + 402, + 756, + 419 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We render multiple viewpoints of the 3D surface representations into 2D projections aligned with the mesh's colored projections. This novel feature graph rendering enables us to capture geometry-texture interactions between the two branches for a comprehensive mesh quality representation.", + "bbox": [ + 511, + 424, + 906, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the texture branch (Fig. 3a), colored projections are rendered after normalizing mesh vertex positions to fit within a unit cube. This standardization allows fixed camera positions for any mesh. Six virtual cameras are placed on the cube's surfaces, facing the object to cover all angles. With PyTorch3D's [30] soft Phong Shader, we render six perpendicular colored projections, employing directional or ambient light to match the conditions of subjective tests.", + "bbox": [ + 511, + 501, + 905, + 622 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The model branch employs the same camera setup to ensure alignment with colored projections. Our novel feature graph rendering customizes PyTorch3D's differentiable renderer to render six perpendicular $3D$ feature projections from the graph, enabling gradient backpropagation to the GCN and base encoder. To focus on raw vertex features interpolated on the mesh surface, we remove shadows and specular effects, setting diffuse and specular reflectivity to zero and using ambient light. Thus we define the mesh as a vertex-color mesh with $C$ -dimensional vertex features (instead of RGB colors) to render the 3D feature projections with hard Phong shader [30].", + "bbox": [ + 511, + 623, + 906, + 805 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Quality Encoder", + "text_level": 1, + "bbox": [ + 511, + 816, + 678, + 832 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We design a quality encoder that processes the aligned projections from model and texture branches to output an expressive quality representation for the mesh. It captures the geometry-texture interactions between the two sets of pro", + "bbox": [ + 511, + 839, + 905, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21417", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fcfdc02f92da17a7a799b09757e5fbbd95a9a22a714c8adf55ed525bd18acc88.jpg", + "image_caption": [ + "(a) Quality encoder." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 537, + 327 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a073be4013f89a4fbe66daa6928f82fe1e3e8f3a63bf7b525c4f011c13425f86.jpg", + "image_caption": [ + "(b) Cross-attention module.", + "Figure 4. (a) The quality encoder. The 3D feature and color projections are divided into valid aligned patches and fed into respective encoders to obtain multiscale 3D surface and color representations. The cross-attention modules capture interactions between these representations, which are then concatenated with 3D feature embeddings directly extracted from the patches to form the final mesh quality representation. (b) The cross-attention module consists of two transformer blocks, where we switch the roles of the two inputs." + ], + "image_footnote": [], + "bbox": [ + 581, + 93, + 893, + 327 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "jections through multi-scale cross-attentions. As discussed in Sec. 1, understanding these interactions is key to a robust and effective MQA. As detailed in Fig. 4a, given model branch's 3D feature projections $\\mathbf{I}_m$ and texture branch's colored projections $\\mathbf{I}_t$ , the quality encoder $\\psi$ outputs", + "bbox": [ + 88, + 440, + 483, + 516 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {f} _ {\\text {m e s h}} = \\psi \\left(\\boldsymbol {I} _ {m}, \\boldsymbol {I} _ {t}; \\boldsymbol {\\theta} _ {\\psi}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 202, + 523, + 480, + 540 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $f_{mesh}$ denotes the final quality representation, and $\\theta_{\\psi}$ denotes the learnable parameters.", + "bbox": [ + 89, + 546, + 482, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Patch Selection. Given two sets of projections, we extract non-overlapping patches and discard those with less than $10\\%$ non-background pixels. This yields $2P$ patches:", + "bbox": [ + 89, + 575, + 483, + 622 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\boldsymbol {I}} _ {m} = \\left\\{\\boldsymbol {I} _ {m, p} \\right\\} _ {p = 1} ^ {P}, \\tilde {\\boldsymbol {I}} _ {t} = \\left\\{\\boldsymbol {I} _ {t, p} \\right\\} _ {p = 1} ^ {P}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 626, + 480, + 646 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $\\tilde{I}_m$ and $\\tilde{I}_t$ denote aligned patches extracted from 3D feature and colored projections, respectively.", + "bbox": [ + 89, + 652, + 482, + 683 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3D Feature & Image Encoding. Next, we feed pairs of aligned patches $I_{m,p}$ and $I_{t,p}$ to a 3D feature encoder $\\phi$ and an image encoder $\\varphi$ , respectively, to extract multiscale quality-aware representations $F_{m,p}$ and $F_{t,p}$ :", + "bbox": [ + 89, + 684, + 483, + 746 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {F} _ {m, p} = \\left\\{\\boldsymbol {f} _ {m, p} ^ {s} | s = 1, \\dots , 5 \\right\\} = \\phi (\\boldsymbol {I} _ {m, p}; \\boldsymbol {\\theta} _ {\\phi}), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 751, + 482, + 776 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {F} _ {t, p} = \\left\\{\\boldsymbol {f} _ {t, p} ^ {s} | s = 1, \\dots , 5 \\right\\} = \\varphi (\\boldsymbol {I} _ {t, p}; \\boldsymbol {\\theta} _ {\\varphi}).\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 772, + 424, + 789 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $\\theta_{\\phi}$ and $\\theta_{\\varphi}$ are learnable parameters of the two encoders, and $s$ denotes scale index. $1\\times 1$ convolutions are used to adjust the channel dimension of representations.", + "bbox": [ + 89, + 794, + 482, + 840 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Cross-attention Modules. We apply cross-attention to capture the interactions between aligned 3D surface representations $\\pmb{f}_{m,p}^{s}$ and color representations $\\pmb{f}_{t,p}^{s}$ at each scale. Figure 4b depicts our proposed module with two transformer", + "bbox": [ + 89, + 840, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "blocks [40], where we alternate the query and key-value roles of $f_{m,p}^{s}$ and $f_{t,p}^{s}$ to explore their interactions and mutual influence. This simulates how one domain's representation affects the impact of the other domain's representation on perceptual quality, as discussed in Sec. 1. We concatenate outputs to form single-scale quality representation $f_{p}^{s}$ .", + "bbox": [ + 511, + 440, + 906, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Mesh Quality Representation. To further exploit the model branch's understanding of mesh's 3D structure and surface properties, we directly extract representation $\\hat{f}_p$ from model branch patches via a convolutional block. $\\hat{f}_p$ is concatenated with the five single-scale representations from cross-attention, and averaged over all patches to obtain the final mesh quality representation $f_{mesh}$ :", + "bbox": [ + 511, + 532, + 905, + 640 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {f} _ {\\text {m e s h}} = \\frac {1}{P} \\sum_ {p = 1} ^ {P} \\left(\\boldsymbol {f} _ {p} ^ {1} \\oplus \\dots \\oplus \\boldsymbol {f} _ {p} ^ {5} \\oplus \\hat {\\boldsymbol {f}} _ {p}\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 647, + 906, + 690 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Quality Regression & Optimization", + "text_level": 1, + "bbox": [ + 511, + 698, + 821, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Fig. 3b, we feed the reference and distorted meshes into HybridMQA separately and map the absolute difference of their quality representations $(|f_{mesh}^{ref} - f_{mesh}^{dis}|)$ to a quality score via an FCN for full-reference MQA. The model is optimized using a loss function with two terms: mean absolute error (MAE) and rank loss. While MAE ensures accurate quality predictions, rank loss helps differentiate closely rated samples within a mini-batch [36].", + "bbox": [ + 511, + 719, + 905, + 840 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "MAE Loss. The MAE loss is defined as", + "bbox": [ + 511, + 840, + 782, + 854 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {m a e} = \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\left| q _ {i} - q _ {i} ^ {\\prime} \\right|, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 624, + 862, + 906, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21418", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $q_{i}$ and $q_{i}^{\\prime}$ denote the predicted and ground truth quality scores of the $i$ -th sample in the batch, with $B$ being the batch size. The ground truth is the mean opinion score (MOS) obtained in subjective tests and normalized to [0, 1].", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Rank Loss. Since ranking is not differentiable, as proposed by Sun et al. [36], we approximate the rank value as", + "bbox": [ + 89, + 152, + 482, + 181 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {r a n k} ^ {i, j} = \\max \\left(0, \\left| q _ {i} ^ {\\prime} - q _ {j} ^ {\\prime} \\right| - e \\left(q _ {i} ^ {\\prime}, q _ {j} ^ {\\prime}\\right) \\cdot \\left(q _ {i} - q _ {j}\\right)\\right), \\\\ e \\left(q _ {i} ^ {\\prime}, q _ {j} ^ {\\prime}\\right) = \\left\\{ \\begin{array}{l l} 1 & q _ {i} ^ {\\prime} \\geq q _ {j} ^ {\\prime} \\\\ - 1 & q _ {i} ^ {\\prime} < q _ {j} ^ {\\prime} \\end{array} , \\right. \\tag {9} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 191, + 480, + 248 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $i$ and $j$ denote two samples in a batch, and $e(q_i', q_j')$ is a sign function. The final rank loss is computed as", + "bbox": [ + 89, + 258, + 480, + 289 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {r a n k} = \\frac {1}{B ^ {2} - B} \\sum_ {i = 1} ^ {B} \\sum_ {\\substack {j = 1 \\\\ j \\neq i}} ^ {B} L _ {r a n k} ^ {i, j}. \\tag{10}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 300, + 480, + 351 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Final Loss. The final loss is a weighted sum of the two:", + "bbox": [ + 89, + 361, + 460, + 376 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL = L _ {m a e} + \\lambda L _ {r a n k}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 388, + 480, + 405 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\lambda$ balances the effect of the two loss terms.", + "bbox": [ + 89, + 416, + 416, + 430 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 445, + 223, + 462 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets & Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 469, + 395, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We validate our method on four publicly available color MQA datasets: Nehme et al. [27], SJTU-TMQA [7], TSMD [47], and CMDM [28] datasets. Nehme et al., the largest available, includes textured meshes with mixed geometric and texture distortions, as do SJTU-TMQA and TSMD. In contrast, CMDM consists of vertex-color meshes with either geometric or color distortions. All datasets use MOS as ground truth.", + "bbox": [ + 89, + 491, + 483, + 611 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use ResNet50 [14] pre-trained on ImageNet [8] as the image encoder $\\varphi$ and a randomly initialized CNN as the 3D feature encoder $\\phi$ . 3D feature and colored projections are rendered at $128 \\times 128$ and $512 \\times 512$ resolutions, with patch sizes of $16 \\times 16$ and $64 \\times 64$ , respectively, to ensure alignment. For data augmentation, we randomly perturb camera angles and flip the patches in training to improve robustness and generalization. More details of the datasets and implementation are provided in the supplementary material.", + "bbox": [ + 89, + 613, + 483, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 757, + 279, + 773 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use 5-fold cross-validation without overlap between train and test source content and report the median performance across five experiments. This strategy is applied to all learning-based methods for fair comparison.", + "bbox": [ + 89, + 780, + 480, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare HybridMQA with 11 model-based and projection-based full-reference MQA methods. Model-based methods include Hausdorff Distance (HD) [4], MSDM2 [18], FMPD [42], GeodesicPSIM [48], and Fu", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "et al. [9]. Projection-based methods include PSNR [43], SSIM [45], VIF [34], LPIPS [52], Graphics-LPIPS [27], and 3D-PSSIM [19]. For a fair comparison, all projection-based methods are evaluated under the same rendering settings as HybridMQA, with no prior assumptions about object orientation. We adopt the Spearman rank-order correlation coefficient (SRCC) and the Pearson linear correlation coefficient (PLCC) to compare all methods. Higher values indicate better performance and a stronger correlation between MOS and predicted quality scores [12]. Further information is provided in the supplementary material.", + "bbox": [ + 511, + 90, + 903, + 257 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3.Quantitative Results", + "text_level": 1, + "bbox": [ + 511, + 270, + 705, + 286 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Overall Performance. Table 1 summarizes the performance comparison, from which we make the following observations: (1) HybridMQA outperforms all model-based and projection-based comparison methods across all datasets, including both textured and vertex-color meshes, demonstrating its effectiveness in the quality assessment of colored meshes. Notably, HybridMQA achieves $6.5\\%$ and $7.7\\%$ performance gain in PLCC over the second-best method, 3D-PSSIM, on Nehme et al. and SJTU-TMQA, respectively; (2) In general, projection-based methods outperform model-based methods which lack access to the object's actual appearance. HybridMQA consolidates the advantages of both types by complementing the textural information extracted from colored projections with 3D representations learned in 3D; (3) Unlike CMDM with single-type distortions, Nehme et al., SJTU-TMQA, and TSMD involve mixed-type and hence more complex distortions. Consequently, while most methods perform well on CMDM, their performance does not extend to the other three datasets. In contrast, HybridMQA consistently performs well on all datasets, showing robustness in assessing complex distortions. This is achieved by HybridMQA's understanding of geometry-texture interactions, which are perturbed uniquely by different distortions.", + "bbox": [ + 511, + 294, + 906, + 655 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Performance by Distortion Type. In Fig. 5, we compare HybridMQA with Graphics-LPIPS based on distortion types in SJTU-TMQA. In addition to the overall superiority of HybridMQA, we observe that while the two methods achieve comparable performance on texture-only distortions (e.g. JPEG [7]), HybridMQA hugely outperforms Graphics-LPIPS on distortions that only affect geometry (e.g. gn [7] and simpNoTex [7, 10]) or both geometry and texture (e.g. qpqtJPEG [7, 11]). This further demonstrates HybridMQA's proficiency in understanding meshes' 3D geometry and its interactions with textural appearance.", + "bbox": [ + 511, + 657, + 905, + 823 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Generalizability. We train models on Nehme et al. and TSMD and test them on SJTU-TMQA. Table 2 shows that HybridMQA significantly outperforms other learning-based methods, showing strong generalizability. Notably, when trained on TSMD and tested on SJTU-TMQA, HybridMQA", + "bbox": [ + 511, + 824, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21419", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c760ae9071fe16e1e65e849605e6d225dbc5a07837708570d06e6c1d4976cd35.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TypeMethodNehmé et al. [27]SJTU-TMQA [7]TSMD [47]CMDM [28]
SRCCPLCCSRCCPLCCSRCCPLCCSRCCPLCC
Model-basedHD [4]0.1070.1750.0600.1400.4460.4620.1890.210
MSDM2 [18]0.3350.3440.0500.1200.0450.2550.4150.517
FMPD [42]0.3910.4040.1560.4580.0770.2180.6150.623
GeodesicPSIM [48]----0.8200.820--
Fu et al. [9]0.6880.696------
Projection-basedPSNR [43]0.3530.3750.2990.2870.7140.7110.8300.839
SSIM [45]0.2100.2260.3940.2890.6730.6740.8520.861
VIF [34]0.5380.5570.4500.4220.8510.8460.8270.837
LPIPS [52]0.6720.6760.7180.7170.7100.7120.8650.918
Graphics-LPIPS [27]0.7220.7460.7900.7620.8340.8120.8590.925
3D-PSSIM [19]0.8820.8420.8420.832--0.8550.854
HybridHybridMQA0.8920.8970.8870.8960.9120.9190.8970.927
", + "bbox": [ + 130, + 88, + 870, + 311 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3a7e63e6e22f7f6203a8f8aac62b63daac8a5a6f7f4c2b0d57a747ceefee61d8.jpg", + "table_caption": [ + "Table 1. SRCC and PLCC scores of MQA methods on four color MQA benchmark datasets. The scores of GeodesicPSIM [48], Fu et al. [9], and 3D-PSSIM [19] are reported directly from their publications as their implementations are not publicly available. Bold and underlined values denote the best and second-best results per column, respectively." + ], + "table_footnote": [], + "table_body": "
Trained onNehmé et al.TSMD
SRCCPLCCSRCCPLCC
LPIPS0.5920.5840.7120.695
Graphics-LPIPS0.7310.7340.7840.773
HybridMQA0.8000.7830.8460.811
", + "bbox": [ + 114, + 387, + 460, + 481 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Generalization evaluation, where models are trained on Nehme et al. and TSMD and tested on SJTU-TMQA.", + "bbox": [ + 89, + 489, + 482, + 518 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f36b5af9e603896d80a9603ea60f5870bf8aaea382cecd7f4dddc146bfc82b11.jpg", + "image_caption": [ + "Figure 5. SRCC/PLCC performance of HybridMQA and Graphics-LPIPS on various distortion types of SJTU-TMQA dataset. Green distortions affect the geometry or both geometry and texture, while others only impact the texture." + ], + "image_footnote": [], + "bbox": [ + 99, + 532, + 475, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "achieves comparable performance to 3D-PSSIM [19], despite 3D-PSSIM being trained directly on SJTU-TMQA.", + "bbox": [ + 89, + 773, + 482, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Qualitative Results", + "text_level": 1, + "bbox": [ + 89, + 816, + 274, + 832 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We apply GradCAM [33] on 3D feature projections to verify that the model branch effectively captures geometry-aware quality representations. As shown in Fig. 6, the highlighted regions with noticeable artifacts align well with hu", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3f4e7f486e3eb16442125b19ad9b3b3ed0bfaa046c773bf97f24baa69bd386c7.jpg", + "image_caption": [ + "Figure 6. GradCAM results on meshes in the model branch. Highlighted regions exhibit more noticeable artifacts, aligning with human perception and showing the model's effectiveness in capturing geometry-aware quality representations." + ], + "image_footnote": [], + "bbox": [ + 519, + 388, + 906, + 577 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "man perception, validating the model branch's ability to extract geometry-aware representations. Additionally, by mapping these results onto the mesh topology via differentiable rendering, our method opens up opportunities for optimizing mesh compression or restoration algorithms. More examples are available in the supplementary material.", + "bbox": [ + 511, + 672, + 906, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 7 shows the effectiveness of our hybrid model in exploiting interactions between representations learned in texture and model branches. GradCAM is applied before and after cross-attention, showing that the two branches focus on different regions, with the model branch effectively highlighting geometric artifacts. Cross-attention then successfully identifies and attends to perceptually important regions by exploring geometry-texture interactions. Note that the actual cross-attention inputs are patches described", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21420", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/44b3247055d76cfe2d14219205c8bfcdf14c4320f5d6d360106ce74a28f08d87.jpg", + "image_caption": [ + "Figure 7. GradCAM before and after cross-attention. The two branches focus on different regions, and cross-attention effectively integrates them by attending to perceptually important regions." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 486, + 439 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "in Sec. 3.3 and 3.4, and the heatmaps are the union of the GradCAM results from the reference and distorted patches. More examples are available in the supplementary material.", + "bbox": [ + 89, + 517, + 482, + 563 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also perform gMAD competition [23] to qualitatively compare HybridMQA with Graphics-LPIPS. The results are available in the supplementary material.", + "bbox": [ + 89, + 563, + 483, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 619, + 256, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct ablation studies on Nehme et al. [27] to evaluate key components and design choices in our model. Further results are available in the supplementary material.", + "bbox": [ + 89, + 643, + 482, + 688 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3D Surface and Textural Representations. To show the contributions of 3D surface and textural representations, we test four configurations: (1) 3D-only, where only 3D representations $\\hat{f}$ and $f_{m}$ are used for mesh quality representation; (2) RGB-only, where only textural representation $f_{t}$ from colored renderings are used; (3) all but excluding $f_{m}$ ; and (4) excluding $\\hat{f}$ . Table 3 shows that all representations contribute to the final performance, with textural representation $f_{t}$ the most significant. Moreover, excluding 3D representations (RGB-only) leads to a significant performance drop, showing the importance of 3D representations and the need to explore their interaction with textural information for accurate MQA. The results also highlight the performance gain of our hybrid approach, as it outperforms both", + "bbox": [ + 89, + 688, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/aa603b712210cc1c06ffed62344f211b0bd882e576b19163908ee2d10c4ce4ee.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
NotesffmftSRCCPLCC
3D-only✓✓-0.5860.595
RGB-only--✓0.8200.846
-✓-✓0.8420.849
--✓✓0.8700.874
Proposed✓✓✓0.8920.897
", + "bbox": [ + 558, + 88, + 864, + 189 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9249fc6cab677618333f5bcfbb7fa8c15b80ce490ae326a5eed170424cd67687.jpg", + "table_caption": [ + "Table 3. Ablation on 3D Surface & Text. Repr. on Nehme et al." + ], + "table_footnote": [], + "table_body": "
ConfigurationsSRCCPLCC
w/o Texture map0.8660.872
w/o Normal map0.8630.866
w/o Vertex map0.8720.873
w/o Base Encoder0.8560.851
w/o GCN0.8660.865
HybridMQA (proposed)0.8920.897
", + "bbox": [ + 570, + 214, + 852, + 327 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Ablation on model branch components on Nehme et al.", + "bbox": [ + 517, + 338, + 898, + 352 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3D-only and RGB-only counterparts.", + "bbox": [ + 511, + 378, + 761, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Model Branch Components. To validate the contributions of model branch components, we test the following configurations: all components but excluding (1) the texture map; (2) the normal map; (3) the vertex map; (4) the Base Encoder; and (5) the GCN. Table 4 shows that all components contribute to learning effective 3D surface representations. Furthermore, the proposed processing units (Base Encoder and GCN) effectively leverage mesh data modalities (2D maps and 3D vertex space) to enhance performance.", + "bbox": [ + 511, + 393, + 906, + 530 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion & Discussion", + "text_level": 1, + "bbox": [ + 511, + 542, + 746, + 558 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present a novel hybrid full-reference MQA method that integrates model- and projection-based approaches for enhanced quality assessment. Our model explores interactions between mesh texture and 3D geometry via cross-attention, enabled by a novel feature rendering process that aligns 3D representations with colored projections. Extensive experiments show the effectiveness and superiority of our method, highlighting the importance of 3D understanding and leveraging geometry-texture interactions for reliable MQA.", + "bbox": [ + 511, + 568, + 906, + 703 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A few limitations present room for improvement. Our method relies on perpendicular viewpoints without considering their varying contributions to perceptual quality. Integrating a semantic-aware module could address this by weighting viewpoints based on importance. Also, memory consumption scales with mesh size as the GCN processes the entire mesh graph, limiting efficiency in real-world applications. Sampling techniques [16, 50, 51] could help reduce memory footprint. Future directions include (1) adapting our work for no-reference MQA; (2) generalizing it to point cloud quality assessment [54, 56] by defining graph edges; and (3) exploring applications in perceptually optimized mesh compression, enhancement, and generation.", + "bbox": [ + 511, + 704, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21421", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Ilyass Aboulaziz, Aladine Chetouani, Mohammed El Hassouni, and Hocine Cherifi. A blind mesh visual quality assessment method based on convolutional neural network. *Electronic Imaging*, 30:1-5, 2018. 3", + "[2] Ilyass Abouelaziz, Aladine Chetouani, Mohammed El Hassouni, Longin Jan Latecki, and Hocine Cherifi. No-reference mesh visual quality assessment via ensemble of convolutional neural networks and compact multi-linear pooling. Pattern Recognition, 100:107174, 2020. 3", + "[3] Martin Alain, Emin Zerman, Cagri Ozcinar, and Giuseppe Valenzise. Chapter 1 - introduction to immersive video technologies. In Immersive Video Technologies, pages 3-24. Academic Press, 2023. 1", + "[4] Nicolas Aspert, Diego Santa-Cruz, and Touradj Ebrahimi. Mesh: measuring errors between surfaces using the hausdorff distance. In Proceedings. IEEE International Conference on Multimedia and Expo, pages 705-708 vol.1, 2002. 2, 6, 7", + "[5] Paolo Cignoni, Claudio Rocchini, and Roberto Scopigno. Metro: Measuring error on simplified surfaces. Computer Graphics Forum, 17(2):167-174, 1998. 2", + "[6] Sahin Coskun, Gokce Nur Yilmaz, Federica Battisti, Musaed Alhussein, and Saiful Islam. Measuring 3d video quality of experience (qoe) using a hybrid metric based on spatial resolution and depth cues. Journal of Imaging, 9(12):281, 2023. 2", + "[7] Bingyang Cui, Qi Yang, Kaifa Yang, Yiling Xu, Xiaozhong Xu, and Shan Liu. Sjtu-tmqa: A quality assessment database for static mesh with texture map. In ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 7875-7879, 2024. 3, 6, 7", + "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 6", + "[9] Chunyang Fu, Xiang Zhang, Thuong Nguyen-Canh, Xiaozhong Xu, Ge Li, and Shan Liu. Surface-sampling based objective quality assessment metrics for meshes. In ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5, 2023. 3, 6, 7", + "[10] Michael Garland and Paul S. Heckbert. Surface simplification using quadric error metrics. In Proceedings of the 24th Annual Conference on Computer Graphics and Interactive Techniques, page 209-216, USA, 1997. ACM Press/Addison-Wesley Publishing Co. 6", + "[11] Google. Draco 3d data compression. https://google.github.io/draco/. Accessed: 2024-09-04. 1,6", + "[12] Video Quality Experts Group et al. Final report from the video quality experts group on the validation of objective models of video quality assessment, phase II. VQEG, 2003. 6", + "[13] Jinjiang Guo, Vincent Vidal, Irene Cheng, Anup Basu, Atilla Baskurt, and Guillaume Lavoue. Subjective and objective" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "visual quality assessment of textured 3d meshes. ACM Trans. Appl. Percept., 14(2), 2016. 2", + "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 6", + "[15] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2012. 3", + "[16] Itai Lang, Asaf Manor, and Shai Avidan. Samplenet: Differentiable point cloud sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7578-7588, 2020. 8", + "[17] Guillaume Lavoué, Elisa Drelie Gelasca, Florent Dupont, Atilla Baskurt, and Touradj Ebrahimi. Perceptually driven 3d distance metrics with application to watermarking. Applications of Digital Image Processing XXIX, 6312:63120L, 2006. 2", + "[18] Guillaume Lavoué. A multiscale metric for 3d mesh visual quality assessment. Computer Graphics Forum, 30(5):1427-1437, 2011. 2, 6, 7", + "[19] Seongmin Lee, Jiwoo Kang, Sanghoon Lee, Weisi Lin, and Alan Conrad Bovik. 3d-pssim: Projective structural similarity for 3d mesh quality assessment robust to topological irregularities. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-16, 2024. 2, 3, 6, 7", + "[20] Jie Li and Pablo Cesar. Chapter 22 - social virtual reality (vr) applications and user experiences. In Immersive Video Technologies, pages 609-648. Academic Press, 2023. 1", + "[21] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Mukund Varma T, Zexiang Xu, and Hao Su. One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. Advances in Neural Information Processing Systems, 36, 2024. 1", + "[22] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10012-10022, 2021. 3", + "[23] Kede Ma, Zhengfang Duanmu, Zhou Wang, Qingbo Wu, Wentao Liu, Hongwei Yong, Hongliang Li, and Lei Zhang. Group maximum differentiation competition: Model comparison with few samples. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(4):851-864, 2020. 8", + "[24] Adrien Maglo, Guillaume Lavoué, Florent Dupont, and Céline Hudelot. 3d mesh compression: Survey, comparisons, and emerging trends. ACM Comput. Surv., 47(3), 2015. 1", + "[25] Adrien Maglo, Guillaume Lavoué, Florent Dupont, and Céline Hudelot. 3d mesh compression: Survey, comparisons, and emerging trends. ACM Comput. Surv., 47(3), 2015. 1", + "[26] Christopher Morris, Martin Ritzert, Matthias Fey, William L. Hamilton, Jan Eric Lenssen, Gaurav Rattan, and Martin Grohe. Weisfeiler and leman go neural: Higher-order graph neural networks. Proceedings of the AAAI Conference on Artificial Intelligence, 33(01):4602-4609, 2019. 4", + "[27] Yana Nehme, Johanna Delanoy, Florent Dupont, JeanPhilippe Farrugia, Patrick Le Callet, and Guillaume Lavoué." + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21422", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Textured mesh quality assessment: Large-scale dataset and deep learning-based quality metric. ACM Trans. Graph., 42 (3), 2023. 2, 3, 6, 7, 8", + "[28] Yana Nehme, Florent Dupont, Jean-Philippe Farrugia, Patrick Le Callet, and Guillaume Lavoué. Visual quality of 3d meshes with diffuse colors in virtual reality: Subjective and objective evaluation. IEEE Transactions on Visualization and Computer Graphics, 27(3):2202-2219, 2021. 2, 3, 6, 7", + "[29] Sai Karthikey Pentapati, Anshul Rai, Arkady Ten, Chaitanya Atluru, and Alan Bovik. Geoscaler: Geometry and rendering-aware downsampling of 3d mesh textures. arXiv preprint arXiv:2311.16581, 2023. 4", + "[30] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3d deep learning with pytorch3d. arXiv:2007.08501, 2020. 4", + "[31] Stephan Reichelt, Ralf Häussler, Gerald Füfterer, and Norbert Leister. Depth cues in human visual perception and their realization in 3d displays. In Three-Dimensional Imaging, Visualization, and Display 2010 and Display Technologies and Applications for Defense, Security, and Avionics IV, pages 92–103. SpIE, 2010. 2", + "[32] Armin Shafiee Sarvestani, Wei Zhou, and Zhou Wang. Perceptual crack detection for rendered 3d textured meshes. In 2024 16th International Conference on Quality of Multimedia Experience (QoMEX), pages 1-7, 2024. 3", + "[33] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 7", + "[34] Hamid R. Sheikh and Alan C. Bovik. Image information and visual quality. IEEE Transactions on Image Processing, 15 (2):430-444, 2006. 3, 6, 7", + "[35] Yawar Siddiqui, Tom Monnier, Filippos Kokkinos, Mahendra Kariya, Yanir Kleiman, Emilien Garreau, Oran Gafni, Natalia Neverova, Andrea Vedaldi, Roman Shapovalov, et al. Meta 3d assetgen: Text-to-mesh generation with high-quality geometry, texture, and pbr materials. arXiv preprint arXiv:2407.02445, 2024. 1", + "[36] Wei Sun, Xiongkuo Min, Wei Lu, and Guangtao Zhai. A deep learning based no-reference quality assessment model forUGC videos. In Proceedings of the 30th ACM International Conference on Multimedia, page 856-865, New York, NY, USA, 2022. Association for Computing Machinery. 5, 6", + "[37] Yuk Ming Tang and Ho Lun Ho. 3d modeling and computer graphics in virtual reality. In mixed reality and three-dimensional computer graphics. IntechOpen, 2020. 1", + "[38] Dihong Tian and Ghassan AlRegib. Batex3: Bit allocation for progressive transmission of textured 3-d models. IEEE Transactions on Circuits and Systems for Video Technology, 18(1):23-35, 2008. 2", + "[39] Fakhri Torkhani, Kai Wang, and Jean-Marc Chassery. A curvature tensor distance for mesh visual quality assessment. In International Conference on Computer Vision and Graphics, pages 253–263. Springer, 2012. 2" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 5", + "[41] Libor Váša and Jan Rus. Dihedral angle mesh error: a fast perception correlated distortion measure for fixed connectivity triangle meshes. Computer Graphics Forum, 31(5):1715-1724, 2012. 2", + "[42] Kai Wang, Fakhri Torkhani, and Annick Montanvert. A fast roughness-based approach to the assessment of 3d mesh visual quality. Computers & Graphics, 36(7):808-818, 2012. Augmented Reality Computer Graphics in China. 2, 6, 7", + "[43] Zhou Wang and Alan C. Bovik. Mean squared error: Love it or leave it? a new look at signal fidelity measures. IEEE Signal Processing Magazine, 26(1):98-117, 2009. 3, 6, 7", + "[44] Zhou Wang, Eero P Simoncelli, and Alan C Bovik. Multiscale structural similarity for image quality assessment. In The Thirty-Seventh Asilomar Conference on Signals, Systems & Computers, 2003, pages 1398-1402 Vol.2, 2003. 2", + "[45] Zhou Wang, Alan C. Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 3, 6, 7", + "[46] Wuyuan Xie, Tengcong Huang, and Miaohui Wang. Mnsnet: multimodal transformer network for 3d surface superresolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12703-12712, 2022. 3", + "[47] Qi Yang, Joel Jung, Haiqiang Wang, Xiaozhong Xu, and Shan Liu. Tsmd: A database for static color mesh quality assessment study. In 2023 IEEE International Conference on Visual Communications and Image Processing (VCIP), pages 1-5, 2023. 3, 6, 7", + "[48] Qi Yang, Joel Jung, Xiaozhong Xu, and Shan Liu. Geodesicpsim: Predicting the quality of static mesh with texture map via geodesic patch similarity. arXiv preprint arXiv:2308.04928, 2023. 2, 3, 6, 7", + "[49] Gareth W. Young, Néill O'Dwyer, and Aljosa Smolic. Chapter 21 - volumetric video as a novel medium for creative storytelling. In Immersive Video Technologies, pages 591-607. Academic Press, 2023. 1", + "[50] Hanqing Zeng, Hongkuan Zhou, Ajitesh Srivastava, Rajgopal Kannan, and Viktor Prasanna. GraphSAINT: Graph sampling based inductive learning method. In International Conference on Learning Representations, 2020. 8", + "[51] Hanqing Zeng, Muhan Zhang, Yinglong Xia, Ajitesh Srivastava, Andrey Malevich, Rajgopal Kannan, Viktor Prasanna, Long Jin, and Ren Chen. Decoupling the depth and scope of graph neural networks. Advances in Neural Information Processing Systems, 34:19665-19679, 2021. 8", + "[52] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3, 6, 7", + "[53] Zicheng Zhang, Wei Sun, Xiongkuo Min, Tao Wang, Wei Lu, and Guangtao Zhai. No-reference quality assessment for" + ], + "bbox": [ + 516, + 90, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21423", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "3d colored point cloud and mesh models. IEEE Transactions on Circuits and Systems for Video Technology, 32(11):7618-7631, 2022. 2, 3", + "[54] Zicheng Zhang, Wei Sun, Xiongkuo Min, Qiyuan Wang, Jun He, Quan Zhou, and Guangtao Zhai. Mm-pcqa: Multi-modal learning for no-reference point cloud quality assessment. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, IJCAI-23, pages 1759–1767. International Joint Conferences on Artificial Intelligence Organization, 2023. Main Track. 8", + "[55] Zicheng Zhang, Wei Sun, Haoning Wu, Yingjie Zhou, Chunyi Li, Zijian Chen, Xiongkuo Min, Guangtao Zhai, and Weisi Lin. Gms-3dqa: Projection-based grid mini-patch sampling for 3d model quality assessment. ACM Trans. Multimedia Comput. Commun. Appl., 20(6), 2024. 2, 3", + "[56] Zicheng Zhang, Haoning Wu, Yingjie Zhou, Chunyi Li, Wei Sun, Chaofeng Chen, Xiongkuo Min, Xiaohong Liu, Weisi Lin, and Guangtao Zhai. Lmm-pcqa: Assisting point cloud quality assessment with lmm. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 7783-7792, 2024. 8", + "[57] Wei Zhou, Qi Yang, Wu Chen, Qiuping Jiang, Guangtao Zhai, and Weisi Lin. Blind quality assessment of dense 3d point clouds with structure guided resampling. ACM Transactions on Multimedia Computing, Communications and Applications, 2024. 2" + ], + "bbox": [ + 91, + 90, + 483, + 457 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "21424", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/HybridMQA_ Exploring Geometry-Texture Interactions for Colored Mesh Quality Assessment/a45af5e0-e0d7-4402-9367-2386d583248b_model.json b/2025/HybridMQA_ Exploring Geometry-Texture Interactions for Colored Mesh Quality Assessment/a45af5e0-e0d7-4402-9367-2386d583248b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5a28f9efd55b30e61d3fcafc4ab236b2ba658e8d --- /dev/null +++ b/2025/HybridMQA_ Exploring Geometry-Texture Interactions for Colored Mesh Quality Assessment/a45af5e0-e0d7-4402-9367-2386d583248b_model.json @@ -0,0 +1,2312 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.126, + 0.13, + 0.873, + 0.177 + ], + "angle": 0, + "content": "HybridMQA: Exploring Geometry-Texture Interactions for Colored Mesh Quality Assessment" + }, + { + "type": "text", + "bbox": [ + 0.2, + 0.204, + 0.802, + 0.257 + ], + "angle": 0, + "content": "Armin Shafiee Sarvestani,\\* Sheyang Tang\\*, Zhou Wang \nDepartment of Electrical and Computer Engineering, University of Waterloo \n{a5shafie, sheyang.tang, zhou.wang}@uwaterloo.ca" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.323, + 0.485, + 0.673 + ], + "angle": 0, + "content": "Mesh quality assessment (MQA) models play a critical role in the design, optimization, and evaluation of mesh operation systems in a wide variety of applications. Current MQA models, whether model-based methods using topology-aware features or projection-based approaches working on rendered 2D projections, often fail to capture the intricate interactions between texture and 3D geometry. We introduce HybridMQA, a first-of-its-kind hybrid full-reference colored MQA framework that integrates model-based and projection-based approaches, capturing complex interactions between textural information and 3D structures for enriched quality representations. Our method employs graph learning to extract detailed 3D representations, which are then projected to 2D using a novel feature rendering process that precisely aligns them with colored projections. This enables the exploration of geometry-texture interactions via cross-attention, producing comprehensive mesh quality representations. Extensive experiments demonstrate HybridMQA's superior performance across diverse datasets, highlighting its ability to effectively leverage geometry-texture interactions for a thorough understanding of mesh quality. Our project website is available at https://arshafiee.github.io/hybridmqa/." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.701, + 0.222, + 0.716 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.726, + 0.483, + 0.879 + ], + "angle": 0, + "content": "Advancements in 3D capture and display technologies have sparked a growing interest in immersive media. 3D meshes, a key form of 3D media, are widely used in applications like virtual and augmented reality [20, 49], gaming, animation, medical modeling [37], and generative 3D content creation [21, 35]. A mesh, comprising triangular faces formed by vertices, is colorized by assigning RGB colors to each vertex (vertex-color mesh) or applying a 2D texture map with UV coordinates (textured mesh). The demand for colored meshes calls for high-quality acquisition [3], compress-" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.293, + 0.903, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.475, + 0.907, + 0.533 + ], + "angle": 0, + "content": "Figure 1. Interactions between texture and geometry. top: complex texture makes the geometry distortion imperceptible. bottom: modifying the geometry affects the appearance of texture distortion. Right half (gray) of each object represents geometry." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.565, + 0.907, + 0.672 + ], + "angle": 0, + "content": "sion [11, 24], and transmission [25]. However, these processes often introduce artifacts that degrade visual quality, highlighting the need for robust mesh quality assessment (MQA) methods. Full-reference (FR) approaches, which take distorted meshes and their pristine references as input and generate a quality score by comparing their visual quality, are essential for accurate mesh quality assessment." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.903 + ], + "angle": 0, + "content": "The perceived quality of a 3D colored mesh is affected by its geometry and texture. Different mesh processing operations cause diverse geometrical and texture distortions that degrade the visual quality of meshes by perturbing the interactions between the object's shape and color. Figure 1 illustrates an example of such geometry-texture interactions, where either the geometry or texture can affect the visual appearance of distortions in the other. In the top row, the easily visible geometry distortion becomes imperceptible when we replace the uniform texture with a complex one. In the bottom row, the noisy texture patterns become smooth when we modify the geometry. This highlights the need for MQA methods that capture these complex geometry-texture interaction—a significant factor largely overlooked by existing methods." + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.218, + 0.9 + ], + "angle": 0, + "content": "*Equal contribution" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21414" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.089, + 0.478, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.263, + 0.483, + 0.361 + ], + "angle": 0, + "content": "Figure 2. Reference and distorted meshes under geometry distortions. Although the distorted meshes (hawk and bowl) have distinct visual qualities (different mean opinion scores (MOS)), Graphics-LPIPS [27] assigns similar scores and reverses their ranking. HybridMQA aligns well with human perception as it understands the mesh's geometry properties. Blue boxes denote the same regions across viewpoints." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.388, + 0.485, + 0.902 + ], + "angle": 0, + "content": "MQA methods may be generally categorized as model-based and projection-based depending on their operation space. Model-based MQA operates directly in the 3D vertex space to extract topology-aware quality features [18, 42]. Their performance is limited due to the lack of access to the object's actual appearance (rendered projections). To compensate, they either (1) apply image quality assessment (IQA) to 2D texture maps [13, 38]; or (2) define color quality descriptors for vertices or faces [28, 48, 53]. However, both fail to capture the final appearance human viewers see. Projection-based MQA, on the other hand, operates on rendered 2D projections [19, 27, 55], hence more effective at assessing texture distortions. Nonetheless, they lack a 3D understanding of the object's topology, which is crucial for accurate quality evaluation. Although underlying geometry distortions may be less noticeable in certain projections (Fig. 1, top row), they become apparent to human viewers when the object is viewed from different angles. Humans intuitively perceive underlying structures: connecting viewpoints, incorporating 3D cues, and detecting distortions that may be imperceptible in isolated 2D projections [6, 31, 57]. This 3D understanding is hard to obtain with projections alone due to the nature of operation space. Fig. 2 illustrates this: the geometry distortion (perturbing vertex positions) is more noticeable along the contours of the hawk and bowl than within their inner regions in 2D projections. However, humans naturally identify viewpoint connections (blue boxes) as objects rotate in 3D, perceiving geometry distortions beneath the texture. Consequently, projection-based MQA metrics (Graphics-LPIPS [27]), which focus on pictorial differences, struggle to detect and penalize these distortions, resulting in wrong rankings (Fig. 2). In contrast, HybridMQA extracts quality-related features from mesh surfaces in 3D, enabling a more accurate MQA." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.393 + ], + "angle": 0, + "content": "In this paper, we propose a novel hybrid FR MQA method, namely HybridMQA, that integrates model- and projection-based approaches to explore geometry-texture interactions for comprehensive MQA. Specifically, a base encoder extracts initial 3D features from 2D texture, normal, and vertex maps, which initialize a feature graph based on mesh connectivity. A graph convolutional network (GCN) then learns detailed 3D surface representations, building a 3D understanding of the object's topology. These surface representations are then complemented by textural information from the mesh's renderings, enabling a hybrid integration of model- and projection-based methods across all mesh data modalities: 2D maps, 3D structure, and colored renderings. Additionally, to explore the intricate geometry-texture interactions between the two operational spaces, we propose a novel feature projection technique that renders 2D projections of surface representations from the graph, precisely aligned with colored renderings. This alignment facilitates the exploration of geometry-texture interactions via cross-attention. Below are our contributions:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.395, + 0.905, + 0.456 + ], + "angle": 0, + "content": "- We propose the first hybrid FR MQA method that consolidates the strengths of model- and projection-based approaches across all mesh data modalities, enabling a comprehensive understanding of mesh quality." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.456, + 0.905, + 0.529 + ], + "angle": 0, + "content": "- We make the first attempt to explore geometry-texture interactions for MQA, drawing meaningful connections between the two domains. Our method shows the importance of leveraging such interactions to build reliable MQA methods." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.531, + 0.905, + 0.591 + ], + "angle": 0, + "content": "- We propose a novel feature projection technique to render 2D projections of 3D surface representations from graph, aligned with the mesh's colored renderings, establishing pixel-to-pixel correspondence to explore interactions." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.591, + 0.905, + 0.651 + ], + "angle": 0, + "content": "- Our model outperforms state-of-the-art FR MQA methods, aligning with human perception and generalizing better. The results highlight the effectiveness of capturing geometry-texture interactions to achieve accurate MQA." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.395, + 0.905, + 0.651 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.669, + 0.663, + 0.685 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.696, + 0.85, + 0.712 + ], + "angle": 0, + "content": "2.1. Model-based Mesh Quality Assessment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.718, + 0.905, + 0.854 + ], + "angle": 0, + "content": "Model-based MQA methods operate directly in 3D, with most existing methods designed for uncolored meshes [4, 5, 17, 18, 39, 41, 42], resulting in suboptimal performance for colored meshes. Early approaches, such as Hausdorff distance [4] and mean squared error (MSE) [5], use Euclidean distance as a quality measure, while Lavoué et al. [17, 18] employ curvature statistics. Other methods use local curvature and roughness pooling [39, 42] or employ dihedral angles as surface quality indicators [41]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "However, the rise of colored meshes drives the need for color integration in MQA methods. Tian and AlRegib [38] and Guo et al. [13] apply MSE and MS-SSIM [44] to" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21415" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.348 + ], + "angle": 0, + "content": "2D texture maps, combining these with geometric descriptors to evaluate colored meshes. Nevertheless, 2D texture maps do not represent the post-rendering appearance of 3D meshes and carry little semantic information, leading to suboptimal performance. In contrast, a second group of methods defines color quality descriptors on per-vertex or per-face color values [9, 28, 48, 53]. Nehme et al. [28] propose CMDM, a model for vertex-color meshes that extracts multi-scale color features from vertex colors. Similarly, Zhang et al. [53] apply statistical measures on vertex colors for no-reference MQA. Fu et al. [9] also sample vertex colors and propose an efficient surface sampling approach to convert meshes into point clouds for quality evaluation. Finally, Yang et al. [48] develop GeodesicPSIM, using surface colors to create textured patches for feature extraction. However, these methods fail to capture the rendered appearance of 3D objects as perceived by ultimate human viewers." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.349, + 0.482, + 0.456 + ], + "angle": 0, + "content": "Overall, while model-based MQA methods benefit from a profound understanding of the object's topology, their performance is constrained by the lack of access to the object's colored appearance. We address this issue by introducing a novel hybrid MQA method that complements the 3D awareness of model-based methods with textural representations derived from the colored appearance of 3D meshes." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.469, + 0.458, + 0.485 + ], + "angle": 0, + "content": "2.2. Projection-based Mesh Quality Assessment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.492, + 0.483, + 0.599 + ], + "angle": 0, + "content": "In projection-based MQA, quality is assessed on rendered 2D projections of 3D meshes, allowing well-established IQA methods such as PSNR [43], SSIM [45], and VIF [34] to be adapted for MQA [7, 32, 47]. However, IQA methods perform poorly in MQA as they are tailored for natural scenes, while 3D meshes involve different distortions, and their 2D projections differ statistically from natural images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.483, + 0.901 + ], + "angle": 0, + "content": "A few projection-based methods have been proposed for meshes without color [1, 2], and they expectedly perform poorly on colored meshes. To address this, Nehme et al. [27] propose Graphics-LPIPS, the first projection-based MQA method designed for colored meshes, which builds on LPIPS [52] and uses pre-trained AlexNet [15] to extract quality features from 2D projections. Similarly, Zhang et al. [55] use Swin Transformer [22] with an efficient mini-patch sampling process for no-reference MQA. Lee et al. [19] introduce 3D-PSSIM, which uses a framework similar to Graphics-LPIPS to extract textural information, while complementing it with geometry-aware information derived from the 2D projection space. Nevertheless, these projection-based methods lack 3D understanding of the object's topology. Our approach overcomes this through a well-defined and end-to-end trainable GCN that operates on vertices in the 3D space. Furthermore, these methods fail to account for interactions between the mesh's textural information and the underlying 3D geometry, which limits their performance in detecting geometry-involved distor" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "tions. We address this by proposing a cross-attention framework that draws connections between the two domains." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.135, + 0.688, + 0.153 + ], + "angle": 0, + "content": "3. Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.161, + 0.907, + 0.389 + ], + "angle": 0, + "content": "Figure 3a shows the overall framework of HybridMQA, consisting of model and texture branches and a quality encoder. In the texture branch, colored projections of the input mesh (reference or distorted) are rendered from six perpendicular viewpoints. In the model branch, 2D normal, vertex, and texture maps are processed by a base encoder to extract initial 3D features, which initialize a graph based on mesh connectivity. A GCN then learns detailed 3D surface representations. At the same viewpoints as the texture branch, we render 3D feature projections from the graph, ensuring pixel-to-pixel alignment with the colored projections. Finally, the quality encoder processes both sets of projections, exploring geometry-texture interactions to produce a comprehensive quality representation of the mesh. Overall, the proposed HybridMQA can be expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.559, + 0.4, + 0.905, + 0.417 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {f} _ {\\text {m e s h}} = \\text {H y b r i d M Q A} (\\mathcal {M} (\\boldsymbol {V}, \\boldsymbol {E}, \\boldsymbol {U}, \\boldsymbol {T} _ {t}); \\boldsymbol {\\theta}). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.429, + 0.906, + 0.549 + ], + "angle": 0, + "content": "Here \\(\\mathcal{M}\\) is the input mesh, \\(\\pmb{\\theta}\\) represents model parameters, and \\(f_{mesh}\\) denotes the final quality representation. The mesh \\(\\mathcal{M}\\) consists of vertices \\(\\pmb{V}\\), vertex connectivity \\(\\pmb{E}\\), UV coordinates \\(\\pmb{U}\\), and the 2D texture map \\(\\pmb{T}_t\\). Vertices and UV coordinates are defined as \\(\\pmb{V} = \\{\\pmb{v}_i\\}_{i=1}^N\\), \\(\\pmb{v}_i \\in \\mathbb{R}^3\\) and \\(\\pmb{U} = \\{\\pmb{u}_i\\}_{i=1}^N\\), \\(\\pmb{u}_i \\in \\mathbb{R}^2\\), where \\(N\\) is the number of vertices, and \\(\\pmb{v}_i\\) and \\(\\pmb{u}_i\\) are the positions of the \\(i\\)-th vertex in 3D space and 2D map, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.55, + 0.906, + 0.626 + ], + "angle": 0, + "content": "Finally, as shown in Fig. 3b, HybridMQA generates quality representations for both reference and distorted meshes. For full-reference quality regression, the absolute differences between these representations are fed into a fully connected network (FCN) to obtain a quality score." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.635, + 0.656, + 0.65 + ], + "angle": 0, + "content": "3.1. Base Encoder" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.658, + 0.907, + 0.84 + ], + "angle": 0, + "content": "Given an input mesh \\(\\mathcal{M}\\), we first project its 3D geometrical attributes (normals and vertex positions) into 2D maps aligned with the texture map using barycentric interpolation and UV mapping. The shared UV coordinates \\(\\pmb{U}\\) ensure alignment between the resulting maps (normal and vertex) and the texture map \\(\\pmb{T}_t\\). These aligned maps are concatenated and processed by a convolutional neural network (CNN) base encoder to jointly capture textural and geometrical information, learning initial quality-aware representations of the mesh's 3D shape and surface (Fig. 3a). This approach of projecting geometrical data into 2D maps has been successfully applied in mesh super-resolution [46]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.785, + 0.864 + ], + "angle": 0, + "content": "3.2. Graph Convolutional Network" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Due to vertex-neighborhood discontinuities in 2D maps, we use graph learning to deepen our model's understanding" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21416" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.699, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.315, + 0.282, + 0.473, + 0.295 + ], + "angle": 0, + "content": "(a) HybridMQA's framework." + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.094, + 0.907, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.743, + 0.283, + 0.878, + 0.295 + ], + "angle": 0, + "content": "(b) FR quality regression." + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.306, + 0.907, + 0.378 + ], + "angle": 0, + "content": "Figure 3. (a) Overview of HybridMQA. In the model branch, a base encoder extracts 3D features from the mesh's 2D maps, initializing a mesh graph. A GCN extracts 3D surface representations, which are rendered as 2D projections aligned with the colored projections from the texture branch. A quality encoder then captures geometry-texture interactions between the two branches, producing the final mesh quality representation. (b) Full-reference (FR) quality regression, where the absolute difference of HybridMQA's mesh quality representations for reference and distorted meshes is mapped to a quality score via a fully connected network (FCN)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.403, + 0.482, + 0.449 + ], + "angle": 0, + "content": "of the mesh's 3D structure via message-passing between neighboring vertices in 3D, enabling them to share useful quality-related information and highlight abnormalities." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.45, + 0.483, + 0.497 + ], + "angle": 0, + "content": "Graph initialization. Given the base encoder's output feature map with \\( C_1 \\) channels, we sample a \\( C_1 \\)-dimensional feature vector for each mesh vertex using UV mapping:" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.511, + 0.483, + 0.531 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {f} _ {i} = \\boldsymbol {T} _ {B E} \\left[ u _ {i} ^ {h}, u _ {i} ^ {w} \\right], i \\in \\{1, \\dots , N \\}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.543, + 0.483, + 0.635 + ], + "angle": 0, + "content": "Here, \\( \\pmb{f}_i \\in \\mathbb{R}^{C_1} \\) is the sampled feature for \\( i \\)-th vertex, \\( u_i^h \\) and \\( u_i^w \\) are its UV coordinates, and \\( T_{BE} \\) denotes base encoder's output feature maps. For vertices with multiple UV mappings, we average the sampled features. This process produces a feature graph with vertices initialized by sampled features and edges defined by mesh connectivity \\( \\pmb{E} \\)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.637, + 0.483, + 0.683 + ], + "angle": 0, + "content": "Graph update. We apply a GCN to integrate features from neighboring vertices to learn surface properties. As proposed by Morris et al. [26], we update vertex features by:" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.702, + 0.483, + 0.737 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {f} _ {i} ^ {\\prime} = \\boldsymbol {\\theta} _ {1} \\boldsymbol {f} _ {i} + \\boldsymbol {\\theta} _ {2} \\sum_ {j \\in \\Gamma (i)} e _ {j, i} \\boldsymbol {f} _ {j}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.749, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Here, \\( f_{i}^{\\prime} \\in \\mathbb{R}^{C_{2}} \\) is the updated feature vector for the \\( i \\)-th vertex, \\( \\Gamma(i) \\) is its 1-ring neighborhood, \\( \\theta_{1}, \\theta_{2} \\in \\mathbb{R}^{C_{2} \\times C_{1}} \\) are learnable weights, and \\( e_{j,i} \\) is the edge weight between vertices \\( i \\) and \\( j \\), defined as the inverse of their Euclidean distance. Through training, the GCN refines vertex feature embeddings to learn quality-aware 3D surface representations, which are later complemented by textural embeddings to deliver a hybrid and comprehensive quality assessment. This approach of graph learning has been effectively applied to mesh texture downsampling [29]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.403, + 0.757, + 0.42 + ], + "angle": 0, + "content": "3.3. Feature & RGB Rendering" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.425, + 0.907, + 0.501 + ], + "angle": 0, + "content": "We render multiple viewpoints of the 3D surface representations into 2D projections aligned with the mesh's colored projections. This novel feature graph rendering enables us to capture geometry-texture interactions between the two branches for a comprehensive mesh quality representation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.502, + 0.906, + 0.623 + ], + "angle": 0, + "content": "In the texture branch (Fig. 3a), colored projections are rendered after normalizing mesh vertex positions to fit within a unit cube. This standardization allows fixed camera positions for any mesh. Six virtual cameras are placed on the cube's surfaces, facing the object to cover all angles. With PyTorch3D's [30] soft Phong Shader, we render six perpendicular colored projections, employing directional or ambient light to match the conditions of subjective tests." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.624, + 0.907, + 0.806 + ], + "angle": 0, + "content": "The model branch employs the same camera setup to ensure alignment with colored projections. Our novel feature graph rendering customizes PyTorch3D's differentiable renderer to render six perpendicular \\(3D\\) feature projections from the graph, enabling gradient backpropagation to the GCN and base encoder. To focus on raw vertex features interpolated on the mesh surface, we remove shadows and specular effects, setting diffuse and specular reflectivity to zero and using ambient light. Thus we define the mesh as a vertex-color mesh with \\(C\\)-dimensional vertex features (instead of RGB colors) to render the 3D feature projections with hard Phong shader [30]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.818, + 0.679, + 0.833 + ], + "angle": 0, + "content": "3.4. Quality Encoder" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.903 + ], + "angle": 0, + "content": "We design a quality encoder that processes the aligned projections from model and texture branches to output an expressive quality representation for the mesh. It captures the geometry-texture interactions between the two sets of pro" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21417" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.089, + 0.538, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.269, + 0.334, + 0.375, + 0.347 + ], + "angle": 0, + "content": "(a) Quality encoder." + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.094, + 0.895, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.669, + 0.334, + 0.815, + 0.346 + ], + "angle": 0, + "content": "(b) Cross-attention module." + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.359, + 0.908, + 0.416 + ], + "angle": 0, + "content": "Figure 4. (a) The quality encoder. The 3D feature and color projections are divided into valid aligned patches and fed into respective encoders to obtain multiscale 3D surface and color representations. The cross-attention modules capture interactions between these representations, which are then concatenated with 3D feature embeddings directly extracted from the patches to form the final mesh quality representation. (b) The cross-attention module consists of two transformer blocks, where we switch the roles of the two inputs." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.441, + 0.484, + 0.517 + ], + "angle": 0, + "content": "jections through multi-scale cross-attentions. As discussed in Sec. 1, understanding these interactions is key to a robust and effective MQA. As detailed in Fig. 4a, given model branch's 3D feature projections \\( \\mathbf{I}_m \\) and texture branch's colored projections \\( \\mathbf{I}_t \\), the quality encoder \\( \\psi \\) outputs" + }, + { + "type": "equation", + "bbox": [ + 0.204, + 0.524, + 0.482, + 0.541 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {f} _ {\\text {m e s h}} = \\psi \\left(\\boldsymbol {I} _ {m}, \\boldsymbol {I} _ {t}; \\boldsymbol {\\theta} _ {\\psi}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.547, + 0.483, + 0.576 + ], + "angle": 0, + "content": "where \\(f_{mesh}\\) denotes the final quality representation, and \\(\\theta_{\\psi}\\) denotes the learnable parameters." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.577, + 0.484, + 0.623 + ], + "angle": 0, + "content": "Patch Selection. Given two sets of projections, we extract non-overlapping patches and discard those with less than \\(10\\%\\) non-background pixels. This yields \\(2P\\) patches:" + }, + { + "type": "equation", + "bbox": [ + 0.171, + 0.627, + 0.482, + 0.647 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\boldsymbol {I}} _ {m} = \\left\\{\\boldsymbol {I} _ {m, p} \\right\\} _ {p = 1} ^ {P}, \\tilde {\\boldsymbol {I}} _ {t} = \\left\\{\\boldsymbol {I} _ {t, p} \\right\\} _ {p = 1} ^ {P}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.653, + 0.483, + 0.684 + ], + "angle": 0, + "content": "Here, \\(\\tilde{I}_m\\) and \\(\\tilde{I}_t\\) denote aligned patches extracted from 3D feature and colored projections, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.685, + 0.484, + 0.747 + ], + "angle": 0, + "content": "3D Feature & Image Encoding. Next, we feed pairs of aligned patches \\( I_{m,p} \\) and \\( I_{t,p} \\) to a 3D feature encoder \\( \\phi \\) and an image encoder \\( \\varphi \\), respectively, to extract multiscale quality-aware representations \\( F_{m,p} \\) and \\( F_{t,p} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.752, + 0.483, + 0.777 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {F} _ {m, p} = \\left\\{\\boldsymbol {f} _ {m, p} ^ {s} | s = 1, \\dots , 5 \\right\\} = \\phi (\\boldsymbol {I} _ {m, p}; \\boldsymbol {\\theta} _ {\\phi}), \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.144, + 0.773, + 0.425, + 0.79 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {F} _ {t, p} = \\left\\{\\boldsymbol {f} _ {t, p} ^ {s} | s = 1, \\dots , 5 \\right\\} = \\varphi (\\boldsymbol {I} _ {t, p}; \\boldsymbol {\\theta} _ {\\varphi}).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.841 + ], + "angle": 0, + "content": "Here, \\(\\theta_{\\phi}\\) and \\(\\theta_{\\varphi}\\) are learnable parameters of the two encoders, and \\(s\\) denotes scale index. \\(1\\times 1\\) convolutions are used to adjust the channel dimension of representations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Cross-attention Modules. We apply cross-attention to capture the interactions between aligned 3D surface representations \\( \\pmb{f}_{m,p}^{s} \\) and color representations \\( \\pmb{f}_{t,p}^{s} \\) at each scale. Figure 4b depicts our proposed module with two transformer" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.441, + 0.907, + 0.534 + ], + "angle": 0, + "content": "blocks [40], where we alternate the query and key-value roles of \\( f_{m,p}^{s} \\) and \\( f_{t,p}^{s} \\) to explore their interactions and mutual influence. This simulates how one domain's representation affects the impact of the other domain's representation on perceptual quality, as discussed in Sec. 1. We concatenate outputs to form single-scale quality representation \\( f_{p}^{s} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.533, + 0.906, + 0.641 + ], + "angle": 0, + "content": "Mesh Quality Representation. To further exploit the model branch's understanding of mesh's 3D structure and surface properties, we directly extract representation \\(\\hat{f}_p\\) from model branch patches via a convolutional block. \\(\\hat{f}_p\\) is concatenated with the five single-scale representations from cross-attention, and averaged over all patches to obtain the final mesh quality representation \\(f_{mesh}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.578, + 0.648, + 0.907, + 0.691 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {f} _ {\\text {m e s h}} = \\frac {1}{P} \\sum_ {p = 1} ^ {P} \\left(\\boldsymbol {f} _ {p} ^ {1} \\oplus \\dots \\oplus \\boldsymbol {f} _ {p} ^ {5} \\oplus \\hat {\\boldsymbol {f}} _ {p}\\right). \\tag {7}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.699, + 0.822, + 0.716 + ], + "angle": 0, + "content": "3.5. Quality Regression & Optimization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.906, + 0.841 + ], + "angle": 0, + "content": "As shown in Fig. 3b, we feed the reference and distorted meshes into HybridMQA separately and map the absolute difference of their quality representations \\((|f_{mesh}^{ref} - f_{mesh}^{dis}|)\\) to a quality score via an FCN for full-reference MQA. The model is optimized using a loss function with two terms: mean absolute error (MAE) and rank loss. While MAE ensures accurate quality predictions, rank loss helps differentiate closely rated samples within a mini-batch [36]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.842, + 0.783, + 0.856 + ], + "angle": 0, + "content": "MAE Loss. The MAE loss is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.625, + 0.863, + 0.907, + 0.904 + ], + "angle": 0, + "content": "\\[\nL _ {m a e} = \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\left| q _ {i} - q _ {i} ^ {\\prime} \\right|, \\tag {8}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21418" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "where \\( q_{i} \\) and \\( q_{i}^{\\prime} \\) denote the predicted and ground truth quality scores of the \\( i \\)-th sample in the batch, with \\( B \\) being the batch size. The ground truth is the mean opinion score (MOS) obtained in subjective tests and normalized to [0, 1]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.153, + 0.483, + 0.183 + ], + "angle": 0, + "content": "Rank Loss. Since ranking is not differentiable, as proposed by Sun et al. [36], we approximate the rank value as" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.192, + 0.482, + 0.249 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {r a n k} ^ {i, j} = \\max \\left(0, \\left| q _ {i} ^ {\\prime} - q _ {j} ^ {\\prime} \\right| - e \\left(q _ {i} ^ {\\prime}, q _ {j} ^ {\\prime}\\right) \\cdot \\left(q _ {i} - q _ {j}\\right)\\right), \\\\ e \\left(q _ {i} ^ {\\prime}, q _ {j} ^ {\\prime}\\right) = \\left\\{ \\begin{array}{l l} 1 & q _ {i} ^ {\\prime} \\geq q _ {j} ^ {\\prime} \\\\ - 1 & q _ {i} ^ {\\prime} < q _ {j} ^ {\\prime} \\end{array} , \\right. \\tag {9} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.259, + 0.482, + 0.29 + ], + "angle": 0, + "content": "where \\(i\\) and \\(j\\) denote two samples in a batch, and \\(e(q_i', q_j')\\) is a sign function. The final rank loss is computed as" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.301, + 0.482, + 0.352 + ], + "angle": 0, + "content": "\\[\nL _ {r a n k} = \\frac {1}{B ^ {2} - B} \\sum_ {i = 1} ^ {B} \\sum_ {\\substack {j = 1 \\\\ j \\neq i}} ^ {B} L _ {r a n k} ^ {i, j}. \\tag{10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.362, + 0.462, + 0.377 + ], + "angle": 0, + "content": "Final Loss. The final loss is a weighted sum of the two:" + }, + { + "type": "equation", + "bbox": [ + 0.212, + 0.39, + 0.482, + 0.406 + ], + "angle": 0, + "content": "\\[\nL = L _ {m a e} + \\lambda L _ {r a n k}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.417, + 0.417, + 0.431 + ], + "angle": 0, + "content": "where \\(\\lambda\\) balances the effect of the two loss terms." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.446, + 0.224, + 0.463 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.47, + 0.396, + 0.485 + ], + "angle": 0, + "content": "4.1. Datasets & Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.492, + 0.484, + 0.612 + ], + "angle": 0, + "content": "We validate our method on four publicly available color MQA datasets: Nehme et al. [27], SJTU-TMQA [7], TSMD [47], and CMDM [28] datasets. Nehme et al., the largest available, includes textured meshes with mixed geometric and texture distortions, as do SJTU-TMQA and TSMD. In contrast, CMDM consists of vertex-color meshes with either geometric or color distortions. All datasets use MOS as ground truth." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.484, + 0.749 + ], + "angle": 0, + "content": "We use ResNet50 [14] pre-trained on ImageNet [8] as the image encoder \\(\\varphi\\) and a randomly initialized CNN as the 3D feature encoder \\(\\phi\\). 3D feature and colored projections are rendered at \\(128 \\times 128\\) and \\(512 \\times 512\\) resolutions, with patch sizes of \\(16 \\times 16\\) and \\(64 \\times 64\\), respectively, to ensure alignment. For data augmentation, we randomly perturb camera angles and flip the patches in training to improve robustness and generalization. More details of the datasets and implementation are provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.758, + 0.281, + 0.774 + ], + "angle": 0, + "content": "4.2. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.482, + 0.84 + ], + "angle": 0, + "content": "We use 5-fold cross-validation without overlap between train and test source content and report the median performance across five experiments. This strategy is applied to all learning-based methods for fair comparison." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We compare HybridMQA with 11 model-based and projection-based full-reference MQA methods. Model-based methods include Hausdorff Distance (HD) [4], MSDM2 [18], FMPD [42], GeodesicPSIM [48], and Fu" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.258 + ], + "angle": 0, + "content": "et al. [9]. Projection-based methods include PSNR [43], SSIM [45], VIF [34], LPIPS [52], Graphics-LPIPS [27], and 3D-PSSIM [19]. For a fair comparison, all projection-based methods are evaluated under the same rendering settings as HybridMQA, with no prior assumptions about object orientation. We adopt the Spearman rank-order correlation coefficient (SRCC) and the Pearson linear correlation coefficient (PLCC) to compare all methods. Higher values indicate better performance and a stronger correlation between MOS and predicted quality scores [12]. Further information is provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.271, + 0.707, + 0.287 + ], + "angle": 0, + "content": "4.3.Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.295, + 0.907, + 0.656 + ], + "angle": 0, + "content": "Overall Performance. Table 1 summarizes the performance comparison, from which we make the following observations: (1) HybridMQA outperforms all model-based and projection-based comparison methods across all datasets, including both textured and vertex-color meshes, demonstrating its effectiveness in the quality assessment of colored meshes. Notably, HybridMQA achieves \\(6.5\\%\\) and \\(7.7\\%\\) performance gain in PLCC over the second-best method, 3D-PSSIM, on Nehme et al. and SJTU-TMQA, respectively; (2) In general, projection-based methods outperform model-based methods which lack access to the object's actual appearance. HybridMQA consolidates the advantages of both types by complementing the textural information extracted from colored projections with 3D representations learned in 3D; (3) Unlike CMDM with single-type distortions, Nehme et al., SJTU-TMQA, and TSMD involve mixed-type and hence more complex distortions. Consequently, while most methods perform well on CMDM, their performance does not extend to the other three datasets. In contrast, HybridMQA consistently performs well on all datasets, showing robustness in assessing complex distortions. This is achieved by HybridMQA's understanding of geometry-texture interactions, which are perturbed uniquely by different distortions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.658, + 0.906, + 0.824 + ], + "angle": 0, + "content": "Performance by Distortion Type. In Fig. 5, we compare HybridMQA with Graphics-LPIPS based on distortion types in SJTU-TMQA. In addition to the overall superiority of HybridMQA, we observe that while the two methods achieve comparable performance on texture-only distortions (e.g. JPEG [7]), HybridMQA hugely outperforms Graphics-LPIPS on distortions that only affect geometry (e.g. gn [7] and simpNoTex [7, 10]) or both geometry and texture (e.g. qpqtJPEG [7, 11]). This further demonstrates HybridMQA's proficiency in understanding meshes' 3D geometry and its interactions with textural appearance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Generalizability. We train models on Nehme et al. and TSMD and test them on SJTU-TMQA. Table 2 shows that HybridMQA significantly outperforms other learning-based methods, showing strong generalizability. Notably, when trained on TSMD and tested on SJTU-TMQA, HybridMQA" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21419" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.131, + 0.089, + 0.872, + 0.313 + ], + "angle": 0, + "content": "
TypeMethodNehmé et al. [27]SJTU-TMQA [7]TSMD [47]CMDM [28]
SRCCPLCCSRCCPLCCSRCCPLCCSRCCPLCC
Model-basedHD [4]0.1070.1750.0600.1400.4460.4620.1890.210
MSDM2 [18]0.3350.3440.0500.1200.0450.2550.4150.517
FMPD [42]0.3910.4040.1560.4580.0770.2180.6150.623
GeodesicPSIM [48]----0.8200.820--
Fu et al. [9]0.6880.696------
Projection-basedPSNR [43]0.3530.3750.2990.2870.7140.7110.8300.839
SSIM [45]0.2100.2260.3940.2890.6730.6740.8520.861
VIF [34]0.5380.5570.4500.4220.8510.8460.8270.837
LPIPS [52]0.6720.6760.7180.7170.7100.7120.8650.918
Graphics-LPIPS [27]0.7220.7460.7900.7620.8340.8120.8590.925
3D-PSSIM [19]0.8820.8420.8420.832--0.8550.854
HybridHybridMQA0.8920.8970.8870.8960.9120.9190.8970.927
" + }, + { + "type": "table_caption", + "bbox": [ + 0.089, + 0.322, + 0.908, + 0.366 + ], + "angle": 0, + "content": "Table 1. SRCC and PLCC scores of MQA methods on four color MQA benchmark datasets. The scores of GeodesicPSIM [48], Fu et al. [9], and 3D-PSSIM [19] are reported directly from their publications as their implementations are not publicly available. Bold and underlined values denote the best and second-best results per column, respectively." + }, + { + "type": "table", + "bbox": [ + 0.115, + 0.388, + 0.462, + 0.482 + ], + "angle": 0, + "content": "
Trained onNehmé et al.TSMD
SRCCPLCCSRCCPLCC
LPIPS0.5920.5840.7120.695
Graphics-LPIPS0.7310.7340.7840.773
HybridMQA0.8000.7830.8460.811
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.491, + 0.483, + 0.52 + ], + "angle": 0, + "content": "Table 2. Generalization evaluation, where models are trained on Nehme et al. and TSMD and tested on SJTU-TMQA." + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.534, + 0.477, + 0.682 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.691, + 0.483, + 0.748 + ], + "angle": 0, + "content": "Figure 5. SRCC/PLCC performance of HybridMQA and Graphics-LPIPS on various distortion types of SJTU-TMQA dataset. Green distortions affect the geometry or both geometry and texture, while others only impact the texture." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.775, + 0.483, + 0.807 + ], + "angle": 0, + "content": "achieves comparable performance to 3D-PSSIM [19], despite 3D-PSSIM being trained directly on SJTU-TMQA." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.818, + 0.275, + 0.833 + ], + "angle": 0, + "content": "4.4. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We apply GradCAM [33] on 3D feature projections to verify that the model branch effectively captures geometry-aware quality representations. As shown in Fig. 6, the highlighted regions with noticeable artifacts align well with hu" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.39, + 0.907, + 0.578 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.591, + 0.907, + 0.648 + ], + "angle": 0, + "content": "Figure 6. GradCAM results on meshes in the model branch. Highlighted regions exhibit more noticeable artifacts, aligning with human perception and showing the model's effectiveness in capturing geometry-aware quality representations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.907, + 0.764 + ], + "angle": 0, + "content": "man perception, validating the model branch's ability to extract geometry-aware representations. Additionally, by mapping these results onto the mesh topology via differentiable rendering, our method opens up opportunities for optimizing mesh compression or restoration algorithms. More examples are available in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Figure 7 shows the effectiveness of our hybrid model in exploiting interactions between representations learned in texture and model branches. GradCAM is applied before and after cross-attention, showing that the two branches focus on different regions, with the model branch effectively highlighting geometric artifacts. Cross-attention then successfully identifies and attends to perceptually important regions by exploring geometry-texture interactions. Note that the actual cross-attention inputs are patches described" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21420" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.488, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.448, + 0.486, + 0.492 + ], + "angle": 0, + "content": "Figure 7. GradCAM before and after cross-attention. The two branches focus on different regions, and cross-attention effectively integrates them by attending to perceptually important regions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.518, + 0.483, + 0.564 + ], + "angle": 0, + "content": "in Sec. 3.3 and 3.4, and the heatmaps are the union of the GradCAM results from the reference and distorted patches. More examples are available in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.564, + 0.484, + 0.61 + ], + "angle": 0, + "content": "We also perform gMAD competition [23] to qualitatively compare HybridMQA with Graphics-LPIPS. The results are available in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.621, + 0.258, + 0.635 + ], + "angle": 0, + "content": "4.5. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.483, + 0.689 + ], + "angle": 0, + "content": "We conduct ablation studies on Nehme et al. [27] to evaluate key components and design choices in our model. Further results are available in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.484, + 0.902 + ], + "angle": 0, + "content": "3D Surface and Textural Representations. To show the contributions of 3D surface and textural representations, we test four configurations: (1) 3D-only, where only 3D representations \\(\\hat{f}\\) and \\(f_{m}\\) are used for mesh quality representation; (2) RGB-only, where only textural representation \\(f_{t}\\) from colored renderings are used; (3) all but excluding \\(f_{m}\\); and (4) excluding \\(\\hat{f}\\). Table 3 shows that all representations contribute to the final performance, with textural representation \\(f_{t}\\) the most significant. Moreover, excluding 3D representations (RGB-only) leads to a significant performance drop, showing the importance of 3D representations and the need to explore their interaction with textural information for accurate MQA. The results also highlight the performance gain of our hybrid approach, as it outperforms both" + }, + { + "type": "table", + "bbox": [ + 0.56, + 0.089, + 0.865, + 0.19 + ], + "angle": 0, + "content": "
NotesffmftSRCCPLCC
3D-only✓✓-0.5860.595
RGB-only--✓0.8200.846
-✓-✓0.8420.849
--✓✓0.8700.874
Proposed✓✓✓0.8920.897
" + }, + { + "type": "table_caption", + "bbox": [ + 0.521, + 0.2, + 0.9, + 0.214 + ], + "angle": 0, + "content": "Table 3. Ablation on 3D Surface & Text. Repr. on Nehme et al." + }, + { + "type": "table", + "bbox": [ + 0.571, + 0.215, + 0.853, + 0.328 + ], + "angle": 0, + "content": "
ConfigurationsSRCCPLCC
w/o Texture map0.8660.872
w/o Normal map0.8630.866
w/o Vertex map0.8720.873
w/o Base Encoder0.8560.851
w/o GCN0.8660.865
HybridMQA (proposed)0.8920.897
" + }, + { + "type": "table_caption", + "bbox": [ + 0.518, + 0.339, + 0.9, + 0.353 + ], + "angle": 0, + "content": "Table 4. Ablation on model branch components on Nehme et al." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.379, + 0.762, + 0.394 + ], + "angle": 0, + "content": "3D-only and RGB-only counterparts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.395, + 0.907, + 0.531 + ], + "angle": 0, + "content": "Model Branch Components. To validate the contributions of model branch components, we test the following configurations: all components but excluding (1) the texture map; (2) the normal map; (3) the vertex map; (4) the Base Encoder; and (5) the GCN. Table 4 shows that all components contribute to learning effective 3D surface representations. Furthermore, the proposed processing units (Base Encoder and GCN) effectively leverage mesh data modalities (2D maps and 3D vertex space) to enhance performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.544, + 0.747, + 0.559 + ], + "angle": 0, + "content": "5. Conclusion & Discussion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.907, + 0.704 + ], + "angle": 0, + "content": "We present a novel hybrid full-reference MQA method that integrates model- and projection-based approaches for enhanced quality assessment. Our model explores interactions between mesh texture and 3D geometry via cross-attention, enabled by a novel feature rendering process that aligns 3D representations with colored projections. Extensive experiments show the effectiveness and superiority of our method, highlighting the importance of 3D understanding and leveraging geometry-texture interactions for reliable MQA." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.909, + 0.903 + ], + "angle": 0, + "content": "A few limitations present room for improvement. Our method relies on perpendicular viewpoints without considering their varying contributions to perceptual quality. Integrating a semantic-aware module could address this by weighting viewpoints based on importance. Also, memory consumption scales with mesh size as the GCN processes the entire mesh graph, limiting efficiency in real-world applications. Sampling techniques [16, 50, 51] could help reduce memory footprint. Future directions include (1) adapting our work for no-reference MQA; (2) generalizing it to point cloud quality assessment [54, 56] by defining graph edges; and (3) exploring applications in perceptually optimized mesh compression, enhancement, and generation." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21421" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.171 + ], + "angle": 0, + "content": "[1] Ilyass Aboulaziz, Aladine Chetouani, Mohammed El Hassouni, and Hocine Cherifi. A blind mesh visual quality assessment method based on convolutional neural network. *Electronic Imaging*, 30:1-5, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.173, + 0.484, + 0.243 + ], + "angle": 0, + "content": "[2] Ilyass Abouelaziz, Aladine Chetouani, Mohammed El Hassouni, Longin Jan Latecki, and Hocine Cherifi. No-reference mesh visual quality assessment via ensemble of convolutional neural networks and compact multi-linear pooling. Pattern Recognition, 100:107174, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.245, + 0.483, + 0.3 + ], + "angle": 0, + "content": "[3] Martin Alain, Emin Zerman, Cagri Ozcinar, and Giuseppe Valenzise. Chapter 1 - introduction to immersive video technologies. In Immersive Video Technologies, pages 3-24. Academic Press, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.302, + 0.483, + 0.37 + ], + "angle": 0, + "content": "[4] Nicolas Aspert, Diego Santa-Cruz, and Touradj Ebrahimi. Mesh: measuring errors between surfaces using the hausdorff distance. In Proceedings. IEEE International Conference on Multimedia and Expo, pages 705-708 vol.1, 2002. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.373, + 0.483, + 0.413 + ], + "angle": 0, + "content": "[5] Paolo Cignoni, Claudio Rocchini, and Roberto Scopigno. Metro: Measuring error on simplified surfaces. Computer Graphics Forum, 17(2):167-174, 1998. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.416, + 0.484, + 0.485 + ], + "angle": 0, + "content": "[6] Sahin Coskun, Gokce Nur Yilmaz, Federica Battisti, Musaed Alhussein, and Saiful Islam. Measuring 3d video quality of experience (qoe) using a hybrid metric based on spatial resolution and depth cues. Journal of Imaging, 9(12):281, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.488, + 0.483, + 0.57 + ], + "angle": 0, + "content": "[7] Bingyang Cui, Qi Yang, Kaifa Yang, Yiling Xu, Xiaozhong Xu, and Shan Liu. Sjtu-tmqa: A quality assessment database for static mesh with texture map. In ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 7875-7879, 2024. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.573, + 0.483, + 0.628 + ], + "angle": 0, + "content": "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.63, + 0.483, + 0.713 + ], + "angle": 0, + "content": "[9] Chunyang Fu, Xiang Zhang, Thuong Nguyen-Canh, Xiaozhong Xu, Ge Li, and Shan Liu. Surface-sampling based objective quality assessment metrics for meshes. In ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5, 2023. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.715, + 0.484, + 0.785 + ], + "angle": 0, + "content": "[10] Michael Garland and Paul S. Heckbert. Surface simplification using quadric error metrics. In Proceedings of the 24th Annual Conference on Computer Graphics and Interactive Techniques, page 209-216, USA, 1997. ACM Press/Addison-Wesley Publishing Co. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.787, + 0.481, + 0.814 + ], + "angle": 0, + "content": "[11] Google. Draco 3d data compression. https://google.github.io/draco/. Accessed: 2024-09-04. 1,6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.816, + 0.483, + 0.87 + ], + "angle": 0, + "content": "[12] Video Quality Experts Group et al. Final report from the video quality experts group on the validation of objective models of video quality assessment, phase II. VQEG, 2003. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[13] Jinjiang Guo, Vincent Vidal, Irene Cheng, Anup Basu, Atilla Baskurt, and Guillaume Lavoue. Subjective and objective" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "visual quality assessment of textured 3d meshes. ACM Trans. Appl. Percept., 14(2), 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.178, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[15] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.234, + 0.905, + 0.288 + ], + "angle": 0, + "content": "[16] Itai Lang, Asaf Manor, and Shai Avidan. Samplenet: Differentiable point cloud sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7578-7588, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.289, + 0.905, + 0.355 + ], + "angle": 0, + "content": "[17] Guillaume Lavoué, Elisa Drelie Gelasca, Florent Dupont, Atilla Baskurt, and Touradj Ebrahimi. Perceptually driven 3d distance metrics with application to watermarking. Applications of Digital Image Processing XXIX, 6312:63120L, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.358, + 0.905, + 0.398 + ], + "angle": 0, + "content": "[18] Guillaume Lavoué. A multiscale metric for 3d mesh visual quality assessment. Computer Graphics Forum, 30(5):1427-1437, 2011. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.401, + 0.905, + 0.469 + ], + "angle": 0, + "content": "[19] Seongmin Lee, Jiwoo Kang, Sanghoon Lee, Weisi Lin, and Alan Conrad Bovik. 3d-pssim: Projective structural similarity for 3d mesh quality assessment robust to topological irregularities. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-16, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.905, + 0.511 + ], + "angle": 0, + "content": "[20] Jie Li and Pablo Cesar. Chapter 22 - social virtual reality (vr) applications and user experiences. In Immersive Video Technologies, pages 609-648. Academic Press, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.512, + 0.905, + 0.58 + ], + "angle": 0, + "content": "[21] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Mukund Varma T, Zexiang Xu, and Hao Su. One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. Advances in Neural Information Processing Systems, 36, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.581, + 0.905, + 0.65 + ], + "angle": 0, + "content": "[22] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10012-10022, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.651, + 0.905, + 0.719 + ], + "angle": 0, + "content": "[23] Kede Ma, Zhengfang Duanmu, Zhou Wang, Qingbo Wu, Wentao Liu, Hongwei Yong, Hongliang Li, and Lei Zhang. Group maximum differentiation competition: Model comparison with few samples. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(4):851-864, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.72, + 0.905, + 0.762 + ], + "angle": 0, + "content": "[24] Adrien Maglo, Guillaume Lavoué, Florent Dupont, and Céline Hudelot. 3d mesh compression: Survey, comparisons, and emerging trends. ACM Comput. Surv., 47(3), 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.763, + 0.905, + 0.803 + ], + "angle": 0, + "content": "[25] Adrien Maglo, Guillaume Lavoué, Florent Dupont, and Céline Hudelot. 3d mesh compression: Survey, comparisons, and emerging trends. ACM Comput. Surv., 47(3), 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.804, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[26] Christopher Morris, Martin Ritzert, Matthias Fey, William L. Hamilton, Jan Eric Lenssen, Gaurav Rattan, and Martin Grohe. Weisfeiler and leman go neural: Higher-order graph neural networks. Proceedings of the AAAI Conference on Artificial Intelligence, 33(01):4602-4609, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[27] Yana Nehme, Johanna Delanoy, Florent Dupont, JeanPhilippe Farrugia, Patrick Le Callet, and Guillaume Lavoué." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21422" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "Textured mesh quality assessment: Large-scale dataset and deep learning-based quality metric. ACM Trans. Graph., 42 (3), 2023. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.135, + 0.483, + 0.217 + ], + "angle": 0, + "content": "[28] Yana Nehme, Florent Dupont, Jean-Philippe Farrugia, Patrick Le Callet, and Guillaume Lavoué. Visual quality of 3d meshes with diffuse colors in virtual reality: Subjective and objective evaluation. IEEE Transactions on Visualization and Computer Graphics, 27(3):2202-2219, 2021. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.218, + 0.482, + 0.273 + ], + "angle": 0, + "content": "[29] Sai Karthikey Pentapati, Anshul Rai, Arkady Ten, Chaitanya Atluru, and Alan Bovik. Geoscaler: Geometry and rendering-aware downsampling of 3d mesh textures. arXiv preprint arXiv:2311.16581, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.482, + 0.328 + ], + "angle": 0, + "content": "[30] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3d deep learning with pytorch3d. arXiv:2007.08501, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.482, + 0.412 + ], + "angle": 0, + "content": "[31] Stephan Reichelt, Ralf Häussler, Gerald Füfterer, and Norbert Leister. Depth cues in human visual perception and their realization in 3d displays. In Three-Dimensional Imaging, Visualization, and Display 2010 and Display Technologies and Applications for Defense, Security, and Avionics IV, pages 92–103. SpIE, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.413, + 0.482, + 0.468 + ], + "angle": 0, + "content": "[32] Armin Shafiee Sarvestani, Wei Zhou, and Zhou Wang. Perceptual crack detection for rendered 3d textured meshes. In 2024 16th International Conference on Quality of Multimedia Experience (QoMEX), pages 1-7, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.482, + 0.55 + ], + "angle": 0, + "content": "[33] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.552, + 0.482, + 0.594 + ], + "angle": 0, + "content": "[34] Hamid R. Sheikh and Alan C. Bovik. Image information and visual quality. IEEE Transactions on Image Processing, 15 (2):430-444, 2006. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.482, + 0.676 + ], + "angle": 0, + "content": "[35] Yawar Siddiqui, Tom Monnier, Filippos Kokkinos, Mahendra Kariya, Yanir Kleiman, Emilien Garreau, Oran Gafni, Natalia Neverova, Andrea Vedaldi, Roman Shapovalov, et al. Meta 3d assetgen: Text-to-mesh generation with high-quality geometry, texture, and pbr materials. arXiv preprint arXiv:2407.02445, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.678, + 0.482, + 0.748 + ], + "angle": 0, + "content": "[36] Wei Sun, Xiongkuo Min, Wei Lu, and Guangtao Zhai. A deep learning based no-reference quality assessment model forUGC videos. In Proceedings of the 30th ACM International Conference on Multimedia, page 856-865, New York, NY, USA, 2022. Association for Computing Machinery. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.482, + 0.789 + ], + "angle": 0, + "content": "[37] Yuk Ming Tang and Ho Lun Ho. 3d modeling and computer graphics in virtual reality. In mixed reality and three-dimensional computer graphics. IntechOpen, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.482, + 0.844 + ], + "angle": 0, + "content": "[38] Dihong Tian and Ghassan AlRegib. Batex3: Bit allocation for progressive transmission of textured 3-d models. IEEE Transactions on Circuits and Systems for Video Technology, 18(1):23-35, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[39] Fakhri Torkhani, Kai Wang, and Jean-Marc Chassery. A curvature tensor distance for mesh visual quality assessment. In International Conference on Computer Vision and Graphics, pages 253–263. Springer, 2012. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.16 + ], + "angle": 0, + "content": "[40] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.162, + 0.905, + 0.217 + ], + "angle": 0, + "content": "[41] Libor Váša and Jan Rus. Dihedral angle mesh error: a fast perception correlated distortion measure for fixed connectivity triangle meshes. Computer Graphics Forum, 31(5):1715-1724, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.218, + 0.905, + 0.274 + ], + "angle": 0, + "content": "[42] Kai Wang, Fakhri Torkhani, and Annick Montanvert. A fast roughness-based approach to the assessment of 3d mesh visual quality. Computers & Graphics, 36(7):808-818, 2012. Augmented Reality Computer Graphics in China. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.905, + 0.315 + ], + "angle": 0, + "content": "[43] Zhou Wang and Alan C. Bovik. Mean squared error: Love it or leave it? a new look at signal fidelity measures. IEEE Signal Processing Magazine, 26(1):98-117, 2009. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.316, + 0.905, + 0.371 + ], + "angle": 0, + "content": "[44] Zhou Wang, Eero P Simoncelli, and Alan C Bovik. Multiscale structural similarity for image quality assessment. In The Thirty-Seventh Asilomar Conference on Signals, Systems & Computers, 2003, pages 1398-1402 Vol.2, 2003. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.371, + 0.905, + 0.427 + ], + "angle": 0, + "content": "[45] Zhou Wang, Alan C. Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.905, + 0.495 + ], + "angle": 0, + "content": "[46] Wuyuan Xie, Tengcong Huang, and Miaohui Wang. Mnsnet: multimodal transformer network for 3d surface superresolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12703-12712, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.497, + 0.905, + 0.566 + ], + "angle": 0, + "content": "[47] Qi Yang, Joel Jung, Haiqiang Wang, Xiaozhong Xu, and Shan Liu. Tsmd: A database for static color mesh quality assessment study. In 2023 IEEE International Conference on Visual Communications and Image Processing (VCIP), pages 1-5, 2023. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.567, + 0.905, + 0.621 + ], + "angle": 0, + "content": "[48] Qi Yang, Joel Jung, Xiaozhong Xu, and Shan Liu. Geodesicpsim: Predicting the quality of static mesh with texture map via geodesic patch similarity. arXiv preprint arXiv:2308.04928, 2023. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.622, + 0.905, + 0.676 + ], + "angle": 0, + "content": "[49] Gareth W. Young, Néill O'Dwyer, and Aljosa Smolic. Chapter 21 - volumetric video as a novel medium for creative storytelling. In Immersive Video Technologies, pages 591-607. Academic Press, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.678, + 0.905, + 0.734 + ], + "angle": 0, + "content": "[50] Hanqing Zeng, Hongkuan Zhou, Ajitesh Srivastava, Rajgopal Kannan, and Viktor Prasanna. GraphSAINT: Graph sampling based inductive learning method. In International Conference on Learning Representations, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.905, + 0.802 + ], + "angle": 0, + "content": "[51] Hanqing Zeng, Muhan Zhang, Yinglong Xia, Ajitesh Srivastava, Andrey Malevich, Rajgopal Kannan, Viktor Prasanna, Long Jin, and Ren Chen. Decoupling the depth and scope of graph neural networks. Advances in Neural Information Processing Systems, 34:19665-19679, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[52] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[53] Zicheng Zhang, Wei Sun, Xiongkuo Min, Tao Wang, Wei Lu, and Guangtao Zhai. No-reference quality assessment for" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21423" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.124, + 0.092, + 0.484, + 0.134 + ], + "angle": 0, + "content": "3d colored point cloud and mesh models. IEEE Transactions on Circuits and Systems for Video Technology, 32(11):7618-7631, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.484, + 0.232 + ], + "angle": 0, + "content": "[54] Zicheng Zhang, Wei Sun, Xiongkuo Min, Qiyuan Wang, Jun He, Quan Zhou, and Guangtao Zhai. Mm-pcqa: Multi-modal learning for no-reference point cloud quality assessment. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, IJCAI-23, pages 1759–1767. International Joint Conferences on Artificial Intelligence Organization, 2023. Main Track. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.484, + 0.303 + ], + "angle": 0, + "content": "[55] Zicheng Zhang, Wei Sun, Haoning Wu, Yingjie Zhou, Chunyi Li, Zijian Chen, Xiongkuo Min, Guangtao Zhai, and Weisi Lin. Gms-3dqa: Projection-based grid mini-patch sampling for 3d model quality assessment. ACM Trans. Multimedia Comput. Commun. Appl., 20(6), 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.304, + 0.484, + 0.387 + ], + "angle": 0, + "content": "[56] Zicheng Zhang, Haoning Wu, Yingjie Zhou, Chunyi Li, Wei Sun, Chaofeng Chen, Xiongkuo Min, Xiaohong Liu, Weisi Lin, and Guangtao Zhai. Lmm-pcqa: Assisting point cloud quality assessment with lmm. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 7783-7792, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.388, + 0.484, + 0.458 + ], + "angle": 0, + "content": "[57] Wei Zhou, Qi Yang, Wu Chen, Qiuping Jiang, Guangtao Zhai, and Weisi Lin. Blind quality assessment of dense 3d point clouds with structure guided resampling. ACM Transactions on Multimedia Computing, Communications and Applications, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.458 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21424" + } + ] +] \ No newline at end of file diff --git a/2025/HybridMQA_ Exploring Geometry-Texture Interactions for Colored Mesh Quality Assessment/a45af5e0-e0d7-4402-9367-2386d583248b_origin.pdf b/2025/HybridMQA_ Exploring Geometry-Texture Interactions for Colored Mesh Quality Assessment/a45af5e0-e0d7-4402-9367-2386d583248b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..26f31bfa0dee64c95152bf72fc68c9c5c470d345 --- /dev/null +++ b/2025/HybridMQA_ Exploring Geometry-Texture Interactions for Colored Mesh Quality Assessment/a45af5e0-e0d7-4402-9367-2386d583248b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6191e5c6156b9324bb5634c7aadc1dd21843b7a26976f16e4c52fd09d30237f3 +size 9776691 diff --git a/2025/HybridMQA_ Exploring Geometry-Texture Interactions for Colored Mesh Quality Assessment/full.md b/2025/HybridMQA_ Exploring Geometry-Texture Interactions for Colored Mesh Quality Assessment/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0d30aef7137c0f7b193c7993891ad8a9bcc719d8 --- /dev/null +++ b/2025/HybridMQA_ Exploring Geometry-Texture Interactions for Colored Mesh Quality Assessment/full.md @@ -0,0 +1,332 @@ +# HybridMQA: Exploring Geometry-Texture Interactions for Colored Mesh Quality Assessment + +Armin Shafiee Sarvestani,\* Sheyang Tang\*, Zhou Wang +Department of Electrical and Computer Engineering, University of Waterloo +{a5shafie, sheyang.tang, zhou.wang}@uwaterloo.ca + +# Abstract + +Mesh quality assessment (MQA) models play a critical role in the design, optimization, and evaluation of mesh operation systems in a wide variety of applications. Current MQA models, whether model-based methods using topology-aware features or projection-based approaches working on rendered 2D projections, often fail to capture the intricate interactions between texture and 3D geometry. We introduce HybridMQA, a first-of-its-kind hybrid full-reference colored MQA framework that integrates model-based and projection-based approaches, capturing complex interactions between textural information and 3D structures for enriched quality representations. Our method employs graph learning to extract detailed 3D representations, which are then projected to 2D using a novel feature rendering process that precisely aligns them with colored projections. This enables the exploration of geometry-texture interactions via cross-attention, producing comprehensive mesh quality representations. Extensive experiments demonstrate HybridMQA's superior performance across diverse datasets, highlighting its ability to effectively leverage geometry-texture interactions for a thorough understanding of mesh quality. Our project website is available at https://arshafiee.github.io/hybridmqa/. + +# 1. Introduction + +Advancements in 3D capture and display technologies have sparked a growing interest in immersive media. 3D meshes, a key form of 3D media, are widely used in applications like virtual and augmented reality [20, 49], gaming, animation, medical modeling [37], and generative 3D content creation [21, 35]. A mesh, comprising triangular faces formed by vertices, is colorized by assigning RGB colors to each vertex (vertex-color mesh) or applying a 2D texture map with UV coordinates (textured mesh). The demand for colored meshes calls for high-quality acquisition [3], compress- + +![](images/17d7cf6dddfddddcc61e1e1c4dd123aa3c5a4a950c08ffd82e3a1e5ddd0b7c53.jpg) +Figure 1. Interactions between texture and geometry. top: complex texture makes the geometry distortion imperceptible. bottom: modifying the geometry affects the appearance of texture distortion. Right half (gray) of each object represents geometry. + +sion [11, 24], and transmission [25]. However, these processes often introduce artifacts that degrade visual quality, highlighting the need for robust mesh quality assessment (MQA) methods. Full-reference (FR) approaches, which take distorted meshes and their pristine references as input and generate a quality score by comparing their visual quality, are essential for accurate mesh quality assessment. + +The perceived quality of a 3D colored mesh is affected by its geometry and texture. Different mesh processing operations cause diverse geometrical and texture distortions that degrade the visual quality of meshes by perturbing the interactions between the object's shape and color. Figure 1 illustrates an example of such geometry-texture interactions, where either the geometry or texture can affect the visual appearance of distortions in the other. In the top row, the easily visible geometry distortion becomes imperceptible when we replace the uniform texture with a complex one. In the bottom row, the noisy texture patterns become smooth when we modify the geometry. This highlights the need for MQA methods that capture these complex geometry-texture interaction—a significant factor largely overlooked by existing methods. + +![](images/3087de5fca1eec8884553d404e47aef7d5dccc11af4dbe186a1baafcf47920ad.jpg) +Figure 2. Reference and distorted meshes under geometry distortions. Although the distorted meshes (hawk and bowl) have distinct visual qualities (different mean opinion scores (MOS)), Graphics-LPIPS [27] assigns similar scores and reverses their ranking. HybridMQA aligns well with human perception as it understands the mesh's geometry properties. Blue boxes denote the same regions across viewpoints. + +MQA methods may be generally categorized as model-based and projection-based depending on their operation space. Model-based MQA operates directly in the 3D vertex space to extract topology-aware quality features [18, 42]. Their performance is limited due to the lack of access to the object's actual appearance (rendered projections). To compensate, they either (1) apply image quality assessment (IQA) to 2D texture maps [13, 38]; or (2) define color quality descriptors for vertices or faces [28, 48, 53]. However, both fail to capture the final appearance human viewers see. Projection-based MQA, on the other hand, operates on rendered 2D projections [19, 27, 55], hence more effective at assessing texture distortions. Nonetheless, they lack a 3D understanding of the object's topology, which is crucial for accurate quality evaluation. Although underlying geometry distortions may be less noticeable in certain projections (Fig. 1, top row), they become apparent to human viewers when the object is viewed from different angles. Humans intuitively perceive underlying structures: connecting viewpoints, incorporating 3D cues, and detecting distortions that may be imperceptible in isolated 2D projections [6, 31, 57]. This 3D understanding is hard to obtain with projections alone due to the nature of operation space. Fig. 2 illustrates this: the geometry distortion (perturbing vertex positions) is more noticeable along the contours of the hawk and bowl than within their inner regions in 2D projections. However, humans naturally identify viewpoint connections (blue boxes) as objects rotate in 3D, perceiving geometry distortions beneath the texture. Consequently, projection-based MQA metrics (Graphics-LPIPS [27]), which focus on pictorial differences, struggle to detect and penalize these distortions, resulting in wrong rankings (Fig. 2). In contrast, HybridMQA extracts quality-related features from mesh surfaces in 3D, enabling a more accurate MQA. + +In this paper, we propose a novel hybrid FR MQA method, namely HybridMQA, that integrates model- and projection-based approaches to explore geometry-texture interactions for comprehensive MQA. Specifically, a base encoder extracts initial 3D features from 2D texture, normal, and vertex maps, which initialize a feature graph based on mesh connectivity. A graph convolutional network (GCN) then learns detailed 3D surface representations, building a 3D understanding of the object's topology. These surface representations are then complemented by textural information from the mesh's renderings, enabling a hybrid integration of model- and projection-based methods across all mesh data modalities: 2D maps, 3D structure, and colored renderings. Additionally, to explore the intricate geometry-texture interactions between the two operational spaces, we propose a novel feature projection technique that renders 2D projections of surface representations from the graph, precisely aligned with colored renderings. This alignment facilitates the exploration of geometry-texture interactions via cross-attention. Below are our contributions: + +- We propose the first hybrid FR MQA method that consolidates the strengths of model- and projection-based approaches across all mesh data modalities, enabling a comprehensive understanding of mesh quality. +- We make the first attempt to explore geometry-texture interactions for MQA, drawing meaningful connections between the two domains. Our method shows the importance of leveraging such interactions to build reliable MQA methods. +- We propose a novel feature projection technique to render 2D projections of 3D surface representations from graph, aligned with the mesh's colored renderings, establishing pixel-to-pixel correspondence to explore interactions. +- Our model outperforms state-of-the-art FR MQA methods, aligning with human perception and generalizing better. The results highlight the effectiveness of capturing geometry-texture interactions to achieve accurate MQA. + +# 2. Related Works + +# 2.1. Model-based Mesh Quality Assessment + +Model-based MQA methods operate directly in 3D, with most existing methods designed for uncolored meshes [4, 5, 17, 18, 39, 41, 42], resulting in suboptimal performance for colored meshes. Early approaches, such as Hausdorff distance [4] and mean squared error (MSE) [5], use Euclidean distance as a quality measure, while Lavoué et al. [17, 18] employ curvature statistics. Other methods use local curvature and roughness pooling [39, 42] or employ dihedral angles as surface quality indicators [41]. + +However, the rise of colored meshes drives the need for color integration in MQA methods. Tian and AlRegib [38] and Guo et al. [13] apply MSE and MS-SSIM [44] to + +2D texture maps, combining these with geometric descriptors to evaluate colored meshes. Nevertheless, 2D texture maps do not represent the post-rendering appearance of 3D meshes and carry little semantic information, leading to suboptimal performance. In contrast, a second group of methods defines color quality descriptors on per-vertex or per-face color values [9, 28, 48, 53]. Nehme et al. [28] propose CMDM, a model for vertex-color meshes that extracts multi-scale color features from vertex colors. Similarly, Zhang et al. [53] apply statistical measures on vertex colors for no-reference MQA. Fu et al. [9] also sample vertex colors and propose an efficient surface sampling approach to convert meshes into point clouds for quality evaluation. Finally, Yang et al. [48] develop GeodesicPSIM, using surface colors to create textured patches for feature extraction. However, these methods fail to capture the rendered appearance of 3D objects as perceived by ultimate human viewers. + +Overall, while model-based MQA methods benefit from a profound understanding of the object's topology, their performance is constrained by the lack of access to the object's colored appearance. We address this issue by introducing a novel hybrid MQA method that complements the 3D awareness of model-based methods with textural representations derived from the colored appearance of 3D meshes. + +# 2.2. Projection-based Mesh Quality Assessment + +In projection-based MQA, quality is assessed on rendered 2D projections of 3D meshes, allowing well-established IQA methods such as PSNR [43], SSIM [45], and VIF [34] to be adapted for MQA [7, 32, 47]. However, IQA methods perform poorly in MQA as they are tailored for natural scenes, while 3D meshes involve different distortions, and their 2D projections differ statistically from natural images. + +A few projection-based methods have been proposed for meshes without color [1, 2], and they expectedly perform poorly on colored meshes. To address this, Nehme et al. [27] propose Graphics-LPIPS, the first projection-based MQA method designed for colored meshes, which builds on LPIPS [52] and uses pre-trained AlexNet [15] to extract quality features from 2D projections. Similarly, Zhang et al. [55] use Swin Transformer [22] with an efficient mini-patch sampling process for no-reference MQA. Lee et al. [19] introduce 3D-PSSIM, which uses a framework similar to Graphics-LPIPS to extract textural information, while complementing it with geometry-aware information derived from the 2D projection space. Nevertheless, these projection-based methods lack 3D understanding of the object's topology. Our approach overcomes this through a well-defined and end-to-end trainable GCN that operates on vertices in the 3D space. Furthermore, these methods fail to account for interactions between the mesh's textural information and the underlying 3D geometry, which limits their performance in detecting geometry-involved distor + +tions. We address this by proposing a cross-attention framework that draws connections between the two domains. + +# 3. Proposed Method + +Figure 3a shows the overall framework of HybridMQA, consisting of model and texture branches and a quality encoder. In the texture branch, colored projections of the input mesh (reference or distorted) are rendered from six perpendicular viewpoints. In the model branch, 2D normal, vertex, and texture maps are processed by a base encoder to extract initial 3D features, which initialize a graph based on mesh connectivity. A GCN then learns detailed 3D surface representations. At the same viewpoints as the texture branch, we render 3D feature projections from the graph, ensuring pixel-to-pixel alignment with the colored projections. Finally, the quality encoder processes both sets of projections, exploring geometry-texture interactions to produce a comprehensive quality representation of the mesh. Overall, the proposed HybridMQA can be expressed as follows: + +$$ +\boldsymbol {f} _ {\text {m e s h}} = \text {H y b r i d M Q A} (\mathcal {M} (\boldsymbol {V}, \boldsymbol {E}, \boldsymbol {U}, \boldsymbol {T} _ {t}); \boldsymbol {\theta}). \tag {1} +$$ + +Here $\mathcal{M}$ is the input mesh, $\pmb{\theta}$ represents model parameters, and $f_{mesh}$ denotes the final quality representation. The mesh $\mathcal{M}$ consists of vertices $\pmb{V}$ , vertex connectivity $\pmb{E}$ , UV coordinates $\pmb{U}$ , and the 2D texture map $\pmb{T}_t$ . Vertices and UV coordinates are defined as $\pmb{V} = \{\pmb{v}_i\}_{i=1}^N$ , $\pmb{v}_i \in \mathbb{R}^3$ and $\pmb{U} = \{\pmb{u}_i\}_{i=1}^N$ , $\pmb{u}_i \in \mathbb{R}^2$ , where $N$ is the number of vertices, and $\pmb{v}_i$ and $\pmb{u}_i$ are the positions of the $i$ -th vertex in 3D space and 2D map, respectively. + +Finally, as shown in Fig. 3b, HybridMQA generates quality representations for both reference and distorted meshes. For full-reference quality regression, the absolute differences between these representations are fed into a fully connected network (FCN) to obtain a quality score. + +# 3.1. Base Encoder + +Given an input mesh $\mathcal{M}$ , we first project its 3D geometrical attributes (normals and vertex positions) into 2D maps aligned with the texture map using barycentric interpolation and UV mapping. The shared UV coordinates $\pmb{U}$ ensure alignment between the resulting maps (normal and vertex) and the texture map $\pmb{T}_t$ . These aligned maps are concatenated and processed by a convolutional neural network (CNN) base encoder to jointly capture textural and geometrical information, learning initial quality-aware representations of the mesh's 3D shape and surface (Fig. 3a). This approach of projecting geometrical data into 2D maps has been successfully applied in mesh super-resolution [46]. + +# 3.2. Graph Convolutional Network + +Due to vertex-neighborhood discontinuities in 2D maps, we use graph learning to deepen our model's understanding + +![](images/779296ae23dab8a424c7f5466b19a516c8295accee3f6b24ed13f197dc53ff24.jpg) +(a) HybridMQA's framework. + +![](images/10c17530b52a03b726ee239cb2cf3168dd1aec8eaf71f3b1aa73683167153677.jpg) +(b) FR quality regression. +Figure 3. (a) Overview of HybridMQA. In the model branch, a base encoder extracts 3D features from the mesh's 2D maps, initializing a mesh graph. A GCN extracts 3D surface representations, which are rendered as 2D projections aligned with the colored projections from the texture branch. A quality encoder then captures geometry-texture interactions between the two branches, producing the final mesh quality representation. (b) Full-reference (FR) quality regression, where the absolute difference of HybridMQA's mesh quality representations for reference and distorted meshes is mapped to a quality score via a fully connected network (FCN). + +of the mesh's 3D structure via message-passing between neighboring vertices in 3D, enabling them to share useful quality-related information and highlight abnormalities. + +Graph initialization. Given the base encoder's output feature map with $C_1$ channels, we sample a $C_1$ -dimensional feature vector for each mesh vertex using UV mapping: + +$$ +\boldsymbol {f} _ {i} = \boldsymbol {T} _ {B E} \left[ u _ {i} ^ {h}, u _ {i} ^ {w} \right], i \in \{1, \dots , N \}. \tag {2} +$$ + +Here, $\pmb{f}_i \in \mathbb{R}^{C_1}$ is the sampled feature for $i$ -th vertex, $u_i^h$ and $u_i^w$ are its UV coordinates, and $T_{BE}$ denotes base encoder's output feature maps. For vertices with multiple UV mappings, we average the sampled features. This process produces a feature graph with vertices initialized by sampled features and edges defined by mesh connectivity $\pmb{E}$ . + +Graph update. We apply a GCN to integrate features from neighboring vertices to learn surface properties. As proposed by Morris et al. [26], we update vertex features by: + +$$ +\boldsymbol {f} _ {i} ^ {\prime} = \boldsymbol {\theta} _ {1} \boldsymbol {f} _ {i} + \boldsymbol {\theta} _ {2} \sum_ {j \in \Gamma (i)} e _ {j, i} \boldsymbol {f} _ {j}. \tag {3} +$$ + +Here, $f_{i}^{\prime} \in \mathbb{R}^{C_{2}}$ is the updated feature vector for the $i$ -th vertex, $\Gamma(i)$ is its 1-ring neighborhood, $\theta_{1}, \theta_{2} \in \mathbb{R}^{C_{2} \times C_{1}}$ are learnable weights, and $e_{j,i}$ is the edge weight between vertices $i$ and $j$ , defined as the inverse of their Euclidean distance. Through training, the GCN refines vertex feature embeddings to learn quality-aware 3D surface representations, which are later complemented by textural embeddings to deliver a hybrid and comprehensive quality assessment. This approach of graph learning has been effectively applied to mesh texture downsampling [29]. + +# 3.3. Feature & RGB Rendering + +We render multiple viewpoints of the 3D surface representations into 2D projections aligned with the mesh's colored projections. This novel feature graph rendering enables us to capture geometry-texture interactions between the two branches for a comprehensive mesh quality representation. + +In the texture branch (Fig. 3a), colored projections are rendered after normalizing mesh vertex positions to fit within a unit cube. This standardization allows fixed camera positions for any mesh. Six virtual cameras are placed on the cube's surfaces, facing the object to cover all angles. With PyTorch3D's [30] soft Phong Shader, we render six perpendicular colored projections, employing directional or ambient light to match the conditions of subjective tests. + +The model branch employs the same camera setup to ensure alignment with colored projections. Our novel feature graph rendering customizes PyTorch3D's differentiable renderer to render six perpendicular $3D$ feature projections from the graph, enabling gradient backpropagation to the GCN and base encoder. To focus on raw vertex features interpolated on the mesh surface, we remove shadows and specular effects, setting diffuse and specular reflectivity to zero and using ambient light. Thus we define the mesh as a vertex-color mesh with $C$ -dimensional vertex features (instead of RGB colors) to render the 3D feature projections with hard Phong shader [30]. + +# 3.4. Quality Encoder + +We design a quality encoder that processes the aligned projections from model and texture branches to output an expressive quality representation for the mesh. It captures the geometry-texture interactions between the two sets of pro + +![](images/fcfdc02f92da17a7a799b09757e5fbbd95a9a22a714c8adf55ed525bd18acc88.jpg) +(a) Quality encoder. + +![](images/a073be4013f89a4fbe66daa6928f82fe1e3e8f3a63bf7b525c4f011c13425f86.jpg) +(b) Cross-attention module. +Figure 4. (a) The quality encoder. The 3D feature and color projections are divided into valid aligned patches and fed into respective encoders to obtain multiscale 3D surface and color representations. The cross-attention modules capture interactions between these representations, which are then concatenated with 3D feature embeddings directly extracted from the patches to form the final mesh quality representation. (b) The cross-attention module consists of two transformer blocks, where we switch the roles of the two inputs. + +jections through multi-scale cross-attentions. As discussed in Sec. 1, understanding these interactions is key to a robust and effective MQA. As detailed in Fig. 4a, given model branch's 3D feature projections $\mathbf{I}_m$ and texture branch's colored projections $\mathbf{I}_t$ , the quality encoder $\psi$ outputs + +$$ +\boldsymbol {f} _ {\text {m e s h}} = \psi \left(\boldsymbol {I} _ {m}, \boldsymbol {I} _ {t}; \boldsymbol {\theta} _ {\psi}\right), \tag {4} +$$ + +where $f_{mesh}$ denotes the final quality representation, and $\theta_{\psi}$ denotes the learnable parameters. + +Patch Selection. Given two sets of projections, we extract non-overlapping patches and discard those with less than $10\%$ non-background pixels. This yields $2P$ patches: + +$$ +\tilde {\boldsymbol {I}} _ {m} = \left\{\boldsymbol {I} _ {m, p} \right\} _ {p = 1} ^ {P}, \tilde {\boldsymbol {I}} _ {t} = \left\{\boldsymbol {I} _ {t, p} \right\} _ {p = 1} ^ {P}. \tag {5} +$$ + +Here, $\tilde{I}_m$ and $\tilde{I}_t$ denote aligned patches extracted from 3D feature and colored projections, respectively. + +3D Feature & Image Encoding. Next, we feed pairs of aligned patches $I_{m,p}$ and $I_{t,p}$ to a 3D feature encoder $\phi$ and an image encoder $\varphi$ , respectively, to extract multiscale quality-aware representations $F_{m,p}$ and $F_{t,p}$ : + +$$ +\boldsymbol {F} _ {m, p} = \left\{\boldsymbol {f} _ {m, p} ^ {s} | s = 1, \dots , 5 \right\} = \phi (\boldsymbol {I} _ {m, p}; \boldsymbol {\theta} _ {\phi}), \tag {6} +$$ + +$$ +\boldsymbol {F} _ {t, p} = \left\{\boldsymbol {f} _ {t, p} ^ {s} | s = 1, \dots , 5 \right\} = \varphi (\boldsymbol {I} _ {t, p}; \boldsymbol {\theta} _ {\varphi}). +$$ + +Here, $\theta_{\phi}$ and $\theta_{\varphi}$ are learnable parameters of the two encoders, and $s$ denotes scale index. $1\times 1$ convolutions are used to adjust the channel dimension of representations. + +Cross-attention Modules. We apply cross-attention to capture the interactions between aligned 3D surface representations $\pmb{f}_{m,p}^{s}$ and color representations $\pmb{f}_{t,p}^{s}$ at each scale. Figure 4b depicts our proposed module with two transformer + +blocks [40], where we alternate the query and key-value roles of $f_{m,p}^{s}$ and $f_{t,p}^{s}$ to explore their interactions and mutual influence. This simulates how one domain's representation affects the impact of the other domain's representation on perceptual quality, as discussed in Sec. 1. We concatenate outputs to form single-scale quality representation $f_{p}^{s}$ . + +Mesh Quality Representation. To further exploit the model branch's understanding of mesh's 3D structure and surface properties, we directly extract representation $\hat{f}_p$ from model branch patches via a convolutional block. $\hat{f}_p$ is concatenated with the five single-scale representations from cross-attention, and averaged over all patches to obtain the final mesh quality representation $f_{mesh}$ : + +$$ +\boldsymbol {f} _ {\text {m e s h}} = \frac {1}{P} \sum_ {p = 1} ^ {P} \left(\boldsymbol {f} _ {p} ^ {1} \oplus \dots \oplus \boldsymbol {f} _ {p} ^ {5} \oplus \hat {\boldsymbol {f}} _ {p}\right). \tag {7} +$$ + +# 3.5. Quality Regression & Optimization + +As shown in Fig. 3b, we feed the reference and distorted meshes into HybridMQA separately and map the absolute difference of their quality representations $(|f_{mesh}^{ref} - f_{mesh}^{dis}|)$ to a quality score via an FCN for full-reference MQA. The model is optimized using a loss function with two terms: mean absolute error (MAE) and rank loss. While MAE ensures accurate quality predictions, rank loss helps differentiate closely rated samples within a mini-batch [36]. + +MAE Loss. The MAE loss is defined as + +$$ +L _ {m a e} = \frac {1}{B} \sum_ {i = 1} ^ {B} \left| q _ {i} - q _ {i} ^ {\prime} \right|, \tag {8} +$$ + +where $q_{i}$ and $q_{i}^{\prime}$ denote the predicted and ground truth quality scores of the $i$ -th sample in the batch, with $B$ being the batch size. The ground truth is the mean opinion score (MOS) obtained in subjective tests and normalized to [0, 1]. + +Rank Loss. Since ranking is not differentiable, as proposed by Sun et al. [36], we approximate the rank value as + +$$ +\begin{array}{l} L _ {r a n k} ^ {i, j} = \max \left(0, \left| q _ {i} ^ {\prime} - q _ {j} ^ {\prime} \right| - e \left(q _ {i} ^ {\prime}, q _ {j} ^ {\prime}\right) \cdot \left(q _ {i} - q _ {j}\right)\right), \\ e \left(q _ {i} ^ {\prime}, q _ {j} ^ {\prime}\right) = \left\{ \begin{array}{l l} 1 & q _ {i} ^ {\prime} \geq q _ {j} ^ {\prime} \\ - 1 & q _ {i} ^ {\prime} < q _ {j} ^ {\prime} \end{array} , \right. \tag {9} \\ \end{array} +$$ + +where $i$ and $j$ denote two samples in a batch, and $e(q_i', q_j')$ is a sign function. The final rank loss is computed as + +$$ +L _ {r a n k} = \frac {1}{B ^ {2} - B} \sum_ {i = 1} ^ {B} \sum_ {\substack {j = 1 \\ j \neq i}} ^ {B} L _ {r a n k} ^ {i, j}. \tag{10} +$$ + +Final Loss. The final loss is a weighted sum of the two: + +$$ +L = L _ {m a e} + \lambda L _ {r a n k}, \tag {11} +$$ + +where $\lambda$ balances the effect of the two loss terms. + +# 4. Experiments + +# 4.1. Datasets & Implementation Details + +We validate our method on four publicly available color MQA datasets: Nehme et al. [27], SJTU-TMQA [7], TSMD [47], and CMDM [28] datasets. Nehme et al., the largest available, includes textured meshes with mixed geometric and texture distortions, as do SJTU-TMQA and TSMD. In contrast, CMDM consists of vertex-color meshes with either geometric or color distortions. All datasets use MOS as ground truth. + +We use ResNet50 [14] pre-trained on ImageNet [8] as the image encoder $\varphi$ and a randomly initialized CNN as the 3D feature encoder $\phi$ . 3D feature and colored projections are rendered at $128 \times 128$ and $512 \times 512$ resolutions, with patch sizes of $16 \times 16$ and $64 \times 64$ , respectively, to ensure alignment. For data augmentation, we randomly perturb camera angles and flip the patches in training to improve robustness and generalization. More details of the datasets and implementation are provided in the supplementary material. + +# 4.2. Experimental Setup + +We use 5-fold cross-validation without overlap between train and test source content and report the median performance across five experiments. This strategy is applied to all learning-based methods for fair comparison. + +We compare HybridMQA with 11 model-based and projection-based full-reference MQA methods. Model-based methods include Hausdorff Distance (HD) [4], MSDM2 [18], FMPD [42], GeodesicPSIM [48], and Fu + +et al. [9]. Projection-based methods include PSNR [43], SSIM [45], VIF [34], LPIPS [52], Graphics-LPIPS [27], and 3D-PSSIM [19]. For a fair comparison, all projection-based methods are evaluated under the same rendering settings as HybridMQA, with no prior assumptions about object orientation. We adopt the Spearman rank-order correlation coefficient (SRCC) and the Pearson linear correlation coefficient (PLCC) to compare all methods. Higher values indicate better performance and a stronger correlation between MOS and predicted quality scores [12]. Further information is provided in the supplementary material. + +# 4.3.Quantitative Results + +Overall Performance. Table 1 summarizes the performance comparison, from which we make the following observations: (1) HybridMQA outperforms all model-based and projection-based comparison methods across all datasets, including both textured and vertex-color meshes, demonstrating its effectiveness in the quality assessment of colored meshes. Notably, HybridMQA achieves $6.5\%$ and $7.7\%$ performance gain in PLCC over the second-best method, 3D-PSSIM, on Nehme et al. and SJTU-TMQA, respectively; (2) In general, projection-based methods outperform model-based methods which lack access to the object's actual appearance. HybridMQA consolidates the advantages of both types by complementing the textural information extracted from colored projections with 3D representations learned in 3D; (3) Unlike CMDM with single-type distortions, Nehme et al., SJTU-TMQA, and TSMD involve mixed-type and hence more complex distortions. Consequently, while most methods perform well on CMDM, their performance does not extend to the other three datasets. In contrast, HybridMQA consistently performs well on all datasets, showing robustness in assessing complex distortions. This is achieved by HybridMQA's understanding of geometry-texture interactions, which are perturbed uniquely by different distortions. + +Performance by Distortion Type. In Fig. 5, we compare HybridMQA with Graphics-LPIPS based on distortion types in SJTU-TMQA. In addition to the overall superiority of HybridMQA, we observe that while the two methods achieve comparable performance on texture-only distortions (e.g. JPEG [7]), HybridMQA hugely outperforms Graphics-LPIPS on distortions that only affect geometry (e.g. gn [7] and simpNoTex [7, 10]) or both geometry and texture (e.g. qpqtJPEG [7, 11]). This further demonstrates HybridMQA's proficiency in understanding meshes' 3D geometry and its interactions with textural appearance. + +Generalizability. We train models on Nehme et al. and TSMD and test them on SJTU-TMQA. Table 2 shows that HybridMQA significantly outperforms other learning-based methods, showing strong generalizability. Notably, when trained on TSMD and tested on SJTU-TMQA, HybridMQA + +
TypeMethodNehmé et al. [27]SJTU-TMQA [7]TSMD [47]CMDM [28]
SRCCPLCCSRCCPLCCSRCCPLCCSRCCPLCC
Model-basedHD [4]0.1070.1750.0600.1400.4460.4620.1890.210
MSDM2 [18]0.3350.3440.0500.1200.0450.2550.4150.517
FMPD [42]0.3910.4040.1560.4580.0770.2180.6150.623
GeodesicPSIM [48]----0.8200.820--
Fu et al. [9]0.6880.696------
Projection-basedPSNR [43]0.3530.3750.2990.2870.7140.7110.8300.839
SSIM [45]0.2100.2260.3940.2890.6730.6740.8520.861
VIF [34]0.5380.5570.4500.4220.8510.8460.8270.837
LPIPS [52]0.6720.6760.7180.7170.7100.7120.8650.918
Graphics-LPIPS [27]0.7220.7460.7900.7620.8340.8120.8590.925
3D-PSSIM [19]0.8820.8420.8420.832--0.8550.854
HybridHybridMQA0.8920.8970.8870.8960.9120.9190.8970.927
+ +Table 1. SRCC and PLCC scores of MQA methods on four color MQA benchmark datasets. The scores of GeodesicPSIM [48], Fu et al. [9], and 3D-PSSIM [19] are reported directly from their publications as their implementations are not publicly available. Bold and underlined values denote the best and second-best results per column, respectively. + +
Trained onNehmé et al.TSMD
SRCCPLCCSRCCPLCC
LPIPS0.5920.5840.7120.695
Graphics-LPIPS0.7310.7340.7840.773
HybridMQA0.8000.7830.8460.811
+ +Table 2. Generalization evaluation, where models are trained on Nehme et al. and TSMD and tested on SJTU-TMQA. + +![](images/f36b5af9e603896d80a9603ea60f5870bf8aaea382cecd7f4dddc146bfc82b11.jpg) +Figure 5. SRCC/PLCC performance of HybridMQA and Graphics-LPIPS on various distortion types of SJTU-TMQA dataset. Green distortions affect the geometry or both geometry and texture, while others only impact the texture. + +achieves comparable performance to 3D-PSSIM [19], despite 3D-PSSIM being trained directly on SJTU-TMQA. + +# 4.4. Qualitative Results + +We apply GradCAM [33] on 3D feature projections to verify that the model branch effectively captures geometry-aware quality representations. As shown in Fig. 6, the highlighted regions with noticeable artifacts align well with hu + +![](images/3f4e7f486e3eb16442125b19ad9b3b3ed0bfaa046c773bf97f24baa69bd386c7.jpg) +Figure 6. GradCAM results on meshes in the model branch. Highlighted regions exhibit more noticeable artifacts, aligning with human perception and showing the model's effectiveness in capturing geometry-aware quality representations. + +man perception, validating the model branch's ability to extract geometry-aware representations. Additionally, by mapping these results onto the mesh topology via differentiable rendering, our method opens up opportunities for optimizing mesh compression or restoration algorithms. More examples are available in the supplementary material. + +Figure 7 shows the effectiveness of our hybrid model in exploiting interactions between representations learned in texture and model branches. GradCAM is applied before and after cross-attention, showing that the two branches focus on different regions, with the model branch effectively highlighting geometric artifacts. Cross-attention then successfully identifies and attends to perceptually important regions by exploring geometry-texture interactions. Note that the actual cross-attention inputs are patches described + +![](images/44b3247055d76cfe2d14219205c8bfcdf14c4320f5d6d360106ce74a28f08d87.jpg) +Figure 7. GradCAM before and after cross-attention. The two branches focus on different regions, and cross-attention effectively integrates them by attending to perceptually important regions. + +in Sec. 3.3 and 3.4, and the heatmaps are the union of the GradCAM results from the reference and distorted patches. More examples are available in the supplementary material. + +We also perform gMAD competition [23] to qualitatively compare HybridMQA with Graphics-LPIPS. The results are available in the supplementary material. + +# 4.5. Ablation Studies + +We conduct ablation studies on Nehme et al. [27] to evaluate key components and design choices in our model. Further results are available in the supplementary material. + +3D Surface and Textural Representations. To show the contributions of 3D surface and textural representations, we test four configurations: (1) 3D-only, where only 3D representations $\hat{f}$ and $f_{m}$ are used for mesh quality representation; (2) RGB-only, where only textural representation $f_{t}$ from colored renderings are used; (3) all but excluding $f_{m}$ ; and (4) excluding $\hat{f}$ . Table 3 shows that all representations contribute to the final performance, with textural representation $f_{t}$ the most significant. Moreover, excluding 3D representations (RGB-only) leads to a significant performance drop, showing the importance of 3D representations and the need to explore their interaction with textural information for accurate MQA. The results also highlight the performance gain of our hybrid approach, as it outperforms both + +
NotesffmftSRCCPLCC
3D-only✓✓-0.5860.595
RGB-only--✓0.8200.846
-✓-✓0.8420.849
--✓✓0.8700.874
Proposed✓✓✓0.8920.897
+ +Table 3. Ablation on 3D Surface & Text. Repr. on Nehme et al. + +
ConfigurationsSRCCPLCC
w/o Texture map0.8660.872
w/o Normal map0.8630.866
w/o Vertex map0.8720.873
w/o Base Encoder0.8560.851
w/o GCN0.8660.865
HybridMQA (proposed)0.8920.897
+ +Table 4. Ablation on model branch components on Nehme et al. + +3D-only and RGB-only counterparts. + +Model Branch Components. To validate the contributions of model branch components, we test the following configurations: all components but excluding (1) the texture map; (2) the normal map; (3) the vertex map; (4) the Base Encoder; and (5) the GCN. Table 4 shows that all components contribute to learning effective 3D surface representations. Furthermore, the proposed processing units (Base Encoder and GCN) effectively leverage mesh data modalities (2D maps and 3D vertex space) to enhance performance. + +# 5. Conclusion & Discussion + +We present a novel hybrid full-reference MQA method that integrates model- and projection-based approaches for enhanced quality assessment. Our model explores interactions between mesh texture and 3D geometry via cross-attention, enabled by a novel feature rendering process that aligns 3D representations with colored projections. Extensive experiments show the effectiveness and superiority of our method, highlighting the importance of 3D understanding and leveraging geometry-texture interactions for reliable MQA. + +A few limitations present room for improvement. Our method relies on perpendicular viewpoints without considering their varying contributions to perceptual quality. Integrating a semantic-aware module could address this by weighting viewpoints based on importance. Also, memory consumption scales with mesh size as the GCN processes the entire mesh graph, limiting efficiency in real-world applications. Sampling techniques [16, 50, 51] could help reduce memory footprint. Future directions include (1) adapting our work for no-reference MQA; (2) generalizing it to point cloud quality assessment [54, 56] by defining graph edges; and (3) exploring applications in perceptually optimized mesh compression, enhancement, and generation. + +# References + +[1] Ilyass Aboulaziz, Aladine Chetouani, Mohammed El Hassouni, and Hocine Cherifi. A blind mesh visual quality assessment method based on convolutional neural network. *Electronic Imaging*, 30:1-5, 2018. 3 +[2] Ilyass Abouelaziz, Aladine Chetouani, Mohammed El Hassouni, Longin Jan Latecki, and Hocine Cherifi. No-reference mesh visual quality assessment via ensemble of convolutional neural networks and compact multi-linear pooling. Pattern Recognition, 100:107174, 2020. 3 +[3] Martin Alain, Emin Zerman, Cagri Ozcinar, and Giuseppe Valenzise. Chapter 1 - introduction to immersive video technologies. In Immersive Video Technologies, pages 3-24. Academic Press, 2023. 1 +[4] Nicolas Aspert, Diego Santa-Cruz, and Touradj Ebrahimi. Mesh: measuring errors between surfaces using the hausdorff distance. In Proceedings. IEEE International Conference on Multimedia and Expo, pages 705-708 vol.1, 2002. 2, 6, 7 +[5] Paolo Cignoni, Claudio Rocchini, and Roberto Scopigno. Metro: Measuring error on simplified surfaces. Computer Graphics Forum, 17(2):167-174, 1998. 2 +[6] Sahin Coskun, Gokce Nur Yilmaz, Federica Battisti, Musaed Alhussein, and Saiful Islam. Measuring 3d video quality of experience (qoe) using a hybrid metric based on spatial resolution and depth cues. Journal of Imaging, 9(12):281, 2023. 2 +[7] Bingyang Cui, Qi Yang, Kaifa Yang, Yiling Xu, Xiaozhong Xu, and Shan Liu. Sjtu-tmqa: A quality assessment database for static mesh with texture map. In ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 7875-7879, 2024. 3, 6, 7 +[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 6 +[9] Chunyang Fu, Xiang Zhang, Thuong Nguyen-Canh, Xiaozhong Xu, Ge Li, and Shan Liu. Surface-sampling based objective quality assessment metrics for meshes. In ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5, 2023. 3, 6, 7 +[10] Michael Garland and Paul S. Heckbert. Surface simplification using quadric error metrics. In Proceedings of the 24th Annual Conference on Computer Graphics and Interactive Techniques, page 209-216, USA, 1997. ACM Press/Addison-Wesley Publishing Co. 6 +[11] Google. Draco 3d data compression. https://google.github.io/draco/. Accessed: 2024-09-04. 1,6 +[12] Video Quality Experts Group et al. Final report from the video quality experts group on the validation of objective models of video quality assessment, phase II. VQEG, 2003. 6 +[13] Jinjiang Guo, Vincent Vidal, Irene Cheng, Anup Basu, Atilla Baskurt, and Guillaume Lavoue. Subjective and objective + +visual quality assessment of textured 3d meshes. ACM Trans. Appl. Percept., 14(2), 2016. 2 +[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 6 +[15] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2012. 3 +[16] Itai Lang, Asaf Manor, and Shai Avidan. Samplenet: Differentiable point cloud sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7578-7588, 2020. 8 +[17] Guillaume Lavoué, Elisa Drelie Gelasca, Florent Dupont, Atilla Baskurt, and Touradj Ebrahimi. Perceptually driven 3d distance metrics with application to watermarking. Applications of Digital Image Processing XXIX, 6312:63120L, 2006. 2 +[18] Guillaume Lavoué. A multiscale metric for 3d mesh visual quality assessment. Computer Graphics Forum, 30(5):1427-1437, 2011. 2, 6, 7 +[19] Seongmin Lee, Jiwoo Kang, Sanghoon Lee, Weisi Lin, and Alan Conrad Bovik. 3d-pssim: Projective structural similarity for 3d mesh quality assessment robust to topological irregularities. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-16, 2024. 2, 3, 6, 7 +[20] Jie Li and Pablo Cesar. Chapter 22 - social virtual reality (vr) applications and user experiences. In Immersive Video Technologies, pages 609-648. Academic Press, 2023. 1 +[21] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Mukund Varma T, Zexiang Xu, and Hao Su. One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. Advances in Neural Information Processing Systems, 36, 2024. 1 +[22] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10012-10022, 2021. 3 +[23] Kede Ma, Zhengfang Duanmu, Zhou Wang, Qingbo Wu, Wentao Liu, Hongwei Yong, Hongliang Li, and Lei Zhang. Group maximum differentiation competition: Model comparison with few samples. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(4):851-864, 2020. 8 +[24] Adrien Maglo, Guillaume Lavoué, Florent Dupont, and Céline Hudelot. 3d mesh compression: Survey, comparisons, and emerging trends. ACM Comput. Surv., 47(3), 2015. 1 +[25] Adrien Maglo, Guillaume Lavoué, Florent Dupont, and Céline Hudelot. 3d mesh compression: Survey, comparisons, and emerging trends. ACM Comput. Surv., 47(3), 2015. 1 +[26] Christopher Morris, Martin Ritzert, Matthias Fey, William L. Hamilton, Jan Eric Lenssen, Gaurav Rattan, and Martin Grohe. Weisfeiler and leman go neural: Higher-order graph neural networks. Proceedings of the AAAI Conference on Artificial Intelligence, 33(01):4602-4609, 2019. 4 +[27] Yana Nehme, Johanna Delanoy, Florent Dupont, JeanPhilippe Farrugia, Patrick Le Callet, and Guillaume Lavoué. + +Textured mesh quality assessment: Large-scale dataset and deep learning-based quality metric. ACM Trans. Graph., 42 (3), 2023. 2, 3, 6, 7, 8 +[28] Yana Nehme, Florent Dupont, Jean-Philippe Farrugia, Patrick Le Callet, and Guillaume Lavoué. Visual quality of 3d meshes with diffuse colors in virtual reality: Subjective and objective evaluation. IEEE Transactions on Visualization and Computer Graphics, 27(3):2202-2219, 2021. 2, 3, 6, 7 +[29] Sai Karthikey Pentapati, Anshul Rai, Arkady Ten, Chaitanya Atluru, and Alan Bovik. Geoscaler: Geometry and rendering-aware downsampling of 3d mesh textures. arXiv preprint arXiv:2311.16581, 2023. 4 +[30] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3d deep learning with pytorch3d. arXiv:2007.08501, 2020. 4 +[31] Stephan Reichelt, Ralf Häussler, Gerald Füfterer, and Norbert Leister. Depth cues in human visual perception and their realization in 3d displays. In Three-Dimensional Imaging, Visualization, and Display 2010 and Display Technologies and Applications for Defense, Security, and Avionics IV, pages 92–103. SpIE, 2010. 2 +[32] Armin Shafiee Sarvestani, Wei Zhou, and Zhou Wang. Perceptual crack detection for rendered 3d textured meshes. In 2024 16th International Conference on Quality of Multimedia Experience (QoMEX), pages 1-7, 2024. 3 +[33] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 7 +[34] Hamid R. Sheikh and Alan C. Bovik. Image information and visual quality. IEEE Transactions on Image Processing, 15 (2):430-444, 2006. 3, 6, 7 +[35] Yawar Siddiqui, Tom Monnier, Filippos Kokkinos, Mahendra Kariya, Yanir Kleiman, Emilien Garreau, Oran Gafni, Natalia Neverova, Andrea Vedaldi, Roman Shapovalov, et al. Meta 3d assetgen: Text-to-mesh generation with high-quality geometry, texture, and pbr materials. arXiv preprint arXiv:2407.02445, 2024. 1 +[36] Wei Sun, Xiongkuo Min, Wei Lu, and Guangtao Zhai. A deep learning based no-reference quality assessment model forUGC videos. In Proceedings of the 30th ACM International Conference on Multimedia, page 856-865, New York, NY, USA, 2022. Association for Computing Machinery. 5, 6 +[37] Yuk Ming Tang and Ho Lun Ho. 3d modeling and computer graphics in virtual reality. In mixed reality and three-dimensional computer graphics. IntechOpen, 2020. 1 +[38] Dihong Tian and Ghassan AlRegib. Batex3: Bit allocation for progressive transmission of textured 3-d models. IEEE Transactions on Circuits and Systems for Video Technology, 18(1):23-35, 2008. 2 +[39] Fakhri Torkhani, Kai Wang, and Jean-Marc Chassery. A curvature tensor distance for mesh visual quality assessment. In International Conference on Computer Vision and Graphics, pages 253–263. Springer, 2012. 2 + +[40] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 5 +[41] Libor Váša and Jan Rus. Dihedral angle mesh error: a fast perception correlated distortion measure for fixed connectivity triangle meshes. Computer Graphics Forum, 31(5):1715-1724, 2012. 2 +[42] Kai Wang, Fakhri Torkhani, and Annick Montanvert. A fast roughness-based approach to the assessment of 3d mesh visual quality. Computers & Graphics, 36(7):808-818, 2012. Augmented Reality Computer Graphics in China. 2, 6, 7 +[43] Zhou Wang and Alan C. Bovik. Mean squared error: Love it or leave it? a new look at signal fidelity measures. IEEE Signal Processing Magazine, 26(1):98-117, 2009. 3, 6, 7 +[44] Zhou Wang, Eero P Simoncelli, and Alan C Bovik. Multiscale structural similarity for image quality assessment. In The Thirty-Seventh Asilomar Conference on Signals, Systems & Computers, 2003, pages 1398-1402 Vol.2, 2003. 2 +[45] Zhou Wang, Alan C. Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 3, 6, 7 +[46] Wuyuan Xie, Tengcong Huang, and Miaohui Wang. Mnsnet: multimodal transformer network for 3d surface superresolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12703-12712, 2022. 3 +[47] Qi Yang, Joel Jung, Haiqiang Wang, Xiaozhong Xu, and Shan Liu. Tsmd: A database for static color mesh quality assessment study. In 2023 IEEE International Conference on Visual Communications and Image Processing (VCIP), pages 1-5, 2023. 3, 6, 7 +[48] Qi Yang, Joel Jung, Xiaozhong Xu, and Shan Liu. Geodesicpsim: Predicting the quality of static mesh with texture map via geodesic patch similarity. arXiv preprint arXiv:2308.04928, 2023. 2, 3, 6, 7 +[49] Gareth W. Young, Néill O'Dwyer, and Aljosa Smolic. Chapter 21 - volumetric video as a novel medium for creative storytelling. In Immersive Video Technologies, pages 591-607. Academic Press, 2023. 1 +[50] Hanqing Zeng, Hongkuan Zhou, Ajitesh Srivastava, Rajgopal Kannan, and Viktor Prasanna. GraphSAINT: Graph sampling based inductive learning method. In International Conference on Learning Representations, 2020. 8 +[51] Hanqing Zeng, Muhan Zhang, Yinglong Xia, Ajitesh Srivastava, Andrey Malevich, Rajgopal Kannan, Viktor Prasanna, Long Jin, and Ren Chen. Decoupling the depth and scope of graph neural networks. Advances in Neural Information Processing Systems, 34:19665-19679, 2021. 8 +[52] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3, 6, 7 +[53] Zicheng Zhang, Wei Sun, Xiongkuo Min, Tao Wang, Wei Lu, and Guangtao Zhai. No-reference quality assessment for + +3d colored point cloud and mesh models. IEEE Transactions on Circuits and Systems for Video Technology, 32(11):7618-7631, 2022. 2, 3 +[54] Zicheng Zhang, Wei Sun, Xiongkuo Min, Qiyuan Wang, Jun He, Quan Zhou, and Guangtao Zhai. Mm-pcqa: Multi-modal learning for no-reference point cloud quality assessment. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, IJCAI-23, pages 1759–1767. International Joint Conferences on Artificial Intelligence Organization, 2023. Main Track. 8 +[55] Zicheng Zhang, Wei Sun, Haoning Wu, Yingjie Zhou, Chunyi Li, Zijian Chen, Xiongkuo Min, Guangtao Zhai, and Weisi Lin. Gms-3dqa: Projection-based grid mini-patch sampling for 3d model quality assessment. ACM Trans. Multimedia Comput. Commun. Appl., 20(6), 2024. 2, 3 +[56] Zicheng Zhang, Haoning Wu, Yingjie Zhou, Chunyi Li, Wei Sun, Chaofeng Chen, Xiongkuo Min, Xiaohong Liu, Weisi Lin, and Guangtao Zhai. Lmm-pcqa: Assisting point cloud quality assessment with lmm. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 7783-7792, 2024. 8 +[57] Wei Zhou, Qi Yang, Wu Chen, Qiuping Jiang, Guangtao Zhai, and Weisi Lin. Blind quality assessment of dense 3d point clouds with structure guided resampling. ACM Transactions on Multimedia Computing, Communications and Applications, 2024. 2 \ No newline at end of file diff --git a/2025/HybridMQA_ Exploring Geometry-Texture Interactions for Colored Mesh Quality Assessment/images.zip b/2025/HybridMQA_ Exploring Geometry-Texture Interactions for Colored Mesh Quality Assessment/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..99260af0de568c85e24fe846ed71b7df0deb2786 --- /dev/null +++ b/2025/HybridMQA_ Exploring Geometry-Texture Interactions for Colored Mesh Quality Assessment/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:943235980458295325f009d7de83baa81a0e4ef76a7208205e627656d75c6d8e +size 612627 diff --git a/2025/HybridMQA_ Exploring Geometry-Texture Interactions for Colored Mesh Quality Assessment/layout.json b/2025/HybridMQA_ Exploring Geometry-Texture Interactions for Colored Mesh Quality Assessment/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e3f4e53482849efecaf353530e47a3ab61c72201 --- /dev/null +++ b/2025/HybridMQA_ Exploring Geometry-Texture Interactions for Colored Mesh Quality Assessment/layout.json @@ -0,0 +1,8769 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 77, + 102, + 534, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 102, + 534, + 140 + ], + "spans": [ + { + "bbox": [ + 77, + 102, + 534, + 140 + ], + "type": "text", + "content": "HybridMQA: Exploring Geometry-Texture Interactions for Colored Mesh Quality Assessment" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 122, + 161, + 490, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 161, + 490, + 203 + ], + "spans": [ + { + "bbox": [ + 122, + 161, + 490, + 203 + ], + "type": "text", + "content": "Armin Shafiee Sarvestani,\\* Sheyang Tang\\*, Zhou Wang \nDepartment of Electrical and Computer Engineering, University of Waterloo \n{a5shafie, sheyang.tang, zhou.wang}@uwaterloo.ca" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 152, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 152, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 255, + 296, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 296, + 533 + ], + "type": "text", + "content": "Mesh quality assessment (MQA) models play a critical role in the design, optimization, and evaluation of mesh operation systems in a wide variety of applications. Current MQA models, whether model-based methods using topology-aware features or projection-based approaches working on rendered 2D projections, often fail to capture the intricate interactions between texture and 3D geometry. We introduce HybridMQA, a first-of-its-kind hybrid full-reference colored MQA framework that integrates model-based and projection-based approaches, capturing complex interactions between textural information and 3D structures for enriched quality representations. Our method employs graph learning to extract detailed 3D representations, which are then projected to 2D using a novel feature rendering process that precisely aligns them with colored projections. This enables the exploration of geometry-texture interactions via cross-attention, producing comprehensive mesh quality representations. Extensive experiments demonstrate HybridMQA's superior performance across diverse datasets, highlighting its ability to effectively leverage geometry-texture interactions for a thorough understanding of mesh quality. Our project website is available at https://arshafiee.github.io/hybridmqa/." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 555, + 135, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 555, + 135, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 555, + 135, + 567 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 574, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 574, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 295, + 696 + ], + "type": "text", + "content": "Advancements in 3D capture and display technologies have sparked a growing interest in immersive media. 3D meshes, a key form of 3D media, are widely used in applications like virtual and augmented reality [20, 49], gaming, animation, medical modeling [37], and generative 3D content creation [21, 35]. A mesh, comprising triangular faces formed by vertices, is colorized by assigning RGB colors to each vertex (vertex-color mesh) or applying a 2D texture map with UV coordinates (textured mesh). The demand for colored meshes calls for high-quality acquisition [3], compress-" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 319, + 232, + 552, + 369 + ], + "blocks": [ + { + "bbox": [ + 319, + 232, + 552, + 369 + ], + "lines": [ + { + "bbox": [ + 319, + 232, + 552, + 369 + ], + "spans": [ + { + "bbox": [ + 319, + 232, + 552, + 369 + ], + "type": "image", + "image_path": "17d7cf6dddfddddcc61e1e1c4dd123aa3c5a4a950c08ffd82e3a1e5ddd0b7c53.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 376, + 555, + 422 + ], + "lines": [ + { + "bbox": [ + 313, + 376, + 555, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 555, + 422 + ], + "type": "text", + "content": "Figure 1. Interactions between texture and geometry. top: complex texture makes the geometry distortion imperceptible. bottom: modifying the geometry affects the appearance of texture distortion. Right half (gray) of each object represents geometry." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 447, + 555, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 555, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 555, + 532 + ], + "type": "text", + "content": "sion [11, 24], and transmission [25]. However, these processes often introduce artifacts that degrade visual quality, highlighting the need for robust mesh quality assessment (MQA) methods. Full-reference (FR) approaches, which take distorted meshes and their pristine references as input and generate a quality score by comparing their visual quality, are essential for accurate mesh quality assessment." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": "The perceived quality of a 3D colored mesh is affected by its geometry and texture. Different mesh processing operations cause diverse geometrical and texture distortions that degrade the visual quality of meshes by perturbing the interactions between the object's shape and color. Figure 1 illustrates an example of such geometry-texture interactions, where either the geometry or texture can affect the visual appearance of distortions in the other. In the top row, the easily visible geometry distortion becomes imperceptible when we replace the uniform texture with a complex one. In the bottom row, the noisy texture patterns become smooth when we modify the geometry. This highlights the need for MQA methods that capture these complex geometry-texture interaction—a significant factor largely overlooked by existing methods." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 133, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 133, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 133, + 712 + ], + "type": "text", + "content": "*Equal contribution" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21414" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 70, + 292, + 200 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 292, + 200 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 292, + 200 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 292, + 200 + ], + "type": "image", + "image_path": "3087de5fca1eec8884553d404e47aef7d5dccc11af4dbe186a1baafcf47920ad.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 208, + 295, + 285 + ], + "lines": [ + { + "bbox": [ + 55, + 208, + 295, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 208, + 295, + 285 + ], + "type": "text", + "content": "Figure 2. Reference and distorted meshes under geometry distortions. Although the distorted meshes (hawk and bowl) have distinct visual qualities (different mean opinion scores (MOS)), Graphics-LPIPS [27] assigns similar scores and reverses their ranking. HybridMQA aligns well with human perception as it understands the mesh's geometry properties. Blue boxes denote the same regions across viewpoints." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 307, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 307, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 307, + 296, + 714 + ], + "type": "text", + "content": "MQA methods may be generally categorized as model-based and projection-based depending on their operation space. Model-based MQA operates directly in the 3D vertex space to extract topology-aware quality features [18, 42]. Their performance is limited due to the lack of access to the object's actual appearance (rendered projections). To compensate, they either (1) apply image quality assessment (IQA) to 2D texture maps [13, 38]; or (2) define color quality descriptors for vertices or faces [28, 48, 53]. However, both fail to capture the final appearance human viewers see. Projection-based MQA, on the other hand, operates on rendered 2D projections [19, 27, 55], hence more effective at assessing texture distortions. Nonetheless, they lack a 3D understanding of the object's topology, which is crucial for accurate quality evaluation. Although underlying geometry distortions may be less noticeable in certain projections (Fig. 1, top row), they become apparent to human viewers when the object is viewed from different angles. Humans intuitively perceive underlying structures: connecting viewpoints, incorporating 3D cues, and detecting distortions that may be imperceptible in isolated 2D projections [6, 31, 57]. This 3D understanding is hard to obtain with projections alone due to the nature of operation space. Fig. 2 illustrates this: the geometry distortion (perturbing vertex positions) is more noticeable along the contours of the hawk and bowl than within their inner regions in 2D projections. However, humans naturally identify viewpoint connections (blue boxes) as objects rotate in 3D, perceiving geometry distortions beneath the texture. Consequently, projection-based MQA metrics (Graphics-LPIPS [27]), which focus on pictorial differences, struggle to detect and penalize these distortions, resulting in wrong rankings (Fig. 2). In contrast, HybridMQA extracts quality-related features from mesh surfaces in 3D, enabling a more accurate MQA." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 553, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 311 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 311 + ], + "type": "text", + "content": "In this paper, we propose a novel hybrid FR MQA method, namely HybridMQA, that integrates model- and projection-based approaches to explore geometry-texture interactions for comprehensive MQA. Specifically, a base encoder extracts initial 3D features from 2D texture, normal, and vertex maps, which initialize a feature graph based on mesh connectivity. A graph convolutional network (GCN) then learns detailed 3D surface representations, building a 3D understanding of the object's topology. These surface representations are then complemented by textural information from the mesh's renderings, enabling a hybrid integration of model- and projection-based methods across all mesh data modalities: 2D maps, 3D structure, and colored renderings. Additionally, to explore the intricate geometry-texture interactions between the two operational spaces, we propose a novel feature projection technique that renders 2D projections of surface representations from the graph, precisely aligned with colored renderings. This alignment facilitates the exploration of geometry-texture interactions via cross-attention. Below are our contributions:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 314, + 312, + 553, + 515 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 314, + 312, + 553, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 312, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 314, + 312, + 553, + 361 + ], + "type": "text", + "content": "- We propose the first hybrid FR MQA method that consolidates the strengths of model- and projection-based approaches across all mesh data modalities, enabling a comprehensive understanding of mesh quality." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 361, + 553, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 361, + 553, + 418 + ], + "spans": [ + { + "bbox": [ + 314, + 361, + 553, + 418 + ], + "type": "text", + "content": "- We make the first attempt to explore geometry-texture interactions for MQA, drawing meaningful connections between the two domains. Our method shows the importance of leveraging such interactions to build reliable MQA methods." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 420, + 553, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 420, + 553, + 468 + ], + "spans": [ + { + "bbox": [ + 314, + 420, + 553, + 468 + ], + "type": "text", + "content": "- We propose a novel feature projection technique to render 2D projections of 3D surface representations from graph, aligned with the mesh's colored renderings, establishing pixel-to-pixel correspondence to explore interactions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 468, + 553, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 468, + 553, + 515 + ], + "spans": [ + { + "bbox": [ + 314, + 468, + 553, + 515 + ], + "type": "text", + "content": "- Our model outperforms state-of-the-art FR MQA methods, aligning with human perception and generalizing better. The results highlight the effectiveness of capturing geometry-texture interactions to achieve accurate MQA." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 529, + 405, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 529, + 405, + 542 + ], + "spans": [ + { + "bbox": [ + 314, + 529, + 405, + 542 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 551, + 520, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 551, + 520, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 551, + 520, + 563 + ], + "type": "text", + "content": "2.1. Model-based Mesh Quality Assessment" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 568, + 553, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 568, + 553, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 568, + 553, + 676 + ], + "type": "text", + "content": "Model-based MQA methods operate directly in 3D, with most existing methods designed for uncolored meshes [4, 5, 17, 18, 39, 41, 42], resulting in suboptimal performance for colored meshes. Early approaches, such as Hausdorff distance [4] and mean squared error (MSE) [5], use Euclidean distance as a quality measure, while Lavoué et al. [17, 18] employ curvature statistics. Other methods use local curvature and roughness pooling [39, 42] or employ dihedral angles as surface quality indicators [41]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "However, the rise of colored meshes drives the need for color integration in MQA methods. Tian and AlRegib [38] and Guo et al. [13] apply MSE and MS-SSIM [44] to" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21415" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 275 + ], + "type": "text", + "content": "2D texture maps, combining these with geometric descriptors to evaluate colored meshes. Nevertheless, 2D texture maps do not represent the post-rendering appearance of 3D meshes and carry little semantic information, leading to suboptimal performance. In contrast, a second group of methods defines color quality descriptors on per-vertex or per-face color values [9, 28, 48, 53]. Nehme et al. [28] propose CMDM, a model for vertex-color meshes that extracts multi-scale color features from vertex colors. Similarly, Zhang et al. [53] apply statistical measures on vertex colors for no-reference MQA. Fu et al. [9] also sample vertex colors and propose an efficient surface sampling approach to convert meshes into point clouds for quality evaluation. Finally, Yang et al. [48] develop GeodesicPSIM, using surface colors to create textured patches for feature extraction. However, these methods fail to capture the rendered appearance of 3D objects as perceived by ultimate human viewers." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 276, + 294, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 276, + 294, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 276, + 294, + 361 + ], + "type": "text", + "content": "Overall, while model-based MQA methods benefit from a profound understanding of the object's topology, their performance is constrained by the lack of access to the object's colored appearance. We address this issue by introducing a novel hybrid MQA method that complements the 3D awareness of model-based methods with textural representations derived from the colored appearance of 3D meshes." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 371, + 280, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 371, + 280, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 371, + 280, + 384 + ], + "type": "text", + "content": "2.2. Projection-based Mesh Quality Assessment" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 389, + 295, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 389, + 295, + 474 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 295, + 474 + ], + "type": "text", + "content": "In projection-based MQA, quality is assessed on rendered 2D projections of 3D meshes, allowing well-established IQA methods such as PSNR [43], SSIM [45], and VIF [34] to be adapted for MQA [7, 32, 47]. However, IQA methods perform poorly in MQA as they are tailored for natural scenes, while 3D meshes involve different distortions, and their 2D projections differ statistically from natural images." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 474, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 295, + 713 + ], + "type": "text", + "content": "A few projection-based methods have been proposed for meshes without color [1, 2], and they expectedly perform poorly on colored meshes. To address this, Nehme et al. [27] propose Graphics-LPIPS, the first projection-based MQA method designed for colored meshes, which builds on LPIPS [52] and uses pre-trained AlexNet [15] to extract quality features from 2D projections. Similarly, Zhang et al. [55] use Swin Transformer [22] with an efficient mini-patch sampling process for no-reference MQA. Lee et al. [19] introduce 3D-PSSIM, which uses a framework similar to Graphics-LPIPS to extract textural information, while complementing it with geometry-aware information derived from the 2D projection space. Nevertheless, these projection-based methods lack 3D understanding of the object's topology. Our approach overcomes this through a well-defined and end-to-end trainable GCN that operates on vertices in the 3D space. Furthermore, these methods fail to account for interactions between the mesh's textural information and the underlying 3D geometry, which limits their performance in detecting geometry-involved distor" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "content": "tions. We address this by proposing a cross-attention framework that draws connections between the two domains." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 106, + 421, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 106, + 421, + 121 + ], + "spans": [ + { + "bbox": [ + 313, + 106, + 421, + 121 + ], + "type": "text", + "content": "3. Proposed Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 127, + 555, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 127, + 555, + 308 + ], + "spans": [ + { + "bbox": [ + 313, + 127, + 555, + 308 + ], + "type": "text", + "content": "Figure 3a shows the overall framework of HybridMQA, consisting of model and texture branches and a quality encoder. In the texture branch, colored projections of the input mesh (reference or distorted) are rendered from six perpendicular viewpoints. In the model branch, 2D normal, vertex, and texture maps are processed by a base encoder to extract initial 3D features, which initialize a graph based on mesh connectivity. A GCN then learns detailed 3D surface representations. At the same viewpoints as the texture branch, we render 3D feature projections from the graph, ensuring pixel-to-pixel alignment with the colored projections. Finally, the quality encoder processes both sets of projections, exploring geometry-texture interactions to produce a comprehensive quality representation of the mesh. Overall, the proposed HybridMQA can be expressed as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 342, + 316, + 553, + 330 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 316, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 342, + 316, + 553, + 330 + ], + "type": "interline_equation", + "content": "\\boldsymbol {f} _ {\\text {m e s h}} = \\text {H y b r i d M Q A} (\\mathcal {M} (\\boldsymbol {V}, \\boldsymbol {E}, \\boldsymbol {U}, \\boldsymbol {T} _ {t}); \\boldsymbol {\\theta}). \\tag {1}", + "image_path": "74c4ad679ebc9bacd902e22816b7e58efad29893e48535eb680b103eed573829.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 339, + 554, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 339, + 554, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 339, + 554, + 434 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 313, + 339, + 554, + 434 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 339, + 554, + 434 + ], + "type": "text", + "content": " is the input mesh, " + }, + { + "bbox": [ + 313, + 339, + 554, + 434 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}" + }, + { + "bbox": [ + 313, + 339, + 554, + 434 + ], + "type": "text", + "content": " represents model parameters, and " + }, + { + "bbox": [ + 313, + 339, + 554, + 434 + ], + "type": "inline_equation", + "content": "f_{mesh}" + }, + { + "bbox": [ + 313, + 339, + 554, + 434 + ], + "type": "text", + "content": " denotes the final quality representation. The mesh " + }, + { + "bbox": [ + 313, + 339, + 554, + 434 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 339, + 554, + 434 + ], + "type": "text", + "content": " consists of vertices " + }, + { + "bbox": [ + 313, + 339, + 554, + 434 + ], + "type": "inline_equation", + "content": "\\pmb{V}" + }, + { + "bbox": [ + 313, + 339, + 554, + 434 + ], + "type": "text", + "content": ", vertex connectivity " + }, + { + "bbox": [ + 313, + 339, + 554, + 434 + ], + "type": "inline_equation", + "content": "\\pmb{E}" + }, + { + "bbox": [ + 313, + 339, + 554, + 434 + ], + "type": "text", + "content": ", UV coordinates " + }, + { + "bbox": [ + 313, + 339, + 554, + 434 + ], + "type": "inline_equation", + "content": "\\pmb{U}" + }, + { + "bbox": [ + 313, + 339, + 554, + 434 + ], + "type": "text", + "content": ", and the 2D texture map " + }, + { + "bbox": [ + 313, + 339, + 554, + 434 + ], + "type": "inline_equation", + "content": "\\pmb{T}_t" + }, + { + "bbox": [ + 313, + 339, + 554, + 434 + ], + "type": "text", + "content": ". Vertices and UV coordinates are defined as " + }, + { + "bbox": [ + 313, + 339, + 554, + 434 + ], + "type": "inline_equation", + "content": "\\pmb{V} = \\{\\pmb{v}_i\\}_{i=1}^N" + }, + { + "bbox": [ + 313, + 339, + 554, + 434 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 339, + 554, + 434 + ], + "type": "inline_equation", + "content": "\\pmb{v}_i \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 339, + 554, + 434 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 339, + 554, + 434 + ], + "type": "inline_equation", + "content": "\\pmb{U} = \\{\\pmb{u}_i\\}_{i=1}^N" + }, + { + "bbox": [ + 313, + 339, + 554, + 434 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 339, + 554, + 434 + ], + "type": "inline_equation", + "content": "\\pmb{u}_i \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 313, + 339, + 554, + 434 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 339, + 554, + 434 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 339, + 554, + 434 + ], + "type": "text", + "content": " is the number of vertices, and " + }, + { + "bbox": [ + 313, + 339, + 554, + 434 + ], + "type": "inline_equation", + "content": "\\pmb{v}_i" + }, + { + "bbox": [ + 313, + 339, + 554, + 434 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 339, + 554, + 434 + ], + "type": "inline_equation", + "content": "\\pmb{u}_i" + }, + { + "bbox": [ + 313, + 339, + 554, + 434 + ], + "type": "text", + "content": " are the positions of the " + }, + { + "bbox": [ + 313, + 339, + 554, + 434 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 339, + 554, + 434 + ], + "type": "text", + "content": "-th vertex in 3D space and 2D map, respectively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 435, + 554, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 435, + 554, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 435, + 554, + 495 + ], + "type": "text", + "content": "Finally, as shown in Fig. 3b, HybridMQA generates quality representations for both reference and distorted meshes. For full-reference quality regression, the absolute differences between these representations are fed into a fully connected network (FCN) to obtain a quality score." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 502, + 401, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 401, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 401, + 514 + ], + "type": "text", + "content": "3.1. Base Encoder" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 521, + 555, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 665 + ], + "type": "text", + "content": "Given an input mesh " + }, + { + "bbox": [ + 313, + 521, + 555, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 521, + 555, + 665 + ], + "type": "text", + "content": ", we first project its 3D geometrical attributes (normals and vertex positions) into 2D maps aligned with the texture map using barycentric interpolation and UV mapping. The shared UV coordinates " + }, + { + "bbox": [ + 313, + 521, + 555, + 665 + ], + "type": "inline_equation", + "content": "\\pmb{U}" + }, + { + "bbox": [ + 313, + 521, + 555, + 665 + ], + "type": "text", + "content": " ensure alignment between the resulting maps (normal and vertex) and the texture map " + }, + { + "bbox": [ + 313, + 521, + 555, + 665 + ], + "type": "inline_equation", + "content": "\\pmb{T}_t" + }, + { + "bbox": [ + 313, + 521, + 555, + 665 + ], + "type": "text", + "content": ". These aligned maps are concatenated and processed by a convolutional neural network (CNN) base encoder to jointly capture textural and geometrical information, learning initial quality-aware representations of the mesh's 3D shape and surface (Fig. 3a). This approach of projecting geometrical data into 2D maps has been successfully applied in mesh super-resolution [46]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 671, + 480, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 480, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 480, + 684 + ], + "type": "text", + "content": "3.2. Graph Convolutional Network" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "Due to vertex-neighborhood discontinuities in 2D maps, we use graph learning to deepen our model's understanding" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21416" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 427, + 221 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 427, + 221 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 427, + 221 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 427, + 221 + ], + "type": "image", + "image_path": "779296ae23dab8a424c7f5466b19a516c8295accee3f6b24ed13f197dc53ff24.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 192, + 223, + 289, + 233 + ], + "lines": [ + { + "bbox": [ + 192, + 223, + 289, + 233 + ], + "spans": [ + { + "bbox": [ + 192, + 223, + 289, + 233 + ], + "type": "text", + "content": "(a) HybridMQA's framework." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 434, + 74, + 555, + 221 + ], + "blocks": [ + { + "bbox": [ + 434, + 74, + 555, + 221 + ], + "lines": [ + { + "bbox": [ + 434, + 74, + 555, + 221 + ], + "spans": [ + { + "bbox": [ + 434, + 74, + 555, + 221 + ], + "type": "image", + "image_path": "10c17530b52a03b726ee239cb2cf3168dd1aec8eaf71f3b1aa73683167153677.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 454, + 224, + 537, + 233 + ], + "lines": [ + { + "bbox": [ + 454, + 224, + 537, + 233 + ], + "spans": [ + { + "bbox": [ + 454, + 224, + 537, + 233 + ], + "type": "text", + "content": "(b) FR quality regression." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 242, + 555, + 299 + ], + "lines": [ + { + "bbox": [ + 54, + 242, + 555, + 299 + ], + "spans": [ + { + "bbox": [ + 54, + 242, + 555, + 299 + ], + "type": "text", + "content": "Figure 3. (a) Overview of HybridMQA. In the model branch, a base encoder extracts 3D features from the mesh's 2D maps, initializing a mesh graph. A GCN extracts 3D surface representations, which are rendered as 2D projections aligned with the colored projections from the texture branch. A quality encoder then captures geometry-texture interactions between the two branches, producing the final mesh quality representation. (b) Full-reference (FR) quality regression, where the absolute difference of HybridMQA's mesh quality representations for reference and distorted meshes is mapped to a quality score via a fully connected network (FCN)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 319, + 294, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 319, + 294, + 355 + ], + "spans": [ + { + "bbox": [ + 54, + 319, + 294, + 355 + ], + "type": "text", + "content": "of the mesh's 3D structure via message-passing between neighboring vertices in 3D, enabling them to share useful quality-related information and highlight abnormalities." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 356, + 295, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 356, + 295, + 393 + ], + "spans": [ + { + "bbox": [ + 54, + 356, + 295, + 393 + ], + "type": "text", + "content": "Graph initialization. Given the base encoder's output feature map with " + }, + { + "bbox": [ + 54, + 356, + 295, + 393 + ], + "type": "inline_equation", + "content": "C_1" + }, + { + "bbox": [ + 54, + 356, + 295, + 393 + ], + "type": "text", + "content": " channels, we sample a " + }, + { + "bbox": [ + 54, + 356, + 295, + 393 + ], + "type": "inline_equation", + "content": "C_1" + }, + { + "bbox": [ + 54, + 356, + 295, + 393 + ], + "type": "text", + "content": "-dimensional feature vector for each mesh vertex using UV mapping:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 100, + 404, + 295, + 420 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 404, + 295, + 420 + ], + "spans": [ + { + "bbox": [ + 100, + 404, + 295, + 420 + ], + "type": "interline_equation", + "content": "\\boldsymbol {f} _ {i} = \\boldsymbol {T} _ {B E} \\left[ u _ {i} ^ {h}, u _ {i} ^ {w} \\right], i \\in \\{1, \\dots , N \\}. \\tag {2}", + "image_path": "73c16af9cf2e3cc9d8c99c427a119352be915a5a045a095bbe10a9589055d532.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 430, + 295, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 430, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 54, + 430, + 295, + 502 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 54, + 430, + 295, + 502 + ], + "type": "inline_equation", + "content": "\\pmb{f}_i \\in \\mathbb{R}^{C_1}" + }, + { + "bbox": [ + 54, + 430, + 295, + 502 + ], + "type": "text", + "content": " is the sampled feature for " + }, + { + "bbox": [ + 54, + 430, + 295, + 502 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 54, + 430, + 295, + 502 + ], + "type": "text", + "content": "-th vertex, " + }, + { + "bbox": [ + 54, + 430, + 295, + 502 + ], + "type": "inline_equation", + "content": "u_i^h" + }, + { + "bbox": [ + 54, + 430, + 295, + 502 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 430, + 295, + 502 + ], + "type": "inline_equation", + "content": "u_i^w" + }, + { + "bbox": [ + 54, + 430, + 295, + 502 + ], + "type": "text", + "content": " are its UV coordinates, and " + }, + { + "bbox": [ + 54, + 430, + 295, + 502 + ], + "type": "inline_equation", + "content": "T_{BE}" + }, + { + "bbox": [ + 54, + 430, + 295, + 502 + ], + "type": "text", + "content": " denotes base encoder's output feature maps. For vertices with multiple UV mappings, we average the sampled features. This process produces a feature graph with vertices initialized by sampled features and edges defined by mesh connectivity " + }, + { + "bbox": [ + 54, + 430, + 295, + 502 + ], + "type": "inline_equation", + "content": "\\pmb{E}" + }, + { + "bbox": [ + 54, + 430, + 295, + 502 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 504, + 295, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 504, + 295, + 540 + ], + "spans": [ + { + "bbox": [ + 54, + 504, + 295, + 540 + ], + "type": "text", + "content": "Graph update. We apply a GCN to integrate features from neighboring vertices to learn surface properties. As proposed by Morris et al. [26], we update vertex features by:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 555, + 295, + 583 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 555, + 295, + 583 + ], + "spans": [ + { + "bbox": [ + 115, + 555, + 295, + 583 + ], + "type": "interline_equation", + "content": "\\boldsymbol {f} _ {i} ^ {\\prime} = \\boldsymbol {\\theta} _ {1} \\boldsymbol {f} _ {i} + \\boldsymbol {\\theta} _ {2} \\sum_ {j \\in \\Gamma (i)} e _ {j, i} \\boldsymbol {f} _ {j}. \\tag {3}", + "image_path": "ed2db9afa466e92e27d6e0f2a16dd93db5b2ae5df2991fbdd67eec72b0e8d562.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 54, + 593, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 593, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 593, + 295, + 715 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 54, + 593, + 295, + 715 + ], + "type": "inline_equation", + "content": "f_{i}^{\\prime} \\in \\mathbb{R}^{C_{2}}" + }, + { + "bbox": [ + 54, + 593, + 295, + 715 + ], + "type": "text", + "content": " is the updated feature vector for the " + }, + { + "bbox": [ + 54, + 593, + 295, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 54, + 593, + 295, + 715 + ], + "type": "text", + "content": "-th vertex, " + }, + { + "bbox": [ + 54, + 593, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\Gamma(i)" + }, + { + "bbox": [ + 54, + 593, + 295, + 715 + ], + "type": "text", + "content": " is its 1-ring neighborhood, " + }, + { + "bbox": [ + 54, + 593, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\theta_{1}, \\theta_{2} \\in \\mathbb{R}^{C_{2} \\times C_{1}}" + }, + { + "bbox": [ + 54, + 593, + 295, + 715 + ], + "type": "text", + "content": " are learnable weights, and " + }, + { + "bbox": [ + 54, + 593, + 295, + 715 + ], + "type": "inline_equation", + "content": "e_{j,i}" + }, + { + "bbox": [ + 54, + 593, + 295, + 715 + ], + "type": "text", + "content": " is the edge weight between vertices " + }, + { + "bbox": [ + 54, + 593, + 295, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 54, + 593, + 295, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 593, + 295, + 715 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 54, + 593, + 295, + 715 + ], + "type": "text", + "content": ", defined as the inverse of their Euclidean distance. Through training, the GCN refines vertex feature embeddings to learn quality-aware 3D surface representations, which are later complemented by textural embeddings to deliver a hybrid and comprehensive quality assessment. This approach of graph learning has been effectively applied to mesh texture downsampling [29]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 319, + 463, + 332 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 463, + 332 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 463, + 332 + ], + "type": "text", + "content": "3.3. Feature & RGB Rendering" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 336, + 555, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 336, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 336, + 555, + 396 + ], + "type": "text", + "content": "We render multiple viewpoints of the 3D surface representations into 2D projections aligned with the mesh's colored projections. This novel feature graph rendering enables us to capture geometry-texture interactions between the two branches for a comprehensive mesh quality representation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 397, + 554, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 397, + 554, + 493 + ], + "spans": [ + { + "bbox": [ + 313, + 397, + 554, + 493 + ], + "type": "text", + "content": "In the texture branch (Fig. 3a), colored projections are rendered after normalizing mesh vertex positions to fit within a unit cube. This standardization allows fixed camera positions for any mesh. Six virtual cameras are placed on the cube's surfaces, facing the object to cover all angles. With PyTorch3D's [30] soft Phong Shader, we render six perpendicular colored projections, employing directional or ambient light to match the conditions of subjective tests." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 494, + 555, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 494, + 555, + 638 + ], + "spans": [ + { + "bbox": [ + 313, + 494, + 555, + 638 + ], + "type": "text", + "content": "The model branch employs the same camera setup to ensure alignment with colored projections. Our novel feature graph rendering customizes PyTorch3D's differentiable renderer to render six perpendicular " + }, + { + "bbox": [ + 313, + 494, + 555, + 638 + ], + "type": "inline_equation", + "content": "3D" + }, + { + "bbox": [ + 313, + 494, + 555, + 638 + ], + "type": "text", + "content": " feature projections from the graph, enabling gradient backpropagation to the GCN and base encoder. To focus on raw vertex features interpolated on the mesh surface, we remove shadows and specular effects, setting diffuse and specular reflectivity to zero and using ambient light. Thus we define the mesh as a vertex-color mesh with " + }, + { + "bbox": [ + 313, + 494, + 555, + 638 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 494, + 555, + 638 + ], + "type": "text", + "content": "-dimensional vertex features (instead of RGB colors) to render the 3D feature projections with hard Phong shader [30]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 647, + 415, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 415, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 415, + 659 + ], + "type": "text", + "content": "3.4. Quality Encoder" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "type": "text", + "content": "We design a quality encoder that processes the aligned projections from model and texture branches to output an expressive quality representation for the mesh. It captures the geometry-texture interactions between the two sets of pro" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21417" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 329, + 259 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 329, + 259 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 329, + 259 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 329, + 259 + ], + "type": "image", + "image_path": "fcfdc02f92da17a7a799b09757e5fbbd95a9a22a714c8adf55ed525bd18acc88.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 164, + 264, + 229, + 274 + ], + "lines": [ + { + "bbox": [ + 164, + 264, + 229, + 274 + ], + "spans": [ + { + "bbox": [ + 164, + 264, + 229, + 274 + ], + "type": "text", + "content": "(a) Quality encoder." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 356, + 74, + 547, + 259 + ], + "blocks": [ + { + "bbox": [ + 356, + 74, + 547, + 259 + ], + "lines": [ + { + "bbox": [ + 356, + 74, + 547, + 259 + ], + "spans": [ + { + "bbox": [ + 356, + 74, + 547, + 259 + ], + "type": "image", + "image_path": "a073be4013f89a4fbe66daa6928f82fe1e3e8f3a63bf7b525c4f011c13425f86.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 409, + 264, + 498, + 274 + ], + "lines": [ + { + "bbox": [ + 409, + 264, + 498, + 274 + ], + "spans": [ + { + "bbox": [ + 409, + 264, + 498, + 274 + ], + "type": "text", + "content": "(b) Cross-attention module." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 284, + 555, + 329 + ], + "lines": [ + { + "bbox": [ + 54, + 284, + 555, + 329 + ], + "spans": [ + { + "bbox": [ + 54, + 284, + 555, + 329 + ], + "type": "text", + "content": "Figure 4. (a) The quality encoder. The 3D feature and color projections are divided into valid aligned patches and fed into respective encoders to obtain multiscale 3D surface and color representations. The cross-attention modules capture interactions between these representations, which are then concatenated with 3D feature embeddings directly extracted from the patches to form the final mesh quality representation. (b) The cross-attention module consists of two transformer blocks, where we switch the roles of the two inputs." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 349, + 296, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 349, + 296, + 409 + ], + "spans": [ + { + "bbox": [ + 54, + 349, + 296, + 409 + ], + "type": "text", + "content": "jections through multi-scale cross-attentions. As discussed in Sec. 1, understanding these interactions is key to a robust and effective MQA. As detailed in Fig. 4a, given model branch's 3D feature projections " + }, + { + "bbox": [ + 54, + 349, + 296, + 409 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_m" + }, + { + "bbox": [ + 54, + 349, + 296, + 409 + ], + "type": "text", + "content": " and texture branch's colored projections " + }, + { + "bbox": [ + 54, + 349, + 296, + 409 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_t" + }, + { + "bbox": [ + 54, + 349, + 296, + 409 + ], + "type": "text", + "content": ", the quality encoder " + }, + { + "bbox": [ + 54, + 349, + 296, + 409 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 54, + 349, + 296, + 409 + ], + "type": "text", + "content": " outputs" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 124, + 415, + 294, + 428 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 415, + 294, + 428 + ], + "spans": [ + { + "bbox": [ + 124, + 415, + 294, + 428 + ], + "type": "interline_equation", + "content": "\\boldsymbol {f} _ {\\text {m e s h}} = \\psi \\left(\\boldsymbol {I} _ {m}, \\boldsymbol {I} _ {t}; \\boldsymbol {\\theta} _ {\\psi}\\right), \\tag {4}", + "image_path": "cb2d5d1c1fa7847a6ce201aa346738bc3a446fc50a9cc3438fca46fbc1da73e8.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 433, + 295, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 433, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 433, + 295, + 456 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 433, + 295, + 456 + ], + "type": "inline_equation", + "content": "f_{mesh}" + }, + { + "bbox": [ + 55, + 433, + 295, + 456 + ], + "type": "text", + "content": " denotes the final quality representation, and " + }, + { + "bbox": [ + 55, + 433, + 295, + 456 + ], + "type": "inline_equation", + "content": "\\theta_{\\psi}" + }, + { + "bbox": [ + 55, + 433, + 295, + 456 + ], + "type": "text", + "content": " denotes the learnable parameters." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 456, + 296, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 296, + 493 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 296, + 493 + ], + "type": "text", + "content": "Patch Selection. Given two sets of projections, we extract non-overlapping patches and discard those with less than " + }, + { + "bbox": [ + 55, + 456, + 296, + 493 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 456, + 296, + 493 + ], + "type": "text", + "content": " non-background pixels. This yields " + }, + { + "bbox": [ + 55, + 456, + 296, + 493 + ], + "type": "inline_equation", + "content": "2P" + }, + { + "bbox": [ + 55, + 456, + 296, + 493 + ], + "type": "text", + "content": " patches:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 104, + 496, + 294, + 512 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 496, + 294, + 512 + ], + "spans": [ + { + "bbox": [ + 104, + 496, + 294, + 512 + ], + "type": "interline_equation", + "content": "\\tilde {\\boldsymbol {I}} _ {m} = \\left\\{\\boldsymbol {I} _ {m, p} \\right\\} _ {p = 1} ^ {P}, \\tilde {\\boldsymbol {I}} _ {t} = \\left\\{\\boldsymbol {I} _ {t, p} \\right\\} _ {p = 1} ^ {P}. \\tag {5}", + "image_path": "722a487536e073781268ac374c5544c8d16724e32e1bff87164774ac155de46f.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 517, + 295, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 295, + 541 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 295, + 541 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 55, + 517, + 295, + 541 + ], + "type": "inline_equation", + "content": "\\tilde{I}_m" + }, + { + "bbox": [ + 55, + 517, + 295, + 541 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 517, + 295, + 541 + ], + "type": "inline_equation", + "content": "\\tilde{I}_t" + }, + { + "bbox": [ + 55, + 517, + 295, + 541 + ], + "type": "text", + "content": " denote aligned patches extracted from 3D feature and colored projections, respectively." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 542, + 296, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 542, + 296, + 591 + ], + "spans": [ + { + "bbox": [ + 55, + 542, + 296, + 591 + ], + "type": "text", + "content": "3D Feature & Image Encoding. Next, we feed pairs of aligned patches " + }, + { + "bbox": [ + 55, + 542, + 296, + 591 + ], + "type": "inline_equation", + "content": "I_{m,p}" + }, + { + "bbox": [ + 55, + 542, + 296, + 591 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 542, + 296, + 591 + ], + "type": "inline_equation", + "content": "I_{t,p}" + }, + { + "bbox": [ + 55, + 542, + 296, + 591 + ], + "type": "text", + "content": " to a 3D feature encoder " + }, + { + "bbox": [ + 55, + 542, + 296, + 591 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 542, + 296, + 591 + ], + "type": "text", + "content": " and an image encoder " + }, + { + "bbox": [ + 55, + 542, + 296, + 591 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 55, + 542, + 296, + 591 + ], + "type": "text", + "content": ", respectively, to extract multiscale quality-aware representations " + }, + { + "bbox": [ + 55, + 542, + 296, + 591 + ], + "type": "inline_equation", + "content": "F_{m,p}" + }, + { + "bbox": [ + 55, + 542, + 296, + 591 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 542, + 296, + 591 + ], + "type": "inline_equation", + "content": "F_{t,p}" + }, + { + "bbox": [ + 55, + 542, + 296, + 591 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 82, + 595, + 295, + 615 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 595, + 295, + 615 + ], + "spans": [ + { + "bbox": [ + 82, + 595, + 295, + 615 + ], + "type": "interline_equation", + "content": "\\boldsymbol {F} _ {m, p} = \\left\\{\\boldsymbol {f} _ {m, p} ^ {s} | s = 1, \\dots , 5 \\right\\} = \\phi (\\boldsymbol {I} _ {m, p}; \\boldsymbol {\\theta} _ {\\phi}), \\tag {6}", + "image_path": "b14299350450ac4bda3026277ad843390990386272766ca743e135bf4cc79f3e.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 88, + 612, + 260, + 625 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 612, + 260, + 625 + ], + "spans": [ + { + "bbox": [ + 88, + 612, + 260, + 625 + ], + "type": "interline_equation", + "content": "\\boldsymbol {F} _ {t, p} = \\left\\{\\boldsymbol {f} _ {t, p} ^ {s} | s = 1, \\dots , 5 \\right\\} = \\varphi (\\boldsymbol {I} _ {t, p}; \\boldsymbol {\\theta} _ {\\varphi}).", + "image_path": "a8f0634076d5b360e71649f96c6c1e3c88048fe8ca88aaa6f4ecefe509b44948.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 629, + 295, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 666 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 55, + 629, + 295, + 666 + ], + "type": "inline_equation", + "content": "\\theta_{\\phi}" + }, + { + "bbox": [ + 55, + 629, + 295, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 629, + 295, + 666 + ], + "type": "inline_equation", + "content": "\\theta_{\\varphi}" + }, + { + "bbox": [ + 55, + 629, + 295, + 666 + ], + "type": "text", + "content": " are learnable parameters of the two encoders, and " + }, + { + "bbox": [ + 55, + 629, + 295, + 666 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 629, + 295, + 666 + ], + "type": "text", + "content": " denotes scale index. " + }, + { + "bbox": [ + 55, + 629, + 295, + 666 + ], + "type": "inline_equation", + "content": "1\\times 1" + }, + { + "bbox": [ + 55, + 629, + 295, + 666 + ], + "type": "text", + "content": " convolutions are used to adjust the channel dimension of representations." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 666, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 296, + 713 + ], + "type": "text", + "content": "Cross-attention Modules. We apply cross-attention to capture the interactions between aligned 3D surface representations " + }, + { + "bbox": [ + 55, + 666, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{m,p}^{s}" + }, + { + "bbox": [ + 55, + 666, + 296, + 713 + ], + "type": "text", + "content": " and color representations " + }, + { + "bbox": [ + 55, + 666, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{t,p}^{s}" + }, + { + "bbox": [ + 55, + 666, + 296, + 713 + ], + "type": "text", + "content": " at each scale. Figure 4b depicts our proposed module with two transformer" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 349, + 555, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 349, + 555, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 349, + 555, + 422 + ], + "type": "text", + "content": "blocks [40], where we alternate the query and key-value roles of " + }, + { + "bbox": [ + 313, + 349, + 555, + 422 + ], + "type": "inline_equation", + "content": "f_{m,p}^{s}" + }, + { + "bbox": [ + 313, + 349, + 555, + 422 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 349, + 555, + 422 + ], + "type": "inline_equation", + "content": "f_{t,p}^{s}" + }, + { + "bbox": [ + 313, + 349, + 555, + 422 + ], + "type": "text", + "content": " to explore their interactions and mutual influence. This simulates how one domain's representation affects the impact of the other domain's representation on perceptual quality, as discussed in Sec. 1. We concatenate outputs to form single-scale quality representation " + }, + { + "bbox": [ + 313, + 349, + 555, + 422 + ], + "type": "inline_equation", + "content": "f_{p}^{s}" + }, + { + "bbox": [ + 313, + 349, + 555, + 422 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 422, + 554, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 554, + 507 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 554, + 507 + ], + "type": "text", + "content": "Mesh Quality Representation. To further exploit the model branch's understanding of mesh's 3D structure and surface properties, we directly extract representation " + }, + { + "bbox": [ + 313, + 422, + 554, + 507 + ], + "type": "inline_equation", + "content": "\\hat{f}_p" + }, + { + "bbox": [ + 313, + 422, + 554, + 507 + ], + "type": "text", + "content": " from model branch patches via a convolutional block. " + }, + { + "bbox": [ + 313, + 422, + 554, + 507 + ], + "type": "inline_equation", + "content": "\\hat{f}_p" + }, + { + "bbox": [ + 313, + 422, + 554, + 507 + ], + "type": "text", + "content": " is concatenated with the five single-scale representations from cross-attention, and averaged over all patches to obtain the final mesh quality representation " + }, + { + "bbox": [ + 313, + 422, + 554, + 507 + ], + "type": "inline_equation", + "content": "f_{mesh}" + }, + { + "bbox": [ + 313, + 422, + 554, + 507 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 353, + 513, + 555, + 547 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 513, + 555, + 547 + ], + "spans": [ + { + "bbox": [ + 353, + 513, + 555, + 547 + ], + "type": "interline_equation", + "content": "\\boldsymbol {f} _ {\\text {m e s h}} = \\frac {1}{P} \\sum_ {p = 1} ^ {P} \\left(\\boldsymbol {f} _ {p} ^ {1} \\oplus \\dots \\oplus \\boldsymbol {f} _ {p} ^ {5} \\oplus \\hat {\\boldsymbol {f}} _ {p}\\right). \\tag {7}", + "image_path": "7c799209ee0fdecba0f8c2caa814d4a91255c5e826533efaa2b8b8dd1fdc842e.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 553, + 503, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 553, + 503, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 553, + 503, + 567 + ], + "type": "text", + "content": "3.5. Quality Regression & Optimization" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 570, + 554, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 554, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 554, + 666 + ], + "type": "text", + "content": "As shown in Fig. 3b, we feed the reference and distorted meshes into HybridMQA separately and map the absolute difference of their quality representations " + }, + { + "bbox": [ + 313, + 570, + 554, + 666 + ], + "type": "inline_equation", + "content": "(|f_{mesh}^{ref} - f_{mesh}^{dis}|)" + }, + { + "bbox": [ + 313, + 570, + 554, + 666 + ], + "type": "text", + "content": " to a quality score via an FCN for full-reference MQA. The model is optimized using a loss function with two terms: mean absolute error (MAE) and rank loss. While MAE ensures accurate quality predictions, rank loss helps differentiate closely rated samples within a mini-batch [36]." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 666, + 479, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 479, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 479, + 677 + ], + "type": "text", + "content": "MAE Loss. The MAE loss is defined as" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 382, + 683, + 555, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 683, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 382, + 683, + 555, + 715 + ], + "type": "interline_equation", + "content": "L _ {m a e} = \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\left| q _ {i} - q _ {i} ^ {\\prime} \\right|, \\tag {8}", + "image_path": "399ef3292e1db726c08d558711fb6cef9c404eacedc7543f1349d7611262d524.jpg" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21418" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "inline_equation", + "content": "q_{i}" + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "inline_equation", + "content": "q_{i}^{\\prime}" + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": " denote the predicted and ground truth quality scores of the " + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "-th sample in the batch, with " + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": " being the batch size. The ground truth is the mean opinion score (MOS) obtained in subjective tests and normalized to [0, 1]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 121, + 295, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 121, + 295, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 121, + 295, + 144 + ], + "type": "text", + "content": "Rank Loss. Since ranking is not differentiable, as proposed by Sun et al. [36], we approximate the rank value as" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 152, + 294, + 197 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 152, + 294, + 197 + ], + "spans": [ + { + "bbox": [ + 67, + 152, + 294, + 197 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {r a n k} ^ {i, j} = \\max \\left(0, \\left| q _ {i} ^ {\\prime} - q _ {j} ^ {\\prime} \\right| - e \\left(q _ {i} ^ {\\prime}, q _ {j} ^ {\\prime}\\right) \\cdot \\left(q _ {i} - q _ {j}\\right)\\right), \\\\ e \\left(q _ {i} ^ {\\prime}, q _ {j} ^ {\\prime}\\right) = \\left\\{ \\begin{array}{l l} 1 & q _ {i} ^ {\\prime} \\geq q _ {j} ^ {\\prime} \\\\ - 1 & q _ {i} ^ {\\prime} < q _ {j} ^ {\\prime} \\end{array} , \\right. \\tag {9} \\\\ \\end{array}", + "image_path": "7f35c1d64c48ce9cab158463305f93b3ace4de7d926532d0be9e5eeb42ccf64c.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 205, + 294, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 205, + 294, + 229 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 294, + 229 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 205, + 294, + 229 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 205, + 294, + 229 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 205, + 294, + 229 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 205, + 294, + 229 + ], + "type": "text", + "content": " denote two samples in a batch, and " + }, + { + "bbox": [ + 55, + 205, + 294, + 229 + ], + "type": "inline_equation", + "content": "e(q_i', q_j')" + }, + { + "bbox": [ + 55, + 205, + 294, + 229 + ], + "type": "text", + "content": " is a sign function. The final rank loss is computed as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 238, + 294, + 278 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 238, + 294, + 278 + ], + "spans": [ + { + "bbox": [ + 106, + 238, + 294, + 278 + ], + "type": "interline_equation", + "content": "L _ {r a n k} = \\frac {1}{B ^ {2} - B} \\sum_ {i = 1} ^ {B} \\sum_ {\\substack {j = 1 \\\\ j \\neq i}} ^ {B} L _ {r a n k} ^ {i, j}. \\tag{10}", + "image_path": "55542702080410d707dd95705ce78985a737101114d263f78c982c2bf2866164.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 286, + 282, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 286, + 282, + 298 + ], + "spans": [ + { + "bbox": [ + 55, + 286, + 282, + 298 + ], + "type": "text", + "content": "Final Loss. The final loss is a weighted sum of the two:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 129, + 308, + 294, + 321 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 308, + 294, + 321 + ], + "spans": [ + { + "bbox": [ + 129, + 308, + 294, + 321 + ], + "type": "interline_equation", + "content": "L = L _ {m a e} + \\lambda L _ {r a n k}, \\tag {11}", + "image_path": "07d903594f5b140727faec1fa73b9734039fe8c71d1f3db16580aeacd6bff535.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 330, + 255, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 330, + 255, + 341 + ], + "spans": [ + { + "bbox": [ + 55, + 330, + 255, + 341 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 330, + 255, + 341 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 330, + 255, + 341 + ], + "type": "text", + "content": " balances the effect of the two loss terms." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 353, + 137, + 366 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 353, + 137, + 366 + ], + "spans": [ + { + "bbox": [ + 55, + 353, + 137, + 366 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 372, + 242, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 372, + 242, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 372, + 242, + 384 + ], + "type": "text", + "content": "4.1. Datasets & Implementation Details" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 389, + 296, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 389, + 296, + 484 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 296, + 484 + ], + "type": "text", + "content": "We validate our method on four publicly available color MQA datasets: Nehme et al. [27], SJTU-TMQA [7], TSMD [47], and CMDM [28] datasets. Nehme et al., the largest available, includes textured meshes with mixed geometric and texture distortions, as do SJTU-TMQA and TSMD. In contrast, CMDM consists of vertex-color meshes with either geometric or color distortions. All datasets use MOS as ground truth." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 486, + 296, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 296, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 296, + 593 + ], + "type": "text", + "content": "We use ResNet50 [14] pre-trained on ImageNet [8] as the image encoder " + }, + { + "bbox": [ + 55, + 486, + 296, + 593 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 55, + 486, + 296, + 593 + ], + "type": "text", + "content": " and a randomly initialized CNN as the 3D feature encoder " + }, + { + "bbox": [ + 55, + 486, + 296, + 593 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 486, + 296, + 593 + ], + "type": "text", + "content": ". 3D feature and colored projections are rendered at " + }, + { + "bbox": [ + 55, + 486, + 296, + 593 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 55, + 486, + 296, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 486, + 296, + 593 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 55, + 486, + 296, + 593 + ], + "type": "text", + "content": " resolutions, with patch sizes of " + }, + { + "bbox": [ + 55, + 486, + 296, + 593 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 55, + 486, + 296, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 486, + 296, + 593 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 55, + 486, + 296, + 593 + ], + "type": "text", + "content": ", respectively, to ensure alignment. For data augmentation, we randomly perturb camera angles and flip the patches in training to improve robustness and generalization. More details of the datasets and implementation are provided in the supplementary material." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 600, + 171, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 171, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 171, + 613 + ], + "type": "text", + "content": "4.2. Experimental Setup" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 618, + 294, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 294, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 294, + 665 + ], + "type": "text", + "content": "We use 5-fold cross-validation without overlap between train and test source content and report the median performance across five experiments. This strategy is applied to all learning-based methods for fair comparison." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "We compare HybridMQA with 11 model-based and projection-based full-reference MQA methods. Model-based methods include Hausdorff Distance (HD) [4], MSDM2 [18], FMPD [42], GeodesicPSIM [48], and Fu" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "content": "et al. [9]. Projection-based methods include PSNR [43], SSIM [45], VIF [34], LPIPS [52], Graphics-LPIPS [27], and 3D-PSSIM [19]. For a fair comparison, all projection-based methods are evaluated under the same rendering settings as HybridMQA, with no prior assumptions about object orientation. We adopt the Spearman rank-order correlation coefficient (SRCC) and the Pearson linear correlation coefficient (PLCC) to compare all methods. Higher values indicate better performance and a stronger correlation between MOS and predicted quality scores [12]. Further information is provided in the supplementary material." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 214, + 432, + 227 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 214, + 432, + 227 + ], + "spans": [ + { + "bbox": [ + 313, + 214, + 432, + 227 + ], + "type": "text", + "content": "4.3.Quantitative Results" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 233, + 555, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 233, + 555, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 233, + 555, + 519 + ], + "type": "text", + "content": "Overall Performance. Table 1 summarizes the performance comparison, from which we make the following observations: (1) HybridMQA outperforms all model-based and projection-based comparison methods across all datasets, including both textured and vertex-color meshes, demonstrating its effectiveness in the quality assessment of colored meshes. Notably, HybridMQA achieves " + }, + { + "bbox": [ + 313, + 233, + 555, + 519 + ], + "type": "inline_equation", + "content": "6.5\\%" + }, + { + "bbox": [ + 313, + 233, + 555, + 519 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 233, + 555, + 519 + ], + "type": "inline_equation", + "content": "7.7\\%" + }, + { + "bbox": [ + 313, + 233, + 555, + 519 + ], + "type": "text", + "content": " performance gain in PLCC over the second-best method, 3D-PSSIM, on Nehme et al. and SJTU-TMQA, respectively; (2) In general, projection-based methods outperform model-based methods which lack access to the object's actual appearance. HybridMQA consolidates the advantages of both types by complementing the textural information extracted from colored projections with 3D representations learned in 3D; (3) Unlike CMDM with single-type distortions, Nehme et al., SJTU-TMQA, and TSMD involve mixed-type and hence more complex distortions. Consequently, while most methods perform well on CMDM, their performance does not extend to the other three datasets. In contrast, HybridMQA consistently performs well on all datasets, showing robustness in assessing complex distortions. This is achieved by HybridMQA's understanding of geometry-texture interactions, which are perturbed uniquely by different distortions." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 521, + 554, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 554, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 554, + 652 + ], + "type": "text", + "content": "Performance by Distortion Type. In Fig. 5, we compare HybridMQA with Graphics-LPIPS based on distortion types in SJTU-TMQA. In addition to the overall superiority of HybridMQA, we observe that while the two methods achieve comparable performance on texture-only distortions (e.g. JPEG [7]), HybridMQA hugely outperforms Graphics-LPIPS on distortions that only affect geometry (e.g. gn [7] and simpNoTex [7, 10]) or both geometry and texture (e.g. qpqtJPEG [7, 11]). This further demonstrates HybridMQA's proficiency in understanding meshes' 3D geometry and its interactions with textural appearance." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "content": "Generalizability. We train models on Nehme et al. and TSMD and test them on SJTU-TMQA. Table 2 shows that HybridMQA significantly outperforms other learning-based methods, showing strong generalizability. Notably, when trained on TSMD and tested on SJTU-TMQA, HybridMQA" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21419" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 80, + 70, + 533, + 247 + ], + "blocks": [ + { + "bbox": [ + 80, + 70, + 533, + 247 + ], + "lines": [ + { + "bbox": [ + 80, + 70, + 533, + 247 + ], + "spans": [ + { + "bbox": [ + 80, + 70, + 533, + 247 + ], + "type": "table", + "html": "
TypeMethodNehmé et al. [27]SJTU-TMQA [7]TSMD [47]CMDM [28]
SRCCPLCCSRCCPLCCSRCCPLCCSRCCPLCC
Model-basedHD [4]0.1070.1750.0600.1400.4460.4620.1890.210
MSDM2 [18]0.3350.3440.0500.1200.0450.2550.4150.517
FMPD [42]0.3910.4040.1560.4580.0770.2180.6150.623
GeodesicPSIM [48]----0.8200.820--
Fu et al. [9]0.6880.696------
Projection-basedPSNR [43]0.3530.3750.2990.2870.7140.7110.8300.839
SSIM [45]0.2100.2260.3940.2890.6730.6740.8520.861
VIF [34]0.5380.5570.4500.4220.8510.8460.8270.837
LPIPS [52]0.6720.6760.7180.7170.7100.7120.8650.918
Graphics-LPIPS [27]0.7220.7460.7900.7620.8340.8120.8590.925
3D-PSSIM [19]0.8820.8420.8420.832--0.8550.854
HybridHybridMQA0.8920.8970.8870.8960.9120.9190.8970.927
", + "image_path": "c760ae9071fe16e1e65e849605e6d225dbc5a07837708570d06e6c1d4976cd35.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 70, + 307, + 282, + 381 + ], + "blocks": [ + { + "bbox": [ + 54, + 255, + 555, + 289 + ], + "lines": [ + { + "bbox": [ + 54, + 255, + 555, + 289 + ], + "spans": [ + { + "bbox": [ + 54, + 255, + 555, + 289 + ], + "type": "text", + "content": "Table 1. SRCC and PLCC scores of MQA methods on four color MQA benchmark datasets. The scores of GeodesicPSIM [48], Fu et al. [9], and 3D-PSSIM [19] are reported directly from their publications as their implementations are not publicly available. Bold and underlined values denote the best and second-best results per column, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 70, + 307, + 282, + 381 + ], + "lines": [ + { + "bbox": [ + 70, + 307, + 282, + 381 + ], + "spans": [ + { + "bbox": [ + 70, + 307, + 282, + 381 + ], + "type": "table", + "html": "
Trained onNehmé et al.TSMD
SRCCPLCCSRCCPLCC
LPIPS0.5920.5840.7120.695
Graphics-LPIPS0.7310.7340.7840.773
HybridMQA0.8000.7830.8460.811
", + "image_path": "3a7e63e6e22f7f6203a8f8aac62b63daac8a5a6f7f4c2b0d57a747ceefee61d8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 388, + 295, + 411 + ], + "lines": [ + { + "bbox": [ + 55, + 388, + 295, + 411 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 295, + 411 + ], + "type": "text", + "content": "Table 2. Generalization evaluation, where models are trained on Nehme et al. and TSMD and tested on SJTU-TMQA." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 61, + 422, + 291, + 540 + ], + "blocks": [ + { + "bbox": [ + 61, + 422, + 291, + 540 + ], + "lines": [ + { + "bbox": [ + 61, + 422, + 291, + 540 + ], + "spans": [ + { + "bbox": [ + 61, + 422, + 291, + 540 + ], + "type": "image", + "image_path": "f36b5af9e603896d80a9603ea60f5870bf8aaea382cecd7f4dddc146bfc82b11.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 547, + 295, + 592 + ], + "lines": [ + { + "bbox": [ + 55, + 547, + 295, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 547, + 295, + 592 + ], + "type": "text", + "content": "Figure 5. SRCC/PLCC performance of HybridMQA and Graphics-LPIPS on various distortion types of SJTU-TMQA dataset. Green distortions affect the geometry or both geometry and texture, while others only impact the texture." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 613, + 295, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 613, + 295, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 613, + 295, + 639 + ], + "type": "text", + "content": "achieves comparable performance to 3D-PSSIM [19], despite 3D-PSSIM being trained directly on SJTU-TMQA." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 647, + 168, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 168, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 168, + 659 + ], + "type": "text", + "content": "4.4. Qualitative Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "We apply GradCAM [33] on 3D feature projections to verify that the model branch effectively captures geometry-aware quality representations. As shown in Fig. 6, the highlighted regions with noticeable artifacts align well with hu" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 318, + 308, + 555, + 457 + ], + "blocks": [ + { + "bbox": [ + 318, + 308, + 555, + 457 + ], + "lines": [ + { + "bbox": [ + 318, + 308, + 555, + 457 + ], + "spans": [ + { + "bbox": [ + 318, + 308, + 555, + 457 + ], + "type": "image", + "image_path": "3f4e7f486e3eb16442125b19ad9b3b3ed0bfaa046c773bf97f24baa69bd386c7.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 468, + 555, + 513 + ], + "lines": [ + { + "bbox": [ + 313, + 468, + 555, + 513 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 555, + 513 + ], + "type": "text", + "content": "Figure 6. GradCAM results on meshes in the model branch. Highlighted regions exhibit more noticeable artifacts, aligning with human perception and showing the model's effectiveness in capturing geometry-aware quality representations." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 533, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 555, + 605 + ], + "type": "text", + "content": "man perception, validating the model branch's ability to extract geometry-aware representations. Additionally, by mapping these results onto the mesh topology via differentiable rendering, our method opens up opportunities for optimizing mesh compression or restoration algorithms. More examples are available in the supplementary material." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "Figure 7 shows the effectiveness of our hybrid model in exploiting interactions between representations learned in texture and model branches. GradCAM is applied before and after cross-attention, showing that the two branches focus on different regions, with the model branch effectively highlighting geometric artifacts. Cross-attention then successfully identifies and attends to perceptually important regions by exploring geometry-texture interactions. Note that the actual cross-attention inputs are patches described" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21420" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 298, + 348 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 298, + 348 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 298, + 348 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 298, + 348 + ], + "type": "image", + "image_path": "44b3247055d76cfe2d14219205c8bfcdf14c4320f5d6d360106ce74a28f08d87.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 354, + 297, + 389 + ], + "lines": [ + { + "bbox": [ + 55, + 354, + 297, + 389 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 297, + 389 + ], + "type": "text", + "content": "Figure 7. GradCAM before and after cross-attention. The two branches focus on different regions, and cross-attention effectively integrates them by attending to perceptually important regions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 410, + 295, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 410, + 295, + 446 + ], + "spans": [ + { + "bbox": [ + 55, + 410, + 295, + 446 + ], + "type": "text", + "content": "in Sec. 3.3 and 3.4, and the heatmaps are the union of the GradCAM results from the reference and distorted patches. More examples are available in the supplementary material." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 446, + 296, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 446, + 296, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 446, + 296, + 483 + ], + "type": "text", + "content": "We also perform gMAD competition [23] to qualitatively compare HybridMQA with Graphics-LPIPS. The results are available in the supplementary material." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 491, + 157, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 157, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 157, + 502 + ], + "type": "text", + "content": "4.5. Ablation Studies" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 510, + 295, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 545 + ], + "type": "text", + "content": "We conduct ablation studies on Nehme et al. [27] to evaluate key components and design choices in our model. Further results are available in the supplementary material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "text", + "content": "3D Surface and Textural Representations. To show the contributions of 3D surface and textural representations, we test four configurations: (1) 3D-only, where only 3D representations " + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\hat{f}" + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "inline_equation", + "content": "f_{m}" + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "text", + "content": " are used for mesh quality representation; (2) RGB-only, where only textural representation " + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "text", + "content": " from colored renderings are used; (3) all but excluding " + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "inline_equation", + "content": "f_{m}" + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "text", + "content": "; and (4) excluding " + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\hat{f}" + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "text", + "content": ". Table 3 shows that all representations contribute to the final performance, with textural representation " + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "text", + "content": " the most significant. Moreover, excluding 3D representations (RGB-only) leads to a significant performance drop, showing the importance of 3D representations and the need to explore their interaction with textural information for accurate MQA. The results also highlight the performance gain of our hybrid approach, as it outperforms both" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 342, + 70, + 529, + 150 + ], + "blocks": [ + { + "bbox": [ + 342, + 70, + 529, + 150 + ], + "lines": [ + { + "bbox": [ + 342, + 70, + 529, + 150 + ], + "spans": [ + { + "bbox": [ + 342, + 70, + 529, + 150 + ], + "type": "table", + "html": "
NotesffmftSRCCPLCC
3D-only✓✓-0.5860.595
RGB-only--✓0.8200.846
-✓-✓0.8420.849
--✓✓0.8700.874
Proposed✓✓✓0.8920.897
", + "image_path": "aa603b712210cc1c06ffed62344f211b0bd882e576b19163908ee2d10c4ce4ee.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 349, + 170, + 522, + 259 + ], + "blocks": [ + { + "bbox": [ + 318, + 158, + 550, + 169 + ], + "lines": [ + { + "bbox": [ + 318, + 158, + 550, + 169 + ], + "spans": [ + { + "bbox": [ + 318, + 158, + 550, + 169 + ], + "type": "text", + "content": "Table 3. Ablation on 3D Surface & Text. Repr. on Nehme et al." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 349, + 170, + 522, + 259 + ], + "lines": [ + { + "bbox": [ + 349, + 170, + 522, + 259 + ], + "spans": [ + { + "bbox": [ + 349, + 170, + 522, + 259 + ], + "type": "table", + "html": "
ConfigurationsSRCCPLCC
w/o Texture map0.8660.872
w/o Normal map0.8630.866
w/o Vertex map0.8720.873
w/o Base Encoder0.8560.851
w/o GCN0.8660.865
HybridMQA (proposed)0.8920.897
", + "image_path": "9249fc6cab677618333f5bcfbb7fa8c15b80ce490ae326a5eed170424cd67687.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 317, + 268, + 550, + 279 + ], + "lines": [ + { + "bbox": [ + 317, + 268, + 550, + 279 + ], + "spans": [ + { + "bbox": [ + 317, + 268, + 550, + 279 + ], + "type": "text", + "content": "Table 4. Ablation on model branch components on Nehme et al." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 300, + 466, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 466, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 466, + 312 + ], + "type": "text", + "content": "3D-only and RGB-only counterparts." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 312, + 555, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 555, + 420 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 555, + 420 + ], + "type": "text", + "content": "Model Branch Components. To validate the contributions of model branch components, we test the following configurations: all components but excluding (1) the texture map; (2) the normal map; (3) the vertex map; (4) the Base Encoder; and (5) the GCN. Table 4 shows that all components contribute to learning effective 3D surface representations. Furthermore, the proposed processing units (Base Encoder and GCN) effectively leverage mesh data modalities (2D maps and 3D vertex space) to enhance performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 430, + 457, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 457, + 442 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 457, + 442 + ], + "type": "text", + "content": "5. Conclusion & Discussion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 450, + 555, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 555, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 555, + 557 + ], + "type": "text", + "content": "We present a novel hybrid full-reference MQA method that integrates model- and projection-based approaches for enhanced quality assessment. Our model explores interactions between mesh texture and 3D geometry via cross-attention, enabled by a novel feature rendering process that aligns 3D representations with colored projections. Extensive experiments show the effectiveness and superiority of our method, highlighting the importance of 3D understanding and leveraging geometry-texture interactions for reliable MQA." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": "A few limitations present room for improvement. Our method relies on perpendicular viewpoints without considering their varying contributions to perceptual quality. Integrating a semantic-aware module could address this by weighting viewpoints based on importance. Also, memory consumption scales with mesh size as the GCN processes the entire mesh graph, limiting efficiency in real-world applications. Sampling techniques [16, 50, 51] could help reduce memory footprint. Future directions include (1) adapting our work for no-reference MQA; (2) generalizing it to point cloud quality assessment [54, 56] by defining graph edges; and (3) exploring applications in perceptually optimized mesh compression, enhancement, and generation." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21421" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "text", + "content": "[1] Ilyass Aboulaziz, Aladine Chetouani, Mohammed El Hassouni, and Hocine Cherifi. A blind mesh visual quality assessment method based on convolutional neural network. *Electronic Imaging*, 30:1-5, 2018. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 137, + 296, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 296, + 192 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 296, + 192 + ], + "type": "text", + "content": "[2] Ilyass Abouelaziz, Aladine Chetouani, Mohammed El Hassouni, Longin Jan Latecki, and Hocine Cherifi. No-reference mesh visual quality assessment via ensemble of convolutional neural networks and compact multi-linear pooling. Pattern Recognition, 100:107174, 2020. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 194, + 295, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 194, + 295, + 237 + ], + "spans": [ + { + "bbox": [ + 62, + 194, + 295, + 237 + ], + "type": "text", + "content": "[3] Martin Alain, Emin Zerman, Cagri Ozcinar, and Giuseppe Valenzise. Chapter 1 - introduction to immersive video technologies. In Immersive Video Technologies, pages 3-24. Academic Press, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 239, + 295, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 239, + 295, + 293 + ], + "spans": [ + { + "bbox": [ + 62, + 239, + 295, + 293 + ], + "type": "text", + "content": "[4] Nicolas Aspert, Diego Santa-Cruz, and Touradj Ebrahimi. Mesh: measuring errors between surfaces using the hausdorff distance. In Proceedings. IEEE International Conference on Multimedia and Expo, pages 705-708 vol.1, 2002. 2, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 295, + 295, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 295, + 295, + 327 + ], + "spans": [ + { + "bbox": [ + 62, + 295, + 295, + 327 + ], + "type": "text", + "content": "[5] Paolo Cignoni, Claudio Rocchini, and Roberto Scopigno. Metro: Measuring error on simplified surfaces. Computer Graphics Forum, 17(2):167-174, 1998. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 329, + 296, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 329, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 62, + 329, + 296, + 384 + ], + "type": "text", + "content": "[6] Sahin Coskun, Gokce Nur Yilmaz, Federica Battisti, Musaed Alhussein, and Saiful Islam. Measuring 3d video quality of experience (qoe) using a hybrid metric based on spatial resolution and depth cues. Journal of Imaging, 9(12):281, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 386, + 295, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 386, + 295, + 451 + ], + "spans": [ + { + "bbox": [ + 62, + 386, + 295, + 451 + ], + "type": "text", + "content": "[7] Bingyang Cui, Qi Yang, Kaifa Yang, Yiling Xu, Xiaozhong Xu, and Shan Liu. Sjtu-tmqa: A quality assessment database for static mesh with texture map. In ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 7875-7879, 2024. 3, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 453, + 295, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 453, + 295, + 497 + ], + "spans": [ + { + "bbox": [ + 62, + 453, + 295, + 497 + ], + "type": "text", + "content": "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 498, + 295, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 498, + 295, + 564 + ], + "spans": [ + { + "bbox": [ + 62, + 498, + 295, + 564 + ], + "type": "text", + "content": "[9] Chunyang Fu, Xiang Zhang, Thuong Nguyen-Canh, Xiaozhong Xu, Ge Li, and Shan Liu. Surface-sampling based objective quality assessment metrics for meshes. In ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5, 2023. 3, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 566, + 296, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 566, + 296, + 621 + ], + "spans": [ + { + "bbox": [ + 57, + 566, + 296, + 621 + ], + "type": "text", + "content": "[10] Michael Garland and Paul S. Heckbert. Surface simplification using quadric error metrics. In Proceedings of the 24th Annual Conference on Computer Graphics and Interactive Techniques, page 209-216, USA, 1997. ACM Press/Addison-Wesley Publishing Co. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 623, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 623, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 57, + 623, + 294, + 644 + ], + "type": "text", + "content": "[11] Google. Draco 3d data compression. https://google.github.io/draco/. Accessed: 2024-09-04. 1,6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 646, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 646, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 646, + 295, + 689 + ], + "type": "text", + "content": "[12] Video Quality Experts Group et al. Final report from the video quality experts group on the validation of objective models of video quality assessment, phase II. VQEG, 2003. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "text", + "content": "[13] Jinjiang Guo, Vincent Vidal, Irene Cheng, Anup Basu, Atilla Baskurt, and Guillaume Lavoue. Subjective and objective" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "visual quality assessment of textured 3d meshes. ACM Trans. Appl. Percept., 14(2), 2016. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 96, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 553, + 139 + ], + "type": "text", + "content": "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 140, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 553, + 183 + ], + "type": "text", + "content": "[15] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2012. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 185, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 185, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 317, + 185, + 553, + 228 + ], + "type": "text", + "content": "[16] Itai Lang, Asaf Manor, and Shai Avidan. Samplenet: Differentiable point cloud sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7578-7588, 2020. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 228, + 553, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 228, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 317, + 228, + 553, + 281 + ], + "type": "text", + "content": "[17] Guillaume Lavoué, Elisa Drelie Gelasca, Florent Dupont, Atilla Baskurt, and Touradj Ebrahimi. Perceptually driven 3d distance metrics with application to watermarking. Applications of Digital Image Processing XXIX, 6312:63120L, 2006. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 283, + 553, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 283, + 553, + 315 + ], + "spans": [ + { + "bbox": [ + 316, + 283, + 553, + 315 + ], + "type": "text", + "content": "[18] Guillaume Lavoué. A multiscale metric for 3d mesh visual quality assessment. Computer Graphics Forum, 30(5):1427-1437, 2011. 2, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 317, + 553, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 317, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 317, + 317, + 553, + 371 + ], + "type": "text", + "content": "[19] Seongmin Lee, Jiwoo Kang, Sanghoon Lee, Weisi Lin, and Alan Conrad Bovik. 3d-pssim: Projective structural similarity for 3d mesh quality assessment robust to topological irregularities. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-16, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 372, + 553, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 553, + 404 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 553, + 404 + ], + "type": "text", + "content": "[20] Jie Li and Pablo Cesar. Chapter 22 - social virtual reality (vr) applications and user experiences. In Immersive Video Technologies, pages 609-648. Academic Press, 2023. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 405, + 553, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 405, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 317, + 405, + 553, + 459 + ], + "type": "text", + "content": "[21] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Mukund Varma T, Zexiang Xu, and Hao Su. One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. Advances in Neural Information Processing Systems, 36, 2024. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 460, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 460, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 317, + 460, + 553, + 514 + ], + "type": "text", + "content": "[22] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10012-10022, 2021. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 515, + 553, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 515, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 317, + 515, + 553, + 569 + ], + "type": "text", + "content": "[23] Kede Ma, Zhengfang Duanmu, Zhou Wang, Qingbo Wu, Wentao Liu, Hongwei Yong, Hongliang Li, and Lei Zhang. Group maximum differentiation competition: Model comparison with few samples. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(4):851-864, 2020. 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 570, + 553, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 570, + 553, + 603 + ], + "spans": [ + { + "bbox": [ + 317, + 570, + 553, + 603 + ], + "type": "text", + "content": "[24] Adrien Maglo, Guillaume Lavoué, Florent Dupont, and Céline Hudelot. 3d mesh compression: Survey, comparisons, and emerging trends. ACM Comput. Surv., 47(3), 2015. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 604, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 604, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 317, + 604, + 553, + 635 + ], + "type": "text", + "content": "[25] Adrien Maglo, Guillaume Lavoué, Florent Dupont, and Céline Hudelot. 3d mesh compression: Survey, comparisons, and emerging trends. ACM Comput. Surv., 47(3), 2015. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 636, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 636, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 317, + 636, + 553, + 690 + ], + "type": "text", + "content": "[26] Christopher Morris, Martin Ritzert, Matthias Fey, William L. Hamilton, Jan Eric Lenssen, Gaurav Rattan, and Martin Grohe. Weisfeiler and leman go neural: Higher-order graph neural networks. Proceedings of the AAAI Conference on Artificial Intelligence, 33(01):4602-4609, 2019. 4" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[27] Yana Nehme, Johanna Delanoy, Florent Dupont, JeanPhilippe Farrugia, Patrick Le Callet, and Guillaume Lavoué." + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21422" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 77, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 294, + 106 + ], + "type": "text", + "content": "Textured mesh quality assessment: Large-scale dataset and deep learning-based quality metric. ACM Trans. Graph., 42 (3), 2023. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 295, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 295, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 295, + 171 + ], + "type": "text", + "content": "[28] Yana Nehme, Florent Dupont, Jean-Philippe Farrugia, Patrick Le Callet, and Guillaume Lavoué. Visual quality of 3d meshes with diffuse colors in virtual reality: Subjective and objective evaluation. IEEE Transactions on Visualization and Computer Graphics, 27(3):2202-2219, 2021. 2, 3, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 172, + 294, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 294, + 216 + ], + "type": "text", + "content": "[29] Sai Karthikey Pentapati, Anshul Rai, Arkady Ten, Chaitanya Atluru, and Alan Bovik. Geoscaler: Geometry and rendering-aware downsampling of 3d mesh textures. arXiv preprint arXiv:2311.16581, 2023. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 217, + 294, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 259 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 259 + ], + "type": "text", + "content": "[30] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3d deep learning with pytorch3d. arXiv:2007.08501, 2020. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 261, + 294, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 294, + 326 + ], + "type": "text", + "content": "[31] Stephan Reichelt, Ralf Häussler, Gerald Füfterer, and Norbert Leister. Depth cues in human visual perception and their realization in 3d displays. In Three-Dimensional Imaging, Visualization, and Display 2010 and Display Technologies and Applications for Defense, Security, and Avionics IV, pages 92–103. SpIE, 2010. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 327, + 294, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 370 + ], + "type": "text", + "content": "[32] Armin Shafiee Sarvestani, Wei Zhou, and Zhou Wang. Perceptual crack detection for rendered 3d textured meshes. In 2024 16th International Conference on Quality of Multimedia Experience (QoMEX), pages 1-7, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 371, + 294, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 294, + 435 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 294, + 435 + ], + "type": "text", + "content": "[33] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 437, + 294, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 470 + ], + "type": "text", + "content": "[34] Hamid R. Sheikh and Alan C. Bovik. Image information and visual quality. IEEE Transactions on Image Processing, 15 (2):430-444, 2006. 3, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 471, + 294, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 294, + 535 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 294, + 535 + ], + "type": "text", + "content": "[35] Yawar Siddiqui, Tom Monnier, Filippos Kokkinos, Mahendra Kariya, Yanir Kleiman, Emilien Garreau, Oran Gafni, Natalia Neverova, Andrea Vedaldi, Roman Shapovalov, et al. Meta 3d assetgen: Text-to-mesh generation with high-quality geometry, texture, and pbr materials. arXiv preprint arXiv:2407.02445, 2024. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 536, + 294, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 536, + 294, + 592 + ], + "spans": [ + { + "bbox": [ + 56, + 536, + 294, + 592 + ], + "type": "text", + "content": "[36] Wei Sun, Xiongkuo Min, Wei Lu, and Guangtao Zhai. A deep learning based no-reference quality assessment model forUGC videos. In Proceedings of the 30th ACM International Conference on Multimedia, page 856-865, New York, NY, USA, 2022. Association for Computing Machinery. 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 592, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 294, + 624 + ], + "type": "text", + "content": "[37] Yuk Ming Tang and Ho Lun Ho. 3d modeling and computer graphics in virtual reality. In mixed reality and three-dimensional computer graphics. IntechOpen, 2020. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 625, + 294, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 294, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 294, + 668 + ], + "type": "text", + "content": "[38] Dihong Tian and Ghassan AlRegib. Batex3: Bit allocation for progressive transmission of textured 3-d models. IEEE Transactions on Circuits and Systems for Video Technology, 18(1):23-35, 2008. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 669, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 294, + 713 + ], + "type": "text", + "content": "[39] Fakhri Torkhani, Kai Wang, and Jean-Marc Chassery. A curvature tensor distance for mesh visual quality assessment. In International Conference on Computer Vision and Graphics, pages 253–263. Springer, 2012. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 126 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 126 + ], + "type": "text", + "content": "[40] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 128, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 128, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 128, + 553, + 171 + ], + "type": "text", + "content": "[41] Libor Váša and Jan Rus. Dihedral angle mesh error: a fast perception correlated distortion measure for fixed connectivity triangle meshes. Computer Graphics Forum, 31(5):1715-1724, 2012. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 172, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 172, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 172, + 553, + 217 + ], + "type": "text", + "content": "[42] Kai Wang, Fakhri Torkhani, and Annick Montanvert. A fast roughness-based approach to the assessment of 3d mesh visual quality. Computers & Graphics, 36(7):808-818, 2012. Augmented Reality Computer Graphics in China. 2, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 217, + 553, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 249 + ], + "type": "text", + "content": "[43] Zhou Wang and Alan C. Bovik. Mean squared error: Love it or leave it? a new look at signal fidelity measures. IEEE Signal Processing Magazine, 26(1):98-117, 2009. 3, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 250, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 250, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 250, + 553, + 293 + ], + "type": "text", + "content": "[44] Zhou Wang, Eero P Simoncelli, and Alan C Bovik. Multiscale structural similarity for image quality assessment. In The Thirty-Seventh Asilomar Conference on Signals, Systems & Computers, 2003, pages 1398-1402 Vol.2, 2003. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 293, + 553, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 293, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 316, + 293, + 553, + 338 + ], + "type": "text", + "content": "[45] Zhou Wang, Alan C. Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 3, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 338, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 553, + 392 + ], + "type": "text", + "content": "[46] Wuyuan Xie, Tengcong Huang, and Miaohui Wang. Mnsnet: multimodal transformer network for 3d surface superresolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12703-12712, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 393, + 553, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 393, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 316, + 393, + 553, + 448 + ], + "type": "text", + "content": "[47] Qi Yang, Joel Jung, Haiqiang Wang, Xiaozhong Xu, and Shan Liu. Tsmd: A database for static color mesh quality assessment study. In 2023 IEEE International Conference on Visual Communications and Image Processing (VCIP), pages 1-5, 2023. 3, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 449, + 553, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 491 + ], + "type": "text", + "content": "[48] Qi Yang, Joel Jung, Xiaozhong Xu, and Shan Liu. Geodesicpsim: Predicting the quality of static mesh with texture map via geodesic patch similarity. arXiv preprint arXiv:2308.04928, 2023. 2, 3, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 492, + 553, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 492, + 553, + 535 + ], + "spans": [ + { + "bbox": [ + 316, + 492, + 553, + 535 + ], + "type": "text", + "content": "[49] Gareth W. Young, Néill O'Dwyer, and Aljosa Smolic. Chapter 21 - volumetric video as a novel medium for creative storytelling. In Immersive Video Technologies, pages 591-607. Academic Press, 2023. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 536, + 553, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 536, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 316, + 536, + 553, + 581 + ], + "type": "text", + "content": "[50] Hanqing Zeng, Hongkuan Zhou, Ajitesh Srivastava, Rajgopal Kannan, and Viktor Prasanna. GraphSAINT: Graph sampling based inductive learning method. In International Conference on Learning Representations, 2020. 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 582, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 553, + 635 + ], + "type": "text", + "content": "[51] Hanqing Zeng, Muhan Zhang, Yinglong Xia, Ajitesh Srivastava, Andrey Malevich, Rajgopal Kannan, Viktor Prasanna, Long Jin, and Ren Chen. Decoupling the depth and scope of graph neural networks. Advances in Neural Information Processing Systems, 34:19665-19679, 2021. 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 635, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 690 + ], + "type": "text", + "content": "[52] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3, 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[53] Zicheng Zhang, Wei Sun, Xiongkuo Min, Tao Wang, Wei Lu, and Guangtao Zhai. No-reference quality assessment for" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21423" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 362 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 75, + 72, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 72, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 296, + 106 + ], + "type": "text", + "content": "3d colored point cloud and mesh models. IEEE Transactions on Circuits and Systems for Video Technology, 32(11):7618-7631, 2022. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 296, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 296, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 296, + 183 + ], + "type": "text", + "content": "[54] Zicheng Zhang, Wei Sun, Xiongkuo Min, Qiyuan Wang, Jun He, Quan Zhou, and Guangtao Zhai. Mm-pcqa: Multi-modal learning for no-reference point cloud quality assessment. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, IJCAI-23, pages 1759–1767. International Joint Conferences on Artificial Intelligence Organization, 2023. Main Track. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 296, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 296, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 296, + 239 + ], + "type": "text", + "content": "[55] Zicheng Zhang, Wei Sun, Haoning Wu, Yingjie Zhou, Chunyi Li, Zijian Chen, Xiongkuo Min, Guangtao Zhai, and Weisi Lin. Gms-3dqa: Projection-based grid mini-patch sampling for 3d model quality assessment. ACM Trans. Multimedia Comput. Commun. Appl., 20(6), 2024. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 240, + 296, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 240, + 296, + 306 + ], + "spans": [ + { + "bbox": [ + 56, + 240, + 296, + 306 + ], + "type": "text", + "content": "[56] Zicheng Zhang, Haoning Wu, Yingjie Zhou, Chunyi Li, Wei Sun, Chaofeng Chen, Xiongkuo Min, Xiaohong Liu, Weisi Lin, and Guangtao Zhai. Lmm-pcqa: Assisting point cloud quality assessment with lmm. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 7783-7792, 2024. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 307, + 296, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 307, + 296, + 362 + ], + "spans": [ + { + "bbox": [ + 56, + 307, + 296, + 362 + ], + "type": "text", + "content": "[57] Wei Zhou, Qi Yang, Wu Chen, Qiuping Jiang, Guangtao Zhai, and Weisi Lin. Blind quality assessment of dense 3d point clouds with structure guided resampling. ACM Transactions on Multimedia Computing, Communications and Applications, 2024. 2" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "21424" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/HyperFree_ A Channel-adaptive and Tuning-free Foundation Model for Hyperspectral Remote Sensing Imagery/9236da20-05f2-4327-b226-ca754d2dead5_content_list.json b/2025/HyperFree_ A Channel-adaptive and Tuning-free Foundation Model for Hyperspectral Remote Sensing Imagery/9236da20-05f2-4327-b226-ca754d2dead5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a16c2d577fa10e61a984513b9d92f5d363979f81 --- /dev/null +++ b/2025/HyperFree_ A Channel-adaptive and Tuning-free Foundation Model for Hyperspectral Remote Sensing Imagery/9236da20-05f2-4327-b226-ca754d2dead5_content_list.json @@ -0,0 +1,1411 @@ +[ + { + "type": "text", + "text": "HyperFree: A Channel-adaptive and Tuning-free Foundation Model for Hyperspectral Remote Sensing Imagery", + "text_level": 1, + "bbox": [ + 137, + 128, + 859, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jingtao Li\\*, Yingyi Liu\\*, Xinyu Wang†, Yunning Peng†, Chen Sun†, Shaoyu Wang², Zhendong Sun†, Tian Ke†, Xiao Jiang†, Tangwei Lu†, Anran Zhao†, Yanfei Zhong†", + "bbox": [ + 156, + 202, + 841, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{Jintaoli, liuyingyi, wangxinyu, pengyunning, sunchen,", + "bbox": [ + 323, + 261, + 669, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sunzhendong, ketian, jiangxiao, lutangwei, anranzha0415, zhongyanfei}@whu.edu.cn", + "bbox": [ + 238, + 277, + 756, + 291 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "wsy1995@snu.ac.kr", + "bbox": [ + 434, + 297, + 557, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 344, + 326, + 359 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Advanced interpretation of hyperspectral remote sensing images benefits many precise Earth observation tasks. Recently, visual foundation models have promoted the remote sensing interpretation but concentrating on RGB and multispectral images. Due to the varied hyperspectral channels, existing foundation models would face image-by-image tuning situation, imposing great pressure on hardware and time resources. In this paper, we propose a tuning-free hyperspectral foundation model called HyperFree, by adapting the existing visual prompt engineering. To process varied channel numbers, we design a learned weight dictionary covering full-spectrum from $0.4 \\sim 2.5\\mu m$ , supporting to build the embedding layer dynamically. To make the prompt design more tractable, HyperFree can generate multiple semantic-aware masks for one prompt by treating feature distance as semantic-similarity. After pre-training HyperFree on constructed large-scale high-resolution hyperspectral images, HyperFree (1 prompt) has shown comparable results with specialized models (5 shots) on 5 tasks and 11 datasets. Code and dataset are accessible at https://rsidea.whu.edu.cn/hyperfree.htm.", + "bbox": [ + 88, + 375, + 483, + 694 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 720, + 220, + 736 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hyperspectral imaging has been recognized by the European Union as one of the Top 100 Disruptive Technologies [58], due to its ability to provide high-resolution spectral data for precise Earth observation across diverse applications [64], including environmental monitoring [31], agriculture [69], and defense [20]. Unlike traditional RGB or multispectral imagery, Hyperspectral imagery (HSI) offers several unique features: high spectral resolution (typically", + "bbox": [ + 89, + 744, + 483, + 867 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/67039194eb890a9a9984274c3a3a021c040dff12b5c1be6bdd1b1d11382ef41d.jpg", + "image_caption": [ + "(a) Mainstream Pre-training and Tuning Paradigm of RSFM" + ], + "image_footnote": [], + "bbox": [ + 522, + 344, + 895, + 421 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b431c2ad7d207f8a9bdd84e93394e2b6a672f38cb29bc8e63c08bd8841c309fa.jpg", + "image_caption": [ + "(b) Prompt Engineering of SAM-variants" + ], + "image_footnote": [], + "bbox": [ + 522, + 445, + 895, + 513 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5629edb7992d3e8971b9a8db18f3f469069c80024c77191cdfecf3fb7355b68b.jpg", + "image_caption": [ + "(c) Prompt Engineering of HyperFree", + "Figure 1. Comparison of existing RSFMs with the proposed HyperFree model. (a) The pre-training and tuning paradigm is the mainstream approach, where fine-tuning is essential for handling unseen HSIs. (b) The prompt engineering paradigm reduces the tuning burden but struggles to handle HSIs with hundreds of spectral channels. (c) HyperFree can directly process unseen HSIs with varying channel counts without fine-tuning, and it can identify multiple masks of the same class using only a single prompt." + ], + "image_footnote": [], + "bbox": [ + 521, + 539, + 895, + 606 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$10\\mathrm{nm}$ ), an extensive wavelength range (400–2500 nm), and dozens to hundreds of spectral channels [64], which enable the detailed identification of diverse targets, while also increasing the complexity of data processing and analysis.", + "bbox": [ + 511, + 758, + 906, + 820 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, remote sensing foundation models (RSFMs) have demonstrated advanced interpretive capabilities, primarily focusing on multispectral and RGB images [4, 7, 13, 25, 38, 49], while also opening new possibilities for analyzing hyperspectral images. RSFMs follow two main", + "bbox": [ + 511, + 820, + 908, + 897 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 107, + 875, + 220, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding authors.", + "bbox": [ + 109, + 887, + 241, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23048", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "paradigms. The predominant paradigm is pre-training and tuning (P-T), illustrated in Figure 1(a), which begins with training a large-scale backbone model in a self-supervised manner, such as masked autoencoders (MAE) [12], followed by fine-tuning with task-specific heads for various downstream applications [7, 13, 38, 49]. In the P-T paradigm, both the pre-training and the fine-tuning stages can be time-consuming. Second paradigm is prompt engineering (P-E), as illustrated in Figure 1(b), which focuses on designing task-specific prompts for a frozen Segment Anything Model (SAM) [17] to directly perform downstream tasks without additional model tuning [4, 25], making it more efficient and user-friendly in practical.", + "bbox": [ + 89, + 90, + 480, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The P-E paradigm is particularly well-suited for processing HSI, as it can be directly applied to diverse, unseen scenes without requiring model fine-tuning. However, it faces two significant challenges when applied to HSI. (i) The first challenge is the variation in spectral parameters across different hyperspectral sensors, including differences in spectral resolution, wavelength range, and the number of spectral channels [10]. For a P-T-based model like HyperSigma [54], although it is pre-trained using HSIs with fixed channels, the challenge can be addressed during its tuning stage by rebuilding the channel embedding layer to align the dimensions with the new data. But for a P-E-based model, the challenge of channel variation requires more careful consideration. (ii) The second challenge is to make prompt input tractable, ensuring a P-E based model can directly process unseen HSIs with least prior knowledge. Specifically, most existing P-E-based RSFMs rely on a frozen SAM [17], which segments remote sensing objects based on specialized models to generate prompts [4, 25]. As shown in Figure 1(b), each prompt typically generates a single mask and all the prompts for target objects are needed by existing P-E-based RSFMs, which largely reduced the model feasibility.", + "bbox": [ + 91, + 287, + 482, + 633 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the above challenges, as shown in Figure 1 (c), we propose a channel-adaptive, tuning-free hyperspectral foundation model, called HyperFree, which can process any unseen HSIs directly in a P-E manner. Specifically, (i) Motivated by the word dictionary in natural language processing, which stores vector embeddings as values corresponding to words and phrases of varying lengths as keys [32], HyperFree introduces a learnable wavelength weight dictionary in the embedding layer, spanning the 400 to 2500 nm spectral range in $10\\mathrm{nm}$ intervals. This allows for the dynamic encoding of unseen HSIs with varying spectral parameters into visual tokens, all with consistent dimensionality. (ii) To make the prompt of HyperFree more tractable, HyperFree is designed to generate multiple semantic-aware masks for one prompt as in Figure 1 (c), where HyperFree maps each prompt and single mask from image space into feature space and treats the feature distance as semantic-", + "bbox": [ + 89, + 635, + 482, + 891 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "similarity. Prompt, mask and feature are interacted adaptively for different downstream task in tuning-free manner with only one prompt or zero shot.", + "bbox": [ + 511, + 90, + 903, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Due to the high acquisition cost and label difficulty, current hyperspectral datasets only have a few images [45, 69], far from the requirements to train the promptable HyperFree. To conquer this problem, we developed the Hyper-Seg data engine, in addition to utilizing well-known multispectral datasets (e.g., fMoW [5] and SpaceNet [50, 51, 59]). This engine automatically generates a large-scale annotated HSI dataset, consisting of nearly 50,000 pairs of HSIs and their corresponding segmentation masks. The original data are sourced from AVIRIS airborne hyperspectral sensors, with a spatial size of $512 \\times 512$ pixels and 224 spectral channels spanning from 400 to $2500\\mathrm{nm}$ at a $10\\mathrm{nm}$ spectral resolution. Unlike other RSFMs trained on unannotated satellite hyperspectral imagery with a $30\\mathrm{m}$ spatial resolution [54], the Hyper-Seg data engine features higher spatial resolution (0.6 to $5.0\\mathrm{m}$ ), along with annotations of 15.44 million masks. It is currently the largest known hyperspectral dataset with high spatial resolution.", + "bbox": [ + 511, + 136, + 903, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To demonstrate the model capabilities, we evaluated HyperFree directly in a P-E manner on 11 datasets across 5 segmentation-related tasks, including multi-class classification [21], one-class classification [66], target detection [65], anomaly detection [47], and change detection [26]. Additionally, we verified HyperFree in a P-T manner on 14 datasets spanning 8 tasks, with additional tasks including hyperspectral denoising [27], hyperspectral unmixing [37], and hyperspectral object tracking [61]. The experimental results show that HyperFree achieves comparable accuracy to state-of-the-art models in a tuning-free manner (using just one prompt), and outperforms most models after tuning.", + "bbox": [ + 511, + 407, + 903, + 603 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as:", + "bbox": [ + 532, + 604, + 805, + 618 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose the first tuning-free hyperspectral foundation model, which can process any hyperspectral image in different tasks with promptable or zero-shot manner.", + "- A weight dictionary that spans the full spectrum, enabling dynamic generation of the embedding layer dynamically according to input wavelengths.", + "- We propose to map both prompts and masks into feature space to identify multiple semantic-aware masks for one prompt, where different interaction workflows are designed for each downstream tasks.", + "- We built the Hyper-Seg data engine to train the HyperFree model and tested it on 11 datasets from 5 tasks in tuning-free manner, 14 datasets from 8 tasks in tuning manner as an extensive experiment." + ], + "bbox": [ + 513, + 619, + 903, + 830 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 844, + 653, + 859 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Hyperspectral image processing. There are diverse tasks to process such precise observation with different targets", + "bbox": [ + 511, + 869, + 903, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23049", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "[31]. (a) Classification: Assigning a specific class to each pixel [21]. (b) One class classification: Discriminating the pixels of target class with labeled binary map [33]. (c) Target detection. Discriminating the pixels of target class with labeled spectrum. (d) Anomaly detection. Discriminating the pixels spectrally deviating from the background [24]. (e) Change detection: Discriminating the changing pixels between two time-steps [26]. Since the output of the above tasks is all pixel-level segmentation maps, it is feasible to unify them using the proposed HyperFree model in tuning-free manner.", + "bbox": [ + 89, + 90, + 480, + 255 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Foundation models. Powered by the transformer architectures, foundation models are characterized by large-scale property in model size and pre-training data, and the transferring ability for downstream tasks [60]. In remote sensing community, some foundation models for multi-spectral images have already been proposed such as SatMAE [7], SepctralEarth [1] and SpectralGPT [13], which trains large ViT-like models [9] with the MAE strategy [12] and massive open-source multi-spectral images (13 channels). In contrast, hyperspectral images have a high acquiring cost [56] and make the large-scale pre-training stage more difficult. A recent study broke the situation by building a largescale hyperspectral dataset HyperGlobal-450K and trained a foundation model called HyperSigma with MAE strategy as well [54]. Our work differs from HyperSigma in two aspects: (i) HyperSigma accepts fixed number of image channels, while HyperFree can deal with any number of channels directly. (ii) HyperSigma needs to be fine-tuned for each downstream task. Differently, HyperFree has ability to process different tasks in fine-tuning free manner.", + "bbox": [ + 91, + 272, + 482, + 574 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Prompt engineering (P-E). Different from the pretraining and then fine-tuning paradigm, P-E can utilize pretrained foundation models to complete downstream tasks directly with prompts [34]. For pre-trained large language models, prompts are mostly hand-crafted text to guide the model generate response for specific task [34, 42]. In computer vision community, SAM follows the same spirit and designs diverse prompts (e.g., point and box) to achieve few-shot transferring performance [17]. Our work extends the visual prompt engineering in hyperspectral data with varying image channels and supports five tasks in tuning-free manner with designed prompt-mask-feature interaction.", + "bbox": [ + 89, + 575, + 482, + 770 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Hyper-Seg Data Engine", + "text_level": 1, + "bbox": [ + 89, + 784, + 318, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Hyperspectral image processing always refers to various dense tasks and a large-scale segmentation dataset is needed to apply the P-E. However, most hyperspectral segmentation datasets only have a few images due to the high acquisition and label cost, which is quite different from the million-scale dataset with natural images such as SA-1B", + "bbox": [ + 89, + 810, + 480, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "[17]. To tackle this, we built a data engine called HyperSeg to generate segmented masks automatically for spectral images and expand the data scale. Hyper-Seg engine firstly separates the spectral image into several groups with three channels according to the prior of key channels position. Each group is then segmented by the SAM-H model [17] and the results are combined by NMS operation to eliminate redundancy and output the final segmentation map. Figure 2 shows the overall workflow, where the nine key channels are selected according to the expert knowledge as in Supp.1.1. Considering the high spatial resolution property of airborne platform, we collected 41946 hyperspectral images from AVIRIS airborne sensor and applied the Hyper-Seg Engine. Besides, we also processed two existing multispectral datasets (fMoW [5] and SpaceNet [50, 51, 59]) to increase the scale further. Compared to the traditional hyperspectral annotation methods which refer to high-resolution RGB imagery, our data engine can utilize more spectral information and obtain a high-precision segmentation mask without labor cost.", + "bbox": [ + 511, + 90, + 903, + 391 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1 shows the comparable statistical information of training datasets between existing RSFMs and proposed HyperFree. Most models only concentrate on the RGB and multispectral images with channel number 3 13. Although the recent model HyperSigma [54] has collected a large-scale hyperspectral images, the spatial resolution (30m/pixel) is largely lower than us (0.6 5m/pixel), and the dataset scale (considering image size and number together) is relatively small. With the designed Hyper-Seg data engine, constructed dataset is the only one with segmentation masks supporting the promptable training. More detailed statistical information of our generated dataset is given in Supp.1.2. The final generated dataset has nearly 150k images and 15.44million masks. We name both the data engine and generated dataset as the Hyper-Seg for simplicity.", + "bbox": [ + 511, + 393, + 903, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. HyperFree Foundation Model", + "text_level": 1, + "bbox": [ + 511, + 632, + 794, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "HyperFree has two customized components upon the metaarchitecture of SAM including channel-adaptive embedding and prompt-mask-feature (PMF) interaction as in Figure 3. Channel-adaptive embedding constructs the embedding layer dynamically according to varying input wavelength, and outputs tokens in same representation space. After the promptable training stage, PMF interaction discriminates the semantic-aware masks for downstream tasks by mapping both prompt and mask into feature space and treating feature distance as semantic similarity.", + "bbox": [ + 511, + 657, + 903, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Channel-adaptive Embedding", + "text_level": 1, + "bbox": [ + 511, + 818, + 779, + 834 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To make HyperFree process any hyperspectral image, channel-adaptive embedding layer is firstly designed to map input image with varying channels into tokens with fixed length, inspired by the tokenizer in NLP community [32].", + "bbox": [ + 511, + 840, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23050", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/428559adf50a8595acfdcb565789207c51467f6398d40bb9b5b971b5ed5f292f.jpg", + "image_caption": [ + "Figure 2. Hyper-Seg Data Engine automatically generates a large-scale dataset comprising nearly 41,900 pairs of airborne HSIs and their corresponding segmentation masks, enabling HyperFree to effectively learn discriminative spectral features." + ], + "image_footnote": [], + "bbox": [ + 212, + 92, + 782, + 287 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/07e945a48bc66c09e257ea552910b4191cb89b59c81457f5fbe71039e59c71a5.jpg", + "table_caption": [ + "Table 1. Comparison of training data used in current RSFMs. The primary features of the proposed HyperFree model include prompt engineering and the use of high-resolution annotated hyperspectral (HRS) imagery." + ], + "table_footnote": [ + "* P-T: Pretraining and Tuning; P-E: Prompt Engineering.", + "* HSR: High Spatial Resolution; MSR: Medium Spatial Resolution; LSR: Low Spatial Resolution.", + "* RGB: Red-Green-Blue Imagery; M: Multispectral Imagery; H: Hyperspectral Imagery" + ], + "table_body": "
DatasetFoundation ModelSpatial Resolution (m)Spectral Resolution (nm)Channel NumberImage SizeImage NumberMain ModalityTraining Paradigm
RGBRVSA [53]0.5 ~ 153\\3110 ~ 316721.00 × 10^6H/M/LSR-RGBP-T
RingMo [48]0.1 ~ 30\\34482.10 × 10^6H/MSR-RGBP-T
ScaleMAE [39]0.31 ~ 3.7\\3\\3.64 × 10^6HSR-RGBP-T
MultispectralSatMAE [6]1015 ~ 18013≈ 45 × 607.13 × 10^5MSR-MP-T
SkySense [11]10/2015 ~ 18010\\\\MSR-MP-T
SpectralGPT [14]10/20/6015 ~ 1801245 ~ 1201.47 × 10^6MSR-MP-T
HyperspectralHyperSigma [54]305/6.5/10150/175644.47 × 10^5LSR-HP-T
HyperFree0.6 ~ 5.0102245124.19 × 10^4HSR-HP-E
", + "bbox": [ + 91, + 383, + 897, + 542 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Tokenizers rely on a pre-defined dictionary to break down a string of text into numerical tokens, where the dictionary has stored the numerical vectors of varied-length words or phrases. Vision transformers always use a convolutional layer to generate patch tokens and what if we construct a layer weight dictionary treating each wavelength as a word? Given the input meta data of wavelengths, we can just loop up the corresponding weights and construct the layer dynamically without re-training it.", + "bbox": [ + 88, + 606, + 482, + 742 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The weight dictionary is built depending on the wavelengths to be processed and the expected token dimension. Since hyperspectral imaging mostly captures the spectral signals from $400\\mathrm{nm}$ to $2500\\mathrm{nm}$ and the $10\\mathrm{nm}$ is of high spectral resolution currently, 221 index channels are designed as the dictionary keys. Denote the output token as a vector of $\\mathbb{R}^j$ corresponding to the patch size $p\\times p$ , each index channel stores a weight matrix of size $\\mathbb{R}^{p\\times p\\times j}$ , and the built dictionary $\\beta$ has a size of $221\\times p\\times p\\times j$ . Given the input hyperspectral image $\\mathbb{R}^{h\\times w\\times n}$ with wavelengths $\\mathbf{b} =$", + "bbox": [ + 88, + 742, + 482, + 893 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$[b_1, b_2, \\ldots, b_n]$ , the corresponding kernel weight can be dynamically generated by combining the searched weight $\\mathbf{W}$ with size $n \\times p \\times p \\times j$ , where $n$ is flexible for each image. For conciseness, we use function $g$ to represent the mapping process from $\\mathbf{b}$ to $\\mathbf{W}$ as kernel weight, and function $f$ to represent the token generation step as in Equation (1).", + "bbox": [ + 511, + 606, + 906, + 699 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf (\\mathbf {X} \\mid \\mathbf {b}, \\beta) = \\operatorname {C o n v} _ {\\mathbf {W}} (\\mathbf {X}) \\quad \\text {w h e r e} \\quad \\mathbf {W} = g (\\mathbf {b}, \\beta) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 534, + 722, + 906, + 741 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the past few decades, many satellites have been launched equipped with spectral sensors, where the key channels are selected by professional teams and practical experience. To utilize these successful priors, we further extract visual tokens in two parallel branches with dynamically constructed layers: one to process key channels and the other to process the split cubes, as shown in Figure 3. The input image $\\mathbf{X}$ is first divided into key channels $\\mathbf{X}_{\\mathbf{k}}$ and the cubes $\\mathbf{X}_{\\mathbf{c}}$ between them, where the key wavelengths are selected by referring to famous spectral sensors (detailed", + "bbox": [ + 509, + 750, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23051", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/479ad1de12bb655da0da7ead4673adc2cf106098b2345ef3da7ea96e200b9d60.jpg", + "image_caption": [ + "(a) Full spectrum Promptable Training" + ], + "image_footnote": [], + "bbox": [ + 127, + 88, + 509, + 359 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4fdc8d8960c1438dd44057abbb04840e2830b01dbba8002eec3d5c77c2e1cc82.jpg", + "image_caption": [ + "(b) Prompt-mask-feature Interactive Inferring", + "Figure 3. (a) Full spectrum Promptable Training. Based on the meta-architecture of promptable segmentation, channel-adaptive embedding layer is developed to process any-channel input, which relies on a learnable weight dictionary and generates the embedding layer dynamically according to the input wavelengths. With the constructed Hyper-Seg dataset, HyperFree is trained with randomly selected channels from full spectrum and point prompts in each iteration. (b) Prompt-mask-feature Interactive Inferring. Keep the trained HyperFree frozen, different PMF interaction workflows are designed to complete the downstream tasks in promptable or zero-shot manner, where we use feature distance to measure the semantic consistency. An interaction instantiation for classification task is given above as an example." + ], + "image_footnote": [], + "bbox": [ + 521, + 87, + 869, + 358 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "in Supp.2). We set two different dictionaries $\\beta_{\\mathbf{k}}$ and $\\beta_{\\mathbf{c}}$ to process $\\mathbf{X}_{\\mathbf{k}}$ and $\\mathbf{X}_{\\mathbf{c}}$ respectively, and the final output tokens $\\mathbf{T} \\in \\mathbb{R}^{(h/p) \\times (w/p) \\times j}$ are computed as the element-wise sum, as shown in Equation (2).", + "bbox": [ + 89, + 488, + 485, + 550 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {T} = f \\left(\\mathbf {X} _ {\\mathbf {k}} \\mid \\mathbf {b} _ {\\mathbf {k}}, \\beta_ {\\mathbf {k}}\\right) + f \\left(\\mathbf {X} _ {\\mathbf {c}} \\mid \\mathbf {b} _ {\\mathbf {c}}, \\beta_ {\\mathbf {c}}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 566, + 482, + 583 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Full-spectrum Promptable Training", + "text_level": 1, + "bbox": [ + 89, + 593, + 403, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The prompt engineering paradigm is adopted to complete downstream tasks directly with designed prompts. Hyper-Free is thus trained in a promptable manner on constructed large-scale Hyper-Seg dataset, where the prompts are randomly chosen at each iteration and the model is expected to segment correctly the corresponding masks. Since hyperspectral image processing tasks are almost conducted at the pixel level, we choose point prompts only at the training stage. Focal loss $L_{f}$ and Dice loss $L_{d}$ are weighted as [17] to supervise the training as Equation (3).", + "bbox": [ + 89, + 616, + 483, + 767 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL = 2 0 L _ {f} + L _ {d} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 230, + 785, + 482, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To train the channel-adaptive embedding layer effectively, we select the subset of channels randomly from the full-spectrum range and process it with corresponding weights only in $\\beta_{\\mathbf{k}}$ and $\\beta_{\\mathrm{c}}$ , which helps the learned dictionary being wavelength-aware and encode the varied channel composition into a unified token space.", + "bbox": [ + 89, + 810, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Prompt-mask-feature (PMF) Interactive Inferring", + "text_level": 1, + "bbox": [ + 511, + 488, + 898, + 505 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After training process, the model can process unseen hyperspectral image with varied channels and segment the corresponding mask given any point prompt. However, many downstream tasks need to segment out all the masks of same semantic category rather than a single mask such as hyperspectral classification and target detection. We tackle this problem by connecting the prompt, mask and feature together in feature space, where prompt tells the model an object reference in task, mask outputs accurate location information, and feature reflects semantic information. In unified feature space, the cosine distance of features is used to measure the semantic similarity.", + "bbox": [ + 511, + 537, + 906, + 719 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Before diving into the specific workflows, we introduce two basic interaction modes with a simple situation. Set one point $(x,y)$ as input prompt for image $\\mathbf{X}$ , the corresponding feature cube $\\mathbf{D}$ and all the generated masks $\\mathbf{M} \\in \\mathbb{R}^{h \\times w \\times k}$ from the trained model, where $\\mathbf{M}$ has $k$ binary maps and each represents a single mask. The first interaction mode maps the prompt $(x,y)$ into a feature vector $\\mathbf{d}_{(x,y)} \\in \\mathbb{R}^j$ as in Equation (4), where the corresponding valid and smallest mask in $\\mathbf{M}$ is used to localize the features in $\\mathbf{D}$ , and the mean feature is treated as $\\mathbf{d}_{(x,y)}$ for the given prompt. This interaction transforms the prompt from image space to feature space, and makes it possible to explore the semantic", + "bbox": [ + 511, + 719, + 908, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23052", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "similarity.", + "bbox": [ + 89, + 90, + 161, + 104 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\text {M o d e} (x, y) \\mid \\mathbf {M}, \\mathbf {D} \\rightarrow \\mathbf {d} _ {(x, y)} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 183, + 114, + 482, + 132 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The second interaction mode selects multiple the masks in $\\mathbf{M}$ , which have similar semantic information with $\\mathbf{d}_{(x,y)}$ as in Equation (5). The selected masks are represented as $\\mathbf{M}_d$ and $\\mathbf{M}_d \\subseteq \\mathbf{M}$ . Specifically, we compute a feature representation for each single mask by localizing in $\\mathbf{D}$ and computing the average, and the cosine similarity between them and $\\mathbf{d}_{(x,y)}$ are treated as the semantic similarity. With some hand-designed thresholds $\\tau$ , $\\mathbf{M}_d$ can be finally segmented. $\\tau$ represents different meanings according to the specific task.", + "bbox": [ + 89, + 140, + 483, + 290 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\text {M o d e} 2: \\left(\\mathbf {d} _ {(x, y)} | \\mathbf {M}, \\mathbf {D}, \\tau\\right)\\rightarrow \\mathbf {M} _ {d} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 306, + 482, + 323 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "By using both interaction modes adaptively, five hyperspectral tasks can be completed with the trained promptable model directly. (a) Hyperspectral classification (HC). Each category needs at least one prompt, and uses Mode 1 and Mode 2 in turn for each category. For a common closed-set setting, $\\tau$ is not necessary since we can use the minimum feature distance between each mask representation and all the prompt representations to decide its category. Figure 3 (b) shows an exemplified workflow. (b) Hyperspectral one-class classification (HOCC). HOCC is similar to H but the difference lies in that HOCC only needs prompts the target category, and its $\\tau$ is instantiated as a class pr parameter to control the mask ratio. (c) Hyperspectral taget detection (HTD). Compared to HOCC, HTD acce the target spectra and it needs to be changed to a pron by selecting the pixel with the smallest cosine distance. is instantiated as a distance threshold to convert the dens map to a binary map. (d) Hyperspectral anomaly detion (HAD). HAD is an unsupervised task and there is $\\mathbf{d}_{(x,y)}$ for HAD. Since HAD is very similar to unsupervis target detection, we use M only and filter out large objec according to mask area to output the final anomaly map.", + "bbox": [ + 89, + 328, + 483, + 660 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Hyperspectral change detection (HCD). Different from the prior tasks, HCD accepts two registered and temporal images to identify the changing locations. The HCD task utilizes both interaction modes for the bi-temporal images. Mode 1 processes the first image to get representation $\\mathbf{d}_{(x,y)}$", + "bbox": [ + 89, + 660, + 483, + 736 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ab5712bd45ec8b0957ff7c3f262bb8419239244063281dbb3c09d2cb7787a65c.jpg", + "table_caption": [ + "Table 2. Instantiation summary of prompts, meaning of $\\tau$ , and interaction modes when HyperFree performs different tasks." + ], + "table_footnote": [], + "table_body": "
TaskPrompt TypeHyperparameter τInteraction Mode
HCOne Point Per CategoriesNo τMode 1, Mode 2
HOCCTarget Category PointClass PriorMode 1, Mode 2
HTDTarget SpectraSimilarity ThresholdMode 1, Mode 2
HADNo PromptMasks Area ThresholdMode 1
HCDNo PromptSimilarity ThresholdMode 1, Mode 2
", + "bbox": [ + 91, + 782, + 483, + 897 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c8766bfd9c3a124d04b58db1c422d193da50f8ee338e2709fe18ccf539f45221.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 540, + 92, + 883, + 416 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8aa3a5b03f95cc7bde125587c9649d741960c2d44980215513318b6be995fecf.jpg", + "image_caption": [ + "Figure 4. Exemplified comparison results for five tuning-free tasks", + "Figure 5. In tuning-free manner, HyperFree (1 prompt) can achieve comparable results with specialized models (5 shots) on 5 tasks and 11 datasets." + ], + "image_footnote": [], + "bbox": [ + 535, + 460, + 883, + 675 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "for each mask, which is input to Mode 2 and compared with the corresponding mask representation with $\\mathbf{M}$ and $\\mathbf{D}$ of the second image. Once the temporal feature distance exceeds the preset $\\tau$ , a change is considered to have happened. The interaction workflow is summarized in Table 2.", + "bbox": [ + 511, + 734, + 906, + 810 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Results", + "text_level": 1, + "bbox": [ + 513, + 823, + 607, + 838 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 848, + 689, + 864 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate the proposed HyperFree on five downstream hyperspectral tasks including HC, HOCC, HTD, HAD and", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23053", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2ad382a0e9a889978bf11da8291f10a1af2ee27d641f466c5387a4e8aef74680.jpg", + "table_caption": [ + "Table 3. Quantitative comparison results on five hyperspectral tasks, where blue numbers indicate the metric ranking of HyperFree. Comparison methods were all trained on each dataset while HyperFree processed the datasets in tuning-free manner." + ], + "table_footnote": [], + "table_body": "
HC Task
LongKou (270 channels)HanChuan (274 channels)
MetricsSVM [30] (5 shot)HybridSN [41] (5 shot)FCN [52] (5 shot)FPGA [68] (5 shot)SSFTT [46] (5 shot)MambaHSI [23] (5 shot)HyperFree (1 prompt)SVM [30] (5 shot)HybridSN [41] (5 shot)FCN [52] (5 shot)FPGA [68] (5 shot)SSFTT [46] (5 shot)MambaHSI [23] (5 shot)HyperFree (1 prompt)
OA82.7748.7886.6791.1889.6692.6593.39(1st)52.6847.7555.5571.4764.8673.3375.47(1st)
AA74.0261.3785.6088.3587.9692.5788.03(3rd)47.7646.1759.7272.0961.2269.3362.35(2nd)
KA78.0435.7282.3088.6687.9590.0091.41(1st)46.8541.3150.1867.5859.6569.1071.56(1st)
HOCC Task
HongHu (270 channels)XiongAn (256 channels)
MetricsOCSVM [43] (5 shot)nnPU [18] (5 shot)BSVM [33] (5 shot)PAN [15] (5 shot)OC Loss [67] (5 shot)T-HOneCls [66] (5 shot)HyperFree (1 prompt)OCSVM [43] (5 shot)nnPU [18] (5 shot)BSVM [33] (5 shot)PAN [15] (5 shot)OC Loss [67] (5 shot)T-HOneCls [66] (5 shot)HyperFree (1 prompt)
F26.3319.1334.8263.6954.7355.9772.52(1st)18.311.7626.3046.3443.0841.3452.50(1st)
P56.4319.7250.7780.0058.2746.5268.61(2nd)39.832.8523.8247.1347.5032.8768.21(1st)
R24.0218.5845.2964.2754.3492.3578.06(2nd)16.081.9857.8353.3247.6160.3865.03(1st)
HTD Task
Airport (205 channels)Cri (46 channels)
MetricsACE [19] (1 shot)CEM [2] (1 shot)GLRT [28] (1 shot)MF [29] (1 shot)HTD-IRN [44] (1 shot)TSTTD [16] (1 shot)HyperFree (1 prompt)ACE [19] (1 shot)CEM [2] (1 shot)GLRT [28] (1 shot)MF [29] (1 shot)HTD-IRN [44] (1 shot)TSTTD [16] (1 shot)HyperFree (1 prompt)
AUC(D,F)0.97940.96030.98010.99160.97450.99290.9806(3rd)0.97350.98930.97370.98910.99750.99870.9791(5th)
AUCODP1.58531.28291.57981.69681.44841.65921.4667(6th)1.20151.45061.20001.45751.39951.61031.4670(2nd)
HAD Task
Beach1 (188 channels)Beach2 (193 channels)
MetricsRXD [40]CRD [22]ADLR [36]LRASR [63]Auto-AD [55]TDD [20]HyperFree (zero shot)RXD [40]CRD [22]ADLR [36]LRASR [63]Auto-AD [55]TDD [20]HyperFree (zero shot)
AUC(D,F)0.98150.94710.45150.74610.95740.98420.9730(3rd)0.90900.85440.79760.82250.94850.96270.9566(2nd)
AUCODP1.25570.97850.56100.85261.12731.13831.9191(1st)1.01770.86700.90640.82801.00971.16881.8697(1st)
HCD Task
River (154 channels)Hermiston (198 channels)
MetricsFC-EF [8] (5 shot)FC-Sc [8] (5 shot)FC-Sd [8] (5 shot)ML-EDAN [35] (5 shot)BIT [3] (5 shot)SST-Former [57] (5 shot)HyperFree (zero shot)FC-EF [8] (5 shot)FC-Sc [8] (5 shot)FC-Sd [8] (5 shot)ML-EDAN [35] (5 shot)BIT [3] (5 shot)SST-Former [57] (5 shot)HyperFree (zero shot)
IoU0.41680.45220.45340.39150.21260.40960.3649(6th)0.37290.37760.48730.32520.52570.53610.5851(1st)
F0.58840.62280.62390.56280.35070.58120.5347(6th)0.54320.54820.65.520.49080.68910.69800.7382(1st)
", + "bbox": [ + 91, + 122, + 908, + 494 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9a200ed64e1c1d13f3364e0573a3884791a5491e2ec04c87dd80a7f62c318031.jpg", + "table_caption": [ + "Table 4. Compared results with SAM using the identical PMF interaction workflows." + ], + "table_footnote": [], + "table_body": "
HCHOCCHTDHADHCD
OAAAKAFPrecisionRecallAUC(D,F)AUCODPAUC(D,F)AUCODPFIoU
SAM+PMF Interaction66.6272.9959.4134.1330.3539.470.79791.10140.96151.88440.59960.4282
HyperFree93.4085.6691.4272.5268.6178.060.97911.46700.95661.86970.73820.5851
", + "bbox": [ + 91, + 527, + 906, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "HCD tasks as in Section 4.3. A total of 11 datasets were used in the evaluation (detailed in Supp.3), and each task had six comparison methods, including both classical and state-of-the-art ones. Since there is no method that can directly complete all five tasks, all comparative methods are tested after being trained on each dataset with five shots per class, while the parameters of HyperFree are fixed for all datasets with 1 prompt or zero-shot as in Table 2. We use ViT-base as backbone to conduct the experiments.", + "bbox": [ + 89, + 606, + 483, + 742 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "HyperFree was trained on the constructed Hyper-Seg dataset using 8 A100 GPUs for approximately 10 hours, which was based on the meta-architecture and parameters of SAM. We use AdamW optimizer with initial learning rate 8e-5 and batch size 16. To help learning the channel-adaptive embedding, we randomly select the image channels at each iteration and match it with corresponding wavelengths. 16 masks and 1 2 points per mask are randomly selected for each input image to train", + "bbox": [ + 89, + 763, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the promptable segmentation.", + "bbox": [ + 511, + 607, + 710, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Main Results", + "text_level": 1, + "bbox": [ + 513, + 622, + 643, + 635 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Most tuning-free quantitative results are reported in Table 3 (1 dataset in supplement due to the space limit) and Figure 5. All the comparative methods are trained with 5 shot for each class and HyperFree processes each dataset directly without tuning. For HC, HOCC and HTD, only one point prompt is provided for each class. HAD and HCD does not need any prompt and deal with images in zero-shot manner. Table 3 shows that HyperFree can surpass most of the specialized models with frozen parameters. On both HC and HCD tasks, HyperFree even outperformed previous SOTA results, including an improvement of about 12 points in $\\mathbf{F}_1$ on HC and about 0.5 points in IoU on HCD task.", + "bbox": [ + 509, + 643, + 906, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We performed qualitative comparison on five exemplified images in Figure 4, where one SOTA model have been visualized for each task comparison. Obviously, the result maps of HyperFree are more complete and smoother than the ones of specialized models. In HC, HOCC, HTD and", + "bbox": [ + 511, + 825, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23054", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e45c2ce6e4623276cf6fba6829e9065536edf0aa9a43cc18adf8188a0a360b94.jpg", + "table_caption": [ + "Table 5. Ablation results on the usage of the learned weight dictionary." + ], + "table_footnote": [], + "table_body": "
Correct WavelengthsβkβcHCHOCCHTDHADHCD
OAAAKAF1PrecisionRecallAUC(D,F)AUCODPAUC(D,F)AUCODPF1IoU
✓✓×73.1864.2365.8565.6247.3064.560.97671.46440.94881.84650.57930.4078
✓×✓66.6371.5259.4463.6258.6178.440.97681.46400.95891.87680.70220.5411
×✓✓75.2538.9866.4457.3064.0954.830.95951.42730.95571.86780.68320.5188
✓✓✓93.4085.6691.4272.5268.6178.060.97911.46700.95661.86970.73820.5851
", + "bbox": [ + 91, + 108, + 903, + 186 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a6d5d6b9a88aab5b71db603328a561f0ef200b15e536b0c55767951a47248281.jpg", + "image_caption": [ + "(a) HyperSigma features" + ], + "image_footnote": [], + "bbox": [ + 91, + 208, + 282, + 330 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3b07dd51824e56815c3a355ffbb7bba73269f2f710e0873a36aca930a8712141.jpg", + "image_caption": [ + "(b) HyperFree features" + ], + "image_footnote": [], + "bbox": [ + 287, + 208, + 480, + 330 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a9efefd3751bea9223204b28fa6a9e2fb9ad4f64cfaadf5bf1fa51d6a1d30c13.jpg", + "image_caption": [ + "Figure 6. t-SNE visualizations of HyperSigma [54] and HyperFree features on WHU-Hi Longkou dataset. We randomly select 1000 pixels for each class.", + "Figure 7. In tuning manner, HyperFree can achieve overall better results with specialized models on 8 tasks and 14 datasets." + ], + "image_footnote": [], + "bbox": [ + 114, + 395, + 460, + 601 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "HAD tasks, HyperFree can generate basically the same result as GT, which are difficulty for existing models.", + "bbox": [ + 89, + 642, + 482, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Analysis Studies", + "text_level": 1, + "bbox": [ + 89, + 681, + 241, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "HyperFree vs SAM. Proposed prompt-mask-feature interaction strategy adapts the pre-trained HyperFree to various downstream tasks. We applied the same strategy to SAM to verify the effectiveness of training on hyperspectral images and reported the results in Table 4. The metrics on most tasks have a decline of about $10\\sim 30$ points, implying a failed processing. We realize it is not a fair comparison and only use it to reflect the difference of hyperspectral data processing.", + "bbox": [ + 89, + 703, + 482, + 839 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "HyperFree vs HyperSigma. HyperSigma is the only open-sourced hyperspectral foundation model [54, 62], which adopts the pre-training and tuning paradigm. Thus, we cannot compare HyperFree with it in tuning-free set-", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ting directly and compare the t-SNE visualization in Figure 6(a) and Figure 6(b). In contrast, proposed HyperFree has a more compact intra-class distance obviously, supporting the tuning-free workflow of PMF interaction.", + "bbox": [ + 511, + 210, + 903, + 270 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Full-Spectrum Weight Dictionary. To verify stored weights are wavelength-aware, we conducted related ablation studies in Table 5. In HC, HOCC and HCD tasks, a drop in accuracy of about $5\\sim 40$ points can be observed if we shuffled the correct wavelength order randomly, showing the wavelength-aware property of learned dictionary. Besides, we proved the built two encoding branches with dictionaries $\\beta_{\\mathbf{k}}$ and $\\beta_{\\mathrm{c}}$ can benefit each other effectively.", + "bbox": [ + 511, + 271, + 905, + 392 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Tuning HyperFree. As an extensive experiment, we have also tested the tuning performance of HyperFree on 8 tasks and 14 datasets in Figure 7, which further included hyperspectral denosing (HD) [27], object tracking (HOT) [61] and unmxing (HU) tasks [37]. The tuning HyperFree has surpassed the previous SOTA models on nearly all tasks.", + "bbox": [ + 511, + 392, + 903, + 482 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "More analysis. For more analytical experiments, please see the supplementary material. (i) Supp.4.1: Complete qualitative comparison of 11 datasets in tuning-free manner. (ii) Supp.4.2: Complete quantitative and qualitative comparison of 14 datasets in tuning manner. (iii) Supp.4.3: Sensitivity analysis about prompt number, prompt location and the hyperparameter $\\tau$ .", + "bbox": [ + 511, + 482, + 905, + 589 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 648, + 637, + 664 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Hyperspectral images can provide rich spectral information and are widely used in resource monitoring and defense fields. However, its unique properties increase the difficulty of foundation model development such as varied-length channels, high sensitivity to imaging conditions and the acquirement difficulty, making the mainstream P-T being low cost-effectiveness due to the image-by-image tuning. Our work tackles this problem by converting to adapt the visual P-E [17]. Differently, we design a learnable weight dictionary covering full-spectrum to generate the channel-adaptive embeddings and PMF interaction to generate semantic-aware masks for different tasks. We hope HyperFree can speed up the era of \"GPT moment\" for hyperspectral intelligent processing.", + "bbox": [ + 511, + 688, + 906, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23055", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 96, + 90, + 256, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 172, + 187, + 188 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Nassim Ait Ali Braham, Conrad M Albrecht, Julien Mairal, Jocelyn Chanussot, Yi Wang, and Xiao Xiang Zhu. Spectral earth: Training hyperspectral foundation models at scale. arXiv preprint arXiv:2408.08447, 2024. 3", + "[2] Chein-I Chang. Constrained energy minimization (cem) for hyperspectral target detection: Theory and generalizations. IEEE Transactions on Geoscience and Remote Sensing, 2024. 7", + "[3] Hao Chen, Zipeng Qi, and Zhenwei Shi. Remote sensing image change detection with transformers. IEEE Transactions on Geoscience and Remote Sensing, 60:1-14, 2021. 7", + "[4] Keyan Chen, Chenyang Liu, Hao Chen, Haotian Zhang, Wenyuan Li, Zhengxia Zou, and Zhenwei Shi. Rsprompter: Learning to prompt for remote sensing instance segmentation based on visual foundation model. IEEE Transactions on Geoscience and Remote Sensing, 2024. 1, 2", + "[5] Gordon Christie, Neil Fendley, James Wilson, and Ryan Mukherjee. Functional map of the world. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6172-6180, 2018. 2, 3", + "[6] Yezhen Cong, Samar Khanna, Chenlin Meng, Patrick Liu, Erik Rozi, Yutong He, Marshall Burke, David Lobell, and Stefano Ermon. Satmae: Pre-training transformers for temporal and multi-spectral satellite imagery. Advances in Neural Information Processing Systems, 35:197-211, 2022. 4", + "[7] Yezhen Cong, Samar Khanna, Chenlin Meng, Patrick Liu, Erik Rozi, Yutong He, Marshall Burke, David B. Lobell, and Stefano Ermon. Satmae: pre-training transformers for temporal and multi-spectral satellite imagery. In Proceedings of the 36th International Conference on Neural Information Processing Systems, Red Hook, NY, USA, 2024. Curran Associates Inc. 1, 2, 3", + "[8] Rodrigo Caye Daudt, Bertr Le Saux, and Alexandre Boulch. Fully convolutional siamese networks for change detection. In 2018 25th IEEE international conference on image processing (ICIP), pages 4063-4067. IEEE, 2018. 7", + "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 3", + "[10] Alexander FH Goetz, Gregg Vane, Jerry E Solomon, and Barrett N Rock. Imaging spectrometry for earth remote sensing. science, 228(4704):1147-1153, 1985. 2", + "[11] Xin Guo, Jiangwei Lao, Bo Dang, Yingying Zhang, Lei Yu, Lixiang Ru, Liheng Zhong, Ziyuan Huang, Kang Wu, Dingxiang Hu, et al. Skysense: A multi-modal remote sensing foundation model towards universal interpretation for" + ], + "bbox": [ + 93, + 198, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "This work was supported by the National Natural Science Foundation of China under Grant No. 42325105 and Grant 424B2010.", + "earth observation imagery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27672-27683, 2024. 4", + "[12] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 2, 3", + "[13] Danfeng Hong, Bing Zhang, Xuyang Li, Yuxuan Li, Chenyu Li, Jing Yao, Naoto Yokoya, Hao Li, Pedram Ghamisi, Xiuping Jia, Antonio Plaza, Paolo Gamba, Jon Atli Benediktsson, and Jocelyn Chanussot. Spectralgpt: Spectral remote sensing foundation model. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(8):5227-5244, 2024. 1, 2, 3", + "[14] Danfeng Hong, Bing Zhang, Xuyang Li, Yuxuan Li, Chenyu Li, Jing Yao, Naoto Yokoya, Hao Li, Pedram Ghamisi, Xi-uping Jia, et al. Spectralgpt: Spectral remote sensing foundation model. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 4", + "[15] Wenpeng Hu, Ran Le, Bing Liu, Feng Ji, Jinwen Ma, Dongyan Zhao, and Rui Yan. Predictive adversarial learning from positive and unlabeled data. In Proceedings of the AAAI conference on artificial intelligence, pages 7806-7814, 2021. 7", + "[16] Jinyue Jiao, Zhiqiang Gong, and Ping Zhong. Triplet spectral-wise transformer network for hyperspectral target detection. IEEE Transactions on Geoscience and Remote Sensing, 2023. 7", + "[17] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 2, 3, 5, 8", + "[18] Ryuichi Kiryo, Gang Niu, Martinus C Du Plessis, and Masashi Sugiyama. Positive-unlabeled learning with nonnegative risk estimator. Advances in neural information processing systems, 30, 2017. 7", + "[19] Shawn Kraut, Louis L Scharf, and L Todd McWhorter. Adaptive subspace detectors. IEEE Transactions on signal processing, 49(1):1-16, 2001. 7", + "[20] Jingtao Li, Xinyu Wang, Shaoyu Wang, Hengwei Zhao, and Yanfei Zhong. One step detection paradigm for hyperspectral anomaly detection via spectral deviation relationship learning. IEEE Transactions on Geoscience and Remote Sensing, 2024. 1, 7", + "[21] Jiaojiao Li, Zhiyuan Zhang, Yuzhe Liu, Rui Song, Yunsong Li, and Qian Du. Swformer: Stochastic windows convolutional transformer for hybrid modality hyperspectral classification. IEEE Transactions on Image Processing, 2024. 2, 3", + "[22] Wei Li and Qian Du. Collaborative representation for hyperspectral anomaly detection. IEEE Transactions on geoscience and remote sensing, 53(3):1463-1474, 2014. 7", + "[23] Yapeng Li, Yong Luo, Lefei Zhang, Zengmao Wang, and Bo Du. Mambahsi: Spatial-spectral mamba for hyperspectral" + ], + "bbox": [ + 93, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23056", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "image classification. IEEE Transactions on Geoscience and Remote Sensing, 62:1-16, 2024. 7", + "[24] Jie Lian, Lizhi Wang, He Sun, and Hua Huang. Gthad: Gated transformer for hyperspectral anomaly detection. IEEE Transactions on Neural Networks and Learning Systems, 2024. 3", + "[25] Nanqing Liu, Xun Xu, Yongyi Su, Haojie Zhang, and Heng-Chao Li. Pointsam: Pointly-supervised segment anything model for remote sensing images. arXiv preprint arXiv:2409.13401, 2024. 1, 2", + "[26] Fulin Luo, Tianyuan Zhou, Jiamin Liu, Tan Guo, Xiwen Gong, and Jinchang Ren. Multiscale diff-changed feature fusion network for hyperspectral image change detection. IEEE Transactions on Geoscience and Remote Sensing, 61: 1-13, 2023. 2, 3", + "[27] Zhaozhi Luo, Xinyu Wang, Petri Pellikka, Janne Heiskanen, and Yanfei Zhong. Unsupervised adaptation learning for real multiplatform hyperspectral image denoising. IEEE transactions on cybernetics, 2024. 2, 8", + "[28] Dimitris Manolakis and Gary Shaw. Detection algorithms for hyperspectral imaging applications. IEEE signal processing magazine, 19(1):29-43, 2002. 7", + "[29] Dimitris Manolakis, Eric Truslow, Michael Pieper, Thomas Cooley, and Michael Brueggeman. Detection algorithms in hyperspectral imaging systems: An overview of practical algorithms. IEEE Signal Processing Magazine, 31(1):24-33, 2013. 7", + "[30] Farid Melgani and Lorenzo Bruzzone. Classification of hyperspectral remote sensing images with support vector machines. IEEE Transactions on geoscience and remote sensing, 42(8):1778-1790, 2004. 7", + "[31] Mohammed Abdulmajeed Moharram and Divya Meena Sundaram. Land use and land cover classification with hyperspectral data: A comprehensive review of methods, challenges and future directions. Neurocomputing, 536:90-113, 2023. 1, 3", + "[32] Aleksandar Petrov, Emanuele La Malfa, Philip Torr, and Adel Bibi. Language model tokenizers introduce unfairness between languages. Advances in Neural Information Processing Systems, 36, 2024. 2, 3", + "[33] Rami Piiroinen, Fabian Ewald Fassnacht, Janne Heiskanen, Eduardo Maeda, Benjamin Mack, and Petri Pellikka. Invasive tree species detection in the eastern arc mountains biodiversity hotspot using one class classification. Remote sensing of environment, 218:119-131, 2018. 3, 7", + "[34] Maciej P Polak and Dane Morgan. Extracting accurate materials data from research papers with conversational language models and prompt engineering. Nature Communications, 15(1):1569, 2024. 3", + "[35] Jiahui Qu, Shaoxiong Hou, Wenqian Dong, Yunsong Li, and Weiying Xie. A multilevel encoder-decoder attention network for change detection in hyperspectral images. IEEE Transactions on Geoscience and Remote Sensing, 60:1-13, 2021. 7", + "[36] Ying Qu, Wei Wang, Rui Guo, Bulent Ayhan, Chiman Kwan, Steven Vance, and Hairong Qi. Hyperspectral anomaly detection through spectral unmixing and dictionary-based low" + ], + "bbox": [ + 91, + 90, + 483, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "rank decomposition. IEEE Transactions on Geoscience and Remote Sensing, 56(8):4391-4405, 2018. 7", + "[37] Behnood Rasti, Alexandre Zouaoui, Julien Mairal, and Jocelyn Chanussot. Image processing and machine learning for hyperspectral unmixing: An overview and the hysupp python package. IEEE Transactions on Geoscience and Remote Sensing, 2024. 2, 8", + "[38] C. J. Reed, R. Gupta, S. Li, S. Brockman, C. Funk, B. Clipp, K. Keutzer, S. Candido, M. Uytendaele, and T. Darrell. Scale-mae: A scale-aware masked autoencoder for multiscale geospatial representation learning. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 4065-4076, Los Alamitos, CA, USA, 2023. IEEE Computer Society. 1, 2", + "[39] Colorado J Reed, Ritwik Gupta, Shufan Li, Sarah Brockman, Christopher Funk, Brian Clipp, Kurt Keutzer, Salvatore Candido, Matt Uytendaele, and Trevor Darrell. Scale-mae: A scale-aware masked autoencoder for multiscale geospatial representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4088-4099, 2023. 4", + "[40] Irving S Reed and Xiaoli Yu. Adaptive multiple-band cfar detection of an optical pattern with unknown spectral distribution. IEEE transactions on acoustics, speech, and signal processing, 38(10):1760-1770, 1990. 7", + "[41] Swalpa Kumar Roy, Gopal Krishna, Shiv Ram Dubey, and Bidyut B Chaudhuri. Hybridsn: Exploring 3-d-2-d cnn feature hierarchy for hyperspectral image classification. IEEE Geoscience and Remote Sensing Letters, 17(2):277-281, 2019. 7", + "[42] Pranab Sahoo, Ayush Kumar Singh, Sriparna Saha, Vinija Jain, Samrat Mondal, and Aman Chadha. A systematic survey of prompt engineering in large language models: Techniques and applications. arXiv preprint arXiv:2402.07927, 2024.3", + "[43] Bernhard Scholkopf, Robert C Williamson, Alex Smola, John Shawe-Taylor, and John Platt. Support vector method for novelty detection. Advances in neural information processing systems, 12, 1999. 7", + "[44] Dunbin Shen, Xiaorui Ma, Wenfeng Kong, Jianjun Liu, Jie Wang, and Hongyu Wang. Hyperspectral target detection based on interpretable representation network. IEEE Transactions on Geoscience and Remote Sensing, 2023. 7", + "[45] Hongjun Su, Zhaoyue Wu, Huihui Zhang, and Qian Du. Hyperspectral anomaly detection: A survey. IEEE Geoscience and Remote Sensing Magazine, 10(1):64-90, 2021. 2", + "[46] Le Sun, Guangrui Zhao, Yuhui Zheng, and Zebin Wu. Spectral-spatial feature tokenization transformer for hyperspectral image classification. IEEE Transactions on Geoscience and Remote Sensing, 60:1-14, 2022. 7", + "[47] Siyu Sun, Jun Liu, and Wei Li. Spatial invariant tensor self-representation model for hyperspectral anomaly detection. IEEE Transactions on Cybernetics, 2023. 2", + "[48] Xian Sun, Peijin Wang, Wanxuan Lu, Zicong Zhu, Xiaonan Lu, Qibin He, Junxi Li, Xue Rong, Zhujun Yang, Hao Chang, et al. Ringmo: A remote sensing foundation model with masked image modeling. IEEE Transactions on Geoscience and Remote Sensing, 61:1-22, 2022. 4" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23057", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Xian Sun, Peijin Wang, Wanxuan Lu, Zicong Zhu, Xiaonan Lu, Qibin He, Junxi Li, Xue Rong, Zhujun Yang, Hao Chang, Qinglin He, Guang Yang, Ruiping Wang, Jiwen Lu, and Kun Fu. Ringmo: A remote sensing foundation model with masked image modeling. IEEE Transactions on Geoscience and Remote Sensing, 61:1-22, 2023. 1, 2", + "[50] Adam Van Etten, Dave Lindenbaum, and Todd M Bacastow. Spacenet: A remote sensing dataset and challenge series. arXiv preprint arXiv:1807.01232, 2018. 2, 3", + "[51] Adam Van Etten, Jacob Shermeyer, Daniel Hogan, Nicholas Weir, and Ryan Lewis. Road network and travel time extraction from multiple look angles with spacenet data. In IGARSS 2020-2020 IEEE International Geoscience and Remote Sensing Symposium, pages 3920-3923. IEEE, 2020. 2, 3", + "[52] Di Wang, Bo Du, and Liangpei Zhang. Fully contextual network for hyperspectral scene parsing. IEEE Transactions on Geoscience and Remote Sensing, 60:1-16, 2021. 7", + "[53] Di Wang, Qiming Zhang, Yufei Xu, Jing Zhang, Bo Du, Dacheng Tao, and Liangpei Zhang. Advancing plain vision transformer toward remote sensing foundation model. IEEE Transactions on Geoscience and Remote Sensing, 61:1-15, 2023. 4", + "[54] Di Wang, Meiqi Hu, Yao Jin, Yuchun Miao, Jiaqi Yang, Yichu Xu, Xiaolei Qin, Jiaqi Ma, Lingyu Sun, Chenxing Li, et al. Hypersigma: Hyperspectral intelligence comprehension foundation model. arXiv preprint arXiv:2406.11519, 2024. 2, 3, 4, 8", + "[55] Shaoyu Wang, Xinyu Wang, Liangpei Zhang, and Yanfei Zhong. Auto-ad: Autonomous hyperspectral anomaly detection network based on fully convolutional autoencoder. IEEE Transactions on Geoscience and Remote Sensing, 60:1-14, 2021. 7", + "[56] Sheng Wang, Kaiyu Guan, Chenhui Zhang, Qu Zhou, Sibo Wang, Xiaocui Wu, Chongya Jiang, Bin Peng, Weiye Mei, Kaiyuan Li, et al. Cross-scale sensing of field-level crop residue cover: Integrating field photos, airborne hyperspectral imaging, and satellite data. Remote Sensing of Environment, 285:113366, 2023. 3", + "[57] Yanheng Wang, Danfeng Hong, Jianjun Sha, Lianru Gao, Lian Liu, Yonggang Zhang, and Xianhui Rong. Spectral-spatial-temporal transformers for hyperspectral image change detection. IEEE Transactions on Geoscience and Remote Sensing, 60:1-14, 2022. 7", + "[58] Philine Warnke, Kerstin Cuhls, Ulrich Schmoch, Lea Daniel, Liviu Andreescu, Bianca Dragomir, Radu Gheorghiu, Catalina Baboschi, Adrian Curaj, Marjukka Parkkinen, et al. 100 radical innovation breakthroughs for the future. 2019. 1", + "[59] Nicholas Weir, David Lindenbaum, Alexei Bastidas, Adam Van Etten, Sean McPherson, Jacob Shermeyer, Varun Kumar, and Hanlin Tang. Spacenet mvoi: A multi-view overhead imagery dataset. In Proceedings of the IEEE/cvf international conference on computer vision, pages 992-1001, 2019. 2, 3", + "[60] Walter F Wiggins and Ali S Tejani. On the opportunities and risks of foundation models for natural language processing in" + ], + "bbox": [ + 91, + 90, + 480, + 902 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "radiology. Radiology: Artificial Intelligence, 4(4):e220119, 2022.3", + "[61] Yinan Wu, Licheng Jiao, Xu Liu, Fang Liu, Shuyuan Yang, and Lingling Li. Domain adaptation-aware transformer for hyperspectral object tracking. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 2, 8", + "[62] Zhitong Xiong, Yi Wang, Fahong Zhang, Adam J Stewart, Joëlle Hanna, Damian Borth, Ioannis Papoutsis, Bertrand Le Saux, Gustau Camps-Valls, and Xiao Xiang Zhu. Neural plasticity-inspired foundation model for observing the earth crossing modalities. arXiv e-prints, pages arXiv-2403, 2024. 8", + "[63] Yang Xu, Zebin Wu, Jun Li, Antonio Plaza, and Zhihui Wei. Anomaly detection in hyperspectral images based on low-rank and sparse representation. IEEE Transactions on Geoscience and Remote Sensing, 54(4):1990-2000, 2015. 7", + "[64] Zongyin Yang, Tom Albrow-Owen, Weiwei Cai, and Tawfique Hasan. Miniaturization of optical spectrometers. Science, 371(6528):eabe0722, 2021. 1", + "[65] Alina Zare, Changzhe Jiao, and Taylor Glenn. Discriminative multiple instance hyperspectral target characterization. IEEE transactions on pattern analysis and machine intelligence, 40(10):2342-2354, 2017. 2", + "[66] Hengwei Zhao, Xinyu Wang, Jingtao Li, and Yanfei Zhong. Class prior-free positive-unlabeled learning with taylor variational loss for hyperspectral remote sensing imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16827-16836, 2023. 2, 7", + "[67] Hengwei Zhao, Yanfei Zhong, Xinyu Wang, and Hong Shu. One-class risk estimation for one-class hyperspectral image classification. IEEE Transactions on Geoscience and Remote Sensing, 2023. 7", + "[68] Zhuo Zheng, Yanfei Zhong, Ailong Ma, and Liangpei Zhang. Fpga: Fast patch-free global learning framework for fully end-to-end hyperspectral image classification. IEEE Transactions on Geoscience and Remote Sensing, 58(8):5612-5626, 2020. 7", + "[69] Yanfei Zhong, Xin Hu, Chang Luo, Xinyu Wang, Ji Zhao, and Liangpei Zhang. Whu-hi: Uav-borne hyperspectral with high spatial resolution (h2) benchmark datasets and classifier for precise crop identification based on deep convolutional neural network with crf. Remote Sensing of Environment, 250:112012, 2020. 1, 2" + ], + "bbox": [ + 516, + 90, + 903, + 696 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "23058", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/HyperFree_ A Channel-adaptive and Tuning-free Foundation Model for Hyperspectral Remote Sensing Imagery/9236da20-05f2-4327-b226-ca754d2dead5_model.json b/2025/HyperFree_ A Channel-adaptive and Tuning-free Foundation Model for Hyperspectral Remote Sensing Imagery/9236da20-05f2-4327-b226-ca754d2dead5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9f7e6d81e175fa19b677da4b797ef074e4baca6d --- /dev/null +++ b/2025/HyperFree_ A Channel-adaptive and Tuning-free Foundation Model for Hyperspectral Remote Sensing Imagery/9236da20-05f2-4327-b226-ca754d2dead5_model.json @@ -0,0 +1,2367 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.138, + 0.13, + 0.861, + 0.178 + ], + "angle": 0, + "content": "HyperFree: A Channel-adaptive and Tuning-free Foundation Model for Hyperspectral Remote Sensing Imagery" + }, + { + "type": "text", + "bbox": [ + 0.157, + 0.203, + 0.842, + 0.258 + ], + "angle": 0, + "content": "Jingtao Li\\*, Yingyi Liu\\*, Xinyu Wang†, Yunning Peng†, Chen Sun†, Shaoyu Wang², Zhendong Sun†, Tian Ke†, Xiao Jiang†, Tangwei Lu†, Anran Zhao†, Yanfei Zhong†" + }, + { + "type": "text", + "bbox": [ + 0.325, + 0.262, + 0.671, + 0.275 + ], + "angle": 0, + "content": "{Jintaoli, liuyingyi, wangxinyu, pengyunning, sunchen," + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.279, + 0.758, + 0.292 + ], + "angle": 0, + "content": "sunzhendong, ketian, jiangxiao, lutangwei, anranzha0415, zhongyanfei}@whu.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.435, + 0.298, + 0.558, + 0.31 + ], + "angle": 0, + "content": "wsy1995@snu.ac.kr" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.345, + 0.327, + 0.361 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.376, + 0.485, + 0.695 + ], + "angle": 0, + "content": "Advanced interpretation of hyperspectral remote sensing images benefits many precise Earth observation tasks. Recently, visual foundation models have promoted the remote sensing interpretation but concentrating on RGB and multispectral images. Due to the varied hyperspectral channels, existing foundation models would face image-by-image tuning situation, imposing great pressure on hardware and time resources. In this paper, we propose a tuning-free hyperspectral foundation model called HyperFree, by adapting the existing visual prompt engineering. To process varied channel numbers, we design a learned weight dictionary covering full-spectrum from \\(0.4 \\sim 2.5\\mu m\\), supporting to build the embedding layer dynamically. To make the prompt design more tractable, HyperFree can generate multiple semantic-aware masks for one prompt by treating feature distance as semantic-similarity. After pre-training HyperFree on constructed large-scale high-resolution hyperspectral images, HyperFree (1 prompt) has shown comparable results with specialized models (5 shots) on 5 tasks and 11 datasets. Code and dataset are accessible at https://rsidea.whu.edu.cn/hyperfree.htm." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.721, + 0.222, + 0.737 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.746, + 0.484, + 0.868 + ], + "angle": 0, + "content": "Hyperspectral imaging has been recognized by the European Union as one of the Top 100 Disruptive Technologies [58], due to its ability to provide high-resolution spectral data for precise Earth observation across diverse applications [64], including environmental monitoring [31], agriculture [69], and defense [20]. Unlike traditional RGB or multispectral imagery, Hyperspectral imagery (HSI) offers several unique features: high spectral resolution (typically" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.345, + 0.897, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.588, + 0.424, + 0.864, + 0.435 + ], + "angle": 0, + "content": "(a) Mainstream Pre-training and Tuning Paradigm of RSFM" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.446, + 0.896, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.633, + 0.519, + 0.82, + 0.53 + ], + "angle": 0, + "content": "(b) Prompt Engineering of SAM-variants" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.54, + 0.896, + 0.607 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.638, + 0.615, + 0.81, + 0.627 + ], + "angle": 0, + "content": "(c) Prompt Engineering of HyperFree" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.639, + 0.907, + 0.752 + ], + "angle": 0, + "content": "Figure 1. Comparison of existing RSFMs with the proposed HyperFree model. (a) The pre-training and tuning paradigm is the mainstream approach, where fine-tuning is essential for handling unseen HSIs. (b) The prompt engineering paradigm reduces the tuning burden but struggles to handle HSIs with hundreds of spectral channels. (c) HyperFree can directly process unseen HSIs with varying channel counts without fine-tuning, and it can identify multiple masks of the same class using only a single prompt." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.76, + 0.907, + 0.821 + ], + "angle": 0, + "content": "\\(10\\mathrm{nm}\\)), an extensive wavelength range (400–2500 nm), and dozens to hundreds of spectral channels [64], which enable the detailed identification of diverse targets, while also increasing the complexity of data processing and analysis." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.821, + 0.909, + 0.898 + ], + "angle": 0, + "content": "Recently, remote sensing foundation models (RSFMs) have demonstrated advanced interpretive capabilities, primarily focusing on multispectral and RGB images [4, 7, 13, 25, 38, 49], while also opening new possibilities for analyzing hyperspectral images. RSFMs follow two main" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.221, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.242, + 0.9 + ], + "angle": 0, + "content": "† Corresponding authors." + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.242, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23048" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.288 + ], + "angle": 0, + "content": "paradigms. The predominant paradigm is pre-training and tuning (P-T), illustrated in Figure 1(a), which begins with training a large-scale backbone model in a self-supervised manner, such as masked autoencoders (MAE) [12], followed by fine-tuning with task-specific heads for various downstream applications [7, 13, 38, 49]. In the P-T paradigm, both the pre-training and the fine-tuning stages can be time-consuming. Second paradigm is prompt engineering (P-E), as illustrated in Figure 1(b), which focuses on designing task-specific prompts for a frozen Segment Anything Model (SAM) [17] to directly perform downstream tasks without additional model tuning [4, 25], making it more efficient and user-friendly in practical." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.289, + 0.483, + 0.635 + ], + "angle": 0, + "content": "The P-E paradigm is particularly well-suited for processing HSI, as it can be directly applied to diverse, unseen scenes without requiring model fine-tuning. However, it faces two significant challenges when applied to HSI. (i) The first challenge is the variation in spectral parameters across different hyperspectral sensors, including differences in spectral resolution, wavelength range, and the number of spectral channels [10]. For a P-T-based model like HyperSigma [54], although it is pre-trained using HSIs with fixed channels, the challenge can be addressed during its tuning stage by rebuilding the channel embedding layer to align the dimensions with the new data. But for a P-E-based model, the challenge of channel variation requires more careful consideration. (ii) The second challenge is to make prompt input tractable, ensuring a P-E based model can directly process unseen HSIs with least prior knowledge. Specifically, most existing P-E-based RSFMs rely on a frozen SAM [17], which segments remote sensing objects based on specialized models to generate prompts [4, 25]. As shown in Figure 1(b), each prompt typically generates a single mask and all the prompts for target objects are needed by existing P-E-based RSFMs, which largely reduced the model feasibility." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.636, + 0.483, + 0.892 + ], + "angle": 0, + "content": "To address the above challenges, as shown in Figure 1 (c), we propose a channel-adaptive, tuning-free hyperspectral foundation model, called HyperFree, which can process any unseen HSIs directly in a P-E manner. Specifically, (i) Motivated by the word dictionary in natural language processing, which stores vector embeddings as values corresponding to words and phrases of varying lengths as keys [32], HyperFree introduces a learnable wavelength weight dictionary in the embedding layer, spanning the 400 to 2500 nm spectral range in \\(10\\mathrm{nm}\\) intervals. This allows for the dynamic encoding of unseen HSIs with varying spectral parameters into visual tokens, all with consistent dimensionality. (ii) To make the prompt of HyperFree more tractable, HyperFree is designed to generate multiple semantic-aware masks for one prompt as in Figure 1 (c), where HyperFree maps each prompt and single mask from image space into feature space and treats the feature distance as semantic-" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "similarity. Prompt, mask and feature are interacted adaptively for different downstream task in tuning-free manner with only one prompt or zero shot." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.137, + 0.905, + 0.408 + ], + "angle": 0, + "content": "Due to the high acquisition cost and label difficulty, current hyperspectral datasets only have a few images [45, 69], far from the requirements to train the promptable HyperFree. To conquer this problem, we developed the Hyper-Seg data engine, in addition to utilizing well-known multispectral datasets (e.g., fMoW [5] and SpaceNet [50, 51, 59]). This engine automatically generates a large-scale annotated HSI dataset, consisting of nearly 50,000 pairs of HSIs and their corresponding segmentation masks. The original data are sourced from AVIRIS airborne hyperspectral sensors, with a spatial size of \\(512 \\times 512\\) pixels and 224 spectral channels spanning from 400 to \\(2500\\mathrm{nm}\\) at a \\(10\\mathrm{nm}\\) spectral resolution. Unlike other RSFMs trained on unannotated satellite hyperspectral imagery with a \\(30\\mathrm{m}\\) spatial resolution [54], the Hyper-Seg data engine features higher spatial resolution (0.6 to \\(5.0\\mathrm{m}\\)), along with annotations of 15.44 million masks. It is currently the largest known hyperspectral dataset with high spatial resolution." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.409, + 0.905, + 0.604 + ], + "angle": 0, + "content": "To demonstrate the model capabilities, we evaluated HyperFree directly in a P-E manner on 11 datasets across 5 segmentation-related tasks, including multi-class classification [21], one-class classification [66], target detection [65], anomaly detection [47], and change detection [26]. Additionally, we verified HyperFree in a P-T manner on 14 datasets spanning 8 tasks, with additional tasks including hyperspectral denoising [27], hyperspectral unmixing [37], and hyperspectral object tracking [61]. The experimental results show that HyperFree achieves comparable accuracy to state-of-the-art models in a tuning-free manner (using just one prompt), and outperforms most models after tuning." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.606, + 0.807, + 0.619 + ], + "angle": 0, + "content": "Our contributions can be summarized as:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.621, + 0.905, + 0.665 + ], + "angle": 0, + "content": "- We propose the first tuning-free hyperspectral foundation model, which can process any hyperspectral image in different tasks with promptable or zero-shot manner." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.666, + 0.905, + 0.71 + ], + "angle": 0, + "content": "- A weight dictionary that spans the full spectrum, enabling dynamic generation of the embedding layer dynamically according to input wavelengths." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.711, + 0.905, + 0.77 + ], + "angle": 0, + "content": "- We propose to map both prompts and masks into feature space to identify multiple semantic-aware masks for one prompt, where different interaction workflows are designed for each downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.771, + 0.905, + 0.832 + ], + "angle": 0, + "content": "- We built the Hyper-Seg data engine to train the HyperFree model and tested it on 11 datasets from 5 tasks in tuning-free manner, 14 datasets from 8 tasks in tuning manner as an extensive experiment." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.621, + 0.905, + 0.832 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.845, + 0.655, + 0.861 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Hyperspectral image processing. There are diverse tasks to process such precise observation with different targets" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23049" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.256 + ], + "angle": 0, + "content": "[31]. (a) Classification: Assigning a specific class to each pixel [21]. (b) One class classification: Discriminating the pixels of target class with labeled binary map [33]. (c) Target detection. Discriminating the pixels of target class with labeled spectrum. (d) Anomaly detection. Discriminating the pixels spectrally deviating from the background [24]. (e) Change detection: Discriminating the changing pixels between two time-steps [26]. Since the output of the above tasks is all pixel-level segmentation maps, it is feasible to unify them using the proposed HyperFree model in tuning-free manner." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.273, + 0.483, + 0.575 + ], + "angle": 0, + "content": "Foundation models. Powered by the transformer architectures, foundation models are characterized by large-scale property in model size and pre-training data, and the transferring ability for downstream tasks [60]. In remote sensing community, some foundation models for multi-spectral images have already been proposed such as SatMAE [7], SepctralEarth [1] and SpectralGPT [13], which trains large ViT-like models [9] with the MAE strategy [12] and massive open-source multi-spectral images (13 channels). In contrast, hyperspectral images have a high acquiring cost [56] and make the large-scale pre-training stage more difficult. A recent study broke the situation by building a largescale hyperspectral dataset HyperGlobal-450K and trained a foundation model called HyperSigma with MAE strategy as well [54]. Our work differs from HyperSigma in two aspects: (i) HyperSigma accepts fixed number of image channels, while HyperFree can deal with any number of channels directly. (ii) HyperSigma needs to be fine-tuned for each downstream task. Differently, HyperFree has ability to process different tasks in fine-tuning free manner." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.576, + 0.483, + 0.771 + ], + "angle": 0, + "content": "Prompt engineering (P-E). Different from the pretraining and then fine-tuning paradigm, P-E can utilize pretrained foundation models to complete downstream tasks directly with prompts [34]. For pre-trained large language models, prompts are mostly hand-crafted text to guide the model generate response for specific task [34, 42]. In computer vision community, SAM follows the same spirit and designs diverse prompts (e.g., point and box) to achieve few-shot transferring performance [17]. Our work extends the visual prompt engineering in hyperspectral data with varying image channels and supports five tasks in tuning-free manner with designed prompt-mask-feature interaction." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.785, + 0.32, + 0.803 + ], + "angle": 0, + "content": "3. Hyper-Seg Data Engine" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.482, + 0.901 + ], + "angle": 0, + "content": "Hyperspectral image processing always refers to various dense tasks and a large-scale segmentation dataset is needed to apply the P-E. However, most hyperspectral segmentation datasets only have a few images due to the high acquisition and label cost, which is quite different from the million-scale dataset with natural images such as SA-1B" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.392 + ], + "angle": 0, + "content": "[17]. To tackle this, we built a data engine called HyperSeg to generate segmented masks automatically for spectral images and expand the data scale. Hyper-Seg engine firstly separates the spectral image into several groups with three channels according to the prior of key channels position. Each group is then segmented by the SAM-H model [17] and the results are combined by NMS operation to eliminate redundancy and output the final segmentation map. Figure 2 shows the overall workflow, where the nine key channels are selected according to the expert knowledge as in Supp.1.1. Considering the high spatial resolution property of airborne platform, we collected 41946 hyperspectral images from AVIRIS airborne sensor and applied the Hyper-Seg Engine. Besides, we also processed two existing multispectral datasets (fMoW [5] and SpaceNet [50, 51, 59]) to increase the scale further. Compared to the traditional hyperspectral annotation methods which refer to high-resolution RGB imagery, our data engine can utilize more spectral information and obtain a high-precision segmentation mask without labor cost." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.394, + 0.905, + 0.621 + ], + "angle": 0, + "content": "Table 1 shows the comparable statistical information of training datasets between existing RSFMs and proposed HyperFree. Most models only concentrate on the RGB and multispectral images with channel number 3 13. Although the recent model HyperSigma [54] has collected a large-scale hyperspectral images, the spatial resolution (30m/pixel) is largely lower than us (0.6 5m/pixel), and the dataset scale (considering image size and number together) is relatively small. With the designed Hyper-Seg data engine, constructed dataset is the only one with segmentation masks supporting the promptable training. More detailed statistical information of our generated dataset is given in Supp.1.2. The final generated dataset has nearly 150k images and 15.44million masks. We name both the data engine and generated dataset as the Hyper-Seg for simplicity." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.633, + 0.795, + 0.65 + ], + "angle": 0, + "content": "4. HyperFree Foundation Model" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.905, + 0.81 + ], + "angle": 0, + "content": "HyperFree has two customized components upon the metaarchitecture of SAM including channel-adaptive embedding and prompt-mask-feature (PMF) interaction as in Figure 3. Channel-adaptive embedding constructs the embedding layer dynamically according to varying input wavelength, and outputs tokens in same representation space. After the promptable training stage, PMF interaction discriminates the semantic-aware masks for downstream tasks by mapping both prompt and mask into feature space and treating feature distance as semantic similarity." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.819, + 0.78, + 0.835 + ], + "angle": 0, + "content": "4.1. Channel-adaptive Embedding" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.905, + 0.901 + ], + "angle": 0, + "content": "To make HyperFree process any hyperspectral image, channel-adaptive embedding layer is firstly designed to map input image with varying channels into tokens with fixed length, inspired by the tokenizer in NLP community [32]." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23050" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.214, + 0.093, + 0.784, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.302, + 0.908, + 0.334 + ], + "angle": 0, + "content": "Figure 2. Hyper-Seg Data Engine automatically generates a large-scale dataset comprising nearly 41,900 pairs of airborne HSIs and their corresponding segmentation masks, enabling HyperFree to effectively learn discriminative spectral features." + }, + { + "type": "table_caption", + "bbox": [ + 0.089, + 0.346, + 0.908, + 0.377 + ], + "angle": 0, + "content": "Table 1. Comparison of training data used in current RSFMs. The primary features of the proposed HyperFree model include prompt engineering and the use of high-resolution annotated hyperspectral (HRS) imagery." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.385, + 0.898, + 0.543 + ], + "angle": 0, + "content": "
DatasetFoundation ModelSpatial Resolution (m)Spectral Resolution (nm)Channel NumberImage SizeImage NumberMain ModalityTraining Paradigm
RGBRVSA [53]0.5 ~ 153\\3110 ~ 316721.00 × 10^6H/M/LSR-RGBP-T
RingMo [48]0.1 ~ 30\\34482.10 × 10^6H/MSR-RGBP-T
ScaleMAE [39]0.31 ~ 3.7\\3\\3.64 × 10^6HSR-RGBP-T
MultispectralSatMAE [6]1015 ~ 18013≈ 45 × 607.13 × 10^5MSR-MP-T
SkySense [11]10/2015 ~ 18010\\\\MSR-MP-T
SpectralGPT [14]10/20/6015 ~ 1801245 ~ 1201.47 × 10^6MSR-MP-T
HyperspectralHyperSigma [54]305/6.5/10150/175644.47 × 10^5LSR-HP-T
HyperFree0.6 ~ 5.0102245124.19 × 10^4HSR-HP-E
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.093, + 0.545, + 0.394, + 0.558 + ], + "angle": 0, + "content": "* P-T: Pretraining and Tuning; P-E: Prompt Engineering." + }, + { + "type": "table_footnote", + "bbox": [ + 0.093, + 0.558, + 0.611, + 0.569 + ], + "angle": 0, + "content": "* HSR: High Spatial Resolution; MSR: Medium Spatial Resolution; LSR: Low Spatial Resolution." + }, + { + "type": "table_footnote", + "bbox": [ + 0.093, + 0.569, + 0.558, + 0.582 + ], + "angle": 0, + "content": "* RGB: Red-Green-Blue Imagery; M: Multispectral Imagery; H: Hyperspectral Imagery" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.545, + 0.611, + 0.582 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.607, + 0.483, + 0.743 + ], + "angle": 0, + "content": "Tokenizers rely on a pre-defined dictionary to break down a string of text into numerical tokens, where the dictionary has stored the numerical vectors of varied-length words or phrases. Vision transformers always use a convolutional layer to generate patch tokens and what if we construct a layer weight dictionary treating each wavelength as a word? Given the input meta data of wavelengths, we can just loop up the corresponding weights and construct the layer dynamically without re-training it." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.743, + 0.483, + 0.895 + ], + "angle": 0, + "content": "The weight dictionary is built depending on the wavelengths to be processed and the expected token dimension. Since hyperspectral imaging mostly captures the spectral signals from \\(400\\mathrm{nm}\\) to \\(2500\\mathrm{nm}\\) and the \\(10\\mathrm{nm}\\) is of high spectral resolution currently, 221 index channels are designed as the dictionary keys. Denote the output token as a vector of \\(\\mathbb{R}^j\\) corresponding to the patch size \\(p\\times p\\), each index channel stores a weight matrix of size \\(\\mathbb{R}^{p\\times p\\times j}\\), and the built dictionary \\(\\beta\\) has a size of \\(221\\times p\\times p\\times j\\). Given the input hyperspectral image \\(\\mathbb{R}^{h\\times w\\times n}\\) with wavelengths \\(\\mathbf{b} =\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.607, + 0.907, + 0.7 + ], + "angle": 0, + "content": "\\([b_1, b_2, \\ldots, b_n]\\), the corresponding kernel weight can be dynamically generated by combining the searched weight \\(\\mathbf{W}\\) with size \\(n \\times p \\times p \\times j\\), where \\(n\\) is flexible for each image. For conciseness, we use function \\(g\\) to represent the mapping process from \\(\\mathbf{b}\\) to \\(\\mathbf{W}\\) as kernel weight, and function \\(f\\) to represent the token generation step as in Equation (1)." + }, + { + "type": "equation", + "bbox": [ + 0.535, + 0.723, + 0.907, + 0.742 + ], + "angle": 0, + "content": "\\[\nf (\\mathbf {X} \\mid \\mathbf {b}, \\beta) = \\operatorname {C o n v} _ {\\mathbf {W}} (\\mathbf {X}) \\quad \\text {w h e r e} \\quad \\mathbf {W} = g (\\mathbf {b}, \\beta) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In the past few decades, many satellites have been launched equipped with spectral sensors, where the key channels are selected by professional teams and practical experience. To utilize these successful priors, we further extract visual tokens in two parallel branches with dynamically constructed layers: one to process key channels and the other to process the split cubes, as shown in Figure 3. The input image \\(\\mathbf{X}\\) is first divided into key channels \\(\\mathbf{X}_{\\mathbf{k}}\\) and the cubes \\(\\mathbf{X}_{\\mathbf{c}}\\) between them, where the key wavelengths are selected by referring to famous spectral sensors (detailed" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23051" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.129, + 0.089, + 0.511, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.363, + 0.424, + 0.374 + ], + "angle": 0, + "content": "(a) Full spectrum Promptable Training" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.088, + 0.87, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.574, + 0.362, + 0.825, + 0.375 + ], + "angle": 0, + "content": "(b) Prompt-mask-feature Interactive Inferring" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.385, + 0.907, + 0.47 + ], + "angle": 0, + "content": "Figure 3. (a) Full spectrum Promptable Training. Based on the meta-architecture of promptable segmentation, channel-adaptive embedding layer is developed to process any-channel input, which relies on a learnable weight dictionary and generates the embedding layer dynamically according to the input wavelengths. With the constructed Hyper-Seg dataset, HyperFree is trained with randomly selected channels from full spectrum and point prompts in each iteration. (b) Prompt-mask-feature Interactive Inferring. Keep the trained HyperFree frozen, different PMF interaction workflows are designed to complete the downstream tasks in promptable or zero-shot manner, where we use feature distance to measure the semantic consistency. An interaction instantiation for classification task is given above as an example." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.489, + 0.486, + 0.551 + ], + "angle": 0, + "content": "in Supp.2). We set two different dictionaries \\(\\beta_{\\mathbf{k}}\\) and \\(\\beta_{\\mathbf{c}}\\) to process \\(\\mathbf{X}_{\\mathbf{k}}\\) and \\(\\mathbf{X}_{\\mathbf{c}}\\) respectively, and the final output tokens \\(\\mathbf{T} \\in \\mathbb{R}^{(h/p) \\times (w/p) \\times j}\\) are computed as the element-wise sum, as shown in Equation (2)." + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.567, + 0.483, + 0.584 + ], + "angle": 0, + "content": "\\[\n\\mathbf {T} = f \\left(\\mathbf {X} _ {\\mathbf {k}} \\mid \\mathbf {b} _ {\\mathbf {k}}, \\beta_ {\\mathbf {k}}\\right) + f \\left(\\mathbf {X} _ {\\mathbf {c}} \\mid \\mathbf {b} _ {\\mathbf {c}}, \\beta_ {\\mathbf {c}}\\right) \\tag {2}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.594, + 0.405, + 0.611 + ], + "angle": 0, + "content": "4.2. Full-spectrum Promptable Training" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.617, + 0.484, + 0.768 + ], + "angle": 0, + "content": "The prompt engineering paradigm is adopted to complete downstream tasks directly with designed prompts. Hyper-Free is thus trained in a promptable manner on constructed large-scale Hyper-Seg dataset, where the prompts are randomly chosen at each iteration and the model is expected to segment correctly the corresponding masks. Since hyperspectral image processing tasks are almost conducted at the pixel level, we choose point prompts only at the training stage. Focal loss \\( L_{f} \\) and Dice loss \\( L_{d} \\) are weighted as [17] to supervise the training as Equation (3)." + }, + { + "type": "equation", + "bbox": [ + 0.231, + 0.786, + 0.483, + 0.802 + ], + "angle": 0, + "content": "\\[\nL = 2 0 L _ {f} + L _ {d} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.485, + 0.903 + ], + "angle": 0, + "content": "To train the channel-adaptive embedding layer effectively, we select the subset of channels randomly from the full-spectrum range and process it with corresponding weights only in \\(\\beta_{\\mathbf{k}}\\) and \\(\\beta_{\\mathrm{c}}\\), which helps the learned dictionary being wavelength-aware and encode the varied channel composition into a unified token space." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.489, + 0.9, + 0.506 + ], + "angle": 0, + "content": "4.3. Prompt-mask-feature (PMF) Interactive Inferring" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.907, + 0.72 + ], + "angle": 0, + "content": "After training process, the model can process unseen hyperspectral image with varied channels and segment the corresponding mask given any point prompt. However, many downstream tasks need to segment out all the masks of same semantic category rather than a single mask such as hyperspectral classification and target detection. We tackle this problem by connecting the prompt, mask and feature together in feature space, where prompt tells the model an object reference in task, mask outputs accurate location information, and feature reflects semantic information. In unified feature space, the cosine distance of features is used to measure the semantic similarity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Before diving into the specific workflows, we introduce two basic interaction modes with a simple situation. Set one point \\((x,y)\\) as input prompt for image \\(\\mathbf{X}\\), the corresponding feature cube \\(\\mathbf{D}\\) and all the generated masks \\(\\mathbf{M} \\in \\mathbb{R}^{h \\times w \\times k}\\) from the trained model, where \\(\\mathbf{M}\\) has \\(k\\) binary maps and each represents a single mask. The first interaction mode maps the prompt \\((x,y)\\) into a feature vector \\(\\mathbf{d}_{(x,y)} \\in \\mathbb{R}^j\\) as in Equation (4), where the corresponding valid and smallest mask in \\(\\mathbf{M}\\) is used to localize the features in \\(\\mathbf{D}\\), and the mean feature is treated as \\(\\mathbf{d}_{(x,y)}\\) for the given prompt. This interaction transforms the prompt from image space to feature space, and makes it possible to explore the semantic" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23052" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.162, + 0.106 + ], + "angle": 0, + "content": "similarity." + }, + { + "type": "equation", + "bbox": [ + 0.184, + 0.116, + 0.483, + 0.133 + ], + "angle": 0, + "content": "\\[\n\\text {M o d e} (x, y) \\mid \\mathbf {M}, \\mathbf {D} \\rightarrow \\mathbf {d} _ {(x, y)} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.141, + 0.484, + 0.291 + ], + "angle": 0, + "content": "The second interaction mode selects multiple the masks in \\(\\mathbf{M}\\), which have similar semantic information with \\(\\mathbf{d}_{(x,y)}\\) as in Equation (5). The selected masks are represented as \\(\\mathbf{M}_d\\) and \\(\\mathbf{M}_d \\subseteq \\mathbf{M}\\). Specifically, we compute a feature representation for each single mask by localizing in \\(\\mathbf{D}\\) and computing the average, and the cosine similarity between them and \\(\\mathbf{d}_{(x,y)}\\) are treated as the semantic similarity. With some hand-designed thresholds \\(\\tau\\), \\(\\mathbf{M}_d\\) can be finally segmented. \\(\\tau\\) represents different meanings according to the specific task." + }, + { + "type": "equation", + "bbox": [ + 0.177, + 0.307, + 0.483, + 0.324 + ], + "angle": 0, + "content": "\\[\n\\text {M o d e} 2: \\left(\\mathbf {d} _ {(x, y)} | \\mathbf {M}, \\mathbf {D}, \\tau\\right)\\rightarrow \\mathbf {M} _ {d} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.329, + 0.485, + 0.661 + ], + "angle": 0, + "content": "By using both interaction modes adaptively, five hyperspectral tasks can be completed with the trained promptable model directly. (a) Hyperspectral classification (HC). Each category needs at least one prompt, and uses Mode 1 and Mode 2 in turn for each category. For a common closed-set setting, \\(\\tau\\) is not necessary since we can use the minimum feature distance between each mask representation and all the prompt representations to decide its category. Figure 3 (b) shows an exemplified workflow. (b) Hyperspectral one-class classification (HOCC). HOCC is similar to H but the difference lies in that HOCC only needs prompts the target category, and its \\(\\tau\\) is instantiated as a class pr parameter to control the mask ratio. (c) Hyperspectral taget detection (HTD). Compared to HOCC, HTD acce the target spectra and it needs to be changed to a pron by selecting the pixel with the smallest cosine distance. is instantiated as a distance threshold to convert the dens map to a binary map. (d) Hyperspectral anomaly detion (HAD). HAD is an unsupervised task and there is \\(\\mathbf{d}_{(x,y)}\\) for HAD. Since HAD is very similar to unsupervis target detection, we use M only and filter out large objec according to mask area to output the final anomaly map." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.661, + 0.484, + 0.737 + ], + "angle": 0, + "content": "Hyperspectral change detection (HCD). Different from the prior tasks, HCD accepts two registered and temporal images to identify the changing locations. The HCD task utilizes both interaction modes for the bi-temporal images. Mode 1 processes the first image to get representation \\(\\mathbf{d}_{(x,y)}\\)" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.745, + 0.483, + 0.772 + ], + "angle": 0, + "content": "Table 2. Instantiation summary of prompts, meaning of \\( \\tau \\), and interaction modes when HyperFree performs different tasks." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.784, + 0.485, + 0.898 + ], + "angle": 0, + "content": "
TaskPrompt TypeHyperparameter τInteraction Mode
HCOne Point Per CategoriesNo τMode 1, Mode 2
HOCCTarget Category PointClass PriorMode 1, Mode 2
HTDTarget SpectraSimilarity ThresholdMode 1, Mode 2
HADNo PromptMasks Area ThresholdMode 1
HCDNo PromptSimilarity ThresholdMode 1, Mode 2
" + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.093, + 0.885, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.434, + 0.907, + 0.459 + ], + "angle": 0, + "content": "Figure 4. Exemplified comparison results for five tuning-free tasks" + }, + { + "type": "image", + "bbox": [ + 0.537, + 0.461, + 0.885, + 0.676 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.684, + 0.907, + 0.725 + ], + "angle": 0, + "content": "Figure 5. In tuning-free manner, HyperFree (1 prompt) can achieve comparable results with specialized models (5 shots) on 5 tasks and 11 datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.736, + 0.907, + 0.811 + ], + "angle": 0, + "content": "for each mask, which is input to Mode 2 and compared with the corresponding mask representation with \\(\\mathbf{M}\\) and \\(\\mathbf{D}\\) of the second image. Once the temporal feature distance exceeds the preset \\(\\tau\\), a change is considered to have happened. The interaction workflow is summarized in Table 2." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.824, + 0.608, + 0.839 + ], + "angle": 0, + "content": "5. Results" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.691, + 0.866 + ], + "angle": 0, + "content": "5.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "We evaluate the proposed HyperFree on five downstream hyperspectral tasks including HC, HOCC, HTD, HAD and" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23053" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.083, + 0.907, + 0.112 + ], + "angle": 0, + "content": "Table 3. Quantitative comparison results on five hyperspectral tasks, where blue numbers indicate the metric ranking of HyperFree. Comparison methods were all trained on each dataset while HyperFree processed the datasets in tuning-free manner." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.123, + 0.909, + 0.495 + ], + "angle": 0, + "content": "
HC Task
LongKou (270 channels)HanChuan (274 channels)
MetricsSVM [30] (5 shot)HybridSN [41] (5 shot)FCN [52] (5 shot)FPGA [68] (5 shot)SSFTT [46] (5 shot)MambaHSI [23] (5 shot)HyperFree (1 prompt)SVM [30] (5 shot)HybridSN [41] (5 shot)FCN [52] (5 shot)FPGA [68] (5 shot)SSFTT [46] (5 shot)MambaHSI [23] (5 shot)HyperFree (1 prompt)
OA82.7748.7886.6791.1889.6692.6593.39(1st)52.6847.7555.5571.4764.8673.3375.47(1st)
AA74.0261.3785.6088.3587.9692.5788.03(3rd)47.7646.1759.7272.0961.2269.3362.35(2nd)
KA78.0435.7282.3088.6687.9590.0091.41(1st)46.8541.3150.1867.5859.6569.1071.56(1st)
HOCC Task
HongHu (270 channels)XiongAn (256 channels)
MetricsOCSVM [43] (5 shot)nnPU [18] (5 shot)BSVM [33] (5 shot)PAN [15] (5 shot)OC Loss [67] (5 shot)T-HOneCls [66] (5 shot)HyperFree (1 prompt)OCSVM [43] (5 shot)nnPU [18] (5 shot)BSVM [33] (5 shot)PAN [15] (5 shot)OC Loss [67] (5 shot)T-HOneCls [66] (5 shot)HyperFree (1 prompt)
F26.3319.1334.8263.6954.7355.9772.52(1st)18.311.7626.3046.3443.0841.3452.50(1st)
P56.4319.7250.7780.0058.2746.5268.61(2nd)39.832.8523.8247.1347.5032.8768.21(1st)
R24.0218.5845.2964.2754.3492.3578.06(2nd)16.081.9857.8353.3247.6160.3865.03(1st)
HTD Task
Airport (205 channels)Cri (46 channels)
MetricsACE [19] (1 shot)CEM [2] (1 shot)GLRT [28] (1 shot)MF [29] (1 shot)HTD-IRN [44] (1 shot)TSTTD [16] (1 shot)HyperFree (1 prompt)ACE [19] (1 shot)CEM [2] (1 shot)GLRT [28] (1 shot)MF [29] (1 shot)HTD-IRN [44] (1 shot)TSTTD [16] (1 shot)HyperFree (1 prompt)
AUC(D,F)0.97940.96030.98010.99160.97450.99290.9806(3rd)0.97350.98930.97370.98910.99750.99870.9791(5th)
AUCODP1.58531.28291.57981.69681.44841.65921.4667(6th)1.20151.45061.20001.45751.39951.61031.4670(2nd)
HAD Task
Beach1 (188 channels)Beach2 (193 channels)
MetricsRXD [40]CRD [22]ADLR [36]LRASR [63]Auto-AD [55]TDD [20]HyperFree (zero shot)RXD [40]CRD [22]ADLR [36]LRASR [63]Auto-AD [55]TDD [20]HyperFree (zero shot)
AUC(D,F)0.98150.94710.45150.74610.95740.98420.9730(3rd)0.90900.85440.79760.82250.94850.96270.9566(2nd)
AUCODP1.25570.97850.56100.85261.12731.13831.9191(1st)1.01770.86700.90640.82801.00971.16881.8697(1st)
HCD Task
River (154 channels)Hermiston (198 channels)
MetricsFC-EF [8] (5 shot)FC-Sc [8] (5 shot)FC-Sd [8] (5 shot)ML-EDAN [35] (5 shot)BIT [3] (5 shot)SST-Former [57] (5 shot)HyperFree (zero shot)FC-EF [8] (5 shot)FC-Sc [8] (5 shot)FC-Sd [8] (5 shot)ML-EDAN [35] (5 shot)BIT [3] (5 shot)SST-Former [57] (5 shot)HyperFree (zero shot)
IoU0.41680.45220.45340.39150.21260.40960.3649(6th)0.37290.37760.48730.32520.52570.53610.5851(1st)
F0.58840.62280.62390.56280.35070.58120.5347(6th)0.54320.54820.65.520.49080.68910.69800.7382(1st)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.246, + 0.503, + 0.75, + 0.516 + ], + "angle": 0, + "content": "Table 4. Compared results with SAM using the identical PMF interaction workflows." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.528, + 0.907, + 0.582 + ], + "angle": 0, + "content": "
HCHOCCHTDHADHCD
OAAAKAFPrecisionRecallAUC(D,F)AUCODPAUC(D,F)AUCODPFIoU
SAM+PMF Interaction66.6272.9959.4134.1330.3539.470.79791.10140.96151.88440.59960.4282
HyperFree93.4085.6691.4272.5268.6178.060.97911.46700.95661.86970.73820.5851
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.607, + 0.484, + 0.743 + ], + "angle": 0, + "content": "HCD tasks as in Section 4.3. A total of 11 datasets were used in the evaluation (detailed in Supp.3), and each task had six comparison methods, including both classical and state-of-the-art ones. Since there is no method that can directly complete all five tasks, all comparative methods are tested after being trained on each dataset with five shots per class, while the parameters of HyperFree are fixed for all datasets with 1 prompt or zero-shot as in Table 2. We use ViT-base as backbone to conduct the experiments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.902 + ], + "angle": 0, + "content": "HyperFree was trained on the constructed Hyper-Seg dataset using 8 A100 GPUs for approximately 10 hours, which was based on the meta-architecture and parameters of SAM. We use AdamW optimizer with initial learning rate 8e-5 and batch size 16. To help learning the channel-adaptive embedding, we randomly select the image channels at each iteration and match it with corresponding wavelengths. 16 masks and 1 2 points per mask are randomly selected for each input image to train" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.608, + 0.712, + 0.621 + ], + "angle": 0, + "content": "the promptable segmentation." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.623, + 0.644, + 0.636 + ], + "angle": 0, + "content": "5.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.644, + 0.907, + 0.825 + ], + "angle": 0, + "content": "Most tuning-free quantitative results are reported in Table 3 (1 dataset in supplement due to the space limit) and Figure 5. All the comparative methods are trained with 5 shot for each class and HyperFree processes each dataset directly without tuning. For HC, HOCC and HTD, only one point prompt is provided for each class. HAD and HCD does not need any prompt and deal with images in zero-shot manner. Table 3 shows that HyperFree can surpass most of the specialized models with frozen parameters. On both HC and HCD tasks, HyperFree even outperformed previous SOTA results, including an improvement of about 12 points in \\(\\mathbf{F}_1\\) on HC and about 0.5 points in IoU on HCD task." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.901 + ], + "angle": 0, + "content": "We performed qualitative comparison on five exemplified images in Figure 4, where one SOTA model have been visualized for each task comparison. Obviously, the result maps of HyperFree are more complete and smoother than the ones of specialized models. In HC, HOCC, HTD and" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23054" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.289, + 0.083, + 0.707, + 0.098 + ], + "angle": 0, + "content": "Table 5. Ablation results on the usage of the learned weight dictionary." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.109, + 0.905, + 0.187 + ], + "angle": 0, + "content": "
Correct WavelengthsβkβcHCHOCCHTDHADHCD
OAAAKAF1PrecisionRecallAUC(D,F)AUCODPAUC(D,F)AUCODPF1IoU
✓✓×73.1864.2365.8565.6247.3064.560.97671.46440.94881.84650.57930.4078
✓×✓66.6371.5259.4463.6258.6178.440.97681.46400.95891.87680.70220.5411
×✓✓75.2538.9866.4457.3064.0954.830.95951.42730.95571.86780.68320.5188
✓✓✓93.4085.6691.4272.5268.6178.060.97911.46700.95661.86970.73820.5851
" + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.209, + 0.283, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.123, + 0.335, + 0.253, + 0.347 + ], + "angle": 0, + "content": "(a) HyperSigma features" + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.209, + 0.482, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.327, + 0.334, + 0.448, + 0.346 + ], + "angle": 0, + "content": "(b) HyperFree features" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.352, + 0.483, + 0.394 + ], + "angle": 0, + "content": "Figure 6. t-SNE visualizations of HyperSigma [54] and HyperFree features on WHU-Hi Longkou dataset. We randomly select 1000 pixels for each class." + }, + { + "type": "image", + "bbox": [ + 0.115, + 0.396, + 0.462, + 0.602 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.612, + 0.483, + 0.64 + ], + "angle": 0, + "content": "Figure 7. In tuning manner, HyperFree can achieve overall better results with specialized models on 8 tasks and 14 datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.483, + 0.674 + ], + "angle": 0, + "content": "HAD tasks, HyperFree can generate basically the same result as GT, which are difficulty for existing models." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.682, + 0.242, + 0.697 + ], + "angle": 0, + "content": "5.3. Analysis Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.483, + 0.84 + ], + "angle": 0, + "content": "HyperFree vs SAM. Proposed prompt-mask-feature interaction strategy adapts the pre-trained HyperFree to various downstream tasks. We applied the same strategy to SAM to verify the effectiveness of training on hyperspectral images and reported the results in Table 4. The metrics on most tasks have a decline of about \\(10\\sim 30\\) points, implying a failed processing. We realize it is not a fair comparison and only use it to reflect the difference of hyperspectral data processing." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "HyperFree vs HyperSigma. HyperSigma is the only open-sourced hyperspectral foundation model [54, 62], which adopts the pre-training and tuning paradigm. Thus, we cannot compare HyperFree with it in tuning-free set-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.212, + 0.905, + 0.271 + ], + "angle": 0, + "content": "ting directly and compare the t-SNE visualization in Figure 6(a) and Figure 6(b). In contrast, proposed HyperFree has a more compact intra-class distance obviously, supporting the tuning-free workflow of PMF interaction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.272, + 0.906, + 0.393 + ], + "angle": 0, + "content": "Full-Spectrum Weight Dictionary. To verify stored weights are wavelength-aware, we conducted related ablation studies in Table 5. In HC, HOCC and HCD tasks, a drop in accuracy of about \\(5\\sim 40\\) points can be observed if we shuffled the correct wavelength order randomly, showing the wavelength-aware property of learned dictionary. Besides, we proved the built two encoding branches with dictionaries \\(\\beta_{\\mathbf{k}}\\) and \\(\\beta_{\\mathrm{c}}\\) can benefit each other effectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.393, + 0.905, + 0.483 + ], + "angle": 0, + "content": "Tuning HyperFree. As an extensive experiment, we have also tested the tuning performance of HyperFree on 8 tasks and 14 datasets in Figure 7, which further included hyperspectral denosing (HD) [27], object tracking (HOT) [61] and unmxing (HU) tasks [37]. The tuning HyperFree has surpassed the previous SOTA models on nearly all tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.483, + 0.906, + 0.59 + ], + "angle": 0, + "content": "More analysis. For more analytical experiments, please see the supplementary material. (i) Supp.4.1: Complete qualitative comparison of 11 datasets in tuning-free manner. (ii) Supp.4.2: Complete quantitative and qualitative comparison of 14 datasets in tuning manner. (iii) Supp.4.3: Sensitivity analysis about prompt number, prompt location and the hyperparameter \\(\\tau\\)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.65, + 0.638, + 0.665 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Hyperspectral images can provide rich spectral information and are widely used in resource monitoring and defense fields. However, its unique properties increase the difficulty of foundation model development such as varied-length channels, high sensitivity to imaging conditions and the acquirement difficulty, making the mainstream P-T being low cost-effectiveness due to the image-by-image tuning. Our work tackles this problem by converting to adapt the visual P-E [17]. Differently, we design a learnable weight dictionary covering full-spectrum to generate the channel-adaptive embeddings and PMF interaction to generate semantic-aware masks for different tasks. We hope HyperFree can speed up the era of \"GPT moment\" for hyperspectral intelligent processing." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23055" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.097, + 0.091, + 0.258, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.486, + 0.161 + ], + "angle": 0, + "content": "This work was supported by the National Natural Science Foundation of China under Grant No. 42325105 and Grant 424B2010." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.173, + 0.188, + 0.189 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.199, + 0.484, + 0.254 + ], + "angle": 0, + "content": "[1] Nassim Ait Ali Braham, Conrad M Albrecht, Julien Mairal, Jocelyn Chanussot, Yi Wang, and Xiao Xiang Zhu. Spectral earth: Training hyperspectral foundation models at scale. arXiv preprint arXiv:2408.08447, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.256, + 0.484, + 0.31 + ], + "angle": 0, + "content": "[2] Chein-I Chang. Constrained energy minimization (cem) for hyperspectral target detection: Theory and generalizations. IEEE Transactions on Geoscience and Remote Sensing, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.312, + 0.483, + 0.353 + ], + "angle": 0, + "content": "[3] Hao Chen, Zipeng Qi, and Zhenwei Shi. Remote sensing image change detection with transformers. IEEE Transactions on Geoscience and Remote Sensing, 60:1-14, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.354, + 0.483, + 0.423 + ], + "angle": 0, + "content": "[4] Keyan Chen, Chenyang Liu, Hao Chen, Haotian Zhang, Wenyuan Li, Zhengxia Zou, and Zhenwei Shi. Rsprompter: Learning to prompt for remote sensing instance segmentation based on visual foundation model. IEEE Transactions on Geoscience and Remote Sensing, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.424, + 0.483, + 0.48 + ], + "angle": 0, + "content": "[5] Gordon Christie, Neil Fendley, James Wilson, and Ryan Mukherjee. Functional map of the world. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6172-6180, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.481, + 0.484, + 0.55 + ], + "angle": 0, + "content": "[6] Yezhen Cong, Samar Khanna, Chenlin Meng, Patrick Liu, Erik Rozi, Yutong He, Marshall Burke, David Lobell, and Stefano Ermon. Satmae: Pre-training transformers for temporal and multi-spectral satellite imagery. Advances in Neural Information Processing Systems, 35:197-211, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.551, + 0.484, + 0.647 + ], + "angle": 0, + "content": "[7] Yezhen Cong, Samar Khanna, Chenlin Meng, Patrick Liu, Erik Rozi, Yutong He, Marshall Burke, David B. Lobell, and Stefano Ermon. Satmae: pre-training transformers for temporal and multi-spectral satellite imagery. In Proceedings of the 36th International Conference on Neural Information Processing Systems, Red Hook, NY, USA, 2024. Curran Associates Inc. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.649, + 0.483, + 0.704 + ], + "angle": 0, + "content": "[8] Rodrigo Caye Daudt, Bertr Le Saux, and Alexandre Boulch. Fully convolutional siamese networks for change detection. In 2018 25th IEEE international conference on image processing (ICIP), pages 4063-4067. IEEE, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.705, + 0.483, + 0.801 + ], + "angle": 0, + "content": "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.483, + 0.845 + ], + "angle": 0, + "content": "[10] Alexander FH Goetz, Gregg Vane, Jerry E Solomon, and Barrett N Rock. Imaging spectrometry for earth remote sensing. science, 228(4704):1147-1153, 1985. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[11] Xin Guo, Jiangwei Lao, Bo Dang, Yingying Zhang, Lei Yu, Lixiang Ru, Liheng Zhong, Ziyuan Huang, Kang Wu, Dingxiang Hu, et al. Skysense: A multi-modal remote sensing foundation model towards universal interpretation for" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.199, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "earth observation imagery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27672-27683, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[12] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[13] Danfeng Hong, Bing Zhang, Xuyang Li, Yuxuan Li, Chenyu Li, Jing Yao, Naoto Yokoya, Hao Li, Pedram Ghamisi, Xiuping Jia, Antonio Plaza, Paolo Gamba, Jon Atli Benediktsson, and Jocelyn Chanussot. Spectralgpt: Spectral remote sensing foundation model. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(8):5227-5244, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.304, + 0.905, + 0.374 + ], + "angle": 0, + "content": "[14] Danfeng Hong, Bing Zhang, Xuyang Li, Yuxuan Li, Chenyu Li, Jing Yao, Naoto Yokoya, Hao Li, Pedram Ghamisi, Xi-uping Jia, et al. Spectralgpt: Spectral remote sensing foundation model. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.375, + 0.905, + 0.444 + ], + "angle": 0, + "content": "[15] Wenpeng Hu, Ran Le, Bing Liu, Feng Ji, Jinwen Ma, Dongyan Zhao, and Rui Yan. Predictive adversarial learning from positive and unlabeled data. In Proceedings of the AAAI conference on artificial intelligence, pages 7806-7814, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.446, + 0.905, + 0.502 + ], + "angle": 0, + "content": "[16] Jinyue Jiao, Zhiqiang Gong, and Ping Zhong. Triplet spectral-wise transformer network for hyperspectral target detection. IEEE Transactions on Geoscience and Remote Sensing, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.504, + 0.905, + 0.585 + ], + "angle": 0, + "content": "[17] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 2, 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.588, + 0.905, + 0.643 + ], + "angle": 0, + "content": "[18] Ryuichi Kiryo, Gang Niu, Martinus C Du Plessis, and Masashi Sugiyama. Positive-unlabeled learning with nonnegative risk estimator. Advances in neural information processing systems, 30, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.645, + 0.905, + 0.687 + ], + "angle": 0, + "content": "[19] Shawn Kraut, Louis L Scharf, and L Todd McWhorter. Adaptive subspace detectors. IEEE Transactions on signal processing, 49(1):1-16, 2001. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.688, + 0.905, + 0.757 + ], + "angle": 0, + "content": "[20] Jingtao Li, Xinyu Wang, Shaoyu Wang, Hengwei Zhao, and Yanfei Zhong. One step detection paradigm for hyperspectral anomaly detection via spectral deviation relationship learning. IEEE Transactions on Geoscience and Remote Sensing, 2024. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.758, + 0.905, + 0.828 + ], + "angle": 0, + "content": "[21] Jiaojiao Li, Zhiyuan Zhang, Yuzhe Liu, Rui Song, Yunsong Li, and Qian Du. Swformer: Stochastic windows convolutional transformer for hybrid modality hyperspectral classification. IEEE Transactions on Image Processing, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.83, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[22] Wei Li and Qian Du. Collaborative representation for hyperspectral anomaly detection. IEEE Transactions on geoscience and remote sensing, 53(3):1463-1474, 2014. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[23] Yapeng Li, Yong Luo, Lefei Zhang, Zengmao Wang, and Bo Du. Mambahsi: Spatial-spectral mamba for hyperspectral" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "23056" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.12 + ], + "angle": 0, + "content": "image classification. IEEE Transactions on Geoscience and Remote Sensing, 62:1-16, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.177 + ], + "angle": 0, + "content": "[24] Jie Lian, Lizhi Wang, He Sun, and Hua Huang. Gthad: Gated transformer for hyperspectral anomaly detection. IEEE Transactions on Neural Networks and Learning Systems, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.484, + 0.234 + ], + "angle": 0, + "content": "[25] Nanqing Liu, Xun Xu, Yongyi Su, Haojie Zhang, and Heng-Chao Li. Pointsam: Pointly-supervised segment anything model for remote sensing images. arXiv preprint arXiv:2409.13401, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.236, + 0.483, + 0.303 + ], + "angle": 0, + "content": "[26] Fulin Luo, Tianyuan Zhou, Jiamin Liu, Tan Guo, Xiwen Gong, and Jinchang Ren. Multiscale diff-changed feature fusion network for hyperspectral image change detection. IEEE Transactions on Geoscience and Remote Sensing, 61: 1-13, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.307, + 0.483, + 0.361 + ], + "angle": 0, + "content": "[27] Zhaozhi Luo, Xinyu Wang, Petri Pellikka, Janne Heiskanen, and Yanfei Zhong. Unsupervised adaptation learning for real multiplatform hyperspectral image denoising. IEEE transactions on cybernetics, 2024. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.363, + 0.483, + 0.404 + ], + "angle": 0, + "content": "[28] Dimitris Manolakis and Gary Shaw. Detection algorithms for hyperspectral imaging applications. IEEE signal processing magazine, 19(1):29-43, 2002. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.483, + 0.473 + ], + "angle": 0, + "content": "[29] Dimitris Manolakis, Eric Truslow, Michael Pieper, Thomas Cooley, and Michael Brueggeman. Detection algorithms in hyperspectral imaging systems: An overview of practical algorithms. IEEE Signal Processing Magazine, 31(1):24-33, 2013. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.477, + 0.483, + 0.532 + ], + "angle": 0, + "content": "[30] Farid Melgani and Lorenzo Bruzzone. Classification of hyperspectral remote sensing images with support vector machines. IEEE Transactions on geoscience and remote sensing, 42(8):1778-1790, 2004. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.534, + 0.483, + 0.601 + ], + "angle": 0, + "content": "[31] Mohammed Abdulmajeed Moharram and Divya Meena Sundaram. Land use and land cover classification with hyperspectral data: A comprehensive review of methods, challenges and future directions. Neurocomputing, 536:90-113, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.605, + 0.483, + 0.659 + ], + "angle": 0, + "content": "[32] Aleksandar Petrov, Emanuele La Malfa, Philip Torr, and Adel Bibi. Language model tokenizers introduce unfairness between languages. Advances in Neural Information Processing Systems, 36, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.483, + 0.73 + ], + "angle": 0, + "content": "[33] Rami Piiroinen, Fabian Ewald Fassnacht, Janne Heiskanen, Eduardo Maeda, Benjamin Mack, and Petri Pellikka. Invasive tree species detection in the eastern arc mountains biodiversity hotspot using one class classification. Remote sensing of environment, 218:119-131, 2018. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.733, + 0.483, + 0.787 + ], + "angle": 0, + "content": "[34] Maciej P Polak and Dane Morgan. Extracting accurate materials data from research papers with conversational language models and prompt engineering. Nature Communications, 15(1):1569, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.483, + 0.856 + ], + "angle": 0, + "content": "[35] Jiahui Qu, Shaoxiong Hou, Wenqian Dong, Yunsong Li, and Weiying Xie. A multilevel encoder-decoder attention network for change detection in hyperspectral images. IEEE Transactions on Geoscience and Remote Sensing, 60:1-13, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[36] Ying Qu, Wei Wang, Rui Guo, Bulent Ayhan, Chiman Kwan, Steven Vance, and Hairong Qi. Hyperspectral anomaly detection through spectral unmixing and dictionary-based low" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.906, + 0.12 + ], + "angle": 0, + "content": "rank decomposition. IEEE Transactions on Geoscience and Remote Sensing, 56(8):4391-4405, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.906, + 0.189 + ], + "angle": 0, + "content": "[37] Behnood Rasti, Alexandre Zouaoui, Julien Mairal, and Jocelyn Chanussot. Image processing and machine learning for hyperspectral unmixing: An overview and the hysupp python package. IEEE Transactions on Geoscience and Remote Sensing, 2024. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.191, + 0.906, + 0.287 + ], + "angle": 0, + "content": "[38] C. J. Reed, R. Gupta, S. Li, S. Brockman, C. Funk, B. Clipp, K. Keutzer, S. Candido, M. Uytendaele, and T. Darrell. Scale-mae: A scale-aware masked autoencoder for multiscale geospatial representation learning. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 4065-4076, Los Alamitos, CA, USA, 2023. IEEE Computer Society. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.288, + 0.906, + 0.384 + ], + "angle": 0, + "content": "[39] Colorado J Reed, Ritwik Gupta, Shufan Li, Sarah Brockman, Christopher Funk, Brian Clipp, Kurt Keutzer, Salvatore Candido, Matt Uytendaele, and Trevor Darrell. Scale-mae: A scale-aware masked autoencoder for multiscale geospatial representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4088-4099, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.385, + 0.906, + 0.44 + ], + "angle": 0, + "content": "[40] Irving S Reed and Xiaoli Yu. Adaptive multiple-band cfar detection of an optical pattern with unknown spectral distribution. IEEE transactions on acoustics, speech, and signal processing, 38(10):1760-1770, 1990. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.906, + 0.509 + ], + "angle": 0, + "content": "[41] Swalpa Kumar Roy, Gopal Krishna, Shiv Ram Dubey, and Bidyut B Chaudhuri. Hybridsn: Exploring 3-d-2-d cnn feature hierarchy for hyperspectral image classification. IEEE Geoscience and Remote Sensing Letters, 17(2):277-281, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.511, + 0.906, + 0.579 + ], + "angle": 0, + "content": "[42] Pranab Sahoo, Ayush Kumar Singh, Sriparna Saha, Vinija Jain, Samrat Mondal, and Aman Chadha. A systematic survey of prompt engineering in large language models: Techniques and applications. arXiv preprint arXiv:2402.07927, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.581, + 0.906, + 0.635 + ], + "angle": 0, + "content": "[43] Bernhard Scholkopf, Robert C Williamson, Alex Smola, John Shawe-Taylor, and John Platt. Support vector method for novelty detection. Advances in neural information processing systems, 12, 1999. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.637, + 0.906, + 0.691 + ], + "angle": 0, + "content": "[44] Dunbin Shen, Xiaorui Ma, Wenfeng Kong, Jianjun Liu, Jie Wang, and Hongyu Wang. Hyperspectral target detection based on interpretable representation network. IEEE Transactions on Geoscience and Remote Sensing, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.692, + 0.906, + 0.734 + ], + "angle": 0, + "content": "[45] Hongjun Su, Zhaoyue Wu, Huihui Zhang, and Qian Du. Hyperspectral anomaly detection: A survey. IEEE Geoscience and Remote Sensing Magazine, 10(1):64-90, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.906, + 0.789 + ], + "angle": 0, + "content": "[46] Le Sun, Guangrui Zhao, Yuhui Zheng, and Zebin Wu. Spectral-spatial feature tokenization transformer for hyperspectral image classification. IEEE Transactions on Geoscience and Remote Sensing, 60:1-14, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.906, + 0.831 + ], + "angle": 0, + "content": "[47] Siyu Sun, Jun Liu, and Wei Li. Spatial invariant tensor self-representation model for hyperspectral anomaly detection. IEEE Transactions on Cybernetics, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[48] Xian Sun, Peijin Wang, Wanxuan Lu, Zicong Zhu, Xiaonan Lu, Qibin He, Junxi Li, Xue Rong, Zhujun Yang, Hao Chang, et al. Ringmo: A remote sensing foundation model with masked image modeling. IEEE Transactions on Geoscience and Remote Sensing, 61:1-22, 2022. 4" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23057" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.176 + ], + "angle": 0, + "content": "[49] Xian Sun, Peijin Wang, Wanxuan Lu, Zicong Zhu, Xiaonan Lu, Qibin He, Junxi Li, Xue Rong, Zhujun Yang, Hao Chang, Qinglin He, Guang Yang, Ruiping Wang, Jiwen Lu, and Kun Fu. Ringmo: A remote sensing foundation model with masked image modeling. IEEE Transactions on Geoscience and Remote Sensing, 61:1-22, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.178, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[50] Adam Van Etten, Dave Lindenbaum, and Todd M Bacastow. Spacenet: A remote sensing dataset and challenge series. arXiv preprint arXiv:1807.01232, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.302 + ], + "angle": 0, + "content": "[51] Adam Van Etten, Jacob Shermeyer, Daniel Hogan, Nicholas Weir, and Ryan Lewis. Road network and travel time extraction from multiple look angles with spacenet data. In IGARSS 2020-2020 IEEE International Geoscience and Remote Sensing Symposium, pages 3920-3923. IEEE, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.305, + 0.482, + 0.347 + ], + "angle": 0, + "content": "[52] Di Wang, Bo Du, and Liangpei Zhang. Fully contextual network for hyperspectral scene parsing. IEEE Transactions on Geoscience and Remote Sensing, 60:1-16, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[53] Di Wang, Qiming Zhang, Yufei Xu, Jing Zhang, Bo Du, Dacheng Tao, and Liangpei Zhang. Advancing plain vision transformer toward remote sensing foundation model. IEEE Transactions on Geoscience and Remote Sensing, 61:1-15, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.482, + 0.488 + ], + "angle": 0, + "content": "[54] Di Wang, Meiqi Hu, Yao Jin, Yuchun Miao, Jiaqi Yang, Yichu Xu, Xiaolei Qin, Jiaqi Ma, Lingyu Sun, Chenxing Li, et al. Hypersigma: Hyperspectral intelligence comprehension foundation model. arXiv preprint arXiv:2406.11519, 2024. 2, 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.482, + 0.559 + ], + "angle": 0, + "content": "[55] Shaoyu Wang, Xinyu Wang, Liangpei Zhang, and Yanfei Zhong. Auto-ad: Autonomous hyperspectral anomaly detection network based on fully convolutional autoencoder. IEEE Transactions on Geoscience and Remote Sensing, 60:1-14, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.482, + 0.644 + ], + "angle": 0, + "content": "[56] Sheng Wang, Kaiyu Guan, Chenhui Zhang, Qu Zhou, Sibo Wang, Xiaocui Wu, Chongya Jiang, Bin Peng, Weiye Mei, Kaiyuan Li, et al. Cross-scale sensing of field-level crop residue cover: Integrating field photos, airborne hyperspectral imaging, and satellite data. Remote Sensing of Environment, 285:113366, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.482, + 0.716 + ], + "angle": 0, + "content": "[57] Yanheng Wang, Danfeng Hong, Jianjun Sha, Lianru Gao, Lian Liu, Yonggang Zhang, and Xianhui Rong. Spectral-spatial-temporal transformers for hyperspectral image change detection. IEEE Transactions on Geoscience and Remote Sensing, 60:1-14, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.482, + 0.786 + ], + "angle": 0, + "content": "[58] Philine Warnke, Kerstin Cuhls, Ulrich Schmoch, Lea Daniel, Liviu Andreescu, Bianca Dragomir, Radu Gheorghiu, Catalina Baboschi, Adrian Curaj, Marjukka Parkkinen, et al. 100 radical innovation breakthroughs for the future. 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[59] Nicholas Weir, David Lindenbaum, Alexei Bastidas, Adam Van Etten, Sean McPherson, Jacob Shermeyer, Varun Kumar, and Hanlin Tang. Spacenet mvoi: A multi-view overhead imagery dataset. In Proceedings of the IEEE/cvf international conference on computer vision, pages 992-1001, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.903 + ], + "angle": 0, + "content": "[60] Walter F Wiggins and Ali S Tejani. On the opportunities and risks of foundation models for natural language processing in" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.092, + 0.905, + 0.119 + ], + "angle": 0, + "content": "radiology. Radiology: Artificial Intelligence, 4(4):e220119, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[61] Yinan Wu, Licheng Jiao, Xu Liu, Fang Liu, Shuyuan Yang, and Lingling Li. Domain adaptation-aware transformer for hyperspectral object tracking. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[62] Zhitong Xiong, Yi Wang, Fahong Zhang, Adam J Stewart, Joëlle Hanna, Damian Borth, Ioannis Papoutsis, Bertrand Le Saux, Gustau Camps-Valls, and Xiao Xiang Zhu. Neural plasticity-inspired foundation model for observing the earth crossing modalities. arXiv e-prints, pages arXiv-2403, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[63] Yang Xu, Zebin Wu, Jun Li, Antonio Plaza, and Zhihui Wei. Anomaly detection in hyperspectral images based on low-rank and sparse representation. IEEE Transactions on Geoscience and Remote Sensing, 54(4):1990-2000, 2015. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[64] Zongyin Yang, Tom Albrow-Owen, Weiwei Cai, and Tawfique Hasan. Miniaturization of optical spectrometers. Science, 371(6528):eabe0722, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[65] Alina Zare, Changzhe Jiao, and Taylor Glenn. Discriminative multiple instance hyperspectral target characterization. IEEE transactions on pattern analysis and machine intelligence, 40(10):2342-2354, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.419, + 0.905, + 0.488 + ], + "angle": 0, + "content": "[66] Hengwei Zhao, Xinyu Wang, Jingtao Li, and Yanfei Zhong. Class prior-free positive-unlabeled learning with taylor variational loss for hyperspectral remote sensing imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16827-16836, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.905, + 0.544 + ], + "angle": 0, + "content": "[67] Hengwei Zhao, Yanfei Zhong, Xinyu Wang, and Hong Shu. One-class risk estimation for one-class hyperspectral image classification. IEEE Transactions on Geoscience and Remote Sensing, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.546, + 0.905, + 0.613 + ], + "angle": 0, + "content": "[68] Zhuo Zheng, Yanfei Zhong, Ailong Ma, and Liangpei Zhang. Fpga: Fast patch-free global learning framework for fully end-to-end hyperspectral image classification. IEEE Transactions on Geoscience and Remote Sensing, 58(8):5612-5626, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.616, + 0.905, + 0.698 + ], + "angle": 0, + "content": "[69] Yanfei Zhong, Xin Hu, Chang Luo, Xinyu Wang, Ji Zhao, and Liangpei Zhang. Whu-hi: Uav-borne hyperspectral with high spatial resolution (h2) benchmark datasets and classifier for precise crop identification based on deep convolutional neural network with crf. Remote Sensing of Environment, 250:112012, 2020. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.698 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23058" + } + ] +] \ No newline at end of file diff --git a/2025/HyperFree_ A Channel-adaptive and Tuning-free Foundation Model for Hyperspectral Remote Sensing Imagery/9236da20-05f2-4327-b226-ca754d2dead5_origin.pdf b/2025/HyperFree_ A Channel-adaptive and Tuning-free Foundation Model for Hyperspectral Remote Sensing Imagery/9236da20-05f2-4327-b226-ca754d2dead5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..703d21f3b1a7862e91c2344c0b0ce2cceccae8c1 --- /dev/null +++ b/2025/HyperFree_ A Channel-adaptive and Tuning-free Foundation Model for Hyperspectral Remote Sensing Imagery/9236da20-05f2-4327-b226-ca754d2dead5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ff2426e9e0b7535b0acfbf5e67c975b6ec751ca9bf40db37dc14bb00dd204128 +size 3366841 diff --git a/2025/HyperFree_ A Channel-adaptive and Tuning-free Foundation Model for Hyperspectral Remote Sensing Imagery/full.md b/2025/HyperFree_ A Channel-adaptive and Tuning-free Foundation Model for Hyperspectral Remote Sensing Imagery/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4eb233cc03ce15d2ad0085815050a688a90b0f2d --- /dev/null +++ b/2025/HyperFree_ A Channel-adaptive and Tuning-free Foundation Model for Hyperspectral Remote Sensing Imagery/full.md @@ -0,0 +1,300 @@ +# HyperFree: A Channel-adaptive and Tuning-free Foundation Model for Hyperspectral Remote Sensing Imagery + +Jingtao Li\*, Yingyi Liu\*, Xinyu Wang†, Yunning Peng†, Chen Sun†, Shaoyu Wang², Zhendong Sun†, Tian Ke†, Xiao Jiang†, Tangwei Lu†, Anran Zhao†, Yanfei Zhong† + +{Jintaoli, liuyingyi, wangxinyu, pengyunning, sunchen, + +sunzhendong, ketian, jiangxiao, lutangwei, anranzha0415, zhongyanfei}@whu.edu.cn + +wsy1995@snu.ac.kr + +# Abstract + +Advanced interpretation of hyperspectral remote sensing images benefits many precise Earth observation tasks. Recently, visual foundation models have promoted the remote sensing interpretation but concentrating on RGB and multispectral images. Due to the varied hyperspectral channels, existing foundation models would face image-by-image tuning situation, imposing great pressure on hardware and time resources. In this paper, we propose a tuning-free hyperspectral foundation model called HyperFree, by adapting the existing visual prompt engineering. To process varied channel numbers, we design a learned weight dictionary covering full-spectrum from $0.4 \sim 2.5\mu m$ , supporting to build the embedding layer dynamically. To make the prompt design more tractable, HyperFree can generate multiple semantic-aware masks for one prompt by treating feature distance as semantic-similarity. After pre-training HyperFree on constructed large-scale high-resolution hyperspectral images, HyperFree (1 prompt) has shown comparable results with specialized models (5 shots) on 5 tasks and 11 datasets. Code and dataset are accessible at https://rsidea.whu.edu.cn/hyperfree.htm. + +# 1. Introduction + +Hyperspectral imaging has been recognized by the European Union as one of the Top 100 Disruptive Technologies [58], due to its ability to provide high-resolution spectral data for precise Earth observation across diverse applications [64], including environmental monitoring [31], agriculture [69], and defense [20]. Unlike traditional RGB or multispectral imagery, Hyperspectral imagery (HSI) offers several unique features: high spectral resolution (typically + +![](images/67039194eb890a9a9984274c3a3a021c040dff12b5c1be6bdd1b1d11382ef41d.jpg) +(a) Mainstream Pre-training and Tuning Paradigm of RSFM + +![](images/b431c2ad7d207f8a9bdd84e93394e2b6a672f38cb29bc8e63c08bd8841c309fa.jpg) +(b) Prompt Engineering of SAM-variants + +![](images/5629edb7992d3e8971b9a8db18f3f469069c80024c77191cdfecf3fb7355b68b.jpg) +(c) Prompt Engineering of HyperFree +Figure 1. Comparison of existing RSFMs with the proposed HyperFree model. (a) The pre-training and tuning paradigm is the mainstream approach, where fine-tuning is essential for handling unseen HSIs. (b) The prompt engineering paradigm reduces the tuning burden but struggles to handle HSIs with hundreds of spectral channels. (c) HyperFree can directly process unseen HSIs with varying channel counts without fine-tuning, and it can identify multiple masks of the same class using only a single prompt. + +$10\mathrm{nm}$ ), an extensive wavelength range (400–2500 nm), and dozens to hundreds of spectral channels [64], which enable the detailed identification of diverse targets, while also increasing the complexity of data processing and analysis. + +Recently, remote sensing foundation models (RSFMs) have demonstrated advanced interpretive capabilities, primarily focusing on multispectral and RGB images [4, 7, 13, 25, 38, 49], while also opening new possibilities for analyzing hyperspectral images. RSFMs follow two main + +paradigms. The predominant paradigm is pre-training and tuning (P-T), illustrated in Figure 1(a), which begins with training a large-scale backbone model in a self-supervised manner, such as masked autoencoders (MAE) [12], followed by fine-tuning with task-specific heads for various downstream applications [7, 13, 38, 49]. In the P-T paradigm, both the pre-training and the fine-tuning stages can be time-consuming. Second paradigm is prompt engineering (P-E), as illustrated in Figure 1(b), which focuses on designing task-specific prompts for a frozen Segment Anything Model (SAM) [17] to directly perform downstream tasks without additional model tuning [4, 25], making it more efficient and user-friendly in practical. + +The P-E paradigm is particularly well-suited for processing HSI, as it can be directly applied to diverse, unseen scenes without requiring model fine-tuning. However, it faces two significant challenges when applied to HSI. (i) The first challenge is the variation in spectral parameters across different hyperspectral sensors, including differences in spectral resolution, wavelength range, and the number of spectral channels [10]. For a P-T-based model like HyperSigma [54], although it is pre-trained using HSIs with fixed channels, the challenge can be addressed during its tuning stage by rebuilding the channel embedding layer to align the dimensions with the new data. But for a P-E-based model, the challenge of channel variation requires more careful consideration. (ii) The second challenge is to make prompt input tractable, ensuring a P-E based model can directly process unseen HSIs with least prior knowledge. Specifically, most existing P-E-based RSFMs rely on a frozen SAM [17], which segments remote sensing objects based on specialized models to generate prompts [4, 25]. As shown in Figure 1(b), each prompt typically generates a single mask and all the prompts for target objects are needed by existing P-E-based RSFMs, which largely reduced the model feasibility. + +To address the above challenges, as shown in Figure 1 (c), we propose a channel-adaptive, tuning-free hyperspectral foundation model, called HyperFree, which can process any unseen HSIs directly in a P-E manner. Specifically, (i) Motivated by the word dictionary in natural language processing, which stores vector embeddings as values corresponding to words and phrases of varying lengths as keys [32], HyperFree introduces a learnable wavelength weight dictionary in the embedding layer, spanning the 400 to 2500 nm spectral range in $10\mathrm{nm}$ intervals. This allows for the dynamic encoding of unseen HSIs with varying spectral parameters into visual tokens, all with consistent dimensionality. (ii) To make the prompt of HyperFree more tractable, HyperFree is designed to generate multiple semantic-aware masks for one prompt as in Figure 1 (c), where HyperFree maps each prompt and single mask from image space into feature space and treats the feature distance as semantic- + +similarity. Prompt, mask and feature are interacted adaptively for different downstream task in tuning-free manner with only one prompt or zero shot. + +Due to the high acquisition cost and label difficulty, current hyperspectral datasets only have a few images [45, 69], far from the requirements to train the promptable HyperFree. To conquer this problem, we developed the Hyper-Seg data engine, in addition to utilizing well-known multispectral datasets (e.g., fMoW [5] and SpaceNet [50, 51, 59]). This engine automatically generates a large-scale annotated HSI dataset, consisting of nearly 50,000 pairs of HSIs and their corresponding segmentation masks. The original data are sourced from AVIRIS airborne hyperspectral sensors, with a spatial size of $512 \times 512$ pixels and 224 spectral channels spanning from 400 to $2500\mathrm{nm}$ at a $10\mathrm{nm}$ spectral resolution. Unlike other RSFMs trained on unannotated satellite hyperspectral imagery with a $30\mathrm{m}$ spatial resolution [54], the Hyper-Seg data engine features higher spatial resolution (0.6 to $5.0\mathrm{m}$ ), along with annotations of 15.44 million masks. It is currently the largest known hyperspectral dataset with high spatial resolution. + +To demonstrate the model capabilities, we evaluated HyperFree directly in a P-E manner on 11 datasets across 5 segmentation-related tasks, including multi-class classification [21], one-class classification [66], target detection [65], anomaly detection [47], and change detection [26]. Additionally, we verified HyperFree in a P-T manner on 14 datasets spanning 8 tasks, with additional tasks including hyperspectral denoising [27], hyperspectral unmixing [37], and hyperspectral object tracking [61]. The experimental results show that HyperFree achieves comparable accuracy to state-of-the-art models in a tuning-free manner (using just one prompt), and outperforms most models after tuning. + +Our contributions can be summarized as: + +- We propose the first tuning-free hyperspectral foundation model, which can process any hyperspectral image in different tasks with promptable or zero-shot manner. +- A weight dictionary that spans the full spectrum, enabling dynamic generation of the embedding layer dynamically according to input wavelengths. +- We propose to map both prompts and masks into feature space to identify multiple semantic-aware masks for one prompt, where different interaction workflows are designed for each downstream tasks. +- We built the Hyper-Seg data engine to train the HyperFree model and tested it on 11 datasets from 5 tasks in tuning-free manner, 14 datasets from 8 tasks in tuning manner as an extensive experiment. + +# 2. Related Work + +Hyperspectral image processing. There are diverse tasks to process such precise observation with different targets + +[31]. (a) Classification: Assigning a specific class to each pixel [21]. (b) One class classification: Discriminating the pixels of target class with labeled binary map [33]. (c) Target detection. Discriminating the pixels of target class with labeled spectrum. (d) Anomaly detection. Discriminating the pixels spectrally deviating from the background [24]. (e) Change detection: Discriminating the changing pixels between two time-steps [26]. Since the output of the above tasks is all pixel-level segmentation maps, it is feasible to unify them using the proposed HyperFree model in tuning-free manner. + +Foundation models. Powered by the transformer architectures, foundation models are characterized by large-scale property in model size and pre-training data, and the transferring ability for downstream tasks [60]. In remote sensing community, some foundation models for multi-spectral images have already been proposed such as SatMAE [7], SepctralEarth [1] and SpectralGPT [13], which trains large ViT-like models [9] with the MAE strategy [12] and massive open-source multi-spectral images (13 channels). In contrast, hyperspectral images have a high acquiring cost [56] and make the large-scale pre-training stage more difficult. A recent study broke the situation by building a largescale hyperspectral dataset HyperGlobal-450K and trained a foundation model called HyperSigma with MAE strategy as well [54]. Our work differs from HyperSigma in two aspects: (i) HyperSigma accepts fixed number of image channels, while HyperFree can deal with any number of channels directly. (ii) HyperSigma needs to be fine-tuned for each downstream task. Differently, HyperFree has ability to process different tasks in fine-tuning free manner. + +Prompt engineering (P-E). Different from the pretraining and then fine-tuning paradigm, P-E can utilize pretrained foundation models to complete downstream tasks directly with prompts [34]. For pre-trained large language models, prompts are mostly hand-crafted text to guide the model generate response for specific task [34, 42]. In computer vision community, SAM follows the same spirit and designs diverse prompts (e.g., point and box) to achieve few-shot transferring performance [17]. Our work extends the visual prompt engineering in hyperspectral data with varying image channels and supports five tasks in tuning-free manner with designed prompt-mask-feature interaction. + +# 3. Hyper-Seg Data Engine + +Hyperspectral image processing always refers to various dense tasks and a large-scale segmentation dataset is needed to apply the P-E. However, most hyperspectral segmentation datasets only have a few images due to the high acquisition and label cost, which is quite different from the million-scale dataset with natural images such as SA-1B + +[17]. To tackle this, we built a data engine called HyperSeg to generate segmented masks automatically for spectral images and expand the data scale. Hyper-Seg engine firstly separates the spectral image into several groups with three channels according to the prior of key channels position. Each group is then segmented by the SAM-H model [17] and the results are combined by NMS operation to eliminate redundancy and output the final segmentation map. Figure 2 shows the overall workflow, where the nine key channels are selected according to the expert knowledge as in Supp.1.1. Considering the high spatial resolution property of airborne platform, we collected 41946 hyperspectral images from AVIRIS airborne sensor and applied the Hyper-Seg Engine. Besides, we also processed two existing multispectral datasets (fMoW [5] and SpaceNet [50, 51, 59]) to increase the scale further. Compared to the traditional hyperspectral annotation methods which refer to high-resolution RGB imagery, our data engine can utilize more spectral information and obtain a high-precision segmentation mask without labor cost. + +Table 1 shows the comparable statistical information of training datasets between existing RSFMs and proposed HyperFree. Most models only concentrate on the RGB and multispectral images with channel number 3 13. Although the recent model HyperSigma [54] has collected a large-scale hyperspectral images, the spatial resolution (30m/pixel) is largely lower than us (0.6 5m/pixel), and the dataset scale (considering image size and number together) is relatively small. With the designed Hyper-Seg data engine, constructed dataset is the only one with segmentation masks supporting the promptable training. More detailed statistical information of our generated dataset is given in Supp.1.2. The final generated dataset has nearly 150k images and 15.44million masks. We name both the data engine and generated dataset as the Hyper-Seg for simplicity. + +# 4. HyperFree Foundation Model + +HyperFree has two customized components upon the metaarchitecture of SAM including channel-adaptive embedding and prompt-mask-feature (PMF) interaction as in Figure 3. Channel-adaptive embedding constructs the embedding layer dynamically according to varying input wavelength, and outputs tokens in same representation space. After the promptable training stage, PMF interaction discriminates the semantic-aware masks for downstream tasks by mapping both prompt and mask into feature space and treating feature distance as semantic similarity. + +# 4.1. Channel-adaptive Embedding + +To make HyperFree process any hyperspectral image, channel-adaptive embedding layer is firstly designed to map input image with varying channels into tokens with fixed length, inspired by the tokenizer in NLP community [32]. + +![](images/428559adf50a8595acfdcb565789207c51467f6398d40bb9b5b971b5ed5f292f.jpg) +Figure 2. Hyper-Seg Data Engine automatically generates a large-scale dataset comprising nearly 41,900 pairs of airborne HSIs and their corresponding segmentation masks, enabling HyperFree to effectively learn discriminative spectral features. + +Table 1. Comparison of training data used in current RSFMs. The primary features of the proposed HyperFree model include prompt engineering and the use of high-resolution annotated hyperspectral (HRS) imagery. + +
DatasetFoundation ModelSpatial Resolution (m)Spectral Resolution (nm)Channel NumberImage SizeImage NumberMain ModalityTraining Paradigm
RGBRVSA [53]0.5 ~ 153\3110 ~ 316721.00 × 10^6H/M/LSR-RGBP-T
RingMo [48]0.1 ~ 30\34482.10 × 10^6H/MSR-RGBP-T
ScaleMAE [39]0.31 ~ 3.7\3\3.64 × 10^6HSR-RGBP-T
MultispectralSatMAE [6]1015 ~ 18013≈ 45 × 607.13 × 10^5MSR-MP-T
SkySense [11]10/2015 ~ 18010\\MSR-MP-T
SpectralGPT [14]10/20/6015 ~ 1801245 ~ 1201.47 × 10^6MSR-MP-T
HyperspectralHyperSigma [54]305/6.5/10150/175644.47 × 10^5LSR-HP-T
HyperFree0.6 ~ 5.0102245124.19 × 10^4HSR-HP-E
+ +* P-T: Pretraining and Tuning; P-E: Prompt Engineering. +* HSR: High Spatial Resolution; MSR: Medium Spatial Resolution; LSR: Low Spatial Resolution. +* RGB: Red-Green-Blue Imagery; M: Multispectral Imagery; H: Hyperspectral Imagery + +Tokenizers rely on a pre-defined dictionary to break down a string of text into numerical tokens, where the dictionary has stored the numerical vectors of varied-length words or phrases. Vision transformers always use a convolutional layer to generate patch tokens and what if we construct a layer weight dictionary treating each wavelength as a word? Given the input meta data of wavelengths, we can just loop up the corresponding weights and construct the layer dynamically without re-training it. + +The weight dictionary is built depending on the wavelengths to be processed and the expected token dimension. Since hyperspectral imaging mostly captures the spectral signals from $400\mathrm{nm}$ to $2500\mathrm{nm}$ and the $10\mathrm{nm}$ is of high spectral resolution currently, 221 index channels are designed as the dictionary keys. Denote the output token as a vector of $\mathbb{R}^j$ corresponding to the patch size $p\times p$ , each index channel stores a weight matrix of size $\mathbb{R}^{p\times p\times j}$ , and the built dictionary $\beta$ has a size of $221\times p\times p\times j$ . Given the input hyperspectral image $\mathbb{R}^{h\times w\times n}$ with wavelengths $\mathbf{b} =$ + +$[b_1, b_2, \ldots, b_n]$ , the corresponding kernel weight can be dynamically generated by combining the searched weight $\mathbf{W}$ with size $n \times p \times p \times j$ , where $n$ is flexible for each image. For conciseness, we use function $g$ to represent the mapping process from $\mathbf{b}$ to $\mathbf{W}$ as kernel weight, and function $f$ to represent the token generation step as in Equation (1). + +$$ +f (\mathbf {X} \mid \mathbf {b}, \beta) = \operatorname {C o n v} _ {\mathbf {W}} (\mathbf {X}) \quad \text {w h e r e} \quad \mathbf {W} = g (\mathbf {b}, \beta) \tag {1} +$$ + +In the past few decades, many satellites have been launched equipped with spectral sensors, where the key channels are selected by professional teams and practical experience. To utilize these successful priors, we further extract visual tokens in two parallel branches with dynamically constructed layers: one to process key channels and the other to process the split cubes, as shown in Figure 3. The input image $\mathbf{X}$ is first divided into key channels $\mathbf{X}_{\mathbf{k}}$ and the cubes $\mathbf{X}_{\mathbf{c}}$ between them, where the key wavelengths are selected by referring to famous spectral sensors (detailed + +![](images/479ad1de12bb655da0da7ead4673adc2cf106098b2345ef3da7ea96e200b9d60.jpg) +(a) Full spectrum Promptable Training + +![](images/4fdc8d8960c1438dd44057abbb04840e2830b01dbba8002eec3d5c77c2e1cc82.jpg) +(b) Prompt-mask-feature Interactive Inferring +Figure 3. (a) Full spectrum Promptable Training. Based on the meta-architecture of promptable segmentation, channel-adaptive embedding layer is developed to process any-channel input, which relies on a learnable weight dictionary and generates the embedding layer dynamically according to the input wavelengths. With the constructed Hyper-Seg dataset, HyperFree is trained with randomly selected channels from full spectrum and point prompts in each iteration. (b) Prompt-mask-feature Interactive Inferring. Keep the trained HyperFree frozen, different PMF interaction workflows are designed to complete the downstream tasks in promptable or zero-shot manner, where we use feature distance to measure the semantic consistency. An interaction instantiation for classification task is given above as an example. + +in Supp.2). We set two different dictionaries $\beta_{\mathbf{k}}$ and $\beta_{\mathbf{c}}$ to process $\mathbf{X}_{\mathbf{k}}$ and $\mathbf{X}_{\mathbf{c}}$ respectively, and the final output tokens $\mathbf{T} \in \mathbb{R}^{(h/p) \times (w/p) \times j}$ are computed as the element-wise sum, as shown in Equation (2). + +$$ +\mathbf {T} = f \left(\mathbf {X} _ {\mathbf {k}} \mid \mathbf {b} _ {\mathbf {k}}, \beta_ {\mathbf {k}}\right) + f \left(\mathbf {X} _ {\mathbf {c}} \mid \mathbf {b} _ {\mathbf {c}}, \beta_ {\mathbf {c}}\right) \tag {2} +$$ + +# 4.2. Full-spectrum Promptable Training + +The prompt engineering paradigm is adopted to complete downstream tasks directly with designed prompts. Hyper-Free is thus trained in a promptable manner on constructed large-scale Hyper-Seg dataset, where the prompts are randomly chosen at each iteration and the model is expected to segment correctly the corresponding masks. Since hyperspectral image processing tasks are almost conducted at the pixel level, we choose point prompts only at the training stage. Focal loss $L_{f}$ and Dice loss $L_{d}$ are weighted as [17] to supervise the training as Equation (3). + +$$ +L = 2 0 L _ {f} + L _ {d} \tag {3} +$$ + +To train the channel-adaptive embedding layer effectively, we select the subset of channels randomly from the full-spectrum range and process it with corresponding weights only in $\beta_{\mathbf{k}}$ and $\beta_{\mathrm{c}}$ , which helps the learned dictionary being wavelength-aware and encode the varied channel composition into a unified token space. + +# 4.3. Prompt-mask-feature (PMF) Interactive Inferring + +After training process, the model can process unseen hyperspectral image with varied channels and segment the corresponding mask given any point prompt. However, many downstream tasks need to segment out all the masks of same semantic category rather than a single mask such as hyperspectral classification and target detection. We tackle this problem by connecting the prompt, mask and feature together in feature space, where prompt tells the model an object reference in task, mask outputs accurate location information, and feature reflects semantic information. In unified feature space, the cosine distance of features is used to measure the semantic similarity. + +Before diving into the specific workflows, we introduce two basic interaction modes with a simple situation. Set one point $(x,y)$ as input prompt for image $\mathbf{X}$ , the corresponding feature cube $\mathbf{D}$ and all the generated masks $\mathbf{M} \in \mathbb{R}^{h \times w \times k}$ from the trained model, where $\mathbf{M}$ has $k$ binary maps and each represents a single mask. The first interaction mode maps the prompt $(x,y)$ into a feature vector $\mathbf{d}_{(x,y)} \in \mathbb{R}^j$ as in Equation (4), where the corresponding valid and smallest mask in $\mathbf{M}$ is used to localize the features in $\mathbf{D}$ , and the mean feature is treated as $\mathbf{d}_{(x,y)}$ for the given prompt. This interaction transforms the prompt from image space to feature space, and makes it possible to explore the semantic + +similarity. + +$$ +\text {M o d e} (x, y) \mid \mathbf {M}, \mathbf {D} \rightarrow \mathbf {d} _ {(x, y)} \tag {4} +$$ + +The second interaction mode selects multiple the masks in $\mathbf{M}$ , which have similar semantic information with $\mathbf{d}_{(x,y)}$ as in Equation (5). The selected masks are represented as $\mathbf{M}_d$ and $\mathbf{M}_d \subseteq \mathbf{M}$ . Specifically, we compute a feature representation for each single mask by localizing in $\mathbf{D}$ and computing the average, and the cosine similarity between them and $\mathbf{d}_{(x,y)}$ are treated as the semantic similarity. With some hand-designed thresholds $\tau$ , $\mathbf{M}_d$ can be finally segmented. $\tau$ represents different meanings according to the specific task. + +$$ +\text {M o d e} 2: \left(\mathbf {d} _ {(x, y)} | \mathbf {M}, \mathbf {D}, \tau\right)\rightarrow \mathbf {M} _ {d} \tag {5} +$$ + +By using both interaction modes adaptively, five hyperspectral tasks can be completed with the trained promptable model directly. (a) Hyperspectral classification (HC). Each category needs at least one prompt, and uses Mode 1 and Mode 2 in turn for each category. For a common closed-set setting, $\tau$ is not necessary since we can use the minimum feature distance between each mask representation and all the prompt representations to decide its category. Figure 3 (b) shows an exemplified workflow. (b) Hyperspectral one-class classification (HOCC). HOCC is similar to H but the difference lies in that HOCC only needs prompts the target category, and its $\tau$ is instantiated as a class pr parameter to control the mask ratio. (c) Hyperspectral taget detection (HTD). Compared to HOCC, HTD acce the target spectra and it needs to be changed to a pron by selecting the pixel with the smallest cosine distance. is instantiated as a distance threshold to convert the dens map to a binary map. (d) Hyperspectral anomaly detion (HAD). HAD is an unsupervised task and there is $\mathbf{d}_{(x,y)}$ for HAD. Since HAD is very similar to unsupervis target detection, we use M only and filter out large objec according to mask area to output the final anomaly map. + +Hyperspectral change detection (HCD). Different from the prior tasks, HCD accepts two registered and temporal images to identify the changing locations. The HCD task utilizes both interaction modes for the bi-temporal images. Mode 1 processes the first image to get representation $\mathbf{d}_{(x,y)}$ + +Table 2. Instantiation summary of prompts, meaning of $\tau$ , and interaction modes when HyperFree performs different tasks. + +
TaskPrompt TypeHyperparameter τInteraction Mode
HCOne Point Per CategoriesNo τMode 1, Mode 2
HOCCTarget Category PointClass PriorMode 1, Mode 2
HTDTarget SpectraSimilarity ThresholdMode 1, Mode 2
HADNo PromptMasks Area ThresholdMode 1
HCDNo PromptSimilarity ThresholdMode 1, Mode 2
+ +![](images/c8766bfd9c3a124d04b58db1c422d193da50f8ee338e2709fe18ccf539f45221.jpg) + +![](images/8aa3a5b03f95cc7bde125587c9649d741960c2d44980215513318b6be995fecf.jpg) +Figure 4. Exemplified comparison results for five tuning-free tasks +Figure 5. In tuning-free manner, HyperFree (1 prompt) can achieve comparable results with specialized models (5 shots) on 5 tasks and 11 datasets. + +for each mask, which is input to Mode 2 and compared with the corresponding mask representation with $\mathbf{M}$ and $\mathbf{D}$ of the second image. Once the temporal feature distance exceeds the preset $\tau$ , a change is considered to have happened. The interaction workflow is summarized in Table 2. + +# 5. Results + +# 5.1. Experimental Setup + +We evaluate the proposed HyperFree on five downstream hyperspectral tasks including HC, HOCC, HTD, HAD and + +Table 3. Quantitative comparison results on five hyperspectral tasks, where blue numbers indicate the metric ranking of HyperFree. Comparison methods were all trained on each dataset while HyperFree processed the datasets in tuning-free manner. + +
HC Task
LongKou (270 channels)HanChuan (274 channels)
MetricsSVM [30] (5 shot)HybridSN [41] (5 shot)FCN [52] (5 shot)FPGA [68] (5 shot)SSFTT [46] (5 shot)MambaHSI [23] (5 shot)HyperFree (1 prompt)SVM [30] (5 shot)HybridSN [41] (5 shot)FCN [52] (5 shot)FPGA [68] (5 shot)SSFTT [46] (5 shot)MambaHSI [23] (5 shot)HyperFree (1 prompt)
OA82.7748.7886.6791.1889.6692.6593.39(1st)52.6847.7555.5571.4764.8673.3375.47(1st)
AA74.0261.3785.6088.3587.9692.5788.03(3rd)47.7646.1759.7272.0961.2269.3362.35(2nd)
KA78.0435.7282.3088.6687.9590.0091.41(1st)46.8541.3150.1867.5859.6569.1071.56(1st)
HOCC Task
HongHu (270 channels)XiongAn (256 channels)
MetricsOCSVM [43] (5 shot)nnPU [18] (5 shot)BSVM [33] (5 shot)PAN [15] (5 shot)OC Loss [67] (5 shot)T-HOneCls [66] (5 shot)HyperFree (1 prompt)OCSVM [43] (5 shot)nnPU [18] (5 shot)BSVM [33] (5 shot)PAN [15] (5 shot)OC Loss [67] (5 shot)T-HOneCls [66] (5 shot)HyperFree (1 prompt)
F26.3319.1334.8263.6954.7355.9772.52(1st)18.311.7626.3046.3443.0841.3452.50(1st)
P56.4319.7250.7780.0058.2746.5268.61(2nd)39.832.8523.8247.1347.5032.8768.21(1st)
R24.0218.5845.2964.2754.3492.3578.06(2nd)16.081.9857.8353.3247.6160.3865.03(1st)
HTD Task
Airport (205 channels)Cri (46 channels)
MetricsACE [19] (1 shot)CEM [2] (1 shot)GLRT [28] (1 shot)MF [29] (1 shot)HTD-IRN [44] (1 shot)TSTTD [16] (1 shot)HyperFree (1 prompt)ACE [19] (1 shot)CEM [2] (1 shot)GLRT [28] (1 shot)MF [29] (1 shot)HTD-IRN [44] (1 shot)TSTTD [16] (1 shot)HyperFree (1 prompt)
AUC(D,F)0.97940.96030.98010.99160.97450.99290.9806(3rd)0.97350.98930.97370.98910.99750.99870.9791(5th)
AUCODP1.58531.28291.57981.69681.44841.65921.4667(6th)1.20151.45061.20001.45751.39951.61031.4670(2nd)
HAD Task
Beach1 (188 channels)Beach2 (193 channels)
MetricsRXD [40]CRD [22]ADLR [36]LRASR [63]Auto-AD [55]TDD [20]HyperFree (zero shot)RXD [40]CRD [22]ADLR [36]LRASR [63]Auto-AD [55]TDD [20]HyperFree (zero shot)
AUC(D,F)0.98150.94710.45150.74610.95740.98420.9730(3rd)0.90900.85440.79760.82250.94850.96270.9566(2nd)
AUCODP1.25570.97850.56100.85261.12731.13831.9191(1st)1.01770.86700.90640.82801.00971.16881.8697(1st)
HCD Task
River (154 channels)Hermiston (198 channels)
MetricsFC-EF [8] (5 shot)FC-Sc [8] (5 shot)FC-Sd [8] (5 shot)ML-EDAN [35] (5 shot)BIT [3] (5 shot)SST-Former [57] (5 shot)HyperFree (zero shot)FC-EF [8] (5 shot)FC-Sc [8] (5 shot)FC-Sd [8] (5 shot)ML-EDAN [35] (5 shot)BIT [3] (5 shot)SST-Former [57] (5 shot)HyperFree (zero shot)
IoU0.41680.45220.45340.39150.21260.40960.3649(6th)0.37290.37760.48730.32520.52570.53610.5851(1st)
F0.58840.62280.62390.56280.35070.58120.5347(6th)0.54320.54820.65.520.49080.68910.69800.7382(1st)
+ +Table 4. Compared results with SAM using the identical PMF interaction workflows. + +
HCHOCCHTDHADHCD
OAAAKAFPrecisionRecallAUC(D,F)AUCODPAUC(D,F)AUCODPFIoU
SAM+PMF Interaction66.6272.9959.4134.1330.3539.470.79791.10140.96151.88440.59960.4282
HyperFree93.4085.6691.4272.5268.6178.060.97911.46700.95661.86970.73820.5851
+ +HCD tasks as in Section 4.3. A total of 11 datasets were used in the evaluation (detailed in Supp.3), and each task had six comparison methods, including both classical and state-of-the-art ones. Since there is no method that can directly complete all five tasks, all comparative methods are tested after being trained on each dataset with five shots per class, while the parameters of HyperFree are fixed for all datasets with 1 prompt or zero-shot as in Table 2. We use ViT-base as backbone to conduct the experiments. + +HyperFree was trained on the constructed Hyper-Seg dataset using 8 A100 GPUs for approximately 10 hours, which was based on the meta-architecture and parameters of SAM. We use AdamW optimizer with initial learning rate 8e-5 and batch size 16. To help learning the channel-adaptive embedding, we randomly select the image channels at each iteration and match it with corresponding wavelengths. 16 masks and 1 2 points per mask are randomly selected for each input image to train + +the promptable segmentation. + +# 5.2. Main Results + +Most tuning-free quantitative results are reported in Table 3 (1 dataset in supplement due to the space limit) and Figure 5. All the comparative methods are trained with 5 shot for each class and HyperFree processes each dataset directly without tuning. For HC, HOCC and HTD, only one point prompt is provided for each class. HAD and HCD does not need any prompt and deal with images in zero-shot manner. Table 3 shows that HyperFree can surpass most of the specialized models with frozen parameters. On both HC and HCD tasks, HyperFree even outperformed previous SOTA results, including an improvement of about 12 points in $\mathbf{F}_1$ on HC and about 0.5 points in IoU on HCD task. + +We performed qualitative comparison on five exemplified images in Figure 4, where one SOTA model have been visualized for each task comparison. Obviously, the result maps of HyperFree are more complete and smoother than the ones of specialized models. In HC, HOCC, HTD and + +Table 5. Ablation results on the usage of the learned weight dictionary. + +
Correct WavelengthsβkβcHCHOCCHTDHADHCD
OAAAKAF1PrecisionRecallAUC(D,F)AUCODPAUC(D,F)AUCODPF1IoU
✓✓×73.1864.2365.8565.6247.3064.560.97671.46440.94881.84650.57930.4078
✓×✓66.6371.5259.4463.6258.6178.440.97681.46400.95891.87680.70220.5411
×✓✓75.2538.9866.4457.3064.0954.830.95951.42730.95571.86780.68320.5188
✓✓✓93.4085.6691.4272.5268.6178.060.97911.46700.95661.86970.73820.5851
+ +![](images/a6d5d6b9a88aab5b71db603328a561f0ef200b15e536b0c55767951a47248281.jpg) +(a) HyperSigma features + +![](images/3b07dd51824e56815c3a355ffbb7bba73269f2f710e0873a36aca930a8712141.jpg) +(b) HyperFree features + +![](images/a9efefd3751bea9223204b28fa6a9e2fb9ad4f64cfaadf5bf1fa51d6a1d30c13.jpg) +Figure 6. t-SNE visualizations of HyperSigma [54] and HyperFree features on WHU-Hi Longkou dataset. We randomly select 1000 pixels for each class. +Figure 7. In tuning manner, HyperFree can achieve overall better results with specialized models on 8 tasks and 14 datasets. + +HAD tasks, HyperFree can generate basically the same result as GT, which are difficulty for existing models. + +# 5.3. Analysis Studies + +HyperFree vs SAM. Proposed prompt-mask-feature interaction strategy adapts the pre-trained HyperFree to various downstream tasks. We applied the same strategy to SAM to verify the effectiveness of training on hyperspectral images and reported the results in Table 4. The metrics on most tasks have a decline of about $10\sim 30$ points, implying a failed processing. We realize it is not a fair comparison and only use it to reflect the difference of hyperspectral data processing. + +HyperFree vs HyperSigma. HyperSigma is the only open-sourced hyperspectral foundation model [54, 62], which adopts the pre-training and tuning paradigm. Thus, we cannot compare HyperFree with it in tuning-free set- + +ting directly and compare the t-SNE visualization in Figure 6(a) and Figure 6(b). In contrast, proposed HyperFree has a more compact intra-class distance obviously, supporting the tuning-free workflow of PMF interaction. + +Full-Spectrum Weight Dictionary. To verify stored weights are wavelength-aware, we conducted related ablation studies in Table 5. In HC, HOCC and HCD tasks, a drop in accuracy of about $5\sim 40$ points can be observed if we shuffled the correct wavelength order randomly, showing the wavelength-aware property of learned dictionary. Besides, we proved the built two encoding branches with dictionaries $\beta_{\mathbf{k}}$ and $\beta_{\mathrm{c}}$ can benefit each other effectively. + +Tuning HyperFree. As an extensive experiment, we have also tested the tuning performance of HyperFree on 8 tasks and 14 datasets in Figure 7, which further included hyperspectral denosing (HD) [27], object tracking (HOT) [61] and unmxing (HU) tasks [37]. The tuning HyperFree has surpassed the previous SOTA models on nearly all tasks. + +More analysis. For more analytical experiments, please see the supplementary material. (i) Supp.4.1: Complete qualitative comparison of 11 datasets in tuning-free manner. (ii) Supp.4.2: Complete quantitative and qualitative comparison of 14 datasets in tuning manner. (iii) Supp.4.3: Sensitivity analysis about prompt number, prompt location and the hyperparameter $\tau$ . + +# 6. Conclusion + +Hyperspectral images can provide rich spectral information and are widely used in resource monitoring and defense fields. However, its unique properties increase the difficulty of foundation model development such as varied-length channels, high sensitivity to imaging conditions and the acquirement difficulty, making the mainstream P-T being low cost-effectiveness due to the image-by-image tuning. Our work tackles this problem by converting to adapt the visual P-E [17]. Differently, we design a learnable weight dictionary covering full-spectrum to generate the channel-adaptive embeddings and PMF interaction to generate semantic-aware masks for different tasks. We hope HyperFree can speed up the era of "GPT moment" for hyperspectral intelligent processing. + +# Acknowledgement + +# References + +[1] Nassim Ait Ali Braham, Conrad M Albrecht, Julien Mairal, Jocelyn Chanussot, Yi Wang, and Xiao Xiang Zhu. Spectral earth: Training hyperspectral foundation models at scale. arXiv preprint arXiv:2408.08447, 2024. 3 +[2] Chein-I Chang. Constrained energy minimization (cem) for hyperspectral target detection: Theory and generalizations. IEEE Transactions on Geoscience and Remote Sensing, 2024. 7 +[3] Hao Chen, Zipeng Qi, and Zhenwei Shi. Remote sensing image change detection with transformers. IEEE Transactions on Geoscience and Remote Sensing, 60:1-14, 2021. 7 +[4] Keyan Chen, Chenyang Liu, Hao Chen, Haotian Zhang, Wenyuan Li, Zhengxia Zou, and Zhenwei Shi. Rsprompter: Learning to prompt for remote sensing instance segmentation based on visual foundation model. IEEE Transactions on Geoscience and Remote Sensing, 2024. 1, 2 +[5] Gordon Christie, Neil Fendley, James Wilson, and Ryan Mukherjee. Functional map of the world. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6172-6180, 2018. 2, 3 +[6] Yezhen Cong, Samar Khanna, Chenlin Meng, Patrick Liu, Erik Rozi, Yutong He, Marshall Burke, David Lobell, and Stefano Ermon. Satmae: Pre-training transformers for temporal and multi-spectral satellite imagery. Advances in Neural Information Processing Systems, 35:197-211, 2022. 4 +[7] Yezhen Cong, Samar Khanna, Chenlin Meng, Patrick Liu, Erik Rozi, Yutong He, Marshall Burke, David B. Lobell, and Stefano Ermon. Satmae: pre-training transformers for temporal and multi-spectral satellite imagery. In Proceedings of the 36th International Conference on Neural Information Processing Systems, Red Hook, NY, USA, 2024. Curran Associates Inc. 1, 2, 3 +[8] Rodrigo Caye Daudt, Bertr Le Saux, and Alexandre Boulch. Fully convolutional siamese networks for change detection. In 2018 25th IEEE international conference on image processing (ICIP), pages 4063-4067. IEEE, 2018. 7 +[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 3 +[10] Alexander FH Goetz, Gregg Vane, Jerry E Solomon, and Barrett N Rock. Imaging spectrometry for earth remote sensing. science, 228(4704):1147-1153, 1985. 2 +[11] Xin Guo, Jiangwei Lao, Bo Dang, Yingying Zhang, Lei Yu, Lixiang Ru, Liheng Zhong, Ziyuan Huang, Kang Wu, Dingxiang Hu, et al. Skysense: A multi-modal remote sensing foundation model towards universal interpretation for + +This work was supported by the National Natural Science Foundation of China under Grant No. 42325105 and Grant 424B2010. +earth observation imagery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27672-27683, 2024. 4 +[12] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 2, 3 +[13] Danfeng Hong, Bing Zhang, Xuyang Li, Yuxuan Li, Chenyu Li, Jing Yao, Naoto Yokoya, Hao Li, Pedram Ghamisi, Xiuping Jia, Antonio Plaza, Paolo Gamba, Jon Atli Benediktsson, and Jocelyn Chanussot. Spectralgpt: Spectral remote sensing foundation model. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(8):5227-5244, 2024. 1, 2, 3 +[14] Danfeng Hong, Bing Zhang, Xuyang Li, Yuxuan Li, Chenyu Li, Jing Yao, Naoto Yokoya, Hao Li, Pedram Ghamisi, Xi-uping Jia, et al. Spectralgpt: Spectral remote sensing foundation model. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 4 +[15] Wenpeng Hu, Ran Le, Bing Liu, Feng Ji, Jinwen Ma, Dongyan Zhao, and Rui Yan. Predictive adversarial learning from positive and unlabeled data. In Proceedings of the AAAI conference on artificial intelligence, pages 7806-7814, 2021. 7 +[16] Jinyue Jiao, Zhiqiang Gong, and Ping Zhong. Triplet spectral-wise transformer network for hyperspectral target detection. IEEE Transactions on Geoscience and Remote Sensing, 2023. 7 +[17] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 2, 3, 5, 8 +[18] Ryuichi Kiryo, Gang Niu, Martinus C Du Plessis, and Masashi Sugiyama. Positive-unlabeled learning with nonnegative risk estimator. Advances in neural information processing systems, 30, 2017. 7 +[19] Shawn Kraut, Louis L Scharf, and L Todd McWhorter. Adaptive subspace detectors. IEEE Transactions on signal processing, 49(1):1-16, 2001. 7 +[20] Jingtao Li, Xinyu Wang, Shaoyu Wang, Hengwei Zhao, and Yanfei Zhong. One step detection paradigm for hyperspectral anomaly detection via spectral deviation relationship learning. IEEE Transactions on Geoscience and Remote Sensing, 2024. 1, 7 +[21] Jiaojiao Li, Zhiyuan Zhang, Yuzhe Liu, Rui Song, Yunsong Li, and Qian Du. Swformer: Stochastic windows convolutional transformer for hybrid modality hyperspectral classification. IEEE Transactions on Image Processing, 2024. 2, 3 +[22] Wei Li and Qian Du. Collaborative representation for hyperspectral anomaly detection. IEEE Transactions on geoscience and remote sensing, 53(3):1463-1474, 2014. 7 +[23] Yapeng Li, Yong Luo, Lefei Zhang, Zengmao Wang, and Bo Du. Mambahsi: Spatial-spectral mamba for hyperspectral + +image classification. IEEE Transactions on Geoscience and Remote Sensing, 62:1-16, 2024. 7 +[24] Jie Lian, Lizhi Wang, He Sun, and Hua Huang. Gthad: Gated transformer for hyperspectral anomaly detection. IEEE Transactions on Neural Networks and Learning Systems, 2024. 3 +[25] Nanqing Liu, Xun Xu, Yongyi Su, Haojie Zhang, and Heng-Chao Li. Pointsam: Pointly-supervised segment anything model for remote sensing images. arXiv preprint arXiv:2409.13401, 2024. 1, 2 +[26] Fulin Luo, Tianyuan Zhou, Jiamin Liu, Tan Guo, Xiwen Gong, and Jinchang Ren. Multiscale diff-changed feature fusion network for hyperspectral image change detection. IEEE Transactions on Geoscience and Remote Sensing, 61: 1-13, 2023. 2, 3 +[27] Zhaozhi Luo, Xinyu Wang, Petri Pellikka, Janne Heiskanen, and Yanfei Zhong. Unsupervised adaptation learning for real multiplatform hyperspectral image denoising. IEEE transactions on cybernetics, 2024. 2, 8 +[28] Dimitris Manolakis and Gary Shaw. Detection algorithms for hyperspectral imaging applications. IEEE signal processing magazine, 19(1):29-43, 2002. 7 +[29] Dimitris Manolakis, Eric Truslow, Michael Pieper, Thomas Cooley, and Michael Brueggeman. Detection algorithms in hyperspectral imaging systems: An overview of practical algorithms. IEEE Signal Processing Magazine, 31(1):24-33, 2013. 7 +[30] Farid Melgani and Lorenzo Bruzzone. Classification of hyperspectral remote sensing images with support vector machines. IEEE Transactions on geoscience and remote sensing, 42(8):1778-1790, 2004. 7 +[31] Mohammed Abdulmajeed Moharram and Divya Meena Sundaram. Land use and land cover classification with hyperspectral data: A comprehensive review of methods, challenges and future directions. Neurocomputing, 536:90-113, 2023. 1, 3 +[32] Aleksandar Petrov, Emanuele La Malfa, Philip Torr, and Adel Bibi. Language model tokenizers introduce unfairness between languages. Advances in Neural Information Processing Systems, 36, 2024. 2, 3 +[33] Rami Piiroinen, Fabian Ewald Fassnacht, Janne Heiskanen, Eduardo Maeda, Benjamin Mack, and Petri Pellikka. Invasive tree species detection in the eastern arc mountains biodiversity hotspot using one class classification. Remote sensing of environment, 218:119-131, 2018. 3, 7 +[34] Maciej P Polak and Dane Morgan. Extracting accurate materials data from research papers with conversational language models and prompt engineering. Nature Communications, 15(1):1569, 2024. 3 +[35] Jiahui Qu, Shaoxiong Hou, Wenqian Dong, Yunsong Li, and Weiying Xie. A multilevel encoder-decoder attention network for change detection in hyperspectral images. IEEE Transactions on Geoscience and Remote Sensing, 60:1-13, 2021. 7 +[36] Ying Qu, Wei Wang, Rui Guo, Bulent Ayhan, Chiman Kwan, Steven Vance, and Hairong Qi. Hyperspectral anomaly detection through spectral unmixing and dictionary-based low + +rank decomposition. IEEE Transactions on Geoscience and Remote Sensing, 56(8):4391-4405, 2018. 7 +[37] Behnood Rasti, Alexandre Zouaoui, Julien Mairal, and Jocelyn Chanussot. Image processing and machine learning for hyperspectral unmixing: An overview and the hysupp python package. IEEE Transactions on Geoscience and Remote Sensing, 2024. 2, 8 +[38] C. J. Reed, R. Gupta, S. Li, S. Brockman, C. Funk, B. Clipp, K. Keutzer, S. Candido, M. Uytendaele, and T. Darrell. Scale-mae: A scale-aware masked autoencoder for multiscale geospatial representation learning. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 4065-4076, Los Alamitos, CA, USA, 2023. IEEE Computer Society. 1, 2 +[39] Colorado J Reed, Ritwik Gupta, Shufan Li, Sarah Brockman, Christopher Funk, Brian Clipp, Kurt Keutzer, Salvatore Candido, Matt Uytendaele, and Trevor Darrell. Scale-mae: A scale-aware masked autoencoder for multiscale geospatial representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4088-4099, 2023. 4 +[40] Irving S Reed and Xiaoli Yu. Adaptive multiple-band cfar detection of an optical pattern with unknown spectral distribution. IEEE transactions on acoustics, speech, and signal processing, 38(10):1760-1770, 1990. 7 +[41] Swalpa Kumar Roy, Gopal Krishna, Shiv Ram Dubey, and Bidyut B Chaudhuri. Hybridsn: Exploring 3-d-2-d cnn feature hierarchy for hyperspectral image classification. IEEE Geoscience and Remote Sensing Letters, 17(2):277-281, 2019. 7 +[42] Pranab Sahoo, Ayush Kumar Singh, Sriparna Saha, Vinija Jain, Samrat Mondal, and Aman Chadha. A systematic survey of prompt engineering in large language models: Techniques and applications. arXiv preprint arXiv:2402.07927, 2024.3 +[43] Bernhard Scholkopf, Robert C Williamson, Alex Smola, John Shawe-Taylor, and John Platt. Support vector method for novelty detection. Advances in neural information processing systems, 12, 1999. 7 +[44] Dunbin Shen, Xiaorui Ma, Wenfeng Kong, Jianjun Liu, Jie Wang, and Hongyu Wang. Hyperspectral target detection based on interpretable representation network. IEEE Transactions on Geoscience and Remote Sensing, 2023. 7 +[45] Hongjun Su, Zhaoyue Wu, Huihui Zhang, and Qian Du. Hyperspectral anomaly detection: A survey. IEEE Geoscience and Remote Sensing Magazine, 10(1):64-90, 2021. 2 +[46] Le Sun, Guangrui Zhao, Yuhui Zheng, and Zebin Wu. Spectral-spatial feature tokenization transformer for hyperspectral image classification. IEEE Transactions on Geoscience and Remote Sensing, 60:1-14, 2022. 7 +[47] Siyu Sun, Jun Liu, and Wei Li. Spatial invariant tensor self-representation model for hyperspectral anomaly detection. IEEE Transactions on Cybernetics, 2023. 2 +[48] Xian Sun, Peijin Wang, Wanxuan Lu, Zicong Zhu, Xiaonan Lu, Qibin He, Junxi Li, Xue Rong, Zhujun Yang, Hao Chang, et al. Ringmo: A remote sensing foundation model with masked image modeling. IEEE Transactions on Geoscience and Remote Sensing, 61:1-22, 2022. 4 + +[49] Xian Sun, Peijin Wang, Wanxuan Lu, Zicong Zhu, Xiaonan Lu, Qibin He, Junxi Li, Xue Rong, Zhujun Yang, Hao Chang, Qinglin He, Guang Yang, Ruiping Wang, Jiwen Lu, and Kun Fu. Ringmo: A remote sensing foundation model with masked image modeling. IEEE Transactions on Geoscience and Remote Sensing, 61:1-22, 2023. 1, 2 +[50] Adam Van Etten, Dave Lindenbaum, and Todd M Bacastow. Spacenet: A remote sensing dataset and challenge series. arXiv preprint arXiv:1807.01232, 2018. 2, 3 +[51] Adam Van Etten, Jacob Shermeyer, Daniel Hogan, Nicholas Weir, and Ryan Lewis. Road network and travel time extraction from multiple look angles with spacenet data. In IGARSS 2020-2020 IEEE International Geoscience and Remote Sensing Symposium, pages 3920-3923. IEEE, 2020. 2, 3 +[52] Di Wang, Bo Du, and Liangpei Zhang. Fully contextual network for hyperspectral scene parsing. IEEE Transactions on Geoscience and Remote Sensing, 60:1-16, 2021. 7 +[53] Di Wang, Qiming Zhang, Yufei Xu, Jing Zhang, Bo Du, Dacheng Tao, and Liangpei Zhang. Advancing plain vision transformer toward remote sensing foundation model. IEEE Transactions on Geoscience and Remote Sensing, 61:1-15, 2023. 4 +[54] Di Wang, Meiqi Hu, Yao Jin, Yuchun Miao, Jiaqi Yang, Yichu Xu, Xiaolei Qin, Jiaqi Ma, Lingyu Sun, Chenxing Li, et al. Hypersigma: Hyperspectral intelligence comprehension foundation model. arXiv preprint arXiv:2406.11519, 2024. 2, 3, 4, 8 +[55] Shaoyu Wang, Xinyu Wang, Liangpei Zhang, and Yanfei Zhong. Auto-ad: Autonomous hyperspectral anomaly detection network based on fully convolutional autoencoder. IEEE Transactions on Geoscience and Remote Sensing, 60:1-14, 2021. 7 +[56] Sheng Wang, Kaiyu Guan, Chenhui Zhang, Qu Zhou, Sibo Wang, Xiaocui Wu, Chongya Jiang, Bin Peng, Weiye Mei, Kaiyuan Li, et al. Cross-scale sensing of field-level crop residue cover: Integrating field photos, airborne hyperspectral imaging, and satellite data. Remote Sensing of Environment, 285:113366, 2023. 3 +[57] Yanheng Wang, Danfeng Hong, Jianjun Sha, Lianru Gao, Lian Liu, Yonggang Zhang, and Xianhui Rong. Spectral-spatial-temporal transformers for hyperspectral image change detection. IEEE Transactions on Geoscience and Remote Sensing, 60:1-14, 2022. 7 +[58] Philine Warnke, Kerstin Cuhls, Ulrich Schmoch, Lea Daniel, Liviu Andreescu, Bianca Dragomir, Radu Gheorghiu, Catalina Baboschi, Adrian Curaj, Marjukka Parkkinen, et al. 100 radical innovation breakthroughs for the future. 2019. 1 +[59] Nicholas Weir, David Lindenbaum, Alexei Bastidas, Adam Van Etten, Sean McPherson, Jacob Shermeyer, Varun Kumar, and Hanlin Tang. Spacenet mvoi: A multi-view overhead imagery dataset. In Proceedings of the IEEE/cvf international conference on computer vision, pages 992-1001, 2019. 2, 3 +[60] Walter F Wiggins and Ali S Tejani. On the opportunities and risks of foundation models for natural language processing in + +radiology. Radiology: Artificial Intelligence, 4(4):e220119, 2022.3 +[61] Yinan Wu, Licheng Jiao, Xu Liu, Fang Liu, Shuyuan Yang, and Lingling Li. Domain adaptation-aware transformer for hyperspectral object tracking. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 2, 8 +[62] Zhitong Xiong, Yi Wang, Fahong Zhang, Adam J Stewart, Joëlle Hanna, Damian Borth, Ioannis Papoutsis, Bertrand Le Saux, Gustau Camps-Valls, and Xiao Xiang Zhu. Neural plasticity-inspired foundation model for observing the earth crossing modalities. arXiv e-prints, pages arXiv-2403, 2024. 8 +[63] Yang Xu, Zebin Wu, Jun Li, Antonio Plaza, and Zhihui Wei. Anomaly detection in hyperspectral images based on low-rank and sparse representation. IEEE Transactions on Geoscience and Remote Sensing, 54(4):1990-2000, 2015. 7 +[64] Zongyin Yang, Tom Albrow-Owen, Weiwei Cai, and Tawfique Hasan. Miniaturization of optical spectrometers. Science, 371(6528):eabe0722, 2021. 1 +[65] Alina Zare, Changzhe Jiao, and Taylor Glenn. Discriminative multiple instance hyperspectral target characterization. IEEE transactions on pattern analysis and machine intelligence, 40(10):2342-2354, 2017. 2 +[66] Hengwei Zhao, Xinyu Wang, Jingtao Li, and Yanfei Zhong. Class prior-free positive-unlabeled learning with taylor variational loss for hyperspectral remote sensing imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16827-16836, 2023. 2, 7 +[67] Hengwei Zhao, Yanfei Zhong, Xinyu Wang, and Hong Shu. One-class risk estimation for one-class hyperspectral image classification. IEEE Transactions on Geoscience and Remote Sensing, 2023. 7 +[68] Zhuo Zheng, Yanfei Zhong, Ailong Ma, and Liangpei Zhang. Fpga: Fast patch-free global learning framework for fully end-to-end hyperspectral image classification. IEEE Transactions on Geoscience and Remote Sensing, 58(8):5612-5626, 2020. 7 +[69] Yanfei Zhong, Xin Hu, Chang Luo, Xinyu Wang, Ji Zhao, and Liangpei Zhang. Whu-hi: Uav-borne hyperspectral with high spatial resolution (h2) benchmark datasets and classifier for precise crop identification based on deep convolutional neural network with crf. Remote Sensing of Environment, 250:112012, 2020. 1, 2 \ No newline at end of file diff --git a/2025/HyperFree_ A Channel-adaptive and Tuning-free Foundation Model for Hyperspectral Remote Sensing Imagery/images.zip b/2025/HyperFree_ A Channel-adaptive and Tuning-free Foundation Model for Hyperspectral Remote Sensing Imagery/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..124b0b218e2cfda0d76c5bb951f446e8d9eed370 --- /dev/null +++ b/2025/HyperFree_ A Channel-adaptive and Tuning-free Foundation Model for Hyperspectral Remote Sensing Imagery/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:03d8360087e2f58e3f02f687d8097831beb6990d27095396f4175af4b3dbb161 +size 817967 diff --git a/2025/HyperFree_ A Channel-adaptive and Tuning-free Foundation Model for Hyperspectral Remote Sensing Imagery/layout.json b/2025/HyperFree_ A Channel-adaptive and Tuning-free Foundation Model for Hyperspectral Remote Sensing Imagery/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b8681d405329e0e3ec30a5f1c025f5f4a2744b7e --- /dev/null +++ b/2025/HyperFree_ A Channel-adaptive and Tuning-free Foundation Model for Hyperspectral Remote Sensing Imagery/layout.json @@ -0,0 +1,8559 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 84, + 102, + 526, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 102, + 526, + 140 + ], + "spans": [ + { + "bbox": [ + 84, + 102, + 526, + 140 + ], + "type": "text", + "content": "HyperFree: A Channel-adaptive and Tuning-free Foundation Model for Hyperspectral Remote Sensing Imagery" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 96, + 160, + 515, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 160, + 515, + 204 + ], + "spans": [ + { + "bbox": [ + 96, + 160, + 515, + 204 + ], + "type": "text", + "content": "Jingtao Li\\*, Yingyi Liu\\*, Xinyu Wang†, Yunning Peng†, Chen Sun†, Shaoyu Wang², Zhendong Sun†, Tian Ke†, Xiao Jiang†, Tangwei Lu†, Anran Zhao†, Yanfei Zhong†" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 198, + 207, + 410, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 207, + 410, + 217 + ], + "spans": [ + { + "bbox": [ + 198, + 207, + 410, + 217 + ], + "type": "text", + "content": "{Jintaoli, liuyingyi, wangxinyu, pengyunning, sunchen," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 146, + 220, + 463, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 220, + 463, + 231 + ], + "spans": [ + { + "bbox": [ + 146, + 220, + 463, + 231 + ], + "type": "text", + "content": "sunzhendong, ketian, jiangxiao, lutangwei, anranzha0415, zhongyanfei}@whu.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 266, + 236, + 341, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 236, + 341, + 245 + ], + "spans": [ + { + "bbox": [ + 266, + 236, + 341, + 245 + ], + "type": "text", + "content": "wsy1995@snu.ac.kr" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "spans": [ + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 297, + 296, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 297, + 296, + 550 + ], + "spans": [ + { + "bbox": [ + 54, + 297, + 296, + 550 + ], + "type": "text", + "content": "Advanced interpretation of hyperspectral remote sensing images benefits many precise Earth observation tasks. Recently, visual foundation models have promoted the remote sensing interpretation but concentrating on RGB and multispectral images. Due to the varied hyperspectral channels, existing foundation models would face image-by-image tuning situation, imposing great pressure on hardware and time resources. In this paper, we propose a tuning-free hyperspectral foundation model called HyperFree, by adapting the existing visual prompt engineering. To process varied channel numbers, we design a learned weight dictionary covering full-spectrum from " + }, + { + "bbox": [ + 54, + 297, + 296, + 550 + ], + "type": "inline_equation", + "content": "0.4 \\sim 2.5\\mu m" + }, + { + "bbox": [ + 54, + 297, + 296, + 550 + ], + "type": "text", + "content": ", supporting to build the embedding layer dynamically. To make the prompt design more tractable, HyperFree can generate multiple semantic-aware masks for one prompt by treating feature distance as semantic-similarity. After pre-training HyperFree on constructed large-scale high-resolution hyperspectral images, HyperFree (1 prompt) has shown comparable results with specialized models (5 shots) on 5 tasks and 11 datasets. Code and dataset are accessible at https://rsidea.whu.edu.cn/hyperfree.htm." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 571, + 135, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 571, + 135, + 583 + ], + "spans": [ + { + "bbox": [ + 55, + 571, + 135, + 583 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 590, + 296, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 296, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 296, + 687 + ], + "type": "text", + "content": "Hyperspectral imaging has been recognized by the European Union as one of the Top 100 Disruptive Technologies [58], due to its ability to provide high-resolution spectral data for precise Earth observation across diverse applications [64], including environmental monitoring [31], agriculture [69], and defense [20]. Unlike traditional RGB or multispectral imagery, Hyperspectral imagery (HSI) offers several unique features: high spectral resolution (typically" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 320, + 273, + 548, + 334 + ], + "blocks": [ + { + "bbox": [ + 320, + 273, + 548, + 334 + ], + "lines": [ + { + "bbox": [ + 320, + 273, + 548, + 334 + ], + "spans": [ + { + "bbox": [ + 320, + 273, + 548, + 334 + ], + "type": "image", + "image_path": "67039194eb890a9a9984274c3a3a021c040dff12b5c1be6bdd1b1d11382ef41d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 359, + 335, + 528, + 344 + ], + "lines": [ + { + "bbox": [ + 359, + 335, + 528, + 344 + ], + "spans": [ + { + "bbox": [ + 359, + 335, + 528, + 344 + ], + "type": "text", + "content": "(a) Mainstream Pre-training and Tuning Paradigm of RSFM" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 320, + 353, + 548, + 407 + ], + "blocks": [ + { + "bbox": [ + 320, + 353, + 548, + 407 + ], + "lines": [ + { + "bbox": [ + 320, + 353, + 548, + 407 + ], + "spans": [ + { + "bbox": [ + 320, + 353, + 548, + 407 + ], + "type": "image", + "image_path": "b431c2ad7d207f8a9bdd84e93394e2b6a672f38cb29bc8e63c08bd8841c309fa.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 387, + 411, + 501, + 419 + ], + "lines": [ + { + "bbox": [ + 387, + 411, + 501, + 419 + ], + "spans": [ + { + "bbox": [ + 387, + 411, + 501, + 419 + ], + "type": "text", + "content": "(b) Prompt Engineering of SAM-variants" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 319, + 427, + 548, + 480 + ], + "blocks": [ + { + "bbox": [ + 319, + 427, + 548, + 480 + ], + "lines": [ + { + "bbox": [ + 319, + 427, + 548, + 480 + ], + "spans": [ + { + "bbox": [ + 319, + 427, + 548, + 480 + ], + "type": "image", + "image_path": "5629edb7992d3e8971b9a8db18f3f469069c80024c77191cdfecf3fb7355b68b.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 390, + 487, + 495, + 496 + ], + "lines": [ + { + "bbox": [ + 390, + 487, + 495, + 496 + ], + "spans": [ + { + "bbox": [ + 390, + 487, + 495, + 496 + ], + "type": "text", + "content": "(c) Prompt Engineering of HyperFree" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 506, + 555, + 595 + ], + "lines": [ + { + "bbox": [ + 313, + 506, + 555, + 595 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 555, + 595 + ], + "type": "text", + "content": "Figure 1. Comparison of existing RSFMs with the proposed HyperFree model. (a) The pre-training and tuning paradigm is the mainstream approach, where fine-tuning is essential for handling unseen HSIs. (b) The prompt engineering paradigm reduces the tuning burden but struggles to handle HSIs with hundreds of spectral channels. (c) HyperFree can directly process unseen HSIs with varying channel counts without fine-tuning, and it can identify multiple masks of the same class using only a single prompt." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 601, + 555, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 601, + 555, + 650 + ], + "spans": [ + { + "bbox": [ + 313, + 601, + 555, + 650 + ], + "type": "inline_equation", + "content": "10\\mathrm{nm}" + }, + { + "bbox": [ + 313, + 601, + 555, + 650 + ], + "type": "text", + "content": "), an extensive wavelength range (400–2500 nm), and dozens to hundreds of spectral channels [64], which enable the detailed identification of diverse targets, while also increasing the complexity of data processing and analysis." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 650, + 556, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 650, + 556, + 711 + ], + "spans": [ + { + "bbox": [ + 313, + 650, + 556, + 711 + ], + "type": "text", + "content": "Recently, remote sensing foundation models (RSFMs) have demonstrated advanced interpretive capabilities, primarily focusing on multispectral and RGB images [4, 7, 13, 25, 38, 49], while also opening new possibilities for analyzing hyperspectral images. RSFMs follow two main" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 67, + 703, + 148, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 148, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 148, + 712 + ], + "type": "text", + "content": "† Corresponding authors." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23048" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "content": "paradigms. The predominant paradigm is pre-training and tuning (P-T), illustrated in Figure 1(a), which begins with training a large-scale backbone model in a self-supervised manner, such as masked autoencoders (MAE) [12], followed by fine-tuning with task-specific heads for various downstream applications [7, 13, 38, 49]. In the P-T paradigm, both the pre-training and the fine-tuning stages can be time-consuming. Second paradigm is prompt engineering (P-E), as illustrated in Figure 1(b), which focuses on designing task-specific prompts for a frozen Segment Anything Model (SAM) [17] to directly perform downstream tasks without additional model tuning [4, 25], making it more efficient and user-friendly in practical." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 228, + 295, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 295, + 502 + ], + "type": "text", + "content": "The P-E paradigm is particularly well-suited for processing HSI, as it can be directly applied to diverse, unseen scenes without requiring model fine-tuning. However, it faces two significant challenges when applied to HSI. (i) The first challenge is the variation in spectral parameters across different hyperspectral sensors, including differences in spectral resolution, wavelength range, and the number of spectral channels [10]. For a P-T-based model like HyperSigma [54], although it is pre-trained using HSIs with fixed channels, the challenge can be addressed during its tuning stage by rebuilding the channel embedding layer to align the dimensions with the new data. But for a P-E-based model, the challenge of channel variation requires more careful consideration. (ii) The second challenge is to make prompt input tractable, ensuring a P-E based model can directly process unseen HSIs with least prior knowledge. Specifically, most existing P-E-based RSFMs rely on a frozen SAM [17], which segments remote sensing objects based on specialized models to generate prompts [4, 25]. As shown in Figure 1(b), each prompt typically generates a single mask and all the prompts for target objects are needed by existing P-E-based RSFMs, which largely reduced the model feasibility." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 503, + 295, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 503, + 295, + 706 + ], + "spans": [ + { + "bbox": [ + 55, + 503, + 295, + 706 + ], + "type": "text", + "content": "To address the above challenges, as shown in Figure 1 (c), we propose a channel-adaptive, tuning-free hyperspectral foundation model, called HyperFree, which can process any unseen HSIs directly in a P-E manner. Specifically, (i) Motivated by the word dictionary in natural language processing, which stores vector embeddings as values corresponding to words and phrases of varying lengths as keys [32], HyperFree introduces a learnable wavelength weight dictionary in the embedding layer, spanning the 400 to 2500 nm spectral range in " + }, + { + "bbox": [ + 55, + 503, + 295, + 706 + ], + "type": "inline_equation", + "content": "10\\mathrm{nm}" + }, + { + "bbox": [ + 55, + 503, + 295, + 706 + ], + "type": "text", + "content": " intervals. This allows for the dynamic encoding of unseen HSIs with varying spectral parameters into visual tokens, all with consistent dimensionality. (ii) To make the prompt of HyperFree more tractable, HyperFree is designed to generate multiple semantic-aware masks for one prompt as in Figure 1 (c), where HyperFree maps each prompt and single mask from image space into feature space and treats the feature distance as semantic-" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": "similarity. Prompt, mask and feature are interacted adaptively for different downstream task in tuning-free manner with only one prompt or zero shot." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 108, + 553, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 108, + 553, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 108, + 553, + 323 + ], + "type": "text", + "content": "Due to the high acquisition cost and label difficulty, current hyperspectral datasets only have a few images [45, 69], far from the requirements to train the promptable HyperFree. To conquer this problem, we developed the Hyper-Seg data engine, in addition to utilizing well-known multispectral datasets (e.g., fMoW [5] and SpaceNet [50, 51, 59]). This engine automatically generates a large-scale annotated HSI dataset, consisting of nearly 50,000 pairs of HSIs and their corresponding segmentation masks. The original data are sourced from AVIRIS airborne hyperspectral sensors, with a spatial size of " + }, + { + "bbox": [ + 313, + 108, + 553, + 323 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 313, + 108, + 553, + 323 + ], + "type": "text", + "content": " pixels and 224 spectral channels spanning from 400 to " + }, + { + "bbox": [ + 313, + 108, + 553, + 323 + ], + "type": "inline_equation", + "content": "2500\\mathrm{nm}" + }, + { + "bbox": [ + 313, + 108, + 553, + 323 + ], + "type": "text", + "content": " at a " + }, + { + "bbox": [ + 313, + 108, + 553, + 323 + ], + "type": "inline_equation", + "content": "10\\mathrm{nm}" + }, + { + "bbox": [ + 313, + 108, + 553, + 323 + ], + "type": "text", + "content": " spectral resolution. Unlike other RSFMs trained on unannotated satellite hyperspectral imagery with a " + }, + { + "bbox": [ + 313, + 108, + 553, + 323 + ], + "type": "inline_equation", + "content": "30\\mathrm{m}" + }, + { + "bbox": [ + 313, + 108, + 553, + 323 + ], + "type": "text", + "content": " spatial resolution [54], the Hyper-Seg data engine features higher spatial resolution (0.6 to " + }, + { + "bbox": [ + 313, + 108, + 553, + 323 + ], + "type": "inline_equation", + "content": "5.0\\mathrm{m}" + }, + { + "bbox": [ + 313, + 108, + 553, + 323 + ], + "type": "text", + "content": "), along with annotations of 15.44 million masks. It is currently the largest known hyperspectral dataset with high spatial resolution." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 323, + 553, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 553, + 478 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 553, + 478 + ], + "type": "text", + "content": "To demonstrate the model capabilities, we evaluated HyperFree directly in a P-E manner on 11 datasets across 5 segmentation-related tasks, including multi-class classification [21], one-class classification [66], target detection [65], anomaly detection [47], and change detection [26]. Additionally, we verified HyperFree in a P-T manner on 14 datasets spanning 8 tasks, with additional tasks including hyperspectral denoising [27], hyperspectral unmixing [37], and hyperspectral object tracking [61]. The experimental results show that HyperFree achieves comparable accuracy to state-of-the-art models in a tuning-free manner (using just one prompt), and outperforms most models after tuning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 326, + 479, + 493, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 479, + 493, + 490 + ], + "spans": [ + { + "bbox": [ + 326, + 479, + 493, + 490 + ], + "type": "text", + "content": "Our contributions can be summarized as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 491, + 553, + 658 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 314, + 491, + 553, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 491, + 553, + 526 + ], + "spans": [ + { + "bbox": [ + 314, + 491, + 553, + 526 + ], + "type": "text", + "content": "- We propose the first tuning-free hyperspectral foundation model, which can process any hyperspectral image in different tasks with promptable or zero-shot manner." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 527, + 553, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 527, + 553, + 562 + ], + "spans": [ + { + "bbox": [ + 314, + 527, + 553, + 562 + ], + "type": "text", + "content": "- A weight dictionary that spans the full spectrum, enabling dynamic generation of the embedding layer dynamically according to input wavelengths." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 563, + 553, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 563, + 553, + 609 + ], + "spans": [ + { + "bbox": [ + 314, + 563, + 553, + 609 + ], + "type": "text", + "content": "- We propose to map both prompts and masks into feature space to identify multiple semantic-aware masks for one prompt, where different interaction workflows are designed for each downstream tasks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 610, + 553, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 610, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 314, + 610, + 553, + 658 + ], + "type": "text", + "content": "- We built the Hyper-Seg data engine to train the HyperFree model and tested it on 11 datasets from 5 tasks in tuning-free manner, 14 datasets from 8 tasks in tuning manner as an extensive experiment." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 669, + 400, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 669, + 400, + 681 + ], + "spans": [ + { + "bbox": [ + 314, + 669, + 400, + 681 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": "Hyperspectral image processing. There are diverse tasks to process such precise observation with different targets" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23049" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 202 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 202 + ], + "type": "text", + "content": "[31]. (a) Classification: Assigning a specific class to each pixel [21]. (b) One class classification: Discriminating the pixels of target class with labeled binary map [33]. (c) Target detection. Discriminating the pixels of target class with labeled spectrum. (d) Anomaly detection. Discriminating the pixels spectrally deviating from the background [24]. (e) Change detection: Discriminating the changing pixels between two time-steps [26]. Since the output of the above tasks is all pixel-level segmentation maps, it is feasible to unify them using the proposed HyperFree model in tuning-free manner." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 216, + 295, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 216, + 295, + 455 + ], + "spans": [ + { + "bbox": [ + 56, + 216, + 295, + 455 + ], + "type": "text", + "content": "Foundation models. Powered by the transformer architectures, foundation models are characterized by large-scale property in model size and pre-training data, and the transferring ability for downstream tasks [60]. In remote sensing community, some foundation models for multi-spectral images have already been proposed such as SatMAE [7], SepctralEarth [1] and SpectralGPT [13], which trains large ViT-like models [9] with the MAE strategy [12] and massive open-source multi-spectral images (13 channels). In contrast, hyperspectral images have a high acquiring cost [56] and make the large-scale pre-training stage more difficult. A recent study broke the situation by building a largescale hyperspectral dataset HyperGlobal-450K and trained a foundation model called HyperSigma with MAE strategy as well [54]. Our work differs from HyperSigma in two aspects: (i) HyperSigma accepts fixed number of image channels, while HyperFree can deal with any number of channels directly. (ii) HyperSigma needs to be fine-tuned for each downstream task. Differently, HyperFree has ability to process different tasks in fine-tuning free manner." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 456, + 295, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 295, + 610 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 295, + 610 + ], + "type": "text", + "content": "Prompt engineering (P-E). Different from the pretraining and then fine-tuning paradigm, P-E can utilize pretrained foundation models to complete downstream tasks directly with prompts [34]. For pre-trained large language models, prompts are mostly hand-crafted text to guide the model generate response for specific task [34, 42]. In computer vision community, SAM follows the same spirit and designs diverse prompts (e.g., point and box) to achieve few-shot transferring performance [17]. Our work extends the visual prompt engineering in hyperspectral data with varying image channels and supports five tasks in tuning-free manner with designed prompt-mask-feature interaction." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 621, + 195, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 621, + 195, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 621, + 195, + 635 + ], + "type": "text", + "content": "3. Hyper-Seg Data Engine" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 642, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 294, + 713 + ], + "type": "text", + "content": "Hyperspectral image processing always refers to various dense tasks and a large-scale segmentation dataset is needed to apply the P-E. However, most hyperspectral segmentation datasets only have a few images due to the high acquisition and label cost, which is quite different from the million-scale dataset with natural images such as SA-1B" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 310 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 310 + ], + "type": "text", + "content": "[17]. To tackle this, we built a data engine called HyperSeg to generate segmented masks automatically for spectral images and expand the data scale. Hyper-Seg engine firstly separates the spectral image into several groups with three channels according to the prior of key channels position. Each group is then segmented by the SAM-H model [17] and the results are combined by NMS operation to eliminate redundancy and output the final segmentation map. Figure 2 shows the overall workflow, where the nine key channels are selected according to the expert knowledge as in Supp.1.1. Considering the high spatial resolution property of airborne platform, we collected 41946 hyperspectral images from AVIRIS airborne sensor and applied the Hyper-Seg Engine. Besides, we also processed two existing multispectral datasets (fMoW [5] and SpaceNet [50, 51, 59]) to increase the scale further. Compared to the traditional hyperspectral annotation methods which refer to high-resolution RGB imagery, our data engine can utilize more spectral information and obtain a high-precision segmentation mask without labor cost." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 312, + 553, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 553, + 491 + ], + "type": "text", + "content": "Table 1 shows the comparable statistical information of training datasets between existing RSFMs and proposed HyperFree. Most models only concentrate on the RGB and multispectral images with channel number 3 13. Although the recent model HyperSigma [54] has collected a large-scale hyperspectral images, the spatial resolution (30m/pixel) is largely lower than us (0.6 5m/pixel), and the dataset scale (considering image size and number together) is relatively small. With the designed Hyper-Seg data engine, constructed dataset is the only one with segmentation masks supporting the promptable training. More detailed statistical information of our generated dataset is given in Supp.1.2. The final generated dataset has nearly 150k images and 15.44million masks. We name both the data engine and generated dataset as the Hyper-Seg for simplicity." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 501, + 486, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 501, + 486, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 501, + 486, + 514 + ], + "type": "text", + "content": "4. HyperFree Foundation Model" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 521, + 553, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 553, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 553, + 641 + ], + "type": "text", + "content": "HyperFree has two customized components upon the metaarchitecture of SAM including channel-adaptive embedding and prompt-mask-feature (PMF) interaction as in Figure 3. Channel-adaptive embedding constructs the embedding layer dynamically according to varying input wavelength, and outputs tokens in same representation space. After the promptable training stage, PMF interaction discriminates the semantic-aware masks for downstream tasks by mapping both prompt and mask into feature space and treating feature distance as semantic similarity." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 648, + 477, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 648, + 477, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 648, + 477, + 661 + ], + "type": "text", + "content": "4.1. Channel-adaptive Embedding" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "type": "text", + "content": "To make HyperFree process any hyperspectral image, channel-adaptive embedding layer is firstly designed to map input image with varying channels into tokens with fixed length, inspired by the tokenizer in NLP community [32]." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23050" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 130, + 73, + 479, + 228 + ], + "blocks": [ + { + "bbox": [ + 130, + 73, + 479, + 228 + ], + "lines": [ + { + "bbox": [ + 130, + 73, + 479, + 228 + ], + "spans": [ + { + "bbox": [ + 130, + 73, + 479, + 228 + ], + "type": "image", + "image_path": "428559adf50a8595acfdcb565789207c51467f6398d40bb9b5b971b5ed5f292f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 239, + 555, + 264 + ], + "lines": [ + { + "bbox": [ + 54, + 239, + 555, + 264 + ], + "spans": [ + { + "bbox": [ + 54, + 239, + 555, + 264 + ], + "type": "text", + "content": "Figure 2. Hyper-Seg Data Engine automatically generates a large-scale dataset comprising nearly 41,900 pairs of airborne HSIs and their corresponding segmentation masks, enabling HyperFree to effectively learn discriminative spectral features." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 304, + 549, + 430 + ], + "blocks": [ + { + "bbox": [ + 54, + 274, + 555, + 298 + ], + "lines": [ + { + "bbox": [ + 54, + 274, + 555, + 298 + ], + "spans": [ + { + "bbox": [ + 54, + 274, + 555, + 298 + ], + "type": "text", + "content": "Table 1. Comparison of training data used in current RSFMs. The primary features of the proposed HyperFree model include prompt engineering and the use of high-resolution annotated hyperspectral (HRS) imagery." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 304, + 549, + 430 + ], + "lines": [ + { + "bbox": [ + 56, + 304, + 549, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 304, + 549, + 430 + ], + "type": "table", + "html": "
DatasetFoundation ModelSpatial Resolution (m)Spectral Resolution (nm)Channel NumberImage SizeImage NumberMain ModalityTraining Paradigm
RGBRVSA [53]0.5 ~ 153\\3110 ~ 316721.00 × 10^6H/M/LSR-RGBP-T
RingMo [48]0.1 ~ 30\\34482.10 × 10^6H/MSR-RGBP-T
ScaleMAE [39]0.31 ~ 3.7\\3\\3.64 × 10^6HSR-RGBP-T
MultispectralSatMAE [6]1015 ~ 18013≈ 45 × 607.13 × 10^5MSR-MP-T
SkySense [11]10/2015 ~ 18010\\\\MSR-MP-T
SpectralGPT [14]10/20/6015 ~ 1801245 ~ 1201.47 × 10^6MSR-MP-T
HyperspectralHyperSigma [54]305/6.5/10150/175644.47 × 10^5LSR-HP-T
HyperFree0.6 ~ 5.0102245124.19 × 10^4HSR-HP-E
", + "image_path": "07e945a48bc66c09e257ea552910b4191cb89b59c81457f5fbe71039e59c71a5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 56, + 431, + 241, + 441 + ], + "lines": [ + { + "bbox": [ + 56, + 431, + 241, + 441 + ], + "spans": [ + { + "bbox": [ + 56, + 431, + 241, + 441 + ], + "type": "text", + "content": "* P-T: Pretraining and Tuning; P-E: Prompt Engineering." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 56, + 441, + 373, + 450 + ], + "lines": [ + { + "bbox": [ + 56, + 441, + 373, + 450 + ], + "spans": [ + { + "bbox": [ + 56, + 441, + 373, + 450 + ], + "type": "text", + "content": "* HSR: High Spatial Resolution; MSR: Medium Spatial Resolution; LSR: Low Spatial Resolution." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 56, + 450, + 341, + 460 + ], + "lines": [ + { + "bbox": [ + 56, + 450, + 341, + 460 + ], + "spans": [ + { + "bbox": [ + 56, + 450, + 341, + 460 + ], + "type": "text", + "content": "* RGB: Red-Green-Blue Imagery; M: Multispectral Imagery; H: Hyperspectral Imagery" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 480, + 295, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 480, + 295, + 588 + ], + "spans": [ + { + "bbox": [ + 54, + 480, + 295, + 588 + ], + "type": "text", + "content": "Tokenizers rely on a pre-defined dictionary to break down a string of text into numerical tokens, where the dictionary has stored the numerical vectors of varied-length words or phrases. Vision transformers always use a convolutional layer to generate patch tokens and what if we construct a layer weight dictionary treating each wavelength as a word? Given the input meta data of wavelengths, we can just loop up the corresponding weights and construct the layer dynamically without re-training it." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 588, + 295, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 588, + 295, + 708 + ], + "spans": [ + { + "bbox": [ + 54, + 588, + 295, + 708 + ], + "type": "text", + "content": "The weight dictionary is built depending on the wavelengths to be processed and the expected token dimension. Since hyperspectral imaging mostly captures the spectral signals from " + }, + { + "bbox": [ + 54, + 588, + 295, + 708 + ], + "type": "inline_equation", + "content": "400\\mathrm{nm}" + }, + { + "bbox": [ + 54, + 588, + 295, + 708 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 54, + 588, + 295, + 708 + ], + "type": "inline_equation", + "content": "2500\\mathrm{nm}" + }, + { + "bbox": [ + 54, + 588, + 295, + 708 + ], + "type": "text", + "content": " and the " + }, + { + "bbox": [ + 54, + 588, + 295, + 708 + ], + "type": "inline_equation", + "content": "10\\mathrm{nm}" + }, + { + "bbox": [ + 54, + 588, + 295, + 708 + ], + "type": "text", + "content": " is of high spectral resolution currently, 221 index channels are designed as the dictionary keys. Denote the output token as a vector of " + }, + { + "bbox": [ + 54, + 588, + 295, + 708 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^j" + }, + { + "bbox": [ + 54, + 588, + 295, + 708 + ], + "type": "text", + "content": " corresponding to the patch size " + }, + { + "bbox": [ + 54, + 588, + 295, + 708 + ], + "type": "inline_equation", + "content": "p\\times p" + }, + { + "bbox": [ + 54, + 588, + 295, + 708 + ], + "type": "text", + "content": ", each index channel stores a weight matrix of size " + }, + { + "bbox": [ + 54, + 588, + 295, + 708 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{p\\times p\\times j}" + }, + { + "bbox": [ + 54, + 588, + 295, + 708 + ], + "type": "text", + "content": ", and the built dictionary " + }, + { + "bbox": [ + 54, + 588, + 295, + 708 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 54, + 588, + 295, + 708 + ], + "type": "text", + "content": " has a size of " + }, + { + "bbox": [ + 54, + 588, + 295, + 708 + ], + "type": "inline_equation", + "content": "221\\times p\\times p\\times j" + }, + { + "bbox": [ + 54, + 588, + 295, + 708 + ], + "type": "text", + "content": ". Given the input hyperspectral image " + }, + { + "bbox": [ + 54, + 588, + 295, + 708 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{h\\times w\\times n}" + }, + { + "bbox": [ + 54, + 588, + 295, + 708 + ], + "type": "text", + "content": " with wavelengths " + }, + { + "bbox": [ + 54, + 588, + 295, + 708 + ], + "type": "inline_equation", + "content": "\\mathbf{b} =" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 480, + 555, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 555, + 554 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 555, + 554 + ], + "type": "inline_equation", + "content": "[b_1, b_2, \\ldots, b_n]" + }, + { + "bbox": [ + 313, + 480, + 555, + 554 + ], + "type": "text", + "content": ", the corresponding kernel weight can be dynamically generated by combining the searched weight " + }, + { + "bbox": [ + 313, + 480, + 555, + 554 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 313, + 480, + 555, + 554 + ], + "type": "text", + "content": " with size " + }, + { + "bbox": [ + 313, + 480, + 555, + 554 + ], + "type": "inline_equation", + "content": "n \\times p \\times p \\times j" + }, + { + "bbox": [ + 313, + 480, + 555, + 554 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 480, + 555, + 554 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 480, + 555, + 554 + ], + "type": "text", + "content": " is flexible for each image. For conciseness, we use function " + }, + { + "bbox": [ + 313, + 480, + 555, + 554 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 313, + 480, + 555, + 554 + ], + "type": "text", + "content": " to represent the mapping process from " + }, + { + "bbox": [ + 313, + 480, + 555, + 554 + ], + "type": "inline_equation", + "content": "\\mathbf{b}" + }, + { + "bbox": [ + 313, + 480, + 555, + 554 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 480, + 555, + 554 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 313, + 480, + 555, + 554 + ], + "type": "text", + "content": " as kernel weight, and function " + }, + { + "bbox": [ + 313, + 480, + 555, + 554 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 480, + 555, + 554 + ], + "type": "text", + "content": " to represent the token generation step as in Equation (1)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 327, + 572, + 555, + 587 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 572, + 555, + 587 + ], + "spans": [ + { + "bbox": [ + 327, + 572, + 555, + 587 + ], + "type": "interline_equation", + "content": "f (\\mathbf {X} \\mid \\mathbf {b}, \\beta) = \\operatorname {C o n v} _ {\\mathbf {W}} (\\mathbf {X}) \\quad \\text {w h e r e} \\quad \\mathbf {W} = g (\\mathbf {b}, \\beta) \\tag {1}", + "image_path": "567c5f1051be72e04186c20e1c109c5b00c03ed110bf9dfcefc6b19c9389108d.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 594, + 556, + 715 + ], + "type": "text", + "content": "In the past few decades, many satellites have been launched equipped with spectral sensors, where the key channels are selected by professional teams and practical experience. To utilize these successful priors, we further extract visual tokens in two parallel branches with dynamically constructed layers: one to process key channels and the other to process the split cubes, as shown in Figure 3. The input image " + }, + { + "bbox": [ + 312, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 312, + 594, + 556, + 715 + ], + "type": "text", + "content": " is first divided into key channels " + }, + { + "bbox": [ + 312, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\mathbf{k}}" + }, + { + "bbox": [ + 312, + 594, + 556, + 715 + ], + "type": "text", + "content": " and the cubes " + }, + { + "bbox": [ + 312, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\mathbf{c}}" + }, + { + "bbox": [ + 312, + 594, + 556, + 715 + ], + "type": "text", + "content": " between them, where the key wavelengths are selected by referring to famous spectral sensors (detailed" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23051" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 78, + 70, + 312, + 285 + ], + "blocks": [ + { + "bbox": [ + 78, + 70, + 312, + 285 + ], + "lines": [ + { + "bbox": [ + 78, + 70, + 312, + 285 + ], + "spans": [ + { + "bbox": [ + 78, + 70, + 312, + 285 + ], + "type": "image", + "image_path": "479ad1de12bb655da0da7ead4673adc2cf106098b2345ef3da7ea96e200b9d60.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 287, + 259, + 296 + ], + "lines": [ + { + "bbox": [ + 130, + 287, + 259, + 296 + ], + "spans": [ + { + "bbox": [ + 130, + 287, + 259, + 296 + ], + "type": "text", + "content": "(a) Full spectrum Promptable Training" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 319, + 69, + 532, + 284 + ], + "blocks": [ + { + "bbox": [ + 319, + 69, + 532, + 284 + ], + "lines": [ + { + "bbox": [ + 319, + 69, + 532, + 284 + ], + "spans": [ + { + "bbox": [ + 319, + 69, + 532, + 284 + ], + "type": "image", + "image_path": "4fdc8d8960c1438dd44057abbb04840e2830b01dbba8002eec3d5c77c2e1cc82.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 351, + 286, + 504, + 297 + ], + "lines": [ + { + "bbox": [ + 351, + 286, + 504, + 297 + ], + "spans": [ + { + "bbox": [ + 351, + 286, + 504, + 297 + ], + "type": "text", + "content": "(b) Prompt-mask-feature Interactive Inferring" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 304, + 555, + 372 + ], + "lines": [ + { + "bbox": [ + 54, + 304, + 555, + 372 + ], + "spans": [ + { + "bbox": [ + 54, + 304, + 555, + 372 + ], + "type": "text", + "content": "Figure 3. (a) Full spectrum Promptable Training. Based on the meta-architecture of promptable segmentation, channel-adaptive embedding layer is developed to process any-channel input, which relies on a learnable weight dictionary and generates the embedding layer dynamically according to the input wavelengths. With the constructed Hyper-Seg dataset, HyperFree is trained with randomly selected channels from full spectrum and point prompts in each iteration. (b) Prompt-mask-feature Interactive Inferring. Keep the trained HyperFree frozen, different PMF interaction workflows are designed to complete the downstream tasks in promptable or zero-shot manner, where we use feature distance to measure the semantic consistency. An interaction instantiation for classification task is given above as an example." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 387, + 297, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 387, + 297, + 436 + ], + "spans": [ + { + "bbox": [ + 55, + 387, + 297, + 436 + ], + "type": "text", + "content": "in Supp.2). We set two different dictionaries " + }, + { + "bbox": [ + 55, + 387, + 297, + 436 + ], + "type": "inline_equation", + "content": "\\beta_{\\mathbf{k}}" + }, + { + "bbox": [ + 55, + 387, + 297, + 436 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 387, + 297, + 436 + ], + "type": "inline_equation", + "content": "\\beta_{\\mathbf{c}}" + }, + { + "bbox": [ + 55, + 387, + 297, + 436 + ], + "type": "text", + "content": " to process " + }, + { + "bbox": [ + 55, + 387, + 297, + 436 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\mathbf{k}}" + }, + { + "bbox": [ + 55, + 387, + 297, + 436 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 387, + 297, + 436 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\mathbf{c}}" + }, + { + "bbox": [ + 55, + 387, + 297, + 436 + ], + "type": "text", + "content": " respectively, and the final output tokens " + }, + { + "bbox": [ + 55, + 387, + 297, + 436 + ], + "type": "inline_equation", + "content": "\\mathbf{T} \\in \\mathbb{R}^{(h/p) \\times (w/p) \\times j}" + }, + { + "bbox": [ + 55, + 387, + 297, + 436 + ], + "type": "text", + "content": " are computed as the element-wise sum, as shown in Equation (2)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 100, + 449, + 295, + 462 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 449, + 295, + 462 + ], + "spans": [ + { + "bbox": [ + 100, + 449, + 295, + 462 + ], + "type": "interline_equation", + "content": "\\mathbf {T} = f \\left(\\mathbf {X} _ {\\mathbf {k}} \\mid \\mathbf {b} _ {\\mathbf {k}}, \\beta_ {\\mathbf {k}}\\right) + f \\left(\\mathbf {X} _ {\\mathbf {c}} \\mid \\mathbf {b} _ {\\mathbf {c}}, \\beta_ {\\mathbf {c}}\\right) \\tag {2}", + "image_path": "e55941939cea0f5fe6a12223f79ce24b5c30e30955e696d8bd9916596ac20ffc.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 470, + 247, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 470, + 247, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 470, + 247, + 483 + ], + "type": "text", + "content": "4.2. Full-spectrum Promptable Training" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 488, + 296, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 488, + 296, + 608 + ], + "spans": [ + { + "bbox": [ + 55, + 488, + 296, + 608 + ], + "type": "text", + "content": "The prompt engineering paradigm is adopted to complete downstream tasks directly with designed prompts. Hyper-Free is thus trained in a promptable manner on constructed large-scale Hyper-Seg dataset, where the prompts are randomly chosen at each iteration and the model is expected to segment correctly the corresponding masks. Since hyperspectral image processing tasks are almost conducted at the pixel level, we choose point prompts only at the training stage. Focal loss " + }, + { + "bbox": [ + 55, + 488, + 296, + 608 + ], + "type": "inline_equation", + "content": "L_{f}" + }, + { + "bbox": [ + 55, + 488, + 296, + 608 + ], + "type": "text", + "content": " and Dice loss " + }, + { + "bbox": [ + 55, + 488, + 296, + 608 + ], + "type": "inline_equation", + "content": "L_{d}" + }, + { + "bbox": [ + 55, + 488, + 296, + 608 + ], + "type": "text", + "content": " are weighted as [17] to supervise the training as Equation (3)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 622, + 295, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 622, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 141, + 622, + 295, + 635 + ], + "type": "interline_equation", + "content": "L = 2 0 L _ {f} + L _ {d} \\tag {3}", + "image_path": "ff6b27ac6e0a1a95f7eb36b63867354f1568ecfd60e9fa659869ab8922bdee11.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "content": "To train the channel-adaptive embedding layer effectively, we select the subset of channels randomly from the full-spectrum range and process it with corresponding weights only in " + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\beta_{\\mathbf{k}}" + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\beta_{\\mathrm{c}}" + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "content": ", which helps the learned dictionary being wavelength-aware and encode the varied channel composition into a unified token space." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 387, + 550, + 400 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 387, + 550, + 400 + ], + "spans": [ + { + "bbox": [ + 313, + 387, + 550, + 400 + ], + "type": "text", + "content": "4.3. Prompt-mask-feature (PMF) Interactive Inferring" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 426, + 555, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 570 + ], + "type": "text", + "content": "After training process, the model can process unseen hyperspectral image with varied channels and segment the corresponding mask given any point prompt. However, many downstream tasks need to segment out all the masks of same semantic category rather than a single mask such as hyperspectral classification and target detection. We tackle this problem by connecting the prompt, mask and feature together in feature space, where prompt tells the model an object reference in task, mask outputs accurate location information, and feature reflects semantic information. In unified feature space, the cosine distance of features is used to measure the semantic similarity." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": "Before diving into the specific workflows, we introduce two basic interaction modes with a simple situation. Set one point " + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": " as input prompt for image " + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": ", the corresponding feature cube " + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": " and all the generated masks " + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{M} \\in \\mathbb{R}^{h \\times w \\times k}" + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": " from the trained model, where " + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": " has " + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": " binary maps and each represents a single mask. The first interaction mode maps the prompt " + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": " into a feature vector " + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_{(x,y)} \\in \\mathbb{R}^j" + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": " as in Equation (4), where the corresponding valid and smallest mask in " + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": " is used to localize the features in " + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": ", and the mean feature is treated as " + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_{(x,y)}" + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": " for the given prompt. This interaction transforms the prompt from image space to feature space, and makes it possible to explore the semantic" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23052" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 99, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 99, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 99, + 83 + ], + "type": "text", + "content": "similarity." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 112, + 91, + 295, + 105 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 91, + 295, + 105 + ], + "spans": [ + { + "bbox": [ + 112, + 91, + 295, + 105 + ], + "type": "interline_equation", + "content": "\\text {M o d e} (x, y) \\mid \\mathbf {M}, \\mathbf {D} \\rightarrow \\mathbf {d} _ {(x, y)} \\tag {4}", + "image_path": "14d0b54515a80a12f77ac6ff5b3a323a47b1231ff3b1ad7e957c553f676a0205.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 111, + 296, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 111, + 296, + 230 + ], + "spans": [ + { + "bbox": [ + 55, + 111, + 296, + 230 + ], + "type": "text", + "content": "The second interaction mode selects multiple the masks in " + }, + { + "bbox": [ + 55, + 111, + 296, + 230 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 55, + 111, + 296, + 230 + ], + "type": "text", + "content": ", which have similar semantic information with " + }, + { + "bbox": [ + 55, + 111, + 296, + 230 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_{(x,y)}" + }, + { + "bbox": [ + 55, + 111, + 296, + 230 + ], + "type": "text", + "content": " as in Equation (5). The selected masks are represented as " + }, + { + "bbox": [ + 55, + 111, + 296, + 230 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_d" + }, + { + "bbox": [ + 55, + 111, + 296, + 230 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 111, + 296, + 230 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_d \\subseteq \\mathbf{M}" + }, + { + "bbox": [ + 55, + 111, + 296, + 230 + ], + "type": "text", + "content": ". Specifically, we compute a feature representation for each single mask by localizing in " + }, + { + "bbox": [ + 55, + 111, + 296, + 230 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 55, + 111, + 296, + 230 + ], + "type": "text", + "content": " and computing the average, and the cosine similarity between them and " + }, + { + "bbox": [ + 55, + 111, + 296, + 230 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_{(x,y)}" + }, + { + "bbox": [ + 55, + 111, + 296, + 230 + ], + "type": "text", + "content": " are treated as the semantic similarity. With some hand-designed thresholds " + }, + { + "bbox": [ + 55, + 111, + 296, + 230 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 111, + 296, + 230 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 111, + 296, + 230 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_d" + }, + { + "bbox": [ + 55, + 111, + 296, + 230 + ], + "type": "text", + "content": " can be finally segmented. " + }, + { + "bbox": [ + 55, + 111, + 296, + 230 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 111, + 296, + 230 + ], + "type": "text", + "content": " represents different meanings according to the specific task." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 108, + 243, + 295, + 256 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 243, + 295, + 256 + ], + "spans": [ + { + "bbox": [ + 108, + 243, + 295, + 256 + ], + "type": "interline_equation", + "content": "\\text {M o d e} 2: \\left(\\mathbf {d} _ {(x, y)} | \\mathbf {M}, \\mathbf {D}, \\tau\\right)\\rightarrow \\mathbf {M} _ {d} \\tag {5}", + "image_path": "6b1590486e08fe6cd05cc24cc3327e2cfdaceafff037a760e2fbd3d8d3aa845b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 260, + 296, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 260, + 296, + 523 + ], + "spans": [ + { + "bbox": [ + 55, + 260, + 296, + 523 + ], + "type": "text", + "content": "By using both interaction modes adaptively, five hyperspectral tasks can be completed with the trained promptable model directly. (a) Hyperspectral classification (HC). Each category needs at least one prompt, and uses Mode 1 and Mode 2 in turn for each category. For a common closed-set setting, " + }, + { + "bbox": [ + 55, + 260, + 296, + 523 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 260, + 296, + 523 + ], + "type": "text", + "content": " is not necessary since we can use the minimum feature distance between each mask representation and all the prompt representations to decide its category. Figure 3 (b) shows an exemplified workflow. (b) Hyperspectral one-class classification (HOCC). HOCC is similar to H but the difference lies in that HOCC only needs prompts the target category, and its " + }, + { + "bbox": [ + 55, + 260, + 296, + 523 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 260, + 296, + 523 + ], + "type": "text", + "content": " is instantiated as a class pr parameter to control the mask ratio. (c) Hyperspectral taget detection (HTD). Compared to HOCC, HTD acce the target spectra and it needs to be changed to a pron by selecting the pixel with the smallest cosine distance. is instantiated as a distance threshold to convert the dens map to a binary map. (d) Hyperspectral anomaly detion (HAD). HAD is an unsupervised task and there is " + }, + { + "bbox": [ + 55, + 260, + 296, + 523 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_{(x,y)}" + }, + { + "bbox": [ + 55, + 260, + 296, + 523 + ], + "type": "text", + "content": " for HAD. Since HAD is very similar to unsupervis target detection, we use M only and filter out large objec according to mask area to output the final anomaly map." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 523, + 296, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 523, + 296, + 583 + ], + "spans": [ + { + "bbox": [ + 55, + 523, + 296, + 583 + ], + "type": "text", + "content": "Hyperspectral change detection (HCD). Different from the prior tasks, HCD accepts two registered and temporal images to identify the changing locations. The HCD task utilizes both interaction modes for the bi-temporal images. Mode 1 processes the first image to get representation " + }, + { + "bbox": [ + 55, + 523, + 296, + 583 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_{(x,y)}" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 56, + 620, + 296, + 711 + ], + "blocks": [ + { + "bbox": [ + 55, + 590, + 295, + 611 + ], + "lines": [ + { + "bbox": [ + 55, + 590, + 295, + 611 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 295, + 611 + ], + "type": "text", + "content": "Table 2. Instantiation summary of prompts, meaning of " + }, + { + "bbox": [ + 55, + 590, + 295, + 611 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 590, + 295, + 611 + ], + "type": "text", + "content": ", and interaction modes when HyperFree performs different tasks." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 620, + 296, + 711 + ], + "lines": [ + { + "bbox": [ + 56, + 620, + 296, + 711 + ], + "spans": [ + { + "bbox": [ + 56, + 620, + 296, + 711 + ], + "type": "table", + "html": "
TaskPrompt TypeHyperparameter τInteraction Mode
HCOne Point Per CategoriesNo τMode 1, Mode 2
HOCCTarget Category PointClass PriorMode 1, Mode 2
HTDTarget SpectraSimilarity ThresholdMode 1, Mode 2
HADNo PromptMasks Area ThresholdMode 1
HCDNo PromptSimilarity ThresholdMode 1, Mode 2
", + "image_path": "ab5712bd45ec8b0957ff7c3f262bb8419239244063281dbb3c09d2cb7787a65c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 331, + 73, + 541, + 330 + ], + "blocks": [ + { + "bbox": [ + 331, + 73, + 541, + 330 + ], + "lines": [ + { + "bbox": [ + 331, + 73, + 541, + 330 + ], + "spans": [ + { + "bbox": [ + 331, + 73, + 541, + 330 + ], + "type": "image", + "image_path": "c8766bfd9c3a124d04b58db1c422d193da50f8ee338e2709fe18ccf539f45221.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 328, + 365, + 541, + 535 + ], + "blocks": [ + { + "bbox": [ + 313, + 343, + 555, + 363 + ], + "lines": [ + { + "bbox": [ + 313, + 343, + 555, + 363 + ], + "spans": [ + { + "bbox": [ + 313, + 343, + 555, + 363 + ], + "type": "text", + "content": "Figure 4. Exemplified comparison results for five tuning-free tasks" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 328, + 365, + 541, + 535 + ], + "lines": [ + { + "bbox": [ + 328, + 365, + 541, + 535 + ], + "spans": [ + { + "bbox": [ + 328, + 365, + 541, + 535 + ], + "type": "image", + "image_path": "8aa3a5b03f95cc7bde125587c9649d741960c2d44980215513318b6be995fecf.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 541, + 555, + 574 + ], + "lines": [ + { + "bbox": [ + 313, + 541, + 555, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 541, + 555, + 574 + ], + "type": "text", + "content": "Figure 5. In tuning-free manner, HyperFree (1 prompt) can achieve comparable results with specialized models (5 shots) on 5 tasks and 11 datasets." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 582, + 555, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 642 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 642 + ], + "type": "text", + "content": "for each mask, which is input to Mode 2 and compared with the corresponding mask representation with " + }, + { + "bbox": [ + 313, + 582, + 555, + 642 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 313, + 582, + 555, + 642 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 582, + 555, + 642 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 313, + 582, + 555, + 642 + ], + "type": "text", + "content": " of the second image. Once the temporal feature distance exceeds the preset " + }, + { + "bbox": [ + 313, + 582, + 555, + 642 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 582, + 555, + 642 + ], + "type": "text", + "content": ", a change is considered to have happened. The interaction workflow is summarized in Table 2." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 652, + 372, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 652, + 372, + 664 + ], + "spans": [ + { + "bbox": [ + 314, + 652, + 372, + 664 + ], + "type": "text", + "content": "5. Results" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 672, + 422, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 422, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 422, + 685 + ], + "type": "text", + "content": "5.1. Experimental Setup" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "We evaluate the proposed HyperFree on five downstream hyperspectral tasks including HC, HOCC, HTD, HAD and" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23053" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 97, + 556, + 392 + ], + "blocks": [ + { + "bbox": [ + 55, + 65, + 555, + 88 + ], + "lines": [ + { + "bbox": [ + 55, + 65, + 555, + 88 + ], + "spans": [ + { + "bbox": [ + 55, + 65, + 555, + 88 + ], + "type": "text", + "content": "Table 3. Quantitative comparison results on five hyperspectral tasks, where blue numbers indicate the metric ranking of HyperFree. Comparison methods were all trained on each dataset while HyperFree processed the datasets in tuning-free manner." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 97, + 556, + 392 + ], + "lines": [ + { + "bbox": [ + 56, + 97, + 556, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 97, + 556, + 392 + ], + "type": "table", + "html": "
HC Task
LongKou (270 channels)HanChuan (274 channels)
MetricsSVM [30] (5 shot)HybridSN [41] (5 shot)FCN [52] (5 shot)FPGA [68] (5 shot)SSFTT [46] (5 shot)MambaHSI [23] (5 shot)HyperFree (1 prompt)SVM [30] (5 shot)HybridSN [41] (5 shot)FCN [52] (5 shot)FPGA [68] (5 shot)SSFTT [46] (5 shot)MambaHSI [23] (5 shot)HyperFree (1 prompt)
OA82.7748.7886.6791.1889.6692.6593.39(1st)52.6847.7555.5571.4764.8673.3375.47(1st)
AA74.0261.3785.6088.3587.9692.5788.03(3rd)47.7646.1759.7272.0961.2269.3362.35(2nd)
KA78.0435.7282.3088.6687.9590.0091.41(1st)46.8541.3150.1867.5859.6569.1071.56(1st)
HOCC Task
HongHu (270 channels)XiongAn (256 channels)
MetricsOCSVM [43] (5 shot)nnPU [18] (5 shot)BSVM [33] (5 shot)PAN [15] (5 shot)OC Loss [67] (5 shot)T-HOneCls [66] (5 shot)HyperFree (1 prompt)OCSVM [43] (5 shot)nnPU [18] (5 shot)BSVM [33] (5 shot)PAN [15] (5 shot)OC Loss [67] (5 shot)T-HOneCls [66] (5 shot)HyperFree (1 prompt)
F26.3319.1334.8263.6954.7355.9772.52(1st)18.311.7626.3046.3443.0841.3452.50(1st)
P56.4319.7250.7780.0058.2746.5268.61(2nd)39.832.8523.8247.1347.5032.8768.21(1st)
R24.0218.5845.2964.2754.3492.3578.06(2nd)16.081.9857.8353.3247.6160.3865.03(1st)
HTD Task
Airport (205 channels)Cri (46 channels)
MetricsACE [19] (1 shot)CEM [2] (1 shot)GLRT [28] (1 shot)MF [29] (1 shot)HTD-IRN [44] (1 shot)TSTTD [16] (1 shot)HyperFree (1 prompt)ACE [19] (1 shot)CEM [2] (1 shot)GLRT [28] (1 shot)MF [29] (1 shot)HTD-IRN [44] (1 shot)TSTTD [16] (1 shot)HyperFree (1 prompt)
AUC(D,F)0.97940.96030.98010.99160.97450.99290.9806(3rd)0.97350.98930.97370.98910.99750.99870.9791(5th)
AUCODP1.58531.28291.57981.69681.44841.65921.4667(6th)1.20151.45061.20001.45751.39951.61031.4670(2nd)
HAD Task
Beach1 (188 channels)Beach2 (193 channels)
MetricsRXD [40]CRD [22]ADLR [36]LRASR [63]Auto-AD [55]TDD [20]HyperFree (zero shot)RXD [40]CRD [22]ADLR [36]LRASR [63]Auto-AD [55]TDD [20]HyperFree (zero shot)
AUC(D,F)0.98150.94710.45150.74610.95740.98420.9730(3rd)0.90900.85440.79760.82250.94850.96270.9566(2nd)
AUCODP1.25570.97850.56100.85261.12731.13831.9191(1st)1.01770.86700.90640.82801.00971.16881.8697(1st)
HCD Task
River (154 channels)Hermiston (198 channels)
MetricsFC-EF [8] (5 shot)FC-Sc [8] (5 shot)FC-Sd [8] (5 shot)ML-EDAN [35] (5 shot)BIT [3] (5 shot)SST-Former [57] (5 shot)HyperFree (zero shot)FC-EF [8] (5 shot)FC-Sc [8] (5 shot)FC-Sd [8] (5 shot)ML-EDAN [35] (5 shot)BIT [3] (5 shot)SST-Former [57] (5 shot)HyperFree (zero shot)
IoU0.41680.45220.45340.39150.21260.40960.3649(6th)0.37290.37760.48730.32520.52570.53610.5851(1st)
F0.58840.62280.62390.56280.35070.58120.5347(6th)0.54320.54820.65.520.49080.68910.69800.7382(1st)
", + "image_path": "2ad382a0e9a889978bf11da8291f10a1af2ee27d641f466c5387a4e8aef74680.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 56, + 418, + 555, + 460 + ], + "blocks": [ + { + "bbox": [ + 150, + 398, + 459, + 408 + ], + "lines": [ + { + "bbox": [ + 150, + 398, + 459, + 408 + ], + "spans": [ + { + "bbox": [ + 150, + 398, + 459, + 408 + ], + "type": "text", + "content": "Table 4. Compared results with SAM using the identical PMF interaction workflows." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 418, + 555, + 460 + ], + "lines": [ + { + "bbox": [ + 56, + 418, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 56, + 418, + 555, + 460 + ], + "type": "table", + "html": "
HCHOCCHTDHADHCD
OAAAKAFPrecisionRecallAUC(D,F)AUCODPAUC(D,F)AUCODPFIoU
SAM+PMF Interaction66.6272.9959.4134.1330.3539.470.79791.10140.96151.88440.59960.4282
HyperFree93.4085.6691.4272.5268.6178.060.97911.46700.95661.86970.73820.5851
", + "image_path": "9a200ed64e1c1d13f3364e0573a3884791a5491e2ec04c87dd80a7f62c318031.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 480, + 296, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 480, + 296, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 480, + 296, + 588 + ], + "type": "text", + "content": "HCD tasks as in Section 4.3. A total of 11 datasets were used in the evaluation (detailed in Supp.3), and each task had six comparison methods, including both classical and state-of-the-art ones. Since there is no method that can directly complete all five tasks, all comparative methods are tested after being trained on each dataset with five shots per class, while the parameters of HyperFree are fixed for all datasets with 1 prompt or zero-shot as in Table 2. We use ViT-base as backbone to conduct the experiments." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": "HyperFree was trained on the constructed Hyper-Seg dataset using 8 A100 GPUs for approximately 10 hours, which was based on the meta-architecture and parameters of SAM. We use AdamW optimizer with initial learning rate 8e-5 and batch size 16. To help learning the channel-adaptive embedding, we randomly select the image channels at each iteration and match it with corresponding wavelengths. 16 masks and 1 2 points per mask are randomly selected for each input image to train" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 481, + 435, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 481, + 435, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 481, + 435, + 491 + ], + "type": "text", + "content": "the promptable segmentation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 493, + 394, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 493, + 394, + 503 + ], + "spans": [ + { + "bbox": [ + 314, + 493, + 394, + 503 + ], + "type": "text", + "content": "5.2. Main Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 312, + 510, + 555, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 510, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 312, + 510, + 555, + 653 + ], + "type": "text", + "content": "Most tuning-free quantitative results are reported in Table 3 (1 dataset in supplement due to the space limit) and Figure 5. All the comparative methods are trained with 5 shot for each class and HyperFree processes each dataset directly without tuning. For HC, HOCC and HTD, only one point prompt is provided for each class. HAD and HCD does not need any prompt and deal with images in zero-shot manner. Table 3 shows that HyperFree can surpass most of the specialized models with frozen parameters. On both HC and HCD tasks, HyperFree even outperformed previous SOTA results, including an improvement of about 12 points in " + }, + { + "bbox": [ + 312, + 510, + 555, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_1" + }, + { + "bbox": [ + 312, + 510, + 555, + 653 + ], + "type": "text", + "content": " on HC and about 0.5 points in IoU on HCD task." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "text", + "content": "We performed qualitative comparison on five exemplified images in Figure 4, where one SOTA model have been visualized for each task comparison. Obviously, the result maps of HyperFree are more complete and smoother than the ones of specialized models. In HC, HOCC, HTD and" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23054" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 86, + 553, + 148 + ], + "blocks": [ + { + "bbox": [ + 176, + 65, + 432, + 77 + ], + "lines": [ + { + "bbox": [ + 176, + 65, + 432, + 77 + ], + "spans": [ + { + "bbox": [ + 176, + 65, + 432, + 77 + ], + "type": "text", + "content": "Table 5. Ablation results on the usage of the learned weight dictionary." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 86, + 553, + 148 + ], + "lines": [ + { + "bbox": [ + 56, + 86, + 553, + 148 + ], + "spans": [ + { + "bbox": [ + 56, + 86, + 553, + 148 + ], + "type": "table", + "html": "
Correct WavelengthsβkβcHCHOCCHTDHADHCD
OAAAKAF1PrecisionRecallAUC(D,F)AUCODPAUC(D,F)AUCODPF1IoU
✓✓×73.1864.2365.8565.6247.3064.560.97671.46440.94881.84650.57930.4078
✓×✓66.6371.5259.4463.6258.6178.440.97681.46400.95891.87680.70220.5411
×✓✓75.2538.9866.4457.3064.0954.830.95951.42730.95571.86780.68320.5188
✓✓✓93.4085.6691.4272.5268.6178.060.97911.46700.95661.86970.73820.5851
", + "image_path": "e45c2ce6e4623276cf6fba6829e9065536edf0aa9a43cc18adf8188a0a360b94.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 56, + 165, + 173, + 262 + ], + "blocks": [ + { + "bbox": [ + 56, + 165, + 173, + 262 + ], + "lines": [ + { + "bbox": [ + 56, + 165, + 173, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 165, + 173, + 262 + ], + "type": "image", + "image_path": "a6d5d6b9a88aab5b71db603328a561f0ef200b15e536b0c55767951a47248281.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 75, + 265, + 154, + 274 + ], + "lines": [ + { + "bbox": [ + 75, + 265, + 154, + 274 + ], + "spans": [ + { + "bbox": [ + 75, + 265, + 154, + 274 + ], + "type": "text", + "content": "(a) HyperSigma features" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 176, + 165, + 294, + 262 + ], + "blocks": [ + { + "bbox": [ + 176, + 165, + 294, + 262 + ], + "lines": [ + { + "bbox": [ + 176, + 165, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 176, + 165, + 294, + 262 + ], + "type": "image", + "image_path": "3b07dd51824e56815c3a355ffbb7bba73269f2f710e0873a36aca930a8712141.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 200, + 264, + 274, + 274 + ], + "lines": [ + { + "bbox": [ + 200, + 264, + 274, + 274 + ], + "spans": [ + { + "bbox": [ + 200, + 264, + 274, + 274 + ], + "type": "text", + "content": "(b) HyperFree features" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 70, + 313, + 282, + 476 + ], + "blocks": [ + { + "bbox": [ + 55, + 278, + 295, + 312 + ], + "lines": [ + { + "bbox": [ + 55, + 278, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 278, + 295, + 312 + ], + "type": "text", + "content": "Figure 6. t-SNE visualizations of HyperSigma [54] and HyperFree features on WHU-Hi Longkou dataset. We randomly select 1000 pixels for each class." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 70, + 313, + 282, + 476 + ], + "lines": [ + { + "bbox": [ + 70, + 313, + 282, + 476 + ], + "spans": [ + { + "bbox": [ + 70, + 313, + 282, + 476 + ], + "type": "image", + "image_path": "a9efefd3751bea9223204b28fa6a9e2fb9ad4f64cfaadf5bf1fa51d6a1d30c13.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 484, + 295, + 506 + ], + "lines": [ + { + "bbox": [ + 55, + 484, + 295, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 484, + 295, + 506 + ], + "type": "text", + "content": "Figure 7. In tuning manner, HyperFree can achieve overall better results with specialized models on 8 tasks and 14 datasets." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 509, + 295, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 295, + 533 + ], + "type": "text", + "content": "HAD tasks, HyperFree can generate basically the same result as GT, which are difficulty for existing models." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 540, + 148, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 148, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 148, + 552 + ], + "type": "text", + "content": "5.3. Analysis Studies" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 557, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 295, + 665 + ], + "type": "text", + "content": "HyperFree vs SAM. Proposed prompt-mask-feature interaction strategy adapts the pre-trained HyperFree to various downstream tasks. We applied the same strategy to SAM to verify the effectiveness of training on hyperspectral images and reported the results in Table 4. The metrics on most tasks have a decline of about " + }, + { + "bbox": [ + 55, + 557, + 295, + 665 + ], + "type": "inline_equation", + "content": "10\\sim 30" + }, + { + "bbox": [ + 55, + 557, + 295, + 665 + ], + "type": "text", + "content": " points, implying a failed processing. We realize it is not a fair comparison and only use it to reflect the difference of hyperspectral data processing." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "HyperFree vs HyperSigma. HyperSigma is the only open-sourced hyperspectral foundation model [54, 62], which adopts the pre-training and tuning paradigm. Thus, we cannot compare HyperFree with it in tuning-free set-" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 167, + 553, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 167, + 553, + 214 + ], + "spans": [ + { + "bbox": [ + 313, + 167, + 553, + 214 + ], + "type": "text", + "content": "ting directly and compare the t-SNE visualization in Figure 6(a) and Figure 6(b). In contrast, proposed HyperFree has a more compact intra-class distance obviously, supporting the tuning-free workflow of PMF interaction." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 215, + 554, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 215, + 554, + 311 + ], + "spans": [ + { + "bbox": [ + 313, + 215, + 554, + 311 + ], + "type": "text", + "content": "Full-Spectrum Weight Dictionary. To verify stored weights are wavelength-aware, we conducted related ablation studies in Table 5. In HC, HOCC and HCD tasks, a drop in accuracy of about " + }, + { + "bbox": [ + 313, + 215, + 554, + 311 + ], + "type": "inline_equation", + "content": "5\\sim 40" + }, + { + "bbox": [ + 313, + 215, + 554, + 311 + ], + "type": "text", + "content": " points can be observed if we shuffled the correct wavelength order randomly, showing the wavelength-aware property of learned dictionary. Besides, we proved the built two encoding branches with dictionaries " + }, + { + "bbox": [ + 313, + 215, + 554, + 311 + ], + "type": "inline_equation", + "content": "\\beta_{\\mathbf{k}}" + }, + { + "bbox": [ + 313, + 215, + 554, + 311 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 215, + 554, + 311 + ], + "type": "inline_equation", + "content": "\\beta_{\\mathrm{c}}" + }, + { + "bbox": [ + 313, + 215, + 554, + 311 + ], + "type": "text", + "content": " can benefit each other effectively." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 311, + 553, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 311, + 553, + 382 + ], + "spans": [ + { + "bbox": [ + 313, + 311, + 553, + 382 + ], + "type": "text", + "content": "Tuning HyperFree. As an extensive experiment, we have also tested the tuning performance of HyperFree on 8 tasks and 14 datasets in Figure 7, which further included hyperspectral denosing (HD) [27], object tracking (HOT) [61] and unmxing (HU) tasks [37]. The tuning HyperFree has surpassed the previous SOTA models on nearly all tasks." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 382, + 554, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 382, + 554, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 382, + 554, + 467 + ], + "type": "text", + "content": "More analysis. For more analytical experiments, please see the supplementary material. (i) Supp.4.1: Complete qualitative comparison of 11 datasets in tuning-free manner. (ii) Supp.4.2: Complete quantitative and qualitative comparison of 14 datasets in tuning manner. (iii) Supp.4.3: Sensitivity analysis about prompt number, prompt location and the hyperparameter " + }, + { + "bbox": [ + 313, + 382, + 554, + 467 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 382, + 554, + 467 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 514, + 390, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 514, + 390, + 526 + ], + "spans": [ + { + "bbox": [ + 314, + 514, + 390, + 526 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 545, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 714 + ], + "type": "text", + "content": "Hyperspectral images can provide rich spectral information and are widely used in resource monitoring and defense fields. However, its unique properties increase the difficulty of foundation model development such as varied-length channels, high sensitivity to imaging conditions and the acquirement difficulty, making the mainstream P-T being low cost-effectiveness due to the image-by-image tuning. Our work tackles this problem by converting to adapt the visual P-E [17]. Differently, we design a learnable weight dictionary covering full-spectrum to generate the channel-adaptive embeddings and PMF interaction to generate semantic-aware masks for different tasks. We hope HyperFree can speed up the era of \"GPT moment\" for hyperspectral intelligent processing." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23055" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 59, + 72, + 157, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 72, + 157, + 85 + ], + "spans": [ + { + "bbox": [ + 59, + 72, + 157, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 137, + 115, + 149 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 137, + 115, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 137, + 115, + 149 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 157, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 157, + 296, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 157, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 61, + 157, + 296, + 201 + ], + "type": "text", + "content": "[1] Nassim Ait Ali Braham, Conrad M Albrecht, Julien Mairal, Jocelyn Chanussot, Yi Wang, and Xiao Xiang Zhu. Spectral earth: Training hyperspectral foundation models at scale. arXiv preprint arXiv:2408.08447, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 202, + 296, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 202, + 296, + 245 + ], + "spans": [ + { + "bbox": [ + 61, + 202, + 296, + 245 + ], + "type": "text", + "content": "[2] Chein-I Chang. Constrained energy minimization (cem) for hyperspectral target detection: Theory and generalizations. IEEE Transactions on Geoscience and Remote Sensing, 2024. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 247, + 295, + 279 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 247, + 295, + 279 + ], + "spans": [ + { + "bbox": [ + 62, + 247, + 295, + 279 + ], + "type": "text", + "content": "[3] Hao Chen, Zipeng Qi, and Zhenwei Shi. Remote sensing image change detection with transformers. IEEE Transactions on Geoscience and Remote Sensing, 60:1-14, 2021. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 280, + 295, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 280, + 295, + 335 + ], + "spans": [ + { + "bbox": [ + 62, + 280, + 295, + 335 + ], + "type": "text", + "content": "[4] Keyan Chen, Chenyang Liu, Hao Chen, Haotian Zhang, Wenyuan Li, Zhengxia Zou, and Zhenwei Shi. Rsprompter: Learning to prompt for remote sensing instance segmentation based on visual foundation model. IEEE Transactions on Geoscience and Remote Sensing, 2024. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 335, + 295, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 335, + 295, + 380 + ], + "spans": [ + { + "bbox": [ + 62, + 335, + 295, + 380 + ], + "type": "text", + "content": "[5] Gordon Christie, Neil Fendley, James Wilson, and Ryan Mukherjee. Functional map of the world. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6172-6180, 2018. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 380, + 296, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 296, + 435 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 296, + 435 + ], + "type": "text", + "content": "[6] Yezhen Cong, Samar Khanna, Chenlin Meng, Patrick Liu, Erik Rozi, Yutong He, Marshall Burke, David Lobell, and Stefano Ermon. Satmae: Pre-training transformers for temporal and multi-spectral satellite imagery. Advances in Neural Information Processing Systems, 35:197-211, 2022. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 436, + 296, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 436, + 296, + 512 + ], + "spans": [ + { + "bbox": [ + 62, + 436, + 296, + 512 + ], + "type": "text", + "content": "[7] Yezhen Cong, Samar Khanna, Chenlin Meng, Patrick Liu, Erik Rozi, Yutong He, Marshall Burke, David B. Lobell, and Stefano Ermon. Satmae: pre-training transformers for temporal and multi-spectral satellite imagery. In Proceedings of the 36th International Conference on Neural Information Processing Systems, Red Hook, NY, USA, 2024. Curran Associates Inc. 1, 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 514, + 295, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 514, + 295, + 557 + ], + "spans": [ + { + "bbox": [ + 62, + 514, + 295, + 557 + ], + "type": "text", + "content": "[8] Rodrigo Caye Daudt, Bertr Le Saux, and Alexandre Boulch. Fully convolutional siamese networks for change detection. In 2018 25th IEEE international conference on image processing (ICIP), pages 4063-4067. IEEE, 2018. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 558, + 295, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 558, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 62, + 558, + 295, + 634 + ], + "type": "text", + "content": "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 635, + 295, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 295, + 669 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 295, + 669 + ], + "type": "text", + "content": "[10] Alexander FH Goetz, Gregg Vane, Jerry E Solomon, and Barrett N Rock. Imaging spectrometry for earth remote sensing. science, 228(4704):1147-1153, 1985. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "text", + "content": "[11] Xin Guo, Jiangwei Lao, Bo Dang, Yingying Zhang, Lei Yu, Lixiang Ru, Liheng Zhong, Ziyuan Huang, Kang Wu, Dingxiang Hu, et al. Skysense: A multi-modal remote sensing foundation model towards universal interpretation for" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 57, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 55, + 91, + 297, + 127 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 127 + ], + "type": "text", + "content": "This work was supported by the National Natural Science Foundation of China under Grant No. 42325105 and Grant 424B2010." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "text", + "content": "earth observation imagery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27672-27683, 2024. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "type": "text", + "content": "[12] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 163, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 239 + ], + "type": "text", + "content": "[13] Danfeng Hong, Bing Zhang, Xuyang Li, Yuxuan Li, Chenyu Li, Jing Yao, Naoto Yokoya, Hao Li, Pedram Ghamisi, Xiuping Jia, Antonio Plaza, Paolo Gamba, Jon Atli Benediktsson, and Jocelyn Chanussot. Spectralgpt: Spectral remote sensing foundation model. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(8):5227-5244, 2024. 1, 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 240, + 553, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 240, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 316, + 240, + 553, + 296 + ], + "type": "text", + "content": "[14] Danfeng Hong, Bing Zhang, Xuyang Li, Yuxuan Li, Chenyu Li, Jing Yao, Naoto Yokoya, Hao Li, Pedram Ghamisi, Xi-uping Jia, et al. Spectralgpt: Spectral remote sensing foundation model. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 297, + 553, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 351 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 351 + ], + "type": "text", + "content": "[15] Wenpeng Hu, Ran Le, Bing Liu, Feng Ji, Jinwen Ma, Dongyan Zhao, and Rui Yan. Predictive adversarial learning from positive and unlabeled data. In Proceedings of the AAAI conference on artificial intelligence, pages 7806-7814, 2021. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 353, + 553, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 353, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 316, + 353, + 553, + 397 + ], + "type": "text", + "content": "[16] Jinyue Jiao, Zhiqiang Gong, and Ping Zhong. Triplet spectral-wise transformer network for hyperspectral target detection. IEEE Transactions on Geoscience and Remote Sensing, 2023. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 399, + 553, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 399, + 553, + 463 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 553, + 463 + ], + "type": "text", + "content": "[17] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 2, 3, 5, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 465, + 553, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 465, + 553, + 509 + ], + "spans": [ + { + "bbox": [ + 317, + 465, + 553, + 509 + ], + "type": "text", + "content": "[18] Ryuichi Kiryo, Gang Niu, Martinus C Du Plessis, and Masashi Sugiyama. Positive-unlabeled learning with nonnegative risk estimator. Advances in neural information processing systems, 30, 2017. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 510, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 510, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 510, + 553, + 544 + ], + "type": "text", + "content": "[19] Shawn Kraut, Louis L Scharf, and L Todd McWhorter. Adaptive subspace detectors. IEEE Transactions on signal processing, 49(1):1-16, 2001. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 544, + 553, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 544, + 553, + 599 + ], + "spans": [ + { + "bbox": [ + 316, + 544, + 553, + 599 + ], + "type": "text", + "content": "[20] Jingtao Li, Xinyu Wang, Shaoyu Wang, Hengwei Zhao, and Yanfei Zhong. One step detection paradigm for hyperspectral anomaly detection via spectral deviation relationship learning. IEEE Transactions on Geoscience and Remote Sensing, 2024. 1, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 600, + 553, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 600, + 553, + 655 + ], + "spans": [ + { + "bbox": [ + 316, + 600, + 553, + 655 + ], + "type": "text", + "content": "[21] Jiaojiao Li, Zhiyuan Zhang, Yuzhe Liu, Rui Song, Yunsong Li, and Qian Du. Swformer: Stochastic windows convolutional transformer for hybrid modality hyperspectral classification. IEEE Transactions on Image Processing, 2024. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 657, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 657, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 657, + 553, + 690 + ], + "type": "text", + "content": "[22] Wei Li and Qian Du. Collaborative representation for hyperspectral anomaly detection. IEEE Transactions on geoscience and remote sensing, 53(3):1463-1474, 2014. 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[23] Yapeng Li, Yong Luo, Lefei Zhang, Zengmao Wang, and Bo Du. Mambahsi: Spatial-spectral mamba for hyperspectral" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "23056" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "text", + "content": "image classification. IEEE Transactions on Geoscience and Remote Sensing, 62:1-16, 2024. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "type": "text", + "content": "[24] Jie Lian, Lizhi Wang, He Sun, and Hua Huang. Gthad: Gated transformer for hyperspectral anomaly detection. IEEE Transactions on Neural Networks and Learning Systems, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 296, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 296, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 296, + 185 + ], + "type": "text", + "content": "[25] Nanqing Liu, Xun Xu, Yongyi Su, Haojie Zhang, and Heng-Chao Li. Pointsam: Pointly-supervised segment anything model for remote sensing images. arXiv preprint arXiv:2409.13401, 2024. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 295, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 295, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 295, + 239 + ], + "type": "text", + "content": "[26] Fulin Luo, Tianyuan Zhou, Jiamin Liu, Tan Guo, Xiwen Gong, and Jinchang Ren. Multiscale diff-changed feature fusion network for hyperspectral image change detection. IEEE Transactions on Geoscience and Remote Sensing, 61: 1-13, 2023. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 243, + 295, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 295, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 295, + 285 + ], + "type": "text", + "content": "[27] Zhaozhi Luo, Xinyu Wang, Petri Pellikka, Janne Heiskanen, and Yanfei Zhong. Unsupervised adaptation learning for real multiplatform hyperspectral image denoising. IEEE transactions on cybernetics, 2024. 2, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 287, + 295, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 287, + 295, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 287, + 295, + 319 + ], + "type": "text", + "content": "[28] Dimitris Manolakis and Gary Shaw. Detection algorithms for hyperspectral imaging applications. IEEE signal processing magazine, 19(1):29-43, 2002. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 321, + 295, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 295, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 295, + 374 + ], + "type": "text", + "content": "[29] Dimitris Manolakis, Eric Truslow, Michael Pieper, Thomas Cooley, and Michael Brueggeman. Detection algorithms in hyperspectral imaging systems: An overview of practical algorithms. IEEE Signal Processing Magazine, 31(1):24-33, 2013. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 377, + 295, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 377, + 295, + 421 + ], + "spans": [ + { + "bbox": [ + 56, + 377, + 295, + 421 + ], + "type": "text", + "content": "[30] Farid Melgani and Lorenzo Bruzzone. Classification of hyperspectral remote sensing images with support vector machines. IEEE Transactions on geoscience and remote sensing, 42(8):1778-1790, 2004. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 422, + 295, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 295, + 475 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 295, + 475 + ], + "type": "text", + "content": "[31] Mohammed Abdulmajeed Moharram and Divya Meena Sundaram. Land use and land cover classification with hyperspectral data: A comprehensive review of methods, challenges and future directions. Neurocomputing, 536:90-113, 2023. 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 479, + 295, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 479, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 479, + 295, + 521 + ], + "type": "text", + "content": "[32] Aleksandar Petrov, Emanuele La Malfa, Philip Torr, and Adel Bibi. Language model tokenizers introduce unfairness between languages. Advances in Neural Information Processing Systems, 36, 2024. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 523, + 295, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 295, + 578 + ], + "type": "text", + "content": "[33] Rami Piiroinen, Fabian Ewald Fassnacht, Janne Heiskanen, Eduardo Maeda, Benjamin Mack, and Petri Pellikka. Invasive tree species detection in the eastern arc mountains biodiversity hotspot using one class classification. Remote sensing of environment, 218:119-131, 2018. 3, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 580, + 295, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 580, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 580, + 295, + 623 + ], + "type": "text", + "content": "[34] Maciej P Polak and Dane Morgan. Extracting accurate materials data from research papers with conversational language models and prompt engineering. Nature Communications, 15(1):1569, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 624, + 295, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 295, + 677 + ], + "type": "text", + "content": "[35] Jiahui Qu, Shaoxiong Hou, Wenqian Dong, Yunsong Li, and Weiying Xie. A multilevel encoder-decoder attention network for change detection in hyperspectral images. IEEE Transactions on Geoscience and Remote Sensing, 60:1-13, 2021. 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 680, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 295, + 714 + ], + "type": "text", + "content": "[36] Ying Qu, Wei Wang, Rui Guo, Bulent Ayhan, Chiman Kwan, Steven Vance, and Hairong Qi. Hyperspectral anomaly detection through spectral unmixing and dictionary-based low" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 554, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 554, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 554, + 95 + ], + "type": "text", + "content": "rank decomposition. IEEE Transactions on Geoscience and Remote Sensing, 56(8):4391-4405, 2018. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 554, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 554, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 554, + 149 + ], + "type": "text", + "content": "[37] Behnood Rasti, Alexandre Zouaoui, Julien Mairal, and Jocelyn Chanussot. Image processing and machine learning for hyperspectral unmixing: An overview and the hysupp python package. IEEE Transactions on Geoscience and Remote Sensing, 2024. 2, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 151, + 554, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 151, + 554, + 227 + ], + "spans": [ + { + "bbox": [ + 316, + 151, + 554, + 227 + ], + "type": "text", + "content": "[38] C. J. Reed, R. Gupta, S. Li, S. Brockman, C. Funk, B. Clipp, K. Keutzer, S. Candido, M. Uytendaele, and T. Darrell. Scale-mae: A scale-aware masked autoencoder for multiscale geospatial representation learning. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 4065-4076, Los Alamitos, CA, USA, 2023. IEEE Computer Society. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 228, + 554, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 554, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 554, + 304 + ], + "type": "text", + "content": "[39] Colorado J Reed, Ritwik Gupta, Shufan Li, Sarah Brockman, Christopher Funk, Brian Clipp, Kurt Keutzer, Salvatore Candido, Matt Uytendaele, and Trevor Darrell. Scale-mae: A scale-aware masked autoencoder for multiscale geospatial representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4088-4099, 2023. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 304, + 554, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 304, + 554, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 304, + 554, + 348 + ], + "type": "text", + "content": "[40] Irving S Reed and Xiaoli Yu. Adaptive multiple-band cfar detection of an optical pattern with unknown spectral distribution. IEEE transactions on acoustics, speech, and signal processing, 38(10):1760-1770, 1990. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 350, + 554, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 554, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 554, + 403 + ], + "type": "text", + "content": "[41] Swalpa Kumar Roy, Gopal Krishna, Shiv Ram Dubey, and Bidyut B Chaudhuri. Hybridsn: Exploring 3-d-2-d cnn feature hierarchy for hyperspectral image classification. IEEE Geoscience and Remote Sensing Letters, 17(2):277-281, 2019. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 404, + 554, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 404, + 554, + 458 + ], + "spans": [ + { + "bbox": [ + 316, + 404, + 554, + 458 + ], + "type": "text", + "content": "[42] Pranab Sahoo, Ayush Kumar Singh, Sriparna Saha, Vinija Jain, Samrat Mondal, and Aman Chadha. A systematic survey of prompt engineering in large language models: Techniques and applications. arXiv preprint arXiv:2402.07927, 2024.3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 460, + 554, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 554, + 502 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 554, + 502 + ], + "type": "text", + "content": "[43] Bernhard Scholkopf, Robert C Williamson, Alex Smola, John Shawe-Taylor, and John Platt. Support vector method for novelty detection. Advances in neural information processing systems, 12, 1999. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 504, + 554, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 504, + 554, + 547 + ], + "spans": [ + { + "bbox": [ + 316, + 504, + 554, + 547 + ], + "type": "text", + "content": "[44] Dunbin Shen, Xiaorui Ma, Wenfeng Kong, Jianjun Liu, Jie Wang, and Hongyu Wang. Hyperspectral target detection based on interpretable representation network. IEEE Transactions on Geoscience and Remote Sensing, 2023. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 548, + 554, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 554, + 581 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 554, + 581 + ], + "type": "text", + "content": "[45] Hongjun Su, Zhaoyue Wu, Huihui Zhang, and Qian Du. Hyperspectral anomaly detection: A survey. IEEE Geoscience and Remote Sensing Magazine, 10(1):64-90, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 582, + 554, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 554, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 554, + 624 + ], + "type": "text", + "content": "[46] Le Sun, Guangrui Zhao, Yuhui Zheng, and Zebin Wu. Spectral-spatial feature tokenization transformer for hyperspectral image classification. IEEE Transactions on Geoscience and Remote Sensing, 60:1-14, 2022. 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 625, + 554, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 554, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 554, + 658 + ], + "type": "text", + "content": "[47] Siyu Sun, Jun Liu, and Wei Li. Spatial invariant tensor self-representation model for hyperspectral anomaly detection. IEEE Transactions on Cybernetics, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 658, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 554, + 713 + ], + "type": "text", + "content": "[48] Xian Sun, Peijin Wang, Wanxuan Lu, Zicong Zhu, Xiaonan Lu, Qibin He, Junxi Li, Xue Rong, Zhujun Yang, Hao Chang, et al. Ringmo: A remote sensing foundation model with masked image modeling. IEEE Transactions on Geoscience and Remote Sensing, 61:1-22, 2022. 4" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23057" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 715 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 139 + ], + "type": "text", + "content": "[49] Xian Sun, Peijin Wang, Wanxuan Lu, Zicong Zhu, Xiaonan Lu, Qibin He, Junxi Li, Xue Rong, Zhujun Yang, Hao Chang, Qinglin He, Guang Yang, Ruiping Wang, Jiwen Lu, and Kun Fu. Ringmo: A remote sensing foundation model with masked image modeling. IEEE Transactions on Geoscience and Remote Sensing, 61:1-22, 2023. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 173 + ], + "type": "text", + "content": "[50] Adam Van Etten, Dave Lindenbaum, and Todd M Bacastow. Spacenet: A remote sensing dataset and challenge series. arXiv preprint arXiv:1807.01232, 2018. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 239 + ], + "type": "text", + "content": "[51] Adam Van Etten, Jacob Shermeyer, Daniel Hogan, Nicholas Weir, and Ryan Lewis. Road network and travel time extraction from multiple look angles with spacenet data. In IGARSS 2020-2020 IEEE International Geoscience and Remote Sensing Symposium, pages 3920-3923. IEEE, 2020. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 241, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 294, + 274 + ], + "type": "text", + "content": "[52] Di Wang, Bo Du, and Liangpei Zhang. Fully contextual network for hyperspectral scene parsing. IEEE Transactions on Geoscience and Remote Sensing, 60:1-16, 2021. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 276, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 330 + ], + "type": "text", + "content": "[53] Di Wang, Qiming Zhang, Yufei Xu, Jing Zhang, Bo Du, Dacheng Tao, and Liangpei Zhang. Advancing plain vision transformer toward remote sensing foundation model. IEEE Transactions on Geoscience and Remote Sensing, 61:1-15, 2023. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 332, + 294, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 294, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 294, + 386 + ], + "type": "text", + "content": "[54] Di Wang, Meiqi Hu, Yao Jin, Yuchun Miao, Jiaqi Yang, Yichu Xu, Xiaolei Qin, Jiaqi Ma, Lingyu Sun, Chenxing Li, et al. Hypersigma: Hyperspectral intelligence comprehension foundation model. arXiv preprint arXiv:2406.11519, 2024. 2, 3, 4, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 388, + 294, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 442 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 442 + ], + "type": "text", + "content": "[55] Shaoyu Wang, Xinyu Wang, Liangpei Zhang, and Yanfei Zhong. Auto-ad: Autonomous hyperspectral anomaly detection network based on fully convolutional autoencoder. IEEE Transactions on Geoscience and Remote Sensing, 60:1-14, 2021. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 445, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 510 + ], + "type": "text", + "content": "[56] Sheng Wang, Kaiyu Guan, Chenhui Zhang, Qu Zhou, Sibo Wang, Xiaocui Wu, Chongya Jiang, Bin Peng, Weiye Mei, Kaiyuan Li, et al. Cross-scale sensing of field-level crop residue cover: Integrating field photos, airborne hyperspectral imaging, and satellite data. Remote Sensing of Environment, 285:113366, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 512, + 294, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 294, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 294, + 567 + ], + "type": "text", + "content": "[57] Yanheng Wang, Danfeng Hong, Jianjun Sha, Lianru Gao, Lian Liu, Yonggang Zhang, and Xianhui Rong. Spectral-spatial-temporal transformers for hyperspectral image change detection. IEEE Transactions on Geoscience and Remote Sensing, 60:1-14, 2022. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 568, + 294, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 294, + 622 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 294, + 622 + ], + "type": "text", + "content": "[58] Philine Warnke, Kerstin Cuhls, Ulrich Schmoch, Lea Daniel, Liviu Andreescu, Bianca Dragomir, Radu Gheorghiu, Catalina Baboschi, Adrian Curaj, Marjukka Parkkinen, et al. 100 radical innovation breakthroughs for the future. 2019. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 624, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 689 + ], + "type": "text", + "content": "[59] Nicholas Weir, David Lindenbaum, Alexei Bastidas, Adam Van Etten, Sean McPherson, Jacob Shermeyer, Varun Kumar, and Hanlin Tang. Spacenet mvoi: A multi-view overhead imagery dataset. In Proceedings of the IEEE/cvf international conference on computer vision, pages 992-1001, 2019. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 691, + 294, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 715 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 715 + ], + "type": "text", + "content": "[60] Walter F Wiggins and Ali S Tejani. On the opportunities and risks of foundation models for natural language processing in" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 552 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 333, + 72, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 72, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 333, + 72, + 553, + 94 + ], + "type": "text", + "content": "radiology. Radiology: Artificial Intelligence, 4(4):e220119, 2022.3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "text", + "content": "[61] Yinan Wu, Licheng Jiao, Xu Liu, Fang Liu, Shuyuan Yang, and Lingling Li. Domain adaptation-aware transformer for hyperspectral object tracking. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 2, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 141, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 206 + ], + "type": "text", + "content": "[62] Zhitong Xiong, Yi Wang, Fahong Zhang, Adam J Stewart, Joëlle Hanna, Damian Borth, Ioannis Papoutsis, Bertrand Le Saux, Gustau Camps-Valls, and Xiao Xiang Zhu. Neural plasticity-inspired foundation model for observing the earth crossing modalities. arXiv e-prints, pages arXiv-2403, 2024. 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "type": "text", + "content": "[63] Yang Xu, Zebin Wu, Jun Li, Antonio Plaza, and Zhihui Wei. Anomaly detection in hyperspectral images based on low-rank and sparse representation. IEEE Transactions on Geoscience and Remote Sensing, 54(4):1990-2000, 2015. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 253, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 553, + 285 + ], + "type": "text", + "content": "[64] Zongyin Yang, Tom Albrow-Owen, Weiwei Cai, and Tawfique Hasan. Miniaturization of optical spectrometers. Science, 371(6528):eabe0722, 2021. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 286, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 330 + ], + "type": "text", + "content": "[65] Alina Zare, Changzhe Jiao, and Taylor Glenn. Discriminative multiple instance hyperspectral target characterization. IEEE transactions on pattern analysis and machine intelligence, 40(10):2342-2354, 2017. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 331, + 553, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 386 + ], + "type": "text", + "content": "[66] Hengwei Zhao, Xinyu Wang, Jingtao Li, and Yanfei Zhong. Class prior-free positive-unlabeled learning with taylor variational loss for hyperspectral remote sensing imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16827-16836, 2023. 2, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 388, + 553, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 430 + ], + "type": "text", + "content": "[67] Hengwei Zhao, Yanfei Zhong, Xinyu Wang, and Hong Shu. One-class risk estimation for one-class hyperspectral image classification. IEEE Transactions on Geoscience and Remote Sensing, 2023. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 432, + 553, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 432, + 553, + 485 + ], + "spans": [ + { + "bbox": [ + 316, + 432, + 553, + 485 + ], + "type": "text", + "content": "[68] Zhuo Zheng, Yanfei Zhong, Ailong Ma, and Liangpei Zhang. Fpga: Fast patch-free global learning framework for fully end-to-end hyperspectral image classification. IEEE Transactions on Geoscience and Remote Sensing, 58(8):5612-5626, 2020. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 487, + 553, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 487, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 316, + 487, + 553, + 552 + ], + "type": "text", + "content": "[69] Yanfei Zhong, Xin Hu, Chang Luo, Xinyu Wang, Ji Zhao, and Liangpei Zhang. Whu-hi: Uav-borne hyperspectral with high spatial resolution (h2) benchmark datasets and classifier for precise crop identification based on deep convolutional neural network with crf. Remote Sensing of Environment, 250:112012, 2020. 1, 2" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23058" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/HyperGLM_ HyperGraph for Video Scene Graph Generation and Anticipation/07619f38-6a65-46d1-ae2a-3885ad0e2f6d_content_list.json b/2025/HyperGLM_ HyperGraph for Video Scene Graph Generation and Anticipation/07619f38-6a65-46d1-ae2a-3885ad0e2f6d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b9dc4de7c9a63e1431efbdf0d6cc451074834485 --- /dev/null +++ b/2025/HyperGLM_ HyperGraph for Video Scene Graph Generation and Anticipation/07619f38-6a65-46d1-ae2a-3885ad0e2f6d_content_list.json @@ -0,0 +1,1838 @@ +[ + { + "type": "text", + "text": "HyperGLM: HyperGraph for Video Scene Graph Generation and Anticipation", + "text_level": 1, + "bbox": [ + 99, + 130, + 898, + 155 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Trong-Thuan Nguyen $^{1\\dagger}$ , Pha Nguyen $^{1\\dagger}$ , Jackson Cothren $^{1}$ , Alper Yilmaz $^{2}$ , Khoa Luu $^{1}$ $^{1}$ University of Arkansas $^{2}$ Ohio State University", + "bbox": [ + 158, + 179, + 836, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}\\{$ thuann, panguyen, jcothre, khoaluu\\}@uark.edu $^{2}$ yilmaz.15@osu.edu uark-cviu.github.io/projects/HyperGLM", + "bbox": [ + 204, + 217, + 792, + 252 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7ec01ce8a7d7ad73de9fa91943f0ab8555b159ceb724809c9ac3eb2b093a6f09.jpg", + "image_caption": [ + "What are the roles of the people?" + ], + "image_footnote": [], + "bbox": [ + 140, + 299, + 854, + 359 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5069f06678a48c5320e6bbad12cc3935df7e1314e9541189e070d67711d3fc70.jpg", + "image_caption": [ + "What is the relationship between objects?", + "Person-1 holding the Umbrella-1 and standing beside Person-2. Person-1 and Person-2 are waiting for Person-3 in her yellow dress. Then, Person -1 talking with Person-3.", + "Figure 1. Our HyperGLM framework supports Video Scene Graph Generation, Anticipation, and Reasoning. HyperGLM constructs scene graphs from observed video frames and predicts relationships in unseen frames by leveraging a unified hypergraph for temporal modeling and comprehensive understanding." + ], + "image_footnote": [], + "bbox": [ + 143, + 378, + 418, + 472 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "It appears that Person-1 is picking up his family, as Person-3 is being dropped off from the helicopter in the rainy weather. Person-2 is likely serving as security.", + "bbox": [ + 514, + 381, + 836, + 420 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "What might the next interaction be?", + "bbox": [ + 627, + 422, + 844, + 436 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Person-1 and Person-3 are talking while waiting for another family member to get off the helicopter.", + "bbox": [ + 514, + 440, + 816, + 465 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "What might they do after moving away?", + "bbox": [ + 607, + 468, + 846, + 481 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "After arriving by helicopter, Person-1, Person-3, and another woman will likely move indoors to rest.", + "bbox": [ + 511, + 486, + 821, + 512 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 555, + 326, + 570 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multimodal LLMs have advanced vision-language tasks but still struggle with understanding video scenes. To bridge this gap, Video Scene Graph Generation (VidSGG) has emerged to capture multi-object relationships across video frames. However, prior methods rely on pairwise connections, limiting their ability to handle complex multi-object interactions and reasoning. To this end, we propose the Multimodal Large Language Models (LLMs) on a Scene HyperGraph (HyperGLM), promoting reasoning about multi-way interactions and higher-order relationships. Our approach uniquely integrates entity scene graphs, which capture spatial relationships between objects, with a procedural graph that models their causal transitions, forming a unified HyperGraph. Significantly, HyperGLM enables reasoning by injecting this unified HyperGraph into LLMs. Additionally, we introduce a new Video Scene Graph Reasoning (VSGR) dataset featuring 1.9M frames from third-person, egocentric, and drone views and support five tasks. Empirically, HyperGLM consistently outperforms state-of-the-art methods, effectively modeling and reasoning complex relationships in diverse scenes.", + "bbox": [ + 88, + 579, + 485, + 881 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 555, + 643, + 570 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, Multimodal Large Language Models (LLMs) [2, 10] have set new benchmarks, with vision-language models excelling in diverse multimodal tasks. However, fully understanding dynamic video scenes remains a significant challenge for applications like autonomous driving, intelligent surveillance, human-object interaction, and multimedia analysis. Towards this goal, Video Scene Graph Generation (VidSGG) [18, 52] has emerged as a critical task for capturing multi-object relationships across video frames. In particular, VidSGG enables high-level tasks such as event forecasting [36, 43, 45], video captioning [24, 38, 40], and video question answering [20, 30, 32, 41] by constructing detailed representations of entities and their interactions.", + "bbox": [ + 509, + 580, + 908, + 776 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, prior VidSGG methods and datasets have limited capacity for comprehensive video understanding. Traditional scene graph-based methods [4, 34, 35, 42] only model pairwise object relationships within single frames, making it challenging to capture higher-order relationships and temporal dependencies in real-world scenarios. Additionally, existing benchmark datasets [18, 34, 35, 52] focus is primarily confined to Scene Graph Generation (SGG) and Scene", + "bbox": [ + 509, + 779, + 910, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ Equal contribution.", + "bbox": [ + 109, + 887, + 223, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "29150", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Graph Anticipation (SGA) tasks, lacking annotations for reasoning tasks such as Video Question Answering (VQA), Video Captioning (VC), and Relation Reasoning (RR).", + "bbox": [ + 89, + 90, + 485, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose the Multimodal LLMs on a Scene HyperGraph (HyperGLM) approach to promote reasoning about multi-way interactions and higher-order relationships through a unified HyperGraph and LLMs, as illustrated in Fig. 1. To achieve this goal, we uniquely incorporate entity scene graphs, which capture spatial relationships between objects, with a procedural graph that models their causal transitions across video frames, forming a unified HyperGraph, as shown in Fig. 2. Significantly, our HyperGLM approach allows reasoning by injecting this unified HyperGraph into LLMs. In addition, we introduce a novel Video Scene Graph Reasoning (VSGR) dataset, comprising 1.9 million video frames surpassing existing benchmark datasets [18, 34, 35, 47, 52] in scale and annotation depth. Specifically, our VSGR dataset includes videos from third-person, egocentric, and drone perspectives. It supports five tasks: Scene Graph Generation, Scene Graph Anticipation, Video Question Answering, Video Captioning, and Relation Reasoning. Notably, our VSGR dataset introduces a new Relation Reasoning task, setting it apart from existing video scene graph datasets, as shown in the comparison in Table 1.", + "bbox": [ + 91, + 138, + 486, + 454 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contributions of this Work. This work presents three contributions to advance Video Scene Graph Generation. First, we introduce Multimodal LLMs on a Scene HyperGraph, leveraging hyperedges and LLMs for reasoning about multiway interactions and higher-order relationships. Second, we develop a new Video Scene Graph Reasoning dataset, surpassing existing scale and annotation depth benchmarks. Our VSGR dataset primarily supports five tasks within diverse video scenes. Finally, the proposed HyperGLM consistently outperforms state-of-the-art methods across all five tasks.", + "bbox": [ + 89, + 454, + 483, + 604 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 617, + 232, + 633 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we review advances in scene graph generation and hypergraph applications in computer vision, then discuss existing limitations and the advantages of our approach.", + "bbox": [ + 89, + 642, + 482, + 688 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Scene Graph Generation", + "text_level": 1, + "bbox": [ + 89, + 696, + 316, + 710 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Scene Graph Generation has significantly advanced with transformer-based models [6, 13, 17, 25, 26, 51] that have become benchmarks due to their efficiency and state-of-the-art performance. Recent work [3, 21, 33] focuses on reducing bias and enhancing mean recall for rare predicates by integrating external knowledge and applying unbiased contextual augmentation, particularly in dynamic video contexts. Open-vocabulary methods [15, 27] supported by vision-language models handle unseen object and relationship classes, improving generalization. Additionally, generative models (e.g., diffusion-based methods [9, 53]) leverage scene graphs for efficient image and scene synthesis. More-", + "bbox": [ + 89, + 719, + 485, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "over, spatial-temporal methods [4, 34-36, 42] effectively capture dynamic object relationships in videos. Recently, Large Language Models [23] have been utilized to enhance triplet extraction and alignment in weakly supervised SGG.", + "bbox": [ + 511, + 90, + 906, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. HyperGraphs", + "text_level": 1, + "bbox": [ + 511, + 162, + 658, + 178 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "HyperGraphs have been adopted in computer vision to model complex multimodal data and capture higher-order relationships. Unlike traditional graph-based approaches, HyperGraphs connect multiple nodes through hyperedges, enabling multi-way relationships. They enhance Graph Neural Networks (GNNs) [12, 54] by allowing the modeling of more sophisticated interactions. Recent advancements, such as HyperGraph Convolution [1] and HyperGraph Attention [22], have further improved GNNs by capturing relationships beyond simple pairwise connections. Therefore, HyperGraph-based models effectively handle temporal dependencies and complex interactions, significantly boosting performance in tasks like accident anticipation [43], group activity recognition [28, 55], and video question answering [44, 46, 50].", + "bbox": [ + 511, + 186, + 908, + 398 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Discussion", + "text_level": 1, + "bbox": [ + 511, + 409, + 630, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Limitations in Prior Methods. The methods introduced in Section 2.1, based on Progressive Feature Fusion [39, 51], Batch-Progressive Transformer Encoding [6, 13, 17, 25, 26], Spatial-Temporal Context Integration [4, 34, 42], and Memory-Guided Temporal Consistency [7, 33, 35], have advanced VidSGG. However, these methods struggle to model higher-order relationships and complex temporal dynamics. Specifically, Progressive Feature Fusion and Batch-Progressive Transformer Encoding are limited in capturing long-term temporal dependencies, with the former lacking long-term context due to frame-by-frame processing and the latter only addressing short-term dependencies. Similarly, Spatial-Temporal Context Integration and Memory-Guided Temporal Consistency inadequately represent multi-object interactions across video frames and insufficiently capture the temporal evolution required for higher-order relationships.", + "bbox": [ + 511, + 431, + 908, + 674 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Advantages of Our Approach. Our HyperGLM approach, Multimodal LLMs on a Scene HyperGraph, promotes reasoning about multi-way interactions and high-order relationships. As illustrated in Fig. 2, HyperGLM enhances the model's ability to interpret complex relationships and anticipate intricate video dynamics. Towards this goal, we uniquely integrate entity scene graphs, which is introduced in Sec. 3 to capture spatial interactions between objects in each frame and a procedural graph, which is presented in Sec. 4.1 to model their causal evolution. In our unified approach, hyperedges connect multiple nodes to capture higher-order relationships distinguished from traditional pairwise methods [4, 34, 35, 42]. In addition, the procedural graph enables multi-step transitions for anticipating future interactions or relationships, and reduces bias by generalizing", + "bbox": [ + 511, + 674, + 910, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "29151", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/aecdeabde8cae2c808904b822671766915543edf0224f666bbe937c2a4a3b079.jpg", + "image_caption": [ + "Figure 2. (a) To model the temporal transition, a simple approach can be using two scene graphs $\\mathbf{G}_{\\mathbf{t}}$ and $\\mathbf{G}_{\\mathbf{t} + 1}$ . (b) Another procedure graph can present this temporal modeling. (c) Our unified HyperGraph in Fig. 2c integrates both entity scene graph to capture spatial relationships and the procedural graph to model the temporal evolution. HyperEdge represents person sitting on couch, holding, then playing guitar, whereas holding $\\rightarrow$ playing describes a chain of interactions. HyperGraph is presented in 3D." + ], + "image_footnote": [], + "bbox": [ + 89, + 85, + 500, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "infrequent relationship categories. Furthermore, our HyperGLM approach leverages fundamental mathematical properties: permutation equivariance ensures that the HyperGraph structure remains consistent under any permutation of node labels, and invariance to hyperedge order preserves semantic meanings regardless of the node visit order during random walks which is defined in Alg. 1. The theoretical foundations and mathematical properties are detailed in the Appendices.", + "bbox": [ + 89, + 272, + 483, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Problem Formulation", + "text_level": 1, + "bbox": [ + 89, + 420, + 295, + 435 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we define two tasks, including Scene Graph Generation (SGG) and Scene Graph Anticipation (SGA). Graph vertex sequence is represented as $\\{V_{\\mathbf{G}}\\mid 1\\leq t\\leq T\\}$ , where each set of vertex $V_{\\mathbf{G}}$ contains object features, bounding boxes, and categories. The scene graph for each frame $t$ , denoted by $\\mathbf{G}_t = \\big(V_{\\mathbf{G}},E_{\\mathbf{G}} = \\{r^{ij}\\mid 1\\leq i < j\\leq |V_{\\mathbf{G}}|\\}$ , consists of all pairwise relationships between objects, with $r^{ij}$ representing the relationship category between $v_{i}$ and $v_{j}$ .", + "bbox": [ + 89, + 446, + 483, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We aim to develop a process $p_{\\theta}: (V_{\\mathbf{G}} \\times V_{\\mathbf{G}}) \\to r^{ij}$ to predict the relationship $r^{ij}$ between each object pair $(v_t^i, v_t^j)$ in $V_{\\mathbf{G}}$ . We define task-specific queries $\\mathbf{Q}_{\\mathrm{SGG}}$ , and $\\mathbf{Q}_{\\mathrm{SGA}}$ to direct our unified model to perform on each specific task. The objective for each task is to minimize the negative likelihood of the predicted scene graph $\\mathbf{G}_t$ to the truth predicate set $Y_t = \\{y^{ij} \\mid 1 \\leq i < j \\leq |V_{\\mathbf{G}}|\\}$ on categories indexed by $k$ . Scene Graph Generation (SGG) generates the scene graph for each frame $t$ from $t = 1$ to $t = T$ , which is defined as:", + "bbox": [ + 89, + 566, + 483, + 702 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} \\mathbb {E} _ {\\mathbf {G} _ {t}, Y _ {t}} \\left[ - \\sum_ {\\left(v _ {t} ^ {i, j}\\right)} \\sum_ {k} \\left(y _ {k} ^ {i j} \\log p _ {\\theta} \\left(r ^ {i j} \\mid \\mathbf {Q} _ {\\mathrm {S G G}}\\right) _ {k}\\right) \\right] \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 124, + 707, + 483, + 742 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Scene Graph Anticipation (SGA) anticipates the scene graphs for future frames, generating predictions $\\mathbf{G}_{t + n}$ , where $n$ denotes the anticipation horizon, formulated as:", + "bbox": [ + 89, + 750, + 483, + 794 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} \\mathbb {E} _ {\\mathbf {G} _ {\\leq t}, Y _ {t}} \\left[ - \\sum_ {\\left(v _ {t + n} ^ {i, j}\\right)} \\sum_ {k} \\left(y _ {k} ^ {i j} \\log p _ {\\theta} \\left(r ^ {i j} \\mid \\mathbf {Q} _ {\\mathrm {S G A}}\\right) _ {k}\\right) \\right] \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 811, + 483, + 848 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In Eqns. (1) and (2), $t$ indexes the current frame. While SGG predicts the scene graph at frame $t$ , SGA reads scene graphs up to frame $t$ to forecast the graph in future frame $t + n$ .", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/fda9408f652dfc66d7ebbd26c6062a1f40b9350b6215790cdbe771985ea3814d.jpg", + "image_caption": [ + "Figure 3. Our HyperGLM framework comprises an image encoder, MLP projector, temporal aggregator, unified HyperGraph, and language model. It processes video frames by encoding each frame with the image encoder and MLP, extracting spatio-temporal features through image patch grids to generate $N$ spatial tokens per frame. The temporal aggregator compresses the $T \\times N$ embeddings over time. The MLP projector then transforms these visual embeddings into the language feature space as frame tokens, interleaved with language tokens, and fed into the Large Language Models." + ], + "image_footnote": [], + "bbox": [ + 514, + 87, + 924, + 189 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Our Proposed HyperGLM Approach", + "text_level": 1, + "bbox": [ + 511, + 297, + 848, + 315 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we present our approach, which incorporates a unified HyperGraph into the LLMs, as illustrated in Fig. 3.", + "bbox": [ + 511, + 324, + 906, + 354 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Video Scene (Hyper)Graphs", + "text_level": 1, + "bbox": [ + 511, + 363, + 764, + 380 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Traditional VidSGG methods [4, 34, 42] construct scene graphs $\\mathbf{G}_t$ to represent scene entities and their relationships as in Fig. 2a. However, these graph-based approaches do not capture higher-order relationships with temporal dependencies for video understanding [36]. To model this property, we propose a novel HyperGraph-based framework that constructs a unified HyperGraph $\\mathcal{H} = (V_{\\mathcal{H}}, E_{\\mathcal{H}})$ , representing spatial relationships within individual frames and temporal transitions across frames, illustrated in Fig. 4. HyperGraph extends traditional graph structures by allowing hyperedges to connect more than two nodes, making them particularly effective for modeling these higher-order relationships. To leverage this property, our unified HyperGraph (see Fig. 2c) integrates entity scene graphs $\\mathbf{G}_t$ for each frame, capturing spatial interactions as introduced in Sec. 3, along with a procedural graph $\\mathbf{P}$ that models their causal transitions, as detailed in Sec. 4.1. We unify these components using a random-walk HyperGraph construction, which is defined in Alg. 1 to capture higher-order connectivity patterns and structural semantics, thereby approximating subgraph matching between the entity scene graphs and the procedural graph. This integration allows our approach to model current subjects' interactions and anticipate their future relationships.", + "bbox": [ + 511, + 387, + 908, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Procedural Graph Construction. SGA objective is to predict the set of relationships $E_{\\mathbf{G}}$ in the next frame based on the current set of relationships. To model the temporal evolution of causal relationships between objects across video frames, we introduce a procedural graph $\\mathbf{P} = (V_{\\mathbf{P}},E_{\\mathbf{P}})$ , serving as the temporal counterpart to the entity scene graphs $\\mathbf{G}_t$ as shown in Fig. 2b. In particular, the procedural graph $\\mathbf{P}$ can vary across datasets, modeling relationship transitions.", + "bbox": [ + 511, + 734, + 908, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Toward this goal, we model and denote the set of vertices in $\\mathbf{P}$ as $V_{\\mathbf{P}} = \\{r_t^{ij} \\in V_{\\mathbf{G}} \\mid 1 \\leq t \\leq T\\} \\neq$ representing distinct relationship categories. The set of edges in $\\mathbf{P}$ , i.e.,", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "29152", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8f8900f6112c23b2fe5219cfb188a8be75705feac1da87af2c44f397671aeef8.jpg", + "image_caption": [ + "Figure 4. Our Video Scene HyperGraph, including entity graphs and a procedural graph, as defined in Sec. 4.1. Blue nodes represent entities, while green nodes denote relationships. The entity graph captures spatial relationships (subject $\\rightarrow$ relationship $\\rightarrow$ object), whereas the procedural graph models relationship transitions $(\\rightarrow)$ . Hyperedges are visualized as polygons, encapsulating interactions through chains of relationships. For instance, a hyperedge illustrates a person picking up, holding, opening, and reading a book while sitting on a couch. HyperGraph is presented in 3D, see Supplementary video." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 903, + 289 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$E_{\\mathbf{P}} = \\{(r_m,r_n)\\}$ represent possible causal transitions between these relationships, where an edge $(r_m,r_n)$ indicates that relationship $r_m$ causally lead to relationship $r_n$ . We quantify causal transitions by calculating transition probabilities $w(r_m,r_n)$ via their observed frequencies as in Eqn. (3).", + "bbox": [ + 89, + 349, + 486, + 428 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nw \\left(r _ {m}, r _ {n}\\right) = \\frac {\\sum_ {t = 1} ^ {T - 1} \\sum_ {r _ {t , t + 1} ^ {i j}} \\mathbb {1} \\left(r _ {t} ^ {i j} = r _ {m} \\wedge r _ {t + 1} ^ {i j} = r _ {n}\\right)}{\\sum_ {t = 1} ^ {T - 1} \\sum_ {r _ {t , t + 1} ^ {i j}} \\mathbb {1} \\left(r _ {t} ^ {i j} = r _ {m}\\right)} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 449, + 483, + 531 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbb{1}(\\cdot)$ is the indicator function that counts transitions from relationship $r_m$ at current frame $t$ to $r_n$ at next frame $t + 1$ . Next, self-loops $w(r_m, r_m)$ are removed from the graph, and these probabilities are normalized as in Eqn. (4).", + "bbox": [ + 89, + 541, + 483, + 604 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\sum_ {r _ {n} \\in E _ {\\mathbf {P}}} w \\left(r _ {m}, r _ {n}\\right) = 1 \\quad \\text {f o r a l l} r _ {m} \\in E _ {\\mathbf {P}} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 613, + 483, + 646 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By leveraging these probabilities, the procedural graph $\\mathbf{P}$ enhances the prediction of future relationships. For each relationship $r_t^{ij}$ in frame $t$ , the most probable relationship $r_{t + 1}^{ij}$ in the next frame $t + 1$ is determined as in Eqn. (5).", + "bbox": [ + 89, + 655, + 483, + 717 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nr _ {t + 1} ^ {i j} = \\arg \\max _ {r _ {n}} P \\left(r _ {n} \\mid r _ {t} ^ {i j}, v ^ {i, j}\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 727, + 483, + 751 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $P(r_{n} \\mid r_{t}^{ij}, v^{i,j}) = w(r_{t}^{ij}, r_{n}) \\times v^{i,j}$ is the probability of transitioning from relationship $r_{t}^{ij}$ to $r_{n}$ , looking at object features. This allows the model to anticipate future interactions based on established temporal patterns.", + "bbox": [ + 89, + 763, + 483, + 825 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "HyperGraph Construction. To incorporate spatial relationships between objects in each frame and their causal temporal transitions, we construct a unified HyperGraph $\\mathcal{H}$ that integrates the entity scene graphs $\\mathbf{G}_t$ and the procedural graph $\\mathbf{P}$ . Specifically, a HyperGraph is an augmented", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "representation combining the entity scene graphs with the procedural graph. This unified structure merges spatial and temporal relationships into a single graph, enabling the use of conventional graph algorithms while preserving the complex interactions captured by the HyperGraph. Mathematically, a unified HyperGraph $\\mathcal{H} = (V_{\\mathcal{H}}, E_{\\mathcal{H}})$ is defined as in Eqn. (6).", + "bbox": [ + 511, + 349, + 906, + 441 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {H} = \\left(\\bigcup_ {t = 1} ^ {T} V _ {\\mathbf {G} _ {t}} \\cup V _ {\\mathbf {P}}, \\quad \\bigcup_ {t = 1} ^ {T} E _ {\\mathbf {G} _ {t}} \\cup E _ {\\mathbf {P}}\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 468, + 906, + 511 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $V_{\\mathcal{H}}$ includes all entity nodes $v_{t}^{i}$ from each $\\mathbf{G}_t$ and the relationship type nodes $V_{\\mathbf{P}}$ . The hyperedge set $E_{\\mathcal{H}}$ includes pairwise relationships $E_{\\mathbf{G}_t}$ within each $\\mathbf{G}_t$ , capturing spatial relationships and temporal transition edges $E_{\\mathbf{P}}$ from $\\mathbf{P}$ , modeling the evolution of causal relationships across frames.", + "bbox": [ + 511, + 532, + 906, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Random-walk Algorithm. We employ the random walks algorithm outlined in Alg. 1 to sample representative substructures from the unified HyperGraph $\\mathcal{H}$ , which captures connectivity patterns and mitigates the NP-hardness of exact subgraph matching. Specifically, these walks alternate between nodes and hyperedges, preserving the multi-node connections intrinsic to hyperedges and capturing the complexity of multi-object relationships and their transition across video frames. In each walk, a hyperedge $h_i$ aggregates the visited nodes, thereby encapsulating higher-order relationships. For example, in the entity scene graph $\\mathbf{G}_t$ , a walk might traverse from a \"person\" to a \"cup\" via the \"holding\", resulting in the hyperedge $h_i = \\{\\text{person}, \\text{cup}\\}$ . Similarly, within the procedural graph $\\mathbf{P}$ , a walk might transition through a sequence of interactions such as \"holding\", \"placing\", and \"releasing\", forming the hyperedge $h_i = \\{\\text{holding}, \\text{placing}, \\text{releasing}\\}$ . Therefore, we generate sampled hyperedges $E_{\\text{sampled}} = \\{h_i \\mid i = 1, \\dots, N_w\\}$ as in L25 of Alg. 1 by conducting multiple random walks.", + "bbox": [ + 511, + 613, + 908, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "29153", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Random-walk for HyperGraph Construction." + ], + "code_body": "Require: $\\mathcal{H} = (V_{\\mathcal{H}},E_{\\mathcal{H}})$ , Number of Walks $N_{w}$ , Walk Length $N_{l}$ \nEnsure: $\\mathcal{H}' = (V_{\\mathcal{H}},E_{\\mathcal{H}}^{\\prime})$ \n1: Initialize $E_{\\mathrm{sampled}}\\gets \\emptyset$ \n2: for $i = 1$ to $N_{w}$ do \n3: Select $v_{\\mathrm{start}}\\in V_{\\mathcal{H}}$ uniformly at random \n4: Initialize walk sequence $S\\leftarrow [v_{\\mathrm{start}}]$ \n5: for $j = 1$ to $N_{l}$ do \n6: if $j$ is odd then Node to HyperEdge \n7: $v_{\\mathrm{current}}\\gets S[j]$ \n8: $\\mathcal{E}_{v_{\\mathrm{current}}}\\gets \\{h\\in E_{\\mathcal{H}}\\mid v_{\\mathrm{current}}\\in h\\}$ \n9: if $\\mathcal{E}_{v_{\\mathrm{current}}}\\neq \\emptyset$ then \n10: Select $h_j\\in \\mathcal{E}_{v_{\\mathrm{current}}}$ \n11: Append $h_j$ to $S$ \n12: end if \n13: else HyperEdge to Node \n14: $h_\\mathrm{current}\\gets S[j]$ \n15: $V_{h_{\\mathrm{current}}}\\gets h_{\\mathrm{current}}$ \n16: Select $v_{j}\\in V_{h_{\\mathrm{current}}}$ \n17: Append $v_{j}$ to $S$ \n18: end if \n19: end for \n20: $h_i\\gets \\{v\\in S\\mid v\\in V_H\\}$ ▷Form new HyperEdge \n21: if $h_i\\notin E_{\\mathcal{H}}\\cup E_{\\mathrm{sampled}}$ then \n22: $E_{\\mathrm{sampled}}\\gets E_{\\mathrm{sampled}}\\cup \\{h_i\\}$ \n23: end if \n24: end for \n25: $E_{\\mathcal{H}}'\\gets E_{\\mathcal{H}}\\cup E_{\\mathrm{sampled}}$ \n26: return $\\mathcal{H}' = (V_{\\mathcal{H}},E_{\\mathcal{H}}')$", + "bbox": [ + 89, + 109, + 486, + 554 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Multimodal LLMs on HyperGraphs", + "text_level": 1, + "bbox": [ + 89, + 561, + 405, + 579 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Formulation. Given an input video $\\mathbf{V}$ and a task query $\\mathbf{Q}$ , we aim to generate a target answer sequence $\\mathbf{A}$ of length $L$ . Especially, by modeling a HyperGraph $\\mathcal{H}$ , the target answer $\\mathbf{A}$ is generated via a process illustrated in Fig. 3, defined as:", + "bbox": [ + 89, + 585, + 485, + 648 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np (\\mathbf {A} | \\mathbf {V}, \\mathbf {Q}, \\mathcal {H}) = \\prod_ {i = 0} ^ {L - 1} p \\left(x _ {i} | \\mathbf {V}, \\mathbf {Q}, \\mathcal {H}, x _ {< i}\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 655, + 483, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $x_{< i}$ denotes the preceding token sequence, and $\\mathbf{A} = [x_0,\\dots ,x_i,\\dots ,x_{L - 1}]$ is the sequence of answer tokens reasoning video $\\mathbf{V}$ by question $\\mathbf{Q}$ and HyperGraph $\\mathcal{H}$ in Fig. 4.", + "bbox": [ + 89, + 704, + 483, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As illustrated in Fig. 5, our process begins with the generation, constructing an entity scene graph $\\mathbf{G}_t = (V_t^e, E_{\\mathbf{G}})$ for each frame $t$ to capture detected objects and their relationships. Relationship anticipation employs the procedural graph $\\mathbf{P}$ to model temporal evolution and predict future interactions. The model then engages in reasoning using the HyperGraph $\\mathcal{H}$ and verification through video captioning to ensure contextual relevance and accuracy. It refines understanding through clarification, generating answers by reasoning over the video and $\\mathcal{H}$ . In scene forecasting, it predicts", + "bbox": [ + 88, + 750, + 485, + 901 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/e24b450fbb708082214c556f322598cd7a6660f62faa37051563a4f518dc227c.jpg", + "table_caption": [ + "Table 1. Comparisons of video scene graph datasets. SGG, SGA, VQA, VC, and RR represent Scene Graph Generation, Scene Graph Anticipation, Video Question Answering, Video Captioning, and Relation Reasoning." + ], + "table_footnote": [], + "table_body": "
Dataset#FramesTasksAnnotations
SGGSGAVQAVCRRBboxRelationText
SportsHHI [47]11.4K✓XXXX✓✓X
Action Genome [18]234.3K✓✓XXX✓✓X
AeroEye [35]261.5K✓XXXX✓✓X
ASPIRe [34]1.6M✓XXXX✓✓X
PVSG [52]153K✓X✓✓X✓✓✓
VSGR (Ours)1.9M✓✓✓✓✓✓✓✓
", + "bbox": [ + 511, + 127, + 910, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "subsequent interactions by identifying new relationships and adapting to evolving contexts using transition probabilities in $\\mathbf{P}$ , anticipating future relationships $r_{t+n}^{ij}$ and improving conversational flow. Finally, the hypothesis allows the model to propose and test conjectures about the underlying themes or intentions, synthesizing information from previous steps and leveraging $\\mathcal{H}$ to explain patterns and forecast results.", + "bbox": [ + 511, + 223, + 906, + 329 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Video Scene Graph Reasoning Dataset", + "text_level": 1, + "bbox": [ + 511, + 343, + 859, + 359 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Limitations of Current Datasets. Existing benchmarks (see Table 1) primarily support SGG and SGA, limiting their applicability to reasoning tasks. They suffer from inadequate support for VQA, VC, and RR, shallow annotations that fail to capture intricate object interactions, insufficient modeling of temporal dynamics, and poor multimodal integration. Thus, our VSGR dataset supports SGG, SGA, VQA, VC, and RR, enabling the reasoning capabilities of LLMs.", + "bbox": [ + 511, + 368, + 908, + 491 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Dataset Construction", + "text_level": 1, + "bbox": [ + 511, + 500, + 714, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Data Acquisition Stage. We source videos from the ASPIRe [34] and AeroEye [35] datasets. While the ASPIRe dataset offers diverse, richly annotated videos emphasizing dynamic interactions and temporal changes, the AeroEye dataset offers drone-captured footage across various scenes. Comprehension Tasks via Question-Answering. We introduce tasks that leverage fine-grained relationships from scene graphs, extending Scene Graph Generation to focus on relation understanding and subject/object interpretation using triplets and leverage GPT-4/GPT-3.5 model for language generation.", + "bbox": [ + 511, + 522, + 908, + 689 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Video Captioning (VC). We generate 82,532 video-caption pairs, resulting in about 22 captions per video. Our process involves (1) extracting triplets from cropped video frames focusing on foreground objects, (2) generating background descriptions based on these triplets, and (3) combining the foreground triplets and background descriptions to produce captions. The average length is 893 characters, surpassing the PVSG [52] dataset in quantity and detail.", + "bbox": [ + 511, + 689, + 908, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Video Question Answering (VQA). We develop 74,856 question-answer pairs by designing questions that explore diverse relationships, selecting subject and object categories to verify specific relationships, or choosing triplets to assess their uniqueness. This results in an average of approximately 20 questions per video, which exceeds existing benchmarks", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "29154", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5a024d78c7f75b4f89f88278cb1bfc3c4464f8cca8a020c22486d42daafce9a5.jpg", + "image_caption": [ + "Figure 5. An example of the diversified context within the streaming dialog in our VSGR dataset. Best viewed in color and zooming in." + ], + "image_footnote": [], + "bbox": [ + 96, + 97, + 486, + 171 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Step-1: Generation", + "text_level": 1, + "bbox": [ + 106, + 174, + 222, + 185 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Given bounding boxes: person-1 [657, 116, 155, 577], person-2 [425, 211, 178, 450], ball-1 [534, 437, 42, 54], ball-2 [670, 350, 37, 65]. What are the relationships between these objects?", + "bbox": [ + 107, + 186, + 475, + 218 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The video depicts including person-1 towards the right side and closer to ball-2. person-2 is situated to the left and is nearer to ball-1.", + "bbox": [ + 107, + 223, + 475, + 246 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Step-2: Relationship Anticipation", + "text_level": 1, + "bbox": [ + 106, + 262, + 305, + 273 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "What interactions might occur between the objects in the next three frame based on ?", + "bbox": [ + 107, + 281, + 447, + 304 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The person-1 will interact with ball-2 by picking it up or dribbling it. Similarly, person-2 may engage with ball-1 corresponding to", + "bbox": [ + 109, + 311, + 475, + 334 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Step-3: Reasoning", + "text_level": 1, + "bbox": [ + 106, + 349, + 218, + 361 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "What purpose or context can be inferred within", + "bbox": [ + 107, + 369, + 325, + 380 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "person-1 and person-2 work with ball-1 and ball-2, suggesting drills like dribbling or shooting. Their separation implies independent practice, but their proximity hints at a possible shift to coordinated team drills later in the session", + "bbox": [ + 107, + 388, + 477, + 421 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Step-4: Verification", + "text_level": 1, + "bbox": [ + 514, + 93, + 635, + 104 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Does person-1 move towards ball-2 and begin dribbling, and does person-2 start practicing shots with ball-1 through ?", + "bbox": [ + 517, + 109, + 885, + 132 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Yes, both individuals are focused on separate balls. person-1 continues with ball-1, and person-2 remains with ball-2, indicating individual practice rather than switching activities or interacting with each other.", + "bbox": [ + 517, + 138, + 885, + 169 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Step-5: Clarification", + "text_level": 1, + "bbox": [ + 517, + 176, + 643, + 188 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "If instead, person-1 picks up ball-2 and passes it to person-2, how would that change the dynamics of?", + "bbox": [ + 519, + 190, + 888, + 210 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This would shift the dynamics from individual practice to a collaborative activity, with person-1 passing ball-2 to person-2. The interaction would signal the start of coordinated play, possibly initiating a game involving both ball-1 and ball-2 and highlighting teamwork between person-1 and person-2.", + "bbox": [ + 519, + 212, + 888, + 252 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Step-6: Scene Forecasting", + "text_level": 1, + "bbox": [ + 519, + 262, + 681, + 273 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Given , what are the likely relationship changes between objects in the upcoming phase?", + "bbox": [ + 522, + 276, + 888, + 299 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The next action could involve person-1 passing ball-2 to person-2 while person-2 passes ball-1 to person-1. Alternatively, person-1 might dribble ball-2 toward person-2, initiating a one-on-one game.", + "bbox": [ + 521, + 305, + 888, + 335 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Step-7: Hypothesis", + "text_level": 1, + "bbox": [ + 519, + 349, + 640, + 362 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "What if person-3 [600, 200, 160, 500] enters and approaches the others? How might that affect the interactions?", + "bbox": [ + 521, + 366, + 888, + 387 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "person-1, person-2, and person-3 could engage in coordinated actions, such as three-person passing drills using ball-1 and ball-2. The interaction would evolve from individual practice to collaborative teamwork.", + "bbox": [ + 522, + 393, + 890, + 422 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "such as the MSRVTT-QA [49] and MSVD-QA [48] datasets.", + "bbox": [ + 89, + 457, + 483, + 472 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Relation Reasoning (RR). Using the annotated scene graphs, we produce 61,120 relation reasoning tasks by selecting partial information as an incomplete input. Each task requires the model to deduce relationships among entities and identify the categories of the subject and object. With an average of approximately 16 tasks per video, our VSGR dataset provides a substantial collection for evaluating models' abilities in relational reasoning and scene understanding. Question and Answer Validation. To ensure the quality and complexity of questions, we implement a rigorous validation process that combines generation by LLMs with human refinement. Initially, GPT-4/GPT-3.5 generates queries based on the triplets extracted from the videos. Human annotators are trained with specific guidelines that emphasize clarity, relevance, and appropriate challenge levels, then review and refine questions. They enhance the questions by ensuring they are directly answerable from the video content and require careful reasoning about the depicted interactions and relationships.", + "bbox": [ + 88, + 474, + 483, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We apply strict filtering criteria to improve the quality of the dataset. First, we eliminate questions that do not require video context and can be answered using general world knowledge, ensuring that models must rely on visual information from the videos. Second, we exclude questions that LLMs can answer correctly, increasing the challenge and utility of the dataset in evaluating advanced reasoning abilities. Finally, independent annotators perform a second round to review and evaluate the quality of the refined questions.", + "bbox": [ + 88, + 763, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Dataset Comparison", + "text_level": 1, + "bbox": [ + 511, + 455, + 705, + 473 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As reported in Table 1, our VSGR dataset represents a substantial advancement in video scene graph benchmarks. Our dataset comprises 3,748 videos and 1,841,243 frames, surpassing existing datasets in scale. Unlike other datasets that address only a limited subset of tasks, our dataset offers comprehensive task coverage, facilitating multifaceted evaluations of LLMs. In addition, our ground truth enriches relation annotations with comprehensive textual descriptions, enabling sophisticated reasoning and relationship predictions, as illustrated in Fig. 5. Additionally, our VSGR dataset incorporates diverse viewpoints, including third-person, ego-centric, and drone perspectives, enhancing its generalization.", + "bbox": [ + 511, + 481, + 908, + 662 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Experiment Results", + "text_level": 1, + "bbox": [ + 511, + 681, + 702, + 698 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.1. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 707, + 728, + 723 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We leverage our VSGR dataset across five tasks. In addition, we utilize the PVSG [52] dataset for the SGG task and the Action Genome [18] dataset for the SGA task.", + "bbox": [ + 511, + 732, + 906, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Model Configuration. We operate the CLIP-ViT-L-336 [8, 37] to encode each video frame into ten tokens (one CLS token and nine from $3 \\times 3$ average pooling). These tokens are fed into a two-layer MLP connector to the Mistral7B-Instruct [19] language model. For training, we apply LoRA [16] to all linear layers with a rank of 128 and a scaling factor of 256, omitting vision-language alignment [31]. We train for two epochs with a batch size of 128 over 16", + "bbox": [ + 511, + 779, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "29155", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d0a01bfd9fa68d49d6cdd8ac7d69f94218428a12f7d82e080192790acb0992a8.jpg", + "image_caption": [ + "Figure 6. Comparison of Recall (R) and mean Recall (mR) at different numbers of hyperedges for the SGG task on the VSGR dataset." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 285, + 183 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/046d8a22a37e1a284f7f5f8a29220ae77a2315b14a16d2f5c3501b83a8f87b6d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 285, + 88, + 478, + 183 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/25e240ff134a120ad587ba8710e00a8ea98ea68745151793b640854a22b90a87.jpg", + "table_caption": [ + "Table 2. Comparison (\\%) on the VSGR and Action Genome datasets for the Scene Graph Anticipation (SGA) task at varying video input fractions $\\mathcal{F}$ ." + ], + "table_footnote": [], + "table_body": "
FMethodAction GenomeVSGR
R/mR@10R/mR@20R/mR@50R/mR@10R/mR@20R/mR@50
0.3STTran+ [5]13.9 / 3.521.6 / 7.340.8 / 20.312.0 / 4.019.0 / 8.035.7 / 18.0
DSGDetr+ [11]14.3 / 3.621.8 / 7.641.3 / 21.212.5 / 4.219.5 / 8.336.0 / 19.0
STTran++ [5]15.4 / 6.227.2 / 14.148.6 / 32.214.0 / 6.522.0 / 11.039.0 / 25.1
DSGDetr++ [11]16.8 / 8.429.0 / 16.748.9 / 32.314.5 / 7.022.5 / 12.040.0 / 26.0
SceneSayerODE [36]23.3 / 13.332.5 / 20.145.1 / 33.018.0 / 9.526.0 / 15.042.0 / 30.0
SceneSayerSDE [36]25.9 / 15.635.7 / 23.147.4 / 37.119.5 / 11.027.5 / 17.044.0 / 33.5
HyperGraph (Ours)26.5 / 14.836.2 / 22.149.3 / 37.218.8 / 9.327.2 / 16.342.5 / 27.4
HyperGLM (Ours)27.5 / 15.837.0 / 24.550.0 / 38.019.0 / 10.028.0 / 16.543.0 / 27.5
0.5STTran+ [5]14.9 / 3.722.6 / 7.642.9 / 21.413.5 / 4.221.0 / 8.537.0 / 19.0
DSGDetr+ [11]15.2 / 3.923.1 / 8.043.3 / 22.213.8 / 4.521.5 / 9.038.0 / 20.0
STTran++ [5]16.6 / 6.629.1 / 14.751.5 / 33.415.5 / 7.023.5 / 12.541.0 / 26.5
DSGDetr++ [11]17.4 / 8.430.5 / 17.051.9 / 33.916.0 / 7.524.0 / 13.042.0 / 28.0
SceneSayerODE [36]26.4 / 14.336.6 / 21.449.8 / 36.020.5 / 11.029.5 / 16.546.1 / 32.5
SceneSayerSDE [36]28.4 / 16.338.6 / 25.151.4 / 39.921.5 / 12.531.0 / 18.548.0 / 35.7
HyperGraph (Ours)29.2 / 16.439.3 / 23.252.1 / 38.720.3 / 12.229.8 / 17.146.2 / 33.1
HyperGLM (Ours)30.0 / 17.040.5 / 27.553.5 / 40.521.5 / 11.531.5 / 18.546.5 / 30.0
0.7STTran+ [5]16.6 / 4.225.1 / 8.547.2 / 24.015.0 / 5.024.0 / 10.541.0 / 21.5
DSGDetr+ [11]16.8 / 4.325.3 / 8.847.4 / 24.715.5 / 5.224.5 / 11.042.0 / 22.0
STTran++ [5]19.0 / 7.732.8 / 17.156.8 / 36.817.0 / 8.027.0 / 14.045.0 / 29.0
DSGDetr++ [11]19.8 / 9.534.1 / 19.256.7 / 37.217.5 / 8.528.0 / 15.046.1 / 30.0
SceneSayerODE [36]32.1 / 16.542.8 / 24.455.6 / 39.623.5 / 13.033.5 / 19.051.0 / 36.0
SceneSayerSDE [36]33.3 / 18.144.0 / 27.356.4 / 44.424.5 / 14.535.7 / 21.053.0 / 38.0
HyperGraph (Ours)34.3 / 19.245.2 / 25.357.2 / 42.122.3 / 13.234.2 / 19.350.4 / 35.3
HyperGLM (Ours)35.7 / 19.546.1 / 30.058.2 / 44.025.1 / 13.535.5 / 21.551.0 / 33.5
", + "bbox": [ + 91, + 244, + 483, + 468 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "iterations on $4 \\times$ GPUs, taking approximately six hours.", + "bbox": [ + 89, + 483, + 462, + 498 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Metrics. We evaluate the SGG and SGA tasks using the Recall and mean Recall scores. In addition, we evaluate the VQA and RR tasks by Accuracy, Precision, Recall, and F1 scores. For the VC task, we utilize CIDEr, MENTOR, ROUGE-L, and BLEU-4 scores to validate our performance. Settings. For the SGG and SGA tasks, we adopt the evaluation settings based on [18, 36]. In particular, the model is provided with raw video frames and must detect objects using a pre-trained detector (i.e. Faster R-CNN) and predict or anticipate their relationships. Especially for the SGA, we set the initial video input fraction $(\\mathcal{F})$ to 0.9, following [36].", + "bbox": [ + 89, + 500, + 483, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.2. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 679, + 243, + 695 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Hypergraph Parameters. The number of walks $(N_w)$ and walk length $(N_l)$ are introduced in Alg. 1, directly impacting the capacity to capture high-order relationships by determining the number of hyperedges. Although more hyperedges increase relational diversity, they can also introduce redundancy beyond an optimal point. Specifically, a higher $N_w$ broadens the range of sampled relationships, while a moderate $N_l$ balances depth. Our experiments demonstrate that optimal performances are achieved with 60 hyperedges $(N_w = 60$ and $N_l = 7)$ , shown in Fig. 6. Further experiments on the SGG task are provided in Table 3, and experiments on these parameters are included in the Appendices. Video Input Fraction. We adjust the initial video input", + "bbox": [ + 89, + 703, + 485, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/647a19bc53823d4bbb847b4c5b9ec0b065fbffcde419a7e7f402c11e182d80d4.jpg", + "table_caption": [ + "Table 3. Comparison $(\\%)$ on the VSGR and PVSG datasets for the Scene Graph Generation (SGG) task at Recall (R) and mean Recall (mR)." + ], + "table_footnote": [], + "table_body": "
MethodPVSGVSGR
R/mR@20R/mR@50R/mR@100R/mR@20R/mR@50R/mR@100
Transformer [52]4.0 / 1.84.4 / 1.94.9 / 2.025.7 / 6.334.5 / 6.543.5 / 7.0
HIG [34]4.6 / 1.94.9 / 2.15.4 / 2.223.8 / 5.731.1 / 5.940.4 / 6.9
CYCLO [35]5.8 / 2.06.1 / 2.26.7 / 2.329.4 / 7.136.4 / 7.747.7 / 7.7
HyperGraph (Ours)6.5 / 2.27.0 / 2.47.5 / 2.631.6 / 7.838.8 / 8.350.3 / 8.5
HyperGLM (Ours)7.5 / 2.88.1 / 3.78.5 / 3.935.8 / 9.242.3 / 10.154.7 / 10.4
", + "bbox": [ + 511, + 114, + 906, + 185 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f484caec26019d8937035ef7f7c1b1a6e6cabe23bb3ef410756ccc6c8388f6ba.jpg", + "table_caption": [ + "Table 4. Comparison (%) on the VSGR and Action Genome datasets for the Scene Graph Anticipation (SGA) task at Recall (R) and mean Recall (mR)." + ], + "table_footnote": [], + "table_body": "
MethodAction GenomeVSGR
R/mR@10R/mR@20R/mR@50R/mR@10R/mR@20R/mR@50
STTran+ [5]17.5 / 4.626.8 / 9.249.6 / 24.316.5 / 5.526.0 / 11.043.0 / 23.0
DSGDetr+ [11]17.9 / 4.727.7 / 9.751.4 / 25.917.0 / 6.027.0 / 11.544.5 / 24.5
STTran++ [5]20.2 / 8.935.7 / 18.460.2 / 38.819.0 / 9.530.0 / 15.549.0 / 31.0
DSGDetr++ [11]22.2 / 11.437.1 / 21.061.0 / 39.519.5 / 10.031.0 / 16.050.0 / 32.5
SceneSayerODE [36]36.6 / 17.848.3 / 27.461.3 / 43.426.5 / 14.037.5 / 20.055.5 / 38.0
SceneSayerSDE [36]37.3 / 20.848.6 / 30.961.6 / 46.827.5 / 16.038.5 / 22.058.2 / 40.0
HyperGraph (Ours)37.5 / 19.149.3 / 31.462.3 / 47.428.4 / 17.539.3 / 22.457.6 / 41.5
HyperGLM (Ours)38.8 / 22.351.5 / 33.065.2 / 48.630.2 / 18.141.1 / 23.559.3 / 43.4
", + "bbox": [ + 513, + 219, + 906, + 316 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "fraction, $\\mathcal{F}$ , to 0.3, 0.5, and 0.7 for the SGA task. This adjustment allows the model to learn from varying observed portions and predict the unseen segment. Table 2 indicates that increasing the portion of the seen video improves performance, suggesting that additional visual context is beneficial. In addition, Table 4 further confirms that the default setting at a higher input fraction ( $F = 0.9$ ) leads to optimal performance. We also present additional settings with varying video input fractions for the SGA task in the Appendices.", + "bbox": [ + 511, + 321, + 908, + 458 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.3. Comparison with State-of-the-Arts", + "text_level": 1, + "bbox": [ + 511, + 468, + 816, + 484 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3 demonstrates that HyperGLM significantly outperforms existing methods on the PVSG and VSGR datasets, achieving the highest $R@20$ scores of $7.5\\%$ and $35.8\\%$ , respectively. By leveraging hyperedges connecting multiple nodes within the HyperGraph, HyperGLM effectively captures complex object interactions and spatial dependencies, transcending traditional pairwise methods to generate more accurate and detailed scene graph representations. Notably, integrating a procedural graph within the HyperGraph reduces bias. Our approach enhances mean Recall, reaching improvements of $2.8\\%$ on the PVSG dataset and $9.2\\%$ on the VSGR dataset, thereby addressing the long-tail distribution challenges that have struggled in previous methods [34, 35, 52]. Furthermore, the LLM enhances the capacity of HyperGLM to infer and predict intricate relationships embedded within the unified HyperGraph, resulting in improved performance compared to HyperGraph, which shows a decrease of $4.2\\%$ at $R@20$ on the VSGR dataset.", + "bbox": [ + 511, + 491, + 908, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Table 4, our HyperGLM approach outperforms existing SGA methods on the Action Genome and VSGR datasets, achieving R@10 scores of $35.7\\%$ and $25.1\\%$ , respectively. This improvement stems from integrating procedural graphs that model causal relationships within the HyperGraph structure. In contrast, the SceneSayer [36] method relies on NeuralODE and NeuralSDE to capture the latent dynamics of object interaction evolution. Our procedural graphs enable multi-step transitions that explicitly", + "bbox": [ + 511, + 763, + 910, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "29156", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c3a87d810387a52aaf9465f5d76b7bd91f7cb2b28781fde1fd2757228c01145d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 104, + 74, + 500, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2e9f9ce2e537e4124e4e155635084f5750b6183684a70cba399a13624e2c0002.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 74, + 895, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3389de1b55da6d5d50ae1f3cccb1c8e11ba6b50c06de7816a66ea78b023b6009.jpg", + "image_caption": [ + "a) Scene Graph Generation" + ], + "image_footnote": [], + "bbox": [ + 91, + 176, + 498, + 382 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a4edd7237696b72899c03efdbebcea8fcf847cef256b9ec964a8af052a749164.jpg", + "image_caption": [ + "b) Scene Graph Anticipation", + "Figure 7. Qualitative comparison of our HyperGLM approach versus SceneSayerSDE [36] for SGG and SGA. Red and green edge labels denote incorrect and correct predictions, respectively. Our HyperGLM approach effectively captures the evolving interactions between person-1 and bicycle-1 or person-2 and person-3 over time and anticipates interactions in unseen video frames, while SceneSayerSDE confuses to similar predicates. Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 500, + 176, + 901, + 382 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8acbf083e8546b004500d65226cbdb3f4a7b578a36f96feb499bc1c332877d51.jpg", + "table_caption": [ + "Table 5. Comparison (%) on VSGR for the Video Question Answering." + ], + "table_footnote": [], + "table_body": "
MethodAccuracyPrecisionRecallF1 Score
Video-ChatGPT [32]33.235.132.333.6
Video-LLaVA-7B [30]43.143.841.742.7
MovieChat [41]43.544.242.643.4
Chat-UniVi-7B [20]44.345.643.244.4
HyperGLM (Ours)45.447.244.345.7
", + "bbox": [ + 91, + 448, + 480, + 535 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4a22900c91ca4021cd26016cb333c5a413bfd50478f3163afdc9fdc811ef2663.jpg", + "table_caption": [ + "Table 6. Comparison (%) on VSGR for the Video Captioning." + ], + "table_footnote": [], + "table_body": "
MethodCIDErMENTORROUGE-LBLEU-4
MV-GPT [38]57.137.562.547.2
CoCap [40]54.329.461.842.8
UniVL + MELTR [24]50.528.160.042.1
HyperGLM (Ours)54.530.764.948.8
", + "bbox": [ + 91, + 549, + 482, + 623 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "capture the temporal evolution of relationships by modeling sequential interactions and their dependencies. In contrast, NeuralODE and NeuralSDE primarily focus on continuous-time dynamics, which may limit their effectiveness in handling discrete, multi-step relational changes as illustrated in Fig. 7b. Significantly, the HyperGraph is injected into the LLM, allowing the model to capture complex temporal patterns, resulting in better results than the HyperGraph.", + "bbox": [ + 89, + 628, + 482, + 748 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In addition to the improvements in SGG and SGA shown in Fig. 7, Tables 5, 6, and 7 demonstrate the significant improvements using our HyperGLM approach in the VQA, VC, and RR tasks. In VQA, the hyperedges connect multiple objects, enabling HyperGLM to capture context-rich interactions more effectively, reaching an accuracy of $45.4\\%$ . For VC, the HyperGraph structure supports evolving object connections over time, allowing our HyperGLM approach to generate captions that describe the relationship between objects and capture the temporal flow of interactions coher-", + "bbox": [ + 89, + 750, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ef56cb900e55844f8cc8069c30b1dd4e47d88bb563910d1fb274c5c2b44396c9.jpg", + "table_caption": [ + "Table 7. Comparison (%) on VSGR for the Relation Reasoning." + ], + "table_footnote": [], + "table_body": "
MethodAccuracyPrecisionRecallF1 Score
Video-LLaVA-7B [30]41.342.540.241.3
MA-LMM [14]42.843.741.842.7
LLaMA-VID-7B [29]44.145.243.544.3
HyperGLM (Ours)47.248.446.547.4
", + "bbox": [ + 513, + 448, + 905, + 525 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ently, achieving scores of $54.5\\%$ at CIDEr and $64.9\\%$ at ROUGE-L. In RR task, the HyperGraph effectively manages intricate dependencies between multiple objects, resulting in precise relational inferences with an accuracy of $47.2\\%$ .", + "bbox": [ + 511, + 542, + 906, + 603 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 625, + 633, + 640 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we have introduced HyperGLM, a novel VidSGG method that integrates scene hypergraph information into LLMs for context-aware and precise scene interpretation. Our approach effectively models complex interactions and higher-order relationships. It outperforms leading methods on benchmarks, including PVSG, Action Genome, and our newly collected VSGR dataset across five tasks.", + "bbox": [ + 511, + 652, + 906, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. Although our HyperGLM approach effectively models multi-way interactions, managing many objects and their interactions can complicate relationship structures. As the HyperGraph expands, essential relationships may become obscured, reducing the clarity of scene interpretation.", + "bbox": [ + 511, + 761, + 906, + 837 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment. This material is based upon work supported by the National Science Foundation under Award No. OIA-1946391. We also acknowledge the Arkansas High-Performance Computing Center for providing GPUs.", + "bbox": [ + 511, + 839, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "29157", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Song Bai, Feihu Zhang, and Philip HS Torr. Hypergraph convolution and hypergraph attention. Pattern Recognition, 110:107637, 2021. 2", + "[2] Joya Chen, Zhaoyang Lv, Shiwei Wu, Kevin Qinghong Lin, Chenan Song, Difei Gao, Jia-Wei Liu, Ziteng Gao, Dongxing Mao, and Mike Zheng Shou. Videollm-online: Online video large language model for streaming video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18407-18418, 2024. 1", + "[3] Zhanwen Chen, Saed Rezayi, and Sheng Li. More knowledge, less bias: Unbiasing scene graph generation with explicit ontological adjustment. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4023-4032, 2023. 2", + "[4] Yuren Cong, Wentong Liao, Hanno Ackermann, Bodo Rosenhahn, and Michael Ying Yang. Spatial-temporal transformer for dynamic scene graph generation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16372-16382, 2021. 1, 2, 3", + "[5] Yuren Cong, Wentong Liao, H. Ackermann, M. Yang, and B. Rosenhahn. Spatial-temporal transformer for dynamic scene graph generation. IEEE International Conference on Computer Vision, 2021. 7", + "[6] Yuren Cong, Michael Ying Yang, and Bodo Rosenhahn. Reltr: Relation transformer for scene graph generation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(9):11169-11183, 2023. 2", + "[7] Youming Deng, Yansheng Li, Yongjun Zhang, Xiang Xiang, Jian Wang, Jingdong Chen, and Jiayi Ma. Hierarchical memory learning for fine-grained scene graph generation. In European Conference on Computer Vision, pages 266-283. Springer, 2022. 2", + "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 6", + "[9] Azade Farshad, Yousef Yeganeh, Yu Chi, Chengzhi Shen, Böjrn Ommer, and Nassir Navab. Scenegenie: Scene graph guided diffusion models for image synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 88-98, 2023. 2", + "[10] Hao Fei, Shengqiong Wu, Wei Ji, Hanwang Zhang, Meishan Zhang, Mong-Li Lee, and Wynne Hsu. Video-of-thought: Step-by-step video reasoning from perception to cognition. In *Forty-first International Conference on Machine Learning*, 2024. 1", + "[11] Shengyu Feng, Hesham Mostafa, Marcel Nassar, Somdeb Majumdar, and Subarna Tripathi. Exploiting long-term dependencies for generating dynamic scene graphs. IEEE Workshop/Winter Conference on Applications of Computer Vision, 2021. 7", + "[12] Yue Gao, Yifan Feng, Shuyi Ji, and Rongrong Ji. Hgnn+: General hypergraph neural networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):3181-3199," + ], + "bbox": [ + 93, + 114, + 483, + 890 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "2022. 2", + "[13] Zeeshan Hayden and Xuming He. Dsgg: Dense relation transformer for an end-to-end scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28317-28326, 2024. 2", + "[14] Bo He, Hengduo Li, Young Kyun Jang, Menglin Jia, Xuefei Cao, Ashish Shah, Abhinav Shrivastava, and Ser-Nam Lim. Ma-lmm: Memory-augmented large multimodal model for long-term video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13504-13514, 2024. 8", + "[15] Tao He, Lianli Gao, Jingkuan Song, and Yuan-Fang Li. Towards open-vocabulary scene graph generation with prompt-based finetuning. In European Conference on Computer Vision, pages 56-73. Springer, 2022. 2", + "[16] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 6", + "[17] Jinbae Im, Jeong Yeon Nam, Nokyung Park, Hyungmin Lee, and Seunghyun Park. Egtr: Extracting graph from transformer for scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24229-24238, 2024. 2", + "[18] Jingwei Ji, Ranjay Krishna, Li Fei-Fei, and Juan Carlos Niebles. Action genome: Actions as compositions of spatiotemporal scene graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10236–10247, 2020. 1, 2, 5, 6, 7", + "[19] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 6", + "[20] Peng Jin, Ryuichi Takanobu, Wancai Zhang, Xiaochun Cao, and Li Yuan. Chat-univi: Unified visual representation empowers large language models with image and video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13700-13710, 2024. 1, 8", + "[21] Tianlei Jin, Fangtai Guo, Qiwei Meng, Shiqiang Zhu, Xiangming Xi, Wen Wang, Zonghao Mu, and Wei Song. Fast contextual scene graph generation with unbiased context augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6302-6311, 2023. 2", + "[22] Eun-Sol Kim, Woo Young Kang, Kyoung-Woon On, Yu-Jung Heo, and Byoung-Tak Zhang. Hypergraph attention networks for multimodal learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14581-14590, 2020. 2", + "[23] Kibum Kim, Kanghoon Yoon, Jaehyeong Jeon, Yeunjun In, Jinyoung Moon, Donghyun Kim, and Chanyoung Park. Llm4sgg: Large language models for weakly supervised scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28306-28316, 2024. 2", + "[24] Dohwan Ko, Joonmyung Choi, Hyeong Kyu Choi, Kyoung-" + ], + "bbox": [ + 516, + 92, + 906, + 893 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "29158", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Woon On, Byungseok Roh, and Hyunwoo J Kim. Meltr: Meta loss transformer for learning to fine-tune video foundation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20105-20115, 2023. 1, 8", + "[25] Sanjoy Kundu and Sathyanarayanan N Aakur. Is-ggt: Iterative scene graph generation with generative transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6292-6301, 2023. 2", + "[26] Rongjie Li, Songyang Zhang, and Xuming He. Sgtr: End-to-end scene graph generation with transformer. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19486-19496, 2022. 2", + "[27] Rongjie Li, Songyang Zhang, Dahua Lin, Kai Chen, and Xuming He. From pixels to graphs: Open-vocabulary scene graph generation with vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28076-28086, 2024. 2", + "[28] Wanxin Li, Wei Xie, Zhigang Tu, Wei Wang, and Lianghao Jin. Multi-hyperedge hypergraph for group activity recognition. In 2022 International Joint Conference on Neural Networks (IJCNN), pages 01-07. IEEE, 2022. 2", + "[29] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In European Conference on Computer Vision, pages 323–340. Springer, 2025. 8", + "[30] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 1, 8", + "[31] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 6", + "[32] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Khan. Video-ChatGPT: Towards detailed video understanding via large vision and language models. In Lun-Wei Ku, Andre Martins, and Vivek Srikumar, editors, Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 12585–12602, Bangkok, Thailand, Aug. 2024. Association for Computational Linguistics. 1, 8", + "[33] Sayak Nag, Kyle Min, Subarna Tripathi, and Amit K Roy-Chowdhury. Unbiased scene graph generation in videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22803-22813, 2023. 2", + "[34] Trong-Thuan Nguyen, Pha Nguyen, and Khoa Luu. Hig: Hierarchical interlacement graph approach to scene graph generation in video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 1, 2, 3, 5, 7", + "[35] Trong-Thuan Nguyen, Pha Nguyen, Li Xin, Cothren Jackson, Yilmaz Alper, and Khoa Luu. CYCLO: Cyclic graph transformer approach to multi-object relationship modeling in aerial videos. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 1, 2, 5, 7", + "[36] Rohith Peddi, Saksham Singh, Parag Singla, Vibhav Gogate, et al. Towards scene graph anticipation. In European Conference on Computer Vision. Springer, 2024. 1, 2, 3, 7, 8", + "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya" + ], + "bbox": [ + 91, + 90, + 483, + 892 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 6", + "[38] Paul Hongsuck Seo, Arsha Nagrani, Anurag Arnab, and Cordelia Schmid. End-to-end generative pretraining for multimodal video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17959-17968, 2022. 1, 8", + "[39] Xindi Shang, Yicong Li, Junbin Xiao, Wei Ji, and Tat-Seng Chua. Video visual relation detection via iterative inference. In Proceedings of the 29th ACM international conference on Multimedia, pages 3654-3663, 2021. 2", + "[40] Yaojie Shen, Xin Gu, Kai Xu, Heng Fan, Longyin Wen, and Libo Zhang. Accurate and fast compressed video captioning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15558-15567, 2023. 1, 8", + "[41] Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Haozhe Chi, Xun Guo, Tian Ye, Yanting Zhang, et al. Moviechat: From dense token to sparse memory for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18221-18232, 2024. 1, 8", + "[42] Yao Teng, Limin Wang, Zhifeng Li, and Gangshan Wu. Target adaptive context aggregation for video scene graph generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13688-13697, 2021. 1, 2, 3", + "[43] Nupur Thakur, PrasanthSai Gouripeddi, and Baoxin Li. Graph (graph): A nested graph-based framework for early accident anticipation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 7533-7541, 2024. 1, 2", + "[44] Aisha Urooj, Hilde Kuehne, Bo Wu, Kim Chheu, Walid BousSELham, Chuang Gan, Niels Lobo, and Mubarak Shah. Learning situation hyper-graphs for video question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14879-14889, 2023. 2", + "[45] Jiahao Wang, Guo Chen, Yifei Huang, Limin Wang, and Tong Lu. Memory-and-anticipation transformer for online action understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13824–13835, 2023. 1", + "[46] Yanan Wang, Shuichiro Haruta, Donghuo Zeng, Julio Vizcarra, and Mori Kurokawa. Multi-object event graph representation learning for video question answering. In *Meeting on Image Recognition and Understanding*, 2024. 2", + "[47] Tao Wu, Runyu He, Gangshan Wu, and Limin Wang. Sportshhi: A dataset for human-human interaction detection in sports videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2, 5", + "[48] Dejing Xu, Zhou Zhao, Jun Xiao, Fei Wu, Hanwang Zhang, Xiangnan He, and Yueting Zhuang. Video question answering via gradually refined attention over appearance and motion. In Proceedings of the 25th ACM international conference on Multimedia, pages 1645-1653, 2017. 6", + "[49] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In Proceedings of the IEEE conference on computer vision and" + ], + "bbox": [ + 516, + 90, + 906, + 892 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "29159", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "ref_text", + "text": "pattern recognition, pages 5288-5296, 2016. 6", + "bbox": [ + 124, + 92, + 403, + 104 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[50] Zenan Xu, Wanjun Zhong, Qinliang Su, Zijing Ou, and Fuwei Zhang. Modeling semantic composition with syntactic hypergraph for video question answering. arXiv preprint arXiv:2205.06530, 2022. 2", + "[51] Jingkang Yang, Yi Zhe Ang, Zujin Guo, Kaiyang Zhou, Wayne Zhang, and Ziwei Liu. Panoptic scene graph generation. In European Conference on Computer Vision, pages 178-196. Springer, 2022. 2", + "[52] Jingkang Yang, Wenxuan Peng, Xiangtai Li, Zujin Guo, Liangyu Chen, Bo Li, Zheng Ma, Kaiyang Zhou, Wayne Zhang, Chen Change Loy, et al. Panoptic video scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18675-18685, 2023. 1, 2, 5, 6, 7", + "[53] Guangyao Zhai, Evin Pinar Örnek, Shun-Cheng Wu, Yan Di, Federico Tombari, Nassir Navab, and Benjamin Busam. Commonscenes: Generating commonsense 3d indoor scenes with scene graphs. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[54] Zizhao Zhang, Yifan Feng, Shihui Ying, and Yue Gao. Deep hypergraph structure learning. arXiv preprint arXiv:2208.12547, 2022. 2", + "[55] Xiaolin Zhu, Dongli Wang, Jianxun Li, Rui Su, Qin Wan, and Yan Zhou. Dynamical attention hypergraph convolutional network for group activity recognition. IEEE Transactions on Neural Networks and Learning Systems, 2024. 2" + ], + "bbox": [ + 91, + 106, + 483, + 458 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "29160", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/HyperGLM_ HyperGraph for Video Scene Graph Generation and Anticipation/07619f38-6a65-46d1-ae2a-3885ad0e2f6d_model.json b/2025/HyperGLM_ HyperGraph for Video Scene Graph Generation and Anticipation/07619f38-6a65-46d1-ae2a-3885ad0e2f6d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..bf1eb0a96efd551a6fbc0311aa3b3e46ee059091 --- /dev/null +++ b/2025/HyperGLM_ HyperGraph for Video Scene Graph Generation and Anticipation/07619f38-6a65-46d1-ae2a-3885ad0e2f6d_model.json @@ -0,0 +1,2532 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.1, + 0.131, + 0.9, + 0.156 + ], + "angle": 0, + "content": "HyperGLM: HyperGraph for Video Scene Graph Generation and Anticipation" + }, + { + "type": "text", + "bbox": [ + 0.16, + 0.18, + 0.837, + 0.217 + ], + "angle": 0, + "content": "Trong-Thuan Nguyen\\(^{1\\dagger}\\), Pha Nguyen\\(^{1\\dagger}\\), Jackson Cothren\\(^{1}\\), Alper Yilmaz\\(^{2}\\), Khoa Luu\\(^{1}\\) \n\\(^{1}\\)University of Arkansas \\(^{2}\\)Ohio State University" + }, + { + "type": "text", + "bbox": [ + 0.205, + 0.218, + 0.794, + 0.253 + ], + "angle": 0, + "content": "\\(^{1}\\{\\)thuann, panguyen, jcothre, khoaluu\\}@uark.edu \\(^{2}\\)yilmaz.15@osu.edu uark-cviu.github.io/projects/HyperGLM" + }, + { + "type": "image", + "bbox": [ + 0.141, + 0.3, + 0.855, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.251, + 0.364, + 0.497, + 0.377 + ], + "angle": 0, + "content": "What is the relationship between objects?" + }, + { + "type": "image", + "bbox": [ + 0.144, + 0.38, + 0.419, + 0.473 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.145, + 0.474, + 0.498, + 0.513 + ], + "angle": 0, + "content": "Person-1 holding the Umbrella-1 and standing beside Person-2. Person-1 and Person-2 are waiting for Person-3 in her yellow dress. Then, Person -1 talking with Person-3." + }, + { + "type": "image_caption", + "bbox": [ + 0.649, + 0.362, + 0.849, + 0.375 + ], + "angle": 0, + "content": "What are the roles of the people?" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.382, + 0.838, + 0.421 + ], + "angle": 0, + "content": "It appears that Person-1 is picking up his family, as Person-3 is being dropped off from the helicopter in the rainy weather. Person-2 is likely serving as security." + }, + { + "type": "text", + "bbox": [ + 0.629, + 0.424, + 0.846, + 0.438 + ], + "angle": 0, + "content": "What might the next interaction be?" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.441, + 0.817, + 0.467 + ], + "angle": 0, + "content": "Person-1 and Person-3 are talking while waiting for another family member to get off the helicopter." + }, + { + "type": "text", + "bbox": [ + 0.609, + 0.469, + 0.848, + 0.482 + ], + "angle": 0, + "content": "What might they do after moving away?" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.487, + 0.822, + 0.513 + ], + "angle": 0, + "content": "After arriving by helicopter, Person-1, Person-3, and another woman will likely move indoors to rest." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.518, + 0.908, + 0.545 + ], + "angle": 0, + "content": "Figure 1. Our HyperGLM framework supports Video Scene Graph Generation, Anticipation, and Reasoning. HyperGLM constructs scene graphs from observed video frames and predicts relationships in unseen frames by leveraging a unified hypergraph for temporal modeling and comprehensive understanding." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.556, + 0.327, + 0.571 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.58, + 0.486, + 0.882 + ], + "angle": 0, + "content": "Multimodal LLMs have advanced vision-language tasks but still struggle with understanding video scenes. To bridge this gap, Video Scene Graph Generation (VidSGG) has emerged to capture multi-object relationships across video frames. However, prior methods rely on pairwise connections, limiting their ability to handle complex multi-object interactions and reasoning. To this end, we propose the Multimodal Large Language Models (LLMs) on a Scene HyperGraph (HyperGLM), promoting reasoning about multi-way interactions and higher-order relationships. Our approach uniquely integrates entity scene graphs, which capture spatial relationships between objects, with a procedural graph that models their causal transitions, forming a unified HyperGraph. Significantly, HyperGLM enables reasoning by injecting this unified HyperGraph into LLMs. Additionally, we introduce a new Video Scene Graph Reasoning (VSGR) dataset featuring 1.9M frames from third-person, egocentric, and drone views and support five tasks. Empirically, HyperGLM consistently outperforms state-of-the-art methods, effectively modeling and reasoning complex relationships in diverse scenes." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.556, + 0.645, + 0.571 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.581, + 0.909, + 0.777 + ], + "angle": 0, + "content": "In recent years, Multimodal Large Language Models (LLMs) [2, 10] have set new benchmarks, with vision-language models excelling in diverse multimodal tasks. However, fully understanding dynamic video scenes remains a significant challenge for applications like autonomous driving, intelligent surveillance, human-object interaction, and multimedia analysis. Towards this goal, Video Scene Graph Generation (VidSGG) [18, 52] has emerged as a critical task for capturing multi-object relationships across video frames. In particular, VidSGG enables high-level tasks such as event forecasting [36, 43, 45], video captioning [24, 38, 40], and video question answering [20, 30, 32, 41] by constructing detailed representations of entities and their interactions." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.78, + 0.911, + 0.902 + ], + "angle": 0, + "content": "However, prior VidSGG methods and datasets have limited capacity for comprehensive video understanding. Traditional scene graph-based methods [4, 34, 35, 42] only model pairwise object relationships within single frames, making it challenging to capture higher-order relationships and temporal dependencies in real-world scenarios. Additionally, existing benchmark datasets [18, 34, 35, 52] focus is primarily confined to Scene Graph Generation (SGG) and Scene" + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.225, + 0.901 + ], + "angle": 0, + "content": "\\(\\dagger\\) Equal contribution." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29150" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.137 + ], + "angle": 0, + "content": "Graph Anticipation (SGA) tasks, lacking annotations for reasoning tasks such as Video Question Answering (VQA), Video Captioning (VC), and Relation Reasoning (RR)." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.139, + 0.487, + 0.455 + ], + "angle": 0, + "content": "In this paper, we propose the Multimodal LLMs on a Scene HyperGraph (HyperGLM) approach to promote reasoning about multi-way interactions and higher-order relationships through a unified HyperGraph and LLMs, as illustrated in Fig. 1. To achieve this goal, we uniquely incorporate entity scene graphs, which capture spatial relationships between objects, with a procedural graph that models their causal transitions across video frames, forming a unified HyperGraph, as shown in Fig. 2. Significantly, our HyperGLM approach allows reasoning by injecting this unified HyperGraph into LLMs. In addition, we introduce a novel Video Scene Graph Reasoning (VSGR) dataset, comprising 1.9 million video frames surpassing existing benchmark datasets [18, 34, 35, 47, 52] in scale and annotation depth. Specifically, our VSGR dataset includes videos from third-person, egocentric, and drone perspectives. It supports five tasks: Scene Graph Generation, Scene Graph Anticipation, Video Question Answering, Video Captioning, and Relation Reasoning. Notably, our VSGR dataset introduces a new Relation Reasoning task, setting it apart from existing video scene graph datasets, as shown in the comparison in Table 1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.455, + 0.485, + 0.606 + ], + "angle": 0, + "content": "Contributions of this Work. This work presents three contributions to advance Video Scene Graph Generation. First, we introduce Multimodal LLMs on a Scene HyperGraph, leveraging hyperedges and LLMs for reasoning about multiway interactions and higher-order relationships. Second, we develop a new Video Scene Graph Reasoning dataset, surpassing existing scale and annotation depth benchmarks. Our VSGR dataset primarily supports five tasks within diverse video scenes. Finally, the proposed HyperGLM consistently outperforms state-of-the-art methods across all five tasks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.618, + 0.233, + 0.634 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.483, + 0.689 + ], + "angle": 0, + "content": "In this section, we review advances in scene graph generation and hypergraph applications in computer vision, then discuss existing limitations and the advantages of our approach." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.697, + 0.318, + 0.712 + ], + "angle": 0, + "content": "2.1. Scene Graph Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.486, + 0.901 + ], + "angle": 0, + "content": "Scene Graph Generation has significantly advanced with transformer-based models [6, 13, 17, 25, 26, 51] that have become benchmarks due to their efficiency and state-of-the-art performance. Recent work [3, 21, 33] focuses on reducing bias and enhancing mean recall for rare predicates by integrating external knowledge and applying unbiased contextual augmentation, particularly in dynamic video contexts. Open-vocabulary methods [15, 27] supported by vision-language models handle unseen object and relationship classes, improving generalization. Additionally, generative models (e.g., diffusion-based methods [9, 53]) leverage scene graphs for efficient image and scene synthesis. More-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.152 + ], + "angle": 0, + "content": "over, spatial-temporal methods [4, 34-36, 42] effectively capture dynamic object relationships in videos. Recently, Large Language Models [23] have been utilized to enhance triplet extraction and alignment in weakly supervised SGG." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.163, + 0.659, + 0.179 + ], + "angle": 0, + "content": "2.2. HyperGraphs" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.187, + 0.909, + 0.399 + ], + "angle": 0, + "content": "HyperGraphs have been adopted in computer vision to model complex multimodal data and capture higher-order relationships. Unlike traditional graph-based approaches, HyperGraphs connect multiple nodes through hyperedges, enabling multi-way relationships. They enhance Graph Neural Networks (GNNs) [12, 54] by allowing the modeling of more sophisticated interactions. Recent advancements, such as HyperGraph Convolution [1] and HyperGraph Attention [22], have further improved GNNs by capturing relationships beyond simple pairwise connections. Therefore, HyperGraph-based models effectively handle temporal dependencies and complex interactions, significantly boosting performance in tasks like accident anticipation [43], group activity recognition [28, 55], and video question answering [44, 46, 50]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.41, + 0.632, + 0.423 + ], + "angle": 0, + "content": "2.3. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.433, + 0.909, + 0.675 + ], + "angle": 0, + "content": "Limitations in Prior Methods. The methods introduced in Section 2.1, based on Progressive Feature Fusion [39, 51], Batch-Progressive Transformer Encoding [6, 13, 17, 25, 26], Spatial-Temporal Context Integration [4, 34, 42], and Memory-Guided Temporal Consistency [7, 33, 35], have advanced VidSGG. However, these methods struggle to model higher-order relationships and complex temporal dynamics. Specifically, Progressive Feature Fusion and Batch-Progressive Transformer Encoding are limited in capturing long-term temporal dependencies, with the former lacking long-term context due to frame-by-frame processing and the latter only addressing short-term dependencies. Similarly, Spatial-Temporal Context Integration and Memory-Guided Temporal Consistency inadequately represent multi-object interactions across video frames and insufficiently capture the temporal evolution required for higher-order relationships." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.911, + 0.902 + ], + "angle": 0, + "content": "Advantages of Our Approach. Our HyperGLM approach, Multimodal LLMs on a Scene HyperGraph, promotes reasoning about multi-way interactions and high-order relationships. As illustrated in Fig. 2, HyperGLM enhances the model's ability to interpret complex relationships and anticipate intricate video dynamics. Towards this goal, we uniquely integrate entity scene graphs, which is introduced in Sec. 3 to capture spatial interactions between objects in each frame and a procedural graph, which is presented in Sec. 4.1 to model their causal evolution. In our unified approach, hyperedges connect multiple nodes to capture higher-order relationships distinguished from traditional pairwise methods [4, 34, 35, 42]. In addition, the procedural graph enables multi-step transitions for anticipating future interactions or relationships, and reduces bias by generalizing" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "29151" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.087, + 0.5, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.181, + 0.484, + 0.266 + ], + "angle": 0, + "content": "Figure 2. (a) To model the temporal transition, a simple approach can be using two scene graphs \\(\\mathbf{G}_{\\mathbf{t}}\\) and \\(\\mathbf{G}_{\\mathbf{t} + 1}\\). (b) Another procedure graph can present this temporal modeling. (c) Our unified HyperGraph in Fig. 2c integrates both entity scene graph to capture spatial relationships and the procedural graph to model the temporal evolution. HyperEdge represents person sitting on couch, holding, then playing guitar, whereas holding \\(\\rightarrow\\) playing describes a chain of interactions. HyperGraph is presented in 3D." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.273, + 0.484, + 0.395 + ], + "angle": 0, + "content": "infrequent relationship categories. Furthermore, our HyperGLM approach leverages fundamental mathematical properties: permutation equivariance ensures that the HyperGraph structure remains consistent under any permutation of node labels, and invariance to hyperedge order preserves semantic meanings regardless of the node visit order during random walks which is defined in Alg. 1. The theoretical foundations and mathematical properties are detailed in the Appendices." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.421, + 0.296, + 0.436 + ], + "angle": 0, + "content": "3. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.447, + 0.484, + 0.568 + ], + "angle": 0, + "content": "In this section, we define two tasks, including Scene Graph Generation (SGG) and Scene Graph Anticipation (SGA). Graph vertex sequence is represented as \\(\\{V_{\\mathbf{G}}\\mid 1\\leq t\\leq T\\}\\), where each set of vertex \\(V_{\\mathbf{G}}\\) contains object features, bounding boxes, and categories. The scene graph for each frame \\(t\\), denoted by \\(\\mathbf{G}_t = \\big(V_{\\mathbf{G}},E_{\\mathbf{G}} = \\{r^{ij}\\mid 1\\leq i < j\\leq |V_{\\mathbf{G}}|\\}\\), consists of all pairwise relationships between objects, with \\(r^{ij}\\) representing the relationship category between \\(v_{i}\\) and \\(v_{j}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.567, + 0.484, + 0.703 + ], + "angle": 0, + "content": "We aim to develop a process \\(p_{\\theta}: (V_{\\mathbf{G}} \\times V_{\\mathbf{G}}) \\to r^{ij}\\) to predict the relationship \\(r^{ij}\\) between each object pair \\((v_t^i, v_t^j)\\) in \\(V_{\\mathbf{G}}\\). We define task-specific queries \\(\\mathbf{Q}_{\\mathrm{SGG}}\\), and \\(\\mathbf{Q}_{\\mathrm{SGA}}\\) to direct our unified model to perform on each specific task. The objective for each task is to minimize the negative likelihood of the predicted scene graph \\(\\mathbf{G}_t\\) to the truth predicate set \\(Y_t = \\{y^{ij} \\mid 1 \\leq i < j \\leq |V_{\\mathbf{G}}|\\}\\) on categories indexed by \\(k\\). Scene Graph Generation (SGG) generates the scene graph for each frame \\(t\\) from \\(t = 1\\) to \\(t = T\\), which is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.125, + 0.708, + 0.484, + 0.743 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} \\mathbb {E} _ {\\mathbf {G} _ {t}, Y _ {t}} \\left[ - \\sum_ {\\left(v _ {t} ^ {i, j}\\right)} \\sum_ {k} \\left(y _ {k} ^ {i j} \\log p _ {\\theta} \\left(r ^ {i j} \\mid \\mathbf {Q} _ {\\mathrm {S G G}}\\right) _ {k}\\right) \\right] \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.795 + ], + "angle": 0, + "content": "Scene Graph Anticipation (SGA) anticipates the scene graphs for future frames, generating predictions \\(\\mathbf{G}_{t + n}\\), where \\(n\\) denotes the anticipation horizon, formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.812, + 0.484, + 0.849 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} \\mathbb {E} _ {\\mathbf {G} _ {\\leq t}, Y _ {t}} \\left[ - \\sum_ {\\left(v _ {t + n} ^ {i, j}\\right)} \\sum_ {k} \\left(y _ {k} ^ {i j} \\log p _ {\\theta} \\left(r ^ {i j} \\mid \\mathbf {Q} _ {\\mathrm {S G A}}\\right) _ {k}\\right) \\right] \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In Eqns. (1) and (2), \\( t \\) indexes the current frame. While SGG predicts the scene graph at frame \\( t \\), SGA reads scene graphs up to frame \\( t \\) to forecast the graph in future frame \\( t + n \\)." + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.088, + 0.925, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.192, + 0.907, + 0.289 + ], + "angle": 0, + "content": "Figure 3. Our HyperGLM framework comprises an image encoder, MLP projector, temporal aggregator, unified HyperGraph, and language model. It processes video frames by encoding each frame with the image encoder and MLP, extracting spatio-temporal features through image patch grids to generate \\(N\\) spatial tokens per frame. The temporal aggregator compresses the \\(T \\times N\\) embeddings over time. The MLP projector then transforms these visual embeddings into the language feature space as frame tokens, interleaved with language tokens, and fed into the Large Language Models." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.299, + 0.849, + 0.316 + ], + "angle": 0, + "content": "4. Our Proposed HyperGLM Approach" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.325, + 0.907, + 0.355 + ], + "angle": 0, + "content": "In this section, we present our approach, which incorporates a unified HyperGraph into the LLMs, as illustrated in Fig. 3." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.364, + 0.766, + 0.381 + ], + "angle": 0, + "content": "4.1. Video Scene (Hyper)Graphs" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.388, + 0.909, + 0.735 + ], + "angle": 0, + "content": "Traditional VidSGG methods [4, 34, 42] construct scene graphs \\(\\mathbf{G}_t\\) to represent scene entities and their relationships as in Fig. 2a. However, these graph-based approaches do not capture higher-order relationships with temporal dependencies for video understanding [36]. To model this property, we propose a novel HyperGraph-based framework that constructs a unified HyperGraph \\(\\mathcal{H} = (V_{\\mathcal{H}}, E_{\\mathcal{H}})\\), representing spatial relationships within individual frames and temporal transitions across frames, illustrated in Fig. 4. HyperGraph extends traditional graph structures by allowing hyperedges to connect more than two nodes, making them particularly effective for modeling these higher-order relationships. To leverage this property, our unified HyperGraph (see Fig. 2c) integrates entity scene graphs \\(\\mathbf{G}_t\\) for each frame, capturing spatial interactions as introduced in Sec. 3, along with a procedural graph \\(\\mathbf{P}\\) that models their causal transitions, as detailed in Sec. 4.1. We unify these components using a random-walk HyperGraph construction, which is defined in Alg. 1 to capture higher-order connectivity patterns and structural semantics, thereby approximating subgraph matching between the entity scene graphs and the procedural graph. This integration allows our approach to model current subjects' interactions and anticipate their future relationships." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.735, + 0.909, + 0.856 + ], + "angle": 0, + "content": "Procedural Graph Construction. SGA objective is to predict the set of relationships \\( E_{\\mathbf{G}} \\) in the next frame based on the current set of relationships. To model the temporal evolution of causal relationships between objects across video frames, we introduce a procedural graph \\( \\mathbf{P} = (V_{\\mathbf{P}},E_{\\mathbf{P}}) \\), serving as the temporal counterpart to the entity scene graphs \\( \\mathbf{G}_t \\) as shown in Fig. 2b. In particular, the procedural graph \\( \\mathbf{P} \\) can vary across datasets, modeling relationship transitions." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Toward this goal, we model and denote the set of vertices in \\(\\mathbf{P}\\) as \\(V_{\\mathbf{P}} = \\{r_t^{ij} \\in V_{\\mathbf{G}} \\mid 1 \\leq t \\leq T\\} \\neq\\) representing distinct relationship categories. The set of edges in \\(\\mathbf{P}\\), i.e.," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29152" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.904, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.291, + 0.908, + 0.342 + ], + "angle": 0, + "content": "Figure 4. Our Video Scene HyperGraph, including entity graphs and a procedural graph, as defined in Sec. 4.1. Blue nodes represent entities, while green nodes denote relationships. The entity graph captures spatial relationships (subject \\(\\rightarrow\\) relationship \\(\\rightarrow\\) object), whereas the procedural graph models relationship transitions \\((\\rightarrow)\\). Hyperedges are visualized as polygons, encapsulating interactions through chains of relationships. For instance, a hyperedge illustrates a person picking up, holding, opening, and reading a book while sitting on a couch. HyperGraph is presented in 3D, see Supplementary video." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.351, + 0.487, + 0.429 + ], + "angle": 0, + "content": "\\(E_{\\mathbf{P}} = \\{(r_m,r_n)\\}\\) represent possible causal transitions between these relationships, where an edge \\((r_m,r_n)\\) indicates that relationship \\(r_m\\) causally lead to relationship \\(r_n\\). We quantify causal transitions by calculating transition probabilities \\(w(r_m,r_n)\\) via their observed frequencies as in Eqn. (3)." + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.45, + 0.484, + 0.532 + ], + "angle": 0, + "content": "\\[\nw \\left(r _ {m}, r _ {n}\\right) = \\frac {\\sum_ {t = 1} ^ {T - 1} \\sum_ {r _ {t , t + 1} ^ {i j}} \\mathbb {1} \\left(r _ {t} ^ {i j} = r _ {m} \\wedge r _ {t + 1} ^ {i j} = r _ {n}\\right)}{\\sum_ {t = 1} ^ {T - 1} \\sum_ {r _ {t , t + 1} ^ {i j}} \\mathbb {1} \\left(r _ {t} ^ {i j} = r _ {m}\\right)} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.542, + 0.484, + 0.605 + ], + "angle": 0, + "content": "where \\(\\mathbb{1}(\\cdot)\\) is the indicator function that counts transitions from relationship \\(r_m\\) at current frame \\(t\\) to \\(r_n\\) at next frame \\(t + 1\\). Next, self-loops \\(w(r_m, r_m)\\) are removed from the graph, and these probabilities are normalized as in Eqn. (4)." + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.614, + 0.484, + 0.647 + ], + "angle": 0, + "content": "\\[\n\\sum_ {r _ {n} \\in E _ {\\mathbf {P}}} w \\left(r _ {m}, r _ {n}\\right) = 1 \\quad \\text {f o r a l l} r _ {m} \\in E _ {\\mathbf {P}} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.656, + 0.484, + 0.718 + ], + "angle": 0, + "content": "By leveraging these probabilities, the procedural graph \\(\\mathbf{P}\\) enhances the prediction of future relationships. For each relationship \\(r_t^{ij}\\) in frame \\(t\\), the most probable relationship \\(r_{t + 1}^{ij}\\) in the next frame \\(t + 1\\) is determined as in Eqn. (5)." + }, + { + "type": "equation", + "bbox": [ + 0.175, + 0.728, + 0.484, + 0.752 + ], + "angle": 0, + "content": "\\[\nr _ {t + 1} ^ {i j} = \\arg \\max _ {r _ {n}} P \\left(r _ {n} \\mid r _ {t} ^ {i j}, v ^ {i, j}\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.764, + 0.484, + 0.826 + ], + "angle": 0, + "content": "where \\( P(r_{n} \\mid r_{t}^{ij}, v^{i,j}) = w(r_{t}^{ij}, r_{n}) \\times v^{i,j} \\) is the probability of transitioning from relationship \\( r_{t}^{ij} \\) to \\( r_{n} \\), looking at object features. This allows the model to anticipate future interactions based on established temporal patterns." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "HyperGraph Construction. To incorporate spatial relationships between objects in each frame and their causal temporal transitions, we construct a unified HyperGraph \\(\\mathcal{H}\\) that integrates the entity scene graphs \\(\\mathbf{G}_t\\) and the procedural graph \\(\\mathbf{P}\\). Specifically, a HyperGraph is an augmented" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.351, + 0.908, + 0.442 + ], + "angle": 0, + "content": "representation combining the entity scene graphs with the procedural graph. This unified structure merges spatial and temporal relationships into a single graph, enabling the use of conventional graph algorithms while preserving the complex interactions captured by the HyperGraph. Mathematically, a unified HyperGraph \\(\\mathcal{H} = (V_{\\mathcal{H}}, E_{\\mathcal{H}})\\) is defined as in Eqn. (6)." + }, + { + "type": "equation", + "bbox": [ + 0.572, + 0.469, + 0.907, + 0.512 + ], + "angle": 0, + "content": "\\[\n\\mathcal {H} = \\left(\\bigcup_ {t = 1} ^ {T} V _ {\\mathbf {G} _ {t}} \\cup V _ {\\mathbf {P}}, \\quad \\bigcup_ {t = 1} ^ {T} E _ {\\mathbf {G} _ {t}} \\cup E _ {\\mathbf {P}}\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.533, + 0.908, + 0.61 + ], + "angle": 0, + "content": "where \\(V_{\\mathcal{H}}\\) includes all entity nodes \\(v_{t}^{i}\\) from each \\(\\mathbf{G}_t\\) and the relationship type nodes \\(V_{\\mathbf{P}}\\). The hyperedge set \\(E_{\\mathcal{H}}\\) includes pairwise relationships \\(E_{\\mathbf{G}_t}\\) within each \\(\\mathbf{G}_t\\), capturing spatial relationships and temporal transition edges \\(E_{\\mathbf{P}}\\) from \\(\\mathbf{P}\\), modeling the evolution of causal relationships across frames." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Random-walk Algorithm. We employ the random walks algorithm outlined in Alg. 1 to sample representative substructures from the unified HyperGraph \\(\\mathcal{H}\\), which captures connectivity patterns and mitigates the NP-hardness of exact subgraph matching. Specifically, these walks alternate between nodes and hyperedges, preserving the multi-node connections intrinsic to hyperedges and capturing the complexity of multi-object relationships and their transition across video frames. In each walk, a hyperedge \\(h_i\\) aggregates the visited nodes, thereby encapsulating higher-order relationships. For example, in the entity scene graph \\(\\mathbf{G}_t\\), a walk might traverse from a \"person\" to a \"cup\" via the \"holding\", resulting in the hyperedge \\(h_i = \\{\\text{person}, \\text{cup}\\}\\). Similarly, within the procedural graph \\(\\mathbf{P}\\), a walk might transition through a sequence of interactions such as \"holding\", \"placing\", and \"releasing\", forming the hyperedge \\(h_i = \\{\\text{holding}, \\text{placing}, \\text{releasing}\\}\\). Therefore, we generate sampled hyperedges \\(E_{\\text{sampled}} = \\{h_i \\mid i = 1, \\dots, N_w\\}\\) as in L25 of Alg. 1 by conducting multiple random walks." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29153" + } + ], + [ + { + "type": "code_caption", + "bbox": [ + 0.092, + 0.09, + 0.48, + 0.108 + ], + "angle": 0, + "content": "Algorithm 1 Random-walk for HyperGraph Construction." + }, + { + "type": "algorithm", + "bbox": [ + 0.09, + 0.11, + 0.487, + 0.555 + ], + "angle": 0, + "content": "Require: \\(\\mathcal{H} = (V_{\\mathcal{H}},E_{\\mathcal{H}})\\) , Number of Walks \\(N_{w}\\) , Walk Length \\(N_{l}\\) \nEnsure: \\(\\mathcal{H}' = (V_{\\mathcal{H}},E_{\\mathcal{H}}^{\\prime})\\) \n1: Initialize \\(E_{\\mathrm{sampled}}\\gets \\emptyset\\) \n2: for \\(i = 1\\) to \\(N_{w}\\) do \n3: Select \\(v_{\\mathrm{start}}\\in V_{\\mathcal{H}}\\) uniformly at random \n4: Initialize walk sequence \\(S\\leftarrow [v_{\\mathrm{start}}]\\) \n5: for \\(j = 1\\) to \\(N_{l}\\) do \n6: if \\(j\\) is odd then Node to HyperEdge \n7: \\(v_{\\mathrm{current}}\\gets S[j]\\) \n8: \\(\\mathcal{E}_{v_{\\mathrm{current}}}\\gets \\{h\\in E_{\\mathcal{H}}\\mid v_{\\mathrm{current}}\\in h\\}\\) \n9: if \\(\\mathcal{E}_{v_{\\mathrm{current}}}\\neq \\emptyset\\) then \n10: Select \\(h_j\\in \\mathcal{E}_{v_{\\mathrm{current}}}\\) \n11: Append \\(h_j\\) to \\(S\\) \n12: end if \n13: else HyperEdge to Node \n14: \\(h_\\mathrm{current}\\gets S[j]\\) \n15: \\(V_{h_{\\mathrm{current}}}\\gets h_{\\mathrm{current}}\\) \n16: Select \\(v_{j}\\in V_{h_{\\mathrm{current}}}\\) \n17: Append \\(v_{j}\\) to \\(S\\) \n18: end if \n19: end for \n20: \\(h_i\\gets \\{v\\in S\\mid v\\in V_H\\}\\) ▷Form new HyperEdge \n21: if \\(h_i\\notin E_{\\mathcal{H}}\\cup E_{\\mathrm{sampled}}\\) then \n22: \\(E_{\\mathrm{sampled}}\\gets E_{\\mathrm{sampled}}\\cup \\{h_i\\}\\) \n23: end if \n24: end for \n25: \\(E_{\\mathcal{H}}'\\gets E_{\\mathcal{H}}\\cup E_{\\mathrm{sampled}}\\) \n26: return \\(\\mathcal{H}' = (V_{\\mathcal{H}},E_{\\mathcal{H}}')\\)" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.563, + 0.406, + 0.58 + ], + "angle": 0, + "content": "4.2. Multimodal LLMs on HyperGraphs" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.586, + 0.486, + 0.649 + ], + "angle": 0, + "content": "Formulation. Given an input video \\(\\mathbf{V}\\) and a task query \\(\\mathbf{Q}\\), we aim to generate a target answer sequence \\(\\mathbf{A}\\) of length \\(L\\). Especially, by modeling a HyperGraph \\(\\mathcal{H}\\), the target answer \\(\\mathbf{A}\\) is generated via a process illustrated in Fig. 3, defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.656, + 0.484, + 0.697 + ], + "angle": 0, + "content": "\\[\np (\\mathbf {A} | \\mathbf {V}, \\mathbf {Q}, \\mathcal {H}) = \\prod_ {i = 0} ^ {L - 1} p \\left(x _ {i} | \\mathbf {V}, \\mathbf {Q}, \\mathcal {H}, x _ {< i}\\right) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.485, + 0.75 + ], + "angle": 0, + "content": "where \\(x_{< i}\\) denotes the preceding token sequence, and \\(\\mathbf{A} = [x_0,\\dots ,x_i,\\dots ,x_{L - 1}]\\) is the sequence of answer tokens reasoning video \\(\\mathbf{V}\\) by question \\(\\mathbf{Q}\\) and HyperGraph \\(\\mathcal{H}\\) in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.751, + 0.486, + 0.902 + ], + "angle": 0, + "content": "As illustrated in Fig. 5, our process begins with the generation, constructing an entity scene graph \\(\\mathbf{G}_t = (V_t^e, E_{\\mathbf{G}})\\) for each frame \\(t\\) to capture detected objects and their relationships. Relationship anticipation employs the procedural graph \\(\\mathbf{P}\\) to model temporal evolution and predict future interactions. The model then engages in reasoning using the HyperGraph \\(\\mathcal{H}\\) and verification through video captioning to ensure contextual relevance and accuracy. It refines understanding through clarification, generating answers by reasoning over the video and \\(\\mathcal{H}\\). In scene forecasting, it predicts" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.089, + 0.907, + 0.127 + ], + "angle": 0, + "content": "Table 1. Comparisons of video scene graph datasets. SGG, SGA, VQA, VC, and RR represent Scene Graph Generation, Scene Graph Anticipation, Video Question Answering, Video Captioning, and Relation Reasoning." + }, + { + "type": "table", + "bbox": [ + 0.513, + 0.128, + 0.911, + 0.212 + ], + "angle": 0, + "content": "
Dataset#FramesTasksAnnotations
SGGSGAVQAVCRRBboxRelationText
SportsHHI [47]11.4K✓XXXX✓✓X
Action Genome [18]234.3K✓✓XXX✓✓X
AeroEye [35]261.5K✓XXXX✓✓X
ASPIRe [34]1.6M✓XXXX✓✓X
PVSG [52]153K✓X✓✓X✓✓✓
VSGR (Ours)1.9M✓✓✓✓✓✓✓✓
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.224, + 0.907, + 0.33 + ], + "angle": 0, + "content": "subsequent interactions by identifying new relationships and adapting to evolving contexts using transition probabilities in \\(\\mathbf{P}\\), anticipating future relationships \\(r_{t+n}^{ij}\\) and improving conversational flow. Finally, the hypothesis allows the model to propose and test conjectures about the underlying themes or intentions, synthesizing information from previous steps and leveraging \\(\\mathcal{H}\\) to explain patterns and forecast results." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.344, + 0.86, + 0.361 + ], + "angle": 0, + "content": "5. Video Scene Graph Reasoning Dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.369, + 0.909, + 0.492 + ], + "angle": 0, + "content": "Limitations of Current Datasets. Existing benchmarks (see Table 1) primarily support SGG and SGA, limiting their applicability to reasoning tasks. They suffer from inadequate support for VQA, VC, and RR, shallow annotations that fail to capture intricate object interactions, insufficient modeling of temporal dynamics, and poor multimodal integration. Thus, our VSGR dataset supports SGG, SGA, VQA, VC, and RR, enabling the reasoning capabilities of LLMs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.5, + 0.715, + 0.515 + ], + "angle": 0, + "content": "5.1. Dataset Construction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.523, + 0.909, + 0.69 + ], + "angle": 0, + "content": "Data Acquisition Stage. We source videos from the ASPIRe [34] and AeroEye [35] datasets. While the ASPIRe dataset offers diverse, richly annotated videos emphasizing dynamic interactions and temporal changes, the AeroEye dataset offers drone-captured footage across various scenes. Comprehension Tasks via Question-Answering. We introduce tasks that leverage fine-grained relationships from scene graphs, extending Scene Graph Generation to focus on relation understanding and subject/object interpretation using triplets and leverage GPT-4/GPT-3.5 model for language generation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.909, + 0.81 + ], + "angle": 0, + "content": "Video Captioning (VC). We generate 82,532 video-caption pairs, resulting in about 22 captions per video. Our process involves (1) extracting triplets from cropped video frames focusing on foreground objects, (2) generating background descriptions based on these triplets, and (3) combining the foreground triplets and background descriptions to produce captions. The average length is 893 characters, surpassing the PVSG [52] dataset in quantity and detail." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Video Question Answering (VQA). We develop 74,856 question-answer pairs by designing questions that explore diverse relationships, selecting subject and object categories to verify specific relationships, or choosing triplets to assess their uniqueness. This results in an average of approximately 20 questions per video, which exceeds existing benchmarks" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29154" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.098, + 0.487, + 0.172 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.107, + 0.175, + 0.223, + 0.186 + ], + "angle": 0, + "content": "Step-1: Generation" + }, + { + "type": "text", + "bbox": [ + 0.108, + 0.188, + 0.476, + 0.219 + ], + "angle": 0, + "content": "Given bounding boxes: person-1 [657, 116, 155, 577], person-2 [425, 211, 178, 450], ball-1 [534, 437, 42, 54], ball-2 [670, 350, 37, 65]. What are the relationships between these objects?" + }, + { + "type": "text", + "bbox": [ + 0.108, + 0.224, + 0.476, + 0.247 + ], + "angle": 0, + "content": "The video depicts including person-1 towards the right side and closer to ball-2. person-2 is situated to the left and is nearer to ball-1." + }, + { + "type": "title", + "bbox": [ + 0.107, + 0.263, + 0.307, + 0.274 + ], + "angle": 0, + "content": "Step-2: Relationship Anticipation" + }, + { + "type": "text", + "bbox": [ + 0.108, + 0.282, + 0.449, + 0.305 + ], + "angle": 0, + "content": "What interactions might occur between the objects in the next three frame based on ?" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.313, + 0.476, + 0.335 + ], + "angle": 0, + "content": "The person-1 will interact with ball-2 by picking it up or dribbling it. Similarly, person-2 may engage with ball-1 corresponding to" + }, + { + "type": "title", + "bbox": [ + 0.107, + 0.35, + 0.219, + 0.362 + ], + "angle": 0, + "content": "Step-3: Reasoning" + }, + { + "type": "text", + "bbox": [ + 0.108, + 0.37, + 0.326, + 0.381 + ], + "angle": 0, + "content": "What purpose or context can be inferred within" + }, + { + "type": "text", + "bbox": [ + 0.108, + 0.39, + 0.478, + 0.422 + ], + "angle": 0, + "content": "person-1 and person-2 work with ball-1 and ball-2, suggesting drills like dribbling or shooting. Their separation implies independent practice, but their proximity hints at a possible shift to coordinated team drills later in the session" + }, + { + "type": "title", + "bbox": [ + 0.516, + 0.094, + 0.636, + 0.105 + ], + "angle": 0, + "content": "Step-4: Verification" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.111, + 0.886, + 0.133 + ], + "angle": 0, + "content": "Does person-1 move towards ball-2 and begin dribbling, and does person-2 start practicing shots with ball-1 through ?" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.139, + 0.887, + 0.17 + ], + "angle": 0, + "content": "Yes, both individuals are focused on separate balls. person-1 continues with ball-1, and person-2 remains with ball-2, indicating individual practice rather than switching activities or interacting with each other." + }, + { + "type": "title", + "bbox": [ + 0.518, + 0.177, + 0.644, + 0.189 + ], + "angle": 0, + "content": "Step-5: Clarification" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.191, + 0.889, + 0.212 + ], + "angle": 0, + "content": "If instead, person-1 picks up ball-2 and passes it to person-2, how would that change the dynamics of?" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.213, + 0.89, + 0.253 + ], + "angle": 0, + "content": "This would shift the dynamics from individual practice to a collaborative activity, with person-1 passing ball-2 to person-2. The interaction would signal the start of coordinated play, possibly initiating a game involving both ball-1 and ball-2 and highlighting teamwork between person-1 and person-2." + }, + { + "type": "title", + "bbox": [ + 0.52, + 0.263, + 0.682, + 0.274 + ], + "angle": 0, + "content": "Step-6: Scene Forecasting" + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.277, + 0.89, + 0.3 + ], + "angle": 0, + "content": "Given , what are the likely relationship changes between objects in the upcoming phase?" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.306, + 0.89, + 0.337 + ], + "angle": 0, + "content": "The next action could involve person-1 passing ball-2 to person-2 while person-2 passes ball-1 to person-1. Alternatively, person-1 might dribble ball-2 toward person-2, initiating a one-on-one game." + }, + { + "type": "title", + "bbox": [ + 0.52, + 0.35, + 0.641, + 0.363 + ], + "angle": 0, + "content": "Step-7: Hypothesis" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.367, + 0.89, + 0.388 + ], + "angle": 0, + "content": "What if person-3 [600, 200, 160, 500] enters and approaches the others? How might that affect the interactions?" + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.394, + 0.891, + 0.424 + ], + "angle": 0, + "content": "person-1, person-2, and person-3 could engage in coordinated actions, such as three-person passing drills using ball-1 and ball-2. The interaction would evolve from individual practice to collaborative teamwork." + }, + { + "type": "image_caption", + "bbox": [ + 0.137, + 0.433, + 0.858, + 0.447 + ], + "angle": 0, + "content": "Figure 5. An example of the diversified context within the streaming dialog in our VSGR dataset. Best viewed in color and zooming in." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.458, + 0.484, + 0.473 + ], + "angle": 0, + "content": "such as the MSRVTT-QA [49] and MSVD-QA [48] datasets." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.475, + 0.484, + 0.763 + ], + "angle": 0, + "content": "Relation Reasoning (RR). Using the annotated scene graphs, we produce 61,120 relation reasoning tasks by selecting partial information as an incomplete input. Each task requires the model to deduce relationships among entities and identify the categories of the subject and object. With an average of approximately 16 tasks per video, our VSGR dataset provides a substantial collection for evaluating models' abilities in relational reasoning and scene understanding. Question and Answer Validation. To ensure the quality and complexity of questions, we implement a rigorous validation process that combines generation by LLMs with human refinement. Initially, GPT-4/GPT-3.5 generates queries based on the triplets extracted from the videos. Human annotators are trained with specific guidelines that emphasize clarity, relevance, and appropriate challenge levels, then review and refine questions. They enhance the questions by ensuring they are directly answerable from the video content and require careful reasoning about the depicted interactions and relationships." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.765, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We apply strict filtering criteria to improve the quality of the dataset. First, we eliminate questions that do not require video context and can be answered using general world knowledge, ensuring that models must rely on visual information from the videos. Second, we exclude questions that LLMs can answer correctly, increasing the challenge and utility of the dataset in evaluating advanced reasoning abilities. Finally, independent annotators perform a second round to review and evaluate the quality of the refined questions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.457, + 0.707, + 0.474 + ], + "angle": 0, + "content": "5.2. Dataset Comparison" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.482, + 0.909, + 0.664 + ], + "angle": 0, + "content": "As reported in Table 1, our VSGR dataset represents a substantial advancement in video scene graph benchmarks. Our dataset comprises 3,748 videos and 1,841,243 frames, surpassing existing datasets in scale. Unlike other datasets that address only a limited subset of tasks, our dataset offers comprehensive task coverage, facilitating multifaceted evaluations of LLMs. In addition, our ground truth enriches relation annotations with comprehensive textual descriptions, enabling sophisticated reasoning and relationship predictions, as illustrated in Fig. 5. Additionally, our VSGR dataset incorporates diverse viewpoints, including third-person, ego-centric, and drone perspectives, enhancing its generalization." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.682, + 0.704, + 0.699 + ], + "angle": 0, + "content": "6. Experiment Results" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.708, + 0.73, + 0.724 + ], + "angle": 0, + "content": "6.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.733, + 0.907, + 0.779 + ], + "angle": 0, + "content": "Datasets. We leverage our VSGR dataset across five tasks. In addition, we utilize the PVSG [52] dataset for the SGG task and the Action Genome [18] dataset for the SGA task." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Model Configuration. We operate the CLIP-ViT-L-336 [8, 37] to encode each video frame into ten tokens (one CLS token and nine from \\(3 \\times 3\\) average pooling). These tokens are fed into a two-layer MLP connector to the Mistral7B-Instruct [19] language model. For training, we apply LoRA [16] to all linear layers with a rank of 128 and a scaling factor of 256, omitting vision-language alignment [31]. We train for two epochs with a batch size of 128 over 16" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29155" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.287, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.287, + 0.089, + 0.48, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.186, + 0.483, + 0.21 + ], + "angle": 0, + "content": "Figure 6. Comparison of Recall (R) and mean Recall (mR) at different numbers of hyperedges for the SGG task on the VSGR dataset." + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.219, + 0.483, + 0.244 + ], + "angle": 0, + "content": "Table 2. Comparison (\\%) on the VSGR and Action Genome datasets for the Scene Graph Anticipation (SGA) task at varying video input fractions \\(\\mathcal{F}\\)." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.246, + 0.485, + 0.469 + ], + "angle": 0, + "content": "
FMethodAction GenomeVSGR
R/mR@10R/mR@20R/mR@50R/mR@10R/mR@20R/mR@50
0.3STTran+ [5]13.9 / 3.521.6 / 7.340.8 / 20.312.0 / 4.019.0 / 8.035.7 / 18.0
DSGDetr+ [11]14.3 / 3.621.8 / 7.641.3 / 21.212.5 / 4.219.5 / 8.336.0 / 19.0
STTran++ [5]15.4 / 6.227.2 / 14.148.6 / 32.214.0 / 6.522.0 / 11.039.0 / 25.1
DSGDetr++ [11]16.8 / 8.429.0 / 16.748.9 / 32.314.5 / 7.022.5 / 12.040.0 / 26.0
SceneSayerODE [36]23.3 / 13.332.5 / 20.145.1 / 33.018.0 / 9.526.0 / 15.042.0 / 30.0
SceneSayerSDE [36]25.9 / 15.635.7 / 23.147.4 / 37.119.5 / 11.027.5 / 17.044.0 / 33.5
HyperGraph (Ours)26.5 / 14.836.2 / 22.149.3 / 37.218.8 / 9.327.2 / 16.342.5 / 27.4
HyperGLM (Ours)27.5 / 15.837.0 / 24.550.0 / 38.019.0 / 10.028.0 / 16.543.0 / 27.5
0.5STTran+ [5]14.9 / 3.722.6 / 7.642.9 / 21.413.5 / 4.221.0 / 8.537.0 / 19.0
DSGDetr+ [11]15.2 / 3.923.1 / 8.043.3 / 22.213.8 / 4.521.5 / 9.038.0 / 20.0
STTran++ [5]16.6 / 6.629.1 / 14.751.5 / 33.415.5 / 7.023.5 / 12.541.0 / 26.5
DSGDetr++ [11]17.4 / 8.430.5 / 17.051.9 / 33.916.0 / 7.524.0 / 13.042.0 / 28.0
SceneSayerODE [36]26.4 / 14.336.6 / 21.449.8 / 36.020.5 / 11.029.5 / 16.546.1 / 32.5
SceneSayerSDE [36]28.4 / 16.338.6 / 25.151.4 / 39.921.5 / 12.531.0 / 18.548.0 / 35.7
HyperGraph (Ours)29.2 / 16.439.3 / 23.252.1 / 38.720.3 / 12.229.8 / 17.146.2 / 33.1
HyperGLM (Ours)30.0 / 17.040.5 / 27.553.5 / 40.521.5 / 11.531.5 / 18.546.5 / 30.0
0.7STTran+ [5]16.6 / 4.225.1 / 8.547.2 / 24.015.0 / 5.024.0 / 10.541.0 / 21.5
DSGDetr+ [11]16.8 / 4.325.3 / 8.847.4 / 24.715.5 / 5.224.5 / 11.042.0 / 22.0
STTran++ [5]19.0 / 7.732.8 / 17.156.8 / 36.817.0 / 8.027.0 / 14.045.0 / 29.0
DSGDetr++ [11]19.8 / 9.534.1 / 19.256.7 / 37.217.5 / 8.528.0 / 15.046.1 / 30.0
SceneSayerODE [36]32.1 / 16.542.8 / 24.455.6 / 39.623.5 / 13.033.5 / 19.051.0 / 36.0
SceneSayerSDE [36]33.3 / 18.144.0 / 27.356.4 / 44.424.5 / 14.535.7 / 21.053.0 / 38.0
HyperGraph (Ours)34.3 / 19.245.2 / 25.357.2 / 42.122.3 / 13.234.2 / 19.350.4 / 35.3
HyperGLM (Ours)35.7 / 19.546.1 / 30.058.2 / 44.025.1 / 13.535.5 / 21.551.0 / 33.5
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.484, + 0.464, + 0.499 + ], + "angle": 0, + "content": "iterations on \\(4 \\times\\) GPUs, taking approximately six hours." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.5, + 0.484, + 0.669 + ], + "angle": 0, + "content": "Metrics. We evaluate the SGG and SGA tasks using the Recall and mean Recall scores. In addition, we evaluate the VQA and RR tasks by Accuracy, Precision, Recall, and F1 scores. For the VC task, we utilize CIDEr, MENTOR, ROUGE-L, and BLEU-4 scores to validate our performance. Settings. For the SGG and SGA tasks, we adopt the evaluation settings based on [18, 36]. In particular, the model is provided with raw video frames and must detect objects using a pre-trained detector (i.e. Faster R-CNN) and predict or anticipate their relationships. Especially for the SGA, we set the initial video input fraction \\((\\mathcal{F})\\) to 0.9, following [36]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.68, + 0.245, + 0.696 + ], + "angle": 0, + "content": "6.2. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.486, + 0.901 + ], + "angle": 0, + "content": "Hypergraph Parameters. The number of walks \\((N_w)\\) and walk length \\((N_l)\\) are introduced in Alg. 1, directly impacting the capacity to capture high-order relationships by determining the number of hyperedges. Although more hyperedges increase relational diversity, they can also introduce redundancy beyond an optimal point. Specifically, a higher \\(N_w\\) broadens the range of sampled relationships, while a moderate \\(N_l\\) balances depth. Our experiments demonstrate that optimal performances are achieved with 60 hyperedges \\((N_w = 60\\) and \\(N_l = 7)\\), shown in Fig. 6. Further experiments on the SGG task are provided in Table 3, and experiments on these parameters are included in the Appendices. Video Input Fraction. We adjust the initial video input" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.089, + 0.907, + 0.113 + ], + "angle": 0, + "content": "Table 3. Comparison \\((\\%)\\) on the VSGR and PVSG datasets for the Scene Graph Generation (SGG) task at Recall (R) and mean Recall (mR)." + }, + { + "type": "table", + "bbox": [ + 0.513, + 0.115, + 0.907, + 0.186 + ], + "angle": 0, + "content": "
MethodPVSGVSGR
R/mR@20R/mR@50R/mR@100R/mR@20R/mR@50R/mR@100
Transformer [52]4.0 / 1.84.4 / 1.94.9 / 2.025.7 / 6.334.5 / 6.543.5 / 7.0
HIG [34]4.6 / 1.94.9 / 2.15.4 / 2.223.8 / 5.731.1 / 5.940.4 / 6.9
CYCLO [35]5.8 / 2.06.1 / 2.26.7 / 2.329.4 / 7.136.4 / 7.747.7 / 7.7
HyperGraph (Ours)6.5 / 2.27.0 / 2.47.5 / 2.631.6 / 7.838.8 / 8.350.3 / 8.5
HyperGLM (Ours)7.5 / 2.88.1 / 3.78.5 / 3.935.8 / 9.242.3 / 10.154.7 / 10.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.194, + 0.907, + 0.218 + ], + "angle": 0, + "content": "Table 4. Comparison (%) on the VSGR and Action Genome datasets for the Scene Graph Anticipation (SGA) task at Recall (R) and mean Recall (mR)." + }, + { + "type": "table", + "bbox": [ + 0.514, + 0.22, + 0.907, + 0.317 + ], + "angle": 0, + "content": "
MethodAction GenomeVSGR
R/mR@10R/mR@20R/mR@50R/mR@10R/mR@20R/mR@50
STTran+ [5]17.5 / 4.626.8 / 9.249.6 / 24.316.5 / 5.526.0 / 11.043.0 / 23.0
DSGDetr+ [11]17.9 / 4.727.7 / 9.751.4 / 25.917.0 / 6.027.0 / 11.544.5 / 24.5
STTran++ [5]20.2 / 8.935.7 / 18.460.2 / 38.819.0 / 9.530.0 / 15.549.0 / 31.0
DSGDetr++ [11]22.2 / 11.437.1 / 21.061.0 / 39.519.5 / 10.031.0 / 16.050.0 / 32.5
SceneSayerODE [36]36.6 / 17.848.3 / 27.461.3 / 43.426.5 / 14.037.5 / 20.055.5 / 38.0
SceneSayerSDE [36]37.3 / 20.848.6 / 30.961.6 / 46.827.5 / 16.038.5 / 22.058.2 / 40.0
HyperGraph (Ours)37.5 / 19.149.3 / 31.462.3 / 47.428.4 / 17.539.3 / 22.457.6 / 41.5
HyperGLM (Ours)38.8 / 22.351.5 / 33.065.2 / 48.630.2 / 18.141.1 / 23.559.3 / 43.4
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.322, + 0.909, + 0.459 + ], + "angle": 0, + "content": "fraction, \\(\\mathcal{F}\\), to 0.3, 0.5, and 0.7 for the SGA task. This adjustment allows the model to learn from varying observed portions and predict the unseen segment. Table 2 indicates that increasing the portion of the seen video improves performance, suggesting that additional visual context is beneficial. In addition, Table 4 further confirms that the default setting at a higher input fraction (\\(F = 0.9\\)) leads to optimal performance. We also present additional settings with varying video input fractions for the SGA task in the Appendices." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.469, + 0.818, + 0.485 + ], + "angle": 0, + "content": "6.3. Comparison with State-of-the-Arts" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.492, + 0.909, + 0.764 + ], + "angle": 0, + "content": "Table 3 demonstrates that HyperGLM significantly outperforms existing methods on the PVSG and VSGR datasets, achieving the highest \\( R@20 \\) scores of \\( 7.5\\% \\) and \\( 35.8\\% \\), respectively. By leveraging hyperedges connecting multiple nodes within the HyperGraph, HyperGLM effectively captures complex object interactions and spatial dependencies, transcending traditional pairwise methods to generate more accurate and detailed scene graph representations. Notably, integrating a procedural graph within the HyperGraph reduces bias. Our approach enhances mean Recall, reaching improvements of \\( 2.8\\% \\) on the PVSG dataset and \\( 9.2\\% \\) on the VSGR dataset, thereby addressing the long-tail distribution challenges that have struggled in previous methods [34, 35, 52]. Furthermore, the LLM enhances the capacity of HyperGLM to infer and predict intricate relationships embedded within the unified HyperGraph, resulting in improved performance compared to HyperGraph, which shows a decrease of \\( 4.2\\% \\) at \\( R@20 \\) on the VSGR dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.911, + 0.903 + ], + "angle": 0, + "content": "As shown in Table 4, our HyperGLM approach outperforms existing SGA methods on the Action Genome and VSGR datasets, achieving R@10 scores of \\(35.7\\%\\) and \\(25.1\\%\\), respectively. This improvement stems from integrating procedural graphs that model causal relationships within the HyperGraph structure. In contrast, the SceneSayer [36] method relies on NeuralODE and NeuralSDE to capture the latent dynamics of object interaction evolution. Our procedural graphs enable multi-step transitions that explicitly" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29156" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.105, + 0.075, + 0.501, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.075, + 0.896, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.178, + 0.499, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.223, + 0.384, + 0.371, + 0.395 + ], + "angle": 0, + "content": "a) Scene Graph Generation" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.178, + 0.902, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.618, + 0.385, + 0.774, + 0.395 + ], + "angle": 0, + "content": "b) Scene Graph Anticipation" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.396, + 0.907, + 0.432 + ], + "angle": 0, + "content": "Figure 7. Qualitative comparison of our HyperGLM approach versus SceneSayerSDE [36] for SGG and SGA. Red and green edge labels denote incorrect and correct predictions, respectively. Our HyperGLM approach effectively captures the evolving interactions between person-1 and bicycle-1 or person-2 and person-3 over time and anticipates interactions in unseen video frames, while SceneSayerSDE confuses to similar predicates. Best viewed in color." + }, + { + "type": "table_caption", + "bbox": [ + 0.098, + 0.435, + 0.476, + 0.447 + ], + "angle": 0, + "content": "Table 5. Comparison (%) on VSGR for the Video Question Answering." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.449, + 0.482, + 0.536 + ], + "angle": 0, + "content": "
MethodAccuracyPrecisionRecallF1 Score
Video-ChatGPT [32]33.235.132.333.6
Video-LLaVA-7B [30]43.143.841.742.7
MovieChat [41]43.544.242.643.4
Chat-UniVi-7B [20]44.345.643.244.4
HyperGLM (Ours)45.447.244.345.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.124, + 0.537, + 0.45, + 0.549 + ], + "angle": 0, + "content": "Table 6. Comparison (%) on VSGR for the Video Captioning." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.55, + 0.483, + 0.624 + ], + "angle": 0, + "content": "
MethodCIDErMENTORROUGE-LBLEU-4
MV-GPT [38]57.137.562.547.2
CoCap [40]54.329.461.842.8
UniVL + MELTR [24]50.528.160.042.1
HyperGLM (Ours)54.530.764.948.8
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.749 + ], + "angle": 0, + "content": "capture the temporal evolution of relationships by modeling sequential interactions and their dependencies. In contrast, NeuralODE and NeuralSDE primarily focus on continuous-time dynamics, which may limit their effectiveness in handling discrete, multi-step relational changes as illustrated in Fig. 7b. Significantly, the HyperGraph is injected into the LLM, allowing the model to capture complex temporal patterns, resulting in better results than the HyperGraph." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In addition to the improvements in SGG and SGA shown in Fig. 7, Tables 5, 6, and 7 demonstrate the significant improvements using our HyperGLM approach in the VQA, VC, and RR tasks. In VQA, the hyperedges connect multiple objects, enabling HyperGLM to capture context-rich interactions more effectively, reaching an accuracy of \\(45.4\\%\\). For VC, the HyperGraph structure supports evolving object connections over time, allowing our HyperGLM approach to generate captions that describe the relationship between objects and capture the temporal flow of interactions coher-" + }, + { + "type": "table_caption", + "bbox": [ + 0.54, + 0.435, + 0.878, + 0.447 + ], + "angle": 0, + "content": "Table 7. Comparison (%) on VSGR for the Relation Reasoning." + }, + { + "type": "table", + "bbox": [ + 0.514, + 0.449, + 0.906, + 0.526 + ], + "angle": 0, + "content": "
MethodAccuracyPrecisionRecallF1 Score
Video-LLaVA-7B [30]41.342.540.241.3
MA-LMM [14]42.843.741.842.7
LLaMA-VID-7B [29]44.145.243.544.3
HyperGLM (Ours)47.248.446.547.4
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.543, + 0.907, + 0.604 + ], + "angle": 0, + "content": "ently, achieving scores of \\(54.5\\%\\) at CIDEr and \\(64.9\\%\\) at ROUGE-L. In RR task, the HyperGraph effectively manages intricate dependencies between multiple objects, resulting in precise relational inferences with an accuracy of \\(47.2\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.626, + 0.634, + 0.641 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.654, + 0.907, + 0.759 + ], + "angle": 0, + "content": "In this paper, we have introduced HyperGLM, a novel VidSGG method that integrates scene hypergraph information into LLMs for context-aware and precise scene interpretation. Our approach effectively models complex interactions and higher-order relationships. It outperforms leading methods on benchmarks, including PVSG, Action Genome, and our newly collected VSGR dataset across five tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.762, + 0.907, + 0.838 + ], + "angle": 0, + "content": "Limitations. Although our HyperGLM approach effectively models multi-way interactions, managing many objects and their interactions can complicate relationship structures. As the HyperGraph expands, essential relationships may become obscured, reducing the clarity of scene interpretation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Acknowledgment. This material is based upon work supported by the National Science Foundation under Award No. OIA-1946391. We also acknowledge the Arkansas High-Performance Computing Center for providing GPUs." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29157" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.155 + ], + "angle": 0, + "content": "[1] Song Bai, Feihu Zhang, and Philip HS Torr. Hypergraph convolution and hypergraph attention. Pattern Recognition, 110:107637, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.156, + 0.484, + 0.238 + ], + "angle": 0, + "content": "[2] Joya Chen, Zhaoyang Lv, Shiwei Wu, Kevin Qinghong Lin, Chenan Song, Difei Gao, Jia-Wei Liu, Ziteng Gao, Dongxing Mao, and Mike Zheng Shou. Videollm-online: Online video large language model for streaming video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18407-18418, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.239, + 0.484, + 0.305 + ], + "angle": 0, + "content": "[3] Zhanwen Chen, Saed Rezayi, and Sheng Li. More knowledge, less bias: Unbiasing scene graph generation with explicit ontological adjustment. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4023-4032, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.306, + 0.484, + 0.374 + ], + "angle": 0, + "content": "[4] Yuren Cong, Wentong Liao, Hanno Ackermann, Bodo Rosenhahn, and Michael Ying Yang. Spatial-temporal transformer for dynamic scene graph generation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16372-16382, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.374, + 0.483, + 0.428 + ], + "angle": 0, + "content": "[5] Yuren Cong, Wentong Liao, H. Ackermann, M. Yang, and B. Rosenhahn. Spatial-temporal transformer for dynamic scene graph generation. IEEE International Conference on Computer Vision, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.429, + 0.484, + 0.481 + ], + "angle": 0, + "content": "[6] Yuren Cong, Michael Ying Yang, and Bodo Rosenhahn. Reltr: Relation transformer for scene graph generation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(9):11169-11183, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.482, + 0.484, + 0.55 + ], + "angle": 0, + "content": "[7] Youming Deng, Yansheng Li, Yongjun Zhang, Xiang Xiang, Jian Wang, Jingdong Chen, and Jiayi Ma. Hierarchical memory learning for fine-grained scene graph generation. In European Conference on Computer Vision, pages 266-283. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.55, + 0.484, + 0.645 + ], + "angle": 0, + "content": "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.646, + 0.484, + 0.713 + ], + "angle": 0, + "content": "[9] Azade Farshad, Yousef Yeganeh, Yu Chi, Chengzhi Shen, Böjrn Ommer, and Nassir Navab. Scenegenie: Scene graph guided diffusion models for image synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 88-98, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.714, + 0.484, + 0.781 + ], + "angle": 0, + "content": "[10] Hao Fei, Shengqiong Wu, Wei Ji, Hanwang Zhang, Meishan Zhang, Mong-Li Lee, and Wynne Hsu. Video-of-thought: Step-by-step video reasoning from perception to cognition. In *Forty-first International Conference on Machine Learning*, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.781, + 0.484, + 0.848 + ], + "angle": 0, + "content": "[11] Shengyu Feng, Hesham Mostafa, Marcel Nassar, Somdeb Majumdar, and Subarna Tripathi. Exploiting long-term dependencies for generating dynamic scene graphs. IEEE Workshop/Winter Conference on Applications of Computer Vision, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.849, + 0.484, + 0.891 + ], + "angle": 0, + "content": "[12] Yue Gao, Yifan Feng, Shuyi Ji, and Rongrong Ji. Hgnn+: General hypergraph neural networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):3181-3199," + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.891 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.597, + 0.105 + ], + "angle": 0, + "content": "2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.106, + 0.907, + 0.16 + ], + "angle": 0, + "content": "[13] Zeeshan Hayden and Xuming He. Dsgg: Dense relation transformer for an end-to-end scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28317-28326, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.16, + 0.907, + 0.242 + ], + "angle": 0, + "content": "[14] Bo He, Hengduo Li, Young Kyun Jang, Menglin Jia, Xuefei Cao, Ashish Shah, Abhinav Shrivastava, and Ser-Nam Lim. Ma-lmm: Memory-augmented large multimodal model for long-term video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13504-13514, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.242, + 0.907, + 0.296 + ], + "angle": 0, + "content": "[15] Tao He, Lianli Gao, Jingkuan Song, and Yuan-Fang Li. Towards open-vocabulary scene graph generation with prompt-based finetuning. In European Conference on Computer Vision, pages 56-73. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.296, + 0.907, + 0.363 + ], + "angle": 0, + "content": "[16] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.364, + 0.906, + 0.432 + ], + "angle": 0, + "content": "[17] Jinbae Im, Jeong Yeon Nam, Nokyung Park, Hyungmin Lee, and Seunghyun Park. Egtr: Extracting graph from transformer for scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24229-24238, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.432, + 0.907, + 0.5 + ], + "angle": 0, + "content": "[18] Jingwei Ji, Ranjay Krishna, Li Fei-Fei, and Juan Carlos Niebles. Action genome: Actions as compositions of spatiotemporal scene graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10236–10247, 2020. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.5, + 0.907, + 0.567 + ], + "angle": 0, + "content": "[19] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.568, + 0.907, + 0.648 + ], + "angle": 0, + "content": "[20] Peng Jin, Ryuichi Takanobu, Wancai Zhang, Xiaochun Cao, and Li Yuan. Chat-univi: Unified visual representation empowers large language models with image and video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13700-13710, 2024. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.649, + 0.907, + 0.731 + ], + "angle": 0, + "content": "[21] Tianlei Jin, Fangtai Guo, Qiwei Meng, Shiqiang Zhu, Xiangming Xi, Wen Wang, Zonghao Mu, and Wei Song. Fast contextual scene graph generation with unbiased context augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6302-6311, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.731, + 0.906, + 0.799 + ], + "angle": 0, + "content": "[22] Eun-Sol Kim, Woo Young Kang, Kyoung-Woon On, Yu-Jung Heo, and Byoung-Tak Zhang. Hypergraph attention networks for multimodal learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14581-14590, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.799, + 0.907, + 0.88 + ], + "angle": 0, + "content": "[23] Kibum Kim, Kanghoon Yoon, Jaehyeong Jeon, Yeunjun In, Jinyoung Moon, Donghyun Kim, and Chanyoung Park. Llm4sgg: Large language models for weakly supervised scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28306-28316, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.88, + 0.907, + 0.895 + ], + "angle": 0, + "content": "[24] Dohwan Ko, Joonmyung Choi, Hyeong Kyu Choi, Kyoung-" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.895 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29158" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.16 + ], + "angle": 0, + "content": "Woon On, Byungseok Roh, and Hyunwoo J Kim. Meltr: Meta loss transformer for learning to fine-tune video foundation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20105-20115, 2023. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.161, + 0.484, + 0.215 + ], + "angle": 0, + "content": "[25] Sanjoy Kundu and Sathyanarayanan N Aakur. Is-ggt: Iterative scene graph generation with generative transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6292-6301, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.215, + 0.484, + 0.269 + ], + "angle": 0, + "content": "[26] Rongjie Li, Songyang Zhang, and Xuming He. Sgtr: End-to-end scene graph generation with transformer. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19486-19496, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.269, + 0.483, + 0.338 + ], + "angle": 0, + "content": "[27] Rongjie Li, Songyang Zhang, Dahua Lin, Kai Chen, and Xuming He. From pixels to graphs: Open-vocabulary scene graph generation with vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28076-28086, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.338, + 0.483, + 0.392 + ], + "angle": 0, + "content": "[28] Wanxin Li, Wei Xie, Zhigang Tu, Wei Wang, and Lianghao Jin. Multi-hyperedge hypergraph for group activity recognition. In 2022 International Joint Conference on Neural Networks (IJCNN), pages 01-07. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.392, + 0.484, + 0.445 + ], + "angle": 0, + "content": "[29] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In European Conference on Computer Vision, pages 323–340. Springer, 2025. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.446, + 0.483, + 0.499 + ], + "angle": 0, + "content": "[30] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.499, + 0.483, + 0.54 + ], + "angle": 0, + "content": "[31] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.54, + 0.484, + 0.649 + ], + "angle": 0, + "content": "[32] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Khan. Video-ChatGPT: Towards detailed video understanding via large vision and language models. In Lun-Wei Ku, Andre Martins, and Vivek Srikumar, editors, Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 12585–12602, Bangkok, Thailand, Aug. 2024. Association for Computational Linguistics. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.649, + 0.484, + 0.703 + ], + "angle": 0, + "content": "[33] Sayak Nag, Kyle Min, Subarna Tripathi, and Amit K Roy-Chowdhury. Unbiased scene graph generation in videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22803-22813, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.703, + 0.484, + 0.77 + ], + "angle": 0, + "content": "[34] Trong-Thuan Nguyen, Pha Nguyen, and Khoa Luu. Hig: Hierarchical interlacement graph approach to scene graph generation in video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 1, 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.771, + 0.483, + 0.839 + ], + "angle": 0, + "content": "[35] Trong-Thuan Nguyen, Pha Nguyen, Li Xin, Cothren Jackson, Yilmaz Alper, and Khoa Luu. CYCLO: Cyclic graph transformer approach to multi-object relationship modeling in aerial videos. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 1, 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.839, + 0.483, + 0.88 + ], + "angle": 0, + "content": "[36] Rohith Peddi, Saksham Singh, Parag Singla, Vibhav Gogate, et al. Towards scene graph anticipation. In European Conference on Computer Vision. Springer, 2024. 1, 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.88, + 0.483, + 0.893 + ], + "angle": 0, + "content": "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.893 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.092, + 0.907, + 0.16 + ], + "angle": 0, + "content": "Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.161, + 0.907, + 0.229 + ], + "angle": 0, + "content": "[38] Paul Hongsuck Seo, Arsha Nagrani, Anurag Arnab, and Cordelia Schmid. End-to-end generative pretraining for multimodal video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17959-17968, 2022. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.229, + 0.907, + 0.283 + ], + "angle": 0, + "content": "[39] Xindi Shang, Yicong Li, Junbin Xiao, Wei Ji, and Tat-Seng Chua. Video visual relation detection via iterative inference. In Proceedings of the 29th ACM international conference on Multimedia, pages 3654-3663, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.283, + 0.906, + 0.338 + ], + "angle": 0, + "content": "[40] Yaojie Shen, Xin Gu, Kai Xu, Heng Fan, Longyin Wen, and Libo Zhang. Accurate and fast compressed video captioning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15558-15567, 2023. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.338, + 0.906, + 0.419 + ], + "angle": 0, + "content": "[41] Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Haozhe Chi, Xun Guo, Tian Ye, Yanting Zhang, et al. Moviechat: From dense token to sparse memory for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18221-18232, 2024. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.419, + 0.906, + 0.473 + ], + "angle": 0, + "content": "[42] Yao Teng, Limin Wang, Zhifeng Li, and Gangshan Wu. Target adaptive context aggregation for video scene graph generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13688-13697, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.473, + 0.907, + 0.54 + ], + "angle": 0, + "content": "[43] Nupur Thakur, PrasanthSai Gouripeddi, and Baoxin Li. Graph (graph): A nested graph-based framework for early accident anticipation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 7533-7541, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.54, + 0.907, + 0.609 + ], + "angle": 0, + "content": "[44] Aisha Urooj, Hilde Kuehne, Bo Wu, Kim Chheu, Walid BousSELham, Chuang Gan, Niels Lobo, and Mubarak Shah. Learning situation hyper-graphs for video question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14879-14889, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.609, + 0.907, + 0.677 + ], + "angle": 0, + "content": "[45] Jiahao Wang, Guo Chen, Yifei Huang, Limin Wang, and Tong Lu. Memory-and-anticipation transformer for online action understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13824–13835, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.677, + 0.907, + 0.731 + ], + "angle": 0, + "content": "[46] Yanan Wang, Shuichiro Haruta, Donghuo Zeng, Julio Vizcarra, and Mori Kurokawa. Multi-object event graph representation learning for video question answering. In *Meeting on Image Recognition and Understanding*, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.731, + 0.907, + 0.786 + ], + "angle": 0, + "content": "[47] Tao Wu, Runyu He, Gangshan Wu, and Limin Wang. Sportshhi: A dataset for human-human interaction detection in sports videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.786, + 0.907, + 0.853 + ], + "angle": 0, + "content": "[48] Dejing Xu, Zhou Zhao, Jun Xiao, Fei Wu, Hanwang Zhang, Xiangnan He, and Yueting Zhuang. Video question answering via gradually refined attention over appearance and motion. In Proceedings of the 25th ACM international conference on Multimedia, pages 1645-1653, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.853, + 0.906, + 0.893 + ], + "angle": 0, + "content": "[49] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In Proceedings of the IEEE conference on computer vision and" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.893 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "29159" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.405, + 0.106 + ], + "angle": 0, + "content": "pattern recognition, pages 5288-5296, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.107, + 0.484, + 0.159 + ], + "angle": 0, + "content": "[50] Zenan Xu, Wanjun Zhong, Qinliang Su, Zijing Ou, and Fuwei Zhang. Modeling semantic composition with syntactic hypergraph for video question answering. arXiv preprint arXiv:2205.06530, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.16, + 0.484, + 0.214 + ], + "angle": 0, + "content": "[51] Jingkang Yang, Yi Zhe Ang, Zujin Guo, Kaiyang Zhou, Wayne Zhang, and Ziwei Liu. Panoptic scene graph generation. In European Conference on Computer Vision, pages 178-196. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.215, + 0.484, + 0.295 + ], + "angle": 0, + "content": "[52] Jingkang Yang, Wenxuan Peng, Xiangtai Li, Zujin Guo, Liangyu Chen, Bo Li, Zheng Ma, Kaiyang Zhou, Wayne Zhang, Chen Change Loy, et al. Panoptic video scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18675-18685, 2023. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.296, + 0.484, + 0.363 + ], + "angle": 0, + "content": "[53] Guangyao Zhai, Evin Pinar Örnek, Shun-Cheng Wu, Yan Di, Federico Tombari, Nassir Navab, and Benjamin Busam. Commonscenes: Generating commonsense 3d indoor scenes with scene graphs. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.364, + 0.483, + 0.403 + ], + "angle": 0, + "content": "[54] Zizhao Zhang, Yifan Feng, Shihui Ying, and Yue Gao. Deep hypergraph structure learning. arXiv preprint arXiv:2208.12547, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.405, + 0.482, + 0.459 + ], + "angle": 0, + "content": "[55] Xiaolin Zhu, Dongli Wang, Jianxun Li, Rui Su, Qin Wan, and Yan Zhou. Dynamical attention hypergraph convolutional network for group activity recognition. IEEE Transactions on Neural Networks and Learning Systems, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.107, + 0.484, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "29160" + } + ] +] \ No newline at end of file diff --git a/2025/HyperGLM_ HyperGraph for Video Scene Graph Generation and Anticipation/07619f38-6a65-46d1-ae2a-3885ad0e2f6d_origin.pdf b/2025/HyperGLM_ HyperGraph for Video Scene Graph Generation and Anticipation/07619f38-6a65-46d1-ae2a-3885ad0e2f6d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..741ac0d92dff408945135a543549e810ebca6182 --- /dev/null +++ b/2025/HyperGLM_ HyperGraph for Video Scene Graph Generation and Anticipation/07619f38-6a65-46d1-ae2a-3885ad0e2f6d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7f78f07d1d31121b495a373303a4e71a2f4c0a51d28e740f9302302ceebb892c +size 8241185 diff --git a/2025/HyperGLM_ HyperGraph for Video Scene Graph Generation and Anticipation/full.md b/2025/HyperGLM_ HyperGraph for Video Scene Graph Generation and Anticipation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7532073c06864f9f1bbd4eddab48a969055dc87f --- /dev/null +++ b/2025/HyperGLM_ HyperGraph for Video Scene Graph Generation and Anticipation/full.md @@ -0,0 +1,392 @@ +# HyperGLM: HyperGraph for Video Scene Graph Generation and Anticipation + +Trong-Thuan Nguyen $^{1\dagger}$ , Pha Nguyen $^{1\dagger}$ , Jackson Cothren $^{1}$ , Alper Yilmaz $^{2}$ , Khoa Luu $^{1}$ $^{1}$ University of Arkansas $^{2}$ Ohio State University + +$^{1}\{$ thuann, panguyen, jcothre, khoaluu\}@uark.edu $^{2}$ yilmaz.15@osu.edu uark-cviu.github.io/projects/HyperGLM + +![](images/7ec01ce8a7d7ad73de9fa91943f0ab8555b159ceb724809c9ac3eb2b093a6f09.jpg) +What are the roles of the people? + +![](images/5069f06678a48c5320e6bbad12cc3935df7e1314e9541189e070d67711d3fc70.jpg) +What is the relationship between objects? +Person-1 holding the Umbrella-1 and standing beside Person-2. Person-1 and Person-2 are waiting for Person-3 in her yellow dress. Then, Person -1 talking with Person-3. +Figure 1. Our HyperGLM framework supports Video Scene Graph Generation, Anticipation, and Reasoning. HyperGLM constructs scene graphs from observed video frames and predicts relationships in unseen frames by leveraging a unified hypergraph for temporal modeling and comprehensive understanding. + +It appears that Person-1 is picking up his family, as Person-3 is being dropped off from the helicopter in the rainy weather. Person-2 is likely serving as security. + +What might the next interaction be? + +Person-1 and Person-3 are talking while waiting for another family member to get off the helicopter. + +What might they do after moving away? + +After arriving by helicopter, Person-1, Person-3, and another woman will likely move indoors to rest. + +# Abstract + +Multimodal LLMs have advanced vision-language tasks but still struggle with understanding video scenes. To bridge this gap, Video Scene Graph Generation (VidSGG) has emerged to capture multi-object relationships across video frames. However, prior methods rely on pairwise connections, limiting their ability to handle complex multi-object interactions and reasoning. To this end, we propose the Multimodal Large Language Models (LLMs) on a Scene HyperGraph (HyperGLM), promoting reasoning about multi-way interactions and higher-order relationships. Our approach uniquely integrates entity scene graphs, which capture spatial relationships between objects, with a procedural graph that models their causal transitions, forming a unified HyperGraph. Significantly, HyperGLM enables reasoning by injecting this unified HyperGraph into LLMs. Additionally, we introduce a new Video Scene Graph Reasoning (VSGR) dataset featuring 1.9M frames from third-person, egocentric, and drone views and support five tasks. Empirically, HyperGLM consistently outperforms state-of-the-art methods, effectively modeling and reasoning complex relationships in diverse scenes. + +# 1. Introduction + +In recent years, Multimodal Large Language Models (LLMs) [2, 10] have set new benchmarks, with vision-language models excelling in diverse multimodal tasks. However, fully understanding dynamic video scenes remains a significant challenge for applications like autonomous driving, intelligent surveillance, human-object interaction, and multimedia analysis. Towards this goal, Video Scene Graph Generation (VidSGG) [18, 52] has emerged as a critical task for capturing multi-object relationships across video frames. In particular, VidSGG enables high-level tasks such as event forecasting [36, 43, 45], video captioning [24, 38, 40], and video question answering [20, 30, 32, 41] by constructing detailed representations of entities and their interactions. + +However, prior VidSGG methods and datasets have limited capacity for comprehensive video understanding. Traditional scene graph-based methods [4, 34, 35, 42] only model pairwise object relationships within single frames, making it challenging to capture higher-order relationships and temporal dependencies in real-world scenarios. Additionally, existing benchmark datasets [18, 34, 35, 52] focus is primarily confined to Scene Graph Generation (SGG) and Scene + +Graph Anticipation (SGA) tasks, lacking annotations for reasoning tasks such as Video Question Answering (VQA), Video Captioning (VC), and Relation Reasoning (RR). + +In this paper, we propose the Multimodal LLMs on a Scene HyperGraph (HyperGLM) approach to promote reasoning about multi-way interactions and higher-order relationships through a unified HyperGraph and LLMs, as illustrated in Fig. 1. To achieve this goal, we uniquely incorporate entity scene graphs, which capture spatial relationships between objects, with a procedural graph that models their causal transitions across video frames, forming a unified HyperGraph, as shown in Fig. 2. Significantly, our HyperGLM approach allows reasoning by injecting this unified HyperGraph into LLMs. In addition, we introduce a novel Video Scene Graph Reasoning (VSGR) dataset, comprising 1.9 million video frames surpassing existing benchmark datasets [18, 34, 35, 47, 52] in scale and annotation depth. Specifically, our VSGR dataset includes videos from third-person, egocentric, and drone perspectives. It supports five tasks: Scene Graph Generation, Scene Graph Anticipation, Video Question Answering, Video Captioning, and Relation Reasoning. Notably, our VSGR dataset introduces a new Relation Reasoning task, setting it apart from existing video scene graph datasets, as shown in the comparison in Table 1. + +Contributions of this Work. This work presents three contributions to advance Video Scene Graph Generation. First, we introduce Multimodal LLMs on a Scene HyperGraph, leveraging hyperedges and LLMs for reasoning about multiway interactions and higher-order relationships. Second, we develop a new Video Scene Graph Reasoning dataset, surpassing existing scale and annotation depth benchmarks. Our VSGR dataset primarily supports five tasks within diverse video scenes. Finally, the proposed HyperGLM consistently outperforms state-of-the-art methods across all five tasks. + +# 2. Related Work + +In this section, we review advances in scene graph generation and hypergraph applications in computer vision, then discuss existing limitations and the advantages of our approach. + +# 2.1. Scene Graph Generation + +Scene Graph Generation has significantly advanced with transformer-based models [6, 13, 17, 25, 26, 51] that have become benchmarks due to their efficiency and state-of-the-art performance. Recent work [3, 21, 33] focuses on reducing bias and enhancing mean recall for rare predicates by integrating external knowledge and applying unbiased contextual augmentation, particularly in dynamic video contexts. Open-vocabulary methods [15, 27] supported by vision-language models handle unseen object and relationship classes, improving generalization. Additionally, generative models (e.g., diffusion-based methods [9, 53]) leverage scene graphs for efficient image and scene synthesis. More- + +over, spatial-temporal methods [4, 34-36, 42] effectively capture dynamic object relationships in videos. Recently, Large Language Models [23] have been utilized to enhance triplet extraction and alignment in weakly supervised SGG. + +# 2.2. HyperGraphs + +HyperGraphs have been adopted in computer vision to model complex multimodal data and capture higher-order relationships. Unlike traditional graph-based approaches, HyperGraphs connect multiple nodes through hyperedges, enabling multi-way relationships. They enhance Graph Neural Networks (GNNs) [12, 54] by allowing the modeling of more sophisticated interactions. Recent advancements, such as HyperGraph Convolution [1] and HyperGraph Attention [22], have further improved GNNs by capturing relationships beyond simple pairwise connections. Therefore, HyperGraph-based models effectively handle temporal dependencies and complex interactions, significantly boosting performance in tasks like accident anticipation [43], group activity recognition [28, 55], and video question answering [44, 46, 50]. + +# 2.3. Discussion + +Limitations in Prior Methods. The methods introduced in Section 2.1, based on Progressive Feature Fusion [39, 51], Batch-Progressive Transformer Encoding [6, 13, 17, 25, 26], Spatial-Temporal Context Integration [4, 34, 42], and Memory-Guided Temporal Consistency [7, 33, 35], have advanced VidSGG. However, these methods struggle to model higher-order relationships and complex temporal dynamics. Specifically, Progressive Feature Fusion and Batch-Progressive Transformer Encoding are limited in capturing long-term temporal dependencies, with the former lacking long-term context due to frame-by-frame processing and the latter only addressing short-term dependencies. Similarly, Spatial-Temporal Context Integration and Memory-Guided Temporal Consistency inadequately represent multi-object interactions across video frames and insufficiently capture the temporal evolution required for higher-order relationships. + +Advantages of Our Approach. Our HyperGLM approach, Multimodal LLMs on a Scene HyperGraph, promotes reasoning about multi-way interactions and high-order relationships. As illustrated in Fig. 2, HyperGLM enhances the model's ability to interpret complex relationships and anticipate intricate video dynamics. Towards this goal, we uniquely integrate entity scene graphs, which is introduced in Sec. 3 to capture spatial interactions between objects in each frame and a procedural graph, which is presented in Sec. 4.1 to model their causal evolution. In our unified approach, hyperedges connect multiple nodes to capture higher-order relationships distinguished from traditional pairwise methods [4, 34, 35, 42]. In addition, the procedural graph enables multi-step transitions for anticipating future interactions or relationships, and reduces bias by generalizing + +![](images/aecdeabde8cae2c808904b822671766915543edf0224f666bbe937c2a4a3b079.jpg) +Figure 2. (a) To model the temporal transition, a simple approach can be using two scene graphs $\mathbf{G}_{\mathbf{t}}$ and $\mathbf{G}_{\mathbf{t} + 1}$ . (b) Another procedure graph can present this temporal modeling. (c) Our unified HyperGraph in Fig. 2c integrates both entity scene graph to capture spatial relationships and the procedural graph to model the temporal evolution. HyperEdge represents person sitting on couch, holding, then playing guitar, whereas holding $\rightarrow$ playing describes a chain of interactions. HyperGraph is presented in 3D. + +infrequent relationship categories. Furthermore, our HyperGLM approach leverages fundamental mathematical properties: permutation equivariance ensures that the HyperGraph structure remains consistent under any permutation of node labels, and invariance to hyperedge order preserves semantic meanings regardless of the node visit order during random walks which is defined in Alg. 1. The theoretical foundations and mathematical properties are detailed in the Appendices. + +# 3. Problem Formulation + +In this section, we define two tasks, including Scene Graph Generation (SGG) and Scene Graph Anticipation (SGA). Graph vertex sequence is represented as $\{V_{\mathbf{G}}\mid 1\leq t\leq T\}$ , where each set of vertex $V_{\mathbf{G}}$ contains object features, bounding boxes, and categories. The scene graph for each frame $t$ , denoted by $\mathbf{G}_t = \big(V_{\mathbf{G}},E_{\mathbf{G}} = \{r^{ij}\mid 1\leq i < j\leq |V_{\mathbf{G}}|\}$ , consists of all pairwise relationships between objects, with $r^{ij}$ representing the relationship category between $v_{i}$ and $v_{j}$ . + +We aim to develop a process $p_{\theta}: (V_{\mathbf{G}} \times V_{\mathbf{G}}) \to r^{ij}$ to predict the relationship $r^{ij}$ between each object pair $(v_t^i, v_t^j)$ in $V_{\mathbf{G}}$ . We define task-specific queries $\mathbf{Q}_{\mathrm{SGG}}$ , and $\mathbf{Q}_{\mathrm{SGA}}$ to direct our unified model to perform on each specific task. The objective for each task is to minimize the negative likelihood of the predicted scene graph $\mathbf{G}_t$ to the truth predicate set $Y_t = \{y^{ij} \mid 1 \leq i < j \leq |V_{\mathbf{G}}|\}$ on categories indexed by $k$ . Scene Graph Generation (SGG) generates the scene graph for each frame $t$ from $t = 1$ to $t = T$ , which is defined as: + +$$ +\min _ {\theta} \mathbb {E} _ {\mathbf {G} _ {t}, Y _ {t}} \left[ - \sum_ {\left(v _ {t} ^ {i, j}\right)} \sum_ {k} \left(y _ {k} ^ {i j} \log p _ {\theta} \left(r ^ {i j} \mid \mathbf {Q} _ {\mathrm {S G G}}\right) _ {k}\right) \right] \tag {1} +$$ + +Scene Graph Anticipation (SGA) anticipates the scene graphs for future frames, generating predictions $\mathbf{G}_{t + n}$ , where $n$ denotes the anticipation horizon, formulated as: + +$$ +\min _ {\theta} \mathbb {E} _ {\mathbf {G} _ {\leq t}, Y _ {t}} \left[ - \sum_ {\left(v _ {t + n} ^ {i, j}\right)} \sum_ {k} \left(y _ {k} ^ {i j} \log p _ {\theta} \left(r ^ {i j} \mid \mathbf {Q} _ {\mathrm {S G A}}\right) _ {k}\right) \right] \tag {2} +$$ + +In Eqns. (1) and (2), $t$ indexes the current frame. While SGG predicts the scene graph at frame $t$ , SGA reads scene graphs up to frame $t$ to forecast the graph in future frame $t + n$ . + +![](images/fda9408f652dfc66d7ebbd26c6062a1f40b9350b6215790cdbe771985ea3814d.jpg) +Figure 3. Our HyperGLM framework comprises an image encoder, MLP projector, temporal aggregator, unified HyperGraph, and language model. It processes video frames by encoding each frame with the image encoder and MLP, extracting spatio-temporal features through image patch grids to generate $N$ spatial tokens per frame. The temporal aggregator compresses the $T \times N$ embeddings over time. The MLP projector then transforms these visual embeddings into the language feature space as frame tokens, interleaved with language tokens, and fed into the Large Language Models. + +# 4. Our Proposed HyperGLM Approach + +In this section, we present our approach, which incorporates a unified HyperGraph into the LLMs, as illustrated in Fig. 3. + +# 4.1. Video Scene (Hyper)Graphs + +Traditional VidSGG methods [4, 34, 42] construct scene graphs $\mathbf{G}_t$ to represent scene entities and their relationships as in Fig. 2a. However, these graph-based approaches do not capture higher-order relationships with temporal dependencies for video understanding [36]. To model this property, we propose a novel HyperGraph-based framework that constructs a unified HyperGraph $\mathcal{H} = (V_{\mathcal{H}}, E_{\mathcal{H}})$ , representing spatial relationships within individual frames and temporal transitions across frames, illustrated in Fig. 4. HyperGraph extends traditional graph structures by allowing hyperedges to connect more than two nodes, making them particularly effective for modeling these higher-order relationships. To leverage this property, our unified HyperGraph (see Fig. 2c) integrates entity scene graphs $\mathbf{G}_t$ for each frame, capturing spatial interactions as introduced in Sec. 3, along with a procedural graph $\mathbf{P}$ that models their causal transitions, as detailed in Sec. 4.1. We unify these components using a random-walk HyperGraph construction, which is defined in Alg. 1 to capture higher-order connectivity patterns and structural semantics, thereby approximating subgraph matching between the entity scene graphs and the procedural graph. This integration allows our approach to model current subjects' interactions and anticipate their future relationships. + +Procedural Graph Construction. SGA objective is to predict the set of relationships $E_{\mathbf{G}}$ in the next frame based on the current set of relationships. To model the temporal evolution of causal relationships between objects across video frames, we introduce a procedural graph $\mathbf{P} = (V_{\mathbf{P}},E_{\mathbf{P}})$ , serving as the temporal counterpart to the entity scene graphs $\mathbf{G}_t$ as shown in Fig. 2b. In particular, the procedural graph $\mathbf{P}$ can vary across datasets, modeling relationship transitions. + +Toward this goal, we model and denote the set of vertices in $\mathbf{P}$ as $V_{\mathbf{P}} = \{r_t^{ij} \in V_{\mathbf{G}} \mid 1 \leq t \leq T\} \neq$ representing distinct relationship categories. The set of edges in $\mathbf{P}$ , i.e., + +![](images/8f8900f6112c23b2fe5219cfb188a8be75705feac1da87af2c44f397671aeef8.jpg) +Figure 4. Our Video Scene HyperGraph, including entity graphs and a procedural graph, as defined in Sec. 4.1. Blue nodes represent entities, while green nodes denote relationships. The entity graph captures spatial relationships (subject $\rightarrow$ relationship $\rightarrow$ object), whereas the procedural graph models relationship transitions $(\rightarrow)$ . Hyperedges are visualized as polygons, encapsulating interactions through chains of relationships. For instance, a hyperedge illustrates a person picking up, holding, opening, and reading a book while sitting on a couch. HyperGraph is presented in 3D, see Supplementary video. + +$E_{\mathbf{P}} = \{(r_m,r_n)\}$ represent possible causal transitions between these relationships, where an edge $(r_m,r_n)$ indicates that relationship $r_m$ causally lead to relationship $r_n$ . We quantify causal transitions by calculating transition probabilities $w(r_m,r_n)$ via their observed frequencies as in Eqn. (3). + +$$ +w \left(r _ {m}, r _ {n}\right) = \frac {\sum_ {t = 1} ^ {T - 1} \sum_ {r _ {t , t + 1} ^ {i j}} \mathbb {1} \left(r _ {t} ^ {i j} = r _ {m} \wedge r _ {t + 1} ^ {i j} = r _ {n}\right)}{\sum_ {t = 1} ^ {T - 1} \sum_ {r _ {t , t + 1} ^ {i j}} \mathbb {1} \left(r _ {t} ^ {i j} = r _ {m}\right)} \tag {3} +$$ + +where $\mathbb{1}(\cdot)$ is the indicator function that counts transitions from relationship $r_m$ at current frame $t$ to $r_n$ at next frame $t + 1$ . Next, self-loops $w(r_m, r_m)$ are removed from the graph, and these probabilities are normalized as in Eqn. (4). + +$$ +\sum_ {r _ {n} \in E _ {\mathbf {P}}} w \left(r _ {m}, r _ {n}\right) = 1 \quad \text {f o r a l l} r _ {m} \in E _ {\mathbf {P}} \tag {4} +$$ + +By leveraging these probabilities, the procedural graph $\mathbf{P}$ enhances the prediction of future relationships. For each relationship $r_t^{ij}$ in frame $t$ , the most probable relationship $r_{t + 1}^{ij}$ in the next frame $t + 1$ is determined as in Eqn. (5). + +$$ +r _ {t + 1} ^ {i j} = \arg \max _ {r _ {n}} P \left(r _ {n} \mid r _ {t} ^ {i j}, v ^ {i, j}\right) \tag {5} +$$ + +where $P(r_{n} \mid r_{t}^{ij}, v^{i,j}) = w(r_{t}^{ij}, r_{n}) \times v^{i,j}$ is the probability of transitioning from relationship $r_{t}^{ij}$ to $r_{n}$ , looking at object features. This allows the model to anticipate future interactions based on established temporal patterns. + +HyperGraph Construction. To incorporate spatial relationships between objects in each frame and their causal temporal transitions, we construct a unified HyperGraph $\mathcal{H}$ that integrates the entity scene graphs $\mathbf{G}_t$ and the procedural graph $\mathbf{P}$ . Specifically, a HyperGraph is an augmented + +representation combining the entity scene graphs with the procedural graph. This unified structure merges spatial and temporal relationships into a single graph, enabling the use of conventional graph algorithms while preserving the complex interactions captured by the HyperGraph. Mathematically, a unified HyperGraph $\mathcal{H} = (V_{\mathcal{H}}, E_{\mathcal{H}})$ is defined as in Eqn. (6). + +$$ +\mathcal {H} = \left(\bigcup_ {t = 1} ^ {T} V _ {\mathbf {G} _ {t}} \cup V _ {\mathbf {P}}, \quad \bigcup_ {t = 1} ^ {T} E _ {\mathbf {G} _ {t}} \cup E _ {\mathbf {P}}\right) \tag {6} +$$ + +where $V_{\mathcal{H}}$ includes all entity nodes $v_{t}^{i}$ from each $\mathbf{G}_t$ and the relationship type nodes $V_{\mathbf{P}}$ . The hyperedge set $E_{\mathcal{H}}$ includes pairwise relationships $E_{\mathbf{G}_t}$ within each $\mathbf{G}_t$ , capturing spatial relationships and temporal transition edges $E_{\mathbf{P}}$ from $\mathbf{P}$ , modeling the evolution of causal relationships across frames. + +Random-walk Algorithm. We employ the random walks algorithm outlined in Alg. 1 to sample representative substructures from the unified HyperGraph $\mathcal{H}$ , which captures connectivity patterns and mitigates the NP-hardness of exact subgraph matching. Specifically, these walks alternate between nodes and hyperedges, preserving the multi-node connections intrinsic to hyperedges and capturing the complexity of multi-object relationships and their transition across video frames. In each walk, a hyperedge $h_i$ aggregates the visited nodes, thereby encapsulating higher-order relationships. For example, in the entity scene graph $\mathbf{G}_t$ , a walk might traverse from a "person" to a "cup" via the "holding", resulting in the hyperedge $h_i = \{\text{person}, \text{cup}\}$ . Similarly, within the procedural graph $\mathbf{P}$ , a walk might transition through a sequence of interactions such as "holding", "placing", and "releasing", forming the hyperedge $h_i = \{\text{holding}, \text{placing}, \text{releasing}\}$ . Therefore, we generate sampled hyperedges $E_{\text{sampled}} = \{h_i \mid i = 1, \dots, N_w\}$ as in L25 of Alg. 1 by conducting multiple random walks. + +Algorithm 1 Random-walk for HyperGraph Construction. +Require: $\mathcal{H} = (V_{\mathcal{H}},E_{\mathcal{H}})$ , Number of Walks $N_{w}$ , Walk Length $N_{l}$ +Ensure: $\mathcal{H}' = (V_{\mathcal{H}},E_{\mathcal{H}}^{\prime})$ +1: Initialize $E_{\mathrm{sampled}}\gets \emptyset$ +2: for $i = 1$ to $N_{w}$ do +3: Select $v_{\mathrm{start}}\in V_{\mathcal{H}}$ uniformly at random +4: Initialize walk sequence $S\leftarrow [v_{\mathrm{start}}]$ +5: for $j = 1$ to $N_{l}$ do +6: if $j$ is odd then Node to HyperEdge +7: $v_{\mathrm{current}}\gets S[j]$ +8: $\mathcal{E}_{v_{\mathrm{current}}}\gets \{h\in E_{\mathcal{H}}\mid v_{\mathrm{current}}\in h\}$ +9: if $\mathcal{E}_{v_{\mathrm{current}}}\neq \emptyset$ then +10: Select $h_j\in \mathcal{E}_{v_{\mathrm{current}}}$ +11: Append $h_j$ to $S$ +12: end if +13: else HyperEdge to Node +14: $h_\mathrm{current}\gets S[j]$ +15: $V_{h_{\mathrm{current}}}\gets h_{\mathrm{current}}$ +16: Select $v_{j}\in V_{h_{\mathrm{current}}}$ +17: Append $v_{j}$ to $S$ +18: end if +19: end for +20: $h_i\gets \{v\in S\mid v\in V_H\}$ ▷Form new HyperEdge +21: if $h_i\notin E_{\mathcal{H}}\cup E_{\mathrm{sampled}}$ then +22: $E_{\mathrm{sampled}}\gets E_{\mathrm{sampled}}\cup \{h_i\}$ +23: end if +24: end for +25: $E_{\mathcal{H}}'\gets E_{\mathcal{H}}\cup E_{\mathrm{sampled}}$ +26: return $\mathcal{H}' = (V_{\mathcal{H}},E_{\mathcal{H}}')$ + +# 4.2. Multimodal LLMs on HyperGraphs + +Formulation. Given an input video $\mathbf{V}$ and a task query $\mathbf{Q}$ , we aim to generate a target answer sequence $\mathbf{A}$ of length $L$ . Especially, by modeling a HyperGraph $\mathcal{H}$ , the target answer $\mathbf{A}$ is generated via a process illustrated in Fig. 3, defined as: + +$$ +p (\mathbf {A} | \mathbf {V}, \mathbf {Q}, \mathcal {H}) = \prod_ {i = 0} ^ {L - 1} p \left(x _ {i} | \mathbf {V}, \mathbf {Q}, \mathcal {H}, x _ {< i}\right) \tag {7} +$$ + +where $x_{< i}$ denotes the preceding token sequence, and $\mathbf{A} = [x_0,\dots ,x_i,\dots ,x_{L - 1}]$ is the sequence of answer tokens reasoning video $\mathbf{V}$ by question $\mathbf{Q}$ and HyperGraph $\mathcal{H}$ in Fig. 4. + +As illustrated in Fig. 5, our process begins with the generation, constructing an entity scene graph $\mathbf{G}_t = (V_t^e, E_{\mathbf{G}})$ for each frame $t$ to capture detected objects and their relationships. Relationship anticipation employs the procedural graph $\mathbf{P}$ to model temporal evolution and predict future interactions. The model then engages in reasoning using the HyperGraph $\mathcal{H}$ and verification through video captioning to ensure contextual relevance and accuracy. It refines understanding through clarification, generating answers by reasoning over the video and $\mathcal{H}$ . In scene forecasting, it predicts + +Table 1. Comparisons of video scene graph datasets. SGG, SGA, VQA, VC, and RR represent Scene Graph Generation, Scene Graph Anticipation, Video Question Answering, Video Captioning, and Relation Reasoning. + +
Dataset#FramesTasksAnnotations
SGGSGAVQAVCRRBboxRelationText
SportsHHI [47]11.4K✓XXXX✓✓X
Action Genome [18]234.3K✓✓XXX✓✓X
AeroEye [35]261.5K✓XXXX✓✓X
ASPIRe [34]1.6M✓XXXX✓✓X
PVSG [52]153K✓X✓✓X✓✓✓
VSGR (Ours)1.9M✓✓✓✓✓✓✓✓
+ +subsequent interactions by identifying new relationships and adapting to evolving contexts using transition probabilities in $\mathbf{P}$ , anticipating future relationships $r_{t+n}^{ij}$ and improving conversational flow. Finally, the hypothesis allows the model to propose and test conjectures about the underlying themes or intentions, synthesizing information from previous steps and leveraging $\mathcal{H}$ to explain patterns and forecast results. + +# 5. Video Scene Graph Reasoning Dataset + +Limitations of Current Datasets. Existing benchmarks (see Table 1) primarily support SGG and SGA, limiting their applicability to reasoning tasks. They suffer from inadequate support for VQA, VC, and RR, shallow annotations that fail to capture intricate object interactions, insufficient modeling of temporal dynamics, and poor multimodal integration. Thus, our VSGR dataset supports SGG, SGA, VQA, VC, and RR, enabling the reasoning capabilities of LLMs. + +# 5.1. Dataset Construction + +Data Acquisition Stage. We source videos from the ASPIRe [34] and AeroEye [35] datasets. While the ASPIRe dataset offers diverse, richly annotated videos emphasizing dynamic interactions and temporal changes, the AeroEye dataset offers drone-captured footage across various scenes. Comprehension Tasks via Question-Answering. We introduce tasks that leverage fine-grained relationships from scene graphs, extending Scene Graph Generation to focus on relation understanding and subject/object interpretation using triplets and leverage GPT-4/GPT-3.5 model for language generation. + +Video Captioning (VC). We generate 82,532 video-caption pairs, resulting in about 22 captions per video. Our process involves (1) extracting triplets from cropped video frames focusing on foreground objects, (2) generating background descriptions based on these triplets, and (3) combining the foreground triplets and background descriptions to produce captions. The average length is 893 characters, surpassing the PVSG [52] dataset in quantity and detail. + +Video Question Answering (VQA). We develop 74,856 question-answer pairs by designing questions that explore diverse relationships, selecting subject and object categories to verify specific relationships, or choosing triplets to assess their uniqueness. This results in an average of approximately 20 questions per video, which exceeds existing benchmarks + +![](images/5a024d78c7f75b4f89f88278cb1bfc3c4464f8cca8a020c22486d42daafce9a5.jpg) +Figure 5. An example of the diversified context within the streaming dialog in our VSGR dataset. Best viewed in color and zooming in. + +# Step-1: Generation + +Given bounding boxes: person-1 [657, 116, 155, 577], person-2 [425, 211, 178, 450], ball-1 [534, 437, 42, 54], ball-2 [670, 350, 37, 65]. What are the relationships between these objects? + +The video depicts including person-1 towards the right side and closer to ball-2. person-2 is situated to the left and is nearer to ball-1. + +# Step-2: Relationship Anticipation + +What interactions might occur between the objects in the next three frame based on ? + +The person-1 will interact with ball-2 by picking it up or dribbling it. Similarly, person-2 may engage with ball-1 corresponding to + +# Step-3: Reasoning + +What purpose or context can be inferred within + +person-1 and person-2 work with ball-1 and ball-2, suggesting drills like dribbling or shooting. Their separation implies independent practice, but their proximity hints at a possible shift to coordinated team drills later in the session + +# Step-4: Verification + +Does person-1 move towards ball-2 and begin dribbling, and does person-2 start practicing shots with ball-1 through ? + +Yes, both individuals are focused on separate balls. person-1 continues with ball-1, and person-2 remains with ball-2, indicating individual practice rather than switching activities or interacting with each other. + +# Step-5: Clarification + +If instead, person-1 picks up ball-2 and passes it to person-2, how would that change the dynamics of? + +This would shift the dynamics from individual practice to a collaborative activity, with person-1 passing ball-2 to person-2. The interaction would signal the start of coordinated play, possibly initiating a game involving both ball-1 and ball-2 and highlighting teamwork between person-1 and person-2. + +# Step-6: Scene Forecasting + +Given , what are the likely relationship changes between objects in the upcoming phase? + +The next action could involve person-1 passing ball-2 to person-2 while person-2 passes ball-1 to person-1. Alternatively, person-1 might dribble ball-2 toward person-2, initiating a one-on-one game. + +# Step-7: Hypothesis + +What if person-3 [600, 200, 160, 500] enters and approaches the others? How might that affect the interactions? + +person-1, person-2, and person-3 could engage in coordinated actions, such as three-person passing drills using ball-1 and ball-2. The interaction would evolve from individual practice to collaborative teamwork. + +such as the MSRVTT-QA [49] and MSVD-QA [48] datasets. + +Relation Reasoning (RR). Using the annotated scene graphs, we produce 61,120 relation reasoning tasks by selecting partial information as an incomplete input. Each task requires the model to deduce relationships among entities and identify the categories of the subject and object. With an average of approximately 16 tasks per video, our VSGR dataset provides a substantial collection for evaluating models' abilities in relational reasoning and scene understanding. Question and Answer Validation. To ensure the quality and complexity of questions, we implement a rigorous validation process that combines generation by LLMs with human refinement. Initially, GPT-4/GPT-3.5 generates queries based on the triplets extracted from the videos. Human annotators are trained with specific guidelines that emphasize clarity, relevance, and appropriate challenge levels, then review and refine questions. They enhance the questions by ensuring they are directly answerable from the video content and require careful reasoning about the depicted interactions and relationships. + +We apply strict filtering criteria to improve the quality of the dataset. First, we eliminate questions that do not require video context and can be answered using general world knowledge, ensuring that models must rely on visual information from the videos. Second, we exclude questions that LLMs can answer correctly, increasing the challenge and utility of the dataset in evaluating advanced reasoning abilities. Finally, independent annotators perform a second round to review and evaluate the quality of the refined questions. + +# 5.2. Dataset Comparison + +As reported in Table 1, our VSGR dataset represents a substantial advancement in video scene graph benchmarks. Our dataset comprises 3,748 videos and 1,841,243 frames, surpassing existing datasets in scale. Unlike other datasets that address only a limited subset of tasks, our dataset offers comprehensive task coverage, facilitating multifaceted evaluations of LLMs. In addition, our ground truth enriches relation annotations with comprehensive textual descriptions, enabling sophisticated reasoning and relationship predictions, as illustrated in Fig. 5. Additionally, our VSGR dataset incorporates diverse viewpoints, including third-person, ego-centric, and drone perspectives, enhancing its generalization. + +# 6. Experiment Results + +# 6.1. Implementation Details + +Datasets. We leverage our VSGR dataset across five tasks. In addition, we utilize the PVSG [52] dataset for the SGG task and the Action Genome [18] dataset for the SGA task. + +Model Configuration. We operate the CLIP-ViT-L-336 [8, 37] to encode each video frame into ten tokens (one CLS token and nine from $3 \times 3$ average pooling). These tokens are fed into a two-layer MLP connector to the Mistral7B-Instruct [19] language model. For training, we apply LoRA [16] to all linear layers with a rank of 128 and a scaling factor of 256, omitting vision-language alignment [31]. We train for two epochs with a batch size of 128 over 16 + +![](images/d0a01bfd9fa68d49d6cdd8ac7d69f94218428a12f7d82e080192790acb0992a8.jpg) +Figure 6. Comparison of Recall (R) and mean Recall (mR) at different numbers of hyperedges for the SGG task on the VSGR dataset. + +![](images/046d8a22a37e1a284f7f5f8a29220ae77a2315b14a16d2f5c3501b83a8f87b6d.jpg) + +Table 2. Comparison (\%) on the VSGR and Action Genome datasets for the Scene Graph Anticipation (SGA) task at varying video input fractions $\mathcal{F}$ . + +
FMethodAction GenomeVSGR
R/mR@10R/mR@20R/mR@50R/mR@10R/mR@20R/mR@50
0.3STTran+ [5]13.9 / 3.521.6 / 7.340.8 / 20.312.0 / 4.019.0 / 8.035.7 / 18.0
DSGDetr+ [11]14.3 / 3.621.8 / 7.641.3 / 21.212.5 / 4.219.5 / 8.336.0 / 19.0
STTran++ [5]15.4 / 6.227.2 / 14.148.6 / 32.214.0 / 6.522.0 / 11.039.0 / 25.1
DSGDetr++ [11]16.8 / 8.429.0 / 16.748.9 / 32.314.5 / 7.022.5 / 12.040.0 / 26.0
SceneSayerODE [36]23.3 / 13.332.5 / 20.145.1 / 33.018.0 / 9.526.0 / 15.042.0 / 30.0
SceneSayerSDE [36]25.9 / 15.635.7 / 23.147.4 / 37.119.5 / 11.027.5 / 17.044.0 / 33.5
HyperGraph (Ours)26.5 / 14.836.2 / 22.149.3 / 37.218.8 / 9.327.2 / 16.342.5 / 27.4
HyperGLM (Ours)27.5 / 15.837.0 / 24.550.0 / 38.019.0 / 10.028.0 / 16.543.0 / 27.5
0.5STTran+ [5]14.9 / 3.722.6 / 7.642.9 / 21.413.5 / 4.221.0 / 8.537.0 / 19.0
DSGDetr+ [11]15.2 / 3.923.1 / 8.043.3 / 22.213.8 / 4.521.5 / 9.038.0 / 20.0
STTran++ [5]16.6 / 6.629.1 / 14.751.5 / 33.415.5 / 7.023.5 / 12.541.0 / 26.5
DSGDetr++ [11]17.4 / 8.430.5 / 17.051.9 / 33.916.0 / 7.524.0 / 13.042.0 / 28.0
SceneSayerODE [36]26.4 / 14.336.6 / 21.449.8 / 36.020.5 / 11.029.5 / 16.546.1 / 32.5
SceneSayerSDE [36]28.4 / 16.338.6 / 25.151.4 / 39.921.5 / 12.531.0 / 18.548.0 / 35.7
HyperGraph (Ours)29.2 / 16.439.3 / 23.252.1 / 38.720.3 / 12.229.8 / 17.146.2 / 33.1
HyperGLM (Ours)30.0 / 17.040.5 / 27.553.5 / 40.521.5 / 11.531.5 / 18.546.5 / 30.0
0.7STTran+ [5]16.6 / 4.225.1 / 8.547.2 / 24.015.0 / 5.024.0 / 10.541.0 / 21.5
DSGDetr+ [11]16.8 / 4.325.3 / 8.847.4 / 24.715.5 / 5.224.5 / 11.042.0 / 22.0
STTran++ [5]19.0 / 7.732.8 / 17.156.8 / 36.817.0 / 8.027.0 / 14.045.0 / 29.0
DSGDetr++ [11]19.8 / 9.534.1 / 19.256.7 / 37.217.5 / 8.528.0 / 15.046.1 / 30.0
SceneSayerODE [36]32.1 / 16.542.8 / 24.455.6 / 39.623.5 / 13.033.5 / 19.051.0 / 36.0
SceneSayerSDE [36]33.3 / 18.144.0 / 27.356.4 / 44.424.5 / 14.535.7 / 21.053.0 / 38.0
HyperGraph (Ours)34.3 / 19.245.2 / 25.357.2 / 42.122.3 / 13.234.2 / 19.350.4 / 35.3
HyperGLM (Ours)35.7 / 19.546.1 / 30.058.2 / 44.025.1 / 13.535.5 / 21.551.0 / 33.5
+ +iterations on $4 \times$ GPUs, taking approximately six hours. + +Metrics. We evaluate the SGG and SGA tasks using the Recall and mean Recall scores. In addition, we evaluate the VQA and RR tasks by Accuracy, Precision, Recall, and F1 scores. For the VC task, we utilize CIDEr, MENTOR, ROUGE-L, and BLEU-4 scores to validate our performance. Settings. For the SGG and SGA tasks, we adopt the evaluation settings based on [18, 36]. In particular, the model is provided with raw video frames and must detect objects using a pre-trained detector (i.e. Faster R-CNN) and predict or anticipate their relationships. Especially for the SGA, we set the initial video input fraction $(\mathcal{F})$ to 0.9, following [36]. + +# 6.2. Ablation Study + +Hypergraph Parameters. The number of walks $(N_w)$ and walk length $(N_l)$ are introduced in Alg. 1, directly impacting the capacity to capture high-order relationships by determining the number of hyperedges. Although more hyperedges increase relational diversity, they can also introduce redundancy beyond an optimal point. Specifically, a higher $N_w$ broadens the range of sampled relationships, while a moderate $N_l$ balances depth. Our experiments demonstrate that optimal performances are achieved with 60 hyperedges $(N_w = 60$ and $N_l = 7)$ , shown in Fig. 6. Further experiments on the SGG task are provided in Table 3, and experiments on these parameters are included in the Appendices. Video Input Fraction. We adjust the initial video input + +Table 3. Comparison $(\%)$ on the VSGR and PVSG datasets for the Scene Graph Generation (SGG) task at Recall (R) and mean Recall (mR). + +
MethodPVSGVSGR
R/mR@20R/mR@50R/mR@100R/mR@20R/mR@50R/mR@100
Transformer [52]4.0 / 1.84.4 / 1.94.9 / 2.025.7 / 6.334.5 / 6.543.5 / 7.0
HIG [34]4.6 / 1.94.9 / 2.15.4 / 2.223.8 / 5.731.1 / 5.940.4 / 6.9
CYCLO [35]5.8 / 2.06.1 / 2.26.7 / 2.329.4 / 7.136.4 / 7.747.7 / 7.7
HyperGraph (Ours)6.5 / 2.27.0 / 2.47.5 / 2.631.6 / 7.838.8 / 8.350.3 / 8.5
HyperGLM (Ours)7.5 / 2.88.1 / 3.78.5 / 3.935.8 / 9.242.3 / 10.154.7 / 10.4
+ +Table 4. Comparison (%) on the VSGR and Action Genome datasets for the Scene Graph Anticipation (SGA) task at Recall (R) and mean Recall (mR). + +
MethodAction GenomeVSGR
R/mR@10R/mR@20R/mR@50R/mR@10R/mR@20R/mR@50
STTran+ [5]17.5 / 4.626.8 / 9.249.6 / 24.316.5 / 5.526.0 / 11.043.0 / 23.0
DSGDetr+ [11]17.9 / 4.727.7 / 9.751.4 / 25.917.0 / 6.027.0 / 11.544.5 / 24.5
STTran++ [5]20.2 / 8.935.7 / 18.460.2 / 38.819.0 / 9.530.0 / 15.549.0 / 31.0
DSGDetr++ [11]22.2 / 11.437.1 / 21.061.0 / 39.519.5 / 10.031.0 / 16.050.0 / 32.5
SceneSayerODE [36]36.6 / 17.848.3 / 27.461.3 / 43.426.5 / 14.037.5 / 20.055.5 / 38.0
SceneSayerSDE [36]37.3 / 20.848.6 / 30.961.6 / 46.827.5 / 16.038.5 / 22.058.2 / 40.0
HyperGraph (Ours)37.5 / 19.149.3 / 31.462.3 / 47.428.4 / 17.539.3 / 22.457.6 / 41.5
HyperGLM (Ours)38.8 / 22.351.5 / 33.065.2 / 48.630.2 / 18.141.1 / 23.559.3 / 43.4
+ +fraction, $\mathcal{F}$ , to 0.3, 0.5, and 0.7 for the SGA task. This adjustment allows the model to learn from varying observed portions and predict the unseen segment. Table 2 indicates that increasing the portion of the seen video improves performance, suggesting that additional visual context is beneficial. In addition, Table 4 further confirms that the default setting at a higher input fraction ( $F = 0.9$ ) leads to optimal performance. We also present additional settings with varying video input fractions for the SGA task in the Appendices. + +# 6.3. Comparison with State-of-the-Arts + +Table 3 demonstrates that HyperGLM significantly outperforms existing methods on the PVSG and VSGR datasets, achieving the highest $R@20$ scores of $7.5\%$ and $35.8\%$ , respectively. By leveraging hyperedges connecting multiple nodes within the HyperGraph, HyperGLM effectively captures complex object interactions and spatial dependencies, transcending traditional pairwise methods to generate more accurate and detailed scene graph representations. Notably, integrating a procedural graph within the HyperGraph reduces bias. Our approach enhances mean Recall, reaching improvements of $2.8\%$ on the PVSG dataset and $9.2\%$ on the VSGR dataset, thereby addressing the long-tail distribution challenges that have struggled in previous methods [34, 35, 52]. Furthermore, the LLM enhances the capacity of HyperGLM to infer and predict intricate relationships embedded within the unified HyperGraph, resulting in improved performance compared to HyperGraph, which shows a decrease of $4.2\%$ at $R@20$ on the VSGR dataset. + +As shown in Table 4, our HyperGLM approach outperforms existing SGA methods on the Action Genome and VSGR datasets, achieving R@10 scores of $35.7\%$ and $25.1\%$ , respectively. This improvement stems from integrating procedural graphs that model causal relationships within the HyperGraph structure. In contrast, the SceneSayer [36] method relies on NeuralODE and NeuralSDE to capture the latent dynamics of object interaction evolution. Our procedural graphs enable multi-step transitions that explicitly + +![](images/c3a87d810387a52aaf9465f5d76b7bd91f7cb2b28781fde1fd2757228c01145d.jpg) + +![](images/2e9f9ce2e537e4124e4e155635084f5750b6183684a70cba399a13624e2c0002.jpg) + +![](images/3389de1b55da6d5d50ae1f3cccb1c8e11ba6b50c06de7816a66ea78b023b6009.jpg) +a) Scene Graph Generation + +![](images/a4edd7237696b72899c03efdbebcea8fcf847cef256b9ec964a8af052a749164.jpg) +b) Scene Graph Anticipation +Figure 7. Qualitative comparison of our HyperGLM approach versus SceneSayerSDE [36] for SGG and SGA. Red and green edge labels denote incorrect and correct predictions, respectively. Our HyperGLM approach effectively captures the evolving interactions between person-1 and bicycle-1 or person-2 and person-3 over time and anticipates interactions in unseen video frames, while SceneSayerSDE confuses to similar predicates. Best viewed in color. + +Table 5. Comparison (%) on VSGR for the Video Question Answering. + +
MethodAccuracyPrecisionRecallF1 Score
Video-ChatGPT [32]33.235.132.333.6
Video-LLaVA-7B [30]43.143.841.742.7
MovieChat [41]43.544.242.643.4
Chat-UniVi-7B [20]44.345.643.244.4
HyperGLM (Ours)45.447.244.345.7
+ +Table 6. Comparison (%) on VSGR for the Video Captioning. + +
MethodCIDErMENTORROUGE-LBLEU-4
MV-GPT [38]57.137.562.547.2
CoCap [40]54.329.461.842.8
UniVL + MELTR [24]50.528.160.042.1
HyperGLM (Ours)54.530.764.948.8
+ +capture the temporal evolution of relationships by modeling sequential interactions and their dependencies. In contrast, NeuralODE and NeuralSDE primarily focus on continuous-time dynamics, which may limit their effectiveness in handling discrete, multi-step relational changes as illustrated in Fig. 7b. Significantly, the HyperGraph is injected into the LLM, allowing the model to capture complex temporal patterns, resulting in better results than the HyperGraph. + +In addition to the improvements in SGG and SGA shown in Fig. 7, Tables 5, 6, and 7 demonstrate the significant improvements using our HyperGLM approach in the VQA, VC, and RR tasks. In VQA, the hyperedges connect multiple objects, enabling HyperGLM to capture context-rich interactions more effectively, reaching an accuracy of $45.4\%$ . For VC, the HyperGraph structure supports evolving object connections over time, allowing our HyperGLM approach to generate captions that describe the relationship between objects and capture the temporal flow of interactions coher- + +Table 7. Comparison (%) on VSGR for the Relation Reasoning. + +
MethodAccuracyPrecisionRecallF1 Score
Video-LLaVA-7B [30]41.342.540.241.3
MA-LMM [14]42.843.741.842.7
LLaMA-VID-7B [29]44.145.243.544.3
HyperGLM (Ours)47.248.446.547.4
+ +ently, achieving scores of $54.5\%$ at CIDEr and $64.9\%$ at ROUGE-L. In RR task, the HyperGraph effectively manages intricate dependencies between multiple objects, resulting in precise relational inferences with an accuracy of $47.2\%$ . + +# 7. Conclusion + +In this paper, we have introduced HyperGLM, a novel VidSGG method that integrates scene hypergraph information into LLMs for context-aware and precise scene interpretation. Our approach effectively models complex interactions and higher-order relationships. It outperforms leading methods on benchmarks, including PVSG, Action Genome, and our newly collected VSGR dataset across five tasks. + +Limitations. Although our HyperGLM approach effectively models multi-way interactions, managing many objects and their interactions can complicate relationship structures. As the HyperGraph expands, essential relationships may become obscured, reducing the clarity of scene interpretation. + +Acknowledgment. This material is based upon work supported by the National Science Foundation under Award No. OIA-1946391. We also acknowledge the Arkansas High-Performance Computing Center for providing GPUs. + +# References + +[1] Song Bai, Feihu Zhang, and Philip HS Torr. Hypergraph convolution and hypergraph attention. Pattern Recognition, 110:107637, 2021. 2 +[2] Joya Chen, Zhaoyang Lv, Shiwei Wu, Kevin Qinghong Lin, Chenan Song, Difei Gao, Jia-Wei Liu, Ziteng Gao, Dongxing Mao, and Mike Zheng Shou. Videollm-online: Online video large language model for streaming video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18407-18418, 2024. 1 +[3] Zhanwen Chen, Saed Rezayi, and Sheng Li. More knowledge, less bias: Unbiasing scene graph generation with explicit ontological adjustment. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4023-4032, 2023. 2 +[4] Yuren Cong, Wentong Liao, Hanno Ackermann, Bodo Rosenhahn, and Michael Ying Yang. Spatial-temporal transformer for dynamic scene graph generation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16372-16382, 2021. 1, 2, 3 +[5] Yuren Cong, Wentong Liao, H. Ackermann, M. Yang, and B. Rosenhahn. Spatial-temporal transformer for dynamic scene graph generation. IEEE International Conference on Computer Vision, 2021. 7 +[6] Yuren Cong, Michael Ying Yang, and Bodo Rosenhahn. Reltr: Relation transformer for scene graph generation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(9):11169-11183, 2023. 2 +[7] Youming Deng, Yansheng Li, Yongjun Zhang, Xiang Xiang, Jian Wang, Jingdong Chen, and Jiayi Ma. Hierarchical memory learning for fine-grained scene graph generation. In European Conference on Computer Vision, pages 266-283. Springer, 2022. 2 +[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 6 +[9] Azade Farshad, Yousef Yeganeh, Yu Chi, Chengzhi Shen, Böjrn Ommer, and Nassir Navab. Scenegenie: Scene graph guided diffusion models for image synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 88-98, 2023. 2 +[10] Hao Fei, Shengqiong Wu, Wei Ji, Hanwang Zhang, Meishan Zhang, Mong-Li Lee, and Wynne Hsu. Video-of-thought: Step-by-step video reasoning from perception to cognition. In *Forty-first International Conference on Machine Learning*, 2024. 1 +[11] Shengyu Feng, Hesham Mostafa, Marcel Nassar, Somdeb Majumdar, and Subarna Tripathi. Exploiting long-term dependencies for generating dynamic scene graphs. IEEE Workshop/Winter Conference on Applications of Computer Vision, 2021. 7 +[12] Yue Gao, Yifan Feng, Shuyi Ji, and Rongrong Ji. Hgnn+: General hypergraph neural networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):3181-3199, + +2022. 2 +[13] Zeeshan Hayden and Xuming He. Dsgg: Dense relation transformer for an end-to-end scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28317-28326, 2024. 2 +[14] Bo He, Hengduo Li, Young Kyun Jang, Menglin Jia, Xuefei Cao, Ashish Shah, Abhinav Shrivastava, and Ser-Nam Lim. Ma-lmm: Memory-augmented large multimodal model for long-term video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13504-13514, 2024. 8 +[15] Tao He, Lianli Gao, Jingkuan Song, and Yuan-Fang Li. Towards open-vocabulary scene graph generation with prompt-based finetuning. In European Conference on Computer Vision, pages 56-73. Springer, 2022. 2 +[16] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 6 +[17] Jinbae Im, Jeong Yeon Nam, Nokyung Park, Hyungmin Lee, and Seunghyun Park. Egtr: Extracting graph from transformer for scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24229-24238, 2024. 2 +[18] Jingwei Ji, Ranjay Krishna, Li Fei-Fei, and Juan Carlos Niebles. Action genome: Actions as compositions of spatiotemporal scene graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10236–10247, 2020. 1, 2, 5, 6, 7 +[19] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 6 +[20] Peng Jin, Ryuichi Takanobu, Wancai Zhang, Xiaochun Cao, and Li Yuan. Chat-univi: Unified visual representation empowers large language models with image and video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13700-13710, 2024. 1, 8 +[21] Tianlei Jin, Fangtai Guo, Qiwei Meng, Shiqiang Zhu, Xiangming Xi, Wen Wang, Zonghao Mu, and Wei Song. Fast contextual scene graph generation with unbiased context augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6302-6311, 2023. 2 +[22] Eun-Sol Kim, Woo Young Kang, Kyoung-Woon On, Yu-Jung Heo, and Byoung-Tak Zhang. Hypergraph attention networks for multimodal learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14581-14590, 2020. 2 +[23] Kibum Kim, Kanghoon Yoon, Jaehyeong Jeon, Yeunjun In, Jinyoung Moon, Donghyun Kim, and Chanyoung Park. Llm4sgg: Large language models for weakly supervised scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28306-28316, 2024. 2 +[24] Dohwan Ko, Joonmyung Choi, Hyeong Kyu Choi, Kyoung- + +Woon On, Byungseok Roh, and Hyunwoo J Kim. Meltr: Meta loss transformer for learning to fine-tune video foundation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20105-20115, 2023. 1, 8 +[25] Sanjoy Kundu and Sathyanarayanan N Aakur. Is-ggt: Iterative scene graph generation with generative transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6292-6301, 2023. 2 +[26] Rongjie Li, Songyang Zhang, and Xuming He. Sgtr: End-to-end scene graph generation with transformer. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19486-19496, 2022. 2 +[27] Rongjie Li, Songyang Zhang, Dahua Lin, Kai Chen, and Xuming He. From pixels to graphs: Open-vocabulary scene graph generation with vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28076-28086, 2024. 2 +[28] Wanxin Li, Wei Xie, Zhigang Tu, Wei Wang, and Lianghao Jin. Multi-hyperedge hypergraph for group activity recognition. In 2022 International Joint Conference on Neural Networks (IJCNN), pages 01-07. IEEE, 2022. 2 +[29] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In European Conference on Computer Vision, pages 323–340. Springer, 2025. 8 +[30] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 1, 8 +[31] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 6 +[32] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Khan. Video-ChatGPT: Towards detailed video understanding via large vision and language models. In Lun-Wei Ku, Andre Martins, and Vivek Srikumar, editors, Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 12585–12602, Bangkok, Thailand, Aug. 2024. Association for Computational Linguistics. 1, 8 +[33] Sayak Nag, Kyle Min, Subarna Tripathi, and Amit K Roy-Chowdhury. Unbiased scene graph generation in videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22803-22813, 2023. 2 +[34] Trong-Thuan Nguyen, Pha Nguyen, and Khoa Luu. Hig: Hierarchical interlacement graph approach to scene graph generation in video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 1, 2, 3, 5, 7 +[35] Trong-Thuan Nguyen, Pha Nguyen, Li Xin, Cothren Jackson, Yilmaz Alper, and Khoa Luu. CYCLO: Cyclic graph transformer approach to multi-object relationship modeling in aerial videos. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 1, 2, 5, 7 +[36] Rohith Peddi, Saksham Singh, Parag Singla, Vibhav Gogate, et al. Towards scene graph anticipation. In European Conference on Computer Vision. Springer, 2024. 1, 2, 3, 7, 8 +[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya + +Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 6 +[38] Paul Hongsuck Seo, Arsha Nagrani, Anurag Arnab, and Cordelia Schmid. End-to-end generative pretraining for multimodal video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17959-17968, 2022. 1, 8 +[39] Xindi Shang, Yicong Li, Junbin Xiao, Wei Ji, and Tat-Seng Chua. Video visual relation detection via iterative inference. In Proceedings of the 29th ACM international conference on Multimedia, pages 3654-3663, 2021. 2 +[40] Yaojie Shen, Xin Gu, Kai Xu, Heng Fan, Longyin Wen, and Libo Zhang. Accurate and fast compressed video captioning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15558-15567, 2023. 1, 8 +[41] Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Haozhe Chi, Xun Guo, Tian Ye, Yanting Zhang, et al. Moviechat: From dense token to sparse memory for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18221-18232, 2024. 1, 8 +[42] Yao Teng, Limin Wang, Zhifeng Li, and Gangshan Wu. Target adaptive context aggregation for video scene graph generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13688-13697, 2021. 1, 2, 3 +[43] Nupur Thakur, PrasanthSai Gouripeddi, and Baoxin Li. Graph (graph): A nested graph-based framework for early accident anticipation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 7533-7541, 2024. 1, 2 +[44] Aisha Urooj, Hilde Kuehne, Bo Wu, Kim Chheu, Walid BousSELham, Chuang Gan, Niels Lobo, and Mubarak Shah. Learning situation hyper-graphs for video question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14879-14889, 2023. 2 +[45] Jiahao Wang, Guo Chen, Yifei Huang, Limin Wang, and Tong Lu. Memory-and-anticipation transformer for online action understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13824–13835, 2023. 1 +[46] Yanan Wang, Shuichiro Haruta, Donghuo Zeng, Julio Vizcarra, and Mori Kurokawa. Multi-object event graph representation learning for video question answering. In *Meeting on Image Recognition and Understanding*, 2024. 2 +[47] Tao Wu, Runyu He, Gangshan Wu, and Limin Wang. Sportshhi: A dataset for human-human interaction detection in sports videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2, 5 +[48] Dejing Xu, Zhou Zhao, Jun Xiao, Fei Wu, Hanwang Zhang, Xiangnan He, and Yueting Zhuang. Video question answering via gradually refined attention over appearance and motion. In Proceedings of the 25th ACM international conference on Multimedia, pages 1645-1653, 2017. 6 +[49] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In Proceedings of the IEEE conference on computer vision and + +pattern recognition, pages 5288-5296, 2016. 6 + +[50] Zenan Xu, Wanjun Zhong, Qinliang Su, Zijing Ou, and Fuwei Zhang. Modeling semantic composition with syntactic hypergraph for video question answering. arXiv preprint arXiv:2205.06530, 2022. 2 +[51] Jingkang Yang, Yi Zhe Ang, Zujin Guo, Kaiyang Zhou, Wayne Zhang, and Ziwei Liu. Panoptic scene graph generation. In European Conference on Computer Vision, pages 178-196. Springer, 2022. 2 +[52] Jingkang Yang, Wenxuan Peng, Xiangtai Li, Zujin Guo, Liangyu Chen, Bo Li, Zheng Ma, Kaiyang Zhou, Wayne Zhang, Chen Change Loy, et al. Panoptic video scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18675-18685, 2023. 1, 2, 5, 6, 7 +[53] Guangyao Zhai, Evin Pinar Örnek, Shun-Cheng Wu, Yan Di, Federico Tombari, Nassir Navab, and Benjamin Busam. Commonscenes: Generating commonsense 3d indoor scenes with scene graphs. Advances in Neural Information Processing Systems, 36, 2024. 2 +[54] Zizhao Zhang, Yifan Feng, Shihui Ying, and Yue Gao. Deep hypergraph structure learning. arXiv preprint arXiv:2208.12547, 2022. 2 +[55] Xiaolin Zhu, Dongli Wang, Jianxun Li, Rui Su, Qin Wan, and Yan Zhou. Dynamical attention hypergraph convolutional network for group activity recognition. IEEE Transactions on Neural Networks and Learning Systems, 2024. 2 \ No newline at end of file diff --git a/2025/HyperGLM_ HyperGraph for Video Scene Graph Generation and Anticipation/images.zip b/2025/HyperGLM_ HyperGraph for Video Scene Graph Generation and Anticipation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6cf3469219a21fc8a72417e644d9348101ab8210 --- /dev/null +++ b/2025/HyperGLM_ HyperGraph for Video Scene Graph Generation and Anticipation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b9d12d93f6a0bb45c805f3d26acbeea19242d0d6a4831221bd83d5c6d70ffcf8 +size 608737 diff --git a/2025/HyperGLM_ HyperGraph for Video Scene Graph Generation and Anticipation/layout.json b/2025/HyperGLM_ HyperGraph for Video Scene Graph Generation and Anticipation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6515845d6a3e3d93f94f55747147e0369aad9bd7 --- /dev/null +++ b/2025/HyperGLM_ HyperGraph for Video Scene Graph Generation and Anticipation/layout.json @@ -0,0 +1,11081 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 61, + 103, + 550, + 123 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 103, + 550, + 123 + ], + "spans": [ + { + "bbox": [ + 61, + 103, + 550, + 123 + ], + "type": "text", + "content": "HyperGLM: HyperGraph for Video Scene Graph Generation and Anticipation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 97, + 142, + 512, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 142, + 512, + 171 + ], + "spans": [ + { + "bbox": [ + 97, + 142, + 512, + 171 + ], + "type": "text", + "content": "Trong-Thuan Nguyen" + }, + { + "bbox": [ + 97, + 142, + 512, + 171 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 97, + 142, + 512, + 171 + ], + "type": "text", + "content": ", Pha Nguyen" + }, + { + "bbox": [ + 97, + 142, + 512, + 171 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 97, + 142, + 512, + 171 + ], + "type": "text", + "content": ", Jackson Cothren" + }, + { + "bbox": [ + 97, + 142, + 512, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 97, + 142, + 512, + 171 + ], + "type": "text", + "content": ", Alper Yilmaz" + }, + { + "bbox": [ + 97, + 142, + 512, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 97, + 142, + 512, + 171 + ], + "type": "text", + "content": ", Khoa Luu" + }, + { + "bbox": [ + 97, + 142, + 512, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 97, + 142, + 512, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 97, + 142, + 512, + 171 + ], + "type": "text", + "content": "University of Arkansas " + }, + { + "bbox": [ + 97, + 142, + 512, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 97, + 142, + 512, + 171 + ], + "type": "text", + "content": "Ohio State University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 125, + 172, + 485, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 172, + 485, + 200 + ], + "spans": [ + { + "bbox": [ + 125, + 172, + 485, + 200 + ], + "type": "inline_equation", + "content": "^{1}\\{" + }, + { + "bbox": [ + 125, + 172, + 485, + 200 + ], + "type": "text", + "content": "thuann, panguyen, jcothre, khoaluu\\}@uark.edu " + }, + { + "bbox": [ + 125, + 172, + 485, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 125, + 172, + 485, + 200 + ], + "type": "text", + "content": "yilmaz.15@osu.edu uark-cviu.github.io/projects/HyperGLM" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 86, + 237, + 523, + 285 + ], + "blocks": [ + { + "bbox": [ + 86, + 237, + 523, + 285 + ], + "lines": [ + { + "bbox": [ + 86, + 237, + 523, + 285 + ], + "spans": [ + { + "bbox": [ + 86, + 237, + 523, + 285 + ], + "type": "image", + "image_path": "7ec01ce8a7d7ad73de9fa91943f0ab8555b159ceb724809c9ac3eb2b093a6f09.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 397, + 286, + 519, + 297 + ], + "lines": [ + { + "bbox": [ + 397, + 286, + 519, + 297 + ], + "spans": [ + { + "bbox": [ + 397, + 286, + 519, + 297 + ], + "type": "text", + "content": "What are the roles of the people?" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 88, + 300, + 256, + 374 + ], + "blocks": [ + { + "bbox": [ + 153, + 288, + 304, + 298 + ], + "lines": [ + { + "bbox": [ + 153, + 288, + 304, + 298 + ], + "spans": [ + { + "bbox": [ + 153, + 288, + 304, + 298 + ], + "type": "text", + "content": "What is the relationship between objects?" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 88, + 300, + 256, + 374 + ], + "lines": [ + { + "bbox": [ + 88, + 300, + 256, + 374 + ], + "spans": [ + { + "bbox": [ + 88, + 300, + 256, + 374 + ], + "type": "image", + "image_path": "5069f06678a48c5320e6bbad12cc3935df7e1314e9541189e070d67711d3fc70.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 88, + 375, + 304, + 406 + ], + "lines": [ + { + "bbox": [ + 88, + 375, + 304, + 406 + ], + "spans": [ + { + "bbox": [ + 88, + 375, + 304, + 406 + ], + "type": "text", + "content": "Person-1 holding the Umbrella-1 and standing beside Person-2. Person-1 and Person-2 are waiting for Person-3 in her yellow dress. Then, Person -1 talking with Person-3." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 410, + 555, + 431 + ], + "lines": [ + { + "bbox": [ + 55, + 410, + 555, + 431 + ], + "spans": [ + { + "bbox": [ + 55, + 410, + 555, + 431 + ], + "type": "text", + "content": "Figure 1. Our HyperGLM framework supports Video Scene Graph Generation, Anticipation, and Reasoning. HyperGLM constructs scene graphs from observed video frames and predicts relationships in unseen frames by leveraging a unified hypergraph for temporal modeling and comprehensive understanding." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 315, + 302, + 512, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 302, + 512, + 333 + ], + "spans": [ + { + "bbox": [ + 315, + 302, + 512, + 333 + ], + "type": "text", + "content": "It appears that Person-1 is picking up his family, as Person-3 is being dropped off from the helicopter in the rainy weather. Person-2 is likely serving as security." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 384, + 335, + 517, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 335, + 517, + 346 + ], + "spans": [ + { + "bbox": [ + 384, + 335, + 517, + 346 + ], + "type": "text", + "content": "What might the next interaction be?" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 349, + 500, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 349, + 500, + 369 + ], + "spans": [ + { + "bbox": [ + 315, + 349, + 500, + 369 + ], + "type": "text", + "content": "Person-1 and Person-3 are talking while waiting for another family member to get off the helicopter." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 372, + 371, + 518, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 371, + 518, + 381 + ], + "spans": [ + { + "bbox": [ + 372, + 371, + 518, + 381 + ], + "type": "text", + "content": "What might they do after moving away?" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 385, + 503, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 385, + 503, + 406 + ], + "spans": [ + { + "bbox": [ + 313, + 385, + 503, + 406 + ], + "type": "text", + "content": "After arriving by helicopter, Person-1, Person-3, and another woman will likely move indoors to rest." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 151, + 440, + 200, + 452 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 440, + 200, + 452 + ], + "spans": [ + { + "bbox": [ + 151, + 440, + 200, + 452 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 54, + 459, + 297, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 459, + 297, + 698 + ], + "spans": [ + { + "bbox": [ + 54, + 459, + 297, + 698 + ], + "type": "text", + "content": "Multimodal LLMs have advanced vision-language tasks but still struggle with understanding video scenes. To bridge this gap, Video Scene Graph Generation (VidSGG) has emerged to capture multi-object relationships across video frames. However, prior methods rely on pairwise connections, limiting their ability to handle complex multi-object interactions and reasoning. To this end, we propose the Multimodal Large Language Models (LLMs) on a Scene HyperGraph (HyperGLM), promoting reasoning about multi-way interactions and higher-order relationships. Our approach uniquely integrates entity scene graphs, which capture spatial relationships between objects, with a procedural graph that models their causal transitions, forming a unified HyperGraph. Significantly, HyperGLM enables reasoning by injecting this unified HyperGraph into LLMs. Additionally, we introduce a new Video Scene Graph Reasoning (VSGR) dataset featuring 1.9M frames from third-person, egocentric, and drone views and support five tasks. Empirically, HyperGLM consistently outperforms state-of-the-art methods, effectively modeling and reasoning complex relationships in diverse scenes." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 440, + 394, + 452 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 440, + 394, + 452 + ], + "spans": [ + { + "bbox": [ + 315, + 440, + 394, + 452 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 460, + 556, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 460, + 556, + 615 + ], + "spans": [ + { + "bbox": [ + 312, + 460, + 556, + 615 + ], + "type": "text", + "content": "In recent years, Multimodal Large Language Models (LLMs) [2, 10] have set new benchmarks, with vision-language models excelling in diverse multimodal tasks. However, fully understanding dynamic video scenes remains a significant challenge for applications like autonomous driving, intelligent surveillance, human-object interaction, and multimedia analysis. Towards this goal, Video Scene Graph Generation (VidSGG) [18, 52] has emerged as a critical task for capturing multi-object relationships across video frames. In particular, VidSGG enables high-level tasks such as event forecasting [36, 43, 45], video captioning [24, 38, 40], and video question answering [20, 30, 32, 41] by constructing detailed representations of entities and their interactions." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 617, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 617, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 312, + 617, + 557, + 714 + ], + "type": "text", + "content": "However, prior VidSGG methods and datasets have limited capacity for comprehensive video understanding. Traditional scene graph-based methods [4, 34, 35, 42] only model pairwise object relationships within single frames, making it challenging to capture higher-order relationships and temporal dependencies in real-world scenarios. Additionally, existing benchmark datasets [18, 34, 35, 52] focus is primarily confined to Scene Graph Generation (SGG) and Scene" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 137, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 137, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 137, + 713 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 67, + 703, + 137, + 713 + ], + "type": "text", + "content": " Equal contribution." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29150" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 108 + ], + "type": "text", + "content": "Graph Anticipation (SGA) tasks, lacking annotations for reasoning tasks such as Video Question Answering (VQA), Video Captioning (VC), and Relation Reasoning (RR)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 110, + 298, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 110, + 298, + 360 + ], + "spans": [ + { + "bbox": [ + 56, + 110, + 298, + 360 + ], + "type": "text", + "content": "In this paper, we propose the Multimodal LLMs on a Scene HyperGraph (HyperGLM) approach to promote reasoning about multi-way interactions and higher-order relationships through a unified HyperGraph and LLMs, as illustrated in Fig. 1. To achieve this goal, we uniquely incorporate entity scene graphs, which capture spatial relationships between objects, with a procedural graph that models their causal transitions across video frames, forming a unified HyperGraph, as shown in Fig. 2. Significantly, our HyperGLM approach allows reasoning by injecting this unified HyperGraph into LLMs. In addition, we introduce a novel Video Scene Graph Reasoning (VSGR) dataset, comprising 1.9 million video frames surpassing existing benchmark datasets [18, 34, 35, 47, 52] in scale and annotation depth. Specifically, our VSGR dataset includes videos from third-person, egocentric, and drone perspectives. It supports five tasks: Scene Graph Generation, Scene Graph Anticipation, Video Question Answering, Video Captioning, and Relation Reasoning. Notably, our VSGR dataset introduces a new Relation Reasoning task, setting it apart from existing video scene graph datasets, as shown in the comparison in Table 1." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 360, + 296, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 360, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 360, + 296, + 479 + ], + "type": "text", + "content": "Contributions of this Work. This work presents three contributions to advance Video Scene Graph Generation. First, we introduce Multimodal LLMs on a Scene HyperGraph, leveraging hyperedges and LLMs for reasoning about multiway interactions and higher-order relationships. Second, we develop a new Video Scene Graph Reasoning dataset, surpassing existing scale and annotation depth benchmarks. Our VSGR dataset primarily supports five tasks within diverse video scenes. Finally, the proposed HyperGLM consistently outperforms state-of-the-art methods across all five tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 489, + 142, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 489, + 142, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 489, + 142, + 502 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 509, + 295, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 295, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 295, + 545 + ], + "type": "text", + "content": "In this section, we review advances in scene graph generation and hypergraph applications in computer vision, then discuss existing limitations and the advantages of our approach." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 552, + 194, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 194, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 194, + 563 + ], + "type": "text", + "content": "2.1. Scene Graph Generation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 570, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 297, + 713 + ], + "type": "text", + "content": "Scene Graph Generation has significantly advanced with transformer-based models [6, 13, 17, 25, 26, 51] that have become benchmarks due to their efficiency and state-of-the-art performance. Recent work [3, 21, 33] focuses on reducing bias and enhancing mean recall for rare predicates by integrating external knowledge and applying unbiased contextual augmentation, particularly in dynamic video contexts. Open-vocabulary methods [15, 27] supported by vision-language models handle unseen object and relationship classes, improving generalization. Additionally, generative models (e.g., diffusion-based methods [9, 53]) leverage scene graphs for efficient image and scene synthesis. More-" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 555, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 120 + ], + "type": "text", + "content": "over, spatial-temporal methods [4, 34-36, 42] effectively capture dynamic object relationships in videos. Recently, Large Language Models [23] have been utilized to enhance triplet extraction and alignment in weakly supervised SGG." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 129, + 403, + 141 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 129, + 403, + 141 + ], + "spans": [ + { + "bbox": [ + 313, + 129, + 403, + 141 + ], + "type": "text", + "content": "2.2. HyperGraphs" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 148, + 556, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 148, + 556, + 316 + ], + "spans": [ + { + "bbox": [ + 313, + 148, + 556, + 316 + ], + "type": "text", + "content": "HyperGraphs have been adopted in computer vision to model complex multimodal data and capture higher-order relationships. Unlike traditional graph-based approaches, HyperGraphs connect multiple nodes through hyperedges, enabling multi-way relationships. They enhance Graph Neural Networks (GNNs) [12, 54] by allowing the modeling of more sophisticated interactions. Recent advancements, such as HyperGraph Convolution [1] and HyperGraph Attention [22], have further improved GNNs by capturing relationships beyond simple pairwise connections. Therefore, HyperGraph-based models effectively handle temporal dependencies and complex interactions, significantly boosting performance in tasks like accident anticipation [43], group activity recognition [28, 55], and video question answering [44, 46, 50]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 324, + 386, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 386, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 386, + 335 + ], + "type": "text", + "content": "2.3. Discussion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 342, + 556, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 556, + 534 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 556, + 534 + ], + "type": "text", + "content": "Limitations in Prior Methods. The methods introduced in Section 2.1, based on Progressive Feature Fusion [39, 51], Batch-Progressive Transformer Encoding [6, 13, 17, 25, 26], Spatial-Temporal Context Integration [4, 34, 42], and Memory-Guided Temporal Consistency [7, 33, 35], have advanced VidSGG. However, these methods struggle to model higher-order relationships and complex temporal dynamics. Specifically, Progressive Feature Fusion and Batch-Progressive Transformer Encoding are limited in capturing long-term temporal dependencies, with the former lacking long-term context due to frame-by-frame processing and the latter only addressing short-term dependencies. Similarly, Spatial-Temporal Context Integration and Memory-Guided Temporal Consistency inadequately represent multi-object interactions across video frames and insufficiently capture the temporal evolution required for higher-order relationships." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 534, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 557, + 714 + ], + "type": "text", + "content": "Advantages of Our Approach. Our HyperGLM approach, Multimodal LLMs on a Scene HyperGraph, promotes reasoning about multi-way interactions and high-order relationships. As illustrated in Fig. 2, HyperGLM enhances the model's ability to interpret complex relationships and anticipate intricate video dynamics. Towards this goal, we uniquely integrate entity scene graphs, which is introduced in Sec. 3 to capture spatial interactions between objects in each frame and a procedural graph, which is presented in Sec. 4.1 to model their causal evolution. In our unified approach, hyperedges connect multiple nodes to capture higher-order relationships distinguished from traditional pairwise methods [4, 34, 35, 42]. In addition, the procedural graph enables multi-step transitions for anticipating future interactions or relationships, and reduces bias by generalizing" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29151" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 68, + 306, + 144 + ], + "blocks": [ + { + "bbox": [ + 55, + 68, + 306, + 144 + ], + "lines": [ + { + "bbox": [ + 55, + 68, + 306, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 68, + 306, + 144 + ], + "type": "image", + "image_path": "aecdeabde8cae2c808904b822671766915543edf0224f666bbe937c2a4a3b079.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 143, + 296, + 210 + ], + "lines": [ + { + "bbox": [ + 55, + 143, + 296, + 210 + ], + "spans": [ + { + "bbox": [ + 55, + 143, + 296, + 210 + ], + "type": "text", + "content": "Figure 2. (a) To model the temporal transition, a simple approach can be using two scene graphs " + }, + { + "bbox": [ + 55, + 143, + 296, + 210 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{\\mathbf{t}}" + }, + { + "bbox": [ + 55, + 143, + 296, + 210 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 143, + 296, + 210 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{\\mathbf{t} + 1}" + }, + { + "bbox": [ + 55, + 143, + 296, + 210 + ], + "type": "text", + "content": ". (b) Another procedure graph can present this temporal modeling. (c) Our unified HyperGraph in Fig. 2c integrates both entity scene graph to capture spatial relationships and the procedural graph to model the temporal evolution. HyperEdge represents person sitting on couch, holding, then playing guitar, whereas holding " + }, + { + "bbox": [ + 55, + 143, + 296, + 210 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 143, + 296, + 210 + ], + "type": "text", + "content": " playing describes a chain of interactions. HyperGraph is presented in 3D." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 216, + 296, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 216, + 296, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 216, + 296, + 312 + ], + "type": "text", + "content": "infrequent relationship categories. Furthermore, our HyperGLM approach leverages fundamental mathematical properties: permutation equivariance ensures that the HyperGraph structure remains consistent under any permutation of node labels, and invariance to hyperedge order preserves semantic meanings regardless of the node visit order during random walks which is defined in Alg. 1. The theoretical foundations and mathematical properties are detailed in the Appendices." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 333, + 181, + 345 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 333, + 181, + 345 + ], + "spans": [ + { + "bbox": [ + 55, + 333, + 181, + 345 + ], + "type": "text", + "content": "3. Problem Formulation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 354, + 296, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 296, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 296, + 449 + ], + "type": "text", + "content": "In this section, we define two tasks, including Scene Graph Generation (SGG) and Scene Graph Anticipation (SGA). Graph vertex sequence is represented as " + }, + { + "bbox": [ + 55, + 354, + 296, + 449 + ], + "type": "inline_equation", + "content": "\\{V_{\\mathbf{G}}\\mid 1\\leq t\\leq T\\}" + }, + { + "bbox": [ + 55, + 354, + 296, + 449 + ], + "type": "text", + "content": ", where each set of vertex " + }, + { + "bbox": [ + 55, + 354, + 296, + 449 + ], + "type": "inline_equation", + "content": "V_{\\mathbf{G}}" + }, + { + "bbox": [ + 55, + 354, + 296, + 449 + ], + "type": "text", + "content": " contains object features, bounding boxes, and categories. The scene graph for each frame " + }, + { + "bbox": [ + 55, + 354, + 296, + 449 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 354, + 296, + 449 + ], + "type": "text", + "content": ", denoted by " + }, + { + "bbox": [ + 55, + 354, + 296, + 449 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_t = \\big(V_{\\mathbf{G}},E_{\\mathbf{G}} = \\{r^{ij}\\mid 1\\leq i < j\\leq |V_{\\mathbf{G}}|\\}" + }, + { + "bbox": [ + 55, + 354, + 296, + 449 + ], + "type": "text", + "content": ", consists of all pairwise relationships between objects, with " + }, + { + "bbox": [ + 55, + 354, + 296, + 449 + ], + "type": "inline_equation", + "content": "r^{ij}" + }, + { + "bbox": [ + 55, + 354, + 296, + 449 + ], + "type": "text", + "content": " representing the relationship category between " + }, + { + "bbox": [ + 55, + 354, + 296, + 449 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 55, + 354, + 296, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 354, + 296, + 449 + ], + "type": "inline_equation", + "content": "v_{j}" + }, + { + "bbox": [ + 55, + 354, + 296, + 449 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 449, + 296, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 296, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 296, + 556 + ], + "type": "text", + "content": "We aim to develop a process " + }, + { + "bbox": [ + 55, + 449, + 296, + 556 + ], + "type": "inline_equation", + "content": "p_{\\theta}: (V_{\\mathbf{G}} \\times V_{\\mathbf{G}}) \\to r^{ij}" + }, + { + "bbox": [ + 55, + 449, + 296, + 556 + ], + "type": "text", + "content": " to predict the relationship " + }, + { + "bbox": [ + 55, + 449, + 296, + 556 + ], + "type": "inline_equation", + "content": "r^{ij}" + }, + { + "bbox": [ + 55, + 449, + 296, + 556 + ], + "type": "text", + "content": " between each object pair " + }, + { + "bbox": [ + 55, + 449, + 296, + 556 + ], + "type": "inline_equation", + "content": "(v_t^i, v_t^j)" + }, + { + "bbox": [ + 55, + 449, + 296, + 556 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 55, + 449, + 296, + 556 + ], + "type": "inline_equation", + "content": "V_{\\mathbf{G}}" + }, + { + "bbox": [ + 55, + 449, + 296, + 556 + ], + "type": "text", + "content": ". We define task-specific queries " + }, + { + "bbox": [ + 55, + 449, + 296, + 556 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathrm{SGG}}" + }, + { + "bbox": [ + 55, + 449, + 296, + 556 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 449, + 296, + 556 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathrm{SGA}}" + }, + { + "bbox": [ + 55, + 449, + 296, + 556 + ], + "type": "text", + "content": " to direct our unified model to perform on each specific task. The objective for each task is to minimize the negative likelihood of the predicted scene graph " + }, + { + "bbox": [ + 55, + 449, + 296, + 556 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_t" + }, + { + "bbox": [ + 55, + 449, + 296, + 556 + ], + "type": "text", + "content": " to the truth predicate set " + }, + { + "bbox": [ + 55, + 449, + 296, + 556 + ], + "type": "inline_equation", + "content": "Y_t = \\{y^{ij} \\mid 1 \\leq i < j \\leq |V_{\\mathbf{G}}|\\}" + }, + { + "bbox": [ + 55, + 449, + 296, + 556 + ], + "type": "text", + "content": " on categories indexed by " + }, + { + "bbox": [ + 55, + 449, + 296, + 556 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 449, + 296, + 556 + ], + "type": "text", + "content": ". Scene Graph Generation (SGG) generates the scene graph for each frame " + }, + { + "bbox": [ + 55, + 449, + 296, + 556 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 449, + 296, + 556 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 449, + 296, + 556 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 55, + 449, + 296, + 556 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 449, + 296, + 556 + ], + "type": "inline_equation", + "content": "t = T" + }, + { + "bbox": [ + 55, + 449, + 296, + 556 + ], + "type": "text", + "content": ", which is defined as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 76, + 560, + 296, + 588 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 560, + 296, + 588 + ], + "spans": [ + { + "bbox": [ + 76, + 560, + 296, + 588 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} \\mathbb {E} _ {\\mathbf {G} _ {t}, Y _ {t}} \\left[ - \\sum_ {\\left(v _ {t} ^ {i, j}\\right)} \\sum_ {k} \\left(y _ {k} ^ {i j} \\log p _ {\\theta} \\left(r ^ {i j} \\mid \\mathbf {Q} _ {\\mathrm {S G G}}\\right) _ {k}\\right) \\right] \\tag {1}", + "image_path": "ecbd4ec53a2785e61861da1eb84bd2554acdd3ae50e1b50645052705acdeaefe.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 594, + 296, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 629 + ], + "type": "text", + "content": "Scene Graph Anticipation (SGA) anticipates the scene graphs for future frames, generating predictions " + }, + { + "bbox": [ + 55, + 594, + 296, + 629 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{t + n}" + }, + { + "bbox": [ + 55, + 594, + 296, + 629 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 594, + 296, + 629 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 594, + 296, + 629 + ], + "type": "text", + "content": " denotes the anticipation horizon, formulated as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 643, + 296, + 672 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 643, + 296, + 672 + ], + "spans": [ + { + "bbox": [ + 67, + 643, + 296, + 672 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} \\mathbb {E} _ {\\mathbf {G} _ {\\leq t}, Y _ {t}} \\left[ - \\sum_ {\\left(v _ {t + n} ^ {i, j}\\right)} \\sum_ {k} \\left(y _ {k} ^ {i j} \\log p _ {\\theta} \\left(r ^ {i j} \\mid \\mathbf {Q} _ {\\mathrm {S G A}}\\right) _ {k}\\right) \\right] \\tag {2}", + "image_path": "a7f551e19e3aac0dd4f8c5161d228790b1a8967f455ec1499430cbb8ef11f023.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "In Eqns. (1) and (2), " + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": " indexes the current frame. While SGG predicts the scene graph at frame " + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": ", SGA reads scene graphs up to frame " + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": " to forecast the graph in future frame " + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "inline_equation", + "content": "t + n" + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 315, + 69, + 566, + 150 + ], + "blocks": [ + { + "bbox": [ + 315, + 69, + 566, + 150 + ], + "lines": [ + { + "bbox": [ + 315, + 69, + 566, + 150 + ], + "spans": [ + { + "bbox": [ + 315, + 69, + 566, + 150 + ], + "type": "image", + "image_path": "fda9408f652dfc66d7ebbd26c6062a1f40b9350b6215790cdbe771985ea3814d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 152, + 555, + 228 + ], + "lines": [ + { + "bbox": [ + 313, + 152, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 152, + 555, + 228 + ], + "type": "text", + "content": "Figure 3. Our HyperGLM framework comprises an image encoder, MLP projector, temporal aggregator, unified HyperGraph, and language model. It processes video frames by encoding each frame with the image encoder and MLP, extracting spatio-temporal features through image patch grids to generate " + }, + { + "bbox": [ + 313, + 152, + 555, + 228 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 152, + 555, + 228 + ], + "type": "text", + "content": " spatial tokens per frame. The temporal aggregator compresses the " + }, + { + "bbox": [ + 313, + 152, + 555, + 228 + ], + "type": "inline_equation", + "content": "T \\times N" + }, + { + "bbox": [ + 313, + 152, + 555, + 228 + ], + "type": "text", + "content": " embeddings over time. The MLP projector then transforms these visual embeddings into the language feature space as frame tokens, interleaved with language tokens, and fed into the Large Language Models." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 236, + 519, + 250 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 519, + 250 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 519, + 250 + ], + "type": "text", + "content": "4. Our Proposed HyperGLM Approach" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 257, + 555, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 257, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 313, + 257, + 555, + 281 + ], + "type": "text", + "content": "In this section, we present our approach, which incorporates a unified HyperGraph into the LLMs, as illustrated in Fig. 3." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 288, + 468, + 301 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 288, + 468, + 301 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 468, + 301 + ], + "type": "text", + "content": "4.1. Video Scene (Hyper)Graphs" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 307, + 556, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 307, + 556, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 307, + 556, + 582 + ], + "type": "text", + "content": "Traditional VidSGG methods [4, 34, 42] construct scene graphs " + }, + { + "bbox": [ + 313, + 307, + 556, + 582 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_t" + }, + { + "bbox": [ + 313, + 307, + 556, + 582 + ], + "type": "text", + "content": " to represent scene entities and their relationships as in Fig. 2a. However, these graph-based approaches do not capture higher-order relationships with temporal dependencies for video understanding [36]. To model this property, we propose a novel HyperGraph-based framework that constructs a unified HyperGraph " + }, + { + "bbox": [ + 313, + 307, + 556, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{H} = (V_{\\mathcal{H}}, E_{\\mathcal{H}})" + }, + { + "bbox": [ + 313, + 307, + 556, + 582 + ], + "type": "text", + "content": ", representing spatial relationships within individual frames and temporal transitions across frames, illustrated in Fig. 4. HyperGraph extends traditional graph structures by allowing hyperedges to connect more than two nodes, making them particularly effective for modeling these higher-order relationships. To leverage this property, our unified HyperGraph (see Fig. 2c) integrates entity scene graphs " + }, + { + "bbox": [ + 313, + 307, + 556, + 582 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_t" + }, + { + "bbox": [ + 313, + 307, + 556, + 582 + ], + "type": "text", + "content": " for each frame, capturing spatial interactions as introduced in Sec. 3, along with a procedural graph " + }, + { + "bbox": [ + 313, + 307, + 556, + 582 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 313, + 307, + 556, + 582 + ], + "type": "text", + "content": " that models their causal transitions, as detailed in Sec. 4.1. We unify these components using a random-walk HyperGraph construction, which is defined in Alg. 1 to capture higher-order connectivity patterns and structural semantics, thereby approximating subgraph matching between the entity scene graphs and the procedural graph. This integration allows our approach to model current subjects' interactions and anticipate their future relationships." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 582, + 556, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 677 + ], + "type": "text", + "content": "Procedural Graph Construction. SGA objective is to predict the set of relationships " + }, + { + "bbox": [ + 313, + 582, + 556, + 677 + ], + "type": "inline_equation", + "content": "E_{\\mathbf{G}}" + }, + { + "bbox": [ + 313, + 582, + 556, + 677 + ], + "type": "text", + "content": " in the next frame based on the current set of relationships. To model the temporal evolution of causal relationships between objects across video frames, we introduce a procedural graph " + }, + { + "bbox": [ + 313, + 582, + 556, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{P} = (V_{\\mathbf{P}},E_{\\mathbf{P}})" + }, + { + "bbox": [ + 313, + 582, + 556, + 677 + ], + "type": "text", + "content": ", serving as the temporal counterpart to the entity scene graphs " + }, + { + "bbox": [ + 313, + 582, + 556, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_t" + }, + { + "bbox": [ + 313, + 582, + 556, + 677 + ], + "type": "text", + "content": " as shown in Fig. 2b. In particular, the procedural graph " + }, + { + "bbox": [ + 313, + 582, + 556, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 313, + 582, + 556, + 677 + ], + "type": "text", + "content": " can vary across datasets, modeling relationship transitions." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "Toward this goal, we model and denote the set of vertices in " + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "inline_equation", + "content": "V_{\\mathbf{P}} = \\{r_t^{ij} \\in V_{\\mathbf{G}} \\mid 1 \\leq t \\leq T\\} \\neq" + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": " representing distinct relationship categories. The set of edges in " + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": ", i.e.," + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29152" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 553, + 229 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 553, + 229 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 553, + 229 + ], + "type": "image", + "image_path": "8f8900f6112c23b2fe5219cfb188a8be75705feac1da87af2c44f397671aeef8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 230, + 555, + 270 + ], + "lines": [ + { + "bbox": [ + 55, + 230, + 555, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 230, + 555, + 270 + ], + "type": "text", + "content": "Figure 4. Our Video Scene HyperGraph, including entity graphs and a procedural graph, as defined in Sec. 4.1. Blue nodes represent entities, while green nodes denote relationships. The entity graph captures spatial relationships (subject " + }, + { + "bbox": [ + 55, + 230, + 555, + 270 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 230, + 555, + 270 + ], + "type": "text", + "content": " relationship " + }, + { + "bbox": [ + 55, + 230, + 555, + 270 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 230, + 555, + 270 + ], + "type": "text", + "content": " object), whereas the procedural graph models relationship transitions " + }, + { + "bbox": [ + 55, + 230, + 555, + 270 + ], + "type": "inline_equation", + "content": "(\\rightarrow)" + }, + { + "bbox": [ + 55, + 230, + 555, + 270 + ], + "type": "text", + "content": ". Hyperedges are visualized as polygons, encapsulating interactions through chains of relationships. For instance, a hyperedge illustrates a person picking up, holding, opening, and reading a book while sitting on a couch. HyperGraph is presented in 3D, see Supplementary video." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 277, + 298, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 298, + 339 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 298, + 339 + ], + "type": "inline_equation", + "content": "E_{\\mathbf{P}} = \\{(r_m,r_n)\\}" + }, + { + "bbox": [ + 55, + 277, + 298, + 339 + ], + "type": "text", + "content": " represent possible causal transitions between these relationships, where an edge " + }, + { + "bbox": [ + 55, + 277, + 298, + 339 + ], + "type": "inline_equation", + "content": "(r_m,r_n)" + }, + { + "bbox": [ + 55, + 277, + 298, + 339 + ], + "type": "text", + "content": " indicates that relationship " + }, + { + "bbox": [ + 55, + 277, + 298, + 339 + ], + "type": "inline_equation", + "content": "r_m" + }, + { + "bbox": [ + 55, + 277, + 298, + 339 + ], + "type": "text", + "content": " causally lead to relationship " + }, + { + "bbox": [ + 55, + 277, + 298, + 339 + ], + "type": "inline_equation", + "content": "r_n" + }, + { + "bbox": [ + 55, + 277, + 298, + 339 + ], + "type": "text", + "content": ". We quantify causal transitions by calculating transition probabilities " + }, + { + "bbox": [ + 55, + 277, + 298, + 339 + ], + "type": "inline_equation", + "content": "w(r_m,r_n)" + }, + { + "bbox": [ + 55, + 277, + 298, + 339 + ], + "type": "text", + "content": " via their observed frequencies as in Eqn. (3)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 65, + 356, + 296, + 421 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 356, + 296, + 421 + ], + "spans": [ + { + "bbox": [ + 65, + 356, + 296, + 421 + ], + "type": "interline_equation", + "content": "w \\left(r _ {m}, r _ {n}\\right) = \\frac {\\sum_ {t = 1} ^ {T - 1} \\sum_ {r _ {t , t + 1} ^ {i j}} \\mathbb {1} \\left(r _ {t} ^ {i j} = r _ {m} \\wedge r _ {t + 1} ^ {i j} = r _ {n}\\right)}{\\sum_ {t = 1} ^ {T - 1} \\sum_ {r _ {t , t + 1} ^ {i j}} \\mathbb {1} \\left(r _ {t} ^ {i j} = r _ {m}\\right)} \\tag {3}", + "image_path": "f6d9ab605dbc192583186213bc38ef8169ebf52f4c5ac9a732d6b1e5eef23bba.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 429, + 296, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 429, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 429, + 296, + 479 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 429, + 296, + 479 + ], + "type": "inline_equation", + "content": "\\mathbb{1}(\\cdot)" + }, + { + "bbox": [ + 55, + 429, + 296, + 479 + ], + "type": "text", + "content": " is the indicator function that counts transitions from relationship " + }, + { + "bbox": [ + 55, + 429, + 296, + 479 + ], + "type": "inline_equation", + "content": "r_m" + }, + { + "bbox": [ + 55, + 429, + 296, + 479 + ], + "type": "text", + "content": " at current frame " + }, + { + "bbox": [ + 55, + 429, + 296, + 479 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 429, + 296, + 479 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 429, + 296, + 479 + ], + "type": "inline_equation", + "content": "r_n" + }, + { + "bbox": [ + 55, + 429, + 296, + 479 + ], + "type": "text", + "content": " at next frame " + }, + { + "bbox": [ + 55, + 429, + 296, + 479 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 55, + 429, + 296, + 479 + ], + "type": "text", + "content": ". Next, self-loops " + }, + { + "bbox": [ + 55, + 429, + 296, + 479 + ], + "type": "inline_equation", + "content": "w(r_m, r_m)" + }, + { + "bbox": [ + 55, + 429, + 296, + 479 + ], + "type": "text", + "content": " are removed from the graph, and these probabilities are normalized as in Eqn. (4)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 93, + 486, + 296, + 512 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 486, + 296, + 512 + ], + "spans": [ + { + "bbox": [ + 93, + 486, + 296, + 512 + ], + "type": "interline_equation", + "content": "\\sum_ {r _ {n} \\in E _ {\\mathbf {P}}} w \\left(r _ {m}, r _ {n}\\right) = 1 \\quad \\text {f o r a l l} r _ {m} \\in E _ {\\mathbf {P}} \\tag {4}", + "image_path": "2dda55cf7a0f1a547b56bb1fe5a478bc90e849a56aed2009f8cc3bbbd6cef747.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 519, + 296, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 519, + 296, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 519, + 296, + 568 + ], + "type": "text", + "content": "By leveraging these probabilities, the procedural graph " + }, + { + "bbox": [ + 55, + 519, + 296, + 568 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 55, + 519, + 296, + 568 + ], + "type": "text", + "content": " enhances the prediction of future relationships. For each relationship " + }, + { + "bbox": [ + 55, + 519, + 296, + 568 + ], + "type": "inline_equation", + "content": "r_t^{ij}" + }, + { + "bbox": [ + 55, + 519, + 296, + 568 + ], + "type": "text", + "content": " in frame " + }, + { + "bbox": [ + 55, + 519, + 296, + 568 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 519, + 296, + 568 + ], + "type": "text", + "content": ", the most probable relationship " + }, + { + "bbox": [ + 55, + 519, + 296, + 568 + ], + "type": "inline_equation", + "content": "r_{t + 1}^{ij}" + }, + { + "bbox": [ + 55, + 519, + 296, + 568 + ], + "type": "text", + "content": " in the next frame " + }, + { + "bbox": [ + 55, + 519, + 296, + 568 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 55, + 519, + 296, + 568 + ], + "type": "text", + "content": " is determined as in Eqn. (5)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 107, + 576, + 296, + 595 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 576, + 296, + 595 + ], + "spans": [ + { + "bbox": [ + 107, + 576, + 296, + 595 + ], + "type": "interline_equation", + "content": "r _ {t + 1} ^ {i j} = \\arg \\max _ {r _ {n}} P \\left(r _ {n} \\mid r _ {t} ^ {i j}, v ^ {i, j}\\right) \\tag {5}", + "image_path": "be0a349588b482e22f9e1e8aaba857bcf78149ea451653b49e99d100cd0ba347.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 605, + 296, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 654 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 654 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 605, + 296, + 654 + ], + "type": "inline_equation", + "content": "P(r_{n} \\mid r_{t}^{ij}, v^{i,j}) = w(r_{t}^{ij}, r_{n}) \\times v^{i,j}" + }, + { + "bbox": [ + 55, + 605, + 296, + 654 + ], + "type": "text", + "content": " is the probability of transitioning from relationship " + }, + { + "bbox": [ + 55, + 605, + 296, + 654 + ], + "type": "inline_equation", + "content": "r_{t}^{ij}" + }, + { + "bbox": [ + 55, + 605, + 296, + 654 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 605, + 296, + 654 + ], + "type": "inline_equation", + "content": "r_{n}" + }, + { + "bbox": [ + 55, + 605, + 296, + 654 + ], + "type": "text", + "content": ", looking at object features. This allows the model to anticipate future interactions based on established temporal patterns." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "HyperGraph Construction. To incorporate spatial relationships between objects in each frame and their causal temporal transitions, we construct a unified HyperGraph " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": " that integrates the entity scene graphs " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_t" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": " and the procedural graph " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": ". Specifically, a HyperGraph is an augmented" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 277, + 555, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 555, + 350 + ], + "type": "text", + "content": "representation combining the entity scene graphs with the procedural graph. This unified structure merges spatial and temporal relationships into a single graph, enabling the use of conventional graph algorithms while preserving the complex interactions captured by the HyperGraph. Mathematically, a unified HyperGraph " + }, + { + "bbox": [ + 313, + 277, + 555, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{H} = (V_{\\mathcal{H}}, E_{\\mathcal{H}})" + }, + { + "bbox": [ + 313, + 277, + 555, + 350 + ], + "type": "text", + "content": " is defined as in Eqn. (6)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 350, + 371, + 555, + 405 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 371, + 555, + 405 + ], + "spans": [ + { + "bbox": [ + 350, + 371, + 555, + 405 + ], + "type": "interline_equation", + "content": "\\mathcal {H} = \\left(\\bigcup_ {t = 1} ^ {T} V _ {\\mathbf {G} _ {t}} \\cup V _ {\\mathbf {P}}, \\quad \\bigcup_ {t = 1} ^ {T} E _ {\\mathbf {G} _ {t}} \\cup E _ {\\mathbf {P}}\\right) \\tag {6}", + "image_path": "a552fdfb015534063502de1cf97a51e1a84b84e802e947bd7af90db172075911.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 422, + 555, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 555, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 555, + 483 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 422, + 555, + 483 + ], + "type": "inline_equation", + "content": "V_{\\mathcal{H}}" + }, + { + "bbox": [ + 313, + 422, + 555, + 483 + ], + "type": "text", + "content": " includes all entity nodes " + }, + { + "bbox": [ + 313, + 422, + 555, + 483 + ], + "type": "inline_equation", + "content": "v_{t}^{i}" + }, + { + "bbox": [ + 313, + 422, + 555, + 483 + ], + "type": "text", + "content": " from each " + }, + { + "bbox": [ + 313, + 422, + 555, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_t" + }, + { + "bbox": [ + 313, + 422, + 555, + 483 + ], + "type": "text", + "content": " and the relationship type nodes " + }, + { + "bbox": [ + 313, + 422, + 555, + 483 + ], + "type": "inline_equation", + "content": "V_{\\mathbf{P}}" + }, + { + "bbox": [ + 313, + 422, + 555, + 483 + ], + "type": "text", + "content": ". The hyperedge set " + }, + { + "bbox": [ + 313, + 422, + 555, + 483 + ], + "type": "inline_equation", + "content": "E_{\\mathcal{H}}" + }, + { + "bbox": [ + 313, + 422, + 555, + 483 + ], + "type": "text", + "content": " includes pairwise relationships " + }, + { + "bbox": [ + 313, + 422, + 555, + 483 + ], + "type": "inline_equation", + "content": "E_{\\mathbf{G}_t}" + }, + { + "bbox": [ + 313, + 422, + 555, + 483 + ], + "type": "text", + "content": " within each " + }, + { + "bbox": [ + 313, + 422, + 555, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_t" + }, + { + "bbox": [ + 313, + 422, + 555, + 483 + ], + "type": "text", + "content": ", capturing spatial relationships and temporal transition edges " + }, + { + "bbox": [ + 313, + 422, + 555, + 483 + ], + "type": "inline_equation", + "content": "E_{\\mathbf{P}}" + }, + { + "bbox": [ + 313, + 422, + 555, + 483 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 422, + 555, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 313, + 422, + 555, + 483 + ], + "type": "text", + "content": ", modeling the evolution of causal relationships across frames." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 486, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 556, + 714 + ], + "type": "text", + "content": "Random-walk Algorithm. We employ the random walks algorithm outlined in Alg. 1 to sample representative substructures from the unified HyperGraph " + }, + { + "bbox": [ + 313, + 486, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 313, + 486, + 556, + 714 + ], + "type": "text", + "content": ", which captures connectivity patterns and mitigates the NP-hardness of exact subgraph matching. Specifically, these walks alternate between nodes and hyperedges, preserving the multi-node connections intrinsic to hyperedges and capturing the complexity of multi-object relationships and their transition across video frames. In each walk, a hyperedge " + }, + { + "bbox": [ + 313, + 486, + 556, + 714 + ], + "type": "inline_equation", + "content": "h_i" + }, + { + "bbox": [ + 313, + 486, + 556, + 714 + ], + "type": "text", + "content": " aggregates the visited nodes, thereby encapsulating higher-order relationships. For example, in the entity scene graph " + }, + { + "bbox": [ + 313, + 486, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_t" + }, + { + "bbox": [ + 313, + 486, + 556, + 714 + ], + "type": "text", + "content": ", a walk might traverse from a \"person\" to a \"cup\" via the \"holding\", resulting in the hyperedge " + }, + { + "bbox": [ + 313, + 486, + 556, + 714 + ], + "type": "inline_equation", + "content": "h_i = \\{\\text{person}, \\text{cup}\\}" + }, + { + "bbox": [ + 313, + 486, + 556, + 714 + ], + "type": "text", + "content": ". Similarly, within the procedural graph " + }, + { + "bbox": [ + 313, + 486, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 313, + 486, + 556, + 714 + ], + "type": "text", + "content": ", a walk might transition through a sequence of interactions such as \"holding\", \"placing\", and \"releasing\", forming the hyperedge " + }, + { + "bbox": [ + 313, + 486, + 556, + 714 + ], + "type": "inline_equation", + "content": "h_i = \\{\\text{holding}, \\text{placing}, \\text{releasing}\\}" + }, + { + "bbox": [ + 313, + 486, + 556, + 714 + ], + "type": "text", + "content": ". Therefore, we generate sampled hyperedges " + }, + { + "bbox": [ + 313, + 486, + 556, + 714 + ], + "type": "inline_equation", + "content": "E_{\\text{sampled}} = \\{h_i \\mid i = 1, \\dots, N_w\\}" + }, + { + "bbox": [ + 313, + 486, + 556, + 714 + ], + "type": "text", + "content": " as in L25 of Alg. 1 by conducting multiple random walks." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29153" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 55, + 87, + 298, + 439 + ], + "blocks": [ + { + "bbox": [ + 56, + 71, + 293, + 85 + ], + "lines": [ + { + "bbox": [ + 56, + 71, + 293, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 293, + 85 + ], + "type": "text", + "content": "Algorithm 1 Random-walk for HyperGraph Construction." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 55, + 87, + 298, + 439 + ], + "lines": [ + { + "bbox": [ + 55, + 87, + 298, + 439 + ], + "spans": [ + { + "bbox": [ + 55, + 87, + 298, + 439 + ], + "type": "text", + "content": "Require: " + }, + { + "bbox": [ + 55, + 87, + 298, + 439 + ], + "type": "inline_equation", + "content": "\\mathcal{H} = (V_{\\mathcal{H}},E_{\\mathcal{H}})" + }, + { + "bbox": [ + 55, + 87, + 298, + 439 + ], + "type": "text", + "content": " , Number of Walks " + }, + { + "bbox": [ + 55, + 87, + 298, + 439 + ], + "type": "inline_equation", + "content": "N_{w}" + }, + { + "bbox": [ + 55, + 87, + 298, + 439 + ], + "type": "text", + "content": " , Walk Length " + }, + { + "bbox": [ + 55, + 87, + 298, + 439 + ], + "type": "inline_equation", + "content": "N_{l}" + }, + { + "bbox": [ + 55, + 87, + 298, + 439 + ], + "type": "text", + "content": " \nEnsure: " + }, + { + "bbox": [ + 55, + 87, + 298, + 439 + ], + "type": "inline_equation", + "content": "\\mathcal{H}' = (V_{\\mathcal{H}},E_{\\mathcal{H}}^{\\prime})" + }, + { + "bbox": [ + 55, + 87, + 298, + 439 + ], + "type": "text", + "content": " \n1: Initialize " + }, + { + "bbox": [ + 55, + 87, + 298, + 439 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{sampled}}\\gets \\emptyset" + }, + { + "bbox": [ + 55, + 87, + 298, + 439 + ], + "type": "text", + "content": " \n2: for " + }, + { + "bbox": [ + 55, + 87, + 298, + 439 + ], + "type": "inline_equation", + "content": "i = 1" + }, + { + "bbox": [ + 55, + 87, + 298, + 439 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 87, + 298, + 439 + ], + "type": "inline_equation", + "content": "N_{w}" + }, + { + "bbox": [ + 55, + 87, + 298, + 439 + ], + "type": "text", + "content": " do \n3: Select " + }, + { + "bbox": [ + 55, + 87, + 298, + 439 + ], + "type": "inline_equation", + "content": "v_{\\mathrm{start}}\\in V_{\\mathcal{H}}" + }, + { + "bbox": [ + 55, + 87, + 298, + 439 + ], + "type": "text", + "content": " uniformly at random \n4: Initialize walk sequence " + }, + { + "bbox": [ + 55, + 87, + 298, + 439 + ], + "type": "inline_equation", + "content": "S\\leftarrow [v_{\\mathrm{start}}]" + }, + { + "bbox": [ + 55, + 87, + 298, + 439 + ], + "type": "text", + "content": " \n5: for " + }, + { + "bbox": [ + 55, + 87, + 298, + 439 + ], + "type": "inline_equation", + "content": "j = 1" + }, + { + "bbox": [ + 55, + 87, + 298, + 439 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 87, + 298, + 439 + ], + "type": "inline_equation", + "content": "N_{l}" + }, + { + "bbox": [ + 55, + 87, + 298, + 439 + ], + "type": "text", + "content": " do \n6: if " + }, + { + "bbox": [ + 55, + 87, + 298, + 439 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 87, + 298, + 439 + ], + "type": "text", + "content": " is odd then Node to HyperEdge \n7: " + }, + { + "bbox": [ + 55, + 87, + 298, + 439 + ], + "type": "inline_equation", + "content": "v_{\\mathrm{current}}\\gets S[j]" + }, + { + "bbox": [ + 55, + 87, + 298, + 439 + ], + "type": "text", + "content": " \n8: " + }, + { + "bbox": [ + 55, + 87, + 298, + 439 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{v_{\\mathrm{current}}}\\gets \\{h\\in E_{\\mathcal{H}}\\mid v_{\\mathrm{current}}\\in h\\}" + }, + { + "bbox": [ + 55, + 87, + 298, + 439 + ], + "type": "text", + "content": " \n9: if " + }, + { + "bbox": [ + 55, + 87, + 298, + 439 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{v_{\\mathrm{current}}}\\neq \\emptyset" + }, + { + "bbox": [ + 55, + 87, + 298, + 439 + ], + "type": "text", + "content": " then \n10: Select " + }, + { + "bbox": [ + 55, + 87, + 298, + 439 + ], + "type": "inline_equation", + "content": "h_j\\in \\mathcal{E}_{v_{\\mathrm{current}}}" + }, + { + "bbox": [ + 55, + 87, + 298, + 439 + ], + "type": "text", + "content": " \n11: Append " + }, + { + "bbox": [ + 55, + 87, + 298, + 439 + ], + "type": "inline_equation", + "content": "h_j" + }, + { + "bbox": [ + 55, + 87, + 298, + 439 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 87, + 298, + 439 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 87, + 298, + 439 + ], + "type": "text", + "content": " \n12: end if \n13: else HyperEdge to Node \n14: " + }, + { + "bbox": [ + 55, + 87, + 298, + 439 + ], + "type": "inline_equation", + "content": "h_\\mathrm{current}\\gets S[j]" + }, + { + "bbox": [ + 55, + 87, + 298, + 439 + ], + "type": "text", + "content": " \n15: " + }, + { + "bbox": [ + 55, + 87, + 298, + 439 + ], + "type": "inline_equation", + "content": "V_{h_{\\mathrm{current}}}\\gets h_{\\mathrm{current}}" + }, + { + "bbox": [ + 55, + 87, + 298, + 439 + ], + "type": "text", + "content": " \n16: Select " + }, + { + "bbox": [ + 55, + 87, + 298, + 439 + ], + "type": "inline_equation", + "content": "v_{j}\\in V_{h_{\\mathrm{current}}}" + }, + { + "bbox": [ + 55, + 87, + 298, + 439 + ], + "type": "text", + "content": " \n17: Append " + }, + { + "bbox": [ + 55, + 87, + 298, + 439 + ], + "type": "inline_equation", + "content": "v_{j}" + }, + { + "bbox": [ + 55, + 87, + 298, + 439 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 87, + 298, + 439 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 87, + 298, + 439 + ], + "type": "text", + "content": " \n18: end if \n19: end for \n20: " + }, + { + "bbox": [ + 55, + 87, + 298, + 439 + ], + "type": "inline_equation", + "content": "h_i\\gets \\{v\\in S\\mid v\\in V_H\\}" + }, + { + "bbox": [ + 55, + 87, + 298, + 439 + ], + "type": "text", + "content": " ▷Form new HyperEdge \n21: if " + }, + { + "bbox": [ + 55, + 87, + 298, + 439 + ], + "type": "inline_equation", + "content": "h_i\\notin E_{\\mathcal{H}}\\cup E_{\\mathrm{sampled}}" + }, + { + "bbox": [ + 55, + 87, + 298, + 439 + ], + "type": "text", + "content": " then \n22: " + }, + { + "bbox": [ + 55, + 87, + 298, + 439 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{sampled}}\\gets E_{\\mathrm{sampled}}\\cup \\{h_i\\}" + }, + { + "bbox": [ + 55, + 87, + 298, + 439 + ], + "type": "text", + "content": " \n23: end if \n24: end for \n25: " + }, + { + "bbox": [ + 55, + 87, + 298, + 439 + ], + "type": "inline_equation", + "content": "E_{\\mathcal{H}}'\\gets E_{\\mathcal{H}}\\cup E_{\\mathrm{sampled}}" + }, + { + "bbox": [ + 55, + 87, + 298, + 439 + ], + "type": "text", + "content": " \n26: return " + }, + { + "bbox": [ + 55, + 87, + 298, + 439 + ], + "type": "inline_equation", + "content": "\\mathcal{H}' = (V_{\\mathcal{H}},E_{\\mathcal{H}}')" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "code_body" + } + ], + "index": 1, + "sub_type": "algorithm" + }, + { + "bbox": [ + 55, + 445, + 248, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 248, + 459 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 248, + 459 + ], + "type": "text", + "content": "4.2. Multimodal LLMs on HyperGraphs" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 464, + 297, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 464, + 297, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 297, + 514 + ], + "type": "text", + "content": "Formulation. Given an input video " + }, + { + "bbox": [ + 55, + 464, + 297, + 514 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 55, + 464, + 297, + 514 + ], + "type": "text", + "content": " and a task query " + }, + { + "bbox": [ + 55, + 464, + 297, + 514 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 55, + 464, + 297, + 514 + ], + "type": "text", + "content": ", we aim to generate a target answer sequence " + }, + { + "bbox": [ + 55, + 464, + 297, + 514 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 55, + 464, + 297, + 514 + ], + "type": "text", + "content": " of length " + }, + { + "bbox": [ + 55, + 464, + 297, + 514 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 464, + 297, + 514 + ], + "type": "text", + "content": ". Especially, by modeling a HyperGraph " + }, + { + "bbox": [ + 55, + 464, + 297, + 514 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 55, + 464, + 297, + 514 + ], + "type": "text", + "content": ", the target answer " + }, + { + "bbox": [ + 55, + 464, + 297, + 514 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 55, + 464, + 297, + 514 + ], + "type": "text", + "content": " is generated via a process illustrated in Fig. 3, defined as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 87, + 519, + 296, + 552 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 519, + 296, + 552 + ], + "spans": [ + { + "bbox": [ + 87, + 519, + 296, + 552 + ], + "type": "interline_equation", + "content": "p (\\mathbf {A} | \\mathbf {V}, \\mathbf {Q}, \\mathcal {H}) = \\prod_ {i = 0} ^ {L - 1} p \\left(x _ {i} | \\mathbf {V}, \\mathbf {Q}, \\mathcal {H}, x _ {< i}\\right) \\tag {7}", + "image_path": "ce6d9f9e08a1c4e915ead1ac8e3f0afcfae9f6a093223e6f1cfefa48654210b6.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 558, + 296, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 594 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 558, + 296, + 594 + ], + "type": "inline_equation", + "content": "x_{< i}" + }, + { + "bbox": [ + 55, + 558, + 296, + 594 + ], + "type": "text", + "content": " denotes the preceding token sequence, and " + }, + { + "bbox": [ + 55, + 558, + 296, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{A} = [x_0,\\dots ,x_i,\\dots ,x_{L - 1}]" + }, + { + "bbox": [ + 55, + 558, + 296, + 594 + ], + "type": "text", + "content": " is the sequence of answer tokens reasoning video " + }, + { + "bbox": [ + 55, + 558, + 296, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 55, + 558, + 296, + 594 + ], + "type": "text", + "content": " by question " + }, + { + "bbox": [ + 55, + 558, + 296, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 55, + 558, + 296, + 594 + ], + "type": "text", + "content": " and HyperGraph " + }, + { + "bbox": [ + 55, + 558, + 296, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 55, + 558, + 296, + 594 + ], + "type": "text", + "content": " in Fig. 4." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 594, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 594, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 594, + 297, + 714 + ], + "type": "text", + "content": "As illustrated in Fig. 5, our process begins with the generation, constructing an entity scene graph " + }, + { + "bbox": [ + 54, + 594, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_t = (V_t^e, E_{\\mathbf{G}})" + }, + { + "bbox": [ + 54, + 594, + 297, + 714 + ], + "type": "text", + "content": " for each frame " + }, + { + "bbox": [ + 54, + 594, + 297, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 54, + 594, + 297, + 714 + ], + "type": "text", + "content": " to capture detected objects and their relationships. Relationship anticipation employs the procedural graph " + }, + { + "bbox": [ + 54, + 594, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 54, + 594, + 297, + 714 + ], + "type": "text", + "content": " to model temporal evolution and predict future interactions. The model then engages in reasoning using the HyperGraph " + }, + { + "bbox": [ + 54, + 594, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 54, + 594, + 297, + 714 + ], + "type": "text", + "content": " and verification through video captioning to ensure contextual relevance and accuracy. It refines understanding through clarification, generating answers by reasoning over the video and " + }, + { + "bbox": [ + 54, + 594, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 54, + 594, + 297, + 714 + ], + "type": "text", + "content": ". In scene forecasting, it predicts" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 313, + 101, + 557, + 167 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 555, + 100 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 555, + 100 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 555, + 100 + ], + "type": "text", + "content": "Table 1. Comparisons of video scene graph datasets. SGG, SGA, VQA, VC, and RR represent Scene Graph Generation, Scene Graph Anticipation, Video Question Answering, Video Captioning, and Relation Reasoning." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 101, + 557, + 167 + ], + "lines": [ + { + "bbox": [ + 313, + 101, + 557, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 101, + 557, + 167 + ], + "type": "table", + "html": "
Dataset#FramesTasksAnnotations
SGGSGAVQAVCRRBboxRelationText
SportsHHI [47]11.4K✓XXXX✓✓X
Action Genome [18]234.3K✓✓XXX✓✓X
AeroEye [35]261.5K✓XXXX✓✓X
ASPIRe [34]1.6M✓XXXX✓✓X
PVSG [52]153K✓X✓✓X✓✓✓
VSGR (Ours)1.9M✓✓✓✓✓✓✓✓
", + "image_path": "e24b450fbb708082214c556f322598cd7a6660f62faa37051563a4f518dc227c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 177, + 555, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 177, + 555, + 261 + ], + "spans": [ + { + "bbox": [ + 313, + 177, + 555, + 261 + ], + "type": "text", + "content": "subsequent interactions by identifying new relationships and adapting to evolving contexts using transition probabilities in " + }, + { + "bbox": [ + 313, + 177, + 555, + 261 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 313, + 177, + 555, + 261 + ], + "type": "text", + "content": ", anticipating future relationships " + }, + { + "bbox": [ + 313, + 177, + 555, + 261 + ], + "type": "inline_equation", + "content": "r_{t+n}^{ij}" + }, + { + "bbox": [ + 313, + 177, + 555, + 261 + ], + "type": "text", + "content": " and improving conversational flow. Finally, the hypothesis allows the model to propose and test conjectures about the underlying themes or intentions, synthesizing information from previous steps and leveraging " + }, + { + "bbox": [ + 313, + 177, + 555, + 261 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 313, + 177, + 555, + 261 + ], + "type": "text", + "content": " to explain patterns and forecast results." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 272, + 526, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 272, + 526, + 285 + ], + "spans": [ + { + "bbox": [ + 313, + 272, + 526, + 285 + ], + "type": "text", + "content": "5. Video Scene Graph Reasoning Dataset" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 292, + 556, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 292, + 556, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 292, + 556, + 389 + ], + "type": "text", + "content": "Limitations of Current Datasets. Existing benchmarks (see Table 1) primarily support SGG and SGA, limiting their applicability to reasoning tasks. They suffer from inadequate support for VQA, VC, and RR, shallow annotations that fail to capture intricate object interactions, insufficient modeling of temporal dynamics, and poor multimodal integration. Thus, our VSGR dataset supports SGG, SGA, VQA, VC, and RR, enabling the reasoning capabilities of LLMs." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 396, + 437, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 437, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 437, + 407 + ], + "type": "text", + "content": "5.1. Dataset Construction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 414, + 556, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 414, + 556, + 546 + ], + "spans": [ + { + "bbox": [ + 313, + 414, + 556, + 546 + ], + "type": "text", + "content": "Data Acquisition Stage. We source videos from the ASPIRe [34] and AeroEye [35] datasets. While the ASPIRe dataset offers diverse, richly annotated videos emphasizing dynamic interactions and temporal changes, the AeroEye dataset offers drone-captured footage across various scenes. Comprehension Tasks via Question-Answering. We introduce tasks that leverage fine-grained relationships from scene graphs, extending Scene Graph Generation to focus on relation understanding and subject/object interpretation using triplets and leverage GPT-4/GPT-3.5 model for language generation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 546, + 556, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 556, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 556, + 641 + ], + "type": "text", + "content": "Video Captioning (VC). We generate 82,532 video-caption pairs, resulting in about 22 captions per video. Our process involves (1) extracting triplets from cropped video frames focusing on foreground objects, (2) generating background descriptions based on these triplets, and (3) combining the foreground triplets and background descriptions to produce captions. The average length is 893 characters, surpassing the PVSG [52] dataset in quantity and detail." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "Video Question Answering (VQA). We develop 74,856 question-answer pairs by designing questions that explore diverse relationships, selecting subject and object categories to verify specific relationships, or choosing triplets to assess their uniqueness. This results in an average of approximately 20 questions per video, which exceeds existing benchmarks" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29154" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 77, + 298, + 136 + ], + "blocks": [ + { + "bbox": [ + 59, + 77, + 298, + 136 + ], + "lines": [ + { + "bbox": [ + 59, + 77, + 298, + 136 + ], + "spans": [ + { + "bbox": [ + 59, + 77, + 298, + 136 + ], + "type": "image", + "image_path": "5a024d78c7f75b4f89f88278cb1bfc3c4464f8cca8a020c22486d42daafce9a5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 83, + 342, + 525, + 354 + ], + "lines": [ + { + "bbox": [ + 83, + 342, + 525, + 354 + ], + "spans": [ + { + "bbox": [ + 83, + 342, + 525, + 354 + ], + "type": "text", + "content": "Figure 5. An example of the diversified context within the streaming dialog in our VSGR dataset. Best viewed in color and zooming in." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 138, + 136, + 147 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 138, + 136, + 147 + ], + "spans": [ + { + "bbox": [ + 65, + 138, + 136, + 147 + ], + "type": "text", + "content": "Step-1: Generation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 148, + 291, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 148, + 291, + 173 + ], + "spans": [ + { + "bbox": [ + 66, + 148, + 291, + 173 + ], + "type": "text", + "content": "Given bounding boxes: person-1 [657, 116, 155, 577], person-2 [425, 211, 178, 450], ball-1 [534, 437, 42, 54], ball-2 [670, 350, 37, 65]. What are the relationships between these objects?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 177, + 291, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 177, + 291, + 195 + ], + "spans": [ + { + "bbox": [ + 66, + 177, + 291, + 195 + ], + "type": "text", + "content": "The video depicts including person-1 towards the right side and closer to ball-2. person-2 is situated to the left and is nearer to ball-1." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 65, + 208, + 187, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 208, + 187, + 217 + ], + "spans": [ + { + "bbox": [ + 65, + 208, + 187, + 217 + ], + "type": "text", + "content": "Step-2: Relationship Anticipation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 66, + 223, + 274, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 223, + 274, + 241 + ], + "spans": [ + { + "bbox": [ + 66, + 223, + 274, + 241 + ], + "type": "text", + "content": "What interactions might occur between the objects in the next three frame based on ?" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 247, + 291, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 247, + 291, + 265 + ], + "spans": [ + { + "bbox": [ + 67, + 247, + 291, + 265 + ], + "type": "text", + "content": "The person-1 will interact with ball-2 by picking it up or dribbling it. Similarly, person-2 may engage with ball-1 corresponding to" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 65, + 277, + 134, + 286 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 277, + 134, + 286 + ], + "spans": [ + { + "bbox": [ + 65, + 277, + 134, + 286 + ], + "type": "text", + "content": "Step-3: Reasoning" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 66, + 293, + 199, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 293, + 199, + 301 + ], + "spans": [ + { + "bbox": [ + 66, + 293, + 199, + 301 + ], + "type": "text", + "content": "What purpose or context can be inferred within" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 66, + 308, + 292, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 308, + 292, + 334 + ], + "spans": [ + { + "bbox": [ + 66, + 308, + 292, + 334 + ], + "type": "text", + "content": "person-1 and person-2 work with ball-1 and ball-2, suggesting drills like dribbling or shooting. Their separation implies independent practice, but their proximity hints at a possible shift to coordinated team drills later in the session" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 315, + 74, + 389, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 74, + 389, + 83 + ], + "spans": [ + { + "bbox": [ + 315, + 74, + 389, + 83 + ], + "type": "text", + "content": "Step-4: Verification" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 317, + 87, + 542, + 105 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 87, + 542, + 105 + ], + "spans": [ + { + "bbox": [ + 317, + 87, + 542, + 105 + ], + "type": "text", + "content": "Does person-1 move towards ball-2 and begin dribbling, and does person-2 start practicing shots with ball-1 through ?" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 317, + 110, + 542, + 134 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 110, + 542, + 134 + ], + "spans": [ + { + "bbox": [ + 317, + 110, + 542, + 134 + ], + "type": "text", + "content": "Yes, both individuals are focused on separate balls. person-1 continues with ball-1, and person-2 remains with ball-2, indicating individual practice rather than switching activities or interacting with each other." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 140, + 394, + 149 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 394, + 149 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 394, + 149 + ], + "type": "text", + "content": "Step-5: Clarification" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 318, + 151, + 544, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 151, + 544, + 167 + ], + "spans": [ + { + "bbox": [ + 318, + 151, + 544, + 167 + ], + "type": "text", + "content": "If instead, person-1 picks up ball-2 and passes it to person-2, how would that change the dynamics of?" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 318, + 168, + 544, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 168, + 544, + 200 + ], + "spans": [ + { + "bbox": [ + 318, + 168, + 544, + 200 + ], + "type": "text", + "content": "This would shift the dynamics from individual practice to a collaborative activity, with person-1 passing ball-2 to person-2. The interaction would signal the start of coordinated play, possibly initiating a game involving both ball-1 and ball-2 and highlighting teamwork between person-1 and person-2." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 318, + 208, + 417, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 208, + 417, + 217 + ], + "spans": [ + { + "bbox": [ + 318, + 208, + 417, + 217 + ], + "type": "text", + "content": "Step-6: Scene Forecasting" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 219, + 544, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 219, + 544, + 237 + ], + "spans": [ + { + "bbox": [ + 320, + 219, + 544, + 237 + ], + "type": "text", + "content": "Given , what are the likely relationship changes between objects in the upcoming phase?" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 319, + 242, + 544, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 242, + 544, + 266 + ], + "spans": [ + { + "bbox": [ + 319, + 242, + 544, + 266 + ], + "type": "text", + "content": "The next action could involve person-1 passing ball-2 to person-2 while person-2 passes ball-1 to person-1. Alternatively, person-1 might dribble ball-2 toward person-2, initiating a one-on-one game." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 318, + 277, + 392, + 287 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 277, + 392, + 287 + ], + "spans": [ + { + "bbox": [ + 318, + 277, + 392, + 287 + ], + "type": "text", + "content": "Step-7: Hypothesis" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 319, + 290, + 544, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 290, + 544, + 307 + ], + "spans": [ + { + "bbox": [ + 319, + 290, + 544, + 307 + ], + "type": "text", + "content": "What if person-3 [600, 200, 160, 500] enters and approaches the others? How might that affect the interactions?" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 312, + 545, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 312, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 320, + 312, + 545, + 335 + ], + "type": "text", + "content": "person-1, person-2, and person-3 could engage in coordinated actions, such as three-person passing drills using ball-1 and ball-2. The interaction would evolve from individual practice to collaborative teamwork." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 55, + 362, + 296, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 362, + 296, + 374 + ], + "spans": [ + { + "bbox": [ + 55, + 362, + 296, + 374 + ], + "type": "text", + "content": "such as the MSRVTT-QA [49] and MSVD-QA [48] datasets." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 54, + 376, + 296, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 376, + 296, + 604 + ], + "spans": [ + { + "bbox": [ + 54, + 376, + 296, + 604 + ], + "type": "text", + "content": "Relation Reasoning (RR). Using the annotated scene graphs, we produce 61,120 relation reasoning tasks by selecting partial information as an incomplete input. Each task requires the model to deduce relationships among entities and identify the categories of the subject and object. With an average of approximately 16 tasks per video, our VSGR dataset provides a substantial collection for evaluating models' abilities in relational reasoning and scene understanding. Question and Answer Validation. To ensure the quality and complexity of questions, we implement a rigorous validation process that combines generation by LLMs with human refinement. Initially, GPT-4/GPT-3.5 generates queries based on the triplets extracted from the videos. Human annotators are trained with specific guidelines that emphasize clarity, relevance, and appropriate challenge levels, then review and refine questions. They enhance the questions by ensuring they are directly answerable from the video content and require careful reasoning about the depicted interactions and relationships." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 54, + 605, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 605, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 605, + 296, + 713 + ], + "type": "text", + "content": "We apply strict filtering criteria to improve the quality of the dataset. First, we eliminate questions that do not require video context and can be answered using general world knowledge, ensuring that models must rely on visual information from the videos. Second, we exclude questions that LLMs can answer correctly, increasing the challenge and utility of the dataset in evaluating advanced reasoning abilities. Finally, independent annotators perform a second round to review and evaluate the quality of the refined questions." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 361, + 432, + 375 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 432, + 375 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 432, + 375 + ], + "type": "text", + "content": "5.2. Dataset Comparison" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 381, + 556, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 381, + 556, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 381, + 556, + 525 + ], + "type": "text", + "content": "As reported in Table 1, our VSGR dataset represents a substantial advancement in video scene graph benchmarks. Our dataset comprises 3,748 videos and 1,841,243 frames, surpassing existing datasets in scale. Unlike other datasets that address only a limited subset of tasks, our dataset offers comprehensive task coverage, facilitating multifaceted evaluations of LLMs. In addition, our ground truth enriches relation annotations with comprehensive textual descriptions, enabling sophisticated reasoning and relationship predictions, as illustrated in Fig. 5. Additionally, our VSGR dataset incorporates diverse viewpoints, including third-person, ego-centric, and drone perspectives, enhancing its generalization." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 540, + 430, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 430, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 430, + 553 + ], + "type": "text", + "content": "6. Experiment Results" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 560, + 446, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 560, + 446, + 573 + ], + "spans": [ + { + "bbox": [ + 313, + 560, + 446, + 573 + ], + "type": "text", + "content": "6.1. Implementation Details" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 580, + 555, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 580, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 580, + 555, + 616 + ], + "type": "text", + "content": "Datasets. We leverage our VSGR dataset across five tasks. In addition, we utilize the PVSG [52] dataset for the SGG task and the Action Genome [18] dataset for the SGA task." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "content": "Model Configuration. We operate the CLIP-ViT-L-336 [8, 37] to encode each video frame into ten tokens (one CLS token and nine from " + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "content": " average pooling). These tokens are fed into a two-layer MLP connector to the Mistral7B-Instruct [19] language model. For training, we apply LoRA [16] to all linear layers with a rank of 128 and a scaling factor of 256, omitting vision-language alignment [31]. We train for two epochs with a batch size of 128 over 16" + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29155" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 175, + 145 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 175, + 145 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 175, + 145 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 175, + 145 + ], + "type": "image", + "image_path": "d0a01bfd9fa68d49d6cdd8ac7d69f94218428a12f7d82e080192790acb0992a8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 147, + 295, + 166 + ], + "lines": [ + { + "bbox": [ + 55, + 147, + 295, + 166 + ], + "spans": [ + { + "bbox": [ + 55, + 147, + 295, + 166 + ], + "type": "text", + "content": "Figure 6. Comparison of Recall (R) and mean Recall (mR) at different numbers of hyperedges for the SGG task on the VSGR dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 175, + 70, + 293, + 145 + ], + "blocks": [ + { + "bbox": [ + 175, + 70, + 293, + 145 + ], + "lines": [ + { + "bbox": [ + 175, + 70, + 293, + 145 + ], + "spans": [ + { + "bbox": [ + 175, + 70, + 293, + 145 + ], + "type": "image", + "image_path": "046d8a22a37e1a284f7f5f8a29220ae77a2315b14a16d2f5c3501b83a8f87b6d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 56, + 194, + 296, + 371 + ], + "blocks": [ + { + "bbox": [ + 55, + 173, + 295, + 193 + ], + "lines": [ + { + "bbox": [ + 55, + 173, + 295, + 193 + ], + "spans": [ + { + "bbox": [ + 55, + 173, + 295, + 193 + ], + "type": "text", + "content": "Table 2. Comparison (\\%) on the VSGR and Action Genome datasets for the Scene Graph Anticipation (SGA) task at varying video input fractions " + }, + { + "bbox": [ + 55, + 173, + 295, + 193 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 55, + 173, + 295, + 193 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 194, + 296, + 371 + ], + "lines": [ + { + "bbox": [ + 56, + 194, + 296, + 371 + ], + "spans": [ + { + "bbox": [ + 56, + 194, + 296, + 371 + ], + "type": "table", + "html": "
FMethodAction GenomeVSGR
R/mR@10R/mR@20R/mR@50R/mR@10R/mR@20R/mR@50
0.3STTran+ [5]13.9 / 3.521.6 / 7.340.8 / 20.312.0 / 4.019.0 / 8.035.7 / 18.0
DSGDetr+ [11]14.3 / 3.621.8 / 7.641.3 / 21.212.5 / 4.219.5 / 8.336.0 / 19.0
STTran++ [5]15.4 / 6.227.2 / 14.148.6 / 32.214.0 / 6.522.0 / 11.039.0 / 25.1
DSGDetr++ [11]16.8 / 8.429.0 / 16.748.9 / 32.314.5 / 7.022.5 / 12.040.0 / 26.0
SceneSayerODE [36]23.3 / 13.332.5 / 20.145.1 / 33.018.0 / 9.526.0 / 15.042.0 / 30.0
SceneSayerSDE [36]25.9 / 15.635.7 / 23.147.4 / 37.119.5 / 11.027.5 / 17.044.0 / 33.5
HyperGraph (Ours)26.5 / 14.836.2 / 22.149.3 / 37.218.8 / 9.327.2 / 16.342.5 / 27.4
HyperGLM (Ours)27.5 / 15.837.0 / 24.550.0 / 38.019.0 / 10.028.0 / 16.543.0 / 27.5
0.5STTran+ [5]14.9 / 3.722.6 / 7.642.9 / 21.413.5 / 4.221.0 / 8.537.0 / 19.0
DSGDetr+ [11]15.2 / 3.923.1 / 8.043.3 / 22.213.8 / 4.521.5 / 9.038.0 / 20.0
STTran++ [5]16.6 / 6.629.1 / 14.751.5 / 33.415.5 / 7.023.5 / 12.541.0 / 26.5
DSGDetr++ [11]17.4 / 8.430.5 / 17.051.9 / 33.916.0 / 7.524.0 / 13.042.0 / 28.0
SceneSayerODE [36]26.4 / 14.336.6 / 21.449.8 / 36.020.5 / 11.029.5 / 16.546.1 / 32.5
SceneSayerSDE [36]28.4 / 16.338.6 / 25.151.4 / 39.921.5 / 12.531.0 / 18.548.0 / 35.7
HyperGraph (Ours)29.2 / 16.439.3 / 23.252.1 / 38.720.3 / 12.229.8 / 17.146.2 / 33.1
HyperGLM (Ours)30.0 / 17.040.5 / 27.553.5 / 40.521.5 / 11.531.5 / 18.546.5 / 30.0
0.7STTran+ [5]16.6 / 4.225.1 / 8.547.2 / 24.015.0 / 5.024.0 / 10.541.0 / 21.5
DSGDetr+ [11]16.8 / 4.325.3 / 8.847.4 / 24.715.5 / 5.224.5 / 11.042.0 / 22.0
STTran++ [5]19.0 / 7.732.8 / 17.156.8 / 36.817.0 / 8.027.0 / 14.045.0 / 29.0
DSGDetr++ [11]19.8 / 9.534.1 / 19.256.7 / 37.217.5 / 8.528.0 / 15.046.1 / 30.0
SceneSayerODE [36]32.1 / 16.542.8 / 24.455.6 / 39.623.5 / 13.033.5 / 19.051.0 / 36.0
SceneSayerSDE [36]33.3 / 18.144.0 / 27.356.4 / 44.424.5 / 14.535.7 / 21.053.0 / 38.0
HyperGraph (Ours)34.3 / 19.245.2 / 25.357.2 / 42.122.3 / 13.234.2 / 19.350.4 / 35.3
HyperGLM (Ours)35.7 / 19.546.1 / 30.058.2 / 44.025.1 / 13.535.5 / 21.551.0 / 33.5
", + "image_path": "25e240ff134a120ad587ba8710e00a8ea98ea68745151793b640854a22b90a87.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 383, + 283, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 383, + 283, + 395 + ], + "spans": [ + { + "bbox": [ + 55, + 383, + 283, + 395 + ], + "type": "text", + "content": "iterations on " + }, + { + "bbox": [ + 55, + 383, + 283, + 395 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 55, + 383, + 283, + 395 + ], + "type": "text", + "content": " GPUs, taking approximately six hours." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 396, + 296, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 296, + 529 + ], + "type": "text", + "content": "Metrics. We evaluate the SGG and SGA tasks using the Recall and mean Recall scores. In addition, we evaluate the VQA and RR tasks by Accuracy, Precision, Recall, and F1 scores. For the VC task, we utilize CIDEr, MENTOR, ROUGE-L, and BLEU-4 scores to validate our performance. Settings. For the SGG and SGA tasks, we adopt the evaluation settings based on [18, 36]. In particular, the model is provided with raw video frames and must detect objects using a pre-trained detector (i.e. Faster R-CNN) and predict or anticipate their relationships. Especially for the SGA, we set the initial video input fraction " + }, + { + "bbox": [ + 55, + 396, + 296, + 529 + ], + "type": "inline_equation", + "content": "(\\mathcal{F})" + }, + { + "bbox": [ + 55, + 396, + 296, + 529 + ], + "type": "text", + "content": " to 0.9, following [36]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 538, + 149, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 538, + 149, + 551 + ], + "spans": [ + { + "bbox": [ + 55, + 538, + 149, + 551 + ], + "type": "text", + "content": "6.2. Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 557, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 297, + 713 + ], + "type": "text", + "content": "Hypergraph Parameters. The number of walks " + }, + { + "bbox": [ + 55, + 557, + 297, + 713 + ], + "type": "inline_equation", + "content": "(N_w)" + }, + { + "bbox": [ + 55, + 557, + 297, + 713 + ], + "type": "text", + "content": " and walk length " + }, + { + "bbox": [ + 55, + 557, + 297, + 713 + ], + "type": "inline_equation", + "content": "(N_l)" + }, + { + "bbox": [ + 55, + 557, + 297, + 713 + ], + "type": "text", + "content": " are introduced in Alg. 1, directly impacting the capacity to capture high-order relationships by determining the number of hyperedges. Although more hyperedges increase relational diversity, they can also introduce redundancy beyond an optimal point. Specifically, a higher " + }, + { + "bbox": [ + 55, + 557, + 297, + 713 + ], + "type": "inline_equation", + "content": "N_w" + }, + { + "bbox": [ + 55, + 557, + 297, + 713 + ], + "type": "text", + "content": " broadens the range of sampled relationships, while a moderate " + }, + { + "bbox": [ + 55, + 557, + 297, + 713 + ], + "type": "inline_equation", + "content": "N_l" + }, + { + "bbox": [ + 55, + 557, + 297, + 713 + ], + "type": "text", + "content": " balances depth. Our experiments demonstrate that optimal performances are achieved with 60 hyperedges " + }, + { + "bbox": [ + 55, + 557, + 297, + 713 + ], + "type": "inline_equation", + "content": "(N_w = 60" + }, + { + "bbox": [ + 55, + 557, + 297, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 557, + 297, + 713 + ], + "type": "inline_equation", + "content": "N_l = 7)" + }, + { + "bbox": [ + 55, + 557, + 297, + 713 + ], + "type": "text", + "content": ", shown in Fig. 6. Further experiments on the SGG task are provided in Table 3, and experiments on these parameters are included in the Appendices. Video Input Fraction. We adjust the initial video input" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 313, + 91, + 555, + 147 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 555, + 89 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 555, + 89 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 555, + 89 + ], + "type": "text", + "content": "Table 3. Comparison " + }, + { + "bbox": [ + 313, + 70, + 555, + 89 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 313, + 70, + 555, + 89 + ], + "type": "text", + "content": " on the VSGR and PVSG datasets for the Scene Graph Generation (SGG) task at Recall (R) and mean Recall (mR)." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 91, + 555, + 147 + ], + "lines": [ + { + "bbox": [ + 313, + 91, + 555, + 147 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 555, + 147 + ], + "type": "table", + "html": "
MethodPVSGVSGR
R/mR@20R/mR@50R/mR@100R/mR@20R/mR@50R/mR@100
Transformer [52]4.0 / 1.84.4 / 1.94.9 / 2.025.7 / 6.334.5 / 6.543.5 / 7.0
HIG [34]4.6 / 1.94.9 / 2.15.4 / 2.223.8 / 5.731.1 / 5.940.4 / 6.9
CYCLO [35]5.8 / 2.06.1 / 2.26.7 / 2.329.4 / 7.136.4 / 7.747.7 / 7.7
HyperGraph (Ours)6.5 / 2.27.0 / 2.47.5 / 2.631.6 / 7.838.8 / 8.350.3 / 8.5
HyperGLM (Ours)7.5 / 2.88.1 / 3.78.5 / 3.935.8 / 9.242.3 / 10.154.7 / 10.4
", + "image_path": "647a19bc53823d4bbb847b4c5b9ec0b065fbffcde419a7e7f402c11e182d80d4.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 314, + 174, + 555, + 251 + ], + "blocks": [ + { + "bbox": [ + 313, + 153, + 555, + 172 + ], + "lines": [ + { + "bbox": [ + 313, + 153, + 555, + 172 + ], + "spans": [ + { + "bbox": [ + 313, + 153, + 555, + 172 + ], + "type": "text", + "content": "Table 4. Comparison (%) on the VSGR and Action Genome datasets for the Scene Graph Anticipation (SGA) task at Recall (R) and mean Recall (mR)." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 314, + 174, + 555, + 251 + ], + "lines": [ + { + "bbox": [ + 314, + 174, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 314, + 174, + 555, + 251 + ], + "type": "table", + "html": "
MethodAction GenomeVSGR
R/mR@10R/mR@20R/mR@50R/mR@10R/mR@20R/mR@50
STTran+ [5]17.5 / 4.626.8 / 9.249.6 / 24.316.5 / 5.526.0 / 11.043.0 / 23.0
DSGDetr+ [11]17.9 / 4.727.7 / 9.751.4 / 25.917.0 / 6.027.0 / 11.544.5 / 24.5
STTran++ [5]20.2 / 8.935.7 / 18.460.2 / 38.819.0 / 9.530.0 / 15.549.0 / 31.0
DSGDetr++ [11]22.2 / 11.437.1 / 21.061.0 / 39.519.5 / 10.031.0 / 16.050.0 / 32.5
SceneSayerODE [36]36.6 / 17.848.3 / 27.461.3 / 43.426.5 / 14.037.5 / 20.055.5 / 38.0
SceneSayerSDE [36]37.3 / 20.848.6 / 30.961.6 / 46.827.5 / 16.038.5 / 22.058.2 / 40.0
HyperGraph (Ours)37.5 / 19.149.3 / 31.462.3 / 47.428.4 / 17.539.3 / 22.457.6 / 41.5
HyperGLM (Ours)38.8 / 22.351.5 / 33.065.2 / 48.630.2 / 18.141.1 / 23.559.3 / 43.4
", + "image_path": "f484caec26019d8937035ef7f7c1b1a6e6cabe23bb3ef410756ccc6c8388f6ba.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 255, + 556, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 255, + 556, + 363 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 556, + 363 + ], + "type": "text", + "content": "fraction, " + }, + { + "bbox": [ + 313, + 255, + 556, + 363 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 313, + 255, + 556, + 363 + ], + "type": "text", + "content": ", to 0.3, 0.5, and 0.7 for the SGA task. This adjustment allows the model to learn from varying observed portions and predict the unseen segment. Table 2 indicates that increasing the portion of the seen video improves performance, suggesting that additional visual context is beneficial. In addition, Table 4 further confirms that the default setting at a higher input fraction (" + }, + { + "bbox": [ + 313, + 255, + 556, + 363 + ], + "type": "inline_equation", + "content": "F = 0.9" + }, + { + "bbox": [ + 313, + 255, + 556, + 363 + ], + "type": "text", + "content": ") leads to optimal performance. We also present additional settings with varying video input fractions for the SGA task in the Appendices." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 371, + 500, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 371, + 500, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 500, + 384 + ], + "type": "text", + "content": "6.3. Comparison with State-of-the-Arts" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 389, + 556, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 556, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 556, + 605 + ], + "type": "text", + "content": "Table 3 demonstrates that HyperGLM significantly outperforms existing methods on the PVSG and VSGR datasets, achieving the highest " + }, + { + "bbox": [ + 313, + 389, + 556, + 605 + ], + "type": "inline_equation", + "content": "R@20" + }, + { + "bbox": [ + 313, + 389, + 556, + 605 + ], + "type": "text", + "content": " scores of " + }, + { + "bbox": [ + 313, + 389, + 556, + 605 + ], + "type": "inline_equation", + "content": "7.5\\%" + }, + { + "bbox": [ + 313, + 389, + 556, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 389, + 556, + 605 + ], + "type": "inline_equation", + "content": "35.8\\%" + }, + { + "bbox": [ + 313, + 389, + 556, + 605 + ], + "type": "text", + "content": ", respectively. By leveraging hyperedges connecting multiple nodes within the HyperGraph, HyperGLM effectively captures complex object interactions and spatial dependencies, transcending traditional pairwise methods to generate more accurate and detailed scene graph representations. Notably, integrating a procedural graph within the HyperGraph reduces bias. Our approach enhances mean Recall, reaching improvements of " + }, + { + "bbox": [ + 313, + 389, + 556, + 605 + ], + "type": "inline_equation", + "content": "2.8\\%" + }, + { + "bbox": [ + 313, + 389, + 556, + 605 + ], + "type": "text", + "content": " on the PVSG dataset and " + }, + { + "bbox": [ + 313, + 389, + 556, + 605 + ], + "type": "inline_equation", + "content": "9.2\\%" + }, + { + "bbox": [ + 313, + 389, + 556, + 605 + ], + "type": "text", + "content": " on the VSGR dataset, thereby addressing the long-tail distribution challenges that have struggled in previous methods [34, 35, 52]. Furthermore, the LLM enhances the capacity of HyperGLM to infer and predict intricate relationships embedded within the unified HyperGraph, resulting in improved performance compared to HyperGraph, which shows a decrease of " + }, + { + "bbox": [ + 313, + 389, + 556, + 605 + ], + "type": "inline_equation", + "content": "4.2\\%" + }, + { + "bbox": [ + 313, + 389, + 556, + 605 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 313, + 389, + 556, + 605 + ], + "type": "inline_equation", + "content": "R@20" + }, + { + "bbox": [ + 313, + 389, + 556, + 605 + ], + "type": "text", + "content": " on the VSGR dataset." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 605, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 557, + 715 + ], + "type": "text", + "content": "As shown in Table 4, our HyperGLM approach outperforms existing SGA methods on the Action Genome and VSGR datasets, achieving R@10 scores of " + }, + { + "bbox": [ + 313, + 605, + 557, + 715 + ], + "type": "inline_equation", + "content": "35.7\\%" + }, + { + "bbox": [ + 313, + 605, + 557, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 605, + 557, + 715 + ], + "type": "inline_equation", + "content": "25.1\\%" + }, + { + "bbox": [ + 313, + 605, + 557, + 715 + ], + "type": "text", + "content": ", respectively. This improvement stems from integrating procedural graphs that model causal relationships within the HyperGraph structure. In contrast, the SceneSayer [36] method relies on NeuralODE and NeuralSDE to capture the latent dynamics of object interaction evolution. Our procedural graphs enable multi-step transitions that explicitly" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29156" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 64, + 59, + 306, + 137 + ], + "blocks": [ + { + "bbox": [ + 64, + 59, + 306, + 137 + ], + "lines": [ + { + "bbox": [ + 64, + 59, + 306, + 137 + ], + "spans": [ + { + "bbox": [ + 64, + 59, + 306, + 137 + ], + "type": "image", + "image_path": "c3a87d810387a52aaf9465f5d76b7bd91f7cb2b28781fde1fd2757228c01145d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 307, + 59, + 548, + 137 + ], + "blocks": [ + { + "bbox": [ + 307, + 59, + 548, + 137 + ], + "lines": [ + { + "bbox": [ + 307, + 59, + 548, + 137 + ], + "spans": [ + { + "bbox": [ + 307, + 59, + 548, + 137 + ], + "type": "image", + "image_path": "2e9f9ce2e537e4124e4e155635084f5750b6183684a70cba399a13624e2c0002.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 56, + 140, + 305, + 303 + ], + "blocks": [ + { + "bbox": [ + 56, + 140, + 305, + 303 + ], + "lines": [ + { + "bbox": [ + 56, + 140, + 305, + 303 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 305, + 303 + ], + "type": "image", + "image_path": "3389de1b55da6d5d50ae1f3cccb1c8e11ba6b50c06de7816a66ea78b023b6009.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 136, + 304, + 227, + 312 + ], + "lines": [ + { + "bbox": [ + 136, + 304, + 227, + 312 + ], + "spans": [ + { + "bbox": [ + 136, + 304, + 227, + 312 + ], + "type": "text", + "content": "a) Scene Graph Generation" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 306, + 140, + 552, + 303 + ], + "blocks": [ + { + "bbox": [ + 306, + 140, + 552, + 303 + ], + "lines": [ + { + "bbox": [ + 306, + 140, + 552, + 303 + ], + "spans": [ + { + "bbox": [ + 306, + 140, + 552, + 303 + ], + "type": "image", + "image_path": "a4edd7237696b72899c03efdbebcea8fcf847cef256b9ec964a8af052a749164.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 378, + 304, + 473, + 312 + ], + "lines": [ + { + "bbox": [ + 378, + 304, + 473, + 312 + ], + "spans": [ + { + "bbox": [ + 378, + 304, + 473, + 312 + ], + "type": "text", + "content": "b) Scene Graph Anticipation" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 313, + 555, + 342 + ], + "lines": [ + { + "bbox": [ + 55, + 313, + 555, + 342 + ], + "spans": [ + { + "bbox": [ + 55, + 313, + 555, + 342 + ], + "type": "text", + "content": "Figure 7. Qualitative comparison of our HyperGLM approach versus SceneSayerSDE [36] for SGG and SGA. Red and green edge labels denote incorrect and correct predictions, respectively. Our HyperGLM approach effectively captures the evolving interactions between person-1 and bicycle-1 or person-2 and person-3 over time and anticipates interactions in unseen video frames, while SceneSayerSDE confuses to similar predicates. Best viewed in color." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 56, + 355, + 294, + 424 + ], + "blocks": [ + { + "bbox": [ + 59, + 344, + 291, + 354 + ], + "lines": [ + { + "bbox": [ + 59, + 344, + 291, + 354 + ], + "spans": [ + { + "bbox": [ + 59, + 344, + 291, + 354 + ], + "type": "text", + "content": "Table 5. Comparison (%) on VSGR for the Video Question Answering." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 355, + 294, + 424 + ], + "lines": [ + { + "bbox": [ + 56, + 355, + 294, + 424 + ], + "spans": [ + { + "bbox": [ + 56, + 355, + 294, + 424 + ], + "type": "table", + "html": "
MethodAccuracyPrecisionRecallF1 Score
Video-ChatGPT [32]33.235.132.333.6
Video-LLaVA-7B [30]43.143.841.742.7
MovieChat [41]43.544.242.643.4
Chat-UniVi-7B [20]44.345.643.244.4
HyperGLM (Ours)45.447.244.345.7
", + "image_path": "8acbf083e8546b004500d65226cbdb3f4a7b578a36f96feb499bc1c332877d51.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 56, + 435, + 295, + 494 + ], + "blocks": [ + { + "bbox": [ + 75, + 425, + 275, + 434 + ], + "lines": [ + { + "bbox": [ + 75, + 425, + 275, + 434 + ], + "spans": [ + { + "bbox": [ + 75, + 425, + 275, + 434 + ], + "type": "text", + "content": "Table 6. Comparison (%) on VSGR for the Video Captioning." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 435, + 295, + 494 + ], + "lines": [ + { + "bbox": [ + 56, + 435, + 295, + 494 + ], + "spans": [ + { + "bbox": [ + 56, + 435, + 295, + 494 + ], + "type": "table", + "html": "
MethodCIDErMENTORROUGE-LBLEU-4
MV-GPT [38]57.137.562.547.2
CoCap [40]54.329.461.842.8
UniVL + MELTR [24]50.528.160.042.1
HyperGLM (Ours)54.530.764.948.8
", + "image_path": "4a22900c91ca4021cd26016cb333c5a413bfd50478f3163afdc9fdc811ef2663.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 498, + 295, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 593 + ], + "type": "text", + "content": "capture the temporal evolution of relationships by modeling sequential interactions and their dependencies. In contrast, NeuralODE and NeuralSDE primarily focus on continuous-time dynamics, which may limit their effectiveness in handling discrete, multi-step relational changes as illustrated in Fig. 7b. Significantly, the HyperGraph is injected into the LLM, allowing the model to capture complex temporal patterns, resulting in better results than the HyperGraph." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": "In addition to the improvements in SGG and SGA shown in Fig. 7, Tables 5, 6, and 7 demonstrate the significant improvements using our HyperGLM approach in the VQA, VC, and RR tasks. In VQA, the hyperedges connect multiple objects, enabling HyperGLM to capture context-rich interactions more effectively, reaching an accuracy of " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "45.4\\%" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": ". For VC, the HyperGraph structure supports evolving object connections over time, allowing our HyperGLM approach to generate captions that describe the relationship between objects and capture the temporal flow of interactions coher-" + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 314, + 355, + 554, + 416 + ], + "blocks": [ + { + "bbox": [ + 330, + 344, + 537, + 354 + ], + "lines": [ + { + "bbox": [ + 330, + 344, + 537, + 354 + ], + "spans": [ + { + "bbox": [ + 330, + 344, + 537, + 354 + ], + "type": "text", + "content": "Table 7. Comparison (%) on VSGR for the Relation Reasoning." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 314, + 355, + 554, + 416 + ], + "lines": [ + { + "bbox": [ + 314, + 355, + 554, + 416 + ], + "spans": [ + { + "bbox": [ + 314, + 355, + 554, + 416 + ], + "type": "table", + "html": "
MethodAccuracyPrecisionRecallF1 Score
Video-LLaVA-7B [30]41.342.540.241.3
MA-LMM [14]42.843.741.842.7
LLaMA-VID-7B [29]44.145.243.544.3
HyperGLM (Ours)47.248.446.547.4
", + "image_path": "ef56cb900e55844f8cc8069c30b1dd4e47d88bb563910d1fb274c5c2b44396c9.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 430, + 555, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 555, + 478 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 555, + 478 + ], + "type": "text", + "content": "ently, achieving scores of " + }, + { + "bbox": [ + 313, + 430, + 555, + 478 + ], + "type": "inline_equation", + "content": "54.5\\%" + }, + { + "bbox": [ + 313, + 430, + 555, + 478 + ], + "type": "text", + "content": " at CIDEr and " + }, + { + "bbox": [ + 313, + 430, + 555, + 478 + ], + "type": "inline_equation", + "content": "64.9\\%" + }, + { + "bbox": [ + 313, + 430, + 555, + 478 + ], + "type": "text", + "content": " at ROUGE-L. In RR task, the HyperGraph effectively manages intricate dependencies between multiple objects, resulting in precise relational inferences with an accuracy of " + }, + { + "bbox": [ + 313, + 430, + 555, + 478 + ], + "type": "inline_equation", + "content": "47.2\\%" + }, + { + "bbox": [ + 313, + 430, + 555, + 478 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 495, + 388, + 507 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 495, + 388, + 507 + ], + "spans": [ + { + "bbox": [ + 314, + 495, + 388, + 507 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 517, + 555, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 555, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 555, + 601 + ], + "type": "text", + "content": "In this paper, we have introduced HyperGLM, a novel VidSGG method that integrates scene hypergraph information into LLMs for context-aware and precise scene interpretation. Our approach effectively models complex interactions and higher-order relationships. It outperforms leading methods on benchmarks, including PVSG, Action Genome, and our newly collected VSGR dataset across five tasks." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 603, + 555, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 603, + 555, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 603, + 555, + 663 + ], + "type": "text", + "content": "Limitations. Although our HyperGLM approach effectively models multi-way interactions, managing many objects and their interactions can complicate relationship structures. As the HyperGraph expands, essential relationships may become obscured, reducing the clarity of scene interpretation." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": "Acknowledgment. This material is based upon work supported by the National Science Foundation under Award No. OIA-1946391. We also acknowledge the Arkansas High-Performance Computing Center for providing GPUs." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29157" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 705 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 122 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 122 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 122 + ], + "type": "text", + "content": "[1] Song Bai, Feihu Zhang, and Philip HS Torr. Hypergraph convolution and hypergraph attention. Pattern Recognition, 110:107637, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 123, + 296, + 188 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 123, + 296, + 188 + ], + "spans": [ + { + "bbox": [ + 61, + 123, + 296, + 188 + ], + "type": "text", + "content": "[2] Joya Chen, Zhaoyang Lv, Shiwei Wu, Kevin Qinghong Lin, Chenan Song, Difei Gao, Jia-Wei Liu, Ziteng Gao, Dongxing Mao, and Mike Zheng Shou. Videollm-online: Online video large language model for streaming video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18407-18418, 2024. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 189, + 296, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 189, + 296, + 241 + ], + "spans": [ + { + "bbox": [ + 62, + 189, + 296, + 241 + ], + "type": "text", + "content": "[3] Zhanwen Chen, Saed Rezayi, and Sheng Li. More knowledge, less bias: Unbiasing scene graph generation with explicit ontological adjustment. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4023-4032, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 242, + 296, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 242, + 296, + 296 + ], + "spans": [ + { + "bbox": [ + 62, + 242, + 296, + 296 + ], + "type": "text", + "content": "[4] Yuren Cong, Wentong Liao, Hanno Ackermann, Bodo Rosenhahn, and Michael Ying Yang. Spatial-temporal transformer for dynamic scene graph generation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16372-16382, 2021. 1, 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 296, + 295, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 296, + 295, + 338 + ], + "spans": [ + { + "bbox": [ + 62, + 296, + 295, + 338 + ], + "type": "text", + "content": "[5] Yuren Cong, Wentong Liao, H. Ackermann, M. Yang, and B. Rosenhahn. Spatial-temporal transformer for dynamic scene graph generation. IEEE International Conference on Computer Vision, 2021. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 339, + 296, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 339, + 296, + 380 + ], + "spans": [ + { + "bbox": [ + 62, + 339, + 296, + 380 + ], + "type": "text", + "content": "[6] Yuren Cong, Michael Ying Yang, and Bodo Rosenhahn. Reltr: Relation transformer for scene graph generation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(9):11169-11183, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 381, + 296, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 381, + 296, + 435 + ], + "spans": [ + { + "bbox": [ + 62, + 381, + 296, + 435 + ], + "type": "text", + "content": "[7] Youming Deng, Yansheng Li, Yongjun Zhang, Xiang Xiang, Jian Wang, Jingdong Chen, and Jiayi Ma. Hierarchical memory learning for fine-grained scene graph generation. In European Conference on Computer Vision, pages 266-283. Springer, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 435, + 296, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 435, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 62, + 435, + 296, + 510 + ], + "type": "text", + "content": "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 511, + 296, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 511, + 296, + 564 + ], + "spans": [ + { + "bbox": [ + 62, + 511, + 296, + 564 + ], + "type": "text", + "content": "[9] Azade Farshad, Yousef Yeganeh, Yu Chi, Chengzhi Shen, Böjrn Ommer, and Nassir Navab. Scenegenie: Scene graph guided diffusion models for image synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 88-98, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 565, + 296, + 618 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 565, + 296, + 618 + ], + "spans": [ + { + "bbox": [ + 57, + 565, + 296, + 618 + ], + "type": "text", + "content": "[10] Hao Fei, Shengqiong Wu, Wei Ji, Hanwang Zhang, Meishan Zhang, Mong-Li Lee, and Wynne Hsu. Video-of-thought: Step-by-step video reasoning from perception to cognition. In *Forty-first International Conference on Machine Learning*, 2024. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 618, + 296, + 671 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 618, + 296, + 671 + ], + "spans": [ + { + "bbox": [ + 57, + 618, + 296, + 671 + ], + "type": "text", + "content": "[11] Shengyu Feng, Hesham Mostafa, Marcel Nassar, Somdeb Majumdar, and Subarna Tripathi. Exploiting long-term dependencies for generating dynamic scene graphs. IEEE Workshop/Winter Conference on Applications of Computer Vision, 2021. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 672, + 296, + 705 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 672, + 296, + 705 + ], + "spans": [ + { + "bbox": [ + 57, + 672, + 296, + 705 + ], + "type": "text", + "content": "[12] Yue Gao, Yifan Feng, Shuyi Ji, and Rongrong Ji. Hgnn+: General hypergraph neural networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):3181-3199," + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 708 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 335, + 73, + 365, + 83 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 365, + 83 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 365, + 83 + ], + "type": "text", + "content": "2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 83, + 555, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 83, + 555, + 126 + ], + "spans": [ + { + "bbox": [ + 316, + 83, + 555, + 126 + ], + "type": "text", + "content": "[13] Zeeshan Hayden and Xuming He. Dsgg: Dense relation transformer for an end-to-end scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28317-28326, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 126, + 555, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 126, + 555, + 191 + ], + "spans": [ + { + "bbox": [ + 316, + 126, + 555, + 191 + ], + "type": "text", + "content": "[14] Bo He, Hengduo Li, Young Kyun Jang, Menglin Jia, Xuefei Cao, Ashish Shah, Abhinav Shrivastava, and Ser-Nam Lim. Ma-lmm: Memory-augmented large multimodal model for long-term video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13504-13514, 2024. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 191, + 555, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 191, + 555, + 234 + ], + "spans": [ + { + "bbox": [ + 316, + 191, + 555, + 234 + ], + "type": "text", + "content": "[15] Tao He, Lianli Gao, Jingkuan Song, and Yuan-Fang Li. Towards open-vocabulary scene graph generation with prompt-based finetuning. In European Conference on Computer Vision, pages 56-73. Springer, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 234, + 555, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 234, + 555, + 287 + ], + "spans": [ + { + "bbox": [ + 316, + 234, + 555, + 287 + ], + "type": "text", + "content": "[16] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 288, + 554, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 288, + 554, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 288, + 554, + 342 + ], + "type": "text", + "content": "[17] Jinbae Im, Jeong Yeon Nam, Nokyung Park, Hyungmin Lee, and Seunghyun Park. Egtr: Extracting graph from transformer for scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24229-24238, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 342, + 555, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 555, + 396 + ], + "type": "text", + "content": "[18] Jingwei Ji, Ranjay Krishna, Li Fei-Fei, and Juan Carlos Niebles. Action genome: Actions as compositions of spatiotemporal scene graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10236–10247, 2020. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 396, + 555, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 396, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 316, + 396, + 555, + 449 + ], + "type": "text", + "content": "[19] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 449, + 555, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 555, + 513 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 555, + 513 + ], + "type": "text", + "content": "[20] Peng Jin, Ryuichi Takanobu, Wancai Zhang, Xiaochun Cao, and Li Yuan. Chat-univi: Unified visual representation empowers large language models with image and video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13700-13710, 2024. 1, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 514, + 555, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 514, + 555, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 514, + 555, + 578 + ], + "type": "text", + "content": "[21] Tianlei Jin, Fangtai Guo, Qiwei Meng, Shiqiang Zhu, Xiangming Xi, Wen Wang, Zonghao Mu, and Wei Song. Fast contextual scene graph generation with unbiased context augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6302-6311, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 578, + 554, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 578, + 554, + 632 + ], + "spans": [ + { + "bbox": [ + 316, + 578, + 554, + 632 + ], + "type": "text", + "content": "[22] Eun-Sol Kim, Woo Young Kang, Kyoung-Woon On, Yu-Jung Heo, and Byoung-Tak Zhang. Hypergraph attention networks for multimodal learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14581-14590, 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 632, + 555, + 696 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 632, + 555, + 696 + ], + "spans": [ + { + "bbox": [ + 316, + 632, + 555, + 696 + ], + "type": "text", + "content": "[23] Kibum Kim, Kanghoon Yoon, Jaehyeong Jeon, Yeunjun In, Jinyoung Moon, Donghyun Kim, and Chanyoung Park. Llm4sgg: Large language models for weakly supervised scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28306-28316, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 696, + 555, + 708 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 696, + 555, + 708 + ], + "spans": [ + { + "bbox": [ + 316, + 696, + 555, + 708 + ], + "type": "text", + "content": "[24] Dohwan Ko, Joonmyung Choi, Hyeong Kyu Choi, Kyoung-" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29158" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 707 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 126 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 126 + ], + "type": "text", + "content": "Woon On, Byungseok Roh, and Hyunwoo J Kim. Meltr: Meta loss transformer for learning to fine-tune video foundation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20105-20115, 2023. 1, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 127, + 296, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 127, + 296, + 170 + ], + "spans": [ + { + "bbox": [ + 56, + 127, + 296, + 170 + ], + "type": "text", + "content": "[25] Sanjoy Kundu and Sathyanarayanan N Aakur. Is-ggt: Iterative scene graph generation with generative transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6292-6301, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 170, + 296, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 170, + 296, + 213 + ], + "spans": [ + { + "bbox": [ + 56, + 170, + 296, + 213 + ], + "type": "text", + "content": "[26] Rongjie Li, Songyang Zhang, and Xuming He. Sgtr: End-to-end scene graph generation with transformer. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19486-19496, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 213, + 295, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 213, + 295, + 267 + ], + "spans": [ + { + "bbox": [ + 56, + 213, + 295, + 267 + ], + "type": "text", + "content": "[27] Rongjie Li, Songyang Zhang, Dahua Lin, Kai Chen, and Xuming He. From pixels to graphs: Open-vocabulary scene graph generation with vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28076-28086, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 267, + 295, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 267, + 295, + 310 + ], + "spans": [ + { + "bbox": [ + 56, + 267, + 295, + 310 + ], + "type": "text", + "content": "[28] Wanxin Li, Wei Xie, Zhigang Tu, Wei Wang, and Lianghao Jin. Multi-hyperedge hypergraph for group activity recognition. In 2022 International Joint Conference on Neural Networks (IJCNN), pages 01-07. IEEE, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 310, + 296, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 310, + 296, + 352 + ], + "spans": [ + { + "bbox": [ + 56, + 310, + 296, + 352 + ], + "type": "text", + "content": "[29] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In European Conference on Computer Vision, pages 323–340. Springer, 2025. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 353, + 295, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 353, + 295, + 395 + ], + "spans": [ + { + "bbox": [ + 56, + 353, + 295, + 395 + ], + "type": "text", + "content": "[30] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 1, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 395, + 295, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 395, + 295, + 427 + ], + "spans": [ + { + "bbox": [ + 56, + 395, + 295, + 427 + ], + "type": "text", + "content": "[31] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 427, + 296, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 427, + 296, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 427, + 296, + 514 + ], + "type": "text", + "content": "[32] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Khan. Video-ChatGPT: Towards detailed video understanding via large vision and language models. In Lun-Wei Ku, Andre Martins, and Vivek Srikumar, editors, Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 12585–12602, Bangkok, Thailand, Aug. 2024. Association for Computational Linguistics. 1, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 514, + 296, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 514, + 296, + 556 + ], + "spans": [ + { + "bbox": [ + 56, + 514, + 296, + 556 + ], + "type": "text", + "content": "[33] Sayak Nag, Kyle Min, Subarna Tripathi, and Amit K Roy-Chowdhury. Unbiased scene graph generation in videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22803-22813, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 556, + 296, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 556, + 296, + 609 + ], + "spans": [ + { + "bbox": [ + 56, + 556, + 296, + 609 + ], + "type": "text", + "content": "[34] Trong-Thuan Nguyen, Pha Nguyen, and Khoa Luu. Hig: Hierarchical interlacement graph approach to scene graph generation in video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 1, 2, 3, 5, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 610, + 295, + 664 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 610, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 56, + 610, + 295, + 664 + ], + "type": "text", + "content": "[35] Trong-Thuan Nguyen, Pha Nguyen, Li Xin, Cothren Jackson, Yilmaz Alper, and Khoa Luu. CYCLO: Cyclic graph transformer approach to multi-object relationship modeling in aerial videos. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 1, 2, 5, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 664, + 295, + 696 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 664, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 56, + 664, + 295, + 696 + ], + "type": "text", + "content": "[36] Rohith Peddi, Saksham Singh, Parag Singla, Vibhav Gogate, et al. Towards scene graph anticipation. In European Conference on Computer Vision. Springer, 2024. 1, 2, 3, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 696, + 295, + 707 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 696, + 295, + 707 + ], + "spans": [ + { + "bbox": [ + 56, + 696, + 295, + 707 + ], + "type": "text", + "content": "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 707 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 72, + 555, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 72, + 555, + 126 + ], + "spans": [ + { + "bbox": [ + 333, + 72, + 555, + 126 + ], + "type": "text", + "content": "Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 127, + 555, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 127, + 555, + 181 + ], + "spans": [ + { + "bbox": [ + 316, + 127, + 555, + 181 + ], + "type": "text", + "content": "[38] Paul Hongsuck Seo, Arsha Nagrani, Anurag Arnab, and Cordelia Schmid. End-to-end generative pretraining for multimodal video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17959-17968, 2022. 1, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 181, + 555, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 181, + 555, + 224 + ], + "spans": [ + { + "bbox": [ + 316, + 181, + 555, + 224 + ], + "type": "text", + "content": "[39] Xindi Shang, Yicong Li, Junbin Xiao, Wei Ji, and Tat-Seng Chua. Video visual relation detection via iterative inference. In Proceedings of the 29th ACM international conference on Multimedia, pages 3654-3663, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 224, + 554, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 224, + 554, + 267 + ], + "spans": [ + { + "bbox": [ + 316, + 224, + 554, + 267 + ], + "type": "text", + "content": "[40] Yaojie Shen, Xin Gu, Kai Xu, Heng Fan, Longyin Wen, and Libo Zhang. Accurate and fast compressed video captioning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15558-15567, 2023. 1, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 267, + 554, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 267, + 554, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 267, + 554, + 331 + ], + "type": "text", + "content": "[41] Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Haozhe Chi, Xun Guo, Tian Ye, Yanting Zhang, et al. Moviechat: From dense token to sparse memory for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18221-18232, 2024. 1, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 331, + 554, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 554, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 554, + 374 + ], + "type": "text", + "content": "[42] Yao Teng, Limin Wang, Zhifeng Li, and Gangshan Wu. Target adaptive context aggregation for video scene graph generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13688-13697, 2021. 1, 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 374, + 555, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 374, + 555, + 427 + ], + "spans": [ + { + "bbox": [ + 316, + 374, + 555, + 427 + ], + "type": "text", + "content": "[43] Nupur Thakur, PrasanthSai Gouripeddi, and Baoxin Li. Graph (graph): A nested graph-based framework for early accident anticipation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 7533-7541, 2024. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 427, + 555, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 427, + 555, + 482 + ], + "spans": [ + { + "bbox": [ + 316, + 427, + 555, + 482 + ], + "type": "text", + "content": "[44] Aisha Urooj, Hilde Kuehne, Bo Wu, Kim Chheu, Walid BousSELham, Chuang Gan, Niels Lobo, and Mubarak Shah. Learning situation hyper-graphs for video question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14879-14889, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 482, + 555, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 482, + 555, + 536 + ], + "spans": [ + { + "bbox": [ + 316, + 482, + 555, + 536 + ], + "type": "text", + "content": "[45] Jiahao Wang, Guo Chen, Yifei Huang, Limin Wang, and Tong Lu. Memory-and-anticipation transformer for online action understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13824–13835, 2023. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 536, + 555, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 536, + 555, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 536, + 555, + 578 + ], + "type": "text", + "content": "[46] Yanan Wang, Shuichiro Haruta, Donghuo Zeng, Julio Vizcarra, and Mori Kurokawa. Multi-object event graph representation learning for video question answering. In *Meeting on Image Recognition and Understanding*, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 578, + 555, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 578, + 555, + 622 + ], + "spans": [ + { + "bbox": [ + 316, + 578, + 555, + 622 + ], + "type": "text", + "content": "[47] Tao Wu, Runyu He, Gangshan Wu, and Limin Wang. Sportshhi: A dataset for human-human interaction detection in sports videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 622, + 555, + 675 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 622, + 555, + 675 + ], + "spans": [ + { + "bbox": [ + 316, + 622, + 555, + 675 + ], + "type": "text", + "content": "[48] Dejing Xu, Zhou Zhao, Jun Xiao, Fei Wu, Hanwang Zhang, Xiangnan He, and Yueting Zhuang. Video question answering via gradually refined attention over appearance and motion. In Proceedings of the 25th ACM international conference on Multimedia, pages 1645-1653, 2017. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 675, + 554, + 707 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 675, + 554, + 707 + ], + "spans": [ + { + "bbox": [ + 316, + 675, + 554, + 707 + ], + "type": "text", + "content": "[49] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In Proceedings of the IEEE conference on computer vision and" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "29159" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 76, + 73, + 247, + 83 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 247, + 83 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 247, + 83 + ], + "type": "text", + "content": "pattern recognition, pages 5288-5296, 2016. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 84, + 296, + 363 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 57, + 84, + 296, + 125 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 84, + 296, + 125 + ], + "spans": [ + { + "bbox": [ + 57, + 84, + 296, + 125 + ], + "type": "text", + "content": "[50] Zenan Xu, Wanjun Zhong, Qinliang Su, Zijing Ou, and Fuwei Zhang. Modeling semantic composition with syntactic hypergraph for video question answering. arXiv preprint arXiv:2205.06530, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 126, + 296, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 126, + 296, + 169 + ], + "spans": [ + { + "bbox": [ + 56, + 126, + 296, + 169 + ], + "type": "text", + "content": "[51] Jingkang Yang, Yi Zhe Ang, Zujin Guo, Kaiyang Zhou, Wayne Zhang, and Ziwei Liu. Panoptic scene graph generation. In European Conference on Computer Vision, pages 178-196. Springer, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 170, + 296, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 170, + 296, + 233 + ], + "spans": [ + { + "bbox": [ + 56, + 170, + 296, + 233 + ], + "type": "text", + "content": "[52] Jingkang Yang, Wenxuan Peng, Xiangtai Li, Zujin Guo, Liangyu Chen, Bo Li, Zheng Ma, Kaiyang Zhou, Wayne Zhang, Chen Change Loy, et al. Panoptic video scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18675-18685, 2023. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 234, + 296, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 234, + 296, + 287 + ], + "spans": [ + { + "bbox": [ + 56, + 234, + 296, + 287 + ], + "type": "text", + "content": "[53] Guangyao Zhai, Evin Pinar Örnek, Shun-Cheng Wu, Yan Di, Federico Tombari, Nassir Navab, and Benjamin Busam. Commonscenes: Generating commonsense 3d indoor scenes with scene graphs. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 288, + 295, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 288, + 295, + 319 + ], + "spans": [ + { + "bbox": [ + 57, + 288, + 295, + 319 + ], + "type": "text", + "content": "[54] Zizhao Zhang, Yifan Feng, Shihui Ying, and Yue Gao. Deep hypergraph structure learning. arXiv preprint arXiv:2208.12547, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 320, + 294, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 320, + 294, + 363 + ], + "spans": [ + { + "bbox": [ + 57, + 320, + 294, + 363 + ], + "type": "text", + "content": "[55] Xiaolin Zhu, Dongli Wang, Jianxun Li, Rui Su, Qin Wan, and Yan Zhou. Dynamical attention hypergraph convolutional network for group activity recognition. IEEE Transactions on Neural Networks and Learning Systems, 2024. 2" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29160" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/HyperGS_ Hyperspectral 3D Gaussian Splatting/8e6d9c3d-45db-41da-98c2-e1299df7de27_content_list.json b/2025/HyperGS_ Hyperspectral 3D Gaussian Splatting/8e6d9c3d-45db-41da-98c2-e1299df7de27_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9a897b0255ec8eb0259afe9e02e3b8cbea564c80 --- /dev/null +++ b/2025/HyperGS_ Hyperspectral 3D Gaussian Splatting/8e6d9c3d-45db-41da-98c2-e1299df7de27_content_list.json @@ -0,0 +1,1881 @@ +[ + { + "type": "text", + "text": "HyperGS: Hyperspectral 3D Gaussian Splitting", + "text_level": 1, + "bbox": [ + 241, + 130, + 728, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Christopher Thirgood1, Oscar Mendez1, Erin Ling1, Jon Storey2, Simon Hadfield1 \n1University of Surrey, 2I3D Robotics", + "bbox": [ + 158, + 179, + 807, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{c.thirgood, o.mendez, chao.ling, s.hadfield}@surrey.ac.uk, jstorey@i3drobotics.com", + "bbox": [ + 122, + 219, + 846, + 234 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6e48a9e7b94b09f46ccb225d769b211c2449103ace1770ab95fcd31510082b49.jpg", + "image_caption": [ + "Figure 1. This image represents a novel hyperspectral image at the 70th channel of a 141-channel image predicted from the top three models, HyperGS (Ours), 3DGS and MipNerf360 for the hyperspectral novel view synthesis task. Two random pixel reconstructions taken from the novel view with corresponding border color for its plot and pixel location in the image represented by the stars." + ], + "image_footnote": [], + "bbox": [ + 84, + 273, + 416, + 479 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c67c80e12e0dc0c72fbd1bed1aae5d5e3cf6efca0683a487d09f1d7238a6529e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 421, + 273, + 547, + 367 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/eea12ede0c05c5c776e3bb025e81dd0dbb15895ba9113cc30f08e84c248114c3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 550, + 273, + 676, + 367 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8621a675894e5482f553b4813f7075822336ac91bb6ea4e7bb0dea36b24f207a.jpg", + "image_caption": [ + "Ground Truth", + "3DGS" + ], + "image_footnote": [], + "bbox": [ + 421, + 378, + 547, + 469 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c61d492aa5a94b4f7c99c1c4e0348094c5ffa5822e1643e063272d67133e387a.jpg", + "image_caption": [ + "HyperGS (Ours)" + ], + "image_footnote": [], + "bbox": [ + 550, + 369, + 676, + 469 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e4b9aa598ff9d4db5eb2b869b2007cd4a8ad99f61e1b4353b00ee90f2cfa5393.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 681, + 273, + 885, + 373 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/91ecb10804c88bdb4a18721e416177e703623c0a974adf95025dd6a1a0b088a1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 681, + 373, + 885, + 484 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 550, + 313, + 566 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We introduce HyperGS, a novel framework for Hyperspectral Novel View Synthesis (HNVS), based on a new latent 3D Gaussian Splitting (3DGS) technique. Our approach enables simultaneous spatial and spectral renderings by encoding material properties from multi-view 3D hyperspectral datasets. HyperGS reconstructs high-fidelity views from arbitrary perspectives with improved accuracy and speed, outperforming currently existing methods. To address the challenges of high-dimensional data, we perform view synthesis in a learned latent space, incorporating a pixel-wise adaptive density function and a pruning technique for increased training stability and efficiency. Additionally, we introduce the first HNVS benchmark, implementing a number of new baselines based on recent SOTA RGB-NVS techniques, alongside the small number of prior works on HNVS. We demonstrate HyperGS's robustness through extensive evaluation of real and simulated hyperspectral scenes with a 14db accuracy improvement upon previously published models.", + "bbox": [ + 73, + 595, + 472, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 550, + 632, + 566 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advancements in Novel View Synthesis (NVS), particularly with implicit Neural Radiance Field modeling (NeRFs) [22] and explicit Gaussian models (3DGS)[13], have dramatically improved the fidelity of synthetically rendered views for conventional RGB images. However, RGB imaging is fundamentally limited, as it cannot capture the detailed material properties or non-visible scene characteristics critical for deeper scene understanding.", + "bbox": [ + 496, + 579, + 890, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hyperspectral imaging addresses these limitations by capturing a continuous spectrum of light across narrow bands for each pixel, providing valuable material properties and subtle spectral signatures. This modality is crucial for applications such as remote sensing, medical diagnostics, environmental monitoring, and robotics [27], where spectral information is paramount. The demand for Hyperspectral Novel View Synthesis (HNVS) is driven by the need for accurate, real-time spectral and spatial modeling. However, synthesizing novel views from hyperspectral data presents significant challenges due to its high dimensionality and the requirement for spectral consistency across different perspectives for each pixel.", + "bbox": [ + 496, + 703, + 892, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "5970", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Previous works, such as HS-NeRF [5], have explored HNVS by adapting NeRF architectures to hyperspectral data. However, these methods suffer from unstable training dynamics, slow rendering times, and inefficiencies in handling high-dimensional.", + "bbox": [ + 75, + 90, + 468, + 165 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3DGS has recently emerged as a promising approach for fast NVS in the RGB domain, offering efficient, high-fidelity rendering of 3D scenes through point-based representations. By projecting 3D Gaussians onto the image plane, 3DGS enables smooth, continuous representations of scene geometry while preserving fine surface details. However, despite its efficiency and rendering quality, 3DGS has not yet been successfully extended to accommodate high-dimensional data types. In our benchmark, traditional 3DGS fails to achieve consistent results against NeRF baselines for HNVS (Section 5 and Section 9 of the supplementary materials). HyperGS aims to address these challenges. In summary, the contributions of this paper are:", + "bbox": [ + 75, + 167, + 467, + 362 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We present the first method that successfully integrates view-dependent hyperspectral material information with 3DGS for high-quality HNVS.", + "2. We introduce an adaptive density control and global pruning process that leverages hyperspectral signatures for efficiency alongside a hyperspectral SFM process to stabilize the modeling.", + "3. A comprehensive benchmark of hyperspectral NeRF approaches and adaptations of classical RGB-NVS models for HNVS." + ], + "bbox": [ + 76, + 364, + 467, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Literature review", + "text_level": 1, + "bbox": [ + 76, + 530, + 246, + 544 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Hyperspectral 3D reconstruction", + "text_level": 1, + "bbox": [ + 76, + 555, + 364, + 570 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Efforts to extend 3D reconstruction to multispectral and hyperspectral domains include works from Zia et al. [31] and Liang J et al. [17], which employ point clouds and key point descriptors for structure-from-motion. These methods enhance 3D reconstruction by improving point matching but often result in sparse, noisy point clouds with inadequate occupancy information for multi-view stereo. Shadows and lighting variations degrade geometric quality. A faster approach by Zia et al. [32] instead skips rasterization by mapping hyperspectral images onto preprocessed 3D meshes.", + "bbox": [ + 75, + 579, + 467, + 729 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. NeRFs", + "text_level": 1, + "bbox": [ + 76, + 739, + 165, + 753 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "NeRFs have gained significant attention [8] since Mildenhall et al. [23] introduced the technique. NeRFs rely on classical volume rendering techniques for scene synthesis. Recent advances have sought to adapt NeRFs for spectral and hyperspectral imaging. SpectralNeRF [16] incorporates a spectrum attention mechanism (SAUNet) for realistic low-channel multi-spectral rendering under variable lighting. However, the reliance on attention mechanisms introduces computational overhead, making long-range spectra diffi", + "bbox": [ + 75, + 763, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "cult to capture. Spec-NeRF [15] takes a more efficient approach by adapting TensoRF [4] for reconstruction, with a seperate MLP estimating the camera's spectral sensitivity, focusing on computationally efficient spectral scene reconstructions. HS-NeRF [5] targets hyperspectral novel view synthesis by learning sinusoidal position encoding for each channel, allowing it to interpolate between viewpoints. This approach struggles with full-spectrum fidelity due to its lack of end-to-end training. HS-NeRF also released two multiview hyperspectral datasets, using two different hyperspectral cameras with varying signal-to-noise ratios. These, NeRF-based methods remain limited by their large parameter counts and instability when handling high-dimensional hyperspectral data alongside slower inference times.", + "bbox": [ + 496, + 90, + 890, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. 3DGS", + "text_level": 1, + "bbox": [ + 500, + 344, + 581, + 358 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Advancements in 3DGS improve point cloud efficiency for hyperspectral imaging but rely on discrete thresholds, causing uneven textures and memory issues with high-dimensional HSI data. VDGS [20] employs a hybrid NeRF neural model for color and opacity, improving spectral reconstruction for HSI. However, this model still heavily depends on the 3DGS prediction and only uses opacity estimation from the MLP. Scaffold-GS [19] introduces compressed representations to smooth surfaces but, like 3DGS, it struggles with artifacts in sparse regions. Mip-Splatting [30] addresses rendering quality by applying 3D smoothing and a 2D Mip filter, yet it remains constrained by viewfrustum-based Gaussian selection, limiting its effectiveness in dynamic scenes. Challenges in Gaussian initialization, especially due to poor SfM data, are tackled by RAINGS [12], which incorporates adaptive optimization. InstantSplat [10] improves stability by using simpler point provisioning alongside the use of dust3r [28] which offers better SFM than COLMAP but has difficulty in processing large amounts of images. Meanwhile, Bulo [3] proposes pixel-level error-based densification to ensure consistent quality, and GaussianPro [6] extends pixel-wise loss methods to better align 3D Gaussian normals during rendering, albeit at the cost of longer training times. EfficientGS[18] and LightGaussian[9] reduce model size and computational load through pruning methods but risk losing important visual details, especially when over-decimation occurs without proper ranking thresholds[18]. These methods often prioritize storage efficiency over rendering performance, overlooking challenges in real-time optimization. Although no previous works have explored the adaptation of Gaussian Splatting to Hyperspectral data, our experiments show that a naive adaptation struggles to model fine details. HyperGS solves these issues by operating in a lower-dimensional learned latent space.", + "bbox": [ + 496, + 371, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "5971", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 76, + 89, + 217, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3DGS, developed by Kerbl et al. [13], represents 3D environments using $N$ Gaussian primitives. Each Gaussian is parameterized by its position $x_{i} \\in \\mathbb{R}^{3}$ in world coordinates, a scaling vector $\\mathbf{S}_i \\in \\mathbb{R}^3$ defining the size of the Gaussian along each axis, a rotation quaternion $\\mathcal{R}_i \\in \\mathbb{R}^4$ defining its orientation, opacity $\\sigma_{i} \\in \\mathbb{R}$ , which is controlled by a sigmoid function, and appearance features $f_{i} \\in \\mathbb{R}^{k}$ to represent view-dependent RGB signals. The core of 3DGS lies in transforming these Gaussians from 3D world space into the 2D image plane for rendering. To do this, we first define the 3D covariance matrix of each Gaussian in world space as:", + "bbox": [ + 75, + 114, + 468, + 294 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Sigma_ {i} = R _ {i} S _ {i} ^ {2} R _ {i} ^ {T}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 218, + 294, + 468, + 311 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $R_{i}$ is the rotation matrix derived from the quaternion $\\mathcal{R}_i$ , and $S_{i}$ is the scaling matrix derived from the vector $\\mathbf{S}_i$ .", + "bbox": [ + 76, + 316, + 468, + 347 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To sput the Gaussian onto the 2D image plane, we transform the covariance matrix using the viewing transformation matrix $W$ and the Jacobian matrix $J$ , which handles the projection onto the camera's image plane as", + "bbox": [ + 76, + 347, + 468, + 407 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\Sigma} _ {i} = J W \\Sigma_ {i} W ^ {T} J ^ {T}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 414, + 468, + 431 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "accounting for the perspective distortion and view transformation. When rendering to a particular pixel, the opacity $\\alpha_{i}$ of Gaussian $i$ is computed as:", + "bbox": [ + 76, + 441, + 468, + 487 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {i} = 1 - \\exp \\left(- \\delta_ {i} ^ {T} \\hat {\\Sigma} _ {i} ^ {- 1} \\delta_ {i}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 494, + 468, + 512 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\delta_{i}$ is the distance between the pixel and the projected center of the Gaussian in 2D.", + "bbox": [ + 76, + 520, + 468, + 549 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The transmittance $T_{i}$ , which represents the cumulative transparency along the ray up to the $i$ -th Gaussian. It is computed as:", + "bbox": [ + 76, + 550, + 468, + 595 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nT _ {i} = \\prod_ {j = 1} ^ {i - 1} \\left(1 - \\sigma_ {i} \\alpha_ {j}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 595, + 468, + 637 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The final color $C$ of each pixel, $p$ is obtained by blending the colors of all Gaussian's projected onto that pixel:", + "bbox": [ + 76, + 641, + 468, + 671 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nC (p) = \\sum_ {i \\in N} T _ {i} \\alpha_ {i} c _ {i}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 679, + 468, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $p$ is the coordinate position in the image, $c_{i}$ is the color of the $i$ -th Gaussian, derived from its color appearance coefficients $f_{i}$ . In summary, this process transforms and projects the Gaussian primitives from 3D space into 2D image space and computes the final pixel colors through opacity and transmittance blending.", + "bbox": [ + 76, + 710, + 468, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Methodology", + "text_level": 1, + "bbox": [ + 76, + 814, + 210, + 832 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "HyperGS aims to provide a lightweight, fast rasterization solution to HNVS that is robust to different hyperspectral cameras' sensitivity functions with outstanding accuracy. Our system diagram is seen in Figure 2.", + "bbox": [ + 76, + 839, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Hyperspectral Compression", + "text_level": 1, + "bbox": [ + 500, + 90, + 751, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address the challenges of high-dimensional optimization, we adopt a novel latent space exploration method for the hyperspectral data. The latent space of a pre-trained autoencoder (AE) serves as an exploratory space for our 3DGS system. This approach reduces the computational load during 3DGS optimization by providing a compact, lower-dimensional target. Additionally, the latent space bounds the error by encapsulating the spectral sensitivity of the hyperspectral camera for each channel, enhancing the accuracy and reliability of novel view reconstruction in the spectral domain. During testing, the latent space 3DGS viewpoint estimations are decoded using the frozen AE to produce novel HSIs from which gradient flow will be calculated.", + "bbox": [ + 496, + 113, + 890, + 324 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The compression network is a fast convolutional AE. Both sides are symmetrical and is built using a series of 1D convolutional layers and Squeeze-Excitation (SE) blocks that operate across the spectral dimension of the images. The SE blocks benefit the performance by weighting features between layers for improved spectral discrimination. The model is trained on the pixel level of the scenes dataset. The encoder compresses the high-dimensional hyperspectral data into a latent representation via max-pooling layers, while the decoder reconstructs the original spectral information from this compressed form via upsampling. The architecture omits skip connections to ensure the decoder can function independently during training and testing to decode the LHSI data produced by the 3DGS system. For a detailed layout of the network architecture, please refer to Fig. 3.", + "bbox": [ + 496, + 325, + 890, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The AE minimizes the loss $L_{ae}$ , defined as:", + "bbox": [ + 517, + 566, + 807, + 582 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {a e} = L _ {H u b e r} \\left(C ^ {*} (p), D e c \\left(E n c \\left(C ^ {*} (p)\\right)\\right)\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 595, + 890, + 612 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $C^*$ is the ground truth pixel spectrum. The Huber loss provides a smooth optimization process throughout training, as it handles outliers effectively while maintaining smooth gradients. Since each hyperspectral dataset includes varying levels of signal-to-noise ratios, this approach helps improve robustness against noise-related issues.", + "bbox": [ + 496, + 625, + 890, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.2. Latent Hyperspectral 3D Gaussian Splitting", + "text_level": 1, + "bbox": [ + 498, + 724, + 877, + 742 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Each Gaussian in the splating process is assigned a latent spectral signature $\\mathbf{f}_i\\in \\mathbb{R}^m$ , where $m$ represents the latent space dimensionality. Inspired by many previous RGB techniques [14, 19, 21, 29] an MLP conditioned on view-directional hash encoding is used to predict anisotropic opacity and spectral color information for each Gaussian. This allows view-dependent spectral effects to be modelled upon specific bands. This signature is mapped directly to the pixel values using the decoder model defined in Section 4.1. For camera $d$ , the MLP $\\mathcal{F}$ predicts view-dependent", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "5972", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/51e4291d2d0e351e9e8fb321cc3a7f7e7bf2b36519dc56638cd1ea0b3f291b51.jpg", + "image_caption": [ + "Figure 2. Visual system diagram of our approach. Blue lines indicate the operational flow, while orange lines represent the gradient flow. Our novel modified latent hyperspectral adaptive density function operates within a latent space provided by a frozen autoencoder, which is also responsible for decoding the final images. Latent space novel views are generated through a combination of the latent hyperspectral Gaussian point cloud and a NeRF-style MLP. Gradients are updated based on the decoded images." + ], + "image_footnote": [], + "bbox": [ + 83, + 88, + 890, + 372 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b7612df0e08b2b813be9f51cc446f7e0fca18a6e8d6626c156db84129bf8ede0.jpg", + "image_caption": [ + "Figure 3. Our channel-wise convolutional AE model learns LHSI space representation of the scene. The decoder is only used in training and inference for the 3DGS system after the preprocessing of the dataset is finished." + ], + "image_footnote": [], + "bbox": [ + 81, + 455, + 468, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "spectral effects $\\tilde{f}_{i,d}$ and opacity $\\tilde{\\sigma}_{i,d}$ effects for each Gaussian. Specifically, the MLP takes the centre $\\mathbf{m}_i$ of each Gaussian and the view direction as inputs:", + "bbox": [ + 75, + 657, + 468, + 704 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n[ \\tilde {f} _ {i, d}, \\tilde {\\sigma} _ {i, d} ] = \\mathcal {F} _ {v} (h (\\mathbf {m} _ {i}, \\mathbf {d}); \\Theta), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 713, + 468, + 732 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $h$ is the hash encoding of the inputs while $\\tilde{\\Theta}$ denotes the MLP parameters similar to that of I-NGP [24] and Mip-Nerf360 [1].", + "bbox": [ + 75, + 743, + 468, + 789 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "These view-dependent spectral and opacity effects are multiplied with those stored in the Gaussian Cloud, leading to the updated volumetric rendering equations with transmittance being defined as:", + "bbox": [ + 75, + 789, + 470, + 849 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nT _ {i, d} = \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j} \\sigma_ {i} \\tilde {\\sigma} _ {i, d}). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 859, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The final latent signature $\\hat{C}$ of each pixel is obtained by blending the colors of all Gaussians projected onto that pixel $p$ :", + "bbox": [ + 496, + 458, + 890, + 503 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {C} (p, d) = \\sum_ {i \\in N} T _ {i, d} \\alpha_ {i} f _ {i} \\tilde {f} _ {i, d}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 503, + 890, + 535 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "It is worth emphasizing that these view-dependent spectral effects are applied within the learned latent space before decoding. This helps maintain a low computational complexity while minimizing outliers.", + "bbox": [ + 496, + 541, + 890, + 602 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A decoding operation is then performed to provide the full fidelity prediction from the system using the decoder from Section 4.1.", + "bbox": [ + 496, + 602, + 890, + 646 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nC (p, d) = \\operatorname {D e c} (\\hat {C} (p, d)). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 659, + 890, + 676 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The final rendering ensures that the spectral integrity is maintained throughout the entire process, which is critical for accurately reconstructing material properties. By leveraging the latent space, HyperGS provides a more meaningful and efficient spectral representation of the scene, than the original full hyperspectral image. The original loss proposed in 3DGS for spatial and geometric consistency in rendered images uses a weighted DSSIM and L1 loss. However, this type of loss have been shown to lead to underfitting for HSI in other fields. This is because an L1 loss can produce extreme errors for sensitive spectral bands, destabilizing training. To address this, we employ a weighted loss combining Charbonnier loss and cosine similarity to account for both spectral quality and spatial consistency.", + "bbox": [ + 496, + 688, + 892, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "5973", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c81695b20ce3f44a3a7c92a7eef62682d17dd3def000b421008ce249f9392f8b.jpg", + "image_caption": [ + "Figure 4. Visualization of our re-projection protocol for initializing the SFM point cloud. We estimate the SFM from grayscale channel slices of the hyperspectral image scene with COLMAP. Then, using the average of all views, we re-project each point into LHSI from provided by our AE, providing an optimal initialization of latent spectral color for the 3DGS system." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 467, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The cosine similarity provides a measured angular distance between vectors, making it ideal for comparing spectra. To provide an overall consistency in the spatial and spectral domain we keep the SSIM score in the training from the original 3DGS system. The per pixel HyperGS training loss $L_{d}(p)$ is:", + "bbox": [ + 75, + 316, + 468, + 409 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {d} (p) = (1 - \\lambda) (\\beta L _ {C B} (p) + L _ {C S} (p)) + \\lambda L _ {S S I M} (p), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 421, + 468, + 438 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\beta$ weighs the spectral loss balancing the cosine similarity and charbonnier loss. Thus the total training loss across all pixels and views is:", + "bbox": [ + 76, + 449, + 468, + 494 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL (C, C ^ {*}) = \\sum_ {p = 1} ^ {P} \\sum_ {d \\in V} L _ {d} (p), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 178, + 506, + 468, + 549 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $C$ is the 3DGS systems network prediction after decoding the entire latent image, $C^{*}$ is the ground truth hyperspectral image, and $V$ is the set of training view directions.", + "bbox": [ + 76, + 560, + 468, + 606 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.1 Initialisation", + "text_level": 1, + "bbox": [ + 76, + 625, + 220, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since there is no dedicated hyperspectral Structure-from-Motion (SfM) algorithm compatible with COLMAP, we first convert the hyperspectral images into grayscale images, $I_{\\mathrm{gray}}$ by selecting the spectral channel with the highest foreground intensity variance to preserve key features for the SfM process. Our process is visualized in Figure 4. Using these grayscale images and their corresponding camera projection matrices for each view $d$ , we apply COLMAP to generate a sparse 3D point cloud. Each 3D point $\\mathbf{X} = (X,Y,Z,1)^T$ in world coordinates is related to its pixel coordinate $\\mathbf{p}_d = (x_d,y_d,1)^T$ in the image plane via the camera projection $\\mathbf{K}_d\\mathbf{E}_d$ :", + "bbox": [ + 75, + 648, + 468, + 830 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {p} _ {d} = \\mathbf {K} _ {d} \\mathbf {E} _ {d} \\mathbf {X}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 218, + 842, + 468, + 858 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{K}_d$ are the camera intrinsics and $\\mathbf{E}_d$ the extrinsics. After generating 3D points and recovering camera poses,", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "we re-project the points into the LHSIs. Each 3D point $\\mathbf{X}$ is linked to pixel coordinates $\\mathbf{p}_d$ across all views. We then initialize the Gaussian cloud with one Gaussian centered on each 3D point $\\mathbf{X}$ , with spectral signature $\\mathbf{f}_i$ computed by averaging the latent hyperspectral signatures $H_{d}(y_{d},x_{d})$ across all views $d$ :", + "bbox": [ + 496, + 90, + 890, + 181 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf _ {i} = \\frac {1}{| V |} \\sum_ {v = 1} ^ {V} \\hat {C} _ {d} ^ {*} (\\mathbf {p} _ {d}), \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 619, + 191, + 890, + 232 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{C}_d^*(\\mathbf{p}_d)$ is the latent hyperspectral signature at pixel $p$ for view $d$ , and $V$ is the set of training views. This averaging ensures the spectral information is captured robustly.", + "bbox": [ + 496, + 246, + 890, + 292 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.2 Latent Hyperspectral Densification", + "text_level": 1, + "bbox": [ + 498, + 311, + 799, + 327 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As the number of color channels in HSI data increases, visual discontinuities become more prevalent. Additionally, our initial SFM reconstruction is less dense compared to standard RGB 3DGS. To address these issues, we incorporate an advanced densification method that enhances stability in both results and training. A key component of 3DGS involves determining whether a Gaussian should be split or cloned based on the gradient magnitude of the Normalized Device Coordinates (NDC) across various viewpoints. However, in sparse regions, this can cause artefacts as large Gaussians are highly visible in many viewpoints, leading to inconsistent splitting. This challenge is exacerbated in the hyperspectral domain of 3DGS, where the larger number of channels and highly variable value ranges make it harder to set effective thresholds. To mitigate this, we scale the gradient by the square of the depth relative to the scene's radius, which reduces the influence of Gaussians near the camera. More formally, we define the splitting score as:", + "bbox": [ + 496, + 335, + 890, + 609 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} \\left(g _ {i}\\right) = \\sum_ {d \\in V} \\sum_ {p = 1} ^ {P} h (d, i) \\sqrt {\\left(\\frac {\\partial L _ {d} (p)}{\\partial x _ {i}}\\right) ^ {2} + \\left(\\frac {\\partial L _ {d} (p)}{\\partial y _ {i}}\\right) ^ {2}}, \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 618, + 890, + 676 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the depth-scaling function is:", + "bbox": [ + 500, + 676, + 740, + 691 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nh (d, i) = \\left(\\frac {\\left| \\mathbf {E} _ {d} \\mathbf {X} _ {i} \\right|}{\\beta_ {\\text {f i e l d}} \\times R}\\right) ^ {2}, \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 702, + 890, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\beta_{\\mathrm{field}}$ is a tunable parameter, and $R$ is the scene's radius, based on the largest distance between any pair of cameras.", + "bbox": [ + 498, + 750, + 890, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Gaussian $g_{i}$ is split or cloned if $S(g_{i}) > \\theta_{q}$ . This approach accounts for depth-based scaling and pixel contributions, leading to better hyperspectral scene reconstruction with fewer artifacts near the camera. However, this more expressive densification can result in an excessive number of Gaussians in scenes with high-frequency details (e.g., the \"pinecone\" scene in Figure 1, to stabilize this we deploy a global pruning strategy.", + "bbox": [ + 496, + 780, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5974", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/73dbb521fad41f339471427d908def227f2f85e65b512739e441adbefdfd0972.jpg", + "image_caption": [ + "Figure 5. Visualization of our pixel-wise pruning. Gaussians with poor similarity scores below a threshold are pruned, such as the red-circled Gaussian and the black ground truth spectrum." + ], + "image_footnote": [], + "bbox": [ + 81, + 90, + 459, + 209 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Global Hyperspectral Gaussian Pruning", + "text_level": 1, + "bbox": [ + 76, + 281, + 421, + 299 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In HyperGS, we deploy pixel-wise pruning [25] to preserve the spectral integrity of latent hyperspectral images (LHSIs) while lowering the overall number of primitives to improve the scene quality, model size, and spectral gradient descent. Unlike cross-view pruning methods, which evaluate Gaussians across multiple viewpoints and leads to over-pruning [11, 18], our approach assesses each Gaussian's contribution at the pixel level. This approach retains fine spectral details specific to each pixel. For each Gaussian $g_{i}$ , pruning is based on its pixel-wise importance $\\mathcal{I}$ across all viewpoints and pixels. We compute the importance for every combination of Gaussian, pixel and viewpoint as the spectral difference between the ground truth hyperspectral value $\\hat{C}_d^* (p)$ and the decoded Gaussians LHSI, with the pruning score defined as:", + "bbox": [ + 75, + 304, + 468, + 530 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {I} [ g _ {i}, p, d ] = \\left(1 - \\left| C _ {d} ^ {*} (p) - D e c \\left(f _ {i}\\right) \\right|\\right) \\alpha_ {i} T _ {i}, \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 542, + 468, + 559 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $T_{i}$ is the accumulated transmittance of Gaussian $g_{i}$ . Including $\\alpha_{i}$ and $T_{i}$ ensures that only Gaussians contributing significantly in visibility and spectral accuracy are retained, and also ensures that the score is 0 for Gaussians that do not overlap the given pixel.", + "bbox": [ + 75, + 568, + 468, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following recent works, we do not prune Gaussians based on the average, or the total pixel-wise importance score. Instead, we retain all Gaussians within the top-K importance ranking for any pixel. More formally the filtered Gaussian cloud $\\mathcal{G}$ is", + "bbox": [ + 75, + 645, + 468, + 719 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {G} = \\left\\{g _ {i} \\mid \\exists (p, d), \\operatorname {R a n k} \\left(g _ {i}, \\mathcal {I} [:, p, d ] < \\tau_ {p}\\right) \\right\\}, \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 729, + 468, + 747 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where \"Rank\" is a function that returns the rank of a given Gaussian within the slice of the score tensor for that pixel and view.", + "bbox": [ + 75, + 756, + 468, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 815, + 209, + 832 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first evaluate HyperGS using the HS-NeRF datasets [5]. These datasets differ in signal-to-noise ratios, channel lengths, and the number of images per scene, providing a comprehensive test of model performance. The Bayspec", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "dataset [5] contains around 360 images per scene (3 scenes total), while the SOP dataset [5] has around 40 images per scene (4 scenes total). Due to the lack of continuous HSI datasets for non-object-centric scenes, we also evaluate HyperGS on a synthetic dataset curated from ScanNetv200 [7] (Section 9 in the supplementary materials, where we replace each semantic label with distinct hyperspectral signatures as seen in our supplementary materials). The baselines we have implemented for the HNVS problem include conventional 3DGS (with our reprojected SFM initialisation 4.2.1 and removed SH coefficients), traditional NeRF models adapted for hyperspectral data (NeRF [22], MipNeRF [2] MipNeRF-360 [1], Nerfacto [26] and TensorF [4] (SpecNeRF), and the only existing HNVS method HS-NeRF [5].", + "bbox": [ + 496, + 90, + 890, + 301 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "All experiments were conducted on an NVIDIA A100 80GB GPU, as MipNeRF360 requires significantly more VRAM than the other methods, which could run on a NVIDIA 3090 GPU. Using the A100 also facilitated accurate tracking of training times across all methods. We assess the performance of HyperGS after 60k training iterations, while NeRF methods are trained with the standard 1024 rays per batch and competing methods using rigorous metrics that capture both accuracy and spectral fidelity: PSNR, SSIM, Spectral Angle Mapping (SAM), RMSE. We use a $90\\%$ training and $10\\%$ test split according to the HS-NeRF dataset.", + "bbox": [ + 496, + 303, + 890, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Real HSI", + "text_level": 1, + "bbox": [ + 500, + 496, + 606, + 511 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare two HNVS datasets from HS-NeRF with different spectral channels and noise levels. The SOC710-VP (SOP) camera, covering $370 - 1100\\mathrm{nm}$ , provides high spatial $(696\\times 520)$ and spectral (128 channels) resolution but suffers from poor temporal resolution due to long exposure. The BaySpec GoldenEye camera offers comparable spatial $(640\\times 512)$ and spectral (141 channels) resolution with faster captures, introducing more noise. Tables 1 and 2 show that HyperGS consistently outperforms other methods on unseen images across all scenes.", + "bbox": [ + 496, + 521, + 890, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The BaySpec dataset particularly highlights HyperGS's strengths. Its autoencoder produces bounded errors in latent space, letting HyperGS effectively manage noise by averaging spectra and decoding within well-defined 1D feature structures. This yields more accurate reconstructions than standard end-to-end training methods. Figure 6 further shows how HyperGS better predicts object size and specular reflections compared to other baselines.", + "bbox": [ + 496, + 672, + 890, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Similar to its performance on natural NeRF tasks, MipNeRF360 performed the best among NeRF methods for HNVS. We attribute its performance to its use of positional encoding and the efficient scene representation using tensor decomposition initially employed by TensorF providing more targeted spectral predictions. Our densification and pruning techniques further refine HyperGS's performance", + "bbox": [ + 496, + 795, + 890, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "5975", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a49e2bc0f82459c8fb593d4372c05c484abc4ea2552c2606224741f5d2af7e32.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 94, + 241, + 195 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/12157b5826d1c8eb0ec2fc694f54ac3bd81a1b93eb18273eb0f53adcc4803fc4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 246, + 95, + 395, + 195 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/19027e83ac16bd4a7b32b45b63e0022ee2de1760a6056b8cc6dd2208ef61268f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 400, + 95, + 550, + 195 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/703e124c97e00ef763508d76056331528e4a47819ef5cd428d5974a0950b546e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 555, + 95, + 705, + 195 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9b6399da5fc0a596fb6a4837f8360145f3e0b549d6dfce57d907848182645a70.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 95, + 859, + 195 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c68127b80fafa5b00deea2c061f2cf18838ac7b21ae659baa02c93ed04470294.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 199, + 241, + 292 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/36bc1370b0629e090bfb637c19279218c95b40cd527bb2b3aa67eef1f89cb8b2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 246, + 199, + 397, + 292 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/be510ed92286bd483b60c30bd2e70245cfcb14ec3fc0024714efd651272a255d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 400, + 199, + 550, + 292 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e698b2475801188354984d79756ad257707a031fd0684eb5b3108f1a335f51b4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 555, + 199, + 705, + 292 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6bbe325aa10cf4fc3727a0c93cd375a338f0f93c6029a1c019cc0cfcb9cde93f.jpg", + "image_caption": [ + "Figure 6. Visualisation of the top 4 methods for frame 51 of 359 for the Caladium plant scene from the Bayspec dataset. The top row shows the 70th channel of the 141 channel image predicted, the bottom row provides a raw pixel-wise error heatmap of the scene." + ], + "image_footnote": [], + "bbox": [ + 709, + 195, + 892, + 297 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e89f175cbb169ca19add91b717112882ab03a5214b2715d3b7395437e1a546b5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 333, + 241, + 428 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9ca7e8fddbc7a4a652cf945b8e75f8ea8de544b6213f1100ce8ed48be0425cd9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 246, + 333, + 395, + 428 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/617c68515f5d61d85207e307c57c4afa8234d969986d4736bbdb1d7c15a8fe4d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 400, + 333, + 550, + 428 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4d604f66b83a9f7a46ebc4672cbbd0708aef82715a3d4327141b4479d113952a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 555, + 333, + 705, + 428 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e5ba4e3e1358514855a7043540f6e9f9667e9fa74b67b9bdf72f8f4545002a42.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 333, + 890, + 428 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/18cbc74fb3341a1027b07275d13246f0822c0b07f4a4298f11b2e2fc99f22708.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 436, + 241, + 525 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/24e1860ad484263dfb17faf68fb1d49ba5eaf9fa637c061aebfcc815fbb68722.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 246, + 438, + 395, + 525 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ac66b354eaf0902cd3201aeeb964e6fb6618432a63d6bad548238d3d82290924.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 400, + 438, + 550, + 525 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/41a7de3963f588c8bbd049a5aa51ca7814198601dbfdd0e38349fac7248ad620.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 555, + 438, + 705, + 525 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8eca2ed85e5e17b5fbdc92be03dbd0524171e46e212fe829131fbc464e671517.jpg", + "image_caption": [ + "Figure 7. Visualisation of the top 4 methods for frame 31 of 49 for the Tools plant scene from the SOP dataset. The top row shows the 70th channel of the 128 channel image predicted, the bottom row provides a raw pixel-wise error heatmap of the scene." + ], + "image_footnote": [], + "bbox": [ + 709, + 430, + 892, + 532 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "by effectively filtering poor spectral representations. Meanwhile, the SOP dataset favors both 3DGS approaches. Due to the high number of frames present in the BaySpec camera scenes, the NeRF approaches adapt well to this abundance of data and can provide far better and more stable representations of the scene given the volume of viewpoints. This volume makes predicting the noisier spectra that the BaySpec camera provides easier than 3DGS. In contrast, although the SOP datasets may have smoother spectra, the number of training viewpoints is far smaller. This leads to NeRF approaches evidently struggling with scene scale and spectral reconstruction, as they struggle to understand the scene scale and perform spectral reconstruction due to the lack of data. In contrast, 3DGS approaches excel due to smoother data interpolation and simpler optimization. This is outlined by Figure 7 highlighting the intense error heatmaps from NeRF approaches compared to 3DGS and HyperGS. HyperGS, however, outperforms all methods across all scenes and camera datasets, suggesting that it provides a far more robust and accurate performance than both 3DGS and NeRF methods for HNVS. For visualiza", + "bbox": [ + 75, + 569, + 468, + 885 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/905de4e1555dd323830ac55bbebbcbf4853ed4d1155adb7cde5e46864b72e71e.jpg", + "table_caption": [ + "BaySpec Datasets" + ], + "table_footnote": [], + "table_body": "
MethodAverage Results
PSNR ↑SSIM↑SAM↓RMSE↓
NeRF23.350.60610.04400.0687
MipNeRF22.750.59470.04350.0776
TensoRF24.660.64820.05010.0587
Nerfacto19.120.58660.05510.1174
MipNerf36026.530.74420.02800.0476
HS-NeRF19.820.67140.05340.1071
3DGS22.910.63210.13350.0810
HyperGS27.110.78040.02540.0440
", + "bbox": [ + 506, + 579, + 883, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Quantitative results using the Bayspec datasets against separate hyperspectral methods and baseline NeRF and 3DGS (best bold, second best italic).", + "bbox": [ + 500, + 762, + 890, + 803 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tions of random pixels and spectral images, please refer to Figures 8 and then 9 which showcases HyperGS's superior pixel reconstruction capabilities compared to other benchmarked methods.", + "bbox": [ + 498, + 839, + 890, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "5976", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/93fdca6666af58207f82e65fa5c76f3b5f49e0f34e1fe6d2adbe247383352798.jpg", + "image_caption": [ + "Figure 8. Three random pixel reconstructions taken from test frame 151 of the bayspec Anacampseros scene." + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 348, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ffa9223e76a854222c5389fc41bd6fc4c74c538f20d3e44e74ac659be534e68f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 351, + 90, + 616, + 226 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/51c6f99a93dfbfaedcd6dbdf9eed0ab238e53592c618c701eb7453682797c033.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 619, + 90, + 885, + 220 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9387353f18a257f80946db0266b31f023ac592de2a94d9d6a5cb69c411f1fa75.jpg", + "image_caption": [ + "Figure 9. Three random pixel reconstructions from test frame 31 of the SOP origami scene." + ], + "image_footnote": [], + "bbox": [ + 81, + 244, + 348, + 375 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b8777cd475ed739d5d64250c1d8e8cc0bd03be14bff316740a3f618e9e85924c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 351, + 244, + 616, + 380 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7e3388e5f38d886dd81476a083d3e2dcb95fabf0553c9480a79f4a7dd784e557.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 619, + 244, + 885, + 373 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a4d39e33275aec94209d146ccb910c1ea71e346c8435a089742199786a9bef64.jpg", + "table_caption": [ + "Surface Optics Datasets" + ], + "table_footnote": [ + "Table 2. Quantitative results using the SOP datasets against separate hyperspectral methods and baseline NeRF and 3DGS." + ], + "table_body": "
MethodAverage Results
PSNR ↑SSIM↑SAM↓RMSE↓
NeRF10.890.59050.06250.3479
MipNeRF12.530.54810.05680.3198
TensorF13.000.56960.05950.2780
Nerfacto16.370.69860.03520.1601
MipNeRF36012.280.68240.13690.2658
HS-NeRF14.440.61650.20370.1953
3DGS28.580.96270.03010.0478
HyperGS30.510.97560.004150.0354
", + "bbox": [ + 81, + 422, + 460, + 588 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 633, + 230, + 650 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To evaluate the effectiveness and impact of various components in our HyperGS model, we conducted a series of ablation experiments. These experiments help to understand the contribution of individual components and choices in our model's design. As shown in Table 3, each new feature added to the baseline 3DGS model improves HNVS performance. The joint latent autoencoder-3DGS architecture boosts spectral reconstruction the most while the proposed densification provides greater overall details captured. With the latent space learning and positional encoding, the performance of HyperGS gets a significant boost in performance too. Please refer to our supplementary materials for additional ablations, including analyses of HyperGS performance, pruning strategy scoring functions, autoencoder performance with varying data and latent sizes, the global pruning strategy, and system speed bottlenecks.", + "bbox": [ + 75, + 659, + 472, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7ef34aa10c4f4ce02e255f423a779ee57d3f998a46b16d7a0d9e9574fb12e50a.jpg", + "table_caption": [], + "table_footnote": [ + "Table 3. Adding each new feature to 3DGS significantly improves the model's average performance for the HNVS task in the bayspec dataset while creating a more stable number of primitives (N.Prim) in the point cloud." + ], + "table_body": "
Ablation StepAverage Results for the Bayspec dataset
PSNR ↑SSIM↑SAM↓RMSE↓N.Prim ↓
Base. 3DGS22.910.63200.13350.0810440k
+ Spec. SFM23.050.63310.13100.0799421k
+ Latent space AE24.870.71010.05480.0602500k
+ Densification25.250.73560.03650.05481.3M
+ Pruning25.170.71990.03740.0555412k
+ View dependent MLP27.050.77920.02530.0443309k
+ Custom L.Function27.110.78010.02540.0440310k
", + "bbox": [ + 504, + 410, + 888, + 522 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusions", + "text_level": 1, + "bbox": [ + 500, + 595, + 627, + 611 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduced HyperGS, the first effective approach for hyperspectral novel view synthesis (HNVS). By encoding latent spectral data within Gaussian primitives and performing Gaussian splatting in a learned latent space, HyperGS achieves high-quality, material-aware rendering. Our adaptive density control and pruning techniques efficiently handle latent hyperspectral signatures in 3D Gaussian point clouds, ensuring stable training and superior accuracy compared to benchmark methods. For future work, we aim to extend HyperGS with more comprehensive latent space decoding that corrects and refines both spectral and spatial information, providing richer contextual detail.", + "bbox": [ + 496, + 619, + 893, + 800 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 818, + 687, + 835 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was partially supported by Industrial 3D Robotics (I3D), IS-Instruments and partially funded by the EPSRC under grant agreement EP/S035761/1.", + "bbox": [ + 496, + 844, + 893, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "5977", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] J.T. Barron, B. Mildenhall, D. Verbin, P.P. Srinivasan, and P. Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 4, 6", + "[2] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields, 2021. 6", + "[3] S. R. Bulò, L. Porzi, and P. Kontschieder. Revising densification in gaussian splatting. ArXiv, 2024. 2", + "[4] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision (ECCV), 2022. 2, 6", + "[5] G. Chen, S. K. Narayanan, T. G. Ottou, B. Missaoui, H. Muriki, C. Pradalier, and Y. Chen. Hyperspectral neural radiance fields. ArXiv, 2024. 2, 6", + "[6] Kai Cheng, Xiaoxiao Long, Kaizhi Yang, Yao Yao, Wei Yin, Yuexin Ma, Wenping Wang, and Xuejin Chen. Gaussianpro: 3d gaussian splatting with progressive propagation. arXiv preprint arXiv:2402.14650, 2024. 2", + "[7] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proc. Computer Vision and Pattern Recognition (CVPR), IEEE, 2017. 6, 1", + "[8] F. Dellaert. NeRF explosion 2020. https://dellaert.github.io/NeRF/, 2020. 2", + "[9] Z. Fan, K. Wang, K. Wen, Z. Zhu, D. Xu, and Z. Wang. Lightgaussian: Unbounded 3d gaussian compression with 15x reduction and $200+$ fps. arXiv preprint arXiv:2311.17245, 2023. 2", + "[10] Z. Fan, W. Cong, K. Wen, K. Wang, J. Zhang, X. Ding, D. Xu, B. Ivanovic, M. Pavone, G. Pavlakos, Z. Wang, and Y. Wang. Instantsplat: Sparse-view sfm-free gaussian splatting in seconds. ArXiv, 2024. 2", + "[11] S. Girish, K. Gupta, and A. Shrivastava. Eagles: Efficient accelerated 3d gaussians with lightweight encodings. arXiv preprint arXiv:2312.04564, 2023. 6", + "[12] Jaewoo Jung, Jisang Han, Honggyu An, Jiwon Kang, Seonghoon Park, and Seungryong Kim. Relaxing accurate initialization constraint for 3d gaussian splatting. arXiv preprint arXiv:2403.09413, 2024. 2", + "[13] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (SIGGRAPH Conference Proceedings), 42(4), 2023. 1, 3", + "[14] Byeonghyeon Lee, Howoong Lee, Xiangyu Sun, Usman Ali, and Eunbyung Park. Deblurring 3d gaussian splatting, 2024. 3", + "[15] J. Li, Y. Li, C. Sun, C. Wang, and J. Xiang. Spec-nerf: Multispectral neural radiance fields. ArXiv, 2023. 2", + "[16] R. Li, J. Liu, G. Liu, S. Zhang, B. Zeng, and S. Liu. Spectralnerf: Physically based spectral rendering with neural radiation field. ArXiv, 2023. 2" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Jie Liang, Ali Zia, Jun Zhou, and Xavier Sirault. 3d plant modelling via hyperspectral imaging. In 2013 IEEE International Conference on Computer Vision Workshops, pages 172-177, 2013. 2", + "[18] W. Liu, T. Guan, B. Zhu, L. Ju, Z. Song, D. Li, Y. Wang, and W. Yang. Efficientges: Streamlining gaussian splatting for large-scale high-resolution scene representation. ArXiv, 2024. 2, 6", + "[19] T. Lu, M. Yu, L. Xu, Y. Xiangli, L. Wang, D. Lin, and B. Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. ArXiv, 2023. 2, 3", + "[20] D. Malarz, W. Smolak, J. Tabor, S. Tadeja, and P. Spurek. Gaussian splatting with nef-based color and opacity. ArXiv, 2023. 2", + "[21] D. Malarz, W. Smolak, J. Tabor, S. Tadeja, and P. Spurek. Gaussian splatting with nef-based color and opacity. ArXiv, 2023. 3", + "[22] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 6", + "[23] B. Mildenhall, P. Hedman, R. Martin-Brualla, P.P. Srinivasan, and J.T. Barron. Nerf in the dark: High dynamic range view synthesis from noisy raw images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[24] T. Müller, A. Evans, C. Schied, and A. Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (TOG), 41(4):102:1-102:15, 2022. 4", + "[25] S. Niedermayr, J. Stumpfegger, and R. Westermann. Compressed 3d gaussian splatting for accelerated novel view synthesis. arXiv preprint arXiv:2401.02436, 2023. 6", + "[26] M. Tancik, E. Weber, E. Ng, R. Li, B. Yi, J. Kerr, T. Wang, A. Kristoffersen, J. Austin, K. Salahi, A. Ahuja, D. McAllister, and A. Kanazawa. Nerfstudio: A modular framework for neural radiance field development. arXiv preprint arXiv:2302.04264, 2023. 6", + "[27] Christopher Thirgood, Oscar Mendez, Erin Chao Ling, Jon Storey, and Simon Hadfield. Raspectloc: Raman spectroscopy-dependent robot localisation. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5296-5303, 2023. 1", + "[28] S. Wang, V. Leroy, Y. Cabon, B. Chidlovskii, and J. Revaud. DUSt3R: Geometric 3D Vision Made Easy. arXiv preprint arXiv:2312.14132, 2023. 2", + "[29] Z. Yang, X. Gao, Y. Sun, Y. Huang, X. Lyu, W. Zhou, S. Jiao, X. Qi, and X. Jin. Spec-Gaussian: Anisotropic View-Dependent Appearance for 3D Gaussian Splatting. ArXiv, 2024. 3", + "[30] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19447-19456, 2024. 2", + "[31] A. Zia, J. Liang, J. Zhou, and Y. Gao. 3d reconstruction from hyperspectral images. In 2015 IEEE Winter Conference on" + ], + "bbox": [ + 501, + 92, + 890, + 902 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "5978", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Applications of Computer Vision (WACV), pages 318-325, 2015. 2", + "[32] Ali Zia, Jie Liang, Jun Zhou, and Yongsheng Gao. 3d reconstruction from hyperspectral images. In 2015 IEEE Winter Conference on Applications of Computer Vision, pages 318-325, 2015. 2" + ], + "bbox": [ + 78, + 90, + 468, + 174 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "5979", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/HyperGS_ Hyperspectral 3D Gaussian Splatting/8e6d9c3d-45db-41da-98c2-e1299df7de27_model.json b/2025/HyperGS_ Hyperspectral 3D Gaussian Splatting/8e6d9c3d-45db-41da-98c2-e1299df7de27_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f85d833379ffbf9dae1fe775bd99c08d5f210760 --- /dev/null +++ b/2025/HyperGS_ Hyperspectral 3D Gaussian Splatting/8e6d9c3d-45db-41da-98c2-e1299df7de27_model.json @@ -0,0 +1,2277 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.242, + 0.131, + 0.73, + 0.154 + ], + "angle": 0, + "content": "HyperGS: Hyperspectral 3D Gaussian Splitting" + }, + { + "type": "text", + "bbox": [ + 0.16, + 0.18, + 0.808, + 0.217 + ], + "angle": 0, + "content": "Christopher Thirgood1, Oscar Mendez1, Erin Ling1, Jon Storey2, Simon Hadfield1 \n1University of Surrey, 2I3D Robotics" + }, + { + "type": "text", + "bbox": [ + 0.123, + 0.22, + 0.848, + 0.235 + ], + "angle": 0, + "content": "{c.thirgood, o.mendez, chao.ling, s.hadfield}@surrey.ac.uk, jstorey@i3drobotics.com" + }, + { + "type": "image", + "bbox": [ + 0.085, + 0.274, + 0.418, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.422, + 0.275, + 0.549, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.552, + 0.275, + 0.677, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.447, + 0.37, + 0.526, + 0.38 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.423, + 0.38, + 0.549, + 0.47 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.472, + 0.471, + 0.504, + 0.481 + ], + "angle": 0, + "content": "3DGS" + }, + { + "type": "image", + "bbox": [ + 0.552, + 0.37, + 0.678, + 0.47 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.573, + 0.471, + 0.658, + 0.482 + ], + "angle": 0, + "content": "HyperGS (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.682, + 0.274, + 0.887, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.682, + 0.374, + 0.887, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.498, + 0.895, + 0.541 + ], + "angle": 0, + "content": "Figure 1. This image represents a novel hyperspectral image at the 70th channel of a 141-channel image predicted from the top three models, HyperGS (Ours), 3DGS and MipNerf360 for the hyperspectral novel view synthesis task. Two random pixel reconstructions taken from the novel view with corresponding border color for its plot and pixel location in the image represented by the stars." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.551, + 0.314, + 0.567 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.596, + 0.473, + 0.884 + ], + "angle": 0, + "content": "We introduce HyperGS, a novel framework for Hyperspectral Novel View Synthesis (HNVS), based on a new latent 3D Gaussian Splitting (3DGS) technique. Our approach enables simultaneous spatial and spectral renderings by encoding material properties from multi-view 3D hyperspectral datasets. HyperGS reconstructs high-fidelity views from arbitrary perspectives with improved accuracy and speed, outperforming currently existing methods. To address the challenges of high-dimensional data, we perform view synthesis in a learned latent space, incorporating a pixel-wise adaptive density function and a pruning technique for increased training stability and efficiency. Additionally, we introduce the first HNVS benchmark, implementing a number of new baselines based on recent SOTA RGB-NVS techniques, alongside the small number of prior works on HNVS. We demonstrate HyperGS's robustness through extensive evaluation of real and simulated hyperspectral scenes with a 14db accuracy improvement upon previously published models." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.551, + 0.633, + 0.567 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.58, + 0.892, + 0.701 + ], + "angle": 0, + "content": "Recent advancements in Novel View Synthesis (NVS), particularly with implicit Neural Radiance Field modeling (NeRFs) [22] and explicit Gaussian models (3DGS)[13], have dramatically improved the fidelity of synthetically rendered views for conventional RGB images. However, RGB imaging is fundamentally limited, as it cannot capture the detailed material properties or non-visible scene characteristics critical for deeper scene understanding." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.893, + 0.903 + ], + "angle": 0, + "content": "Hyperspectral imaging addresses these limitations by capturing a continuous spectrum of light across narrow bands for each pixel, providing valuable material properties and subtle spectral signatures. This modality is crucial for applications such as remote sensing, medical diagnostics, environmental monitoring, and robotics [27], where spectral information is paramount. The demand for Hyperspectral Novel View Synthesis (HNVS) is driven by the need for accurate, real-time spectral and spatial modeling. However, synthesizing novel views from hyperspectral data presents significant challenges due to its high dimensionality and the requirement for spectral consistency across different perspectives for each pixel." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5970" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.166 + ], + "angle": 0, + "content": "Previous works, such as HS-NeRF [5], have explored HNVS by adapting NeRF architectures to hyperspectral data. However, these methods suffer from unstable training dynamics, slow rendering times, and inefficiencies in handling high-dimensional." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.468, + 0.363 + ], + "angle": 0, + "content": "3DGS has recently emerged as a promising approach for fast NVS in the RGB domain, offering efficient, high-fidelity rendering of 3D scenes through point-based representations. By projecting 3D Gaussians onto the image plane, 3DGS enables smooth, continuous representations of scene geometry while preserving fine surface details. However, despite its efficiency and rendering quality, 3DGS has not yet been successfully extended to accommodate high-dimensional data types. In our benchmark, traditional 3DGS fails to achieve consistent results against NeRF baselines for HNVS (Section 5 and Section 9 of the supplementary materials). HyperGS aims to address these challenges. In summary, the contributions of this paper are:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.365, + 0.468, + 0.409 + ], + "angle": 0, + "content": "1. We present the first method that successfully integrates view-dependent hyperspectral material information with 3DGS for high-quality HNVS." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.41, + 0.468, + 0.469 + ], + "angle": 0, + "content": "2. We introduce an adaptive density control and global pruning process that leverages hyperspectral signatures for efficiency alongside a hyperspectral SFM process to stabilize the modeling." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.47, + 0.468, + 0.514 + ], + "angle": 0, + "content": "3. A comprehensive benchmark of hyperspectral NeRF approaches and adaptations of classical RGB-NVS models for HNVS." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.365, + 0.468, + 0.514 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.531, + 0.248, + 0.545 + ], + "angle": 0, + "content": "2. Literature review" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.556, + 0.365, + 0.571 + ], + "angle": 0, + "content": "2.1. Hyperspectral 3D reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.58, + 0.468, + 0.731 + ], + "angle": 0, + "content": "Efforts to extend 3D reconstruction to multispectral and hyperspectral domains include works from Zia et al. [31] and Liang J et al. [17], which employ point clouds and key point descriptors for structure-from-motion. These methods enhance 3D reconstruction by improving point matching but often result in sparse, noisy point clouds with inadequate occupancy information for multi-view stereo. Shadows and lighting variations degrade geometric quality. A faster approach by Zia et al. [32] instead skips rasterization by mapping hyperspectral images onto preprocessed 3D meshes." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.741, + 0.166, + 0.755 + ], + "angle": 0, + "content": "2.2. NeRFs" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.468, + 0.901 + ], + "angle": 0, + "content": "NeRFs have gained significant attention [8] since Mildenhall et al. [23] introduced the technique. NeRFs rely on classical volume rendering techniques for scene synthesis. Recent advances have sought to adapt NeRFs for spectral and hyperspectral imaging. SpectralNeRF [16] incorporates a spectrum attention mechanism (SAUNet) for realistic low-channel multi-spectral rendering under variable lighting. However, the reliance on attention mechanisms introduces computational overhead, making long-range spectra diffi" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.303 + ], + "angle": 0, + "content": "cult to capture. Spec-NeRF [15] takes a more efficient approach by adapting TensoRF [4] for reconstruction, with a seperate MLP estimating the camera's spectral sensitivity, focusing on computationally efficient spectral scene reconstructions. HS-NeRF [5] targets hyperspectral novel view synthesis by learning sinusoidal position encoding for each channel, allowing it to interpolate between viewpoints. This approach struggles with full-spectrum fidelity due to its lack of end-to-end training. HS-NeRF also released two multiview hyperspectral datasets, using two different hyperspectral cameras with varying signal-to-noise ratios. These, NeRF-based methods remain limited by their large parameter counts and instability when handling high-dimensional hyperspectral data alongside slower inference times." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.345, + 0.583, + 0.359 + ], + "angle": 0, + "content": "2.3. 3DGS" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.372, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Advancements in 3DGS improve point cloud efficiency for hyperspectral imaging but rely on discrete thresholds, causing uneven textures and memory issues with high-dimensional HSI data. VDGS [20] employs a hybrid NeRF neural model for color and opacity, improving spectral reconstruction for HSI. However, this model still heavily depends on the 3DGS prediction and only uses opacity estimation from the MLP. Scaffold-GS [19] introduces compressed representations to smooth surfaces but, like 3DGS, it struggles with artifacts in sparse regions. Mip-Splatting [30] addresses rendering quality by applying 3D smoothing and a 2D Mip filter, yet it remains constrained by viewfrustum-based Gaussian selection, limiting its effectiveness in dynamic scenes. Challenges in Gaussian initialization, especially due to poor SfM data, are tackled by RAINGS [12], which incorporates adaptive optimization. InstantSplat [10] improves stability by using simpler point provisioning alongside the use of dust3r [28] which offers better SFM than COLMAP but has difficulty in processing large amounts of images. Meanwhile, Bulo [3] proposes pixel-level error-based densification to ensure consistent quality, and GaussianPro [6] extends pixel-wise loss methods to better align 3D Gaussian normals during rendering, albeit at the cost of longer training times. EfficientGS[18] and LightGaussian[9] reduce model size and computational load through pruning methods but risk losing important visual details, especially when over-decimation occurs without proper ranking thresholds[18]. These methods often prioritize storage efficiency over rendering performance, overlooking challenges in real-time optimization. Although no previous works have explored the adaptation of Gaussian Splatting to Hyperspectral data, our experiments show that a naive adaptation struggles to model fine details. HyperGS solves these issues by operating in a lower-dimensional learned latent space." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "5971" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.09, + 0.218, + 0.106 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.115, + 0.47, + 0.295 + ], + "angle": 0, + "content": "3DGS, developed by Kerbl et al. [13], represents 3D environments using \\(N\\) Gaussian primitives. Each Gaussian is parameterized by its position \\(x_{i} \\in \\mathbb{R}^{3}\\) in world coordinates, a scaling vector \\(\\mathbf{S}_i \\in \\mathbb{R}^3\\) defining the size of the Gaussian along each axis, a rotation quaternion \\(\\mathcal{R}_i \\in \\mathbb{R}^4\\) defining its orientation, opacity \\(\\sigma_{i} \\in \\mathbb{R}\\), which is controlled by a sigmoid function, and appearance features \\(f_{i} \\in \\mathbb{R}^{k}\\) to represent view-dependent RGB signals. The core of 3DGS lies in transforming these Gaussians from 3D world space into the 2D image plane for rendering. To do this, we first define the 3D covariance matrix of each Gaussian in world space as:" + }, + { + "type": "equation", + "bbox": [ + 0.22, + 0.295, + 0.469, + 0.313 + ], + "angle": 0, + "content": "\\[\n\\Sigma_ {i} = R _ {i} S _ {i} ^ {2} R _ {i} ^ {T}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.317, + 0.469, + 0.348 + ], + "angle": 0, + "content": "where \\(R_{i}\\) is the rotation matrix derived from the quaternion \\(\\mathcal{R}_i\\), and \\(S_{i}\\) is the scaling matrix derived from the vector \\(\\mathbf{S}_i\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.348, + 0.469, + 0.408 + ], + "angle": 0, + "content": "To sput the Gaussian onto the 2D image plane, we transform the covariance matrix using the viewing transformation matrix \\( W \\) and the Jacobian matrix \\( J \\), which handles the projection onto the camera's image plane as" + }, + { + "type": "equation", + "bbox": [ + 0.201, + 0.415, + 0.469, + 0.433 + ], + "angle": 0, + "content": "\\[\n\\hat {\\Sigma} _ {i} = J W \\Sigma_ {i} W ^ {T} J ^ {T}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.442, + 0.469, + 0.488 + ], + "angle": 0, + "content": "accounting for the perspective distortion and view transformation. When rendering to a particular pixel, the opacity \\(\\alpha_{i}\\) of Gaussian \\(i\\) is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.495, + 0.469, + 0.513 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {i} = 1 - \\exp \\left(- \\delta_ {i} ^ {T} \\hat {\\Sigma} _ {i} ^ {- 1} \\delta_ {i}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.521, + 0.469, + 0.55 + ], + "angle": 0, + "content": "where \\(\\delta_{i}\\) is the distance between the pixel and the projected center of the Gaussian in 2D." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.551, + 0.469, + 0.596 + ], + "angle": 0, + "content": "The transmittance \\( T_{i} \\), which represents the cumulative transparency along the ray up to the \\( i \\)-th Gaussian. It is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.202, + 0.596, + 0.469, + 0.638 + ], + "angle": 0, + "content": "\\[\nT _ {i} = \\prod_ {j = 1} ^ {i - 1} \\left(1 - \\sigma_ {i} \\alpha_ {j}\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.642, + 0.469, + 0.672 + ], + "angle": 0, + "content": "The final color \\(C\\) of each pixel, \\(p\\) is obtained by blending the colors of all Gaussian's projected onto that pixel:" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.68, + 0.469, + 0.711 + ], + "angle": 0, + "content": "\\[\nC (p) = \\sum_ {i \\in N} T _ {i} \\alpha_ {i} c _ {i}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.712, + 0.469, + 0.803 + ], + "angle": 0, + "content": "where \\( p \\) is the coordinate position in the image, \\( c_{i} \\) is the color of the \\( i \\)-th Gaussian, derived from its color appearance coefficients \\( f_{i} \\). In summary, this process transforms and projects the Gaussian primitives from 3D space into 2D image space and computes the final pixel colors through opacity and transmittance blending." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.815, + 0.212, + 0.833 + ], + "angle": 0, + "content": "4. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.469, + 0.902 + ], + "angle": 0, + "content": "HyperGS aims to provide a lightweight, fast rasterization solution to HNVS that is robust to different hyperspectral cameras' sensitivity functions with outstanding accuracy. Our system diagram is seen in Figure 2." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.753, + 0.108 + ], + "angle": 0, + "content": "4.1. Hyperspectral Compression" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.114, + 0.892, + 0.325 + ], + "angle": 0, + "content": "To address the challenges of high-dimensional optimization, we adopt a novel latent space exploration method for the hyperspectral data. The latent space of a pre-trained autoencoder (AE) serves as an exploratory space for our 3DGS system. This approach reduces the computational load during 3DGS optimization by providing a compact, lower-dimensional target. Additionally, the latent space bounds the error by encapsulating the spectral sensitivity of the hyperspectral camera for each channel, enhancing the accuracy and reliability of novel view reconstruction in the spectral domain. During testing, the latent space 3DGS viewpoint estimations are decoded using the frozen AE to produce novel HSIs from which gradient flow will be calculated." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.326, + 0.892, + 0.568 + ], + "angle": 0, + "content": "The compression network is a fast convolutional AE. Both sides are symmetrical and is built using a series of 1D convolutional layers and Squeeze-Excitation (SE) blocks that operate across the spectral dimension of the images. The SE blocks benefit the performance by weighting features between layers for improved spectral discrimination. The model is trained on the pixel level of the scenes dataset. The encoder compresses the high-dimensional hyperspectral data into a latent representation via max-pooling layers, while the decoder reconstructs the original spectral information from this compressed form via upsampling. The architecture omits skip connections to ensure the decoder can function independently during training and testing to decode the LHSI data produced by the 3DGS system. For a detailed layout of the network architecture, please refer to Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.568, + 0.808, + 0.583 + ], + "angle": 0, + "content": "The AE minimizes the loss \\( L_{ae} \\), defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.596, + 0.891, + 0.613 + ], + "angle": 0, + "content": "\\[\nL _ {a e} = L _ {H u b e r} \\left(C ^ {*} (p), D e c \\left(E n c \\left(C ^ {*} (p)\\right)\\right)\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.625, + 0.892, + 0.716 + ], + "angle": 0, + "content": "where \\( C^* \\) is the ground truth pixel spectrum. The Huber loss provides a smooth optimization process throughout training, as it handles outliers effectively while maintaining smooth gradients. Since each hyperspectral dataset includes varying levels of signal-to-noise ratios, this approach helps improve robustness against noise-related issues." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.726, + 0.878, + 0.743 + ], + "angle": 0, + "content": "4.2. Latent Hyperspectral 3D Gaussian Splitting" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Each Gaussian in the splating process is assigned a latent spectral signature \\(\\mathbf{f}_i\\in \\mathbb{R}^m\\) , where \\(m\\) represents the latent space dimensionality. Inspired by many previous RGB techniques [14, 19, 21, 29] an MLP conditioned on view-directional hash encoding is used to predict anisotropic opacity and spectral color information for each Gaussian. This allows view-dependent spectral effects to be modelled upon specific bands. This signature is mapped directly to the pixel values using the decoder model defined in Section 4.1. For camera \\(d\\) , the MLP \\(\\mathcal{F}\\) predicts view-dependent" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5972" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.089, + 0.891, + 0.373 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.384, + 0.893, + 0.442 + ], + "angle": 0, + "content": "Figure 2. Visual system diagram of our approach. Blue lines indicate the operational flow, while orange lines represent the gradient flow. Our novel modified latent hyperspectral adaptive density function operates within a latent space provided by a frozen autoencoder, which is also responsible for decoding the final images. Latent space novel views are generated through a combination of the latent hyperspectral Gaussian point cloud and a NeRF-style MLP. Gradients are updated based on the decoded images." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.457, + 0.47, + 0.576 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.583, + 0.471, + 0.64 + ], + "angle": 0, + "content": "Figure 3. Our channel-wise convolutional AE model learns LHSI space representation of the scene. The decoder is only used in training and inference for the 3DGS system after the preprocessing of the dataset is finished." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.658, + 0.47, + 0.705 + ], + "angle": 0, + "content": "spectral effects \\(\\tilde{f}_{i,d}\\) and opacity \\(\\tilde{\\sigma}_{i,d}\\) effects for each Gaussian. Specifically, the MLP takes the centre \\(\\mathbf{m}_i\\) of each Gaussian and the view direction as inputs:" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.714, + 0.47, + 0.733 + ], + "angle": 0, + "content": "\\[\n[ \\tilde {f} _ {i, d}, \\tilde {\\sigma} _ {i, d} ] = \\mathcal {F} _ {v} (h (\\mathbf {m} _ {i}, \\mathbf {d}); \\Theta), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.744, + 0.47, + 0.79 + ], + "angle": 0, + "content": "where \\( h \\) is the hash encoding of the inputs while \\( \\tilde{\\Theta} \\) denotes the MLP parameters similar to that of I-NGP [24] and Mip-Nerf360 [1]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.79, + 0.471, + 0.851 + ], + "angle": 0, + "content": "These view-dependent spectral and opacity effects are multiplied with those stored in the Gaussian Cloud, leading to the updated volumetric rendering equations with transmittance being defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.183, + 0.861, + 0.47, + 0.904 + ], + "angle": 0, + "content": "\\[\nT _ {i, d} = \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j} \\sigma_ {i} \\tilde {\\sigma} _ {i, d}). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.459, + 0.892, + 0.504 + ], + "angle": 0, + "content": "The final latent signature \\(\\hat{C}\\) of each pixel is obtained by blending the colors of all Gaussians projected onto that pixel \\(p\\):" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.504, + 0.892, + 0.536 + ], + "angle": 0, + "content": "\\[\n\\hat {C} (p, d) = \\sum_ {i \\in N} T _ {i, d} \\alpha_ {i} f _ {i} \\tilde {f} _ {i, d}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.542, + 0.892, + 0.603 + ], + "angle": 0, + "content": "It is worth emphasizing that these view-dependent spectral effects are applied within the learned latent space before decoding. This helps maintain a low computational complexity while minimizing outliers." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.603, + 0.892, + 0.647 + ], + "angle": 0, + "content": "A decoding operation is then performed to provide the full fidelity prediction from the system using the decoder from Section 4.1." + }, + { + "type": "equation", + "bbox": [ + 0.609, + 0.66, + 0.892, + 0.678 + ], + "angle": 0, + "content": "\\[\nC (p, d) = \\operatorname {D e c} (\\hat {C} (p, d)). \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.893, + 0.902 + ], + "angle": 0, + "content": "The final rendering ensures that the spectral integrity is maintained throughout the entire process, which is critical for accurately reconstructing material properties. By leveraging the latent space, HyperGS provides a more meaningful and efficient spectral representation of the scene, than the original full hyperspectral image. The original loss proposed in 3DGS for spatial and geometric consistency in rendered images uses a weighted DSSIM and L1 loss. However, this type of loss have been shown to lead to underfitting for HSI in other fields. This is because an L1 loss can produce extreme errors for sensitive spectral bands, destabilizing training. To address this, we employ a weighted loss combining Charbonnier loss and cosine similarity to account for both spectral quality and spatial consistency." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5973" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.468, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.223, + 0.471, + 0.307 + ], + "angle": 0, + "content": "Figure 4. Visualization of our re-projection protocol for initializing the SFM point cloud. We estimate the SFM from grayscale channel slices of the hyperspectral image scene with COLMAP. Then, using the average of all views, we re-project each point into LHSI from provided by our AE, providing an optimal initialization of latent spectral color for the 3DGS system." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.318, + 0.47, + 0.41 + ], + "angle": 0, + "content": "The cosine similarity provides a measured angular distance between vectors, making it ideal for comparing spectra. To provide an overall consistency in the spatial and spectral domain we keep the SSIM score in the training from the original 3DGS system. The per pixel HyperGS training loss \\( L_{d}(p) \\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.422, + 0.469, + 0.439 + ], + "angle": 0, + "content": "\\[\nL _ {d} (p) = (1 - \\lambda) (\\beta L _ {C B} (p) + L _ {C S} (p)) + \\lambda L _ {S S I M} (p), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.45, + 0.47, + 0.495 + ], + "angle": 0, + "content": "where \\(\\beta\\) weighs the spectral loss balancing the cosine similarity and charbonnier loss. Thus the total training loss across all pixels and views is:" + }, + { + "type": "equation", + "bbox": [ + 0.179, + 0.507, + 0.469, + 0.55 + ], + "angle": 0, + "content": "\\[\nL (C, C ^ {*}) = \\sum_ {p = 1} ^ {P} \\sum_ {d \\in V} L _ {d} (p), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.561, + 0.469, + 0.607 + ], + "angle": 0, + "content": "where \\(C\\) is the 3DGS systems network prediction after decoding the entire latent image, \\(C^{*}\\) is the ground truth hyperspectral image, and \\(V\\) is the set of training view directions." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.625, + 0.221, + 0.639 + ], + "angle": 0, + "content": "4.2.1 Initialisation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.649, + 0.47, + 0.831 + ], + "angle": 0, + "content": "Since there is no dedicated hyperspectral Structure-from-Motion (SfM) algorithm compatible with COLMAP, we first convert the hyperspectral images into grayscale images, \\(I_{\\mathrm{gray}}\\) by selecting the spectral channel with the highest foreground intensity variance to preserve key features for the SfM process. Our process is visualized in Figure 4. Using these grayscale images and their corresponding camera projection matrices for each view \\(d\\), we apply COLMAP to generate a sparse 3D point cloud. Each 3D point \\(\\mathbf{X} = (X,Y,Z,1)^T\\) in world coordinates is related to its pixel coordinate \\(\\mathbf{p}_d = (x_d,y_d,1)^T\\) in the image plane via the camera projection \\(\\mathbf{K}_d\\mathbf{E}_d\\):" + }, + { + "type": "equation", + "bbox": [ + 0.22, + 0.843, + 0.469, + 0.859 + ], + "angle": 0, + "content": "\\[\n\\mathbf {p} _ {d} = \\mathbf {K} _ {d} \\mathbf {E} _ {d} \\mathbf {X}, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathbf{K}_d\\) are the camera intrinsics and \\(\\mathbf{E}_d\\) the extrinsics. After generating 3D points and recovering camera poses," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.182 + ], + "angle": 0, + "content": "we re-project the points into the LHSIs. Each 3D point \\(\\mathbf{X}\\) is linked to pixel coordinates \\(\\mathbf{p}_d\\) across all views. We then initialize the Gaussian cloud with one Gaussian centered on each 3D point \\(\\mathbf{X}\\), with spectral signature \\(\\mathbf{f}_i\\) computed by averaging the latent hyperspectral signatures \\(H_{d}(y_{d},x_{d})\\) across all views \\(d\\):" + }, + { + "type": "equation", + "bbox": [ + 0.62, + 0.192, + 0.892, + 0.233 + ], + "angle": 0, + "content": "\\[\nf _ {i} = \\frac {1}{| V |} \\sum_ {v = 1} ^ {V} \\hat {C} _ {d} ^ {*} (\\mathbf {p} _ {d}), \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.247, + 0.892, + 0.293 + ], + "angle": 0, + "content": "where \\(\\hat{C}_d^*(\\mathbf{p}_d)\\) is the latent hyperspectral signature at pixel \\(p\\) for view \\(d\\), and \\(V\\) is the set of training views. This averaging ensures the spectral information is captured robustly." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.313, + 0.8, + 0.328 + ], + "angle": 0, + "content": "4.2.2 Latent Hyperspectral Densification" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.337, + 0.892, + 0.61 + ], + "angle": 0, + "content": "As the number of color channels in HSI data increases, visual discontinuities become more prevalent. Additionally, our initial SFM reconstruction is less dense compared to standard RGB 3DGS. To address these issues, we incorporate an advanced densification method that enhances stability in both results and training. A key component of 3DGS involves determining whether a Gaussian should be split or cloned based on the gradient magnitude of the Normalized Device Coordinates (NDC) across various viewpoints. However, in sparse regions, this can cause artefacts as large Gaussians are highly visible in many viewpoints, leading to inconsistent splitting. This challenge is exacerbated in the hyperspectral domain of 3DGS, where the larger number of channels and highly variable value ranges make it harder to set effective thresholds. To mitigate this, we scale the gradient by the square of the depth relative to the scene's radius, which reduces the influence of Gaussians near the camera. More formally, we define the splitting score as:" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.619, + 0.891, + 0.677 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} \\left(g _ {i}\\right) = \\sum_ {d \\in V} \\sum_ {p = 1} ^ {P} h (d, i) \\sqrt {\\left(\\frac {\\partial L _ {d} (p)}{\\partial x _ {i}}\\right) ^ {2} + \\left(\\frac {\\partial L _ {d} (p)}{\\partial y _ {i}}\\right) ^ {2}}, \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.677, + 0.741, + 0.692 + ], + "angle": 0, + "content": "where the depth-scaling function is:" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.703, + 0.892, + 0.74 + ], + "angle": 0, + "content": "\\[\nh (d, i) = \\left(\\frac {\\left| \\mathbf {E} _ {d} \\mathbf {X} _ {i} \\right|}{\\beta_ {\\text {f i e l d}} \\times R}\\right) ^ {2}, \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.75, + 0.891, + 0.78 + ], + "angle": 0, + "content": "\\(\\beta_{\\mathrm{field}}\\) is a tunable parameter, and \\(R\\) is the scene's radius, based on the largest distance between any pair of cameras." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.892, + 0.902 + ], + "angle": 0, + "content": "The Gaussian \\( g_{i} \\) is split or cloned if \\( S(g_{i}) > \\theta_{q} \\). This approach accounts for depth-based scaling and pixel contributions, leading to better hyperspectral scene reconstruction with fewer artifacts near the camera. However, this more expressive densification can result in an excessive number of Gaussians in scenes with high-frequency details (e.g., the \"pinecone\" scene in Figure 1, to stabilize this we deploy a global pruning strategy." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5974" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.092, + 0.46, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.223, + 0.47, + 0.266 + ], + "angle": 0, + "content": "Figure 5. Visualization of our pixel-wise pruning. Gaussians with poor similarity scores below a threshold are pruned, such as the red-circled Gaussian and the black ground truth spectrum." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.282, + 0.423, + 0.3 + ], + "angle": 0, + "content": "4.3. Global Hyperspectral Gaussian Pruning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.305, + 0.47, + 0.531 + ], + "angle": 0, + "content": "In HyperGS, we deploy pixel-wise pruning [25] to preserve the spectral integrity of latent hyperspectral images (LHSIs) while lowering the overall number of primitives to improve the scene quality, model size, and spectral gradient descent. Unlike cross-view pruning methods, which evaluate Gaussians across multiple viewpoints and leads to over-pruning [11, 18], our approach assesses each Gaussian's contribution at the pixel level. This approach retains fine spectral details specific to each pixel. For each Gaussian \\( g_{i} \\), pruning is based on its pixel-wise importance \\( \\mathcal{I} \\) across all viewpoints and pixels. We compute the importance for every combination of Gaussian, pixel and viewpoint as the spectral difference between the ground truth hyperspectral value \\( \\hat{C}_d^* (p) \\) and the decoded Gaussians LHSI, with the pruning score defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.111, + 0.543, + 0.469, + 0.56 + ], + "angle": 0, + "content": "\\[\n\\mathcal {I} [ g _ {i}, p, d ] = \\left(1 - \\left| C _ {d} ^ {*} (p) - D e c \\left(f _ {i}\\right) \\right|\\right) \\alpha_ {i} T _ {i}, \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.469, + 0.645 + ], + "angle": 0, + "content": "where \\( T_{i} \\) is the accumulated transmittance of Gaussian \\( g_{i} \\). Including \\( \\alpha_{i} \\) and \\( T_{i} \\) ensures that only Gaussians contributing significantly in visibility and spectral accuracy are retained, and also ensures that the score is 0 for Gaussians that do not overlap the given pixel." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.646, + 0.469, + 0.72 + ], + "angle": 0, + "content": "Following recent works, we do not prune Gaussians based on the average, or the total pixel-wise importance score. Instead, we retain all Gaussians within the top-K importance ranking for any pixel. More formally the filtered Gaussian cloud \\(\\mathcal{G}\\) is" + }, + { + "type": "equation", + "bbox": [ + 0.115, + 0.731, + 0.469, + 0.748 + ], + "angle": 0, + "content": "\\[\n\\mathcal {G} = \\left\\{g _ {i} \\mid \\exists (p, d), \\operatorname {R a n k} \\left(g _ {i}, \\mathcal {I} [:, p, d ] < \\tau_ {p}\\right) \\right\\}, \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.757, + 0.469, + 0.802 + ], + "angle": 0, + "content": "where \"Rank\" is a function that returns the rank of a given Gaussian within the slice of the score tensor for that pixel and view." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.816, + 0.21, + 0.833 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.469, + 0.902 + ], + "angle": 0, + "content": "We first evaluate HyperGS using the HS-NeRF datasets [5]. These datasets differ in signal-to-noise ratios, channel lengths, and the number of images per scene, providing a comprehensive test of model performance. The Bayspec" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.303 + ], + "angle": 0, + "content": "dataset [5] contains around 360 images per scene (3 scenes total), while the SOP dataset [5] has around 40 images per scene (4 scenes total). Due to the lack of continuous HSI datasets for non-object-centric scenes, we also evaluate HyperGS on a synthetic dataset curated from ScanNetv200 [7] (Section 9 in the supplementary materials, where we replace each semantic label with distinct hyperspectral signatures as seen in our supplementary materials). The baselines we have implemented for the HNVS problem include conventional 3DGS (with our reprojected SFM initialisation 4.2.1 and removed SH coefficients), traditional NeRF models adapted for hyperspectral data (NeRF [22], MipNeRF [2] MipNeRF-360 [1], Nerfacto [26] and TensorF [4] (SpecNeRF), and the only existing HNVS method HS-NeRF [5]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.304, + 0.892, + 0.485 + ], + "angle": 0, + "content": "All experiments were conducted on an NVIDIA A100 80GB GPU, as MipNeRF360 requires significantly more VRAM than the other methods, which could run on a NVIDIA 3090 GPU. Using the A100 also facilitated accurate tracking of training times across all methods. We assess the performance of HyperGS after 60k training iterations, while NeRF methods are trained with the standard 1024 rays per batch and competing methods using rigorous metrics that capture both accuracy and spectral fidelity: PSNR, SSIM, Spectral Angle Mapping (SAM), RMSE. We use a \\(90\\%\\) training and \\(10\\%\\) test split according to the HS-NeRF dataset." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.497, + 0.607, + 0.512 + ], + "angle": 0, + "content": "5.1. Real HSI" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.522, + 0.892, + 0.673 + ], + "angle": 0, + "content": "We compare two HNVS datasets from HS-NeRF with different spectral channels and noise levels. The SOC710-VP (SOP) camera, covering \\(370 - 1100\\mathrm{nm}\\), provides high spatial \\((696\\times 520)\\) and spectral (128 channels) resolution but suffers from poor temporal resolution due to long exposure. The BaySpec GoldenEye camera offers comparable spatial \\((640\\times 512)\\) and spectral (141 channels) resolution with faster captures, introducing more noise. Tables 1 and 2 show that HyperGS consistently outperforms other methods on unseen images across all scenes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.673, + 0.892, + 0.794 + ], + "angle": 0, + "content": "The BaySpec dataset particularly highlights HyperGS's strengths. Its autoencoder produces bounded errors in latent space, letting HyperGS effectively manage noise by averaging spectra and decoding within well-defined 1D feature structures. This yields more accurate reconstructions than standard end-to-end training methods. Figure 6 further shows how HyperGS better predicts object size and specular reflections compared to other baselines." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Similar to its performance on natural NeRF tasks, MipNeRF360 performed the best among NeRF methods for HNVS. We attribute its performance to its use of positional encoding and the efficient scene representation using tensor decomposition initially employed by TensorF providing more targeted spectral predictions. Our densification and pruning techniques further refine HyperGS's performance" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5975" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.095, + 0.242, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.247, + 0.096, + 0.397, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.401, + 0.096, + 0.551, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.556, + 0.096, + 0.706, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.096, + 0.861, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.2, + 0.242, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.247, + 0.2, + 0.398, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.401, + 0.2, + 0.551, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.556, + 0.2, + 0.706, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.196, + 0.893, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.298, + 0.892, + 0.325 + ], + "angle": 0, + "content": "Figure 6. Visualisation of the top 4 methods for frame 51 of 359 for the Caladium plant scene from the Bayspec dataset. The top row shows the 70th channel of the 141 channel image predicted, the bottom row provides a raw pixel-wise error heatmap of the scene." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.334, + 0.242, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.247, + 0.334, + 0.397, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.401, + 0.334, + 0.551, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.556, + 0.334, + 0.706, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.334, + 0.892, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.438, + 0.242, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.247, + 0.439, + 0.397, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.401, + 0.439, + 0.551, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.556, + 0.439, + 0.706, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.431, + 0.893, + 0.533 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.534, + 0.892, + 0.561 + ], + "angle": 0, + "content": "Figure 7. Visualisation of the top 4 methods for frame 31 of 49 for the Tools plant scene from the SOP dataset. The top row shows the 70th channel of the 128 channel image predicted, the bottom row provides a raw pixel-wise error heatmap of the scene." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.57, + 0.47, + 0.886 + ], + "angle": 0, + "content": "by effectively filtering poor spectral representations. Meanwhile, the SOP dataset favors both 3DGS approaches. Due to the high number of frames present in the BaySpec camera scenes, the NeRF approaches adapt well to this abundance of data and can provide far better and more stable representations of the scene given the volume of viewpoints. This volume makes predicting the noisier spectra that the BaySpec camera provides easier than 3DGS. In contrast, although the SOP datasets may have smoother spectra, the number of training viewpoints is far smaller. This leads to NeRF approaches evidently struggling with scene scale and spectral reconstruction, as they struggle to understand the scene scale and perform spectral reconstruction due to the lack of data. In contrast, 3DGS approaches excel due to smoother data interpolation and simpler optimization. This is outlined by Figure 7 highlighting the intense error heatmaps from NeRF approaches compared to 3DGS and HyperGS. HyperGS, however, outperforms all methods across all scenes and camera datasets, suggesting that it provides a far more robust and accurate performance than both 3DGS and NeRF methods for HNVS. For visualiza" + }, + { + "type": "table_caption", + "bbox": [ + 0.637, + 0.567, + 0.757, + 0.58 + ], + "angle": 0, + "content": "BaySpec Datasets" + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.58, + 0.884, + 0.751 + ], + "angle": 0, + "content": "
MethodAverage Results
PSNR ↑SSIM↑SAM↓RMSE↓
NeRF23.350.60610.04400.0687
MipNeRF22.750.59470.04350.0776
TensoRF24.660.64820.05010.0587
Nerfacto19.120.58660.05510.1174
MipNerf36026.530.74420.02800.0476
HS-NeRF19.820.67140.05340.1071
3DGS22.910.63210.13350.0810
HyperGS27.110.78040.02540.0440
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.763, + 0.892, + 0.804 + ], + "angle": 0, + "content": "Table 1. Quantitative results using the Bayspec datasets against separate hyperspectral methods and baseline NeRF and 3DGS (best bold, second best italic)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.892, + 0.9 + ], + "angle": 0, + "content": "tions of random pixels and spectral images, please refer to Figures 8 and then 9 which showcases HyperGS's superior pixel reconstruction capabilities compared to other benchmarked methods." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5976" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.349, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.091, + 0.617, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.62, + 0.091, + 0.887, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.162, + 0.23, + 0.806, + 0.243 + ], + "angle": 0, + "content": "Figure 8. Three random pixel reconstructions taken from test frame 151 of the bayspec Anacampseros scene." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.245, + 0.349, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.245, + 0.617, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.62, + 0.245, + 0.887, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.213, + 0.384, + 0.755, + 0.398 + ], + "angle": 0, + "content": "Figure 9. Three random pixel reconstructions from test frame 31 of the SOP origami scene." + }, + { + "type": "table_caption", + "bbox": [ + 0.194, + 0.41, + 0.353, + 0.424 + ], + "angle": 0, + "content": "Surface Optics Datasets" + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.424, + 0.462, + 0.589 + ], + "angle": 0, + "content": "
MethodAverage Results
PSNR ↑SSIM↑SAM↓RMSE↓
NeRF10.890.59050.06250.3479
MipNeRF12.530.54810.05680.3198
TensorF13.000.56960.05950.2780
Nerfacto16.370.69860.03520.1601
MipNeRF36012.280.68240.13690.2658
HS-NeRF14.440.61650.20370.1953
3DGS28.580.96270.03010.0478
HyperGS30.510.97560.004150.0354
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.597, + 0.47, + 0.626 + ], + "angle": 0, + "content": "Table 2. Quantitative results using the SOP datasets against separate hyperspectral methods and baseline NeRF and 3DGS." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.635, + 0.231, + 0.651 + ], + "angle": 0, + "content": "5.2. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.473, + 0.903 + ], + "angle": 0, + "content": "To evaluate the effectiveness and impact of various components in our HyperGS model, we conducted a series of ablation experiments. These experiments help to understand the contribution of individual components and choices in our model's design. As shown in Table 3, each new feature added to the baseline 3DGS model improves HNVS performance. The joint latent autoencoder-3DGS architecture boosts spectral reconstruction the most while the proposed densification provides greater overall details captured. With the latent space learning and positional encoding, the performance of HyperGS gets a significant boost in performance too. Please refer to our supplementary materials for additional ablations, including analyses of HyperGS performance, pruning strategy scoring functions, autoencoder performance with varying data and latent sizes, the global pruning strategy, and system speed bottlenecks." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.411, + 0.89, + 0.523 + ], + "angle": 0, + "content": "
Ablation StepAverage Results for the Bayspec dataset
PSNR ↑SSIM↑SAM↓RMSE↓N.Prim ↓
Base. 3DGS22.910.63200.13350.0810440k
+ Spec. SFM23.050.63310.13100.0799421k
+ Latent space AE24.870.71010.05480.0602500k
+ Densification25.250.73560.03650.05481.3M
+ Pruning25.170.71990.03740.0555412k
+ View dependent MLP27.050.77920.02530.0443309k
+ Custom L.Function27.110.78010.02540.0440310k
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.498, + 0.53, + 0.894, + 0.587 + ], + "angle": 0, + "content": "Table 3. Adding each new feature to 3DGS significantly improves the model's average performance for the HNVS task in the bayspec dataset while creating a more stable number of primitives (N.Prim) in the point cloud." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.597, + 0.628, + 0.612 + ], + "angle": 0, + "content": "6. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.62, + 0.895, + 0.801 + ], + "angle": 0, + "content": "In this paper, we introduced HyperGS, the first effective approach for hyperspectral novel view synthesis (HNVS). By encoding latent spectral data within Gaussian primitives and performing Gaussian splatting in a learned latent space, HyperGS achieves high-quality, material-aware rendering. Our adaptive density control and pruning techniques efficiently handle latent hyperspectral signatures in 3D Gaussian point clouds, ensuring stable training and superior accuracy compared to benchmark methods. For future work, we aim to extend HyperGS with more comprehensive latent space decoding that corrects and refines both spectral and spatial information, providing richer contextual detail." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.819, + 0.688, + 0.836 + ], + "angle": 0, + "content": "7. Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.845, + 0.894, + 0.9 + ], + "angle": 0, + "content": "This work was partially supported by Industrial 3D Robotics (I3D), IS-Instruments and partially funded by the EPSRC under grant agreement EP/S035761/1." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5977" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] J.T. Barron, B. Mildenhall, D. Verbin, P.P. Srinivasan, and P. Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.186, + 0.472, + 0.242 + ], + "angle": 0, + "content": "[2] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.244, + 0.471, + 0.271 + ], + "angle": 0, + "content": "[3] S. R. Bulò, L. Porzi, and P. Kontschieder. Revising densification in gaussian splatting. ArXiv, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.272, + 0.471, + 0.314 + ], + "angle": 0, + "content": "[4] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision (ECCV), 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.315, + 0.471, + 0.357 + ], + "angle": 0, + "content": "[5] G. Chen, S. K. Narayanan, T. G. Ottou, B. Missaoui, H. Muriki, C. Pradalier, and Y. Chen. Hyperspectral neural radiance fields. ArXiv, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.359, + 0.471, + 0.414 + ], + "angle": 0, + "content": "[6] Kai Cheng, Xiaoxiao Long, Kaizhi Yang, Yao Yao, Wei Yin, Yuexin Ma, Wenping Wang, and Xuejin Chen. Gaussianpro: 3d gaussian splatting with progressive propagation. arXiv preprint arXiv:2402.14650, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.416, + 0.471, + 0.485 + ], + "angle": 0, + "content": "[7] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proc. Computer Vision and Pattern Recognition (CVPR), IEEE, 2017. 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.487, + 0.468, + 0.514 + ], + "angle": 0, + "content": "[8] F. Dellaert. NeRF explosion 2020. https://dellaert.github.io/NeRF/, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.515, + 0.471, + 0.571 + ], + "angle": 0, + "content": "[9] Z. Fan, K. Wang, K. Wen, Z. Zhu, D. Xu, and Z. Wang. Lightgaussian: Unbounded 3d gaussian compression with 15x reduction and \\(200+\\) fps. arXiv preprint arXiv:2311.17245, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.573, + 0.47, + 0.627 + ], + "angle": 0, + "content": "[10] Z. Fan, W. Cong, K. Wen, K. Wang, J. Zhang, X. Ding, D. Xu, B. Ivanovic, M. Pavone, G. Pavlakos, Z. Wang, and Y. Wang. Instantsplat: Sparse-view sfm-free gaussian splatting in seconds. ArXiv, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.63, + 0.47, + 0.671 + ], + "angle": 0, + "content": "[11] S. Girish, K. Gupta, and A. Shrivastava. Eagles: Efficient accelerated 3d gaussians with lightweight encodings. arXiv preprint arXiv:2312.04564, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.673, + 0.47, + 0.728 + ], + "angle": 0, + "content": "[12] Jaewoo Jung, Jisang Han, Honggyu An, Jiwon Kang, Seonghoon Park, and Seungryong Kim. Relaxing accurate initialization constraint for 3d gaussian splatting. arXiv preprint arXiv:2403.09413, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.73, + 0.47, + 0.785 + ], + "angle": 0, + "content": "[13] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (SIGGRAPH Conference Proceedings), 42(4), 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.787, + 0.47, + 0.828 + ], + "angle": 0, + "content": "[14] Byeonghyeon Lee, Howoong Lee, Xiangyu Sun, Usman Ali, and Eunbyung Park. Deblurring 3d gaussian splatting, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.83, + 0.47, + 0.857 + ], + "angle": 0, + "content": "[15] J. Li, Y. Li, C. Sun, C. Wang, and J. Xiang. Spec-nerf: Multispectral neural radiance fields. ArXiv, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[16] R. Li, J. Liu, G. Liu, S. Zhang, B. Zeng, and S. Liu. Spectralnerf: Physically based spectral rendering with neural radiation field. ArXiv, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[17] Jie Liang, Ali Zia, Jun Zhou, and Xavier Sirault. 3d plant modelling via hyperspectral imaging. In 2013 IEEE International Conference on Computer Vision Workshops, pages 172-177, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[18] W. Liu, T. Guan, B. Zhu, L. Ju, Z. Song, D. Li, Y. Wang, and W. Yang. Efficientges: Streamlining gaussian splatting for large-scale high-resolution scene representation. ArXiv, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.206, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[19] T. Lu, M. Yu, L. Xu, Y. Xiangli, L. Wang, D. Lin, and B. Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. ArXiv, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[20] D. Malarz, W. Smolak, J. Tabor, S. Tadeja, and P. Spurek. Gaussian splatting with nef-based color and opacity. ArXiv, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.291, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[21] D. Malarz, W. Smolak, J. Tabor, S. Tadeja, and P. Spurek. Gaussian splatting with nef-based color and opacity. ArXiv, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[22] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.391, + 0.892, + 0.461 + ], + "angle": 0, + "content": "[23] B. Mildenhall, P. Hedman, R. Martin-Brualla, P.P. Srinivasan, and J.T. Barron. Nerf in the dark: High dynamic range view synthesis from noisy raw images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.462, + 0.892, + 0.516 + ], + "angle": 0, + "content": "[24] T. Müller, A. Evans, C. Schied, and A. Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (TOG), 41(4):102:1-102:15, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.518, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[25] S. Niedermayr, J. Stumpfegger, and R. Westermann. Compressed 3d gaussian splatting for accelerated novel view synthesis. arXiv preprint arXiv:2401.02436, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.562, + 0.892, + 0.63 + ], + "angle": 0, + "content": "[26] M. Tancik, E. Weber, E. Ng, R. Li, B. Yi, J. Kerr, T. Wang, A. Kristoffersen, J. Austin, K. Salahi, A. Ahuja, D. McAllister, and A. Kanazawa. Nerfstudio: A modular framework for neural radiance field development. arXiv preprint arXiv:2302.04264, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.632, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[27] Christopher Thirgood, Oscar Mendez, Erin Chao Ling, Jon Storey, and Simon Hadfield. Raspectloc: Raman spectroscopy-dependent robot localisation. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5296-5303, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.703, + 0.892, + 0.744 + ], + "angle": 0, + "content": "[28] S. Wang, V. Leroy, Y. Cabon, B. Chidlovskii, and J. Revaud. DUSt3R: Geometric 3D Vision Made Easy. arXiv preprint arXiv:2312.14132, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.746, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[29] Z. Yang, X. Gao, Y. Sun, Y. Huang, X. Lyu, W. Zhou, S. Jiao, X. Qi, and X. Jin. Spec-Gaussian: Anisotropic View-Dependent Appearance for 3D Gaussian Splatting. ArXiv, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[30] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19447-19456, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.903 + ], + "angle": 0, + "content": "[31] A. Zia, J. Liang, J. Zhou, and Y. Gao. 3d reconstruction from hyperspectral images. In 2015 IEEE Winter Conference on" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "5978" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.468, + 0.119 + ], + "angle": 0, + "content": "Applications of Computer Vision (WACV), pages 318-325, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.175 + ], + "angle": 0, + "content": "[32] Ali Zia, Jie Liang, Jun Zhou, and Yongsheng Gao. 3d reconstruction from hyperspectral images. In 2015 IEEE Winter Conference on Applications of Computer Vision, pages 318-325, 2015. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.175 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "5979" + } + ] +] \ No newline at end of file diff --git a/2025/HyperGS_ Hyperspectral 3D Gaussian Splatting/8e6d9c3d-45db-41da-98c2-e1299df7de27_origin.pdf b/2025/HyperGS_ Hyperspectral 3D Gaussian Splatting/8e6d9c3d-45db-41da-98c2-e1299df7de27_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..edd841dff4618009787525d0e1e5f4464bc3314c --- /dev/null +++ b/2025/HyperGS_ Hyperspectral 3D Gaussian Splatting/8e6d9c3d-45db-41da-98c2-e1299df7de27_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b46d33752bb3903a7dd38dbd834bb71a3317926598e7c57f4d1aa5a304b27cc5 +size 1548248 diff --git a/2025/HyperGS_ Hyperspectral 3D Gaussian Splatting/full.md b/2025/HyperGS_ Hyperspectral 3D Gaussian Splatting/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7825a3305a3c887c5f2331a1e63bbe67025518f5 --- /dev/null +++ b/2025/HyperGS_ Hyperspectral 3D Gaussian Splatting/full.md @@ -0,0 +1,370 @@ +# HyperGS: Hyperspectral 3D Gaussian Splitting + +Christopher Thirgood1, Oscar Mendez1, Erin Ling1, Jon Storey2, Simon Hadfield1 +1University of Surrey, 2I3D Robotics + +{c.thirgood, o.mendez, chao.ling, s.hadfield}@surrey.ac.uk, jstorey@i3drobotics.com + +![](images/6e48a9e7b94b09f46ccb225d769b211c2449103ace1770ab95fcd31510082b49.jpg) +Figure 1. This image represents a novel hyperspectral image at the 70th channel of a 141-channel image predicted from the top three models, HyperGS (Ours), 3DGS and MipNerf360 for the hyperspectral novel view synthesis task. Two random pixel reconstructions taken from the novel view with corresponding border color for its plot and pixel location in the image represented by the stars. + +![](images/c67c80e12e0dc0c72fbd1bed1aae5d5e3cf6efca0683a487d09f1d7238a6529e.jpg) + +![](images/eea12ede0c05c5c776e3bb025e81dd0dbb15895ba9113cc30f08e84c248114c3.jpg) + +![](images/8621a675894e5482f553b4813f7075822336ac91bb6ea4e7bb0dea36b24f207a.jpg) +Ground Truth +3DGS + +![](images/c61d492aa5a94b4f7c99c1c4e0348094c5ffa5822e1643e063272d67133e387a.jpg) +HyperGS (Ours) + +![](images/e4b9aa598ff9d4db5eb2b869b2007cd4a8ad99f61e1b4353b00ee90f2cfa5393.jpg) + +![](images/91ecb10804c88bdb4a18721e416177e703623c0a974adf95025dd6a1a0b088a1.jpg) + +# Abstract + +We introduce HyperGS, a novel framework for Hyperspectral Novel View Synthesis (HNVS), based on a new latent 3D Gaussian Splitting (3DGS) technique. Our approach enables simultaneous spatial and spectral renderings by encoding material properties from multi-view 3D hyperspectral datasets. HyperGS reconstructs high-fidelity views from arbitrary perspectives with improved accuracy and speed, outperforming currently existing methods. To address the challenges of high-dimensional data, we perform view synthesis in a learned latent space, incorporating a pixel-wise adaptive density function and a pruning technique for increased training stability and efficiency. Additionally, we introduce the first HNVS benchmark, implementing a number of new baselines based on recent SOTA RGB-NVS techniques, alongside the small number of prior works on HNVS. We demonstrate HyperGS's robustness through extensive evaluation of real and simulated hyperspectral scenes with a 14db accuracy improvement upon previously published models. + +# 1. Introduction + +Recent advancements in Novel View Synthesis (NVS), particularly with implicit Neural Radiance Field modeling (NeRFs) [22] and explicit Gaussian models (3DGS)[13], have dramatically improved the fidelity of synthetically rendered views for conventional RGB images. However, RGB imaging is fundamentally limited, as it cannot capture the detailed material properties or non-visible scene characteristics critical for deeper scene understanding. + +Hyperspectral imaging addresses these limitations by capturing a continuous spectrum of light across narrow bands for each pixel, providing valuable material properties and subtle spectral signatures. This modality is crucial for applications such as remote sensing, medical diagnostics, environmental monitoring, and robotics [27], where spectral information is paramount. The demand for Hyperspectral Novel View Synthesis (HNVS) is driven by the need for accurate, real-time spectral and spatial modeling. However, synthesizing novel views from hyperspectral data presents significant challenges due to its high dimensionality and the requirement for spectral consistency across different perspectives for each pixel. + +Previous works, such as HS-NeRF [5], have explored HNVS by adapting NeRF architectures to hyperspectral data. However, these methods suffer from unstable training dynamics, slow rendering times, and inefficiencies in handling high-dimensional. + +3DGS has recently emerged as a promising approach for fast NVS in the RGB domain, offering efficient, high-fidelity rendering of 3D scenes through point-based representations. By projecting 3D Gaussians onto the image plane, 3DGS enables smooth, continuous representations of scene geometry while preserving fine surface details. However, despite its efficiency and rendering quality, 3DGS has not yet been successfully extended to accommodate high-dimensional data types. In our benchmark, traditional 3DGS fails to achieve consistent results against NeRF baselines for HNVS (Section 5 and Section 9 of the supplementary materials). HyperGS aims to address these challenges. In summary, the contributions of this paper are: + +1. We present the first method that successfully integrates view-dependent hyperspectral material information with 3DGS for high-quality HNVS. +2. We introduce an adaptive density control and global pruning process that leverages hyperspectral signatures for efficiency alongside a hyperspectral SFM process to stabilize the modeling. +3. A comprehensive benchmark of hyperspectral NeRF approaches and adaptations of classical RGB-NVS models for HNVS. + +# 2. Literature review + +# 2.1. Hyperspectral 3D reconstruction + +Efforts to extend 3D reconstruction to multispectral and hyperspectral domains include works from Zia et al. [31] and Liang J et al. [17], which employ point clouds and key point descriptors for structure-from-motion. These methods enhance 3D reconstruction by improving point matching but often result in sparse, noisy point clouds with inadequate occupancy information for multi-view stereo. Shadows and lighting variations degrade geometric quality. A faster approach by Zia et al. [32] instead skips rasterization by mapping hyperspectral images onto preprocessed 3D meshes. + +# 2.2. NeRFs + +NeRFs have gained significant attention [8] since Mildenhall et al. [23] introduced the technique. NeRFs rely on classical volume rendering techniques for scene synthesis. Recent advances have sought to adapt NeRFs for spectral and hyperspectral imaging. SpectralNeRF [16] incorporates a spectrum attention mechanism (SAUNet) for realistic low-channel multi-spectral rendering under variable lighting. However, the reliance on attention mechanisms introduces computational overhead, making long-range spectra diffi + +cult to capture. Spec-NeRF [15] takes a more efficient approach by adapting TensoRF [4] for reconstruction, with a seperate MLP estimating the camera's spectral sensitivity, focusing on computationally efficient spectral scene reconstructions. HS-NeRF [5] targets hyperspectral novel view synthesis by learning sinusoidal position encoding for each channel, allowing it to interpolate between viewpoints. This approach struggles with full-spectrum fidelity due to its lack of end-to-end training. HS-NeRF also released two multiview hyperspectral datasets, using two different hyperspectral cameras with varying signal-to-noise ratios. These, NeRF-based methods remain limited by their large parameter counts and instability when handling high-dimensional hyperspectral data alongside slower inference times. + +# 2.3. 3DGS + +Advancements in 3DGS improve point cloud efficiency for hyperspectral imaging but rely on discrete thresholds, causing uneven textures and memory issues with high-dimensional HSI data. VDGS [20] employs a hybrid NeRF neural model for color and opacity, improving spectral reconstruction for HSI. However, this model still heavily depends on the 3DGS prediction and only uses opacity estimation from the MLP. Scaffold-GS [19] introduces compressed representations to smooth surfaces but, like 3DGS, it struggles with artifacts in sparse regions. Mip-Splatting [30] addresses rendering quality by applying 3D smoothing and a 2D Mip filter, yet it remains constrained by viewfrustum-based Gaussian selection, limiting its effectiveness in dynamic scenes. Challenges in Gaussian initialization, especially due to poor SfM data, are tackled by RAINGS [12], which incorporates adaptive optimization. InstantSplat [10] improves stability by using simpler point provisioning alongside the use of dust3r [28] which offers better SFM than COLMAP but has difficulty in processing large amounts of images. Meanwhile, Bulo [3] proposes pixel-level error-based densification to ensure consistent quality, and GaussianPro [6] extends pixel-wise loss methods to better align 3D Gaussian normals during rendering, albeit at the cost of longer training times. EfficientGS[18] and LightGaussian[9] reduce model size and computational load through pruning methods but risk losing important visual details, especially when over-decimation occurs without proper ranking thresholds[18]. These methods often prioritize storage efficiency over rendering performance, overlooking challenges in real-time optimization. Although no previous works have explored the adaptation of Gaussian Splatting to Hyperspectral data, our experiments show that a naive adaptation struggles to model fine details. HyperGS solves these issues by operating in a lower-dimensional learned latent space. + +# 3. Preliminaries + +3DGS, developed by Kerbl et al. [13], represents 3D environments using $N$ Gaussian primitives. Each Gaussian is parameterized by its position $x_{i} \in \mathbb{R}^{3}$ in world coordinates, a scaling vector $\mathbf{S}_i \in \mathbb{R}^3$ defining the size of the Gaussian along each axis, a rotation quaternion $\mathcal{R}_i \in \mathbb{R}^4$ defining its orientation, opacity $\sigma_{i} \in \mathbb{R}$ , which is controlled by a sigmoid function, and appearance features $f_{i} \in \mathbb{R}^{k}$ to represent view-dependent RGB signals. The core of 3DGS lies in transforming these Gaussians from 3D world space into the 2D image plane for rendering. To do this, we first define the 3D covariance matrix of each Gaussian in world space as: + +$$ +\Sigma_ {i} = R _ {i} S _ {i} ^ {2} R _ {i} ^ {T}, \tag {1} +$$ + +where $R_{i}$ is the rotation matrix derived from the quaternion $\mathcal{R}_i$ , and $S_{i}$ is the scaling matrix derived from the vector $\mathbf{S}_i$ . + +To sput the Gaussian onto the 2D image plane, we transform the covariance matrix using the viewing transformation matrix $W$ and the Jacobian matrix $J$ , which handles the projection onto the camera's image plane as + +$$ +\hat {\Sigma} _ {i} = J W \Sigma_ {i} W ^ {T} J ^ {T}. \tag {2} +$$ + +accounting for the perspective distortion and view transformation. When rendering to a particular pixel, the opacity $\alpha_{i}$ of Gaussian $i$ is computed as: + +$$ +\alpha_ {i} = 1 - \exp \left(- \delta_ {i} ^ {T} \hat {\Sigma} _ {i} ^ {- 1} \delta_ {i}\right), \tag {3} +$$ + +where $\delta_{i}$ is the distance between the pixel and the projected center of the Gaussian in 2D. + +The transmittance $T_{i}$ , which represents the cumulative transparency along the ray up to the $i$ -th Gaussian. It is computed as: + +$$ +T _ {i} = \prod_ {j = 1} ^ {i - 1} \left(1 - \sigma_ {i} \alpha_ {j}\right). \tag {4} +$$ + +The final color $C$ of each pixel, $p$ is obtained by blending the colors of all Gaussian's projected onto that pixel: + +$$ +C (p) = \sum_ {i \in N} T _ {i} \alpha_ {i} c _ {i}, \tag {5} +$$ + +where $p$ is the coordinate position in the image, $c_{i}$ is the color of the $i$ -th Gaussian, derived from its color appearance coefficients $f_{i}$ . In summary, this process transforms and projects the Gaussian primitives from 3D space into 2D image space and computes the final pixel colors through opacity and transmittance blending. + +# 4. Methodology + +HyperGS aims to provide a lightweight, fast rasterization solution to HNVS that is robust to different hyperspectral cameras' sensitivity functions with outstanding accuracy. Our system diagram is seen in Figure 2. + +# 4.1. Hyperspectral Compression + +To address the challenges of high-dimensional optimization, we adopt a novel latent space exploration method for the hyperspectral data. The latent space of a pre-trained autoencoder (AE) serves as an exploratory space for our 3DGS system. This approach reduces the computational load during 3DGS optimization by providing a compact, lower-dimensional target. Additionally, the latent space bounds the error by encapsulating the spectral sensitivity of the hyperspectral camera for each channel, enhancing the accuracy and reliability of novel view reconstruction in the spectral domain. During testing, the latent space 3DGS viewpoint estimations are decoded using the frozen AE to produce novel HSIs from which gradient flow will be calculated. + +The compression network is a fast convolutional AE. Both sides are symmetrical and is built using a series of 1D convolutional layers and Squeeze-Excitation (SE) blocks that operate across the spectral dimension of the images. The SE blocks benefit the performance by weighting features between layers for improved spectral discrimination. The model is trained on the pixel level of the scenes dataset. The encoder compresses the high-dimensional hyperspectral data into a latent representation via max-pooling layers, while the decoder reconstructs the original spectral information from this compressed form via upsampling. The architecture omits skip connections to ensure the decoder can function independently during training and testing to decode the LHSI data produced by the 3DGS system. For a detailed layout of the network architecture, please refer to Fig. 3. + +The AE minimizes the loss $L_{ae}$ , defined as: + +$$ +L _ {a e} = L _ {H u b e r} \left(C ^ {*} (p), D e c \left(E n c \left(C ^ {*} (p)\right)\right)\right), \tag {6} +$$ + +where $C^*$ is the ground truth pixel spectrum. The Huber loss provides a smooth optimization process throughout training, as it handles outliers effectively while maintaining smooth gradients. Since each hyperspectral dataset includes varying levels of signal-to-noise ratios, this approach helps improve robustness against noise-related issues. + +# 4.2. Latent Hyperspectral 3D Gaussian Splitting + +Each Gaussian in the splating process is assigned a latent spectral signature $\mathbf{f}_i\in \mathbb{R}^m$ , where $m$ represents the latent space dimensionality. Inspired by many previous RGB techniques [14, 19, 21, 29] an MLP conditioned on view-directional hash encoding is used to predict anisotropic opacity and spectral color information for each Gaussian. This allows view-dependent spectral effects to be modelled upon specific bands. This signature is mapped directly to the pixel values using the decoder model defined in Section 4.1. For camera $d$ , the MLP $\mathcal{F}$ predicts view-dependent + +![](images/51e4291d2d0e351e9e8fb321cc3a7f7e7bf2b36519dc56638cd1ea0b3f291b51.jpg) +Figure 2. Visual system diagram of our approach. Blue lines indicate the operational flow, while orange lines represent the gradient flow. Our novel modified latent hyperspectral adaptive density function operates within a latent space provided by a frozen autoencoder, which is also responsible for decoding the final images. Latent space novel views are generated through a combination of the latent hyperspectral Gaussian point cloud and a NeRF-style MLP. Gradients are updated based on the decoded images. + +![](images/b7612df0e08b2b813be9f51cc446f7e0fca18a6e8d6626c156db84129bf8ede0.jpg) +Figure 3. Our channel-wise convolutional AE model learns LHSI space representation of the scene. The decoder is only used in training and inference for the 3DGS system after the preprocessing of the dataset is finished. + +spectral effects $\tilde{f}_{i,d}$ and opacity $\tilde{\sigma}_{i,d}$ effects for each Gaussian. Specifically, the MLP takes the centre $\mathbf{m}_i$ of each Gaussian and the view direction as inputs: + +$$ +[ \tilde {f} _ {i, d}, \tilde {\sigma} _ {i, d} ] = \mathcal {F} _ {v} (h (\mathbf {m} _ {i}, \mathbf {d}); \Theta), \tag {7} +$$ + +where $h$ is the hash encoding of the inputs while $\tilde{\Theta}$ denotes the MLP parameters similar to that of I-NGP [24] and Mip-Nerf360 [1]. + +These view-dependent spectral and opacity effects are multiplied with those stored in the Gaussian Cloud, leading to the updated volumetric rendering equations with transmittance being defined as: + +$$ +T _ {i, d} = \prod_ {j = 1} ^ {i - 1} (1 - \alpha_ {j} \sigma_ {i} \tilde {\sigma} _ {i, d}). \tag {8} +$$ + +The final latent signature $\hat{C}$ of each pixel is obtained by blending the colors of all Gaussians projected onto that pixel $p$ : + +$$ +\hat {C} (p, d) = \sum_ {i \in N} T _ {i, d} \alpha_ {i} f _ {i} \tilde {f} _ {i, d}. \tag {9} +$$ + +It is worth emphasizing that these view-dependent spectral effects are applied within the learned latent space before decoding. This helps maintain a low computational complexity while minimizing outliers. + +A decoding operation is then performed to provide the full fidelity prediction from the system using the decoder from Section 4.1. + +$$ +C (p, d) = \operatorname {D e c} (\hat {C} (p, d)). \tag {10} +$$ + +The final rendering ensures that the spectral integrity is maintained throughout the entire process, which is critical for accurately reconstructing material properties. By leveraging the latent space, HyperGS provides a more meaningful and efficient spectral representation of the scene, than the original full hyperspectral image. The original loss proposed in 3DGS for spatial and geometric consistency in rendered images uses a weighted DSSIM and L1 loss. However, this type of loss have been shown to lead to underfitting for HSI in other fields. This is because an L1 loss can produce extreme errors for sensitive spectral bands, destabilizing training. To address this, we employ a weighted loss combining Charbonnier loss and cosine similarity to account for both spectral quality and spatial consistency. + +![](images/c81695b20ce3f44a3a7c92a7eef62682d17dd3def000b421008ce249f9392f8b.jpg) +Figure 4. Visualization of our re-projection protocol for initializing the SFM point cloud. We estimate the SFM from grayscale channel slices of the hyperspectral image scene with COLMAP. Then, using the average of all views, we re-project each point into LHSI from provided by our AE, providing an optimal initialization of latent spectral color for the 3DGS system. + +The cosine similarity provides a measured angular distance between vectors, making it ideal for comparing spectra. To provide an overall consistency in the spatial and spectral domain we keep the SSIM score in the training from the original 3DGS system. The per pixel HyperGS training loss $L_{d}(p)$ is: + +$$ +L _ {d} (p) = (1 - \lambda) (\beta L _ {C B} (p) + L _ {C S} (p)) + \lambda L _ {S S I M} (p), \tag {11} +$$ + +where $\beta$ weighs the spectral loss balancing the cosine similarity and charbonnier loss. Thus the total training loss across all pixels and views is: + +$$ +L (C, C ^ {*}) = \sum_ {p = 1} ^ {P} \sum_ {d \in V} L _ {d} (p), \tag {12} +$$ + +where $C$ is the 3DGS systems network prediction after decoding the entire latent image, $C^{*}$ is the ground truth hyperspectral image, and $V$ is the set of training view directions. + +# 4.2.1 Initialisation + +Since there is no dedicated hyperspectral Structure-from-Motion (SfM) algorithm compatible with COLMAP, we first convert the hyperspectral images into grayscale images, $I_{\mathrm{gray}}$ by selecting the spectral channel with the highest foreground intensity variance to preserve key features for the SfM process. Our process is visualized in Figure 4. Using these grayscale images and their corresponding camera projection matrices for each view $d$ , we apply COLMAP to generate a sparse 3D point cloud. Each 3D point $\mathbf{X} = (X,Y,Z,1)^T$ in world coordinates is related to its pixel coordinate $\mathbf{p}_d = (x_d,y_d,1)^T$ in the image plane via the camera projection $\mathbf{K}_d\mathbf{E}_d$ : + +$$ +\mathbf {p} _ {d} = \mathbf {K} _ {d} \mathbf {E} _ {d} \mathbf {X}, \tag {13} +$$ + +where $\mathbf{K}_d$ are the camera intrinsics and $\mathbf{E}_d$ the extrinsics. After generating 3D points and recovering camera poses, + +we re-project the points into the LHSIs. Each 3D point $\mathbf{X}$ is linked to pixel coordinates $\mathbf{p}_d$ across all views. We then initialize the Gaussian cloud with one Gaussian centered on each 3D point $\mathbf{X}$ , with spectral signature $\mathbf{f}_i$ computed by averaging the latent hyperspectral signatures $H_{d}(y_{d},x_{d})$ across all views $d$ : + +$$ +f _ {i} = \frac {1}{| V |} \sum_ {v = 1} ^ {V} \hat {C} _ {d} ^ {*} (\mathbf {p} _ {d}), \tag {14} +$$ + +where $\hat{C}_d^*(\mathbf{p}_d)$ is the latent hyperspectral signature at pixel $p$ for view $d$ , and $V$ is the set of training views. This averaging ensures the spectral information is captured robustly. + +# 4.2.2 Latent Hyperspectral Densification + +As the number of color channels in HSI data increases, visual discontinuities become more prevalent. Additionally, our initial SFM reconstruction is less dense compared to standard RGB 3DGS. To address these issues, we incorporate an advanced densification method that enhances stability in both results and training. A key component of 3DGS involves determining whether a Gaussian should be split or cloned based on the gradient magnitude of the Normalized Device Coordinates (NDC) across various viewpoints. However, in sparse regions, this can cause artefacts as large Gaussians are highly visible in many viewpoints, leading to inconsistent splitting. This challenge is exacerbated in the hyperspectral domain of 3DGS, where the larger number of channels and highly variable value ranges make it harder to set effective thresholds. To mitigate this, we scale the gradient by the square of the depth relative to the scene's radius, which reduces the influence of Gaussians near the camera. More formally, we define the splitting score as: + +$$ +\mathcal {S} \left(g _ {i}\right) = \sum_ {d \in V} \sum_ {p = 1} ^ {P} h (d, i) \sqrt {\left(\frac {\partial L _ {d} (p)}{\partial x _ {i}}\right) ^ {2} + \left(\frac {\partial L _ {d} (p)}{\partial y _ {i}}\right) ^ {2}}, \tag {15} +$$ + +where the depth-scaling function is: + +$$ +h (d, i) = \left(\frac {\left| \mathbf {E} _ {d} \mathbf {X} _ {i} \right|}{\beta_ {\text {f i e l d}} \times R}\right) ^ {2}, \tag {16} +$$ + +$\beta_{\mathrm{field}}$ is a tunable parameter, and $R$ is the scene's radius, based on the largest distance between any pair of cameras. + +The Gaussian $g_{i}$ is split or cloned if $S(g_{i}) > \theta_{q}$ . This approach accounts for depth-based scaling and pixel contributions, leading to better hyperspectral scene reconstruction with fewer artifacts near the camera. However, this more expressive densification can result in an excessive number of Gaussians in scenes with high-frequency details (e.g., the "pinecone" scene in Figure 1, to stabilize this we deploy a global pruning strategy. + +![](images/73dbb521fad41f339471427d908def227f2f85e65b512739e441adbefdfd0972.jpg) +Figure 5. Visualization of our pixel-wise pruning. Gaussians with poor similarity scores below a threshold are pruned, such as the red-circled Gaussian and the black ground truth spectrum. + +# 4.3. Global Hyperspectral Gaussian Pruning + +In HyperGS, we deploy pixel-wise pruning [25] to preserve the spectral integrity of latent hyperspectral images (LHSIs) while lowering the overall number of primitives to improve the scene quality, model size, and spectral gradient descent. Unlike cross-view pruning methods, which evaluate Gaussians across multiple viewpoints and leads to over-pruning [11, 18], our approach assesses each Gaussian's contribution at the pixel level. This approach retains fine spectral details specific to each pixel. For each Gaussian $g_{i}$ , pruning is based on its pixel-wise importance $\mathcal{I}$ across all viewpoints and pixels. We compute the importance for every combination of Gaussian, pixel and viewpoint as the spectral difference between the ground truth hyperspectral value $\hat{C}_d^* (p)$ and the decoded Gaussians LHSI, with the pruning score defined as: + +$$ +\mathcal {I} [ g _ {i}, p, d ] = \left(1 - \left| C _ {d} ^ {*} (p) - D e c \left(f _ {i}\right) \right|\right) \alpha_ {i} T _ {i}, \tag {17} +$$ + +where $T_{i}$ is the accumulated transmittance of Gaussian $g_{i}$ . Including $\alpha_{i}$ and $T_{i}$ ensures that only Gaussians contributing significantly in visibility and spectral accuracy are retained, and also ensures that the score is 0 for Gaussians that do not overlap the given pixel. + +Following recent works, we do not prune Gaussians based on the average, or the total pixel-wise importance score. Instead, we retain all Gaussians within the top-K importance ranking for any pixel. More formally the filtered Gaussian cloud $\mathcal{G}$ is + +$$ +\mathcal {G} = \left\{g _ {i} \mid \exists (p, d), \operatorname {R a n k} \left(g _ {i}, \mathcal {I} [:, p, d ] < \tau_ {p}\right) \right\}, \tag {18} +$$ + +where "Rank" is a function that returns the rank of a given Gaussian within the slice of the score tensor for that pixel and view. + +# 5. Experiments + +We first evaluate HyperGS using the HS-NeRF datasets [5]. These datasets differ in signal-to-noise ratios, channel lengths, and the number of images per scene, providing a comprehensive test of model performance. The Bayspec + +dataset [5] contains around 360 images per scene (3 scenes total), while the SOP dataset [5] has around 40 images per scene (4 scenes total). Due to the lack of continuous HSI datasets for non-object-centric scenes, we also evaluate HyperGS on a synthetic dataset curated from ScanNetv200 [7] (Section 9 in the supplementary materials, where we replace each semantic label with distinct hyperspectral signatures as seen in our supplementary materials). The baselines we have implemented for the HNVS problem include conventional 3DGS (with our reprojected SFM initialisation 4.2.1 and removed SH coefficients), traditional NeRF models adapted for hyperspectral data (NeRF [22], MipNeRF [2] MipNeRF-360 [1], Nerfacto [26] and TensorF [4] (SpecNeRF), and the only existing HNVS method HS-NeRF [5]. + +All experiments were conducted on an NVIDIA A100 80GB GPU, as MipNeRF360 requires significantly more VRAM than the other methods, which could run on a NVIDIA 3090 GPU. Using the A100 also facilitated accurate tracking of training times across all methods. We assess the performance of HyperGS after 60k training iterations, while NeRF methods are trained with the standard 1024 rays per batch and competing methods using rigorous metrics that capture both accuracy and spectral fidelity: PSNR, SSIM, Spectral Angle Mapping (SAM), RMSE. We use a $90\%$ training and $10\%$ test split according to the HS-NeRF dataset. + +# 5.1. Real HSI + +We compare two HNVS datasets from HS-NeRF with different spectral channels and noise levels. The SOC710-VP (SOP) camera, covering $370 - 1100\mathrm{nm}$ , provides high spatial $(696\times 520)$ and spectral (128 channels) resolution but suffers from poor temporal resolution due to long exposure. The BaySpec GoldenEye camera offers comparable spatial $(640\times 512)$ and spectral (141 channels) resolution with faster captures, introducing more noise. Tables 1 and 2 show that HyperGS consistently outperforms other methods on unseen images across all scenes. + +The BaySpec dataset particularly highlights HyperGS's strengths. Its autoencoder produces bounded errors in latent space, letting HyperGS effectively manage noise by averaging spectra and decoding within well-defined 1D feature structures. This yields more accurate reconstructions than standard end-to-end training methods. Figure 6 further shows how HyperGS better predicts object size and specular reflections compared to other baselines. + +Similar to its performance on natural NeRF tasks, MipNeRF360 performed the best among NeRF methods for HNVS. We attribute its performance to its use of positional encoding and the efficient scene representation using tensor decomposition initially employed by TensorF providing more targeted spectral predictions. Our densification and pruning techniques further refine HyperGS's performance + +![](images/a49e2bc0f82459c8fb593d4372c05c484abc4ea2552c2606224741f5d2af7e32.jpg) + +![](images/12157b5826d1c8eb0ec2fc694f54ac3bd81a1b93eb18273eb0f53adcc4803fc4.jpg) + +![](images/19027e83ac16bd4a7b32b45b63e0022ee2de1760a6056b8cc6dd2208ef61268f.jpg) + +![](images/703e124c97e00ef763508d76056331528e4a47819ef5cd428d5974a0950b546e.jpg) + +![](images/9b6399da5fc0a596fb6a4837f8360145f3e0b549d6dfce57d907848182645a70.jpg) + +![](images/c68127b80fafa5b00deea2c061f2cf18838ac7b21ae659baa02c93ed04470294.jpg) + +![](images/36bc1370b0629e090bfb637c19279218c95b40cd527bb2b3aa67eef1f89cb8b2.jpg) + +![](images/be510ed92286bd483b60c30bd2e70245cfcb14ec3fc0024714efd651272a255d.jpg) + +![](images/e698b2475801188354984d79756ad257707a031fd0684eb5b3108f1a335f51b4.jpg) + +![](images/6bbe325aa10cf4fc3727a0c93cd375a338f0f93c6029a1c019cc0cfcb9cde93f.jpg) +Figure 6. Visualisation of the top 4 methods for frame 51 of 359 for the Caladium plant scene from the Bayspec dataset. The top row shows the 70th channel of the 141 channel image predicted, the bottom row provides a raw pixel-wise error heatmap of the scene. + +![](images/e89f175cbb169ca19add91b717112882ab03a5214b2715d3b7395437e1a546b5.jpg) + +![](images/9ca7e8fddbc7a4a652cf945b8e75f8ea8de544b6213f1100ce8ed48be0425cd9.jpg) + +![](images/617c68515f5d61d85207e307c57c4afa8234d969986d4736bbdb1d7c15a8fe4d.jpg) + +![](images/4d604f66b83a9f7a46ebc4672cbbd0708aef82715a3d4327141b4479d113952a.jpg) + +![](images/e5ba4e3e1358514855a7043540f6e9f9667e9fa74b67b9bdf72f8f4545002a42.jpg) + +![](images/18cbc74fb3341a1027b07275d13246f0822c0b07f4a4298f11b2e2fc99f22708.jpg) + +![](images/24e1860ad484263dfb17faf68fb1d49ba5eaf9fa637c061aebfcc815fbb68722.jpg) + +![](images/ac66b354eaf0902cd3201aeeb964e6fb6618432a63d6bad548238d3d82290924.jpg) + +![](images/41a7de3963f588c8bbd049a5aa51ca7814198601dbfdd0e38349fac7248ad620.jpg) + +![](images/8eca2ed85e5e17b5fbdc92be03dbd0524171e46e212fe829131fbc464e671517.jpg) +Figure 7. Visualisation of the top 4 methods for frame 31 of 49 for the Tools plant scene from the SOP dataset. The top row shows the 70th channel of the 128 channel image predicted, the bottom row provides a raw pixel-wise error heatmap of the scene. + +by effectively filtering poor spectral representations. Meanwhile, the SOP dataset favors both 3DGS approaches. Due to the high number of frames present in the BaySpec camera scenes, the NeRF approaches adapt well to this abundance of data and can provide far better and more stable representations of the scene given the volume of viewpoints. This volume makes predicting the noisier spectra that the BaySpec camera provides easier than 3DGS. In contrast, although the SOP datasets may have smoother spectra, the number of training viewpoints is far smaller. This leads to NeRF approaches evidently struggling with scene scale and spectral reconstruction, as they struggle to understand the scene scale and perform spectral reconstruction due to the lack of data. In contrast, 3DGS approaches excel due to smoother data interpolation and simpler optimization. This is outlined by Figure 7 highlighting the intense error heatmaps from NeRF approaches compared to 3DGS and HyperGS. HyperGS, however, outperforms all methods across all scenes and camera datasets, suggesting that it provides a far more robust and accurate performance than both 3DGS and NeRF methods for HNVS. For visualiza + +BaySpec Datasets + +
MethodAverage Results
PSNR ↑SSIM↑SAM↓RMSE↓
NeRF23.350.60610.04400.0687
MipNeRF22.750.59470.04350.0776
TensoRF24.660.64820.05010.0587
Nerfacto19.120.58660.05510.1174
MipNerf36026.530.74420.02800.0476
HS-NeRF19.820.67140.05340.1071
3DGS22.910.63210.13350.0810
HyperGS27.110.78040.02540.0440
+ +Table 1. Quantitative results using the Bayspec datasets against separate hyperspectral methods and baseline NeRF and 3DGS (best bold, second best italic). + +tions of random pixels and spectral images, please refer to Figures 8 and then 9 which showcases HyperGS's superior pixel reconstruction capabilities compared to other benchmarked methods. + +![](images/93fdca6666af58207f82e65fa5c76f3b5f49e0f34e1fe6d2adbe247383352798.jpg) +Figure 8. Three random pixel reconstructions taken from test frame 151 of the bayspec Anacampseros scene. + +![](images/ffa9223e76a854222c5389fc41bd6fc4c74c538f20d3e44e74ac659be534e68f.jpg) + +![](images/51c6f99a93dfbfaedcd6dbdf9eed0ab238e53592c618c701eb7453682797c033.jpg) + +![](images/9387353f18a257f80946db0266b31f023ac592de2a94d9d6a5cb69c411f1fa75.jpg) +Figure 9. Three random pixel reconstructions from test frame 31 of the SOP origami scene. + +![](images/b8777cd475ed739d5d64250c1d8e8cc0bd03be14bff316740a3f618e9e85924c.jpg) + +![](images/7e3388e5f38d886dd81476a083d3e2dcb95fabf0553c9480a79f4a7dd784e557.jpg) + +Surface Optics Datasets + +
MethodAverage Results
PSNR ↑SSIM↑SAM↓RMSE↓
NeRF10.890.59050.06250.3479
MipNeRF12.530.54810.05680.3198
TensorF13.000.56960.05950.2780
Nerfacto16.370.69860.03520.1601
MipNeRF36012.280.68240.13690.2658
HS-NeRF14.440.61650.20370.1953
3DGS28.580.96270.03010.0478
HyperGS30.510.97560.004150.0354
+ +Table 2. Quantitative results using the SOP datasets against separate hyperspectral methods and baseline NeRF and 3DGS. + +# 5.2. Ablation Study + +To evaluate the effectiveness and impact of various components in our HyperGS model, we conducted a series of ablation experiments. These experiments help to understand the contribution of individual components and choices in our model's design. As shown in Table 3, each new feature added to the baseline 3DGS model improves HNVS performance. The joint latent autoencoder-3DGS architecture boosts spectral reconstruction the most while the proposed densification provides greater overall details captured. With the latent space learning and positional encoding, the performance of HyperGS gets a significant boost in performance too. Please refer to our supplementary materials for additional ablations, including analyses of HyperGS performance, pruning strategy scoring functions, autoencoder performance with varying data and latent sizes, the global pruning strategy, and system speed bottlenecks. + +
Ablation StepAverage Results for the Bayspec dataset
PSNR ↑SSIM↑SAM↓RMSE↓N.Prim ↓
Base. 3DGS22.910.63200.13350.0810440k
+ Spec. SFM23.050.63310.13100.0799421k
+ Latent space AE24.870.71010.05480.0602500k
+ Densification25.250.73560.03650.05481.3M
+ Pruning25.170.71990.03740.0555412k
+ View dependent MLP27.050.77920.02530.0443309k
+ Custom L.Function27.110.78010.02540.0440310k
+ +Table 3. Adding each new feature to 3DGS significantly improves the model's average performance for the HNVS task in the bayspec dataset while creating a more stable number of primitives (N.Prim) in the point cloud. + +# 6. Conclusions + +In this paper, we introduced HyperGS, the first effective approach for hyperspectral novel view synthesis (HNVS). By encoding latent spectral data within Gaussian primitives and performing Gaussian splatting in a learned latent space, HyperGS achieves high-quality, material-aware rendering. Our adaptive density control and pruning techniques efficiently handle latent hyperspectral signatures in 3D Gaussian point clouds, ensuring stable training and superior accuracy compared to benchmark methods. For future work, we aim to extend HyperGS with more comprehensive latent space decoding that corrects and refines both spectral and spatial information, providing richer contextual detail. + +# 7. Acknowledgements + +This work was partially supported by Industrial 3D Robotics (I3D), IS-Instruments and partially funded by the EPSRC under grant agreement EP/S035761/1. + +# References + +[1] J.T. Barron, B. Mildenhall, D. Verbin, P.P. Srinivasan, and P. Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 4, 6 +[2] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields, 2021. 6 +[3] S. R. Bulò, L. Porzi, and P. Kontschieder. Revising densification in gaussian splatting. ArXiv, 2024. 2 +[4] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision (ECCV), 2022. 2, 6 +[5] G. Chen, S. K. Narayanan, T. G. Ottou, B. Missaoui, H. Muriki, C. Pradalier, and Y. Chen. Hyperspectral neural radiance fields. ArXiv, 2024. 2, 6 +[6] Kai Cheng, Xiaoxiao Long, Kaizhi Yang, Yao Yao, Wei Yin, Yuexin Ma, Wenping Wang, and Xuejin Chen. Gaussianpro: 3d gaussian splatting with progressive propagation. arXiv preprint arXiv:2402.14650, 2024. 2 +[7] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proc. Computer Vision and Pattern Recognition (CVPR), IEEE, 2017. 6, 1 +[8] F. Dellaert. NeRF explosion 2020. https://dellaert.github.io/NeRF/, 2020. 2 +[9] Z. Fan, K. Wang, K. Wen, Z. Zhu, D. Xu, and Z. Wang. Lightgaussian: Unbounded 3d gaussian compression with 15x reduction and $200+$ fps. arXiv preprint arXiv:2311.17245, 2023. 2 +[10] Z. Fan, W. Cong, K. Wen, K. Wang, J. Zhang, X. Ding, D. Xu, B. Ivanovic, M. Pavone, G. Pavlakos, Z. Wang, and Y. Wang. Instantsplat: Sparse-view sfm-free gaussian splatting in seconds. ArXiv, 2024. 2 +[11] S. Girish, K. Gupta, and A. Shrivastava. Eagles: Efficient accelerated 3d gaussians with lightweight encodings. arXiv preprint arXiv:2312.04564, 2023. 6 +[12] Jaewoo Jung, Jisang Han, Honggyu An, Jiwon Kang, Seonghoon Park, and Seungryong Kim. Relaxing accurate initialization constraint for 3d gaussian splatting. arXiv preprint arXiv:2403.09413, 2024. 2 +[13] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (SIGGRAPH Conference Proceedings), 42(4), 2023. 1, 3 +[14] Byeonghyeon Lee, Howoong Lee, Xiangyu Sun, Usman Ali, and Eunbyung Park. Deblurring 3d gaussian splatting, 2024. 3 +[15] J. Li, Y. Li, C. Sun, C. Wang, and J. Xiang. Spec-nerf: Multispectral neural radiance fields. ArXiv, 2023. 2 +[16] R. Li, J. Liu, G. Liu, S. Zhang, B. Zeng, and S. Liu. Spectralnerf: Physically based spectral rendering with neural radiation field. ArXiv, 2023. 2 + +[17] Jie Liang, Ali Zia, Jun Zhou, and Xavier Sirault. 3d plant modelling via hyperspectral imaging. In 2013 IEEE International Conference on Computer Vision Workshops, pages 172-177, 2013. 2 +[18] W. Liu, T. Guan, B. Zhu, L. Ju, Z. Song, D. Li, Y. Wang, and W. Yang. Efficientges: Streamlining gaussian splatting for large-scale high-resolution scene representation. ArXiv, 2024. 2, 6 +[19] T. Lu, M. Yu, L. Xu, Y. Xiangli, L. Wang, D. Lin, and B. Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. ArXiv, 2023. 2, 3 +[20] D. Malarz, W. Smolak, J. Tabor, S. Tadeja, and P. Spurek. Gaussian splatting with nef-based color and opacity. ArXiv, 2023. 2 +[21] D. Malarz, W. Smolak, J. Tabor, S. Tadeja, and P. Spurek. Gaussian splatting with nef-based color and opacity. ArXiv, 2023. 3 +[22] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 6 +[23] B. Mildenhall, P. Hedman, R. Martin-Brualla, P.P. Srinivasan, and J.T. Barron. Nerf in the dark: High dynamic range view synthesis from noisy raw images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[24] T. Müller, A. Evans, C. Schied, and A. Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (TOG), 41(4):102:1-102:15, 2022. 4 +[25] S. Niedermayr, J. Stumpfegger, and R. Westermann. Compressed 3d gaussian splatting for accelerated novel view synthesis. arXiv preprint arXiv:2401.02436, 2023. 6 +[26] M. Tancik, E. Weber, E. Ng, R. Li, B. Yi, J. Kerr, T. Wang, A. Kristoffersen, J. Austin, K. Salahi, A. Ahuja, D. McAllister, and A. Kanazawa. Nerfstudio: A modular framework for neural radiance field development. arXiv preprint arXiv:2302.04264, 2023. 6 +[27] Christopher Thirgood, Oscar Mendez, Erin Chao Ling, Jon Storey, and Simon Hadfield. Raspectloc: Raman spectroscopy-dependent robot localisation. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5296-5303, 2023. 1 +[28] S. Wang, V. Leroy, Y. Cabon, B. Chidlovskii, and J. Revaud. DUSt3R: Geometric 3D Vision Made Easy. arXiv preprint arXiv:2312.14132, 2023. 2 +[29] Z. Yang, X. Gao, Y. Sun, Y. Huang, X. Lyu, W. Zhou, S. Jiao, X. Qi, and X. Jin. Spec-Gaussian: Anisotropic View-Dependent Appearance for 3D Gaussian Splatting. ArXiv, 2024. 3 +[30] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19447-19456, 2024. 2 +[31] A. Zia, J. Liang, J. Zhou, and Y. Gao. 3d reconstruction from hyperspectral images. In 2015 IEEE Winter Conference on + +Applications of Computer Vision (WACV), pages 318-325, 2015. 2 +[32] Ali Zia, Jie Liang, Jun Zhou, and Yongsheng Gao. 3d reconstruction from hyperspectral images. In 2015 IEEE Winter Conference on Applications of Computer Vision, pages 318-325, 2015. 2 \ No newline at end of file diff --git a/2025/HyperGS_ Hyperspectral 3D Gaussian Splatting/images.zip b/2025/HyperGS_ Hyperspectral 3D Gaussian Splatting/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4b690453bbe9051e6799213954a346b1e0e977af --- /dev/null +++ b/2025/HyperGS_ Hyperspectral 3D Gaussian Splatting/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5f90bae2db5a61c0247839ba6dcec119bf29dc39a2f57010ba78e5f66932d3d6 +size 707678 diff --git a/2025/HyperGS_ Hyperspectral 3D Gaussian Splatting/layout.json b/2025/HyperGS_ Hyperspectral 3D Gaussian Splatting/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3d9e2d74ba34de6accc5bf48e06e4ce1fad86e1e --- /dev/null +++ b/2025/HyperGS_ Hyperspectral 3D Gaussian Splatting/layout.json @@ -0,0 +1,8816 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 148, + 103, + 446, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 103, + 446, + 121 + ], + "spans": [ + { + "bbox": [ + 148, + 103, + 446, + 121 + ], + "type": "text", + "content": "HyperGS: Hyperspectral 3D Gaussian Splitting" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 97, + 142, + 494, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 142, + 494, + 171 + ], + "spans": [ + { + "bbox": [ + 97, + 142, + 494, + 171 + ], + "type": "text", + "content": "Christopher Thirgood1, Oscar Mendez1, Erin Ling1, Jon Storey2, Simon Hadfield1 \n1University of Surrey, 2I3D Robotics" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 75, + 174, + 518, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 174, + 518, + 186 + ], + "spans": [ + { + "bbox": [ + 75, + 174, + 518, + 186 + ], + "type": "text", + "content": "{c.thirgood, o.mendez, chao.ling, s.hadfield}@surrey.ac.uk, jstorey@i3drobotics.com" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 52, + 217, + 255, + 380 + ], + "blocks": [ + { + "bbox": [ + 52, + 217, + 255, + 380 + ], + "lines": [ + { + "bbox": [ + 52, + 217, + 255, + 380 + ], + "spans": [ + { + "bbox": [ + 52, + 217, + 255, + 380 + ], + "type": "image", + "image_path": "6e48a9e7b94b09f46ccb225d769b211c2449103ace1770ab95fcd31510082b49.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 394, + 547, + 428 + ], + "lines": [ + { + "bbox": [ + 46, + 394, + 547, + 428 + ], + "spans": [ + { + "bbox": [ + 46, + 394, + 547, + 428 + ], + "type": "text", + "content": "Figure 1. This image represents a novel hyperspectral image at the 70th channel of a 141-channel image predicted from the top three models, HyperGS (Ours), 3DGS and MipNerf360 for the hyperspectral novel view synthesis task. Two random pixel reconstructions taken from the novel view with corresponding border color for its plot and pixel location in the image represented by the stars." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 258, + 217, + 335, + 291 + ], + "blocks": [ + { + "bbox": [ + 258, + 217, + 335, + 291 + ], + "lines": [ + { + "bbox": [ + 258, + 217, + 335, + 291 + ], + "spans": [ + { + "bbox": [ + 258, + 217, + 335, + 291 + ], + "type": "image", + "image_path": "c67c80e12e0dc0c72fbd1bed1aae5d5e3cf6efca0683a487d09f1d7238a6529e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 337, + 217, + 414, + 291 + ], + "blocks": [ + { + "bbox": [ + 337, + 217, + 414, + 291 + ], + "lines": [ + { + "bbox": [ + 337, + 217, + 414, + 291 + ], + "spans": [ + { + "bbox": [ + 337, + 217, + 414, + 291 + ], + "type": "image", + "image_path": "eea12ede0c05c5c776e3bb025e81dd0dbb15895ba9113cc30f08e84c248114c3.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 258, + 300, + 335, + 372 + ], + "blocks": [ + { + "bbox": [ + 273, + 293, + 321, + 300 + ], + "lines": [ + { + "bbox": [ + 273, + 293, + 321, + 300 + ], + "spans": [ + { + "bbox": [ + 273, + 293, + 321, + 300 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 258, + 300, + 335, + 372 + ], + "lines": [ + { + "bbox": [ + 258, + 300, + 335, + 372 + ], + "spans": [ + { + "bbox": [ + 258, + 300, + 335, + 372 + ], + "type": "image", + "image_path": "8621a675894e5482f553b4813f7075822336ac91bb6ea4e7bb0dea36b24f207a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 288, + 373, + 308, + 380 + ], + "lines": [ + { + "bbox": [ + 288, + 373, + 308, + 380 + ], + "spans": [ + { + "bbox": [ + 288, + 373, + 308, + 380 + ], + "type": "text", + "content": "3DGS" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 337, + 293, + 414, + 372 + ], + "blocks": [ + { + "bbox": [ + 337, + 293, + 414, + 372 + ], + "lines": [ + { + "bbox": [ + 337, + 293, + 414, + 372 + ], + "spans": [ + { + "bbox": [ + 337, + 293, + 414, + 372 + ], + "type": "image", + "image_path": "c61d492aa5a94b4f7c99c1c4e0348094c5ffa5822e1643e063272d67133e387a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 350, + 373, + 402, + 381 + ], + "lines": [ + { + "bbox": [ + 350, + 373, + 402, + 381 + ], + "spans": [ + { + "bbox": [ + 350, + 373, + 402, + 381 + ], + "type": "text", + "content": "HyperGS (Ours)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 417, + 217, + 542, + 296 + ], + "blocks": [ + { + "bbox": [ + 417, + 217, + 542, + 296 + ], + "lines": [ + { + "bbox": [ + 417, + 217, + 542, + 296 + ], + "spans": [ + { + "bbox": [ + 417, + 217, + 542, + 296 + ], + "type": "image", + "image_path": "e4b9aa598ff9d4db5eb2b869b2007cd4a8ad99f61e1b4353b00ee90f2cfa5393.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 417, + 296, + 542, + 384 + ], + "blocks": [ + { + "bbox": [ + 417, + 296, + 542, + 384 + ], + "lines": [ + { + "bbox": [ + 417, + 296, + 542, + 384 + ], + "spans": [ + { + "bbox": [ + 417, + 296, + 542, + 384 + ], + "type": "image", + "image_path": "91ecb10804c88bdb4a18721e416177e703623c0a974adf95025dd6a1a0b088a1.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 143, + 436, + 192, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 436, + 192, + 449 + ], + "spans": [ + { + "bbox": [ + 143, + 436, + 192, + 449 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 45, + 472, + 289, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 472, + 289, + 700 + ], + "spans": [ + { + "bbox": [ + 45, + 472, + 289, + 700 + ], + "type": "text", + "content": "We introduce HyperGS, a novel framework for Hyperspectral Novel View Synthesis (HNVS), based on a new latent 3D Gaussian Splitting (3DGS) technique. Our approach enables simultaneous spatial and spectral renderings by encoding material properties from multi-view 3D hyperspectral datasets. HyperGS reconstructs high-fidelity views from arbitrary perspectives with improved accuracy and speed, outperforming currently existing methods. To address the challenges of high-dimensional data, we perform view synthesis in a learned latent space, incorporating a pixel-wise adaptive density function and a pruning technique for increased training stability and efficiency. Additionally, we introduce the first HNVS benchmark, implementing a number of new baselines based on recent SOTA RGB-NVS techniques, alongside the small number of prior works on HNVS. We demonstrate HyperGS's robustness through extensive evaluation of real and simulated hyperspectral scenes with a 14db accuracy improvement upon previously published models." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 436, + 387, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 436, + 387, + 449 + ], + "spans": [ + { + "bbox": [ + 307, + 436, + 387, + 449 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 459, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 459, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 459, + 545, + 555 + ], + "type": "text", + "content": "Recent advancements in Novel View Synthesis (NVS), particularly with implicit Neural Radiance Field modeling (NeRFs) [22] and explicit Gaussian models (3DGS)[13], have dramatically improved the fidelity of synthetically rendered views for conventional RGB images. However, RGB imaging is fundamentally limited, as it cannot capture the detailed material properties or non-visible scene characteristics critical for deeper scene understanding." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 557, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 546, + 715 + ], + "type": "text", + "content": "Hyperspectral imaging addresses these limitations by capturing a continuous spectrum of light across narrow bands for each pixel, providing valuable material properties and subtle spectral signatures. This modality is crucial for applications such as remote sensing, medical diagnostics, environmental monitoring, and robotics [27], where spectral information is paramount. The demand for Hyperspectral Novel View Synthesis (HNVS) is driven by the need for accurate, real-time spectral and spatial modeling. However, synthesizing novel views from hyperspectral data presents significant challenges due to its high dimensionality and the requirement for spectral consistency across different perspectives for each pixel." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5970" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "content": "Previous works, such as HS-NeRF [5], have explored HNVS by adapting NeRF architectures to hyperspectral data. However, these methods suffer from unstable training dynamics, slow rendering times, and inefficiencies in handling high-dimensional." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 286, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 286, + 287 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 286, + 287 + ], + "type": "text", + "content": "3DGS has recently emerged as a promising approach for fast NVS in the RGB domain, offering efficient, high-fidelity rendering of 3D scenes through point-based representations. By projecting 3D Gaussians onto the image plane, 3DGS enables smooth, continuous representations of scene geometry while preserving fine surface details. However, despite its efficiency and rendering quality, 3DGS has not yet been successfully extended to accommodate high-dimensional data types. In our benchmark, traditional 3DGS fails to achieve consistent results against NeRF baselines for HNVS (Section 5 and Section 9 of the supplementary materials). HyperGS aims to address these challenges. In summary, the contributions of this paper are:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 289, + 286, + 407 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 47, + 289, + 286, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 289, + 286, + 323 + ], + "spans": [ + { + "bbox": [ + 47, + 289, + 286, + 323 + ], + "type": "text", + "content": "1. We present the first method that successfully integrates view-dependent hyperspectral material information with 3DGS for high-quality HNVS." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 324, + 286, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 324, + 286, + 371 + ], + "spans": [ + { + "bbox": [ + 47, + 324, + 286, + 371 + ], + "type": "text", + "content": "2. We introduce an adaptive density control and global pruning process that leverages hyperspectral signatures for efficiency alongside a hyperspectral SFM process to stabilize the modeling." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 372, + 286, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 372, + 286, + 407 + ], + "spans": [ + { + "bbox": [ + 47, + 372, + 286, + 407 + ], + "type": "text", + "content": "3. A comprehensive benchmark of hyperspectral NeRF approaches and adaptations of classical RGB-NVS models for HNVS." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 420, + 151, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 420, + 151, + 431 + ], + "spans": [ + { + "bbox": [ + 47, + 420, + 151, + 431 + ], + "type": "text", + "content": "2. Literature review" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 440, + 223, + 452 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 440, + 223, + 452 + ], + "spans": [ + { + "bbox": [ + 47, + 440, + 223, + 452 + ], + "type": "text", + "content": "2.1. Hyperspectral 3D reconstruction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 459, + 286, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 459, + 286, + 578 + ], + "spans": [ + { + "bbox": [ + 46, + 459, + 286, + 578 + ], + "type": "text", + "content": "Efforts to extend 3D reconstruction to multispectral and hyperspectral domains include works from Zia et al. [31] and Liang J et al. [17], which employ point clouds and key point descriptors for structure-from-motion. These methods enhance 3D reconstruction by improving point matching but often result in sparse, noisy point clouds with inadequate occupancy information for multi-view stereo. Shadows and lighting variations degrade geometric quality. A faster approach by Zia et al. [32] instead skips rasterization by mapping hyperspectral images onto preprocessed 3D meshes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 586, + 101, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 101, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 101, + 597 + ], + "type": "text", + "content": "2.2. NeRFs" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "content": "NeRFs have gained significant attention [8] since Mildenhall et al. [23] introduced the technique. NeRFs rely on classical volume rendering techniques for scene synthesis. Recent advances have sought to adapt NeRFs for spectral and hyperspectral imaging. SpectralNeRF [16] incorporates a spectrum attention mechanism (SAUNet) for realistic low-channel multi-spectral rendering under variable lighting. However, the reliance on attention mechanisms introduces computational overhead, making long-range spectra diffi" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 239 + ], + "type": "text", + "content": "cult to capture. Spec-NeRF [15] takes a more efficient approach by adapting TensoRF [4] for reconstruction, with a seperate MLP estimating the camera's spectral sensitivity, focusing on computationally efficient spectral scene reconstructions. HS-NeRF [5] targets hyperspectral novel view synthesis by learning sinusoidal position encoding for each channel, allowing it to interpolate between viewpoints. This approach struggles with full-spectrum fidelity due to its lack of end-to-end training. HS-NeRF also released two multiview hyperspectral datasets, using two different hyperspectral cameras with varying signal-to-noise ratios. These, NeRF-based methods remain limited by their large parameter counts and instability when handling high-dimensional hyperspectral data alongside slower inference times." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 273, + 356, + 284 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 273, + 356, + 284 + ], + "spans": [ + { + "bbox": [ + 306, + 273, + 356, + 284 + ], + "type": "text", + "content": "2.3. 3DGS" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 294, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 294, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 294, + 545, + 713 + ], + "type": "text", + "content": "Advancements in 3DGS improve point cloud efficiency for hyperspectral imaging but rely on discrete thresholds, causing uneven textures and memory issues with high-dimensional HSI data. VDGS [20] employs a hybrid NeRF neural model for color and opacity, improving spectral reconstruction for HSI. However, this model still heavily depends on the 3DGS prediction and only uses opacity estimation from the MLP. Scaffold-GS [19] introduces compressed representations to smooth surfaces but, like 3DGS, it struggles with artifacts in sparse regions. Mip-Splatting [30] addresses rendering quality by applying 3D smoothing and a 2D Mip filter, yet it remains constrained by viewfrustum-based Gaussian selection, limiting its effectiveness in dynamic scenes. Challenges in Gaussian initialization, especially due to poor SfM data, are tackled by RAINGS [12], which incorporates adaptive optimization. InstantSplat [10] improves stability by using simpler point provisioning alongside the use of dust3r [28] which offers better SFM than COLMAP but has difficulty in processing large amounts of images. Meanwhile, Bulo [3] proposes pixel-level error-based densification to ensure consistent quality, and GaussianPro [6] extends pixel-wise loss methods to better align 3D Gaussian normals during rendering, albeit at the cost of longer training times. EfficientGS[18] and LightGaussian[9] reduce model size and computational load through pruning methods but risk losing important visual details, especially when over-decimation occurs without proper ranking thresholds[18]. These methods often prioritize storage efficiency over rendering performance, overlooking challenges in real-time optimization. Although no previous works have explored the adaptation of Gaussian Splatting to Hyperspectral data, our experiments show that a naive adaptation struggles to model fine details. HyperGS solves these issues by operating in a lower-dimensional learned latent space." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "5971" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 133, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 133, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 133, + 83 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 287, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 287, + 233 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 287, + 233 + ], + "type": "text", + "content": "3DGS, developed by Kerbl et al. [13], represents 3D environments using " + }, + { + "bbox": [ + 46, + 91, + 287, + 233 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 91, + 287, + 233 + ], + "type": "text", + "content": " Gaussian primitives. Each Gaussian is parameterized by its position " + }, + { + "bbox": [ + 46, + 91, + 287, + 233 + ], + "type": "inline_equation", + "content": "x_{i} \\in \\mathbb{R}^{3}" + }, + { + "bbox": [ + 46, + 91, + 287, + 233 + ], + "type": "text", + "content": " in world coordinates, a scaling vector " + }, + { + "bbox": [ + 46, + 91, + 287, + 233 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_i \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 91, + 287, + 233 + ], + "type": "text", + "content": " defining the size of the Gaussian along each axis, a rotation quaternion " + }, + { + "bbox": [ + 46, + 91, + 287, + 233 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_i \\in \\mathbb{R}^4" + }, + { + "bbox": [ + 46, + 91, + 287, + 233 + ], + "type": "text", + "content": " defining its orientation, opacity " + }, + { + "bbox": [ + 46, + 91, + 287, + 233 + ], + "type": "inline_equation", + "content": "\\sigma_{i} \\in \\mathbb{R}" + }, + { + "bbox": [ + 46, + 91, + 287, + 233 + ], + "type": "text", + "content": ", which is controlled by a sigmoid function, and appearance features " + }, + { + "bbox": [ + 46, + 91, + 287, + 233 + ], + "type": "inline_equation", + "content": "f_{i} \\in \\mathbb{R}^{k}" + }, + { + "bbox": [ + 46, + 91, + 287, + 233 + ], + "type": "text", + "content": " to represent view-dependent RGB signals. The core of 3DGS lies in transforming these Gaussians from 3D world space into the 2D image plane for rendering. To do this, we first define the 3D covariance matrix of each Gaussian in world space as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 134, + 233, + 287, + 247 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 233, + 287, + 247 + ], + "spans": [ + { + "bbox": [ + 134, + 233, + 287, + 247 + ], + "type": "interline_equation", + "content": "\\Sigma_ {i} = R _ {i} S _ {i} ^ {2} R _ {i} ^ {T}, \\tag {1}", + "image_path": "1a66ec3a8716d118d86eead5c7c77206acbcac32daf584162742b8c650dafab7.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 251, + 287, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 251, + 287, + 275 + ], + "spans": [ + { + "bbox": [ + 47, + 251, + 287, + 275 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 251, + 287, + 275 + ], + "type": "inline_equation", + "content": "R_{i}" + }, + { + "bbox": [ + 47, + 251, + 287, + 275 + ], + "type": "text", + "content": " is the rotation matrix derived from the quaternion " + }, + { + "bbox": [ + 47, + 251, + 287, + 275 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_i" + }, + { + "bbox": [ + 47, + 251, + 287, + 275 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 251, + 287, + 275 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 47, + 251, + 287, + 275 + ], + "type": "text", + "content": " is the scaling matrix derived from the vector " + }, + { + "bbox": [ + 47, + 251, + 287, + 275 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_i" + }, + { + "bbox": [ + 47, + 251, + 287, + 275 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 275, + 287, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 275, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 47, + 275, + 287, + 323 + ], + "type": "text", + "content": "To sput the Gaussian onto the 2D image plane, we transform the covariance matrix using the viewing transformation matrix " + }, + { + "bbox": [ + 47, + 275, + 287, + 323 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 47, + 275, + 287, + 323 + ], + "type": "text", + "content": " and the Jacobian matrix " + }, + { + "bbox": [ + 47, + 275, + 287, + 323 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 47, + 275, + 287, + 323 + ], + "type": "text", + "content": ", which handles the projection onto the camera's image plane as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 123, + 328, + 287, + 342 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 328, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 123, + 328, + 287, + 342 + ], + "type": "interline_equation", + "content": "\\hat {\\Sigma} _ {i} = J W \\Sigma_ {i} W ^ {T} J ^ {T}. \\tag {2}", + "image_path": "3c1e98a025371f7cdb644a02f028294ce6f9026519ec76434a61d008bea23ce5.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 350, + 287, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 350, + 287, + 386 + ], + "spans": [ + { + "bbox": [ + 47, + 350, + 287, + 386 + ], + "type": "text", + "content": "accounting for the perspective distortion and view transformation. When rendering to a particular pixel, the opacity " + }, + { + "bbox": [ + 47, + 350, + 287, + 386 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 47, + 350, + 287, + 386 + ], + "type": "text", + "content": " of Gaussian " + }, + { + "bbox": [ + 47, + 350, + 287, + 386 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 350, + 287, + 386 + ], + "type": "text", + "content": " is computed as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 111, + 392, + 287, + 406 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 392, + 287, + 406 + ], + "spans": [ + { + "bbox": [ + 111, + 392, + 287, + 406 + ], + "type": "interline_equation", + "content": "\\alpha_ {i} = 1 - \\exp \\left(- \\delta_ {i} ^ {T} \\hat {\\Sigma} _ {i} ^ {- 1} \\delta_ {i}\\right), \\tag {3}", + "image_path": "78350c7c87cdd3c1c444695df682567d03d86da8fb215315defaeb722f772bf3.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 412, + 287, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 412, + 287, + 435 + ], + "spans": [ + { + "bbox": [ + 47, + 412, + 287, + 435 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 412, + 287, + 435 + ], + "type": "inline_equation", + "content": "\\delta_{i}" + }, + { + "bbox": [ + 47, + 412, + 287, + 435 + ], + "type": "text", + "content": " is the distance between the pixel and the projected center of the Gaussian in 2D." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 436, + 287, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 436, + 287, + 472 + ], + "spans": [ + { + "bbox": [ + 47, + 436, + 287, + 472 + ], + "type": "text", + "content": "The transmittance " + }, + { + "bbox": [ + 47, + 436, + 287, + 472 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 47, + 436, + 287, + 472 + ], + "type": "text", + "content": ", which represents the cumulative transparency along the ray up to the " + }, + { + "bbox": [ + 47, + 436, + 287, + 472 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 436, + 287, + 472 + ], + "type": "text", + "content": "-th Gaussian. It is computed as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 123, + 472, + 287, + 505 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 472, + 287, + 505 + ], + "spans": [ + { + "bbox": [ + 123, + 472, + 287, + 505 + ], + "type": "interline_equation", + "content": "T _ {i} = \\prod_ {j = 1} ^ {i - 1} \\left(1 - \\sigma_ {i} \\alpha_ {j}\\right). \\tag {4}", + "image_path": "48d6da9be5b3d9b948dc3236c26084082a15b30389a3dce086391504fdfb29f8.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 508, + 287, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 508, + 287, + 532 + ], + "spans": [ + { + "bbox": [ + 47, + 508, + 287, + 532 + ], + "type": "text", + "content": "The final color " + }, + { + "bbox": [ + 47, + 508, + 287, + 532 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 47, + 508, + 287, + 532 + ], + "type": "text", + "content": " of each pixel, " + }, + { + "bbox": [ + 47, + 508, + 287, + 532 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 47, + 508, + 287, + 532 + ], + "type": "text", + "content": " is obtained by blending the colors of all Gaussian's projected onto that pixel:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 125, + 538, + 287, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 538, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 125, + 538, + 287, + 563 + ], + "type": "interline_equation", + "content": "C (p) = \\sum_ {i \\in N} T _ {i} \\alpha_ {i} c _ {i}, \\tag {5}", + "image_path": "2d6ec00452129b0e47e6ffa1332073b1b055213ba40cf37dac496e25c03a4f55.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 563, + 287, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 287, + 635 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 563, + 287, + 635 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 47, + 563, + 287, + 635 + ], + "type": "text", + "content": " is the coordinate position in the image, " + }, + { + "bbox": [ + 47, + 563, + 287, + 635 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 47, + 563, + 287, + 635 + ], + "type": "text", + "content": " is the color of the " + }, + { + "bbox": [ + 47, + 563, + 287, + 635 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 563, + 287, + 635 + ], + "type": "text", + "content": "-th Gaussian, derived from its color appearance coefficients " + }, + { + "bbox": [ + 47, + 563, + 287, + 635 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 47, + 563, + 287, + 635 + ], + "type": "text", + "content": ". In summary, this process transforms and projects the Gaussian primitives from 3D space into 2D image space and computes the final pixel colors through opacity and transmittance blending." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 645, + 129, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 645, + 129, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 645, + 129, + 659 + ], + "type": "text", + "content": "4. Methodology" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "text", + "content": "HyperGS aims to provide a lightweight, fast rasterization solution to HNVS that is robust to different hyperspectral cameras' sensitivity functions with outstanding accuracy. Our system diagram is seen in Figure 2." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 72, + 460, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 460, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 460, + 85 + ], + "type": "text", + "content": "4.1. Hyperspectral Compression" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 90, + 545, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 90, + 545, + 257 + ], + "spans": [ + { + "bbox": [ + 304, + 90, + 545, + 257 + ], + "type": "text", + "content": "To address the challenges of high-dimensional optimization, we adopt a novel latent space exploration method for the hyperspectral data. The latent space of a pre-trained autoencoder (AE) serves as an exploratory space for our 3DGS system. This approach reduces the computational load during 3DGS optimization by providing a compact, lower-dimensional target. Additionally, the latent space bounds the error by encapsulating the spectral sensitivity of the hyperspectral camera for each channel, enhancing the accuracy and reliability of novel view reconstruction in the spectral domain. During testing, the latent space 3DGS viewpoint estimations are decoded using the frozen AE to produce novel HSIs from which gradient flow will be calculated." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 258, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 258, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 258, + 545, + 449 + ], + "type": "text", + "content": "The compression network is a fast convolutional AE. Both sides are symmetrical and is built using a series of 1D convolutional layers and Squeeze-Excitation (SE) blocks that operate across the spectral dimension of the images. The SE blocks benefit the performance by weighting features between layers for improved spectral discrimination. The model is trained on the pixel level of the scenes dataset. The encoder compresses the high-dimensional hyperspectral data into a latent representation via max-pooling layers, while the decoder reconstructs the original spectral information from this compressed form via upsampling. The architecture omits skip connections to ensure the decoder can function independently during training and testing to decode the LHSI data produced by the 3DGS system. For a detailed layout of the network architecture, please refer to Fig. 3." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 449, + 494, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 449, + 494, + 461 + ], + "spans": [ + { + "bbox": [ + 317, + 449, + 494, + 461 + ], + "type": "text", + "content": "The AE minimizes the loss " + }, + { + "bbox": [ + 317, + 449, + 494, + 461 + ], + "type": "inline_equation", + "content": "L_{ae}" + }, + { + "bbox": [ + 317, + 449, + 494, + 461 + ], + "type": "text", + "content": ", defined as:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 336, + 472, + 545, + 485 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 472, + 545, + 485 + ], + "spans": [ + { + "bbox": [ + 336, + 472, + 545, + 485 + ], + "type": "interline_equation", + "content": "L _ {a e} = L _ {H u b e r} \\left(C ^ {*} (p), D e c \\left(E n c \\left(C ^ {*} (p)\\right)\\right)\\right), \\tag {6}", + "image_path": "c153277070f263468978375c6305ecc78bdb13eeb1e9e730ebd793251551db6c.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 495, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 545, + 567 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 495, + 545, + 567 + ], + "type": "inline_equation", + "content": "C^*" + }, + { + "bbox": [ + 304, + 495, + 545, + 567 + ], + "type": "text", + "content": " is the ground truth pixel spectrum. The Huber loss provides a smooth optimization process throughout training, as it handles outliers effectively while maintaining smooth gradients. Since each hyperspectral dataset includes varying levels of signal-to-noise ratios, this approach helps improve robustness against noise-related issues." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 574, + 537, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 574, + 537, + 588 + ], + "spans": [ + { + "bbox": [ + 305, + 574, + 537, + 588 + ], + "type": "text", + "content": "4.2. Latent Hyperspectral 3D Gaussian Splitting" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "Each Gaussian in the splating process is assigned a latent spectral signature " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i\\in \\mathbb{R}^m" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " , where " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " represents the latent space dimensionality. Inspired by many previous RGB techniques [14, 19, 21, 29] an MLP conditioned on view-directional hash encoding is used to predict anisotropic opacity and spectral color information for each Gaussian. This allows view-dependent spectral effects to be modelled upon specific bands. This signature is mapped directly to the pixel values using the decoder model defined in Section 4.1. For camera " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " , the MLP " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " predicts view-dependent" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5972" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 70, + 545, + 295 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 545, + 295 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 545, + 295 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 545, + 295 + ], + "type": "image", + "image_path": "51e4291d2d0e351e9e8fb321cc3a7f7e7bf2b36519dc56638cd1ea0b3f291b51.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 304, + 546, + 350 + ], + "lines": [ + { + "bbox": [ + 46, + 304, + 546, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 546, + 350 + ], + "type": "text", + "content": "Figure 2. Visual system diagram of our approach. Blue lines indicate the operational flow, while orange lines represent the gradient flow. Our novel modified latent hyperspectral adaptive density function operates within a latent space provided by a frozen autoencoder, which is also responsible for decoding the final images. Latent space novel views are generated through a combination of the latent hyperspectral Gaussian point cloud and a NeRF-style MLP. Gradients are updated based on the decoded images." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 361, + 287, + 456 + ], + "blocks": [ + { + "bbox": [ + 50, + 361, + 287, + 456 + ], + "lines": [ + { + "bbox": [ + 50, + 361, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 50, + 361, + 287, + 456 + ], + "type": "image", + "image_path": "b7612df0e08b2b813be9f51cc446f7e0fca18a6e8d6626c156db84129bf8ede0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 461, + 288, + 506 + ], + "lines": [ + { + "bbox": [ + 46, + 461, + 288, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 461, + 288, + 506 + ], + "type": "text", + "content": "Figure 3. Our channel-wise convolutional AE model learns LHSI space representation of the scene. The decoder is only used in training and inference for the 3DGS system after the preprocessing of the dataset is finished." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 521, + 287, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 558 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 558 + ], + "type": "text", + "content": "spectral effects " + }, + { + "bbox": [ + 46, + 521, + 287, + 558 + ], + "type": "inline_equation", + "content": "\\tilde{f}_{i,d}" + }, + { + "bbox": [ + 46, + 521, + 287, + 558 + ], + "type": "text", + "content": " and opacity " + }, + { + "bbox": [ + 46, + 521, + 287, + 558 + ], + "type": "inline_equation", + "content": "\\tilde{\\sigma}_{i,d}" + }, + { + "bbox": [ + 46, + 521, + 287, + 558 + ], + "type": "text", + "content": " effects for each Gaussian. Specifically, the MLP takes the centre " + }, + { + "bbox": [ + 46, + 521, + 287, + 558 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_i" + }, + { + "bbox": [ + 46, + 521, + 287, + 558 + ], + "type": "text", + "content": " of each Gaussian and the view direction as inputs:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 102, + 565, + 287, + 580 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 565, + 287, + 580 + ], + "spans": [ + { + "bbox": [ + 102, + 565, + 287, + 580 + ], + "type": "interline_equation", + "content": "[ \\tilde {f} _ {i, d}, \\tilde {\\sigma} _ {i, d} ] = \\mathcal {F} _ {v} (h (\\mathbf {m} _ {i}, \\mathbf {d}); \\Theta), \\tag {7}", + "image_path": "d4aed853a3ee60b1dfe1650f0113773ed016766fe6806266a9cbb4c901e5dae2.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 589, + 287, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 589, + 287, + 625 + ], + "spans": [ + { + "bbox": [ + 46, + 589, + 287, + 625 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 589, + 287, + 625 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 46, + 589, + 287, + 625 + ], + "type": "text", + "content": " is the hash encoding of the inputs while " + }, + { + "bbox": [ + 46, + 589, + 287, + 625 + ], + "type": "inline_equation", + "content": "\\tilde{\\Theta}" + }, + { + "bbox": [ + 46, + 589, + 287, + 625 + ], + "type": "text", + "content": " denotes the MLP parameters similar to that of I-NGP [24] and Mip-Nerf360 [1]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 625, + 288, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 625, + 288, + 673 + ], + "spans": [ + { + "bbox": [ + 46, + 625, + 288, + 673 + ], + "type": "text", + "content": "These view-dependent spectral and opacity effects are multiplied with those stored in the Gaussian Cloud, leading to the updated volumetric rendering equations with transmittance being defined as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 111, + 681, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 681, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 111, + 681, + 287, + 715 + ], + "type": "interline_equation", + "content": "T _ {i, d} = \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j} \\sigma_ {i} \\tilde {\\sigma} _ {i, d}). \\tag {8}", + "image_path": "24a8dba7c4372ff9e6431cb618f6bbb321802bee9a6937973fed8d7e78675f6f.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 363, + 545, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 363, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 304, + 363, + 545, + 399 + ], + "type": "text", + "content": "The final latent signature " + }, + { + "bbox": [ + 304, + 363, + 545, + 399 + ], + "type": "inline_equation", + "content": "\\hat{C}" + }, + { + "bbox": [ + 304, + 363, + 545, + 399 + ], + "type": "text", + "content": " of each pixel is obtained by blending the colors of all Gaussians projected onto that pixel " + }, + { + "bbox": [ + 304, + 363, + 545, + 399 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 363, + 545, + 399 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 367, + 399, + 545, + 424 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 399, + 545, + 424 + ], + "spans": [ + { + "bbox": [ + 367, + 399, + 545, + 424 + ], + "type": "interline_equation", + "content": "\\hat {C} (p, d) = \\sum_ {i \\in N} T _ {i, d} \\alpha_ {i} f _ {i} \\tilde {f} _ {i, d}. \\tag {9}", + "image_path": "cec8318a327b463abecf242cef8b2365feacc9deb974cf627ee1354f2d1a0f28.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 429, + 545, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 429, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 304, + 429, + 545, + 477 + ], + "type": "text", + "content": "It is worth emphasizing that these view-dependent spectral effects are applied within the learned latent space before decoding. This helps maintain a low computational complexity while minimizing outliers." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 477, + 545, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 477, + 545, + 512 + ], + "spans": [ + { + "bbox": [ + 304, + 477, + 545, + 512 + ], + "type": "text", + "content": "A decoding operation is then performed to provide the full fidelity prediction from the system using the decoder from Section 4.1." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 372, + 522, + 545, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 522, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 372, + 522, + 545, + 536 + ], + "type": "interline_equation", + "content": "C (p, d) = \\operatorname {D e c} (\\hat {C} (p, d)). \\tag {10}", + "image_path": "1abd1a8f7cfd9a9cb54ee50ed8589506e7a809ce0a65e2038bc1a8ba268ea3af.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 545, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 546, + 714 + ], + "type": "text", + "content": "The final rendering ensures that the spectral integrity is maintained throughout the entire process, which is critical for accurately reconstructing material properties. By leveraging the latent space, HyperGS provides a more meaningful and efficient spectral representation of the scene, than the original full hyperspectral image. The original loss proposed in 3DGS for spatial and geometric consistency in rendered images uses a weighted DSSIM and L1 loss. However, this type of loss have been shown to lead to underfitting for HSI in other fields. This is because an L1 loss can produce extreme errors for sensitive spectral bands, destabilizing training. To address this, we employ a weighted loss combining Charbonnier loss and cosine similarity to account for both spectral quality and spatial consistency." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5973" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 286, + 167 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 286, + 167 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 286, + 167 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 286, + 167 + ], + "type": "image", + "image_path": "c81695b20ce3f44a3a7c92a7eef62682d17dd3def000b421008ce249f9392f8b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 176, + 288, + 243 + ], + "lines": [ + { + "bbox": [ + 46, + 176, + 288, + 243 + ], + "spans": [ + { + "bbox": [ + 46, + 176, + 288, + 243 + ], + "type": "text", + "content": "Figure 4. Visualization of our re-projection protocol for initializing the SFM point cloud. We estimate the SFM from grayscale channel slices of the hyperspectral image scene with COLMAP. Then, using the average of all views, we re-project each point into LHSI from provided by our AE, providing an optimal initialization of latent spectral color for the 3DGS system." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 251, + 287, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 287, + 324 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 287, + 324 + ], + "type": "text", + "content": "The cosine similarity provides a measured angular distance between vectors, making it ideal for comparing spectra. To provide an overall consistency in the spatial and spectral domain we keep the SSIM score in the training from the original 3DGS system. The per pixel HyperGS training loss " + }, + { + "bbox": [ + 46, + 251, + 287, + 324 + ], + "type": "inline_equation", + "content": "L_{d}(p)" + }, + { + "bbox": [ + 46, + 251, + 287, + 324 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 52, + 334, + 287, + 347 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 334, + 287, + 347 + ], + "spans": [ + { + "bbox": [ + 52, + 334, + 287, + 347 + ], + "type": "interline_equation", + "content": "L _ {d} (p) = (1 - \\lambda) (\\beta L _ {C B} (p) + L _ {C S} (p)) + \\lambda L _ {S S I M} (p), \\tag {11}", + "image_path": "e3f52f64f9463bf723fa797ac2fb85d416b041e9332d89267ac1aa61cf343e75.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 356, + 287, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 356, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 47, + 356, + 287, + 392 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 356, + 287, + 392 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 47, + 356, + 287, + 392 + ], + "type": "text", + "content": " weighs the spectral loss balancing the cosine similarity and charbonnier loss. Thus the total training loss across all pixels and views is:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 109, + 401, + 287, + 435 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 401, + 287, + 435 + ], + "spans": [ + { + "bbox": [ + 109, + 401, + 287, + 435 + ], + "type": "interline_equation", + "content": "L (C, C ^ {*}) = \\sum_ {p = 1} ^ {P} \\sum_ {d \\in V} L _ {d} (p), \\tag {12}", + "image_path": "c4b634e99affc1d18ef6f0ef64309d335eb39518491d2d9e4c25371c7938f573.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 444, + 287, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 444, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 47, + 444, + 287, + 480 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 444, + 287, + 480 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 47, + 444, + 287, + 480 + ], + "type": "text", + "content": " is the 3DGS systems network prediction after decoding the entire latent image, " + }, + { + "bbox": [ + 47, + 444, + 287, + 480 + ], + "type": "inline_equation", + "content": "C^{*}" + }, + { + "bbox": [ + 47, + 444, + 287, + 480 + ], + "type": "text", + "content": " is the ground truth hyperspectral image, and " + }, + { + "bbox": [ + 47, + 444, + 287, + 480 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 47, + 444, + 287, + 480 + ], + "type": "text", + "content": " is the set of training view directions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 495, + 135, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 495, + 135, + 506 + ], + "spans": [ + { + "bbox": [ + 47, + 495, + 135, + 506 + ], + "type": "text", + "content": "4.2.1 Initialisation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 514, + 287, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 514, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 46, + 514, + 287, + 658 + ], + "type": "text", + "content": "Since there is no dedicated hyperspectral Structure-from-Motion (SfM) algorithm compatible with COLMAP, we first convert the hyperspectral images into grayscale images, " + }, + { + "bbox": [ + 46, + 514, + 287, + 658 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{gray}}" + }, + { + "bbox": [ + 46, + 514, + 287, + 658 + ], + "type": "text", + "content": " by selecting the spectral channel with the highest foreground intensity variance to preserve key features for the SfM process. Our process is visualized in Figure 4. Using these grayscale images and their corresponding camera projection matrices for each view " + }, + { + "bbox": [ + 46, + 514, + 287, + 658 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 514, + 287, + 658 + ], + "type": "text", + "content": ", we apply COLMAP to generate a sparse 3D point cloud. Each 3D point " + }, + { + "bbox": [ + 46, + 514, + 287, + 658 + ], + "type": "inline_equation", + "content": "\\mathbf{X} = (X,Y,Z,1)^T" + }, + { + "bbox": [ + 46, + 514, + 287, + 658 + ], + "type": "text", + "content": " in world coordinates is related to its pixel coordinate " + }, + { + "bbox": [ + 46, + 514, + 287, + 658 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_d = (x_d,y_d,1)^T" + }, + { + "bbox": [ + 46, + 514, + 287, + 658 + ], + "type": "text", + "content": " in the image plane via the camera projection " + }, + { + "bbox": [ + 46, + 514, + 287, + 658 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_d\\mathbf{E}_d" + }, + { + "bbox": [ + 46, + 514, + 287, + 658 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 134, + 667, + 287, + 680 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 667, + 287, + 680 + ], + "spans": [ + { + "bbox": [ + 134, + 667, + 287, + 680 + ], + "type": "interline_equation", + "content": "\\mathbf {p} _ {d} = \\mathbf {K} _ {d} \\mathbf {E} _ {d} \\mathbf {X}, \\tag {13}", + "image_path": "d7fee131e6596b3e37c9e16db974d800904ad98aa0d7dbac30b5ece8cd027ad6.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_d" + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": " are the camera intrinsics and " + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_d" + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": " the extrinsics. After generating 3D points and recovering camera poses," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "we re-project the points into the LHSIs. Each 3D point " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " is linked to pixel coordinates " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_d" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " across all views. We then initialize the Gaussian cloud with one Gaussian centered on each 3D point " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": ", with spectral signature " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " computed by averaging the latent hyperspectral signatures " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "H_{d}(y_{d},x_{d})" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " across all views " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 379, + 152, + 545, + 184 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 379, + 152, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 379, + 152, + 545, + 184 + ], + "type": "interline_equation", + "content": "f _ {i} = \\frac {1}{| V |} \\sum_ {v = 1} ^ {V} \\hat {C} _ {d} ^ {*} (\\mathbf {p} _ {d}), \\tag {14}", + "image_path": "64ae4b5ea286973ed7b61c6b267ebf738533eabbe7dfb74eb556efb8da8c0130.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 195, + 545, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 195, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 304, + 195, + 545, + 232 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 195, + 545, + 232 + ], + "type": "inline_equation", + "content": "\\hat{C}_d^*(\\mathbf{p}_d)" + }, + { + "bbox": [ + 304, + 195, + 545, + 232 + ], + "type": "text", + "content": " is the latent hyperspectral signature at pixel " + }, + { + "bbox": [ + 304, + 195, + 545, + 232 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 195, + 545, + 232 + ], + "type": "text", + "content": " for view " + }, + { + "bbox": [ + 304, + 195, + 545, + 232 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 195, + 545, + 232 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 195, + 545, + 232 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 304, + 195, + 545, + 232 + ], + "type": "text", + "content": " is the set of training views. This averaging ensures the spectral information is captured robustly." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 247, + 489, + 259 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 247, + 489, + 259 + ], + "spans": [ + { + "bbox": [ + 305, + 247, + 489, + 259 + ], + "type": "text", + "content": "4.2.2 Latent Hyperspectral Densification" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 266, + 545, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 266, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 266, + 545, + 483 + ], + "type": "text", + "content": "As the number of color channels in HSI data increases, visual discontinuities become more prevalent. Additionally, our initial SFM reconstruction is less dense compared to standard RGB 3DGS. To address these issues, we incorporate an advanced densification method that enhances stability in both results and training. A key component of 3DGS involves determining whether a Gaussian should be split or cloned based on the gradient magnitude of the Normalized Device Coordinates (NDC) across various viewpoints. However, in sparse regions, this can cause artefacts as large Gaussians are highly visible in many viewpoints, leading to inconsistent splitting. This challenge is exacerbated in the hyperspectral domain of 3DGS, where the larger number of channels and highly variable value ranges make it harder to set effective thresholds. To mitigate this, we scale the gradient by the square of the depth relative to the scene's radius, which reduces the influence of Gaussians near the camera. More formally, we define the splitting score as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 490, + 545, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 490, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 317, + 490, + 545, + 536 + ], + "type": "interline_equation", + "content": "\\mathcal {S} \\left(g _ {i}\\right) = \\sum_ {d \\in V} \\sum_ {p = 1} ^ {P} h (d, i) \\sqrt {\\left(\\frac {\\partial L _ {d} (p)}{\\partial x _ {i}}\\right) ^ {2} + \\left(\\frac {\\partial L _ {d} (p)}{\\partial y _ {i}}\\right) ^ {2}}, \\tag {15}", + "image_path": "0b43eab3b2c19ce8838883665fdb88aeba1fc9ec00c54e1b7219c8d933377faf.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 536, + 453, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 536, + 453, + 548 + ], + "spans": [ + { + "bbox": [ + 306, + 536, + 453, + 548 + ], + "type": "text", + "content": "where the depth-scaling function is:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 372, + 556, + 545, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 556, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 372, + 556, + 545, + 586 + ], + "type": "interline_equation", + "content": "h (d, i) = \\left(\\frac {\\left| \\mathbf {E} _ {d} \\mathbf {X} _ {i} \\right|}{\\beta_ {\\text {f i e l d}} \\times R}\\right) ^ {2}, \\tag {16}", + "image_path": "9609958af67f5c10a68e32597f04dc2d21a1da5cca45404d03a82c6b801cddba.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 594, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 594, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 305, + 594, + 545, + 617 + ], + "type": "inline_equation", + "content": "\\beta_{\\mathrm{field}}" + }, + { + "bbox": [ + 305, + 594, + 545, + 617 + ], + "type": "text", + "content": " is a tunable parameter, and " + }, + { + "bbox": [ + 305, + 594, + 545, + 617 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 305, + 594, + 545, + 617 + ], + "type": "text", + "content": " is the scene's radius, based on the largest distance between any pair of cameras." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 618, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 545, + 714 + ], + "type": "text", + "content": "The Gaussian " + }, + { + "bbox": [ + 304, + 618, + 545, + 714 + ], + "type": "inline_equation", + "content": "g_{i}" + }, + { + "bbox": [ + 304, + 618, + 545, + 714 + ], + "type": "text", + "content": " is split or cloned if " + }, + { + "bbox": [ + 304, + 618, + 545, + 714 + ], + "type": "inline_equation", + "content": "S(g_{i}) > \\theta_{q}" + }, + { + "bbox": [ + 304, + 618, + 545, + 714 + ], + "type": "text", + "content": ". This approach accounts for depth-based scaling and pixel contributions, leading to better hyperspectral scene reconstruction with fewer artifacts near the camera. However, this more expressive densification can result in an excessive number of Gaussians in scenes with high-frequency details (e.g., the \"pinecone\" scene in Figure 1, to stabilize this we deploy a global pruning strategy." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5974" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 281, + 166 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 281, + 166 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 281, + 166 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 281, + 166 + ], + "type": "image", + "image_path": "73dbb521fad41f339471427d908def227f2f85e65b512739e441adbefdfd0972.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 176, + 287, + 210 + ], + "lines": [ + { + "bbox": [ + 46, + 176, + 287, + 210 + ], + "spans": [ + { + "bbox": [ + 46, + 176, + 287, + 210 + ], + "type": "text", + "content": "Figure 5. Visualization of our pixel-wise pruning. Gaussians with poor similarity scores below a threshold are pruned, such as the red-circled Gaussian and the black ground truth spectrum." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 223, + 258, + 237 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 223, + 258, + 237 + ], + "spans": [ + { + "bbox": [ + 47, + 223, + 258, + 237 + ], + "type": "text", + "content": "4.3. Global Hyperspectral Gaussian Pruning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 241, + 287, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 241, + 287, + 420 + ], + "spans": [ + { + "bbox": [ + 46, + 241, + 287, + 420 + ], + "type": "text", + "content": "In HyperGS, we deploy pixel-wise pruning [25] to preserve the spectral integrity of latent hyperspectral images (LHSIs) while lowering the overall number of primitives to improve the scene quality, model size, and spectral gradient descent. Unlike cross-view pruning methods, which evaluate Gaussians across multiple viewpoints and leads to over-pruning [11, 18], our approach assesses each Gaussian's contribution at the pixel level. This approach retains fine spectral details specific to each pixel. For each Gaussian " + }, + { + "bbox": [ + 46, + 241, + 287, + 420 + ], + "type": "inline_equation", + "content": "g_{i}" + }, + { + "bbox": [ + 46, + 241, + 287, + 420 + ], + "type": "text", + "content": ", pruning is based on its pixel-wise importance " + }, + { + "bbox": [ + 46, + 241, + 287, + 420 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 46, + 241, + 287, + 420 + ], + "type": "text", + "content": " across all viewpoints and pixels. We compute the importance for every combination of Gaussian, pixel and viewpoint as the spectral difference between the ground truth hyperspectral value " + }, + { + "bbox": [ + 46, + 241, + 287, + 420 + ], + "type": "inline_equation", + "content": "\\hat{C}_d^* (p)" + }, + { + "bbox": [ + 46, + 241, + 287, + 420 + ], + "type": "text", + "content": " and the decoded Gaussians LHSI, with the pruning score defined as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 430, + 287, + 443 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 430, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 67, + 430, + 287, + 443 + ], + "type": "interline_equation", + "content": "\\mathcal {I} [ g _ {i}, p, d ] = \\left(1 - \\left| C _ {d} ^ {*} (p) - D e c \\left(f _ {i}\\right) \\right|\\right) \\alpha_ {i} T _ {i}, \\tag {17}", + "image_path": "20c8deed0d9d28b31bdf5ca540b893fdc950e2fcc64b0c2da10e63ed6b33e355.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 450, + 287, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 287, + 510 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 450, + 287, + 510 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 46, + 450, + 287, + 510 + ], + "type": "text", + "content": " is the accumulated transmittance of Gaussian " + }, + { + "bbox": [ + 46, + 450, + 287, + 510 + ], + "type": "inline_equation", + "content": "g_{i}" + }, + { + "bbox": [ + 46, + 450, + 287, + 510 + ], + "type": "text", + "content": ". Including " + }, + { + "bbox": [ + 46, + 450, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 46, + 450, + 287, + 510 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 450, + 287, + 510 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 46, + 450, + 287, + 510 + ], + "type": "text", + "content": " ensures that only Gaussians contributing significantly in visibility and spectral accuracy are retained, and also ensures that the score is 0 for Gaussians that do not overlap the given pixel." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 511, + 287, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 511, + 287, + 570 + ], + "spans": [ + { + "bbox": [ + 46, + 511, + 287, + 570 + ], + "type": "text", + "content": "Following recent works, we do not prune Gaussians based on the average, or the total pixel-wise importance score. Instead, we retain all Gaussians within the top-K importance ranking for any pixel. More formally the filtered Gaussian cloud " + }, + { + "bbox": [ + 46, + 511, + 287, + 570 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 46, + 511, + 287, + 570 + ], + "type": "text", + "content": " is" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 70, + 578, + 287, + 592 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 578, + 287, + 592 + ], + "spans": [ + { + "bbox": [ + 70, + 578, + 287, + 592 + ], + "type": "interline_equation", + "content": "\\mathcal {G} = \\left\\{g _ {i} \\mid \\exists (p, d), \\operatorname {R a n k} \\left(g _ {i}, \\mathcal {I} [:, p, d ] < \\tau_ {p}\\right) \\right\\}, \\tag {18}", + "image_path": "6d6a2d2652d1a1b815e63066fcbfec6bfe03056319dd863e04d126b658a9a587.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 599, + 287, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 599, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 599, + 287, + 635 + ], + "type": "text", + "content": "where \"Rank\" is a function that returns the rank of a given Gaussian within the slice of the score tensor for that pixel and view." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 646, + 128, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 646, + 128, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 646, + 128, + 659 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "We first evaluate HyperGS using the HS-NeRF datasets [5]. These datasets differ in signal-to-noise ratios, channel lengths, and the number of images per scene, providing a comprehensive test of model performance. The Bayspec" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 239 + ], + "type": "text", + "content": "dataset [5] contains around 360 images per scene (3 scenes total), while the SOP dataset [5] has around 40 images per scene (4 scenes total). Due to the lack of continuous HSI datasets for non-object-centric scenes, we also evaluate HyperGS on a synthetic dataset curated from ScanNetv200 [7] (Section 9 in the supplementary materials, where we replace each semantic label with distinct hyperspectral signatures as seen in our supplementary materials). The baselines we have implemented for the HNVS problem include conventional 3DGS (with our reprojected SFM initialisation 4.2.1 and removed SH coefficients), traditional NeRF models adapted for hyperspectral data (NeRF [22], MipNeRF [2] MipNeRF-360 [1], Nerfacto [26] and TensorF [4] (SpecNeRF), and the only existing HNVS method HS-NeRF [5]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 240, + 545, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 240, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 304, + 240, + 545, + 384 + ], + "type": "text", + "content": "All experiments were conducted on an NVIDIA A100 80GB GPU, as MipNeRF360 requires significantly more VRAM than the other methods, which could run on a NVIDIA 3090 GPU. Using the A100 also facilitated accurate tracking of training times across all methods. We assess the performance of HyperGS after 60k training iterations, while NeRF methods are trained with the standard 1024 rays per batch and competing methods using rigorous metrics that capture both accuracy and spectral fidelity: PSNR, SSIM, Spectral Angle Mapping (SAM), RMSE. We use a " + }, + { + "bbox": [ + 304, + 240, + 545, + 384 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 304, + 240, + 545, + 384 + ], + "type": "text", + "content": " training and " + }, + { + "bbox": [ + 304, + 240, + 545, + 384 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 240, + 545, + 384 + ], + "type": "text", + "content": " test split according to the HS-NeRF dataset." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 393, + 371, + 405 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 393, + 371, + 405 + ], + "spans": [ + { + "bbox": [ + 306, + 393, + 371, + 405 + ], + "type": "text", + "content": "5.1. Real HSI" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 413, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 413, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 413, + 545, + 533 + ], + "type": "text", + "content": "We compare two HNVS datasets from HS-NeRF with different spectral channels and noise levels. The SOC710-VP (SOP) camera, covering " + }, + { + "bbox": [ + 304, + 413, + 545, + 533 + ], + "type": "inline_equation", + "content": "370 - 1100\\mathrm{nm}" + }, + { + "bbox": [ + 304, + 413, + 545, + 533 + ], + "type": "text", + "content": ", provides high spatial " + }, + { + "bbox": [ + 304, + 413, + 545, + 533 + ], + "type": "inline_equation", + "content": "(696\\times 520)" + }, + { + "bbox": [ + 304, + 413, + 545, + 533 + ], + "type": "text", + "content": " and spectral (128 channels) resolution but suffers from poor temporal resolution due to long exposure. The BaySpec GoldenEye camera offers comparable spatial " + }, + { + "bbox": [ + 304, + 413, + 545, + 533 + ], + "type": "inline_equation", + "content": "(640\\times 512)" + }, + { + "bbox": [ + 304, + 413, + 545, + 533 + ], + "type": "text", + "content": " and spectral (141 channels) resolution with faster captures, introducing more noise. Tables 1 and 2 show that HyperGS consistently outperforms other methods on unseen images across all scenes." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 533, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 628 + ], + "type": "text", + "content": "The BaySpec dataset particularly highlights HyperGS's strengths. Its autoencoder produces bounded errors in latent space, letting HyperGS effectively manage noise by averaging spectra and decoding within well-defined 1D feature structures. This yields more accurate reconstructions than standard end-to-end training methods. Figure 6 further shows how HyperGS better predicts object size and specular reflections compared to other baselines." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 630, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 714 + ], + "type": "text", + "content": "Similar to its performance on natural NeRF tasks, MipNeRF360 performed the best among NeRF methods for HNVS. We attribute its performance to its use of positional encoding and the efficient scene representation using tensor decomposition initially employed by TensorF providing more targeted spectral predictions. Our densification and pruning techniques further refine HyperGS's performance" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5975" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 75, + 148, + 155 + ], + "blocks": [ + { + "bbox": [ + 49, + 75, + 148, + 155 + ], + "lines": [ + { + "bbox": [ + 49, + 75, + 148, + 155 + ], + "spans": [ + { + "bbox": [ + 49, + 75, + 148, + 155 + ], + "type": "image", + "image_path": "a49e2bc0f82459c8fb593d4372c05c484abc4ea2552c2606224741f5d2af7e32.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 151, + 76, + 242, + 155 + ], + "blocks": [ + { + "bbox": [ + 151, + 76, + 242, + 155 + ], + "lines": [ + { + "bbox": [ + 151, + 76, + 242, + 155 + ], + "spans": [ + { + "bbox": [ + 151, + 76, + 242, + 155 + ], + "type": "image", + "image_path": "12157b5826d1c8eb0ec2fc694f54ac3bd81a1b93eb18273eb0f53adcc4803fc4.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 245, + 76, + 337, + 155 + ], + "blocks": [ + { + "bbox": [ + 245, + 76, + 337, + 155 + ], + "lines": [ + { + "bbox": [ + 245, + 76, + 337, + 155 + ], + "spans": [ + { + "bbox": [ + 245, + 76, + 337, + 155 + ], + "type": "image", + "image_path": "19027e83ac16bd4a7b32b45b63e0022ee2de1760a6056b8cc6dd2208ef61268f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 340, + 76, + 432, + 155 + ], + "blocks": [ + { + "bbox": [ + 340, + 76, + 432, + 155 + ], + "lines": [ + { + "bbox": [ + 340, + 76, + 432, + 155 + ], + "spans": [ + { + "bbox": [ + 340, + 76, + 432, + 155 + ], + "type": "image", + "image_path": "703e124c97e00ef763508d76056331528e4a47819ef5cd428d5974a0950b546e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 434, + 76, + 526, + 155 + ], + "blocks": [ + { + "bbox": [ + 434, + 76, + 526, + 155 + ], + "lines": [ + { + "bbox": [ + 434, + 76, + 526, + 155 + ], + "spans": [ + { + "bbox": [ + 434, + 76, + 526, + 155 + ], + "type": "image", + "image_path": "9b6399da5fc0a596fb6a4837f8360145f3e0b549d6dfce57d907848182645a70.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 49, + 158, + 148, + 232 + ], + "blocks": [ + { + "bbox": [ + 49, + 158, + 148, + 232 + ], + "lines": [ + { + "bbox": [ + 49, + 158, + 148, + 232 + ], + "spans": [ + { + "bbox": [ + 49, + 158, + 148, + 232 + ], + "type": "image", + "image_path": "c68127b80fafa5b00deea2c061f2cf18838ac7b21ae659baa02c93ed04470294.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 151, + 158, + 243, + 232 + ], + "blocks": [ + { + "bbox": [ + 151, + 158, + 243, + 232 + ], + "lines": [ + { + "bbox": [ + 151, + 158, + 243, + 232 + ], + "spans": [ + { + "bbox": [ + 151, + 158, + 243, + 232 + ], + "type": "image", + "image_path": "36bc1370b0629e090bfb637c19279218c95b40cd527bb2b3aa67eef1f89cb8b2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 245, + 158, + 337, + 232 + ], + "blocks": [ + { + "bbox": [ + 245, + 158, + 337, + 232 + ], + "lines": [ + { + "bbox": [ + 245, + 158, + 337, + 232 + ], + "spans": [ + { + "bbox": [ + 245, + 158, + 337, + 232 + ], + "type": "image", + "image_path": "be510ed92286bd483b60c30bd2e70245cfcb14ec3fc0024714efd651272a255d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 340, + 158, + 432, + 232 + ], + "blocks": [ + { + "bbox": [ + 340, + 158, + 432, + 232 + ], + "lines": [ + { + "bbox": [ + 340, + 158, + 432, + 232 + ], + "spans": [ + { + "bbox": [ + 340, + 158, + 432, + 232 + ], + "type": "image", + "image_path": "e698b2475801188354984d79756ad257707a031fd0684eb5b3108f1a335f51b4.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 434, + 155, + 546, + 236 + ], + "blocks": [ + { + "bbox": [ + 434, + 155, + 546, + 236 + ], + "lines": [ + { + "bbox": [ + 434, + 155, + 546, + 236 + ], + "spans": [ + { + "bbox": [ + 434, + 155, + 546, + 236 + ], + "type": "image", + "image_path": "6bbe325aa10cf4fc3727a0c93cd375a338f0f93c6029a1c019cc0cfcb9cde93f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 236, + 545, + 257 + ], + "lines": [ + { + "bbox": [ + 47, + 236, + 545, + 257 + ], + "spans": [ + { + "bbox": [ + 47, + 236, + 545, + 257 + ], + "type": "text", + "content": "Figure 6. Visualisation of the top 4 methods for frame 51 of 359 for the Caladium plant scene from the Bayspec dataset. The top row shows the 70th channel of the 141 channel image predicted, the bottom row provides a raw pixel-wise error heatmap of the scene." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 50, + 264, + 148, + 339 + ], + "blocks": [ + { + "bbox": [ + 50, + 264, + 148, + 339 + ], + "lines": [ + { + "bbox": [ + 50, + 264, + 148, + 339 + ], + "spans": [ + { + "bbox": [ + 50, + 264, + 148, + 339 + ], + "type": "image", + "image_path": "e89f175cbb169ca19add91b717112882ab03a5214b2715d3b7395437e1a546b5.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 151, + 264, + 242, + 339 + ], + "blocks": [ + { + "bbox": [ + 151, + 264, + 242, + 339 + ], + "lines": [ + { + "bbox": [ + 151, + 264, + 242, + 339 + ], + "spans": [ + { + "bbox": [ + 151, + 264, + 242, + 339 + ], + "type": "image", + "image_path": "9ca7e8fddbc7a4a652cf945b8e75f8ea8de544b6213f1100ce8ed48be0425cd9.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 245, + 264, + 337, + 339 + ], + "blocks": [ + { + "bbox": [ + 245, + 264, + 337, + 339 + ], + "lines": [ + { + "bbox": [ + 245, + 264, + 337, + 339 + ], + "spans": [ + { + "bbox": [ + 245, + 264, + 337, + 339 + ], + "type": "image", + "image_path": "617c68515f5d61d85207e307c57c4afa8234d969986d4736bbdb1d7c15a8fe4d.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 340, + 264, + 432, + 339 + ], + "blocks": [ + { + "bbox": [ + 340, + 264, + 432, + 339 + ], + "lines": [ + { + "bbox": [ + 340, + 264, + 432, + 339 + ], + "spans": [ + { + "bbox": [ + 340, + 264, + 432, + 339 + ], + "type": "image", + "image_path": "4d604f66b83a9f7a46ebc4672cbbd0708aef82715a3d4327141b4479d113952a.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 434, + 264, + 545, + 339 + ], + "blocks": [ + { + "bbox": [ + 434, + 264, + 545, + 339 + ], + "lines": [ + { + "bbox": [ + 434, + 264, + 545, + 339 + ], + "spans": [ + { + "bbox": [ + 434, + 264, + 545, + 339 + ], + "type": "image", + "image_path": "e5ba4e3e1358514855a7043540f6e9f9667e9fa74b67b9bdf72f8f4545002a42.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 50, + 346, + 148, + 416 + ], + "blocks": [ + { + "bbox": [ + 50, + 346, + 148, + 416 + ], + "lines": [ + { + "bbox": [ + 50, + 346, + 148, + 416 + ], + "spans": [ + { + "bbox": [ + 50, + 346, + 148, + 416 + ], + "type": "image", + "image_path": "18cbc74fb3341a1027b07275d13246f0822c0b07f4a4298f11b2e2fc99f22708.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 151, + 347, + 242, + 416 + ], + "blocks": [ + { + "bbox": [ + 151, + 347, + 242, + 416 + ], + "lines": [ + { + "bbox": [ + 151, + 347, + 242, + 416 + ], + "spans": [ + { + "bbox": [ + 151, + 347, + 242, + 416 + ], + "type": "image", + "image_path": "24e1860ad484263dfb17faf68fb1d49ba5eaf9fa637c061aebfcc815fbb68722.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 245, + 347, + 337, + 416 + ], + "blocks": [ + { + "bbox": [ + 245, + 347, + 337, + 416 + ], + "lines": [ + { + "bbox": [ + 245, + 347, + 337, + 416 + ], + "spans": [ + { + "bbox": [ + 245, + 347, + 337, + 416 + ], + "type": "image", + "image_path": "ac66b354eaf0902cd3201aeeb964e6fb6618432a63d6bad548238d3d82290924.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 340, + 347, + 432, + 416 + ], + "blocks": [ + { + "bbox": [ + 340, + 347, + 432, + 416 + ], + "lines": [ + { + "bbox": [ + 340, + 347, + 432, + 416 + ], + "spans": [ + { + "bbox": [ + 340, + 347, + 432, + 416 + ], + "type": "image", + "image_path": "41a7de3963f588c8bbd049a5aa51ca7814198601dbfdd0e38349fac7248ad620.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 434, + 341, + 546, + 422 + ], + "blocks": [ + { + "bbox": [ + 434, + 341, + 546, + 422 + ], + "lines": [ + { + "bbox": [ + 434, + 341, + 546, + 422 + ], + "spans": [ + { + "bbox": [ + 434, + 341, + 546, + 422 + ], + "type": "image", + "image_path": "8eca2ed85e5e17b5fbdc92be03dbd0524171e46e212fe829131fbc464e671517.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 422, + 545, + 444 + ], + "lines": [ + { + "bbox": [ + 46, + 422, + 545, + 444 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 545, + 444 + ], + "type": "text", + "content": "Figure 7. Visualisation of the top 4 methods for frame 31 of 49 for the Tools plant scene from the SOP dataset. The top row shows the 70th channel of the 128 channel image predicted, the bottom row provides a raw pixel-wise error heatmap of the scene." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 46, + 451, + 287, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 451, + 287, + 701 + ], + "spans": [ + { + "bbox": [ + 46, + 451, + 287, + 701 + ], + "type": "text", + "content": "by effectively filtering poor spectral representations. Meanwhile, the SOP dataset favors both 3DGS approaches. Due to the high number of frames present in the BaySpec camera scenes, the NeRF approaches adapt well to this abundance of data and can provide far better and more stable representations of the scene given the volume of viewpoints. This volume makes predicting the noisier spectra that the BaySpec camera provides easier than 3DGS. In contrast, although the SOP datasets may have smoother spectra, the number of training viewpoints is far smaller. This leads to NeRF approaches evidently struggling with scene scale and spectral reconstruction, as they struggle to understand the scene scale and perform spectral reconstruction due to the lack of data. In contrast, 3DGS approaches excel due to smoother data interpolation and simpler optimization. This is outlined by Figure 7 highlighting the intense error heatmaps from NeRF approaches compared to 3DGS and HyperGS. HyperGS, however, outperforms all methods across all scenes and camera datasets, suggesting that it provides a far more robust and accurate performance than both 3DGS and NeRF methods for HNVS. For visualiza" + } + ] + } + ], + "index": 22 + }, + { + "type": "table", + "bbox": [ + 310, + 459, + 541, + 594 + ], + "blocks": [ + { + "bbox": [ + 389, + 449, + 463, + 459 + ], + "lines": [ + { + "bbox": [ + 389, + 449, + 463, + 459 + ], + "spans": [ + { + "bbox": [ + 389, + 449, + 463, + 459 + ], + "type": "text", + "content": "BaySpec Datasets" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 459, + 541, + 594 + ], + "lines": [ + { + "bbox": [ + 310, + 459, + 541, + 594 + ], + "spans": [ + { + "bbox": [ + 310, + 459, + 541, + 594 + ], + "type": "table", + "html": "
MethodAverage Results
PSNR ↑SSIM↑SAM↓RMSE↓
NeRF23.350.60610.04400.0687
MipNeRF22.750.59470.04350.0776
TensoRF24.660.64820.05010.0587
Nerfacto19.120.58660.05510.1174
MipNerf36026.530.74420.02800.0476
HS-NeRF19.820.67140.05340.1071
3DGS22.910.63210.13350.0810
HyperGS27.110.78040.02540.0440
", + "image_path": "905de4e1555dd323830ac55bbebbcbf4853ed4d1155adb7cde5e46864b72e71e.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "table_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 306, + 604, + 545, + 636 + ], + "lines": [ + { + "bbox": [ + 306, + 604, + 545, + 636 + ], + "spans": [ + { + "bbox": [ + 306, + 604, + 545, + 636 + ], + "type": "text", + "content": "Table 1. Quantitative results using the Bayspec datasets against separate hyperspectral methods and baseline NeRF and 3DGS (best bold, second best italic)." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 665, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 545, + 712 + ], + "type": "text", + "content": "tions of random pixels and spectral images, please refer to Figures 8 and then 9 which showcases HyperGS's superior pixel reconstruction capabilities compared to other benchmarked methods." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5976" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 213, + 178 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 213, + 178 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 213, + 178 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 213, + 178 + ], + "type": "image", + "image_path": "93fdca6666af58207f82e65fa5c76f3b5f49e0f34e1fe6d2adbe247383352798.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 99, + 182, + 493, + 192 + ], + "lines": [ + { + "bbox": [ + 99, + 182, + 493, + 192 + ], + "spans": [ + { + "bbox": [ + 99, + 182, + 493, + 192 + ], + "type": "text", + "content": "Figure 8. Three random pixel reconstructions taken from test frame 151 of the bayspec Anacampseros scene." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 215, + 72, + 377, + 179 + ], + "blocks": [ + { + "bbox": [ + 215, + 72, + 377, + 179 + ], + "lines": [ + { + "bbox": [ + 215, + 72, + 377, + 179 + ], + "spans": [ + { + "bbox": [ + 215, + 72, + 377, + 179 + ], + "type": "image", + "image_path": "ffa9223e76a854222c5389fc41bd6fc4c74c538f20d3e44e74ac659be534e68f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 379, + 72, + 542, + 175 + ], + "blocks": [ + { + "bbox": [ + 379, + 72, + 542, + 175 + ], + "lines": [ + { + "bbox": [ + 379, + 72, + 542, + 175 + ], + "spans": [ + { + "bbox": [ + 379, + 72, + 542, + 175 + ], + "type": "image", + "image_path": "51c6f99a93dfbfaedcd6dbdf9eed0ab238e53592c618c701eb7453682797c033.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 50, + 194, + 213, + 297 + ], + "blocks": [ + { + "bbox": [ + 50, + 194, + 213, + 297 + ], + "lines": [ + { + "bbox": [ + 50, + 194, + 213, + 297 + ], + "spans": [ + { + "bbox": [ + 50, + 194, + 213, + 297 + ], + "type": "image", + "image_path": "9387353f18a257f80946db0266b31f023ac592de2a94d9d6a5cb69c411f1fa75.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 304, + 462, + 315 + ], + "lines": [ + { + "bbox": [ + 130, + 304, + 462, + 315 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 462, + 315 + ], + "type": "text", + "content": "Figure 9. Three random pixel reconstructions from test frame 31 of the SOP origami scene." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 215, + 194, + 377, + 301 + ], + "blocks": [ + { + "bbox": [ + 215, + 194, + 377, + 301 + ], + "lines": [ + { + "bbox": [ + 215, + 194, + 377, + 301 + ], + "spans": [ + { + "bbox": [ + 215, + 194, + 377, + 301 + ], + "type": "image", + "image_path": "b8777cd475ed739d5d64250c1d8e8cc0bd03be14bff316740a3f618e9e85924c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 379, + 194, + 542, + 296 + ], + "blocks": [ + { + "bbox": [ + 379, + 194, + 542, + 296 + ], + "lines": [ + { + "bbox": [ + 379, + 194, + 542, + 296 + ], + "spans": [ + { + "bbox": [ + 379, + 194, + 542, + 296 + ], + "type": "image", + "image_path": "7e3388e5f38d886dd81476a083d3e2dcb95fabf0553c9480a79f4a7dd784e557.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 50, + 335, + 282, + 466 + ], + "blocks": [ + { + "bbox": [ + 118, + 324, + 216, + 335 + ], + "lines": [ + { + "bbox": [ + 118, + 324, + 216, + 335 + ], + "spans": [ + { + "bbox": [ + 118, + 324, + 216, + 335 + ], + "type": "text", + "content": "Surface Optics Datasets" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 335, + 282, + 466 + ], + "lines": [ + { + "bbox": [ + 50, + 335, + 282, + 466 + ], + "spans": [ + { + "bbox": [ + 50, + 335, + 282, + 466 + ], + "type": "table", + "html": "
MethodAverage Results
PSNR ↑SSIM↑SAM↓RMSE↓
NeRF10.890.59050.06250.3479
MipNeRF12.530.54810.05680.3198
TensorF13.000.56960.05950.2780
Nerfacto16.370.69860.03520.1601
MipNeRF36012.280.68240.13690.2658
HS-NeRF14.440.61650.20370.1953
3DGS28.580.96270.03010.0478
HyperGS30.510.97560.004150.0354
", + "image_path": "a4d39e33275aec94209d146ccb910c1ea71e346c8435a089742199786a9bef64.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 472, + 287, + 495 + ], + "lines": [ + { + "bbox": [ + 46, + 472, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 287, + 495 + ], + "type": "text", + "content": "Table 2. Quantitative results using the SOP datasets against separate hyperspectral methods and baseline NeRF and 3DGS." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 502, + 141, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 502, + 141, + 515 + ], + "spans": [ + { + "bbox": [ + 47, + 502, + 141, + 515 + ], + "type": "text", + "content": "5.2. Ablation Study" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 522, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 289, + 715 + ], + "type": "text", + "content": "To evaluate the effectiveness and impact of various components in our HyperGS model, we conducted a series of ablation experiments. These experiments help to understand the contribution of individual components and choices in our model's design. As shown in Table 3, each new feature added to the baseline 3DGS model improves HNVS performance. The joint latent autoencoder-3DGS architecture boosts spectral reconstruction the most while the proposed densification provides greater overall details captured. With the latent space learning and positional encoding, the performance of HyperGS gets a significant boost in performance too. Please refer to our supplementary materials for additional ablations, including analyses of HyperGS performance, pruning strategy scoring functions, autoencoder performance with varying data and latent sizes, the global pruning strategy, and system speed bottlenecks." + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 309, + 325, + 544, + 414 + ], + "blocks": [ + { + "bbox": [ + 309, + 325, + 544, + 414 + ], + "lines": [ + { + "bbox": [ + 309, + 325, + 544, + 414 + ], + "spans": [ + { + "bbox": [ + 309, + 325, + 544, + 414 + ], + "type": "table", + "html": "
Ablation StepAverage Results for the Bayspec dataset
PSNR ↑SSIM↑SAM↓RMSE↓N.Prim ↓
Base. 3DGS22.910.63200.13350.0810440k
+ Spec. SFM23.050.63310.13100.0799421k
+ Latent space AE24.870.71010.05480.0602500k
+ Densification25.250.73560.03650.05481.3M
+ Pruning25.170.71990.03740.0555412k
+ View dependent MLP27.050.77920.02530.0443309k
+ Custom L.Function27.110.78010.02540.0440310k
", + "image_path": "7ef34aa10c4f4ce02e255f423a779ee57d3f998a46b16d7a0d9e9574fb12e50a.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 304, + 419, + 547, + 464 + ], + "lines": [ + { + "bbox": [ + 304, + 419, + 547, + 464 + ], + "spans": [ + { + "bbox": [ + 304, + 419, + 547, + 464 + ], + "type": "text", + "content": "Table 3. Adding each new feature to 3DGS significantly improves the model's average performance for the HNVS task in the bayspec dataset while creating a more stable number of primitives (N.Prim) in the point cloud." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 472, + 384, + 484 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 472, + 384, + 484 + ], + "spans": [ + { + "bbox": [ + 306, + 472, + 384, + 484 + ], + "type": "text", + "content": "6. Conclusions" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 491, + 547, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 491, + 547, + 634 + ], + "spans": [ + { + "bbox": [ + 304, + 491, + 547, + 634 + ], + "type": "text", + "content": "In this paper, we introduced HyperGS, the first effective approach for hyperspectral novel view synthesis (HNVS). By encoding latent spectral data within Gaussian primitives and performing Gaussian splatting in a learned latent space, HyperGS achieves high-quality, material-aware rendering. Our adaptive density control and pruning techniques efficiently handle latent hyperspectral signatures in 3D Gaussian point clouds, ensuring stable training and superior accuracy compared to benchmark methods. For future work, we aim to extend HyperGS with more comprehensive latent space decoding that corrects and refines both spectral and spatial information, providing richer contextual detail." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 648, + 421, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 648, + 421, + 662 + ], + "spans": [ + { + "bbox": [ + 306, + 648, + 421, + 662 + ], + "type": "text", + "content": "7. Acknowledgements" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 669, + 547, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 669, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 669, + 547, + 712 + ], + "type": "text", + "content": "This work was partially supported by Industrial 3D Robotics (I3D), IS-Instruments and partially funded by the EPSRC under grant agreement EP/S035761/1." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5977" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] J.T. Barron, B. Mildenhall, D. Verbin, P.P. Srinivasan, and P. Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 4, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 147, + 288, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 147, + 288, + 191 + ], + "spans": [ + { + "bbox": [ + 53, + 147, + 288, + 191 + ], + "type": "text", + "content": "[2] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields, 2021. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 193, + 288, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 193, + 288, + 214 + ], + "spans": [ + { + "bbox": [ + 53, + 193, + 288, + 214 + ], + "type": "text", + "content": "[3] S. R. Bulò, L. Porzi, and P. Kontschieder. Revising densification in gaussian splatting. ArXiv, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 215, + 288, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 215, + 288, + 248 + ], + "spans": [ + { + "bbox": [ + 53, + 215, + 288, + 248 + ], + "type": "text", + "content": "[4] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision (ECCV), 2022. 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 249, + 288, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 249, + 288, + 282 + ], + "spans": [ + { + "bbox": [ + 53, + 249, + 288, + 282 + ], + "type": "text", + "content": "[5] G. Chen, S. K. Narayanan, T. G. Ottou, B. Missaoui, H. Muriki, C. Pradalier, and Y. Chen. Hyperspectral neural radiance fields. ArXiv, 2024. 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 284, + 288, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 284, + 288, + 327 + ], + "spans": [ + { + "bbox": [ + 53, + 284, + 288, + 327 + ], + "type": "text", + "content": "[6] Kai Cheng, Xiaoxiao Long, Kaizhi Yang, Yao Yao, Wei Yin, Yuexin Ma, Wenping Wang, and Xuejin Chen. Gaussianpro: 3d gaussian splatting with progressive propagation. arXiv preprint arXiv:2402.14650, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 329, + 288, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 329, + 288, + 384 + ], + "spans": [ + { + "bbox": [ + 53, + 329, + 288, + 384 + ], + "type": "text", + "content": "[7] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proc. Computer Vision and Pattern Recognition (CVPR), IEEE, 2017. 6, 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 385, + 286, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 385, + 286, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 385, + 286, + 407 + ], + "type": "text", + "content": "[8] F. Dellaert. NeRF explosion 2020. https://dellaert.github.io/NeRF/, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 407, + 288, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 407, + 288, + 452 + ], + "spans": [ + { + "bbox": [ + 53, + 407, + 288, + 452 + ], + "type": "text", + "content": "[9] Z. Fan, K. Wang, K. Wen, Z. Zhu, D. Xu, and Z. Wang. Lightgaussian: Unbounded 3d gaussian compression with 15x reduction and " + }, + { + "bbox": [ + 53, + 407, + 288, + 452 + ], + "type": "inline_equation", + "content": "200+" + }, + { + "bbox": [ + 53, + 407, + 288, + 452 + ], + "type": "text", + "content": " fps. arXiv preprint arXiv:2311.17245, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 453, + 287, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 453, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 48, + 453, + 287, + 496 + ], + "type": "text", + "content": "[10] Z. Fan, W. Cong, K. Wen, K. Wang, J. Zhang, X. Ding, D. Xu, B. Ivanovic, M. Pavone, G. Pavlakos, Z. Wang, and Y. Wang. Instantsplat: Sparse-view sfm-free gaussian splatting in seconds. ArXiv, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 498, + 287, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 498, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 48, + 498, + 287, + 531 + ], + "type": "text", + "content": "[11] S. Girish, K. Gupta, and A. Shrivastava. Eagles: Efficient accelerated 3d gaussians with lightweight encodings. arXiv preprint arXiv:2312.04564, 2023. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 533, + 287, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 533, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 48, + 533, + 287, + 576 + ], + "type": "text", + "content": "[12] Jaewoo Jung, Jisang Han, Honggyu An, Jiwon Kang, Seonghoon Park, and Seungryong Kim. Relaxing accurate initialization constraint for 3d gaussian splatting. arXiv preprint arXiv:2403.09413, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 578, + 287, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 621 + ], + "type": "text", + "content": "[13] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (SIGGRAPH Conference Proceedings), 42(4), 2023. 1, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 623, + 287, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 623, + 287, + 655 + ], + "spans": [ + { + "bbox": [ + 48, + 623, + 287, + 655 + ], + "type": "text", + "content": "[14] Byeonghyeon Lee, Howoong Lee, Xiangyu Sun, Usman Ali, and Eunbyung Park. Deblurring 3d gaussian splatting, 2024. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 657, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 657, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 657, + 287, + 678 + ], + "type": "text", + "content": "[15] J. Li, Y. Li, C. Sun, C. Wang, and J. Xiang. Spec-nerf: Multispectral neural radiance fields. ArXiv, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[16] R. Li, J. Liu, G. Liu, S. Zhang, B. Zeng, and S. Liu. Spectralnerf: Physically based spectral rendering with neural radiation field. ArXiv, 2023. 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 715 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[17] Jie Liang, Ali Zia, Jun Zhou, and Xavier Sirault. 3d plant modelling via hyperspectral imaging. In 2013 IEEE International Conference on Computer Vision Workshops, pages 172-177, 2013. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "text", + "content": "[18] W. Liu, T. Guan, B. Zhu, L. Ju, Z. Song, D. Li, Y. Wang, and W. Yang. Efficientges: Streamlining gaussian splatting for large-scale high-resolution scene representation. ArXiv, 2024. 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 163, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 196 + ], + "type": "text", + "content": "[19] T. Lu, M. Yu, L. Xu, Y. Xiangli, L. Wang, D. Lin, and B. Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. ArXiv, 2023. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 197, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 228 + ], + "type": "text", + "content": "[20] D. Malarz, W. Smolak, J. Tabor, S. Tadeja, and P. Spurek. Gaussian splatting with nef-based color and opacity. ArXiv, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 230, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 230, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 545, + 262 + ], + "type": "text", + "content": "[21] D. Malarz, W. Smolak, J. Tabor, S. Tadeja, and P. Spurek. Gaussian splatting with nef-based color and opacity. ArXiv, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "type": "text", + "content": "[22] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 309, + 545, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 309, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 307, + 309, + 545, + 365 + ], + "type": "text", + "content": "[23] B. Mildenhall, P. Hedman, R. Martin-Brualla, P.P. Srinivasan, and J.T. Barron. Nerf in the dark: High dynamic range view synthesis from noisy raw images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 365, + 545, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 408 + ], + "type": "text", + "content": "[24] T. Müller, A. Evans, C. Schied, and A. Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (TOG), 41(4):102:1-102:15, 2022. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 410, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 410, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 307, + 410, + 545, + 443 + ], + "type": "text", + "content": "[25] S. Niedermayr, J. Stumpfegger, and R. Westermann. Compressed 3d gaussian splatting for accelerated novel view synthesis. arXiv preprint arXiv:2401.02436, 2023. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 445, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 445, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 307, + 445, + 545, + 498 + ], + "type": "text", + "content": "[26] M. Tancik, E. Weber, E. Ng, R. Li, B. Yi, J. Kerr, T. Wang, A. Kristoffersen, J. Austin, K. Salahi, A. Ahuja, D. McAllister, and A. Kanazawa. Nerfstudio: A modular framework for neural radiance field development. arXiv preprint arXiv:2302.04264, 2023. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 500, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 500, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 500, + 545, + 555 + ], + "type": "text", + "content": "[27] Christopher Thirgood, Oscar Mendez, Erin Chao Ling, Jon Storey, and Simon Hadfield. Raspectloc: Raman spectroscopy-dependent robot localisation. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5296-5303, 2023. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 556, + 545, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 556, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 307, + 556, + 545, + 589 + ], + "type": "text", + "content": "[28] S. Wang, V. Leroy, Y. Cabon, B. Chidlovskii, and J. Revaud. DUSt3R: Geometric 3D Vision Made Easy. arXiv preprint arXiv:2312.14132, 2023. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 590, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 545, + 633 + ], + "type": "text", + "content": "[29] Z. Yang, X. Gao, Y. Sun, Y. Huang, X. Lyu, W. Zhou, S. Jiao, X. Qi, and X. Jin. Spec-Gaussian: Anisotropic View-Dependent Appearance for 3D Gaussian Splatting. ArXiv, 2024. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 635, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 689 + ], + "type": "text", + "content": "[30] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19447-19456, 2024. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 691, + 545, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 715 + ], + "type": "text", + "content": "[31] A. Zia, J. Liang, J. Zhou, and Y. Gao. 3d reconstruction from hyperspectral images. In 2015 IEEE Winter Conference on" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "5978" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 138 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "type": "text", + "content": "Applications of Computer Vision (WACV), pages 318-325, 2015. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 138 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 138 + ], + "type": "text", + "content": "[32] Ali Zia, Jie Liang, Jun Zhou, and Yongsheng Gao. 3d reconstruction from hyperspectral images. In 2015 IEEE Winter Conference on Applications of Computer Vision, pages 318-325, 2015. 2" + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "5979" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/HyperLoRA_ Parameter-Efficient Adaptive Generation for Portrait Synthesis/0dcd811d-bb8b-4fcc-9377-9711c569d3e6_content_list.json b/2025/HyperLoRA_ Parameter-Efficient Adaptive Generation for Portrait Synthesis/0dcd811d-bb8b-4fcc-9377-9711c569d3e6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c5b8c445a7255ca5483bb73d6ebcfaf3b3f73f6a --- /dev/null +++ b/2025/HyperLoRA_ Parameter-Efficient Adaptive Generation for Portrait Synthesis/0dcd811d-bb8b-4fcc-9377-9711c569d3e6_content_list.json @@ -0,0 +1,1391 @@ +[ + { + "type": "text", + "text": "HyperLoRA: Parameter-Efficient Adaptive Generation for Portrait Synthesis", + "text_level": 1, + "bbox": [ + 109, + 130, + 888, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mengtian Li*", + "bbox": [ + 155, + 179, + 264, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jinshu Chen*", + "bbox": [ + 303, + 180, + 411, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wanquan Feng*†", + "bbox": [ + 452, + 180, + 586, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Songtao Zhao", + "bbox": [ + 387, + 199, + 500, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qian He", + "bbox": [ + 540, + 199, + 609, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Intelligent Creation, ByteDance", + "bbox": [ + 367, + 222, + 622, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Bingchuan Li", + "bbox": [ + 627, + 181, + 738, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fei Dai", + "bbox": [ + 777, + 181, + 839, + 196 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2b18acae0a80c6b533d7c6bdf8da7ab287eeee0eea655aea0bfb40567ae17b4d.jpg", + "image_caption": [ + "HyperLoRA", + "w.o. fine-tuning" + ], + "image_footnote": [], + "bbox": [ + 119, + 273, + 331, + 501 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8e4f8b93bf2cee9cca76c169c42a560659ad4669d7e9f5ba7c643a48510897f4.jpg", + "image_caption": [ + "Wizard hat" + ], + "image_footnote": [], + "bbox": [ + 343, + 273, + 517, + 410 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/527d66c7ec402cd61404ff6d50c1345cfe4d97b11593ef85cd9996bb03d6b350.jpg", + "image_caption": [ + "Pink wavy hair" + ], + "image_footnote": [], + "bbox": [ + 521, + 273, + 697, + 410 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/33367ff554f711305f906ea5b484aec86bf46f0a046825bd5302cee8de043e9e.jpg", + "image_caption": [ + "Chinese clothing" + ], + "image_footnote": [], + "bbox": [ + 700, + 273, + 877, + 410 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/efb6a94812e77b87348a40623bae2e7ec47691139244f8276874c99ad3678cce.jpg", + "image_caption": [ + "Snow", + "Figure 1. We propose HyperLoRA, a parameter-efficient adaptive method for portrait synthesis. Given an input face image, HyperLoRA generates personalized LoRA weights without online fine-tuning. Due to the natural interpolability of LoRA, it is easy to support multiple inputs by simple averaging. Leveraging the generated LoRA, we can create personalized portraits with high photorealism and fidelity." + ], + "image_footnote": [], + "bbox": [ + 343, + 435, + 519, + 571 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ca26977f5a8df05a64d117118b84870a78868a3eb97e49848649369bf46b9546.jpg", + "image_caption": [ + "Beach" + ], + "image_footnote": [], + "bbox": [ + 521, + 435, + 697, + 570 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/89dcaae03d9e79937c984dcc9d617622b7ce36da044ecb805cb7f7c62da5e4ac.jpg", + "image_caption": [ + "Cyberpunk" + ], + "image_footnote": [], + "bbox": [ + 700, + 435, + 877, + 571 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 654, + 326, + 670 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Personalized portrait synthesis, essential in domains like social entertainment, has recently made significant progress. Person-wise fine-tuning based methods, such as LoRA and DreamBooth, can produce photorealistic outputs but need training on individual samples, consuming time and resources and posing an unstable risk. Adapter based techniques such as IP-Adapter freeze the foundational model parameters and employ a plug-in architecture to enable zero-shot inference, but they often exhibit a lack of naturalness and authenticity, which are not to be overlooked in portrait synthesis tasks. In this paper, we introduce a parameter-efficient adaptive generation method,", + "bbox": [ + 88, + 688, + 485, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "namely HyperLoRA, that uses an adaptive plug-in network to generate LoRA weights, merging the superior performance of LoRA with the zero-shot capability of adapter scheme. Through our carefully designed network structure and training strategy, we achieve zero-shot personalized portrait generation (supporting both single and multiple image inputs) with high photorealism, fidelity, and editability.", + "bbox": [ + 511, + 656, + 908, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 784, + 645, + 800 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The personalized portrait synthesis task [14, 44, 45] aims to generate images sharing the same identity with what is contained in the given single or multiple portrait images. By feeding the identity information into specific generative model, the task enables the users to create diverse, high-quality and creative personalized portraits, thus hold", + "bbox": [ + 511, + 809, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution † Corresponding author", + "bbox": [ + 109, + 887, + 374, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13114", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing substantial importance across various domains, such as entertainment and social media.", + "bbox": [ + 89, + 90, + 480, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "One key challenge of the personalized portrait synthesis task is the fidelity: the expected generated images should retain the given identity to the greatest extent and avoid irrelevant outputs. Another challenge lies in editability: while preserving the given identity, the portrait synthesis model should maintain the abilities of flexibly editing the other contents of the generated images, for example, the composition of the image, the positioning and the clothing of the character. Additionally, the inference speed is crucial, as fast inference is friendly to not only the computational costs but also the user experience.", + "bbox": [ + 88, + 123, + 480, + 289 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Considering the challenges mentioned above, most of the current solutions have different shortcomings. Existing personalized image synthesis methods can generally be divided into tuning-based and tuning-free approaches. Tuning-based methods, such as LoRA (Low-Rank Adaptation) [18] and DreamBooth [35], update model parameters based on specific user-provided data, which can accurately comprehend and preserve the input identity information. Theoretically, these methods are mainly designed for few-shot cases, and customize different models for each groups of provided data with different identities. Thus, disadvantages like high computing costs, unstable training, and lengthy training processes make such methods bloated when dealing with various provided identities.", + "bbox": [ + 91, + 292, + 483, + 503 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Tuning-free methods, such as IP-Adapter [45], typically utilize an additional plug-in structure to encode user identity information. By feeding such identity-related tokens into the frozen foundational model, these methods make influences on the inference process, and ultimately yielding results that retain the given identities. Such additional adapter struture is trained with large portrait datasets in advance, and freezing when inference. Therefore, these methods get rid of the online training, and are convenient for the actual uses. However, for these methods, the interaction between encoded identity information and the foundational model relies on a group of fixed extra cross-attention modules, which might degrade the quality of generated images, leading to unnatural and non-photorealistic facial features, and fail to recover fine-grained details. In summary, the previous methods fail to simultaneously strike a balance between the fidelity, the editability, and the speed.", + "bbox": [ + 91, + 506, + 483, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we introduce a parameter-efficient adaptive generation method, namely HyperLoRA, which achieves great identity fidelity, high editability and fast inference speed. HyperLoRA uses an adaptive plug-in network to encode identity images, which follows the design of tuning-free methods like IP-Adapter. However, unlike such methods which utilize the encoded tokens for the attention computation, we directly generate the LoRA weights with the plug-in network. This strategy not only maintains the zero", + "bbox": [ + 89, + 763, + 480, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "shot capability of the adapter strategy, but also inherits the representation and high-quality generation effect of the LoRA strategy which adjusts the global parameters without real fine-tuning process.", + "bbox": [ + 511, + 90, + 903, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Considering the substantial number of parameters in LoRA, to reduce the training difficulty, we initially project the LoRA parameters onto a linear LoRA parameter space, rather than directly training in the original parameter space. In this way, the generated LoRA weights can be represented as a linear combination of a trainable basis. We employ a perceiver resampler (as in IP-Adapter [45]) to convert the input identity images into combination coefficients for the basis. To encode the input images, we utilize a CLIP ViT [33] and an AntelopeV2 encoder [20] to extract pixel-level and identity-sensitive features. These feature extractors mentioned above stay frozen during all the training process. To avoid overfitting, we decompose LoRA at the parameter level into an ID-LoRA part and a Base-LoRA part, thereby the ID and ID-irreverent information are fitted into different parts correspondingly. The whole training pipeline runs in a multi-stage way: only the Base-LoRA part is trained in the first stage as a warm-up start; and in the second stage, we train the ID-LoRA with CLIP ViT features; in the final stage, we only finetune the attention modules related to ID embeddings. Besides the case where only a single image is given, our proposed method can also handle cases that there are multiple inputs by simply interpolating the predicted LoRA weights. Our contributions can be summarized as follows:", + "bbox": [ + 511, + 152, + 903, + 529 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose HyperLoRA, a parameter-efficient adaptive method for the personalized portrait synthesis task, merging the superior performance of the LoRA strategy with the zero-shot capability of the adapter strategy.", + "- We construct a low-dimension linear LoRA space, which reduces the training difficulty of the plug-in network while ensuring the identity preservation.", + "- We design a strategy that decomposes LoRA weights at the parameter level into a base part and an ID part, thereby effectively decoupling the identity information restored in the given images and significantly reducing overfitting.", + "- We conduct various experiments to demonstrate the capacities of our framework. Experimental results show that our method achieves both high fidelity and editability." + ], + "bbox": [ + 513, + 532, + 903, + 741 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 760, + 653, + 776 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Text-to-Image Synthesis", + "text_level": 1, + "bbox": [ + 513, + 786, + 733, + 801 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-to-image (T2I) synthesis has been rapidly developed for several years. The task aims to generate high-quality images which are highly correlated with given texts or instructions. Some early works [21, 42, 47] employ transformer [43] or GAN [9] architectures to solve this task, while the performance of the solutions is not that satisfy", + "bbox": [ + 511, + 809, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13115", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ing owing to the limited capacities of their pipelines. Recently, rapid progress has been made by the methods adopting diffusion pipelines [5, 16, 17, 22, 40, 41]. With a stepwise noising-denoising design, the generated results of diffusion pipelines have far higher qualities than other self-supervised methods, especially on the datasets with huge intra-domain variance [4, 38]. In this paper, we follow a diffusion pipeline [34] to build up our pipeline to ensure the qualities and the varieties of the outputs.", + "bbox": [ + 89, + 90, + 480, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Tuning-based Personalized Portrait Synthesis", + "text_level": 1, + "bbox": [ + 89, + 238, + 475, + 255 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Tuning-based methods finetune models based on user-provided data to comprehend user identity information. Dreambooth [35] and its derivative versions [6, 30, 36] treat the entire T2I model trainable, and utilize certain special captions to represent the given identities. LoRA [18] and its variants [7, 12, 13, 26, 46, 48] represent another category of tuning-based methods. Unlike Dreambooth, LoRA greatly enhances the flexibility of the fine-tuning scheme by creating several trainable low-rank matrices within the T2I model along with much fewer parameters than the whole model. Nevertheless, a common issue with all the aforementioned methods is that they lack the ability to handle zero-shot cases. Each fine-tuned model matches only a single customized task, thus the fine-tuning incurs a high burden in terms of time, computational resources, and storage resources. In this work, our proposed HyperLoRA does not rely on online training of user-provided data, thereby circumventing the resource consumption problem brought about by the training process.", + "bbox": [ + 89, + 261, + 483, + 549 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Tuning-free Personalized Portrait Synthesis", + "text_level": 1, + "bbox": [ + 89, + 560, + 460, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Tuning-free methods, such as [25, 28, 29, 44, 45], pre-design a plug-in structure containing few parameters that can be inserted into the T2I base model. After pre-training and freezing such parameters on a large-scale dataset, tuning-free methods utilize such insertable structure to extract the identity information features from the given inputs, and inject them into the base model during inference. The online tuning is not necessary for these methods, thereby they can save resource costs and time. Concurrently, the design of the detachable parameters is also rather flexible. However, these methods merely learn the mapping relationship between the input image and the encoded tokens, which are utilized by a group of additional cross-attention modules, and the outputs are simply blended with the original ones. Our proposed HyperLoRA learns the mapping relationship between the input image and the network parameters, which is more essential. The implicit information contained in the network parameters is far more abundant and friendly to the base model than that in tokens. Compared with the methods mentioned above, our method has better adaptability, higher nonlinearity, and better fidelity.", + "bbox": [ + 89, + 583, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 89, + 604, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce the details of our method, overview of which is illustrated in Fig. 2. In Sec. 3.1, we introduce the low-dimensional linear LoRA space. In Sec. 3.2, we introduce the network structure. In Sec. 3.3, we introduce the designed multi-stage training process.", + "bbox": [ + 511, + 114, + 903, + 191 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Low-Dimensional Linear LoRA Space", + "text_level": 1, + "bbox": [ + 511, + 198, + 843, + 214 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The number of parameters in LoRA is typically large, which makes it difficult to generate the LoRA weights directly. Recognizing the interpolative nature of LoRA, we are inspired to exploit its linearity, and map the LoRA parameters onto a low-dimensional linear space. Note that our discussion below and in the rest of the paper is based on the implementation from [37], which is the most widely acknowledged way. Each layer of LoRA includes 8 matrices, which consist of the down and up matrices for query, key, value, and output. We project each matrix onto a $K$ -dimensional basis. Mark the number of LoRA layers as $N$ , then the overall degrees of freedom for the entire LoRA can be calculated as $N \\times 8K$ , which is significantly smaller than the origin parameter number in most cases. In our paper, we use $K = 128$ , and the number of parameters after projection becomes approximately only $1.2\\%$ (0.14M vs. 11.6M when the LoRA rank is 8) of the original one. To decouple the given specific identity information from the irrelevant features, we design to split a LoRA matrix into a base part and an ID part. We denote the basis of the ID part as $\\{\\mathbf{M}_{id}^{k}\\}_{k=1}^{K}$ , the combination coefficients as $\\{\\alpha_{k}\\}_{k=1}^{K}$ . Further, we denote the ID part as:", + "bbox": [ + 511, + 219, + 906, + 551 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} _ {i d} = \\sum_ {k = 1} ^ {K} \\alpha_ {k} \\cdot \\mathbf {M} _ {i d} ^ {k}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 558, + 903, + 599 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Similarly, the base part, denoted as $\\mathbf{M}_{\\mathrm{base}}$ , follows the same format as $\\mathbf{M}_{id}$ :", + "bbox": [ + 511, + 604, + 903, + 635 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} _ {\\text {b a s e}} = \\sum_ {k = 1} ^ {K} \\beta_ {k} \\cdot \\mathbf {M} _ {\\text {b a s e}} ^ {k}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 641, + 903, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Although the mathematical format of Base-LoRA is similar to ID-LoRA, it possesses a different significance. Specifically, we configure $\\mathbf{M}_{\\text {base }}$ to contain identity-irrelevant contents across all the training samples, such as the background and clothing, essentially capturing information from the input image that should not be included in $\\mathbf{M}_{id}$ . Explicitly modeling Base-LoRA can help ID-LoRA learn more essential identity-related information and ignore the irrelevant rest. In summary, we represent a HyperLoRA parameter matrix $\\mathbf{M}$ as:", + "bbox": [ + 511, + 688, + 906, + 839 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} = \\mathbf {M} _ {\\text {b a s e}} + \\mathbf {M} _ {i d} = \\sum_ {k = 1} ^ {K} \\beta_ {k} \\cdot \\mathbf {M} _ {\\text {b a s e}} ^ {k} + \\sum_ {k = 1} ^ {K} \\alpha_ {k} \\cdot \\mathbf {M} _ {i d} ^ {k}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 845, + 903, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13116", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2bf55c3c39ce1ba061b65e543361d922c182064a252efdbf9aa093ed14d4ac88.jpg", + "image_caption": [ + "Figure 2. Overview of HyperLoRA. We explicitly decompose the HyperLoRA into a Hyper ID-LoRA and a Hyper Base-LoRA. The former is designed to learn ID information while the latter is expected to fit others, e.g. background and clothing. Such a design helps to prevent irrelevant features leaking to ID-LoRA. During the training, we fix the weights of the pretrained SDXL base model and encoders, allowing only HyperLoRA modules updated by Backpropagation. At the inference stage, the Hyper ID-LoRA integrated into SDXL generates personalized images while the Hyper Base-LoRA is optional." + ], + "image_footnote": [], + "bbox": [ + 112, + 88, + 883, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "It is noteworthy that although the degrees of freedom for the LoRA parameters are significantly compressed, they still retain the capacity to represent identity information, as Fig. 3 shows. With our designed projection method and the setting of $K = 128$ , the generated results in the illustration of overfitting a face dataset suggest that, it is still capable of reconstructing the identity of the reference image for the compressed LoRA module without obvious degradation.", + "bbox": [ + 88, + 508, + 483, + 630 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Network Structure", + "text_level": 1, + "bbox": [ + 89, + 647, + 272, + 662 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we introduce our network structure, which is shown in Fig. 4. The whole network can be divided into several main components: the image encoder, the preceiver resampler and the LoRA combination.", + "bbox": [ + 88, + 671, + 482, + 731 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Image Encoder. The input image is passed into two image encoders, including a CLIP [33] ViT and an AntelopeV2 [20] encoder. The features extracted by the former are dense and contain strong structural information, while the features extracted by the latter are more abstract identity information. To explicitly distinguish the training processes of the base part and the id part, we deal with the input image of the CLIP in two different ways: for the training of $\\mathbf{M}_{\\text {base }}$ , we first crop the face region of an original image, and inside the cropped image, we blur the facial features and contour, preventing the base part reconstructing the specific face; for", + "bbox": [ + 88, + 734, + 482, + 902 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2b0c08ae050065e427bee9221844bc8915b4c93ae79d96636f63edc0fbc1d081.jpg", + "image_caption": [ + "Reference Image" + ], + "image_footnote": [], + "bbox": [ + 516, + 506, + 638, + 601 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/cd09f1168c4962caf4c722d24de9d8f94b578a06e93fbfdde8c50fe784b565b2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 653, + 506, + 774, + 601 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f828f3e3a907f8abd6595ffc4d6eb5b58fd6452ab80e8ac52ae22c7c21fef325.jpg", + "image_caption": [ + "Compressed LoRA" + ], + "image_footnote": [], + "bbox": [ + 653, + 602, + 774, + 696 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/201fcb3c3f2b820a5eca438797a83da2f9e1cedba3970649ac3e5af15279537c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 782, + 506, + 903, + 601 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c0fdb56353f9074afdb4f8f0fd3bfeea6de005c900c369bbc9c021f189bbc935.jpg", + "image_caption": [ + "Normal LoRA", + "Figure 3. The identity reconstruction ability on the low-dimensional linear LoRA space. In this example, we project LoRA parameters onto 128-dim basis and train on a face dataset (about 400K samples). Compared to normal LoRA, our compressed LoRA can also maintain the identity of the reference image well." + ], + "image_footnote": [], + "bbox": [ + 782, + 602, + 903, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the training of $\\mathbf{M}_{id}$ , we directly feed the cropped face image into CLIP ViT. Meanwhile, it also serves as the input of AntelopeV2 encoder to extract ID embedding. We utilize a simple ID Projector, consisting of a Linear layer followed by a Reshape and a LayerNorm [2], to transform the ID em", + "bbox": [ + 511, + 824, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13117", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/27560f4006184d5e5ab1f86f32f201f2d1baa3ac625da94aa89043686a14b0c7.jpg", + "image_caption": [ + "Figure 4. Network structure of HyperLoRA. We apply a perceiver resampler to convert the image features into a group of LoRA coefficients, thereby generating the whole LoRA by multiplied with LoRA basis. Two independent perceiver resamplers are instantiated for Hyper Base and ID LoRAs. Note that the second attention block (interacting with green tokens from ID Projector) is absent in Hyper Base-LoRA." + ], + "image_footnote": [], + "bbox": [ + 93, + 87, + 486, + 261 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "bedding into some ID tokens. The details of the approaches mentioned above can be found in Fig. 4.", + "bbox": [ + 89, + 396, + 482, + 425 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Perceiver Resampler. The perceiver resampler [1, 45] employs a transformer backbone and incorporates image tokens into the computation through an attention mechanism. Currently, we apply a 4-layer perceiver resampler, each layer of which consists of two attention modules interacting with the tokens from CLIP ViT and ID Projector. Note that we skip the second attention module when training $\\mathbf{M}_{\\text {base }}$ . We obtain the output of the perveiver resampler, the combination coefficients $\\{\\alpha_{k}\\}_{k=1}^{K}$ and $\\{\\beta_{k}\\}_{k=1}^{K}$ , which would be used to compute the output LoRA weights subsequently.", + "bbox": [ + 89, + 426, + 482, + 578 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "LoRA Combination. Both the Base-LoRA basis $\\{\\mathbf{M}_{base}^k\\}_{k = 1}^K$ and the ID-LoRA basis $\\{\\mathbf{M}_{id}^k\\}_{k = 1}^K$ are trainable variables. They are optimized during the training phase and fixed as constants during inference. Once the combination coefficients $\\{\\alpha_{k}\\}_{k = 1}^{K}$ and $\\{\\beta_k\\}_{k = 1}^K$ are predicted, the network can compute the entire LoRA by Eq. 3. It is noteworthy that the combination way can be adjusted flexibly to achieve various effects, for example, if the editability is a priority, the weight of $\\mathbf{M}_{base}$ in Eq. 3 can be reduced to drop the information unrelated to the input identity.", + "bbox": [ + 89, + 579, + 482, + 731 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Training", + "text_level": 1, + "bbox": [ + 89, + 741, + 194, + 757 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we introduce our multi-stage training process. Our approach begins with training the Base-LoRA as an initial warm-up. Subsequently, we expand the training to ID-LoRA with CLIP ViT. Finally, we fine-tune the model to enhance its identity preserving capability by using AntelopeV2 face encoder. For all stages, we employ the same loss term as in DDPM [16].", + "bbox": [ + 89, + 763, + 482, + 868 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Base-LoRA Warm-up. Training the ID-LoRA directly from an open domain base model is quite challenging, as", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "it tends to inadvertently incorporate extraneous details such as composition, background, and attire. Therefore, as described in Sec.3.1, we train non-facial information into the Base-LoRA, which serves as a warm-up in our training process. In this stage, we blur the facial information of the original input image before passing it through the CLIP ViT, and bypass the AntelopeV2 face encoder.", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ID-LoRA with CLIP. After the Base-LoRA warm-up, we add ID-LoRA into the training process. In this stage, we add the AntelopeV2 branch besides the existed CLIP branch, so that the abstract identity information can be efficiently extracted. We observe that training with CLIP features converges quickly but tends to overfit structural details, whereas training with ID embedding converges more slowly. Consequently, in this stage we stop training with CLIP features as long as it reaches the approximate convergence. To stabilize the training state in the early iterations, we initialize the output layers of the attention modules that are fused with ID embeddings to zero, allowing only the CLIP features to guide the early generation. To enable ID-LoRA to effectively represent the identity in the presence of trigger words (e.g., 3 rare words fcsks fxhks fhyks in this paper), while ensuring that the results remain unaffected in their absence, each iteration of this stage is set to one of the following situations: 1). The given prompt contains trigger words, both Base-LoRA and ID-LoRA are enabled, and the reconstruction target is the target image; 2). The given prompt does not contain trigger words, only ID-LoRA is enabled, and the resulting noise prediction should align with that of the original foundational model; 3). The given prompt does not contain trigger words, both Base-LoRA and ID-LoRA are enabled, and the resulting noise prediction should be consistent with that when only Base-LoRA is enabled. This approach minimizes disruption to the base model parameters and further concentrates the parameters on those related to identity information.", + "bbox": [ + 511, + 198, + 906, + 636 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ID-Embedding Fine-tune. After the previous training stages achieve rough convergence, we fix the weights of the CLIP-related attention modules, and continue to fine-tune the attention modules related to ID embedding (the second attention in Fig. 4). The training loss of this stage is exactly the same as in the second stage.", + "bbox": [ + 511, + 637, + 905, + 728 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 744, + 645, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 771, + 730, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To train our HyperLoRA model, we collect a subset from LAION-2B [39] which comprises 4.4 million portrait images. In detail, InsighFace [10] is leveraged to detect faces and we filter out those non-single portrait images. Meanwhile, we generate a caption for each image in our dataset using BLIP-2 [24] to enable Text-to-Image training. During the training, we randomly crop images around the face for", + "bbox": [ + 511, + 794, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13118", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/316e2f65c3194033d69414bc7dd2d0f17aff60d6f9ba146294bd148ed8e018ad.jpg", + "image_caption": [ + "Figure 5. Adapter and LoRA show different tolerances to CFG. The first row: images generated by InstantID where higher CFG leads to oversaturation. The second row: HyperLoRA always yield reasonable portrait images for CFG ranged from 3 to 7." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 483, + 258 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/65a4f2b97319008de7f57fe1f2dee1306698be87dc2700c952eef0949e38a204.jpg", + "image_caption": [ + "Figure 6. Inference with Base-LoRA improves face fidelity but limits the layout of generated images more similar to the cropped region for training." + ], + "image_footnote": [], + "bbox": [ + 91, + 338, + 483, + 460 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "data augmentation and focus more on the face region. Considering memory consumption, we currently set the LoRA rank of ID-LoRA and Base-LoRA to 8 and 4 respectively, which costs about 60G memory when batch size is 1. We further apply the 8-step gradient accumulation to simulate a larger batch size. We train the model based on SDXL-Base-1.0 [32] with AdamW [27] optimizer and learning rate 0.00001. In this 3-stage training process, we optimize the Base-LoRA, ID-LoRA (CLIP) and ID-LoRA (ID embedding) for 20K, 15K and 55K iterations in turn, and set the probability of three situations to 0.9, 0.05 and 0.05 for ID-LoRA training. We implement our method with PyTorch and Diffusers. The total training process is performed on 16 NVIDIA A100 GPUs for about 10 days.", + "bbox": [ + 88, + 537, + 482, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Settings", + "text_level": 1, + "bbox": [ + 89, + 757, + 189, + 773 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We build workflows in ComfyUI [3] to evaluate our HyperLoRA and three SOTA ID preservation methods, i.e. InstantID [44], IP-Adapter [45] (specifically, IP-Adapter-FaceID-Portrait) and PuLID [11]. To ensure a fair comparison, we apply the same seed, prompt and sampler across all methods. Note that we additionally concatenate the lora trigger words as a prefix to the prompt when testing HyperLoRA, and set different Classifier-Free Guidance [15] (CFG) scales", + "bbox": [ + 89, + 779, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "to generate reasonable images. As shown in Fig. 5, Adapter-based methods usually require a lower CFG scale while HyperLoRA allows a wide range of CFG scales. Although these methods are trained on SDXL-Base-1.0, they show sufficient generalization to other SDXL models. Thus, we use LEOSAM [23], an open-source SDXL model for high quality portraits, as the base model for evaluations, which typically improve image aesthetics. In quantitative comparison, we utilize a recent open-source test set, Unsplash-50 [8], which covers diverse genders, ages and races.", + "bbox": [ + 511, + 90, + 906, + 243 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2e0b22a210784a9038de6fb107de7198ef4b30ff340a6788ca557566aaeeeebe.jpg", + "table_caption": [ + "Table 1. Quantitative comparison." + ], + "table_footnote": [], + "table_body": "
FidelityEditability
CLIP-I↑ID Sim.↑CLIP-I↑CLIP-T↑
IP-Adapter0.7640.5660.7250.244
InstantID0.7340.6810.6880.237
PuLID0.7710.6130.8050.259
Arc2Face0.7860.643--
Ours (Full)0.8530.6780.7100.243
Ours (ID)0.8310.6250.7480.252
", + "bbox": [ + 514, + 277, + 901, + 406 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Comparisons", + "text_level": 1, + "bbox": [ + 511, + 425, + 653, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Comparison. We evaluate fidelity and editability as the quantitative metrics for comparison across these methods. For fidelity, we leverage the cosine of CLIP image features (CLIP-I) to measure face fidelity, and the cosine of ID embeddings (ID Sim.) to measure face ID similarity. We generate four images with prompt headshot of a girl/boy, high definition, 4k for each sample in our test set using each method, and compute the average of them. For editability, we utilize CLIP-I to measure the similarity of images generated with and without ID plug-in networks, and the cosine of CLIP image and text features (CLIP-T) to measure image-text similarity. We collect 10 prompts which cover hair, accessory, clothing and background, and evaluate all combinations of prompts and ID images. The specific CLIP model we used is CLIP-L/14 which is different from the text encoder in SDXL, and CurricularFace [19] (not used as ID encoder in all methods) is adopted to extract ID embeddings. We further incorporate Arc2Face [31], a human face base model, into comparison, which is trained using a fixed caption and a cropped region of faces, thus restricting its editability. Ours (Full) in Table. 1 means inference with both Base-LoRA (weight is set to 0.4) and ID-LoRA, while Ours (ID) represents only using ID-LoRA. As shown in Table. 1 and Fig. 6, Base-LoRA helps to improve fidelity but limits editability. Although our design decouples the image features into different LoRAs, it's hard to avoid leaking mutually. Thus, we can adjust the weight of Base-LoRA to adapt to different application scenarios. Our HyperLoRA (Full and ID) achieve the best and second-best face fidelity while InstantID shows superiority", + "bbox": [ + 511, + 446, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13119", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/901c67ee63d2f56aa722f5da50d18b1d1c62692bff126481dd3ab97700430a0f.jpg", + "image_caption": [ + "Figure 7. Qualitative comparison. From top to bottom, we generate images with the scene prompts: white shirt, wolf ears, wavy hair and brown sweater, and play with snowman." + ], + "image_footnote": [], + "bbox": [ + 107, + 84, + 890, + 570 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "in face ID similarity but lower face fidelity. Both these metrics should be considered together to evaluate fidelity, since the face ID similarity is more abstract and face fidelity reflects more details. This is evidenced in the following visual results in Fig. 7. In the aspect of editability, PuLID possesses a clear advantage, the key contribution of which is a well-designed loss function to reduce intrusion into the base model, thereby improving editability. Despite HyperLoRA (ID) inferior to PuLID, it surpasses other methods, meanwhile has good fidelity. It is noteworthy that there is a trade-off between fidelity and editability, and our HyperLoRA reaches a proper balance.", + "bbox": [ + 88, + 628, + 480, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Comparison. Fig. 7 shows the visual comparisons across IP-Adapter, InstantID, PuLID and our HyperLoRA (only ID-LoRA). The skin of portraits generated by IP-Adapter and InstantID has apparent AI-generated texture, which is a little oversaturation and far from photorealism. It is a common shortcoming of Adapter-based meth", + "bbox": [ + 88, + 810, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ods. PuLID improves this problem by weakening the intrusion to base model, outperforming IP-Adapter and InstantID but still suffering from blurring and lack of details. In contrast, LoRA directly modifies the base model weights instead of introducing extra attention modules, usually generating highly detailed and photorealistic images. Our HyperLoRA naturally inherits this advantage from LoRA. Furthermore, the manner in which LoRA injects ID features exhibits higher nonlinearity. Thus, our method achieves better face fidelity, capturing fine-grained details, such as pupil color. The layout of generated images of our HyperLoRA and PuLID is more similar compared with other methods, which is because of the higher CLIP-I and CLIP-T of editability. In comparison, IP-Adapter and InstantID have inferior editability, resulting in unreasonable layout (the 4th row of IP-Adapter) and failure to follow the prompt (the 3rd row of InstantID).", + "bbox": [ + 511, + 628, + 906, + 885 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13120", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a714d4633ffbd913aa075e9f87d52f83a1bd3837da9fd2a77eb96680e3bdad29.jpg", + "image_caption": [ + "Figure 8. Impact of Hyper Base-LoRA. The first row: the model trained with Hyper Base-LoRA generates correct clothing and background. The second row: the model trained without it takes many irrelevant features into the generated images." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 480, + 294 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1edf11bc1c76b8227732b87dc976de0be720b7142128ab48efa25c84d43aab43.jpg", + "image_caption": [ + "Figure 9. Comparison between different encoders. The first row: our CLIP + ID embeddings training strategy successfully generates glasses and mask. The second row: training with only CLIP fails to follow the prompts." + ], + "image_footnote": [], + "bbox": [ + 91, + 376, + 483, + 612 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 708, + 243, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "No Hyper Base-LoRA. In Fig. 8, we demonstrate the effectiveness of Hyper Base-LoRA. As aforementioned, our Hyper Base-LoRA helps to decouple ID information from input images. This is evidenced in the second row of Fig. 8, where the model trained without Hyper Base-LoRA fails to properly reflect the prompts dress, spacesuit and beach. As expected, the background and clothing are more editable if trained with Hyper Base-LoRA, since it prevents the ID-irrelevant features leaking to ID-LoRA.", + "bbox": [ + 89, + 731, + 482, + 868 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Only CLIP Encoder. Our multi-stage training strategy is crucial for successfully training ID-LoRA. Especially, we", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/85ae6a3b2bfecf5bf0980cf8aa94bd135eccbb32cbd9ff7e40abb3f44c522171.jpg", + "image_caption": [ + "Figure 10. Multiple input images contribute to a higher ID similarity and a more realistic and aesthetic synthesized portrait." + ], + "image_footnote": [], + "bbox": [ + 514, + 85, + 905, + 224 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "leverage CLIP ViT features only during the early iterations of training and then switch to ID embeddings to learn more details. Fig. 9 shows that the problem arises when only adopting CLIP as encoder, where we see that the facial elements are challenging to modify. In contrast, predominantly learning facial features with ID embeddings relaxes the structural constraints on the face.", + "bbox": [ + 511, + 292, + 906, + 397 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Multiple Input Images. Thanks to the natural interpolability of LoRA and our low-dimensional linear LoRA space (as discussed in Sec. 3.1), it is straightforward to support portrait synthesis with multiple input images by averaging the LoRA coefficients generated from each single image. Fig. 10 illustrates that multiple ID images help HyperLoRA to identify more stable and pure ID features, thereby improving both ID fidelity and image quality.", + "bbox": [ + 511, + 398, + 906, + 521 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 541, + 633, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have proposed HyperLoRA, the first zero-shot personalized portrait generation method based on LoRA, which is trained in an end-to-end manner. With the advantage of LoRA, our method modifies the base model weights to inject ID information rather than introducing new cross attention as adapter-based methods do, exhibiting higher nonlinearity, thereby yielding more detailed and realistic portrait images. To make HyperLoRA training feasible, we represent the whole LoRA parameters with a linear combination of a group of LoRA basis, which reduces the number of HyperLoRA parameters magnificently. We further decompose the LoRA into two parts explicitly. With our well-designed training strategy, the ID part tends to learn facial features while other aspects, like background and clothing, are fitted into the base part, thereby reducing the risk of overfitting and improving editability. It's recognized that only a rank 8 HyperLoRA is allowed to train when considering the memory limitation of current GPU. Nevertheless, rank 8 is still significantly smaller than what is frequently used in normal LoRA. An advanced LoRA parameters compression method, such as LoRA autoencoder, may enable larger rank HypeLoRA training, unlocking more powerful capability.", + "bbox": [ + 511, + 568, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13121", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 5", + "[2] Jimmy Lei Ba. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 4", + "[3] comfyanonymous. Comfyui: The most powerful and modular diffusion model gui and backend. 6", + "[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 3", + "[5] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 3", + "[6] Zhongyi Fan, Zixin Yin, Gang Li, Yibing Zhan, and Heliang Zheng. Dreambooth++: Boosting subject-driven generation via region-level references packing. In ACM Multimedia 2024, 2024. 3", + "[7] Rinon Gal, Moab Arar, Yuval Atzmon, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Encoder-based domain tuning for fast personalization of text-to-image models. ACM Transactions on Graphics (TOG), 42(4):1-13, 2023. 3", + "[8] Rinon Gal, Or Lichter, Elad Richardson, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Lcmlookahead for encoder-based text-to-image personalization. arXiv preprint arXiv:2404.03620, 2024. 6", + "[9] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2", + "[10] Jia Guo and Jiankang Deng. Insightface: 2d and 3d face analysis project. 5", + "[11] Zinan Guo, Yanze Wu, Zhuowei Chen, Lang Chen, and Qian He. Pulid: Pure and lightning id customization via contrastive alignment. Advances in neural information processing systems, 2024. 6", + "[12] Ligong Han, Yinxiao Li, Han Zhang, Peyman Milanfar, Dimitris Metaxas, and Feng Yang. Svdiff: Compact parameter space for diffusion fine-tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7323-7334, 2023. 3", + "[13] Soufiane Hayou, Nikhil Ghosh, and Bin Yu. Lora+: Efficient low rank adaptation of large models. In *Forty-first International Conference on Machine Learning*, ICML 2024, Vienna, Austria, July 21-27, 2024, 2024. 3", + "[14] Zecheng He, Bo Sun, Felix Juefei-Xu, Haoyu Ma, Ankit Ramchandani, Vincent Cheung, Siddharth Shah, Anmol Kalia, Harihar Subramanyam, Alireza Zareian, et al. Imagine yourself: Tuning-free personalized image generation. arXiv preprint arXiv:2409.13346, 2024. 1", + "[15] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 6" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 3, 5", + "[17] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. *J. Mach. Learn.* Res., 23(47):1-33, 2022. 3", + "[18] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022, 2022. 2, 3", + "[19] Yuge Huang, Yuhan Wang, Ying Tai, Xiaoming Liu, Pengcheng Shen, Shaoxin Li, Jilin Li, and Feiyue Huang. Curricularface: adaptive curriculum learning loss for deep face recognition. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5901-5910, 2020. 6", + "[20] Insightface. Antelopev2, 2023. 2, 4", + "[21] Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up gans for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10124-10134, 2023. 2", + "[22] Diederik Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. Advances in neural information processing systems, 34:21696-21707, 2021. 3", + "[23] LEOSAM. Leosam's helloworld xl, 2023. 6", + "[24] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 5", + "[25] Zhen Li, Mingdeng Cao, Xintao Wang, Zhongang Qi, Ming-Ming Cheng, and Ying Shan. Photomaker: Customizing realistic human photos via stacked id embedding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8640-8650, 2024. 3", + "[26] Shih-Yang Liu, Chien-Yi Wang, Hongxu Yin, Pavlo Molchanov, Yu-Chiang Frank Wang, Kwang-Ting Cheng, and Min-Hung Chen. Dora: Weight-decomposed low-rank adaptation. In *Forty-first International Conference on Machine Learning*, ICML 2024, Vienna, Austria, July 21-27, 2024, 2024. 3", + "[27] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6", + "[28] Jian Ma, Junhao Liang, Chen Chen, and Haonan Lu. Subject-diffusion: Open domain personalized text-to-image generation without test-time fine-tuning. In ACM SIGGRAPH 2024 Conference Papers, pages 1-12, 2024. 3", + "[29] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4296–4304, 2024. 3" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13122", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Lianyu Pang, Jian Yin, Baoquan Zhao, Feize Wu, Fu Lee Wang, Qing Li, and Xudong Mao. Attndreambooth: Towards text-aligned personalized text-to-image generation. arXiv preprint arXiv:2406.05000, 2024. 3", + "[31] Foivos Paraperas Papantoniou, Alexandros Lattas, Stylianos Moschoglou, Jiankang Deng, Bernhard Kainz, and Stefanos Zafeiriou. Arc2face: A foundation model for id-consistent human faces. In European Conference on Computer Vision, pages 241-261. Springer, 2024. 6", + "[32] Dustin Podell, *Zion English*, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. *Sdxl: Improving latent diffusion models for high-resolution image synthesis.* arXiv preprint arXiv:2307.01952, 2023. 6", + "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. 2021. 2, 4", + "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 3", + "[35] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. 2022. 2, 3", + "[36] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdreambooth: Hypernetworks for fast personalization of text-to-image models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6527-6536, 2024. 3", + "[37] Simo Ryu. Low-rank adaptation for fast text-to-image diffusion fine-tuning. 3", + "[38] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 3", + "[39] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278–25294, 2022. 5", + "[40] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 3", + "[41] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 3", + "[42] Ming Tao, Hao Tang, Fei Wu, Xiao-Yuan Jing, Bing-Kun Bao, and Changsheng Xu. Df-gan: A simple and effec" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tive baseline for text-to-image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16515-16525, 2022. 2", + "[43] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 2", + "[44] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, and Anthony Chen. Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519, 2024. 1, 3, 6", + "[45] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. 2023. 1, 2, 3, 5, 6", + "[46] Shih-Ying Yeh, Yu-Guan Hsieh, Zhidong Gao, Bernard BW Yang, Giyeong Oh, and Yanmin Gong. Navigating text-to-image customization: From lycoris fine-tuning to model evaluation. In The Twelfth International Conference on Learning Representations, 2023. 3", + "[47] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2(3):5, 2022. 2", + "[48] Qingru Zhang, Minshuo Chen, Alexander Bukharin, Pengcheng He, Yu Cheng, Weizhu Chen, and Tuo Zhao. Adaptive budget allocation for parameter-efficient finetuning. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023, 2023. 3" + ], + "bbox": [ + 516, + 92, + 906, + 484 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13123", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/HyperLoRA_ Parameter-Efficient Adaptive Generation for Portrait Synthesis/0dcd811d-bb8b-4fcc-9377-9711c569d3e6_model.json b/2025/HyperLoRA_ Parameter-Efficient Adaptive Generation for Portrait Synthesis/0dcd811d-bb8b-4fcc-9377-9711c569d3e6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..88e9e2a0bc45928a37a8808033aff0280dec4c2a --- /dev/null +++ b/2025/HyperLoRA_ Parameter-Efficient Adaptive Generation for Portrait Synthesis/0dcd811d-bb8b-4fcc-9377-9711c569d3e6_model.json @@ -0,0 +1,2068 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.11, + 0.131, + 0.89, + 0.155 + ], + "angle": 0, + "content": "HyperLoRA: Parameter-Efficient Adaptive Generation for Portrait Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.156, + 0.18, + 0.266, + 0.2 + ], + "angle": 0, + "content": "Mengtian Li*" + }, + { + "type": "text", + "bbox": [ + 0.305, + 0.181, + 0.412, + 0.199 + ], + "angle": 0, + "content": "Jinshu Chen*" + }, + { + "type": "text", + "bbox": [ + 0.454, + 0.181, + 0.587, + 0.199 + ], + "angle": 0, + "content": "Wanquan Feng*†" + }, + { + "type": "text", + "bbox": [ + 0.388, + 0.2, + 0.501, + 0.217 + ], + "angle": 0, + "content": "Songtao Zhao" + }, + { + "type": "text", + "bbox": [ + 0.541, + 0.2, + 0.611, + 0.217 + ], + "angle": 0, + "content": "Qian He" + }, + { + "type": "text", + "bbox": [ + 0.369, + 0.223, + 0.624, + 0.241 + ], + "angle": 0, + "content": "Intelligent Creation, ByteDance" + }, + { + "type": "text", + "bbox": [ + 0.628, + 0.182, + 0.74, + 0.199 + ], + "angle": 0, + "content": "Bingchuan Li" + }, + { + "type": "text", + "bbox": [ + 0.778, + 0.182, + 0.841, + 0.198 + ], + "angle": 0, + "content": "Fei Dai" + }, + { + "type": "image", + "bbox": [ + 0.12, + 0.274, + 0.332, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.162, + 0.535, + 0.259, + 0.552 + ], + "angle": 0, + "content": "HyperLoRA" + }, + { + "type": "image_caption", + "bbox": [ + 0.16, + 0.555, + 0.261, + 0.569 + ], + "angle": 0, + "content": "w.o. fine-tuning" + }, + { + "type": "image", + "bbox": [ + 0.344, + 0.275, + 0.519, + 0.411 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.385, + 0.414, + 0.48, + 0.429 + ], + "angle": 0, + "content": "Wizard hat" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.275, + 0.698, + 0.411 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.551, + 0.414, + 0.668, + 0.431 + ], + "angle": 0, + "content": "Pink wavy hair" + }, + { + "type": "image", + "bbox": [ + 0.702, + 0.275, + 0.878, + 0.411 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.722, + 0.414, + 0.856, + 0.431 + ], + "angle": 0, + "content": "Chinese clothing" + }, + { + "type": "image", + "bbox": [ + 0.344, + 0.436, + 0.52, + 0.572 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.411, + 0.576, + 0.454, + 0.589 + ], + "angle": 0, + "content": "Snow" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.436, + 0.698, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.586, + 0.575, + 0.636, + 0.589 + ], + "angle": 0, + "content": "Beach" + }, + { + "type": "image", + "bbox": [ + 0.701, + 0.436, + 0.878, + 0.572 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.747, + 0.574, + 0.834, + 0.591 + ], + "angle": 0, + "content": "Cyberpunk" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.602, + 0.908, + 0.645 + ], + "angle": 0, + "content": "Figure 1. We propose HyperLoRA, a parameter-efficient adaptive method for portrait synthesis. Given an input face image, HyperLoRA generates personalized LoRA weights without online fine-tuning. Due to the natural interpolability of LoRA, it is easy to support multiple inputs by simple averaging. Leveraging the generated LoRA, we can create personalized portraits with high photorealism and fidelity." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.655, + 0.327, + 0.671 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.689, + 0.486, + 0.871 + ], + "angle": 0, + "content": "Personalized portrait synthesis, essential in domains like social entertainment, has recently made significant progress. Person-wise fine-tuning based methods, such as LoRA and DreamBooth, can produce photorealistic outputs but need training on individual samples, consuming time and resources and posing an unstable risk. Adapter based techniques such as IP-Adapter freeze the foundational model parameters and employ a plug-in architecture to enable zero-shot inference, but they often exhibit a lack of naturalness and authenticity, which are not to be overlooked in portrait synthesis tasks. In this paper, we introduce a parameter-efficient adaptive generation method," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.657, + 0.909, + 0.765 + ], + "angle": 0, + "content": "namely HyperLoRA, that uses an adaptive plug-in network to generate LoRA weights, merging the superior performance of LoRA with the zero-shot capability of adapter scheme. Through our carefully designed network structure and training strategy, we achieve zero-shot personalized portrait generation (supporting both single and multiple image inputs) with high photorealism, fidelity, and editability." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.785, + 0.646, + 0.801 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.903 + ], + "angle": 0, + "content": "The personalized portrait synthesis task [14, 44, 45] aims to generate images sharing the same identity with what is contained in the given single or multiple portrait images. By feeding the identity information into specific generative model, the task enables the users to create diverse, high-quality and creative personalized portraits, thus hold" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.375, + 0.901 + ], + "angle": 0, + "content": "* Equal contribution † Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13114" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.121 + ], + "angle": 0, + "content": "ing substantial importance across various domains, such as entertainment and social media." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.124, + 0.482, + 0.29 + ], + "angle": 0, + "content": "One key challenge of the personalized portrait synthesis task is the fidelity: the expected generated images should retain the given identity to the greatest extent and avoid irrelevant outputs. Another challenge lies in editability: while preserving the given identity, the portrait synthesis model should maintain the abilities of flexibly editing the other contents of the generated images, for example, the composition of the image, the positioning and the clothing of the character. Additionally, the inference speed is crucial, as fast inference is friendly to not only the computational costs but also the user experience." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.293, + 0.485, + 0.504 + ], + "angle": 0, + "content": "Considering the challenges mentioned above, most of the current solutions have different shortcomings. Existing personalized image synthesis methods can generally be divided into tuning-based and tuning-free approaches. Tuning-based methods, such as LoRA (Low-Rank Adaptation) [18] and DreamBooth [35], update model parameters based on specific user-provided data, which can accurately comprehend and preserve the input identity information. Theoretically, these methods are mainly designed for few-shot cases, and customize different models for each groups of provided data with different identities. Thus, disadvantages like high computing costs, unstable training, and lengthy training processes make such methods bloated when dealing with various provided identities." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.507, + 0.484, + 0.763 + ], + "angle": 0, + "content": "Tuning-free methods, such as IP-Adapter [45], typically utilize an additional plug-in structure to encode user identity information. By feeding such identity-related tokens into the frozen foundational model, these methods make influences on the inference process, and ultimately yielding results that retain the given identities. Such additional adapter struture is trained with large portrait datasets in advance, and freezing when inference. Therefore, these methods get rid of the online training, and are convenient for the actual uses. However, for these methods, the interaction between encoded identity information and the foundational model relies on a group of fixed extra cross-attention modules, which might degrade the quality of generated images, leading to unnatural and non-photorealistic facial features, and fail to recover fine-grained details. In summary, the previous methods fail to simultaneously strike a balance between the fidelity, the editability, and the speed." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.482, + 0.901 + ], + "angle": 0, + "content": "In this paper, we introduce a parameter-efficient adaptive generation method, namely HyperLoRA, which achieves great identity fidelity, high editability and fast inference speed. HyperLoRA uses an adaptive plug-in network to encode identity images, which follows the design of tuning-free methods like IP-Adapter. However, unlike such methods which utilize the encoded tokens for the attention computation, we directly generate the LoRA weights with the plug-in network. This strategy not only maintains the zero" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "shot capability of the adapter strategy, but also inherits the representation and high-quality generation effect of the LoRA strategy which adjusts the global parameters without real fine-tuning process." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.153, + 0.905, + 0.53 + ], + "angle": 0, + "content": "Considering the substantial number of parameters in LoRA, to reduce the training difficulty, we initially project the LoRA parameters onto a linear LoRA parameter space, rather than directly training in the original parameter space. In this way, the generated LoRA weights can be represented as a linear combination of a trainable basis. We employ a perceiver resampler (as in IP-Adapter [45]) to convert the input identity images into combination coefficients for the basis. To encode the input images, we utilize a CLIP ViT [33] and an AntelopeV2 encoder [20] to extract pixel-level and identity-sensitive features. These feature extractors mentioned above stay frozen during all the training process. To avoid overfitting, we decompose LoRA at the parameter level into an ID-LoRA part and a Base-LoRA part, thereby the ID and ID-irreverent information are fitted into different parts correspondingly. The whole training pipeline runs in a multi-stage way: only the Base-LoRA part is trained in the first stage as a warm-up start; and in the second stage, we train the ID-LoRA with CLIP ViT features; in the final stage, we only finetune the attention modules related to ID embeddings. Besides the case where only a single image is given, our proposed method can also handle cases that there are multiple inputs by simply interpolating the predicted LoRA weights. Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.533, + 0.904, + 0.591 + ], + "angle": 0, + "content": "- We propose HyperLoRA, a parameter-efficient adaptive method for the personalized portrait synthesis task, merging the superior performance of the LoRA strategy with the zero-shot capability of the adapter strategy." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.593, + 0.904, + 0.637 + ], + "angle": 0, + "content": "- We construct a low-dimension linear LoRA space, which reduces the training difficulty of the plug-in network while ensuring the identity preservation." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.638, + 0.904, + 0.697 + ], + "angle": 0, + "content": "- We design a strategy that decomposes LoRA weights at the parameter level into a base part and an ID part, thereby effectively decoupling the identity information restored in the given images and significantly reducing overfitting." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.699, + 0.904, + 0.742 + ], + "angle": 0, + "content": "- We conduct various experiments to demonstrate the capacities of our framework. Experimental results show that our method achieves both high fidelity and editability." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.533, + 0.904, + 0.742 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.761, + 0.655, + 0.777 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.787, + 0.734, + 0.803 + ], + "angle": 0, + "content": "2.1. Text-to-Image Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.81, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Text-to-image (T2I) synthesis has been rapidly developed for several years. The task aims to generate high-quality images which are highly correlated with given texts or instructions. Some early works [21, 42, 47] employ transformer [43] or GAN [9] architectures to solve this task, while the performance of the solutions is not that satisfy" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13115" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.228 + ], + "angle": 0, + "content": "ing owing to the limited capacities of their pipelines. Recently, rapid progress has been made by the methods adopting diffusion pipelines [5, 16, 17, 22, 40, 41]. With a stepwise noising-denoising design, the generated results of diffusion pipelines have far higher qualities than other self-supervised methods, especially on the datasets with huge intra-domain variance [4, 38]. In this paper, we follow a diffusion pipeline [34] to build up our pipeline to ensure the qualities and the varieties of the outputs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.239, + 0.476, + 0.256 + ], + "angle": 0, + "content": "2.2. Tuning-based Personalized Portrait Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.262, + 0.485, + 0.55 + ], + "angle": 0, + "content": "Tuning-based methods finetune models based on user-provided data to comprehend user identity information. Dreambooth [35] and its derivative versions [6, 30, 36] treat the entire T2I model trainable, and utilize certain special captions to represent the given identities. LoRA [18] and its variants [7, 12, 13, 26, 46, 48] represent another category of tuning-based methods. Unlike Dreambooth, LoRA greatly enhances the flexibility of the fine-tuning scheme by creating several trainable low-rank matrices within the T2I model along with much fewer parameters than the whole model. Nevertheless, a common issue with all the aforementioned methods is that they lack the ability to handle zero-shot cases. Each fine-tuned model matches only a single customized task, thus the fine-tuning incurs a high burden in terms of time, computational resources, and storage resources. In this work, our proposed HyperLoRA does not rely on online training of user-provided data, thereby circumventing the resource consumption problem brought about by the training process." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.561, + 0.462, + 0.577 + ], + "angle": 0, + "content": "2.3. Tuning-free Personalized Portrait Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Tuning-free methods, such as [25, 28, 29, 44, 45], pre-design a plug-in structure containing few parameters that can be inserted into the T2I base model. After pre-training and freezing such parameters on a large-scale dataset, tuning-free methods utilize such insertable structure to extract the identity information features from the given inputs, and inject them into the base model during inference. The online tuning is not necessary for these methods, thereby they can save resource costs and time. Concurrently, the design of the detachable parameters is also rather flexible. However, these methods merely learn the mapping relationship between the input image and the encoded tokens, which are utilized by a group of additional cross-attention modules, and the outputs are simply blended with the original ones. Our proposed HyperLoRA learns the mapping relationship between the input image and the network parameters, which is more essential. The implicit information contained in the network parameters is far more abundant and friendly to the base model than that in tokens. Compared with the methods mentioned above, our method has better adaptability, higher nonlinearity, and better fidelity." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.605, + 0.105 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.115, + 0.905, + 0.192 + ], + "angle": 0, + "content": "In this section, we introduce the details of our method, overview of which is illustrated in Fig. 2. In Sec. 3.1, we introduce the low-dimensional linear LoRA space. In Sec. 3.2, we introduce the network structure. In Sec. 3.3, we introduce the designed multi-stage training process." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.199, + 0.844, + 0.215 + ], + "angle": 0, + "content": "3.1. Low-Dimensional Linear LoRA Space" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.22, + 0.907, + 0.553 + ], + "angle": 0, + "content": "The number of parameters in LoRA is typically large, which makes it difficult to generate the LoRA weights directly. Recognizing the interpolative nature of LoRA, we are inspired to exploit its linearity, and map the LoRA parameters onto a low-dimensional linear space. Note that our discussion below and in the rest of the paper is based on the implementation from [37], which is the most widely acknowledged way. Each layer of LoRA includes 8 matrices, which consist of the down and up matrices for query, key, value, and output. We project each matrix onto a \\(K\\)-dimensional basis. Mark the number of LoRA layers as \\(N\\), then the overall degrees of freedom for the entire LoRA can be calculated as \\(N \\times 8K\\), which is significantly smaller than the origin parameter number in most cases. In our paper, we use \\(K = 128\\), and the number of parameters after projection becomes approximately only \\(1.2\\%\\) (0.14M vs. 11.6M when the LoRA rank is 8) of the original one. To decouple the given specific identity information from the irrelevant features, we design to split a LoRA matrix into a base part and an ID part. We denote the basis of the ID part as \\(\\{\\mathbf{M}_{id}^{k}\\}_{k=1}^{K}\\), the combination coefficients as \\(\\{\\alpha_{k}\\}_{k=1}^{K}\\). Further, we denote the ID part as:" + }, + { + "type": "equation", + "bbox": [ + 0.635, + 0.559, + 0.905, + 0.601 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} _ {i d} = \\sum_ {k = 1} ^ {K} \\alpha_ {k} \\cdot \\mathbf {M} _ {i d} ^ {k}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.606, + 0.905, + 0.636 + ], + "angle": 0, + "content": "Similarly, the base part, denoted as \\(\\mathbf{M}_{\\mathrm{base}}\\), follows the same format as \\(\\mathbf{M}_{id}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.622, + 0.642, + 0.905, + 0.683 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} _ {\\text {b a s e}} = \\sum_ {k = 1} ^ {K} \\beta_ {k} \\cdot \\mathbf {M} _ {\\text {b a s e}} ^ {k}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.84 + ], + "angle": 0, + "content": "Although the mathematical format of Base-LoRA is similar to ID-LoRA, it possesses a different significance. Specifically, we configure \\(\\mathbf{M}_{\\text {base }}\\) to contain identity-irrelevant contents across all the training samples, such as the background and clothing, essentially capturing information from the input image that should not be included in \\(\\mathbf{M}_{id}\\). Explicitly modeling Base-LoRA can help ID-LoRA learn more essential identity-related information and ignore the irrelevant rest. In summary, we represent a HyperLoRA parameter matrix \\(\\mathbf{M}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.523, + 0.846, + 0.905, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} = \\mathbf {M} _ {\\text {b a s e}} + \\mathbf {M} _ {i d} = \\sum_ {k = 1} ^ {K} \\beta_ {k} \\cdot \\mathbf {M} _ {\\text {b a s e}} ^ {k} + \\sum_ {k = 1} ^ {K} \\alpha_ {k} \\cdot \\mathbf {M} _ {i d} ^ {k}. \\tag {3}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13116" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.089, + 0.885, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.413, + 0.908, + 0.486 + ], + "angle": 0, + "content": "Figure 2. Overview of HyperLoRA. We explicitly decompose the HyperLoRA into a Hyper ID-LoRA and a Hyper Base-LoRA. The former is designed to learn ID information while the latter is expected to fit others, e.g. background and clothing. Such a design helps to prevent irrelevant features leaking to ID-LoRA. During the training, we fix the weights of the pretrained SDXL base model and encoders, allowing only HyperLoRA modules updated by Backpropagation. At the inference stage, the Hyper ID-LoRA integrated into SDXL generates personalized images while the Hyper Base-LoRA is optional." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.51, + 0.484, + 0.631 + ], + "angle": 0, + "content": "It is noteworthy that although the degrees of freedom for the LoRA parameters are significantly compressed, they still retain the capacity to represent identity information, as Fig. 3 shows. With our designed projection method and the setting of \\( K = 128 \\), the generated results in the illustration of overfitting a face dataset suggest that, it is still capable of reconstructing the identity of the reference image for the compressed LoRA module without obvious degradation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.648, + 0.273, + 0.663 + ], + "angle": 0, + "content": "3.2. Network Structure" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.672, + 0.483, + 0.732 + ], + "angle": 0, + "content": "In this section, we introduce our network structure, which is shown in Fig. 4. The whole network can be divided into several main components: the image encoder, the preceiver resampler and the LoRA combination." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.735, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Image Encoder. The input image is passed into two image encoders, including a CLIP [33] ViT and an AntelopeV2 [20] encoder. The features extracted by the former are dense and contain strong structural information, while the features extracted by the latter are more abstract identity information. To explicitly distinguish the training processes of the base part and the id part, we deal with the input image of the CLIP in two different ways: for the training of \\(\\mathbf{M}_{\\text {base }}\\), we first crop the face region of an original image, and inside the cropped image, we blur the facial features and contour, preventing the base part reconstructing the specific face; for" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.507, + 0.639, + 0.602 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.541, + 0.603, + 0.612, + 0.611 + ], + "angle": 0, + "content": "Reference Image" + }, + { + "type": "image", + "bbox": [ + 0.654, + 0.507, + 0.776, + 0.602 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.654, + 0.603, + 0.776, + 0.697 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.674, + 0.698, + 0.755, + 0.706 + ], + "angle": 0, + "content": "Compressed LoRA" + }, + { + "type": "image", + "bbox": [ + 0.784, + 0.507, + 0.905, + 0.602 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.784, + 0.603, + 0.905, + 0.697 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.817, + 0.698, + 0.875, + 0.706 + ], + "angle": 0, + "content": "Normal LoRA" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.718, + 0.907, + 0.789 + ], + "angle": 0, + "content": "Figure 3. The identity reconstruction ability on the low-dimensional linear LoRA space. In this example, we project LoRA parameters onto 128-dim basis and train on a face dataset (about 400K samples). Compared to normal LoRA, our compressed LoRA can also maintain the identity of the reference image well." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.907, + 0.903 + ], + "angle": 0, + "content": "the training of \\(\\mathbf{M}_{id}\\), we directly feed the cropped face image into CLIP ViT. Meanwhile, it also serves as the input of AntelopeV2 encoder to extract ID embedding. We utilize a simple ID Projector, consisting of a Linear layer followed by a Reshape and a LayerNorm [2], to transform the ID em" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13117" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.088, + 0.488, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.271, + 0.486, + 0.37 + ], + "angle": 0, + "content": "Figure 4. Network structure of HyperLoRA. We apply a perceiver resampler to convert the image features into a group of LoRA coefficients, thereby generating the whole LoRA by multiplied with LoRA basis. Two independent perceiver resamplers are instantiated for Hyper Base and ID LoRAs. Note that the second attention block (interacting with green tokens from ID Projector) is absent in Hyper Base-LoRA." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.397, + 0.483, + 0.426 + ], + "angle": 0, + "content": "bedding into some ID tokens. The details of the approaches mentioned above can be found in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.428, + 0.483, + 0.579 + ], + "angle": 0, + "content": "Perceiver Resampler. The perceiver resampler [1, 45] employs a transformer backbone and incorporates image tokens into the computation through an attention mechanism. Currently, we apply a 4-layer perceiver resampler, each layer of which consists of two attention modules interacting with the tokens from CLIP ViT and ID Projector. Note that we skip the second attention module when training \\(\\mathbf{M}_{\\text {base }}\\). We obtain the output of the perveiver resampler, the combination coefficients \\(\\{\\alpha_{k}\\}_{k=1}^{K}\\) and \\(\\{\\beta_{k}\\}_{k=1}^{K}\\), which would be used to compute the output LoRA weights subsequently." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.58, + 0.483, + 0.732 + ], + "angle": 0, + "content": "LoRA Combination. Both the Base-LoRA basis \\(\\{\\mathbf{M}_{base}^k\\}_{k = 1}^K\\) and the ID-LoRA basis \\(\\{\\mathbf{M}_{id}^k\\}_{k = 1}^K\\) are trainable variables. They are optimized during the training phase and fixed as constants during inference. Once the combination coefficients \\(\\{\\alpha_{k}\\}_{k = 1}^{K}\\) and \\(\\{\\beta_k\\}_{k = 1}^K\\) are predicted, the network can compute the entire LoRA by Eq. 3. It is noteworthy that the combination way can be adjusted flexibly to achieve various effects, for example, if the editability is a priority, the weight of \\(\\mathbf{M}_{base}\\) in Eq. 3 can be reduced to drop the information unrelated to the input identity." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.742, + 0.196, + 0.758 + ], + "angle": 0, + "content": "3.3. Training" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.869 + ], + "angle": 0, + "content": "In this section, we introduce our multi-stage training process. Our approach begins with training the Base-LoRA as an initial warm-up. Subsequently, we expand the training to ID-LoRA with CLIP ViT. Finally, we fine-tune the model to enhance its identity preserving capability by using AntelopeV2 face encoder. For all stages, we employ the same loss term as in DDPM [16]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Base-LoRA Warm-up. Training the ID-LoRA directly from an open domain base model is quite challenging, as" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "it tends to inadvertently incorporate extraneous details such as composition, background, and attire. Therefore, as described in Sec.3.1, we train non-facial information into the Base-LoRA, which serves as a warm-up in our training process. In this stage, we blur the facial information of the original input image before passing it through the CLIP ViT, and bypass the AntelopeV2 face encoder." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.199, + 0.907, + 0.637 + ], + "angle": 0, + "content": "ID-LoRA with CLIP. After the Base-LoRA warm-up, we add ID-LoRA into the training process. In this stage, we add the AntelopeV2 branch besides the existed CLIP branch, so that the abstract identity information can be efficiently extracted. We observe that training with CLIP features converges quickly but tends to overfit structural details, whereas training with ID embedding converges more slowly. Consequently, in this stage we stop training with CLIP features as long as it reaches the approximate convergence. To stabilize the training state in the early iterations, we initialize the output layers of the attention modules that are fused with ID embeddings to zero, allowing only the CLIP features to guide the early generation. To enable ID-LoRA to effectively represent the identity in the presence of trigger words (e.g., 3 rare words fcsks fxhks fhyks in this paper), while ensuring that the results remain unaffected in their absence, each iteration of this stage is set to one of the following situations: 1). The given prompt contains trigger words, both Base-LoRA and ID-LoRA are enabled, and the reconstruction target is the target image; 2). The given prompt does not contain trigger words, only ID-LoRA is enabled, and the resulting noise prediction should align with that of the original foundational model; 3). The given prompt does not contain trigger words, both Base-LoRA and ID-LoRA are enabled, and the resulting noise prediction should be consistent with that when only Base-LoRA is enabled. This approach minimizes disruption to the base model parameters and further concentrates the parameters on those related to identity information." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.638, + 0.906, + 0.729 + ], + "angle": 0, + "content": "ID-Embedding Fine-tune. After the previous training stages achieve rough convergence, we fix the weights of the CLIP-related attention modules, and continue to fine-tune the attention modules related to ID embedding (the second attention in Fig. 4). The training loss of this stage is exactly the same as in the second stage." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.746, + 0.646, + 0.763 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.772, + 0.731, + 0.789 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.901 + ], + "angle": 0, + "content": "To train our HyperLoRA model, we collect a subset from LAION-2B [39] which comprises 4.4 million portrait images. In detail, InsighFace [10] is leveraged to detect faces and we filter out those non-single portrait images. Meanwhile, we generate a caption for each image in our dataset using BLIP-2 [24] to enable Text-to-Image training. During the training, we randomly crop images around the face for" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13118" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.089, + 0.485, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.269, + 0.483, + 0.327 + ], + "angle": 0, + "content": "Figure 5. Adapter and LoRA show different tolerances to CFG. The first row: images generated by InstantID where higher CFG leads to oversaturation. The second row: HyperLoRA always yield reasonable portrait images for CFG ranged from 3 to 7." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.339, + 0.485, + 0.461 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.469, + 0.483, + 0.514 + ], + "angle": 0, + "content": "Figure 6. Inference with Base-LoRA improves face fidelity but limits the layout of generated images more similar to the cropped region for training." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.538, + 0.483, + 0.75 + ], + "angle": 0, + "content": "data augmentation and focus more on the face region. Considering memory consumption, we currently set the LoRA rank of ID-LoRA and Base-LoRA to 8 and 4 respectively, which costs about 60G memory when batch size is 1. We further apply the 8-step gradient accumulation to simulate a larger batch size. We train the model based on SDXL-Base-1.0 [32] with AdamW [27] optimizer and learning rate 0.00001. In this 3-stage training process, we optimize the Base-LoRA, ID-LoRA (CLIP) and ID-LoRA (ID embedding) for 20K, 15K and 55K iterations in turn, and set the probability of three situations to 0.9, 0.05 and 0.05 for ID-LoRA training. We implement our method with PyTorch and Diffusers. The total training process is performed on 16 NVIDIA A100 GPUs for about 10 days." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.758, + 0.191, + 0.775 + ], + "angle": 0, + "content": "4.2. Settings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We build workflows in ComfyUI [3] to evaluate our HyperLoRA and three SOTA ID preservation methods, i.e. InstantID [44], IP-Adapter [45] (specifically, IP-Adapter-FaceID-Portrait) and PuLID [11]. To ensure a fair comparison, we apply the same seed, prompt and sampler across all methods. Note that we additionally concatenate the lora trigger words as a prefix to the prompt when testing HyperLoRA, and set different Classifier-Free Guidance [15] (CFG) scales" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.244 + ], + "angle": 0, + "content": "to generate reasonable images. As shown in Fig. 5, Adapter-based methods usually require a lower CFG scale while HyperLoRA allows a wide range of CFG scales. Although these methods are trained on SDXL-Base-1.0, they show sufficient generalization to other SDXL models. Thus, we use LEOSAM [23], an open-source SDXL model for high quality portraits, as the base model for evaluations, which typically improve image aesthetics. In quantitative comparison, we utilize a recent open-source test set, Unsplash-50 [8], which covers diverse genders, ages and races." + }, + { + "type": "table_caption", + "bbox": [ + 0.607, + 0.254, + 0.813, + 0.269 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.279, + 0.903, + 0.407 + ], + "angle": 0, + "content": "
FidelityEditability
CLIP-I↑ID Sim.↑CLIP-I↑CLIP-T↑
IP-Adapter0.7640.5660.7250.244
InstantID0.7340.6810.6880.237
PuLID0.7710.6130.8050.259
Arc2Face0.7860.643--
Ours (Full)0.8530.6780.7100.243
Ours (ID)0.8310.6250.7480.252
" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.426, + 0.654, + 0.442 + ], + "angle": 0, + "content": "4.3. Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.448, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Quantitative Comparison. We evaluate fidelity and editability as the quantitative metrics for comparison across these methods. For fidelity, we leverage the cosine of CLIP image features (CLIP-I) to measure face fidelity, and the cosine of ID embeddings (ID Sim.) to measure face ID similarity. We generate four images with prompt headshot of a girl/boy, high definition, 4k for each sample in our test set using each method, and compute the average of them. For editability, we utilize CLIP-I to measure the similarity of images generated with and without ID plug-in networks, and the cosine of CLIP image and text features (CLIP-T) to measure image-text similarity. We collect 10 prompts which cover hair, accessory, clothing and background, and evaluate all combinations of prompts and ID images. The specific CLIP model we used is CLIP-L/14 which is different from the text encoder in SDXL, and CurricularFace [19] (not used as ID encoder in all methods) is adopted to extract ID embeddings. We further incorporate Arc2Face [31], a human face base model, into comparison, which is trained using a fixed caption and a cropped region of faces, thus restricting its editability. Ours (Full) in Table. 1 means inference with both Base-LoRA (weight is set to 0.4) and ID-LoRA, while Ours (ID) represents only using ID-LoRA. As shown in Table. 1 and Fig. 6, Base-LoRA helps to improve fidelity but limits editability. Although our design decouples the image features into different LoRAs, it's hard to avoid leaking mutually. Thus, we can adjust the weight of Base-LoRA to adapt to different application scenarios. Our HyperLoRA (Full and ID) achieve the best and second-best face fidelity while InstantID shows superiority" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13119" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.108, + 0.085, + 0.891, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.579, + 0.908, + 0.608 + ], + "angle": 0, + "content": "Figure 7. Qualitative comparison. From top to bottom, we generate images with the scene prompts: white shirt, wolf ears, wavy hair and brown sweater, and play with snowman." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.629, + 0.482, + 0.81 + ], + "angle": 0, + "content": "in face ID similarity but lower face fidelity. Both these metrics should be considered together to evaluate fidelity, since the face ID similarity is more abstract and face fidelity reflects more details. This is evidenced in the following visual results in Fig. 7. In the aspect of editability, PuLID possesses a clear advantage, the key contribution of which is a well-designed loss function to reduce intrusion into the base model, thereby improving editability. Despite HyperLoRA (ID) inferior to PuLID, it surpasses other methods, meanwhile has good fidelity. It is noteworthy that there is a trade-off between fidelity and editability, and our HyperLoRA reaches a proper balance." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.811, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Qualitative Comparison. Fig. 7 shows the visual comparisons across IP-Adapter, InstantID, PuLID and our HyperLoRA (only ID-LoRA). The skin of portraits generated by IP-Adapter and InstantID has apparent AI-generated texture, which is a little oversaturation and far from photorealism. It is a common shortcoming of Adapter-based meth" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.907, + 0.886 + ], + "angle": 0, + "content": "ods. PuLID improves this problem by weakening the intrusion to base model, outperforming IP-Adapter and InstantID but still suffering from blurring and lack of details. In contrast, LoRA directly modifies the base model weights instead of introducing extra attention modules, usually generating highly detailed and photorealistic images. Our HyperLoRA naturally inherits this advantage from LoRA. Furthermore, the manner in which LoRA injects ID features exhibits higher nonlinearity. Thus, our method achieves better face fidelity, capturing fine-grained details, such as pupil color. The layout of generated images of our HyperLoRA and PuLID is more similar compared with other methods, which is because of the higher CLIP-I and CLIP-T of editability. In comparison, IP-Adapter and InstantID have inferior editability, resulting in unreasonable layout (the 4th row of IP-Adapter) and failure to follow the prompt (the 3rd row of InstantID)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13120" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.086, + 0.482, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.304, + 0.483, + 0.361 + ], + "angle": 0, + "content": "Figure 8. Impact of Hyper Base-LoRA. The first row: the model trained with Hyper Base-LoRA generates correct clothing and background. The second row: the model trained without it takes many irrelevant features into the generated images." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.377, + 0.485, + 0.613 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.622, + 0.483, + 0.679 + ], + "angle": 0, + "content": "Figure 9. Comparison between different encoders. The first row: our CLIP + ID embeddings training strategy successfully generates glasses and mask. The second row: training with only CLIP fails to follow the prompts." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.709, + 0.245, + 0.725 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.732, + 0.483, + 0.869 + ], + "angle": 0, + "content": "No Hyper Base-LoRA. In Fig. 8, we demonstrate the effectiveness of Hyper Base-LoRA. As aforementioned, our Hyper Base-LoRA helps to decouple ID information from input images. This is evidenced in the second row of Fig. 8, where the model trained without Hyper Base-LoRA fails to properly reflect the prompts dress, spacesuit and beach. As expected, the background and clothing are more editable if trained with Hyper Base-LoRA, since it prevents the ID-irrelevant features leaking to ID-LoRA." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Only CLIP Encoder. Our multi-stage training strategy is crucial for successfully training ID-LoRA. Especially, we" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.087, + 0.906, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.234, + 0.905, + 0.262 + ], + "angle": 0, + "content": "Figure 10. Multiple input images contribute to a higher ID similarity and a more realistic and aesthetic synthesized portrait." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.293, + 0.907, + 0.398 + ], + "angle": 0, + "content": "leverage CLIP ViT features only during the early iterations of training and then switch to ID embeddings to learn more details. Fig. 9 shows that the problem arises when only adopting CLIP as encoder, where we see that the facial elements are challenging to modify. In contrast, predominantly learning facial features with ID embeddings relaxes the structural constraints on the face." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.4, + 0.907, + 0.522 + ], + "angle": 0, + "content": "Multiple Input Images. Thanks to the natural interpolability of LoRA and our low-dimensional linear LoRA space (as discussed in Sec. 3.1), it is straightforward to support portrait synthesis with multiple input images by averaging the LoRA coefficients generated from each single image. Fig. 10 illustrates that multiple ID images help HyperLoRA to identify more stable and pure ID features, thereby improving both ID fidelity and image quality." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.542, + 0.634, + 0.557 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We have proposed HyperLoRA, the first zero-shot personalized portrait generation method based on LoRA, which is trained in an end-to-end manner. With the advantage of LoRA, our method modifies the base model weights to inject ID information rather than introducing new cross attention as adapter-based methods do, exhibiting higher nonlinearity, thereby yielding more detailed and realistic portrait images. To make HyperLoRA training feasible, we represent the whole LoRA parameters with a linear combination of a group of LoRA basis, which reduces the number of HyperLoRA parameters magnificently. We further decompose the LoRA into two parts explicitly. With our well-designed training strategy, the ID part tends to learn facial features while other aspects, like background and clothing, are fitted into the base part, thereby reducing the risk of overfitting and improving editability. It's recognized that only a rank 8 HyperLoRA is allowed to train when considering the memory limitation of current GPU. Nevertheless, rank 8 is still significantly smaller than what is frequently used in normal LoRA. An advanced LoRA parameters compression method, such as LoRA autoencoder, may enable larger rank HypeLoRA training, unlocking more powerful capability." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13121" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.197 + ], + "angle": 0, + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.2, + 0.484, + 0.227 + ], + "angle": 0, + "content": "[2] Jimmy Lei Ba. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.23, + 0.483, + 0.257 + ], + "angle": 0, + "content": "[3] comfyanonymous. Comfyui: The most powerful and modular diffusion model gui and backend. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.259, + 0.484, + 0.315 + ], + "angle": 0, + "content": "[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.317, + 0.483, + 0.357 + ], + "angle": 0, + "content": "[5] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.36, + 0.483, + 0.413 + ], + "angle": 0, + "content": "[6] Zhongyi Fan, Zixin Yin, Gang Li, Yibing Zhan, and Heliang Zheng. Dreambooth++: Boosting subject-driven generation via region-level references packing. In ACM Multimedia 2024, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.417, + 0.483, + 0.471 + ], + "angle": 0, + "content": "[7] Rinon Gal, Moab Arar, Yuval Atzmon, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Encoder-based domain tuning for fast personalization of text-to-image models. ACM Transactions on Graphics (TOG), 42(4):1-13, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.474, + 0.483, + 0.529 + ], + "angle": 0, + "content": "[8] Rinon Gal, Or Lichter, Elad Richardson, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Lcmlookahead for encoder-based text-to-image personalization. arXiv preprint arXiv:2404.03620, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.531, + 0.483, + 0.586 + ], + "angle": 0, + "content": "[9] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.588, + 0.483, + 0.616 + ], + "angle": 0, + "content": "[10] Jia Guo and Jiankang Deng. Insightface: 2d and 3d face analysis project. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.617, + 0.483, + 0.672 + ], + "angle": 0, + "content": "[11] Zinan Guo, Yanze Wu, Zhuowei Chen, Lang Chen, and Qian He. Pulid: Pure and lightning id customization via contrastive alignment. Advances in neural information processing systems, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.675, + 0.483, + 0.744 + ], + "angle": 0, + "content": "[12] Ligong Han, Yinxiao Li, Han Zhang, Peyman Milanfar, Dimitris Metaxas, and Feng Yang. Svdiff: Compact parameter space for diffusion fine-tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7323-7334, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.746, + 0.483, + 0.8 + ], + "angle": 0, + "content": "[13] Soufiane Hayou, Nikhil Ghosh, and Bin Yu. Lora+: Efficient low rank adaptation of large models. In *Forty-first International Conference on Machine Learning*, ICML 2024, Vienna, Austria, July 21-27, 2024, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[14] Zecheng He, Bo Sun, Felix Juefei-Xu, Haoyu Ma, Ankit Ramchandani, Vincent Cheung, Siddharth Shah, Anmol Kalia, Harihar Subramanyam, Alireza Zareian, et al. Imagine yourself: Tuning-free personalized image generation. arXiv preprint arXiv:2409.13346, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[15] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 6" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.905, + 0.192 + ], + "angle": 0, + "content": "[17] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. *J. Mach. Learn.* Res., 23(47):1-33, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.195, + 0.905, + 0.277 + ], + "angle": 0, + "content": "[18] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.281, + 0.905, + 0.362 + ], + "angle": 0, + "content": "[19] Yuge Huang, Yuhan Wang, Ying Tai, Xiaoming Liu, Pengcheng Shen, Shaoxin Li, Jilin Li, and Feiyue Huang. Curricularface: adaptive curriculum learning loss for deep face recognition. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5901-5910, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.366, + 0.762, + 0.38 + ], + "angle": 0, + "content": "[20] Insightface. Antelopev2, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.383, + 0.907, + 0.451 + ], + "angle": 0, + "content": "[21] Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up gans for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10124-10134, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.454, + 0.905, + 0.495 + ], + "angle": 0, + "content": "[22] Diederik Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. Advances in neural information processing systems, 34:21696-21707, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.498, + 0.812, + 0.512 + ], + "angle": 0, + "content": "[23] LEOSAM. Leosam's helloworld xl, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.515, + 0.905, + 0.582 + ], + "angle": 0, + "content": "[24] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.586, + 0.905, + 0.655 + ], + "angle": 0, + "content": "[25] Zhen Li, Mingdeng Cao, Xintao Wang, Zhongang Qi, Ming-Ming Cheng, and Ying Shan. Photomaker: Customizing realistic human photos via stacked id embedding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8640-8650, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.658, + 0.905, + 0.74 + ], + "angle": 0, + "content": "[26] Shih-Yang Liu, Chien-Yi Wang, Hongxu Yin, Pavlo Molchanov, Yu-Chiang Frank Wang, Kwang-Ting Cheng, and Min-Hung Chen. Dora: Weight-decomposed low-rank adaptation. In *Forty-first International Conference on Machine Learning*, ICML 2024, Vienna, Austria, July 21-27, 2024, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.744, + 0.905, + 0.771 + ], + "angle": 0, + "content": "[27] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.774, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[28] Jian Ma, Junhao Liang, Chen Chen, and Haonan Lu. Subject-diffusion: Open domain personalized text-to-image generation without test-time fine-tuning. In ACM SIGGRAPH 2024 Conference Papers, pages 1-12, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[29] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4296–4304, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13122" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[30] Lianyu Pang, Jian Yin, Baoquan Zhao, Feize Wu, Fu Lee Wang, Qing Li, and Xudong Mao. Attndreambooth: Towards text-aligned personalized text-to-image generation. arXiv preprint arXiv:2406.05000, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[31] Foivos Paraperas Papantoniou, Alexandros Lattas, Stylianos Moschoglou, Jiankang Deng, Bernhard Kainz, and Stefanos Zafeiriou. Arc2face: A foundation model for id-consistent human faces. In European Conference on Computer Vision, pages 241-261. Springer, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.219, + 0.482, + 0.286 + ], + "angle": 0, + "content": "[32] Dustin Podell, *Zion English*, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. *Sdxl: Improving latent diffusion models for high-resolution image synthesis.* arXiv preprint arXiv:2307.01952, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.288, + 0.482, + 0.357 + ], + "angle": 0, + "content": "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.359, + 0.482, + 0.426 + ], + "angle": 0, + "content": "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.428, + 0.482, + 0.482 + ], + "angle": 0, + "content": "[35] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.482, + 0.566 + ], + "angle": 0, + "content": "[36] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdreambooth: Hypernetworks for fast personalization of text-to-image models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6527-6536, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.482, + 0.594 + ], + "angle": 0, + "content": "[37] Simo Ryu. Low-rank adaptation for fast text-to-image diffusion fine-tuning. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.596, + 0.482, + 0.663 + ], + "angle": 0, + "content": "[38] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.665, + 0.482, + 0.747 + ], + "angle": 0, + "content": "[39] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278–25294, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.749, + 0.482, + 0.815 + ], + "angle": 0, + "content": "[40] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[41] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[42] Ming Tao, Hao Tang, Fei Wu, Xiao-Yuan Jing, Bing-Kun Bao, and Changsheng Xu. Df-gan: A simple and effec" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.134 + ], + "angle": 0, + "content": "tive baseline for text-to-image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16515-16525, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.906, + 0.163 + ], + "angle": 0, + "content": "[43] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.906, + 0.219 + ], + "angle": 0, + "content": "[44] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, and Anthony Chen. Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519, 2024. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.906, + 0.262 + ], + "angle": 0, + "content": "[45] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. 2023. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.906, + 0.332 + ], + "angle": 0, + "content": "[46] Shih-Ying Yeh, Yu-Guan Hsieh, Zhidong Gao, Bernard BW Yang, Giyeong Oh, and Yanmin Gong. Navigating text-to-image customization: From lycoris fine-tuning to model evaluation. In The Twelfth International Conference on Learning Representations, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.906, + 0.403 + ], + "angle": 0, + "content": "[47] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2(3):5, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.405, + 0.906, + 0.486 + ], + "angle": 0, + "content": "[48] Qingru Zhang, Minshuo Chen, Alexander Bukharin, Pengcheng He, Yu Cheng, Weizhu Chen, and Tuo Zhao. Adaptive budget allocation for parameter-efficient finetuning. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.486 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13123" + } + ] +] \ No newline at end of file diff --git a/2025/HyperLoRA_ Parameter-Efficient Adaptive Generation for Portrait Synthesis/0dcd811d-bb8b-4fcc-9377-9711c569d3e6_origin.pdf b/2025/HyperLoRA_ Parameter-Efficient Adaptive Generation for Portrait Synthesis/0dcd811d-bb8b-4fcc-9377-9711c569d3e6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..edcea0d4aecacd76a8bf3e8d5c9f2b65c5da4c20 --- /dev/null +++ b/2025/HyperLoRA_ Parameter-Efficient Adaptive Generation for Portrait Synthesis/0dcd811d-bb8b-4fcc-9377-9711c569d3e6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7f189f895ed02fb353d261887f2ee13524ec57ff50005ae5aefa2d1d774c2963 +size 6812853 diff --git a/2025/HyperLoRA_ Parameter-Efficient Adaptive Generation for Portrait Synthesis/full.md b/2025/HyperLoRA_ Parameter-Efficient Adaptive Generation for Portrait Synthesis/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8a813b60feac7b785f9c5e2ce9fa0ff24ef459fb --- /dev/null +++ b/2025/HyperLoRA_ Parameter-Efficient Adaptive Generation for Portrait Synthesis/full.md @@ -0,0 +1,272 @@ +# HyperLoRA: Parameter-Efficient Adaptive Generation for Portrait Synthesis + +Mengtian Li* + +Jinshu Chen* + +Wanquan Feng*† + +Songtao Zhao + +Qian He + +Intelligent Creation, ByteDance + +Bingchuan Li + +Fei Dai + +![](images/2b18acae0a80c6b533d7c6bdf8da7ab287eeee0eea655aea0bfb40567ae17b4d.jpg) +HyperLoRA +w.o. fine-tuning + +![](images/8e4f8b93bf2cee9cca76c169c42a560659ad4669d7e9f5ba7c643a48510897f4.jpg) +Wizard hat + +![](images/527d66c7ec402cd61404ff6d50c1345cfe4d97b11593ef85cd9996bb03d6b350.jpg) +Pink wavy hair + +![](images/33367ff554f711305f906ea5b484aec86bf46f0a046825bd5302cee8de043e9e.jpg) +Chinese clothing + +![](images/efb6a94812e77b87348a40623bae2e7ec47691139244f8276874c99ad3678cce.jpg) +Snow +Figure 1. We propose HyperLoRA, a parameter-efficient adaptive method for portrait synthesis. Given an input face image, HyperLoRA generates personalized LoRA weights without online fine-tuning. Due to the natural interpolability of LoRA, it is easy to support multiple inputs by simple averaging. Leveraging the generated LoRA, we can create personalized portraits with high photorealism and fidelity. + +![](images/ca26977f5a8df05a64d117118b84870a78868a3eb97e49848649369bf46b9546.jpg) +Beach + +![](images/89dcaae03d9e79937c984dcc9d617622b7ce36da044ecb805cb7f7c62da5e4ac.jpg) +Cyberpunk + +# Abstract + +Personalized portrait synthesis, essential in domains like social entertainment, has recently made significant progress. Person-wise fine-tuning based methods, such as LoRA and DreamBooth, can produce photorealistic outputs but need training on individual samples, consuming time and resources and posing an unstable risk. Adapter based techniques such as IP-Adapter freeze the foundational model parameters and employ a plug-in architecture to enable zero-shot inference, but they often exhibit a lack of naturalness and authenticity, which are not to be overlooked in portrait synthesis tasks. In this paper, we introduce a parameter-efficient adaptive generation method, + +namely HyperLoRA, that uses an adaptive plug-in network to generate LoRA weights, merging the superior performance of LoRA with the zero-shot capability of adapter scheme. Through our carefully designed network structure and training strategy, we achieve zero-shot personalized portrait generation (supporting both single and multiple image inputs) with high photorealism, fidelity, and editability. + +# 1. Introduction + +The personalized portrait synthesis task [14, 44, 45] aims to generate images sharing the same identity with what is contained in the given single or multiple portrait images. By feeding the identity information into specific generative model, the task enables the users to create diverse, high-quality and creative personalized portraits, thus hold + +ing substantial importance across various domains, such as entertainment and social media. + +One key challenge of the personalized portrait synthesis task is the fidelity: the expected generated images should retain the given identity to the greatest extent and avoid irrelevant outputs. Another challenge lies in editability: while preserving the given identity, the portrait synthesis model should maintain the abilities of flexibly editing the other contents of the generated images, for example, the composition of the image, the positioning and the clothing of the character. Additionally, the inference speed is crucial, as fast inference is friendly to not only the computational costs but also the user experience. + +Considering the challenges mentioned above, most of the current solutions have different shortcomings. Existing personalized image synthesis methods can generally be divided into tuning-based and tuning-free approaches. Tuning-based methods, such as LoRA (Low-Rank Adaptation) [18] and DreamBooth [35], update model parameters based on specific user-provided data, which can accurately comprehend and preserve the input identity information. Theoretically, these methods are mainly designed for few-shot cases, and customize different models for each groups of provided data with different identities. Thus, disadvantages like high computing costs, unstable training, and lengthy training processes make such methods bloated when dealing with various provided identities. + +Tuning-free methods, such as IP-Adapter [45], typically utilize an additional plug-in structure to encode user identity information. By feeding such identity-related tokens into the frozen foundational model, these methods make influences on the inference process, and ultimately yielding results that retain the given identities. Such additional adapter struture is trained with large portrait datasets in advance, and freezing when inference. Therefore, these methods get rid of the online training, and are convenient for the actual uses. However, for these methods, the interaction between encoded identity information and the foundational model relies on a group of fixed extra cross-attention modules, which might degrade the quality of generated images, leading to unnatural and non-photorealistic facial features, and fail to recover fine-grained details. In summary, the previous methods fail to simultaneously strike a balance between the fidelity, the editability, and the speed. + +In this paper, we introduce a parameter-efficient adaptive generation method, namely HyperLoRA, which achieves great identity fidelity, high editability and fast inference speed. HyperLoRA uses an adaptive plug-in network to encode identity images, which follows the design of tuning-free methods like IP-Adapter. However, unlike such methods which utilize the encoded tokens for the attention computation, we directly generate the LoRA weights with the plug-in network. This strategy not only maintains the zero + +shot capability of the adapter strategy, but also inherits the representation and high-quality generation effect of the LoRA strategy which adjusts the global parameters without real fine-tuning process. + +Considering the substantial number of parameters in LoRA, to reduce the training difficulty, we initially project the LoRA parameters onto a linear LoRA parameter space, rather than directly training in the original parameter space. In this way, the generated LoRA weights can be represented as a linear combination of a trainable basis. We employ a perceiver resampler (as in IP-Adapter [45]) to convert the input identity images into combination coefficients for the basis. To encode the input images, we utilize a CLIP ViT [33] and an AntelopeV2 encoder [20] to extract pixel-level and identity-sensitive features. These feature extractors mentioned above stay frozen during all the training process. To avoid overfitting, we decompose LoRA at the parameter level into an ID-LoRA part and a Base-LoRA part, thereby the ID and ID-irreverent information are fitted into different parts correspondingly. The whole training pipeline runs in a multi-stage way: only the Base-LoRA part is trained in the first stage as a warm-up start; and in the second stage, we train the ID-LoRA with CLIP ViT features; in the final stage, we only finetune the attention modules related to ID embeddings. Besides the case where only a single image is given, our proposed method can also handle cases that there are multiple inputs by simply interpolating the predicted LoRA weights. Our contributions can be summarized as follows: + +- We propose HyperLoRA, a parameter-efficient adaptive method for the personalized portrait synthesis task, merging the superior performance of the LoRA strategy with the zero-shot capability of the adapter strategy. +- We construct a low-dimension linear LoRA space, which reduces the training difficulty of the plug-in network while ensuring the identity preservation. +- We design a strategy that decomposes LoRA weights at the parameter level into a base part and an ID part, thereby effectively decoupling the identity information restored in the given images and significantly reducing overfitting. +- We conduct various experiments to demonstrate the capacities of our framework. Experimental results show that our method achieves both high fidelity and editability. + +# 2. Related Work + +# 2.1. Text-to-Image Synthesis + +Text-to-image (T2I) synthesis has been rapidly developed for several years. The task aims to generate high-quality images which are highly correlated with given texts or instructions. Some early works [21, 42, 47] employ transformer [43] or GAN [9] architectures to solve this task, while the performance of the solutions is not that satisfy + +ing owing to the limited capacities of their pipelines. Recently, rapid progress has been made by the methods adopting diffusion pipelines [5, 16, 17, 22, 40, 41]. With a stepwise noising-denoising design, the generated results of diffusion pipelines have far higher qualities than other self-supervised methods, especially on the datasets with huge intra-domain variance [4, 38]. In this paper, we follow a diffusion pipeline [34] to build up our pipeline to ensure the qualities and the varieties of the outputs. + +# 2.2. Tuning-based Personalized Portrait Synthesis + +Tuning-based methods finetune models based on user-provided data to comprehend user identity information. Dreambooth [35] and its derivative versions [6, 30, 36] treat the entire T2I model trainable, and utilize certain special captions to represent the given identities. LoRA [18] and its variants [7, 12, 13, 26, 46, 48] represent another category of tuning-based methods. Unlike Dreambooth, LoRA greatly enhances the flexibility of the fine-tuning scheme by creating several trainable low-rank matrices within the T2I model along with much fewer parameters than the whole model. Nevertheless, a common issue with all the aforementioned methods is that they lack the ability to handle zero-shot cases. Each fine-tuned model matches only a single customized task, thus the fine-tuning incurs a high burden in terms of time, computational resources, and storage resources. In this work, our proposed HyperLoRA does not rely on online training of user-provided data, thereby circumventing the resource consumption problem brought about by the training process. + +# 2.3. Tuning-free Personalized Portrait Synthesis + +Tuning-free methods, such as [25, 28, 29, 44, 45], pre-design a plug-in structure containing few parameters that can be inserted into the T2I base model. After pre-training and freezing such parameters on a large-scale dataset, tuning-free methods utilize such insertable structure to extract the identity information features from the given inputs, and inject them into the base model during inference. The online tuning is not necessary for these methods, thereby they can save resource costs and time. Concurrently, the design of the detachable parameters is also rather flexible. However, these methods merely learn the mapping relationship between the input image and the encoded tokens, which are utilized by a group of additional cross-attention modules, and the outputs are simply blended with the original ones. Our proposed HyperLoRA learns the mapping relationship between the input image and the network parameters, which is more essential. The implicit information contained in the network parameters is far more abundant and friendly to the base model than that in tokens. Compared with the methods mentioned above, our method has better adaptability, higher nonlinearity, and better fidelity. + +# 3. Method + +In this section, we introduce the details of our method, overview of which is illustrated in Fig. 2. In Sec. 3.1, we introduce the low-dimensional linear LoRA space. In Sec. 3.2, we introduce the network structure. In Sec. 3.3, we introduce the designed multi-stage training process. + +# 3.1. Low-Dimensional Linear LoRA Space + +The number of parameters in LoRA is typically large, which makes it difficult to generate the LoRA weights directly. Recognizing the interpolative nature of LoRA, we are inspired to exploit its linearity, and map the LoRA parameters onto a low-dimensional linear space. Note that our discussion below and in the rest of the paper is based on the implementation from [37], which is the most widely acknowledged way. Each layer of LoRA includes 8 matrices, which consist of the down and up matrices for query, key, value, and output. We project each matrix onto a $K$ -dimensional basis. Mark the number of LoRA layers as $N$ , then the overall degrees of freedom for the entire LoRA can be calculated as $N \times 8K$ , which is significantly smaller than the origin parameter number in most cases. In our paper, we use $K = 128$ , and the number of parameters after projection becomes approximately only $1.2\%$ (0.14M vs. 11.6M when the LoRA rank is 8) of the original one. To decouple the given specific identity information from the irrelevant features, we design to split a LoRA matrix into a base part and an ID part. We denote the basis of the ID part as $\{\mathbf{M}_{id}^{k}\}_{k=1}^{K}$ , the combination coefficients as $\{\alpha_{k}\}_{k=1}^{K}$ . Further, we denote the ID part as: + +$$ +\mathbf {M} _ {i d} = \sum_ {k = 1} ^ {K} \alpha_ {k} \cdot \mathbf {M} _ {i d} ^ {k}. \tag {1} +$$ + +Similarly, the base part, denoted as $\mathbf{M}_{\mathrm{base}}$ , follows the same format as $\mathbf{M}_{id}$ : + +$$ +\mathbf {M} _ {\text {b a s e}} = \sum_ {k = 1} ^ {K} \beta_ {k} \cdot \mathbf {M} _ {\text {b a s e}} ^ {k}. \tag {2} +$$ + +Although the mathematical format of Base-LoRA is similar to ID-LoRA, it possesses a different significance. Specifically, we configure $\mathbf{M}_{\text {base }}$ to contain identity-irrelevant contents across all the training samples, such as the background and clothing, essentially capturing information from the input image that should not be included in $\mathbf{M}_{id}$ . Explicitly modeling Base-LoRA can help ID-LoRA learn more essential identity-related information and ignore the irrelevant rest. In summary, we represent a HyperLoRA parameter matrix $\mathbf{M}$ as: + +$$ +\mathbf {M} = \mathbf {M} _ {\text {b a s e}} + \mathbf {M} _ {i d} = \sum_ {k = 1} ^ {K} \beta_ {k} \cdot \mathbf {M} _ {\text {b a s e}} ^ {k} + \sum_ {k = 1} ^ {K} \alpha_ {k} \cdot \mathbf {M} _ {i d} ^ {k}. \tag {3} +$$ + +![](images/2bf55c3c39ce1ba061b65e543361d922c182064a252efdbf9aa093ed14d4ac88.jpg) +Figure 2. Overview of HyperLoRA. We explicitly decompose the HyperLoRA into a Hyper ID-LoRA and a Hyper Base-LoRA. The former is designed to learn ID information while the latter is expected to fit others, e.g. background and clothing. Such a design helps to prevent irrelevant features leaking to ID-LoRA. During the training, we fix the weights of the pretrained SDXL base model and encoders, allowing only HyperLoRA modules updated by Backpropagation. At the inference stage, the Hyper ID-LoRA integrated into SDXL generates personalized images while the Hyper Base-LoRA is optional. + +It is noteworthy that although the degrees of freedom for the LoRA parameters are significantly compressed, they still retain the capacity to represent identity information, as Fig. 3 shows. With our designed projection method and the setting of $K = 128$ , the generated results in the illustration of overfitting a face dataset suggest that, it is still capable of reconstructing the identity of the reference image for the compressed LoRA module without obvious degradation. + +# 3.2. Network Structure + +In this section, we introduce our network structure, which is shown in Fig. 4. The whole network can be divided into several main components: the image encoder, the preceiver resampler and the LoRA combination. + +Image Encoder. The input image is passed into two image encoders, including a CLIP [33] ViT and an AntelopeV2 [20] encoder. The features extracted by the former are dense and contain strong structural information, while the features extracted by the latter are more abstract identity information. To explicitly distinguish the training processes of the base part and the id part, we deal with the input image of the CLIP in two different ways: for the training of $\mathbf{M}_{\text {base }}$ , we first crop the face region of an original image, and inside the cropped image, we blur the facial features and contour, preventing the base part reconstructing the specific face; for + +![](images/2b0c08ae050065e427bee9221844bc8915b4c93ae79d96636f63edc0fbc1d081.jpg) +Reference Image + +![](images/cd09f1168c4962caf4c722d24de9d8f94b578a06e93fbfdde8c50fe784b565b2.jpg) + +![](images/f828f3e3a907f8abd6595ffc4d6eb5b58fd6452ab80e8ac52ae22c7c21fef325.jpg) +Compressed LoRA + +![](images/201fcb3c3f2b820a5eca438797a83da2f9e1cedba3970649ac3e5af15279537c.jpg) + +![](images/c0fdb56353f9074afdb4f8f0fd3bfeea6de005c900c369bbc9c021f189bbc935.jpg) +Normal LoRA +Figure 3. The identity reconstruction ability on the low-dimensional linear LoRA space. In this example, we project LoRA parameters onto 128-dim basis and train on a face dataset (about 400K samples). Compared to normal LoRA, our compressed LoRA can also maintain the identity of the reference image well. + +the training of $\mathbf{M}_{id}$ , we directly feed the cropped face image into CLIP ViT. Meanwhile, it also serves as the input of AntelopeV2 encoder to extract ID embedding. We utilize a simple ID Projector, consisting of a Linear layer followed by a Reshape and a LayerNorm [2], to transform the ID em + +![](images/27560f4006184d5e5ab1f86f32f201f2d1baa3ac625da94aa89043686a14b0c7.jpg) +Figure 4. Network structure of HyperLoRA. We apply a perceiver resampler to convert the image features into a group of LoRA coefficients, thereby generating the whole LoRA by multiplied with LoRA basis. Two independent perceiver resamplers are instantiated for Hyper Base and ID LoRAs. Note that the second attention block (interacting with green tokens from ID Projector) is absent in Hyper Base-LoRA. + +bedding into some ID tokens. The details of the approaches mentioned above can be found in Fig. 4. + +Perceiver Resampler. The perceiver resampler [1, 45] employs a transformer backbone and incorporates image tokens into the computation through an attention mechanism. Currently, we apply a 4-layer perceiver resampler, each layer of which consists of two attention modules interacting with the tokens from CLIP ViT and ID Projector. Note that we skip the second attention module when training $\mathbf{M}_{\text {base }}$ . We obtain the output of the perveiver resampler, the combination coefficients $\{\alpha_{k}\}_{k=1}^{K}$ and $\{\beta_{k}\}_{k=1}^{K}$ , which would be used to compute the output LoRA weights subsequently. + +LoRA Combination. Both the Base-LoRA basis $\{\mathbf{M}_{base}^k\}_{k = 1}^K$ and the ID-LoRA basis $\{\mathbf{M}_{id}^k\}_{k = 1}^K$ are trainable variables. They are optimized during the training phase and fixed as constants during inference. Once the combination coefficients $\{\alpha_{k}\}_{k = 1}^{K}$ and $\{\beta_k\}_{k = 1}^K$ are predicted, the network can compute the entire LoRA by Eq. 3. It is noteworthy that the combination way can be adjusted flexibly to achieve various effects, for example, if the editability is a priority, the weight of $\mathbf{M}_{base}$ in Eq. 3 can be reduced to drop the information unrelated to the input identity. + +# 3.3. Training + +In this section, we introduce our multi-stage training process. Our approach begins with training the Base-LoRA as an initial warm-up. Subsequently, we expand the training to ID-LoRA with CLIP ViT. Finally, we fine-tune the model to enhance its identity preserving capability by using AntelopeV2 face encoder. For all stages, we employ the same loss term as in DDPM [16]. + +Base-LoRA Warm-up. Training the ID-LoRA directly from an open domain base model is quite challenging, as + +it tends to inadvertently incorporate extraneous details such as composition, background, and attire. Therefore, as described in Sec.3.1, we train non-facial information into the Base-LoRA, which serves as a warm-up in our training process. In this stage, we blur the facial information of the original input image before passing it through the CLIP ViT, and bypass the AntelopeV2 face encoder. + +ID-LoRA with CLIP. After the Base-LoRA warm-up, we add ID-LoRA into the training process. In this stage, we add the AntelopeV2 branch besides the existed CLIP branch, so that the abstract identity information can be efficiently extracted. We observe that training with CLIP features converges quickly but tends to overfit structural details, whereas training with ID embedding converges more slowly. Consequently, in this stage we stop training with CLIP features as long as it reaches the approximate convergence. To stabilize the training state in the early iterations, we initialize the output layers of the attention modules that are fused with ID embeddings to zero, allowing only the CLIP features to guide the early generation. To enable ID-LoRA to effectively represent the identity in the presence of trigger words (e.g., 3 rare words fcsks fxhks fhyks in this paper), while ensuring that the results remain unaffected in their absence, each iteration of this stage is set to one of the following situations: 1). The given prompt contains trigger words, both Base-LoRA and ID-LoRA are enabled, and the reconstruction target is the target image; 2). The given prompt does not contain trigger words, only ID-LoRA is enabled, and the resulting noise prediction should align with that of the original foundational model; 3). The given prompt does not contain trigger words, both Base-LoRA and ID-LoRA are enabled, and the resulting noise prediction should be consistent with that when only Base-LoRA is enabled. This approach minimizes disruption to the base model parameters and further concentrates the parameters on those related to identity information. + +ID-Embedding Fine-tune. After the previous training stages achieve rough convergence, we fix the weights of the CLIP-related attention modules, and continue to fine-tune the attention modules related to ID embedding (the second attention in Fig. 4). The training loss of this stage is exactly the same as in the second stage. + +# 4. Experiments + +# 4.1. Implementation Details + +To train our HyperLoRA model, we collect a subset from LAION-2B [39] which comprises 4.4 million portrait images. In detail, InsighFace [10] is leveraged to detect faces and we filter out those non-single portrait images. Meanwhile, we generate a caption for each image in our dataset using BLIP-2 [24] to enable Text-to-Image training. During the training, we randomly crop images around the face for + +![](images/316e2f65c3194033d69414bc7dd2d0f17aff60d6f9ba146294bd148ed8e018ad.jpg) +Figure 5. Adapter and LoRA show different tolerances to CFG. The first row: images generated by InstantID where higher CFG leads to oversaturation. The second row: HyperLoRA always yield reasonable portrait images for CFG ranged from 3 to 7. + +![](images/65a4f2b97319008de7f57fe1f2dee1306698be87dc2700c952eef0949e38a204.jpg) +Figure 6. Inference with Base-LoRA improves face fidelity but limits the layout of generated images more similar to the cropped region for training. + +data augmentation and focus more on the face region. Considering memory consumption, we currently set the LoRA rank of ID-LoRA and Base-LoRA to 8 and 4 respectively, which costs about 60G memory when batch size is 1. We further apply the 8-step gradient accumulation to simulate a larger batch size. We train the model based on SDXL-Base-1.0 [32] with AdamW [27] optimizer and learning rate 0.00001. In this 3-stage training process, we optimize the Base-LoRA, ID-LoRA (CLIP) and ID-LoRA (ID embedding) for 20K, 15K and 55K iterations in turn, and set the probability of three situations to 0.9, 0.05 and 0.05 for ID-LoRA training. We implement our method with PyTorch and Diffusers. The total training process is performed on 16 NVIDIA A100 GPUs for about 10 days. + +# 4.2. Settings + +We build workflows in ComfyUI [3] to evaluate our HyperLoRA and three SOTA ID preservation methods, i.e. InstantID [44], IP-Adapter [45] (specifically, IP-Adapter-FaceID-Portrait) and PuLID [11]. To ensure a fair comparison, we apply the same seed, prompt and sampler across all methods. Note that we additionally concatenate the lora trigger words as a prefix to the prompt when testing HyperLoRA, and set different Classifier-Free Guidance [15] (CFG) scales + +to generate reasonable images. As shown in Fig. 5, Adapter-based methods usually require a lower CFG scale while HyperLoRA allows a wide range of CFG scales. Although these methods are trained on SDXL-Base-1.0, they show sufficient generalization to other SDXL models. Thus, we use LEOSAM [23], an open-source SDXL model for high quality portraits, as the base model for evaluations, which typically improve image aesthetics. In quantitative comparison, we utilize a recent open-source test set, Unsplash-50 [8], which covers diverse genders, ages and races. + +Table 1. Quantitative comparison. + +
FidelityEditability
CLIP-I↑ID Sim.↑CLIP-I↑CLIP-T↑
IP-Adapter0.7640.5660.7250.244
InstantID0.7340.6810.6880.237
PuLID0.7710.6130.8050.259
Arc2Face0.7860.643--
Ours (Full)0.8530.6780.7100.243
Ours (ID)0.8310.6250.7480.252
+ +# 4.3. Comparisons + +Quantitative Comparison. We evaluate fidelity and editability as the quantitative metrics for comparison across these methods. For fidelity, we leverage the cosine of CLIP image features (CLIP-I) to measure face fidelity, and the cosine of ID embeddings (ID Sim.) to measure face ID similarity. We generate four images with prompt headshot of a girl/boy, high definition, 4k for each sample in our test set using each method, and compute the average of them. For editability, we utilize CLIP-I to measure the similarity of images generated with and without ID plug-in networks, and the cosine of CLIP image and text features (CLIP-T) to measure image-text similarity. We collect 10 prompts which cover hair, accessory, clothing and background, and evaluate all combinations of prompts and ID images. The specific CLIP model we used is CLIP-L/14 which is different from the text encoder in SDXL, and CurricularFace [19] (not used as ID encoder in all methods) is adopted to extract ID embeddings. We further incorporate Arc2Face [31], a human face base model, into comparison, which is trained using a fixed caption and a cropped region of faces, thus restricting its editability. Ours (Full) in Table. 1 means inference with both Base-LoRA (weight is set to 0.4) and ID-LoRA, while Ours (ID) represents only using ID-LoRA. As shown in Table. 1 and Fig. 6, Base-LoRA helps to improve fidelity but limits editability. Although our design decouples the image features into different LoRAs, it's hard to avoid leaking mutually. Thus, we can adjust the weight of Base-LoRA to adapt to different application scenarios. Our HyperLoRA (Full and ID) achieve the best and second-best face fidelity while InstantID shows superiority + +![](images/901c67ee63d2f56aa722f5da50d18b1d1c62692bff126481dd3ab97700430a0f.jpg) +Figure 7. Qualitative comparison. From top to bottom, we generate images with the scene prompts: white shirt, wolf ears, wavy hair and brown sweater, and play with snowman. + +in face ID similarity but lower face fidelity. Both these metrics should be considered together to evaluate fidelity, since the face ID similarity is more abstract and face fidelity reflects more details. This is evidenced in the following visual results in Fig. 7. In the aspect of editability, PuLID possesses a clear advantage, the key contribution of which is a well-designed loss function to reduce intrusion into the base model, thereby improving editability. Despite HyperLoRA (ID) inferior to PuLID, it surpasses other methods, meanwhile has good fidelity. It is noteworthy that there is a trade-off between fidelity and editability, and our HyperLoRA reaches a proper balance. + +Qualitative Comparison. Fig. 7 shows the visual comparisons across IP-Adapter, InstantID, PuLID and our HyperLoRA (only ID-LoRA). The skin of portraits generated by IP-Adapter and InstantID has apparent AI-generated texture, which is a little oversaturation and far from photorealism. It is a common shortcoming of Adapter-based meth + +ods. PuLID improves this problem by weakening the intrusion to base model, outperforming IP-Adapter and InstantID but still suffering from blurring and lack of details. In contrast, LoRA directly modifies the base model weights instead of introducing extra attention modules, usually generating highly detailed and photorealistic images. Our HyperLoRA naturally inherits this advantage from LoRA. Furthermore, the manner in which LoRA injects ID features exhibits higher nonlinearity. Thus, our method achieves better face fidelity, capturing fine-grained details, such as pupil color. The layout of generated images of our HyperLoRA and PuLID is more similar compared with other methods, which is because of the higher CLIP-I and CLIP-T of editability. In comparison, IP-Adapter and InstantID have inferior editability, resulting in unreasonable layout (the 4th row of IP-Adapter) and failure to follow the prompt (the 3rd row of InstantID). + +![](images/a714d4633ffbd913aa075e9f87d52f83a1bd3837da9fd2a77eb96680e3bdad29.jpg) +Figure 8. Impact of Hyper Base-LoRA. The first row: the model trained with Hyper Base-LoRA generates correct clothing and background. The second row: the model trained without it takes many irrelevant features into the generated images. + +![](images/1edf11bc1c76b8227732b87dc976de0be720b7142128ab48efa25c84d43aab43.jpg) +Figure 9. Comparison between different encoders. The first row: our CLIP + ID embeddings training strategy successfully generates glasses and mask. The second row: training with only CLIP fails to follow the prompts. + +# 4.4. Ablation Study + +No Hyper Base-LoRA. In Fig. 8, we demonstrate the effectiveness of Hyper Base-LoRA. As aforementioned, our Hyper Base-LoRA helps to decouple ID information from input images. This is evidenced in the second row of Fig. 8, where the model trained without Hyper Base-LoRA fails to properly reflect the prompts dress, spacesuit and beach. As expected, the background and clothing are more editable if trained with Hyper Base-LoRA, since it prevents the ID-irrelevant features leaking to ID-LoRA. + +Only CLIP Encoder. Our multi-stage training strategy is crucial for successfully training ID-LoRA. Especially, we + +![](images/85ae6a3b2bfecf5bf0980cf8aa94bd135eccbb32cbd9ff7e40abb3f44c522171.jpg) +Figure 10. Multiple input images contribute to a higher ID similarity and a more realistic and aesthetic synthesized portrait. + +leverage CLIP ViT features only during the early iterations of training and then switch to ID embeddings to learn more details. Fig. 9 shows that the problem arises when only adopting CLIP as encoder, where we see that the facial elements are challenging to modify. In contrast, predominantly learning facial features with ID embeddings relaxes the structural constraints on the face. + +Multiple Input Images. Thanks to the natural interpolability of LoRA and our low-dimensional linear LoRA space (as discussed in Sec. 3.1), it is straightforward to support portrait synthesis with multiple input images by averaging the LoRA coefficients generated from each single image. Fig. 10 illustrates that multiple ID images help HyperLoRA to identify more stable and pure ID features, thereby improving both ID fidelity and image quality. + +# 5. Conclusion + +We have proposed HyperLoRA, the first zero-shot personalized portrait generation method based on LoRA, which is trained in an end-to-end manner. With the advantage of LoRA, our method modifies the base model weights to inject ID information rather than introducing new cross attention as adapter-based methods do, exhibiting higher nonlinearity, thereby yielding more detailed and realistic portrait images. To make HyperLoRA training feasible, we represent the whole LoRA parameters with a linear combination of a group of LoRA basis, which reduces the number of HyperLoRA parameters magnificently. We further decompose the LoRA into two parts explicitly. With our well-designed training strategy, the ID part tends to learn facial features while other aspects, like background and clothing, are fitted into the base part, thereby reducing the risk of overfitting and improving editability. It's recognized that only a rank 8 HyperLoRA is allowed to train when considering the memory limitation of current GPU. Nevertheless, rank 8 is still significantly smaller than what is frequently used in normal LoRA. An advanced LoRA parameters compression method, such as LoRA autoencoder, may enable larger rank HypeLoRA training, unlocking more powerful capability. + +# References + +[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 5 +[2] Jimmy Lei Ba. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 4 +[3] comfyanonymous. Comfyui: The most powerful and modular diffusion model gui and backend. 6 +[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 3 +[5] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 3 +[6] Zhongyi Fan, Zixin Yin, Gang Li, Yibing Zhan, and Heliang Zheng. Dreambooth++: Boosting subject-driven generation via region-level references packing. In ACM Multimedia 2024, 2024. 3 +[7] Rinon Gal, Moab Arar, Yuval Atzmon, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Encoder-based domain tuning for fast personalization of text-to-image models. ACM Transactions on Graphics (TOG), 42(4):1-13, 2023. 3 +[8] Rinon Gal, Or Lichter, Elad Richardson, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Lcmlookahead for encoder-based text-to-image personalization. arXiv preprint arXiv:2404.03620, 2024. 6 +[9] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2 +[10] Jia Guo and Jiankang Deng. Insightface: 2d and 3d face analysis project. 5 +[11] Zinan Guo, Yanze Wu, Zhuowei Chen, Lang Chen, and Qian He. Pulid: Pure and lightning id customization via contrastive alignment. Advances in neural information processing systems, 2024. 6 +[12] Ligong Han, Yinxiao Li, Han Zhang, Peyman Milanfar, Dimitris Metaxas, and Feng Yang. Svdiff: Compact parameter space for diffusion fine-tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7323-7334, 2023. 3 +[13] Soufiane Hayou, Nikhil Ghosh, and Bin Yu. Lora+: Efficient low rank adaptation of large models. In *Forty-first International Conference on Machine Learning*, ICML 2024, Vienna, Austria, July 21-27, 2024, 2024. 3 +[14] Zecheng He, Bo Sun, Felix Juefei-Xu, Haoyu Ma, Ankit Ramchandani, Vincent Cheung, Siddharth Shah, Anmol Kalia, Harihar Subramanyam, Alireza Zareian, et al. Imagine yourself: Tuning-free personalized image generation. arXiv preprint arXiv:2409.13346, 2024. 1 +[15] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 6 + +[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 3, 5 +[17] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. *J. Mach. Learn.* Res., 23(47):1-33, 2022. 3 +[18] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022, 2022. 2, 3 +[19] Yuge Huang, Yuhan Wang, Ying Tai, Xiaoming Liu, Pengcheng Shen, Shaoxin Li, Jilin Li, and Feiyue Huang. Curricularface: adaptive curriculum learning loss for deep face recognition. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5901-5910, 2020. 6 +[20] Insightface. Antelopev2, 2023. 2, 4 +[21] Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up gans for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10124-10134, 2023. 2 +[22] Diederik Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. Advances in neural information processing systems, 34:21696-21707, 2021. 3 +[23] LEOSAM. Leosam's helloworld xl, 2023. 6 +[24] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 5 +[25] Zhen Li, Mingdeng Cao, Xintao Wang, Zhongang Qi, Ming-Ming Cheng, and Ying Shan. Photomaker: Customizing realistic human photos via stacked id embedding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8640-8650, 2024. 3 +[26] Shih-Yang Liu, Chien-Yi Wang, Hongxu Yin, Pavlo Molchanov, Yu-Chiang Frank Wang, Kwang-Ting Cheng, and Min-Hung Chen. Dora: Weight-decomposed low-rank adaptation. In *Forty-first International Conference on Machine Learning*, ICML 2024, Vienna, Austria, July 21-27, 2024, 2024. 3 +[27] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6 +[28] Jian Ma, Junhao Liang, Chen Chen, and Haonan Lu. Subject-diffusion: Open domain personalized text-to-image generation without test-time fine-tuning. In ACM SIGGRAPH 2024 Conference Papers, pages 1-12, 2024. 3 +[29] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4296–4304, 2024. 3 + +[30] Lianyu Pang, Jian Yin, Baoquan Zhao, Feize Wu, Fu Lee Wang, Qing Li, and Xudong Mao. Attndreambooth: Towards text-aligned personalized text-to-image generation. arXiv preprint arXiv:2406.05000, 2024. 3 +[31] Foivos Paraperas Papantoniou, Alexandros Lattas, Stylianos Moschoglou, Jiankang Deng, Bernhard Kainz, and Stefanos Zafeiriou. Arc2face: A foundation model for id-consistent human faces. In European Conference on Computer Vision, pages 241-261. Springer, 2024. 6 +[32] Dustin Podell, *Zion English*, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. *Sdxl: Improving latent diffusion models for high-resolution image synthesis.* arXiv preprint arXiv:2307.01952, 2023. 6 +[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. 2021. 2, 4 +[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 3 +[35] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. 2022. 2, 3 +[36] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdreambooth: Hypernetworks for fast personalization of text-to-image models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6527-6536, 2024. 3 +[37] Simo Ryu. Low-rank adaptation for fast text-to-image diffusion fine-tuning. 3 +[38] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 3 +[39] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278–25294, 2022. 5 +[40] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 3 +[41] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 3 +[42] Ming Tao, Hao Tang, Fei Wu, Xiao-Yuan Jing, Bing-Kun Bao, and Changsheng Xu. Df-gan: A simple and effec + +tive baseline for text-to-image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16515-16525, 2022. 2 +[43] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 2 +[44] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, and Anthony Chen. Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519, 2024. 1, 3, 6 +[45] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. 2023. 1, 2, 3, 5, 6 +[46] Shih-Ying Yeh, Yu-Guan Hsieh, Zhidong Gao, Bernard BW Yang, Giyeong Oh, and Yanmin Gong. Navigating text-to-image customization: From lycoris fine-tuning to model evaluation. In The Twelfth International Conference on Learning Representations, 2023. 3 +[47] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2(3):5, 2022. 2 +[48] Qingru Zhang, Minshuo Chen, Alexander Bukharin, Pengcheng He, Yu Cheng, Weizhu Chen, and Tuo Zhao. Adaptive budget allocation for parameter-efficient finetuning. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023, 2023. 3 \ No newline at end of file diff --git a/2025/HyperLoRA_ Parameter-Efficient Adaptive Generation for Portrait Synthesis/images.zip b/2025/HyperLoRA_ Parameter-Efficient Adaptive Generation for Portrait Synthesis/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7a0b4e099c9a8c2f62aa12b6152677a0d33da7aa --- /dev/null +++ b/2025/HyperLoRA_ Parameter-Efficient Adaptive Generation for Portrait Synthesis/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bb65111e9aaeb343cb21671775dec9ce021a42510c383802ec79ab0f2d7b3325 +size 746736 diff --git a/2025/HyperLoRA_ Parameter-Efficient Adaptive Generation for Portrait Synthesis/layout.json b/2025/HyperLoRA_ Parameter-Efficient Adaptive Generation for Portrait Synthesis/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..bfea41f270f56c18a5c11353297743d8f601dfb0 --- /dev/null +++ b/2025/HyperLoRA_ Parameter-Efficient Adaptive Generation for Portrait Synthesis/layout.json @@ -0,0 +1,6875 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 67, + 103, + 544, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 103, + 544, + 122 + ], + "spans": [ + { + "bbox": [ + 67, + 103, + 544, + 122 + ], + "type": "text", + "content": "HyperLoRA: Parameter-Efficient Adaptive Generation for Portrait Synthesis" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 95, + 142, + 162, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 142, + 162, + 158 + ], + "spans": [ + { + "bbox": [ + 95, + 142, + 162, + 158 + ], + "type": "text", + "content": "Mengtian Li*" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 186, + 143, + 252, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 143, + 252, + 157 + ], + "spans": [ + { + "bbox": [ + 186, + 143, + 252, + 157 + ], + "type": "text", + "content": "Jinshu Chen*" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 277, + 143, + 359, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 143, + 359, + 157 + ], + "spans": [ + { + "bbox": [ + 277, + 143, + 359, + 157 + ], + "type": "text", + "content": "Wanquan Feng*†" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 237, + 158, + 306, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 237, + 158, + 306, + 171 + ], + "spans": [ + { + "bbox": [ + 237, + 158, + 306, + 171 + ], + "type": "text", + "content": "Songtao Zhao" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 331, + 158, + 373, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 158, + 373, + 171 + ], + "spans": [ + { + "bbox": [ + 331, + 158, + 373, + 171 + ], + "type": "text", + "content": "Qian He" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 225, + 176, + 381, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 176, + 381, + 190 + ], + "spans": [ + { + "bbox": [ + 225, + 176, + 381, + 190 + ], + "type": "text", + "content": "Intelligent Creation, ByteDance" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 384, + 144, + 452, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 144, + 452, + 157 + ], + "spans": [ + { + "bbox": [ + 384, + 144, + 452, + 157 + ], + "type": "text", + "content": "Bingchuan Li" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 476, + 144, + 514, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 476, + 144, + 514, + 156 + ], + "spans": [ + { + "bbox": [ + 476, + 144, + 514, + 156 + ], + "type": "text", + "content": "Fei Dai" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 73, + 217, + 203, + 397 + ], + "blocks": [ + { + "bbox": [ + 73, + 217, + 203, + 397 + ], + "lines": [ + { + "bbox": [ + 73, + 217, + 203, + 397 + ], + "spans": [ + { + "bbox": [ + 73, + 217, + 203, + 397 + ], + "type": "image", + "image_path": "2b18acae0a80c6b533d7c6bdf8da7ab287eeee0eea655aea0bfb40567ae17b4d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 99, + 423, + 158, + 437 + ], + "lines": [ + { + "bbox": [ + 99, + 423, + 158, + 437 + ], + "spans": [ + { + "bbox": [ + 99, + 423, + 158, + 437 + ], + "type": "text", + "content": "HyperLoRA" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 97, + 439, + 159, + 450 + ], + "lines": [ + { + "bbox": [ + 97, + 439, + 159, + 450 + ], + "spans": [ + { + "bbox": [ + 97, + 439, + 159, + 450 + ], + "type": "text", + "content": "w.o. fine-tuning" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 210, + 217, + 317, + 325 + ], + "blocks": [ + { + "bbox": [ + 210, + 217, + 317, + 325 + ], + "lines": [ + { + "bbox": [ + 210, + 217, + 317, + 325 + ], + "spans": [ + { + "bbox": [ + 210, + 217, + 317, + 325 + ], + "type": "image", + "image_path": "8e4f8b93bf2cee9cca76c169c42a560659ad4669d7e9f5ba7c643a48510897f4.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 235, + 327, + 293, + 339 + ], + "lines": [ + { + "bbox": [ + 235, + 327, + 293, + 339 + ], + "spans": [ + { + "bbox": [ + 235, + 327, + 293, + 339 + ], + "type": "text", + "content": "Wizard hat" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 319, + 217, + 427, + 325 + ], + "blocks": [ + { + "bbox": [ + 319, + 217, + 427, + 325 + ], + "lines": [ + { + "bbox": [ + 319, + 217, + 427, + 325 + ], + "spans": [ + { + "bbox": [ + 319, + 217, + 427, + 325 + ], + "type": "image", + "image_path": "527d66c7ec402cd61404ff6d50c1345cfe4d97b11593ef85cd9996bb03d6b350.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 337, + 327, + 408, + 341 + ], + "lines": [ + { + "bbox": [ + 337, + 327, + 408, + 341 + ], + "spans": [ + { + "bbox": [ + 337, + 327, + 408, + 341 + ], + "type": "text", + "content": "Pink wavy hair" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 429, + 217, + 537, + 325 + ], + "blocks": [ + { + "bbox": [ + 429, + 217, + 537, + 325 + ], + "lines": [ + { + "bbox": [ + 429, + 217, + 537, + 325 + ], + "spans": [ + { + "bbox": [ + 429, + 217, + 537, + 325 + ], + "type": "image", + "image_path": "33367ff554f711305f906ea5b484aec86bf46f0a046825bd5302cee8de043e9e.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 441, + 327, + 523, + 341 + ], + "lines": [ + { + "bbox": [ + 441, + 327, + 523, + 341 + ], + "spans": [ + { + "bbox": [ + 441, + 327, + 523, + 341 + ], + "type": "text", + "content": "Chinese clothing" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 210, + 345, + 318, + 453 + ], + "blocks": [ + { + "bbox": [ + 210, + 345, + 318, + 453 + ], + "lines": [ + { + "bbox": [ + 210, + 345, + 318, + 453 + ], + "spans": [ + { + "bbox": [ + 210, + 345, + 318, + 453 + ], + "type": "image", + "image_path": "efb6a94812e77b87348a40623bae2e7ec47691139244f8276874c99ad3678cce.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 251, + 456, + 277, + 466 + ], + "lines": [ + { + "bbox": [ + 251, + 456, + 277, + 466 + ], + "spans": [ + { + "bbox": [ + 251, + 456, + 277, + 466 + ], + "type": "text", + "content": "Snow" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 476, + 555, + 510 + ], + "lines": [ + { + "bbox": [ + 55, + 476, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 476, + 555, + 510 + ], + "type": "text", + "content": "Figure 1. We propose HyperLoRA, a parameter-efficient adaptive method for portrait synthesis. Given an input face image, HyperLoRA generates personalized LoRA weights without online fine-tuning. Due to the natural interpolability of LoRA, it is easy to support multiple inputs by simple averaging. Leveraging the generated LoRA, we can create personalized portraits with high photorealism and fidelity." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 319, + 345, + 427, + 452 + ], + "blocks": [ + { + "bbox": [ + 319, + 345, + 427, + 452 + ], + "lines": [ + { + "bbox": [ + 319, + 345, + 427, + 452 + ], + "spans": [ + { + "bbox": [ + 319, + 345, + 427, + 452 + ], + "type": "image", + "image_path": "ca26977f5a8df05a64d117118b84870a78868a3eb97e49848649369bf46b9546.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 358, + 455, + 389, + 466 + ], + "lines": [ + { + "bbox": [ + 358, + 455, + 389, + 466 + ], + "spans": [ + { + "bbox": [ + 358, + 455, + 389, + 466 + ], + "type": "text", + "content": "Beach" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 429, + 345, + 537, + 453 + ], + "blocks": [ + { + "bbox": [ + 429, + 345, + 537, + 453 + ], + "lines": [ + { + "bbox": [ + 429, + 345, + 537, + 453 + ], + "spans": [ + { + "bbox": [ + 429, + 345, + 537, + 453 + ], + "type": "image", + "image_path": "89dcaae03d9e79937c984dcc9d617622b7ce36da044ecb805cb7f7c62da5e4ac.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 457, + 454, + 510, + 468 + ], + "lines": [ + { + "bbox": [ + 457, + 454, + 510, + 468 + ], + "spans": [ + { + "bbox": [ + 457, + 454, + 510, + 468 + ], + "type": "text", + "content": "Cyberpunk" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "bbox": [ + 152, + 518, + 200, + 531 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 518, + 200, + 531 + ], + "spans": [ + { + "bbox": [ + 152, + 518, + 200, + 531 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 54, + 545, + 297, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 545, + 297, + 689 + ], + "spans": [ + { + "bbox": [ + 54, + 545, + 297, + 689 + ], + "type": "text", + "content": "Personalized portrait synthesis, essential in domains like social entertainment, has recently made significant progress. Person-wise fine-tuning based methods, such as LoRA and DreamBooth, can produce photorealistic outputs but need training on individual samples, consuming time and resources and posing an unstable risk. Adapter based techniques such as IP-Adapter freeze the foundational model parameters and employ a plug-in architecture to enable zero-shot inference, but they often exhibit a lack of naturalness and authenticity, which are not to be overlooked in portrait synthesis tasks. In this paper, we introduce a parameter-efficient adaptive generation method," + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 520, + 556, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 520, + 556, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 520, + 556, + 605 + ], + "type": "text", + "content": "namely HyperLoRA, that uses an adaptive plug-in network to generate LoRA weights, merging the superior performance of LoRA with the zero-shot capability of adapter scheme. Through our carefully designed network structure and training strategy, we achieve zero-shot personalized portrait generation (supporting both single and multiple image inputs) with high photorealism, fidelity, and editability." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 314, + 621, + 395, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 621, + 395, + 634 + ], + "spans": [ + { + "bbox": [ + 314, + 621, + 395, + 634 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": "The personalized portrait synthesis task [14, 44, 45] aims to generate images sharing the same identity with what is contained in the given single or multiple portrait images. By feeding the identity information into specific generative model, the task enables the users to create diverse, high-quality and creative personalized portraits, thus hold" + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 229, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 229, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 229, + 713 + ], + "type": "text", + "content": "* Equal contribution † Corresponding author" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13114" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "content": "ing substantial importance across various domains, such as entertainment and social media." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 98, + 294, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 98, + 294, + 229 + ], + "spans": [ + { + "bbox": [ + 54, + 98, + 294, + 229 + ], + "type": "text", + "content": "One key challenge of the personalized portrait synthesis task is the fidelity: the expected generated images should retain the given identity to the greatest extent and avoid irrelevant outputs. Another challenge lies in editability: while preserving the given identity, the portrait synthesis model should maintain the abilities of flexibly editing the other contents of the generated images, for example, the composition of the image, the positioning and the clothing of the character. Additionally, the inference speed is crucial, as fast inference is friendly to not only the computational costs but also the user experience." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 232, + 296, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 232, + 296, + 399 + ], + "spans": [ + { + "bbox": [ + 56, + 232, + 296, + 399 + ], + "type": "text", + "content": "Considering the challenges mentioned above, most of the current solutions have different shortcomings. Existing personalized image synthesis methods can generally be divided into tuning-based and tuning-free approaches. Tuning-based methods, such as LoRA (Low-Rank Adaptation) [18] and DreamBooth [35], update model parameters based on specific user-provided data, which can accurately comprehend and preserve the input identity information. Theoretically, these methods are mainly designed for few-shot cases, and customize different models for each groups of provided data with different identities. Thus, disadvantages like high computing costs, unstable training, and lengthy training processes make such methods bloated when dealing with various provided identities." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 401, + 296, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 401, + 296, + 604 + ], + "spans": [ + { + "bbox": [ + 56, + 401, + 296, + 604 + ], + "type": "text", + "content": "Tuning-free methods, such as IP-Adapter [45], typically utilize an additional plug-in structure to encode user identity information. By feeding such identity-related tokens into the frozen foundational model, these methods make influences on the inference process, and ultimately yielding results that retain the given identities. Such additional adapter struture is trained with large portrait datasets in advance, and freezing when inference. Therefore, these methods get rid of the online training, and are convenient for the actual uses. However, for these methods, the interaction between encoded identity information and the foundational model relies on a group of fixed extra cross-attention modules, which might degrade the quality of generated images, leading to unnatural and non-photorealistic facial features, and fail to recover fine-grained details. In summary, the previous methods fail to simultaneously strike a balance between the fidelity, the editability, and the speed." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 605, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 294, + 713 + ], + "type": "text", + "content": "In this paper, we introduce a parameter-efficient adaptive generation method, namely HyperLoRA, which achieves great identity fidelity, high editability and fast inference speed. HyperLoRA uses an adaptive plug-in network to encode identity images, which follows the design of tuning-free methods like IP-Adapter. However, unlike such methods which utilize the encoded tokens for the attention computation, we directly generate the LoRA weights with the plug-in network. This strategy not only maintains the zero" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": "shot capability of the adapter strategy, but also inherits the representation and high-quality generation effect of the LoRA strategy which adjusts the global parameters without real fine-tuning process." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 121, + 553, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 553, + 419 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 553, + 419 + ], + "type": "text", + "content": "Considering the substantial number of parameters in LoRA, to reduce the training difficulty, we initially project the LoRA parameters onto a linear LoRA parameter space, rather than directly training in the original parameter space. In this way, the generated LoRA weights can be represented as a linear combination of a trainable basis. We employ a perceiver resampler (as in IP-Adapter [45]) to convert the input identity images into combination coefficients for the basis. To encode the input images, we utilize a CLIP ViT [33] and an AntelopeV2 encoder [20] to extract pixel-level and identity-sensitive features. These feature extractors mentioned above stay frozen during all the training process. To avoid overfitting, we decompose LoRA at the parameter level into an ID-LoRA part and a Base-LoRA part, thereby the ID and ID-irreverent information are fitted into different parts correspondingly. The whole training pipeline runs in a multi-stage way: only the Base-LoRA part is trained in the first stage as a warm-up start; and in the second stage, we train the ID-LoRA with CLIP ViT features; in the final stage, we only finetune the attention modules related to ID embeddings. Besides the case where only a single image is given, our proposed method can also handle cases that there are multiple inputs by simply interpolating the predicted LoRA weights. Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 422, + 553, + 587 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 314, + 422, + 553, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 422, + 553, + 468 + ], + "spans": [ + { + "bbox": [ + 314, + 422, + 553, + 468 + ], + "type": "text", + "content": "- We propose HyperLoRA, a parameter-efficient adaptive method for the personalized portrait synthesis task, merging the superior performance of the LoRA strategy with the zero-shot capability of the adapter strategy." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 469, + 553, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 469, + 553, + 504 + ], + "spans": [ + { + "bbox": [ + 314, + 469, + 553, + 504 + ], + "type": "text", + "content": "- We construct a low-dimension linear LoRA space, which reduces the training difficulty of the plug-in network while ensuring the identity preservation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 505, + 553, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 505, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 314, + 505, + 553, + 552 + ], + "type": "text", + "content": "- We design a strategy that decomposes LoRA weights at the parameter level into a base part and an ID part, thereby effectively decoupling the identity information restored in the given images and significantly reducing overfitting." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 553, + 553, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 553, + 553, + 587 + ], + "spans": [ + { + "bbox": [ + 314, + 553, + 553, + 587 + ], + "type": "text", + "content": "- We conduct various experiments to demonstrate the capacities of our framework. Experimental results show that our method achieves both high fidelity and editability." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 602, + 400, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 602, + 400, + 615 + ], + "spans": [ + { + "bbox": [ + 314, + 602, + 400, + 615 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 623, + 449, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 623, + 449, + 635 + ], + "spans": [ + { + "bbox": [ + 314, + 623, + 449, + 635 + ], + "type": "text", + "content": "2.1. Text-to-Image Synthesis" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 641, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 553, + 713 + ], + "type": "text", + "content": "Text-to-image (T2I) synthesis has been rapidly developed for several years. The task aims to generate high-quality images which are highly correlated with given texts or instructions. Some early works [21, 42, 47] employ transformer [43] or GAN [9] architectures to solve this task, while the performance of the solutions is not that satisfy" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13115" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "content": "ing owing to the limited capacities of their pipelines. Recently, rapid progress has been made by the methods adopting diffusion pipelines [5, 16, 17, 22, 40, 41]. With a stepwise noising-denoising design, the generated results of diffusion pipelines have far higher qualities than other self-supervised methods, especially on the datasets with huge intra-domain variance [4, 38]. In this paper, we follow a diffusion pipeline [34] to build up our pipeline to ensure the qualities and the varieties of the outputs." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 189, + 291, + 202 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 189, + 291, + 202 + ], + "spans": [ + { + "bbox": [ + 55, + 189, + 291, + 202 + ], + "type": "text", + "content": "2.2. Tuning-based Personalized Portrait Synthesis" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 207, + 296, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 207, + 296, + 435 + ], + "spans": [ + { + "bbox": [ + 55, + 207, + 296, + 435 + ], + "type": "text", + "content": "Tuning-based methods finetune models based on user-provided data to comprehend user identity information. Dreambooth [35] and its derivative versions [6, 30, 36] treat the entire T2I model trainable, and utilize certain special captions to represent the given identities. LoRA [18] and its variants [7, 12, 13, 26, 46, 48] represent another category of tuning-based methods. Unlike Dreambooth, LoRA greatly enhances the flexibility of the fine-tuning scheme by creating several trainable low-rank matrices within the T2I model along with much fewer parameters than the whole model. Nevertheless, a common issue with all the aforementioned methods is that they lack the ability to handle zero-shot cases. Each fine-tuned model matches only a single customized task, thus the fine-tuning incurs a high burden in terms of time, computational resources, and storage resources. In this work, our proposed HyperLoRA does not rely on online training of user-provided data, thereby circumventing the resource consumption problem brought about by the training process." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 444, + 282, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 444, + 282, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 444, + 282, + 456 + ], + "type": "text", + "content": "2.3. Tuning-free Personalized Portrait Synthesis" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 462, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 295, + 714 + ], + "type": "text", + "content": "Tuning-free methods, such as [25, 28, 29, 44, 45], pre-design a plug-in structure containing few parameters that can be inserted into the T2I base model. After pre-training and freezing such parameters on a large-scale dataset, tuning-free methods utilize such insertable structure to extract the identity information features from the given inputs, and inject them into the base model during inference. The online tuning is not necessary for these methods, thereby they can save resource costs and time. Concurrently, the design of the detachable parameters is also rather flexible. However, these methods merely learn the mapping relationship between the input image and the encoded tokens, which are utilized by a group of additional cross-attention modules, and the outputs are simply blended with the original ones. Our proposed HyperLoRA learns the mapping relationship between the input image and the network parameters, which is more essential. The implicit information contained in the network parameters is far more abundant and friendly to the base model than that in tokens. Compared with the methods mentioned above, our method has better adaptability, higher nonlinearity, and better fidelity." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 71, + 370, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 370, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 370, + 83 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 91, + 553, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 553, + 152 + ], + "type": "text", + "content": "In this section, we introduce the details of our method, overview of which is illustrated in Fig. 2. In Sec. 3.1, we introduce the low-dimensional linear LoRA space. In Sec. 3.2, we introduce the network structure. In Sec. 3.3, we introduce the designed multi-stage training process." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 157, + 516, + 170 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 157, + 516, + 170 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 516, + 170 + ], + "type": "text", + "content": "3.1. Low-Dimensional Linear LoRA Space" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 174, + 555, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 174, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 174, + 555, + 437 + ], + "type": "text", + "content": "The number of parameters in LoRA is typically large, which makes it difficult to generate the LoRA weights directly. Recognizing the interpolative nature of LoRA, we are inspired to exploit its linearity, and map the LoRA parameters onto a low-dimensional linear space. Note that our discussion below and in the rest of the paper is based on the implementation from [37], which is the most widely acknowledged way. Each layer of LoRA includes 8 matrices, which consist of the down and up matrices for query, key, value, and output. We project each matrix onto a " + }, + { + "bbox": [ + 313, + 174, + 555, + 437 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 174, + 555, + 437 + ], + "type": "text", + "content": "-dimensional basis. Mark the number of LoRA layers as " + }, + { + "bbox": [ + 313, + 174, + 555, + 437 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 174, + 555, + 437 + ], + "type": "text", + "content": ", then the overall degrees of freedom for the entire LoRA can be calculated as " + }, + { + "bbox": [ + 313, + 174, + 555, + 437 + ], + "type": "inline_equation", + "content": "N \\times 8K" + }, + { + "bbox": [ + 313, + 174, + 555, + 437 + ], + "type": "text", + "content": ", which is significantly smaller than the origin parameter number in most cases. In our paper, we use " + }, + { + "bbox": [ + 313, + 174, + 555, + 437 + ], + "type": "inline_equation", + "content": "K = 128" + }, + { + "bbox": [ + 313, + 174, + 555, + 437 + ], + "type": "text", + "content": ", and the number of parameters after projection becomes approximately only " + }, + { + "bbox": [ + 313, + 174, + 555, + 437 + ], + "type": "inline_equation", + "content": "1.2\\%" + }, + { + "bbox": [ + 313, + 174, + 555, + 437 + ], + "type": "text", + "content": " (0.14M vs. 11.6M when the LoRA rank is 8) of the original one. To decouple the given specific identity information from the irrelevant features, we design to split a LoRA matrix into a base part and an ID part. We denote the basis of the ID part as " + }, + { + "bbox": [ + 313, + 174, + 555, + 437 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{M}_{id}^{k}\\}_{k=1}^{K}" + }, + { + "bbox": [ + 313, + 174, + 555, + 437 + ], + "type": "text", + "content": ", the combination coefficients as " + }, + { + "bbox": [ + 313, + 174, + 555, + 437 + ], + "type": "inline_equation", + "content": "\\{\\alpha_{k}\\}_{k=1}^{K}" + }, + { + "bbox": [ + 313, + 174, + 555, + 437 + ], + "type": "text", + "content": ". Further, we denote the ID part as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 388, + 442, + 553, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 442, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 388, + 442, + 553, + 475 + ], + "type": "interline_equation", + "content": "\\mathbf {M} _ {i d} = \\sum_ {k = 1} ^ {K} \\alpha_ {k} \\cdot \\mathbf {M} _ {i d} ^ {k}. \\tag {1}", + "image_path": "233e58ecfcc7b191badf038738ac4788a99002eb51f8221981789aa00ae3051b.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 479, + 553, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 553, + 503 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 553, + 503 + ], + "type": "text", + "content": "Similarly, the base part, denoted as " + }, + { + "bbox": [ + 313, + 479, + 553, + 503 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathrm{base}}" + }, + { + "bbox": [ + 313, + 479, + 553, + 503 + ], + "type": "text", + "content": ", follows the same format as " + }, + { + "bbox": [ + 313, + 479, + 553, + 503 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{id}" + }, + { + "bbox": [ + 313, + 479, + 553, + 503 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 380, + 508, + 553, + 540 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 508, + 553, + 540 + ], + "spans": [ + { + "bbox": [ + 380, + 508, + 553, + 540 + ], + "type": "interline_equation", + "content": "\\mathbf {M} _ {\\text {b a s e}} = \\sum_ {k = 1} ^ {K} \\beta_ {k} \\cdot \\mathbf {M} _ {\\text {b a s e}} ^ {k}. \\tag {2}", + "image_path": "e0f1fa240ec2e375f365c7027a5c9ed28ccef8b118131ea4537419dcd1a7ce75.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 545, + 555, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 665 + ], + "type": "text", + "content": "Although the mathematical format of Base-LoRA is similar to ID-LoRA, it possesses a different significance. Specifically, we configure " + }, + { + "bbox": [ + 313, + 545, + 555, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\text {base }}" + }, + { + "bbox": [ + 313, + 545, + 555, + 665 + ], + "type": "text", + "content": " to contain identity-irrelevant contents across all the training samples, such as the background and clothing, essentially capturing information from the input image that should not be included in " + }, + { + "bbox": [ + 313, + 545, + 555, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{id}" + }, + { + "bbox": [ + 313, + 545, + 555, + 665 + ], + "type": "text", + "content": ". Explicitly modeling Base-LoRA can help ID-LoRA learn more essential identity-related information and ignore the irrelevant rest. In summary, we represent a HyperLoRA parameter matrix " + }, + { + "bbox": [ + 313, + 545, + 555, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 313, + 545, + 555, + 665 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 320, + 670, + 553, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 670, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 320, + 670, + 553, + 715 + ], + "type": "interline_equation", + "content": "\\mathbf {M} = \\mathbf {M} _ {\\text {b a s e}} + \\mathbf {M} _ {i d} = \\sum_ {k = 1} ^ {K} \\beta_ {k} \\cdot \\mathbf {M} _ {\\text {b a s e}} ^ {k} + \\sum_ {k = 1} ^ {K} \\alpha_ {k} \\cdot \\mathbf {M} _ {i d} ^ {k}. \\tag {3}", + "image_path": "1427d2f71b66b93bb4e07ff014c505fabe24b30a60bd9ec35114189a4f17708f.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "13116" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 70, + 541, + 319 + ], + "blocks": [ + { + "bbox": [ + 69, + 70, + 541, + 319 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 541, + 319 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 541, + 319 + ], + "type": "image", + "image_path": "2bf55c3c39ce1ba061b65e543361d922c182064a252efdbf9aa093ed14d4ac88.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 327, + 555, + 384 + ], + "lines": [ + { + "bbox": [ + 54, + 327, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 54, + 327, + 555, + 384 + ], + "type": "text", + "content": "Figure 2. Overview of HyperLoRA. We explicitly decompose the HyperLoRA into a Hyper ID-LoRA and a Hyper Base-LoRA. The former is designed to learn ID information while the latter is expected to fit others, e.g. background and clothing. Such a design helps to prevent irrelevant features leaking to ID-LoRA. During the training, we fix the weights of the pretrained SDXL base model and encoders, allowing only HyperLoRA modules updated by Backpropagation. At the inference stage, the Hyper ID-LoRA integrated into SDXL generates personalized images while the Hyper Base-LoRA is optional." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 403, + 296, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 403, + 296, + 499 + ], + "spans": [ + { + "bbox": [ + 54, + 403, + 296, + 499 + ], + "type": "text", + "content": "It is noteworthy that although the degrees of freedom for the LoRA parameters are significantly compressed, they still retain the capacity to represent identity information, as Fig. 3 shows. With our designed projection method and the setting of " + }, + { + "bbox": [ + 54, + 403, + 296, + 499 + ], + "type": "inline_equation", + "content": "K = 128" + }, + { + "bbox": [ + 54, + 403, + 296, + 499 + ], + "type": "text", + "content": ", the generated results in the illustration of overfitting a face dataset suggest that, it is still capable of reconstructing the identity of the reference image for the compressed LoRA module without obvious degradation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 513, + 167, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 513, + 167, + 525 + ], + "spans": [ + { + "bbox": [ + 55, + 513, + 167, + 525 + ], + "type": "text", + "content": "3.2. Network Structure" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 532, + 295, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 532, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 54, + 532, + 295, + 579 + ], + "type": "text", + "content": "In this section, we introduce our network structure, which is shown in Fig. 4. The whole network can be divided into several main components: the image encoder, the preceiver resampler and the LoRA combination." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 582, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 582, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 582, + 295, + 715 + ], + "type": "text", + "content": "Image Encoder. The input image is passed into two image encoders, including a CLIP [33] ViT and an AntelopeV2 [20] encoder. The features extracted by the former are dense and contain strong structural information, while the features extracted by the latter are more abstract identity information. To explicitly distinguish the training processes of the base part and the id part, we deal with the input image of the CLIP in two different ways: for the training of " + }, + { + "bbox": [ + 54, + 582, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\text {base }}" + }, + { + "bbox": [ + 54, + 582, + 295, + 715 + ], + "type": "text", + "content": ", we first crop the face region of an original image, and inside the cropped image, we blur the facial features and contour, preventing the base part reconstructing the specific face; for" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 316, + 401, + 391, + 476 + ], + "blocks": [ + { + "bbox": [ + 316, + 401, + 391, + 476 + ], + "lines": [ + { + "bbox": [ + 316, + 401, + 391, + 476 + ], + "spans": [ + { + "bbox": [ + 316, + 401, + 391, + 476 + ], + "type": "image", + "image_path": "2b0c08ae050065e427bee9221844bc8915b4c93ae79d96636f63edc0fbc1d081.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 477, + 374, + 483 + ], + "lines": [ + { + "bbox": [ + 331, + 477, + 374, + 483 + ], + "spans": [ + { + "bbox": [ + 331, + 477, + 374, + 483 + ], + "type": "text", + "content": "Reference Image" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 400, + 401, + 474, + 476 + ], + "blocks": [ + { + "bbox": [ + 400, + 401, + 474, + 476 + ], + "lines": [ + { + "bbox": [ + 400, + 401, + 474, + 476 + ], + "spans": [ + { + "bbox": [ + 400, + 401, + 474, + 476 + ], + "type": "image", + "image_path": "cd09f1168c4962caf4c722d24de9d8f94b578a06e93fbfdde8c50fe784b565b2.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 400, + 477, + 474, + 552 + ], + "blocks": [ + { + "bbox": [ + 400, + 477, + 474, + 552 + ], + "lines": [ + { + "bbox": [ + 400, + 477, + 474, + 552 + ], + "spans": [ + { + "bbox": [ + 400, + 477, + 474, + 552 + ], + "type": "image", + "image_path": "f828f3e3a907f8abd6595ffc4d6eb5b58fd6452ab80e8ac52ae22c7c21fef325.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 412, + 552, + 462, + 559 + ], + "lines": [ + { + "bbox": [ + 412, + 552, + 462, + 559 + ], + "spans": [ + { + "bbox": [ + 412, + 552, + 462, + 559 + ], + "type": "text", + "content": "Compressed LoRA" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 479, + 401, + 553, + 476 + ], + "blocks": [ + { + "bbox": [ + 479, + 401, + 553, + 476 + ], + "lines": [ + { + "bbox": [ + 479, + 401, + 553, + 476 + ], + "spans": [ + { + "bbox": [ + 479, + 401, + 553, + 476 + ], + "type": "image", + "image_path": "201fcb3c3f2b820a5eca438797a83da2f9e1cedba3970649ac3e5af15279537c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 479, + 477, + 553, + 552 + ], + "blocks": [ + { + "bbox": [ + 479, + 477, + 553, + 552 + ], + "lines": [ + { + "bbox": [ + 479, + 477, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 479, + 477, + 553, + 552 + ], + "type": "image", + "image_path": "c0fdb56353f9074afdb4f8f0fd3bfeea6de005c900c369bbc9c021f189bbc935.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 500, + 552, + 535, + 559 + ], + "lines": [ + { + "bbox": [ + 500, + 552, + 535, + 559 + ], + "spans": [ + { + "bbox": [ + 500, + 552, + 535, + 559 + ], + "type": "text", + "content": "Normal LoRA" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 568, + 555, + 624 + ], + "lines": [ + { + "bbox": [ + 313, + 568, + 555, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 568, + 555, + 624 + ], + "type": "text", + "content": "Figure 3. The identity reconstruction ability on the low-dimensional linear LoRA space. In this example, we project LoRA parameters onto 128-dim basis and train on a face dataset (about 400K samples). Compared to normal LoRA, our compressed LoRA can also maintain the identity of the reference image well." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "text", + "content": "the training of " + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{id}" + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "text", + "content": ", we directly feed the cropped face image into CLIP ViT. Meanwhile, it also serves as the input of AntelopeV2 encoder to extract ID embedding. We utilize a simple ID Projector, consisting of a Linear layer followed by a Reshape and a LayerNorm [2], to transform the ID em" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "13117" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 69, + 298, + 207 + ], + "blocks": [ + { + "bbox": [ + 57, + 69, + 298, + 207 + ], + "lines": [ + { + "bbox": [ + 57, + 69, + 298, + 207 + ], + "spans": [ + { + "bbox": [ + 57, + 69, + 298, + 207 + ], + "type": "image", + "image_path": "27560f4006184d5e5ab1f86f32f201f2d1baa3ac625da94aa89043686a14b0c7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 214, + 297, + 293 + ], + "lines": [ + { + "bbox": [ + 55, + 214, + 297, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 214, + 297, + 293 + ], + "type": "text", + "content": "Figure 4. Network structure of HyperLoRA. We apply a perceiver resampler to convert the image features into a group of LoRA coefficients, thereby generating the whole LoRA by multiplied with LoRA basis. Two independent perceiver resamplers are instantiated for Hyper Base and ID LoRAs. Note that the second attention block (interacting with green tokens from ID Projector) is absent in Hyper Base-LoRA." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 314, + 295, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 314, + 295, + 337 + ], + "spans": [ + { + "bbox": [ + 55, + 314, + 295, + 337 + ], + "type": "text", + "content": "bedding into some ID tokens. The details of the approaches mentioned above can be found in Fig. 4." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 338, + 295, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 338, + 295, + 458 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 295, + 458 + ], + "type": "text", + "content": "Perceiver Resampler. The perceiver resampler [1, 45] employs a transformer backbone and incorporates image tokens into the computation through an attention mechanism. Currently, we apply a 4-layer perceiver resampler, each layer of which consists of two attention modules interacting with the tokens from CLIP ViT and ID Projector. Note that we skip the second attention module when training " + }, + { + "bbox": [ + 55, + 338, + 295, + 458 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\text {base }}" + }, + { + "bbox": [ + 55, + 338, + 295, + 458 + ], + "type": "text", + "content": ". We obtain the output of the perveiver resampler, the combination coefficients " + }, + { + "bbox": [ + 55, + 338, + 295, + 458 + ], + "type": "inline_equation", + "content": "\\{\\alpha_{k}\\}_{k=1}^{K}" + }, + { + "bbox": [ + 55, + 338, + 295, + 458 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 338, + 295, + 458 + ], + "type": "inline_equation", + "content": "\\{\\beta_{k}\\}_{k=1}^{K}" + }, + { + "bbox": [ + 55, + 338, + 295, + 458 + ], + "type": "text", + "content": ", which would be used to compute the output LoRA weights subsequently." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 459, + 295, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 459, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 459, + 295, + 579 + ], + "type": "text", + "content": "LoRA Combination. Both the Base-LoRA basis " + }, + { + "bbox": [ + 55, + 459, + 295, + 579 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{M}_{base}^k\\}_{k = 1}^K" + }, + { + "bbox": [ + 55, + 459, + 295, + 579 + ], + "type": "text", + "content": " and the ID-LoRA basis " + }, + { + "bbox": [ + 55, + 459, + 295, + 579 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{M}_{id}^k\\}_{k = 1}^K" + }, + { + "bbox": [ + 55, + 459, + 295, + 579 + ], + "type": "text", + "content": " are trainable variables. They are optimized during the training phase and fixed as constants during inference. Once the combination coefficients " + }, + { + "bbox": [ + 55, + 459, + 295, + 579 + ], + "type": "inline_equation", + "content": "\\{\\alpha_{k}\\}_{k = 1}^{K}" + }, + { + "bbox": [ + 55, + 459, + 295, + 579 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 459, + 295, + 579 + ], + "type": "inline_equation", + "content": "\\{\\beta_k\\}_{k = 1}^K" + }, + { + "bbox": [ + 55, + 459, + 295, + 579 + ], + "type": "text", + "content": " are predicted, the network can compute the entire LoRA by Eq. 3. It is noteworthy that the combination way can be adjusted flexibly to achieve various effects, for example, if the editability is a priority, the weight of " + }, + { + "bbox": [ + 55, + 459, + 295, + 579 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{base}" + }, + { + "bbox": [ + 55, + 459, + 295, + 579 + ], + "type": "text", + "content": " in Eq. 3 can be reduced to drop the information unrelated to the input identity." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 587, + 119, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 587, + 119, + 600 + ], + "spans": [ + { + "bbox": [ + 55, + 587, + 119, + 600 + ], + "type": "text", + "content": "3.3. Training" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 605, + 295, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 688 + ], + "type": "text", + "content": "In this section, we introduce our multi-stage training process. Our approach begins with training the Base-LoRA as an initial warm-up. Subsequently, we expand the training to ID-LoRA with CLIP ViT. Finally, we fine-tune the model to enhance its identity preserving capability by using AntelopeV2 face encoder. For all stages, we employ the same loss term as in DDPM [16]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "Base-LoRA Warm-up. Training the ID-LoRA directly from an open domain base model is quite challenging, as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "it tends to inadvertently incorporate extraneous details such as composition, background, and attire. Therefore, as described in Sec.3.1, we train non-facial information into the Base-LoRA, which serves as a warm-up in our training process. In this stage, we blur the facial information of the original input image before passing it through the CLIP ViT, and bypass the AntelopeV2 face encoder." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 157, + 555, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 157, + 555, + 504 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 555, + 504 + ], + "type": "text", + "content": "ID-LoRA with CLIP. After the Base-LoRA warm-up, we add ID-LoRA into the training process. In this stage, we add the AntelopeV2 branch besides the existed CLIP branch, so that the abstract identity information can be efficiently extracted. We observe that training with CLIP features converges quickly but tends to overfit structural details, whereas training with ID embedding converges more slowly. Consequently, in this stage we stop training with CLIP features as long as it reaches the approximate convergence. To stabilize the training state in the early iterations, we initialize the output layers of the attention modules that are fused with ID embeddings to zero, allowing only the CLIP features to guide the early generation. To enable ID-LoRA to effectively represent the identity in the presence of trigger words (e.g., 3 rare words fcsks fxhks fhyks in this paper), while ensuring that the results remain unaffected in their absence, each iteration of this stage is set to one of the following situations: 1). The given prompt contains trigger words, both Base-LoRA and ID-LoRA are enabled, and the reconstruction target is the target image; 2). The given prompt does not contain trigger words, only ID-LoRA is enabled, and the resulting noise prediction should align with that of the original foundational model; 3). The given prompt does not contain trigger words, both Base-LoRA and ID-LoRA are enabled, and the resulting noise prediction should be consistent with that when only Base-LoRA is enabled. This approach minimizes disruption to the base model parameters and further concentrates the parameters on those related to identity information." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 505, + 554, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 505, + 554, + 577 + ], + "spans": [ + { + "bbox": [ + 313, + 505, + 554, + 577 + ], + "type": "text", + "content": "ID-Embedding Fine-tune. After the previous training stages achieve rough convergence, we fix the weights of the CLIP-related attention modules, and continue to fine-tune the attention modules related to ID embedding (the second attention in Fig. 4). The training loss of this stage is exactly the same as in the second stage." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 590, + 395, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 590, + 395, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 590, + 395, + 604 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 611, + 447, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 611, + 447, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 611, + 447, + 624 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "content": "To train our HyperLoRA model, we collect a subset from LAION-2B [39] which comprises 4.4 million portrait images. In detail, InsighFace [10] is leveraged to detect faces and we filter out those non-single portrait images. Meanwhile, we generate a caption for each image in our dataset using BLIP-2 [24] to enable Text-to-Image training. During the training, we randomly crop images around the face for" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13118" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 296, + 205 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 296, + 205 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 296, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 296, + 205 + ], + "type": "image", + "image_path": "316e2f65c3194033d69414bc7dd2d0f17aff60d6f9ba146294bd148ed8e018ad.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 213, + 295, + 258 + ], + "lines": [ + { + "bbox": [ + 55, + 213, + 295, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 213, + 295, + 258 + ], + "type": "text", + "content": "Figure 5. Adapter and LoRA show different tolerances to CFG. The first row: images generated by InstantID where higher CFG leads to oversaturation. The second row: HyperLoRA always yield reasonable portrait images for CFG ranged from 3 to 7." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 56, + 268, + 296, + 365 + ], + "blocks": [ + { + "bbox": [ + 56, + 268, + 296, + 365 + ], + "lines": [ + { + "bbox": [ + 56, + 268, + 296, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 268, + 296, + 365 + ], + "type": "image", + "image_path": "65a4f2b97319008de7f57fe1f2dee1306698be87dc2700c952eef0949e38a204.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 371, + 295, + 407 + ], + "lines": [ + { + "bbox": [ + 55, + 371, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 371, + 295, + 407 + ], + "type": "text", + "content": "Figure 6. Inference with Base-LoRA improves face fidelity but limits the layout of generated images more similar to the cropped region for training." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 426, + 295, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 426, + 295, + 594 + ], + "spans": [ + { + "bbox": [ + 54, + 426, + 295, + 594 + ], + "type": "text", + "content": "data augmentation and focus more on the face region. Considering memory consumption, we currently set the LoRA rank of ID-LoRA and Base-LoRA to 8 and 4 respectively, which costs about 60G memory when batch size is 1. We further apply the 8-step gradient accumulation to simulate a larger batch size. We train the model based on SDXL-Base-1.0 [32] with AdamW [27] optimizer and learning rate 0.00001. In this 3-stage training process, we optimize the Base-LoRA, ID-LoRA (CLIP) and ID-LoRA (ID embedding) for 20K, 15K and 55K iterations in turn, and set the probability of three situations to 0.9, 0.05 and 0.05 for ID-LoRA training. We implement our method with PyTorch and Diffusers. The total training process is performed on 16 NVIDIA A100 GPUs for about 10 days." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 600, + 116, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 116, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 116, + 613 + ], + "type": "text", + "content": "4.2. Settings" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": "We build workflows in ComfyUI [3] to evaluate our HyperLoRA and three SOTA ID preservation methods, i.e. InstantID [44], IP-Adapter [45] (specifically, IP-Adapter-FaceID-Portrait) and PuLID [11]. To ensure a fair comparison, we apply the same seed, prompt and sampler across all methods. Note that we additionally concatenate the lora trigger words as a prefix to the prompt when testing HyperLoRA, and set different Classifier-Free Guidance [15] (CFG) scales" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": "to generate reasonable images. As shown in Fig. 5, Adapter-based methods usually require a lower CFG scale while HyperLoRA allows a wide range of CFG scales. Although these methods are trained on SDXL-Base-1.0, they show sufficient generalization to other SDXL models. Thus, we use LEOSAM [23], an open-source SDXL model for high quality portraits, as the base model for evaluations, which typically improve image aesthetics. In quantitative comparison, we utilize a recent open-source test set, Unsplash-50 [8], which covers diverse genders, ages and races." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 315, + 220, + 552, + 322 + ], + "blocks": [ + { + "bbox": [ + 371, + 201, + 497, + 213 + ], + "lines": [ + { + "bbox": [ + 371, + 201, + 497, + 213 + ], + "spans": [ + { + "bbox": [ + 371, + 201, + 497, + 213 + ], + "type": "text", + "content": "Table 1. Quantitative comparison." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 220, + 552, + 322 + ], + "lines": [ + { + "bbox": [ + 315, + 220, + 552, + 322 + ], + "spans": [ + { + "bbox": [ + 315, + 220, + 552, + 322 + ], + "type": "table", + "html": "
FidelityEditability
CLIP-I↑ID Sim.↑CLIP-I↑CLIP-T↑
IP-Adapter0.7640.5660.7250.244
InstantID0.7340.6810.6880.237
PuLID0.7710.6130.8050.259
Arc2Face0.7860.643--
Ours (Full)0.8530.6780.7100.243
Ours (ID)0.8310.6250.7480.252
", + "image_path": "2e0b22a210784a9038de6fb107de7198ef4b30ff340a6788ca557566aaeeeebe.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 337, + 400, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 337, + 400, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 337, + 400, + 350 + ], + "type": "text", + "content": "4.3. Comparisons" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 354, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 556, + 715 + ], + "type": "text", + "content": "Quantitative Comparison. We evaluate fidelity and editability as the quantitative metrics for comparison across these methods. For fidelity, we leverage the cosine of CLIP image features (CLIP-I) to measure face fidelity, and the cosine of ID embeddings (ID Sim.) to measure face ID similarity. We generate four images with prompt headshot of a girl/boy, high definition, 4k for each sample in our test set using each method, and compute the average of them. For editability, we utilize CLIP-I to measure the similarity of images generated with and without ID plug-in networks, and the cosine of CLIP image and text features (CLIP-T) to measure image-text similarity. We collect 10 prompts which cover hair, accessory, clothing and background, and evaluate all combinations of prompts and ID images. The specific CLIP model we used is CLIP-L/14 which is different from the text encoder in SDXL, and CurricularFace [19] (not used as ID encoder in all methods) is adopted to extract ID embeddings. We further incorporate Arc2Face [31], a human face base model, into comparison, which is trained using a fixed caption and a cropped region of faces, thus restricting its editability. Ours (Full) in Table. 1 means inference with both Base-LoRA (weight is set to 0.4) and ID-LoRA, while Ours (ID) represents only using ID-LoRA. As shown in Table. 1 and Fig. 6, Base-LoRA helps to improve fidelity but limits editability. Although our design decouples the image features into different LoRAs, it's hard to avoid leaking mutually. Thus, we can adjust the weight of Base-LoRA to adapt to different application scenarios. Our HyperLoRA (Full and ID) achieve the best and second-best face fidelity while InstantID shows superiority" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13119" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 67, + 545, + 452 + ], + "blocks": [ + { + "bbox": [ + 66, + 67, + 545, + 452 + ], + "lines": [ + { + "bbox": [ + 66, + 67, + 545, + 452 + ], + "spans": [ + { + "bbox": [ + 66, + 67, + 545, + 452 + ], + "type": "image", + "image_path": "901c67ee63d2f56aa722f5da50d18b1d1c62692bff126481dd3ab97700430a0f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 458, + 555, + 481 + ], + "lines": [ + { + "bbox": [ + 55, + 458, + 555, + 481 + ], + "spans": [ + { + "bbox": [ + 55, + 458, + 555, + 481 + ], + "type": "text", + "content": "Figure 7. Qualitative comparison. From top to bottom, we generate images with the scene prompts: white shirt, wolf ears, wavy hair and brown sweater, and play with snowman." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 498, + 294, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 498, + 294, + 641 + ], + "spans": [ + { + "bbox": [ + 54, + 498, + 294, + 641 + ], + "type": "text", + "content": "in face ID similarity but lower face fidelity. Both these metrics should be considered together to evaluate fidelity, since the face ID similarity is more abstract and face fidelity reflects more details. This is evidenced in the following visual results in Fig. 7. In the aspect of editability, PuLID possesses a clear advantage, the key contribution of which is a well-designed loss function to reduce intrusion into the base model, thereby improving editability. Despite HyperLoRA (ID) inferior to PuLID, it surpasses other methods, meanwhile has good fidelity. It is noteworthy that there is a trade-off between fidelity and editability, and our HyperLoRA reaches a proper balance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 642, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 642, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 642, + 295, + 714 + ], + "type": "text", + "content": "Qualitative Comparison. Fig. 7 shows the visual comparisons across IP-Adapter, InstantID, PuLID and our HyperLoRA (only ID-LoRA). The skin of portraits generated by IP-Adapter and InstantID has apparent AI-generated texture, which is a little oversaturation and far from photorealism. It is a common shortcoming of Adapter-based meth" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 498, + 555, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 701 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 701 + ], + "type": "text", + "content": "ods. PuLID improves this problem by weakening the intrusion to base model, outperforming IP-Adapter and InstantID but still suffering from blurring and lack of details. In contrast, LoRA directly modifies the base model weights instead of introducing extra attention modules, usually generating highly detailed and photorealistic images. Our HyperLoRA naturally inherits this advantage from LoRA. Furthermore, the manner in which LoRA injects ID features exhibits higher nonlinearity. Thus, our method achieves better face fidelity, capturing fine-grained details, such as pupil color. The layout of generated images of our HyperLoRA and PuLID is more similar compared with other methods, which is because of the higher CLIP-I and CLIP-T of editability. In comparison, IP-Adapter and InstantID have inferior editability, resulting in unreasonable layout (the 4th row of IP-Adapter) and failure to follow the prompt (the 3rd row of InstantID)." + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13120" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 294, + 233 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 294, + 233 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 294, + 233 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 294, + 233 + ], + "type": "image", + "image_path": "a714d4633ffbd913aa075e9f87d52f83a1bd3837da9fd2a77eb96680e3bdad29.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 240, + 295, + 285 + ], + "lines": [ + { + "bbox": [ + 55, + 240, + 295, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 240, + 295, + 285 + ], + "type": "text", + "content": "Figure 8. Impact of Hyper Base-LoRA. The first row: the model trained with Hyper Base-LoRA generates correct clothing and background. The second row: the model trained without it takes many irrelevant features into the generated images." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 56, + 298, + 296, + 485 + ], + "blocks": [ + { + "bbox": [ + 56, + 298, + 296, + 485 + ], + "lines": [ + { + "bbox": [ + 56, + 298, + 296, + 485 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 296, + 485 + ], + "type": "image", + "image_path": "1edf11bc1c76b8227732b87dc976de0be720b7142128ab48efa25c84d43aab43.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 492, + 295, + 537 + ], + "lines": [ + { + "bbox": [ + 55, + 492, + 295, + 537 + ], + "spans": [ + { + "bbox": [ + 55, + 492, + 295, + 537 + ], + "type": "text", + "content": "Figure 9. Comparison between different encoders. The first row: our CLIP + ID embeddings training strategy successfully generates glasses and mask. The second row: training with only CLIP fails to follow the prompts." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 561, + 149, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 561, + 149, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 561, + 149, + 574 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 579, + 295, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 579, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 579, + 295, + 688 + ], + "type": "text", + "content": "No Hyper Base-LoRA. In Fig. 8, we demonstrate the effectiveness of Hyper Base-LoRA. As aforementioned, our Hyper Base-LoRA helps to decouple ID information from input images. This is evidenced in the second row of Fig. 8, where the model trained without Hyper Base-LoRA fails to properly reflect the prompts dress, spacesuit and beach. As expected, the background and clothing are more editable if trained with Hyper Base-LoRA, since it prevents the ID-irrelevant features leaking to ID-LoRA." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "Only CLIP Encoder. Our multi-stage training strategy is crucial for successfully training ID-LoRA. Especially, we" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 315, + 68, + 554, + 178 + ], + "blocks": [ + { + "bbox": [ + 315, + 68, + 554, + 178 + ], + "lines": [ + { + "bbox": [ + 315, + 68, + 554, + 178 + ], + "spans": [ + { + "bbox": [ + 315, + 68, + 554, + 178 + ], + "type": "image", + "image_path": "85ae6a3b2bfecf5bf0980cf8aa94bd135eccbb32cbd9ff7e40abb3f44c522171.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 185, + 553, + 207 + ], + "lines": [ + { + "bbox": [ + 313, + 185, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 313, + 185, + 553, + 207 + ], + "type": "text", + "content": "Figure 10. Multiple input images contribute to a higher ID similarity and a more realistic and aesthetic synthesized portrait." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 232, + 555, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 555, + 315 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 555, + 315 + ], + "type": "text", + "content": "leverage CLIP ViT features only during the early iterations of training and then switch to ID embeddings to learn more details. Fig. 9 shows that the problem arises when only adopting CLIP as encoder, where we see that the facial elements are challenging to modify. In contrast, predominantly learning facial features with ID embeddings relaxes the structural constraints on the face." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 316, + 555, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 555, + 413 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 555, + 413 + ], + "type": "text", + "content": "Multiple Input Images. Thanks to the natural interpolability of LoRA and our low-dimensional linear LoRA space (as discussed in Sec. 3.1), it is straightforward to support portrait synthesis with multiple input images by averaging the LoRA coefficients generated from each single image. Fig. 10 illustrates that multiple ID images help HyperLoRA to identify more stable and pure ID features, thereby improving both ID fidelity and image quality." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 429, + 388, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 429, + 388, + 441 + ], + "spans": [ + { + "bbox": [ + 314, + 429, + 388, + 441 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "type": "text", + "content": "We have proposed HyperLoRA, the first zero-shot personalized portrait generation method based on LoRA, which is trained in an end-to-end manner. With the advantage of LoRA, our method modifies the base model weights to inject ID information rather than introducing new cross attention as adapter-based methods do, exhibiting higher nonlinearity, thereby yielding more detailed and realistic portrait images. To make HyperLoRA training feasible, we represent the whole LoRA parameters with a linear combination of a group of LoRA basis, which reduces the number of HyperLoRA parameters magnificently. We further decompose the LoRA into two parts explicitly. With our well-designed training strategy, the ID part tends to learn facial features while other aspects, like background and clothing, are fitted into the base part, thereby reducing the risk of overfitting and improving editability. It's recognized that only a rank 8 HyperLoRA is allowed to train when considering the memory limitation of current GPU. Nevertheless, rank 8 is still significantly smaller than what is frequently used in normal LoRA. An advanced LoRA parameters compression method, such as LoRA autoencoder, may enable larger rank HypeLoRA training, unlocking more powerful capability." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13121" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "type": "text", + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 158, + 296, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 158, + 296, + 179 + ], + "spans": [ + { + "bbox": [ + 61, + 158, + 296, + 179 + ], + "type": "text", + "content": "[2] Jimmy Lei Ba. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 182, + 295, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 182, + 295, + 203 + ], + "spans": [ + { + "bbox": [ + 62, + 182, + 295, + 203 + ], + "type": "text", + "content": "[3] comfyanonymous. Comfyui: The most powerful and modular diffusion model gui and backend. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 205, + 296, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 205, + 296, + 249 + ], + "spans": [ + { + "bbox": [ + 62, + 205, + 296, + 249 + ], + "type": "text", + "content": "[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 251, + 295, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 251, + 295, + 282 + ], + "spans": [ + { + "bbox": [ + 62, + 251, + 295, + 282 + ], + "type": "text", + "content": "[5] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 285, + 295, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 285, + 295, + 327 + ], + "spans": [ + { + "bbox": [ + 62, + 285, + 295, + 327 + ], + "type": "text", + "content": "[6] Zhongyi Fan, Zixin Yin, Gang Li, Yibing Zhan, and Heliang Zheng. Dreambooth++: Boosting subject-driven generation via region-level references packing. In ACM Multimedia 2024, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 330, + 295, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 330, + 295, + 373 + ], + "spans": [ + { + "bbox": [ + 62, + 330, + 295, + 373 + ], + "type": "text", + "content": "[7] Rinon Gal, Moab Arar, Yuval Atzmon, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Encoder-based domain tuning for fast personalization of text-to-image models. ACM Transactions on Graphics (TOG), 42(4):1-13, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 375, + 295, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 375, + 295, + 418 + ], + "spans": [ + { + "bbox": [ + 62, + 375, + 295, + 418 + ], + "type": "text", + "content": "[8] Rinon Gal, Or Lichter, Elad Richardson, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Lcmlookahead for encoder-based text-to-image personalization. arXiv preprint arXiv:2404.03620, 2024. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 420, + 295, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 420, + 295, + 464 + ], + "spans": [ + { + "bbox": [ + 62, + 420, + 295, + 464 + ], + "type": "text", + "content": "[9] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 465, + 295, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 465, + 295, + 487 + ], + "spans": [ + { + "bbox": [ + 57, + 465, + 295, + 487 + ], + "type": "text", + "content": "[10] Jia Guo and Jiankang Deng. Insightface: 2d and 3d face analysis project. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 488, + 295, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 488, + 295, + 532 + ], + "spans": [ + { + "bbox": [ + 57, + 488, + 295, + 532 + ], + "type": "text", + "content": "[11] Zinan Guo, Yanze Wu, Zhuowei Chen, Lang Chen, and Qian He. Pulid: Pure and lightning id customization via contrastive alignment. Advances in neural information processing systems, 2024. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 534, + 295, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 534, + 295, + 589 + ], + "spans": [ + { + "bbox": [ + 57, + 534, + 295, + 589 + ], + "type": "text", + "content": "[12] Ligong Han, Yinxiao Li, Han Zhang, Peyman Milanfar, Dimitris Metaxas, and Feng Yang. Svdiff: Compact parameter space for diffusion fine-tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7323-7334, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 590, + 295, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 590, + 295, + 633 + ], + "spans": [ + { + "bbox": [ + 57, + 590, + 295, + 633 + ], + "type": "text", + "content": "[13] Soufiane Hayou, Nikhil Ghosh, and Bin Yu. Lora+: Efficient low rank adaptation of large models. In *Forty-first International Conference on Machine Learning*, ICML 2024, Vienna, Austria, July 21-27, 2024, 2024. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "type": "text", + "content": "[14] Zecheng He, Bo Sun, Felix Juefei-Xu, Haoyu Ma, Ankit Ramchandani, Vincent Cheung, Siddharth Shah, Anmol Kalia, Harihar Subramanyam, Alireza Zareian, et al. Imagine yourself: Tuning-free personalized image generation. arXiv preprint arXiv:2409.13346, 2024. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "text", + "content": "[15] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 6" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 3, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 108, + 553, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 553, + 152 + ], + "type": "text", + "content": "[17] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. *J. Mach. Learn.* Res., 23(47):1-33, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 154, + 553, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 154, + 553, + 219 + ], + "spans": [ + { + "bbox": [ + 317, + 154, + 553, + 219 + ], + "type": "text", + "content": "[18] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022, 2022. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 222, + 553, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 222, + 553, + 286 + ], + "spans": [ + { + "bbox": [ + 316, + 222, + 553, + 286 + ], + "type": "text", + "content": "[19] Yuge Huang, Yuhan Wang, Ying Tai, Xiaoming Liu, Pengcheng Shen, Shaoxin Li, Jilin Li, and Feiyue Huang. Curricularface: adaptive curriculum learning loss for deep face recognition. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5901-5910, 2020. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 289, + 466, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 289, + 466, + 300 + ], + "spans": [ + { + "bbox": [ + 316, + 289, + 466, + 300 + ], + "type": "text", + "content": "[20] Insightface. Antelopev2, 2023. 2, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 303, + 555, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 303, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 317, + 303, + 555, + 357 + ], + "type": "text", + "content": "[21] Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up gans for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10124-10134, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 359, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 359, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 317, + 359, + 553, + 392 + ], + "type": "text", + "content": "[22] Diederik Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. Advances in neural information processing systems, 34:21696-21707, 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 394, + 496, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 394, + 496, + 405 + ], + "spans": [ + { + "bbox": [ + 316, + 394, + 496, + 405 + ], + "type": "text", + "content": "[23] LEOSAM. Leosam's helloworld xl, 2023. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 407, + 553, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 407, + 553, + 460 + ], + "spans": [ + { + "bbox": [ + 317, + 407, + 553, + 460 + ], + "type": "text", + "content": "[24] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 464, + 553, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 464, + 553, + 518 + ], + "spans": [ + { + "bbox": [ + 317, + 464, + 553, + 518 + ], + "type": "text", + "content": "[25] Zhen Li, Mingdeng Cao, Xintao Wang, Zhongang Qi, Ming-Ming Cheng, and Ying Shan. Photomaker: Customizing realistic human photos via stacked id embedding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8640-8650, 2024. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 521, + 553, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 521, + 553, + 586 + ], + "spans": [ + { + "bbox": [ + 317, + 521, + 553, + 586 + ], + "type": "text", + "content": "[26] Shih-Yang Liu, Chien-Yi Wang, Hongxu Yin, Pavlo Molchanov, Yu-Chiang Frank Wang, Kwang-Ting Cheng, and Min-Hung Chen. Dora: Weight-decomposed low-rank adaptation. In *Forty-first International Conference on Machine Learning*, ICML 2024, Vienna, Austria, July 21-27, 2024, 2024. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 589, + 553, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 589, + 553, + 610 + ], + "spans": [ + { + "bbox": [ + 316, + 589, + 553, + 610 + ], + "type": "text", + "content": "[27] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 613, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 613, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 317, + 613, + 553, + 656 + ], + "type": "text", + "content": "[28] Jian Ma, Junhao Liang, Chen Chen, and Haonan Lu. Subject-diffusion: Open domain personalized text-to-image generation without test-time fine-tuning. In ACM SIGGRAPH 2024 Conference Papers, pages 1-12, 2024. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 553, + 713 + ], + "type": "text", + "content": "[29] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4296–4304, 2024. 3" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13122" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[30] Lianyu Pang, Jian Yin, Baoquan Zhao, Feize Wu, Fu Lee Wang, Qing Li, and Xudong Mao. Attndreambooth: Towards text-aligned personalized text-to-image generation. arXiv preprint arXiv:2406.05000, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 172 + ], + "type": "text", + "content": "[31] Foivos Paraperas Papantoniou, Alexandros Lattas, Stylianos Moschoglou, Jiankang Deng, Bernhard Kainz, and Stefanos Zafeiriou. Arc2face: A foundation model for id-consistent human faces. In European Conference on Computer Vision, pages 241-261. Springer, 2024. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 173, + 294, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 173, + 294, + 226 + ], + "spans": [ + { + "bbox": [ + 56, + 173, + 294, + 226 + ], + "type": "text", + "content": "[32] Dustin Podell, *Zion English*, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. *Sdxl: Improving latent diffusion models for high-resolution image synthesis.* arXiv preprint arXiv:2307.01952, 2023. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 228, + 294, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 294, + 282 + ], + "type": "text", + "content": "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. 2021. 2, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 284, + 294, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 284, + 294, + 337 + ], + "spans": [ + { + "bbox": [ + 56, + 284, + 294, + 337 + ], + "type": "text", + "content": "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 338, + 294, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 381 + ], + "type": "text", + "content": "[35] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. 2022. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 383, + 294, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 294, + 448 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 294, + 448 + ], + "type": "text", + "content": "[36] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdreambooth: Hypernetworks for fast personalization of text-to-image models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6527-6536, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 449, + 294, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 294, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 294, + 470 + ], + "type": "text", + "content": "[37] Simo Ryu. Low-rank adaptation for fast text-to-image diffusion fine-tuning. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 472, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 472, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 472, + 294, + 525 + ], + "type": "text", + "content": "[38] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 526, + 294, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 526, + 294, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 526, + 294, + 591 + ], + "type": "text", + "content": "[39] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278–25294, 2022. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 593, + 294, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 593, + 294, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 593, + 294, + 645 + ], + "type": "text", + "content": "[40] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 647, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 690 + ], + "type": "text", + "content": "[41] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "text", + "content": "[42] Ming Tao, Hao Tang, Fei Wu, Xiao-Yuan Jing, Bing-Kun Bao, and Changsheng Xu. Df-gan: A simple and effec" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 384 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "text", + "content": "tive baseline for text-to-image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16515-16525, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 554, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 554, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 554, + 129 + ], + "type": "text", + "content": "[43] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 130, + 554, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 554, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 554, + 173 + ], + "type": "text", + "content": "[44] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, and Anthony Chen. Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519, 2024. 1, 3, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 175, + 554, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 554, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 554, + 207 + ], + "type": "text", + "content": "[45] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. 2023. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 209, + 554, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 554, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 554, + 262 + ], + "type": "text", + "content": "[46] Shih-Ying Yeh, Yu-Guan Hsieh, Zhidong Gao, Bernard BW Yang, Giyeong Oh, and Yanmin Gong. Navigating text-to-image customization: From lycoris fine-tuning to model evaluation. In The Twelfth International Conference on Learning Representations, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 264, + 554, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 554, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 554, + 319 + ], + "type": "text", + "content": "[47] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2(3):5, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 320, + 554, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 320, + 554, + 384 + ], + "spans": [ + { + "bbox": [ + 316, + 320, + 554, + 384 + ], + "type": "text", + "content": "[48] Qingru Zhang, Minshuo Chen, Alexander Bukharin, Pengcheng He, Yu Cheng, Weizhu Chen, and Tuo Zhao. Adaptive budget allocation for parameter-efficient finetuning. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023, 2023. 3" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13123" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/HyperNVD_ Accelerating Neural Video Decomposition via Hypernetworks/f624369a-7a34-4be6-8ddf-7450e2fca80c_content_list.json b/2025/HyperNVD_ Accelerating Neural Video Decomposition via Hypernetworks/f624369a-7a34-4be6-8ddf-7450e2fca80c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e6d0c2602998e98797d12e642802e018e3bee2bc --- /dev/null +++ b/2025/HyperNVD_ Accelerating Neural Video Decomposition via Hypernetworks/f624369a-7a34-4be6-8ddf-7450e2fca80c_content_list.json @@ -0,0 +1,1420 @@ +[ + { + "type": "text", + "text": "HyperNVD: Accelerating Neural Video Decomposition via Hypernetworks", + "text_level": 1, + "bbox": [ + 124, + 130, + 875, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Maria Pilligua1,2,* Danna Xue1,2,*† Javier Vazquez-Corral1,2 \n1 Universitat Autònoma de Barcelona 2 Computer Vision Center {mpilligua, dxue, javier.vazquez}@cvc.uab.cat", + "bbox": [ + 233, + 179, + 761, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Decomposing a video into a layer-based representation is crucial for easy video editing for the creative industries, as it enables independent editing of specific layers. Existing video-layer decomposition models rely on implicit neural representations (INRs) trained independently for each video, making the process time-consuming when applied to new videos. Noticing this limitation, we propose a meta-learning strategy to learn a generic video decomposition model to speed up the training on new videos. Our model is based on a hypernetwork architecture which, given a video-encoder embedding, generates the parameters for a compact INR-based neural video decomposition model. Our strategy mitigates the problem of single-video overfitting and, importantly, shortens the convergence of video decomposition on new, unseen videos. Our code is available at: https://hypernvd.github.io/.", + "bbox": [ + 89, + 300, + 485, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 569, + 220, + 585 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video editing plays a crucial role in today's creative industries, from film-making and advertising to social media content creation. It enables creators to bring their visions to life, enhancing storytelling through visual effects, scene adjustments, and content personalization. However, most current video editing techniques face several challenges: they require extensive manual intervention [39], lack consistent control over edits [7, 11], and are time-consuming [4, 24], limiting their scalability for real-world applications.", + "bbox": [ + 88, + 594, + 482, + 729 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Current video editing workflows in professional software resemble traditional 2D animation. Creators import video into editing software [1, 3] and edit in overlaid layers for each video frame. While motion tracking tools can interpolate adjustments between keyframes [10, 18], handling occlusions and ensuring consistency remain challenging since tracking errors make editing cumbersome. These problems highlight the importance of accurate layered representation and frame mappings.", + "bbox": [ + 88, + 731, + 482, + 867 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/35f53945dc588cf7f23ca3d6e777a5bfb6f42e01a6a1044c3f156ed42fbcc19c.jpg", + "image_caption": [ + "Figure 1. Fine-tuning the video decomposition model from our metamodel (HyperNVD) versus training from scratch on unseen videos shows clear advantages. With initialization from our HyperNVD trained on 15 videos, the model converges faster to the same PSNR and ultimately achieves better performance." + ], + "image_footnote": [], + "bbox": [ + 516, + 268, + 903, + 465 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Layer-based video decomposition methods [4, 20, 24, 31, 38] represent video with several texture layers, where each layer consistently represents specific video content (e.g., moving objects or backgrounds). These approaches are mainly based on compact implicit neural representations (INRs), which encode each video individually. The model projects pixel coordinates and frame indices into a canonical 2D texture space, and then, it maps to RGB values to reconstruct each frame. In this way, it is possible to apply simple edits on the canonical texture layer and then propagate them to the entire video, greatly relieving the need for manual adjustments. Furthermore, the parameter sizes and training times are reduced compared to training a generalized video object segmentation model [2] and global flow estimation model [8, 16]. However, due to the nature of INR, the neural video decomposition model lacks generalization and requires tens of minutes to train a $480\\mathrm{p}$ video [4].", + "bbox": [ + 511, + 551, + 906, + 824 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we propose a meta-learning strategy in layer-based video decomposition to speed up the training. More in detail, we introduce a hypernetwork (hypernet) that, based on the video embedding predicted by VideoMAE [34], generates the parameters for a compact INR", + "bbox": [ + 511, + 825, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*These authors contributed equally to this work.", + "bbox": [ + 107, + 875, + 364, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22933", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "model. Our strategy facilitates the generalization to more than one video and also allows for speed-up training on new, unseen videos, as we can use the parameters given by the hypernet to initialize the neural video decomposition model for further fine-tuning. Figure 1 shows three examples of this speed-up. In this figure, we compare the training time and the PSNR obtained for training an unseen video starting from the parameters suggested by our hypernet or from a random parameter set. Our method obtains the same PSNR with more than 30 minutes less.", + "bbox": [ + 89, + 90, + 480, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contributions. Our main contributions are as follows:", + "bbox": [ + 89, + 250, + 452, + 263 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We incorporate a meta-learning strategy to learn a generic video decomposition method. Our strategy consists of learning a hypernet that, given a video-encoder embedding, predicts the parameters for an INR-based video decomposition model.", + "- Our framework relieves the problem of overfitting on a single video and shortens the convergence time of the neural video decomposition model on unseen videos.", + "- The experimental results show the generalizability and appealing video editing ability of our approach." + ], + "bbox": [ + 91, + 267, + 482, + 417 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 430, + 232, + 445 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Layered video decomposition", + "text_level": 1, + "bbox": [ + 89, + 455, + 349, + 470 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In image editing and animation, an image is often separated into multiple layers, each with color and opacity [30, 37], allowing independent adjustments before stacking them for the final result. This layered representation is also applied to video editing for simplifying the workflow and ensuring consistency [15, 28, 36]. Neural video decomposition approaches [4, 17, 38] decompose videos into consistent 2D texture layers across frames and are usually trained in a self-supervised manner using motion and appearance constraints. LNA [17] uses implicit neural representations to reconstruct videos via neural atlases and alpha blending. While encoding each video individually reduces parameter size and rendering time, reconstructing high-frequency details and long training time remain challenges. Omnimatte [21] and Factormatte [12] add appearance priors to model associated effects like shadows, or reflections among others. Hashing-nvd [4] and CoDeF [26] introduce hash encoding to accelerate optimization. However, these models still need more than 40 minutes of training on one single video [4]. We propose to use a hypernet to generate parameters of the video decomposition INR. In this way, we can further shorten the training time of any new, unseen video by initializing its training from the weights given by the hypernet.", + "bbox": [ + 89, + 477, + 482, + 840 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Hypernetwork", + "text_level": 1, + "bbox": [ + 89, + 848, + 238, + 864 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Hypernetworks (or hypernets) [5, 13] is a technique that enhances the adaptability and flexibility of deep learn", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ing models. Instead of learning task-specific models with fixed parameters, a hypernet dynamically generates parameters for a target model based on task [9, 25, 35], or data [19, 22, 29, 32]. By learning a shared, adaptable representation, hypernets help target models generalize better across different conditions. Once trained, they use embeddings as input to modulate the target model's weights and reduce optimization times on unseen tasks [9, 25]. Recent work has further improved input-specific model flexibility, especially in INR applications [6, 19, 22]. The work most closely related to ours is HyP-NeRF [29], which employs a hypernet to generate instance-specific parameters for hash encoding and NeRF, linking text to 3D object reconstructions. This framework relates semantic similarities across models for different instances but focuses on simple 3D objects with limited shape and color variations. In contrast, our video modeling task is more complex and requires attention to multiple objects, detailed motion, and varied appearances.", + "bbox": [ + 511, + 90, + 903, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Accelerate Neural Video Decomposition via Hypernetwork", + "text_level": 1, + "bbox": [ + 511, + 391, + 903, + 426 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Framework overview", + "text_level": 1, + "bbox": [ + 511, + 433, + 712, + 448 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our goal is to learn a generalizable model for neural video decomposition that helps speed up the optimization of unseen videos. Our HyperNVD consists of three main components: i) a video embedding generation model, ii) a hypernet to generate model parameters, and iii) a neural video decomposition (NVD) model. Figure 2 shows an overview of our model. First, we compress the features predicted by a pre-trained video transformer, VideoMAE [34], and use the compressed embedding as input to the hypernet. Then the hypernet predicts the weights and the multiresolution hash encoding parameters of the target NVD model. Our architecture allows the training of multiple videos simultaneously, which enables fast adaptation of unseen videos. This fast adaptation is accomplished by considering the parameters given by the hypernet for further fine-tuning.", + "bbox": [ + 511, + 455, + 906, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural video decomposition (NVD) model. Standard layered video decomposition models based on INR [4] work by taking video coordinates and frame indices $p = (x, y, t)$ as input and decomposing the video into a background layer and a foreground layer.", + "bbox": [ + 511, + 689, + 903, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As shown in the Layer Module of Figure 2, the foreground layer representation $LM_{f}$ is encoded by a mapping module $\\mathcal{M}_f$ , a texture module $\\mathcal{T}_{f}$ , and a residual module $\\mathcal{R}_f$ . The mapping module $\\mathcal{M}f$ transforms the input coordinate $p$ to the texture space coordinate $(u,v)_f^p$ . The texture module $\\mathcal{T}_f$ converts the texture coordinate into an RGB value $c_{f}$ , while the residual module $\\mathcal{R}_f$ predicts a framewise residual coefficient $r_f$ . Only $\\mathcal{T}f$ and $\\mathcal{R}f$ incorporate Multi-resolution Hashing Encoding (MRHE) to accelerate", + "bbox": [ + 511, + 765, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22934", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/30f300b8457157da4a08ad10ad736aa3390c9c5674a9fa58240cdb61da9b12f3.jpg", + "image_caption": [ + "Figure 2. The architecture of our HyperNVD. HyperNVD consists of $i)$ the MAE encoder to generate video embedding, $ii)$ the hypernet to generate model parameters, and $iii)$ the target neural video decomposition (NVD) model. Given an input video, the MAE encoder encodes it to a compact embedding $e$ . Then the hypernet $\\mathcal{H}$ generates the parameters of the NVD model, including the MultiResolution Hash Encoding (MRHE) and the model weights. The NVD model includes two layer modules (foreground and background) for reconstructing different components in the video and an alpha module for predicting the soft mask to blend the layers. The reconstructed frame is generated by adding the foreground and background texture layers with an opacity map. Each layer module consists of a mapping module, a texture module, and a residual module." + ], + "image_footnote": [], + "bbox": [ + 112, + 85, + 885, + 404 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "training and inference. The MRHE (illustrated in the rightmost block of Figure 2) performs grid hash encoding [23]. Here, the input 2D features serve as coordinates within multi-resolution grids, allowing for interpolation-based data sampling. The sampled data is then concatenated and processed through MLPs. The final output of the foreground layer is computed as $\\hat{c}_f^p = c_f^p * r_f^p$ . The background layer follows the same structure.", + "bbox": [ + 88, + 531, + 480, + 651 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The background and foreground representations are later combined by an alpha module $\\mathcal{A}$ . This module takes the coordinate $p$ as input and predicts a soft opacity map for each layer. The final reconstructed RGB value, $\\hat{c}^p$ , is obtained by weighted summation across both layers:", + "bbox": [ + 89, + 652, + 482, + 728 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {c} ^ {p} = \\left(1 - \\alpha^ {p}\\right) \\hat {c} _ {b} ^ {p} + \\alpha^ {p} \\hat {c} _ {f} ^ {p}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 199, + 741, + 482, + 760 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "All the modules in the above model structure are MLPs.", + "bbox": [ + 89, + 771, + 460, + 785 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Note that our explanation can easily be switched to multiple foreground layers based on the masks provided by users; we explained the case of a single foreground layer for simplification.", + "bbox": [ + 89, + 787, + 483, + 848 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Hypernet. Our hypernet, $\\mathcal{H}$ , predicts the INR parameters given a condition embedding $e$ —see the Hypernet block in Figure 2. It comprises a series of MLPs, each dedicated", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "to generating the parameters of a specific layer within the NVD model. The target of $\\mathcal{H}$ is to generate accurate weights for an NVD model capable of robust video decomposition. To accomplish this, we train it using the same loss functions and constraints applied in single-video decomposition models, as the objective for each video remains unchanged. However, by conditioning the hypernet on unique embeddings for each video, the model learns a generalized decomposition capability that extends beyond individual instances.", + "bbox": [ + 511, + 531, + 906, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "VideoMAE encoder. Typically, hypernet inputs use either learnable embeddings trained jointly with the hypernet or embeddings generated from input data. We use the latter, as representing high-dimensional video data in a compact, low-dimensional embedding is particularly challenging. This approach also supports generalization on multiple videos and enables inference on unseen videos, which is crucial for building a robust NVD model. Moreover, this video embedding needs to encapsulate both motion and scene information, help reduce optimization complexity, and accelerate network convergence.", + "bbox": [ + 511, + 696, + 908, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To achieve this, we use the features predicted by VideoMAE [34], a frozen self-supervised masked autoencoder", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22935", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0623f59a7f60469f76d075b1a83f4243b2ba9715297e28cb1ae3ed45aa6b036d.jpg", + "image_caption": [ + "Figure 3. Training of the autoencoder to get a compressed embedding from VideoMAE. First, we extract features $o$ from each 3D patch using the pre-trained VideoMAE. We then train an autoencoder to obtain a compressed embedding $e$ by minimizing the difference between the output $\\hat{o}$ and the input $o$ ." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 483, + 170 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "trained on video data. As shown in Figure 3, to further reduce the dimension of the VideoMAE output, we use an encoder-decoder structure to learn a compressed embedding of size (768, 1) from the original VideoMAE output of size (768, #patches). This ensures the encoder consolidates information from all patches into a single dense embedding while minimizing information loss. The autoencoder is trained using $L_{1}$ loss to enforce the closeness between the reconstructed output $\\hat{o}$ and the original embedding $o$ .", + "bbox": [ + 88, + 270, + 482, + 406 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We take the VideoMAE together with the encoder part as our VideoMAE encoder in HyperNVD (the top-left module in Figure 2). This encoder is pre-trained and frozen during the hypernet training, allowing us to maintain the rich representations it learned while preserving essential information for the decomposition tasks. In this way, the encoder serves as a compact, fixed embedding provider for the hypernet.", + "bbox": [ + 89, + 406, + 483, + 512 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Training and rendering procedure", + "text_level": 1, + "bbox": [ + 89, + 522, + 390, + 539 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Loss functions. Since all the components of our model are fully differentiable, we train them end-to-end via stochastic gradient descent. We use losses from the previous work [4, 17] to facilitate the training process. The losses include:", + "bbox": [ + 89, + 545, + 482, + 604 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Reconstruction loss to ensure the reconstruction quality of the video.", + "- Consistency loss which provides supervision of optical flow for accurate motion representation.", + "- Sparsity loss to prevent duplicate content in different texture layers.", + "- Residual consistency loss to ensure smooth lighting conditions while preventing the residual estimator from absorbing all color details." + ], + "bbox": [ + 89, + 606, + 482, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Two losses are applied just in the initial iterations:", + "bbox": [ + 89, + 742, + 423, + 757 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Rigidity loss to maintain the structural rigidity in the canonical texture layer.", + "- Alpha bootstrapping loss, which supervises coarse masks for foreground separation." + ], + "bbox": [ + 91, + 757, + 482, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since the model is trained in a self-supervised manner, the additional supervision we use are the ground truth object masks and the optical flow predicted by RAFT [33]. Further details are in the supplementary materials.", + "bbox": [ + 89, + 816, + 482, + 878 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training with the hypernet. The objective is to train a", + "bbox": [ + 89, + 885, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "hypernet capable of generating parameters for the NVD model, so only the hypernet weights are trainable, while the NVD model serves as a set of differentiable layers that enable backpropagation to the hypernet, allowing it to learn from the above-mentioned losses. We begin the training process with a pre-training step that configures the mapping network to predict textures that are both aligned and in an initial rectangular shape corresponding to the frame size. Without this initial step, the model could produce incorrectly oriented textures. With this pre-training, the convergence during the subsequent training phase is significantly faster, as the model starts from a more informed initialization.", + "bbox": [ + 511, + 90, + 906, + 287 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fine-tune on new videos. Once our model is trained on a set of videos, we can use the hypernet to generate parameters for a new video and treat these weights as an initialization for the NVD model (see Figure 1 and Section 4.2).", + "bbox": [ + 511, + 295, + 906, + 356 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Video editing and rendering. The NVD model decomposes an input video into multiple layers, each with distinct textures. For video editing, we first perform inference to obtain these texture layers. Then, each layer can be edited individually with image editing software or other techniques. For editing, we render the texture networks using a $1000 \\times 1000$ grid sampling. The rendered texture is then modified, and the colors are sampled through bilinear interpolation for smooth transitions and precise adjustments. By replacing the original layers with the modified ones and mapping the corresponding coordinates back to each video frame, we can generate the edited video with temporally consistent changes.", + "bbox": [ + 511, + 362, + 908, + 559 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 574, + 648, + 590 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Implementation details", + "text_level": 1, + "bbox": [ + 511, + 599, + 728, + 614 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Datasets. We conduct our experiments mainly on the DAVIS dataset [27]. We train HyperNVD with up to 30 videos, while the rest videos are reserved for testing. Each video consists of the first 16 frames at a resolution of $768 \\times 432$ .", + "bbox": [ + 511, + 621, + 906, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training settings. We use a batch size of 100,000 point locations and train each model for around 25,000 iterations and 100 pre-train iterations. The hypernet consists of approximately 290 million learnable parameters. The main network consists of approximately 4.4 million learnable parameters: 1.75 million for each layer module and 965 thousand for the alpha module. Training and evaluation with 1 video and a batch size of 100,000 requires around 22GB of GPU memory, with a checkpoint size of 2.3GB. Training our model takes roughly 2.8 hours for 1 video and 12.5 hours for 3 videos on an NVIDIA RTX 3090 GPU. We optimize all networks simultaneously using the Adam optimizer, with a learning rate of 1e-2 for the hash encoding", + "bbox": [ + 511, + 704, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22936", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/c10367415ef31265382ccd0e0ab6316842557fb16001a33c5b5422c290e983ef.jpg", + "table_caption": [ + "Table 1. Comparison with previous approaches on reconstruction quality for single video training." + ], + "table_footnote": [], + "table_body": "
Methodshikeblackswancar-turnbearlucia
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
Deformable Sprites [38]25.300.79525.750.73230.740.85026.340.81626.540.771
Layered Neural Atlases [17]30.020.88229.960.85334.390.92929.620.91029.920.888
Hashing-nvd [4]29.120.87330.300.84335.430.89531.560.90829.290.866
HyperNVD (Ours)30.060.84629.530.81034.940.90931.580.90130.300.865
", + "bbox": [ + 151, + 109, + 841, + 200 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/67dc6a1be9c14c56fb7915d8995a721800b73395db238721dc1311c5fda323ac.jpg", + "table_caption": [ + "Table 2. Comparison of the reconstruction quality of our model when trained with different numbers of videos." + ], + "table_footnote": [], + "table_body": "
Num of videoshikeblackswancar-turnbearluciakite-walkbusstrollerboatrollerblade
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
130.060.84629.530.81034.940.90931.580.90130.300.86533.350.83732.450.88931.250.83032.770.87228.750.810
327.310.72827.670.75431.870.850--------------
527.130.76027.320.73431.430.84528.880.85427.430.760----------
1027.990.75428.860.73632.470.85229.990.85328.500.76331.750.79729.450.83428.800.77531.070.82826.060.740
1527.650.76827.410.75632.570.87829.710.85228.540.83331.990.79629.600.84428.920.78030.910.82626.130.742
3027.240.79527.300.74531.640.86429.280.86227.310.80230.930.77128.630.82727.000.72029.980.80324.350.686
", + "bbox": [ + 94, + 234, + 901, + 338 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and a learning rate of 5e-4 for the others.", + "bbox": [ + 89, + 364, + 359, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Details of architecture. In the target NVD model, we apply Rectified Linear Unit (ReLU) activations between hidden layers, with outputs passed through a tanh function. This enables the model to map texture to a predefined continuous range. The hypernet comprises a series of MLPs. Each MLP consists of four fully connected layers, with a hidden dimension of 64 and intermediate ReLU activations. Further details about the architecture can be found in the supplementary material.", + "bbox": [ + 89, + 387, + 482, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Quantitative results", + "text_level": 1, + "bbox": [ + 89, + 536, + 277, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Comparison with previous methods. We present the comparison of PSNR and SSIM values between DeformableSprites [38], Layered Neural Atlases [17], Hashing-nvd [4] and our HyperNVD in Table 1. We report the results of our model trained on one single video independently. For a fair comparison, we use the ground truth mask as initial guidance for all the comparing methods. Results in Table 1 demonstrate that incorporating the hypernet does not degrade the model's performance. Ours maintains the effectiveness of the original architecture (i.e. our baseline Hashing-nvd) while introducing the flexibility of a hypernet-based approach.", + "bbox": [ + 89, + 559, + 482, + 742 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Learning multiple videos. To evaluate the model's generalization ability across multiple videos, we train HyperNVD with 1, 3, 5, 10, 15, and 30 videos, respectively. For consistency, videos included in smaller-scale settings are subsets of those in larger-scale settings. Table 2 shows the PSNR of the videos reconstructed by models trained with different numbers of videos. When training with multiple videos, we observe a slight drop in reconstruction quality compared to training on a single video. This is likely due to the model potentially overfitting to the unique characteristics of a sin", + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "gle video. However, the PSNR only decreases by approximately 3dB as the number of videos increases from 1 to 30, which indicates that the model retains strong performance even with larger datasets.", + "bbox": [ + 511, + 364, + 903, + 424 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 4 further illustrates the image quality of the \"hike\" video when training HyperNVD with different numbers of videos. The models trained on multiple videos can still reconstruct the low-frequency parts well, but some high-frequency details are lost. This stable evolution of PSNR indifferent to the number of videos indicates a plateau in performance with larger video sets. Figure 5 shows the reconstruction quality of the video \"hike\" across different iteration steps. The resulting curves are very similar for all the cases trained with multiple videos.", + "bbox": [ + 511, + 424, + 905, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finetuning on new videos. To evaluate the ability of fast adaption on unseen videos, we train the NVD model initialized by the parameters of our metamodel trained with 15 videos and compare it with the model trained from scratch. Note that the NVD model is trained independently without the hypernet. As shown in Table 3, training from our metamodel (i.e. HyperNVD) consistently improves performance by $0.8\\mathrm{dB}$ across different videos compared to training from scratch (i.e. Hashing-nvd). Figure 1 further shows our model reaches the same PSNR on the unseen videos more than 30 minutes faster, and with more stable training. This demonstrates that our metamodel enables faster adaptation to new scenes, which improves the efficiency of video editing workflow. Additional results are provided in the supplementary material.", + "bbox": [ + 511, + 582, + 905, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Qualitative results", + "text_level": 1, + "bbox": [ + 511, + 818, + 691, + 834 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 6 shows examples of our layer decomposition results. For each video frame, we include a) the original frame, b) the predicted background and foreground components with a checkerboard overlay to illustrate texture", + "bbox": [ + 511, + 840, + 903, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22937", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/928622741847b16fd477c88f7746089bc5908473bc8e054c68902b414e834b17.jpg", + "table_caption": [ + "Table 3. Reconstruction quality in PSNR on unseen videos when fine-tuning from our 15-video metamodel versus training from scratch." + ], + "table_footnote": [], + "table_body": "
Methodscowssoccerballsoapboxfamingohorse-jumpgoatswingkite-surflibbysurf
From scratch27.0428.5131.2829.0327.6127.0629.6531.6529.6333.61
From metamodel28.4030.1632.7429.9628.5027.7830.3732.1930.1133.86
Improvements+1.36+1.65+1.46+0.93+0.89+0.72+0.72+0.54+0.48+0.25
", + "bbox": [ + 91, + 111, + 903, + 176 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/06978cdae851dd78eed78193589a8685d03eff59fa2016a09f85d19bfc945cd2.jpg", + "image_caption": [ + "Figure 4. Comparison of the video \"hike\" reconstructed by models trained with different numbers of videos. The top row shows the full frames, while the bottom row zooms in on a specific patch. Models trained on multiple videos capture the primary structure of the frame but tend to lose some high-frequency details." + ], + "image_footnote": [], + "bbox": [ + 91, + 200, + 483, + 309 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e9c832804d1ea254f6e3106c82e5ea2de5365b79e71f8b89d89fd56914c50725.jpg", + "image_caption": [ + "Figure 5. Comparison of the reconstruction quality for the \"hike\" video trained with different numbers of videos. Our HyperNVD model trained on multiple videos shows a 3dB drop in PSNR compared to single-video training. The PSNR remains stable across multiple video training, suggesting a performance plateau." + ], + "image_footnote": [], + "bbox": [ + 93, + 407, + 480, + 539 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "mapping transformations, c) the predicted residual maps, d) the complete canonical texture map, and e) the reconstructed frame. Our model obtains accurate video reconstruction without introducing noticeable texture distortions. The predicted opacity map separates the foreground and background correctly. The residuals learn the changing reflections on the water and shadows as shown in the \"black-swan\" and \"boat\" videos.", + "bbox": [ + 89, + 642, + 482, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To further showcase the effectiveness of our approach for editing, we demonstrate editing examples in Figure 7. In the first two examples, only the background is modified, while in the last example, we edit the foreground texture. For the first example, the \"blackswan\" video, we transform the plants in the background into flowers and adjust the color of the water. The ripples and shadows in the water maintain a natural appearance, and the transitions at the boundaries between the swan and the water are seamless. In the sec-", + "bbox": [ + 89, + 763, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1087a4022c279b99fa3e090b4c46d635baedad32a872bcfdae97b867d00948df.jpg", + "table_caption": [ + "Table 4. Ablations on the input video embedding and learning the MRHE with or without hypernet. 'Num' represents the number of videos used for training. Using the MAE encoder generated embedding is better than a learnable embedding jointly optimized with the hypernet. The video-specific MRHE encodings predicted by the hypernet outperform a universal MRHE encoding." + ], + "table_footnote": [], + "table_body": "
NumEmbed w MAEMRHE w hyperhikecar-turnblackswan
PSNRSSIMPSNRSSIMPSNRSSIM
1X✓30.170.86734.040.90529.950.822
✓X30.230.85833.470.88128.830.797
✓✓30.060.84634.940.90929.530.810
3X✓26.370.76029.980.82926.070.684
✓X26.880.71230.010.78826.640.707
✓✓27.310.72831.870.85027.670.754
", + "bbox": [ + 514, + 292, + 903, + 382 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ond video (\"kite-walk\"), we alter the beach background to a sunset style. In the \"hike\" video, we change the color of the person's shirt to yellow and add text on the backpack. Additional results of decomposition and editing are available in the supplementary material.", + "bbox": [ + 511, + 412, + 905, + 489 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Ablation study", + "text_level": 1, + "bbox": [ + 511, + 503, + 663, + 520 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We do ablations on using different embedding as the input of the model, on learning the multi-resolution hash encoding with and without hypernet, and on using different types of masks as the initial supervision.", + "bbox": [ + 511, + 527, + 905, + 589 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "VideoMAE embedding v.s. Learnable embedding. A well-designed embedding is crucial for a hypernet to generate effective model parameters. Some related works [4, 22] employ a learnable embedding optimized jointly with the hypernet, while we directly grab the embedding from the pre-trained MAE encoder and freeze it during hypernet training. We compare the results of using an optimized embedding versus an embedding generated by the MAE encoder, as shown in the first and third rows of Table 4. The advantage in PSNR for the VideoMAE embedding when training for three videos is more than $1\\mathrm{dB}$ for the 3-video model. For single-video training, the impact of using either embedding type on the final video quality is minimal. Thus, the advantage of using the VideoMAE Encoder becomes more evident when applied to multiple videos. VideoMAE embedding also shows superiority in unseen videos when using the 3-video metamodel as initialization (see the supplementary materials). Furthermore, using VideoMAE enables robust generalization to new, unseen videos, which is essential for the versatility of our model.", + "bbox": [ + 511, + 598, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22938", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f2161798438d153bda39dd8b751a232b936b76521c12b28f8ab623381dc77e8b.jpg", + "image_caption": [ + "Figure 6. Qualitative results on the DAVIS dataset. Displayed are the a) original frames, b) predicted masks, c) residual maps, d) texture maps, e) reconstructed frames, and PSNR values for different videos. A color checkerboard overlay is applied to the masked areas to visualize texture transformations." + ], + "image_footnote": [], + "bbox": [ + 89, + 89, + 903, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Video-based MRHE learned with hypernet v.s. Universal MRHE without hypernet. In this part, we show the benefits of learning multiresolution hash encoding with the hypernet. In multiresolution hash encoding, input 2D features are treated as coordinates within multi-resolution grids and used to sample data from these grids via interpolation. The sampled data are then concatenated and processed through MLPs to generate the final output. In Table 4, we compare the results of i) using hypernet to learn the parameters of the MLPs making these parameters unique for each video (third rows) and ii) directly using the main NVD net to learn MRHE parameters making them common for all videos (second rows). The higher quality results are ob", + "bbox": [ + 89, + 702, + 485, + 898 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tained when MRHE is learned by the hypernet, with benefits that become more apparent as the number of videos used for training increases. This improvement stems from the fact that different video scenes exhibit varying levels of complexity, making video-specific hash grid parameters beneficial when training across multiple videos.", + "bbox": [ + 511, + 702, + 908, + 796 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Accurate mask v.s. Coarse mask. The reference mask is important in our task since it provides a clue about the moving salient objects. We compare the use of manually annotated precise masks (groundtruth from the DAVIS dataset [27]) with coarsely predicted masks from MaskR-CNN [14] as the initial reference for model training. The", + "bbox": [ + 511, + 809, + 910, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22939", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/860a806acd40860df473310c3257e16465ad87b66a4fffea4eb070e29a3d12ad.jpg", + "image_caption": [ + "a) Reconstructed frame" + ], + "image_footnote": [], + "bbox": [ + 93, + 107, + 245, + 175 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/96c32d2d98d118896a81eaeff01f0fe87deeeafa20541f7f8c43576c65f373ba.jpg", + "image_caption": [ + "b) Output 2D textures" + ], + "image_footnote": [], + "bbox": [ + 253, + 107, + 493, + 175 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f226423c8c65e8f647b0602b88091db8e6d2d70bf77d8db774ccc80cf8c29072.jpg", + "image_caption": [ + "c) Edited 2D textures" + ], + "image_footnote": [], + "bbox": [ + 504, + 107, + 745, + 175 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e6cf06e38c69535d9e3a050707bd1fe63f478eb4c5402c210ea4919dd6d0b96e.jpg", + "image_caption": [ + "d) Edited frame" + ], + "image_footnote": [], + "bbox": [ + 753, + 107, + 903, + 175 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ad892634bc5b026b4e79f8f55d45675ec4be5ca506ae599df7cc4e65c63ae4a2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 93, + 181, + 245, + 250 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1501ab164303ef0bc4a7121ab6c73f4f875b23db2e2ed336240546ba27b560cf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 251, + 181, + 493, + 250 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/abc676660e3007ec115fcd879825f60c6dc8e64e2f7487e2f285e47656cbcc15.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 503, + 181, + 745, + 250 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/878aac63c72d77348d8ac237a45a7e2dfb86ec1b9400d45885d879835aed54b0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 751, + 181, + 903, + 250 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/04ccf7e33a0e0ff0f9cd9a04f649aab1798f2d7de7e3b255f661ec4975c7bef1.jpg", + "image_caption": [ + "Figure 7. Edited video results. We make adjustments to one of the texture layers—the background layers (rows 1 and 2), and the foreground layer (row 3)—while keeping the other layer unchanged and blending the edited layers together. The edited video maintains good visual quality and smooth transitions." + ], + "image_footnote": [], + "bbox": [ + 93, + 260, + 245, + 328 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/28559146cd6cbe1b15ecdbbf07c1cc5694e0a22257445cddb9a6058495e10658.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 251, + 260, + 493, + 328 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e1934625b7583635e28f6f2030aac9c6d825527e998e63df489581dd43d0f036.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 503, + 260, + 745, + 327 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b74eb4709b106020eeac7faf5dc599523313455fb3119c7eb29d7d745f0dd89a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 751, + 260, + 903, + 327 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c0c8a48fe53da6667216841c0a58783aa3616ffabf22c2205f85579f37df23f0.jpg", + "table_caption": [ + "Table 5. Ablation on using different masks as initial guidance." + ], + "table_footnote": [], + "table_body": "
Maskhikecar-turnblackswan
PSNRSSIMPSNRSSIMPSNRSSIM
GT30.060.84634.940.90929.530.810
MaskRCNN29.610.84634.920.90329.660.814
", + "bbox": [ + 98, + 420, + 470, + 483 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "predicted alpha map is a soft mask that always includes the corresponding effects, such as reflection or shadows accompanied by the moving objects. Therefore, it is always different from the original object mask. Results in Table 5 show that our model is robust to masks with varying levels of accuracy. Figure 8 illustrates two frames from the \"bear\" and \"blackswan\" videos, where a more precise initial mask improves object boundary details. For example, using a more refined mask as initial guidance help the alpha map to capture intricate structures, such as the feathers on a swan's tail. Thus, for higher-quality editing results, it is essential to use the most detailed masks possible during model training.", + "bbox": [ + 89, + 510, + 483, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Concluding Remarks", + "text_level": 1, + "bbox": [ + 89, + 707, + 292, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present a framework for multiple video training of INR-based video decomposition models using a hypernetwork. We reformulate the task of training a neural video decomposition model on a single video into learning a hypernetwork that generates network parameters based on video embeddings. This approach effectively mitigates the overfitting of the neural video decomposition model to individual videos, enhances generalization capability, and speeds up adaptation to unseen videos. Through experiments, we validate that our proposed model achieves comparable video layer decomposition and rendering capabilities to previous", + "bbox": [ + 89, + 734, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ec1e1a9b5617ad96e148be95a14de72f083029de09603a316fac223bb788c21f.jpg", + "image_caption": [ + "Figure 8. Comparison between using accurate and coarse masks as initial guidance. Using different masks does not significantly affect the reconstruction results, but providing a refined mask as the initial guidance helps the model achieve more detailed decomposition, which can be especially beneficial for finer edits." + ], + "image_footnote": [], + "bbox": [ + 517, + 400, + 901, + 636 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "methods, while significantly reducing training time on new videos and improving video editing efficiency.", + "bbox": [ + 511, + 737, + 903, + 767 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. We acknowledge the Departament de Recerca i Universitats with ref. 2021SGR01499 and the CERCA Program from Generalitat de Catalunya, Grant PID2021-128178OB-I00 funded by MCIN/AEI/10.13039/501100011033, ERDF \"A way of making Europe\", and the grant Catedra ENIA UAB-Cruilla (TSI-100929-2023-2) from the Ministry of Economic Affairs and Digital Transformation of Spain.", + "bbox": [ + 511, + 777, + 906, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22940", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Adobe. After effects. Accessed: 2024-11-07. 1", + "[2] Ali Athar, Alexander Hermans, Jonathon Luiten, Deva Ramanan, and Bastian Leibe. Tarvis: A unified approach for target-based video segmentation. In CVPR, 2023. 1", + "[3] BlackMagicDesign. Davinci resolve 19. Accessed: 2024-11-07. 1", + "[4] Cheng-Hung Chan, Cheng-Yang Yuan, Cheng Sun, and Hwann-Tzong Chen. Hashing neural video decomposition with multiplicative residuals in space-time. In ICCV, 2023. 1, 2, 4, 5, 6", + "[5] Vinod Kumar Chauhan, Jiandong Zhou, Ping Lu, Soheila Molaei, and David A Clifton. A brief review of hypernetworks in deep learning. Artificial Intelligence Review, 57(9): 250, 2024. 2", + "[6] Pei-Ze Chiang, Meng-Shiun Tsai, Hung-Yu Tseng, Wei-Sheng Lai, and Wei-Chen Chiu. Stylizing 3d scene via implicit representation and hypernetwork. In WACV, 2022. 2", + "[7] Yuren Cong, Mengmeng Xu, Christian Simon, Shoufa Chen, Jiawei Ren, Yanping Xie, Juan-Manuel Perez-Rua, Bodo Rosenhahn, Tao Xiang, and Sen He. Flatten: optical flow-guided attention for consistent text-to-video editing. In ICLR, 2023. 1", + "[8] Carl Doersch, Ankush Gupta, Larisa Markeeva, Adria Recasens, Lucas Smaira, Yusuf Aytar, Joao Carreira, Andrew Zisserman, and Yi Yang. Tap-vid: A benchmark for tracking any point in a video. NeurIPS, 2022. 1", + "[9] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In ICML, 2017. 2", + "[10] Foundry. Nuke. Accessed: 2024-11-07. 1", + "[11] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. In ICLR, 2024. 1", + "[12] Zeqi Gu, Wenqi Xian, Noah Snavely, and Abe Davis. Factormatte: Redefining video matting for re-composition tasks. ACM TOG, 42(4):1-14, 2023. 2", + "[13] David Ha, Andrew Dai, and Quoc V Le. Hypernetworks. In ICLR, 2017. 2", + "[14] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, 2017. 7", + "[15] Nebojsa Jojic and Brendan J Frey. Learning flexible sprites in video layers. In CVPR, 2001. 2", + "[16] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together. In ECCV, 2024. 1", + "[17] Yoni Kasten, Dolev Ofri, Oliver Wang, and Tali Dekel. Layered neural atlases for consistent video editing. ACM TOG, 40(6):1-12, 2021. 2, 4, 5", + "[18] KeenTools. Geotracker for adobe after effects. Accessed: 2024-11-07. 1", + "[19] Chiheon Kim, Doyup Lee, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Generalizable implicit neural representations via instance pattern composers. In CVPR, 2023. 2", + "[20] Jaekyeong Lee, Geonung Kim, and Sunghyun Cho. Rna: Video editing with roi-based neural atlas. ACCV, 2024. 1" + ], + "bbox": [ + 93, + 114, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[21] Erika Lu, Forrester Cole, Tali Dekel, Andrew Zisserman, William T Freeman, and Michael Rubinstein. Omnimatte: Associating objects and their effects in video. In CVPR, 2021. 2", + "[22] Shishira R Maiya, Anubhav Gupta, Matthew Gwilliam, Max Ehrlich, and Abhinav Shrivastava. Latent-inr: A flexible framework for implicit representations of videos with discriminative semantics. In ECCV, 2024. 2, 6", + "[23] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM TOG, 41(4):1-15, 2022. 3", + "[24] Seonghyeon Nam, Marcus A Brubaker, and Michael S Brown. Neural image representations for multi-image fusion and layer separation. In ECCV, 2022. 1", + "[25] Alex Nichol, Joshua Achiam, and John Schulman. On first-order meta-learning algorithms. arXiv preprint arXiv:1803.02999, 2018. 2", + "[26] Hao Ouyang, Qiuyu Wang, Yuxi Xiao, Qingyan Bai, Juntao Zhang, Kecheng Zheng, Xiaowei Zhou, Qifeng Chen, and Yujun Shen. Codef: Content deformation fields for temporally consistent video processing. In CVPR, 2024. 2", + "[27] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In CVPR, 2016. 4, 7", + "[28] Alex Rav-Acha, Pushmeet Kohli, Carsten Rother, and Andrew Fitzgibbon. Unwrap mosaics: A new representation for video editing. In SIGGRAPH, 2008. 2", + "[29] Bipasha Sen, Gaurav Singh, Aditya Agarwal, Rohith Agaram, Madhava Krishna, and Srinath Sridhar. Hypnerf: Learning improved nerf priors using a hypernetwork. NeurIPS, 2024. 2", + "[30] Soumyadip Sengupta, Vivek Jayaram, Brian Curless, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Background matting: The world is your green screen. In CVPR, 2020. 2", + "[31] Dmitriy Smirnov, Michael Gharbi, Matthew Fisher, Vitor Guizilini, Alexei Efros, and Justin M Solomon. Marionette: Self-supervised sprite learning. NeurIPS, 2021. 1", + "[32] Shaolin Su, Qingsen Yan, Yu Zhu, Cheng Zhang, Xin Ge, Jinqiu Sun, and Yanning Zhang. Blindly assess image quality in the wild guided by a self-adaptive hyper network. In CVPR, 2020. 2", + "[33] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In ECCV, 2020. 4", + "[34] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. VideoMAE: Masked autoencoders are data-efficient learners for self-supervised video pre-training. In NeurIPS, 2022. 1, 2, 3", + "[35] Johannes von Oswald, Christian Henning, Benjamin F. Grewe, and João Sacramento. Continual learning with hypernetworks. In ICLR, 2020. 2", + "[36] John YA Wang and Edward H Adelson. Representing moving images with layers. IEEE TIP, 3(5):625-638, 1994. 2", + "[37] Ning Xu, Brian Price, Scott Cohen, and Thomas Huang. Deep image matting. In CVPR, 2017. 2" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22941", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Vickie Ye, Zhengqi Li, Richard Tucker, Angjoo Kanazawa, and Noah Snively. Deformable sprites for unsupervised video decomposition. In CVPR, 2022. 1, 2, 5", + "[39] Emilie Yu, Kevin Blackburn-Matzen, Cuong Nguyen, Oliver Wang, Rubaiat Habib Kazi, and Adrien Bousseau. Videodoodles: Hand-drawn animations on videos with scene-aware canvases. ACM TOG, 42(4):1-12, 2023. 1" + ], + "bbox": [ + 89, + 90, + 482, + 189 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22942", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/HyperNVD_ Accelerating Neural Video Decomposition via Hypernetworks/f624369a-7a34-4be6-8ddf-7450e2fca80c_model.json b/2025/HyperNVD_ Accelerating Neural Video Decomposition via Hypernetworks/f624369a-7a34-4be6-8ddf-7450e2fca80c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ce714f3c25ad3faa1462baa725278721bb30a3d4 --- /dev/null +++ b/2025/HyperNVD_ Accelerating Neural Video Decomposition via Hypernetworks/f624369a-7a34-4be6-8ddf-7450e2fca80c_model.json @@ -0,0 +1,2002 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.125, + 0.131, + 0.877, + 0.155 + ], + "angle": 0, + "content": "HyperNVD: Accelerating Neural Video Decomposition via Hypernetworks" + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.18, + 0.763, + 0.234 + ], + "angle": 0, + "content": "Maria Pilligua1,2,* Danna Xue1,2,*† Javier Vazquez-Corral1,2 \n1 Universitat Autònoma de Barcelona 2 Computer Vision Center {mpilligua, dxue, javier.vazquez}@cvc.uab.cat" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.301, + 0.486, + 0.543 + ], + "angle": 0, + "content": "Decomposing a video into a layer-based representation is crucial for easy video editing for the creative industries, as it enables independent editing of specific layers. Existing video-layer decomposition models rely on implicit neural representations (INRs) trained independently for each video, making the process time-consuming when applied to new videos. Noticing this limitation, we propose a meta-learning strategy to learn a generic video decomposition model to speed up the training on new videos. Our model is based on a hypernetwork architecture which, given a video-encoder embedding, generates the parameters for a compact INR-based neural video decomposition model. Our strategy mitigates the problem of single-video overfitting and, importantly, shortens the convergence of video decomposition on new, unseen videos. Our code is available at: https://hypernvd.github.io/." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.57, + 0.222, + 0.587 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.595, + 0.483, + 0.731 + ], + "angle": 0, + "content": "Video editing plays a crucial role in today's creative industries, from film-making and advertising to social media content creation. It enables creators to bring their visions to life, enhancing storytelling through visual effects, scene adjustments, and content personalization. However, most current video editing techniques face several challenges: they require extensive manual intervention [39], lack consistent control over edits [7, 11], and are time-consuming [4, 24], limiting their scalability for real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.732, + 0.483, + 0.868 + ], + "angle": 0, + "content": "Current video editing workflows in professional software resemble traditional 2D animation. Creators import video into editing software [1, 3] and edit in overlaid layers for each video frame. While motion tracking tools can interpolate adjustments between keyframes [10, 18], handling occlusions and ensuring consistency remain challenging since tracking errors make editing cumbersome. These problems highlight the importance of accurate layered representation and frame mappings." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.269, + 0.905, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.471, + 0.908, + 0.542 + ], + "angle": 0, + "content": "Figure 1. Fine-tuning the video decomposition model from our metamodel (HyperNVD) versus training from scratch on unseen videos shows clear advantages. With initialization from our HyperNVD trained on 15 videos, the model converges faster to the same PSNR and ultimately achieves better performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.552, + 0.907, + 0.825 + ], + "angle": 0, + "content": "Layer-based video decomposition methods [4, 20, 24, 31, 38] represent video with several texture layers, where each layer consistently represents specific video content (e.g., moving objects or backgrounds). These approaches are mainly based on compact implicit neural representations (INRs), which encode each video individually. The model projects pixel coordinates and frame indices into a canonical 2D texture space, and then, it maps to RGB values to reconstruct each frame. In this way, it is possible to apply simple edits on the canonical texture layer and then propagate them to the entire video, greatly relieving the need for manual adjustments. Furthermore, the parameter sizes and training times are reduced compared to training a generalized video object segmentation model [2] and global flow estimation model [8, 16]. However, due to the nature of INR, the neural video decomposition model lacks generalization and requires tens of minutes to train a \\(480\\mathrm{p}\\) video [4]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In this paper, we propose a meta-learning strategy in layer-based video decomposition to speed up the training. More in detail, we introduce a hypernetwork (hypernet) that, based on the video embedding predicted by VideoMAE [34], generates the parameters for a compact INR" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.365, + 0.888 + ], + "angle": 0, + "content": "*These authors contributed equally to this work." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.236, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.365, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22933" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.242 + ], + "angle": 0, + "content": "model. Our strategy facilitates the generalization to more than one video and also allows for speed-up training on new, unseen videos, as we can use the parameters given by the hypernet to initialize the neural video decomposition model for further fine-tuning. Figure 1 shows three examples of this speed-up. In this figure, we compare the training time and the PSNR obtained for training an unseen video starting from the parameters suggested by our hypernet or from a random parameter set. Our method obtains the same PSNR with more than 30 minutes less." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.25, + 0.453, + 0.264 + ], + "angle": 0, + "content": "Contributions. Our main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.268, + 0.483, + 0.342 + ], + "angle": 0, + "content": "- We incorporate a meta-learning strategy to learn a generic video decomposition method. Our strategy consists of learning a hypernet that, given a video-encoder embedding, predicts the parameters for an INR-based video decomposition model." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.343, + 0.483, + 0.387 + ], + "angle": 0, + "content": "- Our framework relieves the problem of overfitting on a single video and shortens the convergence time of the neural video decomposition model on unseen videos." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.388, + 0.482, + 0.418 + ], + "angle": 0, + "content": "- The experimental results show the generalizability and appealing video editing ability of our approach." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.268, + 0.483, + 0.418 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.431, + 0.233, + 0.446 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.456, + 0.351, + 0.472 + ], + "angle": 0, + "content": "2.1. Layered video decomposition" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.478, + 0.483, + 0.841 + ], + "angle": 0, + "content": "In image editing and animation, an image is often separated into multiple layers, each with color and opacity [30, 37], allowing independent adjustments before stacking them for the final result. This layered representation is also applied to video editing for simplifying the workflow and ensuring consistency [15, 28, 36]. Neural video decomposition approaches [4, 17, 38] decompose videos into consistent 2D texture layers across frames and are usually trained in a self-supervised manner using motion and appearance constraints. LNA [17] uses implicit neural representations to reconstruct videos via neural atlases and alpha blending. While encoding each video individually reduces parameter size and rendering time, reconstructing high-frequency details and long training time remain challenges. Omnimatte [21] and Factormatte [12] add appearance priors to model associated effects like shadows, or reflections among others. Hashing-nvd [4] and CoDeF [26] introduce hash encoding to accelerate optimization. However, these models still need more than 40 minutes of training on one single video [4]. We propose to use a hypernet to generate parameters of the video decomposition INR. In this way, we can further shorten the training time of any new, unseen video by initializing its training from the weights given by the hypernet." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.24, + 0.866 + ], + "angle": 0, + "content": "2.2. Hypernetwork" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Hypernetworks (or hypernets) [5, 13] is a technique that enhances the adaptability and flexibility of deep learn" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.379 + ], + "angle": 0, + "content": "ing models. Instead of learning task-specific models with fixed parameters, a hypernet dynamically generates parameters for a target model based on task [9, 25, 35], or data [19, 22, 29, 32]. By learning a shared, adaptable representation, hypernets help target models generalize better across different conditions. Once trained, they use embeddings as input to modulate the target model's weights and reduce optimization times on unseen tasks [9, 25]. Recent work has further improved input-specific model flexibility, especially in INR applications [6, 19, 22]. The work most closely related to ours is HyP-NeRF [29], which employs a hypernet to generate instance-specific parameters for hash encoding and NeRF, linking text to 3D object reconstructions. This framework relates semantic similarities across models for different instances but focuses on simple 3D objects with limited shape and color variations. In contrast, our video modeling task is more complex and requires attention to multiple objects, detailed motion, and varied appearances." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.392, + 0.905, + 0.427 + ], + "angle": 0, + "content": "3. Accelerate Neural Video Decomposition via Hypernetwork" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.434, + 0.714, + 0.449 + ], + "angle": 0, + "content": "3.1. Framework overview" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.456, + 0.907, + 0.683 + ], + "angle": 0, + "content": "Our goal is to learn a generalizable model for neural video decomposition that helps speed up the optimization of unseen videos. Our HyperNVD consists of three main components: i) a video embedding generation model, ii) a hypernet to generate model parameters, and iii) a neural video decomposition (NVD) model. Figure 2 shows an overview of our model. First, we compress the features predicted by a pre-trained video transformer, VideoMAE [34], and use the compressed embedding as input to the hypernet. Then the hypernet predicts the weights and the multiresolution hash encoding parameters of the target NVD model. Our architecture allows the training of multiple videos simultaneously, which enables fast adaptation of unseen videos. This fast adaptation is accomplished by considering the parameters given by the hypernet for further fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.905, + 0.765 + ], + "angle": 0, + "content": "Neural video decomposition (NVD) model. Standard layered video decomposition models based on INR [4] work by taking video coordinates and frame indices \\( p = (x, y, t) \\) as input and decomposing the video into a background layer and a foreground layer." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.905, + 0.901 + ], + "angle": 0, + "content": "As shown in the Layer Module of Figure 2, the foreground layer representation \\( LM_{f} \\) is encoded by a mapping module \\( \\mathcal{M}_f \\), a texture module \\( \\mathcal{T}_{f} \\), and a residual module \\( \\mathcal{R}_f \\). The mapping module \\( \\mathcal{M}f \\) transforms the input coordinate \\( p \\) to the texture space coordinate \\( (u,v)_f^p \\). The texture module \\( \\mathcal{T}_f \\) converts the texture coordinate into an RGB value \\( c_{f} \\), while the residual module \\( \\mathcal{R}_f \\) predicts a framewise residual coefficient \\( r_f \\). Only \\( \\mathcal{T}f \\) and \\( \\mathcal{R}f \\) incorporate Multi-resolution Hashing Encoding (MRHE) to accelerate" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22934" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.087, + 0.887, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.415, + 0.908, + 0.513 + ], + "angle": 0, + "content": "Figure 2. The architecture of our HyperNVD. HyperNVD consists of \\( i) \\) the MAE encoder to generate video embedding, \\( ii) \\) the hypernet to generate model parameters, and \\( iii) \\) the target neural video decomposition (NVD) model. Given an input video, the MAE encoder encodes it to a compact embedding \\( e \\). Then the hypernet \\( \\mathcal{H} \\) generates the parameters of the NVD model, including the MultiResolution Hash Encoding (MRHE) and the model weights. The NVD model includes two layer modules (foreground and background) for reconstructing different components in the video and an alpha module for predicting the soft mask to blend the layers. The reconstructed frame is generated by adding the foreground and background texture layers with an opacity map. Each layer module consists of a mapping module, a texture module, and a residual module." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.532, + 0.482, + 0.652 + ], + "angle": 0, + "content": "training and inference. The MRHE (illustrated in the rightmost block of Figure 2) performs grid hash encoding [23]. Here, the input 2D features serve as coordinates within multi-resolution grids, allowing for interpolation-based data sampling. The sampled data is then concatenated and processed through MLPs. The final output of the foreground layer is computed as \\(\\hat{c}_f^p = c_f^p * r_f^p\\). The background layer follows the same structure." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.653, + 0.483, + 0.729 + ], + "angle": 0, + "content": "The background and foreground representations are later combined by an alpha module \\(\\mathcal{A}\\). This module takes the coordinate \\(p\\) as input and predicts a soft opacity map for each layer. The final reconstructed RGB value, \\(\\hat{c}^p\\), is obtained by weighted summation across both layers:" + }, + { + "type": "equation", + "bbox": [ + 0.2, + 0.742, + 0.483, + 0.761 + ], + "angle": 0, + "content": "\\[\n\\hat {c} ^ {p} = \\left(1 - \\alpha^ {p}\\right) \\hat {c} _ {b} ^ {p} + \\alpha^ {p} \\hat {c} _ {f} ^ {p}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.772, + 0.462, + 0.786 + ], + "angle": 0, + "content": "All the modules in the above model structure are MLPs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.788, + 0.484, + 0.849 + ], + "angle": 0, + "content": "Note that our explanation can easily be switched to multiple foreground layers based on the masks provided by users; we explained the case of a single foreground layer for simplification." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Hypernet. Our hypernet, \\(\\mathcal{H}\\), predicts the INR parameters given a condition embedding \\(e\\) —see the Hypernet block in Figure 2. It comprises a series of MLPs, each dedicated" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.532, + 0.907, + 0.682 + ], + "angle": 0, + "content": "to generating the parameters of a specific layer within the NVD model. The target of \\(\\mathcal{H}\\) is to generate accurate weights for an NVD model capable of robust video decomposition. To accomplish this, we train it using the same loss functions and constraints applied in single-video decomposition models, as the objective for each video remains unchanged. However, by conditioning the hypernet on unique embeddings for each video, the model learns a generalized decomposition capability that extends beyond individual instances." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.697, + 0.909, + 0.865 + ], + "angle": 0, + "content": "VideoMAE encoder. Typically, hypernet inputs use either learnable embeddings trained jointly with the hypernet or embeddings generated from input data. We use the latter, as representing high-dimensional video data in a compact, low-dimensional embedding is particularly challenging. This approach also supports generalization on multiple videos and enables inference on unseen videos, which is crucial for building a robust NVD model. Moreover, this video embedding needs to encapsulate both motion and scene information, help reduce optimization complexity, and accelerate network convergence." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.908, + 0.901 + ], + "angle": 0, + "content": "To achieve this, we use the features predicted by VideoMAE [34], a frozen self-supervised masked autoencoder" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22935" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.087, + 0.485, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.182, + 0.483, + 0.252 + ], + "angle": 0, + "content": "Figure 3. Training of the autoencoder to get a compressed embedding from VideoMAE. First, we extract features \\(o\\) from each 3D patch using the pre-trained VideoMAE. We then train an autoencoder to obtain a compressed embedding \\(e\\) by minimizing the difference between the output \\(\\hat{o}\\) and the input \\(o\\)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.271, + 0.483, + 0.407 + ], + "angle": 0, + "content": "trained on video data. As shown in Figure 3, to further reduce the dimension of the VideoMAE output, we use an encoder-decoder structure to learn a compressed embedding of size (768, 1) from the original VideoMAE output of size (768, #patches). This ensures the encoder consolidates information from all patches into a single dense embedding while minimizing information loss. The autoencoder is trained using \\( L_{1} \\) loss to enforce the closeness between the reconstructed output \\( \\hat{o} \\) and the original embedding \\( o \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.407, + 0.484, + 0.513 + ], + "angle": 0, + "content": "We take the VideoMAE together with the encoder part as our VideoMAE encoder in HyperNVD (the top-left module in Figure 2). This encoder is pre-trained and frozen during the hypernet training, allowing us to maintain the rich representations it learned while preserving essential information for the decomposition tasks. In this way, the encoder serves as a compact, fixed embedding provider for the hypernet." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.523, + 0.392, + 0.54 + ], + "angle": 0, + "content": "3.2. Training and rendering procedure" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.546, + 0.483, + 0.606 + ], + "angle": 0, + "content": "Loss functions. Since all the components of our model are fully differentiable, we train them end-to-end via stochastic gradient descent. We use losses from the previous work [4, 17] to facilitate the training process. The losses include:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.607, + 0.483, + 0.636 + ], + "angle": 0, + "content": "- Reconstruction loss to ensure the reconstruction quality of the video." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.637, + 0.483, + 0.667 + ], + "angle": 0, + "content": "- Consistency loss which provides supervision of optical flow for accurate motion representation." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.667, + 0.483, + 0.696 + ], + "angle": 0, + "content": "- Sparsity loss to prevent duplicate content in different texture layers." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.697, + 0.483, + 0.741 + ], + "angle": 0, + "content": "- Residual consistency loss to ensure smooth lighting conditions while preventing the residual estimator from absorbing all color details." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.607, + 0.483, + 0.741 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.743, + 0.424, + 0.758 + ], + "angle": 0, + "content": "Two losses are applied just in the initial iterations:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.758, + 0.483, + 0.788 + ], + "angle": 0, + "content": "- Rigidity loss to maintain the structural rigidity in the canonical texture layer." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.788, + 0.483, + 0.817 + ], + "angle": 0, + "content": "- Alpha bootstrapping loss, which supervises coarse masks for foreground separation." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.758, + 0.483, + 0.817 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.818, + 0.483, + 0.88 + ], + "angle": 0, + "content": "Since the model is trained in a self-supervised manner, the additional supervision we use are the ground truth object masks and the optical flow predicted by RAFT [33]. Further details are in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Training with the hypernet. The objective is to train a" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.288 + ], + "angle": 0, + "content": "hypernet capable of generating parameters for the NVD model, so only the hypernet weights are trainable, while the NVD model serves as a set of differentiable layers that enable backpropagation to the hypernet, allowing it to learn from the above-mentioned losses. We begin the training process with a pre-training step that configures the mapping network to predict textures that are both aligned and in an initial rectangular shape corresponding to the frame size. Without this initial step, the model could produce incorrectly oriented textures. With this pre-training, the convergence during the subsequent training phase is significantly faster, as the model starts from a more informed initialization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.296, + 0.907, + 0.357 + ], + "angle": 0, + "content": "Fine-tune on new videos. Once our model is trained on a set of videos, we can use the hypernet to generate parameters for a new video and treat these weights as an initialization for the NVD model (see Figure 1 and Section 4.2)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.363, + 0.909, + 0.56 + ], + "angle": 0, + "content": "Video editing and rendering. The NVD model decomposes an input video into multiple layers, each with distinct textures. For video editing, we first perform inference to obtain these texture layers. Then, each layer can be edited individually with image editing software or other techniques. For editing, we render the texture networks using a \\(1000 \\times 1000\\) grid sampling. The rendered texture is then modified, and the colors are sampled through bilinear interpolation for smooth transitions and precise adjustments. By replacing the original layers with the modified ones and mapping the corresponding coordinates back to each video frame, we can generate the edited video with temporally consistent changes." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.575, + 0.649, + 0.592 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.6, + 0.729, + 0.615 + ], + "angle": 0, + "content": "4.1. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.622, + 0.907, + 0.697 + ], + "angle": 0, + "content": "Datasets. We conduct our experiments mainly on the DAVIS dataset [27]. We train HyperNVD with up to 30 videos, while the rest videos are reserved for testing. Each video consists of the first 16 frames at a resolution of \\(768 \\times 432\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Training settings. We use a batch size of 100,000 point locations and train each model for around 25,000 iterations and 100 pre-train iterations. The hypernet consists of approximately 290 million learnable parameters. The main network consists of approximately 4.4 million learnable parameters: 1.75 million for each layer module and 965 thousand for the alpha module. Training and evaluation with 1 video and a batch size of 100,000 requires around 22GB of GPU memory, with a checkpoint size of 2.3GB. Training our model takes roughly 2.8 hours for 1 video and 12.5 hours for 3 videos on an NVIDIA RTX 3090 GPU. We optimize all networks simultaneously using the Adam optimizer, with a learning rate of 1e-2 for the hash encoding" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22936" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.209, + 0.089, + 0.788, + 0.104 + ], + "angle": 0, + "content": "Table 1. Comparison with previous approaches on reconstruction quality for single video training." + }, + { + "type": "table", + "bbox": [ + 0.152, + 0.111, + 0.842, + 0.201 + ], + "angle": 0, + "content": "
Methodshikeblackswancar-turnbearlucia
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
Deformable Sprites [38]25.300.79525.750.73230.740.85026.340.81626.540.771
Layered Neural Atlases [17]30.020.88229.960.85334.390.92929.620.91029.920.888
Hashing-nvd [4]29.120.87330.300.84335.430.89531.560.90829.290.866
HyperNVD (Ours)30.060.84629.530.81034.940.90931.580.90130.300.865
" + }, + { + "type": "table_caption", + "bbox": [ + 0.17, + 0.213, + 0.825, + 0.227 + ], + "angle": 0, + "content": "Table 2. Comparison of the reconstruction quality of our model when trained with different numbers of videos." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.235, + 0.902, + 0.339 + ], + "angle": 0, + "content": "
Num of videoshikeblackswancar-turnbearluciakite-walkbusstrollerboatrollerblade
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
130.060.84629.530.81034.940.90931.580.90130.300.86533.350.83732.450.88931.250.83032.770.87228.750.810
327.310.72827.670.75431.870.850--------------
527.130.76027.320.73431.430.84528.880.85427.430.760----------
1027.990.75428.860.73632.470.85229.990.85328.500.76331.750.79729.450.83428.800.77531.070.82826.060.740
1527.650.76827.410.75632.570.87829.710.85228.540.83331.990.79629.600.84428.920.78030.910.82626.130.742
3027.240.79527.300.74531.640.86429.280.86227.310.80230.930.77128.630.82727.000.72029.980.80324.350.686
" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.365, + 0.361, + 0.379 + ], + "angle": 0, + "content": "and a learning rate of 5e-4 for the others." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.388, + 0.483, + 0.525 + ], + "angle": 0, + "content": "Details of architecture. In the target NVD model, we apply Rectified Linear Unit (ReLU) activations between hidden layers, with outputs passed through a tanh function. This enables the model to map texture to a predefined continuous range. The hypernet comprises a series of MLPs. Each MLP consists of four fully connected layers, with a hidden dimension of 64 and intermediate ReLU activations. Further details about the architecture can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.537, + 0.279, + 0.552 + ], + "angle": 0, + "content": "4.2. Quantitative results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.56, + 0.483, + 0.743 + ], + "angle": 0, + "content": "Comparison with previous methods. We present the comparison of PSNR and SSIM values between DeformableSprites [38], Layered Neural Atlases [17], Hashing-nvd [4] and our HyperNVD in Table 1. We report the results of our model trained on one single video independently. For a fair comparison, we use the ground truth mask as initial guidance for all the comparing methods. Results in Table 1 demonstrate that incorporating the hypernet does not degrade the model's performance. Ours maintains the effectiveness of the original architecture (i.e. our baseline Hashing-nvd) while introducing the flexibility of a hypernet-based approach." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Learning multiple videos. To evaluate the model's generalization ability across multiple videos, we train HyperNVD with 1, 3, 5, 10, 15, and 30 videos, respectively. For consistency, videos included in smaller-scale settings are subsets of those in larger-scale settings. Table 2 shows the PSNR of the videos reconstructed by models trained with different numbers of videos. When training with multiple videos, we observe a slight drop in reconstruction quality compared to training on a single video. This is likely due to the model potentially overfitting to the unique characteristics of a sin" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.365, + 0.905, + 0.425 + ], + "angle": 0, + "content": "gle video. However, the PSNR only decreases by approximately 3dB as the number of videos increases from 1 to 30, which indicates that the model retains strong performance even with larger datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.425, + 0.906, + 0.577 + ], + "angle": 0, + "content": "Figure 4 further illustrates the image quality of the \"hike\" video when training HyperNVD with different numbers of videos. The models trained on multiple videos can still reconstruct the low-frequency parts well, but some high-frequency details are lost. This stable evolution of PSNR indifferent to the number of videos indicates a plateau in performance with larger video sets. Figure 5 shows the reconstruction quality of the video \"hike\" across different iteration steps. The resulting curves are very similar for all the cases trained with multiple videos." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.906, + 0.81 + ], + "angle": 0, + "content": "Finetuning on new videos. To evaluate the ability of fast adaption on unseen videos, we train the NVD model initialized by the parameters of our metamodel trained with 15 videos and compare it with the model trained from scratch. Note that the NVD model is trained independently without the hypernet. As shown in Table 3, training from our metamodel (i.e. HyperNVD) consistently improves performance by \\(0.8\\mathrm{dB}\\) across different videos compared to training from scratch (i.e. Hashing-nvd). Figure 1 further shows our model reaches the same PSNR on the unseen videos more than 30 minutes faster, and with more stable training. This demonstrates that our metamodel enables faster adaptation to new scenes, which improves the efficiency of video editing workflow. Additional results are provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.819, + 0.692, + 0.835 + ], + "angle": 0, + "content": "4.3. Qualitative results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.905, + 0.903 + ], + "angle": 0, + "content": "Figure 6 shows examples of our layer decomposition results. For each video frame, we include a) the original frame, b) the predicted background and foreground components with a checkerboard overlay to illustrate texture" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22937" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.096, + 0.09, + 0.9, + 0.104 + ], + "angle": 0, + "content": "Table 3. Reconstruction quality in PSNR on unseen videos when fine-tuning from our 15-video metamodel versus training from scratch." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.112, + 0.905, + 0.178 + ], + "angle": 0, + "content": "
Methodscowssoccerballsoapboxfamingohorse-jumpgoatswingkite-surflibbysurf
From scratch27.0428.5131.2829.0327.6127.0629.6531.6529.6333.61
From metamodel28.4030.1632.7429.9628.5027.7830.3732.1930.1133.86
Improvements+1.36+1.65+1.46+0.93+0.89+0.72+0.72+0.54+0.48+0.25
" + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.202, + 0.484, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.32, + 0.484, + 0.39 + ], + "angle": 0, + "content": "Figure 4. Comparison of the video \"hike\" reconstructed by models trained with different numbers of videos. The top row shows the full frames, while the bottom row zooms in on a specific patch. Models trained on multiple videos capture the primary structure of the frame but tend to lose some high-frequency details." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.409, + 0.482, + 0.54 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.546, + 0.484, + 0.616 + ], + "angle": 0, + "content": "Figure 5. Comparison of the reconstruction quality for the \"hike\" video trained with different numbers of videos. Our HyperNVD model trained on multiple videos shows a 3dB drop in PSNR compared to single-video training. The PSNR remains stable across multiple video training, suggesting a performance plateau." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.483, + 0.763 + ], + "angle": 0, + "content": "mapping transformations, c) the predicted residual maps, d) the complete canonical texture map, and e) the reconstructed frame. Our model obtains accurate video reconstruction without introducing noticeable texture distortions. The predicted opacity map separates the foreground and background correctly. The residuals learn the changing reflections on the water and shadows as shown in the \"black-swan\" and \"boat\" videos." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.901 + ], + "angle": 0, + "content": "To further showcase the effectiveness of our approach for editing, we demonstrate editing examples in Figure 7. In the first two examples, only the background is modified, while in the last example, we edit the foreground texture. For the first example, the \"blackswan\" video, we transform the plants in the background into flowers and adjust the color of the water. The ripples and shadows in the water maintain a natural appearance, and the transitions at the boundaries between the swan and the water are seamless. In the sec-" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.2, + 0.906, + 0.283 + ], + "angle": 0, + "content": "Table 4. Ablations on the input video embedding and learning the MRHE with or without hypernet. 'Num' represents the number of videos used for training. Using the MAE encoder generated embedding is better than a learnable embedding jointly optimized with the hypernet. The video-specific MRHE encodings predicted by the hypernet outperform a universal MRHE encoding." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.294, + 0.905, + 0.383 + ], + "angle": 0, + "content": "
NumEmbed w MAEMRHE w hyperhikecar-turnblackswan
PSNRSSIMPSNRSSIMPSNRSSIM
1X✓30.170.86734.040.90529.950.822
✓X30.230.85833.470.88128.830.797
✓✓30.060.84634.940.90929.530.810
3X✓26.370.76029.980.82926.070.684
✓X26.880.71230.010.78826.640.707
✓✓27.310.72831.870.85027.670.754
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.414, + 0.906, + 0.49 + ], + "angle": 0, + "content": "ond video (\"kite-walk\"), we alter the beach background to a sunset style. In the \"hike\" video, we change the color of the person's shirt to yellow and add text on the backpack. Additional results of decomposition and editing are available in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.505, + 0.664, + 0.521 + ], + "angle": 0, + "content": "4.4. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.529, + 0.906, + 0.59 + ], + "angle": 0, + "content": "We do ablations on using different embedding as the input of the model, on learning the multi-resolution hash encoding with and without hypernet, and on using different types of masks as the initial supervision." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.907, + 0.901 + ], + "angle": 0, + "content": "VideoMAE embedding v.s. Learnable embedding. A well-designed embedding is crucial for a hypernet to generate effective model parameters. Some related works [4, 22] employ a learnable embedding optimized jointly with the hypernet, while we directly grab the embedding from the pre-trained MAE encoder and freeze it during hypernet training. We compare the results of using an optimized embedding versus an embedding generated by the MAE encoder, as shown in the first and third rows of Table 4. The advantage in PSNR for the VideoMAE embedding when training for three videos is more than \\(1\\mathrm{dB}\\) for the 3-video model. For single-video training, the impact of using either embedding type on the final video quality is minimal. Thus, the advantage of using the VideoMAE Encoder becomes more evident when applied to multiple videos. VideoMAE embedding also shows superiority in unseen videos when using the 3-video metamodel as initialization (see the supplementary materials). Furthermore, using VideoMAE enables robust generalization to new, unseen videos, which is essential for the versatility of our model." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22938" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.09, + 0.905, + 0.635 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.643, + 0.908, + 0.687 + ], + "angle": 0, + "content": "Figure 6. Qualitative results on the DAVIS dataset. Displayed are the a) original frames, b) predicted masks, c) residual maps, d) texture maps, e) reconstructed frames, and PSNR values for different videos. A color checkerboard overlay is applied to the masked areas to visualize texture transformations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.703, + 0.486, + 0.9 + ], + "angle": 0, + "content": "Video-based MRHE learned with hypernet v.s. Universal MRHE without hypernet. In this part, we show the benefits of learning multiresolution hash encoding with the hypernet. In multiresolution hash encoding, input 2D features are treated as coordinates within multi-resolution grids and used to sample data from these grids via interpolation. The sampled data are then concatenated and processed through MLPs to generate the final output. In Table 4, we compare the results of i) using hypernet to learn the parameters of the MLPs making these parameters unique for each video (third rows) and ii) directly using the main NVD net to learn MRHE parameters making them common for all videos (second rows). The higher quality results are ob" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.703, + 0.909, + 0.797 + ], + "angle": 0, + "content": "tained when MRHE is learned by the hypernet, with benefits that become more apparent as the number of videos used for training increases. This improvement stems from the fact that different video scenes exhibit varying levels of complexity, making video-specific hash grid parameters beneficial when training across multiple videos." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Accurate mask v.s. Coarse mask. The reference mask is important in our task since it provides a clue about the moving salient objects. We compare the use of manually annotated precise masks (groundtruth from the DAVIS dataset [27]) with coarsely predicted masks from MaskR-CNN [14] as the initial reference for model training. The" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22939" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.094, + 0.093, + 0.247, + 0.105 + ], + "angle": 0, + "content": "a) Reconstructed frame" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.108, + 0.246, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.304, + 0.093, + 0.445, + 0.106 + ], + "angle": 0, + "content": "b) Output 2D textures" + }, + { + "type": "image", + "bbox": [ + 0.254, + 0.108, + 0.495, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.565, + 0.093, + 0.697, + 0.106 + ], + "angle": 0, + "content": "c) Edited 2D textures" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.108, + 0.746, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.78, + 0.093, + 0.879, + 0.106 + ], + "angle": 0, + "content": "d) Edited frame" + }, + { + "type": "image", + "bbox": [ + 0.754, + 0.108, + 0.905, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.183, + 0.246, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.252, + 0.183, + 0.495, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.183, + 0.746, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.753, + 0.183, + 0.905, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.261, + 0.246, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.253, + 0.261, + 0.495, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.261, + 0.746, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.753, + 0.261, + 0.905, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.34, + 0.908, + 0.381 + ], + "angle": 0, + "content": "Figure 7. Edited video results. We make adjustments to one of the texture layers—the background layers (rows 1 and 2), and the foreground layer (row 3)—while keeping the other layer unchanged and blending the edited layers together. The edited video maintains good visual quality and smooth transitions." + }, + { + "type": "table_caption", + "bbox": [ + 0.102, + 0.4, + 0.471, + 0.414 + ], + "angle": 0, + "content": "Table 5. Ablation on using different masks as initial guidance." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.421, + 0.472, + 0.484 + ], + "angle": 0, + "content": "
Maskhikecar-turnblackswan
PSNRSSIMPSNRSSIMPSNRSSIM
GT30.060.84634.940.90929.530.810
MaskRCNN29.610.84634.920.90329.660.814
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.511, + 0.484, + 0.692 + ], + "angle": 0, + "content": "predicted alpha map is a soft mask that always includes the corresponding effects, such as reflection or shadows accompanied by the moving objects. Therefore, it is always different from the original object mask. Results in Table 5 show that our model is robust to masks with varying levels of accuracy. Figure 8 illustrates two frames from the \"bear\" and \"blackswan\" videos, where a more precise initial mask improves object boundary details. For example, using a more refined mask as initial guidance help the alpha map to capture intricate structures, such as the feathers on a swan's tail. Thus, for higher-quality editing results, it is essential to use the most detailed masks possible during model training." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.708, + 0.294, + 0.725 + ], + "angle": 0, + "content": "5. Concluding Remarks" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We present a framework for multiple video training of INR-based video decomposition models using a hypernetwork. We reformulate the task of training a neural video decomposition model on a single video into learning a hypernetwork that generates network parameters based on video embeddings. This approach effectively mitigates the overfitting of the neural video decomposition model to individual videos, enhances generalization capability, and speeds up adaptation to unseen videos. Through experiments, we validate that our proposed model achieves comparable video layer decomposition and rendering capabilities to previous" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.401, + 0.903, + 0.637 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.648, + 0.907, + 0.718 + ], + "angle": 0, + "content": "Figure 8. Comparison between using accurate and coarse masks as initial guidance. Using different masks does not significantly affect the reconstruction results, but providing a refined mask as the initial guidance helps the model achieve more detailed decomposition, which can be especially beneficial for finer edits." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.738, + 0.905, + 0.768 + ], + "angle": 0, + "content": "methods, while significantly reducing training time on new videos and improving video editing efficiency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.779, + 0.907, + 0.9 + ], + "angle": 0, + "content": "Acknowledgement. We acknowledge the Departament de Recerca i Universitats with ref. 2021SGR01499 and the CERCA Program from Generalitat de Catalunya, Grant PID2021-128178OB-I00 funded by MCIN/AEI/10.13039/501100011033, ERDF \"A way of making Europe\", and the grant Catedra ENIA UAB-Cruilla (TSI-100929-2023-2) from the Ministry of Economic Affairs and Digital Transformation of Spain." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22940" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.408, + 0.128 + ], + "angle": 0, + "content": "[1] Adobe. After effects. Accessed: 2024-11-07. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.131, + 0.483, + 0.171 + ], + "angle": 0, + "content": "[2] Ali Athar, Alexander Hermans, Jonathon Luiten, Deva Ramanan, and Bastian Leibe. Tarvis: A unified approach for target-based video segmentation. In CVPR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.174, + 0.482, + 0.199 + ], + "angle": 0, + "content": "[3] BlackMagicDesign. Davinci resolve 19. Accessed: 2024-11-07. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.203, + 0.482, + 0.257 + ], + "angle": 0, + "content": "[4] Cheng-Hung Chan, Cheng-Yang Yuan, Cheng Sun, and Hwann-Tzong Chen. Hashing neural video decomposition with multiplicative residuals in space-time. In ICCV, 2023. 1, 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.26, + 0.482, + 0.313 + ], + "angle": 0, + "content": "[5] Vinod Kumar Chauhan, Jiandong Zhou, Ping Lu, Soheila Molaei, and David A Clifton. A brief review of hypernetworks in deep learning. Artificial Intelligence Review, 57(9): 250, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.316, + 0.482, + 0.357 + ], + "angle": 0, + "content": "[6] Pei-Ze Chiang, Meng-Shiun Tsai, Hung-Yu Tseng, Wei-Sheng Lai, and Wei-Chen Chiu. Stylizing 3d scene via implicit representation and hypernetwork. In WACV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.36, + 0.482, + 0.426 + ], + "angle": 0, + "content": "[7] Yuren Cong, Mengmeng Xu, Christian Simon, Shoufa Chen, Jiawei Ren, Yanping Xie, Juan-Manuel Perez-Rua, Bodo Rosenhahn, Tao Xiang, and Sen He. Flatten: optical flow-guided attention for consistent text-to-video editing. In ICLR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.43, + 0.482, + 0.484 + ], + "angle": 0, + "content": "[8] Carl Doersch, Ankush Gupta, Larisa Markeeva, Adria Recasens, Lucas Smaira, Yusuf Aytar, Joao Carreira, Andrew Zisserman, and Yi Yang. Tap-vid: A benchmark for tracking any point in a video. NeurIPS, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.487, + 0.482, + 0.526 + ], + "angle": 0, + "content": "[9] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In ICML, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.529, + 0.374, + 0.542 + ], + "angle": 0, + "content": "[10] Foundry. Nuke. Accessed: 2024-11-07. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.544, + 0.482, + 0.584 + ], + "angle": 0, + "content": "[11] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. In ICLR, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.587, + 0.482, + 0.627 + ], + "angle": 0, + "content": "[12] Zeqi Gu, Wenqi Xian, Noah Snavely, and Abe Davis. Factormatte: Redefining video matting for re-composition tasks. ACM TOG, 42(4):1-14, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.63, + 0.482, + 0.655 + ], + "angle": 0, + "content": "[13] David Ha, Andrew Dai, and Quoc V Le. Hypernetworks. In ICLR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.659, + 0.482, + 0.684 + ], + "angle": 0, + "content": "[14] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.688, + 0.482, + 0.714 + ], + "angle": 0, + "content": "[15] Nebojsa Jojic and Brendan J Frey. Learning flexible sprites in video layers. In CVPR, 2001. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.717, + 0.482, + 0.757 + ], + "angle": 0, + "content": "[16] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together. In ECCV, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.76, + 0.482, + 0.799 + ], + "angle": 0, + "content": "[17] Yoni Kasten, Dolev Ofri, Oliver Wang, and Tali Dekel. Layered neural atlases for consistent video editing. ACM TOG, 40(6):1-12, 2021. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.482, + 0.828 + ], + "angle": 0, + "content": "[18] KeenTools. Geotracker for adobe after effects. Accessed: 2024-11-07. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.831, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[19] Chiheon Kim, Doyup Lee, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Generalizable implicit neural representations via instance pattern composers. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[20] Jaekyeong Lee, Geonung Kim, and Sunghyun Cho. Rna: Video editing with roi-based neural atlas. ACCV, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[21] Erika Lu, Forrester Cole, Tali Dekel, Andrew Zisserman, William T Freeman, and Michael Rubinstein. Omnimatte: Associating objects and their effects in video. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[22] Shishira R Maiya, Anubhav Gupta, Matthew Gwilliam, Max Ehrlich, and Abhinav Shrivastava. Latent-inr: A flexible framework for implicit representations of videos with discriminative semantics. In ECCV, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.208, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[23] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM TOG, 41(4):1-15, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.252, + 0.905, + 0.292 + ], + "angle": 0, + "content": "[24] Seonghyeon Nam, Marcus A Brubaker, and Michael S Brown. Neural image representations for multi-image fusion and layer separation. In ECCV, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.296, + 0.905, + 0.335 + ], + "angle": 0, + "content": "[25] Alex Nichol, Joshua Achiam, and John Schulman. On first-order meta-learning algorithms. arXiv preprint arXiv:1803.02999, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.339, + 0.905, + 0.393 + ], + "angle": 0, + "content": "[26] Hao Ouyang, Qiuyu Wang, Yuxi Xiao, Qingyan Bai, Juntao Zhang, Kecheng Zheng, Xiaowei Zhou, Qifeng Chen, and Yujun Shen. Codef: Content deformation fields for temporally consistent video processing. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.397, + 0.905, + 0.451 + ], + "angle": 0, + "content": "[27] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In CVPR, 2016. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.454, + 0.905, + 0.494 + ], + "angle": 0, + "content": "[28] Alex Rav-Acha, Pushmeet Kohli, Carsten Rother, and Andrew Fitzgibbon. Unwrap mosaics: A new representation for video editing. In SIGGRAPH, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.498, + 0.905, + 0.551 + ], + "angle": 0, + "content": "[29] Bipasha Sen, Gaurav Singh, Aditya Agarwal, Rohith Agaram, Madhava Krishna, and Srinath Sridhar. Hypnerf: Learning improved nerf priors using a hypernetwork. NeurIPS, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.555, + 0.905, + 0.608 + ], + "angle": 0, + "content": "[30] Soumyadip Sengupta, Vivek Jayaram, Brian Curless, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Background matting: The world is your green screen. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.612, + 0.905, + 0.653 + ], + "angle": 0, + "content": "[31] Dmitriy Smirnov, Michael Gharbi, Matthew Fisher, Vitor Guizilini, Alexei Efros, and Justin M Solomon. Marionette: Self-supervised sprite learning. NeurIPS, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.656, + 0.905, + 0.709 + ], + "angle": 0, + "content": "[32] Shaolin Su, Qingsen Yan, Yu Zhu, Cheng Zhang, Xin Ge, Jinqiu Sun, and Yanning Zhang. Blindly assess image quality in the wild guided by a self-adaptive hyper network. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.713, + 0.905, + 0.74 + ], + "angle": 0, + "content": "[33] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In ECCV, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.743, + 0.905, + 0.796 + ], + "angle": 0, + "content": "[34] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. VideoMAE: Masked autoencoders are data-efficient learners for self-supervised video pre-training. In NeurIPS, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.8, + 0.905, + 0.841 + ], + "angle": 0, + "content": "[35] Johannes von Oswald, Christian Henning, Benjamin F. Grewe, and João Sacramento. Continual learning with hypernetworks. In ICLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.844, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[36] John YA Wang and Edward H Adelson. Representing moving images with layers. IEEE TIP, 3(5):625-638, 1994. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[37] Ning Xu, Brian Price, Scott Cohen, and Thomas Huang. Deep image matting. In CVPR, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22941" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[38] Vickie Ye, Zhengqi Li, Richard Tucker, Angjoo Kanazawa, and Noah Snively. Deformable sprites for unsupervised video decomposition. In CVPR, 2022. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[39] Emilie Yu, Kevin Blackburn-Matzen, Cuong Nguyen, Oliver Wang, Rubaiat Habib Kazi, and Adrien Bousseau. Videodoodles: Hand-drawn animations on videos with scene-aware canvases. ACM TOG, 42(4):1-12, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.092, + 0.483, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22942" + } + ] +] \ No newline at end of file diff --git a/2025/HyperNVD_ Accelerating Neural Video Decomposition via Hypernetworks/f624369a-7a34-4be6-8ddf-7450e2fca80c_origin.pdf b/2025/HyperNVD_ Accelerating Neural Video Decomposition via Hypernetworks/f624369a-7a34-4be6-8ddf-7450e2fca80c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..88b62236a890f4584b51e6bac54b8ecbe004712d --- /dev/null +++ b/2025/HyperNVD_ Accelerating Neural Video Decomposition via Hypernetworks/f624369a-7a34-4be6-8ddf-7450e2fca80c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0c56c020a2fd5ca9485ee99883ff27b35fa4f05ac769aa8b6d74223b8fbc2c8d +size 4554150 diff --git a/2025/HyperNVD_ Accelerating Neural Video Decomposition via Hypernetworks/full.md b/2025/HyperNVD_ Accelerating Neural Video Decomposition via Hypernetworks/full.md new file mode 100644 index 0000000000000000000000000000000000000000..791003be329d529cff92c645a3b358616b475527 --- /dev/null +++ b/2025/HyperNVD_ Accelerating Neural Video Decomposition via Hypernetworks/full.md @@ -0,0 +1,266 @@ +# HyperNVD: Accelerating Neural Video Decomposition via Hypernetworks + +Maria Pilligua1,2,* Danna Xue1,2,*† Javier Vazquez-Corral1,2 +1 Universitat Autònoma de Barcelona 2 Computer Vision Center {mpilligua, dxue, javier.vazquez}@cvc.uab.cat + +# Abstract + +Decomposing a video into a layer-based representation is crucial for easy video editing for the creative industries, as it enables independent editing of specific layers. Existing video-layer decomposition models rely on implicit neural representations (INRs) trained independently for each video, making the process time-consuming when applied to new videos. Noticing this limitation, we propose a meta-learning strategy to learn a generic video decomposition model to speed up the training on new videos. Our model is based on a hypernetwork architecture which, given a video-encoder embedding, generates the parameters for a compact INR-based neural video decomposition model. Our strategy mitigates the problem of single-video overfitting and, importantly, shortens the convergence of video decomposition on new, unseen videos. Our code is available at: https://hypernvd.github.io/. + +# 1. Introduction + +Video editing plays a crucial role in today's creative industries, from film-making and advertising to social media content creation. It enables creators to bring their visions to life, enhancing storytelling through visual effects, scene adjustments, and content personalization. However, most current video editing techniques face several challenges: they require extensive manual intervention [39], lack consistent control over edits [7, 11], and are time-consuming [4, 24], limiting their scalability for real-world applications. + +Current video editing workflows in professional software resemble traditional 2D animation. Creators import video into editing software [1, 3] and edit in overlaid layers for each video frame. While motion tracking tools can interpolate adjustments between keyframes [10, 18], handling occlusions and ensuring consistency remain challenging since tracking errors make editing cumbersome. These problems highlight the importance of accurate layered representation and frame mappings. + +![](images/35f53945dc588cf7f23ca3d6e777a5bfb6f42e01a6a1044c3f156ed42fbcc19c.jpg) +Figure 1. Fine-tuning the video decomposition model from our metamodel (HyperNVD) versus training from scratch on unseen videos shows clear advantages. With initialization from our HyperNVD trained on 15 videos, the model converges faster to the same PSNR and ultimately achieves better performance. + +Layer-based video decomposition methods [4, 20, 24, 31, 38] represent video with several texture layers, where each layer consistently represents specific video content (e.g., moving objects or backgrounds). These approaches are mainly based on compact implicit neural representations (INRs), which encode each video individually. The model projects pixel coordinates and frame indices into a canonical 2D texture space, and then, it maps to RGB values to reconstruct each frame. In this way, it is possible to apply simple edits on the canonical texture layer and then propagate them to the entire video, greatly relieving the need for manual adjustments. Furthermore, the parameter sizes and training times are reduced compared to training a generalized video object segmentation model [2] and global flow estimation model [8, 16]. However, due to the nature of INR, the neural video decomposition model lacks generalization and requires tens of minutes to train a $480\mathrm{p}$ video [4]. + +In this paper, we propose a meta-learning strategy in layer-based video decomposition to speed up the training. More in detail, we introduce a hypernetwork (hypernet) that, based on the video embedding predicted by VideoMAE [34], generates the parameters for a compact INR + +model. Our strategy facilitates the generalization to more than one video and also allows for speed-up training on new, unseen videos, as we can use the parameters given by the hypernet to initialize the neural video decomposition model for further fine-tuning. Figure 1 shows three examples of this speed-up. In this figure, we compare the training time and the PSNR obtained for training an unseen video starting from the parameters suggested by our hypernet or from a random parameter set. Our method obtains the same PSNR with more than 30 minutes less. + +Contributions. Our main contributions are as follows: + +- We incorporate a meta-learning strategy to learn a generic video decomposition method. Our strategy consists of learning a hypernet that, given a video-encoder embedding, predicts the parameters for an INR-based video decomposition model. +- Our framework relieves the problem of overfitting on a single video and shortens the convergence time of the neural video decomposition model on unseen videos. +- The experimental results show the generalizability and appealing video editing ability of our approach. + +# 2. Related Work + +# 2.1. Layered video decomposition + +In image editing and animation, an image is often separated into multiple layers, each with color and opacity [30, 37], allowing independent adjustments before stacking them for the final result. This layered representation is also applied to video editing for simplifying the workflow and ensuring consistency [15, 28, 36]. Neural video decomposition approaches [4, 17, 38] decompose videos into consistent 2D texture layers across frames and are usually trained in a self-supervised manner using motion and appearance constraints. LNA [17] uses implicit neural representations to reconstruct videos via neural atlases and alpha blending. While encoding each video individually reduces parameter size and rendering time, reconstructing high-frequency details and long training time remain challenges. Omnimatte [21] and Factormatte [12] add appearance priors to model associated effects like shadows, or reflections among others. Hashing-nvd [4] and CoDeF [26] introduce hash encoding to accelerate optimization. However, these models still need more than 40 minutes of training on one single video [4]. We propose to use a hypernet to generate parameters of the video decomposition INR. In this way, we can further shorten the training time of any new, unseen video by initializing its training from the weights given by the hypernet. + +# 2.2. Hypernetwork + +Hypernetworks (or hypernets) [5, 13] is a technique that enhances the adaptability and flexibility of deep learn + +ing models. Instead of learning task-specific models with fixed parameters, a hypernet dynamically generates parameters for a target model based on task [9, 25, 35], or data [19, 22, 29, 32]. By learning a shared, adaptable representation, hypernets help target models generalize better across different conditions. Once trained, they use embeddings as input to modulate the target model's weights and reduce optimization times on unseen tasks [9, 25]. Recent work has further improved input-specific model flexibility, especially in INR applications [6, 19, 22]. The work most closely related to ours is HyP-NeRF [29], which employs a hypernet to generate instance-specific parameters for hash encoding and NeRF, linking text to 3D object reconstructions. This framework relates semantic similarities across models for different instances but focuses on simple 3D objects with limited shape and color variations. In contrast, our video modeling task is more complex and requires attention to multiple objects, detailed motion, and varied appearances. + +# 3. Accelerate Neural Video Decomposition via Hypernetwork + +# 3.1. Framework overview + +Our goal is to learn a generalizable model for neural video decomposition that helps speed up the optimization of unseen videos. Our HyperNVD consists of three main components: i) a video embedding generation model, ii) a hypernet to generate model parameters, and iii) a neural video decomposition (NVD) model. Figure 2 shows an overview of our model. First, we compress the features predicted by a pre-trained video transformer, VideoMAE [34], and use the compressed embedding as input to the hypernet. Then the hypernet predicts the weights and the multiresolution hash encoding parameters of the target NVD model. Our architecture allows the training of multiple videos simultaneously, which enables fast adaptation of unseen videos. This fast adaptation is accomplished by considering the parameters given by the hypernet for further fine-tuning. + +Neural video decomposition (NVD) model. Standard layered video decomposition models based on INR [4] work by taking video coordinates and frame indices $p = (x, y, t)$ as input and decomposing the video into a background layer and a foreground layer. + +As shown in the Layer Module of Figure 2, the foreground layer representation $LM_{f}$ is encoded by a mapping module $\mathcal{M}_f$ , a texture module $\mathcal{T}_{f}$ , and a residual module $\mathcal{R}_f$ . The mapping module $\mathcal{M}f$ transforms the input coordinate $p$ to the texture space coordinate $(u,v)_f^p$ . The texture module $\mathcal{T}_f$ converts the texture coordinate into an RGB value $c_{f}$ , while the residual module $\mathcal{R}_f$ predicts a framewise residual coefficient $r_f$ . Only $\mathcal{T}f$ and $\mathcal{R}f$ incorporate Multi-resolution Hashing Encoding (MRHE) to accelerate + +![](images/30f300b8457157da4a08ad10ad736aa3390c9c5674a9fa58240cdb61da9b12f3.jpg) +Figure 2. The architecture of our HyperNVD. HyperNVD consists of $i)$ the MAE encoder to generate video embedding, $ii)$ the hypernet to generate model parameters, and $iii)$ the target neural video decomposition (NVD) model. Given an input video, the MAE encoder encodes it to a compact embedding $e$ . Then the hypernet $\mathcal{H}$ generates the parameters of the NVD model, including the MultiResolution Hash Encoding (MRHE) and the model weights. The NVD model includes two layer modules (foreground and background) for reconstructing different components in the video and an alpha module for predicting the soft mask to blend the layers. The reconstructed frame is generated by adding the foreground and background texture layers with an opacity map. Each layer module consists of a mapping module, a texture module, and a residual module. + +training and inference. The MRHE (illustrated in the rightmost block of Figure 2) performs grid hash encoding [23]. Here, the input 2D features serve as coordinates within multi-resolution grids, allowing for interpolation-based data sampling. The sampled data is then concatenated and processed through MLPs. The final output of the foreground layer is computed as $\hat{c}_f^p = c_f^p * r_f^p$ . The background layer follows the same structure. + +The background and foreground representations are later combined by an alpha module $\mathcal{A}$ . This module takes the coordinate $p$ as input and predicts a soft opacity map for each layer. The final reconstructed RGB value, $\hat{c}^p$ , is obtained by weighted summation across both layers: + +$$ +\hat {c} ^ {p} = \left(1 - \alpha^ {p}\right) \hat {c} _ {b} ^ {p} + \alpha^ {p} \hat {c} _ {f} ^ {p}. \tag {1} +$$ + +All the modules in the above model structure are MLPs. + +Note that our explanation can easily be switched to multiple foreground layers based on the masks provided by users; we explained the case of a single foreground layer for simplification. + +Hypernet. Our hypernet, $\mathcal{H}$ , predicts the INR parameters given a condition embedding $e$ —see the Hypernet block in Figure 2. It comprises a series of MLPs, each dedicated + +to generating the parameters of a specific layer within the NVD model. The target of $\mathcal{H}$ is to generate accurate weights for an NVD model capable of robust video decomposition. To accomplish this, we train it using the same loss functions and constraints applied in single-video decomposition models, as the objective for each video remains unchanged. However, by conditioning the hypernet on unique embeddings for each video, the model learns a generalized decomposition capability that extends beyond individual instances. + +VideoMAE encoder. Typically, hypernet inputs use either learnable embeddings trained jointly with the hypernet or embeddings generated from input data. We use the latter, as representing high-dimensional video data in a compact, low-dimensional embedding is particularly challenging. This approach also supports generalization on multiple videos and enables inference on unseen videos, which is crucial for building a robust NVD model. Moreover, this video embedding needs to encapsulate both motion and scene information, help reduce optimization complexity, and accelerate network convergence. + +To achieve this, we use the features predicted by VideoMAE [34], a frozen self-supervised masked autoencoder + +![](images/0623f59a7f60469f76d075b1a83f4243b2ba9715297e28cb1ae3ed45aa6b036d.jpg) +Figure 3. Training of the autoencoder to get a compressed embedding from VideoMAE. First, we extract features $o$ from each 3D patch using the pre-trained VideoMAE. We then train an autoencoder to obtain a compressed embedding $e$ by minimizing the difference between the output $\hat{o}$ and the input $o$ . + +trained on video data. As shown in Figure 3, to further reduce the dimension of the VideoMAE output, we use an encoder-decoder structure to learn a compressed embedding of size (768, 1) from the original VideoMAE output of size (768, #patches). This ensures the encoder consolidates information from all patches into a single dense embedding while minimizing information loss. The autoencoder is trained using $L_{1}$ loss to enforce the closeness between the reconstructed output $\hat{o}$ and the original embedding $o$ . + +We take the VideoMAE together with the encoder part as our VideoMAE encoder in HyperNVD (the top-left module in Figure 2). This encoder is pre-trained and frozen during the hypernet training, allowing us to maintain the rich representations it learned while preserving essential information for the decomposition tasks. In this way, the encoder serves as a compact, fixed embedding provider for the hypernet. + +# 3.2. Training and rendering procedure + +Loss functions. Since all the components of our model are fully differentiable, we train them end-to-end via stochastic gradient descent. We use losses from the previous work [4, 17] to facilitate the training process. The losses include: + +- Reconstruction loss to ensure the reconstruction quality of the video. +- Consistency loss which provides supervision of optical flow for accurate motion representation. +- Sparsity loss to prevent duplicate content in different texture layers. +- Residual consistency loss to ensure smooth lighting conditions while preventing the residual estimator from absorbing all color details. + +Two losses are applied just in the initial iterations: + +- Rigidity loss to maintain the structural rigidity in the canonical texture layer. +- Alpha bootstrapping loss, which supervises coarse masks for foreground separation. + +Since the model is trained in a self-supervised manner, the additional supervision we use are the ground truth object masks and the optical flow predicted by RAFT [33]. Further details are in the supplementary materials. + +Training with the hypernet. The objective is to train a + +hypernet capable of generating parameters for the NVD model, so only the hypernet weights are trainable, while the NVD model serves as a set of differentiable layers that enable backpropagation to the hypernet, allowing it to learn from the above-mentioned losses. We begin the training process with a pre-training step that configures the mapping network to predict textures that are both aligned and in an initial rectangular shape corresponding to the frame size. Without this initial step, the model could produce incorrectly oriented textures. With this pre-training, the convergence during the subsequent training phase is significantly faster, as the model starts from a more informed initialization. + +Fine-tune on new videos. Once our model is trained on a set of videos, we can use the hypernet to generate parameters for a new video and treat these weights as an initialization for the NVD model (see Figure 1 and Section 4.2). + +Video editing and rendering. The NVD model decomposes an input video into multiple layers, each with distinct textures. For video editing, we first perform inference to obtain these texture layers. Then, each layer can be edited individually with image editing software or other techniques. For editing, we render the texture networks using a $1000 \times 1000$ grid sampling. The rendered texture is then modified, and the colors are sampled through bilinear interpolation for smooth transitions and precise adjustments. By replacing the original layers with the modified ones and mapping the corresponding coordinates back to each video frame, we can generate the edited video with temporally consistent changes. + +# 4. Experiments + +# 4.1. Implementation details + +Datasets. We conduct our experiments mainly on the DAVIS dataset [27]. We train HyperNVD with up to 30 videos, while the rest videos are reserved for testing. Each video consists of the first 16 frames at a resolution of $768 \times 432$ . + +Training settings. We use a batch size of 100,000 point locations and train each model for around 25,000 iterations and 100 pre-train iterations. The hypernet consists of approximately 290 million learnable parameters. The main network consists of approximately 4.4 million learnable parameters: 1.75 million for each layer module and 965 thousand for the alpha module. Training and evaluation with 1 video and a batch size of 100,000 requires around 22GB of GPU memory, with a checkpoint size of 2.3GB. Training our model takes roughly 2.8 hours for 1 video and 12.5 hours for 3 videos on an NVIDIA RTX 3090 GPU. We optimize all networks simultaneously using the Adam optimizer, with a learning rate of 1e-2 for the hash encoding + +Table 1. Comparison with previous approaches on reconstruction quality for single video training. + +
Methodshikeblackswancar-turnbearlucia
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
Deformable Sprites [38]25.300.79525.750.73230.740.85026.340.81626.540.771
Layered Neural Atlases [17]30.020.88229.960.85334.390.92929.620.91029.920.888
Hashing-nvd [4]29.120.87330.300.84335.430.89531.560.90829.290.866
HyperNVD (Ours)30.060.84629.530.81034.940.90931.580.90130.300.865
+ +Table 2. Comparison of the reconstruction quality of our model when trained with different numbers of videos. + +
Num of videoshikeblackswancar-turnbearluciakite-walkbusstrollerboatrollerblade
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
130.060.84629.530.81034.940.90931.580.90130.300.86533.350.83732.450.88931.250.83032.770.87228.750.810
327.310.72827.670.75431.870.850--------------
527.130.76027.320.73431.430.84528.880.85427.430.760----------
1027.990.75428.860.73632.470.85229.990.85328.500.76331.750.79729.450.83428.800.77531.070.82826.060.740
1527.650.76827.410.75632.570.87829.710.85228.540.83331.990.79629.600.84428.920.78030.910.82626.130.742
3027.240.79527.300.74531.640.86429.280.86227.310.80230.930.77128.630.82727.000.72029.980.80324.350.686
+ +and a learning rate of 5e-4 for the others. + +Details of architecture. In the target NVD model, we apply Rectified Linear Unit (ReLU) activations between hidden layers, with outputs passed through a tanh function. This enables the model to map texture to a predefined continuous range. The hypernet comprises a series of MLPs. Each MLP consists of four fully connected layers, with a hidden dimension of 64 and intermediate ReLU activations. Further details about the architecture can be found in the supplementary material. + +# 4.2. Quantitative results + +Comparison with previous methods. We present the comparison of PSNR and SSIM values between DeformableSprites [38], Layered Neural Atlases [17], Hashing-nvd [4] and our HyperNVD in Table 1. We report the results of our model trained on one single video independently. For a fair comparison, we use the ground truth mask as initial guidance for all the comparing methods. Results in Table 1 demonstrate that incorporating the hypernet does not degrade the model's performance. Ours maintains the effectiveness of the original architecture (i.e. our baseline Hashing-nvd) while introducing the flexibility of a hypernet-based approach. + +Learning multiple videos. To evaluate the model's generalization ability across multiple videos, we train HyperNVD with 1, 3, 5, 10, 15, and 30 videos, respectively. For consistency, videos included in smaller-scale settings are subsets of those in larger-scale settings. Table 2 shows the PSNR of the videos reconstructed by models trained with different numbers of videos. When training with multiple videos, we observe a slight drop in reconstruction quality compared to training on a single video. This is likely due to the model potentially overfitting to the unique characteristics of a sin + +gle video. However, the PSNR only decreases by approximately 3dB as the number of videos increases from 1 to 30, which indicates that the model retains strong performance even with larger datasets. + +Figure 4 further illustrates the image quality of the "hike" video when training HyperNVD with different numbers of videos. The models trained on multiple videos can still reconstruct the low-frequency parts well, but some high-frequency details are lost. This stable evolution of PSNR indifferent to the number of videos indicates a plateau in performance with larger video sets. Figure 5 shows the reconstruction quality of the video "hike" across different iteration steps. The resulting curves are very similar for all the cases trained with multiple videos. + +Finetuning on new videos. To evaluate the ability of fast adaption on unseen videos, we train the NVD model initialized by the parameters of our metamodel trained with 15 videos and compare it with the model trained from scratch. Note that the NVD model is trained independently without the hypernet. As shown in Table 3, training from our metamodel (i.e. HyperNVD) consistently improves performance by $0.8\mathrm{dB}$ across different videos compared to training from scratch (i.e. Hashing-nvd). Figure 1 further shows our model reaches the same PSNR on the unseen videos more than 30 minutes faster, and with more stable training. This demonstrates that our metamodel enables faster adaptation to new scenes, which improves the efficiency of video editing workflow. Additional results are provided in the supplementary material. + +# 4.3. Qualitative results + +Figure 6 shows examples of our layer decomposition results. For each video frame, we include a) the original frame, b) the predicted background and foreground components with a checkerboard overlay to illustrate texture + +Table 3. Reconstruction quality in PSNR on unseen videos when fine-tuning from our 15-video metamodel versus training from scratch. + +
Methodscowssoccerballsoapboxfamingohorse-jumpgoatswingkite-surflibbysurf
From scratch27.0428.5131.2829.0327.6127.0629.6531.6529.6333.61
From metamodel28.4030.1632.7429.9628.5027.7830.3732.1930.1133.86
Improvements+1.36+1.65+1.46+0.93+0.89+0.72+0.72+0.54+0.48+0.25
+ +![](images/06978cdae851dd78eed78193589a8685d03eff59fa2016a09f85d19bfc945cd2.jpg) +Figure 4. Comparison of the video "hike" reconstructed by models trained with different numbers of videos. The top row shows the full frames, while the bottom row zooms in on a specific patch. Models trained on multiple videos capture the primary structure of the frame but tend to lose some high-frequency details. + +![](images/e9c832804d1ea254f6e3106c82e5ea2de5365b79e71f8b89d89fd56914c50725.jpg) +Figure 5. Comparison of the reconstruction quality for the "hike" video trained with different numbers of videos. Our HyperNVD model trained on multiple videos shows a 3dB drop in PSNR compared to single-video training. The PSNR remains stable across multiple video training, suggesting a performance plateau. + +mapping transformations, c) the predicted residual maps, d) the complete canonical texture map, and e) the reconstructed frame. Our model obtains accurate video reconstruction without introducing noticeable texture distortions. The predicted opacity map separates the foreground and background correctly. The residuals learn the changing reflections on the water and shadows as shown in the "black-swan" and "boat" videos. + +To further showcase the effectiveness of our approach for editing, we demonstrate editing examples in Figure 7. In the first two examples, only the background is modified, while in the last example, we edit the foreground texture. For the first example, the "blackswan" video, we transform the plants in the background into flowers and adjust the color of the water. The ripples and shadows in the water maintain a natural appearance, and the transitions at the boundaries between the swan and the water are seamless. In the sec- + +Table 4. Ablations on the input video embedding and learning the MRHE with or without hypernet. 'Num' represents the number of videos used for training. Using the MAE encoder generated embedding is better than a learnable embedding jointly optimized with the hypernet. The video-specific MRHE encodings predicted by the hypernet outperform a universal MRHE encoding. + +
NumEmbed w MAEMRHE w hyperhikecar-turnblackswan
PSNRSSIMPSNRSSIMPSNRSSIM
1X✓30.170.86734.040.90529.950.822
✓X30.230.85833.470.88128.830.797
✓✓30.060.84634.940.90929.530.810
3X✓26.370.76029.980.82926.070.684
✓X26.880.71230.010.78826.640.707
✓✓27.310.72831.870.85027.670.754
+ +ond video ("kite-walk"), we alter the beach background to a sunset style. In the "hike" video, we change the color of the person's shirt to yellow and add text on the backpack. Additional results of decomposition and editing are available in the supplementary material. + +# 4.4. Ablation study + +We do ablations on using different embedding as the input of the model, on learning the multi-resolution hash encoding with and without hypernet, and on using different types of masks as the initial supervision. + +VideoMAE embedding v.s. Learnable embedding. A well-designed embedding is crucial for a hypernet to generate effective model parameters. Some related works [4, 22] employ a learnable embedding optimized jointly with the hypernet, while we directly grab the embedding from the pre-trained MAE encoder and freeze it during hypernet training. We compare the results of using an optimized embedding versus an embedding generated by the MAE encoder, as shown in the first and third rows of Table 4. The advantage in PSNR for the VideoMAE embedding when training for three videos is more than $1\mathrm{dB}$ for the 3-video model. For single-video training, the impact of using either embedding type on the final video quality is minimal. Thus, the advantage of using the VideoMAE Encoder becomes more evident when applied to multiple videos. VideoMAE embedding also shows superiority in unseen videos when using the 3-video metamodel as initialization (see the supplementary materials). Furthermore, using VideoMAE enables robust generalization to new, unseen videos, which is essential for the versatility of our model. + +![](images/f2161798438d153bda39dd8b751a232b936b76521c12b28f8ab623381dc77e8b.jpg) +Figure 6. Qualitative results on the DAVIS dataset. Displayed are the a) original frames, b) predicted masks, c) residual maps, d) texture maps, e) reconstructed frames, and PSNR values for different videos. A color checkerboard overlay is applied to the masked areas to visualize texture transformations. + +Video-based MRHE learned with hypernet v.s. Universal MRHE without hypernet. In this part, we show the benefits of learning multiresolution hash encoding with the hypernet. In multiresolution hash encoding, input 2D features are treated as coordinates within multi-resolution grids and used to sample data from these grids via interpolation. The sampled data are then concatenated and processed through MLPs to generate the final output. In Table 4, we compare the results of i) using hypernet to learn the parameters of the MLPs making these parameters unique for each video (third rows) and ii) directly using the main NVD net to learn MRHE parameters making them common for all videos (second rows). The higher quality results are ob + +tained when MRHE is learned by the hypernet, with benefits that become more apparent as the number of videos used for training increases. This improvement stems from the fact that different video scenes exhibit varying levels of complexity, making video-specific hash grid parameters beneficial when training across multiple videos. + +Accurate mask v.s. Coarse mask. The reference mask is important in our task since it provides a clue about the moving salient objects. We compare the use of manually annotated precise masks (groundtruth from the DAVIS dataset [27]) with coarsely predicted masks from MaskR-CNN [14] as the initial reference for model training. The + +![](images/860a806acd40860df473310c3257e16465ad87b66a4fffea4eb070e29a3d12ad.jpg) +a) Reconstructed frame + +![](images/96c32d2d98d118896a81eaeff01f0fe87deeeafa20541f7f8c43576c65f373ba.jpg) +b) Output 2D textures + +![](images/f226423c8c65e8f647b0602b88091db8e6d2d70bf77d8db774ccc80cf8c29072.jpg) +c) Edited 2D textures + +![](images/e6cf06e38c69535d9e3a050707bd1fe63f478eb4c5402c210ea4919dd6d0b96e.jpg) +d) Edited frame + +![](images/ad892634bc5b026b4e79f8f55d45675ec4be5ca506ae599df7cc4e65c63ae4a2.jpg) + +![](images/1501ab164303ef0bc4a7121ab6c73f4f875b23db2e2ed336240546ba27b560cf.jpg) + +![](images/abc676660e3007ec115fcd879825f60c6dc8e64e2f7487e2f285e47656cbcc15.jpg) + +![](images/878aac63c72d77348d8ac237a45a7e2dfb86ec1b9400d45885d879835aed54b0.jpg) + +![](images/04ccf7e33a0e0ff0f9cd9a04f649aab1798f2d7de7e3b255f661ec4975c7bef1.jpg) +Figure 7. Edited video results. We make adjustments to one of the texture layers—the background layers (rows 1 and 2), and the foreground layer (row 3)—while keeping the other layer unchanged and blending the edited layers together. The edited video maintains good visual quality and smooth transitions. + +![](images/28559146cd6cbe1b15ecdbbf07c1cc5694e0a22257445cddb9a6058495e10658.jpg) + +![](images/e1934625b7583635e28f6f2030aac9c6d825527e998e63df489581dd43d0f036.jpg) + +![](images/b74eb4709b106020eeac7faf5dc599523313455fb3119c7eb29d7d745f0dd89a.jpg) + +Table 5. Ablation on using different masks as initial guidance. + +
Maskhikecar-turnblackswan
PSNRSSIMPSNRSSIMPSNRSSIM
GT30.060.84634.940.90929.530.810
MaskRCNN29.610.84634.920.90329.660.814
+ +predicted alpha map is a soft mask that always includes the corresponding effects, such as reflection or shadows accompanied by the moving objects. Therefore, it is always different from the original object mask. Results in Table 5 show that our model is robust to masks with varying levels of accuracy. Figure 8 illustrates two frames from the "bear" and "blackswan" videos, where a more precise initial mask improves object boundary details. For example, using a more refined mask as initial guidance help the alpha map to capture intricate structures, such as the feathers on a swan's tail. Thus, for higher-quality editing results, it is essential to use the most detailed masks possible during model training. + +# 5. Concluding Remarks + +We present a framework for multiple video training of INR-based video decomposition models using a hypernetwork. We reformulate the task of training a neural video decomposition model on a single video into learning a hypernetwork that generates network parameters based on video embeddings. This approach effectively mitigates the overfitting of the neural video decomposition model to individual videos, enhances generalization capability, and speeds up adaptation to unseen videos. Through experiments, we validate that our proposed model achieves comparable video layer decomposition and rendering capabilities to previous + +![](images/ec1e1a9b5617ad96e148be95a14de72f083029de09603a316fac223bb788c21f.jpg) +Figure 8. Comparison between using accurate and coarse masks as initial guidance. Using different masks does not significantly affect the reconstruction results, but providing a refined mask as the initial guidance helps the model achieve more detailed decomposition, which can be especially beneficial for finer edits. + +methods, while significantly reducing training time on new videos and improving video editing efficiency. + +Acknowledgement. We acknowledge the Departament de Recerca i Universitats with ref. 2021SGR01499 and the CERCA Program from Generalitat de Catalunya, Grant PID2021-128178OB-I00 funded by MCIN/AEI/10.13039/501100011033, ERDF "A way of making Europe", and the grant Catedra ENIA UAB-Cruilla (TSI-100929-2023-2) from the Ministry of Economic Affairs and Digital Transformation of Spain. + +# References + +[1] Adobe. After effects. Accessed: 2024-11-07. 1 +[2] Ali Athar, Alexander Hermans, Jonathon Luiten, Deva Ramanan, and Bastian Leibe. Tarvis: A unified approach for target-based video segmentation. In CVPR, 2023. 1 +[3] BlackMagicDesign. Davinci resolve 19. Accessed: 2024-11-07. 1 +[4] Cheng-Hung Chan, Cheng-Yang Yuan, Cheng Sun, and Hwann-Tzong Chen. Hashing neural video decomposition with multiplicative residuals in space-time. In ICCV, 2023. 1, 2, 4, 5, 6 +[5] Vinod Kumar Chauhan, Jiandong Zhou, Ping Lu, Soheila Molaei, and David A Clifton. A brief review of hypernetworks in deep learning. Artificial Intelligence Review, 57(9): 250, 2024. 2 +[6] Pei-Ze Chiang, Meng-Shiun Tsai, Hung-Yu Tseng, Wei-Sheng Lai, and Wei-Chen Chiu. Stylizing 3d scene via implicit representation and hypernetwork. In WACV, 2022. 2 +[7] Yuren Cong, Mengmeng Xu, Christian Simon, Shoufa Chen, Jiawei Ren, Yanping Xie, Juan-Manuel Perez-Rua, Bodo Rosenhahn, Tao Xiang, and Sen He. Flatten: optical flow-guided attention for consistent text-to-video editing. In ICLR, 2023. 1 +[8] Carl Doersch, Ankush Gupta, Larisa Markeeva, Adria Recasens, Lucas Smaira, Yusuf Aytar, Joao Carreira, Andrew Zisserman, and Yi Yang. Tap-vid: A benchmark for tracking any point in a video. NeurIPS, 2022. 1 +[9] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In ICML, 2017. 2 +[10] Foundry. Nuke. Accessed: 2024-11-07. 1 +[11] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. In ICLR, 2024. 1 +[12] Zeqi Gu, Wenqi Xian, Noah Snavely, and Abe Davis. Factormatte: Redefining video matting for re-composition tasks. ACM TOG, 42(4):1-14, 2023. 2 +[13] David Ha, Andrew Dai, and Quoc V Le. Hypernetworks. In ICLR, 2017. 2 +[14] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, 2017. 7 +[15] Nebojsa Jojic and Brendan J Frey. Learning flexible sprites in video layers. In CVPR, 2001. 2 +[16] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together. In ECCV, 2024. 1 +[17] Yoni Kasten, Dolev Ofri, Oliver Wang, and Tali Dekel. Layered neural atlases for consistent video editing. ACM TOG, 40(6):1-12, 2021. 2, 4, 5 +[18] KeenTools. Geotracker for adobe after effects. Accessed: 2024-11-07. 1 +[19] Chiheon Kim, Doyup Lee, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Generalizable implicit neural representations via instance pattern composers. In CVPR, 2023. 2 +[20] Jaekyeong Lee, Geonung Kim, and Sunghyun Cho. Rna: Video editing with roi-based neural atlas. ACCV, 2024. 1 + +[21] Erika Lu, Forrester Cole, Tali Dekel, Andrew Zisserman, William T Freeman, and Michael Rubinstein. Omnimatte: Associating objects and their effects in video. In CVPR, 2021. 2 +[22] Shishira R Maiya, Anubhav Gupta, Matthew Gwilliam, Max Ehrlich, and Abhinav Shrivastava. Latent-inr: A flexible framework for implicit representations of videos with discriminative semantics. In ECCV, 2024. 2, 6 +[23] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM TOG, 41(4):1-15, 2022. 3 +[24] Seonghyeon Nam, Marcus A Brubaker, and Michael S Brown. Neural image representations for multi-image fusion and layer separation. In ECCV, 2022. 1 +[25] Alex Nichol, Joshua Achiam, and John Schulman. On first-order meta-learning algorithms. arXiv preprint arXiv:1803.02999, 2018. 2 +[26] Hao Ouyang, Qiuyu Wang, Yuxi Xiao, Qingyan Bai, Juntao Zhang, Kecheng Zheng, Xiaowei Zhou, Qifeng Chen, and Yujun Shen. Codef: Content deformation fields for temporally consistent video processing. In CVPR, 2024. 2 +[27] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In CVPR, 2016. 4, 7 +[28] Alex Rav-Acha, Pushmeet Kohli, Carsten Rother, and Andrew Fitzgibbon. Unwrap mosaics: A new representation for video editing. In SIGGRAPH, 2008. 2 +[29] Bipasha Sen, Gaurav Singh, Aditya Agarwal, Rohith Agaram, Madhava Krishna, and Srinath Sridhar. Hypnerf: Learning improved nerf priors using a hypernetwork. NeurIPS, 2024. 2 +[30] Soumyadip Sengupta, Vivek Jayaram, Brian Curless, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Background matting: The world is your green screen. In CVPR, 2020. 2 +[31] Dmitriy Smirnov, Michael Gharbi, Matthew Fisher, Vitor Guizilini, Alexei Efros, and Justin M Solomon. Marionette: Self-supervised sprite learning. NeurIPS, 2021. 1 +[32] Shaolin Su, Qingsen Yan, Yu Zhu, Cheng Zhang, Xin Ge, Jinqiu Sun, and Yanning Zhang. Blindly assess image quality in the wild guided by a self-adaptive hyper network. In CVPR, 2020. 2 +[33] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In ECCV, 2020. 4 +[34] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. VideoMAE: Masked autoencoders are data-efficient learners for self-supervised video pre-training. In NeurIPS, 2022. 1, 2, 3 +[35] Johannes von Oswald, Christian Henning, Benjamin F. Grewe, and João Sacramento. Continual learning with hypernetworks. In ICLR, 2020. 2 +[36] John YA Wang and Edward H Adelson. Representing moving images with layers. IEEE TIP, 3(5):625-638, 1994. 2 +[37] Ning Xu, Brian Price, Scott Cohen, and Thomas Huang. Deep image matting. In CVPR, 2017. 2 + +[38] Vickie Ye, Zhengqi Li, Richard Tucker, Angjoo Kanazawa, and Noah Snively. Deformable sprites for unsupervised video decomposition. In CVPR, 2022. 1, 2, 5 +[39] Emilie Yu, Kevin Blackburn-Matzen, Cuong Nguyen, Oliver Wang, Rubaiat Habib Kazi, and Adrien Bousseau. Videodoodles: Hand-drawn animations on videos with scene-aware canvases. ACM TOG, 42(4):1-12, 2023. 1 \ No newline at end of file diff --git a/2025/HyperNVD_ Accelerating Neural Video Decomposition via Hypernetworks/images.zip b/2025/HyperNVD_ Accelerating Neural Video Decomposition via Hypernetworks/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0697c2482d626280f9771d6c31ecc6399d8102b5 --- /dev/null +++ b/2025/HyperNVD_ Accelerating Neural Video Decomposition via Hypernetworks/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:14c9259468fb01d500b85bbf5a0af6cd4d032a642c0b10abcff89a960487b4dd +size 826822 diff --git a/2025/HyperNVD_ Accelerating Neural Video Decomposition via Hypernetworks/layout.json b/2025/HyperNVD_ Accelerating Neural Video Decomposition via Hypernetworks/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..34c03381bbe6443677b9f071b750c5fa50e5db7e --- /dev/null +++ b/2025/HyperNVD_ Accelerating Neural Video Decomposition via Hypernetworks/layout.json @@ -0,0 +1,6963 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 76, + 103, + 536, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 103, + 536, + 122 + ], + "spans": [ + { + "bbox": [ + 76, + 103, + 536, + 122 + ], + "type": "text", + "content": "HyperNVD: Accelerating Neural Video Decomposition via Hypernetworks" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 143, + 142, + 466, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 142, + 466, + 185 + ], + "spans": [ + { + "bbox": [ + 143, + 142, + 466, + 185 + ], + "type": "text", + "content": "Maria Pilligua1,2,* Danna Xue1,2,*† Javier Vazquez-Corral1,2 \n1 Universitat Autònoma de Barcelona 2 Computer Vision Center {mpilligua, dxue, javier.vazquez}@cvc.uab.cat" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 238, + 297, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 238, + 297, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 297, + 430 + ], + "type": "text", + "content": "Decomposing a video into a layer-based representation is crucial for easy video editing for the creative industries, as it enables independent editing of specific layers. Existing video-layer decomposition models rely on implicit neural representations (INRs) trained independently for each video, making the process time-consuming when applied to new videos. Noticing this limitation, we propose a meta-learning strategy to learn a generic video decomposition model to speed up the training on new videos. Our model is based on a hypernetwork architecture which, given a video-encoder embedding, generates the parameters for a compact INR-based neural video decomposition model. Our strategy mitigates the problem of single-video overfitting and, importantly, shortens the convergence of video decomposition on new, unseen videos. Our code is available at: https://hypernvd.github.io/." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 451, + 135, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 451, + 135, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 451, + 135, + 464 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 471, + 295, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 471, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 54, + 471, + 295, + 578 + ], + "type": "text", + "content": "Video editing plays a crucial role in today's creative industries, from film-making and advertising to social media content creation. It enables creators to bring their visions to life, enhancing storytelling through visual effects, scene adjustments, and content personalization. However, most current video editing techniques face several challenges: they require extensive manual intervention [39], lack consistent control over edits [7, 11], and are time-consuming [4, 24], limiting their scalability for real-world applications." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 579, + 295, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 579, + 295, + 687 + ], + "spans": [ + { + "bbox": [ + 54, + 579, + 295, + 687 + ], + "type": "text", + "content": "Current video editing workflows in professional software resemble traditional 2D animation. Creators import video into editing software [1, 3] and edit in overlaid layers for each video frame. While motion tracking tools can interpolate adjustments between keyframes [10, 18], handling occlusions and ensuring consistency remain challenging since tracking errors make editing cumbersome. These problems highlight the importance of accurate layered representation and frame mappings." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 213, + 553, + 369 + ], + "blocks": [ + { + "bbox": [ + 316, + 213, + 553, + 369 + ], + "lines": [ + { + "bbox": [ + 316, + 213, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 316, + 213, + 553, + 369 + ], + "type": "image", + "image_path": "35f53945dc588cf7f23ca3d6e777a5bfb6f42e01a6a1044c3f156ed42fbcc19c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 373, + 555, + 429 + ], + "lines": [ + { + "bbox": [ + 313, + 373, + 555, + 429 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 555, + 429 + ], + "type": "text", + "content": "Figure 1. Fine-tuning the video decomposition model from our metamodel (HyperNVD) versus training from scratch on unseen videos shows clear advantages. With initialization from our HyperNVD trained on 15 videos, the model converges faster to the same PSNR and ultimately achieves better performance." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 437, + 555, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 555, + 653 + ], + "type": "text", + "content": "Layer-based video decomposition methods [4, 20, 24, 31, 38] represent video with several texture layers, where each layer consistently represents specific video content (e.g., moving objects or backgrounds). These approaches are mainly based on compact implicit neural representations (INRs), which encode each video individually. The model projects pixel coordinates and frame indices into a canonical 2D texture space, and then, it maps to RGB values to reconstruct each frame. In this way, it is possible to apply simple edits on the canonical texture layer and then propagate them to the entire video, greatly relieving the need for manual adjustments. Furthermore, the parameter sizes and training times are reduced compared to training a generalized video object segmentation model [2] and global flow estimation model [8, 16]. However, due to the nature of INR, the neural video decomposition model lacks generalization and requires tens of minutes to train a " + }, + { + "bbox": [ + 313, + 437, + 555, + 653 + ], + "type": "inline_equation", + "content": "480\\mathrm{p}" + }, + { + "bbox": [ + 313, + 437, + 555, + 653 + ], + "type": "text", + "content": " video [4]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": "In this paper, we propose a meta-learning strategy in layer-based video decomposition to speed up the training. More in detail, we introduce a hypernetwork (hypernet) that, based on the video embedding predicted by VideoMAE [34], generates the parameters for a compact INR" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 223, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 223, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 223, + 703 + ], + "type": "text", + "content": "*These authors contributed equally to this work." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22933" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": "model. Our strategy facilitates the generalization to more than one video and also allows for speed-up training on new, unseen videos, as we can use the parameters given by the hypernet to initialize the neural video decomposition model for further fine-tuning. Figure 1 shows three examples of this speed-up. In this figure, we compare the training time and the PSNR obtained for training an unseen video starting from the parameters suggested by our hypernet or from a random parameter set. Our method obtains the same PSNR with more than 30 minutes less." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 198, + 277, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 198, + 277, + 209 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 277, + 209 + ], + "type": "text", + "content": "Contributions. Our main contributions are as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 212, + 295, + 331 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 56, + 212, + 295, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 212, + 295, + 270 + ], + "spans": [ + { + "bbox": [ + 56, + 212, + 295, + 270 + ], + "type": "text", + "content": "- We incorporate a meta-learning strategy to learn a generic video decomposition method. Our strategy consists of learning a hypernet that, given a video-encoder embedding, predicts the parameters for an INR-based video decomposition model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 271, + 295, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 271, + 295, + 306 + ], + "spans": [ + { + "bbox": [ + 56, + 271, + 295, + 306 + ], + "type": "text", + "content": "- Our framework relieves the problem of overfitting on a single video and shortens the convergence time of the neural video decomposition model on unseen videos." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 307, + 294, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 307, + 294, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 307, + 294, + 331 + ], + "type": "text", + "content": "- The experimental results show the generalizability and appealing video editing ability of our approach." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 341, + 142, + 353 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 341, + 142, + 353 + ], + "spans": [ + { + "bbox": [ + 55, + 341, + 142, + 353 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 361, + 214, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 214, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 214, + 373 + ], + "type": "text", + "content": "2.1. Layered video decomposition" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 378, + 295, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 378, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 378, + 295, + 666 + ], + "type": "text", + "content": "In image editing and animation, an image is often separated into multiple layers, each with color and opacity [30, 37], allowing independent adjustments before stacking them for the final result. This layered representation is also applied to video editing for simplifying the workflow and ensuring consistency [15, 28, 36]. Neural video decomposition approaches [4, 17, 38] decompose videos into consistent 2D texture layers across frames and are usually trained in a self-supervised manner using motion and appearance constraints. LNA [17] uses implicit neural representations to reconstruct videos via neural atlases and alpha blending. While encoding each video individually reduces parameter size and rendering time, reconstructing high-frequency details and long training time remain challenges. Omnimatte [21] and Factormatte [12] add appearance priors to model associated effects like shadows, or reflections among others. Hashing-nvd [4] and CoDeF [26] introduce hash encoding to accelerate optimization. However, these models still need more than 40 minutes of training on one single video [4]. We propose to use a hypernet to generate parameters of the video decomposition INR. In this way, we can further shorten the training time of any new, unseen video by initializing its training from the weights given by the hypernet." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 672, + 146, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 146, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 146, + 685 + ], + "type": "text", + "content": "2.2. Hypernetwork" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "Hypernetworks (or hypernets) [5, 13] is a technique that enhances the adaptability and flexibility of deep learn" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 553, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 300 + ], + "type": "text", + "content": "ing models. Instead of learning task-specific models with fixed parameters, a hypernet dynamically generates parameters for a target model based on task [9, 25, 35], or data [19, 22, 29, 32]. By learning a shared, adaptable representation, hypernets help target models generalize better across different conditions. Once trained, they use embeddings as input to modulate the target model's weights and reduce optimization times on unseen tasks [9, 25]. Recent work has further improved input-specific model flexibility, especially in INR applications [6, 19, 22]. The work most closely related to ours is HyP-NeRF [29], which employs a hypernet to generate instance-specific parameters for hash encoding and NeRF, linking text to 3D object reconstructions. This framework relates semantic similarities across models for different instances but focuses on simple 3D objects with limited shape and color variations. In contrast, our video modeling task is more complex and requires attention to multiple objects, detailed motion, and varied appearances." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 310, + 553, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 310, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 310, + 553, + 338 + ], + "type": "text", + "content": "3. Accelerate Neural Video Decomposition via Hypernetwork" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 343, + 436, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 343, + 436, + 355 + ], + "spans": [ + { + "bbox": [ + 313, + 343, + 436, + 355 + ], + "type": "text", + "content": "3.1. Framework overview" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 361, + 555, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 555, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 555, + 540 + ], + "type": "text", + "content": "Our goal is to learn a generalizable model for neural video decomposition that helps speed up the optimization of unseen videos. Our HyperNVD consists of three main components: i) a video embedding generation model, ii) a hypernet to generate model parameters, and iii) a neural video decomposition (NVD) model. Figure 2 shows an overview of our model. First, we compress the features predicted by a pre-trained video transformer, VideoMAE [34], and use the compressed embedding as input to the hypernet. Then the hypernet predicts the weights and the multiresolution hash encoding parameters of the target NVD model. Our architecture allows the training of multiple videos simultaneously, which enables fast adaptation of unseen videos. This fast adaptation is accomplished by considering the parameters given by the hypernet for further fine-tuning." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 546, + 553, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 553, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 553, + 605 + ], + "type": "text", + "content": "Neural video decomposition (NVD) model. Standard layered video decomposition models based on INR [4] work by taking video coordinates and frame indices " + }, + { + "bbox": [ + 313, + 546, + 553, + 605 + ], + "type": "inline_equation", + "content": "p = (x, y, t)" + }, + { + "bbox": [ + 313, + 546, + 553, + 605 + ], + "type": "text", + "content": " as input and decomposing the video into a background layer and a foreground layer." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 606, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 553, + 713 + ], + "type": "text", + "content": "As shown in the Layer Module of Figure 2, the foreground layer representation " + }, + { + "bbox": [ + 313, + 606, + 553, + 713 + ], + "type": "inline_equation", + "content": "LM_{f}" + }, + { + "bbox": [ + 313, + 606, + 553, + 713 + ], + "type": "text", + "content": " is encoded by a mapping module " + }, + { + "bbox": [ + 313, + 606, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_f" + }, + { + "bbox": [ + 313, + 606, + 553, + 713 + ], + "type": "text", + "content": ", a texture module " + }, + { + "bbox": [ + 313, + 606, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{f}" + }, + { + "bbox": [ + 313, + 606, + 553, + 713 + ], + "type": "text", + "content": ", and a residual module " + }, + { + "bbox": [ + 313, + 606, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_f" + }, + { + "bbox": [ + 313, + 606, + 553, + 713 + ], + "type": "text", + "content": ". The mapping module " + }, + { + "bbox": [ + 313, + 606, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M}f" + }, + { + "bbox": [ + 313, + 606, + 553, + 713 + ], + "type": "text", + "content": " transforms the input coordinate " + }, + { + "bbox": [ + 313, + 606, + 553, + 713 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 606, + 553, + 713 + ], + "type": "text", + "content": " to the texture space coordinate " + }, + { + "bbox": [ + 313, + 606, + 553, + 713 + ], + "type": "inline_equation", + "content": "(u,v)_f^p" + }, + { + "bbox": [ + 313, + 606, + 553, + 713 + ], + "type": "text", + "content": ". The texture module " + }, + { + "bbox": [ + 313, + 606, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_f" + }, + { + "bbox": [ + 313, + 606, + 553, + 713 + ], + "type": "text", + "content": " converts the texture coordinate into an RGB value " + }, + { + "bbox": [ + 313, + 606, + 553, + 713 + ], + "type": "inline_equation", + "content": "c_{f}" + }, + { + "bbox": [ + 313, + 606, + 553, + 713 + ], + "type": "text", + "content": ", while the residual module " + }, + { + "bbox": [ + 313, + 606, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_f" + }, + { + "bbox": [ + 313, + 606, + 553, + 713 + ], + "type": "text", + "content": " predicts a framewise residual coefficient " + }, + { + "bbox": [ + 313, + 606, + 553, + 713 + ], + "type": "inline_equation", + "content": "r_f" + }, + { + "bbox": [ + 313, + 606, + 553, + 713 + ], + "type": "text", + "content": ". Only " + }, + { + "bbox": [ + 313, + 606, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{T}f" + }, + { + "bbox": [ + 313, + 606, + 553, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 606, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{R}f" + }, + { + "bbox": [ + 313, + 606, + 553, + 713 + ], + "type": "text", + "content": " incorporate Multi-resolution Hashing Encoding (MRHE) to accelerate" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22934" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 68, + 542, + 320 + ], + "blocks": [ + { + "bbox": [ + 69, + 68, + 542, + 320 + ], + "lines": [ + { + "bbox": [ + 69, + 68, + 542, + 320 + ], + "spans": [ + { + "bbox": [ + 69, + 68, + 542, + 320 + ], + "type": "image", + "image_path": "30f300b8457157da4a08ad10ad736aa3390c9c5674a9fa58240cdb61da9b12f3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 328, + 555, + 406 + ], + "lines": [ + { + "bbox": [ + 54, + 328, + 555, + 406 + ], + "spans": [ + { + "bbox": [ + 54, + 328, + 555, + 406 + ], + "type": "text", + "content": "Figure 2. The architecture of our HyperNVD. HyperNVD consists of " + }, + { + "bbox": [ + 54, + 328, + 555, + 406 + ], + "type": "inline_equation", + "content": "i)" + }, + { + "bbox": [ + 54, + 328, + 555, + 406 + ], + "type": "text", + "content": " the MAE encoder to generate video embedding, " + }, + { + "bbox": [ + 54, + 328, + 555, + 406 + ], + "type": "inline_equation", + "content": "ii)" + }, + { + "bbox": [ + 54, + 328, + 555, + 406 + ], + "type": "text", + "content": " the hypernet to generate model parameters, and " + }, + { + "bbox": [ + 54, + 328, + 555, + 406 + ], + "type": "inline_equation", + "content": "iii)" + }, + { + "bbox": [ + 54, + 328, + 555, + 406 + ], + "type": "text", + "content": " the target neural video decomposition (NVD) model. Given an input video, the MAE encoder encodes it to a compact embedding " + }, + { + "bbox": [ + 54, + 328, + 555, + 406 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 54, + 328, + 555, + 406 + ], + "type": "text", + "content": ". Then the hypernet " + }, + { + "bbox": [ + 54, + 328, + 555, + 406 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 54, + 328, + 555, + 406 + ], + "type": "text", + "content": " generates the parameters of the NVD model, including the MultiResolution Hash Encoding (MRHE) and the model weights. The NVD model includes two layer modules (foreground and background) for reconstructing different components in the video and an alpha module for predicting the soft mask to blend the layers. The reconstructed frame is generated by adding the foreground and background texture layers with an opacity map. Each layer module consists of a mapping module, a texture module, and a residual module." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 421, + 294, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 421, + 294, + 516 + ], + "spans": [ + { + "bbox": [ + 54, + 421, + 294, + 516 + ], + "type": "text", + "content": "training and inference. The MRHE (illustrated in the rightmost block of Figure 2) performs grid hash encoding [23]. Here, the input 2D features serve as coordinates within multi-resolution grids, allowing for interpolation-based data sampling. The sampled data is then concatenated and processed through MLPs. The final output of the foreground layer is computed as " + }, + { + "bbox": [ + 54, + 421, + 294, + 516 + ], + "type": "inline_equation", + "content": "\\hat{c}_f^p = c_f^p * r_f^p" + }, + { + "bbox": [ + 54, + 421, + 294, + 516 + ], + "type": "text", + "content": ". The background layer follows the same structure." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 517, + 295, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 295, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 295, + 577 + ], + "type": "text", + "content": "The background and foreground representations are later combined by an alpha module " + }, + { + "bbox": [ + 55, + 517, + 295, + 577 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 55, + 517, + 295, + 577 + ], + "type": "text", + "content": ". This module takes the coordinate " + }, + { + "bbox": [ + 55, + 517, + 295, + 577 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 517, + 295, + 577 + ], + "type": "text", + "content": " as input and predicts a soft opacity map for each layer. The final reconstructed RGB value, " + }, + { + "bbox": [ + 55, + 517, + 295, + 577 + ], + "type": "inline_equation", + "content": "\\hat{c}^p" + }, + { + "bbox": [ + 55, + 517, + 295, + 577 + ], + "type": "text", + "content": ", is obtained by weighted summation across both layers:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 122, + 587, + 295, + 602 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 587, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 122, + 587, + 295, + 602 + ], + "type": "interline_equation", + "content": "\\hat {c} ^ {p} = \\left(1 - \\alpha^ {p}\\right) \\hat {c} _ {b} ^ {p} + \\alpha^ {p} \\hat {c} _ {f} ^ {p}. \\tag {1}", + "image_path": "6c8d984993a9de1a6c2419ae85b2fb570d00ef5f225642d31194b6e9eb2aa2cc.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 611, + 282, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 282, + 622 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 282, + 622 + ], + "type": "text", + "content": "All the modules in the above model structure are MLPs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 624, + 296, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 296, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 296, + 672 + ], + "type": "text", + "content": "Note that our explanation can easily be switched to multiple foreground layers based on the masks provided by users; we explained the case of a single foreground layer for simplification." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Hypernet. Our hypernet, " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": ", predicts the INR parameters given a condition embedding " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": " —see the Hypernet block in Figure 2. It comprises a series of MLPs, each dedicated" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 421, + 555, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 421, + 555, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 421, + 555, + 540 + ], + "type": "text", + "content": "to generating the parameters of a specific layer within the NVD model. The target of " + }, + { + "bbox": [ + 313, + 421, + 555, + 540 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 313, + 421, + 555, + 540 + ], + "type": "text", + "content": " is to generate accurate weights for an NVD model capable of robust video decomposition. To accomplish this, we train it using the same loss functions and constraints applied in single-video decomposition models, as the objective for each video remains unchanged. However, by conditioning the hypernet on unique embeddings for each video, the model learns a generalized decomposition capability that extends beyond individual instances." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 552, + 556, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 556, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 556, + 685 + ], + "type": "text", + "content": "VideoMAE encoder. Typically, hypernet inputs use either learnable embeddings trained jointly with the hypernet or embeddings generated from input data. We use the latter, as representing high-dimensional video data in a compact, low-dimensional embedding is particularly challenging. This approach also supports generalization on multiple videos and enables inference on unseen videos, which is crucial for building a robust NVD model. Moreover, this video embedding needs to encapsulate both motion and scene information, help reduce optimization complexity, and accelerate network convergence." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "To achieve this, we use the features predicted by VideoMAE [34], a frozen self-supervised masked autoencoder" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22935" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 296, + 135 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 296, + 135 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 296, + 135 + ], + "type": "image", + "image_path": "0623f59a7f60469f76d075b1a83f4243b2ba9715297e28cb1ae3ed45aa6b036d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 144, + 295, + 199 + ], + "lines": [ + { + "bbox": [ + 55, + 144, + 295, + 199 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 295, + 199 + ], + "type": "text", + "content": "Figure 3. Training of the autoencoder to get a compressed embedding from VideoMAE. First, we extract features " + }, + { + "bbox": [ + 55, + 144, + 295, + 199 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 55, + 144, + 295, + 199 + ], + "type": "text", + "content": " from each 3D patch using the pre-trained VideoMAE. We then train an autoencoder to obtain a compressed embedding " + }, + { + "bbox": [ + 55, + 144, + 295, + 199 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 55, + 144, + 295, + 199 + ], + "type": "text", + "content": " by minimizing the difference between the output " + }, + { + "bbox": [ + 55, + 144, + 295, + 199 + ], + "type": "inline_equation", + "content": "\\hat{o}" + }, + { + "bbox": [ + 55, + 144, + 295, + 199 + ], + "type": "text", + "content": " and the input " + }, + { + "bbox": [ + 55, + 144, + 295, + 199 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 55, + 144, + 295, + 199 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 214, + 295, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 214, + 295, + 322 + ], + "spans": [ + { + "bbox": [ + 54, + 214, + 295, + 322 + ], + "type": "text", + "content": "trained on video data. As shown in Figure 3, to further reduce the dimension of the VideoMAE output, we use an encoder-decoder structure to learn a compressed embedding of size (768, 1) from the original VideoMAE output of size (768, #patches). This ensures the encoder consolidates information from all patches into a single dense embedding while minimizing information loss. The autoencoder is trained using " + }, + { + "bbox": [ + 54, + 214, + 295, + 322 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 54, + 214, + 295, + 322 + ], + "type": "text", + "content": " loss to enforce the closeness between the reconstructed output " + }, + { + "bbox": [ + 54, + 214, + 295, + 322 + ], + "type": "inline_equation", + "content": "\\hat{o}" + }, + { + "bbox": [ + 54, + 214, + 295, + 322 + ], + "type": "text", + "content": " and the original embedding " + }, + { + "bbox": [ + 54, + 214, + 295, + 322 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 54, + 214, + 295, + 322 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 322, + 296, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 322, + 296, + 406 + ], + "spans": [ + { + "bbox": [ + 55, + 322, + 296, + 406 + ], + "type": "text", + "content": "We take the VideoMAE together with the encoder part as our VideoMAE encoder in HyperNVD (the top-left module in Figure 2). This encoder is pre-trained and frozen during the hypernet training, allowing us to maintain the rich representations it learned while preserving essential information for the decomposition tasks. In this way, the encoder serves as a compact, fixed embedding provider for the hypernet." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 414, + 239, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 414, + 239, + 427 + ], + "spans": [ + { + "bbox": [ + 55, + 414, + 239, + 427 + ], + "type": "text", + "content": "3.2. Training and rendering procedure" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 432, + 295, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 432, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 432, + 295, + 479 + ], + "type": "text", + "content": "Loss functions. Since all the components of our model are fully differentiable, we train them end-to-end via stochastic gradient descent. We use losses from the previous work [4, 17] to facilitate the training process. The losses include:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 480, + 295, + 586 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 55, + 480, + 295, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 480, + 295, + 503 + ], + "spans": [ + { + "bbox": [ + 55, + 480, + 295, + 503 + ], + "type": "text", + "content": "- Reconstruction loss to ensure the reconstruction quality of the video." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 504, + 295, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 295, + 528 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 295, + 528 + ], + "type": "text", + "content": "- Consistency loss which provides supervision of optical flow for accurate motion representation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 528, + 295, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 528, + 295, + 551 + ], + "spans": [ + { + "bbox": [ + 56, + 528, + 295, + 551 + ], + "type": "text", + "content": "- Sparsity loss to prevent duplicate content in different texture layers." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 552, + 295, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 552, + 295, + 586 + ], + "spans": [ + { + "bbox": [ + 56, + 552, + 295, + 586 + ], + "type": "text", + "content": "- Residual consistency loss to ensure smooth lighting conditions while preventing the residual estimator from absorbing all color details." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 588, + 259, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 259, + 600 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 259, + 600 + ], + "type": "text", + "content": "Two losses are applied just in the initial iterations:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 600, + 295, + 647 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 600, + 295, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 600, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 600, + 295, + 624 + ], + "type": "text", + "content": "- Rigidity loss to maintain the structural rigidity in the canonical texture layer." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 624, + 295, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 295, + 647 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 295, + 647 + ], + "type": "text", + "content": "- Alpha bootstrapping loss, which supervises coarse masks for foreground separation." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 647, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 295, + 696 + ], + "type": "text", + "content": "Since the model is trained in a self-supervised manner, the additional supervision we use are the ground truth object masks and the optical flow predicted by RAFT [33]. Further details are in the supplementary materials." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "type": "text", + "content": "Training with the hypernet. The objective is to train a" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "text", + "content": "hypernet capable of generating parameters for the NVD model, so only the hypernet weights are trainable, while the NVD model serves as a set of differentiable layers that enable backpropagation to the hypernet, allowing it to learn from the above-mentioned losses. We begin the training process with a pre-training step that configures the mapping network to predict textures that are both aligned and in an initial rectangular shape corresponding to the frame size. Without this initial step, the model could produce incorrectly oriented textures. With this pre-training, the convergence during the subsequent training phase is significantly faster, as the model starts from a more informed initialization." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 234, + 555, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 234, + 555, + 282 + ], + "spans": [ + { + "bbox": [ + 313, + 234, + 555, + 282 + ], + "type": "text", + "content": "Fine-tune on new videos. Once our model is trained on a set of videos, we can use the hypernet to generate parameters for a new video and treat these weights as an initialization for the NVD model (see Figure 1 and Section 4.2)." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 287, + 556, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 287, + 556, + 443 + ], + "spans": [ + { + "bbox": [ + 313, + 287, + 556, + 443 + ], + "type": "text", + "content": "Video editing and rendering. The NVD model decomposes an input video into multiple layers, each with distinct textures. For video editing, we first perform inference to obtain these texture layers. Then, each layer can be edited individually with image editing software or other techniques. For editing, we render the texture networks using a " + }, + { + "bbox": [ + 313, + 287, + 556, + 443 + ], + "type": "inline_equation", + "content": "1000 \\times 1000" + }, + { + "bbox": [ + 313, + 287, + 556, + 443 + ], + "type": "text", + "content": " grid sampling. The rendered texture is then modified, and the colors are sampled through bilinear interpolation for smooth transitions and precise adjustments. By replacing the original layers with the modified ones and mapping the corresponding coordinates back to each video frame, we can generate the edited video with temporally consistent changes." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 455, + 397, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 455, + 397, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 455, + 397, + 468 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 475, + 446, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 446, + 487 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 446, + 487 + ], + "type": "text", + "content": "4.1. Implementation details" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 492, + 555, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 492, + 555, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 492, + 555, + 552 + ], + "type": "text", + "content": "Datasets. We conduct our experiments mainly on the DAVIS dataset [27]. We train HyperNVD with up to 30 videos, while the rest videos are reserved for testing. Each video consists of the first 16 frames at a resolution of " + }, + { + "bbox": [ + 313, + 492, + 555, + 552 + ], + "type": "inline_equation", + "content": "768 \\times 432" + }, + { + "bbox": [ + 313, + 492, + 555, + 552 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": "Training settings. We use a batch size of 100,000 point locations and train each model for around 25,000 iterations and 100 pre-train iterations. The hypernet consists of approximately 290 million learnable parameters. The main network consists of approximately 4.4 million learnable parameters: 1.75 million for each layer module and 965 thousand for the alpha module. Training and evaluation with 1 video and a batch size of 100,000 requires around 22GB of GPU memory, with a checkpoint size of 2.3GB. Training our model takes roughly 2.8 hours for 1 video and 12.5 hours for 3 videos on an NVIDIA RTX 3090 GPU. We optimize all networks simultaneously using the Adam optimizer, with a learning rate of 1e-2 for the hash encoding" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22936" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 93, + 87, + 515, + 159 + ], + "blocks": [ + { + "bbox": [ + 127, + 70, + 482, + 82 + ], + "lines": [ + { + "bbox": [ + 127, + 70, + 482, + 82 + ], + "spans": [ + { + "bbox": [ + 127, + 70, + 482, + 82 + ], + "type": "text", + "content": "Table 1. Comparison with previous approaches on reconstruction quality for single video training." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 93, + 87, + 515, + 159 + ], + "lines": [ + { + "bbox": [ + 93, + 87, + 515, + 159 + ], + "spans": [ + { + "bbox": [ + 93, + 87, + 515, + 159 + ], + "type": "table", + "html": "
Methodshikeblackswancar-turnbearlucia
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
Deformable Sprites [38]25.300.79525.750.73230.740.85026.340.81626.540.771
Layered Neural Atlases [17]30.020.88229.960.85334.390.92929.620.91029.920.888
Hashing-nvd [4]29.120.87330.300.84335.430.89531.560.90829.290.866
HyperNVD (Ours)30.060.84629.530.81034.940.90931.580.90130.300.865
", + "image_path": "c10367415ef31265382ccd0e0ab6316842557fb16001a33c5b5422c290e983ef.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 58, + 186, + 552, + 268 + ], + "blocks": [ + { + "bbox": [ + 104, + 168, + 504, + 179 + ], + "lines": [ + { + "bbox": [ + 104, + 168, + 504, + 179 + ], + "spans": [ + { + "bbox": [ + 104, + 168, + 504, + 179 + ], + "type": "text", + "content": "Table 2. Comparison of the reconstruction quality of our model when trained with different numbers of videos." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 186, + 552, + 268 + ], + "lines": [ + { + "bbox": [ + 58, + 186, + 552, + 268 + ], + "spans": [ + { + "bbox": [ + 58, + 186, + 552, + 268 + ], + "type": "table", + "html": "
Num of videoshikeblackswancar-turnbearluciakite-walkbusstrollerboatrollerblade
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
130.060.84629.530.81034.940.90931.580.90130.300.86533.350.83732.450.88931.250.83032.770.87228.750.810
327.310.72827.670.75431.870.850--------------
527.130.76027.320.73431.430.84528.880.85427.430.760----------
1027.990.75428.860.73632.470.85229.990.85328.500.76331.750.79729.450.83428.800.77531.070.82826.060.740
1527.650.76827.410.75632.570.87829.710.85228.540.83331.990.79629.600.84428.920.78030.910.82626.130.742
3027.240.79527.300.74531.640.86429.280.86227.310.80230.930.77128.630.82727.000.72029.980.80324.350.686
", + "image_path": "67dc6a1be9c14c56fb7915d8995a721800b73395db238721dc1311c5fda323ac.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 289, + 220, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 220, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 220, + 300 + ], + "type": "text", + "content": "and a learning rate of 5e-4 for the others." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 307, + 295, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 307, + 295, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 307, + 295, + 415 + ], + "type": "text", + "content": "Details of architecture. In the target NVD model, we apply Rectified Linear Unit (ReLU) activations between hidden layers, with outputs passed through a tanh function. This enables the model to map texture to a predefined continuous range. The hypernet comprises a series of MLPs. Each MLP consists of four fully connected layers, with a hidden dimension of 64 and intermediate ReLU activations. Further details about the architecture can be found in the supplementary material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 425, + 170, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 425, + 170, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 425, + 170, + 437 + ], + "type": "text", + "content": "4.2. Quantitative results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 443, + 295, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 443, + 295, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 443, + 295, + 588 + ], + "type": "text", + "content": "Comparison with previous methods. We present the comparison of PSNR and SSIM values between DeformableSprites [38], Layered Neural Atlases [17], Hashing-nvd [4] and our HyperNVD in Table 1. We report the results of our model trained on one single video independently. For a fair comparison, we use the ground truth mask as initial guidance for all the comparing methods. Results in Table 1 demonstrate that incorporating the hypernet does not degrade the model's performance. Ours maintains the effectiveness of the original architecture (i.e. our baseline Hashing-nvd) while introducing the flexibility of a hypernet-based approach." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": "Learning multiple videos. To evaluate the model's generalization ability across multiple videos, we train HyperNVD with 1, 3, 5, 10, 15, and 30 videos, respectively. For consistency, videos included in smaller-scale settings are subsets of those in larger-scale settings. Table 2 shows the PSNR of the videos reconstructed by models trained with different numbers of videos. When training with multiple videos, we observe a slight drop in reconstruction quality compared to training on a single video. This is likely due to the model potentially overfitting to the unique characteristics of a sin" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 289, + 553, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 553, + 336 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 553, + 336 + ], + "type": "text", + "content": "gle video. However, the PSNR only decreases by approximately 3dB as the number of videos increases from 1 to 30, which indicates that the model retains strong performance even with larger datasets." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 336, + 554, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 336, + 554, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 336, + 554, + 456 + ], + "type": "text", + "content": "Figure 4 further illustrates the image quality of the \"hike\" video when training HyperNVD with different numbers of videos. The models trained on multiple videos can still reconstruct the low-frequency parts well, but some high-frequency details are lost. This stable evolution of PSNR indifferent to the number of videos indicates a plateau in performance with larger video sets. Figure 5 shows the reconstruction quality of the video \"hike\" across different iteration steps. The resulting curves are very similar for all the cases trained with multiple videos." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 461, + 554, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 554, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 554, + 641 + ], + "type": "text", + "content": "Finetuning on new videos. To evaluate the ability of fast adaption on unseen videos, we train the NVD model initialized by the parameters of our metamodel trained with 15 videos and compare it with the model trained from scratch. Note that the NVD model is trained independently without the hypernet. As shown in Table 3, training from our metamodel (i.e. HyperNVD) consistently improves performance by " + }, + { + "bbox": [ + 313, + 461, + 554, + 641 + ], + "type": "inline_equation", + "content": "0.8\\mathrm{dB}" + }, + { + "bbox": [ + 313, + 461, + 554, + 641 + ], + "type": "text", + "content": " across different videos compared to training from scratch (i.e. Hashing-nvd). Figure 1 further shows our model reaches the same PSNR on the unseen videos more than 30 minutes faster, and with more stable training. This demonstrates that our metamodel enables faster adaptation to new scenes, which improves the efficiency of video editing workflow. Additional results are provided in the supplementary material." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 648, + 423, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 648, + 423, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 648, + 423, + 661 + ], + "type": "text", + "content": "4.3. Qualitative results" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 666, + 553, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 553, + 715 + ], + "type": "text", + "content": "Figure 6 shows examples of our layer decomposition results. For each video frame, we include a) the original frame, b) the predicted background and foreground components with a checkerboard overlay to illustrate texture" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22937" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 88, + 553, + 140 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 550, + 82 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 550, + 82 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 550, + 82 + ], + "type": "text", + "content": "Table 3. Reconstruction quality in PSNR on unseen videos when fine-tuning from our 15-video metamodel versus training from scratch." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 88, + 553, + 140 + ], + "lines": [ + { + "bbox": [ + 56, + 88, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 88, + 553, + 140 + ], + "type": "table", + "html": "
Methodscowssoccerballsoapboxfamingohorse-jumpgoatswingkite-surflibbysurf
From scratch27.0428.5131.2829.0327.6127.0629.6531.6529.6333.61
From metamodel28.4030.1632.7429.9628.5027.7830.3732.1930.1133.86
Improvements+1.36+1.65+1.46+0.93+0.89+0.72+0.72+0.54+0.48+0.25
", + "image_path": "928622741847b16fd477c88f7746089bc5908473bc8e054c68902b414e834b17.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 56, + 159, + 296, + 245 + ], + "blocks": [ + { + "bbox": [ + 56, + 159, + 296, + 245 + ], + "lines": [ + { + "bbox": [ + 56, + 159, + 296, + 245 + ], + "spans": [ + { + "bbox": [ + 56, + 159, + 296, + 245 + ], + "type": "image", + "image_path": "06978cdae851dd78eed78193589a8685d03eff59fa2016a09f85d19bfc945cd2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 253, + 296, + 308 + ], + "lines": [ + { + "bbox": [ + 55, + 253, + 296, + 308 + ], + "spans": [ + { + "bbox": [ + 55, + 253, + 296, + 308 + ], + "type": "text", + "content": "Figure 4. Comparison of the video \"hike\" reconstructed by models trained with different numbers of videos. The top row shows the full frames, while the bottom row zooms in on a specific patch. Models trained on multiple videos capture the primary structure of the frame but tend to lose some high-frequency details." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 57, + 323, + 294, + 427 + ], + "blocks": [ + { + "bbox": [ + 57, + 323, + 294, + 427 + ], + "lines": [ + { + "bbox": [ + 57, + 323, + 294, + 427 + ], + "spans": [ + { + "bbox": [ + 57, + 323, + 294, + 427 + ], + "type": "image", + "image_path": "e9c832804d1ea254f6e3106c82e5ea2de5365b79e71f8b89d89fd56914c50725.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 432, + 296, + 487 + ], + "lines": [ + { + "bbox": [ + 55, + 432, + 296, + 487 + ], + "spans": [ + { + "bbox": [ + 55, + 432, + 296, + 487 + ], + "type": "text", + "content": "Figure 5. Comparison of the reconstruction quality for the \"hike\" video trained with different numbers of videos. Our HyperNVD model trained on multiple videos shows a 3dB drop in PSNR compared to single-video training. The PSNR remains stable across multiple video training, suggesting a performance plateau." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 509, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 295, + 604 + ], + "type": "text", + "content": "mapping transformations, c) the predicted residual maps, d) the complete canonical texture map, and e) the reconstructed frame. Our model obtains accurate video reconstruction without introducing noticeable texture distortions. The predicted opacity map separates the foreground and background correctly. The residuals learn the changing reflections on the water and shadows as shown in the \"black-swan\" and \"boat\" videos." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": "To further showcase the effectiveness of our approach for editing, we demonstrate editing examples in Figure 7. In the first two examples, only the background is modified, while in the last example, we edit the foreground texture. For the first example, the \"blackswan\" video, we transform the plants in the background into flowers and adjust the color of the water. The ripples and shadows in the water maintain a natural appearance, and the transitions at the boundaries between the swan and the water are seamless. In the sec-" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 315, + 232, + 553, + 303 + ], + "blocks": [ + { + "bbox": [ + 313, + 158, + 554, + 224 + ], + "lines": [ + { + "bbox": [ + 313, + 158, + 554, + 224 + ], + "spans": [ + { + "bbox": [ + 313, + 158, + 554, + 224 + ], + "type": "text", + "content": "Table 4. Ablations on the input video embedding and learning the MRHE with or without hypernet. 'Num' represents the number of videos used for training. Using the MAE encoder generated embedding is better than a learnable embedding jointly optimized with the hypernet. The video-specific MRHE encodings predicted by the hypernet outperform a universal MRHE encoding." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 232, + 553, + 303 + ], + "lines": [ + { + "bbox": [ + 315, + 232, + 553, + 303 + ], + "spans": [ + { + "bbox": [ + 315, + 232, + 553, + 303 + ], + "type": "table", + "html": "
NumEmbed w MAEMRHE w hyperhikecar-turnblackswan
PSNRSSIMPSNRSSIMPSNRSSIM
1X✓30.170.86734.040.90529.950.822
✓X30.230.85833.470.88128.830.797
✓✓30.060.84634.940.90929.530.810
3X✓26.370.76029.980.82926.070.684
✓X26.880.71230.010.78826.640.707
✓✓27.310.72831.870.85027.670.754
", + "image_path": "1087a4022c279b99fa3e090b4c46d635baedad32a872bcfdae97b867d00948df.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 327, + 554, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 554, + 388 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 554, + 388 + ], + "type": "text", + "content": "ond video (\"kite-walk\"), we alter the beach background to a sunset style. In the \"hike\" video, we change the color of the person's shirt to yellow and add text on the backpack. Additional results of decomposition and editing are available in the supplementary material." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 399, + 406, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 406, + 412 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 406, + 412 + ], + "type": "text", + "content": "4.4. Ablation study" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 418, + 554, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 554, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 554, + 467 + ], + "type": "text", + "content": "We do ablations on using different embedding as the input of the model, on learning the multi-resolution hash encoding with and without hypernet, and on using different types of masks as the initial supervision." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 474, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 555, + 713 + ], + "type": "text", + "content": "VideoMAE embedding v.s. Learnable embedding. A well-designed embedding is crucial for a hypernet to generate effective model parameters. Some related works [4, 22] employ a learnable embedding optimized jointly with the hypernet, while we directly grab the embedding from the pre-trained MAE encoder and freeze it during hypernet training. We compare the results of using an optimized embedding versus an embedding generated by the MAE encoder, as shown in the first and third rows of Table 4. The advantage in PSNR for the VideoMAE embedding when training for three videos is more than " + }, + { + "bbox": [ + 313, + 474, + 555, + 713 + ], + "type": "inline_equation", + "content": "1\\mathrm{dB}" + }, + { + "bbox": [ + 313, + 474, + 555, + 713 + ], + "type": "text", + "content": " for the 3-video model. For single-video training, the impact of using either embedding type on the final video quality is minimal. Thus, the advantage of using the VideoMAE Encoder becomes more evident when applied to multiple videos. VideoMAE embedding also shows superiority in unseen videos when using the 3-video metamodel as initialization (see the supplementary materials). Furthermore, using VideoMAE enables robust generalization to new, unseen videos, which is essential for the versatility of our model." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22938" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 71, + 553, + 502 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 553, + 502 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 553, + 502 + ], + "type": "image", + "image_path": "f2161798438d153bda39dd8b751a232b936b76521c12b28f8ab623381dc77e8b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 509, + 555, + 544 + ], + "lines": [ + { + "bbox": [ + 55, + 509, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 555, + 544 + ], + "type": "text", + "content": "Figure 6. Qualitative results on the DAVIS dataset. Displayed are the a) original frames, b) predicted masks, c) residual maps, d) texture maps, e) reconstructed frames, and PSNR values for different videos. A color checkerboard overlay is applied to the masked areas to visualize texture transformations." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 556, + 297, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 556, + 297, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 297, + 712 + ], + "type": "text", + "content": "Video-based MRHE learned with hypernet v.s. Universal MRHE without hypernet. In this part, we show the benefits of learning multiresolution hash encoding with the hypernet. In multiresolution hash encoding, input 2D features are treated as coordinates within multi-resolution grids and used to sample data from these grids via interpolation. The sampled data are then concatenated and processed through MLPs to generate the final output. In Table 4, we compare the results of i) using hypernet to learn the parameters of the MLPs making these parameters unique for each video (third rows) and ii) directly using the main NVD net to learn MRHE parameters making them common for all videos (second rows). The higher quality results are ob" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 556, + 556, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 556, + 556, + 631 + ], + "spans": [ + { + "bbox": [ + 313, + 556, + 556, + 631 + ], + "type": "text", + "content": "tained when MRHE is learned by the hypernet, with benefits that become more apparent as the number of videos used for training increases. This improvement stems from the fact that different video scenes exhibit varying levels of complexity, making video-specific hash grid parameters beneficial when training across multiple videos." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 641, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 557, + 715 + ], + "type": "text", + "content": "Accurate mask v.s. Coarse mask. The reference mask is important in our task since it provides a clue about the moving salient objects. We compare the use of manually annotated precise masks (groundtruth from the DAVIS dataset [27]) with coarsely predicted masks from MaskR-CNN [14] as the initial reference for model training. The" + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22939" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 85, + 150, + 139 + ], + "blocks": [ + { + "bbox": [ + 57, + 73, + 151, + 83 + ], + "lines": [ + { + "bbox": [ + 57, + 73, + 151, + 83 + ], + "spans": [ + { + "bbox": [ + 57, + 73, + 151, + 83 + ], + "type": "text", + "content": "a) Reconstructed frame" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 57, + 85, + 150, + 139 + ], + "lines": [ + { + "bbox": [ + 57, + 85, + 150, + 139 + ], + "spans": [ + { + "bbox": [ + 57, + 85, + 150, + 139 + ], + "type": "image", + "image_path": "860a806acd40860df473310c3257e16465ad87b66a4fffea4eb070e29a3d12ad.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 155, + 85, + 302, + 139 + ], + "blocks": [ + { + "bbox": [ + 186, + 73, + 272, + 83 + ], + "lines": [ + { + "bbox": [ + 186, + 73, + 272, + 83 + ], + "spans": [ + { + "bbox": [ + 186, + 73, + 272, + 83 + ], + "type": "text", + "content": "b) Output 2D textures" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 155, + 85, + 302, + 139 + ], + "lines": [ + { + "bbox": [ + 155, + 85, + 302, + 139 + ], + "spans": [ + { + "bbox": [ + 155, + 85, + 302, + 139 + ], + "type": "image", + "image_path": "96c32d2d98d118896a81eaeff01f0fe87deeeafa20541f7f8c43576c65f373ba.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 309, + 85, + 456, + 139 + ], + "blocks": [ + { + "bbox": [ + 345, + 73, + 426, + 83 + ], + "lines": [ + { + "bbox": [ + 345, + 73, + 426, + 83 + ], + "spans": [ + { + "bbox": [ + 345, + 73, + 426, + 83 + ], + "type": "text", + "content": "c) Edited 2D textures" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 309, + 85, + 456, + 139 + ], + "lines": [ + { + "bbox": [ + 309, + 85, + 456, + 139 + ], + "spans": [ + { + "bbox": [ + 309, + 85, + 456, + 139 + ], + "type": "image", + "image_path": "f226423c8c65e8f647b0602b88091db8e6d2d70bf77d8db774ccc80cf8c29072.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 461, + 85, + 553, + 139 + ], + "blocks": [ + { + "bbox": [ + 477, + 73, + 537, + 83 + ], + "lines": [ + { + "bbox": [ + 477, + 73, + 537, + 83 + ], + "spans": [ + { + "bbox": [ + 477, + 73, + 537, + 83 + ], + "type": "text", + "content": "d) Edited frame" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 461, + 85, + 553, + 139 + ], + "lines": [ + { + "bbox": [ + 461, + 85, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 461, + 85, + 553, + 139 + ], + "type": "image", + "image_path": "e6cf06e38c69535d9e3a050707bd1fe63f478eb4c5402c210ea4919dd6d0b96e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 57, + 144, + 150, + 198 + ], + "blocks": [ + { + "bbox": [ + 57, + 144, + 150, + 198 + ], + "lines": [ + { + "bbox": [ + 57, + 144, + 150, + 198 + ], + "spans": [ + { + "bbox": [ + 57, + 144, + 150, + 198 + ], + "type": "image", + "image_path": "ad892634bc5b026b4e79f8f55d45675ec4be5ca506ae599df7cc4e65c63ae4a2.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 154, + 144, + 302, + 198 + ], + "blocks": [ + { + "bbox": [ + 154, + 144, + 302, + 198 + ], + "lines": [ + { + "bbox": [ + 154, + 144, + 302, + 198 + ], + "spans": [ + { + "bbox": [ + 154, + 144, + 302, + 198 + ], + "type": "image", + "image_path": "1501ab164303ef0bc4a7121ab6c73f4f875b23db2e2ed336240546ba27b560cf.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 308, + 144, + 456, + 198 + ], + "blocks": [ + { + "bbox": [ + 308, + 144, + 456, + 198 + ], + "lines": [ + { + "bbox": [ + 308, + 144, + 456, + 198 + ], + "spans": [ + { + "bbox": [ + 308, + 144, + 456, + 198 + ], + "type": "image", + "image_path": "abc676660e3007ec115fcd879825f60c6dc8e64e2f7487e2f285e47656cbcc15.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 460, + 144, + 553, + 198 + ], + "blocks": [ + { + "bbox": [ + 460, + 144, + 553, + 198 + ], + "lines": [ + { + "bbox": [ + 460, + 144, + 553, + 198 + ], + "spans": [ + { + "bbox": [ + 460, + 144, + 553, + 198 + ], + "type": "image", + "image_path": "878aac63c72d77348d8ac237a45a7e2dfb86ec1b9400d45885d879835aed54b0.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 57, + 206, + 150, + 260 + ], + "blocks": [ + { + "bbox": [ + 57, + 206, + 150, + 260 + ], + "lines": [ + { + "bbox": [ + 57, + 206, + 150, + 260 + ], + "spans": [ + { + "bbox": [ + 57, + 206, + 150, + 260 + ], + "type": "image", + "image_path": "04ccf7e33a0e0ff0f9cd9a04f649aab1798f2d7de7e3b255f661ec4975c7bef1.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 269, + 555, + 301 + ], + "lines": [ + { + "bbox": [ + 55, + 269, + 555, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 269, + 555, + 301 + ], + "type": "text", + "content": "Figure 7. Edited video results. We make adjustments to one of the texture layers—the background layers (rows 1 and 2), and the foreground layer (row 3)—while keeping the other layer unchanged and blending the edited layers together. The edited video maintains good visual quality and smooth transitions." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 154, + 206, + 302, + 260 + ], + "blocks": [ + { + "bbox": [ + 154, + 206, + 302, + 260 + ], + "lines": [ + { + "bbox": [ + 154, + 206, + 302, + 260 + ], + "spans": [ + { + "bbox": [ + 154, + 206, + 302, + 260 + ], + "type": "image", + "image_path": "28559146cd6cbe1b15ecdbbf07c1cc5694e0a22257445cddb9a6058495e10658.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 308, + 206, + 456, + 259 + ], + "blocks": [ + { + "bbox": [ + 308, + 206, + 456, + 259 + ], + "lines": [ + { + "bbox": [ + 308, + 206, + 456, + 259 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 456, + 259 + ], + "type": "image", + "image_path": "e1934625b7583635e28f6f2030aac9c6d825527e998e63df489581dd43d0f036.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 460, + 206, + 553, + 259 + ], + "blocks": [ + { + "bbox": [ + 460, + 206, + 553, + 259 + ], + "lines": [ + { + "bbox": [ + 460, + 206, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 460, + 206, + 553, + 259 + ], + "type": "image", + "image_path": "b74eb4709b106020eeac7faf5dc599523313455fb3119c7eb29d7d745f0dd89a.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 60, + 333, + 288, + 383 + ], + "blocks": [ + { + "bbox": [ + 62, + 316, + 288, + 327 + ], + "lines": [ + { + "bbox": [ + 62, + 316, + 288, + 327 + ], + "spans": [ + { + "bbox": [ + 62, + 316, + 288, + 327 + ], + "type": "text", + "content": "Table 5. Ablation on using different masks as initial guidance." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 333, + 288, + 383 + ], + "lines": [ + { + "bbox": [ + 60, + 333, + 288, + 383 + ], + "spans": [ + { + "bbox": [ + 60, + 333, + 288, + 383 + ], + "type": "table", + "html": "
Maskhikecar-turnblackswan
PSNRSSIMPSNRSSIMPSNRSSIM
GT30.060.84634.940.90929.530.810
MaskRCNN29.610.84634.920.90329.660.814
", + "image_path": "c0c8a48fe53da6667216841c0a58783aa3616ffabf22c2205f85579f37df23f0.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 404, + 296, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 404, + 296, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 404, + 296, + 548 + ], + "type": "text", + "content": "predicted alpha map is a soft mask that always includes the corresponding effects, such as reflection or shadows accompanied by the moving objects. Therefore, it is always different from the original object mask. Results in Table 5 show that our model is robust to masks with varying levels of accuracy. Figure 8 illustrates two frames from the \"bear\" and \"blackswan\" videos, where a more precise initial mask improves object boundary details. For example, using a more refined mask as initial guidance help the alpha map to capture intricate structures, such as the feathers on a swan's tail. Thus, for higher-quality editing results, it is essential to use the most detailed masks possible during model training." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 560, + 179, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 560, + 179, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 560, + 179, + 574 + ], + "type": "text", + "content": "5. Concluding Remarks" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": "We present a framework for multiple video training of INR-based video decomposition models using a hypernetwork. We reformulate the task of training a neural video decomposition model on a single video into learning a hypernetwork that generates network parameters based on video embeddings. This approach effectively mitigates the overfitting of the neural video decomposition model to individual videos, enhances generalization capability, and speeds up adaptation to unseen videos. Through experiments, we validate that our proposed model achieves comparable video layer decomposition and rendering capabilities to previous" + } + ] + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 317, + 317, + 552, + 504 + ], + "blocks": [ + { + "bbox": [ + 317, + 317, + 552, + 504 + ], + "lines": [ + { + "bbox": [ + 317, + 317, + 552, + 504 + ], + "spans": [ + { + "bbox": [ + 317, + 317, + 552, + 504 + ], + "type": "image", + "image_path": "ec1e1a9b5617ad96e148be95a14de72f083029de09603a316fac223bb788c21f.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 513, + 555, + 568 + ], + "lines": [ + { + "bbox": [ + 313, + 513, + 555, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 513, + 555, + 568 + ], + "type": "text", + "content": "Figure 8. Comparison between using accurate and coarse masks as initial guidance. Using different masks does not significantly affect the reconstruction results, but providing a refined mask as the initial guidance helps the model achieve more detailed decomposition, which can be especially beneficial for finer edits." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 584, + 553, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 584, + 553, + 608 + ], + "spans": [ + { + "bbox": [ + 313, + 584, + 553, + 608 + ], + "type": "text", + "content": "methods, while significantly reducing training time on new videos and improving video editing efficiency." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 616, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 616, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 616, + 555, + 712 + ], + "type": "text", + "content": "Acknowledgement. We acknowledge the Departament de Recerca i Universitats with ref. 2021SGR01499 and the CERCA Program from Generalitat de Catalunya, Grant PID2021-128178OB-I00 funded by MCIN/AEI/10.13039/501100011033, ERDF \"A way of making Europe\", and the grant Catedra ENIA UAB-Cruilla (TSI-100929-2023-2) from the Ministry of Economic Affairs and Digital Transformation of Spain." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22940" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 61, + 91, + 249, + 101 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 249, + 101 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 249, + 101 + ], + "type": "text", + "content": "[1] Adobe. After effects. Accessed: 2024-11-07. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 103, + 295, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 103, + 295, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 103, + 295, + 135 + ], + "type": "text", + "content": "[2] Ali Athar, Alexander Hermans, Jonathon Luiten, Deva Ramanan, and Bastian Leibe. Tarvis: A unified approach for target-based video segmentation. In CVPR, 2023. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 137, + 294, + 157 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 137, + 294, + 157 + ], + "spans": [ + { + "bbox": [ + 62, + 137, + 294, + 157 + ], + "type": "text", + "content": "[3] BlackMagicDesign. Davinci resolve 19. Accessed: 2024-11-07. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 160, + 294, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 160, + 294, + 203 + ], + "spans": [ + { + "bbox": [ + 62, + 160, + 294, + 203 + ], + "type": "text", + "content": "[4] Cheng-Hung Chan, Cheng-Yang Yuan, Cheng Sun, and Hwann-Tzong Chen. Hashing neural video decomposition with multiplicative residuals in space-time. In ICCV, 2023. 1, 2, 4, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 205, + 294, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 205, + 294, + 247 + ], + "spans": [ + { + "bbox": [ + 62, + 205, + 294, + 247 + ], + "type": "text", + "content": "[5] Vinod Kumar Chauhan, Jiandong Zhou, Ping Lu, Soheila Molaei, and David A Clifton. A brief review of hypernetworks in deep learning. Artificial Intelligence Review, 57(9): 250, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 250, + 294, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 250, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 62, + 250, + 294, + 282 + ], + "type": "text", + "content": "[6] Pei-Ze Chiang, Meng-Shiun Tsai, Hung-Yu Tseng, Wei-Sheng Lai, and Wei-Chen Chiu. Stylizing 3d scene via implicit representation and hypernetwork. In WACV, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 285, + 294, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 285, + 294, + 337 + ], + "spans": [ + { + "bbox": [ + 62, + 285, + 294, + 337 + ], + "type": "text", + "content": "[7] Yuren Cong, Mengmeng Xu, Christian Simon, Shoufa Chen, Jiawei Ren, Yanping Xie, Juan-Manuel Perez-Rua, Bodo Rosenhahn, Tao Xiang, and Sen He. Flatten: optical flow-guided attention for consistent text-to-video editing. In ICLR, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 340, + 294, + 383 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 340, + 294, + 383 + ], + "spans": [ + { + "bbox": [ + 62, + 340, + 294, + 383 + ], + "type": "text", + "content": "[8] Carl Doersch, Ankush Gupta, Larisa Markeeva, Adria Recasens, Lucas Smaira, Yusuf Aytar, Joao Carreira, Andrew Zisserman, and Yi Yang. Tap-vid: A benchmark for tracking any point in a video. NeurIPS, 2022. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 385, + 294, + 416 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 385, + 294, + 416 + ], + "spans": [ + { + "bbox": [ + 62, + 385, + 294, + 416 + ], + "type": "text", + "content": "[9] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In ICML, 2017. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 418, + 228, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 418, + 228, + 429 + ], + "spans": [ + { + "bbox": [ + 57, + 418, + 228, + 429 + ], + "type": "text", + "content": "[10] Foundry. Nuke. Accessed: 2024-11-07. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 430, + 294, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 430, + 294, + 462 + ], + "spans": [ + { + "bbox": [ + 57, + 430, + 294, + 462 + ], + "type": "text", + "content": "[11] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. In ICLR, 2024. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 464, + 294, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 464, + 294, + 496 + ], + "spans": [ + { + "bbox": [ + 57, + 464, + 294, + 496 + ], + "type": "text", + "content": "[12] Zeqi Gu, Wenqi Xian, Noah Snavely, and Abe Davis. Factormatte: Redefining video matting for re-composition tasks. ACM TOG, 42(4):1-14, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 498, + 294, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 498, + 294, + 518 + ], + "spans": [ + { + "bbox": [ + 57, + 498, + 294, + 518 + ], + "type": "text", + "content": "[13] David Ha, Andrew Dai, and Quoc V Le. Hypernetworks. In ICLR, 2017. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 521, + 294, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 521, + 294, + 541 + ], + "spans": [ + { + "bbox": [ + 57, + 521, + 294, + 541 + ], + "type": "text", + "content": "[14] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, 2017. 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 544, + 294, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 544, + 294, + 565 + ], + "spans": [ + { + "bbox": [ + 57, + 544, + 294, + 565 + ], + "type": "text", + "content": "[15] Nebojsa Jojic and Brendan J Frey. Learning flexible sprites in video layers. In CVPR, 2001. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 567, + 294, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 567, + 294, + 599 + ], + "spans": [ + { + "bbox": [ + 57, + 567, + 294, + 599 + ], + "type": "text", + "content": "[16] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together. In ECCV, 2024. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 57, + 601, + 294, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 601, + 294, + 632 + ], + "spans": [ + { + "bbox": [ + 57, + 601, + 294, + 632 + ], + "type": "text", + "content": "[17] Yoni Kasten, Dolev Ofri, Oliver Wang, and Tali Dekel. Layered neural atlases for consistent video editing. ACM TOG, 40(6):1-12, 2021. 2, 4, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 57, + 635, + 294, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 294, + 655 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 294, + 655 + ], + "type": "text", + "content": "[18] KeenTools. Geotracker for adobe after effects. Accessed: 2024-11-07. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 57, + 658, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 294, + 689 + ], + "type": "text", + "content": "[19] Chiheon Kim, Doyup Lee, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Generalizable implicit neural representations via instance pattern composers. In CVPR, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "type": "text", + "content": "[20] Jaekyeong Lee, Geonung Kim, and Sunghyun Cho. Rna: Video editing with roi-based neural atlas. ACCV, 2024. 1" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 39, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[21] Erika Lu, Forrester Cole, Tali Dekel, Andrew Zisserman, William T Freeman, and Michael Rubinstein. Omnimatte: Associating objects and their effects in video. In CVPR, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 119, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 553, + 161 + ], + "type": "text", + "content": "[22] Shishira R Maiya, Anubhav Gupta, Matthew Gwilliam, Max Ehrlich, and Abhinav Shrivastava. Latent-inr: A flexible framework for implicit representations of videos with discriminative semantics. In ECCV, 2024. 2, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 164, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 164, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 317, + 164, + 553, + 196 + ], + "type": "text", + "content": "[23] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM TOG, 41(4):1-15, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 199, + 553, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 199, + 553, + 231 + ], + "spans": [ + { + "bbox": [ + 317, + 199, + 553, + 231 + ], + "type": "text", + "content": "[24] Seonghyeon Nam, Marcus A Brubaker, and Michael S Brown. Neural image representations for multi-image fusion and layer separation. In ECCV, 2022. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 234, + 553, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 234, + 553, + 265 + ], + "spans": [ + { + "bbox": [ + 317, + 234, + 553, + 265 + ], + "type": "text", + "content": "[25] Alex Nichol, Joshua Achiam, and John Schulman. On first-order meta-learning algorithms. arXiv preprint arXiv:1803.02999, 2018. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 268, + 553, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 268, + 553, + 311 + ], + "spans": [ + { + "bbox": [ + 317, + 268, + 553, + 311 + ], + "type": "text", + "content": "[26] Hao Ouyang, Qiuyu Wang, Yuxi Xiao, Qingyan Bai, Juntao Zhang, Kecheng Zheng, Xiaowei Zhou, Qifeng Chen, and Yujun Shen. Codef: Content deformation fields for temporally consistent video processing. In CVPR, 2024. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 314, + 553, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 314, + 553, + 357 + ], + "spans": [ + { + "bbox": [ + 317, + 314, + 553, + 357 + ], + "type": "text", + "content": "[27] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In CVPR, 2016. 4, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 359, + 553, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 359, + 553, + 391 + ], + "spans": [ + { + "bbox": [ + 317, + 359, + 553, + 391 + ], + "type": "text", + "content": "[28] Alex Rav-Acha, Pushmeet Kohli, Carsten Rother, and Andrew Fitzgibbon. Unwrap mosaics: A new representation for video editing. In SIGGRAPH, 2008. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 394, + 553, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 394, + 553, + 436 + ], + "spans": [ + { + "bbox": [ + 317, + 394, + 553, + 436 + ], + "type": "text", + "content": "[29] Bipasha Sen, Gaurav Singh, Aditya Agarwal, Rohith Agaram, Madhava Krishna, and Srinath Sridhar. Hypnerf: Learning improved nerf priors using a hypernetwork. NeurIPS, 2024. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 439, + 553, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 439, + 553, + 481 + ], + "spans": [ + { + "bbox": [ + 317, + 439, + 553, + 481 + ], + "type": "text", + "content": "[30] Soumyadip Sengupta, Vivek Jayaram, Brian Curless, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Background matting: The world is your green screen. In CVPR, 2020. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 317, + 484, + 553, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 484, + 553, + 517 + ], + "spans": [ + { + "bbox": [ + 317, + 484, + 553, + 517 + ], + "type": "text", + "content": "[31] Dmitriy Smirnov, Michael Gharbi, Matthew Fisher, Vitor Guizilini, Alexei Efros, and Justin M Solomon. Marionette: Self-supervised sprite learning. NeurIPS, 2021. 1" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 317, + 519, + 553, + 561 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 519, + 553, + 561 + ], + "spans": [ + { + "bbox": [ + 317, + 519, + 553, + 561 + ], + "type": "text", + "content": "[32] Shaolin Su, Qingsen Yan, Yu Zhu, Cheng Zhang, Xin Ge, Jinqiu Sun, and Yanning Zhang. Blindly assess image quality in the wild guided by a self-adaptive hyper network. In CVPR, 2020. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 317, + 564, + 553, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 564, + 553, + 586 + ], + "spans": [ + { + "bbox": [ + 317, + 564, + 553, + 586 + ], + "type": "text", + "content": "[33] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In ECCV, 2020. 4" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 317, + 588, + 553, + 630 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 588, + 553, + 630 + ], + "spans": [ + { + "bbox": [ + 317, + 588, + 553, + 630 + ], + "type": "text", + "content": "[34] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. VideoMAE: Masked autoencoders are data-efficient learners for self-supervised video pre-training. In NeurIPS, 2022. 1, 2, 3" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 317, + 633, + 553, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 633, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 317, + 633, + 553, + 666 + ], + "type": "text", + "content": "[35] Johannes von Oswald, Christian Henning, Benjamin F. Grewe, and João Sacramento. Continual learning with hypernetworks. In ICLR, 2020. 2" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 316, + 668, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 668, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 668, + 553, + 689 + ], + "type": "text", + "content": "[36] John YA Wang and Edward H Adelson. Representing moving images with layers. IEEE TIP, 3(5):625-638, 1994. 2" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[37] Ning Xu, Brian Price, Scott Cohen, and Thomas Huang. Deep image matting. In CVPR, 2017. 2" + } + ] + } + ], + "index": 38 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22941" + } + ] + } + ], + "index": 40 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 150 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 106 + ], + "type": "text", + "content": "[38] Vickie Ye, Zhengqi Li, Richard Tucker, Angjoo Kanazawa, and Noah Snively. Deformable sprites for unsupervised video decomposition. In CVPR, 2022. 1, 2, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "type": "text", + "content": "[39] Emilie Yu, Kevin Blackburn-Matzen, Cuong Nguyen, Oliver Wang, Rubaiat Habib Kazi, and Adrien Bousseau. Videodoodles: Hand-drawn animations on videos with scene-aware canvases. ACM TOG, 42(4):1-12, 2023. 1" + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22942" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/HyperNet Fields_ Efficiently Training Hypernetworks without Ground Truth by Learning Weight Trajectories/68fd503f-6b57-47fa-8aca-593407ee13b9_content_list.json b/2025/HyperNet Fields_ Efficiently Training Hypernetworks without Ground Truth by Learning Weight Trajectories/68fd503f-6b57-47fa-8aca-593407ee13b9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..88212927da16b902d384782f740fb38ed3800f07 --- /dev/null +++ b/2025/HyperNet Fields_ Efficiently Training Hypernetworks without Ground Truth by Learning Weight Trajectories/68fd503f-6b57-47fa-8aca-593407ee13b9_content_list.json @@ -0,0 +1,1748 @@ +[ + { + "type": "text", + "text": "HyperNet Fields: Efficiently Training Hypernetworks without Ground Truth by Learning Weight Trajectories", + "text_level": 1, + "bbox": [ + 186, + 130, + 816, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Eric Hedlin $^{1,2*}$ Munawar Hayat $^{2}$ Fatih Porikli $^{2}$ Kwang Moo Yi $^{1}$ Shweta Mahajan $^{2}$ $^{1}$ University of British Columbia ${}^{2}$ Qualcomm AI Research†", + "bbox": [ + 148, + 200, + 849, + 239 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c4f22adc05793672a7e7332c344006a515b952740da570723acf1df772c116ed.jpg", + "image_caption": [ + "Figure 1. Teaser - We propose hypernetwork fields, where we learn the entire weight space trajectory instead of only the final state, which allows us to train without ever needing to know the final converged weights. Our method can be applied to any application of hypernetworks, including diffusion model personalization and modeling of 3D neural representations." + ], + "image_footnote": [], + "bbox": [ + 94, + 273, + 500, + 516 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/91f43ea1bd60ca683d11dc4a81ff813e1a575285286666305481408c8df0b3ec.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 273, + 901, + 513 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 580, + 326, + 597 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To efficiently adapt large models or to train generative models of neural representations, Hypernetworks have drawn interest. While hypernetworks work well, training them is cumbersome, and often requires ground truth optimized weights for each sample. However, obtaining each of these weights is a training problem of its own—one needs to train, e.g., adaptation weights or even an entire neural field for hypernetworks to regress to. In this work, we propose a method to train hypernetworks, without the need for any per-sample ground truth. Our key idea is to learn a Hypernetwork 'Field' and estimate the entire trajectory of network weight training instead of simply its converged state. In other words, we introduce an additional input to the Hypernetwork, the convergence state, which then makes it act as a neural field that models the entire convergence pathway of a task network. A critical benefit in doing so is that the gradient of the estimated weights at any convergence state must then match the gradients of the original task—this constraint alone is sufficient to train the Hypernetwork Field.", + "bbox": [ + 88, + 613, + 485, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We demonstrate the effectiveness of our method through the task of personalized image generation and 3D shape reconstruction from images and point clouds, demonstrating competitive results without any per-sample ground truth.", + "bbox": [ + 511, + 582, + 908, + 643 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 688, + 645, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Various generative modeling tasks [39] require adaptation, thus data-dependent network weights, or in other words, model parameters being adjusted according to various circumstances (conditions and user input). These include personalized text-to-image generation where the weights of the underlying personalization model, such as DreamBooth [39], are fine-tuned on the concept of interest and used to generate target images conditioned on input prompts. Another example is 3D modeling, where neural", + "bbox": [ + 509, + 714, + 906, + 852 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work done as part of a summer internship.", + "bbox": [ + 529, + 863, + 763, + 875 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ Qualcomm AI Research is an initiative of Qualcomm Technologies, Inc.", + "bbox": [ + 514, + 876, + 903, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22129", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "networks are used to model shapes implicitly with signed distance fields [29] or occupancy [27]. In both cases, a dedicated set of neural network weights is required for each personalization outcome or 3D shape. Thus, the applicability of these methods is highly limited by the training time required to obtain these weights.", + "bbox": [ + 88, + 90, + 480, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "It is therefore critical to seek ways to amortize the per-sample optimizations into Hypernetworks [12, 40]—networks that estimate network parameters. By 'skipping ahead' the tedious training processes involved in fine-tuning or shape regressing, they greatly speed up the entire process. Existing methods, however, have a critical shortcoming in that they supervise hypernetworks by directly matching their output to the precomputed, ground truth converged weights corresponding to the task-specific network for each sample in the dataset. Notably, prior to any training of the Hypernetwork, one must prepare converged task-specific network weights for every sample in the training set. This severely limits the capabilities of Hypernetwork methods when scaling to large datasets. As an example, to train HyperDreamBooth [40], it requires 50 days of GPU time on a NVidia RTX 3090, $^{\\ddagger}$ to just prepare the training data alone.", + "bbox": [ + 91, + 183, + 482, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Furthermore, current methods assume a one-to-one mapping between each input sample and its corresponding optimized weights. This assumption is an oversimplification. Neural network optimization is stochastic with various equally plausible solutions. Therefore, enforcing a bijective mapping between the condition and the ultimately converged weights may limit the potential expressiveness of the hypernetwork, as we will demonstrate in Sec. 3.", + "bbox": [ + 89, + 424, + 480, + 545 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these limitations, we instead propose to train a Hypernetwork Field that estimates the entire optimization trajectory of a neural network training, rather than focusing on a single final set of weights for each sample. Specifically, we introduce the convergence state as additional input to the hypernetwork, such that the weights of the task-specific network at a desired state of convergence can be 'queried'. An interesting outcome of doing so is that the gradient of the estimated weights along this pathway must match the gradients of the original task-specific network, as they are both gradients pointing towards convergence. Hence, training of such hypernetwork fields can simply be done by matching the two gradients.", + "bbox": [ + 89, + 546, + 482, + 742 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "By supervising the hypernetwork to match the gradient of the weights with respect to the optimization step, we avoid the need for precomputing target weights. This supervision strategy allows the hypernetwork to estimate a trajectory of parameters that, at each step, reflects a compatible state across all samples in the dataset. Ultimately, at inference time, the estimated parameters which correspond with the final timestep of the hypernetwork, represent a well", + "bbox": [ + 89, + 743, + 482, + 863 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "converged solution for each of the samples, but are now naturally discovered through training of the hypernetwork field itself. Notably, this single forward pass to estimating the converged parameters is all that is necessary for inference, hence no additional compute is required compared to conventional hypernetwork approaches [12, 40].", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our framework is general and is applicable to all hypernetwork training scenarios. To demonstrate the effectiveness of our method, we apply our method to personalized image generation and 3D shape generation. We achieve competitive performance while reducing the training cost by $\\sim 4$ times what would be required of conventional hypernetwork training.", + "bbox": [ + 511, + 183, + 903, + 289 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, our contributions are:", + "bbox": [ + 532, + 291, + 777, + 304 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose hypernetwork fields, an approach to learn the entire optimization pathway of task-specific networks, which allows ground truth-free training without the need for precomputing the target weights;", + "- our approach drastically reduces the compute required for training hypernetworks;", + "- we validate our approach on two very different tasks and datasets—for personalized image generation we achieve results comparable to the state of the art with significantly less compute; and", + "- we further explore the application of our method to 3D shape reconstruction from images an point clouds." + ], + "bbox": [ + 511, + 306, + 903, + 488 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 507, + 653, + 523 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Hypernetworks. Early work on training hypernetworks backpropagates gradients through the task-specific network to the hypernetwork itself, as demonstrated by approaches [16, 31] and subsequent explorations in continual learning [28] and neural architecture search [49]. These methods have shown that allowing gradient flow through both networks enables the hypernetwork to modify parameters for diverse tasks adaptively. However, more recent works have introduced approaches that involve precomputing weights for the task-specific network and directly supervise the hypernetwork to match them, eliminating the need for backpropagation through the task-specific network and improving training stability by alleviating issues with gradient flow. This supervision with precomputed weights has been shown to enhance gradient stability and training performance, as highlighted in frameworks like HyperDiffusion [12], HyperFields [2], and HyperDreamBooth [40]. However, this limits the size of the dataset to which this can be applied, as the precompute requirement is proportional to the number of samples in the dataset. Our method combines the best of both approaches by removing the need for precomputed weights while explicitly providing weight-based supervision in the form of gradients along the convergence path, enabling stable training on large datasets.", + "bbox": [ + 511, + 537, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "$\\ddagger$ HyperDreamBooth [40] was trained on 15k DreamBooth weights, which each take 5 minutes to compute.", + "bbox": [ + 89, + 875, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22130", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Denoising Models. In generative modeling, finding smooth and reliable paths through latent space has become central to producing coherent samples. Examples include diffusion models [17, 43], which iteratively trace a path from noise to data through successive steps, effectively navigating high-dimensional space by reversing a stochastic process. Consistency models [46] refine this by learning non-Markovian trajectories that reach target data distributions in fewer steps by introducing a formulation that allows them to single-step jump to any timestep. Flow-matching models [23] take a similar path-based approach by learning a vector field that guides samples along pre-defined flows, ensuring a direct and smooth transition through the latent space. Similarly to these methods, our method can be thought of as finding a path through this high-dimensional space, specifically in the form of a single-step estimator from initialization to any desired state of convergence, whose path travels strictly along the direction of the steepest gradient descent.", + "bbox": [ + 89, + 90, + 483, + 363 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Methods with gradient-supervision. Several recent approaches in generative modeling and physics-informed learning employ supervision exclusively through gradients rather than direct supervision between input-output pairs to achieve denoising and reconstruction. Score-based generative models [44, 45, 47] use score matching to learn the gradient (or \"score\") of the log density of noisy data distributions, bypassing the need for direct output supervision. Denoising Score Matching (DSM) [47] trains the model to predict this gradient, guiding it to approximate data distributions through iterative denoising and demonstrating improved stability in noise removal. Similarly, Physics-Informed Neural Networks (PINNs) [35-37] leverage known derivatives from partial differential equations (PDEs) in physical simulations, such as heat diffusion or fluid flow, to inform the network solely through spatial or temporal derivatives. This derivative-based guidance encourages convergence as the model adapts to match observed physical behaviour, requiring no explicit target outputs. Energy-Based Models (EBMs) [11, 14, 22] also align with this gradient-driven approach by learning energy landscapes that adjust gradients to represent target distributions, allowing for smooth gradient matching through learned energy fields. Inspired by these methods, our approach supervises only the gradient direction in task-specific weight space, effectively learning a convergence pathway without requiring explicit final state supervision. By supervising gradients within this pathway, our method achieves robust training stability while generalizing over diverse tasks without per-sample ground truth weights.", + "bbox": [ + 91, + 368, + 483, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 842, + 181, + 858 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To remove the precompute requirements associated with training hypernetworks, we present a new framework that", + "bbox": [ + 89, + 869, + 482, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "models the convergence dynamics of task-specific networks into a single, efficient process, which we name Hypernetwork Field.", + "bbox": [ + 511, + 90, + 903, + 136 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Problem formulation. Formally, let $f_{\\theta}$ denote a task-specific neural network, e.g., dreambooth [39], parameterized by $\\theta$ . $f_{\\theta}$ is then typically trained to minimize a loss function $\\mathbb{E}_{\\mathbf{x}}[\\mathcal{L}(\\theta ,\\mathbf{x})]$ , by training for $T$ optimization steps given a dataset of conditions, e.g., image $\\mathbf{x}$ . In our work, we wish to learn a hypernetwork field $H_{\\phi}(t,\\mathbf{x})$ , parameterized by $\\phi$ that takes as input a condition $\\mathbf{x}$ and an optimization step $t$ . The hypernetwork field then predicts the parameters $\\theta_t = H_\\phi (t,\\mathbf{x})$ at optimization step $t$ of the task-specific network $f_{\\theta_t}(\\mathbf{x})$ , given the condition $\\mathbf{x}$ and the timestep $t$ . In the following, we show how to obtain $H_{\\phi}(t,\\mathbf{x})$ without requiring any $\\theta_t$ to supervise the training.", + "bbox": [ + 511, + 140, + 906, + 321 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. HyperNet Fields", + "text_level": 1, + "bbox": [ + 511, + 330, + 676, + 347 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To train the hypernet field $H_{\\phi}(t,\\mathbf{x})$ without the ground-truth targets for it to regress, we take a step back and revisit what standard hypernetworks are optimizing for. Standard approaches for learning hypernetworks [12, 40] utilize the ground-truth weights $\\theta^{*}$ of the task-specific network, corresponding to the optimized parameters for a specific sample $\\mathbf{x}$ as a supervision signal. In other words, they first find", + "bbox": [ + 511, + 353, + 905, + 459 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\theta^ {*} (\\mathbf {x}) = \\underset {\\theta} {\\arg \\min } \\mathcal {L} (\\theta , \\mathbf {x}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 470, + 906, + 496 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "for each $\\mathbf{x}$ , then find $\\phi$ , such that $\\theta^{*}(\\mathbf{x}) \\approx H_{\\phi}(\\cdot, \\mathbf{x}), \\forall \\mathbf{x}$ . This typically requires several iterations of the optimization process and thus incurs a large computational cost for generating the target parameters $\\theta^{*}(\\mathbf{x})$ . Moreover, since this process has to be repeated for each sample independently, the scalability to large datasets is severely limited.", + "bbox": [ + 511, + 507, + 905, + 598 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Note, however, Eq. (1) is typically found through gradient-based optimization, and forms a trajectory of weight parameter space in the $\\{\\theta_t\\}_{t = 1\\dots T}$ , where $T$ is the number of optimization steps. If we now consider our hypernet field $H_{\\phi}(t,\\mathbf{x})$ that we intend to model, we notice an interesting relationship that must hold between the $\\mathcal{L}(\\theta_t,\\mathbf{x})$ and $H_{\\phi}(t,\\mathbf{x})$ . Specifically, consider the gradient of the task-specific weights at an optimization step $t$ , $\\nabla_{\\theta_t}\\mathcal{L}(\\theta_t,\\mathbf{x})$ . Then, if a hypernet field $H_{\\phi}$ has been properly trained, the direction in which $\\nabla_{\\theta_t}\\mathcal{L}(\\theta_t,\\mathbf{x})$ point to must match the direction in which the hypernet trajectory flows, that is, $\\nabla_tH_\\phi (t,\\mathbf{x})$ . In other words, their gradients must match.", + "bbox": [ + 511, + 599, + 905, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In fact, matching the two gradients allows us to circumvent the necessity of per-sample optimization as $\\nabla_{\\theta_t}\\mathcal{L}(\\theta_t,\\mathbf{x})$ can be computed with relatively minimal overhead by using the current hypernet field estimate for $\\theta_t$ . This process can also be understood as amortizing Eq. (1) into the training process of the hypernet field $H_{\\phi}(t,\\mathbf{x})$ , as we are 'skipping ahead' parts of the optimization loop of Eq. (1) with the estimates coming from $H_{\\phi}(t,\\mathbf{x})$ . Through", + "bbox": [ + 511, + 780, + 908, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22131", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/868d0df2f2a67d6a8b042c2f4131997dd8911b70b965b1e961db9efb123e9367.jpg", + "image_caption": [ + "Figure 2. The Hypenet Fields Framework. Left: Our Hypernetwork with gradient-based supervision from a task-specific network - dreambooth [39]. Right: The sampling process to generate personalized images with our hypernetwork framework." + ], + "image_footnote": [], + "bbox": [ + 91, + 89, + 486, + 200 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5327c2512391025e71e269920022c476cc10f1d2ba8bc520d4a5b78a280787e3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 99, + 906, + 174 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "this amortization, the hypernet field captures detailed information about how the weights evolve throughout training, allowing it to effectively predict and generate weights across the training samples. We now further detail the exact training process.", + "bbox": [ + 88, + 265, + 483, + 340 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Gradient-based Supervision", + "text_level": 1, + "bbox": [ + 89, + 349, + 344, + 366 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In each training iteration of the optimization process, the hypernet field generates an estimate of the task-specific parameters $\\hat{\\theta}_t = H_\\phi(t, \\mathbf{x})$ at timestep $t$ based on the input $\\mathbf{x}$ from a dataset $\\mathcal{D}$ and the timestep $t \\sim U(0, T)$ . Given this estimate, we compute the gradient of the loss function with respect to the task-specific weights at that timestep, $\\nabla_{\\hat{\\theta}_t} \\mathcal{L}(\\hat{\\theta}_t, \\mathbf{x})$ , and perform a single optimization step to update the weights to $\\theta_{t+1}$ . This is given by", + "bbox": [ + 89, + 371, + 483, + 492 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {t + 1} = \\hat {\\theta} _ {t} - \\eta \\nabla_ {\\hat {\\theta} _ {t}} \\mathcal {L} _ {\\text {t a s k}} (\\hat {\\theta} _ {t}, \\mathbf {x}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 500, + 480, + 520 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\eta$ is the learning rate. Then, the hypernet field must also follow this update by definition, thus we wish to have $\\hat{\\theta}_{t + 1}\\approx \\theta_{t + 1}$ . Moreover, assuming $\\theta_0 = \\hat{\\theta}_0 = H_\\phi (0,\\mathbf{x})$ , the entire trajectory can simply be matched via minimizing", + "bbox": [ + 89, + 525, + 483, + 587 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\Delta} (\\mathcal {D}) = \\mathbb {E} _ {\\mathbf {x} \\sim \\mathcal {D}, t \\sim U [ 0, T ]} \\| \\theta_ {t + 1} - H _ {\\phi} (\\mathbf {x}, t + 1) \\| ^ {2}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 594, + 480, + 614 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which we name as the gradient matching loss. This loss function essentially minimizes the difference between the gradient step from Eq. (2) and the numerical gradient $\\nabla_t H_\\phi(\\mathbf{x}, t) = H_\\phi(\\mathbf{x}, t + 1) - H_\\phi(\\mathbf{x}, t)$ over the entire dataset for $T$ steps of optimization; thus replicating the underlying training dynamics of the task-specific network.", + "bbox": [ + 89, + 621, + 483, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Once trained, the hypernetwork field $H_{\\phi}(t,\\mathbf{x})$ thus mimics the entire training trajectory of the task-specific network, instead of the point estimate $\\theta^{*}(\\mathbf{x})$ . We summarize this training process in Fig. 2.", + "bbox": [ + 89, + 710, + 483, + 772 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Ensuring a consistent trajectory throughout training. To ensure training stability, we fix the starting position $\\theta_0$ for each sample $\\mathbf{x}$ throughout training, initializing it randomly only once at the beginning. This approach mirrors the typical neural network initialization to random values, followed by convergence toward the optimal parameters.", + "bbox": [ + 89, + 775, + 483, + 866 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Forcing 0 offset at $t = 0$ by construction. Furthermore, Eq. (3) matches the entire trajectory only when $\\theta_0 = \\hat{\\theta}_0 =$", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 Hypernet Fields Training Algorithm", + "text_level": 1, + "bbox": [ + 514, + 265, + 844, + 280 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: Initialize $\\theta_0$", + "2: $H_{\\phi}\\gets$ Hypernetwork() $\\triangleright$ Hypernetwork with parameters $\\phi$", + "3: while not converged do", + "4: $\\mathbf{x} \\gets \\mathrm{next}(\\mathcal{D}) \\triangleright$ Sample condition $\\mathbf{x}$ from dataset", + "5: $t \\gets$ sample from $[0, T)$", + "6: $\\hat{\\theta}_t\\gets H_\\phi (\\mathbf{x},t)$ $\\triangleright$ Estimate weights for timestep $t$", + "7: $\\Delta \\theta_{t} \\gets -\\eta \\nabla_{\\theta_{t}} \\mathcal{L}_{task}(\\hat{\\theta}_{t}, \\mathbf{x})$ $\\triangleright$ GT gradient for $\\hat{\\theta}_{t}$", + "8: $\\Delta \\hat{\\theta}_t\\gets H_\\phi (\\mathbf{x},t + 1) - \\hat{\\theta}_t$ $\\triangleright$ Est gradient for $\\hat{\\theta}_t$", + "9: $\\phi \\gets \\phi -\\eta \\nabla_{\\phi}\\mathrm{MSE}(\\Delta \\theta_t,\\Delta \\hat{\\theta}_t)$ Update $\\phi$", + "10: end while" + ], + "bbox": [ + 521, + 284, + 903, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$H_{\\phi}(0,\\mathbf{x})$ . Enforcing this, however, can easily be done by construction. Inspired by [21, 46], for any timestep $t$ , we parameterize the model with an offset from the input $\\theta_0$ , conditioned on $t$ as,", + "bbox": [ + 511, + 481, + 906, + 542 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nH _ {\\phi} (\\mathbf {x}, t) = \\theta_ {0} + \\frac {t}{T} \\times H _ {\\phi} ^ {\\prime} (\\mathbf {x}, t), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 547, + 906, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $H'$ is the raw hypernetwork field output from a neural network.", + "bbox": [ + 511, + 580, + 906, + 611 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Algorithm", + "text_level": 1, + "bbox": [ + 513, + 619, + 630, + 636 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We summarize our overall optimization process in Algorithm 1. We first randomly initialize and fix the parameters $\\theta_0$ . We then consider a random sample $\\mathbf{x}$ from the dataset and a random timestep $t$ from the optimization trajectory. With these inputs, we use the hypernetwork $H_{\\phi}$ to obtain the weights $\\hat{\\theta}_t$ of the task-specific network. Next, using these weights and the input $\\mathbf{x}$ we compute the gradient step $\\Delta \\theta_t$ for the task-specific network. Additionally, we calculate $\\Delta \\hat{\\theta}_t = \\hat{\\theta}_{t + 1} - \\hat{\\theta}_t$ from the hypernetwork. The parameters $\\phi$ of the hypernetwork are then updated to match $\\Delta \\hat{\\theta}_t$ with the gradients $\\Delta \\theta_t$ , which is equivalent to minimizing Eq. (3). This yields a computationally efficient framework where a single hypernetwork encodes the optimization trajectory and can be queried to get the weights of the task-specific network at any stage in the training process.", + "bbox": [ + 511, + 642, + 906, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For inference, consider DreamBooth [39] as a task-specific network (Fig. 2). To generate personalized images", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22132", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/bf6635a084243a7e83885b702e461a6136b646190890aab57da0c33a4ff58fb6.jpg", + "image_caption": [ + "Input Image" + ], + "image_footnote": [], + "bbox": [ + 93, + 119, + 292, + 275 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/bc84799f410f4b4d5c0fafc035c30dbe39c926db69a31dff0990bf6b97f3c245.jpg", + "image_caption": [ + "As a drawing" + ], + "image_footnote": [], + "bbox": [ + 305, + 119, + 411, + 200 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3528cc8e9e422dc4f24ac3d126460458a8db90771f530c8ddaad2a1431313b1f.jpg", + "image_caption": [ + "As a pixar character" + ], + "image_footnote": [], + "bbox": [ + 413, + 119, + 521, + 200 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/faa5ad6e1a00a651d2723c148b8928493e5aabc8d7fa72d0fcd1f165ca819c94.jpg", + "image_caption": [ + "As a graffiti mural" + ], + "image_footnote": [], + "bbox": [ + 526, + 119, + 632, + 200 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/42dee81f7839c0a8412a0c19424cfd12df7bd0fa67f1fac509f26c8f52290eaf.jpg", + "image_caption": [ + "Flowing river of lost souls" + ], + "image_footnote": [], + "bbox": [ + 637, + 119, + 741, + 200 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/470027181c8788def4962a2cb500abaecdf08e3b0065ecb81162ccd62031193a.jpg", + "image_caption": [ + "As a funkopo pop figure", + "sno" + ], + "image_footnote": [], + "bbox": [ + 746, + 119, + 852, + 200 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/bae4e155d601fde33a7b29b195a91da186fa1ad9bdacfa92f010f860b3c140c3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 305, + 204, + 411, + 286 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1acee2229fa132d679380b66f722ceadf163a3e06fecbe1798566c94e048a9e4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 413, + 204, + 521, + 286 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4c656a8478eac94baea4a5f46a8ec68b5644228a82dd1f1edda604b9979e8a3d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 524, + 204, + 630, + 285 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3c8c531c0d9339fec94b4eaae0a4d62e05a79fe2b46cb15ecc6325dbbaee5298.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 637, + 204, + 741, + 286 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b6a744410b5976439a5c4a79353521ee8dbf4f8a76c6fc2cc4329a1a326533f5.jpg", + "image_caption": [ + "Textual", + "Inversion" + ], + "image_footnote": [], + "bbox": [ + 746, + 204, + 851, + 286 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/180ff43a4471cf43ea816e1243b77b932d198909a60a6b4e64725978f9724516.jpg", + "image_caption": [ + "Figure 3. Qualitative examples (CelebA HQ) - Qualitative examples of personalized human face generation using the CelebA HQ dataset [20, 25] are shown. Our hypernetwork field achieves fast adaptation, producing results comparable to DreamBooth [39] and Textual Inversion [13] while preserving individual features in a personalized manner. Additional examples are provided in the supplementary material." + ], + "image_footnote": [], + "bbox": [ + 305, + 289, + 411, + 371 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/160c7b767ca3e9d702cb1781bc89b5da22bbb54e8474e4dffb4371f6edecdf5f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 413, + 289, + 521, + 371 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9981e2fe21724ccb90b9229b5e656c0c9264a035f0feae44454f90f50f940834.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 524, + 290, + 630, + 371 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ac323e7721d9bfeee374cb1fd5f5d1aa2f780b171239edbf62eebfcd6aceb659.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 637, + 289, + 741, + 371 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/967bdd4ba0b1e0d349229b9f66fea6dd6cf970614cf34580ff19e5c15b7dc857.jpg", + "image_caption": [ + "DreamBooth" + ], + "image_footnote": [], + "bbox": [ + 746, + 289, + 852, + 371 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "for a given source image, we query the hypernetwork field with the source image and the timestep $T$ to obtain the weights for the DreamBooth model. Using these weights and the conditioning prompt, we can generate personalized images in a single step.", + "bbox": [ + 89, + 465, + 482, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 556, + 307, + 573 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use a frozen Vision Transformer (ViT) [9] to encode the images. We then pass this encoding into a hypernetwork with a 6-layer Diffusion Transformer (DiT) [30] with adaptive layer norm conditioning [5] for both the timestep and the image embedding. We set the sequence length of the DiT is equal to the total number of layers in the task-specific network. We then have a dedicated Multi-Layer Perceptron (MLP) per each layer of the task-specific network that decodes to the network weights of the target network.", + "bbox": [ + 89, + 580, + 482, + 717 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 737, + 220, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We consider two diverse applications - image personalization and 3D shape reconstruction as target applications. While we focus on these two, we emphasize that our hypernet fields framework does not make any assumption on the task-specific network and applies to any task. Moreover, the definition of the loss, $L_{task}(\\theta_t, c)$ is not restricted to a particular domain. Thus, our method can be applied to these problems by modifying the task-specific optimization model.", + "bbox": [ + 89, + 763, + 482, + 898 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Application to Personalized Image Generation", + "text_level": 1, + "bbox": [ + 511, + 465, + 901, + 482 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To apply our method to personalized image generation, we follow prior work [40] and evaluate and compare personalized face generation on CelebA HQ [20, 25], as well as on animal faces with the AFHQ dataset [7].", + "bbox": [ + 511, + 492, + 905, + 553 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Task network - DreamBooth. To personalize and generate customized images we use DreamBooth as our task network, as in HyperDreamBooth [40]. DreamBooth builds upon the denoising diffusion probabilistic models [38, 42]. The generation process is guided by a special conditioning token encoding the characteristics of the input conditioning image, which is found by optimizing through diffusion models, which require a non-trivial amount of compute. On the CelebA dataset [25] we follow prior work [40] and use the text condition \"a [V] face\" where the token [V] represents a placeholder for the specific identity being learned. Analogously, for the AFHQ dataset [7] with animal images, we use the text condition \"a [V] animal\" to guide the model. The conditioning token along with the input prompts are then used to synthesize personalized images.", + "bbox": [ + 511, + 560, + 906, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training setup. For the CelebA-HQ dataset [20, 25], we follow prior work [40] and replicate their setup to make results comparable. Specifically, among the $30\\mathrm{k}$ high-quality aligned faces, we use the train/test splits and evaluation prompts from HyperDreamBooth [40], with $29.9\\mathrm{k}$ images in the training set and 100 images in the test set. For the AFHQ dataset [7] we use the live subject prompts from the", + "bbox": [ + 511, + 794, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22133", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d37fc23d057773e660e30ad2344787d7e0a70fa6271e93b2c7605ea59110ab43.jpg", + "image_caption": [ + "Input Image" + ], + "image_footnote": [], + "bbox": [ + 91, + 122, + 325, + 300 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/740ea11955a3f836234c18ec8363fbac2ba1e63670fe667ba45deddfea599964.jpg", + "image_caption": [ + "Wearing a Santa hat" + ], + "image_footnote": [], + "bbox": [ + 336, + 119, + 464, + 220 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0f1ddb8fb62e9660f1fdfb3030a5f5d1fbbb1334c2496fae8e7b5c61c4cc0854.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 336, + 222, + 462, + 321 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8d54b4a78a62ab2973794d8cae98b2258adf95f0fee00522a98944f5a4d09a1e.jpg", + "image_caption": [ + "Figure 4. Qualitative examples (AHFQ) - Qualitative results of personalized animal image generation from the AFHQ dataset [7] demonstrate the adaptability of our hypernetwork field. Compared to DreamBooth [39] and Textual Inversion [13], our method effectively captures the specific animal's characteristics while enabling rapid adaptation. Further examples are available in the supplementary material." + ], + "image_footnote": [], + "bbox": [ + 336, + 324, + 462, + 421 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/02cf65aad15da8f3f5d0c4a482928faa7d3097d0cf57ad6655789991bad1e7f4.jpg", + "image_caption": [ + "Wearing pink sunglasses" + ], + "image_footnote": [], + "bbox": [ + 467, + 119, + 594, + 220 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1093262a36b2a1323f7b7ea4802c725a183eff9e527e079fae21558d1d101a22.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 467, + 222, + 594, + 321 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/726c0f2dbfb85a2bf059a428c7989e141812f35db840385e31f05520f2cad163.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 467, + 324, + 594, + 421 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3e00a45e0053026bfb32743865769cbc566e3b94189fe707b9750685f87c37c2.jpg", + "image_caption": [ + "Wearing a rainbow scarf" + ], + "image_footnote": [], + "bbox": [ + 596, + 119, + 725, + 220 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b2d30e095c7807e0a39bd222b9aa1e31095eacbc11194ff991e3b7da35d0f7fa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 596, + 222, + 723, + 320 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/85f362426127572ed2ac8dbcbbc8042e062dbbd5be08c8067fc8e2e010fce0bb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 596, + 323, + 723, + 421 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/098c1d4258d257a247900c6a2c80493ba1ab6c2fdf76cbfbd318e5e34e05c985.jpg", + "image_caption": [ + "In the snow", + "o" + ], + "image_footnote": [], + "bbox": [ + 728, + 119, + 856, + 220 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/12e0a3ecbac55668b920bdbe44f33d480c877bbc73be130dcdf31873b634ac29.jpg", + "image_caption": [ + "Textual", + "inversion" + ], + "image_footnote": [], + "bbox": [ + 728, + 222, + 856, + 321 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bed2e2b4c033a6f67d2971ddbf5c8dfa59f7277e581422f50f162e28c744a834.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 728, + 323, + 856, + 421 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "DreamBooth dataset [39] so that the results are comparable. Among the 16.1k aligned animal faces, we use the established train and test splits, leaving 14.6k images for training and 1.5k for testing. In addition to the DreamBooth and HyperDreamBooth approach, we compare against Textual Inversion [13].", + "bbox": [ + 89, + 507, + 482, + 598 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For both experiments, we set the number of optimization steps to be $T = 500$ . Our network is trained to output rank-1 Low-Rank Adaptation (LoRA) [18] weights for both the UNet and the text encoder for Stable Diffusion 1.5 [38], which amounts to 273k parameters across 352 layers. We train our model with a batch size of 256 and for 20k iterations on 8 NVidia A100 GPUs, which takes 1.5 days. Similar to the HyperDreamBooth [40] approach once our model is trained, we can perform 50 steps of fast fine-tuning starting from the hypernetwork outputs.", + "bbox": [ + 89, + 602, + 482, + 753 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Note that, unlike existing works that require per-instance fitting, with our method, it is possible to introduce on-the-fly data augmentation. We thus incorporate color, saturation, contrast, and hue augmentations to perturb the appearance of the conditioning signal $\\mathbf{x}$ , including horizontal flips.", + "bbox": [ + 89, + 757, + 482, + 833 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metrics. We use the standard metrics suggested by DreamBooth [40]: CLIP-T, CLIP-I [34], and DINO [6]. CLIP-T measures the alignment between the generated image and the textual prompt, while CLIP-I and DINO eval", + "bbox": [ + 89, + 839, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ba2ce9e7805acc7c7345aa9fb4f8309e7aa606a0682b83f771cd0a2941ec7b62.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodFace Rec.↑DINO↑CLIP-I↑CLIP-T↑
Hyper-DreamBooth0.6550.4730.5770.286
DreamBooth0.6180.4410.5460.282
Textual Inversion0.6230.2890.4720.277
Ours0.3250.6050.6390.268
", + "bbox": [ + 521, + 505, + 901, + 588 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Quantitative results (CelebA HQ) - We report the standard metrics for our method and the baselines. Our method is able to offer comparable performance, with significantly less compute. In terms of the image alignment metrics, our method significantly outperforms other methods.", + "bbox": [ + 511, + 599, + 905, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "uate the visual consistency of the generated image with the reference object's appearance. We also include the identity preservation metric (Face Rec.) suggested by HyperDream-Booth [40].", + "bbox": [ + 511, + 698, + 905, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Compute time. As noted earlier, our method requires 1.5 days to train with 8 GPUs for the AFHQ and CelebA datasets, with around $14.6\\mathrm{k}$ samples for AFHQ and $29.9\\mathrm{k}$ samples for CelebA, amounting to 12 days of GPU time. This is much faster compared to the training of HyperDreamBooth [40] that requires 15k DreamBooth weights to be precomputed, which amounts to 50 days of GPU time, before their hypernetwork can start training. Thus, our method trains in less than a quarter of the pre-processing", + "bbox": [ + 511, + 763, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22134", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ca7379a1512158e9eb1ab24153dcdfaedffb30bb0fcc6009fa52b203147a152e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodDINO↑CLIP-I↑CLIP-T↑
Textual Inversion0.5960.7800.285
DreamBooth0.5600.7630.268
Ours0.6640.8070.277
", + "bbox": [ + 91, + 88, + 483, + 165 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Quantitative results (AFHQ) - We outperform other baselines in terms of image alignment and provide comparable results in terms of CLIP-T, the prompt alignment. We note that our method outperforms DreamBooth [39] for all metrics, benefiting from the fact that our HyperNet Field learns a general trend of how DreamBooth weights behave.", + "bbox": [ + 89, + 176, + 483, + 260 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/51dc2064d770251ca7022afc7ca6a59b081c97cd5494c04410fa8943cfb8faac.jpg", + "image_caption": [ + "Figure 5. Fast training (AFHQ) - We show example outcomes of our method only after 100 iterations of training, as raw output from our Hypernetwork without fast fine-tuning. As shown, even after only 100 iterations, our model is able to generate images that are visually very similar to the conditioning images." + ], + "image_footnote": [], + "bbox": [ + 102, + 273, + 472, + 470 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "time of HyperDreamBooth.", + "bbox": [ + 89, + 578, + 276, + 593 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Since we query our Hypernetwork at a single optimization step $T$ , our method also yields efficient inference. Inference of our method takes only 0.3 seconds, compared to 2 seconds for HyperDreamBooth [40] and 5 minutes for DreamBooth. Additionally, the fast fine-tuning required by both our method and HyperDreamBooth takes just 20 seconds, allowing for quick adaptation with minimal overhead.", + "bbox": [ + 89, + 594, + 483, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results. We provide qualitative highlights in Figs. 3 and 4, and the quantitative summaries in Tables 1 and 2. Our results for image personalization in Figs. 3 and 4 is on par with the target Dreambooth model whose weights we estimate with our HyperNet Fields approach. That is, our approach generates personalized images that are representative of the given input prompt and the conditional image. Again, note that we achieve these personalization results in a single step by querying the hypernetwork once at timestep $T$ . Interestingly, our method demonstrates remarkable performance for preserving the original image context, as measured by DINO and CLIP-I scores (Tab. 1). Further, in the case of the AFHQ dataset (Tab. 2), we outperform", + "bbox": [ + 89, + 703, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/622f847ef9789721e14016289303b3cef1492280a23a9b1e8a5d3561ee2d4e94.jpg", + "image_caption": [ + "Figure 6. Qualitative examples (3D Shapes) – We show example shape reconstruction results converting (left) point clouds to 3D shapes and (right) images to 3D shapes. More qualitative examples are available in the supplementary material." + ], + "image_footnote": [], + "bbox": [ + 517, + 90, + 906, + 330 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "all DreamBooth variants, demonstrating that not only do we remove the compute-heavy optimization requirement of DreamBooth [39], but also by amortizing the behavior of the learned weights with respect to the input conditioning image, we successfully replicate the dynamics of the task-specific dreambooth network yielding improved results. We emphasize that while results are comparable to each other, our method is significantly faster than the baselines.", + "bbox": [ + 511, + 426, + 906, + 547 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fast training. An interesting observation from our training process is that our model exhibits faster adaptation to personalization. As shown in Fig. 5, after only 100 training iterations, our model generates images that are visually similar to the conditioning images, achieving this level of personalization with minimal training time. We hypothesize that this rapid adaptation is due to our model's ability to simulate the optimization pathway, effectively learning the direction in which to adjust network weights to reduce the task-specific loss.", + "bbox": [ + 511, + 551, + 908, + 703 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Application to 3D shape reconstruction", + "text_level": 1, + "bbox": [ + 511, + 712, + 849, + 729 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We further explore the potential of our method for the task of 3D shape reconstruction, specifically the task of estimating the weights of an occupancy network [27]. Additionally, we consider two different conditioning inputs: a rendered image of the shape [24, 50], and a point cloud of the shape [1, 15, 32]. For the image conditioning, we use the 38 renderings provided in Qiu et al. [33]—covering various angles around each object—which we sample one of and encode with a pre-trained ViT [9]. For the point cloud conditioning, we randomly sample 2048 points uniformly from the surface of the ground truth mesh and encode the input point", + "bbox": [ + 511, + 734, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22135", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/af6289ace7e35a25b4d13c097e058ea67458d953ad8fd060fca9f559fd843772.jpg", + "image_caption": [ + "Figure 7. Hypernet Field output example (ShapeNet) - We show the output of our HyperNet Field for various points along the optimization trajectory as well as the intersection over union (IoU) over all values of T. As shown, the HyperNet Field is able to accurately estimate the entire optimization trajectory, starting from network weights that only represent the airplane shape roughly, and converging to one that provides a detailed reconstruction of the airplane. The IoU plot also shows that as we traverse the trajectory, the IoU of the generated shapes to the ground truth shape improves, demonstrating convergence. Despite being shown only one ground truth optimization step at a time, our model is able to represent the entire optimization trajectory from initialization to convergence." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 903, + 195 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "cloud with an off-the-shelf variational autoencoder [48].", + "bbox": [ + 89, + 315, + 460, + 330 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Task network - Occupancy Net. For both cases, the hypernetwork field architecture remains the same as described in Sec. 3.4, and for the task-specific network, we follow an architecture similar to HyperDiffusion [12] —we use a simple 3-layer dense network [41] with ReLU activations and a hidden dimension of 128. The task-specific network, the occupancy network, takes the 3D coordinates of a query point as input and outputs a scalar value representing the occupancy as in Mescheder et al. [27]. We use marching cubes [26] to render the outcome.", + "bbox": [ + 88, + 335, + 483, + 488 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Training setup. To supervise the training, we sample 20k query points, 10k uniformly, and 10k near the ground-truth surface to be used for training. To find the ground-truth occupancy values for these points, we first make the meshes watertight with [19] and use winding numbers [4] on the watertight mesh to determine whether each point lies inside or outside the surface, and compute the occupancy via Barill et al. [3]. We use the mean squared error (MSE) of the occupancy estimates for the task-specific loss. For all cases, we set $T = 10,000$ .", + "bbox": [ + 88, + 494, + 483, + 645 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "2D image conditioning (Objaverse shapes). For 2D image conditioning, as training over the entire Objaverse dataset [8] would be computationally expensive, we simply overfit to a subset of 128 shapes, as we are mostly interested in the potential of our method. We leave training on the full Objaverse dataset as future work. We obtain reasonable shape reconstruction results as shown in Figures 1 and 6 with our method for this simple overfitting demonstration just within 10 hours of training on a single NVidia RTX 3090 GPU.", + "bbox": [ + 88, + 651, + 483, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3D point cloud conditioning (ShapeNet shapes). For 3D point cloud conditioning, we use the entire dataset. As shown in Fig. 6, our method is able to learn to generate occupancy network weights from point clouds. We further illustrate how both our Hypernet Field output and the IoU for shape generation evolves over the optimization process in Fig. 7. As shown, our method simulates the entire train-", + "bbox": [ + 88, + 794, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ing process of the task-specific network for a given point cloud despite only being shown a single ground truth gradient step at a time.", + "bbox": [ + 511, + 315, + 906, + 359 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 377, + 633, + 392 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We presented Hypernetwork Fields, a novel and efficient hypernetwork training framework. Instead of optimizing individual weights per sample, we use gradient supervision and match gradients along the hypernetworks field and the task-specific network at random points along the optimization trajectory. This drastically reduces the precomputation overhead of training hypernetworks, while maintaining comparable performance to state of the art. We have demonstrated the effectiveness of our method on personalized image generation and 3D shape reconstruction. Our framework is general and can be applied to any hypernetwork training scenario, making it a promising direction for scaling hypernetwork applications to larger datasets.", + "bbox": [ + 511, + 402, + 906, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and future work. Our method enables stable hypernetwork training on large datasets. Because of this, the potential of our method for diverse tasks remains underexplored. This may include tasks such as training a hypernetwork to estimate the weights of NeRFs or even personalized adapters for large language models. However, our approach requires as many output neurons as there are weights in the target network, resulting in suboptimal memory usage. Recent work by Dravid et al. [10] has shown that customized DreamBooth adapters can be represented using only 1k parameters. Integrating our method with this compact representation approach could significantly enhance parameter efficiency while maintaining our proposed benefits of efficient hypernetwork training. These would be interesting avenues for future work.", + "bbox": [ + 511, + 604, + 908, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 513, + 847, + 609, + 863 + ], + "page_idx": 7 + }, + { + "type": "ref_text", + "text": "[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitlagkas, and Leonidas Guibas. Learning representations and generative", + "bbox": [ + 519, + 872, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22136", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "models for 3d point clouds. In ICML, 2018. 7", + "[2] Sudarshan Babu, Richard Liu, Avery Zhou, Michael Maire, Greg Shakhnarovich, and Rana Hanocka. Hyperfields: Towards zero-shot generation of nerfs from text. In ICML. PMLR, 2024. 2", + "[3] Gavin Barill, Neil G Dickson, Ryan Schmidt, David IW Levin, and Alec Jacobson. Fast winding numbers for soups and clouds. Transactions On Graphics, 37(4):1-12, 2018. 8", + "[4] Gavin Barill, Neil G. Dickson, Ryan M. Schmidt, David I. W. Levin, and Alec Jacobson. Fast winding numbers for soups and clouds. Transactions On Graphics, 37:1-12, 2018. 8", + "[5] Tom B Brown. Language models are few-shot learners. NeurIPS, 2020. 5", + "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, pages 9650-9660, 2021. 6", + "[7] Yunjoy Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis for multiple domains. In CVPR, 2020. 5, 6, 11", + "[8] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In CVPR, pages 13142-13153, 2023. 8", + "[9] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 5, 7", + "[10] Amir Dravid, Yossi Gandelsman, Kuan-Chieh Wang, Rameen Abdal, Gordon Wetzstein, Alexei A Efros, and Kfir Aberman. Interpreting the weight space of customized diffusion models. arXiv preprint arXiv:2406.09413, 2024. 8", + "[11] Yilun Du and Igor Mordatch. Implicit generation and modeling with energy based models. NeurIPS, 32, 2019. 3", + "[12] Ziya Erkoç, Fangchang Ma, Qi Shan, Matthias Nießner, and Angela Dai. Hyperdiffusion: Generating implicit neural fields with weight-space diffusion. In ICCV, pages 14300-14310, 2023. 2, 3, 8", + "[13] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. *ICLR*, 2023. 5, 6", + "[14] Will Grathwohl, Kuan-Chieh Wang, Jorn-Henrik Jacobsen, David Duvenaud, Kevin Swersky, and Mohammad Norouzi. Learning energy-based models by diffusion recovery likelihood. ICLR, 2021. 3", + "[15] Thibault Groueix, Matthew Fisher, Vladimir G Kim, Bryan Russell, and Mathieu Aubry. Atlasnet: A papier-mâché approach to learning 3d surface generation. In CVPR, 2018. 7", + "[16] David Ha, Andrew Dai, and Quoc V Le. Hypernetworks. ICLR, 2017. 2", + "[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, pages 6840-6851, 2020. 3", + "[18] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In ICLR, 2022. 6" + ], + "bbox": [ + 93, + 92, + 482, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[19] Jingwei Huang, Yichao Zhou, and Leonidas Guibas. Manifoldplus: A robust and scalable watertight manifold surface generation method for triangle soups. arxiv, 2020. 8", + "[20] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. In ICLR, 2018. 5", + "[21] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. NeurIPS, 35:26565-26577, 2022. 4", + "[22] Yann LeCun, Sumit Chopra, Raia Hadsell, Marc'Aurelio Ranzato, and Fu Jie Huang. A tutorial on energy-based learning. Predicting structured data, 1(2):1-31, 2006. 3", + "[23] Yaron Lipman, Justin Solomon, et al. Flow matching for generative modeling. *ICLR*, 2023. 3", + "[24] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9298–9309, 2023. 7", + "[25] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In ICCV, 2015. 5, 11", + "[26] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. In Semin al graphics: pioneering efforts that shaped the field, pages 347-353, 1998. 8", + "[27] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In CVPR, pages 4460-4470, 2019. 2, 7, 8", + "[28] Jonathan Oswald, Christian Henning, Tomas Pfister, and Jürgen Schmidhuber. Continual learning with hypernetworks. In ICLR, 2019. 2", + "[29] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In CVPR, pages 165-174, 2019. 2", + "[30] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, pages 4195-4205, 2023. 5", + "[31] Marcin Przewiezlikowski, Przemyslaw Przybysz, Jacek Tabor, Maciej Zikeba, and Przemyslaw Spurek. Hypermaml: Few-shot adaptation of deep models with hypernetworks. Neurocomputing, 2024. 2", + "[32] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, 2017. 7", + "[33] Lingteng Qiu, Guanying Chen, Xiaodong Gu, Qi Zuo, Mutian Xu, Yushuang Wu, Weihao Yuan, Zilong Dong, Liefeng Bo, and Xiaoguang Han. Richdreamer: A generalizable normal-depth diffusion model for detail richness in text-to-3d. In CVPR, pages 9914–9925, 2024. 7", + "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 6" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22137", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[35] Maziar Raissi, Paris Perdikaris, and George Em Karniadakis. Physics informed deep learning (part i): Data-driven solutions of nonlinear partial differential equations. arxiv, 2017. 3", + "[36] Maziar Raissi, Paris Perdikaris, and George Em Karniadakis. Physics informed deep learning (part ii): Data-driven discovery of nonlinear partial differential equations. arxiv, 2017.", + "[37] Maziar Raissi, Paris Perdikaris, and George Em Karniadakis. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational Physics, 378:686-707, 2019. 3", + "[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 5, 6", + "[39] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, pages 22500-22510, 2023. 1, 3, 4, 5, 6, 7", + "[40] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdream booth: Hypernetworks for fast personalization of text-to-image models. In CVPR, pages 6527-6536, 2024. 2, 3, 5, 6, 7", + "[41] David E Rumelhart, Geoffrey E Hinton, and Ronald J Williams. Learning representations by back-propagating errors. NATURE, 323(6088):533-536, 1986. 8", + "[42] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 35:36479-36494, 2022. 5", + "[43] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, pages 2256-2265, 2015. 3", + "[44] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. NeurIPS, 32, 2019. 3", + "[45] Yang Song, Conor Durkan, Iain Murray, and Stefano Ermon. Maximum likelihood training of score-based diffusion models. NeurIPS, pages 1415-1428, 2021. 3", + "[46] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. arxiv, 2023. 3, 4", + "[47] Pascal Vincent. A connection between score matching and denoising autoencoders. Neural Computation, 23(7):1661-1674, 2011. 3", + "[48] Guandao Yang, Xun Huang, Zekun Hao, Ming-Yu Liu, Serge Belongie, and Bharath Hariharan. Pointflow: 3d point cloud generation with continuous normalizing flows. In ICCV, pages 4541-4550, 2019. 8", + "[49] Linnan Zhang, Zhiding Ren, Seunghyun Ahn, Hakan Bilen, and Kwang In Lee. Graph hypernetworks for neural architecture search. *ICLR*, 2019. 2", + "[50] Jun-Yan Zhu et al. Shap-e: Generating conditional 3d implicit functions. arxiv, 2023. 7" + ], + "bbox": [ + 91, + 92, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22138", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/HyperNet Fields_ Efficiently Training Hypernetworks without Ground Truth by Learning Weight Trajectories/68fd503f-6b57-47fa-8aca-593407ee13b9_model.json b/2025/HyperNet Fields_ Efficiently Training Hypernetworks without Ground Truth by Learning Weight Trajectories/68fd503f-6b57-47fa-8aca-593407ee13b9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..70200eb0912c833c15e7beb0bb9d2dff196dfe39 --- /dev/null +++ b/2025/HyperNet Fields_ Efficiently Training Hypernetworks without Ground Truth by Learning Weight Trajectories/68fd503f-6b57-47fa-8aca-593407ee13b9_model.json @@ -0,0 +1,2541 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.187, + 0.131, + 0.817, + 0.177 + ], + "angle": 0, + "content": "HyperNet Fields: Efficiently Training Hypernetworks without Ground Truth by Learning Weight Trajectories" + }, + { + "type": "text", + "bbox": [ + 0.149, + 0.202, + 0.851, + 0.24 + ], + "angle": 0, + "content": "Eric Hedlin\\(^{1,2*}\\) Munawar Hayat\\(^{2}\\) Fatih Porikli\\(^{2}\\) Kwang Moo Yi\\(^{1}\\) Shweta Mahajan\\(^{2}\\) \n\\(^{1}\\) University of British Columbia \\({}^{2}\\) Qualcomm AI Research†" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.274, + 0.5, + 0.517 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.275, + 0.903, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.527, + 0.908, + 0.57 + ], + "angle": 0, + "content": "Figure 1. Teaser - We propose hypernetwork fields, where we learn the entire weight space trajectory instead of only the final state, which allows us to train without ever needing to know the final converged weights. Our method can be applied to any application of hypernetworks, including diffusion model personalization and modeling of 3D neural representations." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.582, + 0.327, + 0.598 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.614, + 0.486, + 0.902 + ], + "angle": 0, + "content": "To efficiently adapt large models or to train generative models of neural representations, Hypernetworks have drawn interest. While hypernetworks work well, training them is cumbersome, and often requires ground truth optimized weights for each sample. However, obtaining each of these weights is a training problem of its own—one needs to train, e.g., adaptation weights or even an entire neural field for hypernetworks to regress to. In this work, we propose a method to train hypernetworks, without the need for any per-sample ground truth. Our key idea is to learn a Hypernetwork 'Field' and estimate the entire trajectory of network weight training instead of simply its converged state. In other words, we introduce an additional input to the Hypernetwork, the convergence state, which then makes it act as a neural field that models the entire convergence pathway of a task network. A critical benefit in doing so is that the gradient of the estimated weights at any convergence state must then match the gradients of the original task—this constraint alone is sufficient to train the Hypernetwork Field." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.583, + 0.909, + 0.645 + ], + "angle": 0, + "content": "We demonstrate the effectiveness of our method through the task of personalized image generation and 3D shape reconstruction from images and point clouds, demonstrating competitive results without any per-sample ground truth." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.689, + 0.646, + 0.705 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.715, + 0.907, + 0.853 + ], + "angle": 0, + "content": "Various generative modeling tasks [39] require adaptation, thus data-dependent network weights, or in other words, model parameters being adjusted according to various circumstances (conditions and user input). These include personalized text-to-image generation where the weights of the underlying personalization model, such as DreamBooth [39], are fine-tuned on the concept of interest and used to generate target images conditioned on input prompts. Another example is 3D modeling, where neural" + }, + { + "type": "page_footnote", + "bbox": [ + 0.531, + 0.864, + 0.764, + 0.876 + ], + "angle": 0, + "content": "*Work done as part of a summer internship." + }, + { + "type": "page_footnote", + "bbox": [ + 0.516, + 0.877, + 0.905, + 0.899 + ], + "angle": 0, + "content": "\\(\\dagger\\)Qualcomm AI Research is an initiative of Qualcomm Technologies, Inc." + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.864, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22129" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "networks are used to model shapes implicitly with signed distance fields [29] or occupancy [27]. In both cases, a dedicated set of neural network weights is required for each personalization outcome or 3D shape. Thus, the applicability of these methods is highly limited by the training time required to obtain these weights." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.184, + 0.483, + 0.424 + ], + "angle": 0, + "content": "It is therefore critical to seek ways to amortize the per-sample optimizations into Hypernetworks [12, 40]—networks that estimate network parameters. By 'skipping ahead' the tedious training processes involved in fine-tuning or shape regressing, they greatly speed up the entire process. Existing methods, however, have a critical shortcoming in that they supervise hypernetworks by directly matching their output to the precomputed, ground truth converged weights corresponding to the task-specific network for each sample in the dataset. Notably, prior to any training of the Hypernetwork, one must prepare converged task-specific network weights for every sample in the training set. This severely limits the capabilities of Hypernetwork methods when scaling to large datasets. As an example, to train HyperDreamBooth [40], it requires 50 days of GPU time on a NVidia RTX 3090,\\(^{\\ddagger}\\) to just prepare the training data alone." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.425, + 0.482, + 0.546 + ], + "angle": 0, + "content": "Furthermore, current methods assume a one-to-one mapping between each input sample and its corresponding optimized weights. This assumption is an oversimplification. Neural network optimization is stochastic with various equally plausible solutions. Therefore, enforcing a bijective mapping between the condition and the ultimately converged weights may limit the potential expressiveness of the hypernetwork, as we will demonstrate in Sec. 3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.547, + 0.483, + 0.743 + ], + "angle": 0, + "content": "To address these limitations, we instead propose to train a Hypernetwork Field that estimates the entire optimization trajectory of a neural network training, rather than focusing on a single final set of weights for each sample. Specifically, we introduce the convergence state as additional input to the hypernetwork, such that the weights of the task-specific network at a desired state of convergence can be 'queried'. An interesting outcome of doing so is that the gradient of the estimated weights along this pathway must match the gradients of the original task-specific network, as they are both gradients pointing towards convergence. Hence, training of such hypernetwork fields can simply be done by matching the two gradients." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.744, + 0.483, + 0.864 + ], + "angle": 0, + "content": "By supervising the hypernetwork to match the gradient of the weights with respect to the optimization step, we avoid the need for precomputing target weights. This supervision strategy allows the hypernetwork to estimate a trajectory of parameters that, at each step, reflects a compatible state across all samples in the dataset. Ultimately, at inference time, the estimated parameters which correspond with the final timestep of the hypernetwork, represent a well" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.182 + ], + "angle": 0, + "content": "converged solution for each of the samples, but are now naturally discovered through training of the hypernetwork field itself. Notably, this single forward pass to estimating the converged parameters is all that is necessary for inference, hence no additional compute is required compared to conventional hypernetwork approaches [12, 40]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.184, + 0.905, + 0.29 + ], + "angle": 0, + "content": "Our framework is general and is applicable to all hypernetwork training scenarios. To demonstrate the effectiveness of our method, we apply our method to personalized image generation and 3D shape generation. We achieve competitive performance while reducing the training cost by \\(\\sim 4\\) times what would be required of conventional hypernetwork training." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.292, + 0.778, + 0.305 + ], + "angle": 0, + "content": "To summarize, our contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.308, + 0.905, + 0.368 + ], + "angle": 0, + "content": "- We propose hypernetwork fields, an approach to learn the entire optimization pathway of task-specific networks, which allows ground truth-free training without the need for precomputing the target weights;" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.369, + 0.905, + 0.398 + ], + "angle": 0, + "content": "- our approach drastically reduces the compute required for training hypernetworks;" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.399, + 0.905, + 0.459 + ], + "angle": 0, + "content": "- we validate our approach on two very different tasks and datasets—for personalized image generation we achieve results comparable to the state of the art with significantly less compute; and" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.459, + 0.905, + 0.489 + ], + "angle": 0, + "content": "- we further explore the application of our method to 3D shape reconstruction from images an point clouds." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.308, + 0.905, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.508, + 0.655, + 0.525 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Hypernetworks. Early work on training hypernetworks backpropagates gradients through the task-specific network to the hypernetwork itself, as demonstrated by approaches [16, 31] and subsequent explorations in continual learning [28] and neural architecture search [49]. These methods have shown that allowing gradient flow through both networks enables the hypernetwork to modify parameters for diverse tasks adaptively. However, more recent works have introduced approaches that involve precomputing weights for the task-specific network and directly supervise the hypernetwork to match them, eliminating the need for backpropagation through the task-specific network and improving training stability by alleviating issues with gradient flow. This supervision with precomputed weights has been shown to enhance gradient stability and training performance, as highlighted in frameworks like HyperDiffusion [12], HyperFields [2], and HyperDreamBooth [40]. However, this limits the size of the dataset to which this can be applied, as the precompute requirement is proportional to the number of samples in the dataset. Our method combines the best of both approaches by removing the need for precomputed weights while explicitly providing weight-based supervision in the form of gradients along the convergence path, enabling stable training on large datasets." + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.875, + 0.483, + 0.901 + ], + "angle": 0, + "content": "\\(\\ddagger\\)HyperDreamBooth [40] was trained on 15k DreamBooth weights, which each take 5 minutes to compute." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22130" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.484, + 0.364 + ], + "angle": 0, + "content": "Denoising Models. In generative modeling, finding smooth and reliable paths through latent space has become central to producing coherent samples. Examples include diffusion models [17, 43], which iteratively trace a path from noise to data through successive steps, effectively navigating high-dimensional space by reversing a stochastic process. Consistency models [46] refine this by learning non-Markovian trajectories that reach target data distributions in fewer steps by introducing a formulation that allows them to single-step jump to any timestep. Flow-matching models [23] take a similar path-based approach by learning a vector field that guides samples along pre-defined flows, ensuring a direct and smooth transition through the latent space. Similarly to these methods, our method can be thought of as finding a path through this high-dimensional space, specifically in the form of a single-step estimator from initialization to any desired state of convergence, whose path travels strictly along the direction of the steepest gradient descent." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.369, + 0.485, + 0.825 + ], + "angle": 0, + "content": "Methods with gradient-supervision. Several recent approaches in generative modeling and physics-informed learning employ supervision exclusively through gradients rather than direct supervision between input-output pairs to achieve denoising and reconstruction. Score-based generative models [44, 45, 47] use score matching to learn the gradient (or \"score\") of the log density of noisy data distributions, bypassing the need for direct output supervision. Denoising Score Matching (DSM) [47] trains the model to predict this gradient, guiding it to approximate data distributions through iterative denoising and demonstrating improved stability in noise removal. Similarly, Physics-Informed Neural Networks (PINNs) [35-37] leverage known derivatives from partial differential equations (PDEs) in physical simulations, such as heat diffusion or fluid flow, to inform the network solely through spatial or temporal derivatives. This derivative-based guidance encourages convergence as the model adapts to match observed physical behaviour, requiring no explicit target outputs. Energy-Based Models (EBMs) [11, 14, 22] also align with this gradient-driven approach by learning energy landscapes that adjust gradients to represent target distributions, allowing for smooth gradient matching through learned energy fields. Inspired by these methods, our approach supervises only the gradient direction in task-specific weight space, effectively learning a convergence pathway without requiring explicit final state supervision. By supervising gradients within this pathway, our method achieves robust training stability while generalizing over diverse tasks without per-sample ground truth weights." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.843, + 0.182, + 0.859 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.903 + ], + "angle": 0, + "content": "To remove the precompute requirements associated with training hypernetworks, we present a new framework that" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "models the convergence dynamics of task-specific networks into a single, efficient process, which we name Hypernetwork Field." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.141, + 0.907, + 0.323 + ], + "angle": 0, + "content": "Problem formulation. Formally, let \\( f_{\\theta} \\) denote a task-specific neural network, e.g., dreambooth [39], parameterized by \\( \\theta \\). \\( f_{\\theta} \\) is then typically trained to minimize a loss function \\( \\mathbb{E}_{\\mathbf{x}}[\\mathcal{L}(\\theta ,\\mathbf{x})] \\), by training for \\( T \\) optimization steps given a dataset of conditions, e.g., image \\( \\mathbf{x} \\). In our work, we wish to learn a hypernetwork field \\( H_{\\phi}(t,\\mathbf{x}) \\), parameterized by \\( \\phi \\) that takes as input a condition \\( \\mathbf{x} \\) and an optimization step \\( t \\). The hypernetwork field then predicts the parameters \\( \\theta_t = H_\\phi (t,\\mathbf{x}) \\) at optimization step \\( t \\) of the task-specific network \\( f_{\\theta_t}(\\mathbf{x}) \\), given the condition \\( \\mathbf{x} \\) and the timestep \\( t \\). In the following, we show how to obtain \\( H_{\\phi}(t,\\mathbf{x}) \\) without requiring any \\( \\theta_t \\) to supervise the training." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.332, + 0.677, + 0.348 + ], + "angle": 0, + "content": "3.1. HyperNet Fields" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.354, + 0.906, + 0.46 + ], + "angle": 0, + "content": "To train the hypernet field \\( H_{\\phi}(t,\\mathbf{x}) \\) without the ground-truth targets for it to regress, we take a step back and revisit what standard hypernetworks are optimizing for. Standard approaches for learning hypernetworks [12, 40] utilize the ground-truth weights \\( \\theta^{*} \\) of the task-specific network, corresponding to the optimized parameters for a specific sample \\( \\mathbf{x} \\) as a supervision signal. In other words, they first find" + }, + { + "type": "equation", + "bbox": [ + 0.622, + 0.472, + 0.907, + 0.497 + ], + "angle": 0, + "content": "\\[\n\\theta^ {*} (\\mathbf {x}) = \\underset {\\theta} {\\arg \\min } \\mathcal {L} (\\theta , \\mathbf {x}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.906, + 0.599 + ], + "angle": 0, + "content": "for each \\(\\mathbf{x}\\), then find \\(\\phi\\), such that \\(\\theta^{*}(\\mathbf{x}) \\approx H_{\\phi}(\\cdot, \\mathbf{x}), \\forall \\mathbf{x}\\). This typically requires several iterations of the optimization process and thus incurs a large computational cost for generating the target parameters \\(\\theta^{*}(\\mathbf{x})\\). Moreover, since this process has to be repeated for each sample independently, the scalability to large datasets is severely limited." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.6, + 0.906, + 0.78 + ], + "angle": 0, + "content": "Note, however, Eq. (1) is typically found through gradient-based optimization, and forms a trajectory of weight parameter space in the \\(\\{\\theta_t\\}_{t = 1\\dots T}\\), where \\(T\\) is the number of optimization steps. If we now consider our hypernet field \\(H_{\\phi}(t,\\mathbf{x})\\) that we intend to model, we notice an interesting relationship that must hold between the \\(\\mathcal{L}(\\theta_t,\\mathbf{x})\\) and \\(H_{\\phi}(t,\\mathbf{x})\\). Specifically, consider the gradient of the task-specific weights at an optimization step \\(t\\), \\(\\nabla_{\\theta_t}\\mathcal{L}(\\theta_t,\\mathbf{x})\\). Then, if a hypernet field \\(H_{\\phi}\\) has been properly trained, the direction in which \\(\\nabla_{\\theta_t}\\mathcal{L}(\\theta_t,\\mathbf{x})\\) point to must match the direction in which the hypernet trajectory flows, that is, \\(\\nabla_tH_\\phi (t,\\mathbf{x})\\). In other words, their gradients must match." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.909, + 0.902 + ], + "angle": 0, + "content": "In fact, matching the two gradients allows us to circumvent the necessity of per-sample optimization as \\(\\nabla_{\\theta_t}\\mathcal{L}(\\theta_t,\\mathbf{x})\\) can be computed with relatively minimal overhead by using the current hypernet field estimate for \\(\\theta_t\\). This process can also be understood as amortizing Eq. (1) into the training process of the hypernet field \\(H_{\\phi}(t,\\mathbf{x})\\), as we are 'skipping ahead' parts of the optimization loop of Eq. (1) with the estimates coming from \\(H_{\\phi}(t,\\mathbf{x})\\). Through" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "22131" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.09, + 0.488, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.101, + 0.907, + 0.175 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.211, + 0.908, + 0.242 + ], + "angle": 0, + "content": "Figure 2. The Hypenet Fields Framework. Left: Our Hypernetwork with gradient-based supervision from a task-specific network - dreambooth [39]. Right: The sampling process to generate personalized images with our hypernetwork framework." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.266, + 0.484, + 0.342 + ], + "angle": 0, + "content": "this amortization, the hypernet field captures detailed information about how the weights evolve throughout training, allowing it to effectively predict and generate weights across the training samples. We now further detail the exact training process." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.35, + 0.345, + 0.367 + ], + "angle": 0, + "content": "3.2. Gradient-based Supervision" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.372, + 0.484, + 0.493 + ], + "angle": 0, + "content": "In each training iteration of the optimization process, the hypernet field generates an estimate of the task-specific parameters \\(\\hat{\\theta}_t = H_\\phi(t, \\mathbf{x})\\) at timestep \\(t\\) based on the input \\(\\mathbf{x}\\) from a dataset \\(\\mathcal{D}\\) and the timestep \\(t \\sim U(0, T)\\). Given this estimate, we compute the gradient of the loss function with respect to the task-specific weights at that timestep, \\(\\nabla_{\\hat{\\theta}_t} \\mathcal{L}(\\hat{\\theta}_t, \\mathbf{x})\\), and perform a single optimization step to update the weights to \\(\\theta_{t+1}\\). This is given by" + }, + { + "type": "equation", + "bbox": [ + 0.181, + 0.5, + 0.482, + 0.521 + ], + "angle": 0, + "content": "\\[\n\\theta_ {t + 1} = \\hat {\\theta} _ {t} - \\eta \\nabla_ {\\hat {\\theta} _ {t}} \\mathcal {L} _ {\\text {t a s k}} (\\hat {\\theta} _ {t}, \\mathbf {x}), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.526, + 0.484, + 0.588 + ], + "angle": 0, + "content": "where \\(\\eta\\) is the learning rate. Then, the hypernet field must also follow this update by definition, thus we wish to have \\(\\hat{\\theta}_{t + 1}\\approx \\theta_{t + 1}\\). Moreover, assuming \\(\\theta_0 = \\hat{\\theta}_0 = H_\\phi (0,\\mathbf{x})\\), the entire trajectory can simply be matched via minimizing" + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.595, + 0.482, + 0.615 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\Delta} (\\mathcal {D}) = \\mathbb {E} _ {\\mathbf {x} \\sim \\mathcal {D}, t \\sim U [ 0, T ]} \\| \\theta_ {t + 1} - H _ {\\phi} (\\mathbf {x}, t + 1) \\| ^ {2}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.622, + 0.484, + 0.712 + ], + "angle": 0, + "content": "which we name as the gradient matching loss. This loss function essentially minimizes the difference between the gradient step from Eq. (2) and the numerical gradient \\(\\nabla_t H_\\phi(\\mathbf{x}, t) = H_\\phi(\\mathbf{x}, t + 1) - H_\\phi(\\mathbf{x}, t)\\) over the entire dataset for \\(T\\) steps of optimization; thus replicating the underlying training dynamics of the task-specific network." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.712, + 0.484, + 0.773 + ], + "angle": 0, + "content": "Once trained, the hypernetwork field \\( H_{\\phi}(t,\\mathbf{x}) \\) thus mimics the entire training trajectory of the task-specific network, instead of the point estimate \\( \\theta^{*}(\\mathbf{x}) \\). We summarize this training process in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.776, + 0.484, + 0.867 + ], + "angle": 0, + "content": "Ensuring a consistent trajectory throughout training. To ensure training stability, we fix the starting position \\(\\theta_0\\) for each sample \\(\\mathbf{x}\\) throughout training, initializing it randomly only once at the beginning. This approach mirrors the typical neural network initialization to random values, followed by convergence toward the optimal parameters." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Forcing 0 offset at \\( t = 0 \\) by construction. Furthermore, Eq. (3) matches the entire trajectory only when \\( \\theta_0 = \\hat{\\theta}_0 = \\)" + }, + { + "type": "title", + "bbox": [ + 0.516, + 0.266, + 0.845, + 0.281 + ], + "angle": 0, + "content": "Algorithm 1 Hypernet Fields Training Algorithm" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.285, + 0.624, + 0.3 + ], + "angle": 0, + "content": "1: Initialize \\(\\theta_0\\)" + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.301, + 0.905, + 0.33 + ], + "angle": 0, + "content": "2: \\(H_{\\phi}\\gets\\) Hypernetwork() \\(\\triangleright\\) Hypernetwork with parameters \\(\\phi\\)" + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.331, + 0.736, + 0.345 + ], + "angle": 0, + "content": "3: while not converged do" + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.346, + 0.905, + 0.36 + ], + "angle": 0, + "content": "4: \\(\\mathbf{x} \\gets \\mathrm{next}(\\mathcal{D}) \\triangleright\\) Sample condition \\(\\mathbf{x}\\) from dataset" + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.362, + 0.725, + 0.376 + ], + "angle": 0, + "content": "5: \\( t \\gets \\) sample from \\( [0, T) \\)" + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.377, + 0.905, + 0.393 + ], + "angle": 0, + "content": "6: \\(\\hat{\\theta}_t\\gets H_\\phi (\\mathbf{x},t)\\) \\(\\triangleright\\) Estimate weights for timestep \\(t\\)" + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.394, + 0.905, + 0.41 + ], + "angle": 0, + "content": "7: \\(\\Delta \\theta_{t} \\gets -\\eta \\nabla_{\\theta_{t}} \\mathcal{L}_{task}(\\hat{\\theta}_{t}, \\mathbf{x})\\) \\(\\triangleright\\) GT gradient for \\(\\hat{\\theta}_{t}\\)" + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.41, + 0.905, + 0.426 + ], + "angle": 0, + "content": "8: \\(\\Delta \\hat{\\theta}_t\\gets H_\\phi (\\mathbf{x},t + 1) - \\hat{\\theta}_t\\) \\(\\triangleright\\) Est gradient for \\(\\hat{\\theta}_t\\)" + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.426, + 0.905, + 0.443 + ], + "angle": 0, + "content": "9: \\(\\phi \\gets \\phi -\\eta \\nabla_{\\phi}\\mathrm{MSE}(\\Delta \\theta_t,\\Delta \\hat{\\theta}_t)\\) Update \\(\\phi\\)" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.444, + 0.614, + 0.457 + ], + "angle": 0, + "content": "10: end while" + }, + { + "type": "list", + "bbox": [ + 0.522, + 0.285, + 0.905, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.482, + 0.907, + 0.543 + ], + "angle": 0, + "content": "\\(H_{\\phi}(0,\\mathbf{x})\\). Enforcing this, however, can easily be done by construction. Inspired by [21, 46], for any timestep \\(t\\), we parameterize the model with an offset from the input \\(\\theta_0\\), conditioned on \\(t\\) as," + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.548, + 0.907, + 0.577 + ], + "angle": 0, + "content": "\\[\nH _ {\\phi} (\\mathbf {x}, t) = \\theta_ {0} + \\frac {t}{T} \\times H _ {\\phi} ^ {\\prime} (\\mathbf {x}, t), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.582, + 0.907, + 0.612 + ], + "angle": 0, + "content": "where \\( H' \\) is the raw hypernetwork field output from a neural network." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.621, + 0.631, + 0.637 + ], + "angle": 0, + "content": "3.3. Algorithm" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.907, + 0.871 + ], + "angle": 0, + "content": "We summarize our overall optimization process in Algorithm 1. We first randomly initialize and fix the parameters \\(\\theta_0\\). We then consider a random sample \\(\\mathbf{x}\\) from the dataset and a random timestep \\(t\\) from the optimization trajectory. With these inputs, we use the hypernetwork \\(H_{\\phi}\\) to obtain the weights \\(\\hat{\\theta}_t\\) of the task-specific network. Next, using these weights and the input \\(\\mathbf{x}\\) we compute the gradient step \\(\\Delta \\theta_t\\) for the task-specific network. Additionally, we calculate \\(\\Delta \\hat{\\theta}_t = \\hat{\\theta}_{t + 1} - \\hat{\\theta}_t\\) from the hypernetwork. The parameters \\(\\phi\\) of the hypernetwork are then updated to match \\(\\Delta \\hat{\\theta}_t\\) with the gradients \\(\\Delta \\theta_t\\), which is equivalent to minimizing Eq. (3). This yields a computationally efficient framework where a single hypernetwork encodes the optimization trajectory and can be queried to get the weights of the task-specific network at any stage in the training process." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "For inference, consider DreamBooth [39] as a task-specific network (Fig. 2). To generate personalized images" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22132" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.149, + 0.1, + 0.238, + 0.115 + ], + "angle": 0, + "content": "Input Image" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.121, + 0.293, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.322, + 0.108, + 0.397, + 0.119 + ], + "angle": 0, + "content": "As a drawing" + }, + { + "type": "image", + "bbox": [ + 0.307, + 0.12, + 0.412, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.442, + 0.095, + 0.501, + 0.118 + ], + "angle": 0, + "content": "As a pixar character" + }, + { + "type": "image", + "bbox": [ + 0.415, + 0.12, + 0.522, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.549, + 0.095, + 0.618, + 0.118 + ], + "angle": 0, + "content": "As a graffiti mural" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.12, + 0.633, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.655, + 0.095, + 0.732, + 0.118 + ], + "angle": 0, + "content": "Flowing river of lost souls" + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.12, + 0.743, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.761, + 0.095, + 0.848, + 0.119 + ], + "angle": 0, + "content": "As a funkopo pop figure" + }, + { + "type": "image", + "bbox": [ + 0.748, + 0.12, + 0.853, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.862, + 0.147, + 0.878, + 0.174 + ], + "angle": 0, + "content": "sno" + }, + { + "type": "image_caption", + "bbox": [ + 0.862, + 0.221, + 0.879, + 0.267 + ], + "angle": 270, + "content": "Textual" + }, + { + "type": "image_caption", + "bbox": [ + 0.884, + 0.22, + 0.9, + 0.272 + ], + "angle": 270, + "content": "Inversion" + }, + { + "type": "image", + "bbox": [ + 0.307, + 0.205, + 0.412, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.415, + 0.205, + 0.522, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.205, + 0.632, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.205, + 0.742, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.747, + 0.205, + 0.852, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.307, + 0.29, + 0.412, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.415, + 0.29, + 0.522, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.291, + 0.632, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.29, + 0.743, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.748, + 0.29, + 0.853, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.862, + 0.295, + 0.879, + 0.366 + ], + "angle": 270, + "content": "DreamBooth" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.385, + 0.907, + 0.44 + ], + "angle": 0, + "content": "Figure 3. Qualitative examples (CelebA HQ) - Qualitative examples of personalized human face generation using the CelebA HQ dataset [20, 25] are shown. Our hypernetwork field achieves fast adaptation, producing results comparable to DreamBooth [39] and Textual Inversion [13] while preserving individual features in a personalized manner. Additional examples are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.467, + 0.483, + 0.543 + ], + "angle": 0, + "content": "for a given source image, we query the hypernetwork field with the source image and the timestep \\( T \\) to obtain the weights for the DreamBooth model. Using these weights and the conditioning prompt, we can generate personalized images in a single step." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.558, + 0.308, + 0.574 + ], + "angle": 0, + "content": "3.4. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.582, + 0.483, + 0.718 + ], + "angle": 0, + "content": "We use a frozen Vision Transformer (ViT) [9] to encode the images. We then pass this encoding into a hypernetwork with a 6-layer Diffusion Transformer (DiT) [30] with adaptive layer norm conditioning [5] for both the timestep and the image embedding. We set the sequence length of the DiT is equal to the total number of layers in the task-specific network. We then have a dedicated Multi-Layer Perceptron (MLP) per each layer of the task-specific network that decodes to the network weights of the target network." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.738, + 0.222, + 0.755 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.9 + ], + "angle": 0, + "content": "We consider two diverse applications - image personalization and 3D shape reconstruction as target applications. While we focus on these two, we emphasize that our hypernet fields framework does not make any assumption on the task-specific network and applies to any task. Moreover, the definition of the loss, \\( L_{task}(\\theta_t, c) \\) is not restricted to a particular domain. Thus, our method can be applied to these problems by modifying the task-specific optimization model." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.467, + 0.903, + 0.483 + ], + "angle": 0, + "content": "4.1. Application to Personalized Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.493, + 0.906, + 0.554 + ], + "angle": 0, + "content": "To apply our method to personalized image generation, we follow prior work [40] and evaluate and compare personalized face generation on CelebA HQ [20, 25], as well as on animal faces with the AFHQ dataset [7]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.561, + 0.907, + 0.789 + ], + "angle": 0, + "content": "Task network - DreamBooth. To personalize and generate customized images we use DreamBooth as our task network, as in HyperDreamBooth [40]. DreamBooth builds upon the denoising diffusion probabilistic models [38, 42]. The generation process is guided by a special conditioning token encoding the characteristics of the input conditioning image, which is found by optimizing through diffusion models, which require a non-trivial amount of compute. On the CelebA dataset [25] we follow prior work [40] and use the text condition \"a [V] face\" where the token [V] represents a placeholder for the specific identity being learned. Analogously, for the AFHQ dataset [7] with animal images, we use the text condition \"a [V] animal\" to guide the model. The conditioning token along with the input prompts are then used to synthesize personalized images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Training setup. For the CelebA-HQ dataset [20, 25], we follow prior work [40] and replicate their setup to make results comparable. Specifically, among the \\(30\\mathrm{k}\\) high-quality aligned faces, we use the train/test splits and evaluation prompts from HyperDreamBooth [40], with \\(29.9\\mathrm{k}\\) images in the training set and 100 images in the test set. For the AFHQ dataset [7] we use the live subject prompts from the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22133" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.173, + 0.099, + 0.255, + 0.113 + ], + "angle": 0, + "content": "Input Image" + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.123, + 0.326, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.368, + 0.091, + 0.434, + 0.118 + ], + "angle": 0, + "content": "Wearing a Santa hat" + }, + { + "type": "image", + "bbox": [ + 0.337, + 0.121, + 0.465, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.337, + 0.223, + 0.464, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.337, + 0.325, + 0.464, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.487, + 0.091, + 0.576, + 0.12 + ], + "angle": 0, + "content": "Wearing pink sunglasses" + }, + { + "type": "image", + "bbox": [ + 0.468, + 0.121, + 0.596, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.468, + 0.223, + 0.595, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.468, + 0.325, + 0.595, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.616, + 0.091, + 0.711, + 0.119 + ], + "angle": 0, + "content": "Wearing a rainbow scarf" + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.121, + 0.726, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.223, + 0.725, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.324, + 0.725, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.755, + 0.108, + 0.832, + 0.119 + ], + "angle": 0, + "content": "In the snow" + }, + { + "type": "image", + "bbox": [ + 0.729, + 0.121, + 0.857, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.729, + 0.223, + 0.857, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.729, + 0.324, + 0.857, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.86, + 0.157, + 0.876, + 0.184 + ], + "angle": 0, + "content": "o" + }, + { + "type": "image_caption", + "bbox": [ + 0.859, + 0.25, + 0.875, + 0.293 + ], + "angle": 270, + "content": "Textual" + }, + { + "type": "image_caption", + "bbox": [ + 0.879, + 0.248, + 0.894, + 0.295 + ], + "angle": 270, + "content": "inversion" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.439, + 0.907, + 0.482 + ], + "angle": 0, + "content": "Figure 4. Qualitative examples (AHFQ) - Qualitative results of personalized animal image generation from the AFHQ dataset [7] demonstrate the adaptability of our hypernetwork field. Compared to DreamBooth [39] and Textual Inversion [13], our method effectively captures the specific animal's characteristics while enabling rapid adaptation. Further examples are available in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.508, + 0.483, + 0.599 + ], + "angle": 0, + "content": "DreamBooth dataset [39] so that the results are comparable. Among the 16.1k aligned animal faces, we use the established train and test splits, leaving 14.6k images for training and 1.5k for testing. In addition to the DreamBooth and HyperDreamBooth approach, we compare against Textual Inversion [13]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.603, + 0.483, + 0.755 + ], + "angle": 0, + "content": "For both experiments, we set the number of optimization steps to be \\( T = 500 \\). Our network is trained to output rank-1 Low-Rank Adaptation (LoRA) [18] weights for both the UNet and the text encoder for Stable Diffusion 1.5 [38], which amounts to 273k parameters across 352 layers. We train our model with a batch size of 256 and for 20k iterations on 8 NVidia A100 GPUs, which takes 1.5 days. Similar to the HyperDreamBooth [40] approach once our model is trained, we can perform 50 steps of fast fine-tuning starting from the hypernetwork outputs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.758, + 0.483, + 0.834 + ], + "angle": 0, + "content": "Note that, unlike existing works that require per-instance fitting, with our method, it is possible to introduce on-the-fly data augmentation. We thus incorporate color, saturation, contrast, and hue augmentations to perturb the appearance of the conditioning signal \\(\\mathbf{x}\\), including horizontal flips." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Evaluation metrics. We use the standard metrics suggested by DreamBooth [40]: CLIP-T, CLIP-I [34], and DINO [6]. CLIP-T measures the alignment between the generated image and the textual prompt, while CLIP-I and DINO eval" + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.506, + 0.903, + 0.589 + ], + "angle": 0, + "content": "
MethodFace Rec.↑DINO↑CLIP-I↑CLIP-T↑
Hyper-DreamBooth0.6550.4730.5770.286
DreamBooth0.6180.4410.5460.282
Textual Inversion0.6230.2890.4720.277
Ours0.3250.6050.6390.268
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.6, + 0.906, + 0.669 + ], + "angle": 0, + "content": "Table 1. Quantitative results (CelebA HQ) - We report the standard metrics for our method and the baselines. Our method is able to offer comparable performance, with significantly less compute. In terms of the image alignment metrics, our method significantly outperforms other methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.699, + 0.906, + 0.76 + ], + "angle": 0, + "content": "uate the visual consistency of the generated image with the reference object's appearance. We also include the identity preservation metric (Face Rec.) suggested by HyperDream-Booth [40]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Compute time. As noted earlier, our method requires 1.5 days to train with 8 GPUs for the AFHQ and CelebA datasets, with around \\(14.6\\mathrm{k}\\) samples for AFHQ and \\(29.9\\mathrm{k}\\) samples for CelebA, amounting to 12 days of GPU time. This is much faster compared to the training of HyperDreamBooth [40] that requires 15k DreamBooth weights to be precomputed, which amounts to 50 days of GPU time, before their hypernetwork can start training. Thus, our method trains in less than a quarter of the pre-processing" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22134" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.089, + 0.485, + 0.166 + ], + "angle": 0, + "content": "
MethodDINO↑CLIP-I↑CLIP-T↑
Textual Inversion0.5960.7800.285
DreamBooth0.5600.7630.268
Ours0.6640.8070.277
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.177, + 0.485, + 0.261 + ], + "angle": 0, + "content": "Table 2. Quantitative results (AFHQ) - We outperform other baselines in terms of image alignment and provide comparable results in terms of CLIP-T, the prompt alignment. We note that our method outperforms DreamBooth [39] for all metrics, benefiting from the fact that our HyperNet Field learns a general trend of how DreamBooth weights behave." + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.275, + 0.473, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.481, + 0.484, + 0.553 + ], + "angle": 0, + "content": "Figure 5. Fast training (AFHQ) - We show example outcomes of our method only after 100 iterations of training, as raw output from our Hypernetwork without fast fine-tuning. As shown, even after only 100 iterations, our model is able to generate images that are visually very similar to the conditioning images." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.579, + 0.277, + 0.594 + ], + "angle": 0, + "content": "time of HyperDreamBooth." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.595, + 0.484, + 0.701 + ], + "angle": 0, + "content": "Since we query our Hypernetwork at a single optimization step \\( T \\), our method also yields efficient inference. Inference of our method takes only 0.3 seconds, compared to 2 seconds for HyperDreamBooth [40] and 5 minutes for DreamBooth. Additionally, the fast fine-tuning required by both our method and HyperDreamBooth takes just 20 seconds, allowing for quick adaptation with minimal overhead." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Results. We provide qualitative highlights in Figs. 3 and 4, and the quantitative summaries in Tables 1 and 2. Our results for image personalization in Figs. 3 and 4 is on par with the target Dreambooth model whose weights we estimate with our HyperNet Fields approach. That is, our approach generates personalized images that are representative of the given input prompt and the conditional image. Again, note that we achieve these personalization results in a single step by querying the hypernetwork once at timestep \\( T \\). Interestingly, our method demonstrates remarkable performance for preserving the original image context, as measured by DINO and CLIP-I scores (Tab. 1). Further, in the case of the AFHQ dataset (Tab. 2), we outperform" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.091, + 0.907, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.344, + 0.907, + 0.402 + ], + "angle": 0, + "content": "Figure 6. Qualitative examples (3D Shapes) – We show example shape reconstruction results converting (left) point clouds to 3D shapes and (right) images to 3D shapes. More qualitative examples are available in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.427, + 0.907, + 0.548 + ], + "angle": 0, + "content": "all DreamBooth variants, demonstrating that not only do we remove the compute-heavy optimization requirement of DreamBooth [39], but also by amortizing the behavior of the learned weights with respect to the input conditioning image, we successfully replicate the dynamics of the task-specific dreambooth network yielding improved results. We emphasize that while results are comparable to each other, our method is significantly faster than the baselines." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.552, + 0.909, + 0.704 + ], + "angle": 0, + "content": "Fast training. An interesting observation from our training process is that our model exhibits faster adaptation to personalization. As shown in Fig. 5, after only 100 training iterations, our model generates images that are visually similar to the conditioning images, achieving this level of personalization with minimal training time. We hypothesize that this rapid adaptation is due to our model's ability to simulate the optimization pathway, effectively learning the direction in which to adjust network weights to reduce the task-specific loss." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.713, + 0.851, + 0.73 + ], + "angle": 0, + "content": "4.2. Application to 3D shape reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We further explore the potential of our method for the task of 3D shape reconstruction, specifically the task of estimating the weights of an occupancy network [27]. Additionally, we consider two different conditioning inputs: a rendered image of the shape [24, 50], and a point cloud of the shape [1, 15, 32]. For the image conditioning, we use the 38 renderings provided in Qiu et al. [33]—covering various angles around each object—which we sample one of and encode with a pre-trained ViT [9]. For the point cloud conditioning, we randomly sample 2048 points uniformly from the surface of the ground truth mesh and encode the input point" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22135" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.905, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.206, + 0.907, + 0.292 + ], + "angle": 0, + "content": "Figure 7. Hypernet Field output example (ShapeNet) - We show the output of our HyperNet Field for various points along the optimization trajectory as well as the intersection over union (IoU) over all values of T. As shown, the HyperNet Field is able to accurately estimate the entire optimization trajectory, starting from network weights that only represent the airplane shape roughly, and converging to one that provides a detailed reconstruction of the airplane. The IoU plot also shows that as we traverse the trajectory, the IoU of the generated shapes to the ground truth shape improves, demonstrating convergence. Despite being shown only one ground truth optimization step at a time, our model is able to represent the entire optimization trajectory from initialization to convergence." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.316, + 0.462, + 0.332 + ], + "angle": 0, + "content": "cloud with an off-the-shelf variational autoencoder [48]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.337, + 0.484, + 0.489 + ], + "angle": 0, + "content": "Task network - Occupancy Net. For both cases, the hypernetwork field architecture remains the same as described in Sec. 3.4, and for the task-specific network, we follow an architecture similar to HyperDiffusion [12] —we use a simple 3-layer dense network [41] with ReLU activations and a hidden dimension of 128. The task-specific network, the occupancy network, takes the 3D coordinates of a query point as input and outputs a scalar value representing the occupancy as in Mescheder et al. [27]. We use marching cubes [26] to render the outcome." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.495, + 0.484, + 0.646 + ], + "angle": 0, + "content": "Training setup. To supervise the training, we sample 20k query points, 10k uniformly, and 10k near the ground-truth surface to be used for training. To find the ground-truth occupancy values for these points, we first make the meshes watertight with [19] and use winding numbers [4] on the watertight mesh to determine whether each point lies inside or outside the surface, and compute the occupancy via Barill et al. [3]. We use the mean squared error (MSE) of the occupancy estimates for the task-specific loss. For all cases, we set \\( T = 10,000 \\)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.652, + 0.484, + 0.789 + ], + "angle": 0, + "content": "2D image conditioning (Objaverse shapes). For 2D image conditioning, as training over the entire Objaverse dataset [8] would be computationally expensive, we simply overfit to a subset of 128 shapes, as we are mostly interested in the potential of our method. We leave training on the full Objaverse dataset as future work. We obtain reasonable shape reconstruction results as shown in Figures 1 and 6 with our method for this simple overfitting demonstration just within 10 hours of training on a single NVidia RTX 3090 GPU." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.795, + 0.484, + 0.901 + ], + "angle": 0, + "content": "3D point cloud conditioning (ShapeNet shapes). For 3D point cloud conditioning, we use the entire dataset. As shown in Fig. 6, our method is able to learn to generate occupancy network weights from point clouds. We further illustrate how both our Hypernet Field output and the IoU for shape generation evolves over the optimization process in Fig. 7. As shown, our method simulates the entire train-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.316, + 0.907, + 0.361 + ], + "angle": 0, + "content": "ing process of the task-specific network for a given point cloud despite only being shown a single ground truth gradient step at a time." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.378, + 0.634, + 0.393 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.404, + 0.907, + 0.6 + ], + "angle": 0, + "content": "We presented Hypernetwork Fields, a novel and efficient hypernetwork training framework. Instead of optimizing individual weights per sample, we use gradient supervision and match gradients along the hypernetworks field and the task-specific network at random points along the optimization trajectory. This drastically reduces the precomputation overhead of training hypernetworks, while maintaining comparable performance to state of the art. We have demonstrated the effectiveness of our method on personalized image generation and 3D shape reconstruction. Our framework is general and can be applied to any hypernetwork training scenario, making it a promising direction for scaling hypernetwork applications to larger datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.605, + 0.909, + 0.832 + ], + "angle": 0, + "content": "Limitations and future work. Our method enables stable hypernetwork training on large datasets. Because of this, the potential of our method for diverse tasks remains underexplored. This may include tasks such as training a hypernetwork to estimate the weights of NeRFs or even personalized adapters for large language models. However, our approach requires as many output neurons as there are weights in the target network, resulting in suboptimal memory usage. Recent work by Dravid et al. [10] has shown that customized DreamBooth adapters can be represented using only 1k parameters. Integrating our method with this compact representation approach could significantly enhance parameter efficiency while maintaining our proposed benefits of efficient hypernetwork training. These would be interesting avenues for future work." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.848, + 0.611, + 0.864 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.521, + 0.873, + 0.909, + 0.903 + ], + "angle": 0, + "content": "[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitlagkas, and Leonidas Guibas. Learning representations and generative" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22136" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.398, + 0.106 + ], + "angle": 0, + "content": "models for 3d point clouds. In ICML, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.108, + 0.483, + 0.161 + ], + "angle": 0, + "content": "[2] Sudarshan Babu, Richard Liu, Avery Zhou, Michael Maire, Greg Shakhnarovich, and Rana Hanocka. Hyperfields: Towards zero-shot generation of nerfs from text. In ICML. PMLR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.163, + 0.483, + 0.204 + ], + "angle": 0, + "content": "[3] Gavin Barill, Neil G Dickson, Ryan Schmidt, David IW Levin, and Alec Jacobson. Fast winding numbers for soups and clouds. Transactions On Graphics, 37(4):1-12, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.205, + 0.482, + 0.245 + ], + "angle": 0, + "content": "[4] Gavin Barill, Neil G. Dickson, Ryan M. Schmidt, David I. W. Levin, and Alec Jacobson. Fast winding numbers for soups and clouds. Transactions On Graphics, 37:1-12, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.247, + 0.482, + 0.273 + ], + "angle": 0, + "content": "[5] Tom B Brown. Language models are few-shot learners. NeurIPS, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.275, + 0.482, + 0.328 + ], + "angle": 0, + "content": "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, pages 9650-9660, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.33, + 0.482, + 0.37 + ], + "angle": 0, + "content": "[7] Yunjoy Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis for multiple domains. In CVPR, 2020. 5, 6, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.373, + 0.482, + 0.439 + ], + "angle": 0, + "content": "[8] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In CVPR, pages 13142-13153, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.442, + 0.482, + 0.469 + ], + "angle": 0, + "content": "[9] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.47, + 0.482, + 0.524 + ], + "angle": 0, + "content": "[10] Amir Dravid, Yossi Gandelsman, Kuan-Chieh Wang, Rameen Abdal, Gordon Wetzstein, Alexei A Efros, and Kfir Aberman. Interpreting the weight space of customized diffusion models. arXiv preprint arXiv:2406.09413, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.526, + 0.482, + 0.552 + ], + "angle": 0, + "content": "[11] Yilun Du and Igor Mordatch. Implicit generation and modeling with energy based models. NeurIPS, 32, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.554, + 0.482, + 0.607 + ], + "angle": 0, + "content": "[12] Ziya Erkoç, Fangchang Ma, Qi Shan, Matthias Nießner, and Angela Dai. Hyperdiffusion: Generating implicit neural fields with weight-space diffusion. In ICCV, pages 14300-14310, 2023. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.609, + 0.482, + 0.663 + ], + "angle": 0, + "content": "[13] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. *ICLR*, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.665, + 0.482, + 0.718 + ], + "angle": 0, + "content": "[14] Will Grathwohl, Kuan-Chieh Wang, Jorn-Henrik Jacobsen, David Duvenaud, Kevin Swersky, and Mohammad Norouzi. Learning energy-based models by diffusion recovery likelihood. ICLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.721, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[15] Thibault Groueix, Matthew Fisher, Vladimir G Kim, Bryan Russell, and Mathieu Aubry. Atlasnet: A papier-mâché approach to learning 3d surface generation. In CVPR, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.776, + 0.482, + 0.802 + ], + "angle": 0, + "content": "[16] David Ha, Andrew Dai, and Quoc V Le. Hypernetworks. ICLR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.805, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, pages 6840-6851, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[18] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In ICLR, 2022. 6" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[19] Jingwei Huang, Yichao Zhou, and Leonidas Guibas. Manifoldplus: A robust and scalable watertight manifold surface generation method for triangle soups. arxiv, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[20] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. In ICLR, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.179, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[21] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. NeurIPS, 35:26565-26577, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.222, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[22] Yann LeCun, Sumit Chopra, Raia Hadsell, Marc'Aurelio Ranzato, and Fu Jie Huang. A tutorial on energy-based learning. Predicting structured data, 1(2):1-31, 2006. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.264, + 0.905, + 0.291 + ], + "angle": 0, + "content": "[23] Yaron Lipman, Justin Solomon, et al. Flow matching for generative modeling. *ICLR*, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.294, + 0.907, + 0.361 + ], + "angle": 0, + "content": "[24] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9298–9309, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.364, + 0.905, + 0.404 + ], + "angle": 0, + "content": "[25] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In ICCV, 2015. 5, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.407, + 0.905, + 0.46 + ], + "angle": 0, + "content": "[26] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. In Semin al graphics: pioneering efforts that shaped the field, pages 347-353, 1998. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.463, + 0.905, + 0.517 + ], + "angle": 0, + "content": "[27] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In CVPR, pages 4460-4470, 2019. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.52, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[28] Jonathan Oswald, Christian Henning, Tomas Pfister, and Jürgen Schmidhuber. Continual learning with hypernetworks. In ICLR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.563, + 0.905, + 0.617 + ], + "angle": 0, + "content": "[29] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In CVPR, pages 165-174, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.62, + 0.905, + 0.646 + ], + "angle": 0, + "content": "[30] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, pages 4195-4205, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.649, + 0.905, + 0.703 + ], + "angle": 0, + "content": "[31] Marcin Przewiezlikowski, Przemyslaw Przybysz, Jacek Tabor, Maciej Zikeba, and Przemyslaw Spurek. Hypermaml: Few-shot adaptation of deep models with hypernetworks. Neurocomputing, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.705, + 0.905, + 0.746 + ], + "angle": 0, + "content": "[32] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.748, + 0.905, + 0.816 + ], + "angle": 0, + "content": "[33] Lingteng Qiu, Guanying Chen, Xiaodong Gu, Qi Zuo, Mutian Xu, Yushuang Wu, Weihao Yuan, Zilong Dong, Liefeng Bo, and Xiaoguang Han. Richdreamer: A generalizable normal-depth diffusion model for detail richness in text-to-3d. In CVPR, pages 9914–9925, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.818, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22137" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.146 + ], + "angle": 0, + "content": "[35] Maziar Raissi, Paris Perdikaris, and George Em Karniadakis. Physics informed deep learning (part i): Data-driven solutions of nonlinear partial differential equations. arxiv, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.19 + ], + "angle": 0, + "content": "[36] Maziar Raissi, Paris Perdikaris, and George Em Karniadakis. Physics informed deep learning (part ii): Data-driven discovery of nonlinear partial differential equations. arxiv, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[37] Maziar Raissi, Paris Perdikaris, and George Em Karniadakis. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational Physics, 378:686-707, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.303 + ], + "angle": 0, + "content": "[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.305, + 0.482, + 0.374 + ], + "angle": 0, + "content": "[39] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, pages 22500-22510, 2023. 1, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.375, + 0.482, + 0.444 + ], + "angle": 0, + "content": "[40] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdream booth: Hypernetworks for fast personalization of text-to-image models. In CVPR, pages 6527-6536, 2024. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.446, + 0.482, + 0.487 + ], + "angle": 0, + "content": "[41] David E Rumelhart, Geoffrey E Hinton, and Ronald J Williams. Learning representations by back-propagating errors. NATURE, 323(6088):533-536, 1986. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.489, + 0.482, + 0.558 + ], + "angle": 0, + "content": "[42] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 35:36479-36494, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.56, + 0.482, + 0.613 + ], + "angle": 0, + "content": "[43] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, pages 2256-2265, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.616, + 0.482, + 0.655 + ], + "angle": 0, + "content": "[44] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. NeurIPS, 32, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.659, + 0.482, + 0.7 + ], + "angle": 0, + "content": "[45] Yang Song, Conor Durkan, Iain Murray, and Stefano Ermon. Maximum likelihood training of score-based diffusion models. NeurIPS, pages 1415-1428, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.702, + 0.482, + 0.729 + ], + "angle": 0, + "content": "[46] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. arxiv, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.731, + 0.482, + 0.771 + ], + "angle": 0, + "content": "[47] Pascal Vincent. A connection between score matching and denoising autoencoders. Neural Computation, 23(7):1661-1674, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.774, + 0.482, + 0.828 + ], + "angle": 0, + "content": "[48] Guandao Yang, Xun Huang, Zekun Hao, Ming-Yu Liu, Serge Belongie, and Bharath Hariharan. Pointflow: 3d point cloud generation with continuous normalizing flows. In ICCV, pages 4541-4550, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.83, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[49] Linnan Zhang, Zhiding Ren, Seunghyun Ahn, Hakan Bilen, and Kwang In Lee. Graph hypernetworks for neural architecture search. *ICLR*, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.872, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[50] Jun-Yan Zhu et al. Shap-e: Generating conditional 3d implicit functions. arxiv, 2023. 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22138" + } + ] +] \ No newline at end of file diff --git a/2025/HyperNet Fields_ Efficiently Training Hypernetworks without Ground Truth by Learning Weight Trajectories/68fd503f-6b57-47fa-8aca-593407ee13b9_origin.pdf b/2025/HyperNet Fields_ Efficiently Training Hypernetworks without Ground Truth by Learning Weight Trajectories/68fd503f-6b57-47fa-8aca-593407ee13b9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9af52ed95b5bff02615589eae3bcdb89dd6f3155 --- /dev/null +++ b/2025/HyperNet Fields_ Efficiently Training Hypernetworks without Ground Truth by Learning Weight Trajectories/68fd503f-6b57-47fa-8aca-593407ee13b9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bd52683f99ed5669d7642e1fcc8f699a2b5a87fad362f941904851b091b86a68 +size 7099881 diff --git a/2025/HyperNet Fields_ Efficiently Training Hypernetworks without Ground Truth by Learning Weight Trajectories/full.md b/2025/HyperNet Fields_ Efficiently Training Hypernetworks without Ground Truth by Learning Weight Trajectories/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d2100a48d4157cccf4475d1a8b7f1021178ee6eb --- /dev/null +++ b/2025/HyperNet Fields_ Efficiently Training Hypernetworks without Ground Truth by Learning Weight Trajectories/full.md @@ -0,0 +1,339 @@ +# HyperNet Fields: Efficiently Training Hypernetworks without Ground Truth by Learning Weight Trajectories + +Eric Hedlin $^{1,2*}$ Munawar Hayat $^{2}$ Fatih Porikli $^{2}$ Kwang Moo Yi $^{1}$ Shweta Mahajan $^{2}$ $^{1}$ University of British Columbia ${}^{2}$ Qualcomm AI Research† + +![](images/c4f22adc05793672a7e7332c344006a515b952740da570723acf1df772c116ed.jpg) +Figure 1. Teaser - We propose hypernetwork fields, where we learn the entire weight space trajectory instead of only the final state, which allows us to train without ever needing to know the final converged weights. Our method can be applied to any application of hypernetworks, including diffusion model personalization and modeling of 3D neural representations. + +![](images/91f43ea1bd60ca683d11dc4a81ff813e1a575285286666305481408c8df0b3ec.jpg) + +# Abstract + +To efficiently adapt large models or to train generative models of neural representations, Hypernetworks have drawn interest. While hypernetworks work well, training them is cumbersome, and often requires ground truth optimized weights for each sample. However, obtaining each of these weights is a training problem of its own—one needs to train, e.g., adaptation weights or even an entire neural field for hypernetworks to regress to. In this work, we propose a method to train hypernetworks, without the need for any per-sample ground truth. Our key idea is to learn a Hypernetwork 'Field' and estimate the entire trajectory of network weight training instead of simply its converged state. In other words, we introduce an additional input to the Hypernetwork, the convergence state, which then makes it act as a neural field that models the entire convergence pathway of a task network. A critical benefit in doing so is that the gradient of the estimated weights at any convergence state must then match the gradients of the original task—this constraint alone is sufficient to train the Hypernetwork Field. + +We demonstrate the effectiveness of our method through the task of personalized image generation and 3D shape reconstruction from images and point clouds, demonstrating competitive results without any per-sample ground truth. + +# 1. Introduction + +Various generative modeling tasks [39] require adaptation, thus data-dependent network weights, or in other words, model parameters being adjusted according to various circumstances (conditions and user input). These include personalized text-to-image generation where the weights of the underlying personalization model, such as DreamBooth [39], are fine-tuned on the concept of interest and used to generate target images conditioned on input prompts. Another example is 3D modeling, where neural + +networks are used to model shapes implicitly with signed distance fields [29] or occupancy [27]. In both cases, a dedicated set of neural network weights is required for each personalization outcome or 3D shape. Thus, the applicability of these methods is highly limited by the training time required to obtain these weights. + +It is therefore critical to seek ways to amortize the per-sample optimizations into Hypernetworks [12, 40]—networks that estimate network parameters. By 'skipping ahead' the tedious training processes involved in fine-tuning or shape regressing, they greatly speed up the entire process. Existing methods, however, have a critical shortcoming in that they supervise hypernetworks by directly matching their output to the precomputed, ground truth converged weights corresponding to the task-specific network for each sample in the dataset. Notably, prior to any training of the Hypernetwork, one must prepare converged task-specific network weights for every sample in the training set. This severely limits the capabilities of Hypernetwork methods when scaling to large datasets. As an example, to train HyperDreamBooth [40], it requires 50 days of GPU time on a NVidia RTX 3090, $^{\ddagger}$ to just prepare the training data alone. + +Furthermore, current methods assume a one-to-one mapping between each input sample and its corresponding optimized weights. This assumption is an oversimplification. Neural network optimization is stochastic with various equally plausible solutions. Therefore, enforcing a bijective mapping between the condition and the ultimately converged weights may limit the potential expressiveness of the hypernetwork, as we will demonstrate in Sec. 3. + +To address these limitations, we instead propose to train a Hypernetwork Field that estimates the entire optimization trajectory of a neural network training, rather than focusing on a single final set of weights for each sample. Specifically, we introduce the convergence state as additional input to the hypernetwork, such that the weights of the task-specific network at a desired state of convergence can be 'queried'. An interesting outcome of doing so is that the gradient of the estimated weights along this pathway must match the gradients of the original task-specific network, as they are both gradients pointing towards convergence. Hence, training of such hypernetwork fields can simply be done by matching the two gradients. + +By supervising the hypernetwork to match the gradient of the weights with respect to the optimization step, we avoid the need for precomputing target weights. This supervision strategy allows the hypernetwork to estimate a trajectory of parameters that, at each step, reflects a compatible state across all samples in the dataset. Ultimately, at inference time, the estimated parameters which correspond with the final timestep of the hypernetwork, represent a well + +converged solution for each of the samples, but are now naturally discovered through training of the hypernetwork field itself. Notably, this single forward pass to estimating the converged parameters is all that is necessary for inference, hence no additional compute is required compared to conventional hypernetwork approaches [12, 40]. + +Our framework is general and is applicable to all hypernetwork training scenarios. To demonstrate the effectiveness of our method, we apply our method to personalized image generation and 3D shape generation. We achieve competitive performance while reducing the training cost by $\sim 4$ times what would be required of conventional hypernetwork training. + +To summarize, our contributions are: + +- We propose hypernetwork fields, an approach to learn the entire optimization pathway of task-specific networks, which allows ground truth-free training without the need for precomputing the target weights; +- our approach drastically reduces the compute required for training hypernetworks; +- we validate our approach on two very different tasks and datasets—for personalized image generation we achieve results comparable to the state of the art with significantly less compute; and +- we further explore the application of our method to 3D shape reconstruction from images an point clouds. + +# 2. Related Work + +Hypernetworks. Early work on training hypernetworks backpropagates gradients through the task-specific network to the hypernetwork itself, as demonstrated by approaches [16, 31] and subsequent explorations in continual learning [28] and neural architecture search [49]. These methods have shown that allowing gradient flow through both networks enables the hypernetwork to modify parameters for diverse tasks adaptively. However, more recent works have introduced approaches that involve precomputing weights for the task-specific network and directly supervise the hypernetwork to match them, eliminating the need for backpropagation through the task-specific network and improving training stability by alleviating issues with gradient flow. This supervision with precomputed weights has been shown to enhance gradient stability and training performance, as highlighted in frameworks like HyperDiffusion [12], HyperFields [2], and HyperDreamBooth [40]. However, this limits the size of the dataset to which this can be applied, as the precompute requirement is proportional to the number of samples in the dataset. Our method combines the best of both approaches by removing the need for precomputed weights while explicitly providing weight-based supervision in the form of gradients along the convergence path, enabling stable training on large datasets. + +Denoising Models. In generative modeling, finding smooth and reliable paths through latent space has become central to producing coherent samples. Examples include diffusion models [17, 43], which iteratively trace a path from noise to data through successive steps, effectively navigating high-dimensional space by reversing a stochastic process. Consistency models [46] refine this by learning non-Markovian trajectories that reach target data distributions in fewer steps by introducing a formulation that allows them to single-step jump to any timestep. Flow-matching models [23] take a similar path-based approach by learning a vector field that guides samples along pre-defined flows, ensuring a direct and smooth transition through the latent space. Similarly to these methods, our method can be thought of as finding a path through this high-dimensional space, specifically in the form of a single-step estimator from initialization to any desired state of convergence, whose path travels strictly along the direction of the steepest gradient descent. + +Methods with gradient-supervision. Several recent approaches in generative modeling and physics-informed learning employ supervision exclusively through gradients rather than direct supervision between input-output pairs to achieve denoising and reconstruction. Score-based generative models [44, 45, 47] use score matching to learn the gradient (or "score") of the log density of noisy data distributions, bypassing the need for direct output supervision. Denoising Score Matching (DSM) [47] trains the model to predict this gradient, guiding it to approximate data distributions through iterative denoising and demonstrating improved stability in noise removal. Similarly, Physics-Informed Neural Networks (PINNs) [35-37] leverage known derivatives from partial differential equations (PDEs) in physical simulations, such as heat diffusion or fluid flow, to inform the network solely through spatial or temporal derivatives. This derivative-based guidance encourages convergence as the model adapts to match observed physical behaviour, requiring no explicit target outputs. Energy-Based Models (EBMs) [11, 14, 22] also align with this gradient-driven approach by learning energy landscapes that adjust gradients to represent target distributions, allowing for smooth gradient matching through learned energy fields. Inspired by these methods, our approach supervises only the gradient direction in task-specific weight space, effectively learning a convergence pathway without requiring explicit final state supervision. By supervising gradients within this pathway, our method achieves robust training stability while generalizing over diverse tasks without per-sample ground truth weights. + +# 3. Method + +To remove the precompute requirements associated with training hypernetworks, we present a new framework that + +models the convergence dynamics of task-specific networks into a single, efficient process, which we name Hypernetwork Field. + +Problem formulation. Formally, let $f_{\theta}$ denote a task-specific neural network, e.g., dreambooth [39], parameterized by $\theta$ . $f_{\theta}$ is then typically trained to minimize a loss function $\mathbb{E}_{\mathbf{x}}[\mathcal{L}(\theta ,\mathbf{x})]$ , by training for $T$ optimization steps given a dataset of conditions, e.g., image $\mathbf{x}$ . In our work, we wish to learn a hypernetwork field $H_{\phi}(t,\mathbf{x})$ , parameterized by $\phi$ that takes as input a condition $\mathbf{x}$ and an optimization step $t$ . The hypernetwork field then predicts the parameters $\theta_t = H_\phi (t,\mathbf{x})$ at optimization step $t$ of the task-specific network $f_{\theta_t}(\mathbf{x})$ , given the condition $\mathbf{x}$ and the timestep $t$ . In the following, we show how to obtain $H_{\phi}(t,\mathbf{x})$ without requiring any $\theta_t$ to supervise the training. + +# 3.1. HyperNet Fields + +To train the hypernet field $H_{\phi}(t,\mathbf{x})$ without the ground-truth targets for it to regress, we take a step back and revisit what standard hypernetworks are optimizing for. Standard approaches for learning hypernetworks [12, 40] utilize the ground-truth weights $\theta^{*}$ of the task-specific network, corresponding to the optimized parameters for a specific sample $\mathbf{x}$ as a supervision signal. In other words, they first find + +$$ +\theta^ {*} (\mathbf {x}) = \underset {\theta} {\arg \min } \mathcal {L} (\theta , \mathbf {x}), \tag {1} +$$ + +for each $\mathbf{x}$ , then find $\phi$ , such that $\theta^{*}(\mathbf{x}) \approx H_{\phi}(\cdot, \mathbf{x}), \forall \mathbf{x}$ . This typically requires several iterations of the optimization process and thus incurs a large computational cost for generating the target parameters $\theta^{*}(\mathbf{x})$ . Moreover, since this process has to be repeated for each sample independently, the scalability to large datasets is severely limited. + +Note, however, Eq. (1) is typically found through gradient-based optimization, and forms a trajectory of weight parameter space in the $\{\theta_t\}_{t = 1\dots T}$ , where $T$ is the number of optimization steps. If we now consider our hypernet field $H_{\phi}(t,\mathbf{x})$ that we intend to model, we notice an interesting relationship that must hold between the $\mathcal{L}(\theta_t,\mathbf{x})$ and $H_{\phi}(t,\mathbf{x})$ . Specifically, consider the gradient of the task-specific weights at an optimization step $t$ , $\nabla_{\theta_t}\mathcal{L}(\theta_t,\mathbf{x})$ . Then, if a hypernet field $H_{\phi}$ has been properly trained, the direction in which $\nabla_{\theta_t}\mathcal{L}(\theta_t,\mathbf{x})$ point to must match the direction in which the hypernet trajectory flows, that is, $\nabla_tH_\phi (t,\mathbf{x})$ . In other words, their gradients must match. + +In fact, matching the two gradients allows us to circumvent the necessity of per-sample optimization as $\nabla_{\theta_t}\mathcal{L}(\theta_t,\mathbf{x})$ can be computed with relatively minimal overhead by using the current hypernet field estimate for $\theta_t$ . This process can also be understood as amortizing Eq. (1) into the training process of the hypernet field $H_{\phi}(t,\mathbf{x})$ , as we are 'skipping ahead' parts of the optimization loop of Eq. (1) with the estimates coming from $H_{\phi}(t,\mathbf{x})$ . Through + +![](images/868d0df2f2a67d6a8b042c2f4131997dd8911b70b965b1e961db9efb123e9367.jpg) +Figure 2. The Hypenet Fields Framework. Left: Our Hypernetwork with gradient-based supervision from a task-specific network - dreambooth [39]. Right: The sampling process to generate personalized images with our hypernetwork framework. + +![](images/5327c2512391025e71e269920022c476cc10f1d2ba8bc520d4a5b78a280787e3.jpg) + +this amortization, the hypernet field captures detailed information about how the weights evolve throughout training, allowing it to effectively predict and generate weights across the training samples. We now further detail the exact training process. + +# 3.2. Gradient-based Supervision + +In each training iteration of the optimization process, the hypernet field generates an estimate of the task-specific parameters $\hat{\theta}_t = H_\phi(t, \mathbf{x})$ at timestep $t$ based on the input $\mathbf{x}$ from a dataset $\mathcal{D}$ and the timestep $t \sim U(0, T)$ . Given this estimate, we compute the gradient of the loss function with respect to the task-specific weights at that timestep, $\nabla_{\hat{\theta}_t} \mathcal{L}(\hat{\theta}_t, \mathbf{x})$ , and perform a single optimization step to update the weights to $\theta_{t+1}$ . This is given by + +$$ +\theta_ {t + 1} = \hat {\theta} _ {t} - \eta \nabla_ {\hat {\theta} _ {t}} \mathcal {L} _ {\text {t a s k}} (\hat {\theta} _ {t}, \mathbf {x}), \tag {2} +$$ + +where $\eta$ is the learning rate. Then, the hypernet field must also follow this update by definition, thus we wish to have $\hat{\theta}_{t + 1}\approx \theta_{t + 1}$ . Moreover, assuming $\theta_0 = \hat{\theta}_0 = H_\phi (0,\mathbf{x})$ , the entire trajectory can simply be matched via minimizing + +$$ +\mathcal {L} _ {\Delta} (\mathcal {D}) = \mathbb {E} _ {\mathbf {x} \sim \mathcal {D}, t \sim U [ 0, T ]} \| \theta_ {t + 1} - H _ {\phi} (\mathbf {x}, t + 1) \| ^ {2}, \tag {3} +$$ + +which we name as the gradient matching loss. This loss function essentially minimizes the difference between the gradient step from Eq. (2) and the numerical gradient $\nabla_t H_\phi(\mathbf{x}, t) = H_\phi(\mathbf{x}, t + 1) - H_\phi(\mathbf{x}, t)$ over the entire dataset for $T$ steps of optimization; thus replicating the underlying training dynamics of the task-specific network. + +Once trained, the hypernetwork field $H_{\phi}(t,\mathbf{x})$ thus mimics the entire training trajectory of the task-specific network, instead of the point estimate $\theta^{*}(\mathbf{x})$ . We summarize this training process in Fig. 2. + +Ensuring a consistent trajectory throughout training. To ensure training stability, we fix the starting position $\theta_0$ for each sample $\mathbf{x}$ throughout training, initializing it randomly only once at the beginning. This approach mirrors the typical neural network initialization to random values, followed by convergence toward the optimal parameters. + +Forcing 0 offset at $t = 0$ by construction. Furthermore, Eq. (3) matches the entire trajectory only when $\theta_0 = \hat{\theta}_0 =$ + +# Algorithm 1 Hypernet Fields Training Algorithm + +1: Initialize $\theta_0$ +2: $H_{\phi}\gets$ Hypernetwork() $\triangleright$ Hypernetwork with parameters $\phi$ +3: while not converged do +4: $\mathbf{x} \gets \mathrm{next}(\mathcal{D}) \triangleright$ Sample condition $\mathbf{x}$ from dataset +5: $t \gets$ sample from $[0, T)$ +6: $\hat{\theta}_t\gets H_\phi (\mathbf{x},t)$ $\triangleright$ Estimate weights for timestep $t$ +7: $\Delta \theta_{t} \gets -\eta \nabla_{\theta_{t}} \mathcal{L}_{task}(\hat{\theta}_{t}, \mathbf{x})$ $\triangleright$ GT gradient for $\hat{\theta}_{t}$ +8: $\Delta \hat{\theta}_t\gets H_\phi (\mathbf{x},t + 1) - \hat{\theta}_t$ $\triangleright$ Est gradient for $\hat{\theta}_t$ +9: $\phi \gets \phi -\eta \nabla_{\phi}\mathrm{MSE}(\Delta \theta_t,\Delta \hat{\theta}_t)$ Update $\phi$ +10: end while + +$H_{\phi}(0,\mathbf{x})$ . Enforcing this, however, can easily be done by construction. Inspired by [21, 46], for any timestep $t$ , we parameterize the model with an offset from the input $\theta_0$ , conditioned on $t$ as, + +$$ +H _ {\phi} (\mathbf {x}, t) = \theta_ {0} + \frac {t}{T} \times H _ {\phi} ^ {\prime} (\mathbf {x}, t), \tag {4} +$$ + +where $H'$ is the raw hypernetwork field output from a neural network. + +# 3.3. Algorithm + +We summarize our overall optimization process in Algorithm 1. We first randomly initialize and fix the parameters $\theta_0$ . We then consider a random sample $\mathbf{x}$ from the dataset and a random timestep $t$ from the optimization trajectory. With these inputs, we use the hypernetwork $H_{\phi}$ to obtain the weights $\hat{\theta}_t$ of the task-specific network. Next, using these weights and the input $\mathbf{x}$ we compute the gradient step $\Delta \theta_t$ for the task-specific network. Additionally, we calculate $\Delta \hat{\theta}_t = \hat{\theta}_{t + 1} - \hat{\theta}_t$ from the hypernetwork. The parameters $\phi$ of the hypernetwork are then updated to match $\Delta \hat{\theta}_t$ with the gradients $\Delta \theta_t$ , which is equivalent to minimizing Eq. (3). This yields a computationally efficient framework where a single hypernetwork encodes the optimization trajectory and can be queried to get the weights of the task-specific network at any stage in the training process. + +For inference, consider DreamBooth [39] as a task-specific network (Fig. 2). To generate personalized images + +![](images/bf6635a084243a7e83885b702e461a6136b646190890aab57da0c33a4ff58fb6.jpg) +Input Image + +![](images/bc84799f410f4b4d5c0fafc035c30dbe39c926db69a31dff0990bf6b97f3c245.jpg) +As a drawing + +![](images/3528cc8e9e422dc4f24ac3d126460458a8db90771f530c8ddaad2a1431313b1f.jpg) +As a pixar character + +![](images/faa5ad6e1a00a651d2723c148b8928493e5aabc8d7fa72d0fcd1f165ca819c94.jpg) +As a graffiti mural + +![](images/42dee81f7839c0a8412a0c19424cfd12df7bd0fa67f1fac509f26c8f52290eaf.jpg) +Flowing river of lost souls + +![](images/470027181c8788def4962a2cb500abaecdf08e3b0065ecb81162ccd62031193a.jpg) +As a funkopo pop figure +sno + +![](images/bae4e155d601fde33a7b29b195a91da186fa1ad9bdacfa92f010f860b3c140c3.jpg) + +![](images/1acee2229fa132d679380b66f722ceadf163a3e06fecbe1798566c94e048a9e4.jpg) + +![](images/4c656a8478eac94baea4a5f46a8ec68b5644228a82dd1f1edda604b9979e8a3d.jpg) + +![](images/3c8c531c0d9339fec94b4eaae0a4d62e05a79fe2b46cb15ecc6325dbbaee5298.jpg) + +![](images/b6a744410b5976439a5c4a79353521ee8dbf4f8a76c6fc2cc4329a1a326533f5.jpg) +Textual +Inversion + +![](images/180ff43a4471cf43ea816e1243b77b932d198909a60a6b4e64725978f9724516.jpg) +Figure 3. Qualitative examples (CelebA HQ) - Qualitative examples of personalized human face generation using the CelebA HQ dataset [20, 25] are shown. Our hypernetwork field achieves fast adaptation, producing results comparable to DreamBooth [39] and Textual Inversion [13] while preserving individual features in a personalized manner. Additional examples are provided in the supplementary material. + +![](images/160c7b767ca3e9d702cb1781bc89b5da22bbb54e8474e4dffb4371f6edecdf5f.jpg) + +![](images/9981e2fe21724ccb90b9229b5e656c0c9264a035f0feae44454f90f50f940834.jpg) + +![](images/ac323e7721d9bfeee374cb1fd5f5d1aa2f780b171239edbf62eebfcd6aceb659.jpg) + +![](images/967bdd4ba0b1e0d349229b9f66fea6dd6cf970614cf34580ff19e5c15b7dc857.jpg) +DreamBooth + +for a given source image, we query the hypernetwork field with the source image and the timestep $T$ to obtain the weights for the DreamBooth model. Using these weights and the conditioning prompt, we can generate personalized images in a single step. + +# 3.4. Implementation Details + +We use a frozen Vision Transformer (ViT) [9] to encode the images. We then pass this encoding into a hypernetwork with a 6-layer Diffusion Transformer (DiT) [30] with adaptive layer norm conditioning [5] for both the timestep and the image embedding. We set the sequence length of the DiT is equal to the total number of layers in the task-specific network. We then have a dedicated Multi-Layer Perceptron (MLP) per each layer of the task-specific network that decodes to the network weights of the target network. + +# 4. Experiments + +We consider two diverse applications - image personalization and 3D shape reconstruction as target applications. While we focus on these two, we emphasize that our hypernet fields framework does not make any assumption on the task-specific network and applies to any task. Moreover, the definition of the loss, $L_{task}(\theta_t, c)$ is not restricted to a particular domain. Thus, our method can be applied to these problems by modifying the task-specific optimization model. + +# 4.1. Application to Personalized Image Generation + +To apply our method to personalized image generation, we follow prior work [40] and evaluate and compare personalized face generation on CelebA HQ [20, 25], as well as on animal faces with the AFHQ dataset [7]. + +Task network - DreamBooth. To personalize and generate customized images we use DreamBooth as our task network, as in HyperDreamBooth [40]. DreamBooth builds upon the denoising diffusion probabilistic models [38, 42]. The generation process is guided by a special conditioning token encoding the characteristics of the input conditioning image, which is found by optimizing through diffusion models, which require a non-trivial amount of compute. On the CelebA dataset [25] we follow prior work [40] and use the text condition "a [V] face" where the token [V] represents a placeholder for the specific identity being learned. Analogously, for the AFHQ dataset [7] with animal images, we use the text condition "a [V] animal" to guide the model. The conditioning token along with the input prompts are then used to synthesize personalized images. + +Training setup. For the CelebA-HQ dataset [20, 25], we follow prior work [40] and replicate their setup to make results comparable. Specifically, among the $30\mathrm{k}$ high-quality aligned faces, we use the train/test splits and evaluation prompts from HyperDreamBooth [40], with $29.9\mathrm{k}$ images in the training set and 100 images in the test set. For the AFHQ dataset [7] we use the live subject prompts from the + +![](images/d37fc23d057773e660e30ad2344787d7e0a70fa6271e93b2c7605ea59110ab43.jpg) +Input Image + +![](images/740ea11955a3f836234c18ec8363fbac2ba1e63670fe667ba45deddfea599964.jpg) +Wearing a Santa hat + +![](images/0f1ddb8fb62e9660f1fdfb3030a5f5d1fbbb1334c2496fae8e7b5c61c4cc0854.jpg) + +![](images/8d54b4a78a62ab2973794d8cae98b2258adf95f0fee00522a98944f5a4d09a1e.jpg) +Figure 4. Qualitative examples (AHFQ) - Qualitative results of personalized animal image generation from the AFHQ dataset [7] demonstrate the adaptability of our hypernetwork field. Compared to DreamBooth [39] and Textual Inversion [13], our method effectively captures the specific animal's characteristics while enabling rapid adaptation. Further examples are available in the supplementary material. + +![](images/02cf65aad15da8f3f5d0c4a482928faa7d3097d0cf57ad6655789991bad1e7f4.jpg) +Wearing pink sunglasses + +![](images/1093262a36b2a1323f7b7ea4802c725a183eff9e527e079fae21558d1d101a22.jpg) + +![](images/726c0f2dbfb85a2bf059a428c7989e141812f35db840385e31f05520f2cad163.jpg) + +![](images/3e00a45e0053026bfb32743865769cbc566e3b94189fe707b9750685f87c37c2.jpg) +Wearing a rainbow scarf + +![](images/b2d30e095c7807e0a39bd222b9aa1e31095eacbc11194ff991e3b7da35d0f7fa.jpg) + +![](images/85f362426127572ed2ac8dbcbbc8042e062dbbd5be08c8067fc8e2e010fce0bb.jpg) + +![](images/098c1d4258d257a247900c6a2c80493ba1ab6c2fdf76cbfbd318e5e34e05c985.jpg) +In the snow +o + +![](images/12e0a3ecbac55668b920bdbe44f33d480c877bbc73be130dcdf31873b634ac29.jpg) +Textual +inversion + +![](images/bed2e2b4c033a6f67d2971ddbf5c8dfa59f7277e581422f50f162e28c744a834.jpg) + +DreamBooth dataset [39] so that the results are comparable. Among the 16.1k aligned animal faces, we use the established train and test splits, leaving 14.6k images for training and 1.5k for testing. In addition to the DreamBooth and HyperDreamBooth approach, we compare against Textual Inversion [13]. + +For both experiments, we set the number of optimization steps to be $T = 500$ . Our network is trained to output rank-1 Low-Rank Adaptation (LoRA) [18] weights for both the UNet and the text encoder for Stable Diffusion 1.5 [38], which amounts to 273k parameters across 352 layers. We train our model with a batch size of 256 and for 20k iterations on 8 NVidia A100 GPUs, which takes 1.5 days. Similar to the HyperDreamBooth [40] approach once our model is trained, we can perform 50 steps of fast fine-tuning starting from the hypernetwork outputs. + +Note that, unlike existing works that require per-instance fitting, with our method, it is possible to introduce on-the-fly data augmentation. We thus incorporate color, saturation, contrast, and hue augmentations to perturb the appearance of the conditioning signal $\mathbf{x}$ , including horizontal flips. + +Evaluation metrics. We use the standard metrics suggested by DreamBooth [40]: CLIP-T, CLIP-I [34], and DINO [6]. CLIP-T measures the alignment between the generated image and the textual prompt, while CLIP-I and DINO eval + +
MethodFace Rec.↑DINO↑CLIP-I↑CLIP-T↑
Hyper-DreamBooth0.6550.4730.5770.286
DreamBooth0.6180.4410.5460.282
Textual Inversion0.6230.2890.4720.277
Ours0.3250.6050.6390.268
+ +Table 1. Quantitative results (CelebA HQ) - We report the standard metrics for our method and the baselines. Our method is able to offer comparable performance, with significantly less compute. In terms of the image alignment metrics, our method significantly outperforms other methods. + +uate the visual consistency of the generated image with the reference object's appearance. We also include the identity preservation metric (Face Rec.) suggested by HyperDream-Booth [40]. + +Compute time. As noted earlier, our method requires 1.5 days to train with 8 GPUs for the AFHQ and CelebA datasets, with around $14.6\mathrm{k}$ samples for AFHQ and $29.9\mathrm{k}$ samples for CelebA, amounting to 12 days of GPU time. This is much faster compared to the training of HyperDreamBooth [40] that requires 15k DreamBooth weights to be precomputed, which amounts to 50 days of GPU time, before their hypernetwork can start training. Thus, our method trains in less than a quarter of the pre-processing + +
MethodDINO↑CLIP-I↑CLIP-T↑
Textual Inversion0.5960.7800.285
DreamBooth0.5600.7630.268
Ours0.6640.8070.277
+ +Table 2. Quantitative results (AFHQ) - We outperform other baselines in terms of image alignment and provide comparable results in terms of CLIP-T, the prompt alignment. We note that our method outperforms DreamBooth [39] for all metrics, benefiting from the fact that our HyperNet Field learns a general trend of how DreamBooth weights behave. + +![](images/51dc2064d770251ca7022afc7ca6a59b081c97cd5494c04410fa8943cfb8faac.jpg) +Figure 5. Fast training (AFHQ) - We show example outcomes of our method only after 100 iterations of training, as raw output from our Hypernetwork without fast fine-tuning. As shown, even after only 100 iterations, our model is able to generate images that are visually very similar to the conditioning images. + +time of HyperDreamBooth. + +Since we query our Hypernetwork at a single optimization step $T$ , our method also yields efficient inference. Inference of our method takes only 0.3 seconds, compared to 2 seconds for HyperDreamBooth [40] and 5 minutes for DreamBooth. Additionally, the fast fine-tuning required by both our method and HyperDreamBooth takes just 20 seconds, allowing for quick adaptation with minimal overhead. + +Results. We provide qualitative highlights in Figs. 3 and 4, and the quantitative summaries in Tables 1 and 2. Our results for image personalization in Figs. 3 and 4 is on par with the target Dreambooth model whose weights we estimate with our HyperNet Fields approach. That is, our approach generates personalized images that are representative of the given input prompt and the conditional image. Again, note that we achieve these personalization results in a single step by querying the hypernetwork once at timestep $T$ . Interestingly, our method demonstrates remarkable performance for preserving the original image context, as measured by DINO and CLIP-I scores (Tab. 1). Further, in the case of the AFHQ dataset (Tab. 2), we outperform + +![](images/622f847ef9789721e14016289303b3cef1492280a23a9b1e8a5d3561ee2d4e94.jpg) +Figure 6. Qualitative examples (3D Shapes) – We show example shape reconstruction results converting (left) point clouds to 3D shapes and (right) images to 3D shapes. More qualitative examples are available in the supplementary material. + +all DreamBooth variants, demonstrating that not only do we remove the compute-heavy optimization requirement of DreamBooth [39], but also by amortizing the behavior of the learned weights with respect to the input conditioning image, we successfully replicate the dynamics of the task-specific dreambooth network yielding improved results. We emphasize that while results are comparable to each other, our method is significantly faster than the baselines. + +Fast training. An interesting observation from our training process is that our model exhibits faster adaptation to personalization. As shown in Fig. 5, after only 100 training iterations, our model generates images that are visually similar to the conditioning images, achieving this level of personalization with minimal training time. We hypothesize that this rapid adaptation is due to our model's ability to simulate the optimization pathway, effectively learning the direction in which to adjust network weights to reduce the task-specific loss. + +# 4.2. Application to 3D shape reconstruction + +We further explore the potential of our method for the task of 3D shape reconstruction, specifically the task of estimating the weights of an occupancy network [27]. Additionally, we consider two different conditioning inputs: a rendered image of the shape [24, 50], and a point cloud of the shape [1, 15, 32]. For the image conditioning, we use the 38 renderings provided in Qiu et al. [33]—covering various angles around each object—which we sample one of and encode with a pre-trained ViT [9]. For the point cloud conditioning, we randomly sample 2048 points uniformly from the surface of the ground truth mesh and encode the input point + +![](images/af6289ace7e35a25b4d13c097e058ea67458d953ad8fd060fca9f559fd843772.jpg) +Figure 7. Hypernet Field output example (ShapeNet) - We show the output of our HyperNet Field for various points along the optimization trajectory as well as the intersection over union (IoU) over all values of T. As shown, the HyperNet Field is able to accurately estimate the entire optimization trajectory, starting from network weights that only represent the airplane shape roughly, and converging to one that provides a detailed reconstruction of the airplane. The IoU plot also shows that as we traverse the trajectory, the IoU of the generated shapes to the ground truth shape improves, demonstrating convergence. Despite being shown only one ground truth optimization step at a time, our model is able to represent the entire optimization trajectory from initialization to convergence. + +cloud with an off-the-shelf variational autoencoder [48]. + +Task network - Occupancy Net. For both cases, the hypernetwork field architecture remains the same as described in Sec. 3.4, and for the task-specific network, we follow an architecture similar to HyperDiffusion [12] —we use a simple 3-layer dense network [41] with ReLU activations and a hidden dimension of 128. The task-specific network, the occupancy network, takes the 3D coordinates of a query point as input and outputs a scalar value representing the occupancy as in Mescheder et al. [27]. We use marching cubes [26] to render the outcome. + +Training setup. To supervise the training, we sample 20k query points, 10k uniformly, and 10k near the ground-truth surface to be used for training. To find the ground-truth occupancy values for these points, we first make the meshes watertight with [19] and use winding numbers [4] on the watertight mesh to determine whether each point lies inside or outside the surface, and compute the occupancy via Barill et al. [3]. We use the mean squared error (MSE) of the occupancy estimates for the task-specific loss. For all cases, we set $T = 10,000$ . + +2D image conditioning (Objaverse shapes). For 2D image conditioning, as training over the entire Objaverse dataset [8] would be computationally expensive, we simply overfit to a subset of 128 shapes, as we are mostly interested in the potential of our method. We leave training on the full Objaverse dataset as future work. We obtain reasonable shape reconstruction results as shown in Figures 1 and 6 with our method for this simple overfitting demonstration just within 10 hours of training on a single NVidia RTX 3090 GPU. + +3D point cloud conditioning (ShapeNet shapes). For 3D point cloud conditioning, we use the entire dataset. As shown in Fig. 6, our method is able to learn to generate occupancy network weights from point clouds. We further illustrate how both our Hypernet Field output and the IoU for shape generation evolves over the optimization process in Fig. 7. As shown, our method simulates the entire train- + +ing process of the task-specific network for a given point cloud despite only being shown a single ground truth gradient step at a time. + +# 5. Conclusion + +We presented Hypernetwork Fields, a novel and efficient hypernetwork training framework. Instead of optimizing individual weights per sample, we use gradient supervision and match gradients along the hypernetworks field and the task-specific network at random points along the optimization trajectory. This drastically reduces the precomputation overhead of training hypernetworks, while maintaining comparable performance to state of the art. We have demonstrated the effectiveness of our method on personalized image generation and 3D shape reconstruction. Our framework is general and can be applied to any hypernetwork training scenario, making it a promising direction for scaling hypernetwork applications to larger datasets. + +Limitations and future work. Our method enables stable hypernetwork training on large datasets. Because of this, the potential of our method for diverse tasks remains underexplored. This may include tasks such as training a hypernetwork to estimate the weights of NeRFs or even personalized adapters for large language models. However, our approach requires as many output neurons as there are weights in the target network, resulting in suboptimal memory usage. Recent work by Dravid et al. [10] has shown that customized DreamBooth adapters can be represented using only 1k parameters. Integrating our method with this compact representation approach could significantly enhance parameter efficiency while maintaining our proposed benefits of efficient hypernetwork training. These would be interesting avenues for future work. + +# References + +[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitlagkas, and Leonidas Guibas. Learning representations and generative + +models for 3d point clouds. In ICML, 2018. 7 +[2] Sudarshan Babu, Richard Liu, Avery Zhou, Michael Maire, Greg Shakhnarovich, and Rana Hanocka. Hyperfields: Towards zero-shot generation of nerfs from text. In ICML. PMLR, 2024. 2 +[3] Gavin Barill, Neil G Dickson, Ryan Schmidt, David IW Levin, and Alec Jacobson. Fast winding numbers for soups and clouds. Transactions On Graphics, 37(4):1-12, 2018. 8 +[4] Gavin Barill, Neil G. Dickson, Ryan M. Schmidt, David I. W. Levin, and Alec Jacobson. Fast winding numbers for soups and clouds. Transactions On Graphics, 37:1-12, 2018. 8 +[5] Tom B Brown. Language models are few-shot learners. NeurIPS, 2020. 5 +[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, pages 9650-9660, 2021. 6 +[7] Yunjoy Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis for multiple domains. In CVPR, 2020. 5, 6, 11 +[8] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In CVPR, pages 13142-13153, 2023. 8 +[9] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 5, 7 +[10] Amir Dravid, Yossi Gandelsman, Kuan-Chieh Wang, Rameen Abdal, Gordon Wetzstein, Alexei A Efros, and Kfir Aberman. Interpreting the weight space of customized diffusion models. arXiv preprint arXiv:2406.09413, 2024. 8 +[11] Yilun Du and Igor Mordatch. Implicit generation and modeling with energy based models. NeurIPS, 32, 2019. 3 +[12] Ziya Erkoç, Fangchang Ma, Qi Shan, Matthias Nießner, and Angela Dai. Hyperdiffusion: Generating implicit neural fields with weight-space diffusion. In ICCV, pages 14300-14310, 2023. 2, 3, 8 +[13] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. *ICLR*, 2023. 5, 6 +[14] Will Grathwohl, Kuan-Chieh Wang, Jorn-Henrik Jacobsen, David Duvenaud, Kevin Swersky, and Mohammad Norouzi. Learning energy-based models by diffusion recovery likelihood. ICLR, 2021. 3 +[15] Thibault Groueix, Matthew Fisher, Vladimir G Kim, Bryan Russell, and Mathieu Aubry. Atlasnet: A papier-mâché approach to learning 3d surface generation. In CVPR, 2018. 7 +[16] David Ha, Andrew Dai, and Quoc V Le. Hypernetworks. ICLR, 2017. 2 +[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, pages 6840-6851, 2020. 3 +[18] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In ICLR, 2022. 6 + +[19] Jingwei Huang, Yichao Zhou, and Leonidas Guibas. Manifoldplus: A robust and scalable watertight manifold surface generation method for triangle soups. arxiv, 2020. 8 +[20] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. In ICLR, 2018. 5 +[21] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. NeurIPS, 35:26565-26577, 2022. 4 +[22] Yann LeCun, Sumit Chopra, Raia Hadsell, Marc'Aurelio Ranzato, and Fu Jie Huang. A tutorial on energy-based learning. Predicting structured data, 1(2):1-31, 2006. 3 +[23] Yaron Lipman, Justin Solomon, et al. Flow matching for generative modeling. *ICLR*, 2023. 3 +[24] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9298–9309, 2023. 7 +[25] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In ICCV, 2015. 5, 11 +[26] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. In Semin al graphics: pioneering efforts that shaped the field, pages 347-353, 1998. 8 +[27] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In CVPR, pages 4460-4470, 2019. 2, 7, 8 +[28] Jonathan Oswald, Christian Henning, Tomas Pfister, and Jürgen Schmidhuber. Continual learning with hypernetworks. In ICLR, 2019. 2 +[29] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In CVPR, pages 165-174, 2019. 2 +[30] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, pages 4195-4205, 2023. 5 +[31] Marcin Przewiezlikowski, Przemyslaw Przybysz, Jacek Tabor, Maciej Zikeba, and Przemyslaw Spurek. Hypermaml: Few-shot adaptation of deep models with hypernetworks. Neurocomputing, 2024. 2 +[32] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, 2017. 7 +[33] Lingteng Qiu, Guanying Chen, Xiaodong Gu, Qi Zuo, Mutian Xu, Yushuang Wu, Weihao Yuan, Zilong Dong, Liefeng Bo, and Xiaoguang Han. Richdreamer: A generalizable normal-depth diffusion model for detail richness in text-to-3d. In CVPR, pages 9914–9925, 2024. 7 +[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 6 + +[35] Maziar Raissi, Paris Perdikaris, and George Em Karniadakis. Physics informed deep learning (part i): Data-driven solutions of nonlinear partial differential equations. arxiv, 2017. 3 +[36] Maziar Raissi, Paris Perdikaris, and George Em Karniadakis. Physics informed deep learning (part ii): Data-driven discovery of nonlinear partial differential equations. arxiv, 2017. +[37] Maziar Raissi, Paris Perdikaris, and George Em Karniadakis. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational Physics, 378:686-707, 2019. 3 +[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 5, 6 +[39] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, pages 22500-22510, 2023. 1, 3, 4, 5, 6, 7 +[40] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdream booth: Hypernetworks for fast personalization of text-to-image models. In CVPR, pages 6527-6536, 2024. 2, 3, 5, 6, 7 +[41] David E Rumelhart, Geoffrey E Hinton, and Ronald J Williams. Learning representations by back-propagating errors. NATURE, 323(6088):533-536, 1986. 8 +[42] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 35:36479-36494, 2022. 5 +[43] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, pages 2256-2265, 2015. 3 +[44] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. NeurIPS, 32, 2019. 3 +[45] Yang Song, Conor Durkan, Iain Murray, and Stefano Ermon. Maximum likelihood training of score-based diffusion models. NeurIPS, pages 1415-1428, 2021. 3 +[46] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. arxiv, 2023. 3, 4 +[47] Pascal Vincent. A connection between score matching and denoising autoencoders. Neural Computation, 23(7):1661-1674, 2011. 3 +[48] Guandao Yang, Xun Huang, Zekun Hao, Ming-Yu Liu, Serge Belongie, and Bharath Hariharan. Pointflow: 3d point cloud generation with continuous normalizing flows. In ICCV, pages 4541-4550, 2019. 8 +[49] Linnan Zhang, Zhiding Ren, Seunghyun Ahn, Hakan Bilen, and Kwang In Lee. Graph hypernetworks for neural architecture search. *ICLR*, 2019. 2 +[50] Jun-Yan Zhu et al. Shap-e: Generating conditional 3d implicit functions. arxiv, 2023. 7 \ No newline at end of file diff --git a/2025/HyperNet Fields_ Efficiently Training Hypernetworks without Ground Truth by Learning Weight Trajectories/images.zip b/2025/HyperNet Fields_ Efficiently Training Hypernetworks without Ground Truth by Learning Weight Trajectories/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5581a45a443fe5cbd934c3157a35af531d6b2f0c --- /dev/null +++ b/2025/HyperNet Fields_ Efficiently Training Hypernetworks without Ground Truth by Learning Weight Trajectories/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8dcd37f63b73c49f8ff2aeeababcbfb35e4bd80aef3b2a17be4b02c0e5befaf6 +size 633732 diff --git a/2025/HyperNet Fields_ Efficiently Training Hypernetworks without Ground Truth by Learning Weight Trajectories/layout.json b/2025/HyperNet Fields_ Efficiently Training Hypernetworks without Ground Truth by Learning Weight Trajectories/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..58d166aad23cc1177bcfe03b14539b22895d2a28 --- /dev/null +++ b/2025/HyperNet Fields_ Efficiently Training Hypernetworks without Ground Truth by Learning Weight Trajectories/layout.json @@ -0,0 +1,10117 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 114, + 103, + 500, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 103, + 500, + 140 + ], + "spans": [ + { + "bbox": [ + 114, + 103, + 500, + 140 + ], + "type": "text", + "content": "HyperNet Fields: Efficiently Training Hypernetworks without Ground Truth by Learning Weight Trajectories" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 91, + 159, + 520, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 159, + 520, + 190 + ], + "spans": [ + { + "bbox": [ + 91, + 159, + 520, + 190 + ], + "type": "text", + "content": "Eric Hedlin" + }, + { + "bbox": [ + 91, + 159, + 520, + 190 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 91, + 159, + 520, + 190 + ], + "type": "text", + "content": " Munawar Hayat" + }, + { + "bbox": [ + 91, + 159, + 520, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 91, + 159, + 520, + 190 + ], + "type": "text", + "content": " Fatih Porikli" + }, + { + "bbox": [ + 91, + 159, + 520, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 91, + 159, + 520, + 190 + ], + "type": "text", + "content": " Kwang Moo Yi" + }, + { + "bbox": [ + 91, + 159, + 520, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 91, + 159, + 520, + 190 + ], + "type": "text", + "content": " Shweta Mahajan" + }, + { + "bbox": [ + 91, + 159, + 520, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 91, + 159, + 520, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 91, + 159, + 520, + 190 + ], + "type": "text", + "content": " University of British Columbia " + }, + { + "bbox": [ + 91, + 159, + 520, + 190 + ], + "type": "inline_equation", + "content": "{}^{2}" + }, + { + "bbox": [ + 91, + 159, + 520, + 190 + ], + "type": "text", + "content": " Qualcomm AI Research†" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 58, + 217, + 306, + 409 + ], + "blocks": [ + { + "bbox": [ + 58, + 217, + 306, + 409 + ], + "lines": [ + { + "bbox": [ + 58, + 217, + 306, + 409 + ], + "spans": [ + { + "bbox": [ + 58, + 217, + 306, + 409 + ], + "type": "image", + "image_path": "c4f22adc05793672a7e7332c344006a515b952740da570723acf1df772c116ed.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 417, + 555, + 451 + ], + "lines": [ + { + "bbox": [ + 54, + 417, + 555, + 451 + ], + "spans": [ + { + "bbox": [ + 54, + 417, + 555, + 451 + ], + "type": "text", + "content": "Figure 1. Teaser - We propose hypernetwork fields, where we learn the entire weight space trajectory instead of only the final state, which allows us to train without ever needing to know the final converged weights. Our method can be applied to any application of hypernetworks, including diffusion model personalization and modeling of 3D neural representations." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 310, + 217, + 552, + 407 + ], + "blocks": [ + { + "bbox": [ + 310, + 217, + 552, + 407 + ], + "lines": [ + { + "bbox": [ + 310, + 217, + 552, + 407 + ], + "spans": [ + { + "bbox": [ + 310, + 217, + 552, + 407 + ], + "type": "image", + "image_path": "91f43ea1bd60ca683d11dc4a81ff813e1a575285286666305481408c8df0b3ec.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 460, + 200, + 473 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 460, + 200, + 473 + ], + "spans": [ + { + "bbox": [ + 151, + 460, + 200, + 473 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 486, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 486, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 486, + 297, + 714 + ], + "type": "text", + "content": "To efficiently adapt large models or to train generative models of neural representations, Hypernetworks have drawn interest. While hypernetworks work well, training them is cumbersome, and often requires ground truth optimized weights for each sample. However, obtaining each of these weights is a training problem of its own—one needs to train, e.g., adaptation weights or even an entire neural field for hypernetworks to regress to. In this work, we propose a method to train hypernetworks, without the need for any per-sample ground truth. Our key idea is to learn a Hypernetwork 'Field' and estimate the entire trajectory of network weight training instead of simply its converged state. In other words, we introduce an additional input to the Hypernetwork, the convergence state, which then makes it act as a neural field that models the entire convergence pathway of a task network. A critical benefit in doing so is that the gradient of the estimated weights at any convergence state must then match the gradients of the original task—this constraint alone is sufficient to train the Hypernetwork Field." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 461, + 556, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 556, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 556, + 510 + ], + "type": "text", + "content": "We demonstrate the effectiveness of our method through the task of personalized image generation and 3D shape reconstruction from images and point clouds, demonstrating competitive results without any per-sample ground truth." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 545, + 395, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 545, + 395, + 558 + ], + "spans": [ + { + "bbox": [ + 314, + 545, + 395, + 558 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 566, + 555, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 566, + 555, + 675 + ], + "spans": [ + { + "bbox": [ + 312, + 566, + 555, + 675 + ], + "type": "text", + "content": "Various generative modeling tasks [39] require adaptation, thus data-dependent network weights, or in other words, model parameters being adjusted according to various circumstances (conditions and user input). These include personalized text-to-image generation where the weights of the underlying personalization model, such as DreamBooth [39], are fine-tuned on the concept of interest and used to generate target images conditioned on input prompts. Another example is 3D modeling, where neural" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 324, + 684, + 467, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 684, + 467, + 693 + ], + "spans": [ + { + "bbox": [ + 324, + 684, + 467, + 693 + ], + "type": "text", + "content": "*Work done as part of a summer internship." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 315, + 694, + 553, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 694, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 315, + 694, + 553, + 712 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 315, + 694, + 553, + 712 + ], + "type": "text", + "content": "Qualcomm AI Research is an initiative of Qualcomm Technologies, Inc." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22129" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 294, + 144 + ], + "type": "text", + "content": "networks are used to model shapes implicitly with signed distance fields [29] or occupancy [27]. In both cases, a dedicated set of neural network weights is required for each personalization outcome or 3D shape. Thus, the applicability of these methods is highly limited by the training time required to obtain these weights." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 145, + 295, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 145, + 295, + 335 + ], + "spans": [ + { + "bbox": [ + 56, + 145, + 295, + 335 + ], + "type": "text", + "content": "It is therefore critical to seek ways to amortize the per-sample optimizations into Hypernetworks [12, 40]—networks that estimate network parameters. By 'skipping ahead' the tedious training processes involved in fine-tuning or shape regressing, they greatly speed up the entire process. Existing methods, however, have a critical shortcoming in that they supervise hypernetworks by directly matching their output to the precomputed, ground truth converged weights corresponding to the task-specific network for each sample in the dataset. Notably, prior to any training of the Hypernetwork, one must prepare converged task-specific network weights for every sample in the training set. This severely limits the capabilities of Hypernetwork methods when scaling to large datasets. As an example, to train HyperDreamBooth [40], it requires 50 days of GPU time on a NVidia RTX 3090," + }, + { + "bbox": [ + 56, + 145, + 295, + 335 + ], + "type": "inline_equation", + "content": "^{\\ddagger}" + }, + { + "bbox": [ + 56, + 145, + 295, + 335 + ], + "type": "text", + "content": " to just prepare the training data alone." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 336, + 294, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 336, + 294, + 432 + ], + "spans": [ + { + "bbox": [ + 55, + 336, + 294, + 432 + ], + "type": "text", + "content": "Furthermore, current methods assume a one-to-one mapping between each input sample and its corresponding optimized weights. This assumption is an oversimplification. Neural network optimization is stochastic with various equally plausible solutions. Therefore, enforcing a bijective mapping between the condition and the ultimately converged weights may limit the potential expressiveness of the hypernetwork, as we will demonstrate in Sec. 3." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 433, + 295, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 433, + 295, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 433, + 295, + 588 + ], + "type": "text", + "content": "To address these limitations, we instead propose to train a Hypernetwork Field that estimates the entire optimization trajectory of a neural network training, rather than focusing on a single final set of weights for each sample. Specifically, we introduce the convergence state as additional input to the hypernetwork, such that the weights of the task-specific network at a desired state of convergence can be 'queried'. An interesting outcome of doing so is that the gradient of the estimated weights along this pathway must match the gradients of the original task-specific network, as they are both gradients pointing towards convergence. Hence, training of such hypernetwork fields can simply be done by matching the two gradients." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 589, + 295, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 589, + 295, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 589, + 295, + 684 + ], + "type": "text", + "content": "By supervising the hypernetwork to match the gradient of the weights with respect to the optimization step, we avoid the need for precomputing target weights. This supervision strategy allows the hypernetwork to estimate a trajectory of parameters that, at each step, reflects a compatible state across all samples in the dataset. Ultimately, at inference time, the estimated parameters which correspond with the final timestep of the hypernetwork, represent a well" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "converged solution for each of the samples, but are now naturally discovered through training of the hypernetwork field itself. Notably, this single forward pass to estimating the converged parameters is all that is necessary for inference, hence no additional compute is required compared to conventional hypernetwork approaches [12, 40]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 145, + 553, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 145, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 313, + 145, + 553, + 229 + ], + "type": "text", + "content": "Our framework is general and is applicable to all hypernetwork training scenarios. To demonstrate the effectiveness of our method, we apply our method to personalized image generation and 3D shape generation. We achieve competitive performance while reducing the training cost by " + }, + { + "bbox": [ + 313, + 145, + 553, + 229 + ], + "type": "inline_equation", + "content": "\\sim 4" + }, + { + "bbox": [ + 313, + 145, + 553, + 229 + ], + "type": "text", + "content": " times what would be required of conventional hypernetwork training." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 326, + 231, + 476, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 231, + 476, + 241 + ], + "spans": [ + { + "bbox": [ + 326, + 231, + 476, + 241 + ], + "type": "text", + "content": "To summarize, our contributions are:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 243, + 553, + 387 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 314, + 243, + 553, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 243, + 553, + 291 + ], + "spans": [ + { + "bbox": [ + 314, + 243, + 553, + 291 + ], + "type": "text", + "content": "- We propose hypernetwork fields, an approach to learn the entire optimization pathway of task-specific networks, which allows ground truth-free training without the need for precomputing the target weights;" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 292, + 553, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 292, + 553, + 315 + ], + "spans": [ + { + "bbox": [ + 313, + 292, + 553, + 315 + ], + "type": "text", + "content": "- our approach drastically reduces the compute required for training hypernetworks;" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 316, + 553, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 553, + 363 + ], + "type": "text", + "content": "- we validate our approach on two very different tasks and datasets—for personalized image generation we achieve results comparable to the state of the art with significantly less compute; and" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 363, + 553, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 363, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 553, + 387 + ], + "type": "text", + "content": "- we further explore the application of our method to 3D shape reconstruction from images an point clouds." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 402, + 400, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 402, + 400, + 415 + ], + "spans": [ + { + "bbox": [ + 314, + 402, + 400, + 415 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 426, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 715 + ], + "type": "text", + "content": "Hypernetworks. Early work on training hypernetworks backpropagates gradients through the task-specific network to the hypernetwork itself, as demonstrated by approaches [16, 31] and subsequent explorations in continual learning [28] and neural architecture search [49]. These methods have shown that allowing gradient flow through both networks enables the hypernetwork to modify parameters for diverse tasks adaptively. However, more recent works have introduced approaches that involve precomputing weights for the task-specific network and directly supervise the hypernetwork to match them, eliminating the need for backpropagation through the task-specific network and improving training stability by alleviating issues with gradient flow. This supervision with precomputed weights has been shown to enhance gradient stability and training performance, as highlighted in frameworks like HyperDiffusion [12], HyperFields [2], and HyperDreamBooth [40]. However, this limits the size of the dataset to which this can be applied, as the precompute requirement is proportional to the number of samples in the dataset. Our method combines the best of both approaches by removing the need for precomputed weights while explicitly providing weight-based supervision in the form of gradients along the convergence path, enabling stable training on large datasets." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "text", + "content": "HyperDreamBooth [40] was trained on 15k DreamBooth weights, which each take 5 minutes to compute." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22130" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 288 + ], + "type": "text", + "content": "Denoising Models. In generative modeling, finding smooth and reliable paths through latent space has become central to producing coherent samples. Examples include diffusion models [17, 43], which iteratively trace a path from noise to data through successive steps, effectively navigating high-dimensional space by reversing a stochastic process. Consistency models [46] refine this by learning non-Markovian trajectories that reach target data distributions in fewer steps by introducing a formulation that allows them to single-step jump to any timestep. Flow-matching models [23] take a similar path-based approach by learning a vector field that guides samples along pre-defined flows, ensuring a direct and smooth transition through the latent space. Similarly to these methods, our method can be thought of as finding a path through this high-dimensional space, specifically in the form of a single-step estimator from initialization to any desired state of convergence, whose path travels strictly along the direction of the steepest gradient descent." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 292, + 296, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 292, + 296, + 653 + ], + "spans": [ + { + "bbox": [ + 56, + 292, + 296, + 653 + ], + "type": "text", + "content": "Methods with gradient-supervision. Several recent approaches in generative modeling and physics-informed learning employ supervision exclusively through gradients rather than direct supervision between input-output pairs to achieve denoising and reconstruction. Score-based generative models [44, 45, 47] use score matching to learn the gradient (or \"score\") of the log density of noisy data distributions, bypassing the need for direct output supervision. Denoising Score Matching (DSM) [47] trains the model to predict this gradient, guiding it to approximate data distributions through iterative denoising and demonstrating improved stability in noise removal. Similarly, Physics-Informed Neural Networks (PINNs) [35-37] leverage known derivatives from partial differential equations (PDEs) in physical simulations, such as heat diffusion or fluid flow, to inform the network solely through spatial or temporal derivatives. This derivative-based guidance encourages convergence as the model adapts to match observed physical behaviour, requiring no explicit target outputs. Energy-Based Models (EBMs) [11, 14, 22] also align with this gradient-driven approach by learning energy landscapes that adjust gradients to represent target distributions, allowing for smooth gradient matching through learned energy fields. Inspired by these methods, our approach supervises only the gradient direction in task-specific weight space, effectively learning a convergence pathway without requiring explicit final state supervision. By supervising gradients within this pathway, our method achieves robust training stability while generalizing over diverse tasks without per-sample ground truth weights." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 667, + 111, + 680 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 667, + 111, + 680 + ], + "spans": [ + { + "bbox": [ + 55, + 667, + 111, + 680 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": "To remove the precompute requirements associated with training hypernetworks, we present a new framework that" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": "models the convergence dynamics of task-specific networks into a single, efficient process, which we name Hypernetwork Field." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 111, + 555, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 111, + 555, + 255 + ], + "spans": [ + { + "bbox": [ + 313, + 111, + 555, + 255 + ], + "type": "text", + "content": "Problem formulation. Formally, let " + }, + { + "bbox": [ + 313, + 111, + 555, + 255 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 313, + 111, + 555, + 255 + ], + "type": "text", + "content": " denote a task-specific neural network, e.g., dreambooth [39], parameterized by " + }, + { + "bbox": [ + 313, + 111, + 555, + 255 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 111, + 555, + 255 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 111, + 555, + 255 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 313, + 111, + 555, + 255 + ], + "type": "text", + "content": " is then typically trained to minimize a loss function " + }, + { + "bbox": [ + 313, + 111, + 555, + 255 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_{\\mathbf{x}}[\\mathcal{L}(\\theta ,\\mathbf{x})]" + }, + { + "bbox": [ + 313, + 111, + 555, + 255 + ], + "type": "text", + "content": ", by training for " + }, + { + "bbox": [ + 313, + 111, + 555, + 255 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 111, + 555, + 255 + ], + "type": "text", + "content": " optimization steps given a dataset of conditions, e.g., image " + }, + { + "bbox": [ + 313, + 111, + 555, + 255 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 111, + 555, + 255 + ], + "type": "text", + "content": ". In our work, we wish to learn a hypernetwork field " + }, + { + "bbox": [ + 313, + 111, + 555, + 255 + ], + "type": "inline_equation", + "content": "H_{\\phi}(t,\\mathbf{x})" + }, + { + "bbox": [ + 313, + 111, + 555, + 255 + ], + "type": "text", + "content": ", parameterized by " + }, + { + "bbox": [ + 313, + 111, + 555, + 255 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 111, + 555, + 255 + ], + "type": "text", + "content": " that takes as input a condition " + }, + { + "bbox": [ + 313, + 111, + 555, + 255 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 111, + 555, + 255 + ], + "type": "text", + "content": " and an optimization step " + }, + { + "bbox": [ + 313, + 111, + 555, + 255 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 111, + 555, + 255 + ], + "type": "text", + "content": ". The hypernetwork field then predicts the parameters " + }, + { + "bbox": [ + 313, + 111, + 555, + 255 + ], + "type": "inline_equation", + "content": "\\theta_t = H_\\phi (t,\\mathbf{x})" + }, + { + "bbox": [ + 313, + 111, + 555, + 255 + ], + "type": "text", + "content": " at optimization step " + }, + { + "bbox": [ + 313, + 111, + 555, + 255 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 111, + 555, + 255 + ], + "type": "text", + "content": " of the task-specific network " + }, + { + "bbox": [ + 313, + 111, + 555, + 255 + ], + "type": "inline_equation", + "content": "f_{\\theta_t}(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 111, + 555, + 255 + ], + "type": "text", + "content": ", given the condition " + }, + { + "bbox": [ + 313, + 111, + 555, + 255 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 111, + 555, + 255 + ], + "type": "text", + "content": " and the timestep " + }, + { + "bbox": [ + 313, + 111, + 555, + 255 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 111, + 555, + 255 + ], + "type": "text", + "content": ". In the following, we show how to obtain " + }, + { + "bbox": [ + 313, + 111, + 555, + 255 + ], + "type": "inline_equation", + "content": "H_{\\phi}(t,\\mathbf{x})" + }, + { + "bbox": [ + 313, + 111, + 555, + 255 + ], + "type": "text", + "content": " without requiring any " + }, + { + "bbox": [ + 313, + 111, + 555, + 255 + ], + "type": "inline_equation", + "content": "\\theta_t" + }, + { + "bbox": [ + 313, + 111, + 555, + 255 + ], + "type": "text", + "content": " to supervise the training." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 262, + 414, + 275 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 262, + 414, + 275 + ], + "spans": [ + { + "bbox": [ + 313, + 262, + 414, + 275 + ], + "type": "text", + "content": "3.1. HyperNet Fields" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 280, + 554, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 280, + 554, + 364 + ], + "spans": [ + { + "bbox": [ + 313, + 280, + 554, + 364 + ], + "type": "text", + "content": "To train the hypernet field " + }, + { + "bbox": [ + 313, + 280, + 554, + 364 + ], + "type": "inline_equation", + "content": "H_{\\phi}(t,\\mathbf{x})" + }, + { + "bbox": [ + 313, + 280, + 554, + 364 + ], + "type": "text", + "content": " without the ground-truth targets for it to regress, we take a step back and revisit what standard hypernetworks are optimizing for. Standard approaches for learning hypernetworks [12, 40] utilize the ground-truth weights " + }, + { + "bbox": [ + 313, + 280, + 554, + 364 + ], + "type": "inline_equation", + "content": "\\theta^{*}" + }, + { + "bbox": [ + 313, + 280, + 554, + 364 + ], + "type": "text", + "content": " of the task-specific network, corresponding to the optimized parameters for a specific sample " + }, + { + "bbox": [ + 313, + 280, + 554, + 364 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 280, + 554, + 364 + ], + "type": "text", + "content": " as a supervision signal. In other words, they first find" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 380, + 373, + 555, + 393 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 373, + 555, + 393 + ], + "spans": [ + { + "bbox": [ + 380, + 373, + 555, + 393 + ], + "type": "interline_equation", + "content": "\\theta^ {*} (\\mathbf {x}) = \\underset {\\theta} {\\arg \\min } \\mathcal {L} (\\theta , \\mathbf {x}), \\tag {1}", + "image_path": "db0e103db0508ebd752ddd18063ae94f427f12abd1c6f432ff13d0bc671cb9ba.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 402, + 554, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 554, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 554, + 474 + ], + "type": "text", + "content": "for each " + }, + { + "bbox": [ + 313, + 402, + 554, + 474 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 402, + 554, + 474 + ], + "type": "text", + "content": ", then find " + }, + { + "bbox": [ + 313, + 402, + 554, + 474 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 402, + 554, + 474 + ], + "type": "text", + "content": ", such that " + }, + { + "bbox": [ + 313, + 402, + 554, + 474 + ], + "type": "inline_equation", + "content": "\\theta^{*}(\\mathbf{x}) \\approx H_{\\phi}(\\cdot, \\mathbf{x}), \\forall \\mathbf{x}" + }, + { + "bbox": [ + 313, + 402, + 554, + 474 + ], + "type": "text", + "content": ". This typically requires several iterations of the optimization process and thus incurs a large computational cost for generating the target parameters " + }, + { + "bbox": [ + 313, + 402, + 554, + 474 + ], + "type": "inline_equation", + "content": "\\theta^{*}(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 402, + 554, + 474 + ], + "type": "text", + "content": ". Moreover, since this process has to be repeated for each sample independently, the scalability to large datasets is severely limited." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 475, + 554, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 554, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 554, + 617 + ], + "type": "text", + "content": "Note, however, Eq. (1) is typically found through gradient-based optimization, and forms a trajectory of weight parameter space in the " + }, + { + "bbox": [ + 313, + 475, + 554, + 617 + ], + "type": "inline_equation", + "content": "\\{\\theta_t\\}_{t = 1\\dots T}" + }, + { + "bbox": [ + 313, + 475, + 554, + 617 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 475, + 554, + 617 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 475, + 554, + 617 + ], + "type": "text", + "content": " is the number of optimization steps. If we now consider our hypernet field " + }, + { + "bbox": [ + 313, + 475, + 554, + 617 + ], + "type": "inline_equation", + "content": "H_{\\phi}(t,\\mathbf{x})" + }, + { + "bbox": [ + 313, + 475, + 554, + 617 + ], + "type": "text", + "content": " that we intend to model, we notice an interesting relationship that must hold between the " + }, + { + "bbox": [ + 313, + 475, + 554, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\theta_t,\\mathbf{x})" + }, + { + "bbox": [ + 313, + 475, + 554, + 617 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 475, + 554, + 617 + ], + "type": "inline_equation", + "content": "H_{\\phi}(t,\\mathbf{x})" + }, + { + "bbox": [ + 313, + 475, + 554, + 617 + ], + "type": "text", + "content": ". Specifically, consider the gradient of the task-specific weights at an optimization step " + }, + { + "bbox": [ + 313, + 475, + 554, + 617 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 475, + 554, + 617 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 475, + 554, + 617 + ], + "type": "inline_equation", + "content": "\\nabla_{\\theta_t}\\mathcal{L}(\\theta_t,\\mathbf{x})" + }, + { + "bbox": [ + 313, + 475, + 554, + 617 + ], + "type": "text", + "content": ". Then, if a hypernet field " + }, + { + "bbox": [ + 313, + 475, + 554, + 617 + ], + "type": "inline_equation", + "content": "H_{\\phi}" + }, + { + "bbox": [ + 313, + 475, + 554, + 617 + ], + "type": "text", + "content": " has been properly trained, the direction in which " + }, + { + "bbox": [ + 313, + 475, + 554, + 617 + ], + "type": "inline_equation", + "content": "\\nabla_{\\theta_t}\\mathcal{L}(\\theta_t,\\mathbf{x})" + }, + { + "bbox": [ + 313, + 475, + 554, + 617 + ], + "type": "text", + "content": " point to must match the direction in which the hypernet trajectory flows, that is, " + }, + { + "bbox": [ + 313, + 475, + 554, + 617 + ], + "type": "inline_equation", + "content": "\\nabla_tH_\\phi (t,\\mathbf{x})" + }, + { + "bbox": [ + 313, + 475, + 554, + 617 + ], + "type": "text", + "content": ". In other words, their gradients must match." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "text", + "content": "In fact, matching the two gradients allows us to circumvent the necessity of per-sample optimization as " + }, + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\nabla_{\\theta_t}\\mathcal{L}(\\theta_t,\\mathbf{x})" + }, + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "text", + "content": " can be computed with relatively minimal overhead by using the current hypernet field estimate for " + }, + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\theta_t" + }, + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "text", + "content": ". This process can also be understood as amortizing Eq. (1) into the training process of the hypernet field " + }, + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "inline_equation", + "content": "H_{\\phi}(t,\\mathbf{x})" + }, + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "text", + "content": ", as we are 'skipping ahead' parts of the optimization loop of Eq. (1) with the estimates coming from " + }, + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "inline_equation", + "content": "H_{\\phi}(t,\\mathbf{x})" + }, + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "text", + "content": ". Through" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "22131" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 71, + 298, + 159 + ], + "blocks": [ + { + "bbox": [ + 56, + 71, + 298, + 159 + ], + "lines": [ + { + "bbox": [ + 56, + 71, + 298, + 159 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 298, + 159 + ], + "type": "image", + "image_path": "868d0df2f2a67d6a8b042c2f4131997dd8911b70b965b1e961db9efb123e9367.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 167, + 555, + 191 + ], + "lines": [ + { + "bbox": [ + 54, + 167, + 555, + 191 + ], + "spans": [ + { + "bbox": [ + 54, + 167, + 555, + 191 + ], + "type": "text", + "content": "Figure 2. The Hypenet Fields Framework. Left: Our Hypernetwork with gradient-based supervision from a task-specific network - dreambooth [39]. Right: The sampling process to generate personalized images with our hypernetwork framework." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 312, + 79, + 555, + 138 + ], + "blocks": [ + { + "bbox": [ + 312, + 79, + 555, + 138 + ], + "lines": [ + { + "bbox": [ + 312, + 79, + 555, + 138 + ], + "spans": [ + { + "bbox": [ + 312, + 79, + 555, + 138 + ], + "type": "image", + "image_path": "5327c2512391025e71e269920022c476cc10f1d2ba8bc520d4a5b78a280787e3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 210, + 296, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 210, + 296, + 270 + ], + "spans": [ + { + "bbox": [ + 54, + 210, + 296, + 270 + ], + "type": "text", + "content": "this amortization, the hypernet field captures detailed information about how the weights evolve throughout training, allowing it to effectively predict and generate weights across the training samples. We now further detail the exact training process." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 277, + 211, + 290 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 211, + 290 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 211, + 290 + ], + "type": "text", + "content": "3.2. Gradient-based Supervision" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 294, + 296, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 294, + 296, + 390 + ], + "spans": [ + { + "bbox": [ + 55, + 294, + 296, + 390 + ], + "type": "text", + "content": "In each training iteration of the optimization process, the hypernet field generates an estimate of the task-specific parameters " + }, + { + "bbox": [ + 55, + 294, + 296, + 390 + ], + "type": "inline_equation", + "content": "\\hat{\\theta}_t = H_\\phi(t, \\mathbf{x})" + }, + { + "bbox": [ + 55, + 294, + 296, + 390 + ], + "type": "text", + "content": " at timestep " + }, + { + "bbox": [ + 55, + 294, + 296, + 390 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 294, + 296, + 390 + ], + "type": "text", + "content": " based on the input " + }, + { + "bbox": [ + 55, + 294, + 296, + 390 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 294, + 296, + 390 + ], + "type": "text", + "content": " from a dataset " + }, + { + "bbox": [ + 55, + 294, + 296, + 390 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 294, + 296, + 390 + ], + "type": "text", + "content": " and the timestep " + }, + { + "bbox": [ + 55, + 294, + 296, + 390 + ], + "type": "inline_equation", + "content": "t \\sim U(0, T)" + }, + { + "bbox": [ + 55, + 294, + 296, + 390 + ], + "type": "text", + "content": ". Given this estimate, we compute the gradient of the loss function with respect to the task-specific weights at that timestep, " + }, + { + "bbox": [ + 55, + 294, + 296, + 390 + ], + "type": "inline_equation", + "content": "\\nabla_{\\hat{\\theta}_t} \\mathcal{L}(\\hat{\\theta}_t, \\mathbf{x})" + }, + { + "bbox": [ + 55, + 294, + 296, + 390 + ], + "type": "text", + "content": ", and perform a single optimization step to update the weights to " + }, + { + "bbox": [ + 55, + 294, + 296, + 390 + ], + "type": "inline_equation", + "content": "\\theta_{t+1}" + }, + { + "bbox": [ + 55, + 294, + 296, + 390 + ], + "type": "text", + "content": ". This is given by" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 110, + 396, + 294, + 412 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 396, + 294, + 412 + ], + "spans": [ + { + "bbox": [ + 110, + 396, + 294, + 412 + ], + "type": "interline_equation", + "content": "\\theta_ {t + 1} = \\hat {\\theta} _ {t} - \\eta \\nabla_ {\\hat {\\theta} _ {t}} \\mathcal {L} _ {\\text {t a s k}} (\\hat {\\theta} _ {t}, \\mathbf {x}), \\tag {2}", + "image_path": "25546919d936a5a3752ae1916a4f5aad6dff41c0113d7ac658dcdb02bb0234b9.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 416, + 296, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 416, + 296, + 465 + ], + "spans": [ + { + "bbox": [ + 55, + 416, + 296, + 465 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 416, + 296, + 465 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 55, + 416, + 296, + 465 + ], + "type": "text", + "content": " is the learning rate. Then, the hypernet field must also follow this update by definition, thus we wish to have " + }, + { + "bbox": [ + 55, + 416, + 296, + 465 + ], + "type": "inline_equation", + "content": "\\hat{\\theta}_{t + 1}\\approx \\theta_{t + 1}" + }, + { + "bbox": [ + 55, + 416, + 296, + 465 + ], + "type": "text", + "content": ". Moreover, assuming " + }, + { + "bbox": [ + 55, + 416, + 296, + 465 + ], + "type": "inline_equation", + "content": "\\theta_0 = \\hat{\\theta}_0 = H_\\phi (0,\\mathbf{x})" + }, + { + "bbox": [ + 55, + 416, + 296, + 465 + ], + "type": "text", + "content": ", the entire trajectory can simply be matched via minimizing" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 65, + 471, + 294, + 487 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 471, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 65, + 471, + 294, + 487 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\Delta} (\\mathcal {D}) = \\mathbb {E} _ {\\mathbf {x} \\sim \\mathcal {D}, t \\sim U [ 0, T ]} \\| \\theta_ {t + 1} - H _ {\\phi} (\\mathbf {x}, t + 1) \\| ^ {2}, \\tag {3}", + "image_path": "8744d7b551391ab236195cf2426d9d6bbefcffaf1225cf1768545fafde4d484c.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 492, + 296, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 492, + 296, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 492, + 296, + 563 + ], + "type": "text", + "content": "which we name as the gradient matching loss. This loss function essentially minimizes the difference between the gradient step from Eq. (2) and the numerical gradient " + }, + { + "bbox": [ + 55, + 492, + 296, + 563 + ], + "type": "inline_equation", + "content": "\\nabla_t H_\\phi(\\mathbf{x}, t) = H_\\phi(\\mathbf{x}, t + 1) - H_\\phi(\\mathbf{x}, t)" + }, + { + "bbox": [ + 55, + 492, + 296, + 563 + ], + "type": "text", + "content": " over the entire dataset for " + }, + { + "bbox": [ + 55, + 492, + 296, + 563 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 492, + 296, + 563 + ], + "type": "text", + "content": " steps of optimization; thus replicating the underlying training dynamics of the task-specific network." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 563, + 296, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 296, + 612 + ], + "type": "text", + "content": "Once trained, the hypernetwork field " + }, + { + "bbox": [ + 55, + 563, + 296, + 612 + ], + "type": "inline_equation", + "content": "H_{\\phi}(t,\\mathbf{x})" + }, + { + "bbox": [ + 55, + 563, + 296, + 612 + ], + "type": "text", + "content": " thus mimics the entire training trajectory of the task-specific network, instead of the point estimate " + }, + { + "bbox": [ + 55, + 563, + 296, + 612 + ], + "type": "inline_equation", + "content": "\\theta^{*}(\\mathbf{x})" + }, + { + "bbox": [ + 55, + 563, + 296, + 612 + ], + "type": "text", + "content": ". We summarize this training process in Fig. 2." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 614, + 296, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 614, + 296, + 686 + ], + "spans": [ + { + "bbox": [ + 55, + 614, + 296, + 686 + ], + "type": "text", + "content": "Ensuring a consistent trajectory throughout training. To ensure training stability, we fix the starting position " + }, + { + "bbox": [ + 55, + 614, + 296, + 686 + ], + "type": "inline_equation", + "content": "\\theta_0" + }, + { + "bbox": [ + 55, + 614, + 296, + 686 + ], + "type": "text", + "content": " for each sample " + }, + { + "bbox": [ + 55, + 614, + 296, + 686 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 614, + 296, + 686 + ], + "type": "text", + "content": " throughout training, initializing it randomly only once at the beginning. This approach mirrors the typical neural network initialization to random values, followed by convergence toward the optimal parameters." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Forcing 0 offset at " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " by construction. Furthermore, Eq. (3) matches the entire trajectory only when " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\theta_0 = \\hat{\\theta}_0 =" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 315, + 210, + 517, + 222 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 210, + 517, + 222 + ], + "spans": [ + { + "bbox": [ + 315, + 210, + 517, + 222 + ], + "type": "text", + "content": "Algorithm 1 Hypernet Fields Training Algorithm" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 319, + 225, + 553, + 361 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 319, + 225, + 381, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 225, + 381, + 237 + ], + "spans": [ + { + "bbox": [ + 319, + 225, + 381, + 237 + ], + "type": "text", + "content": "1: Initialize " + }, + { + "bbox": [ + 319, + 225, + 381, + 237 + ], + "type": "inline_equation", + "content": "\\theta_0" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 320, + 238, + 553, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 238, + 553, + 261 + ], + "spans": [ + { + "bbox": [ + 320, + 238, + 553, + 261 + ], + "type": "text", + "content": "2: " + }, + { + "bbox": [ + 320, + 238, + 553, + 261 + ], + "type": "inline_equation", + "content": "H_{\\phi}\\gets" + }, + { + "bbox": [ + 320, + 238, + 553, + 261 + ], + "type": "text", + "content": " Hypernetwork() " + }, + { + "bbox": [ + 320, + 238, + 553, + 261 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 320, + 238, + 553, + 261 + ], + "type": "text", + "content": " Hypernetwork with parameters " + }, + { + "bbox": [ + 320, + 238, + 553, + 261 + ], + "type": "inline_equation", + "content": "\\phi" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 262, + 450, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 262, + 450, + 273 + ], + "spans": [ + { + "bbox": [ + 320, + 262, + 450, + 273 + ], + "type": "text", + "content": "3: while not converged do" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 274, + 553, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 274, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 320, + 274, + 553, + 285 + ], + "type": "text", + "content": "4: " + }, + { + "bbox": [ + 320, + 274, + 553, + 285 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\gets \\mathrm{next}(\\mathcal{D}) \\triangleright" + }, + { + "bbox": [ + 320, + 274, + 553, + 285 + ], + "type": "text", + "content": " Sample condition " + }, + { + "bbox": [ + 320, + 274, + 553, + 285 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 320, + 274, + 553, + 285 + ], + "type": "text", + "content": " from dataset" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 286, + 443, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 286, + 443, + 297 + ], + "spans": [ + { + "bbox": [ + 320, + 286, + 443, + 297 + ], + "type": "text", + "content": "5: " + }, + { + "bbox": [ + 320, + 286, + 443, + 297 + ], + "type": "inline_equation", + "content": "t \\gets" + }, + { + "bbox": [ + 320, + 286, + 443, + 297 + ], + "type": "text", + "content": " sample from " + }, + { + "bbox": [ + 320, + 286, + 443, + 297 + ], + "type": "inline_equation", + "content": "[0, T)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 298, + 553, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 298, + 553, + 311 + ], + "spans": [ + { + "bbox": [ + 320, + 298, + 553, + 311 + ], + "type": "text", + "content": "6: " + }, + { + "bbox": [ + 320, + 298, + 553, + 311 + ], + "type": "inline_equation", + "content": "\\hat{\\theta}_t\\gets H_\\phi (\\mathbf{x},t)" + }, + { + "bbox": [ + 320, + 298, + 553, + 311 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 320, + 298, + 553, + 311 + ], + "type": "text", + "content": " Estimate weights for timestep " + }, + { + "bbox": [ + 320, + 298, + 553, + 311 + ], + "type": "inline_equation", + "content": "t" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 312, + 553, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 312, + 553, + 324 + ], + "spans": [ + { + "bbox": [ + 320, + 312, + 553, + 324 + ], + "type": "text", + "content": "7: " + }, + { + "bbox": [ + 320, + 312, + 553, + 324 + ], + "type": "inline_equation", + "content": "\\Delta \\theta_{t} \\gets -\\eta \\nabla_{\\theta_{t}} \\mathcal{L}_{task}(\\hat{\\theta}_{t}, \\mathbf{x})" + }, + { + "bbox": [ + 320, + 312, + 553, + 324 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 320, + 312, + 553, + 324 + ], + "type": "text", + "content": " GT gradient for " + }, + { + "bbox": [ + 320, + 312, + 553, + 324 + ], + "type": "inline_equation", + "content": "\\hat{\\theta}_{t}" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 324, + 553, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 324, + 553, + 337 + ], + "spans": [ + { + "bbox": [ + 320, + 324, + 553, + 337 + ], + "type": "text", + "content": "8: " + }, + { + "bbox": [ + 320, + 324, + 553, + 337 + ], + "type": "inline_equation", + "content": "\\Delta \\hat{\\theta}_t\\gets H_\\phi (\\mathbf{x},t + 1) - \\hat{\\theta}_t" + }, + { + "bbox": [ + 320, + 324, + 553, + 337 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 320, + 324, + 553, + 337 + ], + "type": "text", + "content": " Est gradient for " + }, + { + "bbox": [ + 320, + 324, + 553, + 337 + ], + "type": "inline_equation", + "content": "\\hat{\\theta}_t" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 337, + 553, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 337, + 553, + 350 + ], + "spans": [ + { + "bbox": [ + 320, + 337, + 553, + 350 + ], + "type": "text", + "content": "9: " + }, + { + "bbox": [ + 320, + 337, + 553, + 350 + ], + "type": "inline_equation", + "content": "\\phi \\gets \\phi -\\eta \\nabla_{\\phi}\\mathrm{MSE}(\\Delta \\theta_t,\\Delta \\hat{\\theta}_t)" + }, + { + "bbox": [ + 320, + 337, + 553, + 350 + ], + "type": "text", + "content": " Update " + }, + { + "bbox": [ + 320, + 337, + 553, + 350 + ], + "type": "inline_equation", + "content": "\\phi" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 319, + 351, + 375, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 351, + 375, + 361 + ], + "spans": [ + { + "bbox": [ + 319, + 351, + 375, + 361 + ], + "type": "text", + "content": "10: end while" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 381, + 555, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 381, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 381, + 555, + 430 + ], + "type": "inline_equation", + "content": "H_{\\phi}(0,\\mathbf{x})" + }, + { + "bbox": [ + 313, + 381, + 555, + 430 + ], + "type": "text", + "content": ". Enforcing this, however, can easily be done by construction. Inspired by [21, 46], for any timestep " + }, + { + "bbox": [ + 313, + 381, + 555, + 430 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 381, + 555, + 430 + ], + "type": "text", + "content": ", we parameterize the model with an offset from the input " + }, + { + "bbox": [ + 313, + 381, + 555, + 430 + ], + "type": "inline_equation", + "content": "\\theta_0" + }, + { + "bbox": [ + 313, + 381, + 555, + 430 + ], + "type": "text", + "content": ", conditioned on " + }, + { + "bbox": [ + 313, + 381, + 555, + 430 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 381, + 555, + 430 + ], + "type": "text", + "content": " as," + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 367, + 434, + 555, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 434, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 367, + 434, + 555, + 456 + ], + "type": "interline_equation", + "content": "H _ {\\phi} (\\mathbf {x}, t) = \\theta_ {0} + \\frac {t}{T} \\times H _ {\\phi} ^ {\\prime} (\\mathbf {x}, t), \\tag {4}", + "image_path": "582983c529edc3ab46a5aecd527229a71928f7151c34447170351304d582408f.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 460, + 555, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 555, + 484 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 555, + 484 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 460, + 555, + 484 + ], + "type": "inline_equation", + "content": "H'" + }, + { + "bbox": [ + 313, + 460, + 555, + 484 + ], + "type": "text", + "content": " is the raw hypernetwork field output from a neural network." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 314, + 491, + 386, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 491, + 386, + 504 + ], + "spans": [ + { + "bbox": [ + 314, + 491, + 386, + 504 + ], + "type": "text", + "content": "3.3. Algorithm" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 509, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 555, + 689 + ], + "type": "text", + "content": "We summarize our overall optimization process in Algorithm 1. We first randomly initialize and fix the parameters " + }, + { + "bbox": [ + 313, + 509, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\theta_0" + }, + { + "bbox": [ + 313, + 509, + 555, + 689 + ], + "type": "text", + "content": ". We then consider a random sample " + }, + { + "bbox": [ + 313, + 509, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 509, + 555, + 689 + ], + "type": "text", + "content": " from the dataset and a random timestep " + }, + { + "bbox": [ + 313, + 509, + 555, + 689 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 509, + 555, + 689 + ], + "type": "text", + "content": " from the optimization trajectory. With these inputs, we use the hypernetwork " + }, + { + "bbox": [ + 313, + 509, + 555, + 689 + ], + "type": "inline_equation", + "content": "H_{\\phi}" + }, + { + "bbox": [ + 313, + 509, + 555, + 689 + ], + "type": "text", + "content": " to obtain the weights " + }, + { + "bbox": [ + 313, + 509, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\hat{\\theta}_t" + }, + { + "bbox": [ + 313, + 509, + 555, + 689 + ], + "type": "text", + "content": " of the task-specific network. Next, using these weights and the input " + }, + { + "bbox": [ + 313, + 509, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 509, + 555, + 689 + ], + "type": "text", + "content": " we compute the gradient step " + }, + { + "bbox": [ + 313, + 509, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\Delta \\theta_t" + }, + { + "bbox": [ + 313, + 509, + 555, + 689 + ], + "type": "text", + "content": " for the task-specific network. Additionally, we calculate " + }, + { + "bbox": [ + 313, + 509, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\Delta \\hat{\\theta}_t = \\hat{\\theta}_{t + 1} - \\hat{\\theta}_t" + }, + { + "bbox": [ + 313, + 509, + 555, + 689 + ], + "type": "text", + "content": " from the hypernetwork. The parameters " + }, + { + "bbox": [ + 313, + 509, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 509, + 555, + 689 + ], + "type": "text", + "content": " of the hypernetwork are then updated to match " + }, + { + "bbox": [ + 313, + 509, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\Delta \\hat{\\theta}_t" + }, + { + "bbox": [ + 313, + 509, + 555, + 689 + ], + "type": "text", + "content": " with the gradients " + }, + { + "bbox": [ + 313, + 509, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\Delta \\theta_t" + }, + { + "bbox": [ + 313, + 509, + 555, + 689 + ], + "type": "text", + "content": ", which is equivalent to minimizing Eq. (3). This yields a computationally efficient framework where a single hypernetwork encodes the optimization trajectory and can be queried to get the weights of the task-specific network at any stage in the training process." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "For inference, consider DreamBooth [39] as a task-specific network (Fig. 2). To generate personalized images" + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22132" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 95, + 179, + 218 + ], + "blocks": [ + { + "bbox": [ + 91, + 79, + 145, + 91 + ], + "lines": [ + { + "bbox": [ + 91, + 79, + 145, + 91 + ], + "spans": [ + { + "bbox": [ + 91, + 79, + 145, + 91 + ], + "type": "text", + "content": "Input Image" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 57, + 95, + 179, + 218 + ], + "lines": [ + { + "bbox": [ + 57, + 95, + 179, + 218 + ], + "spans": [ + { + "bbox": [ + 57, + 95, + 179, + 218 + ], + "type": "image", + "image_path": "bf6635a084243a7e83885b702e461a6136b646190890aab57da0c33a4ff58fb6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 187, + 95, + 252, + 159 + ], + "blocks": [ + { + "bbox": [ + 197, + 85, + 242, + 94 + ], + "lines": [ + { + "bbox": [ + 197, + 85, + 242, + 94 + ], + "spans": [ + { + "bbox": [ + 197, + 85, + 242, + 94 + ], + "type": "text", + "content": "As a drawing" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 187, + 95, + 252, + 159 + ], + "lines": [ + { + "bbox": [ + 187, + 95, + 252, + 159 + ], + "spans": [ + { + "bbox": [ + 187, + 95, + 252, + 159 + ], + "type": "image", + "image_path": "bc84799f410f4b4d5c0fafc035c30dbe39c926db69a31dff0990bf6b97f3c245.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 253, + 95, + 319, + 159 + ], + "blocks": [ + { + "bbox": [ + 270, + 75, + 306, + 93 + ], + "lines": [ + { + "bbox": [ + 270, + 75, + 306, + 93 + ], + "spans": [ + { + "bbox": [ + 270, + 75, + 306, + 93 + ], + "type": "text", + "content": "As a pixar character" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 253, + 95, + 319, + 159 + ], + "lines": [ + { + "bbox": [ + 253, + 95, + 319, + 159 + ], + "spans": [ + { + "bbox": [ + 253, + 95, + 319, + 159 + ], + "type": "image", + "image_path": "3528cc8e9e422dc4f24ac3d126460458a8db90771f530c8ddaad2a1431313b1f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 322, + 95, + 387, + 159 + ], + "blocks": [ + { + "bbox": [ + 335, + 75, + 378, + 93 + ], + "lines": [ + { + "bbox": [ + 335, + 75, + 378, + 93 + ], + "spans": [ + { + "bbox": [ + 335, + 75, + 378, + 93 + ], + "type": "text", + "content": "As a graffiti mural" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 322, + 95, + 387, + 159 + ], + "lines": [ + { + "bbox": [ + 322, + 95, + 387, + 159 + ], + "spans": [ + { + "bbox": [ + 322, + 95, + 387, + 159 + ], + "type": "image", + "image_path": "faa5ad6e1a00a651d2723c148b8928493e5aabc8d7fa72d0fcd1f165ca819c94.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 390, + 95, + 454, + 159 + ], + "blocks": [ + { + "bbox": [ + 400, + 75, + 447, + 93 + ], + "lines": [ + { + "bbox": [ + 400, + 75, + 447, + 93 + ], + "spans": [ + { + "bbox": [ + 400, + 75, + 447, + 93 + ], + "type": "text", + "content": "Flowing river of lost souls" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 390, + 95, + 454, + 159 + ], + "lines": [ + { + "bbox": [ + 390, + 95, + 454, + 159 + ], + "spans": [ + { + "bbox": [ + 390, + 95, + 454, + 159 + ], + "type": "image", + "image_path": "42dee81f7839c0a8412a0c19424cfd12df7bd0fa67f1fac509f26c8f52290eaf.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 457, + 95, + 522, + 159 + ], + "blocks": [ + { + "bbox": [ + 465, + 75, + 518, + 94 + ], + "lines": [ + { + "bbox": [ + 465, + 75, + 518, + 94 + ], + "spans": [ + { + "bbox": [ + 465, + 75, + 518, + 94 + ], + "type": "text", + "content": "As a funkopo pop figure" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 457, + 95, + 522, + 159 + ], + "lines": [ + { + "bbox": [ + 457, + 95, + 522, + 159 + ], + "spans": [ + { + "bbox": [ + 457, + 95, + 522, + 159 + ], + "type": "image", + "image_path": "470027181c8788def4962a2cb500abaecdf08e3b0065ecb81162ccd62031193a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 527, + 116, + 537, + 137 + ], + "lines": [ + { + "bbox": [ + 527, + 116, + 537, + 137 + ], + "spans": [ + { + "bbox": [ + 527, + 116, + 537, + 137 + ], + "type": "text", + "content": "sno" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 187, + 162, + 252, + 227 + ], + "blocks": [ + { + "bbox": [ + 187, + 162, + 252, + 227 + ], + "lines": [ + { + "bbox": [ + 187, + 162, + 252, + 227 + ], + "spans": [ + { + "bbox": [ + 187, + 162, + 252, + 227 + ], + "type": "image", + "image_path": "bae4e155d601fde33a7b29b195a91da186fa1ad9bdacfa92f010f860b3c140c3.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 253, + 162, + 319, + 227 + ], + "blocks": [ + { + "bbox": [ + 253, + 162, + 319, + 227 + ], + "lines": [ + { + "bbox": [ + 253, + 162, + 319, + 227 + ], + "spans": [ + { + "bbox": [ + 253, + 162, + 319, + 227 + ], + "type": "image", + "image_path": "1acee2229fa132d679380b66f722ceadf163a3e06fecbe1798566c94e048a9e4.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 321, + 162, + 386, + 226 + ], + "blocks": [ + { + "bbox": [ + 321, + 162, + 386, + 226 + ], + "lines": [ + { + "bbox": [ + 321, + 162, + 386, + 226 + ], + "spans": [ + { + "bbox": [ + 321, + 162, + 386, + 226 + ], + "type": "image", + "image_path": "4c656a8478eac94baea4a5f46a8ec68b5644228a82dd1f1edda604b9979e8a3d.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 390, + 162, + 454, + 227 + ], + "blocks": [ + { + "bbox": [ + 390, + 162, + 454, + 227 + ], + "lines": [ + { + "bbox": [ + 390, + 162, + 454, + 227 + ], + "spans": [ + { + "bbox": [ + 390, + 162, + 454, + 227 + ], + "type": "image", + "image_path": "3c8c531c0d9339fec94b4eaae0a4d62e05a79fe2b46cb15ecc6325dbbaee5298.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 457, + 162, + 521, + 227 + ], + "blocks": [ + { + "bbox": [ + 527, + 175, + 537, + 211 + ], + "lines": [ + { + "bbox": [ + 527, + 175, + 537, + 211 + ], + "spans": [ + { + "bbox": [ + 527, + 175, + 537, + 211 + ], + "type": "text", + "content": "Textual" + } + ] + } + ], + "index": 13, + "angle": 270, + "type": "image_caption" + }, + { + "bbox": [ + 541, + 174, + 550, + 215 + ], + "lines": [ + { + "bbox": [ + 541, + 174, + 550, + 215 + ], + "spans": [ + { + "bbox": [ + 541, + 174, + 550, + 215 + ], + "type": "text", + "content": "Inversion" + } + ] + } + ], + "index": 14, + "angle": 270, + "type": "image_caption" + }, + { + "bbox": [ + 457, + 162, + 521, + 227 + ], + "lines": [ + { + "bbox": [ + 457, + 162, + 521, + 227 + ], + "spans": [ + { + "bbox": [ + 457, + 162, + 521, + 227 + ], + "type": "image", + "image_path": "b6a744410b5976439a5c4a79353521ee8dbf4f8a76c6fc2cc4329a1a326533f5.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 187, + 229, + 252, + 294 + ], + "blocks": [ + { + "bbox": [ + 187, + 229, + 252, + 294 + ], + "lines": [ + { + "bbox": [ + 187, + 229, + 252, + 294 + ], + "spans": [ + { + "bbox": [ + 187, + 229, + 252, + 294 + ], + "type": "image", + "image_path": "180ff43a4471cf43ea816e1243b77b932d198909a60a6b4e64725978f9724516.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 304, + 555, + 348 + ], + "lines": [ + { + "bbox": [ + 55, + 304, + 555, + 348 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 555, + 348 + ], + "type": "text", + "content": "Figure 3. Qualitative examples (CelebA HQ) - Qualitative examples of personalized human face generation using the CelebA HQ dataset [20, 25] are shown. Our hypernetwork field achieves fast adaptation, producing results comparable to DreamBooth [39] and Textual Inversion [13] while preserving individual features in a personalized manner. Additional examples are provided in the supplementary material." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 253, + 229, + 319, + 294 + ], + "blocks": [ + { + "bbox": [ + 253, + 229, + 319, + 294 + ], + "lines": [ + { + "bbox": [ + 253, + 229, + 319, + 294 + ], + "spans": [ + { + "bbox": [ + 253, + 229, + 319, + 294 + ], + "type": "image", + "image_path": "160c7b767ca3e9d702cb1781bc89b5da22bbb54e8474e4dffb4371f6edecdf5f.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 321, + 230, + 386, + 294 + ], + "blocks": [ + { + "bbox": [ + 321, + 230, + 386, + 294 + ], + "lines": [ + { + "bbox": [ + 321, + 230, + 386, + 294 + ], + "spans": [ + { + "bbox": [ + 321, + 230, + 386, + 294 + ], + "type": "image", + "image_path": "9981e2fe21724ccb90b9229b5e656c0c9264a035f0feae44454f90f50f940834.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 390, + 229, + 454, + 294 + ], + "blocks": [ + { + "bbox": [ + 390, + 229, + 454, + 294 + ], + "lines": [ + { + "bbox": [ + 390, + 229, + 454, + 294 + ], + "spans": [ + { + "bbox": [ + 390, + 229, + 454, + 294 + ], + "type": "image", + "image_path": "ac323e7721d9bfeee374cb1fd5f5d1aa2f780b171239edbf62eebfcd6aceb659.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 457, + 229, + 522, + 294 + ], + "blocks": [ + { + "bbox": [ + 457, + 229, + 522, + 294 + ], + "lines": [ + { + "bbox": [ + 457, + 229, + 522, + 294 + ], + "spans": [ + { + "bbox": [ + 457, + 229, + 522, + 294 + ], + "type": "image", + "image_path": "967bdd4ba0b1e0d349229b9f66fea6dd6cf970614cf34580ff19e5c15b7dc857.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 527, + 233, + 537, + 289 + ], + "lines": [ + { + "bbox": [ + 527, + 233, + 537, + 289 + ], + "spans": [ + { + "bbox": [ + 527, + 233, + 537, + 289 + ], + "type": "text", + "content": "DreamBooth" + } + ] + } + ], + "index": 25, + "angle": 270, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "bbox": [ + 55, + 369, + 295, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 369, + 295, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 295, + 430 + ], + "type": "text", + "content": "for a given source image, we query the hypernetwork field with the source image and the timestep " + }, + { + "bbox": [ + 55, + 369, + 295, + 430 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 369, + 295, + 430 + ], + "type": "text", + "content": " to obtain the weights for the DreamBooth model. Using these weights and the conditioning prompt, we can generate personalized images in a single step." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 55, + 441, + 188, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 188, + 454 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 188, + 454 + ], + "type": "text", + "content": "3.4. Implementation Details" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 55, + 460, + 295, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 568 + ], + "type": "text", + "content": "We use a frozen Vision Transformer (ViT) [9] to encode the images. We then pass this encoding into a hypernetwork with a 6-layer Diffusion Transformer (DiT) [30] with adaptive layer norm conditioning [5] for both the timestep and the image embedding. We set the sequence length of the DiT is equal to the total number of layers in the task-specific network. We then have a dedicated Multi-Layer Perceptron (MLP) per each layer of the task-specific network that decodes to the network weights of the target network." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 55, + 584, + 135, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 584, + 135, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 584, + 135, + 597 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 55, + 605, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 712 + ], + "type": "text", + "content": "We consider two diverse applications - image personalization and 3D shape reconstruction as target applications. While we focus on these two, we emphasize that our hypernet fields framework does not make any assumption on the task-specific network and applies to any task. Moreover, the definition of the loss, " + }, + { + "bbox": [ + 55, + 605, + 295, + 712 + ], + "type": "inline_equation", + "content": "L_{task}(\\theta_t, c)" + }, + { + "bbox": [ + 55, + 605, + 295, + 712 + ], + "type": "text", + "content": " is not restricted to a particular domain. Thus, our method can be applied to these problems by modifying the task-specific optimization model." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 369, + 552, + 382 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 552, + 382 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 552, + 382 + ], + "type": "text", + "content": "4.1. Application to Personalized Image Generation" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 390, + 554, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 390, + 554, + 438 + ], + "spans": [ + { + "bbox": [ + 313, + 390, + 554, + 438 + ], + "type": "text", + "content": "To apply our method to personalized image generation, we follow prior work [40] and evaluate and compare personalized face generation on CelebA HQ [20, 25], as well as on animal faces with the AFHQ dataset [7]." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 313, + 444, + 555, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 444, + 555, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 444, + 555, + 624 + ], + "type": "text", + "content": "Task network - DreamBooth. To personalize and generate customized images we use DreamBooth as our task network, as in HyperDreamBooth [40]. DreamBooth builds upon the denoising diffusion probabilistic models [38, 42]. The generation process is guided by a special conditioning token encoding the characteristics of the input conditioning image, which is found by optimizing through diffusion models, which require a non-trivial amount of compute. On the CelebA dataset [25] we follow prior work [40] and use the text condition \"a [V] face\" where the token [V] represents a placeholder for the specific identity being learned. Analogously, for the AFHQ dataset [7] with animal images, we use the text condition \"a [V] animal\" to guide the model. The conditioning token along with the input prompts are then used to synthesize personalized images." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "content": "Training setup. For the CelebA-HQ dataset [20, 25], we follow prior work [40] and replicate their setup to make results comparable. Specifically, among the " + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "inline_equation", + "content": "30\\mathrm{k}" + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "content": " high-quality aligned faces, we use the train/test splits and evaluation prompts from HyperDreamBooth [40], with " + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "inline_equation", + "content": "29.9\\mathrm{k}" + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "content": " images in the training set and 100 images in the test set. For the AFHQ dataset [7] we use the live subject prompts from the" + } + ] + } + ], + "index": 35 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22133" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 97, + 199, + 238 + ], + "blocks": [ + { + "bbox": [ + 105, + 78, + 156, + 89 + ], + "lines": [ + { + "bbox": [ + 105, + 78, + 156, + 89 + ], + "spans": [ + { + "bbox": [ + 105, + 78, + 156, + 89 + ], + "type": "text", + "content": "Input Image" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 56, + 97, + 199, + 238 + ], + "lines": [ + { + "bbox": [ + 56, + 97, + 199, + 238 + ], + "spans": [ + { + "bbox": [ + 56, + 97, + 199, + 238 + ], + "type": "image", + "image_path": "d37fc23d057773e660e30ad2344787d7e0a70fa6271e93b2c7605ea59110ab43.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 206, + 95, + 284, + 175 + ], + "blocks": [ + { + "bbox": [ + 225, + 72, + 265, + 93 + ], + "lines": [ + { + "bbox": [ + 225, + 72, + 265, + 93 + ], + "spans": [ + { + "bbox": [ + 225, + 72, + 265, + 93 + ], + "type": "text", + "content": "Wearing a Santa hat" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 206, + 95, + 284, + 175 + ], + "lines": [ + { + "bbox": [ + 206, + 95, + 284, + 175 + ], + "spans": [ + { + "bbox": [ + 206, + 95, + 284, + 175 + ], + "type": "image", + "image_path": "740ea11955a3f836234c18ec8363fbac2ba1e63670fe667ba45deddfea599964.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 206, + 176, + 283, + 255 + ], + "blocks": [ + { + "bbox": [ + 206, + 176, + 283, + 255 + ], + "lines": [ + { + "bbox": [ + 206, + 176, + 283, + 255 + ], + "spans": [ + { + "bbox": [ + 206, + 176, + 283, + 255 + ], + "type": "image", + "image_path": "0f1ddb8fb62e9660f1fdfb3030a5f5d1fbbb1334c2496fae8e7b5c61c4cc0854.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 206, + 257, + 283, + 334 + ], + "blocks": [ + { + "bbox": [ + 206, + 257, + 283, + 334 + ], + "lines": [ + { + "bbox": [ + 206, + 257, + 283, + 334 + ], + "spans": [ + { + "bbox": [ + 206, + 257, + 283, + 334 + ], + "type": "image", + "image_path": "8d54b4a78a62ab2973794d8cae98b2258adf95f0fee00522a98944f5a4d09a1e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 347, + 555, + 381 + ], + "lines": [ + { + "bbox": [ + 55, + 347, + 555, + 381 + ], + "spans": [ + { + "bbox": [ + 55, + 347, + 555, + 381 + ], + "type": "text", + "content": "Figure 4. Qualitative examples (AHFQ) - Qualitative results of personalized animal image generation from the AFHQ dataset [7] demonstrate the adaptability of our hypernetwork field. Compared to DreamBooth [39] and Textual Inversion [13], our method effectively captures the specific animal's characteristics while enabling rapid adaptation. Further examples are available in the supplementary material." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 286, + 95, + 364, + 175 + ], + "blocks": [ + { + "bbox": [ + 298, + 72, + 352, + 95 + ], + "lines": [ + { + "bbox": [ + 298, + 72, + 352, + 95 + ], + "spans": [ + { + "bbox": [ + 298, + 72, + 352, + 95 + ], + "type": "text", + "content": "Wearing pink sunglasses" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 286, + 95, + 364, + 175 + ], + "lines": [ + { + "bbox": [ + 286, + 95, + 364, + 175 + ], + "spans": [ + { + "bbox": [ + 286, + 95, + 364, + 175 + ], + "type": "image", + "image_path": "02cf65aad15da8f3f5d0c4a482928faa7d3097d0cf57ad6655789991bad1e7f4.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 286, + 176, + 364, + 255 + ], + "blocks": [ + { + "bbox": [ + 286, + 176, + 364, + 255 + ], + "lines": [ + { + "bbox": [ + 286, + 176, + 364, + 255 + ], + "spans": [ + { + "bbox": [ + 286, + 176, + 364, + 255 + ], + "type": "image", + "image_path": "1093262a36b2a1323f7b7ea4802c725a183eff9e527e079fae21558d1d101a22.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 286, + 257, + 364, + 334 + ], + "blocks": [ + { + "bbox": [ + 286, + 257, + 364, + 334 + ], + "lines": [ + { + "bbox": [ + 286, + 257, + 364, + 334 + ], + "spans": [ + { + "bbox": [ + 286, + 257, + 364, + 334 + ], + "type": "image", + "image_path": "726c0f2dbfb85a2bf059a428c7989e141812f35db840385e31f05520f2cad163.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 365, + 95, + 444, + 175 + ], + "blocks": [ + { + "bbox": [ + 376, + 72, + 435, + 94 + ], + "lines": [ + { + "bbox": [ + 376, + 72, + 435, + 94 + ], + "spans": [ + { + "bbox": [ + 376, + 72, + 435, + 94 + ], + "type": "text", + "content": "Wearing a rainbow scarf" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 365, + 95, + 444, + 175 + ], + "lines": [ + { + "bbox": [ + 365, + 95, + 444, + 175 + ], + "spans": [ + { + "bbox": [ + 365, + 95, + 444, + 175 + ], + "type": "image", + "image_path": "3e00a45e0053026bfb32743865769cbc566e3b94189fe707b9750685f87c37c2.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 365, + 176, + 443, + 254 + ], + "blocks": [ + { + "bbox": [ + 365, + 176, + 443, + 254 + ], + "lines": [ + { + "bbox": [ + 365, + 176, + 443, + 254 + ], + "spans": [ + { + "bbox": [ + 365, + 176, + 443, + 254 + ], + "type": "image", + "image_path": "b2d30e095c7807e0a39bd222b9aa1e31095eacbc11194ff991e3b7da35d0f7fa.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 365, + 256, + 443, + 334 + ], + "blocks": [ + { + "bbox": [ + 365, + 256, + 443, + 334 + ], + "lines": [ + { + "bbox": [ + 365, + 256, + 443, + 334 + ], + "spans": [ + { + "bbox": [ + 365, + 256, + 443, + 334 + ], + "type": "image", + "image_path": "85f362426127572ed2ac8dbcbbc8042e062dbbd5be08c8067fc8e2e010fce0bb.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 446, + 95, + 524, + 175 + ], + "blocks": [ + { + "bbox": [ + 462, + 85, + 509, + 94 + ], + "lines": [ + { + "bbox": [ + 462, + 85, + 509, + 94 + ], + "spans": [ + { + "bbox": [ + 462, + 85, + 509, + 94 + ], + "type": "text", + "content": "In the snow" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 446, + 95, + 524, + 175 + ], + "lines": [ + { + "bbox": [ + 446, + 95, + 524, + 175 + ], + "spans": [ + { + "bbox": [ + 446, + 95, + 524, + 175 + ], + "type": "image", + "image_path": "098c1d4258d257a247900c6a2c80493ba1ab6c2fdf76cbfbd318e5e34e05c985.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 526, + 124, + 536, + 145 + ], + "lines": [ + { + "bbox": [ + 526, + 124, + 536, + 145 + ], + "spans": [ + { + "bbox": [ + 526, + 124, + 536, + 145 + ], + "type": "text", + "content": "o" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 446, + 176, + 524, + 255 + ], + "blocks": [ + { + "bbox": [ + 446, + 176, + 524, + 255 + ], + "lines": [ + { + "bbox": [ + 446, + 176, + 524, + 255 + ], + "spans": [ + { + "bbox": [ + 446, + 176, + 524, + 255 + ], + "type": "image", + "image_path": "12e0a3ecbac55668b920bdbe44f33d480c877bbc73be130dcdf31873b634ac29.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 525, + 198, + 535, + 232 + ], + "lines": [ + { + "bbox": [ + 525, + 198, + 535, + 232 + ], + "spans": [ + { + "bbox": [ + 525, + 198, + 535, + 232 + ], + "type": "text", + "content": "Textual" + } + ] + } + ], + "index": 19, + "angle": 270, + "type": "image_caption" + }, + { + "bbox": [ + 537, + 196, + 547, + 233 + ], + "lines": [ + { + "bbox": [ + 537, + 196, + 547, + 233 + ], + "spans": [ + { + "bbox": [ + 537, + 196, + 547, + 233 + ], + "type": "text", + "content": "inversion" + } + ] + } + ], + "index": 20, + "angle": 270, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 446, + 256, + 524, + 334 + ], + "blocks": [ + { + "bbox": [ + 446, + 256, + 524, + 334 + ], + "lines": [ + { + "bbox": [ + 446, + 256, + 524, + 334 + ], + "spans": [ + { + "bbox": [ + 446, + 256, + 524, + 334 + ], + "type": "image", + "image_path": "bed2e2b4c033a6f67d2971ddbf5c8dfa59f7277e581422f50f162e28c744a834.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 402, + 295, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 402, + 295, + 474 + ], + "spans": [ + { + "bbox": [ + 55, + 402, + 295, + 474 + ], + "type": "text", + "content": "DreamBooth dataset [39] so that the results are comparable. Among the 16.1k aligned animal faces, we use the established train and test splits, leaving 14.6k images for training and 1.5k for testing. In addition to the DreamBooth and HyperDreamBooth approach, we compare against Textual Inversion [13]." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 55, + 477, + 295, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 477, + 295, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 477, + 295, + 597 + ], + "type": "text", + "content": "For both experiments, we set the number of optimization steps to be " + }, + { + "bbox": [ + 55, + 477, + 295, + 597 + ], + "type": "inline_equation", + "content": "T = 500" + }, + { + "bbox": [ + 55, + 477, + 295, + 597 + ], + "type": "text", + "content": ". Our network is trained to output rank-1 Low-Rank Adaptation (LoRA) [18] weights for both the UNet and the text encoder for Stable Diffusion 1.5 [38], which amounts to 273k parameters across 352 layers. We train our model with a batch size of 256 and for 20k iterations on 8 NVidia A100 GPUs, which takes 1.5 days. Similar to the HyperDreamBooth [40] approach once our model is trained, we can perform 50 steps of fast fine-tuning starting from the hypernetwork outputs." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 55, + 600, + 295, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 295, + 660 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 295, + 660 + ], + "type": "text", + "content": "Note that, unlike existing works that require per-instance fitting, with our method, it is possible to introduce on-the-fly data augmentation. We thus incorporate color, saturation, contrast, and hue augmentations to perturb the appearance of the conditioning signal " + }, + { + "bbox": [ + 55, + 600, + 295, + 660 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 600, + 295, + 660 + ], + "type": "text", + "content": ", including horizontal flips." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "content": "Evaluation metrics. We use the standard metrics suggested by DreamBooth [40]: CLIP-T, CLIP-I [34], and DINO [6]. CLIP-T measures the alignment between the generated image and the textual prompt, while CLIP-I and DINO eval" + } + ] + } + ], + "index": 25 + }, + { + "type": "table", + "bbox": [ + 319, + 400, + 552, + 466 + ], + "blocks": [ + { + "bbox": [ + 319, + 400, + 552, + 466 + ], + "lines": [ + { + "bbox": [ + 319, + 400, + 552, + 466 + ], + "spans": [ + { + "bbox": [ + 319, + 400, + 552, + 466 + ], + "type": "table", + "html": "
MethodFace Rec.↑DINO↑CLIP-I↑CLIP-T↑
Hyper-DreamBooth0.6550.4730.5770.286
DreamBooth0.6180.4410.5460.282
Textual Inversion0.6230.2890.4720.277
Ours0.3250.6050.6390.268
", + "image_path": "ba2ce9e7805acc7c7345aa9fb4f8309e7aa606a0682b83f771cd0a2941ec7b62.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "table_body" + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 475, + 554, + 529 + ], + "lines": [ + { + "bbox": [ + 313, + 475, + 554, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 554, + 529 + ], + "type": "text", + "content": "Table 1. Quantitative results (CelebA HQ) - We report the standard metrics for our method and the baselines. Our method is able to offer comparable performance, with significantly less compute. In terms of the image alignment metrics, our method significantly outperforms other methods." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 553, + 554, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 553, + 554, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 553, + 554, + 601 + ], + "type": "text", + "content": "uate the visual consistency of the generated image with the reference object's appearance. We also include the identity preservation metric (Face Rec.) suggested by HyperDream-Booth [40]." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": "Compute time. As noted earlier, our method requires 1.5 days to train with 8 GPUs for the AFHQ and CelebA datasets, with around " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "14.6\\mathrm{k}" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": " samples for AFHQ and " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "29.9\\mathrm{k}" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": " samples for CelebA, amounting to 12 days of GPU time. This is much faster compared to the training of HyperDreamBooth [40] that requires 15k DreamBooth weights to be precomputed, which amounts to 50 days of GPU time, before their hypernetwork can start training. Thus, our method trains in less than a quarter of the pre-processing" + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22134" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 296, + 131 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 296, + 131 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 296, + 131 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 296, + 131 + ], + "type": "table", + "html": "
MethodDINO↑CLIP-I↑CLIP-T↑
Textual Inversion0.5960.7800.285
DreamBooth0.5600.7630.268
Ours0.6640.8070.277
", + "image_path": "ca7379a1512158e9eb1ab24153dcdfaedffb30bb0fcc6009fa52b203147a152e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 140, + 296, + 206 + ], + "lines": [ + { + "bbox": [ + 55, + 140, + 296, + 206 + ], + "spans": [ + { + "bbox": [ + 55, + 140, + 296, + 206 + ], + "type": "text", + "content": "Table 2. Quantitative results (AFHQ) - We outperform other baselines in terms of image alignment and provide comparable results in terms of CLIP-T, the prompt alignment. We note that our method outperforms DreamBooth [39] for all metrics, benefiting from the fact that our HyperNet Field learns a general trend of how DreamBooth weights behave." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 63, + 217, + 289, + 373 + ], + "blocks": [ + { + "bbox": [ + 63, + 217, + 289, + 373 + ], + "lines": [ + { + "bbox": [ + 63, + 217, + 289, + 373 + ], + "spans": [ + { + "bbox": [ + 63, + 217, + 289, + 373 + ], + "type": "image", + "image_path": "51dc2064d770251ca7022afc7ca6a59b081c97cd5494c04410fa8943cfb8faac.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 380, + 296, + 437 + ], + "lines": [ + { + "bbox": [ + 55, + 380, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 380, + 296, + 437 + ], + "type": "text", + "content": "Figure 5. Fast training (AFHQ) - We show example outcomes of our method only after 100 iterations of training, as raw output from our Hypernetwork without fast fine-tuning. As shown, even after only 100 iterations, our model is able to generate images that are visually very similar to the conditioning images." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 458, + 169, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 458, + 169, + 470 + ], + "spans": [ + { + "bbox": [ + 55, + 458, + 169, + 470 + ], + "type": "text", + "content": "time of HyperDreamBooth." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 471, + 296, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 471, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 471, + 296, + 555 + ], + "type": "text", + "content": "Since we query our Hypernetwork at a single optimization step " + }, + { + "bbox": [ + 55, + 471, + 296, + 555 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 471, + 296, + 555 + ], + "type": "text", + "content": ", our method also yields efficient inference. Inference of our method takes only 0.3 seconds, compared to 2 seconds for HyperDreamBooth [40] and 5 minutes for DreamBooth. Additionally, the fast fine-tuning required by both our method and HyperDreamBooth takes just 20 seconds, allowing for quick adaptation with minimal overhead." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "text", + "content": "Results. We provide qualitative highlights in Figs. 3 and 4, and the quantitative summaries in Tables 1 and 2. Our results for image personalization in Figs. 3 and 4 is on par with the target Dreambooth model whose weights we estimate with our HyperNet Fields approach. That is, our approach generates personalized images that are representative of the given input prompt and the conditional image. Again, note that we achieve these personalization results in a single step by querying the hypernetwork once at timestep " + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "text", + "content": ". Interestingly, our method demonstrates remarkable performance for preserving the original image context, as measured by DINO and CLIP-I scores (Tab. 1). Further, in the case of the AFHQ dataset (Tab. 2), we outperform" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 555, + 262 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 555, + 262 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 555, + 262 + ], + "type": "image", + "image_path": "622f847ef9789721e14016289303b3cef1492280a23a9b1e8a5d3561ee2d4e94.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 272, + 555, + 318 + ], + "lines": [ + { + "bbox": [ + 313, + 272, + 555, + 318 + ], + "spans": [ + { + "bbox": [ + 313, + 272, + 555, + 318 + ], + "type": "text", + "content": "Figure 6. Qualitative examples (3D Shapes) – We show example shape reconstruction results converting (left) point clouds to 3D shapes and (right) images to 3D shapes. More qualitative examples are available in the supplementary material." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 338, + 555, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 555, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 555, + 434 + ], + "type": "text", + "content": "all DreamBooth variants, demonstrating that not only do we remove the compute-heavy optimization requirement of DreamBooth [39], but also by amortizing the behavior of the learned weights with respect to the input conditioning image, we successfully replicate the dynamics of the task-specific dreambooth network yielding improved results. We emphasize that while results are comparable to each other, our method is significantly faster than the baselines." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 437, + 556, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 556, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 556, + 557 + ], + "type": "text", + "content": "Fast training. An interesting observation from our training process is that our model exhibits faster adaptation to personalization. As shown in Fig. 5, after only 100 training iterations, our model generates images that are visually similar to the conditioning images, achieving this level of personalization with minimal training time. We hypothesize that this rapid adaptation is due to our model's ability to simulate the optimization pathway, effectively learning the direction in which to adjust network weights to reduce the task-specific loss." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 564, + 520, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 564, + 520, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 564, + 520, + 578 + ], + "type": "text", + "content": "4.2. Application to 3D shape reconstruction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": "We further explore the potential of our method for the task of 3D shape reconstruction, specifically the task of estimating the weights of an occupancy network [27]. Additionally, we consider two different conditioning inputs: a rendered image of the shape [24, 50], and a point cloud of the shape [1, 15, 32]. For the image conditioning, we use the 38 renderings provided in Qiu et al. [33]—covering various angles around each object—which we sample one of and encode with a pre-trained ViT [9]. For the point cloud conditioning, we randomly sample 2048 points uniformly from the surface of the ground truth mesh and encode the input point" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22135" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 553, + 155 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 155 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 155 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 155 + ], + "type": "image", + "image_path": "af6289ace7e35a25b4d13c097e058ea67458d953ad8fd060fca9f559fd843772.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 163, + 555, + 231 + ], + "lines": [ + { + "bbox": [ + 54, + 163, + 555, + 231 + ], + "spans": [ + { + "bbox": [ + 54, + 163, + 555, + 231 + ], + "type": "text", + "content": "Figure 7. Hypernet Field output example (ShapeNet) - We show the output of our HyperNet Field for various points along the optimization trajectory as well as the intersection over union (IoU) over all values of T. As shown, the HyperNet Field is able to accurately estimate the entire optimization trajectory, starting from network weights that only represent the airplane shape roughly, and converging to one that provides a detailed reconstruction of the airplane. The IoU plot also shows that as we traverse the trajectory, the IoU of the generated shapes to the ground truth shape improves, demonstrating convergence. Despite being shown only one ground truth optimization step at a time, our model is able to represent the entire optimization trajectory from initialization to convergence." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 250, + 282, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 250, + 282, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 250, + 282, + 262 + ], + "type": "text", + "content": "cloud with an off-the-shelf variational autoencoder [48]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 266, + 296, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 266, + 296, + 387 + ], + "spans": [ + { + "bbox": [ + 54, + 266, + 296, + 387 + ], + "type": "text", + "content": "Task network - Occupancy Net. For both cases, the hypernetwork field architecture remains the same as described in Sec. 3.4, and for the task-specific network, we follow an architecture similar to HyperDiffusion [12] —we use a simple 3-layer dense network [41] with ReLU activations and a hidden dimension of 128. The task-specific network, the occupancy network, takes the 3D coordinates of a query point as input and outputs a scalar value representing the occupancy as in Mescheder et al. [27]. We use marching cubes [26] to render the outcome." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 392, + 296, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 392, + 296, + 511 + ], + "spans": [ + { + "bbox": [ + 54, + 392, + 296, + 511 + ], + "type": "text", + "content": "Training setup. To supervise the training, we sample 20k query points, 10k uniformly, and 10k near the ground-truth surface to be used for training. To find the ground-truth occupancy values for these points, we first make the meshes watertight with [19] and use winding numbers [4] on the watertight mesh to determine whether each point lies inside or outside the surface, and compute the occupancy via Barill et al. [3]. We use the mean squared error (MSE) of the occupancy estimates for the task-specific loss. For all cases, we set " + }, + { + "bbox": [ + 54, + 392, + 296, + 511 + ], + "type": "inline_equation", + "content": "T = 10,000" + }, + { + "bbox": [ + 54, + 392, + 296, + 511 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 516, + 296, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 516, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 54, + 516, + 296, + 624 + ], + "type": "text", + "content": "2D image conditioning (Objaverse shapes). For 2D image conditioning, as training over the entire Objaverse dataset [8] would be computationally expensive, we simply overfit to a subset of 128 shapes, as we are mostly interested in the potential of our method. We leave training on the full Objaverse dataset as future work. We obtain reasonable shape reconstruction results as shown in Figures 1 and 6 with our method for this simple overfitting demonstration just within 10 hours of training on a single NVidia RTX 3090 GPU." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 629, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 629, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 629, + 296, + 713 + ], + "type": "text", + "content": "3D point cloud conditioning (ShapeNet shapes). For 3D point cloud conditioning, we use the entire dataset. As shown in Fig. 6, our method is able to learn to generate occupancy network weights from point clouds. We further illustrate how both our Hypernet Field output and the IoU for shape generation evolves over the optimization process in Fig. 7. As shown, our method simulates the entire train-" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 250, + 555, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 250, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 313, + 250, + 555, + 285 + ], + "type": "text", + "content": "ing process of the task-specific network for a given point cloud despite only being shown a single ground truth gradient step at a time." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 299, + 388, + 311 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 299, + 388, + 311 + ], + "spans": [ + { + "bbox": [ + 313, + 299, + 388, + 311 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 319, + 555, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 555, + 475 + ], + "type": "text", + "content": "We presented Hypernetwork Fields, a novel and efficient hypernetwork training framework. Instead of optimizing individual weights per sample, we use gradient supervision and match gradients along the hypernetworks field and the task-specific network at random points along the optimization trajectory. This drastically reduces the precomputation overhead of training hypernetworks, while maintaining comparable performance to state of the art. We have demonstrated the effectiveness of our method on personalized image generation and 3D shape reconstruction. Our framework is general and can be applied to any hypernetwork training scenario, making it a promising direction for scaling hypernetwork applications to larger datasets." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 479, + 556, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 556, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 556, + 658 + ], + "type": "text", + "content": "Limitations and future work. Our method enables stable hypernetwork training on large datasets. Because of this, the potential of our method for diverse tasks remains underexplored. This may include tasks such as training a hypernetwork to estimate the weights of NeRFs or even personalized adapters for large language models. However, our approach requires as many output neurons as there are weights in the target network, resulting in suboptimal memory usage. Recent work by Dravid et al. [10] has shown that customized DreamBooth adapters can be represented using only 1k parameters. Integrating our method with this compact representation approach could significantly enhance parameter efficiency while maintaining our proposed benefits of efficient hypernetwork training. These would be interesting avenues for future work." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 671, + 373, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 671, + 373, + 684 + ], + "spans": [ + { + "bbox": [ + 314, + 671, + 373, + 684 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 318, + 691, + 556, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 691, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 318, + 691, + 556, + 715 + ], + "type": "text", + "content": "[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitlagkas, and Leonidas Guibas. Learning representations and generative" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22136" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 77, + 73, + 243, + 83 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 243, + 83 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 243, + 83 + ], + "type": "text", + "content": "models for 3d point clouds. In ICML, 2018. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 85, + 295, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 85, + 295, + 127 + ], + "spans": [ + { + "bbox": [ + 61, + 85, + 295, + 127 + ], + "type": "text", + "content": "[2] Sudarshan Babu, Richard Liu, Avery Zhou, Michael Maire, Greg Shakhnarovich, and Rana Hanocka. Hyperfields: Towards zero-shot generation of nerfs from text. In ICML. PMLR, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 129, + 295, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 129, + 295, + 161 + ], + "spans": [ + { + "bbox": [ + 61, + 129, + 295, + 161 + ], + "type": "text", + "content": "[3] Gavin Barill, Neil G Dickson, Ryan Schmidt, David IW Levin, and Alec Jacobson. Fast winding numbers for soups and clouds. Transactions On Graphics, 37(4):1-12, 2018. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 162, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 162, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 62, + 162, + 294, + 194 + ], + "type": "text", + "content": "[4] Gavin Barill, Neil G. Dickson, Ryan M. Schmidt, David I. W. Levin, and Alec Jacobson. Fast winding numbers for soups and clouds. Transactions On Graphics, 37:1-12, 2018. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 195, + 294, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 195, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 62, + 195, + 294, + 216 + ], + "type": "text", + "content": "[5] Tom B Brown. Language models are few-shot learners. NeurIPS, 2020. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 217, + 294, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 217, + 294, + 259 + ], + "spans": [ + { + "bbox": [ + 62, + 217, + 294, + 259 + ], + "type": "text", + "content": "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, pages 9650-9660, 2021. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 261, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 261, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 62, + 261, + 294, + 293 + ], + "type": "text", + "content": "[7] Yunjoy Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis for multiple domains. In CVPR, 2020. 5, 6, 11" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 295, + 294, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 295, + 294, + 347 + ], + "spans": [ + { + "bbox": [ + 62, + 295, + 294, + 347 + ], + "type": "text", + "content": "[8] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In CVPR, pages 13142-13153, 2023. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 350, + 294, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 350, + 294, + 371 + ], + "spans": [ + { + "bbox": [ + 62, + 350, + 294, + 371 + ], + "type": "text", + "content": "[9] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 5, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 372, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 372, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 57, + 372, + 294, + 415 + ], + "type": "text", + "content": "[10] Amir Dravid, Yossi Gandelsman, Kuan-Chieh Wang, Rameen Abdal, Gordon Wetzstein, Alexei A Efros, and Kfir Aberman. Interpreting the weight space of customized diffusion models. arXiv preprint arXiv:2406.09413, 2024. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 416, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 416, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 57, + 416, + 294, + 437 + ], + "type": "text", + "content": "[11] Yilun Du and Igor Mordatch. Implicit generation and modeling with energy based models. NeurIPS, 32, 2019. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 438, + 294, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 438, + 294, + 480 + ], + "spans": [ + { + "bbox": [ + 57, + 438, + 294, + 480 + ], + "type": "text", + "content": "[12] Ziya Erkoç, Fangchang Ma, Qi Shan, Matthias Nießner, and Angela Dai. Hyperdiffusion: Generating implicit neural fields with weight-space diffusion. In ICCV, pages 14300-14310, 2023. 2, 3, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 482, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 482, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 57, + 482, + 294, + 525 + ], + "type": "text", + "content": "[13] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. *ICLR*, 2023. 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 526, + 294, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 526, + 294, + 568 + ], + "spans": [ + { + "bbox": [ + 57, + 526, + 294, + 568 + ], + "type": "text", + "content": "[14] Will Grathwohl, Kuan-Chieh Wang, Jorn-Henrik Jacobsen, David Duvenaud, Kevin Swersky, and Mohammad Norouzi. Learning energy-based models by diffusion recovery likelihood. ICLR, 2021. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 571, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 571, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 57, + 571, + 294, + 612 + ], + "type": "text", + "content": "[15] Thibault Groueix, Matthew Fisher, Vladimir G Kim, Bryan Russell, and Mathieu Aubry. Atlasnet: A papier-mâché approach to learning 3d surface generation. In CVPR, 2018. 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 614, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 614, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 614, + 294, + 635 + ], + "type": "text", + "content": "[16] David Ha, Andrew Dai, and Quoc V Le. Hypernetworks. ICLR, 2017. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 637, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 637, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 637, + 294, + 667 + ], + "type": "text", + "content": "[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, pages 6840-6851, 2020. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "type": "text", + "content": "[18] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In ICLR, 2022. 6" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[19] Jingwei Huang, Yichao Zhou, and Leonidas Guibas. Manifoldplus: A robust and scalable watertight manifold surface generation method for triangle soups. arxiv, 2020. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 108, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 553, + 139 + ], + "type": "text", + "content": "[20] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. In ICLR, 2018. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 141, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 141, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 317, + 141, + 553, + 173 + ], + "type": "text", + "content": "[21] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. NeurIPS, 35:26565-26577, 2022. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 175, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 175, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 317, + 175, + 553, + 207 + ], + "type": "text", + "content": "[22] Yann LeCun, Sumit Chopra, Raia Hadsell, Marc'Aurelio Ranzato, and Fu Jie Huang. A tutorial on energy-based learning. Predicting structured data, 1(2):1-31, 2006. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 209, + 553, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 209, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 317, + 209, + 553, + 230 + ], + "type": "text", + "content": "[23] Yaron Lipman, Justin Solomon, et al. Flow matching for generative modeling. *ICLR*, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 232, + 555, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 232, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 317, + 232, + 555, + 285 + ], + "type": "text", + "content": "[24] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9298–9309, 2023. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 288, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 288, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 288, + 553, + 319 + ], + "type": "text", + "content": "[25] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In ICCV, 2015. 5, 11" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 322, + 553, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 322, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 316, + 322, + 553, + 364 + ], + "type": "text", + "content": "[26] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. In Semin al graphics: pioneering efforts that shaped the field, pages 347-353, 1998. 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 366, + 553, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 366, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 366, + 553, + 409 + ], + "type": "text", + "content": "[27] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In CVPR, pages 4460-4470, 2019. 2, 7, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 411, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 443 + ], + "type": "text", + "content": "[28] Jonathan Oswald, Christian Henning, Tomas Pfister, and Jürgen Schmidhuber. Continual learning with hypernetworks. In ICLR, 2019. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 445, + 553, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 553, + 488 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 553, + 488 + ], + "type": "text", + "content": "[29] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In CVPR, pages 165-174, 2019. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 491, + 553, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 491, + 553, + 511 + ], + "spans": [ + { + "bbox": [ + 317, + 491, + 553, + 511 + ], + "type": "text", + "content": "[30] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, pages 4195-4205, 2023. 5" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 514, + 553, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 514, + 553, + 556 + ], + "spans": [ + { + "bbox": [ + 317, + 514, + 553, + 556 + ], + "type": "text", + "content": "[31] Marcin Przewiezlikowski, Przemyslaw Przybysz, Jacek Tabor, Maciej Zikeba, and Przemyslaw Spurek. Hypermaml: Few-shot adaptation of deep models with hypernetworks. Neurocomputing, 2024. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 317, + 558, + 553, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 558, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 317, + 558, + 553, + 590 + ], + "type": "text", + "content": "[32] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, 2017. 7" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 317, + 592, + 553, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 592, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 317, + 592, + 553, + 646 + ], + "type": "text", + "content": "[33] Lingteng Qiu, Guanying Chen, Xiaodong Gu, Qi Zuo, Mutian Xu, Yushuang Wu, Weihao Yuan, Zilong Dong, Liefeng Bo, and Xiaoguang Han. Richdreamer: A generalizable normal-depth diffusion model for detail richness in text-to-3d. In CVPR, pages 9914–9925, 2024. 7" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 317, + 647, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 553, + 712 + ], + "type": "text", + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 6" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22137" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "text", + "content": "[35] Maziar Raissi, Paris Perdikaris, and George Em Karniadakis. Physics informed deep learning (part i): Data-driven solutions of nonlinear partial differential equations. arxiv, 2017. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "type": "text", + "content": "[36] Maziar Raissi, Paris Perdikaris, and George Em Karniadakis. Physics informed deep learning (part ii): Data-driven discovery of nonlinear partial differential equations. arxiv, 2017." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 206 + ], + "type": "text", + "content": "[37] Maziar Raissi, Paris Perdikaris, and George Em Karniadakis. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational Physics, 378:686-707, 2019. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 239 + ], + "type": "text", + "content": "[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 241, + 294, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 294, + 296 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 294, + 296 + ], + "type": "text", + "content": "[39] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, pages 22500-22510, 2023. 1, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 297, + 294, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 294, + 351 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 294, + 351 + ], + "type": "text", + "content": "[40] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdream booth: Hypernetworks for fast personalization of text-to-image models. In CVPR, pages 6527-6536, 2024. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 353, + 294, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 353, + 294, + 385 + ], + "spans": [ + { + "bbox": [ + 56, + 353, + 294, + 385 + ], + "type": "text", + "content": "[41] David E Rumelhart, Geoffrey E Hinton, and Ronald J Williams. Learning representations by back-propagating errors. NATURE, 323(6088):533-536, 1986. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 387, + 294, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 387, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 56, + 387, + 294, + 441 + ], + "type": "text", + "content": "[42] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 35:36479-36494, 2022. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 443, + 294, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 443, + 294, + 485 + ], + "spans": [ + { + "bbox": [ + 56, + 443, + 294, + 485 + ], + "type": "text", + "content": "[43] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, pages 2256-2265, 2015. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 487, + 294, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 487, + 294, + 518 + ], + "spans": [ + { + "bbox": [ + 56, + 487, + 294, + 518 + ], + "type": "text", + "content": "[44] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. NeurIPS, 32, 2019. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 521, + 294, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 521, + 294, + 554 + ], + "spans": [ + { + "bbox": [ + 56, + 521, + 294, + 554 + ], + "type": "text", + "content": "[45] Yang Song, Conor Durkan, Iain Murray, and Stefano Ermon. Maximum likelihood training of score-based diffusion models. NeurIPS, pages 1415-1428, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 555, + 294, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 555, + 294, + 577 + ], + "spans": [ + { + "bbox": [ + 56, + 555, + 294, + 577 + ], + "type": "text", + "content": "[46] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. arxiv, 2023. 3, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 578, + 294, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 578, + 294, + 610 + ], + "spans": [ + { + "bbox": [ + 56, + 578, + 294, + 610 + ], + "type": "text", + "content": "[47] Pascal Vincent. A connection between score matching and denoising autoencoders. Neural Computation, 23(7):1661-1674, 2011. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 613, + 294, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 294, + 655 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 294, + 655 + ], + "type": "text", + "content": "[48] Guandao Yang, Xun Huang, Zekun Hao, Ming-Yu Liu, Serge Belongie, and Bharath Hariharan. Pointflow: 3d point cloud generation with continuous normalizing flows. In ICCV, pages 4541-4550, 2019. 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 657, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 657, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 657, + 294, + 689 + ], + "type": "text", + "content": "[49] Linnan Zhang, Zhiding Ren, Seunghyun Ahn, Hakan Bilen, and Kwang In Lee. Graph hypernetworks for neural architecture search. *ICLR*, 2019. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 690, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 690, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 690, + 294, + 712 + ], + "type": "text", + "content": "[50] Jun-Yan Zhu et al. Shap-e: Generating conditional 3d implicit functions. arxiv, 2023. 7" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22138" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/HyperPose_ Hypernetwork-Infused Camera Pose Localization and an Extended Cambridge Landmarks Dataset/dd4d95b8-93a1-4dc9-acfe-357b62a866ca_content_list.json b/2025/HyperPose_ Hypernetwork-Infused Camera Pose Localization and an Extended Cambridge Landmarks Dataset/dd4d95b8-93a1-4dc9-acfe-357b62a866ca_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..21cb2dda64dfd095057accd6ff52ba69a223849f --- /dev/null +++ b/2025/HyperPose_ Hypernetwork-Infused Camera Pose Localization and an Extended Cambridge Landmarks Dataset/dd4d95b8-93a1-4dc9-acfe-357b62a866ca_content_list.json @@ -0,0 +1,1304 @@ +[ + { + "type": "text", + "text": "HyperPose: Hypernetwork-Infused Camera Pose Localization and an Extended Cambridge Landmarks Dataset", + "text_level": 1, + "bbox": [ + 98, + 130, + 898, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ron Ferens Yosi Keller", + "bbox": [ + 385, + 203, + 611, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Faculty of Engineering, Bar Ilan University, Ramat-Gan, Israel", + "bbox": [ + 250, + 220, + 745, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{ronferens, yosi.keller}@gmail.com", + "bbox": [ + 346, + 241, + 645, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we propose HyperPose, which utilizes hypernetworks in absolute camera pose regressors. The inherent appearance variations in natural scenes, attributable to environmental conditions, perspective, and lighting, induce a significant domain disparity between the training and test datasets. This disparity degrades the precision of contemporary localization networks. To mitigate this, we advocate for incorporating hypernetworks into single-scene and multiscene camera pose regression models. During inference, the hypernetwork dynamically computes adaptive weights for the localization regression heads based on the particular input image, effectively narrowing the domain gap. Using indoor and outdoor datasets, we evaluate the HyperPose methodology across multiple established absolute pose regression architectures. We also introduce and share the Extended Cambridge Landmarks (ECL), a novel localization dataset, based on the Cambridge Landmarks dataset, showing it in multiple seasons with significantly varying appearance conditions. Our empirical experiments demonstrate that HyperPose yields notable performance enhancements for single- and multi-scene architectures. We have made our source code, pre-trained models1, and the ECL dataset openly available2.", + "bbox": [ + 86, + 323, + 485, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 686, + 220, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In many computer vision applications, such as indoor navigation, augmented reality, and autonomous driving, it is essential to determine the position and orientation of a camera using an image taken by that camera. Current camera localization techniques show various trade-offs between accuracy, runtime, and memory usage. Feature Matching (FM) and Scene Coordinate Regression (SCR) are considered the most accurate localization methods. FM-based methods [15, 27, 37-41, 51, 54] involve hierarchical localization pipelines, which commonly use a coarse-to-fine pro", + "bbox": [ + 89, + 713, + 483, + 864 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0849acdbaaba02d7326f257f5c5c056f7ffb98c9286d996bea8f57dd6249927b.jpg", + "image_caption": [ + "(a) Baseline architecture." + ], + "image_footnote": [], + "bbox": [ + 555, + 294, + 875, + 364 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/104ac34d2fca2452683ed8f52ece7a2f507a453796ac774c7472d2a9569eafae.jpg", + "image_caption": [ + "(b) Baseline architecture with a hypernetwork.", + "Figure 1. Figure 1a illustrates a baseline APR architecture. It consists of a CNN backbone that generates a feature vector encoding the input query image, followed by regression heads that estimate the translation and orientation. Figure 1b shows the extension of this baseline architecture with a hypernetwork that outputs the weights for the regression heads (translation and orientation) during inference." + ], + "image_footnote": [], + "bbox": [ + 555, + 388, + 870, + 511 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cessing pipeline that begins with image retrieval to perform an initial localization based on images similar to the query image. The initial estimate is refined by local feature extraction and matching. To estimate the camera's 6-Degrees-Of-Freedom (6DoF) accurately, the resulting 2D-2D matches are mapped to 3D correspondences in the scene's point cloud and then used to determine the camera pose through Perspective-n-Point (PnP) and RANSAC. SCR-based methods [4-8, 47] directly regress correspondences between 2D pixels and 3D coordinates in the scene. Relative pose regression (RPR) schemes [2, 14, 17, 20, 26, 57, 62] directly estimate the relative pose between a query image and a set of reference images with well-defined ground truth poses, eliminating the need for conventional pipelines. However, an image retrieval scheme is essential to retrieve a set of nearest-neighbor images. Another approach, absolute pose", + "bbox": [ + 511, + 659, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1https://ronferens.github.io/hyperpose/", + "bbox": [ + 107, + 875, + 379, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "2https://ronferens.github.io/extcambridgelandmarks/", + "bbox": [ + 109, + 887, + 460, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "11547", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "regressors (APRs), can estimate the camera pose with a single forward pass using only the query image, resulting in significantly lower latency, although notably less accurate. Additionally, due to their low memory footprint, APRs can be deployed as a standalone application on edge devices with limited computational resources. A survey of visual camera pose localization is given by [43, 65].", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "PoseNet by Kendall et al. [25] was the first APR approach, which leveraged a convolutional backbone and a multilayer perceptron (MLP) head to regress the camera's position and orientation. This simple and computationally efficient approach, enabling execution at 5ms, led to numerous APR methods that improved accuracy by modifying the backbone and MLP architectures [11, 32, 33, 44, 49, 60, 61, 63]. In addition, alternative loss functions and optimization strategies were explored [23, 24, 43]. Indoor and outdoor environments exhibit dynamic variability in illumination, perspective, and object motion. APR models, despite their computational efficiency, often struggle to maintain accuracy under these diverse conditions and show limited generalization to novel scenes, necessitating scene-specific training. Consequently, while architecturally simple, a unified network might yield suboptimal performance when adapting to heterogeneous real-world visual inputs. Recent studies have proposed extending APR from single-scene to multiscene formulations. Blanton et al. [3] implemented a convolutional neural network (CNN) backbone to compute a global descriptor for the input image. Rather than employing a scene-specific multilayer perceptron (MLP), they optimized a set of fully connected (FC) layers, allocating one layer per scene based on a predicted scene identifier. Although introducing a framework to optimize a single model across multiple scenes, this methodology did not achieve state-of-the-art (SOTA) APR accuracy. An alternative approach was presented by [46] using dual-branch transformers [58]. Encoders focused on extracting pose-related features, while decoders transformed encoded scene identifiers into latent pose representations. The decoder outputs are leveraged for scene classification and selection of corresponding position and orientation embeddings, which are used to regress position and orientation vectors.", + "bbox": [ + 91, + 199, + 483, + 713 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Hypernetworks [22] are a deep learning architecture in which an auxiliary network, the hypernetwork, generates the weights for the primary network (main network) based on the current input and downstream task. The main network and hypernetwork are trained end-to-end using backpropagation. The hypernetwork allows for dynamic and adaptive weight modulation of the main network, leading to improved flexibility in response to changing conditions [16, 18, 34, 55].", + "bbox": [ + 89, + 715, + 482, + 852 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose HyperPose, a hypernetwork-based approach for absolute camera pose regression. Extending common APR methods [11, 23, 25, 44, 45, 60],", + "bbox": [ + 89, + 854, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the hypernetwork computes weights for the regression head. This enables adaptive analysis of image content, allowing the regression weights to focus on informative features, rather than relying on a single constant pretrained set of weights. We validated the proposed method by integrating hypernetworks into single- and multiscene APR architectures. For single-scene networks, we present HyperPose with an existing state-of-the-art APR model [61] as well as a simplistic baseline APR using an off-the-shelf backbone [53] and shallow regression heads, inspired by the principles outlined in [25]. For multiscene, we extend a leading multiscene APR model [46] by integrating a hypernetwork architecture. In both instances, hypernetwork-augmented models demonstrated superior performance compared to their original counterparts when evaluated on the benchmark datasets in this study. To evaluate localization robustness, we introduce an Extended Cambridge Landmarks (ECL) dataset - an augmentation of the original benchmark [25] with algorithmically generated environmental shifts. By combining evening, winter, and summer scenes, the ECL incorporates dynamics such as lighting shifts and seasonal transitions.", + "bbox": [ + 511, + 90, + 903, + 421 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Thus, we propose the following contributions:", + "bbox": [ + 535, + 422, + 839, + 436 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose an approach to enhance any existing absolute camera pose regression models by integrating hypernetwork architectures.", + "- We comprehensively evaluated the proposed HyperPose on both single- and multiscene pose regressors, demonstrating improved accuracy achieved by the hypernetwork-enhanced models over the original versions.", + "- The proposed scheme achieves new SOTA accuracy on multiscene APR benchmarks spanning diverse outdoor and indoor localization tasks.", + "- We introduce the Extended Cambridge Landmarks (ECL) dataset, which builds upon the foundation of the original Cambridge Landmarks dataset [25]. Our ECL data set extends existing test scenes by incorporating various appearance conditions." + ], + "bbox": [ + 513, + 439, + 903, + 679 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 693, + 653, + 709 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Camera pose estimation techniques differ by their input during inference and algorithmic attributes.", + "bbox": [ + 511, + 719, + 903, + 748 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D-based or structure-based localization methods are considered the most accurate for camera pose estimation and have demonstrated state-of-the-art performance in leading benchmarks. These methods utilize correspondences between 2D features in an image and 3D world coordinates. [15, 27, 35, 37-41, 51, 54] introduced a two-phase approach for hierarchical pose estimation pipelines. Thus, each query image is first encoded using a CNN trained for image retrieval (IR), based on a pre-mapped image dataset. Tentative correspondences are then estimated by matching local im", + "bbox": [ + 511, + 750, + 906, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "11548", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "age features to 3D matches. The resulting matches are used by PnP-RANSAC to estimate the camera pose. DSAC [8] and its succeeding scheme, DSAC++ [4], utilize a CNN architecture to directly estimate 3D coordinates from 2D positions in an image and only require the query image during inference. The same as in Absolute Pose Regression (APR) methods, separate models must be trained for each scene, achieving state-of-the-art accuracy. Lately, [15] proposes a solution to mitigate this gap by introducing a scene-agnostic architecture.", + "bbox": [ + 89, + 90, + 480, + 241 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Relative Pose Regression (RPR) algorithms estimate the relative pose between an input query image and a set of reference images based on their known location. Calculating the relative pose involves estimating the relative position and orientation between the query image and the reference images [1, 6, 68]. An Image Retrieval (IR) approach is utilized to determine the closest set of neighbor images, and the relative motion is then estimated between the query image and each of the retrieved images. The estimated relative poses are used, along with the known position of the reference images, to estimate the camera's absolute pose. Similar to 3D and structured-based localization methods, RPRs involve a multistep process and require access to a database of images with labeled poses during the inference phase.", + "bbox": [ + 89, + 247, + 482, + 459 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Absolute Pose Regression (APR) directly estimates the camera pose given the input query image. APRs were first introduced by Kendall [25], using a modified version of the truncated GoogLeNet architecture, where the Softmax classification layer was replaced by a series of fully connected layers that output the pose estimate. Although less accurate than classical structure-based methods, APRs offer several advantages, such as faster inference times (milliseconds), reduced memory footprint (megabytes), and do not require feature engineering. Variations of the encoder and MLP architectures have been proposed to improve APR accuracy [11, 23, 33, 33, 44, 49]. Furthermore, modifications to the loss function have been suggested to improve the balance between orientation and position objectives [24] or to incorporate other types of information [9]. For an in-depth examination of the various architectural designs and loss functions utilized in camera pose regression, please refer to [43]. Although demonstrating faster inference times and low memory requirements, absolute pose regression (APR) models typically require dedicated perscene training to achieve competitive accuracy compared to other methods. This limitation has raised concerns regarding the value of APRs versus significantly more accurate techniques such as [4, 6, 8, 36]. To address this issue, recent approaches suggested extending APRs to a multiscene paradigm. [3] introduced Multi-Scene PoseNet (MSPN), which utilizes a CNN to regress an image feature vector. The two branches then process this feature vector. The first branch classifies the related scene using a fully connected", + "bbox": [ + 89, + 463, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "layer with a softmax operator. The estimated scene index selects scene-specific weights from a database. The second branch combines these weights with the feature vector to regress the pose. The model is trained end-to-end by applying both binary cross-entropy (for classification) and camera pose losses for each scene. Inspired by recent successful applications of Transformers [12, 19, 29] to computer vision tasks, [46] suggested a shared convolutional network backbone followed by two network branches consisting of a Transformer and a regression head. The first regresses the input query image's pose, and the other its orientation. The main incentive for utilizing Transformers encoders is to focus on pose-informative features and decoders to transform encoded scene identifiers into latent pose representations.", + "bbox": [ + 511, + 90, + 903, + 301 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Hypernetworks were first presented by Ha et al. [22]. These are neural networks that have been designed to predict the weights of a primary network. The primary network's weights can be adjusted based on specific inputs, resulting in a more expressive and adaptive model. Hypernetworks have been used in various applications such as learning implicit neural representations [16], semantic segmentation [34], 3D scene representation and modeling [28, 48, 50] and continuous learning [59], to name a few.", + "bbox": [ + 511, + 301, + 903, + 439 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. APRs using Hypernetworks", + "text_level": 1, + "bbox": [ + 511, + 450, + 769, + 467 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "HyperPose is a general approach, applicable to both single- and multi-scene scenarios, for enhancing APR schemes. It comprises two key components: the main network (APR) designed to estimate the camera's absolute position and the hypernetwork that computes the regression weights for the main network conditioned on the input query image. The main network performs a forward pass on the input image to localize the capture camera. The camera pose is represented as the tuple $< \\mathbf{x},\\mathbf{q}>$ , where $\\mathbf{x}\\in \\mathbb{R}^3$ is the camera's position in world coordinates and $\\mathbf{q}\\in \\mathbb{R}^4$ is the quaternion encoding of its $3D$ orientation. Following [13], we formulate the hypernetwork $H$ as a parametric function $\\theta_h = H(\\theta ,\\mathbf{x})$ , where $\\theta$ represents the parameters of the hypernetwork, $\\mathbf{x}$ the network's input, and $\\theta_h$ the parameters of the regression heads layer for the input query image $I_q$ .", + "bbox": [ + 511, + 474, + 903, + 702 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Network Architecture", + "text_level": 1, + "bbox": [ + 511, + 709, + 718, + 723 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.1. HyperPose for Single-Scene APRs", + "text_level": 1, + "bbox": [ + 511, + 731, + 795, + 744 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We apply the proposed HyperPose-based approach to single-scene APR using a baseline APR following [25], and the state-of-the-art AtLoc APR [61].", + "bbox": [ + 511, + 750, + 903, + 794 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Baseline APR with Hypernetwork. As depicted in Fig. 1a, the baseline model's main network consists of a CNN backbone that generates a feature vector encoding the input query image. This vector is then processed by two regression heads. The first estimates the camera position, and the other estimates the orientation. Figure 1b illustrates the integration of a hypernetwork into this architecture. The input", + "bbox": [ + 511, + 795, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "11549", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4bc874b428f303788cf723c4f5595bc3997d4dff65879b6e17c5e680ebdf7ba8.jpg", + "image_caption": [ + "Figure 2. MS-HyperPose - The proposed multi-scene absolute pose regression architecture using a hypernetwork. The primary network employs position and orientation Transformers, in a dual-branch architecture, to extract activation maps from the underlying convolutional backbone. The hypernetwork generates the weights for the regression head in the primary network, using the input query image. These adaptive weights, combined with the latent vectors generated by the Transformers estimate the camera pose, consisting of the spatial position (x) and angular orientation (q)." + ], + "image_footnote": [], + "bbox": [ + 114, + 97, + 897, + 406 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "to the hypernetwork is the feature vector generated by the CNN backbone of the main network. This vector is passed through a single fully-connected (FC) layer followed by a Swish activation [36] function. The processed vector serves as an input to a multilayer perceptron (MLP) that outputs the weights for a single layer in the regression heads of the main network. The camera pose is then estimated by the position and orientation regression heads of the main network.", + "bbox": [ + 88, + 517, + 480, + 637 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "AtLoc with Hypernetwork. [61] presents Attention-Guided Camera Localization (AtLoc), an absolute pose regression model based on self-attention. In this architecture, a visual encoder encodes an image into a latent vector. Conditioned on these extracted features, an attention module computes self-attention maps to reweight the representation into a new feature space. A pose regressor then maps the transformed features after the attention operators to estimate the camera pose. We incorporate a hypernetwork into the AtLoc, the same as for the baseline APR model. In this case, the input to the hypernetwork is the reweighted representation feature vector computed by the visual encoder.", + "bbox": [ + 88, + 640, + 482, + 821 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1.2. HyperPose for Multi-Scene APRs", + "text_level": 1, + "bbox": [ + 89, + 834, + 369, + 849 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The suggested multi-scene model, named MS-HyperPose is an extension of the architecture introduced in [45]. As depicted in Fig. 2, the input query image $I_{q}$ is processed by a", + "bbox": [ + 89, + 854, + 482, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "convolutional backbone that computes intermediate activation maps from two different layers. These activation maps are then utilized as inputs to the main network and the hypernetwork. Within the main network, the positional and orientational Transformers process the activation maps and integrate sequential representations into single latent vectors. After converting the activation maps into representation vectors, these vectors, along with the latent vectors are passed through multilayer perceptron (MLP) layers in the hypernetwork to generate the weights $\\theta_h$ for the translation and orientation regression heads. Using the weights $\\theta_h$ produced by the hypernetwork $H$ and the regression heads in the main network, the position, and orientation of the input query image are estimated.", + "bbox": [ + 511, + 517, + 906, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Main Network consists of a convolutional backbone and two separate branches to regress the camera's position and orientation. Each branch includes its particular Transformer and a multilayer perceptron (MLP) head. Studies have demonstrated that substituting the patchify stem, utilized in the early visual processing of a Vision Transformer (ViT), with a convolutional stem enhances optimization stability and elevates performance to a higher level [64]. Therefore, given an input query image $I_{q} \\in \\mathbb{R}^{H \\times W \\times C}$ , a shared convolutional backbone generates activation maps. Two different activation maps are selected", + "bbox": [ + 511, + 734, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "11550", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "from the CNN based on the particular learned task (Fig. 2a). The activation maps, represented as $M \\in \\mathbb{R}^{h \\times w \\times C^{Fn}}$ , are processed and transformed into $\\tilde{M} \\in \\mathbb{R}^{h \\times w \\times C^{Out}}$ using a $1 \\times 1$ convolutional layer followed by flattening (Fig. 2b).", + "bbox": [ + 89, + 90, + 480, + 154 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following [45], we use the Transformer architecture from [12], modifying the standard Encoder and Decoder to add positional encoding at each attention layer. Each Transformer has $L$ identical blocks, each with a multi-head attention (MHA) mechanism and a two-layer MLP with Swish activation [36]. Layer Normalization (LN) is applied before MHA and MLP processing, as in [58], with outputs merged via residual connections and dropout. To preserve spatial information, each activation map position receives a learned encoding. The processed sequence, serving as input to the Transformer is given by (Fig. 2c):", + "bbox": [ + 88, + 155, + 482, + 320 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nE ^ {I n} = \\left[ T, \\tilde {M} \\right] \\in \\mathbb {R} ^ {(h w) \\times C ^ {O u t}}, T \\in \\mathbb {R} ^ {C _ {M} ^ {O u t}}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 343, + 482, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $T$ is the positional encoding of the activation map.", + "bbox": [ + 109, + 375, + 480, + 390 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given a dataset with $N$ scenes, the Transformer Decoders generate embeddings $E^{Out} = \\in \\mathbb{R}^{N\\times C^{Out}}$ (Fig. 2d). For a query image, only one position matches its scene. The combined Transformer outputs (Fig. 2e) pass through a fully connected layer and Log SoftMax, selecting the highest-probability vectors (Fig. 2f).", + "bbox": [ + 89, + 390, + 480, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The selected vectors $E_{x_i}^{Out}$ and $E_{q_i}^{Out}$ are processed through an MLP. Finally, the estimated camera position and orientation are obtained by summing the outputs of the main network and hypernetwork, using weights predicted by the hypernetwork (Fig. 2h and 2i).", + "bbox": [ + 89, + 479, + 480, + 556 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Hypernetworks consist of MLPs, each regressing the weights for a corresponding regression head (position or orientation) in the main network. Their input is the sum of task-specific encoded activation maps $\\tilde{M}_x$ and $\\tilde{M}_q$ , along with global latent vectors $E_{x_i}^{Out}$ and $E_{q_i}^{Out}$ (Fig. 2g). To ensure size compatibility, the encoded activation map passes through a fully connected layer before summation. This vector is then processed by MLPs with Swish activation, generating inputs for another MLP set, each estimating weights for a specific layer in the main network's regression heads:", + "bbox": [ + 89, + 556, + 482, + 720 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {h} ^ {O u t _ {j}} = H \\left(\\theta_ {h} ^ {O u t _ {j}}, E _ {i} ^ {O u t}\\right) \\in \\mathbb {R} ^ {C _ {E _ {i}} ^ {O u t} \\times C _ {H} ^ {O u t _ {j}}}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 744, + 482, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $j$ is the index of the regression head layer.", + "bbox": [ + 109, + 777, + 431, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Optimization Criteria", + "text_level": 1, + "bbox": [ + 89, + 801, + 295, + 819 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The localization error is calculated by comparing the translation position and orientation of the ground truth pose $\\mathbf{p}_{gt} = < \\mathbf{x}_{gt},\\mathbf{q}_{gt}>$ , where $\\mathbf{x}\\in \\mathbb{R}^3$ represents the position of the camera in the world and $\\mathbf{q}\\in \\mathbb{R}^4$ denotes its orientation encoded as a quaternion, with the estimated pose", + "bbox": [ + 89, + 825, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\mathbf{p}_{est} = < \\mathbf{x}_{est},\\mathbf{q}_{est} >$ . The translation error, represented by $L_{x}$ , is calculated by the Euclidean distance between the ground truth and the estimated camera positions. The orientation error, $L_{q}$ , is defined as the Euclidean distance between the ground truth and the estimated unit quaternions [43]:", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {x} = \\left\\| \\mathbf {x} _ {0} - \\mathbf {x} \\right\\| _ {2}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 647, + 196, + 903, + 212 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {q} = \\left\\| \\mathbf {q} _ {0} - \\frac {\\mathbf {q}}{| | \\mathbf {q} | |} \\right\\| _ {2} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 638, + 218, + 903, + 252 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{q}$ is a normalized norm quaternion to ensure it is a valid orientation encoding.", + "bbox": [ + 511, + 256, + 903, + 285 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As a multi-regression optimization problem, we adopt the method outlined in [24] by using different weighting schemes, based on the uncertainty of each task, to balance the individual loss functions defined for the separate objectives. The model optimizes the aggregated loss function, a combination of the weighted individual losses.", + "bbox": [ + 511, + 287, + 905, + 377 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {p} = L _ {x} \\exp (- s _ {x}) + s _ {x} + L _ {q} \\exp (- s _ {q}) + s _ {q}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 404, + 903, + 421 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $S_{x}$ and $S_{q}$ are the learned parameters. As a further optimization, we adopt the approach detailed in [44] and follow a three-step training process. Firstly, the entire network is trained using the aggregated loss in Eq. 5. In the next two stages, each MLP head undergoes separate fin-tuning with its corresponding loss function.", + "bbox": [ + 511, + 431, + 905, + 521 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Implementation Details", + "text_level": 1, + "bbox": [ + 513, + 531, + 730, + 546 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The shared backbone uses EfficientNet-B0 [31, 52]. We used the activation maps of two endpoints (reduction levels) as input for our position and orientation branches: $m_{rdct4} \\in \\mathbb{R}^{14 \\times 14 \\times 112}$ and $m_{rdct3} \\in \\mathbb{R}^{28 \\times 28 \\times 40}$ , respectively. Following [45], we applied linear projections on each activation map to a common depth dimension of $C_M^{Out} = 256$ and learned the positional encoding of the same depth. The regression heads in the main network consist of two-layer MLPs with input and output layers. The input layer dimension is 1024 for both the position and orientation, while the output dimensions are 3 and 4, respectively.", + "bbox": [ + 511, + 551, + 905, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Single-Scene APR. For both the baseline APR and the HyperPose extension of the AtLoc [61] model, the hypernetwork for the position branch generates weights for an MLP with a dimension of 256, while the orientation hypernetwork produces weights of 512 dimensions.", + "bbox": [ + 511, + 719, + 903, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multi-Scene APR. In contrast to [45], we incorporated four-layer (instead of six) Transformers in each branch (position and orientation) within the main network architecture. Each block consists of an MHA layer with four heads, followed by an MLP that preserves the input dimension. A dropout rate of $p = 0.1$ was applied to each MLP layer. The output dimension of each Transformer in the main network", + "bbox": [ + 511, + 795, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "11551", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/0429cb8d91235b68fcdb01e6d4d531c1d5ede97fed85c9d46d8c3f6535327046.jpg", + "table_caption": [ + "Table 1. Comparative analysis of APR architectures with and without the purported hypernetwork addition using the Cambridge Landmarks dataset. We report the median position/orientation error in meters/degrees. The best performance is marked in bold. (*) The authors of AtLoc did not report results on the Cambridge dataset, and we trained the AtLoc model on this dataset." + ], + "table_footnote": [], + "table_body": "
MethodK. CollegeOld HospitalShop FacadeSt. MaryAvg.
AtLoc* [61]1.53,3.212.17,3.990.93,4.612.14,5.761.69,4.39
AtLoc w/ hypernet0.96,3.432.04,3.610.88,4.221.72,5.441.40,4.18
Baseline APR0.89,2.291.49,3.320.74,4.791.40,4.951.13,3.84
Baseline APR w/ hypernet0.77,2.071.47,3.290.72,4.011.37,4.901.08,3.57
", + "bbox": [ + 189, + 141, + 803, + 217 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "was $C_M^{Out} = 256$ . The position and orientation branches of the hypernetwork generate the weights of three regression layers: input, hidden, and output. The number of weights in the position and orientation branches differ. The hypernetwork's position branch produces the weights for an MLP with dimensions 256, 128, and 3, while the orientation branch generates weights for an MLP with dimensions 512, 512, and 4. The weights generated by the hypernetwork are then put into the regression heads of the main network to estimate the position and orientation of the input query image. The input and hidden layers of the regression head utilize the Swish activation function.", + "bbox": [ + 89, + 229, + 483, + 411 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Extended Cambridge Landmarks Dataset", + "text_level": 1, + "bbox": [ + 89, + 426, + 465, + 445 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/dc3759e8b063cfa79df346d6a900aa23b12aa2b490f3d8304796f2c10cf416f9.jpg", + "image_caption": [ + "(a) Original" + ], + "image_footnote": [], + "bbox": [ + 93, + 465, + 282, + 549 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/323c1720ebdf2614c798d4faabf7f1b8418682de06d5453ea43ba4fc2cde6b98.jpg", + "image_caption": [ + "(b)Evening" + ], + "image_footnote": [], + "bbox": [ + 292, + 465, + 482, + 549 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/76acff78b28455974bf3c466a8d1bb63973e185e0580a784b91d7c118780f857.jpg", + "image_caption": [ + "(c) Summer" + ], + "image_footnote": [], + "bbox": [ + 93, + 564, + 282, + 647 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c7466b1bc9048e773f40d442bc0bb3668461139afb0373c26fc7d0ae149a9d07.jpg", + "image_caption": [ + "(d) Winter", + "Figure 3. A sample from the King's College Scene in the proposed Extended Cambridge Landmarks (ECL) Dataset" + ], + "image_footnote": [], + "bbox": [ + 292, + 564, + 482, + 646 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To further assess the limitations of the HyperPose approach, we constructed an augmented version of the Cambridge Landmarks [25] benchmark called the Extended Cambridge Landmarks (ECL) dataset. Leveraging the original scenes, the ECL dataset introduces new localization challenges due to variations in appearance by creating synthetic evening, winter, and summer variants for each location. These scene permutations simulate real-world environmental variations by introducing shifts in lighting, weather patterns, foliage density, and more. Analyzing performance across ECL variations and their original Cambridge counterparts allows us to evaluate the robustness of", + "bbox": [ + 89, + 719, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "models to real-world variations. These augmentations were conducted by utilizing the novel InstructPix2Pix work [10]. Visual samples of the ECL dataset can be found in the supplementary materials. Figure 3 shows sample from the King's College scene in the extended dataset. Additional samples from each scene in the ECL dataset can be found in the supplementary material. Quantitative pose regression analysis between these more challenging conditions and unmodified sequences helps scientifically assess model adaptability. We hypothesize that exposing limitations unseen in conventional test sets will motivate the development of more invariant solutions. As controllable expansions of established benchmarks, programmatically generated datasets can thus guide the development of robust models for real-world use.", + "bbox": [ + 511, + 229, + 906, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experimental Results", + "text_level": 1, + "bbox": [ + 511, + 470, + 718, + 487 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conducted an experimental evaluation using the 7 Scenes [21] and Cambridge Landmarks [25] datasets for evaluating absolute pose regression. The Cambridge Landmarks dataset represents an outdoor urban environment ranging in size from 900 to 5500 square meters. We present results for four of its six scenes. We excluded the remaining two scenes due to limited coverage in prior work. To compare robustness, we also report the results on the Extended Cambridge Landmarks dataset (ECL) (Section 1 in the supplementary materials). The 7 Scenes dataset consists of seven small-scale indoor scenes with spatial extents ranging from approximately 1 to 10 square meters. These datasets present various localization challenges, including differences in scale, indoor versus outdoor settings, repetitive elements, textureless features, significant viewpoint changes, and variations in trajectory between training and test sets. To facilitate a broader analysis of our proposed method, we conduct an additional evaluation on the Oxford RobotCar Dataset [30]. Such dynamic environmental factors assess HyperPose's ability to improve localization accuracy amid diverse, challenging visual contexts. Our experimental procedures adhere to the same training and evaluation protocols as in [9, 61]. Additional information concerning these sequences, and a comprehensive description of the training methodology and hyperparameter selection for the models, are provided in the supplementary material.", + "bbox": [ + 511, + 492, + 908, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate the impact of integrating hypernetworks in", + "bbox": [ + 532, + 885, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "11552", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a77f329e43f19698f0702f810ed8838afb7facfc19e3c84f4c300d38753d13a3.jpg", + "table_caption": [ + "Table 2. Comparison to single-scene state-of-the-art methods - Cambridge Landmarks: We report the median position/orientation error in meters/degrees. The most effective single and multi-scene APRs are marked in bold." + ], + "table_footnote": [], + "table_body": "
K. CollegeHospitalShop FacadeSt. MaryAvg.
Seq.DSAC* [6]0.18,0.3°0.21,0.4°0.05,0.3°0.15,0.5°0.15,0.4°
MapNet [9]1.08,1.9°1.94,3.9°1.49,4.2°2.00,4.5°1.63,3.6°
GL-Net [67]0.59,0.7°1.88,2.8°0.50,2.9°1.90,3.3°1.22,2.4°
IRVLAD [56]2.80,5.7°4.01,7.1°1.11,7.6°2.31,8.0°2.56,7.1°
VLAD+Inter [42]1.48,4.5°2.68,4.6°0.90,4.3°1.62,6.1°1.67,4.9°
RPREssNet [68]0.76,1.9°1.39,2.8°0.84,4.3°1.32,4.7°1.08,3.4°
NC-EssNet [68]0.61,1.6°0.95,2.7°0.7,3.4°1.12,3.6°0.85,2.8°
RelocGNN[21]0.48,1.0°1.14,2.5°0.48,2.5°1.52,3.2°0.91,2.3°
APRPoseNet [25]1.92,5.40°2.31,5.38°1.46,8.08°2.65,8.48°2.08,6.83°
BayesianPN [23]1.74,4.06°2.57,5.14°1.25,7.54°2.11,8.38°1.91,6.28°
LSTM-PN [60]0.99,3.65°1.51,4.29°1.18,7.44°1.52,6.68°1.30,5.57°
SVS-Pose [33]1.06,2.81°1.50,4.03°0.63,5.73°2.11,8.11°1.32,5.17°
GPoseNet [11]1.61,2.29°2.62,3.89°1.14,5.73°2.93,6.46°2.07,4.59°
PoseNetLearn [24]0.99,1.06°2.17,2.94°1.05,3.97°1.49,3.43°1.42,2.85°
GeoPoseNet [24]0.88,1.04°3.20,3.29°0.88,3.78°1.57,3.32°1.63,2.86°
IRPNet [44]1.18,2.19°1.87,3.38°0.72,3.47°1.87,4.94°1.41,3.50°
Baseline APR w/ hypernet0.61,1.84°1.44,3.03°0.70,3.62°1.37,4.85°1.03,2.67°
MS-APRMSPN [3]1.73,3.65°2.55,4.05°2.92,7.49°2.67,6.18°2.47,5.34°
MS-Trans[46]0.83,1.47°1.81,2.39°0.86,3.07°1.62,3.99°1.28,2.73°
MS-HyperPose0.78,1.18°1.84,2.11°0.83,3.13°1.61,3.22°1.27,2.41°
", + "bbox": [ + 171, + 127, + 823, + 468 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "single-scene APRs. To this end, we compare the accuracy of two model variations, with and without hypernetworks, and report median position and orientation errors. Table 1 presents results on the Cambridge dataset. The hypernetwork improves the accuracy of the proposed baseline APR and AtLoc [61] models. The baseline APR model size is $26.7\\mathrm{MB}$ with an inference time of $9.32\\mathrm{ms}$ , increasing to $40.9\\mathrm{MB}$ and $9.45\\mathrm{ms}$ with a hypernetwork. The original published AtLoc model is $97.9\\mathrm{MB}$ with an inference time of $5.14\\mathrm{ms}$ . The hypernetwork-enhanced AtLoc version has a model size of $116.8\\mathrm{MB}$ and an inference time of $5.56\\mathrm{ms}$ . Note that the reported errors refer to the models' performance after the first stage of model training, without applying the two additional optimization steps described in the training details (See supplementary materials). Furthermore, Tables 2 and 3 compare the median position error, and the orientation error, of the proposed baseline APR with hypernetwork architecture, with the results of recent SOTA localization algorithms, using the Cambridge Landmarks and 7Scenes datasets, respectively. For the indoor and outdoor datasets, in addition to the results of leading APRs, we report the median position and orientation errors of structure-based, sequence-based, IR, and RPR methods. Although DSAC* [6] achieves SOTA accuracy, compared to all other APR methods, HyperPose obtains the most accurate average of $1.03m$ , $2.67^{\\circ}$ on the Cambridge outdoor dataset, and $0.19m$ , $7.48^{\\circ}$ on the 7Scenes indoor dataset. These local-", + "bbox": [ + 91, + 481, + 480, + 888 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ization errors show competitive performance compared to leading RPR techniques, with the hypernet-enhanced baseline ranking third and second/fourth (7Scenes/Cambridge) in positional and rotational accuracy, respectively. Next, we compare the proposed MS-HyperPose architecture with other multiscene APR methods. Tables 2 and 3 present median position and orientation errors, respectively, on the Cambridge Landmarks and 7 Scenes datasets. On both indoor and outdoor benchmarks, our MS-HyperPose approach achieves the most accurate average of $1.27\\mathrm{m}$ , $2.41^{\\circ}$ on Cambridge and $0.18\\mathrm{m}$ , $7.12^{\\circ}$ on 7Scenes compared to other single- and multi-scene APRs. Additionally, MS-HyperPose is the most accurate for both positional and rotational errors in 6 out of 8 parameters on Cambridge and 6 out of 14 on 7 Scenes. The runtime of MS-HyperPose is $30.93\\mathrm{ms}$ , similar to the reference architecture of [46]. Its model size is 571MB, notably larger than the 71.1MB of the original. We discuss storage considerations in the Limitation and Future Work section of the supplementary materials.", + "bbox": [ + 511, + 481, + 903, + 781 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4 reports the mean and median position and orientation errors of AtLoc [61] compared to our hypernetwork-enhanced variant using the Oxford RobotCar Dataset [30]. We also compare the enhanced model with [9], the SOTA APR for this dataset using image sequences. Visualizations of the localization results can be found in the supplementary materials.", + "bbox": [ + 511, + 794, + 903, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "11553", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d82d080e52882a87b823d35f2dac963d51d7d16bb07acb42e874db3ff492a24a.jpg", + "table_caption": [ + "Table 3. Comparison to single-scene SOTA methods - 7Scenes: We report the median position/orientation error in meters/degrees. The most effective single and multi-scene APRs are distinguished in bold." + ], + "table_footnote": [], + "table_body": "
ChessFireHeadsOfficePumpkinKitchenStairsAvg.
SCRDSAC* [6]0.02,1.11°0.02,1.24°0.01,1.82°0.03,1.18°0.04,1.41°0.03,1.70°0.04,1.42°0.03,1.41°
Marepo[15]0.02,1.24°0.02,1.39°0.02,2.03°0.03,1.26°0.04,1.48°0.04,1.71°0.06,1.67°0.03,1.54°
Seq.LsG [66]0.09,3.3°0.26,10.9°0.17,12.7°0.18,5.5°0.20,3.7°0.23,4.9°0.23,11.3°0.19,7.74°
MapNet[9]0.08,3.3°0.27,11.7°0.18,13.3°0.17,5.2°0.22,4.0°0.23,4.9°0.30,12.1°0.21,7.78°
GL-Net[67]0.08,2.8°0.26,8.9°0.17,11.4°0.18,13.3°0.15,2.8°0.25,4.5°0.23,8.8°0.19,7.50°
IRVLAD [56]0.21,12.5°0.33,13.8°0.15,14.9°0.28,11.2°0.31,11.2°0.30,11.3°0.25,12.3°0.26,12.46°
VLAD+Inter[42]0.18,10.0°0.33,12.4°0.14,14.3°0.25,10.1°0.26,9.4°0.27,11.1°0.24,14.7°0.24,11.71°
RPRNN-Net [26]0.13,6.5°0.26,12.7°0.14,12.3°0.21,7.4°0.24,6.4°0.24,8.0°0.27,11.8°0.21,9.30°
RelocNet[2]0.12,4.1°0.26,10.4°0.14,10.5°0.18,5.3°0.26,4.2°0.23,5.1°0.28,7.5°0.21,6.73°
EssNet [68]0.13,5.1°0.27,10.1°0.15,9.9°0.21,6.9°0.22,6.1°0.23,6.9°0.32,11.2°0.22,8.03°
NC-EssNet [68]0.12,5.6°0.26,9.6°0.14,10.7°0.20,6.7°0.22,5.7°0.22,6.3°0.31,7.9°0.21,7.50°
RelocGNN[21]0.08,2.7°0.21,7.5°0.13,8.70°0.15,4.1°0.15,3.5°0.19,3.7°0.22,6.5°0.16,5.24°
CamNet [17]0.04,1.7°0.03,1.7°0.05,2.00°0.04,1.60°0.04,1.6°0.04,1.6°0.04,1.5°0.04,1.70°
APRPoseNet [25]0.32,8.12°0.47,14.4°0.29,12.0°0.48,7.68°0.47,8.42°0.59,8.64°0.47,13.8°0.44,10.4°
BayesianPN [23]0.37,7.24°0.43,13.7°0.31,12.0°0.48,8.04°0.61,7.08°0.58,7.54°0.48,13.1°0.47,9.81°
LSTM-PN [60]0.24,5.77°0.34,11.9°0.21,13.7°0.30,8.08°0.33,7.00°0.37,8.83°0.40,13.7°0.31,9.85°
GPoseNet [11]0.20,7.11°0.38,12.3°0.21,13.8°0.28,8.83°0.37,6.94°0.35,8.15°0.37,12.5°0.31,9.95°
PoseNetLearn[24]0.14,4.50°0.27,11.8°0.18,12.1°0.20,5.77°0.25,4.82°0.24,5.52°0.37,10.6°0.24,7.87°
GeoPoseNet[24]0.13,4.48°0.27,11.3°0.17,13.0°0.19,5.55°0.26,4.75°0.23,5.35°0.35,12.4°0.23,8.12°
IRPNet[44]0.13,5.64°0.25,9.67°0.15,13.1°0.24,6.33°0.22,5.78°0.30,7.29°0.34,11.6°0.23,8.49°
AtLoc[61]0.10,4.07°0.25,11.4°0.16,11.8°0.17,5.34°0.21,4.37°0.23,5.42°0.26,10.5°0.20,7.56°
TransBoNet[49]0.11,4.48°0.25,12.5°0.18,14.0°0.20,5.08°0.19,4.77°0.17,5.35°0.30,13.0°0.20,8.45°
Baseline APR w/ hypernet0.09,4.47°0.28,9.44°0.15,11.4°0.16,6.29°0.19,4.62°0.18,6.95°0.28,9.15°0.19,7.48°
MS-APRMSPN[3]0.09,4.76°0.29,10.5°0.16,13.1°0.16,6.80°0.19,5.50°0.21,6.61°0.31,11.6°0.20,7.56°
MS-Trans[46]0.11,4.66°0.24,9.60°0.14,12.2°0.17,5.66°0.18,4.44°0.17,5.94°0.26,8.45°0.18,7.28°
MS-HyperPose0.11,4.34°0.23,9.79°0.13,10.7°0.17,6.05°0.16,5.24°0.17,6.86°0.27,6.00°0.18,7.00°
", + "bbox": [ + 91, + 127, + 916, + 565 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/44d625949ce4e96d03f742c0faea2034a559be6382b7e6c08495f1d5f1f0fd5b.jpg", + "table_caption": [ + "Table 4. Comparative analysis of different APR architectures with and without the proposed hypernetwork to the Oxford RobotCar dataset: We report the mean and median position/orientation error in meters/degrees." + ], + "table_footnote": [], + "table_body": "
MapNet[9]AtLoc[61]AtLoc w/ hypernet
SequenceMeanMedianMeanMedianMeanMedian
LOOP18.76, 3.465.79, 1.548.61, 4.585.68, 2.238.42, 3.304.91, 1.33
LOOP29.84, 3.964.91, 1.678.61, 4.585.68, 2.238.10, 3.044.73, 1.27
FULL141.4, 12.517.94, 6.6829.6, 12.411.10, 5.2829.58, 6.777.86, 1.38
FULL259.3, 14.820.04, 6.3948.2, 11.112.2, 4.6345.96, 9.608.38, 1.45
", + "bbox": [ + 212, + 619, + 779, + 726 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also assess the performance of two configurations of the baseline APR model, with and without hypernetworks on the Extended Cambridge dataset (ECL). Across all scenes and variations, the hypernetwork-enhanced model outperforms the baseline, demonstrating better adaptation to the seasonal changes and diverse lighting conditions inherent in the scenes. The detailed analysis elucidating the robustness and resilience of HyperPose to scene changes is in the supplementary material.", + "bbox": [ + 89, + 739, + 485, + 876 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusions", + "text_level": 1, + "bbox": [ + 511, + 738, + 638, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we proposed a hypernetwork-based approach to improve APR models. Our experiments show that hypernetworks improve accuracy for single-scene APR models. We also present MS-HyperPose, a novel multiscene camera pose regression model enhanced with a hypernetwork, which achieves new SOTA accuracy in indoor and outdoor localization datasets. We introduce the Extended Cambridge Landmarks (ECL) dataset to facilitate research on robust localization methods.", + "bbox": [ + 511, + 763, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "11554", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Eduardo Arnold, Jamie Wynn, Sara Vicente, Guillermo Garcia-Hernando, Aron Monszpart, Victor Prisacariu, Daniyar Turmukhambetov, and Eric Brachmann. Map-free visual relocalization: Metric pose relative to a single image. In Proceedings of the European Conference on Computer Vision (ECCV), pages 690-708. Springer, 2022. 3", + "[2] Vassileios Balntas, Shuda Li, and Victor Prisacariu. Relocnet: Continuous metric learning relocalisation using neural nets. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 1, 8", + "[3] Hunter Blanton, Connor Greenwell, Scott Workman, and Nathan Jacobs. Extending absolute pose regression to multiple scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 38-39, 2020. 2, 3, 7, 8", + "[4] E. Brachmann and C. Rother. Learning less is more - 6D camera localization via 3D surface regression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4654-4662, 2018. 1, 3", + "[5] Eric Brachmann and Carsten Rother. Expert sample consensus applied to camera re-localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7525-7534, 2019.", + "[6] E. Brachmann and C. Rother. Visual Camera Re-Localization From RGB and RGB-D Images Using DSAC. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(09):5847-5865, 2022. 3, 7, 8", + "[7] Eric Brachmann, Frank Michel, Alexander Krull, Michael Ying Yang, Stefan Gumhold, and Carsten Rother. Uncertainty-driven 6D pose estimation of objects and scenes from a single rgb image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3364-3372, 2016.", + "[8] E. Brachmann, A. Krull, S. Nowozin, J. Shotton, F. Michel, S. Gumhold, and C. Rother. DSAC - differentiable RANSAC for camera localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2492-2500, Los Alamitos, CA, USA, 2017. 1, 3", + "[9] Samarth Brahmbhatt, Jinwei Gu, Kihwan Kim, James Hays, and Jan Kautz. Geometry-aware learning of maps for camera localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3, 6, 7, 8", + "[10] T. Brooks, A. Holynski, and A. A. Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18392-18402, Los Alamitos, CA, USA, 2023. IEEE Computer Society. 6", + "[11] Ming Cai, Chunhua Shen, and Ian Reid. A hybrid probabilistic model for camera relocalization. In Proceedings of the British Machine Vision Conference, page 238, 2018. 2, 3, 7, 8", + "[12] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Proceedings" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "of the European Conference on Computer Vision (ECCV), pages 213-229, Cham, 2020. 3, 5", + "[13] Yoav Chai, Raja Giryes, and Lior Wolf. Supervised and unsupervised learning of parameterized color enhancement. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 992-1000, 2020. 3", + "[14] Kefan Chen, Noah Snavely, and Ameesh Makadia. Wide-baseline relative camera pose estimation with directional learning. 2021 IEEE. In CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3257-3267, 2021. 1", + "[15] Shuai Chen, Tommaso Cavallari, Victor Adrian Prisacariu, and Eric Brachmann. Map-relative pose regression for visual re-localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20665-20674, 2024. 1, 2, 3, 8", + "[16] Yinbo Chen and Xiaolong Wang. Transformers as meta-learners for implicit neural representations. In Proceedings of the European Conference on Computer Vision (ECCV), pages 170–187. Springer, 2022. 2, 3", + "[17] Mingyu Ding, Zhe Wang, Jiankai Sun, Jianping Shi, and Ping Luo. CamNet: Coarse-to-fine retrieval for camera relocalization. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2019. 1, 8", + "[18] Siyan Dong, Qingnan Fan, He Wang, Ji Shi, Li Yi, Thomas Funkhouser, Baoquan Chen, and Leonidas J Guibas. Robust neural routing through space partitions for camera relocalization in dynamic indoor environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8544-8554, 2021. 2", + "[19] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In Proceedings of the International Conference on Learning Representations (ICLR), 2021. 3", + "[20] Sovann En, Alexis Lecheryv, and Frédéric Jurie. Rpnet: An end-to-end network for relative camera pose estimation. In Proceedings of the European Conference on Computer Vision (ECCV) Workshops, pages 0–0, 2018. 1", + "[21] B. Glocker, S. Izadi, J. Shotton, and A. Criminisi. Realtime rgb-d camera relocalization. In 2013 IEEE International Symposium on Mixed and Augmented Reality (ISMAR), pages 173-179, 2013. 6, 7, 8", + "[22] David Ha, Andrew Dai, and Quoc V Le. Hypernetworks. arXiv preprint arXiv:1609.09106, 2016. 2, 3", + "[23] Alex Kendall and Roberto Cipolla. Modelling uncertainty in deep learning for camera relocalization. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), pages 4762-4769, 2016. 2, 3, 7, 8", + "[24] A. Kendall and R. Cipolla. Geometric loss functions for camera pose regression with deep learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6555-6564, 2017. 2, 3, 5, 7, 8", + "[25] A. Kendall, M. Grimes, and R. Cipolla. Posenet: A convolutional network for real-time 6-dof camera relocalization. In" + ], + "bbox": [ + 516, + 93, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "11555", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 2938-2946, 2015. 2, 3, 6, 7, 8", + "[26] Z. Laskar, I. Melekhov, S. Kalia, and J. Kannala. Camera relocalization by computing pairwise relative poses using convolutional neural network. In Proceedings of the IEEE International Conference on Computer Vision Workshops (ICCVW), pages 920–929, 2017. 1, 8", + "[27] Xiaotian Li, Shuzhe Wang, Yi Zhao, Jakob Verbeek, and Juho Kannala. Hierarchical scene coordinate classification and regression for visual localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11983-11992, 2020. 1, 2", + "[28] Gidi Littwin and Lior Wolf. Deep meta functionals for shape representation. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 1824-1833, 2019. 3", + "[29] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 10012-10022, 2021. 3", + "[30] Will Maddern, Geoffrey Pascoe, Chris Linegar, and Paul Newman. 1 year, $1000\\mathrm{km}$ : The oxford robotcar dataset. The International Journal of Robotics Research, 36(1):3-15, 2017. 6, 7", + "[31] Luke Melaskyriazi. Efficientnet-pytorch, 2019. 5", + "[32] Iaroslav Melekhov, Juha Ylioinas, Juho Kannala, and Esa Rahtu. Image-based localization using hourglass networks. In Proceedings of the IEEE International Conference on Computer Vision Workshops (ICCVW), pages 870-877, 2017. 2", + "[33] Tayyab Naseer and W. Burgard. Deep regression for monocular camera-based 6-dof global localization in outdoor environments. 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1525-1530, 2017. 2, 3, 7", + "[34] Yuval Nirkin, Lior Wolf, and Tal Hassner. Hyperseg: Patchwise hypernetwork for real-time semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4061-4070, 2021. 2, 3", + "[35] N. Radwan, A. Valada, and W. Burgard. Vlocnet++: Deep multitask learning for semantic visual localization and odometry. IEEE Robotics and Automation Letters, 3(4): 4407-4414, 2018. 2", + "[36] Prajit Ramachandran, Barret Zoph, and Quoc V Le. Searching for activation functions. arXiv preprint arXiv:1710.05941, 2017. 3, 4, 5", + "[37] P. Sarlin, C. Cadena, R. Siegwart, and M. Dymczyk. From coarse to fine: Robust hierarchical localization at large scale. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12708-12717, 2019. 1, 2", + "[38] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4938-4947, 2020." + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Paul-Edouard Sarlin, Ajaykumar Unagar, Mans Larsson, Hugo Germain, Carl Toft, Viktor Larsson, Marc Pollefeys, Vincent Lepetit, Lars Hammarstrand, Fredrik Kahl, et al. Back to the feature: Learning robust camera localization from pixels to pose. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3247–3257, 2021.", + "[40] Torsten Sattler, Michal Havlena, Filip Radenovic, Konrad Schindler, and Marc Pollefeys. Hyperpoints and fine vocabularies for large-scale location recognition. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 2102-2110, 2015.", + "[41] T. Sattler, B. Leibe, and L. Kobbelt. Efficient effective prioritized matching for large-scale image-based localization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(9):1744-1756, 2017. 1, 2", + "[42] T. Sattler, Q. Zhou, M. Pollefeys, and L. Leal-Taixe. Understanding the limitations of cnn-based absolute camera pose regression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3297-3307, 2019. 7, 8", + "[43] Yoli Shavit and Ron Ferens. Introduction to camera pose estimation with deep learning. In arXiv preprint arXiv:1907.05272, 2019. 2, 3, 5", + "[44] Yoli Shavit and Ron Ferens. Do we really need scene-specific pose encoders? In Proceedings of the International Conference on Pattern Recognition (ICPR), pages 3186-3192. IEEE, 2021. 2, 3, 5, 7, 8", + "[45] Yoli Shavit, Ron Ferens, and Yoshi Keller. Learning multiscene absolute pose regression with transformers. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 2713-2722, 2021. 2, 4, 5", + "[46] Yoli Shavit, Ron Ferens, and Yoshi Keller. Learning multiscene absolute pose regression with transformers. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. 2, 3, 7, 8", + "[47] Jamie Shotton, Ben Glocker, Christopher Zach, Shahram Izadi, Antonio Criminisi, and Andrew Fitzgibbon. Scene coordinate regression forests for camera relocalization in rgb images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2013. 1", + "[48] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in Neural Information Processing Systems (NIPS), 33:7462-7473, 2020. 3", + "[49] Xiaogang Song, Hongjuan Li, Li Liang, Weiwei Shi, Guo Xie, Xiaofeng Lu, and Xinhong Hei. Transbonet: Learning camera localization with transformer bottleneck and attention. Pattern Recognition, 146:109975, 2024. 2, 3, 8", + "[50] Przemysław Spurek, Artur Kasymov, Marcin Mazur, Diana Janik, Slawomir K Tadeja, J Tabor, T Trzeciński, et al. Hyperpocket: Generative point cloud completion. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 6848-6853. IEEE, 2022. 3", + "[51] Hajime Taira, Masatoshi Okutomi, Torsten Sattler, Mircea Cimpoi, Marc Pollefeys, Josef Sivic, Tomas Pajdla, and Ak" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "11556", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ihiko Torii. Inloc: Indoor visual localization with dense matching and view synthesis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(4):1293-1307, 2021. 1, 2", + "[52] Mingxing Tan and Quoc Le. EfficientNet: Rethinking model scaling for convolutional neural networks. pages 6105-6114, Long Beach, California, USA, 2019. 5", + "[53] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In Proceedings of the International Conference on Machine Learning (ICML), pages 6105-6114. PMLR, 2019. 2", + "[54] Shitao Tang, Sicong Tang, Andrea Tagliasacchi, Ping Tan, and Yasutaka Furukawa. Neumap: Neural coordinate mapping by auto-transdecoder for camera localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 929-939, 2023. 1, 2", + "[55] Yi Tay, Zhe Zhao, Dara Bahri, Don Metzler, and Da-Cheng Juan. HyperGrid transformers: Towards a single model for multiple tasks. In Proceedings of the International Conference on Learning Representations (ICLR), 2021. 2", + "[56] Akihiko Torii, Relja Arandjelović, Josef Sivic, Masatoshi Okutomi, and Tomas Pajdla. 24/7 place recognition by view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1808-1817, 2015. 7, 8", + "[57] Mehmet Ozgur Turkoglu, Eric Brachmann, Konrad Schindler, Gabriel J Brostow, and Aron Monszpart. Visual camera re-localization using graph neural networks and relative pose supervision. In 2021 International Conference on 3D Vision (3DV), pages 145-155. IEEE, 2021. 1", + "[58] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems (NIPS), pages 5998-6008, 2017. 2, 5", + "[59] Johannes Von Oswald, Christian Henning, João Sacramento, and Benjamin F Grewe. Continual learning with hypernetworks. arXiv preprint arXiv:1906.00695, 2019. 3", + "[60] F. Walch, C. Hazirbas, L. Leal-Taixe, T. Sattler, S. Hilsenbeck, and D. Cremers. Image-based localization using lstms for structured feature correlation. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 627-637, 2017. 2, 7, 8", + "[61] Bing Wang, Changhao Chen, Chris Xiaoxuan Lu, Peijun Zhao, Niki Trigoni, and Andrew Markham. Atloc: Attention guided camera localization. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10393-10401, 2020. 2, 3, 4, 5, 6, 7, 8", + "[62] Dominik Winkelbauer, Maximilian Denninger, and Rudolph Triebel. Learning to localize in new environments from synthetic training data. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 5840-5846. IEEE, 2021. 1", + "[63] J. Wu, L. Ma, and X. Hu. Delving deeper into convolutional neural networks for camera relocalization. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), pages 5644-5651, 2017. 2" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[64] Tete Xiao, Mannat Singh, Eric Mintun, Trevor Darrell, Piotr Dollár, and Ross Girshick. Early convolutions help transformers see better. Advances in Neural Information Processing Systems (NIPS), 34:30392-30400, 2021. 4", + "[65] Meng Xu, Youchen Wang, Bin Xu, Jun Zhang, Jian Ren, Stefan Poslad, and Pengfei Xu. A critical analysis of image-based camera pose estimation techniques. arXiv preprint arXiv:2201.05816, 2022. 2", + "[66] Fei Xue, Xin Wang, Zike Yan, Qiuyuan Wang, Junqiu Wang, and Hongbin Zha. Local supports global: Deep camera relocalization with sequence enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 8", + "[67] Fei Xue, Xin Wu, Shaojun Cai, and Junqiu Wang. Learning multi-view camera relocalization with graph neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 7, 8", + "[68] Qunjie Zhou, Torsten Sattler, Marc Pollefeys, and Laura Leal-Taixe. To learn or not to learn: Visual localization from essential matrices. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), pages 3319-3326. IEEE, 2020. 3, 7, 8" + ], + "bbox": [ + 516, + 90, + 906, + 415 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11557", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/HyperPose_ Hypernetwork-Infused Camera Pose Localization and an Extended Cambridge Landmarks Dataset/dd4d95b8-93a1-4dc9-acfe-357b62a866ca_model.json b/2025/HyperPose_ Hypernetwork-Infused Camera Pose Localization and an Extended Cambridge Landmarks Dataset/dd4d95b8-93a1-4dc9-acfe-357b62a866ca_model.json new file mode 100644 index 0000000000000000000000000000000000000000..54cfa20c8b75f516260c4a271d9c384ee42b7dac --- /dev/null +++ b/2025/HyperPose_ Hypernetwork-Infused Camera Pose Localization and an Extended Cambridge Landmarks Dataset/dd4d95b8-93a1-4dc9-acfe-357b62a866ca_model.json @@ -0,0 +1,2147 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.099, + 0.131, + 0.9, + 0.177 + ], + "angle": 0, + "content": "HyperPose: Hypernetwork-Infused Camera Pose Localization and an Extended Cambridge Landmarks Dataset" + }, + { + "type": "text", + "bbox": [ + 0.386, + 0.204, + 0.612, + 0.221 + ], + "angle": 0, + "content": "Ron Ferens Yosi Keller" + }, + { + "type": "text", + "bbox": [ + 0.251, + 0.222, + 0.746, + 0.24 + ], + "angle": 0, + "content": "Faculty of Engineering, Bar Ilan University, Ramat-Gan, Israel" + }, + { + "type": "text", + "bbox": [ + 0.348, + 0.242, + 0.647, + 0.256 + ], + "angle": 0, + "content": "{ronferens, yosi.keller}@gmail.com" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.324, + 0.486, + 0.673 + ], + "angle": 0, + "content": "In this work, we propose HyperPose, which utilizes hypernetworks in absolute camera pose regressors. The inherent appearance variations in natural scenes, attributable to environmental conditions, perspective, and lighting, induce a significant domain disparity between the training and test datasets. This disparity degrades the precision of contemporary localization networks. To mitigate this, we advocate for incorporating hypernetworks into single-scene and multiscene camera pose regression models. During inference, the hypernetwork dynamically computes adaptive weights for the localization regression heads based on the particular input image, effectively narrowing the domain gap. Using indoor and outdoor datasets, we evaluate the HyperPose methodology across multiple established absolute pose regression architectures. We also introduce and share the Extended Cambridge Landmarks (ECL), a novel localization dataset, based on the Cambridge Landmarks dataset, showing it in multiple seasons with significantly varying appearance conditions. Our empirical experiments demonstrate that HyperPose yields notable performance enhancements for single- and multi-scene architectures. We have made our source code, pre-trained models1, and the ECL dataset openly available2." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.688, + 0.222, + 0.705 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.714, + 0.484, + 0.866 + ], + "angle": 0, + "content": "In many computer vision applications, such as indoor navigation, augmented reality, and autonomous driving, it is essential to determine the position and orientation of a camera using an image taken by that camera. Current camera localization techniques show various trade-offs between accuracy, runtime, and memory usage. Feature Matching (FM) and Scene Coordinate Regression (SCR) are considered the most accurate localization methods. FM-based methods [15, 27, 37-41, 51, 54] involve hierarchical localization pipelines, which commonly use a coarse-to-fine pro" + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.295, + 0.877, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.644, + 0.374, + 0.777, + 0.386 + ], + "angle": 0, + "content": "(a) Baseline architecture." + }, + { + "type": "image", + "bbox": [ + 0.556, + 0.389, + 0.872, + 0.512 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.587, + 0.515, + 0.833, + 0.528 + ], + "angle": 0, + "content": "(b) Baseline architecture with a hypernetwork." + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.54, + 0.908, + 0.637 + ], + "angle": 0, + "content": "Figure 1. Figure 1a illustrates a baseline APR architecture. It consists of a CNN backbone that generates a feature vector encoding the input query image, followed by regression heads that estimate the translation and orientation. Figure 1b shows the extension of this baseline architecture with a hypernetwork that outputs the weights for the regression heads (translation and orientation) during inference." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.909, + 0.903 + ], + "angle": 0, + "content": "cessing pipeline that begins with image retrieval to perform an initial localization based on images similar to the query image. The initial estimate is refined by local feature extraction and matching. To estimate the camera's 6-Degrees-Of-Freedom (6DoF) accurately, the resulting 2D-2D matches are mapped to 3D correspondences in the scene's point cloud and then used to determine the camera pose through Perspective-n-Point (PnP) and RANSAC. SCR-based methods [4-8, 47] directly regress correspondences between 2D pixels and 3D coordinates in the scene. Relative pose regression (RPR) schemes [2, 14, 17, 20, 26, 57, 62] directly estimate the relative pose between a query image and a set of reference images with well-defined ground truth poses, eliminating the need for conventional pipelines. However, an image retrieval scheme is essential to retrieve a set of nearest-neighbor images. Another approach, absolute pose" + }, + { + "type": "page_footnote", + "bbox": [ + 0.108, + 0.875, + 0.38, + 0.888 + ], + "angle": 0, + "content": "1https://ronferens.github.io/hyperpose/" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.461, + 0.9 + ], + "angle": 0, + "content": "2https://ronferens.github.io/extcambridgelandmarks/" + }, + { + "type": "list", + "bbox": [ + 0.108, + 0.875, + 0.461, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11547" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "regressors (APRs), can estimate the camera pose with a single forward pass using only the query image, resulting in significantly lower latency, although notably less accurate. Additionally, due to their low memory footprint, APRs can be deployed as a standalone application on edge devices with limited computational resources. A survey of visual camera pose localization is given by [43, 65]." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.2, + 0.485, + 0.714 + ], + "angle": 0, + "content": "PoseNet by Kendall et al. [25] was the first APR approach, which leveraged a convolutional backbone and a multilayer perceptron (MLP) head to regress the camera's position and orientation. This simple and computationally efficient approach, enabling execution at 5ms, led to numerous APR methods that improved accuracy by modifying the backbone and MLP architectures [11, 32, 33, 44, 49, 60, 61, 63]. In addition, alternative loss functions and optimization strategies were explored [23, 24, 43]. Indoor and outdoor environments exhibit dynamic variability in illumination, perspective, and object motion. APR models, despite their computational efficiency, often struggle to maintain accuracy under these diverse conditions and show limited generalization to novel scenes, necessitating scene-specific training. Consequently, while architecturally simple, a unified network might yield suboptimal performance when adapting to heterogeneous real-world visual inputs. Recent studies have proposed extending APR from single-scene to multiscene formulations. Blanton et al. [3] implemented a convolutional neural network (CNN) backbone to compute a global descriptor for the input image. Rather than employing a scene-specific multilayer perceptron (MLP), they optimized a set of fully connected (FC) layers, allocating one layer per scene based on a predicted scene identifier. Although introducing a framework to optimize a single model across multiple scenes, this methodology did not achieve state-of-the-art (SOTA) APR accuracy. An alternative approach was presented by [46] using dual-branch transformers [58]. Encoders focused on extracting pose-related features, while decoders transformed encoded scene identifiers into latent pose representations. The decoder outputs are leveraged for scene classification and selection of corresponding position and orientation embeddings, which are used to regress position and orientation vectors." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.717, + 0.483, + 0.853 + ], + "angle": 0, + "content": "Hypernetworks [22] are a deep learning architecture in which an auxiliary network, the hypernetwork, generates the weights for the primary network (main network) based on the current input and downstream task. The main network and hypernetwork are trained end-to-end using backpropagation. The hypernetwork allows for dynamic and adaptive weight modulation of the main network, leading to improved flexibility in response to changing conditions [16, 18, 34, 55]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.902 + ], + "angle": 0, + "content": "In this work, we propose HyperPose, a hypernetwork-based approach for absolute camera pose regression. Extending common APR methods [11, 23, 25, 44, 45, 60]," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.422 + ], + "angle": 0, + "content": "the hypernetwork computes weights for the regression head. This enables adaptive analysis of image content, allowing the regression weights to focus on informative features, rather than relying on a single constant pretrained set of weights. We validated the proposed method by integrating hypernetworks into single- and multiscene APR architectures. For single-scene networks, we present HyperPose with an existing state-of-the-art APR model [61] as well as a simplistic baseline APR using an off-the-shelf backbone [53] and shallow regression heads, inspired by the principles outlined in [25]. For multiscene, we extend a leading multiscene APR model [46] by integrating a hypernetwork architecture. In both instances, hypernetwork-augmented models demonstrated superior performance compared to their original counterparts when evaluated on the benchmark datasets in this study. To evaluate localization robustness, we introduce an Extended Cambridge Landmarks (ECL) dataset - an augmentation of the original benchmark [25] with algorithmically generated environmental shifts. By combining evening, winter, and summer scenes, the ECL incorporates dynamics such as lighting shifts and seasonal transitions." + }, + { + "type": "text", + "bbox": [ + 0.536, + 0.424, + 0.84, + 0.438 + ], + "angle": 0, + "content": "Thus, we propose the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.44, + 0.905, + 0.483 + ], + "angle": 0, + "content": "- We propose an approach to enhance any existing absolute camera pose regression models by integrating hypernetwork architectures." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.484, + 0.905, + 0.558 + ], + "angle": 0, + "content": "- We comprehensively evaluated the proposed HyperPose on both single- and multiscene pose regressors, demonstrating improved accuracy achieved by the hypernetwork-enhanced models over the original versions." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.56, + 0.905, + 0.604 + ], + "angle": 0, + "content": "- The proposed scheme achieves new SOTA accuracy on multiscene APR benchmarks spanning diverse outdoor and indoor localization tasks." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.605, + 0.905, + 0.68 + ], + "angle": 0, + "content": "- We introduce the Extended Cambridge Landmarks (ECL) dataset, which builds upon the foundation of the original Cambridge Landmarks dataset [25]. Our ECL data set extends existing test scenes by incorporating various appearance conditions." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.44, + 0.905, + 0.68 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.694, + 0.655, + 0.71 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.72, + 0.905, + 0.749 + ], + "angle": 0, + "content": "Camera pose estimation techniques differ by their input during inference and algorithmic attributes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.902 + ], + "angle": 0, + "content": "3D-based or structure-based localization methods are considered the most accurate for camera pose estimation and have demonstrated state-of-the-art performance in leading benchmarks. These methods utilize correspondences between 2D features in an image and 3D world coordinates. [15, 27, 35, 37-41, 51, 54] introduced a two-phase approach for hierarchical pose estimation pipelines. Thus, each query image is first encoded using a CNN trained for image retrieval (IR), based on a pre-mapped image dataset. Tentative correspondences are then estimated by matching local im" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "11548" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.242 + ], + "angle": 0, + "content": "age features to 3D matches. The resulting matches are used by PnP-RANSAC to estimate the camera pose. DSAC [8] and its succeeding scheme, DSAC++ [4], utilize a CNN architecture to directly estimate 3D coordinates from 2D positions in an image and only require the query image during inference. The same as in Absolute Pose Regression (APR) methods, separate models must be trained for each scene, achieving state-of-the-art accuracy. Lately, [15] proposes a solution to mitigate this gap by introducing a scene-agnostic architecture." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.248, + 0.483, + 0.46 + ], + "angle": 0, + "content": "Relative Pose Regression (RPR) algorithms estimate the relative pose between an input query image and a set of reference images based on their known location. Calculating the relative pose involves estimating the relative position and orientation between the query image and the reference images [1, 6, 68]. An Image Retrieval (IR) approach is utilized to determine the closest set of neighbor images, and the relative motion is then estimated between the query image and each of the retrieved images. The estimated relative poses are used, along with the known position of the reference images, to estimate the camera's absolute pose. Similar to 3D and structured-based localization methods, RPRs involve a multistep process and require access to a database of images with labeled poses during the inference phase." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.464, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Absolute Pose Regression (APR) directly estimates the camera pose given the input query image. APRs were first introduced by Kendall [25], using a modified version of the truncated GoogLeNet architecture, where the Softmax classification layer was replaced by a series of fully connected layers that output the pose estimate. Although less accurate than classical structure-based methods, APRs offer several advantages, such as faster inference times (milliseconds), reduced memory footprint (megabytes), and do not require feature engineering. Variations of the encoder and MLP architectures have been proposed to improve APR accuracy [11, 23, 33, 33, 44, 49]. Furthermore, modifications to the loss function have been suggested to improve the balance between orientation and position objectives [24] or to incorporate other types of information [9]. For an in-depth examination of the various architectural designs and loss functions utilized in camera pose regression, please refer to [43]. Although demonstrating faster inference times and low memory requirements, absolute pose regression (APR) models typically require dedicated perscene training to achieve competitive accuracy compared to other methods. This limitation has raised concerns regarding the value of APRs versus significantly more accurate techniques such as [4, 6, 8, 36]. To address this issue, recent approaches suggested extending APRs to a multiscene paradigm. [3] introduced Multi-Scene PoseNet (MSPN), which utilizes a CNN to regress an image feature vector. The two branches then process this feature vector. The first branch classifies the related scene using a fully connected" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.302 + ], + "angle": 0, + "content": "layer with a softmax operator. The estimated scene index selects scene-specific weights from a database. The second branch combines these weights with the feature vector to regress the pose. The model is trained end-to-end by applying both binary cross-entropy (for classification) and camera pose losses for each scene. Inspired by recent successful applications of Transformers [12, 19, 29] to computer vision tasks, [46] suggested a shared convolutional network backbone followed by two network branches consisting of a Transformer and a regression head. The first regresses the input query image's pose, and the other its orientation. The main incentive for utilizing Transformers encoders is to focus on pose-informative features and decoders to transform encoded scene identifiers into latent pose representations." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.303, + 0.905, + 0.44 + ], + "angle": 0, + "content": "Hypernetworks were first presented by Ha et al. [22]. These are neural networks that have been designed to predict the weights of a primary network. The primary network's weights can be adjusted based on specific inputs, resulting in a more expressive and adaptive model. Hypernetworks have been used in various applications such as learning implicit neural representations [16], semantic segmentation [34], 3D scene representation and modeling [28, 48, 50] and continuous learning [59], to name a few." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.451, + 0.771, + 0.468 + ], + "angle": 0, + "content": "3. APRs using Hypernetworks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.476, + 0.905, + 0.703 + ], + "angle": 0, + "content": "HyperPose is a general approach, applicable to both single- and multi-scene scenarios, for enhancing APR schemes. It comprises two key components: the main network (APR) designed to estimate the camera's absolute position and the hypernetwork that computes the regression weights for the main network conditioned on the input query image. The main network performs a forward pass on the input image to localize the capture camera. The camera pose is represented as the tuple \\(< \\mathbf{x},\\mathbf{q}>\\), where \\(\\mathbf{x}\\in \\mathbb{R}^3\\) is the camera's position in world coordinates and \\(\\mathbf{q}\\in \\mathbb{R}^4\\) is the quaternion encoding of its \\(3D\\) orientation. Following [13], we formulate the hypernetwork \\(H\\) as a parametric function \\(\\theta_h = H(\\theta ,\\mathbf{x})\\), where \\(\\theta\\) represents the parameters of the hypernetwork, \\(\\mathbf{x}\\) the network's input, and \\(\\theta_h\\) the parameters of the regression heads layer for the input query image \\(I_q\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.71, + 0.719, + 0.724 + ], + "angle": 0, + "content": "3.1. Network Architecture" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.732, + 0.796, + 0.746 + ], + "angle": 0, + "content": "3.1.1. HyperPose for Single-Scene APRs" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.905, + 0.795 + ], + "angle": 0, + "content": "We apply the proposed HyperPose-based approach to single-scene APR using a baseline APR following [25], and the state-of-the-art AtLoc APR [61]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Baseline APR with Hypernetwork. As depicted in Fig. 1a, the baseline model's main network consists of a CNN backbone that generates a feature vector encoding the input query image. This vector is then processed by two regression heads. The first estimates the camera position, and the other estimates the orientation. Figure 1b illustrates the integration of a hypernetwork into this architecture. The input" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "11549" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.116, + 0.098, + 0.898, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.421, + 0.908, + 0.496 + ], + "angle": 0, + "content": "Figure 2. MS-HyperPose - The proposed multi-scene absolute pose regression architecture using a hypernetwork. The primary network employs position and orientation Transformers, in a dual-branch architecture, to extract activation maps from the underlying convolutional backbone. The hypernetwork generates the weights for the regression head in the primary network, using the input query image. These adaptive weights, combined with the latent vectors generated by the Transformers estimate the camera pose, consisting of the spatial position (x) and angular orientation (q)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.518, + 0.482, + 0.638 + ], + "angle": 0, + "content": "to the hypernetwork is the feature vector generated by the CNN backbone of the main network. This vector is passed through a single fully-connected (FC) layer followed by a Swish activation [36] function. The processed vector serves as an input to a multilayer perceptron (MLP) that outputs the weights for a single layer in the regression heads of the main network. The camera pose is then estimated by the position and orientation regression heads of the main network." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.641, + 0.483, + 0.822 + ], + "angle": 0, + "content": "AtLoc with Hypernetwork. [61] presents Attention-Guided Camera Localization (AtLoc), an absolute pose regression model based on self-attention. In this architecture, a visual encoder encodes an image into a latent vector. Conditioned on these extracted features, an attention module computes self-attention maps to reweight the representation into a new feature space. A pose regressor then maps the transformed features after the attention operators to estimate the camera pose. We incorporate a hypernetwork into the AtLoc, the same as for the baseline APR model. In this case, the input to the hypernetwork is the reweighted representation feature vector computed by the visual encoder." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.835, + 0.37, + 0.851 + ], + "angle": 0, + "content": "3.1.2. HyperPose for Multi-Scene APRs" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.903 + ], + "angle": 0, + "content": "The suggested multi-scene model, named MS-HyperPose is an extension of the architecture introduced in [45]. As depicted in Fig. 2, the input query image \\( I_{q} \\) is processed by a" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.518, + 0.907, + 0.73 + ], + "angle": 0, + "content": "convolutional backbone that computes intermediate activation maps from two different layers. These activation maps are then utilized as inputs to the main network and the hypernetwork. Within the main network, the positional and orientational Transformers process the activation maps and integrate sequential representations into single latent vectors. After converting the activation maps into representation vectors, these vectors, along with the latent vectors are passed through multilayer perceptron (MLP) layers in the hypernetwork to generate the weights \\(\\theta_h\\) for the translation and orientation regression heads. Using the weights \\(\\theta_h\\) produced by the hypernetwork \\(H\\) and the regression heads in the main network, the position, and orientation of the input query image are estimated." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.903 + ], + "angle": 0, + "content": "The Main Network consists of a convolutional backbone and two separate branches to regress the camera's position and orientation. Each branch includes its particular Transformer and a multilayer perceptron (MLP) head. Studies have demonstrated that substituting the patchify stem, utilized in the early visual processing of a Vision Transformer (ViT), with a convolutional stem enhances optimization stability and elevates performance to a higher level [64]. Therefore, given an input query image \\( I_{q} \\in \\mathbb{R}^{H \\times W \\times C} \\), a shared convolutional backbone generates activation maps. Two different activation maps are selected" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "11550" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.155 + ], + "angle": 0, + "content": "from the CNN based on the particular learned task (Fig. 2a). The activation maps, represented as \\( M \\in \\mathbb{R}^{h \\times w \\times C^{Fn}} \\), are processed and transformed into \\( \\tilde{M} \\in \\mathbb{R}^{h \\times w \\times C^{Out}} \\) using a \\( 1 \\times 1 \\) convolutional layer followed by flattening (Fig. 2b)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.156, + 0.483, + 0.321 + ], + "angle": 0, + "content": "Following [45], we use the Transformer architecture from [12], modifying the standard Encoder and Decoder to add positional encoding at each attention layer. Each Transformer has \\( L \\) identical blocks, each with a multi-head attention (MHA) mechanism and a two-layer MLP with Swish activation [36]. Layer Normalization (LN) is applied before MHA and MLP processing, as in [58], with outputs merged via residual connections and dropout. To preserve spatial information, each activation map position receives a learned encoding. The processed sequence, serving as input to the Transformer is given by (Fig. 2c):" + }, + { + "type": "equation", + "bbox": [ + 0.142, + 0.344, + 0.483, + 0.365 + ], + "angle": 0, + "content": "\\[\nE ^ {I n} = \\left[ T, \\tilde {M} \\right] \\in \\mathbb {R} ^ {(h w) \\times C ^ {O u t}}, T \\in \\mathbb {R} ^ {C _ {M} ^ {O u t}}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.376, + 0.482, + 0.391 + ], + "angle": 0, + "content": "where \\(T\\) is the positional encoding of the activation map." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.391, + 0.482, + 0.48 + ], + "angle": 0, + "content": "Given a dataset with \\(N\\) scenes, the Transformer Decoders generate embeddings \\(E^{Out} = \\in \\mathbb{R}^{N\\times C^{Out}}\\) (Fig. 2d). For a query image, only one position matches its scene. The combined Transformer outputs (Fig. 2e) pass through a fully connected layer and Log SoftMax, selecting the highest-probability vectors (Fig. 2f)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.481, + 0.482, + 0.557 + ], + "angle": 0, + "content": "The selected vectors \\( E_{x_i}^{Out} \\) and \\( E_{q_i}^{Out} \\) are processed through an MLP. Finally, the estimated camera position and orientation are obtained by summing the outputs of the main network and hypernetwork, using weights predicted by the hypernetwork (Fig. 2h and 2i)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.557, + 0.483, + 0.721 + ], + "angle": 0, + "content": "The Hypernetworks consist of MLPs, each regressing the weights for a corresponding regression head (position or orientation) in the main network. Their input is the sum of task-specific encoded activation maps \\(\\tilde{M}_x\\) and \\(\\tilde{M}_q\\), along with global latent vectors \\(E_{x_i}^{Out}\\) and \\(E_{q_i}^{Out}\\) (Fig. 2g). To ensure size compatibility, the encoded activation map passes through a fully connected layer before summation. This vector is then processed by MLPs with Swish activation, generating inputs for another MLP set, each estimating weights for a specific layer in the main network's regression heads:" + }, + { + "type": "equation", + "bbox": [ + 0.141, + 0.745, + 0.483, + 0.769 + ], + "angle": 0, + "content": "\\[\n\\theta_ {h} ^ {O u t _ {j}} = H \\left(\\theta_ {h} ^ {O u t _ {j}}, E _ {i} ^ {O u t}\\right) \\in \\mathbb {R} ^ {C _ {E _ {i}} ^ {O u t} \\times C _ {H} ^ {O u t _ {j}}}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.779, + 0.433, + 0.795 + ], + "angle": 0, + "content": "where \\(j\\) is the index of the regression head layer." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.803, + 0.296, + 0.82 + ], + "angle": 0, + "content": "3.2. Optimization Criteria" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.483, + 0.901 + ], + "angle": 0, + "content": "The localization error is calculated by comparing the translation position and orientation of the ground truth pose \\(\\mathbf{p}_{gt} = < \\mathbf{x}_{gt},\\mathbf{q}_{gt}>\\) , where \\(\\mathbf{x}\\in \\mathbb{R}^3\\) represents the position of the camera in the world and \\(\\mathbf{q}\\in \\mathbb{R}^4\\) denotes its orientation encoded as a quaternion, with the estimated pose" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.182 + ], + "angle": 0, + "content": "\\(\\mathbf{p}_{est} = < \\mathbf{x}_{est},\\mathbf{q}_{est} >\\) . The translation error, represented by \\(L_{x}\\) , is calculated by the Euclidean distance between the ground truth and the estimated camera positions. The orientation error, \\(L_{q}\\) , is defined as the Euclidean distance between the ground truth and the estimated unit quaternions [43]:" + }, + { + "type": "equation", + "bbox": [ + 0.648, + 0.197, + 0.905, + 0.213 + ], + "angle": 0, + "content": "\\[\nL _ {x} = \\left\\| \\mathbf {x} _ {0} - \\mathbf {x} \\right\\| _ {2}, \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.64, + 0.219, + 0.905, + 0.253 + ], + "angle": 0, + "content": "\\[\nL _ {q} = \\left\\| \\mathbf {q} _ {0} - \\frac {\\mathbf {q}}{| | \\mathbf {q} | |} \\right\\| _ {2} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.257, + 0.905, + 0.286 + ], + "angle": 0, + "content": "where \\(\\mathbf{q}\\) is a normalized norm quaternion to ensure it is a valid orientation encoding." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.288, + 0.906, + 0.378 + ], + "angle": 0, + "content": "As a multi-regression optimization problem, we adopt the method outlined in [24] by using different weighting schemes, based on the uncertainty of each task, to balance the individual loss functions defined for the separate objectives. The model optimizes the aggregated loss function, a combination of the weighted individual losses." + }, + { + "type": "equation", + "bbox": [ + 0.552, + 0.405, + 0.905, + 0.422 + ], + "angle": 0, + "content": "\\[\nL _ {p} = L _ {x} \\exp (- s _ {x}) + s _ {x} + L _ {q} \\exp (- s _ {q}) + s _ {q}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.432, + 0.906, + 0.522 + ], + "angle": 0, + "content": "where \\( S_{x} \\) and \\( S_{q} \\) are the learned parameters. As a further optimization, we adopt the approach detailed in [44] and follow a three-step training process. Firstly, the entire network is trained using the aggregated loss in Eq. 5. In the next two stages, each MLP head undergoes separate fin-tuning with its corresponding loss function." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.532, + 0.731, + 0.547 + ], + "angle": 0, + "content": "3.3. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.906, + 0.719 + ], + "angle": 0, + "content": "The shared backbone uses EfficientNet-B0 [31, 52]. We used the activation maps of two endpoints (reduction levels) as input for our position and orientation branches: \\( m_{rdct4} \\in \\mathbb{R}^{14 \\times 14 \\times 112} \\) and \\( m_{rdct3} \\in \\mathbb{R}^{28 \\times 28 \\times 40} \\), respectively. Following [45], we applied linear projections on each activation map to a common depth dimension of \\( C_M^{Out} = 256 \\) and learned the positional encoding of the same depth. The regression heads in the main network consist of two-layer MLPs with input and output layers. The input layer dimension is 1024 for both the position and orientation, while the output dimensions are 3 and 4, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.905, + 0.795 + ], + "angle": 0, + "content": "Single-Scene APR. For both the baseline APR and the HyperPose extension of the AtLoc [61] model, the hypernetwork for the position branch generates weights for an MLP with a dimension of 256, while the orientation hypernetwork produces weights of 512 dimensions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Multi-Scene APR. In contrast to [45], we incorporated four-layer (instead of six) Transformers in each branch (position and orientation) within the main network architecture. Each block consists of an MHA layer with four heads, followed by an MLP that preserves the input dimension. A dropout rate of \\( p = 0.1 \\) was applied to each MLP layer. The output dimension of each Transformer in the main network" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "11551" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.132 + ], + "angle": 0, + "content": "Table 1. Comparative analysis of APR architectures with and without the purported hypernetwork addition using the Cambridge Landmarks dataset. We report the median position/orientation error in meters/degrees. The best performance is marked in bold. (*) The authors of AtLoc did not report results on the Cambridge dataset, and we trained the AtLoc model on this dataset." + }, + { + "type": "table", + "bbox": [ + 0.19, + 0.142, + 0.805, + 0.218 + ], + "angle": 0, + "content": "
MethodK. CollegeOld HospitalShop FacadeSt. MaryAvg.
AtLoc* [61]1.53,3.212.17,3.990.93,4.612.14,5.761.69,4.39
AtLoc w/ hypernet0.96,3.432.04,3.610.88,4.221.72,5.441.40,4.18
Baseline APR0.89,2.291.49,3.320.74,4.791.40,4.951.13,3.84
Baseline APR w/ hypernet0.77,2.071.47,3.290.72,4.011.37,4.901.08,3.57
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.23, + 0.485, + 0.412 + ], + "angle": 0, + "content": "was \\( C_M^{Out} = 256 \\). The position and orientation branches of the hypernetwork generate the weights of three regression layers: input, hidden, and output. The number of weights in the position and orientation branches differ. The hypernetwork's position branch produces the weights for an MLP with dimensions 256, 128, and 3, while the orientation branch generates weights for an MLP with dimensions 512, 512, and 4. The weights generated by the hypernetwork are then put into the regression heads of the main network to estimate the position and orientation of the input query image. The input and hidden layers of the regression head utilize the Swish activation function." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.427, + 0.466, + 0.446 + ], + "angle": 0, + "content": "4. Extended Cambridge Landmarks Dataset" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.467, + 0.283, + 0.55 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.157, + 0.552, + 0.22, + 0.564 + ], + "angle": 0, + "content": "(a) Original" + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.467, + 0.483, + 0.55 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.356, + 0.553, + 0.42, + 0.564 + ], + "angle": 0, + "content": "(b)Evening" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.565, + 0.283, + 0.648 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.156, + 0.651, + 0.22, + 0.661 + ], + "angle": 0, + "content": "(c) Summer" + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.565, + 0.483, + 0.647 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.36, + 0.651, + 0.416, + 0.661 + ], + "angle": 0, + "content": "(d) Winter" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.674, + 0.483, + 0.703 + ], + "angle": 0, + "content": "Figure 3. A sample from the King's College Scene in the proposed Extended Cambridge Landmarks (ECL) Dataset" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.901 + ], + "angle": 0, + "content": "To further assess the limitations of the HyperPose approach, we constructed an augmented version of the Cambridge Landmarks [25] benchmark called the Extended Cambridge Landmarks (ECL) dataset. Leveraging the original scenes, the ECL dataset introduces new localization challenges due to variations in appearance by creating synthetic evening, winter, and summer variants for each location. These scene permutations simulate real-world environmental variations by introducing shifts in lighting, weather patterns, foliage density, and more. Analyzing performance across ECL variations and their original Cambridge counterparts allows us to evaluate the robustness of" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.231, + 0.908, + 0.457 + ], + "angle": 0, + "content": "models to real-world variations. These augmentations were conducted by utilizing the novel InstructPix2Pix work [10]. Visual samples of the ECL dataset can be found in the supplementary materials. Figure 3 shows sample from the King's College scene in the extended dataset. Additional samples from each scene in the ECL dataset can be found in the supplementary material. Quantitative pose regression analysis between these more challenging conditions and unmodified sequences helps scientifically assess model adaptability. We hypothesize that exposing limitations unseen in conventional test sets will motivate the development of more invariant solutions. As controllable expansions of established benchmarks, programmatically generated datasets can thus guide the development of robust models for real-world use." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.471, + 0.72, + 0.488 + ], + "angle": 0, + "content": "5. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.493, + 0.909, + 0.885 + ], + "angle": 0, + "content": "We conducted an experimental evaluation using the 7 Scenes [21] and Cambridge Landmarks [25] datasets for evaluating absolute pose regression. The Cambridge Landmarks dataset represents an outdoor urban environment ranging in size from 900 to 5500 square meters. We present results for four of its six scenes. We excluded the remaining two scenes due to limited coverage in prior work. To compare robustness, we also report the results on the Extended Cambridge Landmarks dataset (ECL) (Section 1 in the supplementary materials). The 7 Scenes dataset consists of seven small-scale indoor scenes with spatial extents ranging from approximately 1 to 10 square meters. These datasets present various localization challenges, including differences in scale, indoor versus outdoor settings, repetitive elements, textureless features, significant viewpoint changes, and variations in trajectory between training and test sets. To facilitate a broader analysis of our proposed method, we conduct an additional evaluation on the Oxford RobotCar Dataset [30]. Such dynamic environmental factors assess HyperPose's ability to improve localization accuracy amid diverse, challenging visual contexts. Our experimental procedures adhere to the same training and evaluation protocols as in [9, 61]. Additional information concerning these sequences, and a comprehensive description of the training methodology and hyperparameter selection for the models, are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.886, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We evaluate the impact of integrating hypernetworks in" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11552" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.119 + ], + "angle": 0, + "content": "Table 2. Comparison to single-scene state-of-the-art methods - Cambridge Landmarks: We report the median position/orientation error in meters/degrees. The most effective single and multi-scene APRs are marked in bold." + }, + { + "type": "table", + "bbox": [ + 0.173, + 0.128, + 0.825, + 0.469 + ], + "angle": 0, + "content": "
K. CollegeHospitalShop FacadeSt. MaryAvg.
Seq.DSAC* [6]0.18,0.3°0.21,0.4°0.05,0.3°0.15,0.5°0.15,0.4°
MapNet [9]1.08,1.9°1.94,3.9°1.49,4.2°2.00,4.5°1.63,3.6°
GL-Net [67]0.59,0.7°1.88,2.8°0.50,2.9°1.90,3.3°1.22,2.4°
IRVLAD [56]2.80,5.7°4.01,7.1°1.11,7.6°2.31,8.0°2.56,7.1°
VLAD+Inter [42]1.48,4.5°2.68,4.6°0.90,4.3°1.62,6.1°1.67,4.9°
RPREssNet [68]0.76,1.9°1.39,2.8°0.84,4.3°1.32,4.7°1.08,3.4°
NC-EssNet [68]0.61,1.6°0.95,2.7°0.7,3.4°1.12,3.6°0.85,2.8°
RelocGNN[21]0.48,1.0°1.14,2.5°0.48,2.5°1.52,3.2°0.91,2.3°
APRPoseNet [25]1.92,5.40°2.31,5.38°1.46,8.08°2.65,8.48°2.08,6.83°
BayesianPN [23]1.74,4.06°2.57,5.14°1.25,7.54°2.11,8.38°1.91,6.28°
LSTM-PN [60]0.99,3.65°1.51,4.29°1.18,7.44°1.52,6.68°1.30,5.57°
SVS-Pose [33]1.06,2.81°1.50,4.03°0.63,5.73°2.11,8.11°1.32,5.17°
GPoseNet [11]1.61,2.29°2.62,3.89°1.14,5.73°2.93,6.46°2.07,4.59°
PoseNetLearn [24]0.99,1.06°2.17,2.94°1.05,3.97°1.49,3.43°1.42,2.85°
GeoPoseNet [24]0.88,1.04°3.20,3.29°0.88,3.78°1.57,3.32°1.63,2.86°
IRPNet [44]1.18,2.19°1.87,3.38°0.72,3.47°1.87,4.94°1.41,3.50°
Baseline APR w/ hypernet0.61,1.84°1.44,3.03°0.70,3.62°1.37,4.85°1.03,2.67°
MS-APRMSPN [3]1.73,3.65°2.55,4.05°2.92,7.49°2.67,6.18°2.47,5.34°
MS-Trans[46]0.83,1.47°1.81,2.39°0.86,3.07°1.62,3.99°1.28,2.73°
MS-HyperPose0.78,1.18°1.84,2.11°0.83,3.13°1.61,3.22°1.27,2.41°
" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.482, + 0.482, + 0.89 + ], + "angle": 0, + "content": "single-scene APRs. To this end, we compare the accuracy of two model variations, with and without hypernetworks, and report median position and orientation errors. Table 1 presents results on the Cambridge dataset. The hypernetwork improves the accuracy of the proposed baseline APR and AtLoc [61] models. The baseline APR model size is \\(26.7\\mathrm{MB}\\) with an inference time of \\(9.32\\mathrm{ms}\\), increasing to \\(40.9\\mathrm{MB}\\) and \\(9.45\\mathrm{ms}\\) with a hypernetwork. The original published AtLoc model is \\(97.9\\mathrm{MB}\\) with an inference time of \\(5.14\\mathrm{ms}\\). The hypernetwork-enhanced AtLoc version has a model size of \\(116.8\\mathrm{MB}\\) and an inference time of \\(5.56\\mathrm{ms}\\). Note that the reported errors refer to the models' performance after the first stage of model training, without applying the two additional optimization steps described in the training details (See supplementary materials). Furthermore, Tables 2 and 3 compare the median position error, and the orientation error, of the proposed baseline APR with hypernetwork architecture, with the results of recent SOTA localization algorithms, using the Cambridge Landmarks and 7Scenes datasets, respectively. For the indoor and outdoor datasets, in addition to the results of leading APRs, we report the median position and orientation errors of structure-based, sequence-based, IR, and RPR methods. Although DSAC* [6] achieves SOTA accuracy, compared to all other APR methods, HyperPose obtains the most accurate average of \\(1.03m\\), \\(2.67^{\\circ}\\) on the Cambridge outdoor dataset, and \\(0.19m\\), \\(7.48^{\\circ}\\) on the 7Scenes indoor dataset. These local-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.482, + 0.905, + 0.782 + ], + "angle": 0, + "content": "ization errors show competitive performance compared to leading RPR techniques, with the hypernet-enhanced baseline ranking third and second/fourth (7Scenes/Cambridge) in positional and rotational accuracy, respectively. Next, we compare the proposed MS-HyperPose architecture with other multiscene APR methods. Tables 2 and 3 present median position and orientation errors, respectively, on the Cambridge Landmarks and 7 Scenes datasets. On both indoor and outdoor benchmarks, our MS-HyperPose approach achieves the most accurate average of \\(1.27\\mathrm{m}\\), \\(2.41^{\\circ}\\) on Cambridge and \\(0.18\\mathrm{m}\\), \\(7.12^{\\circ}\\) on 7Scenes compared to other single- and multi-scene APRs. Additionally, MS-HyperPose is the most accurate for both positional and rotational errors in 6 out of 8 parameters on Cambridge and 6 out of 14 on 7 Scenes. The runtime of MS-HyperPose is \\(30.93\\mathrm{ms}\\), similar to the reference architecture of [46]. Its model size is 571MB, notably larger than the 71.1MB of the original. We discuss storage considerations in the Limitation and Future Work section of the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.905, + 0.9 + ], + "angle": 0, + "content": "Table 4 reports the mean and median position and orientation errors of AtLoc [61] compared to our hypernetwork-enhanced variant using the Oxford RobotCar Dataset [30]. We also compare the enhanced model with [9], the SOTA APR for this dataset using image sequences. Visualizations of the localization results can be found in the supplementary materials." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "11553" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.119 + ], + "angle": 0, + "content": "Table 3. Comparison to single-scene SOTA methods - 7Scenes: We report the median position/orientation error in meters/degrees. The most effective single and multi-scene APRs are distinguished in bold." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.128, + 0.918, + 0.566 + ], + "angle": 0, + "content": "
ChessFireHeadsOfficePumpkinKitchenStairsAvg.
SCRDSAC* [6]0.02,1.11°0.02,1.24°0.01,1.82°0.03,1.18°0.04,1.41°0.03,1.70°0.04,1.42°0.03,1.41°
Marepo[15]0.02,1.24°0.02,1.39°0.02,2.03°0.03,1.26°0.04,1.48°0.04,1.71°0.06,1.67°0.03,1.54°
Seq.LsG [66]0.09,3.3°0.26,10.9°0.17,12.7°0.18,5.5°0.20,3.7°0.23,4.9°0.23,11.3°0.19,7.74°
MapNet[9]0.08,3.3°0.27,11.7°0.18,13.3°0.17,5.2°0.22,4.0°0.23,4.9°0.30,12.1°0.21,7.78°
GL-Net[67]0.08,2.8°0.26,8.9°0.17,11.4°0.18,13.3°0.15,2.8°0.25,4.5°0.23,8.8°0.19,7.50°
IRVLAD [56]0.21,12.5°0.33,13.8°0.15,14.9°0.28,11.2°0.31,11.2°0.30,11.3°0.25,12.3°0.26,12.46°
VLAD+Inter[42]0.18,10.0°0.33,12.4°0.14,14.3°0.25,10.1°0.26,9.4°0.27,11.1°0.24,14.7°0.24,11.71°
RPRNN-Net [26]0.13,6.5°0.26,12.7°0.14,12.3°0.21,7.4°0.24,6.4°0.24,8.0°0.27,11.8°0.21,9.30°
RelocNet[2]0.12,4.1°0.26,10.4°0.14,10.5°0.18,5.3°0.26,4.2°0.23,5.1°0.28,7.5°0.21,6.73°
EssNet [68]0.13,5.1°0.27,10.1°0.15,9.9°0.21,6.9°0.22,6.1°0.23,6.9°0.32,11.2°0.22,8.03°
NC-EssNet [68]0.12,5.6°0.26,9.6°0.14,10.7°0.20,6.7°0.22,5.7°0.22,6.3°0.31,7.9°0.21,7.50°
RelocGNN[21]0.08,2.7°0.21,7.5°0.13,8.70°0.15,4.1°0.15,3.5°0.19,3.7°0.22,6.5°0.16,5.24°
CamNet [17]0.04,1.7°0.03,1.7°0.05,2.00°0.04,1.60°0.04,1.6°0.04,1.6°0.04,1.5°0.04,1.70°
APRPoseNet [25]0.32,8.12°0.47,14.4°0.29,12.0°0.48,7.68°0.47,8.42°0.59,8.64°0.47,13.8°0.44,10.4°
BayesianPN [23]0.37,7.24°0.43,13.7°0.31,12.0°0.48,8.04°0.61,7.08°0.58,7.54°0.48,13.1°0.47,9.81°
LSTM-PN [60]0.24,5.77°0.34,11.9°0.21,13.7°0.30,8.08°0.33,7.00°0.37,8.83°0.40,13.7°0.31,9.85°
GPoseNet [11]0.20,7.11°0.38,12.3°0.21,13.8°0.28,8.83°0.37,6.94°0.35,8.15°0.37,12.5°0.31,9.95°
PoseNetLearn[24]0.14,4.50°0.27,11.8°0.18,12.1°0.20,5.77°0.25,4.82°0.24,5.52°0.37,10.6°0.24,7.87°
GeoPoseNet[24]0.13,4.48°0.27,11.3°0.17,13.0°0.19,5.55°0.26,4.75°0.23,5.35°0.35,12.4°0.23,8.12°
IRPNet[44]0.13,5.64°0.25,9.67°0.15,13.1°0.24,6.33°0.22,5.78°0.30,7.29°0.34,11.6°0.23,8.49°
AtLoc[61]0.10,4.07°0.25,11.4°0.16,11.8°0.17,5.34°0.21,4.37°0.23,5.42°0.26,10.5°0.20,7.56°
TransBoNet[49]0.11,4.48°0.25,12.5°0.18,14.0°0.20,5.08°0.19,4.77°0.17,5.35°0.30,13.0°0.20,8.45°
Baseline APR w/ hypernet0.09,4.47°0.28,9.44°0.15,11.4°0.16,6.29°0.19,4.62°0.18,6.95°0.28,9.15°0.19,7.48°
MS-APRMSPN[3]0.09,4.76°0.29,10.5°0.16,13.1°0.16,6.80°0.19,5.50°0.21,6.61°0.31,11.6°0.20,7.56°
MS-Trans[46]0.11,4.66°0.24,9.60°0.14,12.2°0.17,5.66°0.18,4.44°0.17,5.94°0.26,8.45°0.18,7.28°
MS-HyperPose0.11,4.34°0.23,9.79°0.13,10.7°0.17,6.05°0.16,5.24°0.17,6.86°0.27,6.00°0.18,7.00°
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.58, + 0.907, + 0.61 + ], + "angle": 0, + "content": "Table 4. Comparative analysis of different APR architectures with and without the proposed hypernetwork to the Oxford RobotCar dataset: We report the mean and median position/orientation error in meters/degrees." + }, + { + "type": "table", + "bbox": [ + 0.214, + 0.62, + 0.781, + 0.727 + ], + "angle": 0, + "content": "
MapNet[9]AtLoc[61]AtLoc w/ hypernet
SequenceMeanMedianMeanMedianMeanMedian
LOOP18.76, 3.465.79, 1.548.61, 4.585.68, 2.238.42, 3.304.91, 1.33
LOOP29.84, 3.964.91, 1.678.61, 4.585.68, 2.238.10, 3.044.73, 1.27
FULL141.4, 12.517.94, 6.6829.6, 12.411.10, 5.2829.58, 6.777.86, 1.38
FULL259.3, 14.820.04, 6.3948.2, 11.112.2, 4.6345.96, 9.608.38, 1.45
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.74, + 0.486, + 0.877 + ], + "angle": 0, + "content": "We also assess the performance of two configurations of the baseline APR model, with and without hypernetworks on the Extended Cambridge dataset (ECL). Across all scenes and variations, the hypernetwork-enhanced model outperforms the baseline, demonstrating better adaptation to the seasonal changes and diverse lighting conditions inherent in the scenes. The detailed analysis elucidating the robustness and resilience of HyperPose to scene changes is in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.739, + 0.64, + 0.755 + ], + "angle": 0, + "content": "6. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this work, we proposed a hypernetwork-based approach to improve APR models. Our experiments show that hypernetworks improve accuracy for single-scene APR models. We also present MS-HyperPose, a novel multiscene camera pose regression model enhanced with a hypernetwork, which achieves new SOTA accuracy in indoor and outdoor localization datasets. We introduce the Extended Cambridge Landmarks (ECL) dataset to facilitate research on robust localization methods." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "11554" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.198 + ], + "angle": 0, + "content": "[1] Eduardo Arnold, Jamie Wynn, Sara Vicente, Guillermo Garcia-Hernando, Aron Monszpart, Victor Prisacariu, Daniyar Turmukhambetov, and Eric Brachmann. Map-free visual relocalization: Metric pose relative to a single image. In Proceedings of the European Conference on Computer Vision (ECCV), pages 690-708. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.2, + 0.484, + 0.255 + ], + "angle": 0, + "content": "[2] Vassileios Balntas, Shuda Li, and Victor Prisacariu. Relocnet: Continuous metric learning relocalisation using neural nets. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.257, + 0.483, + 0.324 + ], + "angle": 0, + "content": "[3] Hunter Blanton, Connor Greenwell, Scott Workman, and Nathan Jacobs. Extending absolute pose regression to multiple scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 38-39, 2020. 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.326, + 0.483, + 0.381 + ], + "angle": 0, + "content": "[4] E. Brachmann and C. Rother. Learning less is more - 6D camera localization via 3D surface regression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4654-4662, 2018. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.383, + 0.484, + 0.437 + ], + "angle": 0, + "content": "[5] Eric Brachmann and Carsten Rother. Expert sample consensus applied to camera re-localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7525-7534, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.439, + 0.483, + 0.493 + ], + "angle": 0, + "content": "[6] E. Brachmann and C. Rother. Visual Camera Re-Localization From RGB and RGB-D Images Using DSAC. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(09):5847-5865, 2022. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.496, + 0.483, + 0.577 + ], + "angle": 0, + "content": "[7] Eric Brachmann, Frank Michel, Alexander Krull, Michael Ying Yang, Stefan Gumhold, and Carsten Rother. Uncertainty-driven 6D pose estimation of objects and scenes from a single rgb image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3364-3372, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.58, + 0.483, + 0.661 + ], + "angle": 0, + "content": "[8] E. Brachmann, A. Krull, S. Nowozin, J. Shotton, F. Michel, S. Gumhold, and C. Rother. DSAC - differentiable RANSAC for camera localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2492-2500, Los Alamitos, CA, USA, 2017. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.664, + 0.483, + 0.731 + ], + "angle": 0, + "content": "[9] Samarth Brahmbhatt, Jinwei Gu, Kihwan Kim, James Hays, and Jan Kautz. Geometry-aware learning of maps for camera localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.734, + 0.483, + 0.803 + ], + "angle": 0, + "content": "[10] T. Brooks, A. Holynski, and A. A. Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18392-18402, Los Alamitos, CA, USA, 2023. IEEE Computer Society. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.483, + 0.857 + ], + "angle": 0, + "content": "[11] Ming Cai, Chunhua Shen, and Ian Reid. A hybrid probabilistic model for camera relocalization. In Proceedings of the British Machine Vision Conference, page 238, 2018. 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[12] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Proceedings" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.094, + 0.905, + 0.12 + ], + "angle": 0, + "content": "of the European Conference on Computer Vision (ECCV), pages 213-229, Cham, 2020. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.123, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[13] Yoav Chai, Raja Giryes, and Lior Wolf. Supervised and unsupervised learning of parameterized color enhancement. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 992-1000, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[14] Kefan Chen, Noah Snavely, and Ameesh Makadia. Wide-baseline relative camera pose estimation with directional learning. 2021 IEEE. In CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3257-3267, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.25, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[15] Shuai Chen, Tommaso Cavallari, Victor Adrian Prisacariu, and Eric Brachmann. Map-relative pose regression for visual re-localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20665-20674, 2024. 1, 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.321, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[16] Yinbo Chen and Xiaolong Wang. Transformers as meta-learners for implicit neural representations. In Proceedings of the European Conference on Computer Vision (ECCV), pages 170–187. Springer, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.378, + 0.905, + 0.432 + ], + "angle": 0, + "content": "[17] Mingyu Ding, Zhe Wang, Jiankai Sun, Jianping Shi, and Ping Luo. CamNet: Coarse-to-fine retrieval for camera relocalization. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2019. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.434, + 0.907, + 0.517 + ], + "angle": 0, + "content": "[18] Siyan Dong, Qingnan Fan, He Wang, Ji Shi, Li Yi, Thomas Funkhouser, Baoquan Chen, and Leonidas J Guibas. Robust neural routing through space partitions for camera relocalization in dynamic indoor environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8544-8554, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.519, + 0.905, + 0.615 + ], + "angle": 0, + "content": "[19] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In Proceedings of the International Conference on Learning Representations (ICLR), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.617, + 0.905, + 0.672 + ], + "angle": 0, + "content": "[20] Sovann En, Alexis Lecheryv, and Frédéric Jurie. Rpnet: An end-to-end network for relative camera pose estimation. In Proceedings of the European Conference on Computer Vision (ECCV) Workshops, pages 0–0, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.674, + 0.905, + 0.729 + ], + "angle": 0, + "content": "[21] B. Glocker, S. Izadi, J. Shotton, and A. Criminisi. Realtime rgb-d camera relocalization. In 2013 IEEE International Symposium on Mixed and Augmented Reality (ISMAR), pages 173-179, 2013. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.731, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[22] David Ha, Andrew Dai, and Quoc V Le. Hypernetworks. arXiv preprint arXiv:1609.09106, 2016. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.76, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[23] Alex Kendall and Roberto Cipolla. Modelling uncertainty in deep learning for camera relocalization. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), pages 4762-4769, 2016. 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.817, + 0.907, + 0.871 + ], + "angle": 0, + "content": "[24] A. Kendall and R. Cipolla. Geometric loss functions for camera pose regression with deep learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6555-6564, 2017. 2, 3, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[25] A. Kendall, M. Grimes, and R. Cipolla. Posenet: A convolutional network for real-time 6-dof camera relocalization. In" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.094, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11555" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 2938-2946, 2015. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.483, + 0.189 + ], + "angle": 0, + "content": "[26] Z. Laskar, I. Melekhov, S. Kalia, and J. Kannala. Camera relocalization by computing pairwise relative poses using convolutional neural network. In Proceedings of the IEEE International Conference on Computer Vision Workshops (ICCVW), pages 920–929, 2017. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.483, + 0.259 + ], + "angle": 0, + "content": "[27] Xiaotian Li, Shuzhe Wang, Yi Zhao, Jakob Verbeek, and Juho Kannala. Hierarchical scene coordinate classification and regression for visual localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11983-11992, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.26, + 0.482, + 0.314 + ], + "angle": 0, + "content": "[28] Gidi Littwin and Lior Wolf. Deep meta functionals for shape representation. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 1824-1833, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.482, + 0.384 + ], + "angle": 0, + "content": "[29] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 10012-10022, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.385, + 0.482, + 0.439 + ], + "angle": 0, + "content": "[30] Will Maddern, Geoffrey Pascoe, Chris Linegar, and Paul Newman. 1 year, \\(1000\\mathrm{km}\\): The oxford robotcar dataset. The International Journal of Robotics Research, 36(1):3-15, 2017. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.441, + 0.418, + 0.455 + ], + "angle": 0, + "content": "[31] Luke Melaskyriazi. Efficientnet-pytorch, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.456, + 0.483, + 0.523 + ], + "angle": 0, + "content": "[32] Iaroslav Melekhov, Juha Ylioinas, Juho Kannala, and Esa Rahtu. Image-based localization using hourglass networks. In Proceedings of the IEEE International Conference on Computer Vision Workshops (ICCVW), pages 870-877, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.525, + 0.483, + 0.593 + ], + "angle": 0, + "content": "[33] Tayyab Naseer and W. Burgard. Deep regression for monocular camera-based 6-dof global localization in outdoor environments. 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1525-1530, 2017. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.483, + 0.663 + ], + "angle": 0, + "content": "[34] Yuval Nirkin, Lior Wolf, and Tal Hassner. Hyperseg: Patchwise hypernetwork for real-time semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4061-4070, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.664, + 0.483, + 0.718 + ], + "angle": 0, + "content": "[35] N. Radwan, A. Valada, and W. Burgard. Vlocnet++: Deep multitask learning for semantic visual localization and odometry. IEEE Robotics and Automation Letters, 3(4): 4407-4414, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.483, + 0.761 + ], + "angle": 0, + "content": "[36] Prajit Ramachandran, Barret Zoph, and Quoc V Le. Searching for activation functions. arXiv preprint arXiv:1710.05941, 2017. 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.483, + 0.83 + ], + "angle": 0, + "content": "[37] P. Sarlin, C. Cadena, R. Siegwart, and M. Dymczyk. From coarse to fine: Robust hierarchical localization at large scale. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12708-12717, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[38] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4938-4947, 2020." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[39] Paul-Edouard Sarlin, Ajaykumar Unagar, Mans Larsson, Hugo Germain, Carl Toft, Viktor Larsson, Marc Pollefeys, Vincent Lepetit, Lars Hammarstrand, Fredrik Kahl, et al. Back to the feature: Learning robust camera localization from pixels to pose. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3247–3257, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.19, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[40] Torsten Sattler, Michal Havlena, Filip Radenovic, Konrad Schindler, and Marc Pollefeys. Hyperpoints and fine vocabularies for large-scale location recognition. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 2102-2110, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.26, + 0.905, + 0.314 + ], + "angle": 0, + "content": "[41] T. Sattler, B. Leibe, and L. Kobbelt. Efficient effective prioritized matching for large-scale image-based localization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(9):1744-1756, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.316, + 0.905, + 0.384 + ], + "angle": 0, + "content": "[42] T. Sattler, Q. Zhou, M. Pollefeys, and L. Leal-Taixe. Understanding the limitations of cnn-based absolute camera pose regression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3297-3307, 2019. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.385, + 0.905, + 0.426 + ], + "angle": 0, + "content": "[43] Yoli Shavit and Ron Ferens. Introduction to camera pose estimation with deep learning. In arXiv preprint arXiv:1907.05272, 2019. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.427, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[44] Yoli Shavit and Ron Ferens. Do we really need scene-specific pose encoders? In Proceedings of the International Conference on Pattern Recognition (ICPR), pages 3186-3192. IEEE, 2021. 2, 3, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.905, + 0.538 + ], + "angle": 0, + "content": "[45] Yoli Shavit, Ron Ferens, and Yoshi Keller. Learning multiscene absolute pose regression with transformers. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 2713-2722, 2021. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.905, + 0.593 + ], + "angle": 0, + "content": "[46] Yoli Shavit, Ron Ferens, and Yoshi Keller. Learning multiscene absolute pose regression with transformers. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.595, + 0.905, + 0.676 + ], + "angle": 0, + "content": "[47] Jamie Shotton, Ben Glocker, Christopher Zach, Shahram Izadi, Antonio Criminisi, and Andrew Fitzgibbon. Scene coordinate regression forests for camera relocalization in rgb images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.678, + 0.905, + 0.746 + ], + "angle": 0, + "content": "[48] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in Neural Information Processing Systems (NIPS), 33:7462-7473, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.748, + 0.905, + 0.802 + ], + "angle": 0, + "content": "[49] Xiaogang Song, Hongjuan Li, Li Liang, Weiwei Shi, Guo Xie, Xiaofeng Lu, and Xinhong Hei. Transbonet: Learning camera localization with transformer bottleneck and attention. Pattern Recognition, 146:109975, 2024. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[50] Przemysław Spurek, Artur Kasymov, Marcin Mazur, Diana Janik, Slawomir K Tadeja, J Tabor, T Trzeciński, et al. Hyperpocket: Generative point cloud completion. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 6848-6853. IEEE, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[51] Hajime Taira, Masatoshi Okutomi, Torsten Sattler, Mircea Cimpoi, Marc Pollefeys, Josef Sivic, Tomas Pajdla, and Ak" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11556" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.147 + ], + "angle": 0, + "content": "ihiko Torii. Inloc: Indoor visual localization with dense matching and view synthesis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(4):1293-1307, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.15, + 0.484, + 0.191 + ], + "angle": 0, + "content": "[52] Mingxing Tan and Quoc Le. EfficientNet: Rethinking model scaling for convolutional neural networks. pages 6105-6114, Long Beach, California, USA, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.484, + 0.248 + ], + "angle": 0, + "content": "[53] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In Proceedings of the International Conference on Machine Learning (ICML), pages 6105-6114. PMLR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.483, + 0.319 + ], + "angle": 0, + "content": "[54] Shitao Tang, Sicong Tang, Andrea Tagliasacchi, Ping Tan, and Yasutaka Furukawa. Neumap: Neural coordinate mapping by auto-transdecoder for camera localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 929-939, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.483, + 0.376 + ], + "angle": 0, + "content": "[55] Yi Tay, Zhe Zhao, Dara Bahri, Don Metzler, and Da-Cheng Juan. HyperGrid transformers: Towards a single model for multiple tasks. In Proceedings of the International Conference on Learning Representations (ICLR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.378, + 0.483, + 0.446 + ], + "angle": 0, + "content": "[56] Akihiko Torii, Relja Arandjelović, Josef Sivic, Masatoshi Okutomi, and Tomas Pajdla. 24/7 place recognition by view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1808-1817, 2015. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.449, + 0.483, + 0.517 + ], + "angle": 0, + "content": "[57] Mehmet Ozgur Turkoglu, Eric Brachmann, Konrad Schindler, Gabriel J Brostow, and Aron Monszpart. Visual camera re-localization using graph neural networks and relative pose supervision. In 2021 International Conference on 3D Vision (3DV), pages 145-155. IEEE, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.483, + 0.587 + ], + "angle": 0, + "content": "[58] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems (NIPS), pages 5998-6008, 2017. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.591, + 0.483, + 0.631 + ], + "angle": 0, + "content": "[59] Johannes Von Oswald, Christian Henning, João Sacramento, and Benjamin F Grewe. Continual learning with hypernetworks. arXiv preprint arXiv:1906.00695, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.483, + 0.701 + ], + "angle": 0, + "content": "[60] F. Walch, C. Hazirbas, L. Leal-Taixe, T. Sattler, S. Hilsenbeck, and D. Cremers. Image-based localization using lstms for structured feature correlation. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 627-637, 2017. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.483, + 0.771 + ], + "angle": 0, + "content": "[61] Bing Wang, Changhao Chen, Chris Xiaoxuan Lu, Peijun Zhao, Niki Trigoni, and Andrew Markham. Atloc: Attention guided camera localization. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10393-10401, 2020. 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.483, + 0.843 + ], + "angle": 0, + "content": "[62] Dominik Winkelbauer, Maximilian Denninger, and Rudolph Triebel. Learning to localize in new environments from synthetic training data. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 5840-5846. IEEE, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[63] J. Wu, L. Ma, and X. Hu. Delving deeper into convolutional neural networks for camera relocalization. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), pages 5644-5651, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[64] Tete Xiao, Mannat Singh, Eric Mintun, Trevor Darrell, Piotr Dollár, and Ross Girshick. Early convolutions help transformers see better. Advances in Neural Information Processing Systems (NIPS), 34:30392-30400, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[65] Meng Xu, Youchen Wang, Bin Xu, Jun Zhang, Jian Ren, Stefan Poslad, and Pengfei Xu. A critical analysis of image-based camera pose estimation techniques. arXiv preprint arXiv:2201.05816, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.907, + 0.275 + ], + "angle": 0, + "content": "[66] Fei Xue, Xin Wang, Zike Yan, Qiuyuan Wang, Junqiu Wang, and Hongbin Zha. Local supports global: Deep camera relocalization with sequence enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[67] Fei Xue, Xin Wu, Shaojun Cai, and Junqiu Wang. Learning multi-view camera relocalization with graph neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.347, + 0.905, + 0.416 + ], + "angle": 0, + "content": "[68] Qunjie Zhou, Torsten Sattler, Marc Pollefeys, and Laura Leal-Taixe. To learn or not to learn: Visual localization from essential matrices. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), pages 3319-3326. IEEE, 2020. 3, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11557" + } + ] +] \ No newline at end of file diff --git a/2025/HyperPose_ Hypernetwork-Infused Camera Pose Localization and an Extended Cambridge Landmarks Dataset/dd4d95b8-93a1-4dc9-acfe-357b62a866ca_origin.pdf b/2025/HyperPose_ Hypernetwork-Infused Camera Pose Localization and an Extended Cambridge Landmarks Dataset/dd4d95b8-93a1-4dc9-acfe-357b62a866ca_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d9400c002cac01cd93450854c09ee81d4b8fc347 --- /dev/null +++ b/2025/HyperPose_ Hypernetwork-Infused Camera Pose Localization and an Extended Cambridge Landmarks Dataset/dd4d95b8-93a1-4dc9-acfe-357b62a866ca_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:01534390682f1a25a5c5a5b3320a5c92fc0f87e7c649eb8ebf4cd56c9fcc9a50 +size 7840585 diff --git a/2025/HyperPose_ Hypernetwork-Infused Camera Pose Localization and an Extended Cambridge Landmarks Dataset/full.md b/2025/HyperPose_ Hypernetwork-Infused Camera Pose Localization and an Extended Cambridge Landmarks Dataset/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5135a32ed13584fb5ff8145577c79a10da97e1a2 --- /dev/null +++ b/2025/HyperPose_ Hypernetwork-Infused Camera Pose Localization and an Extended Cambridge Landmarks Dataset/full.md @@ -0,0 +1,272 @@ +# HyperPose: Hypernetwork-Infused Camera Pose Localization and an Extended Cambridge Landmarks Dataset + +Ron Ferens Yosi Keller + +Faculty of Engineering, Bar Ilan University, Ramat-Gan, Israel + +{ronferens, yosi.keller}@gmail.com + +# Abstract + +In this work, we propose HyperPose, which utilizes hypernetworks in absolute camera pose regressors. The inherent appearance variations in natural scenes, attributable to environmental conditions, perspective, and lighting, induce a significant domain disparity between the training and test datasets. This disparity degrades the precision of contemporary localization networks. To mitigate this, we advocate for incorporating hypernetworks into single-scene and multiscene camera pose regression models. During inference, the hypernetwork dynamically computes adaptive weights for the localization regression heads based on the particular input image, effectively narrowing the domain gap. Using indoor and outdoor datasets, we evaluate the HyperPose methodology across multiple established absolute pose regression architectures. We also introduce and share the Extended Cambridge Landmarks (ECL), a novel localization dataset, based on the Cambridge Landmarks dataset, showing it in multiple seasons with significantly varying appearance conditions. Our empirical experiments demonstrate that HyperPose yields notable performance enhancements for single- and multi-scene architectures. We have made our source code, pre-trained models1, and the ECL dataset openly available2. + +# 1. Introduction + +In many computer vision applications, such as indoor navigation, augmented reality, and autonomous driving, it is essential to determine the position and orientation of a camera using an image taken by that camera. Current camera localization techniques show various trade-offs between accuracy, runtime, and memory usage. Feature Matching (FM) and Scene Coordinate Regression (SCR) are considered the most accurate localization methods. FM-based methods [15, 27, 37-41, 51, 54] involve hierarchical localization pipelines, which commonly use a coarse-to-fine pro + +![](images/0849acdbaaba02d7326f257f5c5c056f7ffb98c9286d996bea8f57dd6249927b.jpg) +(a) Baseline architecture. + +![](images/104ac34d2fca2452683ed8f52ece7a2f507a453796ac774c7472d2a9569eafae.jpg) +(b) Baseline architecture with a hypernetwork. +Figure 1. Figure 1a illustrates a baseline APR architecture. It consists of a CNN backbone that generates a feature vector encoding the input query image, followed by regression heads that estimate the translation and orientation. Figure 1b shows the extension of this baseline architecture with a hypernetwork that outputs the weights for the regression heads (translation and orientation) during inference. + +cessing pipeline that begins with image retrieval to perform an initial localization based on images similar to the query image. The initial estimate is refined by local feature extraction and matching. To estimate the camera's 6-Degrees-Of-Freedom (6DoF) accurately, the resulting 2D-2D matches are mapped to 3D correspondences in the scene's point cloud and then used to determine the camera pose through Perspective-n-Point (PnP) and RANSAC. SCR-based methods [4-8, 47] directly regress correspondences between 2D pixels and 3D coordinates in the scene. Relative pose regression (RPR) schemes [2, 14, 17, 20, 26, 57, 62] directly estimate the relative pose between a query image and a set of reference images with well-defined ground truth poses, eliminating the need for conventional pipelines. However, an image retrieval scheme is essential to retrieve a set of nearest-neighbor images. Another approach, absolute pose + +regressors (APRs), can estimate the camera pose with a single forward pass using only the query image, resulting in significantly lower latency, although notably less accurate. Additionally, due to their low memory footprint, APRs can be deployed as a standalone application on edge devices with limited computational resources. A survey of visual camera pose localization is given by [43, 65]. + +PoseNet by Kendall et al. [25] was the first APR approach, which leveraged a convolutional backbone and a multilayer perceptron (MLP) head to regress the camera's position and orientation. This simple and computationally efficient approach, enabling execution at 5ms, led to numerous APR methods that improved accuracy by modifying the backbone and MLP architectures [11, 32, 33, 44, 49, 60, 61, 63]. In addition, alternative loss functions and optimization strategies were explored [23, 24, 43]. Indoor and outdoor environments exhibit dynamic variability in illumination, perspective, and object motion. APR models, despite their computational efficiency, often struggle to maintain accuracy under these diverse conditions and show limited generalization to novel scenes, necessitating scene-specific training. Consequently, while architecturally simple, a unified network might yield suboptimal performance when adapting to heterogeneous real-world visual inputs. Recent studies have proposed extending APR from single-scene to multiscene formulations. Blanton et al. [3] implemented a convolutional neural network (CNN) backbone to compute a global descriptor for the input image. Rather than employing a scene-specific multilayer perceptron (MLP), they optimized a set of fully connected (FC) layers, allocating one layer per scene based on a predicted scene identifier. Although introducing a framework to optimize a single model across multiple scenes, this methodology did not achieve state-of-the-art (SOTA) APR accuracy. An alternative approach was presented by [46] using dual-branch transformers [58]. Encoders focused on extracting pose-related features, while decoders transformed encoded scene identifiers into latent pose representations. The decoder outputs are leveraged for scene classification and selection of corresponding position and orientation embeddings, which are used to regress position and orientation vectors. + +Hypernetworks [22] are a deep learning architecture in which an auxiliary network, the hypernetwork, generates the weights for the primary network (main network) based on the current input and downstream task. The main network and hypernetwork are trained end-to-end using backpropagation. The hypernetwork allows for dynamic and adaptive weight modulation of the main network, leading to improved flexibility in response to changing conditions [16, 18, 34, 55]. + +In this work, we propose HyperPose, a hypernetwork-based approach for absolute camera pose regression. Extending common APR methods [11, 23, 25, 44, 45, 60], + +the hypernetwork computes weights for the regression head. This enables adaptive analysis of image content, allowing the regression weights to focus on informative features, rather than relying on a single constant pretrained set of weights. We validated the proposed method by integrating hypernetworks into single- and multiscene APR architectures. For single-scene networks, we present HyperPose with an existing state-of-the-art APR model [61] as well as a simplistic baseline APR using an off-the-shelf backbone [53] and shallow regression heads, inspired by the principles outlined in [25]. For multiscene, we extend a leading multiscene APR model [46] by integrating a hypernetwork architecture. In both instances, hypernetwork-augmented models demonstrated superior performance compared to their original counterparts when evaluated on the benchmark datasets in this study. To evaluate localization robustness, we introduce an Extended Cambridge Landmarks (ECL) dataset - an augmentation of the original benchmark [25] with algorithmically generated environmental shifts. By combining evening, winter, and summer scenes, the ECL incorporates dynamics such as lighting shifts and seasonal transitions. + +Thus, we propose the following contributions: + +- We propose an approach to enhance any existing absolute camera pose regression models by integrating hypernetwork architectures. +- We comprehensively evaluated the proposed HyperPose on both single- and multiscene pose regressors, demonstrating improved accuracy achieved by the hypernetwork-enhanced models over the original versions. +- The proposed scheme achieves new SOTA accuracy on multiscene APR benchmarks spanning diverse outdoor and indoor localization tasks. +- We introduce the Extended Cambridge Landmarks (ECL) dataset, which builds upon the foundation of the original Cambridge Landmarks dataset [25]. Our ECL data set extends existing test scenes by incorporating various appearance conditions. + +# 2. Related Work + +Camera pose estimation techniques differ by their input during inference and algorithmic attributes. + +3D-based or structure-based localization methods are considered the most accurate for camera pose estimation and have demonstrated state-of-the-art performance in leading benchmarks. These methods utilize correspondences between 2D features in an image and 3D world coordinates. [15, 27, 35, 37-41, 51, 54] introduced a two-phase approach for hierarchical pose estimation pipelines. Thus, each query image is first encoded using a CNN trained for image retrieval (IR), based on a pre-mapped image dataset. Tentative correspondences are then estimated by matching local im + +age features to 3D matches. The resulting matches are used by PnP-RANSAC to estimate the camera pose. DSAC [8] and its succeeding scheme, DSAC++ [4], utilize a CNN architecture to directly estimate 3D coordinates from 2D positions in an image and only require the query image during inference. The same as in Absolute Pose Regression (APR) methods, separate models must be trained for each scene, achieving state-of-the-art accuracy. Lately, [15] proposes a solution to mitigate this gap by introducing a scene-agnostic architecture. + +Relative Pose Regression (RPR) algorithms estimate the relative pose between an input query image and a set of reference images based on their known location. Calculating the relative pose involves estimating the relative position and orientation between the query image and the reference images [1, 6, 68]. An Image Retrieval (IR) approach is utilized to determine the closest set of neighbor images, and the relative motion is then estimated between the query image and each of the retrieved images. The estimated relative poses are used, along with the known position of the reference images, to estimate the camera's absolute pose. Similar to 3D and structured-based localization methods, RPRs involve a multistep process and require access to a database of images with labeled poses during the inference phase. + +Absolute Pose Regression (APR) directly estimates the camera pose given the input query image. APRs were first introduced by Kendall [25], using a modified version of the truncated GoogLeNet architecture, where the Softmax classification layer was replaced by a series of fully connected layers that output the pose estimate. Although less accurate than classical structure-based methods, APRs offer several advantages, such as faster inference times (milliseconds), reduced memory footprint (megabytes), and do not require feature engineering. Variations of the encoder and MLP architectures have been proposed to improve APR accuracy [11, 23, 33, 33, 44, 49]. Furthermore, modifications to the loss function have been suggested to improve the balance between orientation and position objectives [24] or to incorporate other types of information [9]. For an in-depth examination of the various architectural designs and loss functions utilized in camera pose regression, please refer to [43]. Although demonstrating faster inference times and low memory requirements, absolute pose regression (APR) models typically require dedicated perscene training to achieve competitive accuracy compared to other methods. This limitation has raised concerns regarding the value of APRs versus significantly more accurate techniques such as [4, 6, 8, 36]. To address this issue, recent approaches suggested extending APRs to a multiscene paradigm. [3] introduced Multi-Scene PoseNet (MSPN), which utilizes a CNN to regress an image feature vector. The two branches then process this feature vector. The first branch classifies the related scene using a fully connected + +layer with a softmax operator. The estimated scene index selects scene-specific weights from a database. The second branch combines these weights with the feature vector to regress the pose. The model is trained end-to-end by applying both binary cross-entropy (for classification) and camera pose losses for each scene. Inspired by recent successful applications of Transformers [12, 19, 29] to computer vision tasks, [46] suggested a shared convolutional network backbone followed by two network branches consisting of a Transformer and a regression head. The first regresses the input query image's pose, and the other its orientation. The main incentive for utilizing Transformers encoders is to focus on pose-informative features and decoders to transform encoded scene identifiers into latent pose representations. + +Hypernetworks were first presented by Ha et al. [22]. These are neural networks that have been designed to predict the weights of a primary network. The primary network's weights can be adjusted based on specific inputs, resulting in a more expressive and adaptive model. Hypernetworks have been used in various applications such as learning implicit neural representations [16], semantic segmentation [34], 3D scene representation and modeling [28, 48, 50] and continuous learning [59], to name a few. + +# 3. APRs using Hypernetworks + +HyperPose is a general approach, applicable to both single- and multi-scene scenarios, for enhancing APR schemes. It comprises two key components: the main network (APR) designed to estimate the camera's absolute position and the hypernetwork that computes the regression weights for the main network conditioned on the input query image. The main network performs a forward pass on the input image to localize the capture camera. The camera pose is represented as the tuple $< \mathbf{x},\mathbf{q}>$ , where $\mathbf{x}\in \mathbb{R}^3$ is the camera's position in world coordinates and $\mathbf{q}\in \mathbb{R}^4$ is the quaternion encoding of its $3D$ orientation. Following [13], we formulate the hypernetwork $H$ as a parametric function $\theta_h = H(\theta ,\mathbf{x})$ , where $\theta$ represents the parameters of the hypernetwork, $\mathbf{x}$ the network's input, and $\theta_h$ the parameters of the regression heads layer for the input query image $I_q$ . + +# 3.1. Network Architecture + +# 3.1.1. HyperPose for Single-Scene APRs + +We apply the proposed HyperPose-based approach to single-scene APR using a baseline APR following [25], and the state-of-the-art AtLoc APR [61]. + +Baseline APR with Hypernetwork. As depicted in Fig. 1a, the baseline model's main network consists of a CNN backbone that generates a feature vector encoding the input query image. This vector is then processed by two regression heads. The first estimates the camera position, and the other estimates the orientation. Figure 1b illustrates the integration of a hypernetwork into this architecture. The input + +![](images/4bc874b428f303788cf723c4f5595bc3997d4dff65879b6e17c5e680ebdf7ba8.jpg) +Figure 2. MS-HyperPose - The proposed multi-scene absolute pose regression architecture using a hypernetwork. The primary network employs position and orientation Transformers, in a dual-branch architecture, to extract activation maps from the underlying convolutional backbone. The hypernetwork generates the weights for the regression head in the primary network, using the input query image. These adaptive weights, combined with the latent vectors generated by the Transformers estimate the camera pose, consisting of the spatial position (x) and angular orientation (q). + +to the hypernetwork is the feature vector generated by the CNN backbone of the main network. This vector is passed through a single fully-connected (FC) layer followed by a Swish activation [36] function. The processed vector serves as an input to a multilayer perceptron (MLP) that outputs the weights for a single layer in the regression heads of the main network. The camera pose is then estimated by the position and orientation regression heads of the main network. + +AtLoc with Hypernetwork. [61] presents Attention-Guided Camera Localization (AtLoc), an absolute pose regression model based on self-attention. In this architecture, a visual encoder encodes an image into a latent vector. Conditioned on these extracted features, an attention module computes self-attention maps to reweight the representation into a new feature space. A pose regressor then maps the transformed features after the attention operators to estimate the camera pose. We incorporate a hypernetwork into the AtLoc, the same as for the baseline APR model. In this case, the input to the hypernetwork is the reweighted representation feature vector computed by the visual encoder. + +# 3.1.2. HyperPose for Multi-Scene APRs + +The suggested multi-scene model, named MS-HyperPose is an extension of the architecture introduced in [45]. As depicted in Fig. 2, the input query image $I_{q}$ is processed by a + +convolutional backbone that computes intermediate activation maps from two different layers. These activation maps are then utilized as inputs to the main network and the hypernetwork. Within the main network, the positional and orientational Transformers process the activation maps and integrate sequential representations into single latent vectors. After converting the activation maps into representation vectors, these vectors, along with the latent vectors are passed through multilayer perceptron (MLP) layers in the hypernetwork to generate the weights $\theta_h$ for the translation and orientation regression heads. Using the weights $\theta_h$ produced by the hypernetwork $H$ and the regression heads in the main network, the position, and orientation of the input query image are estimated. + +The Main Network consists of a convolutional backbone and two separate branches to regress the camera's position and orientation. Each branch includes its particular Transformer and a multilayer perceptron (MLP) head. Studies have demonstrated that substituting the patchify stem, utilized in the early visual processing of a Vision Transformer (ViT), with a convolutional stem enhances optimization stability and elevates performance to a higher level [64]. Therefore, given an input query image $I_{q} \in \mathbb{R}^{H \times W \times C}$ , a shared convolutional backbone generates activation maps. Two different activation maps are selected + +from the CNN based on the particular learned task (Fig. 2a). The activation maps, represented as $M \in \mathbb{R}^{h \times w \times C^{Fn}}$ , are processed and transformed into $\tilde{M} \in \mathbb{R}^{h \times w \times C^{Out}}$ using a $1 \times 1$ convolutional layer followed by flattening (Fig. 2b). + +Following [45], we use the Transformer architecture from [12], modifying the standard Encoder and Decoder to add positional encoding at each attention layer. Each Transformer has $L$ identical blocks, each with a multi-head attention (MHA) mechanism and a two-layer MLP with Swish activation [36]. Layer Normalization (LN) is applied before MHA and MLP processing, as in [58], with outputs merged via residual connections and dropout. To preserve spatial information, each activation map position receives a learned encoding. The processed sequence, serving as input to the Transformer is given by (Fig. 2c): + +$$ +E ^ {I n} = \left[ T, \tilde {M} \right] \in \mathbb {R} ^ {(h w) \times C ^ {O u t}}, T \in \mathbb {R} ^ {C _ {M} ^ {O u t}}. \tag {1} +$$ + +where $T$ is the positional encoding of the activation map. + +Given a dataset with $N$ scenes, the Transformer Decoders generate embeddings $E^{Out} = \in \mathbb{R}^{N\times C^{Out}}$ (Fig. 2d). For a query image, only one position matches its scene. The combined Transformer outputs (Fig. 2e) pass through a fully connected layer and Log SoftMax, selecting the highest-probability vectors (Fig. 2f). + +The selected vectors $E_{x_i}^{Out}$ and $E_{q_i}^{Out}$ are processed through an MLP. Finally, the estimated camera position and orientation are obtained by summing the outputs of the main network and hypernetwork, using weights predicted by the hypernetwork (Fig. 2h and 2i). + +The Hypernetworks consist of MLPs, each regressing the weights for a corresponding regression head (position or orientation) in the main network. Their input is the sum of task-specific encoded activation maps $\tilde{M}_x$ and $\tilde{M}_q$ , along with global latent vectors $E_{x_i}^{Out}$ and $E_{q_i}^{Out}$ (Fig. 2g). To ensure size compatibility, the encoded activation map passes through a fully connected layer before summation. This vector is then processed by MLPs with Swish activation, generating inputs for another MLP set, each estimating weights for a specific layer in the main network's regression heads: + +$$ +\theta_ {h} ^ {O u t _ {j}} = H \left(\theta_ {h} ^ {O u t _ {j}}, E _ {i} ^ {O u t}\right) \in \mathbb {R} ^ {C _ {E _ {i}} ^ {O u t} \times C _ {H} ^ {O u t _ {j}}}, \tag {2} +$$ + +where $j$ is the index of the regression head layer. + +# 3.2. Optimization Criteria + +The localization error is calculated by comparing the translation position and orientation of the ground truth pose $\mathbf{p}_{gt} = < \mathbf{x}_{gt},\mathbf{q}_{gt}>$ , where $\mathbf{x}\in \mathbb{R}^3$ represents the position of the camera in the world and $\mathbf{q}\in \mathbb{R}^4$ denotes its orientation encoded as a quaternion, with the estimated pose + +$\mathbf{p}_{est} = < \mathbf{x}_{est},\mathbf{q}_{est} >$ . The translation error, represented by $L_{x}$ , is calculated by the Euclidean distance between the ground truth and the estimated camera positions. The orientation error, $L_{q}$ , is defined as the Euclidean distance between the ground truth and the estimated unit quaternions [43]: + +$$ +L _ {x} = \left\| \mathbf {x} _ {0} - \mathbf {x} \right\| _ {2}, \tag {3} +$$ + +$$ +L _ {q} = \left\| \mathbf {q} _ {0} - \frac {\mathbf {q}}{| | \mathbf {q} | |} \right\| _ {2} \tag {4} +$$ + +where $\mathbf{q}$ is a normalized norm quaternion to ensure it is a valid orientation encoding. + +As a multi-regression optimization problem, we adopt the method outlined in [24] by using different weighting schemes, based on the uncertainty of each task, to balance the individual loss functions defined for the separate objectives. The model optimizes the aggregated loss function, a combination of the weighted individual losses. + +$$ +L _ {p} = L _ {x} \exp (- s _ {x}) + s _ {x} + L _ {q} \exp (- s _ {q}) + s _ {q}, \tag {5} +$$ + +where $S_{x}$ and $S_{q}$ are the learned parameters. As a further optimization, we adopt the approach detailed in [44] and follow a three-step training process. Firstly, the entire network is trained using the aggregated loss in Eq. 5. In the next two stages, each MLP head undergoes separate fin-tuning with its corresponding loss function. + +# 3.3. Implementation Details + +The shared backbone uses EfficientNet-B0 [31, 52]. We used the activation maps of two endpoints (reduction levels) as input for our position and orientation branches: $m_{rdct4} \in \mathbb{R}^{14 \times 14 \times 112}$ and $m_{rdct3} \in \mathbb{R}^{28 \times 28 \times 40}$ , respectively. Following [45], we applied linear projections on each activation map to a common depth dimension of $C_M^{Out} = 256$ and learned the positional encoding of the same depth. The regression heads in the main network consist of two-layer MLPs with input and output layers. The input layer dimension is 1024 for both the position and orientation, while the output dimensions are 3 and 4, respectively. + +Single-Scene APR. For both the baseline APR and the HyperPose extension of the AtLoc [61] model, the hypernetwork for the position branch generates weights for an MLP with a dimension of 256, while the orientation hypernetwork produces weights of 512 dimensions. + +Multi-Scene APR. In contrast to [45], we incorporated four-layer (instead of six) Transformers in each branch (position and orientation) within the main network architecture. Each block consists of an MHA layer with four heads, followed by an MLP that preserves the input dimension. A dropout rate of $p = 0.1$ was applied to each MLP layer. The output dimension of each Transformer in the main network + +Table 1. Comparative analysis of APR architectures with and without the purported hypernetwork addition using the Cambridge Landmarks dataset. We report the median position/orientation error in meters/degrees. The best performance is marked in bold. (*) The authors of AtLoc did not report results on the Cambridge dataset, and we trained the AtLoc model on this dataset. + +
MethodK. CollegeOld HospitalShop FacadeSt. MaryAvg.
AtLoc* [61]1.53,3.212.17,3.990.93,4.612.14,5.761.69,4.39
AtLoc w/ hypernet0.96,3.432.04,3.610.88,4.221.72,5.441.40,4.18
Baseline APR0.89,2.291.49,3.320.74,4.791.40,4.951.13,3.84
Baseline APR w/ hypernet0.77,2.071.47,3.290.72,4.011.37,4.901.08,3.57
+ +was $C_M^{Out} = 256$ . The position and orientation branches of the hypernetwork generate the weights of three regression layers: input, hidden, and output. The number of weights in the position and orientation branches differ. The hypernetwork's position branch produces the weights for an MLP with dimensions 256, 128, and 3, while the orientation branch generates weights for an MLP with dimensions 512, 512, and 4. The weights generated by the hypernetwork are then put into the regression heads of the main network to estimate the position and orientation of the input query image. The input and hidden layers of the regression head utilize the Swish activation function. + +# 4. Extended Cambridge Landmarks Dataset + +![](images/dc3759e8b063cfa79df346d6a900aa23b12aa2b490f3d8304796f2c10cf416f9.jpg) +(a) Original + +![](images/323c1720ebdf2614c798d4faabf7f1b8418682de06d5453ea43ba4fc2cde6b98.jpg) +(b)Evening + +![](images/76acff78b28455974bf3c466a8d1bb63973e185e0580a784b91d7c118780f857.jpg) +(c) Summer + +![](images/c7466b1bc9048e773f40d442bc0bb3668461139afb0373c26fc7d0ae149a9d07.jpg) +(d) Winter +Figure 3. A sample from the King's College Scene in the proposed Extended Cambridge Landmarks (ECL) Dataset + +To further assess the limitations of the HyperPose approach, we constructed an augmented version of the Cambridge Landmarks [25] benchmark called the Extended Cambridge Landmarks (ECL) dataset. Leveraging the original scenes, the ECL dataset introduces new localization challenges due to variations in appearance by creating synthetic evening, winter, and summer variants for each location. These scene permutations simulate real-world environmental variations by introducing shifts in lighting, weather patterns, foliage density, and more. Analyzing performance across ECL variations and their original Cambridge counterparts allows us to evaluate the robustness of + +models to real-world variations. These augmentations were conducted by utilizing the novel InstructPix2Pix work [10]. Visual samples of the ECL dataset can be found in the supplementary materials. Figure 3 shows sample from the King's College scene in the extended dataset. Additional samples from each scene in the ECL dataset can be found in the supplementary material. Quantitative pose regression analysis between these more challenging conditions and unmodified sequences helps scientifically assess model adaptability. We hypothesize that exposing limitations unseen in conventional test sets will motivate the development of more invariant solutions. As controllable expansions of established benchmarks, programmatically generated datasets can thus guide the development of robust models for real-world use. + +# 5. Experimental Results + +We conducted an experimental evaluation using the 7 Scenes [21] and Cambridge Landmarks [25] datasets for evaluating absolute pose regression. The Cambridge Landmarks dataset represents an outdoor urban environment ranging in size from 900 to 5500 square meters. We present results for four of its six scenes. We excluded the remaining two scenes due to limited coverage in prior work. To compare robustness, we also report the results on the Extended Cambridge Landmarks dataset (ECL) (Section 1 in the supplementary materials). The 7 Scenes dataset consists of seven small-scale indoor scenes with spatial extents ranging from approximately 1 to 10 square meters. These datasets present various localization challenges, including differences in scale, indoor versus outdoor settings, repetitive elements, textureless features, significant viewpoint changes, and variations in trajectory between training and test sets. To facilitate a broader analysis of our proposed method, we conduct an additional evaluation on the Oxford RobotCar Dataset [30]. Such dynamic environmental factors assess HyperPose's ability to improve localization accuracy amid diverse, challenging visual contexts. Our experimental procedures adhere to the same training and evaluation protocols as in [9, 61]. Additional information concerning these sequences, and a comprehensive description of the training methodology and hyperparameter selection for the models, are provided in the supplementary material. + +We evaluate the impact of integrating hypernetworks in + +Table 2. Comparison to single-scene state-of-the-art methods - Cambridge Landmarks: We report the median position/orientation error in meters/degrees. The most effective single and multi-scene APRs are marked in bold. + +
K. CollegeHospitalShop FacadeSt. MaryAvg.
Seq.DSAC* [6]0.18,0.3°0.21,0.4°0.05,0.3°0.15,0.5°0.15,0.4°
MapNet [9]1.08,1.9°1.94,3.9°1.49,4.2°2.00,4.5°1.63,3.6°
GL-Net [67]0.59,0.7°1.88,2.8°0.50,2.9°1.90,3.3°1.22,2.4°
IRVLAD [56]2.80,5.7°4.01,7.1°1.11,7.6°2.31,8.0°2.56,7.1°
VLAD+Inter [42]1.48,4.5°2.68,4.6°0.90,4.3°1.62,6.1°1.67,4.9°
RPREssNet [68]0.76,1.9°1.39,2.8°0.84,4.3°1.32,4.7°1.08,3.4°
NC-EssNet [68]0.61,1.6°0.95,2.7°0.7,3.4°1.12,3.6°0.85,2.8°
RelocGNN[21]0.48,1.0°1.14,2.5°0.48,2.5°1.52,3.2°0.91,2.3°
APRPoseNet [25]1.92,5.40°2.31,5.38°1.46,8.08°2.65,8.48°2.08,6.83°
BayesianPN [23]1.74,4.06°2.57,5.14°1.25,7.54°2.11,8.38°1.91,6.28°
LSTM-PN [60]0.99,3.65°1.51,4.29°1.18,7.44°1.52,6.68°1.30,5.57°
SVS-Pose [33]1.06,2.81°1.50,4.03°0.63,5.73°2.11,8.11°1.32,5.17°
GPoseNet [11]1.61,2.29°2.62,3.89°1.14,5.73°2.93,6.46°2.07,4.59°
PoseNetLearn [24]0.99,1.06°2.17,2.94°1.05,3.97°1.49,3.43°1.42,2.85°
GeoPoseNet [24]0.88,1.04°3.20,3.29°0.88,3.78°1.57,3.32°1.63,2.86°
IRPNet [44]1.18,2.19°1.87,3.38°0.72,3.47°1.87,4.94°1.41,3.50°
Baseline APR w/ hypernet0.61,1.84°1.44,3.03°0.70,3.62°1.37,4.85°1.03,2.67°
MS-APRMSPN [3]1.73,3.65°2.55,4.05°2.92,7.49°2.67,6.18°2.47,5.34°
MS-Trans[46]0.83,1.47°1.81,2.39°0.86,3.07°1.62,3.99°1.28,2.73°
MS-HyperPose0.78,1.18°1.84,2.11°0.83,3.13°1.61,3.22°1.27,2.41°
+ +single-scene APRs. To this end, we compare the accuracy of two model variations, with and without hypernetworks, and report median position and orientation errors. Table 1 presents results on the Cambridge dataset. The hypernetwork improves the accuracy of the proposed baseline APR and AtLoc [61] models. The baseline APR model size is $26.7\mathrm{MB}$ with an inference time of $9.32\mathrm{ms}$ , increasing to $40.9\mathrm{MB}$ and $9.45\mathrm{ms}$ with a hypernetwork. The original published AtLoc model is $97.9\mathrm{MB}$ with an inference time of $5.14\mathrm{ms}$ . The hypernetwork-enhanced AtLoc version has a model size of $116.8\mathrm{MB}$ and an inference time of $5.56\mathrm{ms}$ . Note that the reported errors refer to the models' performance after the first stage of model training, without applying the two additional optimization steps described in the training details (See supplementary materials). Furthermore, Tables 2 and 3 compare the median position error, and the orientation error, of the proposed baseline APR with hypernetwork architecture, with the results of recent SOTA localization algorithms, using the Cambridge Landmarks and 7Scenes datasets, respectively. For the indoor and outdoor datasets, in addition to the results of leading APRs, we report the median position and orientation errors of structure-based, sequence-based, IR, and RPR methods. Although DSAC* [6] achieves SOTA accuracy, compared to all other APR methods, HyperPose obtains the most accurate average of $1.03m$ , $2.67^{\circ}$ on the Cambridge outdoor dataset, and $0.19m$ , $7.48^{\circ}$ on the 7Scenes indoor dataset. These local- + +ization errors show competitive performance compared to leading RPR techniques, with the hypernet-enhanced baseline ranking third and second/fourth (7Scenes/Cambridge) in positional and rotational accuracy, respectively. Next, we compare the proposed MS-HyperPose architecture with other multiscene APR methods. Tables 2 and 3 present median position and orientation errors, respectively, on the Cambridge Landmarks and 7 Scenes datasets. On both indoor and outdoor benchmarks, our MS-HyperPose approach achieves the most accurate average of $1.27\mathrm{m}$ , $2.41^{\circ}$ on Cambridge and $0.18\mathrm{m}$ , $7.12^{\circ}$ on 7Scenes compared to other single- and multi-scene APRs. Additionally, MS-HyperPose is the most accurate for both positional and rotational errors in 6 out of 8 parameters on Cambridge and 6 out of 14 on 7 Scenes. The runtime of MS-HyperPose is $30.93\mathrm{ms}$ , similar to the reference architecture of [46]. Its model size is 571MB, notably larger than the 71.1MB of the original. We discuss storage considerations in the Limitation and Future Work section of the supplementary materials. + +Table 4 reports the mean and median position and orientation errors of AtLoc [61] compared to our hypernetwork-enhanced variant using the Oxford RobotCar Dataset [30]. We also compare the enhanced model with [9], the SOTA APR for this dataset using image sequences. Visualizations of the localization results can be found in the supplementary materials. + +Table 3. Comparison to single-scene SOTA methods - 7Scenes: We report the median position/orientation error in meters/degrees. The most effective single and multi-scene APRs are distinguished in bold. + +
ChessFireHeadsOfficePumpkinKitchenStairsAvg.
SCRDSAC* [6]0.02,1.11°0.02,1.24°0.01,1.82°0.03,1.18°0.04,1.41°0.03,1.70°0.04,1.42°0.03,1.41°
Marepo[15]0.02,1.24°0.02,1.39°0.02,2.03°0.03,1.26°0.04,1.48°0.04,1.71°0.06,1.67°0.03,1.54°
Seq.LsG [66]0.09,3.3°0.26,10.9°0.17,12.7°0.18,5.5°0.20,3.7°0.23,4.9°0.23,11.3°0.19,7.74°
MapNet[9]0.08,3.3°0.27,11.7°0.18,13.3°0.17,5.2°0.22,4.0°0.23,4.9°0.30,12.1°0.21,7.78°
GL-Net[67]0.08,2.8°0.26,8.9°0.17,11.4°0.18,13.3°0.15,2.8°0.25,4.5°0.23,8.8°0.19,7.50°
IRVLAD [56]0.21,12.5°0.33,13.8°0.15,14.9°0.28,11.2°0.31,11.2°0.30,11.3°0.25,12.3°0.26,12.46°
VLAD+Inter[42]0.18,10.0°0.33,12.4°0.14,14.3°0.25,10.1°0.26,9.4°0.27,11.1°0.24,14.7°0.24,11.71°
RPRNN-Net [26]0.13,6.5°0.26,12.7°0.14,12.3°0.21,7.4°0.24,6.4°0.24,8.0°0.27,11.8°0.21,9.30°
RelocNet[2]0.12,4.1°0.26,10.4°0.14,10.5°0.18,5.3°0.26,4.2°0.23,5.1°0.28,7.5°0.21,6.73°
EssNet [68]0.13,5.1°0.27,10.1°0.15,9.9°0.21,6.9°0.22,6.1°0.23,6.9°0.32,11.2°0.22,8.03°
NC-EssNet [68]0.12,5.6°0.26,9.6°0.14,10.7°0.20,6.7°0.22,5.7°0.22,6.3°0.31,7.9°0.21,7.50°
RelocGNN[21]0.08,2.7°0.21,7.5°0.13,8.70°0.15,4.1°0.15,3.5°0.19,3.7°0.22,6.5°0.16,5.24°
CamNet [17]0.04,1.7°0.03,1.7°0.05,2.00°0.04,1.60°0.04,1.6°0.04,1.6°0.04,1.5°0.04,1.70°
APRPoseNet [25]0.32,8.12°0.47,14.4°0.29,12.0°0.48,7.68°0.47,8.42°0.59,8.64°0.47,13.8°0.44,10.4°
BayesianPN [23]0.37,7.24°0.43,13.7°0.31,12.0°0.48,8.04°0.61,7.08°0.58,7.54°0.48,13.1°0.47,9.81°
LSTM-PN [60]0.24,5.77°0.34,11.9°0.21,13.7°0.30,8.08°0.33,7.00°0.37,8.83°0.40,13.7°0.31,9.85°
GPoseNet [11]0.20,7.11°0.38,12.3°0.21,13.8°0.28,8.83°0.37,6.94°0.35,8.15°0.37,12.5°0.31,9.95°
PoseNetLearn[24]0.14,4.50°0.27,11.8°0.18,12.1°0.20,5.77°0.25,4.82°0.24,5.52°0.37,10.6°0.24,7.87°
GeoPoseNet[24]0.13,4.48°0.27,11.3°0.17,13.0°0.19,5.55°0.26,4.75°0.23,5.35°0.35,12.4°0.23,8.12°
IRPNet[44]0.13,5.64°0.25,9.67°0.15,13.1°0.24,6.33°0.22,5.78°0.30,7.29°0.34,11.6°0.23,8.49°
AtLoc[61]0.10,4.07°0.25,11.4°0.16,11.8°0.17,5.34°0.21,4.37°0.23,5.42°0.26,10.5°0.20,7.56°
TransBoNet[49]0.11,4.48°0.25,12.5°0.18,14.0°0.20,5.08°0.19,4.77°0.17,5.35°0.30,13.0°0.20,8.45°
Baseline APR w/ hypernet0.09,4.47°0.28,9.44°0.15,11.4°0.16,6.29°0.19,4.62°0.18,6.95°0.28,9.15°0.19,7.48°
MS-APRMSPN[3]0.09,4.76°0.29,10.5°0.16,13.1°0.16,6.80°0.19,5.50°0.21,6.61°0.31,11.6°0.20,7.56°
MS-Trans[46]0.11,4.66°0.24,9.60°0.14,12.2°0.17,5.66°0.18,4.44°0.17,5.94°0.26,8.45°0.18,7.28°
MS-HyperPose0.11,4.34°0.23,9.79°0.13,10.7°0.17,6.05°0.16,5.24°0.17,6.86°0.27,6.00°0.18,7.00°
+ +Table 4. Comparative analysis of different APR architectures with and without the proposed hypernetwork to the Oxford RobotCar dataset: We report the mean and median position/orientation error in meters/degrees. + +
MapNet[9]AtLoc[61]AtLoc w/ hypernet
SequenceMeanMedianMeanMedianMeanMedian
LOOP18.76, 3.465.79, 1.548.61, 4.585.68, 2.238.42, 3.304.91, 1.33
LOOP29.84, 3.964.91, 1.678.61, 4.585.68, 2.238.10, 3.044.73, 1.27
FULL141.4, 12.517.94, 6.6829.6, 12.411.10, 5.2829.58, 6.777.86, 1.38
FULL259.3, 14.820.04, 6.3948.2, 11.112.2, 4.6345.96, 9.608.38, 1.45
+ +We also assess the performance of two configurations of the baseline APR model, with and without hypernetworks on the Extended Cambridge dataset (ECL). Across all scenes and variations, the hypernetwork-enhanced model outperforms the baseline, demonstrating better adaptation to the seasonal changes and diverse lighting conditions inherent in the scenes. The detailed analysis elucidating the robustness and resilience of HyperPose to scene changes is in the supplementary material. + +# 6. Conclusions + +In this work, we proposed a hypernetwork-based approach to improve APR models. Our experiments show that hypernetworks improve accuracy for single-scene APR models. We also present MS-HyperPose, a novel multiscene camera pose regression model enhanced with a hypernetwork, which achieves new SOTA accuracy in indoor and outdoor localization datasets. We introduce the Extended Cambridge Landmarks (ECL) dataset to facilitate research on robust localization methods. + +# References + +[1] Eduardo Arnold, Jamie Wynn, Sara Vicente, Guillermo Garcia-Hernando, Aron Monszpart, Victor Prisacariu, Daniyar Turmukhambetov, and Eric Brachmann. Map-free visual relocalization: Metric pose relative to a single image. In Proceedings of the European Conference on Computer Vision (ECCV), pages 690-708. Springer, 2022. 3 +[2] Vassileios Balntas, Shuda Li, and Victor Prisacariu. Relocnet: Continuous metric learning relocalisation using neural nets. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 1, 8 +[3] Hunter Blanton, Connor Greenwell, Scott Workman, and Nathan Jacobs. Extending absolute pose regression to multiple scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 38-39, 2020. 2, 3, 7, 8 +[4] E. Brachmann and C. Rother. Learning less is more - 6D camera localization via 3D surface regression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4654-4662, 2018. 1, 3 +[5] Eric Brachmann and Carsten Rother. Expert sample consensus applied to camera re-localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7525-7534, 2019. +[6] E. Brachmann and C. Rother. Visual Camera Re-Localization From RGB and RGB-D Images Using DSAC. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(09):5847-5865, 2022. 3, 7, 8 +[7] Eric Brachmann, Frank Michel, Alexander Krull, Michael Ying Yang, Stefan Gumhold, and Carsten Rother. Uncertainty-driven 6D pose estimation of objects and scenes from a single rgb image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3364-3372, 2016. +[8] E. Brachmann, A. Krull, S. Nowozin, J. Shotton, F. Michel, S. Gumhold, and C. Rother. DSAC - differentiable RANSAC for camera localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2492-2500, Los Alamitos, CA, USA, 2017. 1, 3 +[9] Samarth Brahmbhatt, Jinwei Gu, Kihwan Kim, James Hays, and Jan Kautz. Geometry-aware learning of maps for camera localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3, 6, 7, 8 +[10] T. Brooks, A. Holynski, and A. A. Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18392-18402, Los Alamitos, CA, USA, 2023. IEEE Computer Society. 6 +[11] Ming Cai, Chunhua Shen, and Ian Reid. A hybrid probabilistic model for camera relocalization. In Proceedings of the British Machine Vision Conference, page 238, 2018. 2, 3, 7, 8 +[12] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Proceedings + +of the European Conference on Computer Vision (ECCV), pages 213-229, Cham, 2020. 3, 5 +[13] Yoav Chai, Raja Giryes, and Lior Wolf. Supervised and unsupervised learning of parameterized color enhancement. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 992-1000, 2020. 3 +[14] Kefan Chen, Noah Snavely, and Ameesh Makadia. Wide-baseline relative camera pose estimation with directional learning. 2021 IEEE. In CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3257-3267, 2021. 1 +[15] Shuai Chen, Tommaso Cavallari, Victor Adrian Prisacariu, and Eric Brachmann. Map-relative pose regression for visual re-localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20665-20674, 2024. 1, 2, 3, 8 +[16] Yinbo Chen and Xiaolong Wang. Transformers as meta-learners for implicit neural representations. In Proceedings of the European Conference on Computer Vision (ECCV), pages 170–187. Springer, 2022. 2, 3 +[17] Mingyu Ding, Zhe Wang, Jiankai Sun, Jianping Shi, and Ping Luo. CamNet: Coarse-to-fine retrieval for camera relocalization. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2019. 1, 8 +[18] Siyan Dong, Qingnan Fan, He Wang, Ji Shi, Li Yi, Thomas Funkhouser, Baoquan Chen, and Leonidas J Guibas. Robust neural routing through space partitions for camera relocalization in dynamic indoor environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8544-8554, 2021. 2 +[19] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In Proceedings of the International Conference on Learning Representations (ICLR), 2021. 3 +[20] Sovann En, Alexis Lecheryv, and Frédéric Jurie. Rpnet: An end-to-end network for relative camera pose estimation. In Proceedings of the European Conference on Computer Vision (ECCV) Workshops, pages 0–0, 2018. 1 +[21] B. Glocker, S. Izadi, J. Shotton, and A. Criminisi. Realtime rgb-d camera relocalization. In 2013 IEEE International Symposium on Mixed and Augmented Reality (ISMAR), pages 173-179, 2013. 6, 7, 8 +[22] David Ha, Andrew Dai, and Quoc V Le. Hypernetworks. arXiv preprint arXiv:1609.09106, 2016. 2, 3 +[23] Alex Kendall and Roberto Cipolla. Modelling uncertainty in deep learning for camera relocalization. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), pages 4762-4769, 2016. 2, 3, 7, 8 +[24] A. Kendall and R. Cipolla. Geometric loss functions for camera pose regression with deep learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6555-6564, 2017. 2, 3, 5, 7, 8 +[25] A. Kendall, M. Grimes, and R. Cipolla. Posenet: A convolutional network for real-time 6-dof camera relocalization. In + +Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 2938-2946, 2015. 2, 3, 6, 7, 8 +[26] Z. Laskar, I. Melekhov, S. Kalia, and J. Kannala. Camera relocalization by computing pairwise relative poses using convolutional neural network. In Proceedings of the IEEE International Conference on Computer Vision Workshops (ICCVW), pages 920–929, 2017. 1, 8 +[27] Xiaotian Li, Shuzhe Wang, Yi Zhao, Jakob Verbeek, and Juho Kannala. Hierarchical scene coordinate classification and regression for visual localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11983-11992, 2020. 1, 2 +[28] Gidi Littwin and Lior Wolf. Deep meta functionals for shape representation. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 1824-1833, 2019. 3 +[29] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 10012-10022, 2021. 3 +[30] Will Maddern, Geoffrey Pascoe, Chris Linegar, and Paul Newman. 1 year, $1000\mathrm{km}$ : The oxford robotcar dataset. The International Journal of Robotics Research, 36(1):3-15, 2017. 6, 7 +[31] Luke Melaskyriazi. Efficientnet-pytorch, 2019. 5 +[32] Iaroslav Melekhov, Juha Ylioinas, Juho Kannala, and Esa Rahtu. Image-based localization using hourglass networks. In Proceedings of the IEEE International Conference on Computer Vision Workshops (ICCVW), pages 870-877, 2017. 2 +[33] Tayyab Naseer and W. Burgard. Deep regression for monocular camera-based 6-dof global localization in outdoor environments. 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1525-1530, 2017. 2, 3, 7 +[34] Yuval Nirkin, Lior Wolf, and Tal Hassner. Hyperseg: Patchwise hypernetwork for real-time semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4061-4070, 2021. 2, 3 +[35] N. Radwan, A. Valada, and W. Burgard. Vlocnet++: Deep multitask learning for semantic visual localization and odometry. IEEE Robotics and Automation Letters, 3(4): 4407-4414, 2018. 2 +[36] Prajit Ramachandran, Barret Zoph, and Quoc V Le. Searching for activation functions. arXiv preprint arXiv:1710.05941, 2017. 3, 4, 5 +[37] P. Sarlin, C. Cadena, R. Siegwart, and M. Dymczyk. From coarse to fine: Robust hierarchical localization at large scale. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12708-12717, 2019. 1, 2 +[38] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4938-4947, 2020. + +[39] Paul-Edouard Sarlin, Ajaykumar Unagar, Mans Larsson, Hugo Germain, Carl Toft, Viktor Larsson, Marc Pollefeys, Vincent Lepetit, Lars Hammarstrand, Fredrik Kahl, et al. Back to the feature: Learning robust camera localization from pixels to pose. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3247–3257, 2021. +[40] Torsten Sattler, Michal Havlena, Filip Radenovic, Konrad Schindler, and Marc Pollefeys. Hyperpoints and fine vocabularies for large-scale location recognition. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 2102-2110, 2015. +[41] T. Sattler, B. Leibe, and L. Kobbelt. Efficient effective prioritized matching for large-scale image-based localization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(9):1744-1756, 2017. 1, 2 +[42] T. Sattler, Q. Zhou, M. Pollefeys, and L. Leal-Taixe. Understanding the limitations of cnn-based absolute camera pose regression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3297-3307, 2019. 7, 8 +[43] Yoli Shavit and Ron Ferens. Introduction to camera pose estimation with deep learning. In arXiv preprint arXiv:1907.05272, 2019. 2, 3, 5 +[44] Yoli Shavit and Ron Ferens. Do we really need scene-specific pose encoders? In Proceedings of the International Conference on Pattern Recognition (ICPR), pages 3186-3192. IEEE, 2021. 2, 3, 5, 7, 8 +[45] Yoli Shavit, Ron Ferens, and Yoshi Keller. Learning multiscene absolute pose regression with transformers. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 2713-2722, 2021. 2, 4, 5 +[46] Yoli Shavit, Ron Ferens, and Yoshi Keller. Learning multiscene absolute pose regression with transformers. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. 2, 3, 7, 8 +[47] Jamie Shotton, Ben Glocker, Christopher Zach, Shahram Izadi, Antonio Criminisi, and Andrew Fitzgibbon. Scene coordinate regression forests for camera relocalization in rgb images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2013. 1 +[48] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in Neural Information Processing Systems (NIPS), 33:7462-7473, 2020. 3 +[49] Xiaogang Song, Hongjuan Li, Li Liang, Weiwei Shi, Guo Xie, Xiaofeng Lu, and Xinhong Hei. Transbonet: Learning camera localization with transformer bottleneck and attention. Pattern Recognition, 146:109975, 2024. 2, 3, 8 +[50] Przemysław Spurek, Artur Kasymov, Marcin Mazur, Diana Janik, Slawomir K Tadeja, J Tabor, T Trzeciński, et al. Hyperpocket: Generative point cloud completion. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 6848-6853. IEEE, 2022. 3 +[51] Hajime Taira, Masatoshi Okutomi, Torsten Sattler, Mircea Cimpoi, Marc Pollefeys, Josef Sivic, Tomas Pajdla, and Ak + +ihiko Torii. Inloc: Indoor visual localization with dense matching and view synthesis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(4):1293-1307, 2021. 1, 2 +[52] Mingxing Tan and Quoc Le. EfficientNet: Rethinking model scaling for convolutional neural networks. pages 6105-6114, Long Beach, California, USA, 2019. 5 +[53] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In Proceedings of the International Conference on Machine Learning (ICML), pages 6105-6114. PMLR, 2019. 2 +[54] Shitao Tang, Sicong Tang, Andrea Tagliasacchi, Ping Tan, and Yasutaka Furukawa. Neumap: Neural coordinate mapping by auto-transdecoder for camera localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 929-939, 2023. 1, 2 +[55] Yi Tay, Zhe Zhao, Dara Bahri, Don Metzler, and Da-Cheng Juan. HyperGrid transformers: Towards a single model for multiple tasks. In Proceedings of the International Conference on Learning Representations (ICLR), 2021. 2 +[56] Akihiko Torii, Relja Arandjelović, Josef Sivic, Masatoshi Okutomi, and Tomas Pajdla. 24/7 place recognition by view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1808-1817, 2015. 7, 8 +[57] Mehmet Ozgur Turkoglu, Eric Brachmann, Konrad Schindler, Gabriel J Brostow, and Aron Monszpart. Visual camera re-localization using graph neural networks and relative pose supervision. In 2021 International Conference on 3D Vision (3DV), pages 145-155. IEEE, 2021. 1 +[58] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems (NIPS), pages 5998-6008, 2017. 2, 5 +[59] Johannes Von Oswald, Christian Henning, João Sacramento, and Benjamin F Grewe. Continual learning with hypernetworks. arXiv preprint arXiv:1906.00695, 2019. 3 +[60] F. Walch, C. Hazirbas, L. Leal-Taixe, T. Sattler, S. Hilsenbeck, and D. Cremers. Image-based localization using lstms for structured feature correlation. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 627-637, 2017. 2, 7, 8 +[61] Bing Wang, Changhao Chen, Chris Xiaoxuan Lu, Peijun Zhao, Niki Trigoni, and Andrew Markham. Atloc: Attention guided camera localization. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10393-10401, 2020. 2, 3, 4, 5, 6, 7, 8 +[62] Dominik Winkelbauer, Maximilian Denninger, and Rudolph Triebel. Learning to localize in new environments from synthetic training data. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 5840-5846. IEEE, 2021. 1 +[63] J. Wu, L. Ma, and X. Hu. Delving deeper into convolutional neural networks for camera relocalization. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), pages 5644-5651, 2017. 2 + +[64] Tete Xiao, Mannat Singh, Eric Mintun, Trevor Darrell, Piotr Dollár, and Ross Girshick. Early convolutions help transformers see better. Advances in Neural Information Processing Systems (NIPS), 34:30392-30400, 2021. 4 +[65] Meng Xu, Youchen Wang, Bin Xu, Jun Zhang, Jian Ren, Stefan Poslad, and Pengfei Xu. A critical analysis of image-based camera pose estimation techniques. arXiv preprint arXiv:2201.05816, 2022. 2 +[66] Fei Xue, Xin Wang, Zike Yan, Qiuyuan Wang, Junqiu Wang, and Hongbin Zha. Local supports global: Deep camera relocalization with sequence enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 8 +[67] Fei Xue, Xin Wu, Shaojun Cai, and Junqiu Wang. Learning multi-view camera relocalization with graph neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 7, 8 +[68] Qunjie Zhou, Torsten Sattler, Marc Pollefeys, and Laura Leal-Taixe. To learn or not to learn: Visual localization from essential matrices. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), pages 3319-3326. IEEE, 2020. 3, 7, 8 \ No newline at end of file diff --git a/2025/HyperPose_ Hypernetwork-Infused Camera Pose Localization and an Extended Cambridge Landmarks Dataset/images.zip b/2025/HyperPose_ Hypernetwork-Infused Camera Pose Localization and an Extended Cambridge Landmarks Dataset/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..73e3fba99646c07ab7209a4112ed565942e5a71e --- /dev/null +++ b/2025/HyperPose_ Hypernetwork-Infused Camera Pose Localization and an Extended Cambridge Landmarks Dataset/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ddfb1935a77dd2919de79ec477710e08ca64523231f7c6f3c6a39d0f9fc1778c +size 685606 diff --git a/2025/HyperPose_ Hypernetwork-Infused Camera Pose Localization and an Extended Cambridge Landmarks Dataset/layout.json b/2025/HyperPose_ Hypernetwork-Infused Camera Pose Localization and an Extended Cambridge Landmarks Dataset/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b8d1c92cdcdae656b1046acd6bd7f25f013f3182 --- /dev/null +++ b/2025/HyperPose_ Hypernetwork-Infused Camera Pose Localization and an Extended Cambridge Landmarks Dataset/layout.json @@ -0,0 +1,7681 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 60, + 103, + 550, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 103, + 550, + 140 + ], + "spans": [ + { + "bbox": [ + 60, + 103, + 550, + 140 + ], + "type": "text", + "content": "HyperPose: Hypernetwork-Infused Camera Pose Localization and an Extended Cambridge Landmarks Dataset" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 236, + 161, + 374, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 161, + 374, + 175 + ], + "spans": [ + { + "bbox": [ + 236, + 161, + 374, + 175 + ], + "type": "text", + "content": "Ron Ferens Yosi Keller" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 153, + 175, + 456, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 175, + 456, + 190 + ], + "spans": [ + { + "bbox": [ + 153, + 175, + 456, + 190 + ], + "type": "text", + "content": "Faculty of Engineering, Bar Ilan University, Ramat-Gan, Israel" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 212, + 191, + 395, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 191, + 395, + 202 + ], + "spans": [ + { + "bbox": [ + 212, + 191, + 395, + 202 + ], + "type": "text", + "content": "{ronferens, yosi.keller}@gmail.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 256, + 297, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 256, + 297, + 533 + ], + "spans": [ + { + "bbox": [ + 53, + 256, + 297, + 533 + ], + "type": "text", + "content": "In this work, we propose HyperPose, which utilizes hypernetworks in absolute camera pose regressors. The inherent appearance variations in natural scenes, attributable to environmental conditions, perspective, and lighting, induce a significant domain disparity between the training and test datasets. This disparity degrades the precision of contemporary localization networks. To mitigate this, we advocate for incorporating hypernetworks into single-scene and multiscene camera pose regression models. During inference, the hypernetwork dynamically computes adaptive weights for the localization regression heads based on the particular input image, effectively narrowing the domain gap. Using indoor and outdoor datasets, we evaluate the HyperPose methodology across multiple established absolute pose regression architectures. We also introduce and share the Extended Cambridge Landmarks (ECL), a novel localization dataset, based on the Cambridge Landmarks dataset, showing it in multiple seasons with significantly varying appearance conditions. Our empirical experiments demonstrate that HyperPose yields notable performance enhancements for single- and multi-scene architectures. We have made our source code, pre-trained models1, and the ECL dataset openly available2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 544, + 135, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 544, + 135, + 558 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 135, + 558 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 565, + 296, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 565, + 296, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 565, + 296, + 685 + ], + "type": "text", + "content": "In many computer vision applications, such as indoor navigation, augmented reality, and autonomous driving, it is essential to determine the position and orientation of a camera using an image taken by that camera. Current camera localization techniques show various trade-offs between accuracy, runtime, and memory usage. Feature Matching (FM) and Scene Coordinate Regression (SCR) are considered the most accurate localization methods. FM-based methods [15, 27, 37-41, 51, 54] involve hierarchical localization pipelines, which commonly use a coarse-to-fine pro" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 340, + 233, + 536, + 289 + ], + "blocks": [ + { + "bbox": [ + 340, + 233, + 536, + 289 + ], + "lines": [ + { + "bbox": [ + 340, + 233, + 536, + 289 + ], + "spans": [ + { + "bbox": [ + 340, + 233, + 536, + 289 + ], + "type": "image", + "image_path": "0849acdbaaba02d7326f257f5c5c056f7ffb98c9286d996bea8f57dd6249927b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 394, + 296, + 475, + 305 + ], + "lines": [ + { + "bbox": [ + 394, + 296, + 475, + 305 + ], + "spans": [ + { + "bbox": [ + 394, + 296, + 475, + 305 + ], + "type": "text", + "content": "(a) Baseline architecture." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 340, + 308, + 533, + 405 + ], + "blocks": [ + { + "bbox": [ + 340, + 308, + 533, + 405 + ], + "lines": [ + { + "bbox": [ + 340, + 308, + 533, + 405 + ], + "spans": [ + { + "bbox": [ + 340, + 308, + 533, + 405 + ], + "type": "image", + "image_path": "104ac34d2fca2452683ed8f52ece7a2f507a453796ac774c7472d2a9569eafae.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 359, + 407, + 509, + 418 + ], + "lines": [ + { + "bbox": [ + 359, + 407, + 509, + 418 + ], + "spans": [ + { + "bbox": [ + 359, + 407, + 509, + 418 + ], + "type": "text", + "content": "(b) Baseline architecture with a hypernetwork." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 427, + 555, + 504 + ], + "lines": [ + { + "bbox": [ + 313, + 427, + 555, + 504 + ], + "spans": [ + { + "bbox": [ + 313, + 427, + 555, + 504 + ], + "type": "text", + "content": "Figure 1. Figure 1a illustrates a baseline APR architecture. It consists of a CNN backbone that generates a feature vector encoding the input query image, followed by regression heads that estimate the translation and orientation. Figure 1b shows the extension of this baseline architecture with a hypernetwork that outputs the weights for the regression heads (translation and orientation) during inference." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 522, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 556, + 715 + ], + "type": "text", + "content": "cessing pipeline that begins with image retrieval to perform an initial localization based on images similar to the query image. The initial estimate is refined by local feature extraction and matching. To estimate the camera's 6-Degrees-Of-Freedom (6DoF) accurately, the resulting 2D-2D matches are mapped to 3D correspondences in the scene's point cloud and then used to determine the camera pose through Perspective-n-Point (PnP) and RANSAC. SCR-based methods [4-8, 47] directly regress correspondences between 2D pixels and 3D coordinates in the scene. Relative pose regression (RPR) schemes [2, 14, 17, 20, 26, 57, 62] directly estimate the relative pose between a query image and a set of reference images with well-defined ground truth poses, eliminating the need for conventional pipelines. However, an image retrieval scheme is essential to retrieve a set of nearest-neighbor images. Another approach, absolute pose" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 232, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 232, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 232, + 703 + ], + "type": "text", + "content": "1https://ronferens.github.io/hyperpose/" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 67, + 703, + 282, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 282, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 282, + 712 + ], + "type": "text", + "content": "2https://ronferens.github.io/extcambridgelandmarks/" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11547" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "regressors (APRs), can estimate the camera pose with a single forward pass using only the query image, resulting in significantly lower latency, although notably less accurate. Additionally, due to their low memory footprint, APRs can be deployed as a standalone application on edge devices with limited computational resources. A survey of visual camera pose localization is given by [43, 65]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 158, + 296, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 158, + 296, + 565 + ], + "spans": [ + { + "bbox": [ + 56, + 158, + 296, + 565 + ], + "type": "text", + "content": "PoseNet by Kendall et al. [25] was the first APR approach, which leveraged a convolutional backbone and a multilayer perceptron (MLP) head to regress the camera's position and orientation. This simple and computationally efficient approach, enabling execution at 5ms, led to numerous APR methods that improved accuracy by modifying the backbone and MLP architectures [11, 32, 33, 44, 49, 60, 61, 63]. In addition, alternative loss functions and optimization strategies were explored [23, 24, 43]. Indoor and outdoor environments exhibit dynamic variability in illumination, perspective, and object motion. APR models, despite their computational efficiency, often struggle to maintain accuracy under these diverse conditions and show limited generalization to novel scenes, necessitating scene-specific training. Consequently, while architecturally simple, a unified network might yield suboptimal performance when adapting to heterogeneous real-world visual inputs. Recent studies have proposed extending APR from single-scene to multiscene formulations. Blanton et al. [3] implemented a convolutional neural network (CNN) backbone to compute a global descriptor for the input image. Rather than employing a scene-specific multilayer perceptron (MLP), they optimized a set of fully connected (FC) layers, allocating one layer per scene based on a predicted scene identifier. Although introducing a framework to optimize a single model across multiple scenes, this methodology did not achieve state-of-the-art (SOTA) APR accuracy. An alternative approach was presented by [46] using dual-branch transformers [58]. Encoders focused on extracting pose-related features, while decoders transformed encoded scene identifiers into latent pose representations. The decoder outputs are leveraged for scene classification and selection of corresponding position and orientation embeddings, which are used to regress position and orientation vectors." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 567, + 295, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 567, + 295, + 675 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 295, + 675 + ], + "type": "text", + "content": "Hypernetworks [22] are a deep learning architecture in which an auxiliary network, the hypernetwork, generates the weights for the primary network (main network) based on the current input and downstream task. The main network and hypernetwork are trained end-to-end using backpropagation. The hypernetwork allows for dynamic and adaptive weight modulation of the main network, leading to improved flexibility in response to changing conditions [16, 18, 34, 55]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "type": "text", + "content": "In this work, we propose HyperPose, a hypernetwork-based approach for absolute camera pose regression. Extending common APR methods [11, 23, 25, 44, 45, 60]," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 334 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 334 + ], + "type": "text", + "content": "the hypernetwork computes weights for the regression head. This enables adaptive analysis of image content, allowing the regression weights to focus on informative features, rather than relying on a single constant pretrained set of weights. We validated the proposed method by integrating hypernetworks into single- and multiscene APR architectures. For single-scene networks, we present HyperPose with an existing state-of-the-art APR model [61] as well as a simplistic baseline APR using an off-the-shelf backbone [53] and shallow regression heads, inspired by the principles outlined in [25]. For multiscene, we extend a leading multiscene APR model [46] by integrating a hypernetwork architecture. In both instances, hypernetwork-augmented models demonstrated superior performance compared to their original counterparts when evaluated on the benchmark datasets in this study. To evaluate localization robustness, we introduce an Extended Cambridge Landmarks (ECL) dataset - an augmentation of the original benchmark [25] with algorithmically generated environmental shifts. By combining evening, winter, and summer scenes, the ECL incorporates dynamics such as lighting shifts and seasonal transitions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 328, + 335, + 514, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 335, + 514, + 346 + ], + "spans": [ + { + "bbox": [ + 328, + 335, + 514, + 346 + ], + "type": "text", + "content": "Thus, we propose the following contributions:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 348, + 553, + 538 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 314, + 348, + 553, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 348, + 553, + 382 + ], + "spans": [ + { + "bbox": [ + 314, + 348, + 553, + 382 + ], + "type": "text", + "content": "- We propose an approach to enhance any existing absolute camera pose regression models by integrating hypernetwork architectures." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 315, + 383, + 553, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 383, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 315, + 383, + 553, + 441 + ], + "type": "text", + "content": "- We comprehensively evaluated the proposed HyperPose on both single- and multiscene pose regressors, demonstrating improved accuracy achieved by the hypernetwork-enhanced models over the original versions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 315, + 443, + 553, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 443, + 553, + 478 + ], + "spans": [ + { + "bbox": [ + 315, + 443, + 553, + 478 + ], + "type": "text", + "content": "- The proposed scheme achieves new SOTA accuracy on multiscene APR benchmarks spanning diverse outdoor and indoor localization tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 315, + 479, + 553, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 479, + 553, + 538 + ], + "spans": [ + { + "bbox": [ + 315, + 479, + 553, + 538 + ], + "type": "text", + "content": "- We introduce the Extended Cambridge Landmarks (ECL) dataset, which builds upon the foundation of the original Cambridge Landmarks dataset [25]. Our ECL data set extends existing test scenes by incorporating various appearance conditions." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 549, + 400, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 549, + 400, + 562 + ], + "spans": [ + { + "bbox": [ + 314, + 549, + 400, + 562 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 570, + 553, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 553, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 553, + 593 + ], + "type": "text", + "content": "Camera pose estimation techniques differ by their input during inference and algorithmic attributes." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": "3D-based or structure-based localization methods are considered the most accurate for camera pose estimation and have demonstrated state-of-the-art performance in leading benchmarks. These methods utilize correspondences between 2D features in an image and 3D world coordinates. [15, 27, 35, 37-41, 51, 54] introduced a two-phase approach for hierarchical pose estimation pipelines. Thus, each query image is first encoded using a CNN trained for image retrieval (IR), based on a pre-mapped image dataset. Tentative correspondences are then estimated by matching local im" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "11548" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": "age features to 3D matches. The resulting matches are used by PnP-RANSAC to estimate the camera pose. DSAC [8] and its succeeding scheme, DSAC++ [4], utilize a CNN architecture to directly estimate 3D coordinates from 2D positions in an image and only require the query image during inference. The same as in Absolute Pose Regression (APR) methods, separate models must be trained for each scene, achieving state-of-the-art accuracy. Lately, [15] proposes a solution to mitigate this gap by introducing a scene-agnostic architecture." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 196, + 295, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 196, + 295, + 364 + ], + "spans": [ + { + "bbox": [ + 55, + 196, + 295, + 364 + ], + "type": "text", + "content": "Relative Pose Regression (RPR) algorithms estimate the relative pose between an input query image and a set of reference images based on their known location. Calculating the relative pose involves estimating the relative position and orientation between the query image and the reference images [1, 6, 68]. An Image Retrieval (IR) approach is utilized to determine the closest set of neighbor images, and the relative motion is then estimated between the query image and each of the retrieved images. The estimated relative poses are used, along with the known position of the reference images, to estimate the camera's absolute pose. Similar to 3D and structured-based localization methods, RPRs involve a multistep process and require access to a database of images with labeled poses during the inference phase." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 367, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 367, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 367, + 295, + 713 + ], + "type": "text", + "content": "Absolute Pose Regression (APR) directly estimates the camera pose given the input query image. APRs were first introduced by Kendall [25], using a modified version of the truncated GoogLeNet architecture, where the Softmax classification layer was replaced by a series of fully connected layers that output the pose estimate. Although less accurate than classical structure-based methods, APRs offer several advantages, such as faster inference times (milliseconds), reduced memory footprint (megabytes), and do not require feature engineering. Variations of the encoder and MLP architectures have been proposed to improve APR accuracy [11, 23, 33, 33, 44, 49]. Furthermore, modifications to the loss function have been suggested to improve the balance between orientation and position objectives [24] or to incorporate other types of information [9]. For an in-depth examination of the various architectural designs and loss functions utilized in camera pose regression, please refer to [43]. Although demonstrating faster inference times and low memory requirements, absolute pose regression (APR) models typically require dedicated perscene training to achieve competitive accuracy compared to other methods. This limitation has raised concerns regarding the value of APRs versus significantly more accurate techniques such as [4, 6, 8, 36]. To address this issue, recent approaches suggested extending APRs to a multiscene paradigm. [3] introduced Multi-Scene PoseNet (MSPN), which utilizes a CNN to regress an image feature vector. The two branches then process this feature vector. The first branch classifies the related scene using a fully connected" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "type": "text", + "content": "layer with a softmax operator. The estimated scene index selects scene-specific weights from a database. The second branch combines these weights with the feature vector to regress the pose. The model is trained end-to-end by applying both binary cross-entropy (for classification) and camera pose losses for each scene. Inspired by recent successful applications of Transformers [12, 19, 29] to computer vision tasks, [46] suggested a shared convolutional network backbone followed by two network branches consisting of a Transformer and a regression head. The first regresses the input query image's pose, and the other its orientation. The main incentive for utilizing Transformers encoders is to focus on pose-informative features and decoders to transform encoded scene identifiers into latent pose representations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 239, + 553, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 553, + 348 + ], + "type": "text", + "content": "Hypernetworks were first presented by Ha et al. [22]. These are neural networks that have been designed to predict the weights of a primary network. The primary network's weights can be adjusted based on specific inputs, resulting in a more expressive and adaptive model. Hypernetworks have been used in various applications such as learning implicit neural representations [16], semantic segmentation [34], 3D scene representation and modeling [28, 48, 50] and continuous learning [59], to name a few." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 357, + 471, + 370 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 471, + 370 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 471, + 370 + ], + "type": "text", + "content": "3. APRs using Hypernetworks" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 376, + 553, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 553, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 553, + 556 + ], + "type": "text", + "content": "HyperPose is a general approach, applicable to both single- and multi-scene scenarios, for enhancing APR schemes. It comprises two key components: the main network (APR) designed to estimate the camera's absolute position and the hypernetwork that computes the regression weights for the main network conditioned on the input query image. The main network performs a forward pass on the input image to localize the capture camera. The camera pose is represented as the tuple " + }, + { + "bbox": [ + 313, + 376, + 553, + 556 + ], + "type": "inline_equation", + "content": "< \\mathbf{x},\\mathbf{q}>" + }, + { + "bbox": [ + 313, + 376, + 553, + 556 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 376, + 553, + 556 + ], + "type": "inline_equation", + "content": "\\mathbf{x}\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 376, + 553, + 556 + ], + "type": "text", + "content": " is the camera's position in world coordinates and " + }, + { + "bbox": [ + 313, + 376, + 553, + 556 + ], + "type": "inline_equation", + "content": "\\mathbf{q}\\in \\mathbb{R}^4" + }, + { + "bbox": [ + 313, + 376, + 553, + 556 + ], + "type": "text", + "content": " is the quaternion encoding of its " + }, + { + "bbox": [ + 313, + 376, + 553, + 556 + ], + "type": "inline_equation", + "content": "3D" + }, + { + "bbox": [ + 313, + 376, + 553, + 556 + ], + "type": "text", + "content": " orientation. Following [13], we formulate the hypernetwork " + }, + { + "bbox": [ + 313, + 376, + 553, + 556 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 376, + 553, + 556 + ], + "type": "text", + "content": " as a parametric function " + }, + { + "bbox": [ + 313, + 376, + 553, + 556 + ], + "type": "inline_equation", + "content": "\\theta_h = H(\\theta ,\\mathbf{x})" + }, + { + "bbox": [ + 313, + 376, + 553, + 556 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 376, + 553, + 556 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 376, + 553, + 556 + ], + "type": "text", + "content": " represents the parameters of the hypernetwork, " + }, + { + "bbox": [ + 313, + 376, + 553, + 556 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 376, + 553, + 556 + ], + "type": "text", + "content": " the network's input, and " + }, + { + "bbox": [ + 313, + 376, + 553, + 556 + ], + "type": "inline_equation", + "content": "\\theta_h" + }, + { + "bbox": [ + 313, + 376, + 553, + 556 + ], + "type": "text", + "content": " the parameters of the regression heads layer for the input query image " + }, + { + "bbox": [ + 313, + 376, + 553, + 556 + ], + "type": "inline_equation", + "content": "I_q" + }, + { + "bbox": [ + 313, + 376, + 553, + 556 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 562, + 440, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 562, + 440, + 573 + ], + "spans": [ + { + "bbox": [ + 313, + 562, + 440, + 573 + ], + "type": "text", + "content": "3.1. Network Architecture" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 579, + 487, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 579, + 487, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 579, + 487, + 590 + ], + "type": "text", + "content": "3.1.1. HyperPose for Single-Scene APRs" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 594, + 553, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 553, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 553, + 629 + ], + "type": "text", + "content": "We apply the proposed HyperPose-based approach to single-scene APR using a baseline APR following [25], and the state-of-the-art AtLoc APR [61]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "content": "Baseline APR with Hypernetwork. As depicted in Fig. 1a, the baseline model's main network consists of a CNN backbone that generates a feature vector encoding the input query image. This vector is then processed by two regression heads. The first estimates the camera position, and the other estimates the orientation. Figure 1b illustrates the integration of a hypernetwork into this architecture. The input" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11549" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 77, + 549, + 322 + ], + "blocks": [ + { + "bbox": [ + 70, + 77, + 549, + 322 + ], + "lines": [ + { + "bbox": [ + 70, + 77, + 549, + 322 + ], + "spans": [ + { + "bbox": [ + 70, + 77, + 549, + 322 + ], + "type": "image", + "image_path": "4bc874b428f303788cf723c4f5595bc3997d4dff65879b6e17c5e680ebdf7ba8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 333, + 555, + 392 + ], + "lines": [ + { + "bbox": [ + 54, + 333, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 54, + 333, + 555, + 392 + ], + "type": "text", + "content": "Figure 2. MS-HyperPose - The proposed multi-scene absolute pose regression architecture using a hypernetwork. The primary network employs position and orientation Transformers, in a dual-branch architecture, to extract activation maps from the underlying convolutional backbone. The hypernetwork generates the weights for the regression head in the primary network, using the input query image. These adaptive weights, combined with the latent vectors generated by the Transformers estimate the camera pose, consisting of the spatial position (x) and angular orientation (q)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 410, + 294, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 410, + 294, + 505 + ], + "spans": [ + { + "bbox": [ + 54, + 410, + 294, + 505 + ], + "type": "text", + "content": "to the hypernetwork is the feature vector generated by the CNN backbone of the main network. This vector is passed through a single fully-connected (FC) layer followed by a Swish activation [36] function. The processed vector serves as an input to a multilayer perceptron (MLP) that outputs the weights for a single layer in the regression heads of the main network. The camera pose is then estimated by the position and orientation regression heads of the main network." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 507, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 507, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 54, + 507, + 295, + 651 + ], + "type": "text", + "content": "AtLoc with Hypernetwork. [61] presents Attention-Guided Camera Localization (AtLoc), an absolute pose regression model based on self-attention. In this architecture, a visual encoder encodes an image into a latent vector. Conditioned on these extracted features, an attention module computes self-attention maps to reweight the representation into a new feature space. A pose regressor then maps the transformed features after the attention operators to estimate the camera pose. We incorporate a hypernetwork into the AtLoc, the same as for the baseline APR model. In this case, the input to the hypernetwork is the reweighted representation feature vector computed by the visual encoder." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 661, + 226, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 661, + 226, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 661, + 226, + 673 + ], + "type": "text", + "content": "3.1.2. HyperPose for Multi-Scene APRs" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "text", + "content": "The suggested multi-scene model, named MS-HyperPose is an extension of the architecture introduced in [45]. As depicted in Fig. 2, the input query image " + }, + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "inline_equation", + "content": "I_{q}" + }, + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "text", + "content": " is processed by a" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 410, + 555, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 410, + 555, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 410, + 555, + 578 + ], + "type": "text", + "content": "convolutional backbone that computes intermediate activation maps from two different layers. These activation maps are then utilized as inputs to the main network and the hypernetwork. Within the main network, the positional and orientational Transformers process the activation maps and integrate sequential representations into single latent vectors. After converting the activation maps into representation vectors, these vectors, along with the latent vectors are passed through multilayer perceptron (MLP) layers in the hypernetwork to generate the weights " + }, + { + "bbox": [ + 313, + 410, + 555, + 578 + ], + "type": "inline_equation", + "content": "\\theta_h" + }, + { + "bbox": [ + 313, + 410, + 555, + 578 + ], + "type": "text", + "content": " for the translation and orientation regression heads. Using the weights " + }, + { + "bbox": [ + 313, + 410, + 555, + 578 + ], + "type": "inline_equation", + "content": "\\theta_h" + }, + { + "bbox": [ + 313, + 410, + 555, + 578 + ], + "type": "text", + "content": " produced by the hypernetwork " + }, + { + "bbox": [ + 313, + 410, + 555, + 578 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 410, + 555, + 578 + ], + "type": "text", + "content": " and the regression heads in the main network, the position, and orientation of the input query image are estimated." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": "The Main Network consists of a convolutional backbone and two separate branches to regress the camera's position and orientation. Each branch includes its particular Transformer and a multilayer perceptron (MLP) head. Studies have demonstrated that substituting the patchify stem, utilized in the early visual processing of a Vision Transformer (ViT), with a convolutional stem enhances optimization stability and elevates performance to a higher level [64]. Therefore, given an input query image " + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "inline_equation", + "content": "I_{q} \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": ", a shared convolutional backbone generates activation maps. Two different activation maps are selected" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "11550" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 122 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 122 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 122 + ], + "type": "text", + "content": "from the CNN based on the particular learned task (Fig. 2a). The activation maps, represented as " + }, + { + "bbox": [ + 55, + 72, + 294, + 122 + ], + "type": "inline_equation", + "content": "M \\in \\mathbb{R}^{h \\times w \\times C^{Fn}}" + }, + { + "bbox": [ + 55, + 72, + 294, + 122 + ], + "type": "text", + "content": ", are processed and transformed into " + }, + { + "bbox": [ + 55, + 72, + 294, + 122 + ], + "type": "inline_equation", + "content": "\\tilde{M} \\in \\mathbb{R}^{h \\times w \\times C^{Out}}" + }, + { + "bbox": [ + 55, + 72, + 294, + 122 + ], + "type": "text", + "content": " using a " + }, + { + "bbox": [ + 55, + 72, + 294, + 122 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 55, + 72, + 294, + 122 + ], + "type": "text", + "content": " convolutional layer followed by flattening (Fig. 2b)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 123, + 295, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 123, + 295, + 254 + ], + "spans": [ + { + "bbox": [ + 54, + 123, + 295, + 254 + ], + "type": "text", + "content": "Following [45], we use the Transformer architecture from [12], modifying the standard Encoder and Decoder to add positional encoding at each attention layer. Each Transformer has " + }, + { + "bbox": [ + 54, + 123, + 295, + 254 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 54, + 123, + 295, + 254 + ], + "type": "text", + "content": " identical blocks, each with a multi-head attention (MHA) mechanism and a two-layer MLP with Swish activation [36]. Layer Normalization (LN) is applied before MHA and MLP processing, as in [58], with outputs merged via residual connections and dropout. To preserve spatial information, each activation map position receives a learned encoding. The processed sequence, serving as input to the Transformer is given by (Fig. 2c):" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 86, + 272, + 295, + 289 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 272, + 295, + 289 + ], + "spans": [ + { + "bbox": [ + 86, + 272, + 295, + 289 + ], + "type": "interline_equation", + "content": "E ^ {I n} = \\left[ T, \\tilde {M} \\right] \\in \\mathbb {R} ^ {(h w) \\times C ^ {O u t}}, T \\in \\mathbb {R} ^ {C _ {M} ^ {O u t}}. \\tag {1}", + "image_path": "af9fd6d17dbc1c3d27a7f3cc4824198d62638fbaeaa0754a0a2dc94e0c0f6d84.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 297, + 294, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 297, + 294, + 309 + ], + "spans": [ + { + "bbox": [ + 67, + 297, + 294, + 309 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 67, + 297, + 294, + 309 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 67, + 297, + 294, + 309 + ], + "type": "text", + "content": " is the positional encoding of the activation map." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 309, + 294, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 309, + 294, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 309, + 294, + 380 + ], + "type": "text", + "content": "Given a dataset with " + }, + { + "bbox": [ + 55, + 309, + 294, + 380 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 309, + 294, + 380 + ], + "type": "text", + "content": " scenes, the Transformer Decoders generate embeddings " + }, + { + "bbox": [ + 55, + 309, + 294, + 380 + ], + "type": "inline_equation", + "content": "E^{Out} = \\in \\mathbb{R}^{N\\times C^{Out}}" + }, + { + "bbox": [ + 55, + 309, + 294, + 380 + ], + "type": "text", + "content": " (Fig. 2d). For a query image, only one position matches its scene. The combined Transformer outputs (Fig. 2e) pass through a fully connected layer and Log SoftMax, selecting the highest-probability vectors (Fig. 2f)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 380, + 294, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 380, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 55, + 380, + 294, + 441 + ], + "type": "text", + "content": "The selected vectors " + }, + { + "bbox": [ + 55, + 380, + 294, + 441 + ], + "type": "inline_equation", + "content": "E_{x_i}^{Out}" + }, + { + "bbox": [ + 55, + 380, + 294, + 441 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 380, + 294, + 441 + ], + "type": "inline_equation", + "content": "E_{q_i}^{Out}" + }, + { + "bbox": [ + 55, + 380, + 294, + 441 + ], + "type": "text", + "content": " are processed through an MLP. Finally, the estimated camera position and orientation are obtained by summing the outputs of the main network and hypernetwork, using weights predicted by the hypernetwork (Fig. 2h and 2i)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 441, + 295, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 295, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 295, + 571 + ], + "type": "text", + "content": "The Hypernetworks consist of MLPs, each regressing the weights for a corresponding regression head (position or orientation) in the main network. Their input is the sum of task-specific encoded activation maps " + }, + { + "bbox": [ + 55, + 441, + 295, + 571 + ], + "type": "inline_equation", + "content": "\\tilde{M}_x" + }, + { + "bbox": [ + 55, + 441, + 295, + 571 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 441, + 295, + 571 + ], + "type": "inline_equation", + "content": "\\tilde{M}_q" + }, + { + "bbox": [ + 55, + 441, + 295, + 571 + ], + "type": "text", + "content": ", along with global latent vectors " + }, + { + "bbox": [ + 55, + 441, + 295, + 571 + ], + "type": "inline_equation", + "content": "E_{x_i}^{Out}" + }, + { + "bbox": [ + 55, + 441, + 295, + 571 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 441, + 295, + 571 + ], + "type": "inline_equation", + "content": "E_{q_i}^{Out}" + }, + { + "bbox": [ + 55, + 441, + 295, + 571 + ], + "type": "text", + "content": " (Fig. 2g). To ensure size compatibility, the encoded activation map passes through a fully connected layer before summation. This vector is then processed by MLPs with Swish activation, generating inputs for another MLP set, each estimating weights for a specific layer in the main network's regression heads:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 86, + 590, + 295, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 590, + 295, + 609 + ], + "spans": [ + { + "bbox": [ + 86, + 590, + 295, + 609 + ], + "type": "interline_equation", + "content": "\\theta_ {h} ^ {O u t _ {j}} = H \\left(\\theta_ {h} ^ {O u t _ {j}}, E _ {i} ^ {O u t}\\right) \\in \\mathbb {R} ^ {C _ {E _ {i}} ^ {O u t} \\times C _ {H} ^ {O u t _ {j}}}, \\tag {2}", + "image_path": "16abd21e3bdafeb9e06b27a2984f1ec9e812485b97ef01b36a74958edaa9a68b.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 616, + 264, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 616, + 264, + 629 + ], + "spans": [ + { + "bbox": [ + 67, + 616, + 264, + 629 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 67, + 616, + 264, + 629 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 67, + 616, + 264, + 629 + ], + "type": "text", + "content": " is the index of the regression head layer." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 635, + 181, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 181, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 181, + 649 + ], + "type": "text", + "content": "3.2. Optimization Criteria" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "content": "The localization error is calculated by comparing the translation position and orientation of the ground truth pose " + }, + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{gt} = < \\mathbf{x}_{gt},\\mathbf{q}_{gt}>" + }, + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "content": " , where " + }, + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "content": " represents the position of the camera in the world and " + }, + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{q}\\in \\mathbb{R}^4" + }, + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "content": " denotes its orientation encoded as a quaternion, with the estimated pose" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{est} = < \\mathbf{x}_{est},\\mathbf{q}_{est} >" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " . The translation error, represented by " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "L_{x}" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " , is calculated by the Euclidean distance between the ground truth and the estimated camera positions. The orientation error, " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "L_{q}" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " , is defined as the Euclidean distance between the ground truth and the estimated unit quaternions [43]:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 396, + 156, + 553, + 168 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 396, + 156, + 553, + 168 + ], + "spans": [ + { + "bbox": [ + 396, + 156, + 553, + 168 + ], + "type": "interline_equation", + "content": "L _ {x} = \\left\\| \\mathbf {x} _ {0} - \\mathbf {x} \\right\\| _ {2}, \\tag {3}", + "image_path": "08a38af49ecba041e8057aa114e99d083cfbc6f0c59738558e3642b8c2ef686d.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 391, + 173, + 553, + 200 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 173, + 553, + 200 + ], + "spans": [ + { + "bbox": [ + 391, + 173, + 553, + 200 + ], + "type": "interline_equation", + "content": "L _ {q} = \\left\\| \\mathbf {q} _ {0} - \\frac {\\mathbf {q}}{| | \\mathbf {q} | |} \\right\\| _ {2} \\tag {4}", + "image_path": "cef188e4de78c1bed26a0233000e724e0b9c867dad1054b74efb4d72a51e0ed6.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 203, + 553, + 226 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 203, + 553, + 226 + ], + "spans": [ + { + "bbox": [ + 313, + 203, + 553, + 226 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 203, + 553, + 226 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 313, + 203, + 553, + 226 + ], + "type": "text", + "content": " is a normalized norm quaternion to ensure it is a valid orientation encoding." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 228, + 554, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 554, + 299 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 554, + 299 + ], + "type": "text", + "content": "As a multi-regression optimization problem, we adopt the method outlined in [24] by using different weighting schemes, based on the uncertainty of each task, to balance the individual loss functions defined for the separate objectives. The model optimizes the aggregated loss function, a combination of the weighted individual losses." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 337, + 320, + 553, + 334 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 320, + 553, + 334 + ], + "spans": [ + { + "bbox": [ + 337, + 320, + 553, + 334 + ], + "type": "interline_equation", + "content": "L _ {p} = L _ {x} \\exp (- s _ {x}) + s _ {x} + L _ {q} \\exp (- s _ {q}) + s _ {q}, \\tag {5}", + "image_path": "38294ea44dbe48279cad7edac9cd0e282e8fc7377b67114a79ab8557a3048529.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 342, + 554, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 554, + 413 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 554, + 413 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 342, + 554, + 413 + ], + "type": "inline_equation", + "content": "S_{x}" + }, + { + "bbox": [ + 313, + 342, + 554, + 413 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 342, + 554, + 413 + ], + "type": "inline_equation", + "content": "S_{q}" + }, + { + "bbox": [ + 313, + 342, + 554, + 413 + ], + "type": "text", + "content": " are the learned parameters. As a further optimization, we adopt the approach detailed in [44] and follow a three-step training process. Firstly, the entire network is trained using the aggregated loss in Eq. 5. In the next two stages, each MLP head undergoes separate fin-tuning with its corresponding loss function." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 421, + 447, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 421, + 447, + 433 + ], + "spans": [ + { + "bbox": [ + 314, + 421, + 447, + 433 + ], + "type": "text", + "content": "3.3. Implementation Details" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 437, + 554, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 554, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 554, + 569 + ], + "type": "text", + "content": "The shared backbone uses EfficientNet-B0 [31, 52]. We used the activation maps of two endpoints (reduction levels) as input for our position and orientation branches: " + }, + { + "bbox": [ + 313, + 437, + 554, + 569 + ], + "type": "inline_equation", + "content": "m_{rdct4} \\in \\mathbb{R}^{14 \\times 14 \\times 112}" + }, + { + "bbox": [ + 313, + 437, + 554, + 569 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 437, + 554, + 569 + ], + "type": "inline_equation", + "content": "m_{rdct3} \\in \\mathbb{R}^{28 \\times 28 \\times 40}" + }, + { + "bbox": [ + 313, + 437, + 554, + 569 + ], + "type": "text", + "content": ", respectively. Following [45], we applied linear projections on each activation map to a common depth dimension of " + }, + { + "bbox": [ + 313, + 437, + 554, + 569 + ], + "type": "inline_equation", + "content": "C_M^{Out} = 256" + }, + { + "bbox": [ + 313, + 437, + 554, + 569 + ], + "type": "text", + "content": " and learned the positional encoding of the same depth. The regression heads in the main network consist of two-layer MLPs with input and output layers. The input layer dimension is 1024 for both the position and orientation, while the output dimensions are 3 and 4, respectively." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 570, + 553, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 553, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 553, + 629 + ], + "type": "text", + "content": "Single-Scene APR. For both the baseline APR and the HyperPose extension of the AtLoc [61] model, the hypernetwork for the position branch generates weights for an MLP with a dimension of 256, while the orientation hypernetwork produces weights of 512 dimensions." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": "Multi-Scene APR. In contrast to [45], we incorporated four-layer (instead of six) Transformers in each branch (position and orientation) within the main network architecture. Each block consists of an MHA layer with four heads, followed by an MLP that preserves the input dimension. A dropout rate of " + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "inline_equation", + "content": "p = 0.1" + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": " was applied to each MLP layer. The output dimension of each Transformer in the main network" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11551" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 116, + 112, + 492, + 172 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 104 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 104 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 104 + ], + "type": "text", + "content": "Table 1. Comparative analysis of APR architectures with and without the purported hypernetwork addition using the Cambridge Landmarks dataset. We report the median position/orientation error in meters/degrees. The best performance is marked in bold. (*) The authors of AtLoc did not report results on the Cambridge dataset, and we trained the AtLoc model on this dataset." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 116, + 112, + 492, + 172 + ], + "lines": [ + { + "bbox": [ + 116, + 112, + 492, + 172 + ], + "spans": [ + { + "bbox": [ + 116, + 112, + 492, + 172 + ], + "type": "table", + "html": "
MethodK. CollegeOld HospitalShop FacadeSt. MaryAvg.
AtLoc* [61]1.53,3.212.17,3.990.93,4.612.14,5.761.69,4.39
AtLoc w/ hypernet0.96,3.432.04,3.610.88,4.221.72,5.441.40,4.18
Baseline APR0.89,2.291.49,3.320.74,4.791.40,4.951.13,3.84
Baseline APR w/ hypernet0.77,2.071.47,3.290.72,4.011.37,4.901.08,3.57
", + "image_path": "0429cb8d91235b68fcdb01e6d4d531c1d5ede97fed85c9d46d8c3f6535327046.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 182, + 296, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 182, + 296, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 182, + 296, + 326 + ], + "type": "text", + "content": "was " + }, + { + "bbox": [ + 55, + 182, + 296, + 326 + ], + "type": "inline_equation", + "content": "C_M^{Out} = 256" + }, + { + "bbox": [ + 55, + 182, + 296, + 326 + ], + "type": "text", + "content": ". The position and orientation branches of the hypernetwork generate the weights of three regression layers: input, hidden, and output. The number of weights in the position and orientation branches differ. The hypernetwork's position branch produces the weights for an MLP with dimensions 256, 128, and 3, while the orientation branch generates weights for an MLP with dimensions 512, 512, and 4. The weights generated by the hypernetwork are then put into the regression heads of the main network to estimate the position and orientation of the input query image. The input and hidden layers of the regression head utilize the Swish activation function." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 338, + 285, + 353 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 338, + 285, + 353 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 285, + 353 + ], + "type": "text", + "content": "4. Extended Cambridge Landmarks Dataset" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 57, + 369, + 173, + 435 + ], + "blocks": [ + { + "bbox": [ + 57, + 369, + 173, + 435 + ], + "lines": [ + { + "bbox": [ + 57, + 369, + 173, + 435 + ], + "spans": [ + { + "bbox": [ + 57, + 369, + 173, + 435 + ], + "type": "image", + "image_path": "dc3759e8b063cfa79df346d6a900aa23b12aa2b490f3d8304796f2c10cf416f9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 96, + 437, + 134, + 446 + ], + "lines": [ + { + "bbox": [ + 96, + 437, + 134, + 446 + ], + "spans": [ + { + "bbox": [ + 96, + 437, + 134, + 446 + ], + "type": "text", + "content": "(a) Original" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 179, + 369, + 295, + 435 + ], + "blocks": [ + { + "bbox": [ + 179, + 369, + 295, + 435 + ], + "lines": [ + { + "bbox": [ + 179, + 369, + 295, + 435 + ], + "spans": [ + { + "bbox": [ + 179, + 369, + 295, + 435 + ], + "type": "image", + "image_path": "323c1720ebdf2614c798d4faabf7f1b8418682de06d5453ea43ba4fc2cde6b98.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 217, + 437, + 257, + 446 + ], + "lines": [ + { + "bbox": [ + 217, + 437, + 257, + 446 + ], + "spans": [ + { + "bbox": [ + 217, + 437, + 257, + 446 + ], + "type": "text", + "content": "(b)Evening" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 57, + 447, + 173, + 513 + ], + "blocks": [ + { + "bbox": [ + 57, + 447, + 173, + 513 + ], + "lines": [ + { + "bbox": [ + 57, + 447, + 173, + 513 + ], + "spans": [ + { + "bbox": [ + 57, + 447, + 173, + 513 + ], + "type": "image", + "image_path": "76acff78b28455974bf3c466a8d1bb63973e185e0580a784b91d7c118780f857.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 95, + 515, + 134, + 523 + ], + "lines": [ + { + "bbox": [ + 95, + 515, + 134, + 523 + ], + "spans": [ + { + "bbox": [ + 95, + 515, + 134, + 523 + ], + "type": "text", + "content": "(c) Summer" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 179, + 447, + 295, + 512 + ], + "blocks": [ + { + "bbox": [ + 179, + 447, + 295, + 512 + ], + "lines": [ + { + "bbox": [ + 179, + 447, + 295, + 512 + ], + "spans": [ + { + "bbox": [ + 179, + 447, + 295, + 512 + ], + "type": "image", + "image_path": "c7466b1bc9048e773f40d442bc0bb3668461139afb0373c26fc7d0ae149a9d07.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 515, + 254, + 523 + ], + "lines": [ + { + "bbox": [ + 220, + 515, + 254, + 523 + ], + "spans": [ + { + "bbox": [ + 220, + 515, + 254, + 523 + ], + "type": "text", + "content": "(d) Winter" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 533, + 295, + 556 + ], + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 556 + ], + "type": "text", + "content": "Figure 3. A sample from the King's College Scene in the proposed Extended Cambridge Landmarks (ECL) Dataset" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": "To further assess the limitations of the HyperPose approach, we constructed an augmented version of the Cambridge Landmarks [25] benchmark called the Extended Cambridge Landmarks (ECL) dataset. Leveraging the original scenes, the ECL dataset introduces new localization challenges due to variations in appearance by creating synthetic evening, winter, and summer variants for each location. These scene permutations simulate real-world environmental variations by introducing shifts in lighting, weather patterns, foliage density, and more. Analyzing performance across ECL variations and their original Cambridge counterparts allows us to evaluate the robustness of" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 182, + 555, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 182, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 555, + 361 + ], + "type": "text", + "content": "models to real-world variations. These augmentations were conducted by utilizing the novel InstructPix2Pix work [10]. Visual samples of the ECL dataset can be found in the supplementary materials. Figure 3 shows sample from the King's College scene in the extended dataset. Additional samples from each scene in the ECL dataset can be found in the supplementary material. Quantitative pose regression analysis between these more challenging conditions and unmodified sequences helps scientifically assess model adaptability. We hypothesize that exposing limitations unseen in conventional test sets will motivate the development of more invariant solutions. As controllable expansions of established benchmarks, programmatically generated datasets can thus guide the development of robust models for real-world use." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 373, + 440, + 386 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 440, + 386 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 440, + 386 + ], + "type": "text", + "content": "5. Experimental Results" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 390, + 556, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 390, + 556, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 390, + 556, + 700 + ], + "type": "text", + "content": "We conducted an experimental evaluation using the 7 Scenes [21] and Cambridge Landmarks [25] datasets for evaluating absolute pose regression. The Cambridge Landmarks dataset represents an outdoor urban environment ranging in size from 900 to 5500 square meters. We present results for four of its six scenes. We excluded the remaining two scenes due to limited coverage in prior work. To compare robustness, we also report the results on the Extended Cambridge Landmarks dataset (ECL) (Section 1 in the supplementary materials). The 7 Scenes dataset consists of seven small-scale indoor scenes with spatial extents ranging from approximately 1 to 10 square meters. These datasets present various localization challenges, including differences in scale, indoor versus outdoor settings, repetitive elements, textureless features, significant viewpoint changes, and variations in trajectory between training and test sets. To facilitate a broader analysis of our proposed method, we conduct an additional evaluation on the Oxford RobotCar Dataset [30]. Such dynamic environmental factors assess HyperPose's ability to improve localization accuracy amid diverse, challenging visual contexts. Our experimental procedures adhere to the same training and evaluation protocols as in [9, 61]. Additional information concerning these sequences, and a comprehensive description of the training methodology and hyperparameter selection for the models, are provided in the supplementary material." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 326, + 701, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 555, + 713 + ], + "type": "text", + "content": "We evaluate the impact of integrating hypernetworks in" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11552" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 105, + 101, + 504, + 371 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 2. Comparison to single-scene state-of-the-art methods - Cambridge Landmarks: We report the median position/orientation error in meters/degrees. The most effective single and multi-scene APRs are marked in bold." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 105, + 101, + 504, + 371 + ], + "lines": [ + { + "bbox": [ + 105, + 101, + 504, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 101, + 504, + 371 + ], + "type": "table", + "html": "
K. CollegeHospitalShop FacadeSt. MaryAvg.
Seq.DSAC* [6]0.18,0.3°0.21,0.4°0.05,0.3°0.15,0.5°0.15,0.4°
MapNet [9]1.08,1.9°1.94,3.9°1.49,4.2°2.00,4.5°1.63,3.6°
GL-Net [67]0.59,0.7°1.88,2.8°0.50,2.9°1.90,3.3°1.22,2.4°
IRVLAD [56]2.80,5.7°4.01,7.1°1.11,7.6°2.31,8.0°2.56,7.1°
VLAD+Inter [42]1.48,4.5°2.68,4.6°0.90,4.3°1.62,6.1°1.67,4.9°
RPREssNet [68]0.76,1.9°1.39,2.8°0.84,4.3°1.32,4.7°1.08,3.4°
NC-EssNet [68]0.61,1.6°0.95,2.7°0.7,3.4°1.12,3.6°0.85,2.8°
RelocGNN[21]0.48,1.0°1.14,2.5°0.48,2.5°1.52,3.2°0.91,2.3°
APRPoseNet [25]1.92,5.40°2.31,5.38°1.46,8.08°2.65,8.48°2.08,6.83°
BayesianPN [23]1.74,4.06°2.57,5.14°1.25,7.54°2.11,8.38°1.91,6.28°
LSTM-PN [60]0.99,3.65°1.51,4.29°1.18,7.44°1.52,6.68°1.30,5.57°
SVS-Pose [33]1.06,2.81°1.50,4.03°0.63,5.73°2.11,8.11°1.32,5.17°
GPoseNet [11]1.61,2.29°2.62,3.89°1.14,5.73°2.93,6.46°2.07,4.59°
PoseNetLearn [24]0.99,1.06°2.17,2.94°1.05,3.97°1.49,3.43°1.42,2.85°
GeoPoseNet [24]0.88,1.04°3.20,3.29°0.88,3.78°1.57,3.32°1.63,2.86°
IRPNet [44]1.18,2.19°1.87,3.38°0.72,3.47°1.87,4.94°1.41,3.50°
Baseline APR w/ hypernet0.61,1.84°1.44,3.03°0.70,3.62°1.37,4.85°1.03,2.67°
MS-APRMSPN [3]1.73,3.65°2.55,4.05°2.92,7.49°2.67,6.18°2.47,5.34°
MS-Trans[46]0.83,1.47°1.81,2.39°0.86,3.07°1.62,3.99°1.28,2.73°
MS-HyperPose0.78,1.18°1.84,2.11°0.83,3.13°1.61,3.22°1.27,2.41°
", + "image_path": "a77f329e43f19698f0702f810ed8838afb7facfc19e3c84f4c300d38753d13a3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 381, + 294, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 381, + 294, + 704 + ], + "spans": [ + { + "bbox": [ + 56, + 381, + 294, + 704 + ], + "type": "text", + "content": "single-scene APRs. To this end, we compare the accuracy of two model variations, with and without hypernetworks, and report median position and orientation errors. Table 1 presents results on the Cambridge dataset. The hypernetwork improves the accuracy of the proposed baseline APR and AtLoc [61] models. The baseline APR model size is " + }, + { + "bbox": [ + 56, + 381, + 294, + 704 + ], + "type": "inline_equation", + "content": "26.7\\mathrm{MB}" + }, + { + "bbox": [ + 56, + 381, + 294, + 704 + ], + "type": "text", + "content": " with an inference time of " + }, + { + "bbox": [ + 56, + 381, + 294, + 704 + ], + "type": "inline_equation", + "content": "9.32\\mathrm{ms}" + }, + { + "bbox": [ + 56, + 381, + 294, + 704 + ], + "type": "text", + "content": ", increasing to " + }, + { + "bbox": [ + 56, + 381, + 294, + 704 + ], + "type": "inline_equation", + "content": "40.9\\mathrm{MB}" + }, + { + "bbox": [ + 56, + 381, + 294, + 704 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 381, + 294, + 704 + ], + "type": "inline_equation", + "content": "9.45\\mathrm{ms}" + }, + { + "bbox": [ + 56, + 381, + 294, + 704 + ], + "type": "text", + "content": " with a hypernetwork. The original published AtLoc model is " + }, + { + "bbox": [ + 56, + 381, + 294, + 704 + ], + "type": "inline_equation", + "content": "97.9\\mathrm{MB}" + }, + { + "bbox": [ + 56, + 381, + 294, + 704 + ], + "type": "text", + "content": " with an inference time of " + }, + { + "bbox": [ + 56, + 381, + 294, + 704 + ], + "type": "inline_equation", + "content": "5.14\\mathrm{ms}" + }, + { + "bbox": [ + 56, + 381, + 294, + 704 + ], + "type": "text", + "content": ". The hypernetwork-enhanced AtLoc version has a model size of " + }, + { + "bbox": [ + 56, + 381, + 294, + 704 + ], + "type": "inline_equation", + "content": "116.8\\mathrm{MB}" + }, + { + "bbox": [ + 56, + 381, + 294, + 704 + ], + "type": "text", + "content": " and an inference time of " + }, + { + "bbox": [ + 56, + 381, + 294, + 704 + ], + "type": "inline_equation", + "content": "5.56\\mathrm{ms}" + }, + { + "bbox": [ + 56, + 381, + 294, + 704 + ], + "type": "text", + "content": ". Note that the reported errors refer to the models' performance after the first stage of model training, without applying the two additional optimization steps described in the training details (See supplementary materials). Furthermore, Tables 2 and 3 compare the median position error, and the orientation error, of the proposed baseline APR with hypernetwork architecture, with the results of recent SOTA localization algorithms, using the Cambridge Landmarks and 7Scenes datasets, respectively. For the indoor and outdoor datasets, in addition to the results of leading APRs, we report the median position and orientation errors of structure-based, sequence-based, IR, and RPR methods. Although DSAC* [6] achieves SOTA accuracy, compared to all other APR methods, HyperPose obtains the most accurate average of " + }, + { + "bbox": [ + 56, + 381, + 294, + 704 + ], + "type": "inline_equation", + "content": "1.03m" + }, + { + "bbox": [ + 56, + 381, + 294, + 704 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 56, + 381, + 294, + 704 + ], + "type": "inline_equation", + "content": "2.67^{\\circ}" + }, + { + "bbox": [ + 56, + 381, + 294, + 704 + ], + "type": "text", + "content": " on the Cambridge outdoor dataset, and " + }, + { + "bbox": [ + 56, + 381, + 294, + 704 + ], + "type": "inline_equation", + "content": "0.19m" + }, + { + "bbox": [ + 56, + 381, + 294, + 704 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 56, + 381, + 294, + 704 + ], + "type": "inline_equation", + "content": "7.48^{\\circ}" + }, + { + "bbox": [ + 56, + 381, + 294, + 704 + ], + "type": "text", + "content": " on the 7Scenes indoor dataset. These local-" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 381, + 553, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 381, + 553, + 619 + ], + "spans": [ + { + "bbox": [ + 313, + 381, + 553, + 619 + ], + "type": "text", + "content": "ization errors show competitive performance compared to leading RPR techniques, with the hypernet-enhanced baseline ranking third and second/fourth (7Scenes/Cambridge) in positional and rotational accuracy, respectively. Next, we compare the proposed MS-HyperPose architecture with other multiscene APR methods. Tables 2 and 3 present median position and orientation errors, respectively, on the Cambridge Landmarks and 7 Scenes datasets. On both indoor and outdoor benchmarks, our MS-HyperPose approach achieves the most accurate average of " + }, + { + "bbox": [ + 313, + 381, + 553, + 619 + ], + "type": "inline_equation", + "content": "1.27\\mathrm{m}" + }, + { + "bbox": [ + 313, + 381, + 553, + 619 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 381, + 553, + 619 + ], + "type": "inline_equation", + "content": "2.41^{\\circ}" + }, + { + "bbox": [ + 313, + 381, + 553, + 619 + ], + "type": "text", + "content": " on Cambridge and " + }, + { + "bbox": [ + 313, + 381, + 553, + 619 + ], + "type": "inline_equation", + "content": "0.18\\mathrm{m}" + }, + { + "bbox": [ + 313, + 381, + 553, + 619 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 381, + 553, + 619 + ], + "type": "inline_equation", + "content": "7.12^{\\circ}" + }, + { + "bbox": [ + 313, + 381, + 553, + 619 + ], + "type": "text", + "content": " on 7Scenes compared to other single- and multi-scene APRs. Additionally, MS-HyperPose is the most accurate for both positional and rotational errors in 6 out of 8 parameters on Cambridge and 6 out of 14 on 7 Scenes. The runtime of MS-HyperPose is " + }, + { + "bbox": [ + 313, + 381, + 553, + 619 + ], + "type": "inline_equation", + "content": "30.93\\mathrm{ms}" + }, + { + "bbox": [ + 313, + 381, + 553, + 619 + ], + "type": "text", + "content": ", similar to the reference architecture of [46]. Its model size is 571MB, notably larger than the 71.1MB of the original. We discuss storage considerations in the Limitation and Future Work section of the supplementary materials." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 629, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 553, + 712 + ], + "type": "text", + "content": "Table 4 reports the mean and median position and orientation errors of AtLoc [61] compared to our hypernetwork-enhanced variant using the Oxford RobotCar Dataset [30]. We also compare the enhanced model with [9], the SOTA APR for this dataset using image sequences. Visualizations of the localization results can be found in the supplementary materials." + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11553" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 101, + 561, + 448 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 3. Comparison to single-scene SOTA methods - 7Scenes: We report the median position/orientation error in meters/degrees. The most effective single and multi-scene APRs are distinguished in bold." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 101, + 561, + 448 + ], + "lines": [ + { + "bbox": [ + 56, + 101, + 561, + 448 + ], + "spans": [ + { + "bbox": [ + 56, + 101, + 561, + 448 + ], + "type": "table", + "html": "
ChessFireHeadsOfficePumpkinKitchenStairsAvg.
SCRDSAC* [6]0.02,1.11°0.02,1.24°0.01,1.82°0.03,1.18°0.04,1.41°0.03,1.70°0.04,1.42°0.03,1.41°
Marepo[15]0.02,1.24°0.02,1.39°0.02,2.03°0.03,1.26°0.04,1.48°0.04,1.71°0.06,1.67°0.03,1.54°
Seq.LsG [66]0.09,3.3°0.26,10.9°0.17,12.7°0.18,5.5°0.20,3.7°0.23,4.9°0.23,11.3°0.19,7.74°
MapNet[9]0.08,3.3°0.27,11.7°0.18,13.3°0.17,5.2°0.22,4.0°0.23,4.9°0.30,12.1°0.21,7.78°
GL-Net[67]0.08,2.8°0.26,8.9°0.17,11.4°0.18,13.3°0.15,2.8°0.25,4.5°0.23,8.8°0.19,7.50°
IRVLAD [56]0.21,12.5°0.33,13.8°0.15,14.9°0.28,11.2°0.31,11.2°0.30,11.3°0.25,12.3°0.26,12.46°
VLAD+Inter[42]0.18,10.0°0.33,12.4°0.14,14.3°0.25,10.1°0.26,9.4°0.27,11.1°0.24,14.7°0.24,11.71°
RPRNN-Net [26]0.13,6.5°0.26,12.7°0.14,12.3°0.21,7.4°0.24,6.4°0.24,8.0°0.27,11.8°0.21,9.30°
RelocNet[2]0.12,4.1°0.26,10.4°0.14,10.5°0.18,5.3°0.26,4.2°0.23,5.1°0.28,7.5°0.21,6.73°
EssNet [68]0.13,5.1°0.27,10.1°0.15,9.9°0.21,6.9°0.22,6.1°0.23,6.9°0.32,11.2°0.22,8.03°
NC-EssNet [68]0.12,5.6°0.26,9.6°0.14,10.7°0.20,6.7°0.22,5.7°0.22,6.3°0.31,7.9°0.21,7.50°
RelocGNN[21]0.08,2.7°0.21,7.5°0.13,8.70°0.15,4.1°0.15,3.5°0.19,3.7°0.22,6.5°0.16,5.24°
CamNet [17]0.04,1.7°0.03,1.7°0.05,2.00°0.04,1.60°0.04,1.6°0.04,1.6°0.04,1.5°0.04,1.70°
APRPoseNet [25]0.32,8.12°0.47,14.4°0.29,12.0°0.48,7.68°0.47,8.42°0.59,8.64°0.47,13.8°0.44,10.4°
BayesianPN [23]0.37,7.24°0.43,13.7°0.31,12.0°0.48,8.04°0.61,7.08°0.58,7.54°0.48,13.1°0.47,9.81°
LSTM-PN [60]0.24,5.77°0.34,11.9°0.21,13.7°0.30,8.08°0.33,7.00°0.37,8.83°0.40,13.7°0.31,9.85°
GPoseNet [11]0.20,7.11°0.38,12.3°0.21,13.8°0.28,8.83°0.37,6.94°0.35,8.15°0.37,12.5°0.31,9.95°
PoseNetLearn[24]0.14,4.50°0.27,11.8°0.18,12.1°0.20,5.77°0.25,4.82°0.24,5.52°0.37,10.6°0.24,7.87°
GeoPoseNet[24]0.13,4.48°0.27,11.3°0.17,13.0°0.19,5.55°0.26,4.75°0.23,5.35°0.35,12.4°0.23,8.12°
IRPNet[44]0.13,5.64°0.25,9.67°0.15,13.1°0.24,6.33°0.22,5.78°0.30,7.29°0.34,11.6°0.23,8.49°
AtLoc[61]0.10,4.07°0.25,11.4°0.16,11.8°0.17,5.34°0.21,4.37°0.23,5.42°0.26,10.5°0.20,7.56°
TransBoNet[49]0.11,4.48°0.25,12.5°0.18,14.0°0.20,5.08°0.19,4.77°0.17,5.35°0.30,13.0°0.20,8.45°
Baseline APR w/ hypernet0.09,4.47°0.28,9.44°0.15,11.4°0.16,6.29°0.19,4.62°0.18,6.95°0.28,9.15°0.19,7.48°
MS-APRMSPN[3]0.09,4.76°0.29,10.5°0.16,13.1°0.16,6.80°0.19,5.50°0.21,6.61°0.31,11.6°0.20,7.56°
MS-Trans[46]0.11,4.66°0.24,9.60°0.14,12.2°0.17,5.66°0.18,4.44°0.17,5.94°0.26,8.45°0.18,7.28°
MS-HyperPose0.11,4.34°0.23,9.79°0.13,10.7°0.17,6.05°0.16,5.24°0.17,6.86°0.27,6.00°0.18,7.00°
", + "image_path": "d82d080e52882a87b823d35f2dac963d51d7d16bb07acb42e874db3ff492a24a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 130, + 491, + 477, + 575 + ], + "blocks": [ + { + "bbox": [ + 55, + 459, + 555, + 483 + ], + "lines": [ + { + "bbox": [ + 55, + 459, + 555, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 459, + 555, + 483 + ], + "type": "text", + "content": "Table 4. Comparative analysis of different APR architectures with and without the proposed hypernetwork to the Oxford RobotCar dataset: We report the mean and median position/orientation error in meters/degrees." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 130, + 491, + 477, + 575 + ], + "lines": [ + { + "bbox": [ + 130, + 491, + 477, + 575 + ], + "spans": [ + { + "bbox": [ + 130, + 491, + 477, + 575 + ], + "type": "table", + "html": "
MapNet[9]AtLoc[61]AtLoc w/ hypernet
SequenceMeanMedianMeanMedianMeanMedian
LOOP18.76, 3.465.79, 1.548.61, 4.585.68, 2.238.42, 3.304.91, 1.33
LOOP29.84, 3.964.91, 1.678.61, 4.585.68, 2.238.10, 3.044.73, 1.27
FULL141.4, 12.517.94, 6.6829.6, 12.411.10, 5.2829.58, 6.777.86, 1.38
FULL259.3, 14.820.04, 6.3948.2, 11.112.2, 4.6345.96, 9.608.38, 1.45
", + "image_path": "44d625949ce4e96d03f742c0faea2034a559be6382b7e6c08495f1d5f1f0fd5b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 586, + 297, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 297, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 297, + 694 + ], + "type": "text", + "content": "We also assess the performance of two configurations of the baseline APR model, with and without hypernetworks on the Extended Cambridge dataset (ECL). Across all scenes and variations, the hypernetwork-enhanced model outperforms the baseline, demonstrating better adaptation to the seasonal changes and diverse lighting conditions inherent in the scenes. The detailed analysis elucidating the robustness and resilience of HyperPose to scene changes is in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 585, + 391, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 585, + 391, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 585, + 391, + 597 + ], + "type": "text", + "content": "6. Conclusions" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": "In this work, we proposed a hypernetwork-based approach to improve APR models. Our experiments show that hypernetworks improve accuracy for single-scene APR models. We also present MS-HyperPose, a novel multiscene camera pose regression model enhanced with a hypernetwork, which achieves new SOTA accuracy in indoor and outdoor localization datasets. We introduce the Extended Cambridge Landmarks (ECL) dataset to facilitate research on robust localization methods." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "11554" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "type": "text", + "content": "[1] Eduardo Arnold, Jamie Wynn, Sara Vicente, Guillermo Garcia-Hernando, Aron Monszpart, Victor Prisacariu, Daniyar Turmukhambetov, and Eric Brachmann. Map-free visual relocalization: Metric pose relative to a single image. In Proceedings of the European Conference on Computer Vision (ECCV), pages 690-708. Springer, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 158, + 296, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 158, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 61, + 158, + 296, + 201 + ], + "type": "text", + "content": "[2] Vassileios Balntas, Shuda Li, and Victor Prisacariu. Relocnet: Continuous metric learning relocalisation using neural nets. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 1, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 203, + 295, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 203, + 295, + 256 + ], + "spans": [ + { + "bbox": [ + 62, + 203, + 295, + 256 + ], + "type": "text", + "content": "[3] Hunter Blanton, Connor Greenwell, Scott Workman, and Nathan Jacobs. Extending absolute pose regression to multiple scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 38-39, 2020. 2, 3, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 258, + 295, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 258, + 295, + 301 + ], + "spans": [ + { + "bbox": [ + 62, + 258, + 295, + 301 + ], + "type": "text", + "content": "[4] E. Brachmann and C. Rother. Learning less is more - 6D camera localization via 3D surface regression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4654-4662, 2018. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 303, + 296, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 303, + 296, + 346 + ], + "spans": [ + { + "bbox": [ + 62, + 303, + 296, + 346 + ], + "type": "text", + "content": "[5] Eric Brachmann and Carsten Rother. Expert sample consensus applied to camera re-localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7525-7534, 2019." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 347, + 295, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 347, + 295, + 390 + ], + "spans": [ + { + "bbox": [ + 62, + 347, + 295, + 390 + ], + "type": "text", + "content": "[6] E. Brachmann and C. Rother. Visual Camera Re-Localization From RGB and RGB-D Images Using DSAC. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(09):5847-5865, 2022. 3, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 392, + 295, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 392, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 62, + 392, + 295, + 456 + ], + "type": "text", + "content": "[7] Eric Brachmann, Frank Michel, Alexander Krull, Michael Ying Yang, Stefan Gumhold, and Carsten Rother. Uncertainty-driven 6D pose estimation of objects and scenes from a single rgb image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3364-3372, 2016." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 459, + 295, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 459, + 295, + 523 + ], + "spans": [ + { + "bbox": [ + 62, + 459, + 295, + 523 + ], + "type": "text", + "content": "[8] E. Brachmann, A. Krull, S. Nowozin, J. Shotton, F. Michel, S. Gumhold, and C. Rother. DSAC - differentiable RANSAC for camera localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2492-2500, Los Alamitos, CA, USA, 2017. 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 525, + 295, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 525, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 62, + 525, + 295, + 578 + ], + "type": "text", + "content": "[9] Samarth Brahmbhatt, Jinwei Gu, Kihwan Kim, James Hays, and Jan Kautz. Geometry-aware learning of maps for camera localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3, 6, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 581, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 581, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 581, + 295, + 635 + ], + "type": "text", + "content": "[10] T. Brooks, A. Holynski, and A. A. Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18392-18402, Los Alamitos, CA, USA, 2023. IEEE Computer Society. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 636, + 295, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 295, + 678 + ], + "type": "text", + "content": "[11] Ming Cai, Chunhua Shen, and Ian Reid. A hybrid probabilistic model for camera relocalization. In Proceedings of the British Machine Vision Conference, page 238, 2018. 2, 3, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "text", + "content": "[12] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Proceedings" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 74, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "type": "text", + "content": "of the European Conference on Computer Vision (ECCV), pages 213-229, Cham, 2020. 3, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 97, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 97, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 317, + 97, + 553, + 149 + ], + "type": "text", + "content": "[13] Yoav Chai, Raja Giryes, and Lior Wolf. Supervised and unsupervised learning of parameterized color enhancement. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 992-1000, 2020. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 152, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 196 + ], + "type": "text", + "content": "[14] Kefan Chen, Noah Snavely, and Ameesh Makadia. Wide-baseline relative camera pose estimation with directional learning. 2021 IEEE. In CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3257-3267, 2021. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 198, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 198, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 317, + 198, + 553, + 251 + ], + "type": "text", + "content": "[15] Shuai Chen, Tommaso Cavallari, Victor Adrian Prisacariu, and Eric Brachmann. Map-relative pose regression for visual re-localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20665-20674, 2024. 1, 2, 3, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 254, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 254, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 317, + 254, + 553, + 297 + ], + "type": "text", + "content": "[16] Yinbo Chen and Xiaolong Wang. Transformers as meta-learners for implicit neural representations. In Proceedings of the European Conference on Computer Vision (ECCV), pages 170–187. Springer, 2022. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 299, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 299, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 317, + 299, + 553, + 342 + ], + "type": "text", + "content": "[17] Mingyu Ding, Zhe Wang, Jiankai Sun, Jianping Shi, and Ping Luo. CamNet: Coarse-to-fine retrieval for camera relocalization. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2019. 1, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 343, + 555, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 343, + 555, + 409 + ], + "spans": [ + { + "bbox": [ + 317, + 343, + 555, + 409 + ], + "type": "text", + "content": "[18] Siyan Dong, Qingnan Fan, He Wang, Ji Shi, Li Yi, Thomas Funkhouser, Baoquan Chen, and Leonidas J Guibas. Robust neural routing through space partitions for camera relocalization in dynamic indoor environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8544-8554, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 411, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 411, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 317, + 411, + 553, + 487 + ], + "type": "text", + "content": "[19] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In Proceedings of the International Conference on Learning Representations (ICLR), 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 488, + 553, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 488, + 553, + 532 + ], + "spans": [ + { + "bbox": [ + 317, + 488, + 553, + 532 + ], + "type": "text", + "content": "[20] Sovann En, Alexis Lecheryv, and Frédéric Jurie. Rpnet: An end-to-end network for relative camera pose estimation. In Proceedings of the European Conference on Computer Vision (ECCV) Workshops, pages 0–0, 2018. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 533, + 553, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 533, + 553, + 577 + ], + "spans": [ + { + "bbox": [ + 317, + 533, + 553, + 577 + ], + "type": "text", + "content": "[21] B. Glocker, S. Izadi, J. Shotton, and A. Criminisi. Realtime rgb-d camera relocalization. In 2013 IEEE International Symposium on Mixed and Augmented Reality (ISMAR), pages 173-179, 2013. 6, 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 578, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 578, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 317, + 578, + 553, + 600 + ], + "type": "text", + "content": "[22] David Ha, Andrew Dai, and Quoc V Le. Hypernetworks. arXiv preprint arXiv:1609.09106, 2016. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 601, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 601, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 317, + 601, + 553, + 645 + ], + "type": "text", + "content": "[23] Alex Kendall and Roberto Cipolla. Modelling uncertainty in deep learning for camera relocalization. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), pages 4762-4769, 2016. 2, 3, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 647, + 555, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 555, + 689 + ], + "type": "text", + "content": "[24] A. Kendall and R. Cipolla. Geometric loss functions for camera pose regression with deep learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6555-6564, 2017. 2, 3, 5, 7, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[25] A. Kendall, M. Grimes, and R. Cipolla. Posenet: A convolutional network for real-time 6-dof camera relocalization. In" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11555" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "text", + "content": "Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 2938-2946, 2015. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 295, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 295, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 295, + 149 + ], + "type": "text", + "content": "[26] Z. Laskar, I. Melekhov, S. Kalia, and J. Kannala. Camera relocalization by computing pairwise relative poses using convolutional neural network. In Proceedings of the IEEE International Conference on Computer Vision Workshops (ICCVW), pages 920–929, 2017. 1, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 295, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 295, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 295, + 205 + ], + "type": "text", + "content": "[27] Xiaotian Li, Shuzhe Wang, Yi Zhao, Jakob Verbeek, and Juho Kannala. Hierarchical scene coordinate classification and regression for visual localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11983-11992, 2020. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 205, + 294, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 205, + 294, + 248 + ], + "spans": [ + { + "bbox": [ + 56, + 205, + 294, + 248 + ], + "type": "text", + "content": "[28] Gidi Littwin and Lior Wolf. Deep meta functionals for shape representation. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 1824-1833, 2019. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 250, + 294, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 294, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 294, + 304 + ], + "type": "text", + "content": "[29] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 10012-10022, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 304, + 294, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 304, + 294, + 347 + ], + "spans": [ + { + "bbox": [ + 56, + 304, + 294, + 347 + ], + "type": "text", + "content": "[30] Will Maddern, Geoffrey Pascoe, Chris Linegar, and Paul Newman. 1 year, " + }, + { + "bbox": [ + 56, + 304, + 294, + 347 + ], + "type": "inline_equation", + "content": "1000\\mathrm{km}" + }, + { + "bbox": [ + 56, + 304, + 294, + 347 + ], + "type": "text", + "content": ": The oxford robotcar dataset. The International Journal of Robotics Research, 36(1):3-15, 2017. 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 349, + 255, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 349, + 255, + 360 + ], + "spans": [ + { + "bbox": [ + 56, + 349, + 255, + 360 + ], + "type": "text", + "content": "[31] Luke Melaskyriazi. Efficientnet-pytorch, 2019. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 361, + 295, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 295, + 414 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 295, + 414 + ], + "type": "text", + "content": "[32] Iaroslav Melekhov, Juha Ylioinas, Juho Kannala, and Esa Rahtu. Image-based localization using hourglass networks. In Proceedings of the IEEE International Conference on Computer Vision Workshops (ICCVW), pages 870-877, 2017. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 415, + 295, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 295, + 469 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 295, + 469 + ], + "type": "text", + "content": "[33] Tayyab Naseer and W. Burgard. Deep regression for monocular camera-based 6-dof global localization in outdoor environments. 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1525-1530, 2017. 2, 3, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 471, + 295, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 295, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 295, + 525 + ], + "type": "text", + "content": "[34] Yuval Nirkin, Lior Wolf, and Tal Hassner. Hyperseg: Patchwise hypernetwork for real-time semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4061-4070, 2021. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 525, + 295, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 525, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 56, + 525, + 295, + 568 + ], + "type": "text", + "content": "[35] N. Radwan, A. Valada, and W. Burgard. Vlocnet++: Deep multitask learning for semantic visual localization and odometry. IEEE Robotics and Automation Letters, 3(4): 4407-4414, 2018. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 570, + 295, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 295, + 602 + ], + "type": "text", + "content": "[36] Prajit Ramachandran, Barret Zoph, and Quoc V Le. Searching for activation functions. arXiv preprint arXiv:1710.05941, 2017. 3, 4, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 603, + 295, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 295, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 295, + 657 + ], + "type": "text", + "content": "[37] P. Sarlin, C. Cadena, R. Siegwart, and M. Dymczyk. From coarse to fine: Robust hierarchical localization at large scale. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12708-12717, 2019. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "type": "text", + "content": "[38] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4938-4947, 2020." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 149 + ], + "type": "text", + "content": "[39] Paul-Edouard Sarlin, Ajaykumar Unagar, Mans Larsson, Hugo Germain, Carl Toft, Viktor Larsson, Marc Pollefeys, Vincent Lepetit, Lars Hammarstrand, Fredrik Kahl, et al. Back to the feature: Learning robust camera localization from pixels to pose. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3247–3257, 2021." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 150, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 150, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 150, + 553, + 205 + ], + "type": "text", + "content": "[40] Torsten Sattler, Michal Havlena, Filip Radenovic, Konrad Schindler, and Marc Pollefeys. Hyperpoints and fine vocabularies for large-scale location recognition. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 2102-2110, 2015." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 205, + 553, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 205, + 553, + 248 + ], + "spans": [ + { + "bbox": [ + 316, + 205, + 553, + 248 + ], + "type": "text", + "content": "[41] T. Sattler, B. Leibe, and L. Kobbelt. Efficient effective prioritized matching for large-scale image-based localization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(9):1744-1756, 2017. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 250, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 250, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 250, + 553, + 304 + ], + "type": "text", + "content": "[42] T. Sattler, Q. Zhou, M. Pollefeys, and L. Leal-Taixe. Understanding the limitations of cnn-based absolute camera pose regression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3297-3307, 2019. 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 304, + 553, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 304, + 553, + 337 + ], + "spans": [ + { + "bbox": [ + 316, + 304, + 553, + 337 + ], + "type": "text", + "content": "[43] Yoli Shavit and Ron Ferens. Introduction to camera pose estimation with deep learning. In arXiv preprint arXiv:1907.05272, 2019. 2, 3, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 338, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 553, + 381 + ], + "type": "text", + "content": "[44] Yoli Shavit and Ron Ferens. Do we really need scene-specific pose encoders? In Proceedings of the International Conference on Pattern Recognition (ICPR), pages 3186-3192. IEEE, 2021. 2, 3, 5, 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 383, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 553, + 426 + ], + "type": "text", + "content": "[45] Yoli Shavit, Ron Ferens, and Yoshi Keller. Learning multiscene absolute pose regression with transformers. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 2713-2722, 2021. 2, 4, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 426, + 553, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 553, + 469 + ], + "type": "text", + "content": "[46] Yoli Shavit, Ron Ferens, and Yoshi Keller. Learning multiscene absolute pose regression with transformers. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. 2, 3, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 471, + 553, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 471, + 553, + 535 + ], + "spans": [ + { + "bbox": [ + 316, + 471, + 553, + 535 + ], + "type": "text", + "content": "[47] Jamie Shotton, Ben Glocker, Christopher Zach, Shahram Izadi, Antonio Criminisi, and Andrew Fitzgibbon. Scene coordinate regression forests for camera relocalization in rgb images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2013. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 536, + 553, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 536, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 316, + 536, + 553, + 590 + ], + "type": "text", + "content": "[48] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in Neural Information Processing Systems (NIPS), 33:7462-7473, 2020. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 592, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 592, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 592, + 553, + 635 + ], + "type": "text", + "content": "[49] Xiaogang Song, Hongjuan Li, Li Liang, Weiwei Shi, Guo Xie, Xiaofeng Lu, and Xinhong Hei. Transbonet: Learning camera localization with transformer bottleneck and attention. Pattern Recognition, 146:109975, 2024. 2, 3, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 636, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 553, + 690 + ], + "type": "text", + "content": "[50] Przemysław Spurek, Artur Kasymov, Marcin Mazur, Diana Janik, Slawomir K Tadeja, J Tabor, T Trzeciński, et al. Hyperpocket: Generative point cloud completion. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 6848-6853. IEEE, 2022. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[51] Hajime Taira, Masatoshi Okutomi, Torsten Sattler, Mircea Cimpoi, Marc Pollefeys, Josef Sivic, Tomas Pajdla, and Ak" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11556" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "type": "text", + "content": "ihiko Torii. Inloc: Indoor visual localization with dense matching and view synthesis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(4):1293-1307, 2021. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 296, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 296, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 296, + 151 + ], + "type": "text", + "content": "[52] Mingxing Tan and Quoc Le. EfficientNet: Rethinking model scaling for convolutional neural networks. pages 6105-6114, Long Beach, California, USA, 2019. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 296, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 296, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 296, + 196 + ], + "type": "text", + "content": "[53] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In Proceedings of the International Conference on Machine Learning (ICML), pages 6105-6114. PMLR, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 295, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 295, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 295, + 252 + ], + "type": "text", + "content": "[54] Shitao Tang, Sicong Tang, Andrea Tagliasacchi, Ping Tan, and Yasutaka Furukawa. Neumap: Neural coordinate mapping by auto-transdecoder for camera localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 929-939, 2023. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 295, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 295, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 295, + 297 + ], + "type": "text", + "content": "[55] Yi Tay, Zhe Zhao, Dara Bahri, Don Metzler, and Da-Cheng Juan. HyperGrid transformers: Towards a single model for multiple tasks. In Proceedings of the International Conference on Learning Representations (ICLR), 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 299, + 295, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 299, + 295, + 353 + ], + "spans": [ + { + "bbox": [ + 56, + 299, + 295, + 353 + ], + "type": "text", + "content": "[56] Akihiko Torii, Relja Arandjelović, Josef Sivic, Masatoshi Okutomi, and Tomas Pajdla. 24/7 place recognition by view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1808-1817, 2015. 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 355, + 295, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 355, + 295, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 355, + 295, + 409 + ], + "type": "text", + "content": "[57] Mehmet Ozgur Turkoglu, Eric Brachmann, Konrad Schindler, Gabriel J Brostow, and Aron Monszpart. Visual camera re-localization using graph neural networks and relative pose supervision. In 2021 International Conference on 3D Vision (3DV), pages 145-155. IEEE, 2021. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 411, + 295, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 295, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 295, + 464 + ], + "type": "text", + "content": "[58] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems (NIPS), pages 5998-6008, 2017. 2, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 468, + 295, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 468, + 295, + 499 + ], + "spans": [ + { + "bbox": [ + 56, + 468, + 295, + 499 + ], + "type": "text", + "content": "[59] Johannes Von Oswald, Christian Henning, João Sacramento, and Benjamin F Grewe. Continual learning with hypernetworks. arXiv preprint arXiv:1906.00695, 2019. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 501, + 295, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 295, + 555 + ], + "type": "text", + "content": "[60] F. Walch, C. Hazirbas, L. Leal-Taixe, T. Sattler, S. Hilsenbeck, and D. Cremers. Image-based localization using lstms for structured feature correlation. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 627-637, 2017. 2, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 557, + 295, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 295, + 610 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 295, + 610 + ], + "type": "text", + "content": "[61] Bing Wang, Changhao Chen, Chris Xiaoxuan Lu, Peijun Zhao, Niki Trigoni, and Andrew Markham. Atloc: Attention guided camera localization. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10393-10401, 2020. 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 613, + 295, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 295, + 667 + ], + "type": "text", + "content": "[62] Dominik Winkelbauer, Maximilian Denninger, and Rudolph Triebel. Learning to localize in new environments from synthetic training data. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 5840-5846. IEEE, 2021. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 295, + 713 + ], + "type": "text", + "content": "[63] J. Wu, L. Ma, and X. Hu. Delving deeper into convolutional neural networks for camera relocalization. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), pages 5644-5651, 2017. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 329 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "text", + "content": "[64] Tete Xiao, Mannat Singh, Eric Mintun, Trevor Darrell, Piotr Dollár, and Ross Girshick. Early convolutions help transformers see better. Advances in Neural Information Processing Systems (NIPS), 34:30392-30400, 2021. 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "type": "text", + "content": "[65] Meng Xu, Youchen Wang, Bin Xu, Jun Zhang, Jian Ren, Stefan Poslad, and Pengfei Xu. A critical analysis of image-based camera pose estimation techniques. arXiv preprint arXiv:2201.05816, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 163, + 555, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 555, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 555, + 217 + ], + "type": "text", + "content": "[66] Fei Xue, Xin Wang, Zike Yan, Qiuyuan Wang, Junqiu Wang, and Hongbin Zha. Local supports global: Deep camera relocalization with sequence enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 219, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 271 + ], + "type": "text", + "content": "[67] Fei Xue, Xin Wu, Shaojun Cai, and Junqiu Wang. Learning multi-view camera relocalization with graph neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 274, + 553, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 274, + 553, + 329 + ], + "spans": [ + { + "bbox": [ + 316, + 274, + 553, + 329 + ], + "type": "text", + "content": "[68] Qunjie Zhou, Torsten Sattler, Marc Pollefeys, and Laura Leal-Taixe. To learn or not to learn: Visual localization from essential matrices. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), pages 3319-3326. IEEE, 2020. 3, 7, 8" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11557" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/HyperSeg_ Hybrid Segmentation Assistant with Fine-grained Visual Perceiver/d3d1df0c-7574-4e1e-b447-81df1bf28440_content_list.json b/2025/HyperSeg_ Hybrid Segmentation Assistant with Fine-grained Visual Perceiver/d3d1df0c-7574-4e1e-b447-81df1bf28440_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..cef9d8cb298f053845954a661d8450d41a633b81 --- /dev/null +++ b/2025/HyperSeg_ Hybrid Segmentation Assistant with Fine-grained Visual Perceiver/d3d1df0c-7574-4e1e-b447-81df1bf28440_content_list.json @@ -0,0 +1,1535 @@ +[ + { + "type": "text", + "text": "HyperSeg: Hybrid Segmentation Assistant with Fine-grained Visual Perceiver", + "text_level": 1, + "bbox": [ + 106, + 130, + 890, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Cong Wei $^{1,2}$ , Yujie Zhong $^{2\\dagger}$ , Haoxian Tan $^{2}$ , Yong Liu $^{1}$ , Jie Hu $^{2}$ , Dengjie Li $^{2}$ , Zheng Zhao $^{2}$ , Yujiu Yang $^{1\\dagger}$ , Tsinghua Shenzhen International Graduate School, Tsinghua University ${}^{2}$ Meituan Inc.", + "bbox": [ + 99, + 179, + 924, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "weic22@mails.tsinghua.edu.cn, jaszhong@hotmail.com, yang.yujiu@sz.tsinghua.edu.cn", + "bbox": [ + 155, + 219, + 866, + 233 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9f5949157c434b915dc18761b6461f836a0be7ebae7519358b02fb1580c66fc7.jpg", + "image_caption": [ + "Figure 1. Illustration of our HyperSeg which can conduct image and video segmentation tasks with various language and visual instructions. Additionally, HyperSeg can handle complicated reasoning perception tasks compared with previous universal segmentation methods. To our knowledge, HyperSeg is the first VLLM-based universal segmentation model with perception and complex reasoning abilities in both image and video domains." + ], + "image_footnote": [], + "bbox": [ + 94, + 250, + 901, + 577 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 671, + 326, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper aims to address universal segmentation for image and video perception with the strong reasoning ability empowered by Visual Large Language Models (VLLMs). Despite significant progress in current unified segmentation methods, limitations in adaptation to both image and video scenarios, as well as the complex reasoning segmentation, make it difficult for them to handle various challenging instructions and achieve an accurate understanding of fine-grained vision-language correlations. We propose HyperSeg, the first VLLM-based universal segmentation model for pixel-level image and video perception, en", + "bbox": [ + 89, + 704, + 485, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "compassing generic segmentation tasks and more complex reasoning perception tasks requiring powerful reasoning abilities and world knowledge. Besides, to fully leverage the recognition capabilities of VLLMs and the fine-grained visual information, HyperSeg incorporates hybrid entity recognition and fine-grained visual perceiver modules for various segmentation tasks. Combined with the temporal adapter, HyperSeg achieves a comprehensive understanding of temporal information. Experimental results validate the effectiveness of our insights in resolving universal image and video segmentation tasks, including the more complex reasoning perception tasks. Our code is available at https://github.com/congvvc/HyperSeg.", + "bbox": [ + 511, + 672, + 908, + 869 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation.", + "bbox": [ + 236, + 0, + 808, + 18 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "Except for this watermark, it is identical to the accepted version;", + "bbox": [ + 323, + 17, + 722, + 30 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 292, + 31, + 754, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Corresponding authors.", + "bbox": [ + 114, + 886, + 246, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8931", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 89, + 222, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Visual segmentation is one of the most significant tasks in computer vision research, which aims to perform accurate pixel-level semantic understanding. Many specialist models [7, 17, 19, 22] have made great progress in specific segmentation tasks while showing limitations in handling diverse and complicated scenarios since new training data, paradigms, and model architectures are required to adapt to new segmentation tasks. Recent works [24, 25, 57] propose a single framework to unify diverse segmentation tasks. Despite promising, they show the inability to tackle text instructions and complex reasoning segmentation tasks needing powerful reasoning capabilities and world knowledge.", + "bbox": [ + 89, + 114, + 485, + 296 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Visual Large Language Models (VLLMs) have exhibited excellent reasoning and conversation abilities, which play a pivotal role in various vision-language co-understanding tasks [2, 8, 23, 29, 61]. However, these methods are based on rudimentary vision-language alignment, which limits their ability to comprehend finer details in visual perception tasks, like pixel-level segmentation. Recent studies [22, 45, 51, 58, 59] enables VLLMs to perform fine-grained visual understanding, like referring and reasoning segmentation. [22, 41, 45] use the special token [SEG] generated by VLLMs as the prompt for the mask decoder to generate segmentation masks while [58, 59] focus on incorporating instance-aware mask tokens into VLLMs. Though impressive, they show limitations to the universal segmentation framework based on VLLMs for both image and video domains and the capabilities of handling more complex video reasoning segmentation tasks.", + "bbox": [ + 89, + 296, + 485, + 551 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we introduce HyperSeg, the first VLLM-based universal segmentation model for pixel-level image and video perception with complex reasoning and conversation capabilities. HyperSeg can conduct diverse image and video segmentation tasks with various elaborate prompts and temporal adapter module. Besides, HyperSeg shows excellent abilities in complicated vision-language reasoning perception tasks needing rich world knowledge, which is significant for real-world understanding and interactions. As shown in Fig. 1, the explored tasks contain both image and video domains. We organize the tasks into two unified prompt formats: (1) text prompts (class names, reasoning questions, and referring languages), (2) visual prompts (box, mask, etc.). Owing to such flexible and cohesive design, HyperSeg benefits from concurrent training on diverse segmentation tasks and vision domains, facilitating the intricate correlations between different instructions and visual concepts. To further enhance fine-grained object perception and video understanding, we introduce three distinct designs in the following.", + "bbox": [ + 89, + 553, + 485, + 854 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Firstly, we incorporate a hybrid entity recognition strategy to enhance the exploitation of VLLM's recognition capacity. Generation-only works [22, 41, 49] solely rely on VLLM for", + "bbox": [ + 89, + 854, + 485, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "object prediction leading to poor performance in complex multi-object segmentation scenarios. Decode-only methods [58, 59] use the prompt embedding and mask tokens decoded by VLLM to obtain class scores for each mask, which makes the mask tokens interact insufficiently with the semantic condition as they ignore the powerful generative capabilities of VLLM. The proposed hybrid entity recognition leverages the VLLM's powerful generative abilities to enhance the mask tokens' comprehension of category semantics while maintaining the final class scores decoding process.", + "bbox": [ + 511, + 90, + 908, + 258 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Secondly, previous VLLMs usually use coarse-level visual features obtained from CLIP [38] series which primarily encode global visual information while overlooking visual details. To enhance VLLMs' ability of capturing visual details efficiently, we use the Fine-grained Visual Perceiver (FVP) to merge multi-scale visual features into fixed-length fine-grained tokens, allowing retrieval of rich visual details from various scales in the hierarchical vision encoder [7].", + "bbox": [ + 511, + 267, + 908, + 388 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Thirdly, recent VLLM-based segmentation methods [22, 58, 59] demonstrate limitations in handling video perception tasks for video temporal understanding. To this end, we propose the temporal adapter for comprehensive video perception which incorporates global prompt aggregation and local space-time information injection for the coalescence of both long-term and short-term vision-language information.", + "bbox": [ + 511, + 398, + 908, + 505 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Extensive experiments on various segmentation benchmarks demonstrate the preeminent segmentation ability of HyperSeg, providing strong evidence of the effectiveness of our insights. Our HyperSeg also exhibits promising performance on common Multi-modal benchmarks. Additionally, we explore the mutual influence among different tasks involving various visual and task types.", + "bbox": [ + 511, + 515, + 910, + 621 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are summarized as follows:", + "bbox": [ + 532, + 631, + 836, + 645 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present HyperSeg, the first VLLM-based universal segmentation model for pixel-level image and video perception, covering a broad spectrum of common segmentation tasks, complex reasoning, and conversation-based vision-language understanding tasks.", + "- We incorporate hybrid entity recognition and fine-grained visual perceiver modules to VLLM, which allow full exploitation of VLLM's semantic recognition capacity and injection of fine-grained visual information to improve diverse detail-aware segmentation tasks. With the temporal adapter, HyperSeg can conduct more challenging video perception tasks, achieving universal segmentation.", + "- HyperSeg demonstrates superior capabilities on multiple segmentation tasks, achieving excellent performance on both generic and complex reasoning benchmarks with only one model." + ], + "bbox": [ + 522, + 656, + 906, + 897 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8932", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 91, + 89, + 232, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Visual Large Language Model. The emergence of Large Language Model (LLM) has significantly contributed to the development of VLMM. In this context, LLMs are enhanced with multimodal comprehension capabilities, allowing the vision-language co-understanding [1, 2, 23, 28, 29, 61]. Several notable examples of LLMs with multimodal comprehension include BLIP-2 [23], Flamingo [1], MiniGPT-4 [61], LLaVA [29], InstructBLIP [10], and Qwen-VL [2]. While these models have demonstrated impressive performance in vision-language tasks, they solely produce textual outputs that describe the entire image. This restricts their applicability in tasks that require the pixel-level detailed understanding.", + "bbox": [ + 89, + 114, + 485, + 311 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Perception with VLLM. Several methods have been proposed to enhance VLLMs with a more detailed comprehension capability [5, 22, 36, 37, 39, 41, 43, 54]. Shikra [5], Ferret [54], Kosmos-2 [36], and VisionLLM [43] are examples that provide grounding capabilities through regression of box coordinates. Conversely, LISA [22], PixelLM [41], GLaMM [39], and PerceptionGPT [37] employ a mask decoder to predict object masks from special tokens. Most of the existing methods utilize a next-token-prediction approach, which restricts their applicability. PSALM [59] makes an important attempt to bring VLLM into visual perception tasks but fails to fully unleash the potential of VLLM. In contrast, our method proposes to use a hybrid strategy to mitigate this problem and keep the capacity in high-level reasoning.", + "bbox": [ + 89, + 311, + 485, + 537 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unified segmentation model. Another line of studies focuses on the integration of various segmentation tasks into a single model. Mask2former [7] proposes a unified architecture that requires separate training on different segmentation tasks. OpenSeeD [57] introduces a text encoder and extends it to the Open-Set setting. Simultaneously, UNINEXT [25] supports referring segmentation with the assistance of text inputs and a text encoder. However, these works fall short of following complicated instructions and reasoning. In this work, we improve the understanding ability toward language by incorporating LLM, while also maintaining the original ability of vision-centric models.", + "bbox": [ + 89, + 537, + 485, + 719 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 731, + 181, + 746 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 89, + 756, + 200, + 771 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overall architecture. The architecture of HyperSeg is illustrated in Fig. 2, which consists of a fine-grained pyramid visual encoder, a light-weight VLLM, and a segmentation predictor to generate segmentation masks, class scores, and instance embedding for video correspondence according to user's instruction. The proposed FVP module fuses multiscale high-resolution visual features $f_{img}$ into a set of fine-grained tokens to ensure the injection of fine-grained visual", + "bbox": [ + 89, + 779, + 485, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "information (Sec 3.3). The VLLM takes three types of inputs: visual tokens encoded by the CLIP encoder, renewed fine-grained tokens, and prompt tokens for diverse instructions. The output embeddings of semantically enhanced mask tokens (Sec 3.2) and prompt tokens are further fed into the segmentation predictor for final segmentation results. Besides, we utilize the space-time information propagation and global prompt aggregation for comprehensive video understanding (Sec 3.4). We train the LLM with LoRA for efficient parameter tuning.", + "bbox": [ + 511, + 90, + 906, + 241 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Visual Large Language Model. We take a light-weight VLLM as our powerful multi-modal feature encoder, which contains a low-resolution vision encoder like CLIP [38] and an efficient LLM.", + "bbox": [ + 511, + 242, + 906, + 301 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, the model takes vision-prompt pairs $\\{(\\mathcal{V},\\mathcal{P})\\}$ as inputs, where $\\mathcal{V}$ is resized to low resolution and then encoded by CLIP encoder $F_{CLIP}$ to get image features $f_{v}$ . The $f_{v}$ is further projected and concatenated with other task-specific tokens to ensure the comprehensive understanding of multi-modal inputs through the fusion process of LLM $F_{LLM}$ , where $G_{c}$ is the projection function and $E_O$ denotes the output embeddings of LLM. Formally,", + "bbox": [ + 511, + 301, + 908, + 425 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf _ {v} = F _ {C L I P} (\\mathcal {V}), E _ {O} = F _ {L L M} \\left(G _ {c} \\left(f _ {v}\\right), P, \\mathcal {P}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 436, + 906, + 453 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $P$ denotes fine-grained tokens. Furthermore, we manually extract semantic enhanced mask tokens $E_{Q}$ and prompt embedding $E_{P}$ from $E_{O}$ , which are further fed into the pretrained segmentation predictor [7] to generate masks, class scores, and instance embedding for final segmentation results.", + "bbox": [ + 511, + 463, + 908, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Prompt design. In order to accommodate the different segmentation tasks, we propose a flexible design for prompt $\\mathcal{P}$ . As illustrated above, we divide $\\mathcal{P}$ into two formats: text prompts and visual prompts. To be specific, $\\mathcal{P}$ contains the instructions $\\mathcal{P}_{\\mathcal{I}}$ and task-specific conditions $\\mathcal{P}_{\\mathcal{C}}$ , where $\\mathcal{P}_{\\mathcal{I}}$ instructs the model to perform different tasks while $\\mathcal{P}_{\\mathcal{C}}$ indicates diverse conditions which are further used as classifiers to calculate the class scores of predicted masks.", + "bbox": [ + 511, + 555, + 908, + 675 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For class-based segmentation tasks like panoptic segmentation, open-vocabulary segmentation (OVS), and video instance segmentation (VIS), $\\mathcal{P}$ can be demonstrated as $\\mathcal{P}_{\\mathcal{I}}$ : \"Please segment all the positive objects according to the following potential categories.\" $\\mathcal{P}_{\\mathcal{C}}$ : \"[category 1, category 2, category 3, ...]", + "bbox": [ + 511, + 676, + 908, + 766 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For referring and reasoning segmentation tasks like referring expression segmentation (RES), reasoning segmentation, referring video object segmentation (R-VOS), and ReasonVOS, $\\mathcal{P}$ can be designed as $\\mathcal{P}_{\\mathcal{I}}$ : \"Can you perform referring or reasoning segmentation according to the language expression?\" $\\mathcal{P}_{\\mathcal{C}}$ : \"[referring / reasoning text]\"", + "bbox": [ + 511, + 765, + 908, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For visual-guided segmentation tasks like interactive segmentation and video object segmentation (VOS), $\\mathcal{P}$ can be designed as $\\mathcal{P}_{\\mathcal{I}}$ : \"Please segment according to the given", + "bbox": [ + 511, + 854, + 908, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8933", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/643707cb3f024ee261eafd8b603adbc81ae0f501316c55fda3c10e361e527f08.jpg", + "image_caption": [ + "Figure 2. Overview of HyperSeg. HyperSeg encodes the visual input in a multi-grained manner and concatenates the prompt for different perception tasks. We feed learnable fine-grained tokens into a Fine-grained Visual Perceiver (FVP) to integrate multi-scale high-resolution image features into LLM for detailed visual learning and to facilitate space-time information propagation for video understanding. Additionally, we use the semantically enhanced mask tokens and prompt embedding to finally generate the segmentation masks and class scores for generic segmentation, and instance embedding for video instance association." + ], + "image_footnote": [], + "bbox": [ + 106, + 88, + 890, + 369 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "visual region reference\" $\\mathcal{P}_{\\mathcal{C}}$ : \"[vision 1, vision 2, vision 3, ...]\". Instead of using an additional region encoder to extract visual reference features [25], we sample the CLIP visual features $f_{v}$ in VLLM according to the region coordinates and perform adaptive average pooling on them to form the final reference features for each visual prompt.", + "bbox": [ + 88, + 465, + 482, + 556 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Segmentation predictor. Segmentation predictor $F_{p}$ generates the masks $m$ , corresponding class scores $z$ , and instance embedding $e$ through the similar process [7, 15] of three inputs: task-specific prompt embedding $\\{E_{\\mathcal{P}}^{k}\\}_{k=1}^{K}$ , the semantically enhanced mask tokens $\\{E_{Q}^{j}\\}_{j=1}^{N}$ and the multi-scale visual features $f_{img}$ , where $K$ and $N$ denote $K$ categories and $N$ mask proposals. Formally,", + "bbox": [ + 88, + 556, + 482, + 664 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{m _ {j}, z _ {j}, e _ {j} \\right\\} _ {j = 1} ^ {N} = F _ {p} \\left(\\left\\{E _ {\\mathcal {P}} ^ {k} \\right\\} _ {k = 1} ^ {K}, \\left\\{E _ {\\mathcal {Q}} ^ {j} \\right\\} _ {j = 1} ^ {N}, f _ {i m g}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 672, + 482, + 694 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $m_j \\in \\mathbb{R}^{H \\times W}$ is the j-th mask proposal, $z_j \\in \\mathbb{R}^K$ denotes the class scores of $m_j$ , and $e_j \\in \\mathbb{R}^D$ denotes the j-th instance embedding obtained from an extra embedding head only for video domain. For video tasks, we adopt a frame-by-frame manner to get frame-level segmentation results for efficient training and inference processes.", + "bbox": [ + 88, + 702, + 482, + 795 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training objectives. The model can be trained jointly on multiple tasks using the unified loss $\\mathcal{L}$ . Specifically, we employ an autoregressive cross-entropy loss $\\mathcal{L}_{text}$ for text prediction, a combination of per-pixel binary cross-entropy loss $\\mathcal{L}_{bce}$ and DICE loss $\\mathcal{L}_{dice}$ for mask supervision $\\mathcal{L}_{mask}$ , a cross-entropy loss $\\mathcal{L}_{cls}$ for category classification, and a contrastive loss $\\mathcal{L}_{ins}$ for instance association of video", + "bbox": [ + 88, + 795, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "sequences following [48]. $\\lambda$ indicates their sum weight respectively. Formally,", + "bbox": [ + 511, + 465, + 906, + 497 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {t e x t}} + \\lambda_ {\\text {m a s k}} \\mathcal {L} _ {\\text {m a s k}} + \\lambda_ {\\text {c l s}} \\mathcal {L} _ {\\text {c l s}} + \\lambda_ {\\text {i n s}} \\mathcal {L} _ {\\text {i n s}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 508, + 906, + 525 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {m a s k}} = \\lambda_ {b c e} \\mathcal {L} _ {b c e} + \\lambda_ {d i c e} \\mathcal {L} _ {d i c e}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 529, + 906, + 545 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Differences between HyperSeg and previous methods. Previous universal segmentation methods [24, 25, 31] lacking of VLLMs show inability in reasoning perception tasks while our HyperSeg demonstrates brilliant reasoning segmentation capability in complex scenarios. Besides, we make a significant generalization of the current VLLM-based segmentation methods [22, 51, 58, 59] for more diverse segmentation tasks in both image and video domains using a single model framework. Moreover, HyperSeg differs from previous methods in the three designs elaborated in the following sections.", + "bbox": [ + 511, + 550, + 908, + 717 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Hybrid Entity Recognition", + "text_level": 1, + "bbox": [ + 511, + 724, + 756, + 742 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig. 3 (a), predicting presented objects in the way of sequence generation (semantic prediction) tends to miss objects or produce repetitive predictions [45]. On the other hand, Fig. 3 (b), only using VLLM to embed class names (prompt tokens) as mask classifier at the decode stage disregards VLLM's powerful semantic recognition capability. Consequently, we propose a hybrid approach that leverages LLM in both generation and decoding processes.", + "bbox": [ + 511, + 750, + 906, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Instead of integrating mask tokens in input sequences and extracting the corresponding embedding from the one-pass", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8934", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/616d3b0e74640482eaf15a74f78b6ab8a220081ac36f42acbfb171dd6c63b94c.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 106, + 88, + 326, + 223 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9237b9284212ed2d8c78097eb968b9f8e7e7166d4be46a6b96974a00c064f644.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 341, + 87, + 630, + 223 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2bdcbae716cfdc5eeb53754d3febe9800e8d3fd80091a509c0208f98438ab8d6.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 635, + 88, + 887, + 223 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/cca5f1eb172bf49413a0bbe6f89b7437e763d28cffb3eaadeee2aa3505fdc45f.jpg", + "image_caption": [ + "Figure 3. The comparison of different recognition strategies. (a) Generation-Only [22, 41]: both the semantic recognition (existing objects) and their mask tokens are generated by LLM. (b) Decode-Only [58, 59]: prompt embedding and mask tokens are decoded from LLM. The present objects are then determined by their similarity scores. (c) Hybrid (ours): prompt embedding is decoded from LLM while the semantically enhanced mask tokens are generated by LLM. Their similarity scores reflect the objects' presence.", + "(a)" + ], + "image_footnote": [], + "bbox": [ + 101, + 304, + 276, + 459 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5455b62a85ea2b1d9e7aac7be0d7b0e4465f1a672d63a1cd8fe80ba8b0508f73.jpg", + "image_caption": [ + "(b)", + "Figure 4. Comparison between previous vision perceiver and our FVP. (a): previous vision perceiver [2, 23] uses the coarse single-scale CLIP visual features which are inadequate for fine-grained perception tasks. (b): FVP encodes the multi-scale visual features into fine-grained tokens." + ], + "image_footnote": [], + "bbox": [ + 279, + 304, + 468, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "forward output of VLLM, we instruct VLLM to generate the mask tokens preceded by the estimated objects' names. As illustrated in Fig. 3 (c), VLLM is compelled to generate all the existing objects in the vision input and then the mask tokens. The semantically enhanced mask tokens contain valuable semantic integrated information about the image, which are subsequently used as input for the segmentation predictor to generate segmentation masks.", + "bbox": [ + 89, + 563, + 483, + 685 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Fine-grained Visual Perceiver", + "text_level": 1, + "bbox": [ + 89, + 695, + 356, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Why twin-tower vision encoder? As shown in Fig. 4, previous VLLMs and VLLM-based segmentation methods usually utilize the pre-trained CLIP encoder to obtain single-scale and low-resolution vision features interacted with diverse languages, which is insufficient for fine-grained image and video segmentation tasks. Therefore, we adopt an extra pyramid vision encoder [7] to inject details-aware visual information.", + "bbox": [ + 89, + 719, + 483, + 837 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, we fuse multi-scale visual features into fine-grained tokens (stated as $P$ in Sec 3.1) which can inject rich fine-grained visual information into the pre-trained VLLMs without excessive computation cost. Formally, given the", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "vision input $\\nu$ , we leverage a pyramid vision encoder [7] $F_{seg}$ to get details-aware image features $f_{img}$ . For the j-th scale and the previous fine-grained tokens $P_{j-1}$ , the FVP module enriches each token through conditional weighted cross-attention:", + "bbox": [ + 511, + 309, + 906, + 383 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {P} _ {j} = \\operatorname {M H C A} \\left(P _ {j - 1}, G _ {p} \\left(f _ {\\text {i m g}} ^ {(j)}\\right)\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 397, + 906, + 419 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP _ {j} = P _ {j - 1} + \\tanh (\\mathrm {M L P} (\\hat {P} _ {j})) \\cdot \\hat {P} _ {j}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 421, + 906, + 440 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where MHCA denotes the Multi-Head Cross-Attention layer, $G_{p}$ is the projection function, tanh is a normalization function and MLP is a Multilayer Perceptron. The component of $\\tanh(\\mathrm{MLP}(\\hat{P}_j))$ is the conditional weight used to multiply the enriched fine-grained tokens $\\hat{P}_j$ before the residual connection to the previous tokens $P_{j - 1}$ . Additionally, we initialize the weight value to zero to ensure the adaptation to diverse multi-scale image features while retaining the training stability.", + "bbox": [ + 511, + 446, + 908, + 583 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Temporal Adapter", + "text_level": 1, + "bbox": [ + 511, + 593, + 692, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Video segmentation entails distinct challenges, requiring reasoning across multiple frames and the maintenance of temporal coherence. Existing VLLM-based methods exhibit limitations in addressing video perception tasks and lack specialized designs for comprehending temporal dynamics in video analysis. To this end, we utilize global prompt aggregation and local space-time information injection in the time dimension to adapt to more complicated video perception tasks.", + "bbox": [ + 511, + 617, + 908, + 751 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Global prompt aggregation. For the current prompt embedding $E_{\\mathcal{P}}$ in the video object mask retrieval process, we leverage the adaptive average pooling strategy along the time dimension to aggregate global object and temporal information of previous $T$ frames.", + "bbox": [ + 511, + 753, + 908, + 829 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nE _ {\\mathcal {P}} = \\operatorname {A v g P o o l} ([ E _ {\\mathcal {P}} ^ {0}, E _ {\\mathcal {P}} ^ {1}, \\dots , E _ {\\mathcal {P}} ^ {T} ]), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 839, + 906, + 859 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Local space-time information injection. We propose a sequential renewal strategy for space-time information prop", + "bbox": [ + 511, + 869, + 908, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8935", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/6340c8cac50ddbc094880be494fde280b0501a971ae7d9c2c0ae71930afdf96d.jpg", + "table_caption": [ + "Table 1. Comparison with the state-of-the-art models on the RefCOCO series and more challenging generalized referring expression segmentation benchmark gRefCOCO. $\\ddagger$ denotes models using pre-trained SAM [21] for mask generation. * means using gRefCOCO for training while other methods are evaluated in zero-shot manners." + ], + "table_footnote": [], + "table_body": "
TypeMethodRefCOCORefCOCO+RefCOCOggRefCOCO
valtestAtestBvaltestAtestBval(U)test(U)valtestAtestB
SegmentationVLT [11]67.570.565.256.361.050.155.057.752.5*62.2*50.5*
CRIS [44]70.573.266.162.368.153.759.960.455.3*63.8*51.0*
LAVT [53]72.775.868.862.168.455.161.262.157.6*65.3*55.0*
PolyFormer-B [30]74.876.671.167.672.959.367.869.1---
VLLM-basedLISA-7B [22]‡74.979.172.365.170.858.167.970.638.7*52.6*44.8*
PixelLM-7B [41]73.076.568.266.371.758.369.370.5---
GSVA-7B [49]‡76.477.472.864.567.758.671.172.061.7*69.2*60.3*
GroundHog-7B [33]78.579.975.770.575.064.974.174.666.7*--
SAM4MLLM-7B [6]‡79.682.876.173.577.865.874.575.666.3*70.1*63.2*
Segmentation NetworkLaSagnA-7B [45]‡76.878.773.866.470.660.170.671.938.150.442.1
OMG-LLaVA [58]78.080.374.169.173.163.072.972.9---
GLaMM [40]‡79.583.276.972.678.764.674.274.9---
PSALM [59]83.684.781.672.975.570.173.874.442.052.450.6
HyperSeg84.885.783.479.083.575.279.478.947.557.352.5
", + "bbox": [ + 98, + 142, + 901, + 366 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7679cb9200bbf36b1c70a776f3de6d2bcaeee4d4ceb23a6eb592e3abfd33f29e.jpg", + "table_caption": [ + "Table 2. Comparison with the state-of-the-art models on more complex and challenging reasoning segmentation benchmarks: ReVOS in video domain and ReasonSeg in image domain. $\\ddagger$ denotes the same meaning as Tab. 1." + ], + "table_footnote": [], + "table_body": "
MethodBackboneReVOS-ReasoningReVOS-ReferringReVOS-OverallReasonSeg
JFJ&FJFJ&FJFJ&FgIoUcIoU
LMPM [12]Swin-T13.324.318.829.039.134.121.231.726.4--
ReferFormer [47]Video-Swin-B21.325.623.431.234.332.726.229.928.1--
LISA-7B [22]‡ViT-H33.838.436.144.347.145.739.142.740.952.954.0
LaSagnA-7B [45]‡ViT-H---------48.847.2
SAM4MLLM-7B [6]‡EfficientViT-SAM-XL1---------46.748.1
TrackGPT-13B [62]‡ViT-H38.142.940.548.350.649.543.246.845.0--
VISA-7B [51]‡ViT-H36.741.739.251.154.752.943.948.246.152.757.8
VISA-13B [51]‡ViT-H38.343.540.952.355.854.145.349.747.5--
HyperSeg-3BSwin-B50.255.853.056.060.958.553.158.455.759.256.7
", + "bbox": [ + 94, + 417, + 903, + 584 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "agation based on fine-grained tokens $P$ to inject object information of adjacent frames. Formally,", + "bbox": [ + 89, + 609, + 485, + 640 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nP _ {t} = G _ {l} \\left[ F _ {L L M} \\left(P _ {t - 1}\\right) \\right], \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 651, + 483, + 670 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $P_{t}$ denotes the time-aware fine-grained tokens of the current $t$ -th frame, $G_{l}$ is the projection function to transfer the previous features to the current space and align the feature dimensions.", + "bbox": [ + 89, + 679, + 483, + 739 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The proposed global prompt aggregation and local spacetime information injection within our temporal adapter facilitate the coalescence of both long-term and short-term vision-language information, which is essential for comprehensive video perception. More details about the temporal adapter are provided in the Supplementary Material.", + "bbox": [ + 89, + 739, + 483, + 832 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 844, + 223, + 862 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We use the one-stage training strategy to train HyperSeg with the multi-dataset and multi-task manners.", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For image segmentation, we use COCO Panoptic [26], RefCOCO series [35, 55], COCO-Interactive, and ReasonSeg [22]. For video segmentation, we utilize DAVIS-2017 datasets [4], Ref-Youtu-VOS [42], YouTube-VIS 2019 [52], and ReVOS [51]. Besides, we use LLAVA-150k [29] to maintain the vision-language conversation capability of VLLM (we show the results on Multi-modal benchmarks in the Supplementary Material).", + "bbox": [ + 511, + 609, + 906, + 731 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details We load the pre-trained weights of Mipha [63] for our VLLM, and Maks2Former [7] for our segmentation predictor. We use three layers of FVP for fine-grained information fusion and utilize LoRA [18] to finetune the LLM efficiently. We train HyperSeg jointly for 160k iterations using a batch size of 32 on 8 NVIDIA A100 GPUs, which means each task takes approximately 16k iterations. We employ the AdamW optimizer with a learning rate of $4 \\times 10^{-5}$ and with a cosine schedule. All the hyper-parameters in the loss $\\mathcal{L}$ are assigned values 1.0.", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8936", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2ab7ef9cb76d3447467b9955fa1e6e4f51571cd37d9174a9e9ca5dc5e52099a9.jpg", + "table_caption": [ + "Table 3. Quantitative results on the closed-set COCO-Panoptic segmentation, open-vocabulary segmentation (-OV) benchmarks." + ], + "table_footnote": [], + "table_body": "
TypeMethodBackboneCOCO-PanopticADE-OVCitys-OVPC59-OVPAS20-OV
PQmIoUPQmIoUPQmIoUmIoU
Segmentation SpecialistMask2former [7]Swin-B55.165.1-----
OneFormer [19]Swin-L57.967.4-----
SEEM [64]DaViT-B56.166.3-----
MaskCLIP [13]ViT-L30.947.615.123.7-45.9-
DeOP [16]ResNet-101c---22.9-48.891.7
SimBaseline [50]ViT-B---20.5-47.788.4
DaTaSeg [15]ViTDet-B52.862.712.318.328.051.1-
VLLM-based Segmentation NetworkOMG-LLaVA [58]ConvNeXt-L53.8------
PSALM [22]Swin-B55.966.613.718.228.848.581.3
HyperSegSwin-B61.277.216.122.331.164.692.1
", + "bbox": [ + 122, + 114, + 870, + 282 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/592873928381c821b545721056b58bf41bef6420aa6a6c76091ecb4863370519.jpg", + "table_caption": [ + "Table 4. Results of common video segmentation benchmarks, including DAVIS17, Ref-YouTube-VOS, Ref-DAVIS17, and YouTube-VIS 2019. $\\ddagger$ denotes the same meaning as Tab. 1." + ], + "table_footnote": [], + "table_body": "
MethodBackboneDAVIS17Ref-YTRef-DAVISYT-VIS
J&FJ&FJ&FmAP
SEEM [64]DaViT-B62.8---
OMG-Seg [24]ConvNeXt-L74.3--56.4
ReferFormer [47]Video-Swin-B-62.961.1-
OnlineRefer [46]Swin-L-63.564.8-
UNINEXT [25]ConvNeXt-L77.266.266.764.3
LISA-7B [22]‡ViT-H-53.964.8-
VISA-13B [51]‡ViT-H-63.070.4-
VideoLISA-3.8B [3]‡ViT-H-63.768.8-
HyperSeg-3BSwin-B77.668.571.253.8
", + "bbox": [ + 91, + 357, + 485, + 481 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1. Comparisons with State-of-the-Arts", + "text_level": 1, + "bbox": [ + 89, + 496, + 403, + 512 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Referring expression segmentation results. We compare HyperSeg with the state-of-the-art methods on the benchmarks RefCOCO+/g [35, 55] and more challenging generalized referring expression segmentation benchmark gRefCOCO [27]. in Tab. 1. Based on the versatile and adaptable design of HyperSeg, our model achieves state-of-the-art performance on all the referring datasets. Specifically, HyperSeg surpasses the current SOTA by a large margin, reaching 79.7 cIoU on RefCOCO+ val (+6.8 over PSALM). Besides, Our model shows superiority in challenging G-RES tasks compared with previous zero-shot methods, demonstrating the robustness and generalization ability of HyperSeg.", + "bbox": [ + 88, + 521, + 485, + 702 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Reasoning segmentation results. We compare HyperSeg with the state-of-the-art methods on image reasoning segmentation (ReasonSeg [22]) and reasoning video object segmentation (ReVOS [51]) in Tab. 2. Our HyperSeg achieves superior performance on reasoning tasks, significantly surpassing previous state-of-the-art methods (+12.1 on ReVOS-Reasoning), which shows HyperSeg powerful reasoning capability of tackling complex scenarios.", + "bbox": [ + 88, + 703, + 485, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Generic image segmentation results. We show the performance of HyperSeg on COCO-Panoptic [26] and open-vocabulary segmentation [9, 14, 34, 60] tasks in Tab. 3. HyperSeg achieves excellent performance compared with both specialist models and VLLMs-based methods on both closed-", + "bbox": [ + 88, + 825, + 485, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "set and open-vocabulary segmentation tasks. Specifically, HyperSeg surpasses the VLLM-based PSALM by a significant margin (+5.3 on COCO PQ, and +10.6 on mIoU), which demonstrates our powerful capabilities of handling complex semantic perception and segmentation tasks. Besides, we show the results of COCO-Interactive in the Supplementary Material.", + "bbox": [ + 511, + 306, + 906, + 411 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Common video segmentation results. We compare HyperSeg with previous video segmentation methods in Tab. 4, including visual-prompted semi-supervised VOS (DAVIS17 val), text-prompted referring video object segmentation (RefYouTube-VOS, Ref-DAVIS17) and video instance segmentation (YouTube-VIS 2019). HyperSeg shows promising results over previous unified segmentation methods [24, 25]. Besides, HyperSeg performs more video perception tasks than previous VLLM-based models [3, 51].", + "bbox": [ + 511, + 412, + 908, + 549 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Ablations", + "text_level": 1, + "bbox": [ + 511, + 559, + 625, + 574 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The mutual influence between different tasks. Our model can be trained and inferred across multiple tasks and datasets simultaneously. We evaluate the mutual impact of different tasks in Tab. 5. The results show that joint training can enhance the model performance compared with the task-specific model. Besides, the performance of video segmentation tasks can be improved significantly by adding the image training datasets. This demonstrates the generalization and self-consistency of our HyperSeg to perform universal segmentation.", + "bbox": [ + 511, + 582, + 908, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of different LLMs and vision backbone. In Tab. 6, we evaluate the effect of different sizes of LLMs and vision backbone. Our HyperSeg achieves excellent performance using smaller LLMs and vision encoder compared with the previous SOTA models like VISA[51] and PSALM[59]. Besides, the performance of HyperSeg can be further improved by using the more powerful LLM (Phi-2-2.7B [20]).", + "bbox": [ + 511, + 734, + 908, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation on the proposed components. We assess the effectiveness of our proposed FVP module and Hybrid Entity Recognition strategy. As shown in Tab. 7, with our fine-grained visual integration and hybrid entity semantic", + "bbox": [ + 511, + 840, + 908, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8937", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/267b39c75ec973fde10018d0e1820eeacc28392968fdea987684bf1230587559.jpg", + "table_caption": [ + "Table 5. The mutual influence between different tasks. Task-specific means training task-specific models only on data from corresponding tasks, Refer+Reason denotes the model is trained on referring and reasoning segmentation data." + ], + "table_footnote": [], + "table_body": "
Task-specificRefer+ReasonVideoImageRefCOCOCOCOReVOSYT-VIS
valtestAtestBPQmIoUReasoningReferringOverallmAP
✓83.885.982.260.875.151.256.653.950.7
✓83.384.980.9--53.157.355.2-
✓85.686.182.460.976.5----
✓-----51.157.054.150.4
✓✓✓84.885.783.461.277.253.058.555.753.8
", + "bbox": [ + 140, + 128, + 857, + 231 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4fd2b4991d8c7b0b097d12f298291f1acbd7f3ee878dbfd8e60f5b711b1bdb27.jpg", + "table_caption": [ + "Table 6. The comparison of different LLMs and backbone usages. w/o CLIP means without using CLIP vision encoder." + ], + "table_footnote": [], + "table_body": "
MethodLLMCOCOReVOSADE-OVPC59-OVPAS20-OV
PQmIoUReasoningReferringOverallmIoUmIoUmIoU
LISA [22]Vicuna-7B--36.145.740.9---
VISA [51]Vicuna-13B--40.954.147.5---
PSALM(w/o CLIP) [22]Phi-1.5-1.3B55.966.6---18.248.581.3
HyperSeg (w/o CLIP)Phi-1.5-1.3B61.176.044.049.746.918.960.090.6
HyperSegPhi-1.5-1.3B60.976.750.857.053.920.361.590.8
HyperSegPhi-2-2.7B61.277.253.058.555.722.364.692.1
", + "bbox": [ + 138, + 268, + 859, + 386 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/baa1662006fdb3c6f6b26ad0695145f4592c20af2922285393a9acf526358f22.jpg", + "table_caption": [ + "Table 7. Ablation on the core components of HyperSeg. FVP and HER denote the proposed Fine-grained Visual Perceiver and Hybrid Entity Recognition modules." + ], + "table_footnote": [], + "table_body": "
FVPHERYT-VISCOCORefCOCO
mAPPQmIoUcIoU
✓48.454.866.282.8
50.855.866.684.6
✓✓52.059.774.684.3
✓✓53.861.277.284.8
", + "bbox": [ + 120, + 454, + 455, + 554 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/de1d8bafefe6bb1bdb91bc7f87d0c1f2b331a37f3c4b03fc894066a7bbba6cb7.jpg", + "table_caption": [ + "Table 8. Ablation on the Fine-grained Visual Perceiver design. CW denotes the Conditional Weight illustrated in Sec. 3.3, and Scale denotes the total scale in the proposed FVP module." + ], + "table_footnote": [], + "table_body": "
CWScaleYT-VISCOCORefCOCO
mAPPQmIoUcIoU
✓single-layer49.755.868.083.7
multi-layers50.458.973.484.5
multi-layers53.861.277.284.8
", + "bbox": [ + 106, + 621, + 464, + 702 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "enhancement, the segmentation accuracy can be enhanced significantly $(+5.4$ on YT-VIS, $+6.4$ on COCO panoptic PQ).", + "bbox": [ + 89, + 718, + 482, + 747 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Design of the Fine-grained Visual Perceiver. In the FVP module, we combine multi-scale visual features into fixed perception queries using the condition-wise cross-attention layers to extract rich visual details from different scales of the pyramid encoder. As shown in Tab. 8, together with the conditional weight and the multi-scale design, our model makes a significant improvement on both image and video segmentation tasks.", + "bbox": [ + 89, + 750, + 482, + 868 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of temporal adapter. We evaluate the effectiveness of the proposed temporal adapter including global prompt ag", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f05bdb66d27944b97d205086827a7d51f7bb0781fa4001b7cf513df4d02f49f3.jpg", + "table_caption": [ + "Table 9. Ablation on the temporal adapter for video tasks." + ], + "table_footnote": [], + "table_body": "
GlobalLocalRef-DAVIS17ReVOSYT-VIS
J&FJ&FmAP
✓67.354.147.9
68.854.548.5
✓69.354.850.2
✓✓71.255.753.8
", + "bbox": [ + 540, + 426, + 879, + 525 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "gregation (global) and local space-time information injection (local) in Tab. 9. Incorporating both global and local components, the temporal adapter significantly enhances model performance across multiple video segmentation tasks.", + "bbox": [ + 511, + 546, + 906, + 608 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 630, + 633, + 645 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this study, we aim to present HyperSeg, the first VLLM-based universal segmentation model designed for pixel-level image and video perception, encompassing a wide range of generic segmentation and complex reasoning tasks. We propose the Hybrid Entity Recognition and Fine-grained Visual Perceiver to leverage the recognition capacity of VLLMs more effectively and enhances the VLLM's ability by capturing diverse levels of visual information without incurring excessive computational costs. With additional Temporal Adapter, HyperSeg can tackle challenging video tasks by incorporating global and local information. HyperSeg surpasses existing methods on complex reasoning segmentation and traditional perception tasks. The insights presented in this work expand the possibilities of VLLMs in visual perception and lay a foundation for future research on the integration of vision-language models.", + "bbox": [ + 511, + 657, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8938", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgments", + "text_level": 1, + "bbox": [ + 89, + 90, + 269, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was partly supported by the National Key Research and Development Program of China (No. 2024YFB2808903) and the Shenzhen Science and Technology Program (JCYJ20220818101001004).", + "bbox": [ + 89, + 114, + 485, + 176 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 191, + 187, + 208 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 3", + "[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 2, 3, 5, 1", + "[3] Zechen Bai, Tong He, Haiyang Mei, Pichao Wang, Ziteng Gao, Joya Chen, Lei Liu, Zheng Zhang, and Mike Zheng Shou. One token to seg them all: Language instructed reasoning segmentation in videos. arXiv preprint arXiv:2409.19603, 2024. 7", + "[4] Sergi Caelles, Alberto Montes, Kevis-Kokitsi Maninis, Yuhua Chen, Luc Van Gool, Federico Perazzi, and Jordi Pont-Tuset. The 2018 davis challenge on video object segmentation. arXiv preprint arXiv:1803.00557, 2018. 6", + "[5] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 3, 1", + "[6] Yi-Chia Chen, Wei-Hua Li, Cheng Sun, Yu-Chiang Frank Wang, and Chu-Song Chen. Sam4mllm: Enhance multimodal large language model for referring expression segmentation. In European Conference on Computer Vision, pages 323-340. Springer, 2025. 6", + "[7] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 2, 3, 4, 5, 6, 7", + "[8] Jaemin Cho, Jie Lei, Hao Tan, and Mohit Bansal. Unifying vision-and-language tasks via text generation. In International Conference on Machine Learning, pages 1931-1942. PMLR, 2021. 2", + "[9] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 7", + "[10] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning, 2023. 3, 1" + ], + "bbox": [ + 93, + 217, + 485, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[11] Henghui Ding, Chang Liu, Suchen Wang, and Xudong Jiang. Vision-language transformer and query generation for referring segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16321-16330, 2021. 6, 3", + "[12] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, and Chen Change Loy. Mevis: A large-scale benchmark for video segmentation with motion expressions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2694-2703, 2023. 6", + "[13] Zheng Ding, Jieke Wang, and Zhuowen Tu. Open-vocabulary universal image segmentation with maskclip. arXiv preprint arXiv:2208.08984, 2022. 7", + "[14] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-338, 2010. 7", + "[15] Xiuye Gu, Yin Cui, Jonathan Huang, Abdullah Rashwan, Xuan Yang, Xingyi Zhou, Golnaz Ghiasi, Weicheng Kuo, Huizhong Chen, Liang-Chieh Chen, et al. Dataseg: Taming a universal multi-dataset multi-task segmentation model. Advances in Neural Information Processing Systems, 36, 2024. 4, 7", + "[16] Cong Han, Yujie Zhong, Dengjie Li, Kai Han, and Lin Ma. Open-vocabulary semantic segmentation with decoupled one-pass network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1086-1096, 2023. 7", + "[17] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 2", + "[18] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 6", + "[19] Jitesh Jain, Jiachen Li, Mang Tik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. Oneformer: One transformer to rule universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2989-2998, 2023. 2, 7, 3", + "[20] Mojan Javaheripi, Sébastien Bubeck, Marah Abdin, Jyoti Aneja, Sebastien Bubeck, Caio César Teodoro Mendes, Weizhu Chen, Allie Del Giorno, Ronen Eldan, Sivakanth Gopi, et al. Phi-2: The surprising power of small language models. *Microsoft Research Blog*, 1:3, 2023. 7, 1", + "[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 6, 1, 2", + "[22] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. ArXiv, abs/2308.00692, 2023. 2, 3, 4, 5, 6, 7, 8", + "[23] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 2, 3, 5, 1" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8939", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Xiangtai Li, Haobo Yuan, Wei Li, Henghui Ding, Size Wu, Wenwei Zhang, Yining Li, Kai Chen, and Chen Change Loy. Omg-seg: Is one model good enough for all segmentation? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27948-27959, 2024. 2, 4, 7, 3", + "[25] Fangjian Lin, Jianlong Yuan, Sitong Wu, Fan Wang, and Zhibin Wang. Uninext: Exploring a unified architecture for vision recognition. In Proceedings of the 31st ACM International Conference on Multimedia, pages 3200-3208, 2023. 2, 3, 4, 7", + "[26] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision, 2014. 6, 7", + "[27] Chang Liu, Henghui Ding, and Xudong Jiang. Gres: Generalized referring expression segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23592-23601, 2023. 7", + "[28] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 3", + "[29] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2, 3, 6, 1", + "[30] Jiang Liu, Hui Ding, Zhaowei Cai, Yuting Zhang, Ravi Kumar Satzoda, Vijay Mahadevan, and R Manmatha. Polyformer: Referring image segmentation as sequential polygon generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18653-18663, 2023. 6, 3", + "[31] Yong Liu, Cairong Zhang, Yitong Wang, Jiahao Wang, Yujiu Yang, and Yansong Tang. Universal segmentation at arbitrary granularity with language instruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3459-3469, 2024. 4", + "[32] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 1", + "[33] Bo Miao, Mohammed Bennamoun, Yongsheng Gao, and Ajmal Mian. Spectrum-guided multi-granularity referring video object segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 920-930, 2023. 6", + "[34] Roozbeh Mottaghi, Xianjie Chen, Xiaobai Liu, Nam-Gyu Cho, Seong-Whan Lee, Sanja Fidler, Raquel Urtasun, and Alan Yuille. The role of context for object detection and semantic segmentation in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 891–898, 2014. 7", + "[35] Varun K Nagaraja, Vlad I Morariu, and Larry S Davis. Modeling context between objects for referring expression understanding. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 792-807. Springer, 2016. 6, 7" + ], + "bbox": [ + 91, + 90, + 485, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[36] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 3", + "[37] Renjie Pi, Lewei Yao, Jiahui Gao, Jipeng Zhang, and Tong Zhang. Perceptiongpt: Effectively fusing visual perception into llm. arXiv preprint arXiv:2311.06612, 2023. 3", + "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3", + "[39] Hanoona Rasheed, Muhammad Maaz, Sahal Shaji, Abdelrahman Shaker, Salman Khan, Hisham Cholakkal, Rao M Anwer, Erix Xing, Ming-Hsuan Yang, and Fahad S Khan. Glamm: Pixel grounding large multimodal model. arXiv preprint arXiv:2311.03356, 2023. 3", + "[40] Hanoona Rasheed, Muhammad Maaz, Sahal Shaji, Abdelrahman Shaker, Salman Khan, Hisham Cholakkal, Rao M Anwer, Eric Xing, Ming-Hsuan Yang, and Fahad S Khan. Glamm: Pixel grounding large multimodal model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13009-13018, 2024. 6", + "[41] Zhongwei Ren, Zhicheng Huang, Yunchao Wei, Yao Zhao, Dongmei Fu, Jiashi Feng, and Xiaojie Jin. Pixel reasoning with large multimodal model. ArXiv, abs/2312.02228, 2023. 2, 3, 5, 6", + "[42] Seonguk Seo, Joon-Young Lee, and Bohyung Han. Urvos: Unified referring video object segmentation network with a large-scale benchmark. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XV 16, pages 208–223. Springer, 2020. 6", + "[43] Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong Lu, Jie Zhou, Yu Qiao, et al. Vision: Large language model is also an open-ended decoder for vision-centric tasks. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[44] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clipdriven referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11686-11695, 2022. 6", + "[45] Cong Wei, Haoxian Tan, Yujie Zhong, Yujiu Yang, and Lin Ma. Lasagna: Language-based segmentation assistant for complex queries. arXiv preprint arXiv:2404.08506, 2024. 2, 4, 6, 3", + "[46] Dongming Wu, Tiancai Wang, Yang Zhang, Xiangyu Zhang, and Jianbing Shen. Onlinerefer: A simple online baseline for referring video object segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2761-2770, 2023. 7, 3", + "[47] Jiannan Wu, Yi Jiang, Peize Sun, Zehuan Yuan, and Ping Luo. Language as queries for referring video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4974-4984, 2022. 6, 7, 3" + ], + "bbox": [ + 516, + 90, + 906, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8940", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[48] Junfeng Wu, Qihao Liu, Yi Jiang, Song Bai, Alan Yuille, and Xiang Bai. In defense of online models for video instance segmentation. In ECCV, 2022. 4", + "[49] Zhuofan Xia, Dongchen Han, Yizeng Han, Xuran Pan, Shiji Song, and Gao Huang. Gsva: Generalized segmentation via multimodal large language models. arXiv preprint arXiv:2312.10103, 2023. 2, 6, 3", + "[50] Mengde Xu, Zheng Zhang, Fangyun Wei, Yutong Lin, Yue Cao, Han Hu, and Xiang Bai. A simple baseline for open-vocabulary semantic segmentation with pre-trained vision-language model. In European Conference on Computer Vision, pages 736-753. Springer, 2022. 7", + "[51] Cilin Yan, Haochen Wang, Shilin Yan, Xiaolong Jiang, Yao Hu, Guoliang Kang, Weidi Xie, and Efstratios Gavves. Visa: Reasoning video object segmentation via large language models. arXiv preprint arXiv:2407.11325, 2024. 2, 4, 6, 7, 8, 3", + "[52] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5188-5197, 2019. 6", + "[53] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. Levt: Language-aware vision transformer for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18155-18165, 2022. 6, 3", + "[54] Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and Yinfei Yang. Ferret: Refer and ground anything anywhere at any granularity. arXiv preprint arXiv:2310.07704, 2023. 3", + "[55] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 69-85. Springer, 2016. 6, 7", + "[56] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 1", + "[57] Hao Zhang, Feng Li, Xueyan Zou, Shilong Liu, Chunyuan Li, Jianwei Yang, and Lei Zhang. A simple framework for open-vocabulary segmentation and detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1020-1031, 2023. 2, 3", + "[58] Tao Zhang, Xiangtai Li, Hao Fei, Haobo Yuan, Shengqiong Wu, Shunping Ji, Chen Change Loy, and Shuicheng Yan. Omg-llava: Bridging image-level, object-level, pixel-level reasoning and understanding. arXiv preprint arXiv:2406.19389, 2024. 2, 4, 5, 6, 7, 3", + "[59] Zheng Zhang, Yeyao Ma, Enming Zhang, and Xiang Bai. Psalm: Pixelwise segmentation with large multi-modal model. arXiv preprint arXiv:2403.14598, 2024. 2, 3, 4, 5, 6, 7, 1", + "[60] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision, 127(3):302-321, 2019. 7", + "[61] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language" + ], + "bbox": [ + 89, + 92, + 485, + 901 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023.2, 3", + "[62] Jiawen Zhu, Zhi-Qi Cheng, Jun-Yan He, Chenyang Li, Bin Luo, Huchuan Lu, Yifeng Geng, and Xuansong Xie. Tracking with human-intent reasoning. arXiv preprint arXiv:2312.17448, 2023. 6", + "[63] Minjie Zhu, Yichen Zhu, Xin Liu, Ning Liu, Zhiyuan Xu, Chaomin Shen, Yaxin Peng, Zhicai Ou, Feifei Feng, and Jian Tang. A comprehensive overhaul of multimodal assistant with small language models. arXiv preprint arXiv:2403.06199, 2024. 6", + "[64] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Wang, Lijuan Wang, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. Advances in Neural Information Processing Systems, 36, 2024. 7, 1, 2, 3" + ], + "bbox": [ + 516, + 92, + 906, + 303 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "8941", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/HyperSeg_ Hybrid Segmentation Assistant with Fine-grained Visual Perceiver/d3d1df0c-7574-4e1e-b447-81df1bf28440_model.json b/2025/HyperSeg_ Hybrid Segmentation Assistant with Fine-grained Visual Perceiver/d3d1df0c-7574-4e1e-b447-81df1bf28440_model.json new file mode 100644 index 0000000000000000000000000000000000000000..73d06a20da10b0a63f8be3d6acb75c4795c96c0b --- /dev/null +++ b/2025/HyperSeg_ Hybrid Segmentation Assistant with Fine-grained Visual Perceiver/d3d1df0c-7574-4e1e-b447-81df1bf28440_model.json @@ -0,0 +1,2323 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.809, + 0.02 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + }, + { + "type": "header", + "bbox": [ + 0.325, + 0.018, + 0.723, + 0.031 + ], + "angle": 0, + "content": "Except for this watermark, it is identical to the accepted version;" + }, + { + "type": "header", + "bbox": [ + 0.294, + 0.032, + 0.755, + 0.047 + ], + "angle": 0, + "content": "the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.107, + 0.131, + 0.892, + 0.154 + ], + "angle": 0, + "content": "HyperSeg: Hybrid Segmentation Assistant with Fine-grained Visual Perceiver" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.18, + 0.926, + 0.217 + ], + "angle": 0, + "content": "Cong Wei\\(^{1,2}\\), Yujie Zhong\\(^{2\\dagger}\\), Haoxian Tan\\(^{2}\\), Yong Liu\\(^{1}\\), Jie Hu\\(^{2}\\), Dengjie Li\\(^{2}\\), Zheng Zhao\\(^{2}\\), Yujiu Yang\\(^{1\\dagger}\\), Tsinghua Shenzhen International Graduate School, Tsinghua University \\({}^{2}\\)Meituan Inc." + }, + { + "type": "text", + "bbox": [ + 0.156, + 0.22, + 0.867, + 0.234 + ], + "angle": 0, + "content": "weic22@mails.tsinghua.edu.cn, jaszhong@hotmail.com, yang.yujiu@sz.tsinghua.edu.cn" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.251, + 0.903, + 0.578 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.594, + 0.908, + 0.649 + ], + "angle": 0, + "content": "Figure 1. Illustration of our HyperSeg which can conduct image and video segmentation tasks with various language and visual instructions. Additionally, HyperSeg can handle complicated reasoning perception tasks compared with previous universal segmentation methods. To our knowledge, HyperSeg is the first VLLM-based universal segmentation model with perception and complex reasoning abilities in both image and video domains." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.672, + 0.327, + 0.687 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.486, + 0.871 + ], + "angle": 0, + "content": "This paper aims to address universal segmentation for image and video perception with the strong reasoning ability empowered by Visual Large Language Models (VLLMs). Despite significant progress in current unified segmentation methods, limitations in adaptation to both image and video scenarios, as well as the complex reasoning segmentation, make it difficult for them to handle various challenging instructions and achieve an accurate understanding of fine-grained vision-language correlations. We propose HyperSeg, the first VLLM-based universal segmentation model for pixel-level image and video perception, en" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.673, + 0.909, + 0.87 + ], + "angle": 0, + "content": "compassing generic segmentation tasks and more complex reasoning perception tasks requiring powerful reasoning abilities and world knowledge. Besides, to fully leverage the recognition capabilities of VLLMs and the fine-grained visual information, HyperSeg incorporates hybrid entity recognition and fine-grained visual perceiver modules for various segmentation tasks. Combined with the temporal adapter, HyperSeg achieves a comprehensive understanding of temporal information. Experimental results validate the effectiveness of our insights in resolving universal image and video segmentation tasks, including the more complex reasoning perception tasks. Our code is available at https://github.com/congvvc/HyperSeg." + }, + { + "type": "page_footnote", + "bbox": [ + 0.116, + 0.887, + 0.248, + 0.9 + ], + "angle": 0, + "content": "†Corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8931" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.223, + 0.107 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.486, + 0.297 + ], + "angle": 0, + "content": "Visual segmentation is one of the most significant tasks in computer vision research, which aims to perform accurate pixel-level semantic understanding. Many specialist models [7, 17, 19, 22] have made great progress in specific segmentation tasks while showing limitations in handling diverse and complicated scenarios since new training data, paradigms, and model architectures are required to adapt to new segmentation tasks. Recent works [24, 25, 57] propose a single framework to unify diverse segmentation tasks. Despite promising, they show the inability to tackle text instructions and complex reasoning segmentation tasks needing powerful reasoning capabilities and world knowledge." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.297, + 0.486, + 0.553 + ], + "angle": 0, + "content": "Visual Large Language Models (VLLMs) have exhibited excellent reasoning and conversation abilities, which play a pivotal role in various vision-language co-understanding tasks [2, 8, 23, 29, 61]. However, these methods are based on rudimentary vision-language alignment, which limits their ability to comprehend finer details in visual perception tasks, like pixel-level segmentation. Recent studies [22, 45, 51, 58, 59] enables VLLMs to perform fine-grained visual understanding, like referring and reasoning segmentation. [22, 41, 45] use the special token [SEG] generated by VLLMs as the prompt for the mask decoder to generate segmentation masks while [58, 59] focus on incorporating instance-aware mask tokens into VLLMs. Though impressive, they show limitations to the universal segmentation framework based on VLLMs for both image and video domains and the capabilities of handling more complex video reasoning segmentation tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.554, + 0.486, + 0.856 + ], + "angle": 0, + "content": "To this end, we introduce HyperSeg, the first VLLM-based universal segmentation model for pixel-level image and video perception with complex reasoning and conversation capabilities. HyperSeg can conduct diverse image and video segmentation tasks with various elaborate prompts and temporal adapter module. Besides, HyperSeg shows excellent abilities in complicated vision-language reasoning perception tasks needing rich world knowledge, which is significant for real-world understanding and interactions. As shown in Fig. 1, the explored tasks contain both image and video domains. We organize the tasks into two unified prompt formats: (1) text prompts (class names, reasoning questions, and referring languages), (2) visual prompts (box, mask, etc.). Owing to such flexible and cohesive design, HyperSeg benefits from concurrent training on diverse segmentation tasks and vision domains, facilitating the intricate correlations between different instructions and visual concepts. To further enhance fine-grained object perception and video understanding, we introduce three distinct designs in the following." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.486, + 0.901 + ], + "angle": 0, + "content": "Firstly, we incorporate a hybrid entity recognition strategy to enhance the exploitation of VLLM's recognition capacity. Generation-only works [22, 41, 49] solely rely on VLLM for" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.909, + 0.259 + ], + "angle": 0, + "content": "object prediction leading to poor performance in complex multi-object segmentation scenarios. Decode-only methods [58, 59] use the prompt embedding and mask tokens decoded by VLLM to obtain class scores for each mask, which makes the mask tokens interact insufficiently with the semantic condition as they ignore the powerful generative capabilities of VLLM. The proposed hybrid entity recognition leverages the VLLM's powerful generative abilities to enhance the mask tokens' comprehension of category semantics while maintaining the final class scores decoding process." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.268, + 0.909, + 0.39 + ], + "angle": 0, + "content": "Secondly, previous VLLMs usually use coarse-level visual features obtained from CLIP [38] series which primarily encode global visual information while overlooking visual details. To enhance VLLMs' ability of capturing visual details efficiently, we use the Fine-grained Visual Perceiver (FVP) to merge multi-scale visual features into fixed-length fine-grained tokens, allowing retrieval of rich visual details from various scales in the hierarchical vision encoder [7]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.399, + 0.909, + 0.506 + ], + "angle": 0, + "content": "Thirdly, recent VLLM-based segmentation methods [22, 58, 59] demonstrate limitations in handling video perception tasks for video temporal understanding. To this end, we propose the temporal adapter for comprehensive video perception which incorporates global prompt aggregation and local space-time information injection for the coalescence of both long-term and short-term vision-language information." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.516, + 0.911, + 0.622 + ], + "angle": 0, + "content": "Extensive experiments on various segmentation benchmarks demonstrate the preeminent segmentation ability of HyperSeg, providing strong evidence of the effectiveness of our insights. Our HyperSeg also exhibits promising performance on common Multi-modal benchmarks. Additionally, we explore the mutual influence among different tasks involving various visual and task types." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.632, + 0.837, + 0.646 + ], + "angle": 0, + "content": "Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.524, + 0.657, + 0.907, + 0.732 + ], + "angle": 0, + "content": "- We present HyperSeg, the first VLLM-based universal segmentation model for pixel-level image and video perception, covering a broad spectrum of common segmentation tasks, complex reasoning, and conversation-based vision-language understanding tasks." + }, + { + "type": "text", + "bbox": [ + 0.524, + 0.733, + 0.907, + 0.837 + ], + "angle": 0, + "content": "- We incorporate hybrid entity recognition and fine-grained visual perceiver modules to VLLM, which allow full exploitation of VLLM's semantic recognition capacity and injection of fine-grained visual information to improve diverse detail-aware segmentation tasks. With the temporal adapter, HyperSeg can conduct more challenging video perception tasks, achieving universal segmentation." + }, + { + "type": "text", + "bbox": [ + 0.524, + 0.838, + 0.905, + 0.898 + ], + "angle": 0, + "content": "- HyperSeg demonstrates superior capabilities on multiple segmentation tasks, achieving excellent performance on both generic and complex reasoning benchmarks with only one model." + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.657, + 0.907, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8932" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.09, + 0.233, + 0.106 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.486, + 0.312 + ], + "angle": 0, + "content": "Visual Large Language Model. The emergence of Large Language Model (LLM) has significantly contributed to the development of VLMM. In this context, LLMs are enhanced with multimodal comprehension capabilities, allowing the vision-language co-understanding [1, 2, 23, 28, 29, 61]. Several notable examples of LLMs with multimodal comprehension include BLIP-2 [23], Flamingo [1], MiniGPT-4 [61], LLaVA [29], InstructBLIP [10], and Qwen-VL [2]. While these models have demonstrated impressive performance in vision-language tasks, they solely produce textual outputs that describe the entire image. This restricts their applicability in tasks that require the pixel-level detailed understanding." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.312, + 0.486, + 0.538 + ], + "angle": 0, + "content": "Perception with VLLM. Several methods have been proposed to enhance VLLMs with a more detailed comprehension capability [5, 22, 36, 37, 39, 41, 43, 54]. Shikra [5], Ferret [54], Kosmos-2 [36], and VisionLLM [43] are examples that provide grounding capabilities through regression of box coordinates. Conversely, LISA [22], PixelLM [41], GLaMM [39], and PerceptionGPT [37] employ a mask decoder to predict object masks from special tokens. Most of the existing methods utilize a next-token-prediction approach, which restricts their applicability. PSALM [59] makes an important attempt to bring VLLM into visual perception tasks but fails to fully unleash the potential of VLLM. In contrast, our method proposes to use a hybrid strategy to mitigate this problem and keep the capacity in high-level reasoning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.486, + 0.72 + ], + "angle": 0, + "content": "Unified segmentation model. Another line of studies focuses on the integration of various segmentation tasks into a single model. Mask2former [7] proposes a unified architecture that requires separate training on different segmentation tasks. OpenSeeD [57] introduces a text encoder and extends it to the Open-Set setting. Simultaneously, UNINEXT [25] supports referring segmentation with the assistance of text inputs and a text encoder. However, these works fall short of following complicated instructions and reasoning. In this work, we improve the understanding ability toward language by incorporating LLM, while also maintaining the original ability of vision-centric models." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.732, + 0.183, + 0.747 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.757, + 0.201, + 0.772 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Overall architecture. The architecture of HyperSeg is illustrated in Fig. 2, which consists of a fine-grained pyramid visual encoder, a light-weight VLLM, and a segmentation predictor to generate segmentation masks, class scores, and instance embedding for video correspondence according to user's instruction. The proposed FVP module fuses multiscale high-resolution visual features \\(f_{img}\\) into a set of fine-grained tokens to ensure the injection of fine-grained visual" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.242 + ], + "angle": 0, + "content": "information (Sec 3.3). The VLLM takes three types of inputs: visual tokens encoded by the CLIP encoder, renewed fine-grained tokens, and prompt tokens for diverse instructions. The output embeddings of semantically enhanced mask tokens (Sec 3.2) and prompt tokens are further fed into the segmentation predictor for final segmentation results. Besides, we utilize the space-time information propagation and global prompt aggregation for comprehensive video understanding (Sec 3.4). We train the LLM with LoRA for efficient parameter tuning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.243, + 0.908, + 0.302 + ], + "angle": 0, + "content": "Visual Large Language Model. We take a light-weight VLLM as our powerful multi-modal feature encoder, which contains a low-resolution vision encoder like CLIP [38] and an efficient LLM." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.303, + 0.909, + 0.426 + ], + "angle": 0, + "content": "Specifically, the model takes vision-prompt pairs \\(\\{(\\mathcal{V},\\mathcal{P})\\}\\) as inputs, where \\(\\mathcal{V}\\) is resized to low resolution and then encoded by CLIP encoder \\(F_{CLIP}\\) to get image features \\(f_{v}\\). The \\(f_{v}\\) is further projected and concatenated with other task-specific tokens to ensure the comprehensive understanding of multi-modal inputs through the fusion process of LLM \\(F_{LLM}\\), where \\(G_{c}\\) is the projection function and \\(E_O\\) denotes the output embeddings of LLM. Formally," + }, + { + "type": "equation", + "bbox": [ + 0.554, + 0.437, + 0.908, + 0.454 + ], + "angle": 0, + "content": "\\[\nf _ {v} = F _ {C L I P} (\\mathcal {V}), E _ {O} = F _ {L L M} \\left(G _ {c} \\left(f _ {v}\\right), P, \\mathcal {P}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.464, + 0.909, + 0.553 + ], + "angle": 0, + "content": "where \\( P \\) denotes fine-grained tokens. Furthermore, we manually extract semantic enhanced mask tokens \\( E_{Q} \\) and prompt embedding \\( E_{P} \\) from \\( E_{O} \\), which are further fed into the pretrained segmentation predictor [7] to generate masks, class scores, and instance embedding for final segmentation results." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.556, + 0.909, + 0.676 + ], + "angle": 0, + "content": "Prompt design. In order to accommodate the different segmentation tasks, we propose a flexible design for prompt \\(\\mathcal{P}\\). As illustrated above, we divide \\(\\mathcal{P}\\) into two formats: text prompts and visual prompts. To be specific, \\(\\mathcal{P}\\) contains the instructions \\(\\mathcal{P}_{\\mathcal{I}}\\) and task-specific conditions \\(\\mathcal{P}_{\\mathcal{C}}\\), where \\(\\mathcal{P}_{\\mathcal{I}}\\) instructs the model to perform different tasks while \\(\\mathcal{P}_{\\mathcal{C}}\\) indicates diverse conditions which are further used as classifiers to calculate the class scores of predicted masks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.677, + 0.909, + 0.767 + ], + "angle": 0, + "content": "For class-based segmentation tasks like panoptic segmentation, open-vocabulary segmentation (OVS), and video instance segmentation (VIS), \\(\\mathcal{P}\\) can be demonstrated as \\(\\mathcal{P}_{\\mathcal{I}}\\): \"Please segment all the positive objects according to the following potential categories.\" \\(\\mathcal{P}_{\\mathcal{C}}\\): \"[category 1, category 2, category 3, ...]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.909, + 0.856 + ], + "angle": 0, + "content": "For referring and reasoning segmentation tasks like referring expression segmentation (RES), reasoning segmentation, referring video object segmentation (R-VOS), and ReasonVOS, \\(\\mathcal{P}\\) can be designed as \\(\\mathcal{P}_{\\mathcal{I}}\\): \"Can you perform referring or reasoning segmentation according to the language expression?\" \\(\\mathcal{P}_{\\mathcal{C}}\\): \"[referring / reasoning text]\"" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.901 + ], + "angle": 0, + "content": "For visual-guided segmentation tasks like interactive segmentation and video object segmentation (VOS), \\(\\mathcal{P}\\) can be designed as \\(\\mathcal{P}_{\\mathcal{I}}\\): \"Please segment according to the given" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8933" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.089, + 0.891, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.381, + 0.908, + 0.452 + ], + "angle": 0, + "content": "Figure 2. Overview of HyperSeg. HyperSeg encodes the visual input in a multi-grained manner and concatenates the prompt for different perception tasks. We feed learnable fine-grained tokens into a Fine-grained Visual Perceiver (FVP) to integrate multi-scale high-resolution image features into LLM for detailed visual learning and to facilitate space-time information propagation for video understanding. Additionally, we use the semantically enhanced mask tokens and prompt embedding to finally generate the segmentation masks and class scores for generic segmentation, and instance embedding for video instance association." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.467, + 0.483, + 0.557 + ], + "angle": 0, + "content": "visual region reference\" \\(\\mathcal{P}_{\\mathcal{C}}\\) : \"[vision 1, vision 2, vision 3, ...]\". Instead of using an additional region encoder to extract visual reference features [25], we sample the CLIP visual features \\(f_{v}\\) in VLLM according to the region coordinates and perform adaptive average pooling on them to form the final reference features for each visual prompt." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.558, + 0.483, + 0.665 + ], + "angle": 0, + "content": "Segmentation predictor. Segmentation predictor \\( F_{p} \\) generates the masks \\( m \\), corresponding class scores \\( z \\), and instance embedding \\( e \\) through the similar process [7, 15] of three inputs: task-specific prompt embedding \\( \\{E_{\\mathcal{P}}^{k}\\}_{k=1}^{K} \\), the semantically enhanced mask tokens \\( \\{E_{Q}^{j}\\}_{j=1}^{N} \\) and the multi-scale visual features \\( f_{img} \\), where \\( K \\) and \\( N \\) denote \\( K \\) categories and \\( N \\) mask proposals. Formally," + }, + { + "type": "equation", + "bbox": [ + 0.109, + 0.674, + 0.483, + 0.695 + ], + "angle": 0, + "content": "\\[\n\\left\\{m _ {j}, z _ {j}, e _ {j} \\right\\} _ {j = 1} ^ {N} = F _ {p} \\left(\\left\\{E _ {\\mathcal {P}} ^ {k} \\right\\} _ {k = 1} ^ {K}, \\left\\{E _ {\\mathcal {Q}} ^ {j} \\right\\} _ {j = 1} ^ {N}, f _ {i m g}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.703, + 0.483, + 0.796 + ], + "angle": 0, + "content": "where \\(m_j \\in \\mathbb{R}^{H \\times W}\\) is the j-th mask proposal, \\(z_j \\in \\mathbb{R}^K\\) denotes the class scores of \\(m_j\\), and \\(e_j \\in \\mathbb{R}^D\\) denotes the j-th instance embedding obtained from an extra embedding head only for video domain. For video tasks, we adopt a frame-by-frame manner to get frame-level segmentation results for efficient training and inference processes." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.796, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Training objectives. The model can be trained jointly on multiple tasks using the unified loss \\(\\mathcal{L}\\). Specifically, we employ an autoregressive cross-entropy loss \\(\\mathcal{L}_{text}\\) for text prediction, a combination of per-pixel binary cross-entropy loss \\(\\mathcal{L}_{bce}\\) and DICE loss \\(\\mathcal{L}_{dice}\\) for mask supervision \\(\\mathcal{L}_{mask}\\), a cross-entropy loss \\(\\mathcal{L}_{cls}\\) for category classification, and a contrastive loss \\(\\mathcal{L}_{ins}\\) for instance association of video" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.467, + 0.907, + 0.498 + ], + "angle": 0, + "content": "sequences following [48]. \\(\\lambda\\) indicates their sum weight respectively. Formally," + }, + { + "type": "equation", + "bbox": [ + 0.531, + 0.509, + 0.907, + 0.526 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {t e x t}} + \\lambda_ {\\text {m a s k}} \\mathcal {L} _ {\\text {m a s k}} + \\lambda_ {\\text {c l s}} \\mathcal {L} _ {\\text {c l s}} + \\lambda_ {\\text {i n s}} \\mathcal {L} _ {\\text {i n s}}, \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.53, + 0.907, + 0.546 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {m a s k}} = \\lambda_ {b c e} \\mathcal {L} _ {b c e} + \\lambda_ {d i c e} \\mathcal {L} _ {d i c e}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.551, + 0.909, + 0.718 + ], + "angle": 0, + "content": "Differences between HyperSeg and previous methods. Previous universal segmentation methods [24, 25, 31] lacking of VLLMs show inability in reasoning perception tasks while our HyperSeg demonstrates brilliant reasoning segmentation capability in complex scenarios. Besides, we make a significant generalization of the current VLLM-based segmentation methods [22, 51, 58, 59] for more diverse segmentation tasks in both image and video domains using a single model framework. Moreover, HyperSeg differs from previous methods in the three designs elaborated in the following sections." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.726, + 0.757, + 0.743 + ], + "angle": 0, + "content": "3.2. Hybrid Entity Recognition" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.871 + ], + "angle": 0, + "content": "As shown in Fig. 3 (a), predicting presented objects in the way of sequence generation (semantic prediction) tends to miss objects or produce repetitive predictions [45]. On the other hand, Fig. 3 (b), only using VLLM to embed class names (prompt tokens) as mask classifier at the decode stage disregards VLLM's powerful semantic recognition capability. Consequently, we propose a hybrid approach that leverages LLM in both generation and decoding processes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Instead of integrating mask tokens in input sequences and extracting the corresponding embedding from the one-pass" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8934" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.089, + 0.327, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.197, + 0.226, + 0.214, + 0.236 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.342, + 0.088, + 0.631, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.469, + 0.226, + 0.486, + 0.236 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.089, + 0.888, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.779, + 0.226, + 0.796, + 0.236 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.238, + 0.907, + 0.296 + ], + "angle": 0, + "content": "Figure 3. The comparison of different recognition strategies. (a) Generation-Only [22, 41]: both the semantic recognition (existing objects) and their mask tokens are generated by LLM. (b) Decode-Only [58, 59]: prompt embedding and mask tokens are decoded from LLM. The present objects are then determined by their similarity scores. (c) Hybrid (ours): prompt embedding is decoded from LLM while the semantically enhanced mask tokens are generated by LLM. Their similarity scores reflect the objects' presence." + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.305, + 0.277, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.153, + 0.462, + 0.171, + 0.473 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.281, + 0.305, + 0.47, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.387, + 0.462, + 0.406, + 0.473 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.478, + 0.484, + 0.547 + ], + "angle": 0, + "content": "Figure 4. Comparison between previous vision perceiver and our FVP. (a): previous vision perceiver [2, 23] uses the coarse single-scale CLIP visual features which are inadequate for fine-grained perception tasks. (b): FVP encodes the multi-scale visual features into fine-grained tokens." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.564, + 0.484, + 0.686 + ], + "angle": 0, + "content": "forward output of VLLM, we instruct VLLM to generate the mask tokens preceded by the estimated objects' names. As illustrated in Fig. 3 (c), VLLM is compelled to generate all the existing objects in the vision input and then the mask tokens. The semantically enhanced mask tokens contain valuable semantic integrated information about the image, which are subsequently used as input for the segmentation predictor to generate segmentation masks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.696, + 0.357, + 0.712 + ], + "angle": 0, + "content": "3.3. Fine-grained Visual Perceiver" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.838 + ], + "angle": 0, + "content": "Why twin-tower vision encoder? As shown in Fig. 4, previous VLLMs and VLLM-based segmentation methods usually utilize the pre-trained CLIP encoder to obtain single-scale and low-resolution vision features interacted with diverse languages, which is insufficient for fine-grained image and video segmentation tasks. Therefore, we adopt an extra pyramid vision encoder [7] to inject details-aware visual information." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Specifically, we fuse multi-scale visual features into fine-grained tokens (stated as \\( P \\) in Sec 3.1) which can inject rich fine-grained visual information into the pre-trained VLLMs without excessive computation cost. Formally, given the" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.31, + 0.907, + 0.385 + ], + "angle": 0, + "content": "vision input \\(\\nu\\), we leverage a pyramid vision encoder [7] \\(F_{seg}\\) to get details-aware image features \\(f_{img}\\). For the j-th scale and the previous fine-grained tokens \\(P_{j-1}\\), the FVP module enriches each token through conditional weighted cross-attention:" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.398, + 0.907, + 0.42 + ], + "angle": 0, + "content": "\\[\n\\hat {P} _ {j} = \\operatorname {M H C A} \\left(P _ {j - 1}, G _ {p} \\left(f _ {\\text {i m g}} ^ {(j)}\\right)\\right), \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.593, + 0.422, + 0.907, + 0.441 + ], + "angle": 0, + "content": "\\[\nP _ {j} = P _ {j - 1} + \\tanh (\\mathrm {M L P} (\\hat {P} _ {j})) \\cdot \\hat {P} _ {j}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.447, + 0.909, + 0.584 + ], + "angle": 0, + "content": "where MHCA denotes the Multi-Head Cross-Attention layer, \\( G_{p} \\) is the projection function, tanh is a normalization function and MLP is a Multilayer Perceptron. The component of \\( \\tanh(\\mathrm{MLP}(\\hat{P}_j)) \\) is the conditional weight used to multiply the enriched fine-grained tokens \\( \\hat{P}_j \\) before the residual connection to the previous tokens \\( P_{j - 1} \\). Additionally, we initialize the weight value to zero to ensure the adaptation to diverse multi-scale image features while retaining the training stability." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.594, + 0.694, + 0.61 + ], + "angle": 0, + "content": "3.4. Temporal Adapter" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.618, + 0.909, + 0.752 + ], + "angle": 0, + "content": "Video segmentation entails distinct challenges, requiring reasoning across multiple frames and the maintenance of temporal coherence. Existing VLLM-based methods exhibit limitations in addressing video perception tasks and lack specialized designs for comprehending temporal dynamics in video analysis. To this end, we utilize global prompt aggregation and local space-time information injection in the time dimension to adapt to more complicated video perception tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.754, + 0.909, + 0.83 + ], + "angle": 0, + "content": "Global prompt aggregation. For the current prompt embedding \\( E_{\\mathcal{P}} \\) in the video object mask retrieval process, we leverage the adaptive average pooling strategy along the time dimension to aggregate global object and temporal information of previous \\( T \\) frames." + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.84, + 0.907, + 0.86 + ], + "angle": 0, + "content": "\\[\nE _ {\\mathcal {P}} = \\operatorname {A v g P o o l} ([ E _ {\\mathcal {P}} ^ {0}, E _ {\\mathcal {P}} ^ {1}, \\dots , E _ {\\mathcal {P}} ^ {T} ]), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Local space-time information injection. We propose a sequential renewal strategy for space-time information prop" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8935" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.131 + ], + "angle": 0, + "content": "Table 1. Comparison with the state-of-the-art models on the RefCOCO series and more challenging generalized referring expression segmentation benchmark gRefCOCO. \\(\\ddagger\\) denotes models using pre-trained SAM [21] for mask generation. * means using gRefCOCO for training while other methods are evaluated in zero-shot manners." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.143, + 0.903, + 0.367 + ], + "angle": 0, + "content": "
TypeMethodRefCOCORefCOCO+RefCOCOggRefCOCO
valtestAtestBvaltestAtestBval(U)test(U)valtestAtestB
SegmentationVLT [11]67.570.565.256.361.050.155.057.752.5*62.2*50.5*
CRIS [44]70.573.266.162.368.153.759.960.455.3*63.8*51.0*
LAVT [53]72.775.868.862.168.455.161.262.157.6*65.3*55.0*
PolyFormer-B [30]74.876.671.167.672.959.367.869.1---
VLLM-basedLISA-7B [22]‡74.979.172.365.170.858.167.970.638.7*52.6*44.8*
PixelLM-7B [41]73.076.568.266.371.758.369.370.5---
GSVA-7B [49]‡76.477.472.864.567.758.671.172.061.7*69.2*60.3*
GroundHog-7B [33]78.579.975.770.575.064.974.174.666.7*--
SAM4MLLM-7B [6]‡79.682.876.173.577.865.874.575.666.3*70.1*63.2*
Segmentation NetworkLaSagnA-7B [45]‡76.878.773.866.470.660.170.671.938.150.442.1
OMG-LLaVA [58]78.080.374.169.173.163.072.972.9---
GLaMM [40]‡79.583.276.972.678.764.674.274.9---
PSALM [59]83.684.781.672.975.570.173.874.442.052.450.6
HyperSeg84.885.783.479.083.575.279.478.947.557.352.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.379, + 0.907, + 0.409 + ], + "angle": 0, + "content": "Table 2. Comparison with the state-of-the-art models on more complex and challenging reasoning segmentation benchmarks: ReVOS in video domain and ReasonSeg in image domain. \\(\\ddagger\\) denotes the same meaning as Tab. 1." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.419, + 0.905, + 0.585 + ], + "angle": 0, + "content": "
MethodBackboneReVOS-ReasoningReVOS-ReferringReVOS-OverallReasonSeg
JFJ&FJFJ&FJFJ&FgIoUcIoU
LMPM [12]Swin-T13.324.318.829.039.134.121.231.726.4--
ReferFormer [47]Video-Swin-B21.325.623.431.234.332.726.229.928.1--
LISA-7B [22]‡ViT-H33.838.436.144.347.145.739.142.740.952.954.0
LaSagnA-7B [45]‡ViT-H---------48.847.2
SAM4MLLM-7B [6]‡EfficientViT-SAM-XL1---------46.748.1
TrackGPT-13B [62]‡ViT-H38.142.940.548.350.649.543.246.845.0--
VISA-7B [51]‡ViT-H36.741.739.251.154.752.943.948.246.152.757.8
VISA-13B [51]‡ViT-H38.343.540.952.355.854.145.349.747.5--
HyperSeg-3BSwin-B50.255.853.056.060.958.553.158.455.759.256.7
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.61, + 0.486, + 0.641 + ], + "angle": 0, + "content": "agation based on fine-grained tokens \\(P\\) to inject object information of adjacent frames. Formally," + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.652, + 0.484, + 0.671 + ], + "angle": 0, + "content": "\\[\nP _ {t} = G _ {l} \\left[ F _ {L L M} \\left(P _ {t - 1}\\right) \\right], \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.68, + 0.484, + 0.74 + ], + "angle": 0, + "content": "where \\(P_{t}\\) denotes the time-aware fine-grained tokens of the current \\(t\\)-th frame, \\(G_{l}\\) is the projection function to transfer the previous features to the current space and align the feature dimensions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.741, + 0.484, + 0.833 + ], + "angle": 0, + "content": "The proposed global prompt aggregation and local spacetime information injection within our temporal adapter facilitate the coalescence of both long-term and short-term vision-language information, which is essential for comprehensive video perception. More details about the temporal adapter are provided in the Supplementary Material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.845, + 0.224, + 0.863 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Datasets. We use the one-stage training strategy to train HyperSeg with the multi-dataset and multi-task manners." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.61, + 0.907, + 0.732 + ], + "angle": 0, + "content": "For image segmentation, we use COCO Panoptic [26], RefCOCO series [35, 55], COCO-Interactive, and ReasonSeg [22]. For video segmentation, we utilize DAVIS-2017 datasets [4], Ref-Youtu-VOS [42], YouTube-VIS 2019 [52], and ReVOS [51]. Besides, we use LLAVA-150k [29] to maintain the vision-language conversation capability of VLLM (we show the results on Multi-modal benchmarks in the Supplementary Material)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Implementation details We load the pre-trained weights of Mipha [63] for our VLLM, and Maks2Former [7] for our segmentation predictor. We use three layers of FVP for fine-grained information fusion and utilize LoRA [18] to finetune the LLM efficiently. We train HyperSeg jointly for 160k iterations using a batch size of 32 on 8 NVIDIA A100 GPUs, which means each task takes approximately 16k iterations. We employ the AdamW optimizer with a learning rate of \\(4 \\times 10^{-5}\\) and with a cosine schedule. All the hyper-parameters in the loss \\(\\mathcal{L}\\) are assigned values 1.0." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8936" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.12, + 0.09, + 0.877, + 0.105 + ], + "angle": 0, + "content": "Table 3. Quantitative results on the closed-set COCO-Panoptic segmentation, open-vocabulary segmentation (-OV) benchmarks." + }, + { + "type": "table", + "bbox": [ + 0.124, + 0.115, + 0.872, + 0.283 + ], + "angle": 0, + "content": "
TypeMethodBackboneCOCO-PanopticADE-OVCitys-OVPC59-OVPAS20-OV
PQmIoUPQmIoUPQmIoUmIoU
Segmentation SpecialistMask2former [7]Swin-B55.165.1-----
OneFormer [19]Swin-L57.967.4-----
SEEM [64]DaViT-B56.166.3-----
MaskCLIP [13]ViT-L30.947.615.123.7-45.9-
DeOP [16]ResNet-101c---22.9-48.891.7
SimBaseline [50]ViT-B---20.5-47.788.4
DaTaSeg [15]ViTDet-B52.862.712.318.328.051.1-
VLLM-based Segmentation NetworkOMG-LLaVA [58]ConvNeXt-L53.8------
PSALM [22]Swin-B55.966.613.718.228.848.581.3
HyperSegSwin-B61.277.216.122.331.164.692.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.305, + 0.486, + 0.348 + ], + "angle": 0, + "content": "Table 4. Results of common video segmentation benchmarks, including DAVIS17, Ref-YouTube-VOS, Ref-DAVIS17, and YouTube-VIS 2019. \\(\\ddagger\\) denotes the same meaning as Tab. 1." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.358, + 0.486, + 0.482 + ], + "angle": 0, + "content": "
MethodBackboneDAVIS17Ref-YTRef-DAVISYT-VIS
J&FJ&FJ&FmAP
SEEM [64]DaViT-B62.8---
OMG-Seg [24]ConvNeXt-L74.3--56.4
ReferFormer [47]Video-Swin-B-62.961.1-
OnlineRefer [46]Swin-L-63.564.8-
UNINEXT [25]ConvNeXt-L77.266.266.764.3
LISA-7B [22]‡ViT-H-53.964.8-
VISA-13B [51]‡ViT-H-63.070.4-
VideoLISA-3.8B [3]‡ViT-H-63.768.8-
HyperSeg-3BSwin-B77.668.571.253.8
" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.497, + 0.405, + 0.513 + ], + "angle": 0, + "content": "4.1. Comparisons with State-of-the-Arts" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.522, + 0.486, + 0.703 + ], + "angle": 0, + "content": "Referring expression segmentation results. We compare HyperSeg with the state-of-the-art methods on the benchmarks RefCOCO+/g [35, 55] and more challenging generalized referring expression segmentation benchmark gRefCOCO [27]. in Tab. 1. Based on the versatile and adaptable design of HyperSeg, our model achieves state-of-the-art performance on all the referring datasets. Specifically, HyperSeg surpasses the current SOTA by a large margin, reaching 79.7 cIoU on RefCOCO+ val (+6.8 over PSALM). Besides, Our model shows superiority in challenging G-RES tasks compared with previous zero-shot methods, demonstrating the robustness and generalization ability of HyperSeg." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.704, + 0.486, + 0.825 + ], + "angle": 0, + "content": "Reasoning segmentation results. We compare HyperSeg with the state-of-the-art methods on image reasoning segmentation (ReasonSeg [22]) and reasoning video object segmentation (ReVOS [51]) in Tab. 2. Our HyperSeg achieves superior performance on reasoning tasks, significantly surpassing previous state-of-the-art methods (+12.1 on ReVOS-Reasoning), which shows HyperSeg powerful reasoning capability of tackling complex scenarios." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.826, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Generic image segmentation results. We show the performance of HyperSeg on COCO-Panoptic [26] and open-vocabulary segmentation [9, 14, 34, 60] tasks in Tab. 3. HyperSeg achieves excellent performance compared with both specialist models and VLLMs-based methods on both closed-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.308, + 0.908, + 0.412 + ], + "angle": 0, + "content": "set and open-vocabulary segmentation tasks. Specifically, HyperSeg surpasses the VLLM-based PSALM by a significant margin (+5.3 on COCO PQ, and +10.6 on mIoU), which demonstrates our powerful capabilities of handling complex semantic perception and segmentation tasks. Besides, we show the results of COCO-Interactive in the Supplementary Material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.413, + 0.909, + 0.55 + ], + "angle": 0, + "content": "Common video segmentation results. We compare HyperSeg with previous video segmentation methods in Tab. 4, including visual-prompted semi-supervised VOS (DAVIS17 val), text-prompted referring video object segmentation (RefYouTube-VOS, Ref-DAVIS17) and video instance segmentation (YouTube-VIS 2019). HyperSeg shows promising results over previous unified segmentation methods [24, 25]. Besides, HyperSeg performs more video perception tasks than previous VLLM-based models [3, 51]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.56, + 0.627, + 0.575 + ], + "angle": 0, + "content": "4.2. Ablations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.909, + 0.734 + ], + "angle": 0, + "content": "The mutual influence between different tasks. Our model can be trained and inferred across multiple tasks and datasets simultaneously. We evaluate the mutual impact of different tasks in Tab. 5. The results show that joint training can enhance the model performance compared with the task-specific model. Besides, the performance of video segmentation tasks can be improved significantly by adding the image training datasets. This demonstrates the generalization and self-consistency of our HyperSeg to perform universal segmentation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.84 + ], + "angle": 0, + "content": "Effect of different LLMs and vision backbone. In Tab. 6, we evaluate the effect of different sizes of LLMs and vision backbone. Our HyperSeg achieves excellent performance using smaller LLMs and vision encoder compared with the previous SOTA models like VISA[51] and PSALM[59]. Besides, the performance of HyperSeg can be further improved by using the more powerful LLM (Phi-2-2.7B [20])." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Ablation on the proposed components. We assess the effectiveness of our proposed FVP module and Hybrid Entity Recognition strategy. As shown in Tab. 7, with our fine-grained visual integration and hybrid entity semantic" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8937" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 5. The mutual influence between different tasks. Task-specific means training task-specific models only on data from corresponding tasks, Refer+Reason denotes the model is trained on referring and reasoning segmentation data." + }, + { + "type": "table", + "bbox": [ + 0.141, + 0.129, + 0.859, + 0.232 + ], + "angle": 0, + "content": "
Task-specificRefer+ReasonVideoImageRefCOCOCOCOReVOSYT-VIS
valtestAtestBPQmIoUReasoningReferringOverallmAP
✓83.885.982.260.875.151.256.653.950.7
✓83.384.980.9--53.157.355.2-
✓85.686.182.460.976.5----
✓-----51.157.054.150.4
✓✓✓84.885.783.461.277.253.058.555.753.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.145, + 0.244, + 0.851, + 0.26 + ], + "angle": 0, + "content": "Table 6. The comparison of different LLMs and backbone usages. w/o CLIP means without using CLIP vision encoder." + }, + { + "type": "table", + "bbox": [ + 0.14, + 0.27, + 0.86, + 0.387 + ], + "angle": 0, + "content": "
MethodLLMCOCOReVOSADE-OVPC59-OVPAS20-OV
PQmIoUReasoningReferringOverallmIoUmIoUmIoU
LISA [22]Vicuna-7B--36.145.740.9---
VISA [51]Vicuna-13B--40.954.147.5---
PSALM(w/o CLIP) [22]Phi-1.5-1.3B55.966.6---18.248.581.3
HyperSeg (w/o CLIP)Phi-1.5-1.3B61.176.044.049.746.918.960.090.6
HyperSegPhi-1.5-1.3B60.976.750.857.053.920.361.590.8
HyperSegPhi-2-2.7B61.277.253.058.555.722.364.692.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.402, + 0.486, + 0.445 + ], + "angle": 0, + "content": "Table 7. Ablation on the core components of HyperSeg. FVP and HER denote the proposed Fine-grained Visual Perceiver and Hybrid Entity Recognition modules." + }, + { + "type": "table", + "bbox": [ + 0.122, + 0.455, + 0.456, + 0.555 + ], + "angle": 0, + "content": "
FVPHERYT-VISCOCORefCOCO
mAPPQmIoUcIoU
✓48.454.866.282.8
50.855.866.684.6
✓✓52.059.774.684.3
✓✓53.861.277.284.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.569, + 0.483, + 0.612 + ], + "angle": 0, + "content": "Table 8. Ablation on the Fine-grained Visual Perceiver design. CW denotes the Conditional Weight illustrated in Sec. 3.3, and Scale denotes the total scale in the proposed FVP module." + }, + { + "type": "table", + "bbox": [ + 0.107, + 0.622, + 0.465, + 0.703 + ], + "angle": 0, + "content": "
CWScaleYT-VISCOCORefCOCO
mAPPQmIoUcIoU
✓single-layer49.755.868.083.7
multi-layers50.458.973.484.5
multi-layers53.861.277.284.8
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.719, + 0.483, + 0.748 + ], + "angle": 0, + "content": "enhancement, the segmentation accuracy can be enhanced significantly \\((+5.4\\) on YT-VIS, \\(+6.4\\) on COCO panoptic PQ)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.869 + ], + "angle": 0, + "content": "Design of the Fine-grained Visual Perceiver. In the FVP module, we combine multi-scale visual features into fixed perception queries using the condition-wise cross-attention layers to extract rich visual details from different scales of the pyramid encoder. As shown in Tab. 8, together with the conditional weight and the multi-scale design, our model makes a significant improvement on both image and video segmentation tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Effect of temporal adapter. We evaluate the effectiveness of the proposed temporal adapter including global prompt ag" + }, + { + "type": "table_caption", + "bbox": [ + 0.538, + 0.402, + 0.88, + 0.417 + ], + "angle": 0, + "content": "Table 9. Ablation on the temporal adapter for video tasks." + }, + { + "type": "table", + "bbox": [ + 0.541, + 0.428, + 0.88, + 0.526 + ], + "angle": 0, + "content": "
GlobalLocalRef-DAVIS17ReVOSYT-VIS
J&FJ&FmAP
✓67.354.147.9
68.854.548.5
✓69.354.850.2
✓✓71.255.753.8
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.547, + 0.908, + 0.609 + ], + "angle": 0, + "content": "gregation (global) and local space-time information injection (local) in Tab. 9. Incorporating both global and local components, the temporal adapter significantly enhances model performance across multiple video segmentation tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.631, + 0.634, + 0.646 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.909, + 0.901 + ], + "angle": 0, + "content": "In this study, we aim to present HyperSeg, the first VLLM-based universal segmentation model designed for pixel-level image and video perception, encompassing a wide range of generic segmentation and complex reasoning tasks. We propose the Hybrid Entity Recognition and Fine-grained Visual Perceiver to leverage the recognition capacity of VLLMs more effectively and enhances the VLLM's ability by capturing diverse levels of visual information without incurring excessive computational costs. With additional Temporal Adapter, HyperSeg can tackle challenging video tasks by incorporating global and local information. HyperSeg surpasses existing methods on complex reasoning segmentation and traditional perception tasks. The insights presented in this work expand the possibilities of VLLMs in visual perception and lay a foundation for future research on the integration of vision-language models." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8938" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.27, + 0.108 + ], + "angle": 0, + "content": "6. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.486, + 0.178 + ], + "angle": 0, + "content": "This work was partly supported by the National Key Research and Development Program of China (No. 2024YFB2808903) and the Shenzhen Science and Technology Program (JCYJ20220818101001004)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.193, + 0.188, + 0.209 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.218, + 0.486, + 0.3 + ], + "angle": 0, + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.303, + 0.485, + 0.371 + ], + "angle": 0, + "content": "[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 2, 3, 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.375, + 0.484, + 0.442 + ], + "angle": 0, + "content": "[3] Zechen Bai, Tong He, Haiyang Mei, Pichao Wang, Ziteng Gao, Joya Chen, Lei Liu, Zheng Zhang, and Mike Zheng Shou. One token to seg them all: Language instructed reasoning segmentation in videos. arXiv preprint arXiv:2409.19603, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.446, + 0.484, + 0.5 + ], + "angle": 0, + "content": "[4] Sergi Caelles, Alberto Montes, Kevis-Kokitsi Maninis, Yuhua Chen, Luc Van Gool, Federico Perazzi, and Jordi Pont-Tuset. The 2018 davis challenge on video object segmentation. arXiv preprint arXiv:1803.00557, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.504, + 0.484, + 0.558 + ], + "angle": 0, + "content": "[5] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 3, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.562, + 0.484, + 0.629 + ], + "angle": 0, + "content": "[6] Yi-Chia Chen, Wei-Hua Li, Cheng Sun, Yu-Chiang Frank Wang, and Chu-Song Chen. Sam4mllm: Enhance multimodal large language model for referring expression segmentation. In European Conference on Computer Vision, pages 323-340. Springer, 2025. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.633, + 0.484, + 0.701 + ], + "angle": 0, + "content": "[7] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.704, + 0.484, + 0.757 + ], + "angle": 0, + "content": "[8] Jaemin Cho, Jie Lei, Hao Tan, and Mohit Bansal. Unifying vision-and-language tasks via text generation. In International Conference on Machine Learning, pages 1931-1942. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.761, + 0.484, + 0.842 + ], + "angle": 0, + "content": "[9] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[10] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning, 2023. 3, 1" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.218, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.161 + ], + "angle": 0, + "content": "[11] Henghui Ding, Chang Liu, Suchen Wang, and Xudong Jiang. Vision-language transformer and query generation for referring segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16321-16330, 2021. 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.908, + 0.232 + ], + "angle": 0, + "content": "[12] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, and Chen Change Loy. Mevis: A large-scale benchmark for video segmentation with motion expressions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2694-2703, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.234, + 0.907, + 0.272 + ], + "angle": 0, + "content": "[13] Zheng Ding, Jieke Wang, and Zhuowen Tu. Open-vocabulary universal image segmentation with maskclip. arXiv preprint arXiv:2208.08984, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.907, + 0.328 + ], + "angle": 0, + "content": "[14] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-338, 2010. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.331, + 0.908, + 0.412 + ], + "angle": 0, + "content": "[15] Xiuye Gu, Yin Cui, Jonathan Huang, Abdullah Rashwan, Xuan Yang, Xingyi Zhou, Golnaz Ghiasi, Weicheng Kuo, Huizhong Chen, Liang-Chieh Chen, et al. Dataseg: Taming a universal multi-dataset multi-task segmentation model. Advances in Neural Information Processing Systems, 36, 2024. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.907, + 0.468 + ], + "angle": 0, + "content": "[16] Cong Han, Yujie Zhong, Dengjie Li, Kai Han, and Lin Ma. Open-vocabulary semantic segmentation with decoupled one-pass network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1086-1096, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.907, + 0.51 + ], + "angle": 0, + "content": "[17] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.512, + 0.908, + 0.565 + ], + "angle": 0, + "content": "[18] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.568, + 0.908, + 0.636 + ], + "angle": 0, + "content": "[19] Jitesh Jain, Jiachen Li, Mang Tik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. Oneformer: One transformer to rule universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2989-2998, 2023. 2, 7, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.638, + 0.908, + 0.705 + ], + "angle": 0, + "content": "[20] Mojan Javaheripi, Sébastien Bubeck, Marah Abdin, Jyoti Aneja, Sebastien Bubeck, Caio César Teodoro Mendes, Weizhu Chen, Allie Del Giorno, Ronen Eldan, Sivakanth Gopi, et al. Phi-2: The surprising power of small language models. *Microsoft Research Blog*, 1:3, 2023. 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.707, + 0.908, + 0.786 + ], + "angle": 0, + "content": "[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 6, 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.907, + 0.843 + ], + "angle": 0, + "content": "[22] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. ArXiv, abs/2308.00692, 2023. 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.907, + 0.9 + ], + "angle": 0, + "content": "[23] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 2, 3, 5, 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8939" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.175 + ], + "angle": 0, + "content": "[24] Xiangtai Li, Haobo Yuan, Wei Li, Henghui Ding, Size Wu, Wenwei Zhang, Yining Li, Kai Chen, and Chen Change Loy. Omg-seg: Is one model good enough for all segmentation? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27948-27959, 2024. 2, 4, 7, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.484, + 0.245 + ], + "angle": 0, + "content": "[25] Fangjian Lin, Jianlong Yuan, Sitong Wu, Fan Wang, and Zhibin Wang. Uninext: Exploring a unified architecture for vision recognition. In Proceedings of the 31st ACM International Conference on Multimedia, pages 3200-3208, 2023. 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.247, + 0.484, + 0.302 + ], + "angle": 0, + "content": "[26] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision, 2014. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.303, + 0.484, + 0.357 + ], + "angle": 0, + "content": "[27] Chang Liu, Henghui Ding, and Xudong Jiang. Gres: Generalized referring expression segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23592-23601, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.359, + 0.484, + 0.399 + ], + "angle": 0, + "content": "[28] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.401, + 0.484, + 0.441 + ], + "angle": 0, + "content": "[29] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2, 3, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.443, + 0.484, + 0.524 + ], + "angle": 0, + "content": "[30] Jiang Liu, Hui Ding, Zhaowei Cai, Yuting Zhang, Ravi Kumar Satzoda, Vijay Mahadevan, and R Manmatha. Polyformer: Referring image segmentation as sequential polygon generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18653-18663, 2023. 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.526, + 0.484, + 0.594 + ], + "angle": 0, + "content": "[31] Yong Liu, Cairong Zhang, Yitong Wang, Jiahao Wang, Yujiu Yang, and Yansong Tang. Universal segmentation at arbitrary granularity with language instruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3459-3469, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.596, + 0.484, + 0.664 + ], + "angle": 0, + "content": "[32] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.665, + 0.484, + 0.733 + ], + "angle": 0, + "content": "[33] Bo Miao, Mohammed Bennamoun, Yongsheng Gao, and Ajmal Mian. Spectrum-guided multi-granularity referring video object segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 920-930, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.484, + 0.817 + ], + "angle": 0, + "content": "[34] Roozbeh Mottaghi, Xianjie Chen, Xiaobai Liu, Nam-Gyu Cho, Seong-Whan Lee, Sanja Fidler, Raquel Urtasun, and Alan Yuille. The role of context for object detection and semantic segmentation in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 891–898, 2014. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[35] Varun K Nagaraja, Vlad I Morariu, and Larry S Davis. Modeling context between objects for referring expression understanding. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 792-807. Springer, 2016. 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.148 + ], + "angle": 0, + "content": "[36] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.906, + 0.19 + ], + "angle": 0, + "content": "[37] Renjie Pi, Lewei Yao, Jiahui Gao, Jipeng Zhang, and Tong Zhang. Perceptiongpt: Effectively fusing visual perception into llm. arXiv preprint arXiv:2311.06612, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.191, + 0.907, + 0.272 + ], + "angle": 0, + "content": "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.907, + 0.343 + ], + "angle": 0, + "content": "[39] Hanoona Rasheed, Muhammad Maaz, Sahal Shaji, Abdelrahman Shaker, Salman Khan, Hisham Cholakkal, Rao M Anwer, Erix Xing, Ming-Hsuan Yang, and Fahad S Khan. Glamm: Pixel grounding large multimodal model. arXiv preprint arXiv:2311.03356, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.907, + 0.427 + ], + "angle": 0, + "content": "[40] Hanoona Rasheed, Muhammad Maaz, Sahal Shaji, Abdelrahman Shaker, Salman Khan, Hisham Cholakkal, Rao M Anwer, Eric Xing, Ming-Hsuan Yang, and Fahad S Khan. Glamm: Pixel grounding large multimodal model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13009-13018, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.907, + 0.482 + ], + "angle": 0, + "content": "[41] Zhongwei Ren, Zhicheng Huang, Yunchao Wei, Yao Zhao, Dongmei Fu, Jiashi Feng, and Xiaojie Jin. Pixel reasoning with large multimodal model. ArXiv, abs/2312.02228, 2023. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.907, + 0.565 + ], + "angle": 0, + "content": "[42] Seonguk Seo, Joon-Young Lee, and Bohyung Han. Urvos: Unified referring video object segmentation network with a large-scale benchmark. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XV 16, pages 208–223. Springer, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.567, + 0.907, + 0.636 + ], + "angle": 0, + "content": "[43] Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong Lu, Jie Zhou, Yu Qiao, et al. Vision: Large language model is also an open-ended decoder for vision-centric tasks. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.637, + 0.907, + 0.705 + ], + "angle": 0, + "content": "[44] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clipdriven referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11686-11695, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.707, + 0.907, + 0.761 + ], + "angle": 0, + "content": "[45] Cong Wei, Haoxian Tan, Yujie Zhong, Yujiu Yang, and Lin Ma. Lasagna: Language-based segmentation assistant for complex queries. arXiv preprint arXiv:2404.08506, 2024. 2, 4, 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.762, + 0.907, + 0.831 + ], + "angle": 0, + "content": "[46] Dongming Wu, Tiancai Wang, Yang Zhang, Xiangyu Zhang, and Jianbing Shen. Onlinerefer: A simple online baseline for referring video object segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2761-2770, 2023. 7, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.907, + 0.9 + ], + "angle": 0, + "content": "[47] Jiannan Wu, Yi Jiang, Peize Sun, Zehuan Yuan, and Ping Luo. Language as queries for referring video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4974-4984, 2022. 6, 7, 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "8940" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[48] Junfeng Wu, Qihao Liu, Yi Jiang, Song Bai, Alan Yuille, and Xiang Bai. In defense of online models for video instance segmentation. In ECCV, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.134, + 0.483, + 0.189 + ], + "angle": 0, + "content": "[49] Zhuofan Xia, Dongchen Han, Yizeng Han, Xuran Pan, Shiji Song, and Gao Huang. Gsva: Generalized segmentation via multimodal large language models. arXiv preprint arXiv:2312.10103, 2023. 2, 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.19, + 0.486, + 0.259 + ], + "angle": 0, + "content": "[50] Mengde Xu, Zheng Zhang, Fangyun Wei, Yutong Lin, Yue Cao, Han Hu, and Xiang Bai. A simple baseline for open-vocabulary semantic segmentation with pre-trained vision-language model. In European Conference on Computer Vision, pages 736-753. Springer, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.261, + 0.485, + 0.327 + ], + "angle": 0, + "content": "[51] Cilin Yan, Haochen Wang, Shilin Yan, Xiaolong Jiang, Yao Hu, Guoliang Kang, Weidi Xie, and Efstratios Gavves. Visa: Reasoning video object segmentation via large language models. arXiv preprint arXiv:2407.11325, 2024. 2, 4, 6, 7, 8, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.33, + 0.483, + 0.372 + ], + "angle": 0, + "content": "[52] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5188-5197, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.372, + 0.485, + 0.441 + ], + "angle": 0, + "content": "[53] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. Levt: Language-aware vision transformer for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18155-18165, 2022. 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.442, + 0.483, + 0.496 + ], + "angle": 0, + "content": "[54] Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and Yinfei Yang. Ferret: Refer and ground anything anywhere at any granularity. arXiv preprint arXiv:2310.07704, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.497, + 0.485, + 0.579 + ], + "angle": 0, + "content": "[55] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 69-85. Springer, 2016. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.581, + 0.483, + 0.635 + ], + "angle": 0, + "content": "[56] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.636, + 0.485, + 0.704 + ], + "angle": 0, + "content": "[57] Hao Zhang, Feng Li, Xueyan Zou, Shilong Liu, Chunyuan Li, Jianwei Yang, and Lei Zhang. A simple framework for open-vocabulary segmentation and detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1020-1031, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.706, + 0.483, + 0.773 + ], + "angle": 0, + "content": "[58] Tao Zhang, Xiangtai Li, Hao Fei, Haobo Yuan, Shengqiong Wu, Shunping Ji, Chen Change Loy, and Shuicheng Yan. Omg-llava: Bridging image-level, object-level, pixel-level reasoning and understanding. arXiv preprint arXiv:2406.19389, 2024. 2, 4, 5, 6, 7, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.776, + 0.483, + 0.817 + ], + "angle": 0, + "content": "[59] Zheng Zhang, Yeyao Ma, Enming Zhang, and Xiang Bai. Psalm: Pixelwise segmentation with large multi-modal model. arXiv preprint arXiv:2403.14598, 2024. 2, 3, 4, 5, 6, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.483, + 0.872 + ], + "angle": 0, + "content": "[60] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision, 127(3):302-321, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[61] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.093, + 0.486, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023.2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.907, + 0.176 + ], + "angle": 0, + "content": "[62] Jiawen Zhu, Zhi-Qi Cheng, Jun-Yan He, Chenyang Li, Bin Luo, Huchuan Lu, Yifeng Geng, and Xuansong Xie. Tracking with human-intent reasoning. arXiv preprint arXiv:2312.17448, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.906, + 0.246 + ], + "angle": 0, + "content": "[63] Minjie Zhu, Yichen Zhu, Xin Liu, Ning Liu, Zhiyuan Xu, Chaomin Shen, Yaxin Peng, Zhicai Ou, Feifei Feng, and Jian Tang. A comprehensive overhaul of multimodal assistant with small language models. arXiv preprint arXiv:2403.06199, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.906, + 0.304 + ], + "angle": 0, + "content": "[64] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Wang, Lijuan Wang, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. Advances in Neural Information Processing Systems, 36, 2024. 7, 1, 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.956 + ], + "angle": 0, + "content": "8941" + } + ] +] \ No newline at end of file diff --git a/2025/HyperSeg_ Hybrid Segmentation Assistant with Fine-grained Visual Perceiver/d3d1df0c-7574-4e1e-b447-81df1bf28440_origin.pdf b/2025/HyperSeg_ Hybrid Segmentation Assistant with Fine-grained Visual Perceiver/d3d1df0c-7574-4e1e-b447-81df1bf28440_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..551967c15adb2acc6d494069c28bd3d4eaf8459d --- /dev/null +++ b/2025/HyperSeg_ Hybrid Segmentation Assistant with Fine-grained Visual Perceiver/d3d1df0c-7574-4e1e-b447-81df1bf28440_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4b57675caf2d974c5c4d0994c3bc32ffcd9715eb49533a43ac3d6f42e05cab9d +size 2550490 diff --git a/2025/HyperSeg_ Hybrid Segmentation Assistant with Fine-grained Visual Perceiver/full.md b/2025/HyperSeg_ Hybrid Segmentation Assistant with Fine-grained Visual Perceiver/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7085899a63c986e49c3ad36b71e2861c7eee2ae4 --- /dev/null +++ b/2025/HyperSeg_ Hybrid Segmentation Assistant with Fine-grained Visual Perceiver/full.md @@ -0,0 +1,317 @@ +# HyperSeg: Hybrid Segmentation Assistant with Fine-grained Visual Perceiver + +Cong Wei $^{1,2}$ , Yujie Zhong $^{2\dagger}$ , Haoxian Tan $^{2}$ , Yong Liu $^{1}$ , Jie Hu $^{2}$ , Dengjie Li $^{2}$ , Zheng Zhao $^{2}$ , Yujiu Yang $^{1\dagger}$ , Tsinghua Shenzhen International Graduate School, Tsinghua University ${}^{2}$ Meituan Inc. + +weic22@mails.tsinghua.edu.cn, jaszhong@hotmail.com, yang.yujiu@sz.tsinghua.edu.cn + +![](images/9f5949157c434b915dc18761b6461f836a0be7ebae7519358b02fb1580c66fc7.jpg) +Figure 1. Illustration of our HyperSeg which can conduct image and video segmentation tasks with various language and visual instructions. Additionally, HyperSeg can handle complicated reasoning perception tasks compared with previous universal segmentation methods. To our knowledge, HyperSeg is the first VLLM-based universal segmentation model with perception and complex reasoning abilities in both image and video domains. + +# Abstract + +This paper aims to address universal segmentation for image and video perception with the strong reasoning ability empowered by Visual Large Language Models (VLLMs). Despite significant progress in current unified segmentation methods, limitations in adaptation to both image and video scenarios, as well as the complex reasoning segmentation, make it difficult for them to handle various challenging instructions and achieve an accurate understanding of fine-grained vision-language correlations. We propose HyperSeg, the first VLLM-based universal segmentation model for pixel-level image and video perception, en + +compassing generic segmentation tasks and more complex reasoning perception tasks requiring powerful reasoning abilities and world knowledge. Besides, to fully leverage the recognition capabilities of VLLMs and the fine-grained visual information, HyperSeg incorporates hybrid entity recognition and fine-grained visual perceiver modules for various segmentation tasks. Combined with the temporal adapter, HyperSeg achieves a comprehensive understanding of temporal information. Experimental results validate the effectiveness of our insights in resolving universal image and video segmentation tasks, including the more complex reasoning perception tasks. Our code is available at https://github.com/congvvc/HyperSeg. + +# 1. Introduction + +Visual segmentation is one of the most significant tasks in computer vision research, which aims to perform accurate pixel-level semantic understanding. Many specialist models [7, 17, 19, 22] have made great progress in specific segmentation tasks while showing limitations in handling diverse and complicated scenarios since new training data, paradigms, and model architectures are required to adapt to new segmentation tasks. Recent works [24, 25, 57] propose a single framework to unify diverse segmentation tasks. Despite promising, they show the inability to tackle text instructions and complex reasoning segmentation tasks needing powerful reasoning capabilities and world knowledge. + +Visual Large Language Models (VLLMs) have exhibited excellent reasoning and conversation abilities, which play a pivotal role in various vision-language co-understanding tasks [2, 8, 23, 29, 61]. However, these methods are based on rudimentary vision-language alignment, which limits their ability to comprehend finer details in visual perception tasks, like pixel-level segmentation. Recent studies [22, 45, 51, 58, 59] enables VLLMs to perform fine-grained visual understanding, like referring and reasoning segmentation. [22, 41, 45] use the special token [SEG] generated by VLLMs as the prompt for the mask decoder to generate segmentation masks while [58, 59] focus on incorporating instance-aware mask tokens into VLLMs. Though impressive, they show limitations to the universal segmentation framework based on VLLMs for both image and video domains and the capabilities of handling more complex video reasoning segmentation tasks. + +To this end, we introduce HyperSeg, the first VLLM-based universal segmentation model for pixel-level image and video perception with complex reasoning and conversation capabilities. HyperSeg can conduct diverse image and video segmentation tasks with various elaborate prompts and temporal adapter module. Besides, HyperSeg shows excellent abilities in complicated vision-language reasoning perception tasks needing rich world knowledge, which is significant for real-world understanding and interactions. As shown in Fig. 1, the explored tasks contain both image and video domains. We organize the tasks into two unified prompt formats: (1) text prompts (class names, reasoning questions, and referring languages), (2) visual prompts (box, mask, etc.). Owing to such flexible and cohesive design, HyperSeg benefits from concurrent training on diverse segmentation tasks and vision domains, facilitating the intricate correlations between different instructions and visual concepts. To further enhance fine-grained object perception and video understanding, we introduce three distinct designs in the following. + +Firstly, we incorporate a hybrid entity recognition strategy to enhance the exploitation of VLLM's recognition capacity. Generation-only works [22, 41, 49] solely rely on VLLM for + +object prediction leading to poor performance in complex multi-object segmentation scenarios. Decode-only methods [58, 59] use the prompt embedding and mask tokens decoded by VLLM to obtain class scores for each mask, which makes the mask tokens interact insufficiently with the semantic condition as they ignore the powerful generative capabilities of VLLM. The proposed hybrid entity recognition leverages the VLLM's powerful generative abilities to enhance the mask tokens' comprehension of category semantics while maintaining the final class scores decoding process. + +Secondly, previous VLLMs usually use coarse-level visual features obtained from CLIP [38] series which primarily encode global visual information while overlooking visual details. To enhance VLLMs' ability of capturing visual details efficiently, we use the Fine-grained Visual Perceiver (FVP) to merge multi-scale visual features into fixed-length fine-grained tokens, allowing retrieval of rich visual details from various scales in the hierarchical vision encoder [7]. + +Thirdly, recent VLLM-based segmentation methods [22, 58, 59] demonstrate limitations in handling video perception tasks for video temporal understanding. To this end, we propose the temporal adapter for comprehensive video perception which incorporates global prompt aggregation and local space-time information injection for the coalescence of both long-term and short-term vision-language information. + +Extensive experiments on various segmentation benchmarks demonstrate the preeminent segmentation ability of HyperSeg, providing strong evidence of the effectiveness of our insights. Our HyperSeg also exhibits promising performance on common Multi-modal benchmarks. Additionally, we explore the mutual influence among different tasks involving various visual and task types. + +Our contributions are summarized as follows: + +- We present HyperSeg, the first VLLM-based universal segmentation model for pixel-level image and video perception, covering a broad spectrum of common segmentation tasks, complex reasoning, and conversation-based vision-language understanding tasks. +- We incorporate hybrid entity recognition and fine-grained visual perceiver modules to VLLM, which allow full exploitation of VLLM's semantic recognition capacity and injection of fine-grained visual information to improve diverse detail-aware segmentation tasks. With the temporal adapter, HyperSeg can conduct more challenging video perception tasks, achieving universal segmentation. +- HyperSeg demonstrates superior capabilities on multiple segmentation tasks, achieving excellent performance on both generic and complex reasoning benchmarks with only one model. + +# 2. Related Work + +Visual Large Language Model. The emergence of Large Language Model (LLM) has significantly contributed to the development of VLMM. In this context, LLMs are enhanced with multimodal comprehension capabilities, allowing the vision-language co-understanding [1, 2, 23, 28, 29, 61]. Several notable examples of LLMs with multimodal comprehension include BLIP-2 [23], Flamingo [1], MiniGPT-4 [61], LLaVA [29], InstructBLIP [10], and Qwen-VL [2]. While these models have demonstrated impressive performance in vision-language tasks, they solely produce textual outputs that describe the entire image. This restricts their applicability in tasks that require the pixel-level detailed understanding. + +Perception with VLLM. Several methods have been proposed to enhance VLLMs with a more detailed comprehension capability [5, 22, 36, 37, 39, 41, 43, 54]. Shikra [5], Ferret [54], Kosmos-2 [36], and VisionLLM [43] are examples that provide grounding capabilities through regression of box coordinates. Conversely, LISA [22], PixelLM [41], GLaMM [39], and PerceptionGPT [37] employ a mask decoder to predict object masks from special tokens. Most of the existing methods utilize a next-token-prediction approach, which restricts their applicability. PSALM [59] makes an important attempt to bring VLLM into visual perception tasks but fails to fully unleash the potential of VLLM. In contrast, our method proposes to use a hybrid strategy to mitigate this problem and keep the capacity in high-level reasoning. + +Unified segmentation model. Another line of studies focuses on the integration of various segmentation tasks into a single model. Mask2former [7] proposes a unified architecture that requires separate training on different segmentation tasks. OpenSeeD [57] introduces a text encoder and extends it to the Open-Set setting. Simultaneously, UNINEXT [25] supports referring segmentation with the assistance of text inputs and a text encoder. However, these works fall short of following complicated instructions and reasoning. In this work, we improve the understanding ability toward language by incorporating LLM, while also maintaining the original ability of vision-centric models. + +# 3. Method + +# 3.1. Overview + +Overall architecture. The architecture of HyperSeg is illustrated in Fig. 2, which consists of a fine-grained pyramid visual encoder, a light-weight VLLM, and a segmentation predictor to generate segmentation masks, class scores, and instance embedding for video correspondence according to user's instruction. The proposed FVP module fuses multiscale high-resolution visual features $f_{img}$ into a set of fine-grained tokens to ensure the injection of fine-grained visual + +information (Sec 3.3). The VLLM takes three types of inputs: visual tokens encoded by the CLIP encoder, renewed fine-grained tokens, and prompt tokens for diverse instructions. The output embeddings of semantically enhanced mask tokens (Sec 3.2) and prompt tokens are further fed into the segmentation predictor for final segmentation results. Besides, we utilize the space-time information propagation and global prompt aggregation for comprehensive video understanding (Sec 3.4). We train the LLM with LoRA for efficient parameter tuning. + +Visual Large Language Model. We take a light-weight VLLM as our powerful multi-modal feature encoder, which contains a low-resolution vision encoder like CLIP [38] and an efficient LLM. + +Specifically, the model takes vision-prompt pairs $\{(\mathcal{V},\mathcal{P})\}$ as inputs, where $\mathcal{V}$ is resized to low resolution and then encoded by CLIP encoder $F_{CLIP}$ to get image features $f_{v}$ . The $f_{v}$ is further projected and concatenated with other task-specific tokens to ensure the comprehensive understanding of multi-modal inputs through the fusion process of LLM $F_{LLM}$ , where $G_{c}$ is the projection function and $E_O$ denotes the output embeddings of LLM. Formally, + +$$ +f _ {v} = F _ {C L I P} (\mathcal {V}), E _ {O} = F _ {L L M} \left(G _ {c} \left(f _ {v}\right), P, \mathcal {P}\right), \tag {1} +$$ + +where $P$ denotes fine-grained tokens. Furthermore, we manually extract semantic enhanced mask tokens $E_{Q}$ and prompt embedding $E_{P}$ from $E_{O}$ , which are further fed into the pretrained segmentation predictor [7] to generate masks, class scores, and instance embedding for final segmentation results. + +Prompt design. In order to accommodate the different segmentation tasks, we propose a flexible design for prompt $\mathcal{P}$ . As illustrated above, we divide $\mathcal{P}$ into two formats: text prompts and visual prompts. To be specific, $\mathcal{P}$ contains the instructions $\mathcal{P}_{\mathcal{I}}$ and task-specific conditions $\mathcal{P}_{\mathcal{C}}$ , where $\mathcal{P}_{\mathcal{I}}$ instructs the model to perform different tasks while $\mathcal{P}_{\mathcal{C}}$ indicates diverse conditions which are further used as classifiers to calculate the class scores of predicted masks. + +For class-based segmentation tasks like panoptic segmentation, open-vocabulary segmentation (OVS), and video instance segmentation (VIS), $\mathcal{P}$ can be demonstrated as $\mathcal{P}_{\mathcal{I}}$ : "Please segment all the positive objects according to the following potential categories." $\mathcal{P}_{\mathcal{C}}$ : "[category 1, category 2, category 3, ...] + +For referring and reasoning segmentation tasks like referring expression segmentation (RES), reasoning segmentation, referring video object segmentation (R-VOS), and ReasonVOS, $\mathcal{P}$ can be designed as $\mathcal{P}_{\mathcal{I}}$ : "Can you perform referring or reasoning segmentation according to the language expression?" $\mathcal{P}_{\mathcal{C}}$ : "[referring / reasoning text]" + +For visual-guided segmentation tasks like interactive segmentation and video object segmentation (VOS), $\mathcal{P}$ can be designed as $\mathcal{P}_{\mathcal{I}}$ : "Please segment according to the given + +![](images/643707cb3f024ee261eafd8b603adbc81ae0f501316c55fda3c10e361e527f08.jpg) +Figure 2. Overview of HyperSeg. HyperSeg encodes the visual input in a multi-grained manner and concatenates the prompt for different perception tasks. We feed learnable fine-grained tokens into a Fine-grained Visual Perceiver (FVP) to integrate multi-scale high-resolution image features into LLM for detailed visual learning and to facilitate space-time information propagation for video understanding. Additionally, we use the semantically enhanced mask tokens and prompt embedding to finally generate the segmentation masks and class scores for generic segmentation, and instance embedding for video instance association. + +visual region reference" $\mathcal{P}_{\mathcal{C}}$ : "[vision 1, vision 2, vision 3, ...]". Instead of using an additional region encoder to extract visual reference features [25], we sample the CLIP visual features $f_{v}$ in VLLM according to the region coordinates and perform adaptive average pooling on them to form the final reference features for each visual prompt. + +Segmentation predictor. Segmentation predictor $F_{p}$ generates the masks $m$ , corresponding class scores $z$ , and instance embedding $e$ through the similar process [7, 15] of three inputs: task-specific prompt embedding $\{E_{\mathcal{P}}^{k}\}_{k=1}^{K}$ , the semantically enhanced mask tokens $\{E_{Q}^{j}\}_{j=1}^{N}$ and the multi-scale visual features $f_{img}$ , where $K$ and $N$ denote $K$ categories and $N$ mask proposals. Formally, + +$$ +\left\{m _ {j}, z _ {j}, e _ {j} \right\} _ {j = 1} ^ {N} = F _ {p} \left(\left\{E _ {\mathcal {P}} ^ {k} \right\} _ {k = 1} ^ {K}, \left\{E _ {\mathcal {Q}} ^ {j} \right\} _ {j = 1} ^ {N}, f _ {i m g}\right), \tag {2} +$$ + +where $m_j \in \mathbb{R}^{H \times W}$ is the j-th mask proposal, $z_j \in \mathbb{R}^K$ denotes the class scores of $m_j$ , and $e_j \in \mathbb{R}^D$ denotes the j-th instance embedding obtained from an extra embedding head only for video domain. For video tasks, we adopt a frame-by-frame manner to get frame-level segmentation results for efficient training and inference processes. + +Training objectives. The model can be trained jointly on multiple tasks using the unified loss $\mathcal{L}$ . Specifically, we employ an autoregressive cross-entropy loss $\mathcal{L}_{text}$ for text prediction, a combination of per-pixel binary cross-entropy loss $\mathcal{L}_{bce}$ and DICE loss $\mathcal{L}_{dice}$ for mask supervision $\mathcal{L}_{mask}$ , a cross-entropy loss $\mathcal{L}_{cls}$ for category classification, and a contrastive loss $\mathcal{L}_{ins}$ for instance association of video + +sequences following [48]. $\lambda$ indicates their sum weight respectively. Formally, + +$$ +\mathcal {L} = \mathcal {L} _ {\text {t e x t}} + \lambda_ {\text {m a s k}} \mathcal {L} _ {\text {m a s k}} + \lambda_ {\text {c l s}} \mathcal {L} _ {\text {c l s}} + \lambda_ {\text {i n s}} \mathcal {L} _ {\text {i n s}}, \tag {3} +$$ + +$$ +\mathcal {L} _ {\text {m a s k}} = \lambda_ {b c e} \mathcal {L} _ {b c e} + \lambda_ {d i c e} \mathcal {L} _ {d i c e}, \tag {4} +$$ + +Differences between HyperSeg and previous methods. Previous universal segmentation methods [24, 25, 31] lacking of VLLMs show inability in reasoning perception tasks while our HyperSeg demonstrates brilliant reasoning segmentation capability in complex scenarios. Besides, we make a significant generalization of the current VLLM-based segmentation methods [22, 51, 58, 59] for more diverse segmentation tasks in both image and video domains using a single model framework. Moreover, HyperSeg differs from previous methods in the three designs elaborated in the following sections. + +# 3.2. Hybrid Entity Recognition + +As shown in Fig. 3 (a), predicting presented objects in the way of sequence generation (semantic prediction) tends to miss objects or produce repetitive predictions [45]. On the other hand, Fig. 3 (b), only using VLLM to embed class names (prompt tokens) as mask classifier at the decode stage disregards VLLM's powerful semantic recognition capability. Consequently, we propose a hybrid approach that leverages LLM in both generation and decoding processes. + +Instead of integrating mask tokens in input sequences and extracting the corresponding embedding from the one-pass + +![](images/616d3b0e74640482eaf15a74f78b6ab8a220081ac36f42acbfb171dd6c63b94c.jpg) +(a) + +![](images/9237b9284212ed2d8c78097eb968b9f8e7e7166d4be46a6b96974a00c064f644.jpg) +(b) + +![](images/2bdcbae716cfdc5eeb53754d3febe9800e8d3fd80091a509c0208f98438ab8d6.jpg) +(c) + +![](images/cca5f1eb172bf49413a0bbe6f89b7437e763d28cffb3eaadeee2aa3505fdc45f.jpg) +Figure 3. The comparison of different recognition strategies. (a) Generation-Only [22, 41]: both the semantic recognition (existing objects) and their mask tokens are generated by LLM. (b) Decode-Only [58, 59]: prompt embedding and mask tokens are decoded from LLM. The present objects are then determined by their similarity scores. (c) Hybrid (ours): prompt embedding is decoded from LLM while the semantically enhanced mask tokens are generated by LLM. Their similarity scores reflect the objects' presence. +(a) + +![](images/5455b62a85ea2b1d9e7aac7be0d7b0e4465f1a672d63a1cd8fe80ba8b0508f73.jpg) +(b) +Figure 4. Comparison between previous vision perceiver and our FVP. (a): previous vision perceiver [2, 23] uses the coarse single-scale CLIP visual features which are inadequate for fine-grained perception tasks. (b): FVP encodes the multi-scale visual features into fine-grained tokens. + +forward output of VLLM, we instruct VLLM to generate the mask tokens preceded by the estimated objects' names. As illustrated in Fig. 3 (c), VLLM is compelled to generate all the existing objects in the vision input and then the mask tokens. The semantically enhanced mask tokens contain valuable semantic integrated information about the image, which are subsequently used as input for the segmentation predictor to generate segmentation masks. + +# 3.3. Fine-grained Visual Perceiver + +Why twin-tower vision encoder? As shown in Fig. 4, previous VLLMs and VLLM-based segmentation methods usually utilize the pre-trained CLIP encoder to obtain single-scale and low-resolution vision features interacted with diverse languages, which is insufficient for fine-grained image and video segmentation tasks. Therefore, we adopt an extra pyramid vision encoder [7] to inject details-aware visual information. + +Specifically, we fuse multi-scale visual features into fine-grained tokens (stated as $P$ in Sec 3.1) which can inject rich fine-grained visual information into the pre-trained VLLMs without excessive computation cost. Formally, given the + +vision input $\nu$ , we leverage a pyramid vision encoder [7] $F_{seg}$ to get details-aware image features $f_{img}$ . For the j-th scale and the previous fine-grained tokens $P_{j-1}$ , the FVP module enriches each token through conditional weighted cross-attention: + +$$ +\hat {P} _ {j} = \operatorname {M H C A} \left(P _ {j - 1}, G _ {p} \left(f _ {\text {i m g}} ^ {(j)}\right)\right), \tag {5} +$$ + +$$ +P _ {j} = P _ {j - 1} + \tanh (\mathrm {M L P} (\hat {P} _ {j})) \cdot \hat {P} _ {j}, \tag {6} +$$ + +where MHCA denotes the Multi-Head Cross-Attention layer, $G_{p}$ is the projection function, tanh is a normalization function and MLP is a Multilayer Perceptron. The component of $\tanh(\mathrm{MLP}(\hat{P}_j))$ is the conditional weight used to multiply the enriched fine-grained tokens $\hat{P}_j$ before the residual connection to the previous tokens $P_{j - 1}$ . Additionally, we initialize the weight value to zero to ensure the adaptation to diverse multi-scale image features while retaining the training stability. + +# 3.4. Temporal Adapter + +Video segmentation entails distinct challenges, requiring reasoning across multiple frames and the maintenance of temporal coherence. Existing VLLM-based methods exhibit limitations in addressing video perception tasks and lack specialized designs for comprehending temporal dynamics in video analysis. To this end, we utilize global prompt aggregation and local space-time information injection in the time dimension to adapt to more complicated video perception tasks. + +Global prompt aggregation. For the current prompt embedding $E_{\mathcal{P}}$ in the video object mask retrieval process, we leverage the adaptive average pooling strategy along the time dimension to aggregate global object and temporal information of previous $T$ frames. + +$$ +E _ {\mathcal {P}} = \operatorname {A v g P o o l} ([ E _ {\mathcal {P}} ^ {0}, E _ {\mathcal {P}} ^ {1}, \dots , E _ {\mathcal {P}} ^ {T} ]), \tag {7} +$$ + +Local space-time information injection. We propose a sequential renewal strategy for space-time information prop + +Table 1. Comparison with the state-of-the-art models on the RefCOCO series and more challenging generalized referring expression segmentation benchmark gRefCOCO. $\ddagger$ denotes models using pre-trained SAM [21] for mask generation. * means using gRefCOCO for training while other methods are evaluated in zero-shot manners. + +
TypeMethodRefCOCORefCOCO+RefCOCOggRefCOCO
valtestAtestBvaltestAtestBval(U)test(U)valtestAtestB
SegmentationVLT [11]67.570.565.256.361.050.155.057.752.5*62.2*50.5*
CRIS [44]70.573.266.162.368.153.759.960.455.3*63.8*51.0*
LAVT [53]72.775.868.862.168.455.161.262.157.6*65.3*55.0*
PolyFormer-B [30]74.876.671.167.672.959.367.869.1---
VLLM-basedLISA-7B [22]‡74.979.172.365.170.858.167.970.638.7*52.6*44.8*
PixelLM-7B [41]73.076.568.266.371.758.369.370.5---
GSVA-7B [49]‡76.477.472.864.567.758.671.172.061.7*69.2*60.3*
GroundHog-7B [33]78.579.975.770.575.064.974.174.666.7*--
SAM4MLLM-7B [6]‡79.682.876.173.577.865.874.575.666.3*70.1*63.2*
Segmentation NetworkLaSagnA-7B [45]‡76.878.773.866.470.660.170.671.938.150.442.1
OMG-LLaVA [58]78.080.374.169.173.163.072.972.9---
GLaMM [40]‡79.583.276.972.678.764.674.274.9---
PSALM [59]83.684.781.672.975.570.173.874.442.052.450.6
HyperSeg84.885.783.479.083.575.279.478.947.557.352.5
+ +Table 2. Comparison with the state-of-the-art models on more complex and challenging reasoning segmentation benchmarks: ReVOS in video domain and ReasonSeg in image domain. $\ddagger$ denotes the same meaning as Tab. 1. + +
MethodBackboneReVOS-ReasoningReVOS-ReferringReVOS-OverallReasonSeg
JFJ&FJFJ&FJFJ&FgIoUcIoU
LMPM [12]Swin-T13.324.318.829.039.134.121.231.726.4--
ReferFormer [47]Video-Swin-B21.325.623.431.234.332.726.229.928.1--
LISA-7B [22]‡ViT-H33.838.436.144.347.145.739.142.740.952.954.0
LaSagnA-7B [45]‡ViT-H---------48.847.2
SAM4MLLM-7B [6]‡EfficientViT-SAM-XL1---------46.748.1
TrackGPT-13B [62]‡ViT-H38.142.940.548.350.649.543.246.845.0--
VISA-7B [51]‡ViT-H36.741.739.251.154.752.943.948.246.152.757.8
VISA-13B [51]‡ViT-H38.343.540.952.355.854.145.349.747.5--
HyperSeg-3BSwin-B50.255.853.056.060.958.553.158.455.759.256.7
+ +agation based on fine-grained tokens $P$ to inject object information of adjacent frames. Formally, + +$$ +P _ {t} = G _ {l} \left[ F _ {L L M} \left(P _ {t - 1}\right) \right], \tag {8} +$$ + +where $P_{t}$ denotes the time-aware fine-grained tokens of the current $t$ -th frame, $G_{l}$ is the projection function to transfer the previous features to the current space and align the feature dimensions. + +The proposed global prompt aggregation and local spacetime information injection within our temporal adapter facilitate the coalescence of both long-term and short-term vision-language information, which is essential for comprehensive video perception. More details about the temporal adapter are provided in the Supplementary Material. + +# 4. Experiments + +Datasets. We use the one-stage training strategy to train HyperSeg with the multi-dataset and multi-task manners. + +For image segmentation, we use COCO Panoptic [26], RefCOCO series [35, 55], COCO-Interactive, and ReasonSeg [22]. For video segmentation, we utilize DAVIS-2017 datasets [4], Ref-Youtu-VOS [42], YouTube-VIS 2019 [52], and ReVOS [51]. Besides, we use LLAVA-150k [29] to maintain the vision-language conversation capability of VLLM (we show the results on Multi-modal benchmarks in the Supplementary Material). + +Implementation details We load the pre-trained weights of Mipha [63] for our VLLM, and Maks2Former [7] for our segmentation predictor. We use three layers of FVP for fine-grained information fusion and utilize LoRA [18] to finetune the LLM efficiently. We train HyperSeg jointly for 160k iterations using a batch size of 32 on 8 NVIDIA A100 GPUs, which means each task takes approximately 16k iterations. We employ the AdamW optimizer with a learning rate of $4 \times 10^{-5}$ and with a cosine schedule. All the hyper-parameters in the loss $\mathcal{L}$ are assigned values 1.0. + +Table 3. Quantitative results on the closed-set COCO-Panoptic segmentation, open-vocabulary segmentation (-OV) benchmarks. + +
TypeMethodBackboneCOCO-PanopticADE-OVCitys-OVPC59-OVPAS20-OV
PQmIoUPQmIoUPQmIoUmIoU
Segmentation SpecialistMask2former [7]Swin-B55.165.1-----
OneFormer [19]Swin-L57.967.4-----
SEEM [64]DaViT-B56.166.3-----
MaskCLIP [13]ViT-L30.947.615.123.7-45.9-
DeOP [16]ResNet-101c---22.9-48.891.7
SimBaseline [50]ViT-B---20.5-47.788.4
DaTaSeg [15]ViTDet-B52.862.712.318.328.051.1-
VLLM-based Segmentation NetworkOMG-LLaVA [58]ConvNeXt-L53.8------
PSALM [22]Swin-B55.966.613.718.228.848.581.3
HyperSegSwin-B61.277.216.122.331.164.692.1
+ +Table 4. Results of common video segmentation benchmarks, including DAVIS17, Ref-YouTube-VOS, Ref-DAVIS17, and YouTube-VIS 2019. $\ddagger$ denotes the same meaning as Tab. 1. + +
MethodBackboneDAVIS17Ref-YTRef-DAVISYT-VIS
J&FJ&FJ&FmAP
SEEM [64]DaViT-B62.8---
OMG-Seg [24]ConvNeXt-L74.3--56.4
ReferFormer [47]Video-Swin-B-62.961.1-
OnlineRefer [46]Swin-L-63.564.8-
UNINEXT [25]ConvNeXt-L77.266.266.764.3
LISA-7B [22]‡ViT-H-53.964.8-
VISA-13B [51]‡ViT-H-63.070.4-
VideoLISA-3.8B [3]‡ViT-H-63.768.8-
HyperSeg-3BSwin-B77.668.571.253.8
+ +# 4.1. Comparisons with State-of-the-Arts + +Referring expression segmentation results. We compare HyperSeg with the state-of-the-art methods on the benchmarks RefCOCO+/g [35, 55] and more challenging generalized referring expression segmentation benchmark gRefCOCO [27]. in Tab. 1. Based on the versatile and adaptable design of HyperSeg, our model achieves state-of-the-art performance on all the referring datasets. Specifically, HyperSeg surpasses the current SOTA by a large margin, reaching 79.7 cIoU on RefCOCO+ val (+6.8 over PSALM). Besides, Our model shows superiority in challenging G-RES tasks compared with previous zero-shot methods, demonstrating the robustness and generalization ability of HyperSeg. + +Reasoning segmentation results. We compare HyperSeg with the state-of-the-art methods on image reasoning segmentation (ReasonSeg [22]) and reasoning video object segmentation (ReVOS [51]) in Tab. 2. Our HyperSeg achieves superior performance on reasoning tasks, significantly surpassing previous state-of-the-art methods (+12.1 on ReVOS-Reasoning), which shows HyperSeg powerful reasoning capability of tackling complex scenarios. + +Generic image segmentation results. We show the performance of HyperSeg on COCO-Panoptic [26] and open-vocabulary segmentation [9, 14, 34, 60] tasks in Tab. 3. HyperSeg achieves excellent performance compared with both specialist models and VLLMs-based methods on both closed- + +set and open-vocabulary segmentation tasks. Specifically, HyperSeg surpasses the VLLM-based PSALM by a significant margin (+5.3 on COCO PQ, and +10.6 on mIoU), which demonstrates our powerful capabilities of handling complex semantic perception and segmentation tasks. Besides, we show the results of COCO-Interactive in the Supplementary Material. + +Common video segmentation results. We compare HyperSeg with previous video segmentation methods in Tab. 4, including visual-prompted semi-supervised VOS (DAVIS17 val), text-prompted referring video object segmentation (RefYouTube-VOS, Ref-DAVIS17) and video instance segmentation (YouTube-VIS 2019). HyperSeg shows promising results over previous unified segmentation methods [24, 25]. Besides, HyperSeg performs more video perception tasks than previous VLLM-based models [3, 51]. + +# 4.2. Ablations + +The mutual influence between different tasks. Our model can be trained and inferred across multiple tasks and datasets simultaneously. We evaluate the mutual impact of different tasks in Tab. 5. The results show that joint training can enhance the model performance compared with the task-specific model. Besides, the performance of video segmentation tasks can be improved significantly by adding the image training datasets. This demonstrates the generalization and self-consistency of our HyperSeg to perform universal segmentation. + +Effect of different LLMs and vision backbone. In Tab. 6, we evaluate the effect of different sizes of LLMs and vision backbone. Our HyperSeg achieves excellent performance using smaller LLMs and vision encoder compared with the previous SOTA models like VISA[51] and PSALM[59]. Besides, the performance of HyperSeg can be further improved by using the more powerful LLM (Phi-2-2.7B [20]). + +Ablation on the proposed components. We assess the effectiveness of our proposed FVP module and Hybrid Entity Recognition strategy. As shown in Tab. 7, with our fine-grained visual integration and hybrid entity semantic + +Table 5. The mutual influence between different tasks. Task-specific means training task-specific models only on data from corresponding tasks, Refer+Reason denotes the model is trained on referring and reasoning segmentation data. + +
Task-specificRefer+ReasonVideoImageRefCOCOCOCOReVOSYT-VIS
valtestAtestBPQmIoUReasoningReferringOverallmAP
✓83.885.982.260.875.151.256.653.950.7
✓83.384.980.9--53.157.355.2-
✓85.686.182.460.976.5----
✓-----51.157.054.150.4
✓✓✓84.885.783.461.277.253.058.555.753.8
+ +Table 6. The comparison of different LLMs and backbone usages. w/o CLIP means without using CLIP vision encoder. + +
MethodLLMCOCOReVOSADE-OVPC59-OVPAS20-OV
PQmIoUReasoningReferringOverallmIoUmIoUmIoU
LISA [22]Vicuna-7B--36.145.740.9---
VISA [51]Vicuna-13B--40.954.147.5---
PSALM(w/o CLIP) [22]Phi-1.5-1.3B55.966.6---18.248.581.3
HyperSeg (w/o CLIP)Phi-1.5-1.3B61.176.044.049.746.918.960.090.6
HyperSegPhi-1.5-1.3B60.976.750.857.053.920.361.590.8
HyperSegPhi-2-2.7B61.277.253.058.555.722.364.692.1
+ +Table 7. Ablation on the core components of HyperSeg. FVP and HER denote the proposed Fine-grained Visual Perceiver and Hybrid Entity Recognition modules. + +
FVPHERYT-VISCOCORefCOCO
mAPPQmIoUcIoU
✓48.454.866.282.8
50.855.866.684.6
✓✓52.059.774.684.3
✓✓53.861.277.284.8
+ +Table 8. Ablation on the Fine-grained Visual Perceiver design. CW denotes the Conditional Weight illustrated in Sec. 3.3, and Scale denotes the total scale in the proposed FVP module. + +
CWScaleYT-VISCOCORefCOCO
mAPPQmIoUcIoU
✓single-layer49.755.868.083.7
multi-layers50.458.973.484.5
multi-layers53.861.277.284.8
+ +enhancement, the segmentation accuracy can be enhanced significantly $(+5.4$ on YT-VIS, $+6.4$ on COCO panoptic PQ). + +Design of the Fine-grained Visual Perceiver. In the FVP module, we combine multi-scale visual features into fixed perception queries using the condition-wise cross-attention layers to extract rich visual details from different scales of the pyramid encoder. As shown in Tab. 8, together with the conditional weight and the multi-scale design, our model makes a significant improvement on both image and video segmentation tasks. + +Effect of temporal adapter. We evaluate the effectiveness of the proposed temporal adapter including global prompt ag + +Table 9. Ablation on the temporal adapter for video tasks. + +
GlobalLocalRef-DAVIS17ReVOSYT-VIS
J&FJ&FmAP
✓67.354.147.9
68.854.548.5
✓69.354.850.2
✓✓71.255.753.8
+ +gregation (global) and local space-time information injection (local) in Tab. 9. Incorporating both global and local components, the temporal adapter significantly enhances model performance across multiple video segmentation tasks. + +# 5. Conclusion + +In this study, we aim to present HyperSeg, the first VLLM-based universal segmentation model designed for pixel-level image and video perception, encompassing a wide range of generic segmentation and complex reasoning tasks. We propose the Hybrid Entity Recognition and Fine-grained Visual Perceiver to leverage the recognition capacity of VLLMs more effectively and enhances the VLLM's ability by capturing diverse levels of visual information without incurring excessive computational costs. With additional Temporal Adapter, HyperSeg can tackle challenging video tasks by incorporating global and local information. HyperSeg surpasses existing methods on complex reasoning segmentation and traditional perception tasks. The insights presented in this work expand the possibilities of VLLMs in visual perception and lay a foundation for future research on the integration of vision-language models. + +# 6. Acknowledgments + +This work was partly supported by the National Key Research and Development Program of China (No. 2024YFB2808903) and the Shenzhen Science and Technology Program (JCYJ20220818101001004). + +# References + +[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 3 +[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 2, 3, 5, 1 +[3] Zechen Bai, Tong He, Haiyang Mei, Pichao Wang, Ziteng Gao, Joya Chen, Lei Liu, Zheng Zhang, and Mike Zheng Shou. One token to seg them all: Language instructed reasoning segmentation in videos. arXiv preprint arXiv:2409.19603, 2024. 7 +[4] Sergi Caelles, Alberto Montes, Kevis-Kokitsi Maninis, Yuhua Chen, Luc Van Gool, Federico Perazzi, and Jordi Pont-Tuset. The 2018 davis challenge on video object segmentation. arXiv preprint arXiv:1803.00557, 2018. 6 +[5] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 3, 1 +[6] Yi-Chia Chen, Wei-Hua Li, Cheng Sun, Yu-Chiang Frank Wang, and Chu-Song Chen. Sam4mllm: Enhance multimodal large language model for referring expression segmentation. In European Conference on Computer Vision, pages 323-340. Springer, 2025. 6 +[7] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 2, 3, 4, 5, 6, 7 +[8] Jaemin Cho, Jie Lei, Hao Tan, and Mohit Bansal. Unifying vision-and-language tasks via text generation. In International Conference on Machine Learning, pages 1931-1942. PMLR, 2021. 2 +[9] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 7 +[10] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning, 2023. 3, 1 + +[11] Henghui Ding, Chang Liu, Suchen Wang, and Xudong Jiang. Vision-language transformer and query generation for referring segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16321-16330, 2021. 6, 3 +[12] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, and Chen Change Loy. Mevis: A large-scale benchmark for video segmentation with motion expressions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2694-2703, 2023. 6 +[13] Zheng Ding, Jieke Wang, and Zhuowen Tu. Open-vocabulary universal image segmentation with maskclip. arXiv preprint arXiv:2208.08984, 2022. 7 +[14] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-338, 2010. 7 +[15] Xiuye Gu, Yin Cui, Jonathan Huang, Abdullah Rashwan, Xuan Yang, Xingyi Zhou, Golnaz Ghiasi, Weicheng Kuo, Huizhong Chen, Liang-Chieh Chen, et al. Dataseg: Taming a universal multi-dataset multi-task segmentation model. Advances in Neural Information Processing Systems, 36, 2024. 4, 7 +[16] Cong Han, Yujie Zhong, Dengjie Li, Kai Han, and Lin Ma. Open-vocabulary semantic segmentation with decoupled one-pass network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1086-1096, 2023. 7 +[17] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 2 +[18] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 6 +[19] Jitesh Jain, Jiachen Li, Mang Tik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. Oneformer: One transformer to rule universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2989-2998, 2023. 2, 7, 3 +[20] Mojan Javaheripi, Sébastien Bubeck, Marah Abdin, Jyoti Aneja, Sebastien Bubeck, Caio César Teodoro Mendes, Weizhu Chen, Allie Del Giorno, Ronen Eldan, Sivakanth Gopi, et al. Phi-2: The surprising power of small language models. *Microsoft Research Blog*, 1:3, 2023. 7, 1 +[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 6, 1, 2 +[22] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. ArXiv, abs/2308.00692, 2023. 2, 3, 4, 5, 6, 7, 8 +[23] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 2, 3, 5, 1 + +[24] Xiangtai Li, Haobo Yuan, Wei Li, Henghui Ding, Size Wu, Wenwei Zhang, Yining Li, Kai Chen, and Chen Change Loy. Omg-seg: Is one model good enough for all segmentation? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27948-27959, 2024. 2, 4, 7, 3 +[25] Fangjian Lin, Jianlong Yuan, Sitong Wu, Fan Wang, and Zhibin Wang. Uninext: Exploring a unified architecture for vision recognition. In Proceedings of the 31st ACM International Conference on Multimedia, pages 3200-3208, 2023. 2, 3, 4, 7 +[26] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision, 2014. 6, 7 +[27] Chang Liu, Henghui Ding, and Xudong Jiang. Gres: Generalized referring expression segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23592-23601, 2023. 7 +[28] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 3 +[29] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2, 3, 6, 1 +[30] Jiang Liu, Hui Ding, Zhaowei Cai, Yuting Zhang, Ravi Kumar Satzoda, Vijay Mahadevan, and R Manmatha. Polyformer: Referring image segmentation as sequential polygon generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18653-18663, 2023. 6, 3 +[31] Yong Liu, Cairong Zhang, Yitong Wang, Jiahao Wang, Yujiu Yang, and Yansong Tang. Universal segmentation at arbitrary granularity with language instruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3459-3469, 2024. 4 +[32] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 1 +[33] Bo Miao, Mohammed Bennamoun, Yongsheng Gao, and Ajmal Mian. Spectrum-guided multi-granularity referring video object segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 920-930, 2023. 6 +[34] Roozbeh Mottaghi, Xianjie Chen, Xiaobai Liu, Nam-Gyu Cho, Seong-Whan Lee, Sanja Fidler, Raquel Urtasun, and Alan Yuille. The role of context for object detection and semantic segmentation in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 891–898, 2014. 7 +[35] Varun K Nagaraja, Vlad I Morariu, and Larry S Davis. Modeling context between objects for referring expression understanding. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 792-807. Springer, 2016. 6, 7 + +[36] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 3 +[37] Renjie Pi, Lewei Yao, Jiahui Gao, Jipeng Zhang, and Tong Zhang. Perceptiongpt: Effectively fusing visual perception into llm. arXiv preprint arXiv:2311.06612, 2023. 3 +[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3 +[39] Hanoona Rasheed, Muhammad Maaz, Sahal Shaji, Abdelrahman Shaker, Salman Khan, Hisham Cholakkal, Rao M Anwer, Erix Xing, Ming-Hsuan Yang, and Fahad S Khan. Glamm: Pixel grounding large multimodal model. arXiv preprint arXiv:2311.03356, 2023. 3 +[40] Hanoona Rasheed, Muhammad Maaz, Sahal Shaji, Abdelrahman Shaker, Salman Khan, Hisham Cholakkal, Rao M Anwer, Eric Xing, Ming-Hsuan Yang, and Fahad S Khan. Glamm: Pixel grounding large multimodal model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13009-13018, 2024. 6 +[41] Zhongwei Ren, Zhicheng Huang, Yunchao Wei, Yao Zhao, Dongmei Fu, Jiashi Feng, and Xiaojie Jin. Pixel reasoning with large multimodal model. ArXiv, abs/2312.02228, 2023. 2, 3, 5, 6 +[42] Seonguk Seo, Joon-Young Lee, and Bohyung Han. Urvos: Unified referring video object segmentation network with a large-scale benchmark. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XV 16, pages 208–223. Springer, 2020. 6 +[43] Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong Lu, Jie Zhou, Yu Qiao, et al. Vision: Large language model is also an open-ended decoder for vision-centric tasks. Advances in Neural Information Processing Systems, 36, 2024. 3 +[44] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clipdriven referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11686-11695, 2022. 6 +[45] Cong Wei, Haoxian Tan, Yujie Zhong, Yujiu Yang, and Lin Ma. Lasagna: Language-based segmentation assistant for complex queries. arXiv preprint arXiv:2404.08506, 2024. 2, 4, 6, 3 +[46] Dongming Wu, Tiancai Wang, Yang Zhang, Xiangyu Zhang, and Jianbing Shen. Onlinerefer: A simple online baseline for referring video object segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2761-2770, 2023. 7, 3 +[47] Jiannan Wu, Yi Jiang, Peize Sun, Zehuan Yuan, and Ping Luo. Language as queries for referring video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4974-4984, 2022. 6, 7, 3 + +[48] Junfeng Wu, Qihao Liu, Yi Jiang, Song Bai, Alan Yuille, and Xiang Bai. In defense of online models for video instance segmentation. In ECCV, 2022. 4 +[49] Zhuofan Xia, Dongchen Han, Yizeng Han, Xuran Pan, Shiji Song, and Gao Huang. Gsva: Generalized segmentation via multimodal large language models. arXiv preprint arXiv:2312.10103, 2023. 2, 6, 3 +[50] Mengde Xu, Zheng Zhang, Fangyun Wei, Yutong Lin, Yue Cao, Han Hu, and Xiang Bai. A simple baseline for open-vocabulary semantic segmentation with pre-trained vision-language model. In European Conference on Computer Vision, pages 736-753. Springer, 2022. 7 +[51] Cilin Yan, Haochen Wang, Shilin Yan, Xiaolong Jiang, Yao Hu, Guoliang Kang, Weidi Xie, and Efstratios Gavves. Visa: Reasoning video object segmentation via large language models. arXiv preprint arXiv:2407.11325, 2024. 2, 4, 6, 7, 8, 3 +[52] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5188-5197, 2019. 6 +[53] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. Levt: Language-aware vision transformer for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18155-18165, 2022. 6, 3 +[54] Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and Yinfei Yang. Ferret: Refer and ground anything anywhere at any granularity. arXiv preprint arXiv:2310.07704, 2023. 3 +[55] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 69-85. Springer, 2016. 6, 7 +[56] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 1 +[57] Hao Zhang, Feng Li, Xueyan Zou, Shilong Liu, Chunyuan Li, Jianwei Yang, and Lei Zhang. A simple framework for open-vocabulary segmentation and detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1020-1031, 2023. 2, 3 +[58] Tao Zhang, Xiangtai Li, Hao Fei, Haobo Yuan, Shengqiong Wu, Shunping Ji, Chen Change Loy, and Shuicheng Yan. Omg-llava: Bridging image-level, object-level, pixel-level reasoning and understanding. arXiv preprint arXiv:2406.19389, 2024. 2, 4, 5, 6, 7, 3 +[59] Zheng Zhang, Yeyao Ma, Enming Zhang, and Xiang Bai. Psalm: Pixelwise segmentation with large multi-modal model. arXiv preprint arXiv:2403.14598, 2024. 2, 3, 4, 5, 6, 7, 1 +[60] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision, 127(3):302-321, 2019. 7 +[61] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language + +understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023.2, 3 +[62] Jiawen Zhu, Zhi-Qi Cheng, Jun-Yan He, Chenyang Li, Bin Luo, Huchuan Lu, Yifeng Geng, and Xuansong Xie. Tracking with human-intent reasoning. arXiv preprint arXiv:2312.17448, 2023. 6 +[63] Minjie Zhu, Yichen Zhu, Xin Liu, Ning Liu, Zhiyuan Xu, Chaomin Shen, Yaxin Peng, Zhicai Ou, Feifei Feng, and Jian Tang. A comprehensive overhaul of multimodal assistant with small language models. arXiv preprint arXiv:2403.06199, 2024. 6 +[64] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Wang, Lijuan Wang, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. Advances in Neural Information Processing Systems, 36, 2024. 7, 1, 2, 3 \ No newline at end of file diff --git a/2025/HyperSeg_ Hybrid Segmentation Assistant with Fine-grained Visual Perceiver/images.zip b/2025/HyperSeg_ Hybrid Segmentation Assistant with Fine-grained Visual Perceiver/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ad38ca8a273a0d9c51b8b29809411b026fee663a --- /dev/null +++ b/2025/HyperSeg_ Hybrid Segmentation Assistant with Fine-grained Visual Perceiver/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8c05a3a7881ec2ca2ed62309e1d9cf846d905923b71670466ffda38e1ca59983 +size 800009 diff --git a/2025/HyperSeg_ Hybrid Segmentation Assistant with Fine-grained Visual Perceiver/layout.json b/2025/HyperSeg_ Hybrid Segmentation Assistant with Fine-grained Visual Perceiver/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3d6495964d542609e0515c0d89d1bffeaf8ea1f7 --- /dev/null +++ b/2025/HyperSeg_ Hybrid Segmentation Assistant with Fine-grained Visual Perceiver/layout.json @@ -0,0 +1,8780 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 65, + 103, + 545, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 103, + 545, + 121 + ], + "spans": [ + { + "bbox": [ + 65, + 103, + 545, + 121 + ], + "type": "text", + "content": "HyperSeg: Hybrid Segmentation Assistant with Fine-grained Visual Perceiver" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 142, + 566, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 142, + 566, + 171 + ], + "spans": [ + { + "bbox": [ + 61, + 142, + 566, + 171 + ], + "type": "text", + "content": "Cong Wei" + }, + { + "bbox": [ + 61, + 142, + 566, + 171 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 61, + 142, + 566, + 171 + ], + "type": "text", + "content": ", Yujie Zhong" + }, + { + "bbox": [ + 61, + 142, + 566, + 171 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 61, + 142, + 566, + 171 + ], + "type": "text", + "content": ", Haoxian Tan" + }, + { + "bbox": [ + 61, + 142, + 566, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 142, + 566, + 171 + ], + "type": "text", + "content": ", Yong Liu" + }, + { + "bbox": [ + 61, + 142, + 566, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 142, + 566, + 171 + ], + "type": "text", + "content": ", Jie Hu" + }, + { + "bbox": [ + 61, + 142, + 566, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 142, + 566, + 171 + ], + "type": "text", + "content": ", Dengjie Li" + }, + { + "bbox": [ + 61, + 142, + 566, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 142, + 566, + 171 + ], + "type": "text", + "content": ", Zheng Zhao" + }, + { + "bbox": [ + 61, + 142, + 566, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 142, + 566, + 171 + ], + "type": "text", + "content": ", Yujiu Yang" + }, + { + "bbox": [ + 61, + 142, + 566, + 171 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 61, + 142, + 566, + 171 + ], + "type": "text", + "content": ", Tsinghua Shenzhen International Graduate School, Tsinghua University " + }, + { + "bbox": [ + 61, + 142, + 566, + 171 + ], + "type": "inline_equation", + "content": "{}^{2}" + }, + { + "bbox": [ + 61, + 142, + 566, + 171 + ], + "type": "text", + "content": "Meituan Inc." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 95, + 174, + 530, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 174, + 530, + 185 + ], + "spans": [ + { + "bbox": [ + 95, + 174, + 530, + 185 + ], + "type": "text", + "content": "weic22@mails.tsinghua.edu.cn, jaszhong@hotmail.com, yang.yujiu@sz.tsinghua.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 58, + 198, + 552, + 457 + ], + "blocks": [ + { + "bbox": [ + 58, + 198, + 552, + 457 + ], + "lines": [ + { + "bbox": [ + 58, + 198, + 552, + 457 + ], + "spans": [ + { + "bbox": [ + 58, + 198, + 552, + 457 + ], + "type": "image", + "image_path": "9f5949157c434b915dc18761b6461f836a0be7ebae7519358b02fb1580c66fc7.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 470, + 555, + 514 + ], + "lines": [ + { + "bbox": [ + 55, + 470, + 555, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 470, + 555, + 514 + ], + "type": "text", + "content": "Figure 1. Illustration of our HyperSeg which can conduct image and video segmentation tasks with various language and visual instructions. Additionally, HyperSeg can handle complicated reasoning perception tasks compared with previous universal segmentation methods. To our knowledge, HyperSeg is the first VLLM-based universal segmentation model with perception and complex reasoning abilities in both image and video domains." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 152, + 532, + 200, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 532, + 200, + 544 + ], + "spans": [ + { + "bbox": [ + 152, + 532, + 200, + 544 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 558, + 297, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 297, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 297, + 689 + ], + "type": "text", + "content": "This paper aims to address universal segmentation for image and video perception with the strong reasoning ability empowered by Visual Large Language Models (VLLMs). Despite significant progress in current unified segmentation methods, limitations in adaptation to both image and video scenarios, as well as the complex reasoning segmentation, make it difficult for them to handle various challenging instructions and achieve an accurate understanding of fine-grained vision-language correlations. We propose HyperSeg, the first VLLM-based universal segmentation model for pixel-level image and video perception, en" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 533, + 556, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 556, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 556, + 689 + ], + "type": "text", + "content": "compassing generic segmentation tasks and more complex reasoning perception tasks requiring powerful reasoning abilities and world knowledge. Besides, to fully leverage the recognition capabilities of VLLMs and the fine-grained visual information, HyperSeg incorporates hybrid entity recognition and fine-grained visual perceiver modules for various segmentation tasks. Combined with the temporal adapter, HyperSeg achieves a comprehensive understanding of temporal information. Experimental results validate the effectiveness of our insights in resolving universal image and video segmentation tasks, including the more complex reasoning perception tasks. Our code is available at https://github.com/congvvc/HyperSeg." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 495, + 15 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 495, + 15 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 495, + 15 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 198, + 14, + 442, + 24 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 14, + 442, + 24 + ], + "spans": [ + { + "bbox": [ + 198, + 14, + 442, + 24 + ], + "type": "text", + "content": "Except for this watermark, it is identical to the accepted version;" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "spans": [ + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "type": "text", + "content": "the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 70, + 702, + 151, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 702, + 151, + 712 + ], + "spans": [ + { + "bbox": [ + 70, + 702, + 151, + 712 + ], + "type": "text", + "content": "†Corresponding authors." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "8931" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 136, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 136, + 84 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 136, + 84 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 235 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 235 + ], + "type": "text", + "content": "Visual segmentation is one of the most significant tasks in computer vision research, which aims to perform accurate pixel-level semantic understanding. Many specialist models [7, 17, 19, 22] have made great progress in specific segmentation tasks while showing limitations in handling diverse and complicated scenarios since new training data, paradigms, and model architectures are required to adapt to new segmentation tasks. Recent works [24, 25, 57] propose a single framework to unify diverse segmentation tasks. Despite promising, they show the inability to tackle text instructions and complex reasoning segmentation tasks needing powerful reasoning capabilities and world knowledge." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 235, + 297, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 235, + 297, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 235, + 297, + 437 + ], + "type": "text", + "content": "Visual Large Language Models (VLLMs) have exhibited excellent reasoning and conversation abilities, which play a pivotal role in various vision-language co-understanding tasks [2, 8, 23, 29, 61]. However, these methods are based on rudimentary vision-language alignment, which limits their ability to comprehend finer details in visual perception tasks, like pixel-level segmentation. Recent studies [22, 45, 51, 58, 59] enables VLLMs to perform fine-grained visual understanding, like referring and reasoning segmentation. [22, 41, 45] use the special token [SEG] generated by VLLMs as the prompt for the mask decoder to generate segmentation masks while [58, 59] focus on incorporating instance-aware mask tokens into VLLMs. Though impressive, they show limitations to the universal segmentation framework based on VLLMs for both image and video domains and the capabilities of handling more complex video reasoning segmentation tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 438, + 297, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 438, + 297, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 438, + 297, + 677 + ], + "type": "text", + "content": "To this end, we introduce HyperSeg, the first VLLM-based universal segmentation model for pixel-level image and video perception with complex reasoning and conversation capabilities. HyperSeg can conduct diverse image and video segmentation tasks with various elaborate prompts and temporal adapter module. Besides, HyperSeg shows excellent abilities in complicated vision-language reasoning perception tasks needing rich world knowledge, which is significant for real-world understanding and interactions. As shown in Fig. 1, the explored tasks contain both image and video domains. We organize the tasks into two unified prompt formats: (1) text prompts (class names, reasoning questions, and referring languages), (2) visual prompts (box, mask, etc.). Owing to such flexible and cohesive design, HyperSeg benefits from concurrent training on diverse segmentation tasks and vision domains, facilitating the intricate correlations between different instructions and visual concepts. To further enhance fine-grained object perception and video understanding, we introduce three distinct designs in the following." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 677, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 297, + 713 + ], + "type": "text", + "content": "Firstly, we incorporate a hybrid entity recognition strategy to enhance the exploitation of VLLM's recognition capacity. Generation-only works [22, 41, 49] solely rely on VLLM for" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 556, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 556, + 205 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 556, + 205 + ], + "type": "text", + "content": "object prediction leading to poor performance in complex multi-object segmentation scenarios. Decode-only methods [58, 59] use the prompt embedding and mask tokens decoded by VLLM to obtain class scores for each mask, which makes the mask tokens interact insufficiently with the semantic condition as they ignore the powerful generative capabilities of VLLM. The proposed hybrid entity recognition leverages the VLLM's powerful generative abilities to enhance the mask tokens' comprehension of category semantics while maintaining the final class scores decoding process." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 212, + 556, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 212, + 556, + 308 + ], + "spans": [ + { + "bbox": [ + 313, + 212, + 556, + 308 + ], + "type": "text", + "content": "Secondly, previous VLLMs usually use coarse-level visual features obtained from CLIP [38] series which primarily encode global visual information while overlooking visual details. To enhance VLLMs' ability of capturing visual details efficiently, we use the Fine-grained Visual Perceiver (FVP) to merge multi-scale visual features into fixed-length fine-grained tokens, allowing retrieval of rich visual details from various scales in the hierarchical vision encoder [7]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 316, + 556, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 556, + 400 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 556, + 400 + ], + "type": "text", + "content": "Thirdly, recent VLLM-based segmentation methods [22, 58, 59] demonstrate limitations in handling video perception tasks for video temporal understanding. To this end, we propose the temporal adapter for comprehensive video perception which incorporates global prompt aggregation and local space-time information injection for the coalescence of both long-term and short-term vision-language information." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 408, + 557, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 408, + 557, + 492 + ], + "spans": [ + { + "bbox": [ + 313, + 408, + 557, + 492 + ], + "type": "text", + "content": "Extensive experiments on various segmentation benchmarks demonstrate the preeminent segmentation ability of HyperSeg, providing strong evidence of the effectiveness of our insights. Our HyperSeg also exhibits promising performance on common Multi-modal benchmarks. Additionally, we explore the mutual influence among different tasks involving various visual and task types." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 326, + 500, + 512, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 500, + 512, + 511 + ], + "spans": [ + { + "bbox": [ + 326, + 500, + 512, + 511 + ], + "type": "text", + "content": "Our contributions are summarized as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 320, + 520, + 555, + 711 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 320, + 520, + 555, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 520, + 555, + 579 + ], + "spans": [ + { + "bbox": [ + 320, + 520, + 555, + 579 + ], + "type": "text", + "content": "- We present HyperSeg, the first VLLM-based universal segmentation model for pixel-level image and video perception, covering a broad spectrum of common segmentation tasks, complex reasoning, and conversation-based vision-language understanding tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 320, + 580, + 555, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 580, + 555, + 662 + ], + "spans": [ + { + "bbox": [ + 320, + 580, + 555, + 662 + ], + "type": "text", + "content": "- We incorporate hybrid entity recognition and fine-grained visual perceiver modules to VLLM, which allow full exploitation of VLLM's semantic recognition capacity and injection of fine-grained visual information to improve diverse detail-aware segmentation tasks. With the temporal adapter, HyperSeg can conduct more challenging video perception tasks, achieving universal segmentation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 320, + 663, + 553, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 663, + 553, + 711 + ], + "spans": [ + { + "bbox": [ + 320, + 663, + 553, + 711 + ], + "type": "text", + "content": "- HyperSeg demonstrates superior capabilities on multiple segmentation tasks, achieving excellent performance on both generic and complex reasoning benchmarks with only one model." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8932" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 142, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 142, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 142, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 247 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 247 + ], + "type": "text", + "content": "Visual Large Language Model. The emergence of Large Language Model (LLM) has significantly contributed to the development of VLMM. In this context, LLMs are enhanced with multimodal comprehension capabilities, allowing the vision-language co-understanding [1, 2, 23, 28, 29, 61]. Several notable examples of LLMs with multimodal comprehension include BLIP-2 [23], Flamingo [1], MiniGPT-4 [61], LLaVA [29], InstructBLIP [10], and Qwen-VL [2]. While these models have demonstrated impressive performance in vision-language tasks, they solely produce textual outputs that describe the entire image. This restricts their applicability in tasks that require the pixel-level detailed understanding." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 247, + 297, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 247, + 297, + 426 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 297, + 426 + ], + "type": "text", + "content": "Perception with VLLM. Several methods have been proposed to enhance VLLMs with a more detailed comprehension capability [5, 22, 36, 37, 39, 41, 43, 54]. Shikra [5], Ferret [54], Kosmos-2 [36], and VisionLLM [43] are examples that provide grounding capabilities through regression of box coordinates. Conversely, LISA [22], PixelLM [41], GLaMM [39], and PerceptionGPT [37] employ a mask decoder to predict object masks from special tokens. Most of the existing methods utilize a next-token-prediction approach, which restricts their applicability. PSALM [59] makes an important attempt to bring VLLM into visual perception tasks but fails to fully unleash the potential of VLLM. In contrast, our method proposes to use a hybrid strategy to mitigate this problem and keep the capacity in high-level reasoning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 426, + 297, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 297, + 570 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 297, + 570 + ], + "type": "text", + "content": "Unified segmentation model. Another line of studies focuses on the integration of various segmentation tasks into a single model. Mask2former [7] proposes a unified architecture that requires separate training on different segmentation tasks. OpenSeeD [57] introduces a text encoder and extends it to the Open-Set setting. Simultaneously, UNINEXT [25] supports referring segmentation with the assistance of text inputs and a text encoder. However, these works fall short of following complicated instructions and reasoning. In this work, we improve the understanding ability toward language by incorporating LLM, while also maintaining the original ability of vision-centric models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 579, + 111, + 591 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 579, + 111, + 591 + ], + "spans": [ + { + "bbox": [ + 55, + 579, + 111, + 591 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 599, + 123, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 599, + 123, + 611 + ], + "spans": [ + { + "bbox": [ + 55, + 599, + 123, + 611 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 617, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 297, + 715 + ], + "type": "text", + "content": "Overall architecture. The architecture of HyperSeg is illustrated in Fig. 2, which consists of a fine-grained pyramid visual encoder, a light-weight VLLM, and a segmentation predictor to generate segmentation masks, class scores, and instance embedding for video correspondence according to user's instruction. The proposed FVP module fuses multiscale high-resolution visual features " + }, + { + "bbox": [ + 55, + 617, + 297, + 715 + ], + "type": "inline_equation", + "content": "f_{img}" + }, + { + "bbox": [ + 55, + 617, + 297, + 715 + ], + "type": "text", + "content": " into a set of fine-grained tokens to ensure the injection of fine-grained visual" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 555, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 191 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 191 + ], + "type": "text", + "content": "information (Sec 3.3). The VLLM takes three types of inputs: visual tokens encoded by the CLIP encoder, renewed fine-grained tokens, and prompt tokens for diverse instructions. The output embeddings of semantically enhanced mask tokens (Sec 3.2) and prompt tokens are further fed into the segmentation predictor for final segmentation results. Besides, we utilize the space-time information propagation and global prompt aggregation for comprehensive video understanding (Sec 3.4). We train the LLM with LoRA for efficient parameter tuning." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 192, + 555, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 192, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 192, + 555, + 239 + ], + "type": "text", + "content": "Visual Large Language Model. We take a light-weight VLLM as our powerful multi-modal feature encoder, which contains a low-resolution vision encoder like CLIP [38] and an efficient LLM." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 239, + 556, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 556, + 337 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 556, + 337 + ], + "type": "text", + "content": "Specifically, the model takes vision-prompt pairs " + }, + { + "bbox": [ + 313, + 239, + 556, + 337 + ], + "type": "inline_equation", + "content": "\\{(\\mathcal{V},\\mathcal{P})\\}" + }, + { + "bbox": [ + 313, + 239, + 556, + 337 + ], + "type": "text", + "content": " as inputs, where " + }, + { + "bbox": [ + 313, + 239, + 556, + 337 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 313, + 239, + 556, + 337 + ], + "type": "text", + "content": " is resized to low resolution and then encoded by CLIP encoder " + }, + { + "bbox": [ + 313, + 239, + 556, + 337 + ], + "type": "inline_equation", + "content": "F_{CLIP}" + }, + { + "bbox": [ + 313, + 239, + 556, + 337 + ], + "type": "text", + "content": " to get image features " + }, + { + "bbox": [ + 313, + 239, + 556, + 337 + ], + "type": "inline_equation", + "content": "f_{v}" + }, + { + "bbox": [ + 313, + 239, + 556, + 337 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 313, + 239, + 556, + 337 + ], + "type": "inline_equation", + "content": "f_{v}" + }, + { + "bbox": [ + 313, + 239, + 556, + 337 + ], + "type": "text", + "content": " is further projected and concatenated with other task-specific tokens to ensure the comprehensive understanding of multi-modal inputs through the fusion process of LLM " + }, + { + "bbox": [ + 313, + 239, + 556, + 337 + ], + "type": "inline_equation", + "content": "F_{LLM}" + }, + { + "bbox": [ + 313, + 239, + 556, + 337 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 239, + 556, + 337 + ], + "type": "inline_equation", + "content": "G_{c}" + }, + { + "bbox": [ + 313, + 239, + 556, + 337 + ], + "type": "text", + "content": " is the projection function and " + }, + { + "bbox": [ + 313, + 239, + 556, + 337 + ], + "type": "inline_equation", + "content": "E_O" + }, + { + "bbox": [ + 313, + 239, + 556, + 337 + ], + "type": "text", + "content": " denotes the output embeddings of LLM. Formally," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 339, + 346, + 555, + 359 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 346, + 555, + 359 + ], + "spans": [ + { + "bbox": [ + 339, + 346, + 555, + 359 + ], + "type": "interline_equation", + "content": "f _ {v} = F _ {C L I P} (\\mathcal {V}), E _ {O} = F _ {L L M} \\left(G _ {c} \\left(f _ {v}\\right), P, \\mathcal {P}\\right), \\tag {1}", + "image_path": "c9b178f83932bf7c10200b7913980b85fa3168ab88484b40d6207c9f4e7a54ad.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 367, + 556, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 367, + 556, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 367, + 556, + 437 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 367, + 556, + 437 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 367, + 556, + 437 + ], + "type": "text", + "content": " denotes fine-grained tokens. Furthermore, we manually extract semantic enhanced mask tokens " + }, + { + "bbox": [ + 313, + 367, + 556, + 437 + ], + "type": "inline_equation", + "content": "E_{Q}" + }, + { + "bbox": [ + 313, + 367, + 556, + 437 + ], + "type": "text", + "content": " and prompt embedding " + }, + { + "bbox": [ + 313, + 367, + 556, + 437 + ], + "type": "inline_equation", + "content": "E_{P}" + }, + { + "bbox": [ + 313, + 367, + 556, + 437 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 367, + 556, + 437 + ], + "type": "inline_equation", + "content": "E_{O}" + }, + { + "bbox": [ + 313, + 367, + 556, + 437 + ], + "type": "text", + "content": ", which are further fed into the pretrained segmentation predictor [7] to generate masks, class scores, and instance embedding for final segmentation results." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 440, + 556, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 440, + 556, + 535 + ], + "spans": [ + { + "bbox": [ + 313, + 440, + 556, + 535 + ], + "type": "text", + "content": "Prompt design. In order to accommodate the different segmentation tasks, we propose a flexible design for prompt " + }, + { + "bbox": [ + 313, + 440, + 556, + 535 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 313, + 440, + 556, + 535 + ], + "type": "text", + "content": ". As illustrated above, we divide " + }, + { + "bbox": [ + 313, + 440, + 556, + 535 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 313, + 440, + 556, + 535 + ], + "type": "text", + "content": " into two formats: text prompts and visual prompts. To be specific, " + }, + { + "bbox": [ + 313, + 440, + 556, + 535 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 313, + 440, + 556, + 535 + ], + "type": "text", + "content": " contains the instructions " + }, + { + "bbox": [ + 313, + 440, + 556, + 535 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathcal{I}}" + }, + { + "bbox": [ + 313, + 440, + 556, + 535 + ], + "type": "text", + "content": " and task-specific conditions " + }, + { + "bbox": [ + 313, + 440, + 556, + 535 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathcal{C}}" + }, + { + "bbox": [ + 313, + 440, + 556, + 535 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 440, + 556, + 535 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathcal{I}}" + }, + { + "bbox": [ + 313, + 440, + 556, + 535 + ], + "type": "text", + "content": " instructs the model to perform different tasks while " + }, + { + "bbox": [ + 313, + 440, + 556, + 535 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathcal{C}}" + }, + { + "bbox": [ + 313, + 440, + 556, + 535 + ], + "type": "text", + "content": " indicates diverse conditions which are further used as classifiers to calculate the class scores of predicted masks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 536, + 556, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 536, + 556, + 607 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 556, + 607 + ], + "type": "text", + "content": "For class-based segmentation tasks like panoptic segmentation, open-vocabulary segmentation (OVS), and video instance segmentation (VIS), " + }, + { + "bbox": [ + 313, + 536, + 556, + 607 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 313, + 536, + 556, + 607 + ], + "type": "text", + "content": " can be demonstrated as " + }, + { + "bbox": [ + 313, + 536, + 556, + 607 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathcal{I}}" + }, + { + "bbox": [ + 313, + 536, + 556, + 607 + ], + "type": "text", + "content": ": \"Please segment all the positive objects according to the following potential categories.\" " + }, + { + "bbox": [ + 313, + 536, + 556, + 607 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathcal{C}}" + }, + { + "bbox": [ + 313, + 536, + 556, + 607 + ], + "type": "text", + "content": ": \"[category 1, category 2, category 3, ...]" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 606, + 556, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 556, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 556, + 677 + ], + "type": "text", + "content": "For referring and reasoning segmentation tasks like referring expression segmentation (RES), reasoning segmentation, referring video object segmentation (R-VOS), and ReasonVOS, " + }, + { + "bbox": [ + 313, + 606, + 556, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 313, + 606, + 556, + 677 + ], + "type": "text", + "content": " can be designed as " + }, + { + "bbox": [ + 313, + 606, + 556, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathcal{I}}" + }, + { + "bbox": [ + 313, + 606, + 556, + 677 + ], + "type": "text", + "content": ": \"Can you perform referring or reasoning segmentation according to the language expression?\" " + }, + { + "bbox": [ + 313, + 606, + 556, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathcal{C}}" + }, + { + "bbox": [ + 313, + 606, + 556, + 677 + ], + "type": "text", + "content": ": \"[referring / reasoning text]\"" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "content": "For visual-guided segmentation tasks like interactive segmentation and video object segmentation (VOS), " + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "content": " can be designed as " + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathcal{I}}" + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "content": ": \"Please segment according to the given" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8933" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 70, + 545, + 293 + ], + "blocks": [ + { + "bbox": [ + 65, + 70, + 545, + 293 + ], + "lines": [ + { + "bbox": [ + 65, + 70, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 65, + 70, + 545, + 293 + ], + "type": "image", + "image_path": "643707cb3f024ee261eafd8b603adbc81ae0f501316c55fda3c10e361e527f08.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 301, + 555, + 357 + ], + "lines": [ + { + "bbox": [ + 54, + 301, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 54, + 301, + 555, + 357 + ], + "type": "text", + "content": "Figure 2. Overview of HyperSeg. HyperSeg encodes the visual input in a multi-grained manner and concatenates the prompt for different perception tasks. We feed learnable fine-grained tokens into a Fine-grained Visual Perceiver (FVP) to integrate multi-scale high-resolution image features into LLM for detailed visual learning and to facilitate space-time information propagation for video understanding. Additionally, we use the semantically enhanced mask tokens and prompt embedding to finally generate the segmentation masks and class scores for generic segmentation, and instance embedding for video instance association." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 369, + 295, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 369, + 295, + 441 + ], + "spans": [ + { + "bbox": [ + 54, + 369, + 295, + 441 + ], + "type": "text", + "content": "visual region reference\" " + }, + { + "bbox": [ + 54, + 369, + 295, + 441 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathcal{C}}" + }, + { + "bbox": [ + 54, + 369, + 295, + 441 + ], + "type": "text", + "content": " : \"[vision 1, vision 2, vision 3, ...]\". Instead of using an additional region encoder to extract visual reference features [25], we sample the CLIP visual features " + }, + { + "bbox": [ + 54, + 369, + 295, + 441 + ], + "type": "inline_equation", + "content": "f_{v}" + }, + { + "bbox": [ + 54, + 369, + 295, + 441 + ], + "type": "text", + "content": " in VLLM according to the region coordinates and perform adaptive average pooling on them to form the final reference features for each visual prompt." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 441, + 295, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 441, + 295, + 526 + ], + "spans": [ + { + "bbox": [ + 54, + 441, + 295, + 526 + ], + "type": "text", + "content": "Segmentation predictor. Segmentation predictor " + }, + { + "bbox": [ + 54, + 441, + 295, + 526 + ], + "type": "inline_equation", + "content": "F_{p}" + }, + { + "bbox": [ + 54, + 441, + 295, + 526 + ], + "type": "text", + "content": " generates the masks " + }, + { + "bbox": [ + 54, + 441, + 295, + 526 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 54, + 441, + 295, + 526 + ], + "type": "text", + "content": ", corresponding class scores " + }, + { + "bbox": [ + 54, + 441, + 295, + 526 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 54, + 441, + 295, + 526 + ], + "type": "text", + "content": ", and instance embedding " + }, + { + "bbox": [ + 54, + 441, + 295, + 526 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 54, + 441, + 295, + 526 + ], + "type": "text", + "content": " through the similar process [7, 15] of three inputs: task-specific prompt embedding " + }, + { + "bbox": [ + 54, + 441, + 295, + 526 + ], + "type": "inline_equation", + "content": "\\{E_{\\mathcal{P}}^{k}\\}_{k=1}^{K}" + }, + { + "bbox": [ + 54, + 441, + 295, + 526 + ], + "type": "text", + "content": ", the semantically enhanced mask tokens " + }, + { + "bbox": [ + 54, + 441, + 295, + 526 + ], + "type": "inline_equation", + "content": "\\{E_{Q}^{j}\\}_{j=1}^{N}" + }, + { + "bbox": [ + 54, + 441, + 295, + 526 + ], + "type": "text", + "content": " and the multi-scale visual features " + }, + { + "bbox": [ + 54, + 441, + 295, + 526 + ], + "type": "inline_equation", + "content": "f_{img}" + }, + { + "bbox": [ + 54, + 441, + 295, + 526 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 54, + 441, + 295, + 526 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 54, + 441, + 295, + 526 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 441, + 295, + 526 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 54, + 441, + 295, + 526 + ], + "type": "text", + "content": " denote " + }, + { + "bbox": [ + 54, + 441, + 295, + 526 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 54, + 441, + 295, + 526 + ], + "type": "text", + "content": " categories and " + }, + { + "bbox": [ + 54, + 441, + 295, + 526 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 54, + 441, + 295, + 526 + ], + "type": "text", + "content": " mask proposals. Formally," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 533, + 295, + 550 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 533, + 295, + 550 + ], + "spans": [ + { + "bbox": [ + 66, + 533, + 295, + 550 + ], + "type": "interline_equation", + "content": "\\left\\{m _ {j}, z _ {j}, e _ {j} \\right\\} _ {j = 1} ^ {N} = F _ {p} \\left(\\left\\{E _ {\\mathcal {P}} ^ {k} \\right\\} _ {k = 1} ^ {K}, \\left\\{E _ {\\mathcal {Q}} ^ {j} \\right\\} _ {j = 1} ^ {N}, f _ {i m g}\\right), \\tag {2}", + "image_path": "8c5782df5fb3709549f9fd242c978c1deade9d1c45d083c40d417368aa228405.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 556, + 295, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 556, + 295, + 630 + ], + "spans": [ + { + "bbox": [ + 54, + 556, + 295, + 630 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 54, + 556, + 295, + 630 + ], + "type": "inline_equation", + "content": "m_j \\in \\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 54, + 556, + 295, + 630 + ], + "type": "text", + "content": " is the j-th mask proposal, " + }, + { + "bbox": [ + 54, + 556, + 295, + 630 + ], + "type": "inline_equation", + "content": "z_j \\in \\mathbb{R}^K" + }, + { + "bbox": [ + 54, + 556, + 295, + 630 + ], + "type": "text", + "content": " denotes the class scores of " + }, + { + "bbox": [ + 54, + 556, + 295, + 630 + ], + "type": "inline_equation", + "content": "m_j" + }, + { + "bbox": [ + 54, + 556, + 295, + 630 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 54, + 556, + 295, + 630 + ], + "type": "inline_equation", + "content": "e_j \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 54, + 556, + 295, + 630 + ], + "type": "text", + "content": " denotes the j-th instance embedding obtained from an extra embedding head only for video domain. For video tasks, we adopt a frame-by-frame manner to get frame-level segmentation results for efficient training and inference processes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 630, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 630, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 630, + 295, + 713 + ], + "type": "text", + "content": "Training objectives. The model can be trained jointly on multiple tasks using the unified loss " + }, + { + "bbox": [ + 54, + 630, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 54, + 630, + 295, + 713 + ], + "type": "text", + "content": ". Specifically, we employ an autoregressive cross-entropy loss " + }, + { + "bbox": [ + 54, + 630, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{text}" + }, + { + "bbox": [ + 54, + 630, + 295, + 713 + ], + "type": "text", + "content": " for text prediction, a combination of per-pixel binary cross-entropy loss " + }, + { + "bbox": [ + 54, + 630, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{bce}" + }, + { + "bbox": [ + 54, + 630, + 295, + 713 + ], + "type": "text", + "content": " and DICE loss " + }, + { + "bbox": [ + 54, + 630, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{dice}" + }, + { + "bbox": [ + 54, + 630, + 295, + 713 + ], + "type": "text", + "content": " for mask supervision " + }, + { + "bbox": [ + 54, + 630, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{mask}" + }, + { + "bbox": [ + 54, + 630, + 295, + 713 + ], + "type": "text", + "content": ", a cross-entropy loss " + }, + { + "bbox": [ + 54, + 630, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cls}" + }, + { + "bbox": [ + 54, + 630, + 295, + 713 + ], + "type": "text", + "content": " for category classification, and a contrastive loss " + }, + { + "bbox": [ + 54, + 630, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ins}" + }, + { + "bbox": [ + 54, + 630, + 295, + 713 + ], + "type": "text", + "content": " for instance association of video" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 369, + 555, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 555, + 394 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 555, + 394 + ], + "type": "text", + "content": "sequences following [48]. " + }, + { + "bbox": [ + 313, + 369, + 555, + 394 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 369, + 555, + 394 + ], + "type": "text", + "content": " indicates their sum weight respectively. Formally," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 324, + 403, + 555, + 416 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 403, + 555, + 416 + ], + "spans": [ + { + "bbox": [ + 324, + 403, + 555, + 416 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {t e x t}} + \\lambda_ {\\text {m a s k}} \\mathcal {L} _ {\\text {m a s k}} + \\lambda_ {\\text {c l s}} \\mathcal {L} _ {\\text {c l s}} + \\lambda_ {\\text {i n s}} \\mathcal {L} _ {\\text {i n s}}, \\tag {3}", + "image_path": "d01ed70dd9c3478233f0e3ac634f197b61a69e1ad85764b0f8bed96e9d7cf232.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 366, + 419, + 555, + 432 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 419, + 555, + 432 + ], + "spans": [ + { + "bbox": [ + 366, + 419, + 555, + 432 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {m a s k}} = \\lambda_ {b c e} \\mathcal {L} _ {b c e} + \\lambda_ {d i c e} \\mathcal {L} _ {d i c e}, \\tag {4}", + "image_path": "8774bf849674fa292ab5e3bd4fb28c256605c088ceca12ccbb09d7b96b69304e.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 436, + 556, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 436, + 556, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 436, + 556, + 568 + ], + "type": "text", + "content": "Differences between HyperSeg and previous methods. Previous universal segmentation methods [24, 25, 31] lacking of VLLMs show inability in reasoning perception tasks while our HyperSeg demonstrates brilliant reasoning segmentation capability in complex scenarios. Besides, we make a significant generalization of the current VLLM-based segmentation methods [22, 51, 58, 59] for more diverse segmentation tasks in both image and video domains using a single model framework. Moreover, HyperSeg differs from previous methods in the three designs elaborated in the following sections." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 574, + 463, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 574, + 463, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 574, + 463, + 588 + ], + "type": "text", + "content": "3.2. Hybrid Entity Recognition" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 594, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 689 + ], + "type": "text", + "content": "As shown in Fig. 3 (a), predicting presented objects in the way of sequence generation (semantic prediction) tends to miss objects or produce repetitive predictions [45]. On the other hand, Fig. 3 (b), only using VLLM to embed class names (prompt tokens) as mask classifier at the decode stage disregards VLLM's powerful semantic recognition capability. Consequently, we propose a hybrid approach that leverages LLM in both generation and decoding processes." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "Instead of integrating mask tokens in input sequences and extracting the corresponding embedding from the one-pass" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8934" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 70, + 200, + 177 + ], + "blocks": [ + { + "bbox": [ + 65, + 70, + 200, + 177 + ], + "lines": [ + { + "bbox": [ + 65, + 70, + 200, + 177 + ], + "spans": [ + { + "bbox": [ + 65, + 70, + 200, + 177 + ], + "type": "image", + "image_path": "616d3b0e74640482eaf15a74f78b6ab8a220081ac36f42acbfb171dd6c63b94c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 120, + 178, + 130, + 186 + ], + "lines": [ + { + "bbox": [ + 120, + 178, + 130, + 186 + ], + "spans": [ + { + "bbox": [ + 120, + 178, + 130, + 186 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 209, + 69, + 386, + 177 + ], + "blocks": [ + { + "bbox": [ + 209, + 69, + 386, + 177 + ], + "lines": [ + { + "bbox": [ + 209, + 69, + 386, + 177 + ], + "spans": [ + { + "bbox": [ + 209, + 69, + 386, + 177 + ], + "type": "image", + "image_path": "9237b9284212ed2d8c78097eb968b9f8e7e7166d4be46a6b96974a00c064f644.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 287, + 178, + 297, + 186 + ], + "lines": [ + { + "bbox": [ + 287, + 178, + 297, + 186 + ], + "spans": [ + { + "bbox": [ + 287, + 178, + 297, + 186 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 389, + 70, + 543, + 177 + ], + "blocks": [ + { + "bbox": [ + 389, + 70, + 543, + 177 + ], + "lines": [ + { + "bbox": [ + 389, + 70, + 543, + 177 + ], + "spans": [ + { + "bbox": [ + 389, + 70, + 543, + 177 + ], + "type": "image", + "image_path": "2bdcbae716cfdc5eeb53754d3febe9800e8d3fd80091a509c0208f98438ab8d6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 476, + 178, + 487, + 186 + ], + "lines": [ + { + "bbox": [ + 476, + 178, + 487, + 186 + ], + "spans": [ + { + "bbox": [ + 476, + 178, + 487, + 186 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 62, + 241, + 169, + 364 + ], + "blocks": [ + { + "bbox": [ + 55, + 188, + 555, + 234 + ], + "lines": [ + { + "bbox": [ + 55, + 188, + 555, + 234 + ], + "spans": [ + { + "bbox": [ + 55, + 188, + 555, + 234 + ], + "type": "text", + "content": "Figure 3. The comparison of different recognition strategies. (a) Generation-Only [22, 41]: both the semantic recognition (existing objects) and their mask tokens are generated by LLM. (b) Decode-Only [58, 59]: prompt embedding and mask tokens are decoded from LLM. The present objects are then determined by their similarity scores. (c) Hybrid (ours): prompt embedding is decoded from LLM while the semantically enhanced mask tokens are generated by LLM. Their similarity scores reflect the objects' presence." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 62, + 241, + 169, + 364 + ], + "lines": [ + { + "bbox": [ + 62, + 241, + 169, + 364 + ], + "spans": [ + { + "bbox": [ + 62, + 241, + 169, + 364 + ], + "type": "image", + "image_path": "cca5f1eb172bf49413a0bbe6f89b7437e763d28cffb3eaadeee2aa3505fdc45f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 93, + 365, + 104, + 374 + ], + "lines": [ + { + "bbox": [ + 93, + 365, + 104, + 374 + ], + "spans": [ + { + "bbox": [ + 93, + 365, + 104, + 374 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 171, + 241, + 287, + 365 + ], + "blocks": [ + { + "bbox": [ + 171, + 241, + 287, + 365 + ], + "lines": [ + { + "bbox": [ + 171, + 241, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 171, + 241, + 287, + 365 + ], + "type": "image", + "image_path": "5455b62a85ea2b1d9e7aac7be0d7b0e4465f1a672d63a1cd8fe80ba8b0508f73.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 236, + 365, + 248, + 374 + ], + "lines": [ + { + "bbox": [ + 236, + 365, + 248, + 374 + ], + "spans": [ + { + "bbox": [ + 236, + 365, + 248, + 374 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 378, + 296, + 433 + ], + "lines": [ + { + "bbox": [ + 55, + 378, + 296, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 378, + 296, + 433 + ], + "type": "text", + "content": "Figure 4. Comparison between previous vision perceiver and our FVP. (a): previous vision perceiver [2, 23] uses the coarse single-scale CLIP visual features which are inadequate for fine-grained perception tasks. (b): FVP encodes the multi-scale visual features into fine-grained tokens." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 446, + 296, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 446, + 296, + 543 + ], + "spans": [ + { + "bbox": [ + 55, + 446, + 296, + 543 + ], + "type": "text", + "content": "forward output of VLLM, we instruct VLLM to generate the mask tokens preceded by the estimated objects' names. As illustrated in Fig. 3 (c), VLLM is compelled to generate all the existing objects in the vision input and then the mask tokens. The semantically enhanced mask tokens contain valuable semantic integrated information about the image, which are subsequently used as input for the segmentation predictor to generate segmentation masks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 551, + 218, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 551, + 218, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 551, + 218, + 563 + ], + "type": "text", + "content": "3.3. Fine-grained Visual Perceiver" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 570, + 296, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 663 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 663 + ], + "type": "text", + "content": "Why twin-tower vision encoder? As shown in Fig. 4, previous VLLMs and VLLM-based segmentation methods usually utilize the pre-trained CLIP encoder to obtain single-scale and low-resolution vision features interacted with diverse languages, which is insufficient for fine-grained image and video segmentation tasks. Therefore, we adopt an extra pyramid vision encoder [7] to inject details-aware visual information." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Specifically, we fuse multi-scale visual features into fine-grained tokens (stated as " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " in Sec 3.1) which can inject rich fine-grained visual information into the pre-trained VLLMs without excessive computation cost. Formally, given the" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 245, + 555, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 245, + 555, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 245, + 555, + 304 + ], + "type": "text", + "content": "vision input " + }, + { + "bbox": [ + 313, + 245, + 555, + 304 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 313, + 245, + 555, + 304 + ], + "type": "text", + "content": ", we leverage a pyramid vision encoder [7] " + }, + { + "bbox": [ + 313, + 245, + 555, + 304 + ], + "type": "inline_equation", + "content": "F_{seg}" + }, + { + "bbox": [ + 313, + 245, + 555, + 304 + ], + "type": "text", + "content": " to get details-aware image features " + }, + { + "bbox": [ + 313, + 245, + 555, + 304 + ], + "type": "inline_equation", + "content": "f_{img}" + }, + { + "bbox": [ + 313, + 245, + 555, + 304 + ], + "type": "text", + "content": ". For the j-th scale and the previous fine-grained tokens " + }, + { + "bbox": [ + 313, + 245, + 555, + 304 + ], + "type": "inline_equation", + "content": "P_{j-1}" + }, + { + "bbox": [ + 313, + 245, + 555, + 304 + ], + "type": "text", + "content": ", the FVP module enriches each token through conditional weighted cross-attention:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 367, + 315, + 555, + 332 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 315, + 555, + 332 + ], + "spans": [ + { + "bbox": [ + 367, + 315, + 555, + 332 + ], + "type": "interline_equation", + "content": "\\hat {P} _ {j} = \\operatorname {M H C A} \\left(P _ {j - 1}, G _ {p} \\left(f _ {\\text {i m g}} ^ {(j)}\\right)\\right), \\tag {5}", + "image_path": "f8bee75b2b63943a9ee4cc50d3bf886e7b3c7b30208227d8854de2fe54f400f4.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 362, + 334, + 555, + 349 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 334, + 555, + 349 + ], + "spans": [ + { + "bbox": [ + 362, + 334, + 555, + 349 + ], + "type": "interline_equation", + "content": "P _ {j} = P _ {j - 1} + \\tanh (\\mathrm {M L P} (\\hat {P} _ {j})) \\cdot \\hat {P} _ {j}, \\tag {6}", + "image_path": "00bc0d8ec27d55430664c875c363683bbeec02c403e3ed3406c06d884d56d11e.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 354, + 556, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 556, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 556, + 462 + ], + "type": "text", + "content": "where MHCA denotes the Multi-Head Cross-Attention layer, " + }, + { + "bbox": [ + 313, + 354, + 556, + 462 + ], + "type": "inline_equation", + "content": "G_{p}" + }, + { + "bbox": [ + 313, + 354, + 556, + 462 + ], + "type": "text", + "content": " is the projection function, tanh is a normalization function and MLP is a Multilayer Perceptron. The component of " + }, + { + "bbox": [ + 313, + 354, + 556, + 462 + ], + "type": "inline_equation", + "content": "\\tanh(\\mathrm{MLP}(\\hat{P}_j))" + }, + { + "bbox": [ + 313, + 354, + 556, + 462 + ], + "type": "text", + "content": " is the conditional weight used to multiply the enriched fine-grained tokens " + }, + { + "bbox": [ + 313, + 354, + 556, + 462 + ], + "type": "inline_equation", + "content": "\\hat{P}_j" + }, + { + "bbox": [ + 313, + 354, + 556, + 462 + ], + "type": "text", + "content": " before the residual connection to the previous tokens " + }, + { + "bbox": [ + 313, + 354, + 556, + 462 + ], + "type": "inline_equation", + "content": "P_{j - 1}" + }, + { + "bbox": [ + 313, + 354, + 556, + 462 + ], + "type": "text", + "content": ". Additionally, we initialize the weight value to zero to ensure the adaptation to diverse multi-scale image features while retaining the training stability." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 470, + 424, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 470, + 424, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 470, + 424, + 483 + ], + "type": "text", + "content": "3.4. Temporal Adapter" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 489, + 556, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 489, + 556, + 595 + ], + "spans": [ + { + "bbox": [ + 313, + 489, + 556, + 595 + ], + "type": "text", + "content": "Video segmentation entails distinct challenges, requiring reasoning across multiple frames and the maintenance of temporal coherence. Existing VLLM-based methods exhibit limitations in addressing video perception tasks and lack specialized designs for comprehending temporal dynamics in video analysis. To this end, we utilize global prompt aggregation and local space-time information injection in the time dimension to adapt to more complicated video perception tasks." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 597, + 556, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 597, + 556, + 657 + ], + "spans": [ + { + "bbox": [ + 313, + 597, + 556, + 657 + ], + "type": "text", + "content": "Global prompt aggregation. For the current prompt embedding " + }, + { + "bbox": [ + 313, + 597, + 556, + 657 + ], + "type": "inline_equation", + "content": "E_{\\mathcal{P}}" + }, + { + "bbox": [ + 313, + 597, + 556, + 657 + ], + "type": "text", + "content": " in the video object mask retrieval process, we leverage the adaptive average pooling strategy along the time dimension to aggregate global object and temporal information of previous " + }, + { + "bbox": [ + 313, + 597, + 556, + 657 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 597, + 556, + 657 + ], + "type": "text", + "content": " frames." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 359, + 665, + 555, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 665, + 555, + 681 + ], + "spans": [ + { + "bbox": [ + 359, + 665, + 555, + 681 + ], + "type": "interline_equation", + "content": "E _ {\\mathcal {P}} = \\operatorname {A v g P o o l} ([ E _ {\\mathcal {P}} ^ {0}, E _ {\\mathcal {P}} ^ {1}, \\dots , E _ {\\mathcal {P}} ^ {T} ]), \\tag {7}", + "image_path": "f439156efbdcd7377461702d010fe2a99b3979cc23d42dfa68a020c3dff98c73.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "text", + "content": "Local space-time information injection. We propose a sequential renewal strategy for space-time information prop" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8935" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 60, + 113, + 552, + 290 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "type": "text", + "content": "Table 1. Comparison with the state-of-the-art models on the RefCOCO series and more challenging generalized referring expression segmentation benchmark gRefCOCO. " + }, + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "type": "text", + "content": " denotes models using pre-trained SAM [21] for mask generation. * means using gRefCOCO for training while other methods are evaluated in zero-shot manners." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 113, + 552, + 290 + ], + "lines": [ + { + "bbox": [ + 60, + 113, + 552, + 290 + ], + "spans": [ + { + "bbox": [ + 60, + 113, + 552, + 290 + ], + "type": "table", + "html": "
TypeMethodRefCOCORefCOCO+RefCOCOggRefCOCO
valtestAtestBvaltestAtestBval(U)test(U)valtestAtestB
SegmentationVLT [11]67.570.565.256.361.050.155.057.752.5*62.2*50.5*
CRIS [44]70.573.266.162.368.153.759.960.455.3*63.8*51.0*
LAVT [53]72.775.868.862.168.455.161.262.157.6*65.3*55.0*
PolyFormer-B [30]74.876.671.167.672.959.367.869.1---
VLLM-basedLISA-7B [22]‡74.979.172.365.170.858.167.970.638.7*52.6*44.8*
PixelLM-7B [41]73.076.568.266.371.758.369.370.5---
GSVA-7B [49]‡76.477.472.864.567.758.671.172.061.7*69.2*60.3*
GroundHog-7B [33]78.579.975.770.575.064.974.174.666.7*--
SAM4MLLM-7B [6]‡79.682.876.173.577.865.874.575.666.3*70.1*63.2*
Segmentation NetworkLaSagnA-7B [45]‡76.878.773.866.470.660.170.671.938.150.442.1
OMG-LLaVA [58]78.080.374.169.173.163.072.972.9---
GLaMM [40]‡79.583.276.972.678.764.674.274.9---
PSALM [59]83.684.781.672.975.570.173.874.442.052.450.6
HyperSeg84.885.783.479.083.575.279.478.947.557.352.5
", + "image_path": "6340c8cac50ddbc094880be494fde280b0501a971ae7d9c2c0ae71930afdf96d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 58, + 331, + 553, + 463 + ], + "blocks": [ + { + "bbox": [ + 55, + 300, + 555, + 323 + ], + "lines": [ + { + "bbox": [ + 55, + 300, + 555, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 555, + 323 + ], + "type": "text", + "content": "Table 2. Comparison with the state-of-the-art models on more complex and challenging reasoning segmentation benchmarks: ReVOS in video domain and ReasonSeg in image domain. " + }, + { + "bbox": [ + 55, + 300, + 555, + 323 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 55, + 300, + 555, + 323 + ], + "type": "text", + "content": " denotes the same meaning as Tab. 1." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 331, + 553, + 463 + ], + "lines": [ + { + "bbox": [ + 58, + 331, + 553, + 463 + ], + "spans": [ + { + "bbox": [ + 58, + 331, + 553, + 463 + ], + "type": "table", + "html": "
MethodBackboneReVOS-ReasoningReVOS-ReferringReVOS-OverallReasonSeg
JFJ&FJFJ&FJFJ&FgIoUcIoU
LMPM [12]Swin-T13.324.318.829.039.134.121.231.726.4--
ReferFormer [47]Video-Swin-B21.325.623.431.234.332.726.229.928.1--
LISA-7B [22]‡ViT-H33.838.436.144.347.145.739.142.740.952.954.0
LaSagnA-7B [45]‡ViT-H---------48.847.2
SAM4MLLM-7B [6]‡EfficientViT-SAM-XL1---------46.748.1
TrackGPT-13B [62]‡ViT-H38.142.940.548.350.649.543.246.845.0--
VISA-7B [51]‡ViT-H36.741.739.251.154.752.943.948.246.152.757.8
VISA-13B [51]‡ViT-H38.343.540.952.355.854.145.349.747.5--
HyperSeg-3BSwin-B50.255.853.056.060.958.553.158.455.759.256.7
", + "image_path": "7679cb9200bbf36b1c70a776f3de6d2bcaeee4d4ceb23a6eb592e3abfd33f29e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 483, + 297, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 297, + 507 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 297, + 507 + ], + "type": "text", + "content": "agation based on fine-grained tokens " + }, + { + "bbox": [ + 55, + 483, + 297, + 507 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 483, + 297, + 507 + ], + "type": "text", + "content": " to inject object information of adjacent frames. Formally," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 125, + 516, + 296, + 531 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 516, + 296, + 531 + ], + "spans": [ + { + "bbox": [ + 125, + 516, + 296, + 531 + ], + "type": "interline_equation", + "content": "P _ {t} = G _ {l} \\left[ F _ {L L M} \\left(P _ {t - 1}\\right) \\right], \\tag {8}", + "image_path": "e9f253567efbc6aaabee43dd53735e01190357b5a8e81387ee76b120bf5e2a99.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 538, + 296, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 538, + 296, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 538, + 296, + 586 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 538, + 296, + 586 + ], + "type": "inline_equation", + "content": "P_{t}" + }, + { + "bbox": [ + 55, + 538, + 296, + 586 + ], + "type": "text", + "content": " denotes the time-aware fine-grained tokens of the current " + }, + { + "bbox": [ + 55, + 538, + 296, + 586 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 538, + 296, + 586 + ], + "type": "text", + "content": "-th frame, " + }, + { + "bbox": [ + 55, + 538, + 296, + 586 + ], + "type": "inline_equation", + "content": "G_{l}" + }, + { + "bbox": [ + 55, + 538, + 296, + 586 + ], + "type": "text", + "content": " is the projection function to transfer the previous features to the current space and align the feature dimensions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 586, + 296, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 296, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 296, + 659 + ], + "type": "text", + "content": "The proposed global prompt aggregation and local spacetime information injection within our temporal adapter facilitate the coalescence of both long-term and short-term vision-language information, which is essential for comprehensive video perception. More details about the temporal adapter are provided in the Supplementary Material." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 669, + 137, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 669, + 137, + 683 + ], + "spans": [ + { + "bbox": [ + 55, + 669, + 137, + 683 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Datasets. We use the one-stage training strategy to train HyperSeg with the multi-dataset and multi-task manners." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 483, + 555, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 555, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 555, + 579 + ], + "type": "text", + "content": "For image segmentation, we use COCO Panoptic [26], RefCOCO series [35, 55], COCO-Interactive, and ReasonSeg [22]. For video segmentation, we utilize DAVIS-2017 datasets [4], Ref-Youtu-VOS [42], YouTube-VIS 2019 [52], and ReVOS [51]. Besides, we use LLAVA-150k [29] to maintain the vision-language conversation capability of VLLM (we show the results on Multi-modal benchmarks in the Supplementary Material)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "Implementation details We load the pre-trained weights of Mipha [63] for our VLLM, and Maks2Former [7] for our segmentation predictor. We use three layers of FVP for fine-grained information fusion and utilize LoRA [18] to finetune the LLM efficiently. We train HyperSeg jointly for 160k iterations using a batch size of 32 on 8 NVIDIA A100 GPUs, which means each task takes approximately 16k iterations. We employ the AdamW optimizer with a learning rate of " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "4 \\times 10^{-5}" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": " and with a cosine schedule. All the hyper-parameters in the loss " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": " are assigned values 1.0." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8936" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 75, + 91, + 533, + 224 + ], + "blocks": [ + { + "bbox": [ + 73, + 71, + 536, + 83 + ], + "lines": [ + { + "bbox": [ + 73, + 71, + 536, + 83 + ], + "spans": [ + { + "bbox": [ + 73, + 71, + 536, + 83 + ], + "type": "text", + "content": "Table 3. Quantitative results on the closed-set COCO-Panoptic segmentation, open-vocabulary segmentation (-OV) benchmarks." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 75, + 91, + 533, + 224 + ], + "lines": [ + { + "bbox": [ + 75, + 91, + 533, + 224 + ], + "spans": [ + { + "bbox": [ + 75, + 91, + 533, + 224 + ], + "type": "table", + "html": "
TypeMethodBackboneCOCO-PanopticADE-OVCitys-OVPC59-OVPAS20-OV
PQmIoUPQmIoUPQmIoUmIoU
Segmentation SpecialistMask2former [7]Swin-B55.165.1-----
OneFormer [19]Swin-L57.967.4-----
SEEM [64]DaViT-B56.166.3-----
MaskCLIP [13]ViT-L30.947.615.123.7-45.9-
DeOP [16]ResNet-101c---22.9-48.891.7
SimBaseline [50]ViT-B---20.5-47.788.4
DaTaSeg [15]ViTDet-B52.862.712.318.328.051.1-
VLLM-based Segmentation NetworkOMG-LLaVA [58]ConvNeXt-L53.8------
PSALM [22]Swin-B55.966.613.718.228.848.581.3
HyperSegSwin-B61.277.216.122.331.164.692.1
", + "image_path": "2ab7ef9cb76d3447467b9955fa1e6e4f51571cd37d9174a9e9ca5dc5e52099a9.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 56, + 283, + 297, + 381 + ], + "blocks": [ + { + "bbox": [ + 55, + 241, + 297, + 275 + ], + "lines": [ + { + "bbox": [ + 55, + 241, + 297, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 241, + 297, + 275 + ], + "type": "text", + "content": "Table 4. Results of common video segmentation benchmarks, including DAVIS17, Ref-YouTube-VOS, Ref-DAVIS17, and YouTube-VIS 2019. " + }, + { + "bbox": [ + 55, + 241, + 297, + 275 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 55, + 241, + 297, + 275 + ], + "type": "text", + "content": " denotes the same meaning as Tab. 1." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 283, + 297, + 381 + ], + "lines": [ + { + "bbox": [ + 56, + 283, + 297, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 283, + 297, + 381 + ], + "type": "table", + "html": "
MethodBackboneDAVIS17Ref-YTRef-DAVISYT-VIS
J&FJ&FJ&FmAP
SEEM [64]DaViT-B62.8---
OMG-Seg [24]ConvNeXt-L74.3--56.4
ReferFormer [47]Video-Swin-B-62.961.1-
OnlineRefer [46]Swin-L-63.564.8-
UNINEXT [25]ConvNeXt-L77.266.266.764.3
LISA-7B [22]‡ViT-H-53.964.8-
VISA-13B [51]‡ViT-H-63.070.4-
VideoLISA-3.8B [3]‡ViT-H-63.768.8-
HyperSeg-3BSwin-B77.668.571.253.8
", + "image_path": "592873928381c821b545721056b58bf41bef6420aa6a6c76091ecb4863370519.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 393, + 247, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 393, + 247, + 406 + ], + "spans": [ + { + "bbox": [ + 55, + 393, + 247, + 406 + ], + "type": "text", + "content": "4.1. Comparisons with State-of-the-Arts" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 413, + 297, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 413, + 297, + 556 + ], + "spans": [ + { + "bbox": [ + 54, + 413, + 297, + 556 + ], + "type": "text", + "content": "Referring expression segmentation results. We compare HyperSeg with the state-of-the-art methods on the benchmarks RefCOCO+/g [35, 55] and more challenging generalized referring expression segmentation benchmark gRefCOCO [27]. in Tab. 1. Based on the versatile and adaptable design of HyperSeg, our model achieves state-of-the-art performance on all the referring datasets. Specifically, HyperSeg surpasses the current SOTA by a large margin, reaching 79.7 cIoU on RefCOCO+ val (+6.8 over PSALM). Besides, Our model shows superiority in challenging G-RES tasks compared with previous zero-shot methods, demonstrating the robustness and generalization ability of HyperSeg." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 557, + 297, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 557, + 297, + 653 + ], + "spans": [ + { + "bbox": [ + 54, + 557, + 297, + 653 + ], + "type": "text", + "content": "Reasoning segmentation results. We compare HyperSeg with the state-of-the-art methods on image reasoning segmentation (ReasonSeg [22]) and reasoning video object segmentation (ReVOS [51]) in Tab. 2. Our HyperSeg achieves superior performance on reasoning tasks, significantly surpassing previous state-of-the-art methods (+12.1 on ReVOS-Reasoning), which shows HyperSeg powerful reasoning capability of tackling complex scenarios." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 654, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 654, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 654, + 297, + 714 + ], + "type": "text", + "content": "Generic image segmentation results. We show the performance of HyperSeg on COCO-Panoptic [26] and open-vocabulary segmentation [9, 14, 34, 60] tasks in Tab. 3. HyperSeg achieves excellent performance compared with both specialist models and VLLMs-based methods on both closed-" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 243, + 555, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 555, + 326 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 555, + 326 + ], + "type": "text", + "content": "set and open-vocabulary segmentation tasks. Specifically, HyperSeg surpasses the VLLM-based PSALM by a significant margin (+5.3 on COCO PQ, and +10.6 on mIoU), which demonstrates our powerful capabilities of handling complex semantic perception and segmentation tasks. Besides, we show the results of COCO-Interactive in the Supplementary Material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 327, + 556, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 556, + 435 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 556, + 435 + ], + "type": "text", + "content": "Common video segmentation results. We compare HyperSeg with previous video segmentation methods in Tab. 4, including visual-prompted semi-supervised VOS (DAVIS17 val), text-prompted referring video object segmentation (RefYouTube-VOS, Ref-DAVIS17) and video instance segmentation (YouTube-VIS 2019). HyperSeg shows promising results over previous unified segmentation methods [24, 25]. Besides, HyperSeg performs more video perception tasks than previous VLLM-based models [3, 51]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 443, + 383, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 443, + 383, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 443, + 383, + 455 + ], + "type": "text", + "content": "4.2. Ablations" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 461, + 556, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 556, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 556, + 581 + ], + "type": "text", + "content": "The mutual influence between different tasks. Our model can be trained and inferred across multiple tasks and datasets simultaneously. We evaluate the mutual impact of different tasks in Tab. 5. The results show that joint training can enhance the model performance compared with the task-specific model. Besides, the performance of video segmentation tasks can be improved significantly by adding the image training datasets. This demonstrates the generalization and self-consistency of our HyperSeg to perform universal segmentation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 582, + 556, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 665 + ], + "type": "text", + "content": "Effect of different LLMs and vision backbone. In Tab. 6, we evaluate the effect of different sizes of LLMs and vision backbone. Our HyperSeg achieves excellent performance using smaller LLMs and vision encoder compared with the previous SOTA models like VISA[51] and PSALM[59]. Besides, the performance of HyperSeg can be further improved by using the more powerful LLM (Phi-2-2.7B [20])." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 666, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 556, + 714 + ], + "type": "text", + "content": "Ablation on the proposed components. We assess the effectiveness of our proposed FVP module and Hybrid Entity Recognition strategy. As shown in Tab. 7, with our fine-grained visual integration and hybrid entity semantic" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8937" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 86, + 102, + 525, + 183 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 5. The mutual influence between different tasks. Task-specific means training task-specific models only on data from corresponding tasks, Refer+Reason denotes the model is trained on referring and reasoning segmentation data." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 86, + 102, + 525, + 183 + ], + "lines": [ + { + "bbox": [ + 86, + 102, + 525, + 183 + ], + "spans": [ + { + "bbox": [ + 86, + 102, + 525, + 183 + ], + "type": "table", + "html": "
Task-specificRefer+ReasonVideoImageRefCOCOCOCOReVOSYT-VIS
valtestAtestBPQmIoUReasoningReferringOverallmAP
✓83.885.982.260.875.151.256.653.950.7
✓83.384.980.9--53.157.355.2-
✓85.686.182.460.976.5----
✓-----51.157.054.150.4
✓✓✓84.885.783.461.277.253.058.555.753.8
", + "image_path": "267b39c75ec973fde10018d0e1820eeacc28392968fdea987684bf1230587559.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 85, + 213, + 526, + 306 + ], + "blocks": [ + { + "bbox": [ + 88, + 193, + 520, + 205 + ], + "lines": [ + { + "bbox": [ + 88, + 193, + 520, + 205 + ], + "spans": [ + { + "bbox": [ + 88, + 193, + 520, + 205 + ], + "type": "text", + "content": "Table 6. The comparison of different LLMs and backbone usages. w/o CLIP means without using CLIP vision encoder." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 85, + 213, + 526, + 306 + ], + "lines": [ + { + "bbox": [ + 85, + 213, + 526, + 306 + ], + "spans": [ + { + "bbox": [ + 85, + 213, + 526, + 306 + ], + "type": "table", + "html": "
MethodLLMCOCOReVOSADE-OVPC59-OVPAS20-OV
PQmIoUReasoningReferringOverallmIoUmIoUmIoU
LISA [22]Vicuna-7B--36.145.740.9---
VISA [51]Vicuna-13B--40.954.147.5---
PSALM(w/o CLIP) [22]Phi-1.5-1.3B55.966.6---18.248.581.3
HyperSeg (w/o CLIP)Phi-1.5-1.3B61.176.044.049.746.918.960.090.6
HyperSegPhi-1.5-1.3B60.976.750.857.053.920.361.590.8
HyperSegPhi-2-2.7B61.277.253.058.555.722.364.692.1
", + "image_path": "4fd2b4991d8c7b0b097d12f298291f1acbd7f3ee878dbfd8e60f5b711b1bdb27.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 74, + 360, + 279, + 439 + ], + "blocks": [ + { + "bbox": [ + 55, + 318, + 297, + 352 + ], + "lines": [ + { + "bbox": [ + 55, + 318, + 297, + 352 + ], + "spans": [ + { + "bbox": [ + 55, + 318, + 297, + 352 + ], + "type": "text", + "content": "Table 7. Ablation on the core components of HyperSeg. FVP and HER denote the proposed Fine-grained Visual Perceiver and Hybrid Entity Recognition modules." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 74, + 360, + 279, + 439 + ], + "lines": [ + { + "bbox": [ + 74, + 360, + 279, + 439 + ], + "spans": [ + { + "bbox": [ + 74, + 360, + 279, + 439 + ], + "type": "table", + "html": "
FVPHERYT-VISCOCORefCOCO
mAPPQmIoUcIoU
✓48.454.866.282.8
50.855.866.684.6
✓✓52.059.774.684.3
✓✓53.861.277.284.8
", + "image_path": "baa1662006fdb3c6f6b26ad0695145f4592c20af2922285393a9acf526358f22.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 65, + 492, + 284, + 556 + ], + "blocks": [ + { + "bbox": [ + 55, + 450, + 295, + 484 + ], + "lines": [ + { + "bbox": [ + 55, + 450, + 295, + 484 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 295, + 484 + ], + "type": "text", + "content": "Table 8. Ablation on the Fine-grained Visual Perceiver design. CW denotes the Conditional Weight illustrated in Sec. 3.3, and Scale denotes the total scale in the proposed FVP module." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 65, + 492, + 284, + 556 + ], + "lines": [ + { + "bbox": [ + 65, + 492, + 284, + 556 + ], + "spans": [ + { + "bbox": [ + 65, + 492, + 284, + 556 + ], + "type": "table", + "html": "
CWScaleYT-VISCOCORefCOCO
mAPPQmIoUcIoU
✓single-layer49.755.868.083.7
multi-layers50.458.973.484.5
multi-layers53.861.277.284.8
", + "image_path": "de1d8bafefe6bb1bdb91bc7f87d0c1f2b331a37f3c4b03fc894066a7bbba6cb7.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 569, + 295, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 569, + 295, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 569, + 295, + 592 + ], + "type": "text", + "content": "enhancement, the segmentation accuracy can be enhanced significantly " + }, + { + "bbox": [ + 55, + 569, + 295, + 592 + ], + "type": "inline_equation", + "content": "(+5.4" + }, + { + "bbox": [ + 55, + 569, + 295, + 592 + ], + "type": "text", + "content": " on YT-VIS, " + }, + { + "bbox": [ + 55, + 569, + 295, + 592 + ], + "type": "inline_equation", + "content": "+6.4" + }, + { + "bbox": [ + 55, + 569, + 295, + 592 + ], + "type": "text", + "content": " on COCO panoptic PQ)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 594, + 295, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 688 + ], + "type": "text", + "content": "Design of the Fine-grained Visual Perceiver. In the FVP module, we combine multi-scale visual features into fixed perception queries using the condition-wise cross-attention layers to extract rich visual details from different scales of the pyramid encoder. As shown in Tab. 8, together with the conditional weight and the multi-scale design, our model makes a significant improvement on both image and video segmentation tasks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Effect of temporal adapter. We evaluate the effectiveness of the proposed temporal adapter including global prompt ag" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 331, + 338, + 538, + 416 + ], + "blocks": [ + { + "bbox": [ + 329, + 318, + 538, + 330 + ], + "lines": [ + { + "bbox": [ + 329, + 318, + 538, + 330 + ], + "spans": [ + { + "bbox": [ + 329, + 318, + 538, + 330 + ], + "type": "text", + "content": "Table 9. Ablation on the temporal adapter for video tasks." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 331, + 338, + 538, + 416 + ], + "lines": [ + { + "bbox": [ + 331, + 338, + 538, + 416 + ], + "spans": [ + { + "bbox": [ + 331, + 338, + 538, + 416 + ], + "type": "table", + "html": "
GlobalLocalRef-DAVIS17ReVOSYT-VIS
J&FJ&FmAP
✓67.354.147.9
68.854.548.5
✓69.354.850.2
✓✓71.255.753.8
", + "image_path": "f05bdb66d27944b97d205086827a7d51f7bb0781fa4001b7cf513df4d02f49f3.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 433, + 555, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 433, + 555, + 482 + ], + "spans": [ + { + "bbox": [ + 313, + 433, + 555, + 482 + ], + "type": "text", + "content": "gregation (global) and local space-time information injection (local) in Tab. 9. Incorporating both global and local components, the temporal adapter significantly enhances model performance across multiple video segmentation tasks." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 499, + 388, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 499, + 388, + 511 + ], + "spans": [ + { + "bbox": [ + 313, + 499, + 388, + 511 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "text", + "content": "In this study, we aim to present HyperSeg, the first VLLM-based universal segmentation model designed for pixel-level image and video perception, encompassing a wide range of generic segmentation and complex reasoning tasks. We propose the Hybrid Entity Recognition and Fine-grained Visual Perceiver to leverage the recognition capacity of VLLMs more effectively and enhances the VLLM's ability by capturing diverse levels of visual information without incurring excessive computational costs. With additional Temporal Adapter, HyperSeg can tackle challenging video tasks by incorporating global and local information. HyperSeg surpasses existing methods on complex reasoning segmentation and traditional perception tasks. The insights presented in this work expand the possibilities of VLLMs in visual perception and lay a foundation for future research on the integration of vision-language models." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8938" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 165, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 165, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 165, + 85 + ], + "type": "text", + "content": "6. Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 140 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 140 + ], + "type": "text", + "content": "This work was partly supported by the National Key Research and Development Program of China (No. 2024YFB2808903) and the Shenzhen Science and Technology Program (JCYJ20220818101001004)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 115, + 165 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 115, + 165 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 115, + 165 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 172, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 172, + 297, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 172, + 297, + 237 + ], + "spans": [ + { + "bbox": [ + 61, + 172, + 297, + 237 + ], + "type": "text", + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 239, + 296, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 239, + 296, + 293 + ], + "spans": [ + { + "bbox": [ + 61, + 239, + 296, + 293 + ], + "type": "text", + "content": "[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 2, 3, 5, 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 297, + 296, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 297, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 61, + 297, + 296, + 350 + ], + "type": "text", + "content": "[3] Zechen Bai, Tong He, Haiyang Mei, Pichao Wang, Ziteng Gao, Joya Chen, Lei Liu, Zheng Zhang, and Mike Zheng Shou. One token to seg them all: Language instructed reasoning segmentation in videos. arXiv preprint arXiv:2409.19603, 2024. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 353, + 296, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 353, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 62, + 353, + 296, + 396 + ], + "type": "text", + "content": "[4] Sergi Caelles, Alberto Montes, Kevis-Kokitsi Maninis, Yuhua Chen, Luc Van Gool, Federico Perazzi, and Jordi Pont-Tuset. The 2018 davis challenge on video object segmentation. arXiv preprint arXiv:1803.00557, 2018. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 399, + 296, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 399, + 296, + 441 + ], + "spans": [ + { + "bbox": [ + 62, + 399, + 296, + 441 + ], + "type": "text", + "content": "[5] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 3, 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 445, + 296, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 445, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 62, + 445, + 296, + 498 + ], + "type": "text", + "content": "[6] Yi-Chia Chen, Wei-Hua Li, Cheng Sun, Yu-Chiang Frank Wang, and Chu-Song Chen. Sam4mllm: Enhance multimodal large language model for referring expression segmentation. In European Conference on Computer Vision, pages 323-340. Springer, 2025. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 501, + 296, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 501, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 62, + 501, + 296, + 555 + ], + "type": "text", + "content": "[7] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 557, + 296, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 557, + 296, + 599 + ], + "spans": [ + { + "bbox": [ + 62, + 557, + 296, + 599 + ], + "type": "text", + "content": "[8] Jaemin Cho, Jie Lei, Hao Tan, and Mohit Bansal. Unifying vision-and-language tasks via text generation. In International Conference on Machine Learning, pages 1931-1942. PMLR, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 602, + 296, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 602, + 296, + 666 + ], + "spans": [ + { + "bbox": [ + 62, + 602, + 296, + 666 + ], + "type": "text", + "content": "[9] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "type": "text", + "content": "[10] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning, 2023. 3, 1" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 127 + ], + "type": "text", + "content": "[11] Henghui Ding, Chang Liu, Suchen Wang, and Xudong Jiang. Vision-language transformer and query generation for referring segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16321-16330, 2021. 6, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 129, + 555, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 555, + 183 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 555, + 183 + ], + "type": "text", + "content": "[12] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, and Chen Change Loy. Mevis: A large-scale benchmark for video segmentation with motion expressions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2694-2703, 2023. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 185, + 555, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 185, + 555, + 215 + ], + "spans": [ + { + "bbox": [ + 317, + 185, + 555, + 215 + ], + "type": "text", + "content": "[13] Zheng Ding, Jieke Wang, and Zhuowen Tu. Open-vocabulary universal image segmentation with maskclip. arXiv preprint arXiv:2208.08984, 2022. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 217, + 555, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 555, + 259 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 555, + 259 + ], + "type": "text", + "content": "[14] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-338, 2010. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 262, + 555, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 262, + 555, + 326 + ], + "spans": [ + { + "bbox": [ + 316, + 262, + 555, + 326 + ], + "type": "text", + "content": "[15] Xiuye Gu, Yin Cui, Jonathan Huang, Abdullah Rashwan, Xuan Yang, Xingyi Zhou, Golnaz Ghiasi, Weicheng Kuo, Huizhong Chen, Liang-Chieh Chen, et al. Dataseg: Taming a universal multi-dataset multi-task segmentation model. Advances in Neural Information Processing Systems, 36, 2024. 4, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 327, + 555, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 555, + 370 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 555, + 370 + ], + "type": "text", + "content": "[16] Cong Han, Yujie Zhong, Dengjie Li, Kai Han, and Lin Ma. Open-vocabulary semantic segmentation with decoupled one-pass network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1086-1096, 2023. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 372, + 555, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 555, + 403 + ], + "type": "text", + "content": "[17] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 405, + 555, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 405, + 555, + 447 + ], + "spans": [ + { + "bbox": [ + 317, + 405, + 555, + 447 + ], + "type": "text", + "content": "[18] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 449, + 555, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 449, + 555, + 503 + ], + "spans": [ + { + "bbox": [ + 317, + 449, + 555, + 503 + ], + "type": "text", + "content": "[19] Jitesh Jain, Jiachen Li, Mang Tik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. Oneformer: One transformer to rule universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2989-2998, 2023. 2, 7, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 505, + 555, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 505, + 555, + 558 + ], + "spans": [ + { + "bbox": [ + 317, + 505, + 555, + 558 + ], + "type": "text", + "content": "[20] Mojan Javaheripi, Sébastien Bubeck, Marah Abdin, Jyoti Aneja, Sebastien Bubeck, Caio César Teodoro Mendes, Weizhu Chen, Allie Del Giorno, Ronen Eldan, Sivakanth Gopi, et al. Phi-2: The surprising power of small language models. *Microsoft Research Blog*, 1:3, 2023. 7, 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 559, + 555, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 559, + 555, + 622 + ], + "spans": [ + { + "bbox": [ + 317, + 559, + 555, + 622 + ], + "type": "text", + "content": "[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 6, 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 625, + 555, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 555, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 555, + 667 + ], + "type": "text", + "content": "[22] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. ArXiv, abs/2308.00692, 2023. 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 670, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 555, + 712 + ], + "type": "text", + "content": "[23] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 2, 3, 5, 1" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8939" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 138 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 138 + ], + "type": "text", + "content": "[24] Xiangtai Li, Haobo Yuan, Wei Li, Henghui Ding, Size Wu, Wenwei Zhang, Yining Li, Kai Chen, and Chen Change Loy. Omg-seg: Is one model good enough for all segmentation? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27948-27959, 2024. 2, 4, 7, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 296, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 296, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 296, + 194 + ], + "type": "text", + "content": "[25] Fangjian Lin, Jianlong Yuan, Sitong Wu, Fan Wang, and Zhibin Wang. Uninext: Exploring a unified architecture for vision recognition. In Proceedings of the 31st ACM International Conference on Multimedia, pages 3200-3208, 2023. 2, 3, 4, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 195, + 296, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 195, + 296, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 195, + 296, + 239 + ], + "type": "text", + "content": "[26] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision, 2014. 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 239, + 296, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 296, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 296, + 282 + ], + "type": "text", + "content": "[27] Chang Liu, Henghui Ding, and Xudong Jiang. Gres: Generalized referring expression segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23592-23601, 2023. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 284, + 296, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 284, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 284, + 296, + 316 + ], + "type": "text", + "content": "[28] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 317, + 296, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 317, + 296, + 349 + ], + "spans": [ + { + "bbox": [ + 56, + 317, + 296, + 349 + ], + "type": "text", + "content": "[29] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2, 3, 6, 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 350, + 296, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 296, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 296, + 415 + ], + "type": "text", + "content": "[30] Jiang Liu, Hui Ding, Zhaowei Cai, Yuting Zhang, Ravi Kumar Satzoda, Vijay Mahadevan, and R Manmatha. Polyformer: Referring image segmentation as sequential polygon generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18653-18663, 2023. 6, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 416, + 296, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 416, + 296, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 416, + 296, + 470 + ], + "type": "text", + "content": "[31] Yong Liu, Cairong Zhang, Yitong Wang, Jiahao Wang, Yujiu Yang, and Yansong Tang. Universal segmentation at arbitrary granularity with language instruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3459-3469, 2024. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 472, + 296, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 472, + 296, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 472, + 296, + 525 + ], + "type": "text", + "content": "[32] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 526, + 296, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 526, + 296, + 580 + ], + "spans": [ + { + "bbox": [ + 56, + 526, + 296, + 580 + ], + "type": "text", + "content": "[33] Bo Miao, Mohammed Bennamoun, Yongsheng Gao, and Ajmal Mian. Spectrum-guided multi-granularity referring video object segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 920-930, 2023. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 582, + 296, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 296, + 647 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 296, + 647 + ], + "type": "text", + "content": "[34] Roozbeh Mottaghi, Xianjie Chen, Xiaobai Liu, Nam-Gyu Cho, Seong-Whan Lee, Sanja Fidler, Raquel Urtasun, and Alan Yuille. The role of context for object detection and semantic segmentation in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 891–898, 2014. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 647, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 296, + 712 + ], + "type": "text", + "content": "[35] Varun K Nagaraja, Vlad I Morariu, and Larry S Davis. Modeling context between objects for referring expression understanding. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 792-807. Springer, 2016. 6, 7" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 316, + 72, + 555, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 555, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 555, + 117 + ], + "type": "text", + "content": "[36] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 118, + 554, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 554, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 554, + 150 + ], + "type": "text", + "content": "[37] Renjie Pi, Lewei Yao, Jiahui Gao, Jipeng Zhang, and Tong Zhang. Perceptiongpt: Effectively fusing visual perception into llm. arXiv preprint arXiv:2311.06612, 2023. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 151, + 555, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 151, + 555, + 215 + ], + "spans": [ + { + "bbox": [ + 316, + 151, + 555, + 215 + ], + "type": "text", + "content": "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 217, + 555, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 555, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 555, + 271 + ], + "type": "text", + "content": "[39] Hanoona Rasheed, Muhammad Maaz, Sahal Shaji, Abdelrahman Shaker, Salman Khan, Hisham Cholakkal, Rao M Anwer, Erix Xing, Ming-Hsuan Yang, and Fahad S Khan. Glamm: Pixel grounding large multimodal model. arXiv preprint arXiv:2311.03356, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 272, + 555, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 555, + 338 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 555, + 338 + ], + "type": "text", + "content": "[40] Hanoona Rasheed, Muhammad Maaz, Sahal Shaji, Abdelrahman Shaker, Salman Khan, Hisham Cholakkal, Rao M Anwer, Eric Xing, Ming-Hsuan Yang, and Fahad S Khan. Glamm: Pixel grounding large multimodal model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13009-13018, 2024. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 338, + 555, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 555, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 555, + 381 + ], + "type": "text", + "content": "[41] Zhongwei Ren, Zhicheng Huang, Yunchao Wei, Yao Zhao, Dongmei Fu, Jiashi Feng, and Xiaojie Jin. Pixel reasoning with large multimodal model. ArXiv, abs/2312.02228, 2023. 2, 3, 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 383, + 555, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 555, + 447 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 555, + 447 + ], + "type": "text", + "content": "[42] Seonguk Seo, Joon-Young Lee, and Bohyung Han. Urvos: Unified referring video object segmentation network with a large-scale benchmark. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XV 16, pages 208–223. Springer, 2020. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 449, + 555, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 555, + 503 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 555, + 503 + ], + "type": "text", + "content": "[43] Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong Lu, Jie Zhou, Yu Qiao, et al. Vision: Large language model is also an open-ended decoder for vision-centric tasks. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 504, + 555, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 504, + 555, + 558 + ], + "spans": [ + { + "bbox": [ + 316, + 504, + 555, + 558 + ], + "type": "text", + "content": "[44] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clipdriven referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11686-11695, 2022. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 559, + 555, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 555, + 602 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 555, + 602 + ], + "type": "text", + "content": "[45] Cong Wei, Haoxian Tan, Yujie Zhong, Yujiu Yang, and Lin Ma. Lasagna: Language-based segmentation assistant for complex queries. arXiv preprint arXiv:2404.08506, 2024. 2, 4, 6, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 603, + 555, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 603, + 555, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 603, + 555, + 658 + ], + "type": "text", + "content": "[46] Dongming Wu, Tiancai Wang, Yang Zhang, Xiangyu Zhang, and Jianbing Shen. Onlinerefer: A simple online baseline for referring video object segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2761-2770, 2023. 7, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 658, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 555, + 712 + ], + "type": "text", + "content": "[47] Jiannan Wu, Yi Jiang, Peize Sun, Zehuan Yuan, and Ping Luo. Language as queries for referring video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4974-4984, 2022. 6, 7, 3" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "8940" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 73, + 297, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[48] Junfeng Wu, Qihao Liu, Yi Jiang, Song Bai, Alan Yuille, and Xiang Bai. In defense of online models for video instance segmentation. In ECCV, 2022. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 106, + 295, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 106, + 295, + 149 + ], + "spans": [ + { + "bbox": [ + 55, + 106, + 295, + 149 + ], + "type": "text", + "content": "[49] Zhuofan Xia, Dongchen Han, Yizeng Han, Xuran Pan, Shiji Song, and Gao Huang. Gsva: Generalized segmentation via multimodal large language models. arXiv preprint arXiv:2312.10103, 2023. 2, 6, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 150, + 297, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 150, + 297, + 205 + ], + "spans": [ + { + "bbox": [ + 57, + 150, + 297, + 205 + ], + "type": "text", + "content": "[50] Mengde Xu, Zheng Zhang, Fangyun Wei, Yutong Lin, Yue Cao, Han Hu, and Xiang Bai. A simple baseline for open-vocabulary semantic segmentation with pre-trained vision-language model. In European Conference on Computer Vision, pages 736-753. Springer, 2022. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 206, + 296, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 206, + 296, + 258 + ], + "spans": [ + { + "bbox": [ + 57, + 206, + 296, + 258 + ], + "type": "text", + "content": "[51] Cilin Yan, Haochen Wang, Shilin Yan, Xiaolong Jiang, Yao Hu, Guoliang Kang, Weidi Xie, and Efstratios Gavves. Visa: Reasoning video object segmentation via large language models. arXiv preprint arXiv:2407.11325, 2024. 2, 4, 6, 7, 8, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 261, + 295, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 261, + 295, + 294 + ], + "spans": [ + { + "bbox": [ + 57, + 261, + 295, + 294 + ], + "type": "text", + "content": "[52] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5188-5197, 2019. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 294, + 296, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 294, + 296, + 349 + ], + "spans": [ + { + "bbox": [ + 57, + 294, + 296, + 349 + ], + "type": "text", + "content": "[53] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. Levt: Language-aware vision transformer for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18155-18165, 2022. 6, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 350, + 295, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 350, + 295, + 392 + ], + "spans": [ + { + "bbox": [ + 57, + 350, + 295, + 392 + ], + "type": "text", + "content": "[54] Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and Yinfei Yang. Ferret: Refer and ground anything anywhere at any granularity. arXiv preprint arXiv:2310.07704, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 393, + 296, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 393, + 296, + 458 + ], + "spans": [ + { + "bbox": [ + 57, + 393, + 296, + 458 + ], + "type": "text", + "content": "[55] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 69-85. Springer, 2016. 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 460, + 295, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 460, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 57, + 460, + 295, + 502 + ], + "type": "text", + "content": "[56] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 503, + 296, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 503, + 296, + 557 + ], + "spans": [ + { + "bbox": [ + 57, + 503, + 296, + 557 + ], + "type": "text", + "content": "[57] Hao Zhang, Feng Li, Xueyan Zou, Shilong Liu, Chunyuan Li, Jianwei Yang, and Lei Zhang. A simple framework for open-vocabulary segmentation and detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1020-1031, 2023. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 559, + 295, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 559, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 57, + 559, + 295, + 612 + ], + "type": "text", + "content": "[58] Tao Zhang, Xiangtai Li, Hao Fei, Haobo Yuan, Shengqiong Wu, Shunping Ji, Chen Change Loy, and Shuicheng Yan. Omg-llava: Bridging image-level, object-level, pixel-level reasoning and understanding. arXiv preprint arXiv:2406.19389, 2024. 2, 4, 5, 6, 7, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 614, + 295, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 614, + 295, + 647 + ], + "spans": [ + { + "bbox": [ + 57, + 614, + 295, + 647 + ], + "type": "text", + "content": "[59] Zheng Zhang, Yeyao Ma, Enming Zhang, and Xiang Bai. Psalm: Pixelwise segmentation with large multi-modal model. arXiv preprint arXiv:2403.14598, 2024. 2, 3, 4, 5, 6, 7, 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 647, + 295, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 690 + ], + "type": "text", + "content": "[60] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision, 127(3):302-321, 2019. 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "text", + "content": "[61] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 240 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023.2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 555, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 555, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 555, + 139 + ], + "type": "text", + "content": "[62] Jiawen Zhu, Zhi-Qi Cheng, Jun-Yan He, Chenyang Li, Bin Luo, Huchuan Lu, Yifeng Geng, and Xuansong Xie. Tracking with human-intent reasoning. arXiv preprint arXiv:2312.17448, 2023. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 141, + 554, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 554, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 554, + 194 + ], + "type": "text", + "content": "[63] Minjie Zhu, Yichen Zhu, Xin Liu, Ning Liu, Zhiyuan Xu, Chaomin Shen, Yaxin Peng, Zhicai Ou, Feifei Feng, and Jian Tang. A comprehensive overhaul of multimodal assistant with small language models. arXiv preprint arXiv:2403.06199, 2024. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 197, + 554, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 554, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 554, + 240 + ], + "type": "text", + "content": "[64] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Wang, Lijuan Wang, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. Advances in Neural Information Processing Systems, 36, 2024. 7, 1, 2, 3" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "8941" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Hyperbolic Category Discovery/103e7fd4-0c6b-413e-8aae-04fa73b4770b_content_list.json b/2025/Hyperbolic Category Discovery/103e7fd4-0c6b-413e-8aae-04fa73b4770b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..db1fa1a592ee6c26a015f4c9ef07845a93b2d48d --- /dev/null +++ b/2025/Hyperbolic Category Discovery/103e7fd4-0c6b-413e-8aae-04fa73b4770b_content_list.json @@ -0,0 +1,1529 @@ +[ + { + "type": "text", + "text": "Hyperbolic Category Discovery", + "text_level": 1, + "bbox": [ + 338, + 130, + 658, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuanpei Liu* Zhenqi He* Kai Han†", + "bbox": [ + 318, + 179, + 676, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual AI Lab, The University of Hong Kong", + "bbox": [ + 318, + 199, + 676, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{ypliu0,zhenqi_he}@connect.hku.hk kaihanx@hku.hk", + "bbox": [ + 274, + 219, + 723, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generalized Category Discovery (GCD) is an intriguing open-world problem that has garnered increasing attention. Given a dataset that includes both labelled and unlabelled images, GCD aims to categorize all images in the unlabelled subset, regardless of whether they belong to known or unknown classes. In GCD, the common practice typically involves applying a spherical projection operator at the end of the self-supervised pretrained backbone, operating within Euclidean or spherical space. However, both of these spaces have been shown to be suboptimal for encoding samples that possesses hierarchical structures. In contrast, hyperbolic space exhibits exponential volume growth relative to radius, making it inherently strong at capturing the hierarchical structure of samples from both seen and unseen categories. Therefore, we propose to tackle the category discovery challenge in the hyperbolic space. We introduce HypCD, a simple Hyperbolic framework for learning hierarchy-aware representations and classifiers for generalized Category Discovery. HypCD first transforms the Euclidean embedding space of the backbone network into hyperbolic space, facilitating subsequent representation and classification learning by considering both hyperbolic distance and the angle between samples. This approach is particularly helpful for knowledge transfer from known to unknown categories in GCD. We thoroughly evaluate HypCD on public GCD benchmarks, by applying it to various baseline and state-of-the-art methods, consistently achieving significant improvements. Project page: https://visual-ai.github.io/hypcd/", + "bbox": [ + 86, + 300, + 485, + 738 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 765, + 220, + 781 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, category discovery – initially explored as novel category discovery (NCD) [24] and subsequently extended to generalized category discovery (GCD) [51] – has emerged as an intriguing open-world problem, gaining increasing attention. GCD addresses the challenges posed", + "bbox": [ + 89, + 790, + 483, + 867 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/416dc20d2276f8a148df659ee60187b9a177c9883a5d60fe24c761b343233734.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 268, + 901, + 416 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/95e747909c7a03310812b638ffe4b705297935e68713167428b247f101f6ffed.jpg", + "image_caption": [ + "Figure 1. (a) Spherical-based vs. Hyperbolic-based methods, where hyperbolic space better accommodates variations in scale and improves connections between samples. (b) Average ACC comparison of our method and previous SOTA across 'All', 'Old', and 'New' categories on the SSB [52] benchmark using DINO [7]." + ], + "image_footnote": [], + "bbox": [ + 519, + 420, + 901, + 566 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "by partially labelled datasets, where the unlabelled subset may contain instances from both seen and unseen classes. The goal is to leverage knowledge from labelled data to effectively categorize the unlabelled data. Based on the way to predict category index, existing GCD methods can be broadly classified into two types: non-parametric methods [27, 44, 45, 51] and parametric methods [37, 55, 59]. Non-parametric methods predict category index based on feature clustering while parametric methods utilize a parametric classifier.", + "bbox": [ + 509, + 655, + 906, + 805 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As shown in the GCD literature [51], object parts are effective for knowledge transfer from seen to unseen categories, which is crucial for novel category discovery. Methods have been developed to explicitly learn better local features by learning pixel-level prompts around local image regions [55] or utilizing part-level features [63]. However,", + "bbox": [ + 511, + 809, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 107, + 875, + 220, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9891", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8dd1299c9dfe5ab718c05656990a2dab7d455edcb9c7b884c1663db1d255da64.jpg", + "image_caption": [ + "(a) Inter-Category Relation" + ], + "image_footnote": [], + "bbox": [ + 96, + 89, + 328, + 190 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5279474e7184cab0d740b639eb6e62a1756d61fae0bcc1f1a11522bb04a2083c.jpg", + "image_caption": [ + "(b) Intra-Category Relation", + "Figure 2. Hierarchical relations in GCD. (a) Inter-category relationships within the Stanford-Cars dataset [33]. (b) Intra-category relationships within CUB [54] dataset." + ], + "image_footnote": [], + "bbox": [ + 336, + 90, + 470, + 191 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "these methods consider object parts as rigid image patches of the same size, without considering the hierarchical nature of the object parts and the scale discrepancy of the same parts in different images, thus unavoidably restricting the performance for GCD (see Fig. 1(a)), in which the objects often have distinct poses, scales and appearance. To address this problem, one possible solution is to learn image embeddings possessing hierarchical constraints or following tree-like structures. This has been proven to be effective in many tasks. For example, in image retrieval and clustering, the hierarchy constraint may arise from whole-fragment relation [29, 31]. Intuitively, in category discovery, which can be regarded as a transfer clustering task [24], we hypothesize that an embedding space that captures the hierarchical relations of object parts can also facilitate the discovery of new categories. Indeed, the hierarchical relations have been studied in GCD, such as [27, 41, 44, 45, 57]. However, they study from a substantially different perspective: inter-category hierarchy (see Fig. 2(a)). This only considers the hierarchy of different semantic classes from coarse to fine levels. Additionally, these methods require the relationships and number of levels in the hierarchy to be predefined, resulting in a lack of flexibility and scalability. Moreover, these methods are unable to capture more complex hierarchies, such as the compositional parts of an object (see Fig. 2(b)). This is particularly the case because existing methods [45, 51, 55, 59], no matter whether they consider the hierarchy or not, learn the image embeddings in a spherical space. This follows the common practice of applying a spherical projection operator at the end of the self-supervised feature backbone [7, 40]. Consequently, all subsequent operations, including distance calculations, are performed under either Euclidean or spherical geometry, resulting in limited awareness of hierarchical object parts.", + "bbox": [ + 91, + 263, + 483, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we study the overlooked perspective in category discovery: instead of learning in the Euclidean or spherical space, we advocate a space that captures the hierarchical structure of each data point. In spherical space, both the radius and volume are constant, whereas in Euclidean space, the volume grows polynomially with respect to the radius. Both of these spaces have been shown to be suboptimal for encoding samples that possess hierarchical", + "bbox": [ + 89, + 780, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "structures [10, 15, 31]. In contrast, hyperbolic space possesses a distinctive property where its volume grows exponentially relative to the radius. This characteristic makes hyperbolic space particularly suitable for embedding tree-like data, enhancing its representational power. Learning representations in hyperbolic space has proven to be effective in various computer vision tasks, including object recognition [15], object detection [32], semantic segmentation [58], and anomaly detection [35]. Inspired by these successes, we aim to realize the idea of learning hierarchy-aware representations to facilitate knowledge transfer in category discovery, thereby unleashing the potential of hyperbolic representations.", + "bbox": [ + 511, + 90, + 903, + 286 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To achieve this goal, we propose a simple yet effective framework, HypCD, to properly learn the hierarchy-aware representation and classifier for category discovery through the lens of hyperbolic geometry. In this framework, we adapt our framework to popular parametric [59] and non-parametric [51] GCD baselines as well as the state-of-the-art (SOTA) method SelEx [45], obtaining substantial improvements for them, establishing the new SOTA (see Fig. 1(b)). Firstly, starting from the self-supervised backbone pretrained in Euclidean space, we propose to map the Euclidean representation to a constrained Poincaré ball through feature clipping and exponential mapping. Secondly, we implement the hyperbolic representation learning and build a hyperbolic classifier on the Poincaré ball, considering both angle and distance between samples in hyperbolic space. Thirdly, to assign labels to unlabelled data after training, for non-parametric methods, we apply semi-supervised $k$ -means following the common practice; for parametric methods, we employ a hyperbolic classifier to make predictions. Despite its simplicity, our framework achieves significant performance improvements with two different pretrained weights (DINO [7] and DINOv2 [40]) on the public GCD datasets, including the coarse-grained classification datasets CIFAR-10 [34], CIFAR-100 [34], and ImageNet-100 [13], as well as the fine-grained SSB [50] benchmark.", + "bbox": [ + 511, + 287, + 903, + 679 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, we make the following contributions in this paper: (i) We identify the existing GCD methods' common shortcoming in encoding the hierarchical structure and propose to incorporate the hyperbolic geometry into the embedding space to address this limitation; (ii) We propose a simple yet effective framework, called HypCD, for incorporating the hyperbolic geometry in representation learning and classification for category discovery; and (iii) Through extensive experiments on public GCD benchmarks by applying HypCD to baseline and SOTA methods, our method consistently demonstrates effectiveness and superiority.", + "bbox": [ + 511, + 680, + 905, + 845 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 859, + 653, + 875 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Category Discovery. Novel category discovery (NCD) is", + "bbox": [ + 511, + 885, + 903, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9892", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "initially introduced in [24] to establish a realistic framework for transferring knowledge from seen categories to cluster unseen categories, by considering it as a transfer clustering problem. Many subsequent methods have been proposed to advance the field [17, 25, 26, 28, 63, 65]. Generalized category discovery [51] (GCD) relaxes the assumption in NCD by considering unlabelled data containing samples from both known and unknown classes [51]. Further investigations [6, 8, 27, 30, 37, 42, 56] have explored a variety of strategies to address the challenges posed by GCD. One notable approach, SimGCD [59], proposes to learn a parametric classifier enhanced by mean entropy regularization, thereby improving performance. In another vein, GPC [64] employs Gaussian mixture models to jointly learn robust representations while simultaneously estimating the number of unknown categories. SPT-Net [55] presents a spatial prompt tuning method that enables models to concentrate more effectively on specific object parts, thus enhancing knowledge transfer in GCD. Most recently, SelEx [45] has been proposed, leveraging hierarchical semi-supervised $k$ -means to achieve SOTA results on fine-grained datasets. Additionally, various efforts are focused on addressing category discovery from multiple perspectives. For instance, [28] emphasizes multi-modal category discovery; [62] and [8] explore a continual setting; [43] studies category discovery in a federated setting; and [56] examines GCD in the presence of domain shifts.", + "bbox": [ + 89, + 90, + 483, + 498 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Hyperbolic Geometry. Hyperbolic space, defined as a non-Euclidean manifold with exponential volume growth in relation to its radius, is inherently aligned with the embedding of tree-like and hierarchical data structures in visual recognition tasks. Significant advancements in this area include [31], which presents a hyperbolic image embedding technique by projecting model outputs into hyperbolic space, and [15], which integrates hyperbolic geometry into various vision transformer architectures, showcasing performance that surpasses their Euclidean counterparts. Hyperbolic methods have also been developed on diverse tasks such as image classification [15, 23, 31], action recognition [18], few-shot learning [20] and object segmentation [21, 60]. Moreover, recent developments have introduced hyperbolic geometry for neural networks including fully connected layers [46], convolutional neural networks [3], graph neural networks [9, 36], and attention network [22], thereby facilitating a deeper integration of hyperbolic geometry into deep learning regime.", + "bbox": [ + 89, + 500, + 483, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 799, + 181, + 814 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first introduce the task in Sec. 3.1, then move to a review of baselines in Sec. 3.2. Afterwards, the geometry mapping and training details of our framework are described in Sec. 3.3 and Sec. 3.4. Lastly, the label assignment details are outlined in Sec. 3.5.", + "bbox": [ + 89, + 824, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Statement", + "text_level": 1, + "bbox": [ + 513, + 90, + 697, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "GCD aims to learn a model capable of accurately classifying unlabelled samples from known categories while simultaneously clustering those from unknown categories. Consider an unlabelled dataset denoted as $\\mathbf{D}_u = \\left\\{\\left(\\mathbf{x}_i^u,y_i^u\\right)\\right\\} \\in$ $\\mathbf{X}\\times \\mathbf{Y}_u$ and a labelled dataset represented as $\\mathbf{D}_l =$ $\\left\\{\\left(\\mathbf{x}_i^l,y_i^l\\right)\\right\\} \\in \\mathbf{X}\\times \\mathbf{Y}_l,$ where $\\mathbf{Y}_u$ and $\\mathbf{Y}_l$ denote the respective label sets. The unlabelled dataset comprises samples from both known and unknown categories, i.e., specifically $\\mathbf{Y}_l\\subset \\mathbf{Y}_u$ . Let the number of labelled categories be denoted by $M = |\\mathbf{Y}_l|$ . We assume that the total number of categories, $K = |\\mathbf{Y}_l\\cup \\mathbf{Y}_u|$ , is known, as posited in prior works [26, 53, 59]. In scenarios where this information is unavailable, alternative methods such as those proposed in [24, 51] can be employed to yield a reliable estimation.", + "bbox": [ + 511, + 112, + 906, + 324 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Review of Baselines", + "text_level": 1, + "bbox": [ + 513, + 332, + 700, + 347 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Non-parametric Baseline. [51] formalizes the GCD task and proposes a non-parametric baseline. The approach involves finetuning the pre-trained DINO [7] model [14] to enhance the learned representation. The loss function comprises a supervised contrastive loss, which operates on the labelled samples, and a self-supervised contrastive loss, which operates on all the samples. Specifically, given two randomly augmented views $\\mathbf{x}_i$ and $\\mathbf{x}_i'$ for the same image in a mini-batch $B$ , the self-supervised contrastive loss is:", + "bbox": [ + 511, + 354, + 906, + 491 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r e p} ^ {u} = \\frac {1}{| B |} \\sum_ {i \\in B} - \\log \\frac {\\exp \\left(\\mathbf {z} _ {i} \\cdot \\mathbf {z} _ {i} ^ {\\prime} / \\tau_ {r}\\right)}{\\sum_ {j} ^ {j \\neq i} \\exp \\left(\\mathbf {z} _ {i} \\cdot \\mathbf {z} _ {j} ^ {\\prime} / \\tau_ {r}\\right)}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 500, + 906, + 527 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the feature $\\mathbf{z}_i = \\rho_r(\\phi (\\mathbf{x}_i))$ is a $\\ell_2$ -normalized vector and $\\mathbf{z}_i^\\prime$ represents feature from another view $\\mathbf{x}_i^\\prime$ . Here, $\\phi$ refers to the backbone network, $\\rho_r$ denotes the projection head, and $\\tau_{r}$ stands as the temperature parameter used for scaling the features. The supervised contrastive loss for labelled samples is:", + "bbox": [ + 511, + 535, + 905, + 626 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r e p} ^ {s} = \\frac {1}{| B _ {l} |} \\sum_ {i \\in B _ {l}} \\frac {1}{| N _ {i} |} \\sum_ {q \\in N _ {i}} - \\log \\frac {\\exp \\left(\\mathbf {z} _ {i} \\cdot \\mathbf {z} _ {q} / \\tau_ {r}\\right)}{\\sum_ {j} ^ {j \\neq i} \\exp \\left(\\mathbf {z} _ {i} \\cdot \\mathbf {z} _ {j} / \\tau_ {r}\\right)}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 635, + 903, + 665 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $N_{i}$ is the index set for all other images in the labelled mini-batch $B_{l} \\subset B$ having the same label as $\\mathbf{x}_{i}$ . The overall representation learning loss is then: $\\mathcal{L}_{rep} = (1 - \\lambda_b)\\mathcal{L}_{rep}^u + \\lambda_b\\mathcal{L}_{rep}^s$ , where $\\lambda_{b}$ is a balance factor.", + "bbox": [ + 511, + 672, + 906, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Parametric Baseline. [59] introduces a robust parametric GCD baseline, which has been widely adopted in the field ever since [53, 55]. This method employs a parametric classifier implemented in a self-distillation framework [7]. The classifier is randomly initialized with $K$ normalized category prototypes $\\mathbf{C} = \\{\\mathbf{c}_1,\\dots,\\mathbf{c}_K\\}$ . For a randomly augmented view $\\mathbf{x}_i$ and its corresponding normalized hidden feature vector $\\mathbf{h}_i = \\phi (\\mathbf{x}_i) / ||\\phi (\\mathbf{x}_i)||$ , the output probability for the $k$ -th category is given by:", + "bbox": [ + 511, + 734, + 905, + 871 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {p} _ {i} ^ {(k)} = \\frac {\\exp \\left(\\mathbf {h} _ {i} \\cdot \\mathbf {c} _ {k} / \\tau_ {s}\\right)}{\\sum_ {j = 1} ^ {K} \\exp \\left(\\mathbf {h} _ {i} \\cdot \\mathbf {c} _ {j} / \\tau_ {s}\\right)}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 878, + 903, + 904 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9893", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\tau_{s}$ is the scaling temperature for the 'student' view. The soft label $\\mathbf{q}_i$ is generated by the teacher' view with a sharper temperature $\\tau_{t}$ based on another augmented view in a similar manner. The self-distillation loss for the two views is then computed using the cross-entropy loss function: $\\ell_{ce}(\\mathbf{q}',\\mathbf{p}) = -\\sum_{j = 1}^{K}\\mathbf{q}'^{(j)}\\log \\mathbf{p}^{(j)}$ . The unsupervised loss is then computed by aggregating contributions from all samples in the mini-batch $B$ as follows:", + "bbox": [ + 89, + 90, + 483, + 212 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c l s} ^ {u} = \\frac {1}{| B |} \\sum_ {i \\in B} \\ell_ {c e} \\left(\\mathbf {q} _ {i} ^ {\\prime}, \\mathbf {p} _ {i}\\right) - \\xi \\mathcal {H} (\\bar {\\mathbf {p}}), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 223, + 482, + 244 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\overline{\\mathbf{p}} = \\frac{1}{2|B|}\\sum_{i\\in B}(\\mathbf{p}_i + \\mathbf{p}_i')$ denotes the mean prediction across the mini-batch. The mean entropy is defined as: $\\mathcal{H}(\\overline{\\mathbf{p}}) = -\\sum_{j = 1}^{K}\\overline{\\mathbf{p}}^{(j)}\\log \\overline{\\mathbf{p}}^{(j)}$ , weighted by $\\xi$ .", + "bbox": [ + 89, + 257, + 482, + 305 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the labelled samples, the supervised classification loss is written as $\\mathcal{L}_{cls}^{s} = \\frac{1}{|B_{l}|}\\sum_{i\\in B_{l}}\\ell_{ce}(\\mathbf{p}_{i},\\mathbf{y}_{i})$ , where $\\mathbf{y}_i$ represents the one-hot vector corresponding to the ground-truth label $y_{i}$ . The overall objective is $\\mathcal{L}_{cls} = (1 - \\lambda_b)\\mathcal{L}_{cls}^u + \\lambda_b\\mathcal{L}_{cls}^s$ . Integrating this with the representation learning loss $\\mathcal{L}_{rep}$ adopted from [51], the comprehensive training objective is expressed as: $\\mathcal{L}_{gcd} = \\mathcal{L}_{cls} + \\mathcal{L}_{rep}$ . Through training with $\\mathcal{L}_{gcd}$ on both $\\mathbf{D}_l$ and $\\mathbf{D}_u$ , the classifier is empowered to directly predict labels for the unlabelled samples after the training process concludes.", + "bbox": [ + 89, + 305, + 483, + 458 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Hyperbolic Space for Category Discovery", + "text_level": 1, + "bbox": [ + 89, + 468, + 444, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As previously discussed, object parts are critical for facilitating knowledge transfer from labelled categories to unseen ones in GCD. Each sample inherently contains object parts that reside within a hierarchical structure. Moreover, existing GCD methods [44, 45] emphasize the intercategory hierarchy to enhance the clustering performance of unlabelled samples in Euclidean or spherical spaces. However, the geometry of representation space limits their ability to effectively capture other kinds of hierarchy [31]. In contrast, hyperbolic space, characterized by its property of exponential volume growth with respect to the radius [15], emerges as a more suitable space for GCD.", + "bbox": [ + 88, + 489, + 483, + 671 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Hyperbolic space $\\mathbb{H}^n$ is defined as an $n$ -dimensional Riemannian manifold exhibiting constant negative curvature, and it encompasses several analytic models [5]. Following previous literature [15, 31], we employ the Poincaré ball [39] model. In this model, the hyperbolic space is represented as an $n$ -dimensional ball $\\mathbb{D}_c^n = \\{\\mathbf{a}\\in \\mathbb{R}^n\\mid c\\| \\mathbf{a}\\| ^2 < 1\\}$ with curvature value $-c^2$ , where $c$ is the non-negative curvature parameter. The manifold is equipped with the Riemannian metric $g^{\\mathbb{D}} = \\lambda_c^2 g^{\\mathbb{E}}$ where $\\lambda_{c}(\\mathbf{a}) = \\frac{2}{1 - c\\|\\mathbf{a}\\|^2}$ is the conformal factor and $g^{\\mathbb{E}}$ is the identity metric $\\mathbf{I}_n$ in Euclidean space. In this way, the local distances are scaled by the factor $\\lambda_{c}$ approaching infinity near the boundary of the ball. This gives rise to the exponential expansion property of hyperbolic spaces, unlike the polynomial expansion in Euclidean space. However, hyperbolic", + "bbox": [ + 89, + 672, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "space is not vector space and thus operations such as addition can not be directly conducted. To address this problem, we leverage the gyrovector formalism [48]. For a pair of points $\\mathbf{a},\\mathbf{b}\\in \\mathbb{D}_c^n$ , their Möbius addition is defined as:", + "bbox": [ + 511, + 90, + 906, + 152 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {a} \\oplus_ {c} \\mathbf {b} = \\frac {(1 + 2 c \\langle \\mathbf {a} , \\mathbf {b} \\rangle + c \\| \\mathbf {b} \\| ^ {2}) \\mathbf {a} + (1 - c \\| \\mathbf {a} \\| ^ {2}) \\mathbf {b}}{1 + 2 c \\langle \\mathbf {a} , \\mathbf {b} \\rangle + c ^ {2} \\| \\mathbf {a} \\| ^ {2} \\| \\mathbf {b} \\| ^ {2}}. \\qquad (5)\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 161, + 906, + 186 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The hyperbolic distance between them is then:", + "bbox": [ + 511, + 190, + 823, + 204 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {D} _ {\\mathbb {H}} (\\mathbf {a}, \\mathbf {b}) = \\frac {2}{\\sqrt {c}} \\operatorname {a r c t a n h} \\left(\\sqrt {c} \\| - \\mathbf {a} \\oplus_ {c} \\mathbf {b} \\|\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 215, + 906, + 236 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "When $c\\to 0$ , the hyperbolic distance (Eq. 6) closes to the Euclidean distance $\\lim_{c\\to 0}\\mathcal{D}_{\\mathbb{H}}(\\mathbf{a},\\mathbf{b}) = 2\\| \\mathbf{a} - \\mathbf{b}\\|$", + "bbox": [ + 511, + 244, + 905, + 275 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. HypCD", + "text_level": 1, + "bbox": [ + 511, + 284, + 609, + 300 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As illustrated in Fig. 3, we propose a unified framework, HypCD, for category discovery in hyperbolic space, incorporating both parametric [59] and non-parametric GCD approaches. Given two randomly augmented views, we initially obtain the respective Euclidean feature vectors $\\mathbf{z}_i$ and $\\mathbf{z}_i'$ through a self-supervised pretrained backbone [7, 40]. Subsequently, the feature embeddings are mapped into hyperbolic space $\\mathbb{H}^n$ using exponential mapping, facilitating representation learning within this exponentially growing space to more effectively capture and utilize the hierarchical relationships inherent in the training data.", + "bbox": [ + 511, + 306, + 905, + 470 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The exponential mapping [31] serves as a bijective projection between Euclidean space $\\mathbb{E}^n$ and hyperbolic space $\\mathbb{H}^n$ . The projection of tangent vector $\\mathbf{z}$ from $\\mathbb{E}^n$ to $\\mathbb{H}^n$ is formulated as:", + "bbox": [ + 511, + 472, + 905, + 532 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\exp_ {\\mathbf {o}} ^ {c} (\\mathbf {z}) = \\mathbf {o} \\oplus_ {c} \\left(\\tanh \\left(\\sqrt {c} \\frac {\\lambda_ {\\mathbf {o}} ^ {c} \\| \\mathbf {z} \\|}{2}\\right) \\frac {\\mathbf {z}}{\\sqrt {c} \\| \\mathbf {z} \\|}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 539, + 906, + 574 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\oplus_{c}$ is the Mobius addition, as introduced in Eq. 5 and $\\mathbf{o}$ represents the base point of the mapping. To address the issue of gradient vanishing [23] near the boundary of the Poincaré ball during training, we implement a feature clipping operation in $\\mathbb{E}^n$ prior to the exponential mapping. The operation is defined as: $\\mathcal{C}(\\mathbf{z}) = \\min \\{1,\\frac{r}{||\\mathbf{z}||}\\} \\cdot \\mathbf{z}$ , where $r$ denotes the clipping value. For the feature vector $\\mathbf{z}_i$ in $\\mathbb{E}^n$ , the corresponding mapped feature in $\\mathbb{H}^n$ is expressed as $\\mathcal{M}(\\mathbf{z}_i) = \\exp_{\\mathbf{o}}^{c}(\\mathcal{C}(\\mathbf{z}_{i}))$ . The same operation will also be applied to the other feature vector $\\mathbf{z}_i'$ .", + "bbox": [ + 511, + 582, + 905, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As described in Sec.3.2, both parametric [51] and nonparametric [59] baselines utilize the same representation learning method. In our framework, we implement a consistent representation learning solution in hyperbolic space for them (Fig.3(a)). For parametric approaches, a hyperbolic parametric classifier is employed (Fig. 3(b)). We will introduce these components in detail subsequently.", + "bbox": [ + 511, + 734, + 905, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Hyperbolic Representation Learning. Following prior attempts [45, 51, 59], we incorporate both self-supervised and supervised contrastive learning into our framework. However, our approach uniquely operates within hyperbolic", + "bbox": [ + 511, + 840, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9894", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ed52ceb0b1900ddee19b23f6af1d6331bef6de88be7adb5c9f5c0c05431eca90.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 92, + 596, + 217 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7ab57b7d5ec05ca663bcf6556fe08cf1a501351ba238d6bd7a0c18438767f5e7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 598, + 93, + 893, + 215 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2138d4663106c5d07f1aedce0e931db7dfed1e924a16e028f65d8579b2ac8c94.jpg", + "image_caption": [ + "Figure 3. Overall pipeline of our HypCD framework for parametric and non-parametric GCD baselines. (a) Hyperbolic representation learning. (b) Hyperbolic classifier. (c) Non-parametric label assignment. (d) Parametric label assignment." + ], + "image_footnote": [], + "bbox": [ + 94, + 218, + 594, + 340 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8a985e64d6e865db5b5b45af74fbc24dae17c89265f3d9603a7eac2402797f71.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 599, + 218, + 895, + 338 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "space. Furthermore, unlike previous GCD methods that exclusively utilize cosine distance [51, 55, 59] (angle-based) or Euclidean distance [45] (distance-based) for calculating pairwise similarity, we propose a hybrid approach that combines both distance-based and angle-based losses. Such integration has been shown to be more effective for model optimization in hyperbolic space [18]. First, the unified form of self-supervised contrastive loss can be defined as:", + "bbox": [ + 89, + 383, + 483, + 505 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {u} = \\frac {1}{| B |} \\sum_ {i \\in B} - \\log \\frac {\\exp \\left(\\mathcal {S} \\left(\\mathcal {M} \\left(\\mathbf {z} _ {i}\\right) , \\mathcal {M} \\left(\\mathbf {z} _ {i} ^ {\\prime}\\right)\\right) / \\tau_ {r}\\right)}{\\sum_ {j} ^ {j \\neq i} \\exp \\left(\\mathcal {S} \\left(\\mathcal {M} \\left(\\mathbf {z} _ {i}\\right) , \\mathcal {M} \\left(\\mathbf {z} _ {j} ^ {\\prime}\\right)\\right) / \\tau_ {r}\\right)}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 513, + 482, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Similarly, the supervised contrastive loss is unified as:", + "bbox": [ + 89, + 551, + 449, + 566 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {s} = \\frac {1}{| B _ {l} |} \\sum_ {i \\in B _ {l}} \\frac {1}{| N _ {i} |} \\sum_ {q \\in N _ {i}} - \\log \\frac {\\exp \\left(\\mathcal {S} \\left(\\mathcal {M} \\left(\\mathbf {z} _ {i}\\right) , \\mathcal {M} \\left(\\mathbf {z} _ {q}\\right)\\right) / \\tau_ {r}\\right)}{\\sum_ {j \\neq i} \\exp \\left(\\mathcal {S} \\left(\\mathcal {M} \\left(\\mathbf {z} _ {i}\\right) , \\mathcal {M} \\left(\\mathbf {z} _ {j}\\right)\\right) / \\tau_ {r}\\right)}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 574, + 482, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $S$ denotes the similarity function, which can be either distance-based or angle-based. For distance-based contrastive loss, we utilize $S_{d} = -\\mathcal{D}_{\\mathbb{H}}$ as the similarity function, which is formally computed using negative Euclidean distance in prior methods [45]. For angle-based contrastive loss, we employ the cosine similarity, formulated as:", + "bbox": [ + 89, + 612, + 483, + 703 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} _ {a} \\left(\\mathcal {M} \\left(\\mathbf {z} _ {i}\\right), \\mathcal {M} \\left(\\mathbf {z} _ {i} ^ {\\prime}\\right)\\right) = \\frac {\\mathcal {M} \\left(\\mathbf {z} _ {i}\\right) \\cdot \\mathcal {M} \\left(\\mathbf {z} _ {i} ^ {\\prime}\\right)}{\\left| \\left| \\mathcal {M} \\left(\\mathbf {z} _ {i}\\right) \\right| \\right| \\cdot \\left| \\left| \\mathcal {M} \\left(\\mathbf {z} _ {i} ^ {\\prime}\\right) \\right| \\right|}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 710, + 480, + 736 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since hyperbolic space is conformal with Euclidean space, cosine similarity remains equivalent in both $\\mathbb{E}^n$ and $\\mathbb{H}^n$ .", + "bbox": [ + 89, + 743, + 482, + 772 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The final supervised and self-supervised hyperbolic contrastive loss is composed of both distance-based and angle-based losses:", + "bbox": [ + 89, + 773, + 483, + 816 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {h r e p} ^ {s} = \\alpha_ {d} \\mathcal {L} _ {d i s} ^ {s} + (1 - \\alpha_ {d}) \\mathcal {L} _ {a n g} ^ {s}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 824, + 480, + 849 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {h r e p} ^ {u} = \\alpha_ {d} \\mathcal {L} _ {d i s} ^ {u} + (1 - \\alpha_ {d}) \\mathcal {L} _ {a n g} ^ {u},\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 845, + 401, + 863 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{hrep}^s$ and $\\mathcal{L}_{hrep}^u$ represent the supervised and self-supervised hyperbolic contrastive loss, respectively. The", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "terms $\\mathcal{L}_{dis}$ and $\\mathcal{L}_{ang}$ correspond to distance-based and angle-based contrastive loss, respectively, obtained by substituting $S$ with $S_{d}$ and $S_{a}$ . Additionally, $\\alpha_{d}$ is the loss weight of distance-based loss. The overall training objective for hyperbolic representation learning is:", + "bbox": [ + 511, + 383, + 905, + 460 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r e p} ^ {\\mathbb {H}} = \\left(1 - \\lambda_ {b} ^ {\\mathbb {H}}\\right) \\mathcal {L} _ {h r e p} ^ {u} + \\lambda_ {b} ^ {\\mathbb {H}} \\mathcal {L} _ {h r e p} ^ {s}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 467, + 903, + 488 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_b^{\\mathbb{H}}$ serves as the balancing factor between the supervised and unsupervised losses.", + "bbox": [ + 511, + 494, + 903, + 526 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Hyperbolic Classifier. To enhance the parametric baseline with hyperbolic geometry, we replace the conventional Euclidean classification head—traditionally reliant on a multilayer perceptron (MLP) in Euclidean space—with its hyperbolic counterpart, the hyperbolic feed forward network (HypFFN). The hyperbolic linear layer [19] exhibits greater alignment with the baseline [59], and we experimentally find that it outperforms the hyperbolic multinomial logistic regression layer. Consider the last linear layer of the MLP; similar to its Euclidean counterpart, the hyperbolic linear layer is parameterized by a weight matrix $\\mathbf{w} \\in \\mathbb{R}^{I \\times K}$ and a bias vector $\\mathbf{s} \\in \\mathbb{R}^{1 \\times K}$ , where $I$ denotes the input feature dimension. Given the hyperbolic feature $\\mathbf{z}_i^{\\mathbb{H}} = \\mathcal{M}(\\mathbf{z}_i) \\in \\mathbb{R}^{1 \\times I}$ , the linear layer operates as HypLinear $(\\mathbf{z}_i^{\\mathbb{H}}, \\mathbf{w}, \\mathbf{s}) = \\operatorname{Proj}[(\\mathbf{w} \\otimes_c \\mathbf{z}_i^{\\mathbb{H}}) \\oplus_c \\mathbf{s}]$ , where $\\oplus_c$ follows Eq. 5. The Möbius matrix-vector multiplication $\\mathbf{v}_i = \\mathbf{w} \\otimes_c \\mathbf{z}_i^{\\mathbb{H}}$ is defined as:", + "bbox": [ + 511, + 527, + 906, + 785 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {1}{\\sqrt {c}} \\tanh \\left(\\frac {\\| \\mathbf {z} _ {i} ^ {\\mathbb {H}} \\mathbf {w} \\| _ {2}}{\\| \\mathbf {z} _ {i} ^ {\\mathbb {H}} \\| _ {2}} \\tanh ^ {- 1} (\\sqrt {c} \\| \\mathbf {z} _ {i} ^ {\\mathbb {H}} \\| _ {2})\\right) \\frac {\\mathbf {z} _ {i} ^ {\\mathbb {H}} \\mathbf {w}}{\\| \\mathbf {z} _ {i} ^ {\\mathbb {H}} \\mathbf {w} \\| _ {2}}. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 791, + 903, + 818 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To ensure numerical stability [19], a safe projection is operated on the result manifold and represented as:", + "bbox": [ + 511, + 824, + 903, + 854 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {P r o j} (\\mathbf {v} _ {i}) = \\left\\{ \\begin{array}{l l} \\frac {\\mathbf {v} _ {i}}{\\| \\mathbf {v} _ {i} \\| _ {2}} \\times \\frac {1 - 1 0 ^ {- 3}}{\\sqrt {c}}, & \\frac {1 - 1 0 ^ {- 3}}{\\sqrt {c}} < \\| \\mathbf {v} _ {i} \\| _ {2} \\\\ \\mathbf {v} _ {i}, & \\text {o t h e r w i s e} \\end{array} . \\right. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 862, + 903, + 904 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9895", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This integration allows our hyperbolic classifier to be seamlessly incorporated into the baseline [59] by substituting the original MLP with HypFFN. For each point in $\\mathbb{H}^n$ , the tangent space at that point serves as a Euclidean subspace, enabling straightforward adaptation of Euclidean operations within this space [10]. Consequently, the cross-entropy loss for the hyperbolic classifier can be expressed as: $\\ell_{ce}^{\\mathbb{H}} = \\ell_{ce}(\\mathrm{HypFFN}(\\mathbf{z}_i^{\\mathbb{H}}),\\mathbf{y}_i)$ . Additionally, we can define the hyperbolic counterpart $\\mathcal{H}^{\\mathbb{H}}$ for the mean entropy $\\mathcal{H}$ . By substituting the original $\\ell_{ce}$ and $\\mathcal{H}$ with our derived $\\ell_{ce}^{\\mathbb{H}}$ and $\\mathcal{H}^{\\mathbb{H}}$ , respectively, we can readily compute the final hyperbolic classifier loss $\\mathcal{L}_{cls}^{\\mathbb{H}}$ , as detailed in Sec. 3.2.", + "bbox": [ + 89, + 90, + 483, + 273 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.5.Label Assignment", + "text_level": 1, + "bbox": [ + 89, + 284, + 266, + 299 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Existing approaches typically employ either a parametric classification head or non-parametric methods, such as semi-supervised $k$ -means [51], for label assignment. In this paper, we do not independently assess these two methods; rather, we integrate both within the HypCD framework as shown in Fig. 3(c) and (d). For non-parametric approaches, including [51] and the recent SelEx [45], we retain the original label assignment strategy by applying semi-supervised $k$ -means clustering directly to feature representations extracted by $\\phi$ in $\\mathbb{E}^n$ . Our empirical results indicate that training in hyperbolic space allows for the transfer of hierarchical structure encoding from $\\mathbb{H}^n$ to $\\mathbb{E}^n$ . Moreover, we find that the operations of $k$ -means in $\\mathbb{E}^n$ are significantly more efficient while maintaining comparable performance. For the parametric baseline exemplified by SimGCD [59], we utilize the hyperbolic classification head to conduct classification within hyperbolic space using the trained hyperbolic classifier HypFFN. Both design choices are theoretically supported by the property of hyperbolic geometry of encoding hierarchical structures, facilitating a more intuitive and effective representation and classifier for GCD.", + "bbox": [ + 89, + 306, + 483, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 89, + 640, + 215, + 656 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Setups and Implementations", + "text_level": 1, + "bbox": [ + 89, + 665, + 346, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We thoroughly evaluate our method across diverse benchmarks, including the generic image recognition datasets CIFAR-10 and CIFAR-100 [34], as well as ImageNet-100 [13]. Additionally, we assess our approach on the Semantic Shift Benchmark (SSB)[52], which includes fine-grained datasets such as CUB[54], StanfordCars [33], and FGVC-Aircraft [38]. For each dataset, we adhere to the data split scheme detailed in [51]. The method involves sampling a subset of all classes as the known ('Old') classes $\\mathbf{Y}_l$ . Subsequently, $50\\%$ of the images from these known classes are utilized to construct $\\mathbf{D}_l$ , while the remaining images are designated as the unlabelled data $\\mathbf{D}_u$ . Evaluation Metrics. We evaluate the performance using the clustering accuracy (ACC) as defined in the litera", + "bbox": [ + 89, + 688, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ture [51]. The $ACC$ on $\\mathbf{D}_u$ is computed as follows, given the ground truth $y_{i}$ and the predicted labels $\\hat{y}_i$ : $ACC = \\frac{1}{|\\mathbf{D}_u|}\\sum_{i = 1}^{|\\mathbf{D}_u|}\\mathbb{1}(y_i = h(\\hat{y}_i))$ , where $h$ denotes the optimal permutation that aligns the predicted cluster assignments with the ground-truth labels. The $ACC$ values for the 'All', 'Old' and 'New' classes are reported separately.", + "bbox": [ + 511, + 90, + 903, + 184 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. We evaluate HypCD against the non-parametric baseline GCD [51], the parametric baseline SimGCD [59], and the SOTA method SelEx [45], utilizing both DINO [7] and DINoV2 [40] pretrained weights. Detailed information regarding SelEx can be found in the supplementary materials. For GCD [51], the output dimension of the projection head $\\rho_{r}$ is 256. In the case of SimGCD [59], the feature dimension from backbone $\\phi$ is 768. $\\rho_{r}$ and the final block of $\\phi$ are optimized using the SGD optimizer, with an initial learning rate of 0.1, which is decayed to 0.001 over time according to a cosine annealing schedule. The HypFFN is optimized using the Riemannian Adam optimizer [4], with a constant learning rate of 0.01. All models are trained for 200 epochs using a batch size of 128. The curvature parameter $c$ is set to 0.05 for the fine-grained datasets and 0.01 for the generic datasets. Following baselines, the balancing factor $\\lambda_b^{\\mathbb{H}}$ is set to 0.35. By default, the loss weight $\\alpha_{d}$ increases linearly from 0 to 1.0.", + "bbox": [ + 511, + 185, + 906, + 457 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Quantitative Comparison", + "text_level": 1, + "bbox": [ + 511, + 467, + 746, + 483 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our method with recent GCD methods (including ORCA [6], GCD [51], XCon [16], OpenCon [47], PromptCAL [61], DCCL [42], GPC [64], CiPR [27], SimGCD [59], $\\mu$ GCD [53], InfoSieve [44], SPTNet [55], CMS [12], AMEND [2] and SelEx [45]) and report the results in Tab. 1. The evaluation encompasses performance on the SSB benchmark [52] and generic datasets [13, 34]. The hyperbolic methods applying our HypCD framework are indicated by the 'Hyp-' prefix.", + "bbox": [ + 511, + 488, + 906, + 626 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b3dc1a23b1372253764360b526da3ee5f3d8e0e583c844b07bad0b328d1d1f54.jpg", + "image_caption": [ + "Figure 4. Comparison of baseline and hyperbolic counterparts on the SSB. Left: 'All' ACC (higher is better). Right: Discrepancy between 'Old' and 'New' ACC (smaller is better)." + ], + "image_footnote": [], + "bbox": [ + 519, + 643, + 705, + 785 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/16a3a827d444bb117b2bbcf99c6f41a005eb7276158bb6c533c0e4a90a503dc0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 728, + 645, + 901, + 785 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results on SSB. The performance of the GCD methods on the SSB benchmark, utilizing both DINO [7] and DINOv2 [40] pretrained weights, is summarized in the left section of Tab. 1. Besides, we provide a comparative analy-", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9896", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/46aa01a5db04dda235b0a323e66ab90a73e8c1c137c7b141175f767bf997abdf.jpg", + "table_caption": [ + "Table 1. Comparison of GCD methods on the SSB [52] benchmark, CIFAR-10 [34], CIFAR-100 [34] and ImageNet-100 [13] datasets. Results are reported in ACC across the 'All', 'Old' and 'New' categories." + ], + "table_footnote": [ + "\\*results from our implementation." + ], + "table_body": "
MethodCUB [54]Stanford-Cars [33]FGVC-Aircraft [38]CIFAR-10 [34]CIFAR-100 [34]ImageNet-100 [13]
AllOldNewAllOldNewAllOldNewAllOldNewAllOldNewAllOldNew
DINOORCA [6]36.343.832.631.632.031.431.942.226.969.077.452.073.592.663.981.886.279.6
XCon [16]52.154.351.040.558.831.747.744.449.496.097.395.474.281.260.377.693.569.7
OpenCon [47]54.763.854.749.178.632.7---------84.093.881.2
PromptCAL [61]62.964.462.150.270.140.652.252.252.397.996.698.581.284.275.383.192.778.3
DCCL [42]63.560.864.943.155.736.2---96.396.596.975.376.870.280.590.576.2
GPC [64]52.055.547.538.258.927.443.340.744.890.697.687.075.484.660.175.393.466.7
PIM [11]62.775.756.243.166.931.6---94.797.493.378.384.266.583.195.377.0
μGCD [53]65.768.064.656.568.150.953.855.453.0---------
InfoSieve [44]69.477.965.255.774.846.456.363.752.594.897.793.478.382.270.580.593.873.8
CiPR [27]57.158.755.647.061.540.1---97.797.597.781.582.479.780.584.978.3
SPTNet [55]65.868.865.159.079.249.359.361.858.197.395.098.681.384.375.685.493.281.4
CMS [12]68.276.564.056.976.147.656.063.452.3---82.385.775.584.795.679.2
AMEND [2]64.975.659.652.861.848.356.473.348.296.894.697.881.079.983.383.292.978.3
GCD [51]51.356.648.739.057.629.945.041.146.991.597.988.273.076.266.574.189.866.3
Hyp-GCD61.067.058.050.860.945.848.243.650.592.997.590.674.080.062.080.492.574.4
SimGCD [59]60.365.657.753.871.945.054.259.151.897.195.198.180.181.277.883.093.177.9
Hyp-SimGCD64.865.864.262.873.457.758.758.958.596.895.997.282.483.181.286.593.783.0
SelEx [45]73.675.372.858.575.650.357.164.753.395.998.194.882.385.376.383.193.677.8
Hyp-SelEx79.875.881.862.980.054.765.967.365.196.797.696.382.485.177.086.894.682.8
DINOv2μGCD [53]74.075.973.176.191.068.966.368.765.1---------
CiPR [27]78.373.480.866.777.061.8---99.098.799.290.389.093.188.287.688.5
SPTNet [55]76.379.574.6------------90.196.187.1
GCD [51]71.971.272.365.767.864.755.447.959.297.899.097.179.684.569.978.589.573.0
Hyp-GCD75.675.175.972.880.469.162.770.059.097.598.996.884.587.578.582.992.478.2
SimGCD [59]71.578.168.371.581.966.663.969.960.998.796.799.788.589.287.289.995.587.1
Hyp-SimGCD77.677.977.482.585.881.076.470.379.498.997.799.591.590.094.691.996.289.8
SelEx [45]87.485.188.582.293.776.779.882.378.698.5*98.8*98.5*87.7*90.8*81.5*90.9*96.2*88.3*
Hyp-SelEx90.785.393.483.893.379.283.482.084.198.698.198.988.691.582.892.396.490.2
", + "bbox": [ + 91, + 109, + 906, + 503 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sis between the three baseline methods and their hyperbolic counterparts with DINO backbone in Fig. 4. Our hyperbolic methods consistently outperform their Euclidean counterparts, with particularly strong results observed when utilizing the DINOv2 backbone. Among the evaluated methods, Hyp-SelEx achieves the highest average accuracy (ACC) across all datasets, notably excelling on the CUB dataset, where it records an accuracy of $79.8\\%$ for the 'All' classes with DINO and $90.7\\%$ with DINOv2, establishing it as the leading approach. More strikingly, on the Stanford-Cars dataset, our Hyp-GCD method outperforms the baseline by $11.8\\%$ , $13.3\\%$ and $15.9\\%$ in terms of ACC for the 'All', 'Old' and 'New' categories, respectively. Fig. 4 (left) illustrates the average ACC on 'All' categories across the three datasets in the SSB benchmark, indicating that our hyperbolic methods surpass the baseline by a margin of at least $6.0\\%$ . Furthermore, as shown in Fig. 4 (right), hyperbolic methods exhibit a consistently smaller ACC gap between 'Old' and 'New' classes, highlighting the effectiveness of HypCD in enhancing knowledge transfer from known to unseen categories. Additionally, DINOv2 outperforms DINO across all methods, underscoring its ability to capture complex data representations more effectively.", + "bbox": [ + 91, + 526, + 480, + 872 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on Generic Datasets. In the right section of Tab. 1,", + "bbox": [ + 91, + 885, + 480, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "we present the results on three widely used generic datasets: CIFAR-10 [34], CIFAR-100 [34], and ImageNet-100 [13]. Our methods demonstrate consistent improvements across all cases, regardless of the backbone employed. Notably, these enhancements are especially significant on CIFAR-100 and ImageNet-100, which present greater challenges compared to CIFAR-10, where performance is nearly saturated. For CIFAR-100, Hyp-SimGCD and Hyp-SelEx achieve the highest accuracy of $82.4\\%$ for the 'All' categories using DINO, while Hyp-SimGCD ranks first with an accuracy of $91.5\\%$ on this metric when utilizing DINOv2, significantly surpassing baseline methods and the previous SOTA. Results on ImageNet-100 further validate the effectiveness of hyperbolic methods; Hyp-SelEx achieves the highest performance across 'All', 'Old', and 'New' categories with both DINO and DINOv2, outperforming the baseline by a margin of up to $3.7\\%$ .", + "bbox": [ + 511, + 526, + 903, + 782 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Impact of Hyperparameters", + "text_level": 1, + "bbox": [ + 511, + 800, + 764, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Manifold Curvature. Building on previous studies [1, 31] that explore the application of hyperbolic geometry across various tasks, the curvature parameter $c$ (as discussed in Sec. 3.3) is a crucial factor influencing performance and may yield different optimal values across datasets and meth", + "bbox": [ + 511, + 825, + 903, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9897", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f8c5ccd41fb9249ca009f1eff4ac1f6c73eabe907395a74ad51682f00b22de3d.jpg", + "table_caption": [ + "Table 2. Experimental results using different $c$ , $r$ and $\\alpha_{d}^{\\max}$ values in Hyp-SimGCD with DINO [7] pre-trained backbone. Results on the CUB [54] and CIFAR-100 [34] datasets are reported." + ], + "table_footnote": [], + "table_body": "
parameterStanford-Cars [33]CIFAR-100 [34]
AllOldNewAllOldNew
c = 0.0161.474.455.182.483.181.2
c = 0.0562.873.457.781.684.076.7
c = 0.162.375.156.181.182.378.8
r = 1.060.072.953.782.483.181.2
r = 1.561.275.754.281.282.478.8
r = 2.362.873.457.780.181.178.3
αdmax= 0.159.677.551.081.382.479.1
αdmax= 0.562.077.254.682.483.181.2
αdmax= 1.062.873.457.778.983.569.7
", + "bbox": [ + 101, + 130, + 464, + 292 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ods. Intuitively, as the value of $c$ approaches 0, the radius tends toward infinity, causing the Poincaré ball to flatten and resemble Euclidean space; conversely, larger values of $c$ correspond to a steeper configuration. The widely accepted range for $c$ is between 0.01 and 0.3 [15], with larger values exceeding this range resulting in performance degradation. In our experiments, we evaluate different curvature values of 0.01, 0.05, and 0.1 using Hyp-SimGCD, as presented in Tab. 2. Our findings indicate that the optimal curvature values differ between generic and fine-grained datasets. For fine-grained datasets such as CUB [54], the optimal value is 0.05, while for generic datasets like CIFAR-100, a value of 0.01 proves to be more effective.", + "bbox": [ + 88, + 306, + 485, + 503 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Clipping Value. As articulated in [23], feature clipping has emerged as a standard technique for training hyperbolic neural networks. In our framework, we also observe that it plays a crucial role in category discovery performance. In line with the methodology outlined in [23], we investigate a range of clipping values, specifically 1.0, 1.5, and 2.3. The results shown in the second row of Tab. 2 demonstrate that optimal clipping values vary between fine-grained and generic datasets. For fine-grained datasets like CUB [54], the optimal clipping value is determined to be 2.3. Conversely, for generic datasets like CIFAR-100 [34], a clipping value of 1.0 is shown to be more effective.", + "bbox": [ + 89, + 505, + 483, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Loss Weight. As detailed in Sec. 3.4, we implement a hybrid contrastive loss that combines both distance-based and angle-based components, which is essential for effective optimization in hyperbolic space. A loss weight, denoted as $\\alpha_{d}$ , is introduced to regulate the balance between these two types of losses and linearly increasing from 0. In the initial stages, the model prioritizes optimizing the angle between sample points and progressively shifts focus toward optimizing the hyperbolic distance. Consistent with the observations for curvature and clipping values, the optimal max value $\\alpha_{d}^{\\max}$ , varies considerably between coarse-grained and fine-grained datasets. For fine-grained datasets such as CUB [54], an optimal value of 1.0 is observed, whereas for more generic datasets like CIFAR-100 [34], a", + "bbox": [ + 89, + 688, + 485, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8d885524221404d04d0b3ce20ff4b5ee561b5084e32d531a58210487e8c3e8c0.jpg", + "image_caption": [ + "Figure 5. T-SNE [49] comparison between SimGCD [59] and our Hyp-SimGCD using 40 randomly sampled instances from 10 randomly selected categories of the Stanford-Cars dataset [33]." + ], + "image_footnote": [], + "bbox": [ + 513, + 92, + 903, + 265 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "value of 0.5 is found to yield better performance.", + "bbox": [ + 511, + 316, + 839, + 333 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Qualitative Comparison", + "text_level": 1, + "bbox": [ + 511, + 340, + 735, + 358 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Fig. 5, we present a t-SNE [49] visualization of features extracted from the backbone, represented as $\\mathbf{z}_i = \\phi(\\mathbf{x}_i)$ . This visualization compares SimGCD with our Hyp-SimGCD. On the left side of the figure, the clusters generated by SimGCD appear dispersed. Data points from Class 42, highlighted in pink, are spread across multiple areas, indicating significant overlap and a lack of compactness. In contrast, Hyp-SimGCD creates more distinct and tightly clustered groups, concentrating the data points of Class 42 in a more confined area. This comparison implies that Hyp-SimGCD enhances both intra-class compactness and inter-class separation through our hyperbolic representation and classifier learning method. Importantly, even within the original Euclidean space of the backbone network, Hyp-SimGCD exhibits robust clustering performance, which arises from the properties of hyperbolic space in encoding hierarchical structures.", + "bbox": [ + 509, + 363, + 906, + 621 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 633, + 633, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we investigate a previously overlooked perspective in GCD by utilizing a representation space that captures the hierarchical structure of each sample, instead of the conventional Euclidean or spherical spaces. Our approach leverages the distinctive properties of hyperbolic space, where the volume increases exponentially with radius. This characteristic makes hyperbolic space especially suitable for modelling data possessing hierarchical structures, thereby enhancing representational capacity for category discovery. We propose a simple yet effective framework, HypCD, for integrating hyperbolic geometry into GCD methods. Through extensive experiments with parametric and non-parametric GCD baselines and the SOTA method, our framework consistently demonstrates superior performance on public benchmarks, underscoring the effectiveness of hyperbolic space for category discovery.", + "bbox": [ + 509, + 659, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9898", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work is supported by National Natural Science Foundation of China (Grant No. 62306251), Hong Kong Research Grant Council - Early Career Scheme (Grant No. 27208022) and General Research Fund (Grant No. 17211024), and HKU Seed Fund for Basic Research.", + "bbox": [ + 89, + 90, + 485, + 181 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 193, + 187, + 209 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Mina Ghadimi Atigh, Julian Schoep, Erman Acar, Nanne Van Noord, and Pascal Mettes. Hyperbolic image segmentation. In CVPR, 2022. 7", + "[2] Anwesha Banerjee, Liyana Sahir Kallooriyakath, and Soma Biswas. Amend: Adaptive margin and expanded neighborhood for efficient generalized category discovery. In WACV, 2024. 6, 7", + "[3] Ahmad Bdeir, Kristian Schwethelm, and Niels Landwehr. Fully hyperbolic convolutional neural networks for computer vision. In ICLR, 2024. 3", + "[4] Gary Becigneul and Octavian-Eugen Ganea. Riemannian adaptive optimization methods. In ICLR, 2019. 6", + "[5] James W Cannon, William J Floyd, Richard Kenyon, Walter R Parry, et al. Hyperbolic geometry. Flavors of geometry, 31(59-115):2, 1997. 4", + "[6] Kaidi Cao, Maria Brbic, and Jure Leskovec. Open-world semi-supervised learning. In ICLR, 2022. 3, 6, 7", + "[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 1, 2, 3, 4, 6, 8", + "[8] Fernando Julio Cendra, Bingchen Zhao, and Kai Han. Promptcd: Learning gaussian mixture prompt pool for continual category discovery. In ECCV, 2024. 3", + "[9] Ines Chami, Rex Ying, Christopher Ré, and Jure Leskovec. Hyperbolic graph convolutional neural networks. In NeurIPS, 2019. 3", + "[10] Weize Chen, Xu Han, Yankai Lin, Hexu Zhao, Zhiyuan Liu, Peng Li, Maosong Sun, and Jie Zhou. Fully hyperbolic neural networks. In ACL, 2022. 2, 6", + "[11] Florent Chiaroni, Jose Dolz, Ziko Imtiaz Masud, Amar Mitiche, and Ismail Ben Ayed. Parametric information maximization for generalized category discovery. In ICCV, 2023. 7", + "[12] Sua Choi, Dahiun Kang, and Minsu Cho. Contrastive mean-shift learning for generalized category discovery. In CVPR, 2024. 6, 7", + "[13] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 2, 6, 7", + "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 3", + "[15] Aleksandr Ermolov, Leyla Mirvakhabova, Valentin Khrulkov, Nicu Sebe, and Ivan Oseledets. Hyperbolic vision transformers: Combining improvements in metric learning. In CVPR, 2022. 2, 3, 4, 8" + ], + "bbox": [ + 93, + 218, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Yixin Fei, Zhongkai Zhao, Siwei Yang, and Bingchen Zhao. Xcon: Learning with experts for fine-grained category discovery. In BMVC, 2022. 6, 7", + "[17] Enrico Fini, Enver Sangineto, Stéphane Lathuiliere, Zhun Zhong, Moin Nabi, and Elisa Ricci. A unified objective for novel class discovery. In ICCV, 2021. 3", + "[18] Luca Franco, Paolo Mandica, Bharti Munjal, and Fabio Galasso. Hyperbolic self-paced learning for self-supervised skeleton-based action representations. In ICLR, 2023. 3, 5", + "[19] Octavian Ganea, Gary Bécigneul, and Thomas Hofmann. Hyperbolic neural networks. In NeurIPS, 2018. 5", + "[20] Zhi Gao, Yuwei Wu, Yunde Jia, and Mehrtash Harandi. Curvature generation in curved spaces for few-shot learning. In ICCV, 2021. 3", + "[21] Mina GhadimiAtigh, Julian Schoep, Erman Acar, Nanne van Noord, and Pascal Mettes. Hyperbolic image segmentation. In CVPR, 2022. 3", + "[22] Caglar Gulcehre, Misha Denil, Mateusz Malinowski, Ali Razavi, Razvan Pascanu, Karl Moritz Hermann, Peter Battaglia, Victor Bapst, David Raposo, Adam Santoro, and Nando de Freitas. Hyperbolic attention networks. In ICLR, 2019. 3", + "[23] Yunhui Guo, Xudong Wang, Yubei Chen, and Stella X. Yu. Clipped hyperbolic classifiers are super-hyperbolic classifiers. In CVPR, 2022. 3, 4, 8", + "[24] Kai Han, Andrea Vedaldi, and Andrew Zisserman. Learning to discover novel visual categories via deep transfer clustering. In ICCV, 2019. 1, 2, 3", + "[25] Kai Han, Sylvestre-Alvise Rebuffi, Sebastien Ehrhardt, Andrea Vedaldi, and Andrew Zisserman. Automatically discovering and learning new visual categories with ranking statistics. In ICLR, 2020. 3", + "[26] Kai Han, Sylvestre-Alvise Rebuffi, Sebastien Ehrhardt, Andrea Vedaldi, and Andrew Zisserman. Autonovel: Automatically discovering and learning novel visual categories. IEEE TPAMI, 2021. 3", + "[27] Shaozhe Hao, Kai Han, and Kwan-Yee K Wong. Cipr: An efficient framework with cross-instance positive relations for generalized category discovery. TMLR, 2024. 1, 2, 3, 6, 7", + "[28] Xuhui Jia, Kai Han, Yukun Zhu, and Bradley Green. Joint representation learning and novel category discovery on single-and multi-modal data. In ICCV, 2021. 3", + "[29] Justin Johnson, Ranjay Krishna, Michael Stark, Li-Jia Li, David Shamma, Michael Bernstein, and Li Fei-Fei. Image retrieval using scene graphs. In CVPR, 2015. 2", + "[30] KJ Joseph, Sujoy Paul, Gaurav Aggarwal, Soma Biswas, Piyush Rai, Kai Han, and Vineeth N Balasubramanian. Novel class discovery without forgetting. In ECCV, 2022. 3", + "[31] Valentin Khrulkov, Leyla Mirvakhabova, Evgeniya Ustinova, Ivan Oseledets, and Victor Lempitsky. Hyperbolic image embeddings. In CVPR, 2020. 2, 3, 4, 7", + "[32] Fanjie Kong, Yanbei Chen, Jiarui Cai, and Davide Modolo. Hyperbolic learning with synthetic captions for open-world detection. In CVPR, 2024. 2", + "[33] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In ICCV workshop, 2013. 2, 6, 7, 8" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9899", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[34] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 2, 6, 7, 8", + "[35] Huimin Li, Zhentao Chen, Yunhao Xu, and Junlin Hu. Hyperbolic anomaly detection. In CVPR, 2024. 2", + "[36] Qi Liu, Maximilian Nickel, and Douwe Kiela. Hyperbolic graph neural networks. In NeurIPS, 2019. 3", + "[37] Yuanpei Liu and Kai Han. Debgcd: Debiased learning with distribution guidance for generalized category discovery. In ICLR, 2025. 1, 3", + "[38] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 6, 7", + "[39] Maximilian Nickel and Douwe Kiela. Poincaré embeddings for learning hierarchical representations. In NeurIPS, 2017. 4", + "[40] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 4, 6", + "[41] Jona Otholt, Christoph Meinel, and Haojin Yang. Guided cluster aggregation: A hierarchical approach to generalized category discovery. In WACV, pages 2618-2627, 2024. 2", + "[42] Nan Pu, Zhun Zhong, and Nicu Sebe. Dynamic conceptions of contrastive learning for generalized category discovery. In CVPR, 2023. 3, 6, 7", + "[43] Nan Pu, Wenjing Li, Xingyuan Ji, Yalan Qin, Nicu Sebe, and Zhun Zhong. Federated generalized category discovery. In CVPR, 2024. 3", + "[44] Sarah Rastegar, Hazel Doughty, and Cees Snoek. Learn to categorize or categorize to learn? self-coding for generalized category discovery. In NeurIPS, 2023. 1, 2, 4, 6, 7", + "[45] Sarah Rastegar, Mohammadreza Salehi, Yuki M Asano, Hazel Doughty, and Cees G M Snoek. Selex: Self-expertise in fine-grained generalized category discovery. In ECCV, 2024. 1, 2, 3, 4, 5, 6, 7", + "[46] Ryohei Shimizu, Yusuke Mukuta, and Tatsuya Harada. Hyperbolic neural networks++. In *ICLR*, 2021. 3", + "[47] Yiyou Sun and Yixuan Li. Opencon: Open-world contrastive learning. TMLR, 2022. 6, 7", + "[48] Abraham Albert Ungar. A gyrovector space approach to hyperbolic geometry. Synthesis Lectures on Mathematics and Statistics, 1(1):1-194, 2008. 4", + "[49] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. JMLR, 2008. 8", + "[50] Sagar Vaze, Kai Han, Andrea Vedaldi, and Andrew Zisserman. Open-set recognition: A good closed-set classifier is all you need? In ICLR, 2022. 2", + "[51] Sagar Vaze, Kai Han, Andrea Vedaldi, and Andrew Zisserman. Generalized category discovery. In CVPR, 2022. 1, 2, 3, 4, 5, 6, 7", + "[52] Sagar Vaze, Kai Han, Andrea Vedaldi, and Andrew Zisserman. The semantic shift benchmark. In ICML workshop, 2022. 1, 6, 7", + "[53] Sagar Vaze, Andrea Vedaldi, and Andrew Zisserman. No representation rules them all in category discovery. In NeurIPS, 2023. 3, 6, 7" + ], + "bbox": [ + 91, + 92, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[54] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 2, 6, 7, 8", + "[55] Hongjun Wang, Sagar Vaze, and Kai Han. Sptnet: An efficient alternative framework for generalized category discovery with spatial prompt tuning. In ICLR, 2024. 1, 2, 3, 5, 6, 7", + "[56] Hongjun Wang, Sagar Vaze, and Kai Han. Hilo: A learning framework for generalized category discovery robust to domain shifts. In ICLR, 2025. 3", + "[57] Yu Wang, Zhun Zhong, Pengchong Qiao, Xuxin Cheng, Xiawu Zheng, Chang Liu, Nicu Sebe, Rongrong Ji, and Jie Chen. Discover and align taxonomic context priors for open-world semi-supervised learning. In NeurIPS, 2024. 2", + "[58] Simon Weber, Bar Zongür, Nikita Araslanov, and Daniel Cremers. Flattening the parent bias: Hierarchical semantic segmentation in the poincaré ball. In CVPR, 2024. 2", + "[59] Xin Wen, Bingchen Zhao, and Xiaojuan Qi. Parametric classification for generalized category discovery: A baseline study. In ICCV, 2023. 1, 2, 3, 4, 5, 6, 7, 8", + "[60] Zhenzhen Weng, Mehmet Giray Ogut, Shai Limonchik, and Serena Yeung. Unsupervised discovery of the long-tail in instance segmentation using hierarchical self-supervision. In CVPR, 2021. 3", + "[61] Sheng Zhang, Salman Khan, Zhiqiang Shen, Muzammal Naseer, Guangyi Chen, and Fahad Shahbaz Khan. Promptable: Contrastive affinity learning via auxiliary prompts for generalized novel category discovery. In CVPR, 2023. 6, 7", + "[62] Xinwei Zhang, Jianwen Jiang, Yutong Feng, Zhi-fan Wu, Xibin Zhao, Hai Wan, Mingqian Tang, Rong Jin, and Yue Gao. Grow and merge: a unified framework for continuous categories discovery. In NeurIPS, 2022. 3", + "[63] Bingchen Zhao and Kai Han. Novel visual category discovery with dual ranking statistics and mutual knowledge distillation. In NeurIPS, 2021. 1, 3", + "[64] Bingchen Zhao, Xin Wen, and Kai Han. Learning semi-supervised gaussian mixture models for generalized category discovery. In ICCV, 2023. 3, 6, 7", + "[65] Zhun Zhong, Enrico Fini, Subhankar Roy, Zhiming Luo, Elisa Ricci, and Nicu Sebe. Neighborhood contrastive learning for novel class discovery. In CVPR, 2021. 3" + ], + "bbox": [ + 516, + 92, + 903, + 672 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9900", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Hyperbolic Category Discovery/103e7fd4-0c6b-413e-8aae-04fa73b4770b_model.json b/2025/Hyperbolic Category Discovery/103e7fd4-0c6b-413e-8aae-04fa73b4770b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2655069fc82e931d05bec2ca05d4b7226a49d3c2 --- /dev/null +++ b/2025/Hyperbolic Category Discovery/103e7fd4-0c6b-413e-8aae-04fa73b4770b_model.json @@ -0,0 +1,2233 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.339, + 0.131, + 0.66, + 0.154 + ], + "angle": 0, + "content": "Hyperbolic Category Discovery" + }, + { + "type": "text", + "bbox": [ + 0.319, + 0.18, + 0.677, + 0.2 + ], + "angle": 0, + "content": "Yuanpei Liu* Zhenqi He* Kai Han†" + }, + { + "type": "text", + "bbox": [ + 0.32, + 0.2, + 0.678, + 0.218 + ], + "angle": 0, + "content": "Visual AI Lab, The University of Hong Kong" + }, + { + "type": "text", + "bbox": [ + 0.275, + 0.22, + 0.724, + 0.234 + ], + "angle": 0, + "content": "{ypliu0,zhenqi_he}@connect.hku.hk kaihanx@hku.hk" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.301, + 0.486, + 0.739 + ], + "angle": 0, + "content": "Generalized Category Discovery (GCD) is an intriguing open-world problem that has garnered increasing attention. Given a dataset that includes both labelled and unlabelled images, GCD aims to categorize all images in the unlabelled subset, regardless of whether they belong to known or unknown classes. In GCD, the common practice typically involves applying a spherical projection operator at the end of the self-supervised pretrained backbone, operating within Euclidean or spherical space. However, both of these spaces have been shown to be suboptimal for encoding samples that possesses hierarchical structures. In contrast, hyperbolic space exhibits exponential volume growth relative to radius, making it inherently strong at capturing the hierarchical structure of samples from both seen and unseen categories. Therefore, we propose to tackle the category discovery challenge in the hyperbolic space. We introduce HypCD, a simple Hyperbolic framework for learning hierarchy-aware representations and classifiers for generalized Category Discovery. HypCD first transforms the Euclidean embedding space of the backbone network into hyperbolic space, facilitating subsequent representation and classification learning by considering both hyperbolic distance and the angle between samples. This approach is particularly helpful for knowledge transfer from known to unknown categories in GCD. We thoroughly evaluate HypCD on public GCD benchmarks, by applying it to various baseline and state-of-the-art methods, consistently achieving significant improvements. Project page: https://visual-ai.github.io/hypcd/" + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.766, + 0.222, + 0.782 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.791, + 0.484, + 0.868 + ], + "angle": 0, + "content": "Recently, category discovery – initially explored as novel category discovery (NCD) [24] and subsequently extended to generalized category discovery (GCD) [51] – has emerged as an intriguing open-world problem, gaining increasing attention. GCD addresses the challenges posed" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.269, + 0.902, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.421, + 0.902, + 0.567 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.569, + 0.907, + 0.64 + ], + "angle": 0, + "content": "Figure 1. (a) Spherical-based vs. Hyperbolic-based methods, where hyperbolic space better accommodates variations in scale and improves connections between samples. (b) Average ACC comparison of our method and previous SOTA across 'All', 'Old', and 'New' categories on the SSB [52] benchmark using DINO [7]." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.656, + 0.907, + 0.806 + ], + "angle": 0, + "content": "by partially labelled datasets, where the unlabelled subset may contain instances from both seen and unseen classes. The goal is to leverage knowledge from labelled data to effectively categorize the unlabelled data. Based on the way to predict category index, existing GCD methods can be broadly classified into two types: non-parametric methods [27, 44, 45, 51] and parametric methods [37, 55, 59]. Non-parametric methods predict category index based on feature clustering while parametric methods utilize a parametric classifier." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.902 + ], + "angle": 0, + "content": "As shown in the GCD literature [51], object parts are effective for knowledge transfer from seen to unseen categories, which is crucial for novel category discovery. Methods have been developed to explicitly learn better local features by learning pixel-level prompts around local image regions [55] or utilizing part-level features [63]. However," + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.221, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.236, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.236, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "9891" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.09, + 0.329, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.146, + 0.193, + 0.273, + 0.204 + ], + "angle": 0, + "content": "(a) Inter-Category Relation" + }, + { + "type": "image", + "bbox": [ + 0.338, + 0.091, + 0.471, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.342, + 0.193, + 0.468, + 0.204 + ], + "angle": 0, + "content": "(b) Intra-Category Relation" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.206, + 0.483, + 0.248 + ], + "angle": 0, + "content": "Figure 2. Hierarchical relations in GCD. (a) Inter-category relationships within the Stanford-Cars dataset [33]. (b) Intra-category relationships within CUB [54] dataset." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.264, + 0.485, + 0.779 + ], + "angle": 0, + "content": "these methods consider object parts as rigid image patches of the same size, without considering the hierarchical nature of the object parts and the scale discrepancy of the same parts in different images, thus unavoidably restricting the performance for GCD (see Fig. 1(a)), in which the objects often have distinct poses, scales and appearance. To address this problem, one possible solution is to learn image embeddings possessing hierarchical constraints or following tree-like structures. This has been proven to be effective in many tasks. For example, in image retrieval and clustering, the hierarchy constraint may arise from whole-fragment relation [29, 31]. Intuitively, in category discovery, which can be regarded as a transfer clustering task [24], we hypothesize that an embedding space that captures the hierarchical relations of object parts can also facilitate the discovery of new categories. Indeed, the hierarchical relations have been studied in GCD, such as [27, 41, 44, 45, 57]. However, they study from a substantially different perspective: inter-category hierarchy (see Fig. 2(a)). This only considers the hierarchy of different semantic classes from coarse to fine levels. Additionally, these methods require the relationships and number of levels in the hierarchy to be predefined, resulting in a lack of flexibility and scalability. Moreover, these methods are unable to capture more complex hierarchies, such as the compositional parts of an object (see Fig. 2(b)). This is particularly the case because existing methods [45, 51, 55, 59], no matter whether they consider the hierarchy or not, learn the image embeddings in a spherical space. This follows the common practice of applying a spherical projection operator at the end of the self-supervised feature backbone [7, 40]. Consequently, all subsequent operations, including distance calculations, are performed under either Euclidean or spherical geometry, resulting in limited awareness of hierarchical object parts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.902 + ], + "angle": 0, + "content": "In this work, we study the overlooked perspective in category discovery: instead of learning in the Euclidean or spherical space, we advocate a space that captures the hierarchical structure of each data point. In spherical space, both the radius and volume are constant, whereas in Euclidean space, the volume grows polynomially with respect to the radius. Both of these spaces have been shown to be suboptimal for encoding samples that possess hierarchical" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.287 + ], + "angle": 0, + "content": "structures [10, 15, 31]. In contrast, hyperbolic space possesses a distinctive property where its volume grows exponentially relative to the radius. This characteristic makes hyperbolic space particularly suitable for embedding tree-like data, enhancing its representational power. Learning representations in hyperbolic space has proven to be effective in various computer vision tasks, including object recognition [15], object detection [32], semantic segmentation [58], and anomaly detection [35]. Inspired by these successes, we aim to realize the idea of learning hierarchy-aware representations to facilitate knowledge transfer in category discovery, thereby unleashing the potential of hyperbolic representations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.289, + 0.905, + 0.68 + ], + "angle": 0, + "content": "To achieve this goal, we propose a simple yet effective framework, HypCD, to properly learn the hierarchy-aware representation and classifier for category discovery through the lens of hyperbolic geometry. In this framework, we adapt our framework to popular parametric [59] and non-parametric [51] GCD baselines as well as the state-of-the-art (SOTA) method SelEx [45], obtaining substantial improvements for them, establishing the new SOTA (see Fig. 1(b)). Firstly, starting from the self-supervised backbone pretrained in Euclidean space, we propose to map the Euclidean representation to a constrained Poincaré ball through feature clipping and exponential mapping. Secondly, we implement the hyperbolic representation learning and build a hyperbolic classifier on the Poincaré ball, considering both angle and distance between samples in hyperbolic space. Thirdly, to assign labels to unlabelled data after training, for non-parametric methods, we apply semi-supervised \\(k\\)-means following the common practice; for parametric methods, we employ a hyperbolic classifier to make predictions. Despite its simplicity, our framework achieves significant performance improvements with two different pretrained weights (DINO [7] and DINOv2 [40]) on the public GCD datasets, including the coarse-grained classification datasets CIFAR-10 [34], CIFAR-100 [34], and ImageNet-100 [13], as well as the fine-grained SSB [50] benchmark." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.681, + 0.906, + 0.846 + ], + "angle": 0, + "content": "In summary, we make the following contributions in this paper: (i) We identify the existing GCD methods' common shortcoming in encoding the hierarchical structure and propose to incorporate the hyperbolic geometry into the embedding space to address this limitation; (ii) We propose a simple yet effective framework, called HypCD, for incorporating the hyperbolic geometry in representation learning and classification for category discovery; and (iii) Through extensive experiments on public GCD benchmarks by applying HypCD to baseline and SOTA methods, our method consistently demonstrates effectiveness and superiority." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.86, + 0.655, + 0.876 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Category Discovery. Novel category discovery (NCD) is" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9892" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.499 + ], + "angle": 0, + "content": "initially introduced in [24] to establish a realistic framework for transferring knowledge from seen categories to cluster unseen categories, by considering it as a transfer clustering problem. Many subsequent methods have been proposed to advance the field [17, 25, 26, 28, 63, 65]. Generalized category discovery [51] (GCD) relaxes the assumption in NCD by considering unlabelled data containing samples from both known and unknown classes [51]. Further investigations [6, 8, 27, 30, 37, 42, 56] have explored a variety of strategies to address the challenges posed by GCD. One notable approach, SimGCD [59], proposes to learn a parametric classifier enhanced by mean entropy regularization, thereby improving performance. In another vein, GPC [64] employs Gaussian mixture models to jointly learn robust representations while simultaneously estimating the number of unknown categories. SPT-Net [55] presents a spatial prompt tuning method that enables models to concentrate more effectively on specific object parts, thus enhancing knowledge transfer in GCD. Most recently, SelEx [45] has been proposed, leveraging hierarchical semi-supervised \\( k \\)-means to achieve SOTA results on fine-grained datasets. Additionally, various efforts are focused on addressing category discovery from multiple perspectives. For instance, [28] emphasizes multi-modal category discovery; [62] and [8] explore a continual setting; [43] studies category discovery in a federated setting; and [56] examines GCD in the presence of domain shifts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.5, + 0.484, + 0.788 + ], + "angle": 0, + "content": "Hyperbolic Geometry. Hyperbolic space, defined as a non-Euclidean manifold with exponential volume growth in relation to its radius, is inherently aligned with the embedding of tree-like and hierarchical data structures in visual recognition tasks. Significant advancements in this area include [31], which presents a hyperbolic image embedding technique by projecting model outputs into hyperbolic space, and [15], which integrates hyperbolic geometry into various vision transformer architectures, showcasing performance that surpasses their Euclidean counterparts. Hyperbolic methods have also been developed on diverse tasks such as image classification [15, 23, 31], action recognition [18], few-shot learning [20] and object segmentation [21, 60]. Moreover, recent developments have introduced hyperbolic geometry for neural networks including fully connected layers [46], convolutional neural networks [3], graph neural networks [9, 36], and attention network [22], thereby facilitating a deeper integration of hyperbolic geometry into deep learning regime." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.8, + 0.182, + 0.815 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In this section, we first introduce the task in Sec. 3.1, then move to a review of baselines in Sec. 3.2. Afterwards, the geometry mapping and training details of our framework are described in Sec. 3.3 and Sec. 3.4. Lastly, the label assignment details are outlined in Sec. 3.5." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.698, + 0.105 + ], + "angle": 0, + "content": "3.1. Problem Statement" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.113, + 0.907, + 0.325 + ], + "angle": 0, + "content": "GCD aims to learn a model capable of accurately classifying unlabelled samples from known categories while simultaneously clustering those from unknown categories. Consider an unlabelled dataset denoted as \\(\\mathbf{D}_u = \\left\\{\\left(\\mathbf{x}_i^u,y_i^u\\right)\\right\\} \\in\\) \\(\\mathbf{X}\\times \\mathbf{Y}_u\\) and a labelled dataset represented as \\(\\mathbf{D}_l =\\) \\(\\left\\{\\left(\\mathbf{x}_i^l,y_i^l\\right)\\right\\} \\in \\mathbf{X}\\times \\mathbf{Y}_l,\\) where \\(\\mathbf{Y}_u\\) and \\(\\mathbf{Y}_l\\) denote the respective label sets. The unlabelled dataset comprises samples from both known and unknown categories, i.e., specifically \\(\\mathbf{Y}_l\\subset \\mathbf{Y}_u\\) . Let the number of labelled categories be denoted by \\(M = |\\mathbf{Y}_l|\\) . We assume that the total number of categories, \\(K = |\\mathbf{Y}_l\\cup \\mathbf{Y}_u|\\) , is known, as posited in prior works [26, 53, 59]. In scenarios where this information is unavailable, alternative methods such as those proposed in [24, 51] can be employed to yield a reliable estimation." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.333, + 0.702, + 0.348 + ], + "angle": 0, + "content": "3.2. Review of Baselines" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.355, + 0.907, + 0.492 + ], + "angle": 0, + "content": "Non-parametric Baseline. [51] formalizes the GCD task and proposes a non-parametric baseline. The approach involves finetuning the pre-trained DINO [7] model [14] to enhance the learned representation. The loss function comprises a supervised contrastive loss, which operates on the labelled samples, and a self-supervised contrastive loss, which operates on all the samples. Specifically, given two randomly augmented views \\(\\mathbf{x}_i\\) and \\(\\mathbf{x}_i'\\) for the same image in a mini-batch \\(B\\), the self-supervised contrastive loss is:" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.501, + 0.907, + 0.528 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r e p} ^ {u} = \\frac {1}{| B |} \\sum_ {i \\in B} - \\log \\frac {\\exp \\left(\\mathbf {z} _ {i} \\cdot \\mathbf {z} _ {i} ^ {\\prime} / \\tau_ {r}\\right)}{\\sum_ {j} ^ {j \\neq i} \\exp \\left(\\mathbf {z} _ {i} \\cdot \\mathbf {z} _ {j} ^ {\\prime} / \\tau_ {r}\\right)}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.536, + 0.906, + 0.627 + ], + "angle": 0, + "content": "where the feature \\(\\mathbf{z}_i = \\rho_r(\\phi (\\mathbf{x}_i))\\) is a \\(\\ell_2\\)-normalized vector and \\(\\mathbf{z}_i^\\prime\\) represents feature from another view \\(\\mathbf{x}_i^\\prime\\). Here, \\(\\phi\\) refers to the backbone network, \\(\\rho_r\\) denotes the projection head, and \\(\\tau_{r}\\) stands as the temperature parameter used for scaling the features. The supervised contrastive loss for labelled samples is:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.636, + 0.905, + 0.666 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r e p} ^ {s} = \\frac {1}{| B _ {l} |} \\sum_ {i \\in B _ {l}} \\frac {1}{| N _ {i} |} \\sum_ {q \\in N _ {i}} - \\log \\frac {\\exp \\left(\\mathbf {z} _ {i} \\cdot \\mathbf {z} _ {q} / \\tau_ {r}\\right)}{\\sum_ {j} ^ {j \\neq i} \\exp \\left(\\mathbf {z} _ {i} \\cdot \\mathbf {z} _ {j} / \\tau_ {r}\\right)}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.674, + 0.907, + 0.735 + ], + "angle": 0, + "content": "where \\(N_{i}\\) is the index set for all other images in the labelled mini-batch \\(B_{l} \\subset B\\) having the same label as \\(\\mathbf{x}_{i}\\). The overall representation learning loss is then: \\(\\mathcal{L}_{rep} = (1 - \\lambda_b)\\mathcal{L}_{rep}^u + \\lambda_b\\mathcal{L}_{rep}^s\\), where \\(\\lambda_{b}\\) is a balance factor." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.735, + 0.906, + 0.872 + ], + "angle": 0, + "content": "Parametric Baseline. [59] introduces a robust parametric GCD baseline, which has been widely adopted in the field ever since [53, 55]. This method employs a parametric classifier implemented in a self-distillation framework [7]. The classifier is randomly initialized with \\(K\\) normalized category prototypes \\(\\mathbf{C} = \\{\\mathbf{c}_1,\\dots,\\mathbf{c}_K\\}\\). For a randomly augmented view \\(\\mathbf{x}_i\\) and its corresponding normalized hidden feature vector \\(\\mathbf{h}_i = \\phi (\\mathbf{x}_i) / ||\\phi (\\mathbf{x}_i)||\\), the output probability for the \\(k\\)-th category is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.88, + 0.905, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\mathbf {p} _ {i} ^ {(k)} = \\frac {\\exp \\left(\\mathbf {h} _ {i} \\cdot \\mathbf {c} _ {k} / \\tau_ {s}\\right)}{\\sum_ {j = 1} ^ {K} \\exp \\left(\\mathbf {h} _ {i} \\cdot \\mathbf {c} _ {j} / \\tau_ {s}\\right)}, \\tag {3}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9893" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.213 + ], + "angle": 0, + "content": "where \\(\\tau_{s}\\) is the scaling temperature for the 'student' view. The soft label \\(\\mathbf{q}_i\\) is generated by the teacher' view with a sharper temperature \\(\\tau_{t}\\) based on another augmented view in a similar manner. The self-distillation loss for the two views is then computed using the cross-entropy loss function: \\(\\ell_{ce}(\\mathbf{q}',\\mathbf{p}) = -\\sum_{j = 1}^{K}\\mathbf{q}'^{(j)}\\log \\mathbf{p}^{(j)}\\). The unsupervised loss is then computed by aggregating contributions from all samples in the mini-batch \\(B\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.155, + 0.224, + 0.483, + 0.246 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c l s} ^ {u} = \\frac {1}{| B |} \\sum_ {i \\in B} \\ell_ {c e} \\left(\\mathbf {q} _ {i} ^ {\\prime}, \\mathbf {p} _ {i}\\right) - \\xi \\mathcal {H} (\\bar {\\mathbf {p}}), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.258, + 0.483, + 0.306 + ], + "angle": 0, + "content": "where \\(\\overline{\\mathbf{p}} = \\frac{1}{2|B|}\\sum_{i\\in B}(\\mathbf{p}_i + \\mathbf{p}_i')\\) denotes the mean prediction across the mini-batch. The mean entropy is defined as: \\(\\mathcal{H}(\\overline{\\mathbf{p}}) = -\\sum_{j = 1}^{K}\\overline{\\mathbf{p}}^{(j)}\\log \\overline{\\mathbf{p}}^{(j)}\\), weighted by \\(\\xi\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.306, + 0.484, + 0.459 + ], + "angle": 0, + "content": "For the labelled samples, the supervised classification loss is written as \\(\\mathcal{L}_{cls}^{s} = \\frac{1}{|B_{l}|}\\sum_{i\\in B_{l}}\\ell_{ce}(\\mathbf{p}_{i},\\mathbf{y}_{i})\\), where \\(\\mathbf{y}_i\\) represents the one-hot vector corresponding to the ground-truth label \\(y_{i}\\). The overall objective is \\(\\mathcal{L}_{cls} = (1 - \\lambda_b)\\mathcal{L}_{cls}^u + \\lambda_b\\mathcal{L}_{cls}^s\\). Integrating this with the representation learning loss \\(\\mathcal{L}_{rep}\\) adopted from [51], the comprehensive training objective is expressed as: \\(\\mathcal{L}_{gcd} = \\mathcal{L}_{cls} + \\mathcal{L}_{rep}\\). Through training with \\(\\mathcal{L}_{gcd}\\) on both \\(\\mathbf{D}_l\\) and \\(\\mathbf{D}_u\\), the classifier is empowered to directly predict labels for the unlabelled samples after the training process concludes." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.469, + 0.445, + 0.485 + ], + "angle": 0, + "content": "3.3. Hyperbolic Space for Category Discovery" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.491, + 0.484, + 0.672 + ], + "angle": 0, + "content": "As previously discussed, object parts are critical for facilitating knowledge transfer from labelled categories to unseen ones in GCD. Each sample inherently contains object parts that reside within a hierarchical structure. Moreover, existing GCD methods [44, 45] emphasize the intercategory hierarchy to enhance the clustering performance of unlabelled samples in Euclidean or spherical spaces. However, the geometry of representation space limits their ability to effectively capture other kinds of hierarchy [31]. In contrast, hyperbolic space, characterized by its property of exponential volume growth with respect to the radius [15], emerges as a more suitable space for GCD." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.673, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Hyperbolic space \\(\\mathbb{H}^n\\) is defined as an \\(n\\)-dimensional Riemannian manifold exhibiting constant negative curvature, and it encompasses several analytic models [5]. Following previous literature [15, 31], we employ the Poincaré ball [39] model. In this model, the hyperbolic space is represented as an \\(n\\)-dimensional ball \\(\\mathbb{D}_c^n = \\{\\mathbf{a}\\in \\mathbb{R}^n\\mid c\\| \\mathbf{a}\\| ^2 < 1\\}\\) with curvature value \\(-c^2\\), where \\(c\\) is the non-negative curvature parameter. The manifold is equipped with the Riemannian metric \\(g^{\\mathbb{D}} = \\lambda_c^2 g^{\\mathbb{E}}\\) where \\(\\lambda_{c}(\\mathbf{a}) = \\frac{2}{1 - c\\|\\mathbf{a}\\|^2}\\) is the conformal factor and \\(g^{\\mathbb{E}}\\) is the identity metric \\(\\mathbf{I}_n\\) in Euclidean space. In this way, the local distances are scaled by the factor \\(\\lambda_{c}\\) approaching infinity near the boundary of the ball. This gives rise to the exponential expansion property of hyperbolic spaces, unlike the polynomial expansion in Euclidean space. However, hyperbolic" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.153 + ], + "angle": 0, + "content": "space is not vector space and thus operations such as addition can not be directly conducted. To address this problem, we leverage the gyrovector formalism [48]. For a pair of points \\(\\mathbf{a},\\mathbf{b}\\in \\mathbb{D}_c^n\\), their Möbius addition is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.162, + 0.907, + 0.188 + ], + "angle": 0, + "content": "\\[\n\\mathbf {a} \\oplus_ {c} \\mathbf {b} = \\frac {(1 + 2 c \\langle \\mathbf {a} , \\mathbf {b} \\rangle + c \\| \\mathbf {b} \\| ^ {2}) \\mathbf {a} + (1 - c \\| \\mathbf {a} \\| ^ {2}) \\mathbf {b}}{1 + 2 c \\langle \\mathbf {a} , \\mathbf {b} \\rangle + c ^ {2} \\| \\mathbf {a} \\| ^ {2} \\| \\mathbf {b} \\| ^ {2}}. \\qquad (5)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.191, + 0.825, + 0.205 + ], + "angle": 0, + "content": "The hyperbolic distance between them is then:" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.216, + 0.907, + 0.237 + ], + "angle": 0, + "content": "\\[\n\\mathcal {D} _ {\\mathbb {H}} (\\mathbf {a}, \\mathbf {b}) = \\frac {2}{\\sqrt {c}} \\operatorname {a r c t a n h} \\left(\\sqrt {c} \\| - \\mathbf {a} \\oplus_ {c} \\mathbf {b} \\|\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.245, + 0.906, + 0.276 + ], + "angle": 0, + "content": "When \\(c\\to 0\\) , the hyperbolic distance (Eq. 6) closes to the Euclidean distance \\(\\lim_{c\\to 0}\\mathcal{D}_{\\mathbb{H}}(\\mathbf{a},\\mathbf{b}) = 2\\| \\mathbf{a} - \\mathbf{b}\\|\\)" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.285, + 0.611, + 0.301 + ], + "angle": 0, + "content": "3.4. HypCD" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.307, + 0.906, + 0.472 + ], + "angle": 0, + "content": "As illustrated in Fig. 3, we propose a unified framework, HypCD, for category discovery in hyperbolic space, incorporating both parametric [59] and non-parametric GCD approaches. Given two randomly augmented views, we initially obtain the respective Euclidean feature vectors \\(\\mathbf{z}_i\\) and \\(\\mathbf{z}_i'\\) through a self-supervised pretrained backbone [7, 40]. Subsequently, the feature embeddings are mapped into hyperbolic space \\(\\mathbb{H}^n\\) using exponential mapping, facilitating representation learning within this exponentially growing space to more effectively capture and utilize the hierarchical relationships inherent in the training data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.473, + 0.906, + 0.533 + ], + "angle": 0, + "content": "The exponential mapping [31] serves as a bijective projection between Euclidean space \\(\\mathbb{E}^n\\) and hyperbolic space \\(\\mathbb{H}^n\\). The projection of tangent vector \\(\\mathbf{z}\\) from \\(\\mathbb{E}^n\\) to \\(\\mathbb{H}^n\\) is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.537, + 0.54, + 0.907, + 0.575 + ], + "angle": 0, + "content": "\\[\n\\exp_ {\\mathbf {o}} ^ {c} (\\mathbf {z}) = \\mathbf {o} \\oplus_ {c} \\left(\\tanh \\left(\\sqrt {c} \\frac {\\lambda_ {\\mathbf {o}} ^ {c} \\| \\mathbf {z} \\|}{2}\\right) \\frac {\\mathbf {z}}{\\sqrt {c} \\| \\mathbf {z} \\|}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.906, + 0.734 + ], + "angle": 0, + "content": "where \\(\\oplus_{c}\\) is the Mobius addition, as introduced in Eq. 5 and \\(\\mathbf{o}\\) represents the base point of the mapping. To address the issue of gradient vanishing [23] near the boundary of the Poincaré ball during training, we implement a feature clipping operation in \\(\\mathbb{E}^n\\) prior to the exponential mapping. The operation is defined as: \\(\\mathcal{C}(\\mathbf{z}) = \\min \\{1,\\frac{r}{||\\mathbf{z}||}\\} \\cdot \\mathbf{z}\\), where \\(r\\) denotes the clipping value. For the feature vector \\(\\mathbf{z}_i\\) in \\(\\mathbb{E}^n\\), the corresponding mapped feature in \\(\\mathbb{H}^n\\) is expressed as \\(\\mathcal{M}(\\mathbf{z}_i) = \\exp_{\\mathbf{o}}^{c}(\\mathcal{C}(\\mathbf{z}_{i}))\\). The same operation will also be applied to the other feature vector \\(\\mathbf{z}_i'\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.906, + 0.84 + ], + "angle": 0, + "content": "As described in Sec.3.2, both parametric [51] and nonparametric [59] baselines utilize the same representation learning method. In our framework, we implement a consistent representation learning solution in hyperbolic space for them (Fig.3(a)). For parametric approaches, a hyperbolic parametric classifier is employed (Fig. 3(b)). We will introduce these components in detail subsequently." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Hyperbolic Representation Learning. Following prior attempts [45, 51, 59], we incorporate both self-supervised and supervised contrastive learning into our framework. However, our approach uniquely operates within hyperbolic" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9894" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.093, + 0.597, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.094, + 0.895, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.219, + 0.596, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.601, + 0.219, + 0.896, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.345, + 0.905, + 0.375 + ], + "angle": 0, + "content": "Figure 3. Overall pipeline of our HypCD framework for parametric and non-parametric GCD baselines. (a) Hyperbolic representation learning. (b) Hyperbolic classifier. (c) Non-parametric label assignment. (d) Parametric label assignment." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.385, + 0.484, + 0.506 + ], + "angle": 0, + "content": "space. Furthermore, unlike previous GCD methods that exclusively utilize cosine distance [51, 55, 59] (angle-based) or Euclidean distance [45] (distance-based) for calculating pairwise similarity, we propose a hybrid approach that combines both distance-based and angle-based losses. Such integration has been shown to be more effective for model optimization in hyperbolic space [18]. First, the unified form of self-supervised contrastive loss can be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.116, + 0.514, + 0.483, + 0.543 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {u} = \\frac {1}{| B |} \\sum_ {i \\in B} - \\log \\frac {\\exp \\left(\\mathcal {S} \\left(\\mathcal {M} \\left(\\mathbf {z} _ {i}\\right) , \\mathcal {M} \\left(\\mathbf {z} _ {i} ^ {\\prime}\\right)\\right) / \\tau_ {r}\\right)}{\\sum_ {j} ^ {j \\neq i} \\exp \\left(\\mathcal {S} \\left(\\mathcal {M} \\left(\\mathbf {z} _ {i}\\right) , \\mathcal {M} \\left(\\mathbf {z} _ {j} ^ {\\prime}\\right)\\right) / \\tau_ {r}\\right)}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.552, + 0.45, + 0.567 + ], + "angle": 0, + "content": "Similarly, the supervised contrastive loss is unified as:" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.575, + 0.483, + 0.605 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {s} = \\frac {1}{| B _ {l} |} \\sum_ {i \\in B _ {l}} \\frac {1}{| N _ {i} |} \\sum_ {q \\in N _ {i}} - \\log \\frac {\\exp \\left(\\mathcal {S} \\left(\\mathcal {M} \\left(\\mathbf {z} _ {i}\\right) , \\mathcal {M} \\left(\\mathbf {z} _ {q}\\right)\\right) / \\tau_ {r}\\right)}{\\sum_ {j \\neq i} \\exp \\left(\\mathcal {S} \\left(\\mathcal {M} \\left(\\mathbf {z} _ {i}\\right) , \\mathcal {M} \\left(\\mathbf {z} _ {j}\\right)\\right) / \\tau_ {r}\\right)}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.613, + 0.484, + 0.704 + ], + "angle": 0, + "content": "where \\(S\\) denotes the similarity function, which can be either distance-based or angle-based. For distance-based contrastive loss, we utilize \\(S_{d} = -\\mathcal{D}_{\\mathbb{H}}\\) as the similarity function, which is formally computed using negative Euclidean distance in prior methods [45]. For angle-based contrastive loss, we employ the cosine similarity, formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.712, + 0.482, + 0.737 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} _ {a} \\left(\\mathcal {M} \\left(\\mathbf {z} _ {i}\\right), \\mathcal {M} \\left(\\mathbf {z} _ {i} ^ {\\prime}\\right)\\right) = \\frac {\\mathcal {M} \\left(\\mathbf {z} _ {i}\\right) \\cdot \\mathcal {M} \\left(\\mathbf {z} _ {i} ^ {\\prime}\\right)}{\\left| \\left| \\mathcal {M} \\left(\\mathbf {z} _ {i}\\right) \\right| \\right| \\cdot \\left| \\left| \\mathcal {M} \\left(\\mathbf {z} _ {i} ^ {\\prime}\\right) \\right| \\right|}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.744, + 0.483, + 0.773 + ], + "angle": 0, + "content": "Since hyperbolic space is conformal with Euclidean space, cosine similarity remains equivalent in both \\(\\mathbb{E}^n\\) and \\(\\mathbb{H}^n\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.774, + 0.484, + 0.817 + ], + "angle": 0, + "content": "The final supervised and self-supervised hyperbolic contrastive loss is composed of both distance-based and angle-based losses:" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.825, + 0.482, + 0.85 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {h r e p} ^ {s} = \\alpha_ {d} \\mathcal {L} _ {d i s} ^ {s} + (1 - \\alpha_ {d}) \\mathcal {L} _ {a n g} ^ {s}, \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.847, + 0.403, + 0.864 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {h r e p} ^ {u} = \\alpha_ {d} \\mathcal {L} _ {d i s} ^ {u} + (1 - \\alpha_ {d}) \\mathcal {L} _ {a n g} ^ {u},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{hrep}^s\\) and \\(\\mathcal{L}_{hrep}^u\\) represent the supervised and self-supervised hyperbolic contrastive loss, respectively. The" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.385, + 0.906, + 0.462 + ], + "angle": 0, + "content": "terms \\(\\mathcal{L}_{dis}\\) and \\(\\mathcal{L}_{ang}\\) correspond to distance-based and angle-based contrastive loss, respectively, obtained by substituting \\(S\\) with \\(S_{d}\\) and \\(S_{a}\\). Additionally, \\(\\alpha_{d}\\) is the loss weight of distance-based loss. The overall training objective for hyperbolic representation learning is:" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.468, + 0.905, + 0.489 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r e p} ^ {\\mathbb {H}} = \\left(1 - \\lambda_ {b} ^ {\\mathbb {H}}\\right) \\mathcal {L} _ {h r e p} ^ {u} + \\lambda_ {b} ^ {\\mathbb {H}} \\mathcal {L} _ {h r e p} ^ {s}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.496, + 0.905, + 0.527 + ], + "angle": 0, + "content": "where \\(\\lambda_b^{\\mathbb{H}}\\) serves as the balancing factor between the supervised and unsupervised losses." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.528, + 0.907, + 0.786 + ], + "angle": 0, + "content": "Hyperbolic Classifier. To enhance the parametric baseline with hyperbolic geometry, we replace the conventional Euclidean classification head—traditionally reliant on a multilayer perceptron (MLP) in Euclidean space—with its hyperbolic counterpart, the hyperbolic feed forward network (HypFFN). The hyperbolic linear layer [19] exhibits greater alignment with the baseline [59], and we experimentally find that it outperforms the hyperbolic multinomial logistic regression layer. Consider the last linear layer of the MLP; similar to its Euclidean counterpart, the hyperbolic linear layer is parameterized by a weight matrix \\(\\mathbf{w} \\in \\mathbb{R}^{I \\times K}\\) and a bias vector \\(\\mathbf{s} \\in \\mathbb{R}^{1 \\times K}\\), where \\(I\\) denotes the input feature dimension. Given the hyperbolic feature \\(\\mathbf{z}_i^{\\mathbb{H}} = \\mathcal{M}(\\mathbf{z}_i) \\in \\mathbb{R}^{1 \\times I}\\), the linear layer operates as HypLinear \\((\\mathbf{z}_i^{\\mathbb{H}}, \\mathbf{w}, \\mathbf{s}) = \\operatorname{Proj}[(\\mathbf{w} \\otimes_c \\mathbf{z}_i^{\\mathbb{H}}) \\oplus_c \\mathbf{s}]\\), where \\(\\oplus_c\\) follows Eq. 5. The Möbius matrix-vector multiplication \\(\\mathbf{v}_i = \\mathbf{w} \\otimes_c \\mathbf{z}_i^{\\mathbb{H}}\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.792, + 0.905, + 0.819 + ], + "angle": 0, + "content": "\\[\n\\frac {1}{\\sqrt {c}} \\tanh \\left(\\frac {\\| \\mathbf {z} _ {i} ^ {\\mathbb {H}} \\mathbf {w} \\| _ {2}}{\\| \\mathbf {z} _ {i} ^ {\\mathbb {H}} \\| _ {2}} \\tanh ^ {- 1} (\\sqrt {c} \\| \\mathbf {z} _ {i} ^ {\\mathbb {H}} \\| _ {2})\\right) \\frac {\\mathbf {z} _ {i} ^ {\\mathbb {H}} \\mathbf {w}}{\\| \\mathbf {z} _ {i} ^ {\\mathbb {H}} \\mathbf {w} \\| _ {2}}. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.905, + 0.856 + ], + "angle": 0, + "content": "To ensure numerical stability [19], a safe projection is operated on the result manifold and represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.863, + 0.905, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\operatorname {P r o j} (\\mathbf {v} _ {i}) = \\left\\{ \\begin{array}{l l} \\frac {\\mathbf {v} _ {i}}{\\| \\mathbf {v} _ {i} \\| _ {2}} \\times \\frac {1 - 1 0 ^ {- 3}}{\\sqrt {c}}, & \\frac {1 - 1 0 ^ {- 3}}{\\sqrt {c}} < \\| \\mathbf {v} _ {i} \\| _ {2} \\\\ \\mathbf {v} _ {i}, & \\text {o t h e r w i s e} \\end{array} . \\right. \\tag {14}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9895" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.274 + ], + "angle": 0, + "content": "This integration allows our hyperbolic classifier to be seamlessly incorporated into the baseline [59] by substituting the original MLP with HypFFN. For each point in \\(\\mathbb{H}^n\\), the tangent space at that point serves as a Euclidean subspace, enabling straightforward adaptation of Euclidean operations within this space [10]. Consequently, the cross-entropy loss for the hyperbolic classifier can be expressed as: \\(\\ell_{ce}^{\\mathbb{H}} = \\ell_{ce}(\\mathrm{HypFFN}(\\mathbf{z}_i^{\\mathbb{H}}),\\mathbf{y}_i)\\). Additionally, we can define the hyperbolic counterpart \\(\\mathcal{H}^{\\mathbb{H}}\\) for the mean entropy \\(\\mathcal{H}\\). By substituting the original \\(\\ell_{ce}\\) and \\(\\mathcal{H}\\) with our derived \\(\\ell_{ce}^{\\mathbb{H}}\\) and \\(\\mathcal{H}^{\\mathbb{H}}\\), respectively, we can readily compute the final hyperbolic classifier loss \\(\\mathcal{L}_{cls}^{\\mathbb{H}}\\), as detailed in Sec. 3.2." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.285, + 0.267, + 0.3 + ], + "angle": 0, + "content": "3.5.Label Assignment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.307, + 0.485, + 0.626 + ], + "angle": 0, + "content": "Existing approaches typically employ either a parametric classification head or non-parametric methods, such as semi-supervised \\( k \\)-means [51], for label assignment. In this paper, we do not independently assess these two methods; rather, we integrate both within the HypCD framework as shown in Fig. 3(c) and (d). For non-parametric approaches, including [51] and the recent SelEx [45], we retain the original label assignment strategy by applying semi-supervised \\( k \\)-means clustering directly to feature representations extracted by \\( \\phi \\) in \\( \\mathbb{E}^n \\). Our empirical results indicate that training in hyperbolic space allows for the transfer of hierarchical structure encoding from \\( \\mathbb{H}^n \\) to \\( \\mathbb{E}^n \\). Moreover, we find that the operations of \\( k \\)-means in \\( \\mathbb{E}^n \\) are significantly more efficient while maintaining comparable performance. For the parametric baseline exemplified by SimGCD [59], we utilize the hyperbolic classification head to conduct classification within hyperbolic space using the trained hyperbolic classifier HypFFN. Both design choices are theoretically supported by the property of hyperbolic geometry of encoding hierarchical structures, facilitating a more intuitive and effective representation and classifier for GCD." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.641, + 0.216, + 0.657 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.666, + 0.347, + 0.683 + ], + "angle": 0, + "content": "4.1. Setups and Implementations" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Datasets. We thoroughly evaluate our method across diverse benchmarks, including the generic image recognition datasets CIFAR-10 and CIFAR-100 [34], as well as ImageNet-100 [13]. Additionally, we assess our approach on the Semantic Shift Benchmark (SSB)[52], which includes fine-grained datasets such as CUB[54], StanfordCars [33], and FGVC-Aircraft [38]. For each dataset, we adhere to the data split scheme detailed in [51]. The method involves sampling a subset of all classes as the known ('Old') classes \\(\\mathbf{Y}_l\\). Subsequently, \\(50\\%\\) of the images from these known classes are utilized to construct \\(\\mathbf{D}_l\\), while the remaining images are designated as the unlabelled data \\(\\mathbf{D}_u\\). Evaluation Metrics. We evaluate the performance using the clustering accuracy (ACC) as defined in the litera" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.185 + ], + "angle": 0, + "content": "ture [51]. The \\(ACC\\) on \\(\\mathbf{D}_u\\) is computed as follows, given the ground truth \\(y_{i}\\) and the predicted labels \\(\\hat{y}_i\\): \\(ACC = \\frac{1}{|\\mathbf{D}_u|}\\sum_{i = 1}^{|\\mathbf{D}_u|}\\mathbb{1}(y_i = h(\\hat{y}_i))\\), where \\(h\\) denotes the optimal permutation that aligns the predicted cluster assignments with the ground-truth labels. The \\(ACC\\) values for the 'All', 'Old' and 'New' classes are reported separately." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.186, + 0.907, + 0.458 + ], + "angle": 0, + "content": "Implementation Details. We evaluate HypCD against the non-parametric baseline GCD [51], the parametric baseline SimGCD [59], and the SOTA method SelEx [45], utilizing both DINO [7] and DINoV2 [40] pretrained weights. Detailed information regarding SelEx can be found in the supplementary materials. For GCD [51], the output dimension of the projection head \\(\\rho_{r}\\) is 256. In the case of SimGCD [59], the feature dimension from backbone \\(\\phi\\) is 768. \\(\\rho_{r}\\) and the final block of \\(\\phi\\) are optimized using the SGD optimizer, with an initial learning rate of 0.1, which is decayed to 0.001 over time according to a cosine annealing schedule. The HypFFN is optimized using the Riemannian Adam optimizer [4], with a constant learning rate of 0.01. All models are trained for 200 epochs using a batch size of 128. The curvature parameter \\(c\\) is set to 0.05 for the fine-grained datasets and 0.01 for the generic datasets. Following baselines, the balancing factor \\(\\lambda_b^{\\mathbb{H}}\\) is set to 0.35. By default, the loss weight \\(\\alpha_{d}\\) increases linearly from 0 to 1.0." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.468, + 0.747, + 0.484 + ], + "angle": 0, + "content": "4.2. Quantitative Comparison" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.489, + 0.907, + 0.627 + ], + "angle": 0, + "content": "We compare our method with recent GCD methods (including ORCA [6], GCD [51], XCon [16], OpenCon [47], PromptCAL [61], DCCL [42], GPC [64], CiPR [27], SimGCD [59], \\(\\mu\\)GCD [53], InfoSieve [44], SPTNet [55], CMS [12], AMEND [2] and SelEx [45]) and report the results in Tab. 1. The evaluation encompasses performance on the SSB benchmark [52] and generic datasets [13, 34]. The hyperbolic methods applying our HypCD framework are indicated by the 'Hyp-' prefix." + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.645, + 0.707, + 0.786 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.729, + 0.646, + 0.902, + 0.786 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.787, + 0.907, + 0.829 + ], + "angle": 0, + "content": "Figure 4. Comparison of baseline and hyperbolic counterparts on the SSB. Left: 'All' ACC (higher is better). Right: Discrepancy between 'Old' and 'New' ACC (smaller is better)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Results on SSB. The performance of the GCD methods on the SSB benchmark, utilizing both DINO [7] and DINOv2 [40] pretrained weights, is summarized in the left section of Tab. 1. Besides, we provide a comparative analy-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9896" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.083, + 0.905, + 0.111 + ], + "angle": 0, + "content": "Table 1. Comparison of GCD methods on the SSB [52] benchmark, CIFAR-10 [34], CIFAR-100 [34] and ImageNet-100 [13] datasets. Results are reported in ACC across the 'All', 'Old' and 'New' categories." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.111, + 0.907, + 0.505 + ], + "angle": 0, + "content": "
MethodCUB [54]Stanford-Cars [33]FGVC-Aircraft [38]CIFAR-10 [34]CIFAR-100 [34]ImageNet-100 [13]
AllOldNewAllOldNewAllOldNewAllOldNewAllOldNewAllOldNew
DINOORCA [6]36.343.832.631.632.031.431.942.226.969.077.452.073.592.663.981.886.279.6
XCon [16]52.154.351.040.558.831.747.744.449.496.097.395.474.281.260.377.693.569.7
OpenCon [47]54.763.854.749.178.632.7---------84.093.881.2
PromptCAL [61]62.964.462.150.270.140.652.252.252.397.996.698.581.284.275.383.192.778.3
DCCL [42]63.560.864.943.155.736.2---96.396.596.975.376.870.280.590.576.2
GPC [64]52.055.547.538.258.927.443.340.744.890.697.687.075.484.660.175.393.466.7
PIM [11]62.775.756.243.166.931.6---94.797.493.378.384.266.583.195.377.0
μGCD [53]65.768.064.656.568.150.953.855.453.0---------
InfoSieve [44]69.477.965.255.774.846.456.363.752.594.897.793.478.382.270.580.593.873.8
CiPR [27]57.158.755.647.061.540.1---97.797.597.781.582.479.780.584.978.3
SPTNet [55]65.868.865.159.079.249.359.361.858.197.395.098.681.384.375.685.493.281.4
CMS [12]68.276.564.056.976.147.656.063.452.3---82.385.775.584.795.679.2
AMEND [2]64.975.659.652.861.848.356.473.348.296.894.697.881.079.983.383.292.978.3
GCD [51]51.356.648.739.057.629.945.041.146.991.597.988.273.076.266.574.189.866.3
Hyp-GCD61.067.058.050.860.945.848.243.650.592.997.590.674.080.062.080.492.574.4
SimGCD [59]60.365.657.753.871.945.054.259.151.897.195.198.180.181.277.883.093.177.9
Hyp-SimGCD64.865.864.262.873.457.758.758.958.596.895.997.282.483.181.286.593.783.0
SelEx [45]73.675.372.858.575.650.357.164.753.395.998.194.882.385.376.383.193.677.8
Hyp-SelEx79.875.881.862.980.054.765.967.365.196.797.696.382.485.177.086.894.682.8
DINOv2μGCD [53]74.075.973.176.191.068.966.368.765.1---------
CiPR [27]78.373.480.866.777.061.8---99.098.799.290.389.093.188.287.688.5
SPTNet [55]76.379.574.6------------90.196.187.1
GCD [51]71.971.272.365.767.864.755.447.959.297.899.097.179.684.569.978.589.573.0
Hyp-GCD75.675.175.972.880.469.162.770.059.097.598.996.884.587.578.582.992.478.2
SimGCD [59]71.578.168.371.581.966.663.969.960.998.796.799.788.589.287.289.995.587.1
Hyp-SimGCD77.677.977.482.585.881.076.470.379.498.997.799.591.590.094.691.996.289.8
SelEx [45]87.485.188.582.293.776.779.882.378.698.5*98.8*98.5*87.7*90.8*81.5*90.9*96.2*88.3*
Hyp-SelEx90.785.393.483.893.379.283.482.084.198.698.198.988.691.582.892.396.490.2
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.094, + 0.506, + 0.274, + 0.516 + ], + "angle": 0, + "content": "\\*results from our implementation." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.527, + 0.482, + 0.873 + ], + "angle": 0, + "content": "sis between the three baseline methods and their hyperbolic counterparts with DINO backbone in Fig. 4. Our hyperbolic methods consistently outperform their Euclidean counterparts, with particularly strong results observed when utilizing the DINOv2 backbone. Among the evaluated methods, Hyp-SelEx achieves the highest average accuracy (ACC) across all datasets, notably excelling on the CUB dataset, where it records an accuracy of \\(79.8\\%\\) for the 'All' classes with DINO and \\(90.7\\%\\) with DINOv2, establishing it as the leading approach. More strikingly, on the Stanford-Cars dataset, our Hyp-GCD method outperforms the baseline by \\(11.8\\%\\), \\(13.3\\%\\) and \\(15.9\\%\\) in terms of ACC for the 'All', 'Old' and 'New' categories, respectively. Fig. 4 (left) illustrates the average ACC on 'All' categories across the three datasets in the SSB benchmark, indicating that our hyperbolic methods surpass the baseline by a margin of at least \\(6.0\\%\\). Furthermore, as shown in Fig. 4 (right), hyperbolic methods exhibit a consistently smaller ACC gap between 'Old' and 'New' classes, highlighting the effectiveness of HypCD in enhancing knowledge transfer from known to unseen categories. Additionally, DINOv2 outperforms DINO across all methods, underscoring its ability to capture complex data representations more effectively." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.886, + 0.482, + 0.901 + ], + "angle": 0, + "content": "Results on Generic Datasets. In the right section of Tab. 1," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.527, + 0.905, + 0.783 + ], + "angle": 0, + "content": "we present the results on three widely used generic datasets: CIFAR-10 [34], CIFAR-100 [34], and ImageNet-100 [13]. Our methods demonstrate consistent improvements across all cases, regardless of the backbone employed. Notably, these enhancements are especially significant on CIFAR-100 and ImageNet-100, which present greater challenges compared to CIFAR-10, where performance is nearly saturated. For CIFAR-100, Hyp-SimGCD and Hyp-SelEx achieve the highest accuracy of \\(82.4\\%\\) for the 'All' categories using DINO, while Hyp-SimGCD ranks first with an accuracy of \\(91.5\\%\\) on this metric when utilizing DINOv2, significantly surpassing baseline methods and the previous SOTA. Results on ImageNet-100 further validate the effectiveness of hyperbolic methods; Hyp-SelEx achieves the highest performance across 'All', 'Old', and 'New' categories with both DINO and DINOv2, outperforming the baseline by a margin of up to \\(3.7\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.801, + 0.766, + 0.817 + ], + "angle": 0, + "content": "4.3. Impact of Hyperparameters" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Manifold Curvature. Building on previous studies [1, 31] that explore the application of hyperbolic geometry across various tasks, the curvature parameter \\( c \\) (as discussed in Sec. 3.3) is a crucial factor influencing performance and may yield different optimal values across datasets and meth" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9897" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.089, + 0.486, + 0.131 + ], + "angle": 0, + "content": "Table 2. Experimental results using different \\( c \\), \\( r \\) and \\( \\alpha_{d}^{\\max} \\) values in Hyp-SimGCD with DINO [7] pre-trained backbone. Results on the CUB [54] and CIFAR-100 [34] datasets are reported." + }, + { + "type": "table", + "bbox": [ + 0.102, + 0.131, + 0.465, + 0.294 + ], + "angle": 0, + "content": "
parameterStanford-Cars [33]CIFAR-100 [34]
AllOldNewAllOldNew
c = 0.0161.474.455.182.483.181.2
c = 0.0562.873.457.781.684.076.7
c = 0.162.375.156.181.182.378.8
r = 1.060.072.953.782.483.181.2
r = 1.561.275.754.281.282.478.8
r = 2.362.873.457.780.181.178.3
αdmax= 0.159.677.551.081.382.479.1
αdmax= 0.562.077.254.682.483.181.2
αdmax= 1.062.873.457.778.983.569.7
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.308, + 0.486, + 0.504 + ], + "angle": 0, + "content": "ods. Intuitively, as the value of \\( c \\) approaches 0, the radius tends toward infinity, causing the Poincaré ball to flatten and resemble Euclidean space; conversely, larger values of \\( c \\) correspond to a steeper configuration. The widely accepted range for \\( c \\) is between 0.01 and 0.3 [15], with larger values exceeding this range resulting in performance degradation. In our experiments, we evaluate different curvature values of 0.01, 0.05, and 0.1 using Hyp-SimGCD, as presented in Tab. 2. Our findings indicate that the optimal curvature values differ between generic and fine-grained datasets. For fine-grained datasets such as CUB [54], the optimal value is 0.05, while for generic datasets like CIFAR-100, a value of 0.01 proves to be more effective." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.506, + 0.485, + 0.688 + ], + "angle": 0, + "content": "Clipping Value. As articulated in [23], feature clipping has emerged as a standard technique for training hyperbolic neural networks. In our framework, we also observe that it plays a crucial role in category discovery performance. In line with the methodology outlined in [23], we investigate a range of clipping values, specifically 1.0, 1.5, and 2.3. The results shown in the second row of Tab. 2 demonstrate that optimal clipping values vary between fine-grained and generic datasets. For fine-grained datasets like CUB [54], the optimal clipping value is determined to be 2.3. Conversely, for generic datasets like CIFAR-100 [34], a clipping value of 1.0 is shown to be more effective." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Loss Weight. As detailed in Sec. 3.4, we implement a hybrid contrastive loss that combines both distance-based and angle-based components, which is essential for effective optimization in hyperbolic space. A loss weight, denoted as \\(\\alpha_{d}\\), is introduced to regulate the balance between these two types of losses and linearly increasing from 0. In the initial stages, the model prioritizes optimizing the angle between sample points and progressively shifts focus toward optimizing the hyperbolic distance. Consistent with the observations for curvature and clipping values, the optimal max value \\(\\alpha_{d}^{\\max}\\), varies considerably between coarse-grained and fine-grained datasets. For fine-grained datasets such as CUB [54], an optimal value of 1.0 is observed, whereas for more generic datasets like CIFAR-100 [34], a" + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.093, + 0.905, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.266, + 0.907, + 0.308 + ], + "angle": 0, + "content": "Figure 5. T-SNE [49] comparison between SimGCD [59] and our Hyp-SimGCD using 40 randomly sampled instances from 10 randomly selected categories of the Stanford-Cars dataset [33]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.318, + 0.84, + 0.334 + ], + "angle": 0, + "content": "value of 0.5 is found to yield better performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.342, + 0.736, + 0.359 + ], + "angle": 0, + "content": "4.4. Qualitative Comparison" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.364, + 0.908, + 0.622 + ], + "angle": 0, + "content": "In Fig. 5, we present a t-SNE [49] visualization of features extracted from the backbone, represented as \\(\\mathbf{z}_i = \\phi(\\mathbf{x}_i)\\). This visualization compares SimGCD with our Hyp-SimGCD. On the left side of the figure, the clusters generated by SimGCD appear dispersed. Data points from Class 42, highlighted in pink, are spread across multiple areas, indicating significant overlap and a lack of compactness. In contrast, Hyp-SimGCD creates more distinct and tightly clustered groups, concentrating the data points of Class 42 in a more confined area. This comparison implies that Hyp-SimGCD enhances both intra-class compactness and inter-class separation through our hyperbolic representation and classifier learning method. Importantly, even within the original Euclidean space of the backbone network, Hyp-SimGCD exhibits robust clustering performance, which arises from the properties of hyperbolic space in encoding hierarchical structures." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.634, + 0.634, + 0.65 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.66, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In this paper, we investigate a previously overlooked perspective in GCD by utilizing a representation space that captures the hierarchical structure of each sample, instead of the conventional Euclidean or spherical spaces. Our approach leverages the distinctive properties of hyperbolic space, where the volume increases exponentially with radius. This characteristic makes hyperbolic space especially suitable for modelling data possessing hierarchical structures, thereby enhancing representational capacity for category discovery. We propose a simple yet effective framework, HypCD, for integrating hyperbolic geometry into GCD methods. Through extensive experiments with parametric and non-parametric GCD baselines and the SOTA method, our framework consistently demonstrates superior performance on public benchmarks, underscoring the effectiveness of hyperbolic space for category discovery." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9898" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.182 + ], + "angle": 0, + "content": "Acknowledgements. This work is supported by National Natural Science Foundation of China (Grant No. 62306251), Hong Kong Research Grant Council - Early Career Scheme (Grant No. 27208022) and General Research Fund (Grant No. 17211024), and HKU Seed Fund for Basic Research." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.194, + 0.188, + 0.21 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.219, + 0.484, + 0.26 + ], + "angle": 0, + "content": "[1] Mina Ghadimi Atigh, Julian Schoep, Erman Acar, Nanne Van Noord, and Pascal Mettes. Hyperbolic image segmentation. In CVPR, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.261, + 0.484, + 0.316 + ], + "angle": 0, + "content": "[2] Anwesha Banerjee, Liyana Sahir Kallooriyakath, and Soma Biswas. Amend: Adaptive margin and expanded neighborhood for efficient generalized category discovery. In WACV, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.317, + 0.484, + 0.357 + ], + "angle": 0, + "content": "[3] Ahmad Bdeir, Kristian Schwethelm, and Niels Landwehr. Fully hyperbolic convolutional neural networks for computer vision. In ICLR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.359, + 0.483, + 0.385 + ], + "angle": 0, + "content": "[4] Gary Becigneul and Octavian-Eugen Ganea. Riemannian adaptive optimization methods. In ICLR, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.386, + 0.482, + 0.427 + ], + "angle": 0, + "content": "[5] James W Cannon, William J Floyd, Richard Kenyon, Walter R Parry, et al. Hyperbolic geometry. Flavors of geometry, 31(59-115):2, 1997. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.428, + 0.482, + 0.455 + ], + "angle": 0, + "content": "[6] Kaidi Cao, Maria Brbic, and Jure Leskovec. Open-world semi-supervised learning. In ICLR, 2022. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.457, + 0.482, + 0.511 + ], + "angle": 0, + "content": "[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 1, 2, 3, 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.512, + 0.482, + 0.553 + ], + "angle": 0, + "content": "[8] Fernando Julio Cendra, Bingchen Zhao, and Kai Han. Promptcd: Learning gaussian mixture prompt pool for continual category discovery. In ECCV, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.554, + 0.482, + 0.594 + ], + "angle": 0, + "content": "[9] Ines Chami, Rex Ying, Christopher Ré, and Jure Leskovec. Hyperbolic graph convolutional neural networks. In NeurIPS, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.596, + 0.482, + 0.636 + ], + "angle": 0, + "content": "[10] Weize Chen, Xu Han, Yankai Lin, Hexu Zhao, Zhiyuan Liu, Peng Li, Maosong Sun, and Jie Zhou. Fully hyperbolic neural networks. In ACL, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.638, + 0.482, + 0.69 + ], + "angle": 0, + "content": "[11] Florent Chiaroni, Jose Dolz, Ziko Imtiaz Masud, Amar Mitiche, and Ismail Ben Ayed. Parametric information maximization for generalized category discovery. In ICCV, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.692, + 0.482, + 0.734 + ], + "angle": 0, + "content": "[12] Sua Choi, Dahiun Kang, and Minsu Cho. Contrastive mean-shift learning for generalized category discovery. In CVPR, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.735, + 0.482, + 0.775 + ], + "angle": 0, + "content": "[13] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.777, + 0.482, + 0.845 + ], + "angle": 0, + "content": "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[15] Aleksandr Ermolov, Leyla Mirvakhabova, Valentin Khrulkov, Nicu Sebe, and Ivan Oseledets. Hyperbolic vision transformers: Combining improvements in metric learning. In CVPR, 2022. 2, 3, 4, 8" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.219, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[16] Yixin Fei, Zhongkai Zhao, Siwei Yang, and Bingchen Zhao. Xcon: Learning with experts for fine-grained category discovery. In BMVC, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[17] Enrico Fini, Enver Sangineto, Stéphane Lathuiliere, Zhun Zhong, Moin Nabi, and Elisa Ricci. A unified objective for novel class discovery. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.178, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[18] Luca Franco, Paolo Mandica, Bharti Munjal, and Fabio Galasso. Hyperbolic self-paced learning for self-supervised skeleton-based action representations. In ICLR, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.219, + 0.904, + 0.246 + ], + "angle": 0, + "content": "[19] Octavian Ganea, Gary Bécigneul, and Thomas Hofmann. Hyperbolic neural networks. In NeurIPS, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.247, + 0.905, + 0.286 + ], + "angle": 0, + "content": "[20] Zhi Gao, Yuwei Wu, Yunde Jia, and Mehrtash Harandi. Curvature generation in curved spaces for few-shot learning. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.288, + 0.905, + 0.328 + ], + "angle": 0, + "content": "[21] Mina GhadimiAtigh, Julian Schoep, Erman Acar, Nanne van Noord, and Pascal Mettes. Hyperbolic image segmentation. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.905, + 0.397 + ], + "angle": 0, + "content": "[22] Caglar Gulcehre, Misha Denil, Mateusz Malinowski, Ali Razavi, Razvan Pascanu, Karl Moritz Hermann, Peter Battaglia, Victor Bapst, David Raposo, Adam Santoro, and Nando de Freitas. Hyperbolic attention networks. In ICLR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[23] Yunhui Guo, Xudong Wang, Yubei Chen, and Stella X. Yu. Clipped hyperbolic classifiers are super-hyperbolic classifiers. In CVPR, 2022. 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[24] Kai Han, Andrea Vedaldi, and Andrew Zisserman. Learning to discover novel visual categories via deep transfer clustering. In ICCV, 2019. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.905, + 0.537 + ], + "angle": 0, + "content": "[25] Kai Han, Sylvestre-Alvise Rebuffi, Sebastien Ehrhardt, Andrea Vedaldi, and Andrew Zisserman. Automatically discovering and learning new visual categories with ranking statistics. In ICLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.905, + 0.593 + ], + "angle": 0, + "content": "[26] Kai Han, Sylvestre-Alvise Rebuffi, Sebastien Ehrhardt, Andrea Vedaldi, and Andrew Zisserman. Autonovel: Automatically discovering and learning novel visual categories. IEEE TPAMI, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.595, + 0.905, + 0.636 + ], + "angle": 0, + "content": "[27] Shaozhe Hao, Kai Han, and Kwan-Yee K Wong. Cipr: An efficient framework with cross-instance positive relations for generalized category discovery. TMLR, 2024. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.637, + 0.905, + 0.677 + ], + "angle": 0, + "content": "[28] Xuhui Jia, Kai Han, Yukun Zhu, and Bradley Green. Joint representation learning and novel category discovery on single-and multi-modal data. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.905, + 0.719 + ], + "angle": 0, + "content": "[29] Justin Johnson, Ranjay Krishna, Michael Stark, Li-Jia Li, David Shamma, Michael Bernstein, and Li Fei-Fei. Image retrieval using scene graphs. In CVPR, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.721, + 0.905, + 0.773 + ], + "angle": 0, + "content": "[30] KJ Joseph, Sujoy Paul, Gaurav Aggarwal, Soma Biswas, Piyush Rai, Kai Han, and Vineeth N Balasubramanian. Novel class discovery without forgetting. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.905, + 0.816 + ], + "angle": 0, + "content": "[31] Valentin Khrulkov, Leyla Mirvakhabova, Evgeniya Ustinova, Ivan Oseledets, and Victor Lempitsky. Hyperbolic image embeddings. In CVPR, 2020. 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[32] Fanjie Kong, Yanbei Chen, Jiarui Cai, and Davide Modolo. Hyperbolic learning with synthetic captions for open-world detection. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[33] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In ICCV workshop, 2013. 2, 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9899" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "[34] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.482, + 0.148 + ], + "angle": 0, + "content": "[35] Huimin Li, Zhentao Chen, Yunhao Xu, and Junlin Hu. Hyperbolic anomaly detection. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.176 + ], + "angle": 0, + "content": "[36] Qi Liu, Maximilian Nickel, and Douwe Kiela. Hyperbolic graph neural networks. In NeurIPS, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.482, + 0.217 + ], + "angle": 0, + "content": "[37] Yuanpei Liu and Kai Han. Debgcd: Debiased learning with distribution guidance for generalized category discovery. In ICLR, 2025. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.218, + 0.482, + 0.272 + ], + "angle": 0, + "content": "[38] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.482, + 0.314 + ], + "angle": 0, + "content": "[39] Maximilian Nickel and Douwe Kiela. Poincaré embeddings for learning hierarchical representations. In NeurIPS, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.482, + 0.384 + ], + "angle": 0, + "content": "[40] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.385, + 0.482, + 0.427 + ], + "angle": 0, + "content": "[41] Jona Otholt, Christoph Meinel, and Haojin Yang. Guided cluster aggregation: A hierarchical approach to generalized category discovery. In WACV, pages 2618-2627, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.428, + 0.482, + 0.468 + ], + "angle": 0, + "content": "[42] Nan Pu, Zhun Zhong, and Nicu Sebe. Dynamic conceptions of contrastive learning for generalized category discovery. In CVPR, 2023. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.482, + 0.51 + ], + "angle": 0, + "content": "[43] Nan Pu, Wenjing Li, Xingyuan Ji, Yalan Qin, Nicu Sebe, and Zhun Zhong. Federated generalized category discovery. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.482, + 0.552 + ], + "angle": 0, + "content": "[44] Sarah Rastegar, Hazel Doughty, and Cees Snoek. Learn to categorize or categorize to learn? self-coding for generalized category discovery. In NeurIPS, 2023. 1, 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.482, + 0.607 + ], + "angle": 0, + "content": "[45] Sarah Rastegar, Mohammadreza Salehi, Yuki M Asano, Hazel Doughty, and Cees G M Snoek. Selex: Self-expertise in fine-grained generalized category discovery. In ECCV, 2024. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.608, + 0.482, + 0.636 + ], + "angle": 0, + "content": "[46] Ryohei Shimizu, Yusuke Mukuta, and Tatsuya Harada. Hyperbolic neural networks++. In *ICLR*, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.637, + 0.482, + 0.664 + ], + "angle": 0, + "content": "[47] Yiyou Sun and Yixuan Li. Opencon: Open-world contrastive learning. TMLR, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.665, + 0.482, + 0.705 + ], + "angle": 0, + "content": "[48] Abraham Albert Ungar. A gyrovector space approach to hyperbolic geometry. Synthesis Lectures on Mathematics and Statistics, 1(1):1-194, 2008. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.706, + 0.482, + 0.734 + ], + "angle": 0, + "content": "[49] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. JMLR, 2008. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.482, + 0.774 + ], + "angle": 0, + "content": "[50] Sagar Vaze, Kai Han, Andrea Vedaldi, and Andrew Zisserman. Open-set recognition: A good closed-set classifier is all you need? In ICLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.482, + 0.816 + ], + "angle": 0, + "content": "[51] Sagar Vaze, Kai Han, Andrea Vedaldi, and Andrew Zisserman. Generalized category discovery. In CVPR, 2022. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.858 + ], + "angle": 0, + "content": "[52] Sagar Vaze, Kai Han, Andrea Vedaldi, and Andrew Zisserman. The semantic shift benchmark. In ICML workshop, 2022. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[53] Sagar Vaze, Andrea Vedaldi, and Andrew Zisserman. No representation rules them all in category discovery. In NeurIPS, 2023. 3, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[54] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[55] Hongjun Wang, Sagar Vaze, and Kai Han. Sptnet: An efficient alternative framework for generalized category discovery with spatial prompt tuning. In ICLR, 2024. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[56] Hongjun Wang, Sagar Vaze, and Kai Han. Hilo: A learning framework for generalized category discovery robust to domain shifts. In ICLR, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.905, + 0.29 + ], + "angle": 0, + "content": "[57] Yu Wang, Zhun Zhong, Pengchong Qiao, Xuxin Cheng, Xiawu Zheng, Chang Liu, Nicu Sebe, Rongrong Ji, and Jie Chen. Discover and align taxonomic context priors for open-world semi-supervised learning. In NeurIPS, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.292, + 0.905, + 0.333 + ], + "angle": 0, + "content": "[58] Simon Weber, Bar Zongür, Nikita Araslanov, and Daniel Cremers. Flattening the parent bias: Hierarchical semantic segmentation in the poincaré ball. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.376 + ], + "angle": 0, + "content": "[59] Xin Wen, Bingchen Zhao, and Xiaojuan Qi. Parametric classification for generalized category discovery: A baseline study. In ICCV, 2023. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[60] Zhenzhen Weng, Mehmet Giray Ogut, Shai Limonchik, and Serena Yeung. Unsupervised discovery of the long-tail in instance segmentation using hierarchical self-supervision. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.433, + 0.905, + 0.489 + ], + "angle": 0, + "content": "[61] Sheng Zhang, Salman Khan, Zhiqiang Shen, Muzammal Naseer, Guangyi Chen, and Fahad Shahbaz Khan. Promptable: Contrastive affinity learning via auxiliary prompts for generalized novel category discovery. In CVPR, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.905, + 0.545 + ], + "angle": 0, + "content": "[62] Xinwei Zhang, Jianwen Jiang, Yutong Feng, Zhi-fan Wu, Xibin Zhao, Hai Wan, Mingqian Tang, Rong Jin, and Yue Gao. Grow and merge: a unified framework for continuous categories discovery. In NeurIPS, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.547, + 0.905, + 0.588 + ], + "angle": 0, + "content": "[63] Bingchen Zhao and Kai Han. Novel visual category discovery with dual ranking statistics and mutual knowledge distillation. In NeurIPS, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.59, + 0.905, + 0.631 + ], + "angle": 0, + "content": "[64] Bingchen Zhao, Xin Wen, and Kai Han. Learning semi-supervised gaussian mixture models for generalized category discovery. In ICCV, 2023. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.632, + 0.905, + 0.674 + ], + "angle": 0, + "content": "[65] Zhun Zhong, Enrico Fini, Subhankar Roy, Zhiming Luo, Elisa Ricci, and Nicu Sebe. Neighborhood contrastive learning for novel class discovery. In CVPR, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.674 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "9900" + } + ] +] \ No newline at end of file diff --git a/2025/Hyperbolic Category Discovery/103e7fd4-0c6b-413e-8aae-04fa73b4770b_origin.pdf b/2025/Hyperbolic Category Discovery/103e7fd4-0c6b-413e-8aae-04fa73b4770b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4580195598fcb166c27584c6abf34c8d58146377 --- /dev/null +++ b/2025/Hyperbolic Category Discovery/103e7fd4-0c6b-413e-8aae-04fa73b4770b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b2073eeeefea05ac253e9639709df91e19b2c898324b1f9f1303493260754663 +size 9762732 diff --git a/2025/Hyperbolic Category Discovery/full.md b/2025/Hyperbolic Category Discovery/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7fde6961f157ddac9a1f8e9d72a4ece2fc13a46f --- /dev/null +++ b/2025/Hyperbolic Category Discovery/full.md @@ -0,0 +1,325 @@ +# Hyperbolic Category Discovery + +Yuanpei Liu* Zhenqi He* Kai Han† + +Visual AI Lab, The University of Hong Kong + +{ypliu0,zhenqi_he}@connect.hku.hk kaihanx@hku.hk + +# Abstract + +Generalized Category Discovery (GCD) is an intriguing open-world problem that has garnered increasing attention. Given a dataset that includes both labelled and unlabelled images, GCD aims to categorize all images in the unlabelled subset, regardless of whether they belong to known or unknown classes. In GCD, the common practice typically involves applying a spherical projection operator at the end of the self-supervised pretrained backbone, operating within Euclidean or spherical space. However, both of these spaces have been shown to be suboptimal for encoding samples that possesses hierarchical structures. In contrast, hyperbolic space exhibits exponential volume growth relative to radius, making it inherently strong at capturing the hierarchical structure of samples from both seen and unseen categories. Therefore, we propose to tackle the category discovery challenge in the hyperbolic space. We introduce HypCD, a simple Hyperbolic framework for learning hierarchy-aware representations and classifiers for generalized Category Discovery. HypCD first transforms the Euclidean embedding space of the backbone network into hyperbolic space, facilitating subsequent representation and classification learning by considering both hyperbolic distance and the angle between samples. This approach is particularly helpful for knowledge transfer from known to unknown categories in GCD. We thoroughly evaluate HypCD on public GCD benchmarks, by applying it to various baseline and state-of-the-art methods, consistently achieving significant improvements. Project page: https://visual-ai.github.io/hypcd/ + +# 1. Introduction + +Recently, category discovery – initially explored as novel category discovery (NCD) [24] and subsequently extended to generalized category discovery (GCD) [51] – has emerged as an intriguing open-world problem, gaining increasing attention. GCD addresses the challenges posed + +![](images/416dc20d2276f8a148df659ee60187b9a177c9883a5d60fe24c761b343233734.jpg) + +![](images/95e747909c7a03310812b638ffe4b705297935e68713167428b247f101f6ffed.jpg) +Figure 1. (a) Spherical-based vs. Hyperbolic-based methods, where hyperbolic space better accommodates variations in scale and improves connections between samples. (b) Average ACC comparison of our method and previous SOTA across 'All', 'Old', and 'New' categories on the SSB [52] benchmark using DINO [7]. + +by partially labelled datasets, where the unlabelled subset may contain instances from both seen and unseen classes. The goal is to leverage knowledge from labelled data to effectively categorize the unlabelled data. Based on the way to predict category index, existing GCD methods can be broadly classified into two types: non-parametric methods [27, 44, 45, 51] and parametric methods [37, 55, 59]. Non-parametric methods predict category index based on feature clustering while parametric methods utilize a parametric classifier. + +As shown in the GCD literature [51], object parts are effective for knowledge transfer from seen to unseen categories, which is crucial for novel category discovery. Methods have been developed to explicitly learn better local features by learning pixel-level prompts around local image regions [55] or utilizing part-level features [63]. However, + +![](images/8dd1299c9dfe5ab718c05656990a2dab7d455edcb9c7b884c1663db1d255da64.jpg) +(a) Inter-Category Relation + +![](images/5279474e7184cab0d740b639eb6e62a1756d61fae0bcc1f1a11522bb04a2083c.jpg) +(b) Intra-Category Relation +Figure 2. Hierarchical relations in GCD. (a) Inter-category relationships within the Stanford-Cars dataset [33]. (b) Intra-category relationships within CUB [54] dataset. + +these methods consider object parts as rigid image patches of the same size, without considering the hierarchical nature of the object parts and the scale discrepancy of the same parts in different images, thus unavoidably restricting the performance for GCD (see Fig. 1(a)), in which the objects often have distinct poses, scales and appearance. To address this problem, one possible solution is to learn image embeddings possessing hierarchical constraints or following tree-like structures. This has been proven to be effective in many tasks. For example, in image retrieval and clustering, the hierarchy constraint may arise from whole-fragment relation [29, 31]. Intuitively, in category discovery, which can be regarded as a transfer clustering task [24], we hypothesize that an embedding space that captures the hierarchical relations of object parts can also facilitate the discovery of new categories. Indeed, the hierarchical relations have been studied in GCD, such as [27, 41, 44, 45, 57]. However, they study from a substantially different perspective: inter-category hierarchy (see Fig. 2(a)). This only considers the hierarchy of different semantic classes from coarse to fine levels. Additionally, these methods require the relationships and number of levels in the hierarchy to be predefined, resulting in a lack of flexibility and scalability. Moreover, these methods are unable to capture more complex hierarchies, such as the compositional parts of an object (see Fig. 2(b)). This is particularly the case because existing methods [45, 51, 55, 59], no matter whether they consider the hierarchy or not, learn the image embeddings in a spherical space. This follows the common practice of applying a spherical projection operator at the end of the self-supervised feature backbone [7, 40]. Consequently, all subsequent operations, including distance calculations, are performed under either Euclidean or spherical geometry, resulting in limited awareness of hierarchical object parts. + +In this work, we study the overlooked perspective in category discovery: instead of learning in the Euclidean or spherical space, we advocate a space that captures the hierarchical structure of each data point. In spherical space, both the radius and volume are constant, whereas in Euclidean space, the volume grows polynomially with respect to the radius. Both of these spaces have been shown to be suboptimal for encoding samples that possess hierarchical + +structures [10, 15, 31]. In contrast, hyperbolic space possesses a distinctive property where its volume grows exponentially relative to the radius. This characteristic makes hyperbolic space particularly suitable for embedding tree-like data, enhancing its representational power. Learning representations in hyperbolic space has proven to be effective in various computer vision tasks, including object recognition [15], object detection [32], semantic segmentation [58], and anomaly detection [35]. Inspired by these successes, we aim to realize the idea of learning hierarchy-aware representations to facilitate knowledge transfer in category discovery, thereby unleashing the potential of hyperbolic representations. + +To achieve this goal, we propose a simple yet effective framework, HypCD, to properly learn the hierarchy-aware representation and classifier for category discovery through the lens of hyperbolic geometry. In this framework, we adapt our framework to popular parametric [59] and non-parametric [51] GCD baselines as well as the state-of-the-art (SOTA) method SelEx [45], obtaining substantial improvements for them, establishing the new SOTA (see Fig. 1(b)). Firstly, starting from the self-supervised backbone pretrained in Euclidean space, we propose to map the Euclidean representation to a constrained Poincaré ball through feature clipping and exponential mapping. Secondly, we implement the hyperbolic representation learning and build a hyperbolic classifier on the Poincaré ball, considering both angle and distance between samples in hyperbolic space. Thirdly, to assign labels to unlabelled data after training, for non-parametric methods, we apply semi-supervised $k$ -means following the common practice; for parametric methods, we employ a hyperbolic classifier to make predictions. Despite its simplicity, our framework achieves significant performance improvements with two different pretrained weights (DINO [7] and DINOv2 [40]) on the public GCD datasets, including the coarse-grained classification datasets CIFAR-10 [34], CIFAR-100 [34], and ImageNet-100 [13], as well as the fine-grained SSB [50] benchmark. + +In summary, we make the following contributions in this paper: (i) We identify the existing GCD methods' common shortcoming in encoding the hierarchical structure and propose to incorporate the hyperbolic geometry into the embedding space to address this limitation; (ii) We propose a simple yet effective framework, called HypCD, for incorporating the hyperbolic geometry in representation learning and classification for category discovery; and (iii) Through extensive experiments on public GCD benchmarks by applying HypCD to baseline and SOTA methods, our method consistently demonstrates effectiveness and superiority. + +# 2. Related Work + +Category Discovery. Novel category discovery (NCD) is + +initially introduced in [24] to establish a realistic framework for transferring knowledge from seen categories to cluster unseen categories, by considering it as a transfer clustering problem. Many subsequent methods have been proposed to advance the field [17, 25, 26, 28, 63, 65]. Generalized category discovery [51] (GCD) relaxes the assumption in NCD by considering unlabelled data containing samples from both known and unknown classes [51]. Further investigations [6, 8, 27, 30, 37, 42, 56] have explored a variety of strategies to address the challenges posed by GCD. One notable approach, SimGCD [59], proposes to learn a parametric classifier enhanced by mean entropy regularization, thereby improving performance. In another vein, GPC [64] employs Gaussian mixture models to jointly learn robust representations while simultaneously estimating the number of unknown categories. SPT-Net [55] presents a spatial prompt tuning method that enables models to concentrate more effectively on specific object parts, thus enhancing knowledge transfer in GCD. Most recently, SelEx [45] has been proposed, leveraging hierarchical semi-supervised $k$ -means to achieve SOTA results on fine-grained datasets. Additionally, various efforts are focused on addressing category discovery from multiple perspectives. For instance, [28] emphasizes multi-modal category discovery; [62] and [8] explore a continual setting; [43] studies category discovery in a federated setting; and [56] examines GCD in the presence of domain shifts. + +Hyperbolic Geometry. Hyperbolic space, defined as a non-Euclidean manifold with exponential volume growth in relation to its radius, is inherently aligned with the embedding of tree-like and hierarchical data structures in visual recognition tasks. Significant advancements in this area include [31], which presents a hyperbolic image embedding technique by projecting model outputs into hyperbolic space, and [15], which integrates hyperbolic geometry into various vision transformer architectures, showcasing performance that surpasses their Euclidean counterparts. Hyperbolic methods have also been developed on diverse tasks such as image classification [15, 23, 31], action recognition [18], few-shot learning [20] and object segmentation [21, 60]. Moreover, recent developments have introduced hyperbolic geometry for neural networks including fully connected layers [46], convolutional neural networks [3], graph neural networks [9, 36], and attention network [22], thereby facilitating a deeper integration of hyperbolic geometry into deep learning regime. + +# 3. Method + +In this section, we first introduce the task in Sec. 3.1, then move to a review of baselines in Sec. 3.2. Afterwards, the geometry mapping and training details of our framework are described in Sec. 3.3 and Sec. 3.4. Lastly, the label assignment details are outlined in Sec. 3.5. + +# 3.1. Problem Statement + +GCD aims to learn a model capable of accurately classifying unlabelled samples from known categories while simultaneously clustering those from unknown categories. Consider an unlabelled dataset denoted as $\mathbf{D}_u = \left\{\left(\mathbf{x}_i^u,y_i^u\right)\right\} \in$ $\mathbf{X}\times \mathbf{Y}_u$ and a labelled dataset represented as $\mathbf{D}_l =$ $\left\{\left(\mathbf{x}_i^l,y_i^l\right)\right\} \in \mathbf{X}\times \mathbf{Y}_l,$ where $\mathbf{Y}_u$ and $\mathbf{Y}_l$ denote the respective label sets. The unlabelled dataset comprises samples from both known and unknown categories, i.e., specifically $\mathbf{Y}_l\subset \mathbf{Y}_u$ . Let the number of labelled categories be denoted by $M = |\mathbf{Y}_l|$ . We assume that the total number of categories, $K = |\mathbf{Y}_l\cup \mathbf{Y}_u|$ , is known, as posited in prior works [26, 53, 59]. In scenarios where this information is unavailable, alternative methods such as those proposed in [24, 51] can be employed to yield a reliable estimation. + +# 3.2. Review of Baselines + +Non-parametric Baseline. [51] formalizes the GCD task and proposes a non-parametric baseline. The approach involves finetuning the pre-trained DINO [7] model [14] to enhance the learned representation. The loss function comprises a supervised contrastive loss, which operates on the labelled samples, and a self-supervised contrastive loss, which operates on all the samples. Specifically, given two randomly augmented views $\mathbf{x}_i$ and $\mathbf{x}_i'$ for the same image in a mini-batch $B$ , the self-supervised contrastive loss is: + +$$ +\mathcal {L} _ {r e p} ^ {u} = \frac {1}{| B |} \sum_ {i \in B} - \log \frac {\exp \left(\mathbf {z} _ {i} \cdot \mathbf {z} _ {i} ^ {\prime} / \tau_ {r}\right)}{\sum_ {j} ^ {j \neq i} \exp \left(\mathbf {z} _ {i} \cdot \mathbf {z} _ {j} ^ {\prime} / \tau_ {r}\right)}, \tag {1} +$$ + +where the feature $\mathbf{z}_i = \rho_r(\phi (\mathbf{x}_i))$ is a $\ell_2$ -normalized vector and $\mathbf{z}_i^\prime$ represents feature from another view $\mathbf{x}_i^\prime$ . Here, $\phi$ refers to the backbone network, $\rho_r$ denotes the projection head, and $\tau_{r}$ stands as the temperature parameter used for scaling the features. The supervised contrastive loss for labelled samples is: + +$$ +\mathcal {L} _ {r e p} ^ {s} = \frac {1}{| B _ {l} |} \sum_ {i \in B _ {l}} \frac {1}{| N _ {i} |} \sum_ {q \in N _ {i}} - \log \frac {\exp \left(\mathbf {z} _ {i} \cdot \mathbf {z} _ {q} / \tau_ {r}\right)}{\sum_ {j} ^ {j \neq i} \exp \left(\mathbf {z} _ {i} \cdot \mathbf {z} _ {j} / \tau_ {r}\right)}, \tag {2} +$$ + +where $N_{i}$ is the index set for all other images in the labelled mini-batch $B_{l} \subset B$ having the same label as $\mathbf{x}_{i}$ . The overall representation learning loss is then: $\mathcal{L}_{rep} = (1 - \lambda_b)\mathcal{L}_{rep}^u + \lambda_b\mathcal{L}_{rep}^s$ , where $\lambda_{b}$ is a balance factor. + +Parametric Baseline. [59] introduces a robust parametric GCD baseline, which has been widely adopted in the field ever since [53, 55]. This method employs a parametric classifier implemented in a self-distillation framework [7]. The classifier is randomly initialized with $K$ normalized category prototypes $\mathbf{C} = \{\mathbf{c}_1,\dots,\mathbf{c}_K\}$ . For a randomly augmented view $\mathbf{x}_i$ and its corresponding normalized hidden feature vector $\mathbf{h}_i = \phi (\mathbf{x}_i) / ||\phi (\mathbf{x}_i)||$ , the output probability for the $k$ -th category is given by: + +$$ +\mathbf {p} _ {i} ^ {(k)} = \frac {\exp \left(\mathbf {h} _ {i} \cdot \mathbf {c} _ {k} / \tau_ {s}\right)}{\sum_ {j = 1} ^ {K} \exp \left(\mathbf {h} _ {i} \cdot \mathbf {c} _ {j} / \tau_ {s}\right)}, \tag {3} +$$ + +where $\tau_{s}$ is the scaling temperature for the 'student' view. The soft label $\mathbf{q}_i$ is generated by the teacher' view with a sharper temperature $\tau_{t}$ based on another augmented view in a similar manner. The self-distillation loss for the two views is then computed using the cross-entropy loss function: $\ell_{ce}(\mathbf{q}',\mathbf{p}) = -\sum_{j = 1}^{K}\mathbf{q}'^{(j)}\log \mathbf{p}^{(j)}$ . The unsupervised loss is then computed by aggregating contributions from all samples in the mini-batch $B$ as follows: + +$$ +\mathcal {L} _ {c l s} ^ {u} = \frac {1}{| B |} \sum_ {i \in B} \ell_ {c e} \left(\mathbf {q} _ {i} ^ {\prime}, \mathbf {p} _ {i}\right) - \xi \mathcal {H} (\bar {\mathbf {p}}), \tag {4} +$$ + +where $\overline{\mathbf{p}} = \frac{1}{2|B|}\sum_{i\in B}(\mathbf{p}_i + \mathbf{p}_i')$ denotes the mean prediction across the mini-batch. The mean entropy is defined as: $\mathcal{H}(\overline{\mathbf{p}}) = -\sum_{j = 1}^{K}\overline{\mathbf{p}}^{(j)}\log \overline{\mathbf{p}}^{(j)}$ , weighted by $\xi$ . + +For the labelled samples, the supervised classification loss is written as $\mathcal{L}_{cls}^{s} = \frac{1}{|B_{l}|}\sum_{i\in B_{l}}\ell_{ce}(\mathbf{p}_{i},\mathbf{y}_{i})$ , where $\mathbf{y}_i$ represents the one-hot vector corresponding to the ground-truth label $y_{i}$ . The overall objective is $\mathcal{L}_{cls} = (1 - \lambda_b)\mathcal{L}_{cls}^u + \lambda_b\mathcal{L}_{cls}^s$ . Integrating this with the representation learning loss $\mathcal{L}_{rep}$ adopted from [51], the comprehensive training objective is expressed as: $\mathcal{L}_{gcd} = \mathcal{L}_{cls} + \mathcal{L}_{rep}$ . Through training with $\mathcal{L}_{gcd}$ on both $\mathbf{D}_l$ and $\mathbf{D}_u$ , the classifier is empowered to directly predict labels for the unlabelled samples after the training process concludes. + +# 3.3. Hyperbolic Space for Category Discovery + +As previously discussed, object parts are critical for facilitating knowledge transfer from labelled categories to unseen ones in GCD. Each sample inherently contains object parts that reside within a hierarchical structure. Moreover, existing GCD methods [44, 45] emphasize the intercategory hierarchy to enhance the clustering performance of unlabelled samples in Euclidean or spherical spaces. However, the geometry of representation space limits their ability to effectively capture other kinds of hierarchy [31]. In contrast, hyperbolic space, characterized by its property of exponential volume growth with respect to the radius [15], emerges as a more suitable space for GCD. + +Hyperbolic space $\mathbb{H}^n$ is defined as an $n$ -dimensional Riemannian manifold exhibiting constant negative curvature, and it encompasses several analytic models [5]. Following previous literature [15, 31], we employ the Poincaré ball [39] model. In this model, the hyperbolic space is represented as an $n$ -dimensional ball $\mathbb{D}_c^n = \{\mathbf{a}\in \mathbb{R}^n\mid c\| \mathbf{a}\| ^2 < 1\}$ with curvature value $-c^2$ , where $c$ is the non-negative curvature parameter. The manifold is equipped with the Riemannian metric $g^{\mathbb{D}} = \lambda_c^2 g^{\mathbb{E}}$ where $\lambda_{c}(\mathbf{a}) = \frac{2}{1 - c\|\mathbf{a}\|^2}$ is the conformal factor and $g^{\mathbb{E}}$ is the identity metric $\mathbf{I}_n$ in Euclidean space. In this way, the local distances are scaled by the factor $\lambda_{c}$ approaching infinity near the boundary of the ball. This gives rise to the exponential expansion property of hyperbolic spaces, unlike the polynomial expansion in Euclidean space. However, hyperbolic + +space is not vector space and thus operations such as addition can not be directly conducted. To address this problem, we leverage the gyrovector formalism [48]. For a pair of points $\mathbf{a},\mathbf{b}\in \mathbb{D}_c^n$ , their Möbius addition is defined as: + +$$ +\mathbf {a} \oplus_ {c} \mathbf {b} = \frac {(1 + 2 c \langle \mathbf {a} , \mathbf {b} \rangle + c \| \mathbf {b} \| ^ {2}) \mathbf {a} + (1 - c \| \mathbf {a} \| ^ {2}) \mathbf {b}}{1 + 2 c \langle \mathbf {a} , \mathbf {b} \rangle + c ^ {2} \| \mathbf {a} \| ^ {2} \| \mathbf {b} \| ^ {2}}. \qquad (5) +$$ + +The hyperbolic distance between them is then: + +$$ +\mathcal {D} _ {\mathbb {H}} (\mathbf {a}, \mathbf {b}) = \frac {2}{\sqrt {c}} \operatorname {a r c t a n h} \left(\sqrt {c} \| - \mathbf {a} \oplus_ {c} \mathbf {b} \|\right) \tag {6} +$$ + +When $c\to 0$ , the hyperbolic distance (Eq. 6) closes to the Euclidean distance $\lim_{c\to 0}\mathcal{D}_{\mathbb{H}}(\mathbf{a},\mathbf{b}) = 2\| \mathbf{a} - \mathbf{b}\|$ + +# 3.4. HypCD + +As illustrated in Fig. 3, we propose a unified framework, HypCD, for category discovery in hyperbolic space, incorporating both parametric [59] and non-parametric GCD approaches. Given two randomly augmented views, we initially obtain the respective Euclidean feature vectors $\mathbf{z}_i$ and $\mathbf{z}_i'$ through a self-supervised pretrained backbone [7, 40]. Subsequently, the feature embeddings are mapped into hyperbolic space $\mathbb{H}^n$ using exponential mapping, facilitating representation learning within this exponentially growing space to more effectively capture and utilize the hierarchical relationships inherent in the training data. + +The exponential mapping [31] serves as a bijective projection between Euclidean space $\mathbb{E}^n$ and hyperbolic space $\mathbb{H}^n$ . The projection of tangent vector $\mathbf{z}$ from $\mathbb{E}^n$ to $\mathbb{H}^n$ is formulated as: + +$$ +\exp_ {\mathbf {o}} ^ {c} (\mathbf {z}) = \mathbf {o} \oplus_ {c} \left(\tanh \left(\sqrt {c} \frac {\lambda_ {\mathbf {o}} ^ {c} \| \mathbf {z} \|}{2}\right) \frac {\mathbf {z}}{\sqrt {c} \| \mathbf {z} \|}\right), \tag {7} +$$ + +where $\oplus_{c}$ is the Mobius addition, as introduced in Eq. 5 and $\mathbf{o}$ represents the base point of the mapping. To address the issue of gradient vanishing [23] near the boundary of the Poincaré ball during training, we implement a feature clipping operation in $\mathbb{E}^n$ prior to the exponential mapping. The operation is defined as: $\mathcal{C}(\mathbf{z}) = \min \{1,\frac{r}{||\mathbf{z}||}\} \cdot \mathbf{z}$ , where $r$ denotes the clipping value. For the feature vector $\mathbf{z}_i$ in $\mathbb{E}^n$ , the corresponding mapped feature in $\mathbb{H}^n$ is expressed as $\mathcal{M}(\mathbf{z}_i) = \exp_{\mathbf{o}}^{c}(\mathcal{C}(\mathbf{z}_{i}))$ . The same operation will also be applied to the other feature vector $\mathbf{z}_i'$ . + +As described in Sec.3.2, both parametric [51] and nonparametric [59] baselines utilize the same representation learning method. In our framework, we implement a consistent representation learning solution in hyperbolic space for them (Fig.3(a)). For parametric approaches, a hyperbolic parametric classifier is employed (Fig. 3(b)). We will introduce these components in detail subsequently. + +Hyperbolic Representation Learning. Following prior attempts [45, 51, 59], we incorporate both self-supervised and supervised contrastive learning into our framework. However, our approach uniquely operates within hyperbolic + +![](images/ed52ceb0b1900ddee19b23f6af1d6331bef6de88be7adb5c9f5c0c05431eca90.jpg) + +![](images/7ab57b7d5ec05ca663bcf6556fe08cf1a501351ba238d6bd7a0c18438767f5e7.jpg) + +![](images/2138d4663106c5d07f1aedce0e931db7dfed1e924a16e028f65d8579b2ac8c94.jpg) +Figure 3. Overall pipeline of our HypCD framework for parametric and non-parametric GCD baselines. (a) Hyperbolic representation learning. (b) Hyperbolic classifier. (c) Non-parametric label assignment. (d) Parametric label assignment. + +![](images/8a985e64d6e865db5b5b45af74fbc24dae17c89265f3d9603a7eac2402797f71.jpg) + +space. Furthermore, unlike previous GCD methods that exclusively utilize cosine distance [51, 55, 59] (angle-based) or Euclidean distance [45] (distance-based) for calculating pairwise similarity, we propose a hybrid approach that combines both distance-based and angle-based losses. Such integration has been shown to be more effective for model optimization in hyperbolic space [18]. First, the unified form of self-supervised contrastive loss can be defined as: + +$$ +\mathcal {L} ^ {u} = \frac {1}{| B |} \sum_ {i \in B} - \log \frac {\exp \left(\mathcal {S} \left(\mathcal {M} \left(\mathbf {z} _ {i}\right) , \mathcal {M} \left(\mathbf {z} _ {i} ^ {\prime}\right)\right) / \tau_ {r}\right)}{\sum_ {j} ^ {j \neq i} \exp \left(\mathcal {S} \left(\mathcal {M} \left(\mathbf {z} _ {i}\right) , \mathcal {M} \left(\mathbf {z} _ {j} ^ {\prime}\right)\right) / \tau_ {r}\right)}. \tag {8} +$$ + +Similarly, the supervised contrastive loss is unified as: + +$$ +\mathcal {L} ^ {s} = \frac {1}{| B _ {l} |} \sum_ {i \in B _ {l}} \frac {1}{| N _ {i} |} \sum_ {q \in N _ {i}} - \log \frac {\exp \left(\mathcal {S} \left(\mathcal {M} \left(\mathbf {z} _ {i}\right) , \mathcal {M} \left(\mathbf {z} _ {q}\right)\right) / \tau_ {r}\right)}{\sum_ {j \neq i} \exp \left(\mathcal {S} \left(\mathcal {M} \left(\mathbf {z} _ {i}\right) , \mathcal {M} \left(\mathbf {z} _ {j}\right)\right) / \tau_ {r}\right)}, \tag {9} +$$ + +where $S$ denotes the similarity function, which can be either distance-based or angle-based. For distance-based contrastive loss, we utilize $S_{d} = -\mathcal{D}_{\mathbb{H}}$ as the similarity function, which is formally computed using negative Euclidean distance in prior methods [45]. For angle-based contrastive loss, we employ the cosine similarity, formulated as: + +$$ +\mathcal {S} _ {a} \left(\mathcal {M} \left(\mathbf {z} _ {i}\right), \mathcal {M} \left(\mathbf {z} _ {i} ^ {\prime}\right)\right) = \frac {\mathcal {M} \left(\mathbf {z} _ {i}\right) \cdot \mathcal {M} \left(\mathbf {z} _ {i} ^ {\prime}\right)}{\left| \left| \mathcal {M} \left(\mathbf {z} _ {i}\right) \right| \right| \cdot \left| \left| \mathcal {M} \left(\mathbf {z} _ {i} ^ {\prime}\right) \right| \right|}. \tag {10} +$$ + +Since hyperbolic space is conformal with Euclidean space, cosine similarity remains equivalent in both $\mathbb{E}^n$ and $\mathbb{H}^n$ . + +The final supervised and self-supervised hyperbolic contrastive loss is composed of both distance-based and angle-based losses: + +$$ +\mathcal {L} _ {h r e p} ^ {s} = \alpha_ {d} \mathcal {L} _ {d i s} ^ {s} + (1 - \alpha_ {d}) \mathcal {L} _ {a n g} ^ {s}, \tag {11} +$$ + +$$ +\mathcal {L} _ {h r e p} ^ {u} = \alpha_ {d} \mathcal {L} _ {d i s} ^ {u} + (1 - \alpha_ {d}) \mathcal {L} _ {a n g} ^ {u}, +$$ + +where $\mathcal{L}_{hrep}^s$ and $\mathcal{L}_{hrep}^u$ represent the supervised and self-supervised hyperbolic contrastive loss, respectively. The + +terms $\mathcal{L}_{dis}$ and $\mathcal{L}_{ang}$ correspond to distance-based and angle-based contrastive loss, respectively, obtained by substituting $S$ with $S_{d}$ and $S_{a}$ . Additionally, $\alpha_{d}$ is the loss weight of distance-based loss. The overall training objective for hyperbolic representation learning is: + +$$ +\mathcal {L} _ {r e p} ^ {\mathbb {H}} = \left(1 - \lambda_ {b} ^ {\mathbb {H}}\right) \mathcal {L} _ {h r e p} ^ {u} + \lambda_ {b} ^ {\mathbb {H}} \mathcal {L} _ {h r e p} ^ {s}, \tag {12} +$$ + +where $\lambda_b^{\mathbb{H}}$ serves as the balancing factor between the supervised and unsupervised losses. + +Hyperbolic Classifier. To enhance the parametric baseline with hyperbolic geometry, we replace the conventional Euclidean classification head—traditionally reliant on a multilayer perceptron (MLP) in Euclidean space—with its hyperbolic counterpart, the hyperbolic feed forward network (HypFFN). The hyperbolic linear layer [19] exhibits greater alignment with the baseline [59], and we experimentally find that it outperforms the hyperbolic multinomial logistic regression layer. Consider the last linear layer of the MLP; similar to its Euclidean counterpart, the hyperbolic linear layer is parameterized by a weight matrix $\mathbf{w} \in \mathbb{R}^{I \times K}$ and a bias vector $\mathbf{s} \in \mathbb{R}^{1 \times K}$ , where $I$ denotes the input feature dimension. Given the hyperbolic feature $\mathbf{z}_i^{\mathbb{H}} = \mathcal{M}(\mathbf{z}_i) \in \mathbb{R}^{1 \times I}$ , the linear layer operates as HypLinear $(\mathbf{z}_i^{\mathbb{H}}, \mathbf{w}, \mathbf{s}) = \operatorname{Proj}[(\mathbf{w} \otimes_c \mathbf{z}_i^{\mathbb{H}}) \oplus_c \mathbf{s}]$ , where $\oplus_c$ follows Eq. 5. The Möbius matrix-vector multiplication $\mathbf{v}_i = \mathbf{w} \otimes_c \mathbf{z}_i^{\mathbb{H}}$ is defined as: + +$$ +\frac {1}{\sqrt {c}} \tanh \left(\frac {\| \mathbf {z} _ {i} ^ {\mathbb {H}} \mathbf {w} \| _ {2}}{\| \mathbf {z} _ {i} ^ {\mathbb {H}} \| _ {2}} \tanh ^ {- 1} (\sqrt {c} \| \mathbf {z} _ {i} ^ {\mathbb {H}} \| _ {2})\right) \frac {\mathbf {z} _ {i} ^ {\mathbb {H}} \mathbf {w}}{\| \mathbf {z} _ {i} ^ {\mathbb {H}} \mathbf {w} \| _ {2}}. \tag {13} +$$ + +To ensure numerical stability [19], a safe projection is operated on the result manifold and represented as: + +$$ +\operatorname {P r o j} (\mathbf {v} _ {i}) = \left\{ \begin{array}{l l} \frac {\mathbf {v} _ {i}}{\| \mathbf {v} _ {i} \| _ {2}} \times \frac {1 - 1 0 ^ {- 3}}{\sqrt {c}}, & \frac {1 - 1 0 ^ {- 3}}{\sqrt {c}} < \| \mathbf {v} _ {i} \| _ {2} \\ \mathbf {v} _ {i}, & \text {o t h e r w i s e} \end{array} . \right. \tag {14} +$$ + +This integration allows our hyperbolic classifier to be seamlessly incorporated into the baseline [59] by substituting the original MLP with HypFFN. For each point in $\mathbb{H}^n$ , the tangent space at that point serves as a Euclidean subspace, enabling straightforward adaptation of Euclidean operations within this space [10]. Consequently, the cross-entropy loss for the hyperbolic classifier can be expressed as: $\ell_{ce}^{\mathbb{H}} = \ell_{ce}(\mathrm{HypFFN}(\mathbf{z}_i^{\mathbb{H}}),\mathbf{y}_i)$ . Additionally, we can define the hyperbolic counterpart $\mathcal{H}^{\mathbb{H}}$ for the mean entropy $\mathcal{H}$ . By substituting the original $\ell_{ce}$ and $\mathcal{H}$ with our derived $\ell_{ce}^{\mathbb{H}}$ and $\mathcal{H}^{\mathbb{H}}$ , respectively, we can readily compute the final hyperbolic classifier loss $\mathcal{L}_{cls}^{\mathbb{H}}$ , as detailed in Sec. 3.2. + +# 3.5.Label Assignment + +Existing approaches typically employ either a parametric classification head or non-parametric methods, such as semi-supervised $k$ -means [51], for label assignment. In this paper, we do not independently assess these two methods; rather, we integrate both within the HypCD framework as shown in Fig. 3(c) and (d). For non-parametric approaches, including [51] and the recent SelEx [45], we retain the original label assignment strategy by applying semi-supervised $k$ -means clustering directly to feature representations extracted by $\phi$ in $\mathbb{E}^n$ . Our empirical results indicate that training in hyperbolic space allows for the transfer of hierarchical structure encoding from $\mathbb{H}^n$ to $\mathbb{E}^n$ . Moreover, we find that the operations of $k$ -means in $\mathbb{E}^n$ are significantly more efficient while maintaining comparable performance. For the parametric baseline exemplified by SimGCD [59], we utilize the hyperbolic classification head to conduct classification within hyperbolic space using the trained hyperbolic classifier HypFFN. Both design choices are theoretically supported by the property of hyperbolic geometry of encoding hierarchical structures, facilitating a more intuitive and effective representation and classifier for GCD. + +# 4. Experiment + +# 4.1. Setups and Implementations + +Datasets. We thoroughly evaluate our method across diverse benchmarks, including the generic image recognition datasets CIFAR-10 and CIFAR-100 [34], as well as ImageNet-100 [13]. Additionally, we assess our approach on the Semantic Shift Benchmark (SSB)[52], which includes fine-grained datasets such as CUB[54], StanfordCars [33], and FGVC-Aircraft [38]. For each dataset, we adhere to the data split scheme detailed in [51]. The method involves sampling a subset of all classes as the known ('Old') classes $\mathbf{Y}_l$ . Subsequently, $50\%$ of the images from these known classes are utilized to construct $\mathbf{D}_l$ , while the remaining images are designated as the unlabelled data $\mathbf{D}_u$ . Evaluation Metrics. We evaluate the performance using the clustering accuracy (ACC) as defined in the litera + +ture [51]. The $ACC$ on $\mathbf{D}_u$ is computed as follows, given the ground truth $y_{i}$ and the predicted labels $\hat{y}_i$ : $ACC = \frac{1}{|\mathbf{D}_u|}\sum_{i = 1}^{|\mathbf{D}_u|}\mathbb{1}(y_i = h(\hat{y}_i))$ , where $h$ denotes the optimal permutation that aligns the predicted cluster assignments with the ground-truth labels. The $ACC$ values for the 'All', 'Old' and 'New' classes are reported separately. + +Implementation Details. We evaluate HypCD against the non-parametric baseline GCD [51], the parametric baseline SimGCD [59], and the SOTA method SelEx [45], utilizing both DINO [7] and DINoV2 [40] pretrained weights. Detailed information regarding SelEx can be found in the supplementary materials. For GCD [51], the output dimension of the projection head $\rho_{r}$ is 256. In the case of SimGCD [59], the feature dimension from backbone $\phi$ is 768. $\rho_{r}$ and the final block of $\phi$ are optimized using the SGD optimizer, with an initial learning rate of 0.1, which is decayed to 0.001 over time according to a cosine annealing schedule. The HypFFN is optimized using the Riemannian Adam optimizer [4], with a constant learning rate of 0.01. All models are trained for 200 epochs using a batch size of 128. The curvature parameter $c$ is set to 0.05 for the fine-grained datasets and 0.01 for the generic datasets. Following baselines, the balancing factor $\lambda_b^{\mathbb{H}}$ is set to 0.35. By default, the loss weight $\alpha_{d}$ increases linearly from 0 to 1.0. + +# 4.2. Quantitative Comparison + +We compare our method with recent GCD methods (including ORCA [6], GCD [51], XCon [16], OpenCon [47], PromptCAL [61], DCCL [42], GPC [64], CiPR [27], SimGCD [59], $\mu$ GCD [53], InfoSieve [44], SPTNet [55], CMS [12], AMEND [2] and SelEx [45]) and report the results in Tab. 1. The evaluation encompasses performance on the SSB benchmark [52] and generic datasets [13, 34]. The hyperbolic methods applying our HypCD framework are indicated by the 'Hyp-' prefix. + +![](images/b3dc1a23b1372253764360b526da3ee5f3d8e0e583c844b07bad0b328d1d1f54.jpg) +Figure 4. Comparison of baseline and hyperbolic counterparts on the SSB. Left: 'All' ACC (higher is better). Right: Discrepancy between 'Old' and 'New' ACC (smaller is better). + +![](images/16a3a827d444bb117b2bbcf99c6f41a005eb7276158bb6c533c0e4a90a503dc0.jpg) + +Results on SSB. The performance of the GCD methods on the SSB benchmark, utilizing both DINO [7] and DINOv2 [40] pretrained weights, is summarized in the left section of Tab. 1. Besides, we provide a comparative analy- + +Table 1. Comparison of GCD methods on the SSB [52] benchmark, CIFAR-10 [34], CIFAR-100 [34] and ImageNet-100 [13] datasets. Results are reported in ACC across the 'All', 'Old' and 'New' categories. + +
MethodCUB [54]Stanford-Cars [33]FGVC-Aircraft [38]CIFAR-10 [34]CIFAR-100 [34]ImageNet-100 [13]
AllOldNewAllOldNewAllOldNewAllOldNewAllOldNewAllOldNew
DINOORCA [6]36.343.832.631.632.031.431.942.226.969.077.452.073.592.663.981.886.279.6
XCon [16]52.154.351.040.558.831.747.744.449.496.097.395.474.281.260.377.693.569.7
OpenCon [47]54.763.854.749.178.632.7---------84.093.881.2
PromptCAL [61]62.964.462.150.270.140.652.252.252.397.996.698.581.284.275.383.192.778.3
DCCL [42]63.560.864.943.155.736.2---96.396.596.975.376.870.280.590.576.2
GPC [64]52.055.547.538.258.927.443.340.744.890.697.687.075.484.660.175.393.466.7
PIM [11]62.775.756.243.166.931.6---94.797.493.378.384.266.583.195.377.0
μGCD [53]65.768.064.656.568.150.953.855.453.0---------
InfoSieve [44]69.477.965.255.774.846.456.363.752.594.897.793.478.382.270.580.593.873.8
CiPR [27]57.158.755.647.061.540.1---97.797.597.781.582.479.780.584.978.3
SPTNet [55]65.868.865.159.079.249.359.361.858.197.395.098.681.384.375.685.493.281.4
CMS [12]68.276.564.056.976.147.656.063.452.3---82.385.775.584.795.679.2
AMEND [2]64.975.659.652.861.848.356.473.348.296.894.697.881.079.983.383.292.978.3
GCD [51]51.356.648.739.057.629.945.041.146.991.597.988.273.076.266.574.189.866.3
Hyp-GCD61.067.058.050.860.945.848.243.650.592.997.590.674.080.062.080.492.574.4
SimGCD [59]60.365.657.753.871.945.054.259.151.897.195.198.180.181.277.883.093.177.9
Hyp-SimGCD64.865.864.262.873.457.758.758.958.596.895.997.282.483.181.286.593.783.0
SelEx [45]73.675.372.858.575.650.357.164.753.395.998.194.882.385.376.383.193.677.8
Hyp-SelEx79.875.881.862.980.054.765.967.365.196.797.696.382.485.177.086.894.682.8
DINOv2μGCD [53]74.075.973.176.191.068.966.368.765.1---------
CiPR [27]78.373.480.866.777.061.8---99.098.799.290.389.093.188.287.688.5
SPTNet [55]76.379.574.6------------90.196.187.1
GCD [51]71.971.272.365.767.864.755.447.959.297.899.097.179.684.569.978.589.573.0
Hyp-GCD75.675.175.972.880.469.162.770.059.097.598.996.884.587.578.582.992.478.2
SimGCD [59]71.578.168.371.581.966.663.969.960.998.796.799.788.589.287.289.995.587.1
Hyp-SimGCD77.677.977.482.585.881.076.470.379.498.997.799.591.590.094.691.996.289.8
SelEx [45]87.485.188.582.293.776.779.882.378.698.5*98.8*98.5*87.7*90.8*81.5*90.9*96.2*88.3*
Hyp-SelEx90.785.393.483.893.379.283.482.084.198.698.198.988.691.582.892.396.490.2
+ +\*results from our implementation. + +sis between the three baseline methods and their hyperbolic counterparts with DINO backbone in Fig. 4. Our hyperbolic methods consistently outperform their Euclidean counterparts, with particularly strong results observed when utilizing the DINOv2 backbone. Among the evaluated methods, Hyp-SelEx achieves the highest average accuracy (ACC) across all datasets, notably excelling on the CUB dataset, where it records an accuracy of $79.8\%$ for the 'All' classes with DINO and $90.7\%$ with DINOv2, establishing it as the leading approach. More strikingly, on the Stanford-Cars dataset, our Hyp-GCD method outperforms the baseline by $11.8\%$ , $13.3\%$ and $15.9\%$ in terms of ACC for the 'All', 'Old' and 'New' categories, respectively. Fig. 4 (left) illustrates the average ACC on 'All' categories across the three datasets in the SSB benchmark, indicating that our hyperbolic methods surpass the baseline by a margin of at least $6.0\%$ . Furthermore, as shown in Fig. 4 (right), hyperbolic methods exhibit a consistently smaller ACC gap between 'Old' and 'New' classes, highlighting the effectiveness of HypCD in enhancing knowledge transfer from known to unseen categories. Additionally, DINOv2 outperforms DINO across all methods, underscoring its ability to capture complex data representations more effectively. + +Results on Generic Datasets. In the right section of Tab. 1, + +we present the results on three widely used generic datasets: CIFAR-10 [34], CIFAR-100 [34], and ImageNet-100 [13]. Our methods demonstrate consistent improvements across all cases, regardless of the backbone employed. Notably, these enhancements are especially significant on CIFAR-100 and ImageNet-100, which present greater challenges compared to CIFAR-10, where performance is nearly saturated. For CIFAR-100, Hyp-SimGCD and Hyp-SelEx achieve the highest accuracy of $82.4\%$ for the 'All' categories using DINO, while Hyp-SimGCD ranks first with an accuracy of $91.5\%$ on this metric when utilizing DINOv2, significantly surpassing baseline methods and the previous SOTA. Results on ImageNet-100 further validate the effectiveness of hyperbolic methods; Hyp-SelEx achieves the highest performance across 'All', 'Old', and 'New' categories with both DINO and DINOv2, outperforming the baseline by a margin of up to $3.7\%$ . + +# 4.3. Impact of Hyperparameters + +Manifold Curvature. Building on previous studies [1, 31] that explore the application of hyperbolic geometry across various tasks, the curvature parameter $c$ (as discussed in Sec. 3.3) is a crucial factor influencing performance and may yield different optimal values across datasets and meth + +Table 2. Experimental results using different $c$ , $r$ and $\alpha_{d}^{\max}$ values in Hyp-SimGCD with DINO [7] pre-trained backbone. Results on the CUB [54] and CIFAR-100 [34] datasets are reported. + +
parameterStanford-Cars [33]CIFAR-100 [34]
AllOldNewAllOldNew
c = 0.0161.474.455.182.483.181.2
c = 0.0562.873.457.781.684.076.7
c = 0.162.375.156.181.182.378.8
r = 1.060.072.953.782.483.181.2
r = 1.561.275.754.281.282.478.8
r = 2.362.873.457.780.181.178.3
αdmax= 0.159.677.551.081.382.479.1
αdmax= 0.562.077.254.682.483.181.2
αdmax= 1.062.873.457.778.983.569.7
+ +ods. Intuitively, as the value of $c$ approaches 0, the radius tends toward infinity, causing the Poincaré ball to flatten and resemble Euclidean space; conversely, larger values of $c$ correspond to a steeper configuration. The widely accepted range for $c$ is between 0.01 and 0.3 [15], with larger values exceeding this range resulting in performance degradation. In our experiments, we evaluate different curvature values of 0.01, 0.05, and 0.1 using Hyp-SimGCD, as presented in Tab. 2. Our findings indicate that the optimal curvature values differ between generic and fine-grained datasets. For fine-grained datasets such as CUB [54], the optimal value is 0.05, while for generic datasets like CIFAR-100, a value of 0.01 proves to be more effective. + +Clipping Value. As articulated in [23], feature clipping has emerged as a standard technique for training hyperbolic neural networks. In our framework, we also observe that it plays a crucial role in category discovery performance. In line with the methodology outlined in [23], we investigate a range of clipping values, specifically 1.0, 1.5, and 2.3. The results shown in the second row of Tab. 2 demonstrate that optimal clipping values vary between fine-grained and generic datasets. For fine-grained datasets like CUB [54], the optimal clipping value is determined to be 2.3. Conversely, for generic datasets like CIFAR-100 [34], a clipping value of 1.0 is shown to be more effective. + +Loss Weight. As detailed in Sec. 3.4, we implement a hybrid contrastive loss that combines both distance-based and angle-based components, which is essential for effective optimization in hyperbolic space. A loss weight, denoted as $\alpha_{d}$ , is introduced to regulate the balance between these two types of losses and linearly increasing from 0. In the initial stages, the model prioritizes optimizing the angle between sample points and progressively shifts focus toward optimizing the hyperbolic distance. Consistent with the observations for curvature and clipping values, the optimal max value $\alpha_{d}^{\max}$ , varies considerably between coarse-grained and fine-grained datasets. For fine-grained datasets such as CUB [54], an optimal value of 1.0 is observed, whereas for more generic datasets like CIFAR-100 [34], a + +![](images/8d885524221404d04d0b3ce20ff4b5ee561b5084e32d531a58210487e8c3e8c0.jpg) +Figure 5. T-SNE [49] comparison between SimGCD [59] and our Hyp-SimGCD using 40 randomly sampled instances from 10 randomly selected categories of the Stanford-Cars dataset [33]. + +value of 0.5 is found to yield better performance. + +# 4.4. Qualitative Comparison + +In Fig. 5, we present a t-SNE [49] visualization of features extracted from the backbone, represented as $\mathbf{z}_i = \phi(\mathbf{x}_i)$ . This visualization compares SimGCD with our Hyp-SimGCD. On the left side of the figure, the clusters generated by SimGCD appear dispersed. Data points from Class 42, highlighted in pink, are spread across multiple areas, indicating significant overlap and a lack of compactness. In contrast, Hyp-SimGCD creates more distinct and tightly clustered groups, concentrating the data points of Class 42 in a more confined area. This comparison implies that Hyp-SimGCD enhances both intra-class compactness and inter-class separation through our hyperbolic representation and classifier learning method. Importantly, even within the original Euclidean space of the backbone network, Hyp-SimGCD exhibits robust clustering performance, which arises from the properties of hyperbolic space in encoding hierarchical structures. + +# 5. Conclusion + +In this paper, we investigate a previously overlooked perspective in GCD by utilizing a representation space that captures the hierarchical structure of each sample, instead of the conventional Euclidean or spherical spaces. Our approach leverages the distinctive properties of hyperbolic space, where the volume increases exponentially with radius. This characteristic makes hyperbolic space especially suitable for modelling data possessing hierarchical structures, thereby enhancing representational capacity for category discovery. We propose a simple yet effective framework, HypCD, for integrating hyperbolic geometry into GCD methods. Through extensive experiments with parametric and non-parametric GCD baselines and the SOTA method, our framework consistently demonstrates superior performance on public benchmarks, underscoring the effectiveness of hyperbolic space for category discovery. + +Acknowledgements. This work is supported by National Natural Science Foundation of China (Grant No. 62306251), Hong Kong Research Grant Council - Early Career Scheme (Grant No. 27208022) and General Research Fund (Grant No. 17211024), and HKU Seed Fund for Basic Research. + +# References + +[1] Mina Ghadimi Atigh, Julian Schoep, Erman Acar, Nanne Van Noord, and Pascal Mettes. Hyperbolic image segmentation. In CVPR, 2022. 7 +[2] Anwesha Banerjee, Liyana Sahir Kallooriyakath, and Soma Biswas. Amend: Adaptive margin and expanded neighborhood for efficient generalized category discovery. In WACV, 2024. 6, 7 +[3] Ahmad Bdeir, Kristian Schwethelm, and Niels Landwehr. Fully hyperbolic convolutional neural networks for computer vision. In ICLR, 2024. 3 +[4] Gary Becigneul and Octavian-Eugen Ganea. Riemannian adaptive optimization methods. In ICLR, 2019. 6 +[5] James W Cannon, William J Floyd, Richard Kenyon, Walter R Parry, et al. Hyperbolic geometry. Flavors of geometry, 31(59-115):2, 1997. 4 +[6] Kaidi Cao, Maria Brbic, and Jure Leskovec. Open-world semi-supervised learning. In ICLR, 2022. 3, 6, 7 +[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 1, 2, 3, 4, 6, 8 +[8] Fernando Julio Cendra, Bingchen Zhao, and Kai Han. Promptcd: Learning gaussian mixture prompt pool for continual category discovery. In ECCV, 2024. 3 +[9] Ines Chami, Rex Ying, Christopher Ré, and Jure Leskovec. Hyperbolic graph convolutional neural networks. In NeurIPS, 2019. 3 +[10] Weize Chen, Xu Han, Yankai Lin, Hexu Zhao, Zhiyuan Liu, Peng Li, Maosong Sun, and Jie Zhou. Fully hyperbolic neural networks. In ACL, 2022. 2, 6 +[11] Florent Chiaroni, Jose Dolz, Ziko Imtiaz Masud, Amar Mitiche, and Ismail Ben Ayed. Parametric information maximization for generalized category discovery. In ICCV, 2023. 7 +[12] Sua Choi, Dahiun Kang, and Minsu Cho. Contrastive mean-shift learning for generalized category discovery. In CVPR, 2024. 6, 7 +[13] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 2, 6, 7 +[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 3 +[15] Aleksandr Ermolov, Leyla Mirvakhabova, Valentin Khrulkov, Nicu Sebe, and Ivan Oseledets. Hyperbolic vision transformers: Combining improvements in metric learning. In CVPR, 2022. 2, 3, 4, 8 + +[16] Yixin Fei, Zhongkai Zhao, Siwei Yang, and Bingchen Zhao. Xcon: Learning with experts for fine-grained category discovery. In BMVC, 2022. 6, 7 +[17] Enrico Fini, Enver Sangineto, Stéphane Lathuiliere, Zhun Zhong, Moin Nabi, and Elisa Ricci. A unified objective for novel class discovery. In ICCV, 2021. 3 +[18] Luca Franco, Paolo Mandica, Bharti Munjal, and Fabio Galasso. Hyperbolic self-paced learning for self-supervised skeleton-based action representations. In ICLR, 2023. 3, 5 +[19] Octavian Ganea, Gary Bécigneul, and Thomas Hofmann. Hyperbolic neural networks. In NeurIPS, 2018. 5 +[20] Zhi Gao, Yuwei Wu, Yunde Jia, and Mehrtash Harandi. Curvature generation in curved spaces for few-shot learning. In ICCV, 2021. 3 +[21] Mina GhadimiAtigh, Julian Schoep, Erman Acar, Nanne van Noord, and Pascal Mettes. Hyperbolic image segmentation. In CVPR, 2022. 3 +[22] Caglar Gulcehre, Misha Denil, Mateusz Malinowski, Ali Razavi, Razvan Pascanu, Karl Moritz Hermann, Peter Battaglia, Victor Bapst, David Raposo, Adam Santoro, and Nando de Freitas. Hyperbolic attention networks. In ICLR, 2019. 3 +[23] Yunhui Guo, Xudong Wang, Yubei Chen, and Stella X. Yu. Clipped hyperbolic classifiers are super-hyperbolic classifiers. In CVPR, 2022. 3, 4, 8 +[24] Kai Han, Andrea Vedaldi, and Andrew Zisserman. Learning to discover novel visual categories via deep transfer clustering. In ICCV, 2019. 1, 2, 3 +[25] Kai Han, Sylvestre-Alvise Rebuffi, Sebastien Ehrhardt, Andrea Vedaldi, and Andrew Zisserman. Automatically discovering and learning new visual categories with ranking statistics. In ICLR, 2020. 3 +[26] Kai Han, Sylvestre-Alvise Rebuffi, Sebastien Ehrhardt, Andrea Vedaldi, and Andrew Zisserman. Autonovel: Automatically discovering and learning novel visual categories. IEEE TPAMI, 2021. 3 +[27] Shaozhe Hao, Kai Han, and Kwan-Yee K Wong. Cipr: An efficient framework with cross-instance positive relations for generalized category discovery. TMLR, 2024. 1, 2, 3, 6, 7 +[28] Xuhui Jia, Kai Han, Yukun Zhu, and Bradley Green. Joint representation learning and novel category discovery on single-and multi-modal data. In ICCV, 2021. 3 +[29] Justin Johnson, Ranjay Krishna, Michael Stark, Li-Jia Li, David Shamma, Michael Bernstein, and Li Fei-Fei. Image retrieval using scene graphs. In CVPR, 2015. 2 +[30] KJ Joseph, Sujoy Paul, Gaurav Aggarwal, Soma Biswas, Piyush Rai, Kai Han, and Vineeth N Balasubramanian. Novel class discovery without forgetting. In ECCV, 2022. 3 +[31] Valentin Khrulkov, Leyla Mirvakhabova, Evgeniya Ustinova, Ivan Oseledets, and Victor Lempitsky. Hyperbolic image embeddings. In CVPR, 2020. 2, 3, 4, 7 +[32] Fanjie Kong, Yanbei Chen, Jiarui Cai, and Davide Modolo. Hyperbolic learning with synthetic captions for open-world detection. In CVPR, 2024. 2 +[33] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In ICCV workshop, 2013. 2, 6, 7, 8 + +[34] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 2, 6, 7, 8 +[35] Huimin Li, Zhentao Chen, Yunhao Xu, and Junlin Hu. Hyperbolic anomaly detection. In CVPR, 2024. 2 +[36] Qi Liu, Maximilian Nickel, and Douwe Kiela. Hyperbolic graph neural networks. In NeurIPS, 2019. 3 +[37] Yuanpei Liu and Kai Han. Debgcd: Debiased learning with distribution guidance for generalized category discovery. In ICLR, 2025. 1, 3 +[38] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 6, 7 +[39] Maximilian Nickel and Douwe Kiela. Poincaré embeddings for learning hierarchical representations. In NeurIPS, 2017. 4 +[40] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 4, 6 +[41] Jona Otholt, Christoph Meinel, and Haojin Yang. Guided cluster aggregation: A hierarchical approach to generalized category discovery. In WACV, pages 2618-2627, 2024. 2 +[42] Nan Pu, Zhun Zhong, and Nicu Sebe. Dynamic conceptions of contrastive learning for generalized category discovery. In CVPR, 2023. 3, 6, 7 +[43] Nan Pu, Wenjing Li, Xingyuan Ji, Yalan Qin, Nicu Sebe, and Zhun Zhong. Federated generalized category discovery. In CVPR, 2024. 3 +[44] Sarah Rastegar, Hazel Doughty, and Cees Snoek. Learn to categorize or categorize to learn? self-coding for generalized category discovery. In NeurIPS, 2023. 1, 2, 4, 6, 7 +[45] Sarah Rastegar, Mohammadreza Salehi, Yuki M Asano, Hazel Doughty, and Cees G M Snoek. Selex: Self-expertise in fine-grained generalized category discovery. In ECCV, 2024. 1, 2, 3, 4, 5, 6, 7 +[46] Ryohei Shimizu, Yusuke Mukuta, and Tatsuya Harada. Hyperbolic neural networks++. In *ICLR*, 2021. 3 +[47] Yiyou Sun and Yixuan Li. Opencon: Open-world contrastive learning. TMLR, 2022. 6, 7 +[48] Abraham Albert Ungar. A gyrovector space approach to hyperbolic geometry. Synthesis Lectures on Mathematics and Statistics, 1(1):1-194, 2008. 4 +[49] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. JMLR, 2008. 8 +[50] Sagar Vaze, Kai Han, Andrea Vedaldi, and Andrew Zisserman. Open-set recognition: A good closed-set classifier is all you need? In ICLR, 2022. 2 +[51] Sagar Vaze, Kai Han, Andrea Vedaldi, and Andrew Zisserman. Generalized category discovery. In CVPR, 2022. 1, 2, 3, 4, 5, 6, 7 +[52] Sagar Vaze, Kai Han, Andrea Vedaldi, and Andrew Zisserman. The semantic shift benchmark. In ICML workshop, 2022. 1, 6, 7 +[53] Sagar Vaze, Andrea Vedaldi, and Andrew Zisserman. No representation rules them all in category discovery. In NeurIPS, 2023. 3, 6, 7 + +[54] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 2, 6, 7, 8 +[55] Hongjun Wang, Sagar Vaze, and Kai Han. Sptnet: An efficient alternative framework for generalized category discovery with spatial prompt tuning. In ICLR, 2024. 1, 2, 3, 5, 6, 7 +[56] Hongjun Wang, Sagar Vaze, and Kai Han. Hilo: A learning framework for generalized category discovery robust to domain shifts. In ICLR, 2025. 3 +[57] Yu Wang, Zhun Zhong, Pengchong Qiao, Xuxin Cheng, Xiawu Zheng, Chang Liu, Nicu Sebe, Rongrong Ji, and Jie Chen. Discover and align taxonomic context priors for open-world semi-supervised learning. In NeurIPS, 2024. 2 +[58] Simon Weber, Bar Zongür, Nikita Araslanov, and Daniel Cremers. Flattening the parent bias: Hierarchical semantic segmentation in the poincaré ball. In CVPR, 2024. 2 +[59] Xin Wen, Bingchen Zhao, and Xiaojuan Qi. Parametric classification for generalized category discovery: A baseline study. In ICCV, 2023. 1, 2, 3, 4, 5, 6, 7, 8 +[60] Zhenzhen Weng, Mehmet Giray Ogut, Shai Limonchik, and Serena Yeung. Unsupervised discovery of the long-tail in instance segmentation using hierarchical self-supervision. In CVPR, 2021. 3 +[61] Sheng Zhang, Salman Khan, Zhiqiang Shen, Muzammal Naseer, Guangyi Chen, and Fahad Shahbaz Khan. Promptable: Contrastive affinity learning via auxiliary prompts for generalized novel category discovery. In CVPR, 2023. 6, 7 +[62] Xinwei Zhang, Jianwen Jiang, Yutong Feng, Zhi-fan Wu, Xibin Zhao, Hai Wan, Mingqian Tang, Rong Jin, and Yue Gao. Grow and merge: a unified framework for continuous categories discovery. In NeurIPS, 2022. 3 +[63] Bingchen Zhao and Kai Han. Novel visual category discovery with dual ranking statistics and mutual knowledge distillation. In NeurIPS, 2021. 1, 3 +[64] Bingchen Zhao, Xin Wen, and Kai Han. Learning semi-supervised gaussian mixture models for generalized category discovery. In ICCV, 2023. 3, 6, 7 +[65] Zhun Zhong, Enrico Fini, Subhankar Roy, Zhiming Luo, Elisa Ricci, and Nicu Sebe. Neighborhood contrastive learning for novel class discovery. In CVPR, 2021. 3 \ No newline at end of file diff --git a/2025/Hyperbolic Category Discovery/images.zip b/2025/Hyperbolic Category Discovery/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5de89ef73ffc5e38631d92044338e9fa8af7c882 --- /dev/null +++ b/2025/Hyperbolic Category Discovery/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:796f622c41fca24917361897fe4567040c374d33dbdb2e72c841572df86ded7f +size 597990 diff --git a/2025/Hyperbolic Category Discovery/layout.json b/2025/Hyperbolic Category Discovery/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..bfd69a64f8a51147c4f48164e09e35b45ce1d635 --- /dev/null +++ b/2025/Hyperbolic Category Discovery/layout.json @@ -0,0 +1,9883 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 207, + 103, + 403, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 103, + 403, + 121 + ], + "spans": [ + { + "bbox": [ + 207, + 103, + 403, + 121 + ], + "type": "text", + "content": "Hyperbolic Category Discovery" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 195, + 142, + 414, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 142, + 414, + 158 + ], + "spans": [ + { + "bbox": [ + 195, + 142, + 414, + 158 + ], + "type": "text", + "content": "Yuanpei Liu* Zhenqi He* Kai Han†" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 195, + 158, + 414, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 158, + 414, + 172 + ], + "spans": [ + { + "bbox": [ + 195, + 158, + 414, + 172 + ], + "type": "text", + "content": "Visual AI Lab, The University of Hong Kong" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 168, + 174, + 443, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 174, + 443, + 185 + ], + "spans": [ + { + "bbox": [ + 168, + 174, + 443, + 185 + ], + "type": "text", + "content": "{ypliu0,zhenqi_he}@connect.hku.hk kaihanx@hku.hk" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 238, + 297, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 238, + 297, + 585 + ], + "spans": [ + { + "bbox": [ + 53, + 238, + 297, + 585 + ], + "type": "text", + "content": "Generalized Category Discovery (GCD) is an intriguing open-world problem that has garnered increasing attention. Given a dataset that includes both labelled and unlabelled images, GCD aims to categorize all images in the unlabelled subset, regardless of whether they belong to known or unknown classes. In GCD, the common practice typically involves applying a spherical projection operator at the end of the self-supervised pretrained backbone, operating within Euclidean or spherical space. However, both of these spaces have been shown to be suboptimal for encoding samples that possesses hierarchical structures. In contrast, hyperbolic space exhibits exponential volume growth relative to radius, making it inherently strong at capturing the hierarchical structure of samples from both seen and unseen categories. Therefore, we propose to tackle the category discovery challenge in the hyperbolic space. We introduce HypCD, a simple Hyperbolic framework for learning hierarchy-aware representations and classifiers for generalized Category Discovery. HypCD first transforms the Euclidean embedding space of the backbone network into hyperbolic space, facilitating subsequent representation and classification learning by considering both hyperbolic distance and the angle between samples. This approach is particularly helpful for knowledge transfer from known to unknown categories in GCD. We thoroughly evaluate HypCD on public GCD benchmarks, by applying it to various baseline and state-of-the-art methods, consistently achieving significant improvements. Project page: https://visual-ai.github.io/hypcd/" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 606, + 135, + 619 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 606, + 135, + 619 + ], + "spans": [ + { + "bbox": [ + 56, + 606, + 135, + 619 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 626, + 296, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 626, + 296, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 626, + 296, + 687 + ], + "type": "text", + "content": "Recently, category discovery – initially explored as novel category discovery (NCD) [24] and subsequently extended to generalized category discovery (GCD) [51] – has emerged as an intriguing open-world problem, gaining increasing attention. GCD addresses the challenges posed" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 213, + 552, + 330 + ], + "blocks": [ + { + "bbox": [ + 317, + 213, + 552, + 330 + ], + "lines": [ + { + "bbox": [ + 317, + 213, + 552, + 330 + ], + "spans": [ + { + "bbox": [ + 317, + 213, + 552, + 330 + ], + "type": "image", + "image_path": "416dc20d2276f8a148df659ee60187b9a177c9883a5d60fe24c761b343233734.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 318, + 333, + 552, + 449 + ], + "blocks": [ + { + "bbox": [ + 318, + 333, + 552, + 449 + ], + "lines": [ + { + "bbox": [ + 318, + 333, + 552, + 449 + ], + "spans": [ + { + "bbox": [ + 318, + 333, + 552, + 449 + ], + "type": "image", + "image_path": "95e747909c7a03310812b638ffe4b705297935e68713167428b247f101f6ffed.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 450, + 555, + 506 + ], + "lines": [ + { + "bbox": [ + 313, + 450, + 555, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 555, + 506 + ], + "type": "text", + "content": "Figure 1. (a) Spherical-based vs. Hyperbolic-based methods, where hyperbolic space better accommodates variations in scale and improves connections between samples. (b) Average ACC comparison of our method and previous SOTA across 'All', 'Old', and 'New' categories on the SSB [52] benchmark using DINO [7]." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 519, + 555, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 519, + 555, + 638 + ], + "spans": [ + { + "bbox": [ + 312, + 519, + 555, + 638 + ], + "type": "text", + "content": "by partially labelled datasets, where the unlabelled subset may contain instances from both seen and unseen classes. The goal is to leverage knowledge from labelled data to effectively categorize the unlabelled data. Based on the way to predict category index, existing GCD methods can be broadly classified into two types: non-parametric methods [27, 44, 45, 51] and parametric methods [37, 55, 59]. Non-parametric methods predict category index based on feature clustering while parametric methods utilize a parametric classifier." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "content": "As shown in the GCD literature [51], object parts are effective for knowledge transfer from seen to unseen categories, which is crucial for novel category discovery. Methods have been developed to explicitly learn better local features by learning pixel-level prompts around local image regions [55] or utilizing part-level features [63]. However," + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "9891" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 201, + 151 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 201, + 151 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 201, + 151 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 201, + 151 + ], + "type": "image", + "image_path": "8dd1299c9dfe5ab718c05656990a2dab7d455edcb9c7b884c1663db1d255da64.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 89, + 152, + 167, + 161 + ], + "lines": [ + { + "bbox": [ + 89, + 152, + 167, + 161 + ], + "spans": [ + { + "bbox": [ + 89, + 152, + 167, + 161 + ], + "type": "text", + "content": "(a) Inter-Category Relation" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 206, + 72, + 288, + 152 + ], + "blocks": [ + { + "bbox": [ + 206, + 72, + 288, + 152 + ], + "lines": [ + { + "bbox": [ + 206, + 72, + 288, + 152 + ], + "spans": [ + { + "bbox": [ + 206, + 72, + 288, + 152 + ], + "type": "image", + "image_path": "5279474e7184cab0d740b639eb6e62a1756d61fae0bcc1f1a11522bb04a2083c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 209, + 152, + 286, + 161 + ], + "lines": [ + { + "bbox": [ + 209, + 152, + 286, + 161 + ], + "spans": [ + { + "bbox": [ + 209, + 152, + 286, + 161 + ], + "type": "text", + "content": "(b) Intra-Category Relation" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 163, + 295, + 196 + ], + "lines": [ + { + "bbox": [ + 55, + 163, + 295, + 196 + ], + "spans": [ + { + "bbox": [ + 55, + 163, + 295, + 196 + ], + "type": "text", + "content": "Figure 2. Hierarchical relations in GCD. (a) Inter-category relationships within the Stanford-Cars dataset [33]. (b) Intra-category relationships within CUB [54] dataset." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 296, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 296, + 616 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 296, + 616 + ], + "type": "text", + "content": "these methods consider object parts as rigid image patches of the same size, without considering the hierarchical nature of the object parts and the scale discrepancy of the same parts in different images, thus unavoidably restricting the performance for GCD (see Fig. 1(a)), in which the objects often have distinct poses, scales and appearance. To address this problem, one possible solution is to learn image embeddings possessing hierarchical constraints or following tree-like structures. This has been proven to be effective in many tasks. For example, in image retrieval and clustering, the hierarchy constraint may arise from whole-fragment relation [29, 31]. Intuitively, in category discovery, which can be regarded as a transfer clustering task [24], we hypothesize that an embedding space that captures the hierarchical relations of object parts can also facilitate the discovery of new categories. Indeed, the hierarchical relations have been studied in GCD, such as [27, 41, 44, 45, 57]. However, they study from a substantially different perspective: inter-category hierarchy (see Fig. 2(a)). This only considers the hierarchy of different semantic classes from coarse to fine levels. Additionally, these methods require the relationships and number of levels in the hierarchy to be predefined, resulting in a lack of flexibility and scalability. Moreover, these methods are unable to capture more complex hierarchies, such as the compositional parts of an object (see Fig. 2(b)). This is particularly the case because existing methods [45, 51, 55, 59], no matter whether they consider the hierarchy or not, learn the image embeddings in a spherical space. This follows the common practice of applying a spherical projection operator at the end of the self-supervised feature backbone [7, 40]. Consequently, all subsequent operations, including distance calculations, are performed under either Euclidean or spherical geometry, resulting in limited awareness of hierarchical object parts." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "content": "In this work, we study the overlooked perspective in category discovery: instead of learning in the Euclidean or spherical space, we advocate a space that captures the hierarchical structure of each data point. In spherical space, both the radius and volume are constant, whereas in Euclidean space, the volume grows polynomially with respect to the radius. Both of these spaces have been shown to be suboptimal for encoding samples that possess hierarchical" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "content": "structures [10, 15, 31]. In contrast, hyperbolic space possesses a distinctive property where its volume grows exponentially relative to the radius. This characteristic makes hyperbolic space particularly suitable for embedding tree-like data, enhancing its representational power. Learning representations in hyperbolic space has proven to be effective in various computer vision tasks, including object recognition [15], object detection [32], semantic segmentation [58], and anomaly detection [35]. Inspired by these successes, we aim to realize the idea of learning hierarchy-aware representations to facilitate knowledge transfer in category discovery, thereby unleashing the potential of hyperbolic representations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 228, + 553, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 553, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 553, + 538 + ], + "type": "text", + "content": "To achieve this goal, we propose a simple yet effective framework, HypCD, to properly learn the hierarchy-aware representation and classifier for category discovery through the lens of hyperbolic geometry. In this framework, we adapt our framework to popular parametric [59] and non-parametric [51] GCD baselines as well as the state-of-the-art (SOTA) method SelEx [45], obtaining substantial improvements for them, establishing the new SOTA (see Fig. 1(b)). Firstly, starting from the self-supervised backbone pretrained in Euclidean space, we propose to map the Euclidean representation to a constrained Poincaré ball through feature clipping and exponential mapping. Secondly, we implement the hyperbolic representation learning and build a hyperbolic classifier on the Poincaré ball, considering both angle and distance between samples in hyperbolic space. Thirdly, to assign labels to unlabelled data after training, for non-parametric methods, we apply semi-supervised " + }, + { + "bbox": [ + 313, + 228, + 553, + 538 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 228, + 553, + 538 + ], + "type": "text", + "content": "-means following the common practice; for parametric methods, we employ a hyperbolic classifier to make predictions. Despite its simplicity, our framework achieves significant performance improvements with two different pretrained weights (DINO [7] and DINOv2 [40]) on the public GCD datasets, including the coarse-grained classification datasets CIFAR-10 [34], CIFAR-100 [34], and ImageNet-100 [13], as well as the fine-grained SSB [50] benchmark." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 539, + 554, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 539, + 554, + 670 + ], + "spans": [ + { + "bbox": [ + 313, + 539, + 554, + 670 + ], + "type": "text", + "content": "In summary, we make the following contributions in this paper: (i) We identify the existing GCD methods' common shortcoming in encoding the hierarchical structure and propose to incorporate the hyperbolic geometry into the embedding space to address this limitation; (ii) We propose a simple yet effective framework, called HypCD, for incorporating the hyperbolic geometry in representation learning and classification for category discovery; and (iii) Through extensive experiments on public GCD benchmarks by applying HypCD to baseline and SOTA methods, our method consistently demonstrates effectiveness and superiority." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 681, + 400, + 693 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 681, + 400, + 693 + ], + "spans": [ + { + "bbox": [ + 314, + 681, + 400, + 693 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "type": "text", + "content": "Category Discovery. Novel category discovery (NCD) is" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9892" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 395 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 395 + ], + "type": "text", + "content": "initially introduced in [24] to establish a realistic framework for transferring knowledge from seen categories to cluster unseen categories, by considering it as a transfer clustering problem. Many subsequent methods have been proposed to advance the field [17, 25, 26, 28, 63, 65]. Generalized category discovery [51] (GCD) relaxes the assumption in NCD by considering unlabelled data containing samples from both known and unknown classes [51]. Further investigations [6, 8, 27, 30, 37, 42, 56] have explored a variety of strategies to address the challenges posed by GCD. One notable approach, SimGCD [59], proposes to learn a parametric classifier enhanced by mean entropy regularization, thereby improving performance. In another vein, GPC [64] employs Gaussian mixture models to jointly learn robust representations while simultaneously estimating the number of unknown categories. SPT-Net [55] presents a spatial prompt tuning method that enables models to concentrate more effectively on specific object parts, thus enhancing knowledge transfer in GCD. Most recently, SelEx [45] has been proposed, leveraging hierarchical semi-supervised " + }, + { + "bbox": [ + 55, + 72, + 296, + 395 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 72, + 296, + 395 + ], + "type": "text", + "content": "-means to achieve SOTA results on fine-grained datasets. Additionally, various efforts are focused on addressing category discovery from multiple perspectives. For instance, [28] emphasizes multi-modal category discovery; [62] and [8] explore a continual setting; [43] studies category discovery in a federated setting; and [56] examines GCD in the presence of domain shifts." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 396, + 296, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 296, + 624 + ], + "type": "text", + "content": "Hyperbolic Geometry. Hyperbolic space, defined as a non-Euclidean manifold with exponential volume growth in relation to its radius, is inherently aligned with the embedding of tree-like and hierarchical data structures in visual recognition tasks. Significant advancements in this area include [31], which presents a hyperbolic image embedding technique by projecting model outputs into hyperbolic space, and [15], which integrates hyperbolic geometry into various vision transformer architectures, showcasing performance that surpasses their Euclidean counterparts. Hyperbolic methods have also been developed on diverse tasks such as image classification [15, 23, 31], action recognition [18], few-shot learning [20] and object segmentation [21, 60]. Moreover, recent developments have introduced hyperbolic geometry for neural networks including fully connected layers [46], convolutional neural networks [3], graph neural networks [9, 36], and attention network [22], thereby facilitating a deeper integration of hyperbolic geometry into deep learning regime." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 633, + 111, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 633, + 111, + 645 + ], + "spans": [ + { + "bbox": [ + 55, + 633, + 111, + 645 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": "In this section, we first introduce the task in Sec. 3.1, then move to a review of baselines in Sec. 3.2. Afterwards, the geometry mapping and training details of our framework are described in Sec. 3.3 and Sec. 3.4. Lastly, the label assignment details are outlined in Sec. 3.5." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 314, + 72, + 427, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 427, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 427, + 83 + ], + "type": "text", + "content": "3.1. Problem Statement" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 89, + 555, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 555, + 257 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 555, + 257 + ], + "type": "text", + "content": "GCD aims to learn a model capable of accurately classifying unlabelled samples from known categories while simultaneously clustering those from unknown categories. Consider an unlabelled dataset denoted as " + }, + { + "bbox": [ + 313, + 89, + 555, + 257 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_u = \\left\\{\\left(\\mathbf{x}_i^u,y_i^u\\right)\\right\\} \\in" + }, + { + "bbox": [ + 313, + 89, + 555, + 257 + ], + "type": "inline_equation", + "content": "\\mathbf{X}\\times \\mathbf{Y}_u" + }, + { + "bbox": [ + 313, + 89, + 555, + 257 + ], + "type": "text", + "content": " and a labelled dataset represented as " + }, + { + "bbox": [ + 313, + 89, + 555, + 257 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_l =" + }, + { + "bbox": [ + 313, + 89, + 555, + 257 + ], + "type": "inline_equation", + "content": "\\left\\{\\left(\\mathbf{x}_i^l,y_i^l\\right)\\right\\} \\in \\mathbf{X}\\times \\mathbf{Y}_l," + }, + { + "bbox": [ + 313, + 89, + 555, + 257 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 89, + 555, + 257 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_u" + }, + { + "bbox": [ + 313, + 89, + 555, + 257 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 89, + 555, + 257 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_l" + }, + { + "bbox": [ + 313, + 89, + 555, + 257 + ], + "type": "text", + "content": " denote the respective label sets. The unlabelled dataset comprises samples from both known and unknown categories, i.e., specifically " + }, + { + "bbox": [ + 313, + 89, + 555, + 257 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_l\\subset \\mathbf{Y}_u" + }, + { + "bbox": [ + 313, + 89, + 555, + 257 + ], + "type": "text", + "content": " . Let the number of labelled categories be denoted by " + }, + { + "bbox": [ + 313, + 89, + 555, + 257 + ], + "type": "inline_equation", + "content": "M = |\\mathbf{Y}_l|" + }, + { + "bbox": [ + 313, + 89, + 555, + 257 + ], + "type": "text", + "content": " . We assume that the total number of categories, " + }, + { + "bbox": [ + 313, + 89, + 555, + 257 + ], + "type": "inline_equation", + "content": "K = |\\mathbf{Y}_l\\cup \\mathbf{Y}_u|" + }, + { + "bbox": [ + 313, + 89, + 555, + 257 + ], + "type": "text", + "content": " , is known, as posited in prior works [26, 53, 59]. In scenarios where this information is unavailable, alternative methods such as those proposed in [24, 51] can be employed to yield a reliable estimation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 263, + 429, + 275 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 263, + 429, + 275 + ], + "spans": [ + { + "bbox": [ + 314, + 263, + 429, + 275 + ], + "type": "text", + "content": "3.2. Review of Baselines" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 281, + 555, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 555, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 555, + 389 + ], + "type": "text", + "content": "Non-parametric Baseline. [51] formalizes the GCD task and proposes a non-parametric baseline. The approach involves finetuning the pre-trained DINO [7] model [14] to enhance the learned representation. The loss function comprises a supervised contrastive loss, which operates on the labelled samples, and a self-supervised contrastive loss, which operates on all the samples. Specifically, given two randomly augmented views " + }, + { + "bbox": [ + 313, + 281, + 555, + 389 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 313, + 281, + 555, + 389 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 281, + 555, + 389 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i'" + }, + { + "bbox": [ + 313, + 281, + 555, + 389 + ], + "type": "text", + "content": " for the same image in a mini-batch " + }, + { + "bbox": [ + 313, + 281, + 555, + 389 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 281, + 555, + 389 + ], + "type": "text", + "content": ", the self-supervised contrastive loss is:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 347, + 396, + 555, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 396, + 555, + 418 + ], + "spans": [ + { + "bbox": [ + 347, + 396, + 555, + 418 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r e p} ^ {u} = \\frac {1}{| B |} \\sum_ {i \\in B} - \\log \\frac {\\exp \\left(\\mathbf {z} _ {i} \\cdot \\mathbf {z} _ {i} ^ {\\prime} / \\tau_ {r}\\right)}{\\sum_ {j} ^ {j \\neq i} \\exp \\left(\\mathbf {z} _ {i} \\cdot \\mathbf {z} _ {j} ^ {\\prime} / \\tau_ {r}\\right)}, \\tag {1}", + "image_path": "7c5e2a5423e023095b79fa245dead1f7b273b5615eaffd27e4016b766f69a3f6.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 424, + 554, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 424, + 554, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 424, + 554, + 496 + ], + "type": "text", + "content": "where the feature " + }, + { + "bbox": [ + 313, + 424, + 554, + 496 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i = \\rho_r(\\phi (\\mathbf{x}_i))" + }, + { + "bbox": [ + 313, + 424, + 554, + 496 + ], + "type": "text", + "content": " is a " + }, + { + "bbox": [ + 313, + 424, + 554, + 496 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 313, + 424, + 554, + 496 + ], + "type": "text", + "content": "-normalized vector and " + }, + { + "bbox": [ + 313, + 424, + 554, + 496 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i^\\prime" + }, + { + "bbox": [ + 313, + 424, + 554, + 496 + ], + "type": "text", + "content": " represents feature from another view " + }, + { + "bbox": [ + 313, + 424, + 554, + 496 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i^\\prime" + }, + { + "bbox": [ + 313, + 424, + 554, + 496 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 313, + 424, + 554, + 496 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 424, + 554, + 496 + ], + "type": "text", + "content": " refers to the backbone network, " + }, + { + "bbox": [ + 313, + 424, + 554, + 496 + ], + "type": "inline_equation", + "content": "\\rho_r" + }, + { + "bbox": [ + 313, + 424, + 554, + 496 + ], + "type": "text", + "content": " denotes the projection head, and " + }, + { + "bbox": [ + 313, + 424, + 554, + 496 + ], + "type": "inline_equation", + "content": "\\tau_{r}" + }, + { + "bbox": [ + 313, + 424, + 554, + 496 + ], + "type": "text", + "content": " stands as the temperature parameter used for scaling the features. The supervised contrastive loss for labelled samples is:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 319, + 503, + 553, + 527 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 503, + 553, + 527 + ], + "spans": [ + { + "bbox": [ + 319, + 503, + 553, + 527 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r e p} ^ {s} = \\frac {1}{| B _ {l} |} \\sum_ {i \\in B _ {l}} \\frac {1}{| N _ {i} |} \\sum_ {q \\in N _ {i}} - \\log \\frac {\\exp \\left(\\mathbf {z} _ {i} \\cdot \\mathbf {z} _ {q} / \\tau_ {r}\\right)}{\\sum_ {j} ^ {j \\neq i} \\exp \\left(\\mathbf {z} _ {i} \\cdot \\mathbf {z} _ {j} / \\tau_ {r}\\right)}, \\tag {2}", + "image_path": "93095dca5f8149595cb7b1e7f5b396c73756b76fe62a57cd4a53725363d18f5e.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 533, + 555, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 555, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 555, + 582 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 533, + 555, + 582 + ], + "type": "inline_equation", + "content": "N_{i}" + }, + { + "bbox": [ + 313, + 533, + 555, + 582 + ], + "type": "text", + "content": " is the index set for all other images in the labelled mini-batch " + }, + { + "bbox": [ + 313, + 533, + 555, + 582 + ], + "type": "inline_equation", + "content": "B_{l} \\subset B" + }, + { + "bbox": [ + 313, + 533, + 555, + 582 + ], + "type": "text", + "content": " having the same label as " + }, + { + "bbox": [ + 313, + 533, + 555, + 582 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{i}" + }, + { + "bbox": [ + 313, + 533, + 555, + 582 + ], + "type": "text", + "content": ". The overall representation learning loss is then: " + }, + { + "bbox": [ + 313, + 533, + 555, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rep} = (1 - \\lambda_b)\\mathcal{L}_{rep}^u + \\lambda_b\\mathcal{L}_{rep}^s" + }, + { + "bbox": [ + 313, + 533, + 555, + 582 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 533, + 555, + 582 + ], + "type": "inline_equation", + "content": "\\lambda_{b}" + }, + { + "bbox": [ + 313, + 533, + 555, + 582 + ], + "type": "text", + "content": " is a balance factor." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 582, + 554, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 554, + 690 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 554, + 690 + ], + "type": "text", + "content": "Parametric Baseline. [59] introduces a robust parametric GCD baseline, which has been widely adopted in the field ever since [53, 55]. This method employs a parametric classifier implemented in a self-distillation framework [7]. The classifier is randomly initialized with " + }, + { + "bbox": [ + 313, + 582, + 554, + 690 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 582, + 554, + 690 + ], + "type": "text", + "content": " normalized category prototypes " + }, + { + "bbox": [ + 313, + 582, + 554, + 690 + ], + "type": "inline_equation", + "content": "\\mathbf{C} = \\{\\mathbf{c}_1,\\dots,\\mathbf{c}_K\\}" + }, + { + "bbox": [ + 313, + 582, + 554, + 690 + ], + "type": "text", + "content": ". For a randomly augmented view " + }, + { + "bbox": [ + 313, + 582, + 554, + 690 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 313, + 582, + 554, + 690 + ], + "type": "text", + "content": " and its corresponding normalized hidden feature vector " + }, + { + "bbox": [ + 313, + 582, + 554, + 690 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_i = \\phi (\\mathbf{x}_i) / ||\\phi (\\mathbf{x}_i)||" + }, + { + "bbox": [ + 313, + 582, + 554, + 690 + ], + "type": "text", + "content": ", the output probability for the " + }, + { + "bbox": [ + 313, + 582, + 554, + 690 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 582, + 554, + 690 + ], + "type": "text", + "content": "-th category is given by:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 377, + 696, + 553, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 696, + 553, + 716 + ], + "spans": [ + { + "bbox": [ + 377, + 696, + 553, + 716 + ], + "type": "interline_equation", + "content": "\\mathbf {p} _ {i} ^ {(k)} = \\frac {\\exp \\left(\\mathbf {h} _ {i} \\cdot \\mathbf {c} _ {k} / \\tau_ {s}\\right)}{\\sum_ {j = 1} ^ {K} \\exp \\left(\\mathbf {h} _ {i} \\cdot \\mathbf {c} _ {j} / \\tau_ {s}\\right)}, \\tag {3}", + "image_path": "6ecc3907b81514f66951101ed76441dd8b63b11702358cf9f6f83cb94a0f18d6.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9893" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "inline_equation", + "content": "\\tau_{s}" + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": " is the scaling temperature for the 'student' view. The soft label " + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_i" + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": " is generated by the teacher' view with a sharper temperature " + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "inline_equation", + "content": "\\tau_{t}" + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": " based on another augmented view in a similar manner. The self-distillation loss for the two views is then computed using the cross-entropy loss function: " + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "inline_equation", + "content": "\\ell_{ce}(\\mathbf{q}',\\mathbf{p}) = -\\sum_{j = 1}^{K}\\mathbf{q}'^{(j)}\\log \\mathbf{p}^{(j)}" + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": ". The unsupervised loss is then computed by aggregating contributions from all samples in the mini-batch " + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 94, + 177, + 295, + 194 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 177, + 295, + 194 + ], + "spans": [ + { + "bbox": [ + 94, + 177, + 295, + 194 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c l s} ^ {u} = \\frac {1}{| B |} \\sum_ {i \\in B} \\ell_ {c e} \\left(\\mathbf {q} _ {i} ^ {\\prime}, \\mathbf {p} _ {i}\\right) - \\xi \\mathcal {H} (\\bar {\\mathbf {p}}), \\tag {4}", + "image_path": "13ae2b984ccdd3fb32c5ff1bc52e96623a4217de38e8606561f25cd77bbd1fdb.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 204, + 295, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 204, + 295, + 242 + ], + "spans": [ + { + "bbox": [ + 55, + 204, + 295, + 242 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 204, + 295, + 242 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{p}} = \\frac{1}{2|B|}\\sum_{i\\in B}(\\mathbf{p}_i + \\mathbf{p}_i')" + }, + { + "bbox": [ + 55, + 204, + 295, + 242 + ], + "type": "text", + "content": " denotes the mean prediction across the mini-batch. The mean entropy is defined as: " + }, + { + "bbox": [ + 55, + 204, + 295, + 242 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(\\overline{\\mathbf{p}}) = -\\sum_{j = 1}^{K}\\overline{\\mathbf{p}}^{(j)}\\log \\overline{\\mathbf{p}}^{(j)}" + }, + { + "bbox": [ + 55, + 204, + 295, + 242 + ], + "type": "text", + "content": ", weighted by " + }, + { + "bbox": [ + 55, + 204, + 295, + 242 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 55, + 204, + 295, + 242 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 242, + 296, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 242, + 296, + 363 + ], + "spans": [ + { + "bbox": [ + 55, + 242, + 296, + 363 + ], + "type": "text", + "content": "For the labelled samples, the supervised classification loss is written as " + }, + { + "bbox": [ + 55, + 242, + 296, + 363 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cls}^{s} = \\frac{1}{|B_{l}|}\\sum_{i\\in B_{l}}\\ell_{ce}(\\mathbf{p}_{i},\\mathbf{y}_{i})" + }, + { + "bbox": [ + 55, + 242, + 296, + 363 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 242, + 296, + 363 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_i" + }, + { + "bbox": [ + 55, + 242, + 296, + 363 + ], + "type": "text", + "content": " represents the one-hot vector corresponding to the ground-truth label " + }, + { + "bbox": [ + 55, + 242, + 296, + 363 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 55, + 242, + 296, + 363 + ], + "type": "text", + "content": ". The overall objective is " + }, + { + "bbox": [ + 55, + 242, + 296, + 363 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cls} = (1 - \\lambda_b)\\mathcal{L}_{cls}^u + \\lambda_b\\mathcal{L}_{cls}^s" + }, + { + "bbox": [ + 55, + 242, + 296, + 363 + ], + "type": "text", + "content": ". Integrating this with the representation learning loss " + }, + { + "bbox": [ + 55, + 242, + 296, + 363 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rep}" + }, + { + "bbox": [ + 55, + 242, + 296, + 363 + ], + "type": "text", + "content": " adopted from [51], the comprehensive training objective is expressed as: " + }, + { + "bbox": [ + 55, + 242, + 296, + 363 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{gcd} = \\mathcal{L}_{cls} + \\mathcal{L}_{rep}" + }, + { + "bbox": [ + 55, + 242, + 296, + 363 + ], + "type": "text", + "content": ". Through training with " + }, + { + "bbox": [ + 55, + 242, + 296, + 363 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{gcd}" + }, + { + "bbox": [ + 55, + 242, + 296, + 363 + ], + "type": "text", + "content": " on both " + }, + { + "bbox": [ + 55, + 242, + 296, + 363 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_l" + }, + { + "bbox": [ + 55, + 242, + 296, + 363 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 242, + 296, + 363 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_u" + }, + { + "bbox": [ + 55, + 242, + 296, + 363 + ], + "type": "text", + "content": ", the classifier is empowered to directly predict labels for the unlabelled samples after the training process concludes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 371, + 272, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 371, + 272, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 371, + 272, + 384 + ], + "type": "text", + "content": "3.3. Hyperbolic Space for Category Discovery" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 388, + 296, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 388, + 296, + 532 + ], + "spans": [ + { + "bbox": [ + 54, + 388, + 296, + 532 + ], + "type": "text", + "content": "As previously discussed, object parts are critical for facilitating knowledge transfer from labelled categories to unseen ones in GCD. Each sample inherently contains object parts that reside within a hierarchical structure. Moreover, existing GCD methods [44, 45] emphasize the intercategory hierarchy to enhance the clustering performance of unlabelled samples in Euclidean or spherical spaces. However, the geometry of representation space limits their ability to effectively capture other kinds of hierarchy [31]. In contrast, hyperbolic space, characterized by its property of exponential volume growth with respect to the radius [15], emerges as a more suitable space for GCD." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 533, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 296, + 714 + ], + "type": "text", + "content": "Hyperbolic space " + }, + { + "bbox": [ + 55, + 533, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbb{H}^n" + }, + { + "bbox": [ + 55, + 533, + 296, + 714 + ], + "type": "text", + "content": " is defined as an " + }, + { + "bbox": [ + 55, + 533, + 296, + 714 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 533, + 296, + 714 + ], + "type": "text", + "content": "-dimensional Riemannian manifold exhibiting constant negative curvature, and it encompasses several analytic models [5]. Following previous literature [15, 31], we employ the Poincaré ball [39] model. In this model, the hyperbolic space is represented as an " + }, + { + "bbox": [ + 55, + 533, + 296, + 714 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 533, + 296, + 714 + ], + "type": "text", + "content": "-dimensional ball " + }, + { + "bbox": [ + 55, + 533, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_c^n = \\{\\mathbf{a}\\in \\mathbb{R}^n\\mid c\\| \\mathbf{a}\\| ^2 < 1\\}" + }, + { + "bbox": [ + 55, + 533, + 296, + 714 + ], + "type": "text", + "content": " with curvature value " + }, + { + "bbox": [ + 55, + 533, + 296, + 714 + ], + "type": "inline_equation", + "content": "-c^2" + }, + { + "bbox": [ + 55, + 533, + 296, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 533, + 296, + 714 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 533, + 296, + 714 + ], + "type": "text", + "content": " is the non-negative curvature parameter. The manifold is equipped with the Riemannian metric " + }, + { + "bbox": [ + 55, + 533, + 296, + 714 + ], + "type": "inline_equation", + "content": "g^{\\mathbb{D}} = \\lambda_c^2 g^{\\mathbb{E}}" + }, + { + "bbox": [ + 55, + 533, + 296, + 714 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 533, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{c}(\\mathbf{a}) = \\frac{2}{1 - c\\|\\mathbf{a}\\|^2}" + }, + { + "bbox": [ + 55, + 533, + 296, + 714 + ], + "type": "text", + "content": " is the conformal factor and " + }, + { + "bbox": [ + 55, + 533, + 296, + 714 + ], + "type": "inline_equation", + "content": "g^{\\mathbb{E}}" + }, + { + "bbox": [ + 55, + 533, + 296, + 714 + ], + "type": "text", + "content": " is the identity metric " + }, + { + "bbox": [ + 55, + 533, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_n" + }, + { + "bbox": [ + 55, + 533, + 296, + 714 + ], + "type": "text", + "content": " in Euclidean space. In this way, the local distances are scaled by the factor " + }, + { + "bbox": [ + 55, + 533, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{c}" + }, + { + "bbox": [ + 55, + 533, + 296, + 714 + ], + "type": "text", + "content": " approaching infinity near the boundary of the ball. This gives rise to the exponential expansion property of hyperbolic spaces, unlike the polynomial expansion in Euclidean space. However, hyperbolic" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 555, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 121 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 121 + ], + "type": "text", + "content": "space is not vector space and thus operations such as addition can not be directly conducted. To address this problem, we leverage the gyrovector formalism [48]. For a pair of points " + }, + { + "bbox": [ + 313, + 72, + 555, + 121 + ], + "type": "inline_equation", + "content": "\\mathbf{a},\\mathbf{b}\\in \\mathbb{D}_c^n" + }, + { + "bbox": [ + 313, + 72, + 555, + 121 + ], + "type": "text", + "content": ", their Möbius addition is defined as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 348, + 128, + 555, + 148 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 128, + 555, + 148 + ], + "spans": [ + { + "bbox": [ + 348, + 128, + 555, + 148 + ], + "type": "interline_equation", + "content": "\\mathbf {a} \\oplus_ {c} \\mathbf {b} = \\frac {(1 + 2 c \\langle \\mathbf {a} , \\mathbf {b} \\rangle + c \\| \\mathbf {b} \\| ^ {2}) \\mathbf {a} + (1 - c \\| \\mathbf {a} \\| ^ {2}) \\mathbf {b}}{1 + 2 c \\langle \\mathbf {a} , \\mathbf {b} \\rangle + c ^ {2} \\| \\mathbf {a} \\| ^ {2} \\| \\mathbf {b} \\| ^ {2}}. \\qquad (5)", + "image_path": "b600a93628ffb7ba4d1cd9fa529b9b3c647396ebdadd6655365b7054019c8030.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 151, + 504, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 151, + 504, + 162 + ], + "spans": [ + { + "bbox": [ + 313, + 151, + 504, + 162 + ], + "type": "text", + "content": "The hyperbolic distance between them is then:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 348, + 171, + 555, + 187 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 171, + 555, + 187 + ], + "spans": [ + { + "bbox": [ + 348, + 171, + 555, + 187 + ], + "type": "interline_equation", + "content": "\\mathcal {D} _ {\\mathbb {H}} (\\mathbf {a}, \\mathbf {b}) = \\frac {2}{\\sqrt {c}} \\operatorname {a r c t a n h} \\left(\\sqrt {c} \\| - \\mathbf {a} \\oplus_ {c} \\mathbf {b} \\|\\right) \\tag {6}", + "image_path": "cb70794bbdee5f192b5e4079e0b8921f55ece4f6c901fd2da13916937149e8b8.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 194, + 554, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 194, + 554, + 218 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 554, + 218 + ], + "type": "text", + "content": "When " + }, + { + "bbox": [ + 313, + 194, + 554, + 218 + ], + "type": "inline_equation", + "content": "c\\to 0" + }, + { + "bbox": [ + 313, + 194, + 554, + 218 + ], + "type": "text", + "content": " , the hyperbolic distance (Eq. 6) closes to the Euclidean distance " + }, + { + "bbox": [ + 313, + 194, + 554, + 218 + ], + "type": "inline_equation", + "content": "\\lim_{c\\to 0}\\mathcal{D}_{\\mathbb{H}}(\\mathbf{a},\\mathbf{b}) = 2\\| \\mathbf{a} - \\mathbf{b}\\|" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 225, + 373, + 238 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 225, + 373, + 238 + ], + "spans": [ + { + "bbox": [ + 313, + 225, + 373, + 238 + ], + "type": "text", + "content": "3.4. HypCD" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 243, + 554, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 554, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 554, + 373 + ], + "type": "text", + "content": "As illustrated in Fig. 3, we propose a unified framework, HypCD, for category discovery in hyperbolic space, incorporating both parametric [59] and non-parametric GCD approaches. Given two randomly augmented views, we initially obtain the respective Euclidean feature vectors " + }, + { + "bbox": [ + 313, + 243, + 554, + 373 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i" + }, + { + "bbox": [ + 313, + 243, + 554, + 373 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 243, + 554, + 373 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i'" + }, + { + "bbox": [ + 313, + 243, + 554, + 373 + ], + "type": "text", + "content": " through a self-supervised pretrained backbone [7, 40]. Subsequently, the feature embeddings are mapped into hyperbolic space " + }, + { + "bbox": [ + 313, + 243, + 554, + 373 + ], + "type": "inline_equation", + "content": "\\mathbb{H}^n" + }, + { + "bbox": [ + 313, + 243, + 554, + 373 + ], + "type": "text", + "content": " using exponential mapping, facilitating representation learning within this exponentially growing space to more effectively capture and utilize the hierarchical relationships inherent in the training data." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 374, + 554, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 374, + 554, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 374, + 554, + 422 + ], + "type": "text", + "content": "The exponential mapping [31] serves as a bijective projection between Euclidean space " + }, + { + "bbox": [ + 313, + 374, + 554, + 422 + ], + "type": "inline_equation", + "content": "\\mathbb{E}^n" + }, + { + "bbox": [ + 313, + 374, + 554, + 422 + ], + "type": "text", + "content": " and hyperbolic space " + }, + { + "bbox": [ + 313, + 374, + 554, + 422 + ], + "type": "inline_equation", + "content": "\\mathbb{H}^n" + }, + { + "bbox": [ + 313, + 374, + 554, + 422 + ], + "type": "text", + "content": ". The projection of tangent vector " + }, + { + "bbox": [ + 313, + 374, + 554, + 422 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 313, + 374, + 554, + 422 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 374, + 554, + 422 + ], + "type": "inline_equation", + "content": "\\mathbb{E}^n" + }, + { + "bbox": [ + 313, + 374, + 554, + 422 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 374, + 554, + 422 + ], + "type": "inline_equation", + "content": "\\mathbb{H}^n" + }, + { + "bbox": [ + 313, + 374, + 554, + 422 + ], + "type": "text", + "content": " is formulated as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 328, + 427, + 555, + 455 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 427, + 555, + 455 + ], + "spans": [ + { + "bbox": [ + 328, + 427, + 555, + 455 + ], + "type": "interline_equation", + "content": "\\exp_ {\\mathbf {o}} ^ {c} (\\mathbf {z}) = \\mathbf {o} \\oplus_ {c} \\left(\\tanh \\left(\\sqrt {c} \\frac {\\lambda_ {\\mathbf {o}} ^ {c} \\| \\mathbf {z} \\|}{2}\\right) \\frac {\\mathbf {z}}{\\sqrt {c} \\| \\mathbf {z} \\|}\\right), \\tag {7}", + "image_path": "12145f6c8c1f0bb36b8444678431f9653a098ddaed50b00442963f46fd9bf9c3.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 461, + 554, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 554, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 554, + 581 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 461, + 554, + 581 + ], + "type": "inline_equation", + "content": "\\oplus_{c}" + }, + { + "bbox": [ + 313, + 461, + 554, + 581 + ], + "type": "text", + "content": " is the Mobius addition, as introduced in Eq. 5 and " + }, + { + "bbox": [ + 313, + 461, + 554, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{o}" + }, + { + "bbox": [ + 313, + 461, + 554, + 581 + ], + "type": "text", + "content": " represents the base point of the mapping. To address the issue of gradient vanishing [23] near the boundary of the Poincaré ball during training, we implement a feature clipping operation in " + }, + { + "bbox": [ + 313, + 461, + 554, + 581 + ], + "type": "inline_equation", + "content": "\\mathbb{E}^n" + }, + { + "bbox": [ + 313, + 461, + 554, + 581 + ], + "type": "text", + "content": " prior to the exponential mapping. The operation is defined as: " + }, + { + "bbox": [ + 313, + 461, + 554, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{C}(\\mathbf{z}) = \\min \\{1,\\frac{r}{||\\mathbf{z}||}\\} \\cdot \\mathbf{z}" + }, + { + "bbox": [ + 313, + 461, + 554, + 581 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 461, + 554, + 581 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 461, + 554, + 581 + ], + "type": "text", + "content": " denotes the clipping value. For the feature vector " + }, + { + "bbox": [ + 313, + 461, + 554, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i" + }, + { + "bbox": [ + 313, + 461, + 554, + 581 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 313, + 461, + 554, + 581 + ], + "type": "inline_equation", + "content": "\\mathbb{E}^n" + }, + { + "bbox": [ + 313, + 461, + 554, + 581 + ], + "type": "text", + "content": ", the corresponding mapped feature in " + }, + { + "bbox": [ + 313, + 461, + 554, + 581 + ], + "type": "inline_equation", + "content": "\\mathbb{H}^n" + }, + { + "bbox": [ + 313, + 461, + 554, + 581 + ], + "type": "text", + "content": " is expressed as " + }, + { + "bbox": [ + 313, + 461, + 554, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{M}(\\mathbf{z}_i) = \\exp_{\\mathbf{o}}^{c}(\\mathcal{C}(\\mathbf{z}_{i}))" + }, + { + "bbox": [ + 313, + 461, + 554, + 581 + ], + "type": "text", + "content": ". The same operation will also be applied to the other feature vector " + }, + { + "bbox": [ + 313, + 461, + 554, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i'" + }, + { + "bbox": [ + 313, + 461, + 554, + 581 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 582, + 554, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 554, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 554, + 665 + ], + "type": "text", + "content": "As described in Sec.3.2, both parametric [51] and nonparametric [59] baselines utilize the same representation learning method. In our framework, we implement a consistent representation learning solution in hyperbolic space for them (Fig.3(a)). For parametric approaches, a hyperbolic parametric classifier is employed (Fig. 3(b)). We will introduce these components in detail subsequently." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 666, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 554, + 714 + ], + "type": "text", + "content": "Hyperbolic Representation Learning. Following prior attempts [45, 51, 59], we incorporate both self-supervised and supervised contrastive learning into our framework. However, our approach uniquely operates within hyperbolic" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9894" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 73, + 365, + 172 + ], + "blocks": [ + { + "bbox": [ + 58, + 73, + 365, + 172 + ], + "lines": [ + { + "bbox": [ + 58, + 73, + 365, + 172 + ], + "spans": [ + { + "bbox": [ + 58, + 73, + 365, + 172 + ], + "type": "image", + "image_path": "ed52ceb0b1900ddee19b23f6af1d6331bef6de88be7adb5c9f5c0c05431eca90.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 366, + 74, + 547, + 171 + ], + "blocks": [ + { + "bbox": [ + 366, + 74, + 547, + 171 + ], + "lines": [ + { + "bbox": [ + 366, + 74, + 547, + 171 + ], + "spans": [ + { + "bbox": [ + 366, + 74, + 547, + 171 + ], + "type": "image", + "image_path": "7ab57b7d5ec05ca663bcf6556fe08cf1a501351ba238d6bd7a0c18438767f5e7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 58, + 173, + 364, + 270 + ], + "blocks": [ + { + "bbox": [ + 58, + 173, + 364, + 270 + ], + "lines": [ + { + "bbox": [ + 58, + 173, + 364, + 270 + ], + "spans": [ + { + "bbox": [ + 58, + 173, + 364, + 270 + ], + "type": "image", + "image_path": "2138d4663106c5d07f1aedce0e931db7dfed1e924a16e028f65d8579b2ac8c94.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 273, + 553, + 297 + ], + "lines": [ + { + "bbox": [ + 55, + 273, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 55, + 273, + 553, + 297 + ], + "type": "text", + "content": "Figure 3. Overall pipeline of our HypCD framework for parametric and non-parametric GCD baselines. (a) Hyperbolic representation learning. (b) Hyperbolic classifier. (c) Non-parametric label assignment. (d) Parametric label assignment." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 367, + 173, + 548, + 268 + ], + "blocks": [ + { + "bbox": [ + 367, + 173, + 548, + 268 + ], + "lines": [ + { + "bbox": [ + 367, + 173, + 548, + 268 + ], + "spans": [ + { + "bbox": [ + 367, + 173, + 548, + 268 + ], + "type": "image", + "image_path": "8a985e64d6e865db5b5b45af74fbc24dae17c89265f3d9603a7eac2402797f71.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 304, + 296, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 304, + 296, + 400 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 296, + 400 + ], + "type": "text", + "content": "space. Furthermore, unlike previous GCD methods that exclusively utilize cosine distance [51, 55, 59] (angle-based) or Euclidean distance [45] (distance-based) for calculating pairwise similarity, we propose a hybrid approach that combines both distance-based and angle-based losses. Such integration has been shown to be more effective for model optimization in hyperbolic space [18]. First, the unified form of self-supervised contrastive loss can be defined as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 70, + 407, + 295, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 407, + 295, + 430 + ], + "spans": [ + { + "bbox": [ + 70, + 407, + 295, + 430 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {u} = \\frac {1}{| B |} \\sum_ {i \\in B} - \\log \\frac {\\exp \\left(\\mathcal {S} \\left(\\mathcal {M} \\left(\\mathbf {z} _ {i}\\right) , \\mathcal {M} \\left(\\mathbf {z} _ {i} ^ {\\prime}\\right)\\right) / \\tau_ {r}\\right)}{\\sum_ {j} ^ {j \\neq i} \\exp \\left(\\mathcal {S} \\left(\\mathcal {M} \\left(\\mathbf {z} _ {i}\\right) , \\mathcal {M} \\left(\\mathbf {z} _ {j} ^ {\\prime}\\right)\\right) / \\tau_ {r}\\right)}. \\tag {8}", + "image_path": "65258b36cc17ca5bb7e556af2b470e8e8ca48605e01d5b4f13118b47cfe4aa88.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 437, + 275, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 275, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 275, + 449 + ], + "type": "text", + "content": "Similarly, the supervised contrastive loss is unified as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 455, + 295, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 455, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 62, + 455, + 295, + 479 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {s} = \\frac {1}{| B _ {l} |} \\sum_ {i \\in B _ {l}} \\frac {1}{| N _ {i} |} \\sum_ {q \\in N _ {i}} - \\log \\frac {\\exp \\left(\\mathcal {S} \\left(\\mathcal {M} \\left(\\mathbf {z} _ {i}\\right) , \\mathcal {M} \\left(\\mathbf {z} _ {q}\\right)\\right) / \\tau_ {r}\\right)}{\\sum_ {j \\neq i} \\exp \\left(\\mathcal {S} \\left(\\mathcal {M} \\left(\\mathbf {z} _ {i}\\right) , \\mathcal {M} \\left(\\mathbf {z} _ {j}\\right)\\right) / \\tau_ {r}\\right)}, \\tag {9}", + "image_path": "a35ef58c1e8cdf1d7c7be5675345c23d0bcc8fbccf1ce9560568d51f80edd43a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 485, + 296, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 485, + 296, + 557 + ], + "spans": [ + { + "bbox": [ + 55, + 485, + 296, + 557 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 485, + 296, + 557 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 485, + 296, + 557 + ], + "type": "text", + "content": " denotes the similarity function, which can be either distance-based or angle-based. For distance-based contrastive loss, we utilize " + }, + { + "bbox": [ + 55, + 485, + 296, + 557 + ], + "type": "inline_equation", + "content": "S_{d} = -\\mathcal{D}_{\\mathbb{H}}" + }, + { + "bbox": [ + 55, + 485, + 296, + 557 + ], + "type": "text", + "content": " as the similarity function, which is formally computed using negative Euclidean distance in prior methods [45]. For angle-based contrastive loss, we employ the cosine similarity, formulated as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 91, + 563, + 294, + 583 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 563, + 294, + 583 + ], + "spans": [ + { + "bbox": [ + 91, + 563, + 294, + 583 + ], + "type": "interline_equation", + "content": "\\mathcal {S} _ {a} \\left(\\mathcal {M} \\left(\\mathbf {z} _ {i}\\right), \\mathcal {M} \\left(\\mathbf {z} _ {i} ^ {\\prime}\\right)\\right) = \\frac {\\mathcal {M} \\left(\\mathbf {z} _ {i}\\right) \\cdot \\mathcal {M} \\left(\\mathbf {z} _ {i} ^ {\\prime}\\right)}{\\left| \\left| \\mathcal {M} \\left(\\mathbf {z} _ {i}\\right) \\right| \\right| \\cdot \\left| \\left| \\mathcal {M} \\left(\\mathbf {z} _ {i} ^ {\\prime}\\right) \\right| \\right|}. \\tag {10}", + "image_path": "c67a6db170aead8e8a75ab8d0a511bcbb07498a6c25f287b133ff9ca337c74b4.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 589, + 295, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 589, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 589, + 295, + 612 + ], + "type": "text", + "content": "Since hyperbolic space is conformal with Euclidean space, cosine similarity remains equivalent in both " + }, + { + "bbox": [ + 55, + 589, + 295, + 612 + ], + "type": "inline_equation", + "content": "\\mathbb{E}^n" + }, + { + "bbox": [ + 55, + 589, + 295, + 612 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 589, + 295, + 612 + ], + "type": "inline_equation", + "content": "\\mathbb{H}^n" + }, + { + "bbox": [ + 55, + 589, + 295, + 612 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 613, + 296, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 613, + 296, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 613, + 296, + 647 + ], + "type": "text", + "content": "The final supervised and self-supervised hyperbolic contrastive loss is composed of both distance-based and angle-based losses:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 104, + 653, + 294, + 673 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 653, + 294, + 673 + ], + "spans": [ + { + "bbox": [ + 104, + 653, + 294, + 673 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {h r e p} ^ {s} = \\alpha_ {d} \\mathcal {L} _ {d i s} ^ {s} + (1 - \\alpha_ {d}) \\mathcal {L} _ {a n g} ^ {s}, \\tag {11}", + "image_path": "dbdb3c464d203bc12edcce30fa8316ebf92dc065f7ee668d21d163b9eadf171b.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 670, + 246, + 684 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 670, + 246, + 684 + ], + "spans": [ + { + "bbox": [ + 106, + 670, + 246, + 684 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {h r e p} ^ {u} = \\alpha_ {d} \\mathcal {L} _ {d i s} ^ {u} + (1 - \\alpha_ {d}) \\mathcal {L} _ {a n g} ^ {u},", + "image_path": "221a4140fe6646557c7804c5320c8673241e243bfaf2d5dfe7a9172ed06ec0eb.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{hrep}^s" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{hrep}^u" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " represent the supervised and self-supervised hyperbolic contrastive loss, respectively. The" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 304, + 554, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 554, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 554, + 365 + ], + "type": "text", + "content": "terms " + }, + { + "bbox": [ + 313, + 304, + 554, + 365 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{dis}" + }, + { + "bbox": [ + 313, + 304, + 554, + 365 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 304, + 554, + 365 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ang}" + }, + { + "bbox": [ + 313, + 304, + 554, + 365 + ], + "type": "text", + "content": " correspond to distance-based and angle-based contrastive loss, respectively, obtained by substituting " + }, + { + "bbox": [ + 313, + 304, + 554, + 365 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 304, + 554, + 365 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 304, + 554, + 365 + ], + "type": "inline_equation", + "content": "S_{d}" + }, + { + "bbox": [ + 313, + 304, + 554, + 365 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 304, + 554, + 365 + ], + "type": "inline_equation", + "content": "S_{a}" + }, + { + "bbox": [ + 313, + 304, + 554, + 365 + ], + "type": "text", + "content": ". Additionally, " + }, + { + "bbox": [ + 313, + 304, + 554, + 365 + ], + "type": "inline_equation", + "content": "\\alpha_{d}" + }, + { + "bbox": [ + 313, + 304, + 554, + 365 + ], + "type": "text", + "content": " is the loss weight of distance-based loss. The overall training objective for hyperbolic representation learning is:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 361, + 370, + 553, + 387 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 370, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 361, + 370, + 553, + 387 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r e p} ^ {\\mathbb {H}} = \\left(1 - \\lambda_ {b} ^ {\\mathbb {H}}\\right) \\mathcal {L} _ {h r e p} ^ {u} + \\lambda_ {b} ^ {\\mathbb {H}} \\mathcal {L} _ {h r e p} ^ {s}, \\tag {12}", + "image_path": "eae061624a17823f0f2e7bafcad44242a0400df2a93edebef3ae7e3f4b611477.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 392, + 553, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 553, + 417 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 553, + 417 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 392, + 553, + 417 + ], + "type": "inline_equation", + "content": "\\lambda_b^{\\mathbb{H}}" + }, + { + "bbox": [ + 313, + 392, + 553, + 417 + ], + "type": "text", + "content": " serves as the balancing factor between the supervised and unsupervised losses." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 418, + 555, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 555, + 622 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 555, + 622 + ], + "type": "text", + "content": "Hyperbolic Classifier. To enhance the parametric baseline with hyperbolic geometry, we replace the conventional Euclidean classification head—traditionally reliant on a multilayer perceptron (MLP) in Euclidean space—with its hyperbolic counterpart, the hyperbolic feed forward network (HypFFN). The hyperbolic linear layer [19] exhibits greater alignment with the baseline [59], and we experimentally find that it outperforms the hyperbolic multinomial logistic regression layer. Consider the last linear layer of the MLP; similar to its Euclidean counterpart, the hyperbolic linear layer is parameterized by a weight matrix " + }, + { + "bbox": [ + 313, + 418, + 555, + 622 + ], + "type": "inline_equation", + "content": "\\mathbf{w} \\in \\mathbb{R}^{I \\times K}" + }, + { + "bbox": [ + 313, + 418, + 555, + 622 + ], + "type": "text", + "content": " and a bias vector " + }, + { + "bbox": [ + 313, + 418, + 555, + 622 + ], + "type": "inline_equation", + "content": "\\mathbf{s} \\in \\mathbb{R}^{1 \\times K}" + }, + { + "bbox": [ + 313, + 418, + 555, + 622 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 418, + 555, + 622 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 418, + 555, + 622 + ], + "type": "text", + "content": " denotes the input feature dimension. Given the hyperbolic feature " + }, + { + "bbox": [ + 313, + 418, + 555, + 622 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i^{\\mathbb{H}} = \\mathcal{M}(\\mathbf{z}_i) \\in \\mathbb{R}^{1 \\times I}" + }, + { + "bbox": [ + 313, + 418, + 555, + 622 + ], + "type": "text", + "content": ", the linear layer operates as HypLinear " + }, + { + "bbox": [ + 313, + 418, + 555, + 622 + ], + "type": "inline_equation", + "content": "(\\mathbf{z}_i^{\\mathbb{H}}, \\mathbf{w}, \\mathbf{s}) = \\operatorname{Proj}[(\\mathbf{w} \\otimes_c \\mathbf{z}_i^{\\mathbb{H}}) \\oplus_c \\mathbf{s}]" + }, + { + "bbox": [ + 313, + 418, + 555, + 622 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 418, + 555, + 622 + ], + "type": "inline_equation", + "content": "\\oplus_c" + }, + { + "bbox": [ + 313, + 418, + 555, + 622 + ], + "type": "text", + "content": " follows Eq. 5. The Möbius matrix-vector multiplication " + }, + { + "bbox": [ + 313, + 418, + 555, + 622 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i = \\mathbf{w} \\otimes_c \\mathbf{z}_i^{\\mathbb{H}}" + }, + { + "bbox": [ + 313, + 418, + 555, + 622 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 330, + 627, + 553, + 648 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 627, + 553, + 648 + ], + "spans": [ + { + "bbox": [ + 330, + 627, + 553, + 648 + ], + "type": "interline_equation", + "content": "\\frac {1}{\\sqrt {c}} \\tanh \\left(\\frac {\\| \\mathbf {z} _ {i} ^ {\\mathbb {H}} \\mathbf {w} \\| _ {2}}{\\| \\mathbf {z} _ {i} ^ {\\mathbb {H}} \\| _ {2}} \\tanh ^ {- 1} (\\sqrt {c} \\| \\mathbf {z} _ {i} ^ {\\mathbb {H}} \\| _ {2})\\right) \\frac {\\mathbf {z} _ {i} ^ {\\mathbb {H}} \\mathbf {w}}{\\| \\mathbf {z} _ {i} ^ {\\mathbb {H}} \\mathbf {w} \\| _ {2}}. \\tag {13}", + "image_path": "09e4c7a63b79ca2f0a2d1a278256355aa885de38c61cbd8dff061852bdd7095d.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 653, + 553, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 553, + 677 + ], + "type": "text", + "content": "To ensure numerical stability [19], a safe projection is operated on the result manifold and represented as:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 322, + 683, + 553, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 683, + 553, + 716 + ], + "spans": [ + { + "bbox": [ + 322, + 683, + 553, + 716 + ], + "type": "interline_equation", + "content": "\\operatorname {P r o j} (\\mathbf {v} _ {i}) = \\left\\{ \\begin{array}{l l} \\frac {\\mathbf {v} _ {i}}{\\| \\mathbf {v} _ {i} \\| _ {2}} \\times \\frac {1 - 1 0 ^ {- 3}}{\\sqrt {c}}, & \\frac {1 - 1 0 ^ {- 3}}{\\sqrt {c}} < \\| \\mathbf {v} _ {i} \\| _ {2} \\\\ \\mathbf {v} _ {i}, & \\text {o t h e r w i s e} \\end{array} . \\right. \\tag {14}", + "image_path": "c7ff1e0e49f09b1dccc48b9b30627e91e25d8fbf73e1482b9c048954d89ca800.jpg" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9895" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 217 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 217 + ], + "type": "text", + "content": "This integration allows our hyperbolic classifier to be seamlessly incorporated into the baseline [59] by substituting the original MLP with HypFFN. For each point in " + }, + { + "bbox": [ + 55, + 72, + 296, + 217 + ], + "type": "inline_equation", + "content": "\\mathbb{H}^n" + }, + { + "bbox": [ + 55, + 72, + 296, + 217 + ], + "type": "text", + "content": ", the tangent space at that point serves as a Euclidean subspace, enabling straightforward adaptation of Euclidean operations within this space [10]. Consequently, the cross-entropy loss for the hyperbolic classifier can be expressed as: " + }, + { + "bbox": [ + 55, + 72, + 296, + 217 + ], + "type": "inline_equation", + "content": "\\ell_{ce}^{\\mathbb{H}} = \\ell_{ce}(\\mathrm{HypFFN}(\\mathbf{z}_i^{\\mathbb{H}}),\\mathbf{y}_i)" + }, + { + "bbox": [ + 55, + 72, + 296, + 217 + ], + "type": "text", + "content": ". Additionally, we can define the hyperbolic counterpart " + }, + { + "bbox": [ + 55, + 72, + 296, + 217 + ], + "type": "inline_equation", + "content": "\\mathcal{H}^{\\mathbb{H}}" + }, + { + "bbox": [ + 55, + 72, + 296, + 217 + ], + "type": "text", + "content": " for the mean entropy " + }, + { + "bbox": [ + 55, + 72, + 296, + 217 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 55, + 72, + 296, + 217 + ], + "type": "text", + "content": ". By substituting the original " + }, + { + "bbox": [ + 55, + 72, + 296, + 217 + ], + "type": "inline_equation", + "content": "\\ell_{ce}" + }, + { + "bbox": [ + 55, + 72, + 296, + 217 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 217 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 55, + 72, + 296, + 217 + ], + "type": "text", + "content": " with our derived " + }, + { + "bbox": [ + 55, + 72, + 296, + 217 + ], + "type": "inline_equation", + "content": "\\ell_{ce}^{\\mathbb{H}}" + }, + { + "bbox": [ + 55, + 72, + 296, + 217 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 217 + ], + "type": "inline_equation", + "content": "\\mathcal{H}^{\\mathbb{H}}" + }, + { + "bbox": [ + 55, + 72, + 296, + 217 + ], + "type": "text", + "content": ", respectively, we can readily compute the final hyperbolic classifier loss " + }, + { + "bbox": [ + 55, + 72, + 296, + 217 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cls}^{\\mathbb{H}}" + }, + { + "bbox": [ + 55, + 72, + 296, + 217 + ], + "type": "text", + "content": ", as detailed in Sec. 3.2." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 225, + 163, + 237 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 225, + 163, + 237 + ], + "spans": [ + { + "bbox": [ + 55, + 225, + 163, + 237 + ], + "type": "text", + "content": "3.5.Label Assignment" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 243, + 296, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 243, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 296, + 495 + ], + "type": "text", + "content": "Existing approaches typically employ either a parametric classification head or non-parametric methods, such as semi-supervised " + }, + { + "bbox": [ + 55, + 243, + 296, + 495 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 243, + 296, + 495 + ], + "type": "text", + "content": "-means [51], for label assignment. In this paper, we do not independently assess these two methods; rather, we integrate both within the HypCD framework as shown in Fig. 3(c) and (d). For non-parametric approaches, including [51] and the recent SelEx [45], we retain the original label assignment strategy by applying semi-supervised " + }, + { + "bbox": [ + 55, + 243, + 296, + 495 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 243, + 296, + 495 + ], + "type": "text", + "content": "-means clustering directly to feature representations extracted by " + }, + { + "bbox": [ + 55, + 243, + 296, + 495 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 243, + 296, + 495 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 55, + 243, + 296, + 495 + ], + "type": "inline_equation", + "content": "\\mathbb{E}^n" + }, + { + "bbox": [ + 55, + 243, + 296, + 495 + ], + "type": "text", + "content": ". Our empirical results indicate that training in hyperbolic space allows for the transfer of hierarchical structure encoding from " + }, + { + "bbox": [ + 55, + 243, + 296, + 495 + ], + "type": "inline_equation", + "content": "\\mathbb{H}^n" + }, + { + "bbox": [ + 55, + 243, + 296, + 495 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 243, + 296, + 495 + ], + "type": "inline_equation", + "content": "\\mathbb{E}^n" + }, + { + "bbox": [ + 55, + 243, + 296, + 495 + ], + "type": "text", + "content": ". Moreover, we find that the operations of " + }, + { + "bbox": [ + 55, + 243, + 296, + 495 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 243, + 296, + 495 + ], + "type": "text", + "content": "-means in " + }, + { + "bbox": [ + 55, + 243, + 296, + 495 + ], + "type": "inline_equation", + "content": "\\mathbb{E}^n" + }, + { + "bbox": [ + 55, + 243, + 296, + 495 + ], + "type": "text", + "content": " are significantly more efficient while maintaining comparable performance. For the parametric baseline exemplified by SimGCD [59], we utilize the hyperbolic classification head to conduct classification within hyperbolic space using the trained hyperbolic classifier HypFFN. Both design choices are theoretically supported by the property of hyperbolic geometry of encoding hierarchical structures, facilitating a more intuitive and effective representation and classifier for GCD." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 507, + 132, + 520 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 507, + 132, + 520 + ], + "spans": [ + { + "bbox": [ + 55, + 507, + 132, + 520 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 527, + 212, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 527, + 212, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 527, + 212, + 540 + ], + "type": "text", + "content": "4.1. Setups and Implementations" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": "Datasets. We thoroughly evaluate our method across diverse benchmarks, including the generic image recognition datasets CIFAR-10 and CIFAR-100 [34], as well as ImageNet-100 [13]. Additionally, we assess our approach on the Semantic Shift Benchmark (SSB)[52], which includes fine-grained datasets such as CUB[54], StanfordCars [33], and FGVC-Aircraft [38]. For each dataset, we adhere to the data split scheme detailed in [51]. The method involves sampling a subset of all classes as the known ('Old') classes " + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_l" + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": ". Subsequently, " + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": " of the images from these known classes are utilized to construct " + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_l" + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": ", while the remaining images are designated as the unlabelled data " + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_u" + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": ". Evaluation Metrics. We evaluate the performance using the clustering accuracy (ACC) as defined in the litera" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 146 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 146 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 146 + ], + "type": "text", + "content": "ture [51]. The " + }, + { + "bbox": [ + 313, + 72, + 553, + 146 + ], + "type": "inline_equation", + "content": "ACC" + }, + { + "bbox": [ + 313, + 72, + 553, + 146 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 313, + 72, + 553, + 146 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_u" + }, + { + "bbox": [ + 313, + 72, + 553, + 146 + ], + "type": "text", + "content": " is computed as follows, given the ground truth " + }, + { + "bbox": [ + 313, + 72, + 553, + 146 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 313, + 72, + 553, + 146 + ], + "type": "text", + "content": " and the predicted labels " + }, + { + "bbox": [ + 313, + 72, + 553, + 146 + ], + "type": "inline_equation", + "content": "\\hat{y}_i" + }, + { + "bbox": [ + 313, + 72, + 553, + 146 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 313, + 72, + 553, + 146 + ], + "type": "inline_equation", + "content": "ACC = \\frac{1}{|\\mathbf{D}_u|}\\sum_{i = 1}^{|\\mathbf{D}_u|}\\mathbb{1}(y_i = h(\\hat{y}_i))" + }, + { + "bbox": [ + 313, + 72, + 553, + 146 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 72, + 553, + 146 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 72, + 553, + 146 + ], + "type": "text", + "content": " denotes the optimal permutation that aligns the predicted cluster assignments with the ground-truth labels. The " + }, + { + "bbox": [ + 313, + 72, + 553, + 146 + ], + "type": "inline_equation", + "content": "ACC" + }, + { + "bbox": [ + 313, + 72, + 553, + 146 + ], + "type": "text", + "content": " values for the 'All', 'Old' and 'New' classes are reported separately." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 147, + 555, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 147, + 555, + 362 + ], + "spans": [ + { + "bbox": [ + 313, + 147, + 555, + 362 + ], + "type": "text", + "content": "Implementation Details. We evaluate HypCD against the non-parametric baseline GCD [51], the parametric baseline SimGCD [59], and the SOTA method SelEx [45], utilizing both DINO [7] and DINoV2 [40] pretrained weights. Detailed information regarding SelEx can be found in the supplementary materials. For GCD [51], the output dimension of the projection head " + }, + { + "bbox": [ + 313, + 147, + 555, + 362 + ], + "type": "inline_equation", + "content": "\\rho_{r}" + }, + { + "bbox": [ + 313, + 147, + 555, + 362 + ], + "type": "text", + "content": " is 256. In the case of SimGCD [59], the feature dimension from backbone " + }, + { + "bbox": [ + 313, + 147, + 555, + 362 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 147, + 555, + 362 + ], + "type": "text", + "content": " is 768. " + }, + { + "bbox": [ + 313, + 147, + 555, + 362 + ], + "type": "inline_equation", + "content": "\\rho_{r}" + }, + { + "bbox": [ + 313, + 147, + 555, + 362 + ], + "type": "text", + "content": " and the final block of " + }, + { + "bbox": [ + 313, + 147, + 555, + 362 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 147, + 555, + 362 + ], + "type": "text", + "content": " are optimized using the SGD optimizer, with an initial learning rate of 0.1, which is decayed to 0.001 over time according to a cosine annealing schedule. The HypFFN is optimized using the Riemannian Adam optimizer [4], with a constant learning rate of 0.01. All models are trained for 200 epochs using a batch size of 128. The curvature parameter " + }, + { + "bbox": [ + 313, + 147, + 555, + 362 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 147, + 555, + 362 + ], + "type": "text", + "content": " is set to 0.05 for the fine-grained datasets and 0.01 for the generic datasets. Following baselines, the balancing factor " + }, + { + "bbox": [ + 313, + 147, + 555, + 362 + ], + "type": "inline_equation", + "content": "\\lambda_b^{\\mathbb{H}}" + }, + { + "bbox": [ + 313, + 147, + 555, + 362 + ], + "type": "text", + "content": " is set to 0.35. By default, the loss weight " + }, + { + "bbox": [ + 313, + 147, + 555, + 362 + ], + "type": "inline_equation", + "content": "\\alpha_{d}" + }, + { + "bbox": [ + 313, + 147, + 555, + 362 + ], + "type": "text", + "content": " increases linearly from 0 to 1.0." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 370, + 457, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 370, + 457, + 383 + ], + "spans": [ + { + "bbox": [ + 313, + 370, + 457, + 383 + ], + "type": "text", + "content": "4.2. Quantitative Comparison" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 387, + 555, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 387, + 555, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 387, + 555, + 496 + ], + "type": "text", + "content": "We compare our method with recent GCD methods (including ORCA [6], GCD [51], XCon [16], OpenCon [47], PromptCAL [61], DCCL [42], GPC [64], CiPR [27], SimGCD [59], " + }, + { + "bbox": [ + 313, + 387, + 555, + 496 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 313, + 387, + 555, + 496 + ], + "type": "text", + "content": "GCD [53], InfoSieve [44], SPTNet [55], CMS [12], AMEND [2] and SelEx [45]) and report the results in Tab. 1. The evaluation encompasses performance on the SSB benchmark [52] and generic datasets [13, 34]. The hyperbolic methods applying our HypCD framework are indicated by the 'Hyp-' prefix." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 318, + 510, + 432, + 622 + ], + "blocks": [ + { + "bbox": [ + 318, + 510, + 432, + 622 + ], + "lines": [ + { + "bbox": [ + 318, + 510, + 432, + 622 + ], + "spans": [ + { + "bbox": [ + 318, + 510, + 432, + 622 + ], + "type": "image", + "image_path": "b3dc1a23b1372253764360b526da3ee5f3d8e0e583c844b07bad0b328d1d1f54.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 623, + 555, + 656 + ], + "lines": [ + { + "bbox": [ + 313, + 623, + 555, + 656 + ], + "spans": [ + { + "bbox": [ + 313, + 623, + 555, + 656 + ], + "type": "text", + "content": "Figure 4. Comparison of baseline and hyperbolic counterparts on the SSB. Left: 'All' ACC (higher is better). Right: Discrepancy between 'Old' and 'New' ACC (smaller is better)." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 446, + 511, + 552, + 622 + ], + "blocks": [ + { + "bbox": [ + 446, + 511, + 552, + 622 + ], + "lines": [ + { + "bbox": [ + 446, + 511, + 552, + 622 + ], + "spans": [ + { + "bbox": [ + 446, + 511, + 552, + 622 + ], + "type": "image", + "image_path": "16a3a827d444bb117b2bbcf99c6f41a005eb7276158bb6c533c0e4a90a503dc0.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "Results on SSB. The performance of the GCD methods on the SSB benchmark, utilizing both DINO [7] and DINOv2 [40] pretrained weights, is summarized in the left section of Tab. 1. Besides, we provide a comparative analy-" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9896" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 87, + 555, + 399 + ], + "blocks": [ + { + "bbox": [ + 55, + 65, + 553, + 87 + ], + "lines": [ + { + "bbox": [ + 55, + 65, + 553, + 87 + ], + "spans": [ + { + "bbox": [ + 55, + 65, + 553, + 87 + ], + "type": "text", + "content": "Table 1. Comparison of GCD methods on the SSB [52] benchmark, CIFAR-10 [34], CIFAR-100 [34] and ImageNet-100 [13] datasets. Results are reported in ACC across the 'All', 'Old' and 'New' categories." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 87, + 555, + 399 + ], + "lines": [ + { + "bbox": [ + 56, + 87, + 555, + 399 + ], + "spans": [ + { + "bbox": [ + 56, + 87, + 555, + 399 + ], + "type": "table", + "html": "
MethodCUB [54]Stanford-Cars [33]FGVC-Aircraft [38]CIFAR-10 [34]CIFAR-100 [34]ImageNet-100 [13]
AllOldNewAllOldNewAllOldNewAllOldNewAllOldNewAllOldNew
DINOORCA [6]36.343.832.631.632.031.431.942.226.969.077.452.073.592.663.981.886.279.6
XCon [16]52.154.351.040.558.831.747.744.449.496.097.395.474.281.260.377.693.569.7
OpenCon [47]54.763.854.749.178.632.7---------84.093.881.2
PromptCAL [61]62.964.462.150.270.140.652.252.252.397.996.698.581.284.275.383.192.778.3
DCCL [42]63.560.864.943.155.736.2---96.396.596.975.376.870.280.590.576.2
GPC [64]52.055.547.538.258.927.443.340.744.890.697.687.075.484.660.175.393.466.7
PIM [11]62.775.756.243.166.931.6---94.797.493.378.384.266.583.195.377.0
μGCD [53]65.768.064.656.568.150.953.855.453.0---------
InfoSieve [44]69.477.965.255.774.846.456.363.752.594.897.793.478.382.270.580.593.873.8
CiPR [27]57.158.755.647.061.540.1---97.797.597.781.582.479.780.584.978.3
SPTNet [55]65.868.865.159.079.249.359.361.858.197.395.098.681.384.375.685.493.281.4
CMS [12]68.276.564.056.976.147.656.063.452.3---82.385.775.584.795.679.2
AMEND [2]64.975.659.652.861.848.356.473.348.296.894.697.881.079.983.383.292.978.3
GCD [51]51.356.648.739.057.629.945.041.146.991.597.988.273.076.266.574.189.866.3
Hyp-GCD61.067.058.050.860.945.848.243.650.592.997.590.674.080.062.080.492.574.4
SimGCD [59]60.365.657.753.871.945.054.259.151.897.195.198.180.181.277.883.093.177.9
Hyp-SimGCD64.865.864.262.873.457.758.758.958.596.895.997.282.483.181.286.593.783.0
SelEx [45]73.675.372.858.575.650.357.164.753.395.998.194.882.385.376.383.193.677.8
Hyp-SelEx79.875.881.862.980.054.765.967.365.196.797.696.382.485.177.086.894.682.8
DINOv2μGCD [53]74.075.973.176.191.068.966.368.765.1---------
CiPR [27]78.373.480.866.777.061.8---99.098.799.290.389.093.188.287.688.5
SPTNet [55]76.379.574.6------------90.196.187.1
GCD [51]71.971.272.365.767.864.755.447.959.297.899.097.179.684.569.978.589.573.0
Hyp-GCD75.675.175.972.880.469.162.770.059.097.598.996.884.587.578.582.992.478.2
SimGCD [59]71.578.168.371.581.966.663.969.960.998.796.799.788.589.287.289.995.587.1
Hyp-SimGCD77.677.977.482.585.881.076.470.379.498.997.799.591.590.094.691.996.289.8
SelEx [45]87.485.188.582.293.776.779.882.378.698.5*98.8*98.5*87.7*90.8*81.5*90.9*96.2*88.3*
Hyp-SelEx90.785.393.483.893.379.283.482.084.198.698.198.988.691.582.892.396.490.2
", + "image_path": "46aa01a5db04dda235b0a323e66ab90a73e8c1c137c7b141175f767bf997abdf.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 57, + 400, + 167, + 408 + ], + "lines": [ + { + "bbox": [ + 57, + 400, + 167, + 408 + ], + "spans": [ + { + "bbox": [ + 57, + 400, + 167, + 408 + ], + "type": "text", + "content": "\\*results from our implementation." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 417, + 294, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 417, + 294, + 691 + ], + "spans": [ + { + "bbox": [ + 56, + 417, + 294, + 691 + ], + "type": "text", + "content": "sis between the three baseline methods and their hyperbolic counterparts with DINO backbone in Fig. 4. Our hyperbolic methods consistently outperform their Euclidean counterparts, with particularly strong results observed when utilizing the DINOv2 backbone. Among the evaluated methods, Hyp-SelEx achieves the highest average accuracy (ACC) across all datasets, notably excelling on the CUB dataset, where it records an accuracy of " + }, + { + "bbox": [ + 56, + 417, + 294, + 691 + ], + "type": "inline_equation", + "content": "79.8\\%" + }, + { + "bbox": [ + 56, + 417, + 294, + 691 + ], + "type": "text", + "content": " for the 'All' classes with DINO and " + }, + { + "bbox": [ + 56, + 417, + 294, + 691 + ], + "type": "inline_equation", + "content": "90.7\\%" + }, + { + "bbox": [ + 56, + 417, + 294, + 691 + ], + "type": "text", + "content": " with DINOv2, establishing it as the leading approach. More strikingly, on the Stanford-Cars dataset, our Hyp-GCD method outperforms the baseline by " + }, + { + "bbox": [ + 56, + 417, + 294, + 691 + ], + "type": "inline_equation", + "content": "11.8\\%" + }, + { + "bbox": [ + 56, + 417, + 294, + 691 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 56, + 417, + 294, + 691 + ], + "type": "inline_equation", + "content": "13.3\\%" + }, + { + "bbox": [ + 56, + 417, + 294, + 691 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 417, + 294, + 691 + ], + "type": "inline_equation", + "content": "15.9\\%" + }, + { + "bbox": [ + 56, + 417, + 294, + 691 + ], + "type": "text", + "content": " in terms of ACC for the 'All', 'Old' and 'New' categories, respectively. Fig. 4 (left) illustrates the average ACC on 'All' categories across the three datasets in the SSB benchmark, indicating that our hyperbolic methods surpass the baseline by a margin of at least " + }, + { + "bbox": [ + 56, + 417, + 294, + 691 + ], + "type": "inline_equation", + "content": "6.0\\%" + }, + { + "bbox": [ + 56, + 417, + 294, + 691 + ], + "type": "text", + "content": ". Furthermore, as shown in Fig. 4 (right), hyperbolic methods exhibit a consistently smaller ACC gap between 'Old' and 'New' classes, highlighting the effectiveness of HypCD in enhancing knowledge transfer from known to unseen categories. Additionally, DINOv2 outperforms DINO across all methods, underscoring its ability to capture complex data representations more effectively." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 701, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 701, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 701, + 294, + 713 + ], + "type": "text", + "content": "Results on Generic Datasets. In the right section of Tab. 1," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 417, + 553, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 417, + 553, + 620 + ], + "spans": [ + { + "bbox": [ + 313, + 417, + 553, + 620 + ], + "type": "text", + "content": "we present the results on three widely used generic datasets: CIFAR-10 [34], CIFAR-100 [34], and ImageNet-100 [13]. Our methods demonstrate consistent improvements across all cases, regardless of the backbone employed. Notably, these enhancements are especially significant on CIFAR-100 and ImageNet-100, which present greater challenges compared to CIFAR-10, where performance is nearly saturated. For CIFAR-100, Hyp-SimGCD and Hyp-SelEx achieve the highest accuracy of " + }, + { + "bbox": [ + 313, + 417, + 553, + 620 + ], + "type": "inline_equation", + "content": "82.4\\%" + }, + { + "bbox": [ + 313, + 417, + 553, + 620 + ], + "type": "text", + "content": " for the 'All' categories using DINO, while Hyp-SimGCD ranks first with an accuracy of " + }, + { + "bbox": [ + 313, + 417, + 553, + 620 + ], + "type": "inline_equation", + "content": "91.5\\%" + }, + { + "bbox": [ + 313, + 417, + 553, + 620 + ], + "type": "text", + "content": " on this metric when utilizing DINOv2, significantly surpassing baseline methods and the previous SOTA. Results on ImageNet-100 further validate the effectiveness of hyperbolic methods; Hyp-SelEx achieves the highest performance across 'All', 'Old', and 'New' categories with both DINO and DINOv2, outperforming the baseline by a margin of up to " + }, + { + "bbox": [ + 313, + 417, + 553, + 620 + ], + "type": "inline_equation", + "content": "3.7\\%" + }, + { + "bbox": [ + 313, + 417, + 553, + 620 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 634, + 468, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 634, + 468, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 634, + 468, + 647 + ], + "type": "text", + "content": "4.3. Impact of Hyperparameters" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": "Manifold Curvature. Building on previous studies [1, 31] that explore the application of hyperbolic geometry across various tasks, the curvature parameter " + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": " (as discussed in Sec. 3.3) is a crucial factor influencing performance and may yield different optimal values across datasets and meth" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9897" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 62, + 103, + 284, + 232 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 103 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 103 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 103 + ], + "type": "text", + "content": "Table 2. Experimental results using different " + }, + { + "bbox": [ + 55, + 70, + 297, + 103 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 70, + 297, + 103 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 70, + 297, + 103 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 70, + 297, + 103 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 70, + 297, + 103 + ], + "type": "inline_equation", + "content": "\\alpha_{d}^{\\max}" + }, + { + "bbox": [ + 55, + 70, + 297, + 103 + ], + "type": "text", + "content": " values in Hyp-SimGCD with DINO [7] pre-trained backbone. Results on the CUB [54] and CIFAR-100 [34] datasets are reported." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 62, + 103, + 284, + 232 + ], + "lines": [ + { + "bbox": [ + 62, + 103, + 284, + 232 + ], + "spans": [ + { + "bbox": [ + 62, + 103, + 284, + 232 + ], + "type": "table", + "html": "
parameterStanford-Cars [33]CIFAR-100 [34]
AllOldNewAllOldNew
c = 0.0161.474.455.182.483.181.2
c = 0.0562.873.457.781.684.076.7
c = 0.162.375.156.181.182.378.8
r = 1.060.072.953.782.483.181.2
r = 1.561.275.754.281.282.478.8
r = 2.362.873.457.780.181.178.3
αdmax= 0.159.677.551.081.382.479.1
αdmax= 0.562.077.254.682.483.181.2
αdmax= 1.062.873.457.778.983.569.7
", + "image_path": "f8c5ccd41fb9249ca009f1eff4ac1f6c73eabe907395a74ad51682f00b22de3d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 243, + 297, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 243, + 297, + 399 + ], + "spans": [ + { + "bbox": [ + 54, + 243, + 297, + 399 + ], + "type": "text", + "content": "ods. Intuitively, as the value of " + }, + { + "bbox": [ + 54, + 243, + 297, + 399 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 54, + 243, + 297, + 399 + ], + "type": "text", + "content": " approaches 0, the radius tends toward infinity, causing the Poincaré ball to flatten and resemble Euclidean space; conversely, larger values of " + }, + { + "bbox": [ + 54, + 243, + 297, + 399 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 54, + 243, + 297, + 399 + ], + "type": "text", + "content": " correspond to a steeper configuration. The widely accepted range for " + }, + { + "bbox": [ + 54, + 243, + 297, + 399 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 54, + 243, + 297, + 399 + ], + "type": "text", + "content": " is between 0.01 and 0.3 [15], with larger values exceeding this range resulting in performance degradation. In our experiments, we evaluate different curvature values of 0.01, 0.05, and 0.1 using Hyp-SimGCD, as presented in Tab. 2. Our findings indicate that the optimal curvature values differ between generic and fine-grained datasets. For fine-grained datasets such as CUB [54], the optimal value is 0.05, while for generic datasets like CIFAR-100, a value of 0.01 proves to be more effective." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 400, + 296, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 400, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 400, + 296, + 544 + ], + "type": "text", + "content": "Clipping Value. As articulated in [23], feature clipping has emerged as a standard technique for training hyperbolic neural networks. In our framework, we also observe that it plays a crucial role in category discovery performance. In line with the methodology outlined in [23], we investigate a range of clipping values, specifically 1.0, 1.5, and 2.3. The results shown in the second row of Tab. 2 demonstrate that optimal clipping values vary between fine-grained and generic datasets. For fine-grained datasets like CUB [54], the optimal clipping value is determined to be 2.3. Conversely, for generic datasets like CIFAR-100 [34], a clipping value of 1.0 is shown to be more effective." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 545, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 297, + 714 + ], + "type": "text", + "content": "Loss Weight. As detailed in Sec. 3.4, we implement a hybrid contrastive loss that combines both distance-based and angle-based components, which is essential for effective optimization in hyperbolic space. A loss weight, denoted as " + }, + { + "bbox": [ + 55, + 545, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\alpha_{d}" + }, + { + "bbox": [ + 55, + 545, + 297, + 714 + ], + "type": "text", + "content": ", is introduced to regulate the balance between these two types of losses and linearly increasing from 0. In the initial stages, the model prioritizes optimizing the angle between sample points and progressively shifts focus toward optimizing the hyperbolic distance. Consistent with the observations for curvature and clipping values, the optimal max value " + }, + { + "bbox": [ + 55, + 545, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\alpha_{d}^{\\max}" + }, + { + "bbox": [ + 55, + 545, + 297, + 714 + ], + "type": "text", + "content": ", varies considerably between coarse-grained and fine-grained datasets. For fine-grained datasets such as CUB [54], an optimal value of 1.0 is observed, whereas for more generic datasets like CIFAR-100 [34], a" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 314, + 73, + 553, + 210 + ], + "blocks": [ + { + "bbox": [ + 314, + 73, + 553, + 210 + ], + "lines": [ + { + "bbox": [ + 314, + 73, + 553, + 210 + ], + "spans": [ + { + "bbox": [ + 314, + 73, + 553, + 210 + ], + "type": "image", + "image_path": "8d885524221404d04d0b3ce20ff4b5ee561b5084e32d531a58210487e8c3e8c0.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 210, + 555, + 243 + ], + "lines": [ + { + "bbox": [ + 313, + 210, + 555, + 243 + ], + "spans": [ + { + "bbox": [ + 313, + 210, + 555, + 243 + ], + "type": "text", + "content": "Figure 5. T-SNE [49] comparison between SimGCD [59] and our Hyp-SimGCD using 40 randomly sampled instances from 10 randomly selected categories of the Stanford-Cars dataset [33]." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 251, + 514, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 514, + 264 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 514, + 264 + ], + "type": "text", + "content": "value of 0.5 is found to yield better performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 270, + 450, + 284 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 450, + 284 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 450, + 284 + ], + "type": "text", + "content": "4.4. Qualitative Comparison" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 312, + 288, + 555, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 288, + 555, + 492 + ], + "spans": [ + { + "bbox": [ + 312, + 288, + 555, + 492 + ], + "type": "text", + "content": "In Fig. 5, we present a t-SNE [49] visualization of features extracted from the backbone, represented as " + }, + { + "bbox": [ + 312, + 288, + 555, + 492 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i = \\phi(\\mathbf{x}_i)" + }, + { + "bbox": [ + 312, + 288, + 555, + 492 + ], + "type": "text", + "content": ". This visualization compares SimGCD with our Hyp-SimGCD. On the left side of the figure, the clusters generated by SimGCD appear dispersed. Data points from Class 42, highlighted in pink, are spread across multiple areas, indicating significant overlap and a lack of compactness. In contrast, Hyp-SimGCD creates more distinct and tightly clustered groups, concentrating the data points of Class 42 in a more confined area. This comparison implies that Hyp-SimGCD enhances both intra-class compactness and inter-class separation through our hyperbolic representation and classifier learning method. Importantly, even within the original Euclidean space of the backbone network, Hyp-SimGCD exhibits robust clustering performance, which arises from the properties of hyperbolic space in encoding hierarchical structures." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 502, + 388, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 388, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 388, + 514 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 522, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 522, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 522, + 556, + 715 + ], + "type": "text", + "content": "In this paper, we investigate a previously overlooked perspective in GCD by utilizing a representation space that captures the hierarchical structure of each sample, instead of the conventional Euclidean or spherical spaces. Our approach leverages the distinctive properties of hyperbolic space, where the volume increases exponentially with radius. This characteristic makes hyperbolic space especially suitable for modelling data possessing hierarchical structures, thereby enhancing representational capacity for category discovery. We propose a simple yet effective framework, HypCD, for integrating hyperbolic geometry into GCD methods. Through extensive experiments with parametric and non-parametric GCD baselines and the SOTA method, our framework consistently demonstrates superior performance on public benchmarks, underscoring the effectiveness of hyperbolic space for category discovery." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9898" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 144 + ], + "type": "text", + "content": "Acknowledgements. This work is supported by National Natural Science Foundation of China (Grant No. 62306251), Hong Kong Research Grant Council - Early Career Scheme (Grant No. 27208022) and General Research Fund (Grant No. 17211024), and HKU Seed Fund for Basic Research." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 153, + 115, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 153, + 115, + 166 + ], + "spans": [ + { + "bbox": [ + 56, + 153, + 115, + 166 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 173, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 61, + 173, + 296, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 173, + 296, + 205 + ], + "spans": [ + { + "bbox": [ + 61, + 173, + 296, + 205 + ], + "type": "text", + "content": "[1] Mina Ghadimi Atigh, Julian Schoep, Erman Acar, Nanne Van Noord, and Pascal Mettes. Hyperbolic image segmentation. In CVPR, 2022. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 206, + 296, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 206, + 296, + 250 + ], + "spans": [ + { + "bbox": [ + 61, + 206, + 296, + 250 + ], + "type": "text", + "content": "[2] Anwesha Banerjee, Liyana Sahir Kallooriyakath, and Soma Biswas. Amend: Adaptive margin and expanded neighborhood for efficient generalized category discovery. In WACV, 2024. 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 251, + 296, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 251, + 296, + 282 + ], + "spans": [ + { + "bbox": [ + 62, + 251, + 296, + 282 + ], + "type": "text", + "content": "[3] Ahmad Bdeir, Kristian Schwethelm, and Niels Landwehr. Fully hyperbolic convolutional neural networks for computer vision. In ICLR, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 284, + 295, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 284, + 295, + 304 + ], + "spans": [ + { + "bbox": [ + 62, + 284, + 295, + 304 + ], + "type": "text", + "content": "[4] Gary Becigneul and Octavian-Eugen Ganea. Riemannian adaptive optimization methods. In ICLR, 2019. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 305, + 294, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 305, + 294, + 338 + ], + "spans": [ + { + "bbox": [ + 62, + 305, + 294, + 338 + ], + "type": "text", + "content": "[5] James W Cannon, William J Floyd, Richard Kenyon, Walter R Parry, et al. Hyperbolic geometry. Flavors of geometry, 31(59-115):2, 1997. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 338, + 294, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 338, + 294, + 360 + ], + "spans": [ + { + "bbox": [ + 62, + 338, + 294, + 360 + ], + "type": "text", + "content": "[6] Kaidi Cao, Maria Brbic, and Jure Leskovec. Open-world semi-supervised learning. In ICLR, 2022. 3, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 361, + 294, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 361, + 294, + 404 + ], + "spans": [ + { + "bbox": [ + 62, + 361, + 294, + 404 + ], + "type": "text", + "content": "[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 1, 2, 3, 4, 6, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 405, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 405, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 62, + 405, + 294, + 437 + ], + "type": "text", + "content": "[8] Fernando Julio Cendra, Bingchen Zhao, and Kai Han. Promptcd: Learning gaussian mixture prompt pool for continual category discovery. In ECCV, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 438, + 294, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 438, + 294, + 470 + ], + "spans": [ + { + "bbox": [ + 62, + 438, + 294, + 470 + ], + "type": "text", + "content": "[9] Ines Chami, Rex Ying, Christopher Ré, and Jure Leskovec. Hyperbolic graph convolutional neural networks. In NeurIPS, 2019. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 472, + 294, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 472, + 294, + 503 + ], + "spans": [ + { + "bbox": [ + 57, + 472, + 294, + 503 + ], + "type": "text", + "content": "[10] Weize Chen, Xu Han, Yankai Lin, Hexu Zhao, Zhiyuan Liu, Peng Li, Maosong Sun, and Jie Zhou. Fully hyperbolic neural networks. In ACL, 2022. 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 505, + 294, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 505, + 294, + 546 + ], + "spans": [ + { + "bbox": [ + 57, + 505, + 294, + 546 + ], + "type": "text", + "content": "[11] Florent Chiaroni, Jose Dolz, Ziko Imtiaz Masud, Amar Mitiche, and Ismail Ben Ayed. Parametric information maximization for generalized category discovery. In ICCV, 2023. 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 548, + 294, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 548, + 294, + 581 + ], + "spans": [ + { + "bbox": [ + 57, + 548, + 294, + 581 + ], + "type": "text", + "content": "[12] Sua Choi, Dahiun Kang, and Minsu Cho. Contrastive mean-shift learning for generalized category discovery. In CVPR, 2024. 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 582, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 582, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 57, + 582, + 294, + 613 + ], + "type": "text", + "content": "[13] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 2, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 615, + 294, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 615, + 294, + 669 + ], + "spans": [ + { + "bbox": [ + 57, + 615, + 294, + 669 + ], + "type": "text", + "content": "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "type": "text", + "content": "[15] Aleksandr Ermolov, Leyla Mirvakhabova, Valentin Khrulkov, Nicu Sebe, and Ivan Oseledets. Hyperbolic vision transformers: Combining improvements in metric learning. In CVPR, 2022. 2, 3, 4, 8" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[16] Yixin Fei, Zhongkai Zhao, Siwei Yang, and Bingchen Zhao. Xcon: Learning with experts for fine-grained category discovery. In BMVC, 2022. 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 107, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 139 + ], + "type": "text", + "content": "[17] Enrico Fini, Enver Sangineto, Stéphane Lathuiliere, Zhun Zhong, Moin Nabi, and Elisa Ricci. A unified objective for novel class discovery. In ICCV, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 140, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 553, + 172 + ], + "type": "text", + "content": "[18] Luca Franco, Paolo Mandica, Bharti Munjal, and Fabio Galasso. Hyperbolic self-paced learning for self-supervised skeleton-based action representations. In ICLR, 2023. 3, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 173, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 173, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 553, + 194 + ], + "type": "text", + "content": "[19] Octavian Ganea, Gary Bécigneul, and Thomas Hofmann. Hyperbolic neural networks. In NeurIPS, 2018. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 195, + 553, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 195, + 553, + 226 + ], + "spans": [ + { + "bbox": [ + 317, + 195, + 553, + 226 + ], + "type": "text", + "content": "[20] Zhi Gao, Yuwei Wu, Yunde Jia, and Mehrtash Harandi. Curvature generation in curved spaces for few-shot learning. In ICCV, 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 228, + 553, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 259 + ], + "type": "text", + "content": "[21] Mina GhadimiAtigh, Julian Schoep, Erman Acar, Nanne van Noord, and Pascal Mettes. Hyperbolic image segmentation. In CVPR, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 261, + 553, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 553, + 314 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 553, + 314 + ], + "type": "text", + "content": "[22] Caglar Gulcehre, Misha Denil, Mateusz Malinowski, Ali Razavi, Razvan Pascanu, Karl Moritz Hermann, Peter Battaglia, Victor Bapst, David Raposo, Adam Santoro, and Nando de Freitas. Hyperbolic attention networks. In ICLR, 2019. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 316, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 553, + 348 + ], + "type": "text", + "content": "[23] Yunhui Guo, Xudong Wang, Yubei Chen, and Stella X. Yu. Clipped hyperbolic classifiers are super-hyperbolic classifiers. In CVPR, 2022. 3, 4, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 350, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 381 + ], + "type": "text", + "content": "[24] Kai Han, Andrea Vedaldi, and Andrew Zisserman. Learning to discover novel visual categories via deep transfer clustering. In ICCV, 2019. 1, 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 383, + 553, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 553, + 425 + ], + "type": "text", + "content": "[25] Kai Han, Sylvestre-Alvise Rebuffi, Sebastien Ehrhardt, Andrea Vedaldi, and Andrew Zisserman. Automatically discovering and learning new visual categories with ranking statistics. In ICLR, 2020. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 426, + 553, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 553, + 469 + ], + "type": "text", + "content": "[26] Kai Han, Sylvestre-Alvise Rebuffi, Sebastien Ehrhardt, Andrea Vedaldi, and Andrew Zisserman. Autonovel: Automatically discovering and learning novel visual categories. IEEE TPAMI, 2021. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 471, + 553, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 471, + 553, + 503 + ], + "spans": [ + { + "bbox": [ + 316, + 471, + 553, + 503 + ], + "type": "text", + "content": "[27] Shaozhe Hao, Kai Han, and Kwan-Yee K Wong. Cipr: An efficient framework with cross-instance positive relations for generalized category discovery. TMLR, 2024. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 504, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 504, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 316, + 504, + 553, + 536 + ], + "type": "text", + "content": "[28] Xuhui Jia, Kai Han, Yukun Zhu, and Bradley Green. Joint representation learning and novel category discovery on single-and multi-modal data. In ICCV, 2021. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 537, + 553, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 553, + 569 + ], + "type": "text", + "content": "[29] Justin Johnson, Ranjay Krishna, Michael Stark, Li-Jia Li, David Shamma, Michael Bernstein, and Li Fei-Fei. Image retrieval using scene graphs. In CVPR, 2015. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 571, + 553, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 571, + 553, + 612 + ], + "spans": [ + { + "bbox": [ + 316, + 571, + 553, + 612 + ], + "type": "text", + "content": "[30] KJ Joseph, Sujoy Paul, Gaurav Aggarwal, Soma Biswas, Piyush Rai, Kai Han, and Vineeth N Balasubramanian. Novel class discovery without forgetting. In ECCV, 2022. 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 614, + 553, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 553, + 646 + ], + "type": "text", + "content": "[31] Valentin Khrulkov, Leyla Mirvakhabova, Evgeniya Ustinova, Ivan Oseledets, and Victor Lempitsky. Hyperbolic image embeddings. In CVPR, 2020. 2, 3, 4, 7" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 647, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 679 + ], + "type": "text", + "content": "[32] Fanjie Kong, Yanbei Chen, Jiarui Cai, and Davide Modolo. Hyperbolic learning with synthetic captions for open-world detection. In CVPR, 2024. 2" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[33] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In ICCV workshop, 2013. 2, 6, 7, 8" + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9899" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 95 + ], + "type": "text", + "content": "[34] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 2, 6, 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 294, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 294, + 117 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 294, + 117 + ], + "type": "text", + "content": "[35] Huimin Li, Zhentao Chen, Yunhao Xu, and Junlin Hu. Hyperbolic anomaly detection. In CVPR, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 118, + 294, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 139 + ], + "type": "text", + "content": "[36] Qi Liu, Maximilian Nickel, and Douwe Kiela. Hyperbolic graph neural networks. In NeurIPS, 2019. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 140, + 294, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 171 + ], + "type": "text", + "content": "[37] Yuanpei Liu and Kai Han. Debgcd: Debiased learning with distribution guidance for generalized category discovery. In ICLR, 2025. 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 172, + 294, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 294, + 215 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 294, + 215 + ], + "type": "text", + "content": "[38] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 217, + 294, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 248 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 248 + ], + "type": "text", + "content": "[39] Maximilian Nickel and Douwe Kiela. Poincaré embeddings for learning hierarchical representations. In NeurIPS, 2017. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 250, + 294, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 294, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 294, + 304 + ], + "type": "text", + "content": "[40] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 4, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 304, + 294, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 304, + 294, + 338 + ], + "spans": [ + { + "bbox": [ + 56, + 304, + 294, + 338 + ], + "type": "text", + "content": "[41] Jona Otholt, Christoph Meinel, and Haojin Yang. Guided cluster aggregation: A hierarchical approach to generalized category discovery. In WACV, pages 2618-2627, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 338, + 294, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 370 + ], + "type": "text", + "content": "[42] Nan Pu, Zhun Zhong, and Nicu Sebe. Dynamic conceptions of contrastive learning for generalized category discovery. In CVPR, 2023. 3, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 371, + 294, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 294, + 403 + ], + "type": "text", + "content": "[43] Nan Pu, Wenjing Li, Xingyuan Ji, Yalan Qin, Nicu Sebe, and Zhun Zhong. Federated generalized category discovery. In CVPR, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 404, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 294, + 437 + ], + "type": "text", + "content": "[44] Sarah Rastegar, Hazel Doughty, and Cees Snoek. Learn to categorize or categorize to learn? self-coding for generalized category discovery. In NeurIPS, 2023. 1, 2, 4, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 437, + 294, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 480 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 480 + ], + "type": "text", + "content": "[45] Sarah Rastegar, Mohammadreza Salehi, Yuki M Asano, Hazel Doughty, and Cees G M Snoek. Selex: Self-expertise in fine-grained generalized category discovery. In ECCV, 2024. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 481, + 294, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 481, + 294, + 503 + ], + "spans": [ + { + "bbox": [ + 56, + 481, + 294, + 503 + ], + "type": "text", + "content": "[46] Ryohei Shimizu, Yusuke Mukuta, and Tatsuya Harada. Hyperbolic neural networks++. In *ICLR*, 2021. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 504, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 294, + 525 + ], + "type": "text", + "content": "[47] Yiyou Sun and Yixuan Li. Opencon: Open-world contrastive learning. TMLR, 2022. 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 526, + 294, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 526, + 294, + 558 + ], + "spans": [ + { + "bbox": [ + 56, + 526, + 294, + 558 + ], + "type": "text", + "content": "[48] Abraham Albert Ungar. A gyrovector space approach to hyperbolic geometry. Synthesis Lectures on Mathematics and Statistics, 1(1):1-194, 2008. 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 559, + 294, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 294, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 294, + 581 + ], + "type": "text", + "content": "[49] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. JMLR, 2008. 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 582, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 294, + 613 + ], + "type": "text", + "content": "[50] Sagar Vaze, Kai Han, Andrea Vedaldi, and Andrew Zisserman. Open-set recognition: A good closed-set classifier is all you need? In ICLR, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 56, + 614, + 294, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 294, + 646 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 294, + 646 + ], + "type": "text", + "content": "[51] Sagar Vaze, Kai Han, Andrea Vedaldi, and Andrew Zisserman. Generalized category discovery. In CVPR, 2022. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 56, + 647, + 294, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 679 + ], + "type": "text", + "content": "[52] Sagar Vaze, Kai Han, Andrea Vedaldi, and Andrew Zisserman. The semantic shift benchmark. In ICML workshop, 2022. 1, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "type": "text", + "content": "[53] Sagar Vaze, Andrea Vedaldi, and Andrew Zisserman. No representation rules them all in category discovery. In NeurIPS, 2023. 3, 6, 7" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 533 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[54] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 2, 6, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 107, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 149 + ], + "type": "text", + "content": "[55] Hongjun Wang, Sagar Vaze, and Kai Han. Sptnet: An efficient alternative framework for generalized category discovery with spatial prompt tuning. In ICLR, 2024. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 152, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 184 + ], + "type": "text", + "content": "[56] Hongjun Wang, Sagar Vaze, and Kai Han. Hilo: A learning framework for generalized category discovery robust to domain shifts. In ICLR, 2025. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 186, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 229 + ], + "type": "text", + "content": "[57] Yu Wang, Zhun Zhong, Pengchong Qiao, Xuxin Cheng, Xiawu Zheng, Chang Liu, Nicu Sebe, Rongrong Ji, and Jie Chen. Discover and align taxonomic context priors for open-world semi-supervised learning. In NeurIPS, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 231, + 553, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 231, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 316, + 231, + 553, + 263 + ], + "type": "text", + "content": "[58] Simon Weber, Bar Zongür, Nikita Araslanov, and Daniel Cremers. Flattening the parent bias: Hierarchical semantic segmentation in the poincaré ball. In CVPR, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 264, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 297 + ], + "type": "text", + "content": "[59] Xin Wen, Bingchen Zhao, and Xiaojuan Qi. Parametric classification for generalized category discovery: A baseline study. In ICCV, 2023. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 298, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 553, + 341 + ], + "type": "text", + "content": "[60] Zhenzhen Weng, Mehmet Giray Ogut, Shai Limonchik, and Serena Yeung. Unsupervised discovery of the long-tail in instance segmentation using hierarchical self-supervision. In CVPR, 2021. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 342, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 553, + 387 + ], + "type": "text", + "content": "[61] Sheng Zhang, Salman Khan, Zhiqiang Shen, Muzammal Naseer, Guangyi Chen, and Fahad Shahbaz Khan. Promptable: Contrastive affinity learning via auxiliary prompts for generalized novel category discovery. In CVPR, 2023. 6, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 388, + 553, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 431 + ], + "type": "text", + "content": "[62] Xinwei Zhang, Jianwen Jiang, Yutong Feng, Zhi-fan Wu, Xibin Zhao, Hai Wan, Mingqian Tang, Rong Jin, and Yue Gao. Grow and merge: a unified framework for continuous categories discovery. In NeurIPS, 2022. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 433, + 553, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 433, + 553, + 465 + ], + "spans": [ + { + "bbox": [ + 316, + 433, + 553, + 465 + ], + "type": "text", + "content": "[63] Bingchen Zhao and Kai Han. Novel visual category discovery with dual ranking statistics and mutual knowledge distillation. In NeurIPS, 2021. 1, 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 467, + 553, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 467, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 316, + 467, + 553, + 499 + ], + "type": "text", + "content": "[64] Bingchen Zhao, Xin Wen, and Kai Han. Learning semi-supervised gaussian mixture models for generalized category discovery. In ICCV, 2023. 3, 6, 7" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 500, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 500, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 500, + 553, + 533 + ], + "type": "text", + "content": "[65] Zhun Zhong, Enrico Fini, Subhankar Roy, Zhiming Luo, Elisa Ricci, and Nicu Sebe. Neighborhood contrastive learning for novel class discovery. In CVPR, 2021. 3" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "9900" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Hyperbolic Safety-Aware Vision-Language Models/5ca9a6d6-45f1-4b74-b091-1c79b0b891b6_content_list.json b/2025/Hyperbolic Safety-Aware Vision-Language Models/5ca9a6d6-45f1-4b74-b091-1c79b0b891b6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a18276352a9d628cb021369d3b732d7d462f3169 --- /dev/null +++ b/2025/Hyperbolic Safety-Aware Vision-Language Models/5ca9a6d6-45f1-4b74-b091-1c79b0b891b6_content_list.json @@ -0,0 +1,1830 @@ +[ + { + "type": "text", + "text": "Hyperbolic Safety-Aware Vision-Language Models", + "text_level": 1, + "bbox": [ + 241, + 130, + 756, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tobia Poppi\\*1,2 Tejaswi Kasarla\\*3 Pascal Mettes3 Lorenzo Baraldi1 Rita Cucchiara1,4 1University of Modena and Reggio Emilia, Italy 2 University of Pisa, Italy 3 University of Amsterdam, Netherlands 4IIT-CNR, Italy", + "bbox": [ + 125, + 179, + 870, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}\\{\\text {name.surname}\\} @ \\text {unimore.it}$ $^{2}\\{\\text {name.surname}\\} @ \\text {phd.unipi.it}$ $^{3}\\{\\text {initial.surname}\\} @ \\text {uva.nl}$", + "bbox": [ + 117, + 234, + 867, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 286, + 326, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Addressing the retrieval of unsafe content from vision-language models such as CLIP is an important step towards real-world integration. Current efforts have relied on unlearning techniques that try to erase the model's knowledge of unsafe concepts. While effective in reducing unwanted outputs, unlearning limits the model's capacity to discern between safe and unsafe content. In this work, we introduce a novel approach that shifts from unlearning to an awareness paradigm by leveraging the inherent hierarchical properties of the hyperbolic space. We propose to encode safe and unsafe content as an entailment hierarchy, where both are placed in different regions of hyperbolic space. Our HySAC, Hyperbolic Safety-Aware CLIP, employs entailment loss functions to model the hierarchical and asymmetrical relations between safe and unsafe image-text pairs. This modelling - ineffective in standard vision-language models due to their reliance on Euclidean embeddings - endows the model with awareness of unsafe content, enabling it to serve as both a multimodal unsafe classifier and a flexible content retriever, with the option to dynamically redirect unsafe queries toward safer alternatives or retain the original output. Extensive experiments show that our approach not only enhances safety recognition but also establishes a more adaptable and interpretable framework for content moderation in vision-language models. Our source code is available at: https://github.com/aimagelab/HySAC", + "bbox": [ + 89, + 316, + 485, + 710 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Warning: This paper features explicit sexual content and other material that some readers may find disturbing, distressing, or offensive.", + "bbox": [ + 89, + 718, + 483, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 791, + 220, + 806 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large-scale vision-language models (VLMs) have achieved remarkable successes in various applications, including cross-model retrieval [58], text-to-image and image-to-text generation [40, 60] and various downstream tasks [45, 67,", + "bbox": [ + 89, + 816, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "74]. Popular VLMs like CLIP [58] and ALIGN [34] leverage vast amounts of web-scraped image-text data to learn rich multimodal representations by aligning visual and textual modalities. However, most large-scale datasets sourced from the web contain unsafe or inappropriate content, such as violence, nudity, or hate speech [5, 6]. The presence of such content not only raises ethical concerns but also introduces risks for real-world applications [7, 30, 75], where exposure to or misuse of this material can lead to legal and societal repercussions. Birhane et al. [7] also show that the increasing dataset scale can exacerbate hateful and unsafe content, as identified in the now removed LAION-5B [66, 69]. Addressing the issue of unsafe content in VLMs is therefore of utmost importance to ensure responsible AI practices.", + "bbox": [ + 511, + 287, + 906, + 513 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent efforts to mitigate unsafe content in vision-language models have led to the development of methods specifically designed for NSFW (Not Safe For Work) content removal. Most of these works [23, 55] have focused on unlearning (i.e., removing) the knowledge of unsafe content from the models. As a recent example, Poppi et al. [55] develop a fine-tuned version of CLIP which unlearns toxic concepts by redirecting their embeddings towards safe regions, so that retrieval always produces safe content even when the model is prompted with unsafe inputs.", + "bbox": [ + 511, + 517, + 908, + 669 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In contrast, we propose an approach for managing unsafe content in VLMs: emphasizing awareness over unlearning. Rather than hiding the flaws of VLMs by ignoring NSFW content, we aim to equip the VLMs with the ability to distinguish between safe and unsafe content. This in turn helps users of the model to expose or redirect NSFW content when necessary, a crucial step toward improving user agency, understanding, and interpretability [19].", + "bbox": [ + 511, + 671, + 908, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Inspired by recent hyperbolic vision-language models [17, 52], we introduce a hyperbolic framework that leverages the geometric properties of hyperbolic space to separate safe and unsafe content effectively. Using a paired dataset of safe and unsafe image-text inputs [55], we adjust the embeddings to create an entailment-based [24] structure. In this setup, safe concepts are positioned closer to the", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution", + "bbox": [ + 109, + 886, + 217, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4222", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "origin of the hyperbolic space, while unsafe concepts are mapped further away. Specifically, we introduce a hyperbolic safe-to-unsafe entailment mechanism that ensures safe content encompasses unsafe representations within conical regions, defining clear safety boundaries and safety traversals to dynamically adjust query embeddings along the hyperbolic space to promote safe retrievals or, alternatively, expose relevant unsafe content when necessary. This framework not only organizes data into safe and unsafe radius-based regions but also enables controlled movement within the space, allowing retrievals to favor safety as required. Experiments demonstrate that HySAC achieves clear improvements in safety awareness, retrieval performance, and NSFW content handling across multiple datasets, with robustness in both safe content redirection and controlled unsafe content accessibility.", + "bbox": [ + 89, + 90, + 480, + 330 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 349, + 232, + 364 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unlearning in vision-language models. Unlearning concepts and content has recently received a lot of attention, empowered by the success of vision-language models. Various approaches have been explored, such as full model retraining, fine-tuning, machine unlearning [9, 27, 28, 56, 57], and differential privacy [29]. Some of these efforts have focused on text-to-image models, with the goal of removing specific styles, concepts, or objects [36, 76].", + "bbox": [ + 89, + 378, + 482, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A particular emphasis has been placed on removing NSFW content, which encompasses inappropriate, unsafe, or illegal material. Schramowsky et al. [64] steer the generation away from NSFW areas, defined by a fixed set of concepts. The embedding of NSFW concepts is applied as negative guidance during the text-conditioning phase, acting as safety guidance. Gandikota et al. [23] erase visual concepts using only their name, with negative guidance serving as a teacher. Poppi et al. [55] focus on removing NSFW concepts from CLIP-like models by fine-tuning the entire model using ViSU, a multimodal dataset containing Safe/NSFW Image/Texts quadruplets. Unlike earlier approaches, which target specific components, their method fine-tunes the entire vision-and-language model, making content removal applicable to downstream tasks.", + "bbox": [ + 89, + 502, + 482, + 727 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We introduce a method for handling NSFW concepts in CLIP-like models by fine-tuning them in hyperbolic space. We exploit the hierarchical properties of hyperbolic geometry, yielding a clear advantage: the model becomes explicitly aware of whether content is safe or NSFW, rather than merely removing knowledge of unsafe content. Similarly to our approach, safe generation works [8, 64], underscore that a deeper understanding of (un-)safety can improve the control of harmful content generation.", + "bbox": [ + 89, + 729, + 482, + 864 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "NSFW concept detection. A related area of research is the automatic detection of NSFW content. Several meth", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ods have been proposed for detecting NSFW and toxic language [10, 31, 44], primarily in social media contexts. DistilBERT [62] has emerged as a promising model for this task, especially when fine-tuned for identifying adult content. Detecting inappropriate language presents a significant challenge, and this complexity extends to visual content, where various techniques have been developed to detect NSFW imagery [5, 22, 49]. In this domain, models like NudeNet [4] specialize in detecting nudity, while Q16 [63] serves as a broader classifier, capable of identifying a wider range of NSFW content. However, identifying inappropriate visual content remains a complex task, given the challenges posed by subtle visual cues, lack of contextual information, and limited data availability.", + "bbox": [ + 511, + 90, + 903, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While these detection methods focus solely on identifying unsafe content, they do not address the problem of retrieving relevant, safe alternatives when an unsafe input is detected. Our method makes it possible to jointly detect NSFW content and provide a mechanism to shift NSFW queries towards safe but relevant alternatives.", + "bbox": [ + 511, + 301, + 903, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Hyperbolic learning. A key advantage of hyperbolic space is its inherent ability to represent hierarchical or tree-like structures with minimal distortion [11, 50, 51, 61]. A comprehensive list of recent advancements is documented in surveys by Mettes et al. [46] and Peng et al. [53]. Foundational works for building neural networks in hyperbolic space [3, 24, 25, 35, 68] led to the use of hyperbolic models across multiple modalities such as images [2, 20, 21, 26, 42, 71], text [18, 38, 70, 77], graphs [12-14, 41] and recommender systems [48, 73].", + "bbox": [ + 511, + 396, + 903, + 546 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent work has shown the strong potential of hyperbolic learning for vision-language models [17, 33, 52] and demonstrated that training with a loss function enforcing entailment cones [24] leads to the emergence of hierarchical structures between the embeddings. While Desai et al. [17] enforce entailment structure across modalities, Pal et al. [52] also explicitly enforce structure within modalities by leveraging object-level compositions and nouns from text. We take inspiration from these approaches and enforce entailment relations across safe and unsafe embeddings, creating an interpretable space for traversing from unsafe regions to safe regions in CLIP models.", + "bbox": [ + 511, + 547, + 903, + 727 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 513, + 739, + 651, + 753 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Throughout this work, we operate in hyperbolic space, a Riemannian manifold with a constant negative curvature. Following Desai et al. [17], we use the Lorentz model, as it is better equipped to deal with numerical instabilities associated with the Poincaré distance metric [39, 51].", + "bbox": [ + 511, + 763, + 903, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The Lorentz model $\\mathbb{L}^n$ is an $n$ -dimensional manifold in which points are represented on the upper sheet of a two-sheeted hyperboloid in $(n + 1)$ -dimensional Minkowski spacetime. Following terminology from general relativity,", + "bbox": [ + 511, + 839, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4223", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3ff6f18357e9d4e4959c4817dac3667620f8df1c865174840c71f5c6e255466b.jpg", + "image_caption": [ + "Figure 1. Overview of our approach. HySAC builds a hyperbolic embedding that manages content safety through an entailment hierarchy. Unsafe text and images are projected to dedicated regions of hyperbolic space, allowing for safety-aware retrieval and classification." + ], + "image_footnote": [], + "bbox": [ + 99, + 88, + 898, + 300 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "for each vector $\\mathbf{p} \\in \\mathbb{R}^{n+1}$ , we refer to the first dimension, the axis of symmetry, as the time-axis, denoted by $p_0$ , and the remaining $n$ -dimensions as the spatial components, denoted $\\tilde{\\pmb{p}}$ . The Lorentz model, $\\mathbb{L}^n = (\\mathcal{L}^n, \\mathfrak{g}_p^\\kappa)$ is given as", + "bbox": [ + 89, + 356, + 483, + 419 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {n} := \\left\\{\\boldsymbol {p} \\in \\mathbb {R} ^ {n + 1}: \\langle \\mathbf {p}, \\mathbf {p} \\rangle_ {\\mathcal {L}} = - \\frac {1}{\\kappa}, p _ {0} = \\sqrt {1 / \\kappa + \\| \\tilde {\\boldsymbol {p}} \\| ^ {2}}, \\kappa > 0 \\right\\}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 426, + 483, + 468 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $-\\kappa \\in \\mathbb{R}$ denotes the curvature with the Riemannian metric $\\mathfrak{g}_p^\\kappa = \\mathrm{diag}(-1,1,\\ldots ,1)$ . The Lorentzian inner product $\\langle .,.\\rangle_{\\mathcal{L}}$ is induced by the metric tensor $\\mathfrak{g}_p^\\kappa$ and is defined for $\\mathbf{p},\\mathbf{q}\\in \\mathbb{L}^n$ as", + "bbox": [ + 89, + 469, + 483, + 531 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\langle \\mathbf {p}, \\mathbf {q} \\rangle_ {\\mathcal {L}} = - p _ {0} q _ {0} + \\langle \\tilde {\\boldsymbol {p}}, \\tilde {\\boldsymbol {q}} \\rangle , \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 537, + 480, + 554 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\langle .,.\\rangle$ is the Euclidean inner product. The Lorentzian inner product induces a norm on the Lorentzian space which can be written as $\\| \\mathbf{p}\\|_{\\mathcal{L}} = \\sqrt{\\langle\\mathbf{p},\\mathbf{p}\\rangle_{\\mathcal{L}}}$ . We now define the common hyperbolic operations in Lorentz space.", + "bbox": [ + 89, + 559, + 483, + 621 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 3.1 (Lorentzian distance). The Lorentzian distance between two points in $\\mathbb{L}^n$ is the length of their shortest path (geodesic) connecting them, computed as", + "bbox": [ + 89, + 626, + 483, + 671 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nd _ {\\mathcal {L}} (\\mathbf {p}, \\mathbf {q}) = \\sqrt {1 / \\kappa} \\cdot \\cosh^ {- 1} (- \\kappa \\langle \\mathbf {p}, \\mathbf {q} \\rangle_ {\\mathcal {L}}). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 676, + 480, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In our work, we use the negative of Lorentzian distance to calculate similarities between multimodal inputs [17, 52].", + "bbox": [ + 89, + 705, + 483, + 736 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 3.2 (Exponential map). Since Lorentz space is a Riemannian manifold, it is locally Euclidean. This is best described through the tangent space $T_{\\mathbf{p}}\\mathbb{L}^n$ , a first-order approximation of the Lorentzian manifold at a given point $\\mathbf{p}\\in \\mathbb{L}^n$ . The exponential map then provides a means to project elements from the tangent space onto the hyperboloid. Given a point $\\mathbf{v}\\in T_{\\mathbf{p}}\\mathbb{L}^n$ , the exponential map is defined as $\\exp_{\\mathbf{p}}^{\\kappa}\\colon T_{\\mathbf{p}}\\mathbb{L}^{n}\\to \\mathbb{L}^{n}$ with the expression", + "bbox": [ + 89, + 741, + 483, + 864 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\exp_ {\\mathbf {p}} ^ {\\kappa} (\\mathbf {v}) = \\cosh (\\sqrt {\\kappa} \\| \\mathbf {v} \\| _ {\\mathbb {L}}) \\mathbf {p} + \\frac {\\sinh (\\sqrt {\\kappa} \\| \\mathbf {v} \\| _ {\\mathbb {L}})}{\\sqrt {\\kappa} \\| \\mathbf {v} \\| _ {\\mathbb {L}}} \\mathbf {v}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 869, + 483, + 905 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In practice, the reference point $\\mathbf{p}$ is set to the origin $\\mathbf{0} = (\\sqrt{1 / \\kappa}, 0, \\dots, 0)^T$ on the hyperboloid, allowing $\\exp_0^\\kappa$ to project Euclidean vectors from the tangent space at $\\mathbf{0}$ directly onto the hyperboloid [17, 35]. In this work, the exponential map is used to project the outputs of the visual and textual encoders to a shared hyperbolic space.", + "bbox": [ + 511, + 366, + 906, + 457 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. HySAC: Hyperbolic Safety-Aware CLIP", + "text_level": 1, + "bbox": [ + 511, + 469, + 875, + 486 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Problem formulation and objective", + "text_level": 1, + "bbox": [ + 511, + 494, + 818, + 510 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Problem setup. Given a dataset $D = \\{(I_i, T_i)\\}_{i=1}^N$ of $N$ image-text pairs, vision-language models (e.g. CLIP [58]) align the visual and textual embeddings obtained from image and text encoders in a shared embedding space. Large-scale datasets employed for training such embedding spaces are often web-scraped and contain unsafe samples [5]. For our problem setup, in order to differentiate between safe and unsafe content, we denote safe image-text pairs in $D$ as $(I_i, T_i)$ and unsafe image-text pairs as $(I_i^\\star, T_i^\\star)$ .", + "bbox": [ + 511, + 518, + 906, + 655 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To make vision-language models aware of inappropriate contents, and enable them to avoid or redirect the representation of such content, we also require a dataset of quadruplets of safe and unsafe image-text pairs, denoted as $D^{\\star} = \\{(I_k, T_k, I_k^{\\star}, T_k^{\\star})\\}_{k=1}^{K}$ [55]. This dataset is generated following the definition of NSFW content of Schramowski et al. [64] containing the following twenty categories: hate, harassment, violence, suffering, humiliation, harm, suicide, sexual, nudity, bodily fluids, blood, obscene gestures, illegal activity, drug use, theft, vandalism, weapons, abuse, brutality, and cruelty. The dataset is constructed such that the unsafe image-text pairs $(I_k^{\\star}, T_k^{\\star})$ are specific cases or modified versions of the safe representations $(I_k, T_k)$ .", + "bbox": [ + 511, + 656, + 908, + 852 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Modelling relations in the hyperbolic space. We first consider the relationship between text and image modalities in the embedding space. Like other hyperbolic vision-", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4224", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "language models [17, 52], we consider text as a general version of images to reflect the natural structure of partial order embeddings [72]. This is enforced in the embedding space by placing text embeddings closer to the origin and image embeddings farther away, defining a modality entailment relationship formally expressed as", + "bbox": [ + 89, + 90, + 483, + 183 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ng _ {T} \\left(T _ {k}\\right) \\ll g _ {I} \\left(I _ {k}\\right), \\quad \\text {a n d} \\quad g _ {T} \\left(T _ {k} ^ {\\star}\\right) \\ll g _ {I} \\left(I _ {k} ^ {\\star}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 191, + 482, + 209 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $g_{I}$ and $g_{T}$ denote the projections of images and text in the hyperbolic model and $\\ll$ indicates that one embedding is closer to the origin than another. Differently from other models, we also impose a relationship between unsafe and safe pairs, by considering unsafe image-text pairs as specific cases of their safe counterparts within $D^{\\star}$ . Hence, we establish a safety entailment to segregate safe from unsafe content, as follows", + "bbox": [ + 89, + 217, + 483, + 338 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ng _ {I} \\left(I _ {k}\\right) \\ll g _ {T} \\left(T _ {k} ^ {\\star}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 220, + 348, + 482, + 366 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By satisfying these inequalities, $g_{I}$ and $g_{T}$ capture both the modal and safety hierarchies within the data, thereby endowing the embedding space with safety-aware properties. To sum up, our objective is to model a hyperbolic vision-language model with the following inequality chain:", + "bbox": [ + 89, + 375, + 483, + 450 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ng _ {T} \\left(T _ {k}\\right) \\ll g _ {I} \\left(I _ {k}\\right) \\ll g _ {T} \\left(T _ {k} ^ {\\star}\\right) \\ll g _ {I} \\left(I _ {k} ^ {\\star}\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 150, + 459, + 482, + 477 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Below we outline our method to create safety-aware vision-language models in hyperbolic space.", + "bbox": [ + 89, + 484, + 483, + 516 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Hyperbolic safety learning", + "text_level": 1, + "bbox": [ + 89, + 523, + 331, + 541 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To optimize Eq. 7 for safety-aware vision-language models, we propose Hyperbolic Safety-Aware CLIP to rearrange the embedding space to separate safe and unsafe regions. Our optimization consists of two components: (1) a hyperbolic safety contrastive component to align image-text pairs over a mini-batch and (2) a hyperbolic safety entailment to align safe and unsafe content. An overview is given in Figure 1.", + "bbox": [ + 89, + 546, + 483, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Hyperbolic safety contrastive learning. CLIP [58] and Safe-CLIP [55] rely on contrastive objectives to align and distribute the multimodal data. We utilize hyperbolic embeddings to align the visual and textual data. Specifically, we project the visual and textual embeddings from a pre-trained vision-language model onto a hyperboloid [17] through an exponential map (Eq. 4). Let $f_{I}(\\cdot)$ and $f_{T}(\\cdot)$ represent any Euclidean encoders for image and text. Then, $g_{I}(I_{k}) = \\exp_{0}^{\\kappa}(\\alpha_{img} \\cdot f_{I}(I_{k}))$ and $g_{T}(T_{k}) = \\exp_{0}^{\\kappa}(\\alpha_{txt} \\cdot f_{T}(T_{k}))$ represent the hyperbolic representations of a safe image-text pair, $(I_{k}, T_{k})$ . Similarly, $g_{I}(I_{k}^{\\star}) = \\exp_{0}^{\\kappa}(\\alpha_{img} \\cdot f_{I}(I_{k}^{\\star}))$ and $g_{T}(T_{k}^{\\star}) = \\exp_{0}^{\\kappa}(\\alpha_{txt} \\cdot f_{T}(T_{k}^{\\star}))$ represent the hyperbolic representations of an unsafe image-text pair $(I_{k}^{\\star}, T_{k}^{\\star})$ . $\\alpha_{img}$ and $\\alpha_{txt}$ are learnable projection scalars.", + "bbox": [ + 89, + 657, + 483, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To align representations in hyperbolic space, the similarity for the image-text and text-image contrastive loss is", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "based on negative Lorentzian distance (Eq. 3) between $g_{I}(.)$ and $g_{T}(.)$ . We compute the hyperbolic safety contrastive loss over safe image-text pairs $(I_k, T_k)$ in a batch $B$ as", + "bbox": [ + 511, + 90, + 905, + 137 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {c o n t}} ^ {*} (I, T) = - \\sum_ {i \\in B} \\log \\frac {\\exp \\left(d _ {\\mathcal {L}} \\left(g _ {I} \\left(I _ {i}\\right) , g _ {T} \\left(T _ {i}\\right)\\right) / \\tau\\right)}{\\sum_ {k = 1 , k \\neq i} ^ {B} \\exp \\left(d _ {\\mathcal {L}} \\left(g _ {I} \\left(I _ {i}\\right) , g _ {T} \\left(T _ {k}\\right)\\right) / \\tau\\right)}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 147, + 905, + 185 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\tau$ denotes a temperature hyperparameter. A similar contrastive loss is employed to preserve the multimodal structure between unsafe image-text pairs $(I_k^{\\star}, T_k^{\\star})$ . Two additional contrastive losses between cross-safety modalities ensure the alignment of quadruplets in the embedding space. The final contrastive loss is formulated as", + "bbox": [ + 511, + 186, + 906, + 276 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {\\mathrm {h S C}} (I, T, I ^ {\\star}, T ^ {\\star}) = L _ {\\text {c o n t}} ^ {\\star} (I, T) + L _ {\\text {c o n t}} ^ {\\star} \\left(I ^ {\\star}, T ^ {\\star}\\right) \\tag {9} \\\\ + L _ {c o n t} ^ {\\star} (I, T ^ {\\star}) + L _ {c o n t} ^ {\\star} (I ^ {\\star}, T). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 286, + 903, + 323 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Hyperbolic safety entailment learning. Hyperbolic entailment cones, introduced by Ganea et al. [24], generalize partial ordered embeddings [72] to any Riemannian manifold. Entailment cones induce a partial order between concepts in a dataset $\\mathcal{X}$ such that for any pair $(\\mathbf{p},\\mathbf{q})\\in \\mathcal{X}$ , if $\\mathbf{p}$ is a sub-concept of $\\mathbf{q}$ , then $\\mathbf{q}$ entails $\\mathbf{p}$ within a conical region $\\mathfrak{S}_{\\mathbf{q}}$ defined by $\\mathbf{q}$ . For the Lorentz model, $\\mathbb{L}^n$ , the half-aperture of each conical region $\\mathfrak{S}_{\\mathbf{q}}$ is defined as [17, 38]", + "bbox": [ + 511, + 345, + 906, + 467 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\omega (\\mathbf {q}) = \\sin^ {- 1} \\left(\\frac {2 K}{\\sqrt {\\kappa} \\| \\tilde {\\boldsymbol {q}} \\|}\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 477, + 905, + 512 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $-\\kappa$ is the curvature of space, and the constant $K = 0.1$ limits values near the origin [24]. To preserve partial order between image-text relationships, we add entailment from safe text to safe image and from unsafe text to unsafe images, effectively implementing Eq. 5. Specifically, a safe image $I_{k}$ must lie within the cone defined by its corresponding safe text, $T_{k}$ , characterized by the half-aperture $\\omega (T_k)$ . Similarly, an unsafe image $I_{k}^{\\star}$ must lie within the cone defined by its corresponding unsafe text $T_{k}^{\\star}$ . This is enforced through an entailment loss formulated for image-text representations by Le et al. [38] and Desai et al. [17], as", + "bbox": [ + 511, + 521, + 906, + 704 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\begin{array}{l} L _ {e n t} ^ {\\star} (I, T) = \\max \\left(0, \\phi \\left(I _ {k}, T _ {k}\\right) - \\eta \\omega \\left(T _ {k}\\right)\\right) \\text {a n d} \\\\ I _ {*} ^ {\\star} \\left(I _ {*} ^ {\\star}, T _ {*} ^ {\\star}\\right) = \\left(0 + \\left(I _ {*} ^ {\\star}, T _ {*} ^ {\\star}\\right) - \\left(T _ {*} ^ {\\star}\\right)\\right) \\end{array} \\tag {11} \\\\ L _ {e n t} ^ {\\star} \\left(I ^ {\\star}, T ^ {\\star}\\right) = \\max \\left(0, \\phi \\left(I _ {k} ^ {\\star}, T _ {k} ^ {\\star}\\right) - \\eta \\omega \\left(T _ {k} ^ {\\star}\\right)\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 713, + 903, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\phi$ is the exterior angle (between lines $I_{k}T_{k}$ and $\\mathbf{0}T_{k}$ or between lines $I_k^\\star T_k^\\star$ and $\\mathbf{0}T_k^\\star$ ) given by", + "bbox": [ + 511, + 760, + 905, + 792 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\phi \\left(I _ {k}, T _ {k}\\right) = \\cos^ {- 1} \\left(\\frac {I _ {k 0} + T _ {k 0} \\kappa \\left\\langle I _ {k} , T _ {k} \\right\\rangle_ {\\mathcal {L}}}{\\left\\| \\widetilde {T _ {k}} \\right\\| \\sqrt {\\left(\\kappa \\left\\langle I _ {k} , T _ {k} \\right\\rangle_ {\\mathcal {L}}\\right) ^ {2} - 1}}\\right). \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 803, + 905, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $\\eta$ is a threshold for the half-aperture, $\\omega (T_{k})$ [52]. Intuitively, the entailment loss $L_{ent}^{\\star}$ penalizes images $I_{k}$", + "bbox": [ + 511, + 869, + 905, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4225", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "that lie outside the cone $\\mathfrak{S}_{T_k}$ defined by their corresponding text caption $T_{k}$ . Finally, to model the safety hierarchy, we enforce that safe concepts entail unsafe ones, reflecting that safe data are more general and unsafe data are more specific. Specifically, we enforce that a safe image $I_{k}$ entails an unsafe text $T_{k}^{\\star}$ , meaning that the unsafe text $T_{k}^{\\star}$ must lie within the cone defined by the safe image $I_{k}$ , characterized by the half-aperture $\\omega (I_k)$ . This gives us the safety-entailment defined in Eq. 6. This is implemented as", + "bbox": [ + 89, + 90, + 483, + 227 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {e n t} ^ {\\star} \\left(T ^ {\\star}, I\\right) = \\max \\left(0, \\phi \\left(T _ {k} ^ {\\star}, I _ {k}\\right) - \\eta \\omega \\left(I _ {k}\\right)\\right). \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 124, + 234, + 482, + 251 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The overall entailment loss to satisfy Eq. 7 is defined as", + "bbox": [ + 93, + 257, + 465, + 272 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {h S E}} (I, T, I ^ {\\star}, T ^ {\\star}) = L _ {e n t} ^ {\\star} (I, T) + L _ {e n t} ^ {\\star} \\left(T ^ {\\star}, I\\right) + L _ {e n t} ^ {\\star} \\left(I ^ {\\star}, T ^ {\\star}\\right). \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 279, + 482, + 309 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Combined loss function. We integrate the contrastive with the entailment losses to obtain the total loss used to fine-tune the model on the dataset $D^{\\star}$ :", + "bbox": [ + 89, + 327, + 483, + 371 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL (I, T, I ^ {\\star}, T ^ {\\star}) = L _ {\\mathrm {h S C}} (I, T, I ^ {\\star}, T ^ {\\star}) + L _ {\\mathrm {h S E}} (I, T, I ^ {\\star}, T ^ {\\star}). \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 378, + 483, + 409 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our proposal allows the model to differentiate between safe and unsafe embeddings based on their distance from the origin. Safe content is closer to the center, while NSFW content is farther away. This geometric arrangement not only enables the model to detect unsafe content but also allows dynamic manipulation of embeddings. NSFW queries can be redirected toward the safe region, effectively retrieving outputs that prioritize safety and providing more precise control over content retrieval.", + "bbox": [ + 89, + 410, + 483, + 545 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Safety traversals and evaluation", + "text_level": 1, + "bbox": [ + 89, + 553, + 372, + 569 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The hyperbolic safety-aware training results in a restructuring of the shared embedding space of the vision-language model. To obtain safe but relevant retrieval outputs from unsafe queries or vice-versa, we introduce a traversal mechanism to adjust query embeddings in hyperbolic space, enhancing their similarity with either safe or unsafe content, depending on the retrieval task. This traversal involves moving the query embeddings along the line connecting them to the origin of the hyperboloid, altering their hyperbolic distance from the root. By adjusting the embeddings' positions, we align them with regions in the embedding space that correspond to the desired content type.", + "bbox": [ + 89, + 575, + 483, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Traversal Definition. Given an embedding $\\mathbf{q}$ , our method computes the distance from a predefined root feature $\\mathbf{r}$ in hyperbolic space using the Lorentzian distance function $d_{\\mathcal{L}}(\\mathbf{q},\\mathbf{r})$ . For each type of content $X\\in \\{T,I,T^{\\star},I^{\\star}\\}$ , we compute the mean distance $\\mu_X$ from the root feature $\\mathbf{r}$ based on the distribution of each category. The boundary for each type is then defined as", + "bbox": [ + 89, + 758, + 483, + 864 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tau_ {X} = \\mu_ {X} + \\tanh \\left(\\frac {\\mu_ {X} - \\alpha}{\\kappa}\\right) + 1, \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 871, + 482, + 905 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\kappa$ is the negative curvature, and $\\alpha$ is a constant set empirically to 0.8. This shift accounts for the curvature of space, ensuring the boundaries are appropriately adjusted for effective traversal. Defining four bounds allows more nuanced control over traversal depending on the retrieval task. To retrieve a content type $X$ , the query is moved along the Euclidean direction vector $\\mathbf{v}_{\\mathrm{dir}} = \\mathbf{q} - \\mathbf{r}$ toward the root feature $\\mathbf{r}$ until it reaches the corresponding boundary $\\tau_{X}$ (e.g. $\\tau_{T}$ for safe text). The target position $\\mathbf{q}^{*}$ is given as", + "bbox": [ + 511, + 90, + 906, + 227 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {q} ^ {*} = \\mathbf {r} + \\tau_ {X} \\cdot \\frac {\\mathbf {v} _ {\\mathrm {d i r}}}{\\| \\mathbf {v} _ {\\mathrm {d i r}} \\|} \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 635, + 236, + 906, + 267 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "allowing the embeddings to be repositioned to match the target content type while maintaining semantic alignment.", + "bbox": [ + 513, + 275, + 906, + 306 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 513, + 316, + 645, + 334 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Training Details", + "text_level": 1, + "bbox": [ + 513, + 342, + 673, + 358 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. Our experiments are mainly conducted on the ViSU dataset [55], containing 165k quadruplets of safe and unsafe image-text pairs. We also evaluate our model on three real-world NSFW image datasets: NudeNet [4], NSFW data source URLs $^1$ and SMID [15].", + "bbox": [ + 511, + 364, + 905, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. Our safety comparisons include the original CLIP [58] and the state-of-the-art Safe-CLIP [55]. CLIP was trained using a private dataset of 400M image-text pairs [66] which has unsafe data [5]. Safe-CLIP is finetuned on the ViSU dataset [55] to redirect unsafe content to safe correspondent one via contrastive losses and cosine similarities, aiming to unlearn NSFW concepts.", + "bbox": [ + 511, + 440, + 905, + 545 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Models. Our visual and textual encoders are the same as CLIP [58], with VIT-L/14 as visual encoder, to maintain fair comparison to Safe-CLIP [55]. During training, both the visual and textual encoder are fine-tuned using low-rank decomposition [32] with low-rank factor $r = 16$ .", + "bbox": [ + 511, + 546, + 905, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "**Optimization.** We use AdamW [43] with weight decay 0.2 and $(\\beta_{1},\\beta_{2}) = (0.9,0.98)$ . We disable weight decay for all gains, biases, and learnable scalars. The model is finetuned for 20 epochs with batch size 256. The maximum learning rate is $8\\times 10^{-4}$ . We use mixed precision [47] to accelerate training, except computing exponential map and losses for HySAC in FP32 precision for numerical stability.", + "bbox": [ + 511, + 621, + 905, + 727 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Initialization. We initialize image and text encoders akin to CLIP, along with pre-trained weights. We initialize the softmax temperature as $\\tau = 0.07$ and clamp it to a minimum value of 0.01. For HySAC, we initialize the learnable projection scalars $\\alpha_{img} = \\alpha_{txt} = 1 / \\sqrt{512}$ , the curvature parameter $c = 1.0$ and clamp it in [0.1, 10.0] to prevent training instability. All scalars are learned in logarithmic space as $\\log(1 / \\tau)$ , $\\log(c)$ , $\\log(\\alpha_{img})$ and $\\log(\\alpha_{txt})$ .", + "bbox": [ + 511, + 727, + 905, + 848 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Further details on the training setup are provided in the supplementary B.", + "bbox": [ + 511, + 848, + 905, + 878 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "$^{1}$ https://github.com/EBazarov/nsfw_data_source_url", + "bbox": [ + 531, + 887, + 810, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4226", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/0ea2b4b3cccf232c6405186e13caef8b6aa2a418e9de8123d8e2fb8b2047e285.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelText-to-Image (T-to-I)Image-to-Text (I-to-T)Text-to-Image (T*-to-I ∪ I*)Image-to-Text (I*-to-T ∪ T*)
R@1R@10R@20R@1R@10R@20R@1R@10R@20R@1R@10R@20
CLIP [58]36.871.681.539.874.283.52.024.833.24.632.940.6
MERU [17]14.943.054.214.742.353.82.215.221.54.422.629.4
HyCoCLIP [52]34.371.280.634.471.382.22.825.333.28.237.845.7
Safe-CLIP [55]45.981.889.745.382.389.88.046.958.019.162.971.1
MERU*50.084.191.151.285.392.32.339.949.45.747.954.7
HyCoCLIP*47.781.989.146.782.790.41.532.742.36.945.253.6
HySAC49.884.190.748.284.291.230.562.871.842.173.379.8
", + "bbox": [ + 94, + 88, + 903, + 234 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7d82a5cf2f84427091b543ff3e88b6249b721691d2dd62444470e1c37bddbe16.jpg", + "table_caption": [ + "Table 1. Safe content retrieval performance on ViSU test set. Across all tasks and recall rates, HySAC improves over existing safety unlearning CLIP and hyperbolic CLIP models, highlighting that our approach is able to navigate unsafe image or text inputs towards relevant but safe retrieval outputs. * CLIP fine-tuned in hyperbolic space on ViSU training set with MERU/HyCoCLIP losses." + ], + "table_footnote": [], + "table_body": "
ModelText-to-Image (T* -to-I*)Image-to-Text (I* -to-T*)Text-to-Image (T* -to-I* ∪ I)Image-to-Text (I* -to-T* ∪ T)
R@1R@10R@20R@1R@10R@20R@1R@10R@20R@1R@10R@20
CLIP [58]73.194.997.672.895.297.768.492.395.967.193.396.7
MERU [17]29.462.472.225.857.767.823.554.064.319.551.161.2
HyCoCLIP [52]69.593.195.865.091.195.063.789.793.755.288.092.7
Safe-CLIP [55]58.086.291.456.085.191.047.780.085.832.177.184.6
HySAC81.498.499.482.297.899.281.198.499.480.597.298.9
", + "bbox": [ + 93, + 290, + 903, + 402 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Unsafe content retrieval performance on ViSU test set. Akin to safe content retrieval, our approach performs best. This is a result of our objective, as we assign different content to different regions, enabling us to maintain valuable safety information.", + "bbox": [ + 89, + 405, + 903, + 434 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Experimental Results", + "text_level": 1, + "bbox": [ + 89, + 449, + 294, + 465 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To assess the performance of our proposed model, HySAC, we measure its safety awareness and its ability to handle unsafe content effectively, while retaining both safe and unsafe knowledge. In the supplementary D, we report the zero-shot generalization of our method.", + "bbox": [ + 89, + 470, + 482, + 547 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2.1. Safety retrieval comparison", + "text_level": 1, + "bbox": [ + 89, + 555, + 328, + 570 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate our model on the capability of retrieving safe and unsafe items, in comparison to CLIP [58] and Safe-CLIP [55]. We also provide a comparison to recent hyperbolic VLMs, namely MERU [17] and HyCoCLIP [52]. Safe-CLIP and HySAC fine-tuned from CLIP on ViSU [55], MERU trained on RedCaps [16], and HyCoCLIP on GRIT [54]. For a fair comparison, we additionally fine-tuned the CLIP model on the ViSU dataset in hyperbolic space, using MERU and HyCoCLIP losses. All the models are evaluated on the ViSU test set.", + "bbox": [ + 88, + 574, + 482, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Retrieval tasks are defined as text-to-image and image-to-text, where the goal is to find the most relevant counterpart for a given query. Recall@K measures the fraction of queries where the correct item appears in the top-K retrieved results. To assess the safe retrieval performance of HySAC, we measure recall exclusively on safe content in both visual and textual elements ( $T$ -to- $I$ and $I$ -to- $T$ ). This step is crucial to verify that the original CLIP model's retrieval capabilities are retained after finetuning in hyperbolic space with our training method. Then, to evaluate", + "bbox": [ + 88, + 724, + 482, + 876 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the safety-awareness capabilities of HySAC, we introduce a distinct setup in which NSFW elements are used as queries, while the retrievable items include both safe and unsafe elements ( $T^{*}$ -to- $I \\cup I^{*}$ and $I^{*}$ -to- $T \\cup T^{*}$ ). During these experiments, a retrieval is deemed correct only if the query retrieves its safe counterpart, thereby validating the model's ability to redirect unsafe queries towards safe items.", + "bbox": [ + 511, + 449, + 903, + 556 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "When retrieving with HySAC, the threshold $\\tau$ for moving query embeddings is computed using the mean distance of safe embeddings from the origin, adjusting the query embedding towards the safe region. Results are reported in Table 1, where we observe that HySAC consistently improves over both unlearning and existing hyperbolic models and features the highest recalls across all settings and rates. CLIP hyperbolic models finetuned on ViSU data (MERU\\* and HyCoCLIP\\*) perform well on safe-only retrieval, while our method achieves high performance on both safe-only and unsafe-safe retrieval, due to our safety-aware design.", + "bbox": [ + 511, + 561, + 905, + 729 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Table 2, we instead perform analyses for unsafe content retrieval. First, this involves the text-to-image and image-to-text retrieval on only unsafe elements ( $T^{*}$ -to- $I^{*}$ and $I^{*}$ -to- $T^{*}$ ). Second, instead, we perform retrieval by using unsafe elements as queries and both safe and unsafe items as retrievable items ( $T^{*}$ -to- $I^{*} \\cup I$ and $I^{*}$ -to- $T^{*} \\cup T$ ), and deem the retrieval correct only if the query retrieves its corresponding unsafe one. This setup tests the model's ability to function as a content moderator and also showcases its capacity to provide user autonomy in content retrieval decisions. In these tests, the traversal mechanism in", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "2The box data needed for HyCoCLIP was extracted using Kosmos-2.", + "bbox": [ + 107, + 886, + 472, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4227", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/648e8418109949de6e574dd295c73a730d1452a7c47aae6150a045333fd262b8.jpg", + "image_caption": [ + "Figure 2. Distributions of embedding distances from the root. We embed all ViSU training samples and visualize their distance distribution from the root. While CLIP and Safe-CLIP do not separate between texts and images, MERU does. HySAC, instead, also differentiates between safe and unsafe content." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 480, + 311 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/53401a90ee3370d824641525bb56a4fb9afe35a6b12c6fd34d9d04f97c0f75f4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Model(T-to-I)(I-to-T)(T*-to-I ∪ I*)(I*-to-T ∪ T*)
R@1R@10R@1R@10R@1R@10R@1R@10
w/o Ent52.384.950.884.74.149.05.564.5
w/o S-Ent51.084.249.884.31.439.17.463.7
HySAC49.884.148.284.230.562.842.173.3
", + "bbox": [ + 94, + 395, + 480, + 467 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "HySAC uses adjusted parameters to move in the unsafe direction, targeting the retrieval of NSFW content. Here too, HySAC achieves the best recall across all settings, demonstrating that HySAC not only prioritizes safety by navigating away from NSFW content when required but also ensures that users can access NSFW content under controlled conditions, better than existing competitors.", + "bbox": [ + 89, + 523, + 483, + 630 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2.2. Analysis of HySAC", + "text_level": 1, + "bbox": [ + 89, + 638, + 272, + 654 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Assessing HySAC embedding space. Further, we validate the organization of the embedding space as outlined in Equation 7. Our goal is to confirm that the embeddings for safe content are positioned closer to the origin of hyperbolic space, while those for unsafe content are further away, following the proposed hierarchy. Specifically, for each $X \\in \\{T, I, T^{\\star}, I^{\\star}\\}$ of the training-set of $D^{\\star}$ , we compute the distances $d_{\\mathcal{L}}(X, \\mathbf{r})$ from the root feature $\\mathbf{r}$ .", + "bbox": [ + 89, + 657, + 482, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A visualization is reported in Figure 2, where we show the distribution of embeddings in terms of their distance to the root feature. The comparative analysis is done across four different models: HySAC, CLIP, Safe-CLIP, and MERU. For both the hyperbolic models the root feature is the origin of the hyperboloid. For the Euclidean models, since the origin does not lie on the hypersphere, the root is empirically estimated as the embedding that has the least", + "bbox": [ + 89, + 780, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4ed46ecf25cbf454115a8040f6e2ded4888f5d7f9c7e7859dceacdcb01e41524.jpg", + "table_caption": [ + "Table 3. Ablation study on loss components. We evaluate HySAC against two ablations that remove loss components. Results are in the same setting of Table 1." + ], + "table_footnote": [], + "table_body": "
Model% Safe (Text-to-Image)% Safe (Image-to-Text)
NudeNetNSFW URLsSMIDNudeNetNSFW URLsSMID
CLIP78.279.755.233.344.059.1
Safe-CLIP92.692.683.475.276.465.6
HySAC96.293.980.184.495.197.9
", + "bbox": [ + 517, + 88, + 903, + 161 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Retrieval performance on real NSFW images. Rate of safe images retrieved using unsafe prompts from the ViSU test set. The retrievable set includes safe and unsafe real images, with the latter from LAION-400M and the former from NSFW sources.", + "bbox": [ + 511, + 167, + 906, + 222 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "distance from all embeddings of the training set, i.e. the $\\ell_2$ normalization of the average of all embeddings. As it can be seen, the distribution clearly shows four peak distributions for HySAC, each one representing one of the $X$ content types, elucidating the efficacy of our approach in maintaining a clear separation between safe, unsafe, textual, and visual content within the embedding space.", + "bbox": [ + 511, + 233, + 903, + 339 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation Study. In Table 3, we validate the effectiveness of the key components in HySAC, by comparing its full configuration with variants where specific losses are disabled. In particular, we employ one variant which only keeps contrastive losses (denoted as \"w/o Ent\") and one that omits the safety-entailment loss (\"w/o S-Ent\"). Results show that while removing these components slightly improves performance in scenarios involving only safe content, likely due to reduced spatial constraints, their absence significantly undermines the model's effectiveness in dealing with unsafe content, especially in unsafe-to-safe retrieval. These results underscore the essential roles that both the modality- and safety-entailment losses play in enhancing the safety awareness of the proposed model.", + "bbox": [ + 511, + 339, + 906, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Retrieval on real NSFW datasets. To further analyze the safety of HySAC, we conduct retrieval tests using NudeNet [4], NSFW data source URLs and SMID [15]. The first two datasets primarily contain nudity and pornographic content, whereas SMID includes a broader range of inappropriate content, such as violence, harm, and discrimination. We randomly select 1000 images from each dataset to serve as visual elements and use 5k NSFW captions from the ViSU test set. Both the image-to-text and text-to-image retrieval tasks also incorporate 10,000 randomly chosen retrievable safe items from LAION-400M [66].", + "bbox": [ + 511, + 551, + 905, + 717 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results, displayed in Table 4, contrast the performance of HySAC with that of CLIP [58] and Safe-CLIP [55]. We report the proportion of safe retrieval outputs when NSFW queries are used, highlighting the capability of HySAC to enhance safety in retrieval results. Traversing the embedding space towards safety prioritizes safer alternatives, which may adjust relevance in some cases. Due to the absence of datasets with real unsafe data and correlated safe alternatives, this aspect could not be evaluated. Notably, HySAC demonstrates improved performance in securing safer content compared to Safe-CLIP across most datasets for both text-to-image and image-to-text scenarios.", + "bbox": [ + 511, + 719, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4228", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/137086e44026790740324abec669e2161d42a69cb8feb260ddbe55b70414e761.jpg", + "image_caption": [ + "A train filled with toxic chemicals crashing into a bridge, causing a massive explosion and destroying everything in its path.", + "Figure 3. Qualitative traversal results. HySAC traverses towards the root feature, retrieving the top-1 text at each interpolation point. This traversal effectively transitions from unsafe to safe captions, demonstrating the model's ability to ensure safety-aware content retrieval." + ], + "image_footnote": [], + "bbox": [ + 151, + 98, + 256, + 181 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/915170f5b969afbdb3f9a7a8a4a6055527505e925b4ff33f0a9a73ba9b56ee67.jpg", + "image_caption": [ + "The train is going over the bridge over a river.", + "Several people on skis on a snowy slope, enjoying a drug-fueled party.", + "Several people in their ski gear are in the snow.", + "[ROOT]" + ], + "image_footnote": [], + "bbox": [ + 346, + 98, + 452, + 179 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b50845a781859b6ce8160e0ae58ff6b6ec0468e2a8e7a11038a16d70115d12aa.jpg", + "image_caption": [ + "[ROOT]", + "A random Black man is standing in front of a double door with a gun in his hand and a look of hatred in his eyes.", + "A person is holding an umbrella while standing in the garage.", + "[ROOT]" + ], + "image_footnote": [], + "bbox": [ + 545, + 98, + 651, + 180 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5d60419f6c4e62fcf50255a2214760fac3ec688e3827055b5e251bd2f4515c55.jpg", + "image_caption": [ + "Two naked women sitting next to each other while flashing their genitals.", + "Sultry women.", + "A couple of women sitting next to each other in front of a cake.", + "[ROOT]" + ], + "image_footnote": [], + "bbox": [ + 748, + 98, + 852, + 180 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/acfea7d169b2a6f6d04df6e1c1ecbe3307d5bc39640eab9f56498c7b04ad1462.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelNudeNetMixed NSFW
AccFPRFNRAccFPRFNR
NSFW-CNN [37]85.30.014.766.54.535.9
CLIP-classifier [65]97.30.02.776.90.111.0
CLIP-distance [59]86.40.013.677.82.022.1
NudeNet [4]91.20.08.876.94.524.6
Q16 [63]28.50.071.565.38.329.4
HySAC99.50.00.578.916.56.8
", + "bbox": [ + 94, + 356, + 480, + 463 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. NSFW classification. Comparison between HySAC and other NSFW classifiers. Metrics reported in percentages.", + "bbox": [ + 89, + 472, + 482, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Classifying NSFW content. The structure of the embedding space in HySAC also supports the classification of NSFW content. We evaluate this using the NudeNet [4] and Mixed NSFW datasets, comparing against classifiers such as NSFW-CNN [37], CLIP-classifier [65], CLIP-distance [59], NudeNet [4], and Q16 [63]. NudeNet only contains nudity, while Mixed NSFW includes different NSFW categories from various online sources and safe images from PASS [1]. We sample a 1,000-image subset from NudeNet and 442 images from Mixed NSFW, balanced between safe and unsafe. Further details of these datasets are provided in the supplementary C.", + "bbox": [ + 88, + 508, + 482, + 690 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results in Table 5 show that HySAC achieves competitive or superior results in NSFW content classification, despite not being explicitly designed for safety classification. The norm threshold, set to the ViSU dataset mean (Figure 2), differentiates safe from unsafe content.", + "bbox": [ + 89, + 691, + 482, + 766 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2.3. Visualizing the safety traversals", + "text_level": 1, + "bbox": [ + 89, + 775, + 356, + 789 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We examine traversal paths for safe text retrievals, starting with an unsafe image embedding as the query. The traversals are along the geodesic of hyperbolic space from the image to the origin of the hyperboloid, denoted [ROOT].", + "bbox": [ + 89, + 794, + 482, + 854 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The input query is an unsafe image taken from ViSU test set and the text retrieval space consists of a mix of safe and unsafe captions of ViSU test set, metadata-based caption", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "from pexels . com, and a curated list of unsafe words3. To create visualization shown in Figure 3, each retrieved text output is selected only once across all interpolation points, ensuring unique retrievals. Results show that, with HySAC, as the query nears the origin, retrieved content shifts from unsafe to safe while preserving semantic relevance. This progression illustrates the model's capability to effectively navigate the embedding space along relevant paths. For more on the experimental setup and traversal visualizations, see the supplementary E.", + "bbox": [ + 511, + 359, + 906, + 511 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Other ablation studies. For additional ablations on embedding space geometry and hyperparameter evaluation, we refer the reader to the supplementary D.", + "bbox": [ + 511, + 515, + 906, + 561 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 577, + 633, + 592 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper introduces hyperbolic safety-aware vision-language models. Where recent literature focuses on removing or unlearning unsafe image-text content, we bring a perspective of awareness. By modelling unsafe image-text content as specific cases of their safe counterparts, we can divide the space into safe and unsafe regions. We show that hyperbolic space is a natural solution for this hierarchical relation, and propose a hyperbolic CLIP model with safety entailment learning and traversal. Our approach not only results in better retrieval of relevant safe outputs given unsafe inputs but also provides more robustness and comes with an NSFW classifier as a free by-product. Dealing with NSFW data in vision-language models is an important open research problem with real-world implications, from ethical to legal and societal concerns. By opting for awareness, we find that safety recognition improves. Given the importance of the ethical implications of this work, we provide a thorough discussion in the supplementary A.", + "bbox": [ + 511, + 603, + 906, + 875 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "$^{3}$ github.com/LDNOOBW/", + "bbox": [ + 531, + 886, + 674, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4229", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 259, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We acknowledge the CINECA award under the ISCRA initiative, for the availability of high-performance computing resources. This work has been supported by the EU Horizon projects \"ELIAS - European Lighthouse of AI for Sustainability\" (No. 101120237) and \"European Lighthouse on Safe and Secure AI (ELSA)\" (No. 101070617), co-funded by the European Union. Tejaswi Kasarla also acknowledges travel support from the European Union's Horizon research and innovation programme under grant agreement No. 951847 (ELISE) and No. 101120237 (ELIAS).", + "bbox": [ + 89, + 114, + 485, + 267 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 277, + 187, + 292 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yuki M. Asano, Christian Rupprecht, Andrew Zisserman, and Andrea Vedaldi. Pass: An imagenet replacement for self-supervised pretraining without humans. NeurIPS Track on Datasets and Benchmarks, 2021. 8", + "[2] Mina Ghadimi Atigh, Julian Schoep, Erman Acar, Nanne Van Noord, and Pascal Mettes. Hyperbolic image segmentation. In CVPR, pages 4453-4462, 2022. 2", + "[3] Gary Bécigneul and Octavian-Eugen Ganea. Riemannian adaptive optimization methods. arXiv preprint arXiv:1810.00760, 2018. 2", + "[4] P Bedapudi. NudeNet: Neural Nets for Nudity Classification, Detection, and Selective Censoring, 2019. 2, 5, 7, 8", + "[5] Abeba Birhane and Vinay Uday Prabhu. Large image datasets: A pyrrhic win for computer vision? In WACV, pages 1536-1546. IEEE, 2021. 1, 2, 3, 5", + "[6] Abeba Birhane, Vinay Uday Prabhu, and Emmanuel Kahembwe. Multimodal datasets: misogyny, pornography, and malignant stereotypes. arXiv preprint arXiv:2110.01963, 2021. 1", + "[7] Abeba Birhane, Sanghyun Han, Vishnu Boddeti, Sasha Luccioni, et al. Into the laion's den: Investigating hate in multimodal datasets. NeurIPS, 36, 2024. 1", + "[8] Manuel Brack, Felix Friedrich, Patrick Schramowski, and Kristian Kersting. Mitigating inappropriateness in image generation: Can there be value in reflecting the world's ugliness? arXiv preprint arXiv:2305.18398, 2023. 2", + "[9] Yinzhi Cao and Junfeng Yang. Towards Making Systems Forget with Machine Unlearning. In IEEE Symposium on Security and Privacy, 2015. 2", + "[10] Francesco Cauteruccio, Enrico Corradini, Giorgio Terracina, Domenico Ursino, and Luca Virgili. Extraction and analysis of text patterns from nsfw adult content in reddit. Data & Knowledge Engineering, 138:101979, 2022. 2", + "[11] Benjamin Paul Chamberlain, James Clough, and Marc Peter Deisenroth. Neural embeddings of graphs in hyperbolic space. arXiv preprint arXiv:1705.10359, 2017. 2", + "[12] Ines Chami, Adva Wolf, Da-Cheng Juan, Frederic Sala, Sujith Ravi, and Christopher Ré. Low-dimensional hyperbolic knowledge graph embeddings. arXiv preprint arXiv:2005.00545, 2020. 2", + "[13] Nurendra Choudhary, Nikhil Rao, Sumeet Katariya, Karthik Subbian, and Chandan K Reddy. Self-supervised hyperboloid representations from logical queries over knowledge" + ], + "bbox": [ + 93, + 303, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "graphs. In Proceedings of the Web Conference 2021, pages 1373-1384, 2021.", + "[14] Nurendra Choudhary, Nikhil Rao, and Chandan Reddy. Hyperbolic graph neural networks at scale: a meta learning approach. NeurIPS, 36, 2024. 2", + "[15] Damien L Crone, Stefan Bode, Carsten Murawski, and Simon M Laham. The socio-moral image database (smid): A novel stimulus set for the study of social, moral and affective processes. *PloS one*, 13(1):e0190954, 2018. 5, 7", + "[16] Karan Desai, Gaurav Kaul, Zubin Aysola, and Justin Johnson. RedCaps: Web-curated image-text data created by the people, for the people. In NeurIPS Datasets and Benchmarks, 2021. 6", + "[17] Karan Desai, Maximilian Nickel, Tanmay Rajpurohit, Justin Johnson, and Shanmukha Ramakrishna Vedantam. Hyperbolic image-text representations. In ICML, pages 7694-7731. PMLR, 2023. 1, 2, 3, 4, 6", + "[18] Bhuwan Dhingra, Christopher J Shallue, Mohammad Norouzi, Andrew M Dai, and George E Dahl. Embedding text in hyperbolic spaces. arXiv preprint arXiv:1806.04313, 2018. 2", + "[19] Upol Ehsan, Q Vera Liao, Samir Passi, Mark O Riedl, and Hal Daumé III. Seamful xai: Operationalizing seamful design in explainable ai. Proceedings of the ACM on Human-Computer Interaction, 8(CSCW1):1-29, 2024. 1", + "[20] Aleksandr Ermolov, Leyla Mirvakhabova, Valentin Khrulkov, Nicu Sebe, and Ivan Oseledets. Hyperbolic vision transformers: Combining improvements in metric learning. In CVPR, pages 7409-7419, 2022. 2", + "[21] Luca Franco, Paolo Mandica, Bharti Munjal, and Fabio Galasso. Hyperbolic self-paced learning for self-supervised skeleton-based action representations. In ICLR, 2023. 2", + "[22] Shreyansh Gandhi, Samrat Kokkula, Abon Chaudhuri, Alessandro Magnani, Theban Stanley, Behzad Ahmadi, Venkatesh Kandaswamy, Omer Ovenc, and Shie Mannor. Scalable Detection of Offensive and Non-compliant Content/Logo in Product Images. In WACV, 2020. 2", + "[23] Rohit Gandikota, Joanna Materzynska, Jaden Fiotto-Kaufman, and David Bau. Erasing Concepts from Diffusion Models. In ICCV, 2023. 1, 2", + "[24] Octavian Ganea, Gary Bécigneul, and Thomas Hofmann. Hyperbolic entailment cones for learning hierarchical embeddings. In ICML, pages 1646-1655. PMLR, 2018. 1, 2, 4", + "[25] Octavian Ganea, Gary Bécigneul, and Thomas Hofmann. Hyperbolic neural networks. NeurIPS, 31, 2018. 2", + "[26] Songwei Ge, Shlok Mishra, Simon Kornblith, Chun-Liang Li, and David Jacobs. Hyperbolic contrastive learning for visual representations beyond objects. In CVPR, pages 6840-6849, 2023. 2", + "[27] Antonio Ginart, Melody Guan, Gregory Valiant, and James Y Zou. Making AI Forget You: Data Deletion in Machine Learning. In NeurIPS, 2019. 2", + "[28] Aditya Golatkar, Alessandro Achille, and Stefano Soatto. Eternal Sunshine of the Spotless Net: Selective Forgetting in Deep Networks. In CVPR, 2020. 2" + ], + "bbox": [ + 516, + 93, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4230", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Aditya Golatkar, Alessandro Achille, Yu-Xiang Wang, Aaron Roth, Michael Kearns, and Stefano Soatto. Mixed Differential Privacy in Computer Vision. In CVPR, 2022. 2", + "[30] Kimia Hamidieh, Haoran Zhang, Walter Gerych, Thomas Hartvigsen, and Marzyeh Ghassemi. Identifying implicit social biases in vision-language models. In Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society, pages 547-561, 2024. 1", + "[31] Ahmad Fathan Hidayatullah, Anisa Miladya Hakim, and Abdullah Aziz Sembada. Adult Content Classification on Indonesian Tweets using LSTM Neural Network. In ICACSIS, 2019. 2", + "[32] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 5", + "[33] Sarah Ibrahimi, Mina Ghadimi Atigh, Nanne Van Noord, Pascal Mettes, and Marcel Worring. Intriguing properties of hyperbolic embeddings in vision-language models. Transactions on Machine Learning Research, 2024. 2", + "[34] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, pages 4904-4916. PMLR, 2021. 1", + "[35] Valentin Khrulkov, Leyla Mirvakhabova, Evgeniya Ustinova, Ivan Oseledets, and Victor Lempitsky. Hyperbolic image embeddings. In CVPR, pages 6418-6428, 2020. 2, 3", + "[36] Nupur Kumari, Bingliang Zhang, Sheng-Yu Wang, Eli Shechtman, Richard Zhang, and Jun-Yan Zhu. Ablating concepts in text-to-image diffusion models. In ICCV, 2023. 2", + "[37] Gant Laborde. Deep nn for nsfw detection. https://github.com/GantMan/nsfw_model, 2020.8", + "[38] Matthew Le, Stephen Roller, Laetitia Papaxanthos, Douwe Kiela, and Maximilian Nickel. Inferring concept hierarchies from text corpora via hyperbolic embeddings. In ACL, pages 3231-3241, 2019. 2, 4", + "[39] Ya-Wei Eileen Lin, Ronald R Coifman, Gal Mishne, and Ronen Talmon. Hyperbolic diffusion embedding and distance for hierarchical representation learning. In ICML, pages 21003–21025. PMLR, 2023. 2", + "[40] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 36, 2024. 1", + "[41] Qi Liu, Maximilian Nickel, and Douwe Kiela. Hyperbolic graph neural networks. NeurIPS, 32, 2019. 2", + "[42] Shaoteng Liu, Jingjing Chen, Liangming Pan, Chong-Wah Ngo, Tat-Seng Chua, and Yu-Gang Jiang. Hyperbolic visual embedding learning for zero-shot recognition. In CVPR, pages 9273-9281, 2020. 2", + "[43] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 5", + "[44] Todor Markov, Chong Zhang, Sandhini Agarwal, Florentine Eloundou Nekoul, Theodore Lee, Steven Adler, Angela Jiang, and Lilian Weng. A Holistic Approach to Undesired Content Detection in the Real World. In AAAI, 2023. 2", + "[45] Joanna Materzyńska, Antonio Torralba, and David Bau. Disentangling visual and written concepts in clip. In CVPR, pages 16410-16419, 2022. 1" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[46] Pascal Mettes, Mina Ghadimi Atigh, Martin Keller-Ressel, Jeffrey Gu, and Serena Yeung. Hyperbolic deep learning in computer vision: A survey. IJCV, pages 1-25, 2024. 2", + "[47] Paulius Micikevicius, Sharan Narang, Jonah Alben, Gregory Diamos, Erich Elsen, David Garcia, Boris Ginsburg, Michael Houston, Oleksii Kuchaiev, Ganesh Venkatesh, et al. Mixed precision training. In ICLR, 2018. 5", + "[48] Leyla Mirvakhabova, Evgeny Frolov, Valentin Khrulkov, Ivan Oseledets, and Alexander Tuzhilin. Performance of hyperbolic geometry models on top-n recommendation tasks. In Proceedings of the 14th ACM Conference on Recommender Systems, pages 527-532, 2020. 2", + "[49] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards Photorealistic Image Generation and Editing with Text-Guided Diffusion Models. arXiv preprint arXiv:2112.10741, 2021. 2", + "[50] Maximillian Nickel and Douwe Kiela. Poincaré embeddings for learning hierarchical representations. NeurIPS, 30, 2017. 2", + "[51] Maximillian Nickel and Douwe Kiela. Learning continuous hierarchies in the lorentz model of hyperbolic geometry. In ICML, pages 3779-3788. PMLR, 2018. 2", + "[52] Avik Pal, Max van Spengler, Guido Maria D'Amely di Melendugno, Alessandro Flaborea, Fabio Galasso, and Pascal Mettes. Compositional entailment learning for hyperbolic vision-language models. arXiv preprint arXiv:2410.06912, 2024. 1, 2, 3, 4, 6", + "[53] Wei Peng, Tuomas Varanka, Abdelrahman Mostafa, Henglin Shi, and Guoying Zhao. Hyperbolic deep neural networks: A survey. IEEE Trans. PAMI, 44(12):10023-10044, 2021. 2", + "[54] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 6", + "[55] Samuele Poppi, Tobia Poppi, Federico Cocchi, Marcella Cornia, Lorenzo Baraldi, and Rita Cucchiara. Safe-CLIP: Removing NSFW Concepts from Vision-and-Language Models. In ECCV, 2024. 1, 2, 3, 4, 5, 6, 7", + "[56] Samuele Poppi, Sara Sarto, Marcella Cornia, Lorenzo Baraldi, and Rita Cucchiara. Multi-Class Unlearning for Image Classification via Weight Filtering. IEEE Intelligent Systems, 2024. 2", + "[57] Samuele Poppi, Sara Sarto, Marcella Cornia, Lorenzo Baraldi, and Rita Cucchiara. Unlearning vision transformers without retaining data via low-rank decompositions. In ICPR, 2024. 2", + "[58] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 1, 3, 4, 5, 6, 7", + "[59] Javier Rando, Daniel Paleka, David Lindner, Lennart Heim, and Florian Tramer. Red-teaming the stable diffusion safety filter. In NeurIPS ML Safety Workshop, 2022. 8" + ], + "bbox": [ + 514, + 90, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4231", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[60] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 1", + "[61] Frederic Sala, Chris De Sa, Albert Gu, and Christopher Ré. Representation tradeoffs for hyperbolic embeddings. In ICML, pages 4460-4469. PMLR, 2018. 2", + "[62] Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108, 2019. 2", + "[63] Patrick Schramowski, Christopher Tauchmann, and Kristian Kersting. Can Machines Help Us Answering Question 16 in Datasheets, and In Turn Reflecting on Inappropriate Content? In ACM FAccT, 2022. 2, 8", + "[64] Patrick Schramowski, Manuel Brack, Björn Deiseroth, and Kristian Kersting. Safe Latent Diffusion: Mitigating Inappropriate Degeneration in Diffusion Models. In CVPR, 2023. 2, 3", + "[65] Christoph Schuhmann. Clip based nsfw detector. https://github.com/LAION-AI/CLIP-based-NSFW-Detector/tree/main, 2022.8", + "[66] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 1, 5, 7", + "[67] Sheng Shen, Liunian Harold Li, Hao Tan, Mohit Bansal, Anna Rohrbach, Kai-Wei Chang, Zhewei Yao, and Kurt Keutzer. How much can clip benefit vision-and-language tasks? arXiv preprint arXiv:2107.06383, 2021. 1", + "[68] Ryohei Shimizu, Yusuke Mukuta, and Tatsuya Harada. Hyperbolic neural networks++. ICLR, 2021. 2", + "[69] David Thiel. Identifying and eliminating csam in generative ml training data and models. 2023. 1", + "[70] Alexandru Tifrea, Gary Bécigneul, and Octavian-Eugen Ganea. Poincaré glove: Hyperbolic word embeddings. arXiv preprint arXiv:1810.06546, 2018. 2", + "[71] Max van Spengler, Erwin Berkhout, and Pascal Mettes. Poincare resnet. In ICCV, pages 5419-5428, 2023. 2", + "[72] Ivan Vendrov, Ryan Kiros, Sanja Fidler, and Raquel Urtasun. Order-embeddings of images and language. arXiv preprint arXiv:1511.06361, 2015. 4", + "[73] Liping Wang, Fenyu Hu, Shu Wu, and Liang Wang. Fully hyperbolic graph convolution network for recommendation. In Proceedings of the 30th ACM international conference on information & knowledge management, pages 3483-3487, 2021. 2", + "[74] Mengmeng Wang, Jiazheng Xing, and Yong Liu. Actionclip: A new paradigm for video action recognition. arXiv preprint arXiv:2109.08472, 2021. 1", + "[75] Robert Wolfe, Yiwei Yang, Bill Howe, and Aylin Caliskan. Contrastive language-vision ai models pretrained on web-scraped multimodal data exhibit sexual objectification bias. In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency, pages 1174-1185, 2023. 1" + ], + "bbox": [ + 91, + 92, + 480, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[76] Eric Zhang, Kai Wang, Xingqian Xu, Zhangyang Wang, and Humphrey Shi. Forget-Me-Not: Learning to Forget in Text-to-Image Diffusion Models. CVPR, 2024. 2", + "[77] Yudong Zhu, Di Zhou, Jinghui Xiao, Xin Jiang, Xiao Chen, and Qun Liu. Hypertext: Endowing fasttext with hyperbolic geometry. arXiv preprint arXiv:2010.16143, 2020. 2" + ], + "bbox": [ + 514, + 92, + 903, + 176 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "4232", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Hyperbolic Safety-Aware Vision-Language Models/5ca9a6d6-45f1-4b74-b091-1c79b0b891b6_model.json b/2025/Hyperbolic Safety-Aware Vision-Language Models/5ca9a6d6-45f1-4b74-b091-1c79b0b891b6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..967ea6ebf59271976cd8faa32bee925cb4a063fc --- /dev/null +++ b/2025/Hyperbolic Safety-Aware Vision-Language Models/5ca9a6d6-45f1-4b74-b091-1c79b0b891b6_model.json @@ -0,0 +1,2730 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.243, + 0.131, + 0.757, + 0.154 + ], + "angle": 0, + "content": "Hyperbolic Safety-Aware Vision-Language Models" + }, + { + "type": "text", + "bbox": [ + 0.126, + 0.18, + 0.871, + 0.234 + ], + "angle": 0, + "content": "Tobia Poppi\\*1,2 Tejaswi Kasarla\\*3 Pascal Mettes3 Lorenzo Baraldi1 Rita Cucchiara1,4 1University of Modena and Reggio Emilia, Italy 2 University of Pisa, Italy 3 University of Amsterdam, Netherlands 4IIT-CNR, Italy" + }, + { + "type": "text", + "bbox": [ + 0.119, + 0.236, + 0.869, + 0.253 + ], + "angle": 0, + "content": "\\(^{1}\\{\\text {name.surname}\\} @ \\text {unimore.it}\\) \\(^{2}\\{\\text {name.surname}\\} @ \\text {phd.unipi.it}\\) \\(^{3}\\{\\text {initial.surname}\\} @ \\text {uva.nl}\\)" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.287, + 0.327, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.318, + 0.486, + 0.712 + ], + "angle": 0, + "content": "Addressing the retrieval of unsafe content from vision-language models such as CLIP is an important step towards real-world integration. Current efforts have relied on unlearning techniques that try to erase the model's knowledge of unsafe concepts. While effective in reducing unwanted outputs, unlearning limits the model's capacity to discern between safe and unsafe content. In this work, we introduce a novel approach that shifts from unlearning to an awareness paradigm by leveraging the inherent hierarchical properties of the hyperbolic space. We propose to encode safe and unsafe content as an entailment hierarchy, where both are placed in different regions of hyperbolic space. Our HySAC, Hyperbolic Safety-Aware CLIP, employs entailment loss functions to model the hierarchical and asymmetrical relations between safe and unsafe image-text pairs. This modelling - ineffective in standard vision-language models due to their reliance on Euclidean embeddings - endows the model with awareness of unsafe content, enabling it to serve as both a multimodal unsafe classifier and a flexible content retriever, with the option to dynamically redirect unsafe queries toward safer alternatives or retain the original output. Extensive experiments show that our approach not only enhances safety recognition but also establishes a more adaptable and interpretable framework for content moderation in vision-language models. Our source code is available at: https://github.com/aimagelab/HySAC" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.719, + 0.485, + 0.765 + ], + "angle": 0, + "content": "Warning: This paper features explicit sexual content and other material that some readers may find disturbing, distressing, or offensive." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.792, + 0.222, + 0.808 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.817, + 0.485, + 0.879 + ], + "angle": 0, + "content": "Large-scale vision-language models (VLMs) have achieved remarkable successes in various applications, including cross-model retrieval [58], text-to-image and image-to-text generation [40, 60] and various downstream tasks [45, 67," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.288, + 0.908, + 0.515 + ], + "angle": 0, + "content": "74]. Popular VLMs like CLIP [58] and ALIGN [34] leverage vast amounts of web-scraped image-text data to learn rich multimodal representations by aligning visual and textual modalities. However, most large-scale datasets sourced from the web contain unsafe or inappropriate content, such as violence, nudity, or hate speech [5, 6]. The presence of such content not only raises ethical concerns but also introduces risks for real-world applications [7, 30, 75], where exposure to or misuse of this material can lead to legal and societal repercussions. Birhane et al. [7] also show that the increasing dataset scale can exacerbate hateful and unsafe content, as identified in the now removed LAION-5B [66, 69]. Addressing the issue of unsafe content in VLMs is therefore of utmost importance to ensure responsible AI practices." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.518, + 0.909, + 0.67 + ], + "angle": 0, + "content": "Recent efforts to mitigate unsafe content in vision-language models have led to the development of methods specifically designed for NSFW (Not Safe For Work) content removal. Most of these works [23, 55] have focused on unlearning (i.e., removing) the knowledge of unsafe content from the models. As a recent example, Poppi et al. [55] develop a fine-tuned version of CLIP which unlearns toxic concepts by redirecting their embeddings towards safe regions, so that retrieval always produces safe content even when the model is prompted with unsafe inputs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.672, + 0.909, + 0.793 + ], + "angle": 0, + "content": "In contrast, we propose an approach for managing unsafe content in VLMs: emphasizing awareness over unlearning. Rather than hiding the flaws of VLMs by ignoring NSFW content, we aim to equip the VLMs with the ability to distinguish between safe and unsafe content. This in turn helps users of the model to expose or redirect NSFW content when necessary, a crucial step toward improving user agency, understanding, and interpretability [19]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Inspired by recent hyperbolic vision-language models [17, 52], we introduce a hyperbolic framework that leverages the geometric properties of hyperbolic space to separate safe and unsafe content effectively. Using a paired dataset of safe and unsafe image-text inputs [55], we adjust the embeddings to create an entailment-based [24] structure. In this setup, safe concepts are positioned closer to the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.887, + 0.218, + 0.9 + ], + "angle": 0, + "content": "*Equal contribution" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4222" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.332 + ], + "angle": 0, + "content": "origin of the hyperbolic space, while unsafe concepts are mapped further away. Specifically, we introduce a hyperbolic safe-to-unsafe entailment mechanism that ensures safe content encompasses unsafe representations within conical regions, defining clear safety boundaries and safety traversals to dynamically adjust query embeddings along the hyperbolic space to promote safe retrievals or, alternatively, expose relevant unsafe content when necessary. This framework not only organizes data into safe and unsafe radius-based regions but also enables controlled movement within the space, allowing retrievals to favor safety as required. Experiments demonstrate that HySAC achieves clear improvements in safety awareness, retrieval performance, and NSFW content handling across multiple datasets, with robustness in both safe content redirection and controlled unsafe content accessibility." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.35, + 0.233, + 0.365 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.38, + 0.483, + 0.501 + ], + "angle": 0, + "content": "Unlearning in vision-language models. Unlearning concepts and content has recently received a lot of attention, empowered by the success of vision-language models. Various approaches have been explored, such as full model retraining, fine-tuning, machine unlearning [9, 27, 28, 56, 57], and differential privacy [29]. Some of these efforts have focused on text-to-image models, with the goal of removing specific styles, concepts, or objects [36, 76]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.503, + 0.483, + 0.728 + ], + "angle": 0, + "content": "A particular emphasis has been placed on removing NSFW content, which encompasses inappropriate, unsafe, or illegal material. Schramowsky et al. [64] steer the generation away from NSFW areas, defined by a fixed set of concepts. The embedding of NSFW concepts is applied as negative guidance during the text-conditioning phase, acting as safety guidance. Gandikota et al. [23] erase visual concepts using only their name, with negative guidance serving as a teacher. Poppi et al. [55] focus on removing NSFW concepts from CLIP-like models by fine-tuning the entire model using ViSU, a multimodal dataset containing Safe/NSFW Image/Texts quadruplets. Unlike earlier approaches, which target specific components, their method fine-tunes the entire vision-and-language model, making content removal applicable to downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.73, + 0.483, + 0.865 + ], + "angle": 0, + "content": "We introduce a method for handling NSFW concepts in CLIP-like models by fine-tuning them in hyperbolic space. We exploit the hierarchical properties of hyperbolic geometry, yielding a clear advantage: the model becomes explicitly aware of whether content is safe or NSFW, rather than merely removing knowledge of unsafe content. Similarly to our approach, safe generation works [8, 64], underscore that a deeper understanding of (un-)safety can improve the control of harmful content generation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "NSFW concept detection. A related area of research is the automatic detection of NSFW content. Several meth" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.302 + ], + "angle": 0, + "content": "ods have been proposed for detecting NSFW and toxic language [10, 31, 44], primarily in social media contexts. DistilBERT [62] has emerged as a promising model for this task, especially when fine-tuned for identifying adult content. Detecting inappropriate language presents a significant challenge, and this complexity extends to visual content, where various techniques have been developed to detect NSFW imagery [5, 22, 49]. In this domain, models like NudeNet [4] specialize in detecting nudity, while Q16 [63] serves as a broader classifier, capable of identifying a wider range of NSFW content. However, identifying inappropriate visual content remains a complex task, given the challenges posed by subtle visual cues, lack of contextual information, and limited data availability." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.303, + 0.905, + 0.394 + ], + "angle": 0, + "content": "While these detection methods focus solely on identifying unsafe content, they do not address the problem of retrieving relevant, safe alternatives when an unsafe input is detected. Our method makes it possible to jointly detect NSFW content and provide a mechanism to shift NSFW queries towards safe but relevant alternatives." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.397, + 0.905, + 0.547 + ], + "angle": 0, + "content": "Hyperbolic learning. A key advantage of hyperbolic space is its inherent ability to represent hierarchical or tree-like structures with minimal distortion [11, 50, 51, 61]. A comprehensive list of recent advancements is documented in surveys by Mettes et al. [46] and Peng et al. [53]. Foundational works for building neural networks in hyperbolic space [3, 24, 25, 35, 68] led to the use of hyperbolic models across multiple modalities such as images [2, 20, 21, 26, 42, 71], text [18, 38, 70, 77], graphs [12-14, 41] and recommender systems [48, 73]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.548, + 0.905, + 0.728 + ], + "angle": 0, + "content": "Recent work has shown the strong potential of hyperbolic learning for vision-language models [17, 33, 52] and demonstrated that training with a loss function enforcing entailment cones [24] leads to the emergence of hierarchical structures between the embeddings. While Desai et al. [17] enforce entailment structure across modalities, Pal et al. [52] also explicitly enforce structure within modalities by leveraging object-level compositions and nouns from text. We take inspiration from these approaches and enforce entailment relations across safe and unsafe embeddings, creating an interpretable space for traversing from unsafe regions to safe regions in CLIP models." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.74, + 0.652, + 0.755 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.765, + 0.905, + 0.84 + ], + "angle": 0, + "content": "Throughout this work, we operate in hyperbolic space, a Riemannian manifold with a constant negative curvature. Following Desai et al. [17], we use the Lorentz model, as it is better equipped to deal with numerical instabilities associated with the Poincaré distance metric [39, 51]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.84, + 0.905, + 0.901 + ], + "angle": 0, + "content": "The Lorentz model \\(\\mathbb{L}^n\\) is an \\(n\\)-dimensional manifold in which points are represented on the upper sheet of a two-sheeted hyperboloid in \\((n + 1)\\)-dimensional Minkowski spacetime. Following terminology from general relativity," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4223" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.089, + 0.9, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.313, + 0.908, + 0.342 + ], + "angle": 0, + "content": "Figure 1. Overview of our approach. HySAC builds a hyperbolic embedding that manages content safety through an entailment hierarchy. Unsafe text and images are projected to dedicated regions of hyperbolic space, allowing for safety-aware retrieval and classification." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.357, + 0.484, + 0.42 + ], + "angle": 0, + "content": "for each vector \\(\\mathbf{p} \\in \\mathbb{R}^{n+1}\\), we refer to the first dimension, the axis of symmetry, as the time-axis, denoted by \\(p_0\\), and the remaining \\(n\\)-dimensions as the spatial components, denoted \\(\\tilde{\\pmb{p}}\\). The Lorentz model, \\(\\mathbb{L}^n = (\\mathcal{L}^n, \\mathfrak{g}_p^\\kappa)\\) is given as" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.428, + 0.484, + 0.469 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {n} := \\left\\{\\boldsymbol {p} \\in \\mathbb {R} ^ {n + 1}: \\langle \\mathbf {p}, \\mathbf {p} \\rangle_ {\\mathcal {L}} = - \\frac {1}{\\kappa}, p _ {0} = \\sqrt {1 / \\kappa + \\| \\tilde {\\boldsymbol {p}} \\| ^ {2}}, \\kappa > 0 \\right\\}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.47, + 0.484, + 0.532 + ], + "angle": 0, + "content": "where \\(-\\kappa \\in \\mathbb{R}\\) denotes the curvature with the Riemannian metric \\(\\mathfrak{g}_p^\\kappa = \\mathrm{diag}(-1,1,\\ldots ,1)\\). The Lorentzian inner product \\(\\langle .,.\\rangle_{\\mathcal{L}}\\) is induced by the metric tensor \\(\\mathfrak{g}_p^\\kappa\\) and is defined for \\(\\mathbf{p},\\mathbf{q}\\in \\mathbb{L}^n\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.196, + 0.538, + 0.482, + 0.555 + ], + "angle": 0, + "content": "\\[\n\\langle \\mathbf {p}, \\mathbf {q} \\rangle_ {\\mathcal {L}} = - p _ {0} q _ {0} + \\langle \\tilde {\\boldsymbol {p}}, \\tilde {\\boldsymbol {q}} \\rangle , \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.56, + 0.484, + 0.622 + ], + "angle": 0, + "content": "where \\(\\langle .,.\\rangle\\) is the Euclidean inner product. The Lorentzian inner product induces a norm on the Lorentzian space which can be written as \\(\\| \\mathbf{p}\\|_{\\mathcal{L}} = \\sqrt{\\langle\\mathbf{p},\\mathbf{p}\\rangle_{\\mathcal{L}}}\\). We now define the common hyperbolic operations in Lorentz space." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.627, + 0.484, + 0.672 + ], + "angle": 0, + "content": "Definition 3.1 (Lorentzian distance). The Lorentzian distance between two points in \\(\\mathbb{L}^n\\) is the length of their shortest path (geodesic) connecting them, computed as" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.678, + 0.482, + 0.697 + ], + "angle": 0, + "content": "\\[\nd _ {\\mathcal {L}} (\\mathbf {p}, \\mathbf {q}) = \\sqrt {1 / \\kappa} \\cdot \\cosh^ {- 1} (- \\kappa \\langle \\mathbf {p}, \\mathbf {q} \\rangle_ {\\mathcal {L}}). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.706, + 0.484, + 0.737 + ], + "angle": 0, + "content": "In our work, we use the negative of Lorentzian distance to calculate similarities between multimodal inputs [17, 52]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.742, + 0.484, + 0.866 + ], + "angle": 0, + "content": "Definition 3.2 (Exponential map). Since Lorentz space is a Riemannian manifold, it is locally Euclidean. This is best described through the tangent space \\( T_{\\mathbf{p}}\\mathbb{L}^n \\), a first-order approximation of the Lorentzian manifold at a given point \\( \\mathbf{p}\\in \\mathbb{L}^n \\). The exponential map then provides a means to project elements from the tangent space onto the hyperboloid. Given a point \\( \\mathbf{v}\\in T_{\\mathbf{p}}\\mathbb{L}^n \\), the exponential map is defined as \\( \\exp_{\\mathbf{p}}^{\\kappa}\\colon T_{\\mathbf{p}}\\mathbb{L}^{n}\\to \\mathbb{L}^{n} \\) with the expression" + }, + { + "type": "equation", + "bbox": [ + 0.109, + 0.87, + 0.484, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\exp_ {\\mathbf {p}} ^ {\\kappa} (\\mathbf {v}) = \\cosh (\\sqrt {\\kappa} \\| \\mathbf {v} \\| _ {\\mathbb {L}}) \\mathbf {p} + \\frac {\\sinh (\\sqrt {\\kappa} \\| \\mathbf {v} \\| _ {\\mathbb {L}})}{\\sqrt {\\kappa} \\| \\mathbf {v} \\| _ {\\mathbb {L}}} \\mathbf {v}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.367, + 0.907, + 0.458 + ], + "angle": 0, + "content": "In practice, the reference point \\(\\mathbf{p}\\) is set to the origin \\(\\mathbf{0} = (\\sqrt{1 / \\kappa}, 0, \\dots, 0)^T\\) on the hyperboloid, allowing \\(\\exp_0^\\kappa\\) to project Euclidean vectors from the tangent space at \\(\\mathbf{0}\\) directly onto the hyperboloid [17, 35]. In this work, the exponential map is used to project the outputs of the visual and textual encoders to a shared hyperbolic space." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.47, + 0.876, + 0.487 + ], + "angle": 0, + "content": "4. HySAC: Hyperbolic Safety-Aware CLIP" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.495, + 0.819, + 0.511 + ], + "angle": 0, + "content": "4.1. Problem formulation and objective" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.519, + 0.907, + 0.656 + ], + "angle": 0, + "content": "Problem setup. Given a dataset \\( D = \\{(I_i, T_i)\\}_{i=1}^N \\) of \\( N \\) image-text pairs, vision-language models (e.g. CLIP [58]) align the visual and textual embeddings obtained from image and text encoders in a shared embedding space. Large-scale datasets employed for training such embedding spaces are often web-scraped and contain unsafe samples [5]. For our problem setup, in order to differentiate between safe and unsafe content, we denote safe image-text pairs in \\( D \\) as \\( (I_i, T_i) \\) and unsafe image-text pairs as \\( (I_i^\\star, T_i^\\star) \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.657, + 0.909, + 0.853 + ], + "angle": 0, + "content": "To make vision-language models aware of inappropriate contents, and enable them to avoid or redirect the representation of such content, we also require a dataset of quadruplets of safe and unsafe image-text pairs, denoted as \\( D^{\\star} = \\{(I_k, T_k, I_k^{\\star}, T_k^{\\star})\\}_{k=1}^{K} \\) [55]. This dataset is generated following the definition of NSFW content of Schramowski et al. [64] containing the following twenty categories: hate, harassment, violence, suffering, humiliation, harm, suicide, sexual, nudity, bodily fluids, blood, obscene gestures, illegal activity, drug use, theft, vandalism, weapons, abuse, brutality, and cruelty. The dataset is constructed such that the unsafe image-text pairs \\( (I_k^{\\star}, T_k^{\\star}) \\) are specific cases or modified versions of the safe representations \\( (I_k, T_k) \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.901 + ], + "angle": 0, + "content": "Modelling relations in the hyperbolic space. We first consider the relationship between text and image modalities in the embedding space. Like other hyperbolic vision-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4224" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.184 + ], + "angle": 0, + "content": "language models [17, 52], we consider text as a general version of images to reflect the natural structure of partial order embeddings [72]. This is enforced in the embedding space by placing text embeddings closer to the origin and image embeddings farther away, defining a modality entailment relationship formally expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.192, + 0.483, + 0.21 + ], + "angle": 0, + "content": "\\[\ng _ {T} \\left(T _ {k}\\right) \\ll g _ {I} \\left(I _ {k}\\right), \\quad \\text {a n d} \\quad g _ {T} \\left(T _ {k} ^ {\\star}\\right) \\ll g _ {I} \\left(I _ {k} ^ {\\star}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.218, + 0.484, + 0.339 + ], + "angle": 0, + "content": "where \\( g_{I} \\) and \\( g_{T} \\) denote the projections of images and text in the hyperbolic model and \\( \\ll \\) indicates that one embedding is closer to the origin than another. Differently from other models, we also impose a relationship between unsafe and safe pairs, by considering unsafe image-text pairs as specific cases of their safe counterparts within \\( D^{\\star} \\). Hence, we establish a safety entailment to segregate safe from unsafe content, as follows" + }, + { + "type": "equation", + "bbox": [ + 0.222, + 0.349, + 0.483, + 0.367 + ], + "angle": 0, + "content": "\\[\ng _ {I} \\left(I _ {k}\\right) \\ll g _ {T} \\left(T _ {k} ^ {\\star}\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.375, + 0.484, + 0.452 + ], + "angle": 0, + "content": "By satisfying these inequalities, \\( g_{I} \\) and \\( g_{T} \\) capture both the modal and safety hierarchies within the data, thereby endowing the embedding space with safety-aware properties. To sum up, our objective is to model a hyperbolic vision-language model with the following inequality chain:" + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.46, + 0.483, + 0.478 + ], + "angle": 0, + "content": "\\[\ng _ {T} \\left(T _ {k}\\right) \\ll g _ {I} \\left(I _ {k}\\right) \\ll g _ {T} \\left(T _ {k} ^ {\\star}\\right) \\ll g _ {I} \\left(I _ {k} ^ {\\star}\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.486, + 0.484, + 0.517 + ], + "angle": 0, + "content": "Below we outline our method to create safety-aware vision-language models in hyperbolic space." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.525, + 0.332, + 0.542 + ], + "angle": 0, + "content": "4.2. Hyperbolic safety learning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.547, + 0.484, + 0.654 + ], + "angle": 0, + "content": "To optimize Eq. 7 for safety-aware vision-language models, we propose Hyperbolic Safety-Aware CLIP to rearrange the embedding space to separate safe and unsafe regions. Our optimization consists of two components: (1) a hyperbolic safety contrastive component to align image-text pairs over a mini-batch and (2) a hyperbolic safety entailment to align safe and unsafe content. An overview is given in Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.484, + 0.871 + ], + "angle": 0, + "content": "Hyperbolic safety contrastive learning. CLIP [58] and Safe-CLIP [55] rely on contrastive objectives to align and distribute the multimodal data. We utilize hyperbolic embeddings to align the visual and textual data. Specifically, we project the visual and textual embeddings from a pre-trained vision-language model onto a hyperboloid [17] through an exponential map (Eq. 4). Let \\( f_{I}(\\cdot) \\) and \\( f_{T}(\\cdot) \\) represent any Euclidean encoders for image and text. Then, \\( g_{I}(I_{k}) = \\exp_{0}^{\\kappa}(\\alpha_{img} \\cdot f_{I}(I_{k})) \\) and \\( g_{T}(T_{k}) = \\exp_{0}^{\\kappa}(\\alpha_{txt} \\cdot f_{T}(T_{k})) \\) represent the hyperbolic representations of a safe image-text pair, \\( (I_{k}, T_{k}) \\). Similarly, \\( g_{I}(I_{k}^{\\star}) = \\exp_{0}^{\\kappa}(\\alpha_{img} \\cdot f_{I}(I_{k}^{\\star})) \\) and \\( g_{T}(T_{k}^{\\star}) = \\exp_{0}^{\\kappa}(\\alpha_{txt} \\cdot f_{T}(T_{k}^{\\star})) \\) represent the hyperbolic representations of an unsafe image-text pair \\( (I_{k}^{\\star}, T_{k}^{\\star}) \\). \\( \\alpha_{img} \\) and \\( \\alpha_{txt} \\) are learnable projection scalars." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "To align representations in hyperbolic space, the similarity for the image-text and text-image contrastive loss is" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.138 + ], + "angle": 0, + "content": "based on negative Lorentzian distance (Eq. 3) between \\( g_{I}(.) \\) and \\( g_{T}(.) \\). We compute the hyperbolic safety contrastive loss over safe image-text pairs \\( (I_k, T_k) \\) in a batch \\( B \\) as" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.148, + 0.906, + 0.186 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {c o n t}} ^ {*} (I, T) = - \\sum_ {i \\in B} \\log \\frac {\\exp \\left(d _ {\\mathcal {L}} \\left(g _ {I} \\left(I _ {i}\\right) , g _ {T} \\left(T _ {i}\\right)\\right) / \\tau\\right)}{\\sum_ {k = 1 , k \\neq i} ^ {B} \\exp \\left(d _ {\\mathcal {L}} \\left(g _ {I} \\left(I _ {i}\\right) , g _ {T} \\left(T _ {k}\\right)\\right) / \\tau\\right)}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.187, + 0.907, + 0.277 + ], + "angle": 0, + "content": "where \\(\\tau\\) denotes a temperature hyperparameter. A similar contrastive loss is employed to preserve the multimodal structure between unsafe image-text pairs \\((I_k^{\\star}, T_k^{\\star})\\). Two additional contrastive losses between cross-safety modalities ensure the alignment of quadruplets in the embedding space. The final contrastive loss is formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.531, + 0.287, + 0.905, + 0.324 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {\\mathrm {h S C}} (I, T, I ^ {\\star}, T ^ {\\star}) = L _ {\\text {c o n t}} ^ {\\star} (I, T) + L _ {\\text {c o n t}} ^ {\\star} \\left(I ^ {\\star}, T ^ {\\star}\\right) \\tag {9} \\\\ + L _ {c o n t} ^ {\\star} (I, T ^ {\\star}) + L _ {c o n t} ^ {\\star} (I ^ {\\star}, T). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.346, + 0.907, + 0.468 + ], + "angle": 0, + "content": "Hyperbolic safety entailment learning. Hyperbolic entailment cones, introduced by Ganea et al. [24], generalize partial ordered embeddings [72] to any Riemannian manifold. Entailment cones induce a partial order between concepts in a dataset \\(\\mathcal{X}\\) such that for any pair \\((\\mathbf{p},\\mathbf{q})\\in \\mathcal{X}\\), if \\(\\mathbf{p}\\) is a sub-concept of \\(\\mathbf{q}\\), then \\(\\mathbf{q}\\) entails \\(\\mathbf{p}\\) within a conical region \\(\\mathfrak{S}_{\\mathbf{q}}\\) defined by \\(\\mathbf{q}\\). For the Lorentz model, \\(\\mathbb{L}^n\\), the half-aperture of each conical region \\(\\mathfrak{S}_{\\mathbf{q}}\\) is defined as [17, 38]" + }, + { + "type": "equation", + "bbox": [ + 0.616, + 0.478, + 0.906, + 0.513 + ], + "angle": 0, + "content": "\\[\n\\omega (\\mathbf {q}) = \\sin^ {- 1} \\left(\\frac {2 K}{\\sqrt {\\kappa} \\| \\tilde {\\boldsymbol {q}} \\|}\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.522, + 0.907, + 0.705 + ], + "angle": 0, + "content": "where \\(-\\kappa\\) is the curvature of space, and the constant \\(K = 0.1\\) limits values near the origin [24]. To preserve partial order between image-text relationships, we add entailment from safe text to safe image and from unsafe text to unsafe images, effectively implementing Eq. 5. Specifically, a safe image \\(I_{k}\\) must lie within the cone defined by its corresponding safe text, \\(T_{k}\\), characterized by the half-aperture \\(\\omega (T_k)\\). Similarly, an unsafe image \\(I_{k}^{\\star}\\) must lie within the cone defined by its corresponding unsafe text \\(T_{k}^{\\star}\\). This is enforced through an entailment loss formulated for image-text representations by Le et al. [38] and Desai et al. [17], as" + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.714, + 0.905, + 0.751 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\begin{array}{l} L _ {e n t} ^ {\\star} (I, T) = \\max \\left(0, \\phi \\left(I _ {k}, T _ {k}\\right) - \\eta \\omega \\left(T _ {k}\\right)\\right) \\text {a n d} \\\\ I _ {*} ^ {\\star} \\left(I _ {*} ^ {\\star}, T _ {*} ^ {\\star}\\right) = \\left(0 + \\left(I _ {*} ^ {\\star}, T _ {*} ^ {\\star}\\right) - \\left(T _ {*} ^ {\\star}\\right)\\right) \\end{array} \\tag {11} \\\\ L _ {e n t} ^ {\\star} \\left(I ^ {\\star}, T ^ {\\star}\\right) = \\max \\left(0, \\phi \\left(I _ {k} ^ {\\star}, T _ {k} ^ {\\star}\\right) - \\eta \\omega \\left(T _ {k} ^ {\\star}\\right)\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.761, + 0.906, + 0.793 + ], + "angle": 0, + "content": "where \\(\\phi\\) is the exterior angle (between lines \\(I_{k}T_{k}\\) and \\(\\mathbf{0}T_{k}\\) or between lines \\(I_k^\\star T_k^\\star\\) and \\(\\mathbf{0}T_k^\\star\\)) given by" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.804, + 0.906, + 0.855 + ], + "angle": 0, + "content": "\\[\n\\phi \\left(I _ {k}, T _ {k}\\right) = \\cos^ {- 1} \\left(\\frac {I _ {k 0} + T _ {k 0} \\kappa \\left\\langle I _ {k} , T _ {k} \\right\\rangle_ {\\mathcal {L}}}{\\left\\| \\widetilde {T _ {k}} \\right\\| \\sqrt {\\left(\\kappa \\left\\langle I _ {k} , T _ {k} \\right\\rangle_ {\\mathcal {L}}\\right) ^ {2} - 1}}\\right). \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.906, + 0.903 + ], + "angle": 0, + "content": "Here, \\(\\eta\\) is a threshold for the half-aperture, \\(\\omega (T_{k})\\) [52]. Intuitively, the entailment loss \\(L_{ent}^{\\star}\\) penalizes images \\(I_{k}\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4225" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.228 + ], + "angle": 0, + "content": "that lie outside the cone \\(\\mathfrak{S}_{T_k}\\) defined by their corresponding text caption \\(T_{k}\\). Finally, to model the safety hierarchy, we enforce that safe concepts entail unsafe ones, reflecting that safe data are more general and unsafe data are more specific. Specifically, we enforce that a safe image \\(I_{k}\\) entails an unsafe text \\(T_{k}^{\\star}\\), meaning that the unsafe text \\(T_{k}^{\\star}\\) must lie within the cone defined by the safe image \\(I_{k}\\), characterized by the half-aperture \\(\\omega (I_k)\\). This gives us the safety-entailment defined in Eq. 6. This is implemented as" + }, + { + "type": "equation", + "bbox": [ + 0.125, + 0.235, + 0.483, + 0.252 + ], + "angle": 0, + "content": "\\[\nL _ {e n t} ^ {\\star} \\left(T ^ {\\star}, I\\right) = \\max \\left(0, \\phi \\left(T _ {k} ^ {\\star}, I _ {k}\\right) - \\eta \\omega \\left(I _ {k}\\right)\\right). \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.258, + 0.466, + 0.273 + ], + "angle": 0, + "content": "The overall entailment loss to satisfy Eq. 7 is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.092, + 0.28, + 0.483, + 0.31 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {h S E}} (I, T, I ^ {\\star}, T ^ {\\star}) = L _ {e n t} ^ {\\star} (I, T) + L _ {e n t} ^ {\\star} \\left(T ^ {\\star}, I\\right) + L _ {e n t} ^ {\\star} \\left(I ^ {\\star}, T ^ {\\star}\\right). \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.328, + 0.484, + 0.372 + ], + "angle": 0, + "content": "Combined loss function. We integrate the contrastive with the entailment losses to obtain the total loss used to fine-tune the model on the dataset \\(D^{\\star}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.092, + 0.38, + 0.484, + 0.41 + ], + "angle": 0, + "content": "\\[\nL (I, T, I ^ {\\star}, T ^ {\\star}) = L _ {\\mathrm {h S C}} (I, T, I ^ {\\star}, T ^ {\\star}) + L _ {\\mathrm {h S E}} (I, T, I ^ {\\star}, T ^ {\\star}). \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.411, + 0.484, + 0.546 + ], + "angle": 0, + "content": "Our proposal allows the model to differentiate between safe and unsafe embeddings based on their distance from the origin. Safe content is closer to the center, while NSFW content is farther away. This geometric arrangement not only enables the model to detect unsafe content but also allows dynamic manipulation of embeddings. NSFW queries can be redirected toward the safe region, effectively retrieving outputs that prioritize safety and providing more precise control over content retrieval." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.554, + 0.373, + 0.57 + ], + "angle": 0, + "content": "4.3. Safety traversals and evaluation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.577, + 0.484, + 0.759 + ], + "angle": 0, + "content": "The hyperbolic safety-aware training results in a restructuring of the shared embedding space of the vision-language model. To obtain safe but relevant retrieval outputs from unsafe queries or vice-versa, we introduce a traversal mechanism to adjust query embeddings in hyperbolic space, enhancing their similarity with either safe or unsafe content, depending on the retrieval task. This traversal involves moving the query embeddings along the line connecting them to the origin of the hyperboloid, altering their hyperbolic distance from the root. By adjusting the embeddings' positions, we align them with regions in the embedding space that correspond to the desired content type." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.76, + 0.484, + 0.866 + ], + "angle": 0, + "content": "Traversal Definition. Given an embedding \\(\\mathbf{q}\\), our method computes the distance from a predefined root feature \\(\\mathbf{r}\\) in hyperbolic space using the Lorentzian distance function \\(d_{\\mathcal{L}}(\\mathbf{q},\\mathbf{r})\\). For each type of content \\(X\\in \\{T,I,T^{\\star},I^{\\star}\\}\\), we compute the mean distance \\(\\mu_X\\) from the root feature \\(\\mathbf{r}\\) based on the distribution of each category. The boundary for each type is then defined as" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.872, + 0.483, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\tau_ {X} = \\mu_ {X} + \\tanh \\left(\\frac {\\mu_ {X} - \\alpha}{\\kappa}\\right) + 1, \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.228 + ], + "angle": 0, + "content": "where \\(\\kappa\\) is the negative curvature, and \\(\\alpha\\) is a constant set empirically to 0.8. This shift accounts for the curvature of space, ensuring the boundaries are appropriately adjusted for effective traversal. Defining four bounds allows more nuanced control over traversal depending on the retrieval task. To retrieve a content type \\(X\\), the query is moved along the Euclidean direction vector \\(\\mathbf{v}_{\\mathrm{dir}} = \\mathbf{q} - \\mathbf{r}\\) toward the root feature \\(\\mathbf{r}\\) until it reaches the corresponding boundary \\(\\tau_{X}\\) (e.g. \\(\\tau_{T}\\) for safe text). The target position \\(\\mathbf{q}^{*}\\) is given as" + }, + { + "type": "equation", + "bbox": [ + 0.636, + 0.237, + 0.907, + 0.268 + ], + "angle": 0, + "content": "\\[\n\\mathbf {q} ^ {*} = \\mathbf {r} + \\tau_ {X} \\cdot \\frac {\\mathbf {v} _ {\\mathrm {d i r}}}{\\| \\mathbf {v} _ {\\mathrm {d i r}} \\|} \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.276, + 0.907, + 0.307 + ], + "angle": 0, + "content": "allowing the embeddings to be repositioned to match the target content type while maintaining semantic alignment." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.318, + 0.646, + 0.335 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.343, + 0.674, + 0.359 + ], + "angle": 0, + "content": "5.1. Training Details" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.366, + 0.906, + 0.44 + ], + "angle": 0, + "content": "Datasets. Our experiments are mainly conducted on the ViSU dataset [55], containing 165k quadruplets of safe and unsafe image-text pairs. We also evaluate our model on three real-world NSFW image datasets: NudeNet [4], NSFW data source URLs\\(^1\\) and SMID [15]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.441, + 0.906, + 0.546 + ], + "angle": 0, + "content": "Baselines. Our safety comparisons include the original CLIP [58] and the state-of-the-art Safe-CLIP [55]. CLIP was trained using a private dataset of 400M image-text pairs [66] which has unsafe data [5]. Safe-CLIP is finetuned on the ViSU dataset [55] to redirect unsafe content to safe correspondent one via contrastive losses and cosine similarities, aiming to unlearn NSFW concepts." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.547, + 0.906, + 0.621 + ], + "angle": 0, + "content": "Models. Our visual and textual encoders are the same as CLIP [58], with VIT-L/14 as visual encoder, to maintain fair comparison to Safe-CLIP [55]. During training, both the visual and textual encoder are fine-tuned using low-rank decomposition [32] with low-rank factor \\( r = 16 \\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.622, + 0.906, + 0.728 + ], + "angle": 0, + "content": "**Optimization.** We use AdamW [43] with weight decay 0.2 and \\((\\beta_{1},\\beta_{2}) = (0.9,0.98)\\). We disable weight decay for all gains, biases, and learnable scalars. The model is finetuned for 20 epochs with batch size 256. The maximum learning rate is \\(8\\times 10^{-4}\\). We use mixed precision [47] to accelerate training, except computing exponential map and losses for HySAC in FP32 precision for numerical stability." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.728, + 0.906, + 0.849 + ], + "angle": 0, + "content": "Initialization. We initialize image and text encoders akin to CLIP, along with pre-trained weights. We initialize the softmax temperature as \\(\\tau = 0.07\\) and clamp it to a minimum value of 0.01. For HySAC, we initialize the learnable projection scalars \\(\\alpha_{img} = \\alpha_{txt} = 1 / \\sqrt{512}\\), the curvature parameter \\(c = 1.0\\) and clamp it in [0.1, 10.0] to prevent training instability. All scalars are learned in logarithmic space as \\(\\log(1 / \\tau)\\), \\(\\log(c)\\), \\(\\log(\\alpha_{img})\\) and \\(\\log(\\alpha_{txt})\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.849, + 0.906, + 0.879 + ], + "angle": 0, + "content": "Further details on the training setup are provided in the supplementary B." + }, + { + "type": "page_footnote", + "bbox": [ + 0.532, + 0.888, + 0.812, + 0.9 + ], + "angle": 0, + "content": "\\(^{1}\\)https://github.com/EBazarov/nsfw_data_source_url" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4226" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.904, + 0.235 + ], + "angle": 0, + "content": "
ModelText-to-Image (T-to-I)Image-to-Text (I-to-T)Text-to-Image (T*-to-I ∪ I*)Image-to-Text (I*-to-T ∪ T*)
R@1R@10R@20R@1R@10R@20R@1R@10R@20R@1R@10R@20
CLIP [58]36.871.681.539.874.283.52.024.833.24.632.940.6
MERU [17]14.943.054.214.742.353.82.215.221.54.422.629.4
HyCoCLIP [52]34.371.280.634.471.382.22.825.333.28.237.845.7
Safe-CLIP [55]45.981.889.745.382.389.88.046.958.019.162.971.1
MERU*50.084.191.151.285.392.32.339.949.45.747.954.7
HyCoCLIP*47.781.989.146.782.790.41.532.742.36.945.253.6
HySAC49.884.190.748.284.291.230.562.871.842.173.379.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.238, + 0.905, + 0.281 + ], + "angle": 0, + "content": "Table 1. Safe content retrieval performance on ViSU test set. Across all tasks and recall rates, HySAC improves over existing safety unlearning CLIP and hyperbolic CLIP models, highlighting that our approach is able to navigate unsafe image or text inputs towards relevant but safe retrieval outputs. * CLIP fine-tuned in hyperbolic space on ViSU training set with MERU/HyCoCLIP losses." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.291, + 0.905, + 0.403 + ], + "angle": 0, + "content": "
ModelText-to-Image (T* -to-I*)Image-to-Text (I* -to-T*)Text-to-Image (T* -to-I* ∪ I)Image-to-Text (I* -to-T* ∪ T)
R@1R@10R@20R@1R@10R@20R@1R@10R@20R@1R@10R@20
CLIP [58]73.194.997.672.895.297.768.492.395.967.193.396.7
MERU [17]29.462.472.225.857.767.823.554.064.319.551.161.2
HyCoCLIP [52]69.593.195.865.091.195.063.789.793.755.288.092.7
Safe-CLIP [55]58.086.291.456.085.191.047.780.085.832.177.184.6
HySAC81.498.499.482.297.899.281.198.499.480.597.298.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.406, + 0.905, + 0.435 + ], + "angle": 0, + "content": "Table 2. Unsafe content retrieval performance on ViSU test set. Akin to safe content retrieval, our approach performs best. This is a result of our objective, as we assign different content to different regions, enabling us to maintain valuable safety information." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.45, + 0.295, + 0.466 + ], + "angle": 0, + "content": "5.2. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.472, + 0.483, + 0.548 + ], + "angle": 0, + "content": "To assess the performance of our proposed model, HySAC, we measure its safety awareness and its ability to handle unsafe content effectively, while retaining both safe and unsafe knowledge. In the supplementary D, we report the zero-shot generalization of our method." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.556, + 0.33, + 0.571 + ], + "angle": 0, + "content": "5.2.1. Safety retrieval comparison" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.575, + 0.483, + 0.725 + ], + "angle": 0, + "content": "We evaluate our model on the capability of retrieving safe and unsafe items, in comparison to CLIP [58] and Safe-CLIP [55]. We also provide a comparison to recent hyperbolic VLMs, namely MERU [17] and HyCoCLIP [52]. Safe-CLIP and HySAC fine-tuned from CLIP on ViSU [55], MERU trained on RedCaps [16], and HyCoCLIP on GRIT [54]. For a fair comparison, we additionally fine-tuned the CLIP model on the ViSU dataset in hyperbolic space, using MERU and HyCoCLIP losses. All the models are evaluated on the ViSU test set." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.726, + 0.483, + 0.877 + ], + "angle": 0, + "content": "Retrieval tasks are defined as text-to-image and image-to-text, where the goal is to find the most relevant counterpart for a given query. Recall@K measures the fraction of queries where the correct item appears in the top-K retrieved results. To assess the safe retrieval performance of HySAC, we measure recall exclusively on safe content in both visual and textual elements (\\(T\\)-to-\\(I\\) and \\(I\\)-to-\\(T\\)). This step is crucial to verify that the original CLIP model's retrieval capabilities are retained after finetuning in hyperbolic space with our training method. Then, to evaluate" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.45, + 0.905, + 0.557 + ], + "angle": 0, + "content": "the safety-awareness capabilities of HySAC, we introduce a distinct setup in which NSFW elements are used as queries, while the retrievable items include both safe and unsafe elements (\\(T^{*}\\)-to-\\(I \\cup I^{*}\\) and \\(I^{*}\\)-to-\\(T \\cup T^{*}\\)). During these experiments, a retrieval is deemed correct only if the query retrieves its safe counterpart, thereby validating the model's ability to redirect unsafe queries towards safe items." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.562, + 0.906, + 0.73 + ], + "angle": 0, + "content": "When retrieving with HySAC, the threshold \\(\\tau\\) for moving query embeddings is computed using the mean distance of safe embeddings from the origin, adjusting the query embedding towards the safe region. Results are reported in Table 1, where we observe that HySAC consistently improves over both unlearning and existing hyperbolic models and features the highest recalls across all settings and rates. CLIP hyperbolic models finetuned on ViSU data (MERU\\* and HyCoCLIP\\*) perform well on safe-only retrieval, while our method achieves high performance on both safe-only and unsafe-safe retrieval, due to our safety-aware design." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In Table 2, we instead perform analyses for unsafe content retrieval. First, this involves the text-to-image and image-to-text retrieval on only unsafe elements (\\(T^{*}\\)-to-\\(I^{*}\\) and \\(I^{*}\\)-to-\\(T^{*}\\)). Second, instead, we perform retrieval by using unsafe elements as queries and both safe and unsafe items as retrievable items (\\(T^{*}\\)-to-\\(I^{*} \\cup I\\) and \\(I^{*}\\)-to-\\(T^{*} \\cup T\\)), and deem the retrieval correct only if the query retrieves its corresponding unsafe one. This setup tests the model's ability to function as a content moderator and also showcases its capacity to provide user autonomy in content retrieval decisions. In these tests, the traversal mechanism in" + }, + { + "type": "page_footnote", + "bbox": [ + 0.108, + 0.887, + 0.473, + 0.901 + ], + "angle": 0, + "content": "2The box data needed for HyCoCLIP was extracted using Kosmos-2." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4227" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.482, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.319, + 0.483, + 0.388 + ], + "angle": 0, + "content": "Figure 2. Distributions of embedding distances from the root. We embed all ViSU training samples and visualize their distance distribution from the root. While CLIP and Safe-CLIP do not separate between texts and images, MERU does. HySAC, instead, also differentiates between safe and unsafe content." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.396, + 0.481, + 0.468 + ], + "angle": 0, + "content": "
Model(T-to-I)(I-to-T)(T*-to-I ∪ I*)(I*-to-T ∪ T*)
R@1R@10R@1R@10R@1R@10R@1R@10
w/o Ent52.384.950.884.74.149.05.564.5
w/o S-Ent51.084.249.884.31.439.17.463.7
HySAC49.884.148.284.230.562.842.173.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.473, + 0.483, + 0.515 + ], + "angle": 0, + "content": "Table 3. Ablation study on loss components. We evaluate HySAC against two ablations that remove loss components. Results are in the same setting of Table 1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.525, + 0.484, + 0.631 + ], + "angle": 0, + "content": "HySAC uses adjusted parameters to move in the unsafe direction, targeting the retrieval of NSFW content. Here too, HySAC achieves the best recall across all settings, demonstrating that HySAC not only prioritizes safety by navigating away from NSFW content when required but also ensures that users can access NSFW content under controlled conditions, better than existing competitors." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.64, + 0.273, + 0.655 + ], + "angle": 0, + "content": "5.2.2. Analysis of HySAC" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.483, + 0.78 + ], + "angle": 0, + "content": "Assessing HySAC embedding space. Further, we validate the organization of the embedding space as outlined in Equation 7. Our goal is to confirm that the embeddings for safe content are positioned closer to the origin of hyperbolic space, while those for unsafe content are further away, following the proposed hierarchy. Specifically, for each \\( X \\in \\{T, I, T^{\\star}, I^{\\star}\\} \\) of the training-set of \\( D^{\\star} \\), we compute the distances \\( d_{\\mathcal{L}}(X, \\mathbf{r}) \\) from the root feature \\( \\mathbf{r} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.484, + 0.902 + ], + "angle": 0, + "content": "A visualization is reported in Figure 2, where we show the distribution of embeddings in terms of their distance to the root feature. The comparative analysis is done across four different models: HySAC, CLIP, Safe-CLIP, and MERU. For both the hyperbolic models the root feature is the origin of the hyperboloid. For the Euclidean models, since the origin does not lie on the hypersphere, the root is empirically estimated as the embedding that has the least" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.089, + 0.905, + 0.162 + ], + "angle": 0, + "content": "
Model% Safe (Text-to-Image)% Safe (Image-to-Text)
NudeNetNSFW URLsSMIDNudeNetNSFW URLsSMID
CLIP78.279.755.233.344.059.1
Safe-CLIP92.692.683.475.276.465.6
HySAC96.293.980.184.495.197.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.168, + 0.907, + 0.223 + ], + "angle": 0, + "content": "Table 4. Retrieval performance on real NSFW images. Rate of safe images retrieved using unsafe prompts from the ViSU test set. The retrievable set includes safe and unsafe real images, with the latter from LAION-400M and the former from NSFW sources." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.234, + 0.905, + 0.34 + ], + "angle": 0, + "content": "distance from all embeddings of the training set, i.e. the \\(\\ell_2\\) normalization of the average of all embeddings. As it can be seen, the distribution clearly shows four peak distributions for HySAC, each one representing one of the \\(X\\) content types, elucidating the efficacy of our approach in maintaining a clear separation between safe, unsafe, textual, and visual content within the embedding space." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.34, + 0.907, + 0.552 + ], + "angle": 0, + "content": "Ablation Study. In Table 3, we validate the effectiveness of the key components in HySAC, by comparing its full configuration with variants where specific losses are disabled. In particular, we employ one variant which only keeps contrastive losses (denoted as \"w/o Ent\") and one that omits the safety-entailment loss (\"w/o S-Ent\"). Results show that while removing these components slightly improves performance in scenarios involving only safe content, likely due to reduced spatial constraints, their absence significantly undermines the model's effectiveness in dealing with unsafe content, especially in unsafe-to-safe retrieval. These results underscore the essential roles that both the modality- and safety-entailment losses play in enhancing the safety awareness of the proposed model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.906, + 0.718 + ], + "angle": 0, + "content": "Retrieval on real NSFW datasets. To further analyze the safety of HySAC, we conduct retrieval tests using NudeNet [4], NSFW data source URLs and SMID [15]. The first two datasets primarily contain nudity and pornographic content, whereas SMID includes a broader range of inappropriate content, such as violence, harm, and discrimination. We randomly select 1000 images from each dataset to serve as visual elements and use 5k NSFW captions from the ViSU test set. Both the image-to-text and text-to-image retrieval tasks also incorporate 10,000 randomly chosen retrievable safe items from LAION-400M [66]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Results, displayed in Table 4, contrast the performance of HySAC with that of CLIP [58] and Safe-CLIP [55]. We report the proportion of safe retrieval outputs when NSFW queries are used, highlighting the capability of HySAC to enhance safety in retrieval results. Traversing the embedding space towards safety prioritizes safer alternatives, which may adjust relevance in some cases. Due to the absence of datasets with real unsafe data and correlated safe alternatives, this aspect could not be evaluated. Notably, HySAC demonstrates improved performance in securing safer content compared to Safe-CLIP across most datasets for both text-to-image and image-to-text scenarios." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4228" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.152, + 0.099, + 0.258, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.109, + 0.188, + 0.296, + 0.229 + ], + "angle": 0, + "content": "A train filled with toxic chemicals crashing into a bridge, causing a massive explosion and destroying everything in its path." + }, + { + "type": "image_caption", + "bbox": [ + 0.153, + 0.24, + 0.258, + 0.266 + ], + "angle": 0, + "content": "The train is going over the bridge over a river." + }, + { + "type": "image_caption", + "bbox": [ + 0.185, + 0.276, + 0.221, + 0.287 + ], + "angle": 0, + "content": "[ROOT]" + }, + { + "type": "image", + "bbox": [ + 0.347, + 0.099, + 0.454, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.313, + 0.195, + 0.491, + 0.222 + ], + "angle": 0, + "content": "Several people on skis on a snowy slope, enjoying a drug-fueled party." + }, + { + "type": "image_caption", + "bbox": [ + 0.347, + 0.24, + 0.457, + 0.266 + ], + "angle": 0, + "content": "Several people in their ski gear are in the snow." + }, + { + "type": "image_caption", + "bbox": [ + 0.382, + 0.276, + 0.419, + 0.287 + ], + "angle": 0, + "content": "[ROOT]" + }, + { + "type": "image", + "bbox": [ + 0.547, + 0.099, + 0.652, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.511, + 0.187, + 0.692, + 0.228 + ], + "angle": 0, + "content": "A random Black man is standing in front of a double door with a gun in his hand and a look of hatred in his eyes." + }, + { + "type": "image_caption", + "bbox": [ + 0.529, + 0.24, + 0.674, + 0.266 + ], + "angle": 0, + "content": "A person is holding an umbrella while standing in the garage." + }, + { + "type": "image_caption", + "bbox": [ + 0.583, + 0.276, + 0.619, + 0.287 + ], + "angle": 0, + "content": "[ROOT]" + }, + { + "type": "image", + "bbox": [ + 0.749, + 0.099, + 0.854, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.716, + 0.187, + 0.89, + 0.212 + ], + "angle": 0, + "content": "Two naked women sitting next to each other while flashing their genitals." + }, + { + "type": "image_caption", + "bbox": [ + 0.768, + 0.22, + 0.834, + 0.23 + ], + "angle": 0, + "content": "Sultry women." + }, + { + "type": "image_caption", + "bbox": [ + 0.724, + 0.24, + 0.882, + 0.266 + ], + "angle": 0, + "content": "A couple of women sitting next to each other in front of a cake." + }, + { + "type": "image_caption", + "bbox": [ + 0.782, + 0.276, + 0.818, + 0.287 + ], + "angle": 0, + "content": "[ROOT]" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.309, + 0.906, + 0.339 + ], + "angle": 0, + "content": "Figure 3. Qualitative traversal results. HySAC traverses towards the root feature, retrieving the top-1 text at each interpolation point. This traversal effectively transitions from unsafe to safe captions, demonstrating the model's ability to ensure safety-aware content retrieval." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.357, + 0.481, + 0.464 + ], + "angle": 0, + "content": "
ModelNudeNetMixed NSFW
AccFPRFNRAccFPRFNR
NSFW-CNN [37]85.30.014.766.54.535.9
CLIP-classifier [65]97.30.02.776.90.111.0
CLIP-distance [59]86.40.013.677.82.022.1
NudeNet [4]91.20.08.876.94.524.6
Q16 [63]28.50.071.565.38.329.4
HySAC99.50.00.578.916.56.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.473, + 0.483, + 0.501 + ], + "angle": 0, + "content": "Table 5. NSFW classification. Comparison between HySAC and other NSFW classifiers. Metrics reported in percentages." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.51, + 0.483, + 0.691 + ], + "angle": 0, + "content": "Classifying NSFW content. The structure of the embedding space in HySAC also supports the classification of NSFW content. We evaluate this using the NudeNet [4] and Mixed NSFW datasets, comparing against classifiers such as NSFW-CNN [37], CLIP-classifier [65], CLIP-distance [59], NudeNet [4], and Q16 [63]. NudeNet only contains nudity, while Mixed NSFW includes different NSFW categories from various online sources and safe images from PASS [1]. We sample a 1,000-image subset from NudeNet and 442 images from Mixed NSFW, balanced between safe and unsafe. Further details of these datasets are provided in the supplementary C." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.692, + 0.483, + 0.767 + ], + "angle": 0, + "content": "Results in Table 5 show that HySAC achieves competitive or superior results in NSFW content classification, despite not being explicitly designed for safety classification. The norm threshold, set to the ViSU dataset mean (Figure 2), differentiates safe from unsafe content." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.776, + 0.357, + 0.79 + ], + "angle": 0, + "content": "5.2.3. Visualizing the safety traversals" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.856 + ], + "angle": 0, + "content": "We examine traversal paths for safe text retrievals, starting with an unsafe image embedding as the query. The traversals are along the geodesic of hyperbolic space from the image to the origin of the hyperboloid, denoted [ROOT]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "The input query is an unsafe image taken from ViSU test set and the text retrieval space consists of a mix of safe and unsafe captions of ViSU test set, metadata-based caption" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.36, + 0.907, + 0.512 + ], + "angle": 0, + "content": "from pexels . com, and a curated list of unsafe words3. To create visualization shown in Figure 3, each retrieved text output is selected only once across all interpolation points, ensuring unique retrievals. Results show that, with HySAC, as the query nears the origin, retrieved content shifts from unsafe to safe while preserving semantic relevance. This progression illustrates the model's capability to effectively navigate the embedding space along relevant paths. For more on the experimental setup and traversal visualizations, see the supplementary E." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.516, + 0.907, + 0.562 + ], + "angle": 0, + "content": "Other ablation studies. For additional ablations on embedding space geometry and hyperparameter evaluation, we refer the reader to the supplementary D." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.578, + 0.634, + 0.593 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.604, + 0.907, + 0.875 + ], + "angle": 0, + "content": "This paper introduces hyperbolic safety-aware vision-language models. Where recent literature focuses on removing or unlearning unsafe image-text content, we bring a perspective of awareness. By modelling unsafe image-text content as specific cases of their safe counterparts, we can divide the space into safe and unsafe regions. We show that hyperbolic space is a natural solution for this hierarchical relation, and propose a hyperbolic CLIP model with safety entailment learning and traversal. Our approach not only results in better retrieval of relevant safe outputs given unsafe inputs but also provides more robustness and comes with an NSFW classifier as a free by-product. Dealing with NSFW data in vision-language models is an important open research problem with real-world implications, from ethical to legal and societal concerns. By opting for awareness, we find that safety recognition improves. Given the importance of the ethical implications of this work, we provide a thorough discussion in the supplementary A." + }, + { + "type": "page_footnote", + "bbox": [ + 0.532, + 0.887, + 0.676, + 0.9 + ], + "angle": 0, + "content": "\\(^{3}\\)github.com/LDNOOBW/" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4229" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.26, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.486, + 0.268 + ], + "angle": 0, + "content": "We acknowledge the CINECA award under the ISCRA initiative, for the availability of high-performance computing resources. This work has been supported by the EU Horizon projects \"ELIAS - European Lighthouse of AI for Sustainability\" (No. 101120237) and \"European Lighthouse on Safe and Secure AI (ELSA)\" (No. 101070617), co-funded by the European Union. Tejaswi Kasarla also acknowledges travel support from the European Union's Horizon research and innovation programme under grant agreement No. 951847 (ELISE) and No. 101120237 (ELIAS)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.278, + 0.188, + 0.294 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.304, + 0.484, + 0.358 + ], + "angle": 0, + "content": "[1] Yuki M. Asano, Christian Rupprecht, Andrew Zisserman, and Andrea Vedaldi. Pass: An imagenet replacement for self-supervised pretraining without humans. NeurIPS Track on Datasets and Benchmarks, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.36, + 0.484, + 0.4 + ], + "angle": 0, + "content": "[2] Mina Ghadimi Atigh, Julian Schoep, Erman Acar, Nanne Van Noord, and Pascal Mettes. Hyperbolic image segmentation. In CVPR, pages 4453-4462, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.402, + 0.484, + 0.441 + ], + "angle": 0, + "content": "[3] Gary Bécigneul and Octavian-Eugen Ganea. Riemannian adaptive optimization methods. arXiv preprint arXiv:1810.00760, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.444, + 0.483, + 0.47 + ], + "angle": 0, + "content": "[4] P Bedapudi. NudeNet: Neural Nets for Nudity Classification, Detection, and Selective Censoring, 2019. 2, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.472, + 0.483, + 0.511 + ], + "angle": 0, + "content": "[5] Abeba Birhane and Vinay Uday Prabhu. Large image datasets: A pyrrhic win for computer vision? In WACV, pages 1536-1546. IEEE, 2021. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.513, + 0.483, + 0.566 + ], + "angle": 0, + "content": "[6] Abeba Birhane, Vinay Uday Prabhu, and Emmanuel Kahembwe. Multimodal datasets: misogyny, pornography, and malignant stereotypes. arXiv preprint arXiv:2110.01963, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.568, + 0.483, + 0.608 + ], + "angle": 0, + "content": "[7] Abeba Birhane, Sanghyun Han, Vishnu Boddeti, Sasha Luccioni, et al. Into the laion's den: Investigating hate in multimodal datasets. NeurIPS, 36, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.611, + 0.483, + 0.664 + ], + "angle": 0, + "content": "[8] Manuel Brack, Felix Friedrich, Patrick Schramowski, and Kristian Kersting. Mitigating inappropriateness in image generation: Can there be value in reflecting the world's ugliness? arXiv preprint arXiv:2305.18398, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.666, + 0.483, + 0.706 + ], + "angle": 0, + "content": "[9] Yinzhi Cao and Junfeng Yang. Towards Making Systems Forget with Machine Unlearning. In IEEE Symposium on Security and Privacy, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.707, + 0.483, + 0.762 + ], + "angle": 0, + "content": "[10] Francesco Cauteruccio, Enrico Corradini, Giorgio Terracina, Domenico Ursino, and Luca Virgili. Extraction and analysis of text patterns from nsfw adult content in reddit. Data & Knowledge Engineering, 138:101979, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.763, + 0.483, + 0.803 + ], + "angle": 0, + "content": "[11] Benjamin Paul Chamberlain, James Clough, and Marc Peter Deisenroth. Neural embeddings of graphs in hyperbolic space. arXiv preprint arXiv:1705.10359, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.805, + 0.483, + 0.858 + ], + "angle": 0, + "content": "[12] Ines Chami, Adva Wolf, Da-Cheng Juan, Frederic Sala, Sujith Ravi, and Christopher Ré. Low-dimensional hyperbolic knowledge graph embeddings. arXiv preprint arXiv:2005.00545, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[13] Nurendra Choudhary, Nikhil Rao, Sumeet Katariya, Karthik Subbian, and Chandan K Reddy. Self-supervised hyperboloid representations from logical queries over knowledge" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.304, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.094, + 0.905, + 0.12 + ], + "angle": 0, + "content": "graphs. In Proceedings of the Web Conference 2021, pages 1373-1384, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.905, + 0.164 + ], + "angle": 0, + "content": "[14] Nurendra Choudhary, Nikhil Rao, and Chandan Reddy. Hyperbolic graph neural networks at scale: a meta learning approach. NeurIPS, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.167, + 0.905, + 0.222 + ], + "angle": 0, + "content": "[15] Damien L Crone, Stefan Bode, Carsten Murawski, and Simon M Laham. The socio-moral image database (smid): A novel stimulus set for the study of social, moral and affective processes. *PloS one*, 13(1):e0190954, 2018. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.224, + 0.905, + 0.279 + ], + "angle": 0, + "content": "[16] Karan Desai, Gaurav Kaul, Zubin Aysola, and Justin Johnson. RedCaps: Web-curated image-text data created by the people, for the people. In NeurIPS Datasets and Benchmarks, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.282, + 0.905, + 0.336 + ], + "angle": 0, + "content": "[17] Karan Desai, Maximilian Nickel, Tanmay Rajpurohit, Justin Johnson, and Shanmukha Ramakrishna Vedantam. Hyperbolic image-text representations. In ICML, pages 7694-7731. PMLR, 2023. 1, 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.34, + 0.905, + 0.393 + ], + "angle": 0, + "content": "[18] Bhuwan Dhingra, Christopher J Shallue, Mohammad Norouzi, Andrew M Dai, and George E Dahl. Embedding text in hyperbolic spaces. arXiv preprint arXiv:1806.04313, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.397, + 0.905, + 0.452 + ], + "angle": 0, + "content": "[19] Upol Ehsan, Q Vera Liao, Samir Passi, Mark O Riedl, and Hal Daumé III. Seamful xai: Operationalizing seamful design in explainable ai. Proceedings of the ACM on Human-Computer Interaction, 8(CSCW1):1-29, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.455, + 0.905, + 0.51 + ], + "angle": 0, + "content": "[20] Aleksandr Ermolov, Leyla Mirvakhabova, Valentin Khrulkov, Nicu Sebe, and Ivan Oseledets. Hyperbolic vision transformers: Combining improvements in metric learning. In CVPR, pages 7409-7419, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.512, + 0.905, + 0.553 + ], + "angle": 0, + "content": "[21] Luca Franco, Paolo Mandica, Bharti Munjal, and Fabio Galasso. Hyperbolic self-paced learning for self-supervised skeleton-based action representations. In ICLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.556, + 0.905, + 0.624 + ], + "angle": 0, + "content": "[22] Shreyansh Gandhi, Samrat Kokkula, Abon Chaudhuri, Alessandro Magnani, Theban Stanley, Behzad Ahmadi, Venkatesh Kandaswamy, Omer Ovenc, and Shie Mannor. Scalable Detection of Offensive and Non-compliant Content/Logo in Product Images. In WACV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.627, + 0.905, + 0.668 + ], + "angle": 0, + "content": "[23] Rohit Gandikota, Joanna Materzynska, Jaden Fiotto-Kaufman, and David Bau. Erasing Concepts from Diffusion Models. In ICCV, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.671, + 0.905, + 0.724 + ], + "angle": 0, + "content": "[24] Octavian Ganea, Gary Bécigneul, and Thomas Hofmann. Hyperbolic entailment cones for learning hierarchical embeddings. In ICML, pages 1646-1655. PMLR, 2018. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.729, + 0.905, + 0.756 + ], + "angle": 0, + "content": "[25] Octavian Ganea, Gary Bécigneul, and Thomas Hofmann. Hyperbolic neural networks. NeurIPS, 31, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.758, + 0.905, + 0.813 + ], + "angle": 0, + "content": "[26] Songwei Ge, Shlok Mishra, Simon Kornblith, Chun-Liang Li, and David Jacobs. Hyperbolic contrastive learning for visual representations beyond objects. In CVPR, pages 6840-6849, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.816, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[27] Antonio Ginart, Melody Guan, Gregory Valiant, and James Y Zou. Making AI Forget You: Data Deletion in Machine Learning. In NeurIPS, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[28] Aditya Golatkar, Alessandro Achille, and Stefano Soatto. Eternal Sunshine of the Spotless Net: Selective Forgetting in Deep Networks. In CVPR, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.094, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "4230" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[29] Aditya Golatkar, Alessandro Achille, Yu-Xiang Wang, Aaron Roth, Michael Kearns, and Stefano Soatto. Mixed Differential Privacy in Computer Vision. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.483, + 0.203 + ], + "angle": 0, + "content": "[30] Kimia Hamidieh, Haoran Zhang, Walter Gerych, Thomas Hartvigsen, and Marzyeh Ghassemi. Identifying implicit social biases in vision-language models. In Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society, pages 547-561, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.204, + 0.482, + 0.258 + ], + "angle": 0, + "content": "[31] Ahmad Fathan Hidayatullah, Anisa Miladya Hakim, and Abdullah Aziz Sembada. Adult Content Classification on Indonesian Tweets using LSTM Neural Network. In ICACSIS, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.259, + 0.482, + 0.315 + ], + "angle": 0, + "content": "[32] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.482, + 0.371 + ], + "angle": 0, + "content": "[33] Sarah Ibrahimi, Mina Ghadimi Atigh, Nanne Van Noord, Pascal Mettes, and Marcel Worring. Intriguing properties of hyperbolic embeddings in vision-language models. Transactions on Machine Learning Research, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.371, + 0.482, + 0.439 + ], + "angle": 0, + "content": "[34] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, pages 4904-4916. PMLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.44, + 0.482, + 0.482 + ], + "angle": 0, + "content": "[35] Valentin Khrulkov, Leyla Mirvakhabova, Evgeniya Ustinova, Ivan Oseledets, and Victor Lempitsky. Hyperbolic image embeddings. In CVPR, pages 6418-6428, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.483, + 0.482, + 0.524 + ], + "angle": 0, + "content": "[36] Nupur Kumari, Bingliang Zhang, Sheng-Yu Wang, Eli Shechtman, Richard Zhang, and Jun-Yan Zhu. Ablating concepts in text-to-image diffusion models. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.525, + 0.482, + 0.552 + ], + "angle": 0, + "content": "[37] Gant Laborde. Deep nn for nsfw detection. https://github.com/GantMan/nsfw_model, 2020.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.482, + 0.607 + ], + "angle": 0, + "content": "[38] Matthew Le, Stephen Roller, Laetitia Papaxanthos, Douwe Kiela, and Maximilian Nickel. Inferring concept hierarchies from text corpora via hyperbolic embeddings. In ACL, pages 3231-3241, 2019. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.608, + 0.482, + 0.662 + ], + "angle": 0, + "content": "[39] Ya-Wei Eileen Lin, Ronald R Coifman, Gal Mishne, and Ronen Talmon. Hyperbolic diffusion embedding and distance for hierarchical representation learning. In ICML, pages 21003–21025. PMLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.663, + 0.482, + 0.692 + ], + "angle": 0, + "content": "[40] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 36, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.482, + 0.719 + ], + "angle": 0, + "content": "[41] Qi Liu, Maximilian Nickel, and Douwe Kiela. Hyperbolic graph neural networks. NeurIPS, 32, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.482, + 0.775 + ], + "angle": 0, + "content": "[42] Shaoteng Liu, Jingjing Chen, Liangming Pan, Chong-Wah Ngo, Tat-Seng Chua, and Yu-Gang Jiang. Hyperbolic visual embedding learning for zero-shot recognition. In CVPR, pages 9273-9281, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.482, + 0.803 + ], + "angle": 0, + "content": "[43] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.482, + 0.858 + ], + "angle": 0, + "content": "[44] Todor Markov, Chong Zhang, Sandhini Agarwal, Florentine Eloundou Nekoul, Theodore Lee, Steven Adler, Angela Jiang, and Lilian Weng. A Holistic Approach to Undesired Content Detection in the Real World. In AAAI, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[45] Joanna Materzyńska, Antonio Torralba, and David Bau. Disentangling visual and written concepts in clip. In CVPR, pages 16410-16419, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[46] Pascal Mettes, Mina Ghadimi Atigh, Martin Keller-Ressel, Jeffrey Gu, and Serena Yeung. Hyperbolic deep learning in computer vision: A survey. IJCV, pages 1-25, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.137, + 0.905, + 0.192 + ], + "angle": 0, + "content": "[47] Paulius Micikevicius, Sharan Narang, Jonah Alben, Gregory Diamos, Erich Elsen, David Garcia, Boris Ginsburg, Michael Houston, Oleksii Kuchaiev, Ganesh Venkatesh, et al. Mixed precision training. In ICLR, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.194, + 0.905, + 0.264 + ], + "angle": 0, + "content": "[48] Leyla Mirvakhabova, Evgeny Frolov, Valentin Khrulkov, Ivan Oseledets, and Alexander Tuzhilin. Performance of hyperbolic geometry models on top-n recommendation tasks. In Proceedings of the 14th ACM Conference on Recommender Systems, pages 527-532, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.265, + 0.905, + 0.336 + ], + "angle": 0, + "content": "[49] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards Photorealistic Image Generation and Editing with Text-Guided Diffusion Models. arXiv preprint arXiv:2112.10741, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.338, + 0.905, + 0.378 + ], + "angle": 0, + "content": "[50] Maximillian Nickel and Douwe Kiela. Poincaré embeddings for learning hierarchical representations. NeurIPS, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.381, + 0.905, + 0.424 + ], + "angle": 0, + "content": "[51] Maximillian Nickel and Douwe Kiela. Learning continuous hierarchies in the lorentz model of hyperbolic geometry. In ICML, pages 3779-3788. PMLR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.426, + 0.905, + 0.495 + ], + "angle": 0, + "content": "[52] Avik Pal, Max van Spengler, Guido Maria D'Amely di Melendugno, Alessandro Flaborea, Fabio Galasso, and Pascal Mettes. Compositional entailment learning for hyperbolic vision-language models. arXiv preprint arXiv:2410.06912, 2024. 1, 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.497, + 0.905, + 0.539 + ], + "angle": 0, + "content": "[53] Wei Peng, Tuomas Varanka, Abdelrahman Mostafa, Henglin Shi, and Guoying Zhao. Hyperbolic deep neural networks: A survey. IEEE Trans. PAMI, 44(12):10023-10044, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.541, + 0.905, + 0.598 + ], + "angle": 0, + "content": "[54] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.6, + 0.905, + 0.655 + ], + "angle": 0, + "content": "[55] Samuele Poppi, Tobia Poppi, Federico Cocchi, Marcella Cornia, Lorenzo Baraldi, and Rita Cucchiara. Safe-CLIP: Removing NSFW Concepts from Vision-and-Language Models. In ECCV, 2024. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.657, + 0.905, + 0.713 + ], + "angle": 0, + "content": "[56] Samuele Poppi, Sara Sarto, Marcella Cornia, Lorenzo Baraldi, and Rita Cucchiara. Multi-Class Unlearning for Image Classification via Weight Filtering. IEEE Intelligent Systems, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.714, + 0.905, + 0.77 + ], + "angle": 0, + "content": "[57] Samuele Poppi, Sara Sarto, Marcella Cornia, Lorenzo Baraldi, and Rita Cucchiara. Unlearning vision transformers without retaining data via low-rank decompositions. In ICPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.772, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[58] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 1, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.858, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[59] Javier Rando, Daniel Paleka, David Lindner, Lennart Heim, and Florian Tramer. Red-teaming the stable diffusion safety filter. In NeurIPS ML Safety Workshop, 2022. 8" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.956 + ], + "angle": 0, + "content": "4231" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.146 + ], + "angle": 0, + "content": "[60] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.19 + ], + "angle": 0, + "content": "[61] Frederic Sala, Chris De Sa, Albert Gu, and Christopher Ré. Representation tradeoffs for hyperbolic embeddings. In ICML, pages 4460-4469. PMLR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[62] Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.303 + ], + "angle": 0, + "content": "[63] Patrick Schramowski, Christopher Tauchmann, and Kristian Kersting. Can Machines Help Us Answering Question 16 in Datasheets, and In Turn Reflecting on Inappropriate Content? In ACM FAccT, 2022. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.305, + 0.482, + 0.358 + ], + "angle": 0, + "content": "[64] Patrick Schramowski, Manuel Brack, Björn Deiseroth, and Kristian Kersting. Safe Latent Diffusion: Mitigating Inappropriate Degeneration in Diffusion Models. In CVPR, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.482, + 0.403 + ], + "angle": 0, + "content": "[65] Christoph Schuhmann. Clip based nsfw detector. https://github.com/LAION-AI/CLIP-based-NSFW-Detector/tree/main, 2022.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[66] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 1, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.482, + 0.531 + ], + "angle": 0, + "content": "[67] Sheng Shen, Liunian Harold Li, Hao Tan, Mohit Bansal, Anna Rohrbach, Kai-Wei Chang, Zhewei Yao, and Kurt Keutzer. How much can clip benefit vision-and-language tasks? arXiv preprint arXiv:2107.06383, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.532, + 0.482, + 0.56 + ], + "angle": 0, + "content": "[68] Ryohei Shimizu, Yusuke Mukuta, and Tatsuya Harada. Hyperbolic neural networks++. ICLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.482, + 0.588 + ], + "angle": 0, + "content": "[69] David Thiel. Identifying and eliminating csam in generative ml training data and models. 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.482, + 0.631 + ], + "angle": 0, + "content": "[70] Alexandru Tifrea, Gary Bécigneul, and Octavian-Eugen Ganea. Poincaré glove: Hyperbolic word embeddings. arXiv preprint arXiv:1810.06546, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.482, + 0.66 + ], + "angle": 0, + "content": "[71] Max van Spengler, Erwin Berkhout, and Pascal Mettes. Poincare resnet. In ICCV, pages 5419-5428, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.662, + 0.482, + 0.702 + ], + "angle": 0, + "content": "[72] Ivan Vendrov, Ryan Kiros, Sanja Fidler, and Raquel Urtasun. Order-embeddings of images and language. arXiv preprint arXiv:1511.06361, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.482, + 0.771 + ], + "angle": 0, + "content": "[73] Liping Wang, Fenyu Hu, Shu Wu, and Liang Wang. Fully hyperbolic graph convolution network for recommendation. In Proceedings of the 30th ACM international conference on information & knowledge management, pages 3483-3487, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.482, + 0.815 + ], + "angle": 0, + "content": "[74] Mengmeng Wang, Jiazheng Xing, and Yong Liu. Actionclip: A new paradigm for video action recognition. arXiv preprint arXiv:2109.08472, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[75] Robert Wolfe, Yiwei Yang, Bill Howe, and Aylin Caliskan. Contrastive language-vision ai models pretrained on web-scraped multimodal data exhibit sexual objectification bias. In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency, pages 1174-1185, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[76] Eric Zhang, Kai Wang, Xingqian Xu, Zhangyang Wang, and Humphrey Shi. Forget-Me-Not: Learning to Forget in Text-to-Image Diffusion Models. CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[77] Yudong Zhu, Di Zhou, Jinghui Xiao, Xin Jiang, Xiao Chen, and Qun Liu. Hypertext: Endowing fasttext with hyperbolic geometry. arXiv preprint arXiv:2010.16143, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.093, + 0.905, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4232" + } + ] +] \ No newline at end of file diff --git a/2025/Hyperbolic Safety-Aware Vision-Language Models/5ca9a6d6-45f1-4b74-b091-1c79b0b891b6_origin.pdf b/2025/Hyperbolic Safety-Aware Vision-Language Models/5ca9a6d6-45f1-4b74-b091-1c79b0b891b6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d4fd917375fbf81456d9823e29af988ece1eda6a --- /dev/null +++ b/2025/Hyperbolic Safety-Aware Vision-Language Models/5ca9a6d6-45f1-4b74-b091-1c79b0b891b6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8c8ee30003c7633aeaefb48f1dc154b1e7163c63ae53913edafce5149cad2585 +size 4148026 diff --git a/2025/Hyperbolic Safety-Aware Vision-Language Models/full.md b/2025/Hyperbolic Safety-Aware Vision-Language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0daad2a402abf23b3a1b31906cfd54279a2c1d41 --- /dev/null +++ b/2025/Hyperbolic Safety-Aware Vision-Language Models/full.md @@ -0,0 +1,391 @@ +# Hyperbolic Safety-Aware Vision-Language Models + +Tobia Poppi\*1,2 Tejaswi Kasarla\*3 Pascal Mettes3 Lorenzo Baraldi1 Rita Cucchiara1,4 1University of Modena and Reggio Emilia, Italy 2 University of Pisa, Italy 3 University of Amsterdam, Netherlands 4IIT-CNR, Italy + +$^{1}\{\text {name.surname}\} @ \text {unimore.it}$ $^{2}\{\text {name.surname}\} @ \text {phd.unipi.it}$ $^{3}\{\text {initial.surname}\} @ \text {uva.nl}$ + +# Abstract + +Addressing the retrieval of unsafe content from vision-language models such as CLIP is an important step towards real-world integration. Current efforts have relied on unlearning techniques that try to erase the model's knowledge of unsafe concepts. While effective in reducing unwanted outputs, unlearning limits the model's capacity to discern between safe and unsafe content. In this work, we introduce a novel approach that shifts from unlearning to an awareness paradigm by leveraging the inherent hierarchical properties of the hyperbolic space. We propose to encode safe and unsafe content as an entailment hierarchy, where both are placed in different regions of hyperbolic space. Our HySAC, Hyperbolic Safety-Aware CLIP, employs entailment loss functions to model the hierarchical and asymmetrical relations between safe and unsafe image-text pairs. This modelling - ineffective in standard vision-language models due to their reliance on Euclidean embeddings - endows the model with awareness of unsafe content, enabling it to serve as both a multimodal unsafe classifier and a flexible content retriever, with the option to dynamically redirect unsafe queries toward safer alternatives or retain the original output. Extensive experiments show that our approach not only enhances safety recognition but also establishes a more adaptable and interpretable framework for content moderation in vision-language models. Our source code is available at: https://github.com/aimagelab/HySAC + +Warning: This paper features explicit sexual content and other material that some readers may find disturbing, distressing, or offensive. + +# 1. Introduction + +Large-scale vision-language models (VLMs) have achieved remarkable successes in various applications, including cross-model retrieval [58], text-to-image and image-to-text generation [40, 60] and various downstream tasks [45, 67, + +74]. Popular VLMs like CLIP [58] and ALIGN [34] leverage vast amounts of web-scraped image-text data to learn rich multimodal representations by aligning visual and textual modalities. However, most large-scale datasets sourced from the web contain unsafe or inappropriate content, such as violence, nudity, or hate speech [5, 6]. The presence of such content not only raises ethical concerns but also introduces risks for real-world applications [7, 30, 75], where exposure to or misuse of this material can lead to legal and societal repercussions. Birhane et al. [7] also show that the increasing dataset scale can exacerbate hateful and unsafe content, as identified in the now removed LAION-5B [66, 69]. Addressing the issue of unsafe content in VLMs is therefore of utmost importance to ensure responsible AI practices. + +Recent efforts to mitigate unsafe content in vision-language models have led to the development of methods specifically designed for NSFW (Not Safe For Work) content removal. Most of these works [23, 55] have focused on unlearning (i.e., removing) the knowledge of unsafe content from the models. As a recent example, Poppi et al. [55] develop a fine-tuned version of CLIP which unlearns toxic concepts by redirecting their embeddings towards safe regions, so that retrieval always produces safe content even when the model is prompted with unsafe inputs. + +In contrast, we propose an approach for managing unsafe content in VLMs: emphasizing awareness over unlearning. Rather than hiding the flaws of VLMs by ignoring NSFW content, we aim to equip the VLMs with the ability to distinguish between safe and unsafe content. This in turn helps users of the model to expose or redirect NSFW content when necessary, a crucial step toward improving user agency, understanding, and interpretability [19]. + +Inspired by recent hyperbolic vision-language models [17, 52], we introduce a hyperbolic framework that leverages the geometric properties of hyperbolic space to separate safe and unsafe content effectively. Using a paired dataset of safe and unsafe image-text inputs [55], we adjust the embeddings to create an entailment-based [24] structure. In this setup, safe concepts are positioned closer to the + +origin of the hyperbolic space, while unsafe concepts are mapped further away. Specifically, we introduce a hyperbolic safe-to-unsafe entailment mechanism that ensures safe content encompasses unsafe representations within conical regions, defining clear safety boundaries and safety traversals to dynamically adjust query embeddings along the hyperbolic space to promote safe retrievals or, alternatively, expose relevant unsafe content when necessary. This framework not only organizes data into safe and unsafe radius-based regions but also enables controlled movement within the space, allowing retrievals to favor safety as required. Experiments demonstrate that HySAC achieves clear improvements in safety awareness, retrieval performance, and NSFW content handling across multiple datasets, with robustness in both safe content redirection and controlled unsafe content accessibility. + +# 2. Related Work + +Unlearning in vision-language models. Unlearning concepts and content has recently received a lot of attention, empowered by the success of vision-language models. Various approaches have been explored, such as full model retraining, fine-tuning, machine unlearning [9, 27, 28, 56, 57], and differential privacy [29]. Some of these efforts have focused on text-to-image models, with the goal of removing specific styles, concepts, or objects [36, 76]. + +A particular emphasis has been placed on removing NSFW content, which encompasses inappropriate, unsafe, or illegal material. Schramowsky et al. [64] steer the generation away from NSFW areas, defined by a fixed set of concepts. The embedding of NSFW concepts is applied as negative guidance during the text-conditioning phase, acting as safety guidance. Gandikota et al. [23] erase visual concepts using only their name, with negative guidance serving as a teacher. Poppi et al. [55] focus on removing NSFW concepts from CLIP-like models by fine-tuning the entire model using ViSU, a multimodal dataset containing Safe/NSFW Image/Texts quadruplets. Unlike earlier approaches, which target specific components, their method fine-tunes the entire vision-and-language model, making content removal applicable to downstream tasks. + +We introduce a method for handling NSFW concepts in CLIP-like models by fine-tuning them in hyperbolic space. We exploit the hierarchical properties of hyperbolic geometry, yielding a clear advantage: the model becomes explicitly aware of whether content is safe or NSFW, rather than merely removing knowledge of unsafe content. Similarly to our approach, safe generation works [8, 64], underscore that a deeper understanding of (un-)safety can improve the control of harmful content generation. + +NSFW concept detection. A related area of research is the automatic detection of NSFW content. Several meth + +ods have been proposed for detecting NSFW and toxic language [10, 31, 44], primarily in social media contexts. DistilBERT [62] has emerged as a promising model for this task, especially when fine-tuned for identifying adult content. Detecting inappropriate language presents a significant challenge, and this complexity extends to visual content, where various techniques have been developed to detect NSFW imagery [5, 22, 49]. In this domain, models like NudeNet [4] specialize in detecting nudity, while Q16 [63] serves as a broader classifier, capable of identifying a wider range of NSFW content. However, identifying inappropriate visual content remains a complex task, given the challenges posed by subtle visual cues, lack of contextual information, and limited data availability. + +While these detection methods focus solely on identifying unsafe content, they do not address the problem of retrieving relevant, safe alternatives when an unsafe input is detected. Our method makes it possible to jointly detect NSFW content and provide a mechanism to shift NSFW queries towards safe but relevant alternatives. + +Hyperbolic learning. A key advantage of hyperbolic space is its inherent ability to represent hierarchical or tree-like structures with minimal distortion [11, 50, 51, 61]. A comprehensive list of recent advancements is documented in surveys by Mettes et al. [46] and Peng et al. [53]. Foundational works for building neural networks in hyperbolic space [3, 24, 25, 35, 68] led to the use of hyperbolic models across multiple modalities such as images [2, 20, 21, 26, 42, 71], text [18, 38, 70, 77], graphs [12-14, 41] and recommender systems [48, 73]. + +Recent work has shown the strong potential of hyperbolic learning for vision-language models [17, 33, 52] and demonstrated that training with a loss function enforcing entailment cones [24] leads to the emergence of hierarchical structures between the embeddings. While Desai et al. [17] enforce entailment structure across modalities, Pal et al. [52] also explicitly enforce structure within modalities by leveraging object-level compositions and nouns from text. We take inspiration from these approaches and enforce entailment relations across safe and unsafe embeddings, creating an interpretable space for traversing from unsafe regions to safe regions in CLIP models. + +# 3. Preliminaries + +Throughout this work, we operate in hyperbolic space, a Riemannian manifold with a constant negative curvature. Following Desai et al. [17], we use the Lorentz model, as it is better equipped to deal with numerical instabilities associated with the Poincaré distance metric [39, 51]. + +The Lorentz model $\mathbb{L}^n$ is an $n$ -dimensional manifold in which points are represented on the upper sheet of a two-sheeted hyperboloid in $(n + 1)$ -dimensional Minkowski spacetime. Following terminology from general relativity, + +![](images/3ff6f18357e9d4e4959c4817dac3667620f8df1c865174840c71f5c6e255466b.jpg) +Figure 1. Overview of our approach. HySAC builds a hyperbolic embedding that manages content safety through an entailment hierarchy. Unsafe text and images are projected to dedicated regions of hyperbolic space, allowing for safety-aware retrieval and classification. + +for each vector $\mathbf{p} \in \mathbb{R}^{n+1}$ , we refer to the first dimension, the axis of symmetry, as the time-axis, denoted by $p_0$ , and the remaining $n$ -dimensions as the spatial components, denoted $\tilde{\pmb{p}}$ . The Lorentz model, $\mathbb{L}^n = (\mathcal{L}^n, \mathfrak{g}_p^\kappa)$ is given as + +$$ +\mathcal {L} ^ {n} := \left\{\boldsymbol {p} \in \mathbb {R} ^ {n + 1}: \langle \mathbf {p}, \mathbf {p} \rangle_ {\mathcal {L}} = - \frac {1}{\kappa}, p _ {0} = \sqrt {1 / \kappa + \| \tilde {\boldsymbol {p}} \| ^ {2}}, \kappa > 0 \right\}, \tag {1} +$$ + +where $-\kappa \in \mathbb{R}$ denotes the curvature with the Riemannian metric $\mathfrak{g}_p^\kappa = \mathrm{diag}(-1,1,\ldots ,1)$ . The Lorentzian inner product $\langle .,.\rangle_{\mathcal{L}}$ is induced by the metric tensor $\mathfrak{g}_p^\kappa$ and is defined for $\mathbf{p},\mathbf{q}\in \mathbb{L}^n$ as + +$$ +\langle \mathbf {p}, \mathbf {q} \rangle_ {\mathcal {L}} = - p _ {0} q _ {0} + \langle \tilde {\boldsymbol {p}}, \tilde {\boldsymbol {q}} \rangle , \tag {2} +$$ + +where $\langle .,.\rangle$ is the Euclidean inner product. The Lorentzian inner product induces a norm on the Lorentzian space which can be written as $\| \mathbf{p}\|_{\mathcal{L}} = \sqrt{\langle\mathbf{p},\mathbf{p}\rangle_{\mathcal{L}}}$ . We now define the common hyperbolic operations in Lorentz space. + +Definition 3.1 (Lorentzian distance). The Lorentzian distance between two points in $\mathbb{L}^n$ is the length of their shortest path (geodesic) connecting them, computed as + +$$ +d _ {\mathcal {L}} (\mathbf {p}, \mathbf {q}) = \sqrt {1 / \kappa} \cdot \cosh^ {- 1} (- \kappa \langle \mathbf {p}, \mathbf {q} \rangle_ {\mathcal {L}}). \tag {3} +$$ + +In our work, we use the negative of Lorentzian distance to calculate similarities between multimodal inputs [17, 52]. + +Definition 3.2 (Exponential map). Since Lorentz space is a Riemannian manifold, it is locally Euclidean. This is best described through the tangent space $T_{\mathbf{p}}\mathbb{L}^n$ , a first-order approximation of the Lorentzian manifold at a given point $\mathbf{p}\in \mathbb{L}^n$ . The exponential map then provides a means to project elements from the tangent space onto the hyperboloid. Given a point $\mathbf{v}\in T_{\mathbf{p}}\mathbb{L}^n$ , the exponential map is defined as $\exp_{\mathbf{p}}^{\kappa}\colon T_{\mathbf{p}}\mathbb{L}^{n}\to \mathbb{L}^{n}$ with the expression + +$$ +\exp_ {\mathbf {p}} ^ {\kappa} (\mathbf {v}) = \cosh (\sqrt {\kappa} \| \mathbf {v} \| _ {\mathbb {L}}) \mathbf {p} + \frac {\sinh (\sqrt {\kappa} \| \mathbf {v} \| _ {\mathbb {L}})}{\sqrt {\kappa} \| \mathbf {v} \| _ {\mathbb {L}}} \mathbf {v}. \tag {4} +$$ + +In practice, the reference point $\mathbf{p}$ is set to the origin $\mathbf{0} = (\sqrt{1 / \kappa}, 0, \dots, 0)^T$ on the hyperboloid, allowing $\exp_0^\kappa$ to project Euclidean vectors from the tangent space at $\mathbf{0}$ directly onto the hyperboloid [17, 35]. In this work, the exponential map is used to project the outputs of the visual and textual encoders to a shared hyperbolic space. + +# 4. HySAC: Hyperbolic Safety-Aware CLIP + +# 4.1. Problem formulation and objective + +Problem setup. Given a dataset $D = \{(I_i, T_i)\}_{i=1}^N$ of $N$ image-text pairs, vision-language models (e.g. CLIP [58]) align the visual and textual embeddings obtained from image and text encoders in a shared embedding space. Large-scale datasets employed for training such embedding spaces are often web-scraped and contain unsafe samples [5]. For our problem setup, in order to differentiate between safe and unsafe content, we denote safe image-text pairs in $D$ as $(I_i, T_i)$ and unsafe image-text pairs as $(I_i^\star, T_i^\star)$ . + +To make vision-language models aware of inappropriate contents, and enable them to avoid or redirect the representation of such content, we also require a dataset of quadruplets of safe and unsafe image-text pairs, denoted as $D^{\star} = \{(I_k, T_k, I_k^{\star}, T_k^{\star})\}_{k=1}^{K}$ [55]. This dataset is generated following the definition of NSFW content of Schramowski et al. [64] containing the following twenty categories: hate, harassment, violence, suffering, humiliation, harm, suicide, sexual, nudity, bodily fluids, blood, obscene gestures, illegal activity, drug use, theft, vandalism, weapons, abuse, brutality, and cruelty. The dataset is constructed such that the unsafe image-text pairs $(I_k^{\star}, T_k^{\star})$ are specific cases or modified versions of the safe representations $(I_k, T_k)$ . + +Modelling relations in the hyperbolic space. We first consider the relationship between text and image modalities in the embedding space. Like other hyperbolic vision- + +language models [17, 52], we consider text as a general version of images to reflect the natural structure of partial order embeddings [72]. This is enforced in the embedding space by placing text embeddings closer to the origin and image embeddings farther away, defining a modality entailment relationship formally expressed as + +$$ +g _ {T} \left(T _ {k}\right) \ll g _ {I} \left(I _ {k}\right), \quad \text {a n d} \quad g _ {T} \left(T _ {k} ^ {\star}\right) \ll g _ {I} \left(I _ {k} ^ {\star}\right), \tag {5} +$$ + +where $g_{I}$ and $g_{T}$ denote the projections of images and text in the hyperbolic model and $\ll$ indicates that one embedding is closer to the origin than another. Differently from other models, we also impose a relationship between unsafe and safe pairs, by considering unsafe image-text pairs as specific cases of their safe counterparts within $D^{\star}$ . Hence, we establish a safety entailment to segregate safe from unsafe content, as follows + +$$ +g _ {I} \left(I _ {k}\right) \ll g _ {T} \left(T _ {k} ^ {\star}\right). \tag {6} +$$ + +By satisfying these inequalities, $g_{I}$ and $g_{T}$ capture both the modal and safety hierarchies within the data, thereby endowing the embedding space with safety-aware properties. To sum up, our objective is to model a hyperbolic vision-language model with the following inequality chain: + +$$ +g _ {T} \left(T _ {k}\right) \ll g _ {I} \left(I _ {k}\right) \ll g _ {T} \left(T _ {k} ^ {\star}\right) \ll g _ {I} \left(I _ {k} ^ {\star}\right). \tag {7} +$$ + +Below we outline our method to create safety-aware vision-language models in hyperbolic space. + +# 4.2. Hyperbolic safety learning + +To optimize Eq. 7 for safety-aware vision-language models, we propose Hyperbolic Safety-Aware CLIP to rearrange the embedding space to separate safe and unsafe regions. Our optimization consists of two components: (1) a hyperbolic safety contrastive component to align image-text pairs over a mini-batch and (2) a hyperbolic safety entailment to align safe and unsafe content. An overview is given in Figure 1. + +Hyperbolic safety contrastive learning. CLIP [58] and Safe-CLIP [55] rely on contrastive objectives to align and distribute the multimodal data. We utilize hyperbolic embeddings to align the visual and textual data. Specifically, we project the visual and textual embeddings from a pre-trained vision-language model onto a hyperboloid [17] through an exponential map (Eq. 4). Let $f_{I}(\cdot)$ and $f_{T}(\cdot)$ represent any Euclidean encoders for image and text. Then, $g_{I}(I_{k}) = \exp_{0}^{\kappa}(\alpha_{img} \cdot f_{I}(I_{k}))$ and $g_{T}(T_{k}) = \exp_{0}^{\kappa}(\alpha_{txt} \cdot f_{T}(T_{k}))$ represent the hyperbolic representations of a safe image-text pair, $(I_{k}, T_{k})$ . Similarly, $g_{I}(I_{k}^{\star}) = \exp_{0}^{\kappa}(\alpha_{img} \cdot f_{I}(I_{k}^{\star}))$ and $g_{T}(T_{k}^{\star}) = \exp_{0}^{\kappa}(\alpha_{txt} \cdot f_{T}(T_{k}^{\star}))$ represent the hyperbolic representations of an unsafe image-text pair $(I_{k}^{\star}, T_{k}^{\star})$ . $\alpha_{img}$ and $\alpha_{txt}$ are learnable projection scalars. + +To align representations in hyperbolic space, the similarity for the image-text and text-image contrastive loss is + +based on negative Lorentzian distance (Eq. 3) between $g_{I}(.)$ and $g_{T}(.)$ . We compute the hyperbolic safety contrastive loss over safe image-text pairs $(I_k, T_k)$ in a batch $B$ as + +$$ +L _ {\text {c o n t}} ^ {*} (I, T) = - \sum_ {i \in B} \log \frac {\exp \left(d _ {\mathcal {L}} \left(g _ {I} \left(I _ {i}\right) , g _ {T} \left(T _ {i}\right)\right) / \tau\right)}{\sum_ {k = 1 , k \neq i} ^ {B} \exp \left(d _ {\mathcal {L}} \left(g _ {I} \left(I _ {i}\right) , g _ {T} \left(T _ {k}\right)\right) / \tau\right)}, \tag {8} +$$ + +where $\tau$ denotes a temperature hyperparameter. A similar contrastive loss is employed to preserve the multimodal structure between unsafe image-text pairs $(I_k^{\star}, T_k^{\star})$ . Two additional contrastive losses between cross-safety modalities ensure the alignment of quadruplets in the embedding space. The final contrastive loss is formulated as + +$$ +\begin{array}{l} L _ {\mathrm {h S C}} (I, T, I ^ {\star}, T ^ {\star}) = L _ {\text {c o n t}} ^ {\star} (I, T) + L _ {\text {c o n t}} ^ {\star} \left(I ^ {\star}, T ^ {\star}\right) \tag {9} \\ + L _ {c o n t} ^ {\star} (I, T ^ {\star}) + L _ {c o n t} ^ {\star} (I ^ {\star}, T). \\ \end{array} +$$ + +Hyperbolic safety entailment learning. Hyperbolic entailment cones, introduced by Ganea et al. [24], generalize partial ordered embeddings [72] to any Riemannian manifold. Entailment cones induce a partial order between concepts in a dataset $\mathcal{X}$ such that for any pair $(\mathbf{p},\mathbf{q})\in \mathcal{X}$ , if $\mathbf{p}$ is a sub-concept of $\mathbf{q}$ , then $\mathbf{q}$ entails $\mathbf{p}$ within a conical region $\mathfrak{S}_{\mathbf{q}}$ defined by $\mathbf{q}$ . For the Lorentz model, $\mathbb{L}^n$ , the half-aperture of each conical region $\mathfrak{S}_{\mathbf{q}}$ is defined as [17, 38] + +$$ +\omega (\mathbf {q}) = \sin^ {- 1} \left(\frac {2 K}{\sqrt {\kappa} \| \tilde {\boldsymbol {q}} \|}\right), \tag {10} +$$ + +where $-\kappa$ is the curvature of space, and the constant $K = 0.1$ limits values near the origin [24]. To preserve partial order between image-text relationships, we add entailment from safe text to safe image and from unsafe text to unsafe images, effectively implementing Eq. 5. Specifically, a safe image $I_{k}$ must lie within the cone defined by its corresponding safe text, $T_{k}$ , characterized by the half-aperture $\omega (T_k)$ . Similarly, an unsafe image $I_{k}^{\star}$ must lie within the cone defined by its corresponding unsafe text $T_{k}^{\star}$ . This is enforced through an entailment loss formulated for image-text representations by Le et al. [38] and Desai et al. [17], as + +$$ +\begin{array}{l} \begin{array}{l} L _ {e n t} ^ {\star} (I, T) = \max \left(0, \phi \left(I _ {k}, T _ {k}\right) - \eta \omega \left(T _ {k}\right)\right) \text {a n d} \\ I _ {*} ^ {\star} \left(I _ {*} ^ {\star}, T _ {*} ^ {\star}\right) = \left(0 + \left(I _ {*} ^ {\star}, T _ {*} ^ {\star}\right) - \left(T _ {*} ^ {\star}\right)\right) \end{array} \tag {11} \\ L _ {e n t} ^ {\star} \left(I ^ {\star}, T ^ {\star}\right) = \max \left(0, \phi \left(I _ {k} ^ {\star}, T _ {k} ^ {\star}\right) - \eta \omega \left(T _ {k} ^ {\star}\right)\right), \\ \end{array} +$$ + +where $\phi$ is the exterior angle (between lines $I_{k}T_{k}$ and $\mathbf{0}T_{k}$ or between lines $I_k^\star T_k^\star$ and $\mathbf{0}T_k^\star$ ) given by + +$$ +\phi \left(I _ {k}, T _ {k}\right) = \cos^ {- 1} \left(\frac {I _ {k 0} + T _ {k 0} \kappa \left\langle I _ {k} , T _ {k} \right\rangle_ {\mathcal {L}}}{\left\| \widetilde {T _ {k}} \right\| \sqrt {\left(\kappa \left\langle I _ {k} , T _ {k} \right\rangle_ {\mathcal {L}}\right) ^ {2} - 1}}\right). \tag {12} +$$ + +Here, $\eta$ is a threshold for the half-aperture, $\omega (T_{k})$ [52]. Intuitively, the entailment loss $L_{ent}^{\star}$ penalizes images $I_{k}$ + +that lie outside the cone $\mathfrak{S}_{T_k}$ defined by their corresponding text caption $T_{k}$ . Finally, to model the safety hierarchy, we enforce that safe concepts entail unsafe ones, reflecting that safe data are more general and unsafe data are more specific. Specifically, we enforce that a safe image $I_{k}$ entails an unsafe text $T_{k}^{\star}$ , meaning that the unsafe text $T_{k}^{\star}$ must lie within the cone defined by the safe image $I_{k}$ , characterized by the half-aperture $\omega (I_k)$ . This gives us the safety-entailment defined in Eq. 6. This is implemented as + +$$ +L _ {e n t} ^ {\star} \left(T ^ {\star}, I\right) = \max \left(0, \phi \left(T _ {k} ^ {\star}, I _ {k}\right) - \eta \omega \left(I _ {k}\right)\right). \tag {13} +$$ + +The overall entailment loss to satisfy Eq. 7 is defined as + +$$ +L _ {\mathrm {h S E}} (I, T, I ^ {\star}, T ^ {\star}) = L _ {e n t} ^ {\star} (I, T) + L _ {e n t} ^ {\star} \left(T ^ {\star}, I\right) + L _ {e n t} ^ {\star} \left(I ^ {\star}, T ^ {\star}\right). \tag {14} +$$ + +Combined loss function. We integrate the contrastive with the entailment losses to obtain the total loss used to fine-tune the model on the dataset $D^{\star}$ : + +$$ +L (I, T, I ^ {\star}, T ^ {\star}) = L _ {\mathrm {h S C}} (I, T, I ^ {\star}, T ^ {\star}) + L _ {\mathrm {h S E}} (I, T, I ^ {\star}, T ^ {\star}). \tag {15} +$$ + +Our proposal allows the model to differentiate between safe and unsafe embeddings based on their distance from the origin. Safe content is closer to the center, while NSFW content is farther away. This geometric arrangement not only enables the model to detect unsafe content but also allows dynamic manipulation of embeddings. NSFW queries can be redirected toward the safe region, effectively retrieving outputs that prioritize safety and providing more precise control over content retrieval. + +# 4.3. Safety traversals and evaluation + +The hyperbolic safety-aware training results in a restructuring of the shared embedding space of the vision-language model. To obtain safe but relevant retrieval outputs from unsafe queries or vice-versa, we introduce a traversal mechanism to adjust query embeddings in hyperbolic space, enhancing their similarity with either safe or unsafe content, depending on the retrieval task. This traversal involves moving the query embeddings along the line connecting them to the origin of the hyperboloid, altering their hyperbolic distance from the root. By adjusting the embeddings' positions, we align them with regions in the embedding space that correspond to the desired content type. + +Traversal Definition. Given an embedding $\mathbf{q}$ , our method computes the distance from a predefined root feature $\mathbf{r}$ in hyperbolic space using the Lorentzian distance function $d_{\mathcal{L}}(\mathbf{q},\mathbf{r})$ . For each type of content $X\in \{T,I,T^{\star},I^{\star}\}$ , we compute the mean distance $\mu_X$ from the root feature $\mathbf{r}$ based on the distribution of each category. The boundary for each type is then defined as + +$$ +\tau_ {X} = \mu_ {X} + \tanh \left(\frac {\mu_ {X} - \alpha}{\kappa}\right) + 1, \tag {16} +$$ + +where $\kappa$ is the negative curvature, and $\alpha$ is a constant set empirically to 0.8. This shift accounts for the curvature of space, ensuring the boundaries are appropriately adjusted for effective traversal. Defining four bounds allows more nuanced control over traversal depending on the retrieval task. To retrieve a content type $X$ , the query is moved along the Euclidean direction vector $\mathbf{v}_{\mathrm{dir}} = \mathbf{q} - \mathbf{r}$ toward the root feature $\mathbf{r}$ until it reaches the corresponding boundary $\tau_{X}$ (e.g. $\tau_{T}$ for safe text). The target position $\mathbf{q}^{*}$ is given as + +$$ +\mathbf {q} ^ {*} = \mathbf {r} + \tau_ {X} \cdot \frac {\mathbf {v} _ {\mathrm {d i r}}}{\| \mathbf {v} _ {\mathrm {d i r}} \|} \tag {17} +$$ + +allowing the embeddings to be repositioned to match the target content type while maintaining semantic alignment. + +# 5. Experiments + +# 5.1. Training Details + +Datasets. Our experiments are mainly conducted on the ViSU dataset [55], containing 165k quadruplets of safe and unsafe image-text pairs. We also evaluate our model on three real-world NSFW image datasets: NudeNet [4], NSFW data source URLs $^1$ and SMID [15]. + +Baselines. Our safety comparisons include the original CLIP [58] and the state-of-the-art Safe-CLIP [55]. CLIP was trained using a private dataset of 400M image-text pairs [66] which has unsafe data [5]. Safe-CLIP is finetuned on the ViSU dataset [55] to redirect unsafe content to safe correspondent one via contrastive losses and cosine similarities, aiming to unlearn NSFW concepts. + +Models. Our visual and textual encoders are the same as CLIP [58], with VIT-L/14 as visual encoder, to maintain fair comparison to Safe-CLIP [55]. During training, both the visual and textual encoder are fine-tuned using low-rank decomposition [32] with low-rank factor $r = 16$ . + +**Optimization.** We use AdamW [43] with weight decay 0.2 and $(\beta_{1},\beta_{2}) = (0.9,0.98)$ . We disable weight decay for all gains, biases, and learnable scalars. The model is finetuned for 20 epochs with batch size 256. The maximum learning rate is $8\times 10^{-4}$ . We use mixed precision [47] to accelerate training, except computing exponential map and losses for HySAC in FP32 precision for numerical stability. + +Initialization. We initialize image and text encoders akin to CLIP, along with pre-trained weights. We initialize the softmax temperature as $\tau = 0.07$ and clamp it to a minimum value of 0.01. For HySAC, we initialize the learnable projection scalars $\alpha_{img} = \alpha_{txt} = 1 / \sqrt{512}$ , the curvature parameter $c = 1.0$ and clamp it in [0.1, 10.0] to prevent training instability. All scalars are learned in logarithmic space as $\log(1 / \tau)$ , $\log(c)$ , $\log(\alpha_{img})$ and $\log(\alpha_{txt})$ . + +Further details on the training setup are provided in the supplementary B. + +
ModelText-to-Image (T-to-I)Image-to-Text (I-to-T)Text-to-Image (T*-to-I ∪ I*)Image-to-Text (I*-to-T ∪ T*)
R@1R@10R@20R@1R@10R@20R@1R@10R@20R@1R@10R@20
CLIP [58]36.871.681.539.874.283.52.024.833.24.632.940.6
MERU [17]14.943.054.214.742.353.82.215.221.54.422.629.4
HyCoCLIP [52]34.371.280.634.471.382.22.825.333.28.237.845.7
Safe-CLIP [55]45.981.889.745.382.389.88.046.958.019.162.971.1
MERU*50.084.191.151.285.392.32.339.949.45.747.954.7
HyCoCLIP*47.781.989.146.782.790.41.532.742.36.945.253.6
HySAC49.884.190.748.284.291.230.562.871.842.173.379.8
+ +Table 1. Safe content retrieval performance on ViSU test set. Across all tasks and recall rates, HySAC improves over existing safety unlearning CLIP and hyperbolic CLIP models, highlighting that our approach is able to navigate unsafe image or text inputs towards relevant but safe retrieval outputs. * CLIP fine-tuned in hyperbolic space on ViSU training set with MERU/HyCoCLIP losses. + +
ModelText-to-Image (T* -to-I*)Image-to-Text (I* -to-T*)Text-to-Image (T* -to-I* ∪ I)Image-to-Text (I* -to-T* ∪ T)
R@1R@10R@20R@1R@10R@20R@1R@10R@20R@1R@10R@20
CLIP [58]73.194.997.672.895.297.768.492.395.967.193.396.7
MERU [17]29.462.472.225.857.767.823.554.064.319.551.161.2
HyCoCLIP [52]69.593.195.865.091.195.063.789.793.755.288.092.7
Safe-CLIP [55]58.086.291.456.085.191.047.780.085.832.177.184.6
HySAC81.498.499.482.297.899.281.198.499.480.597.298.9
+ +Table 2. Unsafe content retrieval performance on ViSU test set. Akin to safe content retrieval, our approach performs best. This is a result of our objective, as we assign different content to different regions, enabling us to maintain valuable safety information. + +# 5.2. Experimental Results + +To assess the performance of our proposed model, HySAC, we measure its safety awareness and its ability to handle unsafe content effectively, while retaining both safe and unsafe knowledge. In the supplementary D, we report the zero-shot generalization of our method. + +# 5.2.1. Safety retrieval comparison + +We evaluate our model on the capability of retrieving safe and unsafe items, in comparison to CLIP [58] and Safe-CLIP [55]. We also provide a comparison to recent hyperbolic VLMs, namely MERU [17] and HyCoCLIP [52]. Safe-CLIP and HySAC fine-tuned from CLIP on ViSU [55], MERU trained on RedCaps [16], and HyCoCLIP on GRIT [54]. For a fair comparison, we additionally fine-tuned the CLIP model on the ViSU dataset in hyperbolic space, using MERU and HyCoCLIP losses. All the models are evaluated on the ViSU test set. + +Retrieval tasks are defined as text-to-image and image-to-text, where the goal is to find the most relevant counterpart for a given query. Recall@K measures the fraction of queries where the correct item appears in the top-K retrieved results. To assess the safe retrieval performance of HySAC, we measure recall exclusively on safe content in both visual and textual elements ( $T$ -to- $I$ and $I$ -to- $T$ ). This step is crucial to verify that the original CLIP model's retrieval capabilities are retained after finetuning in hyperbolic space with our training method. Then, to evaluate + +the safety-awareness capabilities of HySAC, we introduce a distinct setup in which NSFW elements are used as queries, while the retrievable items include both safe and unsafe elements ( $T^{*}$ -to- $I \cup I^{*}$ and $I^{*}$ -to- $T \cup T^{*}$ ). During these experiments, a retrieval is deemed correct only if the query retrieves its safe counterpart, thereby validating the model's ability to redirect unsafe queries towards safe items. + +When retrieving with HySAC, the threshold $\tau$ for moving query embeddings is computed using the mean distance of safe embeddings from the origin, adjusting the query embedding towards the safe region. Results are reported in Table 1, where we observe that HySAC consistently improves over both unlearning and existing hyperbolic models and features the highest recalls across all settings and rates. CLIP hyperbolic models finetuned on ViSU data (MERU\* and HyCoCLIP\*) perform well on safe-only retrieval, while our method achieves high performance on both safe-only and unsafe-safe retrieval, due to our safety-aware design. + +In Table 2, we instead perform analyses for unsafe content retrieval. First, this involves the text-to-image and image-to-text retrieval on only unsafe elements ( $T^{*}$ -to- $I^{*}$ and $I^{*}$ -to- $T^{*}$ ). Second, instead, we perform retrieval by using unsafe elements as queries and both safe and unsafe items as retrievable items ( $T^{*}$ -to- $I^{*} \cup I$ and $I^{*}$ -to- $T^{*} \cup T$ ), and deem the retrieval correct only if the query retrieves its corresponding unsafe one. This setup tests the model's ability to function as a content moderator and also showcases its capacity to provide user autonomy in content retrieval decisions. In these tests, the traversal mechanism in + +![](images/648e8418109949de6e574dd295c73a730d1452a7c47aae6150a045333fd262b8.jpg) +Figure 2. Distributions of embedding distances from the root. We embed all ViSU training samples and visualize their distance distribution from the root. While CLIP and Safe-CLIP do not separate between texts and images, MERU does. HySAC, instead, also differentiates between safe and unsafe content. + +
Model(T-to-I)(I-to-T)(T*-to-I ∪ I*)(I*-to-T ∪ T*)
R@1R@10R@1R@10R@1R@10R@1R@10
w/o Ent52.384.950.884.74.149.05.564.5
w/o S-Ent51.084.249.884.31.439.17.463.7
HySAC49.884.148.284.230.562.842.173.3
+ +HySAC uses adjusted parameters to move in the unsafe direction, targeting the retrieval of NSFW content. Here too, HySAC achieves the best recall across all settings, demonstrating that HySAC not only prioritizes safety by navigating away from NSFW content when required but also ensures that users can access NSFW content under controlled conditions, better than existing competitors. + +# 5.2.2. Analysis of HySAC + +Assessing HySAC embedding space. Further, we validate the organization of the embedding space as outlined in Equation 7. Our goal is to confirm that the embeddings for safe content are positioned closer to the origin of hyperbolic space, while those for unsafe content are further away, following the proposed hierarchy. Specifically, for each $X \in \{T, I, T^{\star}, I^{\star}\}$ of the training-set of $D^{\star}$ , we compute the distances $d_{\mathcal{L}}(X, \mathbf{r})$ from the root feature $\mathbf{r}$ . + +A visualization is reported in Figure 2, where we show the distribution of embeddings in terms of their distance to the root feature. The comparative analysis is done across four different models: HySAC, CLIP, Safe-CLIP, and MERU. For both the hyperbolic models the root feature is the origin of the hyperboloid. For the Euclidean models, since the origin does not lie on the hypersphere, the root is empirically estimated as the embedding that has the least + +Table 3. Ablation study on loss components. We evaluate HySAC against two ablations that remove loss components. Results are in the same setting of Table 1. + +
Model% Safe (Text-to-Image)% Safe (Image-to-Text)
NudeNetNSFW URLsSMIDNudeNetNSFW URLsSMID
CLIP78.279.755.233.344.059.1
Safe-CLIP92.692.683.475.276.465.6
HySAC96.293.980.184.495.197.9
+ +Table 4. Retrieval performance on real NSFW images. Rate of safe images retrieved using unsafe prompts from the ViSU test set. The retrievable set includes safe and unsafe real images, with the latter from LAION-400M and the former from NSFW sources. + +distance from all embeddings of the training set, i.e. the $\ell_2$ normalization of the average of all embeddings. As it can be seen, the distribution clearly shows four peak distributions for HySAC, each one representing one of the $X$ content types, elucidating the efficacy of our approach in maintaining a clear separation between safe, unsafe, textual, and visual content within the embedding space. + +Ablation Study. In Table 3, we validate the effectiveness of the key components in HySAC, by comparing its full configuration with variants where specific losses are disabled. In particular, we employ one variant which only keeps contrastive losses (denoted as "w/o Ent") and one that omits the safety-entailment loss ("w/o S-Ent"). Results show that while removing these components slightly improves performance in scenarios involving only safe content, likely due to reduced spatial constraints, their absence significantly undermines the model's effectiveness in dealing with unsafe content, especially in unsafe-to-safe retrieval. These results underscore the essential roles that both the modality- and safety-entailment losses play in enhancing the safety awareness of the proposed model. + +Retrieval on real NSFW datasets. To further analyze the safety of HySAC, we conduct retrieval tests using NudeNet [4], NSFW data source URLs and SMID [15]. The first two datasets primarily contain nudity and pornographic content, whereas SMID includes a broader range of inappropriate content, such as violence, harm, and discrimination. We randomly select 1000 images from each dataset to serve as visual elements and use 5k NSFW captions from the ViSU test set. Both the image-to-text and text-to-image retrieval tasks also incorporate 10,000 randomly chosen retrievable safe items from LAION-400M [66]. + +Results, displayed in Table 4, contrast the performance of HySAC with that of CLIP [58] and Safe-CLIP [55]. We report the proportion of safe retrieval outputs when NSFW queries are used, highlighting the capability of HySAC to enhance safety in retrieval results. Traversing the embedding space towards safety prioritizes safer alternatives, which may adjust relevance in some cases. Due to the absence of datasets with real unsafe data and correlated safe alternatives, this aspect could not be evaluated. Notably, HySAC demonstrates improved performance in securing safer content compared to Safe-CLIP across most datasets for both text-to-image and image-to-text scenarios. + +![](images/137086e44026790740324abec669e2161d42a69cb8feb260ddbe55b70414e761.jpg) +A train filled with toxic chemicals crashing into a bridge, causing a massive explosion and destroying everything in its path. +Figure 3. Qualitative traversal results. HySAC traverses towards the root feature, retrieving the top-1 text at each interpolation point. This traversal effectively transitions from unsafe to safe captions, demonstrating the model's ability to ensure safety-aware content retrieval. + +![](images/915170f5b969afbdb3f9a7a8a4a6055527505e925b4ff33f0a9a73ba9b56ee67.jpg) +The train is going over the bridge over a river. +Several people on skis on a snowy slope, enjoying a drug-fueled party. +Several people in their ski gear are in the snow. +[ROOT] + +![](images/b50845a781859b6ce8160e0ae58ff6b6ec0468e2a8e7a11038a16d70115d12aa.jpg) +[ROOT] +A random Black man is standing in front of a double door with a gun in his hand and a look of hatred in his eyes. +A person is holding an umbrella while standing in the garage. +[ROOT] + +![](images/5d60419f6c4e62fcf50255a2214760fac3ec688e3827055b5e251bd2f4515c55.jpg) +Two naked women sitting next to each other while flashing their genitals. +Sultry women. +A couple of women sitting next to each other in front of a cake. +[ROOT] + +
ModelNudeNetMixed NSFW
AccFPRFNRAccFPRFNR
NSFW-CNN [37]85.30.014.766.54.535.9
CLIP-classifier [65]97.30.02.776.90.111.0
CLIP-distance [59]86.40.013.677.82.022.1
NudeNet [4]91.20.08.876.94.524.6
Q16 [63]28.50.071.565.38.329.4
HySAC99.50.00.578.916.56.8
+ +Table 5. NSFW classification. Comparison between HySAC and other NSFW classifiers. Metrics reported in percentages. + +Classifying NSFW content. The structure of the embedding space in HySAC also supports the classification of NSFW content. We evaluate this using the NudeNet [4] and Mixed NSFW datasets, comparing against classifiers such as NSFW-CNN [37], CLIP-classifier [65], CLIP-distance [59], NudeNet [4], and Q16 [63]. NudeNet only contains nudity, while Mixed NSFW includes different NSFW categories from various online sources and safe images from PASS [1]. We sample a 1,000-image subset from NudeNet and 442 images from Mixed NSFW, balanced between safe and unsafe. Further details of these datasets are provided in the supplementary C. + +Results in Table 5 show that HySAC achieves competitive or superior results in NSFW content classification, despite not being explicitly designed for safety classification. The norm threshold, set to the ViSU dataset mean (Figure 2), differentiates safe from unsafe content. + +# 5.2.3. Visualizing the safety traversals + +We examine traversal paths for safe text retrievals, starting with an unsafe image embedding as the query. The traversals are along the geodesic of hyperbolic space from the image to the origin of the hyperboloid, denoted [ROOT]. + +The input query is an unsafe image taken from ViSU test set and the text retrieval space consists of a mix of safe and unsafe captions of ViSU test set, metadata-based caption + +from pexels . com, and a curated list of unsafe words3. To create visualization shown in Figure 3, each retrieved text output is selected only once across all interpolation points, ensuring unique retrievals. Results show that, with HySAC, as the query nears the origin, retrieved content shifts from unsafe to safe while preserving semantic relevance. This progression illustrates the model's capability to effectively navigate the embedding space along relevant paths. For more on the experimental setup and traversal visualizations, see the supplementary E. + +Other ablation studies. For additional ablations on embedding space geometry and hyperparameter evaluation, we refer the reader to the supplementary D. + +# 6. Conclusion + +This paper introduces hyperbolic safety-aware vision-language models. Where recent literature focuses on removing or unlearning unsafe image-text content, we bring a perspective of awareness. By modelling unsafe image-text content as specific cases of their safe counterparts, we can divide the space into safe and unsafe regions. We show that hyperbolic space is a natural solution for this hierarchical relation, and propose a hyperbolic CLIP model with safety entailment learning and traversal. Our approach not only results in better retrieval of relevant safe outputs given unsafe inputs but also provides more robustness and comes with an NSFW classifier as a free by-product. Dealing with NSFW data in vision-language models is an important open research problem with real-world implications, from ethical to legal and societal concerns. By opting for awareness, we find that safety recognition improves. Given the importance of the ethical implications of this work, we provide a thorough discussion in the supplementary A. + +# Acknowledgements + +We acknowledge the CINECA award under the ISCRA initiative, for the availability of high-performance computing resources. This work has been supported by the EU Horizon projects "ELIAS - European Lighthouse of AI for Sustainability" (No. 101120237) and "European Lighthouse on Safe and Secure AI (ELSA)" (No. 101070617), co-funded by the European Union. Tejaswi Kasarla also acknowledges travel support from the European Union's Horizon research and innovation programme under grant agreement No. 951847 (ELISE) and No. 101120237 (ELIAS). + +# References + +[1] Yuki M. Asano, Christian Rupprecht, Andrew Zisserman, and Andrea Vedaldi. Pass: An imagenet replacement for self-supervised pretraining without humans. NeurIPS Track on Datasets and Benchmarks, 2021. 8 +[2] Mina Ghadimi Atigh, Julian Schoep, Erman Acar, Nanne Van Noord, and Pascal Mettes. Hyperbolic image segmentation. In CVPR, pages 4453-4462, 2022. 2 +[3] Gary Bécigneul and Octavian-Eugen Ganea. Riemannian adaptive optimization methods. arXiv preprint arXiv:1810.00760, 2018. 2 +[4] P Bedapudi. NudeNet: Neural Nets for Nudity Classification, Detection, and Selective Censoring, 2019. 2, 5, 7, 8 +[5] Abeba Birhane and Vinay Uday Prabhu. Large image datasets: A pyrrhic win for computer vision? In WACV, pages 1536-1546. IEEE, 2021. 1, 2, 3, 5 +[6] Abeba Birhane, Vinay Uday Prabhu, and Emmanuel Kahembwe. Multimodal datasets: misogyny, pornography, and malignant stereotypes. arXiv preprint arXiv:2110.01963, 2021. 1 +[7] Abeba Birhane, Sanghyun Han, Vishnu Boddeti, Sasha Luccioni, et al. Into the laion's den: Investigating hate in multimodal datasets. NeurIPS, 36, 2024. 1 +[8] Manuel Brack, Felix Friedrich, Patrick Schramowski, and Kristian Kersting. Mitigating inappropriateness in image generation: Can there be value in reflecting the world's ugliness? arXiv preprint arXiv:2305.18398, 2023. 2 +[9] Yinzhi Cao and Junfeng Yang. Towards Making Systems Forget with Machine Unlearning. In IEEE Symposium on Security and Privacy, 2015. 2 +[10] Francesco Cauteruccio, Enrico Corradini, Giorgio Terracina, Domenico Ursino, and Luca Virgili. Extraction and analysis of text patterns from nsfw adult content in reddit. Data & Knowledge Engineering, 138:101979, 2022. 2 +[11] Benjamin Paul Chamberlain, James Clough, and Marc Peter Deisenroth. Neural embeddings of graphs in hyperbolic space. arXiv preprint arXiv:1705.10359, 2017. 2 +[12] Ines Chami, Adva Wolf, Da-Cheng Juan, Frederic Sala, Sujith Ravi, and Christopher Ré. Low-dimensional hyperbolic knowledge graph embeddings. arXiv preprint arXiv:2005.00545, 2020. 2 +[13] Nurendra Choudhary, Nikhil Rao, Sumeet Katariya, Karthik Subbian, and Chandan K Reddy. Self-supervised hyperboloid representations from logical queries over knowledge + +graphs. In Proceedings of the Web Conference 2021, pages 1373-1384, 2021. +[14] Nurendra Choudhary, Nikhil Rao, and Chandan Reddy. Hyperbolic graph neural networks at scale: a meta learning approach. NeurIPS, 36, 2024. 2 +[15] Damien L Crone, Stefan Bode, Carsten Murawski, and Simon M Laham. The socio-moral image database (smid): A novel stimulus set for the study of social, moral and affective processes. *PloS one*, 13(1):e0190954, 2018. 5, 7 +[16] Karan Desai, Gaurav Kaul, Zubin Aysola, and Justin Johnson. RedCaps: Web-curated image-text data created by the people, for the people. In NeurIPS Datasets and Benchmarks, 2021. 6 +[17] Karan Desai, Maximilian Nickel, Tanmay Rajpurohit, Justin Johnson, and Shanmukha Ramakrishna Vedantam. Hyperbolic image-text representations. In ICML, pages 7694-7731. PMLR, 2023. 1, 2, 3, 4, 6 +[18] Bhuwan Dhingra, Christopher J Shallue, Mohammad Norouzi, Andrew M Dai, and George E Dahl. Embedding text in hyperbolic spaces. arXiv preprint arXiv:1806.04313, 2018. 2 +[19] Upol Ehsan, Q Vera Liao, Samir Passi, Mark O Riedl, and Hal Daumé III. Seamful xai: Operationalizing seamful design in explainable ai. Proceedings of the ACM on Human-Computer Interaction, 8(CSCW1):1-29, 2024. 1 +[20] Aleksandr Ermolov, Leyla Mirvakhabova, Valentin Khrulkov, Nicu Sebe, and Ivan Oseledets. Hyperbolic vision transformers: Combining improvements in metric learning. In CVPR, pages 7409-7419, 2022. 2 +[21] Luca Franco, Paolo Mandica, Bharti Munjal, and Fabio Galasso. Hyperbolic self-paced learning for self-supervised skeleton-based action representations. In ICLR, 2023. 2 +[22] Shreyansh Gandhi, Samrat Kokkula, Abon Chaudhuri, Alessandro Magnani, Theban Stanley, Behzad Ahmadi, Venkatesh Kandaswamy, Omer Ovenc, and Shie Mannor. Scalable Detection of Offensive and Non-compliant Content/Logo in Product Images. In WACV, 2020. 2 +[23] Rohit Gandikota, Joanna Materzynska, Jaden Fiotto-Kaufman, and David Bau. Erasing Concepts from Diffusion Models. In ICCV, 2023. 1, 2 +[24] Octavian Ganea, Gary Bécigneul, and Thomas Hofmann. Hyperbolic entailment cones for learning hierarchical embeddings. In ICML, pages 1646-1655. PMLR, 2018. 1, 2, 4 +[25] Octavian Ganea, Gary Bécigneul, and Thomas Hofmann. Hyperbolic neural networks. NeurIPS, 31, 2018. 2 +[26] Songwei Ge, Shlok Mishra, Simon Kornblith, Chun-Liang Li, and David Jacobs. Hyperbolic contrastive learning for visual representations beyond objects. In CVPR, pages 6840-6849, 2023. 2 +[27] Antonio Ginart, Melody Guan, Gregory Valiant, and James Y Zou. Making AI Forget You: Data Deletion in Machine Learning. In NeurIPS, 2019. 2 +[28] Aditya Golatkar, Alessandro Achille, and Stefano Soatto. Eternal Sunshine of the Spotless Net: Selective Forgetting in Deep Networks. In CVPR, 2020. 2 + +[29] Aditya Golatkar, Alessandro Achille, Yu-Xiang Wang, Aaron Roth, Michael Kearns, and Stefano Soatto. Mixed Differential Privacy in Computer Vision. In CVPR, 2022. 2 +[30] Kimia Hamidieh, Haoran Zhang, Walter Gerych, Thomas Hartvigsen, and Marzyeh Ghassemi. Identifying implicit social biases in vision-language models. In Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society, pages 547-561, 2024. 1 +[31] Ahmad Fathan Hidayatullah, Anisa Miladya Hakim, and Abdullah Aziz Sembada. Adult Content Classification on Indonesian Tweets using LSTM Neural Network. In ICACSIS, 2019. 2 +[32] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 5 +[33] Sarah Ibrahimi, Mina Ghadimi Atigh, Nanne Van Noord, Pascal Mettes, and Marcel Worring. Intriguing properties of hyperbolic embeddings in vision-language models. Transactions on Machine Learning Research, 2024. 2 +[34] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, pages 4904-4916. PMLR, 2021. 1 +[35] Valentin Khrulkov, Leyla Mirvakhabova, Evgeniya Ustinova, Ivan Oseledets, and Victor Lempitsky. Hyperbolic image embeddings. In CVPR, pages 6418-6428, 2020. 2, 3 +[36] Nupur Kumari, Bingliang Zhang, Sheng-Yu Wang, Eli Shechtman, Richard Zhang, and Jun-Yan Zhu. Ablating concepts in text-to-image diffusion models. In ICCV, 2023. 2 +[37] Gant Laborde. Deep nn for nsfw detection. https://github.com/GantMan/nsfw_model, 2020.8 +[38] Matthew Le, Stephen Roller, Laetitia Papaxanthos, Douwe Kiela, and Maximilian Nickel. Inferring concept hierarchies from text corpora via hyperbolic embeddings. In ACL, pages 3231-3241, 2019. 2, 4 +[39] Ya-Wei Eileen Lin, Ronald R Coifman, Gal Mishne, and Ronen Talmon. Hyperbolic diffusion embedding and distance for hierarchical representation learning. In ICML, pages 21003–21025. PMLR, 2023. 2 +[40] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 36, 2024. 1 +[41] Qi Liu, Maximilian Nickel, and Douwe Kiela. Hyperbolic graph neural networks. NeurIPS, 32, 2019. 2 +[42] Shaoteng Liu, Jingjing Chen, Liangming Pan, Chong-Wah Ngo, Tat-Seng Chua, and Yu-Gang Jiang. Hyperbolic visual embedding learning for zero-shot recognition. In CVPR, pages 9273-9281, 2020. 2 +[43] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 5 +[44] Todor Markov, Chong Zhang, Sandhini Agarwal, Florentine Eloundou Nekoul, Theodore Lee, Steven Adler, Angela Jiang, and Lilian Weng. A Holistic Approach to Undesired Content Detection in the Real World. In AAAI, 2023. 2 +[45] Joanna Materzyńska, Antonio Torralba, and David Bau. Disentangling visual and written concepts in clip. In CVPR, pages 16410-16419, 2022. 1 + +[46] Pascal Mettes, Mina Ghadimi Atigh, Martin Keller-Ressel, Jeffrey Gu, and Serena Yeung. Hyperbolic deep learning in computer vision: A survey. IJCV, pages 1-25, 2024. 2 +[47] Paulius Micikevicius, Sharan Narang, Jonah Alben, Gregory Diamos, Erich Elsen, David Garcia, Boris Ginsburg, Michael Houston, Oleksii Kuchaiev, Ganesh Venkatesh, et al. Mixed precision training. In ICLR, 2018. 5 +[48] Leyla Mirvakhabova, Evgeny Frolov, Valentin Khrulkov, Ivan Oseledets, and Alexander Tuzhilin. Performance of hyperbolic geometry models on top-n recommendation tasks. In Proceedings of the 14th ACM Conference on Recommender Systems, pages 527-532, 2020. 2 +[49] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards Photorealistic Image Generation and Editing with Text-Guided Diffusion Models. arXiv preprint arXiv:2112.10741, 2021. 2 +[50] Maximillian Nickel and Douwe Kiela. Poincaré embeddings for learning hierarchical representations. NeurIPS, 30, 2017. 2 +[51] Maximillian Nickel and Douwe Kiela. Learning continuous hierarchies in the lorentz model of hyperbolic geometry. In ICML, pages 3779-3788. PMLR, 2018. 2 +[52] Avik Pal, Max van Spengler, Guido Maria D'Amely di Melendugno, Alessandro Flaborea, Fabio Galasso, and Pascal Mettes. Compositional entailment learning for hyperbolic vision-language models. arXiv preprint arXiv:2410.06912, 2024. 1, 2, 3, 4, 6 +[53] Wei Peng, Tuomas Varanka, Abdelrahman Mostafa, Henglin Shi, and Guoying Zhao. Hyperbolic deep neural networks: A survey. IEEE Trans. PAMI, 44(12):10023-10044, 2021. 2 +[54] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 6 +[55] Samuele Poppi, Tobia Poppi, Federico Cocchi, Marcella Cornia, Lorenzo Baraldi, and Rita Cucchiara. Safe-CLIP: Removing NSFW Concepts from Vision-and-Language Models. In ECCV, 2024. 1, 2, 3, 4, 5, 6, 7 +[56] Samuele Poppi, Sara Sarto, Marcella Cornia, Lorenzo Baraldi, and Rita Cucchiara. Multi-Class Unlearning for Image Classification via Weight Filtering. IEEE Intelligent Systems, 2024. 2 +[57] Samuele Poppi, Sara Sarto, Marcella Cornia, Lorenzo Baraldi, and Rita Cucchiara. Unlearning vision transformers without retaining data via low-rank decompositions. In ICPR, 2024. 2 +[58] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 1, 3, 4, 5, 6, 7 +[59] Javier Rando, Daniel Paleka, David Lindner, Lennart Heim, and Florian Tramer. Red-teaming the stable diffusion safety filter. In NeurIPS ML Safety Workshop, 2022. 8 + +[60] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 1 +[61] Frederic Sala, Chris De Sa, Albert Gu, and Christopher Ré. Representation tradeoffs for hyperbolic embeddings. In ICML, pages 4460-4469. PMLR, 2018. 2 +[62] Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108, 2019. 2 +[63] Patrick Schramowski, Christopher Tauchmann, and Kristian Kersting. Can Machines Help Us Answering Question 16 in Datasheets, and In Turn Reflecting on Inappropriate Content? In ACM FAccT, 2022. 2, 8 +[64] Patrick Schramowski, Manuel Brack, Björn Deiseroth, and Kristian Kersting. Safe Latent Diffusion: Mitigating Inappropriate Degeneration in Diffusion Models. In CVPR, 2023. 2, 3 +[65] Christoph Schuhmann. Clip based nsfw detector. https://github.com/LAION-AI/CLIP-based-NSFW-Detector/tree/main, 2022.8 +[66] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 1, 5, 7 +[67] Sheng Shen, Liunian Harold Li, Hao Tan, Mohit Bansal, Anna Rohrbach, Kai-Wei Chang, Zhewei Yao, and Kurt Keutzer. How much can clip benefit vision-and-language tasks? arXiv preprint arXiv:2107.06383, 2021. 1 +[68] Ryohei Shimizu, Yusuke Mukuta, and Tatsuya Harada. Hyperbolic neural networks++. ICLR, 2021. 2 +[69] David Thiel. Identifying and eliminating csam in generative ml training data and models. 2023. 1 +[70] Alexandru Tifrea, Gary Bécigneul, and Octavian-Eugen Ganea. Poincaré glove: Hyperbolic word embeddings. arXiv preprint arXiv:1810.06546, 2018. 2 +[71] Max van Spengler, Erwin Berkhout, and Pascal Mettes. Poincare resnet. In ICCV, pages 5419-5428, 2023. 2 +[72] Ivan Vendrov, Ryan Kiros, Sanja Fidler, and Raquel Urtasun. Order-embeddings of images and language. arXiv preprint arXiv:1511.06361, 2015. 4 +[73] Liping Wang, Fenyu Hu, Shu Wu, and Liang Wang. Fully hyperbolic graph convolution network for recommendation. In Proceedings of the 30th ACM international conference on information & knowledge management, pages 3483-3487, 2021. 2 +[74] Mengmeng Wang, Jiazheng Xing, and Yong Liu. Actionclip: A new paradigm for video action recognition. arXiv preprint arXiv:2109.08472, 2021. 1 +[75] Robert Wolfe, Yiwei Yang, Bill Howe, and Aylin Caliskan. Contrastive language-vision ai models pretrained on web-scraped multimodal data exhibit sexual objectification bias. In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency, pages 1174-1185, 2023. 1 + +[76] Eric Zhang, Kai Wang, Xingqian Xu, Zhangyang Wang, and Humphrey Shi. Forget-Me-Not: Learning to Forget in Text-to-Image Diffusion Models. CVPR, 2024. 2 +[77] Yudong Zhu, Di Zhou, Jinghui Xiao, Xin Jiang, Xiao Chen, and Qun Liu. Hypertext: Endowing fasttext with hyperbolic geometry. arXiv preprint arXiv:2010.16143, 2020. 2 \ No newline at end of file diff --git a/2025/Hyperbolic Safety-Aware Vision-Language Models/images.zip b/2025/Hyperbolic Safety-Aware Vision-Language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6b8132b90251e7646f3e0a5325bd91120c39784a --- /dev/null +++ b/2025/Hyperbolic Safety-Aware Vision-Language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6f46c193f2e40fbd1f08163bb5b5d7e18aad466177bec26d9b025379e6831c68 +size 436130 diff --git a/2025/Hyperbolic Safety-Aware Vision-Language Models/layout.json b/2025/Hyperbolic Safety-Aware Vision-Language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8a9c4afeb3e0c29e84e6d2d3ea60b715d0af5488 --- /dev/null +++ b/2025/Hyperbolic Safety-Aware Vision-Language Models/layout.json @@ -0,0 +1,11035 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 148, + 103, + 463, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 103, + 463, + 121 + ], + "spans": [ + { + "bbox": [ + 148, + 103, + 463, + 121 + ], + "type": "text", + "content": "Hyperbolic Safety-Aware Vision-Language Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 77, + 142, + 533, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 142, + 533, + 185 + ], + "spans": [ + { + "bbox": [ + 77, + 142, + 533, + 185 + ], + "type": "text", + "content": "Tobia Poppi\\*1,2 Tejaswi Kasarla\\*3 Pascal Mettes3 Lorenzo Baraldi1 Rita Cucchiara1,4 1University of Modena and Reggio Emilia, Italy 2 University of Pisa, Italy 3 University of Amsterdam, Netherlands 4IIT-CNR, Italy" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 72, + 186, + 531, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 186, + 531, + 200 + ], + "spans": [ + { + "bbox": [ + 72, + 186, + 531, + 200 + ], + "type": "inline_equation", + "content": "^{1}\\{\\text {name.surname}\\} @ \\text {unimore.it}" + }, + { + "bbox": [ + 72, + 186, + 531, + 200 + ], + "type": "inline_equation", + "content": "^{2}\\{\\text {name.surname}\\} @ \\text {phd.unipi.it}" + }, + { + "bbox": [ + 72, + 186, + 531, + 200 + ], + "type": "inline_equation", + "content": "^{3}\\{\\text {initial.surname}\\} @ \\text {uva.nl}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "spans": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 251, + 297, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 297, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 297, + 563 + ], + "type": "text", + "content": "Addressing the retrieval of unsafe content from vision-language models such as CLIP is an important step towards real-world integration. Current efforts have relied on unlearning techniques that try to erase the model's knowledge of unsafe concepts. While effective in reducing unwanted outputs, unlearning limits the model's capacity to discern between safe and unsafe content. In this work, we introduce a novel approach that shifts from unlearning to an awareness paradigm by leveraging the inherent hierarchical properties of the hyperbolic space. We propose to encode safe and unsafe content as an entailment hierarchy, where both are placed in different regions of hyperbolic space. Our HySAC, Hyperbolic Safety-Aware CLIP, employs entailment loss functions to model the hierarchical and asymmetrical relations between safe and unsafe image-text pairs. This modelling - ineffective in standard vision-language models due to their reliance on Euclidean embeddings - endows the model with awareness of unsafe content, enabling it to serve as both a multimodal unsafe classifier and a flexible content retriever, with the option to dynamically redirect unsafe queries toward safer alternatives or retain the original output. Extensive experiments show that our approach not only enhances safety recognition but also establishes a more adaptable and interpretable framework for content moderation in vision-language models. Our source code is available at: https://github.com/aimagelab/HySAC" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 569, + 296, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 569, + 296, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 569, + 296, + 605 + ], + "type": "text", + "content": "Warning: This paper features explicit sexual content and other material that some readers may find disturbing, distressing, or offensive." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 627, + 135, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 627, + 135, + 639 + ], + "spans": [ + { + "bbox": [ + 56, + 627, + 135, + 639 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 647, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 296, + 696 + ], + "type": "text", + "content": "Large-scale vision-language models (VLMs) have achieved remarkable successes in various applications, including cross-model retrieval [58], text-to-image and image-to-text generation [40, 60] and various downstream tasks [45, 67," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 228, + 555, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 555, + 407 + ], + "type": "text", + "content": "74]. Popular VLMs like CLIP [58] and ALIGN [34] leverage vast amounts of web-scraped image-text data to learn rich multimodal representations by aligning visual and textual modalities. However, most large-scale datasets sourced from the web contain unsafe or inappropriate content, such as violence, nudity, or hate speech [5, 6]. The presence of such content not only raises ethical concerns but also introduces risks for real-world applications [7, 30, 75], where exposure to or misuse of this material can lead to legal and societal repercussions. Birhane et al. [7] also show that the increasing dataset scale can exacerbate hateful and unsafe content, as identified in the now removed LAION-5B [66, 69]. Addressing the issue of unsafe content in VLMs is therefore of utmost importance to ensure responsible AI practices." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 410, + 556, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 410, + 556, + 530 + ], + "spans": [ + { + "bbox": [ + 313, + 410, + 556, + 530 + ], + "type": "text", + "content": "Recent efforts to mitigate unsafe content in vision-language models have led to the development of methods specifically designed for NSFW (Not Safe For Work) content removal. Most of these works [23, 55] have focused on unlearning (i.e., removing) the knowledge of unsafe content from the models. As a recent example, Poppi et al. [55] develop a fine-tuned version of CLIP which unlearns toxic concepts by redirecting their embeddings towards safe regions, so that retrieval always produces safe content even when the model is prompted with unsafe inputs." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 532, + 556, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 532, + 556, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 532, + 556, + 628 + ], + "type": "text", + "content": "In contrast, we propose an approach for managing unsafe content in VLMs: emphasizing awareness over unlearning. Rather than hiding the flaws of VLMs by ignoring NSFW content, we aim to equip the VLMs with the ability to distinguish between safe and unsafe content. This in turn helps users of the model to expose or redirect NSFW content when necessary, a crucial step toward improving user agency, understanding, and interpretability [19]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "Inspired by recent hyperbolic vision-language models [17, 52], we introduce a hyperbolic framework that leverages the geometric properties of hyperbolic space to separate safe and unsafe content effectively. Using a paired dataset of safe and unsafe image-text inputs [55], we adjust the embeddings to create an entailment-based [24] structure. In this setup, safe concepts are positioned closer to the" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 702, + 133, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 133, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 133, + 712 + ], + "type": "text", + "content": "*Equal contribution" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4222" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 262 + ], + "type": "text", + "content": "origin of the hyperbolic space, while unsafe concepts are mapped further away. Specifically, we introduce a hyperbolic safe-to-unsafe entailment mechanism that ensures safe content encompasses unsafe representations within conical regions, defining clear safety boundaries and safety traversals to dynamically adjust query embeddings along the hyperbolic space to promote safe retrievals or, alternatively, expose relevant unsafe content when necessary. This framework not only organizes data into safe and unsafe radius-based regions but also enables controlled movement within the space, allowing retrievals to favor safety as required. Experiments demonstrate that HySAC achieves clear improvements in safety awareness, retrieval performance, and NSFW content handling across multiple datasets, with robustness in both safe content redirection and controlled unsafe content accessibility." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 277, + 142, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 142, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 142, + 289 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 300, + 295, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 295, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 295, + 396 + ], + "type": "text", + "content": "Unlearning in vision-language models. Unlearning concepts and content has recently received a lot of attention, empowered by the success of vision-language models. Various approaches have been explored, such as full model retraining, fine-tuning, machine unlearning [9, 27, 28, 56, 57], and differential privacy [29]. Some of these efforts have focused on text-to-image models, with the goal of removing specific styles, concepts, or objects [36, 76]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 398, + 295, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 398, + 295, + 576 + ], + "spans": [ + { + "bbox": [ + 55, + 398, + 295, + 576 + ], + "type": "text", + "content": "A particular emphasis has been placed on removing NSFW content, which encompasses inappropriate, unsafe, or illegal material. Schramowsky et al. [64] steer the generation away from NSFW areas, defined by a fixed set of concepts. The embedding of NSFW concepts is applied as negative guidance during the text-conditioning phase, acting as safety guidance. Gandikota et al. [23] erase visual concepts using only their name, with negative guidance serving as a teacher. Poppi et al. [55] focus on removing NSFW concepts from CLIP-like models by fine-tuning the entire model using ViSU, a multimodal dataset containing Safe/NSFW Image/Texts quadruplets. Unlike earlier approaches, which target specific components, their method fine-tunes the entire vision-and-language model, making content removal applicable to downstream tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 578, + 295, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 578, + 295, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 578, + 295, + 685 + ], + "type": "text", + "content": "We introduce a method for handling NSFW concepts in CLIP-like models by fine-tuning them in hyperbolic space. We exploit the hierarchical properties of hyperbolic geometry, yielding a clear advantage: the model becomes explicitly aware of whether content is safe or NSFW, rather than merely removing knowledge of unsafe content. Similarly to our approach, safe generation works [8, 64], underscore that a deeper understanding of (un-)safety can improve the control of harmful content generation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "NSFW concept detection. A related area of research is the automatic detection of NSFW content. Several meth" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "type": "text", + "content": "ods have been proposed for detecting NSFW and toxic language [10, 31, 44], primarily in social media contexts. DistilBERT [62] has emerged as a promising model for this task, especially when fine-tuned for identifying adult content. Detecting inappropriate language presents a significant challenge, and this complexity extends to visual content, where various techniques have been developed to detect NSFW imagery [5, 22, 49]. In this domain, models like NudeNet [4] specialize in detecting nudity, while Q16 [63] serves as a broader classifier, capable of identifying a wider range of NSFW content. However, identifying inappropriate visual content remains a complex task, given the challenges posed by subtle visual cues, lack of contextual information, and limited data availability." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 239, + 553, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 553, + 312 + ], + "type": "text", + "content": "While these detection methods focus solely on identifying unsafe content, they do not address the problem of retrieving relevant, safe alternatives when an unsafe input is detected. Our method makes it possible to jointly detect NSFW content and provide a mechanism to shift NSFW queries towards safe but relevant alternatives." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 314, + 553, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 314, + 553, + 433 + ], + "spans": [ + { + "bbox": [ + 313, + 314, + 553, + 433 + ], + "type": "text", + "content": "Hyperbolic learning. A key advantage of hyperbolic space is its inherent ability to represent hierarchical or tree-like structures with minimal distortion [11, 50, 51, 61]. A comprehensive list of recent advancements is documented in surveys by Mettes et al. [46] and Peng et al. [53]. Foundational works for building neural networks in hyperbolic space [3, 24, 25, 35, 68] led to the use of hyperbolic models across multiple modalities such as images [2, 20, 21, 26, 42, 71], text [18, 38, 70, 77], graphs [12-14, 41] and recommender systems [48, 73]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 434, + 553, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 553, + 576 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 553, + 576 + ], + "type": "text", + "content": "Recent work has shown the strong potential of hyperbolic learning for vision-language models [17, 33, 52] and demonstrated that training with a loss function enforcing entailment cones [24] leads to the emergence of hierarchical structures between the embeddings. While Desai et al. [17] enforce entailment structure across modalities, Pal et al. [52] also explicitly enforce structure within modalities by leveraging object-level compositions and nouns from text. We take inspiration from these approaches and enforce entailment relations across safe and unsafe embeddings, creating an interpretable space for traversing from unsafe regions to safe regions in CLIP models." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 586, + 399, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 586, + 399, + 597 + ], + "spans": [ + { + "bbox": [ + 314, + 586, + 399, + 597 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 605, + 553, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 553, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 553, + 665 + ], + "type": "text", + "content": "Throughout this work, we operate in hyperbolic space, a Riemannian manifold with a constant negative curvature. Following Desai et al. [17], we use the Lorentz model, as it is better equipped to deal with numerical instabilities associated with the Poincaré distance metric [39, 51]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": "The Lorentz model " + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathbb{L}^n" + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": " is an " + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": "-dimensional manifold in which points are represented on the upper sheet of a two-sheeted hyperboloid in " + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "inline_equation", + "content": "(n + 1)" + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": "-dimensional Minkowski spacetime. Following terminology from general relativity," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4223" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 550, + 238 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 550, + 238 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 550, + 238 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 550, + 238 + ], + "type": "image", + "image_path": "3ff6f18357e9d4e4959c4817dac3667620f8df1c865174840c71f5c6e255466b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 247, + 555, + 270 + ], + "lines": [ + { + "bbox": [ + 55, + 247, + 555, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 555, + 270 + ], + "type": "text", + "content": "Figure 1. Overview of our approach. HySAC builds a hyperbolic embedding that manages content safety through an entailment hierarchy. Unsafe text and images are projected to dedicated regions of hyperbolic space, allowing for safety-aware retrieval and classification." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 282, + 296, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 282, + 296, + 332 + ], + "spans": [ + { + "bbox": [ + 55, + 282, + 296, + 332 + ], + "type": "text", + "content": "for each vector " + }, + { + "bbox": [ + 55, + 282, + 296, + 332 + ], + "type": "inline_equation", + "content": "\\mathbf{p} \\in \\mathbb{R}^{n+1}" + }, + { + "bbox": [ + 55, + 282, + 296, + 332 + ], + "type": "text", + "content": ", we refer to the first dimension, the axis of symmetry, as the time-axis, denoted by " + }, + { + "bbox": [ + 55, + 282, + 296, + 332 + ], + "type": "inline_equation", + "content": "p_0" + }, + { + "bbox": [ + 55, + 282, + 296, + 332 + ], + "type": "text", + "content": ", and the remaining " + }, + { + "bbox": [ + 55, + 282, + 296, + 332 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 282, + 296, + 332 + ], + "type": "text", + "content": "-dimensions as the spatial components, denoted " + }, + { + "bbox": [ + 55, + 282, + 296, + 332 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{p}}" + }, + { + "bbox": [ + 55, + 282, + 296, + 332 + ], + "type": "text", + "content": ". The Lorentz model, " + }, + { + "bbox": [ + 55, + 282, + 296, + 332 + ], + "type": "inline_equation", + "content": "\\mathbb{L}^n = (\\mathcal{L}^n, \\mathfrak{g}_p^\\kappa)" + }, + { + "bbox": [ + 55, + 282, + 296, + 332 + ], + "type": "text", + "content": " is given as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 338, + 296, + 371 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 338, + 296, + 371 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 296, + 371 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {n} := \\left\\{\\boldsymbol {p} \\in \\mathbb {R} ^ {n + 1}: \\langle \\mathbf {p}, \\mathbf {p} \\rangle_ {\\mathcal {L}} = - \\frac {1}{\\kappa}, p _ {0} = \\sqrt {1 / \\kappa + \\| \\tilde {\\boldsymbol {p}} \\| ^ {2}}, \\kappa > 0 \\right\\}, \\tag {1}", + "image_path": "cdeb23418aba0aa748b52d48ce61bf6bba32ab1c7da1120d7c69ab398e86b042.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 372, + 296, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 372, + 296, + 421 + ], + "spans": [ + { + "bbox": [ + 55, + 372, + 296, + 421 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 372, + 296, + 421 + ], + "type": "inline_equation", + "content": "-\\kappa \\in \\mathbb{R}" + }, + { + "bbox": [ + 55, + 372, + 296, + 421 + ], + "type": "text", + "content": " denotes the curvature with the Riemannian metric " + }, + { + "bbox": [ + 55, + 372, + 296, + 421 + ], + "type": "inline_equation", + "content": "\\mathfrak{g}_p^\\kappa = \\mathrm{diag}(-1,1,\\ldots ,1)" + }, + { + "bbox": [ + 55, + 372, + 296, + 421 + ], + "type": "text", + "content": ". The Lorentzian inner product " + }, + { + "bbox": [ + 55, + 372, + 296, + 421 + ], + "type": "inline_equation", + "content": "\\langle .,.\\rangle_{\\mathcal{L}}" + }, + { + "bbox": [ + 55, + 372, + 296, + 421 + ], + "type": "text", + "content": " is induced by the metric tensor " + }, + { + "bbox": [ + 55, + 372, + 296, + 421 + ], + "type": "inline_equation", + "content": "\\mathfrak{g}_p^\\kappa" + }, + { + "bbox": [ + 55, + 372, + 296, + 421 + ], + "type": "text", + "content": " and is defined for " + }, + { + "bbox": [ + 55, + 372, + 296, + 421 + ], + "type": "inline_equation", + "content": "\\mathbf{p},\\mathbf{q}\\in \\mathbb{L}^n" + }, + { + "bbox": [ + 55, + 372, + 296, + 421 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 119, + 426, + 294, + 439 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 426, + 294, + 439 + ], + "spans": [ + { + "bbox": [ + 119, + 426, + 294, + 439 + ], + "type": "interline_equation", + "content": "\\langle \\mathbf {p}, \\mathbf {q} \\rangle_ {\\mathcal {L}} = - p _ {0} q _ {0} + \\langle \\tilde {\\boldsymbol {p}}, \\tilde {\\boldsymbol {q}} \\rangle , \\tag {2}", + "image_path": "425fff5552e77747a4aeffde7c61723cca9ce9efa4acb692c6c4c9544bd94bd0.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 443, + 296, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 443, + 296, + 492 + ], + "spans": [ + { + "bbox": [ + 55, + 443, + 296, + 492 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 443, + 296, + 492 + ], + "type": "inline_equation", + "content": "\\langle .,.\\rangle" + }, + { + "bbox": [ + 55, + 443, + 296, + 492 + ], + "type": "text", + "content": " is the Euclidean inner product. The Lorentzian inner product induces a norm on the Lorentzian space which can be written as " + }, + { + "bbox": [ + 55, + 443, + 296, + 492 + ], + "type": "inline_equation", + "content": "\\| \\mathbf{p}\\|_{\\mathcal{L}} = \\sqrt{\\langle\\mathbf{p},\\mathbf{p}\\rangle_{\\mathcal{L}}}" + }, + { + "bbox": [ + 55, + 443, + 296, + 492 + ], + "type": "text", + "content": ". We now define the common hyperbolic operations in Lorentz space." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 496, + 296, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 496, + 296, + 532 + ], + "spans": [ + { + "bbox": [ + 55, + 496, + 296, + 532 + ], + "type": "text", + "content": "Definition 3.1 (Lorentzian distance). The Lorentzian distance between two points in " + }, + { + "bbox": [ + 55, + 496, + 296, + 532 + ], + "type": "inline_equation", + "content": "\\mathbb{L}^n" + }, + { + "bbox": [ + 55, + 496, + 296, + 532 + ], + "type": "text", + "content": " is the length of their shortest path (geodesic) connecting them, computed as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 91, + 536, + 294, + 552 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 536, + 294, + 552 + ], + "spans": [ + { + "bbox": [ + 91, + 536, + 294, + 552 + ], + "type": "interline_equation", + "content": "d _ {\\mathcal {L}} (\\mathbf {p}, \\mathbf {q}) = \\sqrt {1 / \\kappa} \\cdot \\cosh^ {- 1} (- \\kappa \\langle \\mathbf {p}, \\mathbf {q} \\rangle_ {\\mathcal {L}}). \\tag {3}", + "image_path": "0b6aa5dc16048f4eb73a93413d092c4b4a3957094a631401df290547747258a8.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 559, + 296, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 559, + 296, + 583 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 296, + 583 + ], + "type": "text", + "content": "In our work, we use the negative of Lorentzian distance to calculate similarities between multimodal inputs [17, 52]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 587, + 296, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 587, + 296, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 587, + 296, + 685 + ], + "type": "text", + "content": "Definition 3.2 (Exponential map). Since Lorentz space is a Riemannian manifold, it is locally Euclidean. This is best described through the tangent space " + }, + { + "bbox": [ + 55, + 587, + 296, + 685 + ], + "type": "inline_equation", + "content": "T_{\\mathbf{p}}\\mathbb{L}^n" + }, + { + "bbox": [ + 55, + 587, + 296, + 685 + ], + "type": "text", + "content": ", a first-order approximation of the Lorentzian manifold at a given point " + }, + { + "bbox": [ + 55, + 587, + 296, + 685 + ], + "type": "inline_equation", + "content": "\\mathbf{p}\\in \\mathbb{L}^n" + }, + { + "bbox": [ + 55, + 587, + 296, + 685 + ], + "type": "text", + "content": ". The exponential map then provides a means to project elements from the tangent space onto the hyperboloid. Given a point " + }, + { + "bbox": [ + 55, + 587, + 296, + 685 + ], + "type": "inline_equation", + "content": "\\mathbf{v}\\in T_{\\mathbf{p}}\\mathbb{L}^n" + }, + { + "bbox": [ + 55, + 587, + 296, + 685 + ], + "type": "text", + "content": ", the exponential map is defined as " + }, + { + "bbox": [ + 55, + 587, + 296, + 685 + ], + "type": "inline_equation", + "content": "\\exp_{\\mathbf{p}}^{\\kappa}\\colon T_{\\mathbf{p}}\\mathbb{L}^{n}\\to \\mathbb{L}^{n}" + }, + { + "bbox": [ + 55, + 587, + 296, + 685 + ], + "type": "text", + "content": " with the expression" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 66, + 689, + 296, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 689, + 296, + 717 + ], + "spans": [ + { + "bbox": [ + 66, + 689, + 296, + 717 + ], + "type": "interline_equation", + "content": "\\exp_ {\\mathbf {p}} ^ {\\kappa} (\\mathbf {v}) = \\cosh (\\sqrt {\\kappa} \\| \\mathbf {v} \\| _ {\\mathbb {L}}) \\mathbf {p} + \\frac {\\sinh (\\sqrt {\\kappa} \\| \\mathbf {v} \\| _ {\\mathbb {L}})}{\\sqrt {\\kappa} \\| \\mathbf {v} \\| _ {\\mathbb {L}}} \\mathbf {v}. \\tag {4}", + "image_path": "d96c8b1d27b7ab74cde843fddfa72a4b87b08364d16faf96173cb6726ab0623c.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 290, + 555, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 290, + 555, + 362 + ], + "spans": [ + { + "bbox": [ + 313, + 290, + 555, + 362 + ], + "type": "text", + "content": "In practice, the reference point " + }, + { + "bbox": [ + 313, + 290, + 555, + 362 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 313, + 290, + 555, + 362 + ], + "type": "text", + "content": " is set to the origin " + }, + { + "bbox": [ + 313, + 290, + 555, + 362 + ], + "type": "inline_equation", + "content": "\\mathbf{0} = (\\sqrt{1 / \\kappa}, 0, \\dots, 0)^T" + }, + { + "bbox": [ + 313, + 290, + 555, + 362 + ], + "type": "text", + "content": " on the hyperboloid, allowing " + }, + { + "bbox": [ + 313, + 290, + 555, + 362 + ], + "type": "inline_equation", + "content": "\\exp_0^\\kappa" + }, + { + "bbox": [ + 313, + 290, + 555, + 362 + ], + "type": "text", + "content": " to project Euclidean vectors from the tangent space at " + }, + { + "bbox": [ + 313, + 290, + 555, + 362 + ], + "type": "inline_equation", + "content": "\\mathbf{0}" + }, + { + "bbox": [ + 313, + 290, + 555, + 362 + ], + "type": "text", + "content": " directly onto the hyperboloid [17, 35]. In this work, the exponential map is used to project the outputs of the visual and textual encoders to a shared hyperbolic space." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 372, + 536, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 372, + 536, + 385 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 536, + 385 + ], + "type": "text", + "content": "4. HySAC: Hyperbolic Safety-Aware CLIP" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 392, + 501, + 404 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 501, + 404 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 501, + 404 + ], + "type": "text", + "content": "4.1. Problem formulation and objective" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 411, + 555, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 555, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 555, + 519 + ], + "type": "text", + "content": "Problem setup. Given a dataset " + }, + { + "bbox": [ + 313, + 411, + 555, + 519 + ], + "type": "inline_equation", + "content": "D = \\{(I_i, T_i)\\}_{i=1}^N" + }, + { + "bbox": [ + 313, + 411, + 555, + 519 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 411, + 555, + 519 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 411, + 555, + 519 + ], + "type": "text", + "content": " image-text pairs, vision-language models (e.g. CLIP [58]) align the visual and textual embeddings obtained from image and text encoders in a shared embedding space. Large-scale datasets employed for training such embedding spaces are often web-scraped and contain unsafe samples [5]. For our problem setup, in order to differentiate between safe and unsafe content, we denote safe image-text pairs in " + }, + { + "bbox": [ + 313, + 411, + 555, + 519 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 411, + 555, + 519 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 411, + 555, + 519 + ], + "type": "inline_equation", + "content": "(I_i, T_i)" + }, + { + "bbox": [ + 313, + 411, + 555, + 519 + ], + "type": "text", + "content": " and unsafe image-text pairs as " + }, + { + "bbox": [ + 313, + 411, + 555, + 519 + ], + "type": "inline_equation", + "content": "(I_i^\\star, T_i^\\star)" + }, + { + "bbox": [ + 313, + 411, + 555, + 519 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 520, + 556, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 520, + 556, + 675 + ], + "spans": [ + { + "bbox": [ + 313, + 520, + 556, + 675 + ], + "type": "text", + "content": "To make vision-language models aware of inappropriate contents, and enable them to avoid or redirect the representation of such content, we also require a dataset of quadruplets of safe and unsafe image-text pairs, denoted as " + }, + { + "bbox": [ + 313, + 520, + 556, + 675 + ], + "type": "inline_equation", + "content": "D^{\\star} = \\{(I_k, T_k, I_k^{\\star}, T_k^{\\star})\\}_{k=1}^{K}" + }, + { + "bbox": [ + 313, + 520, + 556, + 675 + ], + "type": "text", + "content": " [55]. This dataset is generated following the definition of NSFW content of Schramowski et al. [64] containing the following twenty categories: hate, harassment, violence, suffering, humiliation, harm, suicide, sexual, nudity, bodily fluids, blood, obscene gestures, illegal activity, drug use, theft, vandalism, weapons, abuse, brutality, and cruelty. The dataset is constructed such that the unsafe image-text pairs " + }, + { + "bbox": [ + 313, + 520, + 556, + 675 + ], + "type": "inline_equation", + "content": "(I_k^{\\star}, T_k^{\\star})" + }, + { + "bbox": [ + 313, + 520, + 556, + 675 + ], + "type": "text", + "content": " are specific cases or modified versions of the safe representations " + }, + { + "bbox": [ + 313, + 520, + 556, + 675 + ], + "type": "inline_equation", + "content": "(I_k, T_k)" + }, + { + "bbox": [ + 313, + 520, + 556, + 675 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "Modelling relations in the hyperbolic space. We first consider the relationship between text and image modalities in the embedding space. Like other hyperbolic vision-" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4224" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "text", + "content": "language models [17, 52], we consider text as a general version of images to reflect the natural structure of partial order embeddings [72]. This is enforced in the embedding space by placing text embeddings closer to the origin and image embeddings farther away, defining a modality entailment relationship formally expressed as" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 80, + 152, + 295, + 166 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 152, + 295, + 166 + ], + "spans": [ + { + "bbox": [ + 80, + 152, + 295, + 166 + ], + "type": "interline_equation", + "content": "g _ {T} \\left(T _ {k}\\right) \\ll g _ {I} \\left(I _ {k}\\right), \\quad \\text {a n d} \\quad g _ {T} \\left(T _ {k} ^ {\\star}\\right) \\ll g _ {I} \\left(I _ {k} ^ {\\star}\\right), \\tag {5}", + "image_path": "13140756fa7530793db7e57f7d9c71b632c74f2ca56f71afe8007fb786292277.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 172, + 296, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 172, + 296, + 268 + ], + "spans": [ + { + "bbox": [ + 55, + 172, + 296, + 268 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 172, + 296, + 268 + ], + "type": "inline_equation", + "content": "g_{I}" + }, + { + "bbox": [ + 55, + 172, + 296, + 268 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 172, + 296, + 268 + ], + "type": "inline_equation", + "content": "g_{T}" + }, + { + "bbox": [ + 55, + 172, + 296, + 268 + ], + "type": "text", + "content": " denote the projections of images and text in the hyperbolic model and " + }, + { + "bbox": [ + 55, + 172, + 296, + 268 + ], + "type": "inline_equation", + "content": "\\ll" + }, + { + "bbox": [ + 55, + 172, + 296, + 268 + ], + "type": "text", + "content": " indicates that one embedding is closer to the origin than another. Differently from other models, we also impose a relationship between unsafe and safe pairs, by considering unsafe image-text pairs as specific cases of their safe counterparts within " + }, + { + "bbox": [ + 55, + 172, + 296, + 268 + ], + "type": "inline_equation", + "content": "D^{\\star}" + }, + { + "bbox": [ + 55, + 172, + 296, + 268 + ], + "type": "text", + "content": ". Hence, we establish a safety entailment to segregate safe from unsafe content, as follows" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 135, + 276, + 295, + 290 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 276, + 295, + 290 + ], + "spans": [ + { + "bbox": [ + 135, + 276, + 295, + 290 + ], + "type": "interline_equation", + "content": "g _ {I} \\left(I _ {k}\\right) \\ll g _ {T} \\left(T _ {k} ^ {\\star}\\right). \\tag {6}", + "image_path": "0e204faa08c7a6cd10079c169beccaa578af4bd6899e34392a38593729ff7414.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 297, + 296, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 297, + 296, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 296, + 357 + ], + "type": "text", + "content": "By satisfying these inequalities, " + }, + { + "bbox": [ + 55, + 297, + 296, + 357 + ], + "type": "inline_equation", + "content": "g_{I}" + }, + { + "bbox": [ + 55, + 297, + 296, + 357 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 297, + 296, + 357 + ], + "type": "inline_equation", + "content": "g_{T}" + }, + { + "bbox": [ + 55, + 297, + 296, + 357 + ], + "type": "text", + "content": " capture both the modal and safety hierarchies within the data, thereby endowing the embedding space with safety-aware properties. To sum up, our objective is to model a hyperbolic vision-language model with the following inequality chain:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 92, + 364, + 295, + 378 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 364, + 295, + 378 + ], + "spans": [ + { + "bbox": [ + 92, + 364, + 295, + 378 + ], + "type": "interline_equation", + "content": "g _ {T} \\left(T _ {k}\\right) \\ll g _ {I} \\left(I _ {k}\\right) \\ll g _ {T} \\left(T _ {k} ^ {\\star}\\right) \\ll g _ {I} \\left(I _ {k} ^ {\\star}\\right). \\tag {7}", + "image_path": "0426f7d4de551d4d9a21b47048f1191a9382cdee46c11ba7d875e4d88c8cece3.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 384, + 296, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 296, + 409 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 296, + 409 + ], + "type": "text", + "content": "Below we outline our method to create safety-aware vision-language models in hyperbolic space." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 415, + 203, + 429 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 203, + 429 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 203, + 429 + ], + "type": "text", + "content": "4.2. Hyperbolic safety learning" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 433, + 296, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 433, + 296, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 433, + 296, + 517 + ], + "type": "text", + "content": "To optimize Eq. 7 for safety-aware vision-language models, we propose Hyperbolic Safety-Aware CLIP to rearrange the embedding space to separate safe and unsafe regions. Our optimization consists of two components: (1) a hyperbolic safety contrastive component to align image-text pairs over a mini-batch and (2) a hyperbolic safety entailment to align safe and unsafe content. An overview is given in Figure 1." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 521, + 296, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 296, + 689 + ], + "type": "text", + "content": "Hyperbolic safety contrastive learning. CLIP [58] and Safe-CLIP [55] rely on contrastive objectives to align and distribute the multimodal data. We utilize hyperbolic embeddings to align the visual and textual data. Specifically, we project the visual and textual embeddings from a pre-trained vision-language model onto a hyperboloid [17] through an exponential map (Eq. 4). Let " + }, + { + "bbox": [ + 55, + 521, + 296, + 689 + ], + "type": "inline_equation", + "content": "f_{I}(\\cdot)" + }, + { + "bbox": [ + 55, + 521, + 296, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 521, + 296, + 689 + ], + "type": "inline_equation", + "content": "f_{T}(\\cdot)" + }, + { + "bbox": [ + 55, + 521, + 296, + 689 + ], + "type": "text", + "content": " represent any Euclidean encoders for image and text. Then, " + }, + { + "bbox": [ + 55, + 521, + 296, + 689 + ], + "type": "inline_equation", + "content": "g_{I}(I_{k}) = \\exp_{0}^{\\kappa}(\\alpha_{img} \\cdot f_{I}(I_{k}))" + }, + { + "bbox": [ + 55, + 521, + 296, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 521, + 296, + 689 + ], + "type": "inline_equation", + "content": "g_{T}(T_{k}) = \\exp_{0}^{\\kappa}(\\alpha_{txt} \\cdot f_{T}(T_{k}))" + }, + { + "bbox": [ + 55, + 521, + 296, + 689 + ], + "type": "text", + "content": " represent the hyperbolic representations of a safe image-text pair, " + }, + { + "bbox": [ + 55, + 521, + 296, + 689 + ], + "type": "inline_equation", + "content": "(I_{k}, T_{k})" + }, + { + "bbox": [ + 55, + 521, + 296, + 689 + ], + "type": "text", + "content": ". Similarly, " + }, + { + "bbox": [ + 55, + 521, + 296, + 689 + ], + "type": "inline_equation", + "content": "g_{I}(I_{k}^{\\star}) = \\exp_{0}^{\\kappa}(\\alpha_{img} \\cdot f_{I}(I_{k}^{\\star}))" + }, + { + "bbox": [ + 55, + 521, + 296, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 521, + 296, + 689 + ], + "type": "inline_equation", + "content": "g_{T}(T_{k}^{\\star}) = \\exp_{0}^{\\kappa}(\\alpha_{txt} \\cdot f_{T}(T_{k}^{\\star}))" + }, + { + "bbox": [ + 55, + 521, + 296, + 689 + ], + "type": "text", + "content": " represent the hyperbolic representations of an unsafe image-text pair " + }, + { + "bbox": [ + 55, + 521, + 296, + 689 + ], + "type": "inline_equation", + "content": "(I_{k}^{\\star}, T_{k}^{\\star})" + }, + { + "bbox": [ + 55, + 521, + 296, + 689 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 521, + 296, + 689 + ], + "type": "inline_equation", + "content": "\\alpha_{img}" + }, + { + "bbox": [ + 55, + 521, + 296, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 521, + 296, + 689 + ], + "type": "inline_equation", + "content": "\\alpha_{txt}" + }, + { + "bbox": [ + 55, + 521, + 296, + 689 + ], + "type": "text", + "content": " are learnable projection scalars." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "To align representations in hyperbolic space, the similarity for the image-text and text-image contrastive loss is" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": "based on negative Lorentzian distance (Eq. 3) between " + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "inline_equation", + "content": "g_{I}(.)" + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "inline_equation", + "content": "g_{T}(.)" + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": ". We compute the hyperbolic safety contrastive loss over safe image-text pairs " + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "inline_equation", + "content": "(I_k, T_k)" + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": " in a batch " + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 117, + 554, + 147 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 117, + 554, + 147 + ], + "spans": [ + { + "bbox": [ + 314, + 117, + 554, + 147 + ], + "type": "interline_equation", + "content": "L _ {\\text {c o n t}} ^ {*} (I, T) = - \\sum_ {i \\in B} \\log \\frac {\\exp \\left(d _ {\\mathcal {L}} \\left(g _ {I} \\left(I _ {i}\\right) , g _ {T} \\left(T _ {i}\\right)\\right) / \\tau\\right)}{\\sum_ {k = 1 , k \\neq i} ^ {B} \\exp \\left(d _ {\\mathcal {L}} \\left(g _ {I} \\left(I _ {i}\\right) , g _ {T} \\left(T _ {k}\\right)\\right) / \\tau\\right)}, \\tag {8}", + "image_path": "8cd8b550342b12b5a06820abfcff3bc5ccd8848ff5c5e4367425310798817ebf.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 148, + 555, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 148, + 555, + 219 + ], + "spans": [ + { + "bbox": [ + 313, + 148, + 555, + 219 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 148, + 555, + 219 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 148, + 555, + 219 + ], + "type": "text", + "content": " denotes a temperature hyperparameter. A similar contrastive loss is employed to preserve the multimodal structure between unsafe image-text pairs " + }, + { + "bbox": [ + 313, + 148, + 555, + 219 + ], + "type": "inline_equation", + "content": "(I_k^{\\star}, T_k^{\\star})" + }, + { + "bbox": [ + 313, + 148, + 555, + 219 + ], + "type": "text", + "content": ". Two additional contrastive losses between cross-safety modalities ensure the alignment of quadruplets in the embedding space. The final contrastive loss is formulated as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 324, + 227, + 553, + 256 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 227, + 553, + 256 + ], + "spans": [ + { + "bbox": [ + 324, + 227, + 553, + 256 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {\\mathrm {h S C}} (I, T, I ^ {\\star}, T ^ {\\star}) = L _ {\\text {c o n t}} ^ {\\star} (I, T) + L _ {\\text {c o n t}} ^ {\\star} \\left(I ^ {\\star}, T ^ {\\star}\\right) \\tag {9} \\\\ + L _ {c o n t} ^ {\\star} (I, T ^ {\\star}) + L _ {c o n t} ^ {\\star} (I ^ {\\star}, T). \\\\ \\end{array}", + "image_path": "e4e22538df56b18fa4d0d83a8b81221bed1b2b04f6e0a56f1b70b87dc5a0ee2f.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 274, + 555, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 274, + 555, + 370 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 555, + 370 + ], + "type": "text", + "content": "Hyperbolic safety entailment learning. Hyperbolic entailment cones, introduced by Ganea et al. [24], generalize partial ordered embeddings [72] to any Riemannian manifold. Entailment cones induce a partial order between concepts in a dataset " + }, + { + "bbox": [ + 313, + 274, + 555, + 370 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 313, + 274, + 555, + 370 + ], + "type": "text", + "content": " such that for any pair " + }, + { + "bbox": [ + 313, + 274, + 555, + 370 + ], + "type": "inline_equation", + "content": "(\\mathbf{p},\\mathbf{q})\\in \\mathcal{X}" + }, + { + "bbox": [ + 313, + 274, + 555, + 370 + ], + "type": "text", + "content": ", if " + }, + { + "bbox": [ + 313, + 274, + 555, + 370 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 313, + 274, + 555, + 370 + ], + "type": "text", + "content": " is a sub-concept of " + }, + { + "bbox": [ + 313, + 274, + 555, + 370 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 313, + 274, + 555, + 370 + ], + "type": "text", + "content": ", then " + }, + { + "bbox": [ + 313, + 274, + 555, + 370 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 313, + 274, + 555, + 370 + ], + "type": "text", + "content": " entails " + }, + { + "bbox": [ + 313, + 274, + 555, + 370 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 313, + 274, + 555, + 370 + ], + "type": "text", + "content": " within a conical region " + }, + { + "bbox": [ + 313, + 274, + 555, + 370 + ], + "type": "inline_equation", + "content": "\\mathfrak{S}_{\\mathbf{q}}" + }, + { + "bbox": [ + 313, + 274, + 555, + 370 + ], + "type": "text", + "content": " defined by " + }, + { + "bbox": [ + 313, + 274, + 555, + 370 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 313, + 274, + 555, + 370 + ], + "type": "text", + "content": ". For the Lorentz model, " + }, + { + "bbox": [ + 313, + 274, + 555, + 370 + ], + "type": "inline_equation", + "content": "\\mathbb{L}^n" + }, + { + "bbox": [ + 313, + 274, + 555, + 370 + ], + "type": "text", + "content": ", the half-aperture of each conical region " + }, + { + "bbox": [ + 313, + 274, + 555, + 370 + ], + "type": "inline_equation", + "content": "\\mathfrak{S}_{\\mathbf{q}}" + }, + { + "bbox": [ + 313, + 274, + 555, + 370 + ], + "type": "text", + "content": " is defined as [17, 38]" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 376, + 378, + 554, + 406 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 378, + 554, + 406 + ], + "spans": [ + { + "bbox": [ + 376, + 378, + 554, + 406 + ], + "type": "interline_equation", + "content": "\\omega (\\mathbf {q}) = \\sin^ {- 1} \\left(\\frac {2 K}{\\sqrt {\\kappa} \\| \\tilde {\\boldsymbol {q}} \\|}\\right), \\tag {10}", + "image_path": "ea4612b2709c4bf0773f73e6c3a706b3c49803d52e6c6295e090ce09b1292d5d.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 413, + 555, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 555, + 558 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 555, + 558 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 413, + 555, + 558 + ], + "type": "inline_equation", + "content": "-\\kappa" + }, + { + "bbox": [ + 313, + 413, + 555, + 558 + ], + "type": "text", + "content": " is the curvature of space, and the constant " + }, + { + "bbox": [ + 313, + 413, + 555, + 558 + ], + "type": "inline_equation", + "content": "K = 0.1" + }, + { + "bbox": [ + 313, + 413, + 555, + 558 + ], + "type": "text", + "content": " limits values near the origin [24]. To preserve partial order between image-text relationships, we add entailment from safe text to safe image and from unsafe text to unsafe images, effectively implementing Eq. 5. Specifically, a safe image " + }, + { + "bbox": [ + 313, + 413, + 555, + 558 + ], + "type": "inline_equation", + "content": "I_{k}" + }, + { + "bbox": [ + 313, + 413, + 555, + 558 + ], + "type": "text", + "content": " must lie within the cone defined by its corresponding safe text, " + }, + { + "bbox": [ + 313, + 413, + 555, + 558 + ], + "type": "inline_equation", + "content": "T_{k}" + }, + { + "bbox": [ + 313, + 413, + 555, + 558 + ], + "type": "text", + "content": ", characterized by the half-aperture " + }, + { + "bbox": [ + 313, + 413, + 555, + 558 + ], + "type": "inline_equation", + "content": "\\omega (T_k)" + }, + { + "bbox": [ + 313, + 413, + 555, + 558 + ], + "type": "text", + "content": ". Similarly, an unsafe image " + }, + { + "bbox": [ + 313, + 413, + 555, + 558 + ], + "type": "inline_equation", + "content": "I_{k}^{\\star}" + }, + { + "bbox": [ + 313, + 413, + 555, + 558 + ], + "type": "text", + "content": " must lie within the cone defined by its corresponding unsafe text " + }, + { + "bbox": [ + 313, + 413, + 555, + 558 + ], + "type": "inline_equation", + "content": "T_{k}^{\\star}" + }, + { + "bbox": [ + 313, + 413, + 555, + 558 + ], + "type": "text", + "content": ". This is enforced through an entailment loss formulated for image-text representations by Le et al. [38] and Desai et al. [17], as" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 329, + 565, + 553, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 565, + 553, + 594 + ], + "spans": [ + { + "bbox": [ + 329, + 565, + 553, + 594 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\begin{array}{l} L _ {e n t} ^ {\\star} (I, T) = \\max \\left(0, \\phi \\left(I _ {k}, T _ {k}\\right) - \\eta \\omega \\left(T _ {k}\\right)\\right) \\text {a n d} \\\\ I _ {*} ^ {\\star} \\left(I _ {*} ^ {\\star}, T _ {*} ^ {\\star}\\right) = \\left(0 + \\left(I _ {*} ^ {\\star}, T _ {*} ^ {\\star}\\right) - \\left(T _ {*} ^ {\\star}\\right)\\right) \\end{array} \\tag {11} \\\\ L _ {e n t} ^ {\\star} \\left(I ^ {\\star}, T ^ {\\star}\\right) = \\max \\left(0, \\phi \\left(I _ {k} ^ {\\star}, T _ {k} ^ {\\star}\\right) - \\eta \\omega \\left(T _ {k} ^ {\\star}\\right)\\right), \\\\ \\end{array}", + "image_path": "73dea444aed9c127db7e68bae49491c7e6e50108932b22ee044dc23dace2948b.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 602, + 554, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 602, + 554, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 602, + 554, + 628 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 602, + 554, + 628 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 602, + 554, + 628 + ], + "type": "text", + "content": " is the exterior angle (between lines " + }, + { + "bbox": [ + 313, + 602, + 554, + 628 + ], + "type": "inline_equation", + "content": "I_{k}T_{k}" + }, + { + "bbox": [ + 313, + 602, + 554, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 602, + 554, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{0}T_{k}" + }, + { + "bbox": [ + 313, + 602, + 554, + 628 + ], + "type": "text", + "content": " or between lines " + }, + { + "bbox": [ + 313, + 602, + 554, + 628 + ], + "type": "inline_equation", + "content": "I_k^\\star T_k^\\star" + }, + { + "bbox": [ + 313, + 602, + 554, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 602, + 554, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{0}T_k^\\star" + }, + { + "bbox": [ + 313, + 602, + 554, + 628 + ], + "type": "text", + "content": ") given by" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 319, + 636, + 554, + 677 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 636, + 554, + 677 + ], + "spans": [ + { + "bbox": [ + 319, + 636, + 554, + 677 + ], + "type": "interline_equation", + "content": "\\phi \\left(I _ {k}, T _ {k}\\right) = \\cos^ {- 1} \\left(\\frac {I _ {k 0} + T _ {k 0} \\kappa \\left\\langle I _ {k} , T _ {k} \\right\\rangle_ {\\mathcal {L}}}{\\left\\| \\widetilde {T _ {k}} \\right\\| \\sqrt {\\left(\\kappa \\left\\langle I _ {k} , T _ {k} \\right\\rangle_ {\\mathcal {L}}\\right) ^ {2} - 1}}\\right). \\tag {12}", + "image_path": "279ca691616430f17715857dfc61f9e4bf5e9ed6dc892d278093ceef6a3a6a2f.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "type": "text", + "content": " is a threshold for the half-aperture, " + }, + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\omega (T_{k})" + }, + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "type": "text", + "content": " [52]. Intuitively, the entailment loss " + }, + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "type": "inline_equation", + "content": "L_{ent}^{\\star}" + }, + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "type": "text", + "content": " penalizes images " + }, + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "type": "inline_equation", + "content": "I_{k}" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4225" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": "that lie outside the cone " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\mathfrak{S}_{T_k}" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " defined by their corresponding text caption " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "T_{k}" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": ". Finally, to model the safety hierarchy, we enforce that safe concepts entail unsafe ones, reflecting that safe data are more general and unsafe data are more specific. Specifically, we enforce that a safe image " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "I_{k}" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " entails an unsafe text " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "T_{k}^{\\star}" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": ", meaning that the unsafe text " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "T_{k}^{\\star}" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " must lie within the cone defined by the safe image " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "I_{k}" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": ", characterized by the half-aperture " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\omega (I_k)" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": ". This gives us the safety-entailment defined in Eq. 6. This is implemented as" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 76, + 186, + 295, + 199 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 186, + 295, + 199 + ], + "spans": [ + { + "bbox": [ + 76, + 186, + 295, + 199 + ], + "type": "interline_equation", + "content": "L _ {e n t} ^ {\\star} \\left(T ^ {\\star}, I\\right) = \\max \\left(0, \\phi \\left(T _ {k} ^ {\\star}, I _ {k}\\right) - \\eta \\omega \\left(I _ {k}\\right)\\right). \\tag {13}", + "image_path": "a185b29b28075d47ed46fc246b90073e90f1041086e86a60edb3322e07b98c68.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 204, + 285, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 204, + 285, + 216 + ], + "spans": [ + { + "bbox": [ + 57, + 204, + 285, + 216 + ], + "type": "text", + "content": "The overall entailment loss to satisfy Eq. 7 is defined as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 221, + 295, + 245 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 221, + 295, + 245 + ], + "spans": [ + { + "bbox": [ + 56, + 221, + 295, + 245 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {h S E}} (I, T, I ^ {\\star}, T ^ {\\star}) = L _ {e n t} ^ {\\star} (I, T) + L _ {e n t} ^ {\\star} \\left(T ^ {\\star}, I\\right) + L _ {e n t} ^ {\\star} \\left(I ^ {\\star}, T ^ {\\star}\\right). \\tag {14}", + "image_path": "bf1b4a160dfa5865b23ebaeeb79c3b0b5439d7fbcf02bd91d0e5f27a428f5845.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 259, + 296, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 259, + 296, + 294 + ], + "spans": [ + { + "bbox": [ + 55, + 259, + 296, + 294 + ], + "type": "text", + "content": "Combined loss function. We integrate the contrastive with the entailment losses to obtain the total loss used to fine-tune the model on the dataset " + }, + { + "bbox": [ + 55, + 259, + 296, + 294 + ], + "type": "inline_equation", + "content": "D^{\\star}" + }, + { + "bbox": [ + 55, + 259, + 296, + 294 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 300, + 296, + 324 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 300, + 296, + 324 + ], + "spans": [ + { + "bbox": [ + 56, + 300, + 296, + 324 + ], + "type": "interline_equation", + "content": "L (I, T, I ^ {\\star}, T ^ {\\star}) = L _ {\\mathrm {h S C}} (I, T, I ^ {\\star}, T ^ {\\star}) + L _ {\\mathrm {h S E}} (I, T, I ^ {\\star}, T ^ {\\star}). \\tag {15}", + "image_path": "7cb821a727a63394a8c36914a5ef0785221898f0f64b444d829a0692f946dff4.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 325, + 296, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 325, + 296, + 432 + ], + "spans": [ + { + "bbox": [ + 55, + 325, + 296, + 432 + ], + "type": "text", + "content": "Our proposal allows the model to differentiate between safe and unsafe embeddings based on their distance from the origin. Safe content is closer to the center, while NSFW content is farther away. This geometric arrangement not only enables the model to detect unsafe content but also allows dynamic manipulation of embeddings. NSFW queries can be redirected toward the safe region, effectively retrieving outputs that prioritize safety and providing more precise control over content retrieval." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 438, + 228, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 438, + 228, + 451 + ], + "spans": [ + { + "bbox": [ + 55, + 438, + 228, + 451 + ], + "type": "text", + "content": "4.3. Safety traversals and evaluation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 456, + 296, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 296, + 601 + ], + "type": "text", + "content": "The hyperbolic safety-aware training results in a restructuring of the shared embedding space of the vision-language model. To obtain safe but relevant retrieval outputs from unsafe queries or vice-versa, we introduce a traversal mechanism to adjust query embeddings in hyperbolic space, enhancing their similarity with either safe or unsafe content, depending on the retrieval task. This traversal involves moving the query embeddings along the line connecting them to the origin of the hyperboloid, altering their hyperbolic distance from the root. By adjusting the embeddings' positions, we align them with regions in the embedding space that correspond to the desired content type." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 601, + 296, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 601, + 296, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 601, + 296, + 685 + ], + "type": "text", + "content": "Traversal Definition. Given an embedding " + }, + { + "bbox": [ + 55, + 601, + 296, + 685 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 55, + 601, + 296, + 685 + ], + "type": "text", + "content": ", our method computes the distance from a predefined root feature " + }, + { + "bbox": [ + 55, + 601, + 296, + 685 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 55, + 601, + 296, + 685 + ], + "type": "text", + "content": " in hyperbolic space using the Lorentzian distance function " + }, + { + "bbox": [ + 55, + 601, + 296, + 685 + ], + "type": "inline_equation", + "content": "d_{\\mathcal{L}}(\\mathbf{q},\\mathbf{r})" + }, + { + "bbox": [ + 55, + 601, + 296, + 685 + ], + "type": "text", + "content": ". For each type of content " + }, + { + "bbox": [ + 55, + 601, + 296, + 685 + ], + "type": "inline_equation", + "content": "X\\in \\{T,I,T^{\\star},I^{\\star}\\}" + }, + { + "bbox": [ + 55, + 601, + 296, + 685 + ], + "type": "text", + "content": ", we compute the mean distance " + }, + { + "bbox": [ + 55, + 601, + 296, + 685 + ], + "type": "inline_equation", + "content": "\\mu_X" + }, + { + "bbox": [ + 55, + 601, + 296, + 685 + ], + "type": "text", + "content": " from the root feature " + }, + { + "bbox": [ + 55, + 601, + 296, + 685 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 55, + 601, + 296, + 685 + ], + "type": "text", + "content": " based on the distribution of each category. The boundary for each type is then defined as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 102, + 690, + 295, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 690, + 295, + 717 + ], + "spans": [ + { + "bbox": [ + 102, + 690, + 295, + 717 + ], + "type": "interline_equation", + "content": "\\tau_ {X} = \\mu_ {X} + \\tanh \\left(\\frac {\\mu_ {X} - \\alpha}{\\kappa}\\right) + 1, \\tag {16}", + "image_path": "a6f7a3f211b6447680d5aab9ea48e5ef4c88ceb0f57732d7c944ddb4e8c5546f.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": " is the negative curvature, and " + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": " is a constant set empirically to 0.8. This shift accounts for the curvature of space, ensuring the boundaries are appropriately adjusted for effective traversal. Defining four bounds allows more nuanced control over traversal depending on the retrieval task. To retrieve a content type " + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": ", the query is moved along the Euclidean direction vector " + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\mathrm{dir}} = \\mathbf{q} - \\mathbf{r}" + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": " toward the root feature " + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": " until it reaches the corresponding boundary " + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "inline_equation", + "content": "\\tau_{X}" + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": " (e.g. " + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "inline_equation", + "content": "\\tau_{T}" + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": " for safe text). The target position " + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "inline_equation", + "content": "\\mathbf{q}^{*}" + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": " is given as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 389, + 187, + 555, + 212 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 187, + 555, + 212 + ], + "spans": [ + { + "bbox": [ + 389, + 187, + 555, + 212 + ], + "type": "interline_equation", + "content": "\\mathbf {q} ^ {*} = \\mathbf {r} + \\tau_ {X} \\cdot \\frac {\\mathbf {v} _ {\\mathrm {d i r}}}{\\| \\mathbf {v} _ {\\mathrm {d i r}} \\|} \\tag {17}", + "image_path": "b7d5a8581305e10509684cde3a5ddf147161f8e38441e9ada7fff8aefa4991b6.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 218, + 555, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 218, + 555, + 243 + ], + "spans": [ + { + "bbox": [ + 314, + 218, + 555, + 243 + ], + "type": "text", + "content": "allowing the embeddings to be repositioned to match the target content type while maintaining semantic alignment." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 251, + 395, + 265 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 251, + 395, + 265 + ], + "spans": [ + { + "bbox": [ + 314, + 251, + 395, + 265 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 271, + 412, + 284 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 271, + 412, + 284 + ], + "spans": [ + { + "bbox": [ + 314, + 271, + 412, + 284 + ], + "type": "text", + "content": "5.1. Training Details" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 289, + 554, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 554, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 554, + 348 + ], + "type": "text", + "content": "Datasets. Our experiments are mainly conducted on the ViSU dataset [55], containing 165k quadruplets of safe and unsafe image-text pairs. We also evaluate our model on three real-world NSFW image datasets: NudeNet [4], NSFW data source URLs" + }, + { + "bbox": [ + 313, + 289, + 554, + 348 + ], + "type": "inline_equation", + "content": "^1" + }, + { + "bbox": [ + 313, + 289, + 554, + 348 + ], + "type": "text", + "content": " and SMID [15]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 349, + 554, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 349, + 554, + 432 + ], + "spans": [ + { + "bbox": [ + 313, + 349, + 554, + 432 + ], + "type": "text", + "content": "Baselines. Our safety comparisons include the original CLIP [58] and the state-of-the-art Safe-CLIP [55]. CLIP was trained using a private dataset of 400M image-text pairs [66] which has unsafe data [5]. Safe-CLIP is finetuned on the ViSU dataset [55] to redirect unsafe content to safe correspondent one via contrastive losses and cosine similarities, aiming to unlearn NSFW concepts." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 433, + 554, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 433, + 554, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 433, + 554, + 491 + ], + "type": "text", + "content": "Models. Our visual and textual encoders are the same as CLIP [58], with VIT-L/14 as visual encoder, to maintain fair comparison to Safe-CLIP [55]. During training, both the visual and textual encoder are fine-tuned using low-rank decomposition [32] with low-rank factor " + }, + { + "bbox": [ + 313, + 433, + 554, + 491 + ], + "type": "inline_equation", + "content": "r = 16" + }, + { + "bbox": [ + 313, + 433, + 554, + 491 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 492, + 554, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 492, + 554, + 576 + ], + "spans": [ + { + "bbox": [ + 313, + 492, + 554, + 576 + ], + "type": "text", + "content": "**Optimization.** We use AdamW [43] with weight decay 0.2 and " + }, + { + "bbox": [ + 313, + 492, + 554, + 576 + ], + "type": "inline_equation", + "content": "(\\beta_{1},\\beta_{2}) = (0.9,0.98)" + }, + { + "bbox": [ + 313, + 492, + 554, + 576 + ], + "type": "text", + "content": ". We disable weight decay for all gains, biases, and learnable scalars. The model is finetuned for 20 epochs with batch size 256. The maximum learning rate is " + }, + { + "bbox": [ + 313, + 492, + 554, + 576 + ], + "type": "inline_equation", + "content": "8\\times 10^{-4}" + }, + { + "bbox": [ + 313, + 492, + 554, + 576 + ], + "type": "text", + "content": ". We use mixed precision [47] to accelerate training, except computing exponential map and losses for HySAC in FP32 precision for numerical stability." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 576, + 554, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 576, + 554, + 672 + ], + "spans": [ + { + "bbox": [ + 313, + 576, + 554, + 672 + ], + "type": "text", + "content": "Initialization. We initialize image and text encoders akin to CLIP, along with pre-trained weights. We initialize the softmax temperature as " + }, + { + "bbox": [ + 313, + 576, + 554, + 672 + ], + "type": "inline_equation", + "content": "\\tau = 0.07" + }, + { + "bbox": [ + 313, + 576, + 554, + 672 + ], + "type": "text", + "content": " and clamp it to a minimum value of 0.01. For HySAC, we initialize the learnable projection scalars " + }, + { + "bbox": [ + 313, + 576, + 554, + 672 + ], + "type": "inline_equation", + "content": "\\alpha_{img} = \\alpha_{txt} = 1 / \\sqrt{512}" + }, + { + "bbox": [ + 313, + 576, + 554, + 672 + ], + "type": "text", + "content": ", the curvature parameter " + }, + { + "bbox": [ + 313, + 576, + 554, + 672 + ], + "type": "inline_equation", + "content": "c = 1.0" + }, + { + "bbox": [ + 313, + 576, + 554, + 672 + ], + "type": "text", + "content": " and clamp it in [0.1, 10.0] to prevent training instability. All scalars are learned in logarithmic space as " + }, + { + "bbox": [ + 313, + 576, + 554, + 672 + ], + "type": "inline_equation", + "content": "\\log(1 / \\tau)" + }, + { + "bbox": [ + 313, + 576, + 554, + 672 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 576, + 554, + 672 + ], + "type": "inline_equation", + "content": "\\log(c)" + }, + { + "bbox": [ + 313, + 576, + 554, + 672 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 576, + 554, + 672 + ], + "type": "inline_equation", + "content": "\\log(\\alpha_{img})" + }, + { + "bbox": [ + 313, + 576, + 554, + 672 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 576, + 554, + 672 + ], + "type": "inline_equation", + "content": "\\log(\\alpha_{txt})" + }, + { + "bbox": [ + 313, + 576, + 554, + 672 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 672, + 554, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 554, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 554, + 696 + ], + "type": "text", + "content": "Further details on the training setup are provided in the supplementary B." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 325, + 703, + 496, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 703, + 496, + 712 + ], + "spans": [ + { + "bbox": [ + 325, + 703, + 496, + 712 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 325, + 703, + 496, + 712 + ], + "type": "text", + "content": "https://github.com/EBazarov/nsfw_data_source_url" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4226" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 553, + 186 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 186 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 186 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 186 + ], + "type": "table", + "html": "
ModelText-to-Image (T-to-I)Image-to-Text (I-to-T)Text-to-Image (T*-to-I ∪ I*)Image-to-Text (I*-to-T ∪ T*)
R@1R@10R@20R@1R@10R@20R@1R@10R@20R@1R@10R@20
CLIP [58]36.871.681.539.874.283.52.024.833.24.632.940.6
MERU [17]14.943.054.214.742.353.82.215.221.54.422.629.4
HyCoCLIP [52]34.371.280.634.471.382.22.825.333.28.237.845.7
Safe-CLIP [55]45.981.889.745.382.389.88.046.958.019.162.971.1
MERU*50.084.191.151.285.392.32.339.949.45.747.954.7
HyCoCLIP*47.781.989.146.782.790.41.532.742.36.945.253.6
HySAC49.884.190.748.284.291.230.562.871.842.173.379.8
", + "image_path": "0ea2b4b3cccf232c6405186e13caef8b6aa2a418e9de8123d8e2fb8b2047e285.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 230, + 553, + 319 + ], + "blocks": [ + { + "bbox": [ + 55, + 188, + 553, + 222 + ], + "lines": [ + { + "bbox": [ + 55, + 188, + 553, + 222 + ], + "spans": [ + { + "bbox": [ + 55, + 188, + 553, + 222 + ], + "type": "text", + "content": "Table 1. Safe content retrieval performance on ViSU test set. Across all tasks and recall rates, HySAC improves over existing safety unlearning CLIP and hyperbolic CLIP models, highlighting that our approach is able to navigate unsafe image or text inputs towards relevant but safe retrieval outputs. * CLIP fine-tuned in hyperbolic space on ViSU training set with MERU/HyCoCLIP losses." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 230, + 553, + 319 + ], + "lines": [ + { + "bbox": [ + 57, + 230, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 57, + 230, + 553, + 319 + ], + "type": "table", + "html": "
ModelText-to-Image (T* -to-I*)Image-to-Text (I* -to-T*)Text-to-Image (T* -to-I* ∪ I)Image-to-Text (I* -to-T* ∪ T)
R@1R@10R@20R@1R@10R@20R@1R@10R@20R@1R@10R@20
CLIP [58]73.194.997.672.895.297.768.492.395.967.193.396.7
MERU [17]29.462.472.225.857.767.823.554.064.319.551.161.2
HyCoCLIP [52]69.593.195.865.091.195.063.789.793.755.288.092.7
Safe-CLIP [55]58.086.291.456.085.191.047.780.085.832.177.184.6
HySAC81.498.499.482.297.899.281.198.499.480.597.298.9
", + "image_path": "7d82a5cf2f84427091b543ff3e88b6249b721691d2dd62444470e1c37bddbe16.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 321, + 553, + 344 + ], + "lines": [ + { + "bbox": [ + 55, + 321, + 553, + 344 + ], + "spans": [ + { + "bbox": [ + 55, + 321, + 553, + 344 + ], + "type": "text", + "content": "Table 2. Unsafe content retrieval performance on ViSU test set. Akin to safe content retrieval, our approach performs best. This is a result of our objective, as we assign different content to different regions, enabling us to maintain valuable safety information." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 356, + 180, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 356, + 180, + 369 + ], + "spans": [ + { + "bbox": [ + 55, + 356, + 180, + 369 + ], + "type": "text", + "content": "5.2. Experimental Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 373, + 295, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 295, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 295, + 434 + ], + "type": "text", + "content": "To assess the performance of our proposed model, HySAC, we measure its safety awareness and its ability to handle unsafe content effectively, while retaining both safe and unsafe knowledge. In the supplementary D, we report the zero-shot generalization of our method." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 440, + 201, + 452 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 440, + 201, + 452 + ], + "spans": [ + { + "bbox": [ + 55, + 440, + 201, + 452 + ], + "type": "text", + "content": "5.2.1. Safety retrieval comparison" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 455, + 295, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 455, + 295, + 574 + ], + "spans": [ + { + "bbox": [ + 54, + 455, + 295, + 574 + ], + "type": "text", + "content": "We evaluate our model on the capability of retrieving safe and unsafe items, in comparison to CLIP [58] and Safe-CLIP [55]. We also provide a comparison to recent hyperbolic VLMs, namely MERU [17] and HyCoCLIP [52]. Safe-CLIP and HySAC fine-tuned from CLIP on ViSU [55], MERU trained on RedCaps [16], and HyCoCLIP on GRIT [54]. For a fair comparison, we additionally fine-tuned the CLIP model on the ViSU dataset in hyperbolic space, using MERU and HyCoCLIP losses. All the models are evaluated on the ViSU test set." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 574, + 295, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 574, + 295, + 694 + ], + "spans": [ + { + "bbox": [ + 54, + 574, + 295, + 694 + ], + "type": "text", + "content": "Retrieval tasks are defined as text-to-image and image-to-text, where the goal is to find the most relevant counterpart for a given query. Recall@K measures the fraction of queries where the correct item appears in the top-K retrieved results. To assess the safe retrieval performance of HySAC, we measure recall exclusively on safe content in both visual and textual elements (" + }, + { + "bbox": [ + 54, + 574, + 295, + 694 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 54, + 574, + 295, + 694 + ], + "type": "text", + "content": "-to-" + }, + { + "bbox": [ + 54, + 574, + 295, + 694 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 54, + 574, + 295, + 694 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 574, + 295, + 694 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 54, + 574, + 295, + 694 + ], + "type": "text", + "content": "-to-" + }, + { + "bbox": [ + 54, + 574, + 295, + 694 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 54, + 574, + 295, + 694 + ], + "type": "text", + "content": "). This step is crucial to verify that the original CLIP model's retrieval capabilities are retained after finetuning in hyperbolic space with our training method. Then, to evaluate" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 356, + 553, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 356, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 356, + 553, + 441 + ], + "type": "text", + "content": "the safety-awareness capabilities of HySAC, we introduce a distinct setup in which NSFW elements are used as queries, while the retrievable items include both safe and unsafe elements (" + }, + { + "bbox": [ + 313, + 356, + 553, + 441 + ], + "type": "inline_equation", + "content": "T^{*}" + }, + { + "bbox": [ + 313, + 356, + 553, + 441 + ], + "type": "text", + "content": "-to-" + }, + { + "bbox": [ + 313, + 356, + 553, + 441 + ], + "type": "inline_equation", + "content": "I \\cup I^{*}" + }, + { + "bbox": [ + 313, + 356, + 553, + 441 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 356, + 553, + 441 + ], + "type": "inline_equation", + "content": "I^{*}" + }, + { + "bbox": [ + 313, + 356, + 553, + 441 + ], + "type": "text", + "content": "-to-" + }, + { + "bbox": [ + 313, + 356, + 553, + 441 + ], + "type": "inline_equation", + "content": "T \\cup T^{*}" + }, + { + "bbox": [ + 313, + 356, + 553, + 441 + ], + "type": "text", + "content": "). During these experiments, a retrieval is deemed correct only if the query retrieves its safe counterpart, thereby validating the model's ability to redirect unsafe queries towards safe items." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 445, + 554, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 554, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 554, + 578 + ], + "type": "text", + "content": "When retrieving with HySAC, the threshold " + }, + { + "bbox": [ + 313, + 445, + 554, + 578 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 445, + 554, + 578 + ], + "type": "text", + "content": " for moving query embeddings is computed using the mean distance of safe embeddings from the origin, adjusting the query embedding towards the safe region. Results are reported in Table 1, where we observe that HySAC consistently improves over both unlearning and existing hyperbolic models and features the highest recalls across all settings and rates. CLIP hyperbolic models finetuned on ViSU data (MERU\\* and HyCoCLIP\\*) perform well on safe-only retrieval, while our method achieves high performance on both safe-only and unsafe-safe retrieval, due to our safety-aware design." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "In Table 2, we instead perform analyses for unsafe content retrieval. First, this involves the text-to-image and image-to-text retrieval on only unsafe elements (" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "T^{*}" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "-to-" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "I^{*}" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "I^{*}" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "-to-" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "T^{*}" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "). Second, instead, we perform retrieval by using unsafe elements as queries and both safe and unsafe items as retrievable items (" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "T^{*}" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "-to-" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "I^{*} \\cup I" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "I^{*}" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "-to-" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "T^{*} \\cup T" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "), and deem the retrieval correct only if the query retrieves its corresponding unsafe one. This setup tests the model's ability to function as a content moderator and also showcases its capacity to provide user autonomy in content retrieval decisions. In these tests, the traversal mechanism in" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 66, + 702, + 289, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 289, + 713 + ], + "type": "text", + "content": "2The box data needed for HyCoCLIP was extracted using Kosmos-2." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4227" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 294, + 247 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 294, + 247 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 294, + 247 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 294, + 247 + ], + "type": "image", + "image_path": "648e8418109949de6e574dd295c73a730d1452a7c47aae6150a045333fd262b8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 252, + 295, + 307 + ], + "lines": [ + { + "bbox": [ + 55, + 252, + 295, + 307 + ], + "spans": [ + { + "bbox": [ + 55, + 252, + 295, + 307 + ], + "type": "text", + "content": "Figure 2. Distributions of embedding distances from the root. We embed all ViSU training samples and visualize their distance distribution from the root. While CLIP and Safe-CLIP do not separate between texts and images, MERU does. HySAC, instead, also differentiates between safe and unsafe content." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 313, + 294, + 370 + ], + "blocks": [ + { + "bbox": [ + 58, + 313, + 294, + 370 + ], + "lines": [ + { + "bbox": [ + 58, + 313, + 294, + 370 + ], + "spans": [ + { + "bbox": [ + 58, + 313, + 294, + 370 + ], + "type": "table", + "html": "
Model(T-to-I)(I-to-T)(T*-to-I ∪ I*)(I*-to-T ∪ T*)
R@1R@10R@1R@10R@1R@10R@1R@10
w/o Ent52.384.950.884.74.149.05.564.5
w/o S-Ent51.084.249.884.31.439.17.463.7
HySAC49.884.148.284.230.562.842.173.3
", + "image_path": "53401a90ee3370d824641525bb56a4fb9afe35a6b12c6fd34d9d04f97c0f75f4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 415, + 296, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 296, + 499 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 296, + 499 + ], + "type": "text", + "content": "HySAC uses adjusted parameters to move in the unsafe direction, targeting the retrieval of NSFW content. Here too, HySAC achieves the best recall across all settings, demonstrating that HySAC not only prioritizes safety by navigating away from NSFW content when required but also ensures that users can access NSFW content under controlled conditions, better than existing competitors." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 506, + 167, + 518 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 506, + 167, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 167, + 518 + ], + "type": "text", + "content": "5.2.2. Analysis of HySAC" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 521, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 617 + ], + "type": "text", + "content": "Assessing HySAC embedding space. Further, we validate the organization of the embedding space as outlined in Equation 7. Our goal is to confirm that the embeddings for safe content are positioned closer to the origin of hyperbolic space, while those for unsafe content are further away, following the proposed hierarchy. Specifically, for each " + }, + { + "bbox": [ + 55, + 521, + 295, + 617 + ], + "type": "inline_equation", + "content": "X \\in \\{T, I, T^{\\star}, I^{\\star}\\}" + }, + { + "bbox": [ + 55, + 521, + 295, + 617 + ], + "type": "text", + "content": " of the training-set of " + }, + { + "bbox": [ + 55, + 521, + 295, + 617 + ], + "type": "inline_equation", + "content": "D^{\\star}" + }, + { + "bbox": [ + 55, + 521, + 295, + 617 + ], + "type": "text", + "content": ", we compute the distances " + }, + { + "bbox": [ + 55, + 521, + 295, + 617 + ], + "type": "inline_equation", + "content": "d_{\\mathcal{L}}(X, \\mathbf{r})" + }, + { + "bbox": [ + 55, + 521, + 295, + 617 + ], + "type": "text", + "content": " from the root feature " + }, + { + "bbox": [ + 55, + 521, + 295, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 55, + 521, + 295, + 617 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": "A visualization is reported in Figure 2, where we show the distribution of embeddings in terms of their distance to the root feature. The comparative analysis is done across four different models: HySAC, CLIP, Safe-CLIP, and MERU. For both the hyperbolic models the root feature is the origin of the hyperboloid. For the Euclidean models, since the origin does not lie on the hypersphere, the root is empirically estimated as the embedding that has the least" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 317, + 70, + 553, + 128 + ], + "blocks": [ + { + "bbox": [ + 55, + 374, + 295, + 407 + ], + "lines": [ + { + "bbox": [ + 55, + 374, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 374, + 295, + 407 + ], + "type": "text", + "content": "Table 3. Ablation study on loss components. We evaluate HySAC against two ablations that remove loss components. Results are in the same setting of Table 1." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 70, + 553, + 128 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 553, + 128 + ], + "type": "table", + "html": "
Model% Safe (Text-to-Image)% Safe (Image-to-Text)
NudeNetNSFW URLsSMIDNudeNetNSFW URLsSMID
CLIP78.279.755.233.344.059.1
Safe-CLIP92.692.683.475.276.465.6
HySAC96.293.980.184.495.197.9
", + "image_path": "4ed46ecf25cbf454115a8040f6e2ded4888f5d7f9c7e7859dceacdcb01e41524.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 133, + 555, + 176 + ], + "lines": [ + { + "bbox": [ + 313, + 133, + 555, + 176 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 555, + 176 + ], + "type": "text", + "content": "Table 4. Retrieval performance on real NSFW images. Rate of safe images retrieved using unsafe prompts from the ViSU test set. The retrievable set includes safe and unsafe real images, with the latter from LAION-400M and the former from NSFW sources." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 185, + 553, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 185, + 553, + 269 + ], + "spans": [ + { + "bbox": [ + 313, + 185, + 553, + 269 + ], + "type": "text", + "content": "distance from all embeddings of the training set, i.e. the " + }, + { + "bbox": [ + 313, + 185, + 553, + 269 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 313, + 185, + 553, + 269 + ], + "type": "text", + "content": " normalization of the average of all embeddings. As it can be seen, the distribution clearly shows four peak distributions for HySAC, each one representing one of the " + }, + { + "bbox": [ + 313, + 185, + 553, + 269 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 185, + 553, + 269 + ], + "type": "text", + "content": " content types, elucidating the efficacy of our approach in maintaining a clear separation between safe, unsafe, textual, and visual content within the embedding space." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 269, + 555, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 269, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 269, + 555, + 437 + ], + "type": "text", + "content": "Ablation Study. In Table 3, we validate the effectiveness of the key components in HySAC, by comparing its full configuration with variants where specific losses are disabled. In particular, we employ one variant which only keeps contrastive losses (denoted as \"w/o Ent\") and one that omits the safety-entailment loss (\"w/o S-Ent\"). Results show that while removing these components slightly improves performance in scenarios involving only safe content, likely due to reduced spatial constraints, their absence significantly undermines the model's effectiveness in dealing with unsafe content, especially in unsafe-to-safe retrieval. These results underscore the essential roles that both the modality- and safety-entailment losses play in enhancing the safety awareness of the proposed model." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 437, + 554, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 554, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 554, + 568 + ], + "type": "text", + "content": "Retrieval on real NSFW datasets. To further analyze the safety of HySAC, we conduct retrieval tests using NudeNet [4], NSFW data source URLs and SMID [15]. The first two datasets primarily contain nudity and pornographic content, whereas SMID includes a broader range of inappropriate content, such as violence, harm, and discrimination. We randomly select 1000 images from each dataset to serve as visual elements and use 5k NSFW captions from the ViSU test set. Both the image-to-text and text-to-image retrieval tasks also incorporate 10,000 randomly chosen retrievable safe items from LAION-400M [66]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "text", + "content": "Results, displayed in Table 4, contrast the performance of HySAC with that of CLIP [58] and Safe-CLIP [55]. We report the proportion of safe retrieval outputs when NSFW queries are used, highlighting the capability of HySAC to enhance safety in retrieval results. Traversing the embedding space towards safety prioritizes safer alternatives, which may adjust relevance in some cases. Due to the absence of datasets with real unsafe data and correlated safe alternatives, this aspect could not be evaluated. Notably, HySAC demonstrates improved performance in securing safer content compared to Safe-CLIP across most datasets for both text-to-image and image-to-text scenarios." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4228" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 93, + 78, + 157, + 144 + ], + "blocks": [ + { + "bbox": [ + 93, + 78, + 157, + 144 + ], + "lines": [ + { + "bbox": [ + 93, + 78, + 157, + 144 + ], + "spans": [ + { + "bbox": [ + 93, + 78, + 157, + 144 + ], + "type": "image", + "image_path": "137086e44026790740324abec669e2161d42a69cb8feb260ddbe55b70414e761.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 66, + 148, + 181, + 181 + ], + "lines": [ + { + "bbox": [ + 66, + 148, + 181, + 181 + ], + "spans": [ + { + "bbox": [ + 66, + 148, + 181, + 181 + ], + "type": "text", + "content": "A train filled with toxic chemicals crashing into a bridge, causing a massive explosion and destroying everything in its path." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 244, + 554, + 268 + ], + "lines": [ + { + "bbox": [ + 55, + 244, + 554, + 268 + ], + "spans": [ + { + "bbox": [ + 55, + 244, + 554, + 268 + ], + "type": "text", + "content": "Figure 3. Qualitative traversal results. HySAC traverses towards the root feature, retrieving the top-1 text at each interpolation point. This traversal effectively transitions from unsafe to safe captions, demonstrating the model's ability to ensure safety-aware content retrieval." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 212, + 78, + 277, + 142 + ], + "blocks": [ + { + "bbox": [ + 93, + 190, + 157, + 210 + ], + "lines": [ + { + "bbox": [ + 93, + 190, + 157, + 210 + ], + "spans": [ + { + "bbox": [ + 93, + 190, + 157, + 210 + ], + "type": "text", + "content": "The train is going over the bridge over a river." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 212, + 78, + 277, + 142 + ], + "lines": [ + { + "bbox": [ + 212, + 78, + 277, + 142 + ], + "spans": [ + { + "bbox": [ + 212, + 78, + 277, + 142 + ], + "type": "image", + "image_path": "915170f5b969afbdb3f9a7a8a4a6055527505e925b4ff33f0a9a73ba9b56ee67.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 191, + 154, + 300, + 175 + ], + "lines": [ + { + "bbox": [ + 191, + 154, + 300, + 175 + ], + "spans": [ + { + "bbox": [ + 191, + 154, + 300, + 175 + ], + "type": "text", + "content": "Several people on skis on a snowy slope, enjoying a drug-fueled party." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 212, + 190, + 279, + 210 + ], + "lines": [ + { + "bbox": [ + 212, + 190, + 279, + 210 + ], + "spans": [ + { + "bbox": [ + 212, + 190, + 279, + 210 + ], + "type": "text", + "content": "Several people in their ski gear are in the snow." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 233, + 218, + 256, + 227 + ], + "lines": [ + { + "bbox": [ + 233, + 218, + 256, + 227 + ], + "spans": [ + { + "bbox": [ + 233, + 218, + 256, + 227 + ], + "type": "text", + "content": "[ROOT]" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 334, + 78, + 399, + 143 + ], + "blocks": [ + { + "bbox": [ + 113, + 218, + 135, + 227 + ], + "lines": [ + { + "bbox": [ + 113, + 218, + 135, + 227 + ], + "spans": [ + { + "bbox": [ + 113, + 218, + 135, + 227 + ], + "type": "text", + "content": "[ROOT]" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 334, + 78, + 399, + 143 + ], + "lines": [ + { + "bbox": [ + 334, + 78, + 399, + 143 + ], + "spans": [ + { + "bbox": [ + 334, + 78, + 399, + 143 + ], + "type": "image", + "image_path": "b50845a781859b6ce8160e0ae58ff6b6ec0468e2a8e7a11038a16d70115d12aa.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 312, + 148, + 423, + 180 + ], + "lines": [ + { + "bbox": [ + 312, + 148, + 423, + 180 + ], + "spans": [ + { + "bbox": [ + 312, + 148, + 423, + 180 + ], + "type": "text", + "content": "A random Black man is standing in front of a double door with a gun in his hand and a look of hatred in his eyes." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 323, + 190, + 412, + 210 + ], + "lines": [ + { + "bbox": [ + 323, + 190, + 412, + 210 + ], + "spans": [ + { + "bbox": [ + 323, + 190, + 412, + 210 + ], + "type": "text", + "content": "A person is holding an umbrella while standing in the garage." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 356, + 218, + 378, + 227 + ], + "lines": [ + { + "bbox": [ + 356, + 218, + 378, + 227 + ], + "spans": [ + { + "bbox": [ + 356, + 218, + 378, + 227 + ], + "type": "text", + "content": "[ROOT]" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 458, + 78, + 522, + 143 + ], + "blocks": [ + { + "bbox": [ + 458, + 78, + 522, + 143 + ], + "lines": [ + { + "bbox": [ + 458, + 78, + 522, + 143 + ], + "spans": [ + { + "bbox": [ + 458, + 78, + 522, + 143 + ], + "type": "image", + "image_path": "5d60419f6c4e62fcf50255a2214760fac3ec688e3827055b5e251bd2f4515c55.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 438, + 148, + 544, + 167 + ], + "lines": [ + { + "bbox": [ + 438, + 148, + 544, + 167 + ], + "spans": [ + { + "bbox": [ + 438, + 148, + 544, + 167 + ], + "type": "text", + "content": "Two naked women sitting next to each other while flashing their genitals." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 470, + 174, + 510, + 182 + ], + "lines": [ + { + "bbox": [ + 470, + 174, + 510, + 182 + ], + "spans": [ + { + "bbox": [ + 470, + 174, + 510, + 182 + ], + "type": "text", + "content": "Sultry women." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 443, + 190, + 539, + 210 + ], + "lines": [ + { + "bbox": [ + 443, + 190, + 539, + 210 + ], + "spans": [ + { + "bbox": [ + 443, + 190, + 539, + 210 + ], + "type": "text", + "content": "A couple of women sitting next to each other in front of a cake." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 478, + 218, + 500, + 227 + ], + "lines": [ + { + "bbox": [ + 478, + 218, + 500, + 227 + ], + "spans": [ + { + "bbox": [ + 478, + 218, + 500, + 227 + ], + "type": "text", + "content": "[ROOT]" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 58, + 282, + 294, + 367 + ], + "blocks": [ + { + "bbox": [ + 58, + 282, + 294, + 367 + ], + "lines": [ + { + "bbox": [ + 58, + 282, + 294, + 367 + ], + "spans": [ + { + "bbox": [ + 58, + 282, + 294, + 367 + ], + "type": "table", + "html": "
ModelNudeNetMixed NSFW
AccFPRFNRAccFPRFNR
NSFW-CNN [37]85.30.014.766.54.535.9
CLIP-classifier [65]97.30.02.776.90.111.0
CLIP-distance [59]86.40.013.677.82.022.1
NudeNet [4]91.20.08.876.94.524.6
Q16 [63]28.50.071.565.38.329.4
HySAC99.50.00.578.916.56.8
", + "image_path": "acfea7d169b2a6f6d04df6e1c1ecbe3307d5bc39640eab9f56498c7b04ad1462.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 374, + 295, + 396 + ], + "lines": [ + { + "bbox": [ + 55, + 374, + 295, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 374, + 295, + 396 + ], + "type": "text", + "content": "Table 5. NSFW classification. Comparison between HySAC and other NSFW classifiers. Metrics reported in percentages." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 403, + 295, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 403, + 295, + 547 + ], + "spans": [ + { + "bbox": [ + 54, + 403, + 295, + 547 + ], + "type": "text", + "content": "Classifying NSFW content. The structure of the embedding space in HySAC also supports the classification of NSFW content. We evaluate this using the NudeNet [4] and Mixed NSFW datasets, comparing against classifiers such as NSFW-CNN [37], CLIP-classifier [65], CLIP-distance [59], NudeNet [4], and Q16 [63]. NudeNet only contains nudity, while Mixed NSFW includes different NSFW categories from various online sources and safe images from PASS [1]. We sample a 1,000-image subset from NudeNet and 442 images from Mixed NSFW, balanced between safe and unsafe. Further details of these datasets are provided in the supplementary C." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 548, + 295, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 548, + 295, + 607 + ], + "spans": [ + { + "bbox": [ + 55, + 548, + 295, + 607 + ], + "type": "text", + "content": "Results in Table 5 show that HySAC achieves competitive or superior results in NSFW content classification, despite not being explicitly designed for safety classification. The norm threshold, set to the ViSU dataset mean (Figure 2), differentiates safe from unsafe content." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 55, + 614, + 218, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 614, + 218, + 625 + ], + "spans": [ + { + "bbox": [ + 55, + 614, + 218, + 625 + ], + "type": "text", + "content": "5.2.3. Visualizing the safety traversals" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 55, + 629, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 677 + ], + "type": "text", + "content": "We examine traversal paths for safe text retrievals, starting with an unsafe image embedding as the query. The traversals are along the geodesic of hyperbolic space from the image to the origin of the hyperboloid, denoted [ROOT]." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "The input query is an unsafe image taken from ViSU test set and the text retrieval space consists of a mix of safe and unsafe captions of ViSU test set, metadata-based caption" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 285, + 555, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 555, + 405 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 555, + 405 + ], + "type": "text", + "content": "from pexels . com, and a curated list of unsafe words3. To create visualization shown in Figure 3, each retrieved text output is selected only once across all interpolation points, ensuring unique retrievals. Results show that, with HySAC, as the query nears the origin, retrieved content shifts from unsafe to safe while preserving semantic relevance. This progression illustrates the model's capability to effectively navigate the embedding space along relevant paths. For more on the experimental setup and traversal visualizations, see the supplementary E." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 408, + 555, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 408, + 555, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 408, + 555, + 445 + ], + "type": "text", + "content": "Other ablation studies. For additional ablations on embedding space geometry and hyperparameter evaluation, we refer the reader to the supplementary D." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 457, + 388, + 469 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 457, + 388, + 469 + ], + "spans": [ + { + "bbox": [ + 313, + 457, + 388, + 469 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 478, + 555, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 478, + 555, + 693 + ], + "spans": [ + { + "bbox": [ + 313, + 478, + 555, + 693 + ], + "type": "text", + "content": "This paper introduces hyperbolic safety-aware vision-language models. Where recent literature focuses on removing or unlearning unsafe image-text content, we bring a perspective of awareness. By modelling unsafe image-text content as specific cases of their safe counterparts, we can divide the space into safe and unsafe regions. We show that hyperbolic space is a natural solution for this hierarchical relation, and propose a hyperbolic CLIP model with safety entailment learning and traversal. Our approach not only results in better retrieval of relevant safe outputs given unsafe inputs but also provides more robustness and comes with an NSFW classifier as a free by-product. Dealing with NSFW data in vision-language models is an important open research problem with real-world implications, from ethical to legal and societal concerns. By opting for awareness, we find that safety recognition improves. Given the importance of the ethical implications of this work, we provide a thorough discussion in the supplementary A." + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 325, + 702, + 413, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 702, + 413, + 712 + ], + "spans": [ + { + "bbox": [ + 325, + 702, + 413, + 712 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 325, + 702, + 413, + 712 + ], + "type": "text", + "content": "github.com/LDNOOBW/" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4229" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 159, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 159, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 159, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 212 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 212 + ], + "type": "text", + "content": "We acknowledge the CINECA award under the ISCRA initiative, for the availability of high-performance computing resources. This work has been supported by the EU Horizon projects \"ELIAS - European Lighthouse of AI for Sustainability\" (No. 101120237) and \"European Lighthouse on Safe and Secure AI (ELSA)\" (No. 101070617), co-funded by the European Union. Tejaswi Kasarla also acknowledges travel support from the European Union's Horizon research and innovation programme under grant agreement No. 951847 (ELISE) and No. 101120237 (ELIAS)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 220, + 115, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 115, + 232 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 115, + 232 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 240, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 240, + 296, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 240, + 296, + 283 + ], + "spans": [ + { + "bbox": [ + 61, + 240, + 296, + 283 + ], + "type": "text", + "content": "[1] Yuki M. Asano, Christian Rupprecht, Andrew Zisserman, and Andrea Vedaldi. Pass: An imagenet replacement for self-supervised pretraining without humans. NeurIPS Track on Datasets and Benchmarks, 2021. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 285, + 296, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 285, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 62, + 285, + 296, + 316 + ], + "type": "text", + "content": "[2] Mina Ghadimi Atigh, Julian Schoep, Erman Acar, Nanne Van Noord, and Pascal Mettes. Hyperbolic image segmentation. In CVPR, pages 4453-4462, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 318, + 296, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 318, + 296, + 349 + ], + "spans": [ + { + "bbox": [ + 62, + 318, + 296, + 349 + ], + "type": "text", + "content": "[3] Gary Bécigneul and Octavian-Eugen Ganea. Riemannian adaptive optimization methods. arXiv preprint arXiv:1810.00760, 2018. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 351, + 295, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 351, + 295, + 372 + ], + "spans": [ + { + "bbox": [ + 62, + 351, + 295, + 372 + ], + "type": "text", + "content": "[4] P Bedapudi. NudeNet: Neural Nets for Nudity Classification, Detection, and Selective Censoring, 2019. 2, 5, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 373, + 295, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 373, + 295, + 404 + ], + "spans": [ + { + "bbox": [ + 62, + 373, + 295, + 404 + ], + "type": "text", + "content": "[5] Abeba Birhane and Vinay Uday Prabhu. Large image datasets: A pyrrhic win for computer vision? In WACV, pages 1536-1546. IEEE, 2021. 1, 2, 3, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 406, + 295, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 406, + 295, + 448 + ], + "spans": [ + { + "bbox": [ + 62, + 406, + 295, + 448 + ], + "type": "text", + "content": "[6] Abeba Birhane, Vinay Uday Prabhu, and Emmanuel Kahembwe. Multimodal datasets: misogyny, pornography, and malignant stereotypes. arXiv preprint arXiv:2110.01963, 2021. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 449, + 295, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 449, + 295, + 481 + ], + "spans": [ + { + "bbox": [ + 62, + 449, + 295, + 481 + ], + "type": "text", + "content": "[7] Abeba Birhane, Sanghyun Han, Vishnu Boddeti, Sasha Luccioni, et al. Into the laion's den: Investigating hate in multimodal datasets. NeurIPS, 36, 2024. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 483, + 295, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 483, + 295, + 525 + ], + "spans": [ + { + "bbox": [ + 62, + 483, + 295, + 525 + ], + "type": "text", + "content": "[8] Manuel Brack, Felix Friedrich, Patrick Schramowski, and Kristian Kersting. Mitigating inappropriateness in image generation: Can there be value in reflecting the world's ugliness? arXiv preprint arXiv:2305.18398, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 527, + 295, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 527, + 295, + 559 + ], + "spans": [ + { + "bbox": [ + 62, + 527, + 295, + 559 + ], + "type": "text", + "content": "[9] Yinzhi Cao and Junfeng Yang. Towards Making Systems Forget with Machine Unlearning. In IEEE Symposium on Security and Privacy, 2015. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 559, + 295, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 559, + 295, + 603 + ], + "spans": [ + { + "bbox": [ + 57, + 559, + 295, + 603 + ], + "type": "text", + "content": "[10] Francesco Cauteruccio, Enrico Corradini, Giorgio Terracina, Domenico Ursino, and Luca Virgili. Extraction and analysis of text patterns from nsfw adult content in reddit. Data & Knowledge Engineering, 138:101979, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 604, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 604, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 604, + 295, + 635 + ], + "type": "text", + "content": "[11] Benjamin Paul Chamberlain, James Clough, and Marc Peter Deisenroth. Neural embeddings of graphs in hyperbolic space. arXiv preprint arXiv:1705.10359, 2017. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 637, + 295, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 637, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 57, + 637, + 295, + 679 + ], + "type": "text", + "content": "[12] Ines Chami, Adva Wolf, Da-Cheng Juan, Frederic Sala, Sujith Ravi, and Christopher Ré. Low-dimensional hyperbolic knowledge graph embeddings. arXiv preprint arXiv:2005.00545, 2020. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 681, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 714 + ], + "type": "text", + "content": "[13] Nurendra Choudhary, Nikhil Rao, Sumeet Katariya, Karthik Subbian, and Chandan K Reddy. Self-supervised hyperboloid representations from logical queries over knowledge" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 74, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "type": "text", + "content": "graphs. In Proceedings of the Web Conference 2021, pages 1373-1384, 2021." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 97, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 553, + 129 + ], + "type": "text", + "content": "[14] Nurendra Choudhary, Nikhil Rao, and Chandan Reddy. Hyperbolic graph neural networks at scale: a meta learning approach. NeurIPS, 36, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 132, + 553, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 132, + 553, + 175 + ], + "spans": [ + { + "bbox": [ + 316, + 132, + 553, + 175 + ], + "type": "text", + "content": "[15] Damien L Crone, Stefan Bode, Carsten Murawski, and Simon M Laham. The socio-moral image database (smid): A novel stimulus set for the study of social, moral and affective processes. *PloS one*, 13(1):e0190954, 2018. 5, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 177, + 553, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 177, + 553, + 220 + ], + "spans": [ + { + "bbox": [ + 316, + 177, + 553, + 220 + ], + "type": "text", + "content": "[16] Karan Desai, Gaurav Kaul, Zubin Aysola, and Justin Johnson. RedCaps: Web-curated image-text data created by the people, for the people. In NeurIPS Datasets and Benchmarks, 2021. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 223, + 553, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 223, + 553, + 266 + ], + "spans": [ + { + "bbox": [ + 316, + 223, + 553, + 266 + ], + "type": "text", + "content": "[17] Karan Desai, Maximilian Nickel, Tanmay Rajpurohit, Justin Johnson, and Shanmukha Ramakrishna Vedantam. Hyperbolic image-text representations. In ICML, pages 7694-7731. PMLR, 2023. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 269, + 553, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 269, + 553, + 311 + ], + "spans": [ + { + "bbox": [ + 316, + 269, + 553, + 311 + ], + "type": "text", + "content": "[18] Bhuwan Dhingra, Christopher J Shallue, Mohammad Norouzi, Andrew M Dai, and George E Dahl. Embedding text in hyperbolic spaces. arXiv preprint arXiv:1806.04313, 2018. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 314, + 553, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 314, + 553, + 357 + ], + "spans": [ + { + "bbox": [ + 316, + 314, + 553, + 357 + ], + "type": "text", + "content": "[19] Upol Ehsan, Q Vera Liao, Samir Passi, Mark O Riedl, and Hal Daumé III. Seamful xai: Operationalizing seamful design in explainable ai. Proceedings of the ACM on Human-Computer Interaction, 8(CSCW1):1-29, 2024. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 360, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 360, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 360, + 553, + 403 + ], + "type": "text", + "content": "[20] Aleksandr Ermolov, Leyla Mirvakhabova, Valentin Khrulkov, Nicu Sebe, and Ivan Oseledets. Hyperbolic vision transformers: Combining improvements in metric learning. In CVPR, pages 7409-7419, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 405, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 405, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 405, + 553, + 437 + ], + "type": "text", + "content": "[21] Luca Franco, Paolo Mandica, Bharti Munjal, and Fabio Galasso. Hyperbolic self-paced learning for self-supervised skeleton-based action representations. In ICLR, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 440, + 553, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 440, + 553, + 494 + ], + "spans": [ + { + "bbox": [ + 316, + 440, + 553, + 494 + ], + "type": "text", + "content": "[22] Shreyansh Gandhi, Samrat Kokkula, Abon Chaudhuri, Alessandro Magnani, Theban Stanley, Behzad Ahmadi, Venkatesh Kandaswamy, Omer Ovenc, and Shie Mannor. Scalable Detection of Offensive and Non-compliant Content/Logo in Product Images. In WACV, 2020. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 496, + 553, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 496, + 553, + 529 + ], + "spans": [ + { + "bbox": [ + 316, + 496, + 553, + 529 + ], + "type": "text", + "content": "[23] Rohit Gandikota, Joanna Materzynska, Jaden Fiotto-Kaufman, and David Bau. Erasing Concepts from Diffusion Models. In ICCV, 2023. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 531, + 553, + 573 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 531, + 553, + 573 + ], + "spans": [ + { + "bbox": [ + 316, + 531, + 553, + 573 + ], + "type": "text", + "content": "[24] Octavian Ganea, Gary Bécigneul, and Thomas Hofmann. Hyperbolic entailment cones for learning hierarchical embeddings. In ICML, pages 1646-1655. PMLR, 2018. 1, 2, 4" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 577, + 553, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 577, + 553, + 598 + ], + "spans": [ + { + "bbox": [ + 316, + 577, + 553, + 598 + ], + "type": "text", + "content": "[25] Octavian Ganea, Gary Bécigneul, and Thomas Hofmann. Hyperbolic neural networks. NeurIPS, 31, 2018. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 600, + 553, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 600, + 553, + 643 + ], + "spans": [ + { + "bbox": [ + 316, + 600, + 553, + 643 + ], + "type": "text", + "content": "[26] Songwei Ge, Shlok Mishra, Simon Kornblith, Chun-Liang Li, and David Jacobs. Hyperbolic contrastive learning for visual representations beyond objects. In CVPR, pages 6840-6849, 2023. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 646, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 646, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 646, + 553, + 679 + ], + "type": "text", + "content": "[27] Antonio Ginart, Melody Guan, Gregory Valiant, and James Y Zou. Making AI Forget You: Data Deletion in Machine Learning. In NeurIPS, 2019. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[28] Aditya Golatkar, Alessandro Achille, and Stefano Soatto. Eternal Sunshine of the Spotless Net: Selective Forgetting in Deep Networks. In CVPR, 2020. 2" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "4230" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "text", + "content": "[29] Aditya Golatkar, Alessandro Achille, Yu-Xiang Wang, Aaron Roth, Michael Kearns, and Stefano Soatto. Mixed Differential Privacy in Computer Vision. In CVPR, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 295, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 295, + 160 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 295, + 160 + ], + "type": "text", + "content": "[30] Kimia Hamidieh, Haoran Zhang, Walter Gerych, Thomas Hartvigsen, and Marzyeh Ghassemi. Identifying implicit social biases in vision-language models. In Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society, pages 547-561, 2024. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 161, + 294, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 161, + 294, + 204 + ], + "spans": [ + { + "bbox": [ + 56, + 161, + 294, + 204 + ], + "type": "text", + "content": "[31] Ahmad Fathan Hidayatullah, Anisa Miladya Hakim, and Abdullah Aziz Sembada. Adult Content Classification on Indonesian Tweets using LSTM Neural Network. In ICACSIS, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 205, + 294, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 205, + 294, + 249 + ], + "spans": [ + { + "bbox": [ + 56, + 205, + 294, + 249 + ], + "type": "text", + "content": "[32] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 250, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 294, + 293 + ], + "type": "text", + "content": "[33] Sarah Ibrahimi, Mina Ghadimi Atigh, Nanne Van Noord, Pascal Mettes, and Marcel Worring. Intriguing properties of hyperbolic embeddings in vision-language models. Transactions on Machine Learning Research, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 293, + 294, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 293, + 294, + 347 + ], + "spans": [ + { + "bbox": [ + 56, + 293, + 294, + 347 + ], + "type": "text", + "content": "[34] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, pages 4904-4916. PMLR, 2021. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 348, + 294, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 348, + 294, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 348, + 294, + 381 + ], + "type": "text", + "content": "[35] Valentin Khrulkov, Leyla Mirvakhabova, Evgeniya Ustinova, Ivan Oseledets, and Victor Lempitsky. Hyperbolic image embeddings. In CVPR, pages 6418-6428, 2020. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 382, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 382, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 382, + 294, + 415 + ], + "type": "text", + "content": "[36] Nupur Kumari, Bingliang Zhang, Sheng-Yu Wang, Eli Shechtman, Richard Zhang, and Jun-Yan Zhu. Ablating concepts in text-to-image diffusion models. In ICCV, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 415, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 294, + 437 + ], + "type": "text", + "content": "[37] Gant Laborde. Deep nn for nsfw detection. https://github.com/GantMan/nsfw_model, 2020.8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 437, + 294, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 480 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 480 + ], + "type": "text", + "content": "[38] Matthew Le, Stephen Roller, Laetitia Papaxanthos, Douwe Kiela, and Maximilian Nickel. Inferring concept hierarchies from text corpora via hyperbolic embeddings. In ACL, pages 3231-3241, 2019. 2, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 481, + 294, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 481, + 294, + 524 + ], + "spans": [ + { + "bbox": [ + 56, + 481, + 294, + 524 + ], + "type": "text", + "content": "[39] Ya-Wei Eileen Lin, Ronald R Coifman, Gal Mishne, and Ronen Talmon. Hyperbolic diffusion embedding and distance for hierarchical representation learning. In ICML, pages 21003–21025. PMLR, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 525, + 294, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 525, + 294, + 548 + ], + "spans": [ + { + "bbox": [ + 56, + 525, + 294, + 548 + ], + "type": "text", + "content": "[40] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 36, 2024. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 548, + 294, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 569 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 569 + ], + "type": "text", + "content": "[41] Qi Liu, Maximilian Nickel, and Douwe Kiela. Hyperbolic graph neural networks. NeurIPS, 32, 2019. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 570, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 294, + 613 + ], + "type": "text", + "content": "[42] Shaoteng Liu, Jingjing Chen, Liangming Pan, Chong-Wah Ngo, Tat-Seng Chua, and Yu-Gang Jiang. Hyperbolic visual embedding learning for zero-shot recognition. In CVPR, pages 9273-9281, 2020. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 614, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 294, + 635 + ], + "type": "text", + "content": "[43] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 636, + 294, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 294, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 294, + 679 + ], + "type": "text", + "content": "[44] Todor Markov, Chong Zhang, Sandhini Agarwal, Florentine Eloundou Nekoul, Theodore Lee, Steven Adler, Angela Jiang, and Lilian Weng. A Holistic Approach to Undesired Content Detection in the Real World. In AAAI, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "text", + "content": "[45] Joanna Materzyńska, Antonio Torralba, and David Bau. Disentangling visual and written concepts in clip. In CVPR, pages 16410-16419, 2022. 1" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 315, + 72, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 553, + 106 + ], + "type": "text", + "content": "[46] Pascal Mettes, Mina Ghadimi Atigh, Martin Keller-Ressel, Jeffrey Gu, and Serena Yeung. Hyperbolic deep learning in computer vision: A survey. IJCV, pages 1-25, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 108, + 553, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 108, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 315, + 108, + 553, + 152 + ], + "type": "text", + "content": "[47] Paulius Micikevicius, Sharan Narang, Jonah Alben, Gregory Diamos, Erich Elsen, David Garcia, Boris Ginsburg, Michael Houston, Oleksii Kuchaiev, Ganesh Venkatesh, et al. Mixed precision training. In ICLR, 2018. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 153, + 553, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 153, + 553, + 209 + ], + "spans": [ + { + "bbox": [ + 315, + 153, + 553, + 209 + ], + "type": "text", + "content": "[48] Leyla Mirvakhabova, Evgeny Frolov, Valentin Khrulkov, Ivan Oseledets, and Alexander Tuzhilin. Performance of hyperbolic geometry models on top-n recommendation tasks. In Proceedings of the 14th ACM Conference on Recommender Systems, pages 527-532, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 209, + 553, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 209, + 553, + 266 + ], + "spans": [ + { + "bbox": [ + 315, + 209, + 553, + 266 + ], + "type": "text", + "content": "[49] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards Photorealistic Image Generation and Editing with Text-Guided Diffusion Models. arXiv preprint arXiv:2112.10741, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 315, + 267, + 553, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 267, + 553, + 299 + ], + "spans": [ + { + "bbox": [ + 315, + 267, + 553, + 299 + ], + "type": "text", + "content": "[50] Maximillian Nickel and Douwe Kiela. Poincaré embeddings for learning hierarchical representations. NeurIPS, 30, 2017. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 315, + 301, + 553, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 301, + 553, + 335 + ], + "spans": [ + { + "bbox": [ + 315, + 301, + 553, + 335 + ], + "type": "text", + "content": "[51] Maximillian Nickel and Douwe Kiela. Learning continuous hierarchies in the lorentz model of hyperbolic geometry. In ICML, pages 3779-3788. PMLR, 2018. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 315, + 337, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 337, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 315, + 337, + 553, + 392 + ], + "type": "text", + "content": "[52] Avik Pal, Max van Spengler, Guido Maria D'Amely di Melendugno, Alessandro Flaborea, Fabio Galasso, and Pascal Mettes. Compositional entailment learning for hyperbolic vision-language models. arXiv preprint arXiv:2410.06912, 2024. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 315, + 393, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 393, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 315, + 393, + 553, + 426 + ], + "type": "text", + "content": "[53] Wei Peng, Tuomas Varanka, Abdelrahman Mostafa, Henglin Shi, and Guoying Zhao. Hyperbolic deep neural networks: A survey. IEEE Trans. PAMI, 44(12):10023-10044, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 315, + 428, + 553, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 428, + 553, + 473 + ], + "spans": [ + { + "bbox": [ + 315, + 428, + 553, + 473 + ], + "type": "text", + "content": "[54] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 315, + 475, + 553, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 475, + 553, + 518 + ], + "spans": [ + { + "bbox": [ + 315, + 475, + 553, + 518 + ], + "type": "text", + "content": "[55] Samuele Poppi, Tobia Poppi, Federico Cocchi, Marcella Cornia, Lorenzo Baraldi, and Rita Cucchiara. Safe-CLIP: Removing NSFW Concepts from Vision-and-Language Models. In ECCV, 2024. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 315, + 520, + 553, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 520, + 553, + 564 + ], + "spans": [ + { + "bbox": [ + 315, + 520, + 553, + 564 + ], + "type": "text", + "content": "[56] Samuele Poppi, Sara Sarto, Marcella Cornia, Lorenzo Baraldi, and Rita Cucchiara. Multi-Class Unlearning for Image Classification via Weight Filtering. IEEE Intelligent Systems, 2024. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 315, + 565, + 553, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 565, + 553, + 609 + ], + "spans": [ + { + "bbox": [ + 315, + 565, + 553, + 609 + ], + "type": "text", + "content": "[57] Samuele Poppi, Sara Sarto, Marcella Cornia, Lorenzo Baraldi, and Rita Cucchiara. Unlearning vision transformers without retaining data via low-rank decompositions. In ICPR, 2024. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 315, + 611, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 611, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 315, + 611, + 553, + 678 + ], + "type": "text", + "content": "[58] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 1, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 315, + 679, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 679, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 315, + 679, + 553, + 713 + ], + "type": "text", + "content": "[59] Javier Rando, Daniel Paleka, David Lindner, Lennart Heim, and Florian Tramer. Red-teaming the stable diffusion safety filter. In NeurIPS ML Safety Workshop, 2022. 8" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "4231" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "text", + "content": "[60] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "type": "text", + "content": "[61] Frederic Sala, Chris De Sa, Albert Gu, and Christopher Ré. Representation tradeoffs for hyperbolic embeddings. In ICML, pages 4460-4469. PMLR, 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "text", + "content": "[62] Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 239 + ], + "type": "text", + "content": "[63] Patrick Schramowski, Christopher Tauchmann, and Kristian Kersting. Can Machines Help Us Answering Question 16 in Datasheets, and In Turn Reflecting on Inappropriate Content? In ACM FAccT, 2022. 2, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 241, + 294, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 294, + 283 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 294, + 283 + ], + "type": "text", + "content": "[64] Patrick Schramowski, Manuel Brack, Björn Deiseroth, and Kristian Kersting. Safe Latent Diffusion: Mitigating Inappropriate Degeneration in Diffusion Models. In CVPR, 2023. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 286, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 294, + 319 + ], + "type": "text", + "content": "[65] Christoph Schuhmann. Clip based nsfw detector. https://github.com/LAION-AI/CLIP-based-NSFW-Detector/tree/main, 2022.8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 320, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 294, + 374 + ], + "type": "text", + "content": "[66] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 1, 5, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 376, + 294, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 294, + 420 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 294, + 420 + ], + "type": "text", + "content": "[67] Sheng Shen, Liunian Harold Li, Hao Tan, Mohit Bansal, Anna Rohrbach, Kai-Wei Chang, Zhewei Yao, and Kurt Keutzer. How much can clip benefit vision-and-language tasks? arXiv preprint arXiv:2107.06383, 2021. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 421, + 294, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 421, + 294, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 421, + 294, + 443 + ], + "type": "text", + "content": "[68] Ryohei Shimizu, Yusuke Mukuta, and Tatsuya Harada. Hyperbolic neural networks++. ICLR, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 445, + 294, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 465 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 465 + ], + "type": "text", + "content": "[69] David Thiel. Identifying and eliminating csam in generative ml training data and models. 2023. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 467, + 294, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 294, + 499 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 294, + 499 + ], + "type": "text", + "content": "[70] Alexandru Tifrea, Gary Bécigneul, and Octavian-Eugen Ganea. Poincaré glove: Hyperbolic word embeddings. arXiv preprint arXiv:1810.06546, 2018. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 501, + 294, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 294, + 522 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 294, + 522 + ], + "type": "text", + "content": "[71] Max van Spengler, Erwin Berkhout, and Pascal Mettes. Poincare resnet. In ICCV, pages 5419-5428, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 524, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 524, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 524, + 294, + 555 + ], + "type": "text", + "content": "[72] Ivan Vendrov, Ryan Kiros, Sanja Fidler, and Raquel Urtasun. Order-embeddings of images and language. arXiv preprint arXiv:1511.06361, 2015. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 557, + 294, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 294, + 610 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 294, + 610 + ], + "type": "text", + "content": "[73] Liping Wang, Fenyu Hu, Shu Wu, and Liang Wang. Fully hyperbolic graph convolution network for recommendation. In Proceedings of the 30th ACM international conference on information & knowledge management, pages 3483-3487, 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 613, + 294, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 294, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 294, + 645 + ], + "type": "text", + "content": "[74] Mengmeng Wang, Jiazheng Xing, and Yong Liu. Actionclip: A new paradigm for video action recognition. arXiv preprint arXiv:2109.08472, 2021. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "type": "text", + "content": "[75] Robert Wolfe, Yiwei Yang, Bill Howe, and Aylin Caliskan. Contrastive language-vision ai models pretrained on web-scraped multimodal data exhibit sexual objectification bias. In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency, pages 1174-1185, 2023. 1" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 553, + 140 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 315, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 315, + 73, + 553, + 106 + ], + "type": "text", + "content": "[76] Eric Zhang, Kai Wang, Xingqian Xu, Zhangyang Wang, and Humphrey Shi. Forget-Me-Not: Learning to Forget in Text-to-Image Diffusion Models. CVPR, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "type": "text", + "content": "[77] Yudong Zhu, Di Zhou, Jinghui Xiao, Xin Jiang, Xiao Chen, and Qun Liu. Hypertext: Endowing fasttext with hyperbolic geometry. arXiv preprint arXiv:2010.16143, 2020. 2" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4232" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Hyperbolic Uncertainty-Aware Few-Shot Incremental Point Cloud Segmentation/1247e066-a517-4ffe-b97e-a9bff43b89c3_content_list.json b/2025/Hyperbolic Uncertainty-Aware Few-Shot Incremental Point Cloud Segmentation/1247e066-a517-4ffe-b97e-a9bff43b89c3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..99764336e5fc3e9d32148dbed089d395c2482cd4 --- /dev/null +++ b/2025/Hyperbolic Uncertainty-Aware Few-Shot Incremental Point Cloud Segmentation/1247e066-a517-4ffe-b97e-a9bff43b89c3_content_list.json @@ -0,0 +1,1848 @@ +[ + { + "type": "text", + "text": "Hyperbolic Uncertainty-Aware Few-Shot Incremental Point Cloud Segmentation", + "text_level": 1, + "bbox": [ + 93, + 130, + 903, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tanuj Sur† Samrat Mukherjee* Kaizer Rahaman‡ Subbasis Chaudhuri* Muhammad Haris Khan§ Biplab Banerjee*", + "bbox": [ + 109, + 179, + 883, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "†Chennai Mathematical Institute *Indian Institute of Technology Bombay ‡Indian Institute of Technology Kharagpur §Mohamed Bin Zayed University of Artificial Intelligence", + "bbox": [ + 101, + 217, + 908, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 286, + 326, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D point cloud segmentation is essential across a range of applications; however, conventional methods often struggle in evolving environments, particularly when tasked with identifying novel categories under limited supervision. Few-Shot Learning (FSL) and Class Incremental Learning (CIL) have been adapted previously to address these challenges in isolation, yet the combined paradigm of Few-Shot Class Incremental Learning (FSCIL) remains largely unexplored for point cloud segmentation. To address this gap, we introduce Hyperbolic Ideal Prototypes Optimization (HiPO), a novel framework that harnesses hyperbolic embeddings for FSCIL in 3D point clouds. HiPO employs the Poincaré Hyperbolic Sphere as its embedding space, integrating Ideal Prototypes enriched by CLIP-derived class semantics, to capture the hierarchical structure of 3D data. By enforcing orthogonality among prototypes and maximizing representational margins, HiPO constructs a resilient embedding space that mitigates forgetting and enables the seamless integration of new classes, thereby effectively countering overfitting. Extensive evaluations on S3DIS, ScanNetv2, and cross-dataset scenarios demonstrate HiPO's strong performance, significantly surpassing existing approaches in both in-domain and cross-dataset FSCIL tasks for 3D point cloud segmentation.", + "bbox": [ + 88, + 318, + 485, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 704, + 220, + 719 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advancements in deep learning have spurred progress in data-efficient learning, with semi-supervised learning (SSL) [62, 69, 71] and few-shot learning (FSL) [55, 58, 66] emerging as key approaches. SSL utilizes both labeled and unlabeled data to enhance learning, while FSL enables generalization from only a few labeled samples per class. However, both approaches face challenges in real-world, dynamic environments where models must continuously adapt to new classes while retaining prior knowledge", + "bbox": [ + 89, + 729, + 483, + 864 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/70e6e92c326f538e13202a67060afa89ed8e4d8a880cca0433775eef9025159b.jpg", + "image_caption": [ + "Figure 1. Our approach tackles the novel challenge of few-shot class incremental 3D point cloud segmentation. In each session $S^{(t)}, t \\geq 0$ , our model takes an RGB point cloud of an indoor scene $D^{(t)}$ as input and performs semantic segmentation. During the base training session $S^{(0)}$ , sufficient examples are available for each class. In later sessions, only a limited number of labeled point cloud examples are provided for new classes, introducing data imbalance and potential overfitting and forgetting risks." + ], + "image_footnote": [], + "bbox": [ + 526, + 282, + 903, + 445 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "edge—an essential requirement for continual learning (CL) [1, 4, 8, 11, 32, 48, 56]. Few-Shot Class Incremental Learning (FSCIL) [1, 13, 24, 28, 30, 39] combines the data efficiency of FSL with the adaptability of CL, yet remains underdeveloped for complex tasks such as segmentation in unstructured 3D data (see Fig. 1).", + "bbox": [ + 511, + 582, + 906, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Point cloud segmentation is pivotal in applications like autonomous driving and robotics, where it enables precise object recognition and scene understanding [36, 52, 68]. This task assigns semantic labels to each point within a 3D structure, facilitating critical decision-making processes. Current supervised segmentation methods [26, 53, 65, 67] are highly reliant on extensive labeled datasets, which are costly and time-intensive to produce [34]. These methods also assume a fixed set of labels, limiting their adaptability in scenarios where new classes emerge unexpectedly. While FSCIL offers a potential solution, existing methods [1, 24, 72, 78, 79, 81, 85] struggle to represent the complex, hierarchical structures typical of point cloud data due to the limitations of Euclidean embeddings [33, 45, 46].", + "bbox": [ + 511, + 674, + 908, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Capturing hierarchical relationships, such as recogniz", + "bbox": [ + 531, + 885, + 903, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Samrat Mukherjee acknowledges the support of Prime Minister Research Fellowship (PMRF).", + "bbox": [ + 89, + 875, + 482, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "11810", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing subcomponents within an object, demands a framework that can accommodate intricate interrelations. This challenge is intensified in Class Incremental Learning (CIL) [82, 83], where the model must incorporate new classes while preserving previously learned representations—a balance between adaptability and memory known as the stability-plasticity trade-off. In FSCIL, especially under data scarcity, this trade-off is often compromised, leading to overfitting and catastrophic forgetting. Euclidean spaces exacerbate these challenges, as they lack the capacity to model evolving data scenarios [76, 81].", + "bbox": [ + 89, + 90, + 480, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Hyperbolic geometry [25, 38, 42, 57, 74] provides a more suitable framework for hierarchical data [45, 84]. Its negative curvature and exponential volume growth allow it to mirror tree-like data structures, addressing the stability-plasticity trade-off [16] and supporting segmentation in unstructured contexts [27]. Unlike Euclidean spaces, hyperbolic space offers exponentially more volume [44] for accommodating incremental classes, thus reducing overlap and interference (see Fig. 2, Tab. 4). However, efficient optimization of embeddings in the hyperbolic space presents computational challenges when points are near the boundary of the Poincaré ball resulting in vanishing gradient problems [44], and without careful handling, newly added classes can still encroach upon the existing embeddings, aggravating plasticity (see Tab. 3). This highlights the need for an FSCIL framework that uses hyperbolic space embedding hierarchical data in low distortion while ensuring computational efficiency and managing class overlap.", + "bbox": [ + 89, + 260, + 482, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our Proposal: To address these challenges, we introduce Hyperbolic Ideal Prototypes Optimization (HiPo), a novel FSCIL framework tailored for point cloud segmentation. By embedding data within the Poincaré Hyperbolic Sphere [49], HiPo leverages the unique properties of hyperbolic space to capture hierarchical structures and maintain consistent feature-classifier alignment across incremental learning sessions. Through a modified Busemann loss [7], HiPo arranges ideal prototypes orthogonally within the hyperbolic space, maximizing class separation and preserving space for new classes, which minimizes inter-class interference and optimizes the feature space for CIL.", + "bbox": [ + 89, + 535, + 482, + 717 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This design takes advantage of hyperbolic space's exponential scaling, which aligns naturally with complex hierarchies found in point cloud data [3, 27, 45, 57]. Our modified Busemann loss incorporates two components: (1) a prototype alignment term to guide class embeddings towards their respective Hyperbolic Ideal Prototypes, enhancing class separability, and intra-class compactness, and (2) an uncertainty-aware term that penalizes overfitting on novel classes and preventing vanishing gradient problems. Together, these components enable HiPo to effectively address data scarcity, mitigate catastrophic forgetting, and support robust FSCIL under episodic shifts, establishing it", + "bbox": [ + 89, + 719, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6041dc375c4e0f7391ca49c8dcf2da0eaafdb6ef4b254da7fb450ac065593299.jpg", + "image_caption": [ + "Figure 2. Motivation for hyperbolic embeddings: (a) Hyperbolic UMAP [40] of the HiPo embedding space, projected to 2D for the S3DIS dataset [2]. The hyperbolic space's negative curvature aligns with hierarchical data, enabling a more compact representation and clearer class separation than Euclidean space (Dunn index [18] $(\\uparrow)$ for HiPo is 0.69 while for the Euclidean space, it is 0.35), showcasing better class compactness in our case. (b) UMAP of the classifier space in the Euclidean hypersphere" + ], + "image_footnote": [], + "bbox": [ + 539, + 93, + 874, + 237 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "as a powerful solution for incremental learning in 3D point cloud segmentation. Overall, our contributions are:", + "bbox": [ + 511, + 383, + 906, + 412 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We define and address the previously unexplored problem of FSCIL for 3D point cloud segmentation, encompassing both in-domain and cross-dataset scenarios.", + "- We introduce hyperbolic space as a robust embedding framework for FSCIL in 3D segmentation, effectively capturing hierarchical structures and preserving stable feature-class relationships across incremental learning sessions.", + "- We propose a modified Busemann loss with prototype alignment and uncertainty-aware components to ensure optimal feature-classifier alignment with minimized uncertainty in predictions.", + "- Through extensive experiments on benchmark datasets S3DIS [2] and ScanNetv2 [17], HiPo demonstrates superior performance and generalization, setting a new standard for FSCIL in this domain." + ], + "bbox": [ + 511, + 415, + 903, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 662, + 663, + 678 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Point-Cloud Semantic Segmentation: Fully supervised 3D semantic segmentation has been a well-explored area, with methods typically requiring dense point-wise labels [26, 52, 53, 65, 67, 77]. A significant advancement was PointNet [52], which directly processes raw point clouds, removing the need for voxelization but facing challenges in capturing local structures. This limitation spurred improvements like PointNet++ [53], which introduced a hierarchical approach to enhance local feature learning, and DGCNN [65], which incorporated the EdgeConv module to integrate neighborhood information for better geometric representation. More recently, models such as PointTransformer [80] have integrated self-attention mechanisms, inspired by NLP, to elevate feature expressiveness and contextual awareness.", + "bbox": [ + 511, + 688, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "11811", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite these advancements, the dependency on large labeled datasets and the assumption of a fixed class set restrict these methods in practical scenarios where new classes must be incrementally learned with minimal supervision.", + "bbox": [ + 89, + 90, + 482, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "(Few-Shot) Class Incremental Learning: Incremental learning methods for 3D data have predominantly focused on classification tasks rather than segmentation [14, 15, 19, 35, 75]. Approaches such as I3DOL [19] and L3DOC [35] mitigate catastrophic forgetting in 3D object classification through memory and attention mechanisms, yet they struggle as labeled supervision diminishes over time. Few-Shot Class Incremental Learning (FSCIL) [61] has emerged as a more flexible solution, enabling models to learn new classes incrementally with minimal labeled data. Representative methods like CEC [76] use graph-based task linking, while FACT [81] and ALICE [50] create adaptable feature spaces through clustering and cosine similarity. However, these techniques rely on Euclidean embeddings, which have limited capability to represent hierarchical structures intrinsic to 3D data, and frequent classifier updates introduce computational overhead [76]. Open-world semantic segmentation (OWSS) [5, 9] is a related area that focuses on dynamically detecting unknown classes at inference, though our focus on incrementally adding known classes during training sets our approach apart from OWSS.", + "bbox": [ + 91, + 152, + 483, + 469 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Hyperbolic Representations for Hierarchical Data: Hyperbolic geometry has shown promise for hierarchical data representation, with foundational work by Sarkar et al. [57] and Nickel et al. [49] introducing hyperbolic neural networks and Riemannian optimization methods. Hyperbolic embeddings have demonstrated substantial benefits for capturing semantic hierarchies across applications, including image embeddings [27] and FSCIL for image classification [16]. In 3D data, recent research has shown that hyperbolic representations can model complex part-whole relationships effectively. Montanaro et al. [45] applied hyperbolic regularizers to encode hierarchical structures in 3D objects, while Hu et al. [25] leveraged hyperbolic geometry to represent latent structures in point clouds.", + "bbox": [ + 89, + 470, + 482, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our approach leverages hyperbolic embeddings to tackle FSCIL challenges in segmentation, improving adaptability, preserving hierarchies, and ensuring computational efficiency in the point cloud domain.", + "bbox": [ + 89, + 683, + 483, + 744 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Problem Definition & Methodology", + "text_level": 1, + "bbox": [ + 89, + 760, + 410, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $D$ represent a sequence of training datasets $\\{D^{(0)}, D^{(1)}, \\ldots, D^{(T)}\\}$ , where each $D^{(t)} = \\{(P_k^{(t)}, M_k^{(t)})_{k=1}^{|D^{(t)}|}\\}$ corresponds to a training session $S^{(t)}$ , $t \\in \\{0, 1, \\ldots, T\\}$ . The initial session dataset $D^{(0)}$ serves as the base, while $T$ denotes the total number of incremental sessions. In each session $S^{(t)}$ , $P_k^{(t)}$ represents a point cloud sample, and $M_k^{(t)}$ provides its labels within", + "bbox": [ + 89, + 787, + 482, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the label space $C^{(t)} = \\{c_1^{(t)},c_2^{(t)},\\ldots ,c_{|C^{(t)}|}^{(t)}\\}$ where $c_{i}^{(t)}$ is the $i$ -th class in the incremental session $t$ .", + "bbox": [ + 511, + 88, + 903, + 122 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The base session $S^{(0)}$ typically contains an abundant number of samples per class in the label space $C^{(0)}$ . However, in each subsequent incremental session $S^{(t)}, t > 0$ , only a limited number of labeled point clouds $P_{k}^{(t)} \\in \\mathbb{R}^{m \\times d_{0}}$ are provided, where $m$ is the number of points per sample and $d_{0}$ is the feature dimension. Following previous works [1, 24, 72, 78, 79, 81, 85], we assume $|D^{(t)}| = N \\times K$ , where $N$ is the number of classes, and $K$ is the number of samples per class, following an $N$ -way $K$ -shot setting $\\forall t > 0$ . We assume $C^{(t)}$ to be disjoint from previous sessions, i.e., $C^{(t)} \\cap C^{(t')} = \\emptyset$ , $\\forall t' < t$ .", + "bbox": [ + 511, + 123, + 903, + 291 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "During each incremental session, only the current dataset $D^{(t)}$ is accessible, with data from previous sessions unavailable for training. For evaluation in session $t$ , the model is tested across all classes encountered up to that session, covering the label space $\\bigcup_{i=0}^{t} C^{(i)}$ . A solution to the FSCIL problem must jointly address uncharted catastrophic forgetting, data imbalance, and class overfitting judiciously.", + "bbox": [ + 511, + 292, + 903, + 398 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 513, + 405, + 653, + 419 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section mentions the essential backgrounds needed for the proposed framework Hyperbolic Ideal Prototypes Optimization (HiPo). We extensively use the concepts of the Poincaré ball model and its associated Möbius addition, exponential map, and the geodesic distance. We detail their definition and discuss their formulae in the Sup. Mat.", + "bbox": [ + 511, + 426, + 903, + 517 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the Poincaré ball model of hyperbolic geometry, all the points at the boundary of the Poincaré ball represent points at infinity. We initialize a group of classifier prototypes at the boundary of the Poincaré ball termed Ideal Prototypes, following [23], for each of the $n$ individual classes. Assuming $u_{1}, u_{2}, \\dots, u_{n}$ to be the Ideal Prototypes for $n$ classes, they form the boundary of the $n$ -dimensional Poincaré ball of curvature $c$ , $\\mathbb{B}_c^n$ , following [23]:", + "bbox": [ + 511, + 518, + 905, + 640 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {I} _ {n} = \\left\\{u \\in \\mathbb {R} ^ {n}: u _ {1} ^ {2} + u _ {2} ^ {2} + \\dots + u _ {n} ^ {2} = 1 \\right\\} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 565, + 652, + 903, + 670 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since Ideal Prototypes are points at an infinite geodesic distance from all other points in $\\mathbb{B}_c^n$ , we use the Busemann function [7] as a distance measure since computing direct infinite geodesic distance is non-trivial [23].", + "bbox": [ + 511, + 674, + 905, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The pre-assigned Ideal Prototypes form an optimal geometric structure by fixing the target throughout the incremental learning. However, as already pointed out, optimizing Ideal Prototypes in the Poincaré model to address FS-CIL in 3D point cloud segmentation remains unexplored.", + "bbox": [ + 511, + 734, + 903, + 810 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Discussing HiPo", + "text_level": 1, + "bbox": [ + 511, + 818, + 678, + 834 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our model introduces a fixed alignment strategy for feature-classifier relationships in FSCIL, tailored for the hierarchical nature of data such as 3D point clouds. Traditional Euclidean embeddings suffer from compaction [50, 76, 81]", + "bbox": [ + 511, + 839, + 905, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "11812", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/63cf5ba78a7e0238dbb7cc650334e3c091fd8659e767f93bccc19eb1944a7509.jpg", + "image_caption": [ + "Figure 3. HiPo Overview: We show the changes in the Poincaré ball during the initial three sessions in the 3-2T FSCIL setting (three base classes, two incremental classes per session) for the S3DIS dataset [2], designed to easily visualize the working of HiPo. Hyperbolic Ideal Prototypes are optimally positioned to increase the margin of semantically similar classes $(\\mathcal{L}_{\\mathrm{HIP}})$ . As an example, the chair class gets positioned farther from sofa than from table or clutter. We align representations of the newly introduced classes with their corresponding ideal prototypes $(\\mathcal{L}_{\\mathrm{Alignment}})$ regularized by a modified Busemann function $(\\mathcal{L}_{\\mathrm{Uncertainty}})$ . Training is conducted using $\\mathcal{L}_{\\mathrm{HiPo}}$ across sessions." + ], + "image_footnote": [], + "bbox": [ + 94, + 85, + 336, + 281 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0ca66a39feff0e83af01997e2c0c0cc6be61082c8c5238757cd6bcf5c77efcd5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 339, + 85, + 581, + 282 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8792de0dd8c23bfe0c414fffff5e860668178ff2bc3fd3c05c0d29f4f04bdfc6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 583, + 85, + 898, + 282 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and distortion as new classes are added, leading to overlap and forgetting. Hyperbolic space, with its negative curvature and exponential volume, provides a natural fit for preserving hierarchical structures while minimizing inter-class interference. To address catastrophic forgetting, our model anchors feature-classifier relationships within hyperbolic space, ensuring stable alignment between features and classifier prototypes. This fixed alignment prevents the drift of previously learned embeddings as new classes are introduced, preserving the original classifier prototype positions and maintaining consistency across incremental sessions. By leveraging hyperbolic geometry, our approach enables efficient class separation, robust memory retention, and adaptability, making it well-suited for FSCIL in high-dimensional, unstructured data settings. Our method has two main stages:", + "bbox": [ + 88, + 378, + 482, + 619 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Defining Ideal Prototypes: We initialize ideal prototypes for each class on the boundary of $\\mathbb{B}_c^n$ , positioning them mutually orthogonal and at maximal distances to enhance discriminative power, especially among classes with similar characteristics (see Sec. 3.2.1). This fixed configuration allows a dedicated space for each class, preventing the embedding drift observed in Euclidean spaces.", + "- Efficient Projections onto Poincaré ball $\\mathbb{B}_c^n$ : In Hyperbolic embedding space, mean class features are obtained by computing Fréchet [20] means over the feature embedding for the training examples belonging to the classes. Instead of directly computing class features on the Poincaré ball—an intensive task[63]—we obtain the feature vectors in Euclidean space for every class in each session and then project them into hyperbolic space using exponential map $\\exp_0(.)$ [45] (see Sec. 3.2.2). The hyperbolic classifier head $\\mathcal{H}(.)$ consisting of a Möbius layer $\\mathcal{M}$ (also known as hyperbolic feed forward layer [22]), takes the output of the" + ], + "bbox": [ + 89, + 625, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "exponential layer as input and projects mean class features to $\\mathbb{B}_c^n$ . The Poincaré ball addresses few-shot learning using the sample mean prototype as the class representation.", + "bbox": [ + 511, + 378, + 903, + 424 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our framework consists of a point-level classification branch, combining a feature extractor $\\mathcal{E}(.)$ and hyperbolic classifier head $\\mathcal{H}(.)$ , and is trained with two key strategies:", + "bbox": [ + 511, + 425, + 903, + 469 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Feature-Classifier Alignment: For each session $t \\geq 0$ , we align the hyperbolic sample mean features (see Eqn. 4) with the Hyperbolic Ideal Prototypes of their respective classes using alignment loss, $\\mathcal{L}_{\\mathrm{Alignment}}$ . We define $\\mathcal{L}_{\\mathrm{Alignment}}$ using Busemann function $\\mathcal{B}$ and introduce a regularization term, quantifying uncertainty of the class features, to prevent feature collapse around the ideal prototypes, effectively reducing interference with previously seen classes (see Sec. 3.2.3) and vanishing gradient problems [44]. Point-Level Classification: The hyperbolic classifier $\\mathcal{H}(.)$ , consisting of a Möbius layer $\\mathcal{M}$ followed by softmax [45], processes the output of $\\exp_0(\\mathcal{E}(.))$ , with a cross-entropy loss for 3D point cloud segmentation. The joint objective combines this alignment loss $\\mathcal{L}_{\\mathrm{Alignment}}$ and cross-entropy loss $\\mathcal{L}_{\\mathrm{CE}}$ as follows:", + "bbox": [ + 511, + 469, + 905, + 696 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {H I P O}} = \\mathcal {L} _ {\\text {A l i g n m e n t}} + \\mathcal {L} _ {\\mathrm {C E}} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 704, + 903, + 720 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This approach ensures robust feature-classifier alignment across incremental learning sessions, addressing catastrophic forgetting in 3D point cloud segmentation by preserving both old and new class representations in a rehearsal-free manner. The process is illustrated in Fig. 3, and Algorithm 1 and a table summarizing the important variables used, mentioned in the Sup. Mat., supplement these discussions.", + "bbox": [ + 511, + 728, + 905, + 848 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1. Hyperbolic ideal prototypes positioning", + "text_level": 1, + "bbox": [ + 511, + 851, + 834, + 867 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Building upon the goal of maintaining stable feature-classifier alignment, our framework pre-assigns classifier", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "11813", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "prototypes on the boundary of $\\mathbb{B}_c^n$ to define a structured classifier space prior to backbone training. To achieve maximal margin separation between classes, we introduce inductive biases by assigning each class a Hyperbolic Ideal Prototype, distributing these prototypes as uniformly as possible across the Poincaré ball. Inspired by Mettes et al. [41], we incorporate semantic class knowledge to enforce distinct separations, particularly for closely related classes, such as \"chair\" and \"sofa\". To further enhance semantic cohesion, we utilize the CLIP framework [54] due to its robust semantic representations [31, 73] in comparison to embeddings like word2vec [43], Glove [51] etc. By querying CLIP's frozen text encoder with the prompt \"A photo of a [class]\", we obtain text embeddings $S = \\{s_1,\\dots ,s_n\\}$ for each of the $n$ predefined classes. Leveraging the homeomorphism between hyperbolic space $\\mathbb{B}_c^n$ and the hypersphere $\\mathbb{S}^n$ [23], we maximize separation between nearest neighboring embeddings in $\\mathbb{S}^n$ and enforce mutual orthogonality among Ideal Prototypes to mitigate conflict between new and prior knowledge. To achieve this, inspired by RankNet [6], we formulate a Hyperbolic positioning loss, $\\mathcal{L}_{\\mathrm{HIP}}$ , optimizing the ideal prototypes positions as:", + "bbox": [ + 89, + 90, + 483, + 425 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal{L}_{\\text{HIP}} = \\frac{1}{|M|}\\sum_{(i,i^{\\prime},i^{\\prime \\prime})\\in M} - \\bar{W}_{ii^{\\prime}i^{\\prime \\prime}}\\log W_{ii^{\\prime}i^{\\prime \\prime}} - \\\\ (1 - \\bar {W} _ {i i ^ {\\prime} i ^ {\\prime \\prime}}) \\log (1 - W _ {i i ^ {\\prime} i ^ {\\prime \\prime}}) \\tag {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 431, + 480, + 492 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $M$ represents all class triplets, and $\\bar{W}_{ii^{\\prime}i^{\\prime \\prime}} = \\llbracket \\cos \\theta_{\\mathbf{s}_i,\\mathbf{s}_{i^{\\prime}}} \\geq \\cos \\theta_{\\mathbf{s}_i,\\mathbf{s}_{i^{\\prime \\prime}}} \\rrbracket$ denotes triplet ranking order, with $W_{ii^{\\prime}i^{\\prime \\prime}} = \\frac{e^{o_{ii^{\\prime}}i^{\\prime \\prime}}}{1 + e^{o_{ii^{\\prime}}i^{\\prime \\prime}}} \\text{ and } o_{ii^{\\prime}i^{\\prime \\prime}} = \\cos \\theta_{\\mathbf{u}_i,\\mathbf{u}_{i^{\\prime}}} - \\cos \\theta_{\\mathbf{u}_i,\\mathbf{u}_{i^{\\prime \\prime}}}$ , where $\\mathbf{u}_i, \\mathbf{u}_{i^{\\prime}}, \\mathbf{u}_{i^{\\prime \\prime}}$ are ideal prototypes for the corresponding classes $i, i^{\\prime}$ and $i^{\\prime \\prime}$ respectively.", + "bbox": [ + 89, + 497, + 483, + 585 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following Mettes et al. [41], $\\mathcal{L}_{\\mathrm{HIP}}$ optimizes the prototypes' placement on $\\mathbb{B}_c^n$ by aligning the hyperbolic prototypes with semantic ranking. This optimization establishes a classifier structure that facilitates inter-class separability and mutual orthogonality. Hyperbolic space's exponentially larger embedding capacity [44] further ensures that new classes integrate seamlessly without distorting existing embeddings, promoting robust feature organization across learning episodes.", + "bbox": [ + 89, + 585, + 483, + 720 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.2. Hyperbolic class prototype projection", + "text_level": 1, + "bbox": [ + 89, + 727, + 398, + 742 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We model the hierarchical complexity in 3D point clouds using a hyperbolic manifold representation, which preserves nuanced class relationships with minimal distortion.", + "bbox": [ + 89, + 744, + 482, + 790 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "At each session $t \\geq 0$ , we apply the feature extractor $\\mathcal{E}(:, \\theta)$ on input point clouds $\\{(P_k^{(t)})_{k=1}^{|D^{(t)}|}\\}$ to obtain feature vectors $\\pmb{v}$ in Euclidean space. To avoid the computational cost of calculating Fréchet means directly in hyperbolic space [63], we compute class-mean feature vectors in Euclidean space and project them onto the hyperbolic manifold $\\mathbb{B}_c^n$ using $\\exp_0$ and $\\mathcal{M}$ as follows:", + "bbox": [ + 89, + 791, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nv _ {k} ^ {(t)} = \\mathcal {E} \\left(P _ {k} ^ {(t)}; \\theta\\right), \\quad \\forall k \\in \\{1, \\dots , | D ^ {(t)} | \\}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 101, + 846, + 122 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\bar {v} \\left(c _ {i} ^ {(t)}\\right) = \\frac {1}{| D ^ {(t)} |} \\sum_ {k = 1} ^ {| D ^ {(t)} |} v _ {k} ^ {(t)} \\mathbb {1} \\left[ M _ {k} ^ {(t)} = c _ {i} ^ {(t)} \\right] \\tag {4} \\\\ z (c _ {i} ^ {(t)}) = \\mathcal {M} (\\exp_ {0} (\\overline {{v}} (c _ {i} ^ {(t)}))) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 126, + 903, + 193 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $i = \\{1,\\dots ,|C^{(t)}|\\}$ and $\\theta$ denotes learnable parameters. Here, $z(c_{i}^{(t)})$ represents the hyperbolic prototype for each class $c_{i}^{(t)}\\in C^{(t)}$ within the open ball $\\mathbb{B}_c^n = \\{x\\in$ $\\mathbb{R}^n:c\\| x\\| < 1\\}$ . This projection process ensures cohesive, low-distortion representations of class structures as the model adapts incrementally to new classes.", + "bbox": [ + 511, + 198, + 906, + 295 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.3. Uncertainty-aware Busemann learning", + "text_level": 1, + "bbox": [ + 511, + 301, + 831, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Previous works [12, 21, 27] highlight the utility of hyperbolic distance from any point to the origin $\\mathcal{O}_{\\mathbb{B}}$ in $\\mathbb{B}_c^n$ as a measure of uncertainty, providing motivation to utilize the hyperbolic manifold for robust segmentation. In our approach, we align class prototypes with Ideal Prototypes placed at infinity on the hyperbolic boundary. To refine this alignment, we penalize the uncertainty of projected class features $z(c_i^{(t)})$ , represented by their distance $d_{\\mathbb{B}}(z(c_i^{(t)}),\\mathcal{O}_{\\mathbb{B}})$ , thereby preventing embedding drift from $\\mathcal{O}_{\\mathbb{B}}$ and ensuring stability.", + "bbox": [ + 511, + 319, + 906, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "These Ideal Prototypes lie at an infinite geodesic distance from all points within $\\mathbb{B}_c^n$ , making direct prototypical learning impractical. We address this with the Busemann function $\\mathcal{B}$ introduced in [7]. Ghadimi et al. [23] provides a closed-form expression for measuring such infinite geodesic distances. Let $u_{i}$ be the Ideal Prototype for class $c_{i}^{(t)}$ from the set $\\mathcal{I}_n$ , and $\\gamma_{u_i^t}$ a geodesic ray towards $u_{i}^{t}$ , with $z_{i}^{t} = z(c_{i}^{(t)})$ . The Busemann function $\\mathcal{B}$ for a hyperbolic embedding $z_{i}^{t}\\in \\mathbb{B}_{c}^{n}$ relative to $u_{i}^{t}$ is defined as:", + "bbox": [ + 511, + 476, + 905, + 619 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {B} _ {u _ {i} ^ {t}} \\left(z _ {i} ^ {t}\\right) = \\lim _ {q \\rightarrow \\infty} \\left(d _ {\\mathbb {B}} \\left(\\gamma_ {u _ {i} ^ {t}} (q), z _ {i} ^ {t}\\right) - q\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 628, + 906, + 655 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the Poincaré ball $\\mathbb{B}_c^n$ , this simplifies to:", + "bbox": [ + 532, + 659, + 810, + 674 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {B} _ {u _ {i} ^ {t}} \\left(z _ {i} ^ {t}\\right) = \\log \\frac {\\left\\| u _ {i} ^ {t} - z _ {i} ^ {t} \\right\\| ^ {2}}{1 - \\left\\| z _ {i} ^ {t} \\right\\| ^ {2}}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 684, + 903, + 719 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We also incorporate hyperbolic distance as an uncertainty measure for class prototypes in $\\mathbb{B}_c^n$ , which integrates into our loss function as in [23]. The alignment loss introduced in Eq. 2 is defined as:", + "bbox": [ + 511, + 722, + 905, + 782 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {A l g i n m e n t}} = \\mathcal {L} _ {\\mathcal {B}} + \\phi \\mathcal {L} _ {\\text {U n c e r t a i n t y}} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 789, + 903, + 806 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where Busemann loss $\\mathcal{L}_{\\mathcal{B}} = \\mathcal{B}_{u_i^t}(z_i^t)$ , uncertainty loss $\\mathcal{L}_{\\mathrm{Uncertainty}} = d_{\\mathbb{B}}(z_i^t,\\mathcal{O}_{\\mathbb{B}})$ and $\\phi$ is the strength of the regularization effect of uncertainty loss.", + "bbox": [ + 511, + 813, + 905, + 859 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The overall loss function in Eq. 2 can be formulated as:", + "bbox": [ + 532, + 861, + 898, + 876 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {H i P o}} = \\mathcal {L} _ {\\mathcal {B}} + \\phi \\mathcal {L} _ {\\text {U n c e r t a i n t y}} + \\mathcal {L} _ {\\mathrm {C E}}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 885, + 831, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "11814", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "At test time, we employ only the point-level classification branch to segment 3D point clouds.", + "bbox": [ + 89, + 90, + 485, + 122 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experimental Evaluations", + "text_level": 1, + "bbox": [ + 89, + 136, + 334, + 155 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets: We evaluate our method on two widely used benchmarks, S3DIS [2] and ScanNetv2 [17], to assess both in-domain and cross-dataset FSCIL performance. These datasets incorporate large-scale coloured 3D point clouds of indoor spaces which introduce semantics. See Sup. Mat. for additional dataset details. For cross-dataset FSCIL, we create a transition dataset from ScanNetv2 to S3DIS, using ten classes from ScanNetv2 as the base and introducing eight new classes from S3DIS in incremental sessions. Following a strict disjoint setup as in [10], each session includes only novel classes. The number of introduced classes varies per session to assess resilience to catastrophic forgetting, which escalates with more incremental sessions in point cloud recognition [60].", + "bbox": [ + 88, + 162, + 483, + 375 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metrics: We assess model performance using three standard metrics for incremental learning[70]: Average Incremental Accuracy (AIA), which tracks average accuracy across tasks learned sequentially; Last, the accuracy after the final task; and Average Forgetting Rate $(\\mathcal{F}_{Last}^{(T)})$ , which measures performance decline on learned tasks after adapting to the last task. Finally, we report mean Intersection-over-Union $(\\mathrm{mIoU})^1$ to evaluate segmentation performance. See Sup. Mat. for detailed metric definitions. Architecture: Our backbone model employs PointTransformer [77], leveraging recent advances in point cloud recognition, with a Hyperbolic Classifier head $\\mathcal{H}(.)$ for segmentation. We train HiPo using the Riemannian SGD optimizer, facilitated by geoopt [29] for hyperbolic operations. Training is conducted on both $D^{(0)}$ and incremental datasets $D^{(t)}, t > 0$ with a batch size of 16, AdamW optimizer [37], initial learning rate of 0.001, and weight decay of 0.0001 over 100 epochs. Architecture diagram and further hyper-parameter details are available in the Sup. Mat.", + "bbox": [ + 89, + 375, + 483, + 665 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Competitors: We compare our method against various baselines representing distinct FSCIL approaches. For regularization-based methods, we include LwF with Teacher Adaptation [59], and for optimization-based methods, we include C-FSCIL [24]. We also evaluate against regularization-based FSCIL methods specifically designed for 3D point cloud tasks, such as LGKD [72], SDCOT [79], and 3DPC-CISS [70], which have shown effectiveness with the PointTransformer [77] architecture. Comparative results are provided in Tables 1 and 2. Additionally, we adapt BiDist [78], CLOM [85], TEEN [64], FACT [81], and ALICE [50] using the DGCNN [65] backbone for FSCIL in 3D point cloud segmentation. We report the results with the DGCNN [65] backbone in Table 5.", + "bbox": [ + 89, + 665, + 483, + 876 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/038eecf366649802e0a09bacdf7996f2bab9e045ce324bb210c550db31fd3e84.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 90, + 591, + 157 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bc810bb5e6235004291570e9db32336890bef6190cca36c1fa7e35e6be343cac.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 594, + 101, + 668, + 157 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/44ff4e68a9d72ff3fdb4160813e4730863929139856510e83ac99345f0c25498.jpg", + "image_caption": [ + "SDCOT" + ], + "image_footnote": [], + "bbox": [ + 671, + 101, + 746, + 157 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d7fc9e4b09a38c2727df04c1fe888dddbb912a2b587fad7ebc72836b2a6de59d.jpg", + "image_caption": [ + "3DPC-CISS" + ], + "image_footnote": [], + "bbox": [ + 751, + 101, + 825, + 157 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b4a33ea4b4a1916b2987e9fc716386469aa84c4a1d892fb72cbf413380368827.jpg", + "image_caption": [ + "HIPO (Ours)" + ], + "image_footnote": [], + "bbox": [ + 830, + 102, + 903, + 159 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2a70add8cc859c470ac47d259adb1ea9bca6c7cb6927eb57925fc7637af18590.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 162, + 589, + 220 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/66d2ace25701a2759e6ba19ce575d48a4d05f98a5f98cc608858b0f92cb07b5c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 594, + 164, + 668, + 220 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/689ae9ddddceac5aaac41b4ffc222299d136daaa8f43c0ef6121c9784d3e1511.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 674, + 162, + 746, + 220 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5ccdbabf1fceb3a3a0510900afd8dff4b950675bf7bfcdfe15c3973febd84392.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 753, + 164, + 823, + 220 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0d2f5fa6adbc042bddbb0ca62dc9d356fd172f98a6a53a316d878a581085b767.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 831, + 165, + 903, + 222 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7202c6e0419e65a485e5c5e8d5dd359cd6a5ef5f67e58bf8b93c5c637ac8d2c6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 224, + 589, + 282 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3cddb91826033ec2865b8ab68a286af7a82bd16390d1daa8bd396a2106736f56.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 594, + 224, + 668, + 282 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8438730be9ff05adc0e539f3f01f920f5da7ba88dc280bf51a999211acf7d799.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 674, + 224, + 746, + 282 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/94f742444df084e76f69c4bde7648aa71225bb834ceca96684ec01684f1374e1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 751, + 224, + 825, + 282 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/af36710e687a2e9e34fcb8e9532ec24155b01b7fb1c2240f5e3122e7a91e23f1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 831, + 224, + 903, + 282 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/906504720ac575679d1fb765e766e279b1ae7c0d78c833f9442d0b5ff82d6860.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 287, + 589, + 344 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f0b048e14c5365325ffc22e2185a71b48329abe1b96152e3aa7a1a0b03339530.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 594, + 287, + 668, + 344 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/76ffc8307d09e60aa534b7d7ac1a25336137a917f3fd0211a30befcfcb96ee22.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 674, + 287, + 746, + 344 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/76e131ab2723234f675ca50c055ace3bb135891c754cd08860dbec1ef118dad5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 751, + 287, + 825, + 344 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/091f65815f6411c69dfe721f0f41087fdc9561bade3887c56052204cb3ef4a74.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 831, + 287, + 903, + 344 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/74c8fd02f2dea6ca4287fad05266344d10aa971dc7ee7555ddcfe60b010567cc.jpg", + "image_caption": [ + "Figure 4. Qualitative visualization of model performance: Comparing FSCIL for 3D point cloud segmentation results of our method (HiPo) and other baselines for S3DIS dataset [2]. We observe that even with a few-shot setup, our model performs reasonably well in segmenting out small parts like walls, counters, etc., compared to other methods [70, 79]." + ], + "image_footnote": [], + "bbox": [ + 575, + 351, + 709, + 397 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5f6c35fe1cab7a75bf8ec7ac1ebd79287042afe875ce6f02e870616e329ae0cd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 720, + 351, + 833, + 396 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Discussing the main results", + "text_level": 1, + "bbox": [ + 511, + 508, + 756, + 525 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1 presents a detailed comparison of our proposed method, HiPo, against several baseline methods on the S3DIS [2] and ScanNetv2 [17] datasets across various evaluation settings. HiPo consistently outperforms the current state-of-the-art approach, 3DPC-CISS [70], achieving an average improvement of $+2.57\\%$ mIoU on S3DIS and $+2.12\\%$ on $\\mathcal{F}_{Last}^{(T)}$ . This improvement underscores the effectiveness of HiPo's hyperbolic, uncertainty-aware architecture, which mitigates catastrophic forgetting by reducing misalignment between features and prototype representations from earlier sessions. When compared to other baselines, such as LGKD [72] and C-FSCIL [24], HiPo shows substantial gains across all metrics. For instance, in the 10-3T scenario, HiPo achieves $33.40\\%$ mIoU on the Last metric, significantly higher than LGKD's $20.02\\%$ and C-FSCIL's $25.02\\%$ . This consistent superiority highlights HiPo's resilience against forgetting, particularly in incremental sessions with larger class increments, where other methods struggle to retain performance. Furthermore, while SDCOT [79] and LwF [59] suffer from high forgetting rates, HiPo effectively maintains stable performance across incremental tasks, with a $\\mathcal{F}_{Last}^{(T)}$ rate as low as $17.70\\%$ in challenging scenarios, compared to SDCOT's $30.83\\%$ . This robustness can be attributed to HiPo's hyperbolic embed", + "bbox": [ + 511, + 534, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "1We use accuracy and mIoU interchangeably.", + "bbox": [ + 107, + 886, + 346, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "11815", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c2378a8f86726d8baf0fc468e8cea80c492f477959f279185af4ae00526aeb0f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
4-3T7-3T10-3TAverage
Last(↑)AIA(↑)FT(Last)(↓)Last(↑)AIA(↑)FT(Last)(↓)Last(↑)AIA(↑)FT(Last)(↓)Last(↑)AIA(↑)FT(Last)(↓)
LwF [59]4.3322.0523.628.6021.8019.8021.3034.0525.5011.4125.3022.97
C-FSCIL [24]9.0627.9025.1120.0131.9217.8625.0235.3620.6818.0331.0621.88
LGKD [72]12.2431.8025.9419.7630.4115.9720.0232.7125.3817.3431.6422.43
SDCOT [79]5.3028.4230.8312.0029.4626.2023.8035.3023.0013.7031.0626.67
3DPC-CISS [70]25.7040.5019.7324.2034.5015.4529.6038.2017.2026.5037.7317.46
HtPo30.6043.8717.7029.4036.8011.1033.4040.2513.7031.1340.3014.16
", + "bbox": [ + 114, + 88, + 883, + 198 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/27ba110e83a4291187bd08365a383957bc5232603857e4cde8552030dfa6bf25.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
10-2T10-5T15-5TAverage
Last(↑)AIA(↑)FT(Last)(↓)Last(↑)AIA(↑)FT(Last)(↓)Last(↑)AIA(↑)FT(Last)(↓)Last(↑)AIA(↑)FT(Last)(↓)
LwF [59]1.7010.7510.861.0517.5324.725.0621.3832.642.6016.5522.74
C-FSCIL [24]7.1215.9410.5811.1224.1819.5911.5023.5824.159.2521.2318.11
LGKD [72]6.0215.6011.509.5525.5222.4511.1023.8325.468.2221.6519.80
SDCOT [79]0.509.5610.880.4017.8326.151.3019.5036.400.7315.6324.47
3DPC-CISS [70]6.2013.138.328.1022.3021.3012.2024.9525.508.8320.1218.37
HtPo8.5017.1310.3614.6527.2218.8511.9024.4525.1011.6822.2718.12
", + "bbox": [ + 114, + 200, + 883, + 310 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e41b3320d312f5056801e29df1a3648d38e8de24f103c8000715dd00f4106b0e.jpg", + "table_caption": [ + "Table 1. Comparison with literature on the S3DIS [2] (top table) and ScanNetv2 [17] (bottom table) datasets, using the PointTransformer backbone [77] in a 5-shot setting. CIL mIoU (%) is reported. “ $M-NT$ ” indicates that $M$ is the number of classes in the base session $S^{(0)}$ , while $N$ is the number of classes introduced in each incremental session $S^{(t)}$ , $t > 0$ . The highest performance in each column is shown in bold, with the second-highest results underlined. HiPO performs superior compared to other baselines in different incremental settings." + ], + "table_footnote": [], + "table_body": "
Average
Last (↑)AIA (↑)F(T)Last (↓)
LwF [59]8.7822.6023.94
C-FSCIL [24]13.5525.0219.72
LGKD [72]15.6325.5617.51
SDCOT [79]19.2229.2917.52
3DPC-CISS [70]19.9529.1616.01
HtPo20.6029.8116.05
", + "bbox": [ + 145, + 400, + 431, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. SOTA comparison on ScanNetv2 $\\rightarrow$ S3DIS dataset. We compute the average of 10-4T and 10-8T with PointTransformer backbone [77] in 5-shot setting. HIPO achieves superior Last and AIA values with comparable $F_{Last}^{(T)}$ values, highlighting better domain adaptability than other methods.", + "bbox": [ + 88, + 523, + 482, + 594 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ding space, which aligns naturally with hierarchical structures, thereby enhancing inter-class separation and intraclass compactness over time.", + "bbox": [ + 88, + 595, + 482, + 641 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Moreover, Fig. 4 provides qualitative comparisons, where HiPO demonstrates marked improvements over leading baselines, particularly in accurately segmenting challenging structures such as walls and counters. Further analyses and results are available in the Sup. Mat.", + "bbox": [ + 88, + 642, + 482, + 717 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Main ablation analysis", + "text_level": 1, + "bbox": [ + 89, + 727, + 302, + 744 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Cross-dataset class-incremental segmentation: To further assess the effectiveness of our proposed modules, we conduct cross-dataset incremental segmentation experiments (ScanNetv2 $\\rightarrow$ S3DIS). Detailed setup information is in Sup. Mat. As shown in Table 2, the cross-dataset setting introduces a notable increase in forgetting due to domain shift. However, the results underscore HiPo's robustness in managing such domain shifts effectively, with improvements of $0.65\\%$ in Last and $0.52\\%$ in AIA metrics.", + "bbox": [ + 88, + 750, + 482, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation studies on penalty in Busemann loss: We pro", + "bbox": [ + 89, + 885, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "vide a comprehensive ablation study on the regularization constant $\\phi$ introduced in Eq. 7. The regularization strength is varied from $\\phi = 0$ to $\\phi = 1$ in intervals of 0.25 for S3DIS [2] (4-3T, 7-3T, 10-3T) and ScanNetv2 [17] (10-2T, 10-5T, 15-5T). As shown in Table 3, our findings corroborate prior work [23]: without a penalty $(\\phi = 0)$ , samples are overly drawn towards the Ideal Prototypes, resulting in overconfident predictions and lower AIA on test data $\\mathcal{D}_{test}$ in addition to vanishing gradient problems [44]. Conversely, a high penalty $(\\phi = 1)$ increases inter-class confusion, also leading to lower AIA on $\\mathcal{D}_{test}$ . Our final chosen results, highlighted in bold in Table 3, reflect the optimal balance for $\\phi$ .", + "bbox": [ + 511, + 402, + 906, + 599 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison of metrics at different curvatures: Following [47], we experimented with curvature values of $c = 0.0001$ and $c = 0.001$ to implement traditional Euclidean architectures (Poincaré Ball with zero curvature; very small Poincaré ball curvature values are chosen as curvature $c = 0$ is numerically unstable [27]) for the S3DIS [2] (4-3T, 7-3T, 10-3T) and ScanNetv2 [17] (10-2T, 10-5T, 15-5T) setups. Table 4 shows that our HiPO with a Poincaré Ball curvature of $c = -1.00$ consistently outperforms its Euclidean counterpart across most settings. This finding demonstrates that Riemannian geometry leverages the hierarchical structure of point clouds more effectively than Euclidean space, suggesting further exploration in this domain.", + "bbox": [ + 511, + 604, + 908, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Sensitivity to the different number of training samples in each incremental episode: To assess the impact of data availability on model performance, we gradually increase the number of samples per class (shot) from one to all available examples and analyze the Last mIoU and $\\mathcal{F}_{Last}^{(T)}$ results for different shot values in Fig. 5. Our findings indi", + "bbox": [ + 511, + 806, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "11816", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a645460ce21a0458289ce725c7dd5b88b3af1fd7b8d375bc7afdf3010d3009a4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetsSettingsφ = 0.00φ = 0.25φ = 0.50φ = 0.75φ = 1.00
AIA↑FT(St)↓AIA↑FT(St)↓AIA↑FT(St)↓AIA↑FT(St)↓AIA↑FT(St)↓
S3DIS [2]4-3T42.8717.7043.4017.8643.2718.2343.8717.7043.2218.96
7-3T36.1311.6036.4611.2036.8011.1036.7011.1036.0612.10
10-3T39.9514.3040.2513.7040.1014.0040.0714.0540.0514.10
ScanNetv2 [17]10-2T16.7510.5116.8510.5016.9010.0817.1310.3616.9110.76
10-5T27.0818.8727.1318.8727.2218.8527.2118.8527.1618.92
15-5T24.2525.5024.4525.1024.0525.9023.9526.1024.0026.00
", + "bbox": [ + 202, + 88, + 802, + 196 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b9320cd99bb4b9295491e34471e9f9043ed38955314e79bcbdee2fdfba05e1b9.jpg", + "table_caption": [ + "Table 3. Different penalty values in Busemann loss in 5-shot incremental segmentation across the S3DIS [2] dataset and ScanNetV2 [17] dataset, highlighting the regularization benefits introduced by the hyperbolic distance of the projected class prototypes from the origin." + ], + "table_footnote": [], + "table_body": "
DatasetsSettingsc = -1.00 (Ours)c = 0.0001c = 0.001
AIA↑FTLast↓AIA↑FTLast↓AIA↑FTLast↓
S3DIS [2]4-3T43.8717.7042.9519.9343.2317.38
7-3T36.8011.1036.2311.6036.8611.05
10-3T40.2513.7040.2013.8040.2213.75
ScanNetv2 [17]10-2T17.1310.3616.5410.2516.9810.18
10-5T27.2218.8527.1619.0026.8319.92
15-5T24.4525.1024.4025.2024.2025.60
", + "bbox": [ + 94, + 258, + 498, + 359 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/71fd6899a574728d21bd83abf494721a899036eef8cec22063294d65748d6cb4.jpg", + "image_caption": [ + "(a) Last mIoU values" + ], + "image_footnote": [], + "bbox": [ + 109, + 453, + 302, + 588 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0797ba7c14ea84bc856ebcc361ddc407a88046c5c5edd3502ed5da5089f9942e.jpg", + "image_caption": [ + "(b) Forgetting $\\mathcal{F}_{\\text{Last}}^{(T)}$", + "Figure 5. mIoU improves and forgetting reduces with more training examples. Our proposed method outperforms the top-2 baseline methods ([70, 79]) in terms of Last mIoU and forgetting mitigation in the 10-2T setting (Full-data: fully supervised training in incremental fashion)." + ], + "image_footnote": [], + "bbox": [ + 310, + 457, + 478, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "cate that catastrophic forgetting is particularly severe with limited data; however, our method substantially reduces forgetting compared to other baselines. Notably, increasing the number of training examples significantly mitigates forgetting for H1Po, while it exacerbates forgetting for SDCOT.", + "bbox": [ + 89, + 703, + 482, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Experiments with a different backbone: Besides PointTransformer [77], we evaluate our model with the backbone of DGCNN [65]. Unlike current SOTA FSCIL methods—BiDist [78], CLOM [85], TEEN [64], FACT [81], and ALICE [50]—which perform well with DGCNN but degrade with PointTransformer, our method remains effective across both architectures. As shown in Table 5, our approach achieves consistent gains with DGCNN, outper", + "bbox": [ + 89, + 779, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6ed0464f457bed14881938bf17db917fc278df840a7c129cfc0e46e154ab5578.jpg", + "table_caption": [ + "Table 4. Comparison for different curvature values in 5-shot incremental segmentation across S3DIS [2] and ScanNetV2 [17] datasets. This highlights the superior stability-plasticity trade-off of hyperbolic embedding space with negative curvature, where small positive curvature approximates Euclidean space." + ], + "table_footnote": [], + "table_body": "
MethodLast (↑)AIA (↑)FTLast (↓)
LwF [59]9.4526.7524.41
C-FSCIL [24]20.2230.6422.15
LGKD [72]18.6329.1615.75
SDCOT [79]22.7524.5419.94
CLOM [85]21.1834.6718.83
TEEN [64]22.7133.9015.37
FACT [81]24.2726.7021.26
3DPC-CISS [70]22.0324.7419.36
BiDist [78]27.7039.1415.85
HiPo29.9039.8613.87
", + "bbox": [ + 589, + 258, + 831, + 385 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Performance comparison on the S3DIS dataset [2] using the DGCNN backbone [65] in the 5-shot setting. Reported metrics include Last, $\\mathcal{F}_{Last}^{(T)}$ , and AIA for 4-3T and 7-3T settings. HiPo outperforms all baseline methods with DGCNN.", + "bbox": [ + 511, + 393, + 906, + 453 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "forming baselines with $+2.20\\%$ mIoU on Last, $+0.72\\%$ on AIA, and reducing forgetting by $+1.98\\%$ on $\\mathcal{F}_{Last}^{(T)}$ . This demonstrates the model-agnostic nature of HiPO.", + "bbox": [ + 511, + 474, + 906, + 523 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions & Future Directions", + "text_level": 1, + "bbox": [ + 511, + 541, + 815, + 558 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce a previously unexplored problem setting along with a novel solution specifically designed for few-shot class-incremental semantic segmentation of 3D point cloud data, addressing a critical gap in the literature. Recognizing the unique challenges posed by point clouds, we identified limitations in current FSCIL methods and developed HiPo, a model that leverages hyperbolic space to embed point clouds while achieving an improved stability-plasticity trade-off. Our approach consistently outperforms state-of-the-art baselines, demonstrating that hyperbolic geometry's ability to represent hierarchical structures offers a more effective framework for FSCIL in 3D data. HiPo is particularly valuable for real-world applications like autonomous driving, industrial inspection, urban planning, and environmental monitoring, where generalization to unseen classes is essential. Furthermore, our approach enhances safety and efficiency in critical areas such as disaster response and infrastructure maintenance by enabling accurate, timely semantic segmentation of complex 3D environments. Future work will focus on managing possible outliers that may emerge without prior reference across incremental sessions, further strengthening our robustness.", + "bbox": [ + 511, + 568, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "11817", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Noor Ahmed, Anna Kukleva, and Bernt Schiele. Orco: Towards better generalization via orthogonality and contrast for few-shot class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28762-28771, 2024. 1, 3", + "[2] Iro Armeni, Ozan Sener, Amir R Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3d semantic parsing of large-scale indoor spaces. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1534-1543, 2016. 2, 4, 6, 7, 8", + "[3] Mina Ghadimi Atigh, Julian Schoep, Erman Acar, Nanne Van Noord, and Pascal Mettes. Hyperbolic image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4453-4462, 2022. 2", + "[4] Eden Belouadah and Adrian Popescu. II2m: Class incremental learning with dual memory. In Proceedings of the IEEE/CVF international conference on computer vision, pages 583-592, 2019. 1", + "[5] Abhijit Bendale and Terrance Boult. Towards open world recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1893-1902, 2015. 3", + "[6] Chris Burges, Tal Shaked, Erin Renshaw, Ari Lazier, Matt Deeds, Nicole Hamilton, and Greg Hullender. Learning to rank using gradient descent. In Proceedings of the 22nd International Conference on Machine Learning, page 89-96, New York, NY, USA, 2005. Association for Computing Machinery. 5", + "[7] Herbert Busemann. The geometry of geodesics. Courier Corporation, 2005. 2, 3, 5", + "[8] Francisco M Castro, Manuel J Marín-Jiménez, Nicolas Guil, Cordelia Schmid, and Karteek Alahari. End-to-end incremental learning. In Proceedings of the European conference on computer vision (ECCV), pages 233-248, 2018. 1", + "[9] Jun Cen, Peng Yun, Shiwei Zhang, Junhao Cai, Di Luan, Mingqian Tang, Ming Liu, and Michael Yu Wang. Openworld semantic segmentation for lidar point clouds. In European Conference on Computer Vision, pages 318-334. Springer, 2022. 3", + "[10] Fabio Cermelli, Massimiliano Mancini, Samuel Rota Bulo, Elisa Ricci, and Barbara Caputo. Modeling the background for incremental learning in semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9233-9242, 2020. 6", + "[11] Arslan Chaudhry, Puneet K Dokania, Thalaiyasingam Ajthan, and Philip HS Torr. Riemannian walk for incremental learning: Understanding forgetting and intransigence. In Proceedings of the European conference on computer vision (ECCV), pages 532-547, 2018. 1", + "[12] Bike Chen, Wei Peng, Xiaofeng Cao, and Juha Röning. Hyperbolic uncertainty aware semantic segmentation. IEEE Transactions on Intelligent Transportation Systems, 2023. 5", + "[13] Ali Cheraghian, Shafin Rahman, Pengfei Fang, Soumava Kumar Roy, Lars Petersson, and Mehrtash" + ], + "bbox": [ + 93, + 114, + 485, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Harandi. Semantic-aware knowledge distillation for few-shot class-incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2534-2543, 2021. 1", + "[14] Townim Chowdhury, Mahira Jalisha, Ali Cheraghian, and Shafin Rahman. Learning without forgetting for 3d point cloud objects. In Advances in Computational Intelligence: 16th International Work-Conference on Artificial Neural Networks, IWANN 2021, Virtual Event, June 16-18, 2021, Proceedings, Part I 16, pages 484-497. Springer, 2021. 3", + "[15] Townim Chowdhury, Ali Cheraghian, Sameera Ramasinghe, Sahar Ahmadi, Morteza Saberi, and Shafin Rahman. Few-shot class-incremental learning for 3d point cloud objects. In European Conference on Computer Vision, pages 204-220. Springer, 2022. 3", + "[16] Yawen Cui, Zitong Yu, Wei Peng, Qi Tian, and Li Liu. Rethinking few-shot class-incremental learning with open-set hypothesis in hyperbolic geometry. IEEE Transactions on Multimedia, 2023. 2, 3", + "[17] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 2, 6, 7, 8", + "[18] David L. Davies and Donald W. Bouldin. A cluster separation measure. IEEE Transactions on Pattern Analysis and Machine Intelligence, PAMI-1(2):224-227, 1979. 2", + "[19] Jiahua Dong, Yang Cong, Gan Sun, Bingtao Ma, and Lichen Wang. I3dol: Incremental 3d object learning without catastrophic forgetting. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6066-6074, 2021. 3", + "[20] DC Dowson and BV666017 Landau. The frechet distance between multivariate normal distributions. Journal of multivariate analysis, 12(3):450-455, 1982. 4", + "[21] Luca Franco, Paolo Mandica, Konstantinos Kallidromitis, Devin Guillory, Yu-Teng Li, Trevor Darrell, and Fabio Galasso. Hyperbolic active learning for semantic segmentation under domain shift. arXiv preprint arXiv:2306.11180, 2023.5", + "[22] Octavian-Eugen Ganea, Gary Bécigneul, and Thomas Hofmann. Hyperbolic convolutional neural networks. In Advances in Neural Information Processing Systems, pages 4851-4861, 2018. 4", + "[23] Mina Ghadimi Atigh, Martin Keller-Ressel, and Pascal Mettes. Hyperbolic busemann learning with ideal prototypes. Advances in Neural Information Processing Systems, 34:103-115, 2021. 3, 5, 7", + "[24] Michael Hersche, Geethan Karunaratne, Giovanni Cherubini, Luca Benini, Abu Sebastian, and Abbas Rahimi. Constrained few-shot class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9057-9067, 2022. 1, 3, 6, 7, 8", + "[25] Naiwen Hu, Haozhe Cheng, Yifan Xie, Pengcheng Shi, and Jihua Zhu. Hyperbolic image-and-pointcloud contrastive learning for 3d classification. arXiv preprint arXiv:2409.15810, 2024. 2, 3", + "[26] Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, and Andrew Markham." + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "11818", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Randla-net: Efficient semantic segmentation of large-scale point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11108-11117, 2020. 1, 2", + "[27] Valentin Khrulkov, Leyla Mirvakhabova, Evgeniya Ustinova, Ivan Oseledets, and Victor Lempitsky. Hyperbolic image embeddings. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6418-6428, 2020. 2, 3, 5, 7", + "[28] Do-Yeon Kim, Dong-Jun Han, Jun Seo, and Jaekyun Moon. Warping the space: Weight space rotation for class incremental few-shot learning. In The Eleventh International Conference on Learning Representations, 2023. 1", + "[29] Max Kochurov, Rasul Karimov, and Serge Kozlukov. Geoopt: Riemannian optimization in pytorch, 2020. 6", + "[30] Anna Kukleva, Hilde Kuehne, and Bernt Schiele. Generalized and incremental few-shot learning by explicit learning and calibration without forgetting. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9020-9029, 2021. 1", + "[31] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022. 5", + "[32] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40(12):2935-2947, 2017. 1", + "[33] Fangzhou Lin, Yun Yue, Songlin Hou, Xuechu Yu, Yajun Xu, Kazunori D Yamada, and Ziming Zhang. Hyperbolic chamfer distance for point cloud completion. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14595-14606, 2023. 1", + "[34] Pei Lin, Zheng Cai, Jiangtao Zhang, and Lifang Ruan. A survey on point cloud segmentation: From classical methods to deep learning. ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM), 16(3): 1-27, 2020. 1", + "[35] Yuyang Liu, Yang Cong, Gan Sun, Tao Zhang, Jiahua Dong, and Hongsen Liu. L3doc: Lifelong 3d object classification. IEEE Transactions on Image Processing, 30:7486-7498, 2021. 3", + "[36] Zan Liu, Zilong Zhang, Yiyi Lin, Lin Zhu, Han Jin, and Stephen Lin. Pointaugment: An auto-augmentation framework for point cloud classification. arXiv preprint arXiv:2002.10876, 2020. 1", + "[37] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6", + "[38] Paolo Mandica, Luca Franco, Konstantinos Kallidromitis, Suzanne Petryk, and Fabio Galasso. Hyperbolic learning with multimodal large language models. arXiv preprint arXiv:2408.05097, 2024. 2", + "[39] Pratik Mazumder, Pravendra Singh, and Piyush Rai. Few-shot lifelong learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2337-2345, 2021. 1", + "[40] Leland McInnes, John Healy, and James Melville. Umap: Uniform manifold approximation and projection for dimension reduction. arXiv preprint arXiv:1802.03426, 2018. 2" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Pascal Mettes, Elise Van der Pol, and Cees Snoek. Hyperspherical prototype networks. Advances in neural information processing systems, 32, 2019. 5", + "[42] Pascal Mettes, Mina Ghadimi Atigh, Martin Keller-Ressel, Jeffrey Gu, and Serena Yeung. Hyperbolic deep learning in computer vision: A survey. International Journal of Computer Vision, pages 1-25, 2024. 2", + "[43] Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. Distributed representations of words and phrases and their compositionality. Advances in neural information processing systems, 26, 2013. 5", + "[44] Gal Mishne, Zhengchao Wan, Yusu Wang, and Sheng Yang. The numerical stability of hyperbolic representation learning. In International Conference on Machine Learning, pages 24925-24949. PMLR, 2023. 2, 4, 5, 7", + "[45] Antonio Montanaro, Diego Valsesia, and Enrico Magli. Rethinking the compositionality of point clouds through regularization in the hyperbolic space. Advances in Neural Information Processing Systems, 35:33741-33753, 2022. 1, 2, 3, 4", + "[46] Antonio Montanaro, Diego Valsesia, and Enrico Magli. Towards hyperbolic regularizers for point cloud part segmentation. In ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5, 2023. 1", + "[47] Gabriel Moreira, Manuel Marques, João Paulo Costeira, and Alexander Hauptmann. Hyperbolic vs euclidean embeddings in few-shot learning: Two sides of the same coin. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2082-2090, 2024. 7", + "[48] Samrat Mukherjee, Tanuj Sur, Saurish Seksaria, Subhasis Chaudhuri, Gemma Roig, and Biplab Banerjee. Uidaple: Unsupervised incremental domain adaptation through adaptive prompt learning. In ICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5, 2025. 1", + "[49] Maximilian Nickel and Douwe Kiela. Poincaré embeddings for learning hierarchical representations. In Advances in neural information processing systems, pages 6338-6347, 2017. 2, 3", + "[50] Can Peng, Kun Zhao, Tianren Wang, Meng Li, and Brian C Lovell. Few-shot class-incremental learning from an open-set perspective. In European Conference on Computer Vision, pages 382-397. Springer, 2022. 3, 6, 8", + "[51] Jeffrey Pennington, Richard Socher, and Christopher Manning. GloVe: Global vectors for word representation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1532-1543, Doha, Qatar, 2014. Association for Computational Linguistics. 5", + "[52] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 1, 2", + "[53] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "11819", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "point sets in a metric space. Advances in neural information processing systems, 30, 2017. 1, 2", + "[54] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5", + "[55] Sachin Ravi and Hugo Larochelle. Optimization as a model for few-shot learning. arXiv preprint arXiv:1606.04474, 2017. 1", + "[56] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 2001–2010, 2017. 1", + "[57] Rik Sarkar. Low distortion delaunay embedding of trees in hyperbolic plane. In International symposium on graph drawing, pages 355-366. Springer, 2011. 2, 3", + "[58] Yisheng Song, Ting Wang, Puyu Cai, Subrota K Mondal, and Jyoti Prakash Sahoo. A comprehensive survey of few-shot learning: Evolution, applications, challenges, and opportunities. ACM Computing Surveys, 55(13s):1-40, 2023. 1", + "[59] Filip Szatkowski, Mateusz Pyla, Marcin Przewieżlikowski, Sebastian Cygert, Bartlomiej Twardowski, and Tomasz Trzeciński. Adapt your teacher: Improving knowledge distillation for exemplar-free continual learning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1977-1987, 2024. 6, 7, 8", + "[60] Yuwen Tan and Xiang Xiang. Cross-domain few-shot incremental learning for point-cloud recognition. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2307-2316, 2024. 6", + "[61] Xiaoyu Tao, Xiaopeng Hong, Xinyuan Chang, Songlin Dong, Xing Wei, and Yihong Gong. Few-shot class incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12183-12192, 2020. 3", + "[62] Jesper E Van Engelen and Holger H Hoos. A survey on semi-supervised learning. Machine learning, 109(2):373-440, 2020. 1", + "[63] Max van Spengler, Erwin Berkhout, and Pascal Mettes. Poincare resnet. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5419-5428, 2023. 4, 5", + "[64] Qi-Wei Wang, Da-Wei Zhou, Yi-Kai Zhang, De-Chuan Zhan, and Han-Jia Ye. Few-shot class-incremental learning via training-free prototype calibration. Advances in Neural Information Processing Systems, 36, 2024. 6, 8", + "[65] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (tog), 38(5):1-12, 2019. 1, 2, 6, 8", + "[66] Yaqing Wang, Quanming Yao, James T Kwok, and Lionel M Ni. Generalizing from a few examples: A survey on few-shot learning. ACM computing surveys (csur), 53(3):1-34, 2020. 1" + ], + "bbox": [ + 91, + 92, + 482, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[67] Wenxuan Wu, Zhongang Qi, and Li Fuxin. Pointconv: Deep convolutional networks on 3d point clouds. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 9621-9630, 2019. 1, 2", + "[68] Yating Xu, Conghui Hu, Na Zhao, and Gim Hee Lee. Generalized few-shot point cloud segmentation via geometric words. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21506-21515, 2023. 1", + "[69] Xiangli Yang, Zixing Song, Irwin King, and Zenglin Xu. A survey on deep semi-supervised learning. IEEE Transactions on Knowledge and Data Engineering, 35(9):8934-8954, 2022. 1", + "[70] Yuwei Yang, Munawar Hayat, Zhao Jin, Chao Ren, and Yinjie Lei. Geometry and uncertainty-aware 3d point cloud class-incremental semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21759-21768, 2023. 6, 7, 8", + "[71] Zhilu Yang, Ying Zheng, and Jun Zhang. Learning to learn from labeled and unlabeled data with point-wise mutual information. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5704-5711, 2019. 1", + "[72] Ze Yang, Ruibo Li, Evan Ling, Chi Zhang, Yiming Wang, Dezhao Huang, Keng Teck Ma, Minhoe Hur, and Guosheng Lin. Label-guided knowledge distillation for continual semantic segmentation on 2d images and 3d point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18601-18612, 2023. 1, 3, 6, 7, 8", + "[73] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. arXiv preprint arXiv:2111.07783, 2021. 5", + "[74] Yun Yue, Fangzhou Lin, Guanyi Mou, and Ziming Zhang. Understanding hyperbolic metric learning through hard negative sampling. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1891-1903, 2024. 2", + "[75] Maciej Zamorski, Michal Stypulkowski, Konrad Karanowski, Tomasz Trzeciński, and Maciej Zieba. Continual learning on 3d point clouds with random compressed rehearsal. Computer Vision and Image Understanding, 228: 103621, 2023. 3", + "[76] Chi Zhang, Nan Song, Guosheng Lin, Yun Zheng, Pan Pan, and Yinghui Xu. Few-shot incremental learning with continually evolved classifiers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12455-12464, 2021. 2, 3", + "[77] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16259-16268, 2021. 2, 6, 7, 8", + "[78] Linglan Zhao, Jing Lu, Yunlu Xu, Zhanzhan Cheng, Dashan Guo, Yi Niu, and Xiangzhong Fang. Few-shot class incremental learning via class-aware bilateral distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11838-11847, 2023. 1, 3, 6, 8", + "[79] Na Zhao and Gim Hee Lee. Static-dynamic co-teaching for class-incremental 3d object detection. In Proceedings of" + ], + "bbox": [ + 516, + 92, + 906, + 901 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11820", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the AAAI Conference on Artificial Intelligence, pages 3436-3445, 2022. 1, 3, 6, 7, 8", + "[80] Na Zhao, Tat-Seng Chua, and Gim Hee Lee. Few-shot 3d point cloud semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8873-8882, 2021. 2", + "[81] Da-Wei Zhou, Fu-Yun Wang, Han-Jia Ye, Liang Ma, Shiliang Pu, and De-Chuan Zhan. Forward compatible few-shot class-incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9046–9056, 2022. 1, 2, 3, 6, 8", + "[82] Da-Wei Zhou, Zi-Wen Cai, Han-Jia Ye, De-Chuan Zhan, and Ziwei Liu. Revisiting class-incremental learning with pretrained models: Generalizability and adaptivity are all you need. arXiv preprint arXiv:2303.07338, 2023. 2", + "[83] Da-Wei Zhou, Qi-Wei Wang, Zhi-Hong Qi, Han-Jia Ye, De-Chuan Zhan, and Ziwei Liu. Class incremental learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2", + "[84] Yudong Zhu, Di Zhou, Jinghui Xiao, Xin Jiang, Xiao Chen, and Qun Liu. Hypertext: Endowing fasttext with hyperbolic geometry. arXiv preprint arXiv:2010.16143, 2020. 2", + "[85] Yixiong Zou, Shanghang Zhang, Yuhua Li, and Ruixuan Li. Margin-based few-shot class-incremental learning with class-level overfitting mitigation. Advances in neural information processing systems, 35:27267-27279, 2022. 1, 3, 6, 8" + ], + "bbox": [ + 91, + 90, + 483, + 470 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "11821", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/Hyperbolic Uncertainty-Aware Few-Shot Incremental Point Cloud Segmentation/1247e066-a517-4ffe-b97e-a9bff43b89c3_model.json b/2025/Hyperbolic Uncertainty-Aware Few-Shot Incremental Point Cloud Segmentation/1247e066-a517-4ffe-b97e-a9bff43b89c3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..db46161459bee9209831b365a4e702b394249628 --- /dev/null +++ b/2025/Hyperbolic Uncertainty-Aware Few-Shot Incremental Point Cloud Segmentation/1247e066-a517-4ffe-b97e-a9bff43b89c3_model.json @@ -0,0 +1,2831 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.094, + 0.131, + 0.905, + 0.153 + ], + "angle": 0, + "content": "Hyperbolic Uncertainty-Aware Few-Shot Incremental Point Cloud Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.18, + 0.885, + 0.217 + ], + "angle": 0, + "content": "Tanuj Sur† Samrat Mukherjee* Kaizer Rahaman‡ Subbasis Chaudhuri* Muhammad Haris Khan§ Biplab Banerjee*" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.218, + 0.91, + 0.253 + ], + "angle": 0, + "content": "†Chennai Mathematical Institute *Indian Institute of Technology Bombay ‡Indian Institute of Technology Kharagpur §Mohamed Bin Zayed University of Artificial Intelligence" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.287, + 0.327, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.319, + 0.486, + 0.683 + ], + "angle": 0, + "content": "3D point cloud segmentation is essential across a range of applications; however, conventional methods often struggle in evolving environments, particularly when tasked with identifying novel categories under limited supervision. Few-Shot Learning (FSL) and Class Incremental Learning (CIL) have been adapted previously to address these challenges in isolation, yet the combined paradigm of Few-Shot Class Incremental Learning (FSCIL) remains largely unexplored for point cloud segmentation. To address this gap, we introduce Hyperbolic Ideal Prototypes Optimization (HiPO), a novel framework that harnesses hyperbolic embeddings for FSCIL in 3D point clouds. HiPO employs the Poincaré Hyperbolic Sphere as its embedding space, integrating Ideal Prototypes enriched by CLIP-derived class semantics, to capture the hierarchical structure of 3D data. By enforcing orthogonality among prototypes and maximizing representational margins, HiPO constructs a resilient embedding space that mitigates forgetting and enables the seamless integration of new classes, thereby effectively countering overfitting. Extensive evaluations on S3DIS, ScanNetv2, and cross-dataset scenarios demonstrate HiPO's strong performance, significantly surpassing existing approaches in both in-domain and cross-dataset FSCIL tasks for 3D point cloud segmentation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.705, + 0.222, + 0.72 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.73, + 0.485, + 0.866 + ], + "angle": 0, + "content": "Recent advancements in deep learning have spurred progress in data-efficient learning, with semi-supervised learning (SSL) [62, 69, 71] and few-shot learning (FSL) [55, 58, 66] emerging as key approaches. SSL utilizes both labeled and unlabeled data to enhance learning, while FSL enables generalization from only a few labeled samples per class. However, both approaches face challenges in real-world, dynamic environments where models must continuously adapt to new classes while retaining prior knowledge" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.284, + 0.905, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.457, + 0.908, + 0.57 + ], + "angle": 0, + "content": "Figure 1. Our approach tackles the novel challenge of few-shot class incremental 3D point cloud segmentation. In each session \\( S^{(t)}, t \\geq 0 \\), our model takes an RGB point cloud of an indoor scene \\( D^{(t)} \\) as input and performs semantic segmentation. During the base training session \\( S^{(0)} \\), sufficient examples are available for each class. In later sessions, only a limited number of labeled point cloud examples are provided for new classes, introducing data imbalance and potential overfitting and forgetting risks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.907, + 0.674 + ], + "angle": 0, + "content": "edge—an essential requirement for continual learning (CL) [1, 4, 8, 11, 32, 48, 56]. Few-Shot Class Incremental Learning (FSCIL) [1, 13, 24, 28, 30, 39] combines the data efficiency of FSL with the adaptability of CL, yet remains underdeveloped for complex tasks such as segmentation in unstructured 3D data (see Fig. 1)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.885 + ], + "angle": 0, + "content": "Point cloud segmentation is pivotal in applications like autonomous driving and robotics, where it enables precise object recognition and scene understanding [36, 52, 68]. This task assigns semantic labels to each point within a 3D structure, facilitating critical decision-making processes. Current supervised segmentation methods [26, 53, 65, 67] are highly reliant on extensive labeled datasets, which are costly and time-intensive to produce [34]. These methods also assume a fixed set of labels, limiting their adaptability in scenarios where new classes emerge unexpectedly. While FSCIL offers a potential solution, existing methods [1, 24, 72, 78, 79, 81, 85] struggle to represent the complex, hierarchical structures typical of point cloud data due to the limitations of Euclidean embeddings [33, 45, 46]." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Capturing hierarchical relationships, such as recogniz" + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.875, + 0.483, + 0.901 + ], + "angle": 0, + "content": "*Samrat Mukherjee acknowledges the support of Prime Minister Research Fellowship (PMRF)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11810" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.258 + ], + "angle": 0, + "content": "ing subcomponents within an object, demands a framework that can accommodate intricate interrelations. This challenge is intensified in Class Incremental Learning (CIL) [82, 83], where the model must incorporate new classes while preserving previously learned representations—a balance between adaptability and memory known as the stability-plasticity trade-off. In FSCIL, especially under data scarcity, this trade-off is often compromised, leading to overfitting and catastrophic forgetting. Euclidean spaces exacerbate these challenges, as they lack the capacity to model evolving data scenarios [76, 81]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.261, + 0.483, + 0.534 + ], + "angle": 0, + "content": "Hyperbolic geometry [25, 38, 42, 57, 74] provides a more suitable framework for hierarchical data [45, 84]. Its negative curvature and exponential volume growth allow it to mirror tree-like data structures, addressing the stability-plasticity trade-off [16] and supporting segmentation in unstructured contexts [27]. Unlike Euclidean spaces, hyperbolic space offers exponentially more volume [44] for accommodating incremental classes, thus reducing overlap and interference (see Fig. 2, Tab. 4). However, efficient optimization of embeddings in the hyperbolic space presents computational challenges when points are near the boundary of the Poincaré ball resulting in vanishing gradient problems [44], and without careful handling, newly added classes can still encroach upon the existing embeddings, aggravating plasticity (see Tab. 3). This highlights the need for an FSCIL framework that uses hyperbolic space embedding hierarchical data in low distortion while ensuring computational efficiency and managing class overlap." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.536, + 0.483, + 0.718 + ], + "angle": 0, + "content": "Our Proposal: To address these challenges, we introduce Hyperbolic Ideal Prototypes Optimization (HiPo), a novel FSCIL framework tailored for point cloud segmentation. By embedding data within the Poincaré Hyperbolic Sphere [49], HiPo leverages the unique properties of hyperbolic space to capture hierarchical structures and maintain consistent feature-classifier alignment across incremental learning sessions. Through a modified Busemann loss [7], HiPo arranges ideal prototypes orthogonally within the hyperbolic space, maximizing class separation and preserving space for new classes, which minimizes inter-class interference and optimizes the feature space for CIL." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.902 + ], + "angle": 0, + "content": "This design takes advantage of hyperbolic space's exponential scaling, which aligns naturally with complex hierarchies found in point cloud data [3, 27, 45, 57]. Our modified Busemann loss incorporates two components: (1) a prototype alignment term to guide class embeddings towards their respective Hyperbolic Ideal Prototypes, enhancing class separability, and intra-class compactness, and (2) an uncertainty-aware term that penalizes overfitting on novel classes and preventing vanishing gradient problems. Together, these components enable HiPo to effectively address data scarcity, mitigate catastrophic forgetting, and support robust FSCIL under episodic shifts, establishing it" + }, + { + "type": "image", + "bbox": [ + 0.54, + 0.094, + 0.875, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.252, + 0.907, + 0.364 + ], + "angle": 0, + "content": "Figure 2. Motivation for hyperbolic embeddings: (a) Hyperbolic UMAP [40] of the HiPo embedding space, projected to 2D for the S3DIS dataset [2]. The hyperbolic space's negative curvature aligns with hierarchical data, enabling a more compact representation and clearer class separation than Euclidean space (Dunn index [18] \\((\\uparrow)\\) for HiPo is 0.69 while for the Euclidean space, it is 0.35), showcasing better class compactness in our case. (b) UMAP of the classifier space in the Euclidean hypersphere" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.385, + 0.907, + 0.414 + ], + "angle": 0, + "content": "as a powerful solution for incremental learning in 3D point cloud segmentation. Overall, our contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.416, + 0.905, + 0.46 + ], + "angle": 0, + "content": "- We define and address the previously unexplored problem of FSCIL for 3D point cloud segmentation, encompassing both in-domain and cross-dataset scenarios." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.462, + 0.905, + 0.523 + ], + "angle": 0, + "content": "- We introduce hyperbolic space as a robust embedding framework for FSCIL in 3D segmentation, effectively capturing hierarchical structures and preserving stable feature-class relationships across incremental learning sessions." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.525, + 0.905, + 0.584 + ], + "angle": 0, + "content": "- We propose a modified Busemann loss with prototype alignment and uncertainty-aware components to ensure optimal feature-classifier alignment with minimized uncertainty in predictions." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.586, + 0.905, + 0.645 + ], + "angle": 0, + "content": "- Through extensive experiments on benchmark datasets S3DIS [2] and ScanNetv2 [17], HiPo demonstrates superior performance and generalization, setting a new standard for FSCIL in this domain." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.416, + 0.905, + 0.645 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.663, + 0.664, + 0.679 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.901 + ], + "angle": 0, + "content": "3D Point-Cloud Semantic Segmentation: Fully supervised 3D semantic segmentation has been a well-explored area, with methods typically requiring dense point-wise labels [26, 52, 53, 65, 67, 77]. A significant advancement was PointNet [52], which directly processes raw point clouds, removing the need for voxelization but facing challenges in capturing local structures. This limitation spurred improvements like PointNet++ [53], which introduced a hierarchical approach to enhance local feature learning, and DGCNN [65], which incorporated the EdgeConv module to integrate neighborhood information for better geometric representation. More recently, models such as PointTransformer [80] have integrated self-attention mechanisms, inspired by NLP, to elevate feature expressiveness and contextual awareness." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "11811" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.152 + ], + "angle": 0, + "content": "Despite these advancements, the dependency on large labeled datasets and the assumption of a fixed class set restrict these methods in practical scenarios where new classes must be incrementally learned with minimal supervision." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.153, + 0.484, + 0.47 + ], + "angle": 0, + "content": "(Few-Shot) Class Incremental Learning: Incremental learning methods for 3D data have predominantly focused on classification tasks rather than segmentation [14, 15, 19, 35, 75]. Approaches such as I3DOL [19] and L3DOC [35] mitigate catastrophic forgetting in 3D object classification through memory and attention mechanisms, yet they struggle as labeled supervision diminishes over time. Few-Shot Class Incremental Learning (FSCIL) [61] has emerged as a more flexible solution, enabling models to learn new classes incrementally with minimal labeled data. Representative methods like CEC [76] use graph-based task linking, while FACT [81] and ALICE [50] create adaptable feature spaces through clustering and cosine similarity. However, these techniques rely on Euclidean embeddings, which have limited capability to represent hierarchical structures intrinsic to 3D data, and frequent classifier updates introduce computational overhead [76]. Open-world semantic segmentation (OWSS) [5, 9] is a related area that focuses on dynamically detecting unknown classes at inference, though our focus on incrementally adding known classes during training sets our approach apart from OWSS." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.471, + 0.483, + 0.682 + ], + "angle": 0, + "content": "Hyperbolic Representations for Hierarchical Data: Hyperbolic geometry has shown promise for hierarchical data representation, with foundational work by Sarkar et al. [57] and Nickel et al. [49] introducing hyperbolic neural networks and Riemannian optimization methods. Hyperbolic embeddings have demonstrated substantial benefits for capturing semantic hierarchies across applications, including image embeddings [27] and FSCIL for image classification [16]. In 3D data, recent research has shown that hyperbolic representations can model complex part-whole relationships effectively. Montanaro et al. [45] applied hyperbolic regularizers to encode hierarchical structures in 3D objects, while Hu et al. [25] leveraged hyperbolic geometry to represent latent structures in point clouds." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.684, + 0.484, + 0.745 + ], + "angle": 0, + "content": "Our approach leverages hyperbolic embeddings to tackle FSCIL challenges in segmentation, improving adaptability, preserving hierarchies, and ensuring computational efficiency in the point cloud domain." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.761, + 0.411, + 0.779 + ], + "angle": 0, + "content": "3. Problem Definition & Methodology" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.788, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Let \\(D\\) represent a sequence of training datasets \\(\\{D^{(0)}, D^{(1)}, \\ldots, D^{(T)}\\}\\), where each \\(D^{(t)} = \\{(P_k^{(t)}, M_k^{(t)})_{k=1}^{|D^{(t)}|}\\}\\) corresponds to a training session \\(S^{(t)}\\), \\(t \\in \\{0, 1, \\ldots, T\\}\\). The initial session dataset \\(D^{(0)}\\) serves as the base, while \\(T\\) denotes the total number of incremental sessions. In each session \\(S^{(t)}\\), \\(P_k^{(t)}\\) represents a point cloud sample, and \\(M_k^{(t)}\\) provides its labels within" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.089, + 0.905, + 0.123 + ], + "angle": 0, + "content": "the label space \\(C^{(t)} = \\{c_1^{(t)},c_2^{(t)},\\ldots ,c_{|C^{(t)}|}^{(t)}\\}\\) where \\(c_{i}^{(t)}\\) is the \\(i\\)-th class in the incremental session \\(t\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.124, + 0.905, + 0.292 + ], + "angle": 0, + "content": "The base session \\( S^{(0)} \\) typically contains an abundant number of samples per class in the label space \\( C^{(0)} \\). However, in each subsequent incremental session \\( S^{(t)}, t > 0 \\), only a limited number of labeled point clouds \\( P_{k}^{(t)} \\in \\mathbb{R}^{m \\times d_{0}} \\) are provided, where \\( m \\) is the number of points per sample and \\( d_{0} \\) is the feature dimension. Following previous works [1, 24, 72, 78, 79, 81, 85], we assume \\( |D^{(t)}| = N \\times K \\), where \\( N \\) is the number of classes, and \\( K \\) is the number of samples per class, following an \\( N \\)-way \\( K \\)-shot setting \\( \\forall t > 0 \\). We assume \\( C^{(t)} \\) to be disjoint from previous sessions, i.e., \\( C^{(t)} \\cap C^{(t')} = \\emptyset \\), \\( \\forall t' < t \\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.293, + 0.905, + 0.399 + ], + "angle": 0, + "content": "During each incremental session, only the current dataset \\( D^{(t)} \\) is accessible, with data from previous sessions unavailable for training. For evaluation in session \\( t \\), the model is tested across all classes encountered up to that session, covering the label space \\( \\bigcup_{i=0}^{t} C^{(i)} \\). A solution to the FSCIL problem must jointly address uncharted catastrophic forgetting, data imbalance, and class overfitting judiciously." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.406, + 0.655, + 0.42 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.428, + 0.905, + 0.518 + ], + "angle": 0, + "content": "This section mentions the essential backgrounds needed for the proposed framework Hyperbolic Ideal Prototypes Optimization (HiPo). We extensively use the concepts of the Poincaré ball model and its associated Möbius addition, exponential map, and the geodesic distance. We detail their definition and discuss their formulae in the Sup. Mat." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.519, + 0.906, + 0.641 + ], + "angle": 0, + "content": "In the Poincaré ball model of hyperbolic geometry, all the points at the boundary of the Poincaré ball represent points at infinity. We initialize a group of classifier prototypes at the boundary of the Poincaré ball termed Ideal Prototypes, following [23], for each of the \\(n\\) individual classes. Assuming \\(u_{1}, u_{2}, \\dots, u_{n}\\) to be the Ideal Prototypes for \\(n\\) classes, they form the boundary of the \\(n\\)-dimensional Poincaré ball of curvature \\(c\\), \\(\\mathbb{B}_c^n\\), following [23]:" + }, + { + "type": "equation", + "bbox": [ + 0.566, + 0.653, + 0.905, + 0.671 + ], + "angle": 0, + "content": "\\[\n\\mathcal {I} _ {n} = \\left\\{u \\in \\mathbb {R} ^ {n}: u _ {1} ^ {2} + u _ {2} ^ {2} + \\dots + u _ {n} ^ {2} = 1 \\right\\} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.906, + 0.735 + ], + "angle": 0, + "content": "Since Ideal Prototypes are points at an infinite geodesic distance from all other points in \\(\\mathbb{B}_c^n\\), we use the Busemann function [7] as a distance measure since computing direct infinite geodesic distance is non-trivial [23]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.905, + 0.811 + ], + "angle": 0, + "content": "The pre-assigned Ideal Prototypes form an optimal geometric structure by fixing the target throughout the incremental learning. However, as already pointed out, optimizing Ideal Prototypes in the Poincaré model to address FS-CIL in 3D point cloud segmentation remains unexplored." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.819, + 0.679, + 0.835 + ], + "angle": 0, + "content": "3.2. Discussing HiPo" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Our model introduces a fixed alignment strategy for feature-classifier relationships in FSCIL, tailored for the hierarchical nature of data such as 3D point clouds. Traditional Euclidean embeddings suffer from compaction [50, 76, 81]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11812" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.096, + 0.086, + 0.337, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.34, + 0.087, + 0.582, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.584, + 0.087, + 0.899, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.296, + 0.907, + 0.368 + ], + "angle": 0, + "content": "Figure 3. HiPo Overview: We show the changes in the Poincaré ball during the initial three sessions in the 3-2T FSCIL setting (three base classes, two incremental classes per session) for the S3DIS dataset [2], designed to easily visualize the working of HiPo. Hyperbolic Ideal Prototypes are optimally positioned to increase the margin of semantically similar classes \\((\\mathcal{L}_{\\mathrm{HIP}})\\). As an example, the chair class gets positioned farther from sofa than from table or clutter. We align representations of the newly introduced classes with their corresponding ideal prototypes \\((\\mathcal{L}_{\\mathrm{Alignment}})\\) regularized by a modified Busemann function \\((\\mathcal{L}_{\\mathrm{Uncertainty}})\\). Training is conducted using \\(\\mathcal{L}_{\\mathrm{HiPo}}\\) across sessions." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.38, + 0.483, + 0.621 + ], + "angle": 0, + "content": "and distortion as new classes are added, leading to overlap and forgetting. Hyperbolic space, with its negative curvature and exponential volume, provides a natural fit for preserving hierarchical structures while minimizing inter-class interference. To address catastrophic forgetting, our model anchors feature-classifier relationships within hyperbolic space, ensuring stable alignment between features and classifier prototypes. This fixed alignment prevents the drift of previously learned embeddings as new classes are introduced, preserving the original classifier prototype positions and maintaining consistency across incremental sessions. By leveraging hyperbolic geometry, our approach enables efficient class separation, robust memory retention, and adaptability, making it well-suited for FSCIL in high-dimensional, unstructured data settings. Our method has two main stages:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.625, + 0.483, + 0.732 + ], + "angle": 0, + "content": "- Defining Ideal Prototypes: We initialize ideal prototypes for each class on the boundary of \\(\\mathbb{B}_c^n\\), positioning them mutually orthogonal and at maximal distances to enhance discriminative power, especially among classes with similar characteristics (see Sec. 3.2.1). This fixed configuration allows a dedicated space for each class, preventing the embedding drift observed in Euclidean spaces." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.902 + ], + "angle": 0, + "content": "- Efficient Projections onto Poincaré ball \\(\\mathbb{B}_c^n\\): In Hyperbolic embedding space, mean class features are obtained by computing Fréchet [20] means over the feature embedding for the training examples belonging to the classes. Instead of directly computing class features on the Poincaré ball—an intensive task[63]—we obtain the feature vectors in Euclidean space for every class in each session and then project them into hyperbolic space using exponential map \\(\\exp_0(.)\\) [45] (see Sec. 3.2.2). The hyperbolic classifier head \\(\\mathcal{H}(.)\\) consisting of a Möbius layer \\(\\mathcal{M}\\) (also known as hyperbolic feed forward layer [22]), takes the output of the" + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.625, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.38, + 0.905, + 0.425 + ], + "angle": 0, + "content": "exponential layer as input and projects mean class features to \\(\\mathbb{B}_c^n\\). The Poincaré ball addresses few-shot learning using the sample mean prototype as the class representation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.426, + 0.905, + 0.47 + ], + "angle": 0, + "content": "Our framework consists of a point-level classification branch, combining a feature extractor \\(\\mathcal{E}(.)\\) and hyperbolic classifier head \\(\\mathcal{H}(.)\\), and is trained with two key strategies:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.47, + 0.906, + 0.697 + ], + "angle": 0, + "content": "Feature-Classifier Alignment: For each session \\(t \\geq 0\\), we align the hyperbolic sample mean features (see Eqn. 4) with the Hyperbolic Ideal Prototypes of their respective classes using alignment loss, \\(\\mathcal{L}_{\\mathrm{Alignment}}\\). We define \\(\\mathcal{L}_{\\mathrm{Alignment}}\\) using Busemann function \\(\\mathcal{B}\\) and introduce a regularization term, quantifying uncertainty of the class features, to prevent feature collapse around the ideal prototypes, effectively reducing interference with previously seen classes (see Sec. 3.2.3) and vanishing gradient problems [44]. Point-Level Classification: The hyperbolic classifier \\(\\mathcal{H}(.)\\), consisting of a Möbius layer \\(\\mathcal{M}\\) followed by softmax [45], processes the output of \\(\\exp_0(\\mathcal{E}(.))\\), with a cross-entropy loss for 3D point cloud segmentation. The joint objective combines this alignment loss \\(\\mathcal{L}_{\\mathrm{Alignment}}\\) and cross-entropy loss \\(\\mathcal{L}_{\\mathrm{CE}}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.624, + 0.705, + 0.905, + 0.722 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {H I P O}} = \\mathcal {L} _ {\\text {A l i g n m e n t}} + \\mathcal {L} _ {\\mathrm {C E}} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.729, + 0.906, + 0.849 + ], + "angle": 0, + "content": "This approach ensures robust feature-classifier alignment across incremental learning sessions, addressing catastrophic forgetting in 3D point cloud segmentation by preserving both old and new class representations in a rehearsal-free manner. The process is illustrated in Fig. 3, and Algorithm 1 and a table summarizing the important variables used, mentioned in the Sup. Mat., supplement these discussions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.852, + 0.836, + 0.868 + ], + "angle": 0, + "content": "3.2.1. Hyperbolic ideal prototypes positioning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Building upon the goal of maintaining stable feature-classifier alignment, our framework pre-assigns classifier" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11813" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.426 + ], + "angle": 0, + "content": "prototypes on the boundary of \\(\\mathbb{B}_c^n\\) to define a structured classifier space prior to backbone training. To achieve maximal margin separation between classes, we introduce inductive biases by assigning each class a Hyperbolic Ideal Prototype, distributing these prototypes as uniformly as possible across the Poincaré ball. Inspired by Mettes et al. [41], we incorporate semantic class knowledge to enforce distinct separations, particularly for closely related classes, such as \"chair\" and \"sofa\". To further enhance semantic cohesion, we utilize the CLIP framework [54] due to its robust semantic representations [31, 73] in comparison to embeddings like word2vec [43], Glove [51] etc. By querying CLIP's frozen text encoder with the prompt \"A photo of a [class]\", we obtain text embeddings \\(S = \\{s_1,\\dots ,s_n\\}\\) for each of the \\(n\\) predefined classes. Leveraging the homeomorphism between hyperbolic space \\(\\mathbb{B}_c^n\\) and the hypersphere \\(\\mathbb{S}^n\\) [23], we maximize separation between nearest neighboring embeddings in \\(\\mathbb{S}^n\\) and enforce mutual orthogonality among Ideal Prototypes to mitigate conflict between new and prior knowledge. To achieve this, inspired by RankNet [6], we formulate a Hyperbolic positioning loss, \\(\\mathcal{L}_{\\mathrm{HIP}}\\), optimizing the ideal prototypes positions as:" + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.433, + 0.482, + 0.493 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal{L}_{\\text{HIP}} = \\frac{1}{|M|}\\sum_{(i,i^{\\prime},i^{\\prime \\prime})\\in M} - \\bar{W}_{ii^{\\prime}i^{\\prime \\prime}}\\log W_{ii^{\\prime}i^{\\prime \\prime}} - \\\\ (1 - \\bar {W} _ {i i ^ {\\prime} i ^ {\\prime \\prime}}) \\log (1 - W _ {i i ^ {\\prime} i ^ {\\prime \\prime}}) \\tag {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.498, + 0.484, + 0.586 + ], + "angle": 0, + "content": "where \\(M\\) represents all class triplets, and \\(\\bar{W}_{ii^{\\prime}i^{\\prime \\prime}} = \\llbracket \\cos \\theta_{\\mathbf{s}_i,\\mathbf{s}_{i^{\\prime}}} \\geq \\cos \\theta_{\\mathbf{s}_i,\\mathbf{s}_{i^{\\prime \\prime}}} \\rrbracket\\) denotes triplet ranking order, with \\(W_{ii^{\\prime}i^{\\prime \\prime}} = \\frac{e^{o_{ii^{\\prime}}i^{\\prime \\prime}}}{1 + e^{o_{ii^{\\prime}}i^{\\prime \\prime}}} \\text{ and } o_{ii^{\\prime}i^{\\prime \\prime}} = \\cos \\theta_{\\mathbf{u}_i,\\mathbf{u}_{i^{\\prime}}} - \\cos \\theta_{\\mathbf{u}_i,\\mathbf{u}_{i^{\\prime \\prime}}} \\), where \\(\\mathbf{u}_i, \\mathbf{u}_{i^{\\prime}}, \\mathbf{u}_{i^{\\prime \\prime}}\\) are ideal prototypes for the corresponding classes \\(i, i^{\\prime}\\) and \\(i^{\\prime \\prime}\\) respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.587, + 0.484, + 0.722 + ], + "angle": 0, + "content": "Following Mettes et al. [41], \\(\\mathcal{L}_{\\mathrm{HIP}}\\) optimizes the prototypes' placement on \\(\\mathbb{B}_c^n\\) by aligning the hyperbolic prototypes with semantic ranking. This optimization establishes a classifier structure that facilitates inter-class separability and mutual orthogonality. Hyperbolic space's exponentially larger embedding capacity [44] further ensures that new classes integrate seamlessly without distorting existing embeddings, promoting robust feature organization across learning episodes." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.728, + 0.4, + 0.743 + ], + "angle": 0, + "content": "3.2.2. Hyperbolic class prototype projection" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.746, + 0.483, + 0.791 + ], + "angle": 0, + "content": "We model the hierarchical complexity in 3D point clouds using a hyperbolic manifold representation, which preserves nuanced class relationships with minimal distortion." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.792, + 0.484, + 0.902 + ], + "angle": 0, + "content": "At each session \\(t \\geq 0\\), we apply the feature extractor \\(\\mathcal{E}(:, \\theta)\\) on input point clouds \\(\\{(P_k^{(t)})_{k=1}^{|D^{(t)}|}\\}\\) to obtain feature vectors \\(\\pmb{v}\\) in Euclidean space. To avoid the computational cost of calculating Fréchet means directly in hyperbolic space [63], we compute class-mean feature vectors in Euclidean space and project them onto the hyperbolic manifold \\(\\mathbb{B}_c^n\\) using \\(\\exp_0\\) and \\(\\mathcal{M}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.102, + 0.848, + 0.123 + ], + "angle": 0, + "content": "\\[\nv _ {k} ^ {(t)} = \\mathcal {E} \\left(P _ {k} ^ {(t)}; \\theta\\right), \\quad \\forall k \\in \\{1, \\dots , | D ^ {(t)} | \\}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.127, + 0.905, + 0.194 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\bar {v} \\left(c _ {i} ^ {(t)}\\right) = \\frac {1}{| D ^ {(t)} |} \\sum_ {k = 1} ^ {| D ^ {(t)} |} v _ {k} ^ {(t)} \\mathbb {1} \\left[ M _ {k} ^ {(t)} = c _ {i} ^ {(t)} \\right] \\tag {4} \\\\ z (c _ {i} ^ {(t)}) = \\mathcal {M} (\\exp_ {0} (\\overline {{v}} (c _ {i} ^ {(t)}))) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.199, + 0.907, + 0.296 + ], + "angle": 0, + "content": "where \\(i = \\{1,\\dots ,|C^{(t)}|\\}\\) and \\(\\theta\\) denotes learnable parameters. Here, \\(z(c_{i}^{(t)})\\) represents the hyperbolic prototype for each class \\(c_{i}^{(t)}\\in C^{(t)}\\) within the open ball \\(\\mathbb{B}_c^n = \\{x\\in\\) \\(\\mathbb{R}^n:c\\| x\\| < 1\\}\\) . This projection process ensures cohesive, low-distortion representations of class structures as the model adapts incrementally to new classes." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.302, + 0.832, + 0.317 + ], + "angle": 0, + "content": "3.2.3. Uncertainty-aware Busemann learning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.32, + 0.907, + 0.476 + ], + "angle": 0, + "content": "Previous works [12, 21, 27] highlight the utility of hyperbolic distance from any point to the origin \\(\\mathcal{O}_{\\mathbb{B}}\\) in \\(\\mathbb{B}_c^n\\) as a measure of uncertainty, providing motivation to utilize the hyperbolic manifold for robust segmentation. In our approach, we align class prototypes with Ideal Prototypes placed at infinity on the hyperbolic boundary. To refine this alignment, we penalize the uncertainty of projected class features \\(z(c_i^{(t)})\\), represented by their distance \\(d_{\\mathbb{B}}(z(c_i^{(t)}),\\mathcal{O}_{\\mathbb{B}})\\), thereby preventing embedding drift from \\(\\mathcal{O}_{\\mathbb{B}}\\) and ensuring stability." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.477, + 0.906, + 0.621 + ], + "angle": 0, + "content": "These Ideal Prototypes lie at an infinite geodesic distance from all points within \\(\\mathbb{B}_c^n\\), making direct prototypical learning impractical. We address this with the Busemann function \\(\\mathcal{B}\\) introduced in [7]. Ghadimi et al. [23] provides a closed-form expression for measuring such infinite geodesic distances. Let \\(u_{i}\\) be the Ideal Prototype for class \\(c_{i}^{(t)}\\) from the set \\(\\mathcal{I}_n\\), and \\(\\gamma_{u_i^t}\\) a geodesic ray towards \\(u_{i}^{t}\\), with \\(z_{i}^{t} = z(c_{i}^{(t)})\\). The Busemann function \\(\\mathcal{B}\\) for a hyperbolic embedding \\(z_{i}^{t}\\in \\mathbb{B}_{c}^{n}\\) relative to \\(u_{i}^{t}\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.582, + 0.629, + 0.907, + 0.656 + ], + "angle": 0, + "content": "\\[\n\\mathcal {B} _ {u _ {i} ^ {t}} \\left(z _ {i} ^ {t}\\right) = \\lim _ {q \\rightarrow \\infty} \\left(d _ {\\mathbb {B}} \\left(\\gamma_ {u _ {i} ^ {t}} (q), z _ {i} ^ {t}\\right) - q\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.66, + 0.812, + 0.675 + ], + "angle": 0, + "content": "In the Poincaré ball \\(\\mathbb{B}_c^n\\), this simplifies to:" + }, + { + "type": "equation", + "bbox": [ + 0.618, + 0.685, + 0.905, + 0.72 + ], + "angle": 0, + "content": "\\[\n\\mathcal {B} _ {u _ {i} ^ {t}} \\left(z _ {i} ^ {t}\\right) = \\log \\frac {\\left\\| u _ {i} ^ {t} - z _ {i} ^ {t} \\right\\| ^ {2}}{1 - \\left\\| z _ {i} ^ {t} \\right\\| ^ {2}}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.723, + 0.906, + 0.784 + ], + "angle": 0, + "content": "We also incorporate hyperbolic distance as an uncertainty measure for class prototypes in \\(\\mathbb{B}_c^n\\), which integrates into our loss function as in [23]. The alignment loss introduced in Eq. 2 is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.79, + 0.905, + 0.807 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {A l g i n m e n t}} = \\mathcal {L} _ {\\mathcal {B}} + \\phi \\mathcal {L} _ {\\text {U n c e r t a i n t y}} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.814, + 0.906, + 0.861 + ], + "angle": 0, + "content": "where Busemann loss \\(\\mathcal{L}_{\\mathcal{B}} = \\mathcal{B}_{u_i^t}(z_i^t)\\), uncertainty loss \\(\\mathcal{L}_{\\mathrm{Uncertainty}} = d_{\\mathbb{B}}(z_i^t,\\mathcal{O}_{\\mathbb{B}})\\) and \\(\\phi\\) is the strength of the regularization effect of uncertainty loss." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.862, + 0.9, + 0.877 + ], + "angle": 0, + "content": "The overall loss function in Eq. 2 can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.886, + 0.832, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {H i P o}} = \\mathcal {L} _ {\\mathcal {B}} + \\phi \\mathcal {L} _ {\\text {U n c e r t a i n t y}} + \\mathcal {L} _ {\\mathrm {C E}}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11814" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.486, + 0.123 + ], + "angle": 0, + "content": "At test time, we employ only the point-level classification branch to segment 3D point clouds." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.137, + 0.336, + 0.156 + ], + "angle": 0, + "content": "4. Experimental Evaluations" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.163, + 0.485, + 0.375 + ], + "angle": 0, + "content": "Datasets: We evaluate our method on two widely used benchmarks, S3DIS [2] and ScanNetv2 [17], to assess both in-domain and cross-dataset FSCIL performance. These datasets incorporate large-scale coloured 3D point clouds of indoor spaces which introduce semantics. See Sup. Mat. for additional dataset details. For cross-dataset FSCIL, we create a transition dataset from ScanNetv2 to S3DIS, using ten classes from ScanNetv2 as the base and introducing eight new classes from S3DIS in incremental sessions. Following a strict disjoint setup as in [10], each session includes only novel classes. The number of introduced classes varies per session to assess resilience to catastrophic forgetting, which escalates with more incremental sessions in point cloud recognition [60]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.376, + 0.485, + 0.666 + ], + "angle": 0, + "content": "Evaluation metrics: We assess model performance using three standard metrics for incremental learning[70]: Average Incremental Accuracy (AIA), which tracks average accuracy across tasks learned sequentially; Last, the accuracy after the final task; and Average Forgetting Rate \\((\\mathcal{F}_{Last}^{(T)})\\), which measures performance decline on learned tasks after adapting to the last task. Finally, we report mean Intersection-over-Union \\((\\mathrm{mIoU})^1\\) to evaluate segmentation performance. See Sup. Mat. for detailed metric definitions. Architecture: Our backbone model employs PointTransformer [77], leveraging recent advances in point cloud recognition, with a Hyperbolic Classifier head \\(\\mathcal{H}(.)\\) for segmentation. We train HiPo using the Riemannian SGD optimizer, facilitated by geoopt [29] for hyperbolic operations. Training is conducted on both \\(D^{(0)}\\) and incremental datasets \\(D^{(t)}, t > 0\\) with a batch size of 16, AdamW optimizer [37], initial learning rate of 0.001, and weight decay of 0.0001 over 100 epochs. Architecture diagram and further hyper-parameter details are available in the Sup. Mat." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.666, + 0.484, + 0.877 + ], + "angle": 0, + "content": "Competitors: We compare our method against various baselines representing distinct FSCIL approaches. For regularization-based methods, we include LwF with Teacher Adaptation [59], and for optimization-based methods, we include C-FSCIL [24]. We also evaluate against regularization-based FSCIL methods specifically designed for 3D point cloud tasks, such as LGKD [72], SDCOT [79], and 3DPC-CISS [70], which have shown effectiveness with the PointTransformer [77] architecture. Comparative results are provided in Tables 1 and 2. Additionally, we adapt BiDist [78], CLOM [85], TEEN [64], FACT [81], and ALICE [50] using the DGCNN [65] backbone for FSCIL in 3D point cloud segmentation. We report the results with the DGCNN [65] backbone in Table 5." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.091, + 0.593, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.606, + 0.092, + 0.663, + 0.1 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.595, + 0.102, + 0.669, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.694, + 0.092, + 0.729, + 0.101 + ], + "angle": 0, + "content": "SDCOT" + }, + { + "type": "image", + "bbox": [ + 0.672, + 0.102, + 0.748, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.766, + 0.092, + 0.814, + 0.101 + ], + "angle": 0, + "content": "3DPC-CISS" + }, + { + "type": "image", + "bbox": [ + 0.752, + 0.102, + 0.826, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.842, + 0.092, + 0.889, + 0.101 + ], + "angle": 0, + "content": "HIPO (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.831, + 0.103, + 0.905, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.164, + 0.591, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.596, + 0.165, + 0.669, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.675, + 0.164, + 0.748, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.754, + 0.165, + 0.825, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.832, + 0.166, + 0.905, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.226, + 0.591, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.596, + 0.226, + 0.669, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.675, + 0.226, + 0.748, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.753, + 0.226, + 0.826, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.832, + 0.226, + 0.905, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.288, + 0.591, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.596, + 0.288, + 0.669, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.675, + 0.288, + 0.748, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.753, + 0.288, + 0.826, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.832, + 0.288, + 0.905, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.576, + 0.352, + 0.71, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.352, + 0.834, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.408, + 0.907, + 0.492 + ], + "angle": 0, + "content": "Figure 4. Qualitative visualization of model performance: Comparing FSCIL for 3D point cloud segmentation results of our method (HiPo) and other baselines for S3DIS dataset [2]. We observe that even with a few-shot setup, our model performs reasonably well in segmenting out small parts like walls, counters, etc., compared to other methods [70, 79]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.51, + 0.758, + 0.526 + ], + "angle": 0, + "content": "4.1. Discussing the main results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.535, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Table 1 presents a detailed comparison of our proposed method, HiPo, against several baseline methods on the S3DIS [2] and ScanNetv2 [17] datasets across various evaluation settings. HiPo consistently outperforms the current state-of-the-art approach, 3DPC-CISS [70], achieving an average improvement of \\(+2.57\\%\\) mIoU on S3DIS and \\(+2.12\\%\\) on \\(\\mathcal{F}_{Last}^{(T)}\\). This improvement underscores the effectiveness of HiPo's hyperbolic, uncertainty-aware architecture, which mitigates catastrophic forgetting by reducing misalignment between features and prototype representations from earlier sessions. When compared to other baselines, such as LGKD [72] and C-FSCIL [24], HiPo shows substantial gains across all metrics. For instance, in the 10-3T scenario, HiPo achieves \\(33.40\\%\\) mIoU on the Last metric, significantly higher than LGKD's \\(20.02\\%\\) and C-FSCIL's \\(25.02\\%\\). This consistent superiority highlights HiPo's resilience against forgetting, particularly in incremental sessions with larger class increments, where other methods struggle to retain performance. Furthermore, while SDCOT [79] and LwF [59] suffer from high forgetting rates, HiPo effectively maintains stable performance across incremental tasks, with a \\(\\mathcal{F}_{Last}^{(T)}\\) rate as low as \\(17.70\\%\\) in challenging scenarios, compared to SDCOT's \\(30.83\\%\\). This robustness can be attributed to HiPo's hyperbolic embed" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.348, + 0.902 + ], + "angle": 0, + "content": "1We use accuracy and mIoU interchangeably." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "11815" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.115, + 0.089, + 0.885, + 0.199 + ], + "angle": 0, + "content": "
4-3T7-3T10-3TAverage
Last(↑)AIA(↑)FT(Last)(↓)Last(↑)AIA(↑)FT(Last)(↓)Last(↑)AIA(↑)FT(Last)(↓)Last(↑)AIA(↑)FT(Last)(↓)
LwF [59]4.3322.0523.628.6021.8019.8021.3034.0525.5011.4125.3022.97
C-FSCIL [24]9.0627.9025.1120.0131.9217.8625.0235.3620.6818.0331.0621.88
LGKD [72]12.2431.8025.9419.7630.4115.9720.0232.7125.3817.3431.6422.43
SDCOT [79]5.3028.4230.8312.0029.4626.2023.8035.3023.0013.7031.0626.67
3DPC-CISS [70]25.7040.5019.7324.2034.5015.4529.6038.2017.2026.5037.7317.46
HtPo30.6043.8717.7029.4036.8011.1033.4040.2513.7031.1340.3014.16
" + }, + { + "type": "table", + "bbox": [ + 0.115, + 0.201, + 0.884, + 0.311 + ], + "angle": 0, + "content": "
10-2T10-5T15-5TAverage
Last(↑)AIA(↑)FT(Last)(↓)Last(↑)AIA(↑)FT(Last)(↓)Last(↑)AIA(↑)FT(Last)(↓)Last(↑)AIA(↑)FT(Last)(↓)
LwF [59]1.7010.7510.861.0517.5324.725.0621.3832.642.6016.5522.74
C-FSCIL [24]7.1215.9410.5811.1224.1819.5911.5023.5824.159.2521.2318.11
LGKD [72]6.0215.6011.509.5525.5222.4511.1023.8325.468.2221.6519.80
SDCOT [79]0.509.5610.880.4017.8326.151.3019.5036.400.7315.6324.47
3DPC-CISS [70]6.2013.138.328.1022.3021.3012.2024.9525.508.8320.1218.37
HtPo8.5017.1310.3614.6527.2218.8511.9024.4525.1011.6822.2718.12
" + }, + { + "type": "table_caption", + "bbox": [ + 0.089, + 0.321, + 0.907, + 0.381 + ], + "angle": 0, + "content": "Table 1. Comparison with literature on the S3DIS [2] (top table) and ScanNetv2 [17] (bottom table) datasets, using the PointTransformer backbone [77] in a 5-shot setting. CIL mIoU (%) is reported. “\\(M-NT\\)” indicates that \\(M\\) is the number of classes in the base session \\(S^{(0)}\\), while \\(N\\) is the number of classes introduced in each incremental session \\(S^{(t)}\\), \\(t > 0\\). The highest performance in each column is shown in bold, with the second-highest results underlined. HiPO performs superior compared to other baselines in different incremental settings." + }, + { + "type": "table", + "bbox": [ + 0.147, + 0.401, + 0.432, + 0.514 + ], + "angle": 0, + "content": "
Average
Last (↑)AIA (↑)F(T)Last (↓)
LwF [59]8.7822.6023.94
C-FSCIL [24]13.5525.0219.72
LGKD [72]15.6325.5617.51
SDCOT [79]19.2229.2917.52
3DPC-CISS [70]19.9529.1616.01
HtPo20.6029.8116.05
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.524, + 0.483, + 0.595 + ], + "angle": 0, + "content": "Table 2. SOTA comparison on ScanNetv2 \\(\\rightarrow\\) S3DIS dataset. We compute the average of 10-4T and 10-8T with PointTransformer backbone [77] in 5-shot setting. HIPO achieves superior Last and AIA values with comparable \\(F_{Last}^{(T)}\\) values, highlighting better domain adaptability than other methods." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.597, + 0.483, + 0.642 + ], + "angle": 0, + "content": "ding space, which aligns naturally with hierarchical structures, thereby enhancing inter-class separation and intraclass compactness over time." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.643, + 0.483, + 0.718 + ], + "angle": 0, + "content": "Moreover, Fig. 4 provides qualitative comparisons, where HiPO demonstrates marked improvements over leading baselines, particularly in accurately segmenting challenging structures such as walls and counters. Further analyses and results are available in the Sup. Mat." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.728, + 0.303, + 0.745 + ], + "angle": 0, + "content": "4.2. Main ablation analysis" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.75, + 0.483, + 0.884 + ], + "angle": 0, + "content": "Cross-dataset class-incremental segmentation: To further assess the effectiveness of our proposed modules, we conduct cross-dataset incremental segmentation experiments (ScanNetv2 \\(\\rightarrow\\) S3DIS). Detailed setup information is in Sup. Mat. As shown in Table 2, the cross-dataset setting introduces a notable increase in forgetting due to domain shift. However, the results underscore HiPo's robustness in managing such domain shifts effectively, with improvements of \\(0.65\\%\\) in Last and \\(0.52\\%\\) in AIA metrics." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Ablation studies on penalty in Busemann loss: We pro" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.404, + 0.907, + 0.6 + ], + "angle": 0, + "content": "vide a comprehensive ablation study on the regularization constant \\(\\phi\\) introduced in Eq. 7. The regularization strength is varied from \\(\\phi = 0\\) to \\(\\phi = 1\\) in intervals of 0.25 for S3DIS [2] (4-3T, 7-3T, 10-3T) and ScanNetv2 [17] (10-2T, 10-5T, 15-5T). As shown in Table 3, our findings corroborate prior work [23]: without a penalty \\((\\phi = 0)\\), samples are overly drawn towards the Ideal Prototypes, resulting in overconfident predictions and lower AIA on test data \\(\\mathcal{D}_{test}\\) in addition to vanishing gradient problems [44]. Conversely, a high penalty \\((\\phi = 1)\\) increases inter-class confusion, also leading to lower AIA on \\(\\mathcal{D}_{test}\\). Our final chosen results, highlighted in bold in Table 3, reflect the optimal balance for \\(\\phi\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.606, + 0.909, + 0.803 + ], + "angle": 0, + "content": "Comparison of metrics at different curvatures: Following [47], we experimented with curvature values of \\( c = 0.0001 \\) and \\( c = 0.001 \\) to implement traditional Euclidean architectures (Poincaré Ball with zero curvature; very small Poincaré ball curvature values are chosen as curvature \\( c = 0 \\) is numerically unstable [27]) for the S3DIS [2] (4-3T, 7-3T, 10-3T) and ScanNetv2 [17] (10-2T, 10-5T, 15-5T) setups. Table 4 shows that our HiPO with a Poincaré Ball curvature of \\( c = -1.00 \\) consistently outperforms its Euclidean counterpart across most settings. This finding demonstrates that Riemannian geometry leverages the hierarchical structure of point clouds more effectively than Euclidean space, suggesting further exploration in this domain." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.808, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Sensitivity to the different number of training samples in each incremental episode: To assess the impact of data availability on model performance, we gradually increase the number of samples per class (shot) from one to all available examples and analyze the Last mIoU and \\(\\mathcal{F}_{Last}^{(T)}\\) results for different shot values in Fig. 5. Our findings indi" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "11816" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.203, + 0.089, + 0.803, + 0.198 + ], + "angle": 0, + "content": "
DatasetsSettingsφ = 0.00φ = 0.25φ = 0.50φ = 0.75φ = 1.00
AIA↑FT(St)↓AIA↑FT(St)↓AIA↑FT(St)↓AIA↑FT(St)↓AIA↑FT(St)↓
S3DIS [2]4-3T42.8717.7043.4017.8643.2718.2343.8717.7043.2218.96
7-3T36.1311.6036.4611.2036.8011.1036.7011.1036.0612.10
10-3T39.9514.3040.2513.7040.1014.0040.0714.0540.0514.10
ScanNetv2 [17]10-2T16.7510.5116.8510.5016.9010.0817.1310.3616.9110.76
10-5T27.0818.8727.1318.8727.2218.8527.2118.8527.1618.92
15-5T24.2525.5024.4525.1024.0525.9023.9526.1024.0026.00
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.207, + 0.907, + 0.237 + ], + "angle": 0, + "content": "Table 3. Different penalty values in Busemann loss in 5-shot incremental segmentation across the S3DIS [2] dataset and ScanNetV2 [17] dataset, highlighting the regularization benefits introduced by the hyperbolic distance of the projected class prototypes from the origin." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.259, + 0.499, + 0.361 + ], + "angle": 0, + "content": "
DatasetsSettingsc = -1.00 (Ours)c = 0.0001c = 0.001
AIA↑FTLast↓AIA↑FTLast↓AIA↑FTLast↓
S3DIS [2]4-3T43.8717.7042.9519.9343.2317.38
7-3T36.8011.1036.2311.6036.8611.05
10-3T40.2513.7040.2013.8040.2213.75
ScanNetv2 [17]10-2T17.1310.3616.5410.2516.9810.18
10-5T27.2218.8527.1619.0026.8319.92
15-5T24.4525.1024.4025.2024.2025.60
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.37, + 0.484, + 0.442 + ], + "angle": 0, + "content": "Table 4. Comparison for different curvature values in 5-shot incremental segmentation across S3DIS [2] and ScanNetV2 [17] datasets. This highlights the superior stability-plasticity trade-off of hyperbolic embedding space with negative curvature, where small positive curvature approximates Euclidean space." + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.454, + 0.303, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.145, + 0.596, + 0.265, + 0.609 + ], + "angle": 0, + "content": "(a) Last mIoU values" + }, + { + "type": "image", + "bbox": [ + 0.312, + 0.458, + 0.48, + 0.586 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.332, + 0.593, + 0.45, + 0.611 + ], + "angle": 0, + "content": "(b) Forgetting \\(\\mathcal{F}_{\\text{Last}}^{(T)}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.621, + 0.484, + 0.69 + ], + "angle": 0, + "content": "Figure 5. mIoU improves and forgetting reduces with more training examples. Our proposed method outperforms the top-2 baseline methods ([70, 79]) in terms of Last mIoU and forgetting mitigation in the 10-2T setting (Full-data: fully supervised training in incremental fashion)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.483, + 0.779 + ], + "angle": 0, + "content": "cate that catastrophic forgetting is particularly severe with limited data; however, our method substantially reduces forgetting compared to other baselines. Notably, increasing the number of training examples significantly mitigates forgetting for H1Po, while it exacerbates forgetting for SDCOT." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Experiments with a different backbone: Besides PointTransformer [77], we evaluate our model with the backbone of DGCNN [65]. Unlike current SOTA FSCIL methods—BiDist [78], CLOM [85], TEEN [64], FACT [81], and ALICE [50]—which perform well with DGCNN but degrade with PointTransformer, our method remains effective across both architectures. As shown in Table 5, our approach achieves consistent gains with DGCNN, outper" + }, + { + "type": "table", + "bbox": [ + 0.59, + 0.259, + 0.832, + 0.386 + ], + "angle": 0, + "content": "
MethodLast (↑)AIA (↑)FTLast (↓)
LwF [59]9.4526.7524.41
C-FSCIL [24]20.2230.6422.15
LGKD [72]18.6329.1615.75
SDCOT [79]22.7524.5419.94
CLOM [85]21.1834.6718.83
TEEN [64]22.7133.9015.37
FACT [81]24.2726.7021.26
3DPC-CISS [70]22.0324.7419.36
BiDist [78]27.7039.1415.85
HiPo29.9039.8613.87
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.395, + 0.907, + 0.454 + ], + "angle": 0, + "content": "Table 5. Performance comparison on the S3DIS dataset [2] using the DGCNN backbone [65] in the 5-shot setting. Reported metrics include Last, \\(\\mathcal{F}_{Last}^{(T)}\\), and AIA for 4-3T and 7-3T settings. HiPo outperforms all baseline methods with DGCNN." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.476, + 0.907, + 0.525 + ], + "angle": 0, + "content": "forming baselines with \\(+2.20\\%\\) mIoU on Last, \\(+0.72\\%\\) on AIA, and reducing forgetting by \\(+1.98\\%\\) on \\(\\mathcal{F}_{Last}^{(T)}\\). This demonstrates the model-agnostic nature of HiPO." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.542, + 0.816, + 0.559 + ], + "angle": 0, + "content": "5. Conclusions & Future Directions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We introduce a previously unexplored problem setting along with a novel solution specifically designed for few-shot class-incremental semantic segmentation of 3D point cloud data, addressing a critical gap in the literature. Recognizing the unique challenges posed by point clouds, we identified limitations in current FSCIL methods and developed HiPo, a model that leverages hyperbolic space to embed point clouds while achieving an improved stability-plasticity trade-off. Our approach consistently outperforms state-of-the-art baselines, demonstrating that hyperbolic geometry's ability to represent hierarchical structures offers a more effective framework for FSCIL in 3D data. HiPo is particularly valuable for real-world applications like autonomous driving, industrial inspection, urban planning, and environmental monitoring, where generalization to unseen classes is essential. Furthermore, our approach enhances safety and efficiency in critical areas such as disaster response and infrastructure maintenance by enabling accurate, timely semantic segmentation of complex 3D environments. Future work will focus on managing possible outliers that may emerge without prior reference across incremental sessions, further strengthening our robustness." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11817" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.486, + 0.185 + ], + "angle": 0, + "content": "[1] Noor Ahmed, Anna Kukleva, and Bernt Schiele. Orco: Towards better generalization via orthogonality and contrast for few-shot class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28762-28771, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.187, + 0.485, + 0.257 + ], + "angle": 0, + "content": "[2] Iro Armeni, Ozan Sener, Amir R Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3d semantic parsing of large-scale indoor spaces. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1534-1543, 2016. 2, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.259, + 0.483, + 0.327 + ], + "angle": 0, + "content": "[3] Mina Ghadimi Atigh, Julian Schoep, Erman Acar, Nanne Van Noord, and Pascal Mettes. Hyperbolic image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4453-4462, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.329, + 0.484, + 0.384 + ], + "angle": 0, + "content": "[4] Eden Belouadah and Adrian Popescu. II2m: Class incremental learning with dual memory. In Proceedings of the IEEE/CVF international conference on computer vision, pages 583-592, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.387, + 0.483, + 0.441 + ], + "angle": 0, + "content": "[5] Abhijit Bendale and Terrance Boult. Towards open world recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1893-1902, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.444, + 0.483, + 0.527 + ], + "angle": 0, + "content": "[6] Chris Burges, Tal Shaked, Erin Renshaw, Ari Lazier, Matt Deeds, Nicole Hamilton, and Greg Hullender. Learning to rank using gradient descent. In Proceedings of the 22nd International Conference on Machine Learning, page 89-96, New York, NY, USA, 2005. Association for Computing Machinery. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.529, + 0.482, + 0.556 + ], + "angle": 0, + "content": "[7] Herbert Busemann. The geometry of geodesics. Courier Corporation, 2005. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.559, + 0.483, + 0.614 + ], + "angle": 0, + "content": "[8] Francisco M Castro, Manuel J Marín-Jiménez, Nicolas Guil, Cordelia Schmid, and Karteek Alahari. End-to-end incremental learning. In Proceedings of the European conference on computer vision (ECCV), pages 233-248, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.616, + 0.483, + 0.685 + ], + "angle": 0, + "content": "[9] Jun Cen, Peng Yun, Shiwei Zhang, Junhao Cai, Di Luan, Mingqian Tang, Ming Liu, and Michael Yu Wang. Openworld semantic segmentation for lidar point clouds. In European Conference on Computer Vision, pages 318-334. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.687, + 0.483, + 0.757 + ], + "angle": 0, + "content": "[10] Fabio Cermelli, Massimiliano Mancini, Samuel Rota Bulo, Elisa Ricci, and Barbara Caputo. Modeling the background for incremental learning in semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9233-9242, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.758, + 0.483, + 0.828 + ], + "angle": 0, + "content": "[11] Arslan Chaudhry, Puneet K Dokania, Thalaiyasingam Ajthan, and Philip HS Torr. Riemannian walk for incremental learning: Understanding forgetting and intransigence. In Proceedings of the European conference on computer vision (ECCV), pages 532-547, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.83, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[12] Bike Chen, Wei Peng, Xiaofeng Cao, and Juha Röning. Hyperbolic uncertainty aware semantic segmentation. IEEE Transactions on Intelligent Transportation Systems, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[13] Ali Cheraghian, Shafin Rahman, Pengfei Fang, Soumava Kumar Roy, Lars Petersson, and Mehrtash" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.148 + ], + "angle": 0, + "content": "Harandi. Semantic-aware knowledge distillation for few-shot class-incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2534-2543, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[14] Townim Chowdhury, Mahira Jalisha, Ali Cheraghian, and Shafin Rahman. Learning without forgetting for 3d point cloud objects. In Advances in Computational Intelligence: 16th International Work-Conference on Artificial Neural Networks, IWANN 2021, Virtual Event, June 16-18, 2021, Proceedings, Part I 16, pages 484-497. Springer, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.233, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[15] Townim Chowdhury, Ali Cheraghian, Sameera Ramasinghe, Sahar Ahmadi, Morteza Saberi, and Shafin Rahman. Few-shot class-incremental learning for 3d point cloud objects. In European Conference on Computer Vision, pages 204-220. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.905, + 0.356 + ], + "angle": 0, + "content": "[16] Yawen Cui, Zitong Yu, Wei Peng, Qi Tian, and Li Liu. Rethinking few-shot class-incremental learning with open-set hypothesis in hyperbolic geometry. IEEE Transactions on Multimedia, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.358, + 0.905, + 0.427 + ], + "angle": 0, + "content": "[17] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.905, + 0.468 + ], + "angle": 0, + "content": "[18] David L. Davies and Donald W. Bouldin. A cluster separation measure. IEEE Transactions on Pattern Analysis and Machine Intelligence, PAMI-1(2):224-227, 1979. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.469, + 0.905, + 0.525 + ], + "angle": 0, + "content": "[19] Jiahua Dong, Yang Cong, Gan Sun, Bingtao Ma, and Lichen Wang. I3dol: Incremental 3d object learning without catastrophic forgetting. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6066-6074, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.905, + 0.566 + ], + "angle": 0, + "content": "[20] DC Dowson and BV666017 Landau. The frechet distance between multivariate normal distributions. Journal of multivariate analysis, 12(3):450-455, 1982. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.567, + 0.905, + 0.635 + ], + "angle": 0, + "content": "[21] Luca Franco, Paolo Mandica, Konstantinos Kallidromitis, Devin Guillory, Yu-Teng Li, Trevor Darrell, and Fabio Galasso. Hyperbolic active learning for semantic segmentation under domain shift. arXiv preprint arXiv:2306.11180, 2023.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.637, + 0.905, + 0.691 + ], + "angle": 0, + "content": "[22] Octavian-Eugen Ganea, Gary Bécigneul, and Thomas Hofmann. Hyperbolic convolutional neural networks. In Advances in Neural Information Processing Systems, pages 4851-4861, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.692, + 0.905, + 0.747 + ], + "angle": 0, + "content": "[23] Mina Ghadimi Atigh, Martin Keller-Ressel, and Pascal Mettes. Hyperbolic busemann learning with ideal prototypes. Advances in Neural Information Processing Systems, 34:103-115, 2021. 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.748, + 0.905, + 0.816 + ], + "angle": 0, + "content": "[24] Michael Hersche, Geethan Karunaratne, Giovanni Cherubini, Luca Benini, Abu Sebastian, and Abbas Rahimi. Constrained few-shot class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9057-9067, 2022. 1, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[25] Naiwen Hu, Haozhe Cheng, Yifan Xie, Pengcheng Shi, and Jihua Zhu. Hyperbolic image-and-pointcloud contrastive learning for 3d classification. arXiv preprint arXiv:2409.15810, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[26] Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, and Andrew Markham." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "11818" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.147 + ], + "angle": 0, + "content": "Randla-net: Efficient semantic segmentation of large-scale point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11108-11117, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.217 + ], + "angle": 0, + "content": "[27] Valentin Khrulkov, Leyla Mirvakhabova, Evgeniya Ustinova, Ivan Oseledets, and Victor Lempitsky. Hyperbolic image embeddings. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6418-6428, 2020. 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.218, + 0.483, + 0.273 + ], + "angle": 0, + "content": "[28] Do-Yeon Kim, Dong-Jun Han, Jun Seo, and Jaekyun Moon. Warping the space: Weight space rotation for class incremental few-shot learning. In The Eleventh International Conference on Learning Representations, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.274, + 0.482, + 0.302 + ], + "angle": 0, + "content": "[29] Max Kochurov, Rasul Karimov, and Serge Kozlukov. Geoopt: Riemannian optimization in pytorch, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.303, + 0.484, + 0.37 + ], + "angle": 0, + "content": "[30] Anna Kukleva, Hilde Kuehne, and Bernt Schiele. Generalized and incremental few-shot learning by explicit learning and calibration without forgetting. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9020-9029, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.372, + 0.483, + 0.44 + ], + "angle": 0, + "content": "[31] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.442, + 0.482, + 0.482 + ], + "angle": 0, + "content": "[32] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40(12):2935-2947, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.485, + 0.552 + ], + "angle": 0, + "content": "[33] Fangzhou Lin, Yun Yue, Songlin Hou, Xuechu Yu, Yajun Xu, Kazunori D Yamada, and Ziming Zhang. Hyperbolic chamfer distance for point cloud completion. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14595-14606, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.483, + 0.62 + ], + "angle": 0, + "content": "[34] Pei Lin, Zheng Cai, Jiangtao Zhang, and Lifang Ruan. A survey on point cloud segmentation: From classical methods to deep learning. ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM), 16(3): 1-27, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.622, + 0.483, + 0.676 + ], + "angle": 0, + "content": "[35] Yuyang Liu, Yang Cong, Gan Sun, Tao Zhang, Jiahua Dong, and Hongsen Liu. L3doc: Lifelong 3d object classification. IEEE Transactions on Image Processing, 30:7486-7498, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.678, + 0.483, + 0.733 + ], + "angle": 0, + "content": "[36] Zan Liu, Zilong Zhang, Yiyi Lin, Lin Zhu, Han Jin, and Stephen Lin. Pointaugment: An auto-augmentation framework for point cloud classification. arXiv preprint arXiv:2002.10876, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.482, + 0.762 + ], + "angle": 0, + "content": "[37] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.763, + 0.483, + 0.817 + ], + "angle": 0, + "content": "[38] Paolo Mandica, Luca Franco, Konstantinos Kallidromitis, Suzanne Petryk, and Fabio Galasso. Hyperbolic learning with multimodal large language models. arXiv preprint arXiv:2408.05097, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.483, + 0.859 + ], + "angle": 0, + "content": "[39] Pratik Mazumder, Pravendra Singh, and Piyush Rai. Few-shot lifelong learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2337-2345, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[40] Leland McInnes, John Healy, and James Melville. Umap: Uniform manifold approximation and projection for dimension reduction. arXiv preprint arXiv:1802.03426, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.485, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[41] Pascal Mettes, Elise Van der Pol, and Cees Snoek. Hyperspherical prototype networks. Advances in neural information processing systems, 32, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.905, + 0.192 + ], + "angle": 0, + "content": "[42] Pascal Mettes, Mina Ghadimi Atigh, Martin Keller-Ressel, Jeffrey Gu, and Serena Yeung. Hyperbolic deep learning in computer vision: A survey. International Journal of Computer Vision, pages 1-25, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.194, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[43] Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. Distributed representations of words and phrases and their compositionality. Advances in neural information processing systems, 26, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.904, + 0.305 + ], + "angle": 0, + "content": "[44] Gal Mishne, Zhengchao Wan, Yusu Wang, and Sheng Yang. The numerical stability of hyperbolic representation learning. In International Conference on Machine Learning, pages 24925-24949. PMLR, 2023. 2, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.307, + 0.905, + 0.374 + ], + "angle": 0, + "content": "[45] Antonio Montanaro, Diego Valsesia, and Enrico Magli. Rethinking the compositionality of point clouds through regularization in the hyperbolic space. Advances in Neural Information Processing Systems, 35:33741-33753, 2022. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.905, + 0.446 + ], + "angle": 0, + "content": "[46] Antonio Montanaro, Diego Valsesia, and Enrico Magli. Towards hyperbolic regularizers for point cloud part segmentation. In ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.449, + 0.907, + 0.517 + ], + "angle": 0, + "content": "[47] Gabriel Moreira, Manuel Marques, João Paulo Costeira, and Alexander Hauptmann. Hyperbolic vs euclidean embeddings in few-shot learning: Two sides of the same coin. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2082-2090, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.603 + ], + "angle": 0, + "content": "[48] Samrat Mukherjee, Tanuj Sur, Saurish Seksaria, Subhasis Chaudhuri, Gemma Roig, and Biplab Banerjee. Uidaple: Unsupervised incremental domain adaptation through adaptive prompt learning. In ICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.905, + 0.658 + ], + "angle": 0, + "content": "[49] Maximilian Nickel and Douwe Kiela. Poincaré embeddings for learning hierarchical representations. In Advances in neural information processing systems, pages 6338-6347, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[50] Can Peng, Kun Zhao, Tianren Wang, Meng Li, and Brian C Lovell. Few-shot class-incremental learning from an open-set perspective. In European Conference on Computer Vision, pages 382-397. Springer, 2022. 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[51] Jeffrey Pennington, Richard Socher, and Christopher Manning. GloVe: Global vectors for word representation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1532-1543, Doha, Qatar, 2014. Association for Computational Linguistics. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[52] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[53] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11819" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "point sets in a metric space. Advances in neural information processing systems, 30, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[54] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[55] Sachin Ravi and Hugo Larochelle. Optimization as a model for few-shot learning. arXiv preprint arXiv:1606.04474, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.482, + 0.318 + ], + "angle": 0, + "content": "[56] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 2001–2010, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.482, + 0.362 + ], + "angle": 0, + "content": "[57] Rik Sarkar. Low distortion delaunay embedding of trees in hyperbolic plane. In International symposium on graph drawing, pages 355-366. Springer, 2011. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.364, + 0.482, + 0.419 + ], + "angle": 0, + "content": "[58] Yisheng Song, Ting Wang, Puyu Cai, Subrota K Mondal, and Jyoti Prakash Sahoo. A comprehensive survey of few-shot learning: Evolution, applications, challenges, and opportunities. ACM Computing Surveys, 55(13s):1-40, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.421, + 0.483, + 0.503 + ], + "angle": 0, + "content": "[59] Filip Szatkowski, Mateusz Pyla, Marcin Przewieżlikowski, Sebastian Cygert, Bartlomiej Twardowski, and Tomasz Trzeciński. Adapt your teacher: Improving knowledge distillation for exemplar-free continual learning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1977-1987, 2024. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.505, + 0.483, + 0.56 + ], + "angle": 0, + "content": "[60] Yuwen Tan and Xiang Xiang. Cross-domain few-shot incremental learning for point-cloud recognition. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2307-2316, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.482, + 0.63 + ], + "angle": 0, + "content": "[61] Xiaoyu Tao, Xiaopeng Hong, Xinyuan Chang, Songlin Dong, Xing Wei, and Yihong Gong. Few-shot class incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12183-12192, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[62] Jesper E Van Engelen and Holger H Hoos. A survey on semi-supervised learning. Machine learning, 109(2):373-440, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.482, + 0.73 + ], + "angle": 0, + "content": "[63] Max van Spengler, Erwin Berkhout, and Pascal Mettes. Poincare resnet. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5419-5428, 2023. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.733, + 0.482, + 0.788 + ], + "angle": 0, + "content": "[64] Qi-Wei Wang, Da-Wei Zhou, Yi-Kai Zhang, De-Chuan Zhan, and Han-Jia Ye. Few-shot class-incremental learning via training-free prototype calibration. Advances in Neural Information Processing Systems, 36, 2024. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.482, + 0.844 + ], + "angle": 0, + "content": "[65] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (tog), 38(5):1-12, 2019. 1, 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[66] Yaqing Wang, Quanming Yao, James T Kwok, and Lionel M Ni. Generalizing from a few examples: A survey on few-shot learning. ACM computing surveys (csur), 53(3):1-34, 2020. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.148 + ], + "angle": 0, + "content": "[67] Wenxuan Wu, Zhongang Qi, and Li Fuxin. Pointconv: Deep convolutional networks on 3d point clouds. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 9621-9630, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[68] Yating Xu, Conghui Hu, Na Zhao, and Gim Hee Lee. Generalized few-shot point cloud segmentation via geometric words. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21506-21515, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.205, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[69] Xiangli Yang, Zixing Song, Irwin King, and Zenglin Xu. A survey on deep semi-supervised learning. IEEE Transactions on Knowledge and Data Engineering, 35(9):8934-8954, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.907, + 0.329 + ], + "angle": 0, + "content": "[70] Yuwei Yang, Munawar Hayat, Zhao Jin, Chao Ren, and Yinjie Lei. Geometry and uncertainty-aware 3d point cloud class-incremental semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21759-21768, 2023. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.331, + 0.905, + 0.385 + ], + "angle": 0, + "content": "[71] Zhilu Yang, Ying Zheng, and Jun Zhang. Learning to learn from labeled and unlabeled data with point-wise mutual information. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5704-5711, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.386, + 0.905, + 0.469 + ], + "angle": 0, + "content": "[72] Ze Yang, Ruibo Li, Evan Ling, Chi Zhang, Yiming Wang, Dezhao Huang, Keng Teck Ma, Minhoe Hur, and Guosheng Lin. Label-guided knowledge distillation for continual semantic segmentation on 2d images and 3d point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18601-18612, 2023. 1, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.471, + 0.905, + 0.525 + ], + "angle": 0, + "content": "[73] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. arXiv preprint arXiv:2111.07783, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.905, + 0.593 + ], + "angle": 0, + "content": "[74] Yun Yue, Fangzhou Lin, Guanyi Mou, and Ziming Zhang. Understanding hyperbolic metric learning through hard negative sampling. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1891-1903, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.596, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[75] Maciej Zamorski, Michal Stypulkowski, Konrad Karanowski, Tomasz Trzeciński, and Maciej Zieba. Continual learning on 3d point clouds with random compressed rehearsal. Computer Vision and Image Understanding, 228: 103621, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.665, + 0.905, + 0.734 + ], + "angle": 0, + "content": "[76] Chi Zhang, Nan Song, Guosheng Lin, Yun Zheng, Pan Pan, and Yinghui Xu. Few-shot incremental learning with continually evolved classifiers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12455-12464, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.907, + 0.789 + ], + "angle": 0, + "content": "[77] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16259-16268, 2021. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[78] Linglan Zhao, Jing Lu, Yunlu Xu, Zhanzhan Cheng, Dashan Guo, Yi Niu, and Xiangzhong Fang. Few-shot class incremental learning via class-aware bilateral distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11838-11847, 2023. 1, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.907, + 0.902 + ], + "angle": 0, + "content": "[79] Na Zhao and Gim Hee Lee. Static-dynamic co-teaching for class-incremental 3d object detection. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11820" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.12 + ], + "angle": 0, + "content": "the AAAI Conference on Artificial Intelligence, pages 3436-3445, 2022. 1, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.177 + ], + "angle": 0, + "content": "[80] Na Zhao, Tat-Seng Chua, and Gim Hee Lee. Few-shot 3d point cloud semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8873-8882, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.483, + 0.248 + ], + "angle": 0, + "content": "[81] Da-Wei Zhou, Fu-Yun Wang, Han-Jia Ye, Liang Ma, Shiliang Pu, and De-Chuan Zhan. Forward compatible few-shot class-incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9046–9056, 2022. 1, 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.483, + 0.304 + ], + "angle": 0, + "content": "[82] Da-Wei Zhou, Zi-Wen Cai, Han-Jia Ye, De-Chuan Zhan, and Ziwei Liu. Revisiting class-incremental learning with pretrained models: Generalizability and adaptivity are all you need. arXiv preprint arXiv:2303.07338, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.305, + 0.483, + 0.36 + ], + "angle": 0, + "content": "[83] Da-Wei Zhou, Qi-Wei Wang, Zhi-Hong Qi, Han-Jia Ye, De-Chuan Zhan, and Ziwei Liu. Class incremental learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.483, + 0.404 + ], + "angle": 0, + "content": "[84] Yudong Zhu, Di Zhou, Jinghui Xiao, Xin Jiang, Xiao Chen, and Qun Liu. Hypertext: Endowing fasttext with hyperbolic geometry. arXiv preprint arXiv:2010.16143, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.483, + 0.471 + ], + "angle": 0, + "content": "[85] Yixiong Zou, Shanghang Zhang, Yuhua Li, and Ruixuan Li. Margin-based few-shot class-incremental learning with class-level overfitting mitigation. Advances in neural information processing systems, 35:27267-27279, 2022. 1, 3, 6, 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.471 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "11821" + } + ] +] \ No newline at end of file diff --git a/2025/Hyperbolic Uncertainty-Aware Few-Shot Incremental Point Cloud Segmentation/1247e066-a517-4ffe-b97e-a9bff43b89c3_origin.pdf b/2025/Hyperbolic Uncertainty-Aware Few-Shot Incremental Point Cloud Segmentation/1247e066-a517-4ffe-b97e-a9bff43b89c3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..665ff931ffa520afd30ce6ed54ed45a8232061fb --- /dev/null +++ b/2025/Hyperbolic Uncertainty-Aware Few-Shot Incremental Point Cloud Segmentation/1247e066-a517-4ffe-b97e-a9bff43b89c3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9b3c9e406cbb84cf311093fe37d2fcdb9c5f9bc88165cb41fc36b23db7e3f335 +size 8412197 diff --git a/2025/Hyperbolic Uncertainty-Aware Few-Shot Incremental Point Cloud Segmentation/full.md b/2025/Hyperbolic Uncertainty-Aware Few-Shot Incremental Point Cloud Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..87ec2a7a4a2c1adeee492039feab8ef71de559de --- /dev/null +++ b/2025/Hyperbolic Uncertainty-Aware Few-Shot Incremental Point Cloud Segmentation/full.md @@ -0,0 +1,380 @@ +# Hyperbolic Uncertainty-Aware Few-Shot Incremental Point Cloud Segmentation + +Tanuj Sur† Samrat Mukherjee* Kaizer Rahaman‡ Subbasis Chaudhuri* Muhammad Haris Khan§ Biplab Banerjee* + +†Chennai Mathematical Institute *Indian Institute of Technology Bombay ‡Indian Institute of Technology Kharagpur §Mohamed Bin Zayed University of Artificial Intelligence + +# Abstract + +3D point cloud segmentation is essential across a range of applications; however, conventional methods often struggle in evolving environments, particularly when tasked with identifying novel categories under limited supervision. Few-Shot Learning (FSL) and Class Incremental Learning (CIL) have been adapted previously to address these challenges in isolation, yet the combined paradigm of Few-Shot Class Incremental Learning (FSCIL) remains largely unexplored for point cloud segmentation. To address this gap, we introduce Hyperbolic Ideal Prototypes Optimization (HiPO), a novel framework that harnesses hyperbolic embeddings for FSCIL in 3D point clouds. HiPO employs the Poincaré Hyperbolic Sphere as its embedding space, integrating Ideal Prototypes enriched by CLIP-derived class semantics, to capture the hierarchical structure of 3D data. By enforcing orthogonality among prototypes and maximizing representational margins, HiPO constructs a resilient embedding space that mitigates forgetting and enables the seamless integration of new classes, thereby effectively countering overfitting. Extensive evaluations on S3DIS, ScanNetv2, and cross-dataset scenarios demonstrate HiPO's strong performance, significantly surpassing existing approaches in both in-domain and cross-dataset FSCIL tasks for 3D point cloud segmentation. + +# 1. Introduction + +Recent advancements in deep learning have spurred progress in data-efficient learning, with semi-supervised learning (SSL) [62, 69, 71] and few-shot learning (FSL) [55, 58, 66] emerging as key approaches. SSL utilizes both labeled and unlabeled data to enhance learning, while FSL enables generalization from only a few labeled samples per class. However, both approaches face challenges in real-world, dynamic environments where models must continuously adapt to new classes while retaining prior knowledge + +![](images/70e6e92c326f538e13202a67060afa89ed8e4d8a880cca0433775eef9025159b.jpg) +Figure 1. Our approach tackles the novel challenge of few-shot class incremental 3D point cloud segmentation. In each session $S^{(t)}, t \geq 0$ , our model takes an RGB point cloud of an indoor scene $D^{(t)}$ as input and performs semantic segmentation. During the base training session $S^{(0)}$ , sufficient examples are available for each class. In later sessions, only a limited number of labeled point cloud examples are provided for new classes, introducing data imbalance and potential overfitting and forgetting risks. + +edge—an essential requirement for continual learning (CL) [1, 4, 8, 11, 32, 48, 56]. Few-Shot Class Incremental Learning (FSCIL) [1, 13, 24, 28, 30, 39] combines the data efficiency of FSL with the adaptability of CL, yet remains underdeveloped for complex tasks such as segmentation in unstructured 3D data (see Fig. 1). + +Point cloud segmentation is pivotal in applications like autonomous driving and robotics, where it enables precise object recognition and scene understanding [36, 52, 68]. This task assigns semantic labels to each point within a 3D structure, facilitating critical decision-making processes. Current supervised segmentation methods [26, 53, 65, 67] are highly reliant on extensive labeled datasets, which are costly and time-intensive to produce [34]. These methods also assume a fixed set of labels, limiting their adaptability in scenarios where new classes emerge unexpectedly. While FSCIL offers a potential solution, existing methods [1, 24, 72, 78, 79, 81, 85] struggle to represent the complex, hierarchical structures typical of point cloud data due to the limitations of Euclidean embeddings [33, 45, 46]. + +Capturing hierarchical relationships, such as recogniz + +ing subcomponents within an object, demands a framework that can accommodate intricate interrelations. This challenge is intensified in Class Incremental Learning (CIL) [82, 83], where the model must incorporate new classes while preserving previously learned representations—a balance between adaptability and memory known as the stability-plasticity trade-off. In FSCIL, especially under data scarcity, this trade-off is often compromised, leading to overfitting and catastrophic forgetting. Euclidean spaces exacerbate these challenges, as they lack the capacity to model evolving data scenarios [76, 81]. + +Hyperbolic geometry [25, 38, 42, 57, 74] provides a more suitable framework for hierarchical data [45, 84]. Its negative curvature and exponential volume growth allow it to mirror tree-like data structures, addressing the stability-plasticity trade-off [16] and supporting segmentation in unstructured contexts [27]. Unlike Euclidean spaces, hyperbolic space offers exponentially more volume [44] for accommodating incremental classes, thus reducing overlap and interference (see Fig. 2, Tab. 4). However, efficient optimization of embeddings in the hyperbolic space presents computational challenges when points are near the boundary of the Poincaré ball resulting in vanishing gradient problems [44], and without careful handling, newly added classes can still encroach upon the existing embeddings, aggravating plasticity (see Tab. 3). This highlights the need for an FSCIL framework that uses hyperbolic space embedding hierarchical data in low distortion while ensuring computational efficiency and managing class overlap. + +Our Proposal: To address these challenges, we introduce Hyperbolic Ideal Prototypes Optimization (HiPo), a novel FSCIL framework tailored for point cloud segmentation. By embedding data within the Poincaré Hyperbolic Sphere [49], HiPo leverages the unique properties of hyperbolic space to capture hierarchical structures and maintain consistent feature-classifier alignment across incremental learning sessions. Through a modified Busemann loss [7], HiPo arranges ideal prototypes orthogonally within the hyperbolic space, maximizing class separation and preserving space for new classes, which minimizes inter-class interference and optimizes the feature space for CIL. + +This design takes advantage of hyperbolic space's exponential scaling, which aligns naturally with complex hierarchies found in point cloud data [3, 27, 45, 57]. Our modified Busemann loss incorporates two components: (1) a prototype alignment term to guide class embeddings towards their respective Hyperbolic Ideal Prototypes, enhancing class separability, and intra-class compactness, and (2) an uncertainty-aware term that penalizes overfitting on novel classes and preventing vanishing gradient problems. Together, these components enable HiPo to effectively address data scarcity, mitigate catastrophic forgetting, and support robust FSCIL under episodic shifts, establishing it + +![](images/6041dc375c4e0f7391ca49c8dcf2da0eaafdb6ef4b254da7fb450ac065593299.jpg) +Figure 2. Motivation for hyperbolic embeddings: (a) Hyperbolic UMAP [40] of the HiPo embedding space, projected to 2D for the S3DIS dataset [2]. The hyperbolic space's negative curvature aligns with hierarchical data, enabling a more compact representation and clearer class separation than Euclidean space (Dunn index [18] $(\uparrow)$ for HiPo is 0.69 while for the Euclidean space, it is 0.35), showcasing better class compactness in our case. (b) UMAP of the classifier space in the Euclidean hypersphere + +as a powerful solution for incremental learning in 3D point cloud segmentation. Overall, our contributions are: + +- We define and address the previously unexplored problem of FSCIL for 3D point cloud segmentation, encompassing both in-domain and cross-dataset scenarios. +- We introduce hyperbolic space as a robust embedding framework for FSCIL in 3D segmentation, effectively capturing hierarchical structures and preserving stable feature-class relationships across incremental learning sessions. +- We propose a modified Busemann loss with prototype alignment and uncertainty-aware components to ensure optimal feature-classifier alignment with minimized uncertainty in predictions. +- Through extensive experiments on benchmark datasets S3DIS [2] and ScanNetv2 [17], HiPo demonstrates superior performance and generalization, setting a new standard for FSCIL in this domain. + +# 2. Related Works + +3D Point-Cloud Semantic Segmentation: Fully supervised 3D semantic segmentation has been a well-explored area, with methods typically requiring dense point-wise labels [26, 52, 53, 65, 67, 77]. A significant advancement was PointNet [52], which directly processes raw point clouds, removing the need for voxelization but facing challenges in capturing local structures. This limitation spurred improvements like PointNet++ [53], which introduced a hierarchical approach to enhance local feature learning, and DGCNN [65], which incorporated the EdgeConv module to integrate neighborhood information for better geometric representation. More recently, models such as PointTransformer [80] have integrated self-attention mechanisms, inspired by NLP, to elevate feature expressiveness and contextual awareness. + +Despite these advancements, the dependency on large labeled datasets and the assumption of a fixed class set restrict these methods in practical scenarios where new classes must be incrementally learned with minimal supervision. + +(Few-Shot) Class Incremental Learning: Incremental learning methods for 3D data have predominantly focused on classification tasks rather than segmentation [14, 15, 19, 35, 75]. Approaches such as I3DOL [19] and L3DOC [35] mitigate catastrophic forgetting in 3D object classification through memory and attention mechanisms, yet they struggle as labeled supervision diminishes over time. Few-Shot Class Incremental Learning (FSCIL) [61] has emerged as a more flexible solution, enabling models to learn new classes incrementally with minimal labeled data. Representative methods like CEC [76] use graph-based task linking, while FACT [81] and ALICE [50] create adaptable feature spaces through clustering and cosine similarity. However, these techniques rely on Euclidean embeddings, which have limited capability to represent hierarchical structures intrinsic to 3D data, and frequent classifier updates introduce computational overhead [76]. Open-world semantic segmentation (OWSS) [5, 9] is a related area that focuses on dynamically detecting unknown classes at inference, though our focus on incrementally adding known classes during training sets our approach apart from OWSS. + +Hyperbolic Representations for Hierarchical Data: Hyperbolic geometry has shown promise for hierarchical data representation, with foundational work by Sarkar et al. [57] and Nickel et al. [49] introducing hyperbolic neural networks and Riemannian optimization methods. Hyperbolic embeddings have demonstrated substantial benefits for capturing semantic hierarchies across applications, including image embeddings [27] and FSCIL for image classification [16]. In 3D data, recent research has shown that hyperbolic representations can model complex part-whole relationships effectively. Montanaro et al. [45] applied hyperbolic regularizers to encode hierarchical structures in 3D objects, while Hu et al. [25] leveraged hyperbolic geometry to represent latent structures in point clouds. + +Our approach leverages hyperbolic embeddings to tackle FSCIL challenges in segmentation, improving adaptability, preserving hierarchies, and ensuring computational efficiency in the point cloud domain. + +# 3. Problem Definition & Methodology + +Let $D$ represent a sequence of training datasets $\{D^{(0)}, D^{(1)}, \ldots, D^{(T)}\}$ , where each $D^{(t)} = \{(P_k^{(t)}, M_k^{(t)})_{k=1}^{|D^{(t)}|}\}$ corresponds to a training session $S^{(t)}$ , $t \in \{0, 1, \ldots, T\}$ . The initial session dataset $D^{(0)}$ serves as the base, while $T$ denotes the total number of incremental sessions. In each session $S^{(t)}$ , $P_k^{(t)}$ represents a point cloud sample, and $M_k^{(t)}$ provides its labels within + +the label space $C^{(t)} = \{c_1^{(t)},c_2^{(t)},\ldots ,c_{|C^{(t)}|}^{(t)}\}$ where $c_{i}^{(t)}$ is the $i$ -th class in the incremental session $t$ . + +The base session $S^{(0)}$ typically contains an abundant number of samples per class in the label space $C^{(0)}$ . However, in each subsequent incremental session $S^{(t)}, t > 0$ , only a limited number of labeled point clouds $P_{k}^{(t)} \in \mathbb{R}^{m \times d_{0}}$ are provided, where $m$ is the number of points per sample and $d_{0}$ is the feature dimension. Following previous works [1, 24, 72, 78, 79, 81, 85], we assume $|D^{(t)}| = N \times K$ , where $N$ is the number of classes, and $K$ is the number of samples per class, following an $N$ -way $K$ -shot setting $\forall t > 0$ . We assume $C^{(t)}$ to be disjoint from previous sessions, i.e., $C^{(t)} \cap C^{(t')} = \emptyset$ , $\forall t' < t$ . + +During each incremental session, only the current dataset $D^{(t)}$ is accessible, with data from previous sessions unavailable for training. For evaluation in session $t$ , the model is tested across all classes encountered up to that session, covering the label space $\bigcup_{i=0}^{t} C^{(i)}$ . A solution to the FSCIL problem must jointly address uncharted catastrophic forgetting, data imbalance, and class overfitting judiciously. + +# 3.1. Preliminaries + +This section mentions the essential backgrounds needed for the proposed framework Hyperbolic Ideal Prototypes Optimization (HiPo). We extensively use the concepts of the Poincaré ball model and its associated Möbius addition, exponential map, and the geodesic distance. We detail their definition and discuss their formulae in the Sup. Mat. + +In the Poincaré ball model of hyperbolic geometry, all the points at the boundary of the Poincaré ball represent points at infinity. We initialize a group of classifier prototypes at the boundary of the Poincaré ball termed Ideal Prototypes, following [23], for each of the $n$ individual classes. Assuming $u_{1}, u_{2}, \dots, u_{n}$ to be the Ideal Prototypes for $n$ classes, they form the boundary of the $n$ -dimensional Poincaré ball of curvature $c$ , $\mathbb{B}_c^n$ , following [23]: + +$$ +\mathcal {I} _ {n} = \left\{u \in \mathbb {R} ^ {n}: u _ {1} ^ {2} + u _ {2} ^ {2} + \dots + u _ {n} ^ {2} = 1 \right\} \tag {1} +$$ + +Since Ideal Prototypes are points at an infinite geodesic distance from all other points in $\mathbb{B}_c^n$ , we use the Busemann function [7] as a distance measure since computing direct infinite geodesic distance is non-trivial [23]. + +The pre-assigned Ideal Prototypes form an optimal geometric structure by fixing the target throughout the incremental learning. However, as already pointed out, optimizing Ideal Prototypes in the Poincaré model to address FS-CIL in 3D point cloud segmentation remains unexplored. + +# 3.2. Discussing HiPo + +Our model introduces a fixed alignment strategy for feature-classifier relationships in FSCIL, tailored for the hierarchical nature of data such as 3D point clouds. Traditional Euclidean embeddings suffer from compaction [50, 76, 81] + +![](images/63cf5ba78a7e0238dbb7cc650334e3c091fd8659e767f93bccc19eb1944a7509.jpg) +Figure 3. HiPo Overview: We show the changes in the Poincaré ball during the initial three sessions in the 3-2T FSCIL setting (three base classes, two incremental classes per session) for the S3DIS dataset [2], designed to easily visualize the working of HiPo. Hyperbolic Ideal Prototypes are optimally positioned to increase the margin of semantically similar classes $(\mathcal{L}_{\mathrm{HIP}})$ . As an example, the chair class gets positioned farther from sofa than from table or clutter. We align representations of the newly introduced classes with their corresponding ideal prototypes $(\mathcal{L}_{\mathrm{Alignment}})$ regularized by a modified Busemann function $(\mathcal{L}_{\mathrm{Uncertainty}})$ . Training is conducted using $\mathcal{L}_{\mathrm{HiPo}}$ across sessions. + +![](images/0ca66a39feff0e83af01997e2c0c0cc6be61082c8c5238757cd6bcf5c77efcd5.jpg) + +![](images/8792de0dd8c23bfe0c414fffff5e860668178ff2bc3fd3c05c0d29f4f04bdfc6.jpg) + +and distortion as new classes are added, leading to overlap and forgetting. Hyperbolic space, with its negative curvature and exponential volume, provides a natural fit for preserving hierarchical structures while minimizing inter-class interference. To address catastrophic forgetting, our model anchors feature-classifier relationships within hyperbolic space, ensuring stable alignment between features and classifier prototypes. This fixed alignment prevents the drift of previously learned embeddings as new classes are introduced, preserving the original classifier prototype positions and maintaining consistency across incremental sessions. By leveraging hyperbolic geometry, our approach enables efficient class separation, robust memory retention, and adaptability, making it well-suited for FSCIL in high-dimensional, unstructured data settings. Our method has two main stages: + +- Defining Ideal Prototypes: We initialize ideal prototypes for each class on the boundary of $\mathbb{B}_c^n$ , positioning them mutually orthogonal and at maximal distances to enhance discriminative power, especially among classes with similar characteristics (see Sec. 3.2.1). This fixed configuration allows a dedicated space for each class, preventing the embedding drift observed in Euclidean spaces. +- Efficient Projections onto Poincaré ball $\mathbb{B}_c^n$ : In Hyperbolic embedding space, mean class features are obtained by computing Fréchet [20] means over the feature embedding for the training examples belonging to the classes. Instead of directly computing class features on the Poincaré ball—an intensive task[63]—we obtain the feature vectors in Euclidean space for every class in each session and then project them into hyperbolic space using exponential map $\exp_0(.)$ [45] (see Sec. 3.2.2). The hyperbolic classifier head $\mathcal{H}(.)$ consisting of a Möbius layer $\mathcal{M}$ (also known as hyperbolic feed forward layer [22]), takes the output of the + +exponential layer as input and projects mean class features to $\mathbb{B}_c^n$ . The Poincaré ball addresses few-shot learning using the sample mean prototype as the class representation. + +Our framework consists of a point-level classification branch, combining a feature extractor $\mathcal{E}(.)$ and hyperbolic classifier head $\mathcal{H}(.)$ , and is trained with two key strategies: + +Feature-Classifier Alignment: For each session $t \geq 0$ , we align the hyperbolic sample mean features (see Eqn. 4) with the Hyperbolic Ideal Prototypes of their respective classes using alignment loss, $\mathcal{L}_{\mathrm{Alignment}}$ . We define $\mathcal{L}_{\mathrm{Alignment}}$ using Busemann function $\mathcal{B}$ and introduce a regularization term, quantifying uncertainty of the class features, to prevent feature collapse around the ideal prototypes, effectively reducing interference with previously seen classes (see Sec. 3.2.3) and vanishing gradient problems [44]. Point-Level Classification: The hyperbolic classifier $\mathcal{H}(.)$ , consisting of a Möbius layer $\mathcal{M}$ followed by softmax [45], processes the output of $\exp_0(\mathcal{E}(.))$ , with a cross-entropy loss for 3D point cloud segmentation. The joint objective combines this alignment loss $\mathcal{L}_{\mathrm{Alignment}}$ and cross-entropy loss $\mathcal{L}_{\mathrm{CE}}$ as follows: + +$$ +\mathcal {L} _ {\mathrm {H I P O}} = \mathcal {L} _ {\text {A l i g n m e n t}} + \mathcal {L} _ {\mathrm {C E}} \tag {2} +$$ + +This approach ensures robust feature-classifier alignment across incremental learning sessions, addressing catastrophic forgetting in 3D point cloud segmentation by preserving both old and new class representations in a rehearsal-free manner. The process is illustrated in Fig. 3, and Algorithm 1 and a table summarizing the important variables used, mentioned in the Sup. Mat., supplement these discussions. + +# 3.2.1. Hyperbolic ideal prototypes positioning + +Building upon the goal of maintaining stable feature-classifier alignment, our framework pre-assigns classifier + +prototypes on the boundary of $\mathbb{B}_c^n$ to define a structured classifier space prior to backbone training. To achieve maximal margin separation between classes, we introduce inductive biases by assigning each class a Hyperbolic Ideal Prototype, distributing these prototypes as uniformly as possible across the Poincaré ball. Inspired by Mettes et al. [41], we incorporate semantic class knowledge to enforce distinct separations, particularly for closely related classes, such as "chair" and "sofa". To further enhance semantic cohesion, we utilize the CLIP framework [54] due to its robust semantic representations [31, 73] in comparison to embeddings like word2vec [43], Glove [51] etc. By querying CLIP's frozen text encoder with the prompt "A photo of a [class]", we obtain text embeddings $S = \{s_1,\dots ,s_n\}$ for each of the $n$ predefined classes. Leveraging the homeomorphism between hyperbolic space $\mathbb{B}_c^n$ and the hypersphere $\mathbb{S}^n$ [23], we maximize separation between nearest neighboring embeddings in $\mathbb{S}^n$ and enforce mutual orthogonality among Ideal Prototypes to mitigate conflict between new and prior knowledge. To achieve this, inspired by RankNet [6], we formulate a Hyperbolic positioning loss, $\mathcal{L}_{\mathrm{HIP}}$ , optimizing the ideal prototypes positions as: + +$$ +\begin{array}{l} \mathcal{L}_{\text{HIP}} = \frac{1}{|M|}\sum_{(i,i^{\prime},i^{\prime \prime})\in M} - \bar{W}_{ii^{\prime}i^{\prime \prime}}\log W_{ii^{\prime}i^{\prime \prime}} - \\ (1 - \bar {W} _ {i i ^ {\prime} i ^ {\prime \prime}}) \log (1 - W _ {i i ^ {\prime} i ^ {\prime \prime}}) \tag {3} \\ \end{array} +$$ + +where $M$ represents all class triplets, and $\bar{W}_{ii^{\prime}i^{\prime \prime}} = \llbracket \cos \theta_{\mathbf{s}_i,\mathbf{s}_{i^{\prime}}} \geq \cos \theta_{\mathbf{s}_i,\mathbf{s}_{i^{\prime \prime}}} \rrbracket$ denotes triplet ranking order, with $W_{ii^{\prime}i^{\prime \prime}} = \frac{e^{o_{ii^{\prime}}i^{\prime \prime}}}{1 + e^{o_{ii^{\prime}}i^{\prime \prime}}} \text{ and } o_{ii^{\prime}i^{\prime \prime}} = \cos \theta_{\mathbf{u}_i,\mathbf{u}_{i^{\prime}}} - \cos \theta_{\mathbf{u}_i,\mathbf{u}_{i^{\prime \prime}}}$ , where $\mathbf{u}_i, \mathbf{u}_{i^{\prime}}, \mathbf{u}_{i^{\prime \prime}}$ are ideal prototypes for the corresponding classes $i, i^{\prime}$ and $i^{\prime \prime}$ respectively. + +Following Mettes et al. [41], $\mathcal{L}_{\mathrm{HIP}}$ optimizes the prototypes' placement on $\mathbb{B}_c^n$ by aligning the hyperbolic prototypes with semantic ranking. This optimization establishes a classifier structure that facilitates inter-class separability and mutual orthogonality. Hyperbolic space's exponentially larger embedding capacity [44] further ensures that new classes integrate seamlessly without distorting existing embeddings, promoting robust feature organization across learning episodes. + +# 3.2.2. Hyperbolic class prototype projection + +We model the hierarchical complexity in 3D point clouds using a hyperbolic manifold representation, which preserves nuanced class relationships with minimal distortion. + +At each session $t \geq 0$ , we apply the feature extractor $\mathcal{E}(:, \theta)$ on input point clouds $\{(P_k^{(t)})_{k=1}^{|D^{(t)}|}\}$ to obtain feature vectors $\pmb{v}$ in Euclidean space. To avoid the computational cost of calculating Fréchet means directly in hyperbolic space [63], we compute class-mean feature vectors in Euclidean space and project them onto the hyperbolic manifold $\mathbb{B}_c^n$ using $\exp_0$ and $\mathcal{M}$ as follows: + +$$ +v _ {k} ^ {(t)} = \mathcal {E} \left(P _ {k} ^ {(t)}; \theta\right), \quad \forall k \in \{1, \dots , | D ^ {(t)} | \} +$$ + +$$ +\begin{array}{l} \bar {v} \left(c _ {i} ^ {(t)}\right) = \frac {1}{| D ^ {(t)} |} \sum_ {k = 1} ^ {| D ^ {(t)} |} v _ {k} ^ {(t)} \mathbb {1} \left[ M _ {k} ^ {(t)} = c _ {i} ^ {(t)} \right] \tag {4} \\ z (c _ {i} ^ {(t)}) = \mathcal {M} (\exp_ {0} (\overline {{v}} (c _ {i} ^ {(t)}))) \\ \end{array} +$$ + +where $i = \{1,\dots ,|C^{(t)}|\}$ and $\theta$ denotes learnable parameters. Here, $z(c_{i}^{(t)})$ represents the hyperbolic prototype for each class $c_{i}^{(t)}\in C^{(t)}$ within the open ball $\mathbb{B}_c^n = \{x\in$ $\mathbb{R}^n:c\| x\| < 1\}$ . This projection process ensures cohesive, low-distortion representations of class structures as the model adapts incrementally to new classes. + +# 3.2.3. Uncertainty-aware Busemann learning + +Previous works [12, 21, 27] highlight the utility of hyperbolic distance from any point to the origin $\mathcal{O}_{\mathbb{B}}$ in $\mathbb{B}_c^n$ as a measure of uncertainty, providing motivation to utilize the hyperbolic manifold for robust segmentation. In our approach, we align class prototypes with Ideal Prototypes placed at infinity on the hyperbolic boundary. To refine this alignment, we penalize the uncertainty of projected class features $z(c_i^{(t)})$ , represented by their distance $d_{\mathbb{B}}(z(c_i^{(t)}),\mathcal{O}_{\mathbb{B}})$ , thereby preventing embedding drift from $\mathcal{O}_{\mathbb{B}}$ and ensuring stability. + +These Ideal Prototypes lie at an infinite geodesic distance from all points within $\mathbb{B}_c^n$ , making direct prototypical learning impractical. We address this with the Busemann function $\mathcal{B}$ introduced in [7]. Ghadimi et al. [23] provides a closed-form expression for measuring such infinite geodesic distances. Let $u_{i}$ be the Ideal Prototype for class $c_{i}^{(t)}$ from the set $\mathcal{I}_n$ , and $\gamma_{u_i^t}$ a geodesic ray towards $u_{i}^{t}$ , with $z_{i}^{t} = z(c_{i}^{(t)})$ . The Busemann function $\mathcal{B}$ for a hyperbolic embedding $z_{i}^{t}\in \mathbb{B}_{c}^{n}$ relative to $u_{i}^{t}$ is defined as: + +$$ +\mathcal {B} _ {u _ {i} ^ {t}} \left(z _ {i} ^ {t}\right) = \lim _ {q \rightarrow \infty} \left(d _ {\mathbb {B}} \left(\gamma_ {u _ {i} ^ {t}} (q), z _ {i} ^ {t}\right) - q\right) \tag {5} +$$ + +In the Poincaré ball $\mathbb{B}_c^n$ , this simplifies to: + +$$ +\mathcal {B} _ {u _ {i} ^ {t}} \left(z _ {i} ^ {t}\right) = \log \frac {\left\| u _ {i} ^ {t} - z _ {i} ^ {t} \right\| ^ {2}}{1 - \left\| z _ {i} ^ {t} \right\| ^ {2}}. \tag {6} +$$ + +We also incorporate hyperbolic distance as an uncertainty measure for class prototypes in $\mathbb{B}_c^n$ , which integrates into our loss function as in [23]. The alignment loss introduced in Eq. 2 is defined as: + +$$ +\mathcal {L} _ {\text {A l g i n m e n t}} = \mathcal {L} _ {\mathcal {B}} + \phi \mathcal {L} _ {\text {U n c e r t a i n t y}} \tag {7} +$$ + +where Busemann loss $\mathcal{L}_{\mathcal{B}} = \mathcal{B}_{u_i^t}(z_i^t)$ , uncertainty loss $\mathcal{L}_{\mathrm{Uncertainty}} = d_{\mathbb{B}}(z_i^t,\mathcal{O}_{\mathbb{B}})$ and $\phi$ is the strength of the regularization effect of uncertainty loss. + +The overall loss function in Eq. 2 can be formulated as: + +$$ +\mathcal {L} _ {\mathrm {H i P o}} = \mathcal {L} _ {\mathcal {B}} + \phi \mathcal {L} _ {\text {U n c e r t a i n t y}} + \mathcal {L} _ {\mathrm {C E}} +$$ + +At test time, we employ only the point-level classification branch to segment 3D point clouds. + +# 4. Experimental Evaluations + +Datasets: We evaluate our method on two widely used benchmarks, S3DIS [2] and ScanNetv2 [17], to assess both in-domain and cross-dataset FSCIL performance. These datasets incorporate large-scale coloured 3D point clouds of indoor spaces which introduce semantics. See Sup. Mat. for additional dataset details. For cross-dataset FSCIL, we create a transition dataset from ScanNetv2 to S3DIS, using ten classes from ScanNetv2 as the base and introducing eight new classes from S3DIS in incremental sessions. Following a strict disjoint setup as in [10], each session includes only novel classes. The number of introduced classes varies per session to assess resilience to catastrophic forgetting, which escalates with more incremental sessions in point cloud recognition [60]. + +Evaluation metrics: We assess model performance using three standard metrics for incremental learning[70]: Average Incremental Accuracy (AIA), which tracks average accuracy across tasks learned sequentially; Last, the accuracy after the final task; and Average Forgetting Rate $(\mathcal{F}_{Last}^{(T)})$ , which measures performance decline on learned tasks after adapting to the last task. Finally, we report mean Intersection-over-Union $(\mathrm{mIoU})^1$ to evaluate segmentation performance. See Sup. Mat. for detailed metric definitions. Architecture: Our backbone model employs PointTransformer [77], leveraging recent advances in point cloud recognition, with a Hyperbolic Classifier head $\mathcal{H}(.)$ for segmentation. We train HiPo using the Riemannian SGD optimizer, facilitated by geoopt [29] for hyperbolic operations. Training is conducted on both $D^{(0)}$ and incremental datasets $D^{(t)}, t > 0$ with a batch size of 16, AdamW optimizer [37], initial learning rate of 0.001, and weight decay of 0.0001 over 100 epochs. Architecture diagram and further hyper-parameter details are available in the Sup. Mat. + +Competitors: We compare our method against various baselines representing distinct FSCIL approaches. For regularization-based methods, we include LwF with Teacher Adaptation [59], and for optimization-based methods, we include C-FSCIL [24]. We also evaluate against regularization-based FSCIL methods specifically designed for 3D point cloud tasks, such as LGKD [72], SDCOT [79], and 3DPC-CISS [70], which have shown effectiveness with the PointTransformer [77] architecture. Comparative results are provided in Tables 1 and 2. Additionally, we adapt BiDist [78], CLOM [85], TEEN [64], FACT [81], and ALICE [50] using the DGCNN [65] backbone for FSCIL in 3D point cloud segmentation. We report the results with the DGCNN [65] backbone in Table 5. + +![](images/038eecf366649802e0a09bacdf7996f2bab9e045ce324bb210c550db31fd3e84.jpg) + +![](images/bc810bb5e6235004291570e9db32336890bef6190cca36c1fa7e35e6be343cac.jpg) +Ground Truth + +![](images/44ff4e68a9d72ff3fdb4160813e4730863929139856510e83ac99345f0c25498.jpg) +SDCOT + +![](images/d7fc9e4b09a38c2727df04c1fe888dddbb912a2b587fad7ebc72836b2a6de59d.jpg) +3DPC-CISS + +![](images/b4a33ea4b4a1916b2987e9fc716386469aa84c4a1d892fb72cbf413380368827.jpg) +HIPO (Ours) + +![](images/2a70add8cc859c470ac47d259adb1ea9bca6c7cb6927eb57925fc7637af18590.jpg) + +![](images/66d2ace25701a2759e6ba19ce575d48a4d05f98a5f98cc608858b0f92cb07b5c.jpg) + +![](images/689ae9ddddceac5aaac41b4ffc222299d136daaa8f43c0ef6121c9784d3e1511.jpg) + +![](images/5ccdbabf1fceb3a3a0510900afd8dff4b950675bf7bfcdfe15c3973febd84392.jpg) + +![](images/0d2f5fa6adbc042bddbb0ca62dc9d356fd172f98a6a53a316d878a581085b767.jpg) + +![](images/7202c6e0419e65a485e5c5e8d5dd359cd6a5ef5f67e58bf8b93c5c637ac8d2c6.jpg) + +![](images/3cddb91826033ec2865b8ab68a286af7a82bd16390d1daa8bd396a2106736f56.jpg) + +![](images/8438730be9ff05adc0e539f3f01f920f5da7ba88dc280bf51a999211acf7d799.jpg) + +![](images/94f742444df084e76f69c4bde7648aa71225bb834ceca96684ec01684f1374e1.jpg) + +![](images/af36710e687a2e9e34fcb8e9532ec24155b01b7fb1c2240f5e3122e7a91e23f1.jpg) + +![](images/906504720ac575679d1fb765e766e279b1ae7c0d78c833f9442d0b5ff82d6860.jpg) + +![](images/f0b048e14c5365325ffc22e2185a71b48329abe1b96152e3aa7a1a0b03339530.jpg) + +![](images/76ffc8307d09e60aa534b7d7ac1a25336137a917f3fd0211a30befcfcb96ee22.jpg) + +![](images/76e131ab2723234f675ca50c055ace3bb135891c754cd08860dbec1ef118dad5.jpg) + +![](images/091f65815f6411c69dfe721f0f41087fdc9561bade3887c56052204cb3ef4a74.jpg) + +![](images/74c8fd02f2dea6ca4287fad05266344d10aa971dc7ee7555ddcfe60b010567cc.jpg) +Figure 4. Qualitative visualization of model performance: Comparing FSCIL for 3D point cloud segmentation results of our method (HiPo) and other baselines for S3DIS dataset [2]. We observe that even with a few-shot setup, our model performs reasonably well in segmenting out small parts like walls, counters, etc., compared to other methods [70, 79]. + +![](images/5f6c35fe1cab7a75bf8ec7ac1ebd79287042afe875ce6f02e870616e329ae0cd.jpg) + +# 4.1. Discussing the main results + +Table 1 presents a detailed comparison of our proposed method, HiPo, against several baseline methods on the S3DIS [2] and ScanNetv2 [17] datasets across various evaluation settings. HiPo consistently outperforms the current state-of-the-art approach, 3DPC-CISS [70], achieving an average improvement of $+2.57\%$ mIoU on S3DIS and $+2.12\%$ on $\mathcal{F}_{Last}^{(T)}$ . This improvement underscores the effectiveness of HiPo's hyperbolic, uncertainty-aware architecture, which mitigates catastrophic forgetting by reducing misalignment between features and prototype representations from earlier sessions. When compared to other baselines, such as LGKD [72] and C-FSCIL [24], HiPo shows substantial gains across all metrics. For instance, in the 10-3T scenario, HiPo achieves $33.40\%$ mIoU on the Last metric, significantly higher than LGKD's $20.02\%$ and C-FSCIL's $25.02\%$ . This consistent superiority highlights HiPo's resilience against forgetting, particularly in incremental sessions with larger class increments, where other methods struggle to retain performance. Furthermore, while SDCOT [79] and LwF [59] suffer from high forgetting rates, HiPo effectively maintains stable performance across incremental tasks, with a $\mathcal{F}_{Last}^{(T)}$ rate as low as $17.70\%$ in challenging scenarios, compared to SDCOT's $30.83\%$ . This robustness can be attributed to HiPo's hyperbolic embed + +
4-3T7-3T10-3TAverage
Last(↑)AIA(↑)FT(Last)(↓)Last(↑)AIA(↑)FT(Last)(↓)Last(↑)AIA(↑)FT(Last)(↓)Last(↑)AIA(↑)FT(Last)(↓)
LwF [59]4.3322.0523.628.6021.8019.8021.3034.0525.5011.4125.3022.97
C-FSCIL [24]9.0627.9025.1120.0131.9217.8625.0235.3620.6818.0331.0621.88
LGKD [72]12.2431.8025.9419.7630.4115.9720.0232.7125.3817.3431.6422.43
SDCOT [79]5.3028.4230.8312.0029.4626.2023.8035.3023.0013.7031.0626.67
3DPC-CISS [70]25.7040.5019.7324.2034.5015.4529.6038.2017.2026.5037.7317.46
HtPo30.6043.8717.7029.4036.8011.1033.4040.2513.7031.1340.3014.16
+ +
10-2T10-5T15-5TAverage
Last(↑)AIA(↑)FT(Last)(↓)Last(↑)AIA(↑)FT(Last)(↓)Last(↑)AIA(↑)FT(Last)(↓)Last(↑)AIA(↑)FT(Last)(↓)
LwF [59]1.7010.7510.861.0517.5324.725.0621.3832.642.6016.5522.74
C-FSCIL [24]7.1215.9410.5811.1224.1819.5911.5023.5824.159.2521.2318.11
LGKD [72]6.0215.6011.509.5525.5222.4511.1023.8325.468.2221.6519.80
SDCOT [79]0.509.5610.880.4017.8326.151.3019.5036.400.7315.6324.47
3DPC-CISS [70]6.2013.138.328.1022.3021.3012.2024.9525.508.8320.1218.37
HtPo8.5017.1310.3614.6527.2218.8511.9024.4525.1011.6822.2718.12
+ +Table 1. Comparison with literature on the S3DIS [2] (top table) and ScanNetv2 [17] (bottom table) datasets, using the PointTransformer backbone [77] in a 5-shot setting. CIL mIoU (%) is reported. “ $M-NT$ ” indicates that $M$ is the number of classes in the base session $S^{(0)}$ , while $N$ is the number of classes introduced in each incremental session $S^{(t)}$ , $t > 0$ . The highest performance in each column is shown in bold, with the second-highest results underlined. HiPO performs superior compared to other baselines in different incremental settings. + +
Average
Last (↑)AIA (↑)F(T)Last (↓)
LwF [59]8.7822.6023.94
C-FSCIL [24]13.5525.0219.72
LGKD [72]15.6325.5617.51
SDCOT [79]19.2229.2917.52
3DPC-CISS [70]19.9529.1616.01
HtPo20.6029.8116.05
+ +Table 2. SOTA comparison on ScanNetv2 $\rightarrow$ S3DIS dataset. We compute the average of 10-4T and 10-8T with PointTransformer backbone [77] in 5-shot setting. HIPO achieves superior Last and AIA values with comparable $F_{Last}^{(T)}$ values, highlighting better domain adaptability than other methods. + +ding space, which aligns naturally with hierarchical structures, thereby enhancing inter-class separation and intraclass compactness over time. + +Moreover, Fig. 4 provides qualitative comparisons, where HiPO demonstrates marked improvements over leading baselines, particularly in accurately segmenting challenging structures such as walls and counters. Further analyses and results are available in the Sup. Mat. + +# 4.2. Main ablation analysis + +Cross-dataset class-incremental segmentation: To further assess the effectiveness of our proposed modules, we conduct cross-dataset incremental segmentation experiments (ScanNetv2 $\rightarrow$ S3DIS). Detailed setup information is in Sup. Mat. As shown in Table 2, the cross-dataset setting introduces a notable increase in forgetting due to domain shift. However, the results underscore HiPo's robustness in managing such domain shifts effectively, with improvements of $0.65\%$ in Last and $0.52\%$ in AIA metrics. + +Ablation studies on penalty in Busemann loss: We pro + +vide a comprehensive ablation study on the regularization constant $\phi$ introduced in Eq. 7. The regularization strength is varied from $\phi = 0$ to $\phi = 1$ in intervals of 0.25 for S3DIS [2] (4-3T, 7-3T, 10-3T) and ScanNetv2 [17] (10-2T, 10-5T, 15-5T). As shown in Table 3, our findings corroborate prior work [23]: without a penalty $(\phi = 0)$ , samples are overly drawn towards the Ideal Prototypes, resulting in overconfident predictions and lower AIA on test data $\mathcal{D}_{test}$ in addition to vanishing gradient problems [44]. Conversely, a high penalty $(\phi = 1)$ increases inter-class confusion, also leading to lower AIA on $\mathcal{D}_{test}$ . Our final chosen results, highlighted in bold in Table 3, reflect the optimal balance for $\phi$ . + +Comparison of metrics at different curvatures: Following [47], we experimented with curvature values of $c = 0.0001$ and $c = 0.001$ to implement traditional Euclidean architectures (Poincaré Ball with zero curvature; very small Poincaré ball curvature values are chosen as curvature $c = 0$ is numerically unstable [27]) for the S3DIS [2] (4-3T, 7-3T, 10-3T) and ScanNetv2 [17] (10-2T, 10-5T, 15-5T) setups. Table 4 shows that our HiPO with a Poincaré Ball curvature of $c = -1.00$ consistently outperforms its Euclidean counterpart across most settings. This finding demonstrates that Riemannian geometry leverages the hierarchical structure of point clouds more effectively than Euclidean space, suggesting further exploration in this domain. + +Sensitivity to the different number of training samples in each incremental episode: To assess the impact of data availability on model performance, we gradually increase the number of samples per class (shot) from one to all available examples and analyze the Last mIoU and $\mathcal{F}_{Last}^{(T)}$ results for different shot values in Fig. 5. Our findings indi + +
DatasetsSettingsφ = 0.00φ = 0.25φ = 0.50φ = 0.75φ = 1.00
AIA↑FT(St)↓AIA↑FT(St)↓AIA↑FT(St)↓AIA↑FT(St)↓AIA↑FT(St)↓
S3DIS [2]4-3T42.8717.7043.4017.8643.2718.2343.8717.7043.2218.96
7-3T36.1311.6036.4611.2036.8011.1036.7011.1036.0612.10
10-3T39.9514.3040.2513.7040.1014.0040.0714.0540.0514.10
ScanNetv2 [17]10-2T16.7510.5116.8510.5016.9010.0817.1310.3616.9110.76
10-5T27.0818.8727.1318.8727.2218.8527.2118.8527.1618.92
15-5T24.2525.5024.4525.1024.0525.9023.9526.1024.0026.00
+ +Table 3. Different penalty values in Busemann loss in 5-shot incremental segmentation across the S3DIS [2] dataset and ScanNetV2 [17] dataset, highlighting the regularization benefits introduced by the hyperbolic distance of the projected class prototypes from the origin. + +
DatasetsSettingsc = -1.00 (Ours)c = 0.0001c = 0.001
AIA↑FTLast↓AIA↑FTLast↓AIA↑FTLast↓
S3DIS [2]4-3T43.8717.7042.9519.9343.2317.38
7-3T36.8011.1036.2311.6036.8611.05
10-3T40.2513.7040.2013.8040.2213.75
ScanNetv2 [17]10-2T17.1310.3616.5410.2516.9810.18
10-5T27.2218.8527.1619.0026.8319.92
15-5T24.4525.1024.4025.2024.2025.60
+ +![](images/71fd6899a574728d21bd83abf494721a899036eef8cec22063294d65748d6cb4.jpg) +(a) Last mIoU values + +![](images/0797ba7c14ea84bc856ebcc361ddc407a88046c5c5edd3502ed5da5089f9942e.jpg) +(b) Forgetting $\mathcal{F}_{\text{Last}}^{(T)}$ +Figure 5. mIoU improves and forgetting reduces with more training examples. Our proposed method outperforms the top-2 baseline methods ([70, 79]) in terms of Last mIoU and forgetting mitigation in the 10-2T setting (Full-data: fully supervised training in incremental fashion). + +cate that catastrophic forgetting is particularly severe with limited data; however, our method substantially reduces forgetting compared to other baselines. Notably, increasing the number of training examples significantly mitigates forgetting for H1Po, while it exacerbates forgetting for SDCOT. + +Experiments with a different backbone: Besides PointTransformer [77], we evaluate our model with the backbone of DGCNN [65]. Unlike current SOTA FSCIL methods—BiDist [78], CLOM [85], TEEN [64], FACT [81], and ALICE [50]—which perform well with DGCNN but degrade with PointTransformer, our method remains effective across both architectures. As shown in Table 5, our approach achieves consistent gains with DGCNN, outper + +Table 4. Comparison for different curvature values in 5-shot incremental segmentation across S3DIS [2] and ScanNetV2 [17] datasets. This highlights the superior stability-plasticity trade-off of hyperbolic embedding space with negative curvature, where small positive curvature approximates Euclidean space. + +
MethodLast (↑)AIA (↑)FTLast (↓)
LwF [59]9.4526.7524.41
C-FSCIL [24]20.2230.6422.15
LGKD [72]18.6329.1615.75
SDCOT [79]22.7524.5419.94
CLOM [85]21.1834.6718.83
TEEN [64]22.7133.9015.37
FACT [81]24.2726.7021.26
3DPC-CISS [70]22.0324.7419.36
BiDist [78]27.7039.1415.85
HiPo29.9039.8613.87
+ +Table 5. Performance comparison on the S3DIS dataset [2] using the DGCNN backbone [65] in the 5-shot setting. Reported metrics include Last, $\mathcal{F}_{Last}^{(T)}$ , and AIA for 4-3T and 7-3T settings. HiPo outperforms all baseline methods with DGCNN. + +forming baselines with $+2.20\%$ mIoU on Last, $+0.72\%$ on AIA, and reducing forgetting by $+1.98\%$ on $\mathcal{F}_{Last}^{(T)}$ . This demonstrates the model-agnostic nature of HiPO. + +# 5. Conclusions & Future Directions + +We introduce a previously unexplored problem setting along with a novel solution specifically designed for few-shot class-incremental semantic segmentation of 3D point cloud data, addressing a critical gap in the literature. Recognizing the unique challenges posed by point clouds, we identified limitations in current FSCIL methods and developed HiPo, a model that leverages hyperbolic space to embed point clouds while achieving an improved stability-plasticity trade-off. Our approach consistently outperforms state-of-the-art baselines, demonstrating that hyperbolic geometry's ability to represent hierarchical structures offers a more effective framework for FSCIL in 3D data. HiPo is particularly valuable for real-world applications like autonomous driving, industrial inspection, urban planning, and environmental monitoring, where generalization to unseen classes is essential. Furthermore, our approach enhances safety and efficiency in critical areas such as disaster response and infrastructure maintenance by enabling accurate, timely semantic segmentation of complex 3D environments. Future work will focus on managing possible outliers that may emerge without prior reference across incremental sessions, further strengthening our robustness. + +# References + +[1] Noor Ahmed, Anna Kukleva, and Bernt Schiele. Orco: Towards better generalization via orthogonality and contrast for few-shot class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28762-28771, 2024. 1, 3 +[2] Iro Armeni, Ozan Sener, Amir R Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3d semantic parsing of large-scale indoor spaces. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1534-1543, 2016. 2, 4, 6, 7, 8 +[3] Mina Ghadimi Atigh, Julian Schoep, Erman Acar, Nanne Van Noord, and Pascal Mettes. Hyperbolic image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4453-4462, 2022. 2 +[4] Eden Belouadah and Adrian Popescu. II2m: Class incremental learning with dual memory. In Proceedings of the IEEE/CVF international conference on computer vision, pages 583-592, 2019. 1 +[5] Abhijit Bendale and Terrance Boult. Towards open world recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1893-1902, 2015. 3 +[6] Chris Burges, Tal Shaked, Erin Renshaw, Ari Lazier, Matt Deeds, Nicole Hamilton, and Greg Hullender. Learning to rank using gradient descent. In Proceedings of the 22nd International Conference on Machine Learning, page 89-96, New York, NY, USA, 2005. Association for Computing Machinery. 5 +[7] Herbert Busemann. The geometry of geodesics. Courier Corporation, 2005. 2, 3, 5 +[8] Francisco M Castro, Manuel J Marín-Jiménez, Nicolas Guil, Cordelia Schmid, and Karteek Alahari. End-to-end incremental learning. In Proceedings of the European conference on computer vision (ECCV), pages 233-248, 2018. 1 +[9] Jun Cen, Peng Yun, Shiwei Zhang, Junhao Cai, Di Luan, Mingqian Tang, Ming Liu, and Michael Yu Wang. Openworld semantic segmentation for lidar point clouds. In European Conference on Computer Vision, pages 318-334. Springer, 2022. 3 +[10] Fabio Cermelli, Massimiliano Mancini, Samuel Rota Bulo, Elisa Ricci, and Barbara Caputo. Modeling the background for incremental learning in semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9233-9242, 2020. 6 +[11] Arslan Chaudhry, Puneet K Dokania, Thalaiyasingam Ajthan, and Philip HS Torr. Riemannian walk for incremental learning: Understanding forgetting and intransigence. In Proceedings of the European conference on computer vision (ECCV), pages 532-547, 2018. 1 +[12] Bike Chen, Wei Peng, Xiaofeng Cao, and Juha Röning. Hyperbolic uncertainty aware semantic segmentation. IEEE Transactions on Intelligent Transportation Systems, 2023. 5 +[13] Ali Cheraghian, Shafin Rahman, Pengfei Fang, Soumava Kumar Roy, Lars Petersson, and Mehrtash + +Harandi. Semantic-aware knowledge distillation for few-shot class-incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2534-2543, 2021. 1 +[14] Townim Chowdhury, Mahira Jalisha, Ali Cheraghian, and Shafin Rahman. Learning without forgetting for 3d point cloud objects. In Advances in Computational Intelligence: 16th International Work-Conference on Artificial Neural Networks, IWANN 2021, Virtual Event, June 16-18, 2021, Proceedings, Part I 16, pages 484-497. Springer, 2021. 3 +[15] Townim Chowdhury, Ali Cheraghian, Sameera Ramasinghe, Sahar Ahmadi, Morteza Saberi, and Shafin Rahman. Few-shot class-incremental learning for 3d point cloud objects. In European Conference on Computer Vision, pages 204-220. Springer, 2022. 3 +[16] Yawen Cui, Zitong Yu, Wei Peng, Qi Tian, and Li Liu. Rethinking few-shot class-incremental learning with open-set hypothesis in hyperbolic geometry. IEEE Transactions on Multimedia, 2023. 2, 3 +[17] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 2, 6, 7, 8 +[18] David L. Davies and Donald W. Bouldin. A cluster separation measure. IEEE Transactions on Pattern Analysis and Machine Intelligence, PAMI-1(2):224-227, 1979. 2 +[19] Jiahua Dong, Yang Cong, Gan Sun, Bingtao Ma, and Lichen Wang. I3dol: Incremental 3d object learning without catastrophic forgetting. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6066-6074, 2021. 3 +[20] DC Dowson and BV666017 Landau. The frechet distance between multivariate normal distributions. Journal of multivariate analysis, 12(3):450-455, 1982. 4 +[21] Luca Franco, Paolo Mandica, Konstantinos Kallidromitis, Devin Guillory, Yu-Teng Li, Trevor Darrell, and Fabio Galasso. Hyperbolic active learning for semantic segmentation under domain shift. arXiv preprint arXiv:2306.11180, 2023.5 +[22] Octavian-Eugen Ganea, Gary Bécigneul, and Thomas Hofmann. Hyperbolic convolutional neural networks. In Advances in Neural Information Processing Systems, pages 4851-4861, 2018. 4 +[23] Mina Ghadimi Atigh, Martin Keller-Ressel, and Pascal Mettes. Hyperbolic busemann learning with ideal prototypes. Advances in Neural Information Processing Systems, 34:103-115, 2021. 3, 5, 7 +[24] Michael Hersche, Geethan Karunaratne, Giovanni Cherubini, Luca Benini, Abu Sebastian, and Abbas Rahimi. Constrained few-shot class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9057-9067, 2022. 1, 3, 6, 7, 8 +[25] Naiwen Hu, Haozhe Cheng, Yifan Xie, Pengcheng Shi, and Jihua Zhu. Hyperbolic image-and-pointcloud contrastive learning for 3d classification. arXiv preprint arXiv:2409.15810, 2024. 2, 3 +[26] Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, and Andrew Markham. + +Randla-net: Efficient semantic segmentation of large-scale point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11108-11117, 2020. 1, 2 +[27] Valentin Khrulkov, Leyla Mirvakhabova, Evgeniya Ustinova, Ivan Oseledets, and Victor Lempitsky. Hyperbolic image embeddings. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6418-6428, 2020. 2, 3, 5, 7 +[28] Do-Yeon Kim, Dong-Jun Han, Jun Seo, and Jaekyun Moon. Warping the space: Weight space rotation for class incremental few-shot learning. In The Eleventh International Conference on Learning Representations, 2023. 1 +[29] Max Kochurov, Rasul Karimov, and Serge Kozlukov. Geoopt: Riemannian optimization in pytorch, 2020. 6 +[30] Anna Kukleva, Hilde Kuehne, and Bernt Schiele. Generalized and incremental few-shot learning by explicit learning and calibration without forgetting. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9020-9029, 2021. 1 +[31] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022. 5 +[32] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40(12):2935-2947, 2017. 1 +[33] Fangzhou Lin, Yun Yue, Songlin Hou, Xuechu Yu, Yajun Xu, Kazunori D Yamada, and Ziming Zhang. Hyperbolic chamfer distance for point cloud completion. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14595-14606, 2023. 1 +[34] Pei Lin, Zheng Cai, Jiangtao Zhang, and Lifang Ruan. A survey on point cloud segmentation: From classical methods to deep learning. ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM), 16(3): 1-27, 2020. 1 +[35] Yuyang Liu, Yang Cong, Gan Sun, Tao Zhang, Jiahua Dong, and Hongsen Liu. L3doc: Lifelong 3d object classification. IEEE Transactions on Image Processing, 30:7486-7498, 2021. 3 +[36] Zan Liu, Zilong Zhang, Yiyi Lin, Lin Zhu, Han Jin, and Stephen Lin. Pointaugment: An auto-augmentation framework for point cloud classification. arXiv preprint arXiv:2002.10876, 2020. 1 +[37] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6 +[38] Paolo Mandica, Luca Franco, Konstantinos Kallidromitis, Suzanne Petryk, and Fabio Galasso. Hyperbolic learning with multimodal large language models. arXiv preprint arXiv:2408.05097, 2024. 2 +[39] Pratik Mazumder, Pravendra Singh, and Piyush Rai. Few-shot lifelong learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2337-2345, 2021. 1 +[40] Leland McInnes, John Healy, and James Melville. Umap: Uniform manifold approximation and projection for dimension reduction. arXiv preprint arXiv:1802.03426, 2018. 2 + +[41] Pascal Mettes, Elise Van der Pol, and Cees Snoek. Hyperspherical prototype networks. Advances in neural information processing systems, 32, 2019. 5 +[42] Pascal Mettes, Mina Ghadimi Atigh, Martin Keller-Ressel, Jeffrey Gu, and Serena Yeung. Hyperbolic deep learning in computer vision: A survey. International Journal of Computer Vision, pages 1-25, 2024. 2 +[43] Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. Distributed representations of words and phrases and their compositionality. Advances in neural information processing systems, 26, 2013. 5 +[44] Gal Mishne, Zhengchao Wan, Yusu Wang, and Sheng Yang. The numerical stability of hyperbolic representation learning. In International Conference on Machine Learning, pages 24925-24949. PMLR, 2023. 2, 4, 5, 7 +[45] Antonio Montanaro, Diego Valsesia, and Enrico Magli. Rethinking the compositionality of point clouds through regularization in the hyperbolic space. Advances in Neural Information Processing Systems, 35:33741-33753, 2022. 1, 2, 3, 4 +[46] Antonio Montanaro, Diego Valsesia, and Enrico Magli. Towards hyperbolic regularizers for point cloud part segmentation. In ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5, 2023. 1 +[47] Gabriel Moreira, Manuel Marques, João Paulo Costeira, and Alexander Hauptmann. Hyperbolic vs euclidean embeddings in few-shot learning: Two sides of the same coin. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2082-2090, 2024. 7 +[48] Samrat Mukherjee, Tanuj Sur, Saurish Seksaria, Subhasis Chaudhuri, Gemma Roig, and Biplab Banerjee. Uidaple: Unsupervised incremental domain adaptation through adaptive prompt learning. In ICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5, 2025. 1 +[49] Maximilian Nickel and Douwe Kiela. Poincaré embeddings for learning hierarchical representations. In Advances in neural information processing systems, pages 6338-6347, 2017. 2, 3 +[50] Can Peng, Kun Zhao, Tianren Wang, Meng Li, and Brian C Lovell. Few-shot class-incremental learning from an open-set perspective. In European Conference on Computer Vision, pages 382-397. Springer, 2022. 3, 6, 8 +[51] Jeffrey Pennington, Richard Socher, and Christopher Manning. GloVe: Global vectors for word representation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1532-1543, Doha, Qatar, 2014. Association for Computational Linguistics. 5 +[52] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 1, 2 +[53] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on + +point sets in a metric space. Advances in neural information processing systems, 30, 2017. 1, 2 +[54] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5 +[55] Sachin Ravi and Hugo Larochelle. Optimization as a model for few-shot learning. arXiv preprint arXiv:1606.04474, 2017. 1 +[56] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 2001–2010, 2017. 1 +[57] Rik Sarkar. Low distortion delaunay embedding of trees in hyperbolic plane. In International symposium on graph drawing, pages 355-366. Springer, 2011. 2, 3 +[58] Yisheng Song, Ting Wang, Puyu Cai, Subrota K Mondal, and Jyoti Prakash Sahoo. A comprehensive survey of few-shot learning: Evolution, applications, challenges, and opportunities. ACM Computing Surveys, 55(13s):1-40, 2023. 1 +[59] Filip Szatkowski, Mateusz Pyla, Marcin Przewieżlikowski, Sebastian Cygert, Bartlomiej Twardowski, and Tomasz Trzeciński. Adapt your teacher: Improving knowledge distillation for exemplar-free continual learning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1977-1987, 2024. 6, 7, 8 +[60] Yuwen Tan and Xiang Xiang. Cross-domain few-shot incremental learning for point-cloud recognition. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2307-2316, 2024. 6 +[61] Xiaoyu Tao, Xiaopeng Hong, Xinyuan Chang, Songlin Dong, Xing Wei, and Yihong Gong. Few-shot class incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12183-12192, 2020. 3 +[62] Jesper E Van Engelen and Holger H Hoos. A survey on semi-supervised learning. Machine learning, 109(2):373-440, 2020. 1 +[63] Max van Spengler, Erwin Berkhout, and Pascal Mettes. Poincare resnet. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5419-5428, 2023. 4, 5 +[64] Qi-Wei Wang, Da-Wei Zhou, Yi-Kai Zhang, De-Chuan Zhan, and Han-Jia Ye. Few-shot class-incremental learning via training-free prototype calibration. Advances in Neural Information Processing Systems, 36, 2024. 6, 8 +[65] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (tog), 38(5):1-12, 2019. 1, 2, 6, 8 +[66] Yaqing Wang, Quanming Yao, James T Kwok, and Lionel M Ni. Generalizing from a few examples: A survey on few-shot learning. ACM computing surveys (csur), 53(3):1-34, 2020. 1 + +[67] Wenxuan Wu, Zhongang Qi, and Li Fuxin. Pointconv: Deep convolutional networks on 3d point clouds. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 9621-9630, 2019. 1, 2 +[68] Yating Xu, Conghui Hu, Na Zhao, and Gim Hee Lee. Generalized few-shot point cloud segmentation via geometric words. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21506-21515, 2023. 1 +[69] Xiangli Yang, Zixing Song, Irwin King, and Zenglin Xu. A survey on deep semi-supervised learning. IEEE Transactions on Knowledge and Data Engineering, 35(9):8934-8954, 2022. 1 +[70] Yuwei Yang, Munawar Hayat, Zhao Jin, Chao Ren, and Yinjie Lei. Geometry and uncertainty-aware 3d point cloud class-incremental semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21759-21768, 2023. 6, 7, 8 +[71] Zhilu Yang, Ying Zheng, and Jun Zhang. Learning to learn from labeled and unlabeled data with point-wise mutual information. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5704-5711, 2019. 1 +[72] Ze Yang, Ruibo Li, Evan Ling, Chi Zhang, Yiming Wang, Dezhao Huang, Keng Teck Ma, Minhoe Hur, and Guosheng Lin. Label-guided knowledge distillation for continual semantic segmentation on 2d images and 3d point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18601-18612, 2023. 1, 3, 6, 7, 8 +[73] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. arXiv preprint arXiv:2111.07783, 2021. 5 +[74] Yun Yue, Fangzhou Lin, Guanyi Mou, and Ziming Zhang. Understanding hyperbolic metric learning through hard negative sampling. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1891-1903, 2024. 2 +[75] Maciej Zamorski, Michal Stypulkowski, Konrad Karanowski, Tomasz Trzeciński, and Maciej Zieba. Continual learning on 3d point clouds with random compressed rehearsal. Computer Vision and Image Understanding, 228: 103621, 2023. 3 +[76] Chi Zhang, Nan Song, Guosheng Lin, Yun Zheng, Pan Pan, and Yinghui Xu. Few-shot incremental learning with continually evolved classifiers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12455-12464, 2021. 2, 3 +[77] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16259-16268, 2021. 2, 6, 7, 8 +[78] Linglan Zhao, Jing Lu, Yunlu Xu, Zhanzhan Cheng, Dashan Guo, Yi Niu, and Xiangzhong Fang. Few-shot class incremental learning via class-aware bilateral distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11838-11847, 2023. 1, 3, 6, 8 +[79] Na Zhao and Gim Hee Lee. Static-dynamic co-teaching for class-incremental 3d object detection. In Proceedings of + +the AAAI Conference on Artificial Intelligence, pages 3436-3445, 2022. 1, 3, 6, 7, 8 +[80] Na Zhao, Tat-Seng Chua, and Gim Hee Lee. Few-shot 3d point cloud semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8873-8882, 2021. 2 +[81] Da-Wei Zhou, Fu-Yun Wang, Han-Jia Ye, Liang Ma, Shiliang Pu, and De-Chuan Zhan. Forward compatible few-shot class-incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9046–9056, 2022. 1, 2, 3, 6, 8 +[82] Da-Wei Zhou, Zi-Wen Cai, Han-Jia Ye, De-Chuan Zhan, and Ziwei Liu. Revisiting class-incremental learning with pretrained models: Generalizability and adaptivity are all you need. arXiv preprint arXiv:2303.07338, 2023. 2 +[83] Da-Wei Zhou, Qi-Wei Wang, Zhi-Hong Qi, Han-Jia Ye, De-Chuan Zhan, and Ziwei Liu. Class incremental learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2 +[84] Yudong Zhu, Di Zhou, Jinghui Xiao, Xin Jiang, Xiao Chen, and Qun Liu. Hypertext: Endowing fasttext with hyperbolic geometry. arXiv preprint arXiv:2010.16143, 2020. 2 +[85] Yixiong Zou, Shanghang Zhang, Yuhua Li, and Ruixuan Li. Margin-based few-shot class-incremental learning with class-level overfitting mitigation. Advances in neural information processing systems, 35:27267-27279, 2022. 1, 3, 6, 8 \ No newline at end of file diff --git a/2025/Hyperbolic Uncertainty-Aware Few-Shot Incremental Point Cloud Segmentation/images.zip b/2025/Hyperbolic Uncertainty-Aware Few-Shot Incremental Point Cloud Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..78f64c2664dc9cce7d9ac021c5fe45b0cdd5cb3d --- /dev/null +++ b/2025/Hyperbolic Uncertainty-Aware Few-Shot Incremental Point Cloud Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8b84782ae8dd164edd233ba30c8e4b6facc08a1565052fe651eea5c11e849854 +size 564249 diff --git a/2025/Hyperbolic Uncertainty-Aware Few-Shot Incremental Point Cloud Segmentation/layout.json b/2025/Hyperbolic Uncertainty-Aware Few-Shot Incremental Point Cloud Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f561022a08139d16068358f375221060d5ca679a --- /dev/null +++ b/2025/Hyperbolic Uncertainty-Aware Few-Shot Incremental Point Cloud Segmentation/layout.json @@ -0,0 +1,11937 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 57, + 103, + 553, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 103, + 553, + 121 + ], + "spans": [ + { + "bbox": [ + 57, + 103, + 553, + 121 + ], + "type": "text", + "content": "Hyperbolic Uncertainty-Aware Few-Shot Incremental Point Cloud Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 142, + 541, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 142, + 541, + 171 + ], + "spans": [ + { + "bbox": [ + 67, + 142, + 541, + 171 + ], + "type": "text", + "content": "Tanuj Sur† Samrat Mukherjee* Kaizer Rahaman‡ Subbasis Chaudhuri* Muhammad Haris Khan§ Biplab Banerjee*" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 172, + 556, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 172, + 556, + 200 + ], + "spans": [ + { + "bbox": [ + 62, + 172, + 556, + 200 + ], + "type": "text", + "content": "†Chennai Mathematical Institute *Indian Institute of Technology Bombay ‡Indian Institute of Technology Kharagpur §Mohamed Bin Zayed University of Artificial Intelligence" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "spans": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 252, + 297, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 252, + 297, + 540 + ], + "spans": [ + { + "bbox": [ + 54, + 252, + 297, + 540 + ], + "type": "text", + "content": "3D point cloud segmentation is essential across a range of applications; however, conventional methods often struggle in evolving environments, particularly when tasked with identifying novel categories under limited supervision. Few-Shot Learning (FSL) and Class Incremental Learning (CIL) have been adapted previously to address these challenges in isolation, yet the combined paradigm of Few-Shot Class Incremental Learning (FSCIL) remains largely unexplored for point cloud segmentation. To address this gap, we introduce Hyperbolic Ideal Prototypes Optimization (HiPO), a novel framework that harnesses hyperbolic embeddings for FSCIL in 3D point clouds. HiPO employs the Poincaré Hyperbolic Sphere as its embedding space, integrating Ideal Prototypes enriched by CLIP-derived class semantics, to capture the hierarchical structure of 3D data. By enforcing orthogonality among prototypes and maximizing representational margins, HiPO constructs a resilient embedding space that mitigates forgetting and enables the seamless integration of new classes, thereby effectively countering overfitting. Extensive evaluations on S3DIS, ScanNetv2, and cross-dataset scenarios demonstrate HiPO's strong performance, significantly surpassing existing approaches in both in-domain and cross-dataset FSCIL tasks for 3D point cloud segmentation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 558, + 135, + 570 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 135, + 570 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 135, + 570 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 578, + 296, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 578, + 296, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 578, + 296, + 685 + ], + "type": "text", + "content": "Recent advancements in deep learning have spurred progress in data-efficient learning, with semi-supervised learning (SSL) [62, 69, 71] and few-shot learning (FSL) [55, 58, 66] emerging as key approaches. SSL utilizes both labeled and unlabeled data to enhance learning, while FSL enables generalization from only a few labeled samples per class. However, both approaches face challenges in real-world, dynamic environments where models must continuously adapt to new classes while retaining prior knowledge" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 322, + 224, + 553, + 353 + ], + "blocks": [ + { + "bbox": [ + 322, + 224, + 553, + 353 + ], + "lines": [ + { + "bbox": [ + 322, + 224, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 322, + 224, + 553, + 353 + ], + "type": "image", + "image_path": "70e6e92c326f538e13202a67060afa89ed8e4d8a880cca0433775eef9025159b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 361, + 555, + 451 + ], + "lines": [ + { + "bbox": [ + 313, + 361, + 555, + 451 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 555, + 451 + ], + "type": "text", + "content": "Figure 1. Our approach tackles the novel challenge of few-shot class incremental 3D point cloud segmentation. In each session " + }, + { + "bbox": [ + 313, + 361, + 555, + 451 + ], + "type": "inline_equation", + "content": "S^{(t)}, t \\geq 0" + }, + { + "bbox": [ + 313, + 361, + 555, + 451 + ], + "type": "text", + "content": ", our model takes an RGB point cloud of an indoor scene " + }, + { + "bbox": [ + 313, + 361, + 555, + 451 + ], + "type": "inline_equation", + "content": "D^{(t)}" + }, + { + "bbox": [ + 313, + 361, + 555, + 451 + ], + "type": "text", + "content": " as input and performs semantic segmentation. During the base training session " + }, + { + "bbox": [ + 313, + 361, + 555, + 451 + ], + "type": "inline_equation", + "content": "S^{(0)}" + }, + { + "bbox": [ + 313, + 361, + 555, + 451 + ], + "type": "text", + "content": ", sufficient examples are available for each class. In later sessions, only a limited number of labeled point cloud examples are provided for new classes, introducing data imbalance and potential overfitting and forgetting risks." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 461, + 555, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 555, + 533 + ], + "type": "text", + "content": "edge—an essential requirement for continual learning (CL) [1, 4, 8, 11, 32, 48, 56]. Few-Shot Class Incremental Learning (FSCIL) [1, 13, 24, 28, 30, 39] combines the data efficiency of FSL with the adaptability of CL, yet remains underdeveloped for complex tasks such as segmentation in unstructured 3D data (see Fig. 1)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 534, + 556, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 700 + ], + "type": "text", + "content": "Point cloud segmentation is pivotal in applications like autonomous driving and robotics, where it enables precise object recognition and scene understanding [36, 52, 68]. This task assigns semantic labels to each point within a 3D structure, facilitating critical decision-making processes. Current supervised segmentation methods [26, 53, 65, 67] are highly reliant on extensive labeled datasets, which are costly and time-intensive to produce [34]. These methods also assume a fixed set of labels, limiting their adaptability in scenarios where new classes emerge unexpectedly. While FSCIL offers a potential solution, existing methods [1, 24, 72, 78, 79, 81, 85] struggle to represent the complex, hierarchical structures typical of point cloud data due to the limitations of Euclidean embeddings [33, 45, 46]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 325, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 325, + 701, + 553, + 714 + ], + "type": "text", + "content": "Capturing hierarchical relationships, such as recogniz" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "text", + "content": "*Samrat Mukherjee acknowledges the support of Prime Minister Research Fellowship (PMRF)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11810" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "content": "ing subcomponents within an object, demands a framework that can accommodate intricate interrelations. This challenge is intensified in Class Incremental Learning (CIL) [82, 83], where the model must incorporate new classes while preserving previously learned representations—a balance between adaptability and memory known as the stability-plasticity trade-off. In FSCIL, especially under data scarcity, this trade-off is often compromised, leading to overfitting and catastrophic forgetting. Euclidean spaces exacerbate these challenges, as they lack the capacity to model evolving data scenarios [76, 81]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 206, + 295, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 206, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 206, + 295, + 422 + ], + "type": "text", + "content": "Hyperbolic geometry [25, 38, 42, 57, 74] provides a more suitable framework for hierarchical data [45, 84]. Its negative curvature and exponential volume growth allow it to mirror tree-like data structures, addressing the stability-plasticity trade-off [16] and supporting segmentation in unstructured contexts [27]. Unlike Euclidean spaces, hyperbolic space offers exponentially more volume [44] for accommodating incremental classes, thus reducing overlap and interference (see Fig. 2, Tab. 4). However, efficient optimization of embeddings in the hyperbolic space presents computational challenges when points are near the boundary of the Poincaré ball resulting in vanishing gradient problems [44], and without careful handling, newly added classes can still encroach upon the existing embeddings, aggravating plasticity (see Tab. 3). This highlights the need for an FSCIL framework that uses hyperbolic space embedding hierarchical data in low distortion while ensuring computational efficiency and managing class overlap." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 424, + 295, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 424, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 424, + 295, + 568 + ], + "type": "text", + "content": "Our Proposal: To address these challenges, we introduce Hyperbolic Ideal Prototypes Optimization (HiPo), a novel FSCIL framework tailored for point cloud segmentation. By embedding data within the Poincaré Hyperbolic Sphere [49], HiPo leverages the unique properties of hyperbolic space to capture hierarchical structures and maintain consistent feature-classifier alignment across incremental learning sessions. Through a modified Busemann loss [7], HiPo arranges ideal prototypes orthogonally within the hyperbolic space, maximizing class separation and preserving space for new classes, which minimizes inter-class interference and optimizes the feature space for CIL." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": "This design takes advantage of hyperbolic space's exponential scaling, which aligns naturally with complex hierarchies found in point cloud data [3, 27, 45, 57]. Our modified Busemann loss incorporates two components: (1) a prototype alignment term to guide class embeddings towards their respective Hyperbolic Ideal Prototypes, enhancing class separability, and intra-class compactness, and (2) an uncertainty-aware term that penalizes overfitting on novel classes and preventing vanishing gradient problems. Together, these components enable HiPo to effectively address data scarcity, mitigate catastrophic forgetting, and support robust FSCIL under episodic shifts, establishing it" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 330, + 74, + 535, + 188 + ], + "blocks": [ + { + "bbox": [ + 330, + 74, + 535, + 188 + ], + "lines": [ + { + "bbox": [ + 330, + 74, + 535, + 188 + ], + "spans": [ + { + "bbox": [ + 330, + 74, + 535, + 188 + ], + "type": "image", + "image_path": "6041dc375c4e0f7391ca49c8dcf2da0eaafdb6ef4b254da7fb450ac065593299.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 199, + 555, + 288 + ], + "lines": [ + { + "bbox": [ + 313, + 199, + 555, + 288 + ], + "spans": [ + { + "bbox": [ + 313, + 199, + 555, + 288 + ], + "type": "text", + "content": "Figure 2. Motivation for hyperbolic embeddings: (a) Hyperbolic UMAP [40] of the HiPo embedding space, projected to 2D for the S3DIS dataset [2]. The hyperbolic space's negative curvature aligns with hierarchical data, enabling a more compact representation and clearer class separation than Euclidean space (Dunn index [18] " + }, + { + "bbox": [ + 313, + 199, + 555, + 288 + ], + "type": "inline_equation", + "content": "(\\uparrow)" + }, + { + "bbox": [ + 313, + 199, + 555, + 288 + ], + "type": "text", + "content": " for HiPo is 0.69 while for the Euclidean space, it is 0.35), showcasing better class compactness in our case. (b) UMAP of the classifier space in the Euclidean hypersphere" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 304, + 555, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 555, + 327 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 555, + 327 + ], + "type": "text", + "content": "as a powerful solution for incremental learning in 3D point cloud segmentation. Overall, our contributions are:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 329, + 553, + 510 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 313, + 329, + 553, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 329, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 313, + 329, + 553, + 364 + ], + "type": "text", + "content": "- We define and address the previously unexplored problem of FSCIL for 3D point cloud segmentation, encompassing both in-domain and cross-dataset scenarios." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 365, + 553, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 553, + 414 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 553, + 414 + ], + "type": "text", + "content": "- We introduce hyperbolic space as a robust embedding framework for FSCIL in 3D segmentation, effectively capturing hierarchical structures and preserving stable feature-class relationships across incremental learning sessions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 415, + 553, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 553, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 553, + 462 + ], + "type": "text", + "content": "- We propose a modified Busemann loss with prototype alignment and uncertainty-aware components to ensure optimal feature-classifier alignment with minimized uncertainty in predictions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 464, + 553, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 553, + 510 + ], + "type": "text", + "content": "- Through extensive experiments on benchmark datasets S3DIS [2] and ScanNetv2 [17], HiPo demonstrates superior performance and generalization, setting a new standard for FSCIL in this domain." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 525, + 406, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 525, + 406, + 537 + ], + "spans": [ + { + "bbox": [ + 314, + 525, + 406, + 537 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "content": "3D Point-Cloud Semantic Segmentation: Fully supervised 3D semantic segmentation has been a well-explored area, with methods typically requiring dense point-wise labels [26, 52, 53, 65, 67, 77]. A significant advancement was PointNet [52], which directly processes raw point clouds, removing the need for voxelization but facing challenges in capturing local structures. This limitation spurred improvements like PointNet++ [53], which introduced a hierarchical approach to enhance local feature learning, and DGCNN [65], which incorporated the EdgeConv module to integrate neighborhood information for better geometric representation. More recently, models such as PointTransformer [80] have integrated self-attention mechanisms, inspired by NLP, to elevate feature expressiveness and contextual awareness." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11811" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "text", + "content": "Despite these advancements, the dependency on large labeled datasets and the assumption of a fixed class set restrict these methods in practical scenarios where new classes must be incrementally learned with minimal supervision." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 121, + 296, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 121, + 296, + 372 + ], + "spans": [ + { + "bbox": [ + 56, + 121, + 296, + 372 + ], + "type": "text", + "content": "(Few-Shot) Class Incremental Learning: Incremental learning methods for 3D data have predominantly focused on classification tasks rather than segmentation [14, 15, 19, 35, 75]. Approaches such as I3DOL [19] and L3DOC [35] mitigate catastrophic forgetting in 3D object classification through memory and attention mechanisms, yet they struggle as labeled supervision diminishes over time. Few-Shot Class Incremental Learning (FSCIL) [61] has emerged as a more flexible solution, enabling models to learn new classes incrementally with minimal labeled data. Representative methods like CEC [76] use graph-based task linking, while FACT [81] and ALICE [50] create adaptable feature spaces through clustering and cosine similarity. However, these techniques rely on Euclidean embeddings, which have limited capability to represent hierarchical structures intrinsic to 3D data, and frequent classifier updates introduce computational overhead [76]. Open-world semantic segmentation (OWSS) [5, 9] is a related area that focuses on dynamically detecting unknown classes at inference, though our focus on incrementally adding known classes during training sets our approach apart from OWSS." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 373, + 295, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 295, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 295, + 540 + ], + "type": "text", + "content": "Hyperbolic Representations for Hierarchical Data: Hyperbolic geometry has shown promise for hierarchical data representation, with foundational work by Sarkar et al. [57] and Nickel et al. [49] introducing hyperbolic neural networks and Riemannian optimization methods. Hyperbolic embeddings have demonstrated substantial benefits for capturing semantic hierarchies across applications, including image embeddings [27] and FSCIL for image classification [16]. In 3D data, recent research has shown that hyperbolic representations can model complex part-whole relationships effectively. Montanaro et al. [45] applied hyperbolic regularizers to encode hierarchical structures in 3D objects, while Hu et al. [25] leveraged hyperbolic geometry to represent latent structures in point clouds." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 541, + 296, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 541, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 541, + 296, + 590 + ], + "type": "text", + "content": "Our approach leverages hyperbolic embeddings to tackle FSCIL challenges in segmentation, improving adaptability, preserving hierarchies, and ensuring computational efficiency in the point cloud domain." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 602, + 251, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 602, + 251, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 602, + 251, + 616 + ], + "type": "text", + "content": "3. Problem Definition & Methodology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 624, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 295, + 715 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 55, + 624, + 295, + 715 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 624, + 295, + 715 + ], + "type": "text", + "content": " represent a sequence of training datasets " + }, + { + "bbox": [ + 55, + 624, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\{D^{(0)}, D^{(1)}, \\ldots, D^{(T)}\\}" + }, + { + "bbox": [ + 55, + 624, + 295, + 715 + ], + "type": "text", + "content": ", where each " + }, + { + "bbox": [ + 55, + 624, + 295, + 715 + ], + "type": "inline_equation", + "content": "D^{(t)} = \\{(P_k^{(t)}, M_k^{(t)})_{k=1}^{|D^{(t)}|}\\}" + }, + { + "bbox": [ + 55, + 624, + 295, + 715 + ], + "type": "text", + "content": " corresponds to a training session " + }, + { + "bbox": [ + 55, + 624, + 295, + 715 + ], + "type": "inline_equation", + "content": "S^{(t)}" + }, + { + "bbox": [ + 55, + 624, + 295, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 624, + 295, + 715 + ], + "type": "inline_equation", + "content": "t \\in \\{0, 1, \\ldots, T\\}" + }, + { + "bbox": [ + 55, + 624, + 295, + 715 + ], + "type": "text", + "content": ". The initial session dataset " + }, + { + "bbox": [ + 55, + 624, + 295, + 715 + ], + "type": "inline_equation", + "content": "D^{(0)}" + }, + { + "bbox": [ + 55, + 624, + 295, + 715 + ], + "type": "text", + "content": " serves as the base, while " + }, + { + "bbox": [ + 55, + 624, + 295, + 715 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 624, + 295, + 715 + ], + "type": "text", + "content": " denotes the total number of incremental sessions. In each session " + }, + { + "bbox": [ + 55, + 624, + 295, + 715 + ], + "type": "inline_equation", + "content": "S^{(t)}" + }, + { + "bbox": [ + 55, + 624, + 295, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 624, + 295, + 715 + ], + "type": "inline_equation", + "content": "P_k^{(t)}" + }, + { + "bbox": [ + 55, + 624, + 295, + 715 + ], + "type": "text", + "content": " represents a point cloud sample, and " + }, + { + "bbox": [ + 55, + 624, + 295, + 715 + ], + "type": "inline_equation", + "content": "M_k^{(t)}" + }, + { + "bbox": [ + 55, + 624, + 295, + 715 + ], + "type": "text", + "content": " provides its labels within" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 70, + 553, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 70, + 553, + 97 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 553, + 97 + ], + "type": "text", + "content": "the label space " + }, + { + "bbox": [ + 313, + 70, + 553, + 97 + ], + "type": "inline_equation", + "content": "C^{(t)} = \\{c_1^{(t)},c_2^{(t)},\\ldots ,c_{|C^{(t)}|}^{(t)}\\}" + }, + { + "bbox": [ + 313, + 70, + 553, + 97 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 70, + 553, + 97 + ], + "type": "inline_equation", + "content": "c_{i}^{(t)}" + }, + { + "bbox": [ + 313, + 70, + 553, + 97 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 313, + 70, + 553, + 97 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 70, + 553, + 97 + ], + "type": "text", + "content": "-th class in the incremental session " + }, + { + "bbox": [ + 313, + 70, + 553, + 97 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 70, + 553, + 97 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 98, + 553, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 98, + 553, + 231 + ], + "spans": [ + { + "bbox": [ + 313, + 98, + 553, + 231 + ], + "type": "text", + "content": "The base session " + }, + { + "bbox": [ + 313, + 98, + 553, + 231 + ], + "type": "inline_equation", + "content": "S^{(0)}" + }, + { + "bbox": [ + 313, + 98, + 553, + 231 + ], + "type": "text", + "content": " typically contains an abundant number of samples per class in the label space " + }, + { + "bbox": [ + 313, + 98, + 553, + 231 + ], + "type": "inline_equation", + "content": "C^{(0)}" + }, + { + "bbox": [ + 313, + 98, + 553, + 231 + ], + "type": "text", + "content": ". However, in each subsequent incremental session " + }, + { + "bbox": [ + 313, + 98, + 553, + 231 + ], + "type": "inline_equation", + "content": "S^{(t)}, t > 0" + }, + { + "bbox": [ + 313, + 98, + 553, + 231 + ], + "type": "text", + "content": ", only a limited number of labeled point clouds " + }, + { + "bbox": [ + 313, + 98, + 553, + 231 + ], + "type": "inline_equation", + "content": "P_{k}^{(t)} \\in \\mathbb{R}^{m \\times d_{0}}" + }, + { + "bbox": [ + 313, + 98, + 553, + 231 + ], + "type": "text", + "content": " are provided, where " + }, + { + "bbox": [ + 313, + 98, + 553, + 231 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 98, + 553, + 231 + ], + "type": "text", + "content": " is the number of points per sample and " + }, + { + "bbox": [ + 313, + 98, + 553, + 231 + ], + "type": "inline_equation", + "content": "d_{0}" + }, + { + "bbox": [ + 313, + 98, + 553, + 231 + ], + "type": "text", + "content": " is the feature dimension. Following previous works [1, 24, 72, 78, 79, 81, 85], we assume " + }, + { + "bbox": [ + 313, + 98, + 553, + 231 + ], + "type": "inline_equation", + "content": "|D^{(t)}| = N \\times K" + }, + { + "bbox": [ + 313, + 98, + 553, + 231 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 98, + 553, + 231 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 98, + 553, + 231 + ], + "type": "text", + "content": " is the number of classes, and " + }, + { + "bbox": [ + 313, + 98, + 553, + 231 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 98, + 553, + 231 + ], + "type": "text", + "content": " is the number of samples per class, following an " + }, + { + "bbox": [ + 313, + 98, + 553, + 231 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 98, + 553, + 231 + ], + "type": "text", + "content": "-way " + }, + { + "bbox": [ + 313, + 98, + 553, + 231 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 98, + 553, + 231 + ], + "type": "text", + "content": "-shot setting " + }, + { + "bbox": [ + 313, + 98, + 553, + 231 + ], + "type": "inline_equation", + "content": "\\forall t > 0" + }, + { + "bbox": [ + 313, + 98, + 553, + 231 + ], + "type": "text", + "content": ". We assume " + }, + { + "bbox": [ + 313, + 98, + 553, + 231 + ], + "type": "inline_equation", + "content": "C^{(t)}" + }, + { + "bbox": [ + 313, + 98, + 553, + 231 + ], + "type": "text", + "content": " to be disjoint from previous sessions, i.e., " + }, + { + "bbox": [ + 313, + 98, + 553, + 231 + ], + "type": "inline_equation", + "content": "C^{(t)} \\cap C^{(t')} = \\emptyset" + }, + { + "bbox": [ + 313, + 98, + 553, + 231 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 98, + 553, + 231 + ], + "type": "inline_equation", + "content": "\\forall t' < t" + }, + { + "bbox": [ + 313, + 98, + 553, + 231 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 232, + 553, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 553, + 316 + ], + "type": "text", + "content": "During each incremental session, only the current dataset " + }, + { + "bbox": [ + 313, + 232, + 553, + 316 + ], + "type": "inline_equation", + "content": "D^{(t)}" + }, + { + "bbox": [ + 313, + 232, + 553, + 316 + ], + "type": "text", + "content": " is accessible, with data from previous sessions unavailable for training. For evaluation in session " + }, + { + "bbox": [ + 313, + 232, + 553, + 316 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 232, + 553, + 316 + ], + "type": "text", + "content": ", the model is tested across all classes encountered up to that session, covering the label space " + }, + { + "bbox": [ + 313, + 232, + 553, + 316 + ], + "type": "inline_equation", + "content": "\\bigcup_{i=0}^{t} C^{(i)}" + }, + { + "bbox": [ + 313, + 232, + 553, + 316 + ], + "type": "text", + "content": ". A solution to the FSCIL problem must jointly address uncharted catastrophic forgetting, data imbalance, and class overfitting judiciously." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 321, + 400, + 332 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 321, + 400, + 332 + ], + "spans": [ + { + "bbox": [ + 314, + 321, + 400, + 332 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 338, + 553, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 553, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 553, + 410 + ], + "type": "text", + "content": "This section mentions the essential backgrounds needed for the proposed framework Hyperbolic Ideal Prototypes Optimization (HiPo). We extensively use the concepts of the Poincaré ball model and its associated Möbius addition, exponential map, and the geodesic distance. We detail their definition and discuss their formulae in the Sup. Mat." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 411, + 554, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 554, + 507 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 554, + 507 + ], + "type": "text", + "content": "In the Poincaré ball model of hyperbolic geometry, all the points at the boundary of the Poincaré ball represent points at infinity. We initialize a group of classifier prototypes at the boundary of the Poincaré ball termed Ideal Prototypes, following [23], for each of the " + }, + { + "bbox": [ + 313, + 411, + 554, + 507 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 411, + 554, + 507 + ], + "type": "text", + "content": " individual classes. Assuming " + }, + { + "bbox": [ + 313, + 411, + 554, + 507 + ], + "type": "inline_equation", + "content": "u_{1}, u_{2}, \\dots, u_{n}" + }, + { + "bbox": [ + 313, + 411, + 554, + 507 + ], + "type": "text", + "content": " to be the Ideal Prototypes for " + }, + { + "bbox": [ + 313, + 411, + 554, + 507 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 411, + 554, + 507 + ], + "type": "text", + "content": " classes, they form the boundary of the " + }, + { + "bbox": [ + 313, + 411, + 554, + 507 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 411, + 554, + 507 + ], + "type": "text", + "content": "-dimensional Poincaré ball of curvature " + }, + { + "bbox": [ + 313, + 411, + 554, + 507 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 411, + 554, + 507 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 411, + 554, + 507 + ], + "type": "inline_equation", + "content": "\\mathbb{B}_c^n" + }, + { + "bbox": [ + 313, + 411, + 554, + 507 + ], + "type": "text", + "content": ", following [23]:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 346, + 517, + 553, + 531 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 517, + 553, + 531 + ], + "spans": [ + { + "bbox": [ + 346, + 517, + 553, + 531 + ], + "type": "interline_equation", + "content": "\\mathcal {I} _ {n} = \\left\\{u \\in \\mathbb {R} ^ {n}: u _ {1} ^ {2} + u _ {2} ^ {2} + \\dots + u _ {n} ^ {2} = 1 \\right\\} \\tag {1}", + "image_path": "16a4a54edaea86884a8a7d57a32f44d9932cc49f7398274a29e05025132784b7.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 534, + 554, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 554, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 554, + 582 + ], + "type": "text", + "content": "Since Ideal Prototypes are points at an infinite geodesic distance from all other points in " + }, + { + "bbox": [ + 313, + 534, + 554, + 582 + ], + "type": "inline_equation", + "content": "\\mathbb{B}_c^n" + }, + { + "bbox": [ + 313, + 534, + 554, + 582 + ], + "type": "text", + "content": ", we use the Busemann function [7] as a distance measure since computing direct infinite geodesic distance is non-trivial [23]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 582, + 553, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 553, + 642 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 553, + 642 + ], + "type": "text", + "content": "The pre-assigned Ideal Prototypes form an optimal geometric structure by fixing the target throughout the incremental learning. However, as already pointed out, optimizing Ideal Prototypes in the Poincaré model to address FS-CIL in 3D point cloud segmentation remains unexplored." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 648, + 415, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 648, + 415, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 648, + 415, + 661 + ], + "type": "text", + "content": "3.2. Discussing HiPo" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "content": "Our model introduces a fixed alignment strategy for feature-classifier relationships in FSCIL, tailored for the hierarchical nature of data such as 3D point clouds. Traditional Euclidean embeddings suffer from compaction [50, 76, 81]" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11812" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 68, + 206, + 223 + ], + "blocks": [ + { + "bbox": [ + 58, + 68, + 206, + 223 + ], + "lines": [ + { + "bbox": [ + 58, + 68, + 206, + 223 + ], + "spans": [ + { + "bbox": [ + 58, + 68, + 206, + 223 + ], + "type": "image", + "image_path": "63cf5ba78a7e0238dbb7cc650334e3c091fd8659e767f93bccc19eb1944a7509.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 234, + 555, + 291 + ], + "lines": [ + { + "bbox": [ + 54, + 234, + 555, + 291 + ], + "spans": [ + { + "bbox": [ + 54, + 234, + 555, + 291 + ], + "type": "text", + "content": "Figure 3. HiPo Overview: We show the changes in the Poincaré ball during the initial three sessions in the 3-2T FSCIL setting (three base classes, two incremental classes per session) for the S3DIS dataset [2], designed to easily visualize the working of HiPo. Hyperbolic Ideal Prototypes are optimally positioned to increase the margin of semantically similar classes " + }, + { + "bbox": [ + 54, + 234, + 555, + 291 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{HIP}})" + }, + { + "bbox": [ + 54, + 234, + 555, + 291 + ], + "type": "text", + "content": ". As an example, the chair class gets positioned farther from sofa than from table or clutter. We align representations of the newly introduced classes with their corresponding ideal prototypes " + }, + { + "bbox": [ + 54, + 234, + 555, + 291 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{Alignment}})" + }, + { + "bbox": [ + 54, + 234, + 555, + 291 + ], + "type": "text", + "content": " regularized by a modified Busemann function " + }, + { + "bbox": [ + 54, + 234, + 555, + 291 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{Uncertainty}})" + }, + { + "bbox": [ + 54, + 234, + 555, + 291 + ], + "type": "text", + "content": ". Training is conducted using " + }, + { + "bbox": [ + 54, + 234, + 555, + 291 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{HiPo}}" + }, + { + "bbox": [ + 54, + 234, + 555, + 291 + ], + "type": "text", + "content": " across sessions." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 208, + 68, + 356, + 224 + ], + "blocks": [ + { + "bbox": [ + 208, + 68, + 356, + 224 + ], + "lines": [ + { + "bbox": [ + 208, + 68, + 356, + 224 + ], + "spans": [ + { + "bbox": [ + 208, + 68, + 356, + 224 + ], + "type": "image", + "image_path": "0ca66a39feff0e83af01997e2c0c0cc6be61082c8c5238757cd6bcf5c77efcd5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 357, + 68, + 550, + 224 + ], + "blocks": [ + { + "bbox": [ + 357, + 68, + 550, + 224 + ], + "lines": [ + { + "bbox": [ + 357, + 68, + 550, + 224 + ], + "spans": [ + { + "bbox": [ + 357, + 68, + 550, + 224 + ], + "type": "image", + "image_path": "8792de0dd8c23bfe0c414fffff5e860668178ff2bc3fd3c05c0d29f4f04bdfc6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 300, + 295, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 300, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 54, + 300, + 295, + 491 + ], + "type": "text", + "content": "and distortion as new classes are added, leading to overlap and forgetting. Hyperbolic space, with its negative curvature and exponential volume, provides a natural fit for preserving hierarchical structures while minimizing inter-class interference. To address catastrophic forgetting, our model anchors feature-classifier relationships within hyperbolic space, ensuring stable alignment between features and classifier prototypes. This fixed alignment prevents the drift of previously learned embeddings as new classes are introduced, preserving the original classifier prototype positions and maintaining consistency across incremental sessions. By leveraging hyperbolic geometry, our approach enables efficient class separation, robust memory retention, and adaptability, making it well-suited for FSCIL in high-dimensional, unstructured data settings. Our method has two main stages:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 495, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 495, + 295, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 295, + 579 + ], + "type": "text", + "content": "- Defining Ideal Prototypes: We initialize ideal prototypes for each class on the boundary of " + }, + { + "bbox": [ + 55, + 495, + 295, + 579 + ], + "type": "inline_equation", + "content": "\\mathbb{B}_c^n" + }, + { + "bbox": [ + 55, + 495, + 295, + 579 + ], + "type": "text", + "content": ", positioning them mutually orthogonal and at maximal distances to enhance discriminative power, especially among classes with similar characteristics (see Sec. 3.2.1). This fixed configuration allows a dedicated space for each class, preventing the embedding drift observed in Euclidean spaces." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": "- Efficient Projections onto Poincaré ball " + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbb{B}_c^n" + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": ": In Hyperbolic embedding space, mean class features are obtained by computing Fréchet [20] means over the feature embedding for the training examples belonging to the classes. Instead of directly computing class features on the Poincaré ball—an intensive task[63]—we obtain the feature vectors in Euclidean space for every class in each session and then project them into hyperbolic space using exponential map " + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\exp_0(.)" + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": " [45] (see Sec. 3.2.2). The hyperbolic classifier head " + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(.)" + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": " consisting of a Möbius layer " + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": " (also known as hyperbolic feed forward layer [22]), takes the output of the" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 300, + 553, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 553, + 336 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 553, + 336 + ], + "type": "text", + "content": "exponential layer as input and projects mean class features to " + }, + { + "bbox": [ + 313, + 300, + 553, + 336 + ], + "type": "inline_equation", + "content": "\\mathbb{B}_c^n" + }, + { + "bbox": [ + 313, + 300, + 553, + 336 + ], + "type": "text", + "content": ". The Poincaré ball addresses few-shot learning using the sample mean prototype as the class representation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 337, + 553, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 337, + 553, + 372 + ], + "spans": [ + { + "bbox": [ + 313, + 337, + 553, + 372 + ], + "type": "text", + "content": "Our framework consists of a point-level classification branch, combining a feature extractor " + }, + { + "bbox": [ + 313, + 337, + 553, + 372 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(.)" + }, + { + "bbox": [ + 313, + 337, + 553, + 372 + ], + "type": "text", + "content": " and hyperbolic classifier head " + }, + { + "bbox": [ + 313, + 337, + 553, + 372 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(.)" + }, + { + "bbox": [ + 313, + 337, + 553, + 372 + ], + "type": "text", + "content": ", and is trained with two key strategies:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 372, + 554, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 372, + 554, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 554, + 552 + ], + "type": "text", + "content": "Feature-Classifier Alignment: For each session " + }, + { + "bbox": [ + 313, + 372, + 554, + 552 + ], + "type": "inline_equation", + "content": "t \\geq 0" + }, + { + "bbox": [ + 313, + 372, + 554, + 552 + ], + "type": "text", + "content": ", we align the hyperbolic sample mean features (see Eqn. 4) with the Hyperbolic Ideal Prototypes of their respective classes using alignment loss, " + }, + { + "bbox": [ + 313, + 372, + 554, + 552 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Alignment}}" + }, + { + "bbox": [ + 313, + 372, + 554, + 552 + ], + "type": "text", + "content": ". We define " + }, + { + "bbox": [ + 313, + 372, + 554, + 552 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Alignment}}" + }, + { + "bbox": [ + 313, + 372, + 554, + 552 + ], + "type": "text", + "content": " using Busemann function " + }, + { + "bbox": [ + 313, + 372, + 554, + 552 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 313, + 372, + 554, + 552 + ], + "type": "text", + "content": " and introduce a regularization term, quantifying uncertainty of the class features, to prevent feature collapse around the ideal prototypes, effectively reducing interference with previously seen classes (see Sec. 3.2.3) and vanishing gradient problems [44]. Point-Level Classification: The hyperbolic classifier " + }, + { + "bbox": [ + 313, + 372, + 554, + 552 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(.)" + }, + { + "bbox": [ + 313, + 372, + 554, + 552 + ], + "type": "text", + "content": ", consisting of a Möbius layer " + }, + { + "bbox": [ + 313, + 372, + 554, + 552 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 372, + 554, + 552 + ], + "type": "text", + "content": " followed by softmax [45], processes the output of " + }, + { + "bbox": [ + 313, + 372, + 554, + 552 + ], + "type": "inline_equation", + "content": "\\exp_0(\\mathcal{E}(.))" + }, + { + "bbox": [ + 313, + 372, + 554, + 552 + ], + "type": "text", + "content": ", with a cross-entropy loss for 3D point cloud segmentation. The joint objective combines this alignment loss " + }, + { + "bbox": [ + 313, + 372, + 554, + 552 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Alignment}}" + }, + { + "bbox": [ + 313, + 372, + 554, + 552 + ], + "type": "text", + "content": " and cross-entropy loss " + }, + { + "bbox": [ + 313, + 372, + 554, + 552 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CE}}" + }, + { + "bbox": [ + 313, + 372, + 554, + 552 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 381, + 558, + 553, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 558, + 553, + 571 + ], + "spans": [ + { + "bbox": [ + 381, + 558, + 553, + 571 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {H I P O}} = \\mathcal {L} _ {\\text {A l i g n m e n t}} + \\mathcal {L} _ {\\mathrm {C E}} \\tag {2}", + "image_path": "cf7569a9d15a6c419e0d0e5f5d387118f1a098a446830b1731ac23d120725ac3.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 577, + 554, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 577, + 554, + 672 + ], + "spans": [ + { + "bbox": [ + 313, + 577, + 554, + 672 + ], + "type": "text", + "content": "This approach ensures robust feature-classifier alignment across incremental learning sessions, addressing catastrophic forgetting in 3D point cloud segmentation by preserving both old and new class representations in a rehearsal-free manner. The process is illustrated in Fig. 3, and Algorithm 1 and a table summarizing the important variables used, mentioned in the Sup. Mat., supplement these discussions." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 674, + 511, + 687 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 674, + 511, + 687 + ], + "spans": [ + { + "bbox": [ + 313, + 674, + 511, + 687 + ], + "type": "text", + "content": "3.2.1. Hyperbolic ideal prototypes positioning" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "Building upon the goal of maintaining stable feature-classifier alignment, our framework pre-assigns classifier" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "11813" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 337 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 337 + ], + "type": "text", + "content": "prototypes on the boundary of " + }, + { + "bbox": [ + 55, + 72, + 296, + 337 + ], + "type": "inline_equation", + "content": "\\mathbb{B}_c^n" + }, + { + "bbox": [ + 55, + 72, + 296, + 337 + ], + "type": "text", + "content": " to define a structured classifier space prior to backbone training. To achieve maximal margin separation between classes, we introduce inductive biases by assigning each class a Hyperbolic Ideal Prototype, distributing these prototypes as uniformly as possible across the Poincaré ball. Inspired by Mettes et al. [41], we incorporate semantic class knowledge to enforce distinct separations, particularly for closely related classes, such as \"chair\" and \"sofa\". To further enhance semantic cohesion, we utilize the CLIP framework [54] due to its robust semantic representations [31, 73] in comparison to embeddings like word2vec [43], Glove [51] etc. By querying CLIP's frozen text encoder with the prompt \"A photo of a [class]\", we obtain text embeddings " + }, + { + "bbox": [ + 55, + 72, + 296, + 337 + ], + "type": "inline_equation", + "content": "S = \\{s_1,\\dots ,s_n\\}" + }, + { + "bbox": [ + 55, + 72, + 296, + 337 + ], + "type": "text", + "content": " for each of the " + }, + { + "bbox": [ + 55, + 72, + 296, + 337 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 72, + 296, + 337 + ], + "type": "text", + "content": " predefined classes. Leveraging the homeomorphism between hyperbolic space " + }, + { + "bbox": [ + 55, + 72, + 296, + 337 + ], + "type": "inline_equation", + "content": "\\mathbb{B}_c^n" + }, + { + "bbox": [ + 55, + 72, + 296, + 337 + ], + "type": "text", + "content": " and the hypersphere " + }, + { + "bbox": [ + 55, + 72, + 296, + 337 + ], + "type": "inline_equation", + "content": "\\mathbb{S}^n" + }, + { + "bbox": [ + 55, + 72, + 296, + 337 + ], + "type": "text", + "content": " [23], we maximize separation between nearest neighboring embeddings in " + }, + { + "bbox": [ + 55, + 72, + 296, + 337 + ], + "type": "inline_equation", + "content": "\\mathbb{S}^n" + }, + { + "bbox": [ + 55, + 72, + 296, + 337 + ], + "type": "text", + "content": " and enforce mutual orthogonality among Ideal Prototypes to mitigate conflict between new and prior knowledge. To achieve this, inspired by RankNet [6], we formulate a Hyperbolic positioning loss, " + }, + { + "bbox": [ + 55, + 72, + 296, + 337 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{HIP}}" + }, + { + "bbox": [ + 55, + 72, + 296, + 337 + ], + "type": "text", + "content": ", optimizing the ideal prototypes positions as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 342, + 294, + 390 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 342, + 294, + 390 + ], + "spans": [ + { + "bbox": [ + 65, + 342, + 294, + 390 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal{L}_{\\text{HIP}} = \\frac{1}{|M|}\\sum_{(i,i^{\\prime},i^{\\prime \\prime})\\in M} - \\bar{W}_{ii^{\\prime}i^{\\prime \\prime}}\\log W_{ii^{\\prime}i^{\\prime \\prime}} - \\\\ (1 - \\bar {W} _ {i i ^ {\\prime} i ^ {\\prime \\prime}}) \\log (1 - W _ {i i ^ {\\prime} i ^ {\\prime \\prime}}) \\tag {3} \\\\ \\end{array}", + "image_path": "3ba3b6fe4ef1e2f2acd2f459c906dcf6fcd79913aa7f0a5659860eebb1ddaa1a.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 394, + 296, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 394, + 296, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 394, + 296, + 464 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 394, + 296, + 464 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 394, + 296, + 464 + ], + "type": "text", + "content": " represents all class triplets, and " + }, + { + "bbox": [ + 55, + 394, + 296, + 464 + ], + "type": "inline_equation", + "content": "\\bar{W}_{ii^{\\prime}i^{\\prime \\prime}} = \\llbracket \\cos \\theta_{\\mathbf{s}_i,\\mathbf{s}_{i^{\\prime}}} \\geq \\cos \\theta_{\\mathbf{s}_i,\\mathbf{s}_{i^{\\prime \\prime}}} \\rrbracket" + }, + { + "bbox": [ + 55, + 394, + 296, + 464 + ], + "type": "text", + "content": " denotes triplet ranking order, with " + }, + { + "bbox": [ + 55, + 394, + 296, + 464 + ], + "type": "inline_equation", + "content": "W_{ii^{\\prime}i^{\\prime \\prime}} = \\frac{e^{o_{ii^{\\prime}}i^{\\prime \\prime}}}{1 + e^{o_{ii^{\\prime}}i^{\\prime \\prime}}} \\text{ and } o_{ii^{\\prime}i^{\\prime \\prime}} = \\cos \\theta_{\\mathbf{u}_i,\\mathbf{u}_{i^{\\prime}}} - \\cos \\theta_{\\mathbf{u}_i,\\mathbf{u}_{i^{\\prime \\prime}}}" + }, + { + "bbox": [ + 55, + 394, + 296, + 464 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 394, + 296, + 464 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_i, \\mathbf{u}_{i^{\\prime}}, \\mathbf{u}_{i^{\\prime \\prime}}" + }, + { + "bbox": [ + 55, + 394, + 296, + 464 + ], + "type": "text", + "content": " are ideal prototypes for the corresponding classes " + }, + { + "bbox": [ + 55, + 394, + 296, + 464 + ], + "type": "inline_equation", + "content": "i, i^{\\prime}" + }, + { + "bbox": [ + 55, + 394, + 296, + 464 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 394, + 296, + 464 + ], + "type": "inline_equation", + "content": "i^{\\prime \\prime}" + }, + { + "bbox": [ + 55, + 394, + 296, + 464 + ], + "type": "text", + "content": " respectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 464, + 296, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 464, + 296, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 296, + 571 + ], + "type": "text", + "content": "Following Mettes et al. [41], " + }, + { + "bbox": [ + 55, + 464, + 296, + 571 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{HIP}}" + }, + { + "bbox": [ + 55, + 464, + 296, + 571 + ], + "type": "text", + "content": " optimizes the prototypes' placement on " + }, + { + "bbox": [ + 55, + 464, + 296, + 571 + ], + "type": "inline_equation", + "content": "\\mathbb{B}_c^n" + }, + { + "bbox": [ + 55, + 464, + 296, + 571 + ], + "type": "text", + "content": " by aligning the hyperbolic prototypes with semantic ranking. This optimization establishes a classifier structure that facilitates inter-class separability and mutual orthogonality. Hyperbolic space's exponentially larger embedding capacity [44] further ensures that new classes integrate seamlessly without distorting existing embeddings, promoting robust feature organization across learning episodes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 576, + 244, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 244, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 244, + 588 + ], + "type": "text", + "content": "3.2.2. Hyperbolic class prototype projection" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 590, + 295, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 295, + 626 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 295, + 626 + ], + "type": "text", + "content": "We model the hierarchical complexity in 3D point clouds using a hyperbolic manifold representation, which preserves nuanced class relationships with minimal distortion." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 627, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 627, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 627, + 296, + 714 + ], + "type": "text", + "content": "At each session " + }, + { + "bbox": [ + 55, + 627, + 296, + 714 + ], + "type": "inline_equation", + "content": "t \\geq 0" + }, + { + "bbox": [ + 55, + 627, + 296, + 714 + ], + "type": "text", + "content": ", we apply the feature extractor " + }, + { + "bbox": [ + 55, + 627, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(:, \\theta)" + }, + { + "bbox": [ + 55, + 627, + 296, + 714 + ], + "type": "text", + "content": " on input point clouds " + }, + { + "bbox": [ + 55, + 627, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\{(P_k^{(t)})_{k=1}^{|D^{(t)}|}\\}" + }, + { + "bbox": [ + 55, + 627, + 296, + 714 + ], + "type": "text", + "content": " to obtain feature vectors " + }, + { + "bbox": [ + 55, + 627, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{v}" + }, + { + "bbox": [ + 55, + 627, + 296, + 714 + ], + "type": "text", + "content": " in Euclidean space. To avoid the computational cost of calculating Fréchet means directly in hyperbolic space [63], we compute class-mean feature vectors in Euclidean space and project them onto the hyperbolic manifold " + }, + { + "bbox": [ + 55, + 627, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbb{B}_c^n" + }, + { + "bbox": [ + 55, + 627, + 296, + 714 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 55, + 627, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\exp_0" + }, + { + "bbox": [ + 55, + 627, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 627, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 627, + 296, + 714 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 349, + 80, + 518, + 97 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 80, + 518, + 97 + ], + "spans": [ + { + "bbox": [ + 349, + 80, + 518, + 97 + ], + "type": "interline_equation", + "content": "v _ {k} ^ {(t)} = \\mathcal {E} \\left(P _ {k} ^ {(t)}; \\theta\\right), \\quad \\forall k \\in \\{1, \\dots , | D ^ {(t)} | \\}", + "image_path": "062fb64f9d38382b35d65266e86f2abfd6049df4758dd9a220ad601d8439a79f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 350, + 100, + 553, + 153 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 100, + 553, + 153 + ], + "spans": [ + { + "bbox": [ + 350, + 100, + 553, + 153 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\bar {v} \\left(c _ {i} ^ {(t)}\\right) = \\frac {1}{| D ^ {(t)} |} \\sum_ {k = 1} ^ {| D ^ {(t)} |} v _ {k} ^ {(t)} \\mathbb {1} \\left[ M _ {k} ^ {(t)} = c _ {i} ^ {(t)} \\right] \\tag {4} \\\\ z (c _ {i} ^ {(t)}) = \\mathcal {M} (\\exp_ {0} (\\overline {{v}} (c _ {i} ^ {(t)}))) \\\\ \\end{array}", + "image_path": "c1cb0d0e475841b8760270b726577b7408ab2ad916582a8288886f79583bf8e1.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 157, + 555, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 157, + 555, + 234 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 555, + 234 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 157, + 555, + 234 + ], + "type": "inline_equation", + "content": "i = \\{1,\\dots ,|C^{(t)}|\\}" + }, + { + "bbox": [ + 313, + 157, + 555, + 234 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 157, + 555, + 234 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 157, + 555, + 234 + ], + "type": "text", + "content": " denotes learnable parameters. Here, " + }, + { + "bbox": [ + 313, + 157, + 555, + 234 + ], + "type": "inline_equation", + "content": "z(c_{i}^{(t)})" + }, + { + "bbox": [ + 313, + 157, + 555, + 234 + ], + "type": "text", + "content": " represents the hyperbolic prototype for each class " + }, + { + "bbox": [ + 313, + 157, + 555, + 234 + ], + "type": "inline_equation", + "content": "c_{i}^{(t)}\\in C^{(t)}" + }, + { + "bbox": [ + 313, + 157, + 555, + 234 + ], + "type": "text", + "content": " within the open ball " + }, + { + "bbox": [ + 313, + 157, + 555, + 234 + ], + "type": "inline_equation", + "content": "\\mathbb{B}_c^n = \\{x\\in" + }, + { + "bbox": [ + 313, + 157, + 555, + 234 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^n:c\\| x\\| < 1\\}" + }, + { + "bbox": [ + 313, + 157, + 555, + 234 + ], + "type": "text", + "content": " . This projection process ensures cohesive, low-distortion representations of class structures as the model adapts incrementally to new classes." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 239, + 509, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 509, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 509, + 251 + ], + "type": "text", + "content": "3.2.3. Uncertainty-aware Busemann learning" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 253, + 555, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 253, + 555, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 253, + 555, + 376 + ], + "type": "text", + "content": "Previous works [12, 21, 27] highlight the utility of hyperbolic distance from any point to the origin " + }, + { + "bbox": [ + 313, + 253, + 555, + 376 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_{\\mathbb{B}}" + }, + { + "bbox": [ + 313, + 253, + 555, + 376 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 313, + 253, + 555, + 376 + ], + "type": "inline_equation", + "content": "\\mathbb{B}_c^n" + }, + { + "bbox": [ + 313, + 253, + 555, + 376 + ], + "type": "text", + "content": " as a measure of uncertainty, providing motivation to utilize the hyperbolic manifold for robust segmentation. In our approach, we align class prototypes with Ideal Prototypes placed at infinity on the hyperbolic boundary. To refine this alignment, we penalize the uncertainty of projected class features " + }, + { + "bbox": [ + 313, + 253, + 555, + 376 + ], + "type": "inline_equation", + "content": "z(c_i^{(t)})" + }, + { + "bbox": [ + 313, + 253, + 555, + 376 + ], + "type": "text", + "content": ", represented by their distance " + }, + { + "bbox": [ + 313, + 253, + 555, + 376 + ], + "type": "inline_equation", + "content": "d_{\\mathbb{B}}(z(c_i^{(t)}),\\mathcal{O}_{\\mathbb{B}})" + }, + { + "bbox": [ + 313, + 253, + 555, + 376 + ], + "type": "text", + "content": ", thereby preventing embedding drift from " + }, + { + "bbox": [ + 313, + 253, + 555, + 376 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_{\\mathbb{B}}" + }, + { + "bbox": [ + 313, + 253, + 555, + 376 + ], + "type": "text", + "content": " and ensuring stability." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 377, + 554, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 377, + 554, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 554, + 491 + ], + "type": "text", + "content": "These Ideal Prototypes lie at an infinite geodesic distance from all points within " + }, + { + "bbox": [ + 313, + 377, + 554, + 491 + ], + "type": "inline_equation", + "content": "\\mathbb{B}_c^n" + }, + { + "bbox": [ + 313, + 377, + 554, + 491 + ], + "type": "text", + "content": ", making direct prototypical learning impractical. We address this with the Busemann function " + }, + { + "bbox": [ + 313, + 377, + 554, + 491 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 313, + 377, + 554, + 491 + ], + "type": "text", + "content": " introduced in [7]. Ghadimi et al. [23] provides a closed-form expression for measuring such infinite geodesic distances. Let " + }, + { + "bbox": [ + 313, + 377, + 554, + 491 + ], + "type": "inline_equation", + "content": "u_{i}" + }, + { + "bbox": [ + 313, + 377, + 554, + 491 + ], + "type": "text", + "content": " be the Ideal Prototype for class " + }, + { + "bbox": [ + 313, + 377, + 554, + 491 + ], + "type": "inline_equation", + "content": "c_{i}^{(t)}" + }, + { + "bbox": [ + 313, + 377, + 554, + 491 + ], + "type": "text", + "content": " from the set " + }, + { + "bbox": [ + 313, + 377, + 554, + 491 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_n" + }, + { + "bbox": [ + 313, + 377, + 554, + 491 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 377, + 554, + 491 + ], + "type": "inline_equation", + "content": "\\gamma_{u_i^t}" + }, + { + "bbox": [ + 313, + 377, + 554, + 491 + ], + "type": "text", + "content": " a geodesic ray towards " + }, + { + "bbox": [ + 313, + 377, + 554, + 491 + ], + "type": "inline_equation", + "content": "u_{i}^{t}" + }, + { + "bbox": [ + 313, + 377, + 554, + 491 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 313, + 377, + 554, + 491 + ], + "type": "inline_equation", + "content": "z_{i}^{t} = z(c_{i}^{(t)})" + }, + { + "bbox": [ + 313, + 377, + 554, + 491 + ], + "type": "text", + "content": ". The Busemann function " + }, + { + "bbox": [ + 313, + 377, + 554, + 491 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 313, + 377, + 554, + 491 + ], + "type": "text", + "content": " for a hyperbolic embedding " + }, + { + "bbox": [ + 313, + 377, + 554, + 491 + ], + "type": "inline_equation", + "content": "z_{i}^{t}\\in \\mathbb{B}_{c}^{n}" + }, + { + "bbox": [ + 313, + 377, + 554, + 491 + ], + "type": "text", + "content": " relative to " + }, + { + "bbox": [ + 313, + 377, + 554, + 491 + ], + "type": "inline_equation", + "content": "u_{i}^{t}" + }, + { + "bbox": [ + 313, + 377, + 554, + 491 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 356, + 498, + 555, + 519 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 498, + 555, + 519 + ], + "spans": [ + { + "bbox": [ + 356, + 498, + 555, + 519 + ], + "type": "interline_equation", + "content": "\\mathcal {B} _ {u _ {i} ^ {t}} \\left(z _ {i} ^ {t}\\right) = \\lim _ {q \\rightarrow \\infty} \\left(d _ {\\mathbb {B}} \\left(\\gamma_ {u _ {i} ^ {t}} (q), z _ {i} ^ {t}\\right) - q\\right) \\tag {5}", + "image_path": "62956df0a75d0ac60a12936e5fb7f9c93445fba782e04f411474f73924b6bc1d.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 326, + 522, + 496, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 522, + 496, + 534 + ], + "spans": [ + { + "bbox": [ + 326, + 522, + 496, + 534 + ], + "type": "text", + "content": "In the Poincaré ball " + }, + { + "bbox": [ + 326, + 522, + 496, + 534 + ], + "type": "inline_equation", + "content": "\\mathbb{B}_c^n" + }, + { + "bbox": [ + 326, + 522, + 496, + 534 + ], + "type": "text", + "content": ", this simplifies to:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 378, + 542, + 553, + 570 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 542, + 553, + 570 + ], + "spans": [ + { + "bbox": [ + 378, + 542, + 553, + 570 + ], + "type": "interline_equation", + "content": "\\mathcal {B} _ {u _ {i} ^ {t}} \\left(z _ {i} ^ {t}\\right) = \\log \\frac {\\left\\| u _ {i} ^ {t} - z _ {i} ^ {t} \\right\\| ^ {2}}{1 - \\left\\| z _ {i} ^ {t} \\right\\| ^ {2}}. \\tag {6}", + "image_path": "6b00ddccbf51cdb4268d6872163f9e17b6efb754dd34ce9928e4ebc89a09423a.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 572, + 554, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 572, + 554, + 620 + ], + "spans": [ + { + "bbox": [ + 313, + 572, + 554, + 620 + ], + "type": "text", + "content": "We also incorporate hyperbolic distance as an uncertainty measure for class prototypes in " + }, + { + "bbox": [ + 313, + 572, + 554, + 620 + ], + "type": "inline_equation", + "content": "\\mathbb{B}_c^n" + }, + { + "bbox": [ + 313, + 572, + 554, + 620 + ], + "type": "text", + "content": ", which integrates into our loss function as in [23]. The alignment loss introduced in Eq. 2 is defined as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 367, + 625, + 553, + 639 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 625, + 553, + 639 + ], + "spans": [ + { + "bbox": [ + 367, + 625, + 553, + 639 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {A l g i n m e n t}} = \\mathcal {L} _ {\\mathcal {B}} + \\phi \\mathcal {L} _ {\\text {U n c e r t a i n t y}} \\tag {7}", + "image_path": "810ba436acae7440811966c26a22dc30c4a513407d1104e199d036d9ab469954.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 644, + 554, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 644, + 554, + 681 + ], + "spans": [ + { + "bbox": [ + 313, + 644, + 554, + 681 + ], + "type": "text", + "content": "where Busemann loss " + }, + { + "bbox": [ + 313, + 644, + 554, + 681 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathcal{B}} = \\mathcal{B}_{u_i^t}(z_i^t)" + }, + { + "bbox": [ + 313, + 644, + 554, + 681 + ], + "type": "text", + "content": ", uncertainty loss " + }, + { + "bbox": [ + 313, + 644, + 554, + 681 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Uncertainty}} = d_{\\mathbb{B}}(z_i^t,\\mathcal{O}_{\\mathbb{B}})" + }, + { + "bbox": [ + 313, + 644, + 554, + 681 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 644, + 554, + 681 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 644, + 554, + 681 + ], + "type": "text", + "content": " is the strength of the regularization effect of uncertainty loss." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 326, + 682, + 550, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 682, + 550, + 694 + ], + "spans": [ + { + "bbox": [ + 326, + 682, + 550, + 694 + ], + "type": "text", + "content": "The overall loss function in Eq. 2 can be formulated as:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 359, + 701, + 509, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 701, + 509, + 714 + ], + "spans": [ + { + "bbox": [ + 359, + 701, + 509, + 714 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {H i P o}} = \\mathcal {L} _ {\\mathcal {B}} + \\phi \\mathcal {L} _ {\\text {U n c e r t a i n t y}} + \\mathcal {L} _ {\\mathrm {C E}}", + "image_path": "18f40377255533efc671e464ea1ff9c8d419144ab13511215621929af46c11e5.jpg" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11814" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "content": "At test time, we employ only the point-level classification branch to segment 3D point clouds." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 108, + 205, + 123 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 108, + 205, + 123 + ], + "spans": [ + { + "bbox": [ + 55, + 108, + 205, + 123 + ], + "type": "text", + "content": "4. Experimental Evaluations" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 129, + 296, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 129, + 296, + 297 + ], + "spans": [ + { + "bbox": [ + 54, + 129, + 296, + 297 + ], + "type": "text", + "content": "Datasets: We evaluate our method on two widely used benchmarks, S3DIS [2] and ScanNetv2 [17], to assess both in-domain and cross-dataset FSCIL performance. These datasets incorporate large-scale coloured 3D point clouds of indoor spaces which introduce semantics. See Sup. Mat. for additional dataset details. For cross-dataset FSCIL, we create a transition dataset from ScanNetv2 to S3DIS, using ten classes from ScanNetv2 as the base and introducing eight new classes from S3DIS in incremental sessions. Following a strict disjoint setup as in [10], each session includes only novel classes. The number of introduced classes varies per session to assess resilience to catastrophic forgetting, which escalates with more incremental sessions in point cloud recognition [60]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 297, + 296, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 297, + 296, + 527 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 296, + 527 + ], + "type": "text", + "content": "Evaluation metrics: We assess model performance using three standard metrics for incremental learning[70]: Average Incremental Accuracy (AIA), which tracks average accuracy across tasks learned sequentially; Last, the accuracy after the final task; and Average Forgetting Rate " + }, + { + "bbox": [ + 55, + 297, + 296, + 527 + ], + "type": "inline_equation", + "content": "(\\mathcal{F}_{Last}^{(T)})" + }, + { + "bbox": [ + 55, + 297, + 296, + 527 + ], + "type": "text", + "content": ", which measures performance decline on learned tasks after adapting to the last task. Finally, we report mean Intersection-over-Union " + }, + { + "bbox": [ + 55, + 297, + 296, + 527 + ], + "type": "inline_equation", + "content": "(\\mathrm{mIoU})^1" + }, + { + "bbox": [ + 55, + 297, + 296, + 527 + ], + "type": "text", + "content": " to evaluate segmentation performance. See Sup. Mat. for detailed metric definitions. Architecture: Our backbone model employs PointTransformer [77], leveraging recent advances in point cloud recognition, with a Hyperbolic Classifier head " + }, + { + "bbox": [ + 55, + 297, + 296, + 527 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(.)" + }, + { + "bbox": [ + 55, + 297, + 296, + 527 + ], + "type": "text", + "content": " for segmentation. We train HiPo using the Riemannian SGD optimizer, facilitated by geoopt [29] for hyperbolic operations. Training is conducted on both " + }, + { + "bbox": [ + 55, + 297, + 296, + 527 + ], + "type": "inline_equation", + "content": "D^{(0)}" + }, + { + "bbox": [ + 55, + 297, + 296, + 527 + ], + "type": "text", + "content": " and incremental datasets " + }, + { + "bbox": [ + 55, + 297, + 296, + 527 + ], + "type": "inline_equation", + "content": "D^{(t)}, t > 0" + }, + { + "bbox": [ + 55, + 297, + 296, + 527 + ], + "type": "text", + "content": " with a batch size of 16, AdamW optimizer [37], initial learning rate of 0.001, and weight decay of 0.0001 over 100 epochs. Architecture diagram and further hyper-parameter details are available in the Sup. Mat." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 527, + 296, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 527, + 296, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 527, + 296, + 694 + ], + "type": "text", + "content": "Competitors: We compare our method against various baselines representing distinct FSCIL approaches. For regularization-based methods, we include LwF with Teacher Adaptation [59], and for optimization-based methods, we include C-FSCIL [24]. We also evaluate against regularization-based FSCIL methods specifically designed for 3D point cloud tasks, such as LGKD [72], SDCOT [79], and 3DPC-CISS [70], which have shown effectiveness with the PointTransformer [77] architecture. Comparative results are provided in Tables 1 and 2. Additionally, we adapt BiDist [78], CLOM [85], TEEN [64], FACT [81], and ALICE [50] using the DGCNN [65] backbone for FSCIL in 3D point cloud segmentation. We report the results with the DGCNN [65] backbone in Table 5." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 362, + 125 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 362, + 125 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 362, + 125 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 362, + 125 + ], + "type": "image", + "image_path": "038eecf366649802e0a09bacdf7996f2bab9e045ce324bb210c550db31fd3e84.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 364, + 80, + 409, + 125 + ], + "blocks": [ + { + "bbox": [ + 370, + 72, + 405, + 79 + ], + "lines": [ + { + "bbox": [ + 370, + 72, + 405, + 79 + ], + "spans": [ + { + "bbox": [ + 370, + 72, + 405, + 79 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 364, + 80, + 409, + 125 + ], + "lines": [ + { + "bbox": [ + 364, + 80, + 409, + 125 + ], + "spans": [ + { + "bbox": [ + 364, + 80, + 409, + 125 + ], + "type": "image", + "image_path": "bc810bb5e6235004291570e9db32336890bef6190cca36c1fa7e35e6be343cac.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 411, + 80, + 457, + 125 + ], + "blocks": [ + { + "bbox": [ + 424, + 72, + 446, + 79 + ], + "lines": [ + { + "bbox": [ + 424, + 72, + 446, + 79 + ], + "spans": [ + { + "bbox": [ + 424, + 72, + 446, + 79 + ], + "type": "text", + "content": "SDCOT" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 411, + 80, + 457, + 125 + ], + "lines": [ + { + "bbox": [ + 411, + 80, + 457, + 125 + ], + "spans": [ + { + "bbox": [ + 411, + 80, + 457, + 125 + ], + "type": "image", + "image_path": "44ff4e68a9d72ff3fdb4160813e4730863929139856510e83ac99345f0c25498.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 460, + 80, + 505, + 125 + ], + "blocks": [ + { + "bbox": [ + 468, + 72, + 498, + 79 + ], + "lines": [ + { + "bbox": [ + 468, + 72, + 498, + 79 + ], + "spans": [ + { + "bbox": [ + 468, + 72, + 498, + 79 + ], + "type": "text", + "content": "3DPC-CISS" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 460, + 80, + 505, + 125 + ], + "lines": [ + { + "bbox": [ + 460, + 80, + 505, + 125 + ], + "spans": [ + { + "bbox": [ + 460, + 80, + 505, + 125 + ], + "type": "image", + "image_path": "d7fc9e4b09a38c2727df04c1fe888dddbb912a2b587fad7ebc72836b2a6de59d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 508, + 81, + 553, + 126 + ], + "blocks": [ + { + "bbox": [ + 515, + 72, + 544, + 79 + ], + "lines": [ + { + "bbox": [ + 515, + 72, + 544, + 79 + ], + "spans": [ + { + "bbox": [ + 515, + 72, + 544, + 79 + ], + "type": "text", + "content": "HIPO (Ours)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 508, + 81, + 553, + 126 + ], + "lines": [ + { + "bbox": [ + 508, + 81, + 553, + 126 + ], + "spans": [ + { + "bbox": [ + 508, + 81, + 553, + 126 + ], + "type": "image", + "image_path": "b4a33ea4b4a1916b2987e9fc716386469aa84c4a1d892fb72cbf413380368827.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 317, + 129, + 361, + 175 + ], + "blocks": [ + { + "bbox": [ + 317, + 129, + 361, + 175 + ], + "lines": [ + { + "bbox": [ + 317, + 129, + 361, + 175 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 361, + 175 + ], + "type": "image", + "image_path": "2a70add8cc859c470ac47d259adb1ea9bca6c7cb6927eb57925fc7637af18590.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 364, + 130, + 409, + 175 + ], + "blocks": [ + { + "bbox": [ + 364, + 130, + 409, + 175 + ], + "lines": [ + { + "bbox": [ + 364, + 130, + 409, + 175 + ], + "spans": [ + { + "bbox": [ + 364, + 130, + 409, + 175 + ], + "type": "image", + "image_path": "66d2ace25701a2759e6ba19ce575d48a4d05f98a5f98cc608858b0f92cb07b5c.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 413, + 129, + 457, + 175 + ], + "blocks": [ + { + "bbox": [ + 413, + 129, + 457, + 175 + ], + "lines": [ + { + "bbox": [ + 413, + 129, + 457, + 175 + ], + "spans": [ + { + "bbox": [ + 413, + 129, + 457, + 175 + ], + "type": "image", + "image_path": "689ae9ddddceac5aaac41b4ffc222299d136daaa8f43c0ef6121c9784d3e1511.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 461, + 130, + 504, + 175 + ], + "blocks": [ + { + "bbox": [ + 461, + 130, + 504, + 175 + ], + "lines": [ + { + "bbox": [ + 461, + 130, + 504, + 175 + ], + "spans": [ + { + "bbox": [ + 461, + 130, + 504, + 175 + ], + "type": "image", + "image_path": "5ccdbabf1fceb3a3a0510900afd8dff4b950675bf7bfcdfe15c3973febd84392.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 509, + 131, + 553, + 176 + ], + "blocks": [ + { + "bbox": [ + 509, + 131, + 553, + 176 + ], + "lines": [ + { + "bbox": [ + 509, + 131, + 553, + 176 + ], + "spans": [ + { + "bbox": [ + 509, + 131, + 553, + 176 + ], + "type": "image", + "image_path": "0d2f5fa6adbc042bddbb0ca62dc9d356fd172f98a6a53a316d878a581085b767.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 317, + 178, + 361, + 224 + ], + "blocks": [ + { + "bbox": [ + 317, + 178, + 361, + 224 + ], + "lines": [ + { + "bbox": [ + 317, + 178, + 361, + 224 + ], + "spans": [ + { + "bbox": [ + 317, + 178, + 361, + 224 + ], + "type": "image", + "image_path": "7202c6e0419e65a485e5c5e8d5dd359cd6a5ef5f67e58bf8b93c5c637ac8d2c6.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 364, + 178, + 409, + 224 + ], + "blocks": [ + { + "bbox": [ + 364, + 178, + 409, + 224 + ], + "lines": [ + { + "bbox": [ + 364, + 178, + 409, + 224 + ], + "spans": [ + { + "bbox": [ + 364, + 178, + 409, + 224 + ], + "type": "image", + "image_path": "3cddb91826033ec2865b8ab68a286af7a82bd16390d1daa8bd396a2106736f56.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 413, + 178, + 457, + 224 + ], + "blocks": [ + { + "bbox": [ + 413, + 178, + 457, + 224 + ], + "lines": [ + { + "bbox": [ + 413, + 178, + 457, + 224 + ], + "spans": [ + { + "bbox": [ + 413, + 178, + 457, + 224 + ], + "type": "image", + "image_path": "8438730be9ff05adc0e539f3f01f920f5da7ba88dc280bf51a999211acf7d799.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 460, + 178, + 505, + 224 + ], + "blocks": [ + { + "bbox": [ + 460, + 178, + 505, + 224 + ], + "lines": [ + { + "bbox": [ + 460, + 178, + 505, + 224 + ], + "spans": [ + { + "bbox": [ + 460, + 178, + 505, + 224 + ], + "type": "image", + "image_path": "94f742444df084e76f69c4bde7648aa71225bb834ceca96684ec01684f1374e1.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 509, + 178, + 553, + 224 + ], + "blocks": [ + { + "bbox": [ + 509, + 178, + 553, + 224 + ], + "lines": [ + { + "bbox": [ + 509, + 178, + 553, + 224 + ], + "spans": [ + { + "bbox": [ + 509, + 178, + 553, + 224 + ], + "type": "image", + "image_path": "af36710e687a2e9e34fcb8e9532ec24155b01b7fb1c2240f5e3122e7a91e23f1.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 317, + 228, + 361, + 273 + ], + "blocks": [ + { + "bbox": [ + 317, + 228, + 361, + 273 + ], + "lines": [ + { + "bbox": [ + 317, + 228, + 361, + 273 + ], + "spans": [ + { + "bbox": [ + 317, + 228, + 361, + 273 + ], + "type": "image", + "image_path": "906504720ac575679d1fb765e766e279b1ae7c0d78c833f9442d0b5ff82d6860.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 364, + 228, + 409, + 273 + ], + "blocks": [ + { + "bbox": [ + 364, + 228, + 409, + 273 + ], + "lines": [ + { + "bbox": [ + 364, + 228, + 409, + 273 + ], + "spans": [ + { + "bbox": [ + 364, + 228, + 409, + 273 + ], + "type": "image", + "image_path": "f0b048e14c5365325ffc22e2185a71b48329abe1b96152e3aa7a1a0b03339530.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 413, + 228, + 457, + 273 + ], + "blocks": [ + { + "bbox": [ + 413, + 228, + 457, + 273 + ], + "lines": [ + { + "bbox": [ + 413, + 228, + 457, + 273 + ], + "spans": [ + { + "bbox": [ + 413, + 228, + 457, + 273 + ], + "type": "image", + "image_path": "76ffc8307d09e60aa534b7d7ac1a25336137a917f3fd0211a30befcfcb96ee22.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 460, + 228, + 505, + 273 + ], + "blocks": [ + { + "bbox": [ + 460, + 228, + 505, + 273 + ], + "lines": [ + { + "bbox": [ + 460, + 228, + 505, + 273 + ], + "spans": [ + { + "bbox": [ + 460, + 228, + 505, + 273 + ], + "type": "image", + "image_path": "76e131ab2723234f675ca50c055ace3bb135891c754cd08860dbec1ef118dad5.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 509, + 228, + 553, + 273 + ], + "blocks": [ + { + "bbox": [ + 509, + 228, + 553, + 273 + ], + "lines": [ + { + "bbox": [ + 509, + 228, + 553, + 273 + ], + "spans": [ + { + "bbox": [ + 509, + 228, + 553, + 273 + ], + "type": "image", + "image_path": "091f65815f6411c69dfe721f0f41087fdc9561bade3887c56052204cb3ef4a74.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 352, + 278, + 434, + 315 + ], + "blocks": [ + { + "bbox": [ + 352, + 278, + 434, + 315 + ], + "lines": [ + { + "bbox": [ + 352, + 278, + 434, + 315 + ], + "spans": [ + { + "bbox": [ + 352, + 278, + 434, + 315 + ], + "type": "image", + "image_path": "74c8fd02f2dea6ca4287fad05266344d10aa971dc7ee7555ddcfe60b010567cc.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 323, + 555, + 389 + ], + "lines": [ + { + "bbox": [ + 313, + 323, + 555, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 555, + 389 + ], + "type": "text", + "content": "Figure 4. Qualitative visualization of model performance: Comparing FSCIL for 3D point cloud segmentation results of our method (HiPo) and other baselines for S3DIS dataset [2]. We observe that even with a few-shot setup, our model performs reasonably well in segmenting out small parts like walls, counters, etc., compared to other methods [70, 79]." + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 441, + 278, + 510, + 314 + ], + "blocks": [ + { + "bbox": [ + 441, + 278, + 510, + 314 + ], + "lines": [ + { + "bbox": [ + 441, + 278, + 510, + 314 + ], + "spans": [ + { + "bbox": [ + 441, + 278, + 510, + 314 + ], + "type": "image", + "image_path": "5f6c35fe1cab7a75bf8ec7ac1ebd79287042afe875ce6f02e870616e329ae0cd.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 403, + 463, + 416 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 463, + 416 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 463, + 416 + ], + "type": "text", + "content": "4.1. Discussing the main results" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 423, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 555, + 715 + ], + "type": "text", + "content": "Table 1 presents a detailed comparison of our proposed method, HiPo, against several baseline methods on the S3DIS [2] and ScanNetv2 [17] datasets across various evaluation settings. HiPo consistently outperforms the current state-of-the-art approach, 3DPC-CISS [70], achieving an average improvement of " + }, + { + "bbox": [ + 313, + 423, + 555, + 715 + ], + "type": "inline_equation", + "content": "+2.57\\%" + }, + { + "bbox": [ + 313, + 423, + 555, + 715 + ], + "type": "text", + "content": " mIoU on S3DIS and " + }, + { + "bbox": [ + 313, + 423, + 555, + 715 + ], + "type": "inline_equation", + "content": "+2.12\\%" + }, + { + "bbox": [ + 313, + 423, + 555, + 715 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 313, + 423, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{Last}^{(T)}" + }, + { + "bbox": [ + 313, + 423, + 555, + 715 + ], + "type": "text", + "content": ". This improvement underscores the effectiveness of HiPo's hyperbolic, uncertainty-aware architecture, which mitigates catastrophic forgetting by reducing misalignment between features and prototype representations from earlier sessions. When compared to other baselines, such as LGKD [72] and C-FSCIL [24], HiPo shows substantial gains across all metrics. For instance, in the 10-3T scenario, HiPo achieves " + }, + { + "bbox": [ + 313, + 423, + 555, + 715 + ], + "type": "inline_equation", + "content": "33.40\\%" + }, + { + "bbox": [ + 313, + 423, + 555, + 715 + ], + "type": "text", + "content": " mIoU on the Last metric, significantly higher than LGKD's " + }, + { + "bbox": [ + 313, + 423, + 555, + 715 + ], + "type": "inline_equation", + "content": "20.02\\%" + }, + { + "bbox": [ + 313, + 423, + 555, + 715 + ], + "type": "text", + "content": " and C-FSCIL's " + }, + { + "bbox": [ + 313, + 423, + 555, + 715 + ], + "type": "inline_equation", + "content": "25.02\\%" + }, + { + "bbox": [ + 313, + 423, + 555, + 715 + ], + "type": "text", + "content": ". This consistent superiority highlights HiPo's resilience against forgetting, particularly in incremental sessions with larger class increments, where other methods struggle to retain performance. Furthermore, while SDCOT [79] and LwF [59] suffer from high forgetting rates, HiPo effectively maintains stable performance across incremental tasks, with a " + }, + { + "bbox": [ + 313, + 423, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{Last}^{(T)}" + }, + { + "bbox": [ + 313, + 423, + 555, + 715 + ], + "type": "text", + "content": " rate as low as " + }, + { + "bbox": [ + 313, + 423, + 555, + 715 + ], + "type": "inline_equation", + "content": "17.70\\%" + }, + { + "bbox": [ + 313, + 423, + 555, + 715 + ], + "type": "text", + "content": " in challenging scenarios, compared to SDCOT's " + }, + { + "bbox": [ + 313, + 423, + 555, + 715 + ], + "type": "inline_equation", + "content": "30.83\\%" + }, + { + "bbox": [ + 313, + 423, + 555, + 715 + ], + "type": "text", + "content": ". This robustness can be attributed to HiPo's hyperbolic embed" + } + ] + } + ], + "index": 33 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 66, + 702, + 212, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 212, + 714 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 212, + 714 + ], + "type": "text", + "content": "1We use accuracy and mIoU interchangeably." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "11815" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 70, + 541, + 157 + ], + "blocks": [ + { + "bbox": [ + 70, + 70, + 541, + 157 + ], + "lines": [ + { + "bbox": [ + 70, + 70, + 541, + 157 + ], + "spans": [ + { + "bbox": [ + 70, + 70, + 541, + 157 + ], + "type": "table", + "html": "
4-3T7-3T10-3TAverage
Last(↑)AIA(↑)FT(Last)(↓)Last(↑)AIA(↑)FT(Last)(↓)Last(↑)AIA(↑)FT(Last)(↓)Last(↑)AIA(↑)FT(Last)(↓)
LwF [59]4.3322.0523.628.6021.8019.8021.3034.0525.5011.4125.3022.97
C-FSCIL [24]9.0627.9025.1120.0131.9217.8625.0235.3620.6818.0331.0621.88
LGKD [72]12.2431.8025.9419.7630.4115.9720.0232.7125.3817.3431.6422.43
SDCOT [79]5.3028.4230.8312.0029.4626.2023.8035.3023.0013.7031.0626.67
3DPC-CISS [70]25.7040.5019.7324.2034.5015.4529.6038.2017.2026.5037.7317.46
HtPo30.6043.8717.7029.4036.8011.1033.4040.2513.7031.1340.3014.16
", + "image_path": "c2378a8f86726d8baf0fc468e8cea80c492f477959f279185af4ae00526aeb0f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 70, + 159, + 541, + 246 + ], + "blocks": [ + { + "bbox": [ + 70, + 159, + 541, + 246 + ], + "lines": [ + { + "bbox": [ + 70, + 159, + 541, + 246 + ], + "spans": [ + { + "bbox": [ + 70, + 159, + 541, + 246 + ], + "type": "table", + "html": "
10-2T10-5T15-5TAverage
Last(↑)AIA(↑)FT(Last)(↓)Last(↑)AIA(↑)FT(Last)(↓)Last(↑)AIA(↑)FT(Last)(↓)Last(↑)AIA(↑)FT(Last)(↓)
LwF [59]1.7010.7510.861.0517.5324.725.0621.3832.642.6016.5522.74
C-FSCIL [24]7.1215.9410.5811.1224.1819.5911.5023.5824.159.2521.2318.11
LGKD [72]6.0215.6011.509.5525.5222.4511.1023.8325.468.2221.6519.80
SDCOT [79]0.509.5610.880.4017.8326.151.3019.5036.400.7315.6324.47
3DPC-CISS [70]6.2013.138.328.1022.3021.3012.2024.9525.508.8320.1218.37
HtPo8.5017.1310.3614.6527.2218.8511.9024.4525.1011.6822.2718.12
", + "image_path": "27ba110e83a4291187bd08365a383957bc5232603857e4cde8552030dfa6bf25.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 89, + 317, + 264, + 407 + ], + "blocks": [ + { + "bbox": [ + 54, + 254, + 555, + 301 + ], + "lines": [ + { + "bbox": [ + 54, + 254, + 555, + 301 + ], + "spans": [ + { + "bbox": [ + 54, + 254, + 555, + 301 + ], + "type": "text", + "content": "Table 1. Comparison with literature on the S3DIS [2] (top table) and ScanNetv2 [17] (bottom table) datasets, using the PointTransformer backbone [77] in a 5-shot setting. CIL mIoU (%) is reported. “" + }, + { + "bbox": [ + 54, + 254, + 555, + 301 + ], + "type": "inline_equation", + "content": "M-NT" + }, + { + "bbox": [ + 54, + 254, + 555, + 301 + ], + "type": "text", + "content": "” indicates that " + }, + { + "bbox": [ + 54, + 254, + 555, + 301 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 54, + 254, + 555, + 301 + ], + "type": "text", + "content": " is the number of classes in the base session " + }, + { + "bbox": [ + 54, + 254, + 555, + 301 + ], + "type": "inline_equation", + "content": "S^{(0)}" + }, + { + "bbox": [ + 54, + 254, + 555, + 301 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 54, + 254, + 555, + 301 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 54, + 254, + 555, + 301 + ], + "type": "text", + "content": " is the number of classes introduced in each incremental session " + }, + { + "bbox": [ + 54, + 254, + 555, + 301 + ], + "type": "inline_equation", + "content": "S^{(t)}" + }, + { + "bbox": [ + 54, + 254, + 555, + 301 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 254, + 555, + 301 + ], + "type": "inline_equation", + "content": "t > 0" + }, + { + "bbox": [ + 54, + 254, + 555, + 301 + ], + "type": "text", + "content": ". The highest performance in each column is shown in bold, with the second-highest results underlined. HiPO performs superior compared to other baselines in different incremental settings." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 89, + 317, + 264, + 407 + ], + "lines": [ + { + "bbox": [ + 89, + 317, + 264, + 407 + ], + "spans": [ + { + "bbox": [ + 89, + 317, + 264, + 407 + ], + "type": "table", + "html": "
Average
Last (↑)AIA (↑)F(T)Last (↓)
LwF [59]8.7822.6023.94
C-FSCIL [24]13.5525.0219.72
LGKD [72]15.6325.5617.51
SDCOT [79]19.2229.2917.52
3DPC-CISS [70]19.9529.1616.01
HtPo20.6029.8116.05
", + "image_path": "e41b3320d312f5056801e29df1a3648d38e8de24f103c8000715dd00f4106b0e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 415, + 295, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 415, + 295, + 471 + ], + "spans": [ + { + "bbox": [ + 54, + 415, + 295, + 471 + ], + "type": "text", + "content": "Table 2. SOTA comparison on ScanNetv2 " + }, + { + "bbox": [ + 54, + 415, + 295, + 471 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 54, + 415, + 295, + 471 + ], + "type": "text", + "content": " S3DIS dataset. We compute the average of 10-4T and 10-8T with PointTransformer backbone [77] in 5-shot setting. HIPO achieves superior Last and AIA values with comparable " + }, + { + "bbox": [ + 54, + 415, + 295, + 471 + ], + "type": "inline_equation", + "content": "F_{Last}^{(T)}" + }, + { + "bbox": [ + 54, + 415, + 295, + 471 + ], + "type": "text", + "content": " values, highlighting better domain adaptability than other methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 472, + 295, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 472, + 295, + 508 + ], + "spans": [ + { + "bbox": [ + 54, + 472, + 295, + 508 + ], + "type": "text", + "content": "ding space, which aligns naturally with hierarchical structures, thereby enhancing inter-class separation and intraclass compactness over time." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 509, + 295, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 509, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 54, + 509, + 295, + 568 + ], + "type": "text", + "content": "Moreover, Fig. 4 provides qualitative comparisons, where HiPO demonstrates marked improvements over leading baselines, particularly in accurately segmenting challenging structures such as walls and counters. Further analyses and results are available in the Sup. Mat." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 576, + 185, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 185, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 185, + 590 + ], + "type": "text", + "content": "4.2. Main ablation analysis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 594, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 594, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 54, + 594, + 295, + 700 + ], + "type": "text", + "content": "Cross-dataset class-incremental segmentation: To further assess the effectiveness of our proposed modules, we conduct cross-dataset incremental segmentation experiments (ScanNetv2 " + }, + { + "bbox": [ + 54, + 594, + 295, + 700 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 54, + 594, + 295, + 700 + ], + "type": "text", + "content": " S3DIS). Detailed setup information is in Sup. Mat. As shown in Table 2, the cross-dataset setting introduces a notable increase in forgetting due to domain shift. However, the results underscore HiPo's robustness in managing such domain shifts effectively, with improvements of " + }, + { + "bbox": [ + 54, + 594, + 295, + 700 + ], + "type": "inline_equation", + "content": "0.65\\%" + }, + { + "bbox": [ + 54, + 594, + 295, + 700 + ], + "type": "text", + "content": " in Last and " + }, + { + "bbox": [ + 54, + 594, + 295, + 700 + ], + "type": "inline_equation", + "content": "0.52\\%" + }, + { + "bbox": [ + 54, + 594, + 295, + 700 + ], + "type": "text", + "content": " in AIA metrics." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "content": "Ablation studies on penalty in Busemann loss: We pro" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 319, + 555, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 555, + 475 + ], + "type": "text", + "content": "vide a comprehensive ablation study on the regularization constant " + }, + { + "bbox": [ + 313, + 319, + 555, + 475 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 319, + 555, + 475 + ], + "type": "text", + "content": " introduced in Eq. 7. The regularization strength is varied from " + }, + { + "bbox": [ + 313, + 319, + 555, + 475 + ], + "type": "inline_equation", + "content": "\\phi = 0" + }, + { + "bbox": [ + 313, + 319, + 555, + 475 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 319, + 555, + 475 + ], + "type": "inline_equation", + "content": "\\phi = 1" + }, + { + "bbox": [ + 313, + 319, + 555, + 475 + ], + "type": "text", + "content": " in intervals of 0.25 for S3DIS [2] (4-3T, 7-3T, 10-3T) and ScanNetv2 [17] (10-2T, 10-5T, 15-5T). As shown in Table 3, our findings corroborate prior work [23]: without a penalty " + }, + { + "bbox": [ + 313, + 319, + 555, + 475 + ], + "type": "inline_equation", + "content": "(\\phi = 0)" + }, + { + "bbox": [ + 313, + 319, + 555, + 475 + ], + "type": "text", + "content": ", samples are overly drawn towards the Ideal Prototypes, resulting in overconfident predictions and lower AIA on test data " + }, + { + "bbox": [ + 313, + 319, + 555, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{test}" + }, + { + "bbox": [ + 313, + 319, + 555, + 475 + ], + "type": "text", + "content": " in addition to vanishing gradient problems [44]. Conversely, a high penalty " + }, + { + "bbox": [ + 313, + 319, + 555, + 475 + ], + "type": "inline_equation", + "content": "(\\phi = 1)" + }, + { + "bbox": [ + 313, + 319, + 555, + 475 + ], + "type": "text", + "content": " increases inter-class confusion, also leading to lower AIA on " + }, + { + "bbox": [ + 313, + 319, + 555, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{test}" + }, + { + "bbox": [ + 313, + 319, + 555, + 475 + ], + "type": "text", + "content": ". Our final chosen results, highlighted in bold in Table 3, reflect the optimal balance for " + }, + { + "bbox": [ + 313, + 319, + 555, + 475 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 319, + 555, + 475 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 479, + 556, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 556, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 556, + 635 + ], + "type": "text", + "content": "Comparison of metrics at different curvatures: Following [47], we experimented with curvature values of " + }, + { + "bbox": [ + 313, + 479, + 556, + 635 + ], + "type": "inline_equation", + "content": "c = 0.0001" + }, + { + "bbox": [ + 313, + 479, + 556, + 635 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 479, + 556, + 635 + ], + "type": "inline_equation", + "content": "c = 0.001" + }, + { + "bbox": [ + 313, + 479, + 556, + 635 + ], + "type": "text", + "content": " to implement traditional Euclidean architectures (Poincaré Ball with zero curvature; very small Poincaré ball curvature values are chosen as curvature " + }, + { + "bbox": [ + 313, + 479, + 556, + 635 + ], + "type": "inline_equation", + "content": "c = 0" + }, + { + "bbox": [ + 313, + 479, + 556, + 635 + ], + "type": "text", + "content": " is numerically unstable [27]) for the S3DIS [2] (4-3T, 7-3T, 10-3T) and ScanNetv2 [17] (10-2T, 10-5T, 15-5T) setups. Table 4 shows that our HiPO with a Poincaré Ball curvature of " + }, + { + "bbox": [ + 313, + 479, + 556, + 635 + ], + "type": "inline_equation", + "content": "c = -1.00" + }, + { + "bbox": [ + 313, + 479, + 556, + 635 + ], + "type": "text", + "content": " consistently outperforms its Euclidean counterpart across most settings. This finding demonstrates that Riemannian geometry leverages the hierarchical structure of point clouds more effectively than Euclidean space, suggesting further exploration in this domain." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 639, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 639, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 639, + 556, + 715 + ], + "type": "text", + "content": "Sensitivity to the different number of training samples in each incremental episode: To assess the impact of data availability on model performance, we gradually increase the number of samples per class (shot) from one to all available examples and analyze the Last mIoU and " + }, + { + "bbox": [ + 313, + 639, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{Last}^{(T)}" + }, + { + "bbox": [ + 313, + 639, + 556, + 715 + ], + "type": "text", + "content": " results for different shot values in Fig. 5. Our findings indi" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "11816" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 124, + 70, + 491, + 156 + ], + "blocks": [ + { + "bbox": [ + 124, + 70, + 491, + 156 + ], + "lines": [ + { + "bbox": [ + 124, + 70, + 491, + 156 + ], + "spans": [ + { + "bbox": [ + 124, + 70, + 491, + 156 + ], + "type": "table", + "html": "
DatasetsSettingsφ = 0.00φ = 0.25φ = 0.50φ = 0.75φ = 1.00
AIA↑FT(St)↓AIA↑FT(St)↓AIA↑FT(St)↓AIA↑FT(St)↓AIA↑FT(St)↓
S3DIS [2]4-3T42.8717.7043.4017.8643.2718.2343.8717.7043.2218.96
7-3T36.1311.6036.4611.2036.8011.1036.7011.1036.0612.10
10-3T39.9514.3040.2513.7040.1014.0040.0714.0540.0514.10
ScanNetv2 [17]10-2T16.7510.5116.8510.5016.9010.0817.1310.3616.9110.76
10-5T27.0818.8727.1318.8727.2218.8527.2118.8527.1618.92
15-5T24.2525.5024.4525.1024.0525.9023.9526.1024.0026.00
", + "image_path": "a645460ce21a0458289ce725c7dd5b88b3af1fd7b8d375bc7afdf3010d3009a4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 205, + 305, + 285 + ], + "blocks": [ + { + "bbox": [ + 55, + 163, + 555, + 187 + ], + "lines": [ + { + "bbox": [ + 55, + 163, + 555, + 187 + ], + "spans": [ + { + "bbox": [ + 55, + 163, + 555, + 187 + ], + "type": "text", + "content": "Table 3. Different penalty values in Busemann loss in 5-shot incremental segmentation across the S3DIS [2] dataset and ScanNetV2 [17] dataset, highlighting the regularization benefits introduced by the hyperbolic distance of the projected class prototypes from the origin." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 205, + 305, + 285 + ], + "lines": [ + { + "bbox": [ + 58, + 205, + 305, + 285 + ], + "spans": [ + { + "bbox": [ + 58, + 205, + 305, + 285 + ], + "type": "table", + "html": "
DatasetsSettingsc = -1.00 (Ours)c = 0.0001c = 0.001
AIA↑FTLast↓AIA↑FTLast↓AIA↑FTLast↓
S3DIS [2]4-3T43.8717.7042.9519.9343.2317.38
7-3T36.8011.1036.2311.6036.8611.05
10-3T40.2513.7040.2013.8040.2213.75
ScanNetv2 [17]10-2T17.1310.3616.5410.2516.9810.18
10-5T27.2218.8527.1619.0026.8319.92
15-5T24.4525.1024.4025.2024.2025.60
", + "image_path": "b9320cd99bb4b9295491e34471e9f9043ed38955314e79bcbdee2fdfba05e1b9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 67, + 359, + 185, + 466 + ], + "blocks": [ + { + "bbox": [ + 67, + 359, + 185, + 466 + ], + "lines": [ + { + "bbox": [ + 67, + 359, + 185, + 466 + ], + "spans": [ + { + "bbox": [ + 67, + 359, + 185, + 466 + ], + "type": "image", + "image_path": "71fd6899a574728d21bd83abf494721a899036eef8cec22063294d65748d6cb4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 88, + 472, + 162, + 482 + ], + "lines": [ + { + "bbox": [ + 88, + 472, + 162, + 482 + ], + "spans": [ + { + "bbox": [ + 88, + 472, + 162, + 482 + ], + "type": "text", + "content": "(a) Last mIoU values" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 190, + 362, + 293, + 464 + ], + "blocks": [ + { + "bbox": [ + 190, + 362, + 293, + 464 + ], + "lines": [ + { + "bbox": [ + 190, + 362, + 293, + 464 + ], + "spans": [ + { + "bbox": [ + 190, + 362, + 293, + 464 + ], + "type": "image", + "image_path": "0797ba7c14ea84bc856ebcc361ddc407a88046c5c5edd3502ed5da5089f9942e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 203, + 469, + 275, + 483 + ], + "lines": [ + { + "bbox": [ + 203, + 469, + 275, + 483 + ], + "spans": [ + { + "bbox": [ + 203, + 469, + 275, + 483 + ], + "type": "text", + "content": "(b) Forgetting " + }, + { + "bbox": [ + 203, + 469, + 275, + 483 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\text{Last}}^{(T)}" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 491, + 296, + 546 + ], + "lines": [ + { + "bbox": [ + 55, + 491, + 296, + 546 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 296, + 546 + ], + "type": "text", + "content": "Figure 5. mIoU improves and forgetting reduces with more training examples. Our proposed method outperforms the top-2 baseline methods ([70, 79]) in terms of Last mIoU and forgetting mitigation in the 10-2T setting (Full-data: fully supervised training in incremental fashion)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 557, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 295, + 616 + ], + "type": "text", + "content": "cate that catastrophic forgetting is particularly severe with limited data; however, our method substantially reduces forgetting compared to other baselines. Notably, increasing the number of training examples significantly mitigates forgetting for H1Po, while it exacerbates forgetting for SDCOT." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": "Experiments with a different backbone: Besides PointTransformer [77], we evaluate our model with the backbone of DGCNN [65]. Unlike current SOTA FSCIL methods—BiDist [78], CLOM [85], TEEN [64], FACT [81], and ALICE [50]—which perform well with DGCNN but degrade with PointTransformer, our method remains effective across both architectures. As shown in Table 5, our approach achieves consistent gains with DGCNN, outper" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 361, + 205, + 509, + 305 + ], + "blocks": [ + { + "bbox": [ + 55, + 293, + 296, + 350 + ], + "lines": [ + { + "bbox": [ + 55, + 293, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 296, + 350 + ], + "type": "text", + "content": "Table 4. Comparison for different curvature values in 5-shot incremental segmentation across S3DIS [2] and ScanNetV2 [17] datasets. This highlights the superior stability-plasticity trade-off of hyperbolic embedding space with negative curvature, where small positive curvature approximates Euclidean space." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 361, + 205, + 509, + 305 + ], + "lines": [ + { + "bbox": [ + 361, + 205, + 509, + 305 + ], + "spans": [ + { + "bbox": [ + 361, + 205, + 509, + 305 + ], + "type": "table", + "html": "
MethodLast (↑)AIA (↑)FTLast (↓)
LwF [59]9.4526.7524.41
C-FSCIL [24]20.2230.6422.15
LGKD [72]18.6329.1615.75
SDCOT [79]22.7524.5419.94
CLOM [85]21.1834.6718.83
TEEN [64]22.7133.9015.37
FACT [81]24.2726.7021.26
3DPC-CISS [70]22.0324.7419.36
BiDist [78]27.7039.1415.85
HiPo29.9039.8613.87
", + "image_path": "6ed0464f457bed14881938bf17db917fc278df840a7c129cfc0e46e154ab5578.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 312, + 555, + 359 + ], + "lines": [ + { + "bbox": [ + 313, + 312, + 555, + 359 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 555, + 359 + ], + "type": "text", + "content": "Table 5. Performance comparison on the S3DIS dataset [2] using the DGCNN backbone [65] in the 5-shot setting. Reported metrics include Last, " + }, + { + "bbox": [ + 313, + 312, + 555, + 359 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{Last}^{(T)}" + }, + { + "bbox": [ + 313, + 312, + 555, + 359 + ], + "type": "text", + "content": ", and AIA for 4-3T and 7-3T settings. HiPo outperforms all baseline methods with DGCNN." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 376, + 555, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 555, + 415 + ], + "type": "text", + "content": "forming baselines with " + }, + { + "bbox": [ + 313, + 376, + 555, + 415 + ], + "type": "inline_equation", + "content": "+2.20\\%" + }, + { + "bbox": [ + 313, + 376, + 555, + 415 + ], + "type": "text", + "content": " mIoU on Last, " + }, + { + "bbox": [ + 313, + 376, + 555, + 415 + ], + "type": "inline_equation", + "content": "+0.72\\%" + }, + { + "bbox": [ + 313, + 376, + 555, + 415 + ], + "type": "text", + "content": " on AIA, and reducing forgetting by " + }, + { + "bbox": [ + 313, + 376, + 555, + 415 + ], + "type": "inline_equation", + "content": "+1.98\\%" + }, + { + "bbox": [ + 313, + 376, + 555, + 415 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 313, + 376, + 555, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{Last}^{(T)}" + }, + { + "bbox": [ + 313, + 376, + 555, + 415 + ], + "type": "text", + "content": ". This demonstrates the model-agnostic nature of HiPO." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 429, + 499, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 429, + 499, + 442 + ], + "spans": [ + { + "bbox": [ + 313, + 429, + 499, + 442 + ], + "type": "text", + "content": "5. Conclusions & Future Directions" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "text", + "content": "We introduce a previously unexplored problem setting along with a novel solution specifically designed for few-shot class-incremental semantic segmentation of 3D point cloud data, addressing a critical gap in the literature. Recognizing the unique challenges posed by point clouds, we identified limitations in current FSCIL methods and developed HiPo, a model that leverages hyperbolic space to embed point clouds while achieving an improved stability-plasticity trade-off. Our approach consistently outperforms state-of-the-art baselines, demonstrating that hyperbolic geometry's ability to represent hierarchical structures offers a more effective framework for FSCIL in 3D data. HiPo is particularly valuable for real-world applications like autonomous driving, industrial inspection, urban planning, and environmental monitoring, where generalization to unseen classes is essential. Furthermore, our approach enhances safety and efficiency in critical areas such as disaster response and infrastructure maintenance by enabling accurate, timely semantic segmentation of complex 3D environments. Future work will focus on managing possible outliers that may emerge without prior reference across incremental sessions, further strengthening our robustness." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11817" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 297, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 297, + 146 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 297, + 146 + ], + "type": "text", + "content": "[1] Noor Ahmed, Anna Kukleva, and Bernt Schiele. Orco: Towards better generalization via orthogonality and contrast for few-shot class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28762-28771, 2024. 1, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 148, + 296, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 148, + 296, + 203 + ], + "spans": [ + { + "bbox": [ + 61, + 148, + 296, + 203 + ], + "type": "text", + "content": "[2] Iro Armeni, Ozan Sener, Amir R Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3d semantic parsing of large-scale indoor spaces. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1534-1543, 2016. 2, 4, 6, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 205, + 295, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 205, + 295, + 258 + ], + "spans": [ + { + "bbox": [ + 62, + 205, + 295, + 258 + ], + "type": "text", + "content": "[3] Mina Ghadimi Atigh, Julian Schoep, Erman Acar, Nanne Van Noord, and Pascal Mettes. Hyperbolic image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4453-4462, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 260, + 296, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 260, + 296, + 304 + ], + "spans": [ + { + "bbox": [ + 62, + 260, + 296, + 304 + ], + "type": "text", + "content": "[4] Eden Belouadah and Adrian Popescu. II2m: Class incremental learning with dual memory. In Proceedings of the IEEE/CVF international conference on computer vision, pages 583-592, 2019. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 306, + 295, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 306, + 295, + 349 + ], + "spans": [ + { + "bbox": [ + 62, + 306, + 295, + 349 + ], + "type": "text", + "content": "[5] Abhijit Bendale and Terrance Boult. Towards open world recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1893-1902, 2015. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 351, + 295, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 351, + 295, + 417 + ], + "spans": [ + { + "bbox": [ + 62, + 351, + 295, + 417 + ], + "type": "text", + "content": "[6] Chris Burges, Tal Shaked, Erin Renshaw, Ari Lazier, Matt Deeds, Nicole Hamilton, and Greg Hullender. Learning to rank using gradient descent. In Proceedings of the 22nd International Conference on Machine Learning, page 89-96, New York, NY, USA, 2005. Association for Computing Machinery. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 418, + 294, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 418, + 294, + 440 + ], + "spans": [ + { + "bbox": [ + 62, + 418, + 294, + 440 + ], + "type": "text", + "content": "[7] Herbert Busemann. The geometry of geodesics. Courier Corporation, 2005. 2, 3, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 442, + 295, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 442, + 295, + 486 + ], + "spans": [ + { + "bbox": [ + 62, + 442, + 295, + 486 + ], + "type": "text", + "content": "[8] Francisco M Castro, Manuel J Marín-Jiménez, Nicolas Guil, Cordelia Schmid, and Karteek Alahari. End-to-end incremental learning. In Proceedings of the European conference on computer vision (ECCV), pages 233-248, 2018. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 487, + 295, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 487, + 295, + 542 + ], + "spans": [ + { + "bbox": [ + 62, + 487, + 295, + 542 + ], + "type": "text", + "content": "[9] Jun Cen, Peng Yun, Shiwei Zhang, Junhao Cai, Di Luan, Mingqian Tang, Ming Liu, and Michael Yu Wang. Openworld semantic segmentation for lidar point clouds. In European Conference on Computer Vision, pages 318-334. Springer, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 544, + 295, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 544, + 295, + 599 + ], + "spans": [ + { + "bbox": [ + 57, + 544, + 295, + 599 + ], + "type": "text", + "content": "[10] Fabio Cermelli, Massimiliano Mancini, Samuel Rota Bulo, Elisa Ricci, and Barbara Caputo. Modeling the background for incremental learning in semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9233-9242, 2020. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 600, + 295, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 600, + 295, + 655 + ], + "spans": [ + { + "bbox": [ + 57, + 600, + 295, + 655 + ], + "type": "text", + "content": "[11] Arslan Chaudhry, Puneet K Dokania, Thalaiyasingam Ajthan, and Philip HS Torr. Riemannian walk for incremental learning: Understanding forgetting and intransigence. In Proceedings of the European conference on computer vision (ECCV), pages 532-547, 2018. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 657, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 657, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 657, + 295, + 689 + ], + "type": "text", + "content": "[12] Bike Chen, Wei Peng, Xiaofeng Cao, and Juha Röning. Hyperbolic uncertainty aware semantic segmentation. IEEE Transactions on Intelligent Transportation Systems, 2023. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "text", + "content": "[13] Ali Cheraghian, Shafin Rahman, Pengfei Fang, Soumava Kumar Roy, Lars Petersson, and Mehrtash" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 117 + ], + "type": "text", + "content": "Harandi. Semantic-aware knowledge distillation for few-shot class-incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2534-2543, 2021. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 183 + ], + "type": "text", + "content": "[14] Townim Chowdhury, Mahira Jalisha, Ali Cheraghian, and Shafin Rahman. Learning without forgetting for 3d point cloud objects. In Advances in Computational Intelligence: 16th International Work-Conference on Artificial Neural Networks, IWANN 2021, Virtual Event, June 16-18, 2021, Proceedings, Part I 16, pages 484-497. Springer, 2021. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 184, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 184, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 317, + 184, + 553, + 239 + ], + "type": "text", + "content": "[15] Townim Chowdhury, Ali Cheraghian, Sameera Ramasinghe, Sahar Ahmadi, Morteza Saberi, and Shafin Rahman. Few-shot class-incremental learning for 3d point cloud objects. In European Conference on Computer Vision, pages 204-220. Springer, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 239, + 553, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 281 + ], + "type": "text", + "content": "[16] Yawen Cui, Zitong Yu, Wei Peng, Qi Tian, and Li Liu. Rethinking few-shot class-incremental learning with open-set hypothesis in hyperbolic geometry. IEEE Transactions on Multimedia, 2023. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 283, + 553, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 283, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 316, + 283, + 553, + 338 + ], + "type": "text", + "content": "[17] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 2, 6, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 338, + 553, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 553, + 370 + ], + "type": "text", + "content": "[18] David L. Davies and Donald W. Bouldin. A cluster separation measure. IEEE Transactions on Pattern Analysis and Machine Intelligence, PAMI-1(2):224-227, 1979. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 371, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 371, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 371, + 553, + 415 + ], + "type": "text", + "content": "[19] Jiahua Dong, Yang Cong, Gan Sun, Bingtao Ma, and Lichen Wang. I3dol: Incremental 3d object learning without catastrophic forgetting. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6066-6074, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 416, + 553, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 553, + 448 + ], + "type": "text", + "content": "[20] DC Dowson and BV666017 Landau. The frechet distance between multivariate normal distributions. Journal of multivariate analysis, 12(3):450-455, 1982. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 449, + 553, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 449, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 317, + 449, + 553, + 502 + ], + "type": "text", + "content": "[21] Luca Franco, Paolo Mandica, Konstantinos Kallidromitis, Devin Guillory, Yu-Teng Li, Trevor Darrell, and Fabio Galasso. Hyperbolic active learning for semantic segmentation under domain shift. arXiv preprint arXiv:2306.11180, 2023.5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 504, + 553, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 504, + 553, + 547 + ], + "spans": [ + { + "bbox": [ + 317, + 504, + 553, + 547 + ], + "type": "text", + "content": "[22] Octavian-Eugen Ganea, Gary Bécigneul, and Thomas Hofmann. Hyperbolic convolutional neural networks. In Advances in Neural Information Processing Systems, pages 4851-4861, 2018. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 548, + 553, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 548, + 553, + 591 + ], + "spans": [ + { + "bbox": [ + 317, + 548, + 553, + 591 + ], + "type": "text", + "content": "[23] Mina Ghadimi Atigh, Martin Keller-Ressel, and Pascal Mettes. Hyperbolic busemann learning with ideal prototypes. Advances in Neural Information Processing Systems, 34:103-115, 2021. 3, 5, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 592, + 553, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 592, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 317, + 592, + 553, + 646 + ], + "type": "text", + "content": "[24] Michael Hersche, Geethan Karunaratne, Giovanni Cherubini, Luca Benini, Abu Sebastian, and Abbas Rahimi. Constrained few-shot class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9057-9067, 2022. 1, 3, 6, 7, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "text", + "content": "[25] Naiwen Hu, Haozhe Cheng, Yifan Xie, Pengcheng Shi, and Jihua Zhu. Hyperbolic image-and-pointcloud contrastive learning for 3d classification. arXiv preprint arXiv:2409.15810, 2024. 2, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[26] Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, and Andrew Markham." + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11818" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "type": "text", + "content": "Randla-net: Efficient semantic segmentation of large-scale point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11108-11117, 2020. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 171 + ], + "type": "text", + "content": "[27] Valentin Khrulkov, Leyla Mirvakhabova, Evgeniya Ustinova, Ivan Oseledets, and Victor Lempitsky. Hyperbolic image embeddings. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6418-6428, 2020. 2, 3, 5, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 172, + 295, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 295, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 295, + 216 + ], + "type": "text", + "content": "[28] Do-Yeon Kim, Dong-Jun Han, Jun Seo, and Jaekyun Moon. Warping the space: Weight space rotation for class incremental few-shot learning. In The Eleventh International Conference on Learning Representations, 2023. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 217, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 217, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 57, + 217, + 294, + 239 + ], + "type": "text", + "content": "[29] Max Kochurov, Rasul Karimov, and Serge Kozlukov. Geoopt: Riemannian optimization in pytorch, 2020. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 239, + 296, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 296, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 296, + 293 + ], + "type": "text", + "content": "[30] Anna Kukleva, Hilde Kuehne, and Bernt Schiele. Generalized and incremental few-shot learning by explicit learning and calibration without forgetting. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9020-9029, 2021. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 294, + 295, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 295, + 348 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 295, + 348 + ], + "type": "text", + "content": "[31] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 350, + 294, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 350, + 294, + 381 + ], + "spans": [ + { + "bbox": [ + 57, + 350, + 294, + 381 + ], + "type": "text", + "content": "[32] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40(12):2935-2947, 2017. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 383, + 296, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 296, + 437 + ], + "type": "text", + "content": "[33] Fangzhou Lin, Yun Yue, Songlin Hou, Xuechu Yu, Yajun Xu, Kazunori D Yamada, and Ziming Zhang. Hyperbolic chamfer distance for point cloud completion. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14595-14606, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 437, + 295, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 295, + 491 + ], + "type": "text", + "content": "[34] Pei Lin, Zheng Cai, Jiangtao Zhang, and Lifang Ruan. A survey on point cloud segmentation: From classical methods to deep learning. ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM), 16(3): 1-27, 2020. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 492, + 295, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 492, + 295, + 535 + ], + "spans": [ + { + "bbox": [ + 56, + 492, + 295, + 535 + ], + "type": "text", + "content": "[35] Yuyang Liu, Yang Cong, Gan Sun, Tao Zhang, Jiahua Dong, and Hongsen Liu. L3doc: Lifelong 3d object classification. IEEE Transactions on Image Processing, 30:7486-7498, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 536, + 295, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 536, + 295, + 580 + ], + "spans": [ + { + "bbox": [ + 56, + 536, + 295, + 580 + ], + "type": "text", + "content": "[36] Zan Liu, Zilong Zhang, Yiyi Lin, Lin Zhu, Han Jin, and Stephen Lin. Pointaugment: An auto-augmentation framework for point cloud classification. arXiv preprint arXiv:2002.10876, 2020. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 582, + 294, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 294, + 603 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 294, + 603 + ], + "type": "text", + "content": "[37] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 604, + 295, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 604, + 295, + 647 + ], + "spans": [ + { + "bbox": [ + 56, + 604, + 295, + 647 + ], + "type": "text", + "content": "[38] Paolo Mandica, Luca Franco, Konstantinos Kallidromitis, Suzanne Petryk, and Fabio Galasso. Hyperbolic learning with multimodal large language models. arXiv preprint arXiv:2408.05097, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 647, + 295, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 295, + 680 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 295, + 680 + ], + "type": "text", + "content": "[39] Pratik Mazumder, Pravendra Singh, and Piyush Rai. Few-shot lifelong learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2337-2345, 2021. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "type": "text", + "content": "[40] Leland McInnes, John Healy, and James Melville. Umap: Uniform manifold approximation and projection for dimension reduction. arXiv preprint arXiv:1802.03426, 2018. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[41] Pascal Mettes, Elise Van der Pol, and Cees Snoek. Hyperspherical prototype networks. Advances in neural information processing systems, 32, 2019. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 108, + 553, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 553, + 152 + ], + "type": "text", + "content": "[42] Pascal Mettes, Mina Ghadimi Atigh, Martin Keller-Ressel, Jeffrey Gu, and Serena Yeung. Hyperbolic deep learning in computer vision: A survey. International Journal of Computer Vision, pages 1-25, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 153, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 153, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 153, + 553, + 196 + ], + "type": "text", + "content": "[43] Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. Distributed representations of words and phrases and their compositionality. Advances in neural information processing systems, 26, 2013. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 198, + 553, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 241 + ], + "type": "text", + "content": "[44] Gal Mishne, Zhengchao Wan, Yusu Wang, and Sheng Yang. The numerical stability of hyperbolic representation learning. In International Conference on Machine Learning, pages 24925-24949. PMLR, 2023. 2, 4, 5, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 243, + 553, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 553, + 296 + ], + "type": "text", + "content": "[45] Antonio Montanaro, Diego Valsesia, and Enrico Magli. Rethinking the compositionality of point clouds through regularization in the hyperbolic space. Advances in Neural Information Processing Systems, 35:33741-33753, 2022. 1, 2, 3, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 298, + 553, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 553, + 353 + ], + "type": "text", + "content": "[46] Antonio Montanaro, Diego Valsesia, and Enrico Magli. Towards hyperbolic regularizers for point cloud part segmentation. In ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5, 2023. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 355, + 555, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 355, + 555, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 355, + 555, + 409 + ], + "type": "text", + "content": "[47] Gabriel Moreira, Manuel Marques, João Paulo Costeira, and Alexander Hauptmann. Hyperbolic vs euclidean embeddings in few-shot learning: Two sides of the same coin. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2082-2090, 2024. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 411, + 553, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 477 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 477 + ], + "type": "text", + "content": "[48] Samrat Mukherjee, Tanuj Sur, Saurish Seksaria, Subhasis Chaudhuri, Gemma Roig, and Biplab Banerjee. Uidaple: Unsupervised incremental domain adaptation through adaptive prompt learning. In ICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5, 2025. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 478, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 553, + 521 + ], + "type": "text", + "content": "[49] Maximilian Nickel and Douwe Kiela. Poincaré embeddings for learning hierarchical representations. In Advances in neural information processing systems, pages 6338-6347, 2017. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 523, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 553, + 567 + ], + "type": "text", + "content": "[50] Can Peng, Kun Zhao, Tianren Wang, Meng Li, and Brian C Lovell. Few-shot class-incremental learning from an open-set perspective. In European Conference on Computer Vision, pages 382-397. Springer, 2022. 3, 6, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 568, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 553, + 633 + ], + "type": "text", + "content": "[51] Jeffrey Pennington, Richard Socher, and Christopher Manning. GloVe: Global vectors for word representation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1532-1543, Doha, Qatar, 2014. Association for Computational Linguistics. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "text", + "content": "[52] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[53] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11819" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "text", + "content": "point sets in a metric space. Advances in neural information processing systems, 30, 2017. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 162 + ], + "type": "text", + "content": "[54] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 196 + ], + "type": "text", + "content": "[55] Sachin Ravi and Hugo Larochelle. Optimization as a model for few-shot learning. arXiv preprint arXiv:1606.04474, 2017. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 294, + 251 + ], + "type": "text", + "content": "[56] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 2001–2010, 2017. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 294, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 294, + 286 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 294, + 286 + ], + "type": "text", + "content": "[57] Rik Sarkar. Low distortion delaunay embedding of trees in hyperbolic plane. In International symposium on graph drawing, pages 355-366. Springer, 2011. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 288, + 294, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 288, + 294, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 288, + 294, + 331 + ], + "type": "text", + "content": "[58] Yisheng Song, Ting Wang, Puyu Cai, Subrota K Mondal, and Jyoti Prakash Sahoo. A comprehensive survey of few-shot learning: Evolution, applications, challenges, and opportunities. ACM Computing Surveys, 55(13s):1-40, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 333, + 295, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 333, + 295, + 398 + ], + "spans": [ + { + "bbox": [ + 56, + 333, + 295, + 398 + ], + "type": "text", + "content": "[59] Filip Szatkowski, Mateusz Pyla, Marcin Przewieżlikowski, Sebastian Cygert, Bartlomiej Twardowski, and Tomasz Trzeciński. Adapt your teacher: Improving knowledge distillation for exemplar-free continual learning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1977-1987, 2024. 6, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 399, + 295, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 295, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 295, + 443 + ], + "type": "text", + "content": "[60] Yuwen Tan and Xiang Xiang. Cross-domain few-shot incremental learning for point-cloud recognition. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2307-2316, 2024. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 445, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 498 + ], + "type": "text", + "content": "[61] Xiaoyu Tao, Xiaopeng Hong, Xinyuan Chang, Songlin Dong, Xing Wei, and Yihong Gong. Few-shot class incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12183-12192, 2020. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 501, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 294, + 533 + ], + "type": "text", + "content": "[62] Jesper E Van Engelen and Holger H Hoos. A survey on semi-supervised learning. Machine learning, 109(2):373-440, 2020. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 535, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 294, + 578 + ], + "type": "text", + "content": "[63] Max van Spengler, Erwin Berkhout, and Pascal Mettes. Poincare resnet. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5419-5428, 2023. 4, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 580, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 580, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 580, + 294, + 624 + ], + "type": "text", + "content": "[64] Qi-Wei Wang, Da-Wei Zhou, Yi-Kai Zhang, De-Chuan Zhan, and Han-Jia Ye. Few-shot class-incremental learning via training-free prototype calibration. Advances in Neural Information Processing Systems, 36, 2024. 6, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 625, + 294, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 294, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 294, + 668 + ], + "type": "text", + "content": "[65] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (tog), 38(5):1-12, 2019. 1, 2, 6, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[66] Yaqing Wang, Quanming Yao, James T Kwok, and Lionel M Ni. Generalizing from a few examples: A survey on few-shot learning. ACM computing surveys (csur), 53(3):1-34, 2020. 1" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "text", + "content": "[67] Wenxuan Wu, Zhongang Qi, and Li Fuxin. Pointconv: Deep convolutional networks on 3d point clouds. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 9621-9630, 2019. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "type": "text", + "content": "[68] Yating Xu, Conghui Hu, Na Zhao, and Gim Hee Lee. Generalized few-shot point cloud segmentation via geometric words. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21506-21515, 2023. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 162, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 162, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 162, + 553, + 205 + ], + "type": "text", + "content": "[69] Xiangli Yang, Zixing Song, Irwin King, and Zenglin Xu. A survey on deep semi-supervised learning. IEEE Transactions on Knowledge and Data Engineering, 35(9):8934-8954, 2022. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 206, + 555, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 555, + 260 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 555, + 260 + ], + "type": "text", + "content": "[70] Yuwei Yang, Munawar Hayat, Zhao Jin, Chao Ren, and Yinjie Lei. Geometry and uncertainty-aware 3d point cloud class-incremental semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21759-21768, 2023. 6, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 262, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 262, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 262, + 553, + 304 + ], + "type": "text", + "content": "[71] Zhilu Yang, Ying Zheng, and Jun Zhang. Learning to learn from labeled and unlabeled data with point-wise mutual information. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5704-5711, 2019. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 305, + 553, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 305, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 316, + 305, + 553, + 371 + ], + "type": "text", + "content": "[72] Ze Yang, Ruibo Li, Evan Ling, Chi Zhang, Yiming Wang, Dezhao Huang, Keng Teck Ma, Minhoe Hur, and Guosheng Lin. Label-guided knowledge distillation for continual semantic segmentation on 2d images and 3d point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18601-18612, 2023. 1, 3, 6, 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 373, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 373, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 373, + 553, + 415 + ], + "type": "text", + "content": "[73] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. arXiv preprint arXiv:2111.07783, 2021. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 416, + 553, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 553, + 469 + ], + "type": "text", + "content": "[74] Yun Yue, Fangzhou Lin, Guanyi Mou, and Ziming Zhang. Understanding hyperbolic metric learning through hard negative sampling. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1891-1903, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 472, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 472, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 472, + 553, + 525 + ], + "type": "text", + "content": "[75] Maciej Zamorski, Michal Stypulkowski, Konrad Karanowski, Tomasz Trzeciński, and Maciej Zieba. Continual learning on 3d point clouds with random compressed rehearsal. Computer Vision and Image Understanding, 228: 103621, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 526, + 553, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 526, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 316, + 526, + 553, + 581 + ], + "type": "text", + "content": "[76] Chi Zhang, Nan Song, Guosheng Lin, Yun Zheng, Pan Pan, and Yinghui Xu. Few-shot incremental learning with continually evolved classifiers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12455-12464, 2021. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 582, + 555, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 555, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 555, + 624 + ], + "type": "text", + "content": "[77] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16259-16268, 2021. 2, 6, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 625, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 689 + ], + "type": "text", + "content": "[78] Linglan Zhao, Jing Lu, Yunlu Xu, Zhanzhan Cheng, Dashan Guo, Yi Niu, and Xiangzhong Fang. Few-shot class incremental learning via class-aware bilateral distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11838-11847, 2023. 1, 3, 6, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 692, + 555, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 555, + 714 + ], + "type": "text", + "content": "[79] Na Zhao and Gim Hee Lee. Static-dynamic co-teaching for class-incremental 3d object detection. In Proceedings of" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11820" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 373 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "text", + "content": "the AAAI Conference on Artificial Intelligence, pages 3436-3445, 2022. 1, 3, 6, 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "type": "text", + "content": "[80] Na Zhao, Tat-Seng Chua, and Gim Hee Lee. Few-shot 3d point cloud semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8873-8882, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 295, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 295, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 295, + 196 + ], + "type": "text", + "content": "[81] Da-Wei Zhou, Fu-Yun Wang, Han-Jia Ye, Liang Ma, Shiliang Pu, and De-Chuan Zhan. Forward compatible few-shot class-incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9046–9056, 2022. 1, 2, 3, 6, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 295, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 295, + 240 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 295, + 240 + ], + "type": "text", + "content": "[82] Da-Wei Zhou, Zi-Wen Cai, Han-Jia Ye, De-Chuan Zhan, and Ziwei Liu. Revisiting class-incremental learning with pretrained models: Generalizability and adaptivity are all you need. arXiv preprint arXiv:2303.07338, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 241, + 295, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 295, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 295, + 285 + ], + "type": "text", + "content": "[83] Da-Wei Zhou, Qi-Wei Wang, Zhi-Hong Qi, Han-Jia Ye, De-Chuan Zhan, and Ziwei Liu. Class incremental learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 286, + 295, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 295, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 295, + 319 + ], + "type": "text", + "content": "[84] Yudong Zhu, Di Zhou, Jinghui Xiao, Xin Jiang, Xiao Chen, and Qun Liu. Hypertext: Endowing fasttext with hyperbolic geometry. arXiv preprint arXiv:2010.16143, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 320, + 295, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 295, + 373 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 295, + 373 + ], + "type": "text", + "content": "[85] Yixiong Zou, Shanghang Zhang, Yuhua Li, and Ruixuan Li. Margin-based few-shot class-incremental learning with class-level overfitting mitigation. Advances in neural information processing systems, 35:27267-27279, 2022. 1, 3, 6, 8" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "11821" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Hyperdimensional Uncertainty Quantification for Multimodal Uncertainty Fusion in Autonomous Vehicles Perception/8d8555dd-a579-45c4-ad75-fee4bbfad5af_content_list.json b/2025/Hyperdimensional Uncertainty Quantification for Multimodal Uncertainty Fusion in Autonomous Vehicles Perception/8d8555dd-a579-45c4-ad75-fee4bbfad5af_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ae7aa3283dd1c22a3d5e70ca40d9321b390a2523 --- /dev/null +++ b/2025/Hyperdimensional Uncertainty Quantification for Multimodal Uncertainty Fusion in Autonomous Vehicles Perception/8d8555dd-a579-45c4-ad75-fee4bbfad5af_content_list.json @@ -0,0 +1,1628 @@ +[ + { + "type": "text", + "text": "Hyperdimensional Uncertainty Quantification for Multimodal Uncertainty Fusion in Autonomous Vehicles Perception", + "text_level": 1, + "bbox": [ + 91, + 128, + 906, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Luke Chen* Junyao Wang Trier Mortlock Pramod Khargonekar Mohammad Abdullah Al Faruque University of California, Irvine", + "bbox": [ + 99, + 203, + 903, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{panwangc, junyaow4, tmortloc, pramod.khargonekar, alfaruqu}@uci.edu", + "bbox": [ + 197, + 241, + 799, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Uncertainty Quantification (UQ) is crucial for ensuring the reliability of machine learning models deployed in real-world autonomous systems. However, existing approaches typically quantify task-level output prediction uncertainty without considering epistemic uncertainty at the multimodal feature fusion level, leading to sub-optimal outcomes. Additionally, popular uncertainty quantification methods, e.g., Bayesian approximations, remain challenging to deploy in practice due to high computational costs in training and inference. In this paper, we propose HyperDUM, a novel deterministic uncertainty method (DUM) that efficiently quantifies feature-level epistemic uncertainty by leveraging hyperdimensional computing. Our method captures the channel and spatial uncertainties through channel and patch-wise projection and bundling techniques respectively. Multimodal sensor features are then adaptively weighted to mitigate uncertainty propagation and improve feature fusion. Our evaluations show that HyperDUM on average outperforms the state-of-the-art (SOTA) algorithms by up to $2.01\\% / 1.27\\%$ in 3D Object Detection and up to $1.29\\%$ improvement over baselines in semantic segmentation tasks under various types of uncertainties. Notably, HyperDUM requires $2.36 \\times$ less Floating Point Operations and up to $38.30 \\times$ less parameters than SOTA methods, providing an efficient solution for real-world autonomous systems.", + "bbox": [ + 88, + 323, + 485, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 731, + 220, + 748 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Effectively quantifying and addressing uncertainty plays a critical role in ensuring the reliability of ML models in real-world applications, especially where the decisions are consequential, e.g., autonomous driving. Models failing to address uncertainty can produce overconfident outputs, leading to compromised performance [9, 16, 41, 53]. In light of autonomous systems, uncertainties can arise from adverse", + "bbox": [ + 89, + 757, + 483, + 863 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "weather conditions, sensor failures or noises, environment dynamics (movements of other actors, road conditions, vehicle dynamics), and various corner cases [4, 12, 35, 36]. Modern autonomous systems leverage multimodal sensor fusion to broaden the information scope and learn comprehensive representations of the surrounding environment [8, 31, 56]. However, commonly employed sensors such as cameras, lidar, and radar, exhibit varying degrees of robustness and vulnerability across different driving scenarios [51, 54]. For instance, cameras are crucial in delivering rich perceptual information while they are particularly vulnerable to lighting variations and occlusions [60]. Lidar and radar are more robust to visual perturbations but can be affected by multi-path interference and occlusion [2, 28]. Therefore, it is important to consider the uncertainty of each modality before the fusion of multimodal features to achieve optimal results [15].", + "bbox": [ + 511, + 292, + 908, + 532 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Popular Uncertainty Quantification (UQ) methods include deep ensembles, Bayesian neural networks, and Bayesian approximation methods such as Monte Carlo dropout [11, 23, 24]. However, these methods often introduce large training and/or inference overheads compounded by the growing complexity of modern Deep Neural Networks (DNN), ultimately challenging their practicality under real-time computing constraints [7, 22]. Other lines of works such as Conformal Prediction (CP) and Posterior Networks (PostNets) based on evidential deep learning either cannot quantify feature-level epistemic uncertainty or require vast amounts of quality data to accurately estimate the posteriors [1, 5]. Deterministic Uncertainty Methods (DUMs), aiming to efficiently quantify uncertainty using only a single forward pass [10, 34, 42], have become an emerging trend. DUMs rely on statistical/morphological characteristics of latent features to quantify epistemic uncertainty. For instance, [10] employed Prototype Learning (PL) techniques to capture the \"knowledge\" of DNNs by constructing a \"memory bank\" of prototypes to predict uncertainty.", + "bbox": [ + 511, + 536, + 908, + 838 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address these problems, we propose HyperDUM, a novel deterministic uncertainty method for efficient quantification of feature-level epistemic uncertainty. In contrast to existing methods, HyperDUM projects latent features into", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Author. Distribution A: Approved for public release; distribution unlimited. OPSEC # 9537", + "bbox": [ + 89, + 875, + 482, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22306", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "hyperdimensional space to form hyperdimensional prototypes, and then estimates uncertainty based on the distance between features of unseen samples and these prototypes.", + "bbox": [ + 89, + 90, + 485, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To the best of our knowledge, HyperDUM is the first deterministic uncertainty quantification method utilizing hyperdimensional feature prototyping with Channel-wise and Patch-wise projection and bundling to deliver accurate and efficient epistemic uncertainty quantification of latent features. Detailed experimental evaluations on the DeLiVER and aiMotive datasets for semantic segmentation and object detection tasks demonstrate that HyperDUM can efficiently and effectively quantify feature uncertainty and improve fusion performance when compared to standard Bayesian approximation inference baselines and SOTA methods.", + "bbox": [ + 89, + 137, + 485, + 305 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 89, + 320, + 241, + 335 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Uncertainty Quantification", + "text_level": 1, + "bbox": [ + 89, + 345, + 334, + 363 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Predictive uncertainty is composed of three factors, model/epistemic uncertainty, data/aleatoric uncertainty, and distributional uncertainty [32]. Epistemic uncertainty, stemming from the word epistemology which is the study of knowledge, is associated with a model's lack of knowledge of underrepresented data. The scope of this work is to efficiently quantify epistemic uncertainty at the latent feature level to enable uncertainty-weighting prior to feature fusion.", + "bbox": [ + 89, + 369, + 485, + 492 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing UQ approaches are typically derived from Bayesian principles, including Bayesian neural networks which represent model weights as probability distributions, deep ensembles which sample output distributions from diverse models [24], and Monte Carlo dropout which is similar to deep ensembles but utilizes a single model with dropout layers [11]. However, these approaches remain challenging in practice due to enormous training and inference costs [7, 22]. Conformal prediction (CP) is a model-agnostic UQ technique that provides prediction intervals with guaranteed coverage, ensuring the true value falls within the interval at a specified probability [1]. However, while effective in producing reliable task-level uncertainties, CP does not capture uncertainties at intermediate model abstractions, a limitation especially evident in multimodal fusion models, where uncertainty propagates across different modalities before the final output. Evidential deep learning is another line of work, i.e. Posterior Networks which aim to quantify both aleatoric and epistemic uncertainty by directly estimating a posterior distribution over the model's outputs/predictions but is sensitive to data quality for accurately estimating the posteriors [5]. Deterministic uncertainty quantification is a growing research area that aims to efficiently compute epistemic uncertainty of DNNs with a single forward pass [10, 34, 42]. For instance, [34] is a training-free method that injects Gaussian noise into intermediate layers to compute the variance across the representations as a measure of uncertainty. La", + "bbox": [ + 89, + 493, + 485, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tent Deterministic Uncertainty (LDU) constructs prototypes from latent features by enforcing dissimilarity between prototypes and entropy between latent features while correlating uncertainty and the downstream task [10, 36, 46]. Recent research [37] demonstrated that constructing hyperdimensional prototypes can provide comparable performance to existing aleatoric uncertainty quantification solutions, e.g., Bayesian approximation, and offer significant speedups in training and inference. However, it is applied to regression tasks for aleatoric uncertainty. In contrast, our work quantifies epistemic uncertainty on the latent features, extending its applicability to multimodal fusion models and complex tasks such as object detection and semantic segmentation.", + "bbox": [ + 511, + 90, + 908, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Multimodal Uncertainty-Aware Fusion", + "text_level": 1, + "bbox": [ + 511, + 297, + 848, + 313 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multimodal sensor fusion aims to provide more comprehensive representations and address the limitations of unimodal sensor failure cases [8, 31, 56]. Existing approaches dealing with uncertainties fall into two categories: feature-level fusion and output-level fusion. Feature-level fusion, also known as intermediate fusion, applies an uncertainty weighting that scales the features to minimize uncertainty [30, 47]. For instance, [14] computes feature-informativeness and scales the original features based on informativeness. On the other hand, output-level fusion, also known as late-fusion, uses uncertainty to calibrate the outputs of the model, i.e. calibrating softmax probabilities [40]. In contrast to existing uncertainty-based fusion works that commonly rely on Bayesian approximated uncertainty fusion, our proposed HyperDUM is the first to investigate deterministic uncertainty fusion methods in a multimodal autonomous systems setting.", + "bbox": [ + 511, + 320, + 908, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 511, + 575, + 648, + 593 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 511, + 599, + 653, + 616 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given an input sample $\\mathbf{x} = \\{x_{1}, x_{2}, \\dots, x_{M}\\}$ consisting of $M$ multimodal sensor inputs. We have a multimodal model $\\mathbb{M}$ with pre-trained feature encoders $\\mathbf{f} = \\{f_{1}, \\dots, f_{M}\\}$ that extract modality-specific abstract features $z_{m} = f_{m}(x_{m})$ where $1 \\leq m \\leq M$ . Additionally, the model contains fusion blocks $\\mathcal{F}(z_{i}, z_{j})$ where $i, j \\in \\{1, \\dots, M\\}$ such that $i \\neq j$ which combine modality-specific features to learn cross-modal fusion features $g$ . Assuming there are $K$ fusion blocks, the fused features $\\mathbf{g} = \\{g_{1}, \\dots, g_{K}\\}$ are passed to the task head $h(g)$ to output the corresponding task predictions $y$ (i.e. bounding boxes and object classification for object detection, and pixel-wise class labels for semantic segmentation). To account for feature uncertainty, an uncertainty quantification module $\\mathbb{U}_{m}(z_{m})$ is inserted immediately following each feature extractor to predict an uncertainty value $u_{m}$ . Given the uncertainty value, a learnable", + "bbox": [ + 511, + 623, + 908, + 866 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1Cross-modal feature alignment is not required but may be enforced by the backbone.", + "bbox": [ + 511, + 875, + 908, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22307", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d5504aea6ccdf5201820be0635a2f39dd6045f3706781f2363bc9e893a3a7f54.jpg", + "image_caption": [ + "Figure 1. Multimodal model with uncertainty quantification and uncertainty fusion for autonomous driving perception tasks." + ], + "image_footnote": [], + "bbox": [ + 94, + 94, + 897, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "uncertainty weighting module $\\Omega (z_m,u_m)$ reweights the input features to produce an uncertainty-aware feature $\\hat{z}_m$ for each modality. Thus, the input to the fusion blocks now becomes $\\mathcal{F}(\\hat{z}_i,\\hat{z}_j)$ to produce uncertainty-weighted fusion features $\\hat{\\mathbf{g}}$ which are forwarded to be fused and subsequently used by the task head $h(\\hat{\\mathbf{g}})$ to produce the outputs. The system flow of the described multimodal fusion model with uncertainty quantification is illustrated in Figure 1.", + "bbox": [ + 89, + 357, + 483, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.1 Vector Symbolic Architectures/Hyperdimensional Computing", + "text_level": 1, + "bbox": [ + 89, + 498, + 483, + 529 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Vector Symbolic Architectures (VSA) / Hyperdimensional Computing (HDC) are a class of algorithms for creating and handling hyperdimensional representations to mimic cognitive functions [25, 45]. VSA starts with mapping input data to hyperdimensional vectors, or hypervectors $\\mathcal{H}$ , through a projection function $\\phi: \\mathcal{Z} \\to \\mathbf{H}$ , where $\\mathcal{Z} \\in \\mathbb{R}^n$ and $\\mathbf{H} \\in \\mathbb{R}^d$ ( $d \\gg n$ ) is the hypervector space. The projection function depends on the input data representation. This work assumes Euclidean input space. For non-euclidean see [38]. Following the projection, common operations on hypervectors include similarity calculation, bundling, and binding. Similarity $(\\delta(\\cdot, \\cdot))$ calculation measures the distance between two hypervectors. For real-valued hypervectors, a common measure is cosine similarity. Bundling $(\\bigoplus)$ is an element-wise addition of hypervectors, e.g., $\\mathcal{H}^{bundle} = \\mathcal{H}^1 \\bigoplus \\mathcal{H}^2$ , generating a hypervector with the same dimension as inputs. Binding $(\\otimes)$ is an element-wise multiplication associating two hypervectors to create another near-orthogonal hypervector, e.g., $\\mathcal{H}^{bind} = \\mathcal{H}^1 \\bigotimes \\mathcal{H}^2$ where $\\delta(\\mathcal{H}^{bind}, \\mathcal{H}^1) \\approx 0$ and $\\delta(\\mathcal{H}^{bind}, \\mathcal{H}^2) \\approx 0$ .", + "bbox": [ + 91, + 537, + 483, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our work mainly leverages the bundling operation of VSA for the formation of prototypes. In high-dimensional space, bundling mimics how human brains memorize information [48, 57]. For instance, for $\\mathcal{H}^{bundle} = \\mathcal{H}^1 +\\mathcal{H}^2$", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "we have $\\delta (\\mathcal{H}^{bundle},\\mathcal{H}^1)\\gg 0$ while $\\delta (\\mathcal{H}^{bundle},\\mathcal{H}^3)\\approx 0$ $(\\mathcal{H}^3\\neq \\mathcal{H}^1,\\mathcal{H}^2)$ . In other words, VSA can generate prototypical representations by bundling the high-dimensional vectors of individual data samples, making it especially useful for prototype learning. Specifically, suppose we have labeled samples $\\mathcal{D} = \\{(z_1^1,y_1^1),(z_1^2,y_1^2),\\ldots ,(z_M^N,y_M^N)\\}$ where $z_{m}^{i}\\in \\mathcal{Z}$ and $y_{m}^{i}\\in \\{l\\}_{l = 1}^{L}$ , we can form $L$ prototypes by bundling hypervectors corresponding to a particular label:", + "bbox": [ + 511, + 354, + 906, + 491 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {H} _ {m} ^ {l} = \\bigoplus_ {i: y ^ {i} = l} \\phi_ {m} \\left(z _ {m} ^ {i}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 635, + 489, + 906, + 523 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where each hyperdimensional prototype can represent an aspect of the data samples $\\mathcal{D}$ depending on the label definition, i.e. class label (cat and dog) or context label (weather conditions which we use in this work). We then define the notion of uncertainty as the set of similarity between hypervectors and hyperdimensional prototype of modality $m$ as follows:", + "bbox": [ + 511, + 526, + 908, + 618 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {U} _ {m} = \\bigcup_ {l = 1} ^ {L} \\left\\{\\delta \\left(\\mathcal {H} _ {m} ^ {z}, \\mathcal {H} _ {m} ^ {l}\\right) \\right\\}, \\text {w h e r e} | \\mathbb {U} _ {m} | = L \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 565, + 637, + 906, + 678 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbb{U}_m$ is the set of similarity distances between a hypervector $\\mathcal{H}_m^z$ and each hyperdimensional prototype $\\mathcal{H}_m^l$ , $l = 1,2,\\dots ,L$ and used to express the similarity uncertainty. The notion of similarity for uncertainty has been used in other prototype learning methods [10, 26], where the former assigns the variations of similarities between an instance and the prototypes as the belief masses and the latter utilizes similarity to optimize the uncertainty loss.", + "bbox": [ + 511, + 685, + 908, + 806 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. HyperDUM", + "text_level": 1, + "bbox": [ + 511, + 814, + 629, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Following the conventions of VSA for projection and bundling, we propose two modifications for uncertainty quan-", + "bbox": [ + 511, + 837, + 906, + 868 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "2We derive uncertainty quantifiability from similarity/expressivity of VSAs/HDCs based on recent theorems in the Appendix.", + "bbox": [ + 511, + 875, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22308", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tification: 1) Channel-wise and 2) Patch-wise Projection & Bundling (CPB/PPB). CPB modifies the projection and bundling operations to capture the uncertainty of each channel per modality separately, whereas PPB allows the VSA to quantify the spatial uncertainty without losing fine-grained spatial information through patching. Figure 1 visualizes the CPB and PPB operations of the proposed method.", + "bbox": [ + 89, + 90, + 483, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1 Channel-wise Projection & Bundling (CPB)", + "text_level": 1, + "bbox": [ + 89, + 214, + 449, + 229 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Latent features are generally multi-channeled. By convention prototype formation based on Equation 1 projects and bundles all feature dimensions onto a common representation space $\\mathbf{H}$ [52, 55]. Given latent feature $z_{m}^{i} \\in \\mathbb{R}^{C \\times H \\times W}$ with $C$ denoting feature channels and $H, W$ the spatial dimensions, adaptive pooling (max or average) is first used to reduce the spatial dimensions to $z_{m}^{pooled} \\in \\mathbb{R}^{C}$ . Through matrix multiplication and proper dimension initialization of the projection matrix $\\Phi$ , the correct hypervector dimensions can be obtained for $\\mathcal{H}_{m}^{z}$ as follows:", + "bbox": [ + 89, + 237, + 483, + 388 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\phi_ {m} \\left(z _ {m} ^ {i}\\right) = \\Phi^ {(d \\times C)} \\cdot z _ {m} ^ {\\text {p o o l e d} (C)} = \\mathcal {H} _ {m} ^ {z} ^ {(d)} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 393, + 483, + 414 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Another method to handle multi-channels inputs is by projecting each feature dimension separately followed by binding and subsequent bundling to a common space [48]. However, we note that the above operations neglect the fact that different feature channels learn different aspects of the input and may contribute more or less to the overall uncertainty. For example, [27] showed that sharing the same uncertain distribution among different channels is less effective for out-of-distribution generalization, while considering each channel uncertainty separately brings better performances due to the different channel potentials. Therefore, instead of absorbing the channel dimensions during projection and bundling, we retain them as follows:", + "bbox": [ + 89, + 420, + 483, + 617 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\phi_ {m} \\left(z _ {m} ^ {i}\\right) = \\Phi^ {(d \\times C)} \\otimes z _ {m} ^ {\\text {p o o l e d} (C)} = \\mathcal {H} _ {m} ^ {z} ^ {(d \\times C)} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 622, + 483, + 642 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where we use the Einstein summation notation $\\otimes$ (not the binding $\\otimes$ notation) to retain the channel dimension. Consequently, the bundling operation is modified to account for the channel dimensions:", + "bbox": [ + 89, + 647, + 483, + 705 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {H} _ {m} ^ {l} = \\bigoplus_ {i: y ^ {i} = l} ^ {d i m = C} \\phi_ {m} \\left(z _ {m} ^ {i}\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 209, + 710, + 483, + 755 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $dim = C$ indicates the dimension over which the element-wise addition is performed and $\\mathcal{H}_m^l$ is now $\\mathbb{R}^{d\\times C}$ . Additionally, the similarity uncertainty now outputs similarities for each channel dimension by representing Equation 2:", + "bbox": [ + 89, + 760, + 486, + 821 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {U} _ {m} ^ {C} = \\bigcup_ {c = 1} ^ {C} \\left\\{\\bigcup_ {l = 1} ^ {L} \\left\\{\\delta \\left(\\mathcal {H} _ {m} ^ {z} [ c ], \\mathcal {H} _ {m} ^ {l} [ c ]\\right) \\right\\} \\right\\}, \\text {s . t .} | \\mathbb {U} _ {m} ^ {C} | = C \\times L \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 838, + 483, + 896 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We empirically demonstrate through experiments that channel projection and bundling outperforms conventional projection and bundling for the task of multimodal uncertainty-aware feature fusion.", + "bbox": [ + 511, + 90, + 906, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2 Patch-wise Projection & Bundling (PPB)", + "text_level": 1, + "bbox": [ + 511, + 169, + 852, + 184 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Channel projection and bundling enables our method to capture the holistic uncertainties of each feature channel. However, to capture the finer-granularity uncertainties of the spatial dimensions requires a different approach. By convention spatial dimensions are either pooled to a single value and projected to high dimensions as shown by Equation 3 or each spatial dimension is separately projected to high dimensions. The former loses spatial information due to pooling and the latter is computationally infeasible and memory intensive given the large spatial dimensions of deep neural networks.", + "bbox": [ + 511, + 191, + 906, + 344 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inspired by the idea of using image patches to capture spatial information in anomaly detection and image classification [6, 44], we propose PPB which projects and bundles spatial patches of latent features to hyperdimensional space to capture the spatial uncertainties. Specifically, given latent feature $z_{m}^{i}\\in \\mathbb{R}^{C\\times H\\times W}$ , we slice $(H,W)$ into $P$ patches of resolution $(\\tilde{h},\\tilde{w})$ where $\\tilde{h} = H / \\sqrt{(P)}$ , $\\tilde{w} = W / \\sqrt{(P)}$ . This gives us the set of patched features $\\tilde{z}_{m}^{i} = \\{\\tilde{z}_{m}^{i,j}\\}_{j = 1}^{P}$ where each $\\tilde{z}_{m}^{i,j}\\in \\mathbb{R}^{C\\times \\tilde{h}\\times \\tilde{w}}$ are pooled to $\\mathbb{R}^C$ and form the set of pooled patched features $\\tilde{z}_{m}^{pooled}\\in \\mathbb{R}^{C\\times P}$ .", + "bbox": [ + 511, + 345, + 908, + 498 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given patched features $\\tilde{z}_m^{pooled}$ , we project them according to Equation 3 to form the set of patched projections:", + "bbox": [ + 511, + 500, + 906, + 530 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\phi_ {m} \\left(\\tilde {z} _ {m} ^ {i}\\right) = \\Phi^ {(d \\times C)} \\cdot \\tilde {z} _ {m} ^ {\\text {p o o l e d} (C \\times P)} = \\mathcal {H} _ {m} ^ {\\tilde {z}} ^ {(d \\times P)} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 550, + 906, + 571 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Consequently, these projections are bundled for each patch according to Equation 1 to produce the set of patched prototypes as follows:", + "bbox": [ + 511, + 582, + 906, + 627 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {H} _ {\\tilde {m}} ^ {l} = \\bigoplus_ {i: y ^ {i} = l} ^ {d i m = P} \\phi_ {m} \\left(\\tilde {z} _ {m} ^ {i}\\right) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 632, + 638, + 906, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{H}_{\\tilde{m}}^{l}$ represents the patched prototypes of modality $m$ , for each patch and for each context label $i$ , making $|\\mathcal{H}_{\\tilde{m}}^{l}| = d\\times P$", + "bbox": [ + 511, + 688, + 906, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, we can obtain the spatial uncertainties for each patch of the input projection $\\mathcal{H}_m^{\\tilde{z}}$ by modifying Equation 2:", + "bbox": [ + 511, + 734, + 906, + 766 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {U} _ {\\tilde {m}} ^ {P} = \\bigcup_ {p = 1} ^ {P} \\left\\{\\bigcup_ {l = 1} ^ {L} \\left\\{\\delta \\left(\\mathcal {H} _ {\\tilde {m}} ^ {\\tilde {z}} [ p ], \\mathcal {H} _ {\\tilde {m}} ^ {l} [ p ]\\right) \\right\\} \\right\\}, \\text {s . t .} | \\mathbb {U} _ {\\tilde {m}} ^ {P} | = P \\times L \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 786, + 906, + 844 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our final proposed approach combines both channel and patch methods for projection and bundling to capture channel-wise feature uncertainties, as well as finer", + "bbox": [ + 511, + 854, + 908, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22309", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "granularity spatial feature uncertainties, to provide a holistic epistemic uncertainty quantification of each modality.", + "bbox": [ + 89, + 90, + 482, + 122 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 135, + 222, + 152 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 160, + 279, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1.1 Datasets", + "text_level": 1, + "bbox": [ + 89, + 184, + 202, + 198 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate the effectiveness of HyperDUM for uncertainty-aware feature fusion using two multimodal autonomous driving datasets. The aiMotive dataset [33] includes camera, lidar, and radar sensors for long-distance $(\\geq 75m)$ 3D object detection and features diverse scenes across various locations, times, and weather conditions. The DeLiVER dataset [58] contains camera, lidar, event, and depth sensors for semantic segmentation, including adverse corner cases such as motion blur and lidar jitter. Additionally, we integrate corner cases into the aiMotive dataset by applying various lighting effects to camera sensors, causing reduced clarity and scene obstruction. These effects are created through image transformations such as GaussianBlur and exposure adjustments [39, 50]. We also implement foggification [3] to simulate synthetic fogging of lidar data, resulting in noisy point clouds with reduced field of view. These corner case effects can be visualized in Figure 2. Details in the appendix.", + "bbox": [ + 89, + 208, + 485, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1.2 Models", + "text_level": 1, + "bbox": [ + 89, + 484, + 194, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We utilize the pretrained BEVFusion [33] and CMNeXT [58] models as the fusion architectures for aiMotive and DeLiVER, respectively. BEVFusion is built upon VoxelNet [59] and BEVDepth [29] to fuse camera and lidar representations on a unified BEV space. CMNeXT is a four-stage pyramidal fusion model that treats the RGB representation as the primary branch and other modalities as the secondary branch. We integrate the uncertainty quantification module (UQM) at each model's respective pre-fusion points. For aiMotive this is chosen to be at the feature concatenation step, right before the bev Fuse step. For DeLiVER, we insert the UQM immediately after the feature rectification module and right before the feature fusion module. Given that the CMNeXT model has four stages, we instantiate one UQM per stage. Details in the Appendix.", + "bbox": [ + 89, + 508, + 483, + 736 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1.3 Baseline Methods", + "text_level": 1, + "bbox": [ + 89, + 755, + 266, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We implemented different UQ methods and compared their performances against our method. For each implementation, we followed their open-sourced code implementation and hyperparameter settings. Specifically, for Infer-dropout (InfMCD) and infer-noise (InfNoise), we followed their suggestions and chose 10 forward passes and used the variance of the outputs for uncertainty estimation. In addition to variance, we computed predictive entropy, mutual information,", + "bbox": [ + 89, + 779, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and entropy as additional uncertainty estimates following [49]. It was shown that using more uncertainty metrics can provide more effective quantification for different types of uncertainties. For Posterior Network (PostNet), we constructed PostNets for each modality so that they can estimate the modality-specific uncertainty independently and be used in the uncertainty-aware feature fusion process. The use of multiple PostNets is similar to [20] in their construction of separate PostNets for agent's past behavior, road structure map, and social context respectively. For latent deterministic uncertainty (LDU), we set the number of learnable prototypes to the number of scenarios. Specifically, for aiMootive, the scenarios are Highway, Urban, Night, and Rain, while for DeLiVER, the scenarios are Cloudy, Foggy, Night, Rainy, and Sunny. Finally, for all methods, all layers prior to the UQM are kept frozen and we fine-tune the uncertainty weighting layer, which takes the respective uncertainty metrics as input, along with other post-fusion layers, to enable the model to learn the uncertainty-weighted features. Details of the fine-tuning procedures and visualizations of the architectures and UQM insertion points are provided in the Appendix. In all of our experiments the best one is highlighted and bolded and the 2nd best is lightly-highlighted.", + "bbox": [ + 511, + 90, + 908, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Multimodal Uncertainty Fusion for Autonomous Vehicles Perception", + "text_level": 1, + "bbox": [ + 511, + 446, + 906, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.1 3D Object Detection", + "text_level": 1, + "bbox": [ + 511, + 484, + 709, + 500 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/f00aa41568f998fa00e84989d6ff800366e91f98b30df67cf985d30a1e4fe8a9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
UQ MethodHighwayUrbanNightRainMean
BEVFusion [33]72.55/69.0863.72/63.5274.74/73.1342.75/42.8365.67/64.79
InfMCD [34]72.58/69.6562.02/59.5975.05/73.7235.58/37.4164.56/64.58
InfNoise [34]70.11/68.5263.10/60.4275.08/74.1935.58/37.4164.59/64.56
PostNet [5]69.19/68.0261.38/59.8473.32/68.9734.95/37.6063.13/60.21
LDU [10]71.48/69.3362.54/60.1176.49/75.0440.49/41.2464.69/64.73
HyperDUM72.23/69.5864.77/64.4876.69/75.1544.78/45.4866.70/66.00
", + "bbox": [ + 514, + 520, + 906, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1 compares the all-point/11-point interpolation Average Precision (AP) metric for object detection under diverse scenes. It can be seen that apart from the Highway scenario, HyperDUM outperforms all methods across all scenarios. Specifically, HyperDUM improves upon the all-point AP by $2.01\\%$ on average against the state-of-the-art (LDU).", + "bbox": [ + 511, + 638, + 908, + 729 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/93f0be36abd0e4e6fc437cd49876cf34598c99b6bdec3ec41a27be94e08ecef8.jpg", + "table_caption": [ + "Table 1. Results on aiMotive under diverse scenes." + ], + "table_footnote": [], + "table_body": "
UQ MethodMBOEUELFMean
BEVFusion [33]64.10/63.6962.40/59.2663.57/63.2565.07/65.0863.79/62.75
InfMCD [34]62.00/61.9964.47/64.2863.25/63.4964.27/63.8163.50/63.39
InfNoise [34]62.61/62.6164.94/64.7263.18/60.0765.19/64.9963.98/63.10
PostNet [5]60.42/58.9063.12/59.9960.20/59.2564.13/60.6061.97/59.69
LDU [10]62.06/62.2265.07/64.8562.87/60.0265.47/65.1063.87/63.05
HyperDUM64.39/63.9966.16/65.3964.18/63.9165.89/65.2265.16/64.62
", + "bbox": [ + 514, + 739, + 905, + 815 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2. Results on aiMotive under corner cases: Motion Blur (MB), Over-Exposure (OE), Under-Exposure (UE), LiDAR-Fog (LF). (Metrics: all-point AP/11-point interpolation AP)", + "bbox": [ + 511, + 828, + 906, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2 demonstrates the robustness of UQ methods un", + "bbox": [ + 532, + 885, + 908, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22310", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/18f3f884b122647cd429b9c7351cd341d8ce7c342c38bdeb4106a85bfbf62696.jpg", + "image_caption": [ + "Original" + ], + "image_footnote": [], + "bbox": [ + 99, + 88, + 906, + 222 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1ebd5c7b4d839028ec73d946db4182304852d06e5c77471a041ac0bcc3a6af40.jpg", + "image_caption": [ + "Original" + ], + "image_footnote": [], + "bbox": [ + 98, + 236, + 295, + 335 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d04c34863021ca7bd143cc4105ff25333fea4ed304d0673ec6196a77d6aa562d.jpg", + "image_caption": [ + "Motion Blur", + "Figure 2. Corner cases for the aiMotive 3D Object Detection dataset. (Top-Left) Normal Lidar point-cloud birds-eye-view. (Top-Right) Lidar Foggification (LF). (Bottom) In order from left to right (Normal, Motion Blur (MB), Over-Exposure (OE), Under-Exposure (UE))" + ], + "image_footnote": [], + "bbox": [ + 300, + 237, + 496, + 335 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5db4b7439059f0370f743806c814ea632f5fd496cd2d77e7a5f18bb93b3bbc25.jpg", + "image_caption": [ + "Foggification", + "Over-Exposure" + ], + "image_footnote": [], + "bbox": [ + 501, + 237, + 700, + 335 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/575cd4a22a55f882b1974eac25a54afa14eecbefb7fe86ba2282d1e7e2f1d0bf.jpg", + "image_caption": [ + "Under-Exposure" + ], + "image_footnote": [], + "bbox": [ + 704, + 237, + 901, + 335 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "der our injected corner cases. These injections were unseen during training which complicates testing due to distribution shifts. Each corner case is applied to and evaluated on the entire test set across all scenarios. Whereas other methods fail and suffer performance degradations, HyperDUM only degrades $0.51\\% / 0.17\\%$ compared to the baseline w/o injections demonstrating that HyperDUM is more robust even when facing distribution shifts.", + "bbox": [ + 89, + 422, + 483, + 542 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1739e5a183815bfcfea0c48a1d7e1713a51852a76414928bbbbfab2f84fd1344.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
UQ MethodHighwayUrbanNightRainMean
BEVFusion [33]45.59/45.3446.38/45.5344.49/45.8742.65/42.9345.84/46.87
InfMCD [34]45.57/46.2938.90/38.1142.07/41.9726.13/28.4940.74/41.64
InfNoise [34]40.73/40.3842.41/40.9038.80/40.0029.28/32.8140.91/42.04
PostNet [5]41.13/40.6439.47/40.1640.37/41.9228.56/28.9639.37/41.66
LDU [10]43.07/45.0140.17/41.7142.90/42.5231.12/34.0840.65/42.17
HyperDUM44.86/46.0647.89/48.9144.42/46.1240.54/41.6346.52/48.27
", + "bbox": [ + 93, + 585, + 483, + 662 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3 is an experiment on the distant region setting, keeping only predictions and ground truths over $\\geq 75m$ . As expected, all UQ methods experience major performance drops, especially for Rain. This is likely due to the combined uncertainties of distance with rain obstruction causing substantial minimization of the uncertainty-weighted features (the weighting output is between 0-1 times the original feature value). The original BEVFusion model maintains better performance because it is still able to \"guess\" the presence of objects as the features are not minimized by uncertainty.", + "bbox": [ + 89, + 750, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/59caed25341756cdbe980c2b885cd145c740acbbf95f0c36bdaf5cf9c10ffdce.jpg", + "table_caption": [ + "Table 3. Results on aiMotive under diverse scenes in the distant region ( $>75\\mathrm{m}$ ). (Metrics: all-point AP/11-point interpolation AP)" + ], + "table_footnote": [], + "table_body": "
ScenariosUQ Methods (Metrics: mIoU ↑)
CMNeXt [58]InfMCD [34]InfNoise [34]PostNet [5]LDU [10]Gemini [21]HyperDUM (Ours)
Cloudy68.7069.2369.2169.2868.94-69.76+1.06
Foggy65.6666.1866.1066.0465.72-66.85+1.19
Night62.4663.4463.1462.9763.06-64.21+1.75
Rainy67.5068.0868.0968.1667.82-68.71+1.21
Sunny66.5767.0167.1666.9366.75-67.87+1.30
MB62.9163.6163.5563.5563.16-64.28+1.37
OE64.5965.3965.1765.0664.73-65.67+1.08
UE60.0060.3860.4260.2760.29-61.20+1.20
LJ65.9266.1266.2566.3366.40-66.93+1.01
EL65.4866.0566.1766.0665.89-66.80+1.32
Mean66.3066.9066.8666.7766.6066.9067.59+1.29
", + "bbox": [ + 514, + 419, + 906, + 585 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4. DeLiVER mean Intersection over Union (mIoU) performance under adverse weather and corner cases. Lidar-Jitter (LJ), Event Low-resolution (EL).", + "bbox": [ + 511, + 598, + 906, + 640 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.2 Semantic Segmentation", + "text_level": 1, + "bbox": [ + 511, + 672, + 732, + 689 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4 shows the DeLiVER validation set performance on different scenarios and corner cases for the semantic segmentation task. Overall, HyperDUM achieves the best performance over all methods across every weather and corner case scenario. On average HyperDUM improves over the baseline by 1.29. In particular, for the three most difficult scenarios including Under-Exposure, Night and Motion Blur, HyperDUM achieves at least 1.00 improvement over the baseline. For GeminiFusion [21], we only report the mean as they omitted the detailed performance breakdown in their paper. These results demonstrate that our method can better quantify the multimodal feature uncertainties and ultimately improve the fusion features for the semantic segmentation", + "bbox": [ + 509, + 703, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22311", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "task. Additionally, we note that all UQ methods improved over the baseline method. This shows that incorporating UQ with fusion into multimodal pipelines not only enables UQ but also can improve baseline performance.", + "bbox": [ + 89, + 90, + 483, + 151 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f67d38a811f1607ed83ec95873f0566ab5d5ef66ca8a4bc5e4b7680ecc0f5c9c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ScenariosUQ Methods (Metrics: ECE ↓)
CMNeXt [58]InfMCD [34]InfNoise [34]PostNet [5]LDU [10]HyperDUM (Ours)
Cloudy1.18E-021.05E-021.07E-020.99E-021.09E-020.94E-02-0.24
Foggy1.67E-021.48E-021.52E-021.45E-021.56E-021.35E-02-0.30
Night1.88E-021.63E-021.68E-021.71E-021.68E-021.39E-02-0.49
Rainy1.53E-021.37E-021.39E-021.33E-021.43E-021.21E-02-0.32
Sunny1.42E-021.25E-021.27E-021.24E-021.33E-021.12E-02-0.30
MB1.54E-021.32E-021.40E-021.32E-021.44E-021.19E-02-0.35
OE1.49E-021.55E-021.39E-021.42E-021.44E-021.31E-02-0.18
UE1.75E-022.07E-021.69E-021.84E-021.63E-021.53E-02-0.22
LJ1.54E-021.38E-021.39E-021.35E-021.40E-021.28E-02-0.26
EL1.45E-021.17E-021.27E-021.26E-021.27E-021.09E-02-0.36
Mean1.53E-021.35E-021.38E-021.34E-021.41E-021.20E-02-0.33
", + "bbox": [ + 91, + 162, + 486, + 330 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5 shows the Expected Calibration Error (ECE) metric for all methods across the same scenarios. ECE measures how well a model is calibrated by comparing it's predicted probabilities with the actual probabilities of the ground truth distribution. ECE helps us understand how reliable a model's confidence scores are. A higher ECE means the model is more overconfident (or unsure) about its predictions. As shown, HyperDUM has the lowest ECE score over all methods across every weather scenario and corner case with a 0.33 average decrease compared to the baseline. These results demonstrate that our method helps better calibrate the overconfident semantic segmentation model. In fact all UQ methods on average has a lower ECE compared to the baseline demonstrating its usefulness in model calibration.", + "bbox": [ + 88, + 387, + 485, + 599 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Pre vs. Post Uncertainty Feature Learning", + "text_level": 1, + "bbox": [ + 89, + 609, + 452, + 625 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/38adbab919183786e9d369fedb6d5a53eed71b4acf0ac6f1797335784e0b023d.jpg", + "table_caption": [ + "Table 5. DeLiVER Expected Calibration Error (ECE) performance under adverse weather and corner cases." + ], + "table_footnote": [], + "table_body": "
ScenariosUQ Methods (Metrics: mIoU ↑)
CMNeXt [58]InfMCD [34]InfNoise [34]PostNet [5]LDU [10]HyperDUM (Ours)
Cloudy68.768.7868.569.1369.0669.26
Foggy65.6665.7865.5965.8965.9266.40
Night62.4662.762.4462.8263.0863.71
Rainy67.567.6667.3768.0267.8468.29
Sunny66.5766.5266.3966.7866.7867.43
MB62.9163.1162.7963.4063.363.78
OE64.5964.9464.664.9164.9565.20
UE60.0059.7560.0760.1260.360.72
LJ65.9266.1466.0166.1766.2166.44
EL65.4865.765.6665.9166.0366.34
Mean66.3066.5966.4966.6066.6466.98
", + "bbox": [ + 94, + 643, + 477, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6 evaluates the performance of uncertainty feature weighting after the fusion block of Figure 1 or see the archi", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tecture figure in the Appendix. As expected, we found that the overall performance gain is limited/ lower compared to its prefusion uncertainty fusion counterpart. These results align with our motivation that different modalities experience different modality-specific uncertainties. And these uncertainties can propagate to the fusion module, undermining the fusion features and degrading downstream task performance.", + "bbox": [ + 511, + 90, + 908, + 196 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Model Ablation", + "text_level": 1, + "bbox": [ + 513, + 205, + 671, + 220 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b91997c9cbebb6e9604276b6b81a32746629980a52c6b1c8be8478ac2a8d317d.jpg", + "table_caption": [ + "Table 6. DeLiVER mean Intersection over Union (mIoU) performance under adverse weather and corner cases using post-fusion features." + ], + "table_footnote": [], + "table_body": "
Dataset\nMetricaiMotive\nall-point AP/11-point AP ↑DeLiVER\nmIoU ↑
HyperDUM\n- w/o Patch (4x) Proj.-66.70/66.00 (Mean Δ)67.59 (Mean Δ)
-65.67/65.23 (-1.03/-0.77)-67.18 (-0.41)
- w/o Chan Proj.64.43/64.25 (-1.24/-0.98)66.63 (-0.55)
", + "bbox": [ + 535, + 237, + 883, + 297 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 7. Model ablation for aiMotive and DeLiVER.", + "bbox": [ + 550, + 306, + 864, + 320 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For our model ablation, we analyzed the effects of sequentially removing our proposed components and the impact on the overall performance. As seen from Table 7 the removal of the Patch Projection (with 4 patches) results in a $1.03 / 0.77$ decrease in performance and the removal of Channel Projection in a $1.24 / 0.98$ decrease in performance for aiMotive. Similarly for the DeLiVER dataset, the removal of the Patch Projection (with 4 patches) results in a 0.41 decrease in performance and removing Channel Projection results in a 0.55 decrease in performance. Together we see that both methods contribute significantly to the overall performance demonstrating that both channel and spatial uncertainties should be accounted for to better capture the holistic uncertainty of the features before fusion. More ablations in the Appendix.", + "bbox": [ + 511, + 339, + 908, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4.1 Sensitivity Analysis", + "text_level": 1, + "bbox": [ + 511, + 568, + 702, + 584 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/922405aa88be352a81f0f4285c6fbc1fe69fb6c35dc991a9ac0256c67679f586.jpg", + "image_caption": [ + "Figure 3. Analysis on the effects of hyperdimension selection for channel projection on performance and computation (DeLiVER)." + ], + "image_footnote": [], + "bbox": [ + 526, + 609, + 895, + 780 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Here we investigate the tunable parameters for channel and patch projection & bundling (CPB & PPB). For CPB, the only parameter is the channel hyperdimension. For PPB, we have the patch hyperdimension and the number of patches.", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22312", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/198eb036dd032be08ad28d358f231b8de37a1fa24d0312ac4c023745ba753395.jpg", + "image_caption": [ + "Figure 3 shows a sweep over a range of dimensions from 625 to $20\\mathrm{k}$ for CPB. This range of values is chosen based on the HDC expressivity limits (around $10\\mathrm{k}$ ) [48, 57]. We see that performance improves with dimension increase, but plateaus and decreases past $10\\mathrm{k}$ while computation continues to grow. This occurs when large dimensions induce too much sparsity, making similar samples become orthogonal and capturing irrelevant information likened to over-fitting in neural networks [48].", + "Figure 4. Analysis on the effects of patches and patch hyperdimension selection on performance and computation (DeLiVER)." + ], + "image_footnote": [], + "bbox": [ + 101, + 241, + 473, + 385 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 4 shows a sweep over 4, 16 and 64 patch configurations together with the dimension sweep from 625 to 5000. We observe an interesting relation where the number of patches negatively correlates with the per-patch hyperdimension in terms of better performance. This makes sense as we increase the number of patches, the information needed to be encoded becomes smaller due to the reduction in the per-patch spatial resolution as indicated in Section 3.2.2. Therefore, the necessary hyperdimensions needed to encode the information should be reduced, while larger dimensions begin to over-fit. Whereas the opposite is true for less patches, larger hyperdimensions are needed.", + "bbox": [ + 88, + 445, + 486, + 626 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Computation Cost", + "text_level": 1, + "bbox": [ + 89, + 633, + 269, + 648 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f4c3ea05d8b8b731b46d8975a202f86d041adf9ccfd6abe1cda10eda70198016.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetaiMotiveDeLiVER
UQ MethodFLOPs#ParamsFLOPs#Params
InfMCD [34]990.90M4.05K1.84G20.19K
InfNoise [34]990.90M4.05K1.84G20.19K
LDU [10]264.24M44.04M613.42M102.24M
HyperDUM111.84M1.15M338.40M38.85M
- w/o Patch (4x) Proj.84.00M1.13M256.00M38.78M
- w/o Chan Proj.6.96M3.38K20.60M16.83K
", + "bbox": [ + 116, + 667, + 455, + 757 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 8. Comparing inference (FLOPs) and training costs (# of Trainable Parameters).", + "bbox": [ + 89, + 767, + 482, + 795 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finally, the main benefit of HyperDUM is its efficiency for training and inference. We measure the computation costs in terms of Floating point operations (FLOPs) and the number of trainable model parameters in Table 8. We approximate the FLOPs by computing the major contributing operations. Refer to the Appendix for our FLOPs approximations", + "bbox": [ + 89, + 809, + 485, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "for each operation type. For InfMCD and InfNoise these are the entropy, mutual information, and predictive entropy computations. For LDU, it is the cosine similarity between prototypes and matrix multiplication for uncertainty, which was found to be negligible. For HyperDUM, it is the input projection and cosine similarity computations. We see that HyperDUM is up to $2.36 \\times$ and $1.81 \\times$ more FLOP-efficient compared to LDU for aiMotive and DeLiVER respectively. Compared to traditional Bayesian approximation methods, it can be up to $8.86 \\times$ and $5.44 \\times$ more FLOP-efficient. As for the training costs, HyperDUM has $38.30 \\times$ and $2.63 \\times$ less trainable parameters compared to LDU for aiMotive and DeLiVER respectively, making it more memory efficient. This ultimately shows the practicality of HyperDUM as a UQ method for real-world autonomous systems.", + "bbox": [ + 511, + 90, + 906, + 318 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Discussions and Conclusion", + "text_level": 1, + "bbox": [ + 511, + 330, + 769, + 347 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and Future Work: We leveraged cloud computing resources using a single NVIDIA A100 GPU for training. Our method uses the traditional bundling technique, which requires labels for supervised learning of prototypes. The need for labels is the primary limitation of our approach. Although there have been implementations of semi-supervised and unsupervised versions [17, 18], there are no theoretical guarantees and/or justifications for their performance unlike the traditional bundling method [48]. We leave the exploration of these methods for uncertainty quantification to future work. Additionally, the FLOPs of the vector symbolic architectures may be further reduced due to the massively parallel operations of hyperdimensional computation. These optimizations were demonstrated in prior works [13, 19].", + "bbox": [ + 511, + 356, + 908, + 568 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Conclusion: In this work, we propose an efficient method to quantify the epistemic uncertainty of machine learning models at the feature level. HyperDUM projects the latent features as hypervectors to form hyperdimensional prototypes. Similarity between hypervectors of new samples and learned prototypes is used as a proxy to estimate uncertainty. We demonstrate the practicality of HyperDUM through autonomous driving tasks, including object detection and semantic segmentation. Our results show that our method can enhance the performance and robustness of multimodal fusion models under uncertainties induced by various weather conditions and modality-specific corner cases. We achieve this by quantifying the feature uncertainty of each modality and applying an uncertainty-aware weighting layer prior to feature fusion and fine-tuning the post-fusion layers. Finally, we demonstrated that HyperDUM requires significantly less FLOPs and training parameters when compared to existing works, indicating its real-world practicality.", + "bbox": [ + 511, + 568, + 908, + 839 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements: This work acknowledges the support of the Automotive Research Center (ARC), Cooperative Agreement W56HZV-24-2-0001 U.S. Army DEVCOM Ground Vehicle Systems Center (GVSC).", + "bbox": [ + 511, + 839, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22313", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Anastasios N Angelopoulos, Stephen Bates, et al. Conformal prediction: A gentle introduction. Foundations and Trends® in Machine Learning, 16(4):494-591, 2023. 1, 2", + "[2] Sri Hrushikesh Varma Bhupathiraju, Jennifer Sheldon, Luke A Bauer, Vincent Bindschaedler, Takeshi Sugawara, and Sara Rampazzi. Emi-lidar: Uncovering vulnerabilities of lidar sensors in autonomous driving setting using electromagnetic interference. In Proceedings of the 16th ACM Conference on Security and Privacy in Wireless and Mobile Networks, pages 329-340, 2023. 1", + "[3] Mario Bijelic, Tobias Gruber, Fahim Mannan, Florian Kraus, Werner Ritter, Klaus Dietmayer, and Felix Heide. Seeing through fog without seeing fog: Deep multimodal sensor fusion in unseen adverse weather. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11682-11692, 2020. 5, 7", + "[4] Hermann Blum, Paul-Edouard Sarlin, Juan Nieto, Roland Siegwart, and Cesar Cadena. Fishyscapes: A benchmark for safe semantic segmentation in autonomous driving. In proceedings of the IEEE/CVF international conference on computer vision workshops, pages 0–0, 2019. 1", + "[5] Bertrand Charpentier, Daniel Zügner, and Stephan Gümnmann. Posterior network: Uncertainty estimation without ood samples via density-based pseudo-counts. Advances in neural information processing systems, 33:1356-1367, 2020. 1, 2, 5, 6, 7", + "[6] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 4", + "[7] Michael Dusenberry, Ghassen Jerfel, Yeming Wen, Yian Ma, Jasper Snoek, Katherine Heller, Balaji Lakshminarayanan, and Dustin Tran. Efficient and scalable bayesian neural nets with rank-1 factors. In International conference on machine learning, pages 2782-2792. PMLR, 2020. 1, 2", + "[8] Jamil Fayyad, Mohammad A Jaradat, Dominique Gruyer, and Homayoun Najjaran. Deep learning sensor fusion for autonomous vehicle perception and localization: A review. Sensors, 20(15):4220, 2020. 1, 2", + "[9] Florian Fervers, Sebastian Bullinger, Christoph Bodensteiner, Michael Arens, and Rainer Stiefelhagen. Uncertainty-aware vision-based metric cross-view geolocation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21621-21631, 2023. 1", + "[10] Gianni Franchi, Xuanlong Yu, Andrei Bursuc, Emanuel Aldea, Severine Dubuisson, and David Filliat. Latent discriminant deterministic uncertainty. In European Conference on Computer Vision, pages 243-260. Springer, 2022. 1, 2, 3, 5, 6, 7, 8", + "[11] Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In international conference on machine learning, pages 1050-1059. PMLR, 2016. 1, 2", + "[12] Ethan Goan and Clinton Fookes. Uncertainty in real-time semantic segmentation on embedded systems. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4490-4500, 2023. 1" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Saransh Gupta, Justin Morris, Mohsen Imani, Ranganathan Ramkumar, Jeffrey Yu, Aniket Tiwari, Baris Aksanli, and Tajana Simunic Rosing. Thrifty: Training with hyperdimensional computing across flash hierarchy. In Proceedings of the 39th International Conference on Computer-Aided Design, pages 1-9, 2020. 8", + "[14] Zongbo Han, Fan Yang, Junzhou Huang, Changqing Zhang, and Jianhua Yao. Multimodal dynamics: Dynamical fusion for trustworthy multimodal classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20707-20717, 2022. 2", + "[15] Florian Heidecker, Jasmin Breitenstein, Kevin Rösch, Jonas Löhdefink, Maarten Bieshaar, Christoph Stiller, Tim Fingscheidt, and Bernhard Sick. An application-driven conceptualization of corner cases for perception in highly automated driving. In 2021 IEEE Intelligent Vehicles Symposium (IV), pages 644-651. IEEE, 2021. 1", + "[16] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17853-17862, 2023. 1", + "[17] Mohsen Imani, Samuel Bosch, Mojan Javaheripi, Bita Rouhani, Xinyu Wu, Farinaz Koushanfar, and Tajana Rosing. Semihd: Semi-supervised learning using hyperdimensional computing. In 2019 IEEE/ACM International Conference on Computer-Aided Design (ICCAD), pages 1-8. IEEE, 2019. 8", + "[18] Mohsen Imani, Yeseong Kim, Thomas Worley, Saransh Gupta, and Tajana Rosing. Hdcluster: An accurate clustering using brain-inspired high-dimensional computing. In 2019 Design, Automation & Test in Europe Conference & Exhibition (DATE), pages 1591-1594. IEEE, 2019. 8", + "[19] Mohsen Imani, Saikishan Pampana, Saransh Gupta, Minxuan Zhou, Yeseong Kim, and Tajana Rosing. Dual: Acceleration of clustering algorithms using digital-based processing in-memory. In 2020 53rd Annual IEEE/ACM International Symposium on Microarchitecture (MICRO), pages 356-371. IEEE, 2020. 8", + "[20] Masha Itkina and Mykel Kochenderfer. Interpretable self-aware neural networks for robust trajectory prediction. In Conference on Robot Learning, pages 606-617. PMLR, 2023. 5", + "[21] Ding Jia, Jianyuan Guo, Kai Han, Han Wu, Chao Zhang, Chang Xu, and Xinghao Chen. Geminifusion: Efficient pixelwise multimodal fusion for vision transformer. International Conference on Machine Learning (ICML), 2024. 6", + "[22] Xiaotao Jia, Jianlei Yang, Runze Liu, Xueyan Wang, Sorin Dan Cotofana, and Weisheng Zhao. Efficient computation reduction in bayesian neural networks through feature decomposition and memorization. IEEE transactions on neural networks and learning systems, 32(4):1703-1712, 2020. 1, 2", + "[23] Yongchan Kwon, Joong-Ho Won, Beom Joon Kim, and Myunghee Cho Paik. Uncertainty quantification using bayesian neural networks in classification: Application to biomedical image segmentation. Computational Statistics & Data Analysis, 142:106816, 2020. 1" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22314", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scalable predictive uncertainty estimation using deep ensembles. Advances in neural information processing systems, 30, 2017. 1, 2", + "[25] Simon D Levy and Ross Gayler. Vector symbolic architectures: A new building material for artificial general intelligence. In Proceedings of the 2008 conference on artificial general intelligence 2008: Proceedings of the first AGI conference, pages 414-418, 2008. 3", + "[26] Hao Li, Jingkuan Song, Lianli Gao, Xiaosu Zhu, and Hengtao Shen. Prototype-based aleatoric uncertainty quantification for cross-modal retrieval. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[27] Xiaotong Li, Yongxing Dai, Yixiao Ge, Jun Liu, Ying Shan, and Ling-Yu Duan. Uncertainty modeling for out-of-distribution generalization. arXiv preprint arXiv:2202.03958, 2022.4", + "[28] You Li and Javier Ibanez-Guzman. Lidar for autonomous driving: The principles, challenges, and trends for automotive lidar and perception systems. IEEE Signal Processing Magazine, 37(4):50-61, 2020. 1", + "[29] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1477-1485, 2023. 5", + "[30] Ming Liang, Bin Yang, Shenlong Wang, and Raquel Urtasun. Deep continuous fusion for multi-sensor 3d object detection. In Proceedings of the European conference on computer vision (ECCV), pages 641-656, 2018. 2", + "[31] Arnav Vaibhav Malawade, Trier Mortlock, and Mohammad Abdullah Al Faruque. Hydraulics: Context-aware selective sensor fusion for robust and efficient autonomous vehicle perception. In 2022 ACM/IEEE 13th International Conference on Cyber-Physical Systems (ICCPs), pages 68–79. IEEE, 2022. 1, 2", + "[32] Andrey Malinin and Mark Gales. Predictive uncertainty estimation via prior networks. Advances in neural information processing systems, 31, 2018. 2", + "[33] Tamas Matuszka, Ivan Barton, Adam Butykai, Peter Hajas, David Kiss, Domonkos Kovacs, Sandor Kunsagi-Mate, Peter Lengyel, Gabor Nemeth, Leventa PetHo, et al. aimotive dataset: A multimodal dataset for robust autonomous driving with long-range perception. arXiv preprint arXiv:2211.09445, 2022. 5, 6", + "[34] Lu Mi, Hao Wang, Yonglong Tian, Hao He, and Nir N Shavit. Training-free uncertainty estimation for dense regression: Sensitivity as a surrogate. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10042-10050, 2022. 1, 2, 5, 6, 7, 8", + "[35] Trier Mortlock, Arnav Malawade, Kohei Tsujio, and Mohammad Al Faruque. Castnet: a context-aware, spatio-temporal dynamic motion prediction ensemble for autonomous driving. ACM Transactions on Cyber-Physical Systems, 8(2):1–20, 2024. 1", + "[36] Jishnu Mukhoti, Andreas Kirsch, Joost van Amersfoort, Philip HS Torr, and Yarin Gal. Deep deterministic uncertainty:" + ], + "bbox": [ + 91, + 92, + 483, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "A new simple baseline. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24384-24394, 2023. 1, 2", + "[37] Yang Ni, Hanning Chen, Prathyush Poduval, Zhuowen Zou, Pietro Mercati, and Mohsen Imani. Brain-inspired trustworthy hyperdimensional computing with efficient uncertainty quantification. In 2023 IEEE/ACM International Conference on Computer Aided Design (ICCAD), pages 01-09. IEEE, 2023. 2", + "[38] Igor Nunes, Mike Heddes, Tony Givargis, Alexandru Nicolau, and Alex Veidenbaum. Graphhd: Efficient graph classification using hyperdimensional computing. In 2022 Design, Automation & Test in Europe Conference & Exhibition (DATE), pages 1485-1490. IEEE, 2022. 3", + "[39] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch. In NIPS-W, 2017. 5", + "[40] Luis Manuel Pereira, Addisson Salazar, and Luis Vergara. On comparing early and late fusion methods. In International Work-Conference on Artificial Neural Networks, pages 365-378. Springer, 2023. 2", + "[41] Janis Postels, Francesco Ferroni, Huseyin Coskun, Nassir Navab, and Federico Tombari. Sampling-free epistemic uncertainty estimation using approximated variance propagation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2931-2940, 2019. 1", + "[42] Janis Postels, Mattia Segu, Tao Sun, Luca Sieber, Luc Van Gool, Fisher Yu, and Federico Tombari. On the practicality of deterministic epistemic uncertainty. arXiv preprint arXiv:2107.00649, 2021. 1, 2", + "[43] Ali Rahimi and Benjamin Recht. Random features for large-scale kernel machines. Advances in neural information processing systems, 20, 2007. 1", + "[44] Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Schölkopf, Thomas Brox, and Peter Gehler. Towards total recall in industrial anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14318-14328, 2022. 4", + "[45] Kenny Schlegel, Peer Neubert, and Peter Protzel. A comparison of vector symbolic architectures. Artificial Intelligence Review, 55(6):4523-4555, 2022. 3", + "[46] Jiahui She, Yibo Hu, Hailin Shi, Jun Wang, Qiu Shen, and Tao Mei. Dive into ambiguity: Latent distribution mining and pairwise uncertainty estimation for facial expression recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6248-6257, 2021. 2", + "[47] Apoorv Singh. Transformer-based sensor fusion for autonomous driving: A survey. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3312-3317, 2023. 2", + "[48] Anthony Thomas, Sanjoy Dasgupta, and Tajana Rosing. A theoretical perspective on hyperdimensional computing. Journal of Artificial Intelligence Research, 72:215-249, 2021. 3, 4, 8, 1", + "[49] Junjiao Tian, Wesley Cheung, Nathaniel Glaser, Yen-Cheng Liu, and Zsolt Kira. Uno: Uncertainty-aware noisy-or multimodal fusion for unanticipated input degradation. In 2020" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22315", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE International Conference on Robotics and Automation (ICRA), pages 5716-5723. IEEE, 2020. 5, 2", + "[50] Stefan Van der Walt, Johannes L Schonberger, Juan Nunez-Iglesias, François Boulogne, Joshua D Warner, Neil Yager, Emmanuelle Gouillart, and Tony Yu. scikit-image: image processing in python. PeerJ, 2:e453, 2014. 5", + "[51] Jorge Vargas, Suleiman Alsweiss, Onur Token, Rahul Razdan, and Joshua Santos. An overview of autonomous vehicles sensors and their vulnerability to weather conditions. Sensors, 21(16):5397, 2021. 1", + "[52] Junyao Wang, Sitao Huang, and Mohsen Imani. Disthd: A learner-aware dynamic encoding method for hyperdimensional classification. In 2023 60th ACM/IEEE Design Automation Conference (DAC), pages 1-6. IEEE, 2023. 4", + "[53] Junyao Wang, Arnav Vaibhav Malawade, Junhong Zhou, Shih-Yuan Yu, and Mohammad Abdullah Al Faruque. Rs2g: Data-driven scene-graph extraction and embedding for robust autonomous perception and scenario understanding. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 7493-7502, 2024. 1", + "[54] Shaojie Wang, Tong Wu, Ayan Chakrabarti, and Yevgeniy Vorobeychik. Adversarial robustness of deep sensor fusion models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2387-2396, 2022. 1", + "[55] Samuel Wilson, Tobias Fischer, Niko Sünderhauf, and Feras Dayoub. Hyperdimensional feature fusion for out-of-distribution detection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2644-2654, 2023. 4", + "[56] De Jong Yeong, Gustavo Velasco-Hernandez, John Barry, and Joseph Walsh. Sensor and sensor fusion technology in autonomous vehicles: A review. Sensors, 21(6):2140, 2021. 1, 2", + "[57] Tao Yu, Yichi Zhang, Zhiru Zhang, and Christopher M De Sa. Understanding hyperdimensional computing for parallel single-pass learning. Advances in Neural Information Processing Systems, 35:1157-1169, 2022. 3, 8, 1, 7", + "[58] Jiaming Zhang, Ruiping Liu, Hao Shi, Kailun Yang, Simon Reiβ, Kunyu Peng, Haodong Fu, Kaiwei Wang, and Rainer Stiefelhagen. Delivering arbitrary-modal semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1136-1147, 2023. 5, 6, 7", + "[59] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4490-4499, 2018. 5", + "[60] Zijian Zhu, Yichi Zhang, Hai Chen, Yinpeng Dong, Shu Zhao, Wenbo Ding, Jiachen Zhong, and Shibao Zheng. Understanding the robustness of 3d object detection with bird's-eye-view representations in autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21600-21610, 2023. 1" + ], + "bbox": [ + 91, + 90, + 485, + 853 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "22316", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Hyperdimensional Uncertainty Quantification for Multimodal Uncertainty Fusion in Autonomous Vehicles Perception/8d8555dd-a579-45c4-ad75-fee4bbfad5af_model.json b/2025/Hyperdimensional Uncertainty Quantification for Multimodal Uncertainty Fusion in Autonomous Vehicles Perception/8d8555dd-a579-45c4-ad75-fee4bbfad5af_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5c8d753bc3f9db0aa48d5bd5745f47fc5bf5d668 --- /dev/null +++ b/2025/Hyperdimensional Uncertainty Quantification for Multimodal Uncertainty Fusion in Autonomous Vehicles Perception/8d8555dd-a579-45c4-ad75-fee4bbfad5af_model.json @@ -0,0 +1,2323 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.13, + 0.908, + 0.177 + ], + "angle": 0, + "content": "Hyperdimensional Uncertainty Quantification for Multimodal Uncertainty Fusion in Autonomous Vehicles Perception" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.204, + 0.905, + 0.24 + ], + "angle": 0, + "content": "Luke Chen* Junyao Wang Trier Mortlock Pramod Khargonekar Mohammad Abdullah Al Faruque University of California, Irvine" + }, + { + "type": "text", + "bbox": [ + 0.199, + 0.242, + 0.8, + 0.258 + ], + "angle": 0, + "content": "{panwangc, junyaow4, tmortloc, pramod.khargonekar, alfaruqu}@uci.edu" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.324, + 0.486, + 0.702 + ], + "angle": 0, + "content": "Uncertainty Quantification (UQ) is crucial for ensuring the reliability of machine learning models deployed in real-world autonomous systems. However, existing approaches typically quantify task-level output prediction uncertainty without considering epistemic uncertainty at the multimodal feature fusion level, leading to sub-optimal outcomes. Additionally, popular uncertainty quantification methods, e.g., Bayesian approximations, remain challenging to deploy in practice due to high computational costs in training and inference. In this paper, we propose HyperDUM, a novel deterministic uncertainty method (DUM) that efficiently quantifies feature-level epistemic uncertainty by leveraging hyperdimensional computing. Our method captures the channel and spatial uncertainties through channel and patch-wise projection and bundling techniques respectively. Multimodal sensor features are then adaptively weighted to mitigate uncertainty propagation and improve feature fusion. Our evaluations show that HyperDUM on average outperforms the state-of-the-art (SOTA) algorithms by up to \\(2.01\\% / 1.27\\%\\) in 3D Object Detection and up to \\(1.29\\%\\) improvement over baselines in semantic segmentation tasks under various types of uncertainties. Notably, HyperDUM requires \\(2.36 \\times\\) less Floating Point Operations and up to \\(38.30 \\times\\) less parameters than SOTA methods, providing an efficient solution for real-world autonomous systems." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.732, + 0.222, + 0.749 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.758, + 0.485, + 0.864 + ], + "angle": 0, + "content": "Effectively quantifying and addressing uncertainty plays a critical role in ensuring the reliability of ML models in real-world applications, especially where the decisions are consequential, e.g., autonomous driving. Models failing to address uncertainty can produce overconfident outputs, leading to compromised performance [9, 16, 41, 53]. In light of autonomous systems, uncertainties can arise from adverse" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.293, + 0.909, + 0.534 + ], + "angle": 0, + "content": "weather conditions, sensor failures or noises, environment dynamics (movements of other actors, road conditions, vehicle dynamics), and various corner cases [4, 12, 35, 36]. Modern autonomous systems leverage multimodal sensor fusion to broaden the information scope and learn comprehensive representations of the surrounding environment [8, 31, 56]. However, commonly employed sensors such as cameras, lidar, and radar, exhibit varying degrees of robustness and vulnerability across different driving scenarios [51, 54]. For instance, cameras are crucial in delivering rich perceptual information while they are particularly vulnerable to lighting variations and occlusions [60]. Lidar and radar are more robust to visual perturbations but can be affected by multi-path interference and occlusion [2, 28]. Therefore, it is important to consider the uncertainty of each modality before the fusion of multimodal features to achieve optimal results [15]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.537, + 0.909, + 0.839 + ], + "angle": 0, + "content": "Popular Uncertainty Quantification (UQ) methods include deep ensembles, Bayesian neural networks, and Bayesian approximation methods such as Monte Carlo dropout [11, 23, 24]. However, these methods often introduce large training and/or inference overheads compounded by the growing complexity of modern Deep Neural Networks (DNN), ultimately challenging their practicality under real-time computing constraints [7, 22]. Other lines of works such as Conformal Prediction (CP) and Posterior Networks (PostNets) based on evidential deep learning either cannot quantify feature-level epistemic uncertainty or require vast amounts of quality data to accurately estimate the posteriors [1, 5]. Deterministic Uncertainty Methods (DUMs), aiming to efficiently quantify uncertainty using only a single forward pass [10, 34, 42], have become an emerging trend. DUMs rely on statistical/morphological characteristics of latent features to quantify epistemic uncertainty. For instance, [10] employed Prototype Learning (PL) techniques to capture the \"knowledge\" of DNNs by constructing a \"memory bank\" of prototypes to predict uncertainty." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "To address these problems, we propose HyperDUM, a novel deterministic uncertainty method for efficient quantification of feature-level epistemic uncertainty. In contrast to existing methods, HyperDUM projects latent features into" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.875, + 0.483, + 0.9 + ], + "angle": 0, + "content": "*Corresponding Author. Distribution A: Approved for public release; distribution unlimited. OPSEC # 9537" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22306" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.137 + ], + "angle": 0, + "content": "hyperdimensional space to form hyperdimensional prototypes, and then estimates uncertainty based on the distance between features of unseen samples and these prototypes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.138, + 0.486, + 0.306 + ], + "angle": 0, + "content": "To the best of our knowledge, HyperDUM is the first deterministic uncertainty quantification method utilizing hyperdimensional feature prototyping with Channel-wise and Patch-wise projection and bundling to deliver accurate and efficient epistemic uncertainty quantification of latent features. Detailed experimental evaluations on the DeLiVER and aiMotive datasets for semantic segmentation and object detection tasks demonstrate that HyperDUM can efficiently and effectively quantify feature uncertainty and improve fusion performance when compared to standard Bayesian approximation inference baselines and SOTA methods." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.321, + 0.242, + 0.337 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.347, + 0.336, + 0.364 + ], + "angle": 0, + "content": "2.1. Uncertainty Quantification" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.371, + 0.486, + 0.493 + ], + "angle": 0, + "content": "Predictive uncertainty is composed of three factors, model/epistemic uncertainty, data/aleatoric uncertainty, and distributional uncertainty [32]. Epistemic uncertainty, stemming from the word epistemology which is the study of knowledge, is associated with a model's lack of knowledge of underrepresented data. The scope of this work is to efficiently quantify epistemic uncertainty at the latent feature level to enable uncertainty-weighting prior to feature fusion." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.494, + 0.486, + 0.901 + ], + "angle": 0, + "content": "Existing UQ approaches are typically derived from Bayesian principles, including Bayesian neural networks which represent model weights as probability distributions, deep ensembles which sample output distributions from diverse models [24], and Monte Carlo dropout which is similar to deep ensembles but utilizes a single model with dropout layers [11]. However, these approaches remain challenging in practice due to enormous training and inference costs [7, 22]. Conformal prediction (CP) is a model-agnostic UQ technique that provides prediction intervals with guaranteed coverage, ensuring the true value falls within the interval at a specified probability [1]. However, while effective in producing reliable task-level uncertainties, CP does not capture uncertainties at intermediate model abstractions, a limitation especially evident in multimodal fusion models, where uncertainty propagates across different modalities before the final output. Evidential deep learning is another line of work, i.e. Posterior Networks which aim to quantify both aleatoric and epistemic uncertainty by directly estimating a posterior distribution over the model's outputs/predictions but is sensitive to data quality for accurately estimating the posteriors [5]. Deterministic uncertainty quantification is a growing research area that aims to efficiently compute epistemic uncertainty of DNNs with a single forward pass [10, 34, 42]. For instance, [34] is a training-free method that injects Gaussian noise into intermediate layers to compute the variance across the representations as a measure of uncertainty. La" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.909, + 0.289 + ], + "angle": 0, + "content": "tent Deterministic Uncertainty (LDU) constructs prototypes from latent features by enforcing dissimilarity between prototypes and entropy between latent features while correlating uncertainty and the downstream task [10, 36, 46]. Recent research [37] demonstrated that constructing hyperdimensional prototypes can provide comparable performance to existing aleatoric uncertainty quantification solutions, e.g., Bayesian approximation, and offer significant speedups in training and inference. However, it is applied to regression tasks for aleatoric uncertainty. In contrast, our work quantifies epistemic uncertainty on the latent features, extending its applicability to multimodal fusion models and complex tasks such as object detection and semantic segmentation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.298, + 0.849, + 0.314 + ], + "angle": 0, + "content": "2.2. Multimodal Uncertainty-Aware Fusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.321, + 0.909, + 0.563 + ], + "angle": 0, + "content": "Multimodal sensor fusion aims to provide more comprehensive representations and address the limitations of unimodal sensor failure cases [8, 31, 56]. Existing approaches dealing with uncertainties fall into two categories: feature-level fusion and output-level fusion. Feature-level fusion, also known as intermediate fusion, applies an uncertainty weighting that scales the features to minimize uncertainty [30, 47]. For instance, [14] computes feature-informativeness and scales the original features based on informativeness. On the other hand, output-level fusion, also known as late-fusion, uses uncertainty to calibrate the outputs of the model, i.e. calibrating softmax probabilities [40]. In contrast to existing uncertainty-based fusion works that commonly rely on Bayesian approximated uncertainty fusion, our proposed HyperDUM is the first to investigate deterministic uncertainty fusion methods in a multimodal autonomous systems setting." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.576, + 0.65, + 0.594 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.601, + 0.655, + 0.617 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.624, + 0.909, + 0.867 + ], + "angle": 0, + "content": "Given an input sample \\(\\mathbf{x} = \\{x_{1}, x_{2}, \\dots, x_{M}\\}\\) consisting of \\(M\\) multimodal sensor inputs. We have a multimodal model \\(\\mathbb{M}\\) with pre-trained feature encoders \\(\\mathbf{f} = \\{f_{1}, \\dots, f_{M}\\}\\) that extract modality-specific abstract features \\(z_{m} = f_{m}(x_{m})\\) where \\(1 \\leq m \\leq M\\). Additionally, the model contains fusion blocks \\(\\mathcal{F}(z_{i}, z_{j})\\) where \\(i, j \\in \\{1, \\dots, M\\}\\) such that \\(i \\neq j\\) which combine modality-specific features to learn cross-modal fusion features \\(g\\). Assuming there are \\(K\\) fusion blocks, the fused features \\(\\mathbf{g} = \\{g_{1}, \\dots, g_{K}\\}\\) are passed to the task head \\(h(g)\\) to output the corresponding task predictions \\(y\\) (i.e. bounding boxes and object classification for object detection, and pixel-wise class labels for semantic segmentation). To account for feature uncertainty, an uncertainty quantification module \\(\\mathbb{U}_{m}(z_{m})\\) is inserted immediately following each feature extractor to predict an uncertainty value \\(u_{m}\\). Given the uncertainty value, a learnable" + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.875, + 0.909, + 0.9 + ], + "angle": 0, + "content": "1Cross-modal feature alignment is not required but may be enforced by the backbone." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22307" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.095, + 0.898, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.13, + 0.317, + 0.866, + 0.332 + ], + "angle": 0, + "content": "Figure 1. Multimodal model with uncertainty quantification and uncertainty fusion for autonomous driving perception tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.358, + 0.484, + 0.48 + ], + "angle": 0, + "content": "uncertainty weighting module \\(\\Omega (z_m,u_m)\\) reweights the input features to produce an uncertainty-aware feature \\(\\hat{z}_m\\) for each modality. Thus, the input to the fusion blocks now becomes \\(\\mathcal{F}(\\hat{z}_i,\\hat{z}_j)\\) to produce uncertainty-weighted fusion features \\(\\hat{\\mathbf{g}}\\) which are forwarded to be fused and subsequently used by the task head \\(h(\\hat{\\mathbf{g}})\\) to produce the outputs. The system flow of the described multimodal fusion model with uncertainty quantification is illustrated in Figure 1." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.499, + 0.484, + 0.53 + ], + "angle": 0, + "content": "3.1.1 Vector Symbolic Architectures/Hyperdimensional Computing" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.538, + 0.485, + 0.84 + ], + "angle": 0, + "content": "Vector Symbolic Architectures (VSA) / Hyperdimensional Computing (HDC) are a class of algorithms for creating and handling hyperdimensional representations to mimic cognitive functions [25, 45]. VSA starts with mapping input data to hyperdimensional vectors, or hypervectors \\(\\mathcal{H}\\), through a projection function \\(\\phi: \\mathcal{Z} \\to \\mathbf{H}\\), where \\(\\mathcal{Z} \\in \\mathbb{R}^n\\) and \\(\\mathbf{H} \\in \\mathbb{R}^d\\) (\\(d \\gg n\\)) is the hypervector space. The projection function depends on the input data representation. This work assumes Euclidean input space. For non-euclidean see [38]. Following the projection, common operations on hypervectors include similarity calculation, bundling, and binding. Similarity \\((\\delta(\\cdot, \\cdot))\\) calculation measures the distance between two hypervectors. For real-valued hypervectors, a common measure is cosine similarity. Bundling \\((\\bigoplus)\\) is an element-wise addition of hypervectors, e.g., \\(\\mathcal{H}^{bundle} = \\mathcal{H}^1 \\bigoplus \\mathcal{H}^2\\), generating a hypervector with the same dimension as inputs. Binding \\((\\otimes)\\) is an element-wise multiplication associating two hypervectors to create another near-orthogonal hypervector, e.g., \\(\\mathcal{H}^{bind} = \\mathcal{H}^1 \\bigotimes \\mathcal{H}^2\\) where \\(\\delta(\\mathcal{H}^{bind}, \\mathcal{H}^1) \\approx 0\\) and \\(\\delta(\\mathcal{H}^{bind}, \\mathcal{H}^2) \\approx 0\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Our work mainly leverages the bundling operation of VSA for the formation of prototypes. In high-dimensional space, bundling mimics how human brains memorize information [48, 57]. For instance, for \\(\\mathcal{H}^{bundle} = \\mathcal{H}^1 +\\mathcal{H}^2\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.356, + 0.908, + 0.492 + ], + "angle": 0, + "content": "we have \\(\\delta (\\mathcal{H}^{bundle},\\mathcal{H}^1)\\gg 0\\) while \\(\\delta (\\mathcal{H}^{bundle},\\mathcal{H}^3)\\approx 0\\) \\((\\mathcal{H}^3\\neq \\mathcal{H}^1,\\mathcal{H}^2)\\) . In other words, VSA can generate prototypical representations by bundling the high-dimensional vectors of individual data samples, making it especially useful for prototype learning. Specifically, suppose we have labeled samples \\(\\mathcal{D} = \\{(z_1^1,y_1^1),(z_1^2,y_1^2),\\ldots ,(z_M^N,y_M^N)\\}\\) where \\(z_{m}^{i}\\in \\mathcal{Z}\\) and \\(y_{m}^{i}\\in \\{l\\}_{l = 1}^{L}\\) , we can form \\(L\\) prototypes by bundling hypervectors corresponding to a particular label:" + }, + { + "type": "equation", + "bbox": [ + 0.637, + 0.491, + 0.907, + 0.524 + ], + "angle": 0, + "content": "\\[\n\\mathcal {H} _ {m} ^ {l} = \\bigoplus_ {i: y ^ {i} = l} \\phi_ {m} \\left(z _ {m} ^ {i}\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.527, + 0.909, + 0.619 + ], + "angle": 0, + "content": "where each hyperdimensional prototype can represent an aspect of the data samples \\(\\mathcal{D}\\) depending on the label definition, i.e. class label (cat and dog) or context label (weather conditions which we use in this work). We then define the notion of uncertainty as the set of similarity between hypervectors and hyperdimensional prototype of modality \\(m\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.566, + 0.638, + 0.907, + 0.679 + ], + "angle": 0, + "content": "\\[\n\\mathbb {U} _ {m} = \\bigcup_ {l = 1} ^ {L} \\left\\{\\delta \\left(\\mathcal {H} _ {m} ^ {z}, \\mathcal {H} _ {m} ^ {l}\\right) \\right\\}, \\text {w h e r e} | \\mathbb {U} _ {m} | = L \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.686, + 0.909, + 0.807 + ], + "angle": 0, + "content": "where \\(\\mathbb{U}_m\\) is the set of similarity distances between a hypervector \\(\\mathcal{H}_m^z\\) and each hyperdimensional prototype \\(\\mathcal{H}_m^l\\), \\(l = 1,2,\\dots ,L\\) and used to express the similarity uncertainty. The notion of similarity for uncertainty has been used in other prototype learning methods [10, 26], where the former assigns the variations of similarities between an instance and the prototypes as the belief masses and the latter utilizes similarity to optimize the uncertainty loss." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.815, + 0.63, + 0.831 + ], + "angle": 0, + "content": "3.2. HyperDUM" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.838, + 0.908, + 0.869 + ], + "angle": 0, + "content": "Following the conventions of VSA for projection and bundling, we propose two modifications for uncertainty quan-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.512, + 0.875, + 0.908, + 0.901 + ], + "angle": 0, + "content": "2We derive uncertainty quantifiability from similarity/expressivity of VSAs/HDCs based on recent theorems in the Appendix." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22308" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.198 + ], + "angle": 0, + "content": "tification: 1) Channel-wise and 2) Patch-wise Projection & Bundling (CPB/PPB). CPB modifies the projection and bundling operations to capture the uncertainty of each channel per modality separately, whereas PPB allows the VSA to quantify the spatial uncertainty without losing fine-grained spatial information through patching. Figure 1 visualizes the CPB and PPB operations of the proposed method." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.215, + 0.45, + 0.231 + ], + "angle": 0, + "content": "3.2.1 Channel-wise Projection & Bundling (CPB)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.238, + 0.485, + 0.39 + ], + "angle": 0, + "content": "Latent features are generally multi-channeled. By convention prototype formation based on Equation 1 projects and bundles all feature dimensions onto a common representation space \\(\\mathbf{H}\\) [52, 55]. Given latent feature \\(z_{m}^{i} \\in \\mathbb{R}^{C \\times H \\times W}\\) with \\(C\\) denoting feature channels and \\(H, W\\) the spatial dimensions, adaptive pooling (max or average) is first used to reduce the spatial dimensions to \\(z_{m}^{pooled} \\in \\mathbb{R}^{C}\\). Through matrix multiplication and proper dimension initialization of the projection matrix \\(\\Phi\\), the correct hypervector dimensions can be obtained for \\(\\mathcal{H}_{m}^{z}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.395, + 0.484, + 0.415 + ], + "angle": 0, + "content": "\\[\n\\phi_ {m} \\left(z _ {m} ^ {i}\\right) = \\Phi^ {(d \\times C)} \\cdot z _ {m} ^ {\\text {p o o l e d} (C)} = \\mathcal {H} _ {m} ^ {z} ^ {(d)} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.421, + 0.485, + 0.618 + ], + "angle": 0, + "content": "Another method to handle multi-channels inputs is by projecting each feature dimension separately followed by binding and subsequent bundling to a common space [48]. However, we note that the above operations neglect the fact that different feature channels learn different aspects of the input and may contribute more or less to the overall uncertainty. For example, [27] showed that sharing the same uncertain distribution among different channels is less effective for out-of-distribution generalization, while considering each channel uncertainty separately brings better performances due to the different channel potentials. Therefore, instead of absorbing the channel dimensions during projection and bundling, we retain them as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.623, + 0.484, + 0.643 + ], + "angle": 0, + "content": "\\[\n\\phi_ {m} \\left(z _ {m} ^ {i}\\right) = \\Phi^ {(d \\times C)} \\otimes z _ {m} ^ {\\text {p o o l e d} (C)} = \\mathcal {H} _ {m} ^ {z} ^ {(d \\times C)} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.648, + 0.485, + 0.707 + ], + "angle": 0, + "content": "where we use the Einstein summation notation \\(\\otimes\\) (not the binding \\(\\otimes\\) notation) to retain the channel dimension. Consequently, the bundling operation is modified to account for the channel dimensions:" + }, + { + "type": "equation", + "bbox": [ + 0.21, + 0.712, + 0.484, + 0.756 + ], + "angle": 0, + "content": "\\[\n\\mathcal {H} _ {m} ^ {l} = \\bigoplus_ {i: y ^ {i} = l} ^ {d i m = C} \\phi_ {m} \\left(z _ {m} ^ {i}\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.761, + 0.487, + 0.822 + ], + "angle": 0, + "content": "where \\(dim = C\\) indicates the dimension over which the element-wise addition is performed and \\(\\mathcal{H}_m^l\\) is now \\(\\mathbb{R}^{d\\times C}\\). Additionally, the similarity uncertainty now outputs similarities for each channel dimension by representing Equation 2:" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.839, + 0.484, + 0.897 + ], + "angle": 0, + "content": "\\[\n\\mathbb {U} _ {m} ^ {C} = \\bigcup_ {c = 1} ^ {C} \\left\\{\\bigcup_ {l = 1} ^ {L} \\left\\{\\delta \\left(\\mathcal {H} _ {m} ^ {z} [ c ], \\mathcal {H} _ {m} ^ {l} [ c ]\\right) \\right\\} \\right\\}, \\text {s . t .} | \\mathbb {U} _ {m} ^ {C} | = C \\times L \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.152 + ], + "angle": 0, + "content": "We empirically demonstrate through experiments that channel projection and bundling outperforms conventional projection and bundling for the task of multimodal uncertainty-aware feature fusion." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.17, + 0.853, + 0.185 + ], + "angle": 0, + "content": "3.2.2 Patch-wise Projection & Bundling (PPB)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.193, + 0.908, + 0.345 + ], + "angle": 0, + "content": "Channel projection and bundling enables our method to capture the holistic uncertainties of each feature channel. However, to capture the finer-granularity uncertainties of the spatial dimensions requires a different approach. By convention spatial dimensions are either pooled to a single value and projected to high dimensions as shown by Equation 3 or each spatial dimension is separately projected to high dimensions. The former loses spatial information due to pooling and the latter is computationally infeasible and memory intensive given the large spatial dimensions of deep neural networks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.346, + 0.909, + 0.499 + ], + "angle": 0, + "content": "Inspired by the idea of using image patches to capture spatial information in anomaly detection and image classification [6, 44], we propose PPB which projects and bundles spatial patches of latent features to hyperdimensional space to capture the spatial uncertainties. Specifically, given latent feature \\(z_{m}^{i}\\in \\mathbb{R}^{C\\times H\\times W}\\), we slice \\((H,W)\\) into \\(P\\) patches of resolution \\((\\tilde{h},\\tilde{w})\\) where \\(\\tilde{h} = H / \\sqrt{(P)}\\), \\(\\tilde{w} = W / \\sqrt{(P)}\\). This gives us the set of patched features \\(\\tilde{z}_{m}^{i} = \\{\\tilde{z}_{m}^{i,j}\\}_{j = 1}^{P}\\) where each \\(\\tilde{z}_{m}^{i,j}\\in \\mathbb{R}^{C\\times \\tilde{h}\\times \\tilde{w}}\\) are pooled to \\(\\mathbb{R}^C\\) and form the set of pooled patched features \\(\\tilde{z}_{m}^{pooled}\\in \\mathbb{R}^{C\\times P}\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.5, + 0.908, + 0.531 + ], + "angle": 0, + "content": "Given patched features \\(\\tilde{z}_m^{pooled}\\), we project them according to Equation 3 to form the set of patched projections:" + }, + { + "type": "equation", + "bbox": [ + 0.551, + 0.551, + 0.907, + 0.573 + ], + "angle": 0, + "content": "\\[\n\\phi_ {m} \\left(\\tilde {z} _ {m} ^ {i}\\right) = \\Phi^ {(d \\times C)} \\cdot \\tilde {z} _ {m} ^ {\\text {p o o l e d} (C \\times P)} = \\mathcal {H} _ {m} ^ {\\tilde {z}} ^ {(d \\times P)} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.908, + 0.628 + ], + "angle": 0, + "content": "Consequently, these projections are bundled for each patch according to Equation 1 to produce the set of patched prototypes as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.633, + 0.639, + 0.907, + 0.683 + ], + "angle": 0, + "content": "\\[\n\\mathcal {H} _ {\\tilde {m}} ^ {l} = \\bigoplus_ {i: y ^ {i} = l} ^ {d i m = P} \\phi_ {m} \\left(\\tilde {z} _ {m} ^ {i}\\right) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.689, + 0.908, + 0.735 + ], + "angle": 0, + "content": "where \\(\\mathcal{H}_{\\tilde{m}}^{l}\\) represents the patched prototypes of modality \\(m\\) , for each patch and for each context label \\(i\\) , making \\(|\\mathcal{H}_{\\tilde{m}}^{l}| = d\\times P\\)" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.735, + 0.908, + 0.767 + ], + "angle": 0, + "content": "Finally, we can obtain the spatial uncertainties for each patch of the input projection \\(\\mathcal{H}_m^{\\tilde{z}}\\) by modifying Equation 2:" + }, + { + "type": "equation", + "bbox": [ + 0.524, + 0.787, + 0.907, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\mathbb {U} _ {\\tilde {m}} ^ {P} = \\bigcup_ {p = 1} ^ {P} \\left\\{\\bigcup_ {l = 1} ^ {L} \\left\\{\\delta \\left(\\mathcal {H} _ {\\tilde {m}} ^ {\\tilde {z}} [ p ], \\mathcal {H} _ {\\tilde {m}} ^ {l} [ p ]\\right) \\right\\} \\right\\}, \\text {s . t .} | \\mathbb {U} _ {\\tilde {m}} ^ {P} | = P \\times L \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.91, + 0.901 + ], + "angle": 0, + "content": "Our final proposed approach combines both channel and patch methods for projection and bundling to capture channel-wise feature uncertainties, as well as finer" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22309" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.123 + ], + "angle": 0, + "content": "granularity spatial feature uncertainties, to provide a holistic epistemic uncertainty quantification of each modality." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.136, + 0.223, + 0.154 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.161, + 0.281, + 0.178 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.185, + 0.204, + 0.199 + ], + "angle": 0, + "content": "4.1.1 Datasets" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.209, + 0.486, + 0.466 + ], + "angle": 0, + "content": "We evaluate the effectiveness of HyperDUM for uncertainty-aware feature fusion using two multimodal autonomous driving datasets. The aiMotive dataset [33] includes camera, lidar, and radar sensors for long-distance \\((\\geq 75m)\\) 3D object detection and features diverse scenes across various locations, times, and weather conditions. The DeLiVER dataset [58] contains camera, lidar, event, and depth sensors for semantic segmentation, including adverse corner cases such as motion blur and lidar jitter. Additionally, we integrate corner cases into the aiMotive dataset by applying various lighting effects to camera sensors, causing reduced clarity and scene obstruction. These effects are created through image transformations such as GaussianBlur and exposure adjustments [39, 50]. We also implement foggification [3] to simulate synthetic fogging of lidar data, resulting in noisy point clouds with reduced field of view. These corner case effects can be visualized in Figure 2. Details in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.485, + 0.196, + 0.5 + ], + "angle": 0, + "content": "4.1.2 Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.51, + 0.485, + 0.737 + ], + "angle": 0, + "content": "We utilize the pretrained BEVFusion [33] and CMNeXT [58] models as the fusion architectures for aiMotive and DeLiVER, respectively. BEVFusion is built upon VoxelNet [59] and BEVDepth [29] to fuse camera and lidar representations on a unified BEV space. CMNeXT is a four-stage pyramidal fusion model that treats the RGB representation as the primary branch and other modalities as the secondary branch. We integrate the uncertainty quantification module (UQM) at each model's respective pre-fusion points. For aiMotive this is chosen to be at the feature concatenation step, right before the bev Fuse step. For DeLiVER, we insert the UQM immediately after the feature rectification module and right before the feature fusion module. Given that the CMNeXT model has four stages, we instantiate one UQM per stage. Details in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.756, + 0.267, + 0.77 + ], + "angle": 0, + "content": "4.1.3 Baseline Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.485, + 0.901 + ], + "angle": 0, + "content": "We implemented different UQ methods and compared their performances against our method. For each implementation, we followed their open-sourced code implementation and hyperparameter settings. Specifically, for Infer-dropout (InfMCD) and infer-noise (InfNoise), we followed their suggestions and chose 10 forward passes and used the variance of the outputs for uncertainty estimation. In addition to variance, we computed predictive entropy, mutual information," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.909, + 0.44 + ], + "angle": 0, + "content": "and entropy as additional uncertainty estimates following [49]. It was shown that using more uncertainty metrics can provide more effective quantification for different types of uncertainties. For Posterior Network (PostNet), we constructed PostNets for each modality so that they can estimate the modality-specific uncertainty independently and be used in the uncertainty-aware feature fusion process. The use of multiple PostNets is similar to [20] in their construction of separate PostNets for agent's past behavior, road structure map, and social context respectively. For latent deterministic uncertainty (LDU), we set the number of learnable prototypes to the number of scenarios. Specifically, for aiMootive, the scenarios are Highway, Urban, Night, and Rain, while for DeLiVER, the scenarios are Cloudy, Foggy, Night, Rainy, and Sunny. Finally, for all methods, all layers prior to the UQM are kept frozen and we fine-tune the uncertainty weighting layer, which takes the respective uncertainty metrics as input, along with other post-fusion layers, to enable the model to learn the uncertainty-weighted features. Details of the fine-tuning procedures and visualizations of the architectures and UQM insertion points are provided in the Appendix. In all of our experiments the best one is highlighted and bolded and the 2nd best is lightly-highlighted." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.448, + 0.907, + 0.48 + ], + "angle": 0, + "content": "4.2. Multimodal Uncertainty Fusion for Autonomous Vehicles Perception" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.486, + 0.71, + 0.501 + ], + "angle": 0, + "content": "4.2.1 3D Object Detection" + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.521, + 0.907, + 0.597 + ], + "angle": 0, + "content": "
UQ MethodHighwayUrbanNightRainMean
BEVFusion [33]72.55/69.0863.72/63.5274.74/73.1342.75/42.8365.67/64.79
InfMCD [34]72.58/69.6562.02/59.5975.05/73.7235.58/37.4164.56/64.58
InfNoise [34]70.11/68.5263.10/60.4275.08/74.1935.58/37.4164.59/64.56
PostNet [5]69.19/68.0261.38/59.8473.32/68.9734.95/37.6063.13/60.21
LDU [10]71.48/69.3362.54/60.1176.49/75.0440.49/41.2464.69/64.73
HyperDUM72.23/69.5864.77/64.4876.69/75.1544.78/45.4866.70/66.00
" + }, + { + "type": "table_caption", + "bbox": [ + 0.558, + 0.608, + 0.86, + 0.621 + ], + "angle": 0, + "content": "Table 1. Results on aiMotive under diverse scenes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.639, + 0.909, + 0.73 + ], + "angle": 0, + "content": "Table 1 compares the all-point/11-point interpolation Average Precision (AP) metric for object detection under diverse scenes. It can be seen that apart from the Highway scenario, HyperDUM outperforms all methods across all scenarios. Specifically, HyperDUM improves upon the all-point AP by \\(2.01\\%\\) on average against the state-of-the-art (LDU)." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.741, + 0.906, + 0.816 + ], + "angle": 0, + "content": "
UQ MethodMBOEUELFMean
BEVFusion [33]64.10/63.6962.40/59.2663.57/63.2565.07/65.0863.79/62.75
InfMCD [34]62.00/61.9964.47/64.2863.25/63.4964.27/63.8163.50/63.39
InfNoise [34]62.61/62.6164.94/64.7263.18/60.0765.19/64.9963.98/63.10
PostNet [5]60.42/58.9063.12/59.9960.20/59.2564.13/60.6061.97/59.69
LDU [10]62.06/62.2265.07/64.8562.87/60.0265.47/65.1063.87/63.05
HyperDUM64.39/63.9966.16/65.3964.18/63.9165.89/65.2265.16/64.62
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.829, + 0.907, + 0.87 + ], + "angle": 0, + "content": "Table 2. Results on aiMotive under corner cases: Motion Blur (MB), Over-Exposure (OE), Under-Exposure (UE), LiDAR-Fog (LF). (Metrics: all-point AP/11-point interpolation AP)" + }, + { + "type": "table_caption", + "bbox": [ + 0.534, + 0.886, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Table 2 demonstrates the robustness of UQ methods un" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "22310" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.089, + 0.907, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.276, + 0.225, + 0.325, + 0.236 + ], + "angle": 0, + "content": "Original" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.237, + 0.297, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.172, + 0.338, + 0.22, + 0.351 + ], + "angle": 0, + "content": "Original" + }, + { + "type": "image", + "bbox": [ + 0.301, + 0.238, + 0.498, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.365, + 0.339, + 0.436, + 0.35 + ], + "angle": 0, + "content": "Motion Blur" + }, + { + "type": "image_caption", + "bbox": [ + 0.663, + 0.225, + 0.739, + 0.237 + ], + "angle": 0, + "content": "Foggification" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.238, + 0.702, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.562, + 0.339, + 0.647, + 0.351 + ], + "angle": 0, + "content": "Over-Exposure" + }, + { + "type": "image", + "bbox": [ + 0.705, + 0.238, + 0.903, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.761, + 0.339, + 0.853, + 0.351 + ], + "angle": 0, + "content": "Under-Exposure" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.368, + 0.908, + 0.397 + ], + "angle": 0, + "content": "Figure 2. Corner cases for the aiMotive 3D Object Detection dataset. (Top-Left) Normal Lidar point-cloud birds-eye-view. (Top-Right) Lidar Foggification (LF). (Bottom) In order from left to right (Normal, Motion Blur (MB), Over-Exposure (OE), Under-Exposure (UE))" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.423, + 0.484, + 0.543 + ], + "angle": 0, + "content": "der our injected corner cases. These injections were unseen during training which complicates testing due to distribution shifts. Each corner case is applied to and evaluated on the entire test set across all scenarios. Whereas other methods fail and suffer performance degradations, HyperDUM only degrades \\(0.51\\% / 0.17\\%\\) compared to the baseline w/o injections demonstrating that HyperDUM is more robust even when facing distribution shifts." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.587, + 0.484, + 0.663 + ], + "angle": 0, + "content": "
UQ MethodHighwayUrbanNightRainMean
BEVFusion [33]45.59/45.3446.38/45.5344.49/45.8742.65/42.9345.84/46.87
InfMCD [34]45.57/46.2938.90/38.1142.07/41.9726.13/28.4940.74/41.64
InfNoise [34]40.73/40.3842.41/40.9038.80/40.0029.28/32.8140.91/42.04
PostNet [5]41.13/40.6439.47/40.1640.37/41.9228.56/28.9639.37/41.66
LDU [10]43.07/45.0140.17/41.7142.90/42.5231.12/34.0840.65/42.17
HyperDUM44.86/46.0647.89/48.9144.42/46.1240.54/41.6346.52/48.27
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.674, + 0.483, + 0.703 + ], + "angle": 0, + "content": "Table 3. Results on aiMotive under diverse scenes in the distant region (\\(>75\\mathrm{m}\\)). (Metrics: all-point AP/11-point interpolation AP)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Table 3 is an experiment on the distant region setting, keeping only predictions and ground truths over \\(\\geq 75m\\). As expected, all UQ methods experience major performance drops, especially for Rain. This is likely due to the combined uncertainties of distance with rain obstruction causing substantial minimization of the uncertainty-weighted features (the weighting output is between 0-1 times the original feature value). The original BEVFusion model maintains better performance because it is still able to \"guess\" the presence of objects as the features are not minimized by uncertainty." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.42, + 0.907, + 0.587 + ], + "angle": 0, + "content": "
ScenariosUQ Methods (Metrics: mIoU ↑)
CMNeXt [58]InfMCD [34]InfNoise [34]PostNet [5]LDU [10]Gemini [21]HyperDUM (Ours)
Cloudy68.7069.2369.2169.2868.94-69.76+1.06
Foggy65.6666.1866.1066.0465.72-66.85+1.19
Night62.4663.4463.1462.9763.06-64.21+1.75
Rainy67.5068.0868.0968.1667.82-68.71+1.21
Sunny66.5767.0167.1666.9366.75-67.87+1.30
MB62.9163.6163.5563.5563.16-64.28+1.37
OE64.5965.3965.1765.0664.73-65.67+1.08
UE60.0060.3860.4260.2760.29-61.20+1.20
LJ65.9266.1266.2566.3366.40-66.93+1.01
EL65.4866.0566.1766.0665.89-66.80+1.32
Mean66.3066.9066.8666.7766.6066.9067.59+1.29
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.599, + 0.908, + 0.641 + ], + "angle": 0, + "content": "Table 4. DeLiVER mean Intersection over Union (mIoU) performance under adverse weather and corner cases. Lidar-Jitter (LJ), Event Low-resolution (EL)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.674, + 0.733, + 0.69 + ], + "angle": 0, + "content": "4.2.2 Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.704, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Table 4 shows the DeLiVER validation set performance on different scenarios and corner cases for the semantic segmentation task. Overall, HyperDUM achieves the best performance over all methods across every weather and corner case scenario. On average HyperDUM improves over the baseline by 1.29. In particular, for the three most difficult scenarios including Under-Exposure, Night and Motion Blur, HyperDUM achieves at least 1.00 improvement over the baseline. For GeminiFusion [21], we only report the mean as they omitted the detailed performance breakdown in their paper. These results demonstrate that our method can better quantify the multimodal feature uncertainties and ultimately improve the fusion features for the semantic segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22311" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.152 + ], + "angle": 0, + "content": "task. Additionally, we note that all UQ methods improved over the baseline method. This shows that incorporating UQ with fusion into multimodal pipelines not only enables UQ but also can improve baseline performance." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.164, + 0.487, + 0.332 + ], + "angle": 0, + "content": "
ScenariosUQ Methods (Metrics: ECE ↓)
CMNeXt [58]InfMCD [34]InfNoise [34]PostNet [5]LDU [10]HyperDUM (Ours)
Cloudy1.18E-021.05E-021.07E-020.99E-021.09E-020.94E-02-0.24
Foggy1.67E-021.48E-021.52E-021.45E-021.56E-021.35E-02-0.30
Night1.88E-021.63E-021.68E-021.71E-021.68E-021.39E-02-0.49
Rainy1.53E-021.37E-021.39E-021.33E-021.43E-021.21E-02-0.32
Sunny1.42E-021.25E-021.27E-021.24E-021.33E-021.12E-02-0.30
MB1.54E-021.32E-021.40E-021.32E-021.44E-021.19E-02-0.35
OE1.49E-021.55E-021.39E-021.42E-021.44E-021.31E-02-0.18
UE1.75E-022.07E-021.69E-021.84E-021.63E-021.53E-02-0.22
LJ1.54E-021.38E-021.39E-021.35E-021.40E-021.28E-02-0.26
EL1.45E-021.17E-021.27E-021.26E-021.27E-021.09E-02-0.36
Mean1.53E-021.35E-021.38E-021.34E-021.41E-021.20E-02-0.33
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.343, + 0.484, + 0.371 + ], + "angle": 0, + "content": "Table 5. DeLiVER Expected Calibration Error (ECE) performance under adverse weather and corner cases." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.388, + 0.486, + 0.6 + ], + "angle": 0, + "content": "Table 5 shows the Expected Calibration Error (ECE) metric for all methods across the same scenarios. ECE measures how well a model is calibrated by comparing it's predicted probabilities with the actual probabilities of the ground truth distribution. ECE helps us understand how reliable a model's confidence scores are. A higher ECE means the model is more overconfident (or unsure) about its predictions. As shown, HyperDUM has the lowest ECE score over all methods across every weather scenario and corner case with a 0.33 average decrease compared to the baseline. These results demonstrate that our method helps better calibrate the overconfident semantic segmentation model. In fact all UQ methods on average has a lower ECE compared to the baseline demonstrating its usefulness in model calibration." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.61, + 0.454, + 0.626 + ], + "angle": 0, + "content": "4.3. Pre vs. Post Uncertainty Feature Learning" + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.644, + 0.478, + 0.803 + ], + "angle": 0, + "content": "
ScenariosUQ Methods (Metrics: mIoU ↑)
CMNeXt [58]InfMCD [34]InfNoise [34]PostNet [5]LDU [10]HyperDUM (Ours)
Cloudy68.768.7868.569.1369.0669.26
Foggy65.6665.7865.5965.8965.9266.40
Night62.4662.762.4462.8263.0863.71
Rainy67.567.6667.3768.0267.8468.29
Sunny66.5766.5266.3966.7866.7867.43
MB62.9163.1162.7963.4063.363.78
OE64.5964.9464.664.9164.9565.20
UE60.0059.7560.0760.1260.360.72
LJ65.9266.1466.0166.1766.2166.44
EL65.4865.765.6665.9166.0366.34
Mean66.3066.5966.4966.6066.6466.98
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.812, + 0.485, + 0.854 + ], + "angle": 0, + "content": "Table 6. DeLiVER mean Intersection over Union (mIoU) performance under adverse weather and corner cases using post-fusion features." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Table 6 evaluates the performance of uncertainty feature weighting after the fusion block of Figure 1 or see the archi" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.909, + 0.198 + ], + "angle": 0, + "content": "tecture figure in the Appendix. As expected, we found that the overall performance gain is limited/ lower compared to its prefusion uncertainty fusion counterpart. These results align with our motivation that different modalities experience different modality-specific uncertainties. And these uncertainties can propagate to the fusion module, undermining the fusion features and degrading downstream task performance." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.207, + 0.673, + 0.221 + ], + "angle": 0, + "content": "4.4. Model Ablation" + }, + { + "type": "table", + "bbox": [ + 0.536, + 0.238, + 0.884, + 0.298 + ], + "angle": 0, + "content": "
Dataset\nMetricaiMotive\nall-point AP/11-point AP ↑DeLiVER\nmIoU ↑
HyperDUM\n- w/o Patch (4x) Proj.-66.70/66.00 (Mean Δ)67.59 (Mean Δ)
-65.67/65.23 (-1.03/-0.77)-67.18 (-0.41)
- w/o Chan Proj.64.43/64.25 (-1.24/-0.98)66.63 (-0.55)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.551, + 0.307, + 0.866, + 0.321 + ], + "angle": 0, + "content": "Table 7. Model ablation for aiMotive and DeLiVER." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.34, + 0.909, + 0.552 + ], + "angle": 0, + "content": "For our model ablation, we analyzed the effects of sequentially removing our proposed components and the impact on the overall performance. As seen from Table 7 the removal of the Patch Projection (with 4 patches) results in a \\(1.03 / 0.77\\) decrease in performance and the removal of Channel Projection in a \\(1.24 / 0.98\\) decrease in performance for aiMotive. Similarly for the DeLiVER dataset, the removal of the Patch Projection (with 4 patches) results in a 0.41 decrease in performance and removing Channel Projection results in a 0.55 decrease in performance. Together we see that both methods contribute significantly to the overall performance demonstrating that both channel and spatial uncertainties should be accounted for to better capture the holistic uncertainty of the features before fusion. More ablations in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.569, + 0.704, + 0.585 + ], + "angle": 0, + "content": "4.4.1 Sensitivity Analysis" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.61, + 0.897, + 0.781 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.796, + 0.909, + 0.825 + ], + "angle": 0, + "content": "Figure 3. Analysis on the effects of hyperdimension selection for channel projection on performance and computation (DeLiVER)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Here we investigate the tunable parameters for channel and patch projection & bundling (CPB & PPB). For CPB, the only parameter is the channel hyperdimension. For PPB, we have the patch hyperdimension and the number of patches." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22312" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.228 + ], + "angle": 0, + "content": "Figure 3 shows a sweep over a range of dimensions from 625 to \\(20\\mathrm{k}\\) for CPB. This range of values is chosen based on the HDC expressivity limits (around \\(10\\mathrm{k}\\)) [48, 57]. We see that performance improves with dimension increase, but plateaus and decreases past \\(10\\mathrm{k}\\) while computation continues to grow. This occurs when large dimensions induce too much sparsity, making similar samples become orthogonal and capturing irrelevant information likened to over-fitting in neural networks [48]." + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.242, + 0.475, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.402, + 0.485, + 0.431 + ], + "angle": 0, + "content": "Figure 4. Analysis on the effects of patches and patch hyperdimension selection on performance and computation (DeLiVER)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.446, + 0.487, + 0.627 + ], + "angle": 0, + "content": "Figure 4 shows a sweep over 4, 16 and 64 patch configurations together with the dimension sweep from 625 to 5000. We observe an interesting relation where the number of patches negatively correlates with the per-patch hyperdimension in terms of better performance. This makes sense as we increase the number of patches, the information needed to be encoded becomes smaller due to the reduction in the per-patch spatial resolution as indicated in Section 3.2.2. Therefore, the necessary hyperdimensions needed to encode the information should be reduced, while larger dimensions begin to over-fit. Whereas the opposite is true for less patches, larger hyperdimensions are needed." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.634, + 0.27, + 0.65 + ], + "angle": 0, + "content": "4.5. Computation Cost" + }, + { + "type": "table", + "bbox": [ + 0.117, + 0.668, + 0.457, + 0.758 + ], + "angle": 0, + "content": "
DatasetaiMotiveDeLiVER
UQ MethodFLOPs#ParamsFLOPs#Params
InfMCD [34]990.90M4.05K1.84G20.19K
InfNoise [34]990.90M4.05K1.84G20.19K
LDU [10]264.24M44.04M613.42M102.24M
HyperDUM111.84M1.15M338.40M38.85M
- w/o Patch (4x) Proj.84.00M1.13M256.00M38.78M
- w/o Chan Proj.6.96M3.38K20.60M16.83K
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.768, + 0.483, + 0.796 + ], + "angle": 0, + "content": "Table 8. Comparing inference (FLOPs) and training costs (# of Trainable Parameters)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Finally, the main benefit of HyperDUM is its efficiency for training and inference. We measure the computation costs in terms of Floating point operations (FLOPs) and the number of trainable model parameters in Table 8. We approximate the FLOPs by computing the major contributing operations. Refer to the Appendix for our FLOPs approximations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.319 + ], + "angle": 0, + "content": "for each operation type. For InfMCD and InfNoise these are the entropy, mutual information, and predictive entropy computations. For LDU, it is the cosine similarity between prototypes and matrix multiplication for uncertainty, which was found to be negligible. For HyperDUM, it is the input projection and cosine similarity computations. We see that HyperDUM is up to \\(2.36 \\times\\) and \\(1.81 \\times\\) more FLOP-efficient compared to LDU for aiMotive and DeLiVER respectively. Compared to traditional Bayesian approximation methods, it can be up to \\(8.86 \\times\\) and \\(5.44 \\times\\) more FLOP-efficient. As for the training costs, HyperDUM has \\(38.30 \\times\\) and \\(2.63 \\times\\) less trainable parameters compared to LDU for aiMotive and DeLiVER respectively, making it more memory efficient. This ultimately shows the practicality of HyperDUM as a UQ method for real-world autonomous systems." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.332, + 0.771, + 0.348 + ], + "angle": 0, + "content": "5. Discussions and Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.357, + 0.909, + 0.569 + ], + "angle": 0, + "content": "Limitations and Future Work: We leveraged cloud computing resources using a single NVIDIA A100 GPU for training. Our method uses the traditional bundling technique, which requires labels for supervised learning of prototypes. The need for labels is the primary limitation of our approach. Although there have been implementations of semi-supervised and unsupervised versions [17, 18], there are no theoretical guarantees and/or justifications for their performance unlike the traditional bundling method [48]. We leave the exploration of these methods for uncertainty quantification to future work. Additionally, the FLOPs of the vector symbolic architectures may be further reduced due to the massively parallel operations of hyperdimensional computation. These optimizations were demonstrated in prior works [13, 19]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.909, + 0.84 + ], + "angle": 0, + "content": "Conclusion: In this work, we propose an efficient method to quantify the epistemic uncertainty of machine learning models at the feature level. HyperDUM projects the latent features as hypervectors to form hyperdimensional prototypes. Similarity between hypervectors of new samples and learned prototypes is used as a proxy to estimate uncertainty. We demonstrate the practicality of HyperDUM through autonomous driving tasks, including object detection and semantic segmentation. Our results show that our method can enhance the performance and robustness of multimodal fusion models under uncertainties induced by various weather conditions and modality-specific corner cases. We achieve this by quantifying the feature uncertainty of each modality and applying an uncertainty-aware weighting layer prior to feature fusion and fine-tuning the post-fusion layers. Finally, we demonstrated that HyperDUM requires significantly less FLOPs and training parameters when compared to existing works, indicating its real-world practicality." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements: This work acknowledges the support of the Automotive Research Center (ARC), Cooperative Agreement W56HZV-24-2-0001 U.S. Army DEVCOM Ground Vehicle Systems Center (GVSC)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22313" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.484, + 0.156 + ], + "angle": 0, + "content": "[1] Anastasios N Angelopoulos, Stephen Bates, et al. Conformal prediction: A gentle introduction. Foundations and Trends® in Machine Learning, 16(4):494-591, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.158, + 0.484, + 0.255 + ], + "angle": 0, + "content": "[2] Sri Hrushikesh Varma Bhupathiraju, Jennifer Sheldon, Luke A Bauer, Vincent Bindschaedler, Takeshi Sugawara, and Sara Rampazzi. Emi-lidar: Uncovering vulnerabilities of lidar sensors in autonomous driving setting using electromagnetic interference. In Proceedings of the 16th ACM Conference on Security and Privacy in Wireless and Mobile Networks, pages 329-340, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.256, + 0.484, + 0.338 + ], + "angle": 0, + "content": "[3] Mario Bijelic, Tobias Gruber, Fahim Mannan, Florian Kraus, Werner Ritter, Klaus Dietmayer, and Felix Heide. Seeing through fog without seeing fog: Deep multimodal sensor fusion in unseen adverse weather. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11682-11692, 2020. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.34, + 0.483, + 0.409 + ], + "angle": 0, + "content": "[4] Hermann Blum, Paul-Edouard Sarlin, Juan Nieto, Roland Siegwart, and Cesar Cadena. Fishyscapes: A benchmark for safe semantic segmentation in autonomous driving. In proceedings of the IEEE/CVF international conference on computer vision workshops, pages 0–0, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.41, + 0.484, + 0.478 + ], + "angle": 0, + "content": "[5] Bertrand Charpentier, Daniel Zügner, and Stephan Gümnmann. Posterior network: Uncertainty estimation without ood samples via density-based pseudo-counts. Advances in neural information processing systems, 33:1356-1367, 2020. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.48, + 0.484, + 0.521 + ], + "angle": 0, + "content": "[6] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.523, + 0.484, + 0.592 + ], + "angle": 0, + "content": "[7] Michael Dusenberry, Ghassen Jerfel, Yeming Wen, Yian Ma, Jasper Snoek, Katherine Heller, Balaji Lakshminarayanan, and Dustin Tran. Efficient and scalable bayesian neural nets with rank-1 factors. In International conference on machine learning, pages 2782-2792. PMLR, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.593, + 0.484, + 0.648 + ], + "angle": 0, + "content": "[8] Jamil Fayyad, Mohammad A Jaradat, Dominique Gruyer, and Homayoun Najjaran. Deep learning sensor fusion for autonomous vehicle perception and localization: A review. Sensors, 20(15):4220, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.649, + 0.484, + 0.718 + ], + "angle": 0, + "content": "[9] Florian Fervers, Sebastian Bullinger, Christoph Bodensteiner, Michael Arens, and Rainer Stiefelhagen. Uncertainty-aware vision-based metric cross-view geolocation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21621-21631, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.719, + 0.484, + 0.788 + ], + "angle": 0, + "content": "[10] Gianni Franchi, Xuanlong Yu, Andrei Bursuc, Emanuel Aldea, Severine Dubuisson, and David Filliat. Latent discriminant deterministic uncertainty. In European Conference on Computer Vision, pages 243-260. Springer, 2022. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.79, + 0.484, + 0.844 + ], + "angle": 0, + "content": "[11] Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In international conference on machine learning, pages 1050-1059. PMLR, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[12] Ethan Goan and Clinton Fookes. Uncertainty in real-time semantic segmentation on embedded systems. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4490-4500, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.176 + ], + "angle": 0, + "content": "[13] Saransh Gupta, Justin Morris, Mohsen Imani, Ranganathan Ramkumar, Jeffrey Yu, Aniket Tiwari, Baris Aksanli, and Tajana Simunic Rosing. Thrifty: Training with hyperdimensional computing across flash hierarchy. In Proceedings of the 39th International Conference on Computer-Aided Design, pages 1-9, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.179, + 0.907, + 0.247 + ], + "angle": 0, + "content": "[14] Zongbo Han, Fan Yang, Junzhou Huang, Changqing Zhang, and Jianhua Yao. Multimodal dynamics: Dynamical fusion for trustworthy multimodal classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20707-20717, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.249, + 0.907, + 0.332 + ], + "angle": 0, + "content": "[15] Florian Heidecker, Jasmin Breitenstein, Kevin Rösch, Jonas Löhdefink, Maarten Bieshaar, Christoph Stiller, Tim Fingscheidt, and Bernhard Sick. An application-driven conceptualization of corner cases for perception in highly automated driving. In 2021 IEEE Intelligent Vehicles Symposium (IV), pages 644-651. IEEE, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.334, + 0.907, + 0.403 + ], + "angle": 0, + "content": "[16] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17853-17862, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.405, + 0.907, + 0.474 + ], + "angle": 0, + "content": "[17] Mohsen Imani, Samuel Bosch, Mojan Javaheripi, Bita Rouhani, Xinyu Wu, Farinaz Koushanfar, and Tajana Rosing. Semihd: Semi-supervised learning using hyperdimensional computing. In 2019 IEEE/ACM International Conference on Computer-Aided Design (ICCAD), pages 1-8. IEEE, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.476, + 0.907, + 0.546 + ], + "angle": 0, + "content": "[18] Mohsen Imani, Yeseong Kim, Thomas Worley, Saransh Gupta, and Tajana Rosing. Hdcluster: An accurate clustering using brain-inspired high-dimensional computing. In 2019 Design, Automation & Test in Europe Conference & Exhibition (DATE), pages 1591-1594. IEEE, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.548, + 0.907, + 0.63 + ], + "angle": 0, + "content": "[19] Mohsen Imani, Saikishan Pampana, Saransh Gupta, Minxuan Zhou, Yeseong Kim, and Tajana Rosing. Dual: Acceleration of clustering algorithms using digital-based processing in-memory. In 2020 53rd Annual IEEE/ACM International Symposium on Microarchitecture (MICRO), pages 356-371. IEEE, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.632, + 0.907, + 0.686 + ], + "angle": 0, + "content": "[20] Masha Itkina and Mykel Kochenderfer. Interpretable self-aware neural networks for robust trajectory prediction. In Conference on Robot Learning, pages 606-617. PMLR, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.689, + 0.907, + 0.746 + ], + "angle": 0, + "content": "[21] Ding Jia, Jianyuan Guo, Kai Han, Han Wu, Chao Zhang, Chang Xu, and Xinghao Chen. Geminifusion: Efficient pixelwise multimodal fusion for vision transformer. International Conference on Machine Learning (ICML), 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.747, + 0.907, + 0.829 + ], + "angle": 0, + "content": "[22] Xiaotao Jia, Jianlei Yang, Runze Liu, Xueyan Wang, Sorin Dan Cotofana, and Weisheng Zhao. Efficient computation reduction in bayesian neural networks through feature decomposition and memorization. IEEE transactions on neural networks and learning systems, 32(4):1703-1712, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[23] Yongchan Kwon, Joong-Ho Won, Beom Joon Kim, and Myunghee Cho Paik. Uncertainty quantification using bayesian neural networks in classification: Application to biomedical image segmentation. Computational Statistics & Data Analysis, 142:106816, 2020. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22314" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.148 + ], + "angle": 0, + "content": "[24] Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scalable predictive uncertainty estimation using deep ensembles. Advances in neural information processing systems, 30, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.15, + 0.484, + 0.22 + ], + "angle": 0, + "content": "[25] Simon D Levy and Ross Gayler. Vector symbolic architectures: A new building material for artificial general intelligence. In Proceedings of the 2008 conference on artificial general intelligence 2008: Proceedings of the first AGI conference, pages 414-418, 2008. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.483, + 0.276 + ], + "angle": 0, + "content": "[26] Hao Li, Jingkuan Song, Lianli Gao, Xiaosu Zhu, and Hengtao Shen. Prototype-based aleatoric uncertainty quantification for cross-modal retrieval. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.483, + 0.331 + ], + "angle": 0, + "content": "[27] Xiaotong Li, Yongxing Dai, Yixiao Ge, Jun Liu, Ying Shan, and Ling-Yu Duan. Uncertainty modeling for out-of-distribution generalization. arXiv preprint arXiv:2202.03958, 2022.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.483, + 0.389 + ], + "angle": 0, + "content": "[28] You Li and Javier Ibanez-Guzman. Lidar for autonomous driving: The principles, challenges, and trends for automotive lidar and perception systems. IEEE Signal Processing Magazine, 37(4):50-61, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.483, + 0.461 + ], + "angle": 0, + "content": "[29] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1477-1485, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.483, + 0.517 + ], + "angle": 0, + "content": "[30] Ming Liang, Bin Yang, Shenlong Wang, and Raquel Urtasun. Deep continuous fusion for multi-sensor 3d object detection. In Proceedings of the European conference on computer vision (ECCV), pages 641-656, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.483, + 0.602 + ], + "angle": 0, + "content": "[31] Arnav Vaibhav Malawade, Trier Mortlock, and Mohammad Abdullah Al Faruque. Hydraulics: Context-aware selective sensor fusion for robust and efficient autonomous vehicle perception. In 2022 ACM/IEEE 13th International Conference on Cyber-Physical Systems (ICCPs), pages 68–79. IEEE, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.483, + 0.646 + ], + "angle": 0, + "content": "[32] Andrey Malinin and Mark Gales. Predictive uncertainty estimation via prior networks. Advances in neural information processing systems, 31, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.483, + 0.729 + ], + "angle": 0, + "content": "[33] Tamas Matuszka, Ivan Barton, Adam Butykai, Peter Hajas, David Kiss, Domonkos Kovacs, Sandor Kunsagi-Mate, Peter Lengyel, Gabor Nemeth, Leventa PetHo, et al. aimotive dataset: A multimodal dataset for robust autonomous driving with long-range perception. arXiv preprint arXiv:2211.09445, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.483, + 0.801 + ], + "angle": 0, + "content": "[34] Lu Mi, Hao Wang, Yonglong Tian, Hao He, and Nir N Shavit. Training-free uncertainty estimation for dense regression: Sensitivity as a surrogate. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10042-10050, 2022. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.483, + 0.87 + ], + "angle": 0, + "content": "[35] Trier Mortlock, Arnav Malawade, Kohei Tsujio, and Mohammad Al Faruque. Castnet: a context-aware, spatio-temporal dynamic motion prediction ensemble for autonomous driving. ACM Transactions on Cyber-Physical Systems, 8(2):1–20, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[36] Jishnu Mukhoti, Andreas Kirsch, Joost van Amersfoort, Philip HS Torr, and Yarin Gal. Deep deterministic uncertainty:" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.134 + ], + "angle": 0, + "content": "A new simple baseline. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24384-24394, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.135, + 0.907, + 0.216 + ], + "angle": 0, + "content": "[37] Yang Ni, Hanning Chen, Prathyush Poduval, Zhuowen Zou, Pietro Mercati, and Mohsen Imani. Brain-inspired trustworthy hyperdimensional computing with efficient uncertainty quantification. In 2023 IEEE/ACM International Conference on Computer Aided Design (ICCAD), pages 01-09. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.218, + 0.907, + 0.287 + ], + "angle": 0, + "content": "[38] Igor Nunes, Mike Heddes, Tony Givargis, Alexandru Nicolau, and Alex Veidenbaum. Graphhd: Efficient graph classification using hyperdimensional computing. In 2022 Design, Automation & Test in Europe Conference & Exhibition (DATE), pages 1485-1490. IEEE, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.288, + 0.907, + 0.343 + ], + "angle": 0, + "content": "[39] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch. In NIPS-W, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.907, + 0.399 + ], + "angle": 0, + "content": "[40] Luis Manuel Pereira, Addisson Salazar, and Luis Vergara. On comparing early and late fusion methods. In International Work-Conference on Artificial Neural Networks, pages 365-378. Springer, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.907, + 0.468 + ], + "angle": 0, + "content": "[41] Janis Postels, Francesco Ferroni, Huseyin Coskun, Nassir Navab, and Federico Tombari. Sampling-free epistemic uncertainty estimation using approximated variance propagation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2931-2940, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.469, + 0.907, + 0.524 + ], + "angle": 0, + "content": "[42] Janis Postels, Mattia Segu, Tao Sun, Luca Sieber, Luc Van Gool, Fisher Yu, and Federico Tombari. On the practicality of deterministic epistemic uncertainty. arXiv preprint arXiv:2107.00649, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.907, + 0.566 + ], + "angle": 0, + "content": "[43] Ali Rahimi and Benjamin Recht. Random features for large-scale kernel machines. Advances in neural information processing systems, 20, 2007. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.567, + 0.907, + 0.636 + ], + "angle": 0, + "content": "[44] Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Schölkopf, Thomas Brox, and Peter Gehler. Towards total recall in industrial anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14318-14328, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.637, + 0.907, + 0.677 + ], + "angle": 0, + "content": "[45] Kenny Schlegel, Peer Neubert, and Peter Protzel. A comparison of vector symbolic architectures. Artificial Intelligence Review, 55(6):4523-4555, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.907, + 0.748 + ], + "angle": 0, + "content": "[46] Jiahui She, Yibo Hu, Hailin Shi, Jun Wang, Qiu Shen, and Tao Mei. Dive into ambiguity: Latent distribution mining and pairwise uncertainty estimation for facial expression recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6248-6257, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.749, + 0.907, + 0.801 + ], + "angle": 0, + "content": "[47] Apoorv Singh. Transformer-based sensor fusion for autonomous driving: A survey. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3312-3317, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.907, + 0.858 + ], + "angle": 0, + "content": "[48] Anthony Thomas, Sanjoy Dasgupta, and Tajana Rosing. A theoretical perspective on hyperdimensional computing. Journal of Artificial Intelligence Research, 72:215-249, 2021. 3, 4, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[49] Junjiao Tian, Wesley Cheung, Nathaniel Glaser, Yen-Cheng Liu, and Zsolt Kira. Uno: Uncertainty-aware noisy-or multimodal fusion for unanticipated input degradation. In 2020" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22315" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "IEEE International Conference on Robotics and Automation (ICRA), pages 5716-5723. IEEE, 2020. 5, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.486, + 0.178 + ], + "angle": 0, + "content": "[50] Stefan Van der Walt, Johannes L Schonberger, Juan Nunez-Iglesias, François Boulogne, Joshua D Warner, Neil Yager, Emmanuelle Gouillart, and Tony Yu. scikit-image: image processing in python. PeerJ, 2:e453, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.485, + 0.233 + ], + "angle": 0, + "content": "[51] Jorge Vargas, Suleiman Alsweiss, Onur Token, Rahul Razdan, and Joshua Santos. An overview of autonomous vehicles sensors and their vulnerability to weather conditions. Sensors, 21(16):5397, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.485, + 0.29 + ], + "angle": 0, + "content": "[52] Junyao Wang, Sitao Huang, and Mohsen Imani. Disthd: A learner-aware dynamic encoding method for hyperdimensional classification. In 2023 60th ACM/IEEE Design Automation Conference (DAC), pages 1-6. IEEE, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.292, + 0.485, + 0.374 + ], + "angle": 0, + "content": "[53] Junyao Wang, Arnav Vaibhav Malawade, Junhong Zhou, Shih-Yuan Yu, and Mohammad Abdullah Al Faruque. Rs2g: Data-driven scene-graph extraction and embedding for robust autonomous perception and scenario understanding. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 7493-7502, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.376, + 0.485, + 0.444 + ], + "angle": 0, + "content": "[54] Shaojie Wang, Tong Wu, Ayan Chakrabarti, and Yevgeniy Vorobeychik. Adversarial robustness of deep sensor fusion models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2387-2396, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.446, + 0.485, + 0.514 + ], + "angle": 0, + "content": "[55] Samuel Wilson, Tobias Fischer, Niko Sünderhauf, and Feras Dayoub. Hyperdimensional feature fusion for out-of-distribution detection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2644-2654, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.517, + 0.485, + 0.571 + ], + "angle": 0, + "content": "[56] De Jong Yeong, Gustavo Velasco-Hernandez, John Barry, and Joseph Walsh. Sensor and sensor fusion technology in autonomous vehicles: A review. Sensors, 21(6):2140, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.574, + 0.485, + 0.628 + ], + "angle": 0, + "content": "[57] Tao Yu, Yichi Zhang, Zhiru Zhang, and Christopher M De Sa. Understanding hyperdimensional computing for parallel single-pass learning. Advances in Neural Information Processing Systems, 35:1157-1169, 2022. 3, 8, 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.63, + 0.485, + 0.712 + ], + "angle": 0, + "content": "[58] Jiaming Zhang, Ruiping Liu, Hao Shi, Kailun Yang, Simon Reiβ, Kunyu Peng, Haodong Fu, Kaiwei Wang, and Rainer Stiefelhagen. Delivering arbitrary-modal semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1136-1147, 2023. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.715, + 0.485, + 0.769 + ], + "angle": 0, + "content": "[59] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4490-4499, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.771, + 0.485, + 0.854 + ], + "angle": 0, + "content": "[60] Zijian Zhu, Yichi Zhang, Hai Chen, Yinpeng Dong, Shu Zhao, Wenbo Ding, Jiachen Zhong, and Shibao Zheng. Understanding the robustness of 3d object detection with bird's-eye-view representations in autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21600-21610, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.854 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22316" + } + ] +] \ No newline at end of file diff --git a/2025/Hyperdimensional Uncertainty Quantification for Multimodal Uncertainty Fusion in Autonomous Vehicles Perception/8d8555dd-a579-45c4-ad75-fee4bbfad5af_origin.pdf b/2025/Hyperdimensional Uncertainty Quantification for Multimodal Uncertainty Fusion in Autonomous Vehicles Perception/8d8555dd-a579-45c4-ad75-fee4bbfad5af_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..69998b10893c29e38d4e283a4989ee9f28ba2eab --- /dev/null +++ b/2025/Hyperdimensional Uncertainty Quantification for Multimodal Uncertainty Fusion in Autonomous Vehicles Perception/8d8555dd-a579-45c4-ad75-fee4bbfad5af_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:967674a23f96fae38e0771cdadb99635a20d9af7e9acd57bcd12ac4923c74996 +size 6118162 diff --git a/2025/Hyperdimensional Uncertainty Quantification for Multimodal Uncertainty Fusion in Autonomous Vehicles Perception/full.md b/2025/Hyperdimensional Uncertainty Quantification for Multimodal Uncertainty Fusion in Autonomous Vehicles Perception/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9a5afc36604bfac17282aa58eb558005f351ba09 --- /dev/null +++ b/2025/Hyperdimensional Uncertainty Quantification for Multimodal Uncertainty Fusion in Autonomous Vehicles Perception/full.md @@ -0,0 +1,329 @@ +# Hyperdimensional Uncertainty Quantification for Multimodal Uncertainty Fusion in Autonomous Vehicles Perception + +Luke Chen* Junyao Wang Trier Mortlock Pramod Khargonekar Mohammad Abdullah Al Faruque University of California, Irvine + +{panwangc, junyaow4, tmortloc, pramod.khargonekar, alfaruqu}@uci.edu + +# Abstract + +Uncertainty Quantification (UQ) is crucial for ensuring the reliability of machine learning models deployed in real-world autonomous systems. However, existing approaches typically quantify task-level output prediction uncertainty without considering epistemic uncertainty at the multimodal feature fusion level, leading to sub-optimal outcomes. Additionally, popular uncertainty quantification methods, e.g., Bayesian approximations, remain challenging to deploy in practice due to high computational costs in training and inference. In this paper, we propose HyperDUM, a novel deterministic uncertainty method (DUM) that efficiently quantifies feature-level epistemic uncertainty by leveraging hyperdimensional computing. Our method captures the channel and spatial uncertainties through channel and patch-wise projection and bundling techniques respectively. Multimodal sensor features are then adaptively weighted to mitigate uncertainty propagation and improve feature fusion. Our evaluations show that HyperDUM on average outperforms the state-of-the-art (SOTA) algorithms by up to $2.01\% / 1.27\%$ in 3D Object Detection and up to $1.29\%$ improvement over baselines in semantic segmentation tasks under various types of uncertainties. Notably, HyperDUM requires $2.36 \times$ less Floating Point Operations and up to $38.30 \times$ less parameters than SOTA methods, providing an efficient solution for real-world autonomous systems. + +# 1. Introduction + +Effectively quantifying and addressing uncertainty plays a critical role in ensuring the reliability of ML models in real-world applications, especially where the decisions are consequential, e.g., autonomous driving. Models failing to address uncertainty can produce overconfident outputs, leading to compromised performance [9, 16, 41, 53]. In light of autonomous systems, uncertainties can arise from adverse + +weather conditions, sensor failures or noises, environment dynamics (movements of other actors, road conditions, vehicle dynamics), and various corner cases [4, 12, 35, 36]. Modern autonomous systems leverage multimodal sensor fusion to broaden the information scope and learn comprehensive representations of the surrounding environment [8, 31, 56]. However, commonly employed sensors such as cameras, lidar, and radar, exhibit varying degrees of robustness and vulnerability across different driving scenarios [51, 54]. For instance, cameras are crucial in delivering rich perceptual information while they are particularly vulnerable to lighting variations and occlusions [60]. Lidar and radar are more robust to visual perturbations but can be affected by multi-path interference and occlusion [2, 28]. Therefore, it is important to consider the uncertainty of each modality before the fusion of multimodal features to achieve optimal results [15]. + +Popular Uncertainty Quantification (UQ) methods include deep ensembles, Bayesian neural networks, and Bayesian approximation methods such as Monte Carlo dropout [11, 23, 24]. However, these methods often introduce large training and/or inference overheads compounded by the growing complexity of modern Deep Neural Networks (DNN), ultimately challenging their practicality under real-time computing constraints [7, 22]. Other lines of works such as Conformal Prediction (CP) and Posterior Networks (PostNets) based on evidential deep learning either cannot quantify feature-level epistemic uncertainty or require vast amounts of quality data to accurately estimate the posteriors [1, 5]. Deterministic Uncertainty Methods (DUMs), aiming to efficiently quantify uncertainty using only a single forward pass [10, 34, 42], have become an emerging trend. DUMs rely on statistical/morphological characteristics of latent features to quantify epistemic uncertainty. For instance, [10] employed Prototype Learning (PL) techniques to capture the "knowledge" of DNNs by constructing a "memory bank" of prototypes to predict uncertainty. + +To address these problems, we propose HyperDUM, a novel deterministic uncertainty method for efficient quantification of feature-level epistemic uncertainty. In contrast to existing methods, HyperDUM projects latent features into + +hyperdimensional space to form hyperdimensional prototypes, and then estimates uncertainty based on the distance between features of unseen samples and these prototypes. + +To the best of our knowledge, HyperDUM is the first deterministic uncertainty quantification method utilizing hyperdimensional feature prototyping with Channel-wise and Patch-wise projection and bundling to deliver accurate and efficient epistemic uncertainty quantification of latent features. Detailed experimental evaluations on the DeLiVER and aiMotive datasets for semantic segmentation and object detection tasks demonstrate that HyperDUM can efficiently and effectively quantify feature uncertainty and improve fusion performance when compared to standard Bayesian approximation inference baselines and SOTA methods. + +# 2. Related Works + +# 2.1. Uncertainty Quantification + +Predictive uncertainty is composed of three factors, model/epistemic uncertainty, data/aleatoric uncertainty, and distributional uncertainty [32]. Epistemic uncertainty, stemming from the word epistemology which is the study of knowledge, is associated with a model's lack of knowledge of underrepresented data. The scope of this work is to efficiently quantify epistemic uncertainty at the latent feature level to enable uncertainty-weighting prior to feature fusion. + +Existing UQ approaches are typically derived from Bayesian principles, including Bayesian neural networks which represent model weights as probability distributions, deep ensembles which sample output distributions from diverse models [24], and Monte Carlo dropout which is similar to deep ensembles but utilizes a single model with dropout layers [11]. However, these approaches remain challenging in practice due to enormous training and inference costs [7, 22]. Conformal prediction (CP) is a model-agnostic UQ technique that provides prediction intervals with guaranteed coverage, ensuring the true value falls within the interval at a specified probability [1]. However, while effective in producing reliable task-level uncertainties, CP does not capture uncertainties at intermediate model abstractions, a limitation especially evident in multimodal fusion models, where uncertainty propagates across different modalities before the final output. Evidential deep learning is another line of work, i.e. Posterior Networks which aim to quantify both aleatoric and epistemic uncertainty by directly estimating a posterior distribution over the model's outputs/predictions but is sensitive to data quality for accurately estimating the posteriors [5]. Deterministic uncertainty quantification is a growing research area that aims to efficiently compute epistemic uncertainty of DNNs with a single forward pass [10, 34, 42]. For instance, [34] is a training-free method that injects Gaussian noise into intermediate layers to compute the variance across the representations as a measure of uncertainty. La + +tent Deterministic Uncertainty (LDU) constructs prototypes from latent features by enforcing dissimilarity between prototypes and entropy between latent features while correlating uncertainty and the downstream task [10, 36, 46]. Recent research [37] demonstrated that constructing hyperdimensional prototypes can provide comparable performance to existing aleatoric uncertainty quantification solutions, e.g., Bayesian approximation, and offer significant speedups in training and inference. However, it is applied to regression tasks for aleatoric uncertainty. In contrast, our work quantifies epistemic uncertainty on the latent features, extending its applicability to multimodal fusion models and complex tasks such as object detection and semantic segmentation. + +# 2.2. Multimodal Uncertainty-Aware Fusion + +Multimodal sensor fusion aims to provide more comprehensive representations and address the limitations of unimodal sensor failure cases [8, 31, 56]. Existing approaches dealing with uncertainties fall into two categories: feature-level fusion and output-level fusion. Feature-level fusion, also known as intermediate fusion, applies an uncertainty weighting that scales the features to minimize uncertainty [30, 47]. For instance, [14] computes feature-informativeness and scales the original features based on informativeness. On the other hand, output-level fusion, also known as late-fusion, uses uncertainty to calibrate the outputs of the model, i.e. calibrating softmax probabilities [40]. In contrast to existing uncertainty-based fusion works that commonly rely on Bayesian approximated uncertainty fusion, our proposed HyperDUM is the first to investigate deterministic uncertainty fusion methods in a multimodal autonomous systems setting. + +# 3. Methodology + +# 3.1. Preliminaries + +Given an input sample $\mathbf{x} = \{x_{1}, x_{2}, \dots, x_{M}\}$ consisting of $M$ multimodal sensor inputs. We have a multimodal model $\mathbb{M}$ with pre-trained feature encoders $\mathbf{f} = \{f_{1}, \dots, f_{M}\}$ that extract modality-specific abstract features $z_{m} = f_{m}(x_{m})$ where $1 \leq m \leq M$ . Additionally, the model contains fusion blocks $\mathcal{F}(z_{i}, z_{j})$ where $i, j \in \{1, \dots, M\}$ such that $i \neq j$ which combine modality-specific features to learn cross-modal fusion features $g$ . Assuming there are $K$ fusion blocks, the fused features $\mathbf{g} = \{g_{1}, \dots, g_{K}\}$ are passed to the task head $h(g)$ to output the corresponding task predictions $y$ (i.e. bounding boxes and object classification for object detection, and pixel-wise class labels for semantic segmentation). To account for feature uncertainty, an uncertainty quantification module $\mathbb{U}_{m}(z_{m})$ is inserted immediately following each feature extractor to predict an uncertainty value $u_{m}$ . Given the uncertainty value, a learnable + +![](images/d5504aea6ccdf5201820be0635a2f39dd6045f3706781f2363bc9e893a3a7f54.jpg) +Figure 1. Multimodal model with uncertainty quantification and uncertainty fusion for autonomous driving perception tasks. + +uncertainty weighting module $\Omega (z_m,u_m)$ reweights the input features to produce an uncertainty-aware feature $\hat{z}_m$ for each modality. Thus, the input to the fusion blocks now becomes $\mathcal{F}(\hat{z}_i,\hat{z}_j)$ to produce uncertainty-weighted fusion features $\hat{\mathbf{g}}$ which are forwarded to be fused and subsequently used by the task head $h(\hat{\mathbf{g}})$ to produce the outputs. The system flow of the described multimodal fusion model with uncertainty quantification is illustrated in Figure 1. + +# 3.1.1 Vector Symbolic Architectures/Hyperdimensional Computing + +Vector Symbolic Architectures (VSA) / Hyperdimensional Computing (HDC) are a class of algorithms for creating and handling hyperdimensional representations to mimic cognitive functions [25, 45]. VSA starts with mapping input data to hyperdimensional vectors, or hypervectors $\mathcal{H}$ , through a projection function $\phi: \mathcal{Z} \to \mathbf{H}$ , where $\mathcal{Z} \in \mathbb{R}^n$ and $\mathbf{H} \in \mathbb{R}^d$ ( $d \gg n$ ) is the hypervector space. The projection function depends on the input data representation. This work assumes Euclidean input space. For non-euclidean see [38]. Following the projection, common operations on hypervectors include similarity calculation, bundling, and binding. Similarity $(\delta(\cdot, \cdot))$ calculation measures the distance between two hypervectors. For real-valued hypervectors, a common measure is cosine similarity. Bundling $(\bigoplus)$ is an element-wise addition of hypervectors, e.g., $\mathcal{H}^{bundle} = \mathcal{H}^1 \bigoplus \mathcal{H}^2$ , generating a hypervector with the same dimension as inputs. Binding $(\otimes)$ is an element-wise multiplication associating two hypervectors to create another near-orthogonal hypervector, e.g., $\mathcal{H}^{bind} = \mathcal{H}^1 \bigotimes \mathcal{H}^2$ where $\delta(\mathcal{H}^{bind}, \mathcal{H}^1) \approx 0$ and $\delta(\mathcal{H}^{bind}, \mathcal{H}^2) \approx 0$ . + +Our work mainly leverages the bundling operation of VSA for the formation of prototypes. In high-dimensional space, bundling mimics how human brains memorize information [48, 57]. For instance, for $\mathcal{H}^{bundle} = \mathcal{H}^1 +\mathcal{H}^2$ + +we have $\delta (\mathcal{H}^{bundle},\mathcal{H}^1)\gg 0$ while $\delta (\mathcal{H}^{bundle},\mathcal{H}^3)\approx 0$ $(\mathcal{H}^3\neq \mathcal{H}^1,\mathcal{H}^2)$ . In other words, VSA can generate prototypical representations by bundling the high-dimensional vectors of individual data samples, making it especially useful for prototype learning. Specifically, suppose we have labeled samples $\mathcal{D} = \{(z_1^1,y_1^1),(z_1^2,y_1^2),\ldots ,(z_M^N,y_M^N)\}$ where $z_{m}^{i}\in \mathcal{Z}$ and $y_{m}^{i}\in \{l\}_{l = 1}^{L}$ , we can form $L$ prototypes by bundling hypervectors corresponding to a particular label: + +$$ +\mathcal {H} _ {m} ^ {l} = \bigoplus_ {i: y ^ {i} = l} \phi_ {m} \left(z _ {m} ^ {i}\right) \tag {1} +$$ + +where each hyperdimensional prototype can represent an aspect of the data samples $\mathcal{D}$ depending on the label definition, i.e. class label (cat and dog) or context label (weather conditions which we use in this work). We then define the notion of uncertainty as the set of similarity between hypervectors and hyperdimensional prototype of modality $m$ as follows: + +$$ +\mathbb {U} _ {m} = \bigcup_ {l = 1} ^ {L} \left\{\delta \left(\mathcal {H} _ {m} ^ {z}, \mathcal {H} _ {m} ^ {l}\right) \right\}, \text {w h e r e} | \mathbb {U} _ {m} | = L \tag {2} +$$ + +where $\mathbb{U}_m$ is the set of similarity distances between a hypervector $\mathcal{H}_m^z$ and each hyperdimensional prototype $\mathcal{H}_m^l$ , $l = 1,2,\dots ,L$ and used to express the similarity uncertainty. The notion of similarity for uncertainty has been used in other prototype learning methods [10, 26], where the former assigns the variations of similarities between an instance and the prototypes as the belief masses and the latter utilizes similarity to optimize the uncertainty loss. + +# 3.2. HyperDUM + +Following the conventions of VSA for projection and bundling, we propose two modifications for uncertainty quan- + +tification: 1) Channel-wise and 2) Patch-wise Projection & Bundling (CPB/PPB). CPB modifies the projection and bundling operations to capture the uncertainty of each channel per modality separately, whereas PPB allows the VSA to quantify the spatial uncertainty without losing fine-grained spatial information through patching. Figure 1 visualizes the CPB and PPB operations of the proposed method. + +# 3.2.1 Channel-wise Projection & Bundling (CPB) + +Latent features are generally multi-channeled. By convention prototype formation based on Equation 1 projects and bundles all feature dimensions onto a common representation space $\mathbf{H}$ [52, 55]. Given latent feature $z_{m}^{i} \in \mathbb{R}^{C \times H \times W}$ with $C$ denoting feature channels and $H, W$ the spatial dimensions, adaptive pooling (max or average) is first used to reduce the spatial dimensions to $z_{m}^{pooled} \in \mathbb{R}^{C}$ . Through matrix multiplication and proper dimension initialization of the projection matrix $\Phi$ , the correct hypervector dimensions can be obtained for $\mathcal{H}_{m}^{z}$ as follows: + +$$ +\phi_ {m} \left(z _ {m} ^ {i}\right) = \Phi^ {(d \times C)} \cdot z _ {m} ^ {\text {p o o l e d} (C)} = \mathcal {H} _ {m} ^ {z} ^ {(d)} \tag {3} +$$ + +Another method to handle multi-channels inputs is by projecting each feature dimension separately followed by binding and subsequent bundling to a common space [48]. However, we note that the above operations neglect the fact that different feature channels learn different aspects of the input and may contribute more or less to the overall uncertainty. For example, [27] showed that sharing the same uncertain distribution among different channels is less effective for out-of-distribution generalization, while considering each channel uncertainty separately brings better performances due to the different channel potentials. Therefore, instead of absorbing the channel dimensions during projection and bundling, we retain them as follows: + +$$ +\phi_ {m} \left(z _ {m} ^ {i}\right) = \Phi^ {(d \times C)} \otimes z _ {m} ^ {\text {p o o l e d} (C)} = \mathcal {H} _ {m} ^ {z} ^ {(d \times C)} \tag {4} +$$ + +where we use the Einstein summation notation $\otimes$ (not the binding $\otimes$ notation) to retain the channel dimension. Consequently, the bundling operation is modified to account for the channel dimensions: + +$$ +\mathcal {H} _ {m} ^ {l} = \bigoplus_ {i: y ^ {i} = l} ^ {d i m = C} \phi_ {m} \left(z _ {m} ^ {i}\right) \tag {5} +$$ + +where $dim = C$ indicates the dimension over which the element-wise addition is performed and $\mathcal{H}_m^l$ is now $\mathbb{R}^{d\times C}$ . Additionally, the similarity uncertainty now outputs similarities for each channel dimension by representing Equation 2: + +$$ +\mathbb {U} _ {m} ^ {C} = \bigcup_ {c = 1} ^ {C} \left\{\bigcup_ {l = 1} ^ {L} \left\{\delta \left(\mathcal {H} _ {m} ^ {z} [ c ], \mathcal {H} _ {m} ^ {l} [ c ]\right) \right\} \right\}, \text {s . t .} | \mathbb {U} _ {m} ^ {C} | = C \times L \tag {6} +$$ + +We empirically demonstrate through experiments that channel projection and bundling outperforms conventional projection and bundling for the task of multimodal uncertainty-aware feature fusion. + +# 3.2.2 Patch-wise Projection & Bundling (PPB) + +Channel projection and bundling enables our method to capture the holistic uncertainties of each feature channel. However, to capture the finer-granularity uncertainties of the spatial dimensions requires a different approach. By convention spatial dimensions are either pooled to a single value and projected to high dimensions as shown by Equation 3 or each spatial dimension is separately projected to high dimensions. The former loses spatial information due to pooling and the latter is computationally infeasible and memory intensive given the large spatial dimensions of deep neural networks. + +Inspired by the idea of using image patches to capture spatial information in anomaly detection and image classification [6, 44], we propose PPB which projects and bundles spatial patches of latent features to hyperdimensional space to capture the spatial uncertainties. Specifically, given latent feature $z_{m}^{i}\in \mathbb{R}^{C\times H\times W}$ , we slice $(H,W)$ into $P$ patches of resolution $(\tilde{h},\tilde{w})$ where $\tilde{h} = H / \sqrt{(P)}$ , $\tilde{w} = W / \sqrt{(P)}$ . This gives us the set of patched features $\tilde{z}_{m}^{i} = \{\tilde{z}_{m}^{i,j}\}_{j = 1}^{P}$ where each $\tilde{z}_{m}^{i,j}\in \mathbb{R}^{C\times \tilde{h}\times \tilde{w}}$ are pooled to $\mathbb{R}^C$ and form the set of pooled patched features $\tilde{z}_{m}^{pooled}\in \mathbb{R}^{C\times P}$ . + +Given patched features $\tilde{z}_m^{pooled}$ , we project them according to Equation 3 to form the set of patched projections: + +$$ +\phi_ {m} \left(\tilde {z} _ {m} ^ {i}\right) = \Phi^ {(d \times C)} \cdot \tilde {z} _ {m} ^ {\text {p o o l e d} (C \times P)} = \mathcal {H} _ {m} ^ {\tilde {z}} ^ {(d \times P)} \tag {7} +$$ + +Consequently, these projections are bundled for each patch according to Equation 1 to produce the set of patched prototypes as follows: + +$$ +\mathcal {H} _ {\tilde {m}} ^ {l} = \bigoplus_ {i: y ^ {i} = l} ^ {d i m = P} \phi_ {m} \left(\tilde {z} _ {m} ^ {i}\right) \tag {8} +$$ + +where $\mathcal{H}_{\tilde{m}}^{l}$ represents the patched prototypes of modality $m$ , for each patch and for each context label $i$ , making $|\mathcal{H}_{\tilde{m}}^{l}| = d\times P$ + +Finally, we can obtain the spatial uncertainties for each patch of the input projection $\mathcal{H}_m^{\tilde{z}}$ by modifying Equation 2: + +$$ +\mathbb {U} _ {\tilde {m}} ^ {P} = \bigcup_ {p = 1} ^ {P} \left\{\bigcup_ {l = 1} ^ {L} \left\{\delta \left(\mathcal {H} _ {\tilde {m}} ^ {\tilde {z}} [ p ], \mathcal {H} _ {\tilde {m}} ^ {l} [ p ]\right) \right\} \right\}, \text {s . t .} | \mathbb {U} _ {\tilde {m}} ^ {P} | = P \times L \tag {9} +$$ + +Our final proposed approach combines both channel and patch methods for projection and bundling to capture channel-wise feature uncertainties, as well as finer + +granularity spatial feature uncertainties, to provide a holistic epistemic uncertainty quantification of each modality. + +# 4. Experiments + +# 4.1. Experimental Setup + +# 4.1.1 Datasets + +We evaluate the effectiveness of HyperDUM for uncertainty-aware feature fusion using two multimodal autonomous driving datasets. The aiMotive dataset [33] includes camera, lidar, and radar sensors for long-distance $(\geq 75m)$ 3D object detection and features diverse scenes across various locations, times, and weather conditions. The DeLiVER dataset [58] contains camera, lidar, event, and depth sensors for semantic segmentation, including adverse corner cases such as motion blur and lidar jitter. Additionally, we integrate corner cases into the aiMotive dataset by applying various lighting effects to camera sensors, causing reduced clarity and scene obstruction. These effects are created through image transformations such as GaussianBlur and exposure adjustments [39, 50]. We also implement foggification [3] to simulate synthetic fogging of lidar data, resulting in noisy point clouds with reduced field of view. These corner case effects can be visualized in Figure 2. Details in the appendix. + +# 4.1.2 Models + +We utilize the pretrained BEVFusion [33] and CMNeXT [58] models as the fusion architectures for aiMotive and DeLiVER, respectively. BEVFusion is built upon VoxelNet [59] and BEVDepth [29] to fuse camera and lidar representations on a unified BEV space. CMNeXT is a four-stage pyramidal fusion model that treats the RGB representation as the primary branch and other modalities as the secondary branch. We integrate the uncertainty quantification module (UQM) at each model's respective pre-fusion points. For aiMotive this is chosen to be at the feature concatenation step, right before the bev Fuse step. For DeLiVER, we insert the UQM immediately after the feature rectification module and right before the feature fusion module. Given that the CMNeXT model has four stages, we instantiate one UQM per stage. Details in the Appendix. + +# 4.1.3 Baseline Methods + +We implemented different UQ methods and compared their performances against our method. For each implementation, we followed their open-sourced code implementation and hyperparameter settings. Specifically, for Infer-dropout (InfMCD) and infer-noise (InfNoise), we followed their suggestions and chose 10 forward passes and used the variance of the outputs for uncertainty estimation. In addition to variance, we computed predictive entropy, mutual information, + +and entropy as additional uncertainty estimates following [49]. It was shown that using more uncertainty metrics can provide more effective quantification for different types of uncertainties. For Posterior Network (PostNet), we constructed PostNets for each modality so that they can estimate the modality-specific uncertainty independently and be used in the uncertainty-aware feature fusion process. The use of multiple PostNets is similar to [20] in their construction of separate PostNets for agent's past behavior, road structure map, and social context respectively. For latent deterministic uncertainty (LDU), we set the number of learnable prototypes to the number of scenarios. Specifically, for aiMootive, the scenarios are Highway, Urban, Night, and Rain, while for DeLiVER, the scenarios are Cloudy, Foggy, Night, Rainy, and Sunny. Finally, for all methods, all layers prior to the UQM are kept frozen and we fine-tune the uncertainty weighting layer, which takes the respective uncertainty metrics as input, along with other post-fusion layers, to enable the model to learn the uncertainty-weighted features. Details of the fine-tuning procedures and visualizations of the architectures and UQM insertion points are provided in the Appendix. In all of our experiments the best one is highlighted and bolded and the 2nd best is lightly-highlighted. + +# 4.2. Multimodal Uncertainty Fusion for Autonomous Vehicles Perception + +# 4.2.1 3D Object Detection + +
UQ MethodHighwayUrbanNightRainMean
BEVFusion [33]72.55/69.0863.72/63.5274.74/73.1342.75/42.8365.67/64.79
InfMCD [34]72.58/69.6562.02/59.5975.05/73.7235.58/37.4164.56/64.58
InfNoise [34]70.11/68.5263.10/60.4275.08/74.1935.58/37.4164.59/64.56
PostNet [5]69.19/68.0261.38/59.8473.32/68.9734.95/37.6063.13/60.21
LDU [10]71.48/69.3362.54/60.1176.49/75.0440.49/41.2464.69/64.73
HyperDUM72.23/69.5864.77/64.4876.69/75.1544.78/45.4866.70/66.00
+ +Table 1 compares the all-point/11-point interpolation Average Precision (AP) metric for object detection under diverse scenes. It can be seen that apart from the Highway scenario, HyperDUM outperforms all methods across all scenarios. Specifically, HyperDUM improves upon the all-point AP by $2.01\%$ on average against the state-of-the-art (LDU). + +Table 1. Results on aiMotive under diverse scenes. + +
UQ MethodMBOEUELFMean
BEVFusion [33]64.10/63.6962.40/59.2663.57/63.2565.07/65.0863.79/62.75
InfMCD [34]62.00/61.9964.47/64.2863.25/63.4964.27/63.8163.50/63.39
InfNoise [34]62.61/62.6164.94/64.7263.18/60.0765.19/64.9963.98/63.10
PostNet [5]60.42/58.9063.12/59.9960.20/59.2564.13/60.6061.97/59.69
LDU [10]62.06/62.2265.07/64.8562.87/60.0265.47/65.1063.87/63.05
HyperDUM64.39/63.9966.16/65.3964.18/63.9165.89/65.2265.16/64.62
+ +Table 2. Results on aiMotive under corner cases: Motion Blur (MB), Over-Exposure (OE), Under-Exposure (UE), LiDAR-Fog (LF). (Metrics: all-point AP/11-point interpolation AP) + +Table 2 demonstrates the robustness of UQ methods un + +![](images/18f3f884b122647cd429b9c7351cd341d8ce7c342c38bdeb4106a85bfbf62696.jpg) +Original + +![](images/1ebd5c7b4d839028ec73d946db4182304852d06e5c77471a041ac0bcc3a6af40.jpg) +Original + +![](images/d04c34863021ca7bd143cc4105ff25333fea4ed304d0673ec6196a77d6aa562d.jpg) +Motion Blur +Figure 2. Corner cases for the aiMotive 3D Object Detection dataset. (Top-Left) Normal Lidar point-cloud birds-eye-view. (Top-Right) Lidar Foggification (LF). (Bottom) In order from left to right (Normal, Motion Blur (MB), Over-Exposure (OE), Under-Exposure (UE)) + +![](images/5db4b7439059f0370f743806c814ea632f5fd496cd2d77e7a5f18bb93b3bbc25.jpg) +Foggification +Over-Exposure + +![](images/575cd4a22a55f882b1974eac25a54afa14eecbefb7fe86ba2282d1e7e2f1d0bf.jpg) +Under-Exposure + +der our injected corner cases. These injections were unseen during training which complicates testing due to distribution shifts. Each corner case is applied to and evaluated on the entire test set across all scenarios. Whereas other methods fail and suffer performance degradations, HyperDUM only degrades $0.51\% / 0.17\%$ compared to the baseline w/o injections demonstrating that HyperDUM is more robust even when facing distribution shifts. + +
UQ MethodHighwayUrbanNightRainMean
BEVFusion [33]45.59/45.3446.38/45.5344.49/45.8742.65/42.9345.84/46.87
InfMCD [34]45.57/46.2938.90/38.1142.07/41.9726.13/28.4940.74/41.64
InfNoise [34]40.73/40.3842.41/40.9038.80/40.0029.28/32.8140.91/42.04
PostNet [5]41.13/40.6439.47/40.1640.37/41.9228.56/28.9639.37/41.66
LDU [10]43.07/45.0140.17/41.7142.90/42.5231.12/34.0840.65/42.17
HyperDUM44.86/46.0647.89/48.9144.42/46.1240.54/41.6346.52/48.27
+ +Table 3 is an experiment on the distant region setting, keeping only predictions and ground truths over $\geq 75m$ . As expected, all UQ methods experience major performance drops, especially for Rain. This is likely due to the combined uncertainties of distance with rain obstruction causing substantial minimization of the uncertainty-weighted features (the weighting output is between 0-1 times the original feature value). The original BEVFusion model maintains better performance because it is still able to "guess" the presence of objects as the features are not minimized by uncertainty. + +Table 3. Results on aiMotive under diverse scenes in the distant region ( $>75\mathrm{m}$ ). (Metrics: all-point AP/11-point interpolation AP) + +
ScenariosUQ Methods (Metrics: mIoU ↑)
CMNeXt [58]InfMCD [34]InfNoise [34]PostNet [5]LDU [10]Gemini [21]HyperDUM (Ours)
Cloudy68.7069.2369.2169.2868.94-69.76+1.06
Foggy65.6666.1866.1066.0465.72-66.85+1.19
Night62.4663.4463.1462.9763.06-64.21+1.75
Rainy67.5068.0868.0968.1667.82-68.71+1.21
Sunny66.5767.0167.1666.9366.75-67.87+1.30
MB62.9163.6163.5563.5563.16-64.28+1.37
OE64.5965.3965.1765.0664.73-65.67+1.08
UE60.0060.3860.4260.2760.29-61.20+1.20
LJ65.9266.1266.2566.3366.40-66.93+1.01
EL65.4866.0566.1766.0665.89-66.80+1.32
Mean66.3066.9066.8666.7766.6066.9067.59+1.29
+ +Table 4. DeLiVER mean Intersection over Union (mIoU) performance under adverse weather and corner cases. Lidar-Jitter (LJ), Event Low-resolution (EL). + +# 4.2.2 Semantic Segmentation + +Table 4 shows the DeLiVER validation set performance on different scenarios and corner cases for the semantic segmentation task. Overall, HyperDUM achieves the best performance over all methods across every weather and corner case scenario. On average HyperDUM improves over the baseline by 1.29. In particular, for the three most difficult scenarios including Under-Exposure, Night and Motion Blur, HyperDUM achieves at least 1.00 improvement over the baseline. For GeminiFusion [21], we only report the mean as they omitted the detailed performance breakdown in their paper. These results demonstrate that our method can better quantify the multimodal feature uncertainties and ultimately improve the fusion features for the semantic segmentation + +task. Additionally, we note that all UQ methods improved over the baseline method. This shows that incorporating UQ with fusion into multimodal pipelines not only enables UQ but also can improve baseline performance. + +
ScenariosUQ Methods (Metrics: ECE ↓)
CMNeXt [58]InfMCD [34]InfNoise [34]PostNet [5]LDU [10]HyperDUM (Ours)
Cloudy1.18E-021.05E-021.07E-020.99E-021.09E-020.94E-02-0.24
Foggy1.67E-021.48E-021.52E-021.45E-021.56E-021.35E-02-0.30
Night1.88E-021.63E-021.68E-021.71E-021.68E-021.39E-02-0.49
Rainy1.53E-021.37E-021.39E-021.33E-021.43E-021.21E-02-0.32
Sunny1.42E-021.25E-021.27E-021.24E-021.33E-021.12E-02-0.30
MB1.54E-021.32E-021.40E-021.32E-021.44E-021.19E-02-0.35
OE1.49E-021.55E-021.39E-021.42E-021.44E-021.31E-02-0.18
UE1.75E-022.07E-021.69E-021.84E-021.63E-021.53E-02-0.22
LJ1.54E-021.38E-021.39E-021.35E-021.40E-021.28E-02-0.26
EL1.45E-021.17E-021.27E-021.26E-021.27E-021.09E-02-0.36
Mean1.53E-021.35E-021.38E-021.34E-021.41E-021.20E-02-0.33
+ +Table 5 shows the Expected Calibration Error (ECE) metric for all methods across the same scenarios. ECE measures how well a model is calibrated by comparing it's predicted probabilities with the actual probabilities of the ground truth distribution. ECE helps us understand how reliable a model's confidence scores are. A higher ECE means the model is more overconfident (or unsure) about its predictions. As shown, HyperDUM has the lowest ECE score over all methods across every weather scenario and corner case with a 0.33 average decrease compared to the baseline. These results demonstrate that our method helps better calibrate the overconfident semantic segmentation model. In fact all UQ methods on average has a lower ECE compared to the baseline demonstrating its usefulness in model calibration. + +# 4.3. Pre vs. Post Uncertainty Feature Learning + +Table 5. DeLiVER Expected Calibration Error (ECE) performance under adverse weather and corner cases. + +
ScenariosUQ Methods (Metrics: mIoU ↑)
CMNeXt [58]InfMCD [34]InfNoise [34]PostNet [5]LDU [10]HyperDUM (Ours)
Cloudy68.768.7868.569.1369.0669.26
Foggy65.6665.7865.5965.8965.9266.40
Night62.4662.762.4462.8263.0863.71
Rainy67.567.6667.3768.0267.8468.29
Sunny66.5766.5266.3966.7866.7867.43
MB62.9163.1162.7963.4063.363.78
OE64.5964.9464.664.9164.9565.20
UE60.0059.7560.0760.1260.360.72
LJ65.9266.1466.0166.1766.2166.44
EL65.4865.765.6665.9166.0366.34
Mean66.3066.5966.4966.6066.6466.98
+ +Table 6 evaluates the performance of uncertainty feature weighting after the fusion block of Figure 1 or see the archi + +tecture figure in the Appendix. As expected, we found that the overall performance gain is limited/ lower compared to its prefusion uncertainty fusion counterpart. These results align with our motivation that different modalities experience different modality-specific uncertainties. And these uncertainties can propagate to the fusion module, undermining the fusion features and degrading downstream task performance. + +# 4.4. Model Ablation + +Table 6. DeLiVER mean Intersection over Union (mIoU) performance under adverse weather and corner cases using post-fusion features. + +
Dataset +MetricaiMotive +all-point AP/11-point AP ↑DeLiVER +mIoU ↑
HyperDUM +- w/o Patch (4x) Proj.-66.70/66.00 (Mean Δ)67.59 (Mean Δ)
-65.67/65.23 (-1.03/-0.77)-67.18 (-0.41)
- w/o Chan Proj.64.43/64.25 (-1.24/-0.98)66.63 (-0.55)
+ +Table 7. Model ablation for aiMotive and DeLiVER. + +For our model ablation, we analyzed the effects of sequentially removing our proposed components and the impact on the overall performance. As seen from Table 7 the removal of the Patch Projection (with 4 patches) results in a $1.03 / 0.77$ decrease in performance and the removal of Channel Projection in a $1.24 / 0.98$ decrease in performance for aiMotive. Similarly for the DeLiVER dataset, the removal of the Patch Projection (with 4 patches) results in a 0.41 decrease in performance and removing Channel Projection results in a 0.55 decrease in performance. Together we see that both methods contribute significantly to the overall performance demonstrating that both channel and spatial uncertainties should be accounted for to better capture the holistic uncertainty of the features before fusion. More ablations in the Appendix. + +# 4.4.1 Sensitivity Analysis + +![](images/922405aa88be352a81f0f4285c6fbc1fe69fb6c35dc991a9ac0256c67679f586.jpg) +Figure 3. Analysis on the effects of hyperdimension selection for channel projection on performance and computation (DeLiVER). + +Here we investigate the tunable parameters for channel and patch projection & bundling (CPB & PPB). For CPB, the only parameter is the channel hyperdimension. For PPB, we have the patch hyperdimension and the number of patches. + +![](images/198eb036dd032be08ad28d358f231b8de37a1fa24d0312ac4c023745ba753395.jpg) +Figure 3 shows a sweep over a range of dimensions from 625 to $20\mathrm{k}$ for CPB. This range of values is chosen based on the HDC expressivity limits (around $10\mathrm{k}$ ) [48, 57]. We see that performance improves with dimension increase, but plateaus and decreases past $10\mathrm{k}$ while computation continues to grow. This occurs when large dimensions induce too much sparsity, making similar samples become orthogonal and capturing irrelevant information likened to over-fitting in neural networks [48]. +Figure 4. Analysis on the effects of patches and patch hyperdimension selection on performance and computation (DeLiVER). + +Figure 4 shows a sweep over 4, 16 and 64 patch configurations together with the dimension sweep from 625 to 5000. We observe an interesting relation where the number of patches negatively correlates with the per-patch hyperdimension in terms of better performance. This makes sense as we increase the number of patches, the information needed to be encoded becomes smaller due to the reduction in the per-patch spatial resolution as indicated in Section 3.2.2. Therefore, the necessary hyperdimensions needed to encode the information should be reduced, while larger dimensions begin to over-fit. Whereas the opposite is true for less patches, larger hyperdimensions are needed. + +# 4.5. Computation Cost + +
DatasetaiMotiveDeLiVER
UQ MethodFLOPs#ParamsFLOPs#Params
InfMCD [34]990.90M4.05K1.84G20.19K
InfNoise [34]990.90M4.05K1.84G20.19K
LDU [10]264.24M44.04M613.42M102.24M
HyperDUM111.84M1.15M338.40M38.85M
- w/o Patch (4x) Proj.84.00M1.13M256.00M38.78M
- w/o Chan Proj.6.96M3.38K20.60M16.83K
+ +Table 8. Comparing inference (FLOPs) and training costs (# of Trainable Parameters). + +Finally, the main benefit of HyperDUM is its efficiency for training and inference. We measure the computation costs in terms of Floating point operations (FLOPs) and the number of trainable model parameters in Table 8. We approximate the FLOPs by computing the major contributing operations. Refer to the Appendix for our FLOPs approximations + +for each operation type. For InfMCD and InfNoise these are the entropy, mutual information, and predictive entropy computations. For LDU, it is the cosine similarity between prototypes and matrix multiplication for uncertainty, which was found to be negligible. For HyperDUM, it is the input projection and cosine similarity computations. We see that HyperDUM is up to $2.36 \times$ and $1.81 \times$ more FLOP-efficient compared to LDU for aiMotive and DeLiVER respectively. Compared to traditional Bayesian approximation methods, it can be up to $8.86 \times$ and $5.44 \times$ more FLOP-efficient. As for the training costs, HyperDUM has $38.30 \times$ and $2.63 \times$ less trainable parameters compared to LDU for aiMotive and DeLiVER respectively, making it more memory efficient. This ultimately shows the practicality of HyperDUM as a UQ method for real-world autonomous systems. + +# 5. Discussions and Conclusion + +Limitations and Future Work: We leveraged cloud computing resources using a single NVIDIA A100 GPU for training. Our method uses the traditional bundling technique, which requires labels for supervised learning of prototypes. The need for labels is the primary limitation of our approach. Although there have been implementations of semi-supervised and unsupervised versions [17, 18], there are no theoretical guarantees and/or justifications for their performance unlike the traditional bundling method [48]. We leave the exploration of these methods for uncertainty quantification to future work. Additionally, the FLOPs of the vector symbolic architectures may be further reduced due to the massively parallel operations of hyperdimensional computation. These optimizations were demonstrated in prior works [13, 19]. + +Conclusion: In this work, we propose an efficient method to quantify the epistemic uncertainty of machine learning models at the feature level. HyperDUM projects the latent features as hypervectors to form hyperdimensional prototypes. Similarity between hypervectors of new samples and learned prototypes is used as a proxy to estimate uncertainty. We demonstrate the practicality of HyperDUM through autonomous driving tasks, including object detection and semantic segmentation. Our results show that our method can enhance the performance and robustness of multimodal fusion models under uncertainties induced by various weather conditions and modality-specific corner cases. We achieve this by quantifying the feature uncertainty of each modality and applying an uncertainty-aware weighting layer prior to feature fusion and fine-tuning the post-fusion layers. Finally, we demonstrated that HyperDUM requires significantly less FLOPs and training parameters when compared to existing works, indicating its real-world practicality. + +Acknowledgements: This work acknowledges the support of the Automotive Research Center (ARC), Cooperative Agreement W56HZV-24-2-0001 U.S. Army DEVCOM Ground Vehicle Systems Center (GVSC). + +# References + +[1] Anastasios N Angelopoulos, Stephen Bates, et al. Conformal prediction: A gentle introduction. Foundations and Trends® in Machine Learning, 16(4):494-591, 2023. 1, 2 +[2] Sri Hrushikesh Varma Bhupathiraju, Jennifer Sheldon, Luke A Bauer, Vincent Bindschaedler, Takeshi Sugawara, and Sara Rampazzi. Emi-lidar: Uncovering vulnerabilities of lidar sensors in autonomous driving setting using electromagnetic interference. In Proceedings of the 16th ACM Conference on Security and Privacy in Wireless and Mobile Networks, pages 329-340, 2023. 1 +[3] Mario Bijelic, Tobias Gruber, Fahim Mannan, Florian Kraus, Werner Ritter, Klaus Dietmayer, and Felix Heide. Seeing through fog without seeing fog: Deep multimodal sensor fusion in unseen adverse weather. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11682-11692, 2020. 5, 7 +[4] Hermann Blum, Paul-Edouard Sarlin, Juan Nieto, Roland Siegwart, and Cesar Cadena. Fishyscapes: A benchmark for safe semantic segmentation in autonomous driving. In proceedings of the IEEE/CVF international conference on computer vision workshops, pages 0–0, 2019. 1 +[5] Bertrand Charpentier, Daniel Zügner, and Stephan Gümnmann. Posterior network: Uncertainty estimation without ood samples via density-based pseudo-counts. Advances in neural information processing systems, 33:1356-1367, 2020. 1, 2, 5, 6, 7 +[6] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 4 +[7] Michael Dusenberry, Ghassen Jerfel, Yeming Wen, Yian Ma, Jasper Snoek, Katherine Heller, Balaji Lakshminarayanan, and Dustin Tran. Efficient and scalable bayesian neural nets with rank-1 factors. In International conference on machine learning, pages 2782-2792. PMLR, 2020. 1, 2 +[8] Jamil Fayyad, Mohammad A Jaradat, Dominique Gruyer, and Homayoun Najjaran. Deep learning sensor fusion for autonomous vehicle perception and localization: A review. Sensors, 20(15):4220, 2020. 1, 2 +[9] Florian Fervers, Sebastian Bullinger, Christoph Bodensteiner, Michael Arens, and Rainer Stiefelhagen. Uncertainty-aware vision-based metric cross-view geolocation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21621-21631, 2023. 1 +[10] Gianni Franchi, Xuanlong Yu, Andrei Bursuc, Emanuel Aldea, Severine Dubuisson, and David Filliat. Latent discriminant deterministic uncertainty. In European Conference on Computer Vision, pages 243-260. Springer, 2022. 1, 2, 3, 5, 6, 7, 8 +[11] Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In international conference on machine learning, pages 1050-1059. PMLR, 2016. 1, 2 +[12] Ethan Goan and Clinton Fookes. Uncertainty in real-time semantic segmentation on embedded systems. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4490-4500, 2023. 1 + +[13] Saransh Gupta, Justin Morris, Mohsen Imani, Ranganathan Ramkumar, Jeffrey Yu, Aniket Tiwari, Baris Aksanli, and Tajana Simunic Rosing. Thrifty: Training with hyperdimensional computing across flash hierarchy. In Proceedings of the 39th International Conference on Computer-Aided Design, pages 1-9, 2020. 8 +[14] Zongbo Han, Fan Yang, Junzhou Huang, Changqing Zhang, and Jianhua Yao. Multimodal dynamics: Dynamical fusion for trustworthy multimodal classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20707-20717, 2022. 2 +[15] Florian Heidecker, Jasmin Breitenstein, Kevin Rösch, Jonas Löhdefink, Maarten Bieshaar, Christoph Stiller, Tim Fingscheidt, and Bernhard Sick. An application-driven conceptualization of corner cases for perception in highly automated driving. In 2021 IEEE Intelligent Vehicles Symposium (IV), pages 644-651. IEEE, 2021. 1 +[16] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17853-17862, 2023. 1 +[17] Mohsen Imani, Samuel Bosch, Mojan Javaheripi, Bita Rouhani, Xinyu Wu, Farinaz Koushanfar, and Tajana Rosing. Semihd: Semi-supervised learning using hyperdimensional computing. In 2019 IEEE/ACM International Conference on Computer-Aided Design (ICCAD), pages 1-8. IEEE, 2019. 8 +[18] Mohsen Imani, Yeseong Kim, Thomas Worley, Saransh Gupta, and Tajana Rosing. Hdcluster: An accurate clustering using brain-inspired high-dimensional computing. In 2019 Design, Automation & Test in Europe Conference & Exhibition (DATE), pages 1591-1594. IEEE, 2019. 8 +[19] Mohsen Imani, Saikishan Pampana, Saransh Gupta, Minxuan Zhou, Yeseong Kim, and Tajana Rosing. Dual: Acceleration of clustering algorithms using digital-based processing in-memory. In 2020 53rd Annual IEEE/ACM International Symposium on Microarchitecture (MICRO), pages 356-371. IEEE, 2020. 8 +[20] Masha Itkina and Mykel Kochenderfer. Interpretable self-aware neural networks for robust trajectory prediction. In Conference on Robot Learning, pages 606-617. PMLR, 2023. 5 +[21] Ding Jia, Jianyuan Guo, Kai Han, Han Wu, Chao Zhang, Chang Xu, and Xinghao Chen. Geminifusion: Efficient pixelwise multimodal fusion for vision transformer. International Conference on Machine Learning (ICML), 2024. 6 +[22] Xiaotao Jia, Jianlei Yang, Runze Liu, Xueyan Wang, Sorin Dan Cotofana, and Weisheng Zhao. Efficient computation reduction in bayesian neural networks through feature decomposition and memorization. IEEE transactions on neural networks and learning systems, 32(4):1703-1712, 2020. 1, 2 +[23] Yongchan Kwon, Joong-Ho Won, Beom Joon Kim, and Myunghee Cho Paik. Uncertainty quantification using bayesian neural networks in classification: Application to biomedical image segmentation. Computational Statistics & Data Analysis, 142:106816, 2020. 1 + +[24] Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scalable predictive uncertainty estimation using deep ensembles. Advances in neural information processing systems, 30, 2017. 1, 2 +[25] Simon D Levy and Ross Gayler. Vector symbolic architectures: A new building material for artificial general intelligence. In Proceedings of the 2008 conference on artificial general intelligence 2008: Proceedings of the first AGI conference, pages 414-418, 2008. 3 +[26] Hao Li, Jingkuan Song, Lianli Gao, Xiaosu Zhu, and Hengtao Shen. Prototype-based aleatoric uncertainty quantification for cross-modal retrieval. Advances in Neural Information Processing Systems, 36, 2024. 3 +[27] Xiaotong Li, Yongxing Dai, Yixiao Ge, Jun Liu, Ying Shan, and Ling-Yu Duan. Uncertainty modeling for out-of-distribution generalization. arXiv preprint arXiv:2202.03958, 2022.4 +[28] You Li and Javier Ibanez-Guzman. Lidar for autonomous driving: The principles, challenges, and trends for automotive lidar and perception systems. IEEE Signal Processing Magazine, 37(4):50-61, 2020. 1 +[29] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1477-1485, 2023. 5 +[30] Ming Liang, Bin Yang, Shenlong Wang, and Raquel Urtasun. Deep continuous fusion for multi-sensor 3d object detection. In Proceedings of the European conference on computer vision (ECCV), pages 641-656, 2018. 2 +[31] Arnav Vaibhav Malawade, Trier Mortlock, and Mohammad Abdullah Al Faruque. Hydraulics: Context-aware selective sensor fusion for robust and efficient autonomous vehicle perception. In 2022 ACM/IEEE 13th International Conference on Cyber-Physical Systems (ICCPs), pages 68–79. IEEE, 2022. 1, 2 +[32] Andrey Malinin and Mark Gales. Predictive uncertainty estimation via prior networks. Advances in neural information processing systems, 31, 2018. 2 +[33] Tamas Matuszka, Ivan Barton, Adam Butykai, Peter Hajas, David Kiss, Domonkos Kovacs, Sandor Kunsagi-Mate, Peter Lengyel, Gabor Nemeth, Leventa PetHo, et al. aimotive dataset: A multimodal dataset for robust autonomous driving with long-range perception. arXiv preprint arXiv:2211.09445, 2022. 5, 6 +[34] Lu Mi, Hao Wang, Yonglong Tian, Hao He, and Nir N Shavit. Training-free uncertainty estimation for dense regression: Sensitivity as a surrogate. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10042-10050, 2022. 1, 2, 5, 6, 7, 8 +[35] Trier Mortlock, Arnav Malawade, Kohei Tsujio, and Mohammad Al Faruque. Castnet: a context-aware, spatio-temporal dynamic motion prediction ensemble for autonomous driving. ACM Transactions on Cyber-Physical Systems, 8(2):1–20, 2024. 1 +[36] Jishnu Mukhoti, Andreas Kirsch, Joost van Amersfoort, Philip HS Torr, and Yarin Gal. Deep deterministic uncertainty: + +A new simple baseline. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24384-24394, 2023. 1, 2 +[37] Yang Ni, Hanning Chen, Prathyush Poduval, Zhuowen Zou, Pietro Mercati, and Mohsen Imani. Brain-inspired trustworthy hyperdimensional computing with efficient uncertainty quantification. In 2023 IEEE/ACM International Conference on Computer Aided Design (ICCAD), pages 01-09. IEEE, 2023. 2 +[38] Igor Nunes, Mike Heddes, Tony Givargis, Alexandru Nicolau, and Alex Veidenbaum. Graphhd: Efficient graph classification using hyperdimensional computing. In 2022 Design, Automation & Test in Europe Conference & Exhibition (DATE), pages 1485-1490. IEEE, 2022. 3 +[39] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch. In NIPS-W, 2017. 5 +[40] Luis Manuel Pereira, Addisson Salazar, and Luis Vergara. On comparing early and late fusion methods. In International Work-Conference on Artificial Neural Networks, pages 365-378. Springer, 2023. 2 +[41] Janis Postels, Francesco Ferroni, Huseyin Coskun, Nassir Navab, and Federico Tombari. Sampling-free epistemic uncertainty estimation using approximated variance propagation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2931-2940, 2019. 1 +[42] Janis Postels, Mattia Segu, Tao Sun, Luca Sieber, Luc Van Gool, Fisher Yu, and Federico Tombari. On the practicality of deterministic epistemic uncertainty. arXiv preprint arXiv:2107.00649, 2021. 1, 2 +[43] Ali Rahimi and Benjamin Recht. Random features for large-scale kernel machines. Advances in neural information processing systems, 20, 2007. 1 +[44] Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Schölkopf, Thomas Brox, and Peter Gehler. Towards total recall in industrial anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14318-14328, 2022. 4 +[45] Kenny Schlegel, Peer Neubert, and Peter Protzel. A comparison of vector symbolic architectures. Artificial Intelligence Review, 55(6):4523-4555, 2022. 3 +[46] Jiahui She, Yibo Hu, Hailin Shi, Jun Wang, Qiu Shen, and Tao Mei. Dive into ambiguity: Latent distribution mining and pairwise uncertainty estimation for facial expression recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6248-6257, 2021. 2 +[47] Apoorv Singh. Transformer-based sensor fusion for autonomous driving: A survey. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3312-3317, 2023. 2 +[48] Anthony Thomas, Sanjoy Dasgupta, and Tajana Rosing. A theoretical perspective on hyperdimensional computing. Journal of Artificial Intelligence Research, 72:215-249, 2021. 3, 4, 8, 1 +[49] Junjiao Tian, Wesley Cheung, Nathaniel Glaser, Yen-Cheng Liu, and Zsolt Kira. Uno: Uncertainty-aware noisy-or multimodal fusion for unanticipated input degradation. In 2020 + +IEEE International Conference on Robotics and Automation (ICRA), pages 5716-5723. IEEE, 2020. 5, 2 +[50] Stefan Van der Walt, Johannes L Schonberger, Juan Nunez-Iglesias, François Boulogne, Joshua D Warner, Neil Yager, Emmanuelle Gouillart, and Tony Yu. scikit-image: image processing in python. PeerJ, 2:e453, 2014. 5 +[51] Jorge Vargas, Suleiman Alsweiss, Onur Token, Rahul Razdan, and Joshua Santos. An overview of autonomous vehicles sensors and their vulnerability to weather conditions. Sensors, 21(16):5397, 2021. 1 +[52] Junyao Wang, Sitao Huang, and Mohsen Imani. Disthd: A learner-aware dynamic encoding method for hyperdimensional classification. In 2023 60th ACM/IEEE Design Automation Conference (DAC), pages 1-6. IEEE, 2023. 4 +[53] Junyao Wang, Arnav Vaibhav Malawade, Junhong Zhou, Shih-Yuan Yu, and Mohammad Abdullah Al Faruque. Rs2g: Data-driven scene-graph extraction and embedding for robust autonomous perception and scenario understanding. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 7493-7502, 2024. 1 +[54] Shaojie Wang, Tong Wu, Ayan Chakrabarti, and Yevgeniy Vorobeychik. Adversarial robustness of deep sensor fusion models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2387-2396, 2022. 1 +[55] Samuel Wilson, Tobias Fischer, Niko Sünderhauf, and Feras Dayoub. Hyperdimensional feature fusion for out-of-distribution detection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2644-2654, 2023. 4 +[56] De Jong Yeong, Gustavo Velasco-Hernandez, John Barry, and Joseph Walsh. Sensor and sensor fusion technology in autonomous vehicles: A review. Sensors, 21(6):2140, 2021. 1, 2 +[57] Tao Yu, Yichi Zhang, Zhiru Zhang, and Christopher M De Sa. Understanding hyperdimensional computing for parallel single-pass learning. Advances in Neural Information Processing Systems, 35:1157-1169, 2022. 3, 8, 1, 7 +[58] Jiaming Zhang, Ruiping Liu, Hao Shi, Kailun Yang, Simon Reiβ, Kunyu Peng, Haodong Fu, Kaiwei Wang, and Rainer Stiefelhagen. Delivering arbitrary-modal semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1136-1147, 2023. 5, 6, 7 +[59] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4490-4499, 2018. 5 +[60] Zijian Zhu, Yichi Zhang, Hai Chen, Yinpeng Dong, Shu Zhao, Wenbo Ding, Jiachen Zhong, and Shibao Zheng. Understanding the robustness of 3d object detection with bird's-eye-view representations in autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21600-21610, 2023. 1 \ No newline at end of file diff --git a/2025/Hyperdimensional Uncertainty Quantification for Multimodal Uncertainty Fusion in Autonomous Vehicles Perception/images.zip b/2025/Hyperdimensional Uncertainty Quantification for Multimodal Uncertainty Fusion in Autonomous Vehicles Perception/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f7e0f8bc24ad2d73b21ad3af241a8a6299004aad --- /dev/null +++ b/2025/Hyperdimensional Uncertainty Quantification for Multimodal Uncertainty Fusion in Autonomous Vehicles Perception/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d7ac132b09d1a4f640688b2ec1457b60c37eb31a295d8c4fb0bc460fbeee1d7e +size 691963 diff --git a/2025/Hyperdimensional Uncertainty Quantification for Multimodal Uncertainty Fusion in Autonomous Vehicles Perception/layout.json b/2025/Hyperdimensional Uncertainty Quantification for Multimodal Uncertainty Fusion in Autonomous Vehicles Perception/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c9fd88c493247b98b1addb1c62fbb5a8c15d74aa --- /dev/null +++ b/2025/Hyperdimensional Uncertainty Quantification for Multimodal Uncertainty Fusion in Autonomous Vehicles Perception/layout.json @@ -0,0 +1,9008 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 56, + 102, + 555, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 102, + 555, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 102, + 555, + 140 + ], + "type": "text", + "content": "Hyperdimensional Uncertainty Quantification for Multimodal Uncertainty Fusion in Autonomous Vehicles Perception" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 161, + 553, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 161, + 553, + 190 + ], + "spans": [ + { + "bbox": [ + 61, + 161, + 553, + 190 + ], + "type": "text", + "content": "Luke Chen* Junyao Wang Trier Mortlock Pramod Khargonekar Mohammad Abdullah Al Faruque University of California, Irvine" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 121, + 191, + 489, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 191, + 489, + 204 + ], + "spans": [ + { + "bbox": [ + 121, + 191, + 489, + 204 + ], + "type": "text", + "content": "{panwangc, junyaow4, tmortloc, pramod.khargonekar, alfaruqu}@uci.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 256, + 297, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 256, + 297, + 555 + ], + "spans": [ + { + "bbox": [ + 54, + 256, + 297, + 555 + ], + "type": "text", + "content": "Uncertainty Quantification (UQ) is crucial for ensuring the reliability of machine learning models deployed in real-world autonomous systems. However, existing approaches typically quantify task-level output prediction uncertainty without considering epistemic uncertainty at the multimodal feature fusion level, leading to sub-optimal outcomes. Additionally, popular uncertainty quantification methods, e.g., Bayesian approximations, remain challenging to deploy in practice due to high computational costs in training and inference. In this paper, we propose HyperDUM, a novel deterministic uncertainty method (DUM) that efficiently quantifies feature-level epistemic uncertainty by leveraging hyperdimensional computing. Our method captures the channel and spatial uncertainties through channel and patch-wise projection and bundling techniques respectively. Multimodal sensor features are then adaptively weighted to mitigate uncertainty propagation and improve feature fusion. Our evaluations show that HyperDUM on average outperforms the state-of-the-art (SOTA) algorithms by up to " + }, + { + "bbox": [ + 54, + 256, + 297, + 555 + ], + "type": "inline_equation", + "content": "2.01\\% / 1.27\\%" + }, + { + "bbox": [ + 54, + 256, + 297, + 555 + ], + "type": "text", + "content": " in 3D Object Detection and up to " + }, + { + "bbox": [ + 54, + 256, + 297, + 555 + ], + "type": "inline_equation", + "content": "1.29\\%" + }, + { + "bbox": [ + 54, + 256, + 297, + 555 + ], + "type": "text", + "content": " improvement over baselines in semantic segmentation tasks under various types of uncertainties. Notably, HyperDUM requires " + }, + { + "bbox": [ + 54, + 256, + 297, + 555 + ], + "type": "inline_equation", + "content": "2.36 \\times" + }, + { + "bbox": [ + 54, + 256, + 297, + 555 + ], + "type": "text", + "content": " less Floating Point Operations and up to " + }, + { + "bbox": [ + 54, + 256, + 297, + 555 + ], + "type": "inline_equation", + "content": "38.30 \\times" + }, + { + "bbox": [ + 54, + 256, + 297, + 555 + ], + "type": "text", + "content": " less parameters than SOTA methods, providing an efficient solution for real-world autonomous systems." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 579, + 135, + 593 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 135, + 593 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 135, + 593 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 600, + 296, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 296, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 296, + 684 + ], + "type": "text", + "content": "Effectively quantifying and addressing uncertainty plays a critical role in ensuring the reliability of ML models in real-world applications, especially where the decisions are consequential, e.g., autonomous driving. Models failing to address uncertainty can produce overconfident outputs, leading to compromised performance [9, 16, 41, 53]. In light of autonomous systems, uncertainties can arise from adverse" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 232, + 556, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 556, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 556, + 422 + ], + "type": "text", + "content": "weather conditions, sensor failures or noises, environment dynamics (movements of other actors, road conditions, vehicle dynamics), and various corner cases [4, 12, 35, 36]. Modern autonomous systems leverage multimodal sensor fusion to broaden the information scope and learn comprehensive representations of the surrounding environment [8, 31, 56]. However, commonly employed sensors such as cameras, lidar, and radar, exhibit varying degrees of robustness and vulnerability across different driving scenarios [51, 54]. For instance, cameras are crucial in delivering rich perceptual information while they are particularly vulnerable to lighting variations and occlusions [60]. Lidar and radar are more robust to visual perturbations but can be affected by multi-path interference and occlusion [2, 28]. Therefore, it is important to consider the uncertainty of each modality before the fusion of multimodal features to achieve optimal results [15]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 425, + 556, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 425, + 556, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 425, + 556, + 664 + ], + "type": "text", + "content": "Popular Uncertainty Quantification (UQ) methods include deep ensembles, Bayesian neural networks, and Bayesian approximation methods such as Monte Carlo dropout [11, 23, 24]. However, these methods often introduce large training and/or inference overheads compounded by the growing complexity of modern Deep Neural Networks (DNN), ultimately challenging their practicality under real-time computing constraints [7, 22]. Other lines of works such as Conformal Prediction (CP) and Posterior Networks (PostNets) based on evidential deep learning either cannot quantify feature-level epistemic uncertainty or require vast amounts of quality data to accurately estimate the posteriors [1, 5]. Deterministic Uncertainty Methods (DUMs), aiming to efficiently quantify uncertainty using only a single forward pass [10, 34, 42], have become an emerging trend. DUMs rely on statistical/morphological characteristics of latent features to quantify epistemic uncertainty. For instance, [10] employed Prototype Learning (PL) techniques to capture the \"knowledge\" of DNNs by constructing a \"memory bank\" of prototypes to predict uncertainty." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "To address these problems, we propose HyperDUM, a novel deterministic uncertainty method for efficient quantification of feature-level epistemic uncertainty. In contrast to existing methods, HyperDUM projects latent features into" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 693, + 295, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 712 + ], + "type": "text", + "content": "*Corresponding Author. Distribution A: Approved for public release; distribution unlimited. OPSEC # 9537" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22306" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 108 + ], + "type": "text", + "content": "hyperdimensional space to form hyperdimensional prototypes, and then estimates uncertainty based on the distance between features of unseen samples and these prototypes." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 109, + 297, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 109, + 297, + 242 + ], + "spans": [ + { + "bbox": [ + 55, + 109, + 297, + 242 + ], + "type": "text", + "content": "To the best of our knowledge, HyperDUM is the first deterministic uncertainty quantification method utilizing hyperdimensional feature prototyping with Channel-wise and Patch-wise projection and bundling to deliver accurate and efficient epistemic uncertainty quantification of latent features. Detailed experimental evaluations on the DeLiVER and aiMotive datasets for semantic segmentation and object detection tasks demonstrate that HyperDUM can efficiently and effectively quantify feature uncertainty and improve fusion performance when compared to standard Bayesian approximation inference baselines and SOTA methods." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 254, + 148, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 254, + 148, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 254, + 148, + 266 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 274, + 205, + 288 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 274, + 205, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 274, + 205, + 288 + ], + "type": "text", + "content": "2.1. Uncertainty Quantification" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 293, + 297, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 293, + 297, + 390 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 297, + 390 + ], + "type": "text", + "content": "Predictive uncertainty is composed of three factors, model/epistemic uncertainty, data/aleatoric uncertainty, and distributional uncertainty [32]. Epistemic uncertainty, stemming from the word epistemology which is the study of knowledge, is associated with a model's lack of knowledge of underrepresented data. The scope of this work is to efficiently quantify epistemic uncertainty at the latent feature level to enable uncertainty-weighting prior to feature fusion." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 391, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 391, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 391, + 297, + 713 + ], + "type": "text", + "content": "Existing UQ approaches are typically derived from Bayesian principles, including Bayesian neural networks which represent model weights as probability distributions, deep ensembles which sample output distributions from diverse models [24], and Monte Carlo dropout which is similar to deep ensembles but utilizes a single model with dropout layers [11]. However, these approaches remain challenging in practice due to enormous training and inference costs [7, 22]. Conformal prediction (CP) is a model-agnostic UQ technique that provides prediction intervals with guaranteed coverage, ensuring the true value falls within the interval at a specified probability [1]. However, while effective in producing reliable task-level uncertainties, CP does not capture uncertainties at intermediate model abstractions, a limitation especially evident in multimodal fusion models, where uncertainty propagates across different modalities before the final output. Evidential deep learning is another line of work, i.e. Posterior Networks which aim to quantify both aleatoric and epistemic uncertainty by directly estimating a posterior distribution over the model's outputs/predictions but is sensitive to data quality for accurately estimating the posteriors [5]. Deterministic uncertainty quantification is a growing research area that aims to efficiently compute epistemic uncertainty of DNNs with a single forward pass [10, 34, 42]. For instance, [34] is a training-free method that injects Gaussian noise into intermediate layers to compute the variance across the representations as a measure of uncertainty. La" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 556, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 556, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 556, + 228 + ], + "type": "text", + "content": "tent Deterministic Uncertainty (LDU) constructs prototypes from latent features by enforcing dissimilarity between prototypes and entropy between latent features while correlating uncertainty and the downstream task [10, 36, 46]. Recent research [37] demonstrated that constructing hyperdimensional prototypes can provide comparable performance to existing aleatoric uncertainty quantification solutions, e.g., Bayesian approximation, and offer significant speedups in training and inference. However, it is applied to regression tasks for aleatoric uncertainty. In contrast, our work quantifies epistemic uncertainty on the latent features, extending its applicability to multimodal fusion models and complex tasks such as object detection and semantic segmentation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 236, + 519, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 519, + 248 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 519, + 248 + ], + "type": "text", + "content": "2.2. Multimodal Uncertainty-Aware Fusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 254, + 556, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 254, + 556, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 254, + 556, + 445 + ], + "type": "text", + "content": "Multimodal sensor fusion aims to provide more comprehensive representations and address the limitations of unimodal sensor failure cases [8, 31, 56]. Existing approaches dealing with uncertainties fall into two categories: feature-level fusion and output-level fusion. Feature-level fusion, also known as intermediate fusion, applies an uncertainty weighting that scales the features to minimize uncertainty [30, 47]. For instance, [14] computes feature-informativeness and scales the original features based on informativeness. On the other hand, output-level fusion, also known as late-fusion, uses uncertainty to calibrate the outputs of the model, i.e. calibrating softmax probabilities [40]. In contrast to existing uncertainty-based fusion works that commonly rely on Bayesian approximated uncertainty fusion, our proposed HyperDUM is the first to investigate deterministic uncertainty fusion methods in a multimodal autonomous systems setting." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 456, + 397, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 397, + 470 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 397, + 470 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 475, + 400, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 400, + 488 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 400, + 488 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 494, + 556, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 494, + 556, + 686 + ], + "spans": [ + { + "bbox": [ + 313, + 494, + 556, + 686 + ], + "type": "text", + "content": "Given an input sample " + }, + { + "bbox": [ + 313, + 494, + 556, + 686 + ], + "type": "inline_equation", + "content": "\\mathbf{x} = \\{x_{1}, x_{2}, \\dots, x_{M}\\}" + }, + { + "bbox": [ + 313, + 494, + 556, + 686 + ], + "type": "text", + "content": " consisting of " + }, + { + "bbox": [ + 313, + 494, + 556, + 686 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 494, + 556, + 686 + ], + "type": "text", + "content": " multimodal sensor inputs. We have a multimodal model " + }, + { + "bbox": [ + 313, + 494, + 556, + 686 + ], + "type": "inline_equation", + "content": "\\mathbb{M}" + }, + { + "bbox": [ + 313, + 494, + 556, + 686 + ], + "type": "text", + "content": " with pre-trained feature encoders " + }, + { + "bbox": [ + 313, + 494, + 556, + 686 + ], + "type": "inline_equation", + "content": "\\mathbf{f} = \\{f_{1}, \\dots, f_{M}\\}" + }, + { + "bbox": [ + 313, + 494, + 556, + 686 + ], + "type": "text", + "content": " that extract modality-specific abstract features " + }, + { + "bbox": [ + 313, + 494, + 556, + 686 + ], + "type": "inline_equation", + "content": "z_{m} = f_{m}(x_{m})" + }, + { + "bbox": [ + 313, + 494, + 556, + 686 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 494, + 556, + 686 + ], + "type": "inline_equation", + "content": "1 \\leq m \\leq M" + }, + { + "bbox": [ + 313, + 494, + 556, + 686 + ], + "type": "text", + "content": ". Additionally, the model contains fusion blocks " + }, + { + "bbox": [ + 313, + 494, + 556, + 686 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(z_{i}, z_{j})" + }, + { + "bbox": [ + 313, + 494, + 556, + 686 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 494, + 556, + 686 + ], + "type": "inline_equation", + "content": "i, j \\in \\{1, \\dots, M\\}" + }, + { + "bbox": [ + 313, + 494, + 556, + 686 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 313, + 494, + 556, + 686 + ], + "type": "inline_equation", + "content": "i \\neq j" + }, + { + "bbox": [ + 313, + 494, + 556, + 686 + ], + "type": "text", + "content": " which combine modality-specific features to learn cross-modal fusion features " + }, + { + "bbox": [ + 313, + 494, + 556, + 686 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 313, + 494, + 556, + 686 + ], + "type": "text", + "content": ". Assuming there are " + }, + { + "bbox": [ + 313, + 494, + 556, + 686 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 494, + 556, + 686 + ], + "type": "text", + "content": " fusion blocks, the fused features " + }, + { + "bbox": [ + 313, + 494, + 556, + 686 + ], + "type": "inline_equation", + "content": "\\mathbf{g} = \\{g_{1}, \\dots, g_{K}\\}" + }, + { + "bbox": [ + 313, + 494, + 556, + 686 + ], + "type": "text", + "content": " are passed to the task head " + }, + { + "bbox": [ + 313, + 494, + 556, + 686 + ], + "type": "inline_equation", + "content": "h(g)" + }, + { + "bbox": [ + 313, + 494, + 556, + 686 + ], + "type": "text", + "content": " to output the corresponding task predictions " + }, + { + "bbox": [ + 313, + 494, + 556, + 686 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 494, + 556, + 686 + ], + "type": "text", + "content": " (i.e. bounding boxes and object classification for object detection, and pixel-wise class labels for semantic segmentation). To account for feature uncertainty, an uncertainty quantification module " + }, + { + "bbox": [ + 313, + 494, + 556, + 686 + ], + "type": "inline_equation", + "content": "\\mathbb{U}_{m}(z_{m})" + }, + { + "bbox": [ + 313, + 494, + 556, + 686 + ], + "type": "text", + "content": " is inserted immediately following each feature extractor to predict an uncertainty value " + }, + { + "bbox": [ + 313, + 494, + 556, + 686 + ], + "type": "inline_equation", + "content": "u_{m}" + }, + { + "bbox": [ + 313, + 494, + 556, + 686 + ], + "type": "text", + "content": ". Given the uncertainty value, a learnable" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 693, + 556, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 693, + 556, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 693, + 556, + 712 + ], + "type": "text", + "content": "1Cross-modal feature alignment is not required but may be enforced by the backbone." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22307" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 75, + 549, + 240 + ], + "blocks": [ + { + "bbox": [ + 58, + 75, + 549, + 240 + ], + "lines": [ + { + "bbox": [ + 58, + 75, + 549, + 240 + ], + "spans": [ + { + "bbox": [ + 58, + 75, + 549, + 240 + ], + "type": "image", + "image_path": "d5504aea6ccdf5201820be0635a2f39dd6045f3706781f2363bc9e893a3a7f54.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 79, + 251, + 529, + 262 + ], + "lines": [ + { + "bbox": [ + 79, + 251, + 529, + 262 + ], + "spans": [ + { + "bbox": [ + 79, + 251, + 529, + 262 + ], + "type": "text", + "content": "Figure 1. Multimodal model with uncertainty quantification and uncertainty fusion for autonomous driving perception tasks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 283, + 296, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 283, + 296, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 283, + 296, + 380 + ], + "type": "text", + "content": "uncertainty weighting module " + }, + { + "bbox": [ + 55, + 283, + 296, + 380 + ], + "type": "inline_equation", + "content": "\\Omega (z_m,u_m)" + }, + { + "bbox": [ + 55, + 283, + 296, + 380 + ], + "type": "text", + "content": " reweights the input features to produce an uncertainty-aware feature " + }, + { + "bbox": [ + 55, + 283, + 296, + 380 + ], + "type": "inline_equation", + "content": "\\hat{z}_m" + }, + { + "bbox": [ + 55, + 283, + 296, + 380 + ], + "type": "text", + "content": " for each modality. Thus, the input to the fusion blocks now becomes " + }, + { + "bbox": [ + 55, + 283, + 296, + 380 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\hat{z}_i,\\hat{z}_j)" + }, + { + "bbox": [ + 55, + 283, + 296, + 380 + ], + "type": "text", + "content": " to produce uncertainty-weighted fusion features " + }, + { + "bbox": [ + 55, + 283, + 296, + 380 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{g}}" + }, + { + "bbox": [ + 55, + 283, + 296, + 380 + ], + "type": "text", + "content": " which are forwarded to be fused and subsequently used by the task head " + }, + { + "bbox": [ + 55, + 283, + 296, + 380 + ], + "type": "inline_equation", + "content": "h(\\hat{\\mathbf{g}})" + }, + { + "bbox": [ + 55, + 283, + 296, + 380 + ], + "type": "text", + "content": " to produce the outputs. The system flow of the described multimodal fusion model with uncertainty quantification is illustrated in Figure 1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 395, + 296, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 395, + 296, + 419 + ], + "spans": [ + { + "bbox": [ + 55, + 395, + 296, + 419 + ], + "type": "text", + "content": "3.1.1 Vector Symbolic Architectures/Hyperdimensional Computing" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 426, + 296, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 296, + 665 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 296, + 665 + ], + "type": "text", + "content": "Vector Symbolic Architectures (VSA) / Hyperdimensional Computing (HDC) are a class of algorithms for creating and handling hyperdimensional representations to mimic cognitive functions [25, 45]. VSA starts with mapping input data to hyperdimensional vectors, or hypervectors " + }, + { + "bbox": [ + 56, + 426, + 296, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 56, + 426, + 296, + 665 + ], + "type": "text", + "content": ", through a projection function " + }, + { + "bbox": [ + 56, + 426, + 296, + 665 + ], + "type": "inline_equation", + "content": "\\phi: \\mathcal{Z} \\to \\mathbf{H}" + }, + { + "bbox": [ + 56, + 426, + 296, + 665 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 56, + 426, + 296, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{Z} \\in \\mathbb{R}^n" + }, + { + "bbox": [ + 56, + 426, + 296, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 426, + 296, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{H} \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 56, + 426, + 296, + 665 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 56, + 426, + 296, + 665 + ], + "type": "inline_equation", + "content": "d \\gg n" + }, + { + "bbox": [ + 56, + 426, + 296, + 665 + ], + "type": "text", + "content": ") is the hypervector space. The projection function depends on the input data representation. This work assumes Euclidean input space. For non-euclidean see [38]. Following the projection, common operations on hypervectors include similarity calculation, bundling, and binding. Similarity " + }, + { + "bbox": [ + 56, + 426, + 296, + 665 + ], + "type": "inline_equation", + "content": "(\\delta(\\cdot, \\cdot))" + }, + { + "bbox": [ + 56, + 426, + 296, + 665 + ], + "type": "text", + "content": " calculation measures the distance between two hypervectors. For real-valued hypervectors, a common measure is cosine similarity. Bundling " + }, + { + "bbox": [ + 56, + 426, + 296, + 665 + ], + "type": "inline_equation", + "content": "(\\bigoplus)" + }, + { + "bbox": [ + 56, + 426, + 296, + 665 + ], + "type": "text", + "content": " is an element-wise addition of hypervectors, e.g., " + }, + { + "bbox": [ + 56, + 426, + 296, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{H}^{bundle} = \\mathcal{H}^1 \\bigoplus \\mathcal{H}^2" + }, + { + "bbox": [ + 56, + 426, + 296, + 665 + ], + "type": "text", + "content": ", generating a hypervector with the same dimension as inputs. Binding " + }, + { + "bbox": [ + 56, + 426, + 296, + 665 + ], + "type": "inline_equation", + "content": "(\\otimes)" + }, + { + "bbox": [ + 56, + 426, + 296, + 665 + ], + "type": "text", + "content": " is an element-wise multiplication associating two hypervectors to create another near-orthogonal hypervector, e.g., " + }, + { + "bbox": [ + 56, + 426, + 296, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{H}^{bind} = \\mathcal{H}^1 \\bigotimes \\mathcal{H}^2" + }, + { + "bbox": [ + 56, + 426, + 296, + 665 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 56, + 426, + 296, + 665 + ], + "type": "inline_equation", + "content": "\\delta(\\mathcal{H}^{bind}, \\mathcal{H}^1) \\approx 0" + }, + { + "bbox": [ + 56, + 426, + 296, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 426, + 296, + 665 + ], + "type": "inline_equation", + "content": "\\delta(\\mathcal{H}^{bind}, \\mathcal{H}^2) \\approx 0" + }, + { + "bbox": [ + 56, + 426, + 296, + 665 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Our work mainly leverages the bundling operation of VSA for the formation of prototypes. In high-dimensional space, bundling mimics how human brains memorize information [48, 57]. For instance, for " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{H}^{bundle} = \\mathcal{H}^1 +\\mathcal{H}^2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 281, + 555, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 555, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 555, + 389 + ], + "type": "text", + "content": "we have " + }, + { + "bbox": [ + 313, + 281, + 555, + 389 + ], + "type": "inline_equation", + "content": "\\delta (\\mathcal{H}^{bundle},\\mathcal{H}^1)\\gg 0" + }, + { + "bbox": [ + 313, + 281, + 555, + 389 + ], + "type": "text", + "content": " while " + }, + { + "bbox": [ + 313, + 281, + 555, + 389 + ], + "type": "inline_equation", + "content": "\\delta (\\mathcal{H}^{bundle},\\mathcal{H}^3)\\approx 0" + }, + { + "bbox": [ + 313, + 281, + 555, + 389 + ], + "type": "inline_equation", + "content": "(\\mathcal{H}^3\\neq \\mathcal{H}^1,\\mathcal{H}^2)" + }, + { + "bbox": [ + 313, + 281, + 555, + 389 + ], + "type": "text", + "content": " . In other words, VSA can generate prototypical representations by bundling the high-dimensional vectors of individual data samples, making it especially useful for prototype learning. Specifically, suppose we have labeled samples " + }, + { + "bbox": [ + 313, + 281, + 555, + 389 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{(z_1^1,y_1^1),(z_1^2,y_1^2),\\ldots ,(z_M^N,y_M^N)\\}" + }, + { + "bbox": [ + 313, + 281, + 555, + 389 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 281, + 555, + 389 + ], + "type": "inline_equation", + "content": "z_{m}^{i}\\in \\mathcal{Z}" + }, + { + "bbox": [ + 313, + 281, + 555, + 389 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 281, + 555, + 389 + ], + "type": "inline_equation", + "content": "y_{m}^{i}\\in \\{l\\}_{l = 1}^{L}" + }, + { + "bbox": [ + 313, + 281, + 555, + 389 + ], + "type": "text", + "content": " , we can form " + }, + { + "bbox": [ + 313, + 281, + 555, + 389 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 281, + 555, + 389 + ], + "type": "text", + "content": " prototypes by bundling hypervectors corresponding to a particular label:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 389, + 388, + 555, + 415 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 388, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 389, + 388, + 555, + 415 + ], + "type": "interline_equation", + "content": "\\mathcal {H} _ {m} ^ {l} = \\bigoplus_ {i: y ^ {i} = l} \\phi_ {m} \\left(z _ {m} ^ {i}\\right) \\tag {1}", + "image_path": "40c9b0c9a5615e4287c8939240949cb8c13e1b20601ef2b2c2a1fdd3ebf0d82d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 417, + 556, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 417, + 556, + 490 + ], + "spans": [ + { + "bbox": [ + 313, + 417, + 556, + 490 + ], + "type": "text", + "content": "where each hyperdimensional prototype can represent an aspect of the data samples " + }, + { + "bbox": [ + 313, + 417, + 556, + 490 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 417, + 556, + 490 + ], + "type": "text", + "content": " depending on the label definition, i.e. class label (cat and dog) or context label (weather conditions which we use in this work). We then define the notion of uncertainty as the set of similarity between hypervectors and hyperdimensional prototype of modality " + }, + { + "bbox": [ + 313, + 417, + 556, + 490 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 417, + 556, + 490 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 346, + 505, + 555, + 537 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 505, + 555, + 537 + ], + "spans": [ + { + "bbox": [ + 346, + 505, + 555, + 537 + ], + "type": "interline_equation", + "content": "\\mathbb {U} _ {m} = \\bigcup_ {l = 1} ^ {L} \\left\\{\\delta \\left(\\mathcal {H} _ {m} ^ {z}, \\mathcal {H} _ {m} ^ {l}\\right) \\right\\}, \\text {w h e r e} | \\mathbb {U} _ {m} | = L \\tag {2}", + "image_path": "2dd0699d470bb678c853bd9328afdb6421e7cd26b47567889f88357036af9c42.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 543, + 556, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 543, + 556, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 543, + 556, + 639 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 543, + 556, + 639 + ], + "type": "inline_equation", + "content": "\\mathbb{U}_m" + }, + { + "bbox": [ + 313, + 543, + 556, + 639 + ], + "type": "text", + "content": " is the set of similarity distances between a hypervector " + }, + { + "bbox": [ + 313, + 543, + 556, + 639 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_m^z" + }, + { + "bbox": [ + 313, + 543, + 556, + 639 + ], + "type": "text", + "content": " and each hyperdimensional prototype " + }, + { + "bbox": [ + 313, + 543, + 556, + 639 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_m^l" + }, + { + "bbox": [ + 313, + 543, + 556, + 639 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 543, + 556, + 639 + ], + "type": "inline_equation", + "content": "l = 1,2,\\dots ,L" + }, + { + "bbox": [ + 313, + 543, + 556, + 639 + ], + "type": "text", + "content": " and used to express the similarity uncertainty. The notion of similarity for uncertainty has been used in other prototype learning methods [10, 26], where the former assigns the variations of similarities between an instance and the prototypes as the belief masses and the latter utilizes similarity to optimize the uncertainty loss." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 645, + 385, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 645, + 385, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 645, + 385, + 658 + ], + "type": "text", + "content": "3.2. HyperDUM" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 663, + 555, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 663, + 555, + 688 + ], + "spans": [ + { + "bbox": [ + 313, + 663, + 555, + 688 + ], + "type": "text", + "content": "Following the conventions of VSA for projection and bundling, we propose two modifications for uncertainty quan-" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 693, + 555, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 693, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 693, + 555, + 713 + ], + "type": "text", + "content": "2We derive uncertainty quantifiability from similarity/expressivity of VSAs/HDCs based on recent theorems in the Appendix." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22308" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": "tification: 1) Channel-wise and 2) Patch-wise Projection & Bundling (CPB/PPB). CPB modifies the projection and bundling operations to capture the uncertainty of each channel per modality separately, whereas PPB allows the VSA to quantify the spatial uncertainty without losing fine-grained spatial information through patching. Figure 1 visualizes the CPB and PPB operations of the proposed method." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 170, + 275, + 182 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 170, + 275, + 182 + ], + "spans": [ + { + "bbox": [ + 55, + 170, + 275, + 182 + ], + "type": "text", + "content": "3.2.1 Channel-wise Projection & Bundling (CPB)" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 188, + 296, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 188, + 296, + 308 + ], + "spans": [ + { + "bbox": [ + 55, + 188, + 296, + 308 + ], + "type": "text", + "content": "Latent features are generally multi-channeled. By convention prototype formation based on Equation 1 projects and bundles all feature dimensions onto a common representation space " + }, + { + "bbox": [ + 55, + 188, + 296, + 308 + ], + "type": "inline_equation", + "content": "\\mathbf{H}" + }, + { + "bbox": [ + 55, + 188, + 296, + 308 + ], + "type": "text", + "content": " [52, 55]. Given latent feature " + }, + { + "bbox": [ + 55, + 188, + 296, + 308 + ], + "type": "inline_equation", + "content": "z_{m}^{i} \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 55, + 188, + 296, + 308 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 188, + 296, + 308 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 188, + 296, + 308 + ], + "type": "text", + "content": " denoting feature channels and " + }, + { + "bbox": [ + 55, + 188, + 296, + 308 + ], + "type": "inline_equation", + "content": "H, W" + }, + { + "bbox": [ + 55, + 188, + 296, + 308 + ], + "type": "text", + "content": " the spatial dimensions, adaptive pooling (max or average) is first used to reduce the spatial dimensions to " + }, + { + "bbox": [ + 55, + 188, + 296, + 308 + ], + "type": "inline_equation", + "content": "z_{m}^{pooled} \\in \\mathbb{R}^{C}" + }, + { + "bbox": [ + 55, + 188, + 296, + 308 + ], + "type": "text", + "content": ". Through matrix multiplication and proper dimension initialization of the projection matrix " + }, + { + "bbox": [ + 55, + 188, + 296, + 308 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 55, + 188, + 296, + 308 + ], + "type": "text", + "content": ", the correct hypervector dimensions can be obtained for " + }, + { + "bbox": [ + 55, + 188, + 296, + 308 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{m}^{z}" + }, + { + "bbox": [ + 55, + 188, + 296, + 308 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 91, + 312, + 296, + 328 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 312, + 296, + 328 + ], + "spans": [ + { + "bbox": [ + 91, + 312, + 296, + 328 + ], + "type": "interline_equation", + "content": "\\phi_ {m} \\left(z _ {m} ^ {i}\\right) = \\Phi^ {(d \\times C)} \\cdot z _ {m} ^ {\\text {p o o l e d} (C)} = \\mathcal {H} _ {m} ^ {z} ^ {(d)} \\tag {3}", + "image_path": "25971f43f9f560b8d7c7a7353f5b25949bbe02557cc8586b95c46838ba32e897.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 333, + 296, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 333, + 296, + 489 + ], + "spans": [ + { + "bbox": [ + 55, + 333, + 296, + 489 + ], + "type": "text", + "content": "Another method to handle multi-channels inputs is by projecting each feature dimension separately followed by binding and subsequent bundling to a common space [48]. However, we note that the above operations neglect the fact that different feature channels learn different aspects of the input and may contribute more or less to the overall uncertainty. For example, [27] showed that sharing the same uncertain distribution among different channels is less effective for out-of-distribution generalization, while considering each channel uncertainty separately brings better performances due to the different channel potentials. Therefore, instead of absorbing the channel dimensions during projection and bundling, we retain them as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 82, + 493, + 296, + 509 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 493, + 296, + 509 + ], + "spans": [ + { + "bbox": [ + 82, + 493, + 296, + 509 + ], + "type": "interline_equation", + "content": "\\phi_ {m} \\left(z _ {m} ^ {i}\\right) = \\Phi^ {(d \\times C)} \\otimes z _ {m} ^ {\\text {p o o l e d} (C)} = \\mathcal {H} _ {m} ^ {z} ^ {(d \\times C)} \\tag {4}", + "image_path": "41bc98a5773a4b84fd32ca12ad38b7ce06b47f20c417599a8e308a5e5180389b.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 513, + 296, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 513, + 296, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 513, + 296, + 559 + ], + "type": "text", + "content": "where we use the Einstein summation notation " + }, + { + "bbox": [ + 55, + 513, + 296, + 559 + ], + "type": "inline_equation", + "content": "\\otimes" + }, + { + "bbox": [ + 55, + 513, + 296, + 559 + ], + "type": "text", + "content": " (not the binding " + }, + { + "bbox": [ + 55, + 513, + 296, + 559 + ], + "type": "inline_equation", + "content": "\\otimes" + }, + { + "bbox": [ + 55, + 513, + 296, + 559 + ], + "type": "text", + "content": " notation) to retain the channel dimension. Consequently, the bundling operation is modified to account for the channel dimensions:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 128, + 563, + 296, + 598 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 563, + 296, + 598 + ], + "spans": [ + { + "bbox": [ + 128, + 563, + 296, + 598 + ], + "type": "interline_equation", + "content": "\\mathcal {H} _ {m} ^ {l} = \\bigoplus_ {i: y ^ {i} = l} ^ {d i m = C} \\phi_ {m} \\left(z _ {m} ^ {i}\\right) \\tag {5}", + "image_path": "9eb764ae0ddc66d5220b8abff27e7fe802e1489783c939bef7e08ddae3a768b5.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 602, + 298, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 602, + 298, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 602, + 298, + 651 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 602, + 298, + 651 + ], + "type": "inline_equation", + "content": "dim = C" + }, + { + "bbox": [ + 55, + 602, + 298, + 651 + ], + "type": "text", + "content": " indicates the dimension over which the element-wise addition is performed and " + }, + { + "bbox": [ + 55, + 602, + 298, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_m^l" + }, + { + "bbox": [ + 55, + 602, + 298, + 651 + ], + "type": "text", + "content": " is now " + }, + { + "bbox": [ + 55, + 602, + 298, + 651 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{d\\times C}" + }, + { + "bbox": [ + 55, + 602, + 298, + 651 + ], + "type": "text", + "content": ". Additionally, the similarity uncertainty now outputs similarities for each channel dimension by representing Equation 2:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 664, + 296, + 710 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 664, + 296, + 710 + ], + "spans": [ + { + "bbox": [ + 62, + 664, + 296, + 710 + ], + "type": "interline_equation", + "content": "\\mathbb {U} _ {m} ^ {C} = \\bigcup_ {c = 1} ^ {C} \\left\\{\\bigcup_ {l = 1} ^ {L} \\left\\{\\delta \\left(\\mathcal {H} _ {m} ^ {z} [ c ], \\mathcal {H} _ {m} ^ {l} [ c ]\\right) \\right\\} \\right\\}, \\text {s . t .} | \\mathbb {U} _ {m} ^ {C} | = C \\times L \\tag {6}", + "image_path": "8fbe841e965d508742fb9a37e379a5618ea133c08275ef281a64d10f95347aba.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 555, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 120 + ], + "type": "text", + "content": "We empirically demonstrate through experiments that channel projection and bundling outperforms conventional projection and bundling for the task of multimodal uncertainty-aware feature fusion." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 134, + 522, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 134, + 522, + 146 + ], + "spans": [ + { + "bbox": [ + 313, + 134, + 522, + 146 + ], + "type": "text", + "content": "3.2.2 Patch-wise Projection & Bundling (PPB)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 152, + 555, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 152, + 555, + 273 + ], + "spans": [ + { + "bbox": [ + 313, + 152, + 555, + 273 + ], + "type": "text", + "content": "Channel projection and bundling enables our method to capture the holistic uncertainties of each feature channel. However, to capture the finer-granularity uncertainties of the spatial dimensions requires a different approach. By convention spatial dimensions are either pooled to a single value and projected to high dimensions as shown by Equation 3 or each spatial dimension is separately projected to high dimensions. The former loses spatial information due to pooling and the latter is computationally infeasible and memory intensive given the large spatial dimensions of deep neural networks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 274, + 556, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 274, + 556, + 395 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 556, + 395 + ], + "type": "text", + "content": "Inspired by the idea of using image patches to capture spatial information in anomaly detection and image classification [6, 44], we propose PPB which projects and bundles spatial patches of latent features to hyperdimensional space to capture the spatial uncertainties. Specifically, given latent feature " + }, + { + "bbox": [ + 313, + 274, + 556, + 395 + ], + "type": "inline_equation", + "content": "z_{m}^{i}\\in \\mathbb{R}^{C\\times H\\times W}" + }, + { + "bbox": [ + 313, + 274, + 556, + 395 + ], + "type": "text", + "content": ", we slice " + }, + { + "bbox": [ + 313, + 274, + 556, + 395 + ], + "type": "inline_equation", + "content": "(H,W)" + }, + { + "bbox": [ + 313, + 274, + 556, + 395 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 313, + 274, + 556, + 395 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 274, + 556, + 395 + ], + "type": "text", + "content": " patches of resolution " + }, + { + "bbox": [ + 313, + 274, + 556, + 395 + ], + "type": "inline_equation", + "content": "(\\tilde{h},\\tilde{w})" + }, + { + "bbox": [ + 313, + 274, + 556, + 395 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 274, + 556, + 395 + ], + "type": "inline_equation", + "content": "\\tilde{h} = H / \\sqrt{(P)}" + }, + { + "bbox": [ + 313, + 274, + 556, + 395 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 274, + 556, + 395 + ], + "type": "inline_equation", + "content": "\\tilde{w} = W / \\sqrt{(P)}" + }, + { + "bbox": [ + 313, + 274, + 556, + 395 + ], + "type": "text", + "content": ". This gives us the set of patched features " + }, + { + "bbox": [ + 313, + 274, + 556, + 395 + ], + "type": "inline_equation", + "content": "\\tilde{z}_{m}^{i} = \\{\\tilde{z}_{m}^{i,j}\\}_{j = 1}^{P}" + }, + { + "bbox": [ + 313, + 274, + 556, + 395 + ], + "type": "text", + "content": " where each " + }, + { + "bbox": [ + 313, + 274, + 556, + 395 + ], + "type": "inline_equation", + "content": "\\tilde{z}_{m}^{i,j}\\in \\mathbb{R}^{C\\times \\tilde{h}\\times \\tilde{w}}" + }, + { + "bbox": [ + 313, + 274, + 556, + 395 + ], + "type": "text", + "content": " are pooled to " + }, + { + "bbox": [ + 313, + 274, + 556, + 395 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^C" + }, + { + "bbox": [ + 313, + 274, + 556, + 395 + ], + "type": "text", + "content": " and form the set of pooled patched features " + }, + { + "bbox": [ + 313, + 274, + 556, + 395 + ], + "type": "inline_equation", + "content": "\\tilde{z}_{m}^{pooled}\\in \\mathbb{R}^{C\\times P}" + }, + { + "bbox": [ + 313, + 274, + 556, + 395 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 396, + 555, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 555, + 420 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 555, + 420 + ], + "type": "text", + "content": "Given patched features " + }, + { + "bbox": [ + 313, + 396, + 555, + 420 + ], + "type": "inline_equation", + "content": "\\tilde{z}_m^{pooled}" + }, + { + "bbox": [ + 313, + 396, + 555, + 420 + ], + "type": "text", + "content": ", we project them according to Equation 3 to form the set of patched projections:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 337, + 436, + 555, + 453 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 436, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 337, + 436, + 555, + 453 + ], + "type": "interline_equation", + "content": "\\phi_ {m} \\left(\\tilde {z} _ {m} ^ {i}\\right) = \\Phi^ {(d \\times C)} \\cdot \\tilde {z} _ {m} ^ {\\text {p o o l e d} (C \\times P)} = \\mathcal {H} _ {m} ^ {\\tilde {z}} ^ {(d \\times P)} \\tag {7}", + "image_path": "da9000b558959400b4b4fda73f8f3f9f82348a72d8bc85ce65875a96cc8c9e42.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 461, + 555, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 555, + 497 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 555, + 497 + ], + "type": "text", + "content": "Consequently, these projections are bundled for each patch according to Equation 1 to produce the set of patched prototypes as follows:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 387, + 506, + 555, + 540 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 387, + 506, + 555, + 540 + ], + "spans": [ + { + "bbox": [ + 387, + 506, + 555, + 540 + ], + "type": "interline_equation", + "content": "\\mathcal {H} _ {\\tilde {m}} ^ {l} = \\bigoplus_ {i: y ^ {i} = l} ^ {d i m = P} \\phi_ {m} \\left(\\tilde {z} _ {m} ^ {i}\\right) \\tag {8}", + "image_path": "44f5a72d9eb6df0c8048fd8a72948110ca34a1fc5afc0bc6556ad0b70fe1c783.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 545, + 555, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 582 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 545, + 555, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\tilde{m}}^{l}" + }, + { + "bbox": [ + 313, + 545, + 555, + 582 + ], + "type": "text", + "content": " represents the patched prototypes of modality " + }, + { + "bbox": [ + 313, + 545, + 555, + 582 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 545, + 555, + 582 + ], + "type": "text", + "content": " , for each patch and for each context label " + }, + { + "bbox": [ + 313, + 545, + 555, + 582 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 545, + 555, + 582 + ], + "type": "text", + "content": " , making " + }, + { + "bbox": [ + 313, + 545, + 555, + 582 + ], + "type": "inline_equation", + "content": "|\\mathcal{H}_{\\tilde{m}}^{l}| = d\\times P" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 582, + 555, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 607 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 607 + ], + "type": "text", + "content": "Finally, we can obtain the spatial uncertainties for each patch of the input projection " + }, + { + "bbox": [ + 313, + 582, + 555, + 607 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_m^{\\tilde{z}}" + }, + { + "bbox": [ + 313, + 582, + 555, + 607 + ], + "type": "text", + "content": " by modifying Equation 2:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 623, + 555, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 623, + 555, + 669 + ], + "spans": [ + { + "bbox": [ + 320, + 623, + 555, + 669 + ], + "type": "interline_equation", + "content": "\\mathbb {U} _ {\\tilde {m}} ^ {P} = \\bigcup_ {p = 1} ^ {P} \\left\\{\\bigcup_ {l = 1} ^ {L} \\left\\{\\delta \\left(\\mathcal {H} _ {\\tilde {m}} ^ {\\tilde {z}} [ p ], \\mathcal {H} _ {\\tilde {m}} ^ {l} [ p ]\\right) \\right\\} \\right\\}, \\text {s . t .} | \\mathbb {U} _ {\\tilde {m}} ^ {P} | = P \\times L \\tag {9}", + "image_path": "61005644e8965de7423e5ae207303551cfe24d91a5b6e67e9dfc6915beded49d.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "content": "Our final proposed approach combines both channel and patch methods for projection and bundling to capture channel-wise feature uncertainties, as well as finer" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22309" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": "granularity spatial feature uncertainties, to provide a holistic epistemic uncertainty quantification of each modality." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 107, + 136, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 107, + 136, + 121 + ], + "spans": [ + { + "bbox": [ + 55, + 107, + 136, + 121 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 127, + 171, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 127, + 171, + 140 + ], + "spans": [ + { + "bbox": [ + 55, + 127, + 171, + 140 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 146, + 124, + 157 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 146, + 124, + 157 + ], + "spans": [ + { + "bbox": [ + 55, + 146, + 124, + 157 + ], + "type": "text", + "content": "4.1.1 Datasets" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 165, + 297, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 165, + 297, + 369 + ], + "spans": [ + { + "bbox": [ + 55, + 165, + 297, + 369 + ], + "type": "text", + "content": "We evaluate the effectiveness of HyperDUM for uncertainty-aware feature fusion using two multimodal autonomous driving datasets. The aiMotive dataset [33] includes camera, lidar, and radar sensors for long-distance " + }, + { + "bbox": [ + 55, + 165, + 297, + 369 + ], + "type": "inline_equation", + "content": "(\\geq 75m)" + }, + { + "bbox": [ + 55, + 165, + 297, + 369 + ], + "type": "text", + "content": " 3D object detection and features diverse scenes across various locations, times, and weather conditions. The DeLiVER dataset [58] contains camera, lidar, event, and depth sensors for semantic segmentation, including adverse corner cases such as motion blur and lidar jitter. Additionally, we integrate corner cases into the aiMotive dataset by applying various lighting effects to camera sensors, causing reduced clarity and scene obstruction. These effects are created through image transformations such as GaussianBlur and exposure adjustments [39, 50]. We also implement foggification [3] to simulate synthetic fogging of lidar data, resulting in noisy point clouds with reduced field of view. These corner case effects can be visualized in Figure 2. Details in the appendix." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 384, + 119, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 119, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 119, + 396 + ], + "type": "text", + "content": "4.1.2 Models" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 403, + 296, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 296, + 583 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 296, + 583 + ], + "type": "text", + "content": "We utilize the pretrained BEVFusion [33] and CMNeXT [58] models as the fusion architectures for aiMotive and DeLiVER, respectively. BEVFusion is built upon VoxelNet [59] and BEVDepth [29] to fuse camera and lidar representations on a unified BEV space. CMNeXT is a four-stage pyramidal fusion model that treats the RGB representation as the primary branch and other modalities as the secondary branch. We integrate the uncertainty quantification module (UQM) at each model's respective pre-fusion points. For aiMotive this is chosen to be at the feature concatenation step, right before the bev Fuse step. For DeLiVER, we insert the UQM immediately after the feature rectification module and right before the feature fusion module. Given that the CMNeXT model has four stages, we instantiate one UQM per stage. Details in the Appendix." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 598, + 163, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 598, + 163, + 609 + ], + "spans": [ + { + "bbox": [ + 55, + 598, + 163, + 609 + ], + "type": "text", + "content": "4.1.3 Baseline Methods" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": "We implemented different UQ methods and compared their performances against our method. For each implementation, we followed their open-sourced code implementation and hyperparameter settings. Specifically, for Infer-dropout (InfMCD) and infer-noise (InfNoise), we followed their suggestions and chose 10 forward passes and used the variance of the outputs for uncertainty estimation. In addition to variance, we computed predictive entropy, mutual information," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 556, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 556, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 556, + 348 + ], + "type": "text", + "content": "and entropy as additional uncertainty estimates following [49]. It was shown that using more uncertainty metrics can provide more effective quantification for different types of uncertainties. For Posterior Network (PostNet), we constructed PostNets for each modality so that they can estimate the modality-specific uncertainty independently and be used in the uncertainty-aware feature fusion process. The use of multiple PostNets is similar to [20] in their construction of separate PostNets for agent's past behavior, road structure map, and social context respectively. For latent deterministic uncertainty (LDU), we set the number of learnable prototypes to the number of scenarios. Specifically, for aiMootive, the scenarios are Highway, Urban, Night, and Rain, while for DeLiVER, the scenarios are Cloudy, Foggy, Night, Rainy, and Sunny. Finally, for all methods, all layers prior to the UQM are kept frozen and we fine-tune the uncertainty weighting layer, which takes the respective uncertainty metrics as input, along with other post-fusion layers, to enable the model to learn the uncertainty-weighted features. Details of the fine-tuning procedures and visualizations of the architectures and UQM insertion points are provided in the Appendix. In all of our experiments the best one is highlighted and bolded and the 2nd best is lightly-highlighted." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 354, + 555, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 555, + 380 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 555, + 380 + ], + "type": "text", + "content": "4.2. Multimodal Uncertainty Fusion for Autonomous Vehicles Perception" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 384, + 434, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 434, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 434, + 396 + ], + "type": "text", + "content": "4.2.1 3D Object Detection" + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 315, + 412, + 555, + 472 + ], + "blocks": [ + { + "bbox": [ + 315, + 412, + 555, + 472 + ], + "lines": [ + { + "bbox": [ + 315, + 412, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 315, + 412, + 555, + 472 + ], + "type": "table", + "html": "
UQ MethodHighwayUrbanNightRainMean
BEVFusion [33]72.55/69.0863.72/63.5274.74/73.1342.75/42.8365.67/64.79
InfMCD [34]72.58/69.6562.02/59.5975.05/73.7235.58/37.4164.56/64.58
InfNoise [34]70.11/68.5263.10/60.4275.08/74.1935.58/37.4164.59/64.56
PostNet [5]69.19/68.0261.38/59.8473.32/68.9734.95/37.6063.13/60.21
LDU [10]71.48/69.3362.54/60.1176.49/75.0440.49/41.2464.69/64.73
HyperDUM72.23/69.5864.77/64.4876.69/75.1544.78/45.4866.70/66.00
", + "image_path": "f00aa41568f998fa00e84989d6ff800366e91f98b30df67cf985d30a1e4fe8a9.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 506, + 556, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 556, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 556, + 578 + ], + "type": "text", + "content": "Table 1 compares the all-point/11-point interpolation Average Precision (AP) metric for object detection under diverse scenes. It can be seen that apart from the Highway scenario, HyperDUM outperforms all methods across all scenarios. Specifically, HyperDUM improves upon the all-point AP by " + }, + { + "bbox": [ + 313, + 506, + 556, + 578 + ], + "type": "inline_equation", + "content": "2.01\\%" + }, + { + "bbox": [ + 313, + 506, + 556, + 578 + ], + "type": "text", + "content": " on average against the state-of-the-art (LDU)." + } + ] + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 315, + 586, + 554, + 646 + ], + "blocks": [ + { + "bbox": [ + 341, + 481, + 526, + 491 + ], + "lines": [ + { + "bbox": [ + 341, + 481, + 526, + 491 + ], + "spans": [ + { + "bbox": [ + 341, + 481, + 526, + 491 + ], + "type": "text", + "content": "Table 1. Results on aiMotive under diverse scenes." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 586, + 554, + 646 + ], + "lines": [ + { + "bbox": [ + 315, + 586, + 554, + 646 + ], + "spans": [ + { + "bbox": [ + 315, + 586, + 554, + 646 + ], + "type": "table", + "html": "
UQ MethodMBOEUELFMean
BEVFusion [33]64.10/63.6962.40/59.2663.57/63.2565.07/65.0863.79/62.75
InfMCD [34]62.00/61.9964.47/64.2863.25/63.4964.27/63.8163.50/63.39
InfNoise [34]62.61/62.6164.94/64.7263.18/60.0765.19/64.9963.98/63.10
PostNet [5]60.42/58.9063.12/59.9960.20/59.2564.13/60.6061.97/59.69
LDU [10]62.06/62.2265.07/64.8562.87/60.0265.47/65.1063.87/63.05
HyperDUM64.39/63.9966.16/65.3964.18/63.9165.89/65.2265.16/64.62
", + "image_path": "93f0be36abd0e4e6fc437cd49876cf34598c99b6bdec3ec41a27be94e08ecef8.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 656, + 555, + 689 + ], + "lines": [ + { + "bbox": [ + 313, + 656, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 656, + 555, + 689 + ], + "type": "text", + "content": "Table 2. Results on aiMotive under corner cases: Motion Blur (MB), Over-Exposure (OE), Under-Exposure (UE), LiDAR-Fog (LF). (Metrics: all-point AP/11-point interpolation AP)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 326, + 701, + 556, + 713 + ], + "lines": [ + { + "bbox": [ + 326, + 701, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 556, + 713 + ], + "type": "text", + "content": "Table 2 demonstrates the robustness of UQ methods un" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "22310" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 555, + 176 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 555, + 176 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 555, + 176 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 555, + 176 + ], + "type": "image", + "image_path": "18f3f884b122647cd429b9c7351cd341d8ce7c342c38bdeb4106a85bfbf62696.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 168, + 178, + 198, + 186 + ], + "lines": [ + { + "bbox": [ + 168, + 178, + 198, + 186 + ], + "spans": [ + { + "bbox": [ + 168, + 178, + 198, + 186 + ], + "type": "text", + "content": "Original" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 60, + 187, + 181, + 266 + ], + "blocks": [ + { + "bbox": [ + 60, + 187, + 181, + 266 + ], + "lines": [ + { + "bbox": [ + 60, + 187, + 181, + 266 + ], + "spans": [ + { + "bbox": [ + 60, + 187, + 181, + 266 + ], + "type": "image", + "image_path": "1ebd5c7b4d839028ec73d946db4182304852d06e5c77471a041ac0bcc3a6af40.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 105, + 267, + 134, + 277 + ], + "lines": [ + { + "bbox": [ + 105, + 267, + 134, + 277 + ], + "spans": [ + { + "bbox": [ + 105, + 267, + 134, + 277 + ], + "type": "text", + "content": "Original" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 184, + 188, + 304, + 266 + ], + "blocks": [ + { + "bbox": [ + 184, + 188, + 304, + 266 + ], + "lines": [ + { + "bbox": [ + 184, + 188, + 304, + 266 + ], + "spans": [ + { + "bbox": [ + 184, + 188, + 304, + 266 + ], + "type": "image", + "image_path": "d04c34863021ca7bd143cc4105ff25333fea4ed304d0673ec6196a77d6aa562d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 223, + 268, + 266, + 277 + ], + "lines": [ + { + "bbox": [ + 223, + 268, + 266, + 277 + ], + "spans": [ + { + "bbox": [ + 223, + 268, + 266, + 277 + ], + "type": "text", + "content": "Motion Blur" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 291, + 555, + 314 + ], + "lines": [ + { + "bbox": [ + 55, + 291, + 555, + 314 + ], + "spans": [ + { + "bbox": [ + 55, + 291, + 555, + 314 + ], + "type": "text", + "content": "Figure 2. Corner cases for the aiMotive 3D Object Detection dataset. (Top-Left) Normal Lidar point-cloud birds-eye-view. (Top-Right) Lidar Foggification (LF). (Bottom) In order from left to right (Normal, Motion Blur (MB), Over-Exposure (OE), Under-Exposure (UE))" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 307, + 188, + 429, + 266 + ], + "blocks": [ + { + "bbox": [ + 405, + 178, + 452, + 187 + ], + "lines": [ + { + "bbox": [ + 405, + 178, + 452, + 187 + ], + "spans": [ + { + "bbox": [ + 405, + 178, + 452, + 187 + ], + "type": "text", + "content": "Foggification" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 188, + 429, + 266 + ], + "lines": [ + { + "bbox": [ + 307, + 188, + 429, + 266 + ], + "spans": [ + { + "bbox": [ + 307, + 188, + 429, + 266 + ], + "type": "image", + "image_path": "5db4b7439059f0370f743806c814ea632f5fd496cd2d77e7a5f18bb93b3bbc25.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 268, + 395, + 277 + ], + "lines": [ + { + "bbox": [ + 343, + 268, + 395, + 277 + ], + "spans": [ + { + "bbox": [ + 343, + 268, + 395, + 277 + ], + "type": "text", + "content": "Over-Exposure" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 431, + 188, + 552, + 266 + ], + "blocks": [ + { + "bbox": [ + 431, + 188, + 552, + 266 + ], + "lines": [ + { + "bbox": [ + 431, + 188, + 552, + 266 + ], + "spans": [ + { + "bbox": [ + 431, + 188, + 552, + 266 + ], + "type": "image", + "image_path": "575cd4a22a55f882b1974eac25a54afa14eecbefb7fe86ba2282d1e7e2f1d0bf.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 465, + 268, + 522, + 277 + ], + "lines": [ + { + "bbox": [ + 465, + 268, + 522, + 277 + ], + "spans": [ + { + "bbox": [ + 465, + 268, + 522, + 277 + ], + "type": "text", + "content": "Under-Exposure" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 335, + 296, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 335, + 296, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 296, + 430 + ], + "type": "text", + "content": "der our injected corner cases. These injections were unseen during training which complicates testing due to distribution shifts. Each corner case is applied to and evaluated on the entire test set across all scenarios. Whereas other methods fail and suffer performance degradations, HyperDUM only degrades " + }, + { + "bbox": [ + 55, + 335, + 296, + 430 + ], + "type": "inline_equation", + "content": "0.51\\% / 0.17\\%" + }, + { + "bbox": [ + 55, + 335, + 296, + 430 + ], + "type": "text", + "content": " compared to the baseline w/o injections demonstrating that HyperDUM is more robust even when facing distribution shifts." + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 57, + 464, + 296, + 525 + ], + "blocks": [ + { + "bbox": [ + 57, + 464, + 296, + 525 + ], + "lines": [ + { + "bbox": [ + 57, + 464, + 296, + 525 + ], + "spans": [ + { + "bbox": [ + 57, + 464, + 296, + 525 + ], + "type": "table", + "html": "
UQ MethodHighwayUrbanNightRainMean
BEVFusion [33]45.59/45.3446.38/45.5344.49/45.8742.65/42.9345.84/46.87
InfMCD [34]45.57/46.2938.90/38.1142.07/41.9726.13/28.4940.74/41.64
InfNoise [34]40.73/40.3842.41/40.9038.80/40.0029.28/32.8140.91/42.04
PostNet [5]41.13/40.6439.47/40.1640.37/41.9228.56/28.9639.37/41.66
LDU [10]43.07/45.0140.17/41.7142.90/42.5231.12/34.0840.65/42.17
HyperDUM44.86/46.0647.89/48.9144.42/46.1240.54/41.6346.52/48.27
", + "image_path": "1739e5a183815bfcfea0c48a1d7e1713a51852a76414928bbbbfab2f84fd1344.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": "Table 3 is an experiment on the distant region setting, keeping only predictions and ground truths over " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\geq 75m" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": ". As expected, all UQ methods experience major performance drops, especially for Rain. This is likely due to the combined uncertainties of distance with rain obstruction causing substantial minimization of the uncertainty-weighted features (the weighting output is between 0-1 times the original feature value). The original BEVFusion model maintains better performance because it is still able to \"guess\" the presence of objects as the features are not minimized by uncertainty." + } + ] + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 315, + 332, + 555, + 464 + ], + "blocks": [ + { + "bbox": [ + 55, + 533, + 295, + 556 + ], + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 556 + ], + "type": "text", + "content": "Table 3. Results on aiMotive under diverse scenes in the distant region (" + }, + { + "bbox": [ + 55, + 533, + 295, + 556 + ], + "type": "inline_equation", + "content": ">75\\mathrm{m}" + }, + { + "bbox": [ + 55, + 533, + 295, + 556 + ], + "type": "text", + "content": "). (Metrics: all-point AP/11-point interpolation AP)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 332, + 555, + 464 + ], + "lines": [ + { + "bbox": [ + 315, + 332, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 315, + 332, + 555, + 464 + ], + "type": "table", + "html": "
ScenariosUQ Methods (Metrics: mIoU ↑)
CMNeXt [58]InfMCD [34]InfNoise [34]PostNet [5]LDU [10]Gemini [21]HyperDUM (Ours)
Cloudy68.7069.2369.2169.2868.94-69.76+1.06
Foggy65.6666.1866.1066.0465.72-66.85+1.19
Night62.4663.4463.1462.9763.06-64.21+1.75
Rainy67.5068.0868.0968.1667.82-68.71+1.21
Sunny66.5767.0167.1666.9366.75-67.87+1.30
MB62.9163.6163.5563.5563.16-64.28+1.37
OE64.5965.3965.1765.0664.73-65.67+1.08
UE60.0060.3860.4260.2760.29-61.20+1.20
LJ65.9266.1266.2566.3366.40-66.93+1.01
EL65.4866.0566.1766.0665.89-66.80+1.32
Mean66.3066.9066.8666.7766.6066.9067.59+1.29
", + "image_path": "59caed25341756cdbe980c2b885cd145c740acbbf95f0c36bdaf5cf9c10ffdce.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 474, + 555, + 507 + ], + "lines": [ + { + "bbox": [ + 313, + 474, + 555, + 507 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 555, + 507 + ], + "type": "text", + "content": "Table 4. DeLiVER mean Intersection over Union (mIoU) performance under adverse weather and corner cases. Lidar-Jitter (LJ), Event Low-resolution (EL)." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 533, + 448, + 546 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 448, + 546 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 448, + 546 + ], + "type": "text", + "content": "4.2.2 Semantic Segmentation" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 557, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 557, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 557, + 556, + 713 + ], + "type": "text", + "content": "Table 4 shows the DeLiVER validation set performance on different scenarios and corner cases for the semantic segmentation task. Overall, HyperDUM achieves the best performance over all methods across every weather and corner case scenario. On average HyperDUM improves over the baseline by 1.29. In particular, for the three most difficult scenarios including Under-Exposure, Night and Motion Blur, HyperDUM achieves at least 1.00 improvement over the baseline. For GeminiFusion [21], we only report the mean as they omitted the detailed performance breakdown in their paper. These results demonstrate that our method can better quantify the multimodal feature uncertainties and ultimately improve the fusion features for the semantic segmentation" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22311" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": "task. Additionally, we note that all UQ methods improved over the baseline method. This shows that incorporating UQ with fusion into multimodal pipelines not only enables UQ but also can improve baseline performance." + } + ] + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 129, + 298, + 262 + ], + "blocks": [ + { + "bbox": [ + 56, + 129, + 298, + 262 + ], + "lines": [ + { + "bbox": [ + 56, + 129, + 298, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 298, + 262 + ], + "type": "table", + "html": "
ScenariosUQ Methods (Metrics: ECE ↓)
CMNeXt [58]InfMCD [34]InfNoise [34]PostNet [5]LDU [10]HyperDUM (Ours)
Cloudy1.18E-021.05E-021.07E-020.99E-021.09E-020.94E-02-0.24
Foggy1.67E-021.48E-021.52E-021.45E-021.56E-021.35E-02-0.30
Night1.88E-021.63E-021.68E-021.71E-021.68E-021.39E-02-0.49
Rainy1.53E-021.37E-021.39E-021.33E-021.43E-021.21E-02-0.32
Sunny1.42E-021.25E-021.27E-021.24E-021.33E-021.12E-02-0.30
MB1.54E-021.32E-021.40E-021.32E-021.44E-021.19E-02-0.35
OE1.49E-021.55E-021.39E-021.42E-021.44E-021.31E-02-0.18
UE1.75E-022.07E-021.69E-021.84E-021.63E-021.53E-02-0.22
LJ1.54E-021.38E-021.39E-021.35E-021.40E-021.28E-02-0.26
EL1.45E-021.17E-021.27E-021.26E-021.27E-021.09E-02-0.36
Mean1.53E-021.35E-021.38E-021.34E-021.41E-021.20E-02-0.33
", + "image_path": "f67d38a811f1607ed83ec95873f0566ab5d5ef66ca8a4bc5e4b7680ecc0f5c9c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 307, + 297, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 307, + 297, + 475 + ], + "spans": [ + { + "bbox": [ + 54, + 307, + 297, + 475 + ], + "type": "text", + "content": "Table 5 shows the Expected Calibration Error (ECE) metric for all methods across the same scenarios. ECE measures how well a model is calibrated by comparing it's predicted probabilities with the actual probabilities of the ground truth distribution. ECE helps us understand how reliable a model's confidence scores are. A higher ECE means the model is more overconfident (or unsure) about its predictions. As shown, HyperDUM has the lowest ECE score over all methods across every weather scenario and corner case with a 0.33 average decrease compared to the baseline. These results demonstrate that our method helps better calibrate the overconfident semantic segmentation model. In fact all UQ methods on average has a lower ECE compared to the baseline demonstrating its usefulness in model calibration." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 483, + 277, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 277, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 277, + 495 + ], + "type": "text", + "content": "4.3. Pre vs. Post Uncertainty Feature Learning" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 58, + 510, + 292, + 635 + ], + "blocks": [ + { + "bbox": [ + 55, + 271, + 296, + 293 + ], + "lines": [ + { + "bbox": [ + 55, + 271, + 296, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 271, + 296, + 293 + ], + "type": "text", + "content": "Table 5. DeLiVER Expected Calibration Error (ECE) performance under adverse weather and corner cases." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 510, + 292, + 635 + ], + "lines": [ + { + "bbox": [ + 58, + 510, + 292, + 635 + ], + "spans": [ + { + "bbox": [ + 58, + 510, + 292, + 635 + ], + "type": "table", + "html": "
ScenariosUQ Methods (Metrics: mIoU ↑)
CMNeXt [58]InfMCD [34]InfNoise [34]PostNet [5]LDU [10]HyperDUM (Ours)
Cloudy68.768.7868.569.1369.0669.26
Foggy65.6665.7865.5965.8965.9266.40
Night62.4662.762.4462.8263.0863.71
Rainy67.567.6667.3768.0267.8468.29
Sunny66.5766.5266.3966.7866.7867.43
MB62.9163.1162.7963.4063.363.78
OE64.5964.9464.664.9164.9565.20
UE60.0059.7560.0760.1260.360.72
LJ65.9266.1466.0166.1766.2166.44
EL65.4865.765.6665.9166.0366.34
Mean66.3066.5966.4966.6066.6466.98
", + "image_path": "38adbab919183786e9d369fedb6d5a53eed71b4acf0ac6f1797335784e0b023d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "Table 6 evaluates the performance of uncertainty feature weighting after the fusion block of Figure 1 or see the archi" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 556, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 556, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 556, + 156 + ], + "type": "text", + "content": "tecture figure in the Appendix. As expected, we found that the overall performance gain is limited/ lower compared to its prefusion uncertainty fusion counterpart. These results align with our motivation that different modalities experience different modality-specific uncertainties. And these uncertainties can propagate to the fusion module, undermining the fusion features and degrading downstream task performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 163, + 411, + 175 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 163, + 411, + 175 + ], + "spans": [ + { + "bbox": [ + 314, + 163, + 411, + 175 + ], + "type": "text", + "content": "4.4. Model Ablation" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 328, + 188, + 541, + 236 + ], + "blocks": [ + { + "bbox": [ + 55, + 643, + 296, + 676 + ], + "lines": [ + { + "bbox": [ + 55, + 643, + 296, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 643, + 296, + 676 + ], + "type": "text", + "content": "Table 6. DeLiVER mean Intersection over Union (mIoU) performance under adverse weather and corner cases using post-fusion features." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 328, + 188, + 541, + 236 + ], + "lines": [ + { + "bbox": [ + 328, + 188, + 541, + 236 + ], + "spans": [ + { + "bbox": [ + 328, + 188, + 541, + 236 + ], + "type": "table", + "html": "
Dataset\nMetricaiMotive\nall-point AP/11-point AP ↑DeLiVER\nmIoU ↑
HyperDUM\n- w/o Patch (4x) Proj.-66.70/66.00 (Mean Δ)67.59 (Mean Δ)
-65.67/65.23 (-1.03/-0.77)-67.18 (-0.41)
- w/o Chan Proj.64.43/64.25 (-1.24/-0.98)66.63 (-0.55)
", + "image_path": "b91997c9cbebb6e9604276b6b81a32746629980a52c6b1c8be8478ac2a8d317d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 337, + 243, + 529, + 254 + ], + "lines": [ + { + "bbox": [ + 337, + 243, + 529, + 254 + ], + "spans": [ + { + "bbox": [ + 337, + 243, + 529, + 254 + ], + "type": "text", + "content": "Table 7. Model ablation for aiMotive and DeLiVER." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 269, + 556, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 269, + 556, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 269, + 556, + 437 + ], + "type": "text", + "content": "For our model ablation, we analyzed the effects of sequentially removing our proposed components and the impact on the overall performance. As seen from Table 7 the removal of the Patch Projection (with 4 patches) results in a " + }, + { + "bbox": [ + 313, + 269, + 556, + 437 + ], + "type": "inline_equation", + "content": "1.03 / 0.77" + }, + { + "bbox": [ + 313, + 269, + 556, + 437 + ], + "type": "text", + "content": " decrease in performance and the removal of Channel Projection in a " + }, + { + "bbox": [ + 313, + 269, + 556, + 437 + ], + "type": "inline_equation", + "content": "1.24 / 0.98" + }, + { + "bbox": [ + 313, + 269, + 556, + 437 + ], + "type": "text", + "content": " decrease in performance for aiMotive. Similarly for the DeLiVER dataset, the removal of the Patch Projection (with 4 patches) results in a 0.41 decrease in performance and removing Channel Projection results in a 0.55 decrease in performance. Together we see that both methods contribute significantly to the overall performance demonstrating that both channel and spatial uncertainties should be accounted for to better capture the holistic uncertainty of the features before fusion. More ablations in the Appendix." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 450, + 430, + 463 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 430, + 463 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 430, + 463 + ], + "type": "text", + "content": "4.4.1 Sensitivity Analysis" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 322, + 483, + 548, + 618 + ], + "blocks": [ + { + "bbox": [ + 322, + 483, + 548, + 618 + ], + "lines": [ + { + "bbox": [ + 322, + 483, + 548, + 618 + ], + "spans": [ + { + "bbox": [ + 322, + 483, + 548, + 618 + ], + "type": "image", + "image_path": "922405aa88be352a81f0f4285c6fbc1fe69fb6c35dc991a9ac0256c67679f586.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 630, + 556, + 653 + ], + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 653 + ], + "type": "text", + "content": "Figure 3. Analysis on the effects of hyperdimension selection for channel projection on performance and computation (DeLiVER)." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "Here we investigate the tunable parameters for channel and patch projection & bundling (CPB & PPB). For CPB, the only parameter is the channel hyperdimension. For PPB, we have the patch hyperdimension and the number of patches." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22312" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 191, + 290, + 305 + ], + "blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "content": "Figure 3 shows a sweep over a range of dimensions from 625 to " + }, + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "inline_equation", + "content": "20\\mathrm{k}" + }, + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "content": " for CPB. This range of values is chosen based on the HDC expressivity limits (around " + }, + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "inline_equation", + "content": "10\\mathrm{k}" + }, + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "content": ") [48, 57]. We see that performance improves with dimension increase, but plateaus and decreases past " + }, + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "inline_equation", + "content": "10\\mathrm{k}" + }, + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "content": " while computation continues to grow. This occurs when large dimensions induce too much sparsity, making similar samples become orthogonal and capturing irrelevant information likened to over-fitting in neural networks [48]." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 62, + 191, + 290, + 305 + ], + "lines": [ + { + "bbox": [ + 62, + 191, + 290, + 305 + ], + "spans": [ + { + "bbox": [ + 62, + 191, + 290, + 305 + ], + "type": "image", + "image_path": "198eb036dd032be08ad28d358f231b8de37a1fa24d0312ac4c023745ba753395.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 318, + 296, + 341 + ], + "lines": [ + { + "bbox": [ + 55, + 318, + 296, + 341 + ], + "spans": [ + { + "bbox": [ + 55, + 318, + 296, + 341 + ], + "type": "text", + "content": "Figure 4. Analysis on the effects of patches and patch hyperdimension selection on performance and computation (DeLiVER)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 353, + 298, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 353, + 298, + 496 + ], + "spans": [ + { + "bbox": [ + 54, + 353, + 298, + 496 + ], + "type": "text", + "content": "Figure 4 shows a sweep over 4, 16 and 64 patch configurations together with the dimension sweep from 625 to 5000. We observe an interesting relation where the number of patches negatively correlates with the per-patch hyperdimension in terms of better performance. This makes sense as we increase the number of patches, the information needed to be encoded becomes smaller due to the reduction in the per-patch spatial resolution as indicated in Section 3.2.2. Therefore, the necessary hyperdimensions needed to encode the information should be reduced, while larger dimensions begin to over-fit. Whereas the opposite is true for less patches, larger hyperdimensions are needed." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 502, + 165, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 165, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 165, + 514 + ], + "type": "text", + "content": "4.5. Computation Cost" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 71, + 529, + 279, + 600 + ], + "blocks": [ + { + "bbox": [ + 71, + 529, + 279, + 600 + ], + "lines": [ + { + "bbox": [ + 71, + 529, + 279, + 600 + ], + "spans": [ + { + "bbox": [ + 71, + 529, + 279, + 600 + ], + "type": "table", + "html": "
DatasetaiMotiveDeLiVER
UQ MethodFLOPs#ParamsFLOPs#Params
InfMCD [34]990.90M4.05K1.84G20.19K
InfNoise [34]990.90M4.05K1.84G20.19K
LDU [10]264.24M44.04M613.42M102.24M
HyperDUM111.84M1.15M338.40M38.85M
- w/o Patch (4x) Proj.84.00M1.13M256.00M38.78M
- w/o Chan Proj.6.96M3.38K20.60M16.83K
", + "image_path": "f4c3ea05d8b8b731b46d8975a202f86d041adf9ccfd6abe1cda10eda70198016.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 608, + 295, + 630 + ], + "lines": [ + { + "bbox": [ + 55, + 608, + 295, + 630 + ], + "spans": [ + { + "bbox": [ + 55, + 608, + 295, + 630 + ], + "type": "text", + "content": "Table 8. Comparing inference (FLOPs) and training costs (# of Trainable Parameters)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 641, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 297, + 715 + ], + "type": "text", + "content": "Finally, the main benefit of HyperDUM is its efficiency for training and inference. We measure the computation costs in terms of Floating point operations (FLOPs) and the number of trainable model parameters in Table 8. We approximate the FLOPs by computing the major contributing operations. Refer to the Appendix for our FLOPs approximations" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 555, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 252 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 252 + ], + "type": "text", + "content": "for each operation type. For InfMCD and InfNoise these are the entropy, mutual information, and predictive entropy computations. For LDU, it is the cosine similarity between prototypes and matrix multiplication for uncertainty, which was found to be negligible. For HyperDUM, it is the input projection and cosine similarity computations. We see that HyperDUM is up to " + }, + { + "bbox": [ + 313, + 72, + 555, + 252 + ], + "type": "inline_equation", + "content": "2.36 \\times" + }, + { + "bbox": [ + 313, + 72, + 555, + 252 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 555, + 252 + ], + "type": "inline_equation", + "content": "1.81 \\times" + }, + { + "bbox": [ + 313, + 72, + 555, + 252 + ], + "type": "text", + "content": " more FLOP-efficient compared to LDU for aiMotive and DeLiVER respectively. Compared to traditional Bayesian approximation methods, it can be up to " + }, + { + "bbox": [ + 313, + 72, + 555, + 252 + ], + "type": "inline_equation", + "content": "8.86 \\times" + }, + { + "bbox": [ + 313, + 72, + 555, + 252 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 555, + 252 + ], + "type": "inline_equation", + "content": "5.44 \\times" + }, + { + "bbox": [ + 313, + 72, + 555, + 252 + ], + "type": "text", + "content": " more FLOP-efficient. As for the training costs, HyperDUM has " + }, + { + "bbox": [ + 313, + 72, + 555, + 252 + ], + "type": "inline_equation", + "content": "38.30 \\times" + }, + { + "bbox": [ + 313, + 72, + 555, + 252 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 555, + 252 + ], + "type": "inline_equation", + "content": "2.63 \\times" + }, + { + "bbox": [ + 313, + 72, + 555, + 252 + ], + "type": "text", + "content": " less trainable parameters compared to LDU for aiMotive and DeLiVER respectively, making it more memory efficient. This ultimately shows the practicality of HyperDUM as a UQ method for real-world autonomous systems." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 262, + 471, + 275 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 262, + 471, + 275 + ], + "spans": [ + { + "bbox": [ + 313, + 262, + 471, + 275 + ], + "type": "text", + "content": "5. Discussions and Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 282, + 556, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 282, + 556, + 450 + ], + "spans": [ + { + "bbox": [ + 313, + 282, + 556, + 450 + ], + "type": "text", + "content": "Limitations and Future Work: We leveraged cloud computing resources using a single NVIDIA A100 GPU for training. Our method uses the traditional bundling technique, which requires labels for supervised learning of prototypes. The need for labels is the primary limitation of our approach. Although there have been implementations of semi-supervised and unsupervised versions [17, 18], there are no theoretical guarantees and/or justifications for their performance unlike the traditional bundling method [48]. We leave the exploration of these methods for uncertainty quantification to future work. Additionally, the FLOPs of the vector symbolic architectures may be further reduced due to the massively parallel operations of hyperdimensional computation. These optimizations were demonstrated in prior works [13, 19]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 450, + 556, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 556, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 556, + 665 + ], + "type": "text", + "content": "Conclusion: In this work, we propose an efficient method to quantify the epistemic uncertainty of machine learning models at the feature level. HyperDUM projects the latent features as hypervectors to form hyperdimensional prototypes. Similarity between hypervectors of new samples and learned prototypes is used as a proxy to estimate uncertainty. We demonstrate the practicality of HyperDUM through autonomous driving tasks, including object detection and semantic segmentation. Our results show that our method can enhance the performance and robustness of multimodal fusion models under uncertainties induced by various weather conditions and modality-specific corner cases. We achieve this by quantifying the feature uncertainty of each modality and applying an uncertainty-aware weighting layer prior to feature fusion and fine-tuning the post-fusion layers. Finally, we demonstrated that HyperDUM requires significantly less FLOPs and training parameters when compared to existing works, indicating its real-world practicality." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "content": "Acknowledgements: This work acknowledges the support of the Automotive Research Center (ARC), Cooperative Agreement W56HZV-24-2-0001 U.S. Army DEVCOM Ground Vehicle Systems Center (GVSC)." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22313" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "type": "text", + "content": "[1] Anastasios N Angelopoulos, Stephen Bates, et al. Conformal prediction: A gentle introduction. Foundations and Trends® in Machine Learning, 16(4):494-591, 2023. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 296, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 296, + 201 + ], + "type": "text", + "content": "[2] Sri Hrushikesh Varma Bhupathiraju, Jennifer Sheldon, Luke A Bauer, Vincent Bindschaedler, Takeshi Sugawara, and Sara Rampazzi. Emi-lidar: Uncovering vulnerabilities of lidar sensors in autonomous driving setting using electromagnetic interference. In Proceedings of the 16th ACM Conference on Security and Privacy in Wireless and Mobile Networks, pages 329-340, 2023. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 202, + 296, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 202, + 296, + 267 + ], + "spans": [ + { + "bbox": [ + 62, + 202, + 296, + 267 + ], + "type": "text", + "content": "[3] Mario Bijelic, Tobias Gruber, Fahim Mannan, Florian Kraus, Werner Ritter, Klaus Dietmayer, and Felix Heide. Seeing through fog without seeing fog: Deep multimodal sensor fusion in unseen adverse weather. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11682-11692, 2020. 5, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 269, + 295, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 269, + 295, + 323 + ], + "spans": [ + { + "bbox": [ + 62, + 269, + 295, + 323 + ], + "type": "text", + "content": "[4] Hermann Blum, Paul-Edouard Sarlin, Juan Nieto, Roland Siegwart, and Cesar Cadena. Fishyscapes: A benchmark for safe semantic segmentation in autonomous driving. In proceedings of the IEEE/CVF international conference on computer vision workshops, pages 0–0, 2019. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 324, + 296, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 324, + 296, + 378 + ], + "spans": [ + { + "bbox": [ + 62, + 324, + 296, + 378 + ], + "type": "text", + "content": "[5] Bertrand Charpentier, Daniel Zügner, and Stephan Gümnmann. Posterior network: Uncertainty estimation without ood samples via density-based pseudo-counts. Advances in neural information processing systems, 33:1356-1367, 2020. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 380, + 296, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 296, + 412 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 296, + 412 + ], + "type": "text", + "content": "[6] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 414, + 296, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 414, + 296, + 468 + ], + "spans": [ + { + "bbox": [ + 62, + 414, + 296, + 468 + ], + "type": "text", + "content": "[7] Michael Dusenberry, Ghassen Jerfel, Yeming Wen, Yian Ma, Jasper Snoek, Katherine Heller, Balaji Lakshminarayanan, and Dustin Tran. Efficient and scalable bayesian neural nets with rank-1 factors. In International conference on machine learning, pages 2782-2792. PMLR, 2020. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 469, + 296, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 469, + 296, + 513 + ], + "spans": [ + { + "bbox": [ + 62, + 469, + 296, + 513 + ], + "type": "text", + "content": "[8] Jamil Fayyad, Mohammad A Jaradat, Dominique Gruyer, and Homayoun Najjaran. Deep learning sensor fusion for autonomous vehicle perception and localization: A review. Sensors, 20(15):4220, 2020. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 514, + 296, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 514, + 296, + 568 + ], + "spans": [ + { + "bbox": [ + 62, + 514, + 296, + 568 + ], + "type": "text", + "content": "[9] Florian Fervers, Sebastian Bullinger, Christoph Bodensteiner, Michael Arens, and Rainer Stiefelhagen. Uncertainty-aware vision-based metric cross-view geolocation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21621-21631, 2023. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 569, + 296, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 569, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 569, + 296, + 624 + ], + "type": "text", + "content": "[10] Gianni Franchi, Xuanlong Yu, Andrei Bursuc, Emanuel Aldea, Severine Dubuisson, and David Filliat. Latent discriminant deterministic uncertainty. In European Conference on Computer Vision, pages 243-260. Springer, 2022. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 625, + 296, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 625, + 296, + 668 + ], + "spans": [ + { + "bbox": [ + 57, + 625, + 296, + 668 + ], + "type": "text", + "content": "[11] Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In international conference on machine learning, pages 1050-1059. PMLR, 2016. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "text", + "content": "[12] Ethan Goan and Clinton Fookes. Uncertainty in real-time semantic segmentation on embedded systems. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4490-4500, 2023. 1" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 139 + ], + "type": "text", + "content": "[13] Saransh Gupta, Justin Morris, Mohsen Imani, Ranganathan Ramkumar, Jeffrey Yu, Aniket Tiwari, Baris Aksanli, and Tajana Simunic Rosing. Thrifty: Training with hyperdimensional computing across flash hierarchy. In Proceedings of the 39th International Conference on Computer-Aided Design, pages 1-9, 2020. 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 141, + 555, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 141, + 555, + 195 + ], + "spans": [ + { + "bbox": [ + 317, + 141, + 555, + 195 + ], + "type": "text", + "content": "[14] Zongbo Han, Fan Yang, Junzhou Huang, Changqing Zhang, and Jianhua Yao. Multimodal dynamics: Dynamical fusion for trustworthy multimodal classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20707-20717, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 197, + 555, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 197, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 317, + 197, + 555, + 262 + ], + "type": "text", + "content": "[15] Florian Heidecker, Jasmin Breitenstein, Kevin Rösch, Jonas Löhdefink, Maarten Bieshaar, Christoph Stiller, Tim Fingscheidt, and Bernhard Sick. An application-driven conceptualization of corner cases for perception in highly automated driving. In 2021 IEEE Intelligent Vehicles Symposium (IV), pages 644-651. IEEE, 2021. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 264, + 555, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 264, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 317, + 264, + 555, + 319 + ], + "type": "text", + "content": "[16] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17853-17862, 2023. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 320, + 555, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 320, + 555, + 375 + ], + "spans": [ + { + "bbox": [ + 317, + 320, + 555, + 375 + ], + "type": "text", + "content": "[17] Mohsen Imani, Samuel Bosch, Mojan Javaheripi, Bita Rouhani, Xinyu Wu, Farinaz Koushanfar, and Tajana Rosing. Semihd: Semi-supervised learning using hyperdimensional computing. In 2019 IEEE/ACM International Conference on Computer-Aided Design (ICCAD), pages 1-8. IEEE, 2019. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 376, + 555, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 376, + 555, + 432 + ], + "spans": [ + { + "bbox": [ + 317, + 376, + 555, + 432 + ], + "type": "text", + "content": "[18] Mohsen Imani, Yeseong Kim, Thomas Worley, Saransh Gupta, and Tajana Rosing. Hdcluster: An accurate clustering using brain-inspired high-dimensional computing. In 2019 Design, Automation & Test in Europe Conference & Exhibition (DATE), pages 1591-1594. IEEE, 2019. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 434, + 555, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 434, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 317, + 434, + 555, + 498 + ], + "type": "text", + "content": "[19] Mohsen Imani, Saikishan Pampana, Saransh Gupta, Minxuan Zhou, Yeseong Kim, and Tajana Rosing. Dual: Acceleration of clustering algorithms using digital-based processing in-memory. In 2020 53rd Annual IEEE/ACM International Symposium on Microarchitecture (MICRO), pages 356-371. IEEE, 2020. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 500, + 555, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 500, + 555, + 543 + ], + "spans": [ + { + "bbox": [ + 317, + 500, + 555, + 543 + ], + "type": "text", + "content": "[20] Masha Itkina and Mykel Kochenderfer. Interpretable self-aware neural networks for robust trajectory prediction. In Conference on Robot Learning, pages 606-617. PMLR, 2023. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 545, + 555, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 545, + 555, + 590 + ], + "spans": [ + { + "bbox": [ + 317, + 545, + 555, + 590 + ], + "type": "text", + "content": "[21] Ding Jia, Jianyuan Guo, Kai Han, Han Wu, Chao Zhang, Chang Xu, and Xinghao Chen. Geminifusion: Efficient pixelwise multimodal fusion for vision transformer. International Conference on Machine Learning (ICML), 2024. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 591, + 555, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 591, + 555, + 656 + ], + "spans": [ + { + "bbox": [ + 317, + 591, + 555, + 656 + ], + "type": "text", + "content": "[22] Xiaotao Jia, Jianlei Yang, Runze Liu, Xueyan Wang, Sorin Dan Cotofana, and Weisheng Zhao. Efficient computation reduction in bayesian neural networks through feature decomposition and memorization. IEEE transactions on neural networks and learning systems, 32(4):1703-1712, 2020. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 658, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 555, + 713 + ], + "type": "text", + "content": "[23] Yongchan Kwon, Joong-Ho Won, Beom Joon Kim, and Myunghee Cho Paik. Uncertainty quantification using bayesian neural networks in classification: Application to biomedical image segmentation. Computational Statistics & Data Analysis, 142:106816, 2020. 1" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "22314" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 296, + 117 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 296, + 117 + ], + "type": "text", + "content": "[24] Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scalable predictive uncertainty estimation using deep ensembles. Advances in neural information processing systems, 30, 2017. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 296, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 296, + 174 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 296, + 174 + ], + "type": "text", + "content": "[25] Simon D Levy and Ross Gayler. Vector symbolic architectures: A new building material for artificial general intelligence. In Proceedings of the 2008 conference on artificial general intelligence 2008: Proceedings of the first AGI conference, pages 414-418, 2008. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 295, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 295, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 295, + 218 + ], + "type": "text", + "content": "[26] Hao Li, Jingkuan Song, Lianli Gao, Xiaosu Zhu, and Hengtao Shen. Prototype-based aleatoric uncertainty quantification for cross-modal retrieval. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 295, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 295, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 295, + 262 + ], + "type": "text", + "content": "[27] Xiaotong Li, Yongxing Dai, Yixiao Ge, Jun Liu, Ying Shan, and Ling-Yu Duan. Uncertainty modeling for out-of-distribution generalization. arXiv preprint arXiv:2202.03958, 2022.4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 295, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 295, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 295, + 308 + ], + "type": "text", + "content": "[28] You Li and Javier Ibanez-Guzman. Lidar for autonomous driving: The principles, challenges, and trends for automotive lidar and perception systems. IEEE Signal Processing Magazine, 37(4):50-61, 2020. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 309, + 295, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 295, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 295, + 365 + ], + "type": "text", + "content": "[29] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1477-1485, 2023. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 295, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 295, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 295, + 409 + ], + "type": "text", + "content": "[30] Ming Liang, Bin Yang, Shenlong Wang, and Raquel Urtasun. Deep continuous fusion for multi-sensor 3d object detection. In Proceedings of the European conference on computer vision (ECCV), pages 641-656, 2018. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 411, + 295, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 295, + 476 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 295, + 476 + ], + "type": "text", + "content": "[31] Arnav Vaibhav Malawade, Trier Mortlock, and Mohammad Abdullah Al Faruque. Hydraulics: Context-aware selective sensor fusion for robust and efficient autonomous vehicle perception. In 2022 ACM/IEEE 13th International Conference on Cyber-Physical Systems (ICCPs), pages 68–79. IEEE, 2022. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 478, + 295, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 295, + 511 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 295, + 511 + ], + "type": "text", + "content": "[32] Andrey Malinin and Mark Gales. Predictive uncertainty estimation via prior networks. Advances in neural information processing systems, 31, 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 512, + 295, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 295, + 577 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 295, + 577 + ], + "type": "text", + "content": "[33] Tamas Matuszka, Ivan Barton, Adam Butykai, Peter Hajas, David Kiss, Domonkos Kovacs, Sandor Kunsagi-Mate, Peter Lengyel, Gabor Nemeth, Leventa PetHo, et al. aimotive dataset: A multimodal dataset for robust autonomous driving with long-range perception. arXiv preprint arXiv:2211.09445, 2022. 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 579, + 295, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 295, + 634 + ], + "type": "text", + "content": "[34] Lu Mi, Hao Wang, Yonglong Tian, Hao He, and Nir N Shavit. Training-free uncertainty estimation for dense regression: Sensitivity as a surrogate. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10042-10050, 2022. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 635, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 295, + 689 + ], + "type": "text", + "content": "[35] Trier Mortlock, Arnav Malawade, Kohei Tsujio, and Mohammad Al Faruque. Castnet: a context-aware, spatio-temporal dynamic motion prediction ensemble for autonomous driving. ACM Transactions on Cyber-Physical Systems, 8(2):1–20, 2024. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 295, + 714 + ], + "type": "text", + "content": "[36] Jishnu Mukhoti, Andreas Kirsch, Joost van Amersfoort, Philip HS Torr, and Yarin Gal. Deep deterministic uncertainty:" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "text", + "content": "A new simple baseline. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24384-24394, 2023. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 106, + 555, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 106, + 555, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 106, + 555, + 171 + ], + "type": "text", + "content": "[37] Yang Ni, Hanning Chen, Prathyush Poduval, Zhuowen Zou, Pietro Mercati, and Mohsen Imani. Brain-inspired trustworthy hyperdimensional computing with efficient uncertainty quantification. In 2023 IEEE/ACM International Conference on Computer Aided Design (ICCAD), pages 01-09. IEEE, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 172, + 555, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 172, + 555, + 227 + ], + "spans": [ + { + "bbox": [ + 316, + 172, + 555, + 227 + ], + "type": "text", + "content": "[38] Igor Nunes, Mike Heddes, Tony Givargis, Alexandru Nicolau, and Alex Veidenbaum. Graphhd: Efficient graph classification using hyperdimensional computing. In 2022 Design, Automation & Test in Europe Conference & Exhibition (DATE), pages 1485-1490. IEEE, 2022. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 228, + 555, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 555, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 555, + 271 + ], + "type": "text", + "content": "[39] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch. In NIPS-W, 2017. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 272, + 555, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 555, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 555, + 316 + ], + "type": "text", + "content": "[40] Luis Manuel Pereira, Addisson Salazar, and Luis Vergara. On comparing early and late fusion methods. In International Work-Conference on Artificial Neural Networks, pages 365-378. Springer, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 316, + 555, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 555, + 370 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 555, + 370 + ], + "type": "text", + "content": "[41] Janis Postels, Francesco Ferroni, Huseyin Coskun, Nassir Navab, and Federico Tombari. Sampling-free epistemic uncertainty estimation using approximated variance propagation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2931-2940, 2019. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 371, + 555, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 371, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 371, + 555, + 415 + ], + "type": "text", + "content": "[42] Janis Postels, Mattia Segu, Tao Sun, Luca Sieber, Luc Van Gool, Fisher Yu, and Federico Tombari. On the practicality of deterministic epistemic uncertainty. arXiv preprint arXiv:2107.00649, 2021. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 415, + 555, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 555, + 448 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 555, + 448 + ], + "type": "text", + "content": "[43] Ali Rahimi and Benjamin Recht. Random features for large-scale kernel machines. Advances in neural information processing systems, 20, 2007. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 449, + 555, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 555, + 503 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 555, + 503 + ], + "type": "text", + "content": "[44] Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Schölkopf, Thomas Brox, and Peter Gehler. Towards total recall in industrial anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14318-14328, 2022. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 504, + 555, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 504, + 555, + 536 + ], + "spans": [ + { + "bbox": [ + 316, + 504, + 555, + 536 + ], + "type": "text", + "content": "[45] Kenny Schlegel, Peer Neubert, and Peter Protzel. A comparison of vector symbolic architectures. Artificial Intelligence Review, 55(6):4523-4555, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 537, + 555, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 555, + 592 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 555, + 592 + ], + "type": "text", + "content": "[46] Jiahui She, Yibo Hu, Hailin Shi, Jun Wang, Qiu Shen, and Tao Mei. Dive into ambiguity: Latent distribution mining and pairwise uncertainty estimation for facial expression recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6248-6257, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 593, + 555, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 593, + 555, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 593, + 555, + 634 + ], + "type": "text", + "content": "[47] Apoorv Singh. Transformer-based sensor fusion for autonomous driving: A survey. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3312-3317, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 635, + 555, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 555, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 555, + 679 + ], + "type": "text", + "content": "[48] Anthony Thomas, Sanjoy Dasgupta, and Tajana Rosing. A theoretical perspective on hyperdimensional computing. Journal of Artificial Intelligence Research, 72:215-249, 2021. 3, 4, 8, 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 681, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 555, + 713 + ], + "type": "text", + "content": "[49] Junjiao Tian, Wesley Cheung, Nathaniel Glaser, Yen-Cheng Liu, and Zsolt Kira. Uno: Uncertainty-aware noisy-or multimodal fusion for unanticipated input degradation. In 2020" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22315" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 676 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "IEEE International Conference on Robotics and Automation (ICRA), pages 5716-5723. IEEE, 2020. 5, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 297, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 297, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 297, + 140 + ], + "type": "text", + "content": "[50] Stefan Van der Walt, Johannes L Schonberger, Juan Nunez-Iglesias, François Boulogne, Joshua D Warner, Neil Yager, Emmanuelle Gouillart, and Tony Yu. scikit-image: image processing in python. PeerJ, 2:e453, 2014. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 296, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 296, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 296, + 184 + ], + "type": "text", + "content": "[51] Jorge Vargas, Suleiman Alsweiss, Onur Token, Rahul Razdan, and Joshua Santos. An overview of autonomous vehicles sensors and their vulnerability to weather conditions. Sensors, 21(16):5397, 2021. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 296, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 296, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 296, + 229 + ], + "type": "text", + "content": "[52] Junyao Wang, Sitao Huang, and Mohsen Imani. Disthd: A learner-aware dynamic encoding method for hyperdimensional classification. In 2023 60th ACM/IEEE Design Automation Conference (DAC), pages 1-6. IEEE, 2023. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 231, + 296, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 296, + 296 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 296, + 296 + ], + "type": "text", + "content": "[53] Junyao Wang, Arnav Vaibhav Malawade, Junhong Zhou, Shih-Yuan Yu, and Mohammad Abdullah Al Faruque. Rs2g: Data-driven scene-graph extraction and embedding for robust autonomous perception and scenario understanding. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 7493-7502, 2024. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 297, + 296, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 296, + 351 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 296, + 351 + ], + "type": "text", + "content": "[54] Shaojie Wang, Tong Wu, Ayan Chakrabarti, and Yevgeniy Vorobeychik. Adversarial robustness of deep sensor fusion models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2387-2396, 2022. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 353, + 296, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 353, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 353, + 296, + 407 + ], + "type": "text", + "content": "[55] Samuel Wilson, Tobias Fischer, Niko Sünderhauf, and Feras Dayoub. Hyperdimensional feature fusion for out-of-distribution detection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2644-2654, 2023. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 409, + 296, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 409, + 296, + 452 + ], + "spans": [ + { + "bbox": [ + 56, + 409, + 296, + 452 + ], + "type": "text", + "content": "[56] De Jong Yeong, Gustavo Velasco-Hernandez, John Barry, and Joseph Walsh. Sensor and sensor fusion technology in autonomous vehicles: A review. Sensors, 21(6):2140, 2021. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 454, + 296, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 454, + 296, + 497 + ], + "spans": [ + { + "bbox": [ + 56, + 454, + 296, + 497 + ], + "type": "text", + "content": "[57] Tao Yu, Yichi Zhang, Zhiru Zhang, and Christopher M De Sa. Understanding hyperdimensional computing for parallel single-pass learning. Advances in Neural Information Processing Systems, 35:1157-1169, 2022. 3, 8, 1, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 498, + 296, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 498, + 296, + 563 + ], + "spans": [ + { + "bbox": [ + 56, + 498, + 296, + 563 + ], + "type": "text", + "content": "[58] Jiaming Zhang, Ruiping Liu, Hao Shi, Kailun Yang, Simon Reiβ, Kunyu Peng, Haodong Fu, Kaiwei Wang, and Rainer Stiefelhagen. Delivering arbitrary-modal semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1136-1147, 2023. 5, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 566, + 296, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 566, + 296, + 609 + ], + "spans": [ + { + "bbox": [ + 56, + 566, + 296, + 609 + ], + "type": "text", + "content": "[59] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4490-4499, 2018. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 610, + 296, + 676 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 610, + 296, + 676 + ], + "spans": [ + { + "bbox": [ + 56, + 610, + 296, + 676 + ], + "type": "text", + "content": "[60] Zijian Zhu, Yichi Zhang, Hai Chen, Yinpeng Dong, Shu Zhao, Wenbo Ding, Jiachen Zhong, and Shibao Zheng. Understanding the robustness of 3d object detection with bird's-eye-view representations in autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21600-21610, 2023. 1" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "22316" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Hypergraph Vision Transformers_ Images are More than Nodes, More than Edges/b3c66b72-bf5c-47e0-a296-e834f5e49c0f_content_list.json b/2025/Hypergraph Vision Transformers_ Images are More than Nodes, More than Edges/b3c66b72-bf5c-47e0-a296-e834f5e49c0f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c5281b2065c77b4663a609e0a6b086e78e1f6036 --- /dev/null +++ b/2025/Hypergraph Vision Transformers_ Images are More than Nodes, More than Edges/b3c66b72-bf5c-47e0-a296-e834f5e49c0f_content_list.json @@ -0,0 +1,1606 @@ +[ + { + "type": "text", + "text": "Hypergraph Vision Transformers: Images are More than Nodes, More than Edges", + "text_level": 1, + "bbox": [ + 91, + 130, + 906, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Joshua Fixelle \nUniversity of Virginia \njf9fk@virginia.edu", + "bbox": [ + 410, + 181, + 584, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advancements in computer vision have highlighted the scalability of Vision Transformers (ViTs) across various tasks, yet challenges remain in balancing adaptability, computational efficiency, and the ability to model higher-order relationships. Vision Graph Neural Networks (ViGs) offer an alternative by leveraging graph-based methodologies but are hindered by the computational bottlenecks of clustering algorithms used for edge generation. To address these issues, we propose the Hypergraph Vision Transformer (HgVT), which incorporates a hierarchical bipartite hypergraph structure into the vision transformer framework to capture higher-order semantic relationships while maintaining computational efficiency. HgVT leverages population and diversity regularization for dynamic hypergraph construction without clustering, and expert edge pooling to enhance semantic extraction and facilitate graph-based image retrieval. Empirical results demonstrate that HgVT achieves strong performance on image classification and retrieval, positioning it as an efficient framework for semantic-based vision tasks.", + "bbox": [ + 89, + 300, + 485, + 589 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 617, + 220, + 633 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Computer vision has recently transitioned from the historically dominant Convolutional Neural Networks (CNNs) [20, 28, 30] to the increasingly prominent Vision Transformers (ViTs), which have quickly embedded themselves as the new de facto standard [12, 38]. This shift reflects the broader success of transformers in natural language processing [11, 54, 56] and is driven by the remarkable scalability of ViTs across various tasks such as image classification [53, 61], semantic segmentation [26, 63], and image retrieval [2, 29]. While hybrid models like hierarchical attention and CNN-ViT methods [18, 19, 33] aim to balance computational load and flexibility, ViTs still struggle with focusing on salient features rather than comprehensive image understanding [2, 9, 12, 40]. This highlights the need for methods that improve both efficiency and semantic accuracy.", + "bbox": [ + 89, + 643, + 485, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Within the spectrum of novel architectures, Vision Graph Neural Networks (ViGs) [16] and Vision Hypergraph Neural", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Networks (ViHGNNs) [17] leverage graph-based topologies to advance image processing. Unlike CNNs, which harness locality and translation-invariance through densely connected pixel grids and repeated convolutions, both ViTs and ViGs represent images as sets of patches. In ViTs, each patch acts as a vertex within a maximally connected graph, creating semantically weak connections through self-attention. ViGs enhance this by using clustering algorithms to detect edge groupings and applying graph convolutions to these clusters, forming meaningful patch relationships. ViHGNNs extend these capabilities by employing hyperedges that capture complex, higher-order relationships, enriching understanding of the images. These methodologies are depicted in Figure 1.", + "bbox": [ + 511, + 268, + 906, + 467 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While graph-based models like ViG and ViHGNN have introduced notable advancements in visual perception, two critical observations emerge regarding these architectures:", + "bbox": [ + 511, + 470, + 906, + 516 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. In existing vision GNN models [16, 17, 35, 36], edge features are primarily used for basic vertex-to-vertex communication and are not integrated across successive layers: a strategy that could enhance cumulative learning and improve classification accuracy.", + "2. The computational complexities associated with clustering algorithms used for edge generation, such as KNN in ViG and Fuzzy C-Means in ViHGNN, pose significant computational bottlenecks. Approaches like MobileViG [35] and GreedyViG [36] attempt to mitigate these challenges with static graph structures and adding dynamic masking, but do so by trading adaptability for efficiency, failing to achieve a well-balanced solution." + ], + "bbox": [ + 514, + 518, + 908, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In response to limitations in existing graph-based models, we propose the Hypergraph Vision Transformer (HgVT), which advances the hypergraph concept with a bipartite representation where hyperedge features and image patches (vertices) are continuously processed. Unlike traditional models that use graph convolutions, HgVT employs structured multihead attention for efficient vertex-hyperedge message passing and incorporates a dynamic querying mechanism that constructs graph structures in $\\mathcal{O}(|V| \\cdot E)$ time complexity, where $E < |V|$ . This graph structure is then utilized in attention masking to balance structural adaptability with computational efficiency. Furthermore, HgVT integrates", + "bbox": [ + 511, + 719, + 910, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9751", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3d7919bc3e33a1d7a2fc5f7b472d8873ead76dd6883c0ebece9cc42ae71b2beb.jpg", + "image_caption": [ + "(a) CNN" + ], + "image_footnote": [], + "bbox": [ + 205, + 90, + 290, + 145 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0aea464e472c471a64365e7bb55212f6001bb9141c514e309474e24d08c74bf4.jpg", + "image_caption": [ + "(c) ViG" + ], + "image_footnote": [], + "bbox": [ + 303, + 90, + 535, + 152 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/cb6a1a05cf39a3f65ae9e5efed64cbed57528fd570db7d877dba53201c454ee3.jpg", + "image_caption": [ + "(e) HgVT" + ], + "image_footnote": [], + "bbox": [ + 547, + 90, + 795, + 229 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/72cd0db8a09ed4a1847b085b5606afae86b9b78120c967bcdffa2365fb1be12e.jpg", + "image_caption": [ + "(b) ViT" + ], + "image_footnote": [], + "bbox": [ + 209, + 165, + 312, + 234 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/03f6cb301c05c4006f9a09f32ae726be2c2c02775d834efd107a30adb84fa2b1.jpg", + "image_caption": [ + "(d) ViHGNN", + "Figure 1. Comparison of Graph Structures for different methods. Showing (a) CNNs, (b) Vision Transformers, (c) ViG with a KNN clustered GNN, (d) ViHGNN with clustered hyperedges, and (e) our proposed HgVT method. Strong group edges shown with solid lines; weak edges with dashed lines. Hyperedges shown with shaded regions; less dominant hyperedges with gray dashed regions." + ], + "image_footnote": [], + "bbox": [ + 323, + 165, + 542, + 229 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "virtual elements into vertices and hyperedges, forming intermediate links that create a hierarchical semantic structure, as illustrated in Fig. 1e. Attention masking is then used to restrict message passing, and virtual hyperedge features are leveraged for classification. Our contributions are as follows:", + "bbox": [ + 89, + 303, + 485, + 378 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose the Hypergraph Vision Transformer (HgVT), which integrates a hierarchical bipartite hypergraph structure within a vision transformer framework. Our isotropic HgVT-Ti model achieves a Top-1 accuracy of $76.2\\%$ on the ImageNet-1k classification task, surpassing the prior state-of-the-art by $1.9\\%$ , demonstrating the efficacy of hypergraph-based learning within vision transformers.", + "- We introduce population and diversity regularization strategies that enable dynamic hypergraph structure construction in HgVT, allowing the model to self-sparseify connections without relying on traditional clustering techniques.", + "- We implement expert edge pooling, a pooling approach that selects edges based on learned confidence scores, facilitating efficient representation pruning and graph extraction. This approach demonstrates strong semantic clustering behavior across macro-classes and achieves competitive image retrieval performance compared to other feature extractors, while maintaining a more compact model size." + ], + "bbox": [ + 89, + 383, + 483, + 654 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 671, + 232, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision Transformers. Vision Transformers (ViTs) proposed by [12] and refined by [2, 38, 53] use self-attention to process image patches as sequences, scaling to complex datasets and tasks. Recent ViTs have reintroduced spatial hierarchies by leveraging local attention [19, 33], integrating sparse global summaries [68], and employing biomimetic modeling to focus on key regions within images [49]. However, current models tend to focus on the most salient objects and patch-level similarities, ignoring global structure. HgVT addresses this by introducing bipartite hypergraphs to model higher-order relationships for improved semantic understanding.", + "bbox": [ + 89, + 698, + 485, + 864 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Graph-Based Vision Models and Clustering. Graph Neural Networks (GNNs), initially conceptualized by [46], have", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "been applied to vision tasks through Vision Graph Neural Networks (ViGs) [16], which exhibit improved accuracy over ViTs on common vision tasks. ViGs use graph convolutions to model image patch relationships on a graph structure, typically constructed by iterative clustering algorithms such as KNN and Fuzzy C-Means, which introduce computational overhead. Recent methods avoid clustering inefficiencies with static graph structures [35, 36], sacrificing adaptability. HgVT instead introduces a dynamic graph construction method, relying on cosine similarity from learned features to enable efficient, non-iterative, adaptive clustering.", + "bbox": [ + 511, + 303, + 906, + 469 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Hypergraph-Based Methods. While previously used in many computer vision tasks [15, 23, 24], hypergraphs have recently been incorporated into vision GNNs [17, 50], improving their ability to model complex multi-way relationships. However, these methods treat hypergraphs as an intermediate tool rather than producing a hypergraph to represent underlying images, preventing their use in downstream tasks. HgVT instead iteratively refines a hypergraph through subsequent network layers to produce structured representations.", + "bbox": [ + 511, + 474, + 908, + 612 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Hierarchical Hypergraphs", + "text_level": 1, + "bbox": [ + 511, + 625, + 759, + 642 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Graphs and Basic Notations. Graphs are powerful mathematical tools for representing structured information, applicable across diverse disciplines. A graph $\\mathcal{G}$ is defined as a pair $(\\mathcal{V},\\mathcal{E})$ , where $\\mathcal{V} = \\{v_{1},v_{2},\\ldots ,v_{N}\\}$ is a set of vertices, and $\\mathcal{E} = \\{e_{ij}|(v_i,v_j)\\}$ is a set of directed edges. For undirected graphs, the edges are symmetric, such that $e_{ij} = e_{ji}$ . Each edge $e_{ij}$ connects a pair of vertices $v_{i}$ and $v_{j}$ where $v_{i},v_{j}\\in \\mathcal{V}$ . The adjacency matrix $\\hat{\\mathbf{A}}$ is a binary matrix $\\{0,1\\}^{|V|\\times |V|}$ , representing the presence (1) or absence (0) of an edge between each pair of vertices. Similarly, an edge weight matrix can be defined as $\\mathbf{A}\\in \\mathbb{R}^{|V|\\times |V|}$ to quantify the strength or capacity of these connections.", + "bbox": [ + 511, + 648, + 908, + 832 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Graph Convolution Networks (GCNs). Building on this foundation, GCNs utilize vertex feature matrices $\\mathbf{X} \\in \\mathbb{R}^{|V| \\times d}$ to encode vertex properties. Their core mechanism, message passing, updates vertex features through a convolution", + "bbox": [ + 511, + 839, + 908, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9752", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "with a learned projection matrix $\\mathbf{W} \\in \\mathbb{R}^{d \\times d}$ and a non-linear activation, guided by the adjacency matrix $\\hat{\\mathbf{A}}$ , which specifies neighboring vertices. Adjacency features $\\mathbf{X}_{\\mathrm{adj}} \\in \\mathbb{R}^{|V| \\times d_a}$ , typically set as $\\mathbf{X}_{\\mathrm{adj}} = \\mathbf{X}$ , enable dynamic updates to $\\hat{\\mathbf{A}}$ and the edge weight matrix $\\mathbf{A}$ , refining the graph structure through learned interactions. However, GCNs are unable to model multi-vertex relationships as edges in $\\mathcal{E}$ are pairwise.", + "bbox": [ + 89, + 90, + 483, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Hypergraphs and Bipartite Representations", + "text_level": 1, + "bbox": [ + 89, + 207, + 465, + 224 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f02d2b1b429d3ba3aabe6d6c87fcfd98faee49b413293856ad186333b6f851c7.jpg", + "image_caption": [ + "(a) Hypergraph" + ], + "image_footnote": [], + "bbox": [ + 124, + 234, + 256, + 323 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f8d8f04c1031e4eb2eb532b2bca4efa2f9e8ff744965ef4ac57bdba3c12fe7c9.jpg", + "image_caption": [ + "(b) Bipartite Form", + "Figure 2. Comparison of (a) hypergraph and (b) equivalent bipartite representation from Fig. 1d, showing five hyperedges $(e_0 - e_4)$ ." + ], + "image_footnote": [], + "bbox": [ + 279, + 233, + 452, + 323 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To overcome the pairwise limitation inherent in traditional graphs, hypergraphs offer a robust solution by extending the concept of edges to hyperedges, which connect multiple vertices simultaneously. In a hypergraph $\\mathcal{H} = (\\mathcal{V},\\mathcal{E})$ , hyperedges $e_j\\in \\mathcal{E}$ each connect a subset of vertices, defined as $e_j = \\{v_i\\mid v_i\\in \\mathcal{V}$ and $i\\in I_j\\}$ , where $I_{j}$ is the set of indices for vertices that are included in hyperedge $e_j$ . The set $I_{j}$ directly corresponds to the nonzero entries of the $j$ -th column of the incidence matrix $\\mathbf{H}\\in \\{0,1\\}^{|\\mathcal{V}|\\times |\\mathcal{E}|}$ , where $\\mathbf{H}_{ij} = 1$ if vertex $v_{i}$ is included in the hyperedge $e_j$ . This structure captures complex relationships, making hypergraphs valuable for modeling networked systems and group interactions.", + "bbox": [ + 89, + 382, + 483, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Hypergraphs can alternatively be described using a bipartite representation, where the vertex set $\\mathcal{V}$ and hyperedge set $\\mathcal{E}$ form distinct groups linked by the incidence matrix $\\mathbf{H}$ (refer to Fig. 2). This representation results in a new graph $\\mathcal{G}_B = (\\mathcal{V},\\mathcal{E},\\mathcal{E}_B)$ , where $\\mathcal{V}$ represents the original vertices of the hypergraph, and the elements in $\\mathcal{E}$ correspond to hyperedges. The edges in $\\mathcal{E}_B$ , denoted as $\\epsilon_{ve} = (\\nu_v,\\nu_e)$ exist if $\\mathbf{H}_{ve} = 1$ , with $\\nu_{v}\\in \\mathcal{V}$ and $\\nu_{e}\\in \\mathcal{E}$ , linking $\\mathcal{V}$ and $\\mathcal{E}$ .", + "bbox": [ + 89, + 564, + 483, + 684 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the bipartite graph $\\mathcal{G}_B$ , the corresponding adjacency matrix can be simplified as $\\hat{\\mathbf{A}} = \\mathbf{H}$ for $\\mathcal{E} \\to \\mathcal{V}$ interactions, and $\\hat{\\mathbf{A}} = \\mathbf{H}^T$ for $\\mathcal{V} \\to \\mathcal{E}$ . Drawing on principles similar to those in ViHGNN [17], the edge weight matrix $\\mathbf{A}$ can be interpreted as fuzzy membership weights, enabling graded interactions across GNN layers. Complementing this setup, the feature matrices are split into $\\mathbf{X}^{(V)} \\in \\mathbb{R}^{|\\mathcal{V}| \\times d_v}$ and $\\mathbf{X}^{(E)} \\in \\mathbb{R}^{|\\mathcal{E}| \\times d_e}$ , along with their correspond adjacency feature matrices $\\mathbf{X}_{\\mathrm{adj}}^{(V)}$ and $\\mathbf{X}_{\\mathrm{adj}}^{(E)}$ , mirroring traditional GNNs.", + "bbox": [ + 89, + 686, + 483, + 823 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Imposing Hierarchical Structure in Images", + "text_level": 1, + "bbox": [ + 89, + 832, + 457, + 849 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To enhance the capability of hypergraphs in image analysis, we draw inspiration from the register tokens introduced in [8], which act to summarize information that otherwise", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "manifests as noise in areas of low visual significance. Similarly, this work integrates virtual vertices $(v\\mathcal{V})$ , alongside typical image patch vertices $(i\\mathcal{V})$ , and introduces virtual hyperedges $(v\\mathcal{E})$ , alongside primary hyperedges $(p\\mathcal{E})$ , providing levels of semantic feature aggregation and relational abstraction. Notably, these virtual elements, illustrated in Fig. 1e, do not correspond to specific image patches; instead, they are learned embeddings used for semantic summarization and capturing high-level abstract information.", + "bbox": [ + 511, + 90, + 906, + 226 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our proposed hypergraph, constructed from image $I$ as $\\mathcal{G}_B(I)$ , integrates primary and virtual sets, forming $\\mathcal{V} = i\\mathcal{V} \\cup v\\mathcal{V}$ and $\\mathcal{E} = p\\mathcal{E} \\cup v\\mathcal{E}$ , with statically masked communication pathways to enforce a hierarchical structure. Primary hyperedges $(p\\mathcal{E})$ interact with all vertices to support unrestricted semantic aggregation, whereas virtual hyperedges $(v\\mathcal{E})$ , designated for class predictions, connect solely with virtual vertices $(v\\mathcal{V})$ . These restrictions separate visual and abstract information, thereby producing a graph structure suitable for use in downstream applications.", + "bbox": [ + 511, + 227, + 908, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. A Hypergraph Vision Transformer", + "text_level": 1, + "bbox": [ + 511, + 387, + 828, + 404 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The Hypergraph Vision Transformer (HgVT) adapts the architecture of standard Vision Transformers by incorporating bipartite hypergraph features for enhanced image analysis capabilities. Like Vision Transformers, HgVT begins with a patch embedding layer, followed by an isotropic stack of $L \\times \\mathrm{HgVT}$ blocks, culminating in feature pooling and a classifier head. The bipartite hypergraph is represented by four principal feature matrices $-\\mathbf{X}^{(V)}$ , $\\mathbf{X}_{\\mathrm{adj}}^{(V)}$ , $\\mathbf{X}^{(E)}$ , and $\\mathbf{X}_{\\mathrm{adj}}^{(E)}$ which are updated iteratively and in an interleaved fashion within each block. Each block constructs a new adjacency matrix $\\mathbf{A}$ from $\\mathbf{X}_{\\mathrm{adj}}^{(V)}$ and $\\mathbf{X}_{\\mathrm{adj}}^{(E)}$ , enabling flexible adjustments to the hypergraph structure. As illustrated in Fig. 3, this modular process allows for the continuous integration and processing of these matrices within each HgVT block.", + "bbox": [ + 511, + 412, + 908, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Hyperedges as Communication Pools", + "text_level": 1, + "bbox": [ + 511, + 636, + 834, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Each HgVT block processes both vertex and edge information, refining them from the previous block based on a newly constructed graph structure. Initially, adjacency mask computation (detailed in the next section) determines the connectivity for the subsequent processing steps within each block, dynamically adjusting to the updated feature matrices from the previous block. Three attention layers – vertex self-attention, edge aggregate attention, and edge distribution attention – operate sequentially to enhance feature integration and facilitate effective message passing along the hyperedges formed in the adjacency computation step. Finally, separate feed-forward networks process vertex and edge features independently, ensuring specialized treatment for the two distinct sets within the bipartite hypergraph, preserving the unique properties of each set. Operational details of these components are further described in Appendix A.", + "bbox": [ + 511, + 657, + 908, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9753", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/14b90bfc81895302ca98b111966fca9f7608cc4dfe160835ee33c206808492ca.jpg", + "image_caption": [ + "Figure 3. HgVT Architecture Diagram, composed of stacked HgVT blocks with adjacency matrix $\\mathbf{A}$ , vertex features $\\mathbf{X}^{(V)}$ , and hyperedge features $\\mathbf{X}^{(E)}$ . Edge attention flow is shown with gray arrows; input norms and residual adds are omitted for clarity." + ], + "image_footnote": [], + "bbox": [ + 161, + 93, + 831, + 272 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Hypergraph Feature Processing. Within each HgVT block, two distinct point-wise feed-forward networks (FFNs) independently process vertex and hyperedge features, aligning with the bipartite structure of the hypergraph. Each FFN integrates both the element features and their corresponding adjacency features through a fully connected layer, improving the model's ability to synthesize relationships. Processing both feature types within the same FFN layer allows adjacency information to be handled directly, bypassing the need for graph-based message passing and improving computational efficiency. Moreover, parameter overhead can be reduced by optionally tying edge and vertex FFN weights.", + "bbox": [ + 89, + 321, + 485, + 505 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/13b6d70a4bb19ec0622873859df6a17fc3c5cf7349d6cf9dd7b72b6f9789f8a8.jpg", + "image_caption": [ + "(a) Communication Pool." + ], + "image_footnote": [], + "bbox": [ + 96, + 518, + 272, + 613 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/dc882fb198a1561d69ebe306d286227f612cd16894e4067a343795c9c6459166.jpg", + "image_caption": [ + "(b) Attention Sparsity.", + "Figure 4. (a) Hyperedge Communication Pool Flow with edges $\\mathcal{E}$ and member vertices $\\nu_{e}$ ; (b) Attention Sparsity (Mean and std) for HgVT-S on the ImageNet-1k Validation set." + ], + "image_footnote": [], + "bbox": [ + 287, + 515, + 478, + 614 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Hyperedges as Communication Pools. Hypergraph GNNs typically employ a gather $\\rightarrow$ scatter mechanism for processing vertex-hyperedge interactions, whereas HgVT reconceptualizes hyperedges as communication pools that facilitate information flow among vertices and their associated hyperedges. Specifically, vertex self-attention manages vertex-to-vertex $(\\mathcal{V}\\rightarrow \\mathcal{V})$ interactions within hyperedges, edge aggregate attention orchestrates the flow from vertices to hyperedges $(\\mathcal{V}\\rightarrow \\mathcal{E})$ , and edge distribution attention handles the reverse, from hyperedges back to vertices $(\\mathcal{E}\\rightarrow \\mathcal{V})$ . By segmenting the attention operations, HgVT efficiently approximates an all-to-all feature transfer within hyperedges, as illustrated in Fig. 4a, significantly reducing the quadratic complexity associated with full attention mechanisms.", + "bbox": [ + 89, + 689, + 485, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Sparse and Fuzzy Attention. Building upon the dynamic communication pools concept, HgVT employs both sparse and fuzzy attention mechanisms to further optimize computational efficiency. Vertex self-attention is applied selectively to pairs of vertices connected by common hyperedges, as defined by the adjacency matrix $\\hat{\\mathbf{A}}$ , resulting in a sparse attention pattern. As sparsity increases with network depth - demonstrated in Fig. 4b - computational load decreases, while still maintaining compatibility with dense attention during training. Conversely, the edge aggregate and distribution attention mechanisms utilize cross-attention between the vertex and edge feature matrices, $\\mathbf{X}^{(V)}$ and $\\mathbf{X}^{(E)}$ , modulated by the soft adjacency matrix $\\mathbf{A}$ . This modulation, akin to Fuzzy C-Means in ViHGNNs [17], adjusts attention logits based on soft memberships to the individual hyperedges, dynamically adapting to the hypergraph structure and providing a mechanism for gradient flow into the adjacency matrix generation. Furthermore, by thresholding the soft adjacency matrix during inference, the edge attention mechanisms can be converted into a sparse cross-attention mechanism, thereby reducing computational overhead.", + "bbox": [ + 511, + 321, + 908, + 640 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Dynamic Adjacency Formation", + "text_level": 1, + "bbox": [ + 511, + 647, + 790, + 664 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "HgVT dynamically establishes its hypergraph structure to adapt to the varying semantic and spatial structures of different image inputs. It employs cosine similarity, akin to query-key interactions in attention mechanisms, to evaluate the alignment between vertex and hyperedge adjacency features. This approach allows hyperedges to \"query\" vertices for relevant features, providing a scale-invariant assessment that emphasizes the directionality of embedding vectors. The cosine similarity is subsequently transformed into adjacency membership using a sharpened sigmoid function:", + "bbox": [ + 511, + 669, + 908, + 820 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {A} = \\sigma \\left(\\boldsymbol {\\alpha} \\cdot \\tilde {\\mathbf {X}} _ {\\mathrm {a d j}} ^ {(V)} \\left[ \\tilde {\\mathbf {X}} _ {\\mathrm {a d j}} ^ {(E)} \\right] ^ {T}\\right), \\tilde {\\mathbf {X}} _ {\\mathrm {a d j}} ^ {(*)} = \\frac {\\mathbf {X} _ {\\mathrm {a d j}} ^ {(*)}}{| | \\mathbf {X} _ {\\mathrm {a d j}} ^ {(*)} | | _ {2} + \\epsilon} (1)\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 828, + 906, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $\\sigma$ denotes the sigmoid function and $\\alpha = 4$ is a sharpening factor, which pushes values away from zero to estab", + "bbox": [ + 511, + 869, + 908, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9754", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "lish binary-like membership values in matrix $\\mathbf{A}$ . This soft adjacency matrix is further thresholded to create the hard adjacency matrix $\\hat{\\mathbf{A}} = [\\mathbf{A} > 0.5]$ , which provides binary memberships to facilitate sparse attention masking.", + "bbox": [ + 89, + 90, + 483, + 154 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Architecture Scaling", + "text_level": 1, + "bbox": [ + 89, + 162, + 285, + 180 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/541437b8911bbcefb6643f69ba946f1d43e2a7c01dbe1095ac3346e7b814cfd6.jpg", + "table_caption": [ + "Table 1. Scaling variants of our HgVT architecture. All models are trained at $224\\mathrm{x}224$ resolution, except lite variant (HgVT-Lt), trained at $160\\mathrm{x}160$ . Showing count for vertices $(i\\mathcal{V},v\\mathcal{V})$ , hyperedges $(p\\mathcal{E}, v\\mathcal{E})$ , dim for adj. $(d_a)$ and features $(d_f)$ , depth $(L)$ , and heads $(h)$ ." + ], + "table_footnote": [], + "table_body": "
Model|iV||vV||pE||vE|df+daLhParams FLOPS
HgVT-Lt10012326128+641246.8M0.92B
HgVT-Mi19616508128+64945.8M1.39B
HgVT-Ti19616508128+641247.7M1.80B
HgVT-S19616508224+9614723M5.48B
", + "bbox": [ + 93, + 246, + 483, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "HgVT adopts a hybrid scaling strategy inspired by DeiT [53] and ViG [16], with Tab. 1 detailing the transformer scaling hyperparameters and the vertex and edge type allocations, including fixed capacities for non-image vertices $(i\\mathcal{V})$ as proposed by ViHGNN [17]. Our scaling approach aims to balance parameter count and FLOPs with the two main GNN baselines, ViG and ViHGNN. We also introduce a Mini variant (HgVT-Mi) to match DeiT-Ti's computational profile, which we achieve by repeating the final block twice for an effective depth of 9; see Appendix H for details. Finally, we include a Ti-Lite variant (HgVT-Lt) for more efficient ablation studies under a constrained training budget.", + "bbox": [ + 89, + 328, + 485, + 510 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Enforcing Semantic Structure", + "text_level": 1, + "bbox": [ + 89, + 523, + 366, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Feature matrices for virtual vertices and hyperedges, lacking direct input-based initialization, risk converging to homogeneous solutions and collapsed representations. Additionally, the dynamic adjacency calculation alone fails to naturally promote semantic grouping, in contrast to clustering-based approaches commonly used in vision GNNs. To address these issues, we introduce diversity regularization, which enforces orthogonal embeddings, and population regularization, which encourages a structured, sparse hypergraph. For enhanced semantic differentiation of virtual hyperedge features, we further incorporate an expert-based pooling strategy as a more robust alternative to mean pooling.", + "bbox": [ + 89, + 549, + 485, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Diversity-Driven Feature Differentiation", + "text_level": 1, + "bbox": [ + 89, + 742, + 437, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To prevent homogenization of learned feature matrices and to encourage distinct, semantically rich embeddings, we implement a diversity-driven regularization approach. This method, designed to maintain maximum orthogonality among the embeddings of virtual vertices and hyperedges, penalizes the absolute value of the cosine similarity between different feature vectors, aiming for values close to zero. By using normalized embeddings and masking off diagonal elements to preserve self-similarity, the approach prevents", + "bbox": [ + 89, + 763, + 485, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the model from converging to homogeneous solutions or driving individual vectors towards zero magnitude. We then individually penalize $v \\mathcal{V}$ , $\\mathcal{E}$ , and their adjacency features.", + "bbox": [ + 511, + 90, + 906, + 137 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {D} _ {\\mathrm {L}} (\\mathbf {X}) = \\frac {1}{2} \\sum_ {i j} (1 - \\delta_ {i j}) \\cdot \\left| \\tilde {\\mathbf {X}} \\tilde {\\mathbf {X}} ^ {T} \\right| _ {i j}, \\tilde {\\mathbf {X}} = \\frac {\\mathbf {X}}{| | \\mathbf {X} | | _ {2} + \\epsilon} \\quad (2)\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 143, + 901, + 179 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {D I V}} = \\sum_ {x} \\mathrm {D} _ {\\mathrm {L}} (x), x \\in \\{\\mathbf {X} ^ {(: v V)}, \\mathbf {X} _ {\\mathrm {a d j}} ^ {(: v V)}, \\mathbf {X} ^ {(E)}, \\mathbf {X} _ {\\mathrm {a d j}} ^ {(E)} \\} (3)\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 181, + 901, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Where $(:vV)$ represents the subset of $\\mathcal{V}$ containing only the virtual nodes, $\\delta_{ij}$ is the Kronecker delta function, ensuring that self-similarity is not penalized, and $|\\cdot|_{ij}$ denotes the element-wise absolute value, applied to calculate the penalty for non-orthogonal relationships between embeddings.", + "bbox": [ + 511, + 218, + 906, + 295 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Population Regularization: Learned Sparsity", + "text_level": 1, + "bbox": [ + 511, + 301, + 897, + 318 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Unlike clustering methods such as KNN or Fuzzy C-Means, which impose fixed cluster sizes, our model's dynamic adjacency calculation enables flexible and self-adjusting hyperedge populations. To prevent the associated risks of overly sparse or densely connected hypergraphs, we introduce population regularization. This method applies penalties based on the computed soft membership density of each hyperedge derived from the soft adjacency matrix $\\mathbf{A}$ , ensuring each maintains a vertex population within appropriate bounds to avoid overgeneralization and preserve hypergraph integrity.", + "bbox": [ + 511, + 323, + 908, + 474 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP _ {j} = 2 \\cdot \\sum_ {i} \\max (A _ {i j} - 0. 5, 0) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 481, + 906, + 513 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {P O P}} = \\sum_ {j} \\max (P _ {j} - \\beta , 0) + \\max (\\gamma - P _ {j}, 0) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 515, + 906, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $P_{j}$ represents a soft density estimate of vertex connections for the $j^{th}$ hyperedge, only considering non-zero entries of $\\hat{\\mathbf{A}}$ . The hyperparameters $\\beta$ and $\\gamma$ set the upper and lower density limits, ensuring that hyperedges maintain an optimal balance of connections. Penalties are applied if $P_{j}$ either exceeds $\\beta$ or falls below $\\gamma$ , maintaining the desired sparsity and ensuring the structural efficacy of the hypergraph.", + "bbox": [ + 511, + 549, + 908, + 655 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.3. Expert Pooling for Semantic Specialization", + "text_level": 1, + "bbox": [ + 511, + 662, + 879, + 679 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To effectively combine features from multiple virtual hyperedges for classification, our approach utilizes a method akin to expert-choice, where each virtual hyperedge acts as an \"expert\" generating a confidence score. Unlike mean pooling, which risks collapsing distinct features into an average that may dilute individual contributions, this strategy encourages virtual hyperedges to develop unique, semantically meaningful representations. The normalized confidence scores, $P(e)$ , determine the relevance of each hyperedge $e$ 's contribution to the classification task.", + "bbox": [ + 511, + 684, + 908, + 835 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP (e) = \\operatorname {s o f t m a x} \\left(\\mathbf {X} _ {e} ^ {(v E)} \\mathbf {W} _ {e} + b _ {e}\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 840, + 906, + 867 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $(:vE)$ denotes the subset of $(E)$ containing only the virtual hyperedges, and the softmax is computed across the", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9755", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "expert gate set $\\{e\\}$ after projection by $\\mathbf{W} \\in \\mathbb{R}^{d \\times |vE|}$ . During training, $P(e)$ guides the weighted averaging of hyperedge features before subsequent class prediction. We further apply a density loss function [3, 14] to prevent underutilization of any single virtual hyperedge, complemented by a cross-entropy term with label smoothing to increase expert confidence. During inference, top-k routing combines the most relevant hyperedges based on their confidence scores.", + "bbox": [ + 89, + 90, + 483, + 212 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Empirical Evaluation and Performance", + "text_level": 1, + "bbox": [ + 89, + 226, + 444, + 242 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This section evaluates the Hypergraph Vision Transformer using two model configurations from Tab. 1: HgVT-Ti-Lite for targeted ablation studies and scaled variants for benchmarking against comparable classifiers. We apply standard augmentations from DeiT [53] using the Timm library [58], including RandAugment [7], Mixup [64], Cutmix [60], Random Erasing [66], and Repeated Augment [21].", + "bbox": [ + 89, + 252, + 483, + 357 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. For classification in computer vision, we follow standard practices and use the ImageNet-1k dataset [10] at a resolution of $224 \\times 224$ pixels for scaled model evaluations. For ablation studies, we employ ImageNet-100 [51], a 100-class subset of ImageNet-1k with images scaled to $160 \\times 160$ pixels. This selection provides a computationally manageable dataset while maintaining sufficient class variation and larger image sizes compared to datasets like CIFAR-100 (32x32 pixels)[27]. Nevertheless, we find CIFAR-100 useful for assessing the effects of regularization on the hypergraph structure, as detailed in Appendix H.", + "bbox": [ + 89, + 366, + 483, + 532 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training Hyperparameters. Consistent with DeiT, we use the AdamW optimizer with a weight decay of 0.05. Training is conducted on the ImageNet-1k dataset with a batch size of 1024 for 300 epochs following DeiT. For ablations, we train on ImageNet-100 with a batch size of 512 for with a shorter duration of 200 epochs as proposed by [31]. Learning rates follow a cosine-annealing schedule peaking at 1e-3 for both datasets following scaling from DeiT. Furthermore, we omit the use of Exponential Moving Average (EMA) due to its minimal performance improvement (0.1% in DeiT) relative to its overhead per training step. All models were trained using PyTorch on NVIDIA A6000 GPUs; see Appendix I.", + "bbox": [ + 89, + 541, + 483, + 723 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. Following standard protocols, we measure the Top-1 class prediction accuracy to assess overall performance. Additionally, we take advantage of the learned graph structure (extracted from the last layer) on each image, and measure: Hyperedge Entropy (HE), Intra-Cluster Similarity (ICS), Inter-Cluster Distance (ICD), and Silhouette Score (SIL) [45]; further details on structure in Appendix C.", + "bbox": [ + 89, + 732, + 483, + 838 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.1. Evaluation on ImageNet", + "text_level": 1, + "bbox": [ + 89, + 848, + 313, + 863 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tab. 2 presents the ImageNet-1k Top-1 accuracy of HgVT, benchmarked against isotropic models. We limit evaluation", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c6aff34e8e1d152651b42f65de02fb5f4f8383d67f4fe5793bfd522831bd51da.jpg", + "table_caption": [ + "Table 2. ImageNet-1k results for HgVT and other isotropic networks. $\\star$ CNN, $\\clubsuit$ Transformer, $\\star$ GNN, $\\blacksquare$ HGNN, and $\\triangle$ HgVT." + ], + "table_footnote": [], + "table_body": "
ModelParamsFLOPsImNet Top-1ReAL Top-1V2 Top-1
*ResMLP-S12 conv3x3 [52]16.7M3.2B77.084.065.5
*ConvMixer-768/32 [55]21.1M20.9B80.2--
*ConvMixer-1536/20 [55]51.6M51.1B81.4--
*DINOv1-S [2]21.7M4.6B77.0--
*ViT-B/16 [12]86.4M55.5B77.983.6-
*DeiT-Ti [53]5.7M1.3B72.280.160.4
*DeiT-S [53]22.1M4.6B79.885.768.5
*DeiT-B [53]86.4M17.6B81.886.771.5
*ViG-Ti [16]7.1M1.3B73.9--
*ViG-S [16]22.7M4.5B80.4--
*ViG-B [16]86.8M17.7B82.3--
■ViHGNN-Ti [17]8.2M1.8B74.3--
■ViHGNN-S [17]23.2M5.6B81.5--
■ViHGNN-B [17]88.1M19.4B82.9--
▲HgVT-Mi (ours)5.8M1.4B74.481.862.7
▲HgVT-Ti (ours)7.7M1.8B76.283.264.3
▲HgVT-S (ours)22.9M5.5B81.286.770.1
", + "bbox": [ + 517, + 123, + 903, + 349 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "to isotropic models due to the complexity of down-sampling virtual tokens, lacking direct spatial correspondence, needed for pyramidal structures. As a result, we also exclude other pyramidal models from our comparison as they typically outperform isotropic models within the same architecture family [16, 17], making the comparison less meaningful.", + "bbox": [ + 511, + 366, + 903, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "HgVT-Ti demonstrates a clear advantage, surpassing ViHGNN-Ti by $1.9\\%$ in accuracy, with the smaller HgVT-Mi model matching ViHGNN-Ti with only $70\\%$ of the parameters. At the S-scale, we reduce HgVT's depth relative to ViHGNN to accommodate hyperedge attention and adjacency feature overhead, aiming for comparable parameter counts; as a result, HgVT-S only matches ViHGNN-S. Despite this limitation, HgVT-S matches DeiT-B's accuracy on ImageNet Real [1] and achieves competitive performance on ImageNet V2 [43], at nearly a quarter of DeiT-B's size.", + "bbox": [ + 511, + 458, + 906, + 608 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Overall, HgVT achieves strong performance on classification, demonstrating the benefits of integrating hypergraph structures within a vision transformer framework. Additionally, computational evaluation in Appendix B shows a consistent $\\geq 1.4\\times$ speedup over ViG, with adjacency computation exhibiting a $10\\times$ improvement over ViG's KNN-based method. These improvements further highlight HgVT's potential for greater efficiency over prior Vision GNNs.", + "bbox": [ + 511, + 609, + 908, + 731 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.2. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 742, + 678, + 756 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conducted a series of ablations on the ImageNet-100 dataset using the HgVT-Lt model, reporting Top-1 classification accuracy alongside mean hyperedge entropy and silhouette scores to assess the quality of the hypergraph's structure. Notably, we observe a weak anti-correlation between the graph quality metrics and Top-1 accuracy (see Appendix H), indicating opposing objectives. Overall, the ablations are grouped into three categories: regularization, architecture, and pooling methods, with results in Tab. 3.", + "bbox": [ + 511, + 763, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9756", + "bbox": [ + 480, + 945, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/090833d322eedcf24fe1b323ca019d45066e3d0c43472d52cfd14b233de884e5.jpg", + "table_caption": [ + "Table 3. ImageNet-100 ablations with HgVT-Lt. Indicating used (✓) or not used $(\\mathcal{X})$ , and pooling methods: Average (A), Image $(I)$ , Expert $(E)$ , Expert $^+$ Image $(EI)$ , and $EI$ dropping I $(DI)$" + ], + "table_footnote": [], + "table_body": "
Ablation on ↓CLS Dropout Stoch. Decay Diversity PopulationTied FFN Xadj = X dF Mult. Pooled OpEdge EntropyTop-1 Acc.Params (M)
X✓✓✓✓1EI3.320.78084.36
none: HgVT-Lt✓✓✓✓✓1E3.130.75182.23
RegularizationXX✓✓✓1E3.120.74581.89
X✓X✓✓1E1.990.72380.79
X✓✓X✓1E3.890.63981.79
X✓XX✓1E3.580.61081.99
ArchitectureX✓✓✓X1E3.090.74182.89
X✓✓✓X1E2.270.80878.62
X✓✓✓✓1E2.120.78076.95
X✓✓✓✓1.5E2.050.77077.46
PoolingX✓✓✓✓1A3.070.74782.06
X✓✓✓✓1I2.930.76084.08
X✓✓✓✓1E3.130.74982.52
X✓✓✓✓1DI3.320.78080.94
", + "bbox": [ + 94, + 135, + 483, + 359 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Tab. 3, the regularization ablations demonstrate that stochastic path dropout decay [22] improves both Top-1 accuracy and silhouette scores, consistent with ViG and ViHGNN [16, 17]. Omitting Class dropout also boosts accuracy, aligning with DeiT [53]. Furthermore, our proposed diversity and population regularization are essential for preserving graph structure; removing diversity leads to partial representation collapse, while removing population regularization results in near-zero sparsity, effectively turning HgVT into a ViT with increased network complexity.", + "bbox": [ + 88, + 376, + 482, + 527 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the architecture ablations, untying the FFN improves accuracy but significantly increases parameter count, making it an unfavorable tradeoff. Tying adjacency and embedding features $(\\mathbf{X}_{\\mathrm{adj}} = \\mathbf{X})$ reduces parameters and FLOPs but degrades performance, and while untying the FFN or increasing feature dimensionality partially mitigates this, the parameter increase remains suboptimal. This indicates that adjacency and embedding features $(\\mathbf{X}_{\\mathrm{adj}}^{(*)}$ and $\\mathbf{X}^{(*)})$ are similar, yet require dedicated feature spaces to avoid crowding.", + "bbox": [ + 88, + 527, + 482, + 665 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For pooling methods, expert edge pooling outperforms average edge pooling in accuracy, while image pooling achieves the highest accuracy at the cost of degraded graph structure. Combining image and expert pooling recovers lost structure and improves accuracy, with each input focusing on different semantic levels (see Appendix D). Additionally, dropping the pooled image embedding prior to the classifier head maintains moderate performance, indicating that both paths meaningfully contribute to the final prediction.", + "bbox": [ + 88, + 666, + 482, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of Vertex Self-Attention and Patch Embedding. We evaluated the impact of patch embedding methods and vertex self-attention, comparing a convolutional stem (Conv2D-BN-GELU layers [16, 17]) and a simpler patch projection (pixel-shuffled patches with affine projection [2, 12]), across various pooling strategies (average, image, and ex", + "bbox": [ + 88, + 809, + 485, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/13466a7ead897590b93de4e0ea3fbe8f8fb611f86775139cb295dfc67aecd952.jpg", + "table_caption": [ + "Table 4. Top-1 accuracy of HgVT-Lt on ImageNet-100 with $(\\checkmark)$ or without $(X)$ vertex self-attention. Contrasting pooling operations and patch embedding versions: Conv. Stem or Patch Projection." + ], + "table_footnote": [], + "table_body": "
Pooling Op. →AverageImageExpert
Vertex SA →X✓X✓X✓
Conv. Stem78.0282.0682.0584.0878.8782.52
Patch Project64.3072.7670.7676.1762.6271.43
", + "bbox": [ + 529, + 136, + 887, + 200 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "pert), as shown in Tab. 4. The patch projection consistently underperforms the convolutional stem, likely due to the model's small size limiting its effectiveness. Omitting vertex self-attention leads to further degradation, especially without the convolutional stem, suggesting it is crucial for effectively separating features in low-dimensional space. PCA shows that 71/128 channels are needed to explain $95\\%$ of the variance for the convolutional stem, compared to 19/128 for the patch projection, indicating the richer representation captured by the convolutional stem. Notably, image pooling shows the least degradation, likely due to its more direct gradient flow compared to the edge pooling methods.", + "bbox": [ + 511, + 213, + 906, + 396 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.3. Pooling Methods and Graph Structure", + "text_level": 1, + "bbox": [ + 511, + 405, + 846, + 421 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3fc8047445d3c45ca53af5bc68f612b0605d6680d913f7e6c44a4d1f94fb7477.jpg", + "table_caption": [ + "Table 5. Impact of graph structure on pooling operation for HgVT-Lt on ImageNet-100. Measuring graph metrics for image vertices $(i\\mathcal{V})$ and all vertices $(\\mathcal{V})$ , along with features from DINOv2." + ], + "table_footnote": [], + "table_body": "
Pooling Op. →ImageExpertImg. & Expert
Feature Model ↓HEICSICDHEICSICDHEICSICD
HgVT-Lt (iV)3.030.500.313.240.430.373.230.420.36
HgVT-Lt (V)3.200.250.723.390.360.453.390.280.58
DINO2-S (iV)3.040.840.083.250.850.053.250.830.06
DINO2-G (iV)3.040.690.153.250.680.123.250.680.13
", + "bbox": [ + 517, + 474, + 906, + 569 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To evaluate the impact of pooling methods on graph structure, we measured HE, ICS, and ICD using the ImageNet-100 validation set with three strategies: image pooling, expert pooling, and a combined image + expert pooling approach. Metrics used either the image vertex subset $(i\\mathcal{V})$ or the full vertex set $(\\mathcal{V})$ , with features derived from DINOv2 (S and G) [38] and HgVT-Lt using HgVT-Lt's adjacency matrix (A). Results in Tab. 5 show that while all methods achieve similar graph quality using $i\\mathcal{V}$ , image pooling slightly improves similarity. However, including all vertices $(\\mathcal{V})$ consistently increases ICD and entropy while reducing ICS, indicating decreased graph coherence. This effect is more severe with image pooling methods, suggesting that virtual vertices $(v\\mathcal{V})$ act as noisy elements, rather than summarization points.", + "bbox": [ + 511, + 583, + 908, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparing DINOv2 models, all pooling methods align more closely with DINOv2-G, where achieving a balance between ICS and ICD is preferable to maximizing either individually. This trend, along with consistent HE, suggests a focus on higher-level detail, irrespective of the smaller HgVT model size or pooling method. Image pooling shows slightly stronger alignment with both DINO models, indicating that", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9757", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "both high- and low-level semantics are encoded within a single feature space, unlike methods that can use edge channels for high-level concepts. Notably, all expert pooling methods exhibit an emergent macro-class prediction behavior, where each virtual edge $(v\\mathcal{E})$ consistently captures broader taxonomic groups (e.g., dogs, birds). Further representation and macro-class analysis are provided in Appendices D and J.", + "bbox": [ + 89, + 90, + 485, + 196 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.4. Performance on Image Retrieval", + "text_level": 1, + "bbox": [ + 89, + 209, + 377, + 224 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To evaluate the capability of HgVT in capturing semantic structures, we perform image retrieval experiments comparing four methods: pooling similarity (PS), volumetric similarity (VS), adaptive pooling similarity (APS), and adaptive volumetric similarity (AVS). PS ranks the pooled embeddings by cosine similarity (vector search), while the other methods enhance retrieval by leveraging graph structure. Volumetric similarity calculates ellipsoid overlap using an approximate Mahalanobis distance, with pruned hyperedges defining the distribution spread around the pooled embedding (centroid). Adaptive methods further refine retrieval via a graph similarity measure on the pruned hyperedges, re-ranking from a shortlist of $R = 100$ . Computational efficiency in adaptive retrieval is ensured through centroid hash binning and limiting comparisons to the top- $C = 4$ most significant query hyperedges, resulting in a complexity of $\\mathcal{O}(RC)$ . Notably, we prune to 12 hyperedges and use 10 centroid bins. Additional details provided in Appendix G.", + "bbox": [ + 91, + 232, + 483, + 503 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b034752c368d19342c826677dbbb0a5a6595986a1312976eb4a42a9fa35abcc8.jpg", + "table_caption": [ + "Table 6. Image Retrieval on ImageNet-1k with proposed search methods: PS, APS, VS, and AVS. Reporting Top-1 accuracy, mAP@10 (\\%), and 1-NN-hit@10 (\\%) for pooling methods." + ], + "table_footnote": [], + "table_body": "
mAP@10 (%)
ModelTop-1PoolingPSAPSVSAVS
MRL-128 [29]70.52Image64.9465.20--
MRL-256 [29]70.62Image65.0465.20--
HgVT-Ti (ours)76.18Expert70.5669.5970.5369.40
HgVT-Ti (ours)76.18Im&Ex73.2369.5973.0869.49
1-NN-hit@10 (%)
HgVT-Ti (ours)76.18Expert21.2219.1021.2519.56
HgVT-Ti (ours)76.18Im&Ex25.1319.1025.2219.17
", + "bbox": [ + 94, + 559, + 480, + 703 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ImageNet Retrieval. We evaluate retrieval performance on the ImageNet-1K dataset to assess HgVT's ability to capture semantic relationships and compare four retrieval methods: PS, VS, APS, and AVS. The primary metric is mAP@10, with MRL [29] serving as the baseline due to its adaptive re-ranking approach. We also report the 1-NN-CLIP-L hitrate@10, which measures how often the top-1 result ranked by CLIP-L [42] appears in the top-10 retrieved results, offering additional insight into semantic alignment. Results in Tab. 6 show that HgVT-Ti surpasses MRL by over $8\\%$ in retrieval performance, despite being significantly smaller than MRL (ResNet-50) and using a comparably compact", + "bbox": [ + 89, + 719, + 485, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "embedding size ( $d = 2 \\times 128$ ). Among our methods, PS and VS achieve similar results, while APS and AVS underperform, likely due to their focus on exact-feature similarity and ambiguous-class features, which limits alignment with the high-level semantics required by ImageNet's diverse dataset.", + "bbox": [ + 511, + 90, + 906, + 167 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2a2b4870b4a188ee894bf0f2605c6877aa4259b988ac8995dd236b7f7240eb01.jpg", + "table_caption": [ + "Table 7. Image Retrieval on revisited Oxford and Paris; reporting mAP $(\\%)$ . Showing training set and method used. *mAP@100." + ], + "table_footnote": [], + "table_body": "
Dataset →ROxfordRParis
ModelTrain SetMethodMHMH
ALEX+GeM [44]ImNet-1kPS33.810.452.726.0
RN101+R-MAC [44]ImNet-1kPS49.818.574.052.1
DINOv1-S/16 [2]ImNet-1kPS41.813.763.134.4
DINOv1-S/16 [2]GLDv2PS51.524.375.351.6
HgVT-Ti (ours)ImNet-1kVS26.810.555.428.7
VS*25.89.065.628.0
AVS*27.06.864.126.7
HgVT-S (ours)ImNet-1kVS28.012.156.731.1
VS*26.310.265.028.4
AVS*27.410.365.027.1
", + "bbox": [ + 517, + 202, + 903, + 367 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Oxford and Paris Retrieval. To evaluate image retrieval performance beyond simple class retrieval, we use the Oxford and Paris Revisited datasets [39, 41], which provide three splits of increasing difficulty (Easy, Medium, and Hard) for query/database pairs. We report Mean Average Precision (mAP) for the Medium (M) and Hard (H) splits, using mAP@100 for AVS based on short-list ranking, while full mAP and mAP@100 are provided for VS as a baseline comparison. Results, shown in Tab. 7, indicate that HgVT achieves competitive performance with similarly sized feature extractors, though performance on ROxford-M lags. This shortfall may stem from subtle landmark differences better captured by multi-scale Conv-Nets and self-supervised learning, compared to the more salient focus driven by HgVT's classifier training. However, AVS outperforms VS on ROxford-M, demonstrating its ability to uncover finer feature similarities within the hypergraph structure.", + "bbox": [ + 511, + 378, + 908, + 636 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion and Future Directions", + "text_level": 1, + "bbox": [ + 511, + 648, + 821, + 664 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we introduced the Hypergraph Vision Transformer (HgVT), which combines hypergraph structures with vision transformers. Our methods, including diversity and population regularization, along with expert edge pooling, enhance semantic representation and efficiency by enabling dynamic hypergraph construction. Although challenging to tune, these methods avoid reliance on traditional clustering techniques, reflecting a tradeoff between complexity and efficiency. Empirical results show that HgVT achieves strong performance on both image classification and retrieval, positioning HgVT as a competitive framework for semantic vision tasks. Future work will focus on exploring the scalability of hypergraph structures and integrating self-supervised learning to further improve adaptability and better decouple saliency from semantic vision graph generation.", + "bbox": [ + 511, + 674, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9758", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Lucas Beyer, Olivier J. Henaff, Alexander Kolesnikov, Xiaohua Zhai, and Aaron van den Oord. Are we done with imageNet? arXiv preprint arXiv:2006.07159, abs/2006.07159, 2020. 6", + "[2] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jegou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9630-9640, 2021. 1, 2, 6, 7, 8, 4, 5", + "[3] Tianlong Chen, Zhenyu Zhang, Ajay Kumar Jaiswal, Shiwei Liu, and Zhangyang Wang. Sparse moe as the new dropout: Scaling dense and self-slimmable transformers. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. 6", + "[4] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 15", + "[5] Norman Cliff. Dominance statistics: Ordinal analyses to answer ordinal questions. Psychological Bulletin, 114:494-509, 1993. 8", + "[6] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 14", + "[7] Ekin Dogus Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V. Le. Randaugment: Practical automated data augmentation with a reduced search space. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 3008-3017, 2019. 6", + "[8] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. In *The Twelfth International Conference on Learning Representations*, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024. 3", + "[9] Dmitry Demidov, M.H. Sharif, Aliakbar Abdurahimov, Hisham Cholakkal, and Fahad Shahbaz Khan. Salient mask-guided vision transformer for fine-grained classification. In VISIGRAPP, 2023. 1", + "[10] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. In CVPR09, 2009. 6", + "[11] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In North American Chapter of the Association for Computational Linguistics, 2019. 1", + "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at" + ], + "bbox": [ + 93, + 114, + 485, + 900 + ], + "page_idx": 8 + }, + { + "type": "ref_text", + "text": "scale. In International Conference on Learning Representations, 2021. 1, 2, 6, 7, 4, 5", + "bbox": [ + 547, + 93, + 906, + 119 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The PASCAL Visual Object Classes Challenge 2012 (VOC2012) Results. http://www.pascalnetwork.org/challenges/VOC/voc2012/workshop/index.html. 14", + "[14] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. J. Mach. Learn. Res., 23:120:1-120:39, 2022. 6", + "[15] Yue Gao, Meng Wang, Dacheng Tao, Rongrong Ji, and Qionghai Dai. 3-d object retrieval and recognition with hypergraph analysis. IEEE Transactions on Image Processing, 21(9): 4290-4303, 2012. 2", + "[16] Kai Han, Yunhe Wang, Jianyuan Guo, Yehui Tang, and Enhua Wu. Vision GNN: An image is worth graph of nodes. In Advances in Neural Information Processing Systems, 2022. 1, 2, 5, 6, 7, 4", + "[17] Yan Han, Peihao Wang, Souvik Kundu, Ying Ding, and Zhangyang Wang. Vision hgnn: An image is more than a graph of nodes. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 19878-19888, 2023. 1, 2, 3, 4, 5, 6, 7, 27", + "[18] Yunusa Haruna, Shiyin Qin, Abdulrahman Hamman Adama Chukkol, Abdulganiyu Abdu Yusuf, Isah Bello, and Adamu Lawan. Exploring the synergies of hybrid cnns and vits architectures for computer vision: A survey. arXiv preprint arXiv:2402.02941, abs/2402.02941, 2024. 1", + "[19] Ali Hassani, Steven Walton, Jiachen Li, Shen Li, and Humphrey Shi. Neighborhood attention transformer. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2", + "[20] Kaiming He, X. Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2015. 1, 15", + "[21] Elad Hoffer, Tal Ben-Nun, Itay Hubara, Niv Giladi, Torsten Hoefler, and Daniel Soudry. Augment your batch: Improving generalization through instance repetition. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8126-8135, 2020. 6", + "[22] Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Weinberger. Deep networks with stochastic depth, 2016. 7", + "[23] Yuchi Huang, Qingshan Liu, and Dimitris Metaxas. ]video object segmentation by hypergraph cut. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 1738-1745, 2009. 2", + "[24] Yuchi Huang, Qingshan Liu, Shaoting Zhang, and Dimitris N. Metaxas. Image retrieval via probabilistic hypergraph ranking. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 3376-3383, 2010. 2", + "[25] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. 15" + ], + "bbox": [ + 93, + 114, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9759", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3992-4003, 2023. 1", + "[27] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, 2009. 6", + "[28] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2012. 1", + "[29] Aditya Kusupati, Gantavya Bhatt, Aniket Rege, Matthew Wallingford, Aditya Sinha, Vivek Ramanujan, William Howard-Snyder, Kaifeng Chen, Sham M. Kakade, Prateek Jain, and Ali Farhadi. Matryoshka representation learning. In Neural Information Processing Systems, 2022. 1, 8", + "[30] Y. Lecun, L. Bottou, Y. Bengio, and P. Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 1", + "[31] Sangjun Lee, Inwoo Hwang, Gi-Cheon Kang, and Byoung-Tak Zhang. Improving robustness to texture bias via shape-focused augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 4323-4331, 2022. 6", + "[32] Benjamin Lefaudeau, Francisco Massa, Diana Liskovich, Wenhan Xiong, Vittorio Caggiano, Sean Naren, Min Xu, Jieru Hu, Marta Tintore, Susan Zhang, Patrick Labatut, Daniel Haziza, Luca Wehrstedt, Jeremy Reizenstein, and Grigory Sizov. xformers: A modular and hackable transformer modelling library. https://github.com/facebookresearch/xformers, 2022. 28", + "[33] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9992-10002, 2021. 1, 2, 15", + "[34] Leland McInnes, John Healy, Nathaniel Saul, and Lukas Großberger. Umap: Uniform manifold approximation and projection. Journal of Open Source Software, 3(29):861, 2018. 9", + "[35] Mustafa Munir, William Avery, and Radu Marculescu. Mobilevig: Graph-based sparse attention for mobile vision applications. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 2211-2219, 2023. 1, 2", + "[36] Mustafa Munir, William Avery, Md Mostafijur Rahman, and Radu Marculescu. Greedyvig: Dynamic axial graph construction for efficient vision gnns. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6118-6127, 2024. 1, 2", + "[37] NVIDIA. Apex: A pytorch extension. https://github.com/NVIDIA/apex, 2020.28", + "[38] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell" + ], + "bbox": [ + 91, + 90, + 485, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINOv2: Learning robust visual features without supervision. Transactions on Machine Learning Research, 2024. 1, 2, 7, 6, 8, 14, 15", + "[39] James Philbin, Ondrej Chum, Michael Isard, Josef Sivic, and Andrew Zisserman. Lost in quantization: Improving particular object retrieval in large scale image databases. In 2008 IEEE Conference on Computer Vision and Pattern Recognition, pages 1-8, 2008. 8", + "[40] Joan Puigcerver, Carlos Riquelme Ruiz, Basil Mustafa, and Neil Houlsby. From sparse to soft mixtures of experts. In The Twelfth International Conference on Learning Representations, 2024. 1", + "[41] Filip Radenovic, Ahmet Iscen, Giorgos Tolias, Yannis Avrithis, and Ondrej Chum. Revisiting oxford and paris: Large-scale image retrieval benchmarking. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5706-5715, 2018. 8", + "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 8, 6", + "[43] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do imagenet classifiers generalize to imageNet? In International Conference on Machine Learning, 2019. 6", + "[44] Jérôme Revaud, Jon Almazán, Rafael Sampaio de Rezende, and César Roberto de Souza. Learning with average precision: Training image retrieval with a listwise loss. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 5106-5115, 2019. 8", + "[45] Peter J. Rousseeuw. Silhouettes: A graphical aid to the interpretation and validation of cluster analysis. Journal of Computational and Applied Mathematics, 20:53-65, 1987. 6, 7", + "[46] Franco Scarselli, Marco Gori, Ah Chung Tsoi, Markus Hagenbuchner, and Gabriele Monfardini. The graph neural network model. IEEE Transactions on Neural Networks, 20(1):61-80, 2009. 2", + "[47] Noam Shazeer. GLU variants improve transformer. arXiv preprint arXiv:2002.05202, abs/2002.05202, 2020. 4", + "[48] Dai Shi. Transnext: Robust foveal visual perception for vision transformers. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17773-17783, 2023. 14, 15", + "[49] Dai Shi. Transnext: Robust foveal visual perception for vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17773-17783, 2024. 2", + "[50] Sakhinana Sagar Srinivas, Rajat Kumar Sarkar, Sreeja Gangasani, and Venkataramana Runkana. Vision HgNN: An electron-micrograph is worth hypergraph of hypernodes. arXiv preprint arXiv:2408.11351, abs/2408.11351, 2024. 2" + ], + "bbox": [ + 516, + 90, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9760", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive multiview coding. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XI 16, pages 776-794. Springer, 2020. 6", + "[52] Hugo Touvron, Piotr Bojanowski, Mathilde Caron, Matthieu Cord, Alaaeldin El-Nouby, Edouard Grave, Armand Joulin, Gabriel Synnaeve, Jakob Verbeek, and Hervé Jégou. Resmlp: Feedforward networks for image classification with data-efficient training. arXiv preprint arXiv:2105.03404, abs/2105.03404, 2021. 6", + "[53] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herve Jegou. Training data-efficient image transformers amp; distillation through attention. In Proceedings of the 38th International Conference on Machine Learning, pages 10347-10357. PMLR, 2021. 1, 2, 5, 6, 7, 4", + "[54] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurélien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, abs/2302.13971, 2023. 1, 2", + "[55] Asher Trockman and J. Zico Kolter. Patches are all you need? Trans. Mach. Learn. Res., 2023, 2023. 6", + "[56] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 1", + "[57] Petar Velicković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. Graph attention networks. In International Conference on Learning Representations, 2018. 4", + "[58] Ross Wightman. Pytorch image models. https://github.com/rwrightman/pytorch-image-models, 2019.6, 28", + "[59] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In European Conference on Computer Vision. Springer, 2018. 15", + "[60] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Young Joon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6022-6031, 2019. 6", + "[61] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 1204-1213. IEEE, 2022. 1", + "[62] Biao Zhang and Rico Sennrich. Root mean square layer normalization. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 2", + "[63] Bowen Zhang, Zhi Tian, Quan Tang, Xiangxiang Chu, Xiaolin Wei, Chunhua Shen, and Yifan liu. Segvit: Semantic segmentation with plain vision transformers. In Advances in Neural Information Processing Systems, 2022. 1" + ], + "bbox": [ + 91, + 92, + 483, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[64] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In International Conference on Learning Representations, 2018. 6", + "[65] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 6230-6239, 2016. 15", + "[66] Zhun Zhong, Liang Zheng, Guoliang Kang, Shaozi Li, and Yi Yang. Random erasing data augmentation. AAAI, 34: 13001-13008, 2020. 6", + "[67] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision, 127:302 - 321, 2016. 14", + "[68] Lei Zhu, Xinjiang Wang, Zhanghan Ke, Wayne Zhang, and Rynson Lau. Bifformer: Vision transformer with bi-level routing attention. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + ], + "bbox": [ + 516, + 92, + 906, + 372 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "9761", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Hypergraph Vision Transformers_ Images are More than Nodes, More than Edges/b3c66b72-bf5c-47e0-a296-e834f5e49c0f_model.json b/2025/Hypergraph Vision Transformers_ Images are More than Nodes, More than Edges/b3c66b72-bf5c-47e0-a296-e834f5e49c0f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6f42199c8e190a04303d907320219aa8eeb23991 --- /dev/null +++ b/2025/Hypergraph Vision Transformers_ Images are More than Nodes, More than Edges/b3c66b72-bf5c-47e0-a296-e834f5e49c0f_model.json @@ -0,0 +1,2466 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.131, + 0.908, + 0.153 + ], + "angle": 0, + "content": "Hypergraph Vision Transformers: Images are More than Nodes, More than Edges" + }, + { + "type": "text", + "bbox": [ + 0.411, + 0.182, + 0.586, + 0.234 + ], + "angle": 0, + "content": "Joshua Fixelle \nUniversity of Virginia \njf9fk@virginia.edu" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.301, + 0.486, + 0.59 + ], + "angle": 0, + "content": "Recent advancements in computer vision have highlighted the scalability of Vision Transformers (ViTs) across various tasks, yet challenges remain in balancing adaptability, computational efficiency, and the ability to model higher-order relationships. Vision Graph Neural Networks (ViGs) offer an alternative by leveraging graph-based methodologies but are hindered by the computational bottlenecks of clustering algorithms used for edge generation. To address these issues, we propose the Hypergraph Vision Transformer (HgVT), which incorporates a hierarchical bipartite hypergraph structure into the vision transformer framework to capture higher-order semantic relationships while maintaining computational efficiency. HgVT leverages population and diversity regularization for dynamic hypergraph construction without clustering, and expert edge pooling to enhance semantic extraction and facilitate graph-based image retrieval. Empirical results demonstrate that HgVT achieves strong performance on image classification and retrieval, positioning it as an efficient framework for semantic-based vision tasks." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.618, + 0.222, + 0.634 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.486, + 0.87 + ], + "angle": 0, + "content": "Computer vision has recently transitioned from the historically dominant Convolutional Neural Networks (CNNs) [20, 28, 30] to the increasingly prominent Vision Transformers (ViTs), which have quickly embedded themselves as the new de facto standard [12, 38]. This shift reflects the broader success of transformers in natural language processing [11, 54, 56] and is driven by the remarkable scalability of ViTs across various tasks such as image classification [53, 61], semantic segmentation [26, 63], and image retrieval [2, 29]. While hybrid models like hierarchical attention and CNN-ViT methods [18, 19, 33] aim to balance computational load and flexibility, ViTs still struggle with focusing on salient features rather than comprehensive image understanding [2, 9, 12, 40]. This highlights the need for methods that improve both efficiency and semantic accuracy." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Within the spectrum of novel architectures, Vision Graph Neural Networks (ViGs) [16] and Vision Hypergraph Neural" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.27, + 0.908, + 0.468 + ], + "angle": 0, + "content": "Networks (ViHGNNs) [17] leverage graph-based topologies to advance image processing. Unlike CNNs, which harness locality and translation-invariance through densely connected pixel grids and repeated convolutions, both ViTs and ViGs represent images as sets of patches. In ViTs, each patch acts as a vertex within a maximally connected graph, creating semantically weak connections through self-attention. ViGs enhance this by using clustering algorithms to detect edge groupings and applying graph convolutions to these clusters, forming meaningful patch relationships. ViHGNNs extend these capabilities by employing hyperedges that capture complex, higher-order relationships, enriching understanding of the images. These methodologies are depicted in Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.471, + 0.907, + 0.517 + ], + "angle": 0, + "content": "While graph-based models like ViG and ViHGNN have introduced notable advancements in visual perception, two critical observations emerge regarding these architectures:" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.52, + 0.909, + 0.594 + ], + "angle": 0, + "content": "1. In existing vision GNN models [16, 17, 35, 36], edge features are primarily used for basic vertex-to-vertex communication and are not integrated across successive layers: a strategy that could enhance cumulative learning and improve classification accuracy." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.596, + 0.909, + 0.716 + ], + "angle": 0, + "content": "2. The computational complexities associated with clustering algorithms used for edge generation, such as KNN in ViG and Fuzzy C-Means in ViHGNN, pose significant computational bottlenecks. Approaches like MobileViG [35] and GreedyViG [36] attempt to mitigate these challenges with static graph structures and adding dynamic masking, but do so by trading adaptability for efficiency, failing to achieve a well-balanced solution." + }, + { + "type": "list", + "bbox": [ + 0.515, + 0.52, + 0.909, + 0.716 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.911, + 0.902 + ], + "angle": 0, + "content": "In response to limitations in existing graph-based models, we propose the Hypergraph Vision Transformer (HgVT), which advances the hypergraph concept with a bipartite representation where hyperedge features and image patches (vertices) are continuously processed. Unlike traditional models that use graph convolutions, HgVT employs structured multihead attention for efficient vertex-hyperedge message passing and incorporates a dynamic querying mechanism that constructs graph structures in \\(\\mathcal{O}(|V| \\cdot E)\\) time complexity, where \\(E < |V|\\). This graph structure is then utilized in attention masking to balance structural adaptability with computational efficiency. Furthermore, HgVT integrates" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.515, + 0.957 + ], + "angle": 0, + "content": "9751" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.207, + 0.092, + 0.291, + 0.146 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.207, + 0.15, + 0.253, + 0.162 + ], + "angle": 0, + "content": "(a) CNN" + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.092, + 0.536, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.338, + 0.154, + 0.376, + 0.166 + ], + "angle": 0, + "content": "(c) ViG" + }, + { + "type": "image", + "bbox": [ + 0.549, + 0.091, + 0.796, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.573, + 0.23, + 0.625, + 0.242 + ], + "angle": 0, + "content": "(e) HgVT" + }, + { + "type": "image", + "bbox": [ + 0.21, + 0.166, + 0.313, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.207, + 0.23, + 0.245, + 0.241 + ], + "angle": 0, + "content": "(b) ViT" + }, + { + "type": "image", + "bbox": [ + 0.324, + 0.166, + 0.543, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.338, + 0.23, + 0.406, + 0.241 + ], + "angle": 0, + "content": "(d) ViHGNN" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.248, + 0.908, + 0.29 + ], + "angle": 0, + "content": "Figure 1. Comparison of Graph Structures for different methods. Showing (a) CNNs, (b) Vision Transformers, (c) ViG with a KNN clustered GNN, (d) ViHGNN with clustered hyperedges, and (e) our proposed HgVT method. Strong group edges shown with solid lines; weak edges with dashed lines. Hyperedges shown with shaded regions; less dominant hyperedges with gray dashed regions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.304, + 0.486, + 0.379 + ], + "angle": 0, + "content": "virtual elements into vertices and hyperedges, forming intermediate links that create a hierarchical semantic structure, as illustrated in Fig. 1e. Attention masking is then used to restrict message passing, and virtual hyperedge features are leveraged for classification. Our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.384, + 0.485, + 0.489 + ], + "angle": 0, + "content": "- We propose the Hypergraph Vision Transformer (HgVT), which integrates a hierarchical bipartite hypergraph structure within a vision transformer framework. Our isotropic HgVT-Ti model achieves a Top-1 accuracy of \\(76.2\\%\\) on the ImageNet-1k classification task, surpassing the prior state-of-the-art by \\(1.9\\%\\), demonstrating the efficacy of hypergraph-based learning within vision transformers." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.49, + 0.484, + 0.549 + ], + "angle": 0, + "content": "- We introduce population and diversity regularization strategies that enable dynamic hypergraph structure construction in HgVT, allowing the model to self-sparseify connections without relying on traditional clustering techniques." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.55, + 0.485, + 0.655 + ], + "angle": 0, + "content": "- We implement expert edge pooling, a pooling approach that selects edges based on learned confidence scores, facilitating efficient representation pruning and graph extraction. This approach demonstrates strong semantic clustering behavior across macro-classes and achieves competitive image retrieval performance compared to other feature extractors, while maintaining a more compact model size." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.384, + 0.485, + 0.655 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.672, + 0.233, + 0.687 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.699, + 0.486, + 0.865 + ], + "angle": 0, + "content": "Vision Transformers. Vision Transformers (ViTs) proposed by [12] and refined by [2, 38, 53] use self-attention to process image patches as sequences, scaling to complex datasets and tasks. Recent ViTs have reintroduced spatial hierarchies by leveraging local attention [19, 33], integrating sparse global summaries [68], and employing biomimetic modeling to focus on key regions within images [49]. However, current models tend to focus on the most salient objects and patch-level similarities, ignoring global structure. HgVT addresses this by introducing bipartite hypergraphs to model higher-order relationships for improved semantic understanding." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Graph-Based Vision Models and Clustering. Graph Neural Networks (GNNs), initially conceptualized by [46], have" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.304, + 0.908, + 0.47 + ], + "angle": 0, + "content": "been applied to vision tasks through Vision Graph Neural Networks (ViGs) [16], which exhibit improved accuracy over ViTs on common vision tasks. ViGs use graph convolutions to model image patch relationships on a graph structure, typically constructed by iterative clustering algorithms such as KNN and Fuzzy C-Means, which introduce computational overhead. Recent methods avoid clustering inefficiencies with static graph structures [35, 36], sacrificing adaptability. HgVT instead introduces a dynamic graph construction method, relying on cosine similarity from learned features to enable efficient, non-iterative, adaptive clustering." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.476, + 0.909, + 0.613 + ], + "angle": 0, + "content": "Hypergraph-Based Methods. While previously used in many computer vision tasks [15, 23, 24], hypergraphs have recently been incorporated into vision GNNs [17, 50], improving their ability to model complex multi-way relationships. However, these methods treat hypergraphs as an intermediate tool rather than producing a hypergraph to represent underlying images, preventing their use in downstream tasks. HgVT instead iteratively refines a hypergraph through subsequent network layers to produce structured representations." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.625, + 0.761, + 0.643 + ], + "angle": 0, + "content": "3. Hierarchical Hypergraphs" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.65, + 0.909, + 0.833 + ], + "angle": 0, + "content": "Graphs and Basic Notations. Graphs are powerful mathematical tools for representing structured information, applicable across diverse disciplines. A graph \\(\\mathcal{G}\\) is defined as a pair \\((\\mathcal{V},\\mathcal{E})\\), where \\(\\mathcal{V} = \\{v_{1},v_{2},\\ldots ,v_{N}\\}\\) is a set of vertices, and \\(\\mathcal{E} = \\{e_{ij}|(v_i,v_j)\\}\\) is a set of directed edges. For undirected graphs, the edges are symmetric, such that \\(e_{ij} = e_{ji}\\). Each edge \\(e_{ij}\\) connects a pair of vertices \\(v_{i}\\) and \\(v_{j}\\) where \\(v_{i},v_{j}\\in \\mathcal{V}\\). The adjacency matrix \\(\\hat{\\mathbf{A}}\\) is a binary matrix \\(\\{0,1\\}^{|V|\\times |V|}\\), representing the presence (1) or absence (0) of an edge between each pair of vertices. Similarly, an edge weight matrix can be defined as \\(\\mathbf{A}\\in \\mathbb{R}^{|V|\\times |V|}\\) to quantify the strength or capacity of these connections." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Graph Convolution Networks (GCNs). Building on this foundation, GCNs utilize vertex feature matrices \\(\\mathbf{X} \\in \\mathbb{R}^{|V| \\times d}\\) to encode vertex properties. Their core mechanism, message passing, updates vertex features through a convolution" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9752" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.198 + ], + "angle": 0, + "content": "with a learned projection matrix \\(\\mathbf{W} \\in \\mathbb{R}^{d \\times d}\\) and a non-linear activation, guided by the adjacency matrix \\(\\hat{\\mathbf{A}}\\), which specifies neighboring vertices. Adjacency features \\(\\mathbf{X}_{\\mathrm{adj}} \\in \\mathbb{R}^{|V| \\times d_a}\\), typically set as \\(\\mathbf{X}_{\\mathrm{adj}} = \\mathbf{X}\\), enable dynamic updates to \\(\\hat{\\mathbf{A}}\\) and the edge weight matrix \\(\\mathbf{A}\\), refining the graph structure through learned interactions. However, GCNs are unable to model multi-vertex relationships as edges in \\(\\mathcal{E}\\) are pairwise." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.208, + 0.466, + 0.226 + ], + "angle": 0, + "content": "3.1. Hypergraphs and Bipartite Representations" + }, + { + "type": "image", + "bbox": [ + 0.125, + 0.236, + 0.257, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.139, + 0.327, + 0.238, + 0.341 + ], + "angle": 0, + "content": "(a) Hypergraph" + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.234, + 0.454, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.308, + 0.327, + 0.423, + 0.342 + ], + "angle": 0, + "content": "(b) Bipartite Form" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.346, + 0.483, + 0.375 + ], + "angle": 0, + "content": "Figure 2. Comparison of (a) hypergraph and (b) equivalent bipartite representation from Fig. 1d, showing five hyperedges \\((e_0 - e_4)\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.383, + 0.484, + 0.563 + ], + "angle": 0, + "content": "To overcome the pairwise limitation inherent in traditional graphs, hypergraphs offer a robust solution by extending the concept of edges to hyperedges, which connect multiple vertices simultaneously. In a hypergraph \\(\\mathcal{H} = (\\mathcal{V},\\mathcal{E})\\), hyperedges \\(e_j\\in \\mathcal{E}\\) each connect a subset of vertices, defined as \\(e_j = \\{v_i\\mid v_i\\in \\mathcal{V}\\) and \\(i\\in I_j\\}\\), where \\(I_{j}\\) is the set of indices for vertices that are included in hyperedge \\(e_j\\). The set \\(I_{j}\\) directly corresponds to the nonzero entries of the \\(j\\)-th column of the incidence matrix \\(\\mathbf{H}\\in \\{0,1\\}^{|\\mathcal{V}|\\times |\\mathcal{E}|}\\), where \\(\\mathbf{H}_{ij} = 1\\) if vertex \\(v_{i}\\) is included in the hyperedge \\(e_j\\). This structure captures complex relationships, making hypergraphs valuable for modeling networked systems and group interactions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.565, + 0.484, + 0.685 + ], + "angle": 0, + "content": "Hypergraphs can alternatively be described using a bipartite representation, where the vertex set \\(\\mathcal{V}\\) and hyperedge set \\(\\mathcal{E}\\) form distinct groups linked by the incidence matrix \\(\\mathbf{H}\\) (refer to Fig. 2). This representation results in a new graph \\(\\mathcal{G}_B = (\\mathcal{V},\\mathcal{E},\\mathcal{E}_B)\\), where \\(\\mathcal{V}\\) represents the original vertices of the hypergraph, and the elements in \\(\\mathcal{E}\\) correspond to hyperedges. The edges in \\(\\mathcal{E}_B\\), denoted as \\(\\epsilon_{ve} = (\\nu_v,\\nu_e)\\) exist if \\(\\mathbf{H}_{ve} = 1\\), with \\(\\nu_{v}\\in \\mathcal{V}\\) and \\(\\nu_{e}\\in \\mathcal{E}\\), linking \\(\\mathcal{V}\\) and \\(\\mathcal{E}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.687, + 0.484, + 0.824 + ], + "angle": 0, + "content": "In the bipartite graph \\(\\mathcal{G}_B\\), the corresponding adjacency matrix can be simplified as \\(\\hat{\\mathbf{A}} = \\mathbf{H}\\) for \\(\\mathcal{E} \\to \\mathcal{V}\\) interactions, and \\(\\hat{\\mathbf{A}} = \\mathbf{H}^T\\) for \\(\\mathcal{V} \\to \\mathcal{E}\\). Drawing on principles similar to those in ViHGNN [17], the edge weight matrix \\(\\mathbf{A}\\) can be interpreted as fuzzy membership weights, enabling graded interactions across GNN layers. Complementing this setup, the feature matrices are split into \\(\\mathbf{X}^{(V)} \\in \\mathbb{R}^{|\\mathcal{V}| \\times d_v}\\) and \\(\\mathbf{X}^{(E)} \\in \\mathbb{R}^{|\\mathcal{E}| \\times d_e}\\), along with their correspond adjacency feature matrices \\(\\mathbf{X}_{\\mathrm{adj}}^{(V)}\\) and \\(\\mathbf{X}_{\\mathrm{adj}}^{(E)}\\), mirroring traditional GNNs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.458, + 0.85 + ], + "angle": 0, + "content": "3.2. Imposing Hierarchical Structure in Images" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "To enhance the capability of hypergraphs in image analysis, we draw inspiration from the register tokens introduced in [8], which act to summarize information that otherwise" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.227 + ], + "angle": 0, + "content": "manifests as noise in areas of low visual significance. Similarly, this work integrates virtual vertices \\((v\\mathcal{V})\\), alongside typical image patch vertices \\((i\\mathcal{V})\\), and introduces virtual hyperedges \\((v\\mathcal{E})\\), alongside primary hyperedges \\((p\\mathcal{E})\\), providing levels of semantic feature aggregation and relational abstraction. Notably, these virtual elements, illustrated in Fig. 1e, do not correspond to specific image patches; instead, they are learned embeddings used for semantic summarization and capturing high-level abstract information." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.228, + 0.909, + 0.38 + ], + "angle": 0, + "content": "Our proposed hypergraph, constructed from image \\(I\\) as \\(\\mathcal{G}_B(I)\\), integrates primary and virtual sets, forming \\(\\mathcal{V} = i\\mathcal{V} \\cup v\\mathcal{V}\\) and \\(\\mathcal{E} = p\\mathcal{E} \\cup v\\mathcal{E}\\), with statically masked communication pathways to enforce a hierarchical structure. Primary hyperedges \\((p\\mathcal{E})\\) interact with all vertices to support unrestricted semantic aggregation, whereas virtual hyperedges \\((v\\mathcal{E})\\), designated for class predictions, connect solely with virtual vertices \\((v\\mathcal{V})\\). These restrictions separate visual and abstract information, thereby producing a graph structure suitable for use in downstream applications." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.388, + 0.83, + 0.405 + ], + "angle": 0, + "content": "4. A Hypergraph Vision Transformer" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.414, + 0.909, + 0.626 + ], + "angle": 0, + "content": "The Hypergraph Vision Transformer (HgVT) adapts the architecture of standard Vision Transformers by incorporating bipartite hypergraph features for enhanced image analysis capabilities. Like Vision Transformers, HgVT begins with a patch embedding layer, followed by an isotropic stack of \\( L \\times \\mathrm{HgVT} \\) blocks, culminating in feature pooling and a classifier head. The bipartite hypergraph is represented by four principal feature matrices \\( -\\mathbf{X}^{(V)} \\), \\( \\mathbf{X}_{\\mathrm{adj}}^{(V)} \\), \\( \\mathbf{X}^{(E)} \\), and \\( \\mathbf{X}_{\\mathrm{adj}}^{(E)} \\) which are updated iteratively and in an interleaved fashion within each block. Each block constructs a new adjacency matrix \\( \\mathbf{A} \\) from \\( \\mathbf{X}_{\\mathrm{adj}}^{(V)} \\) and \\( \\mathbf{X}_{\\mathrm{adj}}^{(E)} \\), enabling flexible adjustments to the hypergraph structure. As illustrated in Fig. 3, this modular process allows for the continuous integration and processing of these matrices within each HgVT block." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.637, + 0.835, + 0.653 + ], + "angle": 0, + "content": "4.1. Hyperedges as Communication Pools" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Each HgVT block processes both vertex and edge information, refining them from the previous block based on a newly constructed graph structure. Initially, adjacency mask computation (detailed in the next section) determines the connectivity for the subsequent processing steps within each block, dynamically adjusting to the updated feature matrices from the previous block. Three attention layers – vertex self-attention, edge aggregate attention, and edge distribution attention – operate sequentially to enhance feature integration and facilitate effective message passing along the hyperedges formed in the adjacency computation step. Finally, separate feed-forward networks process vertex and edge features independently, ensuring specialized treatment for the two distinct sets within the bipartite hypergraph, preserving the unique properties of each set. Operational details of these components are further described in Appendix A." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9753" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.163, + 0.094, + 0.833, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.286, + 0.907, + 0.318 + ], + "angle": 0, + "content": "Figure 3. HgVT Architecture Diagram, composed of stacked HgVT blocks with adjacency matrix \\(\\mathbf{A}\\), vertex features \\(\\mathbf{X}^{(V)}\\), and hyperedge features \\(\\mathbf{X}^{(E)}\\). Edge attention flow is shown with gray arrows; input norms and residual adds are omitted for clarity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.323, + 0.486, + 0.506 + ], + "angle": 0, + "content": "Hypergraph Feature Processing. Within each HgVT block, two distinct point-wise feed-forward networks (FFNs) independently process vertex and hyperedge features, aligning with the bipartite structure of the hypergraph. Each FFN integrates both the element features and their corresponding adjacency features through a fully connected layer, improving the model's ability to synthesize relationships. Processing both feature types within the same FFN layer allows adjacency information to be handled directly, bypassing the need for graph-based message passing and improving computational efficiency. Moreover, parameter overhead can be reduced by optionally tying edge and vertex FFN weights." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.52, + 0.273, + 0.614 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.116, + 0.618, + 0.251, + 0.63 + ], + "angle": 0, + "content": "(a) Communication Pool." + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.516, + 0.48, + 0.616 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.324, + 0.619, + 0.443, + 0.631 + ], + "angle": 0, + "content": "(b) Attention Sparsity." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.637, + 0.483, + 0.679 + ], + "angle": 0, + "content": "Figure 4. (a) Hyperedge Communication Pool Flow with edges \\(\\mathcal{E}\\) and member vertices \\(\\nu_{e}\\); (b) Attention Sparsity (Mean and std) for HgVT-S on the ImageNet-1k Validation set." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.486, + 0.901 + ], + "angle": 0, + "content": "Hyperedges as Communication Pools. Hypergraph GNNs typically employ a gather \\(\\rightarrow\\) scatter mechanism for processing vertex-hyperedge interactions, whereas HgVT reconceptualizes hyperedges as communication pools that facilitate information flow among vertices and their associated hyperedges. Specifically, vertex self-attention manages vertex-to-vertex \\((\\mathcal{V}\\rightarrow \\mathcal{V})\\) interactions within hyperedges, edge aggregate attention orchestrates the flow from vertices to hyperedges \\((\\mathcal{V}\\rightarrow \\mathcal{E})\\), and edge distribution attention handles the reverse, from hyperedges back to vertices \\((\\mathcal{E}\\rightarrow \\mathcal{V})\\). By segmenting the attention operations, HgVT efficiently approximates an all-to-all feature transfer within hyperedges, as illustrated in Fig. 4a, significantly reducing the quadratic complexity associated with full attention mechanisms." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.323, + 0.909, + 0.641 + ], + "angle": 0, + "content": "Sparse and Fuzzy Attention. Building upon the dynamic communication pools concept, HgVT employs both sparse and fuzzy attention mechanisms to further optimize computational efficiency. Vertex self-attention is applied selectively to pairs of vertices connected by common hyperedges, as defined by the adjacency matrix \\(\\hat{\\mathbf{A}}\\), resulting in a sparse attention pattern. As sparsity increases with network depth - demonstrated in Fig. 4b - computational load decreases, while still maintaining compatibility with dense attention during training. Conversely, the edge aggregate and distribution attention mechanisms utilize cross-attention between the vertex and edge feature matrices, \\(\\mathbf{X}^{(V)}\\) and \\(\\mathbf{X}^{(E)}\\), modulated by the soft adjacency matrix \\(\\mathbf{A}\\). This modulation, akin to Fuzzy C-Means in ViHGNNs [17], adjusts attention logits based on soft memberships to the individual hyperedges, dynamically adapting to the hypergraph structure and providing a mechanism for gradient flow into the adjacency matrix generation. Furthermore, by thresholding the soft adjacency matrix during inference, the edge attention mechanisms can be converted into a sparse cross-attention mechanism, thereby reducing computational overhead." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.648, + 0.792, + 0.665 + ], + "angle": 0, + "content": "4.2. Dynamic Adjacency Formation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.67, + 0.909, + 0.821 + ], + "angle": 0, + "content": "HgVT dynamically establishes its hypergraph structure to adapt to the varying semantic and spatial structures of different image inputs. It employs cosine similarity, akin to query-key interactions in attention mechanisms, to evaluate the alignment between vertex and hyperedge adjacency features. This approach allows hyperedges to \"query\" vertices for relevant features, providing a scale-invariant assessment that emphasizes the directionality of embedding vectors. The cosine similarity is subsequently transformed into adjacency membership using a sharpened sigmoid function:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.829, + 0.907, + 0.865 + ], + "angle": 0, + "content": "\\[\n\\mathbf {A} = \\sigma \\left(\\boldsymbol {\\alpha} \\cdot \\tilde {\\mathbf {X}} _ {\\mathrm {a d j}} ^ {(V)} \\left[ \\tilde {\\mathbf {X}} _ {\\mathrm {a d j}} ^ {(E)} \\right] ^ {T}\\right), \\tilde {\\mathbf {X}} _ {\\mathrm {a d j}} ^ {(*)} = \\frac {\\mathbf {X} _ {\\mathrm {a d j}} ^ {(*)}}{| | \\mathbf {X} _ {\\mathrm {a d j}} ^ {(*)} | | _ {2} + \\epsilon} (1)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.91, + 0.902 + ], + "angle": 0, + "content": "Here, \\(\\sigma\\) denotes the sigmoid function and \\(\\alpha = 4\\) is a sharpening factor, which pushes values away from zero to estab" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9754" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.155 + ], + "angle": 0, + "content": "lish binary-like membership values in matrix \\(\\mathbf{A}\\). This soft adjacency matrix is further thresholded to create the hard adjacency matrix \\(\\hat{\\mathbf{A}} = [\\mathbf{A} > 0.5]\\), which provides binary memberships to facilitate sparse attention masking." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.163, + 0.287, + 0.181 + ], + "angle": 0, + "content": "4.3. Architecture Scaling" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.187, + 0.485, + 0.245 + ], + "angle": 0, + "content": "Table 1. Scaling variants of our HgVT architecture. All models are trained at \\(224\\mathrm{x}224\\) resolution, except lite variant (HgVT-Lt), trained at \\(160\\mathrm{x}160\\). Showing count for vertices \\((i\\mathcal{V},v\\mathcal{V})\\), hyperedges \\((p\\mathcal{E}, v\\mathcal{E})\\), dim for adj. \\((d_a)\\) and features \\((d_f)\\), depth \\((L)\\), and heads \\((h)\\)." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.247, + 0.485, + 0.317 + ], + "angle": 0, + "content": "
Model|iV||vV||pE||vE|df+daLhParams FLOPS
HgVT-Lt10012326128+641246.8M0.92B
HgVT-Mi19616508128+64945.8M1.39B
HgVT-Ti19616508128+641247.7M1.80B
HgVT-S19616508224+9614723M5.48B
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.329, + 0.486, + 0.511 + ], + "angle": 0, + "content": "HgVT adopts a hybrid scaling strategy inspired by DeiT [53] and ViG [16], with Tab. 1 detailing the transformer scaling hyperparameters and the vertex and edge type allocations, including fixed capacities for non-image vertices \\((i\\mathcal{V})\\) as proposed by ViHGNN [17]. Our scaling approach aims to balance parameter count and FLOPs with the two main GNN baselines, ViG and ViHGNN. We also introduce a Mini variant (HgVT-Mi) to match DeiT-Ti's computational profile, which we achieve by repeating the final block twice for an effective depth of 9; see Appendix H for details. Finally, we include a Ti-Lite variant (HgVT-Lt) for more efficient ablation studies under a constrained training budget." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.525, + 0.367, + 0.543 + ], + "angle": 0, + "content": "5. Enforcing Semantic Structure" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.55, + 0.486, + 0.734 + ], + "angle": 0, + "content": "Feature matrices for virtual vertices and hyperedges, lacking direct input-based initialization, risk converging to homogeneous solutions and collapsed representations. Additionally, the dynamic adjacency calculation alone fails to naturally promote semantic grouping, in contrast to clustering-based approaches commonly used in vision GNNs. To address these issues, we introduce diversity regularization, which enforces orthogonal embeddings, and population regularization, which encourages a structured, sparse hypergraph. For enhanced semantic differentiation of virtual hyperedge features, we further incorporate an expert-based pooling strategy as a more robust alternative to mean pooling." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.743, + 0.439, + 0.759 + ], + "angle": 0, + "content": "5.1. Diversity-Driven Feature Differentiation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.486, + 0.903 + ], + "angle": 0, + "content": "To prevent homogenization of learned feature matrices and to encourage distinct, semantically rich embeddings, we implement a diversity-driven regularization approach. This method, designed to maintain maximum orthogonality among the embeddings of virtual vertices and hyperedges, penalizes the absolute value of the cosine similarity between different feature vectors, aiming for values close to zero. By using normalized embeddings and masking off diagonal elements to preserve self-similarity, the approach prevents" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.138 + ], + "angle": 0, + "content": "the model from converging to homogeneous solutions or driving individual vectors towards zero magnitude. We then individually penalize \\( v \\mathcal{V} \\), \\( \\mathcal{E} \\), and their adjacency features." + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.144, + 0.903, + 0.18 + ], + "angle": 0, + "content": "\\[\n\\mathrm {D} _ {\\mathrm {L}} (\\mathbf {X}) = \\frac {1}{2} \\sum_ {i j} (1 - \\delta_ {i j}) \\cdot \\left| \\tilde {\\mathbf {X}} \\tilde {\\mathbf {X}} ^ {T} \\right| _ {i j}, \\tilde {\\mathbf {X}} = \\frac {\\mathbf {X}}{| | \\mathbf {X} | | _ {2} + \\epsilon} \\quad (2)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.182, + 0.903, + 0.213 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {D I V}} = \\sum_ {x} \\mathrm {D} _ {\\mathrm {L}} (x), x \\in \\{\\mathbf {X} ^ {(: v V)}, \\mathbf {X} _ {\\mathrm {a d j}} ^ {(: v V)}, \\mathbf {X} ^ {(E)}, \\mathbf {X} _ {\\mathrm {a d j}} ^ {(E)} \\} (3)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.219, + 0.907, + 0.296 + ], + "angle": 0, + "content": "Where \\((:vV)\\) represents the subset of \\(\\mathcal{V}\\) containing only the virtual nodes, \\(\\delta_{ij}\\) is the Kronecker delta function, ensuring that self-similarity is not penalized, and \\(|\\cdot|_{ij}\\) denotes the element-wise absolute value, applied to calculate the penalty for non-orthogonal relationships between embeddings." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.302, + 0.898, + 0.319 + ], + "angle": 0, + "content": "5.2. Population Regularization: Learned Sparsity" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.324, + 0.909, + 0.476 + ], + "angle": 0, + "content": "Unlike clustering methods such as KNN or Fuzzy C-Means, which impose fixed cluster sizes, our model's dynamic adjacency calculation enables flexible and self-adjusting hyperedge populations. To prevent the associated risks of overly sparse or densely connected hypergraphs, we introduce population regularization. This method applies penalties based on the computed soft membership density of each hyperedge derived from the soft adjacency matrix \\(\\mathbf{A}\\), ensuring each maintains a vertex population within appropriate bounds to avoid overgeneralization and preserve hypergraph integrity." + }, + { + "type": "equation", + "bbox": [ + 0.531, + 0.482, + 0.907, + 0.515 + ], + "angle": 0, + "content": "\\[\nP _ {j} = 2 \\cdot \\sum_ {i} \\max (A _ {i j} - 0. 5, 0) \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.531, + 0.516, + 0.907, + 0.548 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {P O P}} = \\sum_ {j} \\max (P _ {j} - \\beta , 0) + \\max (\\gamma - P _ {j}, 0) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.55, + 0.909, + 0.656 + ], + "angle": 0, + "content": "Here, \\( P_{j} \\) represents a soft density estimate of vertex connections for the \\( j^{th} \\) hyperedge, only considering non-zero entries of \\( \\hat{\\mathbf{A}} \\). The hyperparameters \\( \\beta \\) and \\( \\gamma \\) set the upper and lower density limits, ensuring that hyperedges maintain an optimal balance of connections. Penalties are applied if \\( P_{j} \\) either exceeds \\( \\beta \\) or falls below \\( \\gamma \\), maintaining the desired sparsity and ensuring the structural efficacy of the hypergraph." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.663, + 0.88, + 0.68 + ], + "angle": 0, + "content": "5.3. Expert Pooling for Semantic Specialization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.685, + 0.909, + 0.836 + ], + "angle": 0, + "content": "To effectively combine features from multiple virtual hyperedges for classification, our approach utilizes a method akin to expert-choice, where each virtual hyperedge acts as an \"expert\" generating a confidence score. Unlike mean pooling, which risks collapsing distinct features into an average that may dilute individual contributions, this strategy encourages virtual hyperedges to develop unique, semantically meaningful representations. The normalized confidence scores, \\( P(e) \\), determine the relevance of each hyperedge \\( e \\)'s contribution to the classification task." + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.842, + 0.907, + 0.868 + ], + "angle": 0, + "content": "\\[\nP (e) = \\operatorname {s o f t m a x} \\left(\\mathbf {X} _ {e} ^ {(v E)} \\mathbf {W} _ {e} + b _ {e}\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.902 + ], + "angle": 0, + "content": "Here, \\((:vE)\\) denotes the subset of \\((E)\\) containing only the virtual hyperedges, and the softmax is computed across the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9755" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.485, + 0.213 + ], + "angle": 0, + "content": "expert gate set \\(\\{e\\}\\) after projection by \\(\\mathbf{W} \\in \\mathbb{R}^{d \\times |vE|}\\). During training, \\(P(e)\\) guides the weighted averaging of hyperedge features before subsequent class prediction. We further apply a density loss function [3, 14] to prevent underutilization of any single virtual hyperedge, complemented by a cross-entropy term with label smoothing to increase expert confidence. During inference, top-k routing combines the most relevant hyperedges based on their confidence scores." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.227, + 0.445, + 0.243 + ], + "angle": 0, + "content": "6. Empirical Evaluation and Performance" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.253, + 0.485, + 0.358 + ], + "angle": 0, + "content": "This section evaluates the Hypergraph Vision Transformer using two model configurations from Tab. 1: HgVT-Ti-Lite for targeted ablation studies and scaled variants for benchmarking against comparable classifiers. We apply standard augmentations from DeiT [53] using the Timm library [58], including RandAugment [7], Mixup [64], Cutmix [60], Random Erasing [66], and Repeated Augment [21]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.367, + 0.484, + 0.534 + ], + "angle": 0, + "content": "Datasets. For classification in computer vision, we follow standard practices and use the ImageNet-1k dataset [10] at a resolution of \\(224 \\times 224\\) pixels for scaled model evaluations. For ablation studies, we employ ImageNet-100 [51], a 100-class subset of ImageNet-1k with images scaled to \\(160 \\times 160\\) pixels. This selection provides a computationally manageable dataset while maintaining sufficient class variation and larger image sizes compared to datasets like CIFAR-100 (32x32 pixels)[27]. Nevertheless, we find CIFAR-100 useful for assessing the effects of regularization on the hypergraph structure, as detailed in Appendix H." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.542, + 0.484, + 0.724 + ], + "angle": 0, + "content": "Training Hyperparameters. Consistent with DeiT, we use the AdamW optimizer with a weight decay of 0.05. Training is conducted on the ImageNet-1k dataset with a batch size of 1024 for 300 epochs following DeiT. For ablations, we train on ImageNet-100 with a batch size of 512 for with a shorter duration of 200 epochs as proposed by [31]. Learning rates follow a cosine-annealing schedule peaking at 1e-3 for both datasets following scaling from DeiT. Furthermore, we omit the use of Exponential Moving Average (EMA) due to its minimal performance improvement (0.1% in DeiT) relative to its overhead per training step. All models were trained using PyTorch on NVIDIA A6000 GPUs; see Appendix I." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.733, + 0.484, + 0.839 + ], + "angle": 0, + "content": "Evaluation Metrics. Following standard protocols, we measure the Top-1 class prediction accuracy to assess overall performance. Additionally, we take advantage of the learned graph structure (extracted from the last layer) on each image, and measure: Hyperedge Entropy (HE), Intra-Cluster Similarity (ICS), Inter-Cluster Distance (ICD), and Silhouette Score (SIL) [45]; further details on structure in Appendix C." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.315, + 0.864 + ], + "angle": 0, + "content": "6.1. Evaluation on ImageNet" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Tab. 2 presents the ImageNet-1k Top-1 accuracy of HgVT, benchmarked against isotropic models. We limit evaluation" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.908, + 0.118 + ], + "angle": 0, + "content": "Table 2. ImageNet-1k results for HgVT and other isotropic networks. \\(\\star\\) CNN, \\(\\clubsuit\\) Transformer, \\(\\star\\) GNN, \\(\\blacksquare\\) HGNN, and \\(\\triangle\\) HgVT." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.124, + 0.905, + 0.351 + ], + "angle": 0, + "content": "
ModelParamsFLOPsImNet Top-1ReAL Top-1V2 Top-1
*ResMLP-S12 conv3x3 [52]16.7M3.2B77.084.065.5
*ConvMixer-768/32 [55]21.1M20.9B80.2--
*ConvMixer-1536/20 [55]51.6M51.1B81.4--
*DINOv1-S [2]21.7M4.6B77.0--
*ViT-B/16 [12]86.4M55.5B77.983.6-
*DeiT-Ti [53]5.7M1.3B72.280.160.4
*DeiT-S [53]22.1M4.6B79.885.768.5
*DeiT-B [53]86.4M17.6B81.886.771.5
*ViG-Ti [16]7.1M1.3B73.9--
*ViG-S [16]22.7M4.5B80.4--
*ViG-B [16]86.8M17.7B82.3--
■ViHGNN-Ti [17]8.2M1.8B74.3--
■ViHGNN-S [17]23.2M5.6B81.5--
■ViHGNN-B [17]88.1M19.4B82.9--
▲HgVT-Mi (ours)5.8M1.4B74.481.862.7
▲HgVT-Ti (ours)7.7M1.8B76.283.264.3
▲HgVT-S (ours)22.9M5.5B81.286.770.1
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.367, + 0.905, + 0.457 + ], + "angle": 0, + "content": "to isotropic models due to the complexity of down-sampling virtual tokens, lacking direct spatial correspondence, needed for pyramidal structures. As a result, we also exclude other pyramidal models from our comparison as they typically outperform isotropic models within the same architecture family [16, 17], making the comparison less meaningful." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.459, + 0.907, + 0.609 + ], + "angle": 0, + "content": "HgVT-Ti demonstrates a clear advantage, surpassing ViHGNN-Ti by \\(1.9\\%\\) in accuracy, with the smaller HgVT-Mi model matching ViHGNN-Ti with only \\(70\\%\\) of the parameters. At the S-scale, we reduce HgVT's depth relative to ViHGNN to accommodate hyperedge attention and adjacency feature overhead, aiming for comparable parameter counts; as a result, HgVT-S only matches ViHGNN-S. Despite this limitation, HgVT-S matches DeiT-B's accuracy on ImageNet Real [1] and achieves competitive performance on ImageNet V2 [43], at nearly a quarter of DeiT-B's size." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.61, + 0.909, + 0.732 + ], + "angle": 0, + "content": "Overall, HgVT achieves strong performance on classification, demonstrating the benefits of integrating hypergraph structures within a vision transformer framework. Additionally, computational evaluation in Appendix B shows a consistent \\(\\geq 1.4\\times\\) speedup over ViG, with adjacency computation exhibiting a \\(10\\times\\) improvement over ViG's KNN-based method. These improvements further highlight HgVT's potential for greater efficiency over prior Vision GNNs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.743, + 0.679, + 0.757 + ], + "angle": 0, + "content": "6.2. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.901 + ], + "angle": 0, + "content": "We conducted a series of ablations on the ImageNet-100 dataset using the HgVT-Lt model, reporting Top-1 classification accuracy alongside mean hyperedge entropy and silhouette scores to assess the quality of the hypergraph's structure. Notably, we observe a weak anti-correlation between the graph quality metrics and Top-1 accuracy (see Appendix H), indicating opposing objectives. Overall, the ablations are grouped into three categories: regularization, architecture, and pooling methods, with results in Tab. 3." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9756" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.486, + 0.133 + ], + "angle": 0, + "content": "Table 3. ImageNet-100 ablations with HgVT-Lt. Indicating used (✓) or not used \\((\\mathcal{X})\\) , and pooling methods: Average (A), Image \\((I)\\) , Expert \\((E)\\) , Expert \\(^+\\) Image \\((EI)\\) , and \\(EI\\) dropping I \\((DI)\\)" + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.136, + 0.485, + 0.361 + ], + "angle": 0, + "content": "
Ablation on ↓CLS Dropout Stoch. Decay Diversity PopulationTied FFN Xadj = X dF Mult. Pooled OpEdge EntropyTop-1 Acc.Params (M)
X✓✓✓✓1EI3.320.78084.36
none: HgVT-Lt✓✓✓✓✓1E3.130.75182.23
RegularizationXX✓✓✓1E3.120.74581.89
X✓X✓✓1E1.990.72380.79
X✓✓X✓1E3.890.63981.79
X✓XX✓1E3.580.61081.99
ArchitectureX✓✓✓X1E3.090.74182.89
X✓✓✓X1E2.270.80878.62
X✓✓✓✓1E2.120.78076.95
X✓✓✓✓1.5E2.050.77077.46
PoolingX✓✓✓✓1A3.070.74782.06
X✓✓✓✓1I2.930.76084.08
X✓✓✓✓1E3.130.74982.52
X✓✓✓✓1DI3.320.78080.94
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.377, + 0.483, + 0.528 + ], + "angle": 0, + "content": "As shown in Tab. 3, the regularization ablations demonstrate that stochastic path dropout decay [22] improves both Top-1 accuracy and silhouette scores, consistent with ViG and ViHGNN [16, 17]. Omitting Class dropout also boosts accuracy, aligning with DeiT [53]. Furthermore, our proposed diversity and population regularization are essential for preserving graph structure; removing diversity leads to partial representation collapse, while removing population regularization results in near-zero sparsity, effectively turning HgVT into a ViT with increased network complexity." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.529, + 0.483, + 0.666 + ], + "angle": 0, + "content": "In the architecture ablations, untying the FFN improves accuracy but significantly increases parameter count, making it an unfavorable tradeoff. Tying adjacency and embedding features \\((\\mathbf{X}_{\\mathrm{adj}} = \\mathbf{X})\\) reduces parameters and FLOPs but degrades performance, and while untying the FFN or increasing feature dimensionality partially mitigates this, the parameter increase remains suboptimal. This indicates that adjacency and embedding features \\((\\mathbf{X}_{\\mathrm{adj}}^{(*)}\\) and \\(\\mathbf{X}^{(*)})\\) are similar, yet require dedicated feature spaces to avoid crowding." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.667, + 0.483, + 0.803 + ], + "angle": 0, + "content": "For pooling methods, expert edge pooling outperforms average edge pooling in accuracy, while image pooling achieves the highest accuracy at the cost of degraded graph structure. Combining image and expert pooling recovers lost structure and improves accuracy, with each input focusing on different semantic levels (see Appendix D). Additionally, dropping the pooled image embedding prior to the classifier head maintains moderate performance, indicating that both paths meaningfully contribute to the final prediction." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.81, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Impact of Vertex Self-Attention and Patch Embedding. We evaluated the impact of patch embedding methods and vertex self-attention, comparing a convolutional stem (Conv2D-BN-GELU layers [16, 17]) and a simpler patch projection (pixel-shuffled patches with affine projection [2, 12]), across various pooling strategies (average, image, and ex" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.089, + 0.907, + 0.133 + ], + "angle": 0, + "content": "Table 4. Top-1 accuracy of HgVT-Lt on ImageNet-100 with \\((\\checkmark)\\) or without \\((X)\\) vertex self-attention. Contrasting pooling operations and patch embedding versions: Conv. Stem or Patch Projection." + }, + { + "type": "table", + "bbox": [ + 0.53, + 0.137, + 0.888, + 0.202 + ], + "angle": 0, + "content": "
Pooling Op. →AverageImageExpert
Vertex SA →X✓X✓X✓
Conv. Stem78.0282.0682.0584.0878.8782.52
Patch Project64.3072.7670.7676.1762.6271.43
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.214, + 0.908, + 0.397 + ], + "angle": 0, + "content": "pert), as shown in Tab. 4. The patch projection consistently underperforms the convolutional stem, likely due to the model's small size limiting its effectiveness. Omitting vertex self-attention leads to further degradation, especially without the convolutional stem, suggesting it is crucial for effectively separating features in low-dimensional space. PCA shows that 71/128 channels are needed to explain \\(95\\%\\) of the variance for the convolutional stem, compared to 19/128 for the patch projection, indicating the richer representation captured by the convolutional stem. Notably, image pooling shows the least degradation, likely due to its more direct gradient flow compared to the edge pooling methods." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.406, + 0.847, + 0.422 + ], + "angle": 0, + "content": "6.3. Pooling Methods and Graph Structure" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.429, + 0.907, + 0.471 + ], + "angle": 0, + "content": "Table 5. Impact of graph structure on pooling operation for HgVT-Lt on ImageNet-100. Measuring graph metrics for image vertices \\((i\\mathcal{V})\\) and all vertices \\((\\mathcal{V})\\), along with features from DINOv2." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.476, + 0.907, + 0.57 + ], + "angle": 0, + "content": "
Pooling Op. →ImageExpertImg. & Expert
Feature Model ↓HEICSICDHEICSICDHEICSICD
HgVT-Lt (iV)3.030.500.313.240.430.373.230.420.36
HgVT-Lt (V)3.200.250.723.390.360.453.390.280.58
DINO2-S (iV)3.040.840.083.250.850.053.250.830.06
DINO2-G (iV)3.040.690.153.250.680.123.250.680.13
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.909, + 0.795 + ], + "angle": 0, + "content": "To evaluate the impact of pooling methods on graph structure, we measured HE, ICS, and ICD using the ImageNet-100 validation set with three strategies: image pooling, expert pooling, and a combined image + expert pooling approach. Metrics used either the image vertex subset \\((i\\mathcal{V})\\) or the full vertex set \\((\\mathcal{V})\\), with features derived from DINOv2 (S and G) [38] and HgVT-Lt using HgVT-Lt's adjacency matrix (A). Results in Tab. 5 show that while all methods achieve similar graph quality using \\(i\\mathcal{V}\\), image pooling slightly improves similarity. However, including all vertices \\((\\mathcal{V})\\) consistently increases ICD and entropy while reducing ICS, indicating decreased graph coherence. This effect is more severe with image pooling methods, suggesting that virtual vertices \\((v\\mathcal{V})\\) act as noisy elements, rather than summarization points." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Comparing DINOv2 models, all pooling methods align more closely with DINOv2-G, where achieving a balance between ICS and ICD is preferable to maximizing either individually. This trend, along with consistent HE, suggests a focus on higher-level detail, irrespective of the smaller HgVT model size or pooling method. Image pooling shows slightly stronger alignment with both DINO models, indicating that" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9757" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.198 + ], + "angle": 0, + "content": "both high- and low-level semantics are encoded within a single feature space, unlike methods that can use edge channels for high-level concepts. Notably, all expert pooling methods exhibit an emergent macro-class prediction behavior, where each virtual edge \\((v\\mathcal{E})\\) consistently captures broader taxonomic groups (e.g., dogs, birds). Further representation and macro-class analysis are provided in Appendices D and J." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.21, + 0.379, + 0.226 + ], + "angle": 0, + "content": "6.4. Performance on Image Retrieval" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.233, + 0.485, + 0.504 + ], + "angle": 0, + "content": "To evaluate the capability of HgVT in capturing semantic structures, we perform image retrieval experiments comparing four methods: pooling similarity (PS), volumetric similarity (VS), adaptive pooling similarity (APS), and adaptive volumetric similarity (AVS). PS ranks the pooled embeddings by cosine similarity (vector search), while the other methods enhance retrieval by leveraging graph structure. Volumetric similarity calculates ellipsoid overlap using an approximate Mahalanobis distance, with pruned hyperedges defining the distribution spread around the pooled embedding (centroid). Adaptive methods further refine retrieval via a graph similarity measure on the pruned hyperedges, re-ranking from a shortlist of \\( R = 100 \\). Computational efficiency in adaptive retrieval is ensured through centroid hash binning and limiting comparisons to the top-\\( C = 4 \\) most significant query hyperedges, resulting in a complexity of \\( \\mathcal{O}(RC) \\). Notably, we prune to 12 hyperedges and use 10 centroid bins. Additional details provided in Appendix G." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.513, + 0.485, + 0.556 + ], + "angle": 0, + "content": "Table 6. Image Retrieval on ImageNet-1k with proposed search methods: PS, APS, VS, and AVS. Reporting Top-1 accuracy, mAP@10 (\\%), and 1-NN-hit@10 (\\%) for pooling methods." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.56, + 0.482, + 0.704 + ], + "angle": 0, + "content": "
mAP@10 (%)
ModelTop-1PoolingPSAPSVSAVS
MRL-128 [29]70.52Image64.9465.20--
MRL-256 [29]70.62Image65.0465.20--
HgVT-Ti (ours)76.18Expert70.5669.5970.5369.40
HgVT-Ti (ours)76.18Im&Ex73.2369.5973.0869.49
1-NN-hit@10 (%)
HgVT-Ti (ours)76.18Expert21.2219.1021.2519.56
HgVT-Ti (ours)76.18Im&Ex25.1319.1025.2219.17
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.486, + 0.901 + ], + "angle": 0, + "content": "ImageNet Retrieval. We evaluate retrieval performance on the ImageNet-1K dataset to assess HgVT's ability to capture semantic relationships and compare four retrieval methods: PS, VS, APS, and AVS. The primary metric is mAP@10, with MRL [29] serving as the baseline due to its adaptive re-ranking approach. We also report the 1-NN-CLIP-L hitrate@10, which measures how often the top-1 result ranked by CLIP-L [42] appears in the top-10 retrieved results, offering additional insight into semantic alignment. Results in Tab. 6 show that HgVT-Ti surpasses MRL by over \\(8\\%\\) in retrieval performance, despite being significantly smaller than MRL (ResNet-50) and using a comparably compact" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.169 + ], + "angle": 0, + "content": "embedding size (\\(d = 2 \\times 128\\)). Among our methods, PS and VS achieve similar results, while APS and AVS underperform, likely due to their focus on exact-feature similarity and ambiguous-class features, which limits alignment with the high-level semantics required by ImageNet's diverse dataset." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.173, + 0.907, + 0.201 + ], + "angle": 0, + "content": "Table 7. Image Retrieval on revisited Oxford and Paris; reporting mAP \\((\\%)\\). Showing training set and method used. *mAP@100." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.203, + 0.905, + 0.368 + ], + "angle": 0, + "content": "
Dataset →ROxfordRParis
ModelTrain SetMethodMHMH
ALEX+GeM [44]ImNet-1kPS33.810.452.726.0
RN101+R-MAC [44]ImNet-1kPS49.818.574.052.1
DINOv1-S/16 [2]ImNet-1kPS41.813.763.134.4
DINOv1-S/16 [2]GLDv2PS51.524.375.351.6
HgVT-Ti (ours)ImNet-1kVS26.810.555.428.7
VS*25.89.065.628.0
AVS*27.06.864.126.7
HgVT-S (ours)ImNet-1kVS28.012.156.731.1
VS*26.310.265.028.4
AVS*27.410.365.027.1
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.38, + 0.909, + 0.637 + ], + "angle": 0, + "content": "Oxford and Paris Retrieval. To evaluate image retrieval performance beyond simple class retrieval, we use the Oxford and Paris Revisited datasets [39, 41], which provide three splits of increasing difficulty (Easy, Medium, and Hard) for query/database pairs. We report Mean Average Precision (mAP) for the Medium (M) and Hard (H) splits, using mAP@100 for AVS based on short-list ranking, while full mAP and mAP@100 are provided for VS as a baseline comparison. Results, shown in Tab. 7, indicate that HgVT achieves competitive performance with similarly sized feature extractors, though performance on ROxford-M lags. This shortfall may stem from subtle landmark differences better captured by multi-scale Conv-Nets and self-supervised learning, compared to the more salient focus driven by HgVT's classifier training. However, AVS outperforms VS on ROxford-M, demonstrating its ability to uncover finer feature similarities within the hypergraph structure." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.649, + 0.822, + 0.665 + ], + "angle": 0, + "content": "7. Conclusion and Future Directions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.901 + ], + "angle": 0, + "content": "In this work, we introduced the Hypergraph Vision Transformer (HgVT), which combines hypergraph structures with vision transformers. Our methods, including diversity and population regularization, along with expert edge pooling, enhance semantic representation and efficiency by enabling dynamic hypergraph construction. Although challenging to tune, these methods avoid reliance on traditional clustering techniques, reflecting a tradeoff between complexity and efficiency. Empirical results show that HgVT achieves strong performance on both image classification and retrieval, positioning HgVT as a competitive framework for semantic vision tasks. Future work will focus on exploring the scalability of hypergraph structures and integrating self-supervised learning to further improve adaptability and better decouple saliency from semantic vision graph generation." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9758" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.486, + 0.169 + ], + "angle": 0, + "content": "[1] Lucas Beyer, Olivier J. Henaff, Alexander Kolesnikov, Xiaohua Zhai, and Aaron van den Oord. Are we done with imageNet? arXiv preprint arXiv:2006.07159, abs/2006.07159, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.173, + 0.484, + 0.243 + ], + "angle": 0, + "content": "[2] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jegou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9630-9640, 2021. 1, 2, 6, 7, 8, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.245, + 0.484, + 0.327 + ], + "angle": 0, + "content": "[3] Tianlong Chen, Zhenyu Zhang, Ajay Kumar Jaiswal, Shiwei Liu, and Zhangyang Wang. Sparse moe as the new dropout: Scaling dense and self-slimmable transformers. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.33, + 0.484, + 0.399 + ], + "angle": 0, + "content": "[4] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.402, + 0.485, + 0.441 + ], + "angle": 0, + "content": "[5] Norman Cliff. Dominance statistics: Ordinal analyses to answer ordinal questions. Psychological Bulletin, 114:494-509, 1993. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.445, + 0.484, + 0.526 + ], + "angle": 0, + "content": "[6] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.531, + 0.484, + 0.599 + ], + "angle": 0, + "content": "[7] Ekin Dogus Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V. Le. Randaugment: Practical automated data augmentation with a reduced search space. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 3008-3017, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.602, + 0.484, + 0.669 + ], + "angle": 0, + "content": "[8] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. In *The Twelfth International Conference on Learning Representations*, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.673, + 0.484, + 0.728 + ], + "angle": 0, + "content": "[9] Dmitry Demidov, M.H. Sharif, Aliakbar Abdurahimov, Hisham Cholakkal, and Fahad Shahbaz Khan. Salient mask-guided vision transformer for fine-grained classification. In VISIGRAPP, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.731, + 0.484, + 0.771 + ], + "angle": 0, + "content": "[10] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. In CVPR09, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.775, + 0.484, + 0.829 + ], + "angle": 0, + "content": "[11] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In North American Chapter of the Association for Computational Linguistics, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.094, + 0.908, + 0.12 + ], + "angle": 0, + "content": "scale. In International Conference on Learning Representations, 2021. 1, 2, 6, 7, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.123, + 0.908, + 0.19 + ], + "angle": 0, + "content": "[13] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The PASCAL Visual Object Classes Challenge 2012 (VOC2012) Results. http://www.pascalnetwork.org/challenges/VOC/voc2012/workshop/index.html. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.193, + 0.907, + 0.247 + ], + "angle": 0, + "content": "[14] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. J. Mach. Learn. Res., 23:120:1-120:39, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.907, + 0.304 + ], + "angle": 0, + "content": "[15] Yue Gao, Meng Wang, Dacheng Tao, Rongrong Ji, and Qionghai Dai. 3-d object retrieval and recognition with hypergraph analysis. IEEE Transactions on Image Processing, 21(9): 4290-4303, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.307, + 0.907, + 0.361 + ], + "angle": 0, + "content": "[16] Kai Han, Yunhe Wang, Jianyuan Guo, Yehui Tang, and Enhua Wu. Vision GNN: An image is worth graph of nodes. In Advances in Neural Information Processing Systems, 2022. 1, 2, 5, 6, 7, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.364, + 0.907, + 0.432 + ], + "angle": 0, + "content": "[17] Yan Han, Peihao Wang, Souvik Kundu, Ying Ding, and Zhangyang Wang. Vision hgnn: An image is more than a graph of nodes. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 19878-19888, 2023. 1, 2, 3, 4, 5, 6, 7, 27" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.435, + 0.906, + 0.503 + ], + "angle": 0, + "content": "[18] Yunusa Haruna, Shiyin Qin, Abdulrahman Hamman Adama Chukkol, Abdulganiyu Abdu Yusuf, Isah Bello, and Adamu Lawan. Exploring the synergies of hybrid cnns and vits architectures for computer vision: A survey. arXiv preprint arXiv:2402.02941, abs/2402.02941, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.505, + 0.906, + 0.56 + ], + "angle": 0, + "content": "[19] Ali Hassani, Steven Walton, Jiachen Li, Shen Li, and Humphrey Shi. Neighborhood attention transformer. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.562, + 0.907, + 0.617 + ], + "angle": 0, + "content": "[20] Kaiming He, X. Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2015. 1, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.62, + 0.906, + 0.687 + ], + "angle": 0, + "content": "[21] Elad Hoffer, Tal Ben-Nun, Itay Hubara, Niv Giladi, Torsten Hoefler, and Daniel Soudry. Augment your batch: Improving generalization through instance repetition. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8126-8135, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[22] Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Weinberger. Deep networks with stochastic depth, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.719, + 0.907, + 0.772 + ], + "angle": 0, + "content": "[23] Yuchi Huang, Qingshan Liu, and Dimitris Metaxas. ]video object segmentation by hypergraph cut. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 1738-1745, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.776, + 0.907, + 0.831 + ], + "angle": 0, + "content": "[24] Yuchi Huang, Qingshan Liu, Shaoting Zhang, and Dimitris N. Metaxas. Image retrieval via probabilistic hypergraph ranking. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 3376-3383, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.833, + 0.907, + 0.899 + ], + "angle": 0, + "content": "[25] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. 15" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9759" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.175 + ], + "angle": 0, + "content": "[26] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3992-4003, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.178, + 0.484, + 0.205 + ], + "angle": 0, + "content": "[27] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.484, + 0.262 + ], + "angle": 0, + "content": "[28] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.484, + 0.332 + ], + "angle": 0, + "content": "[29] Aditya Kusupati, Gantavya Bhatt, Aniket Rege, Matthew Wallingford, Aditya Sinha, Vivek Ramanujan, William Howard-Snyder, Kaifeng Chen, Sham M. Kakade, Prateek Jain, and Ali Farhadi. Matryoshka representation learning. In Neural Information Processing Systems, 2022. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.484, + 0.376 + ], + "angle": 0, + "content": "[30] Y. Lecun, L. Bottou, Y. Bengio, and P. Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.378, + 0.484, + 0.446 + ], + "angle": 0, + "content": "[31] Sangjun Lee, Inwoo Hwang, Gi-Cheon Kang, and Byoung-Tak Zhang. Improving robustness to texture bias via shape-focused augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 4323-4331, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.449, + 0.486, + 0.545 + ], + "angle": 0, + "content": "[32] Benjamin Lefaudeau, Francisco Massa, Diana Liskovich, Wenhan Xiong, Vittorio Caggiano, Sean Naren, Min Xu, Jieru Hu, Marta Tintore, Susan Zhang, Patrick Labatut, Daniel Haziza, Luca Wehrstedt, Jeremy Reizenstein, and Grigory Sizov. xformers: A modular and hackable transformer modelling library. https://github.com/facebookresearch/xformers, 2022. 28" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.547, + 0.486, + 0.616 + ], + "angle": 0, + "content": "[33] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9992-10002, 2021. 1, 2, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.617, + 0.486, + 0.671 + ], + "angle": 0, + "content": "[34] Leland McInnes, John Healy, Nathaniel Saul, and Lukas Großberger. Umap: Uniform manifold approximation and projection. Journal of Open Source Software, 3(29):861, 2018. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.675, + 0.486, + 0.744 + ], + "angle": 0, + "content": "[35] Mustafa Munir, William Avery, and Radu Marculescu. Mobilevig: Graph-based sparse attention for mobile vision applications. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 2211-2219, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.486, + 0.814 + ], + "angle": 0, + "content": "[36] Mustafa Munir, William Avery, Md Mostafijur Rahman, and Radu Marculescu. Greedyvig: Dynamic axial graph construction for efficient vision gnns. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6118-6127, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.486, + 0.843 + ], + "angle": 0, + "content": "[37] NVIDIA. Apex: A pytorch extension. https://github.com/NVIDIA/apex, 2020.28" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.486, + 0.901 + ], + "angle": 0, + "content": "[38] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.092, + 0.908, + 0.175 + ], + "angle": 0, + "content": "Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINOv2: Learning robust visual features without supervision. Transactions on Machine Learning Research, 2024. 1, 2, 7, 6, 8, 14, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.908, + 0.247 + ], + "angle": 0, + "content": "[39] James Philbin, Ondrej Chum, Michael Isard, Josef Sivic, and Andrew Zisserman. Lost in quantization: Improving particular object retrieval in large scale image databases. In 2008 IEEE Conference on Computer Vision and Pattern Recognition, pages 1-8, 2008. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.908, + 0.303 + ], + "angle": 0, + "content": "[40] Joan Puigcerver, Carlos Riquelme Ruiz, Basil Mustafa, and Neil Houlsby. From sparse to soft mixtures of experts. In The Twelfth International Conference on Learning Representations, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.907, + 0.374 + ], + "angle": 0, + "content": "[41] Filip Radenovic, Ahmet Iscen, Giorgos Tolias, Yannis Avrithis, and Ondrej Chum. Revisiting oxford and paris: Large-scale image retrieval benchmarking. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5706-5715, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.376, + 0.908, + 0.459 + ], + "angle": 0, + "content": "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 8, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.461, + 0.908, + 0.515 + ], + "angle": 0, + "content": "[43] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do imagenet classifiers generalize to imageNet? In International Conference on Machine Learning, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.517, + 0.908, + 0.585 + ], + "angle": 0, + "content": "[44] Jérôme Revaud, Jon Almazán, Rafael Sampaio de Rezende, and César Roberto de Souza. Learning with average precision: Training image retrieval with a listwise loss. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 5106-5115, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.588, + 0.908, + 0.642 + ], + "angle": 0, + "content": "[45] Peter J. Rousseeuw. Silhouettes: A graphical aid to the interpretation and validation of cluster analysis. Journal of Computational and Applied Mathematics, 20:53-65, 1987. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.645, + 0.908, + 0.699 + ], + "angle": 0, + "content": "[46] Franco Scarselli, Marco Gori, Ah Chung Tsoi, Markus Hagenbuchner, and Gabriele Monfardini. The graph neural network model. IEEE Transactions on Neural Networks, 20(1):61-80, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.702, + 0.907, + 0.73 + ], + "angle": 0, + "content": "[47] Noam Shazeer. GLU variants improve transformer. arXiv preprint arXiv:2002.05202, abs/2002.05202, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.732, + 0.908, + 0.786 + ], + "angle": 0, + "content": "[48] Dai Shi. Transnext: Robust foveal visual perception for vision transformers. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17773-17783, 2023. 14, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.788, + 0.907, + 0.843 + ], + "angle": 0, + "content": "[49] Dai Shi. Transnext: Robust foveal visual perception for vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17773-17783, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.908, + 0.901 + ], + "angle": 0, + "content": "[50] Sakhinana Sagar Srinivas, Rajat Kumar Sarkar, Sreeja Gangasani, and Venkataramana Runkana. Vision HgNN: An electron-micrograph is worth hypergraph of hypernodes. arXiv preprint arXiv:2408.11351, abs/2408.11351, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "9760" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.147 + ], + "angle": 0, + "content": "[51] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive multiview coding. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XI 16, pages 776-794. Springer, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.484, + 0.231 + ], + "angle": 0, + "content": "[52] Hugo Touvron, Piotr Bojanowski, Mathilde Caron, Matthieu Cord, Alaaeldin El-Nouby, Edouard Grave, Armand Joulin, Gabriel Synnaeve, Jakob Verbeek, and Hervé Jégou. Resmlp: Feedforward networks for image classification with data-efficient training. arXiv preprint arXiv:2105.03404, abs/2105.03404, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.484, + 0.314 + ], + "angle": 0, + "content": "[53] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herve Jegou. Training data-efficient image transformers amp; distillation through attention. In Proceedings of the 38th International Conference on Machine Learning, pages 10347-10357. PMLR, 2021. 1, 2, 5, 6, 7, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.484, + 0.411 + ], + "angle": 0, + "content": "[54] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurélien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, abs/2302.13971, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.413, + 0.484, + 0.439 + ], + "angle": 0, + "content": "[55] Asher Trockman and J. Zico Kolter. Patches are all you need? Trans. Mach. Learn. Res., 2023, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.484, + 0.509 + ], + "angle": 0, + "content": "[56] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.484, + 0.564 + ], + "angle": 0, + "content": "[57] Petar Velicković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. Graph attention networks. In International Conference on Learning Representations, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.484, + 0.607 + ], + "angle": 0, + "content": "[58] Ross Wightman. Pytorch image models. https://github.com/rwrightman/pytorch-image-models, 2019.6, 28" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.609, + 0.484, + 0.661 + ], + "angle": 0, + "content": "[59] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In European Conference on Computer Vision. Springer, 2018. 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.665, + 0.484, + 0.733 + ], + "angle": 0, + "content": "[60] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Young Joon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6022-6031, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.484, + 0.802 + ], + "angle": 0, + "content": "[61] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 1204-1213. IEEE, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.484, + 0.845 + ], + "angle": 0, + "content": "[62] Biao Zhang and Rico Sennrich. Root mean square layer normalization. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.847, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[63] Bowen Zhang, Zhi Tian, Quan Tang, Xiangxiang Chu, Xiaolin Wei, Chunhua Shen, and Yifan liu. Segvit: Semantic segmentation with plain vision transformers. In Advances in Neural Information Processing Systems, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.147 + ], + "angle": 0, + "content": "[64] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In International Conference on Learning Representations, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.908, + 0.205 + ], + "angle": 0, + "content": "[65] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 6230-6239, 2016. 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.907, + 0.247 + ], + "angle": 0, + "content": "[66] Zhun Zhong, Liang Zheng, Guoliang Kang, Shaozi Li, and Yi Yang. Random erasing data augmentation. AAAI, 34: 13001-13008, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.908, + 0.303 + ], + "angle": 0, + "content": "[67] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision, 127:302 - 321, 2016. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.908, + 0.373 + ], + "angle": 0, + "content": "[68] Lei Zhu, Xinjiang Wang, Zhanghan Ke, Wayne Zhang, and Rynson Lau. Bifformer: Vision transformer with bi-level routing attention. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.373 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.956 + ], + "angle": 0, + "content": "9761" + } + ] +] \ No newline at end of file diff --git a/2025/Hypergraph Vision Transformers_ Images are More than Nodes, More than Edges/b3c66b72-bf5c-47e0-a296-e834f5e49c0f_origin.pdf b/2025/Hypergraph Vision Transformers_ Images are More than Nodes, More than Edges/b3c66b72-bf5c-47e0-a296-e834f5e49c0f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4e037604373f3be48a3476c1b6c5b69d90b692dc --- /dev/null +++ b/2025/Hypergraph Vision Transformers_ Images are More than Nodes, More than Edges/b3c66b72-bf5c-47e0-a296-e834f5e49c0f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:77e1de36cf0c35e2a84bd846edc9e1a09fed9b1224f0ee8bec37d62bb064daab +size 570246 diff --git a/2025/Hypergraph Vision Transformers_ Images are More than Nodes, More than Edges/full.md b/2025/Hypergraph Vision Transformers_ Images are More than Nodes, More than Edges/full.md new file mode 100644 index 0000000000000000000000000000000000000000..522c2080932904cc05b8aa772322b8bd3ec1df1b --- /dev/null +++ b/2025/Hypergraph Vision Transformers_ Images are More than Nodes, More than Edges/full.md @@ -0,0 +1,336 @@ +# Hypergraph Vision Transformers: Images are More than Nodes, More than Edges + +Joshua Fixelle +University of Virginia +jf9fk@virginia.edu + +# Abstract + +Recent advancements in computer vision have highlighted the scalability of Vision Transformers (ViTs) across various tasks, yet challenges remain in balancing adaptability, computational efficiency, and the ability to model higher-order relationships. Vision Graph Neural Networks (ViGs) offer an alternative by leveraging graph-based methodologies but are hindered by the computational bottlenecks of clustering algorithms used for edge generation. To address these issues, we propose the Hypergraph Vision Transformer (HgVT), which incorporates a hierarchical bipartite hypergraph structure into the vision transformer framework to capture higher-order semantic relationships while maintaining computational efficiency. HgVT leverages population and diversity regularization for dynamic hypergraph construction without clustering, and expert edge pooling to enhance semantic extraction and facilitate graph-based image retrieval. Empirical results demonstrate that HgVT achieves strong performance on image classification and retrieval, positioning it as an efficient framework for semantic-based vision tasks. + +# 1. Introduction + +Computer vision has recently transitioned from the historically dominant Convolutional Neural Networks (CNNs) [20, 28, 30] to the increasingly prominent Vision Transformers (ViTs), which have quickly embedded themselves as the new de facto standard [12, 38]. This shift reflects the broader success of transformers in natural language processing [11, 54, 56] and is driven by the remarkable scalability of ViTs across various tasks such as image classification [53, 61], semantic segmentation [26, 63], and image retrieval [2, 29]. While hybrid models like hierarchical attention and CNN-ViT methods [18, 19, 33] aim to balance computational load and flexibility, ViTs still struggle with focusing on salient features rather than comprehensive image understanding [2, 9, 12, 40]. This highlights the need for methods that improve both efficiency and semantic accuracy. + +Within the spectrum of novel architectures, Vision Graph Neural Networks (ViGs) [16] and Vision Hypergraph Neural + +Networks (ViHGNNs) [17] leverage graph-based topologies to advance image processing. Unlike CNNs, which harness locality and translation-invariance through densely connected pixel grids and repeated convolutions, both ViTs and ViGs represent images as sets of patches. In ViTs, each patch acts as a vertex within a maximally connected graph, creating semantically weak connections through self-attention. ViGs enhance this by using clustering algorithms to detect edge groupings and applying graph convolutions to these clusters, forming meaningful patch relationships. ViHGNNs extend these capabilities by employing hyperedges that capture complex, higher-order relationships, enriching understanding of the images. These methodologies are depicted in Figure 1. + +While graph-based models like ViG and ViHGNN have introduced notable advancements in visual perception, two critical observations emerge regarding these architectures: + +1. In existing vision GNN models [16, 17, 35, 36], edge features are primarily used for basic vertex-to-vertex communication and are not integrated across successive layers: a strategy that could enhance cumulative learning and improve classification accuracy. +2. The computational complexities associated with clustering algorithms used for edge generation, such as KNN in ViG and Fuzzy C-Means in ViHGNN, pose significant computational bottlenecks. Approaches like MobileViG [35] and GreedyViG [36] attempt to mitigate these challenges with static graph structures and adding dynamic masking, but do so by trading adaptability for efficiency, failing to achieve a well-balanced solution. + +In response to limitations in existing graph-based models, we propose the Hypergraph Vision Transformer (HgVT), which advances the hypergraph concept with a bipartite representation where hyperedge features and image patches (vertices) are continuously processed. Unlike traditional models that use graph convolutions, HgVT employs structured multihead attention for efficient vertex-hyperedge message passing and incorporates a dynamic querying mechanism that constructs graph structures in $\mathcal{O}(|V| \cdot E)$ time complexity, where $E < |V|$ . This graph structure is then utilized in attention masking to balance structural adaptability with computational efficiency. Furthermore, HgVT integrates + +![](images/3d7919bc3e33a1d7a2fc5f7b472d8873ead76dd6883c0ebece9cc42ae71b2beb.jpg) +(a) CNN + +![](images/0aea464e472c471a64365e7bb55212f6001bb9141c514e309474e24d08c74bf4.jpg) +(c) ViG + +![](images/cb6a1a05cf39a3f65ae9e5efed64cbed57528fd570db7d877dba53201c454ee3.jpg) +(e) HgVT + +![](images/72cd0db8a09ed4a1847b085b5606afae86b9b78120c967bcdffa2365fb1be12e.jpg) +(b) ViT + +![](images/03f6cb301c05c4006f9a09f32ae726be2c2c02775d834efd107a30adb84fa2b1.jpg) +(d) ViHGNN +Figure 1. Comparison of Graph Structures for different methods. Showing (a) CNNs, (b) Vision Transformers, (c) ViG with a KNN clustered GNN, (d) ViHGNN with clustered hyperedges, and (e) our proposed HgVT method. Strong group edges shown with solid lines; weak edges with dashed lines. Hyperedges shown with shaded regions; less dominant hyperedges with gray dashed regions. + +virtual elements into vertices and hyperedges, forming intermediate links that create a hierarchical semantic structure, as illustrated in Fig. 1e. Attention masking is then used to restrict message passing, and virtual hyperedge features are leveraged for classification. Our contributions are as follows: + +- We propose the Hypergraph Vision Transformer (HgVT), which integrates a hierarchical bipartite hypergraph structure within a vision transformer framework. Our isotropic HgVT-Ti model achieves a Top-1 accuracy of $76.2\%$ on the ImageNet-1k classification task, surpassing the prior state-of-the-art by $1.9\%$ , demonstrating the efficacy of hypergraph-based learning within vision transformers. +- We introduce population and diversity regularization strategies that enable dynamic hypergraph structure construction in HgVT, allowing the model to self-sparseify connections without relying on traditional clustering techniques. +- We implement expert edge pooling, a pooling approach that selects edges based on learned confidence scores, facilitating efficient representation pruning and graph extraction. This approach demonstrates strong semantic clustering behavior across macro-classes and achieves competitive image retrieval performance compared to other feature extractors, while maintaining a more compact model size. + +# 2. Related Work + +Vision Transformers. Vision Transformers (ViTs) proposed by [12] and refined by [2, 38, 53] use self-attention to process image patches as sequences, scaling to complex datasets and tasks. Recent ViTs have reintroduced spatial hierarchies by leveraging local attention [19, 33], integrating sparse global summaries [68], and employing biomimetic modeling to focus on key regions within images [49]. However, current models tend to focus on the most salient objects and patch-level similarities, ignoring global structure. HgVT addresses this by introducing bipartite hypergraphs to model higher-order relationships for improved semantic understanding. + +Graph-Based Vision Models and Clustering. Graph Neural Networks (GNNs), initially conceptualized by [46], have + +been applied to vision tasks through Vision Graph Neural Networks (ViGs) [16], which exhibit improved accuracy over ViTs on common vision tasks. ViGs use graph convolutions to model image patch relationships on a graph structure, typically constructed by iterative clustering algorithms such as KNN and Fuzzy C-Means, which introduce computational overhead. Recent methods avoid clustering inefficiencies with static graph structures [35, 36], sacrificing adaptability. HgVT instead introduces a dynamic graph construction method, relying on cosine similarity from learned features to enable efficient, non-iterative, adaptive clustering. + +Hypergraph-Based Methods. While previously used in many computer vision tasks [15, 23, 24], hypergraphs have recently been incorporated into vision GNNs [17, 50], improving their ability to model complex multi-way relationships. However, these methods treat hypergraphs as an intermediate tool rather than producing a hypergraph to represent underlying images, preventing their use in downstream tasks. HgVT instead iteratively refines a hypergraph through subsequent network layers to produce structured representations. + +# 3. Hierarchical Hypergraphs + +Graphs and Basic Notations. Graphs are powerful mathematical tools for representing structured information, applicable across diverse disciplines. A graph $\mathcal{G}$ is defined as a pair $(\mathcal{V},\mathcal{E})$ , where $\mathcal{V} = \{v_{1},v_{2},\ldots ,v_{N}\}$ is a set of vertices, and $\mathcal{E} = \{e_{ij}|(v_i,v_j)\}$ is a set of directed edges. For undirected graphs, the edges are symmetric, such that $e_{ij} = e_{ji}$ . Each edge $e_{ij}$ connects a pair of vertices $v_{i}$ and $v_{j}$ where $v_{i},v_{j}\in \mathcal{V}$ . The adjacency matrix $\hat{\mathbf{A}}$ is a binary matrix $\{0,1\}^{|V|\times |V|}$ , representing the presence (1) or absence (0) of an edge between each pair of vertices. Similarly, an edge weight matrix can be defined as $\mathbf{A}\in \mathbb{R}^{|V|\times |V|}$ to quantify the strength or capacity of these connections. + +Graph Convolution Networks (GCNs). Building on this foundation, GCNs utilize vertex feature matrices $\mathbf{X} \in \mathbb{R}^{|V| \times d}$ to encode vertex properties. Their core mechanism, message passing, updates vertex features through a convolution + +with a learned projection matrix $\mathbf{W} \in \mathbb{R}^{d \times d}$ and a non-linear activation, guided by the adjacency matrix $\hat{\mathbf{A}}$ , which specifies neighboring vertices. Adjacency features $\mathbf{X}_{\mathrm{adj}} \in \mathbb{R}^{|V| \times d_a}$ , typically set as $\mathbf{X}_{\mathrm{adj}} = \mathbf{X}$ , enable dynamic updates to $\hat{\mathbf{A}}$ and the edge weight matrix $\mathbf{A}$ , refining the graph structure through learned interactions. However, GCNs are unable to model multi-vertex relationships as edges in $\mathcal{E}$ are pairwise. + +# 3.1. Hypergraphs and Bipartite Representations + +![](images/f02d2b1b429d3ba3aabe6d6c87fcfd98faee49b413293856ad186333b6f851c7.jpg) +(a) Hypergraph + +![](images/f8d8f04c1031e4eb2eb532b2bca4efa2f9e8ff744965ef4ac57bdba3c12fe7c9.jpg) +(b) Bipartite Form +Figure 2. Comparison of (a) hypergraph and (b) equivalent bipartite representation from Fig. 1d, showing five hyperedges $(e_0 - e_4)$ . + +To overcome the pairwise limitation inherent in traditional graphs, hypergraphs offer a robust solution by extending the concept of edges to hyperedges, which connect multiple vertices simultaneously. In a hypergraph $\mathcal{H} = (\mathcal{V},\mathcal{E})$ , hyperedges $e_j\in \mathcal{E}$ each connect a subset of vertices, defined as $e_j = \{v_i\mid v_i\in \mathcal{V}$ and $i\in I_j\}$ , where $I_{j}$ is the set of indices for vertices that are included in hyperedge $e_j$ . The set $I_{j}$ directly corresponds to the nonzero entries of the $j$ -th column of the incidence matrix $\mathbf{H}\in \{0,1\}^{|\mathcal{V}|\times |\mathcal{E}|}$ , where $\mathbf{H}_{ij} = 1$ if vertex $v_{i}$ is included in the hyperedge $e_j$ . This structure captures complex relationships, making hypergraphs valuable for modeling networked systems and group interactions. + +Hypergraphs can alternatively be described using a bipartite representation, where the vertex set $\mathcal{V}$ and hyperedge set $\mathcal{E}$ form distinct groups linked by the incidence matrix $\mathbf{H}$ (refer to Fig. 2). This representation results in a new graph $\mathcal{G}_B = (\mathcal{V},\mathcal{E},\mathcal{E}_B)$ , where $\mathcal{V}$ represents the original vertices of the hypergraph, and the elements in $\mathcal{E}$ correspond to hyperedges. The edges in $\mathcal{E}_B$ , denoted as $\epsilon_{ve} = (\nu_v,\nu_e)$ exist if $\mathbf{H}_{ve} = 1$ , with $\nu_{v}\in \mathcal{V}$ and $\nu_{e}\in \mathcal{E}$ , linking $\mathcal{V}$ and $\mathcal{E}$ . + +In the bipartite graph $\mathcal{G}_B$ , the corresponding adjacency matrix can be simplified as $\hat{\mathbf{A}} = \mathbf{H}$ for $\mathcal{E} \to \mathcal{V}$ interactions, and $\hat{\mathbf{A}} = \mathbf{H}^T$ for $\mathcal{V} \to \mathcal{E}$ . Drawing on principles similar to those in ViHGNN [17], the edge weight matrix $\mathbf{A}$ can be interpreted as fuzzy membership weights, enabling graded interactions across GNN layers. Complementing this setup, the feature matrices are split into $\mathbf{X}^{(V)} \in \mathbb{R}^{|\mathcal{V}| \times d_v}$ and $\mathbf{X}^{(E)} \in \mathbb{R}^{|\mathcal{E}| \times d_e}$ , along with their correspond adjacency feature matrices $\mathbf{X}_{\mathrm{adj}}^{(V)}$ and $\mathbf{X}_{\mathrm{adj}}^{(E)}$ , mirroring traditional GNNs. + +# 3.2. Imposing Hierarchical Structure in Images + +To enhance the capability of hypergraphs in image analysis, we draw inspiration from the register tokens introduced in [8], which act to summarize information that otherwise + +manifests as noise in areas of low visual significance. Similarly, this work integrates virtual vertices $(v\mathcal{V})$ , alongside typical image patch vertices $(i\mathcal{V})$ , and introduces virtual hyperedges $(v\mathcal{E})$ , alongside primary hyperedges $(p\mathcal{E})$ , providing levels of semantic feature aggregation and relational abstraction. Notably, these virtual elements, illustrated in Fig. 1e, do not correspond to specific image patches; instead, they are learned embeddings used for semantic summarization and capturing high-level abstract information. + +Our proposed hypergraph, constructed from image $I$ as $\mathcal{G}_B(I)$ , integrates primary and virtual sets, forming $\mathcal{V} = i\mathcal{V} \cup v\mathcal{V}$ and $\mathcal{E} = p\mathcal{E} \cup v\mathcal{E}$ , with statically masked communication pathways to enforce a hierarchical structure. Primary hyperedges $(p\mathcal{E})$ interact with all vertices to support unrestricted semantic aggregation, whereas virtual hyperedges $(v\mathcal{E})$ , designated for class predictions, connect solely with virtual vertices $(v\mathcal{V})$ . These restrictions separate visual and abstract information, thereby producing a graph structure suitable for use in downstream applications. + +# 4. A Hypergraph Vision Transformer + +The Hypergraph Vision Transformer (HgVT) adapts the architecture of standard Vision Transformers by incorporating bipartite hypergraph features for enhanced image analysis capabilities. Like Vision Transformers, HgVT begins with a patch embedding layer, followed by an isotropic stack of $L \times \mathrm{HgVT}$ blocks, culminating in feature pooling and a classifier head. The bipartite hypergraph is represented by four principal feature matrices $-\mathbf{X}^{(V)}$ , $\mathbf{X}_{\mathrm{adj}}^{(V)}$ , $\mathbf{X}^{(E)}$ , and $\mathbf{X}_{\mathrm{adj}}^{(E)}$ which are updated iteratively and in an interleaved fashion within each block. Each block constructs a new adjacency matrix $\mathbf{A}$ from $\mathbf{X}_{\mathrm{adj}}^{(V)}$ and $\mathbf{X}_{\mathrm{adj}}^{(E)}$ , enabling flexible adjustments to the hypergraph structure. As illustrated in Fig. 3, this modular process allows for the continuous integration and processing of these matrices within each HgVT block. + +# 4.1. Hyperedges as Communication Pools + +Each HgVT block processes both vertex and edge information, refining them from the previous block based on a newly constructed graph structure. Initially, adjacency mask computation (detailed in the next section) determines the connectivity for the subsequent processing steps within each block, dynamically adjusting to the updated feature matrices from the previous block. Three attention layers – vertex self-attention, edge aggregate attention, and edge distribution attention – operate sequentially to enhance feature integration and facilitate effective message passing along the hyperedges formed in the adjacency computation step. Finally, separate feed-forward networks process vertex and edge features independently, ensuring specialized treatment for the two distinct sets within the bipartite hypergraph, preserving the unique properties of each set. Operational details of these components are further described in Appendix A. + +![](images/14b90bfc81895302ca98b111966fca9f7608cc4dfe160835ee33c206808492ca.jpg) +Figure 3. HgVT Architecture Diagram, composed of stacked HgVT blocks with adjacency matrix $\mathbf{A}$ , vertex features $\mathbf{X}^{(V)}$ , and hyperedge features $\mathbf{X}^{(E)}$ . Edge attention flow is shown with gray arrows; input norms and residual adds are omitted for clarity. + +Hypergraph Feature Processing. Within each HgVT block, two distinct point-wise feed-forward networks (FFNs) independently process vertex and hyperedge features, aligning with the bipartite structure of the hypergraph. Each FFN integrates both the element features and their corresponding adjacency features through a fully connected layer, improving the model's ability to synthesize relationships. Processing both feature types within the same FFN layer allows adjacency information to be handled directly, bypassing the need for graph-based message passing and improving computational efficiency. Moreover, parameter overhead can be reduced by optionally tying edge and vertex FFN weights. + +![](images/13b6d70a4bb19ec0622873859df6a17fc3c5cf7349d6cf9dd7b72b6f9789f8a8.jpg) +(a) Communication Pool. + +![](images/dc882fb198a1561d69ebe306d286227f612cd16894e4067a343795c9c6459166.jpg) +(b) Attention Sparsity. +Figure 4. (a) Hyperedge Communication Pool Flow with edges $\mathcal{E}$ and member vertices $\nu_{e}$ ; (b) Attention Sparsity (Mean and std) for HgVT-S on the ImageNet-1k Validation set. + +Hyperedges as Communication Pools. Hypergraph GNNs typically employ a gather $\rightarrow$ scatter mechanism for processing vertex-hyperedge interactions, whereas HgVT reconceptualizes hyperedges as communication pools that facilitate information flow among vertices and their associated hyperedges. Specifically, vertex self-attention manages vertex-to-vertex $(\mathcal{V}\rightarrow \mathcal{V})$ interactions within hyperedges, edge aggregate attention orchestrates the flow from vertices to hyperedges $(\mathcal{V}\rightarrow \mathcal{E})$ , and edge distribution attention handles the reverse, from hyperedges back to vertices $(\mathcal{E}\rightarrow \mathcal{V})$ . By segmenting the attention operations, HgVT efficiently approximates an all-to-all feature transfer within hyperedges, as illustrated in Fig. 4a, significantly reducing the quadratic complexity associated with full attention mechanisms. + +Sparse and Fuzzy Attention. Building upon the dynamic communication pools concept, HgVT employs both sparse and fuzzy attention mechanisms to further optimize computational efficiency. Vertex self-attention is applied selectively to pairs of vertices connected by common hyperedges, as defined by the adjacency matrix $\hat{\mathbf{A}}$ , resulting in a sparse attention pattern. As sparsity increases with network depth - demonstrated in Fig. 4b - computational load decreases, while still maintaining compatibility with dense attention during training. Conversely, the edge aggregate and distribution attention mechanisms utilize cross-attention between the vertex and edge feature matrices, $\mathbf{X}^{(V)}$ and $\mathbf{X}^{(E)}$ , modulated by the soft adjacency matrix $\mathbf{A}$ . This modulation, akin to Fuzzy C-Means in ViHGNNs [17], adjusts attention logits based on soft memberships to the individual hyperedges, dynamically adapting to the hypergraph structure and providing a mechanism for gradient flow into the adjacency matrix generation. Furthermore, by thresholding the soft adjacency matrix during inference, the edge attention mechanisms can be converted into a sparse cross-attention mechanism, thereby reducing computational overhead. + +# 4.2. Dynamic Adjacency Formation + +HgVT dynamically establishes its hypergraph structure to adapt to the varying semantic and spatial structures of different image inputs. It employs cosine similarity, akin to query-key interactions in attention mechanisms, to evaluate the alignment between vertex and hyperedge adjacency features. This approach allows hyperedges to "query" vertices for relevant features, providing a scale-invariant assessment that emphasizes the directionality of embedding vectors. The cosine similarity is subsequently transformed into adjacency membership using a sharpened sigmoid function: + +$$ +\mathbf {A} = \sigma \left(\boldsymbol {\alpha} \cdot \tilde {\mathbf {X}} _ {\mathrm {a d j}} ^ {(V)} \left[ \tilde {\mathbf {X}} _ {\mathrm {a d j}} ^ {(E)} \right] ^ {T}\right), \tilde {\mathbf {X}} _ {\mathrm {a d j}} ^ {(*)} = \frac {\mathbf {X} _ {\mathrm {a d j}} ^ {(*)}}{| | \mathbf {X} _ {\mathrm {a d j}} ^ {(*)} | | _ {2} + \epsilon} (1) +$$ + +Here, $\sigma$ denotes the sigmoid function and $\alpha = 4$ is a sharpening factor, which pushes values away from zero to estab + +lish binary-like membership values in matrix $\mathbf{A}$ . This soft adjacency matrix is further thresholded to create the hard adjacency matrix $\hat{\mathbf{A}} = [\mathbf{A} > 0.5]$ , which provides binary memberships to facilitate sparse attention masking. + +# 4.3. Architecture Scaling + +Table 1. Scaling variants of our HgVT architecture. All models are trained at $224\mathrm{x}224$ resolution, except lite variant (HgVT-Lt), trained at $160\mathrm{x}160$ . Showing count for vertices $(i\mathcal{V},v\mathcal{V})$ , hyperedges $(p\mathcal{E}, v\mathcal{E})$ , dim for adj. $(d_a)$ and features $(d_f)$ , depth $(L)$ , and heads $(h)$ . + +
Model|iV||vV||pE||vE|df+daLhParams FLOPS
HgVT-Lt10012326128+641246.8M0.92B
HgVT-Mi19616508128+64945.8M1.39B
HgVT-Ti19616508128+641247.7M1.80B
HgVT-S19616508224+9614723M5.48B
+ +HgVT adopts a hybrid scaling strategy inspired by DeiT [53] and ViG [16], with Tab. 1 detailing the transformer scaling hyperparameters and the vertex and edge type allocations, including fixed capacities for non-image vertices $(i\mathcal{V})$ as proposed by ViHGNN [17]. Our scaling approach aims to balance parameter count and FLOPs with the two main GNN baselines, ViG and ViHGNN. We also introduce a Mini variant (HgVT-Mi) to match DeiT-Ti's computational profile, which we achieve by repeating the final block twice for an effective depth of 9; see Appendix H for details. Finally, we include a Ti-Lite variant (HgVT-Lt) for more efficient ablation studies under a constrained training budget. + +# 5. Enforcing Semantic Structure + +Feature matrices for virtual vertices and hyperedges, lacking direct input-based initialization, risk converging to homogeneous solutions and collapsed representations. Additionally, the dynamic adjacency calculation alone fails to naturally promote semantic grouping, in contrast to clustering-based approaches commonly used in vision GNNs. To address these issues, we introduce diversity regularization, which enforces orthogonal embeddings, and population regularization, which encourages a structured, sparse hypergraph. For enhanced semantic differentiation of virtual hyperedge features, we further incorporate an expert-based pooling strategy as a more robust alternative to mean pooling. + +# 5.1. Diversity-Driven Feature Differentiation + +To prevent homogenization of learned feature matrices and to encourage distinct, semantically rich embeddings, we implement a diversity-driven regularization approach. This method, designed to maintain maximum orthogonality among the embeddings of virtual vertices and hyperedges, penalizes the absolute value of the cosine similarity between different feature vectors, aiming for values close to zero. By using normalized embeddings and masking off diagonal elements to preserve self-similarity, the approach prevents + +the model from converging to homogeneous solutions or driving individual vectors towards zero magnitude. We then individually penalize $v \mathcal{V}$ , $\mathcal{E}$ , and their adjacency features. + +$$ +\mathrm {D} _ {\mathrm {L}} (\mathbf {X}) = \frac {1}{2} \sum_ {i j} (1 - \delta_ {i j}) \cdot \left| \tilde {\mathbf {X}} \tilde {\mathbf {X}} ^ {T} \right| _ {i j}, \tilde {\mathbf {X}} = \frac {\mathbf {X}}{| | \mathbf {X} | | _ {2} + \epsilon} \quad (2) +$$ + +$$ +\mathcal {L} _ {\mathrm {D I V}} = \sum_ {x} \mathrm {D} _ {\mathrm {L}} (x), x \in \{\mathbf {X} ^ {(: v V)}, \mathbf {X} _ {\mathrm {a d j}} ^ {(: v V)}, \mathbf {X} ^ {(E)}, \mathbf {X} _ {\mathrm {a d j}} ^ {(E)} \} (3) +$$ + +Where $(:vV)$ represents the subset of $\mathcal{V}$ containing only the virtual nodes, $\delta_{ij}$ is the Kronecker delta function, ensuring that self-similarity is not penalized, and $|\cdot|_{ij}$ denotes the element-wise absolute value, applied to calculate the penalty for non-orthogonal relationships between embeddings. + +# 5.2. Population Regularization: Learned Sparsity + +Unlike clustering methods such as KNN or Fuzzy C-Means, which impose fixed cluster sizes, our model's dynamic adjacency calculation enables flexible and self-adjusting hyperedge populations. To prevent the associated risks of overly sparse or densely connected hypergraphs, we introduce population regularization. This method applies penalties based on the computed soft membership density of each hyperedge derived from the soft adjacency matrix $\mathbf{A}$ , ensuring each maintains a vertex population within appropriate bounds to avoid overgeneralization and preserve hypergraph integrity. + +$$ +P _ {j} = 2 \cdot \sum_ {i} \max (A _ {i j} - 0. 5, 0) \tag {4} +$$ + +$$ +\mathcal {L} _ {\mathrm {P O P}} = \sum_ {j} \max (P _ {j} - \beta , 0) + \max (\gamma - P _ {j}, 0) \tag {5} +$$ + +Here, $P_{j}$ represents a soft density estimate of vertex connections for the $j^{th}$ hyperedge, only considering non-zero entries of $\hat{\mathbf{A}}$ . The hyperparameters $\beta$ and $\gamma$ set the upper and lower density limits, ensuring that hyperedges maintain an optimal balance of connections. Penalties are applied if $P_{j}$ either exceeds $\beta$ or falls below $\gamma$ , maintaining the desired sparsity and ensuring the structural efficacy of the hypergraph. + +# 5.3. Expert Pooling for Semantic Specialization + +To effectively combine features from multiple virtual hyperedges for classification, our approach utilizes a method akin to expert-choice, where each virtual hyperedge acts as an "expert" generating a confidence score. Unlike mean pooling, which risks collapsing distinct features into an average that may dilute individual contributions, this strategy encourages virtual hyperedges to develop unique, semantically meaningful representations. The normalized confidence scores, $P(e)$ , determine the relevance of each hyperedge $e$ 's contribution to the classification task. + +$$ +P (e) = \operatorname {s o f t m a x} \left(\mathbf {X} _ {e} ^ {(v E)} \mathbf {W} _ {e} + b _ {e}\right) \tag {6} +$$ + +Here, $(:vE)$ denotes the subset of $(E)$ containing only the virtual hyperedges, and the softmax is computed across the + +expert gate set $\{e\}$ after projection by $\mathbf{W} \in \mathbb{R}^{d \times |vE|}$ . During training, $P(e)$ guides the weighted averaging of hyperedge features before subsequent class prediction. We further apply a density loss function [3, 14] to prevent underutilization of any single virtual hyperedge, complemented by a cross-entropy term with label smoothing to increase expert confidence. During inference, top-k routing combines the most relevant hyperedges based on their confidence scores. + +# 6. Empirical Evaluation and Performance + +This section evaluates the Hypergraph Vision Transformer using two model configurations from Tab. 1: HgVT-Ti-Lite for targeted ablation studies and scaled variants for benchmarking against comparable classifiers. We apply standard augmentations from DeiT [53] using the Timm library [58], including RandAugment [7], Mixup [64], Cutmix [60], Random Erasing [66], and Repeated Augment [21]. + +Datasets. For classification in computer vision, we follow standard practices and use the ImageNet-1k dataset [10] at a resolution of $224 \times 224$ pixels for scaled model evaluations. For ablation studies, we employ ImageNet-100 [51], a 100-class subset of ImageNet-1k with images scaled to $160 \times 160$ pixels. This selection provides a computationally manageable dataset while maintaining sufficient class variation and larger image sizes compared to datasets like CIFAR-100 (32x32 pixels)[27]. Nevertheless, we find CIFAR-100 useful for assessing the effects of regularization on the hypergraph structure, as detailed in Appendix H. + +Training Hyperparameters. Consistent with DeiT, we use the AdamW optimizer with a weight decay of 0.05. Training is conducted on the ImageNet-1k dataset with a batch size of 1024 for 300 epochs following DeiT. For ablations, we train on ImageNet-100 with a batch size of 512 for with a shorter duration of 200 epochs as proposed by [31]. Learning rates follow a cosine-annealing schedule peaking at 1e-3 for both datasets following scaling from DeiT. Furthermore, we omit the use of Exponential Moving Average (EMA) due to its minimal performance improvement (0.1% in DeiT) relative to its overhead per training step. All models were trained using PyTorch on NVIDIA A6000 GPUs; see Appendix I. + +Evaluation Metrics. Following standard protocols, we measure the Top-1 class prediction accuracy to assess overall performance. Additionally, we take advantage of the learned graph structure (extracted from the last layer) on each image, and measure: Hyperedge Entropy (HE), Intra-Cluster Similarity (ICS), Inter-Cluster Distance (ICD), and Silhouette Score (SIL) [45]; further details on structure in Appendix C. + +# 6.1. Evaluation on ImageNet + +Tab. 2 presents the ImageNet-1k Top-1 accuracy of HgVT, benchmarked against isotropic models. We limit evaluation + +Table 2. ImageNet-1k results for HgVT and other isotropic networks. $\star$ CNN, $\clubsuit$ Transformer, $\star$ GNN, $\blacksquare$ HGNN, and $\triangle$ HgVT. + +
ModelParamsFLOPsImNet Top-1ReAL Top-1V2 Top-1
*ResMLP-S12 conv3x3 [52]16.7M3.2B77.084.065.5
*ConvMixer-768/32 [55]21.1M20.9B80.2--
*ConvMixer-1536/20 [55]51.6M51.1B81.4--
*DINOv1-S [2]21.7M4.6B77.0--
*ViT-B/16 [12]86.4M55.5B77.983.6-
*DeiT-Ti [53]5.7M1.3B72.280.160.4
*DeiT-S [53]22.1M4.6B79.885.768.5
*DeiT-B [53]86.4M17.6B81.886.771.5
*ViG-Ti [16]7.1M1.3B73.9--
*ViG-S [16]22.7M4.5B80.4--
*ViG-B [16]86.8M17.7B82.3--
■ViHGNN-Ti [17]8.2M1.8B74.3--
■ViHGNN-S [17]23.2M5.6B81.5--
■ViHGNN-B [17]88.1M19.4B82.9--
▲HgVT-Mi (ours)5.8M1.4B74.481.862.7
▲HgVT-Ti (ours)7.7M1.8B76.283.264.3
▲HgVT-S (ours)22.9M5.5B81.286.770.1
+ +to isotropic models due to the complexity of down-sampling virtual tokens, lacking direct spatial correspondence, needed for pyramidal structures. As a result, we also exclude other pyramidal models from our comparison as they typically outperform isotropic models within the same architecture family [16, 17], making the comparison less meaningful. + +HgVT-Ti demonstrates a clear advantage, surpassing ViHGNN-Ti by $1.9\%$ in accuracy, with the smaller HgVT-Mi model matching ViHGNN-Ti with only $70\%$ of the parameters. At the S-scale, we reduce HgVT's depth relative to ViHGNN to accommodate hyperedge attention and adjacency feature overhead, aiming for comparable parameter counts; as a result, HgVT-S only matches ViHGNN-S. Despite this limitation, HgVT-S matches DeiT-B's accuracy on ImageNet Real [1] and achieves competitive performance on ImageNet V2 [43], at nearly a quarter of DeiT-B's size. + +Overall, HgVT achieves strong performance on classification, demonstrating the benefits of integrating hypergraph structures within a vision transformer framework. Additionally, computational evaluation in Appendix B shows a consistent $\geq 1.4\times$ speedup over ViG, with adjacency computation exhibiting a $10\times$ improvement over ViG's KNN-based method. These improvements further highlight HgVT's potential for greater efficiency over prior Vision GNNs. + +# 6.2. Ablation Studies + +We conducted a series of ablations on the ImageNet-100 dataset using the HgVT-Lt model, reporting Top-1 classification accuracy alongside mean hyperedge entropy and silhouette scores to assess the quality of the hypergraph's structure. Notably, we observe a weak anti-correlation between the graph quality metrics and Top-1 accuracy (see Appendix H), indicating opposing objectives. Overall, the ablations are grouped into three categories: regularization, architecture, and pooling methods, with results in Tab. 3. + +Table 3. ImageNet-100 ablations with HgVT-Lt. Indicating used (✓) or not used $(\mathcal{X})$ , and pooling methods: Average (A), Image $(I)$ , Expert $(E)$ , Expert $^+$ Image $(EI)$ , and $EI$ dropping I $(DI)$ + +
Ablation on ↓CLS Dropout Stoch. Decay Diversity PopulationTied FFN Xadj = X dF Mult. Pooled OpEdge EntropyTop-1 Acc.Params (M)
X✓✓✓✓1EI3.320.78084.36
none: HgVT-Lt✓✓✓✓✓1E3.130.75182.23
RegularizationXX✓✓✓1E3.120.74581.89
X✓X✓✓1E1.990.72380.79
X✓✓X✓1E3.890.63981.79
X✓XX✓1E3.580.61081.99
ArchitectureX✓✓✓X1E3.090.74182.89
X✓✓✓X1E2.270.80878.62
X✓✓✓✓1E2.120.78076.95
X✓✓✓✓1.5E2.050.77077.46
PoolingX✓✓✓✓1A3.070.74782.06
X✓✓✓✓1I2.930.76084.08
X✓✓✓✓1E3.130.74982.52
X✓✓✓✓1DI3.320.78080.94
+ +As shown in Tab. 3, the regularization ablations demonstrate that stochastic path dropout decay [22] improves both Top-1 accuracy and silhouette scores, consistent with ViG and ViHGNN [16, 17]. Omitting Class dropout also boosts accuracy, aligning with DeiT [53]. Furthermore, our proposed diversity and population regularization are essential for preserving graph structure; removing diversity leads to partial representation collapse, while removing population regularization results in near-zero sparsity, effectively turning HgVT into a ViT with increased network complexity. + +In the architecture ablations, untying the FFN improves accuracy but significantly increases parameter count, making it an unfavorable tradeoff. Tying adjacency and embedding features $(\mathbf{X}_{\mathrm{adj}} = \mathbf{X})$ reduces parameters and FLOPs but degrades performance, and while untying the FFN or increasing feature dimensionality partially mitigates this, the parameter increase remains suboptimal. This indicates that adjacency and embedding features $(\mathbf{X}_{\mathrm{adj}}^{(*)}$ and $\mathbf{X}^{(*)})$ are similar, yet require dedicated feature spaces to avoid crowding. + +For pooling methods, expert edge pooling outperforms average edge pooling in accuracy, while image pooling achieves the highest accuracy at the cost of degraded graph structure. Combining image and expert pooling recovers lost structure and improves accuracy, with each input focusing on different semantic levels (see Appendix D). Additionally, dropping the pooled image embedding prior to the classifier head maintains moderate performance, indicating that both paths meaningfully contribute to the final prediction. + +Impact of Vertex Self-Attention and Patch Embedding. We evaluated the impact of patch embedding methods and vertex self-attention, comparing a convolutional stem (Conv2D-BN-GELU layers [16, 17]) and a simpler patch projection (pixel-shuffled patches with affine projection [2, 12]), across various pooling strategies (average, image, and ex + +Table 4. Top-1 accuracy of HgVT-Lt on ImageNet-100 with $(\checkmark)$ or without $(X)$ vertex self-attention. Contrasting pooling operations and patch embedding versions: Conv. Stem or Patch Projection. + +
Pooling Op. →AverageImageExpert
Vertex SA →X✓X✓X✓
Conv. Stem78.0282.0682.0584.0878.8782.52
Patch Project64.3072.7670.7676.1762.6271.43
+ +pert), as shown in Tab. 4. The patch projection consistently underperforms the convolutional stem, likely due to the model's small size limiting its effectiveness. Omitting vertex self-attention leads to further degradation, especially without the convolutional stem, suggesting it is crucial for effectively separating features in low-dimensional space. PCA shows that 71/128 channels are needed to explain $95\%$ of the variance for the convolutional stem, compared to 19/128 for the patch projection, indicating the richer representation captured by the convolutional stem. Notably, image pooling shows the least degradation, likely due to its more direct gradient flow compared to the edge pooling methods. + +# 6.3. Pooling Methods and Graph Structure + +Table 5. Impact of graph structure on pooling operation for HgVT-Lt on ImageNet-100. Measuring graph metrics for image vertices $(i\mathcal{V})$ and all vertices $(\mathcal{V})$ , along with features from DINOv2. + +
Pooling Op. →ImageExpertImg. & Expert
Feature Model ↓HEICSICDHEICSICDHEICSICD
HgVT-Lt (iV)3.030.500.313.240.430.373.230.420.36
HgVT-Lt (V)3.200.250.723.390.360.453.390.280.58
DINO2-S (iV)3.040.840.083.250.850.053.250.830.06
DINO2-G (iV)3.040.690.153.250.680.123.250.680.13
+ +To evaluate the impact of pooling methods on graph structure, we measured HE, ICS, and ICD using the ImageNet-100 validation set with three strategies: image pooling, expert pooling, and a combined image + expert pooling approach. Metrics used either the image vertex subset $(i\mathcal{V})$ or the full vertex set $(\mathcal{V})$ , with features derived from DINOv2 (S and G) [38] and HgVT-Lt using HgVT-Lt's adjacency matrix (A). Results in Tab. 5 show that while all methods achieve similar graph quality using $i\mathcal{V}$ , image pooling slightly improves similarity. However, including all vertices $(\mathcal{V})$ consistently increases ICD and entropy while reducing ICS, indicating decreased graph coherence. This effect is more severe with image pooling methods, suggesting that virtual vertices $(v\mathcal{V})$ act as noisy elements, rather than summarization points. + +Comparing DINOv2 models, all pooling methods align more closely with DINOv2-G, where achieving a balance between ICS and ICD is preferable to maximizing either individually. This trend, along with consistent HE, suggests a focus on higher-level detail, irrespective of the smaller HgVT model size or pooling method. Image pooling shows slightly stronger alignment with both DINO models, indicating that + +both high- and low-level semantics are encoded within a single feature space, unlike methods that can use edge channels for high-level concepts. Notably, all expert pooling methods exhibit an emergent macro-class prediction behavior, where each virtual edge $(v\mathcal{E})$ consistently captures broader taxonomic groups (e.g., dogs, birds). Further representation and macro-class analysis are provided in Appendices D and J. + +# 6.4. Performance on Image Retrieval + +To evaluate the capability of HgVT in capturing semantic structures, we perform image retrieval experiments comparing four methods: pooling similarity (PS), volumetric similarity (VS), adaptive pooling similarity (APS), and adaptive volumetric similarity (AVS). PS ranks the pooled embeddings by cosine similarity (vector search), while the other methods enhance retrieval by leveraging graph structure. Volumetric similarity calculates ellipsoid overlap using an approximate Mahalanobis distance, with pruned hyperedges defining the distribution spread around the pooled embedding (centroid). Adaptive methods further refine retrieval via a graph similarity measure on the pruned hyperedges, re-ranking from a shortlist of $R = 100$ . Computational efficiency in adaptive retrieval is ensured through centroid hash binning and limiting comparisons to the top- $C = 4$ most significant query hyperedges, resulting in a complexity of $\mathcal{O}(RC)$ . Notably, we prune to 12 hyperedges and use 10 centroid bins. Additional details provided in Appendix G. + +Table 6. Image Retrieval on ImageNet-1k with proposed search methods: PS, APS, VS, and AVS. Reporting Top-1 accuracy, mAP@10 (\%), and 1-NN-hit@10 (\%) for pooling methods. + +
mAP@10 (%)
ModelTop-1PoolingPSAPSVSAVS
MRL-128 [29]70.52Image64.9465.20--
MRL-256 [29]70.62Image65.0465.20--
HgVT-Ti (ours)76.18Expert70.5669.5970.5369.40
HgVT-Ti (ours)76.18Im&Ex73.2369.5973.0869.49
1-NN-hit@10 (%)
HgVT-Ti (ours)76.18Expert21.2219.1021.2519.56
HgVT-Ti (ours)76.18Im&Ex25.1319.1025.2219.17
+ +ImageNet Retrieval. We evaluate retrieval performance on the ImageNet-1K dataset to assess HgVT's ability to capture semantic relationships and compare four retrieval methods: PS, VS, APS, and AVS. The primary metric is mAP@10, with MRL [29] serving as the baseline due to its adaptive re-ranking approach. We also report the 1-NN-CLIP-L hitrate@10, which measures how often the top-1 result ranked by CLIP-L [42] appears in the top-10 retrieved results, offering additional insight into semantic alignment. Results in Tab. 6 show that HgVT-Ti surpasses MRL by over $8\%$ in retrieval performance, despite being significantly smaller than MRL (ResNet-50) and using a comparably compact + +embedding size ( $d = 2 \times 128$ ). Among our methods, PS and VS achieve similar results, while APS and AVS underperform, likely due to their focus on exact-feature similarity and ambiguous-class features, which limits alignment with the high-level semantics required by ImageNet's diverse dataset. + +Table 7. Image Retrieval on revisited Oxford and Paris; reporting mAP $(\%)$ . Showing training set and method used. *mAP@100. + +
Dataset →ROxfordRParis
ModelTrain SetMethodMHMH
ALEX+GeM [44]ImNet-1kPS33.810.452.726.0
RN101+R-MAC [44]ImNet-1kPS49.818.574.052.1
DINOv1-S/16 [2]ImNet-1kPS41.813.763.134.4
DINOv1-S/16 [2]GLDv2PS51.524.375.351.6
HgVT-Ti (ours)ImNet-1kVS26.810.555.428.7
VS*25.89.065.628.0
AVS*27.06.864.126.7
HgVT-S (ours)ImNet-1kVS28.012.156.731.1
VS*26.310.265.028.4
AVS*27.410.365.027.1
+ +Oxford and Paris Retrieval. To evaluate image retrieval performance beyond simple class retrieval, we use the Oxford and Paris Revisited datasets [39, 41], which provide three splits of increasing difficulty (Easy, Medium, and Hard) for query/database pairs. We report Mean Average Precision (mAP) for the Medium (M) and Hard (H) splits, using mAP@100 for AVS based on short-list ranking, while full mAP and mAP@100 are provided for VS as a baseline comparison. Results, shown in Tab. 7, indicate that HgVT achieves competitive performance with similarly sized feature extractors, though performance on ROxford-M lags. This shortfall may stem from subtle landmark differences better captured by multi-scale Conv-Nets and self-supervised learning, compared to the more salient focus driven by HgVT's classifier training. However, AVS outperforms VS on ROxford-M, demonstrating its ability to uncover finer feature similarities within the hypergraph structure. + +# 7. Conclusion and Future Directions + +In this work, we introduced the Hypergraph Vision Transformer (HgVT), which combines hypergraph structures with vision transformers. Our methods, including diversity and population regularization, along with expert edge pooling, enhance semantic representation and efficiency by enabling dynamic hypergraph construction. Although challenging to tune, these methods avoid reliance on traditional clustering techniques, reflecting a tradeoff between complexity and efficiency. Empirical results show that HgVT achieves strong performance on both image classification and retrieval, positioning HgVT as a competitive framework for semantic vision tasks. Future work will focus on exploring the scalability of hypergraph structures and integrating self-supervised learning to further improve adaptability and better decouple saliency from semantic vision graph generation. + +# References + +[1] Lucas Beyer, Olivier J. Henaff, Alexander Kolesnikov, Xiaohua Zhai, and Aaron van den Oord. Are we done with imageNet? arXiv preprint arXiv:2006.07159, abs/2006.07159, 2020. 6 +[2] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jegou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9630-9640, 2021. 1, 2, 6, 7, 8, 4, 5 +[3] Tianlong Chen, Zhenyu Zhang, Ajay Kumar Jaiswal, Shiwei Liu, and Zhangyang Wang. Sparse moe as the new dropout: Scaling dense and self-slimmable transformers. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. 6 +[4] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 15 +[5] Norman Cliff. Dominance statistics: Ordinal analyses to answer ordinal questions. Psychological Bulletin, 114:494-509, 1993. 8 +[6] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 14 +[7] Ekin Dogus Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V. Le. Randaugment: Practical automated data augmentation with a reduced search space. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 3008-3017, 2019. 6 +[8] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. In *The Twelfth International Conference on Learning Representations*, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024. 3 +[9] Dmitry Demidov, M.H. Sharif, Aliakbar Abdurahimov, Hisham Cholakkal, and Fahad Shahbaz Khan. Salient mask-guided vision transformer for fine-grained classification. In VISIGRAPP, 2023. 1 +[10] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. In CVPR09, 2009. 6 +[11] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In North American Chapter of the Association for Computational Linguistics, 2019. 1 +[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at + +scale. In International Conference on Learning Representations, 2021. 1, 2, 6, 7, 4, 5 + +[13] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The PASCAL Visual Object Classes Challenge 2012 (VOC2012) Results. http://www.pascalnetwork.org/challenges/VOC/voc2012/workshop/index.html. 14 +[14] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. J. Mach. Learn. Res., 23:120:1-120:39, 2022. 6 +[15] Yue Gao, Meng Wang, Dacheng Tao, Rongrong Ji, and Qionghai Dai. 3-d object retrieval and recognition with hypergraph analysis. IEEE Transactions on Image Processing, 21(9): 4290-4303, 2012. 2 +[16] Kai Han, Yunhe Wang, Jianyuan Guo, Yehui Tang, and Enhua Wu. Vision GNN: An image is worth graph of nodes. In Advances in Neural Information Processing Systems, 2022. 1, 2, 5, 6, 7, 4 +[17] Yan Han, Peihao Wang, Souvik Kundu, Ying Ding, and Zhangyang Wang. Vision hgnn: An image is more than a graph of nodes. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 19878-19888, 2023. 1, 2, 3, 4, 5, 6, 7, 27 +[18] Yunusa Haruna, Shiyin Qin, Abdulrahman Hamman Adama Chukkol, Abdulganiyu Abdu Yusuf, Isah Bello, and Adamu Lawan. Exploring the synergies of hybrid cnns and vits architectures for computer vision: A survey. arXiv preprint arXiv:2402.02941, abs/2402.02941, 2024. 1 +[19] Ali Hassani, Steven Walton, Jiachen Li, Shen Li, and Humphrey Shi. Neighborhood attention transformer. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2 +[20] Kaiming He, X. Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2015. 1, 15 +[21] Elad Hoffer, Tal Ben-Nun, Itay Hubara, Niv Giladi, Torsten Hoefler, and Daniel Soudry. Augment your batch: Improving generalization through instance repetition. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8126-8135, 2020. 6 +[22] Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Weinberger. Deep networks with stochastic depth, 2016. 7 +[23] Yuchi Huang, Qingshan Liu, and Dimitris Metaxas. ]video object segmentation by hypergraph cut. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 1738-1745, 2009. 2 +[24] Yuchi Huang, Qingshan Liu, Shaoting Zhang, and Dimitris N. Metaxas. Image retrieval via probabilistic hypergraph ranking. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 3376-3383, 2010. 2 +[25] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. 15 + +[26] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3992-4003, 2023. 1 +[27] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, 2009. 6 +[28] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2012. 1 +[29] Aditya Kusupati, Gantavya Bhatt, Aniket Rege, Matthew Wallingford, Aditya Sinha, Vivek Ramanujan, William Howard-Snyder, Kaifeng Chen, Sham M. Kakade, Prateek Jain, and Ali Farhadi. Matryoshka representation learning. In Neural Information Processing Systems, 2022. 1, 8 +[30] Y. Lecun, L. Bottou, Y. Bengio, and P. Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 1 +[31] Sangjun Lee, Inwoo Hwang, Gi-Cheon Kang, and Byoung-Tak Zhang. Improving robustness to texture bias via shape-focused augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 4323-4331, 2022. 6 +[32] Benjamin Lefaudeau, Francisco Massa, Diana Liskovich, Wenhan Xiong, Vittorio Caggiano, Sean Naren, Min Xu, Jieru Hu, Marta Tintore, Susan Zhang, Patrick Labatut, Daniel Haziza, Luca Wehrstedt, Jeremy Reizenstein, and Grigory Sizov. xformers: A modular and hackable transformer modelling library. https://github.com/facebookresearch/xformers, 2022. 28 +[33] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9992-10002, 2021. 1, 2, 15 +[34] Leland McInnes, John Healy, Nathaniel Saul, and Lukas Großberger. Umap: Uniform manifold approximation and projection. Journal of Open Source Software, 3(29):861, 2018. 9 +[35] Mustafa Munir, William Avery, and Radu Marculescu. Mobilevig: Graph-based sparse attention for mobile vision applications. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 2211-2219, 2023. 1, 2 +[36] Mustafa Munir, William Avery, Md Mostafijur Rahman, and Radu Marculescu. Greedyvig: Dynamic axial graph construction for efficient vision gnns. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6118-6127, 2024. 1, 2 +[37] NVIDIA. Apex: A pytorch extension. https://github.com/NVIDIA/apex, 2020.28 +[38] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell + +Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINOv2: Learning robust visual features without supervision. Transactions on Machine Learning Research, 2024. 1, 2, 7, 6, 8, 14, 15 +[39] James Philbin, Ondrej Chum, Michael Isard, Josef Sivic, and Andrew Zisserman. Lost in quantization: Improving particular object retrieval in large scale image databases. In 2008 IEEE Conference on Computer Vision and Pattern Recognition, pages 1-8, 2008. 8 +[40] Joan Puigcerver, Carlos Riquelme Ruiz, Basil Mustafa, and Neil Houlsby. From sparse to soft mixtures of experts. In The Twelfth International Conference on Learning Representations, 2024. 1 +[41] Filip Radenovic, Ahmet Iscen, Giorgos Tolias, Yannis Avrithis, and Ondrej Chum. Revisiting oxford and paris: Large-scale image retrieval benchmarking. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5706-5715, 2018. 8 +[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 8, 6 +[43] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do imagenet classifiers generalize to imageNet? In International Conference on Machine Learning, 2019. 6 +[44] Jérôme Revaud, Jon Almazán, Rafael Sampaio de Rezende, and César Roberto de Souza. Learning with average precision: Training image retrieval with a listwise loss. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 5106-5115, 2019. 8 +[45] Peter J. Rousseeuw. Silhouettes: A graphical aid to the interpretation and validation of cluster analysis. Journal of Computational and Applied Mathematics, 20:53-65, 1987. 6, 7 +[46] Franco Scarselli, Marco Gori, Ah Chung Tsoi, Markus Hagenbuchner, and Gabriele Monfardini. The graph neural network model. IEEE Transactions on Neural Networks, 20(1):61-80, 2009. 2 +[47] Noam Shazeer. GLU variants improve transformer. arXiv preprint arXiv:2002.05202, abs/2002.05202, 2020. 4 +[48] Dai Shi. Transnext: Robust foveal visual perception for vision transformers. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17773-17783, 2023. 14, 15 +[49] Dai Shi. Transnext: Robust foveal visual perception for vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17773-17783, 2024. 2 +[50] Sakhinana Sagar Srinivas, Rajat Kumar Sarkar, Sreeja Gangasani, and Venkataramana Runkana. Vision HgNN: An electron-micrograph is worth hypergraph of hypernodes. arXiv preprint arXiv:2408.11351, abs/2408.11351, 2024. 2 + +[51] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive multiview coding. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XI 16, pages 776-794. Springer, 2020. 6 +[52] Hugo Touvron, Piotr Bojanowski, Mathilde Caron, Matthieu Cord, Alaaeldin El-Nouby, Edouard Grave, Armand Joulin, Gabriel Synnaeve, Jakob Verbeek, and Hervé Jégou. Resmlp: Feedforward networks for image classification with data-efficient training. arXiv preprint arXiv:2105.03404, abs/2105.03404, 2021. 6 +[53] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herve Jegou. Training data-efficient image transformers amp; distillation through attention. In Proceedings of the 38th International Conference on Machine Learning, pages 10347-10357. PMLR, 2021. 1, 2, 5, 6, 7, 4 +[54] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurélien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, abs/2302.13971, 2023. 1, 2 +[55] Asher Trockman and J. Zico Kolter. Patches are all you need? Trans. Mach. Learn. Res., 2023, 2023. 6 +[56] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 1 +[57] Petar Velicković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. Graph attention networks. In International Conference on Learning Representations, 2018. 4 +[58] Ross Wightman. Pytorch image models. https://github.com/rwrightman/pytorch-image-models, 2019.6, 28 +[59] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In European Conference on Computer Vision. Springer, 2018. 15 +[60] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Young Joon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6022-6031, 2019. 6 +[61] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 1204-1213. IEEE, 2022. 1 +[62] Biao Zhang and Rico Sennrich. Root mean square layer normalization. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 2 +[63] Bowen Zhang, Zhi Tian, Quan Tang, Xiangxiang Chu, Xiaolin Wei, Chunhua Shen, and Yifan liu. Segvit: Semantic segmentation with plain vision transformers. In Advances in Neural Information Processing Systems, 2022. 1 + +[64] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In International Conference on Learning Representations, 2018. 6 +[65] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 6230-6239, 2016. 15 +[66] Zhun Zhong, Liang Zheng, Guoliang Kang, Shaozi Li, and Yi Yang. Random erasing data augmentation. AAAI, 34: 13001-13008, 2020. 6 +[67] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision, 127:302 - 321, 2016. 14 +[68] Lei Zhu, Xinjiang Wang, Zhanghan Ke, Wayne Zhang, and Rynson Lau. Bifformer: Vision transformer with bi-level routing attention. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2 \ No newline at end of file diff --git a/2025/Hypergraph Vision Transformers_ Images are More than Nodes, More than Edges/images.zip b/2025/Hypergraph Vision Transformers_ Images are More than Nodes, More than Edges/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..122945fcc4d6c164c923fbaa97fa71c25994f239 --- /dev/null +++ b/2025/Hypergraph Vision Transformers_ Images are More than Nodes, More than Edges/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:362b1a415756af7a8f5f23f5185080e092b879534b06127397f2a66674c84bc8 +size 470909 diff --git a/2025/Hypergraph Vision Transformers_ Images are More than Nodes, More than Edges/layout.json b/2025/Hypergraph Vision Transformers_ Images are More than Nodes, More than Edges/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f09c212b413d93c0b6c7eb3b8f55f8ba17123f1f --- /dev/null +++ b/2025/Hypergraph Vision Transformers_ Images are More than Nodes, More than Edges/layout.json @@ -0,0 +1,10808 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 56, + 103, + 555, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 103, + 555, + 121 + ], + "spans": [ + { + "bbox": [ + 56, + 103, + 555, + 121 + ], + "type": "text", + "content": "Hypergraph Vision Transformers: Images are More than Nodes, More than Edges" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 251, + 144, + 358, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 251, + 144, + 358, + 185 + ], + "spans": [ + { + "bbox": [ + 251, + 144, + 358, + 185 + ], + "type": "text", + "content": "Joshua Fixelle \nUniversity of Virginia \njf9fk@virginia.edu" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 238, + 297, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 238, + 297, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 297, + 467 + ], + "type": "text", + "content": "Recent advancements in computer vision have highlighted the scalability of Vision Transformers (ViTs) across various tasks, yet challenges remain in balancing adaptability, computational efficiency, and the ability to model higher-order relationships. Vision Graph Neural Networks (ViGs) offer an alternative by leveraging graph-based methodologies but are hindered by the computational bottlenecks of clustering algorithms used for edge generation. To address these issues, we propose the Hypergraph Vision Transformer (HgVT), which incorporates a hierarchical bipartite hypergraph structure into the vision transformer framework to capture higher-order semantic relationships while maintaining computational efficiency. HgVT leverages population and diversity regularization for dynamic hypergraph construction without clustering, and expert edge pooling to enhance semantic extraction and facilitate graph-based image retrieval. Empirical results demonstrate that HgVT achieves strong performance on image classification and retrieval, positioning it as an efficient framework for semantic-based vision tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 489, + 135, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 489, + 135, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 489, + 135, + 502 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 510, + 297, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 297, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 297, + 689 + ], + "type": "text", + "content": "Computer vision has recently transitioned from the historically dominant Convolutional Neural Networks (CNNs) [20, 28, 30] to the increasingly prominent Vision Transformers (ViTs), which have quickly embedded themselves as the new de facto standard [12, 38]. This shift reflects the broader success of transformers in natural language processing [11, 54, 56] and is driven by the remarkable scalability of ViTs across various tasks such as image classification [53, 61], semantic segmentation [26, 63], and image retrieval [2, 29]. While hybrid models like hierarchical attention and CNN-ViT methods [18, 19, 33] aim to balance computational load and flexibility, ViTs still struggle with focusing on salient features rather than comprehensive image understanding [2, 9, 12, 40]. This highlights the need for methods that improve both efficiency and semantic accuracy." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "Within the spectrum of novel architectures, Vision Graph Neural Networks (ViGs) [16] and Vision Hypergraph Neural" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 213, + 555, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 555, + 370 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 555, + 370 + ], + "type": "text", + "content": "Networks (ViHGNNs) [17] leverage graph-based topologies to advance image processing. Unlike CNNs, which harness locality and translation-invariance through densely connected pixel grids and repeated convolutions, both ViTs and ViGs represent images as sets of patches. In ViTs, each patch acts as a vertex within a maximally connected graph, creating semantically weak connections through self-attention. ViGs enhance this by using clustering algorithms to detect edge groupings and applying graph convolutions to these clusters, forming meaningful patch relationships. ViHGNNs extend these capabilities by employing hyperedges that capture complex, higher-order relationships, enriching understanding of the images. These methodologies are depicted in Figure 1." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 373, + 555, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 555, + 409 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 555, + 409 + ], + "type": "text", + "content": "While graph-based models like ViG and ViHGNN have introduced notable advancements in visual perception, two critical observations emerge regarding these architectures:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 411, + 556, + 567 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 315, + 411, + 556, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 411, + 556, + 470 + ], + "spans": [ + { + "bbox": [ + 315, + 411, + 556, + 470 + ], + "type": "text", + "content": "1. In existing vision GNN models [16, 17, 35, 36], edge features are primarily used for basic vertex-to-vertex communication and are not integrated across successive layers: a strategy that could enhance cumulative learning and improve classification accuracy." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 472, + 556, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 472, + 556, + 567 + ], + "spans": [ + { + "bbox": [ + 315, + 472, + 556, + 567 + ], + "type": "text", + "content": "2. The computational complexities associated with clustering algorithms used for edge generation, such as KNN in ViG and Fuzzy C-Means in ViHGNN, pose significant computational bottlenecks. Approaches like MobileViG [35] and GreedyViG [36] attempt to mitigate these challenges with static graph structures and adding dynamic masking, but do so by trading adaptability for efficiency, failing to achieve a well-balanced solution." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 570, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 557, + 714 + ], + "type": "text", + "content": "In response to limitations in existing graph-based models, we propose the Hypergraph Vision Transformer (HgVT), which advances the hypergraph concept with a bipartite representation where hyperedge features and image patches (vertices) are continuously processed. Unlike traditional models that use graph convolutions, HgVT employs structured multihead attention for efficient vertex-hyperedge message passing and incorporates a dynamic querying mechanism that constructs graph structures in " + }, + { + "bbox": [ + 313, + 570, + 557, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(|V| \\cdot E)" + }, + { + "bbox": [ + 313, + 570, + 557, + 714 + ], + "type": "text", + "content": " time complexity, where " + }, + { + "bbox": [ + 313, + 570, + 557, + 714 + ], + "type": "inline_equation", + "content": "E < |V|" + }, + { + "bbox": [ + 313, + 570, + 557, + 714 + ], + "type": "text", + "content": ". This graph structure is then utilized in attention masking to balance structural adaptability with computational efficiency. Furthermore, HgVT integrates" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9751" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 126, + 72, + 178, + 115 + ], + "blocks": [ + { + "bbox": [ + 126, + 72, + 178, + 115 + ], + "lines": [ + { + "bbox": [ + 126, + 72, + 178, + 115 + ], + "spans": [ + { + "bbox": [ + 126, + 72, + 178, + 115 + ], + "type": "image", + "image_path": "3d7919bc3e33a1d7a2fc5f7b472d8873ead76dd6883c0ebece9cc42ae71b2beb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 126, + 118, + 154, + 128 + ], + "lines": [ + { + "bbox": [ + 126, + 118, + 154, + 128 + ], + "spans": [ + { + "bbox": [ + 126, + 118, + 154, + 128 + ], + "type": "text", + "content": "(a) CNN" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 186, + 72, + 328, + 121 + ], + "blocks": [ + { + "bbox": [ + 186, + 72, + 328, + 121 + ], + "lines": [ + { + "bbox": [ + 186, + 72, + 328, + 121 + ], + "spans": [ + { + "bbox": [ + 186, + 72, + 328, + 121 + ], + "type": "image", + "image_path": "0aea464e472c471a64365e7bb55212f6001bb9141c514e309474e24d08c74bf4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 206, + 121, + 230, + 131 + ], + "lines": [ + { + "bbox": [ + 206, + 121, + 230, + 131 + ], + "spans": [ + { + "bbox": [ + 206, + 121, + 230, + 131 + ], + "type": "text", + "content": "(c) ViG" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 335, + 72, + 487, + 182 + ], + "blocks": [ + { + "bbox": [ + 335, + 72, + 487, + 182 + ], + "lines": [ + { + "bbox": [ + 335, + 72, + 487, + 182 + ], + "spans": [ + { + "bbox": [ + 335, + 72, + 487, + 182 + ], + "type": "image", + "image_path": "cb6a1a05cf39a3f65ae9e5efed64cbed57528fd570db7d877dba53201c454ee3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 350, + 182, + 382, + 191 + ], + "lines": [ + { + "bbox": [ + 350, + 182, + 382, + 191 + ], + "spans": [ + { + "bbox": [ + 350, + 182, + 382, + 191 + ], + "type": "text", + "content": "(e) HgVT" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 128, + 131, + 191, + 186 + ], + "blocks": [ + { + "bbox": [ + 128, + 131, + 191, + 186 + ], + "lines": [ + { + "bbox": [ + 128, + 131, + 191, + 186 + ], + "spans": [ + { + "bbox": [ + 128, + 131, + 191, + 186 + ], + "type": "image", + "image_path": "72cd0db8a09ed4a1847b085b5606afae86b9b78120c967bcdffa2365fb1be12e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 126, + 182, + 149, + 190 + ], + "lines": [ + { + "bbox": [ + 126, + 182, + 149, + 190 + ], + "spans": [ + { + "bbox": [ + 126, + 182, + 149, + 190 + ], + "type": "text", + "content": "(b) ViT" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 198, + 131, + 332, + 182 + ], + "blocks": [ + { + "bbox": [ + 198, + 131, + 332, + 182 + ], + "lines": [ + { + "bbox": [ + 198, + 131, + 332, + 182 + ], + "spans": [ + { + "bbox": [ + 198, + 131, + 332, + 182 + ], + "type": "image", + "image_path": "03f6cb301c05c4006f9a09f32ae726be2c2c02775d834efd107a30adb84fa2b1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 206, + 182, + 248, + 190 + ], + "lines": [ + { + "bbox": [ + 206, + 182, + 248, + 190 + ], + "spans": [ + { + "bbox": [ + 206, + 182, + 248, + 190 + ], + "type": "text", + "content": "(d) ViHGNN" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 196, + 555, + 229 + ], + "lines": [ + { + "bbox": [ + 54, + 196, + 555, + 229 + ], + "spans": [ + { + "bbox": [ + 54, + 196, + 555, + 229 + ], + "type": "text", + "content": "Figure 1. Comparison of Graph Structures for different methods. Showing (a) CNNs, (b) Vision Transformers, (c) ViG with a KNN clustered GNN, (d) ViHGNN with clustered hyperedges, and (e) our proposed HgVT method. Strong group edges shown with solid lines; weak edges with dashed lines. Hyperedges shown with shaded regions; less dominant hyperedges with gray dashed regions." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 240, + 297, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 240, + 297, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 240, + 297, + 300 + ], + "type": "text", + "content": "virtual elements into vertices and hyperedges, forming intermediate links that create a hierarchical semantic structure, as illustrated in Fig. 1e. Attention masking is then used to restrict message passing, and virtual hyperedge features are leveraged for classification. Our contributions are as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 304, + 296, + 518 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 55, + 304, + 296, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 304, + 296, + 387 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 296, + 387 + ], + "type": "text", + "content": "- We propose the Hypergraph Vision Transformer (HgVT), which integrates a hierarchical bipartite hypergraph structure within a vision transformer framework. Our isotropic HgVT-Ti model achieves a Top-1 accuracy of " + }, + { + "bbox": [ + 55, + 304, + 296, + 387 + ], + "type": "inline_equation", + "content": "76.2\\%" + }, + { + "bbox": [ + 55, + 304, + 296, + 387 + ], + "type": "text", + "content": " on the ImageNet-1k classification task, surpassing the prior state-of-the-art by " + }, + { + "bbox": [ + 55, + 304, + 296, + 387 + ], + "type": "inline_equation", + "content": "1.9\\%" + }, + { + "bbox": [ + 55, + 304, + 296, + 387 + ], + "type": "text", + "content": ", demonstrating the efficacy of hypergraph-based learning within vision transformers." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 388, + 296, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 296, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 296, + 434 + ], + "type": "text", + "content": "- We introduce population and diversity regularization strategies that enable dynamic hypergraph structure construction in HgVT, allowing the model to self-sparseify connections without relying on traditional clustering techniques." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 435, + 296, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 435, + 296, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 435, + 296, + 518 + ], + "type": "text", + "content": "- We implement expert edge pooling, a pooling approach that selects edges based on learned confidence scores, facilitating efficient representation pruning and graph extraction. This approach demonstrates strong semantic clustering behavior across macro-classes and achieves competitive image retrieval performance compared to other feature extractors, while maintaining a more compact model size." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 532, + 142, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 532, + 142, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 532, + 142, + 544 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 553, + 297, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 553, + 297, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 553, + 297, + 685 + ], + "type": "text", + "content": "Vision Transformers. Vision Transformers (ViTs) proposed by [12] and refined by [2, 38, 53] use self-attention to process image patches as sequences, scaling to complex datasets and tasks. Recent ViTs have reintroduced spatial hierarchies by leveraging local attention [19, 33], integrating sparse global summaries [68], and employing biomimetic modeling to focus on key regions within images [49]. However, current models tend to focus on the most salient objects and patch-level similarities, ignoring global structure. HgVT addresses this by introducing bipartite hypergraphs to model higher-order relationships for improved semantic understanding." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Graph-Based Vision Models and Clustering. Graph Neural Networks (GNNs), initially conceptualized by [46], have" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 240, + 555, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 240, + 555, + 372 + ], + "spans": [ + { + "bbox": [ + 313, + 240, + 555, + 372 + ], + "type": "text", + "content": "been applied to vision tasks through Vision Graph Neural Networks (ViGs) [16], which exhibit improved accuracy over ViTs on common vision tasks. ViGs use graph convolutions to model image patch relationships on a graph structure, typically constructed by iterative clustering algorithms such as KNN and Fuzzy C-Means, which introduce computational overhead. Recent methods avoid clustering inefficiencies with static graph structures [35, 36], sacrificing adaptability. HgVT instead introduces a dynamic graph construction method, relying on cosine similarity from learned features to enable efficient, non-iterative, adaptive clustering." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 376, + 556, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 556, + 485 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 556, + 485 + ], + "type": "text", + "content": "Hypergraph-Based Methods. While previously used in many computer vision tasks [15, 23, 24], hypergraphs have recently been incorporated into vision GNNs [17, 50], improving their ability to model complex multi-way relationships. However, these methods treat hypergraphs as an intermediate tool rather than producing a hypergraph to represent underlying images, preventing their use in downstream tasks. HgVT instead iteratively refines a hypergraph through subsequent network layers to produce structured representations." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 495, + 465, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 465, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 465, + 509 + ], + "type": "text", + "content": "3. Hierarchical Hypergraphs" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 514, + 556, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 556, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 556, + 659 + ], + "type": "text", + "content": "Graphs and Basic Notations. Graphs are powerful mathematical tools for representing structured information, applicable across diverse disciplines. A graph " + }, + { + "bbox": [ + 313, + 514, + 556, + 659 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 313, + 514, + 556, + 659 + ], + "type": "text", + "content": " is defined as a pair " + }, + { + "bbox": [ + 313, + 514, + 556, + 659 + ], + "type": "inline_equation", + "content": "(\\mathcal{V},\\mathcal{E})" + }, + { + "bbox": [ + 313, + 514, + 556, + 659 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 514, + 556, + 659 + ], + "type": "inline_equation", + "content": "\\mathcal{V} = \\{v_{1},v_{2},\\ldots ,v_{N}\\}" + }, + { + "bbox": [ + 313, + 514, + 556, + 659 + ], + "type": "text", + "content": " is a set of vertices, and " + }, + { + "bbox": [ + 313, + 514, + 556, + 659 + ], + "type": "inline_equation", + "content": "\\mathcal{E} = \\{e_{ij}|(v_i,v_j)\\}" + }, + { + "bbox": [ + 313, + 514, + 556, + 659 + ], + "type": "text", + "content": " is a set of directed edges. For undirected graphs, the edges are symmetric, such that " + }, + { + "bbox": [ + 313, + 514, + 556, + 659 + ], + "type": "inline_equation", + "content": "e_{ij} = e_{ji}" + }, + { + "bbox": [ + 313, + 514, + 556, + 659 + ], + "type": "text", + "content": ". Each edge " + }, + { + "bbox": [ + 313, + 514, + 556, + 659 + ], + "type": "inline_equation", + "content": "e_{ij}" + }, + { + "bbox": [ + 313, + 514, + 556, + 659 + ], + "type": "text", + "content": " connects a pair of vertices " + }, + { + "bbox": [ + 313, + 514, + 556, + 659 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 313, + 514, + 556, + 659 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 514, + 556, + 659 + ], + "type": "inline_equation", + "content": "v_{j}" + }, + { + "bbox": [ + 313, + 514, + 556, + 659 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 514, + 556, + 659 + ], + "type": "inline_equation", + "content": "v_{i},v_{j}\\in \\mathcal{V}" + }, + { + "bbox": [ + 313, + 514, + 556, + 659 + ], + "type": "text", + "content": ". The adjacency matrix " + }, + { + "bbox": [ + 313, + 514, + 556, + 659 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{A}}" + }, + { + "bbox": [ + 313, + 514, + 556, + 659 + ], + "type": "text", + "content": " is a binary matrix " + }, + { + "bbox": [ + 313, + 514, + 556, + 659 + ], + "type": "inline_equation", + "content": "\\{0,1\\}^{|V|\\times |V|}" + }, + { + "bbox": [ + 313, + 514, + 556, + 659 + ], + "type": "text", + "content": ", representing the presence (1) or absence (0) of an edge between each pair of vertices. Similarly, an edge weight matrix can be defined as " + }, + { + "bbox": [ + 313, + 514, + 556, + 659 + ], + "type": "inline_equation", + "content": "\\mathbf{A}\\in \\mathbb{R}^{|V|\\times |V|}" + }, + { + "bbox": [ + 313, + 514, + 556, + 659 + ], + "type": "text", + "content": " to quantify the strength or capacity of these connections." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "type": "text", + "content": "Graph Convolution Networks (GCNs). Building on this foundation, GCNs utilize vertex feature matrices " + }, + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in \\mathbb{R}^{|V| \\times d}" + }, + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "type": "text", + "content": " to encode vertex properties. Their core mechanism, message passing, updates vertex features through a convolution" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9752" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": "with a learned projection matrix " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{W} \\in \\mathbb{R}^{d \\times d}" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " and a non-linear activation, guided by the adjacency matrix " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{A}}" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": ", which specifies neighboring vertices. Adjacency features " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\mathrm{adj}} \\in \\mathbb{R}^{|V| \\times d_a}" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": ", typically set as " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\mathrm{adj}} = \\mathbf{X}" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": ", enable dynamic updates to " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{A}}" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " and the edge weight matrix " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": ", refining the graph structure through learned interactions. However, GCNs are unable to model multi-vertex relationships as edges in " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " are pairwise." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 164, + 285, + 178 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 164, + 285, + 178 + ], + "spans": [ + { + "bbox": [ + 55, + 164, + 285, + 178 + ], + "type": "text", + "content": "3.1. Hypergraphs and Bipartite Representations" + } + ] + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 76, + 186, + 157, + 256 + ], + "blocks": [ + { + "bbox": [ + 76, + 186, + 157, + 256 + ], + "lines": [ + { + "bbox": [ + 76, + 186, + 157, + 256 + ], + "spans": [ + { + "bbox": [ + 76, + 186, + 157, + 256 + ], + "type": "image", + "image_path": "f02d2b1b429d3ba3aabe6d6c87fcfd98faee49b413293856ad186333b6f851c7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 85, + 258, + 145, + 270 + ], + "lines": [ + { + "bbox": [ + 85, + 258, + 145, + 270 + ], + "spans": [ + { + "bbox": [ + 85, + 258, + 145, + 270 + ], + "type": "text", + "content": "(a) Hypergraph" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 171, + 185, + 277, + 256 + ], + "blocks": [ + { + "bbox": [ + 171, + 185, + 277, + 256 + ], + "lines": [ + { + "bbox": [ + 171, + 185, + 277, + 256 + ], + "spans": [ + { + "bbox": [ + 171, + 185, + 277, + 256 + ], + "type": "image", + "image_path": "f8d8f04c1031e4eb2eb532b2bca4efa2f9e8ff744965ef4ac57bdba3c12fe7c9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 188, + 258, + 258, + 270 + ], + "lines": [ + { + "bbox": [ + 188, + 258, + 258, + 270 + ], + "spans": [ + { + "bbox": [ + 188, + 258, + 258, + 270 + ], + "type": "text", + "content": "(b) Bipartite Form" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 274, + 295, + 297 + ], + "lines": [ + { + "bbox": [ + 55, + 274, + 295, + 297 + ], + "spans": [ + { + "bbox": [ + 55, + 274, + 295, + 297 + ], + "type": "text", + "content": "Figure 2. Comparison of (a) hypergraph and (b) equivalent bipartite representation from Fig. 1d, showing five hyperedges " + }, + { + "bbox": [ + 55, + 274, + 295, + 297 + ], + "type": "inline_equation", + "content": "(e_0 - e_4)" + }, + { + "bbox": [ + 55, + 274, + 295, + 297 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 303, + 296, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 303, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 303, + 296, + 445 + ], + "type": "text", + "content": "To overcome the pairwise limitation inherent in traditional graphs, hypergraphs offer a robust solution by extending the concept of edges to hyperedges, which connect multiple vertices simultaneously. In a hypergraph " + }, + { + "bbox": [ + 55, + 303, + 296, + 445 + ], + "type": "inline_equation", + "content": "\\mathcal{H} = (\\mathcal{V},\\mathcal{E})" + }, + { + "bbox": [ + 55, + 303, + 296, + 445 + ], + "type": "text", + "content": ", hyperedges " + }, + { + "bbox": [ + 55, + 303, + 296, + 445 + ], + "type": "inline_equation", + "content": "e_j\\in \\mathcal{E}" + }, + { + "bbox": [ + 55, + 303, + 296, + 445 + ], + "type": "text", + "content": " each connect a subset of vertices, defined as " + }, + { + "bbox": [ + 55, + 303, + 296, + 445 + ], + "type": "inline_equation", + "content": "e_j = \\{v_i\\mid v_i\\in \\mathcal{V}" + }, + { + "bbox": [ + 55, + 303, + 296, + 445 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 303, + 296, + 445 + ], + "type": "inline_equation", + "content": "i\\in I_j\\}" + }, + { + "bbox": [ + 55, + 303, + 296, + 445 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 303, + 296, + 445 + ], + "type": "inline_equation", + "content": "I_{j}" + }, + { + "bbox": [ + 55, + 303, + 296, + 445 + ], + "type": "text", + "content": " is the set of indices for vertices that are included in hyperedge " + }, + { + "bbox": [ + 55, + 303, + 296, + 445 + ], + "type": "inline_equation", + "content": "e_j" + }, + { + "bbox": [ + 55, + 303, + 296, + 445 + ], + "type": "text", + "content": ". The set " + }, + { + "bbox": [ + 55, + 303, + 296, + 445 + ], + "type": "inline_equation", + "content": "I_{j}" + }, + { + "bbox": [ + 55, + 303, + 296, + 445 + ], + "type": "text", + "content": " directly corresponds to the nonzero entries of the " + }, + { + "bbox": [ + 55, + 303, + 296, + 445 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 303, + 296, + 445 + ], + "type": "text", + "content": "-th column of the incidence matrix " + }, + { + "bbox": [ + 55, + 303, + 296, + 445 + ], + "type": "inline_equation", + "content": "\\mathbf{H}\\in \\{0,1\\}^{|\\mathcal{V}|\\times |\\mathcal{E}|}" + }, + { + "bbox": [ + 55, + 303, + 296, + 445 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 303, + 296, + 445 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_{ij} = 1" + }, + { + "bbox": [ + 55, + 303, + 296, + 445 + ], + "type": "text", + "content": " if vertex " + }, + { + "bbox": [ + 55, + 303, + 296, + 445 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 55, + 303, + 296, + 445 + ], + "type": "text", + "content": " is included in the hyperedge " + }, + { + "bbox": [ + 55, + 303, + 296, + 445 + ], + "type": "inline_equation", + "content": "e_j" + }, + { + "bbox": [ + 55, + 303, + 296, + 445 + ], + "type": "text", + "content": ". This structure captures complex relationships, making hypergraphs valuable for modeling networked systems and group interactions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 447, + 296, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 447, + 296, + 542 + ], + "spans": [ + { + "bbox": [ + 55, + 447, + 296, + 542 + ], + "type": "text", + "content": "Hypergraphs can alternatively be described using a bipartite representation, where the vertex set " + }, + { + "bbox": [ + 55, + 447, + 296, + 542 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 55, + 447, + 296, + 542 + ], + "type": "text", + "content": " and hyperedge set " + }, + { + "bbox": [ + 55, + 447, + 296, + 542 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 55, + 447, + 296, + 542 + ], + "type": "text", + "content": " form distinct groups linked by the incidence matrix " + }, + { + "bbox": [ + 55, + 447, + 296, + 542 + ], + "type": "inline_equation", + "content": "\\mathbf{H}" + }, + { + "bbox": [ + 55, + 447, + 296, + 542 + ], + "type": "text", + "content": " (refer to Fig. 2). This representation results in a new graph " + }, + { + "bbox": [ + 55, + 447, + 296, + 542 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_B = (\\mathcal{V},\\mathcal{E},\\mathcal{E}_B)" + }, + { + "bbox": [ + 55, + 447, + 296, + 542 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 447, + 296, + 542 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 55, + 447, + 296, + 542 + ], + "type": "text", + "content": " represents the original vertices of the hypergraph, and the elements in " + }, + { + "bbox": [ + 55, + 447, + 296, + 542 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 55, + 447, + 296, + 542 + ], + "type": "text", + "content": " correspond to hyperedges. The edges in " + }, + { + "bbox": [ + 55, + 447, + 296, + 542 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_B" + }, + { + "bbox": [ + 55, + 447, + 296, + 542 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 55, + 447, + 296, + 542 + ], + "type": "inline_equation", + "content": "\\epsilon_{ve} = (\\nu_v,\\nu_e)" + }, + { + "bbox": [ + 55, + 447, + 296, + 542 + ], + "type": "text", + "content": " exist if " + }, + { + "bbox": [ + 55, + 447, + 296, + 542 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_{ve} = 1" + }, + { + "bbox": [ + 55, + 447, + 296, + 542 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 55, + 447, + 296, + 542 + ], + "type": "inline_equation", + "content": "\\nu_{v}\\in \\mathcal{V}" + }, + { + "bbox": [ + 55, + 447, + 296, + 542 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 447, + 296, + 542 + ], + "type": "inline_equation", + "content": "\\nu_{e}\\in \\mathcal{E}" + }, + { + "bbox": [ + 55, + 447, + 296, + 542 + ], + "type": "text", + "content": ", linking " + }, + { + "bbox": [ + 55, + 447, + 296, + 542 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 55, + 447, + 296, + 542 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 447, + 296, + 542 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 55, + 447, + 296, + 542 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 544, + 296, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 296, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 296, + 652 + ], + "type": "text", + "content": "In the bipartite graph " + }, + { + "bbox": [ + 55, + 544, + 296, + 652 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_B" + }, + { + "bbox": [ + 55, + 544, + 296, + 652 + ], + "type": "text", + "content": ", the corresponding adjacency matrix can be simplified as " + }, + { + "bbox": [ + 55, + 544, + 296, + 652 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{A}} = \\mathbf{H}" + }, + { + "bbox": [ + 55, + 544, + 296, + 652 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 55, + 544, + 296, + 652 + ], + "type": "inline_equation", + "content": "\\mathcal{E} \\to \\mathcal{V}" + }, + { + "bbox": [ + 55, + 544, + 296, + 652 + ], + "type": "text", + "content": " interactions, and " + }, + { + "bbox": [ + 55, + 544, + 296, + 652 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{A}} = \\mathbf{H}^T" + }, + { + "bbox": [ + 55, + 544, + 296, + 652 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 55, + 544, + 296, + 652 + ], + "type": "inline_equation", + "content": "\\mathcal{V} \\to \\mathcal{E}" + }, + { + "bbox": [ + 55, + 544, + 296, + 652 + ], + "type": "text", + "content": ". Drawing on principles similar to those in ViHGNN [17], the edge weight matrix " + }, + { + "bbox": [ + 55, + 544, + 296, + 652 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 55, + 544, + 296, + 652 + ], + "type": "text", + "content": " can be interpreted as fuzzy membership weights, enabling graded interactions across GNN layers. Complementing this setup, the feature matrices are split into " + }, + { + "bbox": [ + 55, + 544, + 296, + 652 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^{(V)} \\in \\mathbb{R}^{|\\mathcal{V}| \\times d_v}" + }, + { + "bbox": [ + 55, + 544, + 296, + 652 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 544, + 296, + 652 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^{(E)} \\in \\mathbb{R}^{|\\mathcal{E}| \\times d_e}" + }, + { + "bbox": [ + 55, + 544, + 296, + 652 + ], + "type": "text", + "content": ", along with their correspond adjacency feature matrices " + }, + { + "bbox": [ + 55, + 544, + 296, + 652 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\mathrm{adj}}^{(V)}" + }, + { + "bbox": [ + 55, + 544, + 296, + 652 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 544, + 296, + 652 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\mathrm{adj}}^{(E)}" + }, + { + "bbox": [ + 55, + 544, + 296, + 652 + ], + "type": "text", + "content": ", mirroring traditional GNNs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 659, + 280, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 280, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 280, + 673 + ], + "type": "text", + "content": "3.2. Imposing Hierarchical Structure in Images" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "To enhance the capability of hypergraphs in image analysis, we draw inspiration from the register tokens introduced in [8], which act to summarize information that otherwise" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 555, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 179 + ], + "type": "text", + "content": "manifests as noise in areas of low visual significance. Similarly, this work integrates virtual vertices " + }, + { + "bbox": [ + 313, + 72, + 555, + 179 + ], + "type": "inline_equation", + "content": "(v\\mathcal{V})" + }, + { + "bbox": [ + 313, + 72, + 555, + 179 + ], + "type": "text", + "content": ", alongside typical image patch vertices " + }, + { + "bbox": [ + 313, + 72, + 555, + 179 + ], + "type": "inline_equation", + "content": "(i\\mathcal{V})" + }, + { + "bbox": [ + 313, + 72, + 555, + 179 + ], + "type": "text", + "content": ", and introduces virtual hyperedges " + }, + { + "bbox": [ + 313, + 72, + 555, + 179 + ], + "type": "inline_equation", + "content": "(v\\mathcal{E})" + }, + { + "bbox": [ + 313, + 72, + 555, + 179 + ], + "type": "text", + "content": ", alongside primary hyperedges " + }, + { + "bbox": [ + 313, + 72, + 555, + 179 + ], + "type": "inline_equation", + "content": "(p\\mathcal{E})" + }, + { + "bbox": [ + 313, + 72, + 555, + 179 + ], + "type": "text", + "content": ", providing levels of semantic feature aggregation and relational abstraction. Notably, these virtual elements, illustrated in Fig. 1e, do not correspond to specific image patches; instead, they are learned embeddings used for semantic summarization and capturing high-level abstract information." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 180, + 556, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 180, + 556, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 556, + 300 + ], + "type": "text", + "content": "Our proposed hypergraph, constructed from image " + }, + { + "bbox": [ + 313, + 180, + 556, + 300 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 180, + 556, + 300 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 180, + 556, + 300 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_B(I)" + }, + { + "bbox": [ + 313, + 180, + 556, + 300 + ], + "type": "text", + "content": ", integrates primary and virtual sets, forming " + }, + { + "bbox": [ + 313, + 180, + 556, + 300 + ], + "type": "inline_equation", + "content": "\\mathcal{V} = i\\mathcal{V} \\cup v\\mathcal{V}" + }, + { + "bbox": [ + 313, + 180, + 556, + 300 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 180, + 556, + 300 + ], + "type": "inline_equation", + "content": "\\mathcal{E} = p\\mathcal{E} \\cup v\\mathcal{E}" + }, + { + "bbox": [ + 313, + 180, + 556, + 300 + ], + "type": "text", + "content": ", with statically masked communication pathways to enforce a hierarchical structure. Primary hyperedges " + }, + { + "bbox": [ + 313, + 180, + 556, + 300 + ], + "type": "inline_equation", + "content": "(p\\mathcal{E})" + }, + { + "bbox": [ + 313, + 180, + 556, + 300 + ], + "type": "text", + "content": " interact with all vertices to support unrestricted semantic aggregation, whereas virtual hyperedges " + }, + { + "bbox": [ + 313, + 180, + 556, + 300 + ], + "type": "inline_equation", + "content": "(v\\mathcal{E})" + }, + { + "bbox": [ + 313, + 180, + 556, + 300 + ], + "type": "text", + "content": ", designated for class predictions, connect solely with virtual vertices " + }, + { + "bbox": [ + 313, + 180, + 556, + 300 + ], + "type": "inline_equation", + "content": "(v\\mathcal{V})" + }, + { + "bbox": [ + 313, + 180, + 556, + 300 + ], + "type": "text", + "content": ". These restrictions separate visual and abstract information, thereby producing a graph structure suitable for use in downstream applications." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 307, + 507, + 320 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 307, + 507, + 320 + ], + "spans": [ + { + "bbox": [ + 313, + 307, + 507, + 320 + ], + "type": "text", + "content": "4. A Hypergraph Vision Transformer" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 327, + 556, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 556, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 556, + 495 + ], + "type": "text", + "content": "The Hypergraph Vision Transformer (HgVT) adapts the architecture of standard Vision Transformers by incorporating bipartite hypergraph features for enhanced image analysis capabilities. Like Vision Transformers, HgVT begins with a patch embedding layer, followed by an isotropic stack of " + }, + { + "bbox": [ + 313, + 327, + 556, + 495 + ], + "type": "inline_equation", + "content": "L \\times \\mathrm{HgVT}" + }, + { + "bbox": [ + 313, + 327, + 556, + 495 + ], + "type": "text", + "content": " blocks, culminating in feature pooling and a classifier head. The bipartite hypergraph is represented by four principal feature matrices " + }, + { + "bbox": [ + 313, + 327, + 556, + 495 + ], + "type": "inline_equation", + "content": "-\\mathbf{X}^{(V)}" + }, + { + "bbox": [ + 313, + 327, + 556, + 495 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 327, + 556, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\mathrm{adj}}^{(V)}" + }, + { + "bbox": [ + 313, + 327, + 556, + 495 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 327, + 556, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^{(E)}" + }, + { + "bbox": [ + 313, + 327, + 556, + 495 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 327, + 556, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\mathrm{adj}}^{(E)}" + }, + { + "bbox": [ + 313, + 327, + 556, + 495 + ], + "type": "text", + "content": " which are updated iteratively and in an interleaved fashion within each block. Each block constructs a new adjacency matrix " + }, + { + "bbox": [ + 313, + 327, + 556, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 313, + 327, + 556, + 495 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 327, + 556, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\mathrm{adj}}^{(V)}" + }, + { + "bbox": [ + 313, + 327, + 556, + 495 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 327, + 556, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\mathrm{adj}}^{(E)}" + }, + { + "bbox": [ + 313, + 327, + 556, + 495 + ], + "type": "text", + "content": ", enabling flexible adjustments to the hypergraph structure. As illustrated in Fig. 3, this modular process allows for the continuous integration and processing of these matrices within each HgVT block." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 504, + 511, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 504, + 511, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 504, + 511, + 517 + ], + "type": "text", + "content": "4.1. Hyperedges as Communication Pools" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "text", + "content": "Each HgVT block processes both vertex and edge information, refining them from the previous block based on a newly constructed graph structure. Initially, adjacency mask computation (detailed in the next section) determines the connectivity for the subsequent processing steps within each block, dynamically adjusting to the updated feature matrices from the previous block. Three attention layers – vertex self-attention, edge aggregate attention, and edge distribution attention – operate sequentially to enhance feature integration and facilitate effective message passing along the hyperedges formed in the adjacency computation step. Finally, separate feed-forward networks process vertex and edge features independently, ensuring specialized treatment for the two distinct sets within the bipartite hypergraph, preserving the unique properties of each set. Operational details of these components are further described in Appendix A." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9753" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 99, + 74, + 509, + 216 + ], + "blocks": [ + { + "bbox": [ + 99, + 74, + 509, + 216 + ], + "lines": [ + { + "bbox": [ + 99, + 74, + 509, + 216 + ], + "spans": [ + { + "bbox": [ + 99, + 74, + 509, + 216 + ], + "type": "image", + "image_path": "14b90bfc81895302ca98b111966fca9f7608cc4dfe160835ee33c206808492ca.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 226, + 555, + 251 + ], + "lines": [ + { + "bbox": [ + 54, + 226, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 54, + 226, + 555, + 251 + ], + "type": "text", + "content": "Figure 3. HgVT Architecture Diagram, composed of stacked HgVT blocks with adjacency matrix " + }, + { + "bbox": [ + 54, + 226, + 555, + 251 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 54, + 226, + 555, + 251 + ], + "type": "text", + "content": ", vertex features " + }, + { + "bbox": [ + 54, + 226, + 555, + 251 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^{(V)}" + }, + { + "bbox": [ + 54, + 226, + 555, + 251 + ], + "type": "text", + "content": ", and hyperedge features " + }, + { + "bbox": [ + 54, + 226, + 555, + 251 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^{(E)}" + }, + { + "bbox": [ + 54, + 226, + 555, + 251 + ], + "type": "text", + "content": ". Edge attention flow is shown with gray arrows; input norms and residual adds are omitted for clarity." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 255, + 297, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 297, + 400 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 297, + 400 + ], + "type": "text", + "content": "Hypergraph Feature Processing. Within each HgVT block, two distinct point-wise feed-forward networks (FFNs) independently process vertex and hyperedge features, aligning with the bipartite structure of the hypergraph. Each FFN integrates both the element features and their corresponding adjacency features through a fully connected layer, improving the model's ability to synthesize relationships. Processing both feature types within the same FFN layer allows adjacency information to be handled directly, bypassing the need for graph-based message passing and improving computational efficiency. Moreover, parameter overhead can be reduced by optionally tying edge and vertex FFN weights." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 59, + 411, + 167, + 486 + ], + "blocks": [ + { + "bbox": [ + 59, + 411, + 167, + 486 + ], + "lines": [ + { + "bbox": [ + 59, + 411, + 167, + 486 + ], + "spans": [ + { + "bbox": [ + 59, + 411, + 167, + 486 + ], + "type": "image", + "image_path": "13b6d70a4bb19ec0622873859df6a17fc3c5cf7349d6cf9dd7b72b6f9789f8a8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 70, + 489, + 153, + 498 + ], + "lines": [ + { + "bbox": [ + 70, + 489, + 153, + 498 + ], + "spans": [ + { + "bbox": [ + 70, + 489, + 153, + 498 + ], + "type": "text", + "content": "(a) Communication Pool." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 176, + 408, + 293, + 487 + ], + "blocks": [ + { + "bbox": [ + 176, + 408, + 293, + 487 + ], + "lines": [ + { + "bbox": [ + 176, + 408, + 293, + 487 + ], + "spans": [ + { + "bbox": [ + 176, + 408, + 293, + 487 + ], + "type": "image", + "image_path": "dc882fb198a1561d69ebe306d286227f612cd16894e4067a343795c9c6459166.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 198, + 490, + 271, + 499 + ], + "lines": [ + { + "bbox": [ + 198, + 490, + 271, + 499 + ], + "spans": [ + { + "bbox": [ + 198, + 490, + 271, + 499 + ], + "type": "text", + "content": "(b) Attention Sparsity." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 504, + 295, + 537 + ], + "lines": [ + { + "bbox": [ + 55, + 504, + 295, + 537 + ], + "spans": [ + { + "bbox": [ + 55, + 504, + 295, + 537 + ], + "type": "text", + "content": "Figure 4. (a) Hyperedge Communication Pool Flow with edges " + }, + { + "bbox": [ + 55, + 504, + 295, + 537 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 55, + 504, + 295, + 537 + ], + "type": "text", + "content": " and member vertices " + }, + { + "bbox": [ + 55, + 504, + 295, + 537 + ], + "type": "inline_equation", + "content": "\\nu_{e}" + }, + { + "bbox": [ + 55, + 504, + 295, + 537 + ], + "type": "text", + "content": "; (b) Attention Sparsity (Mean and std) for HgVT-S on the ImageNet-1k Validation set." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 546, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 297, + 713 + ], + "type": "text", + "content": "Hyperedges as Communication Pools. Hypergraph GNNs typically employ a gather " + }, + { + "bbox": [ + 55, + 546, + 297, + 713 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 546, + 297, + 713 + ], + "type": "text", + "content": " scatter mechanism for processing vertex-hyperedge interactions, whereas HgVT reconceptualizes hyperedges as communication pools that facilitate information flow among vertices and their associated hyperedges. Specifically, vertex self-attention manages vertex-to-vertex " + }, + { + "bbox": [ + 55, + 546, + 297, + 713 + ], + "type": "inline_equation", + "content": "(\\mathcal{V}\\rightarrow \\mathcal{V})" + }, + { + "bbox": [ + 55, + 546, + 297, + 713 + ], + "type": "text", + "content": " interactions within hyperedges, edge aggregate attention orchestrates the flow from vertices to hyperedges " + }, + { + "bbox": [ + 55, + 546, + 297, + 713 + ], + "type": "inline_equation", + "content": "(\\mathcal{V}\\rightarrow \\mathcal{E})" + }, + { + "bbox": [ + 55, + 546, + 297, + 713 + ], + "type": "text", + "content": ", and edge distribution attention handles the reverse, from hyperedges back to vertices " + }, + { + "bbox": [ + 55, + 546, + 297, + 713 + ], + "type": "inline_equation", + "content": "(\\mathcal{E}\\rightarrow \\mathcal{V})" + }, + { + "bbox": [ + 55, + 546, + 297, + 713 + ], + "type": "text", + "content": ". By segmenting the attention operations, HgVT efficiently approximates an all-to-all feature transfer within hyperedges, as illustrated in Fig. 4a, significantly reducing the quadratic complexity associated with full attention mechanisms." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 255, + 556, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 255, + 556, + 507 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 556, + 507 + ], + "type": "text", + "content": "Sparse and Fuzzy Attention. Building upon the dynamic communication pools concept, HgVT employs both sparse and fuzzy attention mechanisms to further optimize computational efficiency. Vertex self-attention is applied selectively to pairs of vertices connected by common hyperedges, as defined by the adjacency matrix " + }, + { + "bbox": [ + 313, + 255, + 556, + 507 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{A}}" + }, + { + "bbox": [ + 313, + 255, + 556, + 507 + ], + "type": "text", + "content": ", resulting in a sparse attention pattern. As sparsity increases with network depth - demonstrated in Fig. 4b - computational load decreases, while still maintaining compatibility with dense attention during training. Conversely, the edge aggregate and distribution attention mechanisms utilize cross-attention between the vertex and edge feature matrices, " + }, + { + "bbox": [ + 313, + 255, + 556, + 507 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^{(V)}" + }, + { + "bbox": [ + 313, + 255, + 556, + 507 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 255, + 556, + 507 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^{(E)}" + }, + { + "bbox": [ + 313, + 255, + 556, + 507 + ], + "type": "text", + "content": ", modulated by the soft adjacency matrix " + }, + { + "bbox": [ + 313, + 255, + 556, + 507 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 313, + 255, + 556, + 507 + ], + "type": "text", + "content": ". This modulation, akin to Fuzzy C-Means in ViHGNNs [17], adjusts attention logits based on soft memberships to the individual hyperedges, dynamically adapting to the hypergraph structure and providing a mechanism for gradient flow into the adjacency matrix generation. Furthermore, by thresholding the soft adjacency matrix during inference, the edge attention mechanisms can be converted into a sparse cross-attention mechanism, thereby reducing computational overhead." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 513, + 484, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 513, + 484, + 526 + ], + "spans": [ + { + "bbox": [ + 313, + 513, + 484, + 526 + ], + "type": "text", + "content": "4.2. Dynamic Adjacency Formation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 530, + 556, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 530, + 556, + 650 + ], + "spans": [ + { + "bbox": [ + 313, + 530, + 556, + 650 + ], + "type": "text", + "content": "HgVT dynamically establishes its hypergraph structure to adapt to the varying semantic and spatial structures of different image inputs. It employs cosine similarity, akin to query-key interactions in attention mechanisms, to evaluate the alignment between vertex and hyperedge adjacency features. This approach allows hyperedges to \"query\" vertices for relevant features, providing a scale-invariant assessment that emphasizes the directionality of embedding vectors. The cosine similarity is subsequently transformed into adjacency membership using a sharpened sigmoid function:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 319, + 656, + 555, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 656, + 555, + 685 + ], + "spans": [ + { + "bbox": [ + 319, + 656, + 555, + 685 + ], + "type": "interline_equation", + "content": "\\mathbf {A} = \\sigma \\left(\\boldsymbol {\\alpha} \\cdot \\tilde {\\mathbf {X}} _ {\\mathrm {a d j}} ^ {(V)} \\left[ \\tilde {\\mathbf {X}} _ {\\mathrm {a d j}} ^ {(E)} \\right] ^ {T}\\right), \\tilde {\\mathbf {X}} _ {\\mathrm {a d j}} ^ {(*)} = \\frac {\\mathbf {X} _ {\\mathrm {a d j}} ^ {(*)}}{| | \\mathbf {X} _ {\\mathrm {a d j}} ^ {(*)} | | _ {2} + \\epsilon} (1)", + "image_path": "e6f4e1cde8dd1df8675d01e4a5cc5cb660e340459eb830a5ec2538edfe01d78d.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "text", + "content": " denotes the sigmoid function and " + }, + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\alpha = 4" + }, + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "text", + "content": " is a sharpening factor, which pushes values away from zero to estab" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9754" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "content": "lish binary-like membership values in matrix " + }, + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "content": ". This soft adjacency matrix is further thresholded to create the hard adjacency matrix " + }, + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{A}} = [\\mathbf{A} > 0.5]" + }, + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "content": ", which provides binary memberships to facilitate sparse attention masking." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 129, + 175, + 143 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 129, + 175, + 143 + ], + "spans": [ + { + "bbox": [ + 55, + 129, + 175, + 143 + ], + "type": "text", + "content": "4.3. Architecture Scaling" + } + ] + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 57, + 195, + 296, + 251 + ], + "blocks": [ + { + "bbox": [ + 55, + 148, + 296, + 194 + ], + "lines": [ + { + "bbox": [ + 55, + 148, + 296, + 194 + ], + "spans": [ + { + "bbox": [ + 55, + 148, + 296, + 194 + ], + "type": "text", + "content": "Table 1. Scaling variants of our HgVT architecture. All models are trained at " + }, + { + "bbox": [ + 55, + 148, + 296, + 194 + ], + "type": "inline_equation", + "content": "224\\mathrm{x}224" + }, + { + "bbox": [ + 55, + 148, + 296, + 194 + ], + "type": "text", + "content": " resolution, except lite variant (HgVT-Lt), trained at " + }, + { + "bbox": [ + 55, + 148, + 296, + 194 + ], + "type": "inline_equation", + "content": "160\\mathrm{x}160" + }, + { + "bbox": [ + 55, + 148, + 296, + 194 + ], + "type": "text", + "content": ". Showing count for vertices " + }, + { + "bbox": [ + 55, + 148, + 296, + 194 + ], + "type": "inline_equation", + "content": "(i\\mathcal{V},v\\mathcal{V})" + }, + { + "bbox": [ + 55, + 148, + 296, + 194 + ], + "type": "text", + "content": ", hyperedges " + }, + { + "bbox": [ + 55, + 148, + 296, + 194 + ], + "type": "inline_equation", + "content": "(p\\mathcal{E}, v\\mathcal{E})" + }, + { + "bbox": [ + 55, + 148, + 296, + 194 + ], + "type": "text", + "content": ", dim for adj. " + }, + { + "bbox": [ + 55, + 148, + 296, + 194 + ], + "type": "inline_equation", + "content": "(d_a)" + }, + { + "bbox": [ + 55, + 148, + 296, + 194 + ], + "type": "text", + "content": " and features " + }, + { + "bbox": [ + 55, + 148, + 296, + 194 + ], + "type": "inline_equation", + "content": "(d_f)" + }, + { + "bbox": [ + 55, + 148, + 296, + 194 + ], + "type": "text", + "content": ", depth " + }, + { + "bbox": [ + 55, + 148, + 296, + 194 + ], + "type": "inline_equation", + "content": "(L)" + }, + { + "bbox": [ + 55, + 148, + 296, + 194 + ], + "type": "text", + "content": ", and heads " + }, + { + "bbox": [ + 55, + 148, + 296, + 194 + ], + "type": "inline_equation", + "content": "(h)" + }, + { + "bbox": [ + 55, + 148, + 296, + 194 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 195, + 296, + 251 + ], + "lines": [ + { + "bbox": [ + 57, + 195, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 57, + 195, + 296, + 251 + ], + "type": "table", + "html": "
Model|iV||vV||pE||vE|df+daLhParams FLOPS
HgVT-Lt10012326128+641246.8M0.92B
HgVT-Mi19616508128+64945.8M1.39B
HgVT-Ti19616508128+641247.7M1.80B
HgVT-S19616508224+9614723M5.48B
", + "image_path": "541437b8911bbcefb6643f69ba946f1d43e2a7c01dbe1095ac3346e7b814cfd6.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 260, + 297, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 260, + 297, + 404 + ], + "spans": [ + { + "bbox": [ + 55, + 260, + 297, + 404 + ], + "type": "text", + "content": "HgVT adopts a hybrid scaling strategy inspired by DeiT [53] and ViG [16], with Tab. 1 detailing the transformer scaling hyperparameters and the vertex and edge type allocations, including fixed capacities for non-image vertices " + }, + { + "bbox": [ + 55, + 260, + 297, + 404 + ], + "type": "inline_equation", + "content": "(i\\mathcal{V})" + }, + { + "bbox": [ + 55, + 260, + 297, + 404 + ], + "type": "text", + "content": " as proposed by ViHGNN [17]. Our scaling approach aims to balance parameter count and FLOPs with the two main GNN baselines, ViG and ViHGNN. We also introduce a Mini variant (HgVT-Mi) to match DeiT-Ti's computational profile, which we achieve by repeating the final block twice for an effective depth of 9; see Appendix H for details. Finally, we include a Ti-Lite variant (HgVT-Lt) for more efficient ablation studies under a constrained training budget." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 415, + 224, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 224, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 224, + 430 + ], + "type": "text", + "content": "5. Enforcing Semantic Structure" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 435, + 297, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 435, + 297, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 435, + 297, + 581 + ], + "type": "text", + "content": "Feature matrices for virtual vertices and hyperedges, lacking direct input-based initialization, risk converging to homogeneous solutions and collapsed representations. Additionally, the dynamic adjacency calculation alone fails to naturally promote semantic grouping, in contrast to clustering-based approaches commonly used in vision GNNs. To address these issues, we introduce diversity regularization, which enforces orthogonal embeddings, and population regularization, which encourages a structured, sparse hypergraph. For enhanced semantic differentiation of virtual hyperedge features, we further incorporate an expert-based pooling strategy as a more robust alternative to mean pooling." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 588, + 268, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 268, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 268, + 601 + ], + "type": "text", + "content": "5.1. Diversity-Driven Feature Differentiation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 605, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 297, + 715 + ], + "type": "text", + "content": "To prevent homogenization of learned feature matrices and to encourage distinct, semantically rich embeddings, we implement a diversity-driven regularization approach. This method, designed to maintain maximum orthogonality among the embeddings of virtual vertices and hyperedges, penalizes the absolute value of the cosine similarity between different feature vectors, aiming for values close to zero. By using normalized embeddings and masking off diagonal elements to preserve self-similarity, the approach prevents" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "text", + "content": "the model from converging to homogeneous solutions or driving individual vectors towards zero magnitude. We then individually penalize " + }, + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "inline_equation", + "content": "v \\mathcal{V}" + }, + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "text", + "content": ", and their adjacency features." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 318, + 114, + 552, + 142 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 114, + 552, + 142 + ], + "spans": [ + { + "bbox": [ + 318, + 114, + 552, + 142 + ], + "type": "interline_equation", + "content": "\\mathrm {D} _ {\\mathrm {L}} (\\mathbf {X}) = \\frac {1}{2} \\sum_ {i j} (1 - \\delta_ {i j}) \\cdot \\left| \\tilde {\\mathbf {X}} \\tilde {\\mathbf {X}} ^ {T} \\right| _ {i j}, \\tilde {\\mathbf {X}} = \\frac {\\mathbf {X}}{| | \\mathbf {X} | | _ {2} + \\epsilon} \\quad (2)", + "image_path": "a8be164c307a6b0c5834ca844ca479b0d6e9db7bb300a465367aa4dc02bba054.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 319, + 144, + 552, + 168 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 144, + 552, + 168 + ], + "spans": [ + { + "bbox": [ + 319, + 144, + 552, + 168 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {D I V}} = \\sum_ {x} \\mathrm {D} _ {\\mathrm {L}} (x), x \\in \\{\\mathbf {X} ^ {(: v V)}, \\mathbf {X} _ {\\mathrm {a d j}} ^ {(: v V)}, \\mathbf {X} ^ {(E)}, \\mathbf {X} _ {\\mathrm {a d j}} ^ {(E)} \\} (3)", + "image_path": "f86689c8cbf4568cfb19586d96cd52300859e4c3df311a38a6d469635764d8d0.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 173, + 555, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 173, + 555, + 234 + ], + "spans": [ + { + "bbox": [ + 313, + 173, + 555, + 234 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 313, + 173, + 555, + 234 + ], + "type": "inline_equation", + "content": "(:vV)" + }, + { + "bbox": [ + 313, + 173, + 555, + 234 + ], + "type": "text", + "content": " represents the subset of " + }, + { + "bbox": [ + 313, + 173, + 555, + 234 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 313, + 173, + 555, + 234 + ], + "type": "text", + "content": " containing only the virtual nodes, " + }, + { + "bbox": [ + 313, + 173, + 555, + 234 + ], + "type": "inline_equation", + "content": "\\delta_{ij}" + }, + { + "bbox": [ + 313, + 173, + 555, + 234 + ], + "type": "text", + "content": " is the Kronecker delta function, ensuring that self-similarity is not penalized, and " + }, + { + "bbox": [ + 313, + 173, + 555, + 234 + ], + "type": "inline_equation", + "content": "|\\cdot|_{ij}" + }, + { + "bbox": [ + 313, + 173, + 555, + 234 + ], + "type": "text", + "content": " denotes the element-wise absolute value, applied to calculate the penalty for non-orthogonal relationships between embeddings." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 239, + 549, + 252 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 549, + 252 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 549, + 252 + ], + "type": "text", + "content": "5.2. Population Regularization: Learned Sparsity" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 256, + 556, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 256, + 556, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 256, + 556, + 376 + ], + "type": "text", + "content": "Unlike clustering methods such as KNN or Fuzzy C-Means, which impose fixed cluster sizes, our model's dynamic adjacency calculation enables flexible and self-adjusting hyperedge populations. To prevent the associated risks of overly sparse or densely connected hypergraphs, we introduce population regularization. This method applies penalties based on the computed soft membership density of each hyperedge derived from the soft adjacency matrix " + }, + { + "bbox": [ + 313, + 256, + 556, + 376 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 313, + 256, + 556, + 376 + ], + "type": "text", + "content": ", ensuring each maintains a vertex population within appropriate bounds to avoid overgeneralization and preserve hypergraph integrity." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 324, + 381, + 555, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 381, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 324, + 381, + 555, + 407 + ], + "type": "interline_equation", + "content": "P _ {j} = 2 \\cdot \\sum_ {i} \\max (A _ {i j} - 0. 5, 0) \\tag {4}", + "image_path": "c0eb6a694ec84d941b8caeb9dc40bafb52cbbf7743c6dba0098130742e281ce6.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 324, + 408, + 555, + 434 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 408, + 555, + 434 + ], + "spans": [ + { + "bbox": [ + 324, + 408, + 555, + 434 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {P O P}} = \\sum_ {j} \\max (P _ {j} - \\beta , 0) + \\max (\\gamma - P _ {j}, 0) \\tag {5}", + "image_path": "dce12b67db10a33aaffc11d2427d47bd5bde97577c88b93cb0b1d0996466dec6.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 435, + 556, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 435, + 556, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 435, + 556, + 519 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 435, + 556, + 519 + ], + "type": "inline_equation", + "content": "P_{j}" + }, + { + "bbox": [ + 313, + 435, + 556, + 519 + ], + "type": "text", + "content": " represents a soft density estimate of vertex connections for the " + }, + { + "bbox": [ + 313, + 435, + 556, + 519 + ], + "type": "inline_equation", + "content": "j^{th}" + }, + { + "bbox": [ + 313, + 435, + 556, + 519 + ], + "type": "text", + "content": " hyperedge, only considering non-zero entries of " + }, + { + "bbox": [ + 313, + 435, + 556, + 519 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{A}}" + }, + { + "bbox": [ + 313, + 435, + 556, + 519 + ], + "type": "text", + "content": ". The hyperparameters " + }, + { + "bbox": [ + 313, + 435, + 556, + 519 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 435, + 556, + 519 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 435, + 556, + 519 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 435, + 556, + 519 + ], + "type": "text", + "content": " set the upper and lower density limits, ensuring that hyperedges maintain an optimal balance of connections. Penalties are applied if " + }, + { + "bbox": [ + 313, + 435, + 556, + 519 + ], + "type": "inline_equation", + "content": "P_{j}" + }, + { + "bbox": [ + 313, + 435, + 556, + 519 + ], + "type": "text", + "content": " either exceeds " + }, + { + "bbox": [ + 313, + 435, + 556, + 519 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 435, + 556, + 519 + ], + "type": "text", + "content": " or falls below " + }, + { + "bbox": [ + 313, + 435, + 556, + 519 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 435, + 556, + 519 + ], + "type": "text", + "content": ", maintaining the desired sparsity and ensuring the structural efficacy of the hypergraph." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 525, + 538, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 538, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 538, + 538 + ], + "type": "text", + "content": "5.3. Expert Pooling for Semantic Specialization" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 542, + 556, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 542, + 556, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 542, + 556, + 662 + ], + "type": "text", + "content": "To effectively combine features from multiple virtual hyperedges for classification, our approach utilizes a method akin to expert-choice, where each virtual hyperedge acts as an \"expert\" generating a confidence score. Unlike mean pooling, which risks collapsing distinct features into an average that may dilute individual contributions, this strategy encourages virtual hyperedges to develop unique, semantically meaningful representations. The normalized confidence scores, " + }, + { + "bbox": [ + 313, + 542, + 556, + 662 + ], + "type": "inline_equation", + "content": "P(e)" + }, + { + "bbox": [ + 313, + 542, + 556, + 662 + ], + "type": "text", + "content": ", determine the relevance of each hyperedge " + }, + { + "bbox": [ + 313, + 542, + 556, + 662 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 313, + 542, + 556, + 662 + ], + "type": "text", + "content": "'s contribution to the classification task." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 359, + 666, + 555, + 687 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 666, + 555, + 687 + ], + "spans": [ + { + "bbox": [ + 359, + 666, + 555, + 687 + ], + "type": "interline_equation", + "content": "P (e) = \\operatorname {s o f t m a x} \\left(\\mathbf {X} _ {e} ^ {(v E)} \\mathbf {W} _ {e} + b _ {e}\\right) \\tag {6}", + "image_path": "3eb4b101578c40806d954b76dae771ffe425f720b075dec0ef3d65e9b7a37881.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "inline_equation", + "content": "(:vE)" + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": " denotes the subset of " + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "inline_equation", + "content": "(E)" + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": " containing only the virtual hyperedges, and the softmax is computed across the" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9755" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": "expert gate set " + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "inline_equation", + "content": "\\{e\\}" + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": " after projection by " + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "inline_equation", + "content": "\\mathbf{W} \\in \\mathbb{R}^{d \\times |vE|}" + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": ". During training, " + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "inline_equation", + "content": "P(e)" + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": " guides the weighted averaging of hyperedge features before subsequent class prediction. We further apply a density loss function [3, 14] to prevent underutilization of any single virtual hyperedge, complemented by a cross-entropy term with label smoothing to increase expert confidence. During inference, top-k routing combines the most relevant hyperedges based on their confidence scores." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 179, + 272, + 192 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 179, + 272, + 192 + ], + "spans": [ + { + "bbox": [ + 55, + 179, + 272, + 192 + ], + "type": "text", + "content": "6. Empirical Evaluation and Performance" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 200, + 296, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 200, + 296, + 283 + ], + "spans": [ + { + "bbox": [ + 55, + 200, + 296, + 283 + ], + "type": "text", + "content": "This section evaluates the Hypergraph Vision Transformer using two model configurations from Tab. 1: HgVT-Ti-Lite for targeted ablation studies and scaled variants for benchmarking against comparable classifiers. We apply standard augmentations from DeiT [53] using the Timm library [58], including RandAugment [7], Mixup [64], Cutmix [60], Random Erasing [66], and Repeated Augment [21]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 290, + 296, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 290, + 296, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 290, + 296, + 422 + ], + "type": "text", + "content": "Datasets. For classification in computer vision, we follow standard practices and use the ImageNet-1k dataset [10] at a resolution of " + }, + { + "bbox": [ + 55, + 290, + 296, + 422 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 55, + 290, + 296, + 422 + ], + "type": "text", + "content": " pixels for scaled model evaluations. For ablation studies, we employ ImageNet-100 [51], a 100-class subset of ImageNet-1k with images scaled to " + }, + { + "bbox": [ + 55, + 290, + 296, + 422 + ], + "type": "inline_equation", + "content": "160 \\times 160" + }, + { + "bbox": [ + 55, + 290, + 296, + 422 + ], + "type": "text", + "content": " pixels. This selection provides a computationally manageable dataset while maintaining sufficient class variation and larger image sizes compared to datasets like CIFAR-100 (32x32 pixels)[27]. Nevertheless, we find CIFAR-100 useful for assessing the effects of regularization on the hypergraph structure, as detailed in Appendix H." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 429, + 296, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 429, + 296, + 573 + ], + "spans": [ + { + "bbox": [ + 55, + 429, + 296, + 573 + ], + "type": "text", + "content": "Training Hyperparameters. Consistent with DeiT, we use the AdamW optimizer with a weight decay of 0.05. Training is conducted on the ImageNet-1k dataset with a batch size of 1024 for 300 epochs following DeiT. For ablations, we train on ImageNet-100 with a batch size of 512 for with a shorter duration of 200 epochs as proposed by [31]. Learning rates follow a cosine-annealing schedule peaking at 1e-3 for both datasets following scaling from DeiT. Furthermore, we omit the use of Exponential Moving Average (EMA) due to its minimal performance improvement (0.1% in DeiT) relative to its overhead per training step. All models were trained using PyTorch on NVIDIA A6000 GPUs; see Appendix I." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 580, + 296, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 580, + 296, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 580, + 296, + 664 + ], + "type": "text", + "content": "Evaluation Metrics. Following standard protocols, we measure the Top-1 class prediction accuracy to assess overall performance. Additionally, we take advantage of the learned graph structure (extracted from the last layer) on each image, and measure: Hyperedge Entropy (HE), Intra-Cluster Similarity (ICS), Inter-Cluster Distance (ICD), and Silhouette Score (SIL) [45]; further details on structure in Appendix C." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 672, + 192, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 192, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 192, + 684 + ], + "type": "text", + "content": "6.1. Evaluation on ImageNet" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "Tab. 2 presents the ImageNet-1k Top-1 accuracy of HgVT, benchmarked against isotropic models. We limit evaluation" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 317, + 98, + 553, + 277 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 555, + 93 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 555, + 93 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 555, + 93 + ], + "type": "text", + "content": "Table 2. ImageNet-1k results for HgVT and other isotropic networks. " + }, + { + "bbox": [ + 313, + 71, + 555, + 93 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 313, + 71, + 555, + 93 + ], + "type": "text", + "content": " CNN, " + }, + { + "bbox": [ + 313, + 71, + 555, + 93 + ], + "type": "inline_equation", + "content": "\\clubsuit" + }, + { + "bbox": [ + 313, + 71, + 555, + 93 + ], + "type": "text", + "content": " Transformer, " + }, + { + "bbox": [ + 313, + 71, + 555, + 93 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 313, + 71, + 555, + 93 + ], + "type": "text", + "content": " GNN, " + }, + { + "bbox": [ + 313, + 71, + 555, + 93 + ], + "type": "inline_equation", + "content": "\\blacksquare" + }, + { + "bbox": [ + 313, + 71, + 555, + 93 + ], + "type": "text", + "content": " HGNN, and " + }, + { + "bbox": [ + 313, + 71, + 555, + 93 + ], + "type": "inline_equation", + "content": "\\triangle" + }, + { + "bbox": [ + 313, + 71, + 555, + 93 + ], + "type": "text", + "content": " HgVT." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 98, + 553, + 277 + ], + "lines": [ + { + "bbox": [ + 317, + 98, + 553, + 277 + ], + "spans": [ + { + "bbox": [ + 317, + 98, + 553, + 277 + ], + "type": "table", + "html": "
ModelParamsFLOPsImNet Top-1ReAL Top-1V2 Top-1
*ResMLP-S12 conv3x3 [52]16.7M3.2B77.084.065.5
*ConvMixer-768/32 [55]21.1M20.9B80.2--
*ConvMixer-1536/20 [55]51.6M51.1B81.4--
*DINOv1-S [2]21.7M4.6B77.0--
*ViT-B/16 [12]86.4M55.5B77.983.6-
*DeiT-Ti [53]5.7M1.3B72.280.160.4
*DeiT-S [53]22.1M4.6B79.885.768.5
*DeiT-B [53]86.4M17.6B81.886.771.5
*ViG-Ti [16]7.1M1.3B73.9--
*ViG-S [16]22.7M4.5B80.4--
*ViG-B [16]86.8M17.7B82.3--
■ViHGNN-Ti [17]8.2M1.8B74.3--
■ViHGNN-S [17]23.2M5.6B81.5--
■ViHGNN-B [17]88.1M19.4B82.9--
▲HgVT-Mi (ours)5.8M1.4B74.481.862.7
▲HgVT-Ti (ours)7.7M1.8B76.283.264.3
▲HgVT-S (ours)22.9M5.5B81.286.770.1
", + "image_path": "c6aff34e8e1d152651b42f65de02fb5f4f8383d67f4fe5793bfd522831bd51da.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 290, + 553, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 290, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 290, + 553, + 361 + ], + "type": "text", + "content": "to isotropic models due to the complexity of down-sampling virtual tokens, lacking direct spatial correspondence, needed for pyramidal structures. As a result, we also exclude other pyramidal models from our comparison as they typically outperform isotropic models within the same architecture family [16, 17], making the comparison less meaningful." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 363, + 555, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 363, + 555, + 482 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 555, + 482 + ], + "type": "text", + "content": "HgVT-Ti demonstrates a clear advantage, surpassing ViHGNN-Ti by " + }, + { + "bbox": [ + 313, + 363, + 555, + 482 + ], + "type": "inline_equation", + "content": "1.9\\%" + }, + { + "bbox": [ + 313, + 363, + 555, + 482 + ], + "type": "text", + "content": " in accuracy, with the smaller HgVT-Mi model matching ViHGNN-Ti with only " + }, + { + "bbox": [ + 313, + 363, + 555, + 482 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 313, + 363, + 555, + 482 + ], + "type": "text", + "content": " of the parameters. At the S-scale, we reduce HgVT's depth relative to ViHGNN to accommodate hyperedge attention and adjacency feature overhead, aiming for comparable parameter counts; as a result, HgVT-S only matches ViHGNN-S. Despite this limitation, HgVT-S matches DeiT-B's accuracy on ImageNet Real [1] and achieves competitive performance on ImageNet V2 [43], at nearly a quarter of DeiT-B's size." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 483, + 556, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 556, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 556, + 579 + ], + "type": "text", + "content": "Overall, HgVT achieves strong performance on classification, demonstrating the benefits of integrating hypergraph structures within a vision transformer framework. Additionally, computational evaluation in Appendix B shows a consistent " + }, + { + "bbox": [ + 313, + 483, + 556, + 579 + ], + "type": "inline_equation", + "content": "\\geq 1.4\\times" + }, + { + "bbox": [ + 313, + 483, + 556, + 579 + ], + "type": "text", + "content": " speedup over ViG, with adjacency computation exhibiting a " + }, + { + "bbox": [ + 313, + 483, + 556, + 579 + ], + "type": "inline_equation", + "content": "10\\times" + }, + { + "bbox": [ + 313, + 483, + 556, + 579 + ], + "type": "text", + "content": " improvement over ViG's KNN-based method. These improvements further highlight HgVT's potential for greater efficiency over prior Vision GNNs." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 588, + 415, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 588, + 415, + 599 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 415, + 599 + ], + "type": "text", + "content": "6.2. Ablation Studies" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "content": "We conducted a series of ablations on the ImageNet-100 dataset using the HgVT-Lt model, reporting Top-1 classification accuracy alongside mean hyperedge entropy and silhouette scores to assess the quality of the hypergraph's structure. Notably, we observe a weak anti-correlation between the graph quality metrics and Top-1 accuracy (see Appendix H), indicating opposing objectives. Overall, the ablations are grouped into three categories: regularization, architecture, and pooling methods, with results in Tab. 3." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 315, + 757 + ], + "type": "text", + "content": "9756" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 107, + 296, + 285 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "type": "text", + "content": "Table 3. ImageNet-100 ablations with HgVT-Lt. Indicating used (✓) or not used " + }, + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "type": "inline_equation", + "content": "(\\mathcal{X})" + }, + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "type": "text", + "content": " , and pooling methods: Average (A), Image " + }, + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "type": "inline_equation", + "content": "(I)" + }, + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "type": "text", + "content": " , Expert " + }, + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "type": "inline_equation", + "content": "(E)" + }, + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "type": "text", + "content": " , Expert " + }, + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "type": "text", + "content": " Image " + }, + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "type": "inline_equation", + "content": "(EI)" + }, + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "type": "text", + "content": " , and " + }, + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "type": "inline_equation", + "content": "EI" + }, + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "type": "text", + "content": " dropping I " + }, + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "type": "inline_equation", + "content": "(DI)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 107, + 296, + 285 + ], + "lines": [ + { + "bbox": [ + 58, + 107, + 296, + 285 + ], + "spans": [ + { + "bbox": [ + 58, + 107, + 296, + 285 + ], + "type": "table", + "html": "
Ablation on ↓CLS Dropout Stoch. Decay Diversity PopulationTied FFN Xadj = X dF Mult. Pooled OpEdge EntropyTop-1 Acc.Params (M)
X✓✓✓✓1EI3.320.78084.36
none: HgVT-Lt✓✓✓✓✓1E3.130.75182.23
RegularizationXX✓✓✓1E3.120.74581.89
X✓X✓✓1E1.990.72380.79
X✓✓X✓1E3.890.63981.79
X✓XX✓1E3.580.61081.99
ArchitectureX✓✓✓X1E3.090.74182.89
X✓✓✓X1E2.270.80878.62
X✓✓✓✓1E2.120.78076.95
X✓✓✓✓1.5E2.050.77077.46
PoolingX✓✓✓✓1A3.070.74782.06
X✓✓✓✓1I2.930.76084.08
X✓✓✓✓1E3.130.74982.52
X✓✓✓✓1DI3.320.78080.94
", + "image_path": "090833d322eedcf24fe1b323ca019d45066e3d0c43472d52cfd14b233de884e5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 298, + 295, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 298, + 295, + 418 + ], + "spans": [ + { + "bbox": [ + 54, + 298, + 295, + 418 + ], + "type": "text", + "content": "As shown in Tab. 3, the regularization ablations demonstrate that stochastic path dropout decay [22] improves both Top-1 accuracy and silhouette scores, consistent with ViG and ViHGNN [16, 17]. Omitting Class dropout also boosts accuracy, aligning with DeiT [53]. Furthermore, our proposed diversity and population regularization are essential for preserving graph structure; removing diversity leads to partial representation collapse, while removing population regularization results in near-zero sparsity, effectively turning HgVT into a ViT with increased network complexity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 418, + 295, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 418, + 295, + 527 + ], + "spans": [ + { + "bbox": [ + 54, + 418, + 295, + 527 + ], + "type": "text", + "content": "In the architecture ablations, untying the FFN improves accuracy but significantly increases parameter count, making it an unfavorable tradeoff. Tying adjacency and embedding features " + }, + { + "bbox": [ + 54, + 418, + 295, + 527 + ], + "type": "inline_equation", + "content": "(\\mathbf{X}_{\\mathrm{adj}} = \\mathbf{X})" + }, + { + "bbox": [ + 54, + 418, + 295, + 527 + ], + "type": "text", + "content": " reduces parameters and FLOPs but degrades performance, and while untying the FFN or increasing feature dimensionality partially mitigates this, the parameter increase remains suboptimal. This indicates that adjacency and embedding features " + }, + { + "bbox": [ + 54, + 418, + 295, + 527 + ], + "type": "inline_equation", + "content": "(\\mathbf{X}_{\\mathrm{adj}}^{(*)}" + }, + { + "bbox": [ + 54, + 418, + 295, + 527 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 418, + 295, + 527 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^{(*)})" + }, + { + "bbox": [ + 54, + 418, + 295, + 527 + ], + "type": "text", + "content": " are similar, yet require dedicated feature spaces to avoid crowding." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 528, + 295, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 528, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 54, + 528, + 295, + 635 + ], + "type": "text", + "content": "For pooling methods, expert edge pooling outperforms average edge pooling in accuracy, while image pooling achieves the highest accuracy at the cost of degraded graph structure. Combining image and expert pooling recovers lost structure and improves accuracy, with each input focusing on different semantic levels (see Appendix D). Additionally, dropping the pooled image embedding prior to the classifier head maintains moderate performance, indicating that both paths meaningfully contribute to the final prediction." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 641, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 641, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 641, + 297, + 714 + ], + "type": "text", + "content": "Impact of Vertex Self-Attention and Patch Embedding. We evaluated the impact of patch embedding methods and vertex self-attention, comparing a convolutional stem (Conv2D-BN-GELU layers [16, 17]) and a simpler patch projection (pixel-shuffled patches with affine projection [2, 12]), across various pooling strategies (average, image, and ex" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 324, + 108, + 543, + 159 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 555, + 105 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 555, + 105 + ], + "type": "text", + "content": "Table 4. Top-1 accuracy of HgVT-Lt on ImageNet-100 with " + }, + { + "bbox": [ + 313, + 70, + 555, + 105 + ], + "type": "inline_equation", + "content": "(\\checkmark)" + }, + { + "bbox": [ + 313, + 70, + 555, + 105 + ], + "type": "text", + "content": " or without " + }, + { + "bbox": [ + 313, + 70, + 555, + 105 + ], + "type": "inline_equation", + "content": "(X)" + }, + { + "bbox": [ + 313, + 70, + 555, + 105 + ], + "type": "text", + "content": " vertex self-attention. Contrasting pooling operations and patch embedding versions: Conv. Stem or Patch Projection." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 324, + 108, + 543, + 159 + ], + "lines": [ + { + "bbox": [ + 324, + 108, + 543, + 159 + ], + "spans": [ + { + "bbox": [ + 324, + 108, + 543, + 159 + ], + "type": "table", + "html": "
Pooling Op. →AverageImageExpert
Vertex SA →X✓X✓X✓
Conv. Stem78.0282.0682.0584.0878.8782.52
Patch Project64.3072.7670.7676.1762.6271.43
", + "image_path": "13466a7ead897590b93de4e0ea3fbe8f8fb611f86775139cb295dfc67aecd952.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 169, + 555, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 169, + 555, + 314 + ], + "spans": [ + { + "bbox": [ + 313, + 169, + 555, + 314 + ], + "type": "text", + "content": "pert), as shown in Tab. 4. The patch projection consistently underperforms the convolutional stem, likely due to the model's small size limiting its effectiveness. Omitting vertex self-attention leads to further degradation, especially without the convolutional stem, suggesting it is crucial for effectively separating features in low-dimensional space. PCA shows that 71/128 channels are needed to explain " + }, + { + "bbox": [ + 313, + 169, + 555, + 314 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 313, + 169, + 555, + 314 + ], + "type": "text", + "content": " of the variance for the convolutional stem, compared to 19/128 for the patch projection, indicating the richer representation captured by the convolutional stem. Notably, image pooling shows the least degradation, likely due to its more direct gradient flow compared to the edge pooling methods." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 321, + 518, + 334 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 321, + 518, + 334 + ], + "spans": [ + { + "bbox": [ + 313, + 321, + 518, + 334 + ], + "type": "text", + "content": "6.3. Pooling Methods and Graph Structure" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 317, + 376, + 555, + 451 + ], + "blocks": [ + { + "bbox": [ + 313, + 339, + 555, + 373 + ], + "lines": [ + { + "bbox": [ + 313, + 339, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 339, + 555, + 373 + ], + "type": "text", + "content": "Table 5. Impact of graph structure on pooling operation for HgVT-Lt on ImageNet-100. Measuring graph metrics for image vertices " + }, + { + "bbox": [ + 313, + 339, + 555, + 373 + ], + "type": "inline_equation", + "content": "(i\\mathcal{V})" + }, + { + "bbox": [ + 313, + 339, + 555, + 373 + ], + "type": "text", + "content": " and all vertices " + }, + { + "bbox": [ + 313, + 339, + 555, + 373 + ], + "type": "inline_equation", + "content": "(\\mathcal{V})" + }, + { + "bbox": [ + 313, + 339, + 555, + 373 + ], + "type": "text", + "content": ", along with features from DINOv2." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 376, + 555, + 451 + ], + "lines": [ + { + "bbox": [ + 317, + 376, + 555, + 451 + ], + "spans": [ + { + "bbox": [ + 317, + 376, + 555, + 451 + ], + "type": "table", + "html": "
Pooling Op. →ImageExpertImg. & Expert
Feature Model ↓HEICSICDHEICSICDHEICSICD
HgVT-Lt (iV)3.030.500.313.240.430.373.230.420.36
HgVT-Lt (V)3.200.250.723.390.360.453.390.280.58
DINO2-S (iV)3.040.840.083.250.850.053.250.830.06
DINO2-G (iV)3.040.690.153.250.680.123.250.680.13
", + "image_path": "3fc8047445d3c45ca53af5bc68f612b0605d6680d913f7e6c44a4d1f94fb7477.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 462, + 556, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 556, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 556, + 629 + ], + "type": "text", + "content": "To evaluate the impact of pooling methods on graph structure, we measured HE, ICS, and ICD using the ImageNet-100 validation set with three strategies: image pooling, expert pooling, and a combined image + expert pooling approach. Metrics used either the image vertex subset " + }, + { + "bbox": [ + 313, + 462, + 556, + 629 + ], + "type": "inline_equation", + "content": "(i\\mathcal{V})" + }, + { + "bbox": [ + 313, + 462, + 556, + 629 + ], + "type": "text", + "content": " or the full vertex set " + }, + { + "bbox": [ + 313, + 462, + 556, + 629 + ], + "type": "inline_equation", + "content": "(\\mathcal{V})" + }, + { + "bbox": [ + 313, + 462, + 556, + 629 + ], + "type": "text", + "content": ", with features derived from DINOv2 (S and G) [38] and HgVT-Lt using HgVT-Lt's adjacency matrix (A). Results in Tab. 5 show that while all methods achieve similar graph quality using " + }, + { + "bbox": [ + 313, + 462, + 556, + 629 + ], + "type": "inline_equation", + "content": "i\\mathcal{V}" + }, + { + "bbox": [ + 313, + 462, + 556, + 629 + ], + "type": "text", + "content": ", image pooling slightly improves similarity. However, including all vertices " + }, + { + "bbox": [ + 313, + 462, + 556, + 629 + ], + "type": "inline_equation", + "content": "(\\mathcal{V})" + }, + { + "bbox": [ + 313, + 462, + 556, + 629 + ], + "type": "text", + "content": " consistently increases ICD and entropy while reducing ICS, indicating decreased graph coherence. This effect is more severe with image pooling methods, suggesting that virtual vertices " + }, + { + "bbox": [ + 313, + 462, + 556, + 629 + ], + "type": "inline_equation", + "content": "(v\\mathcal{V})" + }, + { + "bbox": [ + 313, + 462, + 556, + 629 + ], + "type": "text", + "content": " act as noisy elements, rather than summarization points." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "Comparing DINOv2 models, all pooling methods align more closely with DINOv2-G, where achieving a balance between ICS and ICD is preferable to maximizing either individually. This trend, along with consistent HE, suggests a focus on higher-level detail, irrespective of the smaller HgVT model size or pooling method. Image pooling shows slightly stronger alignment with both DINO models, indicating that" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9757" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 156 + ], + "type": "text", + "content": "both high- and low-level semantics are encoded within a single feature space, unlike methods that can use edge channels for high-level concepts. Notably, all expert pooling methods exhibit an emergent macro-class prediction behavior, where each virtual edge " + }, + { + "bbox": [ + 55, + 72, + 297, + 156 + ], + "type": "inline_equation", + "content": "(v\\mathcal{E})" + }, + { + "bbox": [ + 55, + 72, + 297, + 156 + ], + "type": "text", + "content": " consistently captures broader taxonomic groups (e.g., dogs, birds). Further representation and macro-class analysis are provided in Appendices D and J." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 166, + 231, + 178 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 166, + 231, + 178 + ], + "spans": [ + { + "bbox": [ + 55, + 166, + 231, + 178 + ], + "type": "text", + "content": "6.4. Performance on Image Retrieval" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 184, + 296, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 296, + 399 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 296, + 399 + ], + "type": "text", + "content": "To evaluate the capability of HgVT in capturing semantic structures, we perform image retrieval experiments comparing four methods: pooling similarity (PS), volumetric similarity (VS), adaptive pooling similarity (APS), and adaptive volumetric similarity (AVS). PS ranks the pooled embeddings by cosine similarity (vector search), while the other methods enhance retrieval by leveraging graph structure. Volumetric similarity calculates ellipsoid overlap using an approximate Mahalanobis distance, with pruned hyperedges defining the distribution spread around the pooled embedding (centroid). Adaptive methods further refine retrieval via a graph similarity measure on the pruned hyperedges, re-ranking from a shortlist of " + }, + { + "bbox": [ + 56, + 184, + 296, + 399 + ], + "type": "inline_equation", + "content": "R = 100" + }, + { + "bbox": [ + 56, + 184, + 296, + 399 + ], + "type": "text", + "content": ". Computational efficiency in adaptive retrieval is ensured through centroid hash binning and limiting comparisons to the top-" + }, + { + "bbox": [ + 56, + 184, + 296, + 399 + ], + "type": "inline_equation", + "content": "C = 4" + }, + { + "bbox": [ + 56, + 184, + 296, + 399 + ], + "type": "text", + "content": " most significant query hyperedges, resulting in a complexity of " + }, + { + "bbox": [ + 56, + 184, + 296, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(RC)" + }, + { + "bbox": [ + 56, + 184, + 296, + 399 + ], + "type": "text", + "content": ". Notably, we prune to 12 hyperedges and use 10 centroid bins. Additional details provided in Appendix G." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 58, + 443, + 294, + 557 + ], + "blocks": [ + { + "bbox": [ + 55, + 406, + 296, + 440 + ], + "lines": [ + { + "bbox": [ + 55, + 406, + 296, + 440 + ], + "spans": [ + { + "bbox": [ + 55, + 406, + 296, + 440 + ], + "type": "text", + "content": "Table 6. Image Retrieval on ImageNet-1k with proposed search methods: PS, APS, VS, and AVS. Reporting Top-1 accuracy, mAP@10 (\\%), and 1-NN-hit@10 (\\%) for pooling methods." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 443, + 294, + 557 + ], + "lines": [ + { + "bbox": [ + 58, + 443, + 294, + 557 + ], + "spans": [ + { + "bbox": [ + 58, + 443, + 294, + 557 + ], + "type": "table", + "html": "
mAP@10 (%)
ModelTop-1PoolingPSAPSVSAVS
MRL-128 [29]70.52Image64.9465.20--
MRL-256 [29]70.62Image65.0465.20--
HgVT-Ti (ours)76.18Expert70.5669.5970.5369.40
HgVT-Ti (ours)76.18Im&Ex73.2369.5973.0869.49
1-NN-hit@10 (%)
HgVT-Ti (ours)76.18Expert21.2219.1021.2519.56
HgVT-Ti (ours)76.18Im&Ex25.1319.1025.2219.17
", + "image_path": "b034752c368d19342c826677dbbb0a5a6595986a1312976eb4a42a9fa35abcc8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 570, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 297, + 713 + ], + "type": "text", + "content": "ImageNet Retrieval. We evaluate retrieval performance on the ImageNet-1K dataset to assess HgVT's ability to capture semantic relationships and compare four retrieval methods: PS, VS, APS, and AVS. The primary metric is mAP@10, with MRL [29] serving as the baseline due to its adaptive re-ranking approach. We also report the 1-NN-CLIP-L hitrate@10, which measures how often the top-1 result ranked by CLIP-L [42] appears in the top-10 retrieved results, offering additional insight into semantic alignment. Results in Tab. 6 show that HgVT-Ti surpasses MRL by over " + }, + { + "bbox": [ + 55, + 570, + 297, + 713 + ], + "type": "inline_equation", + "content": "8\\%" + }, + { + "bbox": [ + 55, + 570, + 297, + 713 + ], + "type": "text", + "content": " in retrieval performance, despite being significantly smaller than MRL (ResNet-50) and using a comparably compact" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 555, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 133 + ], + "type": "text", + "content": "embedding size (" + }, + { + "bbox": [ + 313, + 72, + 555, + 133 + ], + "type": "inline_equation", + "content": "d = 2 \\times 128" + }, + { + "bbox": [ + 313, + 72, + 555, + 133 + ], + "type": "text", + "content": "). Among our methods, PS and VS achieve similar results, while APS and AVS underperform, likely due to their focus on exact-feature similarity and ambiguous-class features, which limits alignment with the high-level semantics required by ImageNet's diverse dataset." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 317, + 160, + 553, + 291 + ], + "blocks": [ + { + "bbox": [ + 313, + 137, + 555, + 159 + ], + "lines": [ + { + "bbox": [ + 313, + 137, + 555, + 159 + ], + "spans": [ + { + "bbox": [ + 313, + 137, + 555, + 159 + ], + "type": "text", + "content": "Table 7. Image Retrieval on revisited Oxford and Paris; reporting mAP " + }, + { + "bbox": [ + 313, + 137, + 555, + 159 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 313, + 137, + 555, + 159 + ], + "type": "text", + "content": ". Showing training set and method used. *mAP@100." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 160, + 553, + 291 + ], + "lines": [ + { + "bbox": [ + 317, + 160, + 553, + 291 + ], + "spans": [ + { + "bbox": [ + 317, + 160, + 553, + 291 + ], + "type": "table", + "html": "
Dataset →ROxfordRParis
ModelTrain SetMethodMHMH
ALEX+GeM [44]ImNet-1kPS33.810.452.726.0
RN101+R-MAC [44]ImNet-1kPS49.818.574.052.1
DINOv1-S/16 [2]ImNet-1kPS41.813.763.134.4
DINOv1-S/16 [2]GLDv2PS51.524.375.351.6
HgVT-Ti (ours)ImNet-1kVS26.810.555.428.7
VS*25.89.065.628.0
AVS*27.06.864.126.7
HgVT-S (ours)ImNet-1kVS28.012.156.731.1
VS*26.310.265.028.4
AVS*27.410.365.027.1
", + "image_path": "2a2b4870b4a188ee894bf0f2605c6877aa4259b988ac8995dd236b7f7240eb01.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 300, + 556, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 556, + 504 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 556, + 504 + ], + "type": "text", + "content": "Oxford and Paris Retrieval. To evaluate image retrieval performance beyond simple class retrieval, we use the Oxford and Paris Revisited datasets [39, 41], which provide three splits of increasing difficulty (Easy, Medium, and Hard) for query/database pairs. We report Mean Average Precision (mAP) for the Medium (M) and Hard (H) splits, using mAP@100 for AVS based on short-list ranking, while full mAP and mAP@100 are provided for VS as a baseline comparison. Results, shown in Tab. 7, indicate that HgVT achieves competitive performance with similarly sized feature extractors, though performance on ROxford-M lags. This shortfall may stem from subtle landmark differences better captured by multi-scale Conv-Nets and self-supervised learning, compared to the more salient focus driven by HgVT's classifier training. However, AVS outperforms VS on ROxford-M, demonstrating its ability to uncover finer feature similarities within the hypergraph structure." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 514, + 503, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 503, + 526 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 503, + 526 + ], + "type": "text", + "content": "7. Conclusion and Future Directions" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "text", + "content": "In this work, we introduced the Hypergraph Vision Transformer (HgVT), which combines hypergraph structures with vision transformers. Our methods, including diversity and population regularization, along with expert edge pooling, enhance semantic representation and efficiency by enabling dynamic hypergraph construction. Although challenging to tune, these methods avoid reliance on traditional clustering techniques, reflecting a tradeoff between complexity and efficiency. Empirical results show that HgVT achieves strong performance on both image classification and retrieval, positioning HgVT as a competitive framework for semantic vision tasks. Future work will focus on exploring the scalability of hypergraph structures and integrating self-supervised learning to further improve adaptability and better decouple saliency from semantic vision graph generation." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9758" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 297, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 297, + 133 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 297, + 133 + ], + "type": "text", + "content": "[1] Lucas Beyer, Olivier J. Henaff, Alexander Kolesnikov, Xiaohua Zhai, and Aaron van den Oord. Are we done with imageNet? arXiv preprint arXiv:2006.07159, abs/2006.07159, 2020. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 137, + 296, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 296, + 192 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 296, + 192 + ], + "type": "text", + "content": "[2] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jegou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9630-9640, 2021. 1, 2, 6, 7, 8, 4, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 194, + 296, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 194, + 296, + 258 + ], + "spans": [ + { + "bbox": [ + 62, + 194, + 296, + 258 + ], + "type": "text", + "content": "[3] Tianlong Chen, Zhenyu Zhang, Ajay Kumar Jaiswal, Shiwei Liu, and Zhangyang Wang. Sparse moe as the new dropout: Scaling dense and self-slimmable transformers. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 261, + 296, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 261, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 62, + 261, + 296, + 316 + ], + "type": "text", + "content": "[4] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 15" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 318, + 296, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 318, + 296, + 349 + ], + "spans": [ + { + "bbox": [ + 62, + 318, + 296, + 349 + ], + "type": "text", + "content": "[5] Norman Cliff. Dominance statistics: Ordinal analyses to answer ordinal questions. Psychological Bulletin, 114:494-509, 1993. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 352, + 296, + 416 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 352, + 296, + 416 + ], + "spans": [ + { + "bbox": [ + 62, + 352, + 296, + 416 + ], + "type": "text", + "content": "[6] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 14" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 420, + 296, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 420, + 296, + 474 + ], + "spans": [ + { + "bbox": [ + 62, + 420, + 296, + 474 + ], + "type": "text", + "content": "[7] Ekin Dogus Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V. Le. Randaugment: Practical automated data augmentation with a reduced search space. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 3008-3017, 2019. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 476, + 296, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 476, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 62, + 476, + 296, + 529 + ], + "type": "text", + "content": "[8] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. In *The Twelfth International Conference on Learning Representations*, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 533, + 296, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 533, + 296, + 576 + ], + "spans": [ + { + "bbox": [ + 62, + 533, + 296, + 576 + ], + "type": "text", + "content": "[9] Dmitry Demidov, M.H. Sharif, Aliakbar Abdurahimov, Hisham Cholakkal, and Fahad Shahbaz Khan. Salient mask-guided vision transformer for fine-grained classification. In VISIGRAPP, 2023. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 578, + 296, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 578, + 296, + 610 + ], + "spans": [ + { + "bbox": [ + 57, + 578, + 296, + 610 + ], + "type": "text", + "content": "[10] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. In CVPR09, 2009. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 613, + 296, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 296, + 656 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 296, + 656 + ], + "type": "text", + "content": "[11] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In North American Chapter of the Association for Computational Linguistics, 2019. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 658, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 296, + 713 + ], + "type": "text", + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 335, + 74, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 555, + 95 + ], + "type": "text", + "content": "scale. In International Conference on Learning Representations, 2021. 1, 2, 6, 7, 4, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 91, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 317, + 97, + 555, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 97, + 555, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 97, + 555, + 150 + ], + "type": "text", + "content": "[13] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The PASCAL Visual Object Classes Challenge 2012 (VOC2012) Results. http://www.pascalnetwork.org/challenges/VOC/voc2012/workshop/index.html. 14" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 152, + 555, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 152, + 555, + 195 + ], + "spans": [ + { + "bbox": [ + 317, + 152, + 555, + 195 + ], + "type": "text", + "content": "[14] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. J. Mach. Learn. Res., 23:120:1-120:39, 2022. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 198, + 555, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 555, + 240 + ], + "type": "text", + "content": "[15] Yue Gao, Meng Wang, Dacheng Tao, Rongrong Ji, and Qionghai Dai. 3-d object retrieval and recognition with hypergraph analysis. IEEE Transactions on Image Processing, 21(9): 4290-4303, 2012. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 243, + 555, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 555, + 285 + ], + "type": "text", + "content": "[16] Kai Han, Yunhe Wang, Jianyuan Guo, Yehui Tang, and Enhua Wu. Vision GNN: An image is worth graph of nodes. In Advances in Neural Information Processing Systems, 2022. 1, 2, 5, 6, 7, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 288, + 555, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 288, + 555, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 288, + 555, + 342 + ], + "type": "text", + "content": "[17] Yan Han, Peihao Wang, Souvik Kundu, Ying Ding, and Zhangyang Wang. Vision hgnn: An image is more than a graph of nodes. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 19878-19888, 2023. 1, 2, 3, 4, 5, 6, 7, 27" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 344, + 554, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 344, + 554, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 344, + 554, + 398 + ], + "type": "text", + "content": "[18] Yunusa Haruna, Shiyin Qin, Abdulrahman Hamman Adama Chukkol, Abdulganiyu Abdu Yusuf, Isah Bello, and Adamu Lawan. Exploring the synergies of hybrid cnns and vits architectures for computer vision: A survey. arXiv preprint arXiv:2402.02941, abs/2402.02941, 2024. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 399, + 554, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 399, + 554, + 443 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 554, + 443 + ], + "type": "text", + "content": "[19] Ali Hassani, Steven Walton, Jiachen Li, Shen Li, and Humphrey Shi. Neighborhood attention transformer. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 445, + 555, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 555, + 488 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 555, + 488 + ], + "type": "text", + "content": "[20] Kaiming He, X. Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2015. 1, 15" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 491, + 554, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 491, + 554, + 544 + ], + "spans": [ + { + "bbox": [ + 317, + 491, + 554, + 544 + ], + "type": "text", + "content": "[21] Elad Hoffer, Tal Ben-Nun, Itay Hubara, Niv Giladi, Torsten Hoefler, and Daniel Soudry. Augment your batch: Improving generalization through instance repetition. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8126-8135, 2020. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 545, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 553, + 567 + ], + "type": "text", + "content": "[22] Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Weinberger. Deep networks with stochastic depth, 2016. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 569, + 555, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 569, + 555, + 611 + ], + "spans": [ + { + "bbox": [ + 317, + 569, + 555, + 611 + ], + "type": "text", + "content": "[23] Yuchi Huang, Qingshan Liu, and Dimitris Metaxas. ]video object segmentation by hypergraph cut. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 1738-1745, 2009. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 614, + 555, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 614, + 555, + 658 + ], + "spans": [ + { + "bbox": [ + 317, + 614, + 555, + 658 + ], + "type": "text", + "content": "[24] Yuchi Huang, Qingshan Liu, Shaoting Zhang, and Dimitris N. Metaxas. Image retrieval via probabilistic hypergraph ranking. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 3376-3383, 2010. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 659, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 659, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 659, + 555, + 712 + ], + "type": "text", + "content": "[25] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. 15" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9759" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 138 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 138 + ], + "type": "text", + "content": "[26] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3992-4003, 2023. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 296, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 296, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 296, + 162 + ], + "type": "text", + "content": "[27] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, 2009. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 296, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 296, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 296, + 207 + ], + "type": "text", + "content": "[28] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2012. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 296, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 296, + 262 + ], + "type": "text", + "content": "[29] Aditya Kusupati, Gantavya Bhatt, Aniket Rege, Matthew Wallingford, Aditya Sinha, Vivek Ramanujan, William Howard-Snyder, Kaifeng Chen, Sham M. Kakade, Prateek Jain, and Ali Farhadi. Matryoshka representation learning. In Neural Information Processing Systems, 2022. 1, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 296, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 296, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 296, + 297 + ], + "type": "text", + "content": "[30] Y. Lecun, L. Bottou, Y. Bengio, and P. Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 299, + 296, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 299, + 296, + 353 + ], + "spans": [ + { + "bbox": [ + 56, + 299, + 296, + 353 + ], + "type": "text", + "content": "[31] Sangjun Lee, Inwoo Hwang, Gi-Cheon Kang, and Byoung-Tak Zhang. Improving robustness to texture bias via shape-focused augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 4323-4331, 2022. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 355, + 297, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 355, + 297, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 355, + 297, + 431 + ], + "type": "text", + "content": "[32] Benjamin Lefaudeau, Francisco Massa, Diana Liskovich, Wenhan Xiong, Vittorio Caggiano, Sean Naren, Min Xu, Jieru Hu, Marta Tintore, Susan Zhang, Patrick Labatut, Daniel Haziza, Luca Wehrstedt, Jeremy Reizenstein, and Grigory Sizov. xformers: A modular and hackable transformer modelling library. https://github.com/facebookresearch/xformers, 2022. 28" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 433, + 297, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 433, + 297, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 433, + 297, + 487 + ], + "type": "text", + "content": "[33] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9992-10002, 2021. 1, 2, 15" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 488, + 297, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 488, + 297, + 531 + ], + "spans": [ + { + "bbox": [ + 56, + 488, + 297, + 531 + ], + "type": "text", + "content": "[34] Leland McInnes, John Healy, Nathaniel Saul, and Lukas Großberger. Umap: Uniform manifold approximation and projection. Journal of Open Source Software, 3(29):861, 2018. 9" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 534, + 297, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 534, + 297, + 589 + ], + "spans": [ + { + "bbox": [ + 56, + 534, + 297, + 589 + ], + "type": "text", + "content": "[35] Mustafa Munir, William Avery, and Radu Marculescu. Mobilevig: Graph-based sparse attention for mobile vision applications. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 2211-2219, 2023. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 590, + 297, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 297, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 297, + 644 + ], + "type": "text", + "content": "[36] Mustafa Munir, William Avery, Md Mostafijur Rahman, and Radu Marculescu. Greedyvig: Dynamic axial graph construction for efficient vision gnns. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6118-6127, 2024. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 647, + 297, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 297, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 297, + 667 + ], + "type": "text", + "content": "[37] NVIDIA. Apex: A pytorch extension. https://github.com/NVIDIA/apex, 2020.28" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 670, + 297, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 297, + 713 + ], + "type": "text", + "content": "[38] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 72, + 555, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 72, + 555, + 138 + ], + "spans": [ + { + "bbox": [ + 333, + 72, + 555, + 138 + ], + "type": "text", + "content": "Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINOv2: Learning robust visual features without supervision. Transactions on Machine Learning Research, 2024. 1, 2, 7, 6, 8, 14, 15" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 140, + 555, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 555, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 555, + 195 + ], + "type": "text", + "content": "[39] James Philbin, Ondrej Chum, Michael Isard, Josef Sivic, and Andrew Zisserman. Lost in quantization: Improving particular object retrieval in large scale image databases. In 2008 IEEE Conference on Computer Vision and Pattern Recognition, pages 1-8, 2008. 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 197, + 555, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 555, + 239 + ], + "type": "text", + "content": "[40] Joan Puigcerver, Carlos Riquelme Ruiz, Basil Mustafa, and Neil Houlsby. From sparse to soft mixtures of experts. In The Twelfth International Conference on Learning Representations, 2024. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 241, + 555, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 555, + 296 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 555, + 296 + ], + "type": "text", + "content": "[41] Filip Radenovic, Ahmet Iscen, Giorgos Tolias, Yannis Avrithis, and Ondrej Chum. Revisiting oxford and paris: Large-scale image retrieval benchmarking. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5706-5715, 2018. 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 297, + 555, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 555, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 555, + 363 + ], + "type": "text", + "content": "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 8, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 365, + 555, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 555, + 407 + ], + "type": "text", + "content": "[43] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do imagenet classifiers generalize to imageNet? In International Conference on Machine Learning, 2019. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 409, + 555, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 409, + 555, + 463 + ], + "spans": [ + { + "bbox": [ + 316, + 409, + 555, + 463 + ], + "type": "text", + "content": "[44] Jérôme Revaud, Jon Almazán, Rafael Sampaio de Rezende, and César Roberto de Souza. Learning with average precision: Training image retrieval with a listwise loss. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 5106-5115, 2019. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 465, + 555, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 465, + 555, + 508 + ], + "spans": [ + { + "bbox": [ + 316, + 465, + 555, + 508 + ], + "type": "text", + "content": "[45] Peter J. Rousseeuw. Silhouettes: A graphical aid to the interpretation and validation of cluster analysis. Journal of Computational and Applied Mathematics, 20:53-65, 1987. 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 510, + 555, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 510, + 555, + 553 + ], + "spans": [ + { + "bbox": [ + 316, + 510, + 555, + 553 + ], + "type": "text", + "content": "[46] Franco Scarselli, Marco Gori, Ah Chung Tsoi, Markus Hagenbuchner, and Gabriele Monfardini. The graph neural network model. IEEE Transactions on Neural Networks, 20(1):61-80, 2009. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 555, + 555, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 555, + 555, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 555, + 555, + 578 + ], + "type": "text", + "content": "[47] Noam Shazeer. GLU variants improve transformer. arXiv preprint arXiv:2002.05202, abs/2002.05202, 2020. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 579, + 555, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 579, + 555, + 622 + ], + "spans": [ + { + "bbox": [ + 316, + 579, + 555, + 622 + ], + "type": "text", + "content": "[48] Dai Shi. Transnext: Robust foveal visual perception for vision transformers. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17773-17783, 2023. 14, 15" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 624, + 555, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 555, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 555, + 667 + ], + "type": "text", + "content": "[49] Dai Shi. Transnext: Robust foveal visual perception for vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17773-17783, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 670, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 555, + 713 + ], + "type": "text", + "content": "[50] Sakhinana Sagar Srinivas, Rajat Kumar Sarkar, Sreeja Gangasani, and Venkataramana Runkana. Vision HgNN: An electron-micrograph is worth hypergraph of hypernodes. arXiv preprint arXiv:2408.11351, abs/2408.11351, 2024. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "9760" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 296, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 296, + 116 + ], + "type": "text", + "content": "[51] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive multiview coding. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XI 16, pages 776-794. Springer, 2020. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 296, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 296, + 182 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 296, + 182 + ], + "type": "text", + "content": "[52] Hugo Touvron, Piotr Bojanowski, Mathilde Caron, Matthieu Cord, Alaaeldin El-Nouby, Edouard Grave, Armand Joulin, Gabriel Synnaeve, Jakob Verbeek, and Hervé Jégou. Resmlp: Feedforward networks for image classification with data-efficient training. arXiv preprint arXiv:2105.03404, abs/2105.03404, 2021. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 184, + 296, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 296, + 248 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 296, + 248 + ], + "type": "text", + "content": "[53] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herve Jegou. Training data-efficient image transformers amp; distillation through attention. In Proceedings of the 38th International Conference on Machine Learning, pages 10347-10357. PMLR, 2021. 1, 2, 5, 6, 7, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 250, + 296, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 296, + 325 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 296, + 325 + ], + "type": "text", + "content": "[54] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurélien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, abs/2302.13971, 2023. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 327, + 296, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 296, + 347 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 296, + 347 + ], + "type": "text", + "content": "[55] Asher Trockman and J. Zico Kolter. Patches are all you need? Trans. Mach. Learn. Res., 2023, 2023. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 350, + 296, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 296, + 403 + ], + "type": "text", + "content": "[56] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 404, + 296, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 296, + 446 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 296, + 446 + ], + "type": "text", + "content": "[57] Petar Velicković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. Graph attention networks. In International Conference on Learning Representations, 2018. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 449, + 296, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 296, + 480 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 296, + 480 + ], + "type": "text", + "content": "[58] Ross Wightman. Pytorch image models. https://github.com/rwrightman/pytorch-image-models, 2019.6, 28" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 482, + 296, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 482, + 296, + 523 + ], + "spans": [ + { + "bbox": [ + 56, + 482, + 296, + 523 + ], + "type": "text", + "content": "[59] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In European Conference on Computer Vision. Springer, 2018. 15" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 526, + 296, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 526, + 296, + 580 + ], + "spans": [ + { + "bbox": [ + 56, + 526, + 296, + 580 + ], + "type": "text", + "content": "[60] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Young Joon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6022-6031, 2019. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 582, + 296, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 296, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 296, + 635 + ], + "type": "text", + "content": "[61] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 1204-1213. IEEE, 2022. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 636, + 296, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 296, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 296, + 669 + ], + "type": "text", + "content": "[62] Biao Zhang and Rico Sennrich. Root mean square layer normalization. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 670, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 296, + 712 + ], + "type": "text", + "content": "[63] Bowen Zhang, Zhi Tian, Quan Tang, Xiangxiang Chu, Xiaolin Wei, Chunhua Shen, and Yifan liu. Segvit: Semantic segmentation with plain vision transformers. In Advances in Neural Information Processing Systems, 2022. 1" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 295 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "type": "text", + "content": "[64] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In International Conference on Learning Representations, 2018. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 555, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 555, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 555, + 162 + ], + "type": "text", + "content": "[65] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 6230-6239, 2016. 15" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 163, + 555, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 555, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 555, + 195 + ], + "type": "text", + "content": "[66] Zhun Zhong, Liang Zheng, Guoliang Kang, Shaozi Li, and Yi Yang. Random erasing data augmentation. AAAI, 34: 13001-13008, 2020. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 197, + 555, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 555, + 239 + ], + "type": "text", + "content": "[67] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision, 127:302 - 321, 2016. 14" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 241, + 555, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 555, + 295 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 555, + 295 + ], + "type": "text", + "content": "[68] Lei Zhu, Xinjiang Wang, Zhanghan Ke, Wayne Zhang, and Rynson Lau. Bifformer: Vision transformer with bi-level routing attention. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "9761" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Hyperspectral Pansharpening via Diffusion Models with Iteratively Zero-Shot Guidance/f79e03e7-ea12-4e70-879c-79258b8c5e82_content_list.json b/2025/Hyperspectral Pansharpening via Diffusion Models with Iteratively Zero-Shot Guidance/f79e03e7-ea12-4e70-879c-79258b8c5e82_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..19b0bd07c4e8ef3cbffe8e466ddfe6af3e450887 --- /dev/null +++ b/2025/Hyperspectral Pansharpening via Diffusion Models with Iteratively Zero-Shot Guidance/f79e03e7-ea12-4e70-879c-79258b8c5e82_content_list.json @@ -0,0 +1,1660 @@ +[ + { + "type": "text", + "text": "Hyperspectral Pansharpening via Diffusion Models with Iteratively Zero-Shot Guidance", + "text_level": 1, + "bbox": [ + 106, + 128, + 890, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jin-Liang Xiao $^{1,3}$ , Ting-Zhu Huang $^{1*}$ , Liang-Jian Deng $^{1}$ , Guang Lin $^{2,3}$ , Zihan Cao $^{1}$ , Chao Li $^{3}$ , Qibin Zhao $^{3*}$", + "bbox": [ + 99, + 202, + 947, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 University of Electronic Science and Technology of China, Chengdu, China", + "bbox": [ + 215, + 220, + 826, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Tokyo University of Agriculture and Technology, Tokyo, Japan", + "bbox": [ + 264, + 238, + 776, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ RIKEN AIP, Tokyo, Japan", + "bbox": [ + 406, + 256, + 633, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 308, + 326, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hyperspectral pansharpening refers to fusing a panchromatic image (PAN) and a low-resolution hyperspectral image (LR-HSI) to obtain a high-resolution hyperspectral image (HR-HSI). Recently, guiding pre-trained diffusion models (DMs) has demonstrated significant potential in this area, leveraging their powerful representational abilities while avoiding complex training processes. However, these DMs are often trained on RGB images, not well-suited for pansharpening tasks, limited in adapting to the hyperspectral images. In this work, we propose a novel guided diffusion scheme with zero-shot guidance and neural spatial-spectral decomposition (NSSD) to iteratively generate the RGB detail image and map the RGB detail image to target HR-HSI. Specifically, zero-shot guidance employs an auxiliary neural network that trained only with a PAN and LR-HSI to guide pre-trained DMs in generating the RGB detail image, informed by specific prior knowledge. Then, NSSD establishes a spectral mapping from the generated RGB detail image to the final HR-HSI. Extensive experiments are conducted on Pavia, Washington DC, Chukusei, and FR1 datasets to demonstrate that the proposed method significantly enhances the performance of DMs for hyperspectral pansharpening tasks, outperforming existing methods across multiple metrics and achieving improvements in visualization results. The code is available at https://github.com/Jin-liangXiao/DM-zs.", + "bbox": [ + 88, + 340, + 485, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 762, + 220, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hyperspectral images (HSIs), rich in spatial and spectral information, have been extensively used in a variety of applications [22, 46, 48]. In practice, however, the spatial and spectral information in HSIs are often divided across different image types: the panchromatic image (PANs) with high spatial resolution and low-resolution hyperspectral images", + "bbox": [ + 89, + 786, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b768e3fbca1ad5a62a06bdf83fbcf3164d5f2a554d9c1f9958669366b5633d83.jpg", + "image_caption": [ + "Figure 1. The upper row: illustration of the proposed method that includes iterative update of zero-shot guidance for pre-trained diffusion models and spectral mapping from RGB detail image to target HR-HSI. The bottom row: error map comparison between state-of-the-art pre-trained diffusion based hyperspectral pansharpening models (i.e., PLRDiff [30], HIR-Diff [29], and the proposed method)." + ], + "image_footnote": [], + "bbox": [ + 517, + 306, + 906, + 585 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(LR-HSIs) with abundant spectral information, due to hardware constraints. This separation highlights the need for hyperspectral pansharpening, which aims to fuse these image types to generate high-resolution hyperspectral images (HR-HSIs) [26].", + "bbox": [ + 511, + 718, + 906, + 794 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Methods for hyperspectral pansharpening generally fall into two categories: single-image approaches [10, 14, 16, 41, 47] and deep learning (DL) methods [4, 12, 13, 13, 15, 17, 18]. Single-image approaches require only a single image pair instead of an entire dataset for training, which can save significant training costs [36]. While single-image approaches are effective, DL methods can provide stronger", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation.", + "bbox": [ + 236, + 0, + 810, + 18 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "Except for this watermark, it is identical to the accepted version;", + "bbox": [ + 323, + 16, + 723, + 30 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 292, + 31, + 754, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 109, + 887, + 235, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12669", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "representations due to the powerful feature-extraction ability of deep neural networks. However, DL methods come with the trade-off of requiring intensive computational resources for training [7, 18, 37] and consistency in the dataset [40]. Specifically, if there is a distribution shift [8] between the training data and testing data, the performance can degrade significantly, which leads to poor generalization. This issue is particularly serious in hyperspectral pan-sharpening, where hyperspectral data are both more varying and limited than data in general computer vision tasks.", + "bbox": [ + 89, + 90, + 480, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Foundation models, particularly diffusion models (DMs) [21, 34], have gained wide attention in computer vision for their capability of generating high-quality images. In the context of hyperspectral pansharpening, DMs have shown improved generalization due to flexibility of the guidance techniques. Recent studies [29, 30] have demonstrated the potential of DMs, extending pre-trained DMs to hyperspectral pansharpening through specialized guidance. These methods utilize the degenerate process of pansharpening to guide the DM. Due to the powerful representational capabilities, pre-trained DM as a foundation model can be quickly transferred to hyperspectral pansharpening tasks, which can reduce a significant amount of training costs. However, since existing methods lack the consideration of specific prior knowledge of hyperspectral data, the guidance of DMs still needs improvement. The extension from pre-trained DMs to the hyperspectral pansharpening task is non-trivial: We need effective guidance to adapt the pre-trained DMs originally designed with only RGB channels, to manage the complex spectral fidelity while preserving high-resolution spatial details. This adaptation becomes even more challenging when we only have a single (PAN, LR-HSI) image pair. These challenges naturally give rise to two important questions: 1) How can we shift the data distribution of pre-trained DMs to HSIs for hyperspectral pansharpening with limited data through effective guidance? 2) How to accurately transform the RGB information of the generated images to HR-HSIs?", + "bbox": [ + 91, + 247, + 483, + 669 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the aforementioned challenges, we propose a novel guided diffusion scheme with zero-shot guidance and neural spatial-spectral decomposition (NSSD) to iteratively generate an RGB spatial detail image and map the detail image to the target HR-HSI. As shown in the green box of Figure 1, we utilize the pre-trained DMs as the backbone, with all their parameters being frozen. Then, we introduce the specific prior knowledge of testing data by a trainable zero-shot network to guide the DMs to generate the detail image enriched with expected features, as shown in the blue box. Unlike conventional deep neural networks that require extensive training data and complex training processes, the proposed zero-shot network is a lightweight model with few parameters, can be trained naively on a single testing image pair, i.e., a LR-HSI and PAN. This network encodes", + "bbox": [ + 89, + 674, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "specific prior knowledge into pre-trained DMs and guides the DMs to generate detail image more effectively (see Figure 3). Additionally, motivated by the tensor decomposition in single-image approaches [44], we recognize that spatial-spectral correlation is critical for the quality of final data and introduce a neural spatial-spectral decomposition (NSSD) to capture the spatial-spectral correlation and map from the RGB detail image to the final HR-HSI. The NSSD and the guidance of pre-trained DMs are updated iteratively, as displayed in the gray box of Figure 1, making NSSD adaptive to the results of guided DMs. Compared with existing pretrained DMs methods, the proposed method achieves the state-of-the-art result (refer to the bottom row of Figure 1).", + "bbox": [ + 511, + 90, + 903, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contributions are summarized as follows:", + "bbox": [ + 532, + 287, + 870, + 301 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a zero-shot network to efficiently guide the pre-trained DM to generate a detail image enriched with specific prior knowledge.", + "- We introduce a neural spatial-spectral decomposition (NSSD) to iteratively construct the target HR-HSI from the generated detail image.", + "- Experiments on the Pavia, Washington DC, Chikusei, and FR1 datasets demonstrate that the proposed method achieves the state-of-the-art performance." + ], + "bbox": [ + 511, + 304, + 903, + 439 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Preliminaries & Related Works", + "text_level": 1, + "bbox": [ + 511, + 455, + 800, + 470 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we provide definitions, notations, degenerate process of pansharpening, a brief introduction to score-based DMs with guidance, and related works.", + "bbox": [ + 511, + 482, + 903, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Notations", + "text_level": 1, + "bbox": [ + 511, + 539, + 624, + 554 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Throughout the paper, we denote tensors $\\mathcal{X} \\in \\mathbb{R}^{H \\times W \\times S}$ , $\\mathcal{Y} \\in \\mathbb{R}^{h \\times w \\times S}$ , $\\mathbf{P} \\in \\mathbb{R}^{H \\times W}$ as HR-HSI, LR-HSI, PAN, respectively. For tensor $\\mathcal{X}$ , $\\mathbf{X}_{(3)} \\in \\mathbb{R}^{S \\times HW}$ denotes its mode-3 unfolding [24]. Additionally, $\\mathcal{X} \\times_3 \\mathbf{R}$ indicates the mode-3 multiplication between tensor $\\mathcal{X}$ and matrix $\\mathbf{R}$ . $\\mathbf{x} \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})$ means the variable $\\mathbf{x}$ satisfies Gaussian distribution with the mean of $\\mathbf{0}$ and variance of matrix $\\mathbf{I}$ .", + "bbox": [ + 511, + 561, + 903, + 667 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Degenerate Process of Pansharpening", + "text_level": 1, + "bbox": [ + 511, + 680, + 836, + 696 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For the hyperspectral pansharpening task, inputs PAN and LR-HSI are mainly modeled as HR-HSI after the spectral and spatial degradation, respectively [45]. The degenerate process of the hyperspectral pansharpening problem can be mathematically represented as follows:", + "bbox": [ + 511, + 703, + 903, + 779 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {Y} = \\tau (\\mathcal {X}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 661, + 792, + 903, + 815 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} = \\mathcal {X} \\times_ {3} \\mathbf {R},\n$$\n", + "text_format": "latex", + "bbox": [ + 661, + 811, + 758, + 825 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $\\tau$ indicates blurring and down-sampling operations, and $\\mathbf{R} \\in \\mathbb{R}^{1 \\times S}$ denotes the spectral response function [43]. The goal of the task is to obtain the HR-HSI $\\mathcal{X}$ from the LR-HSI $\\mathcal{Y}$ and PAN $\\mathbf{P}$ .", + "bbox": [ + 511, + 840, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12670", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Score-based Diffusion Models with Guidance", + "text_level": 1, + "bbox": [ + 89, + 90, + 472, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion models (DMs) [21, 33] have proven to be a potent class of generative models capable of generating high-quality images. Among them, score-based DMs are a powerful family that models the forward process by stochastic differential equations (SDEs) and the reverse process by corresponding reverse-time SDEs [34]. Given a input $\\mathbf{x}$ , we have", + "bbox": [ + 89, + 112, + 483, + 217 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {d} \\mathbf {x} = f (\\mathbf {x}, t) \\mathrm {d} t + g (t) \\mathrm {d} \\mathbf {w}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 229, + 482, + 247 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {d} \\mathbf {x} = [ f (\\mathbf {x}, t) - g ^ {2} (t) \\nabla_ {\\mathbf {x}} \\log p _ {t} (\\mathbf {x}) ] + g (t) \\mathrm {d} \\bar {\\mathbf {w}}, \\quad (3)\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 268, + 482, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $t$ represents timestep, $f(\\cdot ,t)$ denotes a vector valued function called drift coefficient and $g(t)$ is a scalar valued function called diffusion term. $\\mathbf{w}$ and $\\bar{\\mathbf{w}}$ represent a Wiener process and a reverse-time Wiener process, respectively. Eq. (2), corresponding to the forward SDE, indicates adding noise from $t = 0$ to $T$ , which diffuses the image to pure Gaussian noise. In the reverse process, the corresponding reverse-time SDEs, i.e., Eq. (3), we utilize a learned score network $s_{\\theta}(\\mathbf{x}_t,t)\\coloneqq \\nabla_{\\mathbf{x}}\\log p_t(\\mathbf{x})$ to generate denoised sample $\\mathbf{x}_0$ from a simple Gaussian noise $\\mathbf{x}_T\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ . To more effectively enable DMs to generate the desired images, Dhariwal et al. and Pang et al. [9, 29] proposed the guided DM, which is a valid path for general specific tasks, leveraging the powerful generative capabilities of DMs without the need for complex retraining. The guidance of pre-trained DMs works by introducing gradient in reverse process, which can be represented as", + "bbox": [ + 89, + 297, + 483, + 554 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {s} _ {\\theta} (\\mathbf {x} _ {t}, t) = s _ {\\theta} (\\mathbf {x} _ {t}, t) + s (t) \\cdot \\nabla_ {\\mathbf {x} _ {t}} \\log p (y | \\mathbf {x} _ {t}), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 568, + 482, + 585 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $s(t)$ is the guidance strength in $t$ step, and $p(y|\\mathbf{x}_t)$ is the relationship between estimated $\\mathbf{x}_t$ and condition $y$ , which is crucial for the guidance of DMs on specific tasks.", + "bbox": [ + 89, + 597, + 483, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.4. Related Works", + "text_level": 1, + "bbox": [ + 89, + 652, + 241, + 666 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Leveraging the powerful generative capabilities of DMs, numerous studies have been conducted on hyperspectral pansharpening [5, 25, 39, 42, 50]. However, the performance of these models often depends on complex and time-consuming training processes and substantial training data. Recently, efforts have also been made to bridge the connection between inverse problems and pre-trained DMs. Kawar et al. [23] employed a pre-trained DM to solve linear inverse problems for RGB images. Since pre-trained DMs are primarily trained on RGB data, there remains a gap between the generated outputs and the target hyperspectral images. Miao et al. [27] further extended this approach to hyperspectral image restoration.", + "bbox": [ + 89, + 674, + 482, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For the hyperspectral pansharpening task, a nonlinear inverse problem, different loss functions are proposed to", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "guide DMs. Rui et al. [30] proposed a pre-trained diffusion framework that guides the diffusion process to generate an RGB detail image $\\mathcal{A}$ . This image is then transformed into the final pansharpened hyperspectral image by applying a fixed transformation matrix along the third dimension, as $\\mathcal{X} = \\mathcal{A} \\times_3 \\mathbf{E}$ , where $\\mathcal{X} \\in \\mathbb{R}^{H \\times W \\times S}$ is the target HR-HSI image, $\\mathcal{A} \\in \\mathbb{R}^{H \\times W \\times b}$ is the image generated by a pre-trained DM, and $\\mathbf{E} \\in \\mathbb{R}^{S \\times b}$ is the fixed transform matrix. Pang et al. [29] explore a new choice for estimating the transform matrix $\\mathbf{E}$ and apply total variation (TV) prior to diffusion guided function. However, this issue remains only partially resolved, as their fixed transformation matrix $\\mathbf{E}$ in the overall algorithm causes the final result to depend heavily on the initial estimate of $\\mathbf{E}$ . In contrast, our method introduces an auxiliary neural network to provide the specialized guidance for generating an RGB detail image and iteratively employs a neural spatial-spectral decomposition to adaptively construct the mapping from the RGB detail image to HR-HSI.", + "bbox": [ + 511, + 90, + 906, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 513, + 391, + 648, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we elaborate on the proposed guided diffusion shcheme. As illustrated in Figure 2, the scheme primarily consists of two components: diffusion with zero-shot guidance and neural spatial-spectral decomposition (NSSD). The zero-shot guidance operates on the pre-trained DM to generate an RGB detail image, enabling better preservation of spatial details and adaptation to specific prior knowledge. NSSD establishes the mapping from the detail image to the HR-HSI.", + "bbox": [ + 511, + 416, + 906, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Diffusion with Zero-Shot Guidance", + "text_level": 1, + "bbox": [ + 511, + 561, + 820, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unlike the existing DM-based hyperspectral pansharpening methods [29, 30], we aim to guide the reverse process of DMs for generating the RGB detail image $\\mathcal{A}$ by introducing the specific prior knowledge of testing data. To this end, we design a novel diffusion scheme with zero-shot guidance, where the reverse process of $\\mathcal{A}$ satisfies SDE as follows:", + "bbox": [ + 511, + 583, + 905, + 674 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathrm {d} \\mathcal {A} _ {t} = g (t) \\mathrm {d} \\bar {w} + f (\\mathcal {A}, t) \\mathrm {d} t \\\\ - g ^ {2} (t) \\left[ s _ {\\theta} \\left(x _ {t}, t\\right) + \\nabla_ {\\mathcal {A} _ {t}} \\log p _ {t} (\\tilde {\\mathcal {Y}} | \\mathcal {A} _ {t}) \\right] \\mathrm {d} t, \\tag {5} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 685, + 903, + 723 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $p_t(\\tilde{\\mathcal{Y}}|\\mathcal{A}_t)$ is the pansharpening guidance estimated from the condition, defined as", + "bbox": [ + 511, + 734, + 906, + 766 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathcal {Y}} = (\\mathcal {Y}, \\mathbf {P}, \\mathcal {M}, f _ {\\psi} (\\mathcal {Y}, \\mathbf {P})), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 779, + 906, + 797 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{V}$ and $\\mathbf{P}$ are input image pair LR-HSI and PAN, $\\mathcal{M}$ is a mapping from RGB detail image to HR-HSI (deferred in next subsection), and $f_{\\psi}(\\mathcal{V},\\mathbf{P})$ is an auxiliary neural network that is trained with $\\mathcal{V}$ and $\\mathbf{P}$ . Following [9], we have", + "bbox": [ + 511, + 809, + 905, + 869 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta , \\psi} \\left(\\mathcal {A} _ {t - 1} \\mid \\mathcal {A} _ {t}, \\tilde {\\mathcal {Y}}\\right) \\propto p _ {\\theta} \\left(\\mathcal {A} _ {t - 1} \\mid \\mathcal {A} _ {t}\\right) p _ {\\psi} (\\tilde {\\mathcal {Y}} \\mid \\mathcal {A} _ {t}). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 882, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12671", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/fcc997ad0e0834c3612cc233579d7b350ce28c7a39fdf0d1cfd930cd8615fb14.jpg", + "image_caption": [ + "Figure 2. Illustration of the whole process of the proposed hyperspectral pansharpening method that includes diffusion with zero-shot (ZS) guidance and neural spatial-spectral decomposition (NSSD). The part in the dashed box is the proposed zero-shot guided function that is trained with a single testing image pair. NSSD builds the spectral mapping from the detail image generated by DM to HR-HSI. The guidance of the DM and NSSD are iteratively updated." + ], + "image_footnote": [], + "bbox": [ + 174, + 89, + 823, + 296 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/93f32f0d57dbd6cf88c0f002b0fd4207fc0e1b02cf9311977ef3ddb57173ae1c.jpg", + "image_caption": [ + "Figure 3. The analysis of the proposed zero-shot guidance. The DMs with the proposed guidance can achieve better performance with fewer sampling steps. (The length of the yellow box means the distance of PSNR values between the DM with and without ZS guidance)" + ], + "image_footnote": [], + "bbox": [ + 132, + 392, + 444, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on [6], we define the variable $\\hat{\\mathcal{A}}_0\\coloneqq \\mathbb{E}[\\mathcal{A}_0|\\mathcal{A}_t] =$ $\\mathbb{E}_{\\mathcal{A}_0\\sim p(\\mathcal{A}_0|\\mathcal{A}_t)}[\\mathcal{A}_0]$ , which means", + "bbox": [ + 89, + 638, + 482, + 674 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {A}} _ {0} \\simeq \\frac {1}{\\sqrt {\\bar {\\alpha} _ {t}}} \\left(\\mathcal {A} _ {t} + \\left(1 - \\bar {\\alpha} _ {t}\\right) s _ {\\theta} \\left(\\mathcal {A} _ {t}, t\\right)\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 688, + 482, + 722 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\psi} (\\tilde {\\mathcal {V}} | \\mathcal {A} _ {t}) \\simeq p _ {\\psi} (\\tilde {\\mathcal {V}} | \\hat {\\mathcal {A}} _ {0}), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 202, + 744, + 482, + 766 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\bar{\\alpha}_t\\coloneqq \\prod_{s = 1}^t\\alpha_s$ is the noise scale of DMs. We propose the zero-shot guidance as follows,", + "bbox": [ + 89, + 780, + 482, + 811 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\log p _ {\\psi} (\\tilde {\\mathcal {Y}} | \\hat {\\mathcal {A}} _ {0}) = \\left\\| \\tau (\\mathcal {M} (\\hat {\\mathcal {A}} _ {0})) - \\mathcal {Y} \\right\\| _ {F} ^ {2} \\\\ + \\left\\| \\mathcal {M} \\left(\\hat {\\mathcal {A}} _ {0}\\right) \\times_ {3} \\mathbf {R} - \\mathbf {P} \\right\\| _ {F} ^ {2} + \\lambda_ {1} \\left\\| \\mathcal {M} \\left(\\hat {\\mathcal {A}} _ {0}\\right) - f _ {\\psi} (\\mathcal {Y}, \\mathbf {P}) \\right\\| _ {F} ^ {2}, \\tag {10} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 828, + 488, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda_{1}$ is a weighting scale. Finally, the reverse process is illustrated as:", + "bbox": [ + 511, + 390, + 903, + 419 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {A} _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\mathcal {A} _ {t} - \\frac {\\beta_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} s _ {\\theta} \\left(\\mathcal {A} _ {t}, t\\right)\\right) \\tag {11} \\\\ - \\lambda \\cdot \\frac {\\beta_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}} \\sqrt {\\alpha_ {t}}} \\nabla_ {\\mathcal {A} _ {t}} l o g p _ {\\psi} (\\tilde {\\mathcal {Y}} | \\hat {\\mathcal {A}} _ {0}), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 426, + 903, + 496 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\beta_{t}$ , as a hyperparameter, is generally set to linearly interpolate from 0.0001 to 0.02 [28] and $\\alpha_{t} := 1 - \\beta_{t}$ . The weighting scale $\\lambda$ can balance the effect of guidance.", + "bbox": [ + 511, + 507, + 903, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The rest of this part is how to train the zero-shot network $f_{\\psi}(\\mathcal{V}, \\mathbf{P})$ . The network structure and training process are shown within the dotted box in Figure 2, which has been verified its effectiveness in [11]. Based on the \"invariance among scales\" assumption, we train the network in the reduced-resolution scale. In specific, we use the input LR-HSI as the training ground-truth. Training PAN and training LR-HSI are simulated by PAN and LR-HSI after spatial and spectral degenerate processes, respectively. This approach allows for efficient training of the zero-shot network using only a single testing input pair. Figure 3 shows the change curve of PSNR values with the increase of iteration number. One can observe that the DM guided by zero-shot guidance can achieve better performance with fewer sampling steps. Besides, this single-image training property provides specific prior knowledge to the pre-trained DM.", + "bbox": [ + 511, + 551, + 906, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Neural Spatial-Spectral Decomposition", + "text_level": 1, + "bbox": [ + 511, + 801, + 849, + 819 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This subsection describes neural spatial-spectral decomposition (NSSD), i.e., $\\mathcal{M}$ in Eq. (10). In hyperspectral pan-sharpening, beyond the fidelity term in (1), the spatial-spectral correlation exhibits the low-rank properties characteristic of hyperspectral data [44]. The mode-3 decom", + "bbox": [ + 511, + 825, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12672", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ac02864a374f0b054b5152395b5a575ed2d6a20fcd90661d03ec33e8054a09b3.jpg", + "image_caption": [ + "Figure 4. The structure of neural spatial-spectral decomposition (NSSD) consists of a U-net structure with skip connection and the fully connection structure to extract spatial and spectral information, respectively." + ], + "image_footnote": [], + "bbox": [ + 215, + 88, + 782, + 252 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "position format, i.e., $\\mathcal{X} = \\mathcal{A}\\times_3\\mathbf{E}$ , has shown to be effective [19, 20, 31], which satisfies the following inequation:", + "bbox": [ + 89, + 316, + 480, + 349 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {r a n k} \\left(\\mathbf {X} _ {(3)}\\right) \\leq \\min \\left\\{\\operatorname {r a n k} \\left(\\mathbf {A} _ {(3)}\\right), \\operatorname {r a n k} (\\mathbf {E}) \\right\\} \\leq b, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 359, + 482, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{A} \\in \\mathbb{R}^{H \\times W \\times b}$ , and $\\mathbf{E} \\in \\mathbb{R}^{S \\times b}$ . The low value of the size of $\\mathbf{E}$ limits the expressive ability. Since pre-trained DMs are primarily trained on RGB images, the size of $\\mathcal{A}$ is typically set to $H \\times W \\times 3$ . Additionally, the larger value of $r$ in (12) inevitably increases the time required for the reverse process in the DM.", + "bbox": [ + 89, + 386, + 483, + 477 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To mitigate the drawbacks, we propose a NSSD to capture the spatial-spectral correlation, as follows:", + "bbox": [ + 89, + 479, + 483, + 508 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {X} = \\mathcal {M} (\\mathcal {A}) := \\mathcal {C} _ {\\theta_ {1}} (\\mathcal {A}) \\times_ {3} \\mathbf {D} _ {\\theta_ {2}} (\\mathcal {Y}), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 521, + 482, + 540 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{C}_{\\theta_1}(\\mathcal{A}) \\in \\mathbb{R}^{H \\times W \\times r}$ and $\\mathbf{D}_{\\theta_2}(\\mathcal{Y}) \\in \\mathbb{R}^{S \\times r}$ with hyperparameter $r$ are produced by two neural network with the input $\\mathcal{A} \\in \\mathbb{R}^{H \\times W \\times 3}$ and $\\mathcal{Y} \\in \\mathbb{R}^{h \\times w \\times S}$ , respectively, as illustrated in Figure 4. Specifically, $\\mathcal{C}_{\\theta_1}(\\cdot)$ is the spatial-factor network that extracts the spatial information from the RGB detail image. The structure of $\\mathcal{C}_{\\theta_1}(\\cdot)$ is the U-net type with skip connection, whose detail-extraction capability has been verified [35]. The spectral-factor network, $\\mathbf{D}_{\\theta_2}(\\cdot)$ , consists of the fully connection structure and sigmoid to efficiently absorb spectral information from LR-HSI. NSSD does not require pre-training. It produces the final result through iterative steps using the following loss:", + "bbox": [ + 89, + 550, + 483, + 733 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L o s s = \\left\\| \\tau \\left(\\mathcal {C} _ {\\theta_ {1}} \\times_ {3} \\mathbf {D} _ {\\theta_ {2}}\\right) - \\mathcal {Y} \\right\\| _ {F} ^ {2} \\tag {14} \\\\ + \\lambda_ {2} \\| \\left(\\mathcal {C} _ {\\theta_ {1}} \\times_ {3} \\mathbf {D} _ {\\theta_ {2}}\\right) \\times_ {3} \\mathbf {R} - \\mathbf {P} \\| _ {F} ^ {2}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 743, + 482, + 785 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{2}$ is a balanced parameter. NSSD models the relationship between the generated detail image and the target HR-HSI. Compared with the classic format $\\mathcal{X} = \\mathcal{A} \\times_3 \\mathbf{E}$ , the proposed form $\\mathcal{C}_{\\theta_1}(\\mathcal{A}) \\times_3 \\mathbf{D}_{\\theta_2}(\\mathcal{Y})$ offers a more flexible representation for HR-HSI, particularly for data with higher rank. Furthermore, the NSSD evolves in the guidance of pre-trained DMs, as shown in Eq. (10). The guidance of the", + "bbox": [ + 89, + 794, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1: The Whole Solving Process" + ], + "code_body": "Input: $\\mathcal{A}_T$ sampled from $\\mathcal{N}(0,I)$ ,DM $s_\\theta$ noise scales $\\{\\bar{\\alpha}_t\\}_{t = 1}^T$ $f_{\\psi}$ ,LR-HSI Y,PAN P \nOutput: HR-HSI X \nwhile not converged do for $t = T,T - 1,\\dots 1$ do step 1: estimate $\\hat{\\mathcal{A}}_0$ by Eq. (8) step 2: compute log $p_{\\psi}(\\mathcal{V}|\\mathcal{A}_t)$ by Eq. (10) step 3: sample $\\mathcal{A}_{t - 1}$ by Eq. (11) end NSSD Update: update $\\mathcal{M}$ by minimizing (14) \nend $\\mathcal{X} = \\mathcal{C}_{\\theta_1}(\\mathcal{A})\\times_3\\mathbf{D}_{\\theta_2}(\\mathcal{Y})$ \nreturn X", + "bbox": [ + 526, + 338, + 874, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "DMs and the update of NSSD are mutually adaptive. Thus, NSSD provides a more flexible and DM-compatible representation of the spatial-spectral correlation in HR-HSI.", + "bbox": [ + 511, + 571, + 903, + 618 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Whole Solving Algorithm: Since the guidance of pretrained DM is related to the mapping $\\mathcal{M}$ , the generated detail image is adaptive to the mapping $\\mathcal{M}$ . To enhance the flexibility of mapping and the suitability of diffusion guidance, we alternatively conduct guidance of pre-trained DM and NSSD update as shown in Figure 2. During each guidance step, the parameters of the NSSD are held fixed. The whole solving algorithm is summarized in Algorithm 1.", + "bbox": [ + 511, + 619, + 906, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 753, + 645, + 771 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To demonstrate the effectiveness of the proposed method, we compare it with corresponding state-of-the-art methods, including GSA [1], CNMF [47], HySure [32], ZSL [11], PLRDiff [30], and HIR-Diff [29]. It is worth to be remarked that HIR-Diff [29] is a general inverse problem framework. We rewrite the fidelity for the hyperspectral pansharpening task. The parameters of all methods are adjusted according to the authors' recommendation. The metrics peak", + "bbox": [ + 511, + 779, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12673", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f7f2fd480b50fd617a25808e4186208f0112c88d0c4b54179e9188276ae04a01.jpg", + "image_caption": [ + "Figure 5. Visual comparisons on Pavia dataset. The first row shows the predicted HR-HSIs of different methods, and the second row displays the error maps between predicted HR-HSIs and ground-truth." + ], + "image_footnote": [], + "bbox": [ + 174, + 92, + 823, + 220 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "signal-to-noise ratio (PSNR), the structural similarity index (SSIM) [38], the relative dimensionless global error in synthesis (ERGAS) [2], the spectral angle mapper (SAM) [49], and root mean square error (RMSE) are utilized for assessment. QNR [3] is employed for evaluating real-world data. We assess the results produced by these methods on Pavia1, Washington DC (WDC)2, Chikusei3, and FR1 datasets4. The simulated LR-HSI is gained by applying a $5 \\times 5$ Gaussian blurring kernel and the downsampling factor 4. The simulated PAN image is obtained from groundtruth by multiplying a spectral mean response vector. The downsampling factor of FR1 dataset is 6. The detailed information about these datasets is listed as follows,", + "bbox": [ + 88, + 287, + 482, + 483 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Pavia dataset: Pavia dataset are originally obtained with the size of $610 \\times 340 \\times 115$ . We choose 93 bands by eliminating the bands associated with water vapor absorption and crop the up-left cube with size $256 \\times 256 \\times 93$ .", + "- WDC dataset: WDC dataset is obtained from the Hyperspectral Digital Imagery Collection Experiment (HYDICE) sensor. We eliminate some unusable bands and choose the area with size $256 \\times 256 \\times 191$ .", + "- Chikusei dataset: The Headwall Hyperspec-VNIR-C imaging sensor captures the Chikusei dataset with 128 bands and the spectral range is from 363 to $1018\\mathrm{~nm}$ . We crop a subimage with size $256 \\times 256 \\times 128$ .", + "- FR1 real dataset: The FR1 dataset is a real-world data distributed for the PRISMA contest, for pansharpening at the full spatial resolution. We crop a subimage with PAN size $240 \\times 240$ ." + ], + "bbox": [ + 89, + 484, + 482, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 734, + 307, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "All experiments are conducted on the same PC with Pytorch 2.0 and MATLAB (R2022a) on a computer of 64Gb RAM, Intel(R) Core(TM) i9-10900KF CPU @3.70GHz, and NVIDIA GeForce RTX 4070. The update of NSSD", + "bbox": [ + 88, + 756, + 482, + 816 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1 https://rslab.ut.ac.ir/data", + "2 https://engineering.purdue.edu/biehl/MultiSpec/hyperspectral.html", + "3 https://naotoyokoya.com/Download.html", + "4 https : / / github . com / liangjiandeng / HyperPanCollection" + ], + "bbox": [ + 78, + 825, + 480, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "is optimized by Adam optimizer with 0.001 learning rate (LR). The balanced parameter of guidance $\\lambda_{1}$ is 0.1, and $\\lambda_{2} = 1$ . The diffusion reverse process step number is set to 100. The number of the whole iterations is 20000. The training of the zero-shot neural network is with Adam optimizer with 2000 iterations, and the LR is 0.001 decayed by gamma $= 0.8$ each 1000 iterations. The rank of NSSD for Pavia, WDC, and Chikusei datasets are set to 20, 40, and 40, respectively. For FR1 real dataset, $\\lambda_{1} = \\lambda_{2} = 10$ , and the number of the whole iterations is 10000.", + "bbox": [ + 511, + 287, + 906, + 438 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c8a10cf9d6459bfe580b45cbacfb3d01767069d249b69d204754670f92916ea0.jpg", + "table_caption": [ + "Table 1. Quantitative results for Pavia dataset. (Bold: best; Underline: second best)" + ], + "table_footnote": [], + "table_body": "
MethodPSNR↑SSIM↑ERGAS↓SAM↓RMSE↓
GSA [1]07*TGRS28.3040.8125.8197.75310.368
CNMF [47] 11*TGRS28.8960.8595.4955.9499.558
HySure [32] 14*TGRS25.8470.7587.86110.05914.065
ZSL [11]23'TPAMI32.0600.9053.5495.3186.819
PLRDiff [30] 24'IF33.0600.9053.2915.3346.001
HIR-Diff [29]24'CVPR31.9740.8753.6184.7056.742
Proposed34.2270.9262.8634.4635.284
Ideal value+∞1000
", + "bbox": [ + 514, + 488, + 911, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Results on Pavia Dataset", + "text_level": 1, + "bbox": [ + 511, + 651, + 738, + 666 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this part, we evaluate the effectiveness of the proposed method on the Pavia dataset. Figure 5 shows the visual comparison of these methods in the benchmark. Diffusion-based methods, i.e., PLRDiff [30], HIR-Diff [29], and the proposed method, achieve excellent performance for detail extraction. The residual images, i.e., the second row of Figure 5 obviously verify that our method can preserve more details compared to others. The result produced by HySure [32] leads to spatial distortions. Table 1 provides the quantitative results for compared approaches. For all metrics, the proposed method achieves the best performance.", + "bbox": [ + 511, + 672, + 906, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Results on WDC Dataset", + "text_level": 1, + "bbox": [ + 511, + 848, + 740, + 863 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the WDC dataset, there exist more challenges to preserve the spectral information due to its more spectral", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12674", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3343eb19e5774ef750d27c553b0df5666be8552de592b57c9c57c96a60e79c72.jpg", + "image_caption": [ + "Figure 6. Visual comparisons on WDC dataset. The first row shows the predicted HR-HSIs of different methods, and the second row displays the error maps between predicted HR-HSIs and ground-truth." + ], + "image_footnote": [], + "bbox": [ + 174, + 90, + 823, + 220 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9fb2a19da9eb27ccbdf2baff1f5750f2bdf698789dd63b5822ce4babe14d7e45.jpg", + "image_caption": [ + "Figure 7. Visual comparisons on Chikusei dataset. The first row shows the predicted HR-HSIs of different methods, and the second row displays the error maps between predicted HR-HSIs and ground-truth." + ], + "image_footnote": [], + "bbox": [ + 174, + 276, + 823, + 406 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/845b79e5c9016d3e351d750813349c9df70aaf43526ac100f78768110c5aa74c.jpg", + "table_caption": [ + "Table 2. Quantitative results for WDC dataset. (Bold: best; Underline: second best)" + ], + "table_footnote": [], + "table_body": "
MethodPSNR↑SSIM↑ERGAS↓SAM↓RMSE↓
GSA [1]07*TGRS39.6290.9105.9247.0797.665
CNMF [47] 11*TGRS39.7030.8975.8646.4487.665
HySure [32] 14*TGRS35.0710.7989.92611.89812.928
ZSL [11]23*TPAMI39.6850.9166.3357.8257.301
PLRDiff [30] 24*IF34.7990.83318.2029.5448.667
HIR-Diff [29]24*CVPR38.5060.8447.1026.6868.079
Proposed40.5030.94312.2545.0474.631
Ideal value+∞1000
", + "bbox": [ + 91, + 508, + 488, + 645 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2f8e8dc17a7e342c2a8e0b85751970a39d3dd9a6264de5e5d7fea5e944d9a402.jpg", + "table_caption": [ + "Table 3. Quantitative results for Chikusei dataset. (Bold: best; Underline: second best)" + ], + "table_footnote": [], + "table_body": "
MethodPSNR↑SSIM↑ERGAS↓SAM↓RMSE↓
GSA [1]07*TGRS32.4590.88710.8886.7626.662
CNMF [47] 11*TGRS33.5780.8788.5924.8935.905
HySure [32] 14*TGRS30.2700.75214.1599.8838.417
ZSL [11]23'TPAMI35.2580.9178.3675.6854.632
PLRDiff [30] 24'IF37.1760.9176.9944.8383.651
HIR-Diff [29]24'CVPR36.3890.9056.3333.8944.247
Proposed38.4500.9395.9454.1973.192
Ideal value+∞1000
", + "bbox": [ + 91, + 703, + 488, + 840 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "bands. As shown in Table 2, we provide the visual results of different methods. Our approach also performs best on", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "this dataset. Compared with traditional method CNMF [47], diffusion-based methods PLRDiff [30] and HIR-Diff [29] lead to spectral distortions. As demonstrated in Figure 6, PLRDiff [30] also preserves excellent spatial detail. The proposed method has the advantages of both spatial detail preservation and spectral information extraction.", + "bbox": [ + 511, + 470, + 906, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Results on Chikusei Dataset", + "text_level": 1, + "bbox": [ + 511, + 573, + 764, + 587 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 7 evaluates the effectiveness of the proposed method on the Chikusei dataset. Our method achieves the best result due to the spatial detail preservation. For the detail edge of the image, the proposed method can preserve the spatial detail well. As shown in Table 3, the proposed method demonstrates improvements in PSNR by $1.274\\mathrm{dB} / 2.061\\mathrm{dB}$ , respectively, compared with diffusion-based methods PLRDiff [30] and HIR-Diff [29].", + "bbox": [ + 511, + 595, + 906, + 717 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 726, + 666, + 742 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To demonstrate the effectiveness of the proposed zero-shot guidance and NSSD, we conduct the ablation study on the Pavia dataset. Table 4 shows the results produced by Four models. For the model I, we use the fixed transformation in [30] and regular fidelity to substitute the proposed zero-shot guidance and NSSD, respectively. As shown in Table 4, the results produced by the model with both zero-shot guidance and NSSD are best. Besides, we design more detailed experiments to explore more properties of the two parts.", + "bbox": [ + 511, + 750, + 906, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Zero-shot Guidance: The zero-shot guidance achieves", + "bbox": [ + 532, + 885, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12675", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/779bee8a44ea5cb7c71cdd0cecb10963fc66a3e51134ead4340ec0cfe3e1d7f3.jpg", + "table_caption": [ + "Table 4. Ablation study about the proposed zero-shot (ZS) guidance and neural spatial-spectral decoder (NSSD) on Pavia dataset." + ], + "table_footnote": [], + "table_body": "
AblationZSNSSDPSNRSSIMERGASSAMRMSE
Model I✗✗33.0600.9053.2915.3346.001
Model II✓✗33.2670.9123.2195.2855.877
Model III✗✓33.9720.9212.9464.6055.460
Proposed✓✓34.1790.9252.8774.5335.332
", + "bbox": [ + 91, + 128, + 480, + 204 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7afca9c5e68c05a388f3c312d7e8a110f9e6a28c4987f46857b19e2fa81d142a.jpg", + "image_caption": [ + "Figure 8. Singular value distributions of results produced by PLRDiff [30], HIR-Diff [29], and the proposed method." + ], + "image_footnote": [], + "bbox": [ + 189, + 215, + 385, + 335 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/29b444f201639db0de3bb2ce109cbe60e2201cdcf377958214cba38ab6309a1a.jpg", + "table_caption": [ + "Table 5. Performance of proposed zero-shot guidance compared with conventional regularization term." + ], + "table_footnote": [], + "table_body": "
GuidancePSNRSSIMERGASSAMRMSE
Fidelity32.4700.8963.4555.1146.395
TV32.7700.9023.3565.1566.204
Proposed33.3740.9143.1745.1505.793
", + "bbox": [ + 127, + 426, + 444, + 497 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "excellent guided performance. To demonstrate the effectiveness, we compare it with the conventional total variation regularized term, which is widely used and effective for image processing tasks. As shown in Table 5, the regularized term is helpful for the guidance of DMs. The proposed zero-shot guidance offers more specific prior knowledge for the pre-trained DM to achieve the best performance.", + "bbox": [ + 88, + 521, + 482, + 626 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "NSSD: NSSD describes the spatial-spectral correlation with the low-rank property. Figure 8 shows the singular value distribution of target HR-HSI produced by different methods. We compare the proposed method with NSSD with two fixed Diffusion transformation estimation methods in PLRDiff [30] and HIR-Diff [29]. One can observe that the proposed NSSD provides a flexible transformation and performs the best approximation with the pre-trained DM guidance. Our method can achieve the closest distance between the singular values of ground-truth.", + "bbox": [ + 88, + 627, + 482, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Guided Network", + "text_level": 1, + "bbox": [ + 89, + 787, + 256, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this part, we analyze the influence of zero-shot network structure for guided performance. Table 6 shows the final results with U-Net with skip-connections [35], DSNet [51], and CONet [11]. U-Net contains multi-scale information, DSNet is a structure used in the supervised method, and CONet is a light convolution-based cascade structure.", + "bbox": [ + 89, + 809, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/65365336191c483b322dd090147a2f501f5ad7383a5323a525f0dae189cb4a26.jpg", + "table_caption": [ + "Table 6. Performance of zero-shot guided network with different network structures." + ], + "table_footnote": [], + "table_body": "
Guided StructurePSNRSSIMERGASSAMRMSE
U-Net34.0150.9212.9534.5115.426
DSNet33.7750.9173.0334.6775.609
CONet34.2270.9262.8634.4635.284
", + "bbox": [ + 519, + 128, + 897, + 199 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The proposed method with the above network structures can achieve excellent results. The convolution-based cascade structure achieves the best performance.", + "bbox": [ + 511, + 223, + 906, + 268 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.7. More Discussions", + "text_level": 1, + "bbox": [ + 511, + 277, + 684, + 292 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Real-data Experiment: Table 7 shows the evaluation on the FR1 real-world dataset, which demonstrates the superior performance of the proposed method on real data.", + "bbox": [ + 511, + 300, + 906, + 345 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/284c3279f22065d1a24402d04bba9350471cf314ab309a861719b401754856bb.jpg", + "table_caption": [ + "Table 7. Real data experiment on FR1 dataset." + ], + "table_footnote": [], + "table_body": "
MethodGSACNMFHySureZSLPLRDiffHIR-DiffProposed
QNRT0.5900.6790.6680.8020.6300.6250.832
", + "bbox": [ + 514, + 383, + 897, + 412 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "More Experimental Comparisons: We provide further discussion on running time, downsampling factor, and kernel size, which are shown in the supplementary material.", + "bbox": [ + 511, + 426, + 905, + 472 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations: While the computational cost of DMs with the proposed zero-shot guidance is comparable to that of previous DM-based methods, such as PLRDiff [30], NSSD requires more time to iteratively construct the spectral mapping from the RGB detail image to the HR-HSI, as detailed in the supplementary material. We will explore the use of pre-trained networks to address this limitation in the future.", + "bbox": [ + 511, + 472, + 905, + 579 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 590, + 633, + 607 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduced a novel guided diffusion scheme for hyperspectral pansharpening, combining the iterative update of zero-shot guidance and NSSD. The zero-shot guidance is conducted with an auxiliary convolutional network that can be trained just on one-pair inputs, while NSSD acts as a flexible transformation mapping the generated images by pre-trained DMs to the hyperspectral images. Our method demonstrated significant effectiveness on benchmark datasets, including Pavia, WDC, and Chikusei. Through extensive ablation studies, we analyzed the contributions of the guidance and NSSD and evaluated the impact by different zero-shot network architectures.", + "bbox": [ + 511, + 617, + 906, + 799 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement This work was supported in part by the NSFC (12171072, 12271083, 62273077), Natural Science Foundation of Sichuan Province (2024NSFSC0038), and National Key Research and Development Program of China (2020YFA0714001).", + "bbox": [ + 511, + 816, + 906, + 892 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12676", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] B. Aiazzi, S. Baronti, and M. Selva. Improving component substitution pansharpening through multivariate regression of MS + Pan data. IEEE Trans. Geosci. Remote Sens., 45 (10):3230-3239, 2007. 5, 6, 7", + "[2] L. Alparone, L. Wald, J. Chanussot, C. Thomas, P. Gamba, and L. M. Bruce. Comparison of pansharpening algorithms: Outcome of the 2006 GRS-S data-fusion contest. IEEE Trans. Geosci. Remote Sens., 45(10):3012-3021, 2007. 6", + "[3] L. Alparone, B. Aiazzi, S. Baronti, A. Garzelli, F. Nencini, and M. Selva. Multispectral and panchromatic data fusion assessment without reference. Photogramm. Eng. Remote Sens., 74(2):193-200, 2008. 6", + "[4] W. G. C. Bandara, J. M. J. Valanarasu, and V. M. Patel. Hyperspectral pansharpening based on improved deep image prior and residual reconstruction. EEE Trans. Geosci. Remote Sens., 60:1-16, 2021. 1", + "[5] Z. Cao, S. Cao, L.-J. Deng, X. Wu, J. Hou, and G. Vivone. Diffusion model with disentangled modulations for sharpening multispectral and hyperspectral images. Inf. Fusion, 104: 102158, 2024. 3", + "[6] H. Chung, J. Kim, M. T. McCann, M. L. Klasky, and J. C. Ye. Diffusion posterior sampling for general noisy inverse problems. arXiv preprint arXiv:2209.14687, 2022. 4", + "[7] M. Ciotola, G. Guarino, G. Vivone, G. Poggi, J. Chanussot, A. Plaza, and G. Scarpa. Hyperspectral pansharpening: Critical review, tools and future perspectives. arXiv preprint arXiv:2407.01355, 2024. 2", + "[8] L.-J. Deng, G. Vivone, M. E. Paoletti, G. Scarpa, J. He, Y. Zhang, J. Chanussot, and A. Plaza. Machine learning in pan-sharpening: A benchmark, from shallow to deep networks. IEEE Geosci. Remote Sens. Mag., 10(3):279-315, 2022. 2", + "[9] P. Dhariwal and A. Nichol. Diffusion models beat gans on image synthesis. Adv. Neural Inf. Process. Syst. (NeurIPS), 34:8780-8794, 2021. 3", + "[10] R. Dian, L. Fang, and S. Li. Hyperspectral image superresolution via non-local sparse tensor factorization. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 5344-5353, 2017. 1", + "[11] R. Dian, A. Guo, and S. Li. Zero-shot hyperspectral sharpening. IEEE Trans. Pattern Anal. Mach. Intell., 45(10):12650-12666, 2023. 4, 5, 6, 7, 8", + "[12] W. Dong, T. Zhang, J. Qu, S. Xiao, J. Liang, and Y. Li. Laplacian pyramid dense network for hyperspectral pansharpening. IEEE Trans. Geosci. Remote Sens., 60:1-13, 2021. 1", + "[13] Y. Duan, X. Wu, H. Deng, and L.-J. Deng. Content-adaptive non-local convolution for remote sensing pansharpening. In Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recog., pages 27738-27747, 2024. 1", + "[14] X. Fu, Z. Lin, Y. Huang, and X. Ding. A variational pan-sharpening with local gradient constraints. In Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recog., pages 10265-10274, 2019. 1", + "[15] J. Gao, J. Li, X. Su, M. Jiang, and Q. Yuan. Deep image interpolation: A unified unsupervised framework for pansharpening. In Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recog., pages 609-618, 2022. 1" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] M. Gong, H. Zhang, H. Wang, J. Chen, J. Huang, X. Tian, and J. Ma. Cross-scale domain adaptation with comprehensive information for pansharpening. In Int. Joint Conf. Artif. Intell, pages 785-793, 2024. 1", + "[17] L. He, H. Ye, D. Xi, J. Li, A. Plaza, and M. Zhang. Tree-structured neural network for hyperspectral pansharpening. IEEE J. Sel. Top. Appl. Earth Observ. Remote Sens., 17: 2516-2530, 2023. 1", + "[18] L. He, Z. Fang, J. Li, J. Chanussot, and A. Plaza. Two spectral-spatial implicit neural representations for arbitrary-resolution hyperspectral pansharpening. IEEE Trans. Geosci. Remote Sens., 2024. 1, 2", + "[19] W. He, Q. Yao, C. Li, N. Yokoya, and Q. Zhao. Non-local meets global: An integrated paradigm for hyperspectral denoising. In Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recog., pages 6868-6877, 2019. 5", + "[20] W. He, Q. Yao, C. Li, N. Yokoya, Q. Zhao, H. Zhang, and L. Zhang. Non-local meets global: An iterative paradigm for hyperspectral image restoration. IEEE Trans. Pattern Anal. Mach. Intell., 44(4):2089-2107, 2020. 5", + "[21] Jo. Ho, A. Jain, and P. Abbeel. Denoising diffusion probabilistic models. In Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), pages 6840-6851, 2020. 2, 3", + "[22] S. Huang, M. Ren, Y. Yang, X. Wang, and Y. Wei. MFTN: A multi-scale feature transfer network based on IMatchformer for hyperspectral image super-resolution. In *Forty-first Int. Conf. Mach. Learn.*, 2024. 1", + "[23] B. Kawar, M. Elad, S. Ermon, and J. Song. Denoising diffusion restoration models. Proc. Adv. Neural Inf. Process. Syst., 35:23593-23606, 2022. 3", + "[24] Tamara G Kolda and Brett W Bader. Tensor decompositions and applications. SIAM Rev., 51(3):455-500, 2009. 2", + "[25] S. Li, S. Li, and L. Zhang. Hyperspectral and panchromatic images fusion based on the dual conditional diffusion models. IEEE Trans. Geosci. Remote Sens., 2023. 3", + "[26] L. Loncan, L. B. De Almeida, J. M. Bioucas-Dias, X. Briottet, J. Chanussot, N. Dobigeon, S. Fabre, W. Liao, G. A. Licciardi, M. Simoes, et al. Hyperspectral pansharpening: A review. IEEE Geosci. Remote Sens. Mag., 3(3):27-46, 2015. 1", + "[27] Y. Miao, L. Zhang, L. Zhang, and D. Tao. DDS2M: Self-supervised denoising diffusion spatio-spectral model for hyperspectral image restoration. In Proc. IEEE/CVF Int. Conf. Comput. Vis., pages 12086-12096, 2023. 3", + "[28] A. Quinn Nichol, P. Dhariwal, A. Ramesh, P. Shyam, P. Mishkin, B. Mcgrew, I. Sutskever, and M. Chen. GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. In Int. Conf. Mach. Learn., pages 16784-16804. PMLR, 2022. 4", + "[29] L. Pang, X. Rui, L. Cui, H. Wang, D. Meng, and X. Cao. HIR-Diff: Unsupervised hyperspectral image restoration via improved diffusion models. In Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recog., pages 3005-3014, 2024. 1, 2, 3, 5, 6, 7, 8", + "[30] X. Rui, X. Cao, L. Pang, Z. Zhu, Z. Yue, and D. Meng. Unsupervised hyperspectral pansharpening via low-rank diffusion model. Inf. Fusion, 107:102325, 2024. 1, 2, 3, 5, 6, 7, 8" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12677", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] K. Shi, J. Peng, J. Gao, Y. Luo, and S. Xu. Hyperspectral image denoising via double subspace deep prior. IEEE Trans. Geosci. Remote Sens., 62:1-15, 2024. 5", + "[32] M. Simoes, J. Bioucas-Dias, L. B. Almeida, and J. Chanussot. A convex formulation for hyperspectral image superresolution via subspace-based regularization. IEEE Trans. Geosci. Remote Sens., 53(6):3373-3388, 2014. 5, 6, 7", + "[33] Y. Song and S. Ermon. Generative modeling by estimating gradients of the data distribution. In Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), pages 11895-11907, 2019. 3", + "[34] Y. Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole. Score-based generative modeling through stochastic differential equations. In Int. Conf. Learn. Representations, 2021. 2, 3", + "[35] D. Ulyanov, A. Vedaldi, and V. Lempitsky. Deep image prior. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 9446-9454, 2018. 5, 8", + "[36] G. Vivone, M. Dalla Mura, A. Garzelli, R. Restaino, G. Scarpa, M. O. Ulfarsson, L. Alparone, and J. Chanussot. A new benchmark based on recent advances in multispectral pansharpening: Revisiting pansharpening with classical and emerging pansharpening methods. IEEE Geosci. Remote Sens. Mag., 9(1):53-81, 2020. 1", + "[37] W. Wang, L.-J. Deng, R. Ran, and G. Vivone. A general paradigm with detail-preserving conditional invertible network for image fusion. Int. J. Comput. Vis., 132(4):1029-1054, 2024. 2", + "[38] Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Trans. Image Process., 13(4):600-612, 2004. 6", + "[39] C. Wu, D. Wang, Y. Bai, H. Mao, Y. Li, and Q. Shen. HSR-Diff: Hyperspectral image super-resolution via conditional diffusion models. In Proc. IEEE/CVF Int. Conf. Comput. Vis., pages 7083-7093, 2023. 3", + "[40] J.-L. Xiao, T.-Z. Huang, L.-J. Deng, Z.-C. Wu, X. Wu, and G. Vivone. Variational pansharpening based on coefficient estimation with nonlocal regression. IEEE Trans. Geosci. Remote. Sens., 61:1-15, 2023. 2", + "[41] W. Xie, J. Lei, Y. Cui, Y. Li, and Q. Du. Hyperspectral pansharpening with deep priors. IEEE Trans. Neural Netw. Learn. Syst., 31(5):1529-1543, 2019. 1", + "[42] Y. Xing, L. Qu, S. Zhang, K. Zhang, Y. Zhang, and L. Bruzzone. Crossdiff: Exploring self-supervised representation of pansharpening via cross-predictive diffusion model. IEEE Trans. Image Process., 33:5496-5509, 2024. 3", + "[43] T. Xu, T.-Z. Huang, L.-J. Deng, and N. Yokoya. An iterative regularization method based on tensor subspace representation for hyperspectral image super-resolution. IEEE Trans. Geosci. Remote Sens., 60:1-16, 2022. 2", + "[44] T. Xu, T.-Z. Huang, L.-J. Deng, J.-L. Xiao, C. Broni-Bediako, J. Xia, and N. Yokoya. A coupled tensor double-factor method for hyperspectral and multispectral image fusion. IEEE Trans. Geosci. Remote Sens., 62:1-17, 2024. 2, 4", + "[45] J. Xue, Y.-Q. Zhao, Y. Bu, W. Liao, J. C.-W. Chan, and W. Philips. Spatial-spectral structured sparse low-rank represent" + ], + "bbox": [ + 91, + 90, + 480, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tation for hyperspectral image super-resolution. IEEE Trans. Image Process., 30:3084-3097, 2021. 2", + "[46] J. Yao, D. Hong, J. Chanussot, D. Meng, X. Zhu, and Z. Xu. Cross-attention in coupled unmixing nets for unsupervised hyperspectral super-resolution. In Proc. Eur. Conf. Comput. Vis., pages 208–224. Springer, 2020. 1", + "[47] N. Yokoya, T. Yairi, and A. Iwasaki. Coupled nonnegative matrix factorization unmixing for hyperspectral and multispectral data fusion. IEEE Trans. Geosci. Remote Sens., 50 (2):528-537, 2011. 1, 5, 6, 7", + "[48] Y. Yu, E. Pan, X. Wang, Y. Wu, X. Mei, and J. Ma. Unmixing before fusion: A generalized paradigm for multi-source-based hyperspectral image synthesis. In Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recog., pages 9297-9306, 2024.", + "[49] R. H. Yuhas, A. F. H. Goetz, and J. W. Boardman. Discrimination among semi-arid landscape endmembers using the spectral angle mapper (SAM) algorithm. In Proc. Summaries 3rd Annu. JPL Airborne Geosci. Workshop, pages 147-149, 1992. 6", + "[50] Y. Zhong, X. Wu, L.-J. Deng, and Z. Cao. SSDiff: Spatial-spectral integrated diffusion model for remote sensing pan-sharpening. arXiv preprint arXiv:2404.11537, 2024. 3", + "[51] Y.-W. Zhuo, T.-J. Zhang, J.-F. Hu, H.-X. Dou, T.-Z. Huang, and L.-J. Deng. A deep-shallow fusion network with multi-detail extractor and spectral attention for hyperspectral pan-sharpening. IEEE J. Sel. Top. Appl. Earth Observ. Remote Sens., 15:7539-7555, 2022. 8" + ], + "bbox": [ + 516, + 92, + 903, + 484 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12678", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Hyperspectral Pansharpening via Diffusion Models with Iteratively Zero-Shot Guidance/f79e03e7-ea12-4e70-879c-79258b8c5e82_model.json b/2025/Hyperspectral Pansharpening via Diffusion Models with Iteratively Zero-Shot Guidance/f79e03e7-ea12-4e70-879c-79258b8c5e82_model.json new file mode 100644 index 0000000000000000000000000000000000000000..060b85eb3e6a7a84c61415607d4ad81809b3a493 --- /dev/null +++ b/2025/Hyperspectral Pansharpening via Diffusion Models with Iteratively Zero-Shot Guidance/f79e03e7-ea12-4e70-879c-79258b8c5e82_model.json @@ -0,0 +1,2365 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.02 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + }, + { + "type": "header", + "bbox": [ + 0.325, + 0.017, + 0.725, + 0.031 + ], + "angle": 0, + "content": "Except for this watermark, it is identical to the accepted version;" + }, + { + "type": "header", + "bbox": [ + 0.294, + 0.032, + 0.756, + 0.047 + ], + "angle": 0, + "content": "the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.107, + 0.13, + 0.892, + 0.175 + ], + "angle": 0, + "content": "Hyperspectral Pansharpening via Diffusion Models with Iteratively Zero-Shot Guidance" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.203, + 0.948, + 0.222 + ], + "angle": 0, + "content": "Jin-Liang Xiao\\(^{1,3}\\), Ting-Zhu Huang\\(^{1*}\\), Liang-Jian Deng\\(^{1}\\), Guang Lin\\(^{2,3}\\), Zihan Cao\\(^{1}\\), Chao Li\\(^{3}\\), Qibin Zhao\\(^{3*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.222, + 0.827, + 0.239 + ], + "angle": 0, + "content": "1 University of Electronic Science and Technology of China, Chengdu, China" + }, + { + "type": "text", + "bbox": [ + 0.265, + 0.239, + 0.777, + 0.257 + ], + "angle": 0, + "content": "\\(^{2}\\) Tokyo University of Agriculture and Technology, Tokyo, Japan" + }, + { + "type": "text", + "bbox": [ + 0.408, + 0.257, + 0.634, + 0.275 + ], + "angle": 0, + "content": "\\(^{3}\\) RIKEN AIP, Tokyo, Japan" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.309, + 0.327, + 0.326 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.342, + 0.486, + 0.735 + ], + "angle": 0, + "content": "Hyperspectral pansharpening refers to fusing a panchromatic image (PAN) and a low-resolution hyperspectral image (LR-HSI) to obtain a high-resolution hyperspectral image (HR-HSI). Recently, guiding pre-trained diffusion models (DMs) has demonstrated significant potential in this area, leveraging their powerful representational abilities while avoiding complex training processes. However, these DMs are often trained on RGB images, not well-suited for pansharpening tasks, limited in adapting to the hyperspectral images. In this work, we propose a novel guided diffusion scheme with zero-shot guidance and neural spatial-spectral decomposition (NSSD) to iteratively generate the RGB detail image and map the RGB detail image to target HR-HSI. Specifically, zero-shot guidance employs an auxiliary neural network that trained only with a PAN and LR-HSI to guide pre-trained DMs in generating the RGB detail image, informed by specific prior knowledge. Then, NSSD establishes a spectral mapping from the generated RGB detail image to the final HR-HSI. Extensive experiments are conducted on Pavia, Washington DC, Chukusei, and FR1 datasets to demonstrate that the proposed method significantly enhances the performance of DMs for hyperspectral pansharpening tasks, outperforming existing methods across multiple metrics and achieving improvements in visualization results. The code is available at https://github.com/Jin-liangXiao/DM-zs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.763, + 0.222, + 0.779 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.787, + 0.485, + 0.88 + ], + "angle": 0, + "content": "Hyperspectral images (HSIs), rich in spatial and spectral information, have been extensively used in a variety of applications [22, 46, 48]. In practice, however, the spatial and spectral information in HSIs are often divided across different image types: the panchromatic image (PANs) with high spatial resolution and low-resolution hyperspectral images" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.308, + 0.907, + 0.586 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.595, + 0.909, + 0.693 + ], + "angle": 0, + "content": "Figure 1. The upper row: illustration of the proposed method that includes iterative update of zero-shot guidance for pre-trained diffusion models and spectral mapping from RGB detail image to target HR-HSI. The bottom row: error map comparison between state-of-the-art pre-trained diffusion based hyperspectral pansharpening models (i.e., PLRDiff [30], HIR-Diff [29], and the proposed method)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.719, + 0.907, + 0.795 + ], + "angle": 0, + "content": "(LR-HSIs) with abundant spectral information, due to hardware constraints. This separation highlights the need for hyperspectral pansharpening, which aims to fuse these image types to generate high-resolution hyperspectral images (HR-HSIs) [26]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Methods for hyperspectral pansharpening generally fall into two categories: single-image approaches [10, 14, 16, 41, 47] and deep learning (DL) methods [4, 12, 13, 13, 15, 17, 18]. Single-image approaches require only a single image pair instead of an entire dataset for training, which can save significant training costs [36]. While single-image approaches are effective, DL methods can provide stronger" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.236, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12669" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.243 + ], + "angle": 0, + "content": "representations due to the powerful feature-extraction ability of deep neural networks. However, DL methods come with the trade-off of requiring intensive computational resources for training [7, 18, 37] and consistency in the dataset [40]. Specifically, if there is a distribution shift [8] between the training data and testing data, the performance can degrade significantly, which leads to poor generalization. This issue is particularly serious in hyperspectral pan-sharpening, where hyperspectral data are both more varying and limited than data in general computer vision tasks." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.248, + 0.485, + 0.67 + ], + "angle": 0, + "content": "Foundation models, particularly diffusion models (DMs) [21, 34], have gained wide attention in computer vision for their capability of generating high-quality images. In the context of hyperspectral pansharpening, DMs have shown improved generalization due to flexibility of the guidance techniques. Recent studies [29, 30] have demonstrated the potential of DMs, extending pre-trained DMs to hyperspectral pansharpening through specialized guidance. These methods utilize the degenerate process of pansharpening to guide the DM. Due to the powerful representational capabilities, pre-trained DM as a foundation model can be quickly transferred to hyperspectral pansharpening tasks, which can reduce a significant amount of training costs. However, since existing methods lack the consideration of specific prior knowledge of hyperspectral data, the guidance of DMs still needs improvement. The extension from pre-trained DMs to the hyperspectral pansharpening task is non-trivial: We need effective guidance to adapt the pre-trained DMs originally designed with only RGB channels, to manage the complex spectral fidelity while preserving high-resolution spatial details. This adaptation becomes even more challenging when we only have a single (PAN, LR-HSI) image pair. These challenges naturally give rise to two important questions: 1) How can we shift the data distribution of pre-trained DMs to HSIs for hyperspectral pansharpening with limited data through effective guidance? 2) How to accurately transform the RGB information of the generated images to HR-HSIs?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To address the aforementioned challenges, we propose a novel guided diffusion scheme with zero-shot guidance and neural spatial-spectral decomposition (NSSD) to iteratively generate an RGB spatial detail image and map the detail image to the target HR-HSI. As shown in the green box of Figure 1, we utilize the pre-trained DMs as the backbone, with all their parameters being frozen. Then, we introduce the specific prior knowledge of testing data by a trainable zero-shot network to guide the DMs to generate the detail image enriched with expected features, as shown in the blue box. Unlike conventional deep neural networks that require extensive training data and complex training processes, the proposed zero-shot network is a lightweight model with few parameters, can be trained naively on a single testing image pair, i.e., a LR-HSI and PAN. This network encodes" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.288 + ], + "angle": 0, + "content": "specific prior knowledge into pre-trained DMs and guides the DMs to generate detail image more effectively (see Figure 3). Additionally, motivated by the tensor decomposition in single-image approaches [44], we recognize that spatial-spectral correlation is critical for the quality of final data and introduce a neural spatial-spectral decomposition (NSSD) to capture the spatial-spectral correlation and map from the RGB detail image to the final HR-HSI. The NSSD and the guidance of pre-trained DMs are updated iteratively, as displayed in the gray box of Figure 1, making NSSD adaptive to the results of guided DMs. Compared with existing pretrained DMs methods, the proposed method achieves the state-of-the-art result (refer to the bottom row of Figure 1)." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.289, + 0.872, + 0.303 + ], + "angle": 0, + "content": "The main contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.305, + 0.904, + 0.35 + ], + "angle": 0, + "content": "- We propose a zero-shot network to efficiently guide the pre-trained DM to generate a detail image enriched with specific prior knowledge." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.35, + 0.905, + 0.395 + ], + "angle": 0, + "content": "- We introduce a neural spatial-spectral decomposition (NSSD) to iteratively construct the target HR-HSI from the generated detail image." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.396, + 0.905, + 0.44 + ], + "angle": 0, + "content": "- Experiments on the Pavia, Washington DC, Chikusei, and FR1 datasets demonstrate that the proposed method achieves the state-of-the-art performance." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.305, + 0.905, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.457, + 0.802, + 0.472 + ], + "angle": 0, + "content": "2. Preliminaries & Related Works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.483, + 0.905, + 0.528 + ], + "angle": 0, + "content": "In this section, we provide definitions, notations, degenerate process of pansharpening, a brief introduction to score-based DMs with guidance, and related works." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.54, + 0.625, + 0.555 + ], + "angle": 0, + "content": "2.1. Notations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.562, + 0.905, + 0.669 + ], + "angle": 0, + "content": "Throughout the paper, we denote tensors \\(\\mathcal{X} \\in \\mathbb{R}^{H \\times W \\times S}\\), \\(\\mathcal{Y} \\in \\mathbb{R}^{h \\times w \\times S}\\), \\(\\mathbf{P} \\in \\mathbb{R}^{H \\times W}\\) as HR-HSI, LR-HSI, PAN, respectively. For tensor \\(\\mathcal{X}\\), \\(\\mathbf{X}_{(3)} \\in \\mathbb{R}^{S \\times HW}\\) denotes its mode-3 unfolding [24]. Additionally, \\(\\mathcal{X} \\times_3 \\mathbf{R}\\) indicates the mode-3 multiplication between tensor \\(\\mathcal{X}\\) and matrix \\(\\mathbf{R}\\). \\(\\mathbf{x} \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})\\) means the variable \\(\\mathbf{x}\\) satisfies Gaussian distribution with the mean of \\(\\mathbf{0}\\) and variance of matrix \\(\\mathbf{I}\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.681, + 0.838, + 0.697 + ], + "angle": 0, + "content": "2.2. Degenerate Process of Pansharpening" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.905, + 0.78 + ], + "angle": 0, + "content": "For the hyperspectral pansharpening task, inputs PAN and LR-HSI are mainly modeled as HR-HSI after the spectral and spatial degradation, respectively [45]. The degenerate process of the hyperspectral pansharpening problem can be mathematically represented as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.663, + 0.793, + 0.905, + 0.816 + ], + "angle": 0, + "content": "\\[\n\\mathcal {Y} = \\tau (\\mathcal {X}), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.662, + 0.813, + 0.759, + 0.827 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} = \\mathcal {X} \\times_ {3} \\mathbf {R},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.905, + 0.901 + ], + "angle": 0, + "content": "where \\(\\tau\\) indicates blurring and down-sampling operations, and \\(\\mathbf{R} \\in \\mathbb{R}^{1 \\times S}\\) denotes the spectral response function [43]. The goal of the task is to obtain the HR-HSI \\(\\mathcal{X}\\) from the LR-HSI \\(\\mathcal{Y}\\) and PAN \\(\\mathbf{P}\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12670" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.473, + 0.106 + ], + "angle": 0, + "content": "2.3. Score-based Diffusion Models with Guidance" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.484, + 0.218 + ], + "angle": 0, + "content": "Diffusion models (DMs) [21, 33] have proven to be a potent class of generative models capable of generating high-quality images. Among them, score-based DMs are a powerful family that models the forward process by stochastic differential equations (SDEs) and the reverse process by corresponding reverse-time SDEs [34]. Given a input \\(\\mathbf{x}\\), we have" + }, + { + "type": "equation", + "bbox": [ + 0.196, + 0.231, + 0.483, + 0.248 + ], + "angle": 0, + "content": "\\[\n\\mathrm {d} \\mathbf {x} = f (\\mathbf {x}, t) \\mathrm {d} t + g (t) \\mathrm {d} \\mathbf {w}, \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.27, + 0.483, + 0.289 + ], + "angle": 0, + "content": "\\[\n\\mathrm {d} \\mathbf {x} = [ f (\\mathbf {x}, t) - g ^ {2} (t) \\nabla_ {\\mathbf {x}} \\log p _ {t} (\\mathbf {x}) ] + g (t) \\mathrm {d} \\bar {\\mathbf {w}}, \\quad (3)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.298, + 0.484, + 0.555 + ], + "angle": 0, + "content": "where \\(t\\) represents timestep, \\(f(\\cdot ,t)\\) denotes a vector valued function called drift coefficient and \\(g(t)\\) is a scalar valued function called diffusion term. \\(\\mathbf{w}\\) and \\(\\bar{\\mathbf{w}}\\) represent a Wiener process and a reverse-time Wiener process, respectively. Eq. (2), corresponding to the forward SDE, indicates adding noise from \\(t = 0\\) to \\(T\\), which diffuses the image to pure Gaussian noise. In the reverse process, the corresponding reverse-time SDEs, i.e., Eq. (3), we utilize a learned score network \\(s_{\\theta}(\\mathbf{x}_t,t)\\coloneqq \\nabla_{\\mathbf{x}}\\log p_t(\\mathbf{x})\\) to generate denoised sample \\(\\mathbf{x}_0\\) from a simple Gaussian noise \\(\\mathbf{x}_T\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\). To more effectively enable DMs to generate the desired images, Dhariwal et al. and Pang et al. [9, 29] proposed the guided DM, which is a valid path for general specific tasks, leveraging the powerful generative capabilities of DMs without the need for complex retraining. The guidance of pre-trained DMs works by introducing gradient in reverse process, which can be represented as" + }, + { + "type": "equation", + "bbox": [ + 0.134, + 0.569, + 0.483, + 0.587 + ], + "angle": 0, + "content": "\\[\n\\hat {s} _ {\\theta} (\\mathbf {x} _ {t}, t) = s _ {\\theta} (\\mathbf {x} _ {t}, t) + s (t) \\cdot \\nabla_ {\\mathbf {x} _ {t}} \\log p (y | \\mathbf {x} _ {t}), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.598, + 0.484, + 0.644 + ], + "angle": 0, + "content": "where \\( s(t) \\) is the guidance strength in \\( t \\) step, and \\( p(y|\\mathbf{x}_t) \\) is the relationship between estimated \\( \\mathbf{x}_t \\) and condition \\( y \\), which is crucial for the guidance of DMs on specific tasks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.653, + 0.242, + 0.667 + ], + "angle": 0, + "content": "2.4. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.87 + ], + "angle": 0, + "content": "Leveraging the powerful generative capabilities of DMs, numerous studies have been conducted on hyperspectral pansharpening [5, 25, 39, 42, 50]. However, the performance of these models often depends on complex and time-consuming training processes and substantial training data. Recently, efforts have also been made to bridge the connection between inverse problems and pre-trained DMs. Kawar et al. [23] employed a pre-trained DM to solve linear inverse problems for RGB images. Since pre-trained DMs are primarily trained on RGB data, there remains a gap between the generated outputs and the target hyperspectral images. Miao et al. [27] further extended this approach to hyperspectral image restoration." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "For the hyperspectral pansharpening task, a nonlinear inverse problem, different loss functions are proposed to" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.379 + ], + "angle": 0, + "content": "guide DMs. Rui et al. [30] proposed a pre-trained diffusion framework that guides the diffusion process to generate an RGB detail image \\(\\mathcal{A}\\). This image is then transformed into the final pansharpened hyperspectral image by applying a fixed transformation matrix along the third dimension, as \\(\\mathcal{X} = \\mathcal{A} \\times_3 \\mathbf{E}\\), where \\(\\mathcal{X} \\in \\mathbb{R}^{H \\times W \\times S}\\) is the target HR-HSI image, \\(\\mathcal{A} \\in \\mathbb{R}^{H \\times W \\times b}\\) is the image generated by a pre-trained DM, and \\(\\mathbf{E} \\in \\mathbb{R}^{S \\times b}\\) is the fixed transform matrix. Pang et al. [29] explore a new choice for estimating the transform matrix \\(\\mathbf{E}\\) and apply total variation (TV) prior to diffusion guided function. However, this issue remains only partially resolved, as their fixed transformation matrix \\(\\mathbf{E}\\) in the overall algorithm causes the final result to depend heavily on the initial estimate of \\(\\mathbf{E}\\). In contrast, our method introduces an auxiliary neural network to provide the specialized guidance for generating an RGB detail image and iteratively employs a neural spatial-spectral decomposition to adaptively construct the mapping from the RGB detail image to HR-HSI." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.392, + 0.649, + 0.409 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.417, + 0.907, + 0.553 + ], + "angle": 0, + "content": "In this section, we elaborate on the proposed guided diffusion shcheme. As illustrated in Figure 2, the scheme primarily consists of two components: diffusion with zero-shot guidance and neural spatial-spectral decomposition (NSSD). The zero-shot guidance operates on the pre-trained DM to generate an RGB detail image, enabling better preservation of spatial details and adaptation to specific prior knowledge. NSSD establishes the mapping from the detail image to the HR-HSI." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.562, + 0.821, + 0.577 + ], + "angle": 0, + "content": "3.1. Diffusion with Zero-Shot Guidance" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.906, + 0.675 + ], + "angle": 0, + "content": "Unlike the existing DM-based hyperspectral pansharpening methods [29, 30], we aim to guide the reverse process of DMs for generating the RGB detail image \\(\\mathcal{A}\\) by introducing the specific prior knowledge of testing data. To this end, we design a novel diffusion scheme with zero-shot guidance, where the reverse process of \\(\\mathcal{A}\\) satisfies SDE as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.686, + 0.905, + 0.724 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathrm {d} \\mathcal {A} _ {t} = g (t) \\mathrm {d} \\bar {w} + f (\\mathcal {A}, t) \\mathrm {d} t \\\\ - g ^ {2} (t) \\left[ s _ {\\theta} \\left(x _ {t}, t\\right) + \\nabla_ {\\mathcal {A} _ {t}} \\log p _ {t} (\\tilde {\\mathcal {Y}} | \\mathcal {A} _ {t}) \\right] \\mathrm {d} t, \\tag {5} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.736, + 0.907, + 0.767 + ], + "angle": 0, + "content": "where \\( p_t(\\tilde{\\mathcal{Y}}|\\mathcal{A}_t) \\) is the pansharpening guidance estimated from the condition, defined as" + }, + { + "type": "equation", + "bbox": [ + 0.616, + 0.78, + 0.907, + 0.798 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathcal {Y}} = (\\mathcal {Y}, \\mathbf {P}, \\mathcal {M}, f _ {\\psi} (\\mathcal {Y}, \\mathbf {P})), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.906, + 0.87 + ], + "angle": 0, + "content": "where \\(\\mathcal{V}\\) and \\(\\mathbf{P}\\) are input image pair LR-HSI and PAN, \\(\\mathcal{M}\\) is a mapping from RGB detail image to HR-HSI (deferred in next subsection), and \\(f_{\\psi}(\\mathcal{V},\\mathbf{P})\\) is an auxiliary neural network that is trained with \\(\\mathcal{V}\\) and \\(\\mathbf{P}\\). Following [9], we have" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.883, + 0.907, + 0.902 + ], + "angle": 0, + "content": "\\[\np _ {\\theta , \\psi} \\left(\\mathcal {A} _ {t - 1} \\mid \\mathcal {A} _ {t}, \\tilde {\\mathcal {Y}}\\right) \\propto p _ {\\theta} \\left(\\mathcal {A} _ {t - 1} \\mid \\mathcal {A} _ {t}\\right) p _ {\\psi} (\\tilde {\\mathcal {Y}} \\mid \\mathcal {A} _ {t}). \\tag {7}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12671" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.175, + 0.09, + 0.825, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.308, + 0.908, + 0.366 + ], + "angle": 0, + "content": "Figure 2. Illustration of the whole process of the proposed hyperspectral pansharpening method that includes diffusion with zero-shot (ZS) guidance and neural spatial-spectral decomposition (NSSD). The part in the dashed box is the proposed zero-shot guided function that is trained with a single testing image pair. NSSD builds the spectral mapping from the detail image generated by DM to HR-HSI. The guidance of the DM and NSSD are iteratively updated." + }, + { + "type": "image", + "bbox": [ + 0.133, + 0.393, + 0.445, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.539, + 0.483, + 0.61 + ], + "angle": 0, + "content": "Figure 3. The analysis of the proposed zero-shot guidance. The DMs with the proposed guidance can achieve better performance with fewer sampling steps. (The length of the yellow box means the distance of PSNR values between the DM with and without ZS guidance)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.64, + 0.483, + 0.675 + ], + "angle": 0, + "content": "Based on [6], we define the variable \\(\\hat{\\mathcal{A}}_0\\coloneqq \\mathbb{E}[\\mathcal{A}_0|\\mathcal{A}_t] =\\) \\(\\mathbb{E}_{\\mathcal{A}_0\\sim p(\\mathcal{A}_0|\\mathcal{A}_t)}[\\mathcal{A}_0]\\) , which means" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.689, + 0.483, + 0.723 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {A}} _ {0} \\simeq \\frac {1}{\\sqrt {\\bar {\\alpha} _ {t}}} \\left(\\mathcal {A} _ {t} + \\left(1 - \\bar {\\alpha} _ {t}\\right) s _ {\\theta} \\left(\\mathcal {A} _ {t}, t\\right)\\right), \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.204, + 0.746, + 0.483, + 0.767 + ], + "angle": 0, + "content": "\\[\np _ {\\psi} (\\tilde {\\mathcal {V}} | \\mathcal {A} _ {t}) \\simeq p _ {\\psi} (\\tilde {\\mathcal {V}} | \\hat {\\mathcal {A}} _ {0}), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.813 + ], + "angle": 0, + "content": "where \\(\\bar{\\alpha}_t\\coloneqq \\prod_{s = 1}^t\\alpha_s\\) is the noise scale of DMs. We propose the zero-shot guidance as follows," + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.829, + 0.49, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\log p _ {\\psi} (\\tilde {\\mathcal {Y}} | \\hat {\\mathcal {A}} _ {0}) = \\left\\| \\tau (\\mathcal {M} (\\hat {\\mathcal {A}} _ {0})) - \\mathcal {Y} \\right\\| _ {F} ^ {2} \\\\ + \\left\\| \\mathcal {M} \\left(\\hat {\\mathcal {A}} _ {0}\\right) \\times_ {3} \\mathbf {R} - \\mathbf {P} \\right\\| _ {F} ^ {2} + \\lambda_ {1} \\left\\| \\mathcal {M} \\left(\\hat {\\mathcal {A}} _ {0}\\right) - f _ {\\psi} (\\mathcal {Y}, \\mathbf {P}) \\right\\| _ {F} ^ {2}, \\tag {10} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.391, + 0.905, + 0.42 + ], + "angle": 0, + "content": "where \\(\\lambda_{1}\\) is a weighting scale. Finally, the reverse process is illustrated as:" + }, + { + "type": "equation", + "bbox": [ + 0.523, + 0.428, + 0.905, + 0.497 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {A} _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\mathcal {A} _ {t} - \\frac {\\beta_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} s _ {\\theta} \\left(\\mathcal {A} _ {t}, t\\right)\\right) \\tag {11} \\\\ - \\lambda \\cdot \\frac {\\beta_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}} \\sqrt {\\alpha_ {t}}} \\nabla_ {\\mathcal {A} _ {t}} l o g p _ {\\psi} (\\tilde {\\mathcal {Y}} | \\hat {\\mathcal {A}} _ {0}), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.905, + 0.553 + ], + "angle": 0, + "content": "where \\(\\beta_{t}\\), as a hyperparameter, is generally set to linearly interpolate from 0.0001 to 0.02 [28] and \\(\\alpha_{t} := 1 - \\beta_{t}\\). The weighting scale \\(\\lambda\\) can balance the effect of guidance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.907, + 0.795 + ], + "angle": 0, + "content": "The rest of this part is how to train the zero-shot network \\( f_{\\psi}(\\mathcal{V}, \\mathbf{P}) \\). The network structure and training process are shown within the dotted box in Figure 2, which has been verified its effectiveness in [11]. Based on the \"invariance among scales\" assumption, we train the network in the reduced-resolution scale. In specific, we use the input LR-HSI as the training ground-truth. Training PAN and training LR-HSI are simulated by PAN and LR-HSI after spatial and spectral degenerate processes, respectively. This approach allows for efficient training of the zero-shot network using only a single testing input pair. Figure 3 shows the change curve of PSNR values with the increase of iteration number. One can observe that the DM guided by zero-shot guidance can achieve better performance with fewer sampling steps. Besides, this single-image training property provides specific prior knowledge to the pre-trained DM." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.803, + 0.85, + 0.82 + ], + "angle": 0, + "content": "3.2. Neural Spatial-Spectral Decomposition" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.902 + ], + "angle": 0, + "content": "This subsection describes neural spatial-spectral decomposition (NSSD), i.e., \\(\\mathcal{M}\\) in Eq. (10). In hyperspectral pan-sharpening, beyond the fidelity term in (1), the spatial-spectral correlation exhibits the low-rank properties characteristic of hyperspectral data [44]. The mode-3 decom" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12672" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.216, + 0.089, + 0.784, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.263, + 0.908, + 0.294 + ], + "angle": 0, + "content": "Figure 4. The structure of neural spatial-spectral decomposition (NSSD) consists of a U-net structure with skip connection and the fully connection structure to extract spatial and spectral information, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.318, + 0.482, + 0.351 + ], + "angle": 0, + "content": "position format, i.e., \\(\\mathcal{X} = \\mathcal{A}\\times_3\\mathbf{E}\\), has shown to be effective [19, 20, 31], which satisfies the following inequation:" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.36, + 0.483, + 0.379 + ], + "angle": 0, + "content": "\\[\n\\operatorname {r a n k} \\left(\\mathbf {X} _ {(3)}\\right) \\leq \\min \\left\\{\\operatorname {r a n k} \\left(\\mathbf {A} _ {(3)}\\right), \\operatorname {r a n k} (\\mathbf {E}) \\right\\} \\leq b, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.387, + 0.484, + 0.478 + ], + "angle": 0, + "content": "where \\(\\mathcal{A} \\in \\mathbb{R}^{H \\times W \\times b}\\), and \\(\\mathbf{E} \\in \\mathbb{R}^{S \\times b}\\). The low value of the size of \\(\\mathbf{E}\\) limits the expressive ability. Since pre-trained DMs are primarily trained on RGB images, the size of \\(\\mathcal{A}\\) is typically set to \\(H \\times W \\times 3\\). Additionally, the larger value of \\(r\\) in (12) inevitably increases the time required for the reverse process in the DM." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.48, + 0.484, + 0.51 + ], + "angle": 0, + "content": "To mitigate the drawbacks, we propose a NSSD to capture the spatial-spectral correlation, as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.522, + 0.483, + 0.541 + ], + "angle": 0, + "content": "\\[\n\\mathcal {X} = \\mathcal {M} (\\mathcal {A}) := \\mathcal {C} _ {\\theta_ {1}} (\\mathcal {A}) \\times_ {3} \\mathbf {D} _ {\\theta_ {2}} (\\mathcal {Y}), \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.551, + 0.484, + 0.734 + ], + "angle": 0, + "content": "where \\(\\mathcal{C}_{\\theta_1}(\\mathcal{A}) \\in \\mathbb{R}^{H \\times W \\times r}\\) and \\(\\mathbf{D}_{\\theta_2}(\\mathcal{Y}) \\in \\mathbb{R}^{S \\times r}\\) with hyperparameter \\(r\\) are produced by two neural network with the input \\(\\mathcal{A} \\in \\mathbb{R}^{H \\times W \\times 3}\\) and \\(\\mathcal{Y} \\in \\mathbb{R}^{h \\times w \\times S}\\), respectively, as illustrated in Figure 4. Specifically, \\(\\mathcal{C}_{\\theta_1}(\\cdot)\\) is the spatial-factor network that extracts the spatial information from the RGB detail image. The structure of \\(\\mathcal{C}_{\\theta_1}(\\cdot)\\) is the U-net type with skip connection, whose detail-extraction capability has been verified [35]. The spectral-factor network, \\(\\mathbf{D}_{\\theta_2}(\\cdot)\\), consists of the fully connection structure and sigmoid to efficiently absorb spectral information from LR-HSI. NSSD does not require pre-training. It produces the final result through iterative steps using the following loss:" + }, + { + "type": "equation", + "bbox": [ + 0.106, + 0.744, + 0.483, + 0.786 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L o s s = \\left\\| \\tau \\left(\\mathcal {C} _ {\\theta_ {1}} \\times_ {3} \\mathbf {D} _ {\\theta_ {2}}\\right) - \\mathcal {Y} \\right\\| _ {F} ^ {2} \\tag {14} \\\\ + \\lambda_ {2} \\| \\left(\\mathcal {C} _ {\\theta_ {1}} \\times_ {3} \\mathbf {D} _ {\\theta_ {2}}\\right) \\times_ {3} \\mathbf {R} - \\mathbf {P} \\| _ {F} ^ {2}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\lambda_{2}\\) is a balanced parameter. NSSD models the relationship between the generated detail image and the target HR-HSI. Compared with the classic format \\(\\mathcal{X} = \\mathcal{A} \\times_3 \\mathbf{E}\\), the proposed form \\(\\mathcal{C}_{\\theta_1}(\\mathcal{A}) \\times_3 \\mathbf{D}_{\\theta_2}(\\mathcal{Y})\\) offers a more flexible representation for HR-HSI, particularly for data with higher rank. Furthermore, the NSSD evolves in the guidance of pre-trained DMs, as shown in Eq. (10). The guidance of the" + }, + { + "type": "code_caption", + "bbox": [ + 0.526, + 0.322, + 0.806, + 0.337 + ], + "angle": 0, + "content": "Algorithm 1: The Whole Solving Process" + }, + { + "type": "algorithm", + "bbox": [ + 0.527, + 0.339, + 0.875, + 0.543 + ], + "angle": 0, + "content": "Input: \\(\\mathcal{A}_T\\) sampled from \\(\\mathcal{N}(0,I)\\) ,DM \\(s_\\theta\\) noise scales \\(\\{\\bar{\\alpha}_t\\}_{t = 1}^T\\) \\(f_{\\psi}\\) ,LR-HSI Y,PAN P \nOutput: HR-HSI X \nwhile not converged do for \\(t = T,T - 1,\\dots 1\\) do step 1: estimate \\(\\hat{\\mathcal{A}}_0\\) by Eq. (8) step 2: compute log \\(p_{\\psi}(\\mathcal{V}|\\mathcal{A}_t)\\) by Eq. (10) step 3: sample \\(\\mathcal{A}_{t - 1}\\) by Eq. (11) end NSSD Update: update \\(\\mathcal{M}\\) by minimizing (14) \nend \\(\\mathcal{X} = \\mathcal{C}_{\\theta_1}(\\mathcal{A})\\times_3\\mathbf{D}_{\\theta_2}(\\mathcal{Y})\\) \nreturn X" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.573, + 0.905, + 0.619 + ], + "angle": 0, + "content": "DMs and the update of NSSD are mutually adaptive. Thus, NSSD provides a more flexible and DM-compatible representation of the spatial-spectral correlation in HR-HSI." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.62, + 0.907, + 0.741 + ], + "angle": 0, + "content": "Whole Solving Algorithm: Since the guidance of pretrained DM is related to the mapping \\(\\mathcal{M}\\), the generated detail image is adaptive to the mapping \\(\\mathcal{M}\\). To enhance the flexibility of mapping and the suitability of diffusion guidance, we alternatively conduct guidance of pre-trained DM and NSSD update as shown in Figure 2. During each guidance step, the parameters of the NSSD are held fixed. The whole solving algorithm is summarized in Algorithm 1." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.755, + 0.646, + 0.772 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.902 + ], + "angle": 0, + "content": "To demonstrate the effectiveness of the proposed method, we compare it with corresponding state-of-the-art methods, including GSA [1], CNMF [47], HySure [32], ZSL [11], PLRDiff [30], and HIR-Diff [29]. It is worth to be remarked that HIR-Diff [29] is a general inverse problem framework. We rewrite the fidelity for the hyperspectral pansharpening task. The parameters of all methods are adjusted according to the authors' recommendation. The metrics peak" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12673" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.175, + 0.093, + 0.825, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.233, + 0.908, + 0.263 + ], + "angle": 0, + "content": "Figure 5. Visual comparisons on Pavia dataset. The first row shows the predicted HR-HSIs of different methods, and the second row displays the error maps between predicted HR-HSIs and ground-truth." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.289, + 0.483, + 0.484 + ], + "angle": 0, + "content": "signal-to-noise ratio (PSNR), the structural similarity index (SSIM) [38], the relative dimensionless global error in synthesis (ERGAS) [2], the spectral angle mapper (SAM) [49], and root mean square error (RMSE) are utilized for assessment. QNR [3] is employed for evaluating real-world data. We assess the results produced by these methods on Pavia1, Washington DC (WDC)2, Chikusei3, and FR1 datasets4. The simulated LR-HSI is gained by applying a \\(5 \\times 5\\) Gaussian blurring kernel and the downsampling factor 4. The simulated PAN image is obtained from groundtruth by multiplying a spectral mean response vector. The downsampling factor of FR1 dataset is 6. The detailed information about these datasets is listed as follows," + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.485, + 0.483, + 0.545 + ], + "angle": 0, + "content": "- Pavia dataset: Pavia dataset are originally obtained with the size of \\(610 \\times 340 \\times 115\\). We choose 93 bands by eliminating the bands associated with water vapor absorption and crop the up-left cube with size \\(256 \\times 256 \\times 93\\)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.546, + 0.483, + 0.605 + ], + "angle": 0, + "content": "- WDC dataset: WDC dataset is obtained from the Hyperspectral Digital Imagery Collection Experiment (HYDICE) sensor. We eliminate some unusable bands and choose the area with size \\(256 \\times 256 \\times 191\\)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.606, + 0.483, + 0.665 + ], + "angle": 0, + "content": "- Chikusei dataset: The Headwall Hyperspec-VNIR-C imaging sensor captures the Chikusei dataset with 128 bands and the spectral range is from 363 to \\(1018\\mathrm{~nm}\\). We crop a subimage with size \\(256 \\times 256 \\times 128\\)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.666, + 0.483, + 0.725 + ], + "angle": 0, + "content": "- FR1 real dataset: The FR1 dataset is a real-world data distributed for the PRISMA contest, for pansharpening at the full spatial resolution. We crop a subimage with PAN size \\(240 \\times 240\\)." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.485, + 0.483, + 0.725 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.735, + 0.308, + 0.751 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.757, + 0.483, + 0.817 + ], + "angle": 0, + "content": "All experiments are conducted on the same PC with Pytorch 2.0 and MATLAB (R2022a) on a computer of 64Gb RAM, Intel(R) Core(TM) i9-10900KF CPU @3.70GHz, and NVIDIA GeForce RTX 4070. The update of NSSD" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.826, + 0.308, + 0.838 + ], + "angle": 0, + "content": "1 https://rslab.ut.ac.ir/data" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.839, + 0.482, + 0.864 + ], + "angle": 0, + "content": "2 https://engineering.purdue.edu/biehl/MultiSpec/hyperspectral.html" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.864, + 0.384, + 0.876 + ], + "angle": 0, + "content": "3 https://naotoyokoya.com/Download.html" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.876, + 0.482, + 0.901 + ], + "angle": 0, + "content": "4 https : / / github . com / liangjiandeng / HyperPanCollection" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.826, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.289, + 0.907, + 0.439 + ], + "angle": 0, + "content": "is optimized by Adam optimizer with 0.001 learning rate (LR). The balanced parameter of guidance \\(\\lambda_{1}\\) is 0.1, and \\(\\lambda_{2} = 1\\). The diffusion reverse process step number is set to 100. The number of the whole iterations is 20000. The training of the zero-shot neural network is with Adam optimizer with 2000 iterations, and the LR is 0.001 decayed by gamma \\(= 0.8\\) each 1000 iterations. The rank of NSSD for Pavia, WDC, and Chikusei datasets are set to 20, 40, and 40, respectively. For FR1 real dataset, \\(\\lambda_{1} = \\lambda_{2} = 10\\), and the number of the whole iterations is 10000." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.45, + 0.906, + 0.478 + ], + "angle": 0, + "content": "Table 1. Quantitative results for Pavia dataset. (Bold: best; Underline: second best)" + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.489, + 0.912, + 0.626 + ], + "angle": 0, + "content": "
MethodPSNR↑SSIM↑ERGAS↓SAM↓RMSE↓
GSA [1]07*TGRS28.3040.8125.8197.75310.368
CNMF [47] 11*TGRS28.8960.8595.4955.9499.558
HySure [32] 14*TGRS25.8470.7587.86110.05914.065
ZSL [11]23'TPAMI32.0600.9053.5495.3186.819
PLRDiff [30] 24'IF33.0600.9053.2915.3346.001
HIR-Diff [29]24'CVPR31.9740.8753.6184.7056.742
Proposed34.2270.9262.8634.4635.284
Ideal value+∞1000
" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.652, + 0.74, + 0.667 + ], + "angle": 0, + "content": "4.2. Results on Pavia Dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.907, + 0.84 + ], + "angle": 0, + "content": "In this part, we evaluate the effectiveness of the proposed method on the Pavia dataset. Figure 5 shows the visual comparison of these methods in the benchmark. Diffusion-based methods, i.e., PLRDiff [30], HIR-Diff [29], and the proposed method, achieve excellent performance for detail extraction. The residual images, i.e., the second row of Figure 5 obviously verify that our method can preserve more details compared to others. The result produced by HySure [32] leads to spatial distortions. Table 1 provides the quantitative results for compared approaches. For all metrics, the proposed method achieves the best performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.741, + 0.864 + ], + "angle": 0, + "content": "4.3. Results on WDC Dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "For the WDC dataset, there exist more challenges to preserve the spectral information due to its more spectral" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12674" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.175, + 0.092, + 0.825, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.233, + 0.908, + 0.262 + ], + "angle": 0, + "content": "Figure 6. Visual comparisons on WDC dataset. The first row shows the predicted HR-HSIs of different methods, and the second row displays the error maps between predicted HR-HSIs and ground-truth." + }, + { + "type": "image", + "bbox": [ + 0.175, + 0.277, + 0.825, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.418, + 0.908, + 0.446 + ], + "angle": 0, + "content": "Figure 7. Visual comparisons on Chikusei dataset. The first row shows the predicted HR-HSIs of different methods, and the second row displays the error maps between predicted HR-HSIs and ground-truth." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.47, + 0.483, + 0.498 + ], + "angle": 0, + "content": "Table 2. Quantitative results for WDC dataset. (Bold: best; Underline: second best)" + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.509, + 0.489, + 0.646 + ], + "angle": 0, + "content": "
MethodPSNR↑SSIM↑ERGAS↓SAM↓RMSE↓
GSA [1]07*TGRS39.6290.9105.9247.0797.665
CNMF [47] 11*TGRS39.7030.8975.8646.4487.665
HySure [32] 14*TGRS35.0710.7989.92611.89812.928
ZSL [11]23*TPAMI39.6850.9166.3357.8257.301
PLRDiff [30] 24*IF34.7990.83318.2029.5448.667
HIR-Diff [29]24*CVPR38.5060.8447.1026.6868.079
Proposed40.5030.94312.2545.0474.631
Ideal value+∞1000
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.665, + 0.483, + 0.692 + ], + "angle": 0, + "content": "Table 3. Quantitative results for Chikusei dataset. (Bold: best; Underline: second best)" + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.704, + 0.489, + 0.841 + ], + "angle": 0, + "content": "
MethodPSNR↑SSIM↑ERGAS↓SAM↓RMSE↓
GSA [1]07*TGRS32.4590.88710.8886.7626.662
CNMF [47] 11*TGRS33.5780.8788.5924.8935.905
HySure [32] 14*TGRS30.2700.75214.1599.8838.417
ZSL [11]23'TPAMI35.2580.9178.3675.6854.632
PLRDiff [30] 24'IF37.1760.9176.9944.8383.651
HIR-Diff [29]24'CVPR36.3890.9056.3333.8944.247
Proposed38.4500.9395.9454.1973.192
Ideal value+∞1000
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "bands. As shown in Table 2, we provide the visual results of different methods. Our approach also performs best on" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.472, + 0.907, + 0.563 + ], + "angle": 0, + "content": "this dataset. Compared with traditional method CNMF [47], diffusion-based methods PLRDiff [30] and HIR-Diff [29] lead to spectral distortions. As demonstrated in Figure 6, PLRDiff [30] also preserves excellent spatial detail. The proposed method has the advantages of both spatial detail preservation and spectral information extraction." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.574, + 0.766, + 0.588 + ], + "angle": 0, + "content": "4.4. Results on Chikusei Dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.596, + 0.907, + 0.718 + ], + "angle": 0, + "content": "Figure 7 evaluates the effectiveness of the proposed method on the Chikusei dataset. Our method achieves the best result due to the spatial detail preservation. For the detail edge of the image, the proposed method can preserve the spatial detail well. As shown in Table 3, the proposed method demonstrates improvements in PSNR by \\(1.274\\mathrm{dB} / 2.061\\mathrm{dB}\\), respectively, compared with diffusion-based methods PLRDiff [30] and HIR-Diff [29]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.727, + 0.667, + 0.743 + ], + "angle": 0, + "content": "4.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.885 + ], + "angle": 0, + "content": "To demonstrate the effectiveness of the proposed zero-shot guidance and NSSD, we conduct the ablation study on the Pavia dataset. Table 4 shows the results produced by Four models. For the model I, we use the fixed transformation in [30] and regular fidelity to substitute the proposed zero-shot guidance and NSSD, respectively. As shown in Table 4, the results produced by the model with both zero-shot guidance and NSSD are best. Besides, we design more detailed experiments to explore more properties of the two parts." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.886, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Zero-shot Guidance: The zero-shot guidance achieves" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12675" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.089, + 0.483, + 0.119 + ], + "angle": 0, + "content": "Table 4. Ablation study about the proposed zero-shot (ZS) guidance and neural spatial-spectral decoder (NSSD) on Pavia dataset." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.129, + 0.482, + 0.205 + ], + "angle": 0, + "content": "
AblationZSNSSDPSNRSSIMERGASSAMRMSE
Model I✗✗33.0600.9053.2915.3346.001
Model II✓✗33.2670.9123.2195.2855.877
Model III✗✓33.9720.9212.9464.6055.460
Proposed✓✓34.1790.9252.8774.5335.332
" + }, + { + "type": "image", + "bbox": [ + 0.191, + 0.217, + 0.387, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.345, + 0.483, + 0.375 + ], + "angle": 0, + "content": "Figure 8. Singular value distributions of results produced by PLRDiff [30], HIR-Diff [29], and the proposed method." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.388, + 0.483, + 0.418 + ], + "angle": 0, + "content": "Table 5. Performance of proposed zero-shot guidance compared with conventional regularization term." + }, + { + "type": "table", + "bbox": [ + 0.128, + 0.428, + 0.445, + 0.498 + ], + "angle": 0, + "content": "
GuidancePSNRSSIMERGASSAMRMSE
Fidelity32.4700.8963.4555.1146.395
TV32.7700.9023.3565.1566.204
Proposed33.3740.9143.1745.1505.793
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.522, + 0.483, + 0.627 + ], + "angle": 0, + "content": "excellent guided performance. To demonstrate the effectiveness, we compare it with the conventional total variation regularized term, which is widely used and effective for image processing tasks. As shown in Table 5, the regularized term is helpful for the guidance of DMs. The proposed zero-shot guidance offers more specific prior knowledge for the pre-trained DM to achieve the best performance." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.628, + 0.483, + 0.78 + ], + "angle": 0, + "content": "NSSD: NSSD describes the spatial-spectral correlation with the low-rank property. Figure 8 shows the singular value distribution of target HR-HSI produced by different methods. We compare the proposed method with NSSD with two fixed Diffusion transformation estimation methods in PLRDiff [30] and HIR-Diff [29]. One can observe that the proposed NSSD provides a flexible transformation and performs the best approximation with the pre-trained DM guidance. Our method can achieve the closest distance between the singular values of ground-truth." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.788, + 0.258, + 0.803 + ], + "angle": 0, + "content": "4.6. Guided Network" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In this part, we analyze the influence of zero-shot network structure for guided performance. Table 6 shows the final results with U-Net with skip-connections [35], DSNet [51], and CONet [11]. U-Net contains multi-scale information, DSNet is a structure used in the supervised method, and CONet is a light convolution-based cascade structure." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.089, + 0.907, + 0.119 + ], + "angle": 0, + "content": "Table 6. Performance of zero-shot guided network with different network structures." + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.129, + 0.898, + 0.2 + ], + "angle": 0, + "content": "
Guided StructurePSNRSSIMERGASSAMRMSE
U-Net34.0150.9212.9534.5115.426
DSNet33.7750.9173.0334.6775.609
CONet34.2270.9262.8634.4635.284
" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.224, + 0.907, + 0.27 + ], + "angle": 0, + "content": "The proposed method with the above network structures can achieve excellent results. The convolution-based cascade structure achieves the best performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.279, + 0.686, + 0.294 + ], + "angle": 0, + "content": "4.7. More Discussions" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.301, + 0.907, + 0.346 + ], + "angle": 0, + "content": "Real-data Experiment: Table 7 shows the evaluation on the FR1 real-world dataset, which demonstrates the superior performance of the proposed method on real data." + }, + { + "type": "table_caption", + "bbox": [ + 0.572, + 0.359, + 0.848, + 0.373 + ], + "angle": 0, + "content": "Table 7. Real data experiment on FR1 dataset." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.384, + 0.898, + 0.414 + ], + "angle": 0, + "content": "
MethodGSACNMFHySureZSLPLRDiffHIR-DiffProposed
QNRT0.5900.6790.6680.8020.6300.6250.832
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.428, + 0.906, + 0.473 + ], + "angle": 0, + "content": "More Experimental Comparisons: We provide further discussion on running time, downsampling factor, and kernel size, which are shown in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.473, + 0.906, + 0.58 + ], + "angle": 0, + "content": "Limitations: While the computational cost of DMs with the proposed zero-shot guidance is comparable to that of previous DM-based methods, such as PLRDiff [30], NSSD requires more time to iteratively construct the spectral mapping from the RGB detail image to the HR-HSI, as detailed in the supplementary material. We will explore the use of pre-trained networks to address this limitation in the future." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.592, + 0.634, + 0.608 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.618, + 0.907, + 0.8 + ], + "angle": 0, + "content": "In this paper, we introduced a novel guided diffusion scheme for hyperspectral pansharpening, combining the iterative update of zero-shot guidance and NSSD. The zero-shot guidance is conducted with an auxiliary convolutional network that can be trained just on one-pair inputs, while NSSD acts as a flexible transformation mapping the generated images by pre-trained DMs to the hyperspectral images. Our method demonstrated significant effectiveness on benchmark datasets, including Pavia, WDC, and Chikusei. Through extensive ablation studies, we analyzed the contributions of the guidance and NSSD and evaluated the impact by different zero-shot network architectures." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.817, + 0.907, + 0.893 + ], + "angle": 0, + "content": "Acknowledgement This work was supported in part by the NSFC (12171072, 12271083, 62273077), Natural Science Foundation of Sichuan Province (2024NSFSC0038), and National Key Research and Development Program of China (2020YFA0714001)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "12676" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.17 + ], + "angle": 0, + "content": "[1] B. Aiazzi, S. Baronti, and M. Selva. Improving component substitution pansharpening through multivariate regression of MS + Pan data. IEEE Trans. Geosci. Remote Sens., 45 (10):3230-3239, 2007. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.172, + 0.484, + 0.226 + ], + "angle": 0, + "content": "[2] L. Alparone, L. Wald, J. Chanussot, C. Thomas, P. Gamba, and L. M. Bruce. Comparison of pansharpening algorithms: Outcome of the 2006 GRS-S data-fusion contest. IEEE Trans. Geosci. Remote Sens., 45(10):3012-3021, 2007. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.228, + 0.483, + 0.282 + ], + "angle": 0, + "content": "[3] L. Alparone, B. Aiazzi, S. Baronti, A. Garzelli, F. Nencini, and M. Selva. Multispectral and panchromatic data fusion assessment without reference. Photogramm. Eng. Remote Sens., 74(2):193-200, 2008. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.284, + 0.482, + 0.338 + ], + "angle": 0, + "content": "[4] W. G. C. Bandara, J. M. J. Valanarasu, and V. M. Patel. Hyperspectral pansharpening based on improved deep image prior and residual reconstruction. EEE Trans. Geosci. Remote Sens., 60:1-16, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.34, + 0.482, + 0.394 + ], + "angle": 0, + "content": "[5] Z. Cao, S. Cao, L.-J. Deng, X. Wu, J. Hou, and G. Vivone. Diffusion model with disentangled modulations for sharpening multispectral and hyperspectral images. Inf. Fusion, 104: 102158, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.397, + 0.482, + 0.437 + ], + "angle": 0, + "content": "[6] H. Chung, J. Kim, M. T. McCann, M. L. Klasky, and J. C. Ye. Diffusion posterior sampling for general noisy inverse problems. arXiv preprint arXiv:2209.14687, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.439, + 0.482, + 0.492 + ], + "angle": 0, + "content": "[7] M. Ciotola, G. Guarino, G. Vivone, G. Poggi, J. Chanussot, A. Plaza, and G. Scarpa. Hyperspectral pansharpening: Critical review, tools and future perspectives. arXiv preprint arXiv:2407.01355, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.494, + 0.482, + 0.549 + ], + "angle": 0, + "content": "[8] L.-J. Deng, G. Vivone, M. E. Paoletti, G. Scarpa, J. He, Y. Zhang, J. Chanussot, and A. Plaza. Machine learning in pan-sharpening: A benchmark, from shallow to deep networks. IEEE Geosci. Remote Sens. Mag., 10(3):279-315, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.551, + 0.482, + 0.591 + ], + "angle": 0, + "content": "[9] P. Dhariwal and A. Nichol. Diffusion models beat gans on image synthesis. Adv. Neural Inf. Process. Syst. (NeurIPS), 34:8780-8794, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.593, + 0.482, + 0.647 + ], + "angle": 0, + "content": "[10] R. Dian, L. Fang, and S. Li. Hyperspectral image superresolution via non-local sparse tensor factorization. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 5344-5353, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.649, + 0.482, + 0.69 + ], + "angle": 0, + "content": "[11] R. Dian, A. Guo, and S. Li. Zero-shot hyperspectral sharpening. IEEE Trans. Pattern Anal. Mach. Intell., 45(10):12650-12666, 2023. 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.692, + 0.482, + 0.734 + ], + "angle": 0, + "content": "[12] W. Dong, T. Zhang, J. Qu, S. Xiao, J. Liang, and Y. Li. Laplacian pyramid dense network for hyperspectral pansharpening. IEEE Trans. Geosci. Remote Sens., 60:1-13, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.735, + 0.482, + 0.788 + ], + "angle": 0, + "content": "[13] Y. Duan, X. Wu, H. Deng, and L.-J. Deng. Content-adaptive non-local convolution for remote sensing pansharpening. In Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recog., pages 27738-27747, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.79, + 0.482, + 0.844 + ], + "angle": 0, + "content": "[14] X. Fu, Z. Lin, Y. Huang, and X. Ding. A variational pan-sharpening with local gradient constraints. In Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recog., pages 10265-10274, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[15] J. Gao, J. Li, X. Su, M. Jiang, and Q. Yuan. Deep image interpolation: A unified unsupervised framework for pansharpening. In Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recog., pages 609-618, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[16] M. Gong, H. Zhang, H. Wang, J. Chen, J. Huang, X. Tian, and J. Ma. Cross-scale domain adaptation with comprehensive information for pansharpening. In Int. Joint Conf. Artif. Intell, pages 785-793, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[17] L. He, H. Ye, D. Xi, J. Li, A. Plaza, and M. Zhang. Tree-structured neural network for hyperspectral pansharpening. IEEE J. Sel. Top. Appl. Earth Observ. Remote Sens., 17: 2516-2530, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.207, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[18] L. He, Z. Fang, J. Li, J. Chanussot, and A. Plaza. Two spectral-spatial implicit neural representations for arbitrary-resolution hyperspectral pansharpening. IEEE Trans. Geosci. Remote Sens., 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.263, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[19] W. He, Q. Yao, C. Li, N. Yokoya, and Q. Zhao. Non-local meets global: An integrated paradigm for hyperspectral denoising. In Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recog., pages 6868-6877, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.321, + 0.905, + 0.374 + ], + "angle": 0, + "content": "[20] W. He, Q. Yao, C. Li, N. Yokoya, Q. Zhao, H. Zhang, and L. Zhang. Non-local meets global: An iterative paradigm for hyperspectral image restoration. IEEE Trans. Pattern Anal. Mach. Intell., 44(4):2089-2107, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.377, + 0.905, + 0.418 + ], + "angle": 0, + "content": "[21] Jo. Ho, A. Jain, and P. Abbeel. Denoising diffusion probabilistic models. In Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), pages 6840-6851, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.42, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[22] S. Huang, M. Ren, Y. Yang, X. Wang, and Y. Wei. MFTN: A multi-scale feature transfer network based on IMatchformer for hyperspectral image super-resolution. In *Forty-first Int. Conf. Mach. Learn.*, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.904, + 0.517 + ], + "angle": 0, + "content": "[23] B. Kawar, M. Elad, S. Ermon, and J. Song. Denoising diffusion restoration models. Proc. Adv. Neural Inf. Process. Syst., 35:23593-23606, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.904, + 0.546 + ], + "angle": 0, + "content": "[24] Tamara G Kolda and Brett W Bader. Tensor decompositions and applications. SIAM Rev., 51(3):455-500, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.904, + 0.589 + ], + "angle": 0, + "content": "[25] S. Li, S. Li, and L. Zhang. Hyperspectral and panchromatic images fusion based on the dual conditional diffusion models. IEEE Trans. Geosci. Remote Sens., 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.591, + 0.905, + 0.658 + ], + "angle": 0, + "content": "[26] L. Loncan, L. B. De Almeida, J. M. Bioucas-Dias, X. Briottet, J. Chanussot, N. Dobigeon, S. Fabre, W. Liao, G. A. Licciardi, M. Simoes, et al. Hyperspectral pansharpening: A review. IEEE Geosci. Remote Sens. Mag., 3(3):27-46, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.662, + 0.905, + 0.717 + ], + "angle": 0, + "content": "[27] Y. Miao, L. Zhang, L. Zhang, and D. Tao. DDS2M: Self-supervised denoising diffusion spatio-spectral model for hyperspectral image restoration. In Proc. IEEE/CVF Int. Conf. Comput. Vis., pages 12086-12096, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.719, + 0.905, + 0.786 + ], + "angle": 0, + "content": "[28] A. Quinn Nichol, P. Dhariwal, A. Ramesh, P. Shyam, P. Mishkin, B. Mcgrew, I. Sutskever, and M. Chen. GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. In Int. Conf. Mach. Learn., pages 16784-16804. PMLR, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.789, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[29] L. Pang, X. Rui, L. Cui, H. Wang, D. Meng, and X. Cao. HIR-Diff: Unsupervised hyperspectral image restoration via improved diffusion models. In Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recog., pages 3005-3014, 2024. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[30] X. Rui, X. Cao, L. Pang, Z. Zhu, Z. Yue, and D. Meng. Unsupervised hyperspectral pansharpening via low-rank diffusion model. Inf. Fusion, 107:102325, 2024. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12677" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[31] K. Shi, J. Peng, J. Gao, Y. Luo, and S. Xu. Hyperspectral image denoising via double subspace deep prior. IEEE Trans. Geosci. Remote Sens., 62:1-15, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.19 + ], + "angle": 0, + "content": "[32] M. Simoes, J. Bioucas-Dias, L. B. Almeida, and J. Chanussot. A convex formulation for hyperspectral image superresolution via subspace-based regularization. IEEE Trans. Geosci. Remote Sens., 53(6):3373-3388, 2014. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.482, + 0.234 + ], + "angle": 0, + "content": "[33] Y. Song and S. Ermon. Generative modeling by estimating gradients of the data distribution. In Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), pages 11895-11907, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.236, + 0.482, + 0.29 + ], + "angle": 0, + "content": "[34] Y. Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole. Score-based generative modeling through stochastic differential equations. In Int. Conf. Learn. Representations, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.292, + 0.482, + 0.333 + ], + "angle": 0, + "content": "[35] D. Ulyanov, A. Vedaldi, and V. Lempitsky. Deep image prior. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 9446-9454, 2018. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.335, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[36] G. Vivone, M. Dalla Mura, A. Garzelli, R. Restaino, G. Scarpa, M. O. Ulfarsson, L. Alparone, and J. Chanussot. A new benchmark based on recent advances in multispectral pansharpening: Revisiting pansharpening with classical and emerging pansharpening methods. IEEE Geosci. Remote Sens. Mag., 9(1):53-81, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[37] W. Wang, L.-J. Deng, R. Ran, and G. Vivone. A general paradigm with detail-preserving conditional invertible network for image fusion. Int. J. Comput. Vis., 132(4):1029-1054, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.482, + 0.53 + ], + "angle": 0, + "content": "[38] Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Trans. Image Process., 13(4):600-612, 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.482, + 0.587 + ], + "angle": 0, + "content": "[39] C. Wu, D. Wang, Y. Bai, H. Mao, Y. Li, and Q. Shen. HSR-Diff: Hyperspectral image super-resolution via conditional diffusion models. In Proc. IEEE/CVF Int. Conf. Comput. Vis., pages 7083-7093, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.482, + 0.644 + ], + "angle": 0, + "content": "[40] J.-L. Xiao, T.-Z. Huang, L.-J. Deng, Z.-C. Wu, X. Wu, and G. Vivone. Variational pansharpening based on coefficient estimation with nonlocal regression. IEEE Trans. Geosci. Remote. Sens., 61:1-15, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.482, + 0.687 + ], + "angle": 0, + "content": "[41] W. Xie, J. Lei, Y. Cui, Y. Li, and Q. Du. Hyperspectral pansharpening with deep priors. IEEE Trans. Neural Netw. Learn. Syst., 31(5):1529-1543, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.482, + 0.743 + ], + "angle": 0, + "content": "[42] Y. Xing, L. Qu, S. Zhang, K. Zhang, Y. Zhang, and L. Bruzzone. Crossdiff: Exploring self-supervised representation of pansharpening via cross-predictive diffusion model. IEEE Trans. Image Process., 33:5496-5509, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.482, + 0.8 + ], + "angle": 0, + "content": "[43] T. Xu, T.-Z. Huang, L.-J. Deng, and N. Yokoya. An iterative regularization method based on tensor subspace representation for hyperspectral image super-resolution. IEEE Trans. Geosci. Remote Sens., 60:1-16, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.869 + ], + "angle": 0, + "content": "[44] T. Xu, T.-Z. Huang, L.-J. Deng, J.-L. Xiao, C. Broni-Bediako, J. Xia, and N. Yokoya. A coupled tensor double-factor method for hyperspectral and multispectral image fusion. IEEE Trans. Geosci. Remote Sens., 62:1-17, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[45] J. Xue, Y.-Q. Zhao, Y. Bu, W. Liao, J. C.-W. Chan, and W. Philips. Spatial-spectral structured sparse low-rank represent" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "tation for hyperspectral image super-resolution. IEEE Trans. Image Process., 30:3084-3097, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[46] J. Yao, D. Hong, J. Chanussot, D. Meng, X. Zhu, and Z. Xu. Cross-attention in coupled unmixing nets for unsupervised hyperspectral super-resolution. In Proc. Eur. Conf. Comput. Vis., pages 208–224. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[47] N. Yokoya, T. Yairi, and A. Iwasaki. Coupled nonnegative matrix factorization unmixing for hyperspectral and multispectral data fusion. IEEE Trans. Geosci. Remote Sens., 50 (2):528-537, 2011. 1, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.905, + 0.301 + ], + "angle": 0, + "content": "[48] Y. Yu, E. Pan, X. Wang, Y. Wu, X. Mei, and J. Ma. Unmixing before fusion: A generalized paradigm for multi-source-based hyperspectral image synthesis. In Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recog., pages 9297-9306, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.373 + ], + "angle": 0, + "content": "[49] R. H. Yuhas, A. F. H. Goetz, and J. W. Boardman. Discrimination among semi-arid landscape endmembers using the spectral angle mapper (SAM) algorithm. In Proc. Summaries 3rd Annu. JPL Airborne Geosci. Workshop, pages 147-149, 1992. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.376, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[50] Y. Zhong, X. Wu, L.-J. Deng, and Z. Cao. SSDiff: Spatial-spectral integrated diffusion model for remote sensing pan-sharpening. arXiv preprint arXiv:2404.11537, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.419, + 0.905, + 0.486 + ], + "angle": 0, + "content": "[51] Y.-W. Zhuo, T.-J. Zhang, J.-F. Hu, H.-X. Dou, T.-Z. Huang, and L.-J. Deng. A deep-shallow fusion network with multi-detail extractor and spectral attention for hyperspectral pan-sharpening. IEEE J. Sel. Top. Appl. Earth Observ. Remote Sens., 15:7539-7555, 2022. 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.486 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12678" + } + ] +] \ No newline at end of file diff --git a/2025/Hyperspectral Pansharpening via Diffusion Models with Iteratively Zero-Shot Guidance/f79e03e7-ea12-4e70-879c-79258b8c5e82_origin.pdf b/2025/Hyperspectral Pansharpening via Diffusion Models with Iteratively Zero-Shot Guidance/f79e03e7-ea12-4e70-879c-79258b8c5e82_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..16b37d0cd5c555c88faf357f46f7ceb22eb14040 --- /dev/null +++ b/2025/Hyperspectral Pansharpening via Diffusion Models with Iteratively Zero-Shot Guidance/f79e03e7-ea12-4e70-879c-79258b8c5e82_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:08785acb1d4c85923dc6f5b7842b134a5ceb5365025165844db23498f972db30 +size 3670941 diff --git a/2025/Hyperspectral Pansharpening via Diffusion Models with Iteratively Zero-Shot Guidance/full.md b/2025/Hyperspectral Pansharpening via Diffusion Models with Iteratively Zero-Shot Guidance/full.md new file mode 100644 index 0000000000000000000000000000000000000000..af79330c693cfa11b4a01b05f9f61be31e6e5533 --- /dev/null +++ b/2025/Hyperspectral Pansharpening via Diffusion Models with Iteratively Zero-Shot Guidance/full.md @@ -0,0 +1,350 @@ +# Hyperspectral Pansharpening via Diffusion Models with Iteratively Zero-Shot Guidance + +Jin-Liang Xiao $^{1,3}$ , Ting-Zhu Huang $^{1*}$ , Liang-Jian Deng $^{1}$ , Guang Lin $^{2,3}$ , Zihan Cao $^{1}$ , Chao Li $^{3}$ , Qibin Zhao $^{3*}$ + +1 University of Electronic Science and Technology of China, Chengdu, China + +$^{2}$ Tokyo University of Agriculture and Technology, Tokyo, Japan + +$^{3}$ RIKEN AIP, Tokyo, Japan + +# Abstract + +Hyperspectral pansharpening refers to fusing a panchromatic image (PAN) and a low-resolution hyperspectral image (LR-HSI) to obtain a high-resolution hyperspectral image (HR-HSI). Recently, guiding pre-trained diffusion models (DMs) has demonstrated significant potential in this area, leveraging their powerful representational abilities while avoiding complex training processes. However, these DMs are often trained on RGB images, not well-suited for pansharpening tasks, limited in adapting to the hyperspectral images. In this work, we propose a novel guided diffusion scheme with zero-shot guidance and neural spatial-spectral decomposition (NSSD) to iteratively generate the RGB detail image and map the RGB detail image to target HR-HSI. Specifically, zero-shot guidance employs an auxiliary neural network that trained only with a PAN and LR-HSI to guide pre-trained DMs in generating the RGB detail image, informed by specific prior knowledge. Then, NSSD establishes a spectral mapping from the generated RGB detail image to the final HR-HSI. Extensive experiments are conducted on Pavia, Washington DC, Chukusei, and FR1 datasets to demonstrate that the proposed method significantly enhances the performance of DMs for hyperspectral pansharpening tasks, outperforming existing methods across multiple metrics and achieving improvements in visualization results. The code is available at https://github.com/Jin-liangXiao/DM-zs. + +# 1. Introduction + +Hyperspectral images (HSIs), rich in spatial and spectral information, have been extensively used in a variety of applications [22, 46, 48]. In practice, however, the spatial and spectral information in HSIs are often divided across different image types: the panchromatic image (PANs) with high spatial resolution and low-resolution hyperspectral images + +![](images/b768e3fbca1ad5a62a06bdf83fbcf3164d5f2a554d9c1f9958669366b5633d83.jpg) +Figure 1. The upper row: illustration of the proposed method that includes iterative update of zero-shot guidance for pre-trained diffusion models and spectral mapping from RGB detail image to target HR-HSI. The bottom row: error map comparison between state-of-the-art pre-trained diffusion based hyperspectral pansharpening models (i.e., PLRDiff [30], HIR-Diff [29], and the proposed method). + +(LR-HSIs) with abundant spectral information, due to hardware constraints. This separation highlights the need for hyperspectral pansharpening, which aims to fuse these image types to generate high-resolution hyperspectral images (HR-HSIs) [26]. + +Methods for hyperspectral pansharpening generally fall into two categories: single-image approaches [10, 14, 16, 41, 47] and deep learning (DL) methods [4, 12, 13, 13, 15, 17, 18]. Single-image approaches require only a single image pair instead of an entire dataset for training, which can save significant training costs [36]. While single-image approaches are effective, DL methods can provide stronger + +representations due to the powerful feature-extraction ability of deep neural networks. However, DL methods come with the trade-off of requiring intensive computational resources for training [7, 18, 37] and consistency in the dataset [40]. Specifically, if there is a distribution shift [8] between the training data and testing data, the performance can degrade significantly, which leads to poor generalization. This issue is particularly serious in hyperspectral pan-sharpening, where hyperspectral data are both more varying and limited than data in general computer vision tasks. + +Foundation models, particularly diffusion models (DMs) [21, 34], have gained wide attention in computer vision for their capability of generating high-quality images. In the context of hyperspectral pansharpening, DMs have shown improved generalization due to flexibility of the guidance techniques. Recent studies [29, 30] have demonstrated the potential of DMs, extending pre-trained DMs to hyperspectral pansharpening through specialized guidance. These methods utilize the degenerate process of pansharpening to guide the DM. Due to the powerful representational capabilities, pre-trained DM as a foundation model can be quickly transferred to hyperspectral pansharpening tasks, which can reduce a significant amount of training costs. However, since existing methods lack the consideration of specific prior knowledge of hyperspectral data, the guidance of DMs still needs improvement. The extension from pre-trained DMs to the hyperspectral pansharpening task is non-trivial: We need effective guidance to adapt the pre-trained DMs originally designed with only RGB channels, to manage the complex spectral fidelity while preserving high-resolution spatial details. This adaptation becomes even more challenging when we only have a single (PAN, LR-HSI) image pair. These challenges naturally give rise to two important questions: 1) How can we shift the data distribution of pre-trained DMs to HSIs for hyperspectral pansharpening with limited data through effective guidance? 2) How to accurately transform the RGB information of the generated images to HR-HSIs? + +To address the aforementioned challenges, we propose a novel guided diffusion scheme with zero-shot guidance and neural spatial-spectral decomposition (NSSD) to iteratively generate an RGB spatial detail image and map the detail image to the target HR-HSI. As shown in the green box of Figure 1, we utilize the pre-trained DMs as the backbone, with all their parameters being frozen. Then, we introduce the specific prior knowledge of testing data by a trainable zero-shot network to guide the DMs to generate the detail image enriched with expected features, as shown in the blue box. Unlike conventional deep neural networks that require extensive training data and complex training processes, the proposed zero-shot network is a lightweight model with few parameters, can be trained naively on a single testing image pair, i.e., a LR-HSI and PAN. This network encodes + +specific prior knowledge into pre-trained DMs and guides the DMs to generate detail image more effectively (see Figure 3). Additionally, motivated by the tensor decomposition in single-image approaches [44], we recognize that spatial-spectral correlation is critical for the quality of final data and introduce a neural spatial-spectral decomposition (NSSD) to capture the spatial-spectral correlation and map from the RGB detail image to the final HR-HSI. The NSSD and the guidance of pre-trained DMs are updated iteratively, as displayed in the gray box of Figure 1, making NSSD adaptive to the results of guided DMs. Compared with existing pretrained DMs methods, the proposed method achieves the state-of-the-art result (refer to the bottom row of Figure 1). + +The main contributions are summarized as follows: + +- We propose a zero-shot network to efficiently guide the pre-trained DM to generate a detail image enriched with specific prior knowledge. +- We introduce a neural spatial-spectral decomposition (NSSD) to iteratively construct the target HR-HSI from the generated detail image. +- Experiments on the Pavia, Washington DC, Chikusei, and FR1 datasets demonstrate that the proposed method achieves the state-of-the-art performance. + +# 2. Preliminaries & Related Works + +In this section, we provide definitions, notations, degenerate process of pansharpening, a brief introduction to score-based DMs with guidance, and related works. + +# 2.1. Notations + +Throughout the paper, we denote tensors $\mathcal{X} \in \mathbb{R}^{H \times W \times S}$ , $\mathcal{Y} \in \mathbb{R}^{h \times w \times S}$ , $\mathbf{P} \in \mathbb{R}^{H \times W}$ as HR-HSI, LR-HSI, PAN, respectively. For tensor $\mathcal{X}$ , $\mathbf{X}_{(3)} \in \mathbb{R}^{S \times HW}$ denotes its mode-3 unfolding [24]. Additionally, $\mathcal{X} \times_3 \mathbf{R}$ indicates the mode-3 multiplication between tensor $\mathcal{X}$ and matrix $\mathbf{R}$ . $\mathbf{x} \sim \mathcal{N}(\mathbf{0}, \mathbf{I})$ means the variable $\mathbf{x}$ satisfies Gaussian distribution with the mean of $\mathbf{0}$ and variance of matrix $\mathbf{I}$ . + +# 2.2. Degenerate Process of Pansharpening + +For the hyperspectral pansharpening task, inputs PAN and LR-HSI are mainly modeled as HR-HSI after the spectral and spatial degradation, respectively [45]. The degenerate process of the hyperspectral pansharpening problem can be mathematically represented as follows: + +$$ +\mathcal {Y} = \tau (\mathcal {X}), \tag {1} +$$ + +$$ +\mathbf {P} = \mathcal {X} \times_ {3} \mathbf {R}, +$$ + +where $\tau$ indicates blurring and down-sampling operations, and $\mathbf{R} \in \mathbb{R}^{1 \times S}$ denotes the spectral response function [43]. The goal of the task is to obtain the HR-HSI $\mathcal{X}$ from the LR-HSI $\mathcal{Y}$ and PAN $\mathbf{P}$ . + +# 2.3. Score-based Diffusion Models with Guidance + +Diffusion models (DMs) [21, 33] have proven to be a potent class of generative models capable of generating high-quality images. Among them, score-based DMs are a powerful family that models the forward process by stochastic differential equations (SDEs) and the reverse process by corresponding reverse-time SDEs [34]. Given a input $\mathbf{x}$ , we have + +$$ +\mathrm {d} \mathbf {x} = f (\mathbf {x}, t) \mathrm {d} t + g (t) \mathrm {d} \mathbf {w}, \tag {2} +$$ + +$$ +\mathrm {d} \mathbf {x} = [ f (\mathbf {x}, t) - g ^ {2} (t) \nabla_ {\mathbf {x}} \log p _ {t} (\mathbf {x}) ] + g (t) \mathrm {d} \bar {\mathbf {w}}, \quad (3) +$$ + +where $t$ represents timestep, $f(\cdot ,t)$ denotes a vector valued function called drift coefficient and $g(t)$ is a scalar valued function called diffusion term. $\mathbf{w}$ and $\bar{\mathbf{w}}$ represent a Wiener process and a reverse-time Wiener process, respectively. Eq. (2), corresponding to the forward SDE, indicates adding noise from $t = 0$ to $T$ , which diffuses the image to pure Gaussian noise. In the reverse process, the corresponding reverse-time SDEs, i.e., Eq. (3), we utilize a learned score network $s_{\theta}(\mathbf{x}_t,t)\coloneqq \nabla_{\mathbf{x}}\log p_t(\mathbf{x})$ to generate denoised sample $\mathbf{x}_0$ from a simple Gaussian noise $\mathbf{x}_T\sim \mathcal{N}(\mathbf{0},\mathbf{I})$ . To more effectively enable DMs to generate the desired images, Dhariwal et al. and Pang et al. [9, 29] proposed the guided DM, which is a valid path for general specific tasks, leveraging the powerful generative capabilities of DMs without the need for complex retraining. The guidance of pre-trained DMs works by introducing gradient in reverse process, which can be represented as + +$$ +\hat {s} _ {\theta} (\mathbf {x} _ {t}, t) = s _ {\theta} (\mathbf {x} _ {t}, t) + s (t) \cdot \nabla_ {\mathbf {x} _ {t}} \log p (y | \mathbf {x} _ {t}), \tag {4} +$$ + +where $s(t)$ is the guidance strength in $t$ step, and $p(y|\mathbf{x}_t)$ is the relationship between estimated $\mathbf{x}_t$ and condition $y$ , which is crucial for the guidance of DMs on specific tasks. + +# 2.4. Related Works + +Leveraging the powerful generative capabilities of DMs, numerous studies have been conducted on hyperspectral pansharpening [5, 25, 39, 42, 50]. However, the performance of these models often depends on complex and time-consuming training processes and substantial training data. Recently, efforts have also been made to bridge the connection between inverse problems and pre-trained DMs. Kawar et al. [23] employed a pre-trained DM to solve linear inverse problems for RGB images. Since pre-trained DMs are primarily trained on RGB data, there remains a gap between the generated outputs and the target hyperspectral images. Miao et al. [27] further extended this approach to hyperspectral image restoration. + +For the hyperspectral pansharpening task, a nonlinear inverse problem, different loss functions are proposed to + +guide DMs. Rui et al. [30] proposed a pre-trained diffusion framework that guides the diffusion process to generate an RGB detail image $\mathcal{A}$ . This image is then transformed into the final pansharpened hyperspectral image by applying a fixed transformation matrix along the third dimension, as $\mathcal{X} = \mathcal{A} \times_3 \mathbf{E}$ , where $\mathcal{X} \in \mathbb{R}^{H \times W \times S}$ is the target HR-HSI image, $\mathcal{A} \in \mathbb{R}^{H \times W \times b}$ is the image generated by a pre-trained DM, and $\mathbf{E} \in \mathbb{R}^{S \times b}$ is the fixed transform matrix. Pang et al. [29] explore a new choice for estimating the transform matrix $\mathbf{E}$ and apply total variation (TV) prior to diffusion guided function. However, this issue remains only partially resolved, as their fixed transformation matrix $\mathbf{E}$ in the overall algorithm causes the final result to depend heavily on the initial estimate of $\mathbf{E}$ . In contrast, our method introduces an auxiliary neural network to provide the specialized guidance for generating an RGB detail image and iteratively employs a neural spatial-spectral decomposition to adaptively construct the mapping from the RGB detail image to HR-HSI. + +# 3. Methodology + +In this section, we elaborate on the proposed guided diffusion shcheme. As illustrated in Figure 2, the scheme primarily consists of two components: diffusion with zero-shot guidance and neural spatial-spectral decomposition (NSSD). The zero-shot guidance operates on the pre-trained DM to generate an RGB detail image, enabling better preservation of spatial details and adaptation to specific prior knowledge. NSSD establishes the mapping from the detail image to the HR-HSI. + +# 3.1. Diffusion with Zero-Shot Guidance + +Unlike the existing DM-based hyperspectral pansharpening methods [29, 30], we aim to guide the reverse process of DMs for generating the RGB detail image $\mathcal{A}$ by introducing the specific prior knowledge of testing data. To this end, we design a novel diffusion scheme with zero-shot guidance, where the reverse process of $\mathcal{A}$ satisfies SDE as follows: + +$$ +\begin{array}{l} \mathrm {d} \mathcal {A} _ {t} = g (t) \mathrm {d} \bar {w} + f (\mathcal {A}, t) \mathrm {d} t \\ - g ^ {2} (t) \left[ s _ {\theta} \left(x _ {t}, t\right) + \nabla_ {\mathcal {A} _ {t}} \log p _ {t} (\tilde {\mathcal {Y}} | \mathcal {A} _ {t}) \right] \mathrm {d} t, \tag {5} \\ \end{array} +$$ + +where $p_t(\tilde{\mathcal{Y}}|\mathcal{A}_t)$ is the pansharpening guidance estimated from the condition, defined as + +$$ +\tilde {\mathcal {Y}} = (\mathcal {Y}, \mathbf {P}, \mathcal {M}, f _ {\psi} (\mathcal {Y}, \mathbf {P})), \tag {6} +$$ + +where $\mathcal{V}$ and $\mathbf{P}$ are input image pair LR-HSI and PAN, $\mathcal{M}$ is a mapping from RGB detail image to HR-HSI (deferred in next subsection), and $f_{\psi}(\mathcal{V},\mathbf{P})$ is an auxiliary neural network that is trained with $\mathcal{V}$ and $\mathbf{P}$ . Following [9], we have + +$$ +p _ {\theta , \psi} \left(\mathcal {A} _ {t - 1} \mid \mathcal {A} _ {t}, \tilde {\mathcal {Y}}\right) \propto p _ {\theta} \left(\mathcal {A} _ {t - 1} \mid \mathcal {A} _ {t}\right) p _ {\psi} (\tilde {\mathcal {Y}} \mid \mathcal {A} _ {t}). \tag {7} +$$ + +![](images/fcc997ad0e0834c3612cc233579d7b350ce28c7a39fdf0d1cfd930cd8615fb14.jpg) +Figure 2. Illustration of the whole process of the proposed hyperspectral pansharpening method that includes diffusion with zero-shot (ZS) guidance and neural spatial-spectral decomposition (NSSD). The part in the dashed box is the proposed zero-shot guided function that is trained with a single testing image pair. NSSD builds the spectral mapping from the detail image generated by DM to HR-HSI. The guidance of the DM and NSSD are iteratively updated. + +![](images/93f32f0d57dbd6cf88c0f002b0fd4207fc0e1b02cf9311977ef3ddb57173ae1c.jpg) +Figure 3. The analysis of the proposed zero-shot guidance. The DMs with the proposed guidance can achieve better performance with fewer sampling steps. (The length of the yellow box means the distance of PSNR values between the DM with and without ZS guidance) + +Based on [6], we define the variable $\hat{\mathcal{A}}_0\coloneqq \mathbb{E}[\mathcal{A}_0|\mathcal{A}_t] =$ $\mathbb{E}_{\mathcal{A}_0\sim p(\mathcal{A}_0|\mathcal{A}_t)}[\mathcal{A}_0]$ , which means + +$$ +\hat {\mathcal {A}} _ {0} \simeq \frac {1}{\sqrt {\bar {\alpha} _ {t}}} \left(\mathcal {A} _ {t} + \left(1 - \bar {\alpha} _ {t}\right) s _ {\theta} \left(\mathcal {A} _ {t}, t\right)\right), \tag {8} +$$ + +$$ +p _ {\psi} (\tilde {\mathcal {V}} | \mathcal {A} _ {t}) \simeq p _ {\psi} (\tilde {\mathcal {V}} | \hat {\mathcal {A}} _ {0}), \tag {9} +$$ + +where $\bar{\alpha}_t\coloneqq \prod_{s = 1}^t\alpha_s$ is the noise scale of DMs. We propose the zero-shot guidance as follows, + +$$ +\begin{array}{l} \log p _ {\psi} (\tilde {\mathcal {Y}} | \hat {\mathcal {A}} _ {0}) = \left\| \tau (\mathcal {M} (\hat {\mathcal {A}} _ {0})) - \mathcal {Y} \right\| _ {F} ^ {2} \\ + \left\| \mathcal {M} \left(\hat {\mathcal {A}} _ {0}\right) \times_ {3} \mathbf {R} - \mathbf {P} \right\| _ {F} ^ {2} + \lambda_ {1} \left\| \mathcal {M} \left(\hat {\mathcal {A}} _ {0}\right) - f _ {\psi} (\mathcal {Y}, \mathbf {P}) \right\| _ {F} ^ {2}, \tag {10} \\ \end{array} +$$ + +where $\lambda_{1}$ is a weighting scale. Finally, the reverse process is illustrated as: + +$$ +\begin{array}{l} \mathcal {A} _ {t - 1} = \frac {1}{\sqrt {\alpha_ {t}}} \left(\mathcal {A} _ {t} - \frac {\beta_ {t}}{\sqrt {1 - \bar {\alpha} _ {t}}} s _ {\theta} \left(\mathcal {A} _ {t}, t\right)\right) \tag {11} \\ - \lambda \cdot \frac {\beta_ {t}}{\sqrt {1 - \bar {\alpha} _ {t}} \sqrt {\alpha_ {t}}} \nabla_ {\mathcal {A} _ {t}} l o g p _ {\psi} (\tilde {\mathcal {Y}} | \hat {\mathcal {A}} _ {0}), \\ \end{array} +$$ + +where $\beta_{t}$ , as a hyperparameter, is generally set to linearly interpolate from 0.0001 to 0.02 [28] and $\alpha_{t} := 1 - \beta_{t}$ . The weighting scale $\lambda$ can balance the effect of guidance. + +The rest of this part is how to train the zero-shot network $f_{\psi}(\mathcal{V}, \mathbf{P})$ . The network structure and training process are shown within the dotted box in Figure 2, which has been verified its effectiveness in [11]. Based on the "invariance among scales" assumption, we train the network in the reduced-resolution scale. In specific, we use the input LR-HSI as the training ground-truth. Training PAN and training LR-HSI are simulated by PAN and LR-HSI after spatial and spectral degenerate processes, respectively. This approach allows for efficient training of the zero-shot network using only a single testing input pair. Figure 3 shows the change curve of PSNR values with the increase of iteration number. One can observe that the DM guided by zero-shot guidance can achieve better performance with fewer sampling steps. Besides, this single-image training property provides specific prior knowledge to the pre-trained DM. + +# 3.2. Neural Spatial-Spectral Decomposition + +This subsection describes neural spatial-spectral decomposition (NSSD), i.e., $\mathcal{M}$ in Eq. (10). In hyperspectral pan-sharpening, beyond the fidelity term in (1), the spatial-spectral correlation exhibits the low-rank properties characteristic of hyperspectral data [44]. The mode-3 decom + +![](images/ac02864a374f0b054b5152395b5a575ed2d6a20fcd90661d03ec33e8054a09b3.jpg) +Figure 4. The structure of neural spatial-spectral decomposition (NSSD) consists of a U-net structure with skip connection and the fully connection structure to extract spatial and spectral information, respectively. + +position format, i.e., $\mathcal{X} = \mathcal{A}\times_3\mathbf{E}$ , has shown to be effective [19, 20, 31], which satisfies the following inequation: + +$$ +\operatorname {r a n k} \left(\mathbf {X} _ {(3)}\right) \leq \min \left\{\operatorname {r a n k} \left(\mathbf {A} _ {(3)}\right), \operatorname {r a n k} (\mathbf {E}) \right\} \leq b, \tag {12} +$$ + +where $\mathcal{A} \in \mathbb{R}^{H \times W \times b}$ , and $\mathbf{E} \in \mathbb{R}^{S \times b}$ . The low value of the size of $\mathbf{E}$ limits the expressive ability. Since pre-trained DMs are primarily trained on RGB images, the size of $\mathcal{A}$ is typically set to $H \times W \times 3$ . Additionally, the larger value of $r$ in (12) inevitably increases the time required for the reverse process in the DM. + +To mitigate the drawbacks, we propose a NSSD to capture the spatial-spectral correlation, as follows: + +$$ +\mathcal {X} = \mathcal {M} (\mathcal {A}) := \mathcal {C} _ {\theta_ {1}} (\mathcal {A}) \times_ {3} \mathbf {D} _ {\theta_ {2}} (\mathcal {Y}), \tag {13} +$$ + +where $\mathcal{C}_{\theta_1}(\mathcal{A}) \in \mathbb{R}^{H \times W \times r}$ and $\mathbf{D}_{\theta_2}(\mathcal{Y}) \in \mathbb{R}^{S \times r}$ with hyperparameter $r$ are produced by two neural network with the input $\mathcal{A} \in \mathbb{R}^{H \times W \times 3}$ and $\mathcal{Y} \in \mathbb{R}^{h \times w \times S}$ , respectively, as illustrated in Figure 4. Specifically, $\mathcal{C}_{\theta_1}(\cdot)$ is the spatial-factor network that extracts the spatial information from the RGB detail image. The structure of $\mathcal{C}_{\theta_1}(\cdot)$ is the U-net type with skip connection, whose detail-extraction capability has been verified [35]. The spectral-factor network, $\mathbf{D}_{\theta_2}(\cdot)$ , consists of the fully connection structure and sigmoid to efficiently absorb spectral information from LR-HSI. NSSD does not require pre-training. It produces the final result through iterative steps using the following loss: + +$$ +\begin{array}{l} L o s s = \left\| \tau \left(\mathcal {C} _ {\theta_ {1}} \times_ {3} \mathbf {D} _ {\theta_ {2}}\right) - \mathcal {Y} \right\| _ {F} ^ {2} \tag {14} \\ + \lambda_ {2} \| \left(\mathcal {C} _ {\theta_ {1}} \times_ {3} \mathbf {D} _ {\theta_ {2}}\right) \times_ {3} \mathbf {R} - \mathbf {P} \| _ {F} ^ {2}, \\ \end{array} +$$ + +where $\lambda_{2}$ is a balanced parameter. NSSD models the relationship between the generated detail image and the target HR-HSI. Compared with the classic format $\mathcal{X} = \mathcal{A} \times_3 \mathbf{E}$ , the proposed form $\mathcal{C}_{\theta_1}(\mathcal{A}) \times_3 \mathbf{D}_{\theta_2}(\mathcal{Y})$ offers a more flexible representation for HR-HSI, particularly for data with higher rank. Furthermore, the NSSD evolves in the guidance of pre-trained DMs, as shown in Eq. (10). The guidance of the + +Algorithm 1: The Whole Solving Process +Input: $\mathcal{A}_T$ sampled from $\mathcal{N}(0,I)$ ,DM $s_\theta$ noise scales $\{\bar{\alpha}_t\}_{t = 1}^T$ $f_{\psi}$ ,LR-HSI Y,PAN P +Output: HR-HSI X +while not converged do for $t = T,T - 1,\dots 1$ do step 1: estimate $\hat{\mathcal{A}}_0$ by Eq. (8) step 2: compute log $p_{\psi}(\mathcal{V}|\mathcal{A}_t)$ by Eq. (10) step 3: sample $\mathcal{A}_{t - 1}$ by Eq. (11) end NSSD Update: update $\mathcal{M}$ by minimizing (14) +end $\mathcal{X} = \mathcal{C}_{\theta_1}(\mathcal{A})\times_3\mathbf{D}_{\theta_2}(\mathcal{Y})$ +return X + +DMs and the update of NSSD are mutually adaptive. Thus, NSSD provides a more flexible and DM-compatible representation of the spatial-spectral correlation in HR-HSI. + +Whole Solving Algorithm: Since the guidance of pretrained DM is related to the mapping $\mathcal{M}$ , the generated detail image is adaptive to the mapping $\mathcal{M}$ . To enhance the flexibility of mapping and the suitability of diffusion guidance, we alternatively conduct guidance of pre-trained DM and NSSD update as shown in Figure 2. During each guidance step, the parameters of the NSSD are held fixed. The whole solving algorithm is summarized in Algorithm 1. + +# 4. Experiments + +To demonstrate the effectiveness of the proposed method, we compare it with corresponding state-of-the-art methods, including GSA [1], CNMF [47], HySure [32], ZSL [11], PLRDiff [30], and HIR-Diff [29]. It is worth to be remarked that HIR-Diff [29] is a general inverse problem framework. We rewrite the fidelity for the hyperspectral pansharpening task. The parameters of all methods are adjusted according to the authors' recommendation. The metrics peak + +![](images/f7f2fd480b50fd617a25808e4186208f0112c88d0c4b54179e9188276ae04a01.jpg) +Figure 5. Visual comparisons on Pavia dataset. The first row shows the predicted HR-HSIs of different methods, and the second row displays the error maps between predicted HR-HSIs and ground-truth. + +signal-to-noise ratio (PSNR), the structural similarity index (SSIM) [38], the relative dimensionless global error in synthesis (ERGAS) [2], the spectral angle mapper (SAM) [49], and root mean square error (RMSE) are utilized for assessment. QNR [3] is employed for evaluating real-world data. We assess the results produced by these methods on Pavia1, Washington DC (WDC)2, Chikusei3, and FR1 datasets4. The simulated LR-HSI is gained by applying a $5 \times 5$ Gaussian blurring kernel and the downsampling factor 4. The simulated PAN image is obtained from groundtruth by multiplying a spectral mean response vector. The downsampling factor of FR1 dataset is 6. The detailed information about these datasets is listed as follows, + +- Pavia dataset: Pavia dataset are originally obtained with the size of $610 \times 340 \times 115$ . We choose 93 bands by eliminating the bands associated with water vapor absorption and crop the up-left cube with size $256 \times 256 \times 93$ . +- WDC dataset: WDC dataset is obtained from the Hyperspectral Digital Imagery Collection Experiment (HYDICE) sensor. We eliminate some unusable bands and choose the area with size $256 \times 256 \times 191$ . +- Chikusei dataset: The Headwall Hyperspec-VNIR-C imaging sensor captures the Chikusei dataset with 128 bands and the spectral range is from 363 to $1018\mathrm{~nm}$ . We crop a subimage with size $256 \times 256 \times 128$ . +- FR1 real dataset: The FR1 dataset is a real-world data distributed for the PRISMA contest, for pansharpening at the full spatial resolution. We crop a subimage with PAN size $240 \times 240$ . + +# 4.1. Implementation Details + +All experiments are conducted on the same PC with Pytorch 2.0 and MATLAB (R2022a) on a computer of 64Gb RAM, Intel(R) Core(TM) i9-10900KF CPU @3.70GHz, and NVIDIA GeForce RTX 4070. The update of NSSD + +1 https://rslab.ut.ac.ir/data +2 https://engineering.purdue.edu/biehl/MultiSpec/hyperspectral.html +3 https://naotoyokoya.com/Download.html +4 https : / / github . com / liangjiandeng / HyperPanCollection + +is optimized by Adam optimizer with 0.001 learning rate (LR). The balanced parameter of guidance $\lambda_{1}$ is 0.1, and $\lambda_{2} = 1$ . The diffusion reverse process step number is set to 100. The number of the whole iterations is 20000. The training of the zero-shot neural network is with Adam optimizer with 2000 iterations, and the LR is 0.001 decayed by gamma $= 0.8$ each 1000 iterations. The rank of NSSD for Pavia, WDC, and Chikusei datasets are set to 20, 40, and 40, respectively. For FR1 real dataset, $\lambda_{1} = \lambda_{2} = 10$ , and the number of the whole iterations is 10000. + +Table 1. Quantitative results for Pavia dataset. (Bold: best; Underline: second best) + +
MethodPSNR↑SSIM↑ERGAS↓SAM↓RMSE↓
GSA [1]07*TGRS28.3040.8125.8197.75310.368
CNMF [47] 11*TGRS28.8960.8595.4955.9499.558
HySure [32] 14*TGRS25.8470.7587.86110.05914.065
ZSL [11]23'TPAMI32.0600.9053.5495.3186.819
PLRDiff [30] 24'IF33.0600.9053.2915.3346.001
HIR-Diff [29]24'CVPR31.9740.8753.6184.7056.742
Proposed34.2270.9262.8634.4635.284
Ideal value+∞1000
+ +# 4.2. Results on Pavia Dataset + +In this part, we evaluate the effectiveness of the proposed method on the Pavia dataset. Figure 5 shows the visual comparison of these methods in the benchmark. Diffusion-based methods, i.e., PLRDiff [30], HIR-Diff [29], and the proposed method, achieve excellent performance for detail extraction. The residual images, i.e., the second row of Figure 5 obviously verify that our method can preserve more details compared to others. The result produced by HySure [32] leads to spatial distortions. Table 1 provides the quantitative results for compared approaches. For all metrics, the proposed method achieves the best performance. + +# 4.3. Results on WDC Dataset + +For the WDC dataset, there exist more challenges to preserve the spectral information due to its more spectral + +![](images/3343eb19e5774ef750d27c553b0df5666be8552de592b57c9c57c96a60e79c72.jpg) +Figure 6. Visual comparisons on WDC dataset. The first row shows the predicted HR-HSIs of different methods, and the second row displays the error maps between predicted HR-HSIs and ground-truth. + +![](images/9fb2a19da9eb27ccbdf2baff1f5750f2bdf698789dd63b5822ce4babe14d7e45.jpg) +Figure 7. Visual comparisons on Chikusei dataset. The first row shows the predicted HR-HSIs of different methods, and the second row displays the error maps between predicted HR-HSIs and ground-truth. + +Table 2. Quantitative results for WDC dataset. (Bold: best; Underline: second best) + +
MethodPSNR↑SSIM↑ERGAS↓SAM↓RMSE↓
GSA [1]07*TGRS39.6290.9105.9247.0797.665
CNMF [47] 11*TGRS39.7030.8975.8646.4487.665
HySure [32] 14*TGRS35.0710.7989.92611.89812.928
ZSL [11]23*TPAMI39.6850.9166.3357.8257.301
PLRDiff [30] 24*IF34.7990.83318.2029.5448.667
HIR-Diff [29]24*CVPR38.5060.8447.1026.6868.079
Proposed40.5030.94312.2545.0474.631
Ideal value+∞1000
+ +Table 3. Quantitative results for Chikusei dataset. (Bold: best; Underline: second best) + +
MethodPSNR↑SSIM↑ERGAS↓SAM↓RMSE↓
GSA [1]07*TGRS32.4590.88710.8886.7626.662
CNMF [47] 11*TGRS33.5780.8788.5924.8935.905
HySure [32] 14*TGRS30.2700.75214.1599.8838.417
ZSL [11]23'TPAMI35.2580.9178.3675.6854.632
PLRDiff [30] 24'IF37.1760.9176.9944.8383.651
HIR-Diff [29]24'CVPR36.3890.9056.3333.8944.247
Proposed38.4500.9395.9454.1973.192
Ideal value+∞1000
+ +bands. As shown in Table 2, we provide the visual results of different methods. Our approach also performs best on + +this dataset. Compared with traditional method CNMF [47], diffusion-based methods PLRDiff [30] and HIR-Diff [29] lead to spectral distortions. As demonstrated in Figure 6, PLRDiff [30] also preserves excellent spatial detail. The proposed method has the advantages of both spatial detail preservation and spectral information extraction. + +# 4.4. Results on Chikusei Dataset + +Figure 7 evaluates the effectiveness of the proposed method on the Chikusei dataset. Our method achieves the best result due to the spatial detail preservation. For the detail edge of the image, the proposed method can preserve the spatial detail well. As shown in Table 3, the proposed method demonstrates improvements in PSNR by $1.274\mathrm{dB} / 2.061\mathrm{dB}$ , respectively, compared with diffusion-based methods PLRDiff [30] and HIR-Diff [29]. + +# 4.5. Ablation Study + +To demonstrate the effectiveness of the proposed zero-shot guidance and NSSD, we conduct the ablation study on the Pavia dataset. Table 4 shows the results produced by Four models. For the model I, we use the fixed transformation in [30] and regular fidelity to substitute the proposed zero-shot guidance and NSSD, respectively. As shown in Table 4, the results produced by the model with both zero-shot guidance and NSSD are best. Besides, we design more detailed experiments to explore more properties of the two parts. + +Zero-shot Guidance: The zero-shot guidance achieves + +Table 4. Ablation study about the proposed zero-shot (ZS) guidance and neural spatial-spectral decoder (NSSD) on Pavia dataset. + +
AblationZSNSSDPSNRSSIMERGASSAMRMSE
Model I✗✗33.0600.9053.2915.3346.001
Model II✓✗33.2670.9123.2195.2855.877
Model III✗✓33.9720.9212.9464.6055.460
Proposed✓✓34.1790.9252.8774.5335.332
+ +![](images/7afca9c5e68c05a388f3c312d7e8a110f9e6a28c4987f46857b19e2fa81d142a.jpg) +Figure 8. Singular value distributions of results produced by PLRDiff [30], HIR-Diff [29], and the proposed method. + +Table 5. Performance of proposed zero-shot guidance compared with conventional regularization term. + +
GuidancePSNRSSIMERGASSAMRMSE
Fidelity32.4700.8963.4555.1146.395
TV32.7700.9023.3565.1566.204
Proposed33.3740.9143.1745.1505.793
+ +excellent guided performance. To demonstrate the effectiveness, we compare it with the conventional total variation regularized term, which is widely used and effective for image processing tasks. As shown in Table 5, the regularized term is helpful for the guidance of DMs. The proposed zero-shot guidance offers more specific prior knowledge for the pre-trained DM to achieve the best performance. + +NSSD: NSSD describes the spatial-spectral correlation with the low-rank property. Figure 8 shows the singular value distribution of target HR-HSI produced by different methods. We compare the proposed method with NSSD with two fixed Diffusion transformation estimation methods in PLRDiff [30] and HIR-Diff [29]. One can observe that the proposed NSSD provides a flexible transformation and performs the best approximation with the pre-trained DM guidance. Our method can achieve the closest distance between the singular values of ground-truth. + +# 4.6. Guided Network + +In this part, we analyze the influence of zero-shot network structure for guided performance. Table 6 shows the final results with U-Net with skip-connections [35], DSNet [51], and CONet [11]. U-Net contains multi-scale information, DSNet is a structure used in the supervised method, and CONet is a light convolution-based cascade structure. + +Table 6. Performance of zero-shot guided network with different network structures. + +
Guided StructurePSNRSSIMERGASSAMRMSE
U-Net34.0150.9212.9534.5115.426
DSNet33.7750.9173.0334.6775.609
CONet34.2270.9262.8634.4635.284
+ +The proposed method with the above network structures can achieve excellent results. The convolution-based cascade structure achieves the best performance. + +# 4.7. More Discussions + +Real-data Experiment: Table 7 shows the evaluation on the FR1 real-world dataset, which demonstrates the superior performance of the proposed method on real data. + +Table 7. Real data experiment on FR1 dataset. + +
MethodGSACNMFHySureZSLPLRDiffHIR-DiffProposed
QNRT0.5900.6790.6680.8020.6300.6250.832
+ +More Experimental Comparisons: We provide further discussion on running time, downsampling factor, and kernel size, which are shown in the supplementary material. + +Limitations: While the computational cost of DMs with the proposed zero-shot guidance is comparable to that of previous DM-based methods, such as PLRDiff [30], NSSD requires more time to iteratively construct the spectral mapping from the RGB detail image to the HR-HSI, as detailed in the supplementary material. We will explore the use of pre-trained networks to address this limitation in the future. + +# 5. Conclusion + +In this paper, we introduced a novel guided diffusion scheme for hyperspectral pansharpening, combining the iterative update of zero-shot guidance and NSSD. The zero-shot guidance is conducted with an auxiliary convolutional network that can be trained just on one-pair inputs, while NSSD acts as a flexible transformation mapping the generated images by pre-trained DMs to the hyperspectral images. Our method demonstrated significant effectiveness on benchmark datasets, including Pavia, WDC, and Chikusei. Through extensive ablation studies, we analyzed the contributions of the guidance and NSSD and evaluated the impact by different zero-shot network architectures. + +Acknowledgement This work was supported in part by the NSFC (12171072, 12271083, 62273077), Natural Science Foundation of Sichuan Province (2024NSFSC0038), and National Key Research and Development Program of China (2020YFA0714001). + +# References + +[1] B. Aiazzi, S. Baronti, and M. Selva. Improving component substitution pansharpening through multivariate regression of MS + Pan data. IEEE Trans. Geosci. Remote Sens., 45 (10):3230-3239, 2007. 5, 6, 7 +[2] L. Alparone, L. Wald, J. Chanussot, C. Thomas, P. Gamba, and L. M. Bruce. Comparison of pansharpening algorithms: Outcome of the 2006 GRS-S data-fusion contest. IEEE Trans. Geosci. Remote Sens., 45(10):3012-3021, 2007. 6 +[3] L. Alparone, B. Aiazzi, S. Baronti, A. Garzelli, F. Nencini, and M. Selva. Multispectral and panchromatic data fusion assessment without reference. Photogramm. Eng. Remote Sens., 74(2):193-200, 2008. 6 +[4] W. G. C. Bandara, J. M. J. Valanarasu, and V. M. Patel. Hyperspectral pansharpening based on improved deep image prior and residual reconstruction. EEE Trans. Geosci. Remote Sens., 60:1-16, 2021. 1 +[5] Z. Cao, S. Cao, L.-J. Deng, X. Wu, J. Hou, and G. Vivone. Diffusion model with disentangled modulations for sharpening multispectral and hyperspectral images. Inf. Fusion, 104: 102158, 2024. 3 +[6] H. Chung, J. Kim, M. T. McCann, M. L. Klasky, and J. C. Ye. Diffusion posterior sampling for general noisy inverse problems. arXiv preprint arXiv:2209.14687, 2022. 4 +[7] M. Ciotola, G. Guarino, G. Vivone, G. Poggi, J. Chanussot, A. Plaza, and G. Scarpa. Hyperspectral pansharpening: Critical review, tools and future perspectives. arXiv preprint arXiv:2407.01355, 2024. 2 +[8] L.-J. Deng, G. Vivone, M. E. Paoletti, G. Scarpa, J. He, Y. Zhang, J. Chanussot, and A. Plaza. Machine learning in pan-sharpening: A benchmark, from shallow to deep networks. IEEE Geosci. Remote Sens. Mag., 10(3):279-315, 2022. 2 +[9] P. Dhariwal and A. Nichol. Diffusion models beat gans on image synthesis. Adv. Neural Inf. Process. Syst. (NeurIPS), 34:8780-8794, 2021. 3 +[10] R. Dian, L. Fang, and S. Li. Hyperspectral image superresolution via non-local sparse tensor factorization. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 5344-5353, 2017. 1 +[11] R. Dian, A. Guo, and S. Li. Zero-shot hyperspectral sharpening. IEEE Trans. Pattern Anal. Mach. Intell., 45(10):12650-12666, 2023. 4, 5, 6, 7, 8 +[12] W. Dong, T. Zhang, J. Qu, S. Xiao, J. Liang, and Y. Li. Laplacian pyramid dense network for hyperspectral pansharpening. IEEE Trans. Geosci. Remote Sens., 60:1-13, 2021. 1 +[13] Y. Duan, X. Wu, H. Deng, and L.-J. Deng. Content-adaptive non-local convolution for remote sensing pansharpening. In Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recog., pages 27738-27747, 2024. 1 +[14] X. Fu, Z. Lin, Y. Huang, and X. Ding. A variational pan-sharpening with local gradient constraints. In Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recog., pages 10265-10274, 2019. 1 +[15] J. Gao, J. Li, X. Su, M. Jiang, and Q. Yuan. Deep image interpolation: A unified unsupervised framework for pansharpening. In Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recog., pages 609-618, 2022. 1 + +[16] M. Gong, H. Zhang, H. Wang, J. Chen, J. Huang, X. Tian, and J. Ma. Cross-scale domain adaptation with comprehensive information for pansharpening. In Int. Joint Conf. Artif. Intell, pages 785-793, 2024. 1 +[17] L. He, H. Ye, D. Xi, J. Li, A. Plaza, and M. Zhang. Tree-structured neural network for hyperspectral pansharpening. IEEE J. Sel. Top. Appl. Earth Observ. Remote Sens., 17: 2516-2530, 2023. 1 +[18] L. He, Z. Fang, J. Li, J. Chanussot, and A. Plaza. Two spectral-spatial implicit neural representations for arbitrary-resolution hyperspectral pansharpening. IEEE Trans. Geosci. Remote Sens., 2024. 1, 2 +[19] W. He, Q. Yao, C. Li, N. Yokoya, and Q. Zhao. Non-local meets global: An integrated paradigm for hyperspectral denoising. In Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recog., pages 6868-6877, 2019. 5 +[20] W. He, Q. Yao, C. Li, N. Yokoya, Q. Zhao, H. Zhang, and L. Zhang. Non-local meets global: An iterative paradigm for hyperspectral image restoration. IEEE Trans. Pattern Anal. Mach. Intell., 44(4):2089-2107, 2020. 5 +[21] Jo. Ho, A. Jain, and P. Abbeel. Denoising diffusion probabilistic models. In Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), pages 6840-6851, 2020. 2, 3 +[22] S. Huang, M. Ren, Y. Yang, X. Wang, and Y. Wei. MFTN: A multi-scale feature transfer network based on IMatchformer for hyperspectral image super-resolution. In *Forty-first Int. Conf. Mach. Learn.*, 2024. 1 +[23] B. Kawar, M. Elad, S. Ermon, and J. Song. Denoising diffusion restoration models. Proc. Adv. Neural Inf. Process. Syst., 35:23593-23606, 2022. 3 +[24] Tamara G Kolda and Brett W Bader. Tensor decompositions and applications. SIAM Rev., 51(3):455-500, 2009. 2 +[25] S. Li, S. Li, and L. Zhang. Hyperspectral and panchromatic images fusion based on the dual conditional diffusion models. IEEE Trans. Geosci. Remote Sens., 2023. 3 +[26] L. Loncan, L. B. De Almeida, J. M. Bioucas-Dias, X. Briottet, J. Chanussot, N. Dobigeon, S. Fabre, W. Liao, G. A. Licciardi, M. Simoes, et al. Hyperspectral pansharpening: A review. IEEE Geosci. Remote Sens. Mag., 3(3):27-46, 2015. 1 +[27] Y. Miao, L. Zhang, L. Zhang, and D. Tao. DDS2M: Self-supervised denoising diffusion spatio-spectral model for hyperspectral image restoration. In Proc. IEEE/CVF Int. Conf. Comput. Vis., pages 12086-12096, 2023. 3 +[28] A. Quinn Nichol, P. Dhariwal, A. Ramesh, P. Shyam, P. Mishkin, B. Mcgrew, I. Sutskever, and M. Chen. GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. In Int. Conf. Mach. Learn., pages 16784-16804. PMLR, 2022. 4 +[29] L. Pang, X. Rui, L. Cui, H. Wang, D. Meng, and X. Cao. HIR-Diff: Unsupervised hyperspectral image restoration via improved diffusion models. In Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recog., pages 3005-3014, 2024. 1, 2, 3, 5, 6, 7, 8 +[30] X. Rui, X. Cao, L. Pang, Z. Zhu, Z. Yue, and D. Meng. Unsupervised hyperspectral pansharpening via low-rank diffusion model. Inf. Fusion, 107:102325, 2024. 1, 2, 3, 5, 6, 7, 8 + +[31] K. Shi, J. Peng, J. Gao, Y. Luo, and S. Xu. Hyperspectral image denoising via double subspace deep prior. IEEE Trans. Geosci. Remote Sens., 62:1-15, 2024. 5 +[32] M. Simoes, J. Bioucas-Dias, L. B. Almeida, and J. Chanussot. A convex formulation for hyperspectral image superresolution via subspace-based regularization. IEEE Trans. Geosci. Remote Sens., 53(6):3373-3388, 2014. 5, 6, 7 +[33] Y. Song and S. Ermon. Generative modeling by estimating gradients of the data distribution. In Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), pages 11895-11907, 2019. 3 +[34] Y. Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole. Score-based generative modeling through stochastic differential equations. In Int. Conf. Learn. Representations, 2021. 2, 3 +[35] D. Ulyanov, A. Vedaldi, and V. Lempitsky. Deep image prior. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 9446-9454, 2018. 5, 8 +[36] G. Vivone, M. Dalla Mura, A. Garzelli, R. Restaino, G. Scarpa, M. O. Ulfarsson, L. Alparone, and J. Chanussot. A new benchmark based on recent advances in multispectral pansharpening: Revisiting pansharpening with classical and emerging pansharpening methods. IEEE Geosci. Remote Sens. Mag., 9(1):53-81, 2020. 1 +[37] W. Wang, L.-J. Deng, R. Ran, and G. Vivone. A general paradigm with detail-preserving conditional invertible network for image fusion. Int. J. Comput. Vis., 132(4):1029-1054, 2024. 2 +[38] Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Trans. Image Process., 13(4):600-612, 2004. 6 +[39] C. Wu, D. Wang, Y. Bai, H. Mao, Y. Li, and Q. Shen. HSR-Diff: Hyperspectral image super-resolution via conditional diffusion models. In Proc. IEEE/CVF Int. Conf. Comput. Vis., pages 7083-7093, 2023. 3 +[40] J.-L. Xiao, T.-Z. Huang, L.-J. Deng, Z.-C. Wu, X. Wu, and G. Vivone. Variational pansharpening based on coefficient estimation with nonlocal regression. IEEE Trans. Geosci. Remote. Sens., 61:1-15, 2023. 2 +[41] W. Xie, J. Lei, Y. Cui, Y. Li, and Q. Du. Hyperspectral pansharpening with deep priors. IEEE Trans. Neural Netw. Learn. Syst., 31(5):1529-1543, 2019. 1 +[42] Y. Xing, L. Qu, S. Zhang, K. Zhang, Y. Zhang, and L. Bruzzone. Crossdiff: Exploring self-supervised representation of pansharpening via cross-predictive diffusion model. IEEE Trans. Image Process., 33:5496-5509, 2024. 3 +[43] T. Xu, T.-Z. Huang, L.-J. Deng, and N. Yokoya. An iterative regularization method based on tensor subspace representation for hyperspectral image super-resolution. IEEE Trans. Geosci. Remote Sens., 60:1-16, 2022. 2 +[44] T. Xu, T.-Z. Huang, L.-J. Deng, J.-L. Xiao, C. Broni-Bediako, J. Xia, and N. Yokoya. A coupled tensor double-factor method for hyperspectral and multispectral image fusion. IEEE Trans. Geosci. Remote Sens., 62:1-17, 2024. 2, 4 +[45] J. Xue, Y.-Q. Zhao, Y. Bu, W. Liao, J. C.-W. Chan, and W. Philips. Spatial-spectral structured sparse low-rank represent + +tation for hyperspectral image super-resolution. IEEE Trans. Image Process., 30:3084-3097, 2021. 2 +[46] J. Yao, D. Hong, J. Chanussot, D. Meng, X. Zhu, and Z. Xu. Cross-attention in coupled unmixing nets for unsupervised hyperspectral super-resolution. In Proc. Eur. Conf. Comput. Vis., pages 208–224. Springer, 2020. 1 +[47] N. Yokoya, T. Yairi, and A. Iwasaki. Coupled nonnegative matrix factorization unmixing for hyperspectral and multispectral data fusion. IEEE Trans. Geosci. Remote Sens., 50 (2):528-537, 2011. 1, 5, 6, 7 +[48] Y. Yu, E. Pan, X. Wang, Y. Wu, X. Mei, and J. Ma. Unmixing before fusion: A generalized paradigm for multi-source-based hyperspectral image synthesis. In Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recog., pages 9297-9306, 2024. +[49] R. H. Yuhas, A. F. H. Goetz, and J. W. Boardman. Discrimination among semi-arid landscape endmembers using the spectral angle mapper (SAM) algorithm. In Proc. Summaries 3rd Annu. JPL Airborne Geosci. Workshop, pages 147-149, 1992. 6 +[50] Y. Zhong, X. Wu, L.-J. Deng, and Z. Cao. SSDiff: Spatial-spectral integrated diffusion model for remote sensing pan-sharpening. arXiv preprint arXiv:2404.11537, 2024. 3 +[51] Y.-W. Zhuo, T.-J. Zhang, J.-F. Hu, H.-X. Dou, T.-Z. Huang, and L.-J. Deng. A deep-shallow fusion network with multi-detail extractor and spectral attention for hyperspectral pan-sharpening. IEEE J. Sel. Top. Appl. Earth Observ. Remote Sens., 15:7539-7555, 2022. 8 \ No newline at end of file diff --git a/2025/Hyperspectral Pansharpening via Diffusion Models with Iteratively Zero-Shot Guidance/images.zip b/2025/Hyperspectral Pansharpening via Diffusion Models with Iteratively Zero-Shot Guidance/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..90d50552d945c652acb8c4d3c5af1822dff79379 --- /dev/null +++ b/2025/Hyperspectral Pansharpening via Diffusion Models with Iteratively Zero-Shot Guidance/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:299e9fdf9a9bbf38339262bf43aaed8bab05959be8c779e195d24c4028d206ac +size 786282 diff --git a/2025/Hyperspectral Pansharpening via Diffusion Models with Iteratively Zero-Shot Guidance/layout.json b/2025/Hyperspectral Pansharpening via Diffusion Models with Iteratively Zero-Shot Guidance/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..adfc44466eea8479ed43394e7cb024794ba417a8 --- /dev/null +++ b/2025/Hyperspectral Pansharpening via Diffusion Models with Iteratively Zero-Shot Guidance/layout.json @@ -0,0 +1,9397 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 65, + 102, + 545, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 102, + 545, + 138 + ], + "spans": [ + { + "bbox": [ + 65, + 102, + 545, + 138 + ], + "type": "text", + "content": "Hyperspectral Pansharpening via Diffusion Models with Iteratively Zero-Shot Guidance" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 160, + 580, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 160, + 580, + 175 + ], + "spans": [ + { + "bbox": [ + 61, + 160, + 580, + 175 + ], + "type": "text", + "content": "Jin-Liang Xiao" + }, + { + "bbox": [ + 61, + 160, + 580, + 175 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 61, + 160, + 580, + 175 + ], + "type": "text", + "content": ", Ting-Zhu Huang" + }, + { + "bbox": [ + 61, + 160, + 580, + 175 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 61, + 160, + 580, + 175 + ], + "type": "text", + "content": ", Liang-Jian Deng" + }, + { + "bbox": [ + 61, + 160, + 580, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 160, + 580, + 175 + ], + "type": "text", + "content": ", Guang Lin" + }, + { + "bbox": [ + 61, + 160, + 580, + 175 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 61, + 160, + 580, + 175 + ], + "type": "text", + "content": ", Zihan Cao" + }, + { + "bbox": [ + 61, + 160, + 580, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 160, + 580, + 175 + ], + "type": "text", + "content": ", Chao Li" + }, + { + "bbox": [ + 61, + 160, + 580, + 175 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 61, + 160, + 580, + 175 + ], + "type": "text", + "content": ", Qibin Zhao" + }, + { + "bbox": [ + 61, + 160, + 580, + 175 + ], + "type": "inline_equation", + "content": "^{3*}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 175, + 506, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 175, + 506, + 189 + ], + "spans": [ + { + "bbox": [ + 132, + 175, + 506, + 189 + ], + "type": "text", + "content": "1 University of Electronic Science and Technology of China, Chengdu, China" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 162, + 189, + 475, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 189, + 475, + 203 + ], + "spans": [ + { + "bbox": [ + 162, + 189, + 475, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 162, + 189, + 475, + 203 + ], + "type": "text", + "content": " Tokyo University of Agriculture and Technology, Tokyo, Japan" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 249, + 203, + 388, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 203, + 388, + 217 + ], + "spans": [ + { + "bbox": [ + 249, + 203, + 388, + 217 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 249, + 203, + 388, + 217 + ], + "type": "text", + "content": " RIKEN AIP, Tokyo, Japan" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "spans": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 270, + 297, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 270, + 297, + 582 + ], + "spans": [ + { + "bbox": [ + 54, + 270, + 297, + 582 + ], + "type": "text", + "content": "Hyperspectral pansharpening refers to fusing a panchromatic image (PAN) and a low-resolution hyperspectral image (LR-HSI) to obtain a high-resolution hyperspectral image (HR-HSI). Recently, guiding pre-trained diffusion models (DMs) has demonstrated significant potential in this area, leveraging their powerful representational abilities while avoiding complex training processes. However, these DMs are often trained on RGB images, not well-suited for pansharpening tasks, limited in adapting to the hyperspectral images. In this work, we propose a novel guided diffusion scheme with zero-shot guidance and neural spatial-spectral decomposition (NSSD) to iteratively generate the RGB detail image and map the RGB detail image to target HR-HSI. Specifically, zero-shot guidance employs an auxiliary neural network that trained only with a PAN and LR-HSI to guide pre-trained DMs in generating the RGB detail image, informed by specific prior knowledge. Then, NSSD establishes a spectral mapping from the generated RGB detail image to the final HR-HSI. Extensive experiments are conducted on Pavia, Washington DC, Chukusei, and FR1 datasets to demonstrate that the proposed method significantly enhances the performance of DMs for hyperspectral pansharpening tasks, outperforming existing methods across multiple metrics and achieving improvements in visualization results. The code is available at https://github.com/Jin-liangXiao/DM-zs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 604, + 135, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 604, + 135, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 604, + 135, + 616 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 623, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 623, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 623, + 296, + 696 + ], + "type": "text", + "content": "Hyperspectral images (HSIs), rich in spatial and spectral information, have been extensively used in a variety of applications [22, 46, 48]. In practice, however, the spatial and spectral information in HSIs are often divided across different image types: the panchromatic image (PANs) with high spatial resolution and low-resolution hyperspectral images" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 317, + 243, + 555, + 464 + ], + "blocks": [ + { + "bbox": [ + 317, + 243, + 555, + 464 + ], + "lines": [ + { + "bbox": [ + 317, + 243, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 317, + 243, + 555, + 464 + ], + "type": "image", + "image_path": "b768e3fbca1ad5a62a06bdf83fbcf3164d5f2a554d9c1f9958669366b5633d83.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 471, + 556, + 548 + ], + "lines": [ + { + "bbox": [ + 313, + 471, + 556, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 471, + 556, + 548 + ], + "type": "text", + "content": "Figure 1. The upper row: illustration of the proposed method that includes iterative update of zero-shot guidance for pre-trained diffusion models and spectral mapping from RGB detail image to target HR-HSI. The bottom row: error map comparison between state-of-the-art pre-trained diffusion based hyperspectral pansharpening models (i.e., PLRDiff [30], HIR-Diff [29], and the proposed method)." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 569, + 555, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 569, + 555, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 569, + 555, + 629 + ], + "type": "text", + "content": "(LR-HSIs) with abundant spectral information, due to hardware constraints. This separation highlights the need for hyperspectral pansharpening, which aims to fuse these image types to generate high-resolution hyperspectral images (HR-HSIs) [26]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "Methods for hyperspectral pansharpening generally fall into two categories: single-image approaches [10, 14, 16, 41, 47] and deep learning (DL) methods [4, 12, 13, 13, 15, 17, 18]. Single-image approaches require only a single image pair instead of an entire dataset for training, which can save significant training costs [36]. While single-image approaches are effective, DL methods can provide stronger" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 15 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 15 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 15 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 198, + 13, + 443, + 24 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 13, + 443, + 24 + ], + "spans": [ + { + "bbox": [ + 198, + 13, + 443, + 24 + ], + "type": "text", + "content": "Except for this watermark, it is identical to the accepted version;" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "spans": [ + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "type": "text", + "content": "the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12669" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "type": "text", + "content": "representations due to the powerful feature-extraction ability of deep neural networks. However, DL methods come with the trade-off of requiring intensive computational resources for training [7, 18, 37] and consistency in the dataset [40]. Specifically, if there is a distribution shift [8] between the training data and testing data, the performance can degrade significantly, which leads to poor generalization. This issue is particularly serious in hyperspectral pan-sharpening, where hyperspectral data are both more varying and limited than data in general computer vision tasks." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 196, + 296, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 196, + 296, + 530 + ], + "spans": [ + { + "bbox": [ + 56, + 196, + 296, + 530 + ], + "type": "text", + "content": "Foundation models, particularly diffusion models (DMs) [21, 34], have gained wide attention in computer vision for their capability of generating high-quality images. In the context of hyperspectral pansharpening, DMs have shown improved generalization due to flexibility of the guidance techniques. Recent studies [29, 30] have demonstrated the potential of DMs, extending pre-trained DMs to hyperspectral pansharpening through specialized guidance. These methods utilize the degenerate process of pansharpening to guide the DM. Due to the powerful representational capabilities, pre-trained DM as a foundation model can be quickly transferred to hyperspectral pansharpening tasks, which can reduce a significant amount of training costs. However, since existing methods lack the consideration of specific prior knowledge of hyperspectral data, the guidance of DMs still needs improvement. The extension from pre-trained DMs to the hyperspectral pansharpening task is non-trivial: We need effective guidance to adapt the pre-trained DMs originally designed with only RGB channels, to manage the complex spectral fidelity while preserving high-resolution spatial details. This adaptation becomes even more challenging when we only have a single (PAN, LR-HSI) image pair. These challenges naturally give rise to two important questions: 1) How can we shift the data distribution of pre-trained DMs to HSIs for hyperspectral pansharpening with limited data through effective guidance? 2) How to accurately transform the RGB information of the generated images to HR-HSIs?" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": "To address the aforementioned challenges, we propose a novel guided diffusion scheme with zero-shot guidance and neural spatial-spectral decomposition (NSSD) to iteratively generate an RGB spatial detail image and map the detail image to the target HR-HSI. As shown in the green box of Figure 1, we utilize the pre-trained DMs as the backbone, with all their parameters being frozen. Then, we introduce the specific prior knowledge of testing data by a trainable zero-shot network to guide the DMs to generate the detail image enriched with expected features, as shown in the blue box. Unlike conventional deep neural networks that require extensive training data and complex training processes, the proposed zero-shot network is a lightweight model with few parameters, can be trained naively on a single testing image pair, i.e., a LR-HSI and PAN. This network encodes" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "content": "specific prior knowledge into pre-trained DMs and guides the DMs to generate detail image more effectively (see Figure 3). Additionally, motivated by the tensor decomposition in single-image approaches [44], we recognize that spatial-spectral correlation is critical for the quality of final data and introduce a neural spatial-spectral decomposition (NSSD) to capture the spatial-spectral correlation and map from the RGB detail image to the final HR-HSI. The NSSD and the guidance of pre-trained DMs are updated iteratively, as displayed in the gray box of Figure 1, making NSSD adaptive to the results of guided DMs. Compared with existing pretrained DMs methods, the proposed method achieves the state-of-the-art result (refer to the bottom row of Figure 1)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 326, + 228, + 533, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 228, + 533, + 239 + ], + "spans": [ + { + "bbox": [ + 326, + 228, + 533, + 239 + ], + "type": "text", + "content": "The main contributions are summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 241, + 553, + 348 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 314, + 241, + 553, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 241, + 553, + 277 + ], + "spans": [ + { + "bbox": [ + 314, + 241, + 553, + 277 + ], + "type": "text", + "content": "- We propose a zero-shot network to efficiently guide the pre-trained DM to generate a detail image enriched with specific prior knowledge." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 277, + 553, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 553, + 312 + ], + "type": "text", + "content": "- We introduce a neural spatial-spectral decomposition (NSSD) to iteratively construct the target HR-HSI from the generated detail image." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 313, + 553, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 313, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 313, + 553, + 348 + ], + "type": "text", + "content": "- Experiments on the Pavia, Washington DC, Chikusei, and FR1 datasets demonstrate that the proposed method achieves the state-of-the-art performance." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 361, + 490, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 490, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 490, + 373 + ], + "type": "text", + "content": "2. Preliminaries & Related Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 382, + 553, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 382, + 553, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 382, + 553, + 418 + ], + "type": "text", + "content": "In this section, we provide definitions, notations, degenerate process of pansharpening, a brief introduction to score-based DMs with guidance, and related works." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 427, + 382, + 439 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 427, + 382, + 439 + ], + "spans": [ + { + "bbox": [ + 313, + 427, + 382, + 439 + ], + "type": "text", + "content": "2.1. Notations" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 445, + 553, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 553, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 553, + 529 + ], + "type": "text", + "content": "Throughout the paper, we denote tensors " + }, + { + "bbox": [ + 313, + 445, + 553, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{X} \\in \\mathbb{R}^{H \\times W \\times S}" + }, + { + "bbox": [ + 313, + 445, + 553, + 529 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 445, + 553, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{Y} \\in \\mathbb{R}^{h \\times w \\times S}" + }, + { + "bbox": [ + 313, + 445, + 553, + 529 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 445, + 553, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{P} \\in \\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 313, + 445, + 553, + 529 + ], + "type": "text", + "content": " as HR-HSI, LR-HSI, PAN, respectively. For tensor " + }, + { + "bbox": [ + 313, + 445, + 553, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 313, + 445, + 553, + 529 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 445, + 553, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{(3)} \\in \\mathbb{R}^{S \\times HW}" + }, + { + "bbox": [ + 313, + 445, + 553, + 529 + ], + "type": "text", + "content": " denotes its mode-3 unfolding [24]. Additionally, " + }, + { + "bbox": [ + 313, + 445, + 553, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{X} \\times_3 \\mathbf{R}" + }, + { + "bbox": [ + 313, + 445, + 553, + 529 + ], + "type": "text", + "content": " indicates the mode-3 multiplication between tensor " + }, + { + "bbox": [ + 313, + 445, + 553, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 313, + 445, + 553, + 529 + ], + "type": "text", + "content": " and matrix " + }, + { + "bbox": [ + 313, + 445, + 553, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 313, + 445, + 553, + 529 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 445, + 553, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})" + }, + { + "bbox": [ + 313, + 445, + 553, + 529 + ], + "type": "text", + "content": " means the variable " + }, + { + "bbox": [ + 313, + 445, + 553, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 445, + 553, + 529 + ], + "type": "text", + "content": " satisfies Gaussian distribution with the mean of " + }, + { + "bbox": [ + 313, + 445, + 553, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{0}" + }, + { + "bbox": [ + 313, + 445, + 553, + 529 + ], + "type": "text", + "content": " and variance of matrix " + }, + { + "bbox": [ + 313, + 445, + 553, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 313, + 445, + 553, + 529 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 539, + 512, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 539, + 512, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 539, + 512, + 552 + ], + "type": "text", + "content": "2.2. Degenerate Process of Pansharpening" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 557, + 553, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 553, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 553, + 617 + ], + "type": "text", + "content": "For the hyperspectral pansharpening task, inputs PAN and LR-HSI are mainly modeled as HR-HSI after the spectral and spatial degradation, respectively [45]. The degenerate process of the hyperspectral pansharpening problem can be mathematically represented as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 405, + 628, + 553, + 646 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 628, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 405, + 628, + 553, + 646 + ], + "type": "interline_equation", + "content": "\\mathcal {Y} = \\tau (\\mathcal {X}), \\tag {1}", + "image_path": "c97b23d1bd58ee94338ad9f8c16dcf18fa3b007c240f752bbbdfaa59854da2df.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 405, + 643, + 464, + 654 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 643, + 464, + 654 + ], + "spans": [ + { + "bbox": [ + 405, + 643, + 464, + 654 + ], + "type": "interline_equation", + "content": "\\mathbf {P} = \\mathcal {X} \\times_ {3} \\mathbf {R},", + "image_path": "60ac7967eeea4eb968d7a1bb5e98053f7201766b32f7c758f15a6b62650f1156.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "type": "text", + "content": " indicates blurring and down-sampling operations, and " + }, + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{R} \\in \\mathbb{R}^{1 \\times S}" + }, + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "type": "text", + "content": " denotes the spectral response function [43]. The goal of the task is to obtain the HR-HSI " + }, + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "type": "text", + "content": " from the LR-HSI " + }, + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}" + }, + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "type": "text", + "content": " and PAN " + }, + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12670" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 289, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 289, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 289, + 83 + ], + "type": "text", + "content": "2.3. Score-based Diffusion Models with Guidance" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 296, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 296, + 172 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 296, + 172 + ], + "type": "text", + "content": "Diffusion models (DMs) [21, 33] have proven to be a potent class of generative models capable of generating high-quality images. Among them, score-based DMs are a powerful family that models the forward process by stochastic differential equations (SDEs) and the reverse process by corresponding reverse-time SDEs [34]. Given a input " + }, + { + "bbox": [ + 55, + 89, + 296, + 172 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 89, + 296, + 172 + ], + "type": "text", + "content": ", we have" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 119, + 182, + 295, + 196 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 182, + 295, + 196 + ], + "spans": [ + { + "bbox": [ + 119, + 182, + 295, + 196 + ], + "type": "interline_equation", + "content": "\\mathrm {d} \\mathbf {x} = f (\\mathbf {x}, t) \\mathrm {d} t + g (t) \\mathrm {d} \\mathbf {w}, \\tag {2}", + "image_path": "6d1bb4720ebc0a161bf4e3d12a89130fd5c92ec7e4d56630f3ca55f24b87e2e7.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 79, + 213, + 295, + 228 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 213, + 295, + 228 + ], + "spans": [ + { + "bbox": [ + 79, + 213, + 295, + 228 + ], + "type": "interline_equation", + "content": "\\mathrm {d} \\mathbf {x} = [ f (\\mathbf {x}, t) - g ^ {2} (t) \\nabla_ {\\mathbf {x}} \\log p _ {t} (\\mathbf {x}) ] + g (t) \\mathrm {d} \\bar {\\mathbf {w}}, \\quad (3)", + "image_path": "2d97ab9e4c64cc884ab6488cfff116f0bff106374aeb3441f6e50c3d8dbe4877.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 236, + 296, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 236, + 296, + 439 + ], + "spans": [ + { + "bbox": [ + 55, + 236, + 296, + 439 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 236, + 296, + 439 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 236, + 296, + 439 + ], + "type": "text", + "content": " represents timestep, " + }, + { + "bbox": [ + 55, + 236, + 296, + 439 + ], + "type": "inline_equation", + "content": "f(\\cdot ,t)" + }, + { + "bbox": [ + 55, + 236, + 296, + 439 + ], + "type": "text", + "content": " denotes a vector valued function called drift coefficient and " + }, + { + "bbox": [ + 55, + 236, + 296, + 439 + ], + "type": "inline_equation", + "content": "g(t)" + }, + { + "bbox": [ + 55, + 236, + 296, + 439 + ], + "type": "text", + "content": " is a scalar valued function called diffusion term. " + }, + { + "bbox": [ + 55, + 236, + 296, + 439 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 55, + 236, + 296, + 439 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 236, + 296, + 439 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{w}}" + }, + { + "bbox": [ + 55, + 236, + 296, + 439 + ], + "type": "text", + "content": " represent a Wiener process and a reverse-time Wiener process, respectively. Eq. (2), corresponding to the forward SDE, indicates adding noise from " + }, + { + "bbox": [ + 55, + 236, + 296, + 439 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 55, + 236, + 296, + 439 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 236, + 296, + 439 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 236, + 296, + 439 + ], + "type": "text", + "content": ", which diffuses the image to pure Gaussian noise. In the reverse process, the corresponding reverse-time SDEs, i.e., Eq. (3), we utilize a learned score network " + }, + { + "bbox": [ + 55, + 236, + 296, + 439 + ], + "type": "inline_equation", + "content": "s_{\\theta}(\\mathbf{x}_t,t)\\coloneqq \\nabla_{\\mathbf{x}}\\log p_t(\\mathbf{x})" + }, + { + "bbox": [ + 55, + 236, + 296, + 439 + ], + "type": "text", + "content": " to generate denoised sample " + }, + { + "bbox": [ + 55, + 236, + 296, + 439 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 55, + 236, + 296, + 439 + ], + "type": "text", + "content": " from a simple Gaussian noise " + }, + { + "bbox": [ + 55, + 236, + 296, + 439 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 55, + 236, + 296, + 439 + ], + "type": "text", + "content": ". To more effectively enable DMs to generate the desired images, Dhariwal et al. and Pang et al. [9, 29] proposed the guided DM, which is a valid path for general specific tasks, leveraging the powerful generative capabilities of DMs without the need for complex retraining. The guidance of pre-trained DMs works by introducing gradient in reverse process, which can be represented as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 82, + 450, + 295, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 450, + 295, + 464 + ], + "spans": [ + { + "bbox": [ + 82, + 450, + 295, + 464 + ], + "type": "interline_equation", + "content": "\\hat {s} _ {\\theta} (\\mathbf {x} _ {t}, t) = s _ {\\theta} (\\mathbf {x} _ {t}, t) + s (t) \\cdot \\nabla_ {\\mathbf {x} _ {t}} \\log p (y | \\mathbf {x} _ {t}), \\tag {4}", + "image_path": "9fd4e34ecf19856e2e4891e18873c277b48682f9469a369378c9f6ed85fb2bfd.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 473, + 296, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 296, + 510 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 473, + 296, + 510 + ], + "type": "inline_equation", + "content": "s(t)" + }, + { + "bbox": [ + 55, + 473, + 296, + 510 + ], + "type": "text", + "content": " is the guidance strength in " + }, + { + "bbox": [ + 55, + 473, + 296, + 510 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 473, + 296, + 510 + ], + "type": "text", + "content": " step, and " + }, + { + "bbox": [ + 55, + 473, + 296, + 510 + ], + "type": "inline_equation", + "content": "p(y|\\mathbf{x}_t)" + }, + { + "bbox": [ + 55, + 473, + 296, + 510 + ], + "type": "text", + "content": " is the relationship between estimated " + }, + { + "bbox": [ + 55, + 473, + 296, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 55, + 473, + 296, + 510 + ], + "type": "text", + "content": " and condition " + }, + { + "bbox": [ + 55, + 473, + 296, + 510 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 473, + 296, + 510 + ], + "type": "text", + "content": ", which is crucial for the guidance of DMs on specific tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 517, + 148, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 148, + 528 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 148, + 528 + ], + "type": "text", + "content": "2.4. Related Works" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 534, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 689 + ], + "type": "text", + "content": "Leveraging the powerful generative capabilities of DMs, numerous studies have been conducted on hyperspectral pansharpening [5, 25, 39, 42, 50]. However, the performance of these models often depends on complex and time-consuming training processes and substantial training data. Recently, efforts have also been made to bridge the connection between inverse problems and pre-trained DMs. Kawar et al. [23] employed a pre-trained DM to solve linear inverse problems for RGB images. Since pre-trained DMs are primarily trained on RGB data, there remains a gap between the generated outputs and the target hyperspectral images. Miao et al. [27] further extended this approach to hyperspectral image restoration." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "For the hyperspectral pansharpening task, a nonlinear inverse problem, different loss functions are proposed to" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "type": "text", + "content": "guide DMs. Rui et al. [30] proposed a pre-trained diffusion framework that guides the diffusion process to generate an RGB detail image " + }, + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "type": "text", + "content": ". This image is then transformed into the final pansharpened hyperspectral image by applying a fixed transformation matrix along the third dimension, as " + }, + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "type": "inline_equation", + "content": "\\mathcal{X} = \\mathcal{A} \\times_3 \\mathbf{E}" + }, + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "type": "inline_equation", + "content": "\\mathcal{X} \\in \\mathbb{R}^{H \\times W \\times S}" + }, + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "type": "text", + "content": " is the target HR-HSI image, " + }, + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "type": "inline_equation", + "content": "\\mathcal{A} \\in \\mathbb{R}^{H \\times W \\times b}" + }, + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "type": "text", + "content": " is the image generated by a pre-trained DM, and " + }, + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "type": "inline_equation", + "content": "\\mathbf{E} \\in \\mathbb{R}^{S \\times b}" + }, + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "type": "text", + "content": " is the fixed transform matrix. Pang et al. [29] explore a new choice for estimating the transform matrix " + }, + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "type": "text", + "content": " and apply total variation (TV) prior to diffusion guided function. However, this issue remains only partially resolved, as their fixed transformation matrix " + }, + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "type": "text", + "content": " in the overall algorithm causes the final result to depend heavily on the initial estimate of " + }, + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "type": "text", + "content": ". In contrast, our method introduces an auxiliary neural network to provide the specialized guidance for generating an RGB detail image and iteratively employs a neural spatial-spectral decomposition to adaptively construct the mapping from the RGB detail image to HR-HSI." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 310, + 397, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 310, + 397, + 323 + ], + "spans": [ + { + "bbox": [ + 314, + 310, + 397, + 323 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 330, + 555, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 330, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 330, + 555, + 437 + ], + "type": "text", + "content": "In this section, we elaborate on the proposed guided diffusion shcheme. As illustrated in Figure 2, the scheme primarily consists of two components: diffusion with zero-shot guidance and neural spatial-spectral decomposition (NSSD). The zero-shot guidance operates on the pre-trained DM to generate an RGB detail image, enabling better preservation of spatial details and adaptation to specific prior knowledge. NSSD establishes the mapping from the detail image to the HR-HSI." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 445, + 502, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 502, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 502, + 456 + ], + "type": "text", + "content": "3.1. Diffusion with Zero-Shot Guidance" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 462, + 554, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 554, + 534 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 554, + 534 + ], + "type": "text", + "content": "Unlike the existing DM-based hyperspectral pansharpening methods [29, 30], we aim to guide the reverse process of DMs for generating the RGB detail image " + }, + { + "bbox": [ + 313, + 462, + 554, + 534 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 462, + 554, + 534 + ], + "type": "text", + "content": " by introducing the specific prior knowledge of testing data. To this end, we design a novel diffusion scheme with zero-shot guidance, where the reverse process of " + }, + { + "bbox": [ + 313, + 462, + 554, + 534 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 462, + 554, + 534 + ], + "type": "text", + "content": " satisfies SDE as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 318, + 543, + 553, + 573 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 543, + 553, + 573 + ], + "spans": [ + { + "bbox": [ + 318, + 543, + 553, + 573 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathrm {d} \\mathcal {A} _ {t} = g (t) \\mathrm {d} \\bar {w} + f (\\mathcal {A}, t) \\mathrm {d} t \\\\ - g ^ {2} (t) \\left[ s _ {\\theta} \\left(x _ {t}, t\\right) + \\nabla_ {\\mathcal {A} _ {t}} \\log p _ {t} (\\tilde {\\mathcal {Y}} | \\mathcal {A} _ {t}) \\right] \\mathrm {d} t, \\tag {5} \\\\ \\end{array}", + "image_path": "329e93887cf9489cd5fa73765eea46b788312a2ae5f751fabca807c8b62e69ed.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 582, + 555, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 607 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 607 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 582, + 555, + 607 + ], + "type": "inline_equation", + "content": "p_t(\\tilde{\\mathcal{Y}}|\\mathcal{A}_t)" + }, + { + "bbox": [ + 313, + 582, + 555, + 607 + ], + "type": "text", + "content": " is the pansharpening guidance estimated from the condition, defined as" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 376, + 617, + 555, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 617, + 555, + 632 + ], + "spans": [ + { + "bbox": [ + 376, + 617, + 555, + 632 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathcal {Y}} = (\\mathcal {Y}, \\mathbf {P}, \\mathcal {M}, f _ {\\psi} (\\mathcal {Y}, \\mathbf {P})), \\tag {6}", + "image_path": "82cc67fd22bde7f527072dedb1e3146e48483bcc283099dc045c796b9f5d7725.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 641, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 554, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 641, + 554, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 313, + 641, + 554, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 641, + 554, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 313, + 641, + 554, + 689 + ], + "type": "text", + "content": " are input image pair LR-HSI and PAN, " + }, + { + "bbox": [ + 313, + 641, + 554, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 641, + 554, + 689 + ], + "type": "text", + "content": " is a mapping from RGB detail image to HR-HSI (deferred in next subsection), and " + }, + { + "bbox": [ + 313, + 641, + 554, + 689 + ], + "type": "inline_equation", + "content": "f_{\\psi}(\\mathcal{V},\\mathbf{P})" + }, + { + "bbox": [ + 313, + 641, + 554, + 689 + ], + "type": "text", + "content": " is an auxiliary neural network that is trained with " + }, + { + "bbox": [ + 313, + 641, + 554, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 313, + 641, + 554, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 641, + 554, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 313, + 641, + 554, + 689 + ], + "type": "text", + "content": ". Following [9], we have" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 339, + 699, + 555, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 699, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 339, + 699, + 555, + 714 + ], + "type": "interline_equation", + "content": "p _ {\\theta , \\psi} \\left(\\mathcal {A} _ {t - 1} \\mid \\mathcal {A} _ {t}, \\tilde {\\mathcal {Y}}\\right) \\propto p _ {\\theta} \\left(\\mathcal {A} _ {t - 1} \\mid \\mathcal {A} _ {t}\\right) p _ {\\psi} (\\tilde {\\mathcal {Y}} \\mid \\mathcal {A} _ {t}). \\tag {7}", + "image_path": "b84695d9932f0f4135f17994944d0d7d890e67ec1ddecdf16a8aaa3c10b3faa9.jpg" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12671" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 107, + 71, + 504, + 235 + ], + "blocks": [ + { + "bbox": [ + 107, + 71, + 504, + 235 + ], + "lines": [ + { + "bbox": [ + 107, + 71, + 504, + 235 + ], + "spans": [ + { + "bbox": [ + 107, + 71, + 504, + 235 + ], + "type": "image", + "image_path": "fcc997ad0e0834c3612cc233579d7b350ce28c7a39fdf0d1cfd930cd8615fb14.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 243, + 555, + 289 + ], + "lines": [ + { + "bbox": [ + 54, + 243, + 555, + 289 + ], + "spans": [ + { + "bbox": [ + 54, + 243, + 555, + 289 + ], + "type": "text", + "content": "Figure 2. Illustration of the whole process of the proposed hyperspectral pansharpening method that includes diffusion with zero-shot (ZS) guidance and neural spatial-spectral decomposition (NSSD). The part in the dashed box is the proposed zero-shot guided function that is trained with a single testing image pair. NSSD builds the spectral mapping from the detail image generated by DM to HR-HSI. The guidance of the DM and NSSD are iteratively updated." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 81, + 311, + 272, + 418 + ], + "blocks": [ + { + "bbox": [ + 81, + 311, + 272, + 418 + ], + "lines": [ + { + "bbox": [ + 81, + 311, + 272, + 418 + ], + "spans": [ + { + "bbox": [ + 81, + 311, + 272, + 418 + ], + "type": "image", + "image_path": "93f32f0d57dbd6cf88c0f002b0fd4207fc0e1b02cf9311977ef3ddb57173ae1c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 426, + 295, + 483 + ], + "lines": [ + { + "bbox": [ + 55, + 426, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 295, + 483 + ], + "type": "text", + "content": "Figure 3. The analysis of the proposed zero-shot guidance. The DMs with the proposed guidance can achieve better performance with fewer sampling steps. (The length of the yellow box means the distance of PSNR values between the DM with and without ZS guidance)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 506, + 295, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 506, + 295, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 295, + 534 + ], + "type": "text", + "content": "Based on [6], we define the variable " + }, + { + "bbox": [ + 55, + 506, + 295, + 534 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{A}}_0\\coloneqq \\mathbb{E}[\\mathcal{A}_0|\\mathcal{A}_t] =" + }, + { + "bbox": [ + 55, + 506, + 295, + 534 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_{\\mathcal{A}_0\\sim p(\\mathcal{A}_0|\\mathcal{A}_t)}[\\mathcal{A}_0]" + }, + { + "bbox": [ + 55, + 506, + 295, + 534 + ], + "type": "text", + "content": " , which means" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 97, + 545, + 295, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 545, + 295, + 572 + ], + "spans": [ + { + "bbox": [ + 97, + 545, + 295, + 572 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {A}} _ {0} \\simeq \\frac {1}{\\sqrt {\\bar {\\alpha} _ {t}}} \\left(\\mathcal {A} _ {t} + \\left(1 - \\bar {\\alpha} _ {t}\\right) s _ {\\theta} \\left(\\mathcal {A} _ {t}, t\\right)\\right), \\tag {8}", + "image_path": "7f1b168e60c9d174c54f6177b6d8a38906a884ca0b04c2f9ba46dcb4407ad865.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 124, + 590, + 295, + 607 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 590, + 295, + 607 + ], + "spans": [ + { + "bbox": [ + 124, + 590, + 295, + 607 + ], + "type": "interline_equation", + "content": "p _ {\\psi} (\\tilde {\\mathcal {V}} | \\mathcal {A} _ {t}) \\simeq p _ {\\psi} (\\tilde {\\mathcal {V}} | \\hat {\\mathcal {A}} _ {0}), \\tag {9}", + "image_path": "0ccdfebccbe9c4059ece7d651d1d6f6f7c8749860f49fa891e306a2be829fcce.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 618, + 295, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 643 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 643 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 618, + 295, + 643 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t\\coloneqq \\prod_{s = 1}^t\\alpha_s" + }, + { + "bbox": [ + 55, + 618, + 295, + 643 + ], + "type": "text", + "content": " is the noise scale of DMs. We propose the zero-shot guidance as follows," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 656, + 299, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 656, + 299, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 656, + 299, + 714 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\log p _ {\\psi} (\\tilde {\\mathcal {Y}} | \\hat {\\mathcal {A}} _ {0}) = \\left\\| \\tau (\\mathcal {M} (\\hat {\\mathcal {A}} _ {0})) - \\mathcal {Y} \\right\\| _ {F} ^ {2} \\\\ + \\left\\| \\mathcal {M} \\left(\\hat {\\mathcal {A}} _ {0}\\right) \\times_ {3} \\mathbf {R} - \\mathbf {P} \\right\\| _ {F} ^ {2} + \\lambda_ {1} \\left\\| \\mathcal {M} \\left(\\hat {\\mathcal {A}} _ {0}\\right) - f _ {\\psi} (\\mathcal {Y}, \\mathbf {P}) \\right\\| _ {F} ^ {2}, \\tag {10} \\\\ \\end{array}", + "image_path": "1d33ce6efabb0089177e5cfe12be51e4f76d179f70a968760c397cdd79cf5757.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 309, + 553, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 309, + 553, + 332 + ], + "spans": [ + { + "bbox": [ + 313, + 309, + 553, + 332 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 309, + 553, + 332 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 313, + 309, + 553, + 332 + ], + "type": "text", + "content": " is a weighting scale. Finally, the reverse process is illustrated as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 320, + 338, + 553, + 393 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 338, + 553, + 393 + ], + "spans": [ + { + "bbox": [ + 320, + 338, + 553, + 393 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {A} _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\mathcal {A} _ {t} - \\frac {\\beta_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} s _ {\\theta} \\left(\\mathcal {A} _ {t}, t\\right)\\right) \\tag {11} \\\\ - \\lambda \\cdot \\frac {\\beta_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}} \\sqrt {\\alpha_ {t}}} \\nabla_ {\\mathcal {A} _ {t}} l o g p _ {\\psi} (\\tilde {\\mathcal {Y}} | \\hat {\\mathcal {A}} _ {0}), \\\\ \\end{array}", + "image_path": "babcc53b8094ba4199624033c6210cf6f83f60893e8552405715a916b9ed2cec.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 402, + 553, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 553, + 437 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 402, + 553, + 437 + ], + "type": "inline_equation", + "content": "\\beta_{t}" + }, + { + "bbox": [ + 313, + 402, + 553, + 437 + ], + "type": "text", + "content": ", as a hyperparameter, is generally set to linearly interpolate from 0.0001 to 0.02 [28] and " + }, + { + "bbox": [ + 313, + 402, + 553, + 437 + ], + "type": "inline_equation", + "content": "\\alpha_{t} := 1 - \\beta_{t}" + }, + { + "bbox": [ + 313, + 402, + 553, + 437 + ], + "type": "text", + "content": ". The weighting scale " + }, + { + "bbox": [ + 313, + 402, + 553, + 437 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 402, + 553, + 437 + ], + "type": "text", + "content": " can balance the effect of guidance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 437, + 555, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 555, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 555, + 629 + ], + "type": "text", + "content": "The rest of this part is how to train the zero-shot network " + }, + { + "bbox": [ + 313, + 437, + 555, + 629 + ], + "type": "inline_equation", + "content": "f_{\\psi}(\\mathcal{V}, \\mathbf{P})" + }, + { + "bbox": [ + 313, + 437, + 555, + 629 + ], + "type": "text", + "content": ". The network structure and training process are shown within the dotted box in Figure 2, which has been verified its effectiveness in [11]. Based on the \"invariance among scales\" assumption, we train the network in the reduced-resolution scale. In specific, we use the input LR-HSI as the training ground-truth. Training PAN and training LR-HSI are simulated by PAN and LR-HSI after spatial and spectral degenerate processes, respectively. This approach allows for efficient training of the zero-shot network using only a single testing input pair. Figure 3 shows the change curve of PSNR values with the increase of iteration number. One can observe that the DM guided by zero-shot guidance can achieve better performance with fewer sampling steps. Besides, this single-image training property provides specific prior knowledge to the pre-trained DM." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 635, + 520, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 520, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 520, + 649 + ], + "type": "text", + "content": "3.2. Neural Spatial-Spectral Decomposition" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": "This subsection describes neural spatial-spectral decomposition (NSSD), i.e., " + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": " in Eq. (10). In hyperspectral pan-sharpening, beyond the fidelity term in (1), the spatial-spectral correlation exhibits the low-rank properties characteristic of hyperspectral data [44]. The mode-3 decom" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12672" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 70, + 479, + 200 + ], + "blocks": [ + { + "bbox": [ + 132, + 70, + 479, + 200 + ], + "lines": [ + { + "bbox": [ + 132, + 70, + 479, + 200 + ], + "spans": [ + { + "bbox": [ + 132, + 70, + 479, + 200 + ], + "type": "image", + "image_path": "ac02864a374f0b054b5152395b5a575ed2d6a20fcd90661d03ec33e8054a09b3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 208, + 555, + 232 + ], + "lines": [ + { + "bbox": [ + 54, + 208, + 555, + 232 + ], + "spans": [ + { + "bbox": [ + 54, + 208, + 555, + 232 + ], + "type": "text", + "content": "Figure 4. The structure of neural spatial-spectral decomposition (NSSD) consists of a U-net structure with skip connection and the fully connection structure to extract spatial and spectral information, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 251, + 294, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 294, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 294, + 277 + ], + "type": "text", + "content": "position format, i.e., " + }, + { + "bbox": [ + 55, + 251, + 294, + 277 + ], + "type": "inline_equation", + "content": "\\mathcal{X} = \\mathcal{A}\\times_3\\mathbf{E}" + }, + { + "bbox": [ + 55, + 251, + 294, + 277 + ], + "type": "text", + "content": ", has shown to be effective [19, 20, 31], which satisfies the following inequation:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 71, + 285, + 295, + 300 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 285, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 71, + 285, + 295, + 300 + ], + "type": "interline_equation", + "content": "\\operatorname {r a n k} \\left(\\mathbf {X} _ {(3)}\\right) \\leq \\min \\left\\{\\operatorname {r a n k} \\left(\\mathbf {A} _ {(3)}\\right), \\operatorname {r a n k} (\\mathbf {E}) \\right\\} \\leq b, \\tag {12}", + "image_path": "8686959d7aed739b0a072081dd6c6faef50ca134173d7ee1572c6c7fee29c4d1.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 306, + 296, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 306, + 296, + 378 + ], + "spans": [ + { + "bbox": [ + 55, + 306, + 296, + 378 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 306, + 296, + 378 + ], + "type": "inline_equation", + "content": "\\mathcal{A} \\in \\mathbb{R}^{H \\times W \\times b}" + }, + { + "bbox": [ + 55, + 306, + 296, + 378 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 306, + 296, + 378 + ], + "type": "inline_equation", + "content": "\\mathbf{E} \\in \\mathbb{R}^{S \\times b}" + }, + { + "bbox": [ + 55, + 306, + 296, + 378 + ], + "type": "text", + "content": ". The low value of the size of " + }, + { + "bbox": [ + 55, + 306, + 296, + 378 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 55, + 306, + 296, + 378 + ], + "type": "text", + "content": " limits the expressive ability. Since pre-trained DMs are primarily trained on RGB images, the size of " + }, + { + "bbox": [ + 55, + 306, + 296, + 378 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 55, + 306, + 296, + 378 + ], + "type": "text", + "content": " is typically set to " + }, + { + "bbox": [ + 55, + 306, + 296, + 378 + ], + "type": "inline_equation", + "content": "H \\times W \\times 3" + }, + { + "bbox": [ + 55, + 306, + 296, + 378 + ], + "type": "text", + "content": ". Additionally, the larger value of " + }, + { + "bbox": [ + 55, + 306, + 296, + 378 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 306, + 296, + 378 + ], + "type": "text", + "content": " in (12) inevitably increases the time required for the reverse process in the DM." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 380, + 296, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 380, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 380, + 296, + 403 + ], + "type": "text", + "content": "To mitigate the drawbacks, we propose a NSSD to capture the spatial-spectral correlation, as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 100, + 413, + 295, + 428 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 413, + 295, + 428 + ], + "spans": [ + { + "bbox": [ + 100, + 413, + 295, + 428 + ], + "type": "interline_equation", + "content": "\\mathcal {X} = \\mathcal {M} (\\mathcal {A}) := \\mathcal {C} _ {\\theta_ {1}} (\\mathcal {A}) \\times_ {3} \\mathbf {D} _ {\\theta_ {2}} (\\mathcal {Y}), \\tag {13}", + "image_path": "2c45de6fd175ffbf93930b9bcf0129253a16360b5b496df1480beb0c293c09ce.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 436, + 296, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 436, + 296, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 296, + 581 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 436, + 296, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{\\theta_1}(\\mathcal{A}) \\in \\mathbb{R}^{H \\times W \\times r}" + }, + { + "bbox": [ + 55, + 436, + 296, + 581 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 436, + 296, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{\\theta_2}(\\mathcal{Y}) \\in \\mathbb{R}^{S \\times r}" + }, + { + "bbox": [ + 55, + 436, + 296, + 581 + ], + "type": "text", + "content": " with hyperparameter " + }, + { + "bbox": [ + 55, + 436, + 296, + 581 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 436, + 296, + 581 + ], + "type": "text", + "content": " are produced by two neural network with the input " + }, + { + "bbox": [ + 55, + 436, + 296, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{A} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 55, + 436, + 296, + 581 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 436, + 296, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{Y} \\in \\mathbb{R}^{h \\times w \\times S}" + }, + { + "bbox": [ + 55, + 436, + 296, + 581 + ], + "type": "text", + "content": ", respectively, as illustrated in Figure 4. Specifically, " + }, + { + "bbox": [ + 55, + 436, + 296, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{\\theta_1}(\\cdot)" + }, + { + "bbox": [ + 55, + 436, + 296, + 581 + ], + "type": "text", + "content": " is the spatial-factor network that extracts the spatial information from the RGB detail image. The structure of " + }, + { + "bbox": [ + 55, + 436, + 296, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{\\theta_1}(\\cdot)" + }, + { + "bbox": [ + 55, + 436, + 296, + 581 + ], + "type": "text", + "content": " is the U-net type with skip connection, whose detail-extraction capability has been verified [35]. The spectral-factor network, " + }, + { + "bbox": [ + 55, + 436, + 296, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{\\theta_2}(\\cdot)" + }, + { + "bbox": [ + 55, + 436, + 296, + 581 + ], + "type": "text", + "content": ", consists of the fully connection structure and sigmoid to efficiently absorb spectral information from LR-HSI. NSSD does not require pre-training. It produces the final result through iterative steps using the following loss:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 64, + 589, + 295, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 589, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 64, + 589, + 295, + 622 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L o s s = \\left\\| \\tau \\left(\\mathcal {C} _ {\\theta_ {1}} \\times_ {3} \\mathbf {D} _ {\\theta_ {2}}\\right) - \\mathcal {Y} \\right\\| _ {F} ^ {2} \\tag {14} \\\\ + \\lambda_ {2} \\| \\left(\\mathcal {C} _ {\\theta_ {1}} \\times_ {3} \\mathbf {D} _ {\\theta_ {2}}\\right) \\times_ {3} \\mathbf {R} - \\mathbf {P} \\| _ {F} ^ {2}, \\\\ \\end{array}", + "image_path": "625c3db3fbcfaa90015a07e4321d9c4443e901c9973cf2bb61a3117fe4e39264.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "text", + "content": " is a balanced parameter. NSSD models the relationship between the generated detail image and the target HR-HSI. Compared with the classic format " + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{X} = \\mathcal{A} \\times_3 \\mathbf{E}" + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "text", + "content": ", the proposed form " + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{\\theta_1}(\\mathcal{A}) \\times_3 \\mathbf{D}_{\\theta_2}(\\mathcal{Y})" + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "text", + "content": " offers a more flexible representation for HR-HSI, particularly for data with higher rank. Furthermore, the NSSD evolves in the guidance of pre-trained DMs, as shown in Eq. (10). The guidance of the" + } + ] + } + ], + "index": 9 + }, + { + "type": "code", + "bbox": [ + 322, + 268, + 535, + 430 + ], + "blocks": [ + { + "bbox": [ + 321, + 255, + 493, + 266 + ], + "lines": [ + { + "bbox": [ + 321, + 255, + 493, + 266 + ], + "spans": [ + { + "bbox": [ + 321, + 255, + 493, + 266 + ], + "type": "text", + "content": "Algorithm 1: The Whole Solving Process" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 322, + 268, + 535, + 430 + ], + "lines": [ + { + "bbox": [ + 322, + 268, + 535, + 430 + ], + "spans": [ + { + "bbox": [ + 322, + 268, + 535, + 430 + ], + "type": "text", + "content": "Input: " + }, + { + "bbox": [ + 322, + 268, + 535, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_T" + }, + { + "bbox": [ + 322, + 268, + 535, + 430 + ], + "type": "text", + "content": " sampled from " + }, + { + "bbox": [ + 322, + 268, + 535, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0,I)" + }, + { + "bbox": [ + 322, + 268, + 535, + 430 + ], + "type": "text", + "content": " ,DM " + }, + { + "bbox": [ + 322, + 268, + 535, + 430 + ], + "type": "inline_equation", + "content": "s_\\theta" + }, + { + "bbox": [ + 322, + 268, + 535, + 430 + ], + "type": "text", + "content": " noise scales " + }, + { + "bbox": [ + 322, + 268, + 535, + 430 + ], + "type": "inline_equation", + "content": "\\{\\bar{\\alpha}_t\\}_{t = 1}^T" + }, + { + "bbox": [ + 322, + 268, + 535, + 430 + ], + "type": "inline_equation", + "content": "f_{\\psi}" + }, + { + "bbox": [ + 322, + 268, + 535, + 430 + ], + "type": "text", + "content": " ,LR-HSI Y,PAN P \nOutput: HR-HSI X \nwhile not converged do for " + }, + { + "bbox": [ + 322, + 268, + 535, + 430 + ], + "type": "inline_equation", + "content": "t = T,T - 1,\\dots 1" + }, + { + "bbox": [ + 322, + 268, + 535, + 430 + ], + "type": "text", + "content": " do step 1: estimate " + }, + { + "bbox": [ + 322, + 268, + 535, + 430 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{A}}_0" + }, + { + "bbox": [ + 322, + 268, + 535, + 430 + ], + "type": "text", + "content": " by Eq. (8) step 2: compute log " + }, + { + "bbox": [ + 322, + 268, + 535, + 430 + ], + "type": "inline_equation", + "content": "p_{\\psi}(\\mathcal{V}|\\mathcal{A}_t)" + }, + { + "bbox": [ + 322, + 268, + 535, + 430 + ], + "type": "text", + "content": " by Eq. (10) step 3: sample " + }, + { + "bbox": [ + 322, + 268, + 535, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{t - 1}" + }, + { + "bbox": [ + 322, + 268, + 535, + 430 + ], + "type": "text", + "content": " by Eq. (11) end NSSD Update: update " + }, + { + "bbox": [ + 322, + 268, + 535, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 322, + 268, + 535, + 430 + ], + "type": "text", + "content": " by minimizing (14) \nend " + }, + { + "bbox": [ + 322, + 268, + 535, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{X} = \\mathcal{C}_{\\theta_1}(\\mathcal{A})\\times_3\\mathbf{D}_{\\theta_2}(\\mathcal{Y})" + }, + { + "bbox": [ + 322, + 268, + 535, + 430 + ], + "type": "text", + "content": " \nreturn X" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "code_body" + } + ], + "index": 11, + "sub_type": "algorithm" + }, + { + "bbox": [ + 313, + 453, + 553, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 553, + 490 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 553, + 490 + ], + "type": "text", + "content": "DMs and the update of NSSD are mutually adaptive. Thus, NSSD provides a more flexible and DM-compatible representation of the spatial-spectral correlation in HR-HSI." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 491, + 555, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 555, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 555, + 586 + ], + "type": "text", + "content": "Whole Solving Algorithm: Since the guidance of pretrained DM is related to the mapping " + }, + { + "bbox": [ + 313, + 491, + 555, + 586 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 491, + 555, + 586 + ], + "type": "text", + "content": ", the generated detail image is adaptive to the mapping " + }, + { + "bbox": [ + 313, + 491, + 555, + 586 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 491, + 555, + 586 + ], + "type": "text", + "content": ". To enhance the flexibility of mapping and the suitability of diffusion guidance, we alternatively conduct guidance of pre-trained DM and NSSD update as shown in Figure 2. During each guidance step, the parameters of the NSSD are held fixed. The whole solving algorithm is summarized in Algorithm 1." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 597, + 395, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 597, + 395, + 611 + ], + "spans": [ + { + "bbox": [ + 313, + 597, + 395, + 611 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "content": "To demonstrate the effectiveness of the proposed method, we compare it with corresponding state-of-the-art methods, including GSA [1], CNMF [47], HySure [32], ZSL [11], PLRDiff [30], and HIR-Diff [29]. It is worth to be remarked that HIR-Diff [29] is a general inverse problem framework. We rewrite the fidelity for the hyperspectral pansharpening task. The parameters of all methods are adjusted according to the authors' recommendation. The metrics peak" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12673" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 107, + 73, + 504, + 175 + ], + "blocks": [ + { + "bbox": [ + 107, + 73, + 504, + 175 + ], + "lines": [ + { + "bbox": [ + 107, + 73, + 504, + 175 + ], + "spans": [ + { + "bbox": [ + 107, + 73, + 504, + 175 + ], + "type": "image", + "image_path": "f7f2fd480b50fd617a25808e4186208f0112c88d0c4b54179e9188276ae04a01.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 184, + 555, + 208 + ], + "lines": [ + { + "bbox": [ + 54, + 184, + 555, + 208 + ], + "spans": [ + { + "bbox": [ + 54, + 184, + 555, + 208 + ], + "type": "text", + "content": "Figure 5. Visual comparisons on Pavia dataset. The first row shows the predicted HR-HSIs of different methods, and the second row displays the error maps between predicted HR-HSIs and ground-truth." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 228, + 295, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 228, + 295, + 383 + ], + "spans": [ + { + "bbox": [ + 54, + 228, + 295, + 383 + ], + "type": "text", + "content": "signal-to-noise ratio (PSNR), the structural similarity index (SSIM) [38], the relative dimensionless global error in synthesis (ERGAS) [2], the spectral angle mapper (SAM) [49], and root mean square error (RMSE) are utilized for assessment. QNR [3] is employed for evaluating real-world data. We assess the results produced by these methods on Pavia1, Washington DC (WDC)2, Chikusei3, and FR1 datasets4. The simulated LR-HSI is gained by applying a " + }, + { + "bbox": [ + 54, + 228, + 295, + 383 + ], + "type": "inline_equation", + "content": "5 \\times 5" + }, + { + "bbox": [ + 54, + 228, + 295, + 383 + ], + "type": "text", + "content": " Gaussian blurring kernel and the downsampling factor 4. The simulated PAN image is obtained from groundtruth by multiplying a spectral mean response vector. The downsampling factor of FR1 dataset is 6. The detailed information about these datasets is listed as follows," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 384, + 295, + 574 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 384, + 295, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 295, + 431 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 295, + 431 + ], + "type": "text", + "content": "- Pavia dataset: Pavia dataset are originally obtained with the size of " + }, + { + "bbox": [ + 55, + 384, + 295, + 431 + ], + "type": "inline_equation", + "content": "610 \\times 340 \\times 115" + }, + { + "bbox": [ + 55, + 384, + 295, + 431 + ], + "type": "text", + "content": ". We choose 93 bands by eliminating the bands associated with water vapor absorption and crop the up-left cube with size " + }, + { + "bbox": [ + 55, + 384, + 295, + 431 + ], + "type": "inline_equation", + "content": "256 \\times 256 \\times 93" + }, + { + "bbox": [ + 55, + 384, + 295, + 431 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 432, + 295, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 432, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 432, + 295, + 479 + ], + "type": "text", + "content": "- WDC dataset: WDC dataset is obtained from the Hyperspectral Digital Imagery Collection Experiment (HYDICE) sensor. We eliminate some unusable bands and choose the area with size " + }, + { + "bbox": [ + 55, + 432, + 295, + 479 + ], + "type": "inline_equation", + "content": "256 \\times 256 \\times 191" + }, + { + "bbox": [ + 55, + 432, + 295, + 479 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 479, + 295, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 295, + 526 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 295, + 526 + ], + "type": "text", + "content": "- Chikusei dataset: The Headwall Hyperspec-VNIR-C imaging sensor captures the Chikusei dataset with 128 bands and the spectral range is from 363 to " + }, + { + "bbox": [ + 55, + 479, + 295, + 526 + ], + "type": "inline_equation", + "content": "1018\\mathrm{~nm}" + }, + { + "bbox": [ + 55, + 479, + 295, + 526 + ], + "type": "text", + "content": ". We crop a subimage with size " + }, + { + "bbox": [ + 55, + 479, + 295, + 526 + ], + "type": "inline_equation", + "content": "256 \\times 256 \\times 128" + }, + { + "bbox": [ + 55, + 479, + 295, + 526 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 527, + 295, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 527, + 295, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 527, + 295, + 574 + ], + "type": "text", + "content": "- FR1 real dataset: The FR1 dataset is a real-world data distributed for the PRISMA contest, for pansharpening at the full spatial resolution. We crop a subimage with PAN size " + }, + { + "bbox": [ + 55, + 527, + 295, + 574 + ], + "type": "inline_equation", + "content": "240 \\times 240" + }, + { + "bbox": [ + 55, + 527, + 295, + 574 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 582, + 188, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 188, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 188, + 594 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 599, + 295, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 599, + 295, + 647 + ], + "spans": [ + { + "bbox": [ + 54, + 599, + 295, + 647 + ], + "type": "text", + "content": "All experiments are conducted on the same PC with Pytorch 2.0 and MATLAB (R2022a) on a computer of 64Gb RAM, Intel(R) Core(TM) i9-10900KF CPU @3.70GHz, and NVIDIA GeForce RTX 4070. The update of NSSD" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 654, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 654, + 188, + 663 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 654, + 188, + 663 + ], + "spans": [ + { + "bbox": [ + 48, + 654, + 188, + 663 + ], + "type": "text", + "content": "1 https://rslab.ut.ac.ir/data" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 664, + 294, + 684 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 664, + 294, + 684 + ], + "spans": [ + { + "bbox": [ + 48, + 664, + 294, + 684 + ], + "type": "text", + "content": "2 https://engineering.purdue.edu/biehl/MultiSpec/hyperspectral.html" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 684, + 235, + 693 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 684, + 235, + 693 + ], + "spans": [ + { + "bbox": [ + 48, + 684, + 235, + 693 + ], + "type": "text", + "content": "3 https://naotoyokoya.com/Download.html" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 693, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 693, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 693, + 294, + 713 + ], + "type": "text", + "content": "4 https : / / github . com / liangjiandeng / HyperPanCollection" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 313, + 228, + 555, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 555, + 347 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 555, + 347 + ], + "type": "text", + "content": "is optimized by Adam optimizer with 0.001 learning rate (LR). The balanced parameter of guidance " + }, + { + "bbox": [ + 313, + 228, + 555, + 347 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 313, + 228, + 555, + 347 + ], + "type": "text", + "content": " is 0.1, and " + }, + { + "bbox": [ + 313, + 228, + 555, + 347 + ], + "type": "inline_equation", + "content": "\\lambda_{2} = 1" + }, + { + "bbox": [ + 313, + 228, + 555, + 347 + ], + "type": "text", + "content": ". The diffusion reverse process step number is set to 100. The number of the whole iterations is 20000. The training of the zero-shot neural network is with Adam optimizer with 2000 iterations, and the LR is 0.001 decayed by gamma " + }, + { + "bbox": [ + 313, + 228, + 555, + 347 + ], + "type": "inline_equation", + "content": "= 0.8" + }, + { + "bbox": [ + 313, + 228, + 555, + 347 + ], + "type": "text", + "content": " each 1000 iterations. The rank of NSSD for Pavia, WDC, and Chikusei datasets are set to 20, 40, and 40, respectively. For FR1 real dataset, " + }, + { + "bbox": [ + 313, + 228, + 555, + 347 + ], + "type": "inline_equation", + "content": "\\lambda_{1} = \\lambda_{2} = 10" + }, + { + "bbox": [ + 313, + 228, + 555, + 347 + ], + "type": "text", + "content": ", and the number of the whole iterations is 10000." + } + ] + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 315, + 387, + 558, + 495 + ], + "blocks": [ + { + "bbox": [ + 313, + 356, + 554, + 378 + ], + "lines": [ + { + "bbox": [ + 313, + 356, + 554, + 378 + ], + "spans": [ + { + "bbox": [ + 313, + 356, + 554, + 378 + ], + "type": "text", + "content": "Table 1. Quantitative results for Pavia dataset. (Bold: best; Underline: second best)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 387, + 558, + 495 + ], + "lines": [ + { + "bbox": [ + 315, + 387, + 558, + 495 + ], + "spans": [ + { + "bbox": [ + 315, + 387, + 558, + 495 + ], + "type": "table", + "html": "
MethodPSNR↑SSIM↑ERGAS↓SAM↓RMSE↓
GSA [1]07*TGRS28.3040.8125.8197.75310.368
CNMF [47] 11*TGRS28.8960.8595.4955.9499.558
HySure [32] 14*TGRS25.8470.7587.86110.05914.065
ZSL [11]23'TPAMI32.0600.9053.5495.3186.819
PLRDiff [30] 24'IF33.0600.9053.2915.3346.001
HIR-Diff [29]24'CVPR31.9740.8753.6184.7056.742
Proposed34.2270.9262.8634.4635.284
Ideal value+∞1000
", + "image_path": "c8a10cf9d6459bfe580b45cbacfb3d01767069d249b69d204754670f92916ea0.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 516, + 452, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 516, + 452, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 516, + 452, + 528 + ], + "type": "text", + "content": "4.2. Results on Pavia Dataset" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 533, + 555, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 555, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 555, + 665 + ], + "type": "text", + "content": "In this part, we evaluate the effectiveness of the proposed method on the Pavia dataset. Figure 5 shows the visual comparison of these methods in the benchmark. Diffusion-based methods, i.e., PLRDiff [30], HIR-Diff [29], and the proposed method, achieve excellent performance for detail extraction. The residual images, i.e., the second row of Figure 5 obviously verify that our method can preserve more details compared to others. The result produced by HySure [32] leads to spatial distortions. Table 1 provides the quantitative results for compared approaches. For all metrics, the proposed method achieves the best performance." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 672, + 453, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 453, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 453, + 684 + ], + "type": "text", + "content": "4.3. Results on WDC Dataset" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "For the WDC dataset, there exist more challenges to preserve the spectral information due to its more spectral" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12674" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 107, + 72, + 504, + 175 + ], + "blocks": [ + { + "bbox": [ + 107, + 72, + 504, + 175 + ], + "lines": [ + { + "bbox": [ + 107, + 72, + 504, + 175 + ], + "spans": [ + { + "bbox": [ + 107, + 72, + 504, + 175 + ], + "type": "image", + "image_path": "3343eb19e5774ef750d27c553b0df5666be8552de592b57c9c57c96a60e79c72.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 184, + 555, + 207 + ], + "lines": [ + { + "bbox": [ + 55, + 184, + 555, + 207 + ], + "spans": [ + { + "bbox": [ + 55, + 184, + 555, + 207 + ], + "type": "text", + "content": "Figure 6. Visual comparisons on WDC dataset. The first row shows the predicted HR-HSIs of different methods, and the second row displays the error maps between predicted HR-HSIs and ground-truth." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 107, + 219, + 504, + 322 + ], + "blocks": [ + { + "bbox": [ + 107, + 219, + 504, + 322 + ], + "lines": [ + { + "bbox": [ + 107, + 219, + 504, + 322 + ], + "spans": [ + { + "bbox": [ + 107, + 219, + 504, + 322 + ], + "type": "image", + "image_path": "9fb2a19da9eb27ccbdf2baff1f5750f2bdf698789dd63b5822ce4babe14d7e45.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 331, + 555, + 353 + ], + "lines": [ + { + "bbox": [ + 55, + 331, + 555, + 353 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 555, + 353 + ], + "type": "text", + "content": "Figure 7. Visual comparisons on Chikusei dataset. The first row shows the predicted HR-HSIs of different methods, and the second row displays the error maps between predicted HR-HSIs and ground-truth." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 56, + 403, + 299, + 511 + ], + "blocks": [ + { + "bbox": [ + 55, + 372, + 295, + 394 + ], + "lines": [ + { + "bbox": [ + 55, + 372, + 295, + 394 + ], + "spans": [ + { + "bbox": [ + 55, + 372, + 295, + 394 + ], + "type": "text", + "content": "Table 2. Quantitative results for WDC dataset. (Bold: best; Underline: second best)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 403, + 299, + 511 + ], + "lines": [ + { + "bbox": [ + 56, + 403, + 299, + 511 + ], + "spans": [ + { + "bbox": [ + 56, + 403, + 299, + 511 + ], + "type": "table", + "html": "
MethodPSNR↑SSIM↑ERGAS↓SAM↓RMSE↓
GSA [1]07*TGRS39.6290.9105.9247.0797.665
CNMF [47] 11*TGRS39.7030.8975.8646.4487.665
HySure [32] 14*TGRS35.0710.7989.92611.89812.928
ZSL [11]23*TPAMI39.6850.9166.3357.8257.301
PLRDiff [30] 24*IF34.7990.83318.2029.5448.667
HIR-Diff [29]24*CVPR38.5060.8447.1026.6868.079
Proposed40.5030.94312.2545.0474.631
Ideal value+∞1000
", + "image_path": "845b79e5c9016d3e351d750813349c9df70aaf43526ac100f78768110c5aa74c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 56, + 557, + 299, + 666 + ], + "blocks": [ + { + "bbox": [ + 55, + 526, + 295, + 548 + ], + "lines": [ + { + "bbox": [ + 55, + 526, + 295, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 526, + 295, + 548 + ], + "type": "text", + "content": "Table 3. Quantitative results for Chikusei dataset. (Bold: best; Underline: second best)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 557, + 299, + 666 + ], + "lines": [ + { + "bbox": [ + 56, + 557, + 299, + 666 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 299, + 666 + ], + "type": "table", + "html": "
MethodPSNR↑SSIM↑ERGAS↓SAM↓RMSE↓
GSA [1]07*TGRS32.4590.88710.8886.7626.662
CNMF [47] 11*TGRS33.5780.8788.5924.8935.905
HySure [32] 14*TGRS30.2700.75214.1599.8838.417
ZSL [11]23'TPAMI35.2580.9178.3675.6854.632
PLRDiff [30] 24'IF37.1760.9176.9944.8383.651
HIR-Diff [29]24'CVPR36.3890.9056.3333.8944.247
Proposed38.4500.9395.9454.1973.192
Ideal value+∞1000
", + "image_path": "2f8e8dc17a7e342c2a8e0b85751970a39d3dd9a6264de5e5d7fea5e944d9a402.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "bands. As shown in Table 2, we provide the visual results of different methods. Our approach also performs best on" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 373, + 555, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 555, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 555, + 445 + ], + "type": "text", + "content": "this dataset. Compared with traditional method CNMF [47], diffusion-based methods PLRDiff [30] and HIR-Diff [29] lead to spectral distortions. As demonstrated in Figure 6, PLRDiff [30] also preserves excellent spatial detail. The proposed method has the advantages of both spatial detail preservation and spectral information extraction." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 454, + 468, + 465 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 454, + 468, + 465 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 468, + 465 + ], + "type": "text", + "content": "4.4. Results on Chikusei Dataset" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 472, + 555, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 555, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 555, + 568 + ], + "type": "text", + "content": "Figure 7 evaluates the effectiveness of the proposed method on the Chikusei dataset. Our method achieves the best result due to the spatial detail preservation. For the detail edge of the image, the proposed method can preserve the spatial detail well. As shown in Table 3, the proposed method demonstrates improvements in PSNR by " + }, + { + "bbox": [ + 313, + 472, + 555, + 568 + ], + "type": "inline_equation", + "content": "1.274\\mathrm{dB} / 2.061\\mathrm{dB}" + }, + { + "bbox": [ + 313, + 472, + 555, + 568 + ], + "type": "text", + "content": ", respectively, compared with diffusion-based methods PLRDiff [30] and HIR-Diff [29]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 575, + 408, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 575, + 408, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 575, + 408, + 588 + ], + "type": "text", + "content": "4.5. Ablation Study" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 594, + 555, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 700 + ], + "type": "text", + "content": "To demonstrate the effectiveness of the proposed zero-shot guidance and NSSD, we conduct the ablation study on the Pavia dataset. Table 4 shows the results produced by Four models. For the model I, we use the fixed transformation in [30] and regular fidelity to substitute the proposed zero-shot guidance and NSSD, respectively. As shown in Table 4, the results produced by the model with both zero-shot guidance and NSSD are best. Besides, we design more detailed experiments to explore more properties of the two parts." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 326, + 701, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 554, + 713 + ], + "type": "text", + "content": "Zero-shot Guidance: The zero-shot guidance achieves" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12675" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 102, + 294, + 162 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 295, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 295, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 295, + 94 + ], + "type": "text", + "content": "Table 4. Ablation study about the proposed zero-shot (ZS) guidance and neural spatial-spectral decoder (NSSD) on Pavia dataset." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 102, + 294, + 162 + ], + "lines": [ + { + "bbox": [ + 56, + 102, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 102, + 294, + 162 + ], + "type": "table", + "html": "
AblationZSNSSDPSNRSSIMERGASSAMRMSE
Model I✗✗33.0600.9053.2915.3346.001
Model II✓✗33.2670.9123.2195.2855.877
Model III✗✓33.9720.9212.9464.6055.460
Proposed✓✓34.1790.9252.8774.5335.332
", + "image_path": "779bee8a44ea5cb7c71cdd0cecb10963fc66a3e51134ead4340ec0cfe3e1d7f3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 116, + 171, + 236, + 266 + ], + "blocks": [ + { + "bbox": [ + 116, + 171, + 236, + 266 + ], + "lines": [ + { + "bbox": [ + 116, + 171, + 236, + 266 + ], + "spans": [ + { + "bbox": [ + 116, + 171, + 236, + 266 + ], + "type": "image", + "image_path": "7afca9c5e68c05a388f3c312d7e8a110f9e6a28c4987f46857b19e2fa81d142a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 273, + 295, + 297 + ], + "lines": [ + { + "bbox": [ + 55, + 273, + 295, + 297 + ], + "spans": [ + { + "bbox": [ + 55, + 273, + 295, + 297 + ], + "type": "text", + "content": "Figure 8. Singular value distributions of results produced by PLRDiff [30], HIR-Diff [29], and the proposed method." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 78, + 338, + 272, + 394 + ], + "blocks": [ + { + "bbox": [ + 55, + 307, + 295, + 331 + ], + "lines": [ + { + "bbox": [ + 55, + 307, + 295, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 307, + 295, + 331 + ], + "type": "text", + "content": "Table 5. Performance of proposed zero-shot guidance compared with conventional regularization term." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 78, + 338, + 272, + 394 + ], + "lines": [ + { + "bbox": [ + 78, + 338, + 272, + 394 + ], + "spans": [ + { + "bbox": [ + 78, + 338, + 272, + 394 + ], + "type": "table", + "html": "
GuidancePSNRSSIMERGASSAMRMSE
Fidelity32.4700.8963.4555.1146.395
TV32.7700.9023.3565.1566.204
Proposed33.3740.9143.1745.1505.793
", + "image_path": "29b444f201639db0de3bb2ce109cbe60e2201cdcf377958214cba38ab6309a1a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 413, + 295, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 413, + 295, + 496 + ], + "spans": [ + { + "bbox": [ + 54, + 413, + 295, + 496 + ], + "type": "text", + "content": "excellent guided performance. To demonstrate the effectiveness, we compare it with the conventional total variation regularized term, which is widely used and effective for image processing tasks. As shown in Table 5, the regularized term is helpful for the guidance of DMs. The proposed zero-shot guidance offers more specific prior knowledge for the pre-trained DM to achieve the best performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 497, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 497, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 54, + 497, + 295, + 617 + ], + "type": "text", + "content": "NSSD: NSSD describes the spatial-spectral correlation with the low-rank property. Figure 8 shows the singular value distribution of target HR-HSI produced by different methods. We compare the proposed method with NSSD with two fixed Diffusion transformation estimation methods in PLRDiff [30] and HIR-Diff [29]. One can observe that the proposed NSSD provides a flexible transformation and performs the best approximation with the pre-trained DM guidance. Our method can achieve the closest distance between the singular values of ground-truth." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 624, + 157, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 157, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 157, + 635 + ], + "type": "text", + "content": "4.6. Guided Network" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 641, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 295, + 713 + ], + "type": "text", + "content": "In this part, we analyze the influence of zero-shot network structure for guided performance. Table 6 shows the final results with U-Net with skip-connections [35], DSNet [51], and CONet [11]. U-Net contains multi-scale information, DSNet is a structure used in the supervised method, and CONet is a light convolution-based cascade structure." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 318, + 102, + 549, + 158 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 6. Performance of zero-shot guided network with different network structures." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 102, + 549, + 158 + ], + "lines": [ + { + "bbox": [ + 318, + 102, + 549, + 158 + ], + "spans": [ + { + "bbox": [ + 318, + 102, + 549, + 158 + ], + "type": "table", + "html": "
Guided StructurePSNRSSIMERGASSAMRMSE
U-Net34.0150.9212.9534.5115.426
DSNet33.7750.9173.0334.6775.609
CONet34.2270.9262.8634.4635.284
", + "image_path": "65365336191c483b322dd090147a2f501f5ad7383a5323a525f0dae189cb4a26.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 177, + 555, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 177, + 555, + 213 + ], + "spans": [ + { + "bbox": [ + 313, + 177, + 555, + 213 + ], + "type": "text", + "content": "The proposed method with the above network structures can achieve excellent results. The convolution-based cascade structure achieves the best performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 220, + 419, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 220, + 419, + 232 + ], + "spans": [ + { + "bbox": [ + 313, + 220, + 419, + 232 + ], + "type": "text", + "content": "4.7. More Discussions" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 238, + 555, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 238, + 555, + 274 + ], + "spans": [ + { + "bbox": [ + 313, + 238, + 555, + 274 + ], + "type": "text", + "content": "Real-data Experiment: Table 7 shows the evaluation on the FR1 real-world dataset, which demonstrates the superior performance of the proposed method on real data." + } + ] + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 315, + 304, + 549, + 327 + ], + "blocks": [ + { + "bbox": [ + 350, + 284, + 518, + 295 + ], + "lines": [ + { + "bbox": [ + 350, + 284, + 518, + 295 + ], + "spans": [ + { + "bbox": [ + 350, + 284, + 518, + 295 + ], + "type": "text", + "content": "Table 7. Real data experiment on FR1 dataset." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 304, + 549, + 327 + ], + "lines": [ + { + "bbox": [ + 315, + 304, + 549, + 327 + ], + "spans": [ + { + "bbox": [ + 315, + 304, + 549, + 327 + ], + "type": "table", + "html": "
MethodGSACNMFHySureZSLPLRDiffHIR-DiffProposed
QNRT0.5900.6790.6680.8020.6300.6250.832
", + "image_path": "284c3279f22065d1a24402d04bba9350471cf314ab309a861719b401754856bb.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 338, + 554, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 554, + 374 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 554, + 374 + ], + "type": "text", + "content": "More Experimental Comparisons: We provide further discussion on running time, downsampling factor, and kernel size, which are shown in the supplementary material." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 374, + 554, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 374, + 554, + 459 + ], + "spans": [ + { + "bbox": [ + 313, + 374, + 554, + 459 + ], + "type": "text", + "content": "Limitations: While the computational cost of DMs with the proposed zero-shot guidance is comparable to that of previous DM-based methods, such as PLRDiff [30], NSSD requires more time to iteratively construct the spectral mapping from the RGB detail image to the HR-HSI, as detailed in the supplementary material. We will explore the use of pre-trained networks to address this limitation in the future." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 468, + 388, + 481 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 388, + 481 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 388, + 481 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 489, + 555, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 489, + 555, + 633 + ], + "spans": [ + { + "bbox": [ + 313, + 489, + 555, + 633 + ], + "type": "text", + "content": "In this paper, we introduced a novel guided diffusion scheme for hyperspectral pansharpening, combining the iterative update of zero-shot guidance and NSSD. The zero-shot guidance is conducted with an auxiliary convolutional network that can be trained just on one-pair inputs, while NSSD acts as a flexible transformation mapping the generated images by pre-trained DMs to the hyperspectral images. Our method demonstrated significant effectiveness on benchmark datasets, including Pavia, WDC, and Chikusei. Through extensive ablation studies, we analyzed the contributions of the guidance and NSSD and evaluated the impact by different zero-shot network architectures." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 647, + 555, + 707 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 555, + 707 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 555, + 707 + ], + "type": "text", + "content": "Acknowledgement This work was supported in part by the NSFC (12171072, 12271083, 62273077), Natural Science Foundation of Sichuan Province (2024NSFSC0038), and National Key Research and Development Program of China (2020YFA0714001)." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12676" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "text", + "content": "[1] B. Aiazzi, S. Baronti, and M. Selva. Improving component substitution pansharpening through multivariate regression of MS + Pan data. IEEE Trans. Geosci. Remote Sens., 45 (10):3230-3239, 2007. 5, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 136, + 296, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 136, + 296, + 178 + ], + "spans": [ + { + "bbox": [ + 61, + 136, + 296, + 178 + ], + "type": "text", + "content": "[2] L. Alparone, L. Wald, J. Chanussot, C. Thomas, P. Gamba, and L. M. Bruce. Comparison of pansharpening algorithms: Outcome of the 2006 GRS-S data-fusion contest. IEEE Trans. Geosci. Remote Sens., 45(10):3012-3021, 2007. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 180, + 295, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 180, + 295, + 223 + ], + "spans": [ + { + "bbox": [ + 62, + 180, + 295, + 223 + ], + "type": "text", + "content": "[3] L. Alparone, B. Aiazzi, S. Baronti, A. Garzelli, F. Nencini, and M. Selva. Multispectral and panchromatic data fusion assessment without reference. Photogramm. Eng. Remote Sens., 74(2):193-200, 2008. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 224, + 294, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 224, + 294, + 267 + ], + "spans": [ + { + "bbox": [ + 62, + 224, + 294, + 267 + ], + "type": "text", + "content": "[4] W. G. C. Bandara, J. M. J. Valanarasu, and V. M. Patel. Hyperspectral pansharpening based on improved deep image prior and residual reconstruction. EEE Trans. Geosci. Remote Sens., 60:1-16, 2021. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 269, + 294, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 269, + 294, + 312 + ], + "spans": [ + { + "bbox": [ + 62, + 269, + 294, + 312 + ], + "type": "text", + "content": "[5] Z. Cao, S. Cao, L.-J. Deng, X. Wu, J. Hou, and G. Vivone. Diffusion model with disentangled modulations for sharpening multispectral and hyperspectral images. Inf. Fusion, 104: 102158, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 314, + 294, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 314, + 294, + 346 + ], + "spans": [ + { + "bbox": [ + 62, + 314, + 294, + 346 + ], + "type": "text", + "content": "[6] H. Chung, J. Kim, M. T. McCann, M. L. Klasky, and J. C. Ye. Diffusion posterior sampling for general noisy inverse problems. arXiv preprint arXiv:2209.14687, 2022. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 347, + 294, + 389 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 347, + 294, + 389 + ], + "spans": [ + { + "bbox": [ + 62, + 347, + 294, + 389 + ], + "type": "text", + "content": "[7] M. Ciotola, G. Guarino, G. Vivone, G. Poggi, J. Chanussot, A. Plaza, and G. Scarpa. Hyperspectral pansharpening: Critical review, tools and future perspectives. arXiv preprint arXiv:2407.01355, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 391, + 294, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 391, + 294, + 434 + ], + "spans": [ + { + "bbox": [ + 62, + 391, + 294, + 434 + ], + "type": "text", + "content": "[8] L.-J. Deng, G. Vivone, M. E. Paoletti, G. Scarpa, J. He, Y. Zhang, J. Chanussot, and A. Plaza. Machine learning in pan-sharpening: A benchmark, from shallow to deep networks. IEEE Geosci. Remote Sens. Mag., 10(3):279-315, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 436, + 294, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 436, + 294, + 468 + ], + "spans": [ + { + "bbox": [ + 62, + 436, + 294, + 468 + ], + "type": "text", + "content": "[9] P. Dhariwal and A. Nichol. Diffusion models beat gans on image synthesis. Adv. Neural Inf. Process. Syst. (NeurIPS), 34:8780-8794, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 469, + 294, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 469, + 294, + 512 + ], + "spans": [ + { + "bbox": [ + 57, + 469, + 294, + 512 + ], + "type": "text", + "content": "[10] R. Dian, L. Fang, and S. Li. Hyperspectral image superresolution via non-local sparse tensor factorization. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 5344-5353, 2017. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 514, + 294, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 514, + 294, + 546 + ], + "spans": [ + { + "bbox": [ + 57, + 514, + 294, + 546 + ], + "type": "text", + "content": "[11] R. Dian, A. Guo, and S. Li. Zero-shot hyperspectral sharpening. IEEE Trans. Pattern Anal. Mach. Intell., 45(10):12650-12666, 2023. 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 548, + 294, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 548, + 294, + 581 + ], + "spans": [ + { + "bbox": [ + 57, + 548, + 294, + 581 + ], + "type": "text", + "content": "[12] W. Dong, T. Zhang, J. Qu, S. Xiao, J. Liang, and Y. Li. Laplacian pyramid dense network for hyperspectral pansharpening. IEEE Trans. Geosci. Remote Sens., 60:1-13, 2021. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 582, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 582, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 582, + 294, + 624 + ], + "type": "text", + "content": "[13] Y. Duan, X. Wu, H. Deng, and L.-J. Deng. Content-adaptive non-local convolution for remote sensing pansharpening. In Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recog., pages 27738-27747, 2024. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 625, + 294, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 625, + 294, + 668 + ], + "spans": [ + { + "bbox": [ + 57, + 625, + 294, + 668 + ], + "type": "text", + "content": "[14] X. Fu, Z. Lin, Y. Huang, and X. Ding. A variational pan-sharpening with local gradient constraints. In Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recog., pages 10265-10274, 2019. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "type": "text", + "content": "[15] J. Gao, J. Li, X. Su, M. Jiang, and Q. Yuan. Deep image interpolation: A unified unsupervised framework for pansharpening. In Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recog., pages 609-618, 2022. 1" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[16] M. Gong, H. Zhang, H. Wang, J. Chen, J. Huang, X. Tian, and J. Ma. Cross-scale domain adaptation with comprehensive information for pansharpening. In Int. Joint Conf. Artif. Intell, pages 785-793, 2024. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 118, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 161 + ], + "type": "text", + "content": "[17] L. He, H. Ye, D. Xi, J. Li, A. Plaza, and M. Zhang. Tree-structured neural network for hyperspectral pansharpening. IEEE J. Sel. Top. Appl. Earth Observ. Remote Sens., 17: 2516-2530, 2023. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 163, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 553, + 206 + ], + "type": "text", + "content": "[18] L. He, Z. Fang, J. Li, J. Chanussot, and A. Plaza. Two spectral-spatial implicit neural representations for arbitrary-resolution hyperspectral pansharpening. IEEE Trans. Geosci. Remote Sens., 2024. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 208, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 208, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 317, + 208, + 553, + 251 + ], + "type": "text", + "content": "[19] W. He, Q. Yao, C. Li, N. Yokoya, and Q. Zhao. Non-local meets global: An integrated paradigm for hyperspectral denoising. In Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recog., pages 6868-6877, 2019. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 254, + 553, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 254, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 317, + 254, + 553, + 296 + ], + "type": "text", + "content": "[20] W. He, Q. Yao, C. Li, N. Yokoya, Q. Zhao, H. Zhang, and L. Zhang. Non-local meets global: An iterative paradigm for hyperspectral image restoration. IEEE Trans. Pattern Anal. Mach. Intell., 44(4):2089-2107, 2020. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 298, + 553, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 298, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 317, + 298, + 553, + 331 + ], + "type": "text", + "content": "[21] Jo. Ho, A. Jain, and P. Abbeel. Denoising diffusion probabilistic models. In Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), pages 6840-6851, 2020. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 332, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 332, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 317, + 332, + 553, + 374 + ], + "type": "text", + "content": "[22] S. Huang, M. Ren, Y. Yang, X. Wang, and Y. Wei. MFTN: A multi-scale feature transfer network based on IMatchformer for hyperspectral image super-resolution. In *Forty-first Int. Conf. Mach. Learn.*, 2024. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 376, + 553, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 409 + ], + "type": "text", + "content": "[23] B. Kawar, M. Elad, S. Ermon, and J. Song. Denoising diffusion restoration models. Proc. Adv. Neural Inf. Process. Syst., 35:23593-23606, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 411, + 553, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 432 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 432 + ], + "type": "text", + "content": "[24] Tamara G Kolda and Brett W Bader. Tensor decompositions and applications. SIAM Rev., 51(3):455-500, 2009. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 434, + 553, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 553, + 466 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 553, + 466 + ], + "type": "text", + "content": "[25] S. Li, S. Li, and L. Zhang. Hyperspectral and panchromatic images fusion based on the dual conditional diffusion models. IEEE Trans. Geosci. Remote Sens., 2023. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 468, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 468, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 317, + 468, + 553, + 521 + ], + "type": "text", + "content": "[26] L. Loncan, L. B. De Almeida, J. M. Bioucas-Dias, X. Briottet, J. Chanussot, N. Dobigeon, S. Fabre, W. Liao, G. A. Licciardi, M. Simoes, et al. Hyperspectral pansharpening: A review. IEEE Geosci. Remote Sens. Mag., 3(3):27-46, 2015. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 524, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 524, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 317, + 524, + 553, + 567 + ], + "type": "text", + "content": "[27] Y. Miao, L. Zhang, L. Zhang, and D. Tao. DDS2M: Self-supervised denoising diffusion spatio-spectral model for hyperspectral image restoration. In Proc. IEEE/CVF Int. Conf. Comput. Vis., pages 12086-12096, 2023. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 569, + 553, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 569, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 317, + 569, + 553, + 622 + ], + "type": "text", + "content": "[28] A. Quinn Nichol, P. Dhariwal, A. Ramesh, P. Shyam, P. Mishkin, B. Mcgrew, I. Sutskever, and M. Chen. GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. In Int. Conf. Mach. Learn., pages 16784-16804. PMLR, 2022. 4" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 624, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 624, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 317, + 624, + 553, + 678 + ], + "type": "text", + "content": "[29] L. Pang, X. Rui, L. Cui, H. Wang, D. Meng, and X. Cao. HIR-Diff: Unsupervised hyperspectral image restoration via improved diffusion models. In Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recog., pages 3005-3014, 2024. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "text", + "content": "[30] X. Rui, X. Cao, L. Pang, Z. Zhu, Z. Yue, and D. Meng. Unsupervised hyperspectral pansharpening via low-rank diffusion model. Inf. Fusion, 107:102325, 2024. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12677" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "text", + "content": "[31] K. Shi, J. Peng, J. Gao, Y. Luo, and S. Xu. Hyperspectral image denoising via double subspace deep prior. IEEE Trans. Geosci. Remote Sens., 62:1-15, 2024. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "type": "text", + "content": "[32] M. Simoes, J. Bioucas-Dias, L. B. Almeida, and J. Chanussot. A convex formulation for hyperspectral image superresolution via subspace-based regularization. IEEE Trans. Geosci. Remote Sens., 53(6):3373-3388, 2014. 5, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 185 + ], + "type": "text", + "content": "[33] Y. Song and S. Ermon. Generative modeling by estimating gradients of the data distribution. In Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), pages 11895-11907, 2019. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 294, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 229 + ], + "type": "text", + "content": "[34] Y. Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole. Score-based generative modeling through stochastic differential equations. In Int. Conf. Learn. Representations, 2021. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 231, + 294, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 294, + 263 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 294, + 263 + ], + "type": "text", + "content": "[35] D. Ulyanov, A. Vedaldi, and V. Lempitsky. Deep image prior. In Proc. IEEE Conf. Comput. Vis. Pattern Recog., pages 9446-9454, 2018. 5, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 265, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 265, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 265, + 294, + 330 + ], + "type": "text", + "content": "[36] G. Vivone, M. Dalla Mura, A. Garzelli, R. Restaino, G. Scarpa, M. O. Ulfarsson, L. Alparone, and J. Chanussot. A new benchmark based on recent advances in multispectral pansharpening: Revisiting pansharpening with classical and emerging pansharpening methods. IEEE Geosci. Remote Sens. Mag., 9(1):53-81, 2020. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 332, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 294, + 374 + ], + "type": "text", + "content": "[37] W. Wang, L.-J. Deng, R. Ran, and G. Vivone. A general paradigm with detail-preserving conditional invertible network for image fusion. Int. J. Comput. Vis., 132(4):1029-1054, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 376, + 294, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 294, + 419 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 294, + 419 + ], + "type": "text", + "content": "[38] Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Trans. Image Process., 13(4):600-612, 2004. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 422, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 464 + ], + "type": "text", + "content": "[39] C. Wu, D. Wang, Y. Bai, H. Mao, Y. Li, and Q. Shen. HSR-Diff: Hyperspectral image super-resolution via conditional diffusion models. In Proc. IEEE/CVF Int. Conf. Comput. Vis., pages 7083-7093, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 467, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 294, + 510 + ], + "type": "text", + "content": "[40] J.-L. Xiao, T.-Z. Huang, L.-J. Deng, Z.-C. Wu, X. Wu, and G. Vivone. Variational pansharpening based on coefficient estimation with nonlocal regression. IEEE Trans. Geosci. Remote. Sens., 61:1-15, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 512, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 294, + 544 + ], + "type": "text", + "content": "[41] W. Xie, J. Lei, Y. Cui, Y. Li, and Q. Du. Hyperspectral pansharpening with deep priors. IEEE Trans. Neural Netw. Learn. Syst., 31(5):1529-1543, 2019. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 545, + 294, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 294, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 294, + 588 + ], + "type": "text", + "content": "[42] Y. Xing, L. Qu, S. Zhang, K. Zhang, Y. Zhang, and L. Bruzzone. Crossdiff: Exploring self-supervised representation of pansharpening via cross-predictive diffusion model. IEEE Trans. Image Process., 33:5496-5509, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 590, + 294, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 633 + ], + "type": "text", + "content": "[43] T. Xu, T.-Z. Huang, L.-J. Deng, and N. Yokoya. An iterative regularization method based on tensor subspace representation for hyperspectral image super-resolution. IEEE Trans. Geosci. Remote Sens., 60:1-16, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 635, + 294, + 688 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 688 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 688 + ], + "type": "text", + "content": "[44] T. Xu, T.-Z. Huang, L.-J. Deng, J.-L. Xiao, C. Broni-Bediako, J. Xia, and N. Yokoya. A coupled tensor double-factor method for hyperspectral and multispectral image fusion. IEEE Trans. Geosci. Remote Sens., 62:1-17, 2024. 2, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "type": "text", + "content": "[45] J. Xue, Y.-Q. Zhao, Y. Bu, W. Liao, J. C.-W. Chan, and W. Philips. Spatial-spectral structured sparse low-rank represent" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 384 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "tation for hyperspectral image super-resolution. IEEE Trans. Image Process., 30:3084-3097, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "text", + "content": "[46] J. Yao, D. Hong, J. Chanussot, D. Meng, X. Zhu, and Z. Xu. Cross-attention in coupled unmixing nets for unsupervised hyperspectral super-resolution. In Proc. Eur. Conf. Comput. Vis., pages 208–224. Springer, 2020. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "type": "text", + "content": "[47] N. Yokoya, T. Yairi, and A. Iwasaki. Coupled nonnegative matrix factorization unmixing for hyperspectral and multispectral data fusion. IEEE Trans. Geosci. Remote Sens., 50 (2):528-537, 2011. 1, 5, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 186, + 553, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 238 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 238 + ], + "type": "text", + "content": "[48] Y. Yu, E. Pan, X. Wang, Y. Wu, X. Mei, and J. Ma. Unmixing before fusion: A generalized paradigm for multi-source-based hyperspectral image synthesis. In Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recog., pages 9297-9306, 2024." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 241, + 553, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 295 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 295 + ], + "type": "text", + "content": "[49] R. H. Yuhas, A. F. H. Goetz, and J. W. Boardman. Discrimination among semi-arid landscape endmembers using the spectral angle mapper (SAM) algorithm. In Proc. Summaries 3rd Annu. JPL Airborne Geosci. Workshop, pages 147-149, 1992. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 297, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 330 + ], + "type": "text", + "content": "[50] Y. Zhong, X. Wu, L.-J. Deng, and Z. Cao. SSDiff: Spatial-spectral integrated diffusion model for remote sensing pan-sharpening. arXiv preprint arXiv:2404.11537, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 331, + 553, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 384 + ], + "type": "text", + "content": "[51] Y.-W. Zhuo, T.-J. Zhang, J.-F. Hu, H.-X. Dou, T.-Z. Huang, and L.-J. Deng. A deep-shallow fusion network with multi-detail extractor and spectral attention for hyperspectral pan-sharpening. IEEE J. Sel. Top. Appl. Earth Observ. Remote Sens., 15:7539-7555, 2022. 8" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12678" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/I2VGuard_ Safeguarding Images against Misuse in Diffusion-based Image-to-Video Models/612ebd29-cda5-46e3-b035-7c7393e6f819_content_list.json b/2025/I2VGuard_ Safeguarding Images against Misuse in Diffusion-based Image-to-Video Models/612ebd29-cda5-46e3-b035-7c7393e6f819_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6fa2f789871265a04fd29660d202e3254632ef82 --- /dev/null +++ b/2025/I2VGuard_ Safeguarding Images against Misuse in Diffusion-based Image-to-Video Models/612ebd29-cda5-46e3-b035-7c7393e6f819_content_list.json @@ -0,0 +1,2277 @@ +[ + { + "type": "text", + "text": "I2VGuard: Safeguarding Images against Misuse in Diffusion-based Image-to-Video Models", + "text_level": 1, + "bbox": [ + 147, + 130, + 823, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dongnan Gui $^{1,*}$", + "bbox": [ + 220, + 202, + 349, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xun Guo2", + "bbox": [ + 383, + 203, + 467, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wengang Zhou", + "bbox": [ + 506, + 203, + 637, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yan Lu2", + "bbox": [ + 678, + 203, + 746, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of Science and Technology of China", + "bbox": [ + 171, + 220, + 553, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2Microsoft Research Asia", + "bbox": [ + 593, + 220, + 799, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{gdn2001@mail.,zhwg@}ustc.edu.cn{xunguo，yanlu}@microsoft.com", + "bbox": [ + 205, + 241, + 759, + 256 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cd5a9666902b55dad978753a14b6623d7bc761247732aebf941009b5ed30e5a7.jpg", + "image_caption": [ + "Original Image" + ], + "image_footnote": [], + "bbox": [ + 83, + 272, + 243, + 344 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b190dee41c14ae9a45c74f58645879e1586b85acf5240a675ff4848cb697eae2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 248, + 272, + 408, + 343 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/051a828e773f4ee3851ddc9f28a8a74f795e83c7df6a5b1ad00e8c461a1c1fc4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 408, + 272, + 568, + 343 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cbe52968a8b12665d23cd911ca5608df22960afc66255a44296a065ecff0ed9c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 566, + 272, + 727, + 343 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f1edf07cec993b7d1c4c97a6ca1d8a314aa29a5a280a5b87f782a073f43977cb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 727, + 272, + 887, + 343 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f9a319aa8b7a334532a2d2bceff2c427f58e4a804e497353b5fed06c19dfcf5b.jpg", + "image_caption": [ + "Guarded Image" + ], + "image_footnote": [], + "bbox": [ + 83, + 353, + 243, + 422 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/589d90852a09076afc400ebc93d4e1f68a816aa23cab1c393a3ce95c57a661a3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 248, + 353, + 408, + 422 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cef38a4576704821a26cf49c350e4aeec3a722a3d431d59bd8b57dbd5d5c0a3b.jpg", + "image_caption": [ + "Generated video from original image.", + "Generated video from guarded image." + ], + "image_footnote": [], + "bbox": [ + 408, + 353, + 568, + 422 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f70473bfbc3bab0cc2728e2266eb7ef5280c4ce76af2b67d690dfac815ea1513.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 566, + 353, + 727, + 422 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/175a6c6eb12240bd22927794a1d8232bbd16d3fd7e6cca83b63190e100e6d4aa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 727, + 353, + 887, + 422 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5f3acdcdfb3a1d63d34e5882c9402bd32ccee10bf1547777cfb3af1f6e58bac2.jpg", + "image_caption": [ + "Original Image" + ], + "image_footnote": [], + "bbox": [ + 83, + 440, + 243, + 511 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/512509244fecfd83cd0928919853d3e5ddd552e1fed217f270119af7b49c05b5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 248, + 439, + 408, + 510 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/33dceaebe1057b006cefbab2d6a5950991aef561b10c4e464af994cdf065e5f6.jpg", + "image_caption": [ + "Generated video from original image." + ], + "image_footnote": [], + "bbox": [ + 408, + 439, + 568, + 508 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/22e38137629a7877bed27ee515b2f9ea8363b15c368f8f601ff01fda0595890f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 566, + 439, + 727, + 508 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a0e2340d8dc6e79fa2720cce58e46cfe5d0e1d7c139c11462c689a7806f349c4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 727, + 439, + 887, + 508 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/def9bfe6a5187b68c1ac5b4c089ed6741f1e34b41b7ab85de8b424622890ddc4.jpg", + "image_caption": [ + "Guarded Image", + "Figure 1. Results of our I2VGuard. We present original images, guarded images, and their corresponding SVD-generated videos. All results are generated with the same seed. Our method effectively safeguards images from animation in image-to-video generation." + ], + "image_footnote": [], + "bbox": [ + 83, + 521, + 243, + 590 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c0bac01a0e9ca711bb9935d7687b3e6f683ce4a02bb58963e625b6b1cefd9c36.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 248, + 521, + 408, + 590 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5494f90f7e7b67167e5fa5c9f2a272746a3463b009a1fd1fdc25f4a87a0a6b06.jpg", + "image_caption": [ + "Generated video from guarded image." + ], + "image_footnote": [], + "bbox": [ + 408, + 521, + 568, + 590 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5e2c7109f832796edcd5af975f50944984abdb2bf07d6def1d09a6b1eabfa2d2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 566, + 521, + 727, + 590 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1b71dfa1a84a6fb6a373ad966290ca33dba26e0d2b220cbac53a67ccdc803798.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 727, + 521, + 887, + 590 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 655, + 313, + 671 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advances in image-to-video generation have enabled animation of still images and offered pixel-level controllability. While these models hold great potential to transform single images into vivid and dynamic videos, they also carry risks of misuse that could impact privacy, security, and copyright protection. This paper proposes a novel approach that applies imperceptible perturbations on images to degrade the quality of the generated videos, thereby protecting images from misuse in white-box image-to-video diffusion models. Specifically, we function our approach as an adversarial attack, incorporating spatial, temporal, and diffusion attack modules. The spatial attack shifts image features from", + "bbox": [ + 73, + 688, + 473, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "their original distribution to a lower-quality target distribution, reducing visual fidelity. The temporal attack disrupts coherent motion by interfering with temporal attention maps that guide motion generation. To enhance the robustness of our approach across different models, we further propose a diffusion attack module leveraging contrastive loss. Our approach can be easily integrated with mainstream diffusion-based I2V models. Extensive experiments on SVD, CogVideoX, and ControlNeXt demonstrate that our method significantly impairs generation quality in terms of visual clarity and motion consistency, while introducing only minimal artifacts to the images. To the best of our knowledge, we are the first to explore adversarial attacks on image-to-video generation for security purposes.", + "bbox": [ + 496, + 656, + 895, + 869 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work done during internship at Microsoft Research Asia.", + "bbox": [ + 76, + 886, + 387, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12595", + "bbox": [ + 480, + 944, + 517, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 80, + 90, + 207, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion models [15, 19, 35] have gained significant attention recently, particularly in video generation. Advances in models like Sora [5] and its DiT [33] framework have accelerated the development of video diffusion models, leading to numerous academic [4, 22, 27, 53, 55] and commercial [43-45] applications. The primary types of video generation are text-to-video(T2V) and image-to-video(I2V), with the former supporting greater diversity and the latter offering pixel-level control. Leveraging powerful generative models and techniques from text-to-image generation [35], image-to-video models can now produce visually impressive videos. Other conditional I2V models include CogVideoX [53], which generates videos using both text and image inputs, enabling more versatile content creation, and ControlNeXt [34], which focuses on generating videos conditioned on pose information, offering precise control over motion and structure.", + "bbox": [ + 80, + 114, + 468, + 354 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, the rapid advancement of generative models has also heightened risks related to privacy, security, and copyright [2]. For instance, an adversary with access to a person's image could use image-to-video models to generate misleading or harmful videos, depicting actions the individual never performed, thus compromising individual's privacy and security. Additionally, the unauthorized use of copyrighted images to create animations poses significant risks to intellectual property rights. To address these risks, this work introduces an adversarial attack method designed to prevent the misuse of images in diffusion-based I2V models. Prior research has primarily focused on protecting images from malicious edits [37, 54], which are often viewed as adversarial attacks on image-editing models. Few efforts have focused on safeguarding images from misuse in I2V tasks.", + "bbox": [ + 80, + 357, + 468, + 582 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main challenges in attacking I2V models come from two aspects: first, disrupting temporal consistency with perturbations applied only to individual images; and second, achieving robustness across diverse diffusion models, including UNet and DiT architectures, even under strong conditioning such as poses and text prompts. To address these challenges, our method targets different modules within I2V models and affects various attributes of the generated video. Specifically, we divide our approach into spatial, temporal, and diffusion attacks. The spatial attack focuses on altering the spatial features of the I2V models by targeting the VAE encoder in latent diffusion models. Here, we manipulate image latents, shifting them toward suboptimal values. The temporal attack aims to perturb the temporal features, which are crucial for video motion and consistency. By extracting the temporal attention map, we intentionally disrupt it by pushing it away from its original attention map. This creates a chaotic attention map, leading to temporal inconsistencies in the generated video. The diffusion attack focuses on the diffusion process and targets the output of the denoising module, which includes the UNet in SVD and the trans", + "bbox": [ + 80, + 584, + 468, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "former in CogVideoX. Unlike a standard denoising process, our approach aims to push the predicted outputs away from the original frames, guiding them closer to target frames. To achieve this, we design a contrastive loss that degrades the quality of the predicted frames by treating the target frames as positive samples and the original predicted frames as negative samples.", + "bbox": [ + 501, + 90, + 890, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We conduct a comprehensive qualitative and quantitative analysis of the attack results. Qualitative analysis demonstrates that our method effectively disrupts the original image, leading to low temporal consistency and reduced aesthetic quality in both simple and conditioned I2V models. Quantitative analysis confirms a significant degradation in both spatial quality and temporal consistency of the generated videos. To our knowledge, this is the first adversarial attack method aimed at securing images against misuse in I2V generation.", + "bbox": [ + 503, + 198, + 890, + 345 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 503, + 359, + 638, + 375 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image-to-Video Generation With the great advancements in image generation [15, 19, 31, 35, 40, 41], video generation [3, 5, 20, 22, 39, 43–45, 52, 53] has gathered significant attention and is developing rapidly [18, 48, 49, 51]. Recently, following the success of diffusion models [15, 19, 35, 40] in image generation, diffusion models integrated with UNet [3, 14, 21, 36] or transformers (DiT) [5, 33, 47, 55] have been adopted for video generation. In the field of imaged-to-video generation,AnimateDiff [18] animates personalized text-to-image diffusion models by inserting a temporal motion module, Stable Video Diffusion(SVD) [4] demonstrates strong performance, benefiting from intensive training on large amounts of high-quality data, CogVideoX [53] supports text-image jointly conditional video generation,Animate Anyone [24] and ControlNeXt [34] leverage pose information for controllable image-to-video generation. In this work, we focus on I2V models, which are susceptible to misuse and therefore require robust protection.", + "bbox": [ + 503, + 385, + 890, + 654 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Adversarial Attacks on Diffusion Models Current attacks on diffusion models, which can be broadly categorized [46] into backdoor [9, 12] and adversarial attacks, primarily target image-based diffusion models [13, 16, 17, 50]. In adversarial attacks, the input image or text prompt is perturbed to generate adversarial examples. For instance, AdvDM [30] and Mist [29] utilize adversarial examples to protect human-created artworks, PhotoGuard [37] alters the image by modifying either the encoder or the diffusion outputs against malicious editing, while Glaze [38] perturbs the image to prevent style mimicry, [54] shifts the image away from its original distribution, [56] embeds personal watermarks in the generation of adversarial examples, and DiffusionGuard [11] generates adversarial noise targeting the early stage of the diffusion process. For video-based attacks,", + "bbox": [ + 503, + 674, + 890, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12596", + "bbox": [ + 480, + 946, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/97cf788ca4fa741f68aa6c8c0d9f9e680a5bc2d9eab9c2728ff5ec1e1d2849d4.jpg", + "image_caption": [ + "Figure 2. Overview of our proposed method. The training process begins with a trainable copy $I_{adv}$ of the original image $I_{src}$ . First, we perform inference to obtain the original video $V_{0}$ and get the latent frames $\\mathbf{X}_{0}$ . Noisy latent frames $\\mathbf{X}_{t}$ and the latent image $x_{adv}, x_{src}$ are then processed by the denoising model to predict the original frames. The encoded latent image $x_{adv}$ is utilized for the spatial encoder attack, while the predicted original frames $\\hat{\\mathbf{X}}_{0}$ serve in the diffusion attack to compute contrastive loss. Within the denoising module, we hook into the temporal attention module to extract the temporal attention map. By altering the attacked attention map $A_{adv}$ to diverge from the original $A_{src}$ , we implement the temporal attack." + ], + "image_footnote": [], + "bbox": [ + 94, + 101, + 880, + 406 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "PRIME [28] modifies videos frame-by-frame to prevent malicious video editing, while [32] disrupts style mimicry attacks from video imagery. There are also works that protect the image against other generation, such as text-to-3D generation [42]. In this work, we focus on perturbing the input image in image-to-video diffusion models to perform adversarial attacks. Recent study [23] show that adversarial perturbations fail to reliably protect artists from style mimicry in the field of image generation. Our approach stands out in two ways: it uses a white-box attack strategy, allowing adversarial noise training with both generation and pre-processing models, and it focuses on disrupting temporal consistency rather than mere image fidelity, making the perturbations harder to remove.", + "bbox": [ + 75, + 529, + 472, + 741 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 765, + 168, + 781 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first provide preliminaries on latent diffusion models and problem settings in Sec. 3.1. We then present our attack method, detailing the spatial attack in Sec. 3.2, the temporal attack in Sec. 3.3, and the diffusion attack in Sec. 3.4. An overview of our method is presented in in Fig. 2, which shows our protection procedure involves spatial, temporal, and diffusion attacks.", + "bbox": [ + 75, + 794, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 500, + 529, + 629, + 545 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Latent Diffusion Models In diffusion models, realistic image or video sampling from a target distribution $q(\\cdot)$ is achieved via a sequence of denoising steps in a diffusion process. This approach involves iteratively adding and removing noise to transform an initial sample $\\mathbf{x}_0 \\sim q(\\cdot)$ . Noise is incrementally introduced across steps, creating intermediate states defined by $\\mathbf{x}_{t+1} = \\sqrt{\\overline{\\alpha}_t} \\mathbf{x}_0 + \\sqrt{1 - \\overline{\\alpha}_t} \\epsilon_t$ , where $\\epsilon_t$ is Gaussian noise. The parameters $\\alpha_t$ control the noise transition, leading to a final noisy state with a standard Gaussian distribution. By reversing this process and iteratively denoising each state, the model learns to predict the noise added at each step, ultimately enabling the recovery of samples from $q(\\cdot)$ starting from $\\mathbf{x}_T$ . Latent diffusion models (LDMs), a subclass of diffusion models, operate in a compressed latent space rather than directly in the image space. In LDMs, each input image or video $\\mathbf{x}_0$ is first mapped to a latent representation $\\mathcal{E}(\\mathbf{x}_0)$ via an encoder $\\mathcal{E}$ . The diffusion process then applies in this latent space, producing a series of noisy representations. Training in this latent space enables LDMs to achieve efficient training and faster inference while retaining high-quality outputs.", + "bbox": [ + 496, + 582, + 895, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12597", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Problem Settings Adversarial attacks involve adding subtle, nearly undetectable perturbations $\\delta_{adv}$ to the input data to influence a model's output. In computer vision, adversarial attacks have been widely explored in classifications and image synthesis. However, in image-to-video generation tasks, adversarial attack remains a relatively unexplored area. In this work, we aim to develop adversarial attacks to safeguard images from misused by image-to-video latent diffusion models. To distinguish between the latent and pixel spaces, we represent images and videos in pixel space as $I$ and $V$ , and in latent space as $x$ and $\\mathbf{X}$ , respectively.", + "bbox": [ + 75, + 90, + 472, + 257 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, an adversary can generate a modified image $I_{adv} = I_{src} + \\delta$ from source image $I_{src}$ , crafted to disrupt the generative model $\\mathcal{G}$ , resulting in poor-quality or distorted outputs. This adversarial image is optimized by maximizing the loss between the videos generated from the original and modified images, represented as follows:", + "bbox": [ + 76, + 257, + 472, + 349 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\delta_ {a d v} = \\arg \\max _ {\\| \\delta \\| _ {p} \\leq \\epsilon} \\mathcal {L} (\\mathcal {G} (I _ {s r c} + \\delta , \\mathbf {X} _ {t}, c), \\mathcal {G} (I _ {s r c}, \\mathbf {X} _ {t}, c))\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 359, + 455, + 386 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\delta_{adv}$ represents the computed perturbation, constrained by $\\| \\delta \\| _p\\leq \\epsilon$ to ensure minimal visibility, and $c$ represents other conditions such as text or pose in the imaged-to-video generation process. The constraint on $\\delta_{adv}$ is maintained by a regularization term $\\| I_{adv} - I_{src}\\| ^2$", + "bbox": [ + 75, + 398, + 472, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Spatial Attack", + "text_level": 1, + "bbox": [ + 76, + 483, + 225, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The generated video has two main attributes: spatial features and temporal features. In this section, we focus on attacking the spatial features within individual frames. Previous work [37, 54] has applied adversarial attacks to image editing tasks, focusing solely on spatial features. Our work extends these methods to image-to-video generation task.", + "bbox": [ + 75, + 507, + 468, + 598 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In Latent Diffusion Models (LDMs), an encoder, denoted as $\\mathcal{E}$ , compresses the original image or frames into a latent space representation. An encoder attack aims to compromise this encoder, reducing the effectiveness of generative models.", + "bbox": [ + 75, + 598, + 470, + 659 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {e n c} = \\left\\| \\mathcal {E} \\left(I _ {a d v}\\right) - \\mathcal {E} \\left(I _ {t g}\\right) \\right\\| ^ {2} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 672, + 468, + 691 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, as shown in Eq. (1), the attack transforms the latent representation $\\mathcal{E}(I_{adv})$ into a poor one $\\mathcal{E}(I_{tg})$ , resulting in poor-quality final videos.", + "bbox": [ + 75, + 700, + 468, + 744 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Temporal Attack", + "text_level": 1, + "bbox": [ + 76, + 756, + 243, + 772 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Unlike attacks on image editing tasks, where no temporal features are involved, attacks on video generation rely heavily on temporal information. In this task, disrupting temporal consistency through adversarial attacks is both essential and challenging. To reduce temporal consistency and induce chaotic behavior in the outputs, we target the temporal attention maps within the temporal transformer blocks in video diffusion generator, $\\mathcal{G}$ . Specifically, during training, we insert", + "bbox": [ + 75, + 779, + 472, + 902 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/bc65197530caa0c39ed1dbb219022649ca06a1ae592d05301895de15a36916f1.jpg", + "image_caption": [ + "Figure 3. Temporal self-attention map visualizations. Difference between generated frames from the original image (left) and the guarded image (right) in Fig. 1." + ], + "image_footnote": [], + "bbox": [ + 504, + 90, + 694, + 174 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1258d5cfa6a7d58c434cd3dd88e56b6cbfadef2f95a2eb094cc497a9acea76b7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 90, + 887, + 174 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "a hook, denoted as $\\mathcal{G}$ , within the attention block to extract the self-attention map $A_{adv}$ . We then apply a targeted attack to this attention map, pushing it away from the original map, $A_{src}$ , thereby reducing temporal consistency in the generated outputs. To achieve this, we introduce a temporal attention loss, denoted as $\\mathcal{L}_{att}$ , as shown in Eq. (2), which induces motion-related perturbations and promotes temporal chaos.", + "bbox": [ + 496, + 252, + 893, + 380 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {a t t}} = \\tau_ {2} - \\left| \\left| A _ {\\text {a d v}} - A _ {\\text {s r c}} \\right| \\right| ^ {2} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 381, + 890, + 398 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig. 3, the visualization of the temporal self-attention map on both the original and modified frames highlights this impact: the manipulated attention map misaligns with the original, directly leading to inconsistencies in motion across frames. This visual evidence underscores the effectiveness of our method in disrupting temporal coherence in video generation by distorting the temporal attention mechanism within the model.", + "bbox": [ + 496, + 402, + 893, + 523 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Diffusion Attack", + "text_level": 1, + "bbox": [ + 500, + 531, + 663, + 545 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Although spatial and temporal attacks can disrupt structural and motion information in generated videos, their effectiveness may vary significantly across different models [4, 34, 53], especially when other conditions, e.g., text prompts, are involved. To enhance the robustness and generalization of the perturbation across diffusion models, we introduce a latent space perturbation that uses contrastive loss to shift the predicted noise closer to the target frames while pushing it away from the original frames. Specifically, given the latent representation of the original frames, $\\mathbf{X}_0$ , and the target latent frames, $\\mathbf{X}_{tg}$ , we construct a contrastive objective [8, 10], as shown in Eq. (3).", + "bbox": [ + 496, + 554, + 893, + 736 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {L} _ {c o n} = \\left\\| \\hat {\\mathbf {X}} _ {0, a d v} - \\mathbf {X} _ {t g} \\right\\| ^ {2} + \\max \\left(0, \\tau_ {1} - \\left\\| \\hat {\\mathbf {X}} _ {0, a d v} - \\mathbf {X} _ {0} \\right\\| ^ {2}\\right) (3)\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 742, + 911, + 766 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\tau_{1}$ is the margin distance. The positive pair consists of the predicted frames $\\hat{\\mathbf{X}}_{0,adv}$ and the target frames $\\mathbf{X}_{tg}$ , which we aim to bring closer together, while the negative pair is composed of the predicted frames $\\hat{\\mathbf{X}}_{0,adv}$ and the original frames $\\mathbf{X}_0$ , which we aim to push apart.", + "bbox": [ + 496, + 771, + 890, + 847 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With the hyperparameters $\\alpha, \\beta, \\gamma$ , we define the final training objective as:", + "bbox": [ + 496, + 845, + 890, + 876 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\left\\| I _ {a d v} - I _ {s r c} \\right\\| ^ {2} + \\alpha \\cdot \\mathcal {L} _ {e n c} + \\beta \\cdot \\mathcal {L} _ {c o n} + \\gamma \\cdot \\mathcal {L} _ {a t t} (4)\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 883, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12598", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Adversarial Attack on Image-to-Video Generation" + ], + "code_body": "Input: Source Image $I_{src}$ and Trainable Copy $I_{adv}$ , Target \nImage $I_{tg}$ , Target Video $V_{tg}$ , Image-to-Video Pipeline $\\mathcal{P}$ , \nGenerator $\\mathcal{G}$ , Encoder $\\mathcal{E}$ , Diffusion Scheduler $\\mathcal{S}$ \n(Optional:Condition $c$ ) \nHyperparameters $\\tau_1, \\tau_2, \\alpha, \\beta, \\gamma, \\lambda$ \nEncode image $x_{src}, x_{tg} = \\mathcal{E}(I_{src}), \\mathcal{E}(I_{tg})$ \nGenerate original video $V_0 = \\mathcal{P}(I_{src})$ \nEncode frames $\\mathbf{X}_0, \\mathbf{X}_{tg} = \\mathcal{E}(V_0), \\mathcal{E}(V_{tg})$ \nfor each iteration do \n Encode attacked image $x_{adv} = \\mathcal{E}(I_{adv})$ \n Compute encoder loss: $\\mathcal{L}_{enc} = ||x_{adv} - x_{tg}||^2$ \n Generate noisy frames $\\mathbf{X}_t = \\mathcal{S}(\\mathbf{X}_0)$ \nPredict $\\mathbf{X}_0\\left\\{\\begin{array}{l}\\hat{\\mathbf{X}}_{0,adv}=\\varnothing(\\mathbf{X}_t,x_{adv},c)\\\\ \\hat{\\mathbf{X}}_{0,src}=\\varnothing(\\mathbf{X}_t,x_{src},c)\\\\ A_{src}\\end{array}\\right.$ \nCompute spatial contrastive loss", + "bbox": [ + 91, + 109, + 472, + 358 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c o n} = \\left| \\left| \\hat {\\mathbf {X}} _ {0, a d v} - \\mathbf {X} _ {t g} \\right| \\right| ^ {2} + \\max \\left(0, \\tau_ {1} - \\left| \\left| \\hat {\\mathbf {X}} _ {0, a d v} - \\mathbf {X} _ {0} \\right| \\right| ^ {2}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 369, + 478, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Compute temporal attention loss", + "bbox": [ + 109, + 402, + 326, + 417 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {a t t} = \\tau_ {2} - | | A _ {a d v} - A _ {s r c} | | ^ {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 417, + 385, + 434 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Compute the final loss", + "bbox": [ + 109, + 434, + 261, + 448 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\left\\| I _ {a d v} - I _ {s r c} \\right\\| ^ {2} + \\alpha \\cdot \\mathcal {L} _ {e n c} + \\beta \\cdot \\mathcal {L} _ {c o n} + \\gamma \\cdot \\mathcal {L} _ {a t t}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 448, + 464, + 464 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Update parameters $I_{adv} \\gets I_{adv} - \\lambda \\cdot \\nabla_{I_{adv}}\\mathcal{L}$", + "bbox": [ + 109, + 464, + 418, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "end for", + "text_level": 1, + "bbox": [ + 94, + 479, + 148, + 492 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The complete algorithm is presented in Algorithm 1.", + "bbox": [ + 76, + 530, + 423, + 545 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "It is noteworthy that our method supports both global and local attacks, allowing perturbations to be applied across the entire image or to specific objects within the image. This feature can be implemented by applying a mask to the noise during the attack process.", + "bbox": [ + 75, + 547, + 468, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 648, + 209, + 666 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 76, + 676, + 282, + 694 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Models In the field of video generation, there are few open-source video diffusion models available for experimentation. Stable Video Diffusion [4] (SVD) is a widely used and effective image-to-video diffusion model, which we adopt as the baseline model for our attack and experimental analysis. For other conditional image-to-video models, we employ CogVideoX-5B-I2V [53] for text-image joint conditional video generation. The open-sourced CogVideoX adopts DiT rather than UNet architecture used in SVD, providing enhanced capabilities. We also employ ControlNeXt-SVD-v2 [34] for pose-guided image-to-video generation, which is built upon SVD and provides enhanced control over character animation.", + "bbox": [ + 75, + 703, + 472, + 898 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Data To our knowledge, there are currently no established benchmarks or datasets for image-based adversarial attacks in video generation. In previous work, PhotoGuard [37] reported attack results on over 60 images, while PRIME [28] used an internal dataset containing 35 video clips. In our experiments, we collect a dataset of 300 images and their corresponding generated frames. The selected images and frames depict people and animals with dynamic actions, which could potentially be misused. We exclude images of scenery without main subjects or images that do not yield noticeable temporal movement, as these typically do not produce meaningful motion, rely mainly on camera movements, and may even remain static.", + "bbox": [ + 496, + 90, + 893, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics We validate our experiments from two perspectives: image perturbation and generation effects. Our goal is to introduce minimal perturbations to the images that significantly degrade the generated results. We use peak signal-to-noise ratio (PSNR), structural similarity (SSIM) metrics, and Fréchet Inception Distance (FID) to evaluate the similarity between the attacked and original images. For evaluating video generation results, we use the metrics Subject Consistency, Motion Smoothness, Aesthetic Quality, and Image Quality from VBench [25]. Subject Consistency measures the average cosine similarity of Dino [6] features between each frame and the first frame, assessing temporal coherence. Motion Smoothness calculates the mean absolute error between interpolated and dropped frames. Aesthetic Quality is assessed using the LAION aesthetic predictor [1], and Image Quality evaluates the presence of low-level distortions in generated video frames with MUSIQ [26].", + "bbox": [ + 496, + 305, + 893, + 563 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Setup We control the added noise strength by adjusting the number of training steps and the perturbation threshold at each step. For the loss function, we scale the hyperparameters $\\alpha, \\beta, \\gamma$ to ensure that each loss component remains on a similar scale. During training, we use a black image as the target image $I_{tg}$ and black frames as the target video $V_{tg}$ . For the attention hooks, we focus on the temporal attention block of the last downsampling blocks in the UNet and the middle blocks in the transformer, separately. In terms of computational cost, the SVD model requires 68GB of GPU memory and approximately 2.5 minutes to apply protection, while CogVideoX demands 75GB of GPU memory and 10 minutes. Notably, this protecting time is comparable to the inference time of the I2V model.", + "bbox": [ + 496, + 580, + 893, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Qualitative Analysis", + "text_level": 1, + "bbox": [ + 500, + 801, + 691, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we provide an analysis of attacks on both unconditional and conditional image-to-video generation. Additionally, we present a robustness analysis in Appendix A to demonstrate that our method remains effective in safeguarding images even in the presence of various pre-processing", + "bbox": [ + 496, + 825, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12599", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d7977c29cd41f64b7019d4b8e6088b090b04ddc35f4ae6c64abb0e25a4cd67c7.jpg", + "image_caption": [ + "Original Image" + ], + "image_footnote": [], + "bbox": [ + 96, + 87, + 251, + 155 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a4fb6bbe0037b7b8b327b6811e76664207389d21f521b46ab421dd98efcc7921.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 256, + 88, + 410, + 155 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c331ad1d0f66ce7d87647755fe634be27ff32f03c513646f86c722058429b364.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 410, + 88, + 562, + 155 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4a1f2d50fa091027743d94819fb2e71f1b970a892498221ce1f524f6203b91ff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 562, + 88, + 715, + 155 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a4dd6abbbe134ab289634669e13450abde4a4e2674735a435d8aaccc9cadba67.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 715, + 88, + 867, + 155 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8ece16884112bdb351a6dcd4f04fdcda01708404499381bcbb0ffef0985b5963.jpg", + "image_caption": [ + "Random Noisy Image" + ], + "image_footnote": [], + "bbox": [ + 96, + 166, + 250, + 233 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d36921865af717dfff9a35729d82395e28a3f330892f001b0a8e5f89bcc07168.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 256, + 166, + 410, + 233 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/544778c99c3d7ee5e93e420090969dd962bbc5c13bce8f920311aca932993764.jpg", + "image_caption": [ + "Generated video from original image.", + "Generated video from random noisy image." + ], + "image_footnote": [], + "bbox": [ + 410, + 166, + 562, + 233 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/034c98d6a3d5e62c5f75fa9197cd2368f33aca6c071bd7bd3867777f4b2ad68d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 562, + 166, + 715, + 233 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/75d42566d1ab83cf6009bad08e0d2e14427e7ef53c6ff6cefd0586e21e2e6a6f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 715, + 166, + 867, + 233 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7cc2f8cf26a202fdfdcdce5e9dbd1db035d81f575b2484cc9beac805f9677c8c.jpg", + "image_caption": [ + "PhotoGuard Image" + ], + "image_footnote": [], + "bbox": [ + 96, + 243, + 251, + 311 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1664cf2b2afd8897c1b648bc758281348d2de6e91ae223cb3cceda866c6fded7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 256, + 243, + 410, + 310 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ae9f8e1846b2ec7770ca1f62e524ebc1841933e36e94754fb2f236db612c9fd9.jpg", + "image_caption": [ + "Generated video from PhotoGuard protected image." + ], + "image_footnote": [], + "bbox": [ + 410, + 243, + 562, + 310 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9858540f0f28139820e3ff871239227d880309f123f1b69df22f62dc7b0b3061.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 562, + 243, + 715, + 310 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1603d02838ae62d4efe1ceccb3c1277c1b93c7a53a93d67c202f7f6abbeff186.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 715, + 243, + 867, + 310 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a24e2af1eee20054f2fe7ac1050a06f2d21b1e5d0a95a714d8b14008ed8dcd03.jpg", + "image_caption": [ + "Our Guarded Image" + ], + "image_footnote": [], + "bbox": [ + 96, + 321, + 251, + 388 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/18602c149ff527e0e2199c92e4839924b9d51c5ffb02b5d04776225dc8ceebba.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 256, + 321, + 410, + 388 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/87302b099cdd7cb6f531ecedadfa7b0e662567c93f62bbec2ab49cbf963358d2.jpg", + "image_caption": [ + "Generated video from guarded image." + ], + "image_footnote": [], + "bbox": [ + 410, + 321, + 562, + 388 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f934ebd68b15913493d4a25e030f4858e8a2f5309516fc04e8d785d2dc7ef3fd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 562, + 321, + 715, + 388 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2fdf630e4b34a76095a897010080c7a836f3ef0ec746e3e7dc5152b3bdfdd93f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 715, + 321, + 867, + 388 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cf940a918cddfdb9ef5c5bd97d1814e8f2b440110161aba692b2399b84aa3b86.jpg", + "image_caption": [ + "(a) SVD generation results on original image, random noisy image, PhotoGuard image and our I2VGuard image.", + "Original Image" + ], + "image_footnote": [], + "bbox": [ + 98, + 415, + 251, + 494 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/dd1f4407ed07243aa6586466079aef2f196f60c2b413199eb026c04978326cf6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 256, + 415, + 410, + 494 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f748f1b0a2b25fb25e5006566e2478901b42c9d9e09a7c55d86314d53aa8d3f7.jpg", + "image_caption": [ + "Generated video from original image." + ], + "image_footnote": [], + "bbox": [ + 410, + 415, + 562, + 494 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8010378a9eeccfccf47584822d523e1099125d409e1a5cb3d3cad9baafb70257.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 562, + 415, + 715, + 494 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/56560c5f12c5c4d8e05217fb08870e3f00252483ed16c8bfea762e01b4b07770.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 715, + 415, + 870, + 494 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/86d28d43bf444ee4d9e5e3651dee275a37ae6f94735076c1cc5c99f3ba82ad34.jpg", + "image_caption": [ + "Guarded Image", + "(b) CogVideoX generation results, where we do not set the prompt in the process of video generation.", + "Figure 4. Qualitative results of adversarial attacks on I2V models SVD and CogVideoX. We also include generation results of SVD with random noise and PhotoGuard [37] perturbations for comparison. All generation results are using the same seed." + ], + "image_footnote": [], + "bbox": [ + 98, + 505, + 251, + 584 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7932a5e193773c4df1123d494e7b48d51e1766a469e96ab99593e4d7c32ff534.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 256, + 505, + 410, + 584 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5b5d05439d1fdba01850a3e5637cc39619d82b669b2db3cfe1eca44ae79fa5af.jpg", + "image_caption": [ + "Generated video from guarded image." + ], + "image_footnote": [], + "bbox": [ + 410, + 505, + 562, + 584 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/eb4cb7ed4b319d75d089214454e4ef47b5d9fa66ece7c9b3f788b9823281ccd5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 562, + 505, + 715, + 584 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9f1fc6d2aab65ed05c37ee2c37f2a1b11bc282b182439cc9d5f0db7d9a430a79.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 715, + 505, + 870, + 584 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "techniques and denoising modules [7]. A detailed visualization of the generation effects is provided in Appendix B. Furthermore, an additional comparison between spatial and temporal attacks is presented in Appendix E.", + "bbox": [ + 75, + 680, + 470, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Attack on Image-to-Video Generation We analyze the generation results using two image-to-video generation models: Stable Video Diffusion (SVD) and CogVideoX, as shown in Fig. 4. We observe that the noise added to the guarded image is nearly invisible, yet it causes significant variations in the generated results. In the example shown in Fig. 4a, we additionally illustrate the generated results for both a random noisy image and an image protected by PhotoGuard [37]. The random noise does not affect generation quality, preserv", + "bbox": [ + 75, + 763, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ing both temporal consistency and normal spatial appearance, likely because the VAE filters out such random noise. The PhotoGuard-protected image, however, introduces noise specifically on the person, which slightly affects appearance without compromising temporal consistency. In contrast, our protected image disrupts both spatial content and temporal consistency. Specifically, the woman's head and hair movement appear unnatural, chaotic frames are generated, and unusual textures emerge in the background across all frames. Fig. 4b shows results from CogVideoX under an image-only setting, where no prompt is applied during training and inference. CogVideoX, which is relatively more powerful than SVD, requires relatively stronger noise to affect generation. It can be observed that motion distortions still occur, particu", + "bbox": [ + 496, + 680, + 895, + 893 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12600", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fea02e2f23a4f2b1a73ebe9f0b262fb516bb5a0a5108cd9fbf57adb41d6c27bd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 98, + 87, + 194, + 213 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d2bc0bd02264c0dc82f7a1b2bdad43fc64e3b0f185263a45bed877303d46e498.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 200, + 88, + 295, + 213 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/732b0f92e57e636fc706470137ec74f3a8c27fb4669f6bca3500de7ede1d18f3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 88, + 388, + 213 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/577e63b87b31d2475ca66a4a58fba255ec5c6fe9f43237f8a43b9e63b4056a91.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 392, + 88, + 483, + 213 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b6a98c6ebf9340adf23267014e7b630189296f4c83d3ed69ec3e4228beecec74.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 88, + 578, + 213 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/456e4ccc5b6b1ff8a497e3e1770975c864bbcce35ab5da760030bedd0e296fdb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 581, + 88, + 673, + 213 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9df7b0069d470fc6724ee4dbabf31307c3b7e5975060a1e22c824df1dba5fdbd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 674, + 88, + 767, + 213 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/57bfa7dfaf8144b79ed544440db9dc16afdd4a99e0b874657c086eaea9492029.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 769, + 88, + 862, + 213 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7565a6bd8f58a633fbca595c789a6fc205452c48296bbf1609224ec9e29364b1.jpg", + "image_caption": [ + "Original Image", + "Guarded Image" + ], + "image_footnote": [], + "bbox": [ + 98, + 224, + 192, + 349 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7607e9e47dae44923b471dc0261a80c6b1cfc50094c6cffa795291c39e783b47.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 200, + 224, + 295, + 349 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e1c541087c0921f070ac44991ec1ff560968cbfeb9203b93424887c4554e5ce3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 224, + 388, + 349 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/be1772b4501dc7527efc66417c60b9af54a1f296e9d10abb0e0e33d61512dc3a.jpg", + "image_caption": [ + "Generated video from original image.", + "Generated video from guarded image." + ], + "image_footnote": [], + "bbox": [ + 392, + 224, + 483, + 349 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ecac280cc5a08bf8914a7a47becaec022edf05a9f50aaf1a1a4c136349fc74f5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 224, + 578, + 349 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9033dddd4430a3cd70af6e05ea3f58a21a67c7acd2fbc098fe74192bac60df85.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 581, + 224, + 673, + 349 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/888d2a4950889f5ae6202b6c54ab6235961a0767023d2301e5cc538847f17451.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 674, + 224, + 767, + 349 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/62f0ebbb7f90847cb2a7ff35f058b59634eb1a8fdaf4a065f309f0fbbb3cafd0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 769, + 224, + 862, + 349 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/be2f65d648a05abff5c50122b2d91e9a8290c5072e6c5fa27f5d03301c2bd6d8.jpg", + "image_caption": [ + "Original Image" + ], + "image_footnote": [], + "bbox": [ + 98, + 381, + 251, + 462 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/61e455ee5c51a98aeb1716703f2275f4bf2c84c1d31dba6adbdc29c1cc92acaa.jpg", + "image_caption": [ + "(a) ControlNeXT-SVD-v2 generation results." + ], + "image_footnote": [], + "bbox": [ + 256, + 381, + 411, + 462 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8bf15d03c40781b56d95285ef5d8ce4332f801cec2a8851e8cd393967d21f06b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 413, + 381, + 563, + 462 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/420fa9a3828b40ed26bc4906c75a0010c5c450f026c9c1e64bb36d8cb9d69905.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 565, + 381, + 717, + 462 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8e833d2beb9ba8daa3204eda0cf8ce4bc67df15a308989df51285d46427a4a2e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 718, + 381, + 870, + 462 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ceda9160e87cef3f31702cd15516442d97080d992bfb1249a8534d1fa106f864.jpg", + "image_caption": [ + "Guarded Image", + "Figure 5. Qualitative results of adversarial attacks on conditional I2V models ControlNeXt and CogVideoX. It can be observed that visual features are not controlled effectively, leading to unreasonable generation results. All generation results are using the same seed." + ], + "image_footnote": [], + "bbox": [ + 98, + 474, + 251, + 553 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2f61d492e5569eabeed4034d7ece6397a1d0dece5413fd50ec4f828007f664a3.jpg", + "image_caption": [ + "(b) CogVideoX generation results. The prompt is \"Many ostriches are running\"." + ], + "image_footnote": [], + "bbox": [ + 256, + 474, + 411, + 553 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/33299a55a3ebb1e9ecefcb3b3c7dcad3e67c69a4e8235d7bc810c4cf0dc5067a.jpg", + "image_caption": [ + "Generated video from original image.", + "Generated video from guarded image." + ], + "image_footnote": [], + "bbox": [ + 413, + 474, + 563, + 553 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/16af9bcc9dcaf58f21c379e5380342732a5ea835ef3edbfa9d506bad66a38d45.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 565, + 474, + 717, + 553 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fcaa90bd8af0b969a616a5409f9810deca6602db051438d5fa2c06ddf4e1a814.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 718, + 474, + 870, + 553 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "larly in areas where motion takes place, such as the blurred head movement of the man, with the noise causing the surrounding regions to fail in generation, resulting in a blur. In summary, our method effectively disrupts both spatial content and temporal consistency in generated videos.", + "bbox": [ + 75, + 648, + 470, + 724 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Attack on Conditional Image-to-Video Generation We present an attack on conditional image-to-video models in Fig. 5, specifically examining pose-conditioned ControlNeXt-SVD and text-guided video generation with CogVideoX. In the first row of Fig. 5a, the ControlNeXt model effectively controls the character's motion while maintaining consistent character appearance and background features. However, in the second row, where we guard the original image, the model fails to preserve the character's appearance and fidelity in the background. Noticeable differences", + "bbox": [ + 75, + 750, + 472, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "emerge: for instance, the fruits in the background exhibit blurring, and the character appears deformed. These spatial distortions illustrate the effectiveness of our attack method. Despite this, the pose control remains robust, likely due to the intensive training of the original SVD UNet, which results in strong overfitting to the specific style and hands shape. In the example of CogVideoX shown in Fig. 5b, we observe that adding special texture noise leads to the generation of temporal inconsistency and blurred content compared with the original ones. Specifically, in this case, texture noise is applied around the animals, causing distortions in the generated output, such as headless and transparent ostriches. In our experiments, CogVideoX exhibits strong prompt-following ability, which is due to an additional powerful text encoder compared to SVD. By slightly increasing the noise strength, we are able to successfully execute the attack.", + "bbox": [ + 496, + 648, + 893, + 891 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12601", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d8ba98d75481d9c04a135cd007bdeea3aa5a45760a66d8abb8cc9008a86ead48.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Video SourceModelSubject Consistency(%, ↓)Motion Smoothness(%, ↓)Aesthetic Quality (%, ↓)Image Quality(↓)
Original ImageSVD95.86±2.6297.90±1.4356.76±4.7567.28±6.18
Guarded Image91.57±3.9597.18±1.2153.42±4.9364.38±8.23
Original ImageCogVideoX97.02±1.9699.19±0.2759.94±5.5367.60±6.49
Guarded Image93.50±3.5897.97±0.3253.95±5.6265.24±9.85
", + "bbox": [ + 81, + 88, + 890, + 159 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6adacf430f021ac16e798edc8b8e2370239a2f1144260cf0180f7d0475f0552f.jpg", + "table_caption": [ + "Table 1. Analysis of video generation effects of SVD and CogVideoX from original images and images guarded by our method. Mean and variance of evaluations are reported. We exclude results with extremely high subject consistency and motion smoothness, as these indicate static frames, which are outside the scope of this evaluation. An ablation study on the different attack methods is also presented in ??" + ], + "table_footnote": [], + "table_body": "
Image TypePSNR(↑)SSIM(↑)FID(↓)
Random Noise26.22±0.040.774±0.00314.31
Guarded Image31.83±0.190.868±0.01616.78
", + "bbox": [ + 81, + 234, + 465, + 284 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2. Analysis of perturbations between original and modified images. Random noise is introduced to the original images to establish a baseline, and the mean and variance of the perturbations are reported.", + "bbox": [ + 75, + 294, + 470, + 351 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5909bcccf746fc3a7f012b0e7208f35b2256749cd4166ae92a75209139490e39.jpg", + "image_caption": [ + "Figure 6. Frame-level quantitative evaluation with the same seed." + ], + "image_footnote": [], + "bbox": [ + 80, + 364, + 467, + 477 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Quantitative Analysis", + "text_level": 1, + "bbox": [ + 76, + 531, + 279, + 547 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Image Perturbation Analysis We present the modified image quality results in Tab. 2. Random noise was added to the original images to establish a baseline for comparison. The mean and variance of the added noise in this baseline are generally comparable to the modifications introduced in our attacked images. The results demonstrate that, despite similar perturbation strength, our method achieves significantly higher image quality, as indicated by higher score in both peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM). A qualitative comparison reveals that the noise added to the image possesses its own texture, which is less visible than the random noise added. In terms of image quality and diversity, our approach yields a higher Fréchet Inception Distance (FID). This is expected because the specific texture noise we add aligns with the original image texture, resulting in lower diversity compared to images with added random noise.", + "bbox": [ + 75, + 561, + 470, + 818 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Video Generation Analysis Evaluation results for generated videos of SVD and CogVideoX are shown in Tab. 1. First, we assess the performance of the I2V model on original images, finding that it demonstrates strong subject consis", + "bbox": [ + 75, + 839, + 472, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tency, smooth motion transitions, high aesthetic quality, and well-preserved low-level image quality. This suggests that the I2V models are capable of generating visually coherent and temporally stable content. In particular, CogVideoX performs a better performance than SVD because transformers scale more easily in both data and model size than UNet, making CogVideoX superior to SVD in size and performance. In contrast, our method, which uses a similar noise strength, significantly impacts all evaluated metrics. To be precise, temporal assessments reveal declines around $4\\%$ in subject consistency and around $1\\%$ in motion smoothness, showing that our method effectively disrupts the model's temporal coherence. Specifically, the decline in subject consistency is reflected in irregular movement of the main object's location or the sporadic appearance and disappearance of the object. The decline in motion smoothness is reflected in abrupt, discontinuous frame-to-frame movements. Spatial evaluations also show that our method considerably diminishes aesthetic quality and low-level image quality by $3 \\sim 6\\%$ and $2 \\sim 3\\%$ , respectively. The decrease in aesthetic quality is evident in the appearance of unusual textures or object deformations that detract from the visual appeal. These reductions reflect a successful attack on the image-to-video model. Notably, because CogVideoX employs a 3D VAE, which handles spatiotemporal noise in videos better than the 2D+1D VAE in SVD, CogVideoX shows greater resistance to attacks compared to SVD.", + "bbox": [ + 496, + 238, + 893, + 645 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Additionally, we provide a frame-level quantitative evaluation in Fig. 6. The results indicate that our method effectively disrupts temporal consistency, leading to a propagated deviation from the original generation.", + "bbox": [ + 496, + 646, + 893, + 708 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 723, + 617, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we presented I2VGuard, a novel adversarial method that applies imperceptible perturbations on images to degrade the quality of videos generated by diffusion-based image-to-video models, thereby protecting images from misuse. Concretely, we designed three attack modules, i.e. spatial attack, temporal attack and diffusion attack, to disrupt visual fidelity, temporal consistency, and cross-model robustness, respectively. Extensive experiments on state-of-the-art generative models, i.e. CogVideoX and SVD, demonstrated the effectiveness of our method.", + "bbox": [ + 496, + 750, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12602", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] LAION AI. Laion-aesthetics. 5", + "[2] Ali Asghar, Amna Shifa, and Mamoona Naveed Asghar. Survey on video security: Examining threats, challenges, and future trends. Computers, Materials & Continua, 80(3), 2024. 2", + "[3] Omer Bar-Tal, Hila Chefer, Omer Tov, Charles Herrmann, Roni Paiss, Shiran Zada, Ariel Ephrat, Junhwa Hur, Yuanzhen Li, Tomer Michaeli, et al. Lumiere: A space-time diffusion model for video generation. arXiv preprint arXiv:2401.12945, 2024. 2", + "[4] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2, 4, 5", + "[5] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. 2024. 2", + "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 5", + "[7] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In European conference on computer vision, pages 17-33. Springer, 2022. 6", + "[8] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 4", + "[9] Weixin Chen, Dawn Song, and Bo Li. Trojdiff: Trojan attacks on diffusion models with diverse targets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4035-4044, 2023. 2", + "[10] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In IEEE Conference on Computer Vision and Pattern Recognition, 2021. Computer Vision Foundation / IEEE, 2021. 4", + "[11] June Suk Choi, Kyungmin Lee, Jongheon Jeong, Saining Xie, Jinwoo Shin, and Kimin Lee. Diffusionguard: A robust defense against malicious diffusion-based image editing. In ICML 2024 Next Generation of AI Safety Workshop. 2", + "[12] Sheng-Yen Chou, Pin-Yu Chen, and Tsung-Yi Ho. How to backdoor diffusion models? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4015-4024, 2023. 2", + "[13] Hai Ci, Yiren Song, Pei Yang, Jinheng Xie, and Mike Zheng Shou. Wmadapter: Adding watermark control to latent diffusion models. arXiv preprint arXiv:2406.08337, 2024. 2", + "[14] Özgün Çiçek, Ahmed Abdulkadir, Soeren S Lienkamp, Thomas Brox, and Olaf Ronneberger. 3d u-net: learning dense volumetric segmentation from sparse annotation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2016: 19th International Conference," + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Athens, Greece, October 17-21, 2016, Proceedings, Part II 19, pages 424-432. Springer, 2016. 2", + "[15] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2", + "[16] Weitao Feng, Wenbo Zhou, Jiyan He, Jie Zhang, Tianyi Wei, Guanlin Li, Tianwei Zhang, Weiming Zhang, and Nenghai Yu. Aqualora: Toward white-box protection for customized stable diffusion models via watermark lora. arXiv preprint arXiv:2405.11135, 2024. 2", + "[17] Pierre Fernandez, Guillaume Couairon, Hervé Jégou, Matthijs Douze, and Teddy Furon. The stable signature: Rooting watermarks in latent diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22466-22477, 2023. 2", + "[18] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Imagine your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 2", + "[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2", + "[20] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsanko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2", + "[21] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 2", + "[22] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 2", + "[23] Robert Honig, Javier Rando, Nicholas Carlini, and Florian Tramér. Adversarial perturbations cannot reliably protect artists from generative ai. arXiv preprint arXiv:2406.12027, 2024. 3", + "[24] Li Hu. Animate anyone: Consistent and controllable image-to-video synthesis for character animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8153-8163, 2024. 2", + "[25] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024. 5", + "[26] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5148-5157, 2021. 5", + "[27] PKU-Yuan Lab and Tuzhan AI etc. Open-sora-plan, 2024. 2", + "[28] Guanlin Li, Shuai Yang, Jie Zhang, and Tianwei Zhang. Prime: Protect your videos from malicious editing. arXiv preprint arXiv:2402.01239, 2024. 3, 5" + ], + "bbox": [ + 503, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12603", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Chumeng Liang and Xiaoyu Wu. Mist: Towards improved adversarial examples for diffusion models. arXiv preprint arXiv:2305.12683, 2023. 2", + "[30] Chumeng Liang, Xiaoyu Wu, Yang Hua, Jiaru Zhang, Yiming Xue, Tao Song, Zhengui Xue, Ruhui Ma, and Haibing Guan. Adversarial example does good: Preventing painting imitation from diffusion models via adversarial examples. arXiv preprint arXiv:2302.04578, 2023. 2", + "[31] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International conference on machine learning, pages 8162-8171. PMLR, 2021. 2", + "[32] Josephine Passananti, Stanley Wu, Shawn Shan, Haitao Zheng, and Ben Y Zhao. Disrupting style mimicry attacks on video imagery. arXiv preprint arXiv:2405.06865, 2024. 3", + "[33] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 2", + "[34] Bohao Peng, Jian Wang, Yuechen Zhang, Wenbo Li, MingChang Yang, and Jiaya Jia. Controlnext: Powerful and efficient control for image and video generation. arXiv preprint arXiv:2408.06070, 2024. 2, 4, 5", + "[35] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2", + "[36] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 2", + "[37] Hadi Salman, Alaa Khaddaj, Guillaume Leclerc, Andrew Ilyas, and Aleksander Madry. Raising the cost of malicious ai-powered image editing. In Proceedings of the 40th International Conference on Machine Learning, pages 29894–29918, 2023. 2, 4, 5, 6", + "[38] Shawn Shan, Jenna Cryan, Emily Wenger, Haitao Zheng, Rana Hanocka, and Ben Y Zhao. Glaze: Protecting artists from style mimicry by {Text-to-Image} models. In 32nd USENIX Security Symposium (USENIX Security 23), pages 2187–2204, 2023. 2", + "[39] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2", + "[40] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 2", + "[41] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2", + "[42] Jingwei Sun, Xuchong Zhang, Changfeng Sun, Qicheng Bai, and Hongbin Sun. Latent feature and attention dual erasure" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "attack against multi-view diffusion models for 3d assets protection. arXiv preprint arXiv:2408.11408, 2024. 3", + "[43] Kling AI Team. Kling ai, 2024. 2", + "[44] Luma AI Team. Luma ai, 2024.", + "[45] The Movie Gen team @ Meta. Movie gen: A cast of media foundation models. 2024. 2", + "[46] Vu Tuan Truong, Luan Ba Dang, and Long Bao Le. Attacks and defenses for generative diffusion models: A comprehensive survey. arXiv preprint arXiv:2408.03400, 2024. 2", + "[47] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 2", + "[48] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 2", + "[49] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 2", + "[50] Yuxin Wen, John Kirchenbauer, Jonas Geiping, and Tom Goldstein. Tree-ring watermarks: Fingerprints for diffusion images that are invisible and robust. arXiv preprint arXiv:2305.20030, 2023. 2", + "[51] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7623-7633, 2023. 2", + "[52] Zhen Xing, Qijun Feng, Haoran Chen, Qi Dai, Han Hu, Hang Xu, Zuxuan Wu, and Yu-Gang Jiang. A survey on video diffusion models. ACM Computing Surveys, 2023. 2", + "[53] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 2, 4, 5", + "[54] Jianping Zhang, Zhuoer Xu, Shiwen Cui, Changhua Meng, Weibin Wu, and Michael R Lyu. On the robustness of latent diffusion models. arXiv preprint arXiv:2306.08257, 2023. 2, 4", + "[55] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all, 2024. 2", + "[56] Peifei Zhu, Tsubasa Takahashi, and Hirokatsu Kataoka. Watermark-embedded adversarial examples for copyright protection against diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24420–24430, 2024. 2" + ], + "bbox": [ + 501, + 92, + 893, + 801 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12604", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/I2VGuard_ Safeguarding Images against Misuse in Diffusion-based Image-to-Video Models/612ebd29-cda5-46e3-b035-7c7393e6f819_model.json b/2025/I2VGuard_ Safeguarding Images against Misuse in Diffusion-based Image-to-Video Models/612ebd29-cda5-46e3-b035-7c7393e6f819_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a9a7364f903b4900698ad3c03710bd50f576c769 --- /dev/null +++ b/2025/I2VGuard_ Safeguarding Images against Misuse in Diffusion-based Image-to-Video Models/612ebd29-cda5-46e3-b035-7c7393e6f819_model.json @@ -0,0 +1,3047 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.148, + 0.131, + 0.825, + 0.177 + ], + "angle": 0, + "content": "I2VGuard: Safeguarding Images against Misuse in Diffusion-based Image-to-Video Models" + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.203, + 0.351, + 0.221 + ], + "angle": 0, + "content": "Dongnan Gui\\(^{1,*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.385, + 0.204, + 0.468, + 0.221 + ], + "angle": 0, + "content": "Xun Guo2" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.204, + 0.638, + 0.221 + ], + "angle": 0, + "content": "Wengang Zhou" + }, + { + "type": "text", + "bbox": [ + 0.679, + 0.204, + 0.747, + 0.221 + ], + "angle": 0, + "content": "Yan Lu2" + }, + { + "type": "text", + "bbox": [ + 0.172, + 0.221, + 0.554, + 0.24 + ], + "angle": 0, + "content": "1University of Science and Technology of China" + }, + { + "type": "text", + "bbox": [ + 0.594, + 0.222, + 0.8, + 0.238 + ], + "angle": 0, + "content": "2Microsoft Research Asia" + }, + { + "type": "text", + "bbox": [ + 0.207, + 0.242, + 0.76, + 0.257 + ], + "angle": 0, + "content": "{gdn2001@mail.,zhwg@}ustc.edu.cn{xunguo，yanlu}@microsoft.com" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.273, + 0.245, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.121, + 0.344, + 0.207, + 0.354 + ], + "angle": 0, + "content": "Original Image" + }, + { + "type": "image", + "bbox": [ + 0.249, + 0.273, + 0.41, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.41, + 0.273, + 0.569, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.568, + 0.273, + 0.728, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.728, + 0.273, + 0.888, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.464, + 0.345, + 0.675, + 0.355 + ], + "angle": 0, + "content": "Generated video from original image." + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.354, + 0.245, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.123, + 0.424, + 0.212, + 0.435 + ], + "angle": 0, + "content": "Guarded Image" + }, + { + "type": "image", + "bbox": [ + 0.249, + 0.354, + 0.41, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.41, + 0.354, + 0.569, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.568, + 0.354, + 0.728, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.728, + 0.354, + 0.888, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.461, + 0.424, + 0.678, + 0.435 + ], + "angle": 0, + "content": "Generated video from guarded image." + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.441, + 0.245, + 0.512 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.121, + 0.511, + 0.206, + 0.522 + ], + "angle": 0, + "content": "Original Image" + }, + { + "type": "image", + "bbox": [ + 0.249, + 0.44, + 0.41, + 0.511 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.41, + 0.44, + 0.569, + 0.51 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.568, + 0.44, + 0.728, + 0.51 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.728, + 0.44, + 0.888, + 0.51 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.461, + 0.511, + 0.673, + 0.521 + ], + "angle": 0, + "content": "Generated video from original image." + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.522, + 0.245, + 0.592 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.123, + 0.592, + 0.212, + 0.603 + ], + "angle": 0, + "content": "Guarded Image" + }, + { + "type": "image", + "bbox": [ + 0.249, + 0.522, + 0.41, + 0.592 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.41, + 0.522, + 0.569, + 0.592 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.568, + 0.522, + 0.728, + 0.592 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.728, + 0.522, + 0.888, + 0.592 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.461, + 0.592, + 0.678, + 0.603 + ], + "angle": 0, + "content": "Generated video from guarded image." + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.614, + 0.895, + 0.644 + ], + "angle": 0, + "content": "Figure 1. Results of our I2VGuard. We present original images, guarded images, and their corresponding SVD-generated videos. All results are generated with the same seed. Our method effectively safeguards images from animation in image-to-video generation." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.656, + 0.314, + 0.672 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.689, + 0.475, + 0.871 + ], + "angle": 0, + "content": "Recent advances in image-to-video generation have enabled animation of still images and offered pixel-level controllability. While these models hold great potential to transform single images into vivid and dynamic videos, they also carry risks of misuse that could impact privacy, security, and copyright protection. This paper proposes a novel approach that applies imperceptible perturbations on images to degrade the quality of the generated videos, thereby protecting images from misuse in white-box image-to-video diffusion models. Specifically, we function our approach as an adversarial attack, incorporating spatial, temporal, and diffusion attack modules. The spatial attack shifts image features from" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.657, + 0.896, + 0.87 + ], + "angle": 0, + "content": "their original distribution to a lower-quality target distribution, reducing visual fidelity. The temporal attack disrupts coherent motion by interfering with temporal attention maps that guide motion generation. To enhance the robustness of our approach across different models, we further propose a diffusion attack module leveraging contrastive loss. Our approach can be easily integrated with mainstream diffusion-based I2V models. Extensive experiments on SVD, CogVideoX, and ControlNeXt demonstrate that our method significantly impairs generation quality in terms of visual clarity and motion consistency, while introducing only minimal artifacts to the images. To the best of our knowledge, we are the first to explore adversarial attacks on image-to-video generation for security purposes." + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.887, + 0.388, + 0.902 + ], + "angle": 0, + "content": "*Work done during internship at Microsoft Research Asia." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "12595" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.081, + 0.091, + 0.208, + 0.105 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.116, + 0.47, + 0.356 + ], + "angle": 0, + "content": "Diffusion models [15, 19, 35] have gained significant attention recently, particularly in video generation. Advances in models like Sora [5] and its DiT [33] framework have accelerated the development of video diffusion models, leading to numerous academic [4, 22, 27, 53, 55] and commercial [43-45] applications. The primary types of video generation are text-to-video(T2V) and image-to-video(I2V), with the former supporting greater diversity and the latter offering pixel-level control. Leveraging powerful generative models and techniques from text-to-image generation [35], image-to-video models can now produce visually impressive videos. Other conditional I2V models include CogVideoX [53], which generates videos using both text and image inputs, enabling more versatile content creation, and ControlNeXt [34], which focuses on generating videos conditioned on pose information, offering precise control over motion and structure." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.358, + 0.47, + 0.583 + ], + "angle": 0, + "content": "However, the rapid advancement of generative models has also heightened risks related to privacy, security, and copyright [2]. For instance, an adversary with access to a person's image could use image-to-video models to generate misleading or harmful videos, depicting actions the individual never performed, thus compromising individual's privacy and security. Additionally, the unauthorized use of copyrighted images to create animations poses significant risks to intellectual property rights. To address these risks, this work introduces an adversarial attack method designed to prevent the misuse of images in diffusion-based I2V models. Prior research has primarily focused on protecting images from malicious edits [37, 54], which are often viewed as adversarial attacks on image-editing models. Few efforts have focused on safeguarding images from misuse in I2V tasks." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.585, + 0.47, + 0.899 + ], + "angle": 0, + "content": "The main challenges in attacking I2V models come from two aspects: first, disrupting temporal consistency with perturbations applied only to individual images; and second, achieving robustness across diverse diffusion models, including UNet and DiT architectures, even under strong conditioning such as poses and text prompts. To address these challenges, our method targets different modules within I2V models and affects various attributes of the generated video. Specifically, we divide our approach into spatial, temporal, and diffusion attacks. The spatial attack focuses on altering the spatial features of the I2V models by targeting the VAE encoder in latent diffusion models. Here, we manipulate image latents, shifting them toward suboptimal values. The temporal attack aims to perturb the temporal features, which are crucial for video motion and consistency. By extracting the temporal attention map, we intentionally disrupt it by pushing it away from its original attention map. This creates a chaotic attention map, leading to temporal inconsistencies in the generated video. The diffusion attack focuses on the diffusion process and targets the output of the denoising module, which includes the UNet in SVD and the trans" + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.197 + ], + "angle": 0, + "content": "former in CogVideoX. Unlike a standard denoising process, our approach aims to push the predicted outputs away from the original frames, guiding them closer to target frames. To achieve this, we design a contrastive loss that degrades the quality of the predicted frames by treating the target frames as positive samples and the original predicted frames as negative samples." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.199, + 0.892, + 0.347 + ], + "angle": 0, + "content": "We conduct a comprehensive qualitative and quantitative analysis of the attack results. Qualitative analysis demonstrates that our method effectively disrupts the original image, leading to low temporal consistency and reduced aesthetic quality in both simple and conditioned I2V models. Quantitative analysis confirms a significant degradation in both spatial quality and temporal consistency of the generated videos. To our knowledge, this is the first adversarial attack method aimed at securing images against misuse in I2V generation." + }, + { + "type": "title", + "bbox": [ + 0.504, + 0.361, + 0.64, + 0.375 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.386, + 0.892, + 0.655 + ], + "angle": 0, + "content": "Image-to-Video Generation With the great advancements in image generation [15, 19, 31, 35, 40, 41], video generation [3, 5, 20, 22, 39, 43–45, 52, 53] has gathered significant attention and is developing rapidly [18, 48, 49, 51]. Recently, following the success of diffusion models [15, 19, 35, 40] in image generation, diffusion models integrated with UNet [3, 14, 21, 36] or transformers (DiT) [5, 33, 47, 55] have been adopted for video generation. In the field of imaged-to-video generation,AnimateDiff [18] animates personalized text-to-image diffusion models by inserting a temporal motion module, Stable Video Diffusion(SVD) [4] demonstrates strong performance, benefiting from intensive training on large amounts of high-quality data, CogVideoX [53] supports text-image jointly conditional video generation,Animate Anyone [24] and ControlNeXt [34] leverage pose information for controllable image-to-video generation. In this work, we focus on I2V models, which are susceptible to misuse and therefore require robust protection." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.675, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Adversarial Attacks on Diffusion Models Current attacks on diffusion models, which can be broadly categorized [46] into backdoor [9, 12] and adversarial attacks, primarily target image-based diffusion models [13, 16, 17, 50]. In adversarial attacks, the input image or text prompt is perturbed to generate adversarial examples. For instance, AdvDM [30] and Mist [29] utilize adversarial examples to protect human-created artworks, PhotoGuard [37] alters the image by modifying either the encoder or the diffusion outputs against malicious editing, while Glaze [38] perturbs the image to prevent style mimicry, [54] shifts the image away from its original distribution, [56] embeds personal watermarks in the generation of adversarial examples, and DiffusionGuard [11] generates adversarial noise targeting the early stage of the diffusion process. For video-based attacks," + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.947, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12596" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.096, + 0.102, + 0.881, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.42, + 0.896, + 0.504 + ], + "angle": 0, + "content": "Figure 2. Overview of our proposed method. The training process begins with a trainable copy \\( I_{adv} \\) of the original image \\( I_{src} \\). First, we perform inference to obtain the original video \\( V_{0} \\) and get the latent frames \\( \\mathbf{X}_{0} \\). Noisy latent frames \\( \\mathbf{X}_{t} \\) and the latent image \\( x_{adv}, x_{src} \\) are then processed by the denoising model to predict the original frames. The encoded latent image \\( x_{adv} \\) is utilized for the spatial encoder attack, while the predicted original frames \\( \\hat{\\mathbf{X}}_{0} \\) serve in the diffusion attack to compute contrastive loss. Within the denoising module, we hook into the temporal attention module to extract the temporal attention map. By altering the attacked attention map \\( A_{adv} \\) to diverge from the original \\( A_{src} \\), we implement the temporal attack." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.53, + 0.473, + 0.742 + ], + "angle": 0, + "content": "PRIME [28] modifies videos frame-by-frame to prevent malicious video editing, while [32] disrupts style mimicry attacks from video imagery. There are also works that protect the image against other generation, such as text-to-3D generation [42]. In this work, we focus on perturbing the input image in image-to-video diffusion models to perform adversarial attacks. Recent study [23] show that adversarial perturbations fail to reliably protect artists from style mimicry in the field of image generation. Our approach stands out in two ways: it uses a white-box attack strategy, allowing adversarial noise training with both generation and pre-processing models, and it focuses on disrupting temporal consistency rather than mere image fidelity, making the perturbations harder to remove." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.766, + 0.169, + 0.782 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.903 + ], + "angle": 0, + "content": "In this section, we first provide preliminaries on latent diffusion models and problem settings in Sec. 3.1. We then present our attack method, detailing the spatial attack in Sec. 3.2, the temporal attack in Sec. 3.3, and the diffusion attack in Sec. 3.4. An overview of our method is presented in in Fig. 2, which shows our protection procedure involves spatial, temporal, and diffusion attacks." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.53, + 0.63, + 0.546 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.583, + 0.896, + 0.903 + ], + "angle": 0, + "content": "Latent Diffusion Models In diffusion models, realistic image or video sampling from a target distribution \\( q(\\cdot) \\) is achieved via a sequence of denoising steps in a diffusion process. This approach involves iteratively adding and removing noise to transform an initial sample \\( \\mathbf{x}_0 \\sim q(\\cdot) \\). Noise is incrementally introduced across steps, creating intermediate states defined by \\( \\mathbf{x}_{t+1} = \\sqrt{\\overline{\\alpha}_t} \\mathbf{x}_0 + \\sqrt{1 - \\overline{\\alpha}_t} \\epsilon_t \\), where \\( \\epsilon_t \\) is Gaussian noise. The parameters \\( \\alpha_t \\) control the noise transition, leading to a final noisy state with a standard Gaussian distribution. By reversing this process and iteratively denoising each state, the model learns to predict the noise added at each step, ultimately enabling the recovery of samples from \\( q(\\cdot) \\) starting from \\( \\mathbf{x}_T \\). Latent diffusion models (LDMs), a subclass of diffusion models, operate in a compressed latent space rather than directly in the image space. In LDMs, each input image or video \\( \\mathbf{x}_0 \\) is first mapped to a latent representation \\( \\mathcal{E}(\\mathbf{x}_0) \\) via an encoder \\( \\mathcal{E} \\). The diffusion process then applies in this latent space, producing a series of noisy representations. Training in this latent space enables LDMs to achieve efficient training and faster inference while retaining high-quality outputs." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12597" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.258 + ], + "angle": 0, + "content": "Problem Settings Adversarial attacks involve adding subtle, nearly undetectable perturbations \\(\\delta_{adv}\\) to the input data to influence a model's output. In computer vision, adversarial attacks have been widely explored in classifications and image synthesis. However, in image-to-video generation tasks, adversarial attack remains a relatively unexplored area. In this work, we aim to develop adversarial attacks to safeguard images from misused by image-to-video latent diffusion models. To distinguish between the latent and pixel spaces, we represent images and videos in pixel space as \\(I\\) and \\(V\\), and in latent space as \\(x\\) and \\(\\mathbf{X}\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.258, + 0.473, + 0.35 + ], + "angle": 0, + "content": "Specifically, an adversary can generate a modified image \\( I_{adv} = I_{src} + \\delta \\) from source image \\( I_{src} \\), crafted to disrupt the generative model \\( \\mathcal{G} \\), resulting in poor-quality or distorted outputs. This adversarial image is optimized by maximizing the loss between the videos generated from the original and modified images, represented as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.361, + 0.456, + 0.387 + ], + "angle": 0, + "content": "\\[\n\\delta_ {a d v} = \\arg \\max _ {\\| \\delta \\| _ {p} \\leq \\epsilon} \\mathcal {L} (\\mathcal {G} (I _ {s r c} + \\delta , \\mathbf {X} _ {t}, c), \\mathcal {G} (I _ {s r c}, \\mathbf {X} _ {t}, c))\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.399, + 0.473, + 0.476 + ], + "angle": 0, + "content": "where \\(\\delta_{adv}\\) represents the computed perturbation, constrained by \\(\\| \\delta \\| _p\\leq \\epsilon\\) to ensure minimal visibility, and \\(c\\) represents other conditions such as text or pose in the imaged-to-video generation process. The constraint on \\(\\delta_{adv}\\) is maintained by a regularization term \\(\\| I_{adv} - I_{src}\\| ^2\\)" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.484, + 0.226, + 0.5 + ], + "angle": 0, + "content": "3.2. Spatial Attack" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.508, + 0.47, + 0.599 + ], + "angle": 0, + "content": "The generated video has two main attributes: spatial features and temporal features. In this section, we focus on attacking the spatial features within individual frames. Previous work [37, 54] has applied adversarial attacks to image editing tasks, focusing solely on spatial features. Our work extends these methods to image-to-video generation task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.472, + 0.66 + ], + "angle": 0, + "content": "In Latent Diffusion Models (LDMs), an encoder, denoted as \\(\\mathcal{E}\\), compresses the original image or frames into a latent space representation. An encoder attack aims to compromise this encoder, reducing the effectiveness of generative models." + }, + { + "type": "equation", + "bbox": [ + 0.175, + 0.674, + 0.47, + 0.693 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {e n c} = \\left\\| \\mathcal {E} \\left(I _ {a d v}\\right) - \\mathcal {E} \\left(I _ {t g}\\right) \\right\\| ^ {2} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.701, + 0.47, + 0.746 + ], + "angle": 0, + "content": "Specifically, as shown in Eq. (1), the attack transforms the latent representation \\(\\mathcal{E}(I_{adv})\\) into a poor one \\(\\mathcal{E}(I_{tg})\\), resulting in poor-quality final videos." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.757, + 0.245, + 0.773 + ], + "angle": 0, + "content": "3.3. Temporal Attack" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Unlike attacks on image editing tasks, where no temporal features are involved, attacks on video generation rely heavily on temporal information. In this task, disrupting temporal consistency through adversarial attacks is both essential and challenging. To reduce temporal consistency and induce chaotic behavior in the outputs, we target the temporal attention maps within the temporal transformer blocks in video diffusion generator, \\(\\mathcal{G}\\). Specifically, during training, we insert" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.092, + 0.696, + 0.175 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.092, + 0.888, + 0.175 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.188, + 0.892, + 0.231 + ], + "angle": 0, + "content": "Figure 3. Temporal self-attention map visualizations. Difference between generated frames from the original image (left) and the guarded image (right) in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.253, + 0.894, + 0.381 + ], + "angle": 0, + "content": "a hook, denoted as \\(\\mathcal{G}\\), within the attention block to extract the self-attention map \\(A_{adv}\\). We then apply a targeted attack to this attention map, pushing it away from the original map, \\(A_{src}\\), thereby reducing temporal consistency in the generated outputs. To achieve this, we introduce a temporal attention loss, denoted as \\(\\mathcal{L}_{att}\\), as shown in Eq. (2), which induces motion-related perturbations and promotes temporal chaos." + }, + { + "type": "equation", + "bbox": [ + 0.596, + 0.382, + 0.892, + 0.4 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {a t t}} = \\tau_ {2} - \\left| \\left| A _ {\\text {a d v}} - A _ {\\text {s r c}} \\right| \\right| ^ {2} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.403, + 0.895, + 0.524 + ], + "angle": 0, + "content": "As shown in Fig. 3, the visualization of the temporal self-attention map on both the original and modified frames highlights this impact: the manipulated attention map misaligns with the original, directly leading to inconsistencies in motion across frames. This visual evidence underscores the effectiveness of our method in disrupting temporal coherence in video generation by distorting the temporal attention mechanism within the model." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.532, + 0.665, + 0.546 + ], + "angle": 0, + "content": "3.4. Diffusion Attack" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.555, + 0.895, + 0.737 + ], + "angle": 0, + "content": "Although spatial and temporal attacks can disrupt structural and motion information in generated videos, their effectiveness may vary significantly across different models [4, 34, 53], especially when other conditions, e.g., text prompts, are involved. To enhance the robustness and generalization of the perturbation across diffusion models, we introduce a latent space perturbation that uses contrastive loss to shift the predicted noise closer to the target frames while pushing it away from the original frames. Specifically, given the latent representation of the original frames, \\(\\mathbf{X}_0\\), and the target latent frames, \\(\\mathbf{X}_{tg}\\), we construct a contrastive objective [8, 10], as shown in Eq. (3)." + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.743, + 0.912, + 0.767 + ], + "angle": 0, + "content": "\\[\n\\mathbf {L} _ {c o n} = \\left\\| \\hat {\\mathbf {X}} _ {0, a d v} - \\mathbf {X} _ {t g} \\right\\| ^ {2} + \\max \\left(0, \\tau_ {1} - \\left\\| \\hat {\\mathbf {X}} _ {0, a d v} - \\mathbf {X} _ {0} \\right\\| ^ {2}\\right) (3)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.772, + 0.892, + 0.848 + ], + "angle": 0, + "content": "where \\(\\tau_{1}\\) is the margin distance. The positive pair consists of the predicted frames \\(\\hat{\\mathbf{X}}_{0,adv}\\) and the target frames \\(\\mathbf{X}_{tg}\\), which we aim to bring closer together, while the negative pair is composed of the predicted frames \\(\\hat{\\mathbf{X}}_{0,adv}\\) and the original frames \\(\\mathbf{X}_0\\), which we aim to push apart." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.847, + 0.892, + 0.877 + ], + "angle": 0, + "content": "With the hyperparameters \\(\\alpha, \\beta, \\gamma\\), we define the final training objective as:" + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.884, + 0.894, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\left\\| I _ {a d v} - I _ {s r c} \\right\\| ^ {2} + \\alpha \\cdot \\mathcal {L} _ {e n c} + \\beta \\cdot \\mathcal {L} _ {c o n} + \\gamma \\cdot \\mathcal {L} _ {a t t} (4)\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12598" + } + ], + [ + { + "type": "code_caption", + "bbox": [ + 0.079, + 0.092, + 0.466, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 Adversarial Attack on Image-to-Video Generation" + }, + { + "type": "algorithm", + "bbox": [ + 0.092, + 0.11, + 0.473, + 0.359 + ], + "angle": 0, + "content": "Input: Source Image \\(I_{src}\\) and Trainable Copy \\(I_{adv}\\), Target \nImage \\(I_{tg}\\), Target Video \\(V_{tg}\\), Image-to-Video Pipeline \\(\\mathcal{P}\\), \nGenerator \\(\\mathcal{G}\\), Encoder \\(\\mathcal{E}\\), Diffusion Scheduler \\(\\mathcal{S}\\) \n(Optional:Condition \\(c\\)) \nHyperparameters \\(\\tau_1, \\tau_2, \\alpha, \\beta, \\gamma, \\lambda\\) \nEncode image \\(x_{src}, x_{tg} = \\mathcal{E}(I_{src}), \\mathcal{E}(I_{tg})\\) \nGenerate original video \\(V_0 = \\mathcal{P}(I_{src})\\) \nEncode frames \\(\\mathbf{X}_0, \\mathbf{X}_{tg} = \\mathcal{E}(V_0), \\mathcal{E}(V_{tg})\\) \nfor each iteration do \n Encode attacked image \\(x_{adv} = \\mathcal{E}(I_{adv})\\) \n Compute encoder loss: \\(\\mathcal{L}_{enc} = ||x_{adv} - x_{tg}||^2\\) \n Generate noisy frames \\(\\mathbf{X}_t = \\mathcal{S}(\\mathbf{X}_0)\\) \nPredict \\(\\mathbf{X}_0\\left\\{\\begin{array}{l}\\hat{\\mathbf{X}}_{0,adv}=\\varnothing(\\mathbf{X}_t,x_{adv},c)\\\\ \\hat{\\mathbf{X}}_{0,src}=\\varnothing(\\mathbf{X}_t,x_{src},c)\\\\ A_{src}\\end{array}\\right.\\) \nCompute spatial contrastive loss" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.371, + 0.48, + 0.395 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c o n} = \\left| \\left| \\hat {\\mathbf {X}} _ {0, a d v} - \\mathbf {X} _ {t g} \\right| \\right| ^ {2} + \\max \\left(0, \\tau_ {1} - \\left| \\left| \\hat {\\mathbf {X}} _ {0, a d v} - \\mathbf {X} _ {0} \\right| \\right| ^ {2}\\right)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.404, + 0.327, + 0.419 + ], + "angle": 0, + "content": "Compute temporal attention loss" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.419, + 0.387, + 0.435 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {a t t} = \\tau_ {2} - | | A _ {a d v} - A _ {s r c} | | ^ {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.435, + 0.262, + 0.449 + ], + "angle": 0, + "content": "Compute the final loss" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.449, + 0.465, + 0.465 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\left\\| I _ {a d v} - I _ {s r c} \\right\\| ^ {2} + \\alpha \\cdot \\mathcal {L} _ {e n c} + \\beta \\cdot \\mathcal {L} _ {c o n} + \\gamma \\cdot \\mathcal {L} _ {a t t}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.465, + 0.419, + 0.48 + ], + "angle": 0, + "content": "Update parameters \\(I_{adv} \\gets I_{adv} - \\lambda \\cdot \\nabla_{I_{adv}}\\mathcal{L}\\)" + }, + { + "type": "title", + "bbox": [ + 0.095, + 0.481, + 0.149, + 0.493 + ], + "angle": 0, + "content": "end for" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.531, + 0.424, + 0.546 + ], + "angle": 0, + "content": "The complete algorithm is presented in Algorithm 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.549, + 0.47, + 0.626 + ], + "angle": 0, + "content": "It is noteworthy that our method supports both global and local attacks, allowing perturbations to be applied across the entire image or to specific objects within the image. This feature can be implemented by applying a mask to the noise during the attack process." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.65, + 0.21, + 0.667 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.678, + 0.284, + 0.695 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.473, + 0.9 + ], + "angle": 0, + "content": "Models In the field of video generation, there are few open-source video diffusion models available for experimentation. Stable Video Diffusion [4] (SVD) is a widely used and effective image-to-video diffusion model, which we adopt as the baseline model for our attack and experimental analysis. For other conditional image-to-video models, we employ CogVideoX-5B-I2V [53] for text-image joint conditional video generation. The open-sourced CogVideoX adopts DiT rather than UNet architecture used in SVD, providing enhanced capabilities. We also employ ControlNeXt-SVD-v2 [34] for pose-guided image-to-video generation, which is built upon SVD and provides enhanced control over character animation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.289 + ], + "angle": 0, + "content": "Data To our knowledge, there are currently no established benchmarks or datasets for image-based adversarial attacks in video generation. In previous work, PhotoGuard [37] reported attack results on over 60 images, while PRIME [28] used an internal dataset containing 35 video clips. In our experiments, we collect a dataset of 300 images and their corresponding generated frames. The selected images and frames depict people and animals with dynamic actions, which could potentially be misused. We exclude images of scenery without main subjects or images that do not yield noticeable temporal movement, as these typically do not produce meaningful motion, rely mainly on camera movements, and may even remain static." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.306, + 0.895, + 0.564 + ], + "angle": 0, + "content": "Metrics We validate our experiments from two perspectives: image perturbation and generation effects. Our goal is to introduce minimal perturbations to the images that significantly degrade the generated results. We use peak signal-to-noise ratio (PSNR), structural similarity (SSIM) metrics, and Fréchet Inception Distance (FID) to evaluate the similarity between the attacked and original images. For evaluating video generation results, we use the metrics Subject Consistency, Motion Smoothness, Aesthetic Quality, and Image Quality from VBench [25]. Subject Consistency measures the average cosine similarity of Dino [6] features between each frame and the first frame, assessing temporal coherence. Motion Smoothness calculates the mean absolute error between interpolated and dropped frames. Aesthetic Quality is assessed using the LAION aesthetic predictor [1], and Image Quality evaluates the presence of low-level distortions in generated video frames with MUSIQ [26]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.581, + 0.895, + 0.793 + ], + "angle": 0, + "content": "Setup We control the added noise strength by adjusting the number of training steps and the perturbation threshold at each step. For the loss function, we scale the hyperparameters \\(\\alpha, \\beta, \\gamma\\) to ensure that each loss component remains on a similar scale. During training, we use a black image as the target image \\(I_{tg}\\) and black frames as the target video \\(V_{tg}\\). For the attention hooks, we focus on the temporal attention block of the last downsampling blocks in the UNet and the middle blocks in the transformer, separately. In terms of computational cost, the SVD model requires 68GB of GPU memory and approximately 2.5 minutes to apply protection, while CogVideoX demands 75GB of GPU memory and 10 minutes. Notably, this protecting time is comparable to the inference time of the I2V model." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.692, + 0.818 + ], + "angle": 0, + "content": "4.2. Qualitative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In this section, we provide an analysis of attacks on both unconditional and conditional image-to-video generation. Additionally, we present a robustness analysis in Appendix A to demonstrate that our method remains effective in safeguarding images even in the presence of various pre-processing" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12599" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.088, + 0.252, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.134, + 0.157, + 0.215, + 0.166 + ], + "angle": 0, + "content": "Original Image" + }, + { + "type": "image", + "bbox": [ + 0.257, + 0.089, + 0.411, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.089, + 0.563, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.563, + 0.089, + 0.716, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.716, + 0.089, + 0.869, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.463, + 0.157, + 0.666, + 0.167 + ], + "angle": 0, + "content": "Generated video from original image." + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.167, + 0.251, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.234, + 0.234, + 0.244 + ], + "angle": 0, + "content": "Random Noisy Image" + }, + { + "type": "image", + "bbox": [ + 0.258, + 0.167, + 0.411, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.167, + 0.563, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.563, + 0.167, + 0.716, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.716, + 0.167, + 0.869, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.446, + 0.234, + 0.683, + 0.244 + ], + "angle": 0, + "content": "Generated video from random noisy image." + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.244, + 0.252, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.127, + 0.312, + 0.23, + 0.321 + ], + "angle": 0, + "content": "PhotoGuard Image" + }, + { + "type": "image", + "bbox": [ + 0.258, + 0.244, + 0.411, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.244, + 0.563, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.563, + 0.244, + 0.716, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.716, + 0.244, + 0.869, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.421, + 0.311, + 0.706, + 0.321 + ], + "angle": 0, + "content": "Generated video from PhotoGuard protected image." + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.322, + 0.252, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.125, + 0.389, + 0.232, + 0.399 + ], + "angle": 0, + "content": "Our Guarded Image" + }, + { + "type": "image", + "bbox": [ + 0.258, + 0.322, + 0.411, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.322, + 0.563, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.563, + 0.322, + 0.716, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.716, + 0.322, + 0.869, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.46, + 0.389, + 0.667, + 0.399 + ], + "angle": 0, + "content": "Generated video from guarded image." + }, + { + "type": "image_caption", + "bbox": [ + 0.188, + 0.402, + 0.781, + 0.414 + ], + "angle": 0, + "content": "(a) SVD generation results on original image, random noisy image, PhotoGuard image and our I2VGuard image." + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.416, + 0.252, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.136, + 0.495, + 0.218, + 0.505 + ], + "angle": 0, + "content": "Original Image" + }, + { + "type": "image", + "bbox": [ + 0.258, + 0.416, + 0.411, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.416, + 0.563, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.563, + 0.416, + 0.716, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.716, + 0.416, + 0.871, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.465, + 0.496, + 0.667, + 0.506 + ], + "angle": 0, + "content": "Generated video from original image." + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.506, + 0.252, + 0.585 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.137, + 0.587, + 0.223, + 0.597 + ], + "angle": 0, + "content": "Guarded Image" + }, + { + "type": "image", + "bbox": [ + 0.258, + 0.506, + 0.411, + 0.585 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.506, + 0.563, + 0.585 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.563, + 0.506, + 0.716, + 0.585 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.716, + 0.506, + 0.871, + 0.585 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.463, + 0.587, + 0.669, + 0.598 + ], + "angle": 0, + "content": "Generated video from guarded image." + }, + { + "type": "image_caption", + "bbox": [ + 0.217, + 0.602, + 0.751, + 0.615 + ], + "angle": 0, + "content": "(b) CogVideoX generation results, where we do not set the prompt in the process of video generation." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.627, + 0.892, + 0.656 + ], + "angle": 0, + "content": "Figure 4. Qualitative results of adversarial attacks on I2V models SVD and CogVideoX. We also include generation results of SVD with random noise and PhotoGuard [37] perturbations for comparison. All generation results are using the same seed." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.681, + 0.471, + 0.743 + ], + "angle": 0, + "content": "techniques and denoising modules [7]. A detailed visualization of the generation effects is provided in Appendix B. Furthermore, an additional comparison between spatial and temporal attacks is presented in Appendix E." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Attack on Image-to-Video Generation We analyze the generation results using two image-to-video generation models: Stable Video Diffusion (SVD) and CogVideoX, as shown in Fig. 4. We observe that the noise added to the guarded image is nearly invisible, yet it causes significant variations in the generated results. In the example shown in Fig. 4a, we additionally illustrate the generated results for both a random noisy image and an image protected by PhotoGuard [37]. The random noise does not affect generation quality, preserv" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.681, + 0.896, + 0.894 + ], + "angle": 0, + "content": "ing both temporal consistency and normal spatial appearance, likely because the VAE filters out such random noise. The PhotoGuard-protected image, however, introduces noise specifically on the person, which slightly affects appearance without compromising temporal consistency. In contrast, our protected image disrupts both spatial content and temporal consistency. Specifically, the woman's head and hair movement appear unnatural, chaotic frames are generated, and unusual textures emerge in the background across all frames. Fig. 4b shows results from CogVideoX under an image-only setting, where no prompt is applied during training and inference. CogVideoX, which is relatively more powerful than SVD, requires relatively stronger noise to affect generation. It can be observed that motion distortions still occur, particu" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12600" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.088, + 0.195, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.103, + 0.215, + 0.19, + 0.226 + ], + "angle": 0, + "content": "Original Image" + }, + { + "type": "image", + "bbox": [ + 0.201, + 0.089, + 0.296, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.297, + 0.089, + 0.39, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.393, + 0.089, + 0.485, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.089, + 0.58, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.089, + 0.674, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.676, + 0.089, + 0.769, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.771, + 0.089, + 0.864, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.425, + 0.215, + 0.641, + 0.226 + ], + "angle": 0, + "content": "Generated video from original image." + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.226, + 0.194, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.103, + 0.353, + 0.194, + 0.364 + ], + "angle": 0, + "content": "Guarded Image" + }, + { + "type": "image", + "bbox": [ + 0.201, + 0.226, + 0.296, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.297, + 0.226, + 0.39, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.393, + 0.226, + 0.485, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.226, + 0.58, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.226, + 0.674, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.676, + 0.226, + 0.769, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.771, + 0.226, + 0.864, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.425, + 0.352, + 0.645, + 0.364 + ], + "angle": 0, + "content": "Generated video from guarded image." + }, + { + "type": "image_caption", + "bbox": [ + 0.367, + 0.371, + 0.604, + 0.382 + ], + "angle": 0, + "content": "(a) ControlNeXT-SVD-v2 generation results." + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.382, + 0.253, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.136, + 0.463, + 0.218, + 0.473 + ], + "angle": 0, + "content": "Original Image" + }, + { + "type": "image", + "bbox": [ + 0.258, + 0.382, + 0.413, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.414, + 0.382, + 0.565, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.566, + 0.382, + 0.718, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.382, + 0.871, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.465, + 0.464, + 0.668, + 0.475 + ], + "angle": 0, + "content": "Generated video from original image." + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.475, + 0.253, + 0.554 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.137, + 0.555, + 0.223, + 0.565 + ], + "angle": 0, + "content": "Guarded Image" + }, + { + "type": "image", + "bbox": [ + 0.258, + 0.475, + 0.413, + 0.554 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.414, + 0.475, + 0.565, + 0.554 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.566, + 0.475, + 0.718, + 0.554 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.475, + 0.871, + 0.554 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.463, + 0.555, + 0.669, + 0.566 + ], + "angle": 0, + "content": "Generated video from guarded image." + }, + { + "type": "image_caption", + "bbox": [ + 0.273, + 0.57, + 0.696, + 0.583 + ], + "angle": 0, + "content": "(b) CogVideoX generation results. The prompt is \"Many ostriches are running\"." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.595, + 0.893, + 0.625 + ], + "angle": 0, + "content": "Figure 5. Qualitative results of adversarial attacks on conditional I2V models ControlNeXt and CogVideoX. It can be observed that visual features are not controlled effectively, leading to unreasonable generation results. All generation results are using the same seed." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.65, + 0.472, + 0.726 + ], + "angle": 0, + "content": "larly in areas where motion takes place, such as the blurred head movement of the man, with the noise causing the surrounding regions to fail in generation, resulting in a blur. In summary, our method effectively disrupts both spatial content and temporal consistency in generated videos." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.901 + ], + "angle": 0, + "content": "Attack on Conditional Image-to-Video Generation We present an attack on conditional image-to-video models in Fig. 5, specifically examining pose-conditioned ControlNeXt-SVD and text-guided video generation with CogVideoX. In the first row of Fig. 5a, the ControlNeXt model effectively controls the character's motion while maintaining consistent character appearance and background features. However, in the second row, where we guard the original image, the model fails to preserve the character's appearance and fidelity in the background. Noticeable differences" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.65, + 0.895, + 0.892 + ], + "angle": 0, + "content": "emerge: for instance, the fruits in the background exhibit blurring, and the character appears deformed. These spatial distortions illustrate the effectiveness of our attack method. Despite this, the pose control remains robust, likely due to the intensive training of the original SVD UNet, which results in strong overfitting to the specific style and hands shape. In the example of CogVideoX shown in Fig. 5b, we observe that adding special texture noise leads to the generation of temporal inconsistency and blurred content compared with the original ones. Specifically, in this case, texture noise is applied around the animals, causing distortions in the generated output, such as headless and transparent ostriches. In our experiments, CogVideoX exhibits strong prompt-following ability, which is due to an additional powerful text encoder compared to SVD. By slightly increasing the noise strength, we are able to successfully execute the attack." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12601" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.891, + 0.16 + ], + "angle": 0, + "content": "
Video SourceModelSubject Consistency(%, ↓)Motion Smoothness(%, ↓)Aesthetic Quality (%, ↓)Image Quality(↓)
Original ImageSVD95.86±2.6297.90±1.4356.76±4.7567.28±6.18
Guarded Image91.57±3.9597.18±1.2153.42±4.9364.38±8.23
Original ImageCogVideoX97.02±1.9699.19±0.2759.94±5.5367.60±6.49
Guarded Image93.50±3.5897.97±0.3253.95±5.6265.24±9.85
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.17, + 0.893, + 0.213 + ], + "angle": 0, + "content": "Table 1. Analysis of video generation effects of SVD and CogVideoX from original images and images guarded by our method. Mean and variance of evaluations are reported. We exclude results with extremely high subject consistency and motion smoothness, as these indicate static frames, which are outside the scope of this evaluation. An ablation study on the different attack methods is also presented in ??" + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.235, + 0.466, + 0.285 + ], + "angle": 0, + "content": "
Image TypePSNR(↑)SSIM(↑)FID(↓)
Random Noise26.22±0.040.774±0.00314.31
Guarded Image31.83±0.190.868±0.01616.78
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.295, + 0.472, + 0.352 + ], + "angle": 0, + "content": "Table 2. Analysis of perturbations between original and modified images. Random noise is introduced to the original images to establish a baseline, and the mean and variance of the perturbations are reported." + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.366, + 0.468, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.08, + 0.49, + 0.466, + 0.505 + ], + "angle": 0, + "content": "Figure 6. Frame-level quantitative evaluation with the same seed." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.532, + 0.281, + 0.549 + ], + "angle": 0, + "content": "4.3. Quantitative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.562, + 0.472, + 0.819 + ], + "angle": 0, + "content": "Image Perturbation Analysis We present the modified image quality results in Tab. 2. Random noise was added to the original images to establish a baseline for comparison. The mean and variance of the added noise in this baseline are generally comparable to the modifications introduced in our attacked images. The results demonstrate that, despite similar perturbation strength, our method achieves significantly higher image quality, as indicated by higher score in both peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM). A qualitative comparison reveals that the noise added to the image possesses its own texture, which is less visible than the random noise added. In terms of image quality and diversity, our approach yields a higher Fréchet Inception Distance (FID). This is expected because the specific texture noise we add aligns with the original image texture, resulting in lower diversity compared to images with added random noise." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Video Generation Analysis Evaluation results for generated videos of SVD and CogVideoX are shown in Tab. 1. First, we assess the performance of the I2V model on original images, finding that it demonstrates strong subject consis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.239, + 0.895, + 0.646 + ], + "angle": 0, + "content": "tency, smooth motion transitions, high aesthetic quality, and well-preserved low-level image quality. This suggests that the I2V models are capable of generating visually coherent and temporally stable content. In particular, CogVideoX performs a better performance than SVD because transformers scale more easily in both data and model size than UNet, making CogVideoX superior to SVD in size and performance. In contrast, our method, which uses a similar noise strength, significantly impacts all evaluated metrics. To be precise, temporal assessments reveal declines around \\(4\\%\\) in subject consistency and around \\(1\\%\\) in motion smoothness, showing that our method effectively disrupts the model's temporal coherence. Specifically, the decline in subject consistency is reflected in irregular movement of the main object's location or the sporadic appearance and disappearance of the object. The decline in motion smoothness is reflected in abrupt, discontinuous frame-to-frame movements. Spatial evaluations also show that our method considerably diminishes aesthetic quality and low-level image quality by \\(3 \\sim 6\\%\\) and \\(2 \\sim 3\\%\\), respectively. The decrease in aesthetic quality is evident in the appearance of unusual textures or object deformations that detract from the visual appeal. These reductions reflect a successful attack on the image-to-video model. Notably, because CogVideoX employs a 3D VAE, which handles spatiotemporal noise in videos better than the 2D+1D VAE in SVD, CogVideoX shows greater resistance to attacks compared to SVD." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.647, + 0.895, + 0.709 + ], + "angle": 0, + "content": "Additionally, we provide a frame-level quantitative evaluation in Fig. 6. The results indicate that our method effectively disrupts temporal consistency, leading to a propagated deviation from the original generation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.724, + 0.619, + 0.74 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.901 + ], + "angle": 0, + "content": "In this paper, we presented I2VGuard, a novel adversarial method that applies imperceptible perturbations on images to degrade the quality of videos generated by diffusion-based image-to-video models, thereby protecting images from misuse. Concretely, we designed three attack modules, i.e. spatial attack, temporal attack and diffusion attack, to disrupt visual fidelity, temporal consistency, and cross-model robustness, respectively. Extensive experiments on state-of-the-art generative models, i.e. CogVideoX and SVD, demonstrated the effectiveness of our method." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12602" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.303, + 0.128 + ], + "angle": 0, + "content": "[1] LAION AI. Laion-aesthetics. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.131, + 0.472, + 0.185 + ], + "angle": 0, + "content": "[2] Ali Asghar, Amna Shifa, and Mamoona Naveed Asghar. Survey on video security: Examining threats, challenges, and future trends. Computers, Materials & Continua, 80(3), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.188, + 0.472, + 0.257 + ], + "angle": 0, + "content": "[3] Omer Bar-Tal, Hila Chefer, Omer Tov, Charles Herrmann, Roni Paiss, Shiran Zada, Ariel Ephrat, Junhwa Hur, Yuanzhen Li, Tomer Michaeli, et al. Lumiere: A space-time diffusion model for video generation. arXiv preprint arXiv:2401.12945, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.259, + 0.472, + 0.328 + ], + "angle": 0, + "content": "[4] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.33, + 0.472, + 0.386 + ], + "angle": 0, + "content": "[5] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.388, + 0.472, + 0.457 + ], + "angle": 0, + "content": "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.459, + 0.472, + 0.5 + ], + "angle": 0, + "content": "[7] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In European conference on computer vision, pages 17-33. Springer, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.502, + 0.472, + 0.558 + ], + "angle": 0, + "content": "[8] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.559, + 0.472, + 0.615 + ], + "angle": 0, + "content": "[9] Weixin Chen, Dawn Song, and Bo Li. Trojdiff: Trojan attacks on diffusion models with diverse targets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4035-4044, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.617, + 0.472, + 0.671 + ], + "angle": 0, + "content": "[10] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In IEEE Conference on Computer Vision and Pattern Recognition, 2021. Computer Vision Foundation / IEEE, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.673, + 0.472, + 0.729 + ], + "angle": 0, + "content": "[11] June Suk Choi, Kyungmin Lee, Jongheon Jeong, Saining Xie, Jinwoo Shin, and Kimin Lee. Diffusionguard: A robust defense against malicious diffusion-based image editing. In ICML 2024 Next Generation of AI Safety Workshop. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.731, + 0.472, + 0.786 + ], + "angle": 0, + "content": "[12] Sheng-Yen Chou, Pin-Yu Chen, and Tsung-Yi Ho. How to backdoor diffusion models? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4015-4024, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.788, + 0.472, + 0.83 + ], + "angle": 0, + "content": "[13] Hai Ci, Yiren Song, Pei Yang, Jinheng Xie, and Mike Zheng Shou. Wmadapter: Adding watermark control to latent diffusion models. arXiv preprint arXiv:2406.08337, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.832, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[14] Özgün Çiçek, Ahmed Abdulkadir, Soeren S Lienkamp, Thomas Brox, and Olaf Ronneberger. 3d u-net: learning dense volumetric segmentation from sparse annotation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2016: 19th International Conference," + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Athens, Greece, October 17-21, 2016, Proceedings, Part II 19, pages 424-432. Springer, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[15] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.163, + 0.892, + 0.231 + ], + "angle": 0, + "content": "[16] Weitao Feng, Wenbo Zhou, Jiyan He, Jie Zhang, Tianyi Wei, Guanlin Li, Tianwei Zhang, Weiming Zhang, and Nenghai Yu. Aqualora: Toward white-box protection for customized stable diffusion models via watermark lora. arXiv preprint arXiv:2405.11135, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.232, + 0.894, + 0.3 + ], + "angle": 0, + "content": "[17] Pierre Fernandez, Guillaume Couairon, Hervé Jégou, Matthijs Douze, and Teddy Furon. The stable signature: Rooting watermarks in latent diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22466-22477, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.302, + 0.892, + 0.37 + ], + "angle": 0, + "content": "[18] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Imagine your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.371, + 0.892, + 0.413 + ], + "angle": 0, + "content": "[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.414, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[20] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsanko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.483, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[21] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.539, + 0.892, + 0.592 + ], + "angle": 0, + "content": "[22] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.594, + 0.892, + 0.649 + ], + "angle": 0, + "content": "[23] Robert Honig, Javier Rando, Nicholas Carlini, and Florian Tramér. Adversarial perturbations cannot reliably protect artists from generative ai. arXiv preprint arXiv:2406.12027, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.65, + 0.892, + 0.705 + ], + "angle": 0, + "content": "[24] Li Hu. Animate anyone: Consistent and controllable image-to-video synthesis for character animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8153-8163, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.706, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[25] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.789, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[26] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5148-5157, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.845, + 0.892, + 0.859 + ], + "angle": 0, + "content": "[27] PKU-Yuan Lab and Tuzhan AI etc. Open-sora-plan, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[28] Guanlin Li, Shuai Yang, Jie Zhang, and Tianwei Zhang. Prime: Protect your videos from malicious editing. arXiv preprint arXiv:2402.01239, 2024. 3, 5" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12603" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.133 + ], + "angle": 0, + "content": "[29] Chumeng Liang and Xiaoyu Wu. Mist: Towards improved adversarial examples for diffusion models. arXiv preprint arXiv:2305.12683, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.472, + 0.204 + ], + "angle": 0, + "content": "[30] Chumeng Liang, Xiaoyu Wu, Yang Hua, Jiaru Zhang, Yiming Xue, Tao Song, Zhengui Xue, Ruhui Ma, and Haibing Guan. Adversarial example does good: Preventing painting imitation from diffusion models via adversarial examples. arXiv preprint arXiv:2302.04578, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.205, + 0.472, + 0.258 + ], + "angle": 0, + "content": "[31] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International conference on machine learning, pages 8162-8171. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.26, + 0.472, + 0.302 + ], + "angle": 0, + "content": "[32] Josephine Passananti, Stanley Wu, Shawn Shan, Haitao Zheng, and Ben Y Zhao. Disrupting style mimicry attacks on video imagery. arXiv preprint arXiv:2405.06865, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.302, + 0.472, + 0.357 + ], + "angle": 0, + "content": "[33] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.359, + 0.472, + 0.412 + ], + "angle": 0, + "content": "[34] Bohao Peng, Jian Wang, Yuechen Zhang, Wenbo Li, MingChang Yang, and Jiaya Jia. Controlnext: Powerful and efficient control for image and video generation. arXiv preprint arXiv:2408.06070, 2024. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.413, + 0.472, + 0.483 + ], + "angle": 0, + "content": "[35] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.484, + 0.472, + 0.567 + ], + "angle": 0, + "content": "[36] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.567, + 0.472, + 0.635 + ], + "angle": 0, + "content": "[37] Hadi Salman, Alaa Khaddaj, Guillaume Leclerc, Andrew Ilyas, and Aleksander Madry. Raising the cost of malicious ai-powered image editing. In Proceedings of the 40th International Conference on Machine Learning, pages 29894–29918, 2023. 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.636, + 0.472, + 0.704 + ], + "angle": 0, + "content": "[38] Shawn Shan, Jenna Cryan, Emily Wenger, Haitao Zheng, Rana Hanocka, and Ben Y Zhao. Glaze: Protecting artists from style mimicry by {Text-to-Image} models. In 32nd USENIX Security Symposium (USENIX Security 23), pages 2187–2204, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.705, + 0.472, + 0.762 + ], + "angle": 0, + "content": "[39] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.472, + 0.829 + ], + "angle": 0, + "content": "[40] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.472, + 0.871 + ], + "angle": 0, + "content": "[41] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.872, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[42] Jingwei Sun, Xuchong Zhang, Changfeng Sun, Qicheng Bai, and Hongbin Sun. Latent feature and attention dual erasure" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.894, + 0.12 + ], + "angle": 0, + "content": "attack against multi-view diffusion models for 3d assets protection. arXiv preprint arXiv:2408.11408, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.121, + 0.736, + 0.135 + ], + "angle": 0, + "content": "[43] Kling AI Team. Kling ai, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.725, + 0.15 + ], + "angle": 0, + "content": "[44] Luma AI Team. Luma ai, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.152, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[45] The Movie Gen team @ Meta. Movie gen: A cast of media foundation models. 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.181, + 0.894, + 0.222 + ], + "angle": 0, + "content": "[46] Vu Tuan Truong, Luan Ba Dang, and Long Bao Le. Attacks and defenses for generative diffusion models: A comprehensive survey. arXiv preprint arXiv:2408.03400, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.224, + 0.892, + 0.251 + ], + "angle": 0, + "content": "[47] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.253, + 0.892, + 0.294 + ], + "angle": 0, + "content": "[48] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.295, + 0.892, + 0.363 + ], + "angle": 0, + "content": "[49] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.366, + 0.892, + 0.42 + ], + "angle": 0, + "content": "[50] Yuxin Wen, John Kirchenbauer, Jonas Geiping, and Tom Goldstein. Tree-ring watermarks: Fingerprints for diffusion images that are invisible and robust. arXiv preprint arXiv:2305.20030, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.422, + 0.892, + 0.505 + ], + "angle": 0, + "content": "[51] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7623-7633, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.507, + 0.892, + 0.548 + ], + "angle": 0, + "content": "[52] Zhen Xing, Qijun Feng, Haoran Chen, Qi Dai, Han Hu, Hang Xu, Zuxuan Wu, and Yu-Gang Jiang. A survey on video diffusion models. ACM Computing Surveys, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.549, + 0.892, + 0.618 + ], + "angle": 0, + "content": "[53] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.62, + 0.892, + 0.674 + ], + "angle": 0, + "content": "[54] Jianping Zhang, Zhuoer Xu, Shiwen Cui, Changhua Meng, Weibin Wu, and Michael R Lyu. On the robustness of latent diffusion models. arXiv preprint arXiv:2306.08257, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.676, + 0.892, + 0.731 + ], + "angle": 0, + "content": "[55] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.733, + 0.894, + 0.803 + ], + "angle": 0, + "content": "[56] Peifei Zhu, Tsubasa Takahashi, and Hirokatsu Kataoka. Watermark-embedded adversarial examples for copyright protection against diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24420–24430, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.803 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12604" + } + ] +] \ No newline at end of file diff --git a/2025/I2VGuard_ Safeguarding Images against Misuse in Diffusion-based Image-to-Video Models/612ebd29-cda5-46e3-b035-7c7393e6f819_origin.pdf b/2025/I2VGuard_ Safeguarding Images against Misuse in Diffusion-based Image-to-Video Models/612ebd29-cda5-46e3-b035-7c7393e6f819_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..69d09a90e82ec131fb232421b1f1d69c5db17837 --- /dev/null +++ b/2025/I2VGuard_ Safeguarding Images against Misuse in Diffusion-based Image-to-Video Models/612ebd29-cda5-46e3-b035-7c7393e6f819_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68591e00a4c57365d1f5955abc5bb2623455ad198e21fd333a59cffab2729140 +size 6257816 diff --git a/2025/I2VGuard_ Safeguarding Images against Misuse in Diffusion-based Image-to-Video Models/full.md b/2025/I2VGuard_ Safeguarding Images against Misuse in Diffusion-based Image-to-Video Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..99f3ae73d8e1a79a13e88c7067d1500b02a25b52 --- /dev/null +++ b/2025/I2VGuard_ Safeguarding Images against Misuse in Diffusion-based Image-to-Video Models/full.md @@ -0,0 +1,451 @@ +# I2VGuard: Safeguarding Images against Misuse in Diffusion-based Image-to-Video Models + +Dongnan Gui $^{1,*}$ + +Xun Guo2 + +Wengang Zhou + +Yan Lu2 + +1University of Science and Technology of China + +2Microsoft Research Asia + +{gdn2001@mail.,zhwg@}ustc.edu.cn{xunguo，yanlu}@microsoft.com + +![](images/cd5a9666902b55dad978753a14b6623d7bc761247732aebf941009b5ed30e5a7.jpg) +Original Image + +![](images/b190dee41c14ae9a45c74f58645879e1586b85acf5240a675ff4848cb697eae2.jpg) + +![](images/051a828e773f4ee3851ddc9f28a8a74f795e83c7df6a5b1ad00e8c461a1c1fc4.jpg) + +![](images/cbe52968a8b12665d23cd911ca5608df22960afc66255a44296a065ecff0ed9c.jpg) + +![](images/f1edf07cec993b7d1c4c97a6ca1d8a314aa29a5a280a5b87f782a073f43977cb.jpg) + +![](images/f9a319aa8b7a334532a2d2bceff2c427f58e4a804e497353b5fed06c19dfcf5b.jpg) +Guarded Image + +![](images/589d90852a09076afc400ebc93d4e1f68a816aa23cab1c393a3ce95c57a661a3.jpg) + +![](images/cef38a4576704821a26cf49c350e4aeec3a722a3d431d59bd8b57dbd5d5c0a3b.jpg) +Generated video from original image. +Generated video from guarded image. + +![](images/f70473bfbc3bab0cc2728e2266eb7ef5280c4ce76af2b67d690dfac815ea1513.jpg) + +![](images/175a6c6eb12240bd22927794a1d8232bbd16d3fd7e6cca83b63190e100e6d4aa.jpg) + +![](images/5f3acdcdfb3a1d63d34e5882c9402bd32ccee10bf1547777cfb3af1f6e58bac2.jpg) +Original Image + +![](images/512509244fecfd83cd0928919853d3e5ddd552e1fed217f270119af7b49c05b5.jpg) + +![](images/33dceaebe1057b006cefbab2d6a5950991aef561b10c4e464af994cdf065e5f6.jpg) +Generated video from original image. + +![](images/22e38137629a7877bed27ee515b2f9ea8363b15c368f8f601ff01fda0595890f.jpg) + +![](images/a0e2340d8dc6e79fa2720cce58e46cfe5d0e1d7c139c11462c689a7806f349c4.jpg) + +![](images/def9bfe6a5187b68c1ac5b4c089ed6741f1e34b41b7ab85de8b424622890ddc4.jpg) +Guarded Image +Figure 1. Results of our I2VGuard. We present original images, guarded images, and their corresponding SVD-generated videos. All results are generated with the same seed. Our method effectively safeguards images from animation in image-to-video generation. + +![](images/c0bac01a0e9ca711bb9935d7687b3e6f683ce4a02bb58963e625b6b1cefd9c36.jpg) + +![](images/5494f90f7e7b67167e5fa5c9f2a272746a3463b009a1fd1fdc25f4a87a0a6b06.jpg) +Generated video from guarded image. + +![](images/5e2c7109f832796edcd5af975f50944984abdb2bf07d6def1d09a6b1eabfa2d2.jpg) + +![](images/1b71dfa1a84a6fb6a373ad966290ca33dba26e0d2b220cbac53a67ccdc803798.jpg) + +# Abstract + +Recent advances in image-to-video generation have enabled animation of still images and offered pixel-level controllability. While these models hold great potential to transform single images into vivid and dynamic videos, they also carry risks of misuse that could impact privacy, security, and copyright protection. This paper proposes a novel approach that applies imperceptible perturbations on images to degrade the quality of the generated videos, thereby protecting images from misuse in white-box image-to-video diffusion models. Specifically, we function our approach as an adversarial attack, incorporating spatial, temporal, and diffusion attack modules. The spatial attack shifts image features from + +their original distribution to a lower-quality target distribution, reducing visual fidelity. The temporal attack disrupts coherent motion by interfering with temporal attention maps that guide motion generation. To enhance the robustness of our approach across different models, we further propose a diffusion attack module leveraging contrastive loss. Our approach can be easily integrated with mainstream diffusion-based I2V models. Extensive experiments on SVD, CogVideoX, and ControlNeXt demonstrate that our method significantly impairs generation quality in terms of visual clarity and motion consistency, while introducing only minimal artifacts to the images. To the best of our knowledge, we are the first to explore adversarial attacks on image-to-video generation for security purposes. + +# 1. Introduction + +Diffusion models [15, 19, 35] have gained significant attention recently, particularly in video generation. Advances in models like Sora [5] and its DiT [33] framework have accelerated the development of video diffusion models, leading to numerous academic [4, 22, 27, 53, 55] and commercial [43-45] applications. The primary types of video generation are text-to-video(T2V) and image-to-video(I2V), with the former supporting greater diversity and the latter offering pixel-level control. Leveraging powerful generative models and techniques from text-to-image generation [35], image-to-video models can now produce visually impressive videos. Other conditional I2V models include CogVideoX [53], which generates videos using both text and image inputs, enabling more versatile content creation, and ControlNeXt [34], which focuses on generating videos conditioned on pose information, offering precise control over motion and structure. + +However, the rapid advancement of generative models has also heightened risks related to privacy, security, and copyright [2]. For instance, an adversary with access to a person's image could use image-to-video models to generate misleading or harmful videos, depicting actions the individual never performed, thus compromising individual's privacy and security. Additionally, the unauthorized use of copyrighted images to create animations poses significant risks to intellectual property rights. To address these risks, this work introduces an adversarial attack method designed to prevent the misuse of images in diffusion-based I2V models. Prior research has primarily focused on protecting images from malicious edits [37, 54], which are often viewed as adversarial attacks on image-editing models. Few efforts have focused on safeguarding images from misuse in I2V tasks. + +The main challenges in attacking I2V models come from two aspects: first, disrupting temporal consistency with perturbations applied only to individual images; and second, achieving robustness across diverse diffusion models, including UNet and DiT architectures, even under strong conditioning such as poses and text prompts. To address these challenges, our method targets different modules within I2V models and affects various attributes of the generated video. Specifically, we divide our approach into spatial, temporal, and diffusion attacks. The spatial attack focuses on altering the spatial features of the I2V models by targeting the VAE encoder in latent diffusion models. Here, we manipulate image latents, shifting them toward suboptimal values. The temporal attack aims to perturb the temporal features, which are crucial for video motion and consistency. By extracting the temporal attention map, we intentionally disrupt it by pushing it away from its original attention map. This creates a chaotic attention map, leading to temporal inconsistencies in the generated video. The diffusion attack focuses on the diffusion process and targets the output of the denoising module, which includes the UNet in SVD and the trans + +former in CogVideoX. Unlike a standard denoising process, our approach aims to push the predicted outputs away from the original frames, guiding them closer to target frames. To achieve this, we design a contrastive loss that degrades the quality of the predicted frames by treating the target frames as positive samples and the original predicted frames as negative samples. + +We conduct a comprehensive qualitative and quantitative analysis of the attack results. Qualitative analysis demonstrates that our method effectively disrupts the original image, leading to low temporal consistency and reduced aesthetic quality in both simple and conditioned I2V models. Quantitative analysis confirms a significant degradation in both spatial quality and temporal consistency of the generated videos. To our knowledge, this is the first adversarial attack method aimed at securing images against misuse in I2V generation. + +# 2. Related Work + +Image-to-Video Generation With the great advancements in image generation [15, 19, 31, 35, 40, 41], video generation [3, 5, 20, 22, 39, 43–45, 52, 53] has gathered significant attention and is developing rapidly [18, 48, 49, 51]. Recently, following the success of diffusion models [15, 19, 35, 40] in image generation, diffusion models integrated with UNet [3, 14, 21, 36] or transformers (DiT) [5, 33, 47, 55] have been adopted for video generation. In the field of imaged-to-video generation,AnimateDiff [18] animates personalized text-to-image diffusion models by inserting a temporal motion module, Stable Video Diffusion(SVD) [4] demonstrates strong performance, benefiting from intensive training on large amounts of high-quality data, CogVideoX [53] supports text-image jointly conditional video generation,Animate Anyone [24] and ControlNeXt [34] leverage pose information for controllable image-to-video generation. In this work, we focus on I2V models, which are susceptible to misuse and therefore require robust protection. + +Adversarial Attacks on Diffusion Models Current attacks on diffusion models, which can be broadly categorized [46] into backdoor [9, 12] and adversarial attacks, primarily target image-based diffusion models [13, 16, 17, 50]. In adversarial attacks, the input image or text prompt is perturbed to generate adversarial examples. For instance, AdvDM [30] and Mist [29] utilize adversarial examples to protect human-created artworks, PhotoGuard [37] alters the image by modifying either the encoder or the diffusion outputs against malicious editing, while Glaze [38] perturbs the image to prevent style mimicry, [54] shifts the image away from its original distribution, [56] embeds personal watermarks in the generation of adversarial examples, and DiffusionGuard [11] generates adversarial noise targeting the early stage of the diffusion process. For video-based attacks, + +![](images/97cf788ca4fa741f68aa6c8c0d9f9e680a5bc2d9eab9c2728ff5ec1e1d2849d4.jpg) +Figure 2. Overview of our proposed method. The training process begins with a trainable copy $I_{adv}$ of the original image $I_{src}$ . First, we perform inference to obtain the original video $V_{0}$ and get the latent frames $\mathbf{X}_{0}$ . Noisy latent frames $\mathbf{X}_{t}$ and the latent image $x_{adv}, x_{src}$ are then processed by the denoising model to predict the original frames. The encoded latent image $x_{adv}$ is utilized for the spatial encoder attack, while the predicted original frames $\hat{\mathbf{X}}_{0}$ serve in the diffusion attack to compute contrastive loss. Within the denoising module, we hook into the temporal attention module to extract the temporal attention map. By altering the attacked attention map $A_{adv}$ to diverge from the original $A_{src}$ , we implement the temporal attack. + +PRIME [28] modifies videos frame-by-frame to prevent malicious video editing, while [32] disrupts style mimicry attacks from video imagery. There are also works that protect the image against other generation, such as text-to-3D generation [42]. In this work, we focus on perturbing the input image in image-to-video diffusion models to perform adversarial attacks. Recent study [23] show that adversarial perturbations fail to reliably protect artists from style mimicry in the field of image generation. Our approach stands out in two ways: it uses a white-box attack strategy, allowing adversarial noise training with both generation and pre-processing models, and it focuses on disrupting temporal consistency rather than mere image fidelity, making the perturbations harder to remove. + +# 3. Method + +In this section, we first provide preliminaries on latent diffusion models and problem settings in Sec. 3.1. We then present our attack method, detailing the spatial attack in Sec. 3.2, the temporal attack in Sec. 3.3, and the diffusion attack in Sec. 3.4. An overview of our method is presented in in Fig. 2, which shows our protection procedure involves spatial, temporal, and diffusion attacks. + +# 3.1. Preliminary + +Latent Diffusion Models In diffusion models, realistic image or video sampling from a target distribution $q(\cdot)$ is achieved via a sequence of denoising steps in a diffusion process. This approach involves iteratively adding and removing noise to transform an initial sample $\mathbf{x}_0 \sim q(\cdot)$ . Noise is incrementally introduced across steps, creating intermediate states defined by $\mathbf{x}_{t+1} = \sqrt{\overline{\alpha}_t} \mathbf{x}_0 + \sqrt{1 - \overline{\alpha}_t} \epsilon_t$ , where $\epsilon_t$ is Gaussian noise. The parameters $\alpha_t$ control the noise transition, leading to a final noisy state with a standard Gaussian distribution. By reversing this process and iteratively denoising each state, the model learns to predict the noise added at each step, ultimately enabling the recovery of samples from $q(\cdot)$ starting from $\mathbf{x}_T$ . Latent diffusion models (LDMs), a subclass of diffusion models, operate in a compressed latent space rather than directly in the image space. In LDMs, each input image or video $\mathbf{x}_0$ is first mapped to a latent representation $\mathcal{E}(\mathbf{x}_0)$ via an encoder $\mathcal{E}$ . The diffusion process then applies in this latent space, producing a series of noisy representations. Training in this latent space enables LDMs to achieve efficient training and faster inference while retaining high-quality outputs. + +Problem Settings Adversarial attacks involve adding subtle, nearly undetectable perturbations $\delta_{adv}$ to the input data to influence a model's output. In computer vision, adversarial attacks have been widely explored in classifications and image synthesis. However, in image-to-video generation tasks, adversarial attack remains a relatively unexplored area. In this work, we aim to develop adversarial attacks to safeguard images from misused by image-to-video latent diffusion models. To distinguish between the latent and pixel spaces, we represent images and videos in pixel space as $I$ and $V$ , and in latent space as $x$ and $\mathbf{X}$ , respectively. + +Specifically, an adversary can generate a modified image $I_{adv} = I_{src} + \delta$ from source image $I_{src}$ , crafted to disrupt the generative model $\mathcal{G}$ , resulting in poor-quality or distorted outputs. This adversarial image is optimized by maximizing the loss between the videos generated from the original and modified images, represented as follows: + +$$ +\delta_ {a d v} = \arg \max _ {\| \delta \| _ {p} \leq \epsilon} \mathcal {L} (\mathcal {G} (I _ {s r c} + \delta , \mathbf {X} _ {t}, c), \mathcal {G} (I _ {s r c}, \mathbf {X} _ {t}, c)) +$$ + +where $\delta_{adv}$ represents the computed perturbation, constrained by $\| \delta \| _p\leq \epsilon$ to ensure minimal visibility, and $c$ represents other conditions such as text or pose in the imaged-to-video generation process. The constraint on $\delta_{adv}$ is maintained by a regularization term $\| I_{adv} - I_{src}\| ^2$ + +# 3.2. Spatial Attack + +The generated video has two main attributes: spatial features and temporal features. In this section, we focus on attacking the spatial features within individual frames. Previous work [37, 54] has applied adversarial attacks to image editing tasks, focusing solely on spatial features. Our work extends these methods to image-to-video generation task. + +In Latent Diffusion Models (LDMs), an encoder, denoted as $\mathcal{E}$ , compresses the original image or frames into a latent space representation. An encoder attack aims to compromise this encoder, reducing the effectiveness of generative models. + +$$ +\mathcal {L} _ {e n c} = \left\| \mathcal {E} \left(I _ {a d v}\right) - \mathcal {E} \left(I _ {t g}\right) \right\| ^ {2} \tag {1} +$$ + +Specifically, as shown in Eq. (1), the attack transforms the latent representation $\mathcal{E}(I_{adv})$ into a poor one $\mathcal{E}(I_{tg})$ , resulting in poor-quality final videos. + +# 3.3. Temporal Attack + +Unlike attacks on image editing tasks, where no temporal features are involved, attacks on video generation rely heavily on temporal information. In this task, disrupting temporal consistency through adversarial attacks is both essential and challenging. To reduce temporal consistency and induce chaotic behavior in the outputs, we target the temporal attention maps within the temporal transformer blocks in video diffusion generator, $\mathcal{G}$ . Specifically, during training, we insert + +![](images/bc65197530caa0c39ed1dbb219022649ca06a1ae592d05301895de15a36916f1.jpg) +Figure 3. Temporal self-attention map visualizations. Difference between generated frames from the original image (left) and the guarded image (right) in Fig. 1. + +![](images/1258d5cfa6a7d58c434cd3dd88e56b6cbfadef2f95a2eb094cc497a9acea76b7.jpg) + +a hook, denoted as $\mathcal{G}$ , within the attention block to extract the self-attention map $A_{adv}$ . We then apply a targeted attack to this attention map, pushing it away from the original map, $A_{src}$ , thereby reducing temporal consistency in the generated outputs. To achieve this, we introduce a temporal attention loss, denoted as $\mathcal{L}_{att}$ , as shown in Eq. (2), which induces motion-related perturbations and promotes temporal chaos. + +$$ +\mathcal {L} _ {\text {a t t}} = \tau_ {2} - \left| \left| A _ {\text {a d v}} - A _ {\text {s r c}} \right| \right| ^ {2} \tag {2} +$$ + +As shown in Fig. 3, the visualization of the temporal self-attention map on both the original and modified frames highlights this impact: the manipulated attention map misaligns with the original, directly leading to inconsistencies in motion across frames. This visual evidence underscores the effectiveness of our method in disrupting temporal coherence in video generation by distorting the temporal attention mechanism within the model. + +# 3.4. Diffusion Attack + +Although spatial and temporal attacks can disrupt structural and motion information in generated videos, their effectiveness may vary significantly across different models [4, 34, 53], especially when other conditions, e.g., text prompts, are involved. To enhance the robustness and generalization of the perturbation across diffusion models, we introduce a latent space perturbation that uses contrastive loss to shift the predicted noise closer to the target frames while pushing it away from the original frames. Specifically, given the latent representation of the original frames, $\mathbf{X}_0$ , and the target latent frames, $\mathbf{X}_{tg}$ , we construct a contrastive objective [8, 10], as shown in Eq. (3). + +$$ +\mathbf {L} _ {c o n} = \left\| \hat {\mathbf {X}} _ {0, a d v} - \mathbf {X} _ {t g} \right\| ^ {2} + \max \left(0, \tau_ {1} - \left\| \hat {\mathbf {X}} _ {0, a d v} - \mathbf {X} _ {0} \right\| ^ {2}\right) (3) +$$ + +where $\tau_{1}$ is the margin distance. The positive pair consists of the predicted frames $\hat{\mathbf{X}}_{0,adv}$ and the target frames $\mathbf{X}_{tg}$ , which we aim to bring closer together, while the negative pair is composed of the predicted frames $\hat{\mathbf{X}}_{0,adv}$ and the original frames $\mathbf{X}_0$ , which we aim to push apart. + +With the hyperparameters $\alpha, \beta, \gamma$ , we define the final training objective as: + +$$ +\mathcal {L} = \left\| I _ {a d v} - I _ {s r c} \right\| ^ {2} + \alpha \cdot \mathcal {L} _ {e n c} + \beta \cdot \mathcal {L} _ {c o n} + \gamma \cdot \mathcal {L} _ {a t t} (4) +$$ + +Algorithm 1 Adversarial Attack on Image-to-Video Generation +Input: Source Image $I_{src}$ and Trainable Copy $I_{adv}$ , Target +Image $I_{tg}$ , Target Video $V_{tg}$ , Image-to-Video Pipeline $\mathcal{P}$ , +Generator $\mathcal{G}$ , Encoder $\mathcal{E}$ , Diffusion Scheduler $\mathcal{S}$ +(Optional:Condition $c$ ) +Hyperparameters $\tau_1, \tau_2, \alpha, \beta, \gamma, \lambda$ +Encode image $x_{src}, x_{tg} = \mathcal{E}(I_{src}), \mathcal{E}(I_{tg})$ +Generate original video $V_0 = \mathcal{P}(I_{src})$ +Encode frames $\mathbf{X}_0, \mathbf{X}_{tg} = \mathcal{E}(V_0), \mathcal{E}(V_{tg})$ +for each iteration do + Encode attacked image $x_{adv} = \mathcal{E}(I_{adv})$ + Compute encoder loss: $\mathcal{L}_{enc} = ||x_{adv} - x_{tg}||^2$ + Generate noisy frames $\mathbf{X}_t = \mathcal{S}(\mathbf{X}_0)$ +Predict $\mathbf{X}_0\left\{\begin{array}{l}\hat{\mathbf{X}}_{0,adv}=\varnothing(\mathbf{X}_t,x_{adv},c)\\ \hat{\mathbf{X}}_{0,src}=\varnothing(\mathbf{X}_t,x_{src},c)\\ A_{src}\end{array}\right.$ +Compute spatial contrastive loss + +$$ +\mathcal {L} _ {c o n} = \left| \left| \hat {\mathbf {X}} _ {0, a d v} - \mathbf {X} _ {t g} \right| \right| ^ {2} + \max \left(0, \tau_ {1} - \left| \left| \hat {\mathbf {X}} _ {0, a d v} - \mathbf {X} _ {0} \right| \right| ^ {2}\right) +$$ + +Compute temporal attention loss + +$$ +\mathcal {L} _ {a t t} = \tau_ {2} - | | A _ {a d v} - A _ {s r c} | | ^ {2} +$$ + +Compute the final loss + +$$ +\mathcal {L} = \left\| I _ {a d v} - I _ {s r c} \right\| ^ {2} + \alpha \cdot \mathcal {L} _ {e n c} + \beta \cdot \mathcal {L} _ {c o n} + \gamma \cdot \mathcal {L} _ {a t t} +$$ + +Update parameters $I_{adv} \gets I_{adv} - \lambda \cdot \nabla_{I_{adv}}\mathcal{L}$ + +# end for + +The complete algorithm is presented in Algorithm 1. + +It is noteworthy that our method supports both global and local attacks, allowing perturbations to be applied across the entire image or to specific objects within the image. This feature can be implemented by applying a mask to the noise during the attack process. + +# 4. Experiments + +# 4.1. Experimental Settings + +Models In the field of video generation, there are few open-source video diffusion models available for experimentation. Stable Video Diffusion [4] (SVD) is a widely used and effective image-to-video diffusion model, which we adopt as the baseline model for our attack and experimental analysis. For other conditional image-to-video models, we employ CogVideoX-5B-I2V [53] for text-image joint conditional video generation. The open-sourced CogVideoX adopts DiT rather than UNet architecture used in SVD, providing enhanced capabilities. We also employ ControlNeXt-SVD-v2 [34] for pose-guided image-to-video generation, which is built upon SVD and provides enhanced control over character animation. + +Data To our knowledge, there are currently no established benchmarks or datasets for image-based adversarial attacks in video generation. In previous work, PhotoGuard [37] reported attack results on over 60 images, while PRIME [28] used an internal dataset containing 35 video clips. In our experiments, we collect a dataset of 300 images and their corresponding generated frames. The selected images and frames depict people and animals with dynamic actions, which could potentially be misused. We exclude images of scenery without main subjects or images that do not yield noticeable temporal movement, as these typically do not produce meaningful motion, rely mainly on camera movements, and may even remain static. + +Metrics We validate our experiments from two perspectives: image perturbation and generation effects. Our goal is to introduce minimal perturbations to the images that significantly degrade the generated results. We use peak signal-to-noise ratio (PSNR), structural similarity (SSIM) metrics, and Fréchet Inception Distance (FID) to evaluate the similarity between the attacked and original images. For evaluating video generation results, we use the metrics Subject Consistency, Motion Smoothness, Aesthetic Quality, and Image Quality from VBench [25]. Subject Consistency measures the average cosine similarity of Dino [6] features between each frame and the first frame, assessing temporal coherence. Motion Smoothness calculates the mean absolute error between interpolated and dropped frames. Aesthetic Quality is assessed using the LAION aesthetic predictor [1], and Image Quality evaluates the presence of low-level distortions in generated video frames with MUSIQ [26]. + +Setup We control the added noise strength by adjusting the number of training steps and the perturbation threshold at each step. For the loss function, we scale the hyperparameters $\alpha, \beta, \gamma$ to ensure that each loss component remains on a similar scale. During training, we use a black image as the target image $I_{tg}$ and black frames as the target video $V_{tg}$ . For the attention hooks, we focus on the temporal attention block of the last downsampling blocks in the UNet and the middle blocks in the transformer, separately. In terms of computational cost, the SVD model requires 68GB of GPU memory and approximately 2.5 minutes to apply protection, while CogVideoX demands 75GB of GPU memory and 10 minutes. Notably, this protecting time is comparable to the inference time of the I2V model. + +# 4.2. Qualitative Analysis + +In this section, we provide an analysis of attacks on both unconditional and conditional image-to-video generation. Additionally, we present a robustness analysis in Appendix A to demonstrate that our method remains effective in safeguarding images even in the presence of various pre-processing + +![](images/d7977c29cd41f64b7019d4b8e6088b090b04ddc35f4ae6c64abb0e25a4cd67c7.jpg) +Original Image + +![](images/a4fb6bbe0037b7b8b327b6811e76664207389d21f521b46ab421dd98efcc7921.jpg) + +![](images/c331ad1d0f66ce7d87647755fe634be27ff32f03c513646f86c722058429b364.jpg) + +![](images/4a1f2d50fa091027743d94819fb2e71f1b970a892498221ce1f524f6203b91ff.jpg) + +![](images/a4dd6abbbe134ab289634669e13450abde4a4e2674735a435d8aaccc9cadba67.jpg) + +![](images/8ece16884112bdb351a6dcd4f04fdcda01708404499381bcbb0ffef0985b5963.jpg) +Random Noisy Image + +![](images/d36921865af717dfff9a35729d82395e28a3f330892f001b0a8e5f89bcc07168.jpg) + +![](images/544778c99c3d7ee5e93e420090969dd962bbc5c13bce8f920311aca932993764.jpg) +Generated video from original image. +Generated video from random noisy image. + +![](images/034c98d6a3d5e62c5f75fa9197cd2368f33aca6c071bd7bd3867777f4b2ad68d.jpg) + +![](images/75d42566d1ab83cf6009bad08e0d2e14427e7ef53c6ff6cefd0586e21e2e6a6f.jpg) + +![](images/7cc2f8cf26a202fdfdcdce5e9dbd1db035d81f575b2484cc9beac805f9677c8c.jpg) +PhotoGuard Image + +![](images/1664cf2b2afd8897c1b648bc758281348d2de6e91ae223cb3cceda866c6fded7.jpg) + +![](images/ae9f8e1846b2ec7770ca1f62e524ebc1841933e36e94754fb2f236db612c9fd9.jpg) +Generated video from PhotoGuard protected image. + +![](images/9858540f0f28139820e3ff871239227d880309f123f1b69df22f62dc7b0b3061.jpg) + +![](images/1603d02838ae62d4efe1ceccb3c1277c1b93c7a53a93d67c202f7f6abbeff186.jpg) + +![](images/a24e2af1eee20054f2fe7ac1050a06f2d21b1e5d0a95a714d8b14008ed8dcd03.jpg) +Our Guarded Image + +![](images/18602c149ff527e0e2199c92e4839924b9d51c5ffb02b5d04776225dc8ceebba.jpg) + +![](images/87302b099cdd7cb6f531ecedadfa7b0e662567c93f62bbec2ab49cbf963358d2.jpg) +Generated video from guarded image. + +![](images/f934ebd68b15913493d4a25e030f4858e8a2f5309516fc04e8d785d2dc7ef3fd.jpg) + +![](images/2fdf630e4b34a76095a897010080c7a836f3ef0ec746e3e7dc5152b3bdfdd93f.jpg) + +![](images/cf940a918cddfdb9ef5c5bd97d1814e8f2b440110161aba692b2399b84aa3b86.jpg) +(a) SVD generation results on original image, random noisy image, PhotoGuard image and our I2VGuard image. +Original Image + +![](images/dd1f4407ed07243aa6586466079aef2f196f60c2b413199eb026c04978326cf6.jpg) + +![](images/f748f1b0a2b25fb25e5006566e2478901b42c9d9e09a7c55d86314d53aa8d3f7.jpg) +Generated video from original image. + +![](images/8010378a9eeccfccf47584822d523e1099125d409e1a5cb3d3cad9baafb70257.jpg) + +![](images/56560c5f12c5c4d8e05217fb08870e3f00252483ed16c8bfea762e01b4b07770.jpg) + +![](images/86d28d43bf444ee4d9e5e3651dee275a37ae6f94735076c1cc5c99f3ba82ad34.jpg) +Guarded Image +(b) CogVideoX generation results, where we do not set the prompt in the process of video generation. +Figure 4. Qualitative results of adversarial attacks on I2V models SVD and CogVideoX. We also include generation results of SVD with random noise and PhotoGuard [37] perturbations for comparison. All generation results are using the same seed. + +![](images/7932a5e193773c4df1123d494e7b48d51e1766a469e96ab99593e4d7c32ff534.jpg) + +![](images/5b5d05439d1fdba01850a3e5637cc39619d82b669b2db3cfe1eca44ae79fa5af.jpg) +Generated video from guarded image. + +![](images/eb4cb7ed4b319d75d089214454e4ef47b5d9fa66ece7c9b3f788b9823281ccd5.jpg) + +![](images/9f1fc6d2aab65ed05c37ee2c37f2a1b11bc282b182439cc9d5f0db7d9a430a79.jpg) + +techniques and denoising modules [7]. A detailed visualization of the generation effects is provided in Appendix B. Furthermore, an additional comparison between spatial and temporal attacks is presented in Appendix E. + +Attack on Image-to-Video Generation We analyze the generation results using two image-to-video generation models: Stable Video Diffusion (SVD) and CogVideoX, as shown in Fig. 4. We observe that the noise added to the guarded image is nearly invisible, yet it causes significant variations in the generated results. In the example shown in Fig. 4a, we additionally illustrate the generated results for both a random noisy image and an image protected by PhotoGuard [37]. The random noise does not affect generation quality, preserv + +ing both temporal consistency and normal spatial appearance, likely because the VAE filters out such random noise. The PhotoGuard-protected image, however, introduces noise specifically on the person, which slightly affects appearance without compromising temporal consistency. In contrast, our protected image disrupts both spatial content and temporal consistency. Specifically, the woman's head and hair movement appear unnatural, chaotic frames are generated, and unusual textures emerge in the background across all frames. Fig. 4b shows results from CogVideoX under an image-only setting, where no prompt is applied during training and inference. CogVideoX, which is relatively more powerful than SVD, requires relatively stronger noise to affect generation. It can be observed that motion distortions still occur, particu + +![](images/fea02e2f23a4f2b1a73ebe9f0b262fb516bb5a0a5108cd9fbf57adb41d6c27bd.jpg) + +![](images/d2bc0bd02264c0dc82f7a1b2bdad43fc64e3b0f185263a45bed877303d46e498.jpg) + +![](images/732b0f92e57e636fc706470137ec74f3a8c27fb4669f6bca3500de7ede1d18f3.jpg) + +![](images/577e63b87b31d2475ca66a4a58fba255ec5c6fe9f43237f8a43b9e63b4056a91.jpg) + +![](images/b6a98c6ebf9340adf23267014e7b630189296f4c83d3ed69ec3e4228beecec74.jpg) + +![](images/456e4ccc5b6b1ff8a497e3e1770975c864bbcce35ab5da760030bedd0e296fdb.jpg) + +![](images/9df7b0069d470fc6724ee4dbabf31307c3b7e5975060a1e22c824df1dba5fdbd.jpg) + +![](images/57bfa7dfaf8144b79ed544440db9dc16afdd4a99e0b874657c086eaea9492029.jpg) + +![](images/7565a6bd8f58a633fbca595c789a6fc205452c48296bbf1609224ec9e29364b1.jpg) +Original Image +Guarded Image + +![](images/7607e9e47dae44923b471dc0261a80c6b1cfc50094c6cffa795291c39e783b47.jpg) + +![](images/e1c541087c0921f070ac44991ec1ff560968cbfeb9203b93424887c4554e5ce3.jpg) + +![](images/be1772b4501dc7527efc66417c60b9af54a1f296e9d10abb0e0e33d61512dc3a.jpg) +Generated video from original image. +Generated video from guarded image. + +![](images/ecac280cc5a08bf8914a7a47becaec022edf05a9f50aaf1a1a4c136349fc74f5.jpg) + +![](images/9033dddd4430a3cd70af6e05ea3f58a21a67c7acd2fbc098fe74192bac60df85.jpg) + +![](images/888d2a4950889f5ae6202b6c54ab6235961a0767023d2301e5cc538847f17451.jpg) + +![](images/62f0ebbb7f90847cb2a7ff35f058b59634eb1a8fdaf4a065f309f0fbbb3cafd0.jpg) + +![](images/be2f65d648a05abff5c50122b2d91e9a8290c5072e6c5fa27f5d03301c2bd6d8.jpg) +Original Image + +![](images/61e455ee5c51a98aeb1716703f2275f4bf2c84c1d31dba6adbdc29c1cc92acaa.jpg) +(a) ControlNeXT-SVD-v2 generation results. + +![](images/8bf15d03c40781b56d95285ef5d8ce4332f801cec2a8851e8cd393967d21f06b.jpg) + +![](images/420fa9a3828b40ed26bc4906c75a0010c5c450f026c9c1e64bb36d8cb9d69905.jpg) + +![](images/8e833d2beb9ba8daa3204eda0cf8ce4bc67df15a308989df51285d46427a4a2e.jpg) + +![](images/ceda9160e87cef3f31702cd15516442d97080d992bfb1249a8534d1fa106f864.jpg) +Guarded Image +Figure 5. Qualitative results of adversarial attacks on conditional I2V models ControlNeXt and CogVideoX. It can be observed that visual features are not controlled effectively, leading to unreasonable generation results. All generation results are using the same seed. + +![](images/2f61d492e5569eabeed4034d7ece6397a1d0dece5413fd50ec4f828007f664a3.jpg) +(b) CogVideoX generation results. The prompt is "Many ostriches are running". + +![](images/33299a55a3ebb1e9ecefcb3b3c7dcad3e67c69a4e8235d7bc810c4cf0dc5067a.jpg) +Generated video from original image. +Generated video from guarded image. + +![](images/16af9bcc9dcaf58f21c379e5380342732a5ea835ef3edbfa9d506bad66a38d45.jpg) + +![](images/fcaa90bd8af0b969a616a5409f9810deca6602db051438d5fa2c06ddf4e1a814.jpg) + +larly in areas where motion takes place, such as the blurred head movement of the man, with the noise causing the surrounding regions to fail in generation, resulting in a blur. In summary, our method effectively disrupts both spatial content and temporal consistency in generated videos. + +Attack on Conditional Image-to-Video Generation We present an attack on conditional image-to-video models in Fig. 5, specifically examining pose-conditioned ControlNeXt-SVD and text-guided video generation with CogVideoX. In the first row of Fig. 5a, the ControlNeXt model effectively controls the character's motion while maintaining consistent character appearance and background features. However, in the second row, where we guard the original image, the model fails to preserve the character's appearance and fidelity in the background. Noticeable differences + +emerge: for instance, the fruits in the background exhibit blurring, and the character appears deformed. These spatial distortions illustrate the effectiveness of our attack method. Despite this, the pose control remains robust, likely due to the intensive training of the original SVD UNet, which results in strong overfitting to the specific style and hands shape. In the example of CogVideoX shown in Fig. 5b, we observe that adding special texture noise leads to the generation of temporal inconsistency and blurred content compared with the original ones. Specifically, in this case, texture noise is applied around the animals, causing distortions in the generated output, such as headless and transparent ostriches. In our experiments, CogVideoX exhibits strong prompt-following ability, which is due to an additional powerful text encoder compared to SVD. By slightly increasing the noise strength, we are able to successfully execute the attack. + +
Video SourceModelSubject Consistency(%, ↓)Motion Smoothness(%, ↓)Aesthetic Quality (%, ↓)Image Quality(↓)
Original ImageSVD95.86±2.6297.90±1.4356.76±4.7567.28±6.18
Guarded Image91.57±3.9597.18±1.2153.42±4.9364.38±8.23
Original ImageCogVideoX97.02±1.9699.19±0.2759.94±5.5367.60±6.49
Guarded Image93.50±3.5897.97±0.3253.95±5.6265.24±9.85
+ +Table 1. Analysis of video generation effects of SVD and CogVideoX from original images and images guarded by our method. Mean and variance of evaluations are reported. We exclude results with extremely high subject consistency and motion smoothness, as these indicate static frames, which are outside the scope of this evaluation. An ablation study on the different attack methods is also presented in ?? + +
Image TypePSNR(↑)SSIM(↑)FID(↓)
Random Noise26.22±0.040.774±0.00314.31
Guarded Image31.83±0.190.868±0.01616.78
+ +Table 2. Analysis of perturbations between original and modified images. Random noise is introduced to the original images to establish a baseline, and the mean and variance of the perturbations are reported. + +![](images/5909bcccf746fc3a7f012b0e7208f35b2256749cd4166ae92a75209139490e39.jpg) +Figure 6. Frame-level quantitative evaluation with the same seed. + +# 4.3. Quantitative Analysis + +Image Perturbation Analysis We present the modified image quality results in Tab. 2. Random noise was added to the original images to establish a baseline for comparison. The mean and variance of the added noise in this baseline are generally comparable to the modifications introduced in our attacked images. The results demonstrate that, despite similar perturbation strength, our method achieves significantly higher image quality, as indicated by higher score in both peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM). A qualitative comparison reveals that the noise added to the image possesses its own texture, which is less visible than the random noise added. In terms of image quality and diversity, our approach yields a higher Fréchet Inception Distance (FID). This is expected because the specific texture noise we add aligns with the original image texture, resulting in lower diversity compared to images with added random noise. + +Video Generation Analysis Evaluation results for generated videos of SVD and CogVideoX are shown in Tab. 1. First, we assess the performance of the I2V model on original images, finding that it demonstrates strong subject consis + +tency, smooth motion transitions, high aesthetic quality, and well-preserved low-level image quality. This suggests that the I2V models are capable of generating visually coherent and temporally stable content. In particular, CogVideoX performs a better performance than SVD because transformers scale more easily in both data and model size than UNet, making CogVideoX superior to SVD in size and performance. In contrast, our method, which uses a similar noise strength, significantly impacts all evaluated metrics. To be precise, temporal assessments reveal declines around $4\%$ in subject consistency and around $1\%$ in motion smoothness, showing that our method effectively disrupts the model's temporal coherence. Specifically, the decline in subject consistency is reflected in irregular movement of the main object's location or the sporadic appearance and disappearance of the object. The decline in motion smoothness is reflected in abrupt, discontinuous frame-to-frame movements. Spatial evaluations also show that our method considerably diminishes aesthetic quality and low-level image quality by $3 \sim 6\%$ and $2 \sim 3\%$ , respectively. The decrease in aesthetic quality is evident in the appearance of unusual textures or object deformations that detract from the visual appeal. These reductions reflect a successful attack on the image-to-video model. Notably, because CogVideoX employs a 3D VAE, which handles spatiotemporal noise in videos better than the 2D+1D VAE in SVD, CogVideoX shows greater resistance to attacks compared to SVD. + +Additionally, we provide a frame-level quantitative evaluation in Fig. 6. The results indicate that our method effectively disrupts temporal consistency, leading to a propagated deviation from the original generation. + +# 5. Conclusion + +In this paper, we presented I2VGuard, a novel adversarial method that applies imperceptible perturbations on images to degrade the quality of videos generated by diffusion-based image-to-video models, thereby protecting images from misuse. Concretely, we designed three attack modules, i.e. spatial attack, temporal attack and diffusion attack, to disrupt visual fidelity, temporal consistency, and cross-model robustness, respectively. Extensive experiments on state-of-the-art generative models, i.e. CogVideoX and SVD, demonstrated the effectiveness of our method. + +# References + +[1] LAION AI. Laion-aesthetics. 5 +[2] Ali Asghar, Amna Shifa, and Mamoona Naveed Asghar. Survey on video security: Examining threats, challenges, and future trends. Computers, Materials & Continua, 80(3), 2024. 2 +[3] Omer Bar-Tal, Hila Chefer, Omer Tov, Charles Herrmann, Roni Paiss, Shiran Zada, Ariel Ephrat, Junhwa Hur, Yuanzhen Li, Tomer Michaeli, et al. Lumiere: A space-time diffusion model for video generation. arXiv preprint arXiv:2401.12945, 2024. 2 +[4] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2, 4, 5 +[5] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. 2024. 2 +[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 5 +[7] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In European conference on computer vision, pages 17-33. Springer, 2022. 6 +[8] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 4 +[9] Weixin Chen, Dawn Song, and Bo Li. Trojdiff: Trojan attacks on diffusion models with diverse targets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4035-4044, 2023. 2 +[10] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In IEEE Conference on Computer Vision and Pattern Recognition, 2021. Computer Vision Foundation / IEEE, 2021. 4 +[11] June Suk Choi, Kyungmin Lee, Jongheon Jeong, Saining Xie, Jinwoo Shin, and Kimin Lee. Diffusionguard: A robust defense against malicious diffusion-based image editing. In ICML 2024 Next Generation of AI Safety Workshop. 2 +[12] Sheng-Yen Chou, Pin-Yu Chen, and Tsung-Yi Ho. How to backdoor diffusion models? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4015-4024, 2023. 2 +[13] Hai Ci, Yiren Song, Pei Yang, Jinheng Xie, and Mike Zheng Shou. Wmadapter: Adding watermark control to latent diffusion models. arXiv preprint arXiv:2406.08337, 2024. 2 +[14] Özgün Çiçek, Ahmed Abdulkadir, Soeren S Lienkamp, Thomas Brox, and Olaf Ronneberger. 3d u-net: learning dense volumetric segmentation from sparse annotation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2016: 19th International Conference, + +Athens, Greece, October 17-21, 2016, Proceedings, Part II 19, pages 424-432. Springer, 2016. 2 +[15] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2 +[16] Weitao Feng, Wenbo Zhou, Jiyan He, Jie Zhang, Tianyi Wei, Guanlin Li, Tianwei Zhang, Weiming Zhang, and Nenghai Yu. Aqualora: Toward white-box protection for customized stable diffusion models via watermark lora. arXiv preprint arXiv:2405.11135, 2024. 2 +[17] Pierre Fernandez, Guillaume Couairon, Hervé Jégou, Matthijs Douze, and Teddy Furon. The stable signature: Rooting watermarks in latent diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22466-22477, 2023. 2 +[18] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Imagine your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 2 +[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2 +[20] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsanko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2 +[21] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 2 +[22] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 2 +[23] Robert Honig, Javier Rando, Nicholas Carlini, and Florian Tramér. Adversarial perturbations cannot reliably protect artists from generative ai. arXiv preprint arXiv:2406.12027, 2024. 3 +[24] Li Hu. Animate anyone: Consistent and controllable image-to-video synthesis for character animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8153-8163, 2024. 2 +[25] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024. 5 +[26] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5148-5157, 2021. 5 +[27] PKU-Yuan Lab and Tuzhan AI etc. Open-sora-plan, 2024. 2 +[28] Guanlin Li, Shuai Yang, Jie Zhang, and Tianwei Zhang. Prime: Protect your videos from malicious editing. arXiv preprint arXiv:2402.01239, 2024. 3, 5 + +[29] Chumeng Liang and Xiaoyu Wu. Mist: Towards improved adversarial examples for diffusion models. arXiv preprint arXiv:2305.12683, 2023. 2 +[30] Chumeng Liang, Xiaoyu Wu, Yang Hua, Jiaru Zhang, Yiming Xue, Tao Song, Zhengui Xue, Ruhui Ma, and Haibing Guan. Adversarial example does good: Preventing painting imitation from diffusion models via adversarial examples. arXiv preprint arXiv:2302.04578, 2023. 2 +[31] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International conference on machine learning, pages 8162-8171. PMLR, 2021. 2 +[32] Josephine Passananti, Stanley Wu, Shawn Shan, Haitao Zheng, and Ben Y Zhao. Disrupting style mimicry attacks on video imagery. arXiv preprint arXiv:2405.06865, 2024. 3 +[33] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 2 +[34] Bohao Peng, Jian Wang, Yuechen Zhang, Wenbo Li, MingChang Yang, and Jiaya Jia. Controlnext: Powerful and efficient control for image and video generation. arXiv preprint arXiv:2408.06070, 2024. 2, 4, 5 +[35] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2 +[36] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 2 +[37] Hadi Salman, Alaa Khaddaj, Guillaume Leclerc, Andrew Ilyas, and Aleksander Madry. Raising the cost of malicious ai-powered image editing. In Proceedings of the 40th International Conference on Machine Learning, pages 29894–29918, 2023. 2, 4, 5, 6 +[38] Shawn Shan, Jenna Cryan, Emily Wenger, Haitao Zheng, Rana Hanocka, and Ben Y Zhao. Glaze: Protecting artists from style mimicry by {Text-to-Image} models. In 32nd USENIX Security Symposium (USENIX Security 23), pages 2187–2204, 2023. 2 +[39] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2 +[40] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 2 +[41] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2 +[42] Jingwei Sun, Xuchong Zhang, Changfeng Sun, Qicheng Bai, and Hongbin Sun. Latent feature and attention dual erasure + +attack against multi-view diffusion models for 3d assets protection. arXiv preprint arXiv:2408.11408, 2024. 3 +[43] Kling AI Team. Kling ai, 2024. 2 +[44] Luma AI Team. Luma ai, 2024. +[45] The Movie Gen team @ Meta. Movie gen: A cast of media foundation models. 2024. 2 +[46] Vu Tuan Truong, Luan Ba Dang, and Long Bao Le. Attacks and defenses for generative diffusion models: A comprehensive survey. arXiv preprint arXiv:2408.03400, 2024. 2 +[47] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 2 +[48] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 2 +[49] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 2 +[50] Yuxin Wen, John Kirchenbauer, Jonas Geiping, and Tom Goldstein. Tree-ring watermarks: Fingerprints for diffusion images that are invisible and robust. arXiv preprint arXiv:2305.20030, 2023. 2 +[51] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7623-7633, 2023. 2 +[52] Zhen Xing, Qijun Feng, Haoran Chen, Qi Dai, Han Hu, Hang Xu, Zuxuan Wu, and Yu-Gang Jiang. A survey on video diffusion models. ACM Computing Surveys, 2023. 2 +[53] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 2, 4, 5 +[54] Jianping Zhang, Zhuoer Xu, Shiwen Cui, Changhua Meng, Weibin Wu, and Michael R Lyu. On the robustness of latent diffusion models. arXiv preprint arXiv:2306.08257, 2023. 2, 4 +[55] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all, 2024. 2 +[56] Peifei Zhu, Tsubasa Takahashi, and Hirokatsu Kataoka. Watermark-embedded adversarial examples for copyright protection against diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24420–24430, 2024. 2 \ No newline at end of file diff --git a/2025/I2VGuard_ Safeguarding Images against Misuse in Diffusion-based Image-to-Video Models/images.zip b/2025/I2VGuard_ Safeguarding Images against Misuse in Diffusion-based Image-to-Video Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..579196012558ccbb65d1bb5ff2ab56c31191195c --- /dev/null +++ b/2025/I2VGuard_ Safeguarding Images against Misuse in Diffusion-based Image-to-Video Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:96cc6d8fafab5e24ec1784c8d15f693c5828742c3d6a44da358ec1ee452bc10b +size 913060 diff --git a/2025/I2VGuard_ Safeguarding Images against Misuse in Diffusion-based Image-to-Video Models/layout.json b/2025/I2VGuard_ Safeguarding Images against Misuse in Diffusion-based Image-to-Video Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e77485be0c102ff380f544e3b95eac5c91a20a0e --- /dev/null +++ b/2025/I2VGuard_ Safeguarding Images against Misuse in Diffusion-based Image-to-Video Models/layout.json @@ -0,0 +1,11592 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 90, + 103, + 504, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 103, + 504, + 140 + ], + "spans": [ + { + "bbox": [ + 90, + 103, + 504, + 140 + ], + "type": "text", + "content": "I2VGuard: Safeguarding Images against Misuse in Diffusion-based Image-to-Video Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 135, + 160, + 214, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 160, + 214, + 175 + ], + "spans": [ + { + "bbox": [ + 135, + 160, + 214, + 175 + ], + "type": "text", + "content": "Dongnan Gui" + }, + { + "bbox": [ + 135, + 160, + 214, + 175 + ], + "type": "inline_equation", + "content": "^{1,*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 235, + 161, + 286, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 161, + 286, + 175 + ], + "spans": [ + { + "bbox": [ + 235, + 161, + 286, + 175 + ], + "type": "text", + "content": "Xun Guo2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 310, + 161, + 390, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 161, + 390, + 175 + ], + "spans": [ + { + "bbox": [ + 310, + 161, + 390, + 175 + ], + "type": "text", + "content": "Wengang Zhou" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 415, + 161, + 457, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 415, + 161, + 457, + 175 + ], + "spans": [ + { + "bbox": [ + 415, + 161, + 457, + 175 + ], + "type": "text", + "content": "Yan Lu2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 175, + 339, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 175, + 339, + 190 + ], + "spans": [ + { + "bbox": [ + 105, + 175, + 339, + 190 + ], + "type": "text", + "content": "1University of Science and Technology of China" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 363, + 175, + 489, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 175, + 489, + 188 + ], + "spans": [ + { + "bbox": [ + 363, + 175, + 489, + 188 + ], + "type": "text", + "content": "2Microsoft Research Asia" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 126, + 191, + 465, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 191, + 465, + 203 + ], + "spans": [ + { + "bbox": [ + 126, + 191, + 465, + 203 + ], + "type": "text", + "content": "{gdn2001@mail.,zhwg@}ustc.edu.cn{xunguo，yanlu}@microsoft.com" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 51, + 216, + 149, + 273 + ], + "blocks": [ + { + "bbox": [ + 51, + 216, + 149, + 273 + ], + "lines": [ + { + "bbox": [ + 51, + 216, + 149, + 273 + ], + "spans": [ + { + "bbox": [ + 51, + 216, + 149, + 273 + ], + "type": "image", + "image_path": "cd5a9666902b55dad978753a14b6623d7bc761247732aebf941009b5ed30e5a7.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 74, + 272, + 126, + 280 + ], + "lines": [ + { + "bbox": [ + 74, + 272, + 126, + 280 + ], + "spans": [ + { + "bbox": [ + 74, + 272, + 126, + 280 + ], + "type": "text", + "content": "Original Image" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 152, + 216, + 250, + 272 + ], + "blocks": [ + { + "bbox": [ + 152, + 216, + 250, + 272 + ], + "lines": [ + { + "bbox": [ + 152, + 216, + 250, + 272 + ], + "spans": [ + { + "bbox": [ + 152, + 216, + 250, + 272 + ], + "type": "image", + "image_path": "b190dee41c14ae9a45c74f58645879e1586b85acf5240a675ff4848cb697eae2.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 250, + 216, + 348, + 272 + ], + "blocks": [ + { + "bbox": [ + 250, + 216, + 348, + 272 + ], + "lines": [ + { + "bbox": [ + 250, + 216, + 348, + 272 + ], + "spans": [ + { + "bbox": [ + 250, + 216, + 348, + 272 + ], + "type": "image", + "image_path": "051a828e773f4ee3851ddc9f28a8a74f795e83c7df6a5b1ad00e8c461a1c1fc4.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 347, + 216, + 445, + 272 + ], + "blocks": [ + { + "bbox": [ + 347, + 216, + 445, + 272 + ], + "lines": [ + { + "bbox": [ + 347, + 216, + 445, + 272 + ], + "spans": [ + { + "bbox": [ + 347, + 216, + 445, + 272 + ], + "type": "image", + "image_path": "cbe52968a8b12665d23cd911ca5608df22960afc66255a44296a065ecff0ed9c.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 445, + 216, + 543, + 272 + ], + "blocks": [ + { + "bbox": [ + 445, + 216, + 543, + 272 + ], + "lines": [ + { + "bbox": [ + 445, + 216, + 543, + 272 + ], + "spans": [ + { + "bbox": [ + 445, + 216, + 543, + 272 + ], + "type": "image", + "image_path": "f1edf07cec993b7d1c4c97a6ca1d8a314aa29a5a280a5b87f782a073f43977cb.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 51, + 280, + 149, + 335 + ], + "blocks": [ + { + "bbox": [ + 51, + 280, + 149, + 335 + ], + "lines": [ + { + "bbox": [ + 51, + 280, + 149, + 335 + ], + "spans": [ + { + "bbox": [ + 51, + 280, + 149, + 335 + ], + "type": "image", + "image_path": "f9a319aa8b7a334532a2d2bceff2c427f58e4a804e497353b5fed06c19dfcf5b.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 75, + 335, + 129, + 344 + ], + "lines": [ + { + "bbox": [ + 75, + 335, + 129, + 344 + ], + "spans": [ + { + "bbox": [ + 75, + 335, + 129, + 344 + ], + "type": "text", + "content": "Guarded Image" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 152, + 280, + 250, + 335 + ], + "blocks": [ + { + "bbox": [ + 152, + 280, + 250, + 335 + ], + "lines": [ + { + "bbox": [ + 152, + 280, + 250, + 335 + ], + "spans": [ + { + "bbox": [ + 152, + 280, + 250, + 335 + ], + "type": "image", + "image_path": "589d90852a09076afc400ebc93d4e1f68a816aa23cab1c393a3ce95c57a661a3.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 250, + 280, + 348, + 335 + ], + "blocks": [ + { + "bbox": [ + 283, + 273, + 413, + 281 + ], + "lines": [ + { + "bbox": [ + 283, + 273, + 413, + 281 + ], + "spans": [ + { + "bbox": [ + 283, + 273, + 413, + 281 + ], + "type": "text", + "content": "Generated video from original image." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 250, + 280, + 348, + 335 + ], + "lines": [ + { + "bbox": [ + 250, + 280, + 348, + 335 + ], + "spans": [ + { + "bbox": [ + 250, + 280, + 348, + 335 + ], + "type": "image", + "image_path": "cef38a4576704821a26cf49c350e4aeec3a722a3d431d59bd8b57dbd5d5c0a3b.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 282, + 335, + 414, + 344 + ], + "lines": [ + { + "bbox": [ + 282, + 335, + 414, + 344 + ], + "spans": [ + { + "bbox": [ + 282, + 335, + 414, + 344 + ], + "type": "text", + "content": "Generated video from guarded image." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 347, + 280, + 445, + 335 + ], + "blocks": [ + { + "bbox": [ + 347, + 280, + 445, + 335 + ], + "lines": [ + { + "bbox": [ + 347, + 280, + 445, + 335 + ], + "spans": [ + { + "bbox": [ + 347, + 280, + 445, + 335 + ], + "type": "image", + "image_path": "f70473bfbc3bab0cc2728e2266eb7ef5280c4ce76af2b67d690dfac815ea1513.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 445, + 280, + 543, + 335 + ], + "blocks": [ + { + "bbox": [ + 445, + 280, + 543, + 335 + ], + "lines": [ + { + "bbox": [ + 445, + 280, + 543, + 335 + ], + "spans": [ + { + "bbox": [ + 445, + 280, + 543, + 335 + ], + "type": "image", + "image_path": "175a6c6eb12240bd22927794a1d8232bbd16d3fd7e6cca83b63190e100e6d4aa.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 51, + 349, + 149, + 405 + ], + "blocks": [ + { + "bbox": [ + 51, + 349, + 149, + 405 + ], + "lines": [ + { + "bbox": [ + 51, + 349, + 149, + 405 + ], + "spans": [ + { + "bbox": [ + 51, + 349, + 149, + 405 + ], + "type": "image", + "image_path": "5f3acdcdfb3a1d63d34e5882c9402bd32ccee10bf1547777cfb3af1f6e58bac2.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 74, + 404, + 126, + 413 + ], + "lines": [ + { + "bbox": [ + 74, + 404, + 126, + 413 + ], + "spans": [ + { + "bbox": [ + 74, + 404, + 126, + 413 + ], + "type": "text", + "content": "Original Image" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 152, + 348, + 250, + 404 + ], + "blocks": [ + { + "bbox": [ + 152, + 348, + 250, + 404 + ], + "lines": [ + { + "bbox": [ + 152, + 348, + 250, + 404 + ], + "spans": [ + { + "bbox": [ + 152, + 348, + 250, + 404 + ], + "type": "image", + "image_path": "512509244fecfd83cd0928919853d3e5ddd552e1fed217f270119af7b49c05b5.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 250, + 348, + 348, + 403 + ], + "blocks": [ + { + "bbox": [ + 250, + 348, + 348, + 403 + ], + "lines": [ + { + "bbox": [ + 250, + 348, + 348, + 403 + ], + "spans": [ + { + "bbox": [ + 250, + 348, + 348, + 403 + ], + "type": "image", + "image_path": "33dceaebe1057b006cefbab2d6a5950991aef561b10c4e464af994cdf065e5f6.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 282, + 404, + 411, + 412 + ], + "lines": [ + { + "bbox": [ + 282, + 404, + 411, + 412 + ], + "spans": [ + { + "bbox": [ + 282, + 404, + 411, + 412 + ], + "type": "text", + "content": "Generated video from original image." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 347, + 348, + 445, + 403 + ], + "blocks": [ + { + "bbox": [ + 347, + 348, + 445, + 403 + ], + "lines": [ + { + "bbox": [ + 347, + 348, + 445, + 403 + ], + "spans": [ + { + "bbox": [ + 347, + 348, + 445, + 403 + ], + "type": "image", + "image_path": "22e38137629a7877bed27ee515b2f9ea8363b15c368f8f601ff01fda0595890f.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 445, + 348, + 543, + 403 + ], + "blocks": [ + { + "bbox": [ + 445, + 348, + 543, + 403 + ], + "lines": [ + { + "bbox": [ + 445, + 348, + 543, + 403 + ], + "spans": [ + { + "bbox": [ + 445, + 348, + 543, + 403 + ], + "type": "image", + "image_path": "a0e2340d8dc6e79fa2720cce58e46cfe5d0e1d7c139c11462c689a7806f349c4.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 51, + 413, + 149, + 468 + ], + "blocks": [ + { + "bbox": [ + 51, + 413, + 149, + 468 + ], + "lines": [ + { + "bbox": [ + 51, + 413, + 149, + 468 + ], + "spans": [ + { + "bbox": [ + 51, + 413, + 149, + 468 + ], + "type": "image", + "image_path": "def9bfe6a5187b68c1ac5b4c089ed6741f1e34b41b7ab85de8b424622890ddc4.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 75, + 468, + 129, + 477 + ], + "lines": [ + { + "bbox": [ + 75, + 468, + 129, + 477 + ], + "spans": [ + { + "bbox": [ + 75, + 468, + 129, + 477 + ], + "type": "text", + "content": "Guarded Image" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 45, + 486, + 547, + 510 + ], + "lines": [ + { + "bbox": [ + 45, + 486, + 547, + 510 + ], + "spans": [ + { + "bbox": [ + 45, + 486, + 547, + 510 + ], + "type": "text", + "content": "Figure 1. Results of our I2VGuard. We present original images, guarded images, and their corresponding SVD-generated videos. All results are generated with the same seed. Our method effectively safeguards images from animation in image-to-video generation." + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 152, + 413, + 250, + 468 + ], + "blocks": [ + { + "bbox": [ + 152, + 413, + 250, + 468 + ], + "lines": [ + { + "bbox": [ + 152, + 413, + 250, + 468 + ], + "spans": [ + { + "bbox": [ + 152, + 413, + 250, + 468 + ], + "type": "image", + "image_path": "c0bac01a0e9ca711bb9935d7687b3e6f683ce4a02bb58963e625b6b1cefd9c36.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 250, + 413, + 348, + 468 + ], + "blocks": [ + { + "bbox": [ + 250, + 413, + 348, + 468 + ], + "lines": [ + { + "bbox": [ + 250, + 413, + 348, + 468 + ], + "spans": [ + { + "bbox": [ + 250, + 413, + 348, + 468 + ], + "type": "image", + "image_path": "5494f90f7e7b67167e5fa5c9f2a272746a3463b009a1fd1fdc25f4a87a0a6b06.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 282, + 468, + 414, + 477 + ], + "lines": [ + { + "bbox": [ + 282, + 468, + 414, + 477 + ], + "spans": [ + { + "bbox": [ + 282, + 468, + 414, + 477 + ], + "type": "text", + "content": "Generated video from guarded image." + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 347, + 413, + 445, + 468 + ], + "blocks": [ + { + "bbox": [ + 347, + 413, + 445, + 468 + ], + "lines": [ + { + "bbox": [ + 347, + 413, + 445, + 468 + ], + "spans": [ + { + "bbox": [ + 347, + 413, + 445, + 468 + ], + "type": "image", + "image_path": "5e2c7109f832796edcd5af975f50944984abdb2bf07d6def1d09a6b1eabfa2d2.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 445, + 413, + 543, + 468 + ], + "blocks": [ + { + "bbox": [ + 445, + 413, + 543, + 468 + ], + "lines": [ + { + "bbox": [ + 445, + 413, + 543, + 468 + ], + "spans": [ + { + "bbox": [ + 445, + 413, + 543, + 468 + ], + "type": "image", + "image_path": "1b71dfa1a84a6fb6a373ad966290ca33dba26e0d2b220cbac53a67ccdc803798.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "bbox": [ + 143, + 519, + 192, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 519, + 192, + 532 + ], + "spans": [ + { + "bbox": [ + 143, + 519, + 192, + 532 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 45, + 545, + 290, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 545, + 290, + 689 + ], + "spans": [ + { + "bbox": [ + 45, + 545, + 290, + 689 + ], + "type": "text", + "content": "Recent advances in image-to-video generation have enabled animation of still images and offered pixel-level controllability. While these models hold great potential to transform single images into vivid and dynamic videos, they also carry risks of misuse that could impact privacy, security, and copyright protection. This paper proposes a novel approach that applies imperceptible perturbations on images to degrade the quality of the generated videos, thereby protecting images from misuse in white-box image-to-video diffusion models. Specifically, we function our approach as an adversarial attack, incorporating spatial, temporal, and diffusion attack modules. The spatial attack shifts image features from" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 304, + 520, + 548, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 520, + 548, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 520, + 548, + 689 + ], + "type": "text", + "content": "their original distribution to a lower-quality target distribution, reducing visual fidelity. The temporal attack disrupts coherent motion by interfering with temporal attention maps that guide motion generation. To enhance the robustness of our approach across different models, we further propose a diffusion attack module leveraging contrastive loss. Our approach can be easily integrated with mainstream diffusion-based I2V models. Extensive experiments on SVD, CogVideoX, and ControlNeXt demonstrate that our method significantly impairs generation quality in terms of visual clarity and motion consistency, while introducing only minimal artifacts to the images. To the best of our knowledge, we are the first to explore adversarial attacks on image-to-video generation for security purposes." + } + ] + } + ], + "index": 41 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 702, + 237, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 702, + 237, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 702, + 237, + 714 + ], + "type": "text", + "content": "*Work done during internship at Microsoft Research Asia." + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 294, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 758 + ], + "type": "text", + "content": "12595" + } + ] + } + ], + "index": 43 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 49, + 72, + 127, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 72, + 127, + 83 + ], + "spans": [ + { + "bbox": [ + 49, + 72, + 127, + 83 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 91, + 287, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 91, + 287, + 281 + ], + "spans": [ + { + "bbox": [ + 49, + 91, + 287, + 281 + ], + "type": "text", + "content": "Diffusion models [15, 19, 35] have gained significant attention recently, particularly in video generation. Advances in models like Sora [5] and its DiT [33] framework have accelerated the development of video diffusion models, leading to numerous academic [4, 22, 27, 53, 55] and commercial [43-45] applications. The primary types of video generation are text-to-video(T2V) and image-to-video(I2V), with the former supporting greater diversity and the latter offering pixel-level control. Leveraging powerful generative models and techniques from text-to-image generation [35], image-to-video models can now produce visually impressive videos. Other conditional I2V models include CogVideoX [53], which generates videos using both text and image inputs, enabling more versatile content creation, and ControlNeXt [34], which focuses on generating videos conditioned on pose information, offering precise control over motion and structure." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 283, + 287, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 283, + 287, + 461 + ], + "spans": [ + { + "bbox": [ + 49, + 283, + 287, + 461 + ], + "type": "text", + "content": "However, the rapid advancement of generative models has also heightened risks related to privacy, security, and copyright [2]. For instance, an adversary with access to a person's image could use image-to-video models to generate misleading or harmful videos, depicting actions the individual never performed, thus compromising individual's privacy and security. Additionally, the unauthorized use of copyrighted images to create animations poses significant risks to intellectual property rights. To address these risks, this work introduces an adversarial attack method designed to prevent the misuse of images in diffusion-based I2V models. Prior research has primarily focused on protecting images from malicious edits [37, 54], which are often viewed as adversarial attacks on image-editing models. Few efforts have focused on safeguarding images from misuse in I2V tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 463, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 463, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 463, + 287, + 712 + ], + "type": "text", + "content": "The main challenges in attacking I2V models come from two aspects: first, disrupting temporal consistency with perturbations applied only to individual images; and second, achieving robustness across diverse diffusion models, including UNet and DiT architectures, even under strong conditioning such as poses and text prompts. To address these challenges, our method targets different modules within I2V models and affects various attributes of the generated video. Specifically, we divide our approach into spatial, temporal, and diffusion attacks. The spatial attack focuses on altering the spatial features of the I2V models by targeting the VAE encoder in latent diffusion models. Here, we manipulate image latents, shifting them toward suboptimal values. The temporal attack aims to perturb the temporal features, which are crucial for video motion and consistency. By extracting the temporal attention map, we intentionally disrupt it by pushing it away from its original attention map. This creates a chaotic attention map, leading to temporal inconsistencies in the generated video. The diffusion attack focuses on the diffusion process and targets the output of the denoising module, which includes the UNet in SVD and the trans" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 307, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 156 + ], + "type": "text", + "content": "former in CogVideoX. Unlike a standard denoising process, our approach aims to push the predicted outputs away from the original frames, guiding them closer to target frames. To achieve this, we design a contrastive loss that degrades the quality of the predicted frames by treating the target frames as positive samples and the original predicted frames as negative samples." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 308, + 157, + 545, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 157, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 308, + 157, + 545, + 274 + ], + "type": "text", + "content": "We conduct a comprehensive qualitative and quantitative analysis of the attack results. Qualitative analysis demonstrates that our method effectively disrupts the original image, leading to low temporal consistency and reduced aesthetic quality in both simple and conditioned I2V models. Quantitative analysis confirms a significant degradation in both spatial quality and temporal consistency of the generated videos. To our knowledge, this is the first adversarial attack method aimed at securing images against misuse in I2V generation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 308, + 285, + 391, + 297 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 285, + 391, + 297 + ], + "spans": [ + { + "bbox": [ + 308, + 285, + 391, + 297 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 308, + 305, + 545, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 305, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 308, + 305, + 545, + 518 + ], + "type": "text", + "content": "Image-to-Video Generation With the great advancements in image generation [15, 19, 31, 35, 40, 41], video generation [3, 5, 20, 22, 39, 43–45, 52, 53] has gathered significant attention and is developing rapidly [18, 48, 49, 51]. Recently, following the success of diffusion models [15, 19, 35, 40] in image generation, diffusion models integrated with UNet [3, 14, 21, 36] or transformers (DiT) [5, 33, 47, 55] have been adopted for video generation. In the field of imaged-to-video generation,AnimateDiff [18] animates personalized text-to-image diffusion models by inserting a temporal motion module, Stable Video Diffusion(SVD) [4] demonstrates strong performance, benefiting from intensive training on large amounts of high-quality data, CogVideoX [53] supports text-image jointly conditional video generation,Animate Anyone [24] and ControlNeXt [34] leverage pose information for controllable image-to-video generation. In this work, we focus on I2V models, which are susceptible to misuse and therefore require robust protection." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 308, + 534, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 534, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 534, + 545, + 712 + ], + "type": "text", + "content": "Adversarial Attacks on Diffusion Models Current attacks on diffusion models, which can be broadly categorized [46] into backdoor [9, 12] and adversarial attacks, primarily target image-based diffusion models [13, 16, 17, 50]. In adversarial attacks, the input image or text prompt is perturbed to generate adversarial examples. For instance, AdvDM [30] and Mist [29] utilize adversarial examples to protect human-created artworks, PhotoGuard [37] alters the image by modifying either the encoder or the diffusion outputs against malicious editing, while Glaze [38] perturbs the image to prevent style mimicry, [54] shifts the image away from its original distribution, [56] embeds personal watermarks in the generation of adversarial examples, and DiffusionGuard [11] generates adversarial noise targeting the early stage of the diffusion process. For video-based attacks," + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "type": "text", + "content": "12596" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 80, + 539, + 322 + ], + "blocks": [ + { + "bbox": [ + 58, + 80, + 539, + 322 + ], + "lines": [ + { + "bbox": [ + 58, + 80, + 539, + 322 + ], + "spans": [ + { + "bbox": [ + 58, + 80, + 539, + 322 + ], + "type": "image", + "image_path": "97cf788ca4fa741f68aa6c8c0d9f9e680a5bc2d9eab9c2728ff5ec1e1d2849d4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 332, + 548, + 399 + ], + "lines": [ + { + "bbox": [ + 46, + 332, + 548, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 332, + 548, + 399 + ], + "type": "text", + "content": "Figure 2. Overview of our proposed method. The training process begins with a trainable copy " + }, + { + "bbox": [ + 46, + 332, + 548, + 399 + ], + "type": "inline_equation", + "content": "I_{adv}" + }, + { + "bbox": [ + 46, + 332, + 548, + 399 + ], + "type": "text", + "content": " of the original image " + }, + { + "bbox": [ + 46, + 332, + 548, + 399 + ], + "type": "inline_equation", + "content": "I_{src}" + }, + { + "bbox": [ + 46, + 332, + 548, + 399 + ], + "type": "text", + "content": ". First, we perform inference to obtain the original video " + }, + { + "bbox": [ + 46, + 332, + 548, + 399 + ], + "type": "inline_equation", + "content": "V_{0}" + }, + { + "bbox": [ + 46, + 332, + 548, + 399 + ], + "type": "text", + "content": " and get the latent frames " + }, + { + "bbox": [ + 46, + 332, + 548, + 399 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{0}" + }, + { + "bbox": [ + 46, + 332, + 548, + 399 + ], + "type": "text", + "content": ". Noisy latent frames " + }, + { + "bbox": [ + 46, + 332, + 548, + 399 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{t}" + }, + { + "bbox": [ + 46, + 332, + 548, + 399 + ], + "type": "text", + "content": " and the latent image " + }, + { + "bbox": [ + 46, + 332, + 548, + 399 + ], + "type": "inline_equation", + "content": "x_{adv}, x_{src}" + }, + { + "bbox": [ + 46, + 332, + 548, + 399 + ], + "type": "text", + "content": " are then processed by the denoising model to predict the original frames. The encoded latent image " + }, + { + "bbox": [ + 46, + 332, + 548, + 399 + ], + "type": "inline_equation", + "content": "x_{adv}" + }, + { + "bbox": [ + 46, + 332, + 548, + 399 + ], + "type": "text", + "content": " is utilized for the spatial encoder attack, while the predicted original frames " + }, + { + "bbox": [ + 46, + 332, + 548, + 399 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}_{0}" + }, + { + "bbox": [ + 46, + 332, + 548, + 399 + ], + "type": "text", + "content": " serve in the diffusion attack to compute contrastive loss. Within the denoising module, we hook into the temporal attention module to extract the temporal attention map. By altering the attacked attention map " + }, + { + "bbox": [ + 46, + 332, + 548, + 399 + ], + "type": "inline_equation", + "content": "A_{adv}" + }, + { + "bbox": [ + 46, + 332, + 548, + 399 + ], + "type": "text", + "content": " to diverge from the original " + }, + { + "bbox": [ + 46, + 332, + 548, + 399 + ], + "type": "inline_equation", + "content": "A_{src}" + }, + { + "bbox": [ + 46, + 332, + 548, + 399 + ], + "type": "text", + "content": ", we implement the temporal attack." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 419, + 289, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 419, + 289, + 587 + ], + "spans": [ + { + "bbox": [ + 46, + 419, + 289, + 587 + ], + "type": "text", + "content": "PRIME [28] modifies videos frame-by-frame to prevent malicious video editing, while [32] disrupts style mimicry attacks from video imagery. There are also works that protect the image against other generation, such as text-to-3D generation [42]. In this work, we focus on perturbing the input image in image-to-video diffusion models to perform adversarial attacks. Recent study [23] show that adversarial perturbations fail to reliably protect artists from style mimicry in the field of image generation. Our approach stands out in two ways: it uses a white-box attack strategy, allowing adversarial noise training with both generation and pre-processing models, and it focuses on disrupting temporal consistency rather than mere image fidelity, making the perturbations harder to remove." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 606, + 103, + 619 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 606, + 103, + 619 + ], + "spans": [ + { + "bbox": [ + 47, + 606, + 103, + 619 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "In this section, we first provide preliminaries on latent diffusion models and problem settings in Sec. 3.1. We then present our attack method, detailing the spatial attack in Sec. 3.2, the temporal attack in Sec. 3.3, and the diffusion attack in Sec. 3.4. An overview of our method is presented in in Fig. 2, which shows our protection procedure involves spatial, temporal, and diffusion attacks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 419, + 385, + 432 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 419, + 385, + 432 + ], + "spans": [ + { + "bbox": [ + 306, + 419, + 385, + 432 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 461, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 461, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 548, + 715 + ], + "type": "text", + "content": "Latent Diffusion Models In diffusion models, realistic image or video sampling from a target distribution " + }, + { + "bbox": [ + 304, + 461, + 548, + 715 + ], + "type": "inline_equation", + "content": "q(\\cdot)" + }, + { + "bbox": [ + 304, + 461, + 548, + 715 + ], + "type": "text", + "content": " is achieved via a sequence of denoising steps in a diffusion process. This approach involves iteratively adding and removing noise to transform an initial sample " + }, + { + "bbox": [ + 304, + 461, + 548, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0 \\sim q(\\cdot)" + }, + { + "bbox": [ + 304, + 461, + 548, + 715 + ], + "type": "text", + "content": ". Noise is incrementally introduced across steps, creating intermediate states defined by " + }, + { + "bbox": [ + 304, + 461, + 548, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{t+1} = \\sqrt{\\overline{\\alpha}_t} \\mathbf{x}_0 + \\sqrt{1 - \\overline{\\alpha}_t} \\epsilon_t" + }, + { + "bbox": [ + 304, + 461, + 548, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 461, + 548, + 715 + ], + "type": "inline_equation", + "content": "\\epsilon_t" + }, + { + "bbox": [ + 304, + 461, + 548, + 715 + ], + "type": "text", + "content": " is Gaussian noise. The parameters " + }, + { + "bbox": [ + 304, + 461, + 548, + 715 + ], + "type": "inline_equation", + "content": "\\alpha_t" + }, + { + "bbox": [ + 304, + 461, + 548, + 715 + ], + "type": "text", + "content": " control the noise transition, leading to a final noisy state with a standard Gaussian distribution. By reversing this process and iteratively denoising each state, the model learns to predict the noise added at each step, ultimately enabling the recovery of samples from " + }, + { + "bbox": [ + 304, + 461, + 548, + 715 + ], + "type": "inline_equation", + "content": "q(\\cdot)" + }, + { + "bbox": [ + 304, + 461, + 548, + 715 + ], + "type": "text", + "content": " starting from " + }, + { + "bbox": [ + 304, + 461, + 548, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T" + }, + { + "bbox": [ + 304, + 461, + 548, + 715 + ], + "type": "text", + "content": ". Latent diffusion models (LDMs), a subclass of diffusion models, operate in a compressed latent space rather than directly in the image space. In LDMs, each input image or video " + }, + { + "bbox": [ + 304, + 461, + 548, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 304, + 461, + 548, + 715 + ], + "type": "text", + "content": " is first mapped to a latent representation " + }, + { + "bbox": [ + 304, + 461, + 548, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(\\mathbf{x}_0)" + }, + { + "bbox": [ + 304, + 461, + 548, + 715 + ], + "type": "text", + "content": " via an encoder " + }, + { + "bbox": [ + 304, + 461, + 548, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 304, + 461, + 548, + 715 + ], + "type": "text", + "content": ". The diffusion process then applies in this latent space, producing a series of noisy representations. Training in this latent space enables LDMs to achieve efficient training and faster inference while retaining high-quality outputs." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12597" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": "Problem Settings Adversarial attacks involve adding subtle, nearly undetectable perturbations " + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "inline_equation", + "content": "\\delta_{adv}" + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": " to the input data to influence a model's output. In computer vision, adversarial attacks have been widely explored in classifications and image synthesis. However, in image-to-video generation tasks, adversarial attack remains a relatively unexplored area. In this work, we aim to develop adversarial attacks to safeguard images from misused by image-to-video latent diffusion models. To distinguish between the latent and pixel spaces, we represent images and videos in pixel space as " + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": ", and in latent space as " + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 204, + 289, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 204, + 289, + 277 + ], + "spans": [ + { + "bbox": [ + 47, + 204, + 289, + 277 + ], + "type": "text", + "content": "Specifically, an adversary can generate a modified image " + }, + { + "bbox": [ + 47, + 204, + 289, + 277 + ], + "type": "inline_equation", + "content": "I_{adv} = I_{src} + \\delta" + }, + { + "bbox": [ + 47, + 204, + 289, + 277 + ], + "type": "text", + "content": " from source image " + }, + { + "bbox": [ + 47, + 204, + 289, + 277 + ], + "type": "inline_equation", + "content": "I_{src}" + }, + { + "bbox": [ + 47, + 204, + 289, + 277 + ], + "type": "text", + "content": ", crafted to disrupt the generative model " + }, + { + "bbox": [ + 47, + 204, + 289, + 277 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 47, + 204, + 289, + 277 + ], + "type": "text", + "content": ", resulting in poor-quality or distorted outputs. This adversarial image is optimized by maximizing the loss between the videos generated from the original and modified images, represented as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 285, + 279, + 306 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 285, + 279, + 306 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 279, + 306 + ], + "type": "interline_equation", + "content": "\\delta_ {a d v} = \\arg \\max _ {\\| \\delta \\| _ {p} \\leq \\epsilon} \\mathcal {L} (\\mathcal {G} (I _ {s r c} + \\delta , \\mathbf {X} _ {t}, c), \\mathcal {G} (I _ {s r c}, \\mathbf {X} _ {t}, c))", + "image_path": "664f11d8d71b46b5ce8727f09e1623ecf4cee78cf7d68866045e90bf72a38591.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 316, + 289, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 289, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 289, + 376 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 316, + 289, + 376 + ], + "type": "inline_equation", + "content": "\\delta_{adv}" + }, + { + "bbox": [ + 46, + 316, + 289, + 376 + ], + "type": "text", + "content": " represents the computed perturbation, constrained by " + }, + { + "bbox": [ + 46, + 316, + 289, + 376 + ], + "type": "inline_equation", + "content": "\\| \\delta \\| _p\\leq \\epsilon" + }, + { + "bbox": [ + 46, + 316, + 289, + 376 + ], + "type": "text", + "content": " to ensure minimal visibility, and " + }, + { + "bbox": [ + 46, + 316, + 289, + 376 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 316, + 289, + 376 + ], + "type": "text", + "content": " represents other conditions such as text or pose in the imaged-to-video generation process. The constraint on " + }, + { + "bbox": [ + 46, + 316, + 289, + 376 + ], + "type": "inline_equation", + "content": "\\delta_{adv}" + }, + { + "bbox": [ + 46, + 316, + 289, + 376 + ], + "type": "text", + "content": " is maintained by a regularization term " + }, + { + "bbox": [ + 46, + 316, + 289, + 376 + ], + "type": "inline_equation", + "content": "\\| I_{adv} - I_{src}\\| ^2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 383, + 138, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 383, + 138, + 396 + ], + "spans": [ + { + "bbox": [ + 47, + 383, + 138, + 396 + ], + "type": "text", + "content": "3.2. Spatial Attack" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 402, + 287, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 402, + 287, + 474 + ], + "spans": [ + { + "bbox": [ + 46, + 402, + 287, + 474 + ], + "type": "text", + "content": "The generated video has two main attributes: spatial features and temporal features. In this section, we focus on attacking the spatial features within individual frames. Previous work [37, 54] has applied adversarial attacks to image editing tasks, focusing solely on spatial features. Our work extends these methods to image-to-video generation task." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 474, + 288, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 288, + 522 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 288, + 522 + ], + "type": "text", + "content": "In Latent Diffusion Models (LDMs), an encoder, denoted as " + }, + { + "bbox": [ + 46, + 474, + 288, + 522 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 46, + 474, + 288, + 522 + ], + "type": "text", + "content": ", compresses the original image or frames into a latent space representation. An encoder attack aims to compromise this encoder, reducing the effectiveness of generative models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 107, + 533, + 287, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 533, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 107, + 533, + 287, + 548 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {e n c} = \\left\\| \\mathcal {E} \\left(I _ {a d v}\\right) - \\mathcal {E} \\left(I _ {t g}\\right) \\right\\| ^ {2} \\tag {1}", + "image_path": "c006bddefe1a4b05cc8160ae8d5e4c8778dfba83f6b6fffec113f54d7048e477.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 555, + 287, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 555, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 555, + 287, + 590 + ], + "type": "text", + "content": "Specifically, as shown in Eq. (1), the attack transforms the latent representation " + }, + { + "bbox": [ + 46, + 555, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(I_{adv})" + }, + { + "bbox": [ + 46, + 555, + 287, + 590 + ], + "type": "text", + "content": " into a poor one " + }, + { + "bbox": [ + 46, + 555, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(I_{tg})" + }, + { + "bbox": [ + 46, + 555, + 287, + 590 + ], + "type": "text", + "content": ", resulting in poor-quality final videos." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 599, + 149, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 599, + 149, + 612 + ], + "spans": [ + { + "bbox": [ + 47, + 599, + 149, + 612 + ], + "type": "text", + "content": "3.3. Temporal Attack" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "text", + "content": "Unlike attacks on image editing tasks, where no temporal features are involved, attacks on video generation rely heavily on temporal information. In this task, disrupting temporal consistency through adversarial attacks is both essential and challenging. To reduce temporal consistency and induce chaotic behavior in the outputs, we target the temporal attention maps within the temporal transformer blocks in video diffusion generator, " + }, + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "text", + "content": ". Specifically, during training, we insert" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 309, + 72, + 425, + 138 + ], + "blocks": [ + { + "bbox": [ + 309, + 72, + 425, + 138 + ], + "lines": [ + { + "bbox": [ + 309, + 72, + 425, + 138 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 425, + 138 + ], + "type": "image", + "image_path": "bc65197530caa0c39ed1dbb219022649ca06a1ae592d05301895de15a36916f1.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 148, + 545, + 182 + ], + "lines": [ + { + "bbox": [ + 304, + 148, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 304, + 148, + 545, + 182 + ], + "type": "text", + "content": "Figure 3. Temporal self-attention map visualizations. Difference between generated frames from the original image (left) and the guarded image (right) in Fig. 1." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 427, + 72, + 543, + 138 + ], + "blocks": [ + { + "bbox": [ + 427, + 72, + 543, + 138 + ], + "lines": [ + { + "bbox": [ + 427, + 72, + 543, + 138 + ], + "spans": [ + { + "bbox": [ + 427, + 72, + 543, + 138 + ], + "type": "image", + "image_path": "1258d5cfa6a7d58c434cd3dd88e56b6cbfadef2f95a2eb094cc497a9acea76b7.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 200, + 547, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 200, + 547, + 301 + ], + "spans": [ + { + "bbox": [ + 304, + 200, + 547, + 301 + ], + "type": "text", + "content": "a hook, denoted as " + }, + { + "bbox": [ + 304, + 200, + 547, + 301 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 304, + 200, + 547, + 301 + ], + "type": "text", + "content": ", within the attention block to extract the self-attention map " + }, + { + "bbox": [ + 304, + 200, + 547, + 301 + ], + "type": "inline_equation", + "content": "A_{adv}" + }, + { + "bbox": [ + 304, + 200, + 547, + 301 + ], + "type": "text", + "content": ". We then apply a targeted attack to this attention map, pushing it away from the original map, " + }, + { + "bbox": [ + 304, + 200, + 547, + 301 + ], + "type": "inline_equation", + "content": "A_{src}" + }, + { + "bbox": [ + 304, + 200, + 547, + 301 + ], + "type": "text", + "content": ", thereby reducing temporal consistency in the generated outputs. To achieve this, we introduce a temporal attention loss, denoted as " + }, + { + "bbox": [ + 304, + 200, + 547, + 301 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{att}" + }, + { + "bbox": [ + 304, + 200, + 547, + 301 + ], + "type": "text", + "content": ", as shown in Eq. (2), which induces motion-related perturbations and promotes temporal chaos." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 364, + 302, + 545, + 316 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 302, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 364, + 302, + 545, + 316 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {a t t}} = \\tau_ {2} - \\left| \\left| A _ {\\text {a d v}} - A _ {\\text {s r c}} \\right| \\right| ^ {2} \\tag {2}", + "image_path": "470862cb6032441acee93cc27e2388852f5f142915d96093f07187976c6b3c7b.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 319, + 547, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 319, + 547, + 415 + ], + "spans": [ + { + "bbox": [ + 304, + 319, + 547, + 415 + ], + "type": "text", + "content": "As shown in Fig. 3, the visualization of the temporal self-attention map on both the original and modified frames highlights this impact: the manipulated attention map misaligns with the original, directly leading to inconsistencies in motion across frames. This visual evidence underscores the effectiveness of our method in disrupting temporal coherence in video generation by distorting the temporal attention mechanism within the model." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 421, + 406, + 432 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 421, + 406, + 432 + ], + "spans": [ + { + "bbox": [ + 306, + 421, + 406, + 432 + ], + "type": "text", + "content": "3.4. Diffusion Attack" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 439, + 547, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 439, + 547, + 583 + ], + "spans": [ + { + "bbox": [ + 304, + 439, + 547, + 583 + ], + "type": "text", + "content": "Although spatial and temporal attacks can disrupt structural and motion information in generated videos, their effectiveness may vary significantly across different models [4, 34, 53], especially when other conditions, e.g., text prompts, are involved. To enhance the robustness and generalization of the perturbation across diffusion models, we introduce a latent space perturbation that uses contrastive loss to shift the predicted noise closer to the target frames while pushing it away from the original frames. Specifically, given the latent representation of the original frames, " + }, + { + "bbox": [ + 304, + 439, + 547, + 583 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_0" + }, + { + "bbox": [ + 304, + 439, + 547, + 583 + ], + "type": "text", + "content": ", and the target latent frames, " + }, + { + "bbox": [ + 304, + 439, + 547, + 583 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{tg}" + }, + { + "bbox": [ + 304, + 439, + 547, + 583 + ], + "type": "text", + "content": ", we construct a contrastive objective [8, 10], as shown in Eq. (3)." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 318, + 588, + 558, + 607 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 588, + 558, + 607 + ], + "spans": [ + { + "bbox": [ + 318, + 588, + 558, + 607 + ], + "type": "interline_equation", + "content": "\\mathbf {L} _ {c o n} = \\left\\| \\hat {\\mathbf {X}} _ {0, a d v} - \\mathbf {X} _ {t g} \\right\\| ^ {2} + \\max \\left(0, \\tau_ {1} - \\left\\| \\hat {\\mathbf {X}} _ {0, a d v} - \\mathbf {X} _ {0} \\right\\| ^ {2}\\right) (3)", + "image_path": "00dd4ce27bb78b8c7883f48220257de1598480d7c06bfff92af50964a9e738f5.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 611, + 545, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 611, + 545, + 671 + ], + "spans": [ + { + "bbox": [ + 304, + 611, + 545, + 671 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 611, + 545, + 671 + ], + "type": "inline_equation", + "content": "\\tau_{1}" + }, + { + "bbox": [ + 304, + 611, + 545, + 671 + ], + "type": "text", + "content": " is the margin distance. The positive pair consists of the predicted frames " + }, + { + "bbox": [ + 304, + 611, + 545, + 671 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}_{0,adv}" + }, + { + "bbox": [ + 304, + 611, + 545, + 671 + ], + "type": "text", + "content": " and the target frames " + }, + { + "bbox": [ + 304, + 611, + 545, + 671 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{tg}" + }, + { + "bbox": [ + 304, + 611, + 545, + 671 + ], + "type": "text", + "content": ", which we aim to bring closer together, while the negative pair is composed of the predicted frames " + }, + { + "bbox": [ + 304, + 611, + 545, + 671 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}_{0,adv}" + }, + { + "bbox": [ + 304, + 611, + 545, + 671 + ], + "type": "text", + "content": " and the original frames " + }, + { + "bbox": [ + 304, + 611, + 545, + 671 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_0" + }, + { + "bbox": [ + 304, + 611, + 545, + 671 + ], + "type": "text", + "content": ", which we aim to push apart." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 670, + 545, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 670, + 545, + 694 + ], + "spans": [ + { + "bbox": [ + 304, + 670, + 545, + 694 + ], + "type": "text", + "content": "With the hyperparameters " + }, + { + "bbox": [ + 304, + 670, + 545, + 694 + ], + "type": "inline_equation", + "content": "\\alpha, \\beta, \\gamma" + }, + { + "bbox": [ + 304, + 670, + 545, + 694 + ], + "type": "text", + "content": ", we define the final training objective as:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 310, + 700, + 547, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 700, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 310, + 700, + 547, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\left\\| I _ {a d v} - I _ {s r c} \\right\\| ^ {2} + \\alpha \\cdot \\mathcal {L} _ {e n c} + \\beta \\cdot \\mathcal {L} _ {c o n} + \\gamma \\cdot \\mathcal {L} _ {a t t} (4)", + "image_path": "409082802913f2adcba667b6d0ba609fc03c5f711cf26a08092e128540d5d30f.jpg" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12598" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 56, + 87, + 289, + 284 + ], + "blocks": [ + { + "bbox": [ + 48, + 72, + 285, + 84 + ], + "lines": [ + { + "bbox": [ + 48, + 72, + 285, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 285, + 84 + ], + "type": "text", + "content": "Algorithm 1 Adversarial Attack on Image-to-Video Generation" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 56, + 87, + 289, + 284 + ], + "lines": [ + { + "bbox": [ + 56, + 87, + 289, + 284 + ], + "spans": [ + { + "bbox": [ + 56, + 87, + 289, + 284 + ], + "type": "text", + "content": "Input: Source Image " + }, + { + "bbox": [ + 56, + 87, + 289, + 284 + ], + "type": "inline_equation", + "content": "I_{src}" + }, + { + "bbox": [ + 56, + 87, + 289, + 284 + ], + "type": "text", + "content": " and Trainable Copy " + }, + { + "bbox": [ + 56, + 87, + 289, + 284 + ], + "type": "inline_equation", + "content": "I_{adv}" + }, + { + "bbox": [ + 56, + 87, + 289, + 284 + ], + "type": "text", + "content": ", Target \nImage " + }, + { + "bbox": [ + 56, + 87, + 289, + 284 + ], + "type": "inline_equation", + "content": "I_{tg}" + }, + { + "bbox": [ + 56, + 87, + 289, + 284 + ], + "type": "text", + "content": ", Target Video " + }, + { + "bbox": [ + 56, + 87, + 289, + 284 + ], + "type": "inline_equation", + "content": "V_{tg}" + }, + { + "bbox": [ + 56, + 87, + 289, + 284 + ], + "type": "text", + "content": ", Image-to-Video Pipeline " + }, + { + "bbox": [ + 56, + 87, + 289, + 284 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 56, + 87, + 289, + 284 + ], + "type": "text", + "content": ", \nGenerator " + }, + { + "bbox": [ + 56, + 87, + 289, + 284 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 56, + 87, + 289, + 284 + ], + "type": "text", + "content": ", Encoder " + }, + { + "bbox": [ + 56, + 87, + 289, + 284 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 56, + 87, + 289, + 284 + ], + "type": "text", + "content": ", Diffusion Scheduler " + }, + { + "bbox": [ + 56, + 87, + 289, + 284 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 56, + 87, + 289, + 284 + ], + "type": "text", + "content": " \n(Optional:Condition " + }, + { + "bbox": [ + 56, + 87, + 289, + 284 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 56, + 87, + 289, + 284 + ], + "type": "text", + "content": ") \nHyperparameters " + }, + { + "bbox": [ + 56, + 87, + 289, + 284 + ], + "type": "inline_equation", + "content": "\\tau_1, \\tau_2, \\alpha, \\beta, \\gamma, \\lambda" + }, + { + "bbox": [ + 56, + 87, + 289, + 284 + ], + "type": "text", + "content": " \nEncode image " + }, + { + "bbox": [ + 56, + 87, + 289, + 284 + ], + "type": "inline_equation", + "content": "x_{src}, x_{tg} = \\mathcal{E}(I_{src}), \\mathcal{E}(I_{tg})" + }, + { + "bbox": [ + 56, + 87, + 289, + 284 + ], + "type": "text", + "content": " \nGenerate original video " + }, + { + "bbox": [ + 56, + 87, + 289, + 284 + ], + "type": "inline_equation", + "content": "V_0 = \\mathcal{P}(I_{src})" + }, + { + "bbox": [ + 56, + 87, + 289, + 284 + ], + "type": "text", + "content": " \nEncode frames " + }, + { + "bbox": [ + 56, + 87, + 289, + 284 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_0, \\mathbf{X}_{tg} = \\mathcal{E}(V_0), \\mathcal{E}(V_{tg})" + }, + { + "bbox": [ + 56, + 87, + 289, + 284 + ], + "type": "text", + "content": " \nfor each iteration do \n Encode attacked image " + }, + { + "bbox": [ + 56, + 87, + 289, + 284 + ], + "type": "inline_equation", + "content": "x_{adv} = \\mathcal{E}(I_{adv})" + }, + { + "bbox": [ + 56, + 87, + 289, + 284 + ], + "type": "text", + "content": " \n Compute encoder loss: " + }, + { + "bbox": [ + 56, + 87, + 289, + 284 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{enc} = ||x_{adv} - x_{tg}||^2" + }, + { + "bbox": [ + 56, + 87, + 289, + 284 + ], + "type": "text", + "content": " \n Generate noisy frames " + }, + { + "bbox": [ + 56, + 87, + 289, + 284 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_t = \\mathcal{S}(\\mathbf{X}_0)" + }, + { + "bbox": [ + 56, + 87, + 289, + 284 + ], + "type": "text", + "content": " \nPredict " + }, + { + "bbox": [ + 56, + 87, + 289, + 284 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_0\\left\\{\\begin{array}{l}\\hat{\\mathbf{X}}_{0,adv}=\\varnothing(\\mathbf{X}_t,x_{adv},c)\\\\ \\hat{\\mathbf{X}}_{0,src}=\\varnothing(\\mathbf{X}_t,x_{src},c)\\\\ A_{src}\\end{array}\\right." + }, + { + "bbox": [ + 56, + 87, + 289, + 284 + ], + "type": "text", + "content": " \nCompute spatial contrastive loss" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "code_body" + } + ], + "index": 1, + "sub_type": "algorithm" + }, + { + "bbox": [ + 67, + 293, + 293, + 312 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 293, + 293, + 312 + ], + "spans": [ + { + "bbox": [ + 67, + 293, + 293, + 312 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c o n} = \\left| \\left| \\hat {\\mathbf {X}} _ {0, a d v} - \\mathbf {X} _ {t g} \\right| \\right| ^ {2} + \\max \\left(0, \\tau_ {1} - \\left| \\left| \\hat {\\mathbf {X}} _ {0, a d v} - \\mathbf {X} _ {0} \\right| \\right| ^ {2}\\right)", + "image_path": "fccbaac84ccb100d6dc90553a61444dd66b309dc11d19c0dbac4c2386ff62445.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 319, + 200, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 319, + 200, + 331 + ], + "spans": [ + { + "bbox": [ + 67, + 319, + 200, + 331 + ], + "type": "text", + "content": "Compute temporal attention loss" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 331, + 236, + 344 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 331, + 236, + 344 + ], + "spans": [ + { + "bbox": [ + 115, + 331, + 236, + 344 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {a t t} = \\tau_ {2} - | | A _ {a d v} - A _ {s r c} | | ^ {2}", + "image_path": "b67aee3a3f84c2cf672c19fa96ad15755abcd2b0108893878cc4922a8d8f6f3e.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 344, + 160, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 344, + 160, + 355 + ], + "spans": [ + { + "bbox": [ + 67, + 344, + 160, + 355 + ], + "type": "text", + "content": "Compute the final loss" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 355, + 284, + 368 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 355, + 284, + 368 + ], + "spans": [ + { + "bbox": [ + 67, + 355, + 284, + 368 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\left\\| I _ {a d v} - I _ {s r c} \\right\\| ^ {2} + \\alpha \\cdot \\mathcal {L} _ {e n c} + \\beta \\cdot \\mathcal {L} _ {c o n} + \\gamma \\cdot \\mathcal {L} _ {a t t}", + "image_path": "1f2d41755b92b0efcfdc59b92df163bc4229be244b8532fe52eb7f2b4d2d39a3.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 368, + 256, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 368, + 256, + 380 + ], + "spans": [ + { + "bbox": [ + 67, + 368, + 256, + 380 + ], + "type": "text", + "content": "Update parameters " + }, + { + "bbox": [ + 67, + 368, + 256, + 380 + ], + "type": "inline_equation", + "content": "I_{adv} \\gets I_{adv} - \\lambda \\cdot \\nabla_{I_{adv}}\\mathcal{L}" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 58, + 380, + 91, + 390 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 380, + 91, + 390 + ], + "spans": [ + { + "bbox": [ + 58, + 380, + 91, + 390 + ], + "type": "text", + "content": "end for" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 420, + 259, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 420, + 259, + 432 + ], + "spans": [ + { + "bbox": [ + 47, + 420, + 259, + 432 + ], + "type": "text", + "content": "The complete algorithm is presented in Algorithm 1." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 434, + 287, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 434, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 434, + 287, + 495 + ], + "type": "text", + "content": "It is noteworthy that our method supports both global and local attacks, allowing perturbations to be applied across the entire image or to specific objects within the image. This feature can be implemented by applying a mask to the noise during the attack process." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 514, + 128, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 514, + 128, + 528 + ], + "spans": [ + { + "bbox": [ + 47, + 514, + 128, + 528 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 536, + 173, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 536, + 173, + 550 + ], + "spans": [ + { + "bbox": [ + 47, + 536, + 173, + 550 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 557, + 289, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 289, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 289, + 712 + ], + "type": "text", + "content": "Models In the field of video generation, there are few open-source video diffusion models available for experimentation. Stable Video Diffusion [4] (SVD) is a widely used and effective image-to-video diffusion model, which we adopt as the baseline model for our attack and experimental analysis. For other conditional image-to-video models, we employ CogVideoX-5B-I2V [53] for text-image joint conditional video generation. The open-sourced CogVideoX adopts DiT rather than UNet architecture used in SVD, providing enhanced capabilities. We also employ ControlNeXt-SVD-v2 [34] for pose-guided image-to-video generation, which is built upon SVD and provides enhanced control over character animation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": "Data To our knowledge, there are currently no established benchmarks or datasets for image-based adversarial attacks in video generation. In previous work, PhotoGuard [37] reported attack results on over 60 images, while PRIME [28] used an internal dataset containing 35 video clips. In our experiments, we collect a dataset of 300 images and their corresponding generated frames. The selected images and frames depict people and animals with dynamic actions, which could potentially be misused. We exclude images of scenery without main subjects or images that do not yield noticeable temporal movement, as these typically do not produce meaningful motion, rely mainly on camera movements, and may even remain static." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 242, + 547, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 242, + 547, + 446 + ], + "spans": [ + { + "bbox": [ + 304, + 242, + 547, + 446 + ], + "type": "text", + "content": "Metrics We validate our experiments from two perspectives: image perturbation and generation effects. Our goal is to introduce minimal perturbations to the images that significantly degrade the generated results. We use peak signal-to-noise ratio (PSNR), structural similarity (SSIM) metrics, and Fréchet Inception Distance (FID) to evaluate the similarity between the attacked and original images. For evaluating video generation results, we use the metrics Subject Consistency, Motion Smoothness, Aesthetic Quality, and Image Quality from VBench [25]. Subject Consistency measures the average cosine similarity of Dino [6] features between each frame and the first frame, assessing temporal coherence. Motion Smoothness calculates the mean absolute error between interpolated and dropped frames. Aesthetic Quality is assessed using the LAION aesthetic predictor [1], and Image Quality evaluates the presence of low-level distortions in generated video frames with MUSIQ [26]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 460, + 547, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 547, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 547, + 628 + ], + "type": "text", + "content": "Setup We control the added noise strength by adjusting the number of training steps and the perturbation threshold at each step. For the loss function, we scale the hyperparameters " + }, + { + "bbox": [ + 304, + 460, + 547, + 628 + ], + "type": "inline_equation", + "content": "\\alpha, \\beta, \\gamma" + }, + { + "bbox": [ + 304, + 460, + 547, + 628 + ], + "type": "text", + "content": " to ensure that each loss component remains on a similar scale. During training, we use a black image as the target image " + }, + { + "bbox": [ + 304, + 460, + 547, + 628 + ], + "type": "inline_equation", + "content": "I_{tg}" + }, + { + "bbox": [ + 304, + 460, + 547, + 628 + ], + "type": "text", + "content": " and black frames as the target video " + }, + { + "bbox": [ + 304, + 460, + 547, + 628 + ], + "type": "inline_equation", + "content": "V_{tg}" + }, + { + "bbox": [ + 304, + 460, + 547, + 628 + ], + "type": "text", + "content": ". For the attention hooks, we focus on the temporal attention block of the last downsampling blocks in the UNet and the middle blocks in the transformer, separately. In terms of computational cost, the SVD model requires 68GB of GPU memory and approximately 2.5 minutes to apply protection, while CogVideoX demands 75GB of GPU memory and 10 minutes. Notably, this protecting time is comparable to the inference time of the I2V model." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 635, + 423, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 423, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 423, + 647 + ], + "type": "text", + "content": "4.2. Qualitative Analysis" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": "In this section, we provide an analysis of attacks on both unconditional and conditional image-to-video generation. Additionally, we present a robustness analysis in Appendix A to demonstrate that our method remains effective in safeguarding images even in the presence of various pre-processing" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12599" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 69, + 154, + 123 + ], + "blocks": [ + { + "bbox": [ + 59, + 69, + 154, + 123 + ], + "lines": [ + { + "bbox": [ + 59, + 69, + 154, + 123 + ], + "spans": [ + { + "bbox": [ + 59, + 69, + 154, + 123 + ], + "type": "image", + "image_path": "d7977c29cd41f64b7019d4b8e6088b090b04ddc35f4ae6c64abb0e25a4cd67c7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 82, + 124, + 131, + 131 + ], + "lines": [ + { + "bbox": [ + 82, + 124, + 131, + 131 + ], + "spans": [ + { + "bbox": [ + 82, + 124, + 131, + 131 + ], + "type": "text", + "content": "Original Image" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 157, + 70, + 251, + 123 + ], + "blocks": [ + { + "bbox": [ + 157, + 70, + 251, + 123 + ], + "lines": [ + { + "bbox": [ + 157, + 70, + 251, + 123 + ], + "spans": [ + { + "bbox": [ + 157, + 70, + 251, + 123 + ], + "type": "image", + "image_path": "a4fb6bbe0037b7b8b327b6811e76664207389d21f521b46ab421dd98efcc7921.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 251, + 70, + 344, + 123 + ], + "blocks": [ + { + "bbox": [ + 251, + 70, + 344, + 123 + ], + "lines": [ + { + "bbox": [ + 251, + 70, + 344, + 123 + ], + "spans": [ + { + "bbox": [ + 251, + 70, + 344, + 123 + ], + "type": "image", + "image_path": "c331ad1d0f66ce7d87647755fe634be27ff32f03c513646f86c722058429b364.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 344, + 70, + 438, + 123 + ], + "blocks": [ + { + "bbox": [ + 344, + 70, + 438, + 123 + ], + "lines": [ + { + "bbox": [ + 344, + 70, + 438, + 123 + ], + "spans": [ + { + "bbox": [ + 344, + 70, + 438, + 123 + ], + "type": "image", + "image_path": "4a1f2d50fa091027743d94819fb2e71f1b970a892498221ce1f524f6203b91ff.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 438, + 70, + 531, + 123 + ], + "blocks": [ + { + "bbox": [ + 438, + 70, + 531, + 123 + ], + "lines": [ + { + "bbox": [ + 438, + 70, + 531, + 123 + ], + "spans": [ + { + "bbox": [ + 438, + 70, + 531, + 123 + ], + "type": "image", + "image_path": "a4dd6abbbe134ab289634669e13450abde4a4e2674735a435d8aaccc9cadba67.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 59, + 132, + 153, + 185 + ], + "blocks": [ + { + "bbox": [ + 59, + 132, + 153, + 185 + ], + "lines": [ + { + "bbox": [ + 59, + 132, + 153, + 185 + ], + "spans": [ + { + "bbox": [ + 59, + 132, + 153, + 185 + ], + "type": "image", + "image_path": "8ece16884112bdb351a6dcd4f04fdcda01708404499381bcbb0ffef0985b5963.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 71, + 185, + 143, + 193 + ], + "lines": [ + { + "bbox": [ + 71, + 185, + 143, + 193 + ], + "spans": [ + { + "bbox": [ + 71, + 185, + 143, + 193 + ], + "type": "text", + "content": "Random Noisy Image" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 157, + 132, + 251, + 185 + ], + "blocks": [ + { + "bbox": [ + 157, + 132, + 251, + 185 + ], + "lines": [ + { + "bbox": [ + 157, + 132, + 251, + 185 + ], + "spans": [ + { + "bbox": [ + 157, + 132, + 251, + 185 + ], + "type": "image", + "image_path": "d36921865af717dfff9a35729d82395e28a3f330892f001b0a8e5f89bcc07168.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 251, + 132, + 344, + 185 + ], + "blocks": [ + { + "bbox": [ + 283, + 124, + 407, + 132 + ], + "lines": [ + { + "bbox": [ + 283, + 124, + 407, + 132 + ], + "spans": [ + { + "bbox": [ + 283, + 124, + 407, + 132 + ], + "type": "text", + "content": "Generated video from original image." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 251, + 132, + 344, + 185 + ], + "lines": [ + { + "bbox": [ + 251, + 132, + 344, + 185 + ], + "spans": [ + { + "bbox": [ + 251, + 132, + 344, + 185 + ], + "type": "image", + "image_path": "544778c99c3d7ee5e93e420090969dd962bbc5c13bce8f920311aca932993764.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 272, + 185, + 417, + 193 + ], + "lines": [ + { + "bbox": [ + 272, + 185, + 417, + 193 + ], + "spans": [ + { + "bbox": [ + 272, + 185, + 417, + 193 + ], + "type": "text", + "content": "Generated video from random noisy image." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 344, + 132, + 438, + 185 + ], + "blocks": [ + { + "bbox": [ + 344, + 132, + 438, + 185 + ], + "lines": [ + { + "bbox": [ + 344, + 132, + 438, + 185 + ], + "spans": [ + { + "bbox": [ + 344, + 132, + 438, + 185 + ], + "type": "image", + "image_path": "034c98d6a3d5e62c5f75fa9197cd2368f33aca6c071bd7bd3867777f4b2ad68d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 438, + 132, + 531, + 185 + ], + "blocks": [ + { + "bbox": [ + 438, + 132, + 531, + 185 + ], + "lines": [ + { + "bbox": [ + 438, + 132, + 531, + 185 + ], + "spans": [ + { + "bbox": [ + 438, + 132, + 531, + 185 + ], + "type": "image", + "image_path": "75d42566d1ab83cf6009bad08e0d2e14427e7ef53c6ff6cefd0586e21e2e6a6f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 59, + 193, + 154, + 247 + ], + "blocks": [ + { + "bbox": [ + 59, + 193, + 154, + 247 + ], + "lines": [ + { + "bbox": [ + 59, + 193, + 154, + 247 + ], + "spans": [ + { + "bbox": [ + 59, + 193, + 154, + 247 + ], + "type": "image", + "image_path": "7cc2f8cf26a202fdfdcdce5e9dbd1db035d81f575b2484cc9beac805f9677c8c.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 77, + 247, + 140, + 254 + ], + "lines": [ + { + "bbox": [ + 77, + 247, + 140, + 254 + ], + "spans": [ + { + "bbox": [ + 77, + 247, + 140, + 254 + ], + "type": "text", + "content": "PhotoGuard Image" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 157, + 193, + 251, + 246 + ], + "blocks": [ + { + "bbox": [ + 157, + 193, + 251, + 246 + ], + "lines": [ + { + "bbox": [ + 157, + 193, + 251, + 246 + ], + "spans": [ + { + "bbox": [ + 157, + 193, + 251, + 246 + ], + "type": "image", + "image_path": "1664cf2b2afd8897c1b648bc758281348d2de6e91ae223cb3cceda866c6fded7.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 251, + 193, + 344, + 246 + ], + "blocks": [ + { + "bbox": [ + 251, + 193, + 344, + 246 + ], + "lines": [ + { + "bbox": [ + 251, + 193, + 344, + 246 + ], + "spans": [ + { + "bbox": [ + 251, + 193, + 344, + 246 + ], + "type": "image", + "image_path": "ae9f8e1846b2ec7770ca1f62e524ebc1841933e36e94754fb2f236db612c9fd9.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 257, + 246, + 432, + 254 + ], + "lines": [ + { + "bbox": [ + 257, + 246, + 432, + 254 + ], + "spans": [ + { + "bbox": [ + 257, + 246, + 432, + 254 + ], + "type": "text", + "content": "Generated video from PhotoGuard protected image." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 344, + 193, + 438, + 246 + ], + "blocks": [ + { + "bbox": [ + 344, + 193, + 438, + 246 + ], + "lines": [ + { + "bbox": [ + 344, + 193, + 438, + 246 + ], + "spans": [ + { + "bbox": [ + 344, + 193, + 438, + 246 + ], + "type": "image", + "image_path": "9858540f0f28139820e3ff871239227d880309f123f1b69df22f62dc7b0b3061.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 438, + 193, + 531, + 246 + ], + "blocks": [ + { + "bbox": [ + 438, + 193, + 531, + 246 + ], + "lines": [ + { + "bbox": [ + 438, + 193, + 531, + 246 + ], + "spans": [ + { + "bbox": [ + 438, + 193, + 531, + 246 + ], + "type": "image", + "image_path": "1603d02838ae62d4efe1ceccb3c1277c1b93c7a53a93d67c202f7f6abbeff186.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 59, + 255, + 154, + 308 + ], + "blocks": [ + { + "bbox": [ + 59, + 255, + 154, + 308 + ], + "lines": [ + { + "bbox": [ + 59, + 255, + 154, + 308 + ], + "spans": [ + { + "bbox": [ + 59, + 255, + 154, + 308 + ], + "type": "image", + "image_path": "a24e2af1eee20054f2fe7ac1050a06f2d21b1e5d0a95a714d8b14008ed8dcd03.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 76, + 308, + 141, + 316 + ], + "lines": [ + { + "bbox": [ + 76, + 308, + 141, + 316 + ], + "spans": [ + { + "bbox": [ + 76, + 308, + 141, + 316 + ], + "type": "text", + "content": "Our Guarded Image" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 157, + 255, + 251, + 308 + ], + "blocks": [ + { + "bbox": [ + 157, + 255, + 251, + 308 + ], + "lines": [ + { + "bbox": [ + 157, + 255, + 251, + 308 + ], + "spans": [ + { + "bbox": [ + 157, + 255, + 251, + 308 + ], + "type": "image", + "image_path": "18602c149ff527e0e2199c92e4839924b9d51c5ffb02b5d04776225dc8ceebba.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 251, + 255, + 344, + 308 + ], + "blocks": [ + { + "bbox": [ + 251, + 255, + 344, + 308 + ], + "lines": [ + { + "bbox": [ + 251, + 255, + 344, + 308 + ], + "spans": [ + { + "bbox": [ + 251, + 255, + 344, + 308 + ], + "type": "image", + "image_path": "87302b099cdd7cb6f531ecedadfa7b0e662567c93f62bbec2ab49cbf963358d2.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 281, + 308, + 408, + 316 + ], + "lines": [ + { + "bbox": [ + 281, + 308, + 408, + 316 + ], + "spans": [ + { + "bbox": [ + 281, + 308, + 408, + 316 + ], + "type": "text", + "content": "Generated video from guarded image." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 344, + 255, + 438, + 308 + ], + "blocks": [ + { + "bbox": [ + 344, + 255, + 438, + 308 + ], + "lines": [ + { + "bbox": [ + 344, + 255, + 438, + 308 + ], + "spans": [ + { + "bbox": [ + 344, + 255, + 438, + 308 + ], + "type": "image", + "image_path": "f934ebd68b15913493d4a25e030f4858e8a2f5309516fc04e8d785d2dc7ef3fd.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 438, + 255, + 531, + 308 + ], + "blocks": [ + { + "bbox": [ + 438, + 255, + 531, + 308 + ], + "lines": [ + { + "bbox": [ + 438, + 255, + 531, + 308 + ], + "spans": [ + { + "bbox": [ + 438, + 255, + 531, + 308 + ], + "type": "image", + "image_path": "2fdf630e4b34a76095a897010080c7a836f3ef0ec746e3e7dc5152b3bdfdd93f.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 60, + 329, + 154, + 392 + ], + "blocks": [ + { + "bbox": [ + 115, + 318, + 477, + 327 + ], + "lines": [ + { + "bbox": [ + 115, + 318, + 477, + 327 + ], + "spans": [ + { + "bbox": [ + 115, + 318, + 477, + 327 + ], + "type": "text", + "content": "(a) SVD generation results on original image, random noisy image, PhotoGuard image and our I2VGuard image." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 60, + 329, + 154, + 392 + ], + "lines": [ + { + "bbox": [ + 60, + 329, + 154, + 392 + ], + "spans": [ + { + "bbox": [ + 60, + 329, + 154, + 392 + ], + "type": "image", + "image_path": "cf940a918cddfdb9ef5c5bd97d1814e8f2b440110161aba692b2399b84aa3b86.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 83, + 392, + 133, + 399 + ], + "lines": [ + { + "bbox": [ + 83, + 392, + 133, + 399 + ], + "spans": [ + { + "bbox": [ + 83, + 392, + 133, + 399 + ], + "type": "text", + "content": "Original Image" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 157, + 329, + 251, + 392 + ], + "blocks": [ + { + "bbox": [ + 157, + 329, + 251, + 392 + ], + "lines": [ + { + "bbox": [ + 157, + 329, + 251, + 392 + ], + "spans": [ + { + "bbox": [ + 157, + 329, + 251, + 392 + ], + "type": "image", + "image_path": "dd1f4407ed07243aa6586466079aef2f196f60c2b413199eb026c04978326cf6.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 251, + 329, + 344, + 392 + ], + "blocks": [ + { + "bbox": [ + 251, + 329, + 344, + 392 + ], + "lines": [ + { + "bbox": [ + 251, + 329, + 344, + 392 + ], + "spans": [ + { + "bbox": [ + 251, + 329, + 344, + 392 + ], + "type": "image", + "image_path": "f748f1b0a2b25fb25e5006566e2478901b42c9d9e09a7c55d86314d53aa8d3f7.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 284, + 392, + 408, + 400 + ], + "lines": [ + { + "bbox": [ + 284, + 392, + 408, + 400 + ], + "spans": [ + { + "bbox": [ + 284, + 392, + 408, + 400 + ], + "type": "text", + "content": "Generated video from original image." + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 344, + 329, + 438, + 392 + ], + "blocks": [ + { + "bbox": [ + 344, + 329, + 438, + 392 + ], + "lines": [ + { + "bbox": [ + 344, + 329, + 438, + 392 + ], + "spans": [ + { + "bbox": [ + 344, + 329, + 438, + 392 + ], + "type": "image", + "image_path": "8010378a9eeccfccf47584822d523e1099125d409e1a5cb3d3cad9baafb70257.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 438, + 329, + 533, + 392 + ], + "blocks": [ + { + "bbox": [ + 438, + 329, + 533, + 392 + ], + "lines": [ + { + "bbox": [ + 438, + 329, + 533, + 392 + ], + "spans": [ + { + "bbox": [ + 438, + 329, + 533, + 392 + ], + "type": "image", + "image_path": "56560c5f12c5c4d8e05217fb08870e3f00252483ed16c8bfea762e01b4b07770.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 60, + 400, + 154, + 463 + ], + "blocks": [ + { + "bbox": [ + 60, + 400, + 154, + 463 + ], + "lines": [ + { + "bbox": [ + 60, + 400, + 154, + 463 + ], + "spans": [ + { + "bbox": [ + 60, + 400, + 154, + 463 + ], + "type": "image", + "image_path": "86d28d43bf444ee4d9e5e3651dee275a37ae6f94735076c1cc5c99f3ba82ad34.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 83, + 464, + 136, + 472 + ], + "lines": [ + { + "bbox": [ + 83, + 464, + 136, + 472 + ], + "spans": [ + { + "bbox": [ + 83, + 464, + 136, + 472 + ], + "type": "text", + "content": "Guarded Image" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 476, + 459, + 487 + ], + "lines": [ + { + "bbox": [ + 132, + 476, + 459, + 487 + ], + "spans": [ + { + "bbox": [ + 132, + 476, + 459, + 487 + ], + "type": "text", + "content": "(b) CogVideoX generation results, where we do not set the prompt in the process of video generation." + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 496, + 545, + 519 + ], + "lines": [ + { + "bbox": [ + 46, + 496, + 545, + 519 + ], + "spans": [ + { + "bbox": [ + 46, + 496, + 545, + 519 + ], + "type": "text", + "content": "Figure 4. Qualitative results of adversarial attacks on I2V models SVD and CogVideoX. We also include generation results of SVD with random noise and PhotoGuard [37] perturbations for comparison. All generation results are using the same seed." + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 157, + 400, + 251, + 463 + ], + "blocks": [ + { + "bbox": [ + 157, + 400, + 251, + 463 + ], + "lines": [ + { + "bbox": [ + 157, + 400, + 251, + 463 + ], + "spans": [ + { + "bbox": [ + 157, + 400, + 251, + 463 + ], + "type": "image", + "image_path": "7932a5e193773c4df1123d494e7b48d51e1766a469e96ab99593e4d7c32ff534.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 251, + 400, + 344, + 463 + ], + "blocks": [ + { + "bbox": [ + 251, + 400, + 344, + 463 + ], + "lines": [ + { + "bbox": [ + 251, + 400, + 344, + 463 + ], + "spans": [ + { + "bbox": [ + 251, + 400, + 344, + 463 + ], + "type": "image", + "image_path": "5b5d05439d1fdba01850a3e5637cc39619d82b669b2db3cfe1eca44ae79fa5af.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 283, + 464, + 409, + 473 + ], + "lines": [ + { + "bbox": [ + 283, + 464, + 409, + 473 + ], + "spans": [ + { + "bbox": [ + 283, + 464, + 409, + 473 + ], + "type": "text", + "content": "Generated video from guarded image." + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 344, + 400, + 438, + 463 + ], + "blocks": [ + { + "bbox": [ + 344, + 400, + 438, + 463 + ], + "lines": [ + { + "bbox": [ + 344, + 400, + 438, + 463 + ], + "spans": [ + { + "bbox": [ + 344, + 400, + 438, + 463 + ], + "type": "image", + "image_path": "eb4cb7ed4b319d75d089214454e4ef47b5d9fa66ece7c9b3f788b9823281ccd5.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 438, + 400, + 533, + 463 + ], + "blocks": [ + { + "bbox": [ + 438, + 400, + 533, + 463 + ], + "lines": [ + { + "bbox": [ + 438, + 400, + 533, + 463 + ], + "spans": [ + { + "bbox": [ + 438, + 400, + 533, + 463 + ], + "type": "image", + "image_path": "9f1fc6d2aab65ed05c37ee2c37f2a1b11bc282b182439cc9d5f0db7d9a430a79.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + } + ], + "index": 41 + }, + { + "bbox": [ + 46, + 539, + 288, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 539, + 288, + 588 + ], + "spans": [ + { + "bbox": [ + 46, + 539, + 288, + 588 + ], + "type": "text", + "content": "techniques and denoising modules [7]. A detailed visualization of the generation effects is provided in Appendix B. Furthermore, an additional comparison between spatial and temporal attacks is presented in Appendix E." + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "text", + "content": "Attack on Image-to-Video Generation We analyze the generation results using two image-to-video generation models: Stable Video Diffusion (SVD) and CogVideoX, as shown in Fig. 4. We observe that the noise added to the guarded image is nearly invisible, yet it causes significant variations in the generated results. In the example shown in Fig. 4a, we additionally illustrate the generated results for both a random noisy image and an image protected by PhotoGuard [37]. The random noise does not affect generation quality, preserv" + } + ] + } + ], + "index": 46 + }, + { + "bbox": [ + 304, + 539, + 548, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 539, + 548, + 708 + ], + "spans": [ + { + "bbox": [ + 304, + 539, + 548, + 708 + ], + "type": "text", + "content": "ing both temporal consistency and normal spatial appearance, likely because the VAE filters out such random noise. The PhotoGuard-protected image, however, introduces noise specifically on the person, which slightly affects appearance without compromising temporal consistency. In contrast, our protected image disrupts both spatial content and temporal consistency. Specifically, the woman's head and hair movement appear unnatural, chaotic frames are generated, and unusual textures emerge in the background across all frames. Fig. 4b shows results from CogVideoX under an image-only setting, where no prompt is applied during training and inference. CogVideoX, which is relatively more powerful than SVD, requires relatively stronger noise to affect generation. It can be observed that motion distortions still occur, particu" + } + ] + } + ], + "index": 47 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12600" + } + ] + } + ], + "index": 48 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 69, + 119, + 169 + ], + "blocks": [ + { + "bbox": [ + 60, + 69, + 119, + 169 + ], + "lines": [ + { + "bbox": [ + 60, + 69, + 119, + 169 + ], + "spans": [ + { + "bbox": [ + 60, + 69, + 119, + 169 + ], + "type": "image", + "image_path": "fea02e2f23a4f2b1a73ebe9f0b262fb516bb5a0a5108cd9fbf57adb41d6c27bd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 123, + 70, + 181, + 169 + ], + "blocks": [ + { + "bbox": [ + 123, + 70, + 181, + 169 + ], + "lines": [ + { + "bbox": [ + 123, + 70, + 181, + 169 + ], + "spans": [ + { + "bbox": [ + 123, + 70, + 181, + 169 + ], + "type": "image", + "image_path": "d2bc0bd02264c0dc82f7a1b2bdad43fc64e3b0f185263a45bed877303d46e498.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 181, + 70, + 238, + 169 + ], + "blocks": [ + { + "bbox": [ + 181, + 70, + 238, + 169 + ], + "lines": [ + { + "bbox": [ + 181, + 70, + 238, + 169 + ], + "spans": [ + { + "bbox": [ + 181, + 70, + 238, + 169 + ], + "type": "image", + "image_path": "732b0f92e57e636fc706470137ec74f3a8c27fb4669f6bca3500de7ede1d18f3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 240, + 70, + 296, + 169 + ], + "blocks": [ + { + "bbox": [ + 240, + 70, + 296, + 169 + ], + "lines": [ + { + "bbox": [ + 240, + 70, + 296, + 169 + ], + "spans": [ + { + "bbox": [ + 240, + 70, + 296, + 169 + ], + "type": "image", + "image_path": "577e63b87b31d2475ca66a4a58fba255ec5c6fe9f43237f8a43b9e63b4056a91.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 298, + 70, + 354, + 169 + ], + "blocks": [ + { + "bbox": [ + 298, + 70, + 354, + 169 + ], + "lines": [ + { + "bbox": [ + 298, + 70, + 354, + 169 + ], + "spans": [ + { + "bbox": [ + 298, + 70, + 354, + 169 + ], + "type": "image", + "image_path": "b6a98c6ebf9340adf23267014e7b630189296f4c83d3ed69ec3e4228beecec74.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 356, + 70, + 412, + 169 + ], + "blocks": [ + { + "bbox": [ + 356, + 70, + 412, + 169 + ], + "lines": [ + { + "bbox": [ + 356, + 70, + 412, + 169 + ], + "spans": [ + { + "bbox": [ + 356, + 70, + 412, + 169 + ], + "type": "image", + "image_path": "456e4ccc5b6b1ff8a497e3e1770975c864bbcce35ab5da760030bedd0e296fdb.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 413, + 70, + 470, + 169 + ], + "blocks": [ + { + "bbox": [ + 413, + 70, + 470, + 169 + ], + "lines": [ + { + "bbox": [ + 413, + 70, + 470, + 169 + ], + "spans": [ + { + "bbox": [ + 413, + 70, + 470, + 169 + ], + "type": "image", + "image_path": "9df7b0069d470fc6724ee4dbabf31307c3b7e5975060a1e22c824df1dba5fdbd.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 471, + 70, + 528, + 169 + ], + "blocks": [ + { + "bbox": [ + 471, + 70, + 528, + 169 + ], + "lines": [ + { + "bbox": [ + 471, + 70, + 528, + 169 + ], + "spans": [ + { + "bbox": [ + 471, + 70, + 528, + 169 + ], + "type": "image", + "image_path": "57bfa7dfaf8144b79ed544440db9dc16afdd4a99e0b874657c086eaea9492029.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 60, + 178, + 118, + 277 + ], + "blocks": [ + { + "bbox": [ + 63, + 170, + 116, + 178 + ], + "lines": [ + { + "bbox": [ + 63, + 170, + 116, + 178 + ], + "spans": [ + { + "bbox": [ + 63, + 170, + 116, + 178 + ], + "type": "text", + "content": "Original Image" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 60, + 178, + 118, + 277 + ], + "lines": [ + { + "bbox": [ + 60, + 178, + 118, + 277 + ], + "spans": [ + { + "bbox": [ + 60, + 178, + 118, + 277 + ], + "type": "image", + "image_path": "7565a6bd8f58a633fbca595c789a6fc205452c48296bbf1609224ec9e29364b1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 63, + 279, + 118, + 288 + ], + "lines": [ + { + "bbox": [ + 63, + 279, + 118, + 288 + ], + "spans": [ + { + "bbox": [ + 63, + 279, + 118, + 288 + ], + "type": "text", + "content": "Guarded Image" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 123, + 178, + 181, + 277 + ], + "blocks": [ + { + "bbox": [ + 123, + 178, + 181, + 277 + ], + "lines": [ + { + "bbox": [ + 123, + 178, + 181, + 277 + ], + "spans": [ + { + "bbox": [ + 123, + 178, + 181, + 277 + ], + "type": "image", + "image_path": "7607e9e47dae44923b471dc0261a80c6b1cfc50094c6cffa795291c39e783b47.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 181, + 178, + 238, + 277 + ], + "blocks": [ + { + "bbox": [ + 181, + 178, + 238, + 277 + ], + "lines": [ + { + "bbox": [ + 181, + 178, + 238, + 277 + ], + "spans": [ + { + "bbox": [ + 181, + 178, + 238, + 277 + ], + "type": "image", + "image_path": "e1c541087c0921f070ac44991ec1ff560968cbfeb9203b93424887c4554e5ce3.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 240, + 178, + 296, + 277 + ], + "blocks": [ + { + "bbox": [ + 260, + 170, + 392, + 178 + ], + "lines": [ + { + "bbox": [ + 260, + 170, + 392, + 178 + ], + "spans": [ + { + "bbox": [ + 260, + 170, + 392, + 178 + ], + "type": "text", + "content": "Generated video from original image." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 240, + 178, + 296, + 277 + ], + "lines": [ + { + "bbox": [ + 240, + 178, + 296, + 277 + ], + "spans": [ + { + "bbox": [ + 240, + 178, + 296, + 277 + ], + "type": "image", + "image_path": "be1772b4501dc7527efc66417c60b9af54a1f296e9d10abb0e0e33d61512dc3a.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 260, + 278, + 394, + 288 + ], + "lines": [ + { + "bbox": [ + 260, + 278, + 394, + 288 + ], + "spans": [ + { + "bbox": [ + 260, + 278, + 394, + 288 + ], + "type": "text", + "content": "Generated video from guarded image." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 298, + 178, + 354, + 277 + ], + "blocks": [ + { + "bbox": [ + 298, + 178, + 354, + 277 + ], + "lines": [ + { + "bbox": [ + 298, + 178, + 354, + 277 + ], + "spans": [ + { + "bbox": [ + 298, + 178, + 354, + 277 + ], + "type": "image", + "image_path": "ecac280cc5a08bf8914a7a47becaec022edf05a9f50aaf1a1a4c136349fc74f5.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 356, + 178, + 412, + 277 + ], + "blocks": [ + { + "bbox": [ + 356, + 178, + 412, + 277 + ], + "lines": [ + { + "bbox": [ + 356, + 178, + 412, + 277 + ], + "spans": [ + { + "bbox": [ + 356, + 178, + 412, + 277 + ], + "type": "image", + "image_path": "9033dddd4430a3cd70af6e05ea3f58a21a67c7acd2fbc098fe74192bac60df85.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 413, + 178, + 470, + 277 + ], + "blocks": [ + { + "bbox": [ + 413, + 178, + 470, + 277 + ], + "lines": [ + { + "bbox": [ + 413, + 178, + 470, + 277 + ], + "spans": [ + { + "bbox": [ + 413, + 178, + 470, + 277 + ], + "type": "image", + "image_path": "888d2a4950889f5ae6202b6c54ab6235961a0767023d2301e5cc538847f17451.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 471, + 178, + 528, + 277 + ], + "blocks": [ + { + "bbox": [ + 471, + 178, + 528, + 277 + ], + "lines": [ + { + "bbox": [ + 471, + 178, + 528, + 277 + ], + "spans": [ + { + "bbox": [ + 471, + 178, + 528, + 277 + ], + "type": "image", + "image_path": "62f0ebbb7f90847cb2a7ff35f058b59634eb1a8fdaf4a065f309f0fbbb3cafd0.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 60, + 302, + 154, + 366 + ], + "blocks": [ + { + "bbox": [ + 60, + 302, + 154, + 366 + ], + "lines": [ + { + "bbox": [ + 60, + 302, + 154, + 366 + ], + "spans": [ + { + "bbox": [ + 60, + 302, + 154, + 366 + ], + "type": "image", + "image_path": "be2f65d648a05abff5c50122b2d91e9a8290c5072e6c5fa27f5d03301c2bd6d8.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 83, + 366, + 133, + 374 + ], + "lines": [ + { + "bbox": [ + 83, + 366, + 133, + 374 + ], + "spans": [ + { + "bbox": [ + 83, + 366, + 133, + 374 + ], + "type": "text", + "content": "Original Image" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 157, + 302, + 252, + 366 + ], + "blocks": [ + { + "bbox": [ + 224, + 293, + 369, + 302 + ], + "lines": [ + { + "bbox": [ + 224, + 293, + 369, + 302 + ], + "spans": [ + { + "bbox": [ + 224, + 293, + 369, + 302 + ], + "type": "text", + "content": "(a) ControlNeXT-SVD-v2 generation results." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 157, + 302, + 252, + 366 + ], + "lines": [ + { + "bbox": [ + 157, + 302, + 252, + 366 + ], + "spans": [ + { + "bbox": [ + 157, + 302, + 252, + 366 + ], + "type": "image", + "image_path": "61e455ee5c51a98aeb1716703f2275f4bf2c84c1d31dba6adbdc29c1cc92acaa.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 253, + 302, + 345, + 366 + ], + "blocks": [ + { + "bbox": [ + 253, + 302, + 345, + 366 + ], + "lines": [ + { + "bbox": [ + 253, + 302, + 345, + 366 + ], + "spans": [ + { + "bbox": [ + 253, + 302, + 345, + 366 + ], + "type": "image", + "image_path": "8bf15d03c40781b56d95285ef5d8ce4332f801cec2a8851e8cd393967d21f06b.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 346, + 302, + 439, + 366 + ], + "blocks": [ + { + "bbox": [ + 346, + 302, + 439, + 366 + ], + "lines": [ + { + "bbox": [ + 346, + 302, + 439, + 366 + ], + "spans": [ + { + "bbox": [ + 346, + 302, + 439, + 366 + ], + "type": "image", + "image_path": "420fa9a3828b40ed26bc4906c75a0010c5c450f026c9c1e64bb36d8cb9d69905.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 440, + 302, + 533, + 366 + ], + "blocks": [ + { + "bbox": [ + 440, + 302, + 533, + 366 + ], + "lines": [ + { + "bbox": [ + 440, + 302, + 533, + 366 + ], + "spans": [ + { + "bbox": [ + 440, + 302, + 533, + 366 + ], + "type": "image", + "image_path": "8e833d2beb9ba8daa3204eda0cf8ce4bc67df15a308989df51285d46427a4a2e.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 60, + 376, + 154, + 438 + ], + "blocks": [ + { + "bbox": [ + 60, + 376, + 154, + 438 + ], + "lines": [ + { + "bbox": [ + 60, + 376, + 154, + 438 + ], + "spans": [ + { + "bbox": [ + 60, + 376, + 154, + 438 + ], + "type": "image", + "image_path": "ceda9160e87cef3f31702cd15516442d97080d992bfb1249a8534d1fa106f864.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 83, + 439, + 136, + 447 + ], + "lines": [ + { + "bbox": [ + 83, + 439, + 136, + 447 + ], + "spans": [ + { + "bbox": [ + 83, + 439, + 136, + 447 + ], + "type": "text", + "content": "Guarded Image" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 471, + 546, + 495 + ], + "lines": [ + { + "bbox": [ + 46, + 471, + 546, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 471, + 546, + 495 + ], + "type": "text", + "content": "Figure 5. Qualitative results of adversarial attacks on conditional I2V models ControlNeXt and CogVideoX. It can be observed that visual features are not controlled effectively, leading to unreasonable generation results. All generation results are using the same seed." + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 157, + 376, + 252, + 438 + ], + "blocks": [ + { + "bbox": [ + 157, + 376, + 252, + 438 + ], + "lines": [ + { + "bbox": [ + 157, + 376, + 252, + 438 + ], + "spans": [ + { + "bbox": [ + 157, + 376, + 252, + 438 + ], + "type": "image", + "image_path": "2f61d492e5569eabeed4034d7ece6397a1d0dece5413fd50ec4f828007f664a3.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 167, + 451, + 425, + 461 + ], + "lines": [ + { + "bbox": [ + 167, + 451, + 425, + 461 + ], + "spans": [ + { + "bbox": [ + 167, + 451, + 425, + 461 + ], + "type": "text", + "content": "(b) CogVideoX generation results. The prompt is \"Many ostriches are running\"." + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 253, + 376, + 345, + 438 + ], + "blocks": [ + { + "bbox": [ + 284, + 367, + 408, + 376 + ], + "lines": [ + { + "bbox": [ + 284, + 367, + 408, + 376 + ], + "spans": [ + { + "bbox": [ + 284, + 367, + 408, + 376 + ], + "type": "text", + "content": "Generated video from original image." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 253, + 376, + 345, + 438 + ], + "lines": [ + { + "bbox": [ + 253, + 376, + 345, + 438 + ], + "spans": [ + { + "bbox": [ + 253, + 376, + 345, + 438 + ], + "type": "image", + "image_path": "33299a55a3ebb1e9ecefcb3b3c7dcad3e67c69a4e8235d7bc810c4cf0dc5067a.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 283, + 439, + 409, + 448 + ], + "lines": [ + { + "bbox": [ + 283, + 439, + 409, + 448 + ], + "spans": [ + { + "bbox": [ + 283, + 439, + 409, + 448 + ], + "type": "text", + "content": "Generated video from guarded image." + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 346, + 376, + 439, + 438 + ], + "blocks": [ + { + "bbox": [ + 346, + 376, + 439, + 438 + ], + "lines": [ + { + "bbox": [ + 346, + 376, + 439, + 438 + ], + "spans": [ + { + "bbox": [ + 346, + 376, + 439, + 438 + ], + "type": "image", + "image_path": "16af9bcc9dcaf58f21c379e5380342732a5ea835ef3edbfa9d506bad66a38d45.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 440, + 376, + 533, + 438 + ], + "blocks": [ + { + "bbox": [ + 440, + 376, + 533, + 438 + ], + "lines": [ + { + "bbox": [ + 440, + 376, + 533, + 438 + ], + "spans": [ + { + "bbox": [ + 440, + 376, + 533, + 438 + ], + "type": "image", + "image_path": "fcaa90bd8af0b969a616a5409f9810deca6602db051438d5fa2c06ddf4e1a814.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "bbox": [ + 46, + 514, + 288, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 514, + 288, + 574 + ], + "spans": [ + { + "bbox": [ + 46, + 514, + 288, + 574 + ], + "type": "text", + "content": "larly in areas where motion takes place, such as the blurred head movement of the man, with the noise causing the surrounding regions to fail in generation, resulting in a blur. In summary, our method effectively disrupts both spatial content and temporal consistency in generated videos." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 46, + 594, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 713 + ], + "type": "text", + "content": "Attack on Conditional Image-to-Video Generation We present an attack on conditional image-to-video models in Fig. 5, specifically examining pose-conditioned ControlNeXt-SVD and text-guided video generation with CogVideoX. In the first row of Fig. 5a, the ControlNeXt model effectively controls the character's motion while maintaining consistent character appearance and background features. However, in the second row, where we guard the original image, the model fails to preserve the character's appearance and fidelity in the background. Noticeable differences" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 304, + 514, + 547, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 514, + 547, + 706 + ], + "spans": [ + { + "bbox": [ + 304, + 514, + 547, + 706 + ], + "type": "text", + "content": "emerge: for instance, the fruits in the background exhibit blurring, and the character appears deformed. These spatial distortions illustrate the effectiveness of our attack method. Despite this, the pose control remains robust, likely due to the intensive training of the original SVD UNet, which results in strong overfitting to the specific style and hands shape. In the example of CogVideoX shown in Fig. 5b, we observe that adding special texture noise leads to the generation of temporal inconsistency and blurred content compared with the original ones. Specifically, in this case, texture noise is applied around the animals, causing distortions in the generated output, such as headless and transparent ostriches. In our experiments, CogVideoX exhibits strong prompt-following ability, which is due to an additional powerful text encoder compared to SVD. By slightly increasing the noise strength, we are able to successfully execute the attack." + } + ] + } + ], + "index": 39 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12601" + } + ] + } + ], + "index": 40 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 545, + 126 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 545, + 126 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 545, + 126 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 545, + 126 + ], + "type": "table", + "html": "
Video SourceModelSubject Consistency(%, ↓)Motion Smoothness(%, ↓)Aesthetic Quality (%, ↓)Image Quality(↓)
Original ImageSVD95.86±2.6297.90±1.4356.76±4.7567.28±6.18
Guarded Image91.57±3.9597.18±1.2153.42±4.9364.38±8.23
Original ImageCogVideoX97.02±1.9699.19±0.2759.94±5.5367.60±6.49
Guarded Image93.50±3.5897.97±0.3253.95±5.6265.24±9.85
", + "image_path": "d8ba98d75481d9c04a135cd007bdeea3aa5a45760a66d8abb8cc9008a86ead48.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 186, + 285, + 225 + ], + "blocks": [ + { + "bbox": [ + 46, + 134, + 546, + 168 + ], + "lines": [ + { + "bbox": [ + 46, + 134, + 546, + 168 + ], + "spans": [ + { + "bbox": [ + 46, + 134, + 546, + 168 + ], + "type": "text", + "content": "Table 1. Analysis of video generation effects of SVD and CogVideoX from original images and images guarded by our method. Mean and variance of evaluations are reported. We exclude results with extremely high subject consistency and motion smoothness, as these indicate static frames, which are outside the scope of this evaluation. An ablation study on the different attack methods is also presented in ??" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 186, + 285, + 225 + ], + "lines": [ + { + "bbox": [ + 50, + 186, + 285, + 225 + ], + "spans": [ + { + "bbox": [ + 50, + 186, + 285, + 225 + ], + "type": "table", + "html": "
Image TypePSNR(↑)SSIM(↑)FID(↓)
Random Noise26.22±0.040.774±0.00314.31
Guarded Image31.83±0.190.868±0.01616.78
", + "image_path": "6adacf430f021ac16e798edc8b8e2370239a2f1144260cf0180f7d0475f0552f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 233, + 288, + 278 + ], + "lines": [ + { + "bbox": [ + 46, + 233, + 288, + 278 + ], + "spans": [ + { + "bbox": [ + 46, + 233, + 288, + 278 + ], + "type": "text", + "content": "Table 2. Analysis of perturbations between original and modified images. Random noise is introduced to the original images to establish a baseline, and the mean and variance of the perturbations are reported." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 49, + 289, + 286, + 378 + ], + "blocks": [ + { + "bbox": [ + 49, + 289, + 286, + 378 + ], + "lines": [ + { + "bbox": [ + 49, + 289, + 286, + 378 + ], + "spans": [ + { + "bbox": [ + 49, + 289, + 286, + 378 + ], + "type": "image", + "image_path": "5909bcccf746fc3a7f012b0e7208f35b2256749cd4166ae92a75209139490e39.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 48, + 388, + 285, + 399 + ], + "lines": [ + { + "bbox": [ + 48, + 388, + 285, + 399 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 285, + 399 + ], + "type": "text", + "content": "Figure 6. Frame-level quantitative evaluation with the same seed." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 421, + 171, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 421, + 171, + 434 + ], + "spans": [ + { + "bbox": [ + 47, + 421, + 171, + 434 + ], + "type": "text", + "content": "4.3. Quantitative Analysis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 445, + 288, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 445, + 288, + 648 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 288, + 648 + ], + "type": "text", + "content": "Image Perturbation Analysis We present the modified image quality results in Tab. 2. Random noise was added to the original images to establish a baseline for comparison. The mean and variance of the added noise in this baseline are generally comparable to the modifications introduced in our attacked images. The results demonstrate that, despite similar perturbation strength, our method achieves significantly higher image quality, as indicated by higher score in both peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM). A qualitative comparison reveals that the noise added to the image possesses its own texture, which is less visible than the random noise added. In terms of image quality and diversity, our approach yields a higher Fréchet Inception Distance (FID). This is expected because the specific texture noise we add aligns with the original image texture, resulting in lower diversity compared to images with added random noise." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "type": "text", + "content": "Video Generation Analysis Evaluation results for generated videos of SVD and CogVideoX are shown in Tab. 1. First, we assess the performance of the I2V model on original images, finding that it demonstrates strong subject consis" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 189, + 547, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 189, + 547, + 511 + ], + "spans": [ + { + "bbox": [ + 304, + 189, + 547, + 511 + ], + "type": "text", + "content": "tency, smooth motion transitions, high aesthetic quality, and well-preserved low-level image quality. This suggests that the I2V models are capable of generating visually coherent and temporally stable content. In particular, CogVideoX performs a better performance than SVD because transformers scale more easily in both data and model size than UNet, making CogVideoX superior to SVD in size and performance. In contrast, our method, which uses a similar noise strength, significantly impacts all evaluated metrics. To be precise, temporal assessments reveal declines around " + }, + { + "bbox": [ + 304, + 189, + 547, + 511 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 304, + 189, + 547, + 511 + ], + "type": "text", + "content": " in subject consistency and around " + }, + { + "bbox": [ + 304, + 189, + 547, + 511 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 304, + 189, + 547, + 511 + ], + "type": "text", + "content": " in motion smoothness, showing that our method effectively disrupts the model's temporal coherence. Specifically, the decline in subject consistency is reflected in irregular movement of the main object's location or the sporadic appearance and disappearance of the object. The decline in motion smoothness is reflected in abrupt, discontinuous frame-to-frame movements. Spatial evaluations also show that our method considerably diminishes aesthetic quality and low-level image quality by " + }, + { + "bbox": [ + 304, + 189, + 547, + 511 + ], + "type": "inline_equation", + "content": "3 \\sim 6\\%" + }, + { + "bbox": [ + 304, + 189, + 547, + 511 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 189, + 547, + 511 + ], + "type": "inline_equation", + "content": "2 \\sim 3\\%" + }, + { + "bbox": [ + 304, + 189, + 547, + 511 + ], + "type": "text", + "content": ", respectively. The decrease in aesthetic quality is evident in the appearance of unusual textures or object deformations that detract from the visual appeal. These reductions reflect a successful attack on the image-to-video model. Notably, because CogVideoX employs a 3D VAE, which handles spatiotemporal noise in videos better than the 2D+1D VAE in SVD, CogVideoX shows greater resistance to attacks compared to SVD." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 512, + 547, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 512, + 547, + 561 + ], + "spans": [ + { + "bbox": [ + 304, + 512, + 547, + 561 + ], + "type": "text", + "content": "Additionally, we provide a frame-level quantitative evaluation in Fig. 6. The results indicate that our method effectively disrupts temporal consistency, leading to a propagated deviation from the original generation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 573, + 378, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 573, + 378, + 586 + ], + "spans": [ + { + "bbox": [ + 306, + 573, + 378, + 586 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "content": "In this paper, we presented I2VGuard, a novel adversarial method that applies imperceptible perturbations on images to degrade the quality of videos generated by diffusion-based image-to-video models, thereby protecting images from misuse. Concretely, we designed three attack modules, i.e. spatial attack, temporal attack and diffusion attack, to disrupt visual fidelity, temporal consistency, and cross-model robustness, respectively. Extensive experiments on state-of-the-art generative models, i.e. CogVideoX and SVD, demonstrated the effectiveness of our method." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12602" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 185, + 101 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 185, + 101 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 185, + 101 + ], + "type": "text", + "content": "[1] LAION AI. Laion-aesthetics. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 103, + 288, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 103, + 288, + 146 + ], + "spans": [ + { + "bbox": [ + 53, + 103, + 288, + 146 + ], + "type": "text", + "content": "[2] Ali Asghar, Amna Shifa, and Mamoona Naveed Asghar. Survey on video security: Examining threats, challenges, and future trends. Computers, Materials & Continua, 80(3), 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 148, + 288, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 203 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 203 + ], + "type": "text", + "content": "[3] Omer Bar-Tal, Hila Chefer, Omer Tov, Charles Herrmann, Roni Paiss, Shiran Zada, Ariel Ephrat, Junhwa Hur, Yuanzhen Li, Tomer Michaeli, et al. Lumiere: A space-time diffusion model for video generation. arXiv preprint arXiv:2401.12945, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 205, + 288, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 288, + 259 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 288, + 259 + ], + "type": "text", + "content": "[4] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2, 4, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 261, + 288, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 261, + 288, + 305 + ], + "spans": [ + { + "bbox": [ + 53, + 261, + 288, + 305 + ], + "type": "text", + "content": "[5] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 307, + 288, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 307, + 288, + 361 + ], + "spans": [ + { + "bbox": [ + 53, + 307, + 288, + 361 + ], + "type": "text", + "content": "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 363, + 288, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 363, + 288, + 396 + ], + "spans": [ + { + "bbox": [ + 53, + 363, + 288, + 396 + ], + "type": "text", + "content": "[7] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In European conference on computer vision, pages 17-33. Springer, 2022. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 397, + 288, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 397, + 288, + 441 + ], + "spans": [ + { + "bbox": [ + 53, + 397, + 288, + 441 + ], + "type": "text", + "content": "[8] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 442, + 288, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 442, + 288, + 487 + ], + "spans": [ + { + "bbox": [ + 53, + 442, + 288, + 487 + ], + "type": "text", + "content": "[9] Weixin Chen, Dawn Song, and Bo Li. Trojdiff: Trojan attacks on diffusion models with diverse targets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4035-4044, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 488, + 288, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 488, + 288, + 531 + ], + "spans": [ + { + "bbox": [ + 48, + 488, + 288, + 531 + ], + "type": "text", + "content": "[10] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In IEEE Conference on Computer Vision and Pattern Recognition, 2021. Computer Vision Foundation / IEEE, 2021. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 533, + 288, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 533, + 288, + 577 + ], + "spans": [ + { + "bbox": [ + 48, + 533, + 288, + 577 + ], + "type": "text", + "content": "[11] June Suk Choi, Kyungmin Lee, Jongheon Jeong, Saining Xie, Jinwoo Shin, and Kimin Lee. Diffusionguard: A robust defense against malicious diffusion-based image editing. In ICML 2024 Next Generation of AI Safety Workshop. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 578, + 288, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 288, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 288, + 622 + ], + "type": "text", + "content": "[12] Sheng-Yen Chou, Pin-Yu Chen, and Tsung-Yi Ho. How to backdoor diffusion models? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4015-4024, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 624, + 288, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 288, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 288, + 657 + ], + "type": "text", + "content": "[13] Hai Ci, Yiren Song, Pei Yang, Jinheng Xie, and Mike Zheng Shou. Wmadapter: Adding watermark control to latent diffusion models. arXiv preprint arXiv:2406.08337, 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "type": "text", + "content": "[14] Özgün Çiçek, Ahmed Abdulkadir, Soeren S Lienkamp, Thomas Brox, and Olaf Ronneberger. 3d u-net: learning dense volumetric segmentation from sparse annotation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2016: 19th International Conference," + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "Athens, Greece, October 17-21, 2016, Proceedings, Part II 19, pages 424-432. Springer, 2016. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 95, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 545, + 128 + ], + "type": "text", + "content": "[15] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 129, + 545, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 129, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 308, + 129, + 545, + 182 + ], + "type": "text", + "content": "[16] Weitao Feng, Wenbo Zhou, Jiyan He, Jie Zhang, Tianyi Wei, Guanlin Li, Tianwei Zhang, Weiming Zhang, and Nenghai Yu. Aqualora: Toward white-box protection for customized stable diffusion models via watermark lora. arXiv preprint arXiv:2405.11135, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 183, + 547, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 183, + 547, + 237 + ], + "spans": [ + { + "bbox": [ + 308, + 183, + 547, + 237 + ], + "type": "text", + "content": "[17] Pierre Fernandez, Guillaume Couairon, Hervé Jégou, Matthijs Douze, and Teddy Furon. The stable signature: Rooting watermarks in latent diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22466-22477, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 239, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 239, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 308, + 239, + 545, + 293 + ], + "type": "text", + "content": "[18] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Imagine your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 293, + 545, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 293, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 308, + 293, + 545, + 327 + ], + "type": "text", + "content": "[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 327, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 381 + ], + "type": "text", + "content": "[20] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsanko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 382, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 382, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 308, + 382, + 545, + 425 + ], + "type": "text", + "content": "[21] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 426, + 545, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 545, + 468 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 545, + 468 + ], + "type": "text", + "content": "[22] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 470, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 470, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 470, + 545, + 514 + ], + "type": "text", + "content": "[23] Robert Honig, Javier Rando, Nicholas Carlini, and Florian Tramér. Adversarial perturbations cannot reliably protect artists from generative ai. arXiv preprint arXiv:2406.12027, 2024. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 514, + 545, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 514, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 308, + 514, + 545, + 558 + ], + "type": "text", + "content": "[24] Li Hu. Animate anyone: Consistent and controllable image-to-video synthesis for character animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8153-8163, 2024. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 559, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 624 + ], + "type": "text", + "content": "[25] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 624, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 545, + 668 + ], + "type": "text", + "content": "[26] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5148-5157, 2021. 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 669, + 545, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 669, + 545, + 680 + ], + "spans": [ + { + "bbox": [ + 308, + 669, + 545, + 680 + ], + "type": "text", + "content": "[27] PKU-Yuan Lab and Tuzhan AI etc. Open-sora-plan, 2024. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "text", + "content": "[28] Guanlin Li, Shuai Yang, Jie Zhang, and Tianwei Zhang. Prime: Protect your videos from malicious editing. arXiv preprint arXiv:2402.01239, 2024. 3, 5" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12603" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "text", + "content": "[29] Chumeng Liang and Xiaoyu Wu. Mist: Towards improved adversarial examples for diffusion models. arXiv preprint arXiv:2305.12683, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 288, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 288, + 161 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 288, + 161 + ], + "type": "text", + "content": "[30] Chumeng Liang, Xiaoyu Wu, Yang Hua, Jiaru Zhang, Yiming Xue, Tao Song, Zhengui Xue, Ruhui Ma, and Haibing Guan. Adversarial example does good: Preventing painting imitation from diffusion models via adversarial examples. arXiv preprint arXiv:2302.04578, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 162, + 288, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 162, + 288, + 204 + ], + "spans": [ + { + "bbox": [ + 48, + 162, + 288, + 204 + ], + "type": "text", + "content": "[31] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International conference on machine learning, pages 8162-8171. PMLR, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 205, + 288, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 205, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 205, + 288, + 239 + ], + "type": "text", + "content": "[32] Josephine Passananti, Stanley Wu, Shawn Shan, Haitao Zheng, and Ben Y Zhao. Disrupting style mimicry attacks on video imagery. arXiv preprint arXiv:2405.06865, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 239, + 288, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 288, + 282 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 288, + 282 + ], + "type": "text", + "content": "[33] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 284, + 288, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 284, + 288, + 326 + ], + "spans": [ + { + "bbox": [ + 48, + 284, + 288, + 326 + ], + "type": "text", + "content": "[34] Bohao Peng, Jian Wang, Yuechen Zhang, Wenbo Li, MingChang Yang, and Jiaya Jia. Controlnext: Powerful and efficient control for image and video generation. arXiv preprint arXiv:2408.06070, 2024. 2, 4, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 327, + 288, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 288, + 382 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 288, + 382 + ], + "type": "text", + "content": "[35] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 383, + 288, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 383, + 288, + 449 + ], + "spans": [ + { + "bbox": [ + 48, + 383, + 288, + 449 + ], + "type": "text", + "content": "[36] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 449, + 288, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 288, + 502 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 288, + 502 + ], + "type": "text", + "content": "[37] Hadi Salman, Alaa Khaddaj, Guillaume Leclerc, Andrew Ilyas, and Aleksander Madry. Raising the cost of malicious ai-powered image editing. In Proceedings of the 40th International Conference on Machine Learning, pages 29894–29918, 2023. 2, 4, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 503, + 288, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 503, + 288, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 503, + 288, + 557 + ], + "type": "text", + "content": "[38] Shawn Shan, Jenna Cryan, Emily Wenger, Haitao Zheng, Rana Hanocka, and Ben Y Zhao. Glaze: Protecting artists from style mimicry by {Text-to-Image} models. In 32nd USENIX Security Symposium (USENIX Security 23), pages 2187–2204, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 558, + 288, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 288, + 603 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 288, + 603 + ], + "type": "text", + "content": "[39] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 603, + 288, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 288, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 288, + 656 + ], + "type": "text", + "content": "[40] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 689 + ], + "type": "text", + "content": "[41] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 690, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 690, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 690, + 288, + 713 + ], + "type": "text", + "content": "[42] Jingwei Sun, Xuchong Zhang, Changfeng Sun, Qicheng Bai, and Hongbin Sun. Latent feature and attention dual erasure" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 635 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 326, + 73, + 547, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 547, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 547, + 95 + ], + "type": "text", + "content": "attack against multi-view diffusion models for 3d assets protection. arXiv preprint arXiv:2408.11408, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 95, + 450, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 450, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 450, + 106 + ], + "type": "text", + "content": "[43] Kling AI Team. Kling ai, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 108, + 443, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 443, + 118 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 443, + 118 + ], + "type": "text", + "content": "[44] Luma AI Team. Luma ai, 2024." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 120, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 120, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 120, + 545, + 140 + ], + "type": "text", + "content": "[45] The Movie Gen team @ Meta. Movie gen: A cast of media foundation models. 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 143, + 547, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 143, + 547, + 175 + ], + "spans": [ + { + "bbox": [ + 307, + 143, + 547, + 175 + ], + "type": "text", + "content": "[46] Vu Tuan Truong, Luan Ba Dang, and Long Bao Le. Attacks and defenses for generative diffusion models: A comprehensive survey. arXiv preprint arXiv:2408.03400, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 177, + 545, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 177, + 545, + 198 + ], + "spans": [ + { + "bbox": [ + 307, + 177, + 545, + 198 + ], + "type": "text", + "content": "[47] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 200, + 545, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 200, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 307, + 200, + 545, + 232 + ], + "type": "text", + "content": "[48] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 233, + 545, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 233, + 545, + 287 + ], + "spans": [ + { + "bbox": [ + 307, + 233, + 545, + 287 + ], + "type": "text", + "content": "[49] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 289, + 545, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 289, + 545, + 332 + ], + "spans": [ + { + "bbox": [ + 307, + 289, + 545, + 332 + ], + "type": "text", + "content": "[50] Yuxin Wen, John Kirchenbauer, Jonas Geiping, and Tom Goldstein. Tree-ring watermarks: Fingerprints for diffusion images that are invisible and robust. arXiv preprint arXiv:2305.20030, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 334, + 545, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 334, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 307, + 334, + 545, + 399 + ], + "type": "text", + "content": "[51] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7623-7633, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 401, + 545, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 401, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 307, + 401, + 545, + 434 + ], + "type": "text", + "content": "[52] Zhen Xing, Qijun Feng, Haoran Chen, Qi Dai, Han Hu, Hang Xu, Zuxuan Wu, and Yu-Gang Jiang. A survey on video diffusion models. ACM Computing Surveys, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 434, + 545, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 434, + 545, + 489 + ], + "spans": [ + { + "bbox": [ + 307, + 434, + 545, + 489 + ], + "type": "text", + "content": "[53] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 2, 4, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 491, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 491, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 307, + 491, + 545, + 533 + ], + "type": "text", + "content": "[54] Jianping Zhang, Zhuoer Xu, Shiwen Cui, Changhua Meng, Weibin Wu, and Michael R Lyu. On the robustness of latent diffusion models. arXiv preprint arXiv:2306.08257, 2023. 2, 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 535, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 535, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 535, + 545, + 578 + ], + "type": "text", + "content": "[55] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all, 2024. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 580, + 547, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 580, + 547, + 635 + ], + "spans": [ + { + "bbox": [ + 307, + 580, + 547, + 635 + ], + "type": "text", + "content": "[56] Peifei Zhu, Tsubasa Takahashi, and Hirokatsu Kataoka. Watermark-embedded adversarial examples for copyright protection against diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24420–24430, 2024. 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "12604" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/IAAO_ Interactive Affordance Learning for Articulated Objects in 3D Environments/c94d31b4-1877-4bc2-8b0f-9d4d8e45806a_content_list.json b/2025/IAAO_ Interactive Affordance Learning for Articulated Objects in 3D Environments/c94d31b4-1877-4bc2-8b0f-9d4d8e45806a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c886e8b03597a61746f062328cf78d2ee9bcd796 --- /dev/null +++ b/2025/IAAO_ Interactive Affordance Learning for Articulated Objects in 3D Environments/c94d31b4-1877-4bc2-8b0f-9d4d8e45806a_content_list.json @@ -0,0 +1,1455 @@ +[ + { + "type": "text", + "text": "IAAO: Interactive Affordance Learning for Articulated Objects in 3D Environments", + "text_level": 1, + "bbox": [ + 148, + 128, + 849, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Can Zhang Gim Hee Lee", + "bbox": [ + 377, + 204, + 620, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Department of Computer Science, National University of Singapore", + "bbox": [ + 230, + 222, + 764, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "can.zhang@u.nus.edu gimhee.lee@nus.edu.sg", + "bbox": [ + 297, + 242, + 694, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This work presents IAAO, a novel framework that builds an explicit 3D model for intelligent agents to gain understanding of articulated objects in their environment through interaction. Unlike prior methods that rely on task-specific networks and assumptions about movable parts, our IAAO leverages large foundation models to estimate interactive affordances and part articulations in three stages. We first build hierarchical features and label fields for each object state using 3D Gaussian Splatting (3DGS) by distilling mask features and view-consistent labels from multi-view images. We then perform object- and part-level queries on the 3D Gaussian primitives to identify static and articulated elements, estimating global transformations and local articulation parameters along with affordances. Finally, scenes from different states are merged and refined based on the estimated transformations, enabling robust affordance-based interaction and manipulation of objects. Experimental results demonstrate the effectiveness of our method. Our source code is available at: https://lulusindazc.github.io/IAAOproject/.", + "bbox": [ + 89, + 325, + 483, + 628 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 661, + 220, + 678 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The real world contains various dynamic objects that continuously interact and move through time. Understanding and reconstructing the shapes, poses, and articulations of dynamic objects is crucial for many applications such as robotics, augmented reality and virtual reality (AR/VR), etc. While humans naturally develop an intuitive understanding of how objects move and interact, teaching this skill to an intelligent agent remains a complex challenge due to the diverse and intricate nature of articulated objects. Furthermore, a key capability still missing in intelligent agents is effective interaction with their environment, particularly with functional elements like door handles and light switches. For intelligent agents to interact with functional objects, in addition to object detection, they must also", + "bbox": [ + 89, + 688, + 483, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/27f5e2f12eb4553f6ad1350557013517bc204be00f41b58200ba4c3ecd77e8c2.jpg", + "image_caption": [ + "Figure 1. Our IAAO requires multi-view images of the object or indoor scene from two different joint states (Left). The output is a 3D interactive field which supports interactions with multiple movable parts for fine-grained segmentation (e.g. Case 2: handles) and articulation reconstruction (e.g. Case 1: two articulated doors)." + ], + "image_footnote": [], + "bbox": [ + 509, + 291, + 915, + 422 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "understand the specific physical interactions each object affords—known as affordance [10]. The small size of these elements further intensifies the challenge, making them difficult to detect. In this work, we aim to reconstruct the 3D complex scene and interact with both objects with static geometry and objects with articulations and fine-grained affordances as illustrated in Fig. 1.", + "bbox": [ + 511, + 535, + 906, + 642 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Early approaches [21, 46] to understanding articulated objects rely on supervised learning with costly 3D annotations and category-specific models, which limit generalization to novel objects. Recent methods [16, 22] such as Ditto [16] and PARIS [22] have advanced this field by creating digital twins of objects based on observations in two joint configurations. Since Ditto is trained on multiview point clouds from specific categories, it struggles with objects that differ significantly from its training data. In contrast, PARIS [22] uses multi-view posed images to directly optimize an implicit representation without pretraining, allowing for improved generalization. However, it remains sensitive to initialization and may experience stability issues. Further developments [42, 47] have expanded scalability to objects with multiple moving parts based on implicit neural radiance fields (NeRFs). Nonetheless, several challenges remain unexplored, including generaliz", + "bbox": [ + 511, + 643, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12132", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing to complex scenes, flexibly applied for manipulation, and practical assumptions (e.g. handling unknown camera poses, unaligned static object parts, and occlusion).", + "bbox": [ + 89, + 90, + 480, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In 3D perception, instance segmentation provides essential object-level understanding for agents interacting within scenes. Recent studies [27, 29] explore part-object segmentation to decompose objects into finer components such as cabinet drawers. However, these methods are preliminary and struggle with affordance detection due to the lack of precise small-scale geometry in existing datasets. Open-vocabulary models [6, 18, 35, 43] utilize foundation models such as CLIP [39] to extract semantic features and consolidate them into a multi-view consistent 3D neural field representation. However, they rely on costly multi-scale queries to refine segmentation boundaries, and face challenges in high-dimensional feature encoding.", + "bbox": [ + 88, + 138, + 482, + 335 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the limitations mentioned above, we introduce IAAO, a framework designed for interaction of intelligent agents with objects, parts, and functional affordances, as well as for articulating movable objects. We explicitly leverage 3D semantic information to provide our model with scene understanding and interaction capabilities for articulated objects. Key insights of our approach include: 1) semantic scene reconstruction where SAM masks enable multi-view consistent segmentations from partial views, 2) hierarchical feature fields to efficiently compress dense features from multiple views into 3DGS for object- and part-level interaction in diverse scenes, 3) explicit 3D Gaussian representations for affordance localization via query similarity and motion recovery via 3D point-to-2D pixel correspondences, 4) global-local transformation initialization to enhance robustness in realistic scenarios with potential scene misalignment between object states, and 5) scene state fusion merging two different scene states to enhance object geometry completeness.", + "bbox": [ + 91, + 338, + 482, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, our IAAO enables interactive affordance detection and object articulation via the following steps: 1) 3D model construction. Using multi-view posed images of each state, we model the 3D scene as a set of explicit 3D Gaussian primitives with 3D Gaussian Splatting (3DGS) [17]. 2) Hierarchical feature field construction. We propose an efficient distillation scheme where shape-aware 2D features and dense semantic features are extracted from large foundation models, including CLIP [39], SAM [19], and DINOv2 [33]. 2D features are efficiently distilled into 3D fields with a decoder. 3) Semantic-guided mask association across states. We first cluster SAM-generated masks from all views within each state to obtain view-consistent mask labels. We then compute 3D proposal features by merging the features from the corresponding 2D masks. By comparing pairwise feature similarities, we establish 3D mask-level correspondences across states that result in consistent mask label sets for the entire scene in each", + "bbox": [ + 91, + 628, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "state. 4) Affordance prediction. Given a task description, we encode it with the CLIP text encoder and compare the text embeddings with the feature fields to identify the relevant affordance. 5) Motion recovery. To estimate transformations, we define a set of consistency and matching losses that recover both global scene-level and local part-level transformations that ensure accurate motion representation across different states.", + "bbox": [ + 511, + 90, + 903, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We summarize our main contributions as follows:", + "bbox": [ + 532, + 212, + 870, + 226 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce IAAO, an interactive affordance system utilizing 3D Gaussian fields embedded with hierarchical language-aligned semantics and class-agnostic masks. This system supports manipulation tasks guided by various prompts (point, mask, and language) at both object and part levels, regardless of object categories and shape.", + "- We propose a method to reconstruct motion via global matching at the scene level and local matching for articulated objects, without relying on impractical assumptions about static object alignment or known camera poses.", + "- Our IAAO achieves state-of-the-art performance across extensive experiments on multiple benchmarks, including synthetic, real-world, and indoor scene data.", + "- We show strong model generalization to complex indoor environments and previously unseen articulated objects, with no restrictions on the number of movable parts." + ], + "bbox": [ + 511, + 227, + 903, + 468 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 482, + 653, + 498 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural Fields for 3D Scene Understanding. 3D Scene understanding is primarily characterized by four classic representations, including volumetric fields ([34]), point clouds ([40]), 3D meshes ([7]), and depth maps ([11, 41, 54]). Unlike traditional representations, NeRF (Neural Radiance Fields) [28] introduces a neural implicit field to capture the geometry and appearance of a scene. Through a MultiLayer Perceptron (MLP) that takes 3D positions and 2D view directions as inputs, NeRF learns to implicitly represent the color and radiance of the scenes from a collection of posed images. However, a key drawback of NeRF is the slow training and rendering speed. Recently, 3D Gaussian Splatting (3DGS) [17] has been introduced as an alternative to implicit radiance field representations. In contrast to NeRF, 3DGS explicitly represents radiance fields as a collection of oriented 3D Gaussians. Each Gaussian is defined by its spatial position, opacity, and a covariance matrix, which allows flexible optimization. With efficient differentiable rasterization, 3DGS achieves fast rendering with high-quality results. Various work leverage neural fields as a representation robotic manipulation [18, 44, 45, 48, 58, 60]. Methods that leverage visual foundation models to construct neural feature fields [18, 37, 59] are most relevant to our work. With feature distillation via rendering, we reconstruct a hierarchical feature field for object localization and fine-grained affordance prediction.", + "bbox": [ + 511, + 507, + 906, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12133", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Visual Affordance. Affordance [9] refers to the potential interactions that an object or its parts facilitate for an agent. Affordance prediction involves deducing interaction opportunities from visual representations, e.g. images [4, 8, 25] or 3D models [3, 30, 31]. This field has gained significant interest in robotics as a foundational element for tasks e.g. grasping [26, 55], planning [50] and exploration [31]. However, although these methods excel at predicting affordance regions, they lack detailed interaction information, which is essential for engaging with functional elements. Recent work [2] addresses fine-grained functional elements using comprehensive natural language task descriptions for interaction. However, it depends on high-fidelity point clouds, restricting its practical applicability. We use high-resolution images with a Vision-Language Model (VLM) and 3D motion heuristics to predict affordances, eliminating the need for high-fidelity point clouds.", + "bbox": [ + 89, + 90, + 480, + 347 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Articulation Reasoning by Interaction. As embodied AI continues to evolve, the study of articulated objects has become a critical area of research. Recently, several synthetic [46, 49] and scanned data [15, 24, 27, 36] datasets have been introduced for articulated objects. These datasets contain annotations for part segmentation and motion parameters that enable data-driven approaches in the prediction of motion parameters from 3D point clouds [46, 53]. Recent works have been increasingly focused on real-world scenarios by detecting articulated parts and the motion parameters from images [15, 56] and videos [36]. Interactive perception involves agents gaining new insights from their interactions with the environment, and has been widely used to study object articulations [13, 14, 16, 32]. In this work, we focus on using language, and geometric and semantic scene understanding to achieve articulation reasoning.", + "bbox": [ + 89, + 349, + 482, + 592 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Our Method", + "text_level": 1, + "bbox": [ + 89, + 609, + 220, + 626 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Objective. Given a scene of articulated objects represented by a set of multi-view posed images $\\mathcal{I}^t = \\{I_i^t\\}_{i=1}^{N^t}$ at two states $t$ and $t'$ , the goal is to reconstruct the geometries and semantics of the objects in the scene. We focus on articulated objects which contain one or more movable parts with rotary or prismatic joints. Under the condition of unknown object categorory in the scene, we aim to recover the object part geometry, segmentation and joint articulations in the presence of occlusions, shape discrepancies and motion changes across different scene states. For static objects in the scene, the task degenerates to global transformation estimation between two input scene states for state matching.", + "bbox": [ + 89, + 638, + 482, + 820 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overview. Fig. 2 shows our IAAO framework consisting of: 1) A semantic scene reconstruction stage (cf. Sec. 3.1) where we first extract CLIP features for objects in individual scene state. These features are regularized by class-agnostic masks generated from 2D foundation models, e.g.", + "bbox": [ + 89, + 824, + 482, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SAM [57] at the instance and part level. We then cluster view-inconsistent masks from all views to obtain 3D consistent labels, and build a hierarchical feature field and label field for all masks in each scene state via 3D Gaussian Splatting (3DGS). 2) An affordance and motion prediction stage (cf. Sec. 3.2) where we perform object- and part-level queries by directly operating on 3D primitives represented by 3D Gaussians. Static elements and articulated objects are identified by isolating the groups of 3D Gaussians corresponding to segmented objects and functional parts of the scene, i.e. affordance prediction. For motion recovery, we estimate global transformation from static Gaussian primitives and local articulation parameters from the segmented movable parts of objects. 3) A scene state fusion stage (cf. Sec. 3.3) enabling affordance interaction and object manipulation by merging and refining the reconstructed scenes from two states according to the estimated transformations.", + "bbox": [ + 511, + 90, + 903, + 348 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Semantic Scene Reconstruction", + "text_level": 1, + "bbox": [ + 511, + 362, + 790, + 376 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a multi-view capture $\\mathcal{I}^t$ of scene state $t$ , we construct a 3D model using 3DGS initialized with sparse point cloud generated from Structure from Motion (SfM) [40]. Each point serves as the center of a Gaussian primitive embedded with geometry and appearance parameters. These 3D Gaussian primitives $G^{t} = g_{p,p = 1}^{tP^{t}}$ are then rendered into 2D views via differentiable rasterization for parameter optimization. To enrich the reconstructed model with semantic information, we augment each 3D Gaussian with semantic feature embeddings derived from large foundation models.", + "bbox": [ + 511, + 385, + 906, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "View-Consistent Mask Clustering. For each input image $I_{i}^{t} \\in \\mathcal{I}^{t}$ , we first derive the class-agnostic masks $M_{i}^{t} = \\{m_{i,n}^{t}\\}_{n=1}^{n_{i}^{t}}$ with an off-the-shelf mask predictor, where $n_{i}^{t}$ represents the number of masks in view $i$ . For $M_{i}^{t}$ , we investigated the use of SAM [57] to generate both instance- and part-level masks for arbitrary objects. Unfortunately, mask generation models such as SAM face several challenges: a) class-agnostic masks do not have complete 3D information, b) inconsistent segmentation across images due to variations in viewpoint and appearance, and c) lack of one-to-one correspondence among masks due to oversegmentation for individual object. We aim to generate 3D-consistent labels for all masks in $M^{t}$ by defining a mask label mapping function $\\phi(M^{t}, G^{t}) \\mapsto O^{t}$ . This function maps a set of masks $M_{o}^{t}$ belonging to the same 3D object to a consistent $o^{t} \\in O^{t}$ within the 3D Gaussian primitives $G^{t}$ .", + "bbox": [ + 511, + 547, + 906, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To associate class-agnostic masks across all input views at scene state $t$ , we construct a mask graph to fuse the 2D masks $M^t$ of the entire scene into cohesive 3D instances $O^t$ . In the initial graph $\\mathcal{G}_0^t = (\\mathcal{V}_0^t, \\mathcal{E}_0^t)$ , each node in $\\mathcal{V}_0^t$ represents a mask detected from the views, and each edge in $\\mathcal{E}_0^t$ represents potential matches between pairs of object parts. Following the mask clustering workflow from [52],", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12134", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/adec8629e5093ea4f04cfe7247ed41f44f0f7587894a7e590e29ce644fe54d84.jpg", + "image_caption": [ + "Stage 1: Semantic Scene Reconstruction" + ], + "image_footnote": [], + "bbox": [ + 99, + 109, + 736, + 210 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d03463a4d5b791fc4f27e28523e5f8f7ae47420cf69378f4106587d1870a5935.jpg", + "image_caption": [ + "Stage 3: Scene State Fusion", + "Figure 2. Our IAAO framework. 1) Top: Constructing 3D Gaussian fields in each state. We optimize 3DGS fields with hierarchical mask features, DINOv2 features and 3D-consistent mask labels generated from multi-view images. We also incorporate geometry information from depth images into the 3D Gaussians. 2) Bottom: Affordance and motion prediction. A query prompt is embedded using a pretrained encoder to localize relevant regions in the 3D Gaussians. For motion prediction, we optimize the transformation parameters by applying consistency and matching losses to 2D-3D correspondences between states. 3) Right: Scene fusion. Using the estimated transformations, we merge reconstructed 3DGS models from both states, aligning static and articulated elements." + ], + "image_footnote": [], + "bbox": [ + 753, + 117, + 895, + 327 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/82cd04537932a936e0cf0481341cf4e78840e1c4733b5c9ea5d629700c38e946.jpg", + "image_caption": [ + "Stage 2: Affordance and Motion Prediction" + ], + "image_footnote": [], + "bbox": [ + 99, + 232, + 741, + 329 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "we then cluster the mask nodes and update the edges to facilitate mask association across different views. For each cluster, we combine the corresponding partial sets of Gaussians from individual masks to form a complete 3D instance and establish correspondences between 2D masks and 3D nodes. This results in the final graph $\\mathcal{G}^t = (\\mathcal{V}^t,\\mathcal{E}^t)$ , where each node in $\\mathcal{V}^t$ represents a 3D instance proposal in $O^t$ in scene state $t$ . We maintain a list of associated 2D masks $M_o^t$ for each $o^t$ , discarding severely occluded masks. This 2D-3D relationship enables consistent segmentation labels across all views. However, the mask label mapping $\\phi (M^{t},G^{t})$ may be affected by noise due to varying part segmentation hierarchies across views. To mitigate segmentation ambiguities, we use the labeled masks $M_o^t$ in the final graph to supervise a label field for object part $o^t$ in the given scene state. To this end, we use a cross-entropy classification loss $\\mathcal{L}_{\\mathrm{label}}$ for enhanced consistency.", + "bbox": [ + 91, + 452, + 483, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Per-State Hierarchical Feature Field. We generate a set of feature embeddings from the given image $I_{i}^{t}$ using pretrained foundation models. Specifically, we utilize MaskCLIP [5] to produce mask-level CLIP features at both instance and part levels, and DINOv2 [1, 33] to extract DINO embeddings. To enable object- and part-level interaction, we construct a hierarchical feature field by embedding dense features from multi-view images into the 3D Gaussian primitives. Using semantic hierarchy from SAM, we enhance our 3D semantic field accuracy while making the querying process more efficient. Directly embedding high-dimensional 2D features would result in excess", + "bbox": [ + 89, + 719, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "sive computational and memory costs. Therefore, we build a low-dimensional latent feature field and introduce a decoder to project the rendered features $\\hat{F}^t$ back to the 2D feature space $F^t$ during the differentiable rasterization optimization process as:", + "bbox": [ + 511, + 452, + 906, + 527 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {f e a t} = \\left\\| \\mathcal {D} \\left(\\hat {F} ^ {t} \\left(I _ {i} ^ {t}\\right)\\right) - F ^ {t} \\left(I _ {i} ^ {t}\\right) \\right\\| _ {2} ^ {2}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 537, + 906, + 558 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The decoder $\\mathcal{D}$ is implemented as a small MLP with three output branches: two for instance- and part-level CLIP features, and one for DINO features.", + "bbox": [ + 511, + 566, + 905, + 611 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Semantic Association between Scene States. Upon obtaining the mask label mapping functions $O^t$ and $O^{t'}$ for scene states $t$ and $t'$ , we establish a global 3D mask association which links the same semantic regions across both states into a unified 3D mask using their semantic features. For each 3D proposal in $O^t$ , we select the top- $k$ representative masks from $M_o^t$ and aggregate their rendered features to create the corresponding semantic primitives $F_o^t$ . We then build an affinity matrix capturing the cosine similarity between all 3D proposal pairs $(O^t, O^{t'})$ using their primitives $(F_O^t, F_O^{t'})$ from two scene states. For each query proposal $o^t \\in O^t$ , we define its visual match as the target proposal in $O^{t'}$ with the highest similarity.", + "bbox": [ + 511, + 612, + 906, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Affordance and Motion Prediction", + "text_level": 1, + "bbox": [ + 511, + 818, + 815, + 832 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Unlike NeRF-based methods [6, 18, 23] which require resource-intensive rendering to derive language-embedded features and geometry from implicit MLPs, our IAAO directly uses explicit Gaussian primitives for efficient object", + "bbox": [ + 511, + 839, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12135", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "part localization and manipulation. Given the reconstructed 3D scene $G^{t} = \\{g_{p}^{t}\\}_{p = 1}^{P^{t}}$ at scene state $t$ , where each Gaussian point $g_{p}^{t}$ is embedded with geometric information (coordinates of points $p^t$ ) and additional features (color and semantic embeddings), our goal is to enable scene-aware affordance prediction and object articulation estimation.", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Functional Affordance Prediction. Functional affordances are defined as parts of the scene that facilitate interactions for agents to perform specific tasks. We utilize the reconstructed hierarchical feature and label fields as the 3D representations to predict the masks $\\{m_a\\}_{a = 1}^{K_a}$ and affordance labels $\\{l_a\\}_{a = 1}^{K_a}$ in the scene, where $K_{a}$ is the total number of affordance categories. We first build associations between affordance labels and our reconstructed mask label field, and then infer the point-level affordance of a functional element in the scene. Additionally, for task-specific reasoning, we encode the task description using the CLIP text encoder. To identify the relevant mask for a given description, we calculate the similarity between the embeddings in the feature field and the query embeddings. A mask is retrieved if its similarity score with the description text exceeds a predefined threshold.", + "bbox": [ + 89, + 185, + 482, + 425 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Global and Local Motion Recovery. Given the reconstructed scenes with different object articulations at two states, our objective is to estimate scene-aware transformation and 3D motion primitives for articulated parts. Our approach segments the reconstructed Gaussians $G^{t}$ and $G^{t'}$ into static scene elements and articulated objects via object-part querying for global and per-part local transformation estimation, respectively. 1) Object-level querying in complex scenes: In an indoor scene with a complex background and multiple objects, we start with object-level querying. Positive language queries are used to identify target objects for manipulation, while optional negative queries exclude irrelevant items. This selection process identifies 3D Gaussians with object-level CLIP features that align more closely with the positive queries than the negative ones, following standard CLIP feature-text comparison practices [45]. 2) Part-level querying for targeted objects: For individual objects in the scene, we apply part-level querying to focus on 3D Gaussians within the object. This allows for fine-grained transformation estimation specific to each articulated part. In addition to language, other user input prompts such as masks and points can also be used. Nonetheless, we do not further elaborate on these prompts as they can be easily obtained by a user clicking on the images.", + "bbox": [ + 89, + 429, + 482, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We define the global scene-aware transformation between two different scene states as $\\xi_g^t = (s_g^t,R_g^t,T_g^t)$ where $s_g^t\\in \\mathbb{R}$ represents the scale factor, $R_{g}^{t}\\in \\mathrm{SO}(3)$ is the rotation matrix and $T_{g}^{t}\\in \\mathbb{R}^{3}$ is the translation vector. To initialize the transformation, we estimate an initial alignment using 3D Gaussians derived from static elements in the scene, excluding articulated objects. From the Gaussian", + "bbox": [ + 89, + 795, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "points in both scene states, we select reliable points with opacity values above a threshold (empirically set at 0.7) to serve as input point sets for the coarse registration. Using the GeoTransformer [38], we calculate the scene alignment transformation by matching coarse superpoints and refining with dense point correspondences. Since these point clouds are based on reconstructed 3D Gaussians, the resulting transformation is coarse, providing an initial approximation for $\\xi_{g}^{t}$ that can be further refined.", + "bbox": [ + 511, + 90, + 903, + 227 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After achieving global alignment between two scene states, we define the local transformation for each articulated part as $\\xi_{o}^{t} = (R_{o}^{t},T_{o}^{t})$ , representing the movement of part $o$ from state $t$ to state $t'$ . All points within a part are assumed to follow the same motion parameters. The reverse mapping, which transforms part $o$ from state $t'$ back to state $t$ , is given by the inverse transformation, $\\xi_{o}^{t'} = (R_{o}^{t},T_{o}^{t})^{-1}$ . To facilitate this process, we construct a 3D Gaussian correspondence field, allowing each Gaussian primitive $g_{p}^{t}(o)$ at state $t$ to be repositioned to its corresponding location $p^{t'}$ at state $t'$ based on the motion parameters of part $o^t$ . Formally:", + "bbox": [ + 511, + 227, + 906, + 396 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np ^ {t \\rightarrow t ^ {\\prime}} = \\left\\{\\begin{array}{l l}f \\left(p ^ {t}, \\xi_ {g} ^ {t}\\right) = R _ {g} ^ {t} p ^ {t} + T _ {g} ^ {t},&\\text {i f} p \\text {i n s t a t i c s c e n e ,}\\\\f \\left(f \\left(p ^ {t}, \\xi_ {g} ^ {t}\\right), \\xi_ {o} ^ {t}\\right),&\\text {o t h e r w i s e .}\\end{array}\\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 402, + 903, + 452 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The same rule applies to the backward transformation function $f^{-1}$ for obtaining $p^{t' \\to t}$ (i.e. mapping from $t'$ back to $t$ ) with the transformations $\\xi_g^{t'}$ and $\\xi_o^{t'}$ .", + "bbox": [ + 511, + 460, + 903, + 507 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Establishing accurate dense correspondences between point clouds in two states is challenging due to the sparsity and noise in 3D points derived from the 3DGS model. To optimize the correspondence field, we instead match 3D Gaussian primitives to 2D pixels across the two states, guided by part geometry and segmentation information. Specifically, for a Gaussian primitive $g_{p}^{t}(o)$ of part $o$ at state $t$ , the corresponding 2D pixels at state $t'$ can be treated as observations from novel views. To enhance consistency, we incorporate mask-level features alongside the rendered image consistency loss $\\mathcal{L}_{rgb}$ . For a 3D proposal $o^t$ in graph $\\mathcal{G}^t$ , we identify the part mask list $\\mathcal{M}_o^{t'}$ associated with the corresponding proposal $o^{t'}$ at state $t'$ . Following the splatting process, we render the feature field of $o^t$ to training view $n$ in target scene state $t'$ and decode the mask-level features by $\\mathcal{D}(\\hat{F}_o^t (I_n^{t'}))$ . We formulate the mask feature loss for part $o$ transformed from state $t$ to target state $t'$ as:", + "bbox": [ + 511, + 507, + 906, + 765 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {m a s k} (o ^ {t \\rightarrow t ^ {\\prime}}) = \\\\ \\Sigma_ {n} ^ {N _ {o} ^ {t ^ {\\prime}}} \\frac {M _ {o} ^ {t ^ {\\prime}} \\left(I _ {n} ^ {t ^ {\\prime}}\\right)}{\\left| M _ {o} ^ {t ^ {\\prime}} \\left(I _ {n} ^ {t ^ {\\prime}}\\right) \\right|} \\cdot \\| \\mathcal {D} \\left(\\hat {F} _ {o} ^ {t} \\left(I _ {n} ^ {t ^ {\\prime}}\\right)\\right) - F _ {o} ^ {t ^ {\\prime}} \\left(I _ {n} ^ {t ^ {\\prime}}\\right) \\| _ {2} ^ {2}, \\tag {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 772, + 903, + 830 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $M_{o}^{t^{\\prime}}(I_{n}^{t^{\\prime}})$ and $F_{o}^{t^{\\prime}}(I_{n}^{t^{\\prime}})$ represent the corresponding part mask and 2D mask-level features of part $o$ in view $I_{n}^{t^{\\prime}}$ at the target state $t^{\\prime}$ . $N_{o}^{t}$ denotes the number of part masks for proposal $o^t$ across all training views.", + "bbox": [ + 511, + 838, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12136", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since the mask-level loss provides only coarse guidance for transformation, we further propose to find dense correspondences between the 3D Gaussian primitives and 2D pixels by comparing their DINO features. We begin by computing the feature similarity matrix $\\alpha_{p\\rightarrow o}(I_n^{t'})$ between each pixel in the part mask $M_o^{t'}(I_n^{t'})$ in view $I_n^{t'}$ and the sampled Gaussian point $g_p^t$ from 3D proposal $o^t$ . We then normalize the similarity matrix $\\alpha_{p\\rightarrow o}(I_n^{t'})$ using a softmax over the entire mask to obtain the weight $\\beta_{p\\rightarrow o}(I_n^{t'})$ . Finally, we identify the corresponding 2D pixel $s_{p\\rightarrow o}^{t'}(I_n^{t'})$ for the sampled 3D point $p^t$ using a weighted sum. See the supplementary material for details on the computation. For each point-pixel pair $(p^t,s_{p\\rightarrow o}^{t'}(I_n^{t'}))$ , we filter out pairs that collide by checking whether $s_{p\\rightarrow o}^{t'}(I_n^{t'})$ falls within the same part mask $o^{t'}$ . We then define the matching loss as follows:", + "bbox": [ + 89, + 90, + 483, + 330 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {m a t c h}} = \\left\\| \\pi_ {n} ^ {t ^ {\\prime}} \\left(p ^ {t \\rightarrow t ^ {\\prime}}\\right) - s _ {p \\rightarrow n} ^ {t ^ {\\prime}} \\left(I _ {n} ^ {t ^ {\\prime}}\\right)\\right\\| _ {2} ^ {2}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 342, + 482, + 362 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\pi_n^{t'}$ represents the projection of 3D point $p^{t\\rightarrow t'}$ to view $I_{n}^{t^{\\prime}}$ in scene state $t^\\prime$", + "bbox": [ + 89, + 373, + 482, + 405 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "**Optimization.** For the first stage, each scene is reconstructed independently at two different states. We optionally use depth map for further geometry regularization. In the second stage, we reconstruct motion of global scene and local articulated objects. The total loss is defined as:", + "bbox": [ + 89, + 406, + 483, + 481 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda_ {\\text {c o n s}} \\left(\\mathcal {L} _ {\\text {r g b}} + \\mathcal {L} _ {\\text {m a s k}} + \\mathcal {L} _ {\\text {l a b e l}}\\right) + \\lambda_ {\\text {m a t c h}} \\mathcal {L} _ {\\text {m a t c h}}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 493, + 482, + 511 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\lambda_{\\text{cons}}$ and $\\lambda_{\\text{match}}$ denote the balancing weights of consistency loss and matching loss.", + "bbox": [ + 89, + 521, + 483, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3. Scene State Fusion", + "text_level": 1, + "bbox": [ + 89, + 561, + 272, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After training, we merge the 3DGS models from the two states using the estimated transformations to fill in the occluded regions in each state. Details of our fusion and filtering strategy are provided in the supplementary material. The combined 3D Gaussians facilitate object manipulation by addressing potential artifacts that may appear in occluded areas of the reconstructed scene. Optional fine-tuning can be applied to either the edited regions or the entire scene for further refinement.", + "bbox": [ + 89, + 583, + 483, + 717 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 89, + 733, + 215, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 757, + 281, + 773 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. 1) PARIS two-part object dataset. This dataset contains two-part articulated objects, featuring 10 synthetic objects from PartNet-Mobility [49] and 2 real-world objects captured using MultiScan [27]. Each object comprises a movable part and a static part, observed in two distinct joint states. The dataset includes RGB images, object masks from 100 random viewpoints, and depth images for both synthetic and real-world objects. 2) Synthetic multi-part", + "bbox": [ + 89, + 779, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "object dataset. This dataset includes two synthetic scenes containing multi-part objects from PartNet-Mobility, with each object featuring one static part and multiple movable parts. Each object is observed in two articulation states, and the dataset provides RGB, depth, and mask data from 100 random viewpoints. 3) Indoor scene OmniSim dataset. OmniSim is generated using the OmniGibson [20] simulator with various indoor scene models. By adjusting the rotation of articulated object joints, we generate some scenes for evaluation, each containing RGBD images, interactive object masks and object state metrics at each state.", + "bbox": [ + 511, + 90, + 903, + 257 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. To evaluate articulation models, we use: 1) Axis Ang Err $(^{\\circ})$ means angular error of the predicted joint axis for both revolute and prismatic joints. 2) Axis Pos Err (0.1 m): The shortest distance between the predicted and true joint axes for revolute joints. For cross-state part motion, we use Part Motion Err $(^{\\circ}$ or m). It evaluates joint state precision by calculating the geodesic error for predicted rotations in revolute joints or the Euclidean error for translations in prismatic joints, as specified in PARIS [22]. For object and part mesh reconstruction, we use Chamfer-L1 distance (CD), reporting CD-w for the full surface, CD-s for static parts, and CD-m for movable parts.", + "bbox": [ + 511, + 262, + 905, + 444 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. Ditto [16] and PARIS [22] are models for reconstructing part-level shape and motion in two-part articulated objects, using multi-view data. Ditto assumes one moving part and is pre-trained on specific categories, while PARIS uses a NeRF-based representation to handle unknown objects. $PARIS^{*}$ is an enhanced version with depth supervision, and $PARIS - m^{*}$ extends PARIS to multi-part objects. CSG-reg uses TSDF fusion and Constructive Solid Geometry for part segmentation, with registration for alignment. 3Dseg-reg follows a similar process but employs PA-Conv [51] for part segmentation, reporting results only for trained categories due to limited generalization. DigitalTwinArt [47] divides reconstruction into two stages: 1) It reconstructs object-level shape independently of articulation; 2) It recovers the articulation model by identifying part segmentation and motions through state correspondences.", + "bbox": [ + 511, + 446, + 905, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Setup. There are some key differences on evaluation setup between our model and previous models as we firstly incorporate the semantics into the explicit neural field supporting robust interactions even in complex environment. Unlike prior methods which assume the number of parts is known, our model can detect any articulated objects if enough information for reconstructing object shape is given from multi-view images. To identify corresponding parts for evaluation, prior work need to iterate through all possible pairs of predicted and ground-truth parts, selecting the match with the smallest total Chamfer distance. In our model, we identify the corresponding parts based on the constructed semantic feature and mask label field. We extract the mesh based on the 3D point clouds derived", + "bbox": [ + 511, + 688, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12137", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0517bed912c1e8b9c9a54e9cd9a7a421e055bd7e85774efe6b8b1497953a044f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SimulationReal
FoldChairFridgeLaptopOvenScissorStaplerUSBWasherBladeStorageAllFridgeStorageAll
Axis Ang ↓Ditto [16]89.3589.783.120.964.5089.8689.7789.5179.54*6.3254.221.715.883.80
PARIS [22]8.089.150.020.043.8239.730.1325.3615.380.0310.171.6443.1322.39
PARIS* [22]15.792.930.037.4316.628.170.7118.4041.280.0311.141.9030.1016.00
CSG-reg0.100.270.470.350.280.3011.7871.937.642.829.608.9269.7139.31
3Dseg-reg--2.34-----9.40-----
DigitalTwinArt [47]0.030.070.060.220.110.060.110.430.270.060.142.1018.1110.11
IAAO (Ours)0.020.040.050.170.080.050.090.360.190.060.111.915.828.86
Axis Pos ↓Ditto [16]3.771.02*0.010.135.700.205.410.66--2.111.84-0.92
PARIS [22]0.450.380.000.002.102.272.361.50--1.130.34-0.17
PARIS* [22]0.251.130.000.051.594.673.353.28--1.790.50-0.25
CSG-reg0.020.000.200.180.010.020.012.13--0.321.46-0.73
3Dseg-reg--0.10-----------
DigitalTwinArt [47]0.010.010.000.010.020.010.000.01--0.010.57-0.29
IAAO (Ours)0.010.000.000.000.010.010.010.00--0.010.32-0.16
Part Motion ↓Ditto [16]99.36×5.182.0919.2856.6180.6055.72×0.0939.878.430.384.41
PARIS [22]131.6624.580.030.03120.70110.8064.8560.350.340.3051.362.160.561.36
PARIS* [22]127.3445.260.039.1368.36107.7696.9349.770.360.3050.521.580.571.07
CSG-reg0.130.290.350.580.200.4410.48158.990.050.0417.1614.820.647.73
3Dseg-reg--1.61-----0.15-----
DigitalTwinArt [47]0.160.090.080.110.150.050.110.250.000.000.101.860.201.03
IAAO (Ours)0.130.070.030.040.110.050.090.210.000.000.071.460.150.81
CD-s ↓Ditto [14]33.793.050.252.5239.0741.642.6410.3246.909.1818.9447.0116.0931.55
PARIS [22]9.163.651.941.8812.951.882.6925.391.1912.767.1842.5754.5448.56
PARIS* [22]10.208.8215.581.953.182.482.4812.191.408.676.4611.6420.2515.94
CSG-reg1.691.453.933.263.262.221.954.530.597.062.706.3312.559.44
3Dseg-reg--0.76-----66.31-----
DigitalTwinArt [47]0.180.604.660.402.652.194.804.690.554.692.102.5310.866.69
IAAO (Ours)0.160.510.170.361.981.864.433.700.514.271.802.479.836.15
CD-m ↓Ditto [14]141.110.990.190.9420.6831.2115.8812.89195.932.2042.2050.6020.3535.48
PARIS [22]8.997.760.2128.7046.6419.275.32178.4325.2176.6939.7245.66864.82455.24
PARIS* [22]17.977.230.156.5416.6530.4610.17265.27117.9952.3452.4877.85474.57276.21
CSG-reg1.9121.710.42256.991.956.3629.78436.4226.621.3978.36442.17521.49481.83
3Dseg-reg--1.01-----6.23-----
DigitalTwinArt [47]0.150.270.160.470.412.271.340.361.500.370.731.1426.4613.80
IAAO (Ours)0.130.180.150.410.402.091.180.331.420.250.651.0221.4311.23
CD-w ↓Ditto [14]6.802.160.312.511.702.382.097.2942.043.917.126.5014.0810.29
PARIS [22]1.802.920.3011.7310.493.582.0024.380.608.576.6422.9863.3543.16
PARIS* [22]4.375.530.263.183.905.271.7810.110.587.804.288.9932.1020.55
CSG-reg0.480.980.403.001.701.991.204.480.564.001.885.7114.2910.00
3Dseg-reg--0.81-----0.78-----
DigitalTwinArt [47]0.270.700.354.180.432.191.184.740.363.991.842.199.335.76
IAAO (Ours)0.220.610.273.490.341.961.064.410.343.941.662.029.185.60
", + "bbox": [ + 91, + 87, + 903, + 554 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/83751533b87998ba7768907d728b67924477901c0b950076d77e3a9566afde1e.jpg", + "table_caption": [ + "Table 1. Results on PARIS dataset including both synthetic and real data. 'x': failure case, and \\*\\*: joint axis or position. -': no result is available. Best result in bold." + ], + "table_footnote": [], + "table_body": "
Axis Ang 0↓Axis Ang 1 ↓Axis Pos 0 ↓Axis Pos 1 ↓Part Motion 0 ↓Part Motion 1 ↓CD-s ↓CD-m 0 ↓CD-m 1 ↓CD-w ↓
Fridge-mPARIS*-m [22]34.5215.913.601.6386.21105.868.52526.19160.8615.00
DigitalTwinArt [47]0.160.100.010.000.110.130.610.400.520.89
IAAO (Ours)0.160.080.010.000.090.090.540.440.470.96
Storage-mPARIS*-m [22]43.2626.1810.42-79.840.648.56128.62266.718.66
DigitalTwinArt [47]0.210.880.05-0.130.000.850.213.460.99
IAAO (Ours)0.180.910.04-0.100.000.780.263.350.97
", + "bbox": [ + 91, + 606, + 903, + 694 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Results on multi-part object dataset, averaged over 10 trials with different random seeds. Joint 1 of \"Storage-m\" is solely prismatic with no Axis Pos. '-': no result is available. Best results in bold.", + "bbox": [ + 89, + 702, + 906, + 729 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "from 3DGS model via SuGaR [12] for evaluation. Following [22, 47], we transform our extracted parts with predicted motions to start state $t$ for evaluation.", + "bbox": [ + 89, + 757, + 480, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Quantitative Results", + "text_level": 1, + "bbox": [ + 89, + 816, + 284, + 832 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 1 shows quantitative results from the PARIS Two-Part Object Dataset, covering synthetic and real instances. Ditto performs well on seen categories but struggles with unseen ones. PARIS occasionally fails in shape and articu", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "lation reconstruction, while depth supervision in PARIS* improves shape accuracy for complex and real-world objects but increases articulation errors due to optimization challenges. CSG-reg and 3Dseg-reg handle simple objects well but struggle with complex ones, as segmentation errors affect articulation estimation.", + "bbox": [ + 511, + 757, + 906, + 848 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 2 presents the results for multi-part objects. DigitalTwinArt [47] outperforms PARIS* in shape and articulation reconstruction by incorporating additional supervi", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12138", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/89c9354449d82712c5da8a3ab67c7479c09ada6c7cfb2f93029f474b6e05d0d1.jpg", + "image_caption": [ + "Input State $t$" + ], + "image_footnote": [], + "bbox": [ + 166, + 119, + 228, + 181 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/28dea88c1c409761ac03efc7dde903dfb89b1a47754db10ae7f9139fc7322908.jpg", + "image_caption": [ + "Figure 3. Qualitative analysis of shape reconstruction, part segmentation, and joint prediction results on multi-part object dataset." + ], + "image_footnote": [], + "bbox": [ + 165, + 194, + 243, + 257 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6efbf2f792257d879d606dd525cab3d11e7171ccdea792aa6dd2a762f5743812.jpg", + "image_caption": [ + "Input State $t^{\\prime}$" + ], + "image_footnote": [], + "bbox": [ + 361, + 121, + 431, + 184 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b91e2bc9b5cc41f18b4a764ba040032890c75cd3f83cd9b4a7f3d33c9c0444b5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 352, + 193, + 434, + 258 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7ed2f1a62e2b807252e49a4dcb0d2874bf1e57c923223c9461152b64fd6b7aaf.jpg", + "image_caption": [ + "DigitalTwinArt" + ], + "image_footnote": [], + "bbox": [ + 576, + 114, + 620, + 189 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/13dada9615698bce585a36479a82796c309641397aa2e1a2b74e53412daf17fa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 560, + 195, + 647, + 260 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/037e560b46698e76ad2993d6b34ee4017145eda164add557532a09b933251b3f.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 769, + 114, + 821, + 191 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b84eb0f24ff6a40ca1e10d3f94d80597e779743cfaf49a0d6444b8f4b1a93dfd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 759, + 195, + 844, + 260 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dd0508e1e8dd9913d693ca0808d6dce190c68df45232188a33e61f5b0b27abea.jpg", + "image_caption": [ + "Figure 4. Qualitative results of shape reconstruction, part segmentation, and joint prediction on PARIS." + ], + "image_footnote": [], + "bbox": [ + 106, + 306, + 483, + 462 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f6d070a5e79eaf79715f24ba579dab346a7120340df4ac72474f41ded37c9f64.jpg", + "image_caption": [ + "Figure 5. Motion snapshots on PARIS & multi-part object." + ], + "image_footnote": [], + "bbox": [ + 112, + 505, + 475, + 575 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b2640977660d1bae9da69cdd7570becc2c96823c8753170ed0e1056d770dce37.jpg", + "image_caption": [ + "Figure 6. Motion snapshots on scene-level OmniSim dataset." + ], + "image_footnote": [], + "bbox": [ + 101, + 599, + 473, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "sion on 3D geometries. Our IAAO achieves competitive results by leveraging geometric and semantic information from VLMs, enhancing object part modeling.", + "bbox": [ + 89, + 702, + 483, + 748 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Qualitative Results", + "text_level": 1, + "bbox": [ + 89, + 757, + 272, + 772 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fig. 3 and Fig. 4 show qualitative results on shape reconstruction, part segmentation, and joint prediction for the multi-part object and PARIS two-part datasets. Our IAAO achieves more complete segmentation results while also enabling interaction with affordance and segmentation features. Fig. 5 and 6 show snapshots of the generated motions on several examples at object and scene levels in a simulator. It shows IAAO can produce smooth interpolations of", + "bbox": [ + 88, + 780, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e316cb575ad6a653fb7e95be1df160fc44846afb1e12e0c4b1b9d95d61af5c40.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodAxis Ang ↓Axis Pos ↓Part Motion ↓CD-s ↓CD-m ↓
Proposed0.280.010.110.800.35
w/o matching4.591.388.571.21123.71
w/o mask14.820.010.282.59129.14
w/o rgb0.560.010.170.810.42
w/o label0.540.010.158.56169.73
", + "bbox": [ + 517, + 304, + 901, + 371 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Ablation study on multi-part object dataset.", + "bbox": [ + 552, + 381, + 864, + 395 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "articulated objects with good part geometry modeling.", + "bbox": [ + 511, + 409, + 870, + 425 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 431, + 666, + 449 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We assess the impact of our design choices specifically on multi-part objects, which present greater challenges. We first perform experiments to test out the impact of 2D-3D matching loss ('w/o matching'). As indicated in Tab. 3, the matching loss plays a crucial role in enhancing articulation reconstruction. We also study constraints from mask features and labels, which act as a coarse correspondence mechanism at the superpoint (mask) level. These constraints also enforce consistency within the feature fields of each mask, leading to improved part segmentation and more accurate reconstruction of articulated components.", + "bbox": [ + 511, + 455, + 906, + 621 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 633, + 633, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present IAAO, a framework that facilitates interactive affordance detection and object articulation from two scene sequences. Each scene sequence captures movable parts of objects in different joint states. Our IAAO can handle both simple scenes with single objects and complex indoor scenes without limitations on the number of objects and movable parts. The entire scene is reconstructed using 3D Gaussians, embedded with robust zero-shot generalization capabilities from 2D foundation models. Affordance prediction is achieved through object and part queries on the 3D Gaussian primitives. Motion estimation is achieved by establishing 2D-3D correspondences within each object to track transformations. Static elements serve for global scene alignment while movable elements are for local transformation. Extensive experiments demonstrate that our approach outperforms existing methods.", + "bbox": [ + 511, + 659, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12139", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This research / project is supported by the National Research Foundation (NRF) Singapore, under its NRF-Investigatorship Programme (Award ID. NRF-NRFI09-0008).", + "bbox": [ + 89, + 90, + 485, + 151 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 162, + 187, + 179 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers, 2023. 4", + "[2] Alexandros Delitzas, Ayca Takmaz, Federico Tombari, Robert Sumner, Marc Pollefeys, and Francis Engelmann. Scenefun3d: Fine-grained functionality and affordance understanding in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14531-14542, 2024. 3", + "[3] Shengheng Deng, Xun Xu, Chaozheng Wu, Ke Chen, and Kui Jia. 3d affordancenet: A benchmark for visual object affordance understanding. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1778-1787, 2021. 3", + "[4] Thanh-Toan Do, Anh Nguyen, and Ian Reid. Affordancenet: An end-to-end deep learning approach for object affordance detection. In 2018 IEEE international conference on robotics and automation (ICRA), pages 5882-5889. IEEE, 2018. 3", + "[5] Xiaoyi Dong, Jianmin Bao, Yinglin Zheng, Ting Zhang, Dongdong Chen, Hao Yang, Ming Zeng, Weiming Zhang, Lu Yuan, Dong Chen, et al. Maskclip: Masked self-distillation advances contrastive language-image pretraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10995-11005, 2023. 4", + "[6] Francis Engelmann, Fabian Manhardt, Michael Niemeyer, Keisuke Tateno, Marc Pollefeys, and Federico Tombari. Opennerf: Open set 3d neural scene segmentation with pixelwise features and rendered novel views. arXiv preprint arXiv:2404.03650, 2024. 2, 4", + "[7] Carlos Hernández Esteban and Francis Schmitt. Silhouette and stereo fusion for 3d object modeling. Computer Vision and Image Understanding, 96(3):367-392, 2004. 2", + "[8] Kuan Fang, Te-Lin Wu, Daniel Yang, Silvio Savarese, and Joseph J Lim. Demo2vec: Reasoning object affordances from online videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2139-2147, 2018. 3", + "[9] JJ Gibson. The theory of affordances. Perceiving, acting and knowing: Towards an ecological psychology/Erlbaum, 1977. 3", + "[10] James J Gibson. The ecological approach to visual perception: classic edition. Psychology press, 2014. 1", + "[11] Michael Goesele, Brian Curless, and Steven M Seitz. Multiview stereo revisited. In 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'06), pages 2402-2409. IEEE, 2006. 2", + "[12] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5354-5363, 2024. 7" + ], + "bbox": [ + 93, + 188, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Karol Hausman, Scott Niekum, Sarah Osentoski, and Gaurav S Sukhatme. Active articulation model estimation through interactive perception. In 2015 IEEE International Conference on Robotics and Automation (ICRA), pages 3305-3312. IEEE, 2015. 3", + "[14] Cheng-Chun Hsu, Zhenyu Jiang, and Yuke Zhu. Ditto in the house: Building articulation models of indoor scenes through interactive perception. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 3933-3939. IEEE, 2023. 3, 7", + "[15] Hanxiao Jiang, Yongsen Mao, Manolis Savva, and Angel X Chang. Opd: Single-view 3d openable part detection. In European Conference on Computer Vision, pages 410-426. Springer, 2022. 3", + "[16] Zhenyu Jiang, Cheng-Chun Hsu, and Yuke Zhu. Ditto: Building digital twins of articulated objects from interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5616-5626, 2022. 1, 3, 6, 7", + "[17] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 2", + "[18] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Lerf: Language embedded radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19729-19739, 2023. 2, 4", + "[19] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 2", + "[20] Chengshu Li, Ruohan Zhang, Josiah Wong, Cem Gokmen, Sanjana Srivastava, Roberto Martin-Martín, Chen Wang, Gabrael Levine, Wensi Ai, Benjamin Martinez, et al. Behavior-1k: A human-centered, embodied ai benchmark with 1,000 everyday activities and realistic simulation. arXiv preprint arXiv:2403.09227, 2024. 6", + "[21] Xiaolong Li, He Wang, Li Yi, Leonidas J Guibas, A Lynn Abbott, and Shuran Song. Category-level articulated object pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3706-3715, 2020. 1", + "[22] Jiayi Liu, Ali Mahdavi-Amiri, and Manolis Savva. Paris: Part-level reconstruction and motion analysis for articulated objects. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 352-363, 2023. 1, 6, 7", + "[23] Kunhao Liu, Fangneng Zhan, Jiahui Zhang, Muyu Xu, Yingchen Yu, Abdulmotaleb El Saddik, Christian Theobalt, Eric Xing, and Shijian Lu. Weakly supervised 3d open-vocabulary segmentation. Advances in Neural Information Processing Systems, 36:53433-53456, 2023. 4", + "[24] Liu Liu, Wenqiang Xu, Haoyuan Fu, Sucheng Qian, Qiao-jun Yu, Yang Han, and Cewu Lu. Akb-48: A real-world articulated object knowledge base. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14809–14818, 2022. 3" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12140", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Timo Luddecke and Florentin Worgotter. Learning to segment affordances. In Proceedings of the IEEE International Conference on Computer Vision Workshops, pages 769-776, 2017. 3", + "[26] Priyanka Mandikal and Kristen Grauman. Learning dexterous grasping with object-centric visual affordances. In 2021 IEEE international conference on robotics and automation (ICRA), pages 6169-6176. IEEE, 2021. 3", + "[27] Yongsen Mao, Yiming Zhang, Hanxiao Jiang, Angel Chang, and Manolis Savva. Multiscan: Scalable rgbd scanning for 3d environments with articulated objects. Advances in neural information processing systems, 35:9058-9071, 2022. 2, 3, 6", + "[28] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2", + "[29] Kaichun Mo, Shilin Zhu, Angel X Chang, Li Yi, Subarna Tripathi, Leonidas J Guibas, and Hao Su. Partnet: A largescale benchmark for fine-grained and hierarchical part-level 3d object understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 909-918, 2019. 2", + "[30] Kaichun Mo, Yuzhe Qin, Fanbo Xiang, Hao Su, and Leonidas Guibas. O2o-afford: Annotation-free large-scale object-object affordance learning. In Conference on robot learning, pages 1666-1677. PMLR, 2022. 3", + "[31] Tushar Nagarajan and Kristen Grauman. Learning affordance landscapes for interaction exploration in 3d environments. Advances in Neural Information Processing Systems, 33:2005-2015, 2020. 3", + "[32] Neil Nie, Samir Yitzhak Gadre, Kiana Ehsani, and Shuran Song. Structure from action: Learning interactions for articulated object 3d structure discovery. arXiv preprint arXiv:2207.08997, 2022. 3", + "[33] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 2, 4", + "[34] Sylvain Paris, François X Sillion, and Long Quan. A surface reconstruction method using global graph cut optimization. International Journal of Computer Vision, 66:141-161, 2006. 2", + "[35] Songyou Peng, Kyle Genova, Chiyu Jiang, Andrea Tagliasacchi, Marc Pollefeys, Thomas Funkhouser, et al. Openscene: 3d scene understanding with open vocabularies. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 815-824, 2023. 2", + "[36] Shengyi Qian, Linyi Jin, Chris Rockwell, Siyi Chen, and David F Fouhey. Understanding 3d object articulation in internet videos. In Proceedings of the IEEE/CVF Conference" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "on Computer Vision and Pattern Recognition, pages 1599-1609, 2022. 3", + "[37] Minghan Qin, Wanhua Li, Jiawei Zhou, Haoqian Wang, and Hanspeter Pfister. Langsplat: 3d language gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20051-20060, 2024. 2", + "[38] Zheng Qin, Hao Yu, Changjian Wang, Yulan Guo, Yuxing Peng, and Kai Xu. Geometric transformer for fast and robust point cloud registration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11143-11152, 2022. 5", + "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2", + "[40] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 2, 3", + "[41] Christoph Strecha, Rik Fransens, and Luc Van Gool. Combined depth and outlier estimation in multi-view stereo. In 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'06), pages 2394-2401. IEEE, 2006. 2", + "[42] Archana Swaminathan, Anubhav Gupta, Kamal Gupta, Shishira R Maiya, Vatsal Agarwal, and Abhinav Shrivastava. Leia: Latent view-invariant embeddings for implicit 3d articulation. In European Conference on Computer Vision, pages 210-227. Springer, 2025. 1", + "[43] Ayca Takmaz, Elisabetta Fedele, Robert Sumner, Marc Pollefeys, Federico Tombari, and Francis Engelmann. Openmask3d: Open-vocabulary 3d instance segmentation. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[44] Zhenggang Tang, Balakumar Sundaralingam, Jonathan Tremblay, Bowen Wen, Ye Yuan, Stephen Tyree, Charles Loop, Alexander Schwing, and Stan Birchfield. RGB-only reconstruction of tabletop scenes for collision-free manipulator control. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 1778-1785. IEEE, 2023. 2", + "[45] Can Wang, Mengei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3835-3844, 2022. 2, 5", + "[46] Xiaogang Wang, Bin Zhou, Yahao Shi, Xiaowu Chen, Qinping Zhao, and Kai Xu. Shape2motion: Joint analysis of motion parts and attributes from 3d shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8876-8884, 2019. 1, 3", + "[47] Yijia Weng, Bowen Wen, Jonathan Tremblay, Valts Blukis, Dieter Fox, Leonidas Guibas, and Stan Birchfield. Neural implicit representation for building digital twins of unknown" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12141", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "articulated objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3141-3150, 2024. 1, 6, 7", + "[48] Youngsun Wi, Pete Florence, Andy Zeng, and Nima Fazeli. Virdo: Visio-tactile implicit representations of deformable objects. In 2022 International Conference on Robotics and Automation (ICRA), pages 3583-3590. IEEE, 2022. 2", + "[49] Fanbo Xiang, Yuzhe Qin, Kaichun Mo, Yikuan Xia, Hao Zhu, Fangchen Liu, Minghua Liu, Hanxiao Jiang, Yifu Yuan, He Wang, et al. Sapien: A simulated part-based interactive environment. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11097-11107, 2020. 3, 6", + "[50] Danfei Xu, Ajay Mandlekar, Roberto Martin-Martin, Yuke Zhu, Silvio Savarese, and Li Fei-Fei. Deep affordance foresight: Planning through what can be done in the future. In 2021 IEEE international conference on robotics and automation (ICRA), pages 6206–6213. IEEE, 2021. 3", + "[51] Mutian Xu, Runyu Ding, Hengshuang Zhao, and Xiaojuan Qi. Paconv: Position adaptive convolution with dynamic kernel assembling on point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3173-3182, 2021. 6", + "[52] Mi Yan, Jiazhao Zhang, Yan Zhu, and He Wang. Maskclustering: View consensus based mask graph clustering for open-vocabulary 3d instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28274-28284, 2024. 3", + "[53] Zihao Yan, Ruizhen Hu, Xingguang Yan, Luanmin Chen, Oliver Van Kaick, Hao Zhang, and Hui Huang. RPM-net: recurrent prediction of motion and parts from point cloud. arXiv preprint arXiv:2006.14865, 2020. 3", + "[54] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In Proceedings of the European conference on computer vision (ECCV), pages 767-783, 2018. 2", + "[55] Andy Zeng, Shuran Song, Kuan-Ting Yu, Elliott Donlon, Francois R Hogan, Maria Bauza, Daolin Ma, Orion Taylor, Melody Liu, Eudald Romo, et al. Robotic pick-and-place of novel objects in clutter with multi-affordance grasping and cross-domain image matching. The International Journal of Robotics Research, 41(7):690-705, 2022. 3", + "[56] Vicky Zeng, Tabitha Edith Lee, Jacky Liang, and Oliver Kroemer. Visual identification of articulated object parts. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 2443-2450. IEEE, 2021. 3", + "[57] Chaoning Zhang, Dongshen Han, Yu Qiao, Jung Uk Kim, Sung-Ho Bae, Seungkyu Lee, and Choong Seon Hong. Faster segment anything: Towards lightweight sam for mobile applications. arXiv preprint arXiv:2306.14289, 2023. 3", + "[58] Allan Zhou, Moo Jin Kim, Lirui Wang, Pete Florence, and Chelsea Finn. Nerf in the palm of your hand: Corrective augmentation for robotics via novel-view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17907-17917, 2023. 2" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[59] Shijie Zhou, Haoran Chang, Sicheng Jiang, Zhiwen Fan, Zehao Zhu, Dejia Xu, Pradyumna Chari, Suya You, Zhangyang Wang, and Achuta Kadambi. Feature 3dgs: Supercharging 3d gaussian splattering to enable distilled feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21676-21685, 2024. 2", + "[60] Luyang Zhu, Arsalan Mousavian, Yu Xiang, Hammad Mazhar, Jozef van Eenbergen, Shoubhik Debnath, and Dieter Fox. Rgb-d local implicit function for depth completion of transparent objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4649-4658, 2021. 2" + ], + "bbox": [ + 516, + 90, + 903, + 258 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "12142", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/IAAO_ Interactive Affordance Learning for Articulated Objects in 3D Environments/c94d31b4-1877-4bc2-8b0f-9d4d8e45806a_model.json b/2025/IAAO_ Interactive Affordance Learning for Articulated Objects in 3D Environments/c94d31b4-1877-4bc2-8b0f-9d4d8e45806a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..771788a91e06262ea0724d80602fa765744ac4bb --- /dev/null +++ b/2025/IAAO_ Interactive Affordance Learning for Articulated Objects in 3D Environments/c94d31b4-1877-4bc2-8b0f-9d4d8e45806a_model.json @@ -0,0 +1,2191 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.149, + 0.13, + 0.851, + 0.175 + ], + "angle": 0, + "content": "IAAO: Interactive Affordance Learning for Articulated Objects in 3D Environments" + }, + { + "type": "text", + "bbox": [ + 0.378, + 0.205, + 0.621, + 0.222 + ], + "angle": 0, + "content": "Can Zhang Gim Hee Lee" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.223, + 0.766, + 0.24 + ], + "angle": 0, + "content": "Department of Computer Science, National University of Singapore" + }, + { + "type": "text", + "bbox": [ + 0.299, + 0.243, + 0.695, + 0.257 + ], + "angle": 0, + "content": "can.zhang@u.nus.edu gimhee.lee@nus.edu.sg" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.327, + 0.485, + 0.629 + ], + "angle": 0, + "content": "This work presents IAAO, a novel framework that builds an explicit 3D model for intelligent agents to gain understanding of articulated objects in their environment through interaction. Unlike prior methods that rely on task-specific networks and assumptions about movable parts, our IAAO leverages large foundation models to estimate interactive affordances and part articulations in three stages. We first build hierarchical features and label fields for each object state using 3D Gaussian Splatting (3DGS) by distilling mask features and view-consistent labels from multi-view images. We then perform object- and part-level queries on the 3D Gaussian primitives to identify static and articulated elements, estimating global transformations and local articulation parameters along with affordances. Finally, scenes from different states are merged and refined based on the estimated transformations, enabling robust affordance-based interaction and manipulation of objects. Experimental results demonstrate the effectiveness of our method. Our source code is available at: https://lulusindazc.github.io/IAAOproject/." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.662, + 0.222, + 0.679 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.484, + 0.901 + ], + "angle": 0, + "content": "The real world contains various dynamic objects that continuously interact and move through time. Understanding and reconstructing the shapes, poses, and articulations of dynamic objects is crucial for many applications such as robotics, augmented reality and virtual reality (AR/VR), etc. While humans naturally develop an intuitive understanding of how objects move and interact, teaching this skill to an intelligent agent remains a complex challenge due to the diverse and intricate nature of articulated objects. Furthermore, a key capability still missing in intelligent agents is effective interaction with their environment, particularly with functional elements like door handles and light switches. For intelligent agents to interact with functional objects, in addition to object detection, they must also" + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.292, + 0.916, + 0.423 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.436, + 0.908, + 0.506 + ], + "angle": 0, + "content": "Figure 1. Our IAAO requires multi-view images of the object or indoor scene from two different joint states (Left). The output is a 3D interactive field which supports interactions with multiple movable parts for fine-grained segmentation (e.g. Case 2: handles) and articulation reconstruction (e.g. Case 1: two articulated doors)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.536, + 0.907, + 0.643 + ], + "angle": 0, + "content": "understand the specific physical interactions each object affords—known as affordance [10]. The small size of these elements further intensifies the challenge, making them difficult to detect. In this work, we aim to reconstruct the 3D complex scene and interact with both objects with static geometry and objects with articulations and fine-grained affordances as illustrated in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Early approaches [21, 46] to understanding articulated objects rely on supervised learning with costly 3D annotations and category-specific models, which limit generalization to novel objects. Recent methods [16, 22] such as Ditto [16] and PARIS [22] have advanced this field by creating digital twins of objects based on observations in two joint configurations. Since Ditto is trained on multiview point clouds from specific categories, it struggles with objects that differ significantly from its training data. In contrast, PARIS [22] uses multi-view posed images to directly optimize an implicit representation without pretraining, allowing for improved generalization. However, it remains sensitive to initialization and may experience stability issues. Further developments [42, 47] have expanded scalability to objects with multiple moving parts based on implicit neural radiance fields (NeRFs). Nonetheless, several challenges remain unexplored, including generaliz" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12132" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.137 + ], + "angle": 0, + "content": "ing to complex scenes, flexibly applied for manipulation, and practical assumptions (e.g. handling unknown camera poses, unaligned static object parts, and occlusion)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.14, + 0.483, + 0.337 + ], + "angle": 0, + "content": "In 3D perception, instance segmentation provides essential object-level understanding for agents interacting within scenes. Recent studies [27, 29] explore part-object segmentation to decompose objects into finer components such as cabinet drawers. However, these methods are preliminary and struggle with affordance detection due to the lack of precise small-scale geometry in existing datasets. Open-vocabulary models [6, 18, 35, 43] utilize foundation models such as CLIP [39] to extract semantic features and consolidate them into a multi-view consistent 3D neural field representation. However, they rely on costly multi-scale queries to refine segmentation boundaries, and face challenges in high-dimensional feature encoding." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.339, + 0.483, + 0.626 + ], + "angle": 0, + "content": "To address the limitations mentioned above, we introduce IAAO, a framework designed for interaction of intelligent agents with objects, parts, and functional affordances, as well as for articulating movable objects. We explicitly leverage 3D semantic information to provide our model with scene understanding and interaction capabilities for articulated objects. Key insights of our approach include: 1) semantic scene reconstruction where SAM masks enable multi-view consistent segmentations from partial views, 2) hierarchical feature fields to efficiently compress dense features from multiple views into 3DGS for object- and part-level interaction in diverse scenes, 3) explicit 3D Gaussian representations for affordance localization via query similarity and motion recovery via 3D point-to-2D pixel correspondences, 4) global-local transformation initialization to enhance robustness in realistic scenarios with potential scene misalignment between object states, and 5) scene state fusion merging two different scene states to enhance object geometry completeness." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.629, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Specifically, our IAAO enables interactive affordance detection and object articulation via the following steps: 1) 3D model construction. Using multi-view posed images of each state, we model the 3D scene as a set of explicit 3D Gaussian primitives with 3D Gaussian Splatting (3DGS) [17]. 2) Hierarchical feature field construction. We propose an efficient distillation scheme where shape-aware 2D features and dense semantic features are extracted from large foundation models, including CLIP [39], SAM [19], and DINOv2 [33]. 2D features are efficiently distilled into 3D fields with a decoder. 3) Semantic-guided mask association across states. We first cluster SAM-generated masks from all views within each state to obtain view-consistent mask labels. We then compute 3D proposal features by merging the features from the corresponding 2D masks. By comparing pairwise feature similarities, we establish 3D mask-level correspondences across states that result in consistent mask label sets for the entire scene in each" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.212 + ], + "angle": 0, + "content": "state. 4) Affordance prediction. Given a task description, we encode it with the CLIP text encoder and compare the text embeddings with the feature fields to identify the relevant affordance. 5) Motion recovery. To estimate transformations, we define a set of consistency and matching losses that recover both global scene-level and local part-level transformations that ensure accurate motion representation across different states." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.213, + 0.872, + 0.227 + ], + "angle": 0, + "content": "We summarize our main contributions as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.228, + 0.905, + 0.319 + ], + "angle": 0, + "content": "- We introduce IAAO, an interactive affordance system utilizing 3D Gaussian fields embedded with hierarchical language-aligned semantics and class-agnostic masks. This system supports manipulation tasks guided by various prompts (point, mask, and language) at both object and part levels, regardless of object categories and shape." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.32, + 0.905, + 0.379 + ], + "angle": 0, + "content": "- We propose a method to reconstruct motion via global matching at the scene level and local matching for articulated objects, without relying on impractical assumptions about static object alignment or known camera poses." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.379, + 0.905, + 0.422 + ], + "angle": 0, + "content": "- Our IAAO achieves state-of-the-art performance across extensive experiments on multiple benchmarks, including synthetic, real-world, and indoor scene data." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.424, + 0.905, + 0.469 + ], + "angle": 0, + "content": "- We show strong model generalization to complex indoor environments and previously unseen articulated objects, with no restrictions on the number of movable parts." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.228, + 0.905, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.483, + 0.655, + 0.499 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Neural Fields for 3D Scene Understanding. 3D Scene understanding is primarily characterized by four classic representations, including volumetric fields ([34]), point clouds ([40]), 3D meshes ([7]), and depth maps ([11, 41, 54]). Unlike traditional representations, NeRF (Neural Radiance Fields) [28] introduces a neural implicit field to capture the geometry and appearance of a scene. Through a MultiLayer Perceptron (MLP) that takes 3D positions and 2D view directions as inputs, NeRF learns to implicitly represent the color and radiance of the scenes from a collection of posed images. However, a key drawback of NeRF is the slow training and rendering speed. Recently, 3D Gaussian Splatting (3DGS) [17] has been introduced as an alternative to implicit radiance field representations. In contrast to NeRF, 3DGS explicitly represents radiance fields as a collection of oriented 3D Gaussians. Each Gaussian is defined by its spatial position, opacity, and a covariance matrix, which allows flexible optimization. With efficient differentiable rasterization, 3DGS achieves fast rendering with high-quality results. Various work leverage neural fields as a representation robotic manipulation [18, 44, 45, 48, 58, 60]. Methods that leverage visual foundation models to construct neural feature fields [18, 37, 59] are most relevant to our work. With feature distillation via rendering, we reconstruct a hierarchical feature field for object localization and fine-grained affordance prediction." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12133" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.482, + 0.348 + ], + "angle": 0, + "content": "Visual Affordance. Affordance [9] refers to the potential interactions that an object or its parts facilitate for an agent. Affordance prediction involves deducing interaction opportunities from visual representations, e.g. images [4, 8, 25] or 3D models [3, 30, 31]. This field has gained significant interest in robotics as a foundational element for tasks e.g. grasping [26, 55], planning [50] and exploration [31]. However, although these methods excel at predicting affordance regions, they lack detailed interaction information, which is essential for engaging with functional elements. Recent work [2] addresses fine-grained functional elements using comprehensive natural language task descriptions for interaction. However, it depends on high-fidelity point clouds, restricting its practical applicability. We use high-resolution images with a Vision-Language Model (VLM) and 3D motion heuristics to predict affordances, eliminating the need for high-fidelity point clouds." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.35, + 0.483, + 0.593 + ], + "angle": 0, + "content": "Articulation Reasoning by Interaction. As embodied AI continues to evolve, the study of articulated objects has become a critical area of research. Recently, several synthetic [46, 49] and scanned data [15, 24, 27, 36] datasets have been introduced for articulated objects. These datasets contain annotations for part segmentation and motion parameters that enable data-driven approaches in the prediction of motion parameters from 3D point clouds [46, 53]. Recent works have been increasingly focused on real-world scenarios by detecting articulated parts and the motion parameters from images [15, 56] and videos [36]. Interactive perception involves agents gaining new insights from their interactions with the environment, and has been widely used to study object articulations [13, 14, 16, 32]. In this work, we focus on using language, and geometric and semantic scene understanding to achieve articulation reasoning." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.611, + 0.221, + 0.627 + ], + "angle": 0, + "content": "3. Our Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.639, + 0.483, + 0.821 + ], + "angle": 0, + "content": "Objective. Given a scene of articulated objects represented by a set of multi-view posed images \\(\\mathcal{I}^t = \\{I_i^t\\}_{i=1}^{N^t}\\) at two states \\(t\\) and \\(t'\\), the goal is to reconstruct the geometries and semantics of the objects in the scene. We focus on articulated objects which contain one or more movable parts with rotary or prismatic joints. Under the condition of unknown object categorory in the scene, we aim to recover the object part geometry, segmentation and joint articulations in the presence of occlusions, shape discrepancies and motion changes across different scene states. For static objects in the scene, the task degenerates to global transformation estimation between two input scene states for state matching." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Overview. Fig. 2 shows our IAAO framework consisting of: 1) A semantic scene reconstruction stage (cf. Sec. 3.1) where we first extract CLIP features for objects in individual scene state. These features are regularized by class-agnostic masks generated from 2D foundation models, e.g." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.091, + 0.905, + 0.349 + ], + "angle": 0, + "content": "SAM [57] at the instance and part level. We then cluster view-inconsistent masks from all views to obtain 3D consistent labels, and build a hierarchical feature field and label field for all masks in each scene state via 3D Gaussian Splatting (3DGS). 2) An affordance and motion prediction stage (cf. Sec. 3.2) where we perform object- and part-level queries by directly operating on 3D primitives represented by 3D Gaussians. Static elements and articulated objects are identified by isolating the groups of 3D Gaussians corresponding to segmented objects and functional parts of the scene, i.e. affordance prediction. For motion recovery, we estimate global transformation from static Gaussian primitives and local articulation parameters from the segmented movable parts of objects. 3) A scene state fusion stage (cf. Sec. 3.3) enabling affordance interaction and object manipulation by merging and refining the reconstructed scenes from two states according to the estimated transformations." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.363, + 0.791, + 0.377 + ], + "angle": 0, + "content": "3.1. Semantic Scene Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.386, + 0.907, + 0.543 + ], + "angle": 0, + "content": "Given a multi-view capture \\(\\mathcal{I}^t\\) of scene state \\(t\\), we construct a 3D model using 3DGS initialized with sparse point cloud generated from Structure from Motion (SfM) [40]. Each point serves as the center of a Gaussian primitive embedded with geometry and appearance parameters. These 3D Gaussian primitives \\(G^{t} = g_{p,p = 1}^{tP^{t}}\\) are then rendered into 2D views via differentiable rasterization for parameter optimization. To enrich the reconstructed model with semantic information, we augment each 3D Gaussian with semantic feature embeddings derived from large foundation models." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.548, + 0.907, + 0.794 + ], + "angle": 0, + "content": "View-Consistent Mask Clustering. For each input image \\( I_{i}^{t} \\in \\mathcal{I}^{t} \\), we first derive the class-agnostic masks \\( M_{i}^{t} = \\{m_{i,n}^{t}\\}_{n=1}^{n_{i}^{t}} \\) with an off-the-shelf mask predictor, where \\( n_{i}^{t} \\) represents the number of masks in view \\( i \\). For \\( M_{i}^{t} \\), we investigated the use of SAM [57] to generate both instance- and part-level masks for arbitrary objects. Unfortunately, mask generation models such as SAM face several challenges: a) class-agnostic masks do not have complete 3D information, b) inconsistent segmentation across images due to variations in viewpoint and appearance, and c) lack of one-to-one correspondence among masks due to oversegmentation for individual object. We aim to generate 3D-consistent labels for all masks in \\( M^{t} \\) by defining a mask label mapping function \\( \\phi(M^{t}, G^{t}) \\mapsto O^{t} \\). This function maps a set of masks \\( M_{o}^{t} \\) belonging to the same 3D object to a consistent \\( o^{t} \\in O^{t} \\) within the 3D Gaussian primitives \\( G^{t} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "To associate class-agnostic masks across all input views at scene state \\(t\\), we construct a mask graph to fuse the 2D masks \\(M^t\\) of the entire scene into cohesive 3D instances \\(O^t\\). In the initial graph \\(\\mathcal{G}_0^t = (\\mathcal{V}_0^t, \\mathcal{E}_0^t)\\), each node in \\(\\mathcal{V}_0^t\\) represents a mask detected from the views, and each edge in \\(\\mathcal{E}_0^t\\) represents potential matches between pairs of object parts. Following the mask clustering workflow from [52]," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12134" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.104, + 0.098, + 0.295, + 0.108 + ], + "angle": 0, + "content": "Stage 1: Semantic Scene Reconstruction" + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.111, + 0.737, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.755, + 0.102, + 0.886, + 0.111 + ], + "angle": 0, + "content": "Stage 3: Scene State Fusion" + }, + { + "type": "image", + "bbox": [ + 0.754, + 0.118, + 0.896, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.104, + 0.219, + 0.309, + 0.23 + ], + "angle": 0, + "content": "Stage 2: Affordance and Motion Prediction" + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.233, + 0.742, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.343, + 0.907, + 0.427 + ], + "angle": 0, + "content": "Figure 2. Our IAAO framework. 1) Top: Constructing 3D Gaussian fields in each state. We optimize 3DGS fields with hierarchical mask features, DINOv2 features and 3D-consistent mask labels generated from multi-view images. We also incorporate geometry information from depth images into the 3D Gaussians. 2) Bottom: Affordance and motion prediction. A query prompt is embedded using a pretrained encoder to localize relevant regions in the 3D Gaussians. For motion prediction, we optimize the transformation parameters by applying consistency and matching losses to 2D-3D correspondences between states. 3) Right: Scene fusion. Using the estimated transformations, we merge reconstructed 3DGS models from both states, aligning static and articulated elements." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.453, + 0.485, + 0.711 + ], + "angle": 0, + "content": "we then cluster the mask nodes and update the edges to facilitate mask association across different views. For each cluster, we combine the corresponding partial sets of Gaussians from individual masks to form a complete 3D instance and establish correspondences between 2D masks and 3D nodes. This results in the final graph \\(\\mathcal{G}^t = (\\mathcal{V}^t,\\mathcal{E}^t)\\), where each node in \\(\\mathcal{V}^t\\) represents a 3D instance proposal in \\(O^t\\) in scene state \\(t\\). We maintain a list of associated 2D masks \\(M_o^t\\) for each \\(o^t\\), discarding severely occluded masks. This 2D-3D relationship enables consistent segmentation labels across all views. However, the mask label mapping \\(\\phi (M^{t},G^{t})\\) may be affected by noise due to varying part segmentation hierarchies across views. To mitigate segmentation ambiguities, we use the labeled masks \\(M_o^t\\) in the final graph to supervise a label field for object part \\(o^t\\) in the given scene state. To this end, we use a cross-entropy classification loss \\(\\mathcal{L}_{\\mathrm{label}}\\) for enhanced consistency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Per-State Hierarchical Feature Field. We generate a set of feature embeddings from the given image \\( I_{i}^{t} \\) using pretrained foundation models. Specifically, we utilize MaskCLIP [5] to produce mask-level CLIP features at both instance and part levels, and DINOv2 [1, 33] to extract DINO embeddings. To enable object- and part-level interaction, we construct a hierarchical feature field by embedding dense features from multi-view images into the 3D Gaussian primitives. Using semantic hierarchy from SAM, we enhance our 3D semantic field accuracy while making the querying process more efficient. Directly embedding high-dimensional 2D features would result in excess" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.453, + 0.907, + 0.529 + ], + "angle": 0, + "content": "sive computational and memory costs. Therefore, we build a low-dimensional latent feature field and introduce a decoder to project the rendered features \\(\\hat{F}^t\\) back to the 2D feature space \\(F^t\\) during the differentiable rasterization optimization process as:" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.539, + 0.907, + 0.559 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {f e a t} = \\left\\| \\mathcal {D} \\left(\\hat {F} ^ {t} \\left(I _ {i} ^ {t}\\right)\\right) - F ^ {t} \\left(I _ {i} ^ {t}\\right) \\right\\| _ {2} ^ {2}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.906, + 0.612 + ], + "angle": 0, + "content": "The decoder \\(\\mathcal{D}\\) is implemented as a small MLP with three output branches: two for instance- and part-level CLIP features, and one for DINO features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.613, + 0.907, + 0.81 + ], + "angle": 0, + "content": "Semantic Association between Scene States. Upon obtaining the mask label mapping functions \\( O^t \\) and \\( O^{t'} \\) for scene states \\( t \\) and \\( t' \\), we establish a global 3D mask association which links the same semantic regions across both states into a unified 3D mask using their semantic features. For each 3D proposal in \\( O^t \\), we select the top- \\( k \\) representative masks from \\( M_o^t \\) and aggregate their rendered features to create the corresponding semantic primitives \\( F_o^t \\). We then build an affinity matrix capturing the cosine similarity between all 3D proposal pairs \\( (O^t, O^{t'}) \\) using their primitives \\( (F_O^t, F_O^{t'}) \\) from two scene states. For each query proposal \\( o^t \\in O^t \\), we define its visual match as the target proposal in \\( O^{t'} \\) with the highest similarity." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.819, + 0.816, + 0.833 + ], + "angle": 0, + "content": "3.2. Affordance and Motion Prediction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Unlike NeRF-based methods [6, 18, 23] which require resource-intensive rendering to derive language-embedded features and geometry from implicit MLPs, our IAAO directly uses explicit Gaussian primitives for efficient object" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12135" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "part localization and manipulation. Given the reconstructed 3D scene \\( G^{t} = \\{g_{p}^{t}\\}_{p = 1}^{P^{t}} \\) at scene state \\( t \\), where each Gaussian point \\( g_{p}^{t} \\) is embedded with geometric information (coordinates of points \\( p^t \\)) and additional features (color and semantic embeddings), our goal is to enable scene-aware affordance prediction and object articulation estimation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.186, + 0.483, + 0.426 + ], + "angle": 0, + "content": "Functional Affordance Prediction. Functional affordances are defined as parts of the scene that facilitate interactions for agents to perform specific tasks. We utilize the reconstructed hierarchical feature and label fields as the 3D representations to predict the masks \\(\\{m_a\\}_{a = 1}^{K_a}\\) and affordance labels \\(\\{l_a\\}_{a = 1}^{K_a}\\) in the scene, where \\(K_{a}\\) is the total number of affordance categories. We first build associations between affordance labels and our reconstructed mask label field, and then infer the point-level affordance of a functional element in the scene. Additionally, for task-specific reasoning, we encode the task description using the CLIP text encoder. To identify the relevant mask for a given description, we calculate the similarity between the embeddings in the feature field and the query embeddings. A mask is retrieved if its similarity score with the description text exceeds a predefined threshold." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.43, + 0.483, + 0.793 + ], + "angle": 0, + "content": "Global and Local Motion Recovery. Given the reconstructed scenes with different object articulations at two states, our objective is to estimate scene-aware transformation and 3D motion primitives for articulated parts. Our approach segments the reconstructed Gaussians \\( G^{t} \\) and \\( G^{t'} \\) into static scene elements and articulated objects via object-part querying for global and per-part local transformation estimation, respectively. 1) Object-level querying in complex scenes: In an indoor scene with a complex background and multiple objects, we start with object-level querying. Positive language queries are used to identify target objects for manipulation, while optional negative queries exclude irrelevant items. This selection process identifies 3D Gaussians with object-level CLIP features that align more closely with the positive queries than the negative ones, following standard CLIP feature-text comparison practices [45]. 2) Part-level querying for targeted objects: For individual objects in the scene, we apply part-level querying to focus on 3D Gaussians within the object. This allows for fine-grained transformation estimation specific to each articulated part. In addition to language, other user input prompts such as masks and points can also be used. Nonetheless, we do not further elaborate on these prompts as they can be easily obtained by a user clicking on the images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We define the global scene-aware transformation between two different scene states as \\(\\xi_g^t = (s_g^t,R_g^t,T_g^t)\\) where \\(s_g^t\\in \\mathbb{R}\\) represents the scale factor, \\(R_{g}^{t}\\in \\mathrm{SO}(3)\\) is the rotation matrix and \\(T_{g}^{t}\\in \\mathbb{R}^{3}\\) is the translation vector. To initialize the transformation, we estimate an initial alignment using 3D Gaussians derived from static elements in the scene, excluding articulated objects. From the Gaussian" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.228 + ], + "angle": 0, + "content": "points in both scene states, we select reliable points with opacity values above a threshold (empirically set at 0.7) to serve as input point sets for the coarse registration. Using the GeoTransformer [38], we calculate the scene alignment transformation by matching coarse superpoints and refining with dense point correspondences. Since these point clouds are based on reconstructed 3D Gaussians, the resulting transformation is coarse, providing an initial approximation for \\(\\xi_{g}^{t}\\) that can be further refined." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.228, + 0.907, + 0.397 + ], + "angle": 0, + "content": "After achieving global alignment between two scene states, we define the local transformation for each articulated part as \\(\\xi_{o}^{t} = (R_{o}^{t},T_{o}^{t})\\), representing the movement of part \\(o\\) from state \\(t\\) to state \\(t'\\). All points within a part are assumed to follow the same motion parameters. The reverse mapping, which transforms part \\(o\\) from state \\(t'\\) back to state \\(t\\), is given by the inverse transformation, \\(\\xi_{o}^{t'} = (R_{o}^{t},T_{o}^{t})^{-1}\\). To facilitate this process, we construct a 3D Gaussian correspondence field, allowing each Gaussian primitive \\(g_{p}^{t}(o)\\) at state \\(t\\) to be repositioned to its corresponding location \\(p^{t'}\\) at state \\(t'\\) based on the motion parameters of part \\(o^t\\). Formally:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.404, + 0.905, + 0.453 + ], + "angle": 0, + "content": "\\[\np ^ {t \\rightarrow t ^ {\\prime}} = \\left\\{\\begin{array}{l l}f \\left(p ^ {t}, \\xi_ {g} ^ {t}\\right) = R _ {g} ^ {t} p ^ {t} + T _ {g} ^ {t},&\\text {i f} p \\text {i n s t a t i c s c e n e ,}\\\\f \\left(f \\left(p ^ {t}, \\xi_ {g} ^ {t}\\right), \\xi_ {o} ^ {t}\\right),&\\text {o t h e r w i s e .}\\end{array}\\right. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.462, + 0.905, + 0.508 + ], + "angle": 0, + "content": "The same rule applies to the backward transformation function \\( f^{-1} \\) for obtaining \\( p^{t' \\to t} \\) (i.e. mapping from \\( t' \\) back to \\( t \\)) with the transformations \\( \\xi_g^{t'} \\) and \\( \\xi_o^{t'} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.907, + 0.766 + ], + "angle": 0, + "content": "Establishing accurate dense correspondences between point clouds in two states is challenging due to the sparsity and noise in 3D points derived from the 3DGS model. To optimize the correspondence field, we instead match 3D Gaussian primitives to 2D pixels across the two states, guided by part geometry and segmentation information. Specifically, for a Gaussian primitive \\( g_{p}^{t}(o) \\) of part \\( o \\) at state \\( t \\), the corresponding 2D pixels at state \\( t' \\) can be treated as observations from novel views. To enhance consistency, we incorporate mask-level features alongside the rendered image consistency loss \\( \\mathcal{L}_{rgb} \\). For a 3D proposal \\( o^t \\) in graph \\( \\mathcal{G}^t \\), we identify the part mask list \\( \\mathcal{M}_o^{t'} \\) associated with the corresponding proposal \\( o^{t'} \\) at state \\( t' \\). Following the splatting process, we render the feature field of \\( o^t \\) to training view \\( n \\) in target scene state \\( t' \\) and decode the mask-level features by \\( \\mathcal{D}(\\hat{F}_o^t (I_n^{t'})) \\). We formulate the mask feature loss for part \\( o \\) transformed from state \\( t \\) to target state \\( t' \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.544, + 0.773, + 0.905, + 0.832 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {m a s k} (o ^ {t \\rightarrow t ^ {\\prime}}) = \\\\ \\Sigma_ {n} ^ {N _ {o} ^ {t ^ {\\prime}}} \\frac {M _ {o} ^ {t ^ {\\prime}} \\left(I _ {n} ^ {t ^ {\\prime}}\\right)}{\\left| M _ {o} ^ {t ^ {\\prime}} \\left(I _ {n} ^ {t ^ {\\prime}}\\right) \\right|} \\cdot \\| \\mathcal {D} \\left(\\hat {F} _ {o} ^ {t} \\left(I _ {n} ^ {t ^ {\\prime}}\\right)\\right) - F _ {o} ^ {t ^ {\\prime}} \\left(I _ {n} ^ {t ^ {\\prime}}\\right) \\| _ {2} ^ {2}, \\tag {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.839, + 0.905, + 0.901 + ], + "angle": 0, + "content": "where \\( M_{o}^{t^{\\prime}}(I_{n}^{t^{\\prime}}) \\) and \\( F_{o}^{t^{\\prime}}(I_{n}^{t^{\\prime}}) \\) represent the corresponding part mask and 2D mask-level features of part \\( o \\) in view \\( I_{n}^{t^{\\prime}} \\) at the target state \\( t^{\\prime} \\). \\( N_{o}^{t} \\) denotes the number of part masks for proposal \\( o^t \\) across all training views." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12136" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.332 + ], + "angle": 0, + "content": "Since the mask-level loss provides only coarse guidance for transformation, we further propose to find dense correspondences between the 3D Gaussian primitives and 2D pixels by comparing their DINO features. We begin by computing the feature similarity matrix \\(\\alpha_{p\\rightarrow o}(I_n^{t'})\\) between each pixel in the part mask \\(M_o^{t'}(I_n^{t'})\\) in view \\(I_n^{t'}\\) and the sampled Gaussian point \\(g_p^t\\) from 3D proposal \\(o^t\\). We then normalize the similarity matrix \\(\\alpha_{p\\rightarrow o}(I_n^{t'})\\) using a softmax over the entire mask to obtain the weight \\(\\beta_{p\\rightarrow o}(I_n^{t'})\\). Finally, we identify the corresponding 2D pixel \\(s_{p\\rightarrow o}^{t'}(I_n^{t'})\\) for the sampled 3D point \\(p^t\\) using a weighted sum. See the supplementary material for details on the computation. For each point-pixel pair \\((p^t,s_{p\\rightarrow o}^{t'}(I_n^{t'}))\\), we filter out pairs that collide by checking whether \\(s_{p\\rightarrow o}^{t'}(I_n^{t'})\\) falls within the same part mask \\(o^{t'}\\). We then define the matching loss as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.343, + 0.483, + 0.363 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {m a t c h}} = \\left\\| \\pi_ {n} ^ {t ^ {\\prime}} \\left(p ^ {t \\rightarrow t ^ {\\prime}}\\right) - s _ {p \\rightarrow n} ^ {t ^ {\\prime}} \\left(I _ {n} ^ {t ^ {\\prime}}\\right)\\right\\| _ {2} ^ {2}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.374, + 0.483, + 0.406 + ], + "angle": 0, + "content": "where \\(\\pi_n^{t'}\\) represents the projection of 3D point \\(p^{t\\rightarrow t'}\\) to view \\(I_{n}^{t^{\\prime}}\\) in scene state \\(t^\\prime\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.407, + 0.484, + 0.482 + ], + "angle": 0, + "content": "**Optimization.** For the first stage, each scene is reconstructed independently at two different states. We optionally use depth map for further geometry regularization. In the second stage, we reconstruct motion of global scene and local articulated objects. The total loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.494, + 0.483, + 0.512 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda_ {\\text {c o n s}} \\left(\\mathcal {L} _ {\\text {r g b}} + \\mathcal {L} _ {\\text {m a s k}} + \\mathcal {L} _ {\\text {l a b e l}}\\right) + \\lambda_ {\\text {m a t c h}} \\mathcal {L} _ {\\text {m a t c h}}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.522, + 0.484, + 0.553 + ], + "angle": 0, + "content": "\\(\\lambda_{\\text{cons}}\\) and \\(\\lambda_{\\text{match}}\\) denote the balancing weights of consistency loss and matching loss." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.562, + 0.273, + 0.576 + ], + "angle": 0, + "content": "3.3. Scene State Fusion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.484, + 0.718 + ], + "angle": 0, + "content": "After training, we merge the 3DGS models from the two states using the estimated transformations to fill in the occluded regions in each state. Details of our fusion and filtering strategy are provided in the supplementary material. The combined 3D Gaussians facilitate object manipulation by addressing potential artifacts that may appear in occluded areas of the reconstructed scene. Optional fine-tuning can be applied to either the edited regions or the entire scene for further refinement." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.734, + 0.216, + 0.75 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.758, + 0.282, + 0.774 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Datasets. 1) PARIS two-part object dataset. This dataset contains two-part articulated objects, featuring 10 synthetic objects from PartNet-Mobility [49] and 2 real-world objects captured using MultiScan [27]. Each object comprises a movable part and a static part, observed in two distinct joint states. The dataset includes RGB images, object masks from 100 random viewpoints, and depth images for both synthetic and real-world objects. 2) Synthetic multi-part" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.258 + ], + "angle": 0, + "content": "object dataset. This dataset includes two synthetic scenes containing multi-part objects from PartNet-Mobility, with each object featuring one static part and multiple movable parts. Each object is observed in two articulation states, and the dataset provides RGB, depth, and mask data from 100 random viewpoints. 3) Indoor scene OmniSim dataset. OmniSim is generated using the OmniGibson [20] simulator with various indoor scene models. By adjusting the rotation of articulated object joints, we generate some scenes for evaluation, each containing RGBD images, interactive object masks and object state metrics at each state." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.263, + 0.906, + 0.445 + ], + "angle": 0, + "content": "Metrics. To evaluate articulation models, we use: 1) Axis Ang Err \\((^{\\circ})\\) means angular error of the predicted joint axis for both revolute and prismatic joints. 2) Axis Pos Err (0.1 m): The shortest distance between the predicted and true joint axes for revolute joints. For cross-state part motion, we use Part Motion Err \\((^{\\circ}\\) or m). It evaluates joint state precision by calculating the geodesic error for predicted rotations in revolute joints or the Euclidean error for translations in prismatic joints, as specified in PARIS [22]. For object and part mesh reconstruction, we use Chamfer-L1 distance (CD), reporting CD-w for the full surface, CD-s for static parts, and CD-m for movable parts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.447, + 0.906, + 0.688 + ], + "angle": 0, + "content": "Baselines. Ditto [16] and PARIS [22] are models for reconstructing part-level shape and motion in two-part articulated objects, using multi-view data. Ditto assumes one moving part and is pre-trained on specific categories, while PARIS uses a NeRF-based representation to handle unknown objects. \\(PARIS^{*}\\) is an enhanced version with depth supervision, and \\(PARIS - m^{*}\\) extends PARIS to multi-part objects. CSG-reg uses TSDF fusion and Constructive Solid Geometry for part segmentation, with registration for alignment. 3Dseg-reg follows a similar process but employs PA-Conv [51] for part segmentation, reporting results only for trained categories due to limited generalization. DigitalTwinArt [47] divides reconstruction into two stages: 1) It reconstructs object-level shape independently of articulation; 2) It recovers the articulation model by identifying part segmentation and motions through state correspondences." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Evaluation Setup. There are some key differences on evaluation setup between our model and previous models as we firstly incorporate the semantics into the explicit neural field supporting robust interactions even in complex environment. Unlike prior methods which assume the number of parts is known, our model can detect any articulated objects if enough information for reconstructing object shape is given from multi-view images. To identify corresponding parts for evaluation, prior work need to iterate through all possible pairs of predicted and ground-truth parts, selecting the match with the smallest total Chamfer distance. In our model, we identify the corresponding parts based on the constructed semantic feature and mask label field. We extract the mesh based on the 3D point clouds derived" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12137" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.088, + 0.905, + 0.555 + ], + "angle": 0, + "content": "
SimulationReal
FoldChairFridgeLaptopOvenScissorStaplerUSBWasherBladeStorageAllFridgeStorageAll
Axis Ang ↓Ditto [16]89.3589.783.120.964.5089.8689.7789.5179.54*6.3254.221.715.883.80
PARIS [22]8.089.150.020.043.8239.730.1325.3615.380.0310.171.6443.1322.39
PARIS* [22]15.792.930.037.4316.628.170.7118.4041.280.0311.141.9030.1016.00
CSG-reg0.100.270.470.350.280.3011.7871.937.642.829.608.9269.7139.31
3Dseg-reg--2.34-----9.40-----
DigitalTwinArt [47]0.030.070.060.220.110.060.110.430.270.060.142.1018.1110.11
IAAO (Ours)0.020.040.050.170.080.050.090.360.190.060.111.915.828.86
Axis Pos ↓Ditto [16]3.771.02*0.010.135.700.205.410.66--2.111.84-0.92
PARIS [22]0.450.380.000.002.102.272.361.50--1.130.34-0.17
PARIS* [22]0.251.130.000.051.594.673.353.28--1.790.50-0.25
CSG-reg0.020.000.200.180.010.020.012.13--0.321.46-0.73
3Dseg-reg--0.10-----------
DigitalTwinArt [47]0.010.010.000.010.020.010.000.01--0.010.57-0.29
IAAO (Ours)0.010.000.000.000.010.010.010.00--0.010.32-0.16
Part Motion ↓Ditto [16]99.36×5.182.0919.2856.6180.6055.72×0.0939.878.430.384.41
PARIS [22]131.6624.580.030.03120.70110.8064.8560.350.340.3051.362.160.561.36
PARIS* [22]127.3445.260.039.1368.36107.7696.9349.770.360.3050.521.580.571.07
CSG-reg0.130.290.350.580.200.4410.48158.990.050.0417.1614.820.647.73
3Dseg-reg--1.61-----0.15-----
DigitalTwinArt [47]0.160.090.080.110.150.050.110.250.000.000.101.860.201.03
IAAO (Ours)0.130.070.030.040.110.050.090.210.000.000.071.460.150.81
CD-s ↓Ditto [14]33.793.050.252.5239.0741.642.6410.3246.909.1818.9447.0116.0931.55
PARIS [22]9.163.651.941.8812.951.882.6925.391.1912.767.1842.5754.5448.56
PARIS* [22]10.208.8215.581.953.182.482.4812.191.408.676.4611.6420.2515.94
CSG-reg1.691.453.933.263.262.221.954.530.597.062.706.3312.559.44
3Dseg-reg--0.76-----66.31-----
DigitalTwinArt [47]0.180.604.660.402.652.194.804.690.554.692.102.5310.866.69
IAAO (Ours)0.160.510.170.361.981.864.433.700.514.271.802.479.836.15
CD-m ↓Ditto [14]141.110.990.190.9420.6831.2115.8812.89195.932.2042.2050.6020.3535.48
PARIS [22]8.997.760.2128.7046.6419.275.32178.4325.2176.6939.7245.66864.82455.24
PARIS* [22]17.977.230.156.5416.6530.4610.17265.27117.9952.3452.4877.85474.57276.21
CSG-reg1.9121.710.42256.991.956.3629.78436.4226.621.3978.36442.17521.49481.83
3Dseg-reg--1.01-----6.23-----
DigitalTwinArt [47]0.150.270.160.470.412.271.340.361.500.370.731.1426.4613.80
IAAO (Ours)0.130.180.150.410.402.091.180.331.420.250.651.0221.4311.23
CD-w ↓Ditto [14]6.802.160.312.511.702.382.097.2942.043.917.126.5014.0810.29
PARIS [22]1.802.920.3011.7310.493.582.0024.380.608.576.6422.9863.3543.16
PARIS* [22]4.375.530.263.183.905.271.7810.110.587.804.288.9932.1020.55
CSG-reg0.480.980.403.001.701.991.204.480.564.001.885.7114.2910.00
3Dseg-reg--0.81-----0.78-----
DigitalTwinArt [47]0.270.700.354.180.432.191.184.740.363.991.842.199.335.76
IAAO (Ours)0.220.610.273.490.341.961.064.410.343.941.662.029.185.60
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.565, + 0.907, + 0.594 + ], + "angle": 0, + "content": "Table 1. Results on PARIS dataset including both synthetic and real data. 'x': failure case, and \\*\\*: joint axis or position. -': no result is available. Best result in bold." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.607, + 0.905, + 0.695 + ], + "angle": 0, + "content": "
Axis Ang 0↓Axis Ang 1 ↓Axis Pos 0 ↓Axis Pos 1 ↓Part Motion 0 ↓Part Motion 1 ↓CD-s ↓CD-m 0 ↓CD-m 1 ↓CD-w ↓
Fridge-mPARIS*-m [22]34.5215.913.601.6386.21105.868.52526.19160.8615.00
DigitalTwinArt [47]0.160.100.010.000.110.130.610.400.520.89
IAAO (Ours)0.160.080.010.000.090.090.540.440.470.96
Storage-mPARIS*-m [22]43.2626.1810.42-79.840.648.56128.62266.718.66
DigitalTwinArt [47]0.210.880.05-0.130.000.850.213.460.99
IAAO (Ours)0.180.910.04-0.100.000.780.263.350.97
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.703, + 0.908, + 0.731 + ], + "angle": 0, + "content": "Table 2. Results on multi-part object dataset, averaged over 10 trials with different random seeds. Joint 1 of \"Storage-m\" is solely prismatic with no Axis Pos. '-': no result is available. Best results in bold." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.758, + 0.482, + 0.803 + ], + "angle": 0, + "content": "from 3DGS model via SuGaR [12] for evaluation. Following [22, 47], we transform our extracted parts with predicted motions to start state \\( t \\) for evaluation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.817, + 0.285, + 0.833 + ], + "angle": 0, + "content": "4.2. Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Tab. 1 shows quantitative results from the PARIS Two-Part Object Dataset, covering synthetic and real instances. Ditto performs well on seen categories but struggles with unseen ones. PARIS occasionally fails in shape and articu" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.758, + 0.907, + 0.849 + ], + "angle": 0, + "content": "lation reconstruction, while depth supervision in PARIS* improves shape accuracy for complex and real-world objects but increases articulation errors due to optimization challenges. CSG-reg and 3Dseg-reg handle simple objects well but struggle with complex ones, as segmentation errors affect articulation estimation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Tab. 2 presents the results for multi-part objects. DigitalTwinArt [47] outperforms PARIS* in shape and articulation reconstruction by incorporating additional supervi" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12138" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.159, + 0.092, + 0.237, + 0.107 + ], + "angle": 0, + "content": "Input State \\(t\\)" + }, + { + "type": "image", + "bbox": [ + 0.168, + 0.121, + 0.23, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.166, + 0.195, + 0.245, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.361, + 0.093, + 0.442, + 0.106 + ], + "angle": 0, + "content": "Input State \\(t^{\\prime}\\)" + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.122, + 0.432, + 0.185 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.353, + 0.194, + 0.436, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.557, + 0.093, + 0.643, + 0.107 + ], + "angle": 0, + "content": "DigitalTwinArt" + }, + { + "type": "image", + "bbox": [ + 0.578, + 0.116, + 0.621, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.562, + 0.196, + 0.648, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.777, + 0.094, + 0.81, + 0.105 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.77, + 0.115, + 0.822, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.761, + 0.196, + 0.846, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.116, + 0.267, + 0.879, + 0.281 + ], + "angle": 0, + "content": "Figure 3. Qualitative analysis of shape reconstruction, part segmentation, and joint prediction results on multi-part object dataset." + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.308, + 0.485, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.472, + 0.483, + 0.499 + ], + "angle": 0, + "content": "Figure 4. Qualitative results of shape reconstruction, part segmentation, and joint prediction on PARIS." + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.506, + 0.477, + 0.576 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.112, + 0.58, + 0.462, + 0.594 + ], + "angle": 0, + "content": "Figure 5. Motion snapshots on PARIS & multi-part object." + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.6, + 0.475, + 0.669 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.105, + 0.672, + 0.468, + 0.687 + ], + "angle": 0, + "content": "Figure 6. Motion snapshots on scene-level OmniSim dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.703, + 0.484, + 0.749 + ], + "angle": 0, + "content": "sion on 3D geometries. Our IAAO achieves competitive results by leveraging geometric and semantic information from VLMs, enhancing object part modeling." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.758, + 0.274, + 0.773 + ], + "angle": 0, + "content": "4.3. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.781, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Fig. 3 and Fig. 4 show qualitative results on shape reconstruction, part segmentation, and joint prediction for the multi-part object and PARIS two-part datasets. Our IAAO achieves more complete segmentation results while also enabling interaction with affordance and segmentation features. Fig. 5 and 6 show snapshots of the generated motions on several examples at object and scene levels in a simulator. It shows IAAO can produce smooth interpolations of" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.305, + 0.903, + 0.372 + ], + "angle": 0, + "content": "
MethodAxis Ang ↓Axis Pos ↓Part Motion ↓CD-s ↓CD-m ↓
Proposed0.280.010.110.800.35
w/o matching4.591.388.571.21123.71
w/o mask14.820.010.282.59129.14
w/o rgb0.560.010.170.810.42
w/o label0.540.010.158.56169.73
" + }, + { + "type": "table_caption", + "bbox": [ + 0.553, + 0.382, + 0.866, + 0.396 + ], + "angle": 0, + "content": "Table 3. Ablation study on multi-part object dataset." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.41, + 0.871, + 0.426 + ], + "angle": 0, + "content": "articulated objects with good part geometry modeling." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.433, + 0.667, + 0.45 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.456, + 0.907, + 0.622 + ], + "angle": 0, + "content": "We assess the impact of our design choices specifically on multi-part objects, which present greater challenges. We first perform experiments to test out the impact of 2D-3D matching loss ('w/o matching'). As indicated in Tab. 3, the matching loss plays a crucial role in enhancing articulation reconstruction. We also study constraints from mask features and labels, which act as a coarse correspondence mechanism at the superpoint (mask) level. These constraints also enforce consistency within the feature fields of each mask, leading to improved part segmentation and more accurate reconstruction of articulated components." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.634, + 0.634, + 0.649 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We present IAAO, a framework that facilitates interactive affordance detection and object articulation from two scene sequences. Each scene sequence captures movable parts of objects in different joint states. Our IAAO can handle both simple scenes with single objects and complex indoor scenes without limitations on the number of objects and movable parts. The entire scene is reconstructed using 3D Gaussians, embedded with robust zero-shot generalization capabilities from 2D foundation models. Affordance prediction is achieved through object and part queries on the 3D Gaussian primitives. Motion estimation is achieved by establishing 2D-3D correspondences within each object to track transformations. Static elements serve for global scene alignment while movable elements are for local transformation. Extensive experiments demonstrate that our approach outperforms existing methods." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12139" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.152 + ], + "angle": 0, + "content": "Acknowledgement. This research / project is supported by the National Research Foundation (NRF) Singapore, under its NRF-Investigatorship Programme (Award ID. NRF-NRFI09-0008)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.164, + 0.188, + 0.18 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.189, + 0.484, + 0.218 + ], + "angle": 0, + "content": "[1] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.218, + 0.484, + 0.3 + ], + "angle": 0, + "content": "[2] Alexandros Delitzas, Ayca Takmaz, Federico Tombari, Robert Sumner, Marc Pollefeys, and Francis Engelmann. Scenefun3d: Fine-grained functionality and affordance understanding in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14531-14542, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.301, + 0.484, + 0.37 + ], + "angle": 0, + "content": "[3] Shengheng Deng, Xun Xu, Chaozheng Wu, Ke Chen, and Kui Jia. 3d affordancenet: A benchmark for visual object affordance understanding. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1778-1787, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.371, + 0.484, + 0.426 + ], + "angle": 0, + "content": "[4] Thanh-Toan Do, Anh Nguyen, and Ian Reid. Affordancenet: An end-to-end deep learning approach for object affordance detection. In 2018 IEEE international conference on robotics and automation (ICRA), pages 5882-5889. IEEE, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.427, + 0.484, + 0.522 + ], + "angle": 0, + "content": "[5] Xiaoyi Dong, Jianmin Bao, Yinglin Zheng, Ting Zhang, Dongdong Chen, Hao Yang, Ming Zeng, Weiming Zhang, Lu Yuan, Dong Chen, et al. Maskclip: Masked self-distillation advances contrastive language-image pretraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10995-11005, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.524, + 0.484, + 0.592 + ], + "angle": 0, + "content": "[6] Francis Engelmann, Fabian Manhardt, Michael Niemeyer, Keisuke Tateno, Marc Pollefeys, and Federico Tombari. Opennerf: Open set 3d neural scene segmentation with pixelwise features and rendered novel views. arXiv preprint arXiv:2404.03650, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.594, + 0.484, + 0.636 + ], + "angle": 0, + "content": "[7] Carlos Hernández Esteban and Francis Schmitt. Silhouette and stereo fusion for 3d object modeling. Computer Vision and Image Understanding, 96(3):367-392, 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.636, + 0.484, + 0.704 + ], + "angle": 0, + "content": "[8] Kuan Fang, Te-Lin Wu, Daniel Yang, Silvio Savarese, and Joseph J Lim. Demo2vec: Reasoning object affordances from online videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2139-2147, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.705, + 0.484, + 0.747 + ], + "angle": 0, + "content": "[9] JJ Gibson. The theory of affordances. Perceiving, acting and knowing: Towards an ecological psychology/Erlbaum, 1977. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.748, + 0.483, + 0.775 + ], + "angle": 0, + "content": "[10] James J Gibson. The ecological approach to visual perception: classic edition. Psychology press, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.776, + 0.483, + 0.831 + ], + "angle": 0, + "content": "[11] Michael Goesele, Brian Curless, and Steven M Seitz. Multiview stereo revisited. In 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'06), pages 2402-2409. IEEE, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[12] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5354-5363, 2024. 7" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.189, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.161 + ], + "angle": 0, + "content": "[13] Karol Hausman, Scott Niekum, Sarah Osentoski, and Gaurav S Sukhatme. Active articulation model estimation through interactive perception. In 2015 IEEE International Conference on Robotics and Automation (ICRA), pages 3305-3312. IEEE, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.162, + 0.906, + 0.232 + ], + "angle": 0, + "content": "[14] Cheng-Chun Hsu, Zhenyu Jiang, and Yuke Zhu. Ditto in the house: Building articulation models of indoor scenes through interactive perception. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 3933-3939. IEEE, 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.233, + 0.906, + 0.288 + ], + "angle": 0, + "content": "[15] Hanxiao Jiang, Yongsen Mao, Manolis Savva, and Angel X Chang. Opd: Single-view 3d openable part detection. In European Conference on Computer Vision, pages 410-426. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.289, + 0.906, + 0.356 + ], + "angle": 0, + "content": "[16] Zhenyu Jiang, Cheng-Chun Hsu, and Yuke Zhu. Ditto: Building digital twins of articulated objects from interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5616-5626, 2022. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.358, + 0.906, + 0.412 + ], + "angle": 0, + "content": "[17] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.413, + 0.906, + 0.482 + ], + "angle": 0, + "content": "[18] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Lerf: Language embedded radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19729-19739, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.906, + 0.551 + ], + "angle": 0, + "content": "[19] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.553, + 0.906, + 0.636 + ], + "angle": 0, + "content": "[20] Chengshu Li, Ruohan Zhang, Josiah Wong, Cem Gokmen, Sanjana Srivastava, Roberto Martin-Martín, Chen Wang, Gabrael Levine, Wensi Ai, Benjamin Martinez, et al. Behavior-1k: A human-centered, embodied ai benchmark with 1,000 everyday activities and realistic simulation. arXiv preprint arXiv:2403.09227, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.636, + 0.906, + 0.704 + ], + "angle": 0, + "content": "[21] Xiaolong Li, He Wang, Li Yi, Leonidas J Guibas, A Lynn Abbott, and Shuran Song. Category-level articulated object pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3706-3715, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.705, + 0.906, + 0.762 + ], + "angle": 0, + "content": "[22] Jiayi Liu, Ali Mahdavi-Amiri, and Manolis Savva. Paris: Part-level reconstruction and motion analysis for articulated objects. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 352-363, 2023. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.762, + 0.906, + 0.831 + ], + "angle": 0, + "content": "[23] Kunhao Liu, Fangneng Zhan, Jiahui Zhang, Muyu Xu, Yingchen Yu, Abdulmotaleb El Saddik, Christian Theobalt, Eric Xing, and Shijian Lu. Weakly supervised 3d open-vocabulary segmentation. Advances in Neural Information Processing Systems, 36:53433-53456, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[24] Liu Liu, Wenqiang Xu, Haoyuan Fu, Sucheng Qian, Qiao-jun Yu, Yang Han, and Cewu Lu. Akb-48: A real-world articulated object knowledge base. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14809–14818, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12140" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.146 + ], + "angle": 0, + "content": "[25] Timo Luddecke and Florentin Worgotter. Learning to segment affordances. In Proceedings of the IEEE International Conference on Computer Vision Workshops, pages 769-776, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.15, + 0.482, + 0.205 + ], + "angle": 0, + "content": "[26] Priyanka Mandikal and Kristen Grauman. Learning dexterous grasping with object-centric visual affordances. In 2021 IEEE international conference on robotics and automation (ICRA), pages 6169-6176. IEEE, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.208, + 0.482, + 0.274 + ], + "angle": 0, + "content": "[27] Yongsen Mao, Yiming Zhang, Hanxiao Jiang, Angel Chang, and Manolis Savva. Multiscan: Scalable rgbd scanning for 3d environments with articulated objects. Advances in neural information processing systems, 35:9058-9071, 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.345 + ], + "angle": 0, + "content": "[28] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.482, + 0.431 + ], + "angle": 0, + "content": "[29] Kaichun Mo, Shilin Zhu, Angel X Chang, Li Yi, Subarna Tripathi, Leonidas J Guibas, and Hao Su. Partnet: A largescale benchmark for fine-grained and hierarchical part-level 3d object understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 909-918, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.482, + 0.489 + ], + "angle": 0, + "content": "[30] Kaichun Mo, Yuzhe Qin, Fanbo Xiang, Hao Su, and Leonidas Guibas. O2o-afford: Annotation-free large-scale object-object affordance learning. In Conference on robot learning, pages 1666-1677. PMLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.482, + 0.545 + ], + "angle": 0, + "content": "[31] Tushar Nagarajan and Kristen Grauman. Learning affordance landscapes for interaction exploration in 3d environments. Advances in Neural Information Processing Systems, 33:2005-2015, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.482, + 0.602 + ], + "angle": 0, + "content": "[32] Neil Nie, Samir Yitzhak Gadre, Kiana Ehsani, and Shuran Song. Structure from action: Learning interactions for articulated object 3d structure discovery. arXiv preprint arXiv:2207.08997, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.606, + 0.482, + 0.729 + ], + "angle": 0, + "content": "[33] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.482, + 0.785 + ], + "angle": 0, + "content": "[34] Sylvain Paris, François X Sillion, and Long Quan. A surface reconstruction method using global graph cut optimization. International Journal of Computer Vision, 66:141-161, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[35] Songyou Peng, Kyle Genova, Chiyu Jiang, Andrea Tagliasacchi, Marc Pollefeys, Thomas Funkhouser, et al. Openscene: 3d scene understanding with open vocabularies. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 815-824, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[36] Shengyi Qian, Linyi Jin, Chris Rockwell, Siyi Chen, and David F Fouhey. Understanding 3d object articulation in internet videos. In Proceedings of the IEEE/CVF Conference" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "on Computer Vision and Pattern Recognition, pages 1599-1609, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[37] Minghan Qin, Wanhua Li, Jiawei Zhou, Haoqian Wang, and Hanspeter Pfister. Langsplat: 3d language gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20051-20060, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.194, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[38] Zheng Qin, Hao Yu, Changjian Wang, Yulan Guo, Yuxing Peng, and Kai Xu. Geometric transformer for fast and robust point cloud registration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11143-11152, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.265, + 0.905, + 0.346 + ], + "angle": 0, + "content": "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.349, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[40] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.407, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[41] Christoph Strecha, Rik Fransens, and Luc Van Gool. Combined depth and outlier estimation in multi-view stereo. In 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'06), pages 2394-2401. IEEE, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.478, + 0.905, + 0.546 + ], + "angle": 0, + "content": "[42] Archana Swaminathan, Anubhav Gupta, Kamal Gupta, Shishira R Maiya, Vatsal Agarwal, and Abhinav Shrivastava. Leia: Latent view-invariant embeddings for implicit 3d articulation. In European Conference on Computer Vision, pages 210-227. Springer, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.549, + 0.905, + 0.615 + ], + "angle": 0, + "content": "[43] Ayca Takmaz, Elisabetta Fedele, Robert Sumner, Marc Pollefeys, Federico Tombari, and Francis Engelmann. Openmask3d: Open-vocabulary 3d instance segmentation. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.62, + 0.905, + 0.715 + ], + "angle": 0, + "content": "[44] Zhenggang Tang, Balakumar Sundaralingam, Jonathan Tremblay, Bowen Wen, Ye Yuan, Stephen Tyree, Charles Loop, Alexander Schwing, and Stan Birchfield. RGB-only reconstruction of tabletop scenes for collision-free manipulator control. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 1778-1785. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.718, + 0.905, + 0.786 + ], + "angle": 0, + "content": "[45] Can Wang, Mengei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3835-3844, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.789, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[46] Xiaogang Wang, Bin Zhou, Yahao Shi, Xiaowu Chen, Qinping Zhao, and Kai Xu. Shape2motion: Joint analysis of motion parts and attributes from 3d shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8876-8884, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[47] Yijia Weng, Bowen Wen, Jonathan Tremblay, Valts Blukis, Dieter Fox, Leonidas Guibas, and Stan Birchfield. Neural implicit representation for building digital twins of unknown" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12141" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.123, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "articulated objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3141-3150, 2024. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.191 + ], + "angle": 0, + "content": "[48] Youngsun Wi, Pete Florence, Andy Zeng, and Nima Fazeli. Virdo: Visio-tactile implicit representations of deformable objects. In 2022 International Conference on Robotics and Automation (ICRA), pages 3583-3590. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.482, + 0.275 + ], + "angle": 0, + "content": "[49] Fanbo Xiang, Yuzhe Qin, Kaichun Mo, Yikuan Xia, Hao Zhu, Fangchen Liu, Minghua Liu, Hanxiao Jiang, Yifu Yuan, He Wang, et al. Sapien: A simulated part-based interactive environment. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11097-11107, 2020. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.348 + ], + "angle": 0, + "content": "[50] Danfei Xu, Ajay Mandlekar, Roberto Martin-Martin, Yuke Zhu, Silvio Savarese, and Li Fei-Fei. Deep affordance foresight: Planning through what can be done in the future. In 2021 IEEE international conference on robotics and automation (ICRA), pages 6206–6213. IEEE, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[51] Mutian Xu, Runyu Ding, Hengshuang Zhao, and Xiaojuan Qi. Paconv: Position adaptive convolution with dynamic kernel assembling on point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3173-3182, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.482, + 0.489 + ], + "angle": 0, + "content": "[52] Mi Yan, Jiazhao Zhang, Yan Zhu, and He Wang. Maskclustering: View consensus based mask graph clustering for open-vocabulary 3d instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28274-28284, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.482, + 0.546 + ], + "angle": 0, + "content": "[53] Zihao Yan, Ruizhen Hu, Xingguang Yan, Luanmin Chen, Oliver Van Kaick, Hao Zhang, and Hui Huang. RPM-net: recurrent prediction of motion and parts from point cloud. arXiv preprint arXiv:2006.14865, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.482, + 0.603 + ], + "angle": 0, + "content": "[54] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In Proceedings of the European conference on computer vision (ECCV), pages 767-783, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.606, + 0.482, + 0.687 + ], + "angle": 0, + "content": "[55] Andy Zeng, Shuran Song, Kuan-Ting Yu, Elliott Donlon, Francois R Hogan, Maria Bauza, Daolin Ma, Orion Taylor, Melody Liu, Eudald Romo, et al. Robotic pick-and-place of novel objects in clutter with multi-affordance grasping and cross-domain image matching. The International Journal of Robotics Research, 41(7):690-705, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.69, + 0.482, + 0.758 + ], + "angle": 0, + "content": "[56] Vicky Zeng, Tabitha Edith Lee, Jacky Liang, and Oliver Kroemer. Visual identification of articulated object parts. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 2443-2450. IEEE, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[57] Chaoning Zhang, Dongshen Han, Yu Qiao, Jung Uk Kim, Sung-Ho Bae, Seungkyu Lee, and Choong Seon Hong. Faster segment anything: Towards lightweight sam for mobile applications. arXiv preprint arXiv:2306.14289, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[58] Allan Zhou, Moo Jin Kim, Lirui Wang, Pete Florence, and Chelsea Finn. Nerf in the palm of your hand: Corrective augmentation for robotics via novel-view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17907-17917, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[59] Shijie Zhou, Haoran Chang, Sicheng Jiang, Zhiwen Fan, Zehao Zhu, Dejia Xu, Pradyumna Chari, Suya You, Zhangyang Wang, and Achuta Kadambi. Feature 3dgs: Supercharging 3d gaussian splattering to enable distilled feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21676-21685, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[60] Luyang Zhu, Arsalan Mousavian, Yu Xiang, Hammad Mazhar, Jozef van Eenbergen, Shoubhik Debnath, and Dieter Fox. Rgb-d local implicit function for depth completion of transparent objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4649-4658, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12142" + } + ] +] \ No newline at end of file diff --git a/2025/IAAO_ Interactive Affordance Learning for Articulated Objects in 3D Environments/c94d31b4-1877-4bc2-8b0f-9d4d8e45806a_origin.pdf b/2025/IAAO_ Interactive Affordance Learning for Articulated Objects in 3D Environments/c94d31b4-1877-4bc2-8b0f-9d4d8e45806a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..fdb703aa94374ba88ad5aebb9298388eb31ae738 --- /dev/null +++ b/2025/IAAO_ Interactive Affordance Learning for Articulated Objects in 3D Environments/c94d31b4-1877-4bc2-8b0f-9d4d8e45806a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ee18ad3085e403c0dc73bfcb47f5647f0b5217c55f7a8ec843c2040ce0212721 +size 7664265 diff --git a/2025/IAAO_ Interactive Affordance Learning for Articulated Objects in 3D Environments/full.md b/2025/IAAO_ Interactive Affordance Learning for Articulated Objects in 3D Environments/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9746b413ed1b9022f67580f06c9f88b64aad8464 --- /dev/null +++ b/2025/IAAO_ Interactive Affordance Learning for Articulated Objects in 3D Environments/full.md @@ -0,0 +1,291 @@ +# IAAO: Interactive Affordance Learning for Articulated Objects in 3D Environments + +Can Zhang Gim Hee Lee + +Department of Computer Science, National University of Singapore + +can.zhang@u.nus.edu gimhee.lee@nus.edu.sg + +# Abstract + +This work presents IAAO, a novel framework that builds an explicit 3D model for intelligent agents to gain understanding of articulated objects in their environment through interaction. Unlike prior methods that rely on task-specific networks and assumptions about movable parts, our IAAO leverages large foundation models to estimate interactive affordances and part articulations in three stages. We first build hierarchical features and label fields for each object state using 3D Gaussian Splatting (3DGS) by distilling mask features and view-consistent labels from multi-view images. We then perform object- and part-level queries on the 3D Gaussian primitives to identify static and articulated elements, estimating global transformations and local articulation parameters along with affordances. Finally, scenes from different states are merged and refined based on the estimated transformations, enabling robust affordance-based interaction and manipulation of objects. Experimental results demonstrate the effectiveness of our method. Our source code is available at: https://lulusindazc.github.io/IAAOproject/. + +# 1. Introduction + +The real world contains various dynamic objects that continuously interact and move through time. Understanding and reconstructing the shapes, poses, and articulations of dynamic objects is crucial for many applications such as robotics, augmented reality and virtual reality (AR/VR), etc. While humans naturally develop an intuitive understanding of how objects move and interact, teaching this skill to an intelligent agent remains a complex challenge due to the diverse and intricate nature of articulated objects. Furthermore, a key capability still missing in intelligent agents is effective interaction with their environment, particularly with functional elements like door handles and light switches. For intelligent agents to interact with functional objects, in addition to object detection, they must also + +![](images/27f5e2f12eb4553f6ad1350557013517bc204be00f41b58200ba4c3ecd77e8c2.jpg) +Figure 1. Our IAAO requires multi-view images of the object or indoor scene from two different joint states (Left). The output is a 3D interactive field which supports interactions with multiple movable parts for fine-grained segmentation (e.g. Case 2: handles) and articulation reconstruction (e.g. Case 1: two articulated doors). + +understand the specific physical interactions each object affords—known as affordance [10]. The small size of these elements further intensifies the challenge, making them difficult to detect. In this work, we aim to reconstruct the 3D complex scene and interact with both objects with static geometry and objects with articulations and fine-grained affordances as illustrated in Fig. 1. + +Early approaches [21, 46] to understanding articulated objects rely on supervised learning with costly 3D annotations and category-specific models, which limit generalization to novel objects. Recent methods [16, 22] such as Ditto [16] and PARIS [22] have advanced this field by creating digital twins of objects based on observations in two joint configurations. Since Ditto is trained on multiview point clouds from specific categories, it struggles with objects that differ significantly from its training data. In contrast, PARIS [22] uses multi-view posed images to directly optimize an implicit representation without pretraining, allowing for improved generalization. However, it remains sensitive to initialization and may experience stability issues. Further developments [42, 47] have expanded scalability to objects with multiple moving parts based on implicit neural radiance fields (NeRFs). Nonetheless, several challenges remain unexplored, including generaliz + +ing to complex scenes, flexibly applied for manipulation, and practical assumptions (e.g. handling unknown camera poses, unaligned static object parts, and occlusion). + +In 3D perception, instance segmentation provides essential object-level understanding for agents interacting within scenes. Recent studies [27, 29] explore part-object segmentation to decompose objects into finer components such as cabinet drawers. However, these methods are preliminary and struggle with affordance detection due to the lack of precise small-scale geometry in existing datasets. Open-vocabulary models [6, 18, 35, 43] utilize foundation models such as CLIP [39] to extract semantic features and consolidate them into a multi-view consistent 3D neural field representation. However, they rely on costly multi-scale queries to refine segmentation boundaries, and face challenges in high-dimensional feature encoding. + +To address the limitations mentioned above, we introduce IAAO, a framework designed for interaction of intelligent agents with objects, parts, and functional affordances, as well as for articulating movable objects. We explicitly leverage 3D semantic information to provide our model with scene understanding and interaction capabilities for articulated objects. Key insights of our approach include: 1) semantic scene reconstruction where SAM masks enable multi-view consistent segmentations from partial views, 2) hierarchical feature fields to efficiently compress dense features from multiple views into 3DGS for object- and part-level interaction in diverse scenes, 3) explicit 3D Gaussian representations for affordance localization via query similarity and motion recovery via 3D point-to-2D pixel correspondences, 4) global-local transformation initialization to enhance robustness in realistic scenarios with potential scene misalignment between object states, and 5) scene state fusion merging two different scene states to enhance object geometry completeness. + +Specifically, our IAAO enables interactive affordance detection and object articulation via the following steps: 1) 3D model construction. Using multi-view posed images of each state, we model the 3D scene as a set of explicit 3D Gaussian primitives with 3D Gaussian Splatting (3DGS) [17]. 2) Hierarchical feature field construction. We propose an efficient distillation scheme where shape-aware 2D features and dense semantic features are extracted from large foundation models, including CLIP [39], SAM [19], and DINOv2 [33]. 2D features are efficiently distilled into 3D fields with a decoder. 3) Semantic-guided mask association across states. We first cluster SAM-generated masks from all views within each state to obtain view-consistent mask labels. We then compute 3D proposal features by merging the features from the corresponding 2D masks. By comparing pairwise feature similarities, we establish 3D mask-level correspondences across states that result in consistent mask label sets for the entire scene in each + +state. 4) Affordance prediction. Given a task description, we encode it with the CLIP text encoder and compare the text embeddings with the feature fields to identify the relevant affordance. 5) Motion recovery. To estimate transformations, we define a set of consistency and matching losses that recover both global scene-level and local part-level transformations that ensure accurate motion representation across different states. + +We summarize our main contributions as follows: + +- We introduce IAAO, an interactive affordance system utilizing 3D Gaussian fields embedded with hierarchical language-aligned semantics and class-agnostic masks. This system supports manipulation tasks guided by various prompts (point, mask, and language) at both object and part levels, regardless of object categories and shape. +- We propose a method to reconstruct motion via global matching at the scene level and local matching for articulated objects, without relying on impractical assumptions about static object alignment or known camera poses. +- Our IAAO achieves state-of-the-art performance across extensive experiments on multiple benchmarks, including synthetic, real-world, and indoor scene data. +- We show strong model generalization to complex indoor environments and previously unseen articulated objects, with no restrictions on the number of movable parts. + +# 2. Related Work + +Neural Fields for 3D Scene Understanding. 3D Scene understanding is primarily characterized by four classic representations, including volumetric fields ([34]), point clouds ([40]), 3D meshes ([7]), and depth maps ([11, 41, 54]). Unlike traditional representations, NeRF (Neural Radiance Fields) [28] introduces a neural implicit field to capture the geometry and appearance of a scene. Through a MultiLayer Perceptron (MLP) that takes 3D positions and 2D view directions as inputs, NeRF learns to implicitly represent the color and radiance of the scenes from a collection of posed images. However, a key drawback of NeRF is the slow training and rendering speed. Recently, 3D Gaussian Splatting (3DGS) [17] has been introduced as an alternative to implicit radiance field representations. In contrast to NeRF, 3DGS explicitly represents radiance fields as a collection of oriented 3D Gaussians. Each Gaussian is defined by its spatial position, opacity, and a covariance matrix, which allows flexible optimization. With efficient differentiable rasterization, 3DGS achieves fast rendering with high-quality results. Various work leverage neural fields as a representation robotic manipulation [18, 44, 45, 48, 58, 60]. Methods that leverage visual foundation models to construct neural feature fields [18, 37, 59] are most relevant to our work. With feature distillation via rendering, we reconstruct a hierarchical feature field for object localization and fine-grained affordance prediction. + +Visual Affordance. Affordance [9] refers to the potential interactions that an object or its parts facilitate for an agent. Affordance prediction involves deducing interaction opportunities from visual representations, e.g. images [4, 8, 25] or 3D models [3, 30, 31]. This field has gained significant interest in robotics as a foundational element for tasks e.g. grasping [26, 55], planning [50] and exploration [31]. However, although these methods excel at predicting affordance regions, they lack detailed interaction information, which is essential for engaging with functional elements. Recent work [2] addresses fine-grained functional elements using comprehensive natural language task descriptions for interaction. However, it depends on high-fidelity point clouds, restricting its practical applicability. We use high-resolution images with a Vision-Language Model (VLM) and 3D motion heuristics to predict affordances, eliminating the need for high-fidelity point clouds. + +Articulation Reasoning by Interaction. As embodied AI continues to evolve, the study of articulated objects has become a critical area of research. Recently, several synthetic [46, 49] and scanned data [15, 24, 27, 36] datasets have been introduced for articulated objects. These datasets contain annotations for part segmentation and motion parameters that enable data-driven approaches in the prediction of motion parameters from 3D point clouds [46, 53]. Recent works have been increasingly focused on real-world scenarios by detecting articulated parts and the motion parameters from images [15, 56] and videos [36]. Interactive perception involves agents gaining new insights from their interactions with the environment, and has been widely used to study object articulations [13, 14, 16, 32]. In this work, we focus on using language, and geometric and semantic scene understanding to achieve articulation reasoning. + +# 3. Our Method + +Objective. Given a scene of articulated objects represented by a set of multi-view posed images $\mathcal{I}^t = \{I_i^t\}_{i=1}^{N^t}$ at two states $t$ and $t'$ , the goal is to reconstruct the geometries and semantics of the objects in the scene. We focus on articulated objects which contain one or more movable parts with rotary or prismatic joints. Under the condition of unknown object categorory in the scene, we aim to recover the object part geometry, segmentation and joint articulations in the presence of occlusions, shape discrepancies and motion changes across different scene states. For static objects in the scene, the task degenerates to global transformation estimation between two input scene states for state matching. + +Overview. Fig. 2 shows our IAAO framework consisting of: 1) A semantic scene reconstruction stage (cf. Sec. 3.1) where we first extract CLIP features for objects in individual scene state. These features are regularized by class-agnostic masks generated from 2D foundation models, e.g. + +SAM [57] at the instance and part level. We then cluster view-inconsistent masks from all views to obtain 3D consistent labels, and build a hierarchical feature field and label field for all masks in each scene state via 3D Gaussian Splatting (3DGS). 2) An affordance and motion prediction stage (cf. Sec. 3.2) where we perform object- and part-level queries by directly operating on 3D primitives represented by 3D Gaussians. Static elements and articulated objects are identified by isolating the groups of 3D Gaussians corresponding to segmented objects and functional parts of the scene, i.e. affordance prediction. For motion recovery, we estimate global transformation from static Gaussian primitives and local articulation parameters from the segmented movable parts of objects. 3) A scene state fusion stage (cf. Sec. 3.3) enabling affordance interaction and object manipulation by merging and refining the reconstructed scenes from two states according to the estimated transformations. + +# 3.1. Semantic Scene Reconstruction + +Given a multi-view capture $\mathcal{I}^t$ of scene state $t$ , we construct a 3D model using 3DGS initialized with sparse point cloud generated from Structure from Motion (SfM) [40]. Each point serves as the center of a Gaussian primitive embedded with geometry and appearance parameters. These 3D Gaussian primitives $G^{t} = g_{p,p = 1}^{tP^{t}}$ are then rendered into 2D views via differentiable rasterization for parameter optimization. To enrich the reconstructed model with semantic information, we augment each 3D Gaussian with semantic feature embeddings derived from large foundation models. + +View-Consistent Mask Clustering. For each input image $I_{i}^{t} \in \mathcal{I}^{t}$ , we first derive the class-agnostic masks $M_{i}^{t} = \{m_{i,n}^{t}\}_{n=1}^{n_{i}^{t}}$ with an off-the-shelf mask predictor, where $n_{i}^{t}$ represents the number of masks in view $i$ . For $M_{i}^{t}$ , we investigated the use of SAM [57] to generate both instance- and part-level masks for arbitrary objects. Unfortunately, mask generation models such as SAM face several challenges: a) class-agnostic masks do not have complete 3D information, b) inconsistent segmentation across images due to variations in viewpoint and appearance, and c) lack of one-to-one correspondence among masks due to oversegmentation for individual object. We aim to generate 3D-consistent labels for all masks in $M^{t}$ by defining a mask label mapping function $\phi(M^{t}, G^{t}) \mapsto O^{t}$ . This function maps a set of masks $M_{o}^{t}$ belonging to the same 3D object to a consistent $o^{t} \in O^{t}$ within the 3D Gaussian primitives $G^{t}$ . + +To associate class-agnostic masks across all input views at scene state $t$ , we construct a mask graph to fuse the 2D masks $M^t$ of the entire scene into cohesive 3D instances $O^t$ . In the initial graph $\mathcal{G}_0^t = (\mathcal{V}_0^t, \mathcal{E}_0^t)$ , each node in $\mathcal{V}_0^t$ represents a mask detected from the views, and each edge in $\mathcal{E}_0^t$ represents potential matches between pairs of object parts. Following the mask clustering workflow from [52], + +![](images/adec8629e5093ea4f04cfe7247ed41f44f0f7587894a7e590e29ce644fe54d84.jpg) +Stage 1: Semantic Scene Reconstruction + +![](images/d03463a4d5b791fc4f27e28523e5f8f7ae47420cf69378f4106587d1870a5935.jpg) +Stage 3: Scene State Fusion +Figure 2. Our IAAO framework. 1) Top: Constructing 3D Gaussian fields in each state. We optimize 3DGS fields with hierarchical mask features, DINOv2 features and 3D-consistent mask labels generated from multi-view images. We also incorporate geometry information from depth images into the 3D Gaussians. 2) Bottom: Affordance and motion prediction. A query prompt is embedded using a pretrained encoder to localize relevant regions in the 3D Gaussians. For motion prediction, we optimize the transformation parameters by applying consistency and matching losses to 2D-3D correspondences between states. 3) Right: Scene fusion. Using the estimated transformations, we merge reconstructed 3DGS models from both states, aligning static and articulated elements. + +![](images/82cd04537932a936e0cf0481341cf4e78840e1c4733b5c9ea5d629700c38e946.jpg) +Stage 2: Affordance and Motion Prediction + +we then cluster the mask nodes and update the edges to facilitate mask association across different views. For each cluster, we combine the corresponding partial sets of Gaussians from individual masks to form a complete 3D instance and establish correspondences between 2D masks and 3D nodes. This results in the final graph $\mathcal{G}^t = (\mathcal{V}^t,\mathcal{E}^t)$ , where each node in $\mathcal{V}^t$ represents a 3D instance proposal in $O^t$ in scene state $t$ . We maintain a list of associated 2D masks $M_o^t$ for each $o^t$ , discarding severely occluded masks. This 2D-3D relationship enables consistent segmentation labels across all views. However, the mask label mapping $\phi (M^{t},G^{t})$ may be affected by noise due to varying part segmentation hierarchies across views. To mitigate segmentation ambiguities, we use the labeled masks $M_o^t$ in the final graph to supervise a label field for object part $o^t$ in the given scene state. To this end, we use a cross-entropy classification loss $\mathcal{L}_{\mathrm{label}}$ for enhanced consistency. + +Per-State Hierarchical Feature Field. We generate a set of feature embeddings from the given image $I_{i}^{t}$ using pretrained foundation models. Specifically, we utilize MaskCLIP [5] to produce mask-level CLIP features at both instance and part levels, and DINOv2 [1, 33] to extract DINO embeddings. To enable object- and part-level interaction, we construct a hierarchical feature field by embedding dense features from multi-view images into the 3D Gaussian primitives. Using semantic hierarchy from SAM, we enhance our 3D semantic field accuracy while making the querying process more efficient. Directly embedding high-dimensional 2D features would result in excess + +sive computational and memory costs. Therefore, we build a low-dimensional latent feature field and introduce a decoder to project the rendered features $\hat{F}^t$ back to the 2D feature space $F^t$ during the differentiable rasterization optimization process as: + +$$ +\mathcal {L} _ {f e a t} = \left\| \mathcal {D} \left(\hat {F} ^ {t} \left(I _ {i} ^ {t}\right)\right) - F ^ {t} \left(I _ {i} ^ {t}\right) \right\| _ {2} ^ {2}. \tag {1} +$$ + +The decoder $\mathcal{D}$ is implemented as a small MLP with three output branches: two for instance- and part-level CLIP features, and one for DINO features. + +Semantic Association between Scene States. Upon obtaining the mask label mapping functions $O^t$ and $O^{t'}$ for scene states $t$ and $t'$ , we establish a global 3D mask association which links the same semantic regions across both states into a unified 3D mask using their semantic features. For each 3D proposal in $O^t$ , we select the top- $k$ representative masks from $M_o^t$ and aggregate their rendered features to create the corresponding semantic primitives $F_o^t$ . We then build an affinity matrix capturing the cosine similarity between all 3D proposal pairs $(O^t, O^{t'})$ using their primitives $(F_O^t, F_O^{t'})$ from two scene states. For each query proposal $o^t \in O^t$ , we define its visual match as the target proposal in $O^{t'}$ with the highest similarity. + +# 3.2. Affordance and Motion Prediction + +Unlike NeRF-based methods [6, 18, 23] which require resource-intensive rendering to derive language-embedded features and geometry from implicit MLPs, our IAAO directly uses explicit Gaussian primitives for efficient object + +part localization and manipulation. Given the reconstructed 3D scene $G^{t} = \{g_{p}^{t}\}_{p = 1}^{P^{t}}$ at scene state $t$ , where each Gaussian point $g_{p}^{t}$ is embedded with geometric information (coordinates of points $p^t$ ) and additional features (color and semantic embeddings), our goal is to enable scene-aware affordance prediction and object articulation estimation. + +Functional Affordance Prediction. Functional affordances are defined as parts of the scene that facilitate interactions for agents to perform specific tasks. We utilize the reconstructed hierarchical feature and label fields as the 3D representations to predict the masks $\{m_a\}_{a = 1}^{K_a}$ and affordance labels $\{l_a\}_{a = 1}^{K_a}$ in the scene, where $K_{a}$ is the total number of affordance categories. We first build associations between affordance labels and our reconstructed mask label field, and then infer the point-level affordance of a functional element in the scene. Additionally, for task-specific reasoning, we encode the task description using the CLIP text encoder. To identify the relevant mask for a given description, we calculate the similarity between the embeddings in the feature field and the query embeddings. A mask is retrieved if its similarity score with the description text exceeds a predefined threshold. + +Global and Local Motion Recovery. Given the reconstructed scenes with different object articulations at two states, our objective is to estimate scene-aware transformation and 3D motion primitives for articulated parts. Our approach segments the reconstructed Gaussians $G^{t}$ and $G^{t'}$ into static scene elements and articulated objects via object-part querying for global and per-part local transformation estimation, respectively. 1) Object-level querying in complex scenes: In an indoor scene with a complex background and multiple objects, we start with object-level querying. Positive language queries are used to identify target objects for manipulation, while optional negative queries exclude irrelevant items. This selection process identifies 3D Gaussians with object-level CLIP features that align more closely with the positive queries than the negative ones, following standard CLIP feature-text comparison practices [45]. 2) Part-level querying for targeted objects: For individual objects in the scene, we apply part-level querying to focus on 3D Gaussians within the object. This allows for fine-grained transformation estimation specific to each articulated part. In addition to language, other user input prompts such as masks and points can also be used. Nonetheless, we do not further elaborate on these prompts as they can be easily obtained by a user clicking on the images. + +We define the global scene-aware transformation between two different scene states as $\xi_g^t = (s_g^t,R_g^t,T_g^t)$ where $s_g^t\in \mathbb{R}$ represents the scale factor, $R_{g}^{t}\in \mathrm{SO}(3)$ is the rotation matrix and $T_{g}^{t}\in \mathbb{R}^{3}$ is the translation vector. To initialize the transformation, we estimate an initial alignment using 3D Gaussians derived from static elements in the scene, excluding articulated objects. From the Gaussian + +points in both scene states, we select reliable points with opacity values above a threshold (empirically set at 0.7) to serve as input point sets for the coarse registration. Using the GeoTransformer [38], we calculate the scene alignment transformation by matching coarse superpoints and refining with dense point correspondences. Since these point clouds are based on reconstructed 3D Gaussians, the resulting transformation is coarse, providing an initial approximation for $\xi_{g}^{t}$ that can be further refined. + +After achieving global alignment between two scene states, we define the local transformation for each articulated part as $\xi_{o}^{t} = (R_{o}^{t},T_{o}^{t})$ , representing the movement of part $o$ from state $t$ to state $t'$ . All points within a part are assumed to follow the same motion parameters. The reverse mapping, which transforms part $o$ from state $t'$ back to state $t$ , is given by the inverse transformation, $\xi_{o}^{t'} = (R_{o}^{t},T_{o}^{t})^{-1}$ . To facilitate this process, we construct a 3D Gaussian correspondence field, allowing each Gaussian primitive $g_{p}^{t}(o)$ at state $t$ to be repositioned to its corresponding location $p^{t'}$ at state $t'$ based on the motion parameters of part $o^t$ . Formally: + +$$ +p ^ {t \rightarrow t ^ {\prime}} = \left\{\begin{array}{l l}f \left(p ^ {t}, \xi_ {g} ^ {t}\right) = R _ {g} ^ {t} p ^ {t} + T _ {g} ^ {t},&\text {i f} p \text {i n s t a t i c s c e n e ,}\\f \left(f \left(p ^ {t}, \xi_ {g} ^ {t}\right), \xi_ {o} ^ {t}\right),&\text {o t h e r w i s e .}\end{array}\right. \tag {2} +$$ + +The same rule applies to the backward transformation function $f^{-1}$ for obtaining $p^{t' \to t}$ (i.e. mapping from $t'$ back to $t$ ) with the transformations $\xi_g^{t'}$ and $\xi_o^{t'}$ . + +Establishing accurate dense correspondences between point clouds in two states is challenging due to the sparsity and noise in 3D points derived from the 3DGS model. To optimize the correspondence field, we instead match 3D Gaussian primitives to 2D pixels across the two states, guided by part geometry and segmentation information. Specifically, for a Gaussian primitive $g_{p}^{t}(o)$ of part $o$ at state $t$ , the corresponding 2D pixels at state $t'$ can be treated as observations from novel views. To enhance consistency, we incorporate mask-level features alongside the rendered image consistency loss $\mathcal{L}_{rgb}$ . For a 3D proposal $o^t$ in graph $\mathcal{G}^t$ , we identify the part mask list $\mathcal{M}_o^{t'}$ associated with the corresponding proposal $o^{t'}$ at state $t'$ . Following the splatting process, we render the feature field of $o^t$ to training view $n$ in target scene state $t'$ and decode the mask-level features by $\mathcal{D}(\hat{F}_o^t (I_n^{t'}))$ . We formulate the mask feature loss for part $o$ transformed from state $t$ to target state $t'$ as: + +$$ +\begin{array}{l} \mathcal {L} _ {m a s k} (o ^ {t \rightarrow t ^ {\prime}}) = \\ \Sigma_ {n} ^ {N _ {o} ^ {t ^ {\prime}}} \frac {M _ {o} ^ {t ^ {\prime}} \left(I _ {n} ^ {t ^ {\prime}}\right)}{\left| M _ {o} ^ {t ^ {\prime}} \left(I _ {n} ^ {t ^ {\prime}}\right) \right|} \cdot \| \mathcal {D} \left(\hat {F} _ {o} ^ {t} \left(I _ {n} ^ {t ^ {\prime}}\right)\right) - F _ {o} ^ {t ^ {\prime}} \left(I _ {n} ^ {t ^ {\prime}}\right) \| _ {2} ^ {2}, \tag {3} \\ \end{array} +$$ + +where $M_{o}^{t^{\prime}}(I_{n}^{t^{\prime}})$ and $F_{o}^{t^{\prime}}(I_{n}^{t^{\prime}})$ represent the corresponding part mask and 2D mask-level features of part $o$ in view $I_{n}^{t^{\prime}}$ at the target state $t^{\prime}$ . $N_{o}^{t}$ denotes the number of part masks for proposal $o^t$ across all training views. + +Since the mask-level loss provides only coarse guidance for transformation, we further propose to find dense correspondences between the 3D Gaussian primitives and 2D pixels by comparing their DINO features. We begin by computing the feature similarity matrix $\alpha_{p\rightarrow o}(I_n^{t'})$ between each pixel in the part mask $M_o^{t'}(I_n^{t'})$ in view $I_n^{t'}$ and the sampled Gaussian point $g_p^t$ from 3D proposal $o^t$ . We then normalize the similarity matrix $\alpha_{p\rightarrow o}(I_n^{t'})$ using a softmax over the entire mask to obtain the weight $\beta_{p\rightarrow o}(I_n^{t'})$ . Finally, we identify the corresponding 2D pixel $s_{p\rightarrow o}^{t'}(I_n^{t'})$ for the sampled 3D point $p^t$ using a weighted sum. See the supplementary material for details on the computation. For each point-pixel pair $(p^t,s_{p\rightarrow o}^{t'}(I_n^{t'}))$ , we filter out pairs that collide by checking whether $s_{p\rightarrow o}^{t'}(I_n^{t'})$ falls within the same part mask $o^{t'}$ . We then define the matching loss as follows: + +$$ +\mathcal {L} _ {\text {m a t c h}} = \left\| \pi_ {n} ^ {t ^ {\prime}} \left(p ^ {t \rightarrow t ^ {\prime}}\right) - s _ {p \rightarrow n} ^ {t ^ {\prime}} \left(I _ {n} ^ {t ^ {\prime}}\right)\right\| _ {2} ^ {2}, \tag {4} +$$ + +where $\pi_n^{t'}$ represents the projection of 3D point $p^{t\rightarrow t'}$ to view $I_{n}^{t^{\prime}}$ in scene state $t^\prime$ + +**Optimization.** For the first stage, each scene is reconstructed independently at two different states. We optionally use depth map for further geometry regularization. In the second stage, we reconstruct motion of global scene and local articulated objects. The total loss is defined as: + +$$ +\mathcal {L} = \lambda_ {\text {c o n s}} \left(\mathcal {L} _ {\text {r g b}} + \mathcal {L} _ {\text {m a s k}} + \mathcal {L} _ {\text {l a b e l}}\right) + \lambda_ {\text {m a t c h}} \mathcal {L} _ {\text {m a t c h}}. \tag {5} +$$ + +$\lambda_{\text{cons}}$ and $\lambda_{\text{match}}$ denote the balancing weights of consistency loss and matching loss. + +# 3.3. Scene State Fusion + +After training, we merge the 3DGS models from the two states using the estimated transformations to fill in the occluded regions in each state. Details of our fusion and filtering strategy are provided in the supplementary material. The combined 3D Gaussians facilitate object manipulation by addressing potential artifacts that may appear in occluded areas of the reconstructed scene. Optional fine-tuning can be applied to either the edited regions or the entire scene for further refinement. + +# 4. Experiment + +# 4.1. Experimental Setup + +Datasets. 1) PARIS two-part object dataset. This dataset contains two-part articulated objects, featuring 10 synthetic objects from PartNet-Mobility [49] and 2 real-world objects captured using MultiScan [27]. Each object comprises a movable part and a static part, observed in two distinct joint states. The dataset includes RGB images, object masks from 100 random viewpoints, and depth images for both synthetic and real-world objects. 2) Synthetic multi-part + +object dataset. This dataset includes two synthetic scenes containing multi-part objects from PartNet-Mobility, with each object featuring one static part and multiple movable parts. Each object is observed in two articulation states, and the dataset provides RGB, depth, and mask data from 100 random viewpoints. 3) Indoor scene OmniSim dataset. OmniSim is generated using the OmniGibson [20] simulator with various indoor scene models. By adjusting the rotation of articulated object joints, we generate some scenes for evaluation, each containing RGBD images, interactive object masks and object state metrics at each state. + +Metrics. To evaluate articulation models, we use: 1) Axis Ang Err $(^{\circ})$ means angular error of the predicted joint axis for both revolute and prismatic joints. 2) Axis Pos Err (0.1 m): The shortest distance between the predicted and true joint axes for revolute joints. For cross-state part motion, we use Part Motion Err $(^{\circ}$ or m). It evaluates joint state precision by calculating the geodesic error for predicted rotations in revolute joints or the Euclidean error for translations in prismatic joints, as specified in PARIS [22]. For object and part mesh reconstruction, we use Chamfer-L1 distance (CD), reporting CD-w for the full surface, CD-s for static parts, and CD-m for movable parts. + +Baselines. Ditto [16] and PARIS [22] are models for reconstructing part-level shape and motion in two-part articulated objects, using multi-view data. Ditto assumes one moving part and is pre-trained on specific categories, while PARIS uses a NeRF-based representation to handle unknown objects. $PARIS^{*}$ is an enhanced version with depth supervision, and $PARIS - m^{*}$ extends PARIS to multi-part objects. CSG-reg uses TSDF fusion and Constructive Solid Geometry for part segmentation, with registration for alignment. 3Dseg-reg follows a similar process but employs PA-Conv [51] for part segmentation, reporting results only for trained categories due to limited generalization. DigitalTwinArt [47] divides reconstruction into two stages: 1) It reconstructs object-level shape independently of articulation; 2) It recovers the articulation model by identifying part segmentation and motions through state correspondences. + +Evaluation Setup. There are some key differences on evaluation setup between our model and previous models as we firstly incorporate the semantics into the explicit neural field supporting robust interactions even in complex environment. Unlike prior methods which assume the number of parts is known, our model can detect any articulated objects if enough information for reconstructing object shape is given from multi-view images. To identify corresponding parts for evaluation, prior work need to iterate through all possible pairs of predicted and ground-truth parts, selecting the match with the smallest total Chamfer distance. In our model, we identify the corresponding parts based on the constructed semantic feature and mask label field. We extract the mesh based on the 3D point clouds derived + +
SimulationReal
FoldChairFridgeLaptopOvenScissorStaplerUSBWasherBladeStorageAllFridgeStorageAll
Axis Ang ↓Ditto [16]89.3589.783.120.964.5089.8689.7789.5179.54*6.3254.221.715.883.80
PARIS [22]8.089.150.020.043.8239.730.1325.3615.380.0310.171.6443.1322.39
PARIS* [22]15.792.930.037.4316.628.170.7118.4041.280.0311.141.9030.1016.00
CSG-reg0.100.270.470.350.280.3011.7871.937.642.829.608.9269.7139.31
3Dseg-reg--2.34-----9.40-----
DigitalTwinArt [47]0.030.070.060.220.110.060.110.430.270.060.142.1018.1110.11
IAAO (Ours)0.020.040.050.170.080.050.090.360.190.060.111.915.828.86
Axis Pos ↓Ditto [16]3.771.02*0.010.135.700.205.410.66--2.111.84-0.92
PARIS [22]0.450.380.000.002.102.272.361.50--1.130.34-0.17
PARIS* [22]0.251.130.000.051.594.673.353.28--1.790.50-0.25
CSG-reg0.020.000.200.180.010.020.012.13--0.321.46-0.73
3Dseg-reg--0.10-----------
DigitalTwinArt [47]0.010.010.000.010.020.010.000.01--0.010.57-0.29
IAAO (Ours)0.010.000.000.000.010.010.010.00--0.010.32-0.16
Part Motion ↓Ditto [16]99.36×5.182.0919.2856.6180.6055.72×0.0939.878.430.384.41
PARIS [22]131.6624.580.030.03120.70110.8064.8560.350.340.3051.362.160.561.36
PARIS* [22]127.3445.260.039.1368.36107.7696.9349.770.360.3050.521.580.571.07
CSG-reg0.130.290.350.580.200.4410.48158.990.050.0417.1614.820.647.73
3Dseg-reg--1.61-----0.15-----
DigitalTwinArt [47]0.160.090.080.110.150.050.110.250.000.000.101.860.201.03
IAAO (Ours)0.130.070.030.040.110.050.090.210.000.000.071.460.150.81
CD-s ↓Ditto [14]33.793.050.252.5239.0741.642.6410.3246.909.1818.9447.0116.0931.55
PARIS [22]9.163.651.941.8812.951.882.6925.391.1912.767.1842.5754.5448.56
PARIS* [22]10.208.8215.581.953.182.482.4812.191.408.676.4611.6420.2515.94
CSG-reg1.691.453.933.263.262.221.954.530.597.062.706.3312.559.44
3Dseg-reg--0.76-----66.31-----
DigitalTwinArt [47]0.180.604.660.402.652.194.804.690.554.692.102.5310.866.69
IAAO (Ours)0.160.510.170.361.981.864.433.700.514.271.802.479.836.15
CD-m ↓Ditto [14]141.110.990.190.9420.6831.2115.8812.89195.932.2042.2050.6020.3535.48
PARIS [22]8.997.760.2128.7046.6419.275.32178.4325.2176.6939.7245.66864.82455.24
PARIS* [22]17.977.230.156.5416.6530.4610.17265.27117.9952.3452.4877.85474.57276.21
CSG-reg1.9121.710.42256.991.956.3629.78436.4226.621.3978.36442.17521.49481.83
3Dseg-reg--1.01-----6.23-----
DigitalTwinArt [47]0.150.270.160.470.412.271.340.361.500.370.731.1426.4613.80
IAAO (Ours)0.130.180.150.410.402.091.180.331.420.250.651.0221.4311.23
CD-w ↓Ditto [14]6.802.160.312.511.702.382.097.2942.043.917.126.5014.0810.29
PARIS [22]1.802.920.3011.7310.493.582.0024.380.608.576.6422.9863.3543.16
PARIS* [22]4.375.530.263.183.905.271.7810.110.587.804.288.9932.1020.55
CSG-reg0.480.980.403.001.701.991.204.480.564.001.885.7114.2910.00
3Dseg-reg--0.81-----0.78-----
DigitalTwinArt [47]0.270.700.354.180.432.191.184.740.363.991.842.199.335.76
IAAO (Ours)0.220.610.273.490.341.961.064.410.343.941.662.029.185.60
+ +Table 1. Results on PARIS dataset including both synthetic and real data. 'x': failure case, and \*\*: joint axis or position. -': no result is available. Best result in bold. + +
Axis Ang 0↓Axis Ang 1 ↓Axis Pos 0 ↓Axis Pos 1 ↓Part Motion 0 ↓Part Motion 1 ↓CD-s ↓CD-m 0 ↓CD-m 1 ↓CD-w ↓
Fridge-mPARIS*-m [22]34.5215.913.601.6386.21105.868.52526.19160.8615.00
DigitalTwinArt [47]0.160.100.010.000.110.130.610.400.520.89
IAAO (Ours)0.160.080.010.000.090.090.540.440.470.96
Storage-mPARIS*-m [22]43.2626.1810.42-79.840.648.56128.62266.718.66
DigitalTwinArt [47]0.210.880.05-0.130.000.850.213.460.99
IAAO (Ours)0.180.910.04-0.100.000.780.263.350.97
+ +Table 2. Results on multi-part object dataset, averaged over 10 trials with different random seeds. Joint 1 of "Storage-m" is solely prismatic with no Axis Pos. '-': no result is available. Best results in bold. + +from 3DGS model via SuGaR [12] for evaluation. Following [22, 47], we transform our extracted parts with predicted motions to start state $t$ for evaluation. + +# 4.2. Quantitative Results + +Tab. 1 shows quantitative results from the PARIS Two-Part Object Dataset, covering synthetic and real instances. Ditto performs well on seen categories but struggles with unseen ones. PARIS occasionally fails in shape and articu + +lation reconstruction, while depth supervision in PARIS* improves shape accuracy for complex and real-world objects but increases articulation errors due to optimization challenges. CSG-reg and 3Dseg-reg handle simple objects well but struggle with complex ones, as segmentation errors affect articulation estimation. + +Tab. 2 presents the results for multi-part objects. DigitalTwinArt [47] outperforms PARIS* in shape and articulation reconstruction by incorporating additional supervi + +![](images/89c9354449d82712c5da8a3ab67c7479c09ada6c7cfb2f93029f474b6e05d0d1.jpg) +Input State $t$ + +![](images/28dea88c1c409761ac03efc7dde903dfb89b1a47754db10ae7f9139fc7322908.jpg) +Figure 3. Qualitative analysis of shape reconstruction, part segmentation, and joint prediction results on multi-part object dataset. + +![](images/6efbf2f792257d879d606dd525cab3d11e7171ccdea792aa6dd2a762f5743812.jpg) +Input State $t^{\prime}$ + +![](images/b91e2bc9b5cc41f18b4a764ba040032890c75cd3f83cd9b4a7f3d33c9c0444b5.jpg) + +![](images/7ed2f1a62e2b807252e49a4dcb0d2874bf1e57c923223c9461152b64fd6b7aaf.jpg) +DigitalTwinArt + +![](images/13dada9615698bce585a36479a82796c309641397aa2e1a2b74e53412daf17fa.jpg) + +![](images/037e560b46698e76ad2993d6b34ee4017145eda164add557532a09b933251b3f.jpg) +Ours + +![](images/b84eb0f24ff6a40ca1e10d3f94d80597e779743cfaf49a0d6444b8f4b1a93dfd.jpg) + +![](images/dd0508e1e8dd9913d693ca0808d6dce190c68df45232188a33e61f5b0b27abea.jpg) +Figure 4. Qualitative results of shape reconstruction, part segmentation, and joint prediction on PARIS. + +![](images/f6d070a5e79eaf79715f24ba579dab346a7120340df4ac72474f41ded37c9f64.jpg) +Figure 5. Motion snapshots on PARIS & multi-part object. + +![](images/b2640977660d1bae9da69cdd7570becc2c96823c8753170ed0e1056d770dce37.jpg) +Figure 6. Motion snapshots on scene-level OmniSim dataset. + +sion on 3D geometries. Our IAAO achieves competitive results by leveraging geometric and semantic information from VLMs, enhancing object part modeling. + +# 4.3. Qualitative Results + +Fig. 3 and Fig. 4 show qualitative results on shape reconstruction, part segmentation, and joint prediction for the multi-part object and PARIS two-part datasets. Our IAAO achieves more complete segmentation results while also enabling interaction with affordance and segmentation features. Fig. 5 and 6 show snapshots of the generated motions on several examples at object and scene levels in a simulator. It shows IAAO can produce smooth interpolations of + +
MethodAxis Ang ↓Axis Pos ↓Part Motion ↓CD-s ↓CD-m ↓
Proposed0.280.010.110.800.35
w/o matching4.591.388.571.21123.71
w/o mask14.820.010.282.59129.14
w/o rgb0.560.010.170.810.42
w/o label0.540.010.158.56169.73
+ +Table 3. Ablation study on multi-part object dataset. + +articulated objects with good part geometry modeling. + +# 4.4. Ablation Study + +We assess the impact of our design choices specifically on multi-part objects, which present greater challenges. We first perform experiments to test out the impact of 2D-3D matching loss ('w/o matching'). As indicated in Tab. 3, the matching loss plays a crucial role in enhancing articulation reconstruction. We also study constraints from mask features and labels, which act as a coarse correspondence mechanism at the superpoint (mask) level. These constraints also enforce consistency within the feature fields of each mask, leading to improved part segmentation and more accurate reconstruction of articulated components. + +# 5. Conclusion + +We present IAAO, a framework that facilitates interactive affordance detection and object articulation from two scene sequences. Each scene sequence captures movable parts of objects in different joint states. Our IAAO can handle both simple scenes with single objects and complex indoor scenes without limitations on the number of objects and movable parts. The entire scene is reconstructed using 3D Gaussians, embedded with robust zero-shot generalization capabilities from 2D foundation models. Affordance prediction is achieved through object and part queries on the 3D Gaussian primitives. Motion estimation is achieved by establishing 2D-3D correspondences within each object to track transformations. Static elements serve for global scene alignment while movable elements are for local transformation. Extensive experiments demonstrate that our approach outperforms existing methods. + +Acknowledgement. This research / project is supported by the National Research Foundation (NRF) Singapore, under its NRF-Investigatorship Programme (Award ID. NRF-NRFI09-0008). + +# References + +[1] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers, 2023. 4 +[2] Alexandros Delitzas, Ayca Takmaz, Federico Tombari, Robert Sumner, Marc Pollefeys, and Francis Engelmann. Scenefun3d: Fine-grained functionality and affordance understanding in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14531-14542, 2024. 3 +[3] Shengheng Deng, Xun Xu, Chaozheng Wu, Ke Chen, and Kui Jia. 3d affordancenet: A benchmark for visual object affordance understanding. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1778-1787, 2021. 3 +[4] Thanh-Toan Do, Anh Nguyen, and Ian Reid. Affordancenet: An end-to-end deep learning approach for object affordance detection. In 2018 IEEE international conference on robotics and automation (ICRA), pages 5882-5889. IEEE, 2018. 3 +[5] Xiaoyi Dong, Jianmin Bao, Yinglin Zheng, Ting Zhang, Dongdong Chen, Hao Yang, Ming Zeng, Weiming Zhang, Lu Yuan, Dong Chen, et al. Maskclip: Masked self-distillation advances contrastive language-image pretraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10995-11005, 2023. 4 +[6] Francis Engelmann, Fabian Manhardt, Michael Niemeyer, Keisuke Tateno, Marc Pollefeys, and Federico Tombari. Opennerf: Open set 3d neural scene segmentation with pixelwise features and rendered novel views. arXiv preprint arXiv:2404.03650, 2024. 2, 4 +[7] Carlos Hernández Esteban and Francis Schmitt. Silhouette and stereo fusion for 3d object modeling. Computer Vision and Image Understanding, 96(3):367-392, 2004. 2 +[8] Kuan Fang, Te-Lin Wu, Daniel Yang, Silvio Savarese, and Joseph J Lim. Demo2vec: Reasoning object affordances from online videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2139-2147, 2018. 3 +[9] JJ Gibson. The theory of affordances. Perceiving, acting and knowing: Towards an ecological psychology/Erlbaum, 1977. 3 +[10] James J Gibson. The ecological approach to visual perception: classic edition. Psychology press, 2014. 1 +[11] Michael Goesele, Brian Curless, and Steven M Seitz. Multiview stereo revisited. In 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'06), pages 2402-2409. IEEE, 2006. 2 +[12] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5354-5363, 2024. 7 + +[13] Karol Hausman, Scott Niekum, Sarah Osentoski, and Gaurav S Sukhatme. Active articulation model estimation through interactive perception. In 2015 IEEE International Conference on Robotics and Automation (ICRA), pages 3305-3312. IEEE, 2015. 3 +[14] Cheng-Chun Hsu, Zhenyu Jiang, and Yuke Zhu. Ditto in the house: Building articulation models of indoor scenes through interactive perception. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 3933-3939. IEEE, 2023. 3, 7 +[15] Hanxiao Jiang, Yongsen Mao, Manolis Savva, and Angel X Chang. Opd: Single-view 3d openable part detection. In European Conference on Computer Vision, pages 410-426. Springer, 2022. 3 +[16] Zhenyu Jiang, Cheng-Chun Hsu, and Yuke Zhu. Ditto: Building digital twins of articulated objects from interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5616-5626, 2022. 1, 3, 6, 7 +[17] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 2 +[18] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Lerf: Language embedded radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19729-19739, 2023. 2, 4 +[19] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 2 +[20] Chengshu Li, Ruohan Zhang, Josiah Wong, Cem Gokmen, Sanjana Srivastava, Roberto Martin-Martín, Chen Wang, Gabrael Levine, Wensi Ai, Benjamin Martinez, et al. Behavior-1k: A human-centered, embodied ai benchmark with 1,000 everyday activities and realistic simulation. arXiv preprint arXiv:2403.09227, 2024. 6 +[21] Xiaolong Li, He Wang, Li Yi, Leonidas J Guibas, A Lynn Abbott, and Shuran Song. Category-level articulated object pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3706-3715, 2020. 1 +[22] Jiayi Liu, Ali Mahdavi-Amiri, and Manolis Savva. Paris: Part-level reconstruction and motion analysis for articulated objects. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 352-363, 2023. 1, 6, 7 +[23] Kunhao Liu, Fangneng Zhan, Jiahui Zhang, Muyu Xu, Yingchen Yu, Abdulmotaleb El Saddik, Christian Theobalt, Eric Xing, and Shijian Lu. Weakly supervised 3d open-vocabulary segmentation. Advances in Neural Information Processing Systems, 36:53433-53456, 2023. 4 +[24] Liu Liu, Wenqiang Xu, Haoyuan Fu, Sucheng Qian, Qiao-jun Yu, Yang Han, and Cewu Lu. Akb-48: A real-world articulated object knowledge base. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14809–14818, 2022. 3 + +[25] Timo Luddecke and Florentin Worgotter. Learning to segment affordances. In Proceedings of the IEEE International Conference on Computer Vision Workshops, pages 769-776, 2017. 3 +[26] Priyanka Mandikal and Kristen Grauman. Learning dexterous grasping with object-centric visual affordances. In 2021 IEEE international conference on robotics and automation (ICRA), pages 6169-6176. IEEE, 2021. 3 +[27] Yongsen Mao, Yiming Zhang, Hanxiao Jiang, Angel Chang, and Manolis Savva. Multiscan: Scalable rgbd scanning for 3d environments with articulated objects. Advances in neural information processing systems, 35:9058-9071, 2022. 2, 3, 6 +[28] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2 +[29] Kaichun Mo, Shilin Zhu, Angel X Chang, Li Yi, Subarna Tripathi, Leonidas J Guibas, and Hao Su. Partnet: A largescale benchmark for fine-grained and hierarchical part-level 3d object understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 909-918, 2019. 2 +[30] Kaichun Mo, Yuzhe Qin, Fanbo Xiang, Hao Su, and Leonidas Guibas. O2o-afford: Annotation-free large-scale object-object affordance learning. In Conference on robot learning, pages 1666-1677. PMLR, 2022. 3 +[31] Tushar Nagarajan and Kristen Grauman. Learning affordance landscapes for interaction exploration in 3d environments. Advances in Neural Information Processing Systems, 33:2005-2015, 2020. 3 +[32] Neil Nie, Samir Yitzhak Gadre, Kiana Ehsani, and Shuran Song. Structure from action: Learning interactions for articulated object 3d structure discovery. arXiv preprint arXiv:2207.08997, 2022. 3 +[33] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 2, 4 +[34] Sylvain Paris, François X Sillion, and Long Quan. A surface reconstruction method using global graph cut optimization. International Journal of Computer Vision, 66:141-161, 2006. 2 +[35] Songyou Peng, Kyle Genova, Chiyu Jiang, Andrea Tagliasacchi, Marc Pollefeys, Thomas Funkhouser, et al. Openscene: 3d scene understanding with open vocabularies. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 815-824, 2023. 2 +[36] Shengyi Qian, Linyi Jin, Chris Rockwell, Siyi Chen, and David F Fouhey. Understanding 3d object articulation in internet videos. In Proceedings of the IEEE/CVF Conference + +on Computer Vision and Pattern Recognition, pages 1599-1609, 2022. 3 +[37] Minghan Qin, Wanhua Li, Jiawei Zhou, Haoqian Wang, and Hanspeter Pfister. Langsplat: 3d language gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20051-20060, 2024. 2 +[38] Zheng Qin, Hao Yu, Changjian Wang, Yulan Guo, Yuxing Peng, and Kai Xu. Geometric transformer for fast and robust point cloud registration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11143-11152, 2022. 5 +[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2 +[40] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 2, 3 +[41] Christoph Strecha, Rik Fransens, and Luc Van Gool. Combined depth and outlier estimation in multi-view stereo. In 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'06), pages 2394-2401. IEEE, 2006. 2 +[42] Archana Swaminathan, Anubhav Gupta, Kamal Gupta, Shishira R Maiya, Vatsal Agarwal, and Abhinav Shrivastava. Leia: Latent view-invariant embeddings for implicit 3d articulation. In European Conference on Computer Vision, pages 210-227. Springer, 2025. 1 +[43] Ayca Takmaz, Elisabetta Fedele, Robert Sumner, Marc Pollefeys, Federico Tombari, and Francis Engelmann. Openmask3d: Open-vocabulary 3d instance segmentation. Advances in Neural Information Processing Systems, 36, 2024. 2 +[44] Zhenggang Tang, Balakumar Sundaralingam, Jonathan Tremblay, Bowen Wen, Ye Yuan, Stephen Tyree, Charles Loop, Alexander Schwing, and Stan Birchfield. RGB-only reconstruction of tabletop scenes for collision-free manipulator control. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 1778-1785. IEEE, 2023. 2 +[45] Can Wang, Mengei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3835-3844, 2022. 2, 5 +[46] Xiaogang Wang, Bin Zhou, Yahao Shi, Xiaowu Chen, Qinping Zhao, and Kai Xu. Shape2motion: Joint analysis of motion parts and attributes from 3d shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8876-8884, 2019. 1, 3 +[47] Yijia Weng, Bowen Wen, Jonathan Tremblay, Valts Blukis, Dieter Fox, Leonidas Guibas, and Stan Birchfield. Neural implicit representation for building digital twins of unknown + +articulated objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3141-3150, 2024. 1, 6, 7 +[48] Youngsun Wi, Pete Florence, Andy Zeng, and Nima Fazeli. Virdo: Visio-tactile implicit representations of deformable objects. In 2022 International Conference on Robotics and Automation (ICRA), pages 3583-3590. IEEE, 2022. 2 +[49] Fanbo Xiang, Yuzhe Qin, Kaichun Mo, Yikuan Xia, Hao Zhu, Fangchen Liu, Minghua Liu, Hanxiao Jiang, Yifu Yuan, He Wang, et al. Sapien: A simulated part-based interactive environment. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11097-11107, 2020. 3, 6 +[50] Danfei Xu, Ajay Mandlekar, Roberto Martin-Martin, Yuke Zhu, Silvio Savarese, and Li Fei-Fei. Deep affordance foresight: Planning through what can be done in the future. In 2021 IEEE international conference on robotics and automation (ICRA), pages 6206–6213. IEEE, 2021. 3 +[51] Mutian Xu, Runyu Ding, Hengshuang Zhao, and Xiaojuan Qi. Paconv: Position adaptive convolution with dynamic kernel assembling on point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3173-3182, 2021. 6 +[52] Mi Yan, Jiazhao Zhang, Yan Zhu, and He Wang. Maskclustering: View consensus based mask graph clustering for open-vocabulary 3d instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28274-28284, 2024. 3 +[53] Zihao Yan, Ruizhen Hu, Xingguang Yan, Luanmin Chen, Oliver Van Kaick, Hao Zhang, and Hui Huang. RPM-net: recurrent prediction of motion and parts from point cloud. arXiv preprint arXiv:2006.14865, 2020. 3 +[54] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In Proceedings of the European conference on computer vision (ECCV), pages 767-783, 2018. 2 +[55] Andy Zeng, Shuran Song, Kuan-Ting Yu, Elliott Donlon, Francois R Hogan, Maria Bauza, Daolin Ma, Orion Taylor, Melody Liu, Eudald Romo, et al. Robotic pick-and-place of novel objects in clutter with multi-affordance grasping and cross-domain image matching. The International Journal of Robotics Research, 41(7):690-705, 2022. 3 +[56] Vicky Zeng, Tabitha Edith Lee, Jacky Liang, and Oliver Kroemer. Visual identification of articulated object parts. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 2443-2450. IEEE, 2021. 3 +[57] Chaoning Zhang, Dongshen Han, Yu Qiao, Jung Uk Kim, Sung-Ho Bae, Seungkyu Lee, and Choong Seon Hong. Faster segment anything: Towards lightweight sam for mobile applications. arXiv preprint arXiv:2306.14289, 2023. 3 +[58] Allan Zhou, Moo Jin Kim, Lirui Wang, Pete Florence, and Chelsea Finn. Nerf in the palm of your hand: Corrective augmentation for robotics via novel-view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17907-17917, 2023. 2 + +[59] Shijie Zhou, Haoran Chang, Sicheng Jiang, Zhiwen Fan, Zehao Zhu, Dejia Xu, Pradyumna Chari, Suya You, Zhangyang Wang, and Achuta Kadambi. Feature 3dgs: Supercharging 3d gaussian splattering to enable distilled feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21676-21685, 2024. 2 +[60] Luyang Zhu, Arsalan Mousavian, Yu Xiang, Hammad Mazhar, Jozef van Eenbergen, Shoubhik Debnath, and Dieter Fox. Rgb-d local implicit function for depth completion of transparent objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4649-4658, 2021. 2 \ No newline at end of file diff --git a/2025/IAAO_ Interactive Affordance Learning for Articulated Objects in 3D Environments/images.zip b/2025/IAAO_ Interactive Affordance Learning for Articulated Objects in 3D Environments/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5d6aede476ac1e86d605af68db65817a7f640aa3 --- /dev/null +++ b/2025/IAAO_ Interactive Affordance Learning for Articulated Objects in 3D Environments/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3f6c4e6a601d8304ed9fc7ec38130a7d256374de506fb60de606a0a7ed36fd04 +size 522758 diff --git a/2025/IAAO_ Interactive Affordance Learning for Articulated Objects in 3D Environments/layout.json b/2025/IAAO_ Interactive Affordance Learning for Articulated Objects in 3D Environments/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4335417391fe8e0beafc35a304a4162df1757aa5 --- /dev/null +++ b/2025/IAAO_ Interactive Affordance Learning for Articulated Objects in 3D Environments/layout.json @@ -0,0 +1,9279 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 91, + 102, + 520, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 102, + 520, + 138 + ], + "spans": [ + { + "bbox": [ + 91, + 102, + 520, + 138 + ], + "type": "text", + "content": "IAAO: Interactive Affordance Learning for Articulated Objects in 3D Environments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 231, + 162, + 380, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 162, + 380, + 175 + ], + "spans": [ + { + "bbox": [ + 231, + 162, + 380, + 175 + ], + "type": "text", + "content": "Can Zhang Gim Hee Lee" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 141, + 176, + 468, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 176, + 468, + 190 + ], + "spans": [ + { + "bbox": [ + 141, + 176, + 468, + 190 + ], + "type": "text", + "content": "Department of Computer Science, National University of Singapore" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 182, + 192, + 425, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 192, + 425, + 203 + ], + "spans": [ + { + "bbox": [ + 182, + 192, + 425, + 203 + ], + "type": "text", + "content": "can.zhang@u.nus.edu gimhee.lee@nus.edu.sg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 258, + 296, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 296, + 498 + ], + "type": "text", + "content": "This work presents IAAO, a novel framework that builds an explicit 3D model for intelligent agents to gain understanding of articulated objects in their environment through interaction. Unlike prior methods that rely on task-specific networks and assumptions about movable parts, our IAAO leverages large foundation models to estimate interactive affordances and part articulations in three stages. We first build hierarchical features and label fields for each object state using 3D Gaussian Splatting (3DGS) by distilling mask features and view-consistent labels from multi-view images. We then perform object- and part-level queries on the 3D Gaussian primitives to identify static and articulated elements, estimating global transformations and local articulation parameters along with affordances. Finally, scenes from different states are merged and refined based on the estimated transformations, enabling robust affordance-based interaction and manipulation of objects. Experimental results demonstrate the effectiveness of our method. Our source code is available at: https://lulusindazc.github.io/IAAOproject/." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 524, + 135, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 524, + 135, + 537 + ], + "spans": [ + { + "bbox": [ + 56, + 524, + 135, + 537 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": "The real world contains various dynamic objects that continuously interact and move through time. Understanding and reconstructing the shapes, poses, and articulations of dynamic objects is crucial for many applications such as robotics, augmented reality and virtual reality (AR/VR), etc. While humans naturally develop an intuitive understanding of how objects move and interact, teaching this skill to an intelligent agent remains a complex challenge due to the diverse and intricate nature of articulated objects. Furthermore, a key capability still missing in intelligent agents is effective interaction with their environment, particularly with functional elements like door handles and light switches. For intelligent agents to interact with functional objects, in addition to object detection, they must also" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 312, + 231, + 560, + 335 + ], + "blocks": [ + { + "bbox": [ + 312, + 231, + 560, + 335 + ], + "lines": [ + { + "bbox": [ + 312, + 231, + 560, + 335 + ], + "spans": [ + { + "bbox": [ + 312, + 231, + 560, + 335 + ], + "type": "image", + "image_path": "27f5e2f12eb4553f6ad1350557013517bc204be00f41b58200ba4c3ecd77e8c2.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 345, + 555, + 400 + ], + "lines": [ + { + "bbox": [ + 313, + 345, + 555, + 400 + ], + "spans": [ + { + "bbox": [ + 313, + 345, + 555, + 400 + ], + "type": "text", + "content": "Figure 1. Our IAAO requires multi-view images of the object or indoor scene from two different joint states (Left). The output is a 3D interactive field which supports interactions with multiple movable parts for fine-grained segmentation (e.g. Case 2: handles) and articulation reconstruction (e.g. Case 1: two articulated doors)." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 424, + 555, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 424, + 555, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 424, + 555, + 509 + ], + "type": "text", + "content": "understand the specific physical interactions each object affords—known as affordance [10]. The small size of these elements further intensifies the challenge, making them difficult to detect. In this work, we aim to reconstruct the 3D complex scene and interact with both objects with static geometry and objects with articulations and fine-grained affordances as illustrated in Fig. 1." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": "Early approaches [21, 46] to understanding articulated objects rely on supervised learning with costly 3D annotations and category-specific models, which limit generalization to novel objects. Recent methods [16, 22] such as Ditto [16] and PARIS [22] have advanced this field by creating digital twins of objects based on observations in two joint configurations. Since Ditto is trained on multiview point clouds from specific categories, it struggles with objects that differ significantly from its training data. In contrast, PARIS [22] uses multi-view posed images to directly optimize an implicit representation without pretraining, allowing for improved generalization. However, it remains sensitive to initialization and may experience stability issues. Further developments [42, 47] have expanded scalability to objects with multiple moving parts based on implicit neural radiance fields (NeRFs). Nonetheless, several challenges remain unexplored, including generaliz" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12132" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": "ing to complex scenes, flexibly applied for manipulation, and practical assumptions (e.g. handling unknown camera poses, unaligned static object parts, and occlusion)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 110, + 295, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 110, + 295, + 266 + ], + "spans": [ + { + "bbox": [ + 54, + 110, + 295, + 266 + ], + "type": "text", + "content": "In 3D perception, instance segmentation provides essential object-level understanding for agents interacting within scenes. Recent studies [27, 29] explore part-object segmentation to decompose objects into finer components such as cabinet drawers. However, these methods are preliminary and struggle with affordance detection due to the lack of precise small-scale geometry in existing datasets. Open-vocabulary models [6, 18, 35, 43] utilize foundation models such as CLIP [39] to extract semantic features and consolidate them into a multi-view consistent 3D neural field representation. However, they rely on costly multi-scale queries to refine segmentation boundaries, and face challenges in high-dimensional feature encoding." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 268, + 295, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 268, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 56, + 268, + 295, + 495 + ], + "type": "text", + "content": "To address the limitations mentioned above, we introduce IAAO, a framework designed for interaction of intelligent agents with objects, parts, and functional affordances, as well as for articulating movable objects. We explicitly leverage 3D semantic information to provide our model with scene understanding and interaction capabilities for articulated objects. Key insights of our approach include: 1) semantic scene reconstruction where SAM masks enable multi-view consistent segmentations from partial views, 2) hierarchical feature fields to efficiently compress dense features from multiple views into 3DGS for object- and part-level interaction in diverse scenes, 3) explicit 3D Gaussian representations for affordance localization via query similarity and motion recovery via 3D point-to-2D pixel correspondences, 4) global-local transformation initialization to enhance robustness in realistic scenarios with potential scene misalignment between object states, and 5) scene state fusion merging two different scene states to enhance object geometry completeness." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 498, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 498, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 498, + 295, + 713 + ], + "type": "text", + "content": "Specifically, our IAAO enables interactive affordance detection and object articulation via the following steps: 1) 3D model construction. Using multi-view posed images of each state, we model the 3D scene as a set of explicit 3D Gaussian primitives with 3D Gaussian Splatting (3DGS) [17]. 2) Hierarchical feature field construction. We propose an efficient distillation scheme where shape-aware 2D features and dense semantic features are extracted from large foundation models, including CLIP [39], SAM [19], and DINOv2 [33]. 2D features are efficiently distilled into 3D fields with a decoder. 3) Semantic-guided mask association across states. We first cluster SAM-generated masks from all views within each state to obtain view-consistent mask labels. We then compute 3D proposal features by merging the features from the corresponding 2D masks. By comparing pairwise feature similarities, we establish 3D mask-level correspondences across states that result in consistent mask label sets for the entire scene in each" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": "state. 4) Affordance prediction. Given a task description, we encode it with the CLIP text encoder and compare the text embeddings with the feature fields to identify the relevant affordance. 5) Motion recovery. To estimate transformations, we define a set of consistency and matching losses that recover both global scene-level and local part-level transformations that ensure accurate motion representation across different states." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 326, + 168, + 533, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 168, + 533, + 179 + ], + "spans": [ + { + "bbox": [ + 326, + 168, + 533, + 179 + ], + "type": "text", + "content": "We summarize our main contributions as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 180, + 553, + 371 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 314, + 180, + 553, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 180, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 314, + 180, + 553, + 252 + ], + "type": "text", + "content": "- We introduce IAAO, an interactive affordance system utilizing 3D Gaussian fields embedded with hierarchical language-aligned semantics and class-agnostic masks. This system supports manipulation tasks guided by various prompts (point, mask, and language) at both object and part levels, regardless of object categories and shape." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 253, + 553, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 253, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 253, + 553, + 300 + ], + "type": "text", + "content": "- We propose a method to reconstruct motion via global matching at the scene level and local matching for articulated objects, without relying on impractical assumptions about static object alignment or known camera poses." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 300, + 553, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 553, + 334 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 553, + 334 + ], + "type": "text", + "content": "- Our IAAO achieves state-of-the-art performance across extensive experiments on multiple benchmarks, including synthetic, real-world, and indoor scene data." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 335, + 553, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 553, + 371 + ], + "type": "text", + "content": "- We show strong model generalization to complex indoor environments and previously unseen articulated objects, with no restrictions on the number of movable parts." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 382, + 400, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 382, + 400, + 395 + ], + "spans": [ + { + "bbox": [ + 314, + 382, + 400, + 395 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 402, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 555, + 714 + ], + "type": "text", + "content": "Neural Fields for 3D Scene Understanding. 3D Scene understanding is primarily characterized by four classic representations, including volumetric fields ([34]), point clouds ([40]), 3D meshes ([7]), and depth maps ([11, 41, 54]). Unlike traditional representations, NeRF (Neural Radiance Fields) [28] introduces a neural implicit field to capture the geometry and appearance of a scene. Through a MultiLayer Perceptron (MLP) that takes 3D positions and 2D view directions as inputs, NeRF learns to implicitly represent the color and radiance of the scenes from a collection of posed images. However, a key drawback of NeRF is the slow training and rendering speed. Recently, 3D Gaussian Splatting (3DGS) [17] has been introduced as an alternative to implicit radiance field representations. In contrast to NeRF, 3DGS explicitly represents radiance fields as a collection of oriented 3D Gaussians. Each Gaussian is defined by its spatial position, opacity, and a covariance matrix, which allows flexible optimization. With efficient differentiable rasterization, 3DGS achieves fast rendering with high-quality results. Various work leverage neural fields as a representation robotic manipulation [18, 44, 45, 48, 58, 60]. Methods that leverage visual foundation models to construct neural feature fields [18, 37, 59] are most relevant to our work. With feature distillation via rendering, we reconstruct a hierarchical feature field for object localization and fine-grained affordance prediction." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12133" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 275 + ], + "type": "text", + "content": "Visual Affordance. Affordance [9] refers to the potential interactions that an object or its parts facilitate for an agent. Affordance prediction involves deducing interaction opportunities from visual representations, e.g. images [4, 8, 25] or 3D models [3, 30, 31]. This field has gained significant interest in robotics as a foundational element for tasks e.g. grasping [26, 55], planning [50] and exploration [31]. However, although these methods excel at predicting affordance regions, they lack detailed interaction information, which is essential for engaging with functional elements. Recent work [2] addresses fine-grained functional elements using comprehensive natural language task descriptions for interaction. However, it depends on high-fidelity point clouds, restricting its practical applicability. We use high-resolution images with a Vision-Language Model (VLM) and 3D motion heuristics to predict affordances, eliminating the need for high-fidelity point clouds." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 277, + 295, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 295, + 469 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 295, + 469 + ], + "type": "text", + "content": "Articulation Reasoning by Interaction. As embodied AI continues to evolve, the study of articulated objects has become a critical area of research. Recently, several synthetic [46, 49] and scanned data [15, 24, 27, 36] datasets have been introduced for articulated objects. These datasets contain annotations for part segmentation and motion parameters that enable data-driven approaches in the prediction of motion parameters from 3D point clouds [46, 53]. Recent works have been increasingly focused on real-world scenarios by detecting articulated parts and the motion parameters from images [15, 56] and videos [36]. Interactive perception involves agents gaining new insights from their interactions with the environment, and has been widely used to study object articulations [13, 14, 16, 32]. In this work, we focus on using language, and geometric and semantic scene understanding to achieve articulation reasoning." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 483, + 135, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 135, + 496 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 135, + 496 + ], + "type": "text", + "content": "3. Our Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 506, + 295, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 506, + 295, + 650 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 295, + 650 + ], + "type": "text", + "content": "Objective. Given a scene of articulated objects represented by a set of multi-view posed images " + }, + { + "bbox": [ + 55, + 506, + 295, + 650 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^t = \\{I_i^t\\}_{i=1}^{N^t}" + }, + { + "bbox": [ + 55, + 506, + 295, + 650 + ], + "type": "text", + "content": " at two states " + }, + { + "bbox": [ + 55, + 506, + 295, + 650 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 506, + 295, + 650 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 506, + 295, + 650 + ], + "type": "inline_equation", + "content": "t'" + }, + { + "bbox": [ + 55, + 506, + 295, + 650 + ], + "type": "text", + "content": ", the goal is to reconstruct the geometries and semantics of the objects in the scene. We focus on articulated objects which contain one or more movable parts with rotary or prismatic joints. Under the condition of unknown object categorory in the scene, we aim to recover the object part geometry, segmentation and joint articulations in the presence of occlusions, shape discrepancies and motion changes across different scene states. For static objects in the scene, the task degenerates to global transformation estimation between two input scene states for state matching." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 653, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 715 + ], + "type": "text", + "content": "Overview. Fig. 2 shows our IAAO framework consisting of: 1) A semantic scene reconstruction stage (cf. Sec. 3.1) where we first extract CLIP features for objects in individual scene state. These features are regularized by class-agnostic masks generated from 2D foundation models, e.g." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 276 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 276 + ], + "type": "text", + "content": "SAM [57] at the instance and part level. We then cluster view-inconsistent masks from all views to obtain 3D consistent labels, and build a hierarchical feature field and label field for all masks in each scene state via 3D Gaussian Splatting (3DGS). 2) An affordance and motion prediction stage (cf. Sec. 3.2) where we perform object- and part-level queries by directly operating on 3D primitives represented by 3D Gaussians. Static elements and articulated objects are identified by isolating the groups of 3D Gaussians corresponding to segmented objects and functional parts of the scene, i.e. affordance prediction. For motion recovery, we estimate global transformation from static Gaussian primitives and local articulation parameters from the segmented movable parts of objects. 3) A scene state fusion stage (cf. Sec. 3.3) enabling affordance interaction and object manipulation by merging and refining the reconstructed scenes from two states according to the estimated transformations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 287, + 484, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 287, + 484, + 298 + ], + "spans": [ + { + "bbox": [ + 313, + 287, + 484, + 298 + ], + "type": "text", + "content": "3.1. Semantic Scene Reconstruction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 305, + 555, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 305, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 305, + 555, + 430 + ], + "type": "text", + "content": "Given a multi-view capture " + }, + { + "bbox": [ + 313, + 305, + 555, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^t" + }, + { + "bbox": [ + 313, + 305, + 555, + 430 + ], + "type": "text", + "content": " of scene state " + }, + { + "bbox": [ + 313, + 305, + 555, + 430 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 305, + 555, + 430 + ], + "type": "text", + "content": ", we construct a 3D model using 3DGS initialized with sparse point cloud generated from Structure from Motion (SfM) [40]. Each point serves as the center of a Gaussian primitive embedded with geometry and appearance parameters. These 3D Gaussian primitives " + }, + { + "bbox": [ + 313, + 305, + 555, + 430 + ], + "type": "inline_equation", + "content": "G^{t} = g_{p,p = 1}^{tP^{t}}" + }, + { + "bbox": [ + 313, + 305, + 555, + 430 + ], + "type": "text", + "content": " are then rendered into 2D views via differentiable rasterization for parameter optimization. To enrich the reconstructed model with semantic information, we augment each 3D Gaussian with semantic feature embeddings derived from large foundation models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 434, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 555, + 628 + ], + "type": "text", + "content": "View-Consistent Mask Clustering. For each input image " + }, + { + "bbox": [ + 313, + 434, + 555, + 628 + ], + "type": "inline_equation", + "content": "I_{i}^{t} \\in \\mathcal{I}^{t}" + }, + { + "bbox": [ + 313, + 434, + 555, + 628 + ], + "type": "text", + "content": ", we first derive the class-agnostic masks " + }, + { + "bbox": [ + 313, + 434, + 555, + 628 + ], + "type": "inline_equation", + "content": "M_{i}^{t} = \\{m_{i,n}^{t}\\}_{n=1}^{n_{i}^{t}}" + }, + { + "bbox": [ + 313, + 434, + 555, + 628 + ], + "type": "text", + "content": " with an off-the-shelf mask predictor, where " + }, + { + "bbox": [ + 313, + 434, + 555, + 628 + ], + "type": "inline_equation", + "content": "n_{i}^{t}" + }, + { + "bbox": [ + 313, + 434, + 555, + 628 + ], + "type": "text", + "content": " represents the number of masks in view " + }, + { + "bbox": [ + 313, + 434, + 555, + 628 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 434, + 555, + 628 + ], + "type": "text", + "content": ". For " + }, + { + "bbox": [ + 313, + 434, + 555, + 628 + ], + "type": "inline_equation", + "content": "M_{i}^{t}" + }, + { + "bbox": [ + 313, + 434, + 555, + 628 + ], + "type": "text", + "content": ", we investigated the use of SAM [57] to generate both instance- and part-level masks for arbitrary objects. Unfortunately, mask generation models such as SAM face several challenges: a) class-agnostic masks do not have complete 3D information, b) inconsistent segmentation across images due to variations in viewpoint and appearance, and c) lack of one-to-one correspondence among masks due to oversegmentation for individual object. We aim to generate 3D-consistent labels for all masks in " + }, + { + "bbox": [ + 313, + 434, + 555, + 628 + ], + "type": "inline_equation", + "content": "M^{t}" + }, + { + "bbox": [ + 313, + 434, + 555, + 628 + ], + "type": "text", + "content": " by defining a mask label mapping function " + }, + { + "bbox": [ + 313, + 434, + 555, + 628 + ], + "type": "inline_equation", + "content": "\\phi(M^{t}, G^{t}) \\mapsto O^{t}" + }, + { + "bbox": [ + 313, + 434, + 555, + 628 + ], + "type": "text", + "content": ". This function maps a set of masks " + }, + { + "bbox": [ + 313, + 434, + 555, + 628 + ], + "type": "inline_equation", + "content": "M_{o}^{t}" + }, + { + "bbox": [ + 313, + 434, + 555, + 628 + ], + "type": "text", + "content": " belonging to the same 3D object to a consistent " + }, + { + "bbox": [ + 313, + 434, + 555, + 628 + ], + "type": "inline_equation", + "content": "o^{t} \\in O^{t}" + }, + { + "bbox": [ + 313, + 434, + 555, + 628 + ], + "type": "text", + "content": " within the 3D Gaussian primitives " + }, + { + "bbox": [ + 313, + 434, + 555, + 628 + ], + "type": "inline_equation", + "content": "G^{t}" + }, + { + "bbox": [ + 313, + 434, + 555, + 628 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "To associate class-agnostic masks across all input views at scene state " + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": ", we construct a mask graph to fuse the 2D masks " + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "inline_equation", + "content": "M^t" + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": " of the entire scene into cohesive 3D instances " + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "inline_equation", + "content": "O^t" + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": ". In the initial graph " + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_0^t = (\\mathcal{V}_0^t, \\mathcal{E}_0^t)" + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": ", each node in " + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_0^t" + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": " represents a mask detected from the views, and each edge in " + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_0^t" + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": " represents potential matches between pairs of object parts. Following the mask clustering workflow from [52]," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12134" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 87, + 451, + 167 + ], + "blocks": [ + { + "bbox": [ + 63, + 77, + 180, + 85 + ], + "lines": [ + { + "bbox": [ + 63, + 77, + 180, + 85 + ], + "spans": [ + { + "bbox": [ + 63, + 77, + 180, + 85 + ], + "type": "text", + "content": "Stage 1: Semantic Scene Reconstruction" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 61, + 87, + 451, + 167 + ], + "lines": [ + { + "bbox": [ + 61, + 87, + 451, + 167 + ], + "spans": [ + { + "bbox": [ + 61, + 87, + 451, + 167 + ], + "type": "image", + "image_path": "adec8629e5093ea4f04cfe7247ed41f44f0f7587894a7e590e29ce644fe54d84.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 461, + 93, + 548, + 259 + ], + "blocks": [ + { + "bbox": [ + 462, + 80, + 542, + 87 + ], + "lines": [ + { + "bbox": [ + 462, + 80, + 542, + 87 + ], + "spans": [ + { + "bbox": [ + 462, + 80, + 542, + 87 + ], + "type": "text", + "content": "Stage 3: Scene State Fusion" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 461, + 93, + 548, + 259 + ], + "lines": [ + { + "bbox": [ + 461, + 93, + 548, + 259 + ], + "spans": [ + { + "bbox": [ + 461, + 93, + 548, + 259 + ], + "type": "image", + "image_path": "d03463a4d5b791fc4f27e28523e5f8f7ae47420cf69378f4106587d1870a5935.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 271, + 555, + 338 + ], + "lines": [ + { + "bbox": [ + 55, + 271, + 555, + 338 + ], + "spans": [ + { + "bbox": [ + 55, + 271, + 555, + 338 + ], + "type": "text", + "content": "Figure 2. Our IAAO framework. 1) Top: Constructing 3D Gaussian fields in each state. We optimize 3DGS fields with hierarchical mask features, DINOv2 features and 3D-consistent mask labels generated from multi-view images. We also incorporate geometry information from depth images into the 3D Gaussians. 2) Bottom: Affordance and motion prediction. A query prompt is embedded using a pretrained encoder to localize relevant regions in the 3D Gaussians. For motion prediction, we optimize the transformation parameters by applying consistency and matching losses to 2D-3D correspondences between states. 3) Right: Scene fusion. Using the estimated transformations, we merge reconstructed 3DGS models from both states, aligning static and articulated elements." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 61, + 184, + 454, + 261 + ], + "blocks": [ + { + "bbox": [ + 63, + 173, + 189, + 182 + ], + "lines": [ + { + "bbox": [ + 63, + 173, + 189, + 182 + ], + "spans": [ + { + "bbox": [ + 63, + 173, + 189, + 182 + ], + "type": "text", + "content": "Stage 2: Affordance and Motion Prediction" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 61, + 184, + 454, + 261 + ], + "lines": [ + { + "bbox": [ + 61, + 184, + 454, + 261 + ], + "spans": [ + { + "bbox": [ + 61, + 184, + 454, + 261 + ], + "type": "image", + "image_path": "82cd04537932a936e0cf0481341cf4e78840e1c4733b5c9ea5d629700c38e946.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 358, + 296, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 358, + 296, + 563 + ], + "spans": [ + { + "bbox": [ + 56, + 358, + 296, + 563 + ], + "type": "text", + "content": "we then cluster the mask nodes and update the edges to facilitate mask association across different views. For each cluster, we combine the corresponding partial sets of Gaussians from individual masks to form a complete 3D instance and establish correspondences between 2D masks and 3D nodes. This results in the final graph " + }, + { + "bbox": [ + 56, + 358, + 296, + 563 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^t = (\\mathcal{V}^t,\\mathcal{E}^t)" + }, + { + "bbox": [ + 56, + 358, + 296, + 563 + ], + "type": "text", + "content": ", where each node in " + }, + { + "bbox": [ + 56, + 358, + 296, + 563 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^t" + }, + { + "bbox": [ + 56, + 358, + 296, + 563 + ], + "type": "text", + "content": " represents a 3D instance proposal in " + }, + { + "bbox": [ + 56, + 358, + 296, + 563 + ], + "type": "inline_equation", + "content": "O^t" + }, + { + "bbox": [ + 56, + 358, + 296, + 563 + ], + "type": "text", + "content": " in scene state " + }, + { + "bbox": [ + 56, + 358, + 296, + 563 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 56, + 358, + 296, + 563 + ], + "type": "text", + "content": ". We maintain a list of associated 2D masks " + }, + { + "bbox": [ + 56, + 358, + 296, + 563 + ], + "type": "inline_equation", + "content": "M_o^t" + }, + { + "bbox": [ + 56, + 358, + 296, + 563 + ], + "type": "text", + "content": " for each " + }, + { + "bbox": [ + 56, + 358, + 296, + 563 + ], + "type": "inline_equation", + "content": "o^t" + }, + { + "bbox": [ + 56, + 358, + 296, + 563 + ], + "type": "text", + "content": ", discarding severely occluded masks. This 2D-3D relationship enables consistent segmentation labels across all views. However, the mask label mapping " + }, + { + "bbox": [ + 56, + 358, + 296, + 563 + ], + "type": "inline_equation", + "content": "\\phi (M^{t},G^{t})" + }, + { + "bbox": [ + 56, + 358, + 296, + 563 + ], + "type": "text", + "content": " may be affected by noise due to varying part segmentation hierarchies across views. To mitigate segmentation ambiguities, we use the labeled masks " + }, + { + "bbox": [ + 56, + 358, + 296, + 563 + ], + "type": "inline_equation", + "content": "M_o^t" + }, + { + "bbox": [ + 56, + 358, + 296, + 563 + ], + "type": "text", + "content": " in the final graph to supervise a label field for object part " + }, + { + "bbox": [ + 56, + 358, + 296, + 563 + ], + "type": "inline_equation", + "content": "o^t" + }, + { + "bbox": [ + 56, + 358, + 296, + 563 + ], + "type": "text", + "content": " in the given scene state. To this end, we use a cross-entropy classification loss " + }, + { + "bbox": [ + 56, + 358, + 296, + 563 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{label}}" + }, + { + "bbox": [ + 56, + 358, + 296, + 563 + ], + "type": "text", + "content": " for enhanced consistency." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": "Per-State Hierarchical Feature Field. We generate a set of feature embeddings from the given image " + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "I_{i}^{t}" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": " using pretrained foundation models. Specifically, we utilize MaskCLIP [5] to produce mask-level CLIP features at both instance and part levels, and DINOv2 [1, 33] to extract DINO embeddings. To enable object- and part-level interaction, we construct a hierarchical feature field by embedding dense features from multi-view images into the 3D Gaussian primitives. Using semantic hierarchy from SAM, we enhance our 3D semantic field accuracy while making the querying process more efficient. Directly embedding high-dimensional 2D features would result in excess" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 358, + 555, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 358, + 555, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 358, + 555, + 418 + ], + "type": "text", + "content": "sive computational and memory costs. Therefore, we build a low-dimensional latent feature field and introduce a decoder to project the rendered features " + }, + { + "bbox": [ + 313, + 358, + 555, + 418 + ], + "type": "inline_equation", + "content": "\\hat{F}^t" + }, + { + "bbox": [ + 313, + 358, + 555, + 418 + ], + "type": "text", + "content": " back to the 2D feature space " + }, + { + "bbox": [ + 313, + 358, + 555, + 418 + ], + "type": "inline_equation", + "content": "F^t" + }, + { + "bbox": [ + 313, + 358, + 555, + 418 + ], + "type": "text", + "content": " during the differentiable rasterization optimization process as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 361, + 426, + 555, + 442 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 426, + 555, + 442 + ], + "spans": [ + { + "bbox": [ + 361, + 426, + 555, + 442 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {f e a t} = \\left\\| \\mathcal {D} \\left(\\hat {F} ^ {t} \\left(I _ {i} ^ {t}\\right)\\right) - F ^ {t} \\left(I _ {i} ^ {t}\\right) \\right\\| _ {2} ^ {2}. \\tag {1}", + "image_path": "dadd43ea11f74dc869cc9277af9f51aa761e7423169804795662ae2c624a28da.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 449, + 554, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 554, + 484 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 554, + 484 + ], + "type": "text", + "content": "The decoder " + }, + { + "bbox": [ + 313, + 449, + 554, + 484 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 449, + 554, + 484 + ], + "type": "text", + "content": " is implemented as a small MLP with three output branches: two for instance- and part-level CLIP features, and one for DINO features." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 485, + 555, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 555, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 555, + 641 + ], + "type": "text", + "content": "Semantic Association between Scene States. Upon obtaining the mask label mapping functions " + }, + { + "bbox": [ + 313, + 485, + 555, + 641 + ], + "type": "inline_equation", + "content": "O^t" + }, + { + "bbox": [ + 313, + 485, + 555, + 641 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 485, + 555, + 641 + ], + "type": "inline_equation", + "content": "O^{t'}" + }, + { + "bbox": [ + 313, + 485, + 555, + 641 + ], + "type": "text", + "content": " for scene states " + }, + { + "bbox": [ + 313, + 485, + 555, + 641 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 485, + 555, + 641 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 485, + 555, + 641 + ], + "type": "inline_equation", + "content": "t'" + }, + { + "bbox": [ + 313, + 485, + 555, + 641 + ], + "type": "text", + "content": ", we establish a global 3D mask association which links the same semantic regions across both states into a unified 3D mask using their semantic features. For each 3D proposal in " + }, + { + "bbox": [ + 313, + 485, + 555, + 641 + ], + "type": "inline_equation", + "content": "O^t" + }, + { + "bbox": [ + 313, + 485, + 555, + 641 + ], + "type": "text", + "content": ", we select the top- " + }, + { + "bbox": [ + 313, + 485, + 555, + 641 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 485, + 555, + 641 + ], + "type": "text", + "content": " representative masks from " + }, + { + "bbox": [ + 313, + 485, + 555, + 641 + ], + "type": "inline_equation", + "content": "M_o^t" + }, + { + "bbox": [ + 313, + 485, + 555, + 641 + ], + "type": "text", + "content": " and aggregate their rendered features to create the corresponding semantic primitives " + }, + { + "bbox": [ + 313, + 485, + 555, + 641 + ], + "type": "inline_equation", + "content": "F_o^t" + }, + { + "bbox": [ + 313, + 485, + 555, + 641 + ], + "type": "text", + "content": ". We then build an affinity matrix capturing the cosine similarity between all 3D proposal pairs " + }, + { + "bbox": [ + 313, + 485, + 555, + 641 + ], + "type": "inline_equation", + "content": "(O^t, O^{t'})" + }, + { + "bbox": [ + 313, + 485, + 555, + 641 + ], + "type": "text", + "content": " using their primitives " + }, + { + "bbox": [ + 313, + 485, + 555, + 641 + ], + "type": "inline_equation", + "content": "(F_O^t, F_O^{t'})" + }, + { + "bbox": [ + 313, + 485, + 555, + 641 + ], + "type": "text", + "content": " from two scene states. For each query proposal " + }, + { + "bbox": [ + 313, + 485, + 555, + 641 + ], + "type": "inline_equation", + "content": "o^t \\in O^t" + }, + { + "bbox": [ + 313, + 485, + 555, + 641 + ], + "type": "text", + "content": ", we define its visual match as the target proposal in " + }, + { + "bbox": [ + 313, + 485, + 555, + 641 + ], + "type": "inline_equation", + "content": "O^{t'}" + }, + { + "bbox": [ + 313, + 485, + 555, + 641 + ], + "type": "text", + "content": " with the highest similarity." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 648, + 499, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 648, + 499, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 648, + 499, + 659 + ], + "type": "text", + "content": "3.2. Affordance and Motion Prediction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": "Unlike NeRF-based methods [6, 18, 23] which require resource-intensive rendering to derive language-embedded features and geometry from implicit MLPs, our IAAO directly uses explicit Gaussian primitives for efficient object" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12135" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "part localization and manipulation. Given the reconstructed 3D scene " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "G^{t} = \\{g_{p}^{t}\\}_{p = 1}^{P^{t}}" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": " at scene state " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": ", where each Gaussian point " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "g_{p}^{t}" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": " is embedded with geometric information (coordinates of points " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "p^t" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": ") and additional features (color and semantic embeddings), our goal is to enable scene-aware affordance prediction and object articulation estimation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 147, + 295, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 147, + 295, + 337 + ], + "spans": [ + { + "bbox": [ + 55, + 147, + 295, + 337 + ], + "type": "text", + "content": "Functional Affordance Prediction. Functional affordances are defined as parts of the scene that facilitate interactions for agents to perform specific tasks. We utilize the reconstructed hierarchical feature and label fields as the 3D representations to predict the masks " + }, + { + "bbox": [ + 55, + 147, + 295, + 337 + ], + "type": "inline_equation", + "content": "\\{m_a\\}_{a = 1}^{K_a}" + }, + { + "bbox": [ + 55, + 147, + 295, + 337 + ], + "type": "text", + "content": " and affordance labels " + }, + { + "bbox": [ + 55, + 147, + 295, + 337 + ], + "type": "inline_equation", + "content": "\\{l_a\\}_{a = 1}^{K_a}" + }, + { + "bbox": [ + 55, + 147, + 295, + 337 + ], + "type": "text", + "content": " in the scene, where " + }, + { + "bbox": [ + 55, + 147, + 295, + 337 + ], + "type": "inline_equation", + "content": "K_{a}" + }, + { + "bbox": [ + 55, + 147, + 295, + 337 + ], + "type": "text", + "content": " is the total number of affordance categories. We first build associations between affordance labels and our reconstructed mask label field, and then infer the point-level affordance of a functional element in the scene. Additionally, for task-specific reasoning, we encode the task description using the CLIP text encoder. To identify the relevant mask for a given description, we calculate the similarity between the embeddings in the feature field and the query embeddings. A mask is retrieved if its similarity score with the description text exceeds a predefined threshold." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 340, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 340, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 340, + 295, + 628 + ], + "type": "text", + "content": "Global and Local Motion Recovery. Given the reconstructed scenes with different object articulations at two states, our objective is to estimate scene-aware transformation and 3D motion primitives for articulated parts. Our approach segments the reconstructed Gaussians " + }, + { + "bbox": [ + 55, + 340, + 295, + 628 + ], + "type": "inline_equation", + "content": "G^{t}" + }, + { + "bbox": [ + 55, + 340, + 295, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 340, + 295, + 628 + ], + "type": "inline_equation", + "content": "G^{t'}" + }, + { + "bbox": [ + 55, + 340, + 295, + 628 + ], + "type": "text", + "content": " into static scene elements and articulated objects via object-part querying for global and per-part local transformation estimation, respectively. 1) Object-level querying in complex scenes: In an indoor scene with a complex background and multiple objects, we start with object-level querying. Positive language queries are used to identify target objects for manipulation, while optional negative queries exclude irrelevant items. This selection process identifies 3D Gaussians with object-level CLIP features that align more closely with the positive queries than the negative ones, following standard CLIP feature-text comparison practices [45]. 2) Part-level querying for targeted objects: For individual objects in the scene, we apply part-level querying to focus on 3D Gaussians within the object. This allows for fine-grained transformation estimation specific to each articulated part. In addition to language, other user input prompts such as masks and points can also be used. Nonetheless, we do not further elaborate on these prompts as they can be easily obtained by a user clicking on the images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "content": "We define the global scene-aware transformation between two different scene states as " + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\xi_g^t = (s_g^t,R_g^t,T_g^t)" + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "inline_equation", + "content": "s_g^t\\in \\mathbb{R}" + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "content": " represents the scale factor, " + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "inline_equation", + "content": "R_{g}^{t}\\in \\mathrm{SO}(3)" + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "content": " is the rotation matrix and " + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "inline_equation", + "content": "T_{g}^{t}\\in \\mathbb{R}^{3}" + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "content": " is the translation vector. To initialize the transformation, we estimate an initial alignment using 3D Gaussians derived from static elements in the scene, excluding articulated objects. From the Gaussian" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": "points in both scene states, we select reliable points with opacity values above a threshold (empirically set at 0.7) to serve as input point sets for the coarse registration. Using the GeoTransformer [38], we calculate the scene alignment transformation by matching coarse superpoints and refining with dense point correspondences. Since these point clouds are based on reconstructed 3D Gaussians, the resulting transformation is coarse, providing an initial approximation for " + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "inline_equation", + "content": "\\xi_{g}^{t}" + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": " that can be further refined." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 180, + 555, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 180, + 555, + 314 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 555, + 314 + ], + "type": "text", + "content": "After achieving global alignment between two scene states, we define the local transformation for each articulated part as " + }, + { + "bbox": [ + 313, + 180, + 555, + 314 + ], + "type": "inline_equation", + "content": "\\xi_{o}^{t} = (R_{o}^{t},T_{o}^{t})" + }, + { + "bbox": [ + 313, + 180, + 555, + 314 + ], + "type": "text", + "content": ", representing the movement of part " + }, + { + "bbox": [ + 313, + 180, + 555, + 314 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 313, + 180, + 555, + 314 + ], + "type": "text", + "content": " from state " + }, + { + "bbox": [ + 313, + 180, + 555, + 314 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 180, + 555, + 314 + ], + "type": "text", + "content": " to state " + }, + { + "bbox": [ + 313, + 180, + 555, + 314 + ], + "type": "inline_equation", + "content": "t'" + }, + { + "bbox": [ + 313, + 180, + 555, + 314 + ], + "type": "text", + "content": ". All points within a part are assumed to follow the same motion parameters. The reverse mapping, which transforms part " + }, + { + "bbox": [ + 313, + 180, + 555, + 314 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 313, + 180, + 555, + 314 + ], + "type": "text", + "content": " from state " + }, + { + "bbox": [ + 313, + 180, + 555, + 314 + ], + "type": "inline_equation", + "content": "t'" + }, + { + "bbox": [ + 313, + 180, + 555, + 314 + ], + "type": "text", + "content": " back to state " + }, + { + "bbox": [ + 313, + 180, + 555, + 314 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 180, + 555, + 314 + ], + "type": "text", + "content": ", is given by the inverse transformation, " + }, + { + "bbox": [ + 313, + 180, + 555, + 314 + ], + "type": "inline_equation", + "content": "\\xi_{o}^{t'} = (R_{o}^{t},T_{o}^{t})^{-1}" + }, + { + "bbox": [ + 313, + 180, + 555, + 314 + ], + "type": "text", + "content": ". To facilitate this process, we construct a 3D Gaussian correspondence field, allowing each Gaussian primitive " + }, + { + "bbox": [ + 313, + 180, + 555, + 314 + ], + "type": "inline_equation", + "content": "g_{p}^{t}(o)" + }, + { + "bbox": [ + 313, + 180, + 555, + 314 + ], + "type": "text", + "content": " at state " + }, + { + "bbox": [ + 313, + 180, + 555, + 314 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 180, + 555, + 314 + ], + "type": "text", + "content": " to be repositioned to its corresponding location " + }, + { + "bbox": [ + 313, + 180, + 555, + 314 + ], + "type": "inline_equation", + "content": "p^{t'}" + }, + { + "bbox": [ + 313, + 180, + 555, + 314 + ], + "type": "text", + "content": " at state " + }, + { + "bbox": [ + 313, + 180, + 555, + 314 + ], + "type": "inline_equation", + "content": "t'" + }, + { + "bbox": [ + 313, + 180, + 555, + 314 + ], + "type": "text", + "content": " based on the motion parameters of part " + }, + { + "bbox": [ + 313, + 180, + 555, + 314 + ], + "type": "inline_equation", + "content": "o^t" + }, + { + "bbox": [ + 313, + 180, + 555, + 314 + ], + "type": "text", + "content": ". Formally:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 319, + 553, + 358 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 319, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 314, + 319, + 553, + 358 + ], + "type": "interline_equation", + "content": "p ^ {t \\rightarrow t ^ {\\prime}} = \\left\\{\\begin{array}{l l}f \\left(p ^ {t}, \\xi_ {g} ^ {t}\\right) = R _ {g} ^ {t} p ^ {t} + T _ {g} ^ {t},&\\text {i f} p \\text {i n s t a t i c s c e n e ,}\\\\f \\left(f \\left(p ^ {t}, \\xi_ {g} ^ {t}\\right), \\xi_ {o} ^ {t}\\right),&\\text {o t h e r w i s e .}\\end{array}\\right. \\tag {2}", + "image_path": "91610d4b1365d2591d9a1a65da5b68340240e16ff57c00e70552aea0f8f25e33.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 365, + 553, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 553, + 402 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 553, + 402 + ], + "type": "text", + "content": "The same rule applies to the backward transformation function " + }, + { + "bbox": [ + 313, + 365, + 553, + 402 + ], + "type": "inline_equation", + "content": "f^{-1}" + }, + { + "bbox": [ + 313, + 365, + 553, + 402 + ], + "type": "text", + "content": " for obtaining " + }, + { + "bbox": [ + 313, + 365, + 553, + 402 + ], + "type": "inline_equation", + "content": "p^{t' \\to t}" + }, + { + "bbox": [ + 313, + 365, + 553, + 402 + ], + "type": "text", + "content": " (i.e. mapping from " + }, + { + "bbox": [ + 313, + 365, + 553, + 402 + ], + "type": "inline_equation", + "content": "t'" + }, + { + "bbox": [ + 313, + 365, + 553, + 402 + ], + "type": "text", + "content": " back to " + }, + { + "bbox": [ + 313, + 365, + 553, + 402 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 365, + 553, + 402 + ], + "type": "text", + "content": ") with the transformations " + }, + { + "bbox": [ + 313, + 365, + 553, + 402 + ], + "type": "inline_equation", + "content": "\\xi_g^{t'}" + }, + { + "bbox": [ + 313, + 365, + 553, + 402 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 365, + 553, + 402 + ], + "type": "inline_equation", + "content": "\\xi_o^{t'}" + }, + { + "bbox": [ + 313, + 365, + 553, + 402 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 402, + 555, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 555, + 606 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 555, + 606 + ], + "type": "text", + "content": "Establishing accurate dense correspondences between point clouds in two states is challenging due to the sparsity and noise in 3D points derived from the 3DGS model. To optimize the correspondence field, we instead match 3D Gaussian primitives to 2D pixels across the two states, guided by part geometry and segmentation information. Specifically, for a Gaussian primitive " + }, + { + "bbox": [ + 313, + 402, + 555, + 606 + ], + "type": "inline_equation", + "content": "g_{p}^{t}(o)" + }, + { + "bbox": [ + 313, + 402, + 555, + 606 + ], + "type": "text", + "content": " of part " + }, + { + "bbox": [ + 313, + 402, + 555, + 606 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 313, + 402, + 555, + 606 + ], + "type": "text", + "content": " at state " + }, + { + "bbox": [ + 313, + 402, + 555, + 606 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 402, + 555, + 606 + ], + "type": "text", + "content": ", the corresponding 2D pixels at state " + }, + { + "bbox": [ + 313, + 402, + 555, + 606 + ], + "type": "inline_equation", + "content": "t'" + }, + { + "bbox": [ + 313, + 402, + 555, + 606 + ], + "type": "text", + "content": " can be treated as observations from novel views. To enhance consistency, we incorporate mask-level features alongside the rendered image consistency loss " + }, + { + "bbox": [ + 313, + 402, + 555, + 606 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rgb}" + }, + { + "bbox": [ + 313, + 402, + 555, + 606 + ], + "type": "text", + "content": ". For a 3D proposal " + }, + { + "bbox": [ + 313, + 402, + 555, + 606 + ], + "type": "inline_equation", + "content": "o^t" + }, + { + "bbox": [ + 313, + 402, + 555, + 606 + ], + "type": "text", + "content": " in graph " + }, + { + "bbox": [ + 313, + 402, + 555, + 606 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^t" + }, + { + "bbox": [ + 313, + 402, + 555, + 606 + ], + "type": "text", + "content": ", we identify the part mask list " + }, + { + "bbox": [ + 313, + 402, + 555, + 606 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_o^{t'}" + }, + { + "bbox": [ + 313, + 402, + 555, + 606 + ], + "type": "text", + "content": " associated with the corresponding proposal " + }, + { + "bbox": [ + 313, + 402, + 555, + 606 + ], + "type": "inline_equation", + "content": "o^{t'}" + }, + { + "bbox": [ + 313, + 402, + 555, + 606 + ], + "type": "text", + "content": " at state " + }, + { + "bbox": [ + 313, + 402, + 555, + 606 + ], + "type": "inline_equation", + "content": "t'" + }, + { + "bbox": [ + 313, + 402, + 555, + 606 + ], + "type": "text", + "content": ". Following the splatting process, we render the feature field of " + }, + { + "bbox": [ + 313, + 402, + 555, + 606 + ], + "type": "inline_equation", + "content": "o^t" + }, + { + "bbox": [ + 313, + 402, + 555, + 606 + ], + "type": "text", + "content": " to training view " + }, + { + "bbox": [ + 313, + 402, + 555, + 606 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 402, + 555, + 606 + ], + "type": "text", + "content": " in target scene state " + }, + { + "bbox": [ + 313, + 402, + 555, + 606 + ], + "type": "inline_equation", + "content": "t'" + }, + { + "bbox": [ + 313, + 402, + 555, + 606 + ], + "type": "text", + "content": " and decode the mask-level features by " + }, + { + "bbox": [ + 313, + 402, + 555, + 606 + ], + "type": "inline_equation", + "content": "\\mathcal{D}(\\hat{F}_o^t (I_n^{t'}))" + }, + { + "bbox": [ + 313, + 402, + 555, + 606 + ], + "type": "text", + "content": ". We formulate the mask feature loss for part " + }, + { + "bbox": [ + 313, + 402, + 555, + 606 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 313, + 402, + 555, + 606 + ], + "type": "text", + "content": " transformed from state " + }, + { + "bbox": [ + 313, + 402, + 555, + 606 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 402, + 555, + 606 + ], + "type": "text", + "content": " to target state " + }, + { + "bbox": [ + 313, + 402, + 555, + 606 + ], + "type": "inline_equation", + "content": "t'" + }, + { + "bbox": [ + 313, + 402, + 555, + 606 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 332, + 612, + 553, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 612, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 332, + 612, + 553, + 658 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {m a s k} (o ^ {t \\rightarrow t ^ {\\prime}}) = \\\\ \\Sigma_ {n} ^ {N _ {o} ^ {t ^ {\\prime}}} \\frac {M _ {o} ^ {t ^ {\\prime}} \\left(I _ {n} ^ {t ^ {\\prime}}\\right)}{\\left| M _ {o} ^ {t ^ {\\prime}} \\left(I _ {n} ^ {t ^ {\\prime}}\\right) \\right|} \\cdot \\| \\mathcal {D} \\left(\\hat {F} _ {o} ^ {t} \\left(I _ {n} ^ {t ^ {\\prime}}\\right)\\right) - F _ {o} ^ {t ^ {\\prime}} \\left(I _ {n} ^ {t ^ {\\prime}}\\right) \\| _ {2} ^ {2}, \\tag {3} \\\\ \\end{array}", + "image_path": "f13908c36901a8ac7cb156c776979f63e6617e4ff410bd7111b9a8ce67d590e5.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 664, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 664, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 664, + 553, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 664, + 553, + 713 + ], + "type": "inline_equation", + "content": "M_{o}^{t^{\\prime}}(I_{n}^{t^{\\prime}})" + }, + { + "bbox": [ + 313, + 664, + 553, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 664, + 553, + 713 + ], + "type": "inline_equation", + "content": "F_{o}^{t^{\\prime}}(I_{n}^{t^{\\prime}})" + }, + { + "bbox": [ + 313, + 664, + 553, + 713 + ], + "type": "text", + "content": " represent the corresponding part mask and 2D mask-level features of part " + }, + { + "bbox": [ + 313, + 664, + 553, + 713 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 313, + 664, + 553, + 713 + ], + "type": "text", + "content": " in view " + }, + { + "bbox": [ + 313, + 664, + 553, + 713 + ], + "type": "inline_equation", + "content": "I_{n}^{t^{\\prime}}" + }, + { + "bbox": [ + 313, + 664, + 553, + 713 + ], + "type": "text", + "content": " at the target state " + }, + { + "bbox": [ + 313, + 664, + 553, + 713 + ], + "type": "inline_equation", + "content": "t^{\\prime}" + }, + { + "bbox": [ + 313, + 664, + 553, + 713 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 664, + 553, + 713 + ], + "type": "inline_equation", + "content": "N_{o}^{t}" + }, + { + "bbox": [ + 313, + 664, + 553, + 713 + ], + "type": "text", + "content": " denotes the number of part masks for proposal " + }, + { + "bbox": [ + 313, + 664, + 553, + 713 + ], + "type": "inline_equation", + "content": "o^t" + }, + { + "bbox": [ + 313, + 664, + 553, + 713 + ], + "type": "text", + "content": " across all training views." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12136" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 262 + ], + "type": "text", + "content": "Since the mask-level loss provides only coarse guidance for transformation, we further propose to find dense correspondences between the 3D Gaussian primitives and 2D pixels by comparing their DINO features. We begin by computing the feature similarity matrix " + }, + { + "bbox": [ + 55, + 72, + 296, + 262 + ], + "type": "inline_equation", + "content": "\\alpha_{p\\rightarrow o}(I_n^{t'})" + }, + { + "bbox": [ + 55, + 72, + 296, + 262 + ], + "type": "text", + "content": " between each pixel in the part mask " + }, + { + "bbox": [ + 55, + 72, + 296, + 262 + ], + "type": "inline_equation", + "content": "M_o^{t'}(I_n^{t'})" + }, + { + "bbox": [ + 55, + 72, + 296, + 262 + ], + "type": "text", + "content": " in view " + }, + { + "bbox": [ + 55, + 72, + 296, + 262 + ], + "type": "inline_equation", + "content": "I_n^{t'}" + }, + { + "bbox": [ + 55, + 72, + 296, + 262 + ], + "type": "text", + "content": " and the sampled Gaussian point " + }, + { + "bbox": [ + 55, + 72, + 296, + 262 + ], + "type": "inline_equation", + "content": "g_p^t" + }, + { + "bbox": [ + 55, + 72, + 296, + 262 + ], + "type": "text", + "content": " from 3D proposal " + }, + { + "bbox": [ + 55, + 72, + 296, + 262 + ], + "type": "inline_equation", + "content": "o^t" + }, + { + "bbox": [ + 55, + 72, + 296, + 262 + ], + "type": "text", + "content": ". We then normalize the similarity matrix " + }, + { + "bbox": [ + 55, + 72, + 296, + 262 + ], + "type": "inline_equation", + "content": "\\alpha_{p\\rightarrow o}(I_n^{t'})" + }, + { + "bbox": [ + 55, + 72, + 296, + 262 + ], + "type": "text", + "content": " using a softmax over the entire mask to obtain the weight " + }, + { + "bbox": [ + 55, + 72, + 296, + 262 + ], + "type": "inline_equation", + "content": "\\beta_{p\\rightarrow o}(I_n^{t'})" + }, + { + "bbox": [ + 55, + 72, + 296, + 262 + ], + "type": "text", + "content": ". Finally, we identify the corresponding 2D pixel " + }, + { + "bbox": [ + 55, + 72, + 296, + 262 + ], + "type": "inline_equation", + "content": "s_{p\\rightarrow o}^{t'}(I_n^{t'})" + }, + { + "bbox": [ + 55, + 72, + 296, + 262 + ], + "type": "text", + "content": " for the sampled 3D point " + }, + { + "bbox": [ + 55, + 72, + 296, + 262 + ], + "type": "inline_equation", + "content": "p^t" + }, + { + "bbox": [ + 55, + 72, + 296, + 262 + ], + "type": "text", + "content": " using a weighted sum. See the supplementary material for details on the computation. For each point-pixel pair " + }, + { + "bbox": [ + 55, + 72, + 296, + 262 + ], + "type": "inline_equation", + "content": "(p^t,s_{p\\rightarrow o}^{t'}(I_n^{t'}))" + }, + { + "bbox": [ + 55, + 72, + 296, + 262 + ], + "type": "text", + "content": ", we filter out pairs that collide by checking whether " + }, + { + "bbox": [ + 55, + 72, + 296, + 262 + ], + "type": "inline_equation", + "content": "s_{p\\rightarrow o}^{t'}(I_n^{t'})" + }, + { + "bbox": [ + 55, + 72, + 296, + 262 + ], + "type": "text", + "content": " falls within the same part mask " + }, + { + "bbox": [ + 55, + 72, + 296, + 262 + ], + "type": "inline_equation", + "content": "o^{t'}" + }, + { + "bbox": [ + 55, + 72, + 296, + 262 + ], + "type": "text", + "content": ". We then define the matching loss as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 94, + 271, + 295, + 287 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 271, + 295, + 287 + ], + "spans": [ + { + "bbox": [ + 94, + 271, + 295, + 287 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {m a t c h}} = \\left\\| \\pi_ {n} ^ {t ^ {\\prime}} \\left(p ^ {t \\rightarrow t ^ {\\prime}}\\right) - s _ {p \\rightarrow n} ^ {t ^ {\\prime}} \\left(I _ {n} ^ {t ^ {\\prime}}\\right)\\right\\| _ {2} ^ {2}, \\tag {4}", + "image_path": "bb44dc51e439efb4b25bdf3febeb6ba2cfdf744c8309a5a19130fbb147d5ac0b.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 296, + 295, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 296, + 295, + 321 + ], + "spans": [ + { + "bbox": [ + 55, + 296, + 295, + 321 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 296, + 295, + 321 + ], + "type": "inline_equation", + "content": "\\pi_n^{t'}" + }, + { + "bbox": [ + 55, + 296, + 295, + 321 + ], + "type": "text", + "content": " represents the projection of 3D point " + }, + { + "bbox": [ + 55, + 296, + 295, + 321 + ], + "type": "inline_equation", + "content": "p^{t\\rightarrow t'}" + }, + { + "bbox": [ + 55, + 296, + 295, + 321 + ], + "type": "text", + "content": " to view " + }, + { + "bbox": [ + 55, + 296, + 295, + 321 + ], + "type": "inline_equation", + "content": "I_{n}^{t^{\\prime}}" + }, + { + "bbox": [ + 55, + 296, + 295, + 321 + ], + "type": "text", + "content": " in scene state " + }, + { + "bbox": [ + 55, + 296, + 295, + 321 + ], + "type": "inline_equation", + "content": "t^\\prime" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 322, + 296, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 322, + 296, + 381 + ], + "spans": [ + { + "bbox": [ + 55, + 322, + 296, + 381 + ], + "type": "text", + "content": "**Optimization.** For the first stage, each scene is reconstructed independently at two different states. We optionally use depth map for further geometry regularization. In the second stage, we reconstruct motion of global scene and local articulated objects. The total loss is defined as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 391, + 295, + 405 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 391, + 295, + 405 + ], + "spans": [ + { + "bbox": [ + 61, + 391, + 295, + 405 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda_ {\\text {c o n s}} \\left(\\mathcal {L} _ {\\text {r g b}} + \\mathcal {L} _ {\\text {m a s k}} + \\mathcal {L} _ {\\text {l a b e l}}\\right) + \\lambda_ {\\text {m a t c h}} \\mathcal {L} _ {\\text {m a t c h}}. \\tag {5}", + "image_path": "39f64a6ce140f1f3672f21733a948682808653279a5af24c974ccf5c97ec5b61.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 413, + 296, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 413, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 296, + 437 + ], + "type": "inline_equation", + "content": "\\lambda_{\\text{cons}}" + }, + { + "bbox": [ + 55, + 413, + 296, + 437 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 413, + 296, + 437 + ], + "type": "inline_equation", + "content": "\\lambda_{\\text{match}}" + }, + { + "bbox": [ + 55, + 413, + 296, + 437 + ], + "type": "text", + "content": " denote the balancing weights of consistency loss and matching loss." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 445, + 167, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 167, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 167, + 456 + ], + "type": "text", + "content": "3.3. Scene State Fusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 462, + 296, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 296, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 296, + 568 + ], + "type": "text", + "content": "After training, we merge the 3DGS models from the two states using the estimated transformations to fill in the occluded regions in each state. Details of our fusion and filtering strategy are provided in the supplementary material. The combined 3D Gaussians facilitate object manipulation by addressing potential artifacts that may appear in occluded areas of the reconstructed scene. Optional fine-tuning can be applied to either the edited regions or the entire scene for further refinement." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 581, + 132, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 581, + 132, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 581, + 132, + 594 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 600, + 172, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 172, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 172, + 613 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "content": "Datasets. 1) PARIS two-part object dataset. This dataset contains two-part articulated objects, featuring 10 synthetic objects from PartNet-Mobility [49] and 2 real-world objects captured using MultiScan [27]. Each object comprises a movable part and a static part, observed in two distinct joint states. The dataset includes RGB images, object masks from 100 random viewpoints, and depth images for both synthetic and real-world objects. 2) Synthetic multi-part" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "content": "object dataset. This dataset includes two synthetic scenes containing multi-part objects from PartNet-Mobility, with each object featuring one static part and multiple movable parts. Each object is observed in two articulation states, and the dataset provides RGB, depth, and mask data from 100 random viewpoints. 3) Indoor scene OmniSim dataset. OmniSim is generated using the OmniGibson [20] simulator with various indoor scene models. By adjusting the rotation of articulated object joints, we generate some scenes for evaluation, each containing RGBD images, interactive object masks and object state metrics at each state." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 208, + 554, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 208, + 554, + 352 + ], + "spans": [ + { + "bbox": [ + 313, + 208, + 554, + 352 + ], + "type": "text", + "content": "Metrics. To evaluate articulation models, we use: 1) Axis Ang Err " + }, + { + "bbox": [ + 313, + 208, + 554, + 352 + ], + "type": "inline_equation", + "content": "(^{\\circ})" + }, + { + "bbox": [ + 313, + 208, + 554, + 352 + ], + "type": "text", + "content": " means angular error of the predicted joint axis for both revolute and prismatic joints. 2) Axis Pos Err (0.1 m): The shortest distance between the predicted and true joint axes for revolute joints. For cross-state part motion, we use Part Motion Err " + }, + { + "bbox": [ + 313, + 208, + 554, + 352 + ], + "type": "inline_equation", + "content": "(^{\\circ}" + }, + { + "bbox": [ + 313, + 208, + 554, + 352 + ], + "type": "text", + "content": " or m). It evaluates joint state precision by calculating the geodesic error for predicted rotations in revolute joints or the Euclidean error for translations in prismatic joints, as specified in PARIS [22]. For object and part mesh reconstruction, we use Chamfer-L1 distance (CD), reporting CD-w for the full surface, CD-s for static parts, and CD-m for movable parts." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 354, + 554, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 554, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 554, + 544 + ], + "type": "text", + "content": "Baselines. Ditto [16] and PARIS [22] are models for reconstructing part-level shape and motion in two-part articulated objects, using multi-view data. Ditto assumes one moving part and is pre-trained on specific categories, while PARIS uses a NeRF-based representation to handle unknown objects. " + }, + { + "bbox": [ + 313, + 354, + 554, + 544 + ], + "type": "inline_equation", + "content": "PARIS^{*}" + }, + { + "bbox": [ + 313, + 354, + 554, + 544 + ], + "type": "text", + "content": " is an enhanced version with depth supervision, and " + }, + { + "bbox": [ + 313, + 354, + 554, + 544 + ], + "type": "inline_equation", + "content": "PARIS - m^{*}" + }, + { + "bbox": [ + 313, + 354, + 554, + 544 + ], + "type": "text", + "content": " extends PARIS to multi-part objects. CSG-reg uses TSDF fusion and Constructive Solid Geometry for part segmentation, with registration for alignment. 3Dseg-reg follows a similar process but employs PA-Conv [51] for part segmentation, reporting results only for trained categories due to limited generalization. DigitalTwinArt [47] divides reconstruction into two stages: 1) It reconstructs object-level shape independently of articulation; 2) It recovers the articulation model by identifying part segmentation and motions through state correspondences." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 545, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 554, + 713 + ], + "type": "text", + "content": "Evaluation Setup. There are some key differences on evaluation setup between our model and previous models as we firstly incorporate the semantics into the explicit neural field supporting robust interactions even in complex environment. Unlike prior methods which assume the number of parts is known, our model can detect any articulated objects if enough information for reconstructing object shape is given from multi-view images. To identify corresponding parts for evaluation, prior work need to iterate through all possible pairs of predicted and ground-truth parts, selecting the match with the smallest total Chamfer distance. In our model, we identify the corresponding parts based on the constructed semantic feature and mask label field. We extract the mesh based on the 3D point clouds derived" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12137" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 69, + 553, + 439 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 553, + 439 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 553, + 439 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 553, + 439 + ], + "type": "table", + "html": "
SimulationReal
FoldChairFridgeLaptopOvenScissorStaplerUSBWasherBladeStorageAllFridgeStorageAll
Axis Ang ↓Ditto [16]89.3589.783.120.964.5089.8689.7789.5179.54*6.3254.221.715.883.80
PARIS [22]8.089.150.020.043.8239.730.1325.3615.380.0310.171.6443.1322.39
PARIS* [22]15.792.930.037.4316.628.170.7118.4041.280.0311.141.9030.1016.00
CSG-reg0.100.270.470.350.280.3011.7871.937.642.829.608.9269.7139.31
3Dseg-reg--2.34-----9.40-----
DigitalTwinArt [47]0.030.070.060.220.110.060.110.430.270.060.142.1018.1110.11
IAAO (Ours)0.020.040.050.170.080.050.090.360.190.060.111.915.828.86
Axis Pos ↓Ditto [16]3.771.02*0.010.135.700.205.410.66--2.111.84-0.92
PARIS [22]0.450.380.000.002.102.272.361.50--1.130.34-0.17
PARIS* [22]0.251.130.000.051.594.673.353.28--1.790.50-0.25
CSG-reg0.020.000.200.180.010.020.012.13--0.321.46-0.73
3Dseg-reg--0.10-----------
DigitalTwinArt [47]0.010.010.000.010.020.010.000.01--0.010.57-0.29
IAAO (Ours)0.010.000.000.000.010.010.010.00--0.010.32-0.16
Part Motion ↓Ditto [16]99.36×5.182.0919.2856.6180.6055.72×0.0939.878.430.384.41
PARIS [22]131.6624.580.030.03120.70110.8064.8560.350.340.3051.362.160.561.36
PARIS* [22]127.3445.260.039.1368.36107.7696.9349.770.360.3050.521.580.571.07
CSG-reg0.130.290.350.580.200.4410.48158.990.050.0417.1614.820.647.73
3Dseg-reg--1.61-----0.15-----
DigitalTwinArt [47]0.160.090.080.110.150.050.110.250.000.000.101.860.201.03
IAAO (Ours)0.130.070.030.040.110.050.090.210.000.000.071.460.150.81
CD-s ↓Ditto [14]33.793.050.252.5239.0741.642.6410.3246.909.1818.9447.0116.0931.55
PARIS [22]9.163.651.941.8812.951.882.6925.391.1912.767.1842.5754.5448.56
PARIS* [22]10.208.8215.581.953.182.482.4812.191.408.676.4611.6420.2515.94
CSG-reg1.691.453.933.263.262.221.954.530.597.062.706.3312.559.44
3Dseg-reg--0.76-----66.31-----
DigitalTwinArt [47]0.180.604.660.402.652.194.804.690.554.692.102.5310.866.69
IAAO (Ours)0.160.510.170.361.981.864.433.700.514.271.802.479.836.15
CD-m ↓Ditto [14]141.110.990.190.9420.6831.2115.8812.89195.932.2042.2050.6020.3535.48
PARIS [22]8.997.760.2128.7046.6419.275.32178.4325.2176.6939.7245.66864.82455.24
PARIS* [22]17.977.230.156.5416.6530.4610.17265.27117.9952.3452.4877.85474.57276.21
CSG-reg1.9121.710.42256.991.956.3629.78436.4226.621.3978.36442.17521.49481.83
3Dseg-reg--1.01-----6.23-----
DigitalTwinArt [47]0.150.270.160.470.412.271.340.361.500.370.731.1426.4613.80
IAAO (Ours)0.130.180.150.410.402.091.180.331.420.250.651.0221.4311.23
CD-w ↓Ditto [14]6.802.160.312.511.702.382.097.2942.043.917.126.5014.0810.29
PARIS [22]1.802.920.3011.7310.493.582.0024.380.608.576.6422.9863.3543.16
PARIS* [22]4.375.530.263.183.905.271.7810.110.587.804.288.9932.1020.55
CSG-reg0.480.980.403.001.701.991.204.480.564.001.885.7114.2910.00
3Dseg-reg--0.81-----0.78-----
DigitalTwinArt [47]0.270.700.354.180.432.191.184.740.363.991.842.199.335.76
IAAO (Ours)0.220.610.273.490.341.961.064.410.343.941.662.029.185.60
", + "image_path": "0517bed912c1e8b9c9a54e9cd9a7a421e055bd7e85774efe6b8b1497953a044f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 480, + 553, + 550 + ], + "blocks": [ + { + "bbox": [ + 55, + 447, + 555, + 470 + ], + "lines": [ + { + "bbox": [ + 55, + 447, + 555, + 470 + ], + "spans": [ + { + "bbox": [ + 55, + 447, + 555, + 470 + ], + "type": "text", + "content": "Table 1. Results on PARIS dataset including both synthetic and real data. 'x': failure case, and \\*\\*: joint axis or position. -': no result is available. Best result in bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 480, + 553, + 550 + ], + "lines": [ + { + "bbox": [ + 56, + 480, + 553, + 550 + ], + "spans": [ + { + "bbox": [ + 56, + 480, + 553, + 550 + ], + "type": "table", + "html": "
Axis Ang 0↓Axis Ang 1 ↓Axis Pos 0 ↓Axis Pos 1 ↓Part Motion 0 ↓Part Motion 1 ↓CD-s ↓CD-m 0 ↓CD-m 1 ↓CD-w ↓
Fridge-mPARIS*-m [22]34.5215.913.601.6386.21105.868.52526.19160.8615.00
DigitalTwinArt [47]0.160.100.010.000.110.130.610.400.520.89
IAAO (Ours)0.160.080.010.000.090.090.540.440.470.96
Storage-mPARIS*-m [22]43.2626.1810.42-79.840.648.56128.62266.718.66
DigitalTwinArt [47]0.210.880.05-0.130.000.850.213.460.99
IAAO (Ours)0.180.910.04-0.100.000.780.263.350.97
", + "image_path": "83751533b87998ba7768907d728b67924477901c0b950076d77e3a9566afde1e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 556, + 555, + 578 + ], + "lines": [ + { + "bbox": [ + 55, + 556, + 555, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 555, + 578 + ], + "type": "text", + "content": "Table 2. Results on multi-part object dataset, averaged over 10 trials with different random seeds. Joint 1 of \"Storage-m\" is solely prismatic with no Axis Pos. '-': no result is available. Best results in bold." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 600, + 294, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 294, + 635 + ], + "type": "text", + "content": "from 3DGS model via SuGaR [12] for evaluation. Following [22, 47], we transform our extracted parts with predicted motions to start state " + }, + { + "bbox": [ + 55, + 600, + 294, + 635 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 600, + 294, + 635 + ], + "type": "text", + "content": " for evaluation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 647, + 174, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 174, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 174, + 659 + ], + "type": "text", + "content": "4.2. Quantitative Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "Tab. 1 shows quantitative results from the PARIS Two-Part Object Dataset, covering synthetic and real instances. Ditto performs well on seen categories but struggles with unseen ones. PARIS occasionally fails in shape and articu" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 600, + 555, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 555, + 672 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 555, + 672 + ], + "type": "text", + "content": "lation reconstruction, while depth supervision in PARIS* improves shape accuracy for complex and real-world objects but increases articulation errors due to optimization challenges. CSG-reg and 3Dseg-reg handle simple objects well but struggle with complex ones, as segmentation errors affect articulation estimation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "Tab. 2 presents the results for multi-part objects. DigitalTwinArt [47] outperforms PARIS* in shape and articulation reconstruction by incorporating additional supervi" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "12138" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 102, + 95, + 140, + 144 + ], + "blocks": [ + { + "bbox": [ + 97, + 72, + 145, + 84 + ], + "lines": [ + { + "bbox": [ + 97, + 72, + 145, + 84 + ], + "spans": [ + { + "bbox": [ + 97, + 72, + 145, + 84 + ], + "type": "text", + "content": "Input State " + }, + { + "bbox": [ + 97, + 72, + 145, + 84 + ], + "type": "inline_equation", + "content": "t" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 102, + 95, + 140, + 144 + ], + "lines": [ + { + "bbox": [ + 102, + 95, + 140, + 144 + ], + "spans": [ + { + "bbox": [ + 102, + 95, + 140, + 144 + ], + "type": "image", + "image_path": "89c9354449d82712c5da8a3ab67c7479c09ada6c7cfb2f93029f474b6e05d0d1.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 101, + 154, + 149, + 204 + ], + "blocks": [ + { + "bbox": [ + 101, + 154, + 149, + 204 + ], + "lines": [ + { + "bbox": [ + 101, + 154, + 149, + 204 + ], + "spans": [ + { + "bbox": [ + 101, + 154, + 149, + 204 + ], + "type": "image", + "image_path": "28dea88c1c409761ac03efc7dde903dfb89b1a47754db10ae7f9139fc7322908.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 70, + 211, + 537, + 222 + ], + "lines": [ + { + "bbox": [ + 70, + 211, + 537, + 222 + ], + "spans": [ + { + "bbox": [ + 70, + 211, + 537, + 222 + ], + "type": "text", + "content": "Figure 3. Qualitative analysis of shape reconstruction, part segmentation, and joint prediction results on multi-part object dataset." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 221, + 96, + 264, + 146 + ], + "blocks": [ + { + "bbox": [ + 220, + 73, + 270, + 83 + ], + "lines": [ + { + "bbox": [ + 220, + 73, + 270, + 83 + ], + "spans": [ + { + "bbox": [ + 220, + 73, + 270, + 83 + ], + "type": "text", + "content": "Input State " + }, + { + "bbox": [ + 220, + 73, + 270, + 83 + ], + "type": "inline_equation", + "content": "t^{\\prime}" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 221, + 96, + 264, + 146 + ], + "lines": [ + { + "bbox": [ + 221, + 96, + 264, + 146 + ], + "spans": [ + { + "bbox": [ + 221, + 96, + 264, + 146 + ], + "type": "image", + "image_path": "6efbf2f792257d879d606dd525cab3d11e7171ccdea792aa6dd2a762f5743812.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 216, + 153, + 266, + 205 + ], + "blocks": [ + { + "bbox": [ + 216, + 153, + 266, + 205 + ], + "lines": [ + { + "bbox": [ + 216, + 153, + 266, + 205 + ], + "spans": [ + { + "bbox": [ + 216, + 153, + 266, + 205 + ], + "type": "image", + "image_path": "b91e2bc9b5cc41f18b4a764ba040032890c75cd3f83cd9b4a7f3d33c9c0444b5.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 353, + 91, + 380, + 150 + ], + "blocks": [ + { + "bbox": [ + 340, + 73, + 393, + 84 + ], + "lines": [ + { + "bbox": [ + 340, + 73, + 393, + 84 + ], + "spans": [ + { + "bbox": [ + 340, + 73, + 393, + 84 + ], + "type": "text", + "content": "DigitalTwinArt" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 353, + 91, + 380, + 150 + ], + "lines": [ + { + "bbox": [ + 353, + 91, + 380, + 150 + ], + "spans": [ + { + "bbox": [ + 353, + 91, + 380, + 150 + ], + "type": "image", + "image_path": "7ed2f1a62e2b807252e49a4dcb0d2874bf1e57c923223c9461152b64fd6b7aaf.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 343, + 155, + 396, + 206 + ], + "blocks": [ + { + "bbox": [ + 343, + 155, + 396, + 206 + ], + "lines": [ + { + "bbox": [ + 343, + 155, + 396, + 206 + ], + "spans": [ + { + "bbox": [ + 343, + 155, + 396, + 206 + ], + "type": "image", + "image_path": "13dada9615698bce585a36479a82796c309641397aa2e1a2b74e53412daf17fa.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 471, + 91, + 503, + 152 + ], + "blocks": [ + { + "bbox": [ + 475, + 74, + 495, + 83 + ], + "lines": [ + { + "bbox": [ + 475, + 74, + 495, + 83 + ], + "spans": [ + { + "bbox": [ + 475, + 74, + 495, + 83 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 471, + 91, + 503, + 152 + ], + "lines": [ + { + "bbox": [ + 471, + 91, + 503, + 152 + ], + "spans": [ + { + "bbox": [ + 471, + 91, + 503, + 152 + ], + "type": "image", + "image_path": "037e560b46698e76ad2993d6b34ee4017145eda164add557532a09b933251b3f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 465, + 155, + 517, + 206 + ], + "blocks": [ + { + "bbox": [ + 465, + 155, + 517, + 206 + ], + "lines": [ + { + "bbox": [ + 465, + 155, + 517, + 206 + ], + "spans": [ + { + "bbox": [ + 465, + 155, + 517, + 206 + ], + "type": "image", + "image_path": "b84eb0f24ff6a40ca1e10d3f94d80597e779743cfaf49a0d6444b8f4b1a93dfd.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 65, + 243, + 296, + 366 + ], + "blocks": [ + { + "bbox": [ + 65, + 243, + 296, + 366 + ], + "lines": [ + { + "bbox": [ + 65, + 243, + 296, + 366 + ], + "spans": [ + { + "bbox": [ + 65, + 243, + 296, + 366 + ], + "type": "image", + "image_path": "dd0508e1e8dd9913d693ca0808d6dce190c68df45232188a33e61f5b0b27abea.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 373, + 295, + 395 + ], + "lines": [ + { + "bbox": [ + 55, + 373, + 295, + 395 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 295, + 395 + ], + "type": "text", + "content": "Figure 4. Qualitative results of shape reconstruction, part segmentation, and joint prediction on PARIS." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 69, + 400, + 291, + 456 + ], + "blocks": [ + { + "bbox": [ + 69, + 400, + 291, + 456 + ], + "lines": [ + { + "bbox": [ + 69, + 400, + 291, + 456 + ], + "spans": [ + { + "bbox": [ + 69, + 400, + 291, + 456 + ], + "type": "image", + "image_path": "f6d070a5e79eaf79715f24ba579dab346a7120340df4ac72474f41ded37c9f64.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 68, + 459, + 282, + 470 + ], + "lines": [ + { + "bbox": [ + 68, + 459, + 282, + 470 + ], + "spans": [ + { + "bbox": [ + 68, + 459, + 282, + 470 + ], + "type": "text", + "content": "Figure 5. Motion snapshots on PARIS & multi-part object." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 62, + 475, + 290, + 529 + ], + "blocks": [ + { + "bbox": [ + 62, + 475, + 290, + 529 + ], + "lines": [ + { + "bbox": [ + 62, + 475, + 290, + 529 + ], + "spans": [ + { + "bbox": [ + 62, + 475, + 290, + 529 + ], + "type": "image", + "image_path": "b2640977660d1bae9da69cdd7570becc2c96823c8753170ed0e1056d770dce37.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 532, + 286, + 544 + ], + "lines": [ + { + "bbox": [ + 64, + 532, + 286, + 544 + ], + "spans": [ + { + "bbox": [ + 64, + 532, + 286, + 544 + ], + "type": "text", + "content": "Figure 6. Motion snapshots on scene-level OmniSim dataset." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 556, + 296, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 556, + 296, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 296, + 593 + ], + "type": "text", + "content": "sion on 3D geometries. Our IAAO achieves competitive results by leveraging geometric and semantic information from VLMs, enhancing object part modeling." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 600, + 167, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 167, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 167, + 612 + ], + "type": "text", + "content": "4.3. Qualitative Results" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 54, + 618, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 618, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 618, + 295, + 713 + ], + "type": "text", + "content": "Fig. 3 and Fig. 4 show qualitative results on shape reconstruction, part segmentation, and joint prediction for the multi-part object and PARIS two-part datasets. Our IAAO achieves more complete segmentation results while also enabling interaction with affordance and segmentation features. Fig. 5 and 6 show snapshots of the generated motions on several examples at object and scene levels in a simulator. It shows IAAO can produce smooth interpolations of" + } + ] + } + ], + "index": 21 + }, + { + "type": "table", + "bbox": [ + 317, + 241, + 552, + 294 + ], + "blocks": [ + { + "bbox": [ + 317, + 241, + 552, + 294 + ], + "lines": [ + { + "bbox": [ + 317, + 241, + 552, + 294 + ], + "spans": [ + { + "bbox": [ + 317, + 241, + 552, + 294 + ], + "type": "table", + "html": "
MethodAxis Ang ↓Axis Pos ↓Part Motion ↓CD-s ↓CD-m ↓
Proposed0.280.010.110.800.35
w/o matching4.591.388.571.21123.71
w/o mask14.820.010.282.59129.14
w/o rgb0.560.010.170.810.42
w/o label0.540.010.158.56169.73
", + "image_path": "e316cb575ad6a653fb7e95be1df160fc44846afb1e12e0c4b1b9d95d61af5c40.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "table_body" + } + ], + "index": 22 + }, + { + "bbox": [ + 338, + 302, + 529, + 313 + ], + "lines": [ + { + "bbox": [ + 338, + 302, + 529, + 313 + ], + "spans": [ + { + "bbox": [ + 338, + 302, + 529, + 313 + ], + "type": "text", + "content": "Table 3. Ablation study on multi-part object dataset." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 324, + 533, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 533, + 337 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 533, + 337 + ], + "type": "text", + "content": "articulated objects with good part geometry modeling." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 342, + 408, + 356 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 408, + 356 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 408, + 356 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 361, + 555, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 555, + 492 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 555, + 492 + ], + "type": "text", + "content": "We assess the impact of our design choices specifically on multi-part objects, which present greater challenges. We first perform experiments to test out the impact of 2D-3D matching loss ('w/o matching'). As indicated in Tab. 3, the matching loss plays a crucial role in enhancing articulation reconstruction. We also study constraints from mask features and labels, which act as a coarse correspondence mechanism at the superpoint (mask) level. These constraints also enforce consistency within the feature fields of each mask, leading to improved part segmentation and more accurate reconstruction of articulated components." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 502, + 388, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 388, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 388, + 514 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "type": "text", + "content": "We present IAAO, a framework that facilitates interactive affordance detection and object articulation from two scene sequences. Each scene sequence captures movable parts of objects in different joint states. Our IAAO can handle both simple scenes with single objects and complex indoor scenes without limitations on the number of objects and movable parts. The entire scene is reconstructed using 3D Gaussians, embedded with robust zero-shot generalization capabilities from 2D foundation models. Affordance prediction is achieved through object and part queries on the 3D Gaussian primitives. Motion estimation is achieved by establishing 2D-3D correspondences within each object to track transformations. Static elements serve for global scene alignment while movable elements are for local transformation. Extensive experiments demonstrate that our approach outperforms existing methods." + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12139" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 120 + ], + "type": "text", + "content": "Acknowledgement. This research / project is supported by the National Research Foundation (NRF) Singapore, under its NRF-Investigatorship Programme (Award ID. NRF-NRFI09-0008)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 115, + 142 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 115, + 142 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 115, + 142 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 149, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 149, + 296, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 149, + 296, + 172 + ], + "spans": [ + { + "bbox": [ + 61, + 149, + 296, + 172 + ], + "type": "text", + "content": "[1] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers, 2023. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 172, + 296, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 172, + 296, + 237 + ], + "spans": [ + { + "bbox": [ + 61, + 172, + 296, + 237 + ], + "type": "text", + "content": "[2] Alexandros Delitzas, Ayca Takmaz, Federico Tombari, Robert Sumner, Marc Pollefeys, and Francis Engelmann. Scenefun3d: Fine-grained functionality and affordance understanding in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14531-14542, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 238, + 296, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 238, + 296, + 293 + ], + "spans": [ + { + "bbox": [ + 61, + 238, + 296, + 293 + ], + "type": "text", + "content": "[3] Shengheng Deng, Xun Xu, Chaozheng Wu, Ke Chen, and Kui Jia. 3d affordancenet: A benchmark for visual object affordance understanding. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1778-1787, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 293, + 296, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 293, + 296, + 337 + ], + "spans": [ + { + "bbox": [ + 61, + 293, + 296, + 337 + ], + "type": "text", + "content": "[4] Thanh-Toan Do, Anh Nguyen, and Ian Reid. Affordancenet: An end-to-end deep learning approach for object affordance detection. In 2018 IEEE international conference on robotics and automation (ICRA), pages 5882-5889. IEEE, 2018. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 338, + 296, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 338, + 296, + 413 + ], + "spans": [ + { + "bbox": [ + 61, + 338, + 296, + 413 + ], + "type": "text", + "content": "[5] Xiaoyi Dong, Jianmin Bao, Yinglin Zheng, Ting Zhang, Dongdong Chen, Hao Yang, Ming Zeng, Weiming Zhang, Lu Yuan, Dong Chen, et al. Maskclip: Masked self-distillation advances contrastive language-image pretraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10995-11005, 2023. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 415, + 296, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 415, + 296, + 468 + ], + "spans": [ + { + "bbox": [ + 62, + 415, + 296, + 468 + ], + "type": "text", + "content": "[6] Francis Engelmann, Fabian Manhardt, Michael Niemeyer, Keisuke Tateno, Marc Pollefeys, and Federico Tombari. Opennerf: Open set 3d neural scene segmentation with pixelwise features and rendered novel views. arXiv preprint arXiv:2404.03650, 2024. 2, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 470, + 296, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 470, + 296, + 503 + ], + "spans": [ + { + "bbox": [ + 62, + 470, + 296, + 503 + ], + "type": "text", + "content": "[7] Carlos Hernández Esteban and Francis Schmitt. Silhouette and stereo fusion for 3d object modeling. Computer Vision and Image Understanding, 96(3):367-392, 2004. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 503, + 296, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 503, + 296, + 557 + ], + "spans": [ + { + "bbox": [ + 62, + 503, + 296, + 557 + ], + "type": "text", + "content": "[8] Kuan Fang, Te-Lin Wu, Daniel Yang, Silvio Savarese, and Joseph J Lim. Demo2vec: Reasoning object affordances from online videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2139-2147, 2018. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 558, + 296, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 558, + 296, + 591 + ], + "spans": [ + { + "bbox": [ + 62, + 558, + 296, + 591 + ], + "type": "text", + "content": "[9] JJ Gibson. The theory of affordances. Perceiving, acting and knowing: Towards an ecological psychology/Erlbaum, 1977. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 592, + 295, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 592, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 57, + 592, + 295, + 613 + ], + "type": "text", + "content": "[10] James J Gibson. The ecological approach to visual perception: classic edition. Psychology press, 2014. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 614, + 295, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 614, + 295, + 658 + ], + "spans": [ + { + "bbox": [ + 57, + 614, + 295, + 658 + ], + "type": "text", + "content": "[11] Michael Goesele, Brian Curless, and Steven M Seitz. Multiview stereo revisited. In 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'06), pages 2402-2409. IEEE, 2006. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 658, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 296, + 713 + ], + "type": "text", + "content": "[12] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5354-5363, 2024. 7" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "type": "text", + "content": "[13] Karol Hausman, Scott Niekum, Sarah Osentoski, and Gaurav S Sukhatme. Active articulation model estimation through interactive perception. In 2015 IEEE International Conference on Robotics and Automation (ICRA), pages 3305-3312. IEEE, 2015. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 128, + 554, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 128, + 554, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 128, + 554, + 183 + ], + "type": "text", + "content": "[14] Cheng-Chun Hsu, Zhenyu Jiang, and Yuke Zhu. Ditto in the house: Building articulation models of indoor scenes through interactive perception. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 3933-3939. IEEE, 2023. 3, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 184, + 554, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 184, + 554, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 184, + 554, + 228 + ], + "type": "text", + "content": "[15] Hanxiao Jiang, Yongsen Mao, Manolis Savva, and Angel X Chang. Opd: Single-view 3d openable part detection. In European Conference on Computer Vision, pages 410-426. Springer, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 228, + 554, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 554, + 281 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 554, + 281 + ], + "type": "text", + "content": "[16] Zhenyu Jiang, Cheng-Chun Hsu, and Yuke Zhu. Ditto: Building digital twins of articulated objects from interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5616-5626, 2022. 1, 3, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 283, + 554, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 283, + 554, + 326 + ], + "spans": [ + { + "bbox": [ + 316, + 283, + 554, + 326 + ], + "type": "text", + "content": "[17] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 327, + 554, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 554, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 554, + 381 + ], + "type": "text", + "content": "[18] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Lerf: Language embedded radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19729-19739, 2023. 2, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 383, + 554, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 554, + 436 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 554, + 436 + ], + "type": "text", + "content": "[19] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 437, + 554, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 437, + 554, + 503 + ], + "spans": [ + { + "bbox": [ + 316, + 437, + 554, + 503 + ], + "type": "text", + "content": "[20] Chengshu Li, Ruohan Zhang, Josiah Wong, Cem Gokmen, Sanjana Srivastava, Roberto Martin-Martín, Chen Wang, Gabrael Levine, Wensi Ai, Benjamin Martinez, et al. Behavior-1k: A human-centered, embodied ai benchmark with 1,000 everyday activities and realistic simulation. arXiv preprint arXiv:2403.09227, 2024. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 503, + 554, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 503, + 554, + 557 + ], + "spans": [ + { + "bbox": [ + 316, + 503, + 554, + 557 + ], + "type": "text", + "content": "[21] Xiaolong Li, He Wang, Li Yi, Leonidas J Guibas, A Lynn Abbott, and Shuran Song. Category-level articulated object pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3706-3715, 2020. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 558, + 554, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 558, + 554, + 603 + ], + "spans": [ + { + "bbox": [ + 316, + 558, + 554, + 603 + ], + "type": "text", + "content": "[22] Jiayi Liu, Ali Mahdavi-Amiri, and Manolis Savva. Paris: Part-level reconstruction and motion analysis for articulated objects. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 352-363, 2023. 1, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 603, + 554, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 603, + 554, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 603, + 554, + 658 + ], + "type": "text", + "content": "[23] Kunhao Liu, Fangneng Zhan, Jiahui Zhang, Muyu Xu, Yingchen Yu, Abdulmotaleb El Saddik, Christian Theobalt, Eric Xing, and Shijian Lu. Weakly supervised 3d open-vocabulary segmentation. Advances in Neural Information Processing Systems, 36:53433-53456, 2023. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 658, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 554, + 713 + ], + "type": "text", + "content": "[24] Liu Liu, Wenqiang Xu, Haoyuan Fu, Sucheng Qian, Qiao-jun Yu, Yang Han, and Cewu Lu. Akb-48: A real-world articulated object knowledge base. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14809–14818, 2022. 3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12140" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "text", + "content": "[25] Timo Luddecke and Florentin Worgotter. Learning to segment affordances. In Proceedings of the IEEE International Conference on Computer Vision Workshops, pages 769-776, 2017. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "text", + "content": "[26] Priyanka Mandikal and Kristen Grauman. Learning dexterous grasping with object-centric visual affordances. In 2021 IEEE international conference on robotics and automation (ICRA), pages 6169-6176. IEEE, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 164, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 164, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 294, + 217 + ], + "type": "text", + "content": "[27] Yongsen Mao, Yiming Zhang, Hanxiao Jiang, Angel Chang, and Manolis Savva. Multiscan: Scalable rgbd scanning for 3d environments with articulated objects. Advances in neural information processing systems, 35:9058-9071, 2022. 2, 3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 294, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 273 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 273 + ], + "type": "text", + "content": "[28] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 276, + 294, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 341 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 341 + ], + "type": "text", + "content": "[29] Kaichun Mo, Shilin Zhu, Angel X Chang, Li Yi, Subarna Tripathi, Leonidas J Guibas, and Hao Su. Partnet: A largescale benchmark for fine-grained and hierarchical part-level 3d object understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 909-918, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 343, + 294, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 294, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 294, + 387 + ], + "type": "text", + "content": "[30] Kaichun Mo, Yuzhe Qin, Fanbo Xiang, Hao Su, and Leonidas Guibas. O2o-afford: Annotation-free large-scale object-object affordance learning. In Conference on robot learning, pages 1666-1677. PMLR, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 388, + 294, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 431 + ], + "type": "text", + "content": "[31] Tushar Nagarajan and Kristen Grauman. Learning affordance landscapes for interaction exploration in 3d environments. Advances in Neural Information Processing Systems, 33:2005-2015, 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 434, + 294, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 294, + 476 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 294, + 476 + ], + "type": "text", + "content": "[32] Neil Nie, Samir Yitzhak Gadre, Kiana Ehsani, and Shuran Song. Structure from action: Learning interactions for articulated object 3d structure discovery. arXiv preprint arXiv:2207.08997, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 479, + 294, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 479, + 294, + 577 + ], + "spans": [ + { + "bbox": [ + 56, + 479, + 294, + 577 + ], + "type": "text", + "content": "[33] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 2, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 579, + 294, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 294, + 621 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 294, + 621 + ], + "type": "text", + "content": "[34] Sylvain Paris, François X Sillion, and Long Quan. A surface reconstruction method using global graph cut optimization. International Journal of Computer Vision, 66:141-161, 2006. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 624, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 678 + ], + "type": "text", + "content": "[35] Songyou Peng, Kyle Genova, Chiyu Jiang, Andrea Tagliasacchi, Marc Pollefeys, Thomas Funkhouser, et al. Openscene: 3d scene understanding with open vocabularies. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 815-824, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "text", + "content": "[36] Shengyi Qian, Linyi Jin, Chris Rockwell, Siyi Chen, and David F Fouhey. Understanding 3d object articulation in internet videos. In Proceedings of the IEEE/CVF Conference" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "text", + "content": "on Computer Vision and Pattern Recognition, pages 1599-1609, 2022. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 97, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 553, + 149 + ], + "type": "text", + "content": "[37] Minghan Qin, Wanhua Li, Jiawei Zhou, Haoqian Wang, and Hanspeter Pfister. Langsplat: 3d language gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20051-20060, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 153, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 153, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 153, + 553, + 207 + ], + "type": "text", + "content": "[38] Zheng Qin, Hao Yu, Changjian Wang, Yulan Guo, Yuxing Peng, and Kai Xu. Geometric transformer for fast and robust point cloud registration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11143-11152, 2022. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 209, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 209, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 317, + 209, + 553, + 274 + ], + "type": "text", + "content": "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 276, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 276, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 317, + 276, + 553, + 319 + ], + "type": "text", + "content": "[40] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 322, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 322, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 317, + 322, + 553, + 374 + ], + "type": "text", + "content": "[41] Christoph Strecha, Rik Fransens, and Luc Van Gool. Combined depth and outlier estimation in multi-view stereo. In 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'06), pages 2394-2401. IEEE, 2006. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 378, + 553, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 378, + 553, + 432 + ], + "spans": [ + { + "bbox": [ + 317, + 378, + 553, + 432 + ], + "type": "text", + "content": "[42] Archana Swaminathan, Anubhav Gupta, Kamal Gupta, Shishira R Maiya, Vatsal Agarwal, and Abhinav Shrivastava. Leia: Latent view-invariant embeddings for implicit 3d articulation. In European Conference on Computer Vision, pages 210-227. Springer, 2025. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 434, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 434, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 317, + 434, + 553, + 487 + ], + "type": "text", + "content": "[43] Ayca Takmaz, Elisabetta Fedele, Robert Sumner, Marc Pollefeys, Federico Tombari, and Francis Engelmann. Openmask3d: Open-vocabulary 3d instance segmentation. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 491, + 553, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 491, + 553, + 566 + ], + "spans": [ + { + "bbox": [ + 317, + 491, + 553, + 566 + ], + "type": "text", + "content": "[44] Zhenggang Tang, Balakumar Sundaralingam, Jonathan Tremblay, Bowen Wen, Ye Yuan, Stephen Tyree, Charles Loop, Alexander Schwing, and Stan Birchfield. RGB-only reconstruction of tabletop scenes for collision-free manipulator control. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 1778-1785. IEEE, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 568, + 553, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 568, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 317, + 568, + 553, + 622 + ], + "type": "text", + "content": "[45] Can Wang, Mengei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3835-3844, 2022. 2, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 624, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 624, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 317, + 624, + 553, + 678 + ], + "type": "text", + "content": "[46] Xiaogang Wang, Bin Zhou, Yahao Shi, Xiaowu Chen, Qinping Zhao, and Kai Xu. Shape2motion: Joint analysis of motion parts and attributes from 3d shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8876-8884, 2019. 1, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "type": "text", + "content": "[47] Yijia Weng, Bowen Wen, Jonathan Tremblay, Valts Blukis, Dieter Fox, Leonidas Guibas, and Stan Birchfield. Neural implicit representation for building digital twins of unknown" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12141" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 75, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 294, + 106 + ], + "type": "text", + "content": "articulated objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3141-3150, 2024. 1, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 151 + ], + "type": "text", + "content": "[48] Youngsun Wi, Pete Florence, Andy Zeng, and Nima Fazeli. Virdo: Visio-tactile implicit representations of deformable objects. In 2022 International Conference on Robotics and Automation (ICRA), pages 3583-3590. IEEE, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 217 + ], + "type": "text", + "content": "[49] Fanbo Xiang, Yuzhe Qin, Kaichun Mo, Yikuan Xia, Hao Zhu, Fangchen Liu, Minghua Liu, Hanxiao Jiang, Yifu Yuan, He Wang, et al. Sapien: A simulated part-based interactive environment. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11097-11107, 2020. 3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 294, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 275 + ], + "type": "text", + "content": "[50] Danfei Xu, Ajay Mandlekar, Roberto Martin-Martin, Yuke Zhu, Silvio Savarese, and Li Fei-Fei. Deep affordance foresight: Planning through what can be done in the future. In 2021 IEEE international conference on robotics and automation (ICRA), pages 6206–6213. IEEE, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 276, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 330 + ], + "type": "text", + "content": "[51] Mutian Xu, Runyu Ding, Hengshuang Zhao, and Xiaojuan Qi. Paconv: Position adaptive convolution with dynamic kernel assembling on point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3173-3182, 2021. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 332, + 294, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 294, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 294, + 387 + ], + "type": "text", + "content": "[52] Mi Yan, Jiazhao Zhang, Yan Zhu, and He Wang. Maskclustering: View consensus based mask graph clustering for open-vocabulary 3d instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28274-28284, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 388, + 294, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 432 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 432 + ], + "type": "text", + "content": "[53] Zihao Yan, Ruizhen Hu, Xingguang Yan, Luanmin Chen, Oliver Van Kaick, Hao Zhang, and Hui Huang. RPM-net: recurrent prediction of motion and parts from point cloud. arXiv preprint arXiv:2006.14865, 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 434, + 294, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 294, + 477 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 294, + 477 + ], + "type": "text", + "content": "[54] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In Proceedings of the European conference on computer vision (ECCV), pages 767-783, 2018. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 479, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 479, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 479, + 294, + 544 + ], + "type": "text", + "content": "[55] Andy Zeng, Shuran Song, Kuan-Ting Yu, Elliott Donlon, Francois R Hogan, Maria Bauza, Daolin Ma, Orion Taylor, Melody Liu, Eudald Romo, et al. Robotic pick-and-place of novel objects in clutter with multi-affordance grasping and cross-domain image matching. The International Journal of Robotics Research, 41(7):690-705, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 546, + 294, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 546, + 294, + 600 + ], + "spans": [ + { + "bbox": [ + 56, + 546, + 294, + 600 + ], + "type": "text", + "content": "[56] Vicky Zeng, Tabitha Edith Lee, Jacky Liang, and Oliver Kroemer. Visual identification of articulated object parts. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 2443-2450. IEEE, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 602, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 294, + 656 + ], + "type": "text", + "content": "[57] Chaoning Zhang, Dongshen Han, Yu Qiao, Jung Uk Kim, Sung-Ho Bae, Seungkyu Lee, and Choong Seon Hong. Faster segment anything: Towards lightweight sam for mobile applications. arXiv preprint arXiv:2306.14289, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "text", + "content": "[58] Allan Zhou, Moo Jin Kim, Lirui Wang, Pete Florence, and Chelsea Finn. Nerf in the palm of your hand: Corrective augmentation for robotics via novel-view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17907-17917, 2023. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 205 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 139 + ], + "type": "text", + "content": "[59] Shijie Zhou, Haoran Chang, Sicheng Jiang, Zhiwen Fan, Zehao Zhu, Dejia Xu, Pradyumna Chari, Suya You, Zhangyang Wang, and Achuta Kadambi. Feature 3dgs: Supercharging 3d gaussian splattering to enable distilled feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21676-21685, 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 140, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 205 + ], + "type": "text", + "content": "[60] Luyang Zhu, Arsalan Mousavian, Yu Xiang, Hammad Mazhar, Jozef van Eenbergen, Shoubhik Debnath, and Dieter Fox. Rgb-d local implicit function for depth completion of transparent objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4649-4658, 2021. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12142" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/ICE_ Intrinsic Concept Extraction from a Single Image via Diffusion Models/e9edff49-034e-4138-82ed-91d11ac96969_content_list.json b/2025/ICE_ Intrinsic Concept Extraction from a Single Image via Diffusion Models/e9edff49-034e-4138-82ed-91d11ac96969_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6312115516560f851f2d9c324fc1169c07aef851 --- /dev/null +++ b/2025/ICE_ Intrinsic Concept Extraction from a Single Image via Diffusion Models/e9edff49-034e-4138-82ed-91d11ac96969_content_list.json @@ -0,0 +1,2332 @@ +[ + { + "type": "text", + "text": "ICE: Intrinsic Concept Extraction from a Single Image via Diffusion Models", + "text_level": 1, + "bbox": [ + 112, + 130, + 883, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fernando Julio Cendra Kai Han", + "bbox": [ + 346, + 180, + 648, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual AI Lab, The University of Hong Kong", + "bbox": [ + 318, + 205, + 676, + 224 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "fcendra@connect.hku.hk, kaihanx@hku.hk", + "bbox": [ + 323, + 227, + 661, + 239 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0101f83f761da9715d805a7df8cef39b4842ec23a4cf74794c7c39908a32de13.jpg", + "image_caption": [ + "Figure 1. We showcase a structured approach for defining visual concepts within an image, where object-level concepts are identified and analyzed to reveal their underlying intrinsic attributes, such as color, material, and shape. We present the ICE (Intrinsic Concept Extraction) framework, which leverages Text-to-Image (T2I) models to systematically discover these concepts, providing a more effective method for learning visual concepts." + ], + "image_footnote": [], + "bbox": [ + 96, + 247, + 906, + 465 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 532, + 326, + 547 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The inherent ambiguity in defining visual concepts poses significant challenges for modern generative models, such as the diffusion-based Text-to-Image (T2I) models, in accurately learning concepts from a single image. Existing methods lack a systematic way to reliably extract the interpretable underlying intrinsic concepts. To address this challenge, we present ICE, short for Intrinsic Concept Extraction, a novel framework that exclusively utilizes a T2I model to automatically and systematically extract intrinsic concepts from a single image. ICE consists of two pivotal stages. In the first stage, ICE devises an automatic concept localization module to pinpoint relevant text-based concepts and their corresponding masks within the image. This critical stage streamlines concept initialization and provides precise guidance for subsequent analysis. The second stage delves deeper into each identified mask, decomposing the object-level concepts into intrinsic concepts and general concepts. This decomposition allows for a more granular and interpretable breakdown of visual elements. Our framework demonstrates superior performance on in", + "bbox": [ + 88, + 566, + 485, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "trinsic concept extraction from a single image in an unsupervised manner. Project page: https://visual-ai.github.io/ice", + "bbox": [ + 511, + 534, + 906, + 579 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 614, + 645, + 630 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text-to-Image (T2I) models have undergone significant advancements, enabling the generation of high-quality images from textual prompts and bridging natural language processing with visual content creation. These capabilities have driven applications in creative design, entertainment, and assistive technologies, underscoring the pivotal role of T2I models in generative AI.", + "bbox": [ + 511, + 641, + 906, + 747 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Beyond image generation, emerging research has uncovered that T2I models excel in tasks such as image classification, segmentation, and semantic correspondence. Studies [10, 14, 16, 18, 20, 30, 33, 35, 37] have shown their effectiveness in these areas. This versatility suggests that diffusion-based T2I models encapsulate substantial world knowledge, allowing them to comprehend and recreate visual concepts with precision and detail. Building on this potential, our work explores the use of T2I models like Stable Diffusion for concept learning. Instead of solely generating", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 109, + 886, + 235, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23734", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "images from text, we investigate how these models invert the generative process to systematically define and learn the fundamental concepts that comprise a given unlabelled image, thereby uncovering the essential concepts that enable the synthesis of complex visual scenes.", + "bbox": [ + 89, + 90, + 480, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite their promising potential, concept learning within T2I models faces several inherent challenges. Defining visual concepts remains inherently ambiguous, complicating the reliability with which these models can discover and utilize essential concepts. Previous methods [7, 8, 26] such as Textual Inversion (TI) and DreamBooth require learning multiple similar images to encapsulate a concept. Other frameworks, such as Break-A-Scene [1], MCPL [13], and ConceptExpress [9], focus on extracting object-level concepts within a single image. Moreover, these methods often struggle to provide clear and interpretable concept representations, hindering their effectiveness in capturing the nuanced and multifaceted nature of visual scenes. The ambiguity in defining visual concepts not only impedes reliable concept discovery but also poses significant challenges for human annotators attempting to delineate these concepts manually.", + "bbox": [ + 91, + 185, + 482, + 441 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Addressing these challenges requires a fundamental re-evaluation of how we define and learn visual concepts within an image. Previous works, such as Inspiration Tree [32], have attempted to construct hierarchical representations by organizing tokens into tree structures to capture different concepts within an image. Similarly, LangInt [15] uses Visual Question Answering (VQA) techniques to guide concept learning, but it requires training separate encoders for each concept axis, thereby failing to fully utilize their inherent hierarchical structures. Consequently, there is a need for more systematic and structured methods to define and learn concepts from images, enabling a more granular understanding of the fundamental building blocks of visual representation.", + "bbox": [ + 89, + 459, + 482, + 671 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In response to these challenges, we present ICE, a novel two-stage framework designed to automatically and systematically extract intrinsic concepts using a single T2I model. ICE first devises an automatic concept localization module to retrieves text-based concepts and their corresponding masks. This enhances initialization and provides clear guidance for subsequent processing. The framework then systematically decomposes these object-level concepts into intrinsic and general concepts, enabling a more granular and interpretable breakdown of visual elements. Through comprehensive experiments, ICE demonstrates its effectiveness in leveraging T2I models to systematically uncover diverse visual concepts, facilitating a structured and detailed understanding of image components.", + "bbox": [ + 89, + 689, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 89, + 653, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-to-Image generation. Diffusion-based methods have significantly advanced the field of text-to-image (T2I) generation, pushing the boundaries of image synthesis to unprecedented levels. Early approaches such as DALL-E 2 [24], Imagen [27], and GLIDE [21] demonstrated the potential of diffusion models in generating high-fidelity images from textual descriptions. Latent Diffusion Models (LDMs) [25] further enhanced this capability by operating in a latent space, enabling more efficient and scalable training. Building upon the foundation laid by LDMs, Stable Diffusion (SD) [25] is trained on the LAION-5B dataset [28] and has set new benchmarks in T2I synthesis.", + "bbox": [ + 511, + 114, + 903, + 296 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Beyond basic image generation, diffusion models have been adapted for a variety of specialized tasks. Controllable generation [21, 22, 38] allows for nuanced adjustments based on user input, while applications such as global and local editing [5, 31] enable fine-grained modifications to specific regions of an image. Additionally, these models have been extended to video generation and editing [11, 29, 34], inpainting [19], and scene generation [2, 3], showcasing their versatility across use cases.", + "bbox": [ + 511, + 296, + 903, + 431 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generative concept learning. Generative concept learning aims to decompose and understand the fundamental elements that constitute complex visual scenes. Textual Inversion (TI) [7] was among the first methods to learn single concepts from multiple images by introducing new words and optimizing their corresponding token embeddings. Similarly, DreamBooth [26] extends this approach by leveraging less frequently used words and optimizing both the denoising U-Net and introducing a preservation loss to enhance the quality of the learned concepts.", + "bbox": [ + 511, + 431, + 903, + 582 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Break-A-Scene [1] relies heavily on human-annotated masks to isolate and learn distinct concepts from a single image, limiting its applicability in settings where such annotations are unavailable Concurrent works like MCPL [13] and DisenDiff [39] address similar challenges but require concept-descriptive text captions or predefined class names, making them less feasible for more general tasks. Recent advancements, such as ConceptExpress [9], have reduced reliance on human annotations by automatically identifying and learning multiple concepts within an image. However, these methods primarily treat concepts at the object-level, neglecting intrinsic attributes such as colour and material properties. Inspiration Tree [32] introduces a preliminary approach to concept decomposition by splitting a token into two tokens to force the learning of different concepts. Despite this, it lacks reliable interpretation and structured guidance, resulting in less coherent concept representations. LangInt [15] and CusConcept [36] attempt to learn intrinsic concepts using VQA models [12, 17] to query specific attributes of an image and independently learn these intrinsic concepts. However, these approaches are limited to scenar", + "bbox": [ + 511, + 584, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23735", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4bd400f44d5ac11e467b49aaa80561f356871c34290918e4e750a0f0a26215c3.jpg", + "image_caption": [ + "Figure 2. Concept definition hierarchy illustrating how object-level concepts are decomposed into intrinsic attributes, including object type, colour, material and other intrinsics." + ], + "image_footnote": [], + "bbox": [ + 145, + 90, + 433, + 253 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ios with a single object-level concept per image, failing to handle images with multiple concepts.", + "bbox": [ + 89, + 313, + 483, + 343 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Comparison and positioning of ICE. In contrast to existing methodologies, our proposed framework, ICE, offers a unified and structured approach to automatically and systematically discover intrinsic concepts within an image using a single T2I model. Unlike previous methods, ICE not only identifies object-level concepts but also decomposes them into intrinsic attributes such as colour and material (see Figure 2), providing a more comprehensive and interpretable representation of visual concepts. This two-stage decomposition ensures that both general and specific attributes are captured, enhancing the model's ability to understand and generate nuanced visual content. For an overview of the differences between our framework and existing methods, please refer to Table 1.", + "bbox": [ + 89, + 344, + 483, + 555 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/eb8b7670f54404bea497338a07ab85a2dfd76c9e4fd76c67d04ddcfac2b29d50.jpg", + "table_caption": [ + "Table 1. Comparison of ICE and relevant works." + ], + "table_footnote": [], + "table_body": "
Learned concept(s)Framework details
MethodObject-level conceptsIntrinsic conceptsMulti conceptsSingle imageExtra information
Textual Inversion [7]✓XXX-
Dreambooth [26]✓XXX-
Inspiration Tree [32]X✓XX-
LangInt [15]✓✓X✓VQA-guided
Break-A-Scene [1]✓X✓✓Mask
MCPL [13]✓X✓✓Text
ConceptExpress [9]✓X✓✓-
ICE (Ours)✓✓✓✓-
", + "bbox": [ + 93, + 582, + 483, + 722 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 752, + 230, + 767 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion models have emerged as a powerful framework for generative tasks, particularly in Text-to-Image (T2I) applications. They operate by iteratively denoising data through a defined forward and reverse diffusion process, enabling the generation of high-fidelity images from noise.", + "bbox": [ + 89, + 777, + 482, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Forward process. In the forward diffusion process, Gaussian noise is incrementally added to an image over a series of timesteps $t = 1,2,\\ldots ,T$ , progressively transforming it", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "into pure noise:", + "bbox": [ + 513, + 92, + 620, + 106 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {t - 1}\\right) = \\mathcal {N} \\left(\\mathbf {x} _ {t}; \\sqrt {1 - \\beta_ {t}} \\mathbf {x} _ {t - 1}, \\beta_ {t} \\mathbf {I}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 109, + 903, + 127 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\beta_{t}$ controls the noise variance at each timestep. We define $\\alpha_{t} = 1 - \\beta_{t}$ , and the cumulative product $\\bar{\\alpha}_{t}$ is given by $\\bar{\\alpha}_{t} = \\prod_{s=1}^{t} \\alpha_{s}$ . Here, $\\alpha_{t}$ represents the remaining signal strength at timestep $t$ , while $\\bar{\\alpha}_{t}$ indicates the cumulative effect, controlling the overall noise level in the image.", + "bbox": [ + 511, + 130, + 903, + 205 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Reverse process. The reverse diffusion process aims to reconstruct the original image by predicting and removing the added noise. This is achieved by training a denoising network $\\epsilon_{\\theta}$ to minimize the denoising loss:", + "bbox": [ + 511, + 205, + 903, + 266 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e c o n}} = \\mathbb {E} _ {\\mathbf {x} _ {0}, \\epsilon , t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (\\mathbf {x} _ {t}, t, \\mathcal {E} (\\mathbf {p})) \\| _ {2} ^ {2} \\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 268, + 903, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{x}_t = \\sqrt{\\bar{\\alpha}_t}\\mathbf{x}_0 + \\sqrt{1 - \\bar{\\alpha}_t\\epsilon}$ and $\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ . The term $\\mathbf{p}$ represents the text prompt used for conditioning via text encoder $\\mathcal{E}$ like CLIP [23].", + "bbox": [ + 511, + 289, + 903, + 334 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The denoising loss function $\\mathcal{L}_{\\mathrm{recon}}$ is crucial for training the diffusion model, guiding $\\epsilon_{\\theta}$ to accurately predict the added noise $\\epsilon$ at each timestep. By minimizing this loss, the model learns to generate clean images from noisy inputs, effectively capturing the underlying data distribution.", + "bbox": [ + 511, + 335, + 903, + 411 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Relation to ICE. ICE leverages the inherent capabilities of diffusion-based T2I models to perform concept extraction and decomposition, enabling a more structured and interpretable approach to visual concept learning.", + "bbox": [ + 511, + 411, + 906, + 472 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 513, + 487, + 604, + 502 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The proposed ICE (Intrinsic Concept Extraction) framework is designed to automatically and systematically discover essential concepts within images using a single T2I diffusion model. The framework operates through a two-stage architecture: Stage One: Automatic Concept Localization and Stage Two: Structured Concept Learning as illustrated in Figure 3. In this section, we provide an overview of each stage, detailing the underlying processes, and training objectives.", + "bbox": [ + 511, + 513, + 905, + 650 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Stage One: Automatic Concept Localization", + "text_level": 1, + "bbox": [ + 511, + 660, + 888, + 676 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This stage is designed to extract object-level concepts from an unlabelled input image automatically. This stage leverages off-the-shelf modules integrated within the T2I diffusion model, ensuring a training-free and seamless concept extraction process. The workflow of Automatic Concept Localization is illustrated in Figure 4.", + "bbox": [ + 511, + 685, + 905, + 775 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text-based concept retrieval. Given an unlabelled input image $\\mathbf{x}$ , the first step involves retrieving relevant text-based concepts from the T2I model's vocabulary. This retrieval is facilitated by the Image-to-Text Retriever, based on T2I's CLIP encoders, denoted as $\\mathcal{T}(\\mathbf{x})$ to decompose dense embeddings into sparse, interpretable semantic concepts. Formally, the retrieval process can be expressed as:", + "bbox": [ + 511, + 776, + 905, + 881 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{c _ {i} \\right\\} = \\mathcal {T} (\\mathbf {x}), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 661, + 885, + 903, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23736", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/acbe1f654db1084933b0f42a343c9e759456c5e05637fdfc0e00778580e8489c.jpg", + "image_caption": [ + "Figure 3. Illustration of the proposed ICE (Intrinsic Concept Extraction) framework, which consists of two stages: (1) Automatic Concept Localization, where a diffusion model is employed to extract object-level concepts and their corresponding masks from an image without prior training, and (2) Structured Concept Learning, where these extracted information are further leveraged to uncover essential concepts." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 911, + 239 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3329133e769ffc614e867a77aa1bc7bb92b4e6ec7bc3eaec8dd4b8af0ae5c83a.jpg", + "image_caption": [ + "Stage One: Automatic Concept Localization" + ], + "image_footnote": [], + "bbox": [ + 109, + 311, + 624, + 467 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d3d65dfbd493db0d530da5faf97029653e6144fea818a3acdba50ededb4fa9d3.jpg", + "image_caption": [ + "Stage One's use case diagram", + "Figure 4. Stage One: Automatic Concept Localization. Starting with an unlabelled image $\\mathbf{x}$ , the Image-to-Text concept extractor retrieves the top-1 text-based concept $c_{i}$ using CLIP encoders. A zero-shot segmentor via T2I model generates the corresponding mask $\\mathbf{m}_{i}$ , and the image is updated by removing the masked region. This process iterates until no objects remain in the image." + ], + "image_footnote": [], + "bbox": [ + 635, + 316, + 901, + 467 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\{c_i\\}$ represents the set of top- $k$ text-based concepts extracted from the image $\\mathbf{x}$ .", + "bbox": [ + 89, + 527, + 482, + 556 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Mask generation via T2I zero-shot segmentor. Once the top-1 text concept $c_{i}$ is identified, the framework employs the Segmentor, denoted as $S(\\mathbf{x},c_i)$ , to obtain a corresponding segmentation mask $\\mathbf{m}_i$ . The mask generation process is formalized as:", + "bbox": [ + 89, + 556, + 483, + 632 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {m} _ {i} = \\mathcal {S} (\\mathbf {x}, c _ {i}), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 233, + 633, + 482, + 650 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{m}_i$ delineates the region of $\\mathbf{x}$ associated with the concept $c_{i}$ . In our implementation, without introducing any external pretrained model, we utilize a training-free image segmentor that leverages the self-attention layers within the pre-trained Stable Diffusion model to produce high-quality segmentation masks without any additional training.", + "bbox": [ + 89, + 657, + 482, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Iterative object extraction. With the mask $\\mathbf{m}_i$ obtained, the corresponding region is removed from the image to prepare for the extraction of subsequent concepts. The updated image $\\mathbf{x}'$ is computed as: $\\mathbf{x}' = \\mathbf{x} \\odot (1 - \\mathbf{m}_i)$ , where $\\odot$ denotes element-wise multiplication. This updated image $\\mathbf{x}'$ is then fed back into $\\mathcal{T}$ and $S$ to iteratively extract and mask remaining objects until no objects remain in the image.", + "bbox": [ + 89, + 750, + 482, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The integration of $\\mathcal{T}$ and $S$ within the T2I diffusion model allows for a fully automated and training-free concept extraction process. This design ensures that the frame", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "work does not rely on any external models beyond the pretrained T2I diffusion model, enhancing its scalability and applicability across diverse datasets. The Automatic Concept Extraction module outputs a set of text-based concepts $\\{c_i\\}$ along with their masks $\\{\\mathbf{m}_i\\}$ for a given image $\\mathbf{x}$ . The pseudocode for this stage is provided in the supplementary material.", + "bbox": [ + 511, + 527, + 906, + 632 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Stage Two: Structured Concept Learning", + "text_level": 1, + "bbox": [ + 511, + 645, + 867, + 662 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This stage focuses on decomposing the extracted object-level concepts into the underlying intrinsic concepts, such as colour and material. To ensure an accurate decomposition of the concepts, we divide this stage into two phases: (1) learning object-level concepts and (2) learning intrinsic concepts. Stage two's workflow is illustrated in Figure 5.", + "bbox": [ + 511, + 670, + 906, + 761 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Phase one: learning object-level concepts. Here, the framework learns both concept-specific and instance-specific tokens for each extracted object-level concept. For each object-level concept $c_{i}$ , two tokens are introduced:", + "bbox": [ + 511, + 762, + 905, + 821 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Concept-specific token ( $c_i^{\\text{conspec}}$ ): Represents the general semantic category of the object.", + "- Instance-specific token $(c_{i}^{\\mathrm{inspec}})$ : Captures attributes unique to the specific instance of the object." + ], + "bbox": [ + 511, + 823, + 903, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To differentiate these tokens, we initialize $c_{i}^{\\mathrm{conspec}}$ em", + "bbox": [ + 531, + 883, + 903, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23737", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fdae20ccaa81735089bb133ae4ddc554178ab9c618626302b86af0523d509be7.jpg", + "image_caption": [ + "Figure 5. Stage Two: Structured Concept Learning. This stage is divided into two phases: (1) learning object-level concepts, where concept-specific $(c_i^{\\mathrm{conspec}})$ and instance-specific $(c_i^{\\mathrm{inspec}})$ tokens are learned using an object-level triplet loss $\\mathcal{L}_{\\mathrm{triplet}}^{\\mathrm{obj}}$ , and (2) learning intrinsic concepts, which decomposes object-level concepts into intrinsic attributes $(c_j^{\\mathrm{intrinsic}})$ using an intrinsic triplet loss $\\mathcal{L}_{\\mathrm{triplet}}^{\\mathrm{intrinsic}}$ . This hierarchical approach ensures accurate separation of general semantic categories from specific and intrinsic attributes." + ], + "image_footnote": [], + "bbox": [ + 99, + 89, + 903, + 282 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "bedding with the extracted text-based concept $c_{i}$ which is obtained in the first stage and employ a triplet loss mechanism, where the $c_{i}$ serves as the anchor, $c_{i}^{\\mathrm{conspec}}$ as the positive, and $c_{i}^{\\mathrm{inspec}}$ as the negative. The triplet loss $\\mathcal{L}_{\\mathrm{triplet}}^{\\mathrm{obj}}$ for object-level concept $c_{i}$ is defined as:", + "bbox": [ + 88, + 359, + 483, + 438 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {t r i p l e t}} ^ {\\mathrm {o b j}} = \\max (0, \\| \\mathcal {E} (\\text {a n c h o r}) - \\mathcal {E} \\left(c _ {i} ^ {\\text {c o n s p e c}}\\right) \\| _ {2} ^ {2} \\\\ - \\left\\| \\mathcal {E} (\\text {a n c h o r}) - \\mathcal {E} \\left(c _ {i} ^ {\\text {i n s p e c}}\\right) \\right\\| _ {2} ^ {2} + \\gamma), \\tag {5} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 439, + 480, + 493 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\gamma$ is the margin parameter, and $\\mathcal{E}(\\cdot)$ denotes the T2I's CLIP text encoder. This loss encourages $c_{i}^{\\mathrm{conspec}}$ to be closer to anchor $c_{i}$ than $c_{i}^{\\mathrm{inspec}}$ , effectively distinguishing between general and instance-specific attributes.", + "bbox": [ + 89, + 494, + 483, + 556 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Phase two: learning intrinsic concepts. The second phase delves into decomposing the learned object-level concepts into intrinsic attributes. For each intrinsic concept (e.g., colour, material), we introduce corresponding intrinsic tokens $c_{j}^{\\mathrm{intrinsic}}$ associated with object-level concepts $c_{i}$ . The intrinsic concept tokens are anchored using combined representations of $c_{i}^{\\mathrm{conspec}}$ and $c_{i}^{\\mathrm{inspec}}$ . Specifically, for an intrinsic concept $j$ , the anchor text $\\mathsf{anchor}_{j}$ is formulated as:", + "bbox": [ + 89, + 566, + 483, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "anchorj:\"a intrinsicconcept", + "bbox": [ + 151, + 715, + 423, + 732 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where intrinsic denotes the specific attribute (e.g., color). The triplet loss $\\mathcal{L}_{\\mathrm{triplet}}^{\\mathrm{intrinsic}}$ for intrinsic concept $j$ is then defined as:", + "bbox": [ + 89, + 761, + 482, + 805 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {t r i p l e t}} ^ {\\text {i n t r i n s i c}} = \\max (0, \\| \\mathcal {E} (\\text {a n c h o r} _ {j}) - \\mathcal {E} (c _ {j} ^ {\\text {i n t r i n s i c}}) \\| _ {2} ^ {2} \\\\ \\left. - \\left\\| \\mathcal {E} (\\operatorname {a n c h o r} _ {j}) - \\mathcal {E} \\left(c _ {k} ^ {\\text {i n t r i n s i c}}\\right) \\right\\| _ {2} ^ {2} + \\gamma\\right), \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 806, + 482, + 852 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $c_k^{\\mathrm{intrinsic}}$ represents intrinsic tokens not associated with intrinsic concept $j$ , ensuring that each intrinsic token is uniquely mapped to its corresponding attribute.", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Concept refinement", + "text_level": 1, + "bbox": [ + 511, + 359, + 702, + 376 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Optimizing intrinsic concepts solely through token embeddings is challenging, often leading to misaligned and inconsistent representations of specific visual attributes such as color and texture. To enhance the accuracy and consistency of these intrinsic concepts, after completing Phase two, we perform concept refinement by fine-tuning both the U-Net denoising network and the text encoder $\\mathcal{E}$ for a limited number of training steps. This fine-tuning ensures that the intrinsic concepts are precisely aligned with their corresponding visual attributes.", + "bbox": [ + 511, + 383, + 906, + 534 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Overall training objective", + "text_level": 1, + "bbox": [ + 511, + 547, + 748, + 564 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The overall training objective of the ICE framework integrates multiple loss components to facilitate both reconstruction quality and concept differentiation. The total loss $\\mathcal{L}_{\\mathrm{total}}$ is given by:", + "bbox": [ + 511, + 571, + 905, + 633 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {r e c o n}} + \\lambda_ {\\text {a t t}} \\mathcal {L} _ {\\text {a t t}} + \\lambda_ {\\text {t r i p l e t}} \\mathcal {L} _ {\\text {t r i p l e t}}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 651, + 903, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\mathcal{L}_{\\mathrm{recon}}$ , as defined in Eq.(2), is the reconstruction loss between the predicted noise and the actual noise with text prompt1 conditioning $\\mathbf{p}$ as:", + "bbox": [ + 511, + 678, + 906, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Phase one, p: \" a photo of $c_{i}^{\\mathrm{inspec}}$ & $c_{i}^{\\mathrm{conspec}}$ \"", + "bbox": [ + 557, + 738, + 859, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Phase two, p: “ a photo of $\\mathbb{C}^{\\mathrm{intrinsic}}$ & $c_{i}^{\\mathrm{conspec}}$ ,", + "bbox": [ + 552, + 775, + 867, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbb{C}^{\\mathrm{intrinsic}}$ is a set of $c_{j}^{\\mathrm{intrinsic}}$ tokens. $\\mathcal{L}_{\\mathrm{att}}$ is the attention mask loss, which employs Wasserstein loss, following [9], to align the attention regions with the segmentation masks:", + "bbox": [ + 511, + 808, + 905, + 854 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {a t t}} = \\mathcal {W} \\left(\\mathbf {A} _ {i}, \\mathbf {m} _ {i}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 642, + 857, + 903, + 875 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "1The list of prompt templates is provided in supplementary material.", + "bbox": [ + 529, + 886, + 892, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23738", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{W}$ denotes the Wasserstein distance, $\\mathbf{A}_i$ is the attention map corresponding to concept $c_{i}$ , and $\\mathbf{m}_i$ is the segmentation mask obtained in the first stage. $\\mathcal{L}_{\\mathrm{triplet}}^{\\mathrm{obj}}$ and $\\mathcal{L}_{\\mathrm{triplet}}^{\\mathrm{intrinsic}}$ are the object-level and intrinsic triplet losses for different phases:", + "bbox": [ + 89, + 90, + 480, + 167 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t r i p l e t}} ^ {\\mathrm {o b j}} = \\sum_ {i} \\mathcal {L} _ {\\text {t r i p l e t}, i} ^ {\\mathrm {o b j}}, \\quad \\mathcal {L} _ {\\text {t r i p l e t}} ^ {\\text {i n t r i n s i c}} = \\sum_ {j} \\mathcal {L} _ {\\text {t r i p l e t}, j} ^ {\\text {i n t r i n s i c}}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 171, + 482, + 204 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The weighting factors $\\lambda_{\\mathrm{att}}$ and $\\lambda_{\\mathrm{triplet}}$ balance the contributions of the attention and triplet losses relative to the reconstruction loss in each phase, ensuring stable optimization across all objectives.", + "bbox": [ + 89, + 207, + 483, + 268 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 284, + 222, + 301 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. We implement our ICE framework using Stable Diffusion v2.1 for both stages. In the first stage, for the Image-to-Text Retriever module $\\mathcal{T}$ , we utilize the CLIP-based SpLiCE framework [4], while the T2I zero-shot Segmentor $S$ employs DiffSeg [30]. We adapt the input and output interfaces of these training-free modules to accommodate our automatic concept localization process. Specifically, modifications were made to ensure seamless integration within the framework, allowing for efficient and accurate extraction of text-based concepts and their corresponding masks. The training process in the second stage is divided into two phases: Phase one focuses on learning object-level concepts and is conducted over 400 training steps, followed by Phase two, which targets intrinsic concepts and also comprises 400 training steps. Subsequently, concept refinement is performed with another 300 steps to enhance the quality and coherence of the learned concepts. We assign the weighting factors $\\lambda_{\\mathrm{att}} = 1\\times 10^{-5}$ and $\\lambda_{\\mathrm{triplet}} = 1$ . For the triplet loss margin $\\gamma$ value, please refer to the supplementary material. Finally, we compute a prior preservation loss [26] to promote diverse image generation while retaining class-specific knowledge. All experiments are conducted on a single NVIDIA RTX 3090 GPU.", + "bbox": [ + 91, + 311, + 483, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baseline & evaluation metrics. To evaluate the effectiveness of ICE, we compare our framework with the state-of-the-art method, ConceptExpress [9], on the Unsupervised Concept Extraction (UCE) benchmarks. Specifically, we utilize $D1$ dataset, which is based on the Unsplash datasets, providing a diverse and extensive collection of unlabeled images. Moreover, the UCE benchmarks assess models based on two primary metrics: (1) Concept Similarity, which includes Identity Similarity $(\\mathrm{SIM}^I)$ measuring how accurately individual concepts are recreated, and Compositional Similarity $(\\mathrm{SIM}^C)$ evaluating the overall coherence of the generated image based on the extracted concepts. (2) Classification Accuracy assesses the disentangle-ment and classification performance of the extracted concepts by evaluating how well the concepts can be used to classify images correctly. These metrics provide a compre", + "bbox": [ + 89, + 659, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "hensive evaluation of both the fidelity and utility of the extracted concepts, enabling a robust quantitative comparison between ICE and ConceptExpress.", + "bbox": [ + 511, + 90, + 905, + 137 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Quantitative results", + "text_level": 1, + "bbox": [ + 511, + 148, + 702, + 165 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Table 2 & 3, ICE achieves higher scores in all metrics compared to previous methods, demonstrating its superior ability to recreate and compose visual concepts from unlabelled images accurately.", + "bbox": [ + 511, + 172, + 905, + 234 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4d235afaed4dd2af41858db8f4523ca88faac1f25cf9a6c683f7c3990256e382.jpg", + "table_caption": [ + "Table 2. Performance of ICE and relevant works on Unsupervised Concept Extraction (UCE) benchmarks using CLIP [23] encoder." + ], + "table_footnote": [], + "table_body": "
Method\\(SIM^I\\)(%)\\(SIM^C\\)(%)\\(ACC^1\\)(%)\\(ACC^3\\)(%)
Break-A-Scene [1]0.6270.7730.1740.282
ConceptExpress[9]0.6890.7840.2630.385
ICE (Ours)0.7380.8220.3250.518
", + "bbox": [ + 522, + 276, + 898, + 340 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/779be91fad999f3df2698f3d56ace517fce84deeb4f53bd6787dbcf64a9d49a3.jpg", + "table_caption": [ + "Table 3. Performance of ICE and relevant works on Unsupervised Concept Extraction (UCE) benchmarks using DINO [6] encoder." + ], + "table_footnote": [], + "table_body": "
Method\\( SIM^I \\) (%)\\( SIM^C \\) (%)\\( ACC^1 \\) (%)\\( ACC^3 \\) (%)
Break-A-Scene [1]0.2540.5100.2020.315
ConceptExpress [9]0.3190.5680.3240.470
ICE (Ours)0.6770.7550.4760.638
", + "bbox": [ + 522, + 398, + 897, + 462 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This enhanced performance can be attributed to two key factors: (1) Enhanced Concept Extraction: Our framework effectively extracts text-based concepts and their corresponding masks, resulting in superior masking accuracy as illustrated in Figure 7. Unlike other methods that may only mask object-level regions, ICE also retrieves the most similar text-based descriptors. This dual extraction provides additional guidance and robust initialization for the model, ensuring that both the spatial and semantic aspects of each concept are accurately captured. (2) Structured Concept Learning Approach: ICE's two-stage framework adopts a structured approach to concept learning, systematically decomposing object-level concepts into intrinsic attributes. This structured methodology facilitates a more granular and interpretable breakdown of visual elements, leading to the discovery of richer and more meaningful concepts. Specifically, ICE achieves CLIP Identity Similarity $(\\mathrm{SIM}^I)$ of $73.8\\%$ , compared to ConceptExpress's $68.9\\%$ , indicating a substantial improvement in accurately identifying object-level concepts. Additionally, ICE attains a Compositional Similarity $(\\mathrm{SIM}^C)$ of $82.2\\%$ , surpassing ConceptExpress's $78.4\\%$ , which underscores our framework's enhanced capability in understanding and composing complex visual attributes. These factors collectively contribute to the superior performance of ICE, highlighting the effectiveness of our framework in systematically uncovering and refining both object-level and intrinsic visual concepts from unlabeled images.", + "bbox": [ + 511, + 478, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23739", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d7508154afc88686741b650f3dc41fd4917c2873e85d0785055d78a68f840d5b.jpg", + "image_caption": [ + "Input image $x$" + ], + "image_footnote": [], + "bbox": [ + 98, + 128, + 254, + 251 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Extracted mask $m_{i}$", + "bbox": [ + 285, + 95, + 346, + 125 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bcfbac42c7da04d2a791750997d3cc8355c35fffe87992dcefecde5e2fa75585.jpg", + "image_caption": [ + "\"Bust\"", + "Extracted mask $m_{i}$" + ], + "image_footnote": [], + "bbox": [ + 277, + 128, + 356, + 186 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Retrived text $c_{i}$", + "bbox": [ + 370, + 95, + 424, + 125 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f6e07ac92b04635b532c70939a6b6d65213a3a6a10e02e6f3610d4d5b365b47e.jpg", + "image_caption": [ + "\"Spout\"" + ], + "image_footnote": [], + "bbox": [ + 277, + 191, + 354, + 250 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/44efb1ab860c9c03cc39f6bef074fb38e6150c951a3d7714e04d64caef0174ba.jpg", + "image_caption": [ + "\"Bust\"" + ], + "image_footnote": [], + "bbox": [ + 277, + 258, + 354, + 316 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a6bd3c7a3d32481cc26c3da3485315c0410d89578b08e7671d9251346f5fdeb9.jpg", + "image_caption": [ + "\"Cone\"" + ], + "image_footnote": [], + "bbox": [ + 277, + 321, + 354, + 381 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/36a554f0b73c52cb401821572c61d2e561654a0848a352feabd644cf0c594ec7.jpg", + "image_caption": [ + "\"Ornament\"" + ], + "image_footnote": [], + "bbox": [ + 277, + 392, + 354, + 450 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7cbf3f514973643e3d2650bf1c801e7969dbad7890632b9aeaa29dc9fb4c8428.jpg", + "image_caption": [ + "\"Rat\"" + ], + "image_footnote": [], + "bbox": [ + 277, + 454, + 354, + 513 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e0a25c1868956268280a0560affd7afd21c604d99cc50c9d034b089ae2b1acb5.jpg", + "image_caption": [ + "\"Wooden\"" + ], + "image_footnote": [], + "bbox": [ + 277, + 523, + 354, + 583 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bd82722cf765e4b2415cfdffdfcdf9250fc13de6a696993cefb7ecdc12ad0cf4.jpg", + "image_caption": [ + "\"Wooden\"" + ], + "image_footnote": [], + "bbox": [ + 277, + 585, + 354, + 645 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/49572d6efb2343757f85df62054c1cf6e7577a07c5524fd4505a6e7421c1eb26.jpg", + "image_caption": [ + "\"Pines\"" + ], + "image_footnote": [], + "bbox": [ + 277, + 655, + 354, + 714 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3d9085d5686db6dc4f8bec30c562a1f9c8982e0cd20461f0c480c71842ba444d.jpg", + "image_caption": [ + "\"Chapter\"", + "${13}/{14}$" + ], + "image_footnote": [], + "bbox": [ + 277, + 717, + 354, + 775 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c68b3da89367bb3cce1f3412d22b7a3ac0df4dde391a5170c9c5bd1427a9619b.jpg", + "image_caption": [ + "Figure 6. Qualitative results of the ICE framework demonstrating its systematic concept discovery process. Column 1: Input images. Column 2: Extracted text-based concepts and their corresponding masks obtained from Stage One: Automatic Concept Localization. Columns 3 & 4: Generated images of learned object-level and intrinsic concepts derived from Stage Two: Structured Concept Learning." + ], + "image_footnote": [], + "bbox": [ + 96, + 656, + 254, + 775 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ed concepts a", + "bbox": [ + 277, + 797, + 354, + 810 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ages of learned object-level and int", + "bbox": [ + 277, + 811, + 482, + 825 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Retrived text $c_{i}$", + "bbox": [ + 370, + 95, + 424, + 125 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ea14d6f07c734134a2480a9b1a421271005bffc9a3ef60491ffc3fc67b25898b.jpg", + "image_caption": [ + "Object-level concepts" + ], + "image_footnote": [], + "bbox": [ + 488, + 128, + 563, + 186 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/71beebb5974c120a3cd1cbf916e712f2128075e21fe3434b03225989ab7f949b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 191, + 562, + 250 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d18b86cbf0ac814f23b776aca8e838f72e1ba53d3f1910ab55379ea6aab93ff8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 258, + 562, + 316 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/39b43670b729c78aba84611f4f4d316fdddf1ecf4780286c0397da1e51053af2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 323, + 562, + 382 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/18d784b5b57590c11339e68f51d7667182fb99ce38ca1b9b2fdafab9408053b4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 390, + 562, + 450 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d0d6473201dbb9e31a141a8075d42b2ebc2d54aa9d61984e22c81be92ff9b0f2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 455, + 562, + 513 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/003698f65df7c20d9323d31332dafeca0fd784babfbf5b2c8194c36139a1a5c7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 523, + 562, + 583 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9e223f4985b4b95143057b15b806b8d245964acb9ce24108a4c51097c29ef86e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 585, + 562, + 645 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d9f8f4a05f3210f49cc6e843082f26f4058dc5f87f669793e8122cf507006427.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 655, + 562, + 714 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/911be1d34710dcafd4c87a8b9c736226e484703c420460aa54783acd676f15e2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 717, + 562, + 775 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cb515daad63eeb70026ccda8c5907a2049bbbe26952d0689b52cf6ae88558c0d.jpg", + "image_caption": [ + "Object" + ], + "image_footnote": [], + "bbox": [ + 611, + 128, + 687, + 186 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/372c56ae06ef95feb11eceb1c39c954e800d011ba6c6e23ea8f9df9d8b260ce8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 611, + 190, + 687, + 250 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b3a31eac67b73c57571526039bc44c6b88fe42c743259d894de3624d5f45fa76.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 611, + 258, + 687, + 316 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/21ba1b0515526447a961e128581ed2c924bf6fe1cf36a9f36e9b7063b2607472.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 611, + 323, + 687, + 382 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7f70af42a461f2b5bc9f206d506465fd2b2436bfe187e1cc7dafa2f4e0c48bd4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 611, + 390, + 687, + 450 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/71abfaadb1912d53f675c13b6865281883cffc2ce0233029f005c54860023a30.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 611, + 455, + 687, + 515 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/116bfc79088af567816e3e5949e5a6f69898e8166ff8673a6e0b452013299be9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 611, + 523, + 687, + 582 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/15e96c0aa635874fbd5bab5e9784c78b16d82b892de738c22f23455fbaec8702.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 611, + 585, + 687, + 645 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5eff3e65c8b7bf434cb4e2550c8498edd1800676e1b88bc81e863b9ce32e4145.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 611, + 655, + 687, + 714 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d912698def2d005067d1c0423f85d827f08d3400ebbe20d22834643948cd7252.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 611, + 717, + 687, + 775 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f4cd31a0f838010406b5dde62ce4e57fdf45e7256aef17ae284ce9eca9f454f2.jpg", + "image_caption": [ + "Intrinsic concepts", + "Material" + ], + "image_footnote": [], + "bbox": [ + 710, + 128, + 785, + 186 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a539b214618624a6761bb76ff0feadbc1f69ea6db88ff343e7c275ff65552160.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 190, + 785, + 250 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9d141318618ed837da517facd02e300133b4e2cff29f4b4412243e54864e60d9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 258, + 785, + 316 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1c2e15953df712ac6d87b5c70059d22e8f23181f53bc19af058a15ed5a36fddf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 321, + 785, + 382 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f821de92b0d31d3623158e9d004c01d649939fc0063f9c0949d6bf5b7a813f9e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 390, + 785, + 450 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5eb9b66ace13319b694f46dcaf92935c5ff9adc6526cdfb580d6f9643f485cbc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 455, + 785, + 515 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/059670f93a9891a330ab07f080b5e7c70b6e8fd7b38c8efc964df1aa5a745373.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 523, + 785, + 583 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3b74a28dfd3bd767ce5a031d4546598d8eb8c2092d75dfaa3d1b844686122b6b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 585, + 785, + 645 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e40a0315a20242769a5ef29622f35843d4011be00ba148f8d9ccf159c55c69aa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 655, + 785, + 714 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/85207812dc6184607e49026fe773046655a14953e73d8ba9eddbc21382bc0004.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 717, + 785, + 775 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/170880547fd7af43f1b2326e1e48ef0e29f0438e7b29cd10e3b01404b293a4f3.jpg", + "image_caption": [ + "Colour" + ], + "image_footnote": [], + "bbox": [ + 810, + 128, + 885, + 186 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/19ee7cad0c1ccfa7e1d2a2f9de75265cd061592667f6bb2085046607b21624e3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 810, + 190, + 885, + 250 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/20182432f1f3fe2a0cee4f04b29d9fc0aa524bfde1dbad8226ae1a4b7f0a4bbd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 810, + 258, + 885, + 316 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9661a9fa5f556f129b80e143e606c157c5979a68dce4b299ed8a03a305278aba.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 810, + 323, + 885, + 382 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/aee9329f8e5b9e7c94d761a72233d516961a49b6a35ee5e6d2b47c2e14d2f252.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 810, + 390, + 885, + 450 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b40f12dd3ebea1b7b49cb1039431d7a8545e91480ab9e9b8708f918eccf86bfe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 810, + 455, + 885, + 515 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a34a703320c4b5337f1cee02e110381fe9a26b72622ca8000dcf59f0302bef19.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 810, + 525, + 885, + 583 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0626c06bea49fafd325040eb7abc9e09a7bc8d029723b8bfd4aae1a494e17ea8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 810, + 585, + 885, + 645 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a268e652d80cbc83a806d189a7690e28e990880eabe087160e7ab42acef969a7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 810, + 655, + 885, + 714 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/86a681c17a33a043643391e023952ea9bef8cc901a14fba327bf9d9f2f53b46d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 810, + 717, + 885, + 775 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Qualitative results", + "text_level": 1, + "bbox": [ + 89, + 837, + 267, + 852 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To demonstrate the effectiveness of the proposed ICE framework, we present qualitative results, as illustrated in", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 6, showcasing ICE's ability to accurately extract and decompose visual concepts systematically.", + "bbox": [ + 511, + 837, + 903, + 868 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "First, given an input image $x$ , ICE extracts corresponding concept masks and retrieves relevant text-based descrip", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23740", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/37c036b79aaa8e5c41fe35427099d65316ad33cc1c883b2af4015cf481ba0993.jpg", + "image_caption": [ + "Figure 7. Qualitative comparison of segmentation results between ConceptExpress and our framework. Our framework not only delivers more accurate segmentation of object-level regions but also effectively retrieves the most relevant text-based concepts, demonstrating its unique capabilities over ConceptExpress." + ], + "image_footnote": [], + "bbox": [ + 93, + 89, + 181, + 155 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/686aa1e4224702a41b9b47c9712e321b29f72ef11b255db4276a1ef42e01dd91.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 181, + 93, + 300, + 155 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/af3a018444c013dd344bd8b0aa21a7989964386c7db0e4a8c3576dcde42e645a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 303, + 90, + 390, + 156 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6276d4649ae5cfbe2ea6796f3d990cff27c819e6a3d4d339dd3f9624845148b2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 392, + 93, + 509, + 155 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9d9515a877f54e7a838fb4cd1e387319d9d87a3a688d78d005a4d91b272398cb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 514, + 90, + 601, + 156 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4494398349677432efd5b785a09799a9d67774bde15135ac5202fa533d181a98.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 602, + 93, + 687, + 155 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b18eac24960c1508e208f030115c4da5b3c2fc22977e01f76e6b7af3934312d3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 90, + 781, + 156 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3d503ed9fda1fbdb20388c9edbb457cc6451466e3dbd555a37cf8205a4df6a16.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 781, + 93, + 800, + 155 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dac1f2ad33523c1e9a7866bcb673db1096a894e110046582bb5bea7c6569832e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 802, + 93, + 900, + 155 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0d3018bee43e09eaee90d61aaeb947fcca13e3203b387e11d881d1efc6271a46.jpg", + "table_caption": [ + "Table 4. Ablation study on ICE model components on the $D1$ dataset." + ], + "table_footnote": [], + "table_body": "
Model's componentsCLIP [23] encoderDINO [6] encoder
MethodStage One's maskStage Two's token init.Stage Two learningSIM1(%)SIMC(%)ACC1(%)ACC3(%)SIM1(%)SIMC(%)ACC1(%)ACC3(%)
ConceptExpressXXX0.6890.7840.2630.3850.3190.5680.3240.470
ICE w. mask✓XX0.7100.7810.3070.4560.4930.6010.3950.604
ICE w.o Stage Two✓✓X0.7260.8070.3010.4520.5010.6210.4110.604
ICE w.o text init.✓X✓0.7220.8140.3200.4750.5480.6430.4490.627
ICE (Ours)✓✓✓0.7380.8220.3250.5180.6770.7550.4760.638
", + "bbox": [ + 99, + 218, + 666, + 318 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4cc642f02c60ebe1ba10c89a50da870c4e5e8c4b2e6397ba8928d51db95aa774.jpg", + "table_caption": [ + "Table 5. Quantitative comparison of generated masks on the $D1$ dataset." + ], + "table_footnote": [], + "table_body": "
MethodmIoURecallPrecision
ConceptExpress0.4830.6760.657
ICE (Ours)0.6350.8930.720
", + "bbox": [ + 674, + 255, + 900, + 306 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tors, as shown in the second column of the figure. These masks precisely delineate regions within the image, ensuring accurate localization of each concept. Next, leveraging the extracted masks and retrieved texts, ICE learns distinct object-level concepts, displayed in the third column. For example, the framework identifies objects such as “Bust” and “Cone” and associates them with their respective semantic descriptors. Finally, utilizing the information from the object-level concepts, ICE systematically discovers intrinsic concepts, including attributes like material and colour, through its second-stage processing. This decomposition enables the framework to uncover compositional intrinsic concepts such as the object type, material, and colour, as demonstrated in the fourth column.", + "bbox": [ + 88, + 329, + 483, + 540 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This structured approach highlights the superiority of our framework by effectively leveraging a single T2I model to systematically uncover both object-level and intrinsic visual concepts. The ability to deconstruct an image into its fundamental components not only validates the robustness of ICE but also illustrates the potential of T2I models to facilitate comprehensive and structured concept learning in an unsupervised manner.", + "bbox": [ + 89, + 542, + 483, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Model component analysis", + "text_level": 1, + "bbox": [ + 89, + 676, + 330, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ICE ablation study. To validate the effectiveness of our ICE framework, we conduct an ablation study on its individual components: Stage One's retrieved mask, Stage Two's token embedding initialization using text retrieved from Stage One, and Stage Two's Structured Concept Learning. The results, as shown in Table 4, demonstrate that each component significantly enhances performance. Furthermore, the complete ICE framework outperforms all its variants across every metric in the UCE benchmarks.", + "bbox": [ + 89, + 702, + 482, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ICE Stage One's mask quality. Furthermore, Table 5 compares our Stage One generated masks with ConceptExpress using $D1$ 's manually annotated ground-truth masks. We evaluate the segmentation performance of generated masks", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "against ground-truth masks using the Hungarian algorithm to find the optimal matching between the sets of masks. We evaluate each pair of predicted and ground-truth masks using segmentation criterions, forming a similarity matrix. The Hungarian algorithm is then applied to this matrix to determine the optimal one-to-one correspondence that maximizes the overall similarity. Overall, our Stage One technique achieves better segmentation results across different criterion. Notably, our Stage One masks exhibit significantly higher recall rates, indicating that a greater proportion of the relevant object-level regions in the annotated area are successfully captured. This higher recall demonstrates that ICE's Automatic Concept Localization method in Stage One is more effective at identifying and extracting meaningful concepts, reducing the likelihood of missing important object-level concepts compared to previous methods.", + "bbox": [ + 511, + 329, + 906, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 604, + 632, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce ICE, a novel framework for automatically and systematically extracting intrinsic concepts from a single image via diffusion models. Our framework consists of two stages. In Stage One, we design a concept localization module based on the T2I model that effectively retrieves the most relevant text-based concepts and their corresponding masks. Subsequently, in Stage Two, we introduce a structured concept learning method to systematically decompose object-level concepts into their intrinsic attributes, resulting in richer and more interpretable concept representations. In summary, our approach effectively harnesses Text-to-Image (T2I) models to enhance concept learning, laying the groundwork for further advancements in intrinsic visual concept extraction.", + "bbox": [ + 511, + 637, + 906, + 848 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. This work is supported by the Hong Kong Research Grants Council - General Research Fund (Grant No.: 17211024).", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23741", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Omri Avrahami, Kfir Aberman, Ohad Fried, Daniel Cohen-Or, and Dani Lischinski. Break-a-scene: Extracting multiple concepts from a single image. In SIGGRAPH Asia, 2023. 2, 3, 6", + "[2] Omri Avrahami, Thomas Hayes, Oran Gafni, Sonal Gupta, Yaniv Taigman, Devi Parikh, Dani Lischinski, Ohad Fried, and Xi Yin. Spatext: Spatio-textual representation for controllable image generation. In CVPR, 2023. 2", + "[3] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. In ICML, 2023. 2", + "[4] Usha Bhalla, Alex Oesterling, Suraj Srinivas, Flavio P Calmon, and Himabindu Lakkaraju. Interpreting clip with sparse linear concept embeddings (splice). In NeurIPS, 2024. 6", + "[5] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023. 2", + "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 6, 8", + "[7] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In ICLR, 2023. 2, 3", + "[8] Shaozhe Hao, Kai Han, Shihao Zhao, and Kwan-Yee K Wong. Vico: Detail-preserving visual condition for personalized text-to-image generation. arXiv preprint arXiv:2306.00971, 2023. 2", + "[9] Shaozhe Hao, Kai Han, Zhengyao Lv, Shihao Zhao, and Kwan-Yee K Wong. Conceptexpress: Harnessing diffusion models for single-image unsupervised concept extraction. In ECCV, 2024. 2, 3, 5, 6", + "[10] Eric Hedlin, Gopal Sharma, Shweta Mahajan, Hossam Isack, Abhishek Kar, Andrea Tagliasacchi, and Kwang Moo Yi. Unsupervised semantic correspondence using stable diffusion. In NeurIPS, 2023. 1", + "[11] Jonathan Ho, Tim Salimans, Alexey A Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. In NeurIPS, 2022. 2", + "[12] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Weli-hinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 2", + "[13] Chen Jin, Ryutaro Tanno, Amrutha Saseendran, Tom Diethe, and Philip Teare. An image is worth multiple words: Discovering object level concepts using multi-concept prompt learning. In ICML, 2023. 2, 3", + "[14] Laurynas Karazija, Iro Laina, Andrea Vedaldi, and Christian Rupprecht. Diffusion models for zero-shot open-vocabulary segmentation. In ECCV, 2023. 1", + "[15] Sharon Lee, Yunzhi Zhang, Shangzhe Wu, and Jiajun Wu. Language-informed visual concept learning. In ICLR, 2023. 2, 3" + ], + "bbox": [ + 93, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Alexander C. Li, Mihir Prabhudesai, Shivam Duggal, Ellis Brown, and Deepak Pathak. Your diffusion model is secretly a zero-shot classifier. In ICCV, 2023. 1", + "[17] Dongxu Li, Junnan Li, and Steven CH Hoi. Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. In NeurIPS, 2023. 2", + "[18] Xinghui Li, Jingyi Lu, Kai Han, and Victor Prisacariu. Sd4match: Learning to prompt stable diffusion model for semantic matching. In CVPR, 2024. 1", + "[19] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In CVPR, 2022. 2", + "[20] Minheng Ni, Yabo Zhang, Kailai Feng, Xiaoming Li, Yiwen Guo, and Wangmeng Zuo. Ref-diff: Zero-shot referring image segmentation with generative models. arXiv preprint arXiv:2308.16777, 2023. 1", + "[21] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob Mcgrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In ICML, 2022. 2", + "[22] Zeju Qiu, Weiyang Liu, Haiwen Feng, Yuxuan Xue, Yao Feng, Zhen Liu, Dan Zhang, Adrian Weller, and Bernhard Scholkopf. Controlling text-to-image diffusion by orthogonal finetuning. In NeurIPS, 2023. 2", + "[23] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 3, 6, 8", + "[24] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents. arXiv preprint arXiv:2204.06125, 2022. 2", + "[25] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2", + "[26] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, 2023. 2, 3, 6", + "[27] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, 2022. 2", + "[28] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS, 2022. 2", + "[29] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. In ICLR, 2022. 2", + "[30] Junjiao Tian, Lavisha Aggarwal, Andrea Colaco, Zsolt Kira, and Mar Gonzalez-Franco. Diffuse, attend, and segment:" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23742", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Unsupervised zero-shot segmentation using stable diffusion. In CVPR, 2024. 1, 6", + "[31] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In CVPR, 2023. 2", + "[32] Yael Vinker, Andrey Voynov, Daniel Cohen-Or, and Ariel Shamir. Concept decomposition for visual exploration and inspiration. In SIGGRAPH Asia, 2023. 2, 3", + "[33] Jinglong Wang, Xiawei Li, Jing Zhang, Qingyuan Xu, Qin Zhou, Qian Yu, Lu Sheng, and Dong Xu. Diffusion model is secretly a training-free open vocabulary semantic segmenter. arXiv preprint arXiv:2309.02773, 2023. 1", + "[34] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Weixian Lei, Yuchao Gu, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In ICCV, 2022. 2", + "[35] Changming Xiao, Qi Yang, Feng Zhou, and Changshui Zhang. From text to mask: Localizing entities using the attention of text-to-image diffusion models. In Neurocomputing, 2023. 1", + "[36] Zhi Xu, Shaozhe Hao, and Kai Han. Cusconcept: Customized visual concept decomposition with diffusion models. In WACV, 2025. 2", + "[37] Junyi Zhang, Charles Herrmann, Junhwa Hur, Luisa Polania Cabrera, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. A tale of two features: Stable diffusion complements dino for zero-shot semantic correspondence. In NeurIPS, 2023. 1", + "[38] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 2", + "[39] Yanbing Zhang, Mengping Yang, Qin Zhou, and Zhe Wang. Attention calibration for disentangled text-to-image personalization. In CVPR, 2024. 2" + ], + "bbox": [ + 91, + 90, + 482, + 573 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23743", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/ICE_ Intrinsic Concept Extraction from a Single Image via Diffusion Models/e9edff49-034e-4138-82ed-91d11ac96969_model.json b/2025/ICE_ Intrinsic Concept Extraction from a Single Image via Diffusion Models/e9edff49-034e-4138-82ed-91d11ac96969_model.json new file mode 100644 index 0000000000000000000000000000000000000000..de8d65934721eb71b84a9073ade0138293526ff3 --- /dev/null +++ b/2025/ICE_ Intrinsic Concept Extraction from a Single Image via Diffusion Models/e9edff49-034e-4138-82ed-91d11ac96969_model.json @@ -0,0 +1,2882 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.131, + 0.885, + 0.154 + ], + "angle": 0, + "content": "ICE: Intrinsic Concept Extraction from a Single Image via Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.348, + 0.181, + 0.65, + 0.199 + ], + "angle": 0, + "content": "Fernando Julio Cendra Kai Han" + }, + { + "type": "text", + "bbox": [ + 0.32, + 0.206, + 0.678, + 0.225 + ], + "angle": 0, + "content": "Visual AI Lab, The University of Hong Kong" + }, + { + "type": "text", + "bbox": [ + 0.325, + 0.228, + 0.663, + 0.241 + ], + "angle": 0, + "content": "fcendra@connect.hku.hk, kaihanx@hku.hk" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.248, + 0.907, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.467, + 0.907, + 0.523 + ], + "angle": 0, + "content": "Figure 1. We showcase a structured approach for defining visual concepts within an image, where object-level concepts are identified and analyzed to reveal their underlying intrinsic attributes, such as color, material, and shape. We present the ICE (Intrinsic Concept Extraction) framework, which leverages Text-to-Image (T2I) models to systematically discover these concepts, providing a more effective method for learning visual concepts." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.533, + 0.327, + 0.549 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.567, + 0.486, + 0.87 + ], + "angle": 0, + "content": "The inherent ambiguity in defining visual concepts poses significant challenges for modern generative models, such as the diffusion-based Text-to-Image (T2I) models, in accurately learning concepts from a single image. Existing methods lack a systematic way to reliably extract the interpretable underlying intrinsic concepts. To address this challenge, we present ICE, short for Intrinsic Concept Extraction, a novel framework that exclusively utilizes a T2I model to automatically and systematically extract intrinsic concepts from a single image. ICE consists of two pivotal stages. In the first stage, ICE devises an automatic concept localization module to pinpoint relevant text-based concepts and their corresponding masks within the image. This critical stage streamlines concept initialization and provides precise guidance for subsequent analysis. The second stage delves deeper into each identified mask, decomposing the object-level concepts into intrinsic concepts and general concepts. This decomposition allows for a more granular and interpretable breakdown of visual elements. Our framework demonstrates superior performance on in" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.535, + 0.907, + 0.58 + ], + "angle": 0, + "content": "trinsic concept extraction from a single image in an unsupervised manner. Project page: https://visual-ai.github.io/ice" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.615, + 0.646, + 0.631 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.642, + 0.907, + 0.748 + ], + "angle": 0, + "content": "Text-to-Image (T2I) models have undergone significant advancements, enabling the generation of high-quality images from textual prompts and bridging natural language processing with visual content creation. These capabilities have driven applications in creative design, entertainment, and assistive technologies, underscoring the pivotal role of T2I models in generative AI." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Beyond image generation, emerging research has uncovered that T2I models excel in tasks such as image classification, segmentation, and semantic correspondence. Studies [10, 14, 16, 18, 20, 30, 33, 35, 37] have shown their effectiveness in these areas. This versatility suggests that diffusion-based T2I models encapsulate substantial world knowledge, allowing them to comprehend and recreate visual concepts with precision and detail. Building on this potential, our work explores the use of T2I models like Stable Diffusion for concept learning. Instead of solely generating" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.887, + 0.236, + 0.901 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23734" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.168 + ], + "angle": 0, + "content": "images from text, we investigate how these models invert the generative process to systematically define and learn the fundamental concepts that comprise a given unlabelled image, thereby uncovering the essential concepts that enable the synthesis of complex visual scenes." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.186, + 0.483, + 0.442 + ], + "angle": 0, + "content": "Despite their promising potential, concept learning within T2I models faces several inherent challenges. Defining visual concepts remains inherently ambiguous, complicating the reliability with which these models can discover and utilize essential concepts. Previous methods [7, 8, 26] such as Textual Inversion (TI) and DreamBooth require learning multiple similar images to encapsulate a concept. Other frameworks, such as Break-A-Scene [1], MCPL [13], and ConceptExpress [9], focus on extracting object-level concepts within a single image. Moreover, these methods often struggle to provide clear and interpretable concept representations, hindering their effectiveness in capturing the nuanced and multifaceted nature of visual scenes. The ambiguity in defining visual concepts not only impedes reliable concept discovery but also poses significant challenges for human annotators attempting to delineate these concepts manually." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.46, + 0.483, + 0.672 + ], + "angle": 0, + "content": "Addressing these challenges requires a fundamental re-evaluation of how we define and learn visual concepts within an image. Previous works, such as Inspiration Tree [32], have attempted to construct hierarchical representations by organizing tokens into tree structures to capture different concepts within an image. Similarly, LangInt [15] uses Visual Question Answering (VQA) techniques to guide concept learning, but it requires training separate encoders for each concept axis, thereby failing to fully utilize their inherent hierarchical structures. Consequently, there is a need for more systematic and structured methods to define and learn concepts from images, enabling a more granular understanding of the fundamental building blocks of visual representation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In response to these challenges, we present ICE, a novel two-stage framework designed to automatically and systematically extract intrinsic concepts using a single T2I model. ICE first devises an automatic concept localization module to retrieves text-based concepts and their corresponding masks. This enhances initialization and provides clear guidance for subsequent processing. The framework then systematically decomposes these object-level concepts into intrinsic and general concepts, enabling a more granular and interpretable breakdown of visual elements. Through comprehensive experiments, ICE demonstrates its effectiveness in leveraging T2I models to systematically uncover diverse visual concepts, facilitating a structured and detailed understanding of image components." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.655, + 0.107 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.115, + 0.905, + 0.297 + ], + "angle": 0, + "content": "Text-to-Image generation. Diffusion-based methods have significantly advanced the field of text-to-image (T2I) generation, pushing the boundaries of image synthesis to unprecedented levels. Early approaches such as DALL-E 2 [24], Imagen [27], and GLIDE [21] demonstrated the potential of diffusion models in generating high-fidelity images from textual descriptions. Latent Diffusion Models (LDMs) [25] further enhanced this capability by operating in a latent space, enabling more efficient and scalable training. Building upon the foundation laid by LDMs, Stable Diffusion (SD) [25] is trained on the LAION-5B dataset [28] and has set new benchmarks in T2I synthesis." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.297, + 0.905, + 0.432 + ], + "angle": 0, + "content": "Beyond basic image generation, diffusion models have been adapted for a variety of specialized tasks. Controllable generation [21, 22, 38] allows for nuanced adjustments based on user input, while applications such as global and local editing [5, 31] enable fine-grained modifications to specific regions of an image. Additionally, these models have been extended to video generation and editing [11, 29, 34], inpainting [19], and scene generation [2, 3], showcasing their versatility across use cases." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.433, + 0.905, + 0.583 + ], + "angle": 0, + "content": "Generative concept learning. Generative concept learning aims to decompose and understand the fundamental elements that constitute complex visual scenes. Textual Inversion (TI) [7] was among the first methods to learn single concepts from multiple images by introducing new words and optimizing their corresponding token embeddings. Similarly, DreamBooth [26] extends this approach by leveraging less frequently used words and optimizing both the denoising U-Net and introducing a preservation loss to enhance the quality of the learned concepts." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.585, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Break-A-Scene [1] relies heavily on human-annotated masks to isolate and learn distinct concepts from a single image, limiting its applicability in settings where such annotations are unavailable Concurrent works like MCPL [13] and DisenDiff [39] address similar challenges but require concept-descriptive text captions or predefined class names, making them less feasible for more general tasks. Recent advancements, such as ConceptExpress [9], have reduced reliance on human annotations by automatically identifying and learning multiple concepts within an image. However, these methods primarily treat concepts at the object-level, neglecting intrinsic attributes such as colour and material properties. Inspiration Tree [32] introduces a preliminary approach to concept decomposition by splitting a token into two tokens to force the learning of different concepts. Despite this, it lacks reliable interpretation and structured guidance, resulting in less coherent concept representations. LangInt [15] and CusConcept [36] attempt to learn intrinsic concepts using VQA models [12, 17] to query specific attributes of an image and independently learn these intrinsic concepts. However, these approaches are limited to scenar" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23735" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.146, + 0.092, + 0.434, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.262, + 0.486, + 0.306 + ], + "angle": 0, + "content": "Figure 2. Concept definition hierarchy illustrating how object-level concepts are decomposed into intrinsic attributes, including object type, colour, material and other intrinsics." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.314, + 0.484, + 0.344 + ], + "angle": 0, + "content": "ios with a single object-level concept per image, failing to handle images with multiple concepts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.345, + 0.485, + 0.556 + ], + "angle": 0, + "content": "Comparison and positioning of ICE. In contrast to existing methodologies, our proposed framework, ICE, offers a unified and structured approach to automatically and systematically discover intrinsic concepts within an image using a single T2I model. Unlike previous methods, ICE not only identifies object-level concepts but also decomposes them into intrinsic attributes such as colour and material (see Figure 2), providing a more comprehensive and interpretable representation of visual concepts. This two-stage decomposition ensures that both general and specific attributes are captured, enhancing the model's ability to understand and generate nuanced visual content. For an overview of the differences between our framework and existing methods, please refer to Table 1." + }, + { + "type": "table_caption", + "bbox": [ + 0.14, + 0.568, + 0.429, + 0.581 + ], + "angle": 0, + "content": "Table 1. Comparison of ICE and relevant works." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.583, + 0.485, + 0.723 + ], + "angle": 0, + "content": "
Learned concept(s)Framework details
MethodObject-level conceptsIntrinsic conceptsMulti conceptsSingle imageExtra information
Textual Inversion [7]✓XXX-
Dreambooth [26]✓XXX-
Inspiration Tree [32]X✓XX-
LangInt [15]✓✓X✓VQA-guided
Break-A-Scene [1]✓X✓✓Mask
MCPL [13]✓X✓✓Text
ConceptExpress [9]✓X✓✓-
ICE (Ours)✓✓✓✓-
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.753, + 0.231, + 0.768 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.779, + 0.483, + 0.856 + ], + "angle": 0, + "content": "Diffusion models have emerged as a powerful framework for generative tasks, particularly in Text-to-Image (T2I) applications. They operate by iteratively denoising data through a defined forward and reverse diffusion process, enabling the generation of high-fidelity images from noise." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Forward process. In the forward diffusion process, Gaussian noise is incrementally added to an image over a series of timesteps \\(t = 1,2,\\ldots ,T\\), progressively transforming it" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.093, + 0.622, + 0.107 + ], + "angle": 0, + "content": "into pure noise:" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.11, + 0.905, + 0.128 + ], + "angle": 0, + "content": "\\[\nq \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {t - 1}\\right) = \\mathcal {N} \\left(\\mathbf {x} _ {t}; \\sqrt {1 - \\beta_ {t}} \\mathbf {x} _ {t - 1}, \\beta_ {t} \\mathbf {I}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.131, + 0.905, + 0.206 + ], + "angle": 0, + "content": "where \\(\\beta_{t}\\) controls the noise variance at each timestep. We define \\(\\alpha_{t} = 1 - \\beta_{t}\\), and the cumulative product \\(\\bar{\\alpha}_{t}\\) is given by \\(\\bar{\\alpha}_{t} = \\prod_{s=1}^{t} \\alpha_{s}\\). Here, \\(\\alpha_{t}\\) represents the remaining signal strength at timestep \\(t\\), while \\(\\bar{\\alpha}_{t}\\) indicates the cumulative effect, controlling the overall noise level in the image." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.207, + 0.905, + 0.267 + ], + "angle": 0, + "content": "Reverse process. The reverse diffusion process aims to reconstruct the original image by predicting and removing the added noise. This is achieved by training a denoising network \\(\\epsilon_{\\theta}\\) to minimize the denoising loss:" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.269, + 0.905, + 0.288 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e c o n}} = \\mathbb {E} _ {\\mathbf {x} _ {0}, \\epsilon , t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (\\mathbf {x} _ {t}, t, \\mathcal {E} (\\mathbf {p})) \\| _ {2} ^ {2} \\right], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.29, + 0.905, + 0.335 + ], + "angle": 0, + "content": "where \\(\\mathbf{x}_t = \\sqrt{\\bar{\\alpha}_t}\\mathbf{x}_0 + \\sqrt{1 - \\bar{\\alpha}_t\\epsilon}\\) and \\(\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\). The term \\(\\mathbf{p}\\) represents the text prompt used for conditioning via text encoder \\(\\mathcal{E}\\) like CLIP [23]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.336, + 0.905, + 0.412 + ], + "angle": 0, + "content": "The denoising loss function \\(\\mathcal{L}_{\\mathrm{recon}}\\) is crucial for training the diffusion model, guiding \\(\\epsilon_{\\theta}\\) to accurately predict the added noise \\(\\epsilon\\) at each timestep. By minimizing this loss, the model learns to generate clean images from noisy inputs, effectively capturing the underlying data distribution." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.412, + 0.907, + 0.473 + ], + "angle": 0, + "content": "Relation to ICE. ICE leverages the inherent capabilities of diffusion-based T2I models to perform concept extraction and decomposition, enabling a more structured and interpretable approach to visual concept learning." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.488, + 0.605, + 0.503 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.514, + 0.906, + 0.651 + ], + "angle": 0, + "content": "The proposed ICE (Intrinsic Concept Extraction) framework is designed to automatically and systematically discover essential concepts within images using a single T2I diffusion model. The framework operates through a two-stage architecture: Stage One: Automatic Concept Localization and Stage Two: Structured Concept Learning as illustrated in Figure 3. In this section, we provide an overview of each stage, detailing the underlying processes, and training objectives." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.661, + 0.889, + 0.677 + ], + "angle": 0, + "content": "4.1. Stage One: Automatic Concept Localization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.686, + 0.906, + 0.776 + ], + "angle": 0, + "content": "This stage is designed to extract object-level concepts from an unlabelled input image automatically. This stage leverages off-the-shelf modules integrated within the T2I diffusion model, ensuring a training-free and seamless concept extraction process. The workflow of Automatic Concept Localization is illustrated in Figure 4." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.777, + 0.906, + 0.882 + ], + "angle": 0, + "content": "Text-based concept retrieval. Given an unlabelled input image \\(\\mathbf{x}\\), the first step involves retrieving relevant text-based concepts from the T2I model's vocabulary. This retrieval is facilitated by the Image-to-Text Retriever, based on T2I's CLIP encoders, denoted as \\(\\mathcal{T}(\\mathbf{x})\\) to decompose dense embeddings into sparse, interpretable semantic concepts. Formally, the retrieval process can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.663, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\left\\{c _ {i} \\right\\} = \\mathcal {T} (\\mathbf {x}), \\tag {3}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "23736" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.089, + 0.912, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.243, + 0.908, + 0.288 + ], + "angle": 0, + "content": "Figure 3. Illustration of the proposed ICE (Intrinsic Concept Extraction) framework, which consists of two stages: (1) Automatic Concept Localization, where a diffusion model is employed to extract object-level concepts and their corresponding masks from an image without prior training, and (2) Structured Concept Learning, where these extracted information are further leveraged to uncover essential concepts." + }, + { + "type": "image_caption", + "bbox": [ + 0.1, + 0.293, + 0.401, + 0.31 + ], + "angle": 0, + "content": "Stage One: Automatic Concept Localization" + }, + { + "type": "image", + "bbox": [ + 0.11, + 0.313, + 0.625, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.671, + 0.298, + 0.877, + 0.314 + ], + "angle": 0, + "content": "Stage One's use case diagram" + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.317, + 0.903, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.474, + 0.908, + 0.517 + ], + "angle": 0, + "content": "Figure 4. Stage One: Automatic Concept Localization. Starting with an unlabelled image \\(\\mathbf{x}\\), the Image-to-Text concept extractor retrieves the top-1 text-based concept \\(c_{i}\\) using CLIP encoders. A zero-shot segmentor via T2I model generates the corresponding mask \\(\\mathbf{m}_{i}\\), and the image is updated by removing the masked region. This process iterates until no objects remain in the image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.528, + 0.483, + 0.558 + ], + "angle": 0, + "content": "where \\(\\{c_i\\}\\) represents the set of top-\\(k\\) text-based concepts extracted from the image \\(\\mathbf{x}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.558, + 0.484, + 0.633 + ], + "angle": 0, + "content": "Mask generation via T2I zero-shot segmentor. Once the top-1 text concept \\( c_{i} \\) is identified, the framework employs the Segmentor, denoted as \\( S(\\mathbf{x},c_i) \\), to obtain a corresponding segmentation mask \\( \\mathbf{m}_i \\). The mask generation process is formalized as:" + }, + { + "type": "equation", + "bbox": [ + 0.234, + 0.635, + 0.483, + 0.651 + ], + "angle": 0, + "content": "\\[\n\\mathbf {m} _ {i} = \\mathcal {S} (\\mathbf {x}, c _ {i}), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.483, + 0.749 + ], + "angle": 0, + "content": "where \\(\\mathbf{m}_i\\) delineates the region of \\(\\mathbf{x}\\) associated with the concept \\(c_{i}\\). In our implementation, without introducing any external pretrained model, we utilize a training-free image segmentor that leverages the self-attention layers within the pre-trained Stable Diffusion model to produce high-quality segmentation masks without any additional training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.856 + ], + "angle": 0, + "content": "Iterative object extraction. With the mask \\(\\mathbf{m}_i\\) obtained, the corresponding region is removed from the image to prepare for the extraction of subsequent concepts. The updated image \\(\\mathbf{x}'\\) is computed as: \\(\\mathbf{x}' = \\mathbf{x} \\odot (1 - \\mathbf{m}_i)\\), where \\(\\odot\\) denotes element-wise multiplication. This updated image \\(\\mathbf{x}'\\) is then fed back into \\(\\mathcal{T}\\) and \\(S\\) to iteratively extract and mask remaining objects until no objects remain in the image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "The integration of \\(\\mathcal{T}\\) and \\(S\\) within the T2I diffusion model allows for a fully automated and training-free concept extraction process. This design ensures that the frame" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.528, + 0.907, + 0.633 + ], + "angle": 0, + "content": "work does not rely on any external models beyond the pretrained T2I diffusion model, enhancing its scalability and applicability across diverse datasets. The Automatic Concept Extraction module outputs a set of text-based concepts \\(\\{c_i\\}\\) along with their masks \\(\\{\\mathbf{m}_i\\}\\) for a given image \\(\\mathbf{x}\\). The pseudocode for this stage is provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.646, + 0.869, + 0.663 + ], + "angle": 0, + "content": "4.2. Stage Two: Structured Concept Learning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.671, + 0.907, + 0.762 + ], + "angle": 0, + "content": "This stage focuses on decomposing the extracted object-level concepts into the underlying intrinsic concepts, such as colour and material. To ensure an accurate decomposition of the concepts, we divide this stage into two phases: (1) learning object-level concepts and (2) learning intrinsic concepts. Stage two's workflow is illustrated in Figure 5." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.763, + 0.906, + 0.823 + ], + "angle": 0, + "content": "Phase one: learning object-level concepts. Here, the framework learns both concept-specific and instance-specific tokens for each extracted object-level concept. For each object-level concept \\( c_{i} \\), two tokens are introduced:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.824, + 0.905, + 0.854 + ], + "angle": 0, + "content": "- Concept-specific token (\\( c_i^{\\text{conspec}} \\)): Represents the general semantic category of the object." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.854, + 0.905, + 0.884 + ], + "angle": 0, + "content": "- Instance-specific token \\((c_{i}^{\\mathrm{inspec}})\\): Captures attributes unique to the specific instance of the object." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.824, + 0.905, + 0.884 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.885, + 0.905, + 0.902 + ], + "angle": 0, + "content": "To differentiate these tokens, we initialize \\( c_{i}^{\\mathrm{conspec}} \\) em" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23737" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.09, + 0.905, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.291, + 0.908, + 0.351 + ], + "angle": 0, + "content": "Figure 5. Stage Two: Structured Concept Learning. This stage is divided into two phases: (1) learning object-level concepts, where concept-specific \\((c_i^{\\mathrm{conspec}})\\) and instance-specific \\((c_i^{\\mathrm{inspec}})\\) tokens are learned using an object-level triplet loss \\(\\mathcal{L}_{\\mathrm{triplet}}^{\\mathrm{obj}}\\), and (2) learning intrinsic concepts, which decomposes object-level concepts into intrinsic attributes \\((c_j^{\\mathrm{intrinsic}})\\) using an intrinsic triplet loss \\(\\mathcal{L}_{\\mathrm{triplet}}^{\\mathrm{intrinsic}}\\). This hierarchical approach ensures accurate separation of general semantic categories from specific and intrinsic attributes." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.361, + 0.484, + 0.439 + ], + "angle": 0, + "content": "bedding with the extracted text-based concept \\( c_{i} \\) which is obtained in the first stage and employ a triplet loss mechanism, where the \\( c_{i} \\) serves as the anchor, \\( c_{i}^{\\mathrm{conspec}} \\) as the positive, and \\( c_{i}^{\\mathrm{inspec}} \\) as the negative. The triplet loss \\( \\mathcal{L}_{\\mathrm{triplet}}^{\\mathrm{obj}} \\) for object-level concept \\( c_{i} \\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.113, + 0.44, + 0.482, + 0.494 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {t r i p l e t}} ^ {\\mathrm {o b j}} = \\max (0, \\| \\mathcal {E} (\\text {a n c h o r}) - \\mathcal {E} \\left(c _ {i} ^ {\\text {c o n s p e c}}\\right) \\| _ {2} ^ {2} \\\\ - \\left\\| \\mathcal {E} (\\text {a n c h o r}) - \\mathcal {E} \\left(c _ {i} ^ {\\text {i n s p e c}}\\right) \\right\\| _ {2} ^ {2} + \\gamma), \\tag {5} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.496, + 0.484, + 0.558 + ], + "angle": 0, + "content": "where \\(\\gamma\\) is the margin parameter, and \\(\\mathcal{E}(\\cdot)\\) denotes the T2I's CLIP text encoder. This loss encourages \\(c_{i}^{\\mathrm{conspec}}\\) to be closer to anchor \\(c_{i}\\) than \\(c_{i}^{\\mathrm{inspec}}\\), effectively distinguishing between general and instance-specific attributes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.567, + 0.484, + 0.688 + ], + "angle": 0, + "content": "Phase two: learning intrinsic concepts. The second phase delves into decomposing the learned object-level concepts into intrinsic attributes. For each intrinsic concept (e.g., colour, material), we introduce corresponding intrinsic tokens \\( c_{j}^{\\mathrm{intrinsic}} \\) associated with object-level concepts \\( c_{i} \\). The intrinsic concept tokens are anchored using combined representations of \\( c_{i}^{\\mathrm{conspec}} \\) and \\( c_{i}^{\\mathrm{inspec}} \\). Specifically, for an intrinsic concept \\( j \\), the anchor text \\( \\mathsf{anchor}_{j} \\) is formulated as:" + }, + { + "type": "text", + "bbox": [ + 0.152, + 0.716, + 0.424, + 0.733 + ], + "angle": 0, + "content": "anchorj:\"a intrinsicconcept" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.762, + 0.483, + 0.806 + ], + "angle": 0, + "content": "where intrinsic denotes the specific attribute (e.g., color). The triplet loss \\(\\mathcal{L}_{\\mathrm{triplet}}^{\\mathrm{intrinsic}}\\) for intrinsic concept \\(j\\) is then defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.807, + 0.483, + 0.853 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {t r i p l e t}} ^ {\\text {i n t r i n s i c}} = \\max (0, \\| \\mathcal {E} (\\text {a n c h o r} _ {j}) - \\mathcal {E} (c _ {j} ^ {\\text {i n t r i n s i c}}) \\| _ {2} ^ {2} \\\\ \\left. - \\left\\| \\mathcal {E} (\\operatorname {a n c h o r} _ {j}) - \\mathcal {E} \\left(c _ {k} ^ {\\text {i n t r i n s i c}}\\right) \\right\\| _ {2} ^ {2} + \\gamma\\right), \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.855, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\( c_k^{\\mathrm{intrinsic}} \\) represents intrinsic tokens not associated with intrinsic concept \\( j \\), ensuring that each intrinsic token is uniquely mapped to its corresponding attribute." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.36, + 0.704, + 0.377 + ], + "angle": 0, + "content": "4.3. Concept refinement" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.385, + 0.907, + 0.535 + ], + "angle": 0, + "content": "Optimizing intrinsic concepts solely through token embeddings is challenging, often leading to misaligned and inconsistent representations of specific visual attributes such as color and texture. To enhance the accuracy and consistency of these intrinsic concepts, after completing Phase two, we perform concept refinement by fine-tuning both the U-Net denoising network and the text encoder \\(\\mathcal{E}\\) for a limited number of training steps. This fine-tuning ensures that the intrinsic concepts are precisely aligned with their corresponding visual attributes." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.548, + 0.749, + 0.565 + ], + "angle": 0, + "content": "4.4. Overall training objective" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.572, + 0.906, + 0.634 + ], + "angle": 0, + "content": "The overall training objective of the ICE framework integrates multiple loss components to facilitate both reconstruction quality and concept differentiation. The total loss \\(\\mathcal{L}_{\\mathrm{total}}\\) is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.578, + 0.652, + 0.905, + 0.669 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {r e c o n}} + \\lambda_ {\\text {a t t}} \\mathcal {L} _ {\\text {a t t}} + \\lambda_ {\\text {t r i p l e t}} \\mathcal {L} _ {\\text {t r i p l e t}}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.679, + 0.907, + 0.725 + ], + "angle": 0, + "content": "\\(\\mathcal{L}_{\\mathrm{recon}}\\), as defined in Eq.(2), is the reconstruction loss between the predicted noise and the actual noise with text prompt1 conditioning \\(\\mathbf{p}\\) as:" + }, + { + "type": "text", + "bbox": [ + 0.558, + 0.739, + 0.86, + 0.759 + ], + "angle": 0, + "content": "Phase one, p: \" a photo of \\( c_{i}^{\\mathrm{inspec}} \\) & \\( c_{i}^{\\mathrm{conspec}} \\)\"" + }, + { + "type": "text", + "bbox": [ + 0.553, + 0.776, + 0.868, + 0.795 + ], + "angle": 0, + "content": "Phase two, p: “ a photo of \\(\\mathbb{C}^{\\mathrm{intrinsic}}\\) & \\(c_{i}^{\\mathrm{conspec}}\\)," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.809, + 0.906, + 0.855 + ], + "angle": 0, + "content": "where \\(\\mathbb{C}^{\\mathrm{intrinsic}}\\) is a set of \\(c_{j}^{\\mathrm{intrinsic}}\\) tokens. \\(\\mathcal{L}_{\\mathrm{att}}\\) is the attention mask loss, which employs Wasserstein loss, following [9], to align the attention regions with the segmentation masks:" + }, + { + "type": "equation", + "bbox": [ + 0.643, + 0.858, + 0.905, + 0.875 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {a t t}} = \\mathcal {W} \\left(\\mathbf {A} _ {i}, \\mathbf {m} _ {i}\\right), \\tag {8}\n\\]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.531, + 0.887, + 0.893, + 0.901 + ], + "angle": 0, + "content": "1The list of prompt templates is provided in supplementary material." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23738" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.169 + ], + "angle": 0, + "content": "where \\(\\mathcal{W}\\) denotes the Wasserstein distance, \\(\\mathbf{A}_i\\) is the attention map corresponding to concept \\(c_{i}\\), and \\(\\mathbf{m}_i\\) is the segmentation mask obtained in the first stage. \\(\\mathcal{L}_{\\mathrm{triplet}}^{\\mathrm{obj}}\\) and \\(\\mathcal{L}_{\\mathrm{triplet}}^{\\mathrm{intrinsic}}\\) are the object-level and intrinsic triplet losses for different phases:" + }, + { + "type": "equation", + "bbox": [ + 0.116, + 0.172, + 0.483, + 0.205 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t r i p l e t}} ^ {\\mathrm {o b j}} = \\sum_ {i} \\mathcal {L} _ {\\text {t r i p l e t}, i} ^ {\\mathrm {o b j}}, \\quad \\mathcal {L} _ {\\text {t r i p l e t}} ^ {\\text {i n t r i n s i c}} = \\sum_ {j} \\mathcal {L} _ {\\text {t r i p l e t}, j} ^ {\\text {i n t r i n s i c}}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.208, + 0.484, + 0.269 + ], + "angle": 0, + "content": "The weighting factors \\(\\lambda_{\\mathrm{att}}\\) and \\(\\lambda_{\\mathrm{triplet}}\\) balance the contributions of the attention and triplet losses relative to the reconstruction loss in each phase, ensuring stable optimization across all objectives." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.285, + 0.223, + 0.303 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.312, + 0.485, + 0.658 + ], + "angle": 0, + "content": "Implementation details. We implement our ICE framework using Stable Diffusion v2.1 for both stages. In the first stage, for the Image-to-Text Retriever module \\(\\mathcal{T}\\), we utilize the CLIP-based SpLiCE framework [4], while the T2I zero-shot Segmentor \\(S\\) employs DiffSeg [30]. We adapt the input and output interfaces of these training-free modules to accommodate our automatic concept localization process. Specifically, modifications were made to ensure seamless integration within the framework, allowing for efficient and accurate extraction of text-based concepts and their corresponding masks. The training process in the second stage is divided into two phases: Phase one focuses on learning object-level concepts and is conducted over 400 training steps, followed by Phase two, which targets intrinsic concepts and also comprises 400 training steps. Subsequently, concept refinement is performed with another 300 steps to enhance the quality and coherence of the learned concepts. We assign the weighting factors \\(\\lambda_{\\mathrm{att}} = 1\\times 10^{-5}\\) and \\(\\lambda_{\\mathrm{triplet}} = 1\\). For the triplet loss margin \\(\\gamma\\) value, please refer to the supplementary material. Finally, we compute a prior preservation loss [26] to promote diverse image generation while retaining class-specific knowledge. All experiments are conducted on a single NVIDIA RTX 3090 GPU." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Baseline & evaluation metrics. To evaluate the effectiveness of ICE, we compare our framework with the state-of-the-art method, ConceptExpress [9], on the Unsupervised Concept Extraction (UCE) benchmarks. Specifically, we utilize \\(D1\\) dataset, which is based on the Unsplash datasets, providing a diverse and extensive collection of unlabeled images. Moreover, the UCE benchmarks assess models based on two primary metrics: (1) Concept Similarity, which includes Identity Similarity \\((\\mathrm{SIM}^I)\\) measuring how accurately individual concepts are recreated, and Compositional Similarity \\((\\mathrm{SIM}^C)\\) evaluating the overall coherence of the generated image based on the extracted concepts. (2) Classification Accuracy assesses the disentangle-ment and classification performance of the extracted concepts by evaluating how well the concepts can be used to classify images correctly. These metrics provide a compre" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.138 + ], + "angle": 0, + "content": "hensive evaluation of both the fidelity and utility of the extracted concepts, enabling a robust quantitative comparison between ICE and ConceptExpress." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.15, + 0.704, + 0.166 + ], + "angle": 0, + "content": "5.1. Quantitative results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.174, + 0.906, + 0.235 + ], + "angle": 0, + "content": "As shown in Table 2 & 3, ICE achieves higher scores in all metrics compared to previous methods, demonstrating its superior ability to recreate and compose visual concepts from unlabelled images accurately." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.246, + 0.906, + 0.274 + ], + "angle": 0, + "content": "Table 2. Performance of ICE and relevant works on Unsupervised Concept Extraction (UCE) benchmarks using CLIP [23] encoder." + }, + { + "type": "table", + "bbox": [ + 0.524, + 0.277, + 0.899, + 0.342 + ], + "angle": 0, + "content": "
Method\\(SIM^I\\)(%)\\(SIM^C\\)(%)\\(ACC^1\\)(%)\\(ACC^3\\)(%)
Break-A-Scene [1]0.6270.7730.1740.282
ConceptExpress[9]0.6890.7840.2630.385
ICE (Ours)0.7380.8220.3250.518
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.369, + 0.906, + 0.397 + ], + "angle": 0, + "content": "Table 3. Performance of ICE and relevant works on Unsupervised Concept Extraction (UCE) benchmarks using DINO [6] encoder." + }, + { + "type": "table", + "bbox": [ + 0.524, + 0.399, + 0.898, + 0.463 + ], + "angle": 0, + "content": "
Method\\( SIM^I \\) (%)\\( SIM^C \\) (%)\\( ACC^1 \\) (%)\\( ACC^3 \\) (%)
Break-A-Scene [1]0.2540.5100.2020.315
ConceptExpress [9]0.3190.5680.3240.470
ICE (Ours)0.6770.7550.4760.638
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.479, + 0.907, + 0.901 + ], + "angle": 0, + "content": "This enhanced performance can be attributed to two key factors: (1) Enhanced Concept Extraction: Our framework effectively extracts text-based concepts and their corresponding masks, resulting in superior masking accuracy as illustrated in Figure 7. Unlike other methods that may only mask object-level regions, ICE also retrieves the most similar text-based descriptors. This dual extraction provides additional guidance and robust initialization for the model, ensuring that both the spatial and semantic aspects of each concept are accurately captured. (2) Structured Concept Learning Approach: ICE's two-stage framework adopts a structured approach to concept learning, systematically decomposing object-level concepts into intrinsic attributes. This structured methodology facilitates a more granular and interpretable breakdown of visual elements, leading to the discovery of richer and more meaningful concepts. Specifically, ICE achieves CLIP Identity Similarity \\((\\mathrm{SIM}^I)\\) of \\(73.8\\%\\), compared to ConceptExpress's \\(68.9\\%\\), indicating a substantial improvement in accurately identifying object-level concepts. Additionally, ICE attains a Compositional Similarity \\((\\mathrm{SIM}^C)\\) of \\(82.2\\%\\), surpassing ConceptExpress's \\(78.4\\%\\), which underscores our framework's enhanced capability in understanding and composing complex visual attributes. These factors collectively contribute to the superior performance of ICE, highlighting the effectiveness of our framework in systematically uncovering and refining both object-level and intrinsic visual concepts from unlabeled images." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23739" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.135, + 0.103, + 0.22, + 0.117 + ], + "angle": 0, + "content": "Input image \\(x\\)" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.13, + 0.255, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.287, + 0.097, + 0.347, + 0.126 + ], + "angle": 0, + "content": "Extracted mask \\(m_{i}\\)" + }, + { + "type": "image", + "bbox": [ + 0.279, + 0.129, + 0.357, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.372, + 0.097, + 0.425, + 0.126 + ], + "angle": 0, + "content": "Retrived text \\(c_{i}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.379, + 0.152, + 0.421, + 0.164 + ], + "angle": 0, + "content": "\"Bust\"" + }, + { + "type": "image", + "bbox": [ + 0.279, + 0.192, + 0.356, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.375, + 0.215, + 0.424, + 0.228 + ], + "angle": 0, + "content": "\"Spout\"" + }, + { + "type": "image", + "bbox": [ + 0.279, + 0.259, + 0.356, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.378, + 0.283, + 0.421, + 0.295 + ], + "angle": 0, + "content": "\"Bust\"" + }, + { + "type": "image", + "bbox": [ + 0.279, + 0.323, + 0.356, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.376, + 0.347, + 0.423, + 0.359 + ], + "angle": 0, + "content": "\"Cone\"" + }, + { + "type": "image", + "bbox": [ + 0.279, + 0.393, + 0.356, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.362, + 0.416, + 0.437, + 0.428 + ], + "angle": 0, + "content": "\"Ornament\"" + }, + { + "type": "image", + "bbox": [ + 0.279, + 0.455, + 0.356, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.382, + 0.479, + 0.419, + 0.491 + ], + "angle": 0, + "content": "\"Rat\"" + }, + { + "type": "image", + "bbox": [ + 0.279, + 0.524, + 0.356, + 0.584 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.369, + 0.549, + 0.431, + 0.561 + ], + "angle": 0, + "content": "\"Wooden\"" + }, + { + "type": "image", + "bbox": [ + 0.279, + 0.587, + 0.356, + 0.646 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.369, + 0.611, + 0.431, + 0.622 + ], + "angle": 0, + "content": "\"Wooden\"" + }, + { + "type": "image", + "bbox": [ + 0.279, + 0.656, + 0.356, + 0.715 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.376, + 0.68, + 0.423, + 0.692 + ], + "angle": 0, + "content": "\"Pines\"" + }, + { + "type": "image", + "bbox": [ + 0.279, + 0.718, + 0.356, + 0.776 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.368, + 0.741, + 0.431, + 0.754 + ], + "angle": 0, + "content": "\"Chapter\"" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.657, + 0.255, + 0.776 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.283, + 0.097, + 0.347, + 0.126 + ], + "angle": 0, + "content": "Extracted mask \\(m_{i}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.279, + 0.799, + 0.356, + 0.811 + ], + "angle": 0, + "content": "ed concepts a" + }, + { + "type": "image_caption", + "bbox": [ + 0.279, + 0.813, + 0.483, + 0.826 + ], + "angle": 0, + "content": "ages of learned object-level and int" + }, + { + "type": "image_caption", + "bbox": [ + 0.28, + 0.741, + 0.356, + 0.753 + ], + "angle": 0, + "content": "\\( {13}/{14} \\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.372, + 0.097, + 0.425, + 0.126 + ], + "angle": 0, + "content": "Retrived text \\(c_{i}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.489, + 0.096, + 0.565, + 0.124 + ], + "angle": 0, + "content": "Object-level concepts" + }, + { + "type": "image", + "bbox": [ + 0.489, + 0.129, + 0.565, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.488, + 0.192, + 0.563, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.488, + 0.259, + 0.563, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.488, + 0.324, + 0.563, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.391, + 0.563, + 0.451 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.457, + 0.563, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.525, + 0.563, + 0.584 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.587, + 0.563, + 0.646 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.488, + 0.656, + 0.563, + 0.715 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.488, + 0.718, + 0.563, + 0.776 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.696, + 0.096, + 0.803, + 0.11 + ], + "angle": 0, + "content": "Intrinsic concepts" + }, + { + "type": "image_caption", + "bbox": [ + 0.63, + 0.113, + 0.671, + 0.126 + ], + "angle": 0, + "content": "Object" + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.129, + 0.688, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.191, + 0.688, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.259, + 0.688, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.324, + 0.688, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.391, + 0.688, + 0.451 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.457, + 0.688, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.525, + 0.688, + 0.583 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.587, + 0.688, + 0.646 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.656, + 0.688, + 0.715 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.718, + 0.688, + 0.776 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.724, + 0.113, + 0.776, + 0.126 + ], + "angle": 0, + "content": "Material" + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.129, + 0.787, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.191, + 0.787, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.259, + 0.787, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.323, + 0.787, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.711, + 0.391, + 0.787, + 0.451 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.711, + 0.457, + 0.787, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.711, + 0.525, + 0.787, + 0.584 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.711, + 0.587, + 0.787, + 0.646 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.711, + 0.656, + 0.787, + 0.715 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.711, + 0.718, + 0.787, + 0.776 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.828, + 0.114, + 0.871, + 0.126 + ], + "angle": 0, + "content": "Colour" + }, + { + "type": "image", + "bbox": [ + 0.811, + 0.129, + 0.887, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.811, + 0.191, + 0.887, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.811, + 0.259, + 0.887, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.811, + 0.324, + 0.887, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.811, + 0.391, + 0.887, + 0.451 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.811, + 0.457, + 0.887, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.812, + 0.526, + 0.886, + 0.584 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.811, + 0.587, + 0.887, + 0.646 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.812, + 0.656, + 0.887, + 0.715 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.811, + 0.718, + 0.887, + 0.776 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.785, + 0.905, + 0.828 + ], + "angle": 0, + "content": "Figure 6. Qualitative results of the ICE framework demonstrating its systematic concept discovery process. Column 1: Input images. Column 2: Extracted text-based concepts and their corresponding masks obtained from Stage One: Automatic Concept Localization. Columns 3 & 4: Generated images of learned object-level and intrinsic concepts derived from Stage Two: Structured Concept Learning." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.838, + 0.268, + 0.853 + ], + "angle": 0, + "content": "5.2. Qualitative results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To demonstrate the effectiveness of the proposed ICE framework, we present qualitative results, as illustrated in" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.838, + 0.905, + 0.869 + ], + "angle": 0, + "content": "Figure 6, showcasing ICE's ability to accurately extract and decompose visual concepts systematically." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "First, given an input image \\( x \\), ICE extracts corresponding concept masks and retrieves relevant text-based descrip" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23740" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.09, + 0.182, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.182, + 0.094, + 0.301, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.305, + 0.091, + 0.392, + 0.157 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.393, + 0.094, + 0.51, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.091, + 0.602, + 0.157 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.603, + 0.094, + 0.688, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.69, + 0.091, + 0.782, + 0.157 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.782, + 0.094, + 0.801, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.803, + 0.094, + 0.901, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.161, + 0.908, + 0.202 + ], + "angle": 0, + "content": "Figure 7. Qualitative comparison of segmentation results between ConceptExpress and our framework. Our framework not only delivers more accurate segmentation of object-level regions but also effectively retrieves the most relevant text-based concepts, demonstrating its unique capabilities over ConceptExpress." + }, + { + "type": "table_caption", + "bbox": [ + 0.175, + 0.204, + 0.591, + 0.218 + ], + "angle": 0, + "content": "Table 4. Ablation study on ICE model components on the \\(D1\\) dataset." + }, + { + "type": "table", + "bbox": [ + 0.1, + 0.219, + 0.668, + 0.319 + ], + "angle": 0, + "content": "
Model's componentsCLIP [23] encoderDINO [6] encoder
MethodStage One's maskStage Two's token init.Stage Two learningSIM1(%)SIMC(%)ACC1(%)ACC3(%)SIM1(%)SIMC(%)ACC1(%)ACC3(%)
ConceptExpressXXX0.6890.7840.2630.3850.3190.5680.3240.470
ICE w. mask✓XX0.7100.7810.3070.4560.4930.6010.3950.604
ICE w.o Stage Two✓✓X0.7260.8070.3010.4520.5010.6210.4110.604
ICE w.o text init.✓X✓0.7220.8140.3200.4750.5480.6430.4490.627
ICE (Ours)✓✓✓0.7380.8220.3250.5180.6770.7550.4760.638
" + }, + { + "type": "table_caption", + "bbox": [ + 0.676, + 0.227, + 0.901, + 0.253 + ], + "angle": 0, + "content": "Table 5. Quantitative comparison of generated masks on the \\(D1\\) dataset." + }, + { + "type": "table", + "bbox": [ + 0.676, + 0.256, + 0.901, + 0.307 + ], + "angle": 0, + "content": "
MethodmIoURecallPrecision
ConceptExpress0.4830.6760.657
ICE (Ours)0.6350.8930.720
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.33, + 0.484, + 0.541 + ], + "angle": 0, + "content": "tors, as shown in the second column of the figure. These masks precisely delineate regions within the image, ensuring accurate localization of each concept. Next, leveraging the extracted masks and retrieved texts, ICE learns distinct object-level concepts, displayed in the third column. For example, the framework identifies objects such as “Bust” and “Cone” and associates them with their respective semantic descriptors. Finally, utilizing the information from the object-level concepts, ICE systematically discovers intrinsic concepts, including attributes like material and colour, through its second-stage processing. This decomposition enables the framework to uncover compositional intrinsic concepts such as the object type, material, and colour, as demonstrated in the fourth column." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.543, + 0.484, + 0.664 + ], + "angle": 0, + "content": "This structured approach highlights the superiority of our framework by effectively leveraging a single T2I model to systematically uncover both object-level and intrinsic visual concepts. The ability to deconstruct an image into its fundamental components not only validates the robustness of ICE but also illustrates the potential of T2I models to facilitate comprehensive and structured concept learning in an unsupervised manner." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.678, + 0.331, + 0.695 + ], + "angle": 0, + "content": "5.3. Model component analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.703, + 0.483, + 0.839 + ], + "angle": 0, + "content": "ICE ablation study. To validate the effectiveness of our ICE framework, we conduct an ablation study on its individual components: Stage One's retrieved mask, Stage Two's token embedding initialization using text retrieved from Stage One, and Stage Two's Structured Concept Learning. The results, as shown in Table 4, demonstrate that each component significantly enhances performance. Furthermore, the complete ICE framework outperforms all its variants across every metric in the UCE benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "ICE Stage One's mask quality. Furthermore, Table 5 compares our Stage One generated masks with ConceptExpress using \\(D1\\)'s manually annotated ground-truth masks. We evaluate the segmentation performance of generated masks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.33, + 0.907, + 0.572 + ], + "angle": 0, + "content": "against ground-truth masks using the Hungarian algorithm to find the optimal matching between the sets of masks. We evaluate each pair of predicted and ground-truth masks using segmentation criterions, forming a similarity matrix. The Hungarian algorithm is then applied to this matrix to determine the optimal one-to-one correspondence that maximizes the overall similarity. Overall, our Stage One technique achieves better segmentation results across different criterion. Notably, our Stage One masks exhibit significantly higher recall rates, indicating that a greater proportion of the relevant object-level regions in the annotated area are successfully captured. This higher recall demonstrates that ICE's Automatic Concept Localization method in Stage One is more effective at identifying and extracting meaningful concepts, reducing the likelihood of missing important object-level concepts compared to previous methods." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.606, + 0.633, + 0.621 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.638, + 0.907, + 0.849 + ], + "angle": 0, + "content": "In this paper, we introduce ICE, a novel framework for automatically and systematically extracting intrinsic concepts from a single image via diffusion models. Our framework consists of two stages. In Stage One, we design a concept localization module based on the T2I model that effectively retrieves the most relevant text-based concepts and their corresponding masks. Subsequently, in Stage Two, we introduce a structured concept learning method to systematically decompose object-level concepts into their intrinsic attributes, resulting in richer and more interpretable concept representations. In summary, our approach effectively harnesses Text-to-Image (T2I) models to enhance concept learning, laying the groundwork for further advancements in intrinsic visual concept extraction." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Acknowledgments. This work is supported by the Hong Kong Research Grants Council - General Research Fund (Grant No.: 17211024)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23741" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.17 + ], + "angle": 0, + "content": "[1] Omri Avrahami, Kfir Aberman, Ohad Fried, Daniel Cohen-Or, and Dani Lischinski. Break-a-scene: Extracting multiple concepts from a single image. In SIGGRAPH Asia, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.173, + 0.484, + 0.227 + ], + "angle": 0, + "content": "[2] Omri Avrahami, Thomas Hayes, Oran Gafni, Sonal Gupta, Yaniv Taigman, Devi Parikh, Dani Lischinski, Ohad Fried, and Xi Yin. Spatext: Spatio-textual representation for controllable image generation. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.23, + 0.483, + 0.271 + ], + "angle": 0, + "content": "[3] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. In ICML, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.273, + 0.483, + 0.327 + ], + "angle": 0, + "content": "[4] Usha Bhalla, Alex Oesterling, Suraj Srinivas, Flavio P Calmon, and Himabindu Lakkaraju. Interpreting clip with sparse linear concept embeddings (splice). In NeurIPS, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.33, + 0.483, + 0.37 + ], + "angle": 0, + "content": "[5] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.374, + 0.483, + 0.427 + ], + "angle": 0, + "content": "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.431, + 0.483, + 0.486 + ], + "angle": 0, + "content": "[7] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In ICLR, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.488, + 0.483, + 0.542 + ], + "angle": 0, + "content": "[8] Shaozhe Hao, Kai Han, Shihao Zhao, and Kwan-Yee K Wong. Vico: Detail-preserving visual condition for personalized text-to-image generation. arXiv preprint arXiv:2306.00971, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.545, + 0.483, + 0.599 + ], + "angle": 0, + "content": "[9] Shaozhe Hao, Kai Han, Zhengyao Lv, Shihao Zhao, and Kwan-Yee K Wong. Conceptexpress: Harnessing diffusion models for single-image unsupervised concept extraction. In ECCV, 2024. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.602, + 0.483, + 0.656 + ], + "angle": 0, + "content": "[10] Eric Hedlin, Gopal Sharma, Shweta Mahajan, Hossam Isack, Abhishek Kar, Andrea Tagliasacchi, and Kwang Moo Yi. Unsupervised semantic correspondence using stable diffusion. In NeurIPS, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.659, + 0.483, + 0.699 + ], + "angle": 0, + "content": "[11] Jonathan Ho, Tim Salimans, Alexey A Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.702, + 0.483, + 0.757 + ], + "angle": 0, + "content": "[12] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Weli-hinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.759, + 0.483, + 0.814 + ], + "angle": 0, + "content": "[13] Chen Jin, Ryutaro Tanno, Amrutha Saseendran, Tom Diethe, and Philip Teare. An image is worth multiple words: Discovering object level concepts using multi-concept prompt learning. In ICML, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.483, + 0.857 + ], + "angle": 0, + "content": "[14] Laurynas Karazija, Iro Laina, Andrea Vedaldi, and Christian Rupprecht. Diffusion models for zero-shot open-vocabulary segmentation. In ECCV, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.899 + ], + "angle": 0, + "content": "[15] Sharon Lee, Yunzhi Zhang, Shangzhe Wu, and Jiajun Wu. Language-informed visual concept learning. In ICLR, 2023. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "[16] Alexander C. Li, Mihir Prabhudesai, Shivam Duggal, Ellis Brown, and Deepak Pathak. Your diffusion model is secretly a zero-shot classifier. In ICCV, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[17] Dongxu Li, Junnan Li, and Steven CH Hoi. Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. In NeurIPS, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.178, + 0.905, + 0.217 + ], + "angle": 0, + "content": "[18] Xinghui Li, Jingyi Lu, Kai Han, and Victor Prisacariu. Sd4match: Learning to prompt stable diffusion model for semantic matching. In CVPR, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.219, + 0.905, + 0.272 + ], + "angle": 0, + "content": "[19] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.275, + 0.905, + 0.327 + ], + "angle": 0, + "content": "[20] Minheng Ni, Yabo Zhang, Kailai Feng, Xiaoming Li, Yiwen Guo, and Wangmeng Zuo. Ref-diff: Zero-shot referring image segmentation with generative models. arXiv preprint arXiv:2308.16777, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.331, + 0.905, + 0.397 + ], + "angle": 0, + "content": "[21] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob Mcgrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In ICML, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.401, + 0.905, + 0.453 + ], + "angle": 0, + "content": "[22] Zeju Qiu, Weiyang Liu, Haiwen Feng, Yuxuan Xue, Yao Feng, Zhen Liu, Dan Zhang, Adrian Weller, and Bernhard Scholkopf. Controlling text-to-image diffusion by orthogonal finetuning. In NeurIPS, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.456, + 0.905, + 0.523 + ], + "angle": 0, + "content": "[23] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.905, + 0.579 + ], + "angle": 0, + "content": "[24] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents. arXiv preprint arXiv:2204.06125, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.582, + 0.905, + 0.621 + ], + "angle": 0, + "content": "[25] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.624, + 0.905, + 0.677 + ], + "angle": 0, + "content": "[26] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.679, + 0.905, + 0.747 + ], + "angle": 0, + "content": "[27] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.749, + 0.905, + 0.816 + ], + "angle": 0, + "content": "[28] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.818, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[29] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. In ICLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[30] Junjiao Tian, Lavisha Aggarwal, Andrea Colaco, Zsolt Kira, and Mar Gonzalez-Franco. Diffuse, attend, and segment:" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23742" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "Unsupervised zero-shot segmentation using stable diffusion. In CVPR, 2024. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.163 + ], + "angle": 0, + "content": "[31] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.207 + ], + "angle": 0, + "content": "[32] Yael Vinker, Andrey Voynov, Daniel Cohen-Or, and Ariel Shamir. Concept decomposition for visual exploration and inspiration. In SIGGRAPH Asia, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.208, + 0.482, + 0.263 + ], + "angle": 0, + "content": "[33] Jinglong Wang, Xiawei Li, Jing Zhang, Qingyuan Xu, Qin Zhou, Qian Yu, Lu Sheng, and Dong Xu. Diffusion model is secretly a training-free open vocabulary semantic segmenter. arXiv preprint arXiv:2309.02773, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[34] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Weixian Lei, Yuchao Gu, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In ICCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.39 + ], + "angle": 0, + "content": "[35] Changming Xiao, Qi Yang, Feng Zhou, and Changshui Zhang. From text to mask: Localizing entities using the attention of text-to-image diffusion models. In Neurocomputing, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.482, + 0.432 + ], + "angle": 0, + "content": "[36] Zhi Xu, Shaozhe Hao, and Kai Han. Cusconcept: Customized visual concept decomposition with diffusion models. In WACV, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.482, + 0.502 + ], + "angle": 0, + "content": "[37] Junyi Zhang, Charles Herrmann, Junhwa Hur, Luisa Polania Cabrera, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. A tale of two features: Stable diffusion complements dino for zero-shot semantic correspondence. In NeurIPS, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.504, + 0.482, + 0.533 + ], + "angle": 0, + "content": "[38] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.534, + 0.482, + 0.574 + ], + "angle": 0, + "content": "[39] Yanbing Zhang, Mengping Yang, Qin Zhou, and Zhe Wang. Attention calibration for disentangled text-to-image personalization. In CVPR, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.574 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23743" + } + ] +] \ No newline at end of file diff --git a/2025/ICE_ Intrinsic Concept Extraction from a Single Image via Diffusion Models/e9edff49-034e-4138-82ed-91d11ac96969_origin.pdf b/2025/ICE_ Intrinsic Concept Extraction from a Single Image via Diffusion Models/e9edff49-034e-4138-82ed-91d11ac96969_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..dfb1450ad3b6fa8c92c41cdfb62c083148dd9e7d --- /dev/null +++ b/2025/ICE_ Intrinsic Concept Extraction from a Single Image via Diffusion Models/e9edff49-034e-4138-82ed-91d11ac96969_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d3cf49bb204904616ace59495ce21aecbea31fcf4ad15db24fbacebe2ed0d65e +size 8853053 diff --git a/2025/ICE_ Intrinsic Concept Extraction from a Single Image via Diffusion Models/full.md b/2025/ICE_ Intrinsic Concept Extraction from a Single Image via Diffusion Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f9ee11683333299d165def3b82dc2a6c77965a4e --- /dev/null +++ b/2025/ICE_ Intrinsic Concept Extraction from a Single Image via Diffusion Models/full.md @@ -0,0 +1,436 @@ +# ICE: Intrinsic Concept Extraction from a Single Image via Diffusion Models + +Fernando Julio Cendra Kai Han + +Visual AI Lab, The University of Hong Kong + +fcendra@connect.hku.hk, kaihanx@hku.hk + +![](images/0101f83f761da9715d805a7df8cef39b4842ec23a4cf74794c7c39908a32de13.jpg) +Figure 1. We showcase a structured approach for defining visual concepts within an image, where object-level concepts are identified and analyzed to reveal their underlying intrinsic attributes, such as color, material, and shape. We present the ICE (Intrinsic Concept Extraction) framework, which leverages Text-to-Image (T2I) models to systematically discover these concepts, providing a more effective method for learning visual concepts. + +# Abstract + +The inherent ambiguity in defining visual concepts poses significant challenges for modern generative models, such as the diffusion-based Text-to-Image (T2I) models, in accurately learning concepts from a single image. Existing methods lack a systematic way to reliably extract the interpretable underlying intrinsic concepts. To address this challenge, we present ICE, short for Intrinsic Concept Extraction, a novel framework that exclusively utilizes a T2I model to automatically and systematically extract intrinsic concepts from a single image. ICE consists of two pivotal stages. In the first stage, ICE devises an automatic concept localization module to pinpoint relevant text-based concepts and their corresponding masks within the image. This critical stage streamlines concept initialization and provides precise guidance for subsequent analysis. The second stage delves deeper into each identified mask, decomposing the object-level concepts into intrinsic concepts and general concepts. This decomposition allows for a more granular and interpretable breakdown of visual elements. Our framework demonstrates superior performance on in + +trinsic concept extraction from a single image in an unsupervised manner. Project page: https://visual-ai.github.io/ice + +# 1. Introduction + +Text-to-Image (T2I) models have undergone significant advancements, enabling the generation of high-quality images from textual prompts and bridging natural language processing with visual content creation. These capabilities have driven applications in creative design, entertainment, and assistive technologies, underscoring the pivotal role of T2I models in generative AI. + +Beyond image generation, emerging research has uncovered that T2I models excel in tasks such as image classification, segmentation, and semantic correspondence. Studies [10, 14, 16, 18, 20, 30, 33, 35, 37] have shown their effectiveness in these areas. This versatility suggests that diffusion-based T2I models encapsulate substantial world knowledge, allowing them to comprehend and recreate visual concepts with precision and detail. Building on this potential, our work explores the use of T2I models like Stable Diffusion for concept learning. Instead of solely generating + +images from text, we investigate how these models invert the generative process to systematically define and learn the fundamental concepts that comprise a given unlabelled image, thereby uncovering the essential concepts that enable the synthesis of complex visual scenes. + +Despite their promising potential, concept learning within T2I models faces several inherent challenges. Defining visual concepts remains inherently ambiguous, complicating the reliability with which these models can discover and utilize essential concepts. Previous methods [7, 8, 26] such as Textual Inversion (TI) and DreamBooth require learning multiple similar images to encapsulate a concept. Other frameworks, such as Break-A-Scene [1], MCPL [13], and ConceptExpress [9], focus on extracting object-level concepts within a single image. Moreover, these methods often struggle to provide clear and interpretable concept representations, hindering their effectiveness in capturing the nuanced and multifaceted nature of visual scenes. The ambiguity in defining visual concepts not only impedes reliable concept discovery but also poses significant challenges for human annotators attempting to delineate these concepts manually. + +Addressing these challenges requires a fundamental re-evaluation of how we define and learn visual concepts within an image. Previous works, such as Inspiration Tree [32], have attempted to construct hierarchical representations by organizing tokens into tree structures to capture different concepts within an image. Similarly, LangInt [15] uses Visual Question Answering (VQA) techniques to guide concept learning, but it requires training separate encoders for each concept axis, thereby failing to fully utilize their inherent hierarchical structures. Consequently, there is a need for more systematic and structured methods to define and learn concepts from images, enabling a more granular understanding of the fundamental building blocks of visual representation. + +In response to these challenges, we present ICE, a novel two-stage framework designed to automatically and systematically extract intrinsic concepts using a single T2I model. ICE first devises an automatic concept localization module to retrieves text-based concepts and their corresponding masks. This enhances initialization and provides clear guidance for subsequent processing. The framework then systematically decomposes these object-level concepts into intrinsic and general concepts, enabling a more granular and interpretable breakdown of visual elements. Through comprehensive experiments, ICE demonstrates its effectiveness in leveraging T2I models to systematically uncover diverse visual concepts, facilitating a structured and detailed understanding of image components. + +# 2. Related Work + +Text-to-Image generation. Diffusion-based methods have significantly advanced the field of text-to-image (T2I) generation, pushing the boundaries of image synthesis to unprecedented levels. Early approaches such as DALL-E 2 [24], Imagen [27], and GLIDE [21] demonstrated the potential of diffusion models in generating high-fidelity images from textual descriptions. Latent Diffusion Models (LDMs) [25] further enhanced this capability by operating in a latent space, enabling more efficient and scalable training. Building upon the foundation laid by LDMs, Stable Diffusion (SD) [25] is trained on the LAION-5B dataset [28] and has set new benchmarks in T2I synthesis. + +Beyond basic image generation, diffusion models have been adapted for a variety of specialized tasks. Controllable generation [21, 22, 38] allows for nuanced adjustments based on user input, while applications such as global and local editing [5, 31] enable fine-grained modifications to specific regions of an image. Additionally, these models have been extended to video generation and editing [11, 29, 34], inpainting [19], and scene generation [2, 3], showcasing their versatility across use cases. + +Generative concept learning. Generative concept learning aims to decompose and understand the fundamental elements that constitute complex visual scenes. Textual Inversion (TI) [7] was among the first methods to learn single concepts from multiple images by introducing new words and optimizing their corresponding token embeddings. Similarly, DreamBooth [26] extends this approach by leveraging less frequently used words and optimizing both the denoising U-Net and introducing a preservation loss to enhance the quality of the learned concepts. + +Break-A-Scene [1] relies heavily on human-annotated masks to isolate and learn distinct concepts from a single image, limiting its applicability in settings where such annotations are unavailable Concurrent works like MCPL [13] and DisenDiff [39] address similar challenges but require concept-descriptive text captions or predefined class names, making them less feasible for more general tasks. Recent advancements, such as ConceptExpress [9], have reduced reliance on human annotations by automatically identifying and learning multiple concepts within an image. However, these methods primarily treat concepts at the object-level, neglecting intrinsic attributes such as colour and material properties. Inspiration Tree [32] introduces a preliminary approach to concept decomposition by splitting a token into two tokens to force the learning of different concepts. Despite this, it lacks reliable interpretation and structured guidance, resulting in less coherent concept representations. LangInt [15] and CusConcept [36] attempt to learn intrinsic concepts using VQA models [12, 17] to query specific attributes of an image and independently learn these intrinsic concepts. However, these approaches are limited to scenar + +![](images/4bd400f44d5ac11e467b49aaa80561f356871c34290918e4e750a0f0a26215c3.jpg) +Figure 2. Concept definition hierarchy illustrating how object-level concepts are decomposed into intrinsic attributes, including object type, colour, material and other intrinsics. + +ios with a single object-level concept per image, failing to handle images with multiple concepts. + +Comparison and positioning of ICE. In contrast to existing methodologies, our proposed framework, ICE, offers a unified and structured approach to automatically and systematically discover intrinsic concepts within an image using a single T2I model. Unlike previous methods, ICE not only identifies object-level concepts but also decomposes them into intrinsic attributes such as colour and material (see Figure 2), providing a more comprehensive and interpretable representation of visual concepts. This two-stage decomposition ensures that both general and specific attributes are captured, enhancing the model's ability to understand and generate nuanced visual content. For an overview of the differences between our framework and existing methods, please refer to Table 1. + +Table 1. Comparison of ICE and relevant works. + +
Learned concept(s)Framework details
MethodObject-level conceptsIntrinsic conceptsMulti conceptsSingle imageExtra information
Textual Inversion [7]✓XXX-
Dreambooth [26]✓XXX-
Inspiration Tree [32]X✓XX-
LangInt [15]✓✓X✓VQA-guided
Break-A-Scene [1]✓X✓✓Mask
MCPL [13]✓X✓✓Text
ConceptExpress [9]✓X✓✓-
ICE (Ours)✓✓✓✓-
+ +# 3. Preliminaries + +Diffusion models have emerged as a powerful framework for generative tasks, particularly in Text-to-Image (T2I) applications. They operate by iteratively denoising data through a defined forward and reverse diffusion process, enabling the generation of high-fidelity images from noise. + +Forward process. In the forward diffusion process, Gaussian noise is incrementally added to an image over a series of timesteps $t = 1,2,\ldots ,T$ , progressively transforming it + +into pure noise: + +$$ +q \left(\mathbf {x} _ {t} \mid \mathbf {x} _ {t - 1}\right) = \mathcal {N} \left(\mathbf {x} _ {t}; \sqrt {1 - \beta_ {t}} \mathbf {x} _ {t - 1}, \beta_ {t} \mathbf {I}\right), \tag {1} +$$ + +where $\beta_{t}$ controls the noise variance at each timestep. We define $\alpha_{t} = 1 - \beta_{t}$ , and the cumulative product $\bar{\alpha}_{t}$ is given by $\bar{\alpha}_{t} = \prod_{s=1}^{t} \alpha_{s}$ . Here, $\alpha_{t}$ represents the remaining signal strength at timestep $t$ , while $\bar{\alpha}_{t}$ indicates the cumulative effect, controlling the overall noise level in the image. + +Reverse process. The reverse diffusion process aims to reconstruct the original image by predicting and removing the added noise. This is achieved by training a denoising network $\epsilon_{\theta}$ to minimize the denoising loss: + +$$ +\mathcal {L} _ {\text {r e c o n}} = \mathbb {E} _ {\mathbf {x} _ {0}, \epsilon , t} \left[ \| \epsilon - \epsilon_ {\theta} (\mathbf {x} _ {t}, t, \mathcal {E} (\mathbf {p})) \| _ {2} ^ {2} \right], \tag {2} +$$ + +where $\mathbf{x}_t = \sqrt{\bar{\alpha}_t}\mathbf{x}_0 + \sqrt{1 - \bar{\alpha}_t\epsilon}$ and $\epsilon \sim \mathcal{N}(\mathbf{0},\mathbf{I})$ . The term $\mathbf{p}$ represents the text prompt used for conditioning via text encoder $\mathcal{E}$ like CLIP [23]. + +The denoising loss function $\mathcal{L}_{\mathrm{recon}}$ is crucial for training the diffusion model, guiding $\epsilon_{\theta}$ to accurately predict the added noise $\epsilon$ at each timestep. By minimizing this loss, the model learns to generate clean images from noisy inputs, effectively capturing the underlying data distribution. + +Relation to ICE. ICE leverages the inherent capabilities of diffusion-based T2I models to perform concept extraction and decomposition, enabling a more structured and interpretable approach to visual concept learning. + +# 4. Method + +The proposed ICE (Intrinsic Concept Extraction) framework is designed to automatically and systematically discover essential concepts within images using a single T2I diffusion model. The framework operates through a two-stage architecture: Stage One: Automatic Concept Localization and Stage Two: Structured Concept Learning as illustrated in Figure 3. In this section, we provide an overview of each stage, detailing the underlying processes, and training objectives. + +# 4.1. Stage One: Automatic Concept Localization + +This stage is designed to extract object-level concepts from an unlabelled input image automatically. This stage leverages off-the-shelf modules integrated within the T2I diffusion model, ensuring a training-free and seamless concept extraction process. The workflow of Automatic Concept Localization is illustrated in Figure 4. + +Text-based concept retrieval. Given an unlabelled input image $\mathbf{x}$ , the first step involves retrieving relevant text-based concepts from the T2I model's vocabulary. This retrieval is facilitated by the Image-to-Text Retriever, based on T2I's CLIP encoders, denoted as $\mathcal{T}(\mathbf{x})$ to decompose dense embeddings into sparse, interpretable semantic concepts. Formally, the retrieval process can be expressed as: + +$$ +\left\{c _ {i} \right\} = \mathcal {T} (\mathbf {x}), \tag {3} +$$ + +![](images/acbe1f654db1084933b0f42a343c9e759456c5e05637fdfc0e00778580e8489c.jpg) +Figure 3. Illustration of the proposed ICE (Intrinsic Concept Extraction) framework, which consists of two stages: (1) Automatic Concept Localization, where a diffusion model is employed to extract object-level concepts and their corresponding masks from an image without prior training, and (2) Structured Concept Learning, where these extracted information are further leveraged to uncover essential concepts. + +![](images/3329133e769ffc614e867a77aa1bc7bb92b4e6ec7bc3eaec8dd4b8af0ae5c83a.jpg) +Stage One: Automatic Concept Localization + +![](images/d3d65dfbd493db0d530da5faf97029653e6144fea818a3acdba50ededb4fa9d3.jpg) +Stage One's use case diagram +Figure 4. Stage One: Automatic Concept Localization. Starting with an unlabelled image $\mathbf{x}$ , the Image-to-Text concept extractor retrieves the top-1 text-based concept $c_{i}$ using CLIP encoders. A zero-shot segmentor via T2I model generates the corresponding mask $\mathbf{m}_{i}$ , and the image is updated by removing the masked region. This process iterates until no objects remain in the image. + +where $\{c_i\}$ represents the set of top- $k$ text-based concepts extracted from the image $\mathbf{x}$ . + +Mask generation via T2I zero-shot segmentor. Once the top-1 text concept $c_{i}$ is identified, the framework employs the Segmentor, denoted as $S(\mathbf{x},c_i)$ , to obtain a corresponding segmentation mask $\mathbf{m}_i$ . The mask generation process is formalized as: + +$$ +\mathbf {m} _ {i} = \mathcal {S} (\mathbf {x}, c _ {i}), \tag {4} +$$ + +where $\mathbf{m}_i$ delineates the region of $\mathbf{x}$ associated with the concept $c_{i}$ . In our implementation, without introducing any external pretrained model, we utilize a training-free image segmentor that leverages the self-attention layers within the pre-trained Stable Diffusion model to produce high-quality segmentation masks without any additional training. + +Iterative object extraction. With the mask $\mathbf{m}_i$ obtained, the corresponding region is removed from the image to prepare for the extraction of subsequent concepts. The updated image $\mathbf{x}'$ is computed as: $\mathbf{x}' = \mathbf{x} \odot (1 - \mathbf{m}_i)$ , where $\odot$ denotes element-wise multiplication. This updated image $\mathbf{x}'$ is then fed back into $\mathcal{T}$ and $S$ to iteratively extract and mask remaining objects until no objects remain in the image. + +The integration of $\mathcal{T}$ and $S$ within the T2I diffusion model allows for a fully automated and training-free concept extraction process. This design ensures that the frame + +work does not rely on any external models beyond the pretrained T2I diffusion model, enhancing its scalability and applicability across diverse datasets. The Automatic Concept Extraction module outputs a set of text-based concepts $\{c_i\}$ along with their masks $\{\mathbf{m}_i\}$ for a given image $\mathbf{x}$ . The pseudocode for this stage is provided in the supplementary material. + +# 4.2. Stage Two: Structured Concept Learning + +This stage focuses on decomposing the extracted object-level concepts into the underlying intrinsic concepts, such as colour and material. To ensure an accurate decomposition of the concepts, we divide this stage into two phases: (1) learning object-level concepts and (2) learning intrinsic concepts. Stage two's workflow is illustrated in Figure 5. + +Phase one: learning object-level concepts. Here, the framework learns both concept-specific and instance-specific tokens for each extracted object-level concept. For each object-level concept $c_{i}$ , two tokens are introduced: + +- Concept-specific token ( $c_i^{\text{conspec}}$ ): Represents the general semantic category of the object. +- Instance-specific token $(c_{i}^{\mathrm{inspec}})$ : Captures attributes unique to the specific instance of the object. + +To differentiate these tokens, we initialize $c_{i}^{\mathrm{conspec}}$ em + +![](images/fdae20ccaa81735089bb133ae4ddc554178ab9c618626302b86af0523d509be7.jpg) +Figure 5. Stage Two: Structured Concept Learning. This stage is divided into two phases: (1) learning object-level concepts, where concept-specific $(c_i^{\mathrm{conspec}})$ and instance-specific $(c_i^{\mathrm{inspec}})$ tokens are learned using an object-level triplet loss $\mathcal{L}_{\mathrm{triplet}}^{\mathrm{obj}}$ , and (2) learning intrinsic concepts, which decomposes object-level concepts into intrinsic attributes $(c_j^{\mathrm{intrinsic}})$ using an intrinsic triplet loss $\mathcal{L}_{\mathrm{triplet}}^{\mathrm{intrinsic}}$ . This hierarchical approach ensures accurate separation of general semantic categories from specific and intrinsic attributes. + +bedding with the extracted text-based concept $c_{i}$ which is obtained in the first stage and employ a triplet loss mechanism, where the $c_{i}$ serves as the anchor, $c_{i}^{\mathrm{conspec}}$ as the positive, and $c_{i}^{\mathrm{inspec}}$ as the negative. The triplet loss $\mathcal{L}_{\mathrm{triplet}}^{\mathrm{obj}}$ for object-level concept $c_{i}$ is defined as: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {t r i p l e t}} ^ {\mathrm {o b j}} = \max (0, \| \mathcal {E} (\text {a n c h o r}) - \mathcal {E} \left(c _ {i} ^ {\text {c o n s p e c}}\right) \| _ {2} ^ {2} \\ - \left\| \mathcal {E} (\text {a n c h o r}) - \mathcal {E} \left(c _ {i} ^ {\text {i n s p e c}}\right) \right\| _ {2} ^ {2} + \gamma), \tag {5} \\ \end{array} +$$ + +where $\gamma$ is the margin parameter, and $\mathcal{E}(\cdot)$ denotes the T2I's CLIP text encoder. This loss encourages $c_{i}^{\mathrm{conspec}}$ to be closer to anchor $c_{i}$ than $c_{i}^{\mathrm{inspec}}$ , effectively distinguishing between general and instance-specific attributes. + +Phase two: learning intrinsic concepts. The second phase delves into decomposing the learned object-level concepts into intrinsic attributes. For each intrinsic concept (e.g., colour, material), we introduce corresponding intrinsic tokens $c_{j}^{\mathrm{intrinsic}}$ associated with object-level concepts $c_{i}$ . The intrinsic concept tokens are anchored using combined representations of $c_{i}^{\mathrm{conspec}}$ and $c_{i}^{\mathrm{inspec}}$ . Specifically, for an intrinsic concept $j$ , the anchor text $\mathsf{anchor}_{j}$ is formulated as: + +anchorj:"a intrinsicconcept + +where intrinsic denotes the specific attribute (e.g., color). The triplet loss $\mathcal{L}_{\mathrm{triplet}}^{\mathrm{intrinsic}}$ for intrinsic concept $j$ is then defined as: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {t r i p l e t}} ^ {\text {i n t r i n s i c}} = \max (0, \| \mathcal {E} (\text {a n c h o r} _ {j}) - \mathcal {E} (c _ {j} ^ {\text {i n t r i n s i c}}) \| _ {2} ^ {2} \\ \left. - \left\| \mathcal {E} (\operatorname {a n c h o r} _ {j}) - \mathcal {E} \left(c _ {k} ^ {\text {i n t r i n s i c}}\right) \right\| _ {2} ^ {2} + \gamma\right), \tag {6} \\ \end{array} +$$ + +where $c_k^{\mathrm{intrinsic}}$ represents intrinsic tokens not associated with intrinsic concept $j$ , ensuring that each intrinsic token is uniquely mapped to its corresponding attribute. + +# 4.3. Concept refinement + +Optimizing intrinsic concepts solely through token embeddings is challenging, often leading to misaligned and inconsistent representations of specific visual attributes such as color and texture. To enhance the accuracy and consistency of these intrinsic concepts, after completing Phase two, we perform concept refinement by fine-tuning both the U-Net denoising network and the text encoder $\mathcal{E}$ for a limited number of training steps. This fine-tuning ensures that the intrinsic concepts are precisely aligned with their corresponding visual attributes. + +# 4.4. Overall training objective + +The overall training objective of the ICE framework integrates multiple loss components to facilitate both reconstruction quality and concept differentiation. The total loss $\mathcal{L}_{\mathrm{total}}$ is given by: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {\text {r e c o n}} + \lambda_ {\text {a t t}} \mathcal {L} _ {\text {a t t}} + \lambda_ {\text {t r i p l e t}} \mathcal {L} _ {\text {t r i p l e t}}. \tag {7} +$$ + +$\mathcal{L}_{\mathrm{recon}}$ , as defined in Eq.(2), is the reconstruction loss between the predicted noise and the actual noise with text prompt1 conditioning $\mathbf{p}$ as: + +Phase one, p: " a photo of $c_{i}^{\mathrm{inspec}}$ & $c_{i}^{\mathrm{conspec}}$ " + +Phase two, p: “ a photo of $\mathbb{C}^{\mathrm{intrinsic}}$ & $c_{i}^{\mathrm{conspec}}$ , + +where $\mathbb{C}^{\mathrm{intrinsic}}$ is a set of $c_{j}^{\mathrm{intrinsic}}$ tokens. $\mathcal{L}_{\mathrm{att}}$ is the attention mask loss, which employs Wasserstein loss, following [9], to align the attention regions with the segmentation masks: + +$$ +\mathcal {L} _ {\mathrm {a t t}} = \mathcal {W} \left(\mathbf {A} _ {i}, \mathbf {m} _ {i}\right), \tag {8} +$$ + +where $\mathcal{W}$ denotes the Wasserstein distance, $\mathbf{A}_i$ is the attention map corresponding to concept $c_{i}$ , and $\mathbf{m}_i$ is the segmentation mask obtained in the first stage. $\mathcal{L}_{\mathrm{triplet}}^{\mathrm{obj}}$ and $\mathcal{L}_{\mathrm{triplet}}^{\mathrm{intrinsic}}$ are the object-level and intrinsic triplet losses for different phases: + +$$ +\mathcal {L} _ {\text {t r i p l e t}} ^ {\mathrm {o b j}} = \sum_ {i} \mathcal {L} _ {\text {t r i p l e t}, i} ^ {\mathrm {o b j}}, \quad \mathcal {L} _ {\text {t r i p l e t}} ^ {\text {i n t r i n s i c}} = \sum_ {j} \mathcal {L} _ {\text {t r i p l e t}, j} ^ {\text {i n t r i n s i c}}. \tag {9} +$$ + +The weighting factors $\lambda_{\mathrm{att}}$ and $\lambda_{\mathrm{triplet}}$ balance the contributions of the attention and triplet losses relative to the reconstruction loss in each phase, ensuring stable optimization across all objectives. + +# 5. Experiments + +Implementation details. We implement our ICE framework using Stable Diffusion v2.1 for both stages. In the first stage, for the Image-to-Text Retriever module $\mathcal{T}$ , we utilize the CLIP-based SpLiCE framework [4], while the T2I zero-shot Segmentor $S$ employs DiffSeg [30]. We adapt the input and output interfaces of these training-free modules to accommodate our automatic concept localization process. Specifically, modifications were made to ensure seamless integration within the framework, allowing for efficient and accurate extraction of text-based concepts and their corresponding masks. The training process in the second stage is divided into two phases: Phase one focuses on learning object-level concepts and is conducted over 400 training steps, followed by Phase two, which targets intrinsic concepts and also comprises 400 training steps. Subsequently, concept refinement is performed with another 300 steps to enhance the quality and coherence of the learned concepts. We assign the weighting factors $\lambda_{\mathrm{att}} = 1\times 10^{-5}$ and $\lambda_{\mathrm{triplet}} = 1$ . For the triplet loss margin $\gamma$ value, please refer to the supplementary material. Finally, we compute a prior preservation loss [26] to promote diverse image generation while retaining class-specific knowledge. All experiments are conducted on a single NVIDIA RTX 3090 GPU. + +Baseline & evaluation metrics. To evaluate the effectiveness of ICE, we compare our framework with the state-of-the-art method, ConceptExpress [9], on the Unsupervised Concept Extraction (UCE) benchmarks. Specifically, we utilize $D1$ dataset, which is based on the Unsplash datasets, providing a diverse and extensive collection of unlabeled images. Moreover, the UCE benchmarks assess models based on two primary metrics: (1) Concept Similarity, which includes Identity Similarity $(\mathrm{SIM}^I)$ measuring how accurately individual concepts are recreated, and Compositional Similarity $(\mathrm{SIM}^C)$ evaluating the overall coherence of the generated image based on the extracted concepts. (2) Classification Accuracy assesses the disentangle-ment and classification performance of the extracted concepts by evaluating how well the concepts can be used to classify images correctly. These metrics provide a compre + +hensive evaluation of both the fidelity and utility of the extracted concepts, enabling a robust quantitative comparison between ICE and ConceptExpress. + +# 5.1. Quantitative results + +As shown in Table 2 & 3, ICE achieves higher scores in all metrics compared to previous methods, demonstrating its superior ability to recreate and compose visual concepts from unlabelled images accurately. + +Table 2. Performance of ICE and relevant works on Unsupervised Concept Extraction (UCE) benchmarks using CLIP [23] encoder. + +
Method\(SIM^I\)(%)\(SIM^C\)(%)\(ACC^1\)(%)\(ACC^3\)(%)
Break-A-Scene [1]0.6270.7730.1740.282
ConceptExpress[9]0.6890.7840.2630.385
ICE (Ours)0.7380.8220.3250.518
+ +Table 3. Performance of ICE and relevant works on Unsupervised Concept Extraction (UCE) benchmarks using DINO [6] encoder. + +
Method\( SIM^I \) (%)\( SIM^C \) (%)\( ACC^1 \) (%)\( ACC^3 \) (%)
Break-A-Scene [1]0.2540.5100.2020.315
ConceptExpress [9]0.3190.5680.3240.470
ICE (Ours)0.6770.7550.4760.638
+ +This enhanced performance can be attributed to two key factors: (1) Enhanced Concept Extraction: Our framework effectively extracts text-based concepts and their corresponding masks, resulting in superior masking accuracy as illustrated in Figure 7. Unlike other methods that may only mask object-level regions, ICE also retrieves the most similar text-based descriptors. This dual extraction provides additional guidance and robust initialization for the model, ensuring that both the spatial and semantic aspects of each concept are accurately captured. (2) Structured Concept Learning Approach: ICE's two-stage framework adopts a structured approach to concept learning, systematically decomposing object-level concepts into intrinsic attributes. This structured methodology facilitates a more granular and interpretable breakdown of visual elements, leading to the discovery of richer and more meaningful concepts. Specifically, ICE achieves CLIP Identity Similarity $(\mathrm{SIM}^I)$ of $73.8\%$ , compared to ConceptExpress's $68.9\%$ , indicating a substantial improvement in accurately identifying object-level concepts. Additionally, ICE attains a Compositional Similarity $(\mathrm{SIM}^C)$ of $82.2\%$ , surpassing ConceptExpress's $78.4\%$ , which underscores our framework's enhanced capability in understanding and composing complex visual attributes. These factors collectively contribute to the superior performance of ICE, highlighting the effectiveness of our framework in systematically uncovering and refining both object-level and intrinsic visual concepts from unlabeled images. + +![](images/d7508154afc88686741b650f3dc41fd4917c2873e85d0785055d78a68f840d5b.jpg) +Input image $x$ + +Extracted mask $m_{i}$ + +![](images/bcfbac42c7da04d2a791750997d3cc8355c35fffe87992dcefecde5e2fa75585.jpg) +"Bust" +Extracted mask $m_{i}$ + +Retrived text $c_{i}$ + +![](images/f6e07ac92b04635b532c70939a6b6d65213a3a6a10e02e6f3610d4d5b365b47e.jpg) +"Spout" + +![](images/44efb1ab860c9c03cc39f6bef074fb38e6150c951a3d7714e04d64caef0174ba.jpg) +"Bust" + +![](images/a6bd3c7a3d32481cc26c3da3485315c0410d89578b08e7671d9251346f5fdeb9.jpg) +"Cone" + +![](images/36a554f0b73c52cb401821572c61d2e561654a0848a352feabd644cf0c594ec7.jpg) +"Ornament" + +![](images/7cbf3f514973643e3d2650bf1c801e7969dbad7890632b9aeaa29dc9fb4c8428.jpg) +"Rat" + +![](images/e0a25c1868956268280a0560affd7afd21c604d99cc50c9d034b089ae2b1acb5.jpg) +"Wooden" + +![](images/bd82722cf765e4b2415cfdffdfcdf9250fc13de6a696993cefb7ecdc12ad0cf4.jpg) +"Wooden" + +![](images/49572d6efb2343757f85df62054c1cf6e7577a07c5524fd4505a6e7421c1eb26.jpg) +"Pines" + +![](images/3d9085d5686db6dc4f8bec30c562a1f9c8982e0cd20461f0c480c71842ba444d.jpg) +"Chapter" +${13}/{14}$ + +![](images/c68b3da89367bb3cce1f3412d22b7a3ac0df4dde391a5170c9c5bd1427a9619b.jpg) +Figure 6. Qualitative results of the ICE framework demonstrating its systematic concept discovery process. Column 1: Input images. Column 2: Extracted text-based concepts and their corresponding masks obtained from Stage One: Automatic Concept Localization. Columns 3 & 4: Generated images of learned object-level and intrinsic concepts derived from Stage Two: Structured Concept Learning. + +ed concepts a + +ages of learned object-level and int + +Retrived text $c_{i}$ + +![](images/ea14d6f07c734134a2480a9b1a421271005bffc9a3ef60491ffc3fc67b25898b.jpg) +Object-level concepts + +![](images/71beebb5974c120a3cd1cbf916e712f2128075e21fe3434b03225989ab7f949b.jpg) + +![](images/d18b86cbf0ac814f23b776aca8e838f72e1ba53d3f1910ab55379ea6aab93ff8.jpg) + +![](images/39b43670b729c78aba84611f4f4d316fdddf1ecf4780286c0397da1e51053af2.jpg) + +![](images/18d784b5b57590c11339e68f51d7667182fb99ce38ca1b9b2fdafab9408053b4.jpg) + +![](images/d0d6473201dbb9e31a141a8075d42b2ebc2d54aa9d61984e22c81be92ff9b0f2.jpg) + +![](images/003698f65df7c20d9323d31332dafeca0fd784babfbf5b2c8194c36139a1a5c7.jpg) + +![](images/9e223f4985b4b95143057b15b806b8d245964acb9ce24108a4c51097c29ef86e.jpg) + +![](images/d9f8f4a05f3210f49cc6e843082f26f4058dc5f87f669793e8122cf507006427.jpg) + +![](images/911be1d34710dcafd4c87a8b9c736226e484703c420460aa54783acd676f15e2.jpg) + +![](images/cb515daad63eeb70026ccda8c5907a2049bbbe26952d0689b52cf6ae88558c0d.jpg) +Object + +![](images/372c56ae06ef95feb11eceb1c39c954e800d011ba6c6e23ea8f9df9d8b260ce8.jpg) + +![](images/b3a31eac67b73c57571526039bc44c6b88fe42c743259d894de3624d5f45fa76.jpg) + +![](images/21ba1b0515526447a961e128581ed2c924bf6fe1cf36a9f36e9b7063b2607472.jpg) + +![](images/7f70af42a461f2b5bc9f206d506465fd2b2436bfe187e1cc7dafa2f4e0c48bd4.jpg) + +![](images/71abfaadb1912d53f675c13b6865281883cffc2ce0233029f005c54860023a30.jpg) + +![](images/116bfc79088af567816e3e5949e5a6f69898e8166ff8673a6e0b452013299be9.jpg) + +![](images/15e96c0aa635874fbd5bab5e9784c78b16d82b892de738c22f23455fbaec8702.jpg) + +![](images/5eff3e65c8b7bf434cb4e2550c8498edd1800676e1b88bc81e863b9ce32e4145.jpg) + +![](images/d912698def2d005067d1c0423f85d827f08d3400ebbe20d22834643948cd7252.jpg) + +![](images/f4cd31a0f838010406b5dde62ce4e57fdf45e7256aef17ae284ce9eca9f454f2.jpg) +Intrinsic concepts +Material + +![](images/a539b214618624a6761bb76ff0feadbc1f69ea6db88ff343e7c275ff65552160.jpg) + +![](images/9d141318618ed837da517facd02e300133b4e2cff29f4b4412243e54864e60d9.jpg) + +![](images/1c2e15953df712ac6d87b5c70059d22e8f23181f53bc19af058a15ed5a36fddf.jpg) + +![](images/f821de92b0d31d3623158e9d004c01d649939fc0063f9c0949d6bf5b7a813f9e.jpg) + +![](images/5eb9b66ace13319b694f46dcaf92935c5ff9adc6526cdfb580d6f9643f485cbc.jpg) + +![](images/059670f93a9891a330ab07f080b5e7c70b6e8fd7b38c8efc964df1aa5a745373.jpg) + +![](images/3b74a28dfd3bd767ce5a031d4546598d8eb8c2092d75dfaa3d1b844686122b6b.jpg) + +![](images/e40a0315a20242769a5ef29622f35843d4011be00ba148f8d9ccf159c55c69aa.jpg) + +![](images/85207812dc6184607e49026fe773046655a14953e73d8ba9eddbc21382bc0004.jpg) + +![](images/170880547fd7af43f1b2326e1e48ef0e29f0438e7b29cd10e3b01404b293a4f3.jpg) +Colour + +![](images/19ee7cad0c1ccfa7e1d2a2f9de75265cd061592667f6bb2085046607b21624e3.jpg) + +![](images/20182432f1f3fe2a0cee4f04b29d9fc0aa524bfde1dbad8226ae1a4b7f0a4bbd.jpg) + +![](images/9661a9fa5f556f129b80e143e606c157c5979a68dce4b299ed8a03a305278aba.jpg) + +![](images/aee9329f8e5b9e7c94d761a72233d516961a49b6a35ee5e6d2b47c2e14d2f252.jpg) + +![](images/b40f12dd3ebea1b7b49cb1039431d7a8545e91480ab9e9b8708f918eccf86bfe.jpg) + +![](images/a34a703320c4b5337f1cee02e110381fe9a26b72622ca8000dcf59f0302bef19.jpg) + +![](images/0626c06bea49fafd325040eb7abc9e09a7bc8d029723b8bfd4aae1a494e17ea8.jpg) + +![](images/a268e652d80cbc83a806d189a7690e28e990880eabe087160e7ab42acef969a7.jpg) + +![](images/86a681c17a33a043643391e023952ea9bef8cc901a14fba327bf9d9f2f53b46d.jpg) + +# 5.2. Qualitative results + +To demonstrate the effectiveness of the proposed ICE framework, we present qualitative results, as illustrated in + +Figure 6, showcasing ICE's ability to accurately extract and decompose visual concepts systematically. + +First, given an input image $x$ , ICE extracts corresponding concept masks and retrieves relevant text-based descrip + +![](images/37c036b79aaa8e5c41fe35427099d65316ad33cc1c883b2af4015cf481ba0993.jpg) +Figure 7. Qualitative comparison of segmentation results between ConceptExpress and our framework. Our framework not only delivers more accurate segmentation of object-level regions but also effectively retrieves the most relevant text-based concepts, demonstrating its unique capabilities over ConceptExpress. + +![](images/686aa1e4224702a41b9b47c9712e321b29f72ef11b255db4276a1ef42e01dd91.jpg) + +![](images/af3a018444c013dd344bd8b0aa21a7989964386c7db0e4a8c3576dcde42e645a.jpg) + +![](images/6276d4649ae5cfbe2ea6796f3d990cff27c819e6a3d4d339dd3f9624845148b2.jpg) + +![](images/9d9515a877f54e7a838fb4cd1e387319d9d87a3a688d78d005a4d91b272398cb.jpg) + +![](images/4494398349677432efd5b785a09799a9d67774bde15135ac5202fa533d181a98.jpg) + +![](images/b18eac24960c1508e208f030115c4da5b3c2fc22977e01f76e6b7af3934312d3.jpg) + +![](images/3d503ed9fda1fbdb20388c9edbb457cc6451466e3dbd555a37cf8205a4df6a16.jpg) + +![](images/dac1f2ad33523c1e9a7866bcb673db1096a894e110046582bb5bea7c6569832e.jpg) + +Table 4. Ablation study on ICE model components on the $D1$ dataset. + +
Model's componentsCLIP [23] encoderDINO [6] encoder
MethodStage One's maskStage Two's token init.Stage Two learningSIM1(%)SIMC(%)ACC1(%)ACC3(%)SIM1(%)SIMC(%)ACC1(%)ACC3(%)
ConceptExpressXXX0.6890.7840.2630.3850.3190.5680.3240.470
ICE w. mask✓XX0.7100.7810.3070.4560.4930.6010.3950.604
ICE w.o Stage Two✓✓X0.7260.8070.3010.4520.5010.6210.4110.604
ICE w.o text init.✓X✓0.7220.8140.3200.4750.5480.6430.4490.627
ICE (Ours)✓✓✓0.7380.8220.3250.5180.6770.7550.4760.638
+ +Table 5. Quantitative comparison of generated masks on the $D1$ dataset. + +
MethodmIoURecallPrecision
ConceptExpress0.4830.6760.657
ICE (Ours)0.6350.8930.720
+ +tors, as shown in the second column of the figure. These masks precisely delineate regions within the image, ensuring accurate localization of each concept. Next, leveraging the extracted masks and retrieved texts, ICE learns distinct object-level concepts, displayed in the third column. For example, the framework identifies objects such as “Bust” and “Cone” and associates them with their respective semantic descriptors. Finally, utilizing the information from the object-level concepts, ICE systematically discovers intrinsic concepts, including attributes like material and colour, through its second-stage processing. This decomposition enables the framework to uncover compositional intrinsic concepts such as the object type, material, and colour, as demonstrated in the fourth column. + +This structured approach highlights the superiority of our framework by effectively leveraging a single T2I model to systematically uncover both object-level and intrinsic visual concepts. The ability to deconstruct an image into its fundamental components not only validates the robustness of ICE but also illustrates the potential of T2I models to facilitate comprehensive and structured concept learning in an unsupervised manner. + +# 5.3. Model component analysis + +ICE ablation study. To validate the effectiveness of our ICE framework, we conduct an ablation study on its individual components: Stage One's retrieved mask, Stage Two's token embedding initialization using text retrieved from Stage One, and Stage Two's Structured Concept Learning. The results, as shown in Table 4, demonstrate that each component significantly enhances performance. Furthermore, the complete ICE framework outperforms all its variants across every metric in the UCE benchmarks. + +ICE Stage One's mask quality. Furthermore, Table 5 compares our Stage One generated masks with ConceptExpress using $D1$ 's manually annotated ground-truth masks. We evaluate the segmentation performance of generated masks + +against ground-truth masks using the Hungarian algorithm to find the optimal matching between the sets of masks. We evaluate each pair of predicted and ground-truth masks using segmentation criterions, forming a similarity matrix. The Hungarian algorithm is then applied to this matrix to determine the optimal one-to-one correspondence that maximizes the overall similarity. Overall, our Stage One technique achieves better segmentation results across different criterion. Notably, our Stage One masks exhibit significantly higher recall rates, indicating that a greater proportion of the relevant object-level regions in the annotated area are successfully captured. This higher recall demonstrates that ICE's Automatic Concept Localization method in Stage One is more effective at identifying and extracting meaningful concepts, reducing the likelihood of missing important object-level concepts compared to previous methods. + +# 6. Conclusion + +In this paper, we introduce ICE, a novel framework for automatically and systematically extracting intrinsic concepts from a single image via diffusion models. Our framework consists of two stages. In Stage One, we design a concept localization module based on the T2I model that effectively retrieves the most relevant text-based concepts and their corresponding masks. Subsequently, in Stage Two, we introduce a structured concept learning method to systematically decompose object-level concepts into their intrinsic attributes, resulting in richer and more interpretable concept representations. In summary, our approach effectively harnesses Text-to-Image (T2I) models to enhance concept learning, laying the groundwork for further advancements in intrinsic visual concept extraction. + +Acknowledgments. This work is supported by the Hong Kong Research Grants Council - General Research Fund (Grant No.: 17211024). + +# References + +[1] Omri Avrahami, Kfir Aberman, Ohad Fried, Daniel Cohen-Or, and Dani Lischinski. Break-a-scene: Extracting multiple concepts from a single image. In SIGGRAPH Asia, 2023. 2, 3, 6 +[2] Omri Avrahami, Thomas Hayes, Oran Gafni, Sonal Gupta, Yaniv Taigman, Devi Parikh, Dani Lischinski, Ohad Fried, and Xi Yin. Spatext: Spatio-textual representation for controllable image generation. In CVPR, 2023. 2 +[3] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. In ICML, 2023. 2 +[4] Usha Bhalla, Alex Oesterling, Suraj Srinivas, Flavio P Calmon, and Himabindu Lakkaraju. Interpreting clip with sparse linear concept embeddings (splice). In NeurIPS, 2024. 6 +[5] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023. 2 +[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 6, 8 +[7] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In ICLR, 2023. 2, 3 +[8] Shaozhe Hao, Kai Han, Shihao Zhao, and Kwan-Yee K Wong. Vico: Detail-preserving visual condition for personalized text-to-image generation. arXiv preprint arXiv:2306.00971, 2023. 2 +[9] Shaozhe Hao, Kai Han, Zhengyao Lv, Shihao Zhao, and Kwan-Yee K Wong. Conceptexpress: Harnessing diffusion models for single-image unsupervised concept extraction. In ECCV, 2024. 2, 3, 5, 6 +[10] Eric Hedlin, Gopal Sharma, Shweta Mahajan, Hossam Isack, Abhishek Kar, Andrea Tagliasacchi, and Kwang Moo Yi. Unsupervised semantic correspondence using stable diffusion. In NeurIPS, 2023. 1 +[11] Jonathan Ho, Tim Salimans, Alexey A Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. In NeurIPS, 2022. 2 +[12] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Weli-hinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 2 +[13] Chen Jin, Ryutaro Tanno, Amrutha Saseendran, Tom Diethe, and Philip Teare. An image is worth multiple words: Discovering object level concepts using multi-concept prompt learning. In ICML, 2023. 2, 3 +[14] Laurynas Karazija, Iro Laina, Andrea Vedaldi, and Christian Rupprecht. Diffusion models for zero-shot open-vocabulary segmentation. In ECCV, 2023. 1 +[15] Sharon Lee, Yunzhi Zhang, Shangzhe Wu, and Jiajun Wu. Language-informed visual concept learning. In ICLR, 2023. 2, 3 + +[16] Alexander C. Li, Mihir Prabhudesai, Shivam Duggal, Ellis Brown, and Deepak Pathak. Your diffusion model is secretly a zero-shot classifier. In ICCV, 2023. 1 +[17] Dongxu Li, Junnan Li, and Steven CH Hoi. Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. In NeurIPS, 2023. 2 +[18] Xinghui Li, Jingyi Lu, Kai Han, and Victor Prisacariu. Sd4match: Learning to prompt stable diffusion model for semantic matching. In CVPR, 2024. 1 +[19] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In CVPR, 2022. 2 +[20] Minheng Ni, Yabo Zhang, Kailai Feng, Xiaoming Li, Yiwen Guo, and Wangmeng Zuo. Ref-diff: Zero-shot referring image segmentation with generative models. arXiv preprint arXiv:2308.16777, 2023. 1 +[21] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob Mcgrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In ICML, 2022. 2 +[22] Zeju Qiu, Weiyang Liu, Haiwen Feng, Yuxuan Xue, Yao Feng, Zhen Liu, Dan Zhang, Adrian Weller, and Bernhard Scholkopf. Controlling text-to-image diffusion by orthogonal finetuning. In NeurIPS, 2023. 2 +[23] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 3, 6, 8 +[24] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents. arXiv preprint arXiv:2204.06125, 2022. 2 +[25] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2 +[26] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, 2023. 2, 3, 6 +[27] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, 2022. 2 +[28] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS, 2022. 2 +[29] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. In ICLR, 2022. 2 +[30] Junjiao Tian, Lavisha Aggarwal, Andrea Colaco, Zsolt Kira, and Mar Gonzalez-Franco. Diffuse, attend, and segment: + +Unsupervised zero-shot segmentation using stable diffusion. In CVPR, 2024. 1, 6 +[31] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In CVPR, 2023. 2 +[32] Yael Vinker, Andrey Voynov, Daniel Cohen-Or, and Ariel Shamir. Concept decomposition for visual exploration and inspiration. In SIGGRAPH Asia, 2023. 2, 3 +[33] Jinglong Wang, Xiawei Li, Jing Zhang, Qingyuan Xu, Qin Zhou, Qian Yu, Lu Sheng, and Dong Xu. Diffusion model is secretly a training-free open vocabulary semantic segmenter. arXiv preprint arXiv:2309.02773, 2023. 1 +[34] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Weixian Lei, Yuchao Gu, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In ICCV, 2022. 2 +[35] Changming Xiao, Qi Yang, Feng Zhou, and Changshui Zhang. From text to mask: Localizing entities using the attention of text-to-image diffusion models. In Neurocomputing, 2023. 1 +[36] Zhi Xu, Shaozhe Hao, and Kai Han. Cusconcept: Customized visual concept decomposition with diffusion models. In WACV, 2025. 2 +[37] Junyi Zhang, Charles Herrmann, Junhwa Hur, Luisa Polania Cabrera, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. A tale of two features: Stable diffusion complements dino for zero-shot semantic correspondence. In NeurIPS, 2023. 1 +[38] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 2 +[39] Yanbing Zhang, Mengping Yang, Qin Zhou, and Zhe Wang. Attention calibration for disentangled text-to-image personalization. In CVPR, 2024. 2 \ No newline at end of file diff --git a/2025/ICE_ Intrinsic Concept Extraction from a Single Image via Diffusion Models/images.zip b/2025/ICE_ Intrinsic Concept Extraction from a Single Image via Diffusion Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..548dde36557a0c53048eaeaea7da1bec29a0e158 --- /dev/null +++ b/2025/ICE_ Intrinsic Concept Extraction from a Single Image via Diffusion Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68b9792a2492c24597acc103e932b5c1da106060649d856cf7eb471d61bb4452 +size 660546 diff --git a/2025/ICE_ Intrinsic Concept Extraction from a Single Image via Diffusion Models/layout.json b/2025/ICE_ Intrinsic Concept Extraction from a Single Image via Diffusion Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..de169c03aa1d41f191a8cc210885b59c29a56114 --- /dev/null +++ b/2025/ICE_ Intrinsic Concept Extraction from a Single Image via Diffusion Models/layout.json @@ -0,0 +1,11718 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 69, + 103, + 541, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 103, + 541, + 121 + ], + "spans": [ + { + "bbox": [ + 69, + 103, + 541, + 121 + ], + "type": "text", + "content": "ICE: Intrinsic Concept Extraction from a Single Image via Diffusion Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 212, + 143, + 397, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 143, + 397, + 157 + ], + "spans": [ + { + "bbox": [ + 212, + 143, + 397, + 157 + ], + "type": "text", + "content": "Fernando Julio Cendra Kai Han" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 195, + 163, + 414, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 163, + 414, + 178 + ], + "spans": [ + { + "bbox": [ + 195, + 163, + 414, + 178 + ], + "type": "text", + "content": "Visual AI Lab, The University of Hong Kong" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 198, + 180, + 405, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 180, + 405, + 190 + ], + "spans": [ + { + "bbox": [ + 198, + 180, + 405, + 190 + ], + "type": "text", + "content": "fcendra@connect.hku.hk, kaihanx@hku.hk" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 59, + 196, + 555, + 369 + ], + "blocks": [ + { + "bbox": [ + 59, + 196, + 555, + 369 + ], + "lines": [ + { + "bbox": [ + 59, + 196, + 555, + 369 + ], + "spans": [ + { + "bbox": [ + 59, + 196, + 555, + 369 + ], + "type": "image", + "image_path": "0101f83f761da9715d805a7df8cef39b4842ec23a4cf74794c7c39908a32de13.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 369, + 555, + 414 + ], + "lines": [ + { + "bbox": [ + 55, + 369, + 555, + 414 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 555, + 414 + ], + "type": "text", + "content": "Figure 1. We showcase a structured approach for defining visual concepts within an image, where object-level concepts are identified and analyzed to reveal their underlying intrinsic attributes, such as color, material, and shape. We present the ICE (Intrinsic Concept Extraction) framework, which leverages Text-to-Image (T2I) models to systematically discover these concepts, providing a more effective method for learning visual concepts." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 422, + 200, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 422, + 200, + 434 + ], + "spans": [ + { + "bbox": [ + 151, + 422, + 200, + 434 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 449, + 297, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 449, + 297, + 689 + ], + "spans": [ + { + "bbox": [ + 54, + 449, + 297, + 689 + ], + "type": "text", + "content": "The inherent ambiguity in defining visual concepts poses significant challenges for modern generative models, such as the diffusion-based Text-to-Image (T2I) models, in accurately learning concepts from a single image. Existing methods lack a systematic way to reliably extract the interpretable underlying intrinsic concepts. To address this challenge, we present ICE, short for Intrinsic Concept Extraction, a novel framework that exclusively utilizes a T2I model to automatically and systematically extract intrinsic concepts from a single image. ICE consists of two pivotal stages. In the first stage, ICE devises an automatic concept localization module to pinpoint relevant text-based concepts and their corresponding masks within the image. This critical stage streamlines concept initialization and provides precise guidance for subsequent analysis. The second stage delves deeper into each identified mask, decomposing the object-level concepts into intrinsic concepts and general concepts. This decomposition allows for a more granular and interpretable breakdown of visual elements. Our framework demonstrates superior performance on in" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 423, + 555, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 555, + 459 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 555, + 459 + ], + "type": "text", + "content": "trinsic concept extraction from a single image in an unsupervised manner. Project page: https://visual-ai.github.io/ice" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 487, + 395, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 487, + 395, + 499 + ], + "spans": [ + { + "bbox": [ + 314, + 487, + 395, + 499 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 508, + 555, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 508, + 555, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 508, + 555, + 592 + ], + "type": "text", + "content": "Text-to-Image (T2I) models have undergone significant advancements, enabling the generation of high-quality images from textual prompts and bridging natural language processing with visual content creation. These capabilities have driven applications in creative design, entertainment, and assistive technologies, underscoring the pivotal role of T2I models in generative AI." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "Beyond image generation, emerging research has uncovered that T2I models excel in tasks such as image classification, segmentation, and semantic correspondence. Studies [10, 14, 16, 18, 20, 30, 33, 35, 37] have shown their effectiveness in these areas. This versatility suggests that diffusion-based T2I models encapsulate substantial world knowledge, allowing them to comprehend and recreate visual concepts with precision and detail. Building on this potential, our work explores the use of T2I models like Stable Diffusion for concept learning. Instead of solely generating" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 702, + 144, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 144, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 144, + 713 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23734" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": "images from text, we investigate how these models invert the generative process to systematically define and learn the fundamental concepts that comprise a given unlabelled image, thereby uncovering the essential concepts that enable the synthesis of complex visual scenes." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 147, + 295, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 147, + 295, + 350 + ], + "spans": [ + { + "bbox": [ + 56, + 147, + 295, + 350 + ], + "type": "text", + "content": "Despite their promising potential, concept learning within T2I models faces several inherent challenges. Defining visual concepts remains inherently ambiguous, complicating the reliability with which these models can discover and utilize essential concepts. Previous methods [7, 8, 26] such as Textual Inversion (TI) and DreamBooth require learning multiple similar images to encapsulate a concept. Other frameworks, such as Break-A-Scene [1], MCPL [13], and ConceptExpress [9], focus on extracting object-level concepts within a single image. Moreover, these methods often struggle to provide clear and interpretable concept representations, hindering their effectiveness in capturing the nuanced and multifaceted nature of visual scenes. The ambiguity in defining visual concepts not only impedes reliable concept discovery but also poses significant challenges for human annotators attempting to delineate these concepts manually." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 364, + 295, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 364, + 295, + 532 + ], + "spans": [ + { + "bbox": [ + 55, + 364, + 295, + 532 + ], + "type": "text", + "content": "Addressing these challenges requires a fundamental re-evaluation of how we define and learn visual concepts within an image. Previous works, such as Inspiration Tree [32], have attempted to construct hierarchical representations by organizing tokens into tree structures to capture different concepts within an image. Similarly, LangInt [15] uses Visual Question Answering (VQA) techniques to guide concept learning, but it requires training separate encoders for each concept axis, thereby failing to fully utilize their inherent hierarchical structures. Consequently, there is a need for more systematic and structured methods to define and learn concepts from images, enabling a more granular understanding of the fundamental building blocks of visual representation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "text", + "content": "In response to these challenges, we present ICE, a novel two-stage framework designed to automatically and systematically extract intrinsic concepts using a single T2I model. ICE first devises an automatic concept localization module to retrieves text-based concepts and their corresponding masks. This enhances initialization and provides clear guidance for subsequent processing. The framework then systematically decomposes these object-level concepts into intrinsic and general concepts, enabling a more granular and interpretable breakdown of visual elements. Through comprehensive experiments, ICE demonstrates its effectiveness in leveraging T2I models to systematically uncover diverse visual concepts, facilitating a structured and detailed understanding of image components." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 314, + 71, + 400, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 400, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 400, + 84 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 91, + 553, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 553, + 235 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 553, + 235 + ], + "type": "text", + "content": "Text-to-Image generation. Diffusion-based methods have significantly advanced the field of text-to-image (T2I) generation, pushing the boundaries of image synthesis to unprecedented levels. Early approaches such as DALL-E 2 [24], Imagen [27], and GLIDE [21] demonstrated the potential of diffusion models in generating high-fidelity images from textual descriptions. Latent Diffusion Models (LDMs) [25] further enhanced this capability by operating in a latent space, enabling more efficient and scalable training. Building upon the foundation laid by LDMs, Stable Diffusion (SD) [25] is trained on the LAION-5B dataset [28] and has set new benchmarks in T2I synthesis." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 235, + 553, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 235, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 235, + 553, + 342 + ], + "type": "text", + "content": "Beyond basic image generation, diffusion models have been adapted for a variety of specialized tasks. Controllable generation [21, 22, 38] allows for nuanced adjustments based on user input, while applications such as global and local editing [5, 31] enable fine-grained modifications to specific regions of an image. Additionally, these models have been extended to video generation and editing [11, 29, 34], inpainting [19], and scene generation [2, 3], showcasing their versatility across use cases." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 342, + 553, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 553, + 461 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 553, + 461 + ], + "type": "text", + "content": "Generative concept learning. Generative concept learning aims to decompose and understand the fundamental elements that constitute complex visual scenes. Textual Inversion (TI) [7] was among the first methods to learn single concepts from multiple images by introducing new words and optimizing their corresponding token embeddings. Similarly, DreamBooth [26] extends this approach by leveraging less frequently used words and optimizing both the denoising U-Net and introducing a preservation loss to enhance the quality of the learned concepts." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 463, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 463, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 553, + 713 + ], + "type": "text", + "content": "Break-A-Scene [1] relies heavily on human-annotated masks to isolate and learn distinct concepts from a single image, limiting its applicability in settings where such annotations are unavailable Concurrent works like MCPL [13] and DisenDiff [39] address similar challenges but require concept-descriptive text captions or predefined class names, making them less feasible for more general tasks. Recent advancements, such as ConceptExpress [9], have reduced reliance on human annotations by automatically identifying and learning multiple concepts within an image. However, these methods primarily treat concepts at the object-level, neglecting intrinsic attributes such as colour and material properties. Inspiration Tree [32] introduces a preliminary approach to concept decomposition by splitting a token into two tokens to force the learning of different concepts. Despite this, it lacks reliable interpretation and structured guidance, resulting in less coherent concept representations. LangInt [15] and CusConcept [36] attempt to learn intrinsic concepts using VQA models [12, 17] to query specific attributes of an image and independently learn these intrinsic concepts. However, these approaches are limited to scenar" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23735" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 89, + 72, + 265, + 201 + ], + "blocks": [ + { + "bbox": [ + 89, + 72, + 265, + 201 + ], + "lines": [ + { + "bbox": [ + 89, + 72, + 265, + 201 + ], + "spans": [ + { + "bbox": [ + 89, + 72, + 265, + 201 + ], + "type": "image", + "image_path": "4bd400f44d5ac11e467b49aaa80561f356871c34290918e4e750a0f0a26215c3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 207, + 297, + 242 + ], + "lines": [ + { + "bbox": [ + 55, + 207, + 297, + 242 + ], + "spans": [ + { + "bbox": [ + 55, + 207, + 297, + 242 + ], + "type": "text", + "content": "Figure 2. Concept definition hierarchy illustrating how object-level concepts are decomposed into intrinsic attributes, including object type, colour, material and other intrinsics." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 248, + 296, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 248, + 296, + 272 + ], + "spans": [ + { + "bbox": [ + 55, + 248, + 296, + 272 + ], + "type": "text", + "content": "ios with a single object-level concept per image, failing to handle images with multiple concepts." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 273, + 296, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 273, + 296, + 440 + ], + "spans": [ + { + "bbox": [ + 55, + 273, + 296, + 440 + ], + "type": "text", + "content": "Comparison and positioning of ICE. In contrast to existing methodologies, our proposed framework, ICE, offers a unified and structured approach to automatically and systematically discover intrinsic concepts within an image using a single T2I model. Unlike previous methods, ICE not only identifies object-level concepts but also decomposes them into intrinsic attributes such as colour and material (see Figure 2), providing a more comprehensive and interpretable representation of visual concepts. This two-stage decomposition ensures that both general and specific attributes are captured, enhancing the model's ability to understand and generate nuanced visual content. For an overview of the differences between our framework and existing methods, please refer to Table 1." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 57, + 461, + 296, + 572 + ], + "blocks": [ + { + "bbox": [ + 85, + 449, + 262, + 460 + ], + "lines": [ + { + "bbox": [ + 85, + 449, + 262, + 460 + ], + "spans": [ + { + "bbox": [ + 85, + 449, + 262, + 460 + ], + "type": "text", + "content": "Table 1. Comparison of ICE and relevant works." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 461, + 296, + 572 + ], + "lines": [ + { + "bbox": [ + 57, + 461, + 296, + 572 + ], + "spans": [ + { + "bbox": [ + 57, + 461, + 296, + 572 + ], + "type": "table", + "html": "
Learned concept(s)Framework details
MethodObject-level conceptsIntrinsic conceptsMulti conceptsSingle imageExtra information
Textual Inversion [7]✓XXX-
Dreambooth [26]✓XXX-
Inspiration Tree [32]X✓XX-
LangInt [15]✓✓X✓VQA-guided
Break-A-Scene [1]✓X✓✓Mask
MCPL [13]✓X✓✓Text
ConceptExpress [9]✓X✓✓-
ICE (Ours)✓✓✓✓-
", + "image_path": "eb8b7670f54404bea497338a07ab85a2dfd76c9e4fd76c67d04ddcfac2b29d50.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 596, + 141, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 596, + 141, + 608 + ], + "spans": [ + { + "bbox": [ + 55, + 596, + 141, + 608 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 616, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 295, + 677 + ], + "type": "text", + "content": "Diffusion models have emerged as a powerful framework for generative tasks, particularly in Text-to-Image (T2I) applications. They operate by iteratively denoising data through a defined forward and reverse diffusion process, enabling the generation of high-fidelity images from noise." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "Forward process. In the forward diffusion process, Gaussian noise is incrementally added to an image over a series of timesteps " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "t = 1,2,\\ldots ,T" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": ", progressively transforming it" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 73, + 380, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 73, + 380, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 73, + 380, + 84 + ], + "type": "text", + "content": "into pure noise:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 349, + 87, + 553, + 101 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 87, + 553, + 101 + ], + "spans": [ + { + "bbox": [ + 349, + 87, + 553, + 101 + ], + "type": "interline_equation", + "content": "q \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {t - 1}\\right) = \\mathcal {N} \\left(\\mathbf {x} _ {t}; \\sqrt {1 - \\beta_ {t}} \\mathbf {x} _ {t - 1}, \\beta_ {t} \\mathbf {I}\\right), \\tag {1}", + "image_path": "ba0f50c708c5eca49b9b17c219e7b87eea7a271ef742293ca15aae6c1e988ae1.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 103, + 553, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 103, + 553, + 163 + ], + "spans": [ + { + "bbox": [ + 313, + 103, + 553, + 163 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 103, + 553, + 163 + ], + "type": "inline_equation", + "content": "\\beta_{t}" + }, + { + "bbox": [ + 313, + 103, + 553, + 163 + ], + "type": "text", + "content": " controls the noise variance at each timestep. We define " + }, + { + "bbox": [ + 313, + 103, + 553, + 163 + ], + "type": "inline_equation", + "content": "\\alpha_{t} = 1 - \\beta_{t}" + }, + { + "bbox": [ + 313, + 103, + 553, + 163 + ], + "type": "text", + "content": ", and the cumulative product " + }, + { + "bbox": [ + 313, + 103, + 553, + 163 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_{t}" + }, + { + "bbox": [ + 313, + 103, + 553, + 163 + ], + "type": "text", + "content": " is given by " + }, + { + "bbox": [ + 313, + 103, + 553, + 163 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_{t} = \\prod_{s=1}^{t} \\alpha_{s}" + }, + { + "bbox": [ + 313, + 103, + 553, + 163 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 313, + 103, + 553, + 163 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 313, + 103, + 553, + 163 + ], + "type": "text", + "content": " represents the remaining signal strength at timestep " + }, + { + "bbox": [ + 313, + 103, + 553, + 163 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 103, + 553, + 163 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 313, + 103, + 553, + 163 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_{t}" + }, + { + "bbox": [ + 313, + 103, + 553, + 163 + ], + "type": "text", + "content": " indicates the cumulative effect, controlling the overall noise level in the image." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 163, + 553, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 163, + 553, + 211 + ], + "spans": [ + { + "bbox": [ + 313, + 163, + 553, + 211 + ], + "type": "text", + "content": "Reverse process. The reverse diffusion process aims to reconstruct the original image by predicting and removing the added noise. This is achieved by training a denoising network " + }, + { + "bbox": [ + 313, + 163, + 553, + 211 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 313, + 163, + 553, + 211 + ], + "type": "text", + "content": " to minimize the denoising loss:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 349, + 213, + 553, + 228 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 213, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 349, + 213, + 553, + 228 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e c o n}} = \\mathbb {E} _ {\\mathbf {x} _ {0}, \\epsilon , t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (\\mathbf {x} _ {t}, t, \\mathcal {E} (\\mathbf {p})) \\| _ {2} ^ {2} \\right], \\tag {2}", + "image_path": "d82861ca3d1047233699975a301c6f8768737e5ab94b8ad2cea5ab28acbeccbe.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 229, + 553, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 229, + 553, + 265 + ], + "spans": [ + { + "bbox": [ + 313, + 229, + 553, + 265 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 229, + 553, + 265 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t = \\sqrt{\\bar{\\alpha}_t}\\mathbf{x}_0 + \\sqrt{1 - \\bar{\\alpha}_t\\epsilon}" + }, + { + "bbox": [ + 313, + 229, + 553, + 265 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 229, + 553, + 265 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 313, + 229, + 553, + 265 + ], + "type": "text", + "content": ". The term " + }, + { + "bbox": [ + 313, + 229, + 553, + 265 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 313, + 229, + 553, + 265 + ], + "type": "text", + "content": " represents the text prompt used for conditioning via text encoder " + }, + { + "bbox": [ + 313, + 229, + 553, + 265 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 313, + 229, + 553, + 265 + ], + "type": "text", + "content": " like CLIP [23]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 266, + 553, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 553, + 326 + ], + "type": "text", + "content": "The denoising loss function " + }, + { + "bbox": [ + 313, + 266, + 553, + 326 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{recon}}" + }, + { + "bbox": [ + 313, + 266, + 553, + 326 + ], + "type": "text", + "content": " is crucial for training the diffusion model, guiding " + }, + { + "bbox": [ + 313, + 266, + 553, + 326 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 313, + 266, + 553, + 326 + ], + "type": "text", + "content": " to accurately predict the added noise " + }, + { + "bbox": [ + 313, + 266, + 553, + 326 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 266, + 553, + 326 + ], + "type": "text", + "content": " at each timestep. By minimizing this loss, the model learns to generate clean images from noisy inputs, effectively capturing the underlying data distribution." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 326, + 555, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 326, + 555, + 374 + ], + "spans": [ + { + "bbox": [ + 313, + 326, + 555, + 374 + ], + "type": "text", + "content": "Relation to ICE. ICE leverages the inherent capabilities of diffusion-based T2I models to perform concept extraction and decomposition, enabling a more structured and interpretable approach to visual concept learning." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 386, + 370, + 398 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 386, + 370, + 398 + ], + "spans": [ + { + "bbox": [ + 314, + 386, + 370, + 398 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 407, + 554, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 554, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 554, + 515 + ], + "type": "text", + "content": "The proposed ICE (Intrinsic Concept Extraction) framework is designed to automatically and systematically discover essential concepts within images using a single T2I diffusion model. The framework operates through a two-stage architecture: Stage One: Automatic Concept Localization and Stage Two: Structured Concept Learning as illustrated in Figure 3. In this section, we provide an overview of each stage, detailing the underlying processes, and training objectives." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 523, + 544, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 523, + 544, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 523, + 544, + 536 + ], + "type": "text", + "content": "4.1. Stage One: Automatic Concept Localization" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 543, + 554, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 543, + 554, + 614 + ], + "spans": [ + { + "bbox": [ + 313, + 543, + 554, + 614 + ], + "type": "text", + "content": "This stage is designed to extract object-level concepts from an unlabelled input image automatically. This stage leverages off-the-shelf modules integrated within the T2I diffusion model, ensuring a training-free and seamless concept extraction process. The workflow of Automatic Concept Localization is illustrated in Figure 4." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 615, + 554, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 615, + 554, + 698 + ], + "spans": [ + { + "bbox": [ + 313, + 615, + 554, + 698 + ], + "type": "text", + "content": "Text-based concept retrieval. Given an unlabelled input image " + }, + { + "bbox": [ + 313, + 615, + 554, + 698 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 615, + 554, + 698 + ], + "type": "text", + "content": ", the first step involves retrieving relevant text-based concepts from the T2I model's vocabulary. This retrieval is facilitated by the Image-to-Text Retriever, based on T2I's CLIP encoders, denoted as " + }, + { + "bbox": [ + 313, + 615, + 554, + 698 + ], + "type": "inline_equation", + "content": "\\mathcal{T}(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 615, + 554, + 698 + ], + "type": "text", + "content": " to decompose dense embeddings into sparse, interpretable semantic concepts. Formally, the retrieval process can be expressed as:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 405, + 701, + 553, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 405, + 701, + 553, + 714 + ], + "type": "interline_equation", + "content": "\\left\\{c _ {i} \\right\\} = \\mathcal {T} (\\mathbf {x}), \\tag {3}", + "image_path": "1e0616592f6e3c017fe428e1d2d3379c397d47eb6eb2dfd9aba7cefe0d98e7da.jpg" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "23736" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 558, + 190 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 558, + 190 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 558, + 190 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 558, + 190 + ], + "type": "image", + "image_path": "acbe1f654db1084933b0f42a343c9e759456c5e05637fdfc0e00778580e8489c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 192, + 555, + 228 + ], + "lines": [ + { + "bbox": [ + 55, + 192, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 192, + 555, + 228 + ], + "type": "text", + "content": "Figure 3. Illustration of the proposed ICE (Intrinsic Concept Extraction) framework, which consists of two stages: (1) Automatic Concept Localization, where a diffusion model is employed to extract object-level concepts and their corresponding masks from an image without prior training, and (2) Structured Concept Learning, where these extracted information are further leveraged to uncover essential concepts." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 67, + 247, + 382, + 370 + ], + "blocks": [ + { + "bbox": [ + 61, + 232, + 245, + 245 + ], + "lines": [ + { + "bbox": [ + 61, + 232, + 245, + 245 + ], + "spans": [ + { + "bbox": [ + 61, + 232, + 245, + 245 + ], + "type": "text", + "content": "Stage One: Automatic Concept Localization" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 67, + 247, + 382, + 370 + ], + "lines": [ + { + "bbox": [ + 67, + 247, + 382, + 370 + ], + "spans": [ + { + "bbox": [ + 67, + 247, + 382, + 370 + ], + "type": "image", + "image_path": "3329133e769ffc614e867a77aa1bc7bb92b4e6ec7bc3eaec8dd4b8af0ae5c83a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 389, + 251, + 552, + 370 + ], + "blocks": [ + { + "bbox": [ + 410, + 236, + 536, + 248 + ], + "lines": [ + { + "bbox": [ + 410, + 236, + 536, + 248 + ], + "spans": [ + { + "bbox": [ + 410, + 236, + 536, + 248 + ], + "type": "text", + "content": "Stage One's use case diagram" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 389, + 251, + 552, + 370 + ], + "lines": [ + { + "bbox": [ + 389, + 251, + 552, + 370 + ], + "spans": [ + { + "bbox": [ + 389, + 251, + 552, + 370 + ], + "type": "image", + "image_path": "d3d65dfbd493db0d530da5faf97029653e6144fea818a3acdba50ededb4fa9d3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 375, + 555, + 409 + ], + "lines": [ + { + "bbox": [ + 54, + 375, + 555, + 409 + ], + "spans": [ + { + "bbox": [ + 54, + 375, + 555, + 409 + ], + "type": "text", + "content": "Figure 4. Stage One: Automatic Concept Localization. Starting with an unlabelled image " + }, + { + "bbox": [ + 54, + 375, + 555, + 409 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 54, + 375, + 555, + 409 + ], + "type": "text", + "content": ", the Image-to-Text concept extractor retrieves the top-1 text-based concept " + }, + { + "bbox": [ + 54, + 375, + 555, + 409 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 54, + 375, + 555, + 409 + ], + "type": "text", + "content": " using CLIP encoders. A zero-shot segmentor via T2I model generates the corresponding mask " + }, + { + "bbox": [ + 54, + 375, + 555, + 409 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_{i}" + }, + { + "bbox": [ + 54, + 375, + 555, + 409 + ], + "type": "text", + "content": ", and the image is updated by removing the masked region. This process iterates until no objects remain in the image." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 418, + 295, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 295, + 441 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 295, + 441 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 418, + 295, + 441 + ], + "type": "inline_equation", + "content": "\\{c_i\\}" + }, + { + "bbox": [ + 55, + 418, + 295, + 441 + ], + "type": "text", + "content": " represents the set of top-" + }, + { + "bbox": [ + 55, + 418, + 295, + 441 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 418, + 295, + 441 + ], + "type": "text", + "content": " text-based concepts extracted from the image " + }, + { + "bbox": [ + 55, + 418, + 295, + 441 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 418, + 295, + 441 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 441, + 296, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 296, + 501 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 296, + 501 + ], + "type": "text", + "content": "Mask generation via T2I zero-shot segmentor. Once the top-1 text concept " + }, + { + "bbox": [ + 55, + 441, + 296, + 501 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 55, + 441, + 296, + 501 + ], + "type": "text", + "content": " is identified, the framework employs the Segmentor, denoted as " + }, + { + "bbox": [ + 55, + 441, + 296, + 501 + ], + "type": "inline_equation", + "content": "S(\\mathbf{x},c_i)" + }, + { + "bbox": [ + 55, + 441, + 296, + 501 + ], + "type": "text", + "content": ", to obtain a corresponding segmentation mask " + }, + { + "bbox": [ + 55, + 441, + 296, + 501 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_i" + }, + { + "bbox": [ + 55, + 441, + 296, + 501 + ], + "type": "text", + "content": ". The mask generation process is formalized as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 143, + 502, + 295, + 515 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 502, + 295, + 515 + ], + "spans": [ + { + "bbox": [ + 143, + 502, + 295, + 515 + ], + "type": "interline_equation", + "content": "\\mathbf {m} _ {i} = \\mathcal {S} (\\mathbf {x}, c _ {i}), \\tag {4}", + "image_path": "cd9474547b738268bf7cb30d09d2774b3b44b8afbf3ab2a837e60089d0f6dcb7.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 521, + 295, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 593 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 521, + 295, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_i" + }, + { + "bbox": [ + 55, + 521, + 295, + 593 + ], + "type": "text", + "content": " delineates the region of " + }, + { + "bbox": [ + 55, + 521, + 295, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 521, + 295, + 593 + ], + "type": "text", + "content": " associated with the concept " + }, + { + "bbox": [ + 55, + 521, + 295, + 593 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 55, + 521, + 295, + 593 + ], + "type": "text", + "content": ". In our implementation, without introducing any external pretrained model, we utilize a training-free image segmentor that leverages the self-attention layers within the pre-trained Stable Diffusion model to produce high-quality segmentation masks without any additional training." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 594, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 677 + ], + "type": "text", + "content": "Iterative object extraction. With the mask " + }, + { + "bbox": [ + 55, + 594, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_i" + }, + { + "bbox": [ + 55, + 594, + 295, + 677 + ], + "type": "text", + "content": " obtained, the corresponding region is removed from the image to prepare for the extraction of subsequent concepts. The updated image " + }, + { + "bbox": [ + 55, + 594, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{x}'" + }, + { + "bbox": [ + 55, + 594, + 295, + 677 + ], + "type": "text", + "content": " is computed as: " + }, + { + "bbox": [ + 55, + 594, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{x}' = \\mathbf{x} \\odot (1 - \\mathbf{m}_i)" + }, + { + "bbox": [ + 55, + 594, + 295, + 677 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 594, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 55, + 594, + 295, + 677 + ], + "type": "text", + "content": " denotes element-wise multiplication. This updated image " + }, + { + "bbox": [ + 55, + 594, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{x}'" + }, + { + "bbox": [ + 55, + 594, + 295, + 677 + ], + "type": "text", + "content": " is then fed back into " + }, + { + "bbox": [ + 55, + 594, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 55, + 594, + 295, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 594, + 295, + 677 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 594, + 295, + 677 + ], + "type": "text", + "content": " to iteratively extract and mask remaining objects until no objects remain in the image." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "The integration of " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": " within the T2I diffusion model allows for a fully automated and training-free concept extraction process. This design ensures that the frame" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 418, + 555, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 555, + 501 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 555, + 501 + ], + "type": "text", + "content": "work does not rely on any external models beyond the pretrained T2I diffusion model, enhancing its scalability and applicability across diverse datasets. The Automatic Concept Extraction module outputs a set of text-based concepts " + }, + { + "bbox": [ + 313, + 418, + 555, + 501 + ], + "type": "inline_equation", + "content": "\\{c_i\\}" + }, + { + "bbox": [ + 313, + 418, + 555, + 501 + ], + "type": "text", + "content": " along with their masks " + }, + { + "bbox": [ + 313, + 418, + 555, + 501 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{m}_i\\}" + }, + { + "bbox": [ + 313, + 418, + 555, + 501 + ], + "type": "text", + "content": " for a given image " + }, + { + "bbox": [ + 313, + 418, + 555, + 501 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 418, + 555, + 501 + ], + "type": "text", + "content": ". The pseudocode for this stage is provided in the supplementary material." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 511, + 531, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 511, + 531, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 511, + 531, + 525 + ], + "type": "text", + "content": "4.2. Stage Two: Structured Concept Learning" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 531, + 555, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 531, + 555, + 603 + ], + "spans": [ + { + "bbox": [ + 313, + 531, + 555, + 603 + ], + "type": "text", + "content": "This stage focuses on decomposing the extracted object-level concepts into the underlying intrinsic concepts, such as colour and material. To ensure an accurate decomposition of the concepts, we divide this stage into two phases: (1) learning object-level concepts and (2) learning intrinsic concepts. Stage two's workflow is illustrated in Figure 5." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 604, + 554, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 604, + 554, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 604, + 554, + 651 + ], + "type": "text", + "content": "Phase one: learning object-level concepts. Here, the framework learns both concept-specific and instance-specific tokens for each extracted object-level concept. For each object-level concept " + }, + { + "bbox": [ + 313, + 604, + 554, + 651 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 313, + 604, + 554, + 651 + ], + "type": "text", + "content": ", two tokens are introduced:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 652, + 553, + 700 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 313, + 652, + 553, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 652, + 553, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 652, + 553, + 676 + ], + "type": "text", + "content": "- Concept-specific token (" + }, + { + "bbox": [ + 313, + 652, + 553, + 676 + ], + "type": "inline_equation", + "content": "c_i^{\\text{conspec}}" + }, + { + "bbox": [ + 313, + 652, + 553, + 676 + ], + "type": "text", + "content": "): Represents the general semantic category of the object." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 676, + 553, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 676, + 553, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 676, + 553, + 700 + ], + "type": "text", + "content": "- Instance-specific token " + }, + { + "bbox": [ + 313, + 676, + 553, + 700 + ], + "type": "inline_equation", + "content": "(c_{i}^{\\mathrm{inspec}})" + }, + { + "bbox": [ + 313, + 676, + 553, + 700 + ], + "type": "text", + "content": ": Captures attributes unique to the specific instance of the object." + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 325, + 700, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 700, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 325, + 700, + 553, + 714 + ], + "type": "text", + "content": "To differentiate these tokens, we initialize " + }, + { + "bbox": [ + 325, + 700, + 553, + 714 + ], + "type": "inline_equation", + "content": "c_{i}^{\\mathrm{conspec}}" + }, + { + "bbox": [ + 325, + 700, + 553, + 714 + ], + "type": "text", + "content": " em" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23737" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 71, + 553, + 224 + ], + "blocks": [ + { + "bbox": [ + 61, + 71, + 553, + 224 + ], + "lines": [ + { + "bbox": [ + 61, + 71, + 553, + 224 + ], + "spans": [ + { + "bbox": [ + 61, + 71, + 553, + 224 + ], + "type": "image", + "image_path": "fdae20ccaa81735089bb133ae4ddc554178ab9c618626302b86af0523d509be7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 230, + 555, + 277 + ], + "lines": [ + { + "bbox": [ + 54, + 230, + 555, + 277 + ], + "spans": [ + { + "bbox": [ + 54, + 230, + 555, + 277 + ], + "type": "text", + "content": "Figure 5. Stage Two: Structured Concept Learning. This stage is divided into two phases: (1) learning object-level concepts, where concept-specific " + }, + { + "bbox": [ + 54, + 230, + 555, + 277 + ], + "type": "inline_equation", + "content": "(c_i^{\\mathrm{conspec}})" + }, + { + "bbox": [ + 54, + 230, + 555, + 277 + ], + "type": "text", + "content": " and instance-specific " + }, + { + "bbox": [ + 54, + 230, + 555, + 277 + ], + "type": "inline_equation", + "content": "(c_i^{\\mathrm{inspec}})" + }, + { + "bbox": [ + 54, + 230, + 555, + 277 + ], + "type": "text", + "content": " tokens are learned using an object-level triplet loss " + }, + { + "bbox": [ + 54, + 230, + 555, + 277 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{triplet}}^{\\mathrm{obj}}" + }, + { + "bbox": [ + 54, + 230, + 555, + 277 + ], + "type": "text", + "content": ", and (2) learning intrinsic concepts, which decomposes object-level concepts into intrinsic attributes " + }, + { + "bbox": [ + 54, + 230, + 555, + 277 + ], + "type": "inline_equation", + "content": "(c_j^{\\mathrm{intrinsic}})" + }, + { + "bbox": [ + 54, + 230, + 555, + 277 + ], + "type": "text", + "content": " using an intrinsic triplet loss " + }, + { + "bbox": [ + 54, + 230, + 555, + 277 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{triplet}}^{\\mathrm{intrinsic}}" + }, + { + "bbox": [ + 54, + 230, + 555, + 277 + ], + "type": "text", + "content": ". This hierarchical approach ensures accurate separation of general semantic categories from specific and intrinsic attributes." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 285, + 296, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 285, + 296, + 347 + ], + "spans": [ + { + "bbox": [ + 54, + 285, + 296, + 347 + ], + "type": "text", + "content": "bedding with the extracted text-based concept " + }, + { + "bbox": [ + 54, + 285, + 296, + 347 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 54, + 285, + 296, + 347 + ], + "type": "text", + "content": " which is obtained in the first stage and employ a triplet loss mechanism, where the " + }, + { + "bbox": [ + 54, + 285, + 296, + 347 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 54, + 285, + 296, + 347 + ], + "type": "text", + "content": " serves as the anchor, " + }, + { + "bbox": [ + 54, + 285, + 296, + 347 + ], + "type": "inline_equation", + "content": "c_{i}^{\\mathrm{conspec}}" + }, + { + "bbox": [ + 54, + 285, + 296, + 347 + ], + "type": "text", + "content": " as the positive, and " + }, + { + "bbox": [ + 54, + 285, + 296, + 347 + ], + "type": "inline_equation", + "content": "c_{i}^{\\mathrm{inspec}}" + }, + { + "bbox": [ + 54, + 285, + 296, + 347 + ], + "type": "text", + "content": " as the negative. The triplet loss " + }, + { + "bbox": [ + 54, + 285, + 296, + 347 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{triplet}}^{\\mathrm{obj}}" + }, + { + "bbox": [ + 54, + 285, + 296, + 347 + ], + "type": "text", + "content": " for object-level concept " + }, + { + "bbox": [ + 54, + 285, + 296, + 347 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 54, + 285, + 296, + 347 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 348, + 294, + 391 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 348, + 294, + 391 + ], + "spans": [ + { + "bbox": [ + 69, + 348, + 294, + 391 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {t r i p l e t}} ^ {\\mathrm {o b j}} = \\max (0, \\| \\mathcal {E} (\\text {a n c h o r}) - \\mathcal {E} \\left(c _ {i} ^ {\\text {c o n s p e c}}\\right) \\| _ {2} ^ {2} \\\\ - \\left\\| \\mathcal {E} (\\text {a n c h o r}) - \\mathcal {E} \\left(c _ {i} ^ {\\text {i n s p e c}}\\right) \\right\\| _ {2} ^ {2} + \\gamma), \\tag {5} \\\\ \\end{array}", + "image_path": "ed4d5175ba07ebc467605fd34958c2c9b643be3d825e3c7dc8d891c8b375c1b2.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 392, + 296, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 392, + 296, + 441 + ], + "spans": [ + { + "bbox": [ + 55, + 392, + 296, + 441 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 392, + 296, + 441 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 392, + 296, + 441 + ], + "type": "text", + "content": " is the margin parameter, and " + }, + { + "bbox": [ + 55, + 392, + 296, + 441 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(\\cdot)" + }, + { + "bbox": [ + 55, + 392, + 296, + 441 + ], + "type": "text", + "content": " denotes the T2I's CLIP text encoder. This loss encourages " + }, + { + "bbox": [ + 55, + 392, + 296, + 441 + ], + "type": "inline_equation", + "content": "c_{i}^{\\mathrm{conspec}}" + }, + { + "bbox": [ + 55, + 392, + 296, + 441 + ], + "type": "text", + "content": " to be closer to anchor " + }, + { + "bbox": [ + 55, + 392, + 296, + 441 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 55, + 392, + 296, + 441 + ], + "type": "text", + "content": " than " + }, + { + "bbox": [ + 55, + 392, + 296, + 441 + ], + "type": "inline_equation", + "content": "c_{i}^{\\mathrm{inspec}}" + }, + { + "bbox": [ + 55, + 392, + 296, + 441 + ], + "type": "text", + "content": ", effectively distinguishing between general and instance-specific attributes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 449, + 296, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 296, + 544 + ], + "type": "text", + "content": "Phase two: learning intrinsic concepts. The second phase delves into decomposing the learned object-level concepts into intrinsic attributes. For each intrinsic concept (e.g., colour, material), we introduce corresponding intrinsic tokens " + }, + { + "bbox": [ + 55, + 449, + 296, + 544 + ], + "type": "inline_equation", + "content": "c_{j}^{\\mathrm{intrinsic}}" + }, + { + "bbox": [ + 55, + 449, + 296, + 544 + ], + "type": "text", + "content": " associated with object-level concepts " + }, + { + "bbox": [ + 55, + 449, + 296, + 544 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 55, + 449, + 296, + 544 + ], + "type": "text", + "content": ". The intrinsic concept tokens are anchored using combined representations of " + }, + { + "bbox": [ + 55, + 449, + 296, + 544 + ], + "type": "inline_equation", + "content": "c_{i}^{\\mathrm{conspec}}" + }, + { + "bbox": [ + 55, + 449, + 296, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 449, + 296, + 544 + ], + "type": "inline_equation", + "content": "c_{i}^{\\mathrm{inspec}}" + }, + { + "bbox": [ + 55, + 449, + 296, + 544 + ], + "type": "text", + "content": ". Specifically, for an intrinsic concept " + }, + { + "bbox": [ + 55, + 449, + 296, + 544 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 449, + 296, + 544 + ], + "type": "text", + "content": ", the anchor text " + }, + { + "bbox": [ + 55, + 449, + 296, + 544 + ], + "type": "inline_equation", + "content": "\\mathsf{anchor}_{j}" + }, + { + "bbox": [ + 55, + 449, + 296, + 544 + ], + "type": "text", + "content": " is formulated as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 93, + 567, + 259, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 567, + 259, + 580 + ], + "spans": [ + { + "bbox": [ + 93, + 567, + 259, + 580 + ], + "type": "text", + "content": "anchorj:\"a intrinsicconcept" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 603, + 295, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 603, + 295, + 638 + ], + "spans": [ + { + "bbox": [ + 55, + 603, + 295, + 638 + ], + "type": "text", + "content": "where intrinsic denotes the specific attribute (e.g., color). The triplet loss " + }, + { + "bbox": [ + 55, + 603, + 295, + 638 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{triplet}}^{\\mathrm{intrinsic}}" + }, + { + "bbox": [ + 55, + 603, + 295, + 638 + ], + "type": "text", + "content": " for intrinsic concept " + }, + { + "bbox": [ + 55, + 603, + 295, + 638 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 603, + 295, + 638 + ], + "type": "text", + "content": " is then defined as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 639, + 295, + 675 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 639, + 295, + 675 + ], + "spans": [ + { + "bbox": [ + 62, + 639, + 295, + 675 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {t r i p l e t}} ^ {\\text {i n t r i n s i c}} = \\max (0, \\| \\mathcal {E} (\\text {a n c h o r} _ {j}) - \\mathcal {E} (c _ {j} ^ {\\text {i n t r i n s i c}}) \\| _ {2} ^ {2} \\\\ \\left. - \\left\\| \\mathcal {E} (\\operatorname {a n c h o r} _ {j}) - \\mathcal {E} \\left(c _ {k} ^ {\\text {i n t r i n s i c}}\\right) \\right\\| _ {2} ^ {2} + \\gamma\\right), \\tag {6} \\\\ \\end{array}", + "image_path": "a905801e192e2b9daeab6e0edd9acecc3e090bf60065cce4e4b32ecd9fe5611c.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "c_k^{\\mathrm{intrinsic}}" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " represents intrinsic tokens not associated with intrinsic concept " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": ", ensuring that each intrinsic token is uniquely mapped to its corresponding attribute." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 285, + 430, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 430, + 298 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 430, + 298 + ], + "type": "text", + "content": "4.3. Concept refinement" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 304, + 555, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 555, + 423 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 555, + 423 + ], + "type": "text", + "content": "Optimizing intrinsic concepts solely through token embeddings is challenging, often leading to misaligned and inconsistent representations of specific visual attributes such as color and texture. To enhance the accuracy and consistency of these intrinsic concepts, after completing Phase two, we perform concept refinement by fine-tuning both the U-Net denoising network and the text encoder " + }, + { + "bbox": [ + 313, + 304, + 555, + 423 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 313, + 304, + 555, + 423 + ], + "type": "text", + "content": " for a limited number of training steps. This fine-tuning ensures that the intrinsic concepts are precisely aligned with their corresponding visual attributes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 434, + 458, + 447 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 458, + 447 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 458, + 447 + ], + "type": "text", + "content": "4.4. Overall training objective" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 453, + 554, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 554, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 554, + 502 + ], + "type": "text", + "content": "The overall training objective of the ICE framework integrates multiple loss components to facilitate both reconstruction quality and concept differentiation. The total loss " + }, + { + "bbox": [ + 313, + 453, + 554, + 502 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{total}}" + }, + { + "bbox": [ + 313, + 453, + 554, + 502 + ], + "type": "text", + "content": " is given by:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 353, + 516, + 553, + 529 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 516, + 553, + 529 + ], + "spans": [ + { + "bbox": [ + 353, + 516, + 553, + 529 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {r e c o n}} + \\lambda_ {\\text {a t t}} \\mathcal {L} _ {\\text {a t t}} + \\lambda_ {\\text {t r i p l e t}} \\mathcal {L} _ {\\text {t r i p l e t}}. \\tag {7}", + "image_path": "645772988b55d865075afc2f1128f52bd281cb9b0c2cfca31fbcbfd75eb5e3bf.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 537, + 555, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 537, + 555, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 537, + 555, + 574 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{recon}}" + }, + { + "bbox": [ + 313, + 537, + 555, + 574 + ], + "type": "text", + "content": ", as defined in Eq.(2), is the reconstruction loss between the predicted noise and the actual noise with text prompt1 conditioning " + }, + { + "bbox": [ + 313, + 537, + 555, + 574 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 313, + 537, + 555, + 574 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 341, + 585, + 526, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 585, + 526, + 601 + ], + "spans": [ + { + "bbox": [ + 341, + 585, + 526, + 601 + ], + "type": "text", + "content": "Phase one, p: \" a photo of " + }, + { + "bbox": [ + 341, + 585, + 526, + 601 + ], + "type": "inline_equation", + "content": "c_{i}^{\\mathrm{inspec}}" + }, + { + "bbox": [ + 341, + 585, + 526, + 601 + ], + "type": "text", + "content": " & " + }, + { + "bbox": [ + 341, + 585, + 526, + 601 + ], + "type": "inline_equation", + "content": "c_{i}^{\\mathrm{conspec}}" + }, + { + "bbox": [ + 341, + 585, + 526, + 601 + ], + "type": "text", + "content": "\"" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 338, + 614, + 531, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 614, + 531, + 629 + ], + "spans": [ + { + "bbox": [ + 338, + 614, + 531, + 629 + ], + "type": "text", + "content": "Phase two, p: “ a photo of " + }, + { + "bbox": [ + 338, + 614, + 531, + 629 + ], + "type": "inline_equation", + "content": "\\mathbb{C}^{\\mathrm{intrinsic}}" + }, + { + "bbox": [ + 338, + 614, + 531, + 629 + ], + "type": "text", + "content": " & " + }, + { + "bbox": [ + 338, + 614, + 531, + 629 + ], + "type": "inline_equation", + "content": "c_{i}^{\\mathrm{conspec}}" + }, + { + "bbox": [ + 338, + 614, + 531, + 629 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 640, + 554, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 640, + 554, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 640, + 554, + 677 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 640, + 554, + 677 + ], + "type": "inline_equation", + "content": "\\mathbb{C}^{\\mathrm{intrinsic}}" + }, + { + "bbox": [ + 313, + 640, + 554, + 677 + ], + "type": "text", + "content": " is a set of " + }, + { + "bbox": [ + 313, + 640, + 554, + 677 + ], + "type": "inline_equation", + "content": "c_{j}^{\\mathrm{intrinsic}}" + }, + { + "bbox": [ + 313, + 640, + 554, + 677 + ], + "type": "text", + "content": " tokens. " + }, + { + "bbox": [ + 313, + 640, + 554, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{att}}" + }, + { + "bbox": [ + 313, + 640, + 554, + 677 + ], + "type": "text", + "content": " is the attention mask loss, which employs Wasserstein loss, following [9], to align the attention regions with the segmentation masks:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 393, + 679, + 553, + 693 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 679, + 553, + 693 + ], + "spans": [ + { + "bbox": [ + 393, + 679, + 553, + 693 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {a t t}} = \\mathcal {W} \\left(\\mathbf {A} _ {i}, \\mathbf {m} _ {i}\\right), \\tag {8}", + "image_path": "19ac7f27d17abdd6d5def746eb415405ed3d70c6ae6c1bf58b97c5b9dcfe20f2.jpg" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 324, + 702, + 546, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 702, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 324, + 702, + 546, + 713 + ], + "type": "text", + "content": "1The list of prompt templates is provided in supplementary material." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23738" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": " denotes the Wasserstein distance, " + }, + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_i" + }, + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": " is the attention map corresponding to concept " + }, + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_i" + }, + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": " is the segmentation mask obtained in the first stage. " + }, + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{triplet}}^{\\mathrm{obj}}" + }, + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{triplet}}^{\\mathrm{intrinsic}}" + }, + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": " are the object-level and intrinsic triplet losses for different phases:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 70, + 136, + 295, + 162 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 136, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 70, + 136, + 295, + 162 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t r i p l e t}} ^ {\\mathrm {o b j}} = \\sum_ {i} \\mathcal {L} _ {\\text {t r i p l e t}, i} ^ {\\mathrm {o b j}}, \\quad \\mathcal {L} _ {\\text {t r i p l e t}} ^ {\\text {i n t r i n s i c}} = \\sum_ {j} \\mathcal {L} _ {\\text {t r i p l e t}, j} ^ {\\text {i n t r i n s i c}}. \\tag {9}", + "image_path": "bac9111ce61f42e911709a243e61411044342782245d2091e86161ba3e5e4f1c.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 164, + 296, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 164, + 296, + 213 + ], + "spans": [ + { + "bbox": [ + 55, + 164, + 296, + 213 + ], + "type": "text", + "content": "The weighting factors " + }, + { + "bbox": [ + 55, + 164, + 296, + 213 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{att}}" + }, + { + "bbox": [ + 55, + 164, + 296, + 213 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 164, + 296, + 213 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{triplet}}" + }, + { + "bbox": [ + 55, + 164, + 296, + 213 + ], + "type": "text", + "content": " balance the contributions of the attention and triplet losses relative to the reconstruction loss in each phase, ensuring stable optimization across all objectives." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 225, + 136, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 225, + 136, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 225, + 136, + 239 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 247, + 296, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 247, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 247, + 296, + 521 + ], + "type": "text", + "content": "Implementation details. We implement our ICE framework using Stable Diffusion v2.1 for both stages. In the first stage, for the Image-to-Text Retriever module " + }, + { + "bbox": [ + 56, + 247, + 296, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 56, + 247, + 296, + 521 + ], + "type": "text", + "content": ", we utilize the CLIP-based SpLiCE framework [4], while the T2I zero-shot Segmentor " + }, + { + "bbox": [ + 56, + 247, + 296, + 521 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 56, + 247, + 296, + 521 + ], + "type": "text", + "content": " employs DiffSeg [30]. We adapt the input and output interfaces of these training-free modules to accommodate our automatic concept localization process. Specifically, modifications were made to ensure seamless integration within the framework, allowing for efficient and accurate extraction of text-based concepts and their corresponding masks. The training process in the second stage is divided into two phases: Phase one focuses on learning object-level concepts and is conducted over 400 training steps, followed by Phase two, which targets intrinsic concepts and also comprises 400 training steps. Subsequently, concept refinement is performed with another 300 steps to enhance the quality and coherence of the learned concepts. We assign the weighting factors " + }, + { + "bbox": [ + 56, + 247, + 296, + 521 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{att}} = 1\\times 10^{-5}" + }, + { + "bbox": [ + 56, + 247, + 296, + 521 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 247, + 296, + 521 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{triplet}} = 1" + }, + { + "bbox": [ + 56, + 247, + 296, + 521 + ], + "type": "text", + "content": ". For the triplet loss margin " + }, + { + "bbox": [ + 56, + 247, + 296, + 521 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 56, + 247, + 296, + 521 + ], + "type": "text", + "content": " value, please refer to the supplementary material. Finally, we compute a prior preservation loss [26] to promote diverse image generation while retaining class-specific knowledge. All experiments are conducted on a single NVIDIA RTX 3090 GPU." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "text", + "content": "Baseline & evaluation metrics. To evaluate the effectiveness of ICE, we compare our framework with the state-of-the-art method, ConceptExpress [9], on the Unsupervised Concept Extraction (UCE) benchmarks. Specifically, we utilize " + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "inline_equation", + "content": "D1" + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "text", + "content": " dataset, which is based on the Unsplash datasets, providing a diverse and extensive collection of unlabeled images. Moreover, the UCE benchmarks assess models based on two primary metrics: (1) Concept Similarity, which includes Identity Similarity " + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "inline_equation", + "content": "(\\mathrm{SIM}^I)" + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "text", + "content": " measuring how accurately individual concepts are recreated, and Compositional Similarity " + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "inline_equation", + "content": "(\\mathrm{SIM}^C)" + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "text", + "content": " evaluating the overall coherence of the generated image based on the extracted concepts. (2) Classification Accuracy assesses the disentangle-ment and classification performance of the extracted concepts by evaluating how well the concepts can be used to classify images correctly. These metrics provide a compre" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": "hensive evaluation of both the fidelity and utility of the extracted concepts, enabling a robust quantitative comparison between ICE and ConceptExpress." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 118, + 430, + 131 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 118, + 430, + 131 + ], + "spans": [ + { + "bbox": [ + 313, + 118, + 430, + 131 + ], + "type": "text", + "content": "5.1. Quantitative results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 137, + 554, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 137, + 554, + 186 + ], + "spans": [ + { + "bbox": [ + 313, + 137, + 554, + 186 + ], + "type": "text", + "content": "As shown in Table 2 & 3, ICE achieves higher scores in all metrics compared to previous methods, demonstrating its superior ability to recreate and compose visual concepts from unlabelled images accurately." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 320, + 219, + 550, + 270 + ], + "blocks": [ + { + "bbox": [ + 313, + 194, + 554, + 217 + ], + "lines": [ + { + "bbox": [ + 313, + 194, + 554, + 217 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 554, + 217 + ], + "type": "text", + "content": "Table 2. Performance of ICE and relevant works on Unsupervised Concept Extraction (UCE) benchmarks using CLIP [23] encoder." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 320, + 219, + 550, + 270 + ], + "lines": [ + { + "bbox": [ + 320, + 219, + 550, + 270 + ], + "spans": [ + { + "bbox": [ + 320, + 219, + 550, + 270 + ], + "type": "table", + "html": "
Method\\(SIM^I\\)(%)\\(SIM^C\\)(%)\\(ACC^1\\)(%)\\(ACC^3\\)(%)
Break-A-Scene [1]0.6270.7730.1740.282
ConceptExpress[9]0.6890.7840.2630.385
ICE (Ours)0.7380.8220.3250.518
", + "image_path": "4d235afaed4dd2af41858db8f4523ca88faac1f25cf9a6c683f7c3990256e382.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 320, + 316, + 549, + 366 + ], + "blocks": [ + { + "bbox": [ + 313, + 292, + 554, + 314 + ], + "lines": [ + { + "bbox": [ + 313, + 292, + 554, + 314 + ], + "spans": [ + { + "bbox": [ + 313, + 292, + 554, + 314 + ], + "type": "text", + "content": "Table 3. Performance of ICE and relevant works on Unsupervised Concept Extraction (UCE) benchmarks using DINO [6] encoder." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 320, + 316, + 549, + 366 + ], + "lines": [ + { + "bbox": [ + 320, + 316, + 549, + 366 + ], + "spans": [ + { + "bbox": [ + 320, + 316, + 549, + 366 + ], + "type": "table", + "html": "
Method\\( SIM^I \\) (%)\\( SIM^C \\) (%)\\( ACC^1 \\) (%)\\( ACC^3 \\) (%)
Break-A-Scene [1]0.2540.5100.2020.315
ConceptExpress [9]0.3190.5680.3240.470
ICE (Ours)0.6770.7550.4760.638
", + "image_path": "779be91fad999f3df2698f3d56ace517fce84deeb4f53bd6787dbcf64a9d49a3.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 379, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 379, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 379, + 555, + 713 + ], + "type": "text", + "content": "This enhanced performance can be attributed to two key factors: (1) Enhanced Concept Extraction: Our framework effectively extracts text-based concepts and their corresponding masks, resulting in superior masking accuracy as illustrated in Figure 7. Unlike other methods that may only mask object-level regions, ICE also retrieves the most similar text-based descriptors. This dual extraction provides additional guidance and robust initialization for the model, ensuring that both the spatial and semantic aspects of each concept are accurately captured. (2) Structured Concept Learning Approach: ICE's two-stage framework adopts a structured approach to concept learning, systematically decomposing object-level concepts into intrinsic attributes. This structured methodology facilitates a more granular and interpretable breakdown of visual elements, leading to the discovery of richer and more meaningful concepts. Specifically, ICE achieves CLIP Identity Similarity " + }, + { + "bbox": [ + 313, + 379, + 555, + 713 + ], + "type": "inline_equation", + "content": "(\\mathrm{SIM}^I)" + }, + { + "bbox": [ + 313, + 379, + 555, + 713 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 379, + 555, + 713 + ], + "type": "inline_equation", + "content": "73.8\\%" + }, + { + "bbox": [ + 313, + 379, + 555, + 713 + ], + "type": "text", + "content": ", compared to ConceptExpress's " + }, + { + "bbox": [ + 313, + 379, + 555, + 713 + ], + "type": "inline_equation", + "content": "68.9\\%" + }, + { + "bbox": [ + 313, + 379, + 555, + 713 + ], + "type": "text", + "content": ", indicating a substantial improvement in accurately identifying object-level concepts. Additionally, ICE attains a Compositional Similarity " + }, + { + "bbox": [ + 313, + 379, + 555, + 713 + ], + "type": "inline_equation", + "content": "(\\mathrm{SIM}^C)" + }, + { + "bbox": [ + 313, + 379, + 555, + 713 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 379, + 555, + 713 + ], + "type": "inline_equation", + "content": "82.2\\%" + }, + { + "bbox": [ + 313, + 379, + 555, + 713 + ], + "type": "text", + "content": ", surpassing ConceptExpress's " + }, + { + "bbox": [ + 313, + 379, + 555, + 713 + ], + "type": "inline_equation", + "content": "78.4\\%" + }, + { + "bbox": [ + 313, + 379, + 555, + 713 + ], + "type": "text", + "content": ", which underscores our framework's enhanced capability in understanding and composing complex visual attributes. These factors collectively contribute to the superior performance of ICE, highlighting the effectiveness of our framework in systematically uncovering and refining both object-level and intrinsic visual concepts from unlabeled images." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23739" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 102, + 156, + 199 + ], + "blocks": [ + { + "bbox": [ + 82, + 81, + 134, + 92 + ], + "lines": [ + { + "bbox": [ + 82, + 81, + 134, + 92 + ], + "spans": [ + { + "bbox": [ + 82, + 81, + 134, + 92 + ], + "type": "text", + "content": "Input image " + }, + { + "bbox": [ + 82, + 81, + 134, + 92 + ], + "type": "inline_equation", + "content": "x" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 60, + 102, + 156, + 199 + ], + "lines": [ + { + "bbox": [ + 60, + 102, + 156, + 199 + ], + "spans": [ + { + "bbox": [ + 60, + 102, + 156, + 199 + ], + "type": "image", + "image_path": "d7508154afc88686741b650f3dc41fd4917c2873e85d0785055d78a68f840d5b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 175, + 76, + 212, + 99 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 76, + 212, + 99 + ], + "spans": [ + { + "bbox": [ + 175, + 76, + 212, + 99 + ], + "type": "text", + "content": "Extracted mask " + }, + { + "bbox": [ + 175, + 76, + 212, + 99 + ], + "type": "inline_equation", + "content": "m_{i}" + } + ] + } + ], + "index": 2, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 170, + 102, + 218, + 148 + ], + "blocks": [ + { + "bbox": [ + 170, + 102, + 218, + 148 + ], + "lines": [ + { + "bbox": [ + 170, + 102, + 218, + 148 + ], + "spans": [ + { + "bbox": [ + 170, + 102, + 218, + 148 + ], + "type": "image", + "image_path": "bcfbac42c7da04d2a791750997d3cc8355c35fffe87992dcefecde5e2fa75585.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 231, + 120, + 257, + 129 + ], + "lines": [ + { + "bbox": [ + 231, + 120, + 257, + 129 + ], + "spans": [ + { + "bbox": [ + 231, + 120, + 257, + 129 + ], + "type": "text", + "content": "\"Bust\"" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 173, + 76, + 212, + 99 + ], + "lines": [ + { + "bbox": [ + 173, + 76, + 212, + 99 + ], + "spans": [ + { + "bbox": [ + 173, + 76, + 212, + 99 + ], + "type": "text", + "content": "Extracted mask " + }, + { + "bbox": [ + 173, + 76, + 212, + 99 + ], + "type": "inline_equation", + "content": "m_{i}" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 227, + 76, + 260, + 99 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 76, + 260, + 99 + ], + "spans": [ + { + "bbox": [ + 227, + 76, + 260, + 99 + ], + "type": "text", + "content": "Retrived text " + }, + { + "bbox": [ + 227, + 76, + 260, + 99 + ], + "type": "inline_equation", + "content": "c_{i}" + } + ] + } + ], + "index": 4, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 170, + 152, + 217, + 198 + ], + "blocks": [ + { + "bbox": [ + 170, + 152, + 217, + 198 + ], + "lines": [ + { + "bbox": [ + 170, + 152, + 217, + 198 + ], + "spans": [ + { + "bbox": [ + 170, + 152, + 217, + 198 + ], + "type": "image", + "image_path": "f6e07ac92b04635b532c70939a6b6d65213a3a6a10e02e6f3610d4d5b365b47e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 229, + 170, + 259, + 180 + ], + "lines": [ + { + "bbox": [ + 229, + 170, + 259, + 180 + ], + "spans": [ + { + "bbox": [ + 229, + 170, + 259, + 180 + ], + "type": "text", + "content": "\"Spout\"" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 170, + 205, + 217, + 251 + ], + "blocks": [ + { + "bbox": [ + 170, + 205, + 217, + 251 + ], + "lines": [ + { + "bbox": [ + 170, + 205, + 217, + 251 + ], + "spans": [ + { + "bbox": [ + 170, + 205, + 217, + 251 + ], + "type": "image", + "image_path": "44efb1ab860c9c03cc39f6bef074fb38e6150c951a3d7714e04d64caef0174ba.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 231, + 224, + 257, + 233 + ], + "lines": [ + { + "bbox": [ + 231, + 224, + 257, + 233 + ], + "spans": [ + { + "bbox": [ + 231, + 224, + 257, + 233 + ], + "type": "text", + "content": "\"Bust\"" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 170, + 255, + 217, + 302 + ], + "blocks": [ + { + "bbox": [ + 170, + 255, + 217, + 302 + ], + "lines": [ + { + "bbox": [ + 170, + 255, + 217, + 302 + ], + "spans": [ + { + "bbox": [ + 170, + 255, + 217, + 302 + ], + "type": "image", + "image_path": "a6bd3c7a3d32481cc26c3da3485315c0410d89578b08e7671d9251346f5fdeb9.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 230, + 274, + 258, + 284 + ], + "lines": [ + { + "bbox": [ + 230, + 274, + 258, + 284 + ], + "spans": [ + { + "bbox": [ + 230, + 274, + 258, + 284 + ], + "type": "text", + "content": "\"Cone\"" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 170, + 311, + 217, + 357 + ], + "blocks": [ + { + "bbox": [ + 170, + 311, + 217, + 357 + ], + "lines": [ + { + "bbox": [ + 170, + 311, + 217, + 357 + ], + "spans": [ + { + "bbox": [ + 170, + 311, + 217, + 357 + ], + "type": "image", + "image_path": "36a554f0b73c52cb401821572c61d2e561654a0848a352feabd644cf0c594ec7.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 221, + 329, + 267, + 338 + ], + "lines": [ + { + "bbox": [ + 221, + 329, + 267, + 338 + ], + "spans": [ + { + "bbox": [ + 221, + 329, + 267, + 338 + ], + "type": "text", + "content": "\"Ornament\"" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 170, + 360, + 217, + 407 + ], + "blocks": [ + { + "bbox": [ + 170, + 360, + 217, + 407 + ], + "lines": [ + { + "bbox": [ + 170, + 360, + 217, + 407 + ], + "spans": [ + { + "bbox": [ + 170, + 360, + 217, + 407 + ], + "type": "image", + "image_path": "7cbf3f514973643e3d2650bf1c801e7969dbad7890632b9aeaa29dc9fb4c8428.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 233, + 379, + 256, + 388 + ], + "lines": [ + { + "bbox": [ + 233, + 379, + 256, + 388 + ], + "spans": [ + { + "bbox": [ + 233, + 379, + 256, + 388 + ], + "type": "text", + "content": "\"Rat\"" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 170, + 415, + 217, + 462 + ], + "blocks": [ + { + "bbox": [ + 170, + 415, + 217, + 462 + ], + "lines": [ + { + "bbox": [ + 170, + 415, + 217, + 462 + ], + "spans": [ + { + "bbox": [ + 170, + 415, + 217, + 462 + ], + "type": "image", + "image_path": "e0a25c1868956268280a0560affd7afd21c604d99cc50c9d034b089ae2b1acb5.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 225, + 434, + 263, + 444 + ], + "lines": [ + { + "bbox": [ + 225, + 434, + 263, + 444 + ], + "spans": [ + { + "bbox": [ + 225, + 434, + 263, + 444 + ], + "type": "text", + "content": "\"Wooden\"" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 170, + 464, + 217, + 511 + ], + "blocks": [ + { + "bbox": [ + 170, + 464, + 217, + 511 + ], + "lines": [ + { + "bbox": [ + 170, + 464, + 217, + 511 + ], + "spans": [ + { + "bbox": [ + 170, + 464, + 217, + 511 + ], + "type": "image", + "image_path": "bd82722cf765e4b2415cfdffdfcdf9250fc13de6a696993cefb7ecdc12ad0cf4.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 225, + 483, + 263, + 492 + ], + "lines": [ + { + "bbox": [ + 225, + 483, + 263, + 492 + ], + "spans": [ + { + "bbox": [ + 225, + 483, + 263, + 492 + ], + "type": "text", + "content": "\"Wooden\"" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 170, + 519, + 217, + 566 + ], + "blocks": [ + { + "bbox": [ + 170, + 519, + 217, + 566 + ], + "lines": [ + { + "bbox": [ + 170, + 519, + 217, + 566 + ], + "spans": [ + { + "bbox": [ + 170, + 519, + 217, + 566 + ], + "type": "image", + "image_path": "49572d6efb2343757f85df62054c1cf6e7577a07c5524fd4505a6e7421c1eb26.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 230, + 538, + 258, + 548 + ], + "lines": [ + { + "bbox": [ + 230, + 538, + 258, + 548 + ], + "spans": [ + { + "bbox": [ + 230, + 538, + 258, + 548 + ], + "type": "text", + "content": "\"Pines\"" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 170, + 568, + 217, + 614 + ], + "blocks": [ + { + "bbox": [ + 170, + 568, + 217, + 614 + ], + "lines": [ + { + "bbox": [ + 170, + 568, + 217, + 614 + ], + "spans": [ + { + "bbox": [ + 170, + 568, + 217, + 614 + ], + "type": "image", + "image_path": "3d9085d5686db6dc4f8bec30c562a1f9c8982e0cd20461f0c480c71842ba444d.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 225, + 586, + 263, + 597 + ], + "lines": [ + { + "bbox": [ + 225, + 586, + 263, + 597 + ], + "spans": [ + { + "bbox": [ + 225, + 586, + 263, + 597 + ], + "type": "text", + "content": "\"Chapter\"" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 171, + 586, + 217, + 596 + ], + "lines": [ + { + "bbox": [ + 171, + 586, + 217, + 596 + ], + "spans": [ + { + "bbox": [ + 171, + 586, + 217, + 596 + ], + "type": "inline_equation", + "content": "{13}/{14}" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 59, + 520, + 156, + 614 + ], + "blocks": [ + { + "bbox": [ + 59, + 520, + 156, + 614 + ], + "lines": [ + { + "bbox": [ + 59, + 520, + 156, + 614 + ], + "spans": [ + { + "bbox": [ + 59, + 520, + 156, + 614 + ], + "type": "image", + "image_path": "c68b3da89367bb3cce1f3412d22b7a3ac0df4dde391a5170c9c5bd1427a9619b.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 621, + 553, + 655 + ], + "lines": [ + { + "bbox": [ + 55, + 621, + 553, + 655 + ], + "spans": [ + { + "bbox": [ + 55, + 621, + 553, + 655 + ], + "type": "text", + "content": "Figure 6. Qualitative results of the ICE framework demonstrating its systematic concept discovery process. Column 1: Input images. Column 2: Extracted text-based concepts and their corresponding masks obtained from Stage One: Automatic Concept Localization. Columns 3 & 4: Generated images of learned object-level and intrinsic concepts derived from Stage Two: Structured Concept Learning." + } + ] + } + ], + "index": 75, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "bbox": [ + 170, + 632, + 217, + 642 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 632, + 217, + 642 + ], + "spans": [ + { + "bbox": [ + 170, + 632, + 217, + 642 + ], + "type": "text", + "content": "ed concepts a" + } + ] + } + ], + "index": 26, + "type": "text" + }, + { + "bbox": [ + 170, + 643, + 295, + 654 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 643, + 295, + 654 + ], + "spans": [ + { + "bbox": [ + 170, + 643, + 295, + 654 + ], + "type": "text", + "content": "ages of learned object-level and int" + } + ] + } + ], + "index": 27, + "type": "text" + }, + { + "bbox": [ + 227, + 76, + 260, + 99 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 76, + 260, + 99 + ], + "spans": [ + { + "bbox": [ + 227, + 76, + 260, + 99 + ], + "type": "text", + "content": "Retrived text " + }, + { + "bbox": [ + 227, + 76, + 260, + 99 + ], + "type": "inline_equation", + "content": "c_{i}" + } + ] + } + ], + "index": 29, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 299, + 102, + 345, + 148 + ], + "blocks": [ + { + "bbox": [ + 299, + 76, + 345, + 98 + ], + "lines": [ + { + "bbox": [ + 299, + 76, + 345, + 98 + ], + "spans": [ + { + "bbox": [ + 299, + 76, + 345, + 98 + ], + "type": "text", + "content": "Object-level concepts" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 299, + 102, + 345, + 148 + ], + "lines": [ + { + "bbox": [ + 299, + 102, + 345, + 148 + ], + "spans": [ + { + "bbox": [ + 299, + 102, + 345, + 148 + ], + "type": "image", + "image_path": "ea14d6f07c734134a2480a9b1a421271005bffc9a3ef60491ffc3fc67b25898b.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 298, + 152, + 344, + 198 + ], + "blocks": [ + { + "bbox": [ + 298, + 152, + 344, + 198 + ], + "lines": [ + { + "bbox": [ + 298, + 152, + 344, + 198 + ], + "spans": [ + { + "bbox": [ + 298, + 152, + 344, + 198 + ], + "type": "image", + "image_path": "71beebb5974c120a3cd1cbf916e712f2128075e21fe3434b03225989ab7f949b.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 298, + 205, + 344, + 251 + ], + "blocks": [ + { + "bbox": [ + 298, + 205, + 344, + 251 + ], + "lines": [ + { + "bbox": [ + 298, + 205, + 344, + 251 + ], + "spans": [ + { + "bbox": [ + 298, + 205, + 344, + 251 + ], + "type": "image", + "image_path": "d18b86cbf0ac814f23b776aca8e838f72e1ba53d3f1910ab55379ea6aab93ff8.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 298, + 256, + 344, + 303 + ], + "blocks": [ + { + "bbox": [ + 298, + 256, + 344, + 303 + ], + "lines": [ + { + "bbox": [ + 298, + 256, + 344, + 303 + ], + "spans": [ + { + "bbox": [ + 298, + 256, + 344, + 303 + ], + "type": "image", + "image_path": "39b43670b729c78aba84611f4f4d316fdddf1ecf4780286c0397da1e51053af2.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 298, + 309, + 344, + 357 + ], + "blocks": [ + { + "bbox": [ + 298, + 309, + 344, + 357 + ], + "lines": [ + { + "bbox": [ + 298, + 309, + 344, + 357 + ], + "spans": [ + { + "bbox": [ + 298, + 309, + 344, + 357 + ], + "type": "image", + "image_path": "18d784b5b57590c11339e68f51d7667182fb99ce38ca1b9b2fdafab9408053b4.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 298, + 361, + 344, + 407 + ], + "blocks": [ + { + "bbox": [ + 298, + 361, + 344, + 407 + ], + "lines": [ + { + "bbox": [ + 298, + 361, + 344, + 407 + ], + "spans": [ + { + "bbox": [ + 298, + 361, + 344, + 407 + ], + "type": "image", + "image_path": "d0d6473201dbb9e31a141a8075d42b2ebc2d54aa9d61984e22c81be92ff9b0f2.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 298, + 415, + 344, + 462 + ], + "blocks": [ + { + "bbox": [ + 298, + 415, + 344, + 462 + ], + "lines": [ + { + "bbox": [ + 298, + 415, + 344, + 462 + ], + "spans": [ + { + "bbox": [ + 298, + 415, + 344, + 462 + ], + "type": "image", + "image_path": "003698f65df7c20d9323d31332dafeca0fd784babfbf5b2c8194c36139a1a5c7.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 298, + 464, + 344, + 511 + ], + "blocks": [ + { + "bbox": [ + 298, + 464, + 344, + 511 + ], + "lines": [ + { + "bbox": [ + 298, + 464, + 344, + 511 + ], + "spans": [ + { + "bbox": [ + 298, + 464, + 344, + 511 + ], + "type": "image", + "image_path": "9e223f4985b4b95143057b15b806b8d245964acb9ce24108a4c51097c29ef86e.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 298, + 519, + 344, + 566 + ], + "blocks": [ + { + "bbox": [ + 298, + 519, + 344, + 566 + ], + "lines": [ + { + "bbox": [ + 298, + 519, + 344, + 566 + ], + "spans": [ + { + "bbox": [ + 298, + 519, + 344, + 566 + ], + "type": "image", + "image_path": "d9f8f4a05f3210f49cc6e843082f26f4058dc5f87f669793e8122cf507006427.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 298, + 568, + 344, + 614 + ], + "blocks": [ + { + "bbox": [ + 298, + 568, + 344, + 614 + ], + "lines": [ + { + "bbox": [ + 298, + 568, + 344, + 614 + ], + "spans": [ + { + "bbox": [ + 298, + 568, + 344, + 614 + ], + "type": "image", + "image_path": "911be1d34710dcafd4c87a8b9c736226e484703c420460aa54783acd676f15e2.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 374, + 102, + 421, + 148 + ], + "blocks": [ + { + "bbox": [ + 385, + 89, + 410, + 99 + ], + "lines": [ + { + "bbox": [ + 385, + 89, + 410, + 99 + ], + "spans": [ + { + "bbox": [ + 385, + 89, + 410, + 99 + ], + "type": "text", + "content": "Object" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 374, + 102, + 421, + 148 + ], + "lines": [ + { + "bbox": [ + 374, + 102, + 421, + 148 + ], + "spans": [ + { + "bbox": [ + 374, + 102, + 421, + 148 + ], + "type": "image", + "image_path": "cb515daad63eeb70026ccda8c5907a2049bbbe26952d0689b52cf6ae88558c0d.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 374, + 151, + 421, + 198 + ], + "blocks": [ + { + "bbox": [ + 374, + 151, + 421, + 198 + ], + "lines": [ + { + "bbox": [ + 374, + 151, + 421, + 198 + ], + "spans": [ + { + "bbox": [ + 374, + 151, + 421, + 198 + ], + "type": "image", + "image_path": "372c56ae06ef95feb11eceb1c39c954e800d011ba6c6e23ea8f9df9d8b260ce8.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 374, + 205, + 421, + 251 + ], + "blocks": [ + { + "bbox": [ + 374, + 205, + 421, + 251 + ], + "lines": [ + { + "bbox": [ + 374, + 205, + 421, + 251 + ], + "spans": [ + { + "bbox": [ + 374, + 205, + 421, + 251 + ], + "type": "image", + "image_path": "b3a31eac67b73c57571526039bc44c6b88fe42c743259d894de3624d5f45fa76.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 374, + 256, + 421, + 303 + ], + "blocks": [ + { + "bbox": [ + 374, + 256, + 421, + 303 + ], + "lines": [ + { + "bbox": [ + 374, + 256, + 421, + 303 + ], + "spans": [ + { + "bbox": [ + 374, + 256, + 421, + 303 + ], + "type": "image", + "image_path": "21ba1b0515526447a961e128581ed2c924bf6fe1cf36a9f36e9b7063b2607472.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + } + ], + "index": 46 + }, + { + "type": "image", + "bbox": [ + 374, + 309, + 421, + 357 + ], + "blocks": [ + { + "bbox": [ + 374, + 309, + 421, + 357 + ], + "lines": [ + { + "bbox": [ + 374, + 309, + 421, + 357 + ], + "spans": [ + { + "bbox": [ + 374, + 309, + 421, + 357 + ], + "type": "image", + "image_path": "7f70af42a461f2b5bc9f206d506465fd2b2436bfe187e1cc7dafa2f4e0c48bd4.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + } + ], + "index": 47 + }, + { + "type": "image", + "bbox": [ + 374, + 361, + 421, + 408 + ], + "blocks": [ + { + "bbox": [ + 374, + 361, + 421, + 408 + ], + "lines": [ + { + "bbox": [ + 374, + 361, + 421, + 408 + ], + "spans": [ + { + "bbox": [ + 374, + 361, + 421, + 408 + ], + "type": "image", + "image_path": "71abfaadb1912d53f675c13b6865281883cffc2ce0233029f005c54860023a30.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + } + ], + "index": 48 + }, + { + "type": "image", + "bbox": [ + 374, + 415, + 421, + 461 + ], + "blocks": [ + { + "bbox": [ + 374, + 415, + 421, + 461 + ], + "lines": [ + { + "bbox": [ + 374, + 415, + 421, + 461 + ], + "spans": [ + { + "bbox": [ + 374, + 415, + 421, + 461 + ], + "type": "image", + "image_path": "116bfc79088af567816e3e5949e5a6f69898e8166ff8673a6e0b452013299be9.jpg" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_body" + } + ], + "index": 49 + }, + { + "type": "image", + "bbox": [ + 374, + 464, + 421, + 511 + ], + "blocks": [ + { + "bbox": [ + 374, + 464, + 421, + 511 + ], + "lines": [ + { + "bbox": [ + 374, + 464, + 421, + 511 + ], + "spans": [ + { + "bbox": [ + 374, + 464, + 421, + 511 + ], + "type": "image", + "image_path": "15e96c0aa635874fbd5bab5e9784c78b16d82b892de738c22f23455fbaec8702.jpg" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_body" + } + ], + "index": 50 + }, + { + "type": "image", + "bbox": [ + 374, + 519, + 421, + 566 + ], + "blocks": [ + { + "bbox": [ + 374, + 519, + 421, + 566 + ], + "lines": [ + { + "bbox": [ + 374, + 519, + 421, + 566 + ], + "spans": [ + { + "bbox": [ + 374, + 519, + 421, + 566 + ], + "type": "image", + "image_path": "5eff3e65c8b7bf434cb4e2550c8498edd1800676e1b88bc81e863b9ce32e4145.jpg" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_body" + } + ], + "index": 51 + }, + { + "type": "image", + "bbox": [ + 374, + 568, + 421, + 614 + ], + "blocks": [ + { + "bbox": [ + 374, + 568, + 421, + 614 + ], + "lines": [ + { + "bbox": [ + 374, + 568, + 421, + 614 + ], + "spans": [ + { + "bbox": [ + 374, + 568, + 421, + 614 + ], + "type": "image", + "image_path": "d912698def2d005067d1c0423f85d827f08d3400ebbe20d22834643948cd7252.jpg" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_body" + } + ], + "index": 52 + }, + { + "type": "image", + "bbox": [ + 435, + 102, + 481, + 148 + ], + "blocks": [ + { + "bbox": [ + 425, + 76, + 491, + 87 + ], + "lines": [ + { + "bbox": [ + 425, + 76, + 491, + 87 + ], + "spans": [ + { + "bbox": [ + 425, + 76, + 491, + 87 + ], + "type": "text", + "content": "Intrinsic concepts" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 443, + 89, + 474, + 99 + ], + "lines": [ + { + "bbox": [ + 443, + 89, + 474, + 99 + ], + "spans": [ + { + "bbox": [ + 443, + 89, + 474, + 99 + ], + "type": "text", + "content": "Material" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 435, + 102, + 481, + 148 + ], + "lines": [ + { + "bbox": [ + 435, + 102, + 481, + 148 + ], + "spans": [ + { + "bbox": [ + 435, + 102, + 481, + 148 + ], + "type": "image", + "image_path": "f4cd31a0f838010406b5dde62ce4e57fdf45e7256aef17ae284ce9eca9f454f2.jpg" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_body" + } + ], + "index": 54 + }, + { + "type": "image", + "bbox": [ + 435, + 151, + 481, + 198 + ], + "blocks": [ + { + "bbox": [ + 435, + 151, + 481, + 198 + ], + "lines": [ + { + "bbox": [ + 435, + 151, + 481, + 198 + ], + "spans": [ + { + "bbox": [ + 435, + 151, + 481, + 198 + ], + "type": "image", + "image_path": "a539b214618624a6761bb76ff0feadbc1f69ea6db88ff343e7c275ff65552160.jpg" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_body" + } + ], + "index": 55 + }, + { + "type": "image", + "bbox": [ + 435, + 205, + 481, + 251 + ], + "blocks": [ + { + "bbox": [ + 435, + 205, + 481, + 251 + ], + "lines": [ + { + "bbox": [ + 435, + 205, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 435, + 205, + 481, + 251 + ], + "type": "image", + "image_path": "9d141318618ed837da517facd02e300133b4e2cff29f4b4412243e54864e60d9.jpg" + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_body" + } + ], + "index": 56 + }, + { + "type": "image", + "bbox": [ + 435, + 255, + 481, + 303 + ], + "blocks": [ + { + "bbox": [ + 435, + 255, + 481, + 303 + ], + "lines": [ + { + "bbox": [ + 435, + 255, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 435, + 255, + 481, + 303 + ], + "type": "image", + "image_path": "1c2e15953df712ac6d87b5c70059d22e8f23181f53bc19af058a15ed5a36fddf.jpg" + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "image_body" + } + ], + "index": 57 + }, + { + "type": "image", + "bbox": [ + 435, + 309, + 481, + 357 + ], + "blocks": [ + { + "bbox": [ + 435, + 309, + 481, + 357 + ], + "lines": [ + { + "bbox": [ + 435, + 309, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 435, + 309, + 481, + 357 + ], + "type": "image", + "image_path": "f821de92b0d31d3623158e9d004c01d649939fc0063f9c0949d6bf5b7a813f9e.jpg" + } + ] + } + ], + "index": 58, + "angle": 0, + "type": "image_body" + } + ], + "index": 58 + }, + { + "type": "image", + "bbox": [ + 435, + 361, + 481, + 408 + ], + "blocks": [ + { + "bbox": [ + 435, + 361, + 481, + 408 + ], + "lines": [ + { + "bbox": [ + 435, + 361, + 481, + 408 + ], + "spans": [ + { + "bbox": [ + 435, + 361, + 481, + 408 + ], + "type": "image", + "image_path": "5eb9b66ace13319b694f46dcaf92935c5ff9adc6526cdfb580d6f9643f485cbc.jpg" + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "image_body" + } + ], + "index": 59 + }, + { + "type": "image", + "bbox": [ + 435, + 415, + 481, + 462 + ], + "blocks": [ + { + "bbox": [ + 435, + 415, + 481, + 462 + ], + "lines": [ + { + "bbox": [ + 435, + 415, + 481, + 462 + ], + "spans": [ + { + "bbox": [ + 435, + 415, + 481, + 462 + ], + "type": "image", + "image_path": "059670f93a9891a330ab07f080b5e7c70b6e8fd7b38c8efc964df1aa5a745373.jpg" + } + ] + } + ], + "index": 60, + "angle": 0, + "type": "image_body" + } + ], + "index": 60 + }, + { + "type": "image", + "bbox": [ + 435, + 464, + 481, + 511 + ], + "blocks": [ + { + "bbox": [ + 435, + 464, + 481, + 511 + ], + "lines": [ + { + "bbox": [ + 435, + 464, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 435, + 464, + 481, + 511 + ], + "type": "image", + "image_path": "3b74a28dfd3bd767ce5a031d4546598d8eb8c2092d75dfaa3d1b844686122b6b.jpg" + } + ] + } + ], + "index": 61, + "angle": 0, + "type": "image_body" + } + ], + "index": 61 + }, + { + "type": "image", + "bbox": [ + 435, + 519, + 481, + 566 + ], + "blocks": [ + { + "bbox": [ + 435, + 519, + 481, + 566 + ], + "lines": [ + { + "bbox": [ + 435, + 519, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 435, + 519, + 481, + 566 + ], + "type": "image", + "image_path": "e40a0315a20242769a5ef29622f35843d4011be00ba148f8d9ccf159c55c69aa.jpg" + } + ] + } + ], + "index": 62, + "angle": 0, + "type": "image_body" + } + ], + "index": 62 + }, + { + "type": "image", + "bbox": [ + 435, + 568, + 481, + 614 + ], + "blocks": [ + { + "bbox": [ + 435, + 568, + 481, + 614 + ], + "lines": [ + { + "bbox": [ + 435, + 568, + 481, + 614 + ], + "spans": [ + { + "bbox": [ + 435, + 568, + 481, + 614 + ], + "type": "image", + "image_path": "85207812dc6184607e49026fe773046655a14953e73d8ba9eddbc21382bc0004.jpg" + } + ] + } + ], + "index": 63, + "angle": 0, + "type": "image_body" + } + ], + "index": 63 + }, + { + "type": "image", + "bbox": [ + 496, + 102, + 542, + 148 + ], + "blocks": [ + { + "bbox": [ + 506, + 90, + 533, + 99 + ], + "lines": [ + { + "bbox": [ + 506, + 90, + 533, + 99 + ], + "spans": [ + { + "bbox": [ + 506, + 90, + 533, + 99 + ], + "type": "text", + "content": "Colour" + } + ] + } + ], + "index": 64, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 496, + 102, + 542, + 148 + ], + "lines": [ + { + "bbox": [ + 496, + 102, + 542, + 148 + ], + "spans": [ + { + "bbox": [ + 496, + 102, + 542, + 148 + ], + "type": "image", + "image_path": "170880547fd7af43f1b2326e1e48ef0e29f0438e7b29cd10e3b01404b293a4f3.jpg" + } + ] + } + ], + "index": 65, + "angle": 0, + "type": "image_body" + } + ], + "index": 65 + }, + { + "type": "image", + "bbox": [ + 496, + 151, + 542, + 198 + ], + "blocks": [ + { + "bbox": [ + 496, + 151, + 542, + 198 + ], + "lines": [ + { + "bbox": [ + 496, + 151, + 542, + 198 + ], + "spans": [ + { + "bbox": [ + 496, + 151, + 542, + 198 + ], + "type": "image", + "image_path": "19ee7cad0c1ccfa7e1d2a2f9de75265cd061592667f6bb2085046607b21624e3.jpg" + } + ] + } + ], + "index": 66, + "angle": 0, + "type": "image_body" + } + ], + "index": 66 + }, + { + "type": "image", + "bbox": [ + 496, + 205, + 542, + 251 + ], + "blocks": [ + { + "bbox": [ + 496, + 205, + 542, + 251 + ], + "lines": [ + { + "bbox": [ + 496, + 205, + 542, + 251 + ], + "spans": [ + { + "bbox": [ + 496, + 205, + 542, + 251 + ], + "type": "image", + "image_path": "20182432f1f3fe2a0cee4f04b29d9fc0aa524bfde1dbad8226ae1a4b7f0a4bbd.jpg" + } + ] + } + ], + "index": 67, + "angle": 0, + "type": "image_body" + } + ], + "index": 67 + }, + { + "type": "image", + "bbox": [ + 496, + 256, + 542, + 303 + ], + "blocks": [ + { + "bbox": [ + 496, + 256, + 542, + 303 + ], + "lines": [ + { + "bbox": [ + 496, + 256, + 542, + 303 + ], + "spans": [ + { + "bbox": [ + 496, + 256, + 542, + 303 + ], + "type": "image", + "image_path": "9661a9fa5f556f129b80e143e606c157c5979a68dce4b299ed8a03a305278aba.jpg" + } + ] + } + ], + "index": 68, + "angle": 0, + "type": "image_body" + } + ], + "index": 68 + }, + { + "type": "image", + "bbox": [ + 496, + 309, + 542, + 357 + ], + "blocks": [ + { + "bbox": [ + 496, + 309, + 542, + 357 + ], + "lines": [ + { + "bbox": [ + 496, + 309, + 542, + 357 + ], + "spans": [ + { + "bbox": [ + 496, + 309, + 542, + 357 + ], + "type": "image", + "image_path": "aee9329f8e5b9e7c94d761a72233d516961a49b6a35ee5e6d2b47c2e14d2f252.jpg" + } + ] + } + ], + "index": 69, + "angle": 0, + "type": "image_body" + } + ], + "index": 69 + }, + { + "type": "image", + "bbox": [ + 496, + 361, + 542, + 408 + ], + "blocks": [ + { + "bbox": [ + 496, + 361, + 542, + 408 + ], + "lines": [ + { + "bbox": [ + 496, + 361, + 542, + 408 + ], + "spans": [ + { + "bbox": [ + 496, + 361, + 542, + 408 + ], + "type": "image", + "image_path": "b40f12dd3ebea1b7b49cb1039431d7a8545e91480ab9e9b8708f918eccf86bfe.jpg" + } + ] + } + ], + "index": 70, + "angle": 0, + "type": "image_body" + } + ], + "index": 70 + }, + { + "type": "image", + "bbox": [ + 496, + 416, + 542, + 462 + ], + "blocks": [ + { + "bbox": [ + 496, + 416, + 542, + 462 + ], + "lines": [ + { + "bbox": [ + 496, + 416, + 542, + 462 + ], + "spans": [ + { + "bbox": [ + 496, + 416, + 542, + 462 + ], + "type": "image", + "image_path": "a34a703320c4b5337f1cee02e110381fe9a26b72622ca8000dcf59f0302bef19.jpg" + } + ] + } + ], + "index": 71, + "angle": 0, + "type": "image_body" + } + ], + "index": 71 + }, + { + "type": "image", + "bbox": [ + 496, + 464, + 542, + 511 + ], + "blocks": [ + { + "bbox": [ + 496, + 464, + 542, + 511 + ], + "lines": [ + { + "bbox": [ + 496, + 464, + 542, + 511 + ], + "spans": [ + { + "bbox": [ + 496, + 464, + 542, + 511 + ], + "type": "image", + "image_path": "0626c06bea49fafd325040eb7abc9e09a7bc8d029723b8bfd4aae1a494e17ea8.jpg" + } + ] + } + ], + "index": 72, + "angle": 0, + "type": "image_body" + } + ], + "index": 72 + }, + { + "type": "image", + "bbox": [ + 496, + 519, + 542, + 566 + ], + "blocks": [ + { + "bbox": [ + 496, + 519, + 542, + 566 + ], + "lines": [ + { + "bbox": [ + 496, + 519, + 542, + 566 + ], + "spans": [ + { + "bbox": [ + 496, + 519, + 542, + 566 + ], + "type": "image", + "image_path": "a268e652d80cbc83a806d189a7690e28e990880eabe087160e7ab42acef969a7.jpg" + } + ] + } + ], + "index": 73, + "angle": 0, + "type": "image_body" + } + ], + "index": 73 + }, + { + "type": "image", + "bbox": [ + 496, + 568, + 542, + 614 + ], + "blocks": [ + { + "bbox": [ + 496, + 568, + 542, + 614 + ], + "lines": [ + { + "bbox": [ + 496, + 568, + 542, + 614 + ], + "spans": [ + { + "bbox": [ + 496, + 568, + 542, + 614 + ], + "type": "image", + "image_path": "86a681c17a33a043643391e023952ea9bef8cc901a14fba327bf9d9f2f53b46d.jpg" + } + ] + } + ], + "index": 74, + "angle": 0, + "type": "image_body" + } + ], + "index": 74 + }, + { + "bbox": [ + 55, + 663, + 164, + 675 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 663, + 164, + 675 + ], + "spans": [ + { + "bbox": [ + 55, + 663, + 164, + 675 + ], + "type": "text", + "content": "5.2. Qualitative results" + } + ] + } + ], + "index": 76 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "To demonstrate the effectiveness of the proposed ICE framework, we present qualitative results, as illustrated in" + } + ] + } + ], + "index": 77 + }, + { + "bbox": [ + 313, + 663, + 553, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 663, + 553, + 688 + ], + "spans": [ + { + "bbox": [ + 313, + 663, + 553, + 688 + ], + "type": "text", + "content": "Figure 6, showcasing ICE's ability to accurately extract and decompose visual concepts systematically." + } + ] + } + ], + "index": 78 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "First, given an input image " + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": ", ICE extracts corresponding concept masks and retrieves relevant text-based descrip" + } + ] + } + ], + "index": 79 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23740" + } + ] + } + ], + "index": 80 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 71, + 111, + 123 + ], + "blocks": [ + { + "bbox": [ + 57, + 71, + 111, + 123 + ], + "lines": [ + { + "bbox": [ + 57, + 71, + 111, + 123 + ], + "spans": [ + { + "bbox": [ + 57, + 71, + 111, + 123 + ], + "type": "image", + "image_path": "37c036b79aaa8e5c41fe35427099d65316ad33cc1c883b2af4015cf481ba0993.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 127, + 555, + 159 + ], + "lines": [ + { + "bbox": [ + 55, + 127, + 555, + 159 + ], + "spans": [ + { + "bbox": [ + 55, + 127, + 555, + 159 + ], + "type": "text", + "content": "Figure 7. Qualitative comparison of segmentation results between ConceptExpress and our framework. Our framework not only delivers more accurate segmentation of object-level regions but also effectively retrieves the most relevant text-based concepts, demonstrating its unique capabilities over ConceptExpress." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 111, + 74, + 184, + 123 + ], + "blocks": [ + { + "bbox": [ + 111, + 74, + 184, + 123 + ], + "lines": [ + { + "bbox": [ + 111, + 74, + 184, + 123 + ], + "spans": [ + { + "bbox": [ + 111, + 74, + 184, + 123 + ], + "type": "image", + "image_path": "686aa1e4224702a41b9b47c9712e321b29f72ef11b255db4276a1ef42e01dd91.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 186, + 72, + 239, + 124 + ], + "blocks": [ + { + "bbox": [ + 186, + 72, + 239, + 124 + ], + "lines": [ + { + "bbox": [ + 186, + 72, + 239, + 124 + ], + "spans": [ + { + "bbox": [ + 186, + 72, + 239, + 124 + ], + "type": "image", + "image_path": "af3a018444c013dd344bd8b0aa21a7989964386c7db0e4a8c3576dcde42e645a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 240, + 74, + 312, + 123 + ], + "blocks": [ + { + "bbox": [ + 240, + 74, + 312, + 123 + ], + "lines": [ + { + "bbox": [ + 240, + 74, + 312, + 123 + ], + "spans": [ + { + "bbox": [ + 240, + 74, + 312, + 123 + ], + "type": "image", + "image_path": "6276d4649ae5cfbe2ea6796f3d990cff27c819e6a3d4d339dd3f9624845148b2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 315, + 72, + 368, + 124 + ], + "blocks": [ + { + "bbox": [ + 315, + 72, + 368, + 124 + ], + "lines": [ + { + "bbox": [ + 315, + 72, + 368, + 124 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 368, + 124 + ], + "type": "image", + "image_path": "9d9515a877f54e7a838fb4cd1e387319d9d87a3a688d78d005a4d91b272398cb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 369, + 74, + 421, + 123 + ], + "blocks": [ + { + "bbox": [ + 369, + 74, + 421, + 123 + ], + "lines": [ + { + "bbox": [ + 369, + 74, + 421, + 123 + ], + "spans": [ + { + "bbox": [ + 369, + 74, + 421, + 123 + ], + "type": "image", + "image_path": "4494398349677432efd5b785a09799a9d67774bde15135ac5202fa533d181a98.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 422, + 72, + 478, + 124 + ], + "blocks": [ + { + "bbox": [ + 422, + 72, + 478, + 124 + ], + "lines": [ + { + "bbox": [ + 422, + 72, + 478, + 124 + ], + "spans": [ + { + "bbox": [ + 422, + 72, + 478, + 124 + ], + "type": "image", + "image_path": "b18eac24960c1508e208f030115c4da5b3c2fc22977e01f76e6b7af3934312d3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 478, + 74, + 490, + 123 + ], + "blocks": [ + { + "bbox": [ + 478, + 74, + 490, + 123 + ], + "lines": [ + { + "bbox": [ + 478, + 74, + 490, + 123 + ], + "spans": [ + { + "bbox": [ + 478, + 74, + 490, + 123 + ], + "type": "image", + "image_path": "3d503ed9fda1fbdb20388c9edbb457cc6451466e3dbd555a37cf8205a4df6a16.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 491, + 74, + 551, + 123 + ], + "blocks": [ + { + "bbox": [ + 491, + 74, + 551, + 123 + ], + "lines": [ + { + "bbox": [ + 491, + 74, + 551, + 123 + ], + "spans": [ + { + "bbox": [ + 491, + 74, + 551, + 123 + ], + "type": "image", + "image_path": "dac1f2ad33523c1e9a7866bcb673db1096a894e110046582bb5bea7c6569832e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 61, + 173, + 408, + 252 + ], + "blocks": [ + { + "bbox": [ + 107, + 161, + 361, + 172 + ], + "lines": [ + { + "bbox": [ + 107, + 161, + 361, + 172 + ], + "spans": [ + { + "bbox": [ + 107, + 161, + 361, + 172 + ], + "type": "text", + "content": "Table 4. Ablation study on ICE model components on the " + }, + { + "bbox": [ + 107, + 161, + 361, + 172 + ], + "type": "inline_equation", + "content": "D1" + }, + { + "bbox": [ + 107, + 161, + 361, + 172 + ], + "type": "text", + "content": " dataset." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 173, + 408, + 252 + ], + "lines": [ + { + "bbox": [ + 61, + 173, + 408, + 252 + ], + "spans": [ + { + "bbox": [ + 61, + 173, + 408, + 252 + ], + "type": "table", + "html": "
Model's componentsCLIP [23] encoderDINO [6] encoder
MethodStage One's maskStage Two's token init.Stage Two learningSIM1(%)SIMC(%)ACC1(%)ACC3(%)SIM1(%)SIMC(%)ACC1(%)ACC3(%)
ConceptExpressXXX0.6890.7840.2630.3850.3190.5680.3240.470
ICE w. mask✓XX0.7100.7810.3070.4560.4930.6010.3950.604
ICE w.o Stage Two✓✓X0.7260.8070.3010.4520.5010.6210.4110.604
ICE w.o text init.✓X✓0.7220.8140.3200.4750.5480.6430.4490.627
ICE (Ours)✓✓✓0.7380.8220.3250.5180.6770.7550.4760.638
", + "image_path": "0d3018bee43e09eaee90d61aaeb947fcca13e3203b387e11d881d1efc6271a46.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 413, + 202, + 551, + 243 + ], + "blocks": [ + { + "bbox": [ + 413, + 179, + 551, + 200 + ], + "lines": [ + { + "bbox": [ + 413, + 179, + 551, + 200 + ], + "spans": [ + { + "bbox": [ + 413, + 179, + 551, + 200 + ], + "type": "text", + "content": "Table 5. Quantitative comparison of generated masks on the " + }, + { + "bbox": [ + 413, + 179, + 551, + 200 + ], + "type": "inline_equation", + "content": "D1" + }, + { + "bbox": [ + 413, + 179, + 551, + 200 + ], + "type": "text", + "content": " dataset." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 413, + 202, + 551, + 243 + ], + "lines": [ + { + "bbox": [ + 413, + 202, + 551, + 243 + ], + "spans": [ + { + "bbox": [ + 413, + 202, + 551, + 243 + ], + "type": "table", + "html": "
MethodmIoURecallPrecision
ConceptExpress0.4830.6760.657
ICE (Ours)0.6350.8930.720
", + "image_path": "4cc642f02c60ebe1ba10c89a50da870c4e5e8c4b2e6397ba8928d51db95aa774.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 54, + 261, + 296, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 261, + 296, + 428 + ], + "spans": [ + { + "bbox": [ + 54, + 261, + 296, + 428 + ], + "type": "text", + "content": "tors, as shown in the second column of the figure. These masks precisely delineate regions within the image, ensuring accurate localization of each concept. Next, leveraging the extracted masks and retrieved texts, ICE learns distinct object-level concepts, displayed in the third column. For example, the framework identifies objects such as “Bust” and “Cone” and associates them with their respective semantic descriptors. Finally, utilizing the information from the object-level concepts, ICE systematically discovers intrinsic concepts, including attributes like material and colour, through its second-stage processing. This decomposition enables the framework to uncover compositional intrinsic concepts such as the object type, material, and colour, as demonstrated in the fourth column." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 430, + 296, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 430, + 296, + 525 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 296, + 525 + ], + "type": "text", + "content": "This structured approach highlights the superiority of our framework by effectively leveraging a single T2I model to systematically uncover both object-level and intrinsic visual concepts. The ability to deconstruct an image into its fundamental components not only validates the robustness of ICE but also illustrates the potential of T2I models to facilitate comprehensive and structured concept learning in an unsupervised manner." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 536, + 202, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 536, + 202, + 550 + ], + "spans": [ + { + "bbox": [ + 55, + 536, + 202, + 550 + ], + "type": "text", + "content": "5.3. Model component analysis" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 556, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 556, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 295, + 664 + ], + "type": "text", + "content": "ICE ablation study. To validate the effectiveness of our ICE framework, we conduct an ablation study on its individual components: Stage One's retrieved mask, Stage Two's token embedding initialization using text retrieved from Stage One, and Stage Two's Structured Concept Learning. The results, as shown in Table 4, demonstrate that each component significantly enhances performance. Furthermore, the complete ICE framework outperforms all its variants across every metric in the UCE benchmarks." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "ICE Stage One's mask quality. Furthermore, Table 5 compares our Stage One generated masks with ConceptExpress using " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "D1" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "'s manually annotated ground-truth masks. We evaluate the segmentation performance of generated masks" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 261, + 555, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 261, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 261, + 555, + 453 + ], + "type": "text", + "content": "against ground-truth masks using the Hungarian algorithm to find the optimal matching between the sets of masks. We evaluate each pair of predicted and ground-truth masks using segmentation criterions, forming a similarity matrix. The Hungarian algorithm is then applied to this matrix to determine the optimal one-to-one correspondence that maximizes the overall similarity. Overall, our Stage One technique achieves better segmentation results across different criterion. Notably, our Stage One masks exhibit significantly higher recall rates, indicating that a greater proportion of the relevant object-level regions in the annotated area are successfully captured. This higher recall demonstrates that ICE's Automatic Concept Localization method in Stage One is more effective at identifying and extracting meaningful concepts, reducing the likelihood of missing important object-level concepts compared to previous methods." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 314, + 479, + 387, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 479, + 387, + 491 + ], + "spans": [ + { + "bbox": [ + 314, + 479, + 387, + 491 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 505, + 555, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 505, + 555, + 672 + ], + "spans": [ + { + "bbox": [ + 313, + 505, + 555, + 672 + ], + "type": "text", + "content": "In this paper, we introduce ICE, a novel framework for automatically and systematically extracting intrinsic concepts from a single image via diffusion models. Our framework consists of two stages. In Stage One, we design a concept localization module based on the T2I model that effectively retrieves the most relevant text-based concepts and their corresponding masks. Subsequently, in Stage Two, we introduce a structured concept learning method to systematically decompose object-level concepts into their intrinsic attributes, resulting in richer and more interpretable concept representations. In summary, our approach effectively harnesses Text-to-Image (T2I) models to enhance concept learning, laying the groundwork for further advancements in intrinsic visual concept extraction." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "Acknowledgments. This work is supported by the Hong Kong Research Grants Council - General Research Fund (Grant No.: 17211024)." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23741" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "text", + "content": "[1] Omri Avrahami, Kfir Aberman, Ohad Fried, Daniel Cohen-Or, and Dani Lischinski. Break-a-scene: Extracting multiple concepts from a single image. In SIGGRAPH Asia, 2023. 2, 3, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 137, + 296, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 296, + 179 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 296, + 179 + ], + "type": "text", + "content": "[2] Omri Avrahami, Thomas Hayes, Oran Gafni, Sonal Gupta, Yaniv Taigman, Devi Parikh, Dani Lischinski, Ohad Fried, and Xi Yin. Spatext: Spatio-textual representation for controllable image generation. In CVPR, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 182, + 295, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 182, + 295, + 214 + ], + "spans": [ + { + "bbox": [ + 62, + 182, + 295, + 214 + ], + "type": "text", + "content": "[3] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. In ICML, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 216, + 295, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 216, + 295, + 258 + ], + "spans": [ + { + "bbox": [ + 62, + 216, + 295, + 258 + ], + "type": "text", + "content": "[4] Usha Bhalla, Alex Oesterling, Suraj Srinivas, Flavio P Calmon, and Himabindu Lakkaraju. Interpreting clip with sparse linear concept embeddings (splice). In NeurIPS, 2024. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 261, + 295, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 261, + 295, + 293 + ], + "spans": [ + { + "bbox": [ + 62, + 261, + 295, + 293 + ], + "type": "text", + "content": "[5] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 296, + 295, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 296, + 295, + 338 + ], + "spans": [ + { + "bbox": [ + 62, + 296, + 295, + 338 + ], + "type": "text", + "content": "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 6, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 341, + 295, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 341, + 295, + 384 + ], + "spans": [ + { + "bbox": [ + 62, + 341, + 295, + 384 + ], + "type": "text", + "content": "[7] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In ICLR, 2023. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 386, + 295, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 386, + 295, + 429 + ], + "spans": [ + { + "bbox": [ + 62, + 386, + 295, + 429 + ], + "type": "text", + "content": "[8] Shaozhe Hao, Kai Han, Shihao Zhao, and Kwan-Yee K Wong. Vico: Detail-preserving visual condition for personalized text-to-image generation. arXiv preprint arXiv:2306.00971, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 431, + 295, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 431, + 295, + 474 + ], + "spans": [ + { + "bbox": [ + 62, + 431, + 295, + 474 + ], + "type": "text", + "content": "[9] Shaozhe Hao, Kai Han, Zhengyao Lv, Shihao Zhao, and Kwan-Yee K Wong. Conceptexpress: Harnessing diffusion models for single-image unsupervised concept extraction. In ECCV, 2024. 2, 3, 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 476, + 295, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 476, + 295, + 519 + ], + "spans": [ + { + "bbox": [ + 57, + 476, + 295, + 519 + ], + "type": "text", + "content": "[10] Eric Hedlin, Gopal Sharma, Shweta Mahajan, Hossam Isack, Abhishek Kar, Andrea Tagliasacchi, and Kwang Moo Yi. Unsupervised semantic correspondence using stable diffusion. In NeurIPS, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 521, + 295, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 521, + 295, + 553 + ], + "spans": [ + { + "bbox": [ + 57, + 521, + 295, + 553 + ], + "type": "text", + "content": "[11] Jonathan Ho, Tim Salimans, Alexey A Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 555, + 295, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 555, + 295, + 599 + ], + "spans": [ + { + "bbox": [ + 57, + 555, + 295, + 599 + ], + "type": "text", + "content": "[12] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Weli-hinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 601, + 295, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 601, + 295, + 644 + ], + "spans": [ + { + "bbox": [ + 57, + 601, + 295, + 644 + ], + "type": "text", + "content": "[13] Chen Jin, Ryutaro Tanno, Amrutha Saseendran, Tom Diethe, and Philip Teare. An image is worth multiple words: Discovering object level concepts using multi-concept prompt learning. In ICML, 2023. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 647, + 295, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 678 + ], + "type": "text", + "content": "[14] Laurynas Karazija, Iro Laina, Andrea Vedaldi, and Christian Rupprecht. Diffusion models for zero-shot open-vocabulary segmentation. In ECCV, 2023. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 681, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 712 + ], + "type": "text", + "content": "[15] Sharon Lee, Yunzhi Zhang, Shangzhe Wu, and Jiajun Wu. Language-informed visual concept learning. In ICLR, 2023. 2, 3" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "type": "text", + "content": "[16] Alexander C. Li, Mihir Prabhudesai, Shivam Duggal, Ellis Brown, and Deepak Pathak. Your diffusion model is secretly a zero-shot classifier. In ICCV, 2023. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 107, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 139 + ], + "type": "text", + "content": "[17] Dongxu Li, Junnan Li, and Steven CH Hoi. Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. In NeurIPS, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 140, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 553, + 171 + ], + "type": "text", + "content": "[18] Xinghui Li, Jingyi Lu, Kai Han, and Victor Prisacariu. Sd4match: Learning to prompt stable diffusion model for semantic matching. In CVPR, 2024. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 173, + 553, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 173, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 553, + 215 + ], + "type": "text", + "content": "[19] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In CVPR, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 217, + 553, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 258 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 258 + ], + "type": "text", + "content": "[20] Minheng Ni, Yabo Zhang, Kailai Feng, Xiaoming Li, Yiwen Guo, and Wangmeng Zuo. Ref-diff: Zero-shot referring image segmentation with generative models. arXiv preprint arXiv:2308.16777, 2023. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 262, + 553, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 262, + 553, + 314 + ], + "spans": [ + { + "bbox": [ + 316, + 262, + 553, + 314 + ], + "type": "text", + "content": "[21] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob Mcgrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In ICML, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 317, + 553, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 317, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 316, + 317, + 553, + 358 + ], + "type": "text", + "content": "[22] Zeju Qiu, Weiyang Liu, Haiwen Feng, Yuxuan Xue, Yao Feng, Zhen Liu, Dan Zhang, Adrian Weller, and Bernhard Scholkopf. Controlling text-to-image diffusion by orthogonal finetuning. In NeurIPS, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 361, + 553, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 361, + 553, + 414 + ], + "spans": [ + { + "bbox": [ + 317, + 361, + 553, + 414 + ], + "type": "text", + "content": "[23] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 3, 6, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 416, + 553, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 553, + 458 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 553, + 458 + ], + "type": "text", + "content": "[24] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents. arXiv preprint arXiv:2204.06125, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 460, + 553, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 553, + 491 + ], + "type": "text", + "content": "[25] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 494, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 494, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 317, + 494, + 553, + 536 + ], + "type": "text", + "content": "[26] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, 2023. 2, 3, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 537, + 553, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 537, + 553, + 591 + ], + "spans": [ + { + "bbox": [ + 317, + 537, + 553, + 591 + ], + "type": "text", + "content": "[27] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 593, + 553, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 593, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 317, + 593, + 553, + 646 + ], + "type": "text", + "content": "[28] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 553, + 689 + ], + "type": "text", + "content": "[29] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. In ICLR, 2022. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[30] Junjiao Tian, Lavisha Aggarwal, Andrea Colaco, Zsolt Kira, and Mar Gonzalez-Franco. Diffuse, attend, and segment:" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23742" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 454 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "Unsupervised zero-shot segmentation using stable diffusion. In CVPR, 2024. 1, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 129 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 129 + ], + "type": "text", + "content": "[31] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In CVPR, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 129, + 294, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 163 + ], + "type": "text", + "content": "[32] Yael Vinker, Andrey Voynov, Daniel Cohen-Or, and Ariel Shamir. Concept decomposition for visual exploration and inspiration. In SIGGRAPH Asia, 2023. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 164, + 294, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 164, + 294, + 208 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 294, + 208 + ], + "type": "text", + "content": "[33] Jinglong Wang, Xiawei Li, Jing Zhang, Qingyuan Xu, Qin Zhou, Qian Yu, Lu Sheng, and Dong Xu. Diffusion model is secretly a training-free open vocabulary semantic segmenter. arXiv preprint arXiv:2309.02773, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 209, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 262 + ], + "type": "text", + "content": "[34] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Weixian Lei, Yuchao Gu, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In ICCV, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 264, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 308 + ], + "type": "text", + "content": "[35] Changming Xiao, Qi Yang, Feng Zhou, and Changshui Zhang. From text to mask: Localizing entities using the attention of text-to-image diffusion models. In Neurocomputing, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 309, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 294, + 342 + ], + "type": "text", + "content": "[36] Zhi Xu, Shaozhe Hao, and Kai Han. Cusconcept: Customized visual concept decomposition with diffusion models. In WACV, 2025. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 343, + 294, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 294, + 397 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 294, + 397 + ], + "type": "text", + "content": "[37] Junyi Zhang, Charles Herrmann, Junhwa Hur, Luisa Polania Cabrera, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. A tale of two features: Stable diffusion complements dino for zero-shot semantic correspondence. In NeurIPS, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 399, + 294, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 294, + 422 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 294, + 422 + ], + "type": "text", + "content": "[38] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 422, + 294, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 454 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 454 + ], + "type": "text", + "content": "[39] Yanbing Zhang, Mengping Yang, Qin Zhou, and Zhe Wang. Attention calibration for disentangled text-to-image personalization. In CVPR, 2024. 2" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23743" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/ICP_ Immediate Compensation Pruning for Mid-to-high Sparsity/9ac0c560-9a5b-4372-9074-16698b23f1e1_content_list.json b/2025/ICP_ Immediate Compensation Pruning for Mid-to-high Sparsity/9ac0c560-9a5b-4372-9074-16698b23f1e1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..84221064ec54e88ad47a95fc228a3a5cd6bbac95 --- /dev/null +++ b/2025/ICP_ Immediate Compensation Pruning for Mid-to-high Sparsity/9ac0c560-9a5b-4372-9074-16698b23f1e1_content_list.json @@ -0,0 +1,1508 @@ +[ + { + "type": "text", + "text": "ICP: Immediate Compensation Pruning for Mid-to-high Sparsity", + "text_level": 1, + "bbox": [ + 171, + 130, + 831, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xin Luo $^{1,2}$ Xueming Fu $^{1,2}$ Zihang Jiang $^{1,2*}$ S.Kevin Zhou $^{1,2,3,4*}$", + "bbox": [ + 192, + 178, + 805, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 School of Biomedical Engineering, Division of Life Sciences and Medicine, USTC", + "bbox": [ + 163, + 198, + 833, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 MIRACLE Center, Suzhou Institute for Advance Research, USTC", + "bbox": [ + 228, + 215, + 792, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Key Laboratory of Intelligent Information Processing of CAS, ICT, CAS", + "bbox": [ + 205, + 232, + 790, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ State Key Laboratory of Precision and Intelligent Chemistry, USTC", + "bbox": [ + 223, + 251, + 772, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{luoxin.ustc, xuemingfu, jzh0103}@mail.ustc.edu.cn", + "bbox": [ + 279, + 271, + 717, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{s.kevin.zhou}@gmail.com", + "bbox": [ + 393, + 289, + 604, + 304 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 339, + 326, + 354 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The increasing adoption of large-scale models under 7 billion parameters in both language and vision domains enables inference tasks on a single consumer-grade GPU but makes fine-tuning models of this scale, especially 7B models, challenging. This limits the applicability of pruning methods that require full fine-tuning. Meanwhile, pruning methods that do not require fine-tuning perform well at low sparsity levels (10%-50%) but struggle at mid-to-high sparsity levels (50%-70%), where the error behaves equivalently to that of semi-structured pruning. To address these issues, this paper introduces ICP, which finds a balance between full fine-tuning and zero fine-tuning. First, Sparsity Rearrange is used to reorganize the predefined sparsity levels, followed by Block-wise Compensate Pruning, which alternates pruning and compensation on the model's backbone, fully utilizing inference results while avoiding full model fine-tuning. Experiments show that ICP improves performance at mid-to-high sparsity levels compared to baselines, with only a slight increase in pruning time and no additional peak memory overhead.", + "bbox": [ + 86, + 369, + 483, + 674 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 700, + 220, + 717 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Introduction of Transformer [5, 31] propelled both language and vision models toward larger scales, leading to the development of high-performance models such as the OPT [37] and Llama [17, 28, 29] series for language tasks, and the SAM [13] series for vision tasks. However, their large scale presents significant deployment challenges. In scenarios with downstream tasks or hardware constraints, models under 8 billion (8B) parameters for language tasks and under 650 million parameters for vision tasks are most commonly used [12, 14, 16, 34, 36]. Nevertheless, model compres", + "bbox": [ + 89, + 726, + 482, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sion is necessary to further reduce their size for deployment. Moreover, on a single desktop-class GPU, it is typically only feasible to perform inference with these models, while fine-tuning remains impractical due to hardware limitations.", + "bbox": [ + 511, + 340, + 903, + 400 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pruning is a critical model compression technique, often categorized into multi-shot and one-shot methods based on pruning frequency. Traditional iterative pruning methods, based on the Lottery Ticket Hypothesis [6], involve an iterative prune-and-finetune process to gradually identify a sparse subnetwork that performs comparably to the original model, thereby reducing the number of model parameters. Classic methods like magnitude pruning [8] and more recent approaches such as LLM-Pruner [18], which utilizes block pruning with LoRA [11] fine-tuning, adhere to this hypothesis. However, for large models, global fine-tuning is resource-intensive, demanding significant GPU memory. Although LLM-Pruner reduces memory requirements using LoRA, it still requires storage for intermediate results and gradients, while methods like LLM Surgeon [30], despite avoiding global fine-tuning, also demand substantial memory and time due to the use of second-order derivatives. In contrast, one-shot pruning methods have gained traction by minimizing memory usage during fine-tuning. SparseGPT [7] avoids global fine-tuning by compensating for precision within each basic layer, thus accelerating the pruning process. Wanda [27] further simplifies this by eliminating compensation altogether, using a unique significance score to achieve performance comparable to SparseGPT without the need for fine-tuning. PruneZero[4] constructs an operator search space by combining input, output, and gradient operations to optimize model metrics. While, it requires multiple full inferences across models, which is computationally expensive. In memory-constrained settings, chunked model loading further exacerbates the inference time.", + "bbox": [ + 511, + 401, + 906, + 868 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, both SparseGPT and Wanda have notable limitations. (i) While these methods perform well at lower", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9487", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sparsity levels (10%-50%) when pruning large language models, their effectiveness significantly declines at mid-to-high sparsity levels (50%-70%). (ii) SparseGPT and Wanda show poor performance on vision models. Wanda, in particular, relies on priors that are specific to language models, making it even less effective than SparseGPT for vision tasks.", + "bbox": [ + 89, + 90, + 483, + 196 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e39f481947dbafb8d0c3ad62fe3b5e72ca9d2735a2e7a91c33e4dbeea7f3abbc.jpg", + "image_caption": [ + "(a) SparseGPT", + "Figure 1. Error propagation with respect to layer index." + ], + "image_footnote": [], + "bbox": [ + 94, + 205, + 285, + 323 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7aa0a1012d41d66ec91272005f6e030f48648c9e1d9a1d6f47a547a93f0b0987.jpg", + "image_caption": [ + "(b) Wanda" + ], + "image_footnote": [], + "bbox": [ + 295, + 207, + 483, + 323 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To illustrate why we focus on mid-to-high sparsity levels, we present the error accumulation during the pruning process of OPT-1.3B using SparseGPT and Wanda in Fig.1a and Fig.1b, with errors measured by mean square error (MSE). As shown, both SparseGPT and Wanda prune blocks sequentially without inter-block interaction, leading to faster error accumulation in deeper layers, which aligns with intuitive expectations. Furthermore, the error accumulation trends of semistructured sparsity 2:4 and 4:8 differ from those of unstructured sparsity $50\\%$ , instead aligning more closely with higher sparsity levels. Fig.1a, indicates that incorporating quantization strategies further shifts the error trends toward those of higher sparsity levels. Therefore, mitigating inter-block error propagation at sparsity levels of $50\\% -70\\%$ can effectively enhance the performance of unstructured pruning and hybrid compression strategies.", + "bbox": [ + 88, + 383, + 482, + 626 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose ICP, a large model pruning method that is applicable to both vision and language models. Improve performance at both low- and high-sparsity levels in pruning large language and vision models, without increasing GPU memory usage and pruning time. Specifically, we design a sliding window-based inter-block error compensation method based on the principle of error propagation. As shown in Fig. 2, after pruning each block, the subsequent block compensates for the pruning-induced errors of the previous block before pruning itself. This approach ensures that errors introduced during pruning are compensated for in the next block, preventing error propagation to deeper layers and avoiding the need for full-model fine-tuning during the pruning process. We highlight our contributions as follows:", + "bbox": [ + 89, + 626, + 482, + 851 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. We introduce the concept of inter-block error compensation, demonstrating that this approach can enhance the performance of one-shot pruning methods with limited", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d17cf899e889fca9ac16c06445f3dd6dd8b5b904d568d46f0bcdcf7847d90b0d.jpg", + "image_caption": [ + "Figure 2. Comparison between our proposed method ICP and conventional pruning methods. (a) illustrates a standard pruning approach, where all blocks are pruned simultaneously. (b) demonstrates our method. After pruning one block, the subsequent block adjusts to compensate for the error introduced by the pruning of the previous block." + ], + "image_footnote": [], + "bbox": [ + 555, + 98, + 864, + 258 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "computational resources.", + "bbox": [ + 532, + 393, + 700, + 407 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "2. Building on this concept, we further design a sliding window-based inter-block error compensation strategy. After pruning each layer, we use the next unpruned layer to compensate for the errors induced by pruning the current layer. It improves post-pruning model performance in both vision and language models.", + "3. Extensive experiments on large language models and vision models show that our method significantly enhances post-pruning model performance without increasing the GPU memory consumption." + ], + "bbox": [ + 513, + 407, + 906, + 559 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 592, + 653, + 607 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Transformer Architecture", + "text_level": 1, + "bbox": [ + 513, + 625, + 750, + 638 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Since its introduction [31], Transformer has become the cornerstone of language models due to its superior performance in language tasks. Following the development of GPT [24], researchers have built increasingly large models using Transformer as primary building blocks. The Transformer has also proven effective in vision tasks. For example, SAM [13], which is based on the Vision Transformer, particularly excels in segmentation and serves as a foundational model for various downstream tasks [20, 23, 34].", + "bbox": [ + 511, + 651, + 906, + 787 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Both language and vision models utilize a backbone composed of multiple Transformer blocks. The majority of parameters in these models are concentrated within their respective backbones. These Transformer blocks consist of multi-head self-attention and feed-forward neural networks with two linear layers. Multi-head self-attention is generally considered more critical than the feed-forward network.", + "bbox": [ + 511, + 795, + 908, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9488", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Large Model Pruning", + "text_level": 1, + "bbox": [ + 89, + 90, + 294, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With the emergence of large models, the computational and memory costs of global fine-tuning made iterative pruning methods prohibitively expensive. The speed and lower computational resource requirements of one-shot pruning methods have led to the development of several such techniques for large models, notably SparseGPT and Wanda. SparseGPT [7] is a one-shot rapid pruning method belonging to Optimal Brain Surgeon [10] family. It independently analyzes each basic layer of the model, formulates the pruning problem as a constrained optimization problem, and solves the pruned weight values using the Lagrange multiplier method. SparseGPT builds on the OBS foundation by leveraging the property of Cholesky decomposition to efficiently obtain different orders of Hessian matrices, significantly accelerating the pruning process and reducing memory usage. Essentially, SparseGPT consists of two parts: pruning and compensation, which are integrated during the Lagrange multiplier method's solution process. However, SparseGPT compensates for pruning errors by adjusting unpruned weights within each basic layer, without considering inter-block weight interactions. Wanda [27], introduces a strong prior for language models, assuming that weight importance is related to both the absolute value of the weight and the corresponding output. Wanda designs an importance metric that achieves good results without compensating for pruning losses. However, subsequent studies [33] suggest that while Wanda performs well in many cases, it can occasionally fall short compared to SparseGPT, and its performance in vision models is particularly lacking, where its strong priors may lead to suboptimal results.", + "bbox": [ + 89, + 112, + 485, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 582, + 183, + 598 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our objective is to develop an efficient pruning method for large models, including both vision and language models which should be executable within a reasonable time frame, even under limited computational and memory resources. Additionally, it should sustain robust performance at mid-to-high sparsity levels. In this section, we introduce ICP, a pruning strategy that effectively mitigates performance degradation at mid-to-high sparsity levels and facilitates rapid pruning in memory-constrained environments. The following sections detail our method in two aspects: Blockwise Compensate Pruning and Sparsity Rearrangement.", + "bbox": [ + 89, + 609, + 483, + 776 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Block-wise Compensate Pruning", + "text_level": 1, + "bbox": [ + 89, + 787, + 377, + 803 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose Block-wise Compensate Pruning, which enhances inter-block error compensation by building upon SparseGPT. As depicted in Fig. 3(a), Block-wise Compensate Pruning leverages the modular structure of large models, processing each block sequentially using a sliding window approach. In simple terms, the parameters of the next", + "bbox": [ + 89, + 810, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "block are updated to compensate for the pruning-induced errors in the preceding block. Initially, the sliding window contains two blocks. Since the operations are consistent after each slide, we explain using the first two blocks.", + "bbox": [ + 511, + 90, + 906, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In Fig. 3(a), the sliding window moves from left to right with a window length encompassing two blocks. Initially, the window contains $B^{1}$ and $B^{2}$ . Pruning is first applied to $B^{1}$ , followed by compensating for the pruning-induced error in $B^{1}$ using $B^{2}$ . The detailed calculation procedure for this compensation is outlined in Algorithm 1, as follows: The input data from the calibration set passes through the position encoding module, producing an intermediate result, $z^{0}$ , which is then duplicated to create $z^{0'}$ . At this stage, the feature flow splits into two streams: the \"standard stream\" $z^{0'}$ and the \"error stream\" $z^{0}$ . Before pruning $B^{1}$ , $z^{0'}$ passes through $B^{1}$ , producing an unbiased output, $z^{1'}$ . SparseGPT is then applied to prune $B^{1}$ , and $z^{0}$ passes through the pruned $B^{1}$ , resulting in a biased output, $z^{2}$ . Subsequently, $z^{1'}$ passes through the unpruned $B^{1}$ , yielding an unbiased output, $z^{2'}$ . Here, $z^{1}$ serves as both the output of the pruned $B^{1}$ and the input to $B^{2}$ . Since $z^{1}$ is biased, the output from $B^{1}$ will also be biased, whereas the desired output is $z^{2'}$ . Therefore, $z^{2'}$ serves as the target for $z^{1}$ , forming a labeled feature dataset ( $z^{1}, z^{2'}$ ). To align output of $B^{1}$ with $z^{2'}$ , we adjust $B^{1}$ 's parameters by training it on this feature dataset, reducing the output bias. Finally, the sliding window advances by one block, and the process is repeated: prune the first block and adjust the parameters of the second block using newly formed feature dataset. As above, Block-wise Compensate Pruning progresses sequentially along the backbone, processing one block at a time.", + "bbox": [ + 511, + 155, + 908, + 564 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Note that this method operates with a sliding window containing only two blocks, using the second block to compensate for the pruning-induced errors in the first, rather than relying on additional blocks for error correction. In this setup, once the first block is pruned, the standard stream dataset can pass through the second block to generate labels for the error stream. These labels are then reused during the compensation phase of the next sliding window, eliminating the need to store an additional block for real-time label computation. Consequently, at any given time, at most the weights of a single block reside in memory, optimizing both storage and computation. By contrast, if compensation required two subsequent blocks, the process would either necessitate storing two states of the standard stream or duplicating the next two blocks as a teacher model for distillation-like compensation, both of which would substantially increase memory overhead. It only requires a single block to be loaded into memory at a given moment. As each block is loaded into memory and pruned, the previous block is unloaded before the next block is loaded. This approach significantly reduces memory usage during the pruning and compensation process.", + "bbox": [ + 511, + 569, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9489", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/dd8117578032474846e58faa2ac98734446693915b595917a331516198257a04.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 138, + 89, + 864, + 222 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6ace0e570d1c8586525828bd982e3a6efa6a21802c299a5664646228f474203e.jpg", + "image_caption": [ + "Figure 3. Overview of the proposed method. (a) Block-wise Compensate Pruning for sequential block pruning and compensation; (b) basic structure of a transformer block (showing only weight-intensive and essential layers);" + ], + "image_footnote": [], + "bbox": [ + 138, + 222, + 857, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 Block-wise Compensate Pruning", + "text_level": 1, + "bbox": [ + 91, + 430, + 398, + 445 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "Input: Backbone of large model $\\{B^i\\}_{i = 1}^n$ , i-th block $B^{i}$ , Pruning function Prune (module, hyperparameters), Calibration dataset $x_{j}$ , Position embedding block $PE$ Update function Update (module,(sample, label)), Sparsity s, Sparsity Rearrangement parameter $\\alpha ,\\beta$", + "Output: Pruned backbone $\\{B^i\\}_{i = 1}^n$", + "1: $z^0 = PE(\\{x_j\\})$", + "2: $z_{i}^{0^{\\prime}} = z_{i}^{0} / /$ Copy embeddings", + "3: $z_{j}^{1'} = B^{0}\\left(z_{j}^{0'}\\right) // \\text{Preparation}$", + "4: for $i$ in 1 to $n - 1$ do", + "5: $B^{i} = \\text{Prune}(B^{i}, (s, \\alpha, \\beta))$", + "6: $z_{j}^{i} = B^{i}\\left(z_{j}^{i - 1}\\right) / /$ Forward Propagation", + "7: $z_{j}^{i + 1,i^{\\prime}} = B^{i + 1}\\left(z_{j}^{i^{\\prime}}\\right)$", + "8: $B^{i+1} = \\text{Update}\\left(B^{i+1}, z_{j}^{i}, z_{j}^{i+1'}\\right)$", + "9: end for", + "10: $B^n = \\text{Prune}(B^n, (s, \\alpha, \\beta))$", + "11: return Pruned backbone $\\{B^i\\}_{i=1}^n$" + ], + "bbox": [ + 91, + 448, + 482, + 650 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Sparsity Rearrangement", + "text_level": 1, + "bbox": [ + 89, + 681, + 316, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In many pruning methods, the sparsity of different parts of the model can be adaptively determined through specific rules. However, since our method is based on a sliding window approach for pruning, it is difficult to obtain information about the blocks that are outside the current window. If we attempt to acquire global information, we would either incur significant I/O overhead from loading model segments sequentially or face high memory usage from loading the entire model at once. Some methods even evaluate importance based on gradients, which can add further demands on gradient storage.", + "bbox": [ + 88, + 703, + 482, + 868 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To avoid these issues, fast pruning methods like SparseGPT or Wanda typically assign the same sparsity", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "level uniformly across all parts of the model, equivalent to the target sparsity. However, this approach is clearly suboptimal. Certain prior knowledge can be leveraged to adjust the sparsity distribution, thereby avoiding these problems while improving performance.", + "bbox": [ + 511, + 415, + 906, + 492 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1. Inter-Block Sparsity Rearrangement", + "text_level": 1, + "bbox": [ + 511, + 498, + 815, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Block-wise Compensate Pruning fine-tunes $B^{i+1}$ to compensate for the pruning-induced errors in $B^i$ . However, when pruning the last block, $B^n$ , there is no $B^{n+1}$ available to compensate for the errors in $B^n$ . Therefore, assigning the same number of pruned weights to each block is not optimal. Instead, the pruning distribution should be adjusted across different blocks. To maintain the global target sparsity, we rearrange the sparsity between blocks as follows:", + "bbox": [ + 511, + 516, + 906, + 638 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP _ {B ^ {n}} = \\left| \\theta^ {n} \\right| \\times s \\times \\alpha , \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 635, + 652, + 906, + 670 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP _ {B ^ {j}} = \\left| \\theta^ {j} \\right| \\times s + \\frac {\\left| \\theta^ {n} \\right| \\times s \\times (1 - \\alpha)}{n - 1} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 685, + 905, + 715 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $j = 1, \\ldots, n - 1$ , $\\theta^j$ denotes the parameters in $B^j$ , $|\\theta|$ denotes the number of weights in a single block, and $\\alpha$ is a super-parameter that redistributes a portion of the pruned weights from the last block evenly across all preceding blocks.", + "bbox": [ + 511, + 722, + 906, + 799 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2. Intra-Block Sparsity Rearrangement", + "text_level": 1, + "bbox": [ + 511, + 806, + 816, + 821 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig. 3(b), the majority of the parameters within a Transformer block are contributed by the matrices depicted, which are clearly hierarchical. Therefore, a block can essentially be viewed as a nested function: the QKV matrices represent the innermost function, while the second", + "bbox": [ + 511, + 824, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9490", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "fully connected layer (fc2) represents the outermost function. According to the conclusions drawn in the Theoretical Foundations section, if the same sparsity level is applied uniformly across all layers, the errors introduced by the QKV matrices, which are further from the output of $B^{i}$ , will be more challenging to compensate for in $B^{i + 1}$ . On the other hand, fc2 is closer to the output of $B^{i}$ , making its errors easier to compensate for in $B^{i + 1}$ . Additionally, inspired by previous works, earlier layers tend to converge first [1], and faster convergence often leads to better generalization performance [9]. Therefore, the sparsity distribution should be rearranged by reducing the sparsity of the QKV matrices and increasing the sparsity of fc2. To ensure that the global target sparsity remains unchanged, we rearrange the sparsity according to the following rule:", + "bbox": [ + 89, + 90, + 483, + 318 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP _ {t} = s \\times | W _ {t} ^ {i} | \\times (1 - \\beta), t \\in \\{Q, K, V \\}. (3)\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 325, + 482, + 345 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP _ {f c 2} = s \\times | W _ {f c 2} ^ {i} | + s \\times | W _ {f c 2} ^ {i} | \\times \\beta . \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 366, + 482, + 386 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $P_{t}$ represents the number of pruned weights, $|W|$ denotes the total number of weights in a single layer, and $\\beta$ is a super-parameter that needs to be adjusted.", + "bbox": [ + 89, + 388, + 483, + 434 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 446, + 223, + 463 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 470, + 281, + 488 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To evaluate the effectiveness of our method, we focus on models with 7 billion parameters or fewer, as these are the most commonly used in research and resource-constrained scenarios. Pruning these models is also the most feasible on a single desktop-class GPU. Specifically, we validate our approach using the OPT-125M/1.3B/2.7B/6.7B and Llama 2-7B, as well as the SAM-B/L/H. All experiments were conducted on a single NVIDIA RTX 3090 GPU.", + "bbox": [ + 89, + 493, + 483, + 613 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the compensation process, we employ the Adam optimizer with a Mean Squared Error (MSE) loss function and a batch size of 1. All compensation operations during pruning share a unified learning rate, which is adjusted using the ExponentialLR scheduler with a decay factor of 0.87.", + "bbox": [ + 89, + 614, + 483, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For calibration data, language model calibration dataset follows SparseGPT, using 128 segments of 2048 tokens, randomly sampled from C4 dataset [25]. Similarly, the vision model calibration dataset consists of 128 images, randomly sampled from the second compressed file of the SA-1B dataset [13] provided by Meta.", + "bbox": [ + 89, + 688, + 483, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To comprehensively evaluate the performance of pruned models, we validate OPT using the Wikitext-v2 [21] and PTB datasets [19] from Hugging Face, with perplexity (PPL) as the evaluation metric. We also evaluate the pruned Llama 2-7B model using accuracy on seven ZeroShot tasks: BoolQ [2], RTE [32], HellaSwag [35], Wino-Grande [26], ARC Easy and Challenge [3], and Open-bookQA [22], following the experimental setup introduced", + "bbox": [ + 89, + 780, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "by Sun et al. [27] in their work on Wanda. This allows for a more task-specific evaluation of the pruned model's generalization capabilities. For SAM, we validate the performance using the fourth compressed file of the SA-1B dataset as well as the COCO dataset [15], employing Intersection over Union (IoU) as the primary metric for visual models.", + "bbox": [ + 511, + 90, + 905, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Unstructured Pruning", + "text_level": 1, + "bbox": [ + 511, + 208, + 720, + 224 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Language Models. In Table 1 and Table 2, we report the perplexity performance of the OPT series models on Wikitext and PTB datasets after pruning to various sparsity levels using our method and other baseline approaches. Our method demonstrates superior performance over baselines even at lower sparsity levels, and as sparsity increases, the performance gap widens in favor of our approach. Additionally, as model size decreases, the advantage of our method over the baselines becomes even more pronounced. Table 3 also presents results in comparing the performances of our proposed method against those of baselines on the Llama 2-7B model at $50\\%$ and $70\\%$ sparsity levels. Our method consistently outperforms all baseline approaches at both sparsity levels, achieving the best results across all tasks. These findings highlight the effectiveness of the interblock error compensation pruning method introduced in this paper, particularly under high-sparsity conditions, across different model scales, and for various model architectures.", + "bbox": [ + 511, + 229, + 906, + 501 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/ab1c41eba5571e697d2f7a49dd02485d9dbe945252a052f550921f5e0adb5ae4.jpg", + "table_caption": [ + "Table 1. Perplexity performances of pruned OPT models of different scales at various sparsity levels. The models are calibrated on a subset of the C4 and evaluated on full Wikitext-v2." + ], + "table_footnote": [], + "table_body": "
MethodModel10%20%30%40%50%60%70%80%90%
Magnitude125M28.2529.6534.5154.60193.4920.0380648906614
SparseGPT27.8827.9828.8630.5637.0259.55220.923784666
Wanda27.4827.7628.1130.6738.9275.17328.219184609
ICP (ours)27.3727.8728.6629.6633.9045.7865.20270.41612
Magnitude1.3B14.7215.6224.74388.01713939294431634428871
SparseGPT14.6714.7515.1616.4917.5022.0851.75752.46797
Wanda14.6314.6915.0115.8918.4126.5599.53225816868
ICP (ours)14.5914.6014.6715.0416.4921.1736.78117.11099
Magnitude2.7B12.5913.1315.5830.32265.236047251961416668
SparseGPT12.3212.3912.6612.6613.4616.0426.92138.45818
Wanda12.2312.2612.3612.8614.2119.83387.9590516527
ICP (ours)12.2212.2512.3412.6213.3215.7721.7954.67764.2
Magnitude6.7B10.9211.2712.5331.89968.71263916975255915297
SparseGPT10.8310.7610.7510.9611.5913.4820.4896.2511938
Wanda10.7310.6310.6410.9611.9815.19159.2395416076
ICP (ours)10.5610.5810.5910.8511.2713.0317.8244.691469
", + "bbox": [ + 522, + 556, + 897, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Vision Models. In Tables 4 and Table 5, we report IoU performance of SAM at varying sparsity levels on SA-1B and COCO datasets with single-point interaction. Following experimental setup, SAMs are pruned with only 128 images random sampled from SA-1B as calibration data. The experiments first verify that SparseGPT indeed achieves strong IoU performance on visual models under low sparsity conditions, outperforming both Magnitude and Wanda.", + "bbox": [ + 511, + 779, + 906, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9491", + "bbox": [ + 480, + 944, + 513, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/6ecfa2ae1e2ace5d16c443e11f68593c8de8abbc80400a8fe2cd06158bf558cd.jpg", + "table_caption": [ + "Table 2. Perplexity performancese of pruned OPT models of different scales at various sparsity levels. The models are calibrated on a subset of the C4 and evaluated on PTB." + ], + "table_footnote": [], + "table_body": "
MethodModel10%20%30%40%50%60%70%80%90%
Magnitude125M39.9743.2549.9782.69276.21146343041245382
SparseGPT39.4340.3642.7145.2455.7888.44259.629195374
Wanda39.0539.8341.4344.4157.59109.7399.722144474
ICP (ours)39.1239.1340.9643.4749.4164.1088.85336.33075
Magnitude1.3B20.6621.5139.89823.53171949984671728333126
SparseGPT20.3420.6421.5023.1725.4733.4876.91531.87515
Wanda20.3820.7021.4123.2528.0043.95135.3180912025
ICP (ours)20.1320.4121.1122.1924.6832.1954.33148.83094
Magnitude2.7B18.3819.0622.9646.36262.632077276957512178
SparseGPT18.1518.5018.9818.9820.4424.6742.75146.06107
Wanda17.9718.0418.3219.2521.8633.44450.8546221039
ICP (ours)17.9718.0218.2918.8620.0522.9235.1771.912357
Magnitude6.7B16.1816.7018.5248.85613.493799973157424260
SparseGPT15.7715.8115.9216.2817.4320.2331.86105.08320
Wanda15.7615.7715.9416.4317.9223.61211.3302414998
ICP (ours)15.7615.7415.9416.2217.3020.0728.0464.182549
", + "bbox": [ + 102, + 130, + 470, + 335 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e7ae43f33179b383dc80c8196b60e38393a7495d892ce2067a11fe5010963a94.jpg", + "table_caption": [ + "Table 3. Zero-shot performance (Accuracy, %) of Llama 2-7B pruned by different method on various tasks at sparsity of $50\\%$ and $70\\%$ ." + ], + "table_footnote": [], + "table_body": "
MethodStrategyBQRTEHSWGARC-eARC-cOBQAMean
Dense077.7463.1857.1068.9876.2643.4331.4059.73
Magnitude50%62.9457.0449.1263.3864.0634.6426.8051.14
SparseGPT76.3355.9652.8968.9872.0138.2328.4056.11
Wanda76.4253.4352.4568.6772.2239.3331.0056.22
ICP (ours)76.5057.0453.0369.1172.3940.4431.0056.96
Magnitude70%37.9553.0725.9349.2527.8222.8717.0033.41
SparseGPT64.6553.7933.4558.6443.3121.6717.0041.79
Wanda48.4752.7127.9749.6430.8118.6012.2034.34
ICP (ours)67.7155.2341.4560.2255.5627.0521.4046.94
", + "bbox": [ + 102, + 397, + 470, + 525 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As Wanda is based on strong priors from language models, its underperformance compared to SparseGPT on visual models is within our expectations. Lastly, the method proposed in this paper outperforms all baselines and exhibits a similar trend observed in language model experiments: the higher the sparsity, the greater the advantage of our method over the baseline methods. This demonstrates that the proposed method is not only effective for language models but also remains effective for visual models.", + "bbox": [ + 91, + 556, + 480, + 691 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Semi-Structure Pruning and Quantization", + "text_level": 1, + "bbox": [ + 94, + 709, + 447, + 726 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "On certain specialized hardware, such as specific NVIDIA GPUs or other custom-designed computing architectures, the combination of pruning and quantization can further accelerate inference while maintaining an acceptable performance degradation. To explore this, we conduct tests using a joint compression strategy that combines pruning and quantization. Specifically, we focus on the combination of 2:4 and 4:8 semi-structured sparsity with 3-bit and 4-bit quantization. Since Wanda does not support quantization, we exclude it from experiments involving quantization and instead compared our approach against SparseGPT.", + "bbox": [ + 91, + 734, + 480, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7114bf472e365b28c1b5a1a7722133566c8c0d865df8b664b13df96ddad25fd8.jpg", + "table_caption": [ + "Table 4. Instance segmentation performance of SAM models (IoU, $\\%$ ) at various sparsity using single-point interaction on SA-1B. The models are calibrated on a subset of 128 images sampled from SA-1B." + ], + "table_footnote": [], + "table_body": "
MethodModel10%20%30%40%50%60%70%80%90%
MagnitudeSAM-H74.8074.5773.8772.2669.4561.5128.222.651.70
SparseGPT74.8174.7774.6074.2073.3071.1565.6950.9718.15
Wanda74.8274.7674.3573.6071.6567.6553.0021.841.20
ICP (ours)74.8274.7874.6674.6074.4273.8272.8569.7357.60
MagnitudeSAM-L74.8174.5973.9372.2869.1862.5645.529.924.52
SparseGPT74.8474.7774.6174.1373.2070.7064.4545.487.37
Wanda74.8674.7874.6274.0572.5168.2855.1322.811.31
ICP (ours)74.8574.7974.6974.5574.2773.6472.2668.1753.68
MagnitudeSAM-B72.6972.5671.9970.3667.3860.8746.0125.531.51
SparseGPT72.7272.6772.5072.0871.1168.6863.0247.4918.81
Wanda72.7172.6472.3371.4569.4464.3550.2928.162.18
ICP (ours)72.7272.6972.5672.2871.8571.0669.3364.7851.82
", + "bbox": [ + 526, + 145, + 893, + 303 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/00239242c6a99d2bae0e16b3f4df71be9c71c49cd0639a01cbca81d3eac89d9e.jpg", + "table_caption": [ + "Table 5. Instance segmentation performance of SAM models (IoU, %) at various sparsity using single-point interaction on COCO." + ], + "table_footnote": [], + "table_body": "
MethodModel10%20%30%40%50%60%70%80%90%
MagnitudeSAM-H71.1271.1170.8670.6169.3763.6032.687.887.21
SparseGPT71.1171.1870.9870.9370.4269.5266.7056.8227.57
Wanda71.1771.0170.9870.7369.6366.0553.4522.835.51
ICP (ours)71.1471.1871.1270.9870.8170.2469.4366.6958.41
MagnitudeSAM-L70.5670.5170.2769.5368.0862.6649.1220.6618.12
SparseGPT70.5570.5370.4370.0469.6168.2164.4248.3511.30
Wanda70.5670.5870.4569.9168.6565.5654.6719.896.38
ICP (ours)70.5570.5870.5870.3769.9169.5168.1465.4256.92
MagnitudeSAM-B66.8467.0667.0666.6765.2962.2853.6140.867.48
SparseGPT67.0967.0266.9866.7666.1765.0062.0352.4628.51
Wanda66.9466.9767.0466.4765.3862.3153.5839.327.89
ICP (ours)67.1067.0667.0466.8366.4566.0764.5162.4054.50
", + "bbox": [ + 526, + 347, + 893, + 503 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c9af1e2b1d3edb33b9cdae473daf9e0cbacee1d7d039fba8c0ead6c119ab2cbc.jpg", + "table_caption": [ + "Table 6. Perplexity performance of pruned OPT models at various compression strategies." + ], + "table_footnote": [], + "table_body": "
MethodStrategyOPT-125MOPT-1.3BOPT-2.7BOPT-6.7B
WTPTBWTPTBWTPTBWTPTB
SparseGPT2:459.3593.0423.8738.0917.1126.9814.1521.54
Wanda2:480.01111.9728.2043.3721.1834.5515.9025.09
ICP (ours)2:446.9067.5622.0333.9516.3925.3114.0421.19
SparseGPT4:843.9071.9920.2131.3615.0023.0512.4918.86
Wanda4:853.1879.0222.1734.5616.7926.1313.5520.19
ICP (ours)4:839.9159.3419.1530.4114.8822.8812.3818.53
SparseGPT2:4+int3122.33196.4440.5864.5324.7041.0620.0731.91
ICP (ours)2:4+int374.07114.9327.8846.1419.5730.2818.3826.82
SparseGPT2:4+int470.88107.3925.9641.6918.5630.6815.2423.01
ICP (ours)2:4+int451.2475.3923.4934.4217.0527.0415.1022.54
SparseGPT4:8+int384.02120.1632.4149.2720.2432.9917.0826.36
ICP (ours)4:8+int359.6786.0124.7340.1917.8926.8016.1624.30
SparseGPT4:8+int450.5880.9723.3234.3815.7424.7613.4419.72
ICP (ours)4:8+int442.0360.6020.8631.5815.3323.8013.2419.49
", + "bbox": [ + 535, + 547, + 885, + 761 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Sparsity Rearrangement Ablation", + "text_level": 1, + "bbox": [ + 516, + 787, + 808, + 803 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To validate the Sparsity Rearrangement method based on priors, as proposed in the Method section, we conducted experiments on both Inter-Block Sparsity Rearrangement and Intra-Block Sparsity Rearrangement.", + "bbox": [ + 516, + 809, + 903, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Inter-Block Sparsity Rearrangement. In Fig. 4a and Fig. 4b, we present the performance of OPT-125M and", + "bbox": [ + 516, + 869, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9492", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a3a46e55e2963bef15a2916ef7248ebb6e7a48b0bb6e21b9249399ab1936ae2c.jpg", + "table_caption": [ + "Table 7. Zero-shot Performance (Accuracy, %) of Llama 2-7B Model under various compression strategies." + ], + "table_footnote": [], + "table_body": "
MethodStrategyBQRTEHSWGARC-eARC-cOBQAMean
SparseGPT2:468.0758.8443.3865.5964.1831.9124.8050.97
Wanda2:468.1353.4341.4262.4363.1330.6323.8049.00
ICP (ours)2:469.7258.8446.7564.4866.0433.7026.2052.25
SparseGPT4:871.0754.8748.3567.8068.6934.4727.2053.21
Wanda4:873.0053.7947.0366.9367.4734.2227.0052.78
ICP (ours)4:873.1258.1250.1364.8868.9036.0128.0054.16
SparseGPT2:4+int364.6555.2338.5859.0455.4725.3419.4045.39
ICP (ours)2:4+int366.8254.1543.3660.7759.5528.6721.0047.76
SparseGPT2:4+int466.7354.1542.3465.1961.7428.9223.8048.98
ICP (ours)2:4+int470.4656.6846.2262.9064.8130.6322.8050.64
SparseGPT4:8+int368.5953.4342.1262.3560.1029.1021.6048.18
ICP (ours)4:8+int368.3854.8743.5462.8563.3832.3422.4049.68
SparseGPT4:8+int469.2754.1546.5866.0667.5534.3927.2052.17
ICP (ours)4:8+int472.2961.0148.8662.7566.8833.8728.0053.38
", + "bbox": [ + 102, + 117, + 470, + 315 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/37964a2f8e98a6a70d693ac2aa98bd2654c2b2888b58ef6e03f0a4406777c662.jpg", + "table_caption": [ + "Table 8. Instance segmentation performance of SAM models (IoU, $\\%$ ) at various compression strategies using single-point interaction on SA-1B and COCO." + ], + "table_footnote": [], + "table_body": "
MethodStrategySAM-BSAM-LSAM-H
COCOSA-1BCOCOSA-1BCOCOSA-1B
SparseGPT2:462.8067.5965.0169.0767.0269.62
Wanda2:460.5962.0662.3765.8464.6866.36
ICP (ours)2:464.9571.3067.8373.5669.0973.64
SparseGPT4:863.9369.6966.5671.2767.9871.41
Wanda4:862.6266.0565.0269.7566.8069.55
ICP (ours)4:865.3771.7268.1773.8669.4474.03
SparseGPT2:4+int362.0964.7763.4565.6766.3967.85
ICP (ours)2:4+int364.4769.7167.1372.6068.7573.26
SparseGPT2:4+int462.6666.7664.7868.6966.8169.44
ICP (ours)2:4+int464.8970.6467.6473.0969.0273.72
SparseGPT4:8+int363.0267.2764.7468.0267.2569.69
ICP (ours)4:8+int364.8469.8667.6873.2869.1073.35
SparseGPT4:8+int463.7768.9866.0470.3167.8771.05
ICP (ours)4:8+int465.2470.9868.0273.8269.3473.78
", + "bbox": [ + 120, + 372, + 454, + 585 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "SAM-B pruned by ICP with different $\\alpha$ . For clarity in visual presentation, we conducted experiments with $70\\%$ sparsity on OPT-125M and $90\\%$ sparsity on SAM-B. We observed that in the language model, $\\alpha$ had the greatest impact on the performance of OPT-125M on the PTB dataset, and adjusting $\\alpha$ further improved the model's performance on non-calibration datasets. On vision models, the influence of $\\alpha$ on performance was similar on both datasets, with adjustments to $\\alpha$ enhancing the overall performance of the visual model. Overall, appropriate tuning of $\\alpha$ can further improve the performance of pruned models. Based on experiments, we conclude that $\\alpha$ typically yields optimal results within the range of [0.6, 0.75] for language models, and within [0.8, 0.9] for vision model.", + "bbox": [ + 89, + 612, + 482, + 823 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Intra-Block Sparsity Rearrangement. In Fig. 4c, we illustrate the impact of different $\\beta$ on the language model, using the $70\\%$ sparsity model of OPT-125M as an example. We observed that in the language model, adjusting $\\beta$ can further reduce the model's perplexity. Unlike $\\alpha$ , $\\beta$ not", + "bbox": [ + 89, + 825, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "only improves the model's performance on non-calibration datasets but also enhances its performance on calibration datasets. Overall, appropriately adjusting $\\beta$ can further improve the performance of the pruned model. The optimal value for $\\beta$ typically falls within the range of $[0.6 \\times (1 - s), 0.8 \\times (1 - s)]$ , where $s$ denotes sparsity. In Fig. 4d, we exemplify the effects of varying $\\beta$ on the visual model, specifically using the $90\\%$ sparsity model of SAM-B. Our findings indicate that the influence of $\\beta$ on the visual model is minimal; both on calibration and non-calibration datasets, adjustments to $\\beta$ yield only marginal improvements in the model's performance.", + "bbox": [ + 511, + 90, + 906, + 272 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Calibration Set Size Ablation", + "text_level": 1, + "bbox": [ + 511, + 282, + 774, + 297 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Although our experiments follow the settings of SparseGPT and Wanda, using 128 randomly selected samples for pruning both language and vision models, Fig. 4e and Fig. 4f show the impact of calibration set size on our method's performance. Fig. 1 presents results on the OPT-125M language model at $70\\%$ sparsity, while Fig. 2 shows results for the SAM-B vision model at $90\\%$ sparsity. In both cases, increasing the sample size beyond 128 offers minimal performance gains while significantly increasing GPU memory usage. Thus, like SparseGPT and Wanda, 128 samples remain the optimal choice.", + "bbox": [ + 511, + 305, + 906, + 472 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.6. Compensation Epoch Ablation", + "text_level": 1, + "bbox": [ + 511, + 482, + 785, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Fig.4g and Fig.4h, we illustrate the impact of varying the number of epochs during the compensation on the performance of both vision and language models. It is evident that a single epoch is sufficient to effectively mitigate the error introduced by pruning in earlier layers. While additional epochs can further improve performance, the marginal gains diminish rapidly. Although we report results using 12 epochs in the experiments presented in 4.2 and 4.3 to demonstrate the best achievable performance, selecting a single epoch offers the optimal trade-off between model performance and computational overhead.", + "bbox": [ + 511, + 505, + 908, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.7. Memory and Time Efficiency", + "text_level": 1, + "bbox": [ + 511, + 681, + 772, + 698 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Peak memory usage (PMU) and execution time are critical factors in determining whether a pruning method can be deployed on devices with limited memory capacity and lower cost. In Table 9, we present the PMU and runtime of SparseGPT, Wanda, and our proposed method when pruning OPT-6.7B on an NVIDIA RTX 3090. The experimental setup is consistent with the description in 4.1. Our method employs a commonly used chunked storage strategy, where intermediate outputs of each layer are stored in host memory instead of GPU memory, effectively reducing memory usage. While this approach may slow down the pruning process, as discussed in Section 4.6, a single epoch of fine-tuning is sufficient to significantly improve", + "bbox": [ + 511, + 704, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9493", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cc8dfd7b1ba0b437667dde1e01afc20843dd57529ecb60469c0b5d7fbef34424.jpg", + "image_caption": [ + "(a) Perplexity performance of pruned OPT-125M with different $\\alpha$" + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 354, + 208 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a9cc02cc1bfd6a93eb4d02ce06f676f906372b78ba0d01ecab5ceefbd16fb5f1.jpg", + "image_caption": [ + "(b) Instance segmentation performance (IoU, %) of SAM-B with different $\\alpha$ ." + ], + "image_footnote": [], + "bbox": [ + 366, + 88, + 630, + 208 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bf9c50dd352ccf6218444dfd3632170d98261fe55abf943d3d2d3ab663dfb6ac.jpg", + "image_caption": [ + "(c) Perplexity performance of OPT-125M with different $\\beta$" + ], + "image_footnote": [], + "bbox": [ + 643, + 88, + 903, + 209 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7a96db60711f9674db4441172da3117c8ce3d675ec2542ec1fbd3e190815335e.jpg", + "image_caption": [ + "(d) Instance segmentation performance (IoU, %) of SAM-B with different $\\beta$ ." + ], + "image_footnote": [], + "bbox": [ + 91, + 241, + 354, + 361 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/25bbdfb3d8e95c75b128bb3808b9ee70688a98c43dc6d9f455e0feb7882a7174.jpg", + "image_caption": [ + "(e) Perplexity performance of OPT-125M with different calibration set size." + ], + "image_footnote": [], + "bbox": [ + 367, + 241, + 630, + 361 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/80b480ac31d38361ce2a816ab553da4a67ef8bb15b7edcbfe762295bca57ebce.jpg", + "image_caption": [ + "(f) Instance segmentation performance (IoU, %) of SAM-B with different calibration set size." + ], + "image_footnote": [], + "bbox": [ + 643, + 241, + 905, + 362 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/797e9a14b85ec63bf6edc8cd84fc204df5ae975d440c44637518be3aad9d6e4c.jpg", + "image_caption": [ + "(g) Perplexity performance of pruned OPT-125M with different number of compensation epochs." + ], + "image_footnote": [], + "bbox": [ + 194, + 393, + 455, + 515 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a30956c31be2c5da3c3603cb57eef34222fb9fc886497ad873d9e7b4d78e779d.jpg", + "image_caption": [ + "(h) Segmentation performance of SAM-B with different number of compensation epochs.", + "Figure 4. Performance metrics for pruned OPT-125M and SAM-B models under varying parameters $\\alpha$ , $\\beta$ , calibration dataset sizes, and number of compensation epochs." + ], + "image_footnote": [], + "bbox": [ + 540, + 393, + 805, + 515 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "model performance. Consequently, data only needs to be loaded from host memory once, keeping the additional time overhead within an acceptable range. It is worth noting that we use full-precision fine-tuning rather than Parameter-Efficient Fine-Tuning (PEFT) techniques. If combined with PEFT methods, memory consumption could be further reduced. This advantage enables our method to efficiently prune larger models on desktop-class GPUs, thereby expanding its applicability.", + "bbox": [ + 89, + 609, + 483, + 744 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/3201cc154090a50cfb2082a3d9885b2346923b7c100964d4a2aabc1ea10dc5f0.jpg", + "table_caption": [ + "Table 9. GPU Memory Consumption Comparison" + ], + "table_footnote": [], + "table_body": "
MethodPMUTime(s)
SparseGPT8.3 GB1004
Wanda21 GB143
ICP (1-epoch)7.8 GB1551
ICP (12-epoch)7.8 GB4470
", + "bbox": [ + 161, + 789, + 413, + 875 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 607, + 633, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose an efficient pruning method ICP, which is effective for both vision and language models while consuming limited GPU memory. ICP update each block individually by compensating for the inter-block errors introduced by pruning, then rearranges the sparsity distribution based on prior knowledge of Transformer and pruning method. It significantly improves the performance of models pruned at high sparsity levels without increasing GPU memory consumption. ICP, as it does not rely on strong priors specific to any particular model, demonstrates dual applicability to both vision and language models. This flexibility suggests that it could potentially be extended to multimodal models in the future. However, this work has limitations: (1) the hyperparameters $\\alpha$ and $\\beta$ are not automatically tuned; (2) ICP has not been tested on multimodal models; (3) the potential of individual blocks has not been fully explored. These will be the focus of our future work.", + "bbox": [ + 511, + 643, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9494", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is supported by Natural Science Foundation of China under Grant 62271465, Suzhou Basic Research Program under Grant SYG202338, and Jiangsu Province Science Foundation for Youths (NO. BK20240464).", + "bbox": [ + 89, + 113, + 483, + 170 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 196, + 187, + 213 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yixiong Chen, Alan Yuille, and Zongwei Zhou. Which layer is learning faster? a systematic exploration of layerwise convergence rate for deep neural networks. In The Eleventh International Conference on Learning Representations, 2023. 5", + "[2] Christopher Clark, Kenton Lee, Ming-Wei Chang, Tom Kwiatkowski, Michael Collins, and Kristina Toutanova. Boolq: Exploring the surprising difficulty of natural yes/no questions. arXiv preprint arXiv:1905.10044, 2019. 5", + "[3] Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. Think you have solved question answering? try arc, the ai2 reasoning challenge. arXiv preprint arXiv:1803.05457, 2018. 5", + "[4] Peijie Dong, Lujun Li, Zhenheng Tang, Xiang Liu, Xinglin Pan, Qiang Wang, and Xiaowen Chu. Pruner-zero: Evolving symbolic pruning metric from scratch for large language models. arXiv preprint arXiv:2406.02924, 2024. 1", + "[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1", + "[6] Jonathan Frankle and Michael Carbin. The lottery ticket hypothesis: Finding sparse, trainable neural networks. In International Conference on Learning Representations, 2018. 1", + "[7] Elias Frantar and Dan Alistarh. Sparsegpt: Massive language models can be accurately pruned in one-shot. In International Conference on Machine Learning, pages 10323-10337. PMLR, 2023. 1, 3", + "[8] Song Han, Jeff Pool, John Tran, and William Dally. Learning both weights and connections for efficient neural network. Advances in neural information processing systems, 28, 2015. 1", + "[9] Moritz Hardt, Ben Recht, and Yoram Singer. Train faster, generalize better: Stability of stochastic gradient descent. In International conference on machine learning, pages 1225-1234. PMLR, 2016. 5", + "[10] Babak Hassibi and David Stork. Second order derivatives for network pruning: Optimal brain surgeon. In Advances in Neural Information Processing Systems. Morgan-Kaufmann, 1992. 3", + "[11] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 1" + ], + "bbox": [ + 94, + 223, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Lei Ke, Mingqiao Ye, Martin Danelljan, Yu-Wing Tai, Chi-Keung Tang, Fisher Yu, et al. Segment anything in high quality. Advances in Neural Information Processing Systems, 36, 2024. 1", + "[13] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 1, 2, 5", + "[14] Chunyuan Li, Cliff Wong, Sheng Zhang, Naoto Usuyama, Haotian Liu, Jianwei Yang, Tristan Naumann, Hoifung Poon, and Jianfeng Gao. Llava-med: Training a large language-and-vision assistant for biomedicine in one day. Advances in Neural Information Processing Systems, 36, 2024. 1", + "[15] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision - ECCV 2014, pages 740-755, Cham, 2014. Springer International Publishing. 5", + "[16] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1", + "[17] AI@Meta Llama Team. The llama3 herd of models, 2024. Accessed: 2024-08-01. 1", + "[18] Xinyin Ma, Gongfan Fang, and Xinchao Wang. Llm-pruner: On the structural pruning of large language models. Advances in neural information processing systems, 36:21702-21720, 2023. 1", + "[19] Mitch Marcus, Grace Kim, Mary Ann Marcinkiewicz, Robert MacIntyre, Ann Bies, Mark Ferguson, Karen Katz, and Britta Schasberger. The penn treebank: Annotating predicate argument structure. In Human Language Technology: Proceedings of a Workshop held at Plainsboro, New Jersey, March 8-11, 1994, 1994. 5", + "[20] Maciej A Mazurowski, Haoyu Dong, Hanxue Gu, Jichen Yang, Nicholas Konz, and Yixin Zhang. Segment anything model for medical image analysis: an experimental study. Medical Image Analysis, 89:102918, 2023. 2", + "[21] Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. Pointer sentinel mixture models. arXiv preprint arXiv:1609.07843, 2016. 5", + "[22] Todor Mihaylov, Peter Clark, Tushar Khot, and Ashish Sabharwal. Can a suit of armor conduct electricity? a new dataset for open book question answering. arXiv preprint arXiv:1809.02789, 2018. 5", + "[23] Quan Quan, Fenghe Tang, Zikang Xu, Heqin Zhu, and S Kevin Zhou. Slide-sam: Medical sam meets sliding window. In Medical Imaging with Deep Learning, 2024. 2", + "[24] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018. 2", + "[25] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 5" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9495", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Keisuke Sakaguchi, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. Winogrande: An adversarial winograd schema challenge at scale. Communications of the ACM, 64 (9):99-106, 2021. 5", + "[27] Mingjie Sun, Zhuang Liu, Anna Bair, and J Zico Kolter. A simple and effective pruning approach for large language models. In The Twelfth International Conference on Learning Representations, 2024. 1, 3, 5", + "[28] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1", + "[29] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 1", + "[30] Tycho FA van der Ouderaa, Markus Nagel, Mart Van Baalen, Yuki M Asano, and Tijmen Blankevoort. The IIm surgeon. arXiv preprint arXiv:2312.17244, 2023. 1", + "[31] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 1, 2", + "[32] Alex Wang. Glue: A multi-task benchmark and analysis platform for natural language understanding. arXiv preprint arXiv:1804.07461, 2018. 5", + "[33] Miles Williams and Nikolaos Aletras. How does calibration data affect the post-training pruning and quantization of large language models? arXiv preprint arXiv:2311.09755, 2023. 3", + "[34] Junde Wu, Wei Ji, Yuanpei Liu, Huazhu Fu, Min Xu, Yanwu Xu, and Yueming Jin. Medical sam adapter: Adapting segment anything model for medical image segmentation. arXiv preprint arXiv:2304.12620, 2023. 1, 2", + "[35] Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. Hellaswag: Can a machine really finish your sentence? arXiv preprint arXiv:1905.07830, 2019. 5", + "[36] Renrui Zhang, Jiaming Han, Chris Liu, Peng Gao, Aojun Zhou, Xiangfei Hu, Shilin Yan, Pan Lu, Hongsheng Li, and Yu Qiao. Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199, 2023. 1", + "[37] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 1" + ], + "bbox": [ + 91, + 92, + 480, + 782 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9496", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/ICP_ Immediate Compensation Pruning for Mid-to-high Sparsity/9ac0c560-9a5b-4372-9074-16698b23f1e1_model.json b/2025/ICP_ Immediate Compensation Pruning for Mid-to-high Sparsity/9ac0c560-9a5b-4372-9074-16698b23f1e1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1e1094d4dd482a14814d404594f494a833f6ea3a --- /dev/null +++ b/2025/ICP_ Immediate Compensation Pruning for Mid-to-high Sparsity/9ac0c560-9a5b-4372-9074-16698b23f1e1_model.json @@ -0,0 +1,2167 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.172, + 0.131, + 0.833, + 0.154 + ], + "angle": 0, + "content": "ICP: Immediate Compensation Pruning for Mid-to-high Sparsity" + }, + { + "type": "text", + "bbox": [ + 0.194, + 0.179, + 0.807, + 0.199 + ], + "angle": 0, + "content": "Xin Luo\\(^{1,2}\\) Xueming Fu\\(^{1,2}\\) Zihang Jiang\\(^{1,2*}\\) S.Kevin Zhou\\(^{1,2,3,4*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.165, + 0.199, + 0.834, + 0.216 + ], + "angle": 0, + "content": "1 School of Biomedical Engineering, Division of Life Sciences and Medicine, USTC" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.216, + 0.794, + 0.233 + ], + "angle": 0, + "content": "2 MIRACLE Center, Suzhou Institute for Advance Research, USTC" + }, + { + "type": "text", + "bbox": [ + 0.207, + 0.233, + 0.792, + 0.252 + ], + "angle": 0, + "content": "3 Key Laboratory of Intelligent Information Processing of CAS, ICT, CAS" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.252, + 0.773, + 0.27 + ], + "angle": 0, + "content": "\\(^{4}\\) State Key Laboratory of Precision and Intelligent Chemistry, USTC" + }, + { + "type": "text", + "bbox": [ + 0.28, + 0.272, + 0.718, + 0.288 + ], + "angle": 0, + "content": "{luoxin.ustc, xuemingfu, jzh0103}@mail.ustc.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.394, + 0.29, + 0.605, + 0.305 + ], + "angle": 0, + "content": "{s.kevin.zhou}@gmail.com" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.34, + 0.327, + 0.356 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.371, + 0.485, + 0.675 + ], + "angle": 0, + "content": "The increasing adoption of large-scale models under 7 billion parameters in both language and vision domains enables inference tasks on a single consumer-grade GPU but makes fine-tuning models of this scale, especially 7B models, challenging. This limits the applicability of pruning methods that require full fine-tuning. Meanwhile, pruning methods that do not require fine-tuning perform well at low sparsity levels (10%-50%) but struggle at mid-to-high sparsity levels (50%-70%), where the error behaves equivalently to that of semi-structured pruning. To address these issues, this paper introduces ICP, which finds a balance between full fine-tuning and zero fine-tuning. First, Sparsity Rearrange is used to reorganize the predefined sparsity levels, followed by Block-wise Compensate Pruning, which alternates pruning and compensation on the model's backbone, fully utilizing inference results while avoiding full model fine-tuning. Experiments show that ICP improves performance at mid-to-high sparsity levels compared to baselines, with only a slight increase in pruning time and no additional peak memory overhead." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.702, + 0.222, + 0.718 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.727, + 0.483, + 0.879 + ], + "angle": 0, + "content": "Introduction of Transformer [5, 31] propelled both language and vision models toward larger scales, leading to the development of high-performance models such as the OPT [37] and Llama [17, 28, 29] series for language tasks, and the SAM [13] series for vision tasks. However, their large scale presents significant deployment challenges. In scenarios with downstream tasks or hardware constraints, models under 8 billion (8B) parameters for language tasks and under 650 million parameters for vision tasks are most commonly used [12, 14, 16, 34, 36]. Nevertheless, model compres" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.341, + 0.905, + 0.401 + ], + "angle": 0, + "content": "sion is necessary to further reduce their size for deployment. Moreover, on a single desktop-class GPU, it is typically only feasible to perform inference with these models, while fine-tuning remains impractical due to hardware limitations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.402, + 0.907, + 0.869 + ], + "angle": 0, + "content": "Pruning is a critical model compression technique, often categorized into multi-shot and one-shot methods based on pruning frequency. Traditional iterative pruning methods, based on the Lottery Ticket Hypothesis [6], involve an iterative prune-and-finetune process to gradually identify a sparse subnetwork that performs comparably to the original model, thereby reducing the number of model parameters. Classic methods like magnitude pruning [8] and more recent approaches such as LLM-Pruner [18], which utilizes block pruning with LoRA [11] fine-tuning, adhere to this hypothesis. However, for large models, global fine-tuning is resource-intensive, demanding significant GPU memory. Although LLM-Pruner reduces memory requirements using LoRA, it still requires storage for intermediate results and gradients, while methods like LLM Surgeon [30], despite avoiding global fine-tuning, also demand substantial memory and time due to the use of second-order derivatives. In contrast, one-shot pruning methods have gained traction by minimizing memory usage during fine-tuning. SparseGPT [7] avoids global fine-tuning by compensating for precision within each basic layer, thus accelerating the pruning process. Wanda [27] further simplifies this by eliminating compensation altogether, using a unique significance score to achieve performance comparable to SparseGPT without the need for fine-tuning. PruneZero[4] constructs an operator search space by combining input, output, and gradient operations to optimize model metrics. While, it requires multiple full inferences across models, which is computationally expensive. In memory-constrained settings, chunked model loading further exacerbates the inference time." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "However, both SparseGPT and Wanda have notable limitations. (i) While these methods perform well at lower" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.236, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9487" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.198 + ], + "angle": 0, + "content": "sparsity levels (10%-50%) when pruning large language models, their effectiveness significantly declines at mid-to-high sparsity levels (50%-70%). (ii) SparseGPT and Wanda show poor performance on vision models. Wanda, in particular, relies on priors that are specific to language models, making it even less effective than SparseGPT for vision tasks." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.207, + 0.287, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.147, + 0.327, + 0.227, + 0.339 + ], + "angle": 0, + "content": "(a) SparseGPT" + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.208, + 0.484, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.36, + 0.327, + 0.416, + 0.338 + ], + "angle": 0, + "content": "(b) Wanda" + }, + { + "type": "image_caption", + "bbox": [ + 0.121, + 0.351, + 0.452, + 0.366 + ], + "angle": 0, + "content": "Figure 1. Error propagation with respect to layer index." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.385, + 0.483, + 0.627 + ], + "angle": 0, + "content": "To illustrate why we focus on mid-to-high sparsity levels, we present the error accumulation during the pruning process of OPT-1.3B using SparseGPT and Wanda in Fig.1a and Fig.1b, with errors measured by mean square error (MSE). As shown, both SparseGPT and Wanda prune blocks sequentially without inter-block interaction, leading to faster error accumulation in deeper layers, which aligns with intuitive expectations. Furthermore, the error accumulation trends of semistructured sparsity 2:4 and 4:8 differ from those of unstructured sparsity \\(50\\%\\), instead aligning more closely with higher sparsity levels. Fig.1a, indicates that incorporating quantization strategies further shifts the error trends toward those of higher sparsity levels. Therefore, mitigating inter-block error propagation at sparsity levels of \\(50\\% -70\\%\\) can effectively enhance the performance of unstructured pruning and hybrid compression strategies." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.627, + 0.483, + 0.852 + ], + "angle": 0, + "content": "In this paper, we propose ICP, a large model pruning method that is applicable to both vision and language models. Improve performance at both low- and high-sparsity levels in pruning large language and vision models, without increasing GPU memory usage and pruning time. Specifically, we design a sliding window-based inter-block error compensation method based on the principle of error propagation. As shown in Fig. 2, after pruning each block, the subsequent block compensates for the pruning-induced errors of the previous block before pruning itself. This approach ensures that errors introduced during pruning are compensated for in the next block, preventing error propagation to deeper layers and avoiding the need for full-model fine-tuning during the pruning process. We highlight our contributions as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "1. We introduce the concept of inter-block error compensation, demonstrating that this approach can enhance the performance of one-shot pruning methods with limited" + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.099, + 0.865, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.27, + 0.908, + 0.354 + ], + "angle": 0, + "content": "Figure 2. Comparison between our proposed method ICP and conventional pruning methods. (a) illustrates a standard pruning approach, where all blocks are pruned simultaneously. (b) demonstrates our method. After pruning one block, the subsequent block adjusts to compensate for the error introduced by the pruning of the previous block." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.394, + 0.702, + 0.408 + ], + "angle": 0, + "content": "computational resources." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.409, + 0.907, + 0.499 + ], + "angle": 0, + "content": "2. Building on this concept, we further design a sliding window-based inter-block error compensation strategy. After pruning each layer, we use the next unpruned layer to compensate for the errors induced by pruning the current layer. It improves post-pruning model performance in both vision and language models." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.5, + 0.906, + 0.56 + ], + "angle": 0, + "content": "3. Extensive experiments on large language models and vision models show that our method significantly enhances post-pruning model performance without increasing the GPU memory consumption." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.409, + 0.907, + 0.56 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.593, + 0.655, + 0.608 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.625, + 0.75, + 0.639 + ], + "angle": 0, + "content": "2.1. Transformer Architecture" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.652, + 0.907, + 0.789 + ], + "angle": 0, + "content": "Since its introduction [31], Transformer has become the cornerstone of language models due to its superior performance in language tasks. Following the development of GPT [24], researchers have built increasingly large models using Transformer as primary building blocks. The Transformer has also proven effective in vision tasks. For example, SAM [13], which is based on the Vision Transformer, particularly excels in segmentation and serves as a foundational model for various downstream tasks [20, 23, 34]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Both language and vision models utilize a backbone composed of multiple Transformer blocks. The majority of parameters in these models are concentrated within their respective backbones. These Transformer blocks consist of multi-head self-attention and feed-forward neural networks with two linear layers. Multi-head self-attention is generally considered more critical than the feed-forward network." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9488" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.295, + 0.108 + ], + "angle": 0, + "content": "2.2. Large Model Pruning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.486, + 0.568 + ], + "angle": 0, + "content": "With the emergence of large models, the computational and memory costs of global fine-tuning made iterative pruning methods prohibitively expensive. The speed and lower computational resource requirements of one-shot pruning methods have led to the development of several such techniques for large models, notably SparseGPT and Wanda. SparseGPT [7] is a one-shot rapid pruning method belonging to Optimal Brain Surgeon [10] family. It independently analyzes each basic layer of the model, formulates the pruning problem as a constrained optimization problem, and solves the pruned weight values using the Lagrange multiplier method. SparseGPT builds on the OBS foundation by leveraging the property of Cholesky decomposition to efficiently obtain different orders of Hessian matrices, significantly accelerating the pruning process and reducing memory usage. Essentially, SparseGPT consists of two parts: pruning and compensation, which are integrated during the Lagrange multiplier method's solution process. However, SparseGPT compensates for pruning errors by adjusting unpruned weights within each basic layer, without considering inter-block weight interactions. Wanda [27], introduces a strong prior for language models, assuming that weight importance is related to both the absolute value of the weight and the corresponding output. Wanda designs an importance metric that achieves good results without compensating for pruning losses. However, subsequent studies [33] suggest that while Wanda performs well in many cases, it can occasionally fall short compared to SparseGPT, and its performance in vision models is particularly lacking, where its strong priors may lead to suboptimal results." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.583, + 0.184, + 0.599 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.61, + 0.485, + 0.777 + ], + "angle": 0, + "content": "Our objective is to develop an efficient pruning method for large models, including both vision and language models which should be executable within a reasonable time frame, even under limited computational and memory resources. Additionally, it should sustain robust performance at mid-to-high sparsity levels. In this section, we introduce ICP, a pruning strategy that effectively mitigates performance degradation at mid-to-high sparsity levels and facilitates rapid pruning in memory-constrained environments. The following sections detail our method in two aspects: Blockwise Compensate Pruning and Sparsity Rearrangement." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.788, + 0.379, + 0.804 + ], + "angle": 0, + "content": "3.1. Block-wise Compensate Pruning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.485, + 0.903 + ], + "angle": 0, + "content": "We propose Block-wise Compensate Pruning, which enhances inter-block error compensation by building upon SparseGPT. As depicted in Fig. 3(a), Block-wise Compensate Pruning leverages the modular structure of large models, processing each block sequentially using a sliding window approach. In simple terms, the parameters of the next" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.152 + ], + "angle": 0, + "content": "block are updated to compensate for the pruning-induced errors in the preceding block. Initially, the sliding window contains two blocks. Since the operations are consistent after each slide, we explain using the first two blocks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.156, + 0.909, + 0.565 + ], + "angle": 0, + "content": "In Fig. 3(a), the sliding window moves from left to right with a window length encompassing two blocks. Initially, the window contains \\( B^{1} \\) and \\( B^{2} \\). Pruning is first applied to \\( B^{1} \\), followed by compensating for the pruning-induced error in \\( B^{1} \\) using \\( B^{2} \\). The detailed calculation procedure for this compensation is outlined in Algorithm 1, as follows: The input data from the calibration set passes through the position encoding module, producing an intermediate result, \\( z^{0} \\), which is then duplicated to create \\( z^{0'} \\). At this stage, the feature flow splits into two streams: the \"standard stream\" \\( z^{0'} \\) and the \"error stream\" \\( z^{0} \\). Before pruning \\( B^{1} \\), \\( z^{0'} \\) passes through \\( B^{1} \\), producing an unbiased output, \\( z^{1'} \\). SparseGPT is then applied to prune \\( B^{1} \\), and \\( z^{0} \\) passes through the pruned \\( B^{1} \\), resulting in a biased output, \\( z^{2} \\). Subsequently, \\( z^{1'} \\) passes through the unpruned \\( B^{1} \\), yielding an unbiased output, \\( z^{2'} \\). Here, \\( z^{1} \\) serves as both the output of the pruned \\( B^{1} \\) and the input to \\( B^{2} \\). Since \\( z^{1} \\) is biased, the output from \\( B^{1} \\) will also be biased, whereas the desired output is \\( z^{2'} \\). Therefore, \\( z^{2'} \\) serves as the target for \\( z^{1} \\), forming a labeled feature dataset (\\( z^{1}, z^{2'} \\)). To align output of \\( B^{1} \\) with \\( z^{2'} \\), we adjust \\( B^{1} \\)'s parameters by training it on this feature dataset, reducing the output bias. Finally, the sliding window advances by one block, and the process is repeated: prune the first block and adjust the parameters of the second block using newly formed feature dataset. As above, Block-wise Compensate Pruning progresses sequentially along the backbone, processing one block at a time." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.57, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Note that this method operates with a sliding window containing only two blocks, using the second block to compensate for the pruning-induced errors in the first, rather than relying on additional blocks for error correction. In this setup, once the first block is pruned, the standard stream dataset can pass through the second block to generate labels for the error stream. These labels are then reused during the compensation phase of the next sliding window, eliminating the need to store an additional block for real-time label computation. Consequently, at any given time, at most the weights of a single block reside in memory, optimizing both storage and computation. By contrast, if compensation required two subsequent blocks, the process would either necessitate storing two states of the standard stream or duplicating the next two blocks as a teacher model for distillation-like compensation, both of which would substantially increase memory overhead. It only requires a single block to be loaded into memory at a given moment. As each block is loaded into memory and pruned, the previous block is unloaded before the next block is loaded. This approach significantly reduces memory usage during the pruning and compensation process." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "9489" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.139, + 0.09, + 0.865, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.139, + 0.223, + 0.858, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.361, + 0.908, + 0.391 + ], + "angle": 0, + "content": "Figure 3. Overview of the proposed method. (a) Block-wise Compensate Pruning for sequential block pruning and compensation; (b) basic structure of a transformer block (showing only weight-intensive and essential layers);" + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.431, + 0.4, + 0.446 + ], + "angle": 0, + "content": "Algorithm 1 Block-wise Compensate Pruning" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.449, + 0.483, + 0.491 + ], + "angle": 0, + "content": "Input: Backbone of large model \\(\\{B^i\\}_{i = 1}^n\\) , i-th block \\(B^{i}\\) , Pruning function Prune (module, hyperparameters), Calibration dataset \\(x_{j}\\) , Position embedding block \\(PE\\) Update function Update (module,(sample, label)), Sparsity s, Sparsity Rearrangement parameter \\(\\alpha ,\\beta\\)" + }, + { + "type": "text", + "bbox": [ + 0.095, + 0.491, + 0.267, + 0.502 + ], + "angle": 0, + "content": "Output: Pruned backbone \\(\\{B^i\\}_{i = 1}^n\\)" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.502, + 0.204, + 0.515 + ], + "angle": 0, + "content": "1: \\(z^0 = PE(\\{x_j\\})\\)" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.515, + 0.288, + 0.528 + ], + "angle": 0, + "content": "2: \\( z_{i}^{0^{\\prime}} = z_{i}^{0} / / \\) Copy embeddings" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.528, + 0.3, + 0.547 + ], + "angle": 0, + "content": "3: \\( z_{j}^{1'} = B^{0}\\left(z_{j}^{0'}\\right) // \\text{Preparation} \\)" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.547, + 0.219, + 0.556 + ], + "angle": 0, + "content": "4: for \\( i \\) in 1 to \\( n - 1 \\) do" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.556, + 0.279, + 0.568 + ], + "angle": 0, + "content": "5: \\(B^{i} = \\text{Prune}(B^{i}, (s, \\alpha, \\beta))\\)" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.568, + 0.373, + 0.583 + ], + "angle": 0, + "content": "6: \\(z_{j}^{i} = B^{i}\\left(z_{j}^{i - 1}\\right) / /\\) Forward Propagation" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.583, + 0.241, + 0.601 + ], + "angle": 0, + "content": "7: \\(z_{j}^{i + 1,i^{\\prime}} = B^{i + 1}\\left(z_{j}^{i^{\\prime}}\\right)\\)" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.601, + 0.32, + 0.619 + ], + "angle": 0, + "content": "8: \\( B^{i+1} = \\text{Update}\\left(B^{i+1}, z_{j}^{i}, z_{j}^{i+1'}\\right) \\)" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.619, + 0.15, + 0.628 + ], + "angle": 0, + "content": "9: end for" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.628, + 0.272, + 0.64 + ], + "angle": 0, + "content": "10: \\(B^n = \\text{Prune}(B^n, (s, \\alpha, \\beta))\\)" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.64, + 0.28, + 0.651 + ], + "angle": 0, + "content": "11: return Pruned backbone \\(\\{B^i\\}_{i=1}^n\\)" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.449, + 0.483, + 0.651 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.682, + 0.317, + 0.698 + ], + "angle": 0, + "content": "3.2. Sparsity Rearrangement" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.704, + 0.483, + 0.869 + ], + "angle": 0, + "content": "In many pruning methods, the sparsity of different parts of the model can be adaptively determined through specific rules. However, since our method is based on a sliding window approach for pruning, it is difficult to obtain information about the blocks that are outside the current window. If we attempt to acquire global information, we would either incur significant I/O overhead from loading model segments sequentially or face high memory usage from loading the entire model at once. Some methods even evaluate importance based on gradients, which can add further demands on gradient storage." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "To avoid these issues, fast pruning methods like SparseGPT or Wanda typically assign the same sparsity" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.416, + 0.907, + 0.493 + ], + "angle": 0, + "content": "level uniformly across all parts of the model, equivalent to the target sparsity. However, this approach is clearly suboptimal. Certain prior knowledge can be leveraged to adjust the sparsity distribution, thereby avoiding these problems while improving performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.499, + 0.816, + 0.514 + ], + "angle": 0, + "content": "3.2.1. Inter-Block Sparsity Rearrangement" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.517, + 0.907, + 0.64 + ], + "angle": 0, + "content": "Block-wise Compensate Pruning fine-tunes \\( B^{i+1} \\) to compensate for the pruning-induced errors in \\( B^i \\). However, when pruning the last block, \\( B^n \\), there is no \\( B^{n+1} \\) available to compensate for the errors in \\( B^n \\). Therefore, assigning the same number of pruned weights to each block is not optimal. Instead, the pruning distribution should be adjusted across different blocks. To maintain the global target sparsity, we rearrange the sparsity between blocks as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.637, + 0.653, + 0.907, + 0.671 + ], + "angle": 0, + "content": "\\[\nP _ {B ^ {n}} = \\left| \\theta^ {n} \\right| \\times s \\times \\alpha , \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.582, + 0.686, + 0.906, + 0.717 + ], + "angle": 0, + "content": "\\[\nP _ {B ^ {j}} = \\left| \\theta^ {j} \\right| \\times s + \\frac {\\left| \\theta^ {n} \\right| \\times s \\times (1 - \\alpha)}{n - 1} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.723, + 0.907, + 0.8 + ], + "angle": 0, + "content": "where \\( j = 1, \\ldots, n - 1 \\), \\( \\theta^j \\) denotes the parameters in \\( B^j \\), \\( |\\theta| \\) denotes the number of weights in a single block, and \\( \\alpha \\) is a super-parameter that redistributes a portion of the pruned weights from the last block evenly across all preceding blocks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.807, + 0.818, + 0.822 + ], + "angle": 0, + "content": "3.2.2. Intra-Block Sparsity Rearrangement" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.907, + 0.901 + ], + "angle": 0, + "content": "As shown in Fig. 3(b), the majority of the parameters within a Transformer block are contributed by the matrices depicted, which are clearly hierarchical. Therefore, a block can essentially be viewed as a nested function: the QKV matrices represent the innermost function, while the second" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9490" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.319 + ], + "angle": 0, + "content": "fully connected layer (fc2) represents the outermost function. According to the conclusions drawn in the Theoretical Foundations section, if the same sparsity level is applied uniformly across all layers, the errors introduced by the QKV matrices, which are further from the output of \\( B^{i} \\), will be more challenging to compensate for in \\( B^{i + 1} \\). On the other hand, fc2 is closer to the output of \\( B^{i} \\), making its errors easier to compensate for in \\( B^{i + 1} \\). Additionally, inspired by previous works, earlier layers tend to converge first [1], and faster convergence often leads to better generalization performance [9]. Therefore, the sparsity distribution should be rearranged by reducing the sparsity of the QKV matrices and increasing the sparsity of fc2. To ensure that the global target sparsity remains unchanged, we rearrange the sparsity according to the following rule:" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.327, + 0.483, + 0.346 + ], + "angle": 0, + "content": "\\[\nP _ {t} = s \\times | W _ {t} ^ {i} | \\times (1 - \\beta), t \\in \\{Q, K, V \\}. (3)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.158, + 0.367, + 0.483, + 0.387 + ], + "angle": 0, + "content": "\\[\nP _ {f c 2} = s \\times | W _ {f c 2} ^ {i} | + s \\times | W _ {f c 2} ^ {i} | \\times \\beta . \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.389, + 0.484, + 0.435 + ], + "angle": 0, + "content": "where \\(P_{t}\\) represents the number of pruned weights, \\(|W|\\) denotes the total number of weights in a single layer, and \\(\\beta\\) is a super-parameter that needs to be adjusted." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.447, + 0.224, + 0.464 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.471, + 0.282, + 0.489 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.494, + 0.484, + 0.614 + ], + "angle": 0, + "content": "To evaluate the effectiveness of our method, we focus on models with 7 billion parameters or fewer, as these are the most commonly used in research and resource-constrained scenarios. Pruning these models is also the most feasible on a single desktop-class GPU. Specifically, we validate our approach using the OPT-125M/1.3B/2.7B/6.7B and Llama 2-7B, as well as the SAM-B/L/H. All experiments were conducted on a single NVIDIA RTX 3090 GPU." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.615, + 0.484, + 0.689 + ], + "angle": 0, + "content": "For the compensation process, we employ the Adam optimizer with a Mean Squared Error (MSE) loss function and a batch size of 1. All compensation operations during pruning share a unified learning rate, which is adjusted using the ExponentialLR scheduler with a decay factor of 0.87." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.484, + 0.78 + ], + "angle": 0, + "content": "For calibration data, language model calibration dataset follows SparseGPT, using 128 segments of 2048 tokens, randomly sampled from C4 dataset [25]. Similarly, the vision model calibration dataset consists of 128 images, randomly sampled from the second compressed file of the SA-1B dataset [13] provided by Meta." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To comprehensively evaluate the performance of pruned models, we validate OPT using the Wikitext-v2 [21] and PTB datasets [19] from Hugging Face, with perplexity (PPL) as the evaluation metric. We also evaluate the pruned Llama 2-7B model using accuracy on seven ZeroShot tasks: BoolQ [2], RTE [32], HellaSwag [35], Wino-Grande [26], ARC Easy and Challenge [3], and Open-bookQA [22], following the experimental setup introduced" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.198 + ], + "angle": 0, + "content": "by Sun et al. [27] in their work on Wanda. This allows for a more task-specific evaluation of the pruned model's generalization capabilities. For SAM, we validate the performance using the fourth compressed file of the SA-1B dataset as well as the COCO dataset [15], employing Intersection over Union (IoU) as the primary metric for visual models." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.209, + 0.722, + 0.225 + ], + "angle": 0, + "content": "4.2. Unstructured Pruning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.231, + 0.907, + 0.502 + ], + "angle": 0, + "content": "Language Models. In Table 1 and Table 2, we report the perplexity performance of the OPT series models on Wikitext and PTB datasets after pruning to various sparsity levels using our method and other baseline approaches. Our method demonstrates superior performance over baselines even at lower sparsity levels, and as sparsity increases, the performance gap widens in favor of our approach. Additionally, as model size decreases, the advantage of our method over the baselines becomes even more pronounced. Table 3 also presents results in comparing the performances of our proposed method against those of baselines on the Llama 2-7B model at \\(50\\%\\) and \\(70\\%\\) sparsity levels. Our method consistently outperforms all baseline approaches at both sparsity levels, achieving the best results across all tasks. These findings highlight the effectiveness of the interblock error compensation pruning method introduced in this paper, particularly under high-sparsity conditions, across different model scales, and for various model architectures." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.516, + 0.906, + 0.556 + ], + "angle": 0, + "content": "Table 1. Perplexity performances of pruned OPT models of different scales at various sparsity levels. The models are calibrated on a subset of the C4 and evaluated on full Wikitext-v2." + }, + { + "type": "table", + "bbox": [ + 0.524, + 0.557, + 0.898, + 0.763 + ], + "angle": 0, + "content": "
MethodModel10%20%30%40%50%60%70%80%90%
Magnitude125M28.2529.6534.5154.60193.4920.0380648906614
SparseGPT27.8827.9828.8630.5637.0259.55220.923784666
Wanda27.4827.7628.1130.6738.9275.17328.219184609
ICP (ours)27.3727.8728.6629.6633.9045.7865.20270.41612
Magnitude1.3B14.7215.6224.74388.01713939294431634428871
SparseGPT14.6714.7515.1616.4917.5022.0851.75752.46797
Wanda14.6314.6915.0115.8918.4126.5599.53225816868
ICP (ours)14.5914.6014.6715.0416.4921.1736.78117.11099
Magnitude2.7B12.5913.1315.5830.32265.236047251961416668
SparseGPT12.3212.3912.6612.6613.4616.0426.92138.45818
Wanda12.2312.2612.3612.8614.2119.83387.9590516527
ICP (ours)12.2212.2512.3412.6213.3215.7721.7954.67764.2
Magnitude6.7B10.9211.2712.5331.89968.71263916975255915297
SparseGPT10.8310.7610.7510.9611.5913.4820.4896.2511938
Wanda10.7310.6310.6410.9611.9815.19159.2395416076
ICP (ours)10.5610.5810.5910.8511.2713.0317.8244.691469
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Vision Models. In Tables 4 and Table 5, we report IoU performance of SAM at varying sparsity levels on SA-1B and COCO datasets with single-point interaction. Following experimental setup, SAMs are pruned with only 128 images random sampled from SA-1B as calibration data. The experiments first verify that SparseGPT indeed achieves strong IoU performance on visual models under low sparsity conditions, outperforming both Magnitude and Wanda." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "9491" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.093, + 0.09, + 0.482, + 0.13 + ], + "angle": 0, + "content": "Table 2. Perplexity performancese of pruned OPT models of different scales at various sparsity levels. The models are calibrated on a subset of the C4 and evaluated on PTB." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.131, + 0.472, + 0.337 + ], + "angle": 0, + "content": "
MethodModel10%20%30%40%50%60%70%80%90%
Magnitude125M39.9743.2549.9782.69276.21146343041245382
SparseGPT39.4340.3642.7145.2455.7888.44259.629195374
Wanda39.0539.8341.4344.4157.59109.7399.722144474
ICP (ours)39.1239.1340.9643.4749.4164.1088.85336.33075
Magnitude1.3B20.6621.5139.89823.53171949984671728333126
SparseGPT20.3420.6421.5023.1725.4733.4876.91531.87515
Wanda20.3820.7021.4123.2528.0043.95135.3180912025
ICP (ours)20.1320.4121.1122.1924.6832.1954.33148.83094
Magnitude2.7B18.3819.0622.9646.36262.632077276957512178
SparseGPT18.1518.5018.9818.9820.4424.6742.75146.06107
Wanda17.9718.0418.3219.2521.8633.44450.8546221039
ICP (ours)17.9718.0218.2918.8620.0522.9235.1771.912357
Magnitude6.7B16.1816.7018.5248.85613.493799973157424260
SparseGPT15.7715.8115.9216.2817.4320.2331.86105.08320
Wanda15.7615.7715.9416.4317.9223.61211.3302414998
ICP (ours)15.7615.7415.9416.2217.3020.0728.0464.182549
" + }, + { + "type": "table_caption", + "bbox": [ + 0.093, + 0.355, + 0.482, + 0.396 + ], + "angle": 0, + "content": "Table 3. Zero-shot performance (Accuracy, %) of Llama 2-7B pruned by different method on various tasks at sparsity of \\(50\\%\\) and \\(70\\%\\)." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.398, + 0.472, + 0.526 + ], + "angle": 0, + "content": "
MethodStrategyBQRTEHSWGARC-eARC-cOBQAMean
Dense077.7463.1857.1068.9876.2643.4331.4059.73
Magnitude50%62.9457.0449.1263.3864.0634.6426.8051.14
SparseGPT76.3355.9652.8968.9872.0138.2328.4056.11
Wanda76.4253.4352.4568.6772.2239.3331.0056.22
ICP (ours)76.5057.0453.0369.1172.3940.4431.0056.96
Magnitude70%37.9553.0725.9349.2527.8222.8717.0033.41
SparseGPT64.6553.7933.4558.6443.3121.6717.0041.79
Wanda48.4752.7127.9749.6430.8118.6012.2034.34
ICP (ours)67.7155.2341.4560.2255.5627.0521.4046.94
" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.557, + 0.482, + 0.692 + ], + "angle": 0, + "content": "As Wanda is based on strong priors from language models, its underperformance compared to SparseGPT on visual models is within our expectations. Lastly, the method proposed in this paper outperforms all baselines and exhibits a similar trend observed in language model experiments: the higher the sparsity, the greater the advantage of our method over the baseline methods. This demonstrates that the proposed method is not only effective for language models but also remains effective for visual models." + }, + { + "type": "title", + "bbox": [ + 0.095, + 0.71, + 0.449, + 0.727 + ], + "angle": 0, + "content": "4.3. Semi-Structure Pruning and Quantization" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.735, + 0.482, + 0.901 + ], + "angle": 0, + "content": "On certain specialized hardware, such as specific NVIDIA GPUs or other custom-designed computing architectures, the combination of pruning and quantization can further accelerate inference while maintaining an acceptable performance degradation. To explore this, we conduct tests using a joint compression strategy that combines pruning and quantization. Specifically, we focus on the combination of 2:4 and 4:8 semi-structured sparsity with 3-bit and 4-bit quantization. Since Wanda does not support quantization, we exclude it from experiments involving quantization and instead compared our approach against SparseGPT." + }, + { + "type": "table_caption", + "bbox": [ + 0.517, + 0.09, + 0.905, + 0.144 + ], + "angle": 0, + "content": "Table 4. Instance segmentation performance of SAM models (IoU, \\(\\%\\) ) at various sparsity using single-point interaction on SA-1B. The models are calibrated on a subset of 128 images sampled from SA-1B." + }, + { + "type": "table", + "bbox": [ + 0.527, + 0.146, + 0.895, + 0.304 + ], + "angle": 0, + "content": "
MethodModel10%20%30%40%50%60%70%80%90%
MagnitudeSAM-H74.8074.5773.8772.2669.4561.5128.222.651.70
SparseGPT74.8174.7774.6074.2073.3071.1565.6950.9718.15
Wanda74.8274.7674.3573.6071.6567.6553.0021.841.20
ICP (ours)74.8274.7874.6674.6074.4273.8272.8569.7357.60
MagnitudeSAM-L74.8174.5973.9372.2869.1862.5645.529.924.52
SparseGPT74.8474.7774.6174.1373.2070.7064.4545.487.37
Wanda74.8674.7874.6274.0572.5168.2855.1322.811.31
ICP (ours)74.8574.7974.6974.5574.2773.6472.2668.1753.68
MagnitudeSAM-B72.6972.5671.9970.3667.3860.8746.0125.531.51
SparseGPT72.7272.6772.5072.0871.1168.6863.0247.4918.81
Wanda72.7172.6472.3371.4569.4464.3550.2928.162.18
ICP (ours)72.7272.6972.5672.2871.8571.0669.3364.7851.82
" + }, + { + "type": "table_caption", + "bbox": [ + 0.517, + 0.319, + 0.904, + 0.346 + ], + "angle": 0, + "content": "Table 5. Instance segmentation performance of SAM models (IoU, %) at various sparsity using single-point interaction on COCO." + }, + { + "type": "table", + "bbox": [ + 0.527, + 0.348, + 0.895, + 0.504 + ], + "angle": 0, + "content": "
MethodModel10%20%30%40%50%60%70%80%90%
MagnitudeSAM-H71.1271.1170.8670.6169.3763.6032.687.887.21
SparseGPT71.1171.1870.9870.9370.4269.5266.7056.8227.57
Wanda71.1771.0170.9870.7369.6366.0553.4522.835.51
ICP (ours)71.1471.1871.1270.9870.8170.2469.4366.6958.41
MagnitudeSAM-L70.5670.5170.2769.5368.0862.6649.1220.6618.12
SparseGPT70.5570.5370.4370.0469.6168.2164.4248.3511.30
Wanda70.5670.5870.4569.9168.6565.5654.6719.896.38
ICP (ours)70.5570.5870.5870.3769.9169.5168.1465.4256.92
MagnitudeSAM-B66.8467.0667.0666.6765.2962.2853.6140.867.48
SparseGPT67.0967.0266.9866.7666.1765.0062.0352.4628.51
Wanda66.9466.9767.0466.4765.3862.3153.5839.327.89
ICP (ours)67.1067.0667.0466.8366.4566.0764.5162.4054.50
" + }, + { + "type": "table_caption", + "bbox": [ + 0.517, + 0.52, + 0.905, + 0.547 + ], + "angle": 0, + "content": "Table 6. Perplexity performance of pruned OPT models at various compression strategies." + }, + { + "type": "table", + "bbox": [ + 0.536, + 0.548, + 0.887, + 0.762 + ], + "angle": 0, + "content": "
MethodStrategyOPT-125MOPT-1.3BOPT-2.7BOPT-6.7B
WTPTBWTPTBWTPTBWTPTB
SparseGPT2:459.3593.0423.8738.0917.1126.9814.1521.54
Wanda2:480.01111.9728.2043.3721.1834.5515.9025.09
ICP (ours)2:446.9067.5622.0333.9516.3925.3114.0421.19
SparseGPT4:843.9071.9920.2131.3615.0023.0512.4918.86
Wanda4:853.1879.0222.1734.5616.7926.1313.5520.19
ICP (ours)4:839.9159.3419.1530.4114.8822.8812.3818.53
SparseGPT2:4+int3122.33196.4440.5864.5324.7041.0620.0731.91
ICP (ours)2:4+int374.07114.9327.8846.1419.5730.2818.3826.82
SparseGPT2:4+int470.88107.3925.9641.6918.5630.6815.2423.01
ICP (ours)2:4+int451.2475.3923.4934.4217.0527.0415.1022.54
SparseGPT4:8+int384.02120.1632.4149.2720.2432.9917.0826.36
ICP (ours)4:8+int359.6786.0124.7340.1917.8926.8016.1624.30
SparseGPT4:8+int450.5880.9723.3234.3815.7424.7613.4419.72
ICP (ours)4:8+int442.0360.6020.8631.5815.3323.8013.2419.49
" + }, + { + "type": "title", + "bbox": [ + 0.517, + 0.788, + 0.809, + 0.804 + ], + "angle": 0, + "content": "4.4. Sparsity Rearrangement Ablation" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.81, + 0.905, + 0.87 + ], + "angle": 0, + "content": "To validate the Sparsity Rearrangement method based on priors, as proposed in the Method section, we conducted experiments on both Inter-Block Sparsity Rearrangement and Intra-Block Sparsity Rearrangement." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Inter-Block Sparsity Rearrangement. In Fig. 4a and Fig. 4b, we present the performance of OPT-125M and" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.515, + 0.957 + ], + "angle": 0, + "content": "9492" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.092, + 0.09, + 0.483, + 0.117 + ], + "angle": 0, + "content": "Table 7. Zero-shot Performance (Accuracy, %) of Llama 2-7B Model under various compression strategies." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.118, + 0.472, + 0.316 + ], + "angle": 0, + "content": "
MethodStrategyBQRTEHSWGARC-eARC-cOBQAMean
SparseGPT2:468.0758.8443.3865.5964.1831.9124.8050.97
Wanda2:468.1353.4341.4262.4363.1330.6323.8049.00
ICP (ours)2:469.7258.8446.7564.4866.0433.7026.2052.25
SparseGPT4:871.0754.8748.3567.8068.6934.4727.2053.21
Wanda4:873.0053.7947.0366.9367.4734.2227.0052.78
ICP (ours)4:873.1258.1250.1364.8868.9036.0128.0054.16
SparseGPT2:4+int364.6555.2338.5859.0455.4725.3419.4045.39
ICP (ours)2:4+int366.8254.1543.3660.7759.5528.6721.0047.76
SparseGPT2:4+int466.7354.1542.3465.1961.7428.9223.8048.98
ICP (ours)2:4+int470.4656.6846.2262.9064.8130.6322.8050.64
SparseGPT4:8+int368.5953.4342.1262.3560.1029.1021.6048.18
ICP (ours)4:8+int368.3854.8743.5462.8563.3832.3422.4049.68
SparseGPT4:8+int469.2754.1546.5866.0667.5534.3927.2052.17
ICP (ours)4:8+int472.2961.0148.8662.7566.8833.8728.0053.38
" + }, + { + "type": "table_caption", + "bbox": [ + 0.092, + 0.33, + 0.483, + 0.371 + ], + "angle": 0, + "content": "Table 8. Instance segmentation performance of SAM models (IoU, \\(\\%\\) ) at various compression strategies using single-point interaction on SA-1B and COCO." + }, + { + "type": "table", + "bbox": [ + 0.121, + 0.373, + 0.455, + 0.587 + ], + "angle": 0, + "content": "
MethodStrategySAM-BSAM-LSAM-H
COCOSA-1BCOCOSA-1BCOCOSA-1B
SparseGPT2:462.8067.5965.0169.0767.0269.62
Wanda2:460.5962.0662.3765.8464.6866.36
ICP (ours)2:464.9571.3067.8373.5669.0973.64
SparseGPT4:863.9369.6966.5671.2767.9871.41
Wanda4:862.6266.0565.0269.7566.8069.55
ICP (ours)4:865.3771.7268.1773.8669.4474.03
SparseGPT2:4+int362.0964.7763.4565.6766.3967.85
ICP (ours)2:4+int364.4769.7167.1372.6068.7573.26
SparseGPT2:4+int462.6666.7664.7868.6966.8169.44
ICP (ours)2:4+int464.8970.6467.6473.0969.0273.72
SparseGPT4:8+int363.0267.2764.7468.0267.2569.69
ICP (ours)4:8+int364.8469.8667.6873.2869.1073.35
SparseGPT4:8+int463.7768.9866.0470.3167.8771.05
ICP (ours)4:8+int465.2470.9868.0273.8269.3473.78
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.613, + 0.483, + 0.824 + ], + "angle": 0, + "content": "SAM-B pruned by ICP with different \\(\\alpha\\). For clarity in visual presentation, we conducted experiments with \\(70\\%\\) sparsity on OPT-125M and \\(90\\%\\) sparsity on SAM-B. We observed that in the language model, \\(\\alpha\\) had the greatest impact on the performance of OPT-125M on the PTB dataset, and adjusting \\(\\alpha\\) further improved the model's performance on non-calibration datasets. On vision models, the influence of \\(\\alpha\\) on performance was similar on both datasets, with adjustments to \\(\\alpha\\) enhancing the overall performance of the visual model. Overall, appropriate tuning of \\(\\alpha\\) can further improve the performance of pruned models. Based on experiments, we conclude that \\(\\alpha\\) typically yields optimal results within the range of [0.6, 0.75] for language models, and within [0.8, 0.9] for vision model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Intra-Block Sparsity Rearrangement. In Fig. 4c, we illustrate the impact of different \\(\\beta\\) on the language model, using the \\(70\\%\\) sparsity model of OPT-125M as an example. We observed that in the language model, adjusting \\(\\beta\\) can further reduce the model's perplexity. Unlike \\(\\alpha\\), \\(\\beta\\) not" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.273 + ], + "angle": 0, + "content": "only improves the model's performance on non-calibration datasets but also enhances its performance on calibration datasets. Overall, appropriately adjusting \\(\\beta\\) can further improve the performance of the pruned model. The optimal value for \\(\\beta\\) typically falls within the range of \\([0.6 \\times (1 - s), 0.8 \\times (1 - s)]\\), where \\(s\\) denotes sparsity. In Fig. 4d, we exemplify the effects of varying \\(\\beta\\) on the visual model, specifically using the \\(90\\%\\) sparsity model of SAM-B. Our findings indicate that the influence of \\(\\beta\\) on the visual model is minimal; both on calibration and non-calibration datasets, adjustments to \\(\\beta\\) yield only marginal improvements in the model's performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.284, + 0.776, + 0.299 + ], + "angle": 0, + "content": "4.5. Calibration Set Size Ablation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.306, + 0.907, + 0.473 + ], + "angle": 0, + "content": "Although our experiments follow the settings of SparseGPT and Wanda, using 128 randomly selected samples for pruning both language and vision models, Fig. 4e and Fig. 4f show the impact of calibration set size on our method's performance. Fig. 1 presents results on the OPT-125M language model at \\(70\\%\\) sparsity, while Fig. 2 shows results for the SAM-B vision model at \\(90\\%\\) sparsity. In both cases, increasing the sample size beyond 128 offers minimal performance gains while significantly increasing GPU memory usage. Thus, like SparseGPT and Wanda, 128 samples remain the optimal choice." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.483, + 0.787, + 0.5 + ], + "angle": 0, + "content": "4.6. Compensation Epoch Ablation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.506, + 0.909, + 0.673 + ], + "angle": 0, + "content": "In Fig.4g and Fig.4h, we illustrate the impact of varying the number of epochs during the compensation on the performance of both vision and language models. It is evident that a single epoch is sufficient to effectively mitigate the error introduced by pruning in earlier layers. While additional epochs can further improve performance, the marginal gains diminish rapidly. Although we report results using 12 epochs in the experiments presented in 4.2 and 4.3 to demonstrate the best achievable performance, selecting a single epoch offers the optimal trade-off between model performance and computational overhead." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.682, + 0.774, + 0.699 + ], + "angle": 0, + "content": "4.7. Memory and Time Efficiency" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Peak memory usage (PMU) and execution time are critical factors in determining whether a pruning method can be deployed on devices with limited memory capacity and lower cost. In Table 9, we present the PMU and runtime of SparseGPT, Wanda, and our proposed method when pruning OPT-6.7B on an NVIDIA RTX 3090. The experimental setup is consistent with the description in 4.1. Our method employs a commonly used chunked storage strategy, where intermediate outputs of each layer are stored in host memory instead of GPU memory, effectively reducing memory usage. While this approach may slow down the pruning process, as discussed in Section 4.6, a single epoch of fine-tuning is sufficient to significantly improve" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9493" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.355, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.093, + 0.213, + 0.356, + 0.237 + ], + "angle": 0, + "content": "(a) Perplexity performance of pruned OPT-125M with different \\(\\alpha\\)" + }, + { + "type": "image", + "bbox": [ + 0.367, + 0.089, + 0.632, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.367, + 0.213, + 0.631, + 0.237 + ], + "angle": 0, + "content": "(b) Instance segmentation performance (IoU, %) of SAM-B with different \\(\\alpha\\)." + }, + { + "type": "image", + "bbox": [ + 0.644, + 0.089, + 0.905, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.642, + 0.213, + 0.905, + 0.237 + ], + "angle": 0, + "content": "(c) Perplexity performance of OPT-125M with different \\(\\beta\\)" + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.242, + 0.356, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.092, + 0.365, + 0.356, + 0.39 + ], + "angle": 0, + "content": "(d) Instance segmentation performance (IoU, %) of SAM-B with different \\(\\beta\\)." + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.242, + 0.631, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.367, + 0.365, + 0.63, + 0.39 + ], + "angle": 0, + "content": "(e) Perplexity performance of OPT-125M with different calibration set size." + }, + { + "type": "image", + "bbox": [ + 0.644, + 0.242, + 0.906, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.642, + 0.365, + 0.905, + 0.39 + ], + "angle": 0, + "content": "(f) Instance segmentation performance (IoU, %) of SAM-B with different calibration set size." + }, + { + "type": "image", + "bbox": [ + 0.195, + 0.395, + 0.457, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.194, + 0.518, + 0.456, + 0.543 + ], + "angle": 0, + "content": "(g) Perplexity performance of pruned OPT-125M with different number of compensation epochs." + }, + { + "type": "image", + "bbox": [ + 0.542, + 0.395, + 0.807, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.541, + 0.518, + 0.805, + 0.543 + ], + "angle": 0, + "content": "(h) Segmentation performance of SAM-B with different number of compensation epochs." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.555, + 0.908, + 0.584 + ], + "angle": 0, + "content": "Figure 4. Performance metrics for pruned OPT-125M and SAM-B models under varying parameters \\(\\alpha\\), \\(\\beta\\), calibration dataset sizes, and number of compensation epochs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.61, + 0.484, + 0.746 + ], + "angle": 0, + "content": "model performance. Consequently, data only needs to be loaded from host memory once, keeping the additional time overhead within an acceptable range. It is worth noting that we use full-precision fine-tuning rather than Parameter-Efficient Fine-Tuning (PEFT) techniques. If combined with PEFT methods, memory consumption could be further reduced. This advantage enables our method to efficiently prune larger models on desktop-class GPUs, thereby expanding its applicability." + }, + { + "type": "table_caption", + "bbox": [ + 0.138, + 0.776, + 0.436, + 0.789 + ], + "angle": 0, + "content": "Table 9. GPU Memory Consumption Comparison" + }, + { + "type": "table", + "bbox": [ + 0.163, + 0.79, + 0.415, + 0.876 + ], + "angle": 0, + "content": "
MethodPMUTime(s)
SparseGPT8.3 GB1004
Wanda21 GB143
ICP (1-epoch)7.8 GB1551
ICP (12-epoch)7.8 GB4470
" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.608, + 0.634, + 0.625 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we propose an efficient pruning method ICP, which is effective for both vision and language models while consuming limited GPU memory. ICP update each block individually by compensating for the inter-block errors introduced by pruning, then rearranges the sparsity distribution based on prior knowledge of Transformer and pruning method. It significantly improves the performance of models pruned at high sparsity levels without increasing GPU memory consumption. ICP, as it does not rely on strong priors specific to any particular model, demonstrates dual applicability to both vision and language models. This flexibility suggests that it could potentially be extended to multimodal models in the future. However, this work has limitations: (1) the hyperparameters \\(\\alpha\\) and \\(\\beta\\) are not automatically tuned; (2) ICP has not been tested on multimodal models; (3) the potential of individual blocks has not been fully explored. These will be the focus of our future work." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9494" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.114, + 0.484, + 0.171 + ], + "angle": 0, + "content": "This work is supported by Natural Science Foundation of China under Grant 62271465, Suzhou Basic Research Program under Grant SYG202338, and Jiangsu Province Science Foundation for Youths (NO. BK20240464)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.198, + 0.188, + 0.214 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.224, + 0.484, + 0.293 + ], + "angle": 0, + "content": "[1] Yixiong Chen, Alan Yuille, and Zongwei Zhou. Which layer is learning faster? a systematic exploration of layerwise convergence rate for deep neural networks. In The Eleventh International Conference on Learning Representations, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.295, + 0.484, + 0.35 + ], + "angle": 0, + "content": "[2] Christopher Clark, Kenton Lee, Ming-Wei Chang, Tom Kwiatkowski, Michael Collins, and Kristina Toutanova. Boolq: Exploring the surprising difficulty of natural yes/no questions. arXiv preprint arXiv:1905.10044, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.351, + 0.483, + 0.419 + ], + "angle": 0, + "content": "[3] Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. Think you have solved question answering? try arc, the ai2 reasoning challenge. arXiv preprint arXiv:1803.05457, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.422, + 0.483, + 0.477 + ], + "angle": 0, + "content": "[4] Peijie Dong, Lujun Li, Zhenheng Tang, Xiang Liu, Xinglin Pan, Qiang Wang, and Xiaowen Chu. Pruner-zero: Evolving symbolic pruning metric from scratch for large language models. arXiv preprint arXiv:2406.02924, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.479, + 0.484, + 0.56 + ], + "angle": 0, + "content": "[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.563, + 0.484, + 0.616 + ], + "angle": 0, + "content": "[6] Jonathan Frankle and Michael Carbin. The lottery ticket hypothesis: Finding sparse, trainable neural networks. In International Conference on Learning Representations, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.619, + 0.484, + 0.674 + ], + "angle": 0, + "content": "[7] Elias Frantar and Dan Alistarh. Sparsegpt: Massive language models can be accurately pruned in one-shot. In International Conference on Machine Learning, pages 10323-10337. PMLR, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.676, + 0.484, + 0.731 + ], + "angle": 0, + "content": "[8] Song Han, Jeff Pool, John Tran, and William Dally. Learning both weights and connections for efficient neural network. Advances in neural information processing systems, 28, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.733, + 0.484, + 0.787 + ], + "angle": 0, + "content": "[9] Moritz Hardt, Ben Recht, and Yoram Singer. Train faster, generalize better: Stability of stochastic gradient descent. In International conference on machine learning, pages 1225-1234. PMLR, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.789, + 0.483, + 0.843 + ], + "angle": 0, + "content": "[10] Babak Hassibi and David Stork. Second order derivatives for network pruning: Optimal brain surgeon. In Advances in Neural Information Processing Systems. Morgan-Kaufmann, 1992. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.095, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[11] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.095, + 0.224, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[12] Lei Ke, Mingqiao Ye, Martin Danelljan, Yu-Wing Tai, Chi-Keung Tang, Fisher Yu, et al. Segment anything in high quality. Advances in Neural Information Processing Systems, 36, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[13] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.234, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[14] Chunyuan Li, Cliff Wong, Sheng Zhang, Naoto Usuyama, Haotian Liu, Jianwei Yang, Tristan Naumann, Hoifung Poon, and Jianfeng Gao. Llava-med: Training a large language-and-vision assistant for biomedicine in one day. Advances in Neural Information Processing Systems, 36, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.305, + 0.905, + 0.374 + ], + "angle": 0, + "content": "[15] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision - ECCV 2014, pages 740-755, Cham, 2014. Springer International Publishing. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.376, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[16] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.419, + 0.904, + 0.445 + ], + "angle": 0, + "content": "[17] AI@Meta Llama Team. The llama3 herd of models, 2024. Accessed: 2024-08-01. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.448, + 0.905, + 0.502 + ], + "angle": 0, + "content": "[18] Xinyin Ma, Gongfan Fang, and Xinchao Wang. Llm-pruner: On the structural pruning of large language models. Advances in neural information processing systems, 36:21702-21720, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.504, + 0.905, + 0.587 + ], + "angle": 0, + "content": "[19] Mitch Marcus, Grace Kim, Mary Ann Marcinkiewicz, Robert MacIntyre, Ann Bies, Mark Ferguson, Karen Katz, and Britta Schasberger. The penn treebank: Annotating predicate argument structure. In Human Language Technology: Proceedings of a Workshop held at Plainsboro, New Jersey, March 8-11, 1994, 1994. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.589, + 0.905, + 0.644 + ], + "angle": 0, + "content": "[20] Maciej A Mazurowski, Haoyu Dong, Hanxue Gu, Jichen Yang, Nicholas Konz, and Yixin Zhang. Segment anything model for medical image analysis: an experimental study. Medical Image Analysis, 89:102918, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.647, + 0.905, + 0.687 + ], + "angle": 0, + "content": "[21] Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. Pointer sentinel mixture models. arXiv preprint arXiv:1609.07843, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.689, + 0.905, + 0.743 + ], + "angle": 0, + "content": "[22] Todor Mihaylov, Peter Clark, Tushar Khot, and Ashish Sabharwal. Can a suit of armor conduct electricity? a new dataset for open book question answering. arXiv preprint arXiv:1809.02789, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.905, + 0.787 + ], + "angle": 0, + "content": "[23] Quan Quan, Fenghe Tang, Zikang Xu, Heqin Zhu, and S Kevin Zhou. Slide-sam: Medical sam meets sliding window. In Medical Imaging with Deep Learning, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[24] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[25] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9495" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[26] Keisuke Sakaguchi, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. Winogrande: An adversarial winograd schema challenge at scale. Communications of the ACM, 64 (9):99-106, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.205 + ], + "angle": 0, + "content": "[27] Mingjie Sun, Zhuang Liu, Anna Bair, and J Zico Kolter. A simple and effective pruning approach for large language models. In The Twelfth International Conference on Learning Representations, 2024. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.275 + ], + "angle": 0, + "content": "[28] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.482, + 0.346 + ], + "angle": 0, + "content": "[29] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.348, + 0.482, + 0.388 + ], + "angle": 0, + "content": "[30] Tycho FA van der Ouderaa, Markus Nagel, Mart Van Baalen, Yuki M Asano, and Tijmen Blankevoort. The IIm surgeon. arXiv preprint arXiv:2312.17244, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.39, + 0.482, + 0.445 + ], + "angle": 0, + "content": "[31] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.447, + 0.482, + 0.487 + ], + "angle": 0, + "content": "[32] Alex Wang. Glue: A multi-task benchmark and analysis platform for natural language understanding. arXiv preprint arXiv:1804.07461, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.489, + 0.482, + 0.543 + ], + "angle": 0, + "content": "[33] Miles Williams and Nikolaos Aletras. How does calibration data affect the post-training pruning and quantization of large language models? arXiv preprint arXiv:2311.09755, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.546, + 0.482, + 0.601 + ], + "angle": 0, + "content": "[34] Junde Wu, Wei Ji, Yuanpei Liu, Huazhu Fu, Min Xu, Yanwu Xu, and Yueming Jin. Medical sam adapter: Adapting segment anything model for medical image segmentation. arXiv preprint arXiv:2304.12620, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.603, + 0.482, + 0.643 + ], + "angle": 0, + "content": "[35] Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. Hellaswag: Can a machine really finish your sentence? arXiv preprint arXiv:1905.07830, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.646, + 0.482, + 0.713 + ], + "angle": 0, + "content": "[36] Renrui Zhang, Jiaming Han, Chris Liu, Peng Gao, Aojun Zhou, Xiangfei Hu, Shilin Yan, Pan Lu, Hongsheng Li, and Yu Qiao. Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.716, + 0.482, + 0.783 + ], + "angle": 0, + "content": "[37] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.783 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9496" + } + ] +] \ No newline at end of file diff --git a/2025/ICP_ Immediate Compensation Pruning for Mid-to-high Sparsity/9ac0c560-9a5b-4372-9074-16698b23f1e1_origin.pdf b/2025/ICP_ Immediate Compensation Pruning for Mid-to-high Sparsity/9ac0c560-9a5b-4372-9074-16698b23f1e1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..842697d4daec7dbf6b148f0d6677d5febf3bc56d --- /dev/null +++ b/2025/ICP_ Immediate Compensation Pruning for Mid-to-high Sparsity/9ac0c560-9a5b-4372-9074-16698b23f1e1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:689adc8485bada49ba8f4e35f4d2cfa1175ef70deddc8c5bf83e844065c8c088 +size 693054 diff --git a/2025/ICP_ Immediate Compensation Pruning for Mid-to-high Sparsity/full.md b/2025/ICP_ Immediate Compensation Pruning for Mid-to-high Sparsity/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c13cc6630defce8f5a9410e52bdece79ac9404ef --- /dev/null +++ b/2025/ICP_ Immediate Compensation Pruning for Mid-to-high Sparsity/full.md @@ -0,0 +1,300 @@ +# ICP: Immediate Compensation Pruning for Mid-to-high Sparsity + +Xin Luo $^{1,2}$ Xueming Fu $^{1,2}$ Zihang Jiang $^{1,2*}$ S.Kevin Zhou $^{1,2,3,4*}$ + +1 School of Biomedical Engineering, Division of Life Sciences and Medicine, USTC + +2 MIRACLE Center, Suzhou Institute for Advance Research, USTC + +3 Key Laboratory of Intelligent Information Processing of CAS, ICT, CAS + +$^{4}$ State Key Laboratory of Precision and Intelligent Chemistry, USTC + +{luoxin.ustc, xuemingfu, jzh0103}@mail.ustc.edu.cn + +{s.kevin.zhou}@gmail.com + +# Abstract + +The increasing adoption of large-scale models under 7 billion parameters in both language and vision domains enables inference tasks on a single consumer-grade GPU but makes fine-tuning models of this scale, especially 7B models, challenging. This limits the applicability of pruning methods that require full fine-tuning. Meanwhile, pruning methods that do not require fine-tuning perform well at low sparsity levels (10%-50%) but struggle at mid-to-high sparsity levels (50%-70%), where the error behaves equivalently to that of semi-structured pruning. To address these issues, this paper introduces ICP, which finds a balance between full fine-tuning and zero fine-tuning. First, Sparsity Rearrange is used to reorganize the predefined sparsity levels, followed by Block-wise Compensate Pruning, which alternates pruning and compensation on the model's backbone, fully utilizing inference results while avoiding full model fine-tuning. Experiments show that ICP improves performance at mid-to-high sparsity levels compared to baselines, with only a slight increase in pruning time and no additional peak memory overhead. + +# 1. Introduction + +Introduction of Transformer [5, 31] propelled both language and vision models toward larger scales, leading to the development of high-performance models such as the OPT [37] and Llama [17, 28, 29] series for language tasks, and the SAM [13] series for vision tasks. However, their large scale presents significant deployment challenges. In scenarios with downstream tasks or hardware constraints, models under 8 billion (8B) parameters for language tasks and under 650 million parameters for vision tasks are most commonly used [12, 14, 16, 34, 36]. Nevertheless, model compres + +sion is necessary to further reduce their size for deployment. Moreover, on a single desktop-class GPU, it is typically only feasible to perform inference with these models, while fine-tuning remains impractical due to hardware limitations. + +Pruning is a critical model compression technique, often categorized into multi-shot and one-shot methods based on pruning frequency. Traditional iterative pruning methods, based on the Lottery Ticket Hypothesis [6], involve an iterative prune-and-finetune process to gradually identify a sparse subnetwork that performs comparably to the original model, thereby reducing the number of model parameters. Classic methods like magnitude pruning [8] and more recent approaches such as LLM-Pruner [18], which utilizes block pruning with LoRA [11] fine-tuning, adhere to this hypothesis. However, for large models, global fine-tuning is resource-intensive, demanding significant GPU memory. Although LLM-Pruner reduces memory requirements using LoRA, it still requires storage for intermediate results and gradients, while methods like LLM Surgeon [30], despite avoiding global fine-tuning, also demand substantial memory and time due to the use of second-order derivatives. In contrast, one-shot pruning methods have gained traction by minimizing memory usage during fine-tuning. SparseGPT [7] avoids global fine-tuning by compensating for precision within each basic layer, thus accelerating the pruning process. Wanda [27] further simplifies this by eliminating compensation altogether, using a unique significance score to achieve performance comparable to SparseGPT without the need for fine-tuning. PruneZero[4] constructs an operator search space by combining input, output, and gradient operations to optimize model metrics. While, it requires multiple full inferences across models, which is computationally expensive. In memory-constrained settings, chunked model loading further exacerbates the inference time. + +However, both SparseGPT and Wanda have notable limitations. (i) While these methods perform well at lower + +sparsity levels (10%-50%) when pruning large language models, their effectiveness significantly declines at mid-to-high sparsity levels (50%-70%). (ii) SparseGPT and Wanda show poor performance on vision models. Wanda, in particular, relies on priors that are specific to language models, making it even less effective than SparseGPT for vision tasks. + +![](images/e39f481947dbafb8d0c3ad62fe3b5e72ca9d2735a2e7a91c33e4dbeea7f3abbc.jpg) +(a) SparseGPT +Figure 1. Error propagation with respect to layer index. + +![](images/7aa0a1012d41d66ec91272005f6e030f48648c9e1d9a1d6f47a547a93f0b0987.jpg) +(b) Wanda + +To illustrate why we focus on mid-to-high sparsity levels, we present the error accumulation during the pruning process of OPT-1.3B using SparseGPT and Wanda in Fig.1a and Fig.1b, with errors measured by mean square error (MSE). As shown, both SparseGPT and Wanda prune blocks sequentially without inter-block interaction, leading to faster error accumulation in deeper layers, which aligns with intuitive expectations. Furthermore, the error accumulation trends of semistructured sparsity 2:4 and 4:8 differ from those of unstructured sparsity $50\%$ , instead aligning more closely with higher sparsity levels. Fig.1a, indicates that incorporating quantization strategies further shifts the error trends toward those of higher sparsity levels. Therefore, mitigating inter-block error propagation at sparsity levels of $50\% -70\%$ can effectively enhance the performance of unstructured pruning and hybrid compression strategies. + +In this paper, we propose ICP, a large model pruning method that is applicable to both vision and language models. Improve performance at both low- and high-sparsity levels in pruning large language and vision models, without increasing GPU memory usage and pruning time. Specifically, we design a sliding window-based inter-block error compensation method based on the principle of error propagation. As shown in Fig. 2, after pruning each block, the subsequent block compensates for the pruning-induced errors of the previous block before pruning itself. This approach ensures that errors introduced during pruning are compensated for in the next block, preventing error propagation to deeper layers and avoiding the need for full-model fine-tuning during the pruning process. We highlight our contributions as follows: + +1. We introduce the concept of inter-block error compensation, demonstrating that this approach can enhance the performance of one-shot pruning methods with limited + +![](images/d17cf899e889fca9ac16c06445f3dd6dd8b5b904d568d46f0bcdcf7847d90b0d.jpg) +Figure 2. Comparison between our proposed method ICP and conventional pruning methods. (a) illustrates a standard pruning approach, where all blocks are pruned simultaneously. (b) demonstrates our method. After pruning one block, the subsequent block adjusts to compensate for the error introduced by the pruning of the previous block. + +computational resources. + +2. Building on this concept, we further design a sliding window-based inter-block error compensation strategy. After pruning each layer, we use the next unpruned layer to compensate for the errors induced by pruning the current layer. It improves post-pruning model performance in both vision and language models. +3. Extensive experiments on large language models and vision models show that our method significantly enhances post-pruning model performance without increasing the GPU memory consumption. + +# 2. Related Work + +# 2.1. Transformer Architecture + +Since its introduction [31], Transformer has become the cornerstone of language models due to its superior performance in language tasks. Following the development of GPT [24], researchers have built increasingly large models using Transformer as primary building blocks. The Transformer has also proven effective in vision tasks. For example, SAM [13], which is based on the Vision Transformer, particularly excels in segmentation and serves as a foundational model for various downstream tasks [20, 23, 34]. + +Both language and vision models utilize a backbone composed of multiple Transformer blocks. The majority of parameters in these models are concentrated within their respective backbones. These Transformer blocks consist of multi-head self-attention and feed-forward neural networks with two linear layers. Multi-head self-attention is generally considered more critical than the feed-forward network. + +# 2.2. Large Model Pruning + +With the emergence of large models, the computational and memory costs of global fine-tuning made iterative pruning methods prohibitively expensive. The speed and lower computational resource requirements of one-shot pruning methods have led to the development of several such techniques for large models, notably SparseGPT and Wanda. SparseGPT [7] is a one-shot rapid pruning method belonging to Optimal Brain Surgeon [10] family. It independently analyzes each basic layer of the model, formulates the pruning problem as a constrained optimization problem, and solves the pruned weight values using the Lagrange multiplier method. SparseGPT builds on the OBS foundation by leveraging the property of Cholesky decomposition to efficiently obtain different orders of Hessian matrices, significantly accelerating the pruning process and reducing memory usage. Essentially, SparseGPT consists of two parts: pruning and compensation, which are integrated during the Lagrange multiplier method's solution process. However, SparseGPT compensates for pruning errors by adjusting unpruned weights within each basic layer, without considering inter-block weight interactions. Wanda [27], introduces a strong prior for language models, assuming that weight importance is related to both the absolute value of the weight and the corresponding output. Wanda designs an importance metric that achieves good results without compensating for pruning losses. However, subsequent studies [33] suggest that while Wanda performs well in many cases, it can occasionally fall short compared to SparseGPT, and its performance in vision models is particularly lacking, where its strong priors may lead to suboptimal results. + +# 3. Method + +Our objective is to develop an efficient pruning method for large models, including both vision and language models which should be executable within a reasonable time frame, even under limited computational and memory resources. Additionally, it should sustain robust performance at mid-to-high sparsity levels. In this section, we introduce ICP, a pruning strategy that effectively mitigates performance degradation at mid-to-high sparsity levels and facilitates rapid pruning in memory-constrained environments. The following sections detail our method in two aspects: Blockwise Compensate Pruning and Sparsity Rearrangement. + +# 3.1. Block-wise Compensate Pruning + +We propose Block-wise Compensate Pruning, which enhances inter-block error compensation by building upon SparseGPT. As depicted in Fig. 3(a), Block-wise Compensate Pruning leverages the modular structure of large models, processing each block sequentially using a sliding window approach. In simple terms, the parameters of the next + +block are updated to compensate for the pruning-induced errors in the preceding block. Initially, the sliding window contains two blocks. Since the operations are consistent after each slide, we explain using the first two blocks. + +In Fig. 3(a), the sliding window moves from left to right with a window length encompassing two blocks. Initially, the window contains $B^{1}$ and $B^{2}$ . Pruning is first applied to $B^{1}$ , followed by compensating for the pruning-induced error in $B^{1}$ using $B^{2}$ . The detailed calculation procedure for this compensation is outlined in Algorithm 1, as follows: The input data from the calibration set passes through the position encoding module, producing an intermediate result, $z^{0}$ , which is then duplicated to create $z^{0'}$ . At this stage, the feature flow splits into two streams: the "standard stream" $z^{0'}$ and the "error stream" $z^{0}$ . Before pruning $B^{1}$ , $z^{0'}$ passes through $B^{1}$ , producing an unbiased output, $z^{1'}$ . SparseGPT is then applied to prune $B^{1}$ , and $z^{0}$ passes through the pruned $B^{1}$ , resulting in a biased output, $z^{2}$ . Subsequently, $z^{1'}$ passes through the unpruned $B^{1}$ , yielding an unbiased output, $z^{2'}$ . Here, $z^{1}$ serves as both the output of the pruned $B^{1}$ and the input to $B^{2}$ . Since $z^{1}$ is biased, the output from $B^{1}$ will also be biased, whereas the desired output is $z^{2'}$ . Therefore, $z^{2'}$ serves as the target for $z^{1}$ , forming a labeled feature dataset ( $z^{1}, z^{2'}$ ). To align output of $B^{1}$ with $z^{2'}$ , we adjust $B^{1}$ 's parameters by training it on this feature dataset, reducing the output bias. Finally, the sliding window advances by one block, and the process is repeated: prune the first block and adjust the parameters of the second block using newly formed feature dataset. As above, Block-wise Compensate Pruning progresses sequentially along the backbone, processing one block at a time. + +Note that this method operates with a sliding window containing only two blocks, using the second block to compensate for the pruning-induced errors in the first, rather than relying on additional blocks for error correction. In this setup, once the first block is pruned, the standard stream dataset can pass through the second block to generate labels for the error stream. These labels are then reused during the compensation phase of the next sliding window, eliminating the need to store an additional block for real-time label computation. Consequently, at any given time, at most the weights of a single block reside in memory, optimizing both storage and computation. By contrast, if compensation required two subsequent blocks, the process would either necessitate storing two states of the standard stream or duplicating the next two blocks as a teacher model for distillation-like compensation, both of which would substantially increase memory overhead. It only requires a single block to be loaded into memory at a given moment. As each block is loaded into memory and pruned, the previous block is unloaded before the next block is loaded. This approach significantly reduces memory usage during the pruning and compensation process. + +![](images/dd8117578032474846e58faa2ac98734446693915b595917a331516198257a04.jpg) + +![](images/6ace0e570d1c8586525828bd982e3a6efa6a21802c299a5664646228f474203e.jpg) +Figure 3. Overview of the proposed method. (a) Block-wise Compensate Pruning for sequential block pruning and compensation; (b) basic structure of a transformer block (showing only weight-intensive and essential layers); + +# Algorithm 1 Block-wise Compensate Pruning + +Input: Backbone of large model $\{B^i\}_{i = 1}^n$ , i-th block $B^{i}$ , Pruning function Prune (module, hyperparameters), Calibration dataset $x_{j}$ , Position embedding block $PE$ Update function Update (module,(sample, label)), Sparsity s, Sparsity Rearrangement parameter $\alpha ,\beta$ +Output: Pruned backbone $\{B^i\}_{i = 1}^n$ +1: $z^0 = PE(\{x_j\})$ +2: $z_{i}^{0^{\prime}} = z_{i}^{0} / /$ Copy embeddings +3: $z_{j}^{1'} = B^{0}\left(z_{j}^{0'}\right) // \text{Preparation}$ +4: for $i$ in 1 to $n - 1$ do +5: $B^{i} = \text{Prune}(B^{i}, (s, \alpha, \beta))$ +6: $z_{j}^{i} = B^{i}\left(z_{j}^{i - 1}\right) / /$ Forward Propagation +7: $z_{j}^{i + 1,i^{\prime}} = B^{i + 1}\left(z_{j}^{i^{\prime}}\right)$ +8: $B^{i+1} = \text{Update}\left(B^{i+1}, z_{j}^{i}, z_{j}^{i+1'}\right)$ +9: end for +10: $B^n = \text{Prune}(B^n, (s, \alpha, \beta))$ +11: return Pruned backbone $\{B^i\}_{i=1}^n$ + +# 3.2. Sparsity Rearrangement + +In many pruning methods, the sparsity of different parts of the model can be adaptively determined through specific rules. However, since our method is based on a sliding window approach for pruning, it is difficult to obtain information about the blocks that are outside the current window. If we attempt to acquire global information, we would either incur significant I/O overhead from loading model segments sequentially or face high memory usage from loading the entire model at once. Some methods even evaluate importance based on gradients, which can add further demands on gradient storage. + +To avoid these issues, fast pruning methods like SparseGPT or Wanda typically assign the same sparsity + +level uniformly across all parts of the model, equivalent to the target sparsity. However, this approach is clearly suboptimal. Certain prior knowledge can be leveraged to adjust the sparsity distribution, thereby avoiding these problems while improving performance. + +# 3.2.1. Inter-Block Sparsity Rearrangement + +Block-wise Compensate Pruning fine-tunes $B^{i+1}$ to compensate for the pruning-induced errors in $B^i$ . However, when pruning the last block, $B^n$ , there is no $B^{n+1}$ available to compensate for the errors in $B^n$ . Therefore, assigning the same number of pruned weights to each block is not optimal. Instead, the pruning distribution should be adjusted across different blocks. To maintain the global target sparsity, we rearrange the sparsity between blocks as follows: + +$$ +P _ {B ^ {n}} = \left| \theta^ {n} \right| \times s \times \alpha , \tag {1} +$$ + +$$ +P _ {B ^ {j}} = \left| \theta^ {j} \right| \times s + \frac {\left| \theta^ {n} \right| \times s \times (1 - \alpha)}{n - 1} \tag {2} +$$ + +where $j = 1, \ldots, n - 1$ , $\theta^j$ denotes the parameters in $B^j$ , $|\theta|$ denotes the number of weights in a single block, and $\alpha$ is a super-parameter that redistributes a portion of the pruned weights from the last block evenly across all preceding blocks. + +# 3.2.2. Intra-Block Sparsity Rearrangement + +As shown in Fig. 3(b), the majority of the parameters within a Transformer block are contributed by the matrices depicted, which are clearly hierarchical. Therefore, a block can essentially be viewed as a nested function: the QKV matrices represent the innermost function, while the second + +fully connected layer (fc2) represents the outermost function. According to the conclusions drawn in the Theoretical Foundations section, if the same sparsity level is applied uniformly across all layers, the errors introduced by the QKV matrices, which are further from the output of $B^{i}$ , will be more challenging to compensate for in $B^{i + 1}$ . On the other hand, fc2 is closer to the output of $B^{i}$ , making its errors easier to compensate for in $B^{i + 1}$ . Additionally, inspired by previous works, earlier layers tend to converge first [1], and faster convergence often leads to better generalization performance [9]. Therefore, the sparsity distribution should be rearranged by reducing the sparsity of the QKV matrices and increasing the sparsity of fc2. To ensure that the global target sparsity remains unchanged, we rearrange the sparsity according to the following rule: + +$$ +P _ {t} = s \times | W _ {t} ^ {i} | \times (1 - \beta), t \in \{Q, K, V \}. (3) +$$ + +$$ +P _ {f c 2} = s \times | W _ {f c 2} ^ {i} | + s \times | W _ {f c 2} ^ {i} | \times \beta . \tag {4} +$$ + +where $P_{t}$ represents the number of pruned weights, $|W|$ denotes the total number of weights in a single layer, and $\beta$ is a super-parameter that needs to be adjusted. + +# 4. Experiments + +# 4.1. Experimental Setup + +To evaluate the effectiveness of our method, we focus on models with 7 billion parameters or fewer, as these are the most commonly used in research and resource-constrained scenarios. Pruning these models is also the most feasible on a single desktop-class GPU. Specifically, we validate our approach using the OPT-125M/1.3B/2.7B/6.7B and Llama 2-7B, as well as the SAM-B/L/H. All experiments were conducted on a single NVIDIA RTX 3090 GPU. + +For the compensation process, we employ the Adam optimizer with a Mean Squared Error (MSE) loss function and a batch size of 1. All compensation operations during pruning share a unified learning rate, which is adjusted using the ExponentialLR scheduler with a decay factor of 0.87. + +For calibration data, language model calibration dataset follows SparseGPT, using 128 segments of 2048 tokens, randomly sampled from C4 dataset [25]. Similarly, the vision model calibration dataset consists of 128 images, randomly sampled from the second compressed file of the SA-1B dataset [13] provided by Meta. + +To comprehensively evaluate the performance of pruned models, we validate OPT using the Wikitext-v2 [21] and PTB datasets [19] from Hugging Face, with perplexity (PPL) as the evaluation metric. We also evaluate the pruned Llama 2-7B model using accuracy on seven ZeroShot tasks: BoolQ [2], RTE [32], HellaSwag [35], Wino-Grande [26], ARC Easy and Challenge [3], and Open-bookQA [22], following the experimental setup introduced + +by Sun et al. [27] in their work on Wanda. This allows for a more task-specific evaluation of the pruned model's generalization capabilities. For SAM, we validate the performance using the fourth compressed file of the SA-1B dataset as well as the COCO dataset [15], employing Intersection over Union (IoU) as the primary metric for visual models. + +# 4.2. Unstructured Pruning + +Language Models. In Table 1 and Table 2, we report the perplexity performance of the OPT series models on Wikitext and PTB datasets after pruning to various sparsity levels using our method and other baseline approaches. Our method demonstrates superior performance over baselines even at lower sparsity levels, and as sparsity increases, the performance gap widens in favor of our approach. Additionally, as model size decreases, the advantage of our method over the baselines becomes even more pronounced. Table 3 also presents results in comparing the performances of our proposed method against those of baselines on the Llama 2-7B model at $50\%$ and $70\%$ sparsity levels. Our method consistently outperforms all baseline approaches at both sparsity levels, achieving the best results across all tasks. These findings highlight the effectiveness of the interblock error compensation pruning method introduced in this paper, particularly under high-sparsity conditions, across different model scales, and for various model architectures. + +Table 1. Perplexity performances of pruned OPT models of different scales at various sparsity levels. The models are calibrated on a subset of the C4 and evaluated on full Wikitext-v2. + +
MethodModel10%20%30%40%50%60%70%80%90%
Magnitude125M28.2529.6534.5154.60193.4920.0380648906614
SparseGPT27.8827.9828.8630.5637.0259.55220.923784666
Wanda27.4827.7628.1130.6738.9275.17328.219184609
ICP (ours)27.3727.8728.6629.6633.9045.7865.20270.41612
Magnitude1.3B14.7215.6224.74388.01713939294431634428871
SparseGPT14.6714.7515.1616.4917.5022.0851.75752.46797
Wanda14.6314.6915.0115.8918.4126.5599.53225816868
ICP (ours)14.5914.6014.6715.0416.4921.1736.78117.11099
Magnitude2.7B12.5913.1315.5830.32265.236047251961416668
SparseGPT12.3212.3912.6612.6613.4616.0426.92138.45818
Wanda12.2312.2612.3612.8614.2119.83387.9590516527
ICP (ours)12.2212.2512.3412.6213.3215.7721.7954.67764.2
Magnitude6.7B10.9211.2712.5331.89968.71263916975255915297
SparseGPT10.8310.7610.7510.9611.5913.4820.4896.2511938
Wanda10.7310.6310.6410.9611.9815.19159.2395416076
ICP (ours)10.5610.5810.5910.8511.2713.0317.8244.691469
+ +Vision Models. In Tables 4 and Table 5, we report IoU performance of SAM at varying sparsity levels on SA-1B and COCO datasets with single-point interaction. Following experimental setup, SAMs are pruned with only 128 images random sampled from SA-1B as calibration data. The experiments first verify that SparseGPT indeed achieves strong IoU performance on visual models under low sparsity conditions, outperforming both Magnitude and Wanda. + +Table 2. Perplexity performancese of pruned OPT models of different scales at various sparsity levels. The models are calibrated on a subset of the C4 and evaluated on PTB. + +
MethodModel10%20%30%40%50%60%70%80%90%
Magnitude125M39.9743.2549.9782.69276.21146343041245382
SparseGPT39.4340.3642.7145.2455.7888.44259.629195374
Wanda39.0539.8341.4344.4157.59109.7399.722144474
ICP (ours)39.1239.1340.9643.4749.4164.1088.85336.33075
Magnitude1.3B20.6621.5139.89823.53171949984671728333126
SparseGPT20.3420.6421.5023.1725.4733.4876.91531.87515
Wanda20.3820.7021.4123.2528.0043.95135.3180912025
ICP (ours)20.1320.4121.1122.1924.6832.1954.33148.83094
Magnitude2.7B18.3819.0622.9646.36262.632077276957512178
SparseGPT18.1518.5018.9818.9820.4424.6742.75146.06107
Wanda17.9718.0418.3219.2521.8633.44450.8546221039
ICP (ours)17.9718.0218.2918.8620.0522.9235.1771.912357
Magnitude6.7B16.1816.7018.5248.85613.493799973157424260
SparseGPT15.7715.8115.9216.2817.4320.2331.86105.08320
Wanda15.7615.7715.9416.4317.9223.61211.3302414998
ICP (ours)15.7615.7415.9416.2217.3020.0728.0464.182549
+ +Table 3. Zero-shot performance (Accuracy, %) of Llama 2-7B pruned by different method on various tasks at sparsity of $50\%$ and $70\%$ . + +
MethodStrategyBQRTEHSWGARC-eARC-cOBQAMean
Dense077.7463.1857.1068.9876.2643.4331.4059.73
Magnitude50%62.9457.0449.1263.3864.0634.6426.8051.14
SparseGPT76.3355.9652.8968.9872.0138.2328.4056.11
Wanda76.4253.4352.4568.6772.2239.3331.0056.22
ICP (ours)76.5057.0453.0369.1172.3940.4431.0056.96
Magnitude70%37.9553.0725.9349.2527.8222.8717.0033.41
SparseGPT64.6553.7933.4558.6443.3121.6717.0041.79
Wanda48.4752.7127.9749.6430.8118.6012.2034.34
ICP (ours)67.7155.2341.4560.2255.5627.0521.4046.94
+ +As Wanda is based on strong priors from language models, its underperformance compared to SparseGPT on visual models is within our expectations. Lastly, the method proposed in this paper outperforms all baselines and exhibits a similar trend observed in language model experiments: the higher the sparsity, the greater the advantage of our method over the baseline methods. This demonstrates that the proposed method is not only effective for language models but also remains effective for visual models. + +# 4.3. Semi-Structure Pruning and Quantization + +On certain specialized hardware, such as specific NVIDIA GPUs or other custom-designed computing architectures, the combination of pruning and quantization can further accelerate inference while maintaining an acceptable performance degradation. To explore this, we conduct tests using a joint compression strategy that combines pruning and quantization. Specifically, we focus on the combination of 2:4 and 4:8 semi-structured sparsity with 3-bit and 4-bit quantization. Since Wanda does not support quantization, we exclude it from experiments involving quantization and instead compared our approach against SparseGPT. + +Table 4. Instance segmentation performance of SAM models (IoU, $\%$ ) at various sparsity using single-point interaction on SA-1B. The models are calibrated on a subset of 128 images sampled from SA-1B. + +
MethodModel10%20%30%40%50%60%70%80%90%
MagnitudeSAM-H74.8074.5773.8772.2669.4561.5128.222.651.70
SparseGPT74.8174.7774.6074.2073.3071.1565.6950.9718.15
Wanda74.8274.7674.3573.6071.6567.6553.0021.841.20
ICP (ours)74.8274.7874.6674.6074.4273.8272.8569.7357.60
MagnitudeSAM-L74.8174.5973.9372.2869.1862.5645.529.924.52
SparseGPT74.8474.7774.6174.1373.2070.7064.4545.487.37
Wanda74.8674.7874.6274.0572.5168.2855.1322.811.31
ICP (ours)74.8574.7974.6974.5574.2773.6472.2668.1753.68
MagnitudeSAM-B72.6972.5671.9970.3667.3860.8746.0125.531.51
SparseGPT72.7272.6772.5072.0871.1168.6863.0247.4918.81
Wanda72.7172.6472.3371.4569.4464.3550.2928.162.18
ICP (ours)72.7272.6972.5672.2871.8571.0669.3364.7851.82
+ +Table 5. Instance segmentation performance of SAM models (IoU, %) at various sparsity using single-point interaction on COCO. + +
MethodModel10%20%30%40%50%60%70%80%90%
MagnitudeSAM-H71.1271.1170.8670.6169.3763.6032.687.887.21
SparseGPT71.1171.1870.9870.9370.4269.5266.7056.8227.57
Wanda71.1771.0170.9870.7369.6366.0553.4522.835.51
ICP (ours)71.1471.1871.1270.9870.8170.2469.4366.6958.41
MagnitudeSAM-L70.5670.5170.2769.5368.0862.6649.1220.6618.12
SparseGPT70.5570.5370.4370.0469.6168.2164.4248.3511.30
Wanda70.5670.5870.4569.9168.6565.5654.6719.896.38
ICP (ours)70.5570.5870.5870.3769.9169.5168.1465.4256.92
MagnitudeSAM-B66.8467.0667.0666.6765.2962.2853.6140.867.48
SparseGPT67.0967.0266.9866.7666.1765.0062.0352.4628.51
Wanda66.9466.9767.0466.4765.3862.3153.5839.327.89
ICP (ours)67.1067.0667.0466.8366.4566.0764.5162.4054.50
+ +Table 6. Perplexity performance of pruned OPT models at various compression strategies. + +
MethodStrategyOPT-125MOPT-1.3BOPT-2.7BOPT-6.7B
WTPTBWTPTBWTPTBWTPTB
SparseGPT2:459.3593.0423.8738.0917.1126.9814.1521.54
Wanda2:480.01111.9728.2043.3721.1834.5515.9025.09
ICP (ours)2:446.9067.5622.0333.9516.3925.3114.0421.19
SparseGPT4:843.9071.9920.2131.3615.0023.0512.4918.86
Wanda4:853.1879.0222.1734.5616.7926.1313.5520.19
ICP (ours)4:839.9159.3419.1530.4114.8822.8812.3818.53
SparseGPT2:4+int3122.33196.4440.5864.5324.7041.0620.0731.91
ICP (ours)2:4+int374.07114.9327.8846.1419.5730.2818.3826.82
SparseGPT2:4+int470.88107.3925.9641.6918.5630.6815.2423.01
ICP (ours)2:4+int451.2475.3923.4934.4217.0527.0415.1022.54
SparseGPT4:8+int384.02120.1632.4149.2720.2432.9917.0826.36
ICP (ours)4:8+int359.6786.0124.7340.1917.8926.8016.1624.30
SparseGPT4:8+int450.5880.9723.3234.3815.7424.7613.4419.72
ICP (ours)4:8+int442.0360.6020.8631.5815.3323.8013.2419.49
+ +# 4.4. Sparsity Rearrangement Ablation + +To validate the Sparsity Rearrangement method based on priors, as proposed in the Method section, we conducted experiments on both Inter-Block Sparsity Rearrangement and Intra-Block Sparsity Rearrangement. + +Inter-Block Sparsity Rearrangement. In Fig. 4a and Fig. 4b, we present the performance of OPT-125M and + +Table 7. Zero-shot Performance (Accuracy, %) of Llama 2-7B Model under various compression strategies. + +
MethodStrategyBQRTEHSWGARC-eARC-cOBQAMean
SparseGPT2:468.0758.8443.3865.5964.1831.9124.8050.97
Wanda2:468.1353.4341.4262.4363.1330.6323.8049.00
ICP (ours)2:469.7258.8446.7564.4866.0433.7026.2052.25
SparseGPT4:871.0754.8748.3567.8068.6934.4727.2053.21
Wanda4:873.0053.7947.0366.9367.4734.2227.0052.78
ICP (ours)4:873.1258.1250.1364.8868.9036.0128.0054.16
SparseGPT2:4+int364.6555.2338.5859.0455.4725.3419.4045.39
ICP (ours)2:4+int366.8254.1543.3660.7759.5528.6721.0047.76
SparseGPT2:4+int466.7354.1542.3465.1961.7428.9223.8048.98
ICP (ours)2:4+int470.4656.6846.2262.9064.8130.6322.8050.64
SparseGPT4:8+int368.5953.4342.1262.3560.1029.1021.6048.18
ICP (ours)4:8+int368.3854.8743.5462.8563.3832.3422.4049.68
SparseGPT4:8+int469.2754.1546.5866.0667.5534.3927.2052.17
ICP (ours)4:8+int472.2961.0148.8662.7566.8833.8728.0053.38
+ +Table 8. Instance segmentation performance of SAM models (IoU, $\%$ ) at various compression strategies using single-point interaction on SA-1B and COCO. + +
MethodStrategySAM-BSAM-LSAM-H
COCOSA-1BCOCOSA-1BCOCOSA-1B
SparseGPT2:462.8067.5965.0169.0767.0269.62
Wanda2:460.5962.0662.3765.8464.6866.36
ICP (ours)2:464.9571.3067.8373.5669.0973.64
SparseGPT4:863.9369.6966.5671.2767.9871.41
Wanda4:862.6266.0565.0269.7566.8069.55
ICP (ours)4:865.3771.7268.1773.8669.4474.03
SparseGPT2:4+int362.0964.7763.4565.6766.3967.85
ICP (ours)2:4+int364.4769.7167.1372.6068.7573.26
SparseGPT2:4+int462.6666.7664.7868.6966.8169.44
ICP (ours)2:4+int464.8970.6467.6473.0969.0273.72
SparseGPT4:8+int363.0267.2764.7468.0267.2569.69
ICP (ours)4:8+int364.8469.8667.6873.2869.1073.35
SparseGPT4:8+int463.7768.9866.0470.3167.8771.05
ICP (ours)4:8+int465.2470.9868.0273.8269.3473.78
+ +SAM-B pruned by ICP with different $\alpha$ . For clarity in visual presentation, we conducted experiments with $70\%$ sparsity on OPT-125M and $90\%$ sparsity on SAM-B. We observed that in the language model, $\alpha$ had the greatest impact on the performance of OPT-125M on the PTB dataset, and adjusting $\alpha$ further improved the model's performance on non-calibration datasets. On vision models, the influence of $\alpha$ on performance was similar on both datasets, with adjustments to $\alpha$ enhancing the overall performance of the visual model. Overall, appropriate tuning of $\alpha$ can further improve the performance of pruned models. Based on experiments, we conclude that $\alpha$ typically yields optimal results within the range of [0.6, 0.75] for language models, and within [0.8, 0.9] for vision model. + +Intra-Block Sparsity Rearrangement. In Fig. 4c, we illustrate the impact of different $\beta$ on the language model, using the $70\%$ sparsity model of OPT-125M as an example. We observed that in the language model, adjusting $\beta$ can further reduce the model's perplexity. Unlike $\alpha$ , $\beta$ not + +only improves the model's performance on non-calibration datasets but also enhances its performance on calibration datasets. Overall, appropriately adjusting $\beta$ can further improve the performance of the pruned model. The optimal value for $\beta$ typically falls within the range of $[0.6 \times (1 - s), 0.8 \times (1 - s)]$ , where $s$ denotes sparsity. In Fig. 4d, we exemplify the effects of varying $\beta$ on the visual model, specifically using the $90\%$ sparsity model of SAM-B. Our findings indicate that the influence of $\beta$ on the visual model is minimal; both on calibration and non-calibration datasets, adjustments to $\beta$ yield only marginal improvements in the model's performance. + +# 4.5. Calibration Set Size Ablation + +Although our experiments follow the settings of SparseGPT and Wanda, using 128 randomly selected samples for pruning both language and vision models, Fig. 4e and Fig. 4f show the impact of calibration set size on our method's performance. Fig. 1 presents results on the OPT-125M language model at $70\%$ sparsity, while Fig. 2 shows results for the SAM-B vision model at $90\%$ sparsity. In both cases, increasing the sample size beyond 128 offers minimal performance gains while significantly increasing GPU memory usage. Thus, like SparseGPT and Wanda, 128 samples remain the optimal choice. + +# 4.6. Compensation Epoch Ablation + +In Fig.4g and Fig.4h, we illustrate the impact of varying the number of epochs during the compensation on the performance of both vision and language models. It is evident that a single epoch is sufficient to effectively mitigate the error introduced by pruning in earlier layers. While additional epochs can further improve performance, the marginal gains diminish rapidly. Although we report results using 12 epochs in the experiments presented in 4.2 and 4.3 to demonstrate the best achievable performance, selecting a single epoch offers the optimal trade-off between model performance and computational overhead. + +# 4.7. Memory and Time Efficiency + +Peak memory usage (PMU) and execution time are critical factors in determining whether a pruning method can be deployed on devices with limited memory capacity and lower cost. In Table 9, we present the PMU and runtime of SparseGPT, Wanda, and our proposed method when pruning OPT-6.7B on an NVIDIA RTX 3090. The experimental setup is consistent with the description in 4.1. Our method employs a commonly used chunked storage strategy, where intermediate outputs of each layer are stored in host memory instead of GPU memory, effectively reducing memory usage. While this approach may slow down the pruning process, as discussed in Section 4.6, a single epoch of fine-tuning is sufficient to significantly improve + +![](images/cc8dfd7b1ba0b437667dde1e01afc20843dd57529ecb60469c0b5d7fbef34424.jpg) +(a) Perplexity performance of pruned OPT-125M with different $\alpha$ + +![](images/a9cc02cc1bfd6a93eb4d02ce06f676f906372b78ba0d01ecab5ceefbd16fb5f1.jpg) +(b) Instance segmentation performance (IoU, %) of SAM-B with different $\alpha$ . + +![](images/bf9c50dd352ccf6218444dfd3632170d98261fe55abf943d3d2d3ab663dfb6ac.jpg) +(c) Perplexity performance of OPT-125M with different $\beta$ + +![](images/7a96db60711f9674db4441172da3117c8ce3d675ec2542ec1fbd3e190815335e.jpg) +(d) Instance segmentation performance (IoU, %) of SAM-B with different $\beta$ . + +![](images/25bbdfb3d8e95c75b128bb3808b9ee70688a98c43dc6d9f455e0feb7882a7174.jpg) +(e) Perplexity performance of OPT-125M with different calibration set size. + +![](images/80b480ac31d38361ce2a816ab553da4a67ef8bb15b7edcbfe762295bca57ebce.jpg) +(f) Instance segmentation performance (IoU, %) of SAM-B with different calibration set size. + +![](images/797e9a14b85ec63bf6edc8cd84fc204df5ae975d440c44637518be3aad9d6e4c.jpg) +(g) Perplexity performance of pruned OPT-125M with different number of compensation epochs. + +![](images/a30956c31be2c5da3c3603cb57eef34222fb9fc886497ad873d9e7b4d78e779d.jpg) +(h) Segmentation performance of SAM-B with different number of compensation epochs. +Figure 4. Performance metrics for pruned OPT-125M and SAM-B models under varying parameters $\alpha$ , $\beta$ , calibration dataset sizes, and number of compensation epochs. + +model performance. Consequently, data only needs to be loaded from host memory once, keeping the additional time overhead within an acceptable range. It is worth noting that we use full-precision fine-tuning rather than Parameter-Efficient Fine-Tuning (PEFT) techniques. If combined with PEFT methods, memory consumption could be further reduced. This advantage enables our method to efficiently prune larger models on desktop-class GPUs, thereby expanding its applicability. + +Table 9. GPU Memory Consumption Comparison + +
MethodPMUTime(s)
SparseGPT8.3 GB1004
Wanda21 GB143
ICP (1-epoch)7.8 GB1551
ICP (12-epoch)7.8 GB4470
+ +# 5. Conclusion + +In this paper, we propose an efficient pruning method ICP, which is effective for both vision and language models while consuming limited GPU memory. ICP update each block individually by compensating for the inter-block errors introduced by pruning, then rearranges the sparsity distribution based on prior knowledge of Transformer and pruning method. It significantly improves the performance of models pruned at high sparsity levels without increasing GPU memory consumption. ICP, as it does not rely on strong priors specific to any particular model, demonstrates dual applicability to both vision and language models. This flexibility suggests that it could potentially be extended to multimodal models in the future. However, this work has limitations: (1) the hyperparameters $\alpha$ and $\beta$ are not automatically tuned; (2) ICP has not been tested on multimodal models; (3) the potential of individual blocks has not been fully explored. These will be the focus of our future work. + +# Acknowledgement + +This work is supported by Natural Science Foundation of China under Grant 62271465, Suzhou Basic Research Program under Grant SYG202338, and Jiangsu Province Science Foundation for Youths (NO. BK20240464). + +# References + +[1] Yixiong Chen, Alan Yuille, and Zongwei Zhou. Which layer is learning faster? a systematic exploration of layerwise convergence rate for deep neural networks. In The Eleventh International Conference on Learning Representations, 2023. 5 +[2] Christopher Clark, Kenton Lee, Ming-Wei Chang, Tom Kwiatkowski, Michael Collins, and Kristina Toutanova. Boolq: Exploring the surprising difficulty of natural yes/no questions. arXiv preprint arXiv:1905.10044, 2019. 5 +[3] Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. Think you have solved question answering? try arc, the ai2 reasoning challenge. arXiv preprint arXiv:1803.05457, 2018. 5 +[4] Peijie Dong, Lujun Li, Zhenheng Tang, Xiang Liu, Xinglin Pan, Qiang Wang, and Xiaowen Chu. Pruner-zero: Evolving symbolic pruning metric from scratch for large language models. arXiv preprint arXiv:2406.02924, 2024. 1 +[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1 +[6] Jonathan Frankle and Michael Carbin. The lottery ticket hypothesis: Finding sparse, trainable neural networks. In International Conference on Learning Representations, 2018. 1 +[7] Elias Frantar and Dan Alistarh. Sparsegpt: Massive language models can be accurately pruned in one-shot. In International Conference on Machine Learning, pages 10323-10337. PMLR, 2023. 1, 3 +[8] Song Han, Jeff Pool, John Tran, and William Dally. Learning both weights and connections for efficient neural network. Advances in neural information processing systems, 28, 2015. 1 +[9] Moritz Hardt, Ben Recht, and Yoram Singer. Train faster, generalize better: Stability of stochastic gradient descent. In International conference on machine learning, pages 1225-1234. PMLR, 2016. 5 +[10] Babak Hassibi and David Stork. Second order derivatives for network pruning: Optimal brain surgeon. In Advances in Neural Information Processing Systems. Morgan-Kaufmann, 1992. 3 +[11] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 1 + +[12] Lei Ke, Mingqiao Ye, Martin Danelljan, Yu-Wing Tai, Chi-Keung Tang, Fisher Yu, et al. Segment anything in high quality. Advances in Neural Information Processing Systems, 36, 2024. 1 +[13] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 1, 2, 5 +[14] Chunyuan Li, Cliff Wong, Sheng Zhang, Naoto Usuyama, Haotian Liu, Jianwei Yang, Tristan Naumann, Hoifung Poon, and Jianfeng Gao. Llava-med: Training a large language-and-vision assistant for biomedicine in one day. Advances in Neural Information Processing Systems, 36, 2024. 1 +[15] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision - ECCV 2014, pages 740-755, Cham, 2014. Springer International Publishing. 5 +[16] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1 +[17] AI@Meta Llama Team. The llama3 herd of models, 2024. Accessed: 2024-08-01. 1 +[18] Xinyin Ma, Gongfan Fang, and Xinchao Wang. Llm-pruner: On the structural pruning of large language models. Advances in neural information processing systems, 36:21702-21720, 2023. 1 +[19] Mitch Marcus, Grace Kim, Mary Ann Marcinkiewicz, Robert MacIntyre, Ann Bies, Mark Ferguson, Karen Katz, and Britta Schasberger. The penn treebank: Annotating predicate argument structure. In Human Language Technology: Proceedings of a Workshop held at Plainsboro, New Jersey, March 8-11, 1994, 1994. 5 +[20] Maciej A Mazurowski, Haoyu Dong, Hanxue Gu, Jichen Yang, Nicholas Konz, and Yixin Zhang. Segment anything model for medical image analysis: an experimental study. Medical Image Analysis, 89:102918, 2023. 2 +[21] Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. Pointer sentinel mixture models. arXiv preprint arXiv:1609.07843, 2016. 5 +[22] Todor Mihaylov, Peter Clark, Tushar Khot, and Ashish Sabharwal. Can a suit of armor conduct electricity? a new dataset for open book question answering. arXiv preprint arXiv:1809.02789, 2018. 5 +[23] Quan Quan, Fenghe Tang, Zikang Xu, Heqin Zhu, and S Kevin Zhou. Slide-sam: Medical sam meets sliding window. In Medical Imaging with Deep Learning, 2024. 2 +[24] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018. 2 +[25] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 5 + +[26] Keisuke Sakaguchi, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. Winogrande: An adversarial winograd schema challenge at scale. Communications of the ACM, 64 (9):99-106, 2021. 5 +[27] Mingjie Sun, Zhuang Liu, Anna Bair, and J Zico Kolter. A simple and effective pruning approach for large language models. In The Twelfth International Conference on Learning Representations, 2024. 1, 3, 5 +[28] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1 +[29] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 1 +[30] Tycho FA van der Ouderaa, Markus Nagel, Mart Van Baalen, Yuki M Asano, and Tijmen Blankevoort. The IIm surgeon. arXiv preprint arXiv:2312.17244, 2023. 1 +[31] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 1, 2 +[32] Alex Wang. Glue: A multi-task benchmark and analysis platform for natural language understanding. arXiv preprint arXiv:1804.07461, 2018. 5 +[33] Miles Williams and Nikolaos Aletras. How does calibration data affect the post-training pruning and quantization of large language models? arXiv preprint arXiv:2311.09755, 2023. 3 +[34] Junde Wu, Wei Ji, Yuanpei Liu, Huazhu Fu, Min Xu, Yanwu Xu, and Yueming Jin. Medical sam adapter: Adapting segment anything model for medical image segmentation. arXiv preprint arXiv:2304.12620, 2023. 1, 2 +[35] Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. Hellaswag: Can a machine really finish your sentence? arXiv preprint arXiv:1905.07830, 2019. 5 +[36] Renrui Zhang, Jiaming Han, Chris Liu, Peng Gao, Aojun Zhou, Xiangfei Hu, Shilin Yan, Pan Lu, Hongsheng Li, and Yu Qiao. Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199, 2023. 1 +[37] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 1 \ No newline at end of file diff --git a/2025/ICP_ Immediate Compensation Pruning for Mid-to-high Sparsity/images.zip b/2025/ICP_ Immediate Compensation Pruning for Mid-to-high Sparsity/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a4fd80821acd46c69110d9046a35849cfb9870a6 --- /dev/null +++ b/2025/ICP_ Immediate Compensation Pruning for Mid-to-high Sparsity/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:48c99f46a19d1b48eae32958a9c06226db92f72b7826c444d8599c06f462e936 +size 934479 diff --git a/2025/ICP_ Immediate Compensation Pruning for Mid-to-high Sparsity/layout.json b/2025/ICP_ Immediate Compensation Pruning for Mid-to-high Sparsity/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4a3f7b9a98df5010b7fb9903f17ffbb08841d88e --- /dev/null +++ b/2025/ICP_ Immediate Compensation Pruning for Mid-to-high Sparsity/layout.json @@ -0,0 +1,8823 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 105, + 103, + 509, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 103, + 509, + 121 + ], + "spans": [ + { + "bbox": [ + 105, + 103, + 509, + 121 + ], + "type": "text", + "content": "ICP: Immediate Compensation Pruning for Mid-to-high Sparsity" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 118, + 141, + 493, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 141, + 493, + 157 + ], + "spans": [ + { + "bbox": [ + 118, + 141, + 493, + 157 + ], + "type": "text", + "content": "Xin Luo" + }, + { + "bbox": [ + 118, + 141, + 493, + 157 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 118, + 141, + 493, + 157 + ], + "type": "text", + "content": " Xueming Fu" + }, + { + "bbox": [ + 118, + 141, + 493, + 157 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 118, + 141, + 493, + 157 + ], + "type": "text", + "content": " Zihang Jiang" + }, + { + "bbox": [ + 118, + 141, + 493, + 157 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 118, + 141, + 493, + 157 + ], + "type": "text", + "content": " S.Kevin Zhou" + }, + { + "bbox": [ + 118, + 141, + 493, + 157 + ], + "type": "inline_equation", + "content": "^{1,2,3,4*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 100, + 157, + 510, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 157, + 510, + 171 + ], + "spans": [ + { + "bbox": [ + 100, + 157, + 510, + 171 + ], + "type": "text", + "content": "1 School of Biomedical Engineering, Division of Life Sciences and Medicine, USTC" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 140, + 171, + 485, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 171, + 485, + 184 + ], + "spans": [ + { + "bbox": [ + 140, + 171, + 485, + 184 + ], + "type": "text", + "content": "2 MIRACLE Center, Suzhou Institute for Advance Research, USTC" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 126, + 184, + 484, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 184, + 484, + 199 + ], + "spans": [ + { + "bbox": [ + 126, + 184, + 484, + 199 + ], + "type": "text", + "content": "3 Key Laboratory of Intelligent Information Processing of CAS, ICT, CAS" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 137, + 199, + 473, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 199, + 473, + 213 + ], + "spans": [ + { + "bbox": [ + 137, + 199, + 473, + 213 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 137, + 199, + 473, + 213 + ], + "type": "text", + "content": " State Key Laboratory of Precision and Intelligent Chemistry, USTC" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 171, + 215, + 439, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 215, + 439, + 228 + ], + "spans": [ + { + "bbox": [ + 171, + 215, + 439, + 228 + ], + "type": "text", + "content": "{luoxin.ustc, xuemingfu, jzh0103}@mail.ustc.edu.cn" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 241, + 229, + 370, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 229, + 370, + 241 + ], + "spans": [ + { + "bbox": [ + 241, + 229, + 370, + 241 + ], + "type": "text", + "content": "{s.kevin.zhou}@gmail.com" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 151, + 269, + 200, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 269, + 200, + 281 + ], + "spans": [ + { + "bbox": [ + 151, + 269, + 200, + 281 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 293, + 296, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 293, + 296, + 534 + ], + "spans": [ + { + "bbox": [ + 53, + 293, + 296, + 534 + ], + "type": "text", + "content": "The increasing adoption of large-scale models under 7 billion parameters in both language and vision domains enables inference tasks on a single consumer-grade GPU but makes fine-tuning models of this scale, especially 7B models, challenging. This limits the applicability of pruning methods that require full fine-tuning. Meanwhile, pruning methods that do not require fine-tuning perform well at low sparsity levels (10%-50%) but struggle at mid-to-high sparsity levels (50%-70%), where the error behaves equivalently to that of semi-structured pruning. To address these issues, this paper introduces ICP, which finds a balance between full fine-tuning and zero fine-tuning. First, Sparsity Rearrange is used to reorganize the predefined sparsity levels, followed by Block-wise Compensate Pruning, which alternates pruning and compensation on the model's backbone, fully utilizing inference results while avoiding full model fine-tuning. Experiments show that ICP improves performance at mid-to-high sparsity levels compared to baselines, with only a slight increase in pruning time and no additional peak memory overhead." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 555, + 135, + 568 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 135, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 135, + 568 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 575, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 575, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 575, + 295, + 696 + ], + "type": "text", + "content": "Introduction of Transformer [5, 31] propelled both language and vision models toward larger scales, leading to the development of high-performance models such as the OPT [37] and Llama [17, 28, 29] series for language tasks, and the SAM [13] series for vision tasks. However, their large scale presents significant deployment challenges. In scenarios with downstream tasks or hardware constraints, models under 8 billion (8B) parameters for language tasks and under 650 million parameters for vision tasks are most commonly used [12, 14, 16, 34, 36]. Nevertheless, model compres" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 270, + 553, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 553, + 317 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 553, + 317 + ], + "type": "text", + "content": "sion is necessary to further reduce their size for deployment. Moreover, on a single desktop-class GPU, it is typically only feasible to perform inference with these models, while fine-tuning remains impractical due to hardware limitations." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 318, + 555, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 318, + 555, + 688 + ], + "spans": [ + { + "bbox": [ + 313, + 318, + 555, + 688 + ], + "type": "text", + "content": "Pruning is a critical model compression technique, often categorized into multi-shot and one-shot methods based on pruning frequency. Traditional iterative pruning methods, based on the Lottery Ticket Hypothesis [6], involve an iterative prune-and-finetune process to gradually identify a sparse subnetwork that performs comparably to the original model, thereby reducing the number of model parameters. Classic methods like magnitude pruning [8] and more recent approaches such as LLM-Pruner [18], which utilizes block pruning with LoRA [11] fine-tuning, adhere to this hypothesis. However, for large models, global fine-tuning is resource-intensive, demanding significant GPU memory. Although LLM-Pruner reduces memory requirements using LoRA, it still requires storage for intermediate results and gradients, while methods like LLM Surgeon [30], despite avoiding global fine-tuning, also demand substantial memory and time due to the use of second-order derivatives. In contrast, one-shot pruning methods have gained traction by minimizing memory usage during fine-tuning. SparseGPT [7] avoids global fine-tuning by compensating for precision within each basic layer, thus accelerating the pruning process. Wanda [27] further simplifies this by eliminating compensation altogether, using a unique significance score to achieve performance comparable to SparseGPT without the need for fine-tuning. PruneZero[4] constructs an operator search space by combining input, output, and gradient operations to optimize model metrics. While, it requires multiple full inferences across models, which is computationally expensive. In memory-constrained settings, chunked model loading further exacerbates the inference time." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "However, both SparseGPT and Wanda have notable limitations. (i) While these methods perform well at lower" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9487" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": "sparsity levels (10%-50%) when pruning large language models, their effectiveness significantly declines at mid-to-high sparsity levels (50%-70%). (ii) SparseGPT and Wanda show poor performance on vision models. Wanda, in particular, relies on priors that are specific to language models, making it even less effective than SparseGPT for vision tasks." + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 163, + 175, + 256 + ], + "blocks": [ + { + "bbox": [ + 58, + 163, + 175, + 256 + ], + "lines": [ + { + "bbox": [ + 58, + 163, + 175, + 256 + ], + "spans": [ + { + "bbox": [ + 58, + 163, + 175, + 256 + ], + "type": "image", + "image_path": "e39f481947dbafb8d0c3ad62fe3b5e72ca9d2735a2e7a91c33e4dbeea7f3abbc.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 89, + 258, + 138, + 268 + ], + "lines": [ + { + "bbox": [ + 89, + 258, + 138, + 268 + ], + "spans": [ + { + "bbox": [ + 89, + 258, + 138, + 268 + ], + "type": "text", + "content": "(a) SparseGPT" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 74, + 277, + 276, + 289 + ], + "lines": [ + { + "bbox": [ + 74, + 277, + 276, + 289 + ], + "spans": [ + { + "bbox": [ + 74, + 277, + 276, + 289 + ], + "type": "text", + "content": "Figure 1. Error propagation with respect to layer index." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 181, + 164, + 296, + 256 + ], + "blocks": [ + { + "bbox": [ + 181, + 164, + 296, + 256 + ], + "lines": [ + { + "bbox": [ + 181, + 164, + 296, + 256 + ], + "spans": [ + { + "bbox": [ + 181, + 164, + 296, + 256 + ], + "type": "image", + "image_path": "7aa0a1012d41d66ec91272005f6e030f48648c9e1d9a1d6f47a547a93f0b0987.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 258, + 254, + 267 + ], + "lines": [ + { + "bbox": [ + 220, + 258, + 254, + 267 + ], + "spans": [ + { + "bbox": [ + 220, + 258, + 254, + 267 + ], + "type": "text", + "content": "(b) Wanda" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 304, + 295, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 304, + 295, + 496 + ], + "spans": [ + { + "bbox": [ + 54, + 304, + 295, + 496 + ], + "type": "text", + "content": "To illustrate why we focus on mid-to-high sparsity levels, we present the error accumulation during the pruning process of OPT-1.3B using SparseGPT and Wanda in Fig.1a and Fig.1b, with errors measured by mean square error (MSE). As shown, both SparseGPT and Wanda prune blocks sequentially without inter-block interaction, leading to faster error accumulation in deeper layers, which aligns with intuitive expectations. Furthermore, the error accumulation trends of semistructured sparsity 2:4 and 4:8 differ from those of unstructured sparsity " + }, + { + "bbox": [ + 54, + 304, + 295, + 496 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 54, + 304, + 295, + 496 + ], + "type": "text", + "content": ", instead aligning more closely with higher sparsity levels. Fig.1a, indicates that incorporating quantization strategies further shifts the error trends toward those of higher sparsity levels. Therefore, mitigating inter-block error propagation at sparsity levels of " + }, + { + "bbox": [ + 54, + 304, + 295, + 496 + ], + "type": "inline_equation", + "content": "50\\% -70\\%" + }, + { + "bbox": [ + 54, + 304, + 295, + 496 + ], + "type": "text", + "content": " can effectively enhance the performance of unstructured pruning and hybrid compression strategies." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 496, + 295, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 496, + 295, + 674 + ], + "spans": [ + { + "bbox": [ + 55, + 496, + 295, + 674 + ], + "type": "text", + "content": "In this paper, we propose ICP, a large model pruning method that is applicable to both vision and language models. Improve performance at both low- and high-sparsity levels in pruning large language and vision models, without increasing GPU memory usage and pruning time. Specifically, we design a sliding window-based inter-block error compensation method based on the principle of error propagation. As shown in Fig. 2, after pruning each block, the subsequent block compensates for the pruning-induced errors of the previous block before pruning itself. This approach ensures that errors introduced during pruning are compensated for in the next block, preventing error propagation to deeper layers and avoiding the need for full-model fine-tuning during the pruning process. We highlight our contributions as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "1. We introduce the concept of inter-block error compensation, demonstrating that this approach can enhance the performance of one-shot pruning methods with limited" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 340, + 78, + 529, + 205 + ], + "blocks": [ + { + "bbox": [ + 340, + 78, + 529, + 205 + ], + "lines": [ + { + "bbox": [ + 340, + 78, + 529, + 205 + ], + "spans": [ + { + "bbox": [ + 340, + 78, + 529, + 205 + ], + "type": "image", + "image_path": "d17cf899e889fca9ac16c06445f3dd6dd8b5b904d568d46f0bcdcf7847d90b0d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 213, + 555, + 280 + ], + "lines": [ + { + "bbox": [ + 313, + 213, + 555, + 280 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 555, + 280 + ], + "type": "text", + "content": "Figure 2. Comparison between our proposed method ICP and conventional pruning methods. (a) illustrates a standard pruning approach, where all blocks are pruned simultaneously. (b) demonstrates our method. After pruning one block, the subsequent block adjusts to compensate for the error introduced by the pruning of the previous block." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 326, + 312, + 429, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 312, + 429, + 323 + ], + "spans": [ + { + "bbox": [ + 326, + 312, + 429, + 323 + ], + "type": "text", + "content": "computational resources." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 323, + 555, + 443 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 314, + 323, + 555, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 323, + 555, + 395 + ], + "spans": [ + { + "bbox": [ + 314, + 323, + 555, + 395 + ], + "type": "text", + "content": "2. Building on this concept, we further design a sliding window-based inter-block error compensation strategy. After pruning each layer, we use the next unpruned layer to compensate for the errors induced by pruning the current layer. It improves post-pruning model performance in both vision and language models." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 396, + 554, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 396, + 554, + 443 + ], + "spans": [ + { + "bbox": [ + 314, + 396, + 554, + 443 + ], + "type": "text", + "content": "3. Extensive experiments on large language models and vision models show that our method significantly enhances post-pruning model performance without increasing the GPU memory consumption." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 469, + 400, + 481 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 469, + 400, + 481 + ], + "spans": [ + { + "bbox": [ + 314, + 469, + 400, + 481 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 495, + 459, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 495, + 459, + 506 + ], + "spans": [ + { + "bbox": [ + 314, + 495, + 459, + 506 + ], + "type": "text", + "content": "2.1. Transformer Architecture" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 516, + 555, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 516, + 555, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 516, + 555, + 624 + ], + "type": "text", + "content": "Since its introduction [31], Transformer has become the cornerstone of language models due to its superior performance in language tasks. Following the development of GPT [24], researchers have built increasingly large models using Transformer as primary building blocks. The Transformer has also proven effective in vision tasks. For example, SAM [13], which is based on the Vision Transformer, particularly excels in segmentation and serves as a foundational model for various downstream tasks [20, 23, 34]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 630, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 713 + ], + "type": "text", + "content": "Both language and vision models utilize a backbone composed of multiple Transformer blocks. The majority of parameters in these models are concentrated within their respective backbones. These Transformer blocks consist of multi-head self-attention and feed-forward neural networks with two linear layers. Multi-head self-attention is generally considered more critical than the feed-forward network." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9488" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 180, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 180, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 180, + 85 + ], + "type": "text", + "content": "2.2. Large Model Pruning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 297, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 297, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 297, + 449 + ], + "type": "text", + "content": "With the emergence of large models, the computational and memory costs of global fine-tuning made iterative pruning methods prohibitively expensive. The speed and lower computational resource requirements of one-shot pruning methods have led to the development of several such techniques for large models, notably SparseGPT and Wanda. SparseGPT [7] is a one-shot rapid pruning method belonging to Optimal Brain Surgeon [10] family. It independently analyzes each basic layer of the model, formulates the pruning problem as a constrained optimization problem, and solves the pruned weight values using the Lagrange multiplier method. SparseGPT builds on the OBS foundation by leveraging the property of Cholesky decomposition to efficiently obtain different orders of Hessian matrices, significantly accelerating the pruning process and reducing memory usage. Essentially, SparseGPT consists of two parts: pruning and compensation, which are integrated during the Lagrange multiplier method's solution process. However, SparseGPT compensates for pruning errors by adjusting unpruned weights within each basic layer, without considering inter-block weight interactions. Wanda [27], introduces a strong prior for language models, assuming that weight importance is related to both the absolute value of the weight and the corresponding output. Wanda designs an importance metric that achieves good results without compensating for pruning losses. However, subsequent studies [33] suggest that while Wanda performs well in many cases, it can occasionally fall short compared to SparseGPT, and its performance in vision models is particularly lacking, where its strong priors may lead to suboptimal results." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 461, + 112, + 474 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 461, + 112, + 474 + ], + "spans": [ + { + "bbox": [ + 55, + 461, + 112, + 474 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 483, + 296, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 296, + 615 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 296, + 615 + ], + "type": "text", + "content": "Our objective is to develop an efficient pruning method for large models, including both vision and language models which should be executable within a reasonable time frame, even under limited computational and memory resources. Additionally, it should sustain robust performance at mid-to-high sparsity levels. In this section, we introduce ICP, a pruning strategy that effectively mitigates performance degradation at mid-to-high sparsity levels and facilitates rapid pruning in memory-constrained environments. The following sections detail our method in two aspects: Blockwise Compensate Pruning and Sparsity Rearrangement." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 624, + 231, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 231, + 636 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 231, + 636 + ], + "type": "text", + "content": "3.1. Block-wise Compensate Pruning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "content": "We propose Block-wise Compensate Pruning, which enhances inter-block error compensation by building upon SparseGPT. As depicted in Fig. 3(a), Block-wise Compensate Pruning leverages the modular structure of large models, processing each block sequentially using a sliding window approach. In simple terms, the parameters of the next" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 555, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 120 + ], + "type": "text", + "content": "block are updated to compensate for the pruning-induced errors in the preceding block. Initially, the sliding window contains two blocks. Since the operations are consistent after each slide, we explain using the first two blocks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "spans": [ + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "text", + "content": "In Fig. 3(a), the sliding window moves from left to right with a window length encompassing two blocks. Initially, the window contains " + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "inline_equation", + "content": "B^{1}" + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "inline_equation", + "content": "B^{2}" + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "text", + "content": ". Pruning is first applied to " + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "inline_equation", + "content": "B^{1}" + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "text", + "content": ", followed by compensating for the pruning-induced error in " + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "inline_equation", + "content": "B^{1}" + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "inline_equation", + "content": "B^{2}" + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "text", + "content": ". The detailed calculation procedure for this compensation is outlined in Algorithm 1, as follows: The input data from the calibration set passes through the position encoding module, producing an intermediate result, " + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "inline_equation", + "content": "z^{0}" + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "text", + "content": ", which is then duplicated to create " + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "inline_equation", + "content": "z^{0'}" + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "text", + "content": ". At this stage, the feature flow splits into two streams: the \"standard stream\" " + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "inline_equation", + "content": "z^{0'}" + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "text", + "content": " and the \"error stream\" " + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "inline_equation", + "content": "z^{0}" + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "text", + "content": ". Before pruning " + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "inline_equation", + "content": "B^{1}" + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "inline_equation", + "content": "z^{0'}" + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "text", + "content": " passes through " + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "inline_equation", + "content": "B^{1}" + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "text", + "content": ", producing an unbiased output, " + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "inline_equation", + "content": "z^{1'}" + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "text", + "content": ". SparseGPT is then applied to prune " + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "inline_equation", + "content": "B^{1}" + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "inline_equation", + "content": "z^{0}" + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "text", + "content": " passes through the pruned " + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "inline_equation", + "content": "B^{1}" + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "text", + "content": ", resulting in a biased output, " + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "inline_equation", + "content": "z^{2}" + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "text", + "content": ". Subsequently, " + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "inline_equation", + "content": "z^{1'}" + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "text", + "content": " passes through the unpruned " + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "inline_equation", + "content": "B^{1}" + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "text", + "content": ", yielding an unbiased output, " + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "inline_equation", + "content": "z^{2'}" + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "inline_equation", + "content": "z^{1}" + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "text", + "content": " serves as both the output of the pruned " + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "inline_equation", + "content": "B^{1}" + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "text", + "content": " and the input to " + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "inline_equation", + "content": "B^{2}" + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "text", + "content": ". Since " + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "inline_equation", + "content": "z^{1}" + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "text", + "content": " is biased, the output from " + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "inline_equation", + "content": "B^{1}" + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "text", + "content": " will also be biased, whereas the desired output is " + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "inline_equation", + "content": "z^{2'}" + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "text", + "content": ". Therefore, " + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "inline_equation", + "content": "z^{2'}" + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "text", + "content": " serves as the target for " + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "inline_equation", + "content": "z^{1}" + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "text", + "content": ", forming a labeled feature dataset (" + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "inline_equation", + "content": "z^{1}, z^{2'}" + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "text", + "content": "). To align output of " + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "inline_equation", + "content": "B^{1}" + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "inline_equation", + "content": "z^{2'}" + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "text", + "content": ", we adjust " + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "inline_equation", + "content": "B^{1}" + }, + { + "bbox": [ + 313, + 123, + 556, + 447 + ], + "type": "text", + "content": "'s parameters by training it on this feature dataset, reducing the output bias. Finally, the sliding window advances by one block, and the process is repeated: prune the first block and adjust the parameters of the second block using newly formed feature dataset. As above, Block-wise Compensate Pruning progresses sequentially along the backbone, processing one block at a time." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 451, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 451, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 451, + 556, + 715 + ], + "type": "text", + "content": "Note that this method operates with a sliding window containing only two blocks, using the second block to compensate for the pruning-induced errors in the first, rather than relying on additional blocks for error correction. In this setup, once the first block is pruned, the standard stream dataset can pass through the second block to generate labels for the error stream. These labels are then reused during the compensation phase of the next sliding window, eliminating the need to store an additional block for real-time label computation. Consequently, at any given time, at most the weights of a single block reside in memory, optimizing both storage and computation. By contrast, if compensation required two subsequent blocks, the process would either necessitate storing two states of the standard stream or duplicating the next two blocks as a teacher model for distillation-like compensation, both of which would substantially increase memory overhead. It only requires a single block to be loaded into memory at a given moment. As each block is loaded into memory and pruned, the previous block is unloaded before the next block is loaded. This approach significantly reduces memory usage during the pruning and compensation process." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "9489" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 85, + 71, + 529, + 176 + ], + "blocks": [ + { + "bbox": [ + 85, + 71, + 529, + 176 + ], + "lines": [ + { + "bbox": [ + 85, + 71, + 529, + 176 + ], + "spans": [ + { + "bbox": [ + 85, + 71, + 529, + 176 + ], + "type": "image", + "image_path": "dd8117578032474846e58faa2ac98734446693915b595917a331516198257a04.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 85, + 176, + 525, + 277 + ], + "blocks": [ + { + "bbox": [ + 85, + 176, + 525, + 277 + ], + "lines": [ + { + "bbox": [ + 85, + 176, + 525, + 277 + ], + "spans": [ + { + "bbox": [ + 85, + 176, + 525, + 277 + ], + "type": "image", + "image_path": "6ace0e570d1c8586525828bd982e3a6efa6a21802c299a5664646228f474203e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 285, + 555, + 309 + ], + "lines": [ + { + "bbox": [ + 55, + 285, + 555, + 309 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 555, + 309 + ], + "type": "text", + "content": "Figure 3. Overview of the proposed method. (a) Block-wise Compensate Pruning for sequential block pruning and compensation; (b) basic structure of a transformer block (showing only weight-intensive and essential layers);" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 341, + 244, + 353 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 341, + 244, + 353 + ], + "spans": [ + { + "bbox": [ + 56, + 341, + 244, + 353 + ], + "type": "text", + "content": "Algorithm 1 Block-wise Compensate Pruning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 355, + 295, + 515 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 56, + 355, + 295, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 355, + 295, + 388 + ], + "spans": [ + { + "bbox": [ + 56, + 355, + 295, + 388 + ], + "type": "text", + "content": "Input: Backbone of large model " + }, + { + "bbox": [ + 56, + 355, + 295, + 388 + ], + "type": "inline_equation", + "content": "\\{B^i\\}_{i = 1}^n" + }, + { + "bbox": [ + 56, + 355, + 295, + 388 + ], + "type": "text", + "content": " , i-th block " + }, + { + "bbox": [ + 56, + 355, + 295, + 388 + ], + "type": "inline_equation", + "content": "B^{i}" + }, + { + "bbox": [ + 56, + 355, + 295, + 388 + ], + "type": "text", + "content": " , Pruning function Prune (module, hyperparameters), Calibration dataset " + }, + { + "bbox": [ + 56, + 355, + 295, + 388 + ], + "type": "inline_equation", + "content": "x_{j}" + }, + { + "bbox": [ + 56, + 355, + 295, + 388 + ], + "type": "text", + "content": " , Position embedding block " + }, + { + "bbox": [ + 56, + 355, + 295, + 388 + ], + "type": "inline_equation", + "content": "PE" + }, + { + "bbox": [ + 56, + 355, + 295, + 388 + ], + "type": "text", + "content": " Update function Update (module,(sample, label)), Sparsity s, Sparsity Rearrangement parameter " + }, + { + "bbox": [ + 56, + 355, + 295, + 388 + ], + "type": "inline_equation", + "content": "\\alpha ,\\beta" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 388, + 163, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 388, + 163, + 397 + ], + "spans": [ + { + "bbox": [ + 58, + 388, + 163, + 397 + ], + "type": "text", + "content": "Output: Pruned backbone " + }, + { + "bbox": [ + 58, + 388, + 163, + 397 + ], + "type": "inline_equation", + "content": "\\{B^i\\}_{i = 1}^n" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 397, + 124, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 397, + 124, + 407 + ], + "spans": [ + { + "bbox": [ + 59, + 397, + 124, + 407 + ], + "type": "text", + "content": "1: " + }, + { + "bbox": [ + 59, + 397, + 124, + 407 + ], + "type": "inline_equation", + "content": "z^0 = PE(\\{x_j\\})" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 59, + 407, + 176, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 407, + 176, + 418 + ], + "spans": [ + { + "bbox": [ + 59, + 407, + 176, + 418 + ], + "type": "text", + "content": "2: " + }, + { + "bbox": [ + 59, + 407, + 176, + 418 + ], + "type": "inline_equation", + "content": "z_{i}^{0^{\\prime}} = z_{i}^{0} / /" + }, + { + "bbox": [ + 59, + 407, + 176, + 418 + ], + "type": "text", + "content": " Copy embeddings" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 59, + 418, + 183, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 418, + 183, + 433 + ], + "spans": [ + { + "bbox": [ + 59, + 418, + 183, + 433 + ], + "type": "text", + "content": "3: " + }, + { + "bbox": [ + 59, + 418, + 183, + 433 + ], + "type": "inline_equation", + "content": "z_{j}^{1'} = B^{0}\\left(z_{j}^{0'}\\right) // \\text{Preparation}" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 59, + 433, + 134, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 433, + 134, + 440 + ], + "spans": [ + { + "bbox": [ + 59, + 433, + 134, + 440 + ], + "type": "text", + "content": "4: for " + }, + { + "bbox": [ + 59, + 433, + 134, + 440 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 59, + 433, + 134, + 440 + ], + "type": "text", + "content": " in 1 to " + }, + { + "bbox": [ + 59, + 433, + 134, + 440 + ], + "type": "inline_equation", + "content": "n - 1" + }, + { + "bbox": [ + 59, + 433, + 134, + 440 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 59, + 440, + 170, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 440, + 170, + 449 + ], + "spans": [ + { + "bbox": [ + 59, + 440, + 170, + 449 + ], + "type": "text", + "content": "5: " + }, + { + "bbox": [ + 59, + 440, + 170, + 449 + ], + "type": "inline_equation", + "content": "B^{i} = \\text{Prune}(B^{i}, (s, \\alpha, \\beta))" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 59, + 449, + 228, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 449, + 228, + 461 + ], + "spans": [ + { + "bbox": [ + 59, + 449, + 228, + 461 + ], + "type": "text", + "content": "6: " + }, + { + "bbox": [ + 59, + 449, + 228, + 461 + ], + "type": "inline_equation", + "content": "z_{j}^{i} = B^{i}\\left(z_{j}^{i - 1}\\right) / /" + }, + { + "bbox": [ + 59, + 449, + 228, + 461 + ], + "type": "text", + "content": " Forward Propagation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 59, + 461, + 147, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 461, + 147, + 475 + ], + "spans": [ + { + "bbox": [ + 59, + 461, + 147, + 475 + ], + "type": "text", + "content": "7: " + }, + { + "bbox": [ + 59, + 461, + 147, + 475 + ], + "type": "inline_equation", + "content": "z_{j}^{i + 1,i^{\\prime}} = B^{i + 1}\\left(z_{j}^{i^{\\prime}}\\right)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 59, + 475, + 195, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 475, + 195, + 490 + ], + "spans": [ + { + "bbox": [ + 59, + 475, + 195, + 490 + ], + "type": "text", + "content": "8: " + }, + { + "bbox": [ + 59, + 475, + 195, + 490 + ], + "type": "inline_equation", + "content": "B^{i+1} = \\text{Update}\\left(B^{i+1}, z_{j}^{i}, z_{j}^{i+1'}\\right)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 59, + 490, + 91, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 490, + 91, + 497 + ], + "spans": [ + { + "bbox": [ + 59, + 490, + 91, + 497 + ], + "type": "text", + "content": "9: end for" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 59, + 497, + 166, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 497, + 166, + 506 + ], + "spans": [ + { + "bbox": [ + 59, + 497, + 166, + 506 + ], + "type": "text", + "content": "10: " + }, + { + "bbox": [ + 59, + 497, + 166, + 506 + ], + "type": "inline_equation", + "content": "B^n = \\text{Prune}(B^n, (s, \\alpha, \\beta))" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 59, + 506, + 171, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 506, + 171, + 515 + ], + "spans": [ + { + "bbox": [ + 59, + 506, + 171, + 515 + ], + "type": "text", + "content": "11: return Pruned backbone " + }, + { + "bbox": [ + 59, + 506, + 171, + 515 + ], + "type": "inline_equation", + "content": "\\{B^i\\}_{i=1}^n" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 540, + 194, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 194, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 194, + 552 + ], + "type": "text", + "content": "3.2. Sparsity Rearrangement" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 54, + 557, + 295, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 557, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 54, + 557, + 295, + 688 + ], + "type": "text", + "content": "In many pruning methods, the sparsity of different parts of the model can be adaptively determined through specific rules. However, since our method is based on a sliding window approach for pruning, it is difficult to obtain information about the blocks that are outside the current window. If we attempt to acquire global information, we would either incur significant I/O overhead from loading model segments sequentially or face high memory usage from loading the entire model at once. Some methods even evaluate importance based on gradients, which can add further demands on gradient storage." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "To avoid these issues, fast pruning methods like SparseGPT or Wanda typically assign the same sparsity" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 329, + 555, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 329, + 555, + 390 + ], + "spans": [ + { + "bbox": [ + 313, + 329, + 555, + 390 + ], + "type": "text", + "content": "level uniformly across all parts of the model, equivalent to the target sparsity. However, this approach is clearly suboptimal. Certain prior knowledge can be leveraged to adjust the sparsity distribution, thereby avoiding these problems while improving performance." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 395, + 499, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 395, + 499, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 395, + 499, + 407 + ], + "type": "text", + "content": "3.2.1. Inter-Block Sparsity Rearrangement" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 409, + 555, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 409, + 555, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 409, + 555, + 506 + ], + "type": "text", + "content": "Block-wise Compensate Pruning fine-tunes " + }, + { + "bbox": [ + 313, + 409, + 555, + 506 + ], + "type": "inline_equation", + "content": "B^{i+1}" + }, + { + "bbox": [ + 313, + 409, + 555, + 506 + ], + "type": "text", + "content": " to compensate for the pruning-induced errors in " + }, + { + "bbox": [ + 313, + 409, + 555, + 506 + ], + "type": "inline_equation", + "content": "B^i" + }, + { + "bbox": [ + 313, + 409, + 555, + 506 + ], + "type": "text", + "content": ". However, when pruning the last block, " + }, + { + "bbox": [ + 313, + 409, + 555, + 506 + ], + "type": "inline_equation", + "content": "B^n" + }, + { + "bbox": [ + 313, + 409, + 555, + 506 + ], + "type": "text", + "content": ", there is no " + }, + { + "bbox": [ + 313, + 409, + 555, + 506 + ], + "type": "inline_equation", + "content": "B^{n+1}" + }, + { + "bbox": [ + 313, + 409, + 555, + 506 + ], + "type": "text", + "content": " available to compensate for the errors in " + }, + { + "bbox": [ + 313, + 409, + 555, + 506 + ], + "type": "inline_equation", + "content": "B^n" + }, + { + "bbox": [ + 313, + 409, + 555, + 506 + ], + "type": "text", + "content": ". Therefore, assigning the same number of pruned weights to each block is not optimal. Instead, the pruning distribution should be adjusted across different blocks. To maintain the global target sparsity, we rearrange the sparsity between blocks as follows:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 389, + 517, + 555, + 531 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 517, + 555, + 531 + ], + "spans": [ + { + "bbox": [ + 389, + 517, + 555, + 531 + ], + "type": "interline_equation", + "content": "P _ {B ^ {n}} = \\left| \\theta^ {n} \\right| \\times s \\times \\alpha , \\tag {1}", + "image_path": "d89dd97924e050e1148a537558a52db5437be48daecc5404c129bc3d657a5bf5.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 356, + 543, + 554, + 567 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 543, + 554, + 567 + ], + "spans": [ + { + "bbox": [ + 356, + 543, + 554, + 567 + ], + "type": "interline_equation", + "content": "P _ {B ^ {j}} = \\left| \\theta^ {j} \\right| \\times s + \\frac {\\left| \\theta^ {n} \\right| \\times s \\times (1 - \\alpha)}{n - 1} \\tag {2}", + "image_path": "1304042f4460adc1a258304fa816219d0c12d70ae6249b221dc540730ebcf9bb.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 572, + 555, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 572, + 555, + 633 + ], + "spans": [ + { + "bbox": [ + 313, + 572, + 555, + 633 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 572, + 555, + 633 + ], + "type": "inline_equation", + "content": "j = 1, \\ldots, n - 1" + }, + { + "bbox": [ + 313, + 572, + 555, + 633 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 572, + 555, + 633 + ], + "type": "inline_equation", + "content": "\\theta^j" + }, + { + "bbox": [ + 313, + 572, + 555, + 633 + ], + "type": "text", + "content": " denotes the parameters in " + }, + { + "bbox": [ + 313, + 572, + 555, + 633 + ], + "type": "inline_equation", + "content": "B^j" + }, + { + "bbox": [ + 313, + 572, + 555, + 633 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 572, + 555, + 633 + ], + "type": "inline_equation", + "content": "|\\theta|" + }, + { + "bbox": [ + 313, + 572, + 555, + 633 + ], + "type": "text", + "content": " denotes the number of weights in a single block, and " + }, + { + "bbox": [ + 313, + 572, + 555, + 633 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 572, + 555, + 633 + ], + "type": "text", + "content": " is a super-parameter that redistributes a portion of the pruned weights from the last block evenly across all preceding blocks." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 639, + 500, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 639, + 500, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 639, + 500, + 651 + ], + "type": "text", + "content": "3.2.2. Intra-Block Sparsity Rearrangement" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "type": "text", + "content": "As shown in Fig. 3(b), the majority of the parameters within a Transformer block are contributed by the matrices depicted, which are clearly hierarchical. Therefore, a block can essentially be viewed as a nested function: the QKV matrices represent the innermost function, while the second" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9490" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "type": "text", + "content": "fully connected layer (fc2) represents the outermost function. According to the conclusions drawn in the Theoretical Foundations section, if the same sparsity level is applied uniformly across all layers, the errors introduced by the QKV matrices, which are further from the output of " + }, + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "type": "inline_equation", + "content": "B^{i}" + }, + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "type": "text", + "content": ", will be more challenging to compensate for in " + }, + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "type": "inline_equation", + "content": "B^{i + 1}" + }, + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "type": "text", + "content": ". On the other hand, fc2 is closer to the output of " + }, + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "type": "inline_equation", + "content": "B^{i}" + }, + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "type": "text", + "content": ", making its errors easier to compensate for in " + }, + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "type": "inline_equation", + "content": "B^{i + 1}" + }, + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "type": "text", + "content": ". Additionally, inspired by previous works, earlier layers tend to converge first [1], and faster convergence often leads to better generalization performance [9]. Therefore, the sparsity distribution should be rearranged by reducing the sparsity of the QKV matrices and increasing the sparsity of fc2. To ensure that the global target sparsity remains unchanged, we rearrange the sparsity according to the following rule:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 84, + 258, + 295, + 274 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 258, + 295, + 274 + ], + "spans": [ + { + "bbox": [ + 84, + 258, + 295, + 274 + ], + "type": "interline_equation", + "content": "P _ {t} = s \\times | W _ {t} ^ {i} | \\times (1 - \\beta), t \\in \\{Q, K, V \\}. (3)", + "image_path": "f792003f589bbac6ba12c70093ba78e515f19a43d8281e1ad3e1ce7b879f88f7.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 96, + 290, + 295, + 306 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 290, + 295, + 306 + ], + "spans": [ + { + "bbox": [ + 96, + 290, + 295, + 306 + ], + "type": "interline_equation", + "content": "P _ {f c 2} = s \\times | W _ {f c 2} ^ {i} | + s \\times | W _ {f c 2} ^ {i} | \\times \\beta . \\tag {4}", + "image_path": "378433e44173e4ecb0312e41aaf20bbc4ca5207049fdb5ee02c6c8da767a94ad.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 308, + 296, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 308, + 296, + 344 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 296, + 344 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 308, + 296, + 344 + ], + "type": "inline_equation", + "content": "P_{t}" + }, + { + "bbox": [ + 55, + 308, + 296, + 344 + ], + "type": "text", + "content": " represents the number of pruned weights, " + }, + { + "bbox": [ + 55, + 308, + 296, + 344 + ], + "type": "inline_equation", + "content": "|W|" + }, + { + "bbox": [ + 55, + 308, + 296, + 344 + ], + "type": "text", + "content": " denotes the total number of weights in a single layer, and " + }, + { + "bbox": [ + 55, + 308, + 296, + 344 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 308, + 296, + 344 + ], + "type": "text", + "content": " is a super-parameter that needs to be adjusted." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 354, + 137, + 367 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 137, + 367 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 137, + 367 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 373, + 172, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 172, + 387 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 172, + 387 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 391, + 296, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 391, + 296, + 486 + ], + "spans": [ + { + "bbox": [ + 55, + 391, + 296, + 486 + ], + "type": "text", + "content": "To evaluate the effectiveness of our method, we focus on models with 7 billion parameters or fewer, as these are the most commonly used in research and resource-constrained scenarios. Pruning these models is also the most feasible on a single desktop-class GPU. Specifically, we validate our approach using the OPT-125M/1.3B/2.7B/6.7B and Llama 2-7B, as well as the SAM-B/L/H. All experiments were conducted on a single NVIDIA RTX 3090 GPU." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 487, + 296, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 296, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 296, + 545 + ], + "type": "text", + "content": "For the compensation process, we employ the Adam optimizer with a Mean Squared Error (MSE) loss function and a batch size of 1. All compensation operations during pruning share a unified learning rate, which is adjusted using the ExponentialLR scheduler with a decay factor of 0.87." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 545, + 296, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 617 + ], + "type": "text", + "content": "For calibration data, language model calibration dataset follows SparseGPT, using 128 segments of 2048 tokens, randomly sampled from C4 dataset [25]. Similarly, the vision model calibration dataset consists of 128 images, randomly sampled from the second compressed file of the SA-1B dataset [13] provided by Meta." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": "To comprehensively evaluate the performance of pruned models, we validate OPT using the Wikitext-v2 [21] and PTB datasets [19] from Hugging Face, with perplexity (PPL) as the evaluation metric. We also evaluate the pruned Llama 2-7B model using accuracy on seven ZeroShot tasks: BoolQ [2], RTE [32], HellaSwag [35], Wino-Grande [26], ARC Easy and Challenge [3], and Open-bookQA [22], following the experimental setup introduced" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "text", + "content": "by Sun et al. [27] in their work on Wanda. This allows for a more task-specific evaluation of the pruned model's generalization capabilities. For SAM, we validate the performance using the fourth compressed file of the SA-1B dataset as well as the COCO dataset [15], employing Intersection over Union (IoU) as the primary metric for visual models." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 165, + 441, + 178 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 165, + 441, + 178 + ], + "spans": [ + { + "bbox": [ + 313, + 165, + 441, + 178 + ], + "type": "text", + "content": "4.2. Unstructured Pruning" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 182, + 555, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 182, + 555, + 397 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 555, + 397 + ], + "type": "text", + "content": "Language Models. In Table 1 and Table 2, we report the perplexity performance of the OPT series models on Wikitext and PTB datasets after pruning to various sparsity levels using our method and other baseline approaches. Our method demonstrates superior performance over baselines even at lower sparsity levels, and as sparsity increases, the performance gap widens in favor of our approach. Additionally, as model size decreases, the advantage of our method over the baselines becomes even more pronounced. Table 3 also presents results in comparing the performances of our proposed method against those of baselines on the Llama 2-7B model at " + }, + { + "bbox": [ + 313, + 182, + 555, + 397 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 182, + 555, + 397 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 182, + 555, + 397 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 313, + 182, + 555, + 397 + ], + "type": "text", + "content": " sparsity levels. Our method consistently outperforms all baseline approaches at both sparsity levels, achieving the best results across all tasks. These findings highlight the effectiveness of the interblock error compensation pruning method introduced in this paper, particularly under high-sparsity conditions, across different model scales, and for various model architectures." + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 320, + 441, + 549, + 604 + ], + "blocks": [ + { + "bbox": [ + 313, + 408, + 554, + 440 + ], + "lines": [ + { + "bbox": [ + 313, + 408, + 554, + 440 + ], + "spans": [ + { + "bbox": [ + 313, + 408, + 554, + 440 + ], + "type": "text", + "content": "Table 1. Perplexity performances of pruned OPT models of different scales at various sparsity levels. The models are calibrated on a subset of the C4 and evaluated on full Wikitext-v2." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 320, + 441, + 549, + 604 + ], + "lines": [ + { + "bbox": [ + 320, + 441, + 549, + 604 + ], + "spans": [ + { + "bbox": [ + 320, + 441, + 549, + 604 + ], + "type": "table", + "html": "
MethodModel10%20%30%40%50%60%70%80%90%
Magnitude125M28.2529.6534.5154.60193.4920.0380648906614
SparseGPT27.8827.9828.8630.5637.0259.55220.923784666
Wanda27.4827.7628.1130.6738.9275.17328.219184609
ICP (ours)27.3727.8728.6629.6633.9045.7865.20270.41612
Magnitude1.3B14.7215.6224.74388.01713939294431634428871
SparseGPT14.6714.7515.1616.4917.5022.0851.75752.46797
Wanda14.6314.6915.0115.8918.4126.5599.53225816868
ICP (ours)14.5914.6014.6715.0416.4921.1736.78117.11099
Magnitude2.7B12.5913.1315.5830.32265.236047251961416668
SparseGPT12.3212.3912.6612.6613.4616.0426.92138.45818
Wanda12.2312.2612.3612.8614.2119.83387.9590516527
ICP (ours)12.2212.2512.3412.6213.3215.7721.7954.67764.2
Magnitude6.7B10.9211.2712.5331.89968.71263916975255915297
SparseGPT10.8310.7610.7510.9611.5913.4820.4896.2511938
Wanda10.7310.6310.6410.9611.9815.19159.2395416076
ICP (ours)10.5610.5810.5910.8511.2713.0317.8244.691469
", + "image_path": "ab1c41eba5571e697d2f7a49dd02485d9dbe945252a052f550921f5e0adb5ae4.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "content": "Vision Models. In Tables 4 and Table 5, we report IoU performance of SAM at varying sparsity levels on SA-1B and COCO datasets with single-point interaction. Following experimental setup, SAMs are pruned with only 128 images random sampled from SA-1B as calibration data. The experiments first verify that SparseGPT indeed achieves strong IoU performance on visual models under low sparsity conditions, outperforming both Magnitude and Wanda." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "text", + "content": "9491" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 63, + 103, + 288, + 266 + ], + "blocks": [ + { + "bbox": [ + 56, + 71, + 294, + 102 + ], + "lines": [ + { + "bbox": [ + 56, + 71, + 294, + 102 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 294, + 102 + ], + "type": "text", + "content": "Table 2. Perplexity performancese of pruned OPT models of different scales at various sparsity levels. The models are calibrated on a subset of the C4 and evaluated on PTB." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 103, + 288, + 266 + ], + "lines": [ + { + "bbox": [ + 63, + 103, + 288, + 266 + ], + "spans": [ + { + "bbox": [ + 63, + 103, + 288, + 266 + ], + "type": "table", + "html": "
MethodModel10%20%30%40%50%60%70%80%90%
Magnitude125M39.9743.2549.9782.69276.21146343041245382
SparseGPT39.4340.3642.7145.2455.7888.44259.629195374
Wanda39.0539.8341.4344.4157.59109.7399.722144474
ICP (ours)39.1239.1340.9643.4749.4164.1088.85336.33075
Magnitude1.3B20.6621.5139.89823.53171949984671728333126
SparseGPT20.3420.6421.5023.1725.4733.4876.91531.87515
Wanda20.3820.7021.4123.2528.0043.95135.3180912025
ICP (ours)20.1320.4121.1122.1924.6832.1954.33148.83094
Magnitude2.7B18.3819.0622.9646.36262.632077276957512178
SparseGPT18.1518.5018.9818.9820.4424.6742.75146.06107
Wanda17.9718.0418.3219.2521.8633.44450.8546221039
ICP (ours)17.9718.0218.2918.8620.0522.9235.1771.912357
Magnitude6.7B16.1816.7018.5248.85613.493799973157424260
SparseGPT15.7715.8115.9216.2817.4320.2331.86105.08320
Wanda15.7615.7715.9416.4317.9223.61211.3302414998
ICP (ours)15.7615.7415.9416.2217.3020.0728.0464.182549
", + "image_path": "6ecfa2ae1e2ace5d16c443e11f68593c8de8abbc80400a8fe2cd06158bf558cd.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 63, + 315, + 288, + 416 + ], + "blocks": [ + { + "bbox": [ + 56, + 281, + 294, + 313 + ], + "lines": [ + { + "bbox": [ + 56, + 281, + 294, + 313 + ], + "spans": [ + { + "bbox": [ + 56, + 281, + 294, + 313 + ], + "type": "text", + "content": "Table 3. Zero-shot performance (Accuracy, %) of Llama 2-7B pruned by different method on various tasks at sparsity of " + }, + { + "bbox": [ + 56, + 281, + 294, + 313 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 56, + 281, + 294, + 313 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 281, + 294, + 313 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 56, + 281, + 294, + 313 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 315, + 288, + 416 + ], + "lines": [ + { + "bbox": [ + 63, + 315, + 288, + 416 + ], + "spans": [ + { + "bbox": [ + 63, + 315, + 288, + 416 + ], + "type": "table", + "html": "
MethodStrategyBQRTEHSWGARC-eARC-cOBQAMean
Dense077.7463.1857.1068.9876.2643.4331.4059.73
Magnitude50%62.9457.0449.1263.3864.0634.6426.8051.14
SparseGPT76.3355.9652.8968.9872.0138.2328.4056.11
Wanda76.4253.4352.4568.6772.2239.3331.0056.22
ICP (ours)76.5057.0453.0369.1172.3940.4431.0056.96
Magnitude70%37.9553.0725.9349.2527.8222.8717.0033.41
SparseGPT64.6553.7933.4558.6443.3121.6717.0041.79
Wanda48.4752.7127.9749.6430.8118.6012.2034.34
ICP (ours)67.7155.2341.4560.2255.5627.0521.4046.94
", + "image_path": "e7ae43f33179b383dc80c8196b60e38393a7495d892ce2067a11fe5010963a94.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 441, + 294, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 441, + 294, + 548 + ], + "spans": [ + { + "bbox": [ + 56, + 441, + 294, + 548 + ], + "type": "text", + "content": "As Wanda is based on strong priors from language models, its underperformance compared to SparseGPT on visual models is within our expectations. Lastly, the method proposed in this paper outperforms all baselines and exhibits a similar trend observed in language model experiments: the higher the sparsity, the greater the advantage of our method over the baseline methods. This demonstrates that the proposed method is not only effective for language models but also remains effective for visual models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 562, + 274, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 562, + 274, + 575 + ], + "spans": [ + { + "bbox": [ + 58, + 562, + 274, + 575 + ], + "type": "text", + "content": "4.3. Semi-Structure Pruning and Quantization" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 582, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 294, + 713 + ], + "type": "text", + "content": "On certain specialized hardware, such as specific NVIDIA GPUs or other custom-designed computing architectures, the combination of pruning and quantization can further accelerate inference while maintaining an acceptable performance degradation. To explore this, we conduct tests using a joint compression strategy that combines pruning and quantization. Specifically, we focus on the combination of 2:4 and 4:8 semi-structured sparsity with 3-bit and 4-bit quantization. Since Wanda does not support quantization, we exclude it from experiments involving quantization and instead compared our approach against SparseGPT." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 322, + 115, + 547, + 240 + ], + "blocks": [ + { + "bbox": [ + 316, + 71, + 553, + 114 + ], + "lines": [ + { + "bbox": [ + 316, + 71, + 553, + 114 + ], + "spans": [ + { + "bbox": [ + 316, + 71, + 553, + 114 + ], + "type": "text", + "content": "Table 4. Instance segmentation performance of SAM models (IoU, " + }, + { + "bbox": [ + 316, + 71, + 553, + 114 + ], + "type": "inline_equation", + "content": "\\%" + }, + { + "bbox": [ + 316, + 71, + 553, + 114 + ], + "type": "text", + "content": " ) at various sparsity using single-point interaction on SA-1B. The models are calibrated on a subset of 128 images sampled from SA-1B." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 322, + 115, + 547, + 240 + ], + "lines": [ + { + "bbox": [ + 322, + 115, + 547, + 240 + ], + "spans": [ + { + "bbox": [ + 322, + 115, + 547, + 240 + ], + "type": "table", + "html": "
MethodModel10%20%30%40%50%60%70%80%90%
MagnitudeSAM-H74.8074.5773.8772.2669.4561.5128.222.651.70
SparseGPT74.8174.7774.6074.2073.3071.1565.6950.9718.15
Wanda74.8274.7674.3573.6071.6567.6553.0021.841.20
ICP (ours)74.8274.7874.6674.6074.4273.8272.8569.7357.60
MagnitudeSAM-L74.8174.5973.9372.2869.1862.5645.529.924.52
SparseGPT74.8474.7774.6174.1373.2070.7064.4545.487.37
Wanda74.8674.7874.6274.0572.5168.2855.1322.811.31
ICP (ours)74.8574.7974.6974.5574.2773.6472.2668.1753.68
MagnitudeSAM-B72.6972.5671.9970.3667.3860.8746.0125.531.51
SparseGPT72.7272.6772.5072.0871.1168.6863.0247.4918.81
Wanda72.7172.6472.3371.4569.4464.3550.2928.162.18
ICP (ours)72.7272.6972.5672.2871.8571.0669.3364.7851.82
", + "image_path": "7114bf472e365b28c1b5a1a7722133566c8c0d865df8b664b13df96ddad25fd8.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 322, + 275, + 547, + 399 + ], + "blocks": [ + { + "bbox": [ + 316, + 252, + 553, + 274 + ], + "lines": [ + { + "bbox": [ + 316, + 252, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 252, + 553, + 274 + ], + "type": "text", + "content": "Table 5. Instance segmentation performance of SAM models (IoU, %) at various sparsity using single-point interaction on COCO." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 322, + 275, + 547, + 399 + ], + "lines": [ + { + "bbox": [ + 322, + 275, + 547, + 399 + ], + "spans": [ + { + "bbox": [ + 322, + 275, + 547, + 399 + ], + "type": "table", + "html": "
MethodModel10%20%30%40%50%60%70%80%90%
MagnitudeSAM-H71.1271.1170.8670.6169.3763.6032.687.887.21
SparseGPT71.1171.1870.9870.9370.4269.5266.7056.8227.57
Wanda71.1771.0170.9870.7369.6366.0553.4522.835.51
ICP (ours)71.1471.1871.1270.9870.8170.2469.4366.6958.41
MagnitudeSAM-L70.5670.5170.2769.5368.0862.6649.1220.6618.12
SparseGPT70.5570.5370.4370.0469.6168.2164.4248.3511.30
Wanda70.5670.5870.4569.9168.6565.5654.6719.896.38
ICP (ours)70.5570.5870.5870.3769.9169.5168.1465.4256.92
MagnitudeSAM-B66.8467.0667.0666.6765.2962.2853.6140.867.48
SparseGPT67.0967.0266.9866.7666.1765.0062.0352.4628.51
Wanda66.9466.9767.0466.4765.3862.3153.5839.327.89
ICP (ours)67.1067.0667.0466.8366.4566.0764.5162.4054.50
", + "image_path": "00239242c6a99d2bae0e16b3f4df71be9c71c49cd0639a01cbca81d3eac89d9e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 328, + 434, + 542, + 603 + ], + "blocks": [ + { + "bbox": [ + 316, + 411, + 553, + 433 + ], + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 433 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 433 + ], + "type": "text", + "content": "Table 6. Perplexity performance of pruned OPT models at various compression strategies." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 328, + 434, + 542, + 603 + ], + "lines": [ + { + "bbox": [ + 328, + 434, + 542, + 603 + ], + "spans": [ + { + "bbox": [ + 328, + 434, + 542, + 603 + ], + "type": "table", + "html": "
MethodStrategyOPT-125MOPT-1.3BOPT-2.7BOPT-6.7B
WTPTBWTPTBWTPTBWTPTB
SparseGPT2:459.3593.0423.8738.0917.1126.9814.1521.54
Wanda2:480.01111.9728.2043.3721.1834.5515.9025.09
ICP (ours)2:446.9067.5622.0333.9516.3925.3114.0421.19
SparseGPT4:843.9071.9920.2131.3615.0023.0512.4918.86
Wanda4:853.1879.0222.1734.5616.7926.1313.5520.19
ICP (ours)4:839.9159.3419.1530.4114.8822.8812.3818.53
SparseGPT2:4+int3122.33196.4440.5864.5324.7041.0620.0731.91
ICP (ours)2:4+int374.07114.9327.8846.1419.5730.2818.3826.82
SparseGPT2:4+int470.88107.3925.9641.6918.5630.6815.2423.01
ICP (ours)2:4+int451.2475.3923.4934.4217.0527.0415.1022.54
SparseGPT4:8+int384.02120.1632.4149.2720.2432.9917.0826.36
ICP (ours)4:8+int359.6786.0124.7340.1917.8926.8016.1624.30
SparseGPT4:8+int450.5880.9723.3234.3815.7424.7613.4419.72
ICP (ours)4:8+int442.0360.6020.8631.5815.3323.8013.2419.49
", + "image_path": "c9af1e2b1d3edb33b9cdae473daf9e0cbacee1d7d039fba8c0ead6c119ab2cbc.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 624, + 495, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 495, + 636 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 495, + 636 + ], + "type": "text", + "content": "4.4. Sparsity Rearrangement Ablation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 641, + 553, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 641, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 641, + 553, + 689 + ], + "type": "text", + "content": "To validate the Sparsity Rearrangement method based on priors, as proposed in the Method section, we conducted experiments on both Inter-Block Sparsity Rearrangement and Intra-Block Sparsity Rearrangement." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 689, + 553, + 713 + ], + "type": "text", + "content": "Inter-Block Sparsity Rearrangement. In Fig. 4a and Fig. 4b, we present the performance of OPT-125M and" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "9492" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 63, + 93, + 288, + 250 + ], + "blocks": [ + { + "bbox": [ + 56, + 71, + 295, + 92 + ], + "lines": [ + { + "bbox": [ + 56, + 71, + 295, + 92 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 295, + 92 + ], + "type": "text", + "content": "Table 7. Zero-shot Performance (Accuracy, %) of Llama 2-7B Model under various compression strategies." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 93, + 288, + 250 + ], + "lines": [ + { + "bbox": [ + 63, + 93, + 288, + 250 + ], + "spans": [ + { + "bbox": [ + 63, + 93, + 288, + 250 + ], + "type": "table", + "html": "
MethodStrategyBQRTEHSWGARC-eARC-cOBQAMean
SparseGPT2:468.0758.8443.3865.5964.1831.9124.8050.97
Wanda2:468.1353.4341.4262.4363.1330.6323.8049.00
ICP (ours)2:469.7258.8446.7564.4866.0433.7026.2052.25
SparseGPT4:871.0754.8748.3567.8068.6934.4727.2053.21
Wanda4:873.0053.7947.0366.9367.4734.2227.0052.78
ICP (ours)4:873.1258.1250.1364.8868.9036.0128.0054.16
SparseGPT2:4+int364.6555.2338.5859.0455.4725.3419.4045.39
ICP (ours)2:4+int366.8254.1543.3660.7759.5528.6721.0047.76
SparseGPT2:4+int466.7354.1542.3465.1961.7428.9223.8048.98
ICP (ours)2:4+int470.4656.6846.2262.9064.8130.6322.8050.64
SparseGPT4:8+int368.5953.4342.1262.3560.1029.1021.6048.18
ICP (ours)4:8+int368.3854.8743.5462.8563.3832.3422.4049.68
SparseGPT4:8+int469.2754.1546.5866.0667.5534.3927.2052.17
ICP (ours)4:8+int472.2961.0148.8662.7566.8833.8728.0053.38
", + "image_path": "a3a46e55e2963bef15a2916ef7248ebb6e7a48b0bb6e21b9249399ab1936ae2c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 74, + 295, + 278, + 464 + ], + "blocks": [ + { + "bbox": [ + 56, + 261, + 295, + 293 + ], + "lines": [ + { + "bbox": [ + 56, + 261, + 295, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 295, + 293 + ], + "type": "text", + "content": "Table 8. Instance segmentation performance of SAM models (IoU, " + }, + { + "bbox": [ + 56, + 261, + 295, + 293 + ], + "type": "inline_equation", + "content": "\\%" + }, + { + "bbox": [ + 56, + 261, + 295, + 293 + ], + "type": "text", + "content": " ) at various compression strategies using single-point interaction on SA-1B and COCO." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 74, + 295, + 278, + 464 + ], + "lines": [ + { + "bbox": [ + 74, + 295, + 278, + 464 + ], + "spans": [ + { + "bbox": [ + 74, + 295, + 278, + 464 + ], + "type": "table", + "html": "
MethodStrategySAM-BSAM-LSAM-H
COCOSA-1BCOCOSA-1BCOCOSA-1B
SparseGPT2:462.8067.5965.0169.0767.0269.62
Wanda2:460.5962.0662.3765.8464.6866.36
ICP (ours)2:464.9571.3067.8373.5669.0973.64
SparseGPT4:863.9369.6966.5671.2767.9871.41
Wanda4:862.6266.0565.0269.7566.8069.55
ICP (ours)4:865.3771.7268.1773.8669.4474.03
SparseGPT2:4+int362.0964.7763.4565.6766.3967.85
ICP (ours)2:4+int364.4769.7167.1372.6068.7573.26
SparseGPT2:4+int462.6666.7664.7868.6966.8169.44
ICP (ours)2:4+int464.8970.6467.6473.0969.0273.72
SparseGPT4:8+int363.0267.2764.7468.0267.2569.69
ICP (ours)4:8+int364.8469.8667.6873.2869.1073.35
SparseGPT4:8+int463.7768.9866.0470.3167.8771.05
ICP (ours)4:8+int465.2470.9868.0273.8269.3473.78
", + "image_path": "37964a2f8e98a6a70d693ac2aa98bd2654c2b2888b58ef6e03f0a4406777c662.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 485, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 485, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 485, + 295, + 652 + ], + "type": "text", + "content": "SAM-B pruned by ICP with different " + }, + { + "bbox": [ + 55, + 485, + 295, + 652 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 485, + 295, + 652 + ], + "type": "text", + "content": ". For clarity in visual presentation, we conducted experiments with " + }, + { + "bbox": [ + 55, + 485, + 295, + 652 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 55, + 485, + 295, + 652 + ], + "type": "text", + "content": " sparsity on OPT-125M and " + }, + { + "bbox": [ + 55, + 485, + 295, + 652 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 55, + 485, + 295, + 652 + ], + "type": "text", + "content": " sparsity on SAM-B. We observed that in the language model, " + }, + { + "bbox": [ + 55, + 485, + 295, + 652 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 485, + 295, + 652 + ], + "type": "text", + "content": " had the greatest impact on the performance of OPT-125M on the PTB dataset, and adjusting " + }, + { + "bbox": [ + 55, + 485, + 295, + 652 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 485, + 295, + 652 + ], + "type": "text", + "content": " further improved the model's performance on non-calibration datasets. On vision models, the influence of " + }, + { + "bbox": [ + 55, + 485, + 295, + 652 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 485, + 295, + 652 + ], + "type": "text", + "content": " on performance was similar on both datasets, with adjustments to " + }, + { + "bbox": [ + 55, + 485, + 295, + 652 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 485, + 295, + 652 + ], + "type": "text", + "content": " enhancing the overall performance of the visual model. Overall, appropriate tuning of " + }, + { + "bbox": [ + 55, + 485, + 295, + 652 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 485, + 295, + 652 + ], + "type": "text", + "content": " can further improve the performance of pruned models. Based on experiments, we conclude that " + }, + { + "bbox": [ + 55, + 485, + 295, + 652 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 485, + 295, + 652 + ], + "type": "text", + "content": " typically yields optimal results within the range of [0.6, 0.75] for language models, and within [0.8, 0.9] for vision model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "content": "Intra-Block Sparsity Rearrangement. In Fig. 4c, we illustrate the impact of different " + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "content": " on the language model, using the " + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "content": " sparsity model of OPT-125M as an example. We observed that in the language model, adjusting " + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "content": " can further reduce the model's perplexity. Unlike " + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "content": " not" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": "only improves the model's performance on non-calibration datasets but also enhances its performance on calibration datasets. Overall, appropriately adjusting " + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": " can further improve the performance of the pruned model. The optimal value for " + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": " typically falls within the range of " + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "inline_equation", + "content": "[0.6 \\times (1 - s), 0.8 \\times (1 - s)]" + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": " denotes sparsity. In Fig. 4d, we exemplify the effects of varying " + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": " on the visual model, specifically using the " + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": " sparsity model of SAM-B. Our findings indicate that the influence of " + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": " on the visual model is minimal; both on calibration and non-calibration datasets, adjustments to " + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": " yield only marginal improvements in the model's performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 224, + 474, + 236 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 224, + 474, + 236 + ], + "spans": [ + { + "bbox": [ + 313, + 224, + 474, + 236 + ], + "type": "text", + "content": "4.5. Calibration Set Size Ablation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 242, + 555, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 242, + 555, + 374 + ], + "spans": [ + { + "bbox": [ + 313, + 242, + 555, + 374 + ], + "type": "text", + "content": "Although our experiments follow the settings of SparseGPT and Wanda, using 128 randomly selected samples for pruning both language and vision models, Fig. 4e and Fig. 4f show the impact of calibration set size on our method's performance. Fig. 1 presents results on the OPT-125M language model at " + }, + { + "bbox": [ + 313, + 242, + 555, + 374 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 313, + 242, + 555, + 374 + ], + "type": "text", + "content": " sparsity, while Fig. 2 shows results for the SAM-B vision model at " + }, + { + "bbox": [ + 313, + 242, + 555, + 374 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 313, + 242, + 555, + 374 + ], + "type": "text", + "content": " sparsity. In both cases, increasing the sample size beyond 128 offers minimal performance gains while significantly increasing GPU memory usage. Thus, like SparseGPT and Wanda, 128 samples remain the optimal choice." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 382, + 481, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 382, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 382, + 481, + 396 + ], + "type": "text", + "content": "4.6. Compensation Epoch Ablation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 400, + 556, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 400, + 556, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 400, + 556, + 533 + ], + "type": "text", + "content": "In Fig.4g and Fig.4h, we illustrate the impact of varying the number of epochs during the compensation on the performance of both vision and language models. It is evident that a single epoch is sufficient to effectively mitigate the error introduced by pruning in earlier layers. While additional epochs can further improve performance, the marginal gains diminish rapidly. Although we report results using 12 epochs in the experiments presented in 4.2 and 4.3 to demonstrate the best achievable performance, selecting a single epoch offers the optimal trade-off between model performance and computational overhead." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 540, + 473, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 473, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 473, + 553 + ], + "type": "text", + "content": "4.7. Memory and Time Efficiency" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": "Peak memory usage (PMU) and execution time are critical factors in determining whether a pruning method can be deployed on devices with limited memory capacity and lower cost. In Table 9, we present the PMU and runtime of SparseGPT, Wanda, and our proposed method when pruning OPT-6.7B on an NVIDIA RTX 3090. The experimental setup is consistent with the description in 4.1. Our method employs a commonly used chunked storage strategy, where intermediate outputs of each layer are stored in host memory instead of GPU memory, effectively reducing memory usage. While this approach may slow down the pruning process, as discussed in Section 4.6, a single epoch of fine-tuning is sufficient to significantly improve" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9493" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 217, + 165 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 217, + 165 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 217, + 165 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 217, + 165 + ], + "type": "image", + "image_path": "cc8dfd7b1ba0b437667dde1e01afc20843dd57529ecb60469c0b5d7fbef34424.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 56, + 168, + 217, + 187 + ], + "lines": [ + { + "bbox": [ + 56, + 168, + 217, + 187 + ], + "spans": [ + { + "bbox": [ + 56, + 168, + 217, + 187 + ], + "type": "text", + "content": "(a) Perplexity performance of pruned OPT-125M with different " + }, + { + "bbox": [ + 56, + 168, + 217, + 187 + ], + "type": "inline_equation", + "content": "\\alpha" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 224, + 70, + 386, + 165 + ], + "blocks": [ + { + "bbox": [ + 224, + 70, + 386, + 165 + ], + "lines": [ + { + "bbox": [ + 224, + 70, + 386, + 165 + ], + "spans": [ + { + "bbox": [ + 224, + 70, + 386, + 165 + ], + "type": "image", + "image_path": "a9cc02cc1bfd6a93eb4d02ce06f676f906372b78ba0d01ecab5ceefbd16fb5f1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 224, + 168, + 386, + 187 + ], + "lines": [ + { + "bbox": [ + 224, + 168, + 386, + 187 + ], + "spans": [ + { + "bbox": [ + 224, + 168, + 386, + 187 + ], + "type": "text", + "content": "(b) Instance segmentation performance (IoU, %) of SAM-B with different " + }, + { + "bbox": [ + 224, + 168, + 386, + 187 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 224, + 168, + 386, + 187 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 394, + 70, + 553, + 166 + ], + "blocks": [ + { + "bbox": [ + 394, + 70, + 553, + 166 + ], + "lines": [ + { + "bbox": [ + 394, + 70, + 553, + 166 + ], + "spans": [ + { + "bbox": [ + 394, + 70, + 553, + 166 + ], + "type": "image", + "image_path": "bf9c50dd352ccf6218444dfd3632170d98261fe55abf943d3d2d3ab663dfb6ac.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 392, + 168, + 553, + 187 + ], + "lines": [ + { + "bbox": [ + 392, + 168, + 553, + 187 + ], + "spans": [ + { + "bbox": [ + 392, + 168, + 553, + 187 + ], + "type": "text", + "content": "(c) Perplexity performance of OPT-125M with different " + }, + { + "bbox": [ + 392, + 168, + 553, + 187 + ], + "type": "inline_equation", + "content": "\\beta" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 56, + 191, + 217, + 286 + ], + "blocks": [ + { + "bbox": [ + 56, + 191, + 217, + 286 + ], + "lines": [ + { + "bbox": [ + 56, + 191, + 217, + 286 + ], + "spans": [ + { + "bbox": [ + 56, + 191, + 217, + 286 + ], + "type": "image", + "image_path": "7a96db60711f9674db4441172da3117c8ce3d675ec2542ec1fbd3e190815335e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 56, + 289, + 217, + 308 + ], + "lines": [ + { + "bbox": [ + 56, + 289, + 217, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 289, + 217, + 308 + ], + "type": "text", + "content": "(d) Instance segmentation performance (IoU, %) of SAM-B with different " + }, + { + "bbox": [ + 56, + 289, + 217, + 308 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 56, + 289, + 217, + 308 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 225, + 191, + 386, + 286 + ], + "blocks": [ + { + "bbox": [ + 225, + 191, + 386, + 286 + ], + "lines": [ + { + "bbox": [ + 225, + 191, + 386, + 286 + ], + "spans": [ + { + "bbox": [ + 225, + 191, + 386, + 286 + ], + "type": "image", + "image_path": "25bbdfb3d8e95c75b128bb3808b9ee70688a98c43dc6d9f455e0feb7882a7174.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 224, + 289, + 385, + 308 + ], + "lines": [ + { + "bbox": [ + 224, + 289, + 385, + 308 + ], + "spans": [ + { + "bbox": [ + 224, + 289, + 385, + 308 + ], + "type": "text", + "content": "(e) Perplexity performance of OPT-125M with different calibration set size." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 394, + 191, + 554, + 287 + ], + "blocks": [ + { + "bbox": [ + 394, + 191, + 554, + 287 + ], + "lines": [ + { + "bbox": [ + 394, + 191, + 554, + 287 + ], + "spans": [ + { + "bbox": [ + 394, + 191, + 554, + 287 + ], + "type": "image", + "image_path": "80b480ac31d38361ce2a816ab553da4a67ef8bb15b7edcbfe762295bca57ebce.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 392, + 289, + 553, + 308 + ], + "lines": [ + { + "bbox": [ + 392, + 289, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 392, + 289, + 553, + 308 + ], + "type": "text", + "content": "(f) Instance segmentation performance (IoU, %) of SAM-B with different calibration set size." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 119, + 312, + 279, + 408 + ], + "blocks": [ + { + "bbox": [ + 119, + 312, + 279, + 408 + ], + "lines": [ + { + "bbox": [ + 119, + 312, + 279, + 408 + ], + "spans": [ + { + "bbox": [ + 119, + 312, + 279, + 408 + ], + "type": "image", + "image_path": "797e9a14b85ec63bf6edc8cd84fc204df5ae975d440c44637518be3aad9d6e4c.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 118, + 410, + 279, + 430 + ], + "lines": [ + { + "bbox": [ + 118, + 410, + 279, + 430 + ], + "spans": [ + { + "bbox": [ + 118, + 410, + 279, + 430 + ], + "type": "text", + "content": "(g) Perplexity performance of pruned OPT-125M with different number of compensation epochs." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 331, + 312, + 493, + 408 + ], + "blocks": [ + { + "bbox": [ + 331, + 312, + 493, + 408 + ], + "lines": [ + { + "bbox": [ + 331, + 312, + 493, + 408 + ], + "spans": [ + { + "bbox": [ + 331, + 312, + 493, + 408 + ], + "type": "image", + "image_path": "a30956c31be2c5da3c3603cb57eef34222fb9fc886497ad873d9e7b4d78e779d.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 410, + 492, + 430 + ], + "lines": [ + { + "bbox": [ + 331, + 410, + 492, + 430 + ], + "spans": [ + { + "bbox": [ + 331, + 410, + 492, + 430 + ], + "type": "text", + "content": "(h) Segmentation performance of SAM-B with different number of compensation epochs." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 439, + 555, + 462 + ], + "lines": [ + { + "bbox": [ + 55, + 439, + 555, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 439, + 555, + 462 + ], + "type": "text", + "content": "Figure 4. Performance metrics for pruned OPT-125M and SAM-B models under varying parameters " + }, + { + "bbox": [ + 55, + 439, + 555, + 462 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 439, + 555, + 462 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 439, + 555, + 462 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 439, + 555, + 462 + ], + "type": "text", + "content": ", calibration dataset sizes, and number of compensation epochs." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 483, + 296, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 296, + 590 + ], + "type": "text", + "content": "model performance. Consequently, data only needs to be loaded from host memory once, keeping the additional time overhead within an acceptable range. It is worth noting that we use full-precision fine-tuning rather than Parameter-Efficient Fine-Tuning (PEFT) techniques. If combined with PEFT methods, memory consumption could be further reduced. This advantage enables our method to efficiently prune larger models on desktop-class GPUs, thereby expanding its applicability." + } + ] + } + ], + "index": 17 + }, + { + "type": "table", + "bbox": [ + 99, + 625, + 253, + 693 + ], + "blocks": [ + { + "bbox": [ + 84, + 614, + 266, + 624 + ], + "lines": [ + { + "bbox": [ + 84, + 614, + 266, + 624 + ], + "spans": [ + { + "bbox": [ + 84, + 614, + 266, + 624 + ], + "type": "text", + "content": "Table 9. GPU Memory Consumption Comparison" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 99, + 625, + 253, + 693 + ], + "lines": [ + { + "bbox": [ + 99, + 625, + 253, + 693 + ], + "spans": [ + { + "bbox": [ + 99, + 625, + 253, + 693 + ], + "type": "table", + "html": "
MethodPMUTime(s)
SparseGPT8.3 GB1004
Wanda21 GB143
ICP (1-epoch)7.8 GB1551
ICP (12-epoch)7.8 GB4470
", + "image_path": "3201cc154090a50cfb2082a3d9885b2346923b7c100964d4a2aabc1ea10dc5f0.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 314, + 481, + 388, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 481, + 388, + 495 + ], + "spans": [ + { + "bbox": [ + 314, + 481, + 388, + 495 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we propose an efficient pruning method ICP, which is effective for both vision and language models while consuming limited GPU memory. ICP update each block individually by compensating for the inter-block errors introduced by pruning, then rearranges the sparsity distribution based on prior knowledge of Transformer and pruning method. It significantly improves the performance of models pruned at high sparsity levels without increasing GPU memory consumption. ICP, as it does not rely on strong priors specific to any particular model, demonstrates dual applicability to both vision and language models. This flexibility suggests that it could potentially be extended to multimodal models in the future. However, this work has limitations: (1) the hyperparameters " + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "content": " are not automatically tuned; (2) ICP has not been tested on multimodal models; (3) the potential of individual blocks has not been fully explored. These will be the focus of our future work." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9494" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 296, + 135 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 296, + 135 + ], + "type": "text", + "content": "This work is supported by Natural Science Foundation of China under Grant 62271465, Suzhou Basic Research Program under Grant SYG202338, and Jiangsu Province Science Foundation for Youths (NO. BK20240464)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 156, + 115, + 169 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 156, + 115, + 169 + ], + "spans": [ + { + "bbox": [ + 56, + 156, + 115, + 169 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 177, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 177, + 296, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 177, + 296, + 232 + ], + "spans": [ + { + "bbox": [ + 61, + 177, + 296, + 232 + ], + "type": "text", + "content": "[1] Yixiong Chen, Alan Yuille, and Zongwei Zhou. Which layer is learning faster? a systematic exploration of layerwise convergence rate for deep neural networks. In The Eleventh International Conference on Learning Representations, 2023. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 233, + 296, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 233, + 296, + 277 + ], + "spans": [ + { + "bbox": [ + 61, + 233, + 296, + 277 + ], + "type": "text", + "content": "[2] Christopher Clark, Kenton Lee, Ming-Wei Chang, Tom Kwiatkowski, Michael Collins, and Kristina Toutanova. Boolq: Exploring the surprising difficulty of natural yes/no questions. arXiv preprint arXiv:1905.10044, 2019. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 277, + 295, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 277, + 295, + 331 + ], + "spans": [ + { + "bbox": [ + 62, + 277, + 295, + 331 + ], + "type": "text", + "content": "[3] Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. Think you have solved question answering? try arc, the ai2 reasoning challenge. arXiv preprint arXiv:1803.05457, 2018. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 334, + 295, + 377 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 334, + 295, + 377 + ], + "spans": [ + { + "bbox": [ + 62, + 334, + 295, + 377 + ], + "type": "text", + "content": "[4] Peijie Dong, Lujun Li, Zhenheng Tang, Xiang Liu, Xinglin Pan, Qiang Wang, and Xiaowen Chu. Pruner-zero: Evolving symbolic pruning metric from scratch for large language models. arXiv preprint arXiv:2406.02924, 2024. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 379, + 296, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 379, + 296, + 443 + ], + "spans": [ + { + "bbox": [ + 62, + 379, + 296, + 443 + ], + "type": "text", + "content": "[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 445, + 296, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 445, + 296, + 487 + ], + "spans": [ + { + "bbox": [ + 62, + 445, + 296, + 487 + ], + "type": "text", + "content": "[6] Jonathan Frankle and Michael Carbin. The lottery ticket hypothesis: Finding sparse, trainable neural networks. In International Conference on Learning Representations, 2018. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 490, + 296, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 490, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 62, + 490, + 296, + 533 + ], + "type": "text", + "content": "[7] Elias Frantar and Dan Alistarh. Sparsegpt: Massive language models can be accurately pruned in one-shot. In International Conference on Machine Learning, pages 10323-10337. PMLR, 2023. 1, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 535, + 296, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 535, + 296, + 578 + ], + "spans": [ + { + "bbox": [ + 62, + 535, + 296, + 578 + ], + "type": "text", + "content": "[8] Song Han, Jeff Pool, John Tran, and William Dally. Learning both weights and connections for efficient neural network. Advances in neural information processing systems, 28, 2015. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 580, + 296, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 580, + 296, + 623 + ], + "spans": [ + { + "bbox": [ + 62, + 580, + 296, + 623 + ], + "type": "text", + "content": "[9] Moritz Hardt, Ben Recht, and Yoram Singer. Train faster, generalize better: Stability of stochastic gradient descent. In International conference on machine learning, pages 1225-1234. PMLR, 2016. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 58, + 624, + 295, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 624, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 58, + 624, + 295, + 667 + ], + "type": "text", + "content": "[10] Babak Hassibi and David Stork. Second order derivatives for network pruning: Optimal brain surgeon. In Advances in Neural Information Processing Systems. Morgan-Kaufmann, 1992. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 58, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 670, + 295, + 713 + ], + "type": "text", + "content": "[11] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 1" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[12] Lei Ke, Mingqiao Ye, Martin Danelljan, Yu-Wing Tai, Chi-Keung Tang, Fisher Yu, et al. Segment anything in high quality. Advances in Neural Information Processing Systems, 36, 2024. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 118, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 183 + ], + "type": "text", + "content": "[13] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 1, 2, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 185, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 185, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 317, + 185, + 553, + 239 + ], + "type": "text", + "content": "[14] Chunyuan Li, Cliff Wong, Sheng Zhang, Naoto Usuyama, Haotian Liu, Jianwei Yang, Tristan Naumann, Hoifung Poon, and Jianfeng Gao. Llava-med: Training a large language-and-vision assistant for biomedicine in one day. Advances in Neural Information Processing Systems, 36, 2024. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 241, + 553, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 241, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 317, + 241, + 553, + 296 + ], + "type": "text", + "content": "[15] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision - ECCV 2014, pages 740-755, Cham, 2014. Springer International Publishing. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 297, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 330 + ], + "type": "text", + "content": "[16] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 331, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 352 + ], + "type": "text", + "content": "[17] AI@Meta Llama Team. The llama3 herd of models, 2024. Accessed: 2024-08-01. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 354, + 553, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 354, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 317, + 354, + 553, + 397 + ], + "type": "text", + "content": "[18] Xinyin Ma, Gongfan Fang, and Xinchao Wang. Llm-pruner: On the structural pruning of large language models. Advances in neural information processing systems, 36:21702-21720, 2023. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 399, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 399, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 553, + 464 + ], + "type": "text", + "content": "[19] Mitch Marcus, Grace Kim, Mary Ann Marcinkiewicz, Robert MacIntyre, Ann Bies, Mark Ferguson, Karen Katz, and Britta Schasberger. The penn treebank: Annotating predicate argument structure. In Human Language Technology: Proceedings of a Workshop held at Plainsboro, New Jersey, March 8-11, 1994, 1994. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 466, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 466, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 317, + 466, + 553, + 510 + ], + "type": "text", + "content": "[20] Maciej A Mazurowski, Haoyu Dong, Hanxue Gu, Jichen Yang, Nicholas Konz, and Yixin Zhang. Segment anything model for medical image analysis: an experimental study. Medical Image Analysis, 89:102918, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 512, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 512, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 512, + 553, + 544 + ], + "type": "text", + "content": "[21] Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. Pointer sentinel mixture models. arXiv preprint arXiv:1609.07843, 2016. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 545, + 553, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 545, + 553, + 588 + ], + "spans": [ + { + "bbox": [ + 317, + 545, + 553, + 588 + ], + "type": "text", + "content": "[22] Todor Mihaylov, Peter Clark, Tushar Khot, and Ashish Sabharwal. Can a suit of armor conduct electricity? a new dataset for open book question answering. arXiv preprint arXiv:1809.02789, 2018. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 590, + 553, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 553, + 623 + ], + "type": "text", + "content": "[23] Quan Quan, Fenghe Tang, Zikang Xu, Heqin Zhu, and S Kevin Zhou. Slide-sam: Medical sam meets sliding window. In Medical Imaging with Deep Learning, 2024. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 624, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 657 + ], + "type": "text", + "content": "[24] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 553, + 713 + ], + "type": "text", + "content": "[25] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 5" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9495" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 620 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[26] Keisuke Sakaguchi, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. Winogrande: An adversarial winograd schema challenge at scale. Communications of the ACM, 64 (9):99-106, 2021. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "text", + "content": "[27] Mingjie Sun, Zhuang Liu, Anna Bair, and J Zico Kolter. A simple and effective pruning approach for large language models. In The Twelfth International Conference on Learning Representations, 2024. 1, 3, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "type": "text", + "content": "[28] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 294, + 274 + ], + "type": "text", + "content": "[29] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 275, + 294, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 275, + 294, + 307 + ], + "spans": [ + { + "bbox": [ + 56, + 275, + 294, + 307 + ], + "type": "text", + "content": "[30] Tycho FA van der Ouderaa, Markus Nagel, Mart Van Baalen, Yuki M Asano, and Tijmen Blankevoort. The IIm surgeon. arXiv preprint arXiv:2312.17244, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 308, + 294, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 308, + 294, + 352 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 294, + 352 + ], + "type": "text", + "content": "[31] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 354, + 294, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 294, + 385 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 294, + 385 + ], + "type": "text", + "content": "[32] Alex Wang. Glue: A multi-task benchmark and analysis platform for natural language understanding. arXiv preprint arXiv:1804.07461, 2018. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 387, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 387, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 387, + 294, + 430 + ], + "type": "text", + "content": "[33] Miles Williams and Nikolaos Aletras. How does calibration data affect the post-training pruning and quantization of large language models? arXiv preprint arXiv:2311.09755, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 432, + 294, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 432, + 294, + 475 + ], + "spans": [ + { + "bbox": [ + 56, + 432, + 294, + 475 + ], + "type": "text", + "content": "[34] Junde Wu, Wei Ji, Yuanpei Liu, Huazhu Fu, Min Xu, Yanwu Xu, and Yueming Jin. Medical sam adapter: Adapting segment anything model for medical image segmentation. arXiv preprint arXiv:2304.12620, 2023. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 477, + 294, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 477, + 294, + 509 + ], + "spans": [ + { + "bbox": [ + 56, + 477, + 294, + 509 + ], + "type": "text", + "content": "[35] Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. Hellaswag: Can a machine really finish your sentence? arXiv preprint arXiv:1905.07830, 2019. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 511, + 294, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 511, + 294, + 564 + ], + "spans": [ + { + "bbox": [ + 56, + 511, + 294, + 564 + ], + "type": "text", + "content": "[36] Renrui Zhang, Jiaming Han, Chris Liu, Peng Gao, Aojun Zhou, Xiangfei Hu, Shilin Yan, Pan Lu, Hongsheng Li, and Yu Qiao. Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 567, + 294, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 294, + 620 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 294, + 620 + ], + "type": "text", + "content": "[37] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 1" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9496" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/ICT_ Image-Object Cross-Level Trusted Intervention for Mitigating Object Hallucination in Large Vision-Language Models/a90828a4-8b1b-4072-b5af-bed317d47267_content_list.json b/2025/ICT_ Image-Object Cross-Level Trusted Intervention for Mitigating Object Hallucination in Large Vision-Language Models/a90828a4-8b1b-4072-b5af-bed317d47267_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8e549d475d4659389342c21ceafad8bc0096694d --- /dev/null +++ b/2025/ICT_ Image-Object Cross-Level Trusted Intervention for Mitigating Object Hallucination in Large Vision-Language Models/a90828a4-8b1b-4072-b5af-bed317d47267_content_list.json @@ -0,0 +1,1616 @@ +[ + { + "type": "text", + "text": "ICT: Image-Object Cross-Level Trusted Intervention for Mitigating Object Hallucination in Large Vision-Language Models", + "text_level": 1, + "bbox": [ + 119, + 128, + 879, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Junzhe Chen $^{1,2*}$ , Tianshu Zhang $^{1*}$ , Shiyu Huang $^{3}$ , Yuwei Niu $^{4}$ , Linfeng Zhang $^{5}$ , Lijie Wen $^{1\\dagger}$ , Xuming Hu $^{2\\dagger}$", + "bbox": [ + 233, + 202, + 759, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Tsinghua University, $^{2}$ The Hong Kong University of Science and Technology (Guangzhou), $^{3}$ Zhipu AI, $^{4}$ Chongqing University, $^{5}$ Shanghai Jiao Tong University.", + "bbox": [ + 133, + 239, + 861, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "chenjz24@mails.tsinghua.edu.cn,", + "bbox": [ + 313, + 275, + 676, + 291 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "wenlj@tsinghua.edu.cn, xuminghu97@gmail.com,", + "bbox": [ + 233, + 292, + 751, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 344, + 326, + 359 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite the recent breakthroughs achieved by Large Vision Language Models (LVLMs) in understanding and responding to complex visual-textual contexts, their inherent hallucination tendencies limit their practical application in real-world scenarios that demand high levels of precision. Existing methods typically either fine-tune the LVLMs using additional data, which incurs extra costs in manual annotation and computational resources or perform comparisons at the decoding stage, which may eliminate useful language priors for reasoning while introducing inference time overhead. Therefore, we propose ICT, a lightweight, training-free method that calculates an intervention direction to shift the model's focus towards different levels of visual information, enhancing its attention to high-level and fine-grained visual details. During the forward pass stage, the intervention is applied to the attention heads that encode the overall image information and the fine-grained object details, effectively mitigating the phenomenon of overly language priors, and thereby alleviating hallucinations. Extensive experiments demonstrate that ICT achieves strong performance with a small amount of data and generalizes well across different datasets and models. Our codes are publicly available at:https://github.com/THU-BPM/ICT/.", + "bbox": [ + 88, + 375, + 483, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 750, + 220, + 766 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The recent rapid advancements in Large Vision Language Models (LVLMs) have demonstrated significant potential to tackle complex real-world tasks [5, 27, 29, 52, 56]. However, LVLMs frequently generate text outputs that are inconsistent with the visual input, such as incorrectly determining the presence of objects in an image or inaccurately", + "bbox": [ + 89, + 775, + 482, + 867 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d5ea88bbeb089521af97a2e2f50af36678eb1f491b5330addf3d1535fd0d75e6.jpg", + "image_caption": [ + "Figure 1. Comparison between Contrastive Decoding (top) and our proposed ICT (bottom). In the top example, Contrastive Decoding indiscriminately removes both beneficial and detrimental language priors, leading to hallucinations. In contrast, our approach enhances the model's attention to visual details while preserving useful language priors, allowing it to correctly identify and describe objects in the image." + ], + "image_footnote": [], + "bbox": [ + 517, + 344, + 903, + 539 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "describing the attributes of those objects [55, 73]. This issue is commonly referred to as the object hallucination phenomenon [55], which significantly limits the applicability of LVLMs in high-stakes scenarios that demand precise accuracy, such as autonomous driving [95] and medical surgery [35, 83]. Previous research suggested that the primary reasons for hallucinations in LVLMs are as follows: 1) Large language models (LLMs), such as Vicuna [12], possess significantly stronger parameters and capabilities compared to visual encoders, such as CLIP [68]. This results in an excessively strong language prior, causing the model to overly rely on language cues while neglecting visual input [24, 25, 40, 43, 55]. 2) Current visual decoders often struggle to accurately capture fine-grained visual semantics, resulting in errors in detailed object attributes in an image, such as color and quantity [4, 9, 13, 55].", + "bbox": [ + 511, + 642, + 906, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Based on the two issues outlined above, previous ef", + "bbox": [ + 531, + 885, + 903, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal Contribution.", + "bbox": [ + 114, + 875, + 232, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding authors.", + "bbox": [ + 114, + 887, + 250, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4209", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "forts to mitigate the hallucination phenomenon in LVLMs can be categorized into three main approaches: 1) Finetuning with additional data: This approach involves introducing high-quality annotated data to better align the model's behavior with human interpretations, effectively teaching the model to focus more on visual information [23, 46, 52, 106]. However, this method not only requires costly human annotation but also involves updating model parameters, which demands substantial computational resources, thereby limiting its scalability. 2) Perceptual enhancement: This method incorporates additional information, such as depth maps and segmentation maps [34, 38, 106], as auxiliary inputs to assist the visual encoder in capturing more detailed visual features, thereby reducing hallucinations. However, it often requires manual selection of auxiliary features, which limits its generalizability across different tasks. 3) Contrastive decoding: This method alleviates hallucinations without requiring additional training. It induces hallucinations by applying transformations such as blurring, rotation, or cropping to the original visual input. During the decoding stage, tokens associated with these induced hallucinations are penalized, thus mitigating the influence of language priors [10, 14, 36, 43, 64, 70, 82, 87, 109]. However, methods such as VCD often indiscriminately eliminate all language priors, including those that may be beneficial. As shown in Figure 1, the original model recognizes Curry holding a golf club, while utilizing the language prior that identifies Curry as a basketball player. This language prior is valuable for achieving an accurate interpretation. However, through contrastive decoding, this useful prior is also removed, which can inadvertently result in hallucinations.", + "bbox": [ + 89, + 90, + 485, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the challenges of mitigating hallucinations in LVLMs, we propose Image-Object Cross-Level Trusted Intervention (ICT), a training-free, plug-and-play method applicable during the forward pass. Unlike contrastive decoding, our approach does not eliminate language priors to reduce the model's over-reliance on text modality. Instead, it intervenes during the forward pass to enhance the model's focus on both comprehensive visual information and fine-grained object details. We explored the activation patterns of attention heads when the model produces correct versus hallucinated responses. This analysis allowed us to identify activation value deviations that can shift the model from being \"untrustworthy\" to \"trustworthy\". According to prior research [8, 19, 81], which has shown that different heads in the multi-head attention mechanism encode information at varying levels of granularity, we train binary classifiers for each head to determine which heads encode overall visual information and which capture detailed visual features. During the forward pass, we adjust the activation values of these heads based on their identified levels, thereby enhancing the model's attention to the relevant visual features and", + "bbox": [ + 89, + 583, + 486, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "reducing the likelihood of hallucinations. As illustrated in Figure 1, after applying ICT, the model is able to focus more on the details within the image, such as identifying the man as Curry, while simultaneously utilizing beneficial language priors (e.g., Curry is a basketball player) to infer and arrive at the correct answer. Since the intervention shift vectors are pre-computed, ICT does not introduce additional delays during the forward pass.", + "bbox": [ + 511, + 90, + 906, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our experiments demonstrate that, for both LLaVAv1.5 [52] and Qwen-VL [5], applying ICT results in an average improvement of $6.27\\%$ on POPE benchmark and 67.37 points on MME benchmark. Furthermore, ICT exhibits cross-dataset generalization and model-agnostic generalizability. Our contributions can be summarized as:", + "bbox": [ + 511, + 213, + 908, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We propose Image-Object Cross-Level Trusted Intervention (ICT), a novel, training-free, plug-and-play method that effectively reduces hallucinations in LVLMs by enhancing focus on both overall visual information and fine-grained object details during the forward pass, without eliminating beneficial language priors.", + "bbox": [ + 511, + 304, + 906, + 395 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Unlike existing contrastive decoding approaches, we introduce an intervention mechanism that operates during the forward pass rather than the decoding stage, allowing ICT to be orthogonal and complementary to existing decoding strategies while not introducing any additional latency.", + "bbox": [ + 511, + 396, + 906, + 472 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Extensive experiments on LLaVA-v1.5 and Qwen-VL demonstrate that ICT significantly improves performance on both the POPE and MME benchmarks while maintaining cross-dataset and model-agnostic generalizability.", + "bbox": [ + 511, + 473, + 905, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 549, + 656, + 565 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Large Vision-Language Models", + "text_level": 1, + "bbox": [ + 511, + 575, + 792, + 590 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Following the success of LLM [6, 12, 61, 78, 80], researchers have been exploring the multimodal field. Taking advantage of the remarkable capabilities of LLMs, Large Vision-Language Models incorporate visual encoders and feature projectors into powerful LLMs, enabling them to understand and generate content based on visual and text input. Most of these models undergo two training steps, namely the pre-training stage and the fine-tuning stage. The main purpose of the pre-training stage is to align textual and visual features, while the goal of the fine-tuning stage is to bridge further the modality gap between vision and language, as well as lifting their instruction following capabilities and performance on specific downstream tasks. Early attempts, such as FLamingo [2], Gemini [79], and BLIP-2 [44], have already shown promising results. Recent works like LLaVA-v1.5 [54], Qwen2-VL [84], and xgenmm [92] have advanced this field, dramatically improving the abilities of these models. To enhance the consistency between vision representation and language representation, many efforts have been made, including utilizing higher res", + "bbox": [ + 511, + 598, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4210", + "bbox": [ + 482, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "olution vision encoders, transitioning to larger and more capable LLMs, and adapting reinforcement learning techniques such as RLHF [62, 98], among others. Despite the remarkable progress that has been made, LVLMs still suffer from severe hallucinations, which restricts their potential for large-scale applications in real-world scenarios.", + "bbox": [ + 89, + 90, + 483, + 183 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Mitigating Hallucinations in LVLMs", + "text_level": 1, + "bbox": [ + 89, + 195, + 408, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Many efforts have been made to understand the causes of hallucination [1, 16, 30, 39, 50, 50, 53, 63, 65, 71, 72, 76, 93, 101]. Existing approaches to mitigate hallucinations can be roughly divided into two categories based on the stage they take place. The first category focuses on the training phase. Most of the work falling into this category introduces additional or curated datasets [21, 22, 51, 97], these works usually design a dataset specifically for hallucination-related tasks or improve data cleaning methods [69, 75], or introduce novel training objectives [7, 41, 77]. While effective, these methods typically require extensive training, which is both time-consuming and resource-intensive. The second category focuses on the inferencing phase, usually involves CD-based [47] novel decoding strategies [10, 14, 36, 59, 64, 70, 82, 87, 94, 107, 109]. Another approach is to detect potential hallucinations while generating and correct them [18, 37, 60], among others [45, 86, 100, 103]. In addition, some researchers tackle hallucinations by manipulating attention weight assigned to the image [33, 89, 90, 104, 108], or the prompt-related part of the image [3]. Other works include prompt-based methods [42, 67, 85, 88, 91, 102], utilizing external tools [11, 96, 105] or external knowledge [17, 28, 66] and so on.", + "bbox": [ + 91, + 219, + 483, + 568 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Task Formulation", + "text_level": 1, + "bbox": [ + 89, + 585, + 264, + 601 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given an LVLM parameterized by $\\theta$ , the model processes a textual input $\\pmb{X} = \\{x_{1}, x_{2}, \\ldots, x_{N_{x}}\\}$ and a visual input $\\pmb{V} = \\{v_{1}, v_{2}, \\ldots, v_{N_{v}}\\}$ , where $N_{x}$ and $N_{v}$ denote the sequence lengths of the text and visual inputs, respectively. The model then concatenates the textual and visual sequences to form a unified input $\\pmb{H} = \\text{concat}(\\pmb{V}, \\pmb{X})$ , which is subsequently passed through $L$ layers of a transformer architecture. At each layer, the concatenated input undergoes multi-head attention, which is computed as:", + "bbox": [ + 89, + 612, + 483, + 750 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {H} ^ {(l + 1)} = \\boldsymbol {H} ^ {(l)} + \\sum_ {n = 1} ^ {N} A t t n _ {n} ^ {(l)} \\left(\\boldsymbol {H} ^ {(l)}\\right) \\cdot W _ {o} ^ {(l)}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 763, + 482, + 805 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $Attn_{n}^{(l)}$ denotes the attention operation of the $n$ -th head at the $l$ -th layer, and $W_{o}^{(l)} \\in \\mathbb{R}^{d \\times Nd}$ is the output projection matrix, with $d$ representing the dimensionality of each attention head and $N$ the number of heads. Subsequently, the model autoregressively predicts the next token", + "bbox": [ + 89, + 821, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "based on the output from the $L$ -th layer:", + "bbox": [ + 511, + 90, + 781, + 106 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np \\left(y _ {t} \\mid y _ {< t}\\right) = \\operatorname {S o f t m a x} \\left(f c \\left(\\boldsymbol {H} _ {t} ^ {(L)}\\right)\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 583, + 113, + 906, + 132 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $f c(\\cdot)$ is an affine layer mapping the $Hd$ -dimensional vector to a probability distribution of vocabulary size $M$ .", + "bbox": [ + 511, + 138, + 906, + 170 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Methodology", + "text_level": 1, + "bbox": [ + 511, + 181, + 648, + 198 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce two levels of fine-grained intervention modules. The first module enables the LVLM to focus on the image, thereby reducing over-reliance on language priors. The second module encourages the LVLM to attend more closely to image objects, helping to mitigate the omission of critical objects and reduce hallucinations.", + "bbox": [ + 511, + 205, + 905, + 297 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Intervention at Image-Level", + "text_level": 1, + "bbox": [ + 511, + 305, + 764, + 320 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This module aims to identify the attention heads associated with overall image information and to apply targeted interventions to these heads. This approach enhances the model's focus on visual input while diminishing the influence of language priors.", + "bbox": [ + 511, + 325, + 905, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As illustrated in Figure 2, consider a batch of image-question pairs, denoted as $\\{(X_i,V_i)\\}_{i = 1}^B$ , each labeled with the answer \"Yes\" from the POPE dataset. Here, $X_{i}$ takes the form of \"Is there a/an [object] in the image?\" For each question, we extract the specified object $O_{i}$ and reformulate the question to $q =$ \"What is it in the image?\" Then, for each image $V_{i}$ , we progressively add Gaussian noise following the forward diffusion process [26] to obtain the final blurred image $V_{i}^{\\prime}$ :", + "bbox": [ + 511, + 402, + 905, + 539 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} q (V _ {i} ^ {(t)} \\mid V _ {i} ^ {(t - 1)}) = \\mathcal {N} (V _ {i} ^ {(t)}; \\sqrt {1 - \\beta_ {t}} V _ {i} ^ {(t - 1)}, \\beta_ {t} \\mathbf {I}), \\\\ q \\left(V _ {i} ^ {\\prime} \\mid V _ {i}\\right) = \\prod_ {t = 1} ^ {T} q \\left(V _ {i} ^ {(t)} \\mid V _ {i} ^ {(t - 1)}\\right), \\tag {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 544, + 903, + 608 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\beta_{t}$ is the noise variance at step $t$ . Finally, we construct a modified dataset $\\{(q,V_i,V_i',O_i)\\}_{i=1}^B$ to obtain Image-Level intervention vectors. We then concatenate $O_i$ with the question $q$ to form the answer, resulting in data pairs of trusted and untrusted data for the same sample: $(q + O_i,V_i)$ and $(q + O_i,V_i')$ . For each sample, we treat the representation of the last token as the fused representation of multimodal data and extract the attention activation from a total of $N\\times L$ heads across the $L$ layers for both trusted and untrusted data. These are denoted as $A_{i} = \\{Attn_{n}^{(l)}H_{s}^{(l)}\\}_{n=1,l=1}^{N,L}$ and $A_{i}' = \\{Attn_{n}^{(l)}H_{s}^{(l)}\\}_{n=1,l=1}^{N,L}$ , where $s = N_x + N_v + 1$ . According to the trusted activation $\\{A_{i}\\}_{i=1}^{B}$ and the untrusted activation $\\{A_{i}'\\}_{i=1}^{B}$ obtained from all sample pairs, we can calculate the following activation shift vector $\\{S_{n}^{(l)}\\}_{n=1,l=1}^{N,L}$ that encourages the model to pay more attention to visual information, as follows:", + "bbox": [ + 511, + 614, + 906, + 857 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {S} _ {n} ^ {(l)} = \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\left(\\boldsymbol {A} _ {i, n} ^ {(l)} - \\boldsymbol {A} _ {i, n} ^ {\\prime (l)}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 864, + 906, + 905 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4211", + "bbox": [ + 480, + 944, + 513, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/609bed94047a25603363ccc606374ecd88d35da5bdb7fac0915277db876cde1b.jpg", + "image_caption": [ + "Figure 2. Overview of our proposed ICT method. ICT consists of two levels of intervention: Image-Level and Object-Level. The Image-Level module enhances the model's focus on the overall scene to reduce reliance on language priors, while the Object-Level module focuses on fine-grained details to mitigate hallucinations. We apply targeted activation shifts to selected attention heads identified through binary classifiers trained to distinguish trusted and untrusted data pairs." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 908, + 290 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Next, we train a binary classifier $f_{n}^{(l)}(\\cdot)$ using $B$ sample pairs for each head to detect which heads encode Image-Level information, specifically those that can better distinguish the differences between pairs of trusted and untrusted samples. We then apply activation interventions to these selected heads:", + "bbox": [ + 89, + 364, + 483, + 455 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nA t t n _ {\\mathrm {i m g}} = \\left\\{A t t n _ {n} ^ {(l)} \\mid A t t n _ {n} ^ {(l)} \\in T o p K (\\text {A c c u r a c y} (f _ {n} ^ {(l)} (\\cdot))) \\right\\},\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 465, + 483, + 487 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {H} ^ {(l + 1)} = \\boldsymbol {H} ^ {(l)} + \\sum_ {n = 1} ^ {N} \\left(A t t n _ {n} ^ {(l)} \\left(\\boldsymbol {H} ^ {(l)}\\right) + \\mathbb {I} _ {\\operatorname {i m g}, n} ^ {(l)} \\alpha \\boldsymbol {S} _ {n} ^ {(l)}\\right) \\cdot W _ {o} ^ {(l)}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 489, + 483, + 542 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $K$ is the number of selected intervention heads, $\\mathbb{I}_{\\mathrm{img},n}^{(l)}$ is an indicator function, which is 1 if $Attn_{n}^{(l)} \\in Attn_{\\mathrm{img}}$ and 0 otherwise, and $\\alpha$ denotes the intensity of the intervention. After using $\\{S_n^{(l)}\\}_{n=1,l=1}^{N,L}$ to perform Image-Level interventions on heads that encode image information, the model enhances the trustworthiness of the visual level, places greater attention on visual information, thus mitigates the impact of overly strong language priors.", + "bbox": [ + 89, + 542, + 483, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Intervention at Object-Level", + "text_level": 1, + "bbox": [ + 89, + 681, + 346, + 698 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After enhancing the model's trustworthiness at the Image-Level, a more fine-grained, Object-Level intervention becomes necessary to increase the model's attention to image details, thereby reducing hallucinations caused by the omission of fine details. As shown in Figure 2, similar to Section 4.1, we construct a dataset $\\{(q,V_i,V_i^{\\prime \\prime},O_i)\\}_{i = 1}^{B}$ to obtain Object-Level intervention vectors. We use Grounding DINO [58] to identify the area of object $O_{i}$ in image $V_{i}$ . Gaussian noise is then added selectively to this object region, based on Eq. 3, producing a locally blurred image $V_{i}^{\\prime \\prime}$ . Using both the original image $V_{i}$ and the partially blurred image $V_{i}^{\\prime \\prime}$ , we construct trusted and untrusted data pairs for each sample: $(q + O_i,V_i)$ and $(q + O_i,V_i^{\\prime \\prime})$ and analyze", + "bbox": [ + 89, + 704, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the attention activation values across $N\\times L$ heads, allowing us to compute an Object-Level activation shift vector $\\{S_{n,\\mathrm{obj}}^{(l)}\\}_{n = 1,l = 1}^{N,L}$", + "bbox": [ + 511, + 366, + 906, + 415 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {S} _ {n, \\mathrm {o b j}} ^ {(l)} = \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\left(\\boldsymbol {A} _ {i, n} ^ {(l)} - \\boldsymbol {A} _ {i, n} ^ {\\prime \\prime (l)}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 417, + 906, + 458 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A binary classifier $f_{n,obj}^{(l)}(\\cdot)$ is then trained to identify heads that effectively distinguish trusted from untrusted object-focused samples. Interventions are subsequently applied to the selected heads as follows:", + "bbox": [ + 511, + 468, + 906, + 530 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nA t t n _ {\\mathrm {o b j}} = \\left\\{A t t n _ {n} ^ {(l)} \\mid A t t n _ {n} ^ {(l)} \\in T o p K (\\text {A c c u r a c y} (f _ {n, o b j} ^ {(l)} (\\cdot))) \\right\\},\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 541, + 911, + 564 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {H} ^ {(l + 1)} = \\boldsymbol {H} ^ {(l)} + \\sum_ {n = 1} ^ {N} \\left(A t t n _ {n} ^ {(l)} \\left(\\boldsymbol {H} ^ {(l)}\\right) + \\mathbb {I} _ {\\mathrm {o b j}, n} ^ {(l)} \\beta \\boldsymbol {S} _ {\\mathrm {o b j}, n} ^ {(l)}\\right) \\cdot W _ {o} ^ {(l)}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 566, + 910, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $K$ is the number of intervention heads, $\\mathbb{I}_{\\mathrm{obj},n}^{(l)}$ is an indicator function set to 1 if $Attn_{n}^{(l)} \\in Attn_{\\mathrm{obj}}$ and 0 otherwise, and $\\beta$ controls the intervention intensity.", + "bbox": [ + 511, + 619, + 905, + 671 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, we integrate the Image-Level and Object-Level intervention modules to create a unified approach that reinforces the model's focus on both the overall visual context and finer object-specific details as follows:", + "bbox": [ + 511, + 672, + 906, + 733 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\boldsymbol {H} ^ {(l + 1)} = \\boldsymbol {H} ^ {(l)} + \\sum_ {n = 1} ^ {N} \\left(A t t n _ {n} ^ {(l)} \\left(\\boldsymbol {H} ^ {(l)}\\right) \\right. \\tag {8} \\\\ \\left. \\right.\\left. + \\mathbb {I} _ {\\operatorname {i m g}, n} ^ {(l)} \\alpha \\boldsymbol {S} _ {n} ^ {(l)} + \\mathbb {I} _ {\\operatorname {o b j}, n} ^ {(l)} \\beta \\boldsymbol {S} _ {\\operatorname {o b j}, n} ^ {(l)}\\right) \\cdot W _ {o} ^ {(l)}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 744, + 903, + 813 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By applying these interventions in tandem, the model gains a balanced attention mechanism that mitigates excessive reliance on language priors while enhancing sensitivity to essential visual cues across different levels of granularity, thereby alleviating the occurrence of hallucinations.", + "bbox": [ + 511, + 825, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4212", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 89, + 223, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 114, + 279, + 132 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets and Metrics.", + "text_level": 1, + "bbox": [ + 89, + 138, + 246, + 152 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "POPE [48] is a benchmark designed to assess how effectively LVLMs recognize the presence of specific objects in images, thus identifying Object-Level hallucinations. It uses Yes/No questions based on object annotations, with metrics including Accuracy, Precision, Recall, and F1 score. The dataset is balanced, with $50\\%$ of queries for existing objects and $50\\%$ for non-existing ones, and employs three sampling strategies: random, popular, and adversarial. Drawing from MSCOCO [49], A-OKVQA [74], and GQA [32] datasets, POPE evaluates 27,000 query-answer pairs to gauge model performance.", + "bbox": [ + 88, + 152, + 482, + 319 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "MME [20] benchmark is designed to comprehensively evaluate the performance of LVLMs across various dimensions. It includes ten perception-focused tasks and four cognition-related tasks. MME benchmark specifically addresses Object-Level hallucinations through subsets focusing on object existence and count, while attribute-level hallucinations are assessed via subsets related to object position and color. The evaluation metric is Accuracy.", + "bbox": [ + 89, + 320, + 483, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. We adopt the widely used LLaVA-v1.5 [52] and Qwen-VL [5] models as our baseline LVLMs. We compared two baselines that eliminate language priors of the LVLMs in the decoding stage to alleviate hallucinations: VCD [43] and Opera [31].", + "bbox": [ + 89, + 441, + 483, + 516 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. In our experiments, we utilized 1,500 QA pairs with \"Yes\" responses from the COCO Random subset of the POPE dataset to train the intervention shift vector. Subsequently, we evaluated the ICT approach on two datasets with significant distributional differences: POPE, and MME. This evaluation was conducted to assess the generalizability and robustness of ICT across diverse data distributions. For each attention head, we employed Support Vector Machines (SVMs) [15] as the classifier and performed 2-fold cross-validation to evaluate the classification accuracy. In the experiment, we set $\\alpha = \\beta$ and determined the optimal values of $\\alpha$ , $\\beta$ , and $K$ through a grid search. Detailed hyperparameter configurations are provided in Appendix A. All experiments were conducted on a system equipped with $8 \\times$ H800 GPUs.", + "bbox": [ + 89, + 518, + 483, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Main Results", + "text_level": 1, + "bbox": [ + 89, + 756, + 230, + 770 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Results on POPE. Table 1 presents the results of LLaVA-v1.5 and Qwen-VL on nine subsets of the POPE dataset. From comparing these methods, we can derive the following conclusions: 1) Applying ICT results in an average performance improvement on the F1 score for LLaVA-v1.5 and Qwen-VL of $7.09\\%$ and $5.44\\%$ on 9 subsets respectively. This improvement can be attributed to the fact that ICT does not eliminate language priors, which may provide", + "bbox": [ + 89, + 777, + 482, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "useful information. Instead, it enhances the model's attention to various levels of visual information, thereby reducing the model's tendency to overly rely on language priors and mitigating the occurrence of hallucinations. 2) Individually intervening at both the image level and object level, with average gains of $5.76\\%$ and $5.47\\%$ in F1 score, respectively. This demonstrates that enhancing LVLMs' attention to image information at various levels can effectively mitigate hallucinations. Moreover, since Object-Level intervention also implicitly directs the model's focus toward broader image information, it achieves a comparatively higher performance gain. 3) The intervention shift vector trained using 1,500 samples from the MSCOCO random subset achieves an average improvement on F1 score of $7.67\\%$ on MSCOCO random and an average improvement of $6.09\\%$ across the other eight subsets with different distributions. This indicates that ICT generalizes effectively, and the resulting intervention vector captures a general direction towards trustworthiness rather than fitting a specific dataset.", + "bbox": [ + 511, + 90, + 906, + 380 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Results on MME. Figure 3 presents the experimental results on the MME benchmark, where we followed Leng et al. [43] evaluation setup, focusing on the hallucination subset of MME. Compared to the unmodified model, ICT achieved a score improvement of 80.51 on LLaVA-v1.5 and 34.67 on Qwen-VL, surpassing the performance of VCD. Additionally, on non-hallucination tasks such as commonsense QA, ICT demonstrated a 10-point improvement on LLaVA-v1.5 and a 9.5-point improvement on Qwen-VL. These results indicate that ICT not only effectively reduces hallucinations but also enhances general reasoning capabilities across different models. Moreover, since the objects present in the images of the MME dataset are generally prominent, the variant ICT $w/o$ image provides only limited improvements compared to ICT $w/o$ object. This demonstrates that the absence of Image-Level guidance restricts the model's ability to fully leverage global visual context, thereby reducing its effectiveness in scenarios where the dominant objects play a crucial role.", + "bbox": [ + 511, + 404, + 908, + 691 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/83db3db20d370ddf3feed3811612747a8a58a4c6ea5e3e8026b7fa1600971d08.jpg", + "image_caption": [ + "Figure 3. Comparison of ICT with baseline methods (Vanilla and VCD) on the MME benchmark. The radar chart illustrates improvements across various evaluation categories, including existence, position, count, color, and commonsense QA (CSQA)." + ], + "image_footnote": [], + "bbox": [ + 516, + 715, + 903, + 858 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4213", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/07bf9695ea90362d813b7e1a9618e879f909cb9861e7b977c62fb3787f4dbee2.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. Main results on POPE tasks. We evaluate the accuracy of various LVLMs on the POPE task across the MSCOCO, A-OKVQA, and GQA datasets. Regular represents the setting where direct sampling is applied. ICT $w/o$ image and ICT $w/o$ object correspond to the exclusion of image-level and object-level interventions, respectively. The bold and underlined values indicate the highest and second-highest metrics under each setting, respectively." + ], + "table_body": "
DatasetSettingMethodLLaVA-v1.5Qwen-VL
AccuracyF1 ScoreAccuracyF1 Score
COCORandomRegular83.2981.3384.3782.67
VCD87.7387.1688.6387.81
OPERIA89.2088.8187.3186.92
ICT w/o image89.70 (+6.41)89.93 (+8.60)88.76(+4.39)87.84(+5.17)
ICT w/o object88.50 (+5.21)88.81 (+7.48)89.13(+4.76)88.20(+5.63)
ICT90.11 (+6.82)90.03 (+8.70)89.46(+5.09)89.20(+6.63)
PopularRegular81.8880.0684.1382.06
VCD85.3885.0687.1286.40
OPERIA86.6486.6287.4486.68
ICT w/o image87.43 (+5.55)87.09 (+7.03)87.63(+3.50)86.78(+4.72)
ICT w/o object87.30 (+5.42)86.65 (+6.59)87.10(+2.97)86.28(+4.22)
ICT87.50 (+5.62)87.60 (+7.54)88.16(+4.03)87.33(+5.27)
AdversarialRegular78.9677.5782.2680.37
VCD80.8881.3384.2683.90
OPERIA81.2481.3884.7883.45
ICT w/o image84.13 (+5.17)83.31 (+5.74)85.13(+2.87)84.35(+3.98)
ICT w/o object83.63 (+4.67)82.13 (+4.56)85.23(+2.97)84.62(+4.25)
ICT84.43 (+5.47)83.74 (+6.17)84.96(+2.70)84.42(+4.05)
AOKVQARandomRegular83.4582.5686.6785.59
VCD86.1586.3489.2289.01
OPERIA88.0284.5988.1988.43
ICT w/o image89.60 (+6.15)89.75 (+7.19)89.46(+2.79)89.28(+3.69)
ICT w/o object89.00 (+5.55)88.91 (+6.35)89.26(+2.59)88.95(+3.36)
ICT89.20 (+5.75)89.41 (+6.85)89.46(+2.79)89.03(+3.44)
PopularRegular79.9079.5985.5684.63
VCD81.8582.8287.8587.81
OPERIA83.2284.6787.9187.13
ICT w/o image85.36 (+5.46)85.34 (+5.75)87.70(+2.14)87.68(+3.05)
ICT w/o object84.65 (+4.75)84.32 (+4.73)87.90(+2.34)87.48(+2.85)
ICT85.73 (+5.83)85.34 (+5.75)88.13(+3.57)87.83(+3.20)
AdversarialRegular74.0475.1579.5779.50
VCD74.9777.7381.2782.38
OPERIA73.8277.9180.8281.54
ICT w/o image79.26 (+5.22)80.14 (+4.99)81.50(+1.93)82.10(+2.60)
ICT w/o object77.79 (+3.75)79.53 (+4.38)81.40(+1.83)82.18(+2.68)
ICT79.60 (+5.56)80.43 (+5.28)81.94(+2.37)82.44(+2.94)
GQARandomRegular83.7382.9580.9779.01
VCD86.6586.9985.5985.33
OPERIA88.1388.9186.0285.29
ICT w/o image89.03 (+5.30)88.99 (+6.04)85.20(+4.23)85.90(+6.89)
ICT w/o object88.97 (+5.24)89.30 (+6.35)87.20(+6.23)86.80(+7.79)
ICT89.60 (+5.87)89.44 (+6.49)86.38(+5.41)86.96(+7.95)
PopularRegular78.1778.3775.9974.84
VCD80.7382.2481.8382.23
OPERIA79.2782.1181.9782.12
ICT w/o image84.23 (+6.06)83.92 (+5.55)81.43(+5.44)80.86(+6.02)
ICT w/o object84.70 (+6.53)84.45 (+6.08)83.93(+7.94)82.95(+8.11)
ICT84.70 (+6.53)84.78 (+6.41)82.63(+6.64)82.22(+7.38)
AdversarialRegular75.0876.0675.4674.33
VCD76.0978.7880.0180.75
OPERIA75.0078.7180.2480.64
ICT w/o image81.92 (+6.84)82.11 (+6.05)80.03(+4.57)79.90(+5.47)
ICT w/o object81.65 (+6.57)82.00 (+5.94)81.60(+6.14)81.53(+7.10)
ICT81.50 (+6.42)82.27 (+6.21)80.83(+5.37)80.60(+6.17)
", + "bbox": [ + 151, + 87, + 848, + 821 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4214", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Analysis and Discussions", + "text_level": 1, + "bbox": [ + 89, + 89, + 323, + 107 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.1. Inference Latency across Different Methods", + "text_level": 1, + "bbox": [ + 89, + 114, + 464, + 131 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Table 2, we compare the inference efficiency of two different representative approaches with our proposed method. Notably, while the CD-based methods result in a significant increase in inference time (2.1 to 3.6 times slower), our ICT incurs virtually no impact on inference efficiency. We attribute this to the fact that both VCD and OPERA either require multiple inference runs or involve substantial additional computations during inference. In contrast, ICT introduces no extra computational cost at inference, further highlighting the advantages of our method.", + "bbox": [ + 88, + 136, + 482, + 287 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/536e03754b19c21243e41041bd5458be0c00f6d7983b662c8db62dab90bfca8a.jpg", + "table_caption": [], + "table_footnote": [ + "Table 2. Comparison of the efficiency of different methods in generating tokens of varying lengths on an NVIDIA H800 GPU. Inference times are recorded in milliseconds." + ], + "table_body": "
Method20-Token Len50-Token Len80-Token Len
LLaVA-v1.5405.3 ↑×1.0934.6 ↑×1.01440.0 ↑×1.0
+ VCD988.9 ↑ × 2.42031.7 ↑ × 2.23077.5 ↑ × 2.1
+ OPERA1371.7 ↑ × 3.43294.1 ↑ × 3.55717.3 ↑ × 3.6
+ ICT415.9 ↑ × 1.0931.9 ↑ × 1.01485.5 ↑ × 1.0
", + "bbox": [ + 96, + 290, + 475, + 371 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.2. Generalizability between Different Models", + "text_level": 1, + "bbox": [ + 89, + 421, + 452, + 438 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To verify that the derived activation shift vectors to some extent represent the model's alignment with authenticity, we explored whether these shifts are transferable to other models. Specifically, we applied the activation shift vectors obtained from LLaVA-v1.5 on the COCO random subset to intervene in Qwen-VL. We then evaluated the generalizability of this intervention on the GQA dataset, which has a distribution significantly different from COCO. The results in Table 3 demonstrate that ICT-LLaVA-v1.5 achieves an average improvement of $4.62\\%$ in F1 score and $4.64\\%$ in accuracy compared to the unmodified model. Additionally, we employed t-SNE to visualize the offset vectors across different layers for LLaVA-v1.5 and Qwen-VL. Figure 4 shows the Object-Level and Image-Level shift vectors of multi-head attention in the 16th and 18th layers for both models. From the figure, we can observe that in layer 16, the Image-Level vectors (blue and yellow) are relatively close to each other, indicating a degree of similarity in encoding Image-Level information between the models. However, the Object-Level vectors (red and green) are more distinct, reflecting model-specific characteristics in encoding fine-grained object details. In layer 18, the Image-Level and Object-Level vectors for both models exhibit high similarity, indicating that the resulting activation shift vectors represent universal shift vectors that guide the model to pay more attention to visual information.", + "bbox": [ + 91, + 444, + 482, + 834 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.3. Impact of Hyperparameters on Performance", + "text_level": 1, + "bbox": [ + 89, + 834, + 470, + 849 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our proposed ICT primarily relies on two key hyperparameters: the intervention intensity $\\alpha$ and the number of heads $K$ involved in the intervention. To investigate the", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/12ba79e155dd7a4d0e372b6e33b40098d7e630c396520c34f5c5c0ab79bfac0c.jpg", + "image_caption": [ + "Figure 4. t-SNE visualization of Object-Level and Image-Level offset vectors for LLaVA-v1.5 and Qwen-VL at layers 16 and 18. Impact of these hyperparameters on performance, we fixed one of the parameters and calculated the average F1 score for LLaVA-v1.5 and Qwen-VL on the POPE COCO random subset. The results are presented in Figure 6, we can observe that when $\\alpha$ is negative—indicating that reverse intervention reduces the model's attention to visual information—the model's performance declines. When $\\alpha$ is too small, the model does not receive sufficient intervention, resulting in suboptimal outcomes. Conversely, when $\\alpha$ is excessively large, the intervention becomes too strong, disrupting the model's foundational capabilities and leading to a decrease in performance. For the hyperparameter $K$ , we find that when $K$ is too small, certain attention heads that encode relevant visual information are not adequately intervened upon, resulting in suboptimal performance. Conversely, when $K$ is too large, attention heads that encode irrelevant information are unnecessarily interfered with, which leads to performance degradation." + ], + "image_footnote": [], + "bbox": [ + 516, + 89, + 903, + 202 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.4. Analysis of Attention Heads for Visual Information Encoding", + "text_level": 1, + "bbox": [ + 511, + 521, + 906, + 555 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Figure 7, we present the classification accuracy of the binary classifiers for Image-Level and Object-Level features across the 1,024 multi-head attention heads in the 32 layers of LLaVA-v1.5. This analysis helps identify which heads encode overall visual information and which heads capture fine-grained visual details. From the figure, we can observe", + "bbox": [ + 511, + 559, + 906, + 650 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f8beef3d8e631cb91400e7143775abb381658953d1dfb35fee4da509b3839fae.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SettingMethodAccuracyF1 Score
RandomRegular80.9779.01
VCD85.5985.33
ICT86.3886.96
ICT-LLaVA-v1.585.1083.27
PopularRegular75.9974.84
VCD81.8382.23
ICT82.6382.22
ICT-LLaVA-v1.581.5080.10
AdversarialRegular75.4674.33
VCD80.0180.75
ICT80.8380.60
ICT-LLaVA-v1.579.7378.68
", + "bbox": [ + 542, + 657, + 877, + 848 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Cross-model generalization of ICT. Performance of Qwen-VL on GQA after applying activation shift from LLaVA-v1.5. ICT-LLaVA-v1.5 refers to the results obtained by applying LLaVA-v1.5's activation shift vectors to intervene in Qwen-VL.", + "bbox": [ + 511, + 858, + 908, + 914 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4215", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/061fcca9a3ee2c52e18822df4fdfafda5ac52b8f646caf5be60c6a3657ab698f.jpg", + "image_caption": [ + "VCD: Yes, there is.", + "ICT: No, there isn't.", + "Question: How many uncut fruits are in the image?" + ], + "image_footnote": [], + "bbox": [ + 187, + 108, + 354, + 205 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0a69dc51e3541fb53d1e1ad5e1efc87d97a10cfbe2cf9560213e69c063dd809a.jpg", + "image_caption": [ + "Question: Is there a horse in the image?", + "POPE" + ], + "image_footnote": [], + "bbox": [ + 555, + 119, + 779, + 205 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/44463fb48a92b5fe147d9e585a798d28a936dfd3835b536e7f985137d1e2c41c.jpg", + "image_caption": [ + "VCD: There are two uncut fruits in the image.", + "ICT: There are four uncut fruits in the image.", + "Figure 5. Case Study and Error Analysis of ICT. The vertical axis in the right section of the figure represents the attention ratio of each question text token to the visual tokens." + ], + "image_footnote": [], + "bbox": [ + 187, + 215, + 333, + 328 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d00f3d20228a6c7d87d26059a11fbdfab60dd6780df7a954d3f4ac14926768cc.jpg", + "image_caption": [ + "LLaVA-Bench" + ], + "image_footnote": [], + "bbox": [ + 555, + 244, + 792, + 323 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b1db38f81147a4e3fbbc8512ae0b974d3abe3653d47b72ee36fedf81b1ae1c8b.jpg", + "image_caption": [ + "Figure 6. Effect of hyperparameters $\\alpha$ (left) and $K$ (right) on the F1 score for LLaVA-v1.5 and Qwen-VL models on the POPE COCO random subset." + ], + "image_footnote": [], + "bbox": [ + 91, + 361, + 282, + 479 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8158a5900d2bc9350a72d94b857f4a48acf733e6a778f887c09e0b05cc429b16.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 361, + 480, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "that the attention heads encoding overall visual information are predominantly located in the earlier layers of the model, such as layers 4 and 10. In contrast, the attention heads that capture fine-grained visual details are more concentrated in the later layers, such as layers 20 and 22.", + "bbox": [ + 89, + 529, + 483, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.5. Case Study and Error Analysis", + "text_level": 1, + "bbox": [ + 89, + 618, + 366, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Figure 5, we present a case study on the POPE and LLaVA-Bench datasets, visualizing the attention ratio of each question text token to the visual tokens. As illustrated, after applying ICT, the model allocates a higher proportion of attention to visual tokens, especially to object tokens relevant to the question (e.g., \"horse\" and \"fruits\"). By prior-", + "bbox": [ + 89, + 641, + 483, + 732 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7dbdda3e10270cab46ddf2f6629e41aba07c1ce1f8ba4279fa920d77be7fcb7d.jpg", + "image_caption": [ + "Figure 7. Heatmap of classification accuracy for Image-Level (left) and Object-Level (right) features of LLaVA-v1.5." + ], + "image_footnote": [], + "bbox": [ + 91, + 756, + 480, + 891 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "itizing visual information, ICT correctly identifies the absence of a horse in the image, whereas VCD erroneously concludes that a horse is present due to insufficient attention to visual cues. However, when asked, \"How many uncut fruits are in the picture?\", VCD incorrectly answered \"two\" due to a lack of focus on visual details. Although ICT correctly identifies that there are a total of four fruits in the image, this question requires not only attention to visual content but also reasoning within the text modality. The model needs to not only recognize the overall count of fruits but also focus on the attribute \"uncut\". Due to its failure to incorporate this information, ICT outputs a wrong answer.", + "bbox": [ + 511, + 363, + 906, + 544 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Furthermore, when the text becomes more complex, such as in \"How many uncut fruits are there in the image?\", the increased linguistic complexity demands greater attention from the model. This, in turn, limits the extent to which attention can be allocated to the visual domain. In contrast, function words like \"is\" and \"there\" are less complex and require minimal processing, allowing the model to shift more focus toward visual tokens. As a result, this redistribution of attention can lead to a more noticeable shift in how the model processes visual and textual information.", + "bbox": [ + 511, + 546, + 908, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion and Limitations", + "text_level": 1, + "bbox": [ + 513, + 722, + 769, + 738 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we mitigate the issue of object hallucinations in LVLMs by proposing Image-Object Cross-Level Trusted Intervention (ICT), a training-free, plug-and-play method that enhances the model's focus on both Image-Level and Object-Level visual information during the forward pass. Our experiments show that ICT significantly reduces hallucinations while also improving general reasoning capabilities. The results confirm that ICT effectively mitigates excessive reliance on language priors, enhancing both accuracy and robustness in diverse visual contexts.", + "bbox": [ + 511, + 750, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4216", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by the National Key Research and Development Program of China (Grant No. 2024YFB3309702); Guangdong Provincial Department of Education Project (Grant No.2024KQNCX028); CAAI-Ant Group Research Fund; Scientific Research Projects for the Higher-educational Institutions (Grant No.2024312096), Education Bureau of Guangzhou Municipality; Guangzhou-HKUST(GZ) Joint Funding Program (Grant No.2025A03J3957), Education Bureau of Guangzhou Municipality.", + "bbox": [ + 89, + 113, + 485, + 255 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 281, + 187, + 297 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Renat Aksitov, Chung-Ching Chang, David Reitter, Siamak Shakeri, and Yunhsuan Sung. Characterizing attribution and fluency tradeoffs for retrieval-augmented large language models, 2023. 3", + "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: A visual language model for few-shot learning. In Advances in Neural Information Processing Systems, pages 23716-23736, 2022. 2", + "[3] Wenbin An, Feng Tian, Sicong Leng, Jiahao Nie, Haonan Lin, QianYing Wang, Guang Dai, Ping Chen, and Shijian Lu. Agla: Mitigating object hallucinations in large vision-language models with assembly of global and local attention. arXiv preprint, 2024. 3", + "[4] Wenbin An, Feng Tian, Sicong Leng, Jiahao Nie, Haonan Lin, QianYing Wang, Guang Dai, Ping Chen, and Shijian Lu. Agla: Mitigating object hallucinations in large vision-language models with assembly of global and local attention. arXiv preprint arXiv:2406.12718, 2024. 1", + "[5] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. 2023. 1, 2, 5", + "[6] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D. Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In Advances in Neural Information Processing Systems, pages 1877-1901, 2020. 2", + "[7] Beitao Chen, Xinyu Lyu, Lianli Gao, Jingkuan Song, and Heng Tao Shen. Alleviating hallucinations in large vision-language models through hallucination-induced optimization, 2024. 3", + "[8] Haozhe Chen, Carl Vondrick, and Chengzhi Mao. Selfie: Self-interpretation of large language model embeddings, 2024. 2", + "[9] Zhaorun Chen, Zhuokai Zhao, Hongyin Luo, Huaxiu Yao, Bo Li, and Jiawei Zhou. Halc: Object hallucination reduction via adaptive focal-contrast decoding. In *Forty-first International Conference on Machine Learning*. 1" + ], + "bbox": [ + 107, + 306, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[10] Zhaorun Chen, Zhuokai Zhao, Hongyin Luo, Huaxiu Yao, Bo Li, and Jiawei Zhou. Halc: Object hallucination reduction via adaptive focal-contrast decoding, 2024. http://arxiv.org/abs/2403.00425.2,3", + "[11] I.-Chun Chern, Steffi Chern, Shiqi Chen, Weizhe Yuan, Kehua Feng, Chunting Zhou, Junxian He, Graham Neubig, and Pengfei Liu. Factool: Factuality detection in generative ai - a tool augmented framework for multi-task and multi-domain scenarios, 2023. http://arxiv.org/abs/2307.13528.3", + "[12] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with $90\\%$ * chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2(3):6, 2023. 1, 2", + "[13] Jaemin Cho, Seunghyun Yoon, Ajinkya Kale, Franck Dernoncourt, Trung Bui, and Mohit Bansal. Fine-grained image captioning with clip reward. In Findings of the Association for Computational Linguistics: NAACL 2022, pages 517-527, 2022. 1", + "[14] Yung-Sung Chuang, Yujia Xie, Hongyin Luo, Yoon Kim, James Glass, and Pengcheng He. Dola: Decoding by contrasting layers improves factuality in large language models, 2024. http://arxiv.org/abs/2309.03883.2, 3", + "[15] Corinna Cortes. Support-vector networks. Machine Learning, 1995. 5", + "[16] Li D., Rawat A.S., Zaheer M., Wang X., Lukasik M., Veit A., Yu F., and Kumar S. Large language models with controllable working memory. arXiv preprint arXiv:2211.05110, 2022. 3", + "[17] Hanxing Ding, Liang Pang, Zihao Wei, Huawei Shen, and Xueqi Cheng. Retrieve only when it needs: Adaptive retrieval augmentation for hallucination mitigation in large language models, 2024. http://arxiv.org/abs/2402.10612.3", + "[18] Sebastian Farquhar, Jannik Kossen, Lorenz Kuhn, and Yarin Gal. Detecting hallucinations in large language models using semantic entropy. Nature, 630(8017):625-630, 2024. 3", + "[19] Javier Ferrando, Gabriele Sarti, Arianna Bisazza, and Marta R. Costa-jussà. A primer on the inner workings of transformer-based language models, 2024. 2", + "[20] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023.5", + "[21] Suriya Gunasekar, Yi Zhang, Jyoti Aneja, Cao César Teodoro Mendes, Allie Del Giorno, Sivakanth Gopi, Mojan Javaheripi, Piero Kauffmann, Gustavo de Rosa, Olli Saarikivi, Adil Salim, Shital Shah, Harkirat Singh Behl, Xin Wang, Sébastien Bubeck, Ronen Eldan, Adam Tauman Kalai, Yin Tat Lee, and Yuanzhi Li. Textbooks are all you need, 2023. 3", + "[22] Anisha Gunjal, Jihan Yin, and Erhan Bas. Detecting and preventing hallucinations in large vision language models." + ], + "bbox": [ + 522, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4217", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In Proceedings of the AAAI Conference on Artificial Intelligence, pages 18135-18143, 2024. 3", + "[23] Anisha Gunjal, Jihan Yin, and Erhan Bas. Detecting and preventing hallucinations in large vision language models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 18135-18143, 2024. 2", + "[24] Vipul Gupta, Zhuowan Li, Adam Kortylewski, Chenyu Zhang, Yingwei Li, and Alan Yuille. Swapmix: Diagnosing and regularizing the over-reliance on visual context in visual question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5078-5088, 2022. 1", + "[25] Yudong Han, Liqiang Nie, Jianhua Yin, Jianlong Wu, and Yan Yan. Visual perturbation-aware collaborative learning for overcoming the language prior problem. arXiv preprint arXiv:2207.11850, 2022. 1", + "[26] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3", + "[27] Wenyi Hong, Weihan Wang, Ming Ding, Wenmeng Yu, Qingsong Lv, Yan Wang, Yean Cheng, Shiyu Huang, Junhui Ji, Zhao Xue, et al. Cogvlm2: Visual language models for image and video understanding. arXiv preprint arXiv:2408.16500, 2024. 1", + "[28] Haichuan Hu, Yuhan Sun, and Quanjun Zhang. Lrp4rag: Detecting hallucinations in retrieval-augmented generation via layer-wise relevance propagation, 2024. http:// arxiv.org/abs/2408.15533.3", + "[29] Yingdong Hu, Fanqi Lin, Tong Zhang, Li Yi, and Yang Gao. Look before you leap: Unveiling the power of gpt-4v in robotic vision-language planning. arXiv preprint arXiv:2311.17842, 2023. 1", + "[30] Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, and Ting Liu. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions, 2023. 3", + "[31] Qidong Huang, Xiaoyi Dong, Pan Zhang, Bin Wang, Conghui He, Jiaqi Wang, Dahua Lin, Weiming Zhang, and Nenghai Yu. Opera: Alleviating hallucination in multimodal large language models via over-trust penalty and retrospection-allocation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13418-13427, 2024. 5", + "[32] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 5", + "[33] Fushuo Huo, Wenchao Xu, Zhong Zhang, Haozhao Wang, Zhicheng Chen, and Peilin Zhao. Self-introspective decoding: Alleviating hallucinations for large vision-language models, 2024. 3", + "[34] Jitesh Jain, Jianwei Yang, and Humphrey Shi. Vcoder: Versatile vision encoders for multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27992-28002, 2024. 2" + ], + "bbox": [ + 99, + 92, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[35] Juseong Jin and Chang Wook Jeong. Surgical-llava: Toward surgical scenario understanding via large language and vision models. arXiv preprint arXiv:2410.09750, 2024. 1", + "[36] Junho Kim, Hyunjun Kim, Yeonju Kim, and Yong Man Ro. Code: Contrasting self-generated description to combat hallucination in large multi-modal models, 2024. 2, 3", + "[37] Lorenz Kuhn, Yarin Gal, and Sebastian Farquhar. Semantic uncertainty: Linguistic invariances for uncertainty estimation in natural language generation. arXiv, 2023. 3", + "[38] Junlin Lee, Yequan Wang, Jing Li, and Min Zhang. Multimodal reasoning with multimodal knowledge graph. arXiv preprint arXiv:2406.02030, 2024. 2", + "[39] Katherine Lee, Daphne Ippolito, Andrew Nystrom, Chiyuan Zhang, Douglas Eck, Chris Callison-Burch, and Nicholas Carlini. Deduplicating training data makes language models better. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8424-8445, 2022. 3", + "[40] Kang-il Lee, Minbeom Kim, Seunghyun Yoon, Minsung Kim, Dongryeol Lee, Hyukhun Koh, and Kyomin Jung. Vlind-bench: Measuring language priors in large vision-language models. arXiv preprint arXiv:2406.08702, 2024. 1", + "[41] Nayeon Lee, Wei Ping, Peng Xu, Mostofa Patwary, Pascale N Fung, Mohammad Shoeybi, and Bryan Catanzaro. Factuality enhanced language models for open-ended text generation. In Advances in Neural Information Processing Systems, pages 34586-34599, 2022. 3", + "[42] Seongyun Lee, Sue Hyun Park, Yongrae Jo, and Minjoon Seo. Volcano: Mitigating multimodal hallucination through self-feedback guided revision. arXiv, 2024. 3", + "[43] Sicong Leng, Hang Zhang, Guanzheng Chen, Xin Li, Shijian Lu, Chunyan Miao, and Lidong Bing. Mitigating object hallucinations in large vision-language models through visual contrastive decoding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13872-13882, 2024. 1, 2, 5", + "[44] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In Proceedings of the International Conference on Machine Learning, pages 19730-19742. PMLR, 2023. 2", + "[45] Kenneth Li, Oam Patel, Fernanda Viégas, Hanspeter Pfister, and Martin Wattenberg. Inference-time intervention: Eliciting truthful answers from a language model. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[46] Lei Li, Zhihui Xie, Mukai Li, Shunian Chen, Peiyi Wang, Liang Chen, Yazheng Yang, Benyou Wang, Lingpeng Kong, and Qi Liu. Vlfeedback: A large-scale ai feedback dataset for large vision-language models alignment. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 6227-6246, 2024. 2", + "[47] Xiang Lisa Li, Ari Holtzman, Daniel Fried, Percy Liang, Jason Eisner, Tatsunori Hashimoto, Luke Zettlemoyer, and Mike Lewis. Contrastive decoding: Open-ended text generation as optimization. arXiv preprint, 2023. 3" + ], + "bbox": [ + 522, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4218", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[48] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. In The 2023 Conference on Empirical Methods in Natural Language Processing. 5", + "[49] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5", + "[50] Bingbin Liu, Jordan T. Ash, Surbhi Goel, Akshay Krishnamurthy, and Cyril Zhang. Exposing attention glitches with flip-flop language modeling, 2023. 3", + "[51] Fuxiao Liu, Kevin Lin, Linjie Li, Jianfeng Wang, Yaser Ya-coob, and Lijuan Wang. Mitigating hallucination in large multi-modal models via robust instruction tuning. In The Twelfth International Conference on Learning Representations, 2023. 3", + "[52] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 1, 2, 5", + "[53] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[54] Hanchao Liu, Wenyuan Xue, Yifei Chen, Dapeng Chen, Xiutian Zhao, Ke Wang, Liping Hou, Rongjun Li, and Wei Peng. A survey on hallucination in large vision-language models. arXiv preprint, 2024. 2", + "[55] Hanchao Liu, Wenyuan Xue, Yifei Chen, Dapeng Chen, Xiutian Zhao, Ke Wang, Liping Hou, Rongjun Li, and Wei Peng. A survey on hallucination in large vision-language models. arXiv preprint arXiv:2402.00253, 2024. 1", + "[56] Jiajun Liu, Yibing Wang, Hanghang Ma, Xiaoping Wu, Xiaqi Ma, Xiaoming Wei, Jianbin Jiao, Enhua Wu, and Jie Hu. Kangaroo: A powerful video-language model supporting long-context video input. arXiv preprint arXiv:2408.15542, 2024. 1", + "[57] Jiazhen Liu, Yuhan Fu, Ruobing Xie, Runquan Xie, Xingwu Sun, Fengzong Lian, Zhanhui Kang, and Xirong Li. PhD: A chatgpt-prompted visual hallucination evaluation dataset, 2025. 2", + "[58] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 4", + "[59] Shi Liu, Kecheng Zheng, and Wei Chen. Paying more attention to image: A training-free method for alleviating hallucination in lvlms, 2024. 3", + "[60] Alexander Nikitin, Jannik Kossen, Yarin Gal, and Pekka Marttinen. Kernel language entropy: Fine-grained uncertainty quantification for llms from semantic similarities. arXiv, 2024. 3", + "[61] OpenAI. GPT-4. https://openai.com/gpt-4, 2023.2", + "[62] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, et al." + ], + "bbox": [ + 99, + 90, + 485, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Training language models to follow instructions with human feedback. In Advances in Neural Information Processing Systems, pages 27730-27744, 2022. 3", + "[63] Amandalynne Paullada, Inioluwa Deborah Raji, Emily M. Bender, Emily Denton, and Alex Hanna. Data and its (dis)contents: A survey of dataset development and use in machine learning research. *Patterns*, 2(11):100336, 2021. 3", + "[64] Phuc Phan, Hieu Tran, and Long Phan. Distillation contrastive decoding: Improving llms reasoning with contrastive decoding and distillation, 2024. 2, 3", + "[65] Venkit P.N., Gautam S., Panchanadikar R., Huang T.H., and Wilson S. Nationality bias in text generation. arXiv preprint arXiv:2302.02463, 2023. 3", + "[66] Xiaoye Qu, Qiyuan Chen, Wei Wei, Jishuo Sun, and Jianfeng Dong. Alleviating hallucination in large vision-language models with active retrieval augmentation, 2024. http://arxiv.org/abs/2408.00555.3", + "[67] Xiaoye Qu, Jiashuo Sun, Wei Wei, and Yu Cheng. Look, compare, decide: Alleviating hallucination in large vision-language models via multi-view multi-path reasoning, 2024. http://arxiv.org/abs/2408.17150.3", + "[68] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1", + "[69] Vikas Raunak, Arul Menezes, and Marcin Junczys-Dowmunt. The curious case of hallucinations in neural machine translation. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1172-1183, 2021. 3", + "[70] Kim S., Cho B., Bae S., Ahn S., and Yun S.Y. Vacode: Visual augmented contrastive decoding. arXiv preprint arXiv:2408.05337, 2024. 2, 3", + "[71] Lin S., Hilton J., and Evans O. Truthfulqa: Measuring how models mimic human falsehoods. arXiv preprint arXiv:2109.07958, 2021. 3", + "[72] Yin S., Fu C., Zhao S., Li K., Sun X., Xu T., and Chen E. A survey on multimodal large language models. arXiv preprint arXiv:2306.13549, 2023. 3", + "[73] Pranab Sahoo, Prakash Meharia, Akash Ghosh, Sriparna Saha, Vinija Jain, and Aman Chadha. A comprehensive survey of hallucination in large language, image, video and audio foundation models. Findings of the Association for Computational Linguistics: EMNLP 2024, pages 11709-11724, 2024. 1", + "[74] Dustin Schwenk, Apoorv Khandelwal, Christopher Clark, Kenneth Marino, and Roozbeh Mottaghi. A-okvqa: A benchmark for visual question answering using world knowledge. In European conference on computer vision, pages 146–162. Springer, 2022. 5", + "[75] Lei Shen, Haolan Zhan, Xin Shen, Hongshen Chen, Xiaofang Zhao, and Xiaodan Zhu. Identifying untrustworthy samples: Data filtering for open-domain dialogues with" + ], + "bbox": [ + 524, + 92, + 906, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "4219", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "bayesian optimization. In Proceedings of the 30th ACM International Conference on Information & Knowledge Management, pages 1598-1608, 2021. 3", + "[76] Weijia Shi, Xiaochuang Han, Mike Lewis, Yulia Tsvetkov, Luke Zettlemoyer, and Scott Wen tau Yih. Trusting your evidence: Hallucinate less with context-aware decoding, 2023. 3", + "[77] Weijia Shi, Sewon Min, Maria Lomeli, Chunting Zhou, Margaret Li, Rich James, Xi Victoria Lin, Noah A. Smith, Luke Zettlemoyer, Scott Yih, and Mike Lewis. In-context pretraining: Language modeling beyond document boundaries, 2023. 3", + "[78] GLM T., Zeng A., Xu B., Wang B., Zhang C., Yin D., Rojas D., Feng G., Zhao H., Lai H., and Yu H. Chatglm: A family of large language models from glm-130b to glm-4 all tools. arXiv preprint arXiv:2406.12793, 2024. 2", + "[79] Gemma Team, Thomas Mesnard, Cassidy Hardin, Robert Dadashi, Surya Bhupatiraju, Shreya Pathak, Laurent Sifre, Morgane Rivière, Mihir Sanjay Kale, Juliette Love, Pouya Tafti, Léonard Hussenot, Pier Giuseppe Sessa, Aakanksha Chowdhery, Adam Roberts, Aditya Barua, Alex Botev, Alex Castro-Ros, Ambrose Slone, Amélie Héliou, Andrea Tacchetti, Anna Bulanova, Antonia Paterson, Beth Tsai, Bobak Shahriari, Charline Le Lan, Christopher A. Choquette-Choo, Clément Crepy, Daniel Cer, Daphne Ippolito, David Reid, Elena Buchatskaya, Eric Ni, Eric Noland, Geng Yan, George Tucker, George-Christian Muraru, Grigory Rozhdestvenskiy, Henryk Michalewski, Ian Tenney, Ivan Grishchenko, Jacob Austin, James Keeling, Jane Labanowski, Jean-Baptiste Lespiau, Jeff Stanway, Jenny Brennan, Jeremy Chen, Johan Ferret, Justin Chiu, Justin Mao-Jones, Katherine Lee, Kathy Yu, Katie Millican, Lars Lowe Sjoesund, Lisa Lee, Lucas Dixon, Rachel Reid, Maciej Mikula, Mateo Wirth, Michael Sharman, Nikolai Chinaev, Nithum Thain, Olivier Bachem, Oscar Chang, Oscar Wahltinez, Paige Bailey, Paul Michel, Petko Yotov, Rahma Chaabouni, Ramona Comanescu, Reena Jana, Rohan Anil, Ross McIlroy, Ruibo Liu, Ryan Mullins, Samuel L Smith, Sebastian Borgeaud, Sertan Girgin, Sholto Douglas, Shree Pandya, Siamak Shakeri, Soham De, Ted Klimenko, Tom Hennigan, Vlad Feinberg, Wojciech Stokowiec, Yu hui Chen, Zafarali Ahmed, Zhitao Gong, Tris Warkentin, Ludovic Peran, Minh Giang, Clément Farabet, Oriol Vinyals, Jeff Dean, Koray Kavukcuoglu, Demis Hassabis, Zoubin Ghahramani, Douglas Eck, Joelle Barral, Fernando Pereira, Eli Collins, Armand Joulin, Noah Fiedel, Evan Senter, Alek Andreev, and Kathleen Kenealy. Gemma: Open models based on gemini research and technology, 2024. 2", + "[80] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint, 2023. 2", + "[81] Elena Voita, David Talbot, Fedor Moiseev, Rico Sennrich, and Ivan Titov. Analyzing multi-head self-attention: Specialized heads do the heavy lifting, the rest can be pruned. In Proceedings of the 57th Annual Meeting of the Asso" + ], + "bbox": [ + 99, + 90, + 482, + 898 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "citation for Computational Linguistics, pages 5797-5808, 2019. 2", + "[82] Zhong W., Feng X., Zhao L., Li Q., Huang L., Gu Y., Ma W., Xu Y., and Qin B. Investigating and mitigating the multimodal hallucination snowballing in large vision-language models. arXiv preprint arXiv:2407.00569, 2024. 2, 3", + "[83] Guankun Wang, Long Bai, Wan Jun Nah, Jie Wang, Zhaoxi Zhang, Zhen Chen, Jinlin Wu, Mobarakol Islam, Hongbin Liu, and Hongliang Ren. Surgical-lvIm: Learning to adapt large vision-language model for grounded visual question answering in robotic surgery. arXiv preprint arXiv:2405.10948, 2024. 1", + "[84] P Wang, S Bai, S Tan, S Wang, Z Fan, J Bai, K Chen, X Liu, J Wang, W Ge, and Y Fan. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 2", + "[85] Xintong Wang, Jingheng Pan, Liang Ding, and Chris Biemann. Mitigating hallucinations in large vision-language models with instruction contrastive decoding, 2024. 3", + "[86] Sangmin Woo, Jaehyuk Jang, Donguk Kim, Yubin Choi, and Changick Kim. Ritual: Random image transformations as a universal anti-hallucination lever in lvlms, 2024. http://arxiv.org/abs/2405.17821.3", + "[87] Sangmin Woo, Donguk Kim, Jaehyuk Jang, Yubin Choi, and Changick Kim. Don't miss the forest for the trees: Attentional vision calibration for large vision language models, 2024. 2, 3", + "[88] Junfei Wu, Qiang Liu, Ding Wang, Jinghao Zhang, Shu Wu, Liang Wang, and Tieniu Tan. Logical closed loop: Uncovering object hallucinations in large vision-language models, 2024. http://arxiv.org/abs/2402.11622.3", + "[89] Kai Wu, Boyuan Jiang, Zhengkai Jiang, Qingdong He, Donghao Luo, Shengzhi Wang, Qingwen Liu, and Chengjie Wang. Noiseboost: Alleviating hallucination with noise perturbation for multimodal large language models, 2024. http://arxiv.org/abs/2405.20081.3", + "[90] Xin Xiao, Bohong Wu, Jiacong Wang, Chunyuan Li, Xun Zhou, and Haoyuan Guo. Seeing the image: Prioritizing visual correlation by contrastive alignment, 2024. 3", + "[91] Xiaohan Xu, Chongyang Tao, Tao Shen, Can Xu, Hongbo Xu, Guodong Long, and Jian guang Lou. Re-reading improves reasoning in large language models, 2024. http://arxiv.org/abs/2309.06275.3", + "[92] L Xue, M Shu, A Awadalla, J Wang, A Yan, S Purushwalkam, H Zhou, V Prabhu, Y Dai, M S Ryoo, and S Kendre. xgen-mm (blip-3): A family of open large multimodal models. arXiv preprint arXiv:2408.08872, 2024. 2", + "[93] Onoe Y., Zhang M.J., Choi E., and Durrett G. Entity cloze by date: What lms know about unseen entities. arXiv preprint arXiv:2205.02832, 2022. 3", + "[94] Park Y., Lee D., Choe J., and Chang B. Convis: Contrastive decoding with hallucination visualization for mitigating hallucinations in multimodal large language models. arXiv preprint arXiv:2408.13906, 2024. 3", + "[95] Zhenjie Yang, Xiaosong Jia, Hongyang Li, and Junchi Yan. Llm4drive: A survey of large language models for au" + ], + "bbox": [ + 522, + 92, + 903, + 900 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "4220", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tonomous driving. In NeurIPS 2024 Workshop on OpenWorld Agents, 2023. 1", + "[96] Shukang Yin, Chaoyou Fu, Sirui Zhao, Tong Xu, Hao Wang, Dianbo Sui, Yunhang Shen, Ke Li, Xing Sun, and Enhong Chen. Woodpecker: Hallucination correction for multimodal large language models. arXiv preprint, 2023. 3", + "[97] Qifan Yu, Juncheng Li, Longhui Wei, Liang Pang, Wentao Ye, Bosheng Qin, Siliang Tang, Qi Tian, and Yueting Zhuang. Hallucidctor: Mitigating hallucinatory toxicity in visual instruction data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12944–12953, 2024. 3", + "[98] Tianyu Yu, Yuan Yao, Haoye Zhang, Taiwen He, Yifeng Han, Ganqu Cui, Jinyi Hu, et al. Rlhf-v: Towards trustworthy mllms via behavior alignment from fine-grained correctional human feedback. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13807-13816, 2024. 3", + "[99] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, Cong Wei, Botao Yu, Ruibin Yuan, Renliang Sun, Ming Yin, Boyuan Zheng, Zhenzhu Yang, Yibo Liu, Wenhao Huang, Huan Sun, Yu Su, and Wenhu Chen. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of CVPR, 2024. 2", + "[100] Zihao Yue, Liang Zhang, and Qin Jin. Less is more: Mitigating multimodal hallucination from an eos decision perspective, 2024. http://arxiv.org/abs/2402.14545.3", + "[101] Bai Z., Wang P., Xiao T., He T., Han Z., Zhang Z., and Shou MZ. Hallucination of multimodal large language models: A survey. arXiv preprint arXiv:2404.18930, 2024. 3", + "[102] Han Z., Bai Z., Mei H., Xu Q., Zhang C., and Shou M.Z. Skip: A simple method to reduce hallucination in large vision-language models. arXiv preprint arXiv:2402.01345, 2024. 3", + "[103] Shaolei Zhang, Tian Yu, and Yang Feng. Truthx: Alleviating hallucinations by editing large language models in truthful space, 2024. 3", + "[104] Yi-Fan Zhang, Weichen Yu, Qingsong Wen, Xue Wang, Zhang Zhang, Liang Wang, Rong Jin, and Tieniu Tan. Debiasing multimodal large language models. arXiv preprint, 2024. 3", + "[105] L Zhao, Y Deng, W Zhang, and Q Gu. Mitigating object hallucination in large vision-language models via classifier-free guidance. 2024. 3", + "[106] Zhiyuan Zhao, Bin Wang, Linke Ouyang, Xiaoyi Dong, Jiaqi Wang, and Conghui He. Beyond hallucinations: Enhancing lvlms through hallucination-aware direct preference optimization. arXiv preprint arXiv:2311.16839, 2023. 2", + "[107] Zheng Zhao, Emilio Monti, Jens Lehmann, and Haytham Assem. Enhancing contextual understanding in large language models through contrastive decoding, 2024. 3", + "[108] Lanyun Zhu, Deyi Ji, Tianrun Chen, Peng Xu, Jieping Ye, and Jun Liu. Ibid: Alleviating hallucinations in large" + ], + "bbox": [ + 93, + 92, + 482, + 900 + ], + "page_idx": 12 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "vision-language models via image-biased decoding. arXiv preprint, 2024. 3", + "[109] Wenhao Zhu, Sizhe Liu, Shujian Huang, Shuaijie She, Chris Wendler, and Jiajun Chen. Multilingual contrastive decoding via language-agnostic layers skipping, 2024. http://arxiv.org/abs/2407.10795.2,3" + ], + "bbox": [ + 517, + 92, + 903, + 176 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "4221", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 12 + } +] \ No newline at end of file diff --git a/2025/ICT_ Image-Object Cross-Level Trusted Intervention for Mitigating Object Hallucination in Large Vision-Language Models/a90828a4-8b1b-4072-b5af-bed317d47267_model.json b/2025/ICT_ Image-Object Cross-Level Trusted Intervention for Mitigating Object Hallucination in Large Vision-Language Models/a90828a4-8b1b-4072-b5af-bed317d47267_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ca3ff56af06546a103e7952b4882d0d490ef3aac --- /dev/null +++ b/2025/ICT_ Image-Object Cross-Level Trusted Intervention for Mitigating Object Hallucination in Large Vision-Language Models/a90828a4-8b1b-4072-b5af-bed317d47267_model.json @@ -0,0 +1,2877 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.12, + 0.13, + 0.88, + 0.177 + ], + "angle": 0, + "content": "ICT: Image-Object Cross-Level Trusted Intervention for Mitigating Object Hallucination in Large Vision-Language Models" + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.203, + 0.761, + 0.239 + ], + "angle": 0, + "content": "Junzhe Chen\\(^{1,2*}\\), Tianshu Zhang\\(^{1*}\\), Shiyu Huang\\(^{3}\\), Yuwei Niu\\(^{4}\\), Linfeng Zhang\\(^{5}\\), Lijie Wen\\(^{1\\dagger}\\), Xuming Hu\\(^{2\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.135, + 0.24, + 0.862, + 0.275 + ], + "angle": 0, + "content": "\\(^{1}\\)Tsinghua University, \\(^{2}\\)The Hong Kong University of Science and Technology (Guangzhou), \\(^{3}\\)Zhipu AI, \\(^{4}\\)Chongqing University, \\(^{5}\\)Shanghai Jiao Tong University." + }, + { + "type": "text", + "bbox": [ + 0.315, + 0.276, + 0.678, + 0.292 + ], + "angle": 0, + "content": "chenjz24@mails.tsinghua.edu.cn," + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.294, + 0.752, + 0.31 + ], + "angle": 0, + "content": "wenlj@tsinghua.edu.cn, xuminghu97@gmail.com," + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.345, + 0.327, + 0.361 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.376, + 0.485, + 0.725 + ], + "angle": 0, + "content": "Despite the recent breakthroughs achieved by Large Vision Language Models (LVLMs) in understanding and responding to complex visual-textual contexts, their inherent hallucination tendencies limit their practical application in real-world scenarios that demand high levels of precision. Existing methods typically either fine-tune the LVLMs using additional data, which incurs extra costs in manual annotation and computational resources or perform comparisons at the decoding stage, which may eliminate useful language priors for reasoning while introducing inference time overhead. Therefore, we propose ICT, a lightweight, training-free method that calculates an intervention direction to shift the model's focus towards different levels of visual information, enhancing its attention to high-level and fine-grained visual details. During the forward pass stage, the intervention is applied to the attention heads that encode the overall image information and the fine-grained object details, effectively mitigating the phenomenon of overly language priors, and thereby alleviating hallucinations. Extensive experiments demonstrate that ICT achieves strong performance with a small amount of data and generalizes well across different datasets and models. Our codes are publicly available at:https://github.com/THU-BPM/ICT/." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.751, + 0.222, + 0.767 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.776, + 0.483, + 0.868 + ], + "angle": 0, + "content": "The recent rapid advancements in Large Vision Language Models (LVLMs) have demonstrated significant potential to tackle complex real-world tasks [5, 27, 29, 52, 56]. However, LVLMs frequently generate text outputs that are inconsistent with the visual input, such as incorrectly determining the presence of objects in an image or inaccurately" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.345, + 0.905, + 0.54 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.545, + 0.907, + 0.642 + ], + "angle": 0, + "content": "Figure 1. Comparison between Contrastive Decoding (top) and our proposed ICT (bottom). In the top example, Contrastive Decoding indiscriminately removes both beneficial and detrimental language priors, leading to hallucinations. In contrast, our approach enhances the model's attention to visual details while preserving useful language priors, allowing it to correctly identify and describe objects in the image." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.907, + 0.885 + ], + "angle": 0, + "content": "describing the attributes of those objects [55, 73]. This issue is commonly referred to as the object hallucination phenomenon [55], which significantly limits the applicability of LVLMs in high-stakes scenarios that demand precise accuracy, such as autonomous driving [95] and medical surgery [35, 83]. Previous research suggested that the primary reasons for hallucinations in LVLMs are as follows: 1) Large language models (LLMs), such as Vicuna [12], possess significantly stronger parameters and capabilities compared to visual encoders, such as CLIP [68]. This results in an excessively strong language prior, causing the model to overly rely on language cues while neglecting visual input [24, 25, 40, 43, 55]. 2) Current visual decoders often struggle to accurately capture fine-grained visual semantics, resulting in errors in detailed object attributes in an image, such as color and quantity [4, 9, 13, 55]." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Based on the two issues outlined above, previous ef" + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.875, + 0.233, + 0.888 + ], + "angle": 0, + "content": "* Equal Contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.888, + 0.251, + 0.9 + ], + "angle": 0, + "content": "† Corresponding authors." + }, + { + "type": "list", + "bbox": [ + 0.115, + 0.875, + 0.251, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4209" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.486, + 0.576 + ], + "angle": 0, + "content": "forts to mitigate the hallucination phenomenon in LVLMs can be categorized into three main approaches: 1) Finetuning with additional data: This approach involves introducing high-quality annotated data to better align the model's behavior with human interpretations, effectively teaching the model to focus more on visual information [23, 46, 52, 106]. However, this method not only requires costly human annotation but also involves updating model parameters, which demands substantial computational resources, thereby limiting its scalability. 2) Perceptual enhancement: This method incorporates additional information, such as depth maps and segmentation maps [34, 38, 106], as auxiliary inputs to assist the visual encoder in capturing more detailed visual features, thereby reducing hallucinations. However, it often requires manual selection of auxiliary features, which limits its generalizability across different tasks. 3) Contrastive decoding: This method alleviates hallucinations without requiring additional training. It induces hallucinations by applying transformations such as blurring, rotation, or cropping to the original visual input. During the decoding stage, tokens associated with these induced hallucinations are penalized, thus mitigating the influence of language priors [10, 14, 36, 43, 64, 70, 82, 87, 109]. However, methods such as VCD often indiscriminately eliminate all language priors, including those that may be beneficial. As shown in Figure 1, the original model recognizes Curry holding a golf club, while utilizing the language prior that identifies Curry as a basketball player. This language prior is valuable for achieving an accurate interpretation. However, through contrastive decoding, this useful prior is also removed, which can inadvertently result in hallucinations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.487, + 0.902 + ], + "angle": 0, + "content": "To address the challenges of mitigating hallucinations in LVLMs, we propose Image-Object Cross-Level Trusted Intervention (ICT), a training-free, plug-and-play method applicable during the forward pass. Unlike contrastive decoding, our approach does not eliminate language priors to reduce the model's over-reliance on text modality. Instead, it intervenes during the forward pass to enhance the model's focus on both comprehensive visual information and fine-grained object details. We explored the activation patterns of attention heads when the model produces correct versus hallucinated responses. This analysis allowed us to identify activation value deviations that can shift the model from being \"untrustworthy\" to \"trustworthy\". According to prior research [8, 19, 81], which has shown that different heads in the multi-head attention mechanism encode information at varying levels of granularity, we train binary classifiers for each head to determine which heads encode overall visual information and which capture detailed visual features. During the forward pass, we adjust the activation values of these heads based on their identified levels, thereby enhancing the model's attention to the relevant visual features and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.212 + ], + "angle": 0, + "content": "reducing the likelihood of hallucinations. As illustrated in Figure 1, after applying ICT, the model is able to focus more on the details within the image, such as identifying the man as Curry, while simultaneously utilizing beneficial language priors (e.g., Curry is a basketball player) to infer and arrive at the correct answer. Since the intervention shift vectors are pre-computed, ICT does not introduce additional delays during the forward pass." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.214, + 0.909, + 0.304 + ], + "angle": 0, + "content": "Our experiments demonstrate that, for both LLaVAv1.5 [52] and Qwen-VL [5], applying ICT results in an average improvement of \\(6.27\\%\\) on POPE benchmark and 67.37 points on MME benchmark. Furthermore, ICT exhibits cross-dataset generalization and model-agnostic generalizability. Our contributions can be summarized as:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.305, + 0.907, + 0.396 + ], + "angle": 0, + "content": "- We propose Image-Object Cross-Level Trusted Intervention (ICT), a novel, training-free, plug-and-play method that effectively reduces hallucinations in LVLMs by enhancing focus on both overall visual information and fine-grained object details during the forward pass, without eliminating beneficial language priors." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.397, + 0.907, + 0.473 + ], + "angle": 0, + "content": "- Unlike existing contrastive decoding approaches, we introduce an intervention mechanism that operates during the forward pass rather than the decoding stage, allowing ICT to be orthogonal and complementary to existing decoding strategies while not introducing any additional latency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.474, + 0.906, + 0.534 + ], + "angle": 0, + "content": "- Extensive experiments on LLaVA-v1.5 and Qwen-VL demonstrate that ICT significantly improves performance on both the POPE and MME benchmarks while maintaining cross-dataset and model-agnostic generalizability." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.55, + 0.658, + 0.566 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.576, + 0.793, + 0.592 + ], + "angle": 0, + "content": "2.1. Large Vision-Language Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Following the success of LLM [6, 12, 61, 78, 80], researchers have been exploring the multimodal field. Taking advantage of the remarkable capabilities of LLMs, Large Vision-Language Models incorporate visual encoders and feature projectors into powerful LLMs, enabling them to understand and generate content based on visual and text input. Most of these models undergo two training steps, namely the pre-training stage and the fine-tuning stage. The main purpose of the pre-training stage is to align textual and visual features, while the goal of the fine-tuning stage is to bridge further the modality gap between vision and language, as well as lifting their instruction following capabilities and performance on specific downstream tasks. Early attempts, such as FLamingo [2], Gemini [79], and BLIP-2 [44], have already shown promising results. Recent works like LLaVA-v1.5 [54], Qwen2-VL [84], and xgenmm [92] have advanced this field, dramatically improving the abilities of these models. To enhance the consistency between vision representation and language representation, many efforts have been made, including utilizing higher res" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "4210" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.184 + ], + "angle": 0, + "content": "olution vision encoders, transitioning to larger and more capable LLMs, and adapting reinforcement learning techniques such as RLHF [62, 98], among others. Despite the remarkable progress that has been made, LVLMs still suffer from severe hallucinations, which restricts their potential for large-scale applications in real-world scenarios." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.196, + 0.409, + 0.213 + ], + "angle": 0, + "content": "2.2. Mitigating Hallucinations in LVLMs" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.22, + 0.485, + 0.569 + ], + "angle": 0, + "content": "Many efforts have been made to understand the causes of hallucination [1, 16, 30, 39, 50, 50, 53, 63, 65, 71, 72, 76, 93, 101]. Existing approaches to mitigate hallucinations can be roughly divided into two categories based on the stage they take place. The first category focuses on the training phase. Most of the work falling into this category introduces additional or curated datasets [21, 22, 51, 97], these works usually design a dataset specifically for hallucination-related tasks or improve data cleaning methods [69, 75], or introduce novel training objectives [7, 41, 77]. While effective, these methods typically require extensive training, which is both time-consuming and resource-intensive. The second category focuses on the inferencing phase, usually involves CD-based [47] novel decoding strategies [10, 14, 36, 59, 64, 70, 82, 87, 94, 107, 109]. Another approach is to detect potential hallucinations while generating and correct them [18, 37, 60], among others [45, 86, 100, 103]. In addition, some researchers tackle hallucinations by manipulating attention weight assigned to the image [33, 89, 90, 104, 108], or the prompt-related part of the image [3]. Other works include prompt-based methods [42, 67, 85, 88, 91, 102], utilizing external tools [11, 96, 105] or external knowledge [17, 28, 66] and so on." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.586, + 0.266, + 0.602 + ], + "angle": 0, + "content": "3. Task Formulation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.613, + 0.484, + 0.75 + ], + "angle": 0, + "content": "Given an LVLM parameterized by \\(\\theta\\), the model processes a textual input \\(\\pmb{X} = \\{x_{1}, x_{2}, \\ldots, x_{N_{x}}\\}\\) and a visual input \\(\\pmb{V} = \\{v_{1}, v_{2}, \\ldots, v_{N_{v}}\\}\\), where \\(N_{x}\\) and \\(N_{v}\\) denote the sequence lengths of the text and visual inputs, respectively. The model then concatenates the textual and visual sequences to form a unified input \\(\\pmb{H} = \\text{concat}(\\pmb{V}, \\pmb{X})\\), which is subsequently passed through \\(L\\) layers of a transformer architecture. At each layer, the concatenated input undergoes multi-head attention, which is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.764, + 0.483, + 0.806 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {H} ^ {(l + 1)} = \\boldsymbol {H} ^ {(l)} + \\sum_ {n = 1} ^ {N} A t t n _ {n} ^ {(l)} \\left(\\boldsymbol {H} ^ {(l)}\\right) \\cdot W _ {o} ^ {(l)}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.822, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\( Attn_{n}^{(l)} \\) denotes the attention operation of the \\( n \\)-th head at the \\( l \\)-th layer, and \\( W_{o}^{(l)} \\in \\mathbb{R}^{d \\times Nd} \\) is the output projection matrix, with \\( d \\) representing the dimensionality of each attention head and \\( N \\) the number of heads. Subsequently, the model autoregressively predicts the next token" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.782, + 0.107 + ], + "angle": 0, + "content": "based on the output from the \\(L\\)-th layer:" + }, + { + "type": "equation", + "bbox": [ + 0.584, + 0.114, + 0.907, + 0.133 + ], + "angle": 0, + "content": "\\[\np \\left(y _ {t} \\mid y _ {< t}\\right) = \\operatorname {S o f t m a x} \\left(f c \\left(\\boldsymbol {H} _ {t} ^ {(L)}\\right)\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.14, + 0.907, + 0.171 + ], + "angle": 0, + "content": "where \\( f c(\\cdot) \\) is an affine layer mapping the \\( Hd \\)-dimensional vector to a probability distribution of vocabulary size \\( M \\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.182, + 0.649, + 0.199 + ], + "angle": 0, + "content": "4. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.207, + 0.906, + 0.299 + ], + "angle": 0, + "content": "In this section, we introduce two levels of fine-grained intervention modules. The first module enables the LVLM to focus on the image, thereby reducing over-reliance on language priors. The second module encourages the LVLM to attend more closely to image objects, helping to mitigate the omission of critical objects and reduce hallucinations." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.306, + 0.766, + 0.321 + ], + "angle": 0, + "content": "4.1. Intervention at Image-Level" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.327, + 0.906, + 0.403 + ], + "angle": 0, + "content": "This module aims to identify the attention heads associated with overall image information and to apply targeted interventions to these heads. This approach enhances the model's focus on visual input while diminishing the influence of language priors." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.403, + 0.906, + 0.54 + ], + "angle": 0, + "content": "As illustrated in Figure 2, consider a batch of image-question pairs, denoted as \\(\\{(X_i,V_i)\\}_{i = 1}^B\\), each labeled with the answer \"Yes\" from the POPE dataset. Here, \\(X_{i}\\) takes the form of \"Is there a/an [object] in the image?\" For each question, we extract the specified object \\(O_{i}\\) and reformulate the question to \\(q =\\) \"What is it in the image?\" Then, for each image \\(V_{i}\\), we progressively add Gaussian noise following the forward diffusion process [26] to obtain the final blurred image \\(V_{i}^{\\prime}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.531, + 0.545, + 0.905, + 0.609 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} q (V _ {i} ^ {(t)} \\mid V _ {i} ^ {(t - 1)}) = \\mathcal {N} (V _ {i} ^ {(t)}; \\sqrt {1 - \\beta_ {t}} V _ {i} ^ {(t - 1)}, \\beta_ {t} \\mathbf {I}), \\\\ q \\left(V _ {i} ^ {\\prime} \\mid V _ {i}\\right) = \\prod_ {t = 1} ^ {T} q \\left(V _ {i} ^ {(t)} \\mid V _ {i} ^ {(t - 1)}\\right), \\tag {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.615, + 0.907, + 0.858 + ], + "angle": 0, + "content": "where \\(\\beta_{t}\\) is the noise variance at step \\(t\\). Finally, we construct a modified dataset \\(\\{(q,V_i,V_i',O_i)\\}_{i=1}^B\\) to obtain Image-Level intervention vectors. We then concatenate \\(O_i\\) with the question \\(q\\) to form the answer, resulting in data pairs of trusted and untrusted data for the same sample: \\((q + O_i,V_i)\\) and \\((q + O_i,V_i')\\). For each sample, we treat the representation of the last token as the fused representation of multimodal data and extract the attention activation from a total of \\(N\\times L\\) heads across the \\(L\\) layers for both trusted and untrusted data. These are denoted as \\(A_{i} = \\{Attn_{n}^{(l)}H_{s}^{(l)}\\}_{n=1,l=1}^{N,L}\\) and \\(A_{i}' = \\{Attn_{n}^{(l)}H_{s}^{(l)}\\}_{n=1,l=1}^{N,L}\\), where \\(s = N_x + N_v + 1\\). According to the trusted activation \\(\\{A_{i}\\}_{i=1}^{B}\\) and the untrusted activation \\(\\{A_{i}'\\}_{i=1}^{B}\\) obtained from all sample pairs, we can calculate the following activation shift vector \\(\\{S_{n}^{(l)}\\}_{n=1,l=1}^{N,L}\\) that encourages the model to pay more attention to visual information, as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.865, + 0.907, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {S} _ {n} ^ {(l)} = \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\left(\\boldsymbol {A} _ {i, n} ^ {(l)} - \\boldsymbol {A} _ {i, n} ^ {\\prime (l)}\\right). \\tag {4}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "4211" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.91, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.302, + 0.908, + 0.358 + ], + "angle": 0, + "content": "Figure 2. Overview of our proposed ICT method. ICT consists of two levels of intervention: Image-Level and Object-Level. The Image-Level module enhances the model's focus on the overall scene to reduce reliance on language priors, while the Object-Level module focuses on fine-grained details to mitigate hallucinations. We apply targeted activation shifts to selected attention heads identified through binary classifiers trained to distinguish trusted and untrusted data pairs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.366, + 0.484, + 0.457 + ], + "angle": 0, + "content": "Next, we train a binary classifier \\( f_{n}^{(l)}(\\cdot) \\) using \\( B \\) sample pairs for each head to detect which heads encode Image-Level information, specifically those that can better distinguish the differences between pairs of trusted and untrusted samples. We then apply activation interventions to these selected heads:" + }, + { + "type": "equation", + "bbox": [ + 0.092, + 0.467, + 0.484, + 0.488 + ], + "angle": 0, + "content": "\\[\nA t t n _ {\\mathrm {i m g}} = \\left\\{A t t n _ {n} ^ {(l)} \\mid A t t n _ {n} ^ {(l)} \\in T o p K (\\text {A c c u r a c y} (f _ {n} ^ {(l)} (\\cdot))) \\right\\},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.49, + 0.485, + 0.544 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {H} ^ {(l + 1)} = \\boldsymbol {H} ^ {(l)} + \\sum_ {n = 1} ^ {N} \\left(A t t n _ {n} ^ {(l)} \\left(\\boldsymbol {H} ^ {(l)}\\right) + \\mathbb {I} _ {\\operatorname {i m g}, n} ^ {(l)} \\alpha \\boldsymbol {S} _ {n} ^ {(l)}\\right) \\cdot W _ {o} ^ {(l)}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.544, + 0.484, + 0.673 + ], + "angle": 0, + "content": "where \\(K\\) is the number of selected intervention heads, \\(\\mathbb{I}_{\\mathrm{img},n}^{(l)}\\) is an indicator function, which is 1 if \\(Attn_{n}^{(l)} \\in Attn_{\\mathrm{img}}\\) and 0 otherwise, and \\(\\alpha\\) denotes the intensity of the intervention. After using \\(\\{S_n^{(l)}\\}_{n=1,l=1}^{N,L}\\) to perform Image-Level interventions on heads that encode image information, the model enhances the trustworthiness of the visual level, places greater attention on visual information, thus mitigates the impact of overly strong language priors." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.682, + 0.347, + 0.699 + ], + "angle": 0, + "content": "4.2. Intervention at Object-Level" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.903 + ], + "angle": 0, + "content": "After enhancing the model's trustworthiness at the Image-Level, a more fine-grained, Object-Level intervention becomes necessary to increase the model's attention to image details, thereby reducing hallucinations caused by the omission of fine details. As shown in Figure 2, similar to Section 4.1, we construct a dataset \\(\\{(q,V_i,V_i^{\\prime \\prime},O_i)\\}_{i = 1}^{B}\\) to obtain Object-Level intervention vectors. We use Grounding DINO [58] to identify the area of object \\(O_{i}\\) in image \\(V_{i}\\). Gaussian noise is then added selectively to this object region, based on Eq. 3, producing a locally blurred image \\(V_{i}^{\\prime \\prime}\\). Using both the original image \\(V_{i}\\) and the partially blurred image \\(V_{i}^{\\prime \\prime}\\), we construct trusted and untrusted data pairs for each sample: \\((q + O_i,V_i)\\) and \\((q + O_i,V_i^{\\prime \\prime})\\) and analyze" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.367, + 0.907, + 0.416 + ], + "angle": 0, + "content": "the attention activation values across \\(N\\times L\\) heads, allowing us to compute an Object-Level activation shift vector \\(\\{S_{n,\\mathrm{obj}}^{(l)}\\}_{n = 1,l = 1}^{N,L}\\)" + }, + { + "type": "equation", + "bbox": [ + 0.596, + 0.419, + 0.907, + 0.459 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {S} _ {n, \\mathrm {o b j}} ^ {(l)} = \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\left(\\boldsymbol {A} _ {i, n} ^ {(l)} - \\boldsymbol {A} _ {i, n} ^ {\\prime \\prime (l)}\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.469, + 0.907, + 0.531 + ], + "angle": 0, + "content": "A binary classifier \\( f_{n,obj}^{(l)}(\\cdot) \\) is then trained to identify heads that effectively distinguish trusted from untrusted object-focused samples. Interventions are subsequently applied to the selected heads as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.542, + 0.912, + 0.565 + ], + "angle": 0, + "content": "\\[\nA t t n _ {\\mathrm {o b j}} = \\left\\{A t t n _ {n} ^ {(l)} \\mid A t t n _ {n} ^ {(l)} \\in T o p K (\\text {A c c u r a c y} (f _ {n, o b j} ^ {(l)} (\\cdot))) \\right\\},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.567, + 0.911, + 0.621 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {H} ^ {(l + 1)} = \\boldsymbol {H} ^ {(l)} + \\sum_ {n = 1} ^ {N} \\left(A t t n _ {n} ^ {(l)} \\left(\\boldsymbol {H} ^ {(l)}\\right) + \\mathbb {I} _ {\\mathrm {o b j}, n} ^ {(l)} \\beta \\boldsymbol {S} _ {\\mathrm {o b j}, n} ^ {(l)}\\right) \\cdot W _ {o} ^ {(l)}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.621, + 0.906, + 0.672 + ], + "angle": 0, + "content": "where \\(K\\) is the number of intervention heads, \\(\\mathbb{I}_{\\mathrm{obj},n}^{(l)}\\) is an indicator function set to 1 if \\(Attn_{n}^{(l)} \\in Attn_{\\mathrm{obj}}\\) and 0 otherwise, and \\(\\beta\\) controls the intervention intensity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.673, + 0.907, + 0.734 + ], + "angle": 0, + "content": "Finally, we integrate the Image-Level and Object-Level intervention modules to create a unified approach that reinforces the model's focus on both the overall visual context and finer object-specific details as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.554, + 0.745, + 0.905, + 0.814 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\boldsymbol {H} ^ {(l + 1)} = \\boldsymbol {H} ^ {(l)} + \\sum_ {n = 1} ^ {N} \\left(A t t n _ {n} ^ {(l)} \\left(\\boldsymbol {H} ^ {(l)}\\right) \\right. \\tag {8} \\\\ \\left. \\right.\\left. + \\mathbb {I} _ {\\operatorname {i m g}, n} ^ {(l)} \\alpha \\boldsymbol {S} _ {n} ^ {(l)} + \\mathbb {I} _ {\\operatorname {o b j}, n} ^ {(l)} \\beta \\boldsymbol {S} _ {\\operatorname {o b j}, n} ^ {(l)}\\right) \\cdot W _ {o} ^ {(l)}. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.907, + 0.901 + ], + "angle": 0, + "content": "By applying these interventions in tandem, the model gains a balanced attention mechanism that mitigates excessive reliance on language priors while enhancing sensitivity to essential visual cues across different levels of granularity, thereby alleviating the occurrence of hallucinations." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4212" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.09, + 0.224, + 0.108 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.115, + 0.281, + 0.133 + ], + "angle": 0, + "content": "5.1. Experimental Setup" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.139, + 0.248, + 0.153 + ], + "angle": 0, + "content": "Datasets and Metrics." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.154, + 0.483, + 0.32 + ], + "angle": 0, + "content": "POPE [48] is a benchmark designed to assess how effectively LVLMs recognize the presence of specific objects in images, thus identifying Object-Level hallucinations. It uses Yes/No questions based on object annotations, with metrics including Accuracy, Precision, Recall, and F1 score. The dataset is balanced, with \\(50\\%\\) of queries for existing objects and \\(50\\%\\) for non-existing ones, and employs three sampling strategies: random, popular, and adversarial. Drawing from MSCOCO [49], A-OKVQA [74], and GQA [32] datasets, POPE evaluates 27,000 query-answer pairs to gauge model performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.321, + 0.484, + 0.442 + ], + "angle": 0, + "content": "MME [20] benchmark is designed to comprehensively evaluate the performance of LVLMs across various dimensions. It includes ten perception-focused tasks and four cognition-related tasks. MME benchmark specifically addresses Object-Level hallucinations through subsets focusing on object existence and count, while attribute-level hallucinations are assessed via subsets related to object position and color. The evaluation metric is Accuracy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.442, + 0.484, + 0.517 + ], + "angle": 0, + "content": "Baselines. We adopt the widely used LLaVA-v1.5 [52] and Qwen-VL [5] models as our baseline LVLMs. We compared two baselines that eliminate language priors of the LVLMs in the decoding stage to alleviate hallucinations: VCD [43] and Opera [31]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.519, + 0.484, + 0.746 + ], + "angle": 0, + "content": "Implementation Details. In our experiments, we utilized 1,500 QA pairs with \"Yes\" responses from the COCO Random subset of the POPE dataset to train the intervention shift vector. Subsequently, we evaluated the ICT approach on two datasets with significant distributional differences: POPE, and MME. This evaluation was conducted to assess the generalizability and robustness of ICT across diverse data distributions. For each attention head, we employed Support Vector Machines (SVMs) [15] as the classifier and performed 2-fold cross-validation to evaluate the classification accuracy. In the experiment, we set \\(\\alpha = \\beta\\) and determined the optimal values of \\(\\alpha\\), \\(\\beta\\), and \\(K\\) through a grid search. Detailed hyperparameter configurations are provided in Appendix A. All experiments were conducted on a system equipped with \\(8 \\times\\) H800 GPUs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.757, + 0.231, + 0.771 + ], + "angle": 0, + "content": "5.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.779, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Results on POPE. Table 1 presents the results of LLaVA-v1.5 and Qwen-VL on nine subsets of the POPE dataset. From comparing these methods, we can derive the following conclusions: 1) Applying ICT results in an average performance improvement on the F1 score for LLaVA-v1.5 and Qwen-VL of \\(7.09\\%\\) and \\(5.44\\%\\) on 9 subsets respectively. This improvement can be attributed to the fact that ICT does not eliminate language priors, which may provide" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.381 + ], + "angle": 0, + "content": "useful information. Instead, it enhances the model's attention to various levels of visual information, thereby reducing the model's tendency to overly rely on language priors and mitigating the occurrence of hallucinations. 2) Individually intervening at both the image level and object level, with average gains of \\(5.76\\%\\) and \\(5.47\\%\\) in F1 score, respectively. This demonstrates that enhancing LVLMs' attention to image information at various levels can effectively mitigate hallucinations. Moreover, since Object-Level intervention also implicitly directs the model's focus toward broader image information, it achieves a comparatively higher performance gain. 3) The intervention shift vector trained using 1,500 samples from the MSCOCO random subset achieves an average improvement on F1 score of \\(7.67\\%\\) on MSCOCO random and an average improvement of \\(6.09\\%\\) across the other eight subsets with different distributions. This indicates that ICT generalizes effectively, and the resulting intervention vector captures a general direction towards trustworthiness rather than fitting a specific dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.405, + 0.909, + 0.693 + ], + "angle": 0, + "content": "Results on MME. Figure 3 presents the experimental results on the MME benchmark, where we followed Leng et al. [43] evaluation setup, focusing on the hallucination subset of MME. Compared to the unmodified model, ICT achieved a score improvement of 80.51 on LLaVA-v1.5 and 34.67 on Qwen-VL, surpassing the performance of VCD. Additionally, on non-hallucination tasks such as commonsense QA, ICT demonstrated a 10-point improvement on LLaVA-v1.5 and a 9.5-point improvement on Qwen-VL. These results indicate that ICT not only effectively reduces hallucinations but also enhances general reasoning capabilities across different models. Moreover, since the objects present in the images of the MME dataset are generally prominent, the variant ICT \\( w/o \\) image provides only limited improvements compared to ICT \\( w/o \\) object. This demonstrates that the absence of Image-Level guidance restricts the model's ability to fully leverage global visual context, thereby reducing its effectiveness in scenarios where the dominant objects play a crucial role." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.717, + 0.905, + 0.859 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.863, + 0.909, + 0.919 + ], + "angle": 0, + "content": "Figure 3. Comparison of ICT with baseline methods (Vanilla and VCD) on the MME benchmark. The radar chart illustrates improvements across various evaluation categories, including existence, position, count, color, and commonsense QA (CSQA)." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4213" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.152, + 0.088, + 0.849, + 0.822 + ], + "angle": 0, + "content": "
DatasetSettingMethodLLaVA-v1.5Qwen-VL
AccuracyF1 ScoreAccuracyF1 Score
COCORandomRegular83.2981.3384.3782.67
VCD87.7387.1688.6387.81
OPERIA89.2088.8187.3186.92
ICT w/o image89.70 (+6.41)89.93 (+8.60)88.76(+4.39)87.84(+5.17)
ICT w/o object88.50 (+5.21)88.81 (+7.48)89.13(+4.76)88.20(+5.63)
ICT90.11 (+6.82)90.03 (+8.70)89.46(+5.09)89.20(+6.63)
PopularRegular81.8880.0684.1382.06
VCD85.3885.0687.1286.40
OPERIA86.6486.6287.4486.68
ICT w/o image87.43 (+5.55)87.09 (+7.03)87.63(+3.50)86.78(+4.72)
ICT w/o object87.30 (+5.42)86.65 (+6.59)87.10(+2.97)86.28(+4.22)
ICT87.50 (+5.62)87.60 (+7.54)88.16(+4.03)87.33(+5.27)
AdversarialRegular78.9677.5782.2680.37
VCD80.8881.3384.2683.90
OPERIA81.2481.3884.7883.45
ICT w/o image84.13 (+5.17)83.31 (+5.74)85.13(+2.87)84.35(+3.98)
ICT w/o object83.63 (+4.67)82.13 (+4.56)85.23(+2.97)84.62(+4.25)
ICT84.43 (+5.47)83.74 (+6.17)84.96(+2.70)84.42(+4.05)
AOKVQARandomRegular83.4582.5686.6785.59
VCD86.1586.3489.2289.01
OPERIA88.0284.5988.1988.43
ICT w/o image89.60 (+6.15)89.75 (+7.19)89.46(+2.79)89.28(+3.69)
ICT w/o object89.00 (+5.55)88.91 (+6.35)89.26(+2.59)88.95(+3.36)
ICT89.20 (+5.75)89.41 (+6.85)89.46(+2.79)89.03(+3.44)
PopularRegular79.9079.5985.5684.63
VCD81.8582.8287.8587.81
OPERIA83.2284.6787.9187.13
ICT w/o image85.36 (+5.46)85.34 (+5.75)87.70(+2.14)87.68(+3.05)
ICT w/o object84.65 (+4.75)84.32 (+4.73)87.90(+2.34)87.48(+2.85)
ICT85.73 (+5.83)85.34 (+5.75)88.13(+3.57)87.83(+3.20)
AdversarialRegular74.0475.1579.5779.50
VCD74.9777.7381.2782.38
OPERIA73.8277.9180.8281.54
ICT w/o image79.26 (+5.22)80.14 (+4.99)81.50(+1.93)82.10(+2.60)
ICT w/o object77.79 (+3.75)79.53 (+4.38)81.40(+1.83)82.18(+2.68)
ICT79.60 (+5.56)80.43 (+5.28)81.94(+2.37)82.44(+2.94)
GQARandomRegular83.7382.9580.9779.01
VCD86.6586.9985.5985.33
OPERIA88.1388.9186.0285.29
ICT w/o image89.03 (+5.30)88.99 (+6.04)85.20(+4.23)85.90(+6.89)
ICT w/o object88.97 (+5.24)89.30 (+6.35)87.20(+6.23)86.80(+7.79)
ICT89.60 (+5.87)89.44 (+6.49)86.38(+5.41)86.96(+7.95)
PopularRegular78.1778.3775.9974.84
VCD80.7382.2481.8382.23
OPERIA79.2782.1181.9782.12
ICT w/o image84.23 (+6.06)83.92 (+5.55)81.43(+5.44)80.86(+6.02)
ICT w/o object84.70 (+6.53)84.45 (+6.08)83.93(+7.94)82.95(+8.11)
ICT84.70 (+6.53)84.78 (+6.41)82.63(+6.64)82.22(+7.38)
AdversarialRegular75.0876.0675.4674.33
VCD76.0978.7880.0180.75
OPERIA75.0078.7180.2480.64
ICT w/o image81.92 (+6.84)82.11 (+6.05)80.03(+4.57)79.90(+5.47)
ICT w/o object81.65 (+6.57)82.00 (+5.94)81.60(+6.14)81.53(+7.10)
ICT81.50 (+6.42)82.27 (+6.21)80.83(+5.37)80.60(+6.17)
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.83, + 0.907, + 0.888 + ], + "angle": 0, + "content": "Table 1. Main results on POPE tasks. We evaluate the accuracy of various LVLMs on the POPE task across the MSCOCO, A-OKVQA, and GQA datasets. Regular represents the setting where direct sampling is applied. ICT \\( w/o \\) image and ICT \\( w/o \\) object correspond to the exclusion of image-level and object-level interventions, respectively. The bold and underlined values indicate the highest and second-highest metrics under each setting, respectively." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4214" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.09, + 0.325, + 0.108 + ], + "angle": 0, + "content": "6. Analysis and Discussions" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.115, + 0.465, + 0.132 + ], + "angle": 0, + "content": "6.1. Inference Latency across Different Methods" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.137, + 0.483, + 0.289 + ], + "angle": 0, + "content": "In Table 2, we compare the inference efficiency of two different representative approaches with our proposed method. Notably, while the CD-based methods result in a significant increase in inference time (2.1 to 3.6 times slower), our ICT incurs virtually no impact on inference efficiency. We attribute this to the fact that both VCD and OPERA either require multiple inference runs or involve substantial additional computations during inference. In contrast, ICT introduces no extra computational cost at inference, further highlighting the advantages of our method." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.291, + 0.476, + 0.372 + ], + "angle": 0, + "content": "
Method20-Token Len50-Token Len80-Token Len
LLaVA-v1.5405.3 ↑×1.0934.6 ↑×1.01440.0 ↑×1.0
+ VCD988.9 ↑ × 2.42031.7 ↑ × 2.23077.5 ↑ × 2.1
+ OPERA1371.7 ↑ × 3.43294.1 ↑ × 3.55717.3 ↑ × 3.6
+ ICT415.9 ↑ × 1.0931.9 ↑ × 1.01485.5 ↑ × 1.0
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.373, + 0.483, + 0.414 + ], + "angle": 0, + "content": "Table 2. Comparison of the efficiency of different methods in generating tokens of varying lengths on an NVIDIA H800 GPU. Inference times are recorded in milliseconds." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.422, + 0.453, + 0.439 + ], + "angle": 0, + "content": "6.2. Generalizability between Different Models" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.445, + 0.483, + 0.835 + ], + "angle": 0, + "content": "To verify that the derived activation shift vectors to some extent represent the model's alignment with authenticity, we explored whether these shifts are transferable to other models. Specifically, we applied the activation shift vectors obtained from LLaVA-v1.5 on the COCO random subset to intervene in Qwen-VL. We then evaluated the generalizability of this intervention on the GQA dataset, which has a distribution significantly different from COCO. The results in Table 3 demonstrate that ICT-LLaVA-v1.5 achieves an average improvement of \\(4.62\\%\\) in F1 score and \\(4.64\\%\\) in accuracy compared to the unmodified model. Additionally, we employed t-SNE to visualize the offset vectors across different layers for LLaVA-v1.5 and Qwen-VL. Figure 4 shows the Object-Level and Image-Level shift vectors of multi-head attention in the 16th and 18th layers for both models. From the figure, we can observe that in layer 16, the Image-Level vectors (blue and yellow) are relatively close to each other, indicating a degree of similarity in encoding Image-Level information between the models. However, the Object-Level vectors (red and green) are more distinct, reflecting model-specific characteristics in encoding fine-grained object details. In layer 18, the Image-Level and Object-Level vectors for both models exhibit high similarity, indicating that the resulting activation shift vectors represent universal shift vectors that guide the model to pay more attention to visual information." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.835, + 0.471, + 0.851 + ], + "angle": 0, + "content": "6.3. Impact of Hyperparameters on Performance" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Our proposed ICT primarily relies on two key hyperparameters: the intervention intensity \\(\\alpha\\) and the number of heads \\(K\\) involved in the intervention. To investigate the" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.09, + 0.905, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.207, + 0.908, + 0.508 + ], + "angle": 0, + "content": "Figure 4. t-SNE visualization of Object-Level and Image-Level offset vectors for LLaVA-v1.5 and Qwen-VL at layers 16 and 18. Impact of these hyperparameters on performance, we fixed one of the parameters and calculated the average F1 score for LLaVA-v1.5 and Qwen-VL on the POPE COCO random subset. The results are presented in Figure 6, we can observe that when \\(\\alpha\\) is negative—indicating that reverse intervention reduces the model's attention to visual information—the model's performance declines. When \\(\\alpha\\) is too small, the model does not receive sufficient intervention, resulting in suboptimal outcomes. Conversely, when \\(\\alpha\\) is excessively large, the intervention becomes too strong, disrupting the model's foundational capabilities and leading to a decrease in performance. For the hyperparameter \\(K\\), we find that when \\(K\\) is too small, certain attention heads that encode relevant visual information are not adequately intervened upon, resulting in suboptimal performance. Conversely, when \\(K\\) is too large, attention heads that encode irrelevant information are unnecessarily interfered with, which leads to performance degradation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.522, + 0.907, + 0.556 + ], + "angle": 0, + "content": "6.4. Analysis of Attention Heads for Visual Information Encoding" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.56, + 0.907, + 0.651 + ], + "angle": 0, + "content": "In Figure 7, we present the classification accuracy of the binary classifiers for Image-Level and Object-Level features across the 1,024 multi-head attention heads in the 32 layers of LLaVA-v1.5. This analysis helps identify which heads encode overall visual information and which heads capture fine-grained visual details. From the figure, we can observe" + }, + { + "type": "table", + "bbox": [ + 0.544, + 0.659, + 0.878, + 0.849 + ], + "angle": 0, + "content": "
SettingMethodAccuracyF1 Score
RandomRegular80.9779.01
VCD85.5985.33
ICT86.3886.96
ICT-LLaVA-v1.585.1083.27
PopularRegular75.9974.84
VCD81.8382.23
ICT82.6382.22
ICT-LLaVA-v1.581.5080.10
AdversarialRegular75.4674.33
VCD80.0180.75
ICT80.8380.60
ICT-LLaVA-v1.579.7378.68
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.859, + 0.909, + 0.915 + ], + "angle": 0, + "content": "Table 3. Cross-model generalization of ICT. Performance of Qwen-VL on GQA after applying activation shift from LLaVA-v1.5. ICT-LLaVA-v1.5 refers to the results obtained by applying LLaVA-v1.5's activation shift vectors to intervene in Qwen-VL." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4215" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.189, + 0.109, + 0.355, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.379, + 0.113, + 0.519, + 0.139 + ], + "angle": 0, + "content": "Question: Is there a horse in the image?" + }, + { + "type": "image_caption", + "bbox": [ + 0.389, + 0.152, + 0.513, + 0.164 + ], + "angle": 0, + "content": "VCD: Yes, there is." + }, + { + "type": "image_caption", + "bbox": [ + 0.38, + 0.186, + 0.522, + 0.198 + ], + "angle": 0, + "content": "ICT: No, there isn't." + }, + { + "type": "image_caption", + "bbox": [ + 0.648, + 0.107, + 0.686, + 0.118 + ], + "angle": 0, + "content": "POPE" + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.12, + 0.78, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.189, + 0.217, + 0.334, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.354, + 0.223, + 0.528, + 0.248 + ], + "angle": 0, + "content": "Question: How many uncut fruits are in the image?" + }, + { + "type": "image_caption", + "bbox": [ + 0.354, + 0.259, + 0.528, + 0.284 + ], + "angle": 0, + "content": "VCD: There are two uncut fruits in the image." + }, + { + "type": "image_caption", + "bbox": [ + 0.356, + 0.297, + 0.532, + 0.323 + ], + "angle": 0, + "content": "ICT: There are four uncut fruits in the image." + }, + { + "type": "image_caption", + "bbox": [ + 0.629, + 0.228, + 0.723, + 0.24 + ], + "angle": 0, + "content": "LLaVA-Bench" + }, + { + "type": "image", + "bbox": [ + 0.556, + 0.246, + 0.794, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.33, + 0.907, + 0.358 + ], + "angle": 0, + "content": "Figure 5. Case Study and Error Analysis of ICT. The vertical axis in the right section of the figure represents the attention ratio of each question text token to the visual tokens." + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.362, + 0.283, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.362, + 0.482, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.487, + 0.483, + 0.527 + ], + "angle": 0, + "content": "Figure 6. Effect of hyperparameters \\(\\alpha\\) (left) and \\(K\\) (right) on the F1 score for LLaVA-v1.5 and Qwen-VL models on the POPE COCO random subset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.53, + 0.484, + 0.605 + ], + "angle": 0, + "content": "that the attention heads encoding overall visual information are predominantly located in the earlier layers of the model, such as layers 4 and 10. In contrast, the attention heads that capture fine-grained visual details are more concentrated in the later layers, such as layers 20 and 22." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.619, + 0.367, + 0.634 + ], + "angle": 0, + "content": "6.5. Case Study and Error Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.642, + 0.484, + 0.733 + ], + "angle": 0, + "content": "In Figure 5, we present a case study on the POPE and LLaVA-Bench datasets, visualizing the attention ratio of each question text token to the visual tokens. As illustrated, after applying ICT, the model allocates a higher proportion of attention to visual tokens, especially to object tokens relevant to the question (e.g., \"horse\" and \"fruits\"). By prior-" + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.757, + 0.482, + 0.892 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.898, + 0.483, + 0.926 + ], + "angle": 0, + "content": "Figure 7. Heatmap of classification accuracy for Image-Level (left) and Object-Level (right) features of LLaVA-v1.5." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.364, + 0.907, + 0.545 + ], + "angle": 0, + "content": "itizing visual information, ICT correctly identifies the absence of a horse in the image, whereas VCD erroneously concludes that a horse is present due to insufficient attention to visual cues. However, when asked, \"How many uncut fruits are in the picture?\", VCD incorrectly answered \"two\" due to a lack of focus on visual details. Although ICT correctly identifies that there are a total of four fruits in the image, this question requires not only attention to visual content but also reasoning within the text modality. The model needs to not only recognize the overall count of fruits but also focus on the attribute \"uncut\". Due to its failure to incorporate this information, ICT outputs a wrong answer." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.547, + 0.909, + 0.698 + ], + "angle": 0, + "content": "Furthermore, when the text becomes more complex, such as in \"How many uncut fruits are there in the image?\", the increased linguistic complexity demands greater attention from the model. This, in turn, limits the extent to which attention can be allocated to the visual domain. In contrast, function words like \"is\" and \"there\" are less complex and require minimal processing, allowing the model to shift more focus toward visual tokens. As a result, this redistribution of attention can lead to a more noticeable shift in how the model processes visual and textual information." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.723, + 0.771, + 0.739 + ], + "angle": 0, + "content": "7. Conclusion and Limitations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.901 + ], + "angle": 0, + "content": "In this paper, we mitigate the issue of object hallucinations in LVLMs by proposing Image-Object Cross-Level Trusted Intervention (ICT), a training-free, plug-and-play method that enhances the model's focus on both Image-Level and Object-Level visual information during the forward pass. Our experiments show that ICT significantly reduces hallucinations while also improving general reasoning capabilities. The results confirm that ICT effectively mitigates excessive reliance on language priors, enhancing both accuracy and robustness in diverse visual contexts." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4216" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.486, + 0.256 + ], + "angle": 0, + "content": "This work was supported by the National Key Research and Development Program of China (Grant No. 2024YFB3309702); Guangdong Provincial Department of Education Project (Grant No.2024KQNCX028); CAAI-Ant Group Research Fund; Scientific Research Projects for the Higher-educational Institutions (Grant No.2024312096), Education Bureau of Guangzhou Municipality; Guangzhou-HKUST(GZ) Joint Funding Program (Grant No.2025A03J3957), Education Bureau of Guangzhou Municipality." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.282, + 0.188, + 0.298 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.307, + 0.482, + 0.363 + ], + "angle": 0, + "content": "[1] Renat Aksitov, Chung-Ching Chang, David Reitter, Siamak Shakeri, and Yunhsuan Sung. Characterizing attribution and fluency tradeoffs for retrieval-augmented large language models, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.365, + 0.483, + 0.447 + ], + "angle": 0, + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: A visual language model for few-shot learning. In Advances in Neural Information Processing Systems, pages 23716-23736, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.449, + 0.483, + 0.517 + ], + "angle": 0, + "content": "[3] Wenbin An, Feng Tian, Sicong Leng, Jiahao Nie, Haonan Lin, QianYing Wang, Guang Dai, Ping Chen, and Shijian Lu. Agla: Mitigating object hallucinations in large vision-language models with assembly of global and local attention. arXiv preprint, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.52, + 0.482, + 0.588 + ], + "angle": 0, + "content": "[4] Wenbin An, Feng Tian, Sicong Leng, Jiahao Nie, Haonan Lin, QianYing Wang, Guang Dai, Ping Chen, and Shijian Lu. Agla: Mitigating object hallucinations in large vision-language models with assembly of global and local attention. arXiv preprint arXiv:2406.12718, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.591, + 0.482, + 0.658 + ], + "angle": 0, + "content": "[5] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. 2023. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.661, + 0.482, + 0.743 + ], + "angle": 0, + "content": "[6] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D. Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In Advances in Neural Information Processing Systems, pages 1877-1901, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.746, + 0.482, + 0.8 + ], + "angle": 0, + "content": "[7] Beitao Chen, Xinyu Lyu, Lianli Gao, Jingkuan Song, and Heng Tao Shen. Alleviating hallucinations in large vision-language models through hallucination-induced optimization, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.803, + 0.482, + 0.842 + ], + "angle": 0, + "content": "[8] Haozhe Chen, Carl Vondrick, and Chengzhi Mao. Selfie: Self-interpretation of large language model embeddings, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[9] Zhaorun Chen, Zhuokai Zhao, Hongyin Luo, Huaxiu Yao, Bo Li, and Jiawei Zhou. Halc: Object hallucination reduction via adaptive focal-contrast decoding. In *Forty-first International Conference on Machine Learning*. 1" + }, + { + "type": "list", + "bbox": [ + 0.108, + 0.307, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[10] Zhaorun Chen, Zhuokai Zhao, Hongyin Luo, Huaxiu Yao, Bo Li, and Jiawei Zhou. Halc: Object hallucination reduction via adaptive focal-contrast decoding, 2024. http://arxiv.org/abs/2403.00425.2,3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.149, + 0.905, + 0.231 + ], + "angle": 0, + "content": "[11] I.-Chun Chern, Steffi Chern, Shiqi Chen, Weizhe Yuan, Kehua Feng, Chunting Zhou, Junxian He, Graham Neubig, and Pengfei Liu. Factool: Factuality detection in generative ai - a tool augmented framework for multi-task and multi-domain scenarios, 2023. http://arxiv.org/abs/2307.13528.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.233, + 0.905, + 0.314 + ], + "angle": 0, + "content": "[12] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with \\(90\\%\\) * chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2(3):6, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.316, + 0.904, + 0.383 + ], + "angle": 0, + "content": "[13] Jaemin Cho, Seunghyun Yoon, Ajinkya Kale, Franck Dernoncourt, Trung Bui, and Mohit Bansal. Fine-grained image captioning with clip reward. In Findings of the Association for Computational Linguistics: NAACL 2022, pages 517-527, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.385, + 0.904, + 0.453 + ], + "angle": 0, + "content": "[14] Yung-Sung Chuang, Yujia Xie, Hongyin Luo, Yoon Kim, James Glass, and Pengcheng He. Dola: Decoding by contrasting layers improves factuality in large language models, 2024. http://arxiv.org/abs/2309.03883.2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.455, + 0.904, + 0.482 + ], + "angle": 0, + "content": "[15] Corinna Cortes. Support-vector networks. Machine Learning, 1995. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.484, + 0.905, + 0.537 + ], + "angle": 0, + "content": "[16] Li D., Rawat A.S., Zaheer M., Wang X., Lukasik M., Veit A., Yu F., and Kumar S. Large language models with controllable working memory. arXiv preprint arXiv:2211.05110, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.539, + 0.905, + 0.606 + ], + "angle": 0, + "content": "[17] Hanxing Ding, Liang Pang, Zihao Wei, Huawei Shen, and Xueqi Cheng. Retrieve only when it needs: Adaptive retrieval augmentation for hallucination mitigation in large language models, 2024. http://arxiv.org/abs/2402.10612.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.609, + 0.905, + 0.661 + ], + "angle": 0, + "content": "[18] Sebastian Farquhar, Jannik Kossen, Lorenz Kuhn, and Yarin Gal. Detecting hallucinations in large language models using semantic entropy. Nature, 630(8017):625-630, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.664, + 0.905, + 0.705 + ], + "angle": 0, + "content": "[19] Javier Ferrando, Gabriele Sarti, Arianna Bisazza, and Marta R. Costa-jussà. A primer on the inner workings of transformer-based language models, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.707, + 0.905, + 0.774 + ], + "angle": 0, + "content": "[20] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.776, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[21] Suriya Gunasekar, Yi Zhang, Jyoti Aneja, Cao César Teodoro Mendes, Allie Del Giorno, Sivakanth Gopi, Mojan Javaheripi, Piero Kauffmann, Gustavo de Rosa, Olli Saarikivi, Adil Salim, Shital Shah, Harkirat Singh Behl, Xin Wang, Sébastien Bubeck, Ronen Eldan, Adam Tauman Kalai, Yin Tat Lee, and Yuanzhi Li. Textbooks are all you need, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[22] Anisha Gunjal, Jihan Yin, and Erhan Bas. Detecting and preventing hallucinations in large vision language models." + }, + { + "type": "list", + "bbox": [ + 0.523, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4217" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.133, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "In Proceedings of the AAAI Conference on Artificial Intelligence, pages 18135-18143, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.122, + 0.483, + 0.176 + ], + "angle": 0, + "content": "[23] Anisha Gunjal, Jihan Yin, and Erhan Bas. Detecting and preventing hallucinations in large vision language models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 18135-18143, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.177, + 0.482, + 0.26 + ], + "angle": 0, + "content": "[24] Vipul Gupta, Zhuowan Li, Adam Kortylewski, Chenyu Zhang, Yingwei Li, and Alan Yuille. Swapmix: Diagnosing and regularizing the over-reliance on visual context in visual question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5078-5088, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.261, + 0.483, + 0.315 + ], + "angle": 0, + "content": "[25] Yudong Han, Liqiang Nie, Jianhua Yin, Jianlong Wu, and Yan Yan. Visual perturbation-aware collaborative learning for overcoming the language prior problem. arXiv preprint arXiv:2207.11850, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.317, + 0.483, + 0.357 + ], + "angle": 0, + "content": "[26] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.359, + 0.482, + 0.426 + ], + "angle": 0, + "content": "[27] Wenyi Hong, Weihan Wang, Ming Ding, Wenmeng Yu, Qingsong Lv, Yan Wang, Yean Cheng, Shiyu Huang, Junhui Ji, Zhao Xue, et al. Cogvlm2: Visual language models for image and video understanding. arXiv preprint arXiv:2408.16500, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.428, + 0.482, + 0.482 + ], + "angle": 0, + "content": "[28] Haichuan Hu, Yuhan Sun, and Quanjun Zhang. Lrp4rag: Detecting hallucinations in retrieval-augmented generation via layer-wise relevance propagation, 2024. http:// arxiv.org/abs/2408.15533.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.484, + 0.482, + 0.538 + ], + "angle": 0, + "content": "[29] Yingdong Hu, Fanqi Lin, Tong Zhang, Li Yi, and Yang Gao. Look before you leap: Unveiling the power of gpt-4v in robotic vision-language planning. arXiv preprint arXiv:2311.17842, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.54, + 0.482, + 0.609 + ], + "angle": 0, + "content": "[30] Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, and Ting Liu. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.61, + 0.482, + 0.706 + ], + "angle": 0, + "content": "[31] Qidong Huang, Xiaoyi Dong, Pan Zhang, Bin Wang, Conghui He, Jiaqi Wang, Dahua Lin, Weiming Zhang, and Nenghai Yu. Opera: Alleviating hallucination in multimodal large language models via over-trust penalty and retrospection-allocation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13418-13427, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.707, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[32] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.776, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[33] Fushuo Huo, Wenchao Xu, Zhong Zhang, Haozhao Wang, Zhicheng Chen, and Peilin Zhao. Self-introspective decoding: Alleviating hallucinations for large vision-language models, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.832, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[34] Jitesh Jain, Jianwei Yang, and Humphrey Shi. Vcoder: Versatile vision encoders for multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27992-28002, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[35] Juseong Jin and Chang Wook Jeong. Surgical-llava: Toward surgical scenario understanding via large language and vision models. arXiv preprint arXiv:2410.09750, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.15, + 0.905, + 0.191 + ], + "angle": 0, + "content": "[36] Junho Kim, Hyunjun Kim, Yeonju Kim, and Yong Man Ro. Code: Contrasting self-generated description to combat hallucination in large multi-modal models, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.193, + 0.905, + 0.235 + ], + "angle": 0, + "content": "[37] Lorenz Kuhn, Yarin Gal, and Sebastian Farquhar. Semantic uncertainty: Linguistic invariances for uncertainty estimation in natural language generation. arXiv, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.236, + 0.905, + 0.277 + ], + "angle": 0, + "content": "[38] Junlin Lee, Yequan Wang, Jing Li, and Min Zhang. Multimodal reasoning with multimodal knowledge graph. arXiv preprint arXiv:2406.02030, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.28, + 0.905, + 0.362 + ], + "angle": 0, + "content": "[39] Katherine Lee, Daphne Ippolito, Andrew Nystrom, Chiyuan Zhang, Douglas Eck, Chris Callison-Burch, and Nicholas Carlini. Deduplicating training data makes language models better. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8424-8445, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.365, + 0.905, + 0.433 + ], + "angle": 0, + "content": "[40] Kang-il Lee, Minbeom Kim, Seunghyun Yoon, Minsung Kim, Dongryeol Lee, Hyukhun Koh, and Kyomin Jung. Vlind-bench: Measuring language priors in large vision-language models. arXiv preprint arXiv:2406.08702, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.435, + 0.905, + 0.504 + ], + "angle": 0, + "content": "[41] Nayeon Lee, Wei Ping, Peng Xu, Mostofa Patwary, Pascale N Fung, Mohammad Shoeybi, and Bryan Catanzaro. Factuality enhanced language models for open-ended text generation. In Advances in Neural Information Processing Systems, pages 34586-34599, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.506, + 0.905, + 0.547 + ], + "angle": 0, + "content": "[42] Seongyun Lee, Sue Hyun Park, Yongrae Jo, and Minjoon Seo. Volcano: Mitigating multimodal hallucination through self-feedback guided revision. arXiv, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.549, + 0.905, + 0.631 + ], + "angle": 0, + "content": "[43] Sicong Leng, Hang Zhang, Guanzheng Chen, Xin Li, Shijian Lu, Chunyan Miao, and Lidong Bing. Mitigating object hallucinations in large vision-language models through visual contrastive decoding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13872-13882, 2024. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.633, + 0.905, + 0.702 + ], + "angle": 0, + "content": "[44] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In Proceedings of the International Conference on Machine Learning, pages 19730-19742. PMLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.704, + 0.905, + 0.759 + ], + "angle": 0, + "content": "[45] Kenneth Li, Oam Patel, Fernanda Viégas, Hanspeter Pfister, and Martin Wattenberg. Inference-time intervention: Eliciting truthful answers from a language model. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.761, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[46] Lei Li, Zhihui Xie, Mukai Li, Shunian Chen, Peiyi Wang, Liang Chen, Yazheng Yang, Benyou Wang, Lingpeng Kong, and Qi Liu. Vlfeedback: A large-scale ai feedback dataset for large vision-language models alignment. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 6227-6246, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[47] Xiang Lisa Li, Ari Holtzman, Daniel Fried, Percy Liang, Jason Eisner, Tatsunori Hashimoto, Luke Zettlemoyer, and Mike Lewis. Contrastive decoding: Open-ended text generation as optimization. arXiv preprint, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4218" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.092, + 0.486, + 0.148 + ], + "angle": 0, + "content": "[48] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. In The 2023 Conference on Empirical Methods in Natural Language Processing. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.149, + 0.483, + 0.233 + ], + "angle": 0, + "content": "[49] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.234, + 0.483, + 0.276 + ], + "angle": 0, + "content": "[50] Bingbin Liu, Jordan T. Ash, Surbhi Goel, Akshay Krishnamurthy, and Cyril Zhang. Exposing attention glitches with flip-flop language modeling, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.277, + 0.483, + 0.345 + ], + "angle": 0, + "content": "[51] Fuxiao Liu, Kevin Lin, Linjie Li, Jianfeng Wang, Yaser Ya-coob, and Lijuan Wang. Mitigating hallucination in large multi-modal models via robust instruction tuning. In The Twelfth International Conference on Learning Representations, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.347, + 0.483, + 0.388 + ], + "angle": 0, + "content": "[52] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.39, + 0.483, + 0.432 + ], + "angle": 0, + "content": "[53] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.433, + 0.483, + 0.489 + ], + "angle": 0, + "content": "[54] Hanchao Liu, Wenyuan Xue, Yifei Chen, Dapeng Chen, Xiutian Zhao, Ke Wang, Liping Hou, Rongjun Li, and Wei Peng. A survey on hallucination in large vision-language models. arXiv preprint, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.49, + 0.483, + 0.545 + ], + "angle": 0, + "content": "[55] Hanchao Liu, Wenyuan Xue, Yifei Chen, Dapeng Chen, Xiutian Zhao, Ke Wang, Liping Hou, Rongjun Li, and Wei Peng. A survey on hallucination in large vision-language models. arXiv preprint arXiv:2402.00253, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.547, + 0.483, + 0.615 + ], + "angle": 0, + "content": "[56] Jiajun Liu, Yibing Wang, Hanghang Ma, Xiaoping Wu, Xiaqi Ma, Xiaoming Wei, Jianbin Jiao, Enhua Wu, and Jie Hu. Kangaroo: A powerful video-language model supporting long-context video input. arXiv preprint arXiv:2408.15542, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.617, + 0.483, + 0.672 + ], + "angle": 0, + "content": "[57] Jiazhen Liu, Yuhan Fu, Ruobing Xie, Runquan Xie, Xingwu Sun, Fengzong Lian, Zhanhui Kang, and Xirong Li. PhD: A chatgpt-prompted visual hallucination evaluation dataset, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.673, + 0.483, + 0.743 + ], + "angle": 0, + "content": "[58] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.745, + 0.483, + 0.786 + ], + "angle": 0, + "content": "[59] Shi Liu, Kecheng Zheng, and Wei Chen. Paying more attention to image: A training-free method for alleviating hallucination in lvlms, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.788, + 0.483, + 0.842 + ], + "angle": 0, + "content": "[60] Alexander Nikitin, Jannik Kossen, Yarin Gal, and Pekka Marttinen. Kernel language entropy: Fine-grained uncertainty quantification for llms from semantic similarities. arXiv, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.845, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[61] OpenAI. GPT-4. https://openai.com/gpt-4, 2023.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[62] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, et al." + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.092, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.553, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "Training language models to follow instructions with human feedback. In Advances in Neural Information Processing Systems, pages 27730-27744, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.136, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[63] Amandalynne Paullada, Inioluwa Deborah Raji, Emily M. Bender, Emily Denton, and Alex Hanna. Data and its (dis)contents: A survey of dataset development and use in machine learning research. *Patterns*, 2(11):100336, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.207, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[64] Phuc Phan, Hieu Tran, and Long Phan. Distillation contrastive decoding: Improving llms reasoning with contrastive decoding and distillation, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.249, + 0.905, + 0.29 + ], + "angle": 0, + "content": "[65] Venkit P.N., Gautam S., Panchanadikar R., Huang T.H., and Wilson S. Nationality bias in text generation. arXiv preprint arXiv:2302.02463, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.292, + 0.905, + 0.348 + ], + "angle": 0, + "content": "[66] Xiaoye Qu, Qiyuan Chen, Wei Wei, Jishuo Sun, and Jianfeng Dong. Alleviating hallucination in large vision-language models with active retrieval augmentation, 2024. http://arxiv.org/abs/2408.00555.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.349, + 0.905, + 0.405 + ], + "angle": 0, + "content": "[67] Xiaoye Qu, Jiashuo Sun, Wei Wei, and Yu Cheng. Look, compare, decide: Alleviating hallucination in large vision-language models via multi-view multi-path reasoning, 2024. http://arxiv.org/abs/2408.17150.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.407, + 0.905, + 0.489 + ], + "angle": 0, + "content": "[68] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.491, + 0.907, + 0.572 + ], + "angle": 0, + "content": "[69] Vikas Raunak, Arul Menezes, and Marcin Junczys-Dowmunt. The curious case of hallucinations in neural machine translation. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1172-1183, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.575, + 0.905, + 0.616 + ], + "angle": 0, + "content": "[70] Kim S., Cho B., Bae S., Ahn S., and Yun S.Y. Vacode: Visual augmented contrastive decoding. arXiv preprint arXiv:2408.05337, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.618, + 0.905, + 0.659 + ], + "angle": 0, + "content": "[71] Lin S., Hilton J., and Evans O. Truthfulqa: Measuring how models mimic human falsehoods. arXiv preprint arXiv:2109.07958, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.661, + 0.905, + 0.703 + ], + "angle": 0, + "content": "[72] Yin S., Fu C., Zhao S., Li K., Sun X., Xu T., and Chen E. A survey on multimodal large language models. arXiv preprint arXiv:2306.13549, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.704, + 0.905, + 0.786 + ], + "angle": 0, + "content": "[73] Pranab Sahoo, Prakash Meharia, Akash Ghosh, Sriparna Saha, Vinija Jain, and Aman Chadha. A comprehensive survey of hallucination in large language, image, video and audio foundation models. Findings of the Association for Computational Linguistics: EMNLP 2024, pages 11709-11724, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.789, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[74] Dustin Schwenk, Apoorv Khandelwal, Christopher Clark, Kenneth Marino, and Roozbeh Mottaghi. A-okvqa: A benchmark for visual question answering using world knowledge. In European conference on computer vision, pages 146–162. Springer, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[75] Lei Shen, Haolan Zhan, Xin Shen, Hongshen Chen, Xiaofang Zhao, and Xiaodan Zhu. Identifying untrustworthy samples: Data filtering for open-domain dialogues with" + }, + { + "type": "list", + "bbox": [ + 0.525, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4219" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.131, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "bayesian optimization. In Proceedings of the 30th ACM International Conference on Information & Knowledge Management, pages 1598-1608, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.136, + 0.483, + 0.189 + ], + "angle": 0, + "content": "[76] Weijia Shi, Xiaochuang Han, Mike Lewis, Yulia Tsvetkov, Luke Zettlemoyer, and Scott Wen tau Yih. Trusting your evidence: Hallucinate less with context-aware decoding, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.192, + 0.482, + 0.259 + ], + "angle": 0, + "content": "[77] Weijia Shi, Sewon Min, Maria Lomeli, Chunting Zhou, Margaret Li, Rich James, Xi Victoria Lin, Noah A. Smith, Luke Zettlemoyer, Scott Yih, and Mike Lewis. In-context pretraining: Language modeling beyond document boundaries, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.262, + 0.482, + 0.316 + ], + "angle": 0, + "content": "[78] GLM T., Zeng A., Xu B., Wang B., Zhang C., Yin D., Rojas D., Feng G., Zhao H., Lai H., and Yu H. Chatglm: A family of large language models from glm-130b to glm-4 all tools. arXiv preprint arXiv:2406.12793, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.318, + 0.483, + 0.773 + ], + "angle": 0, + "content": "[79] Gemma Team, Thomas Mesnard, Cassidy Hardin, Robert Dadashi, Surya Bhupatiraju, Shreya Pathak, Laurent Sifre, Morgane Rivière, Mihir Sanjay Kale, Juliette Love, Pouya Tafti, Léonard Hussenot, Pier Giuseppe Sessa, Aakanksha Chowdhery, Adam Roberts, Aditya Barua, Alex Botev, Alex Castro-Ros, Ambrose Slone, Amélie Héliou, Andrea Tacchetti, Anna Bulanova, Antonia Paterson, Beth Tsai, Bobak Shahriari, Charline Le Lan, Christopher A. Choquette-Choo, Clément Crepy, Daniel Cer, Daphne Ippolito, David Reid, Elena Buchatskaya, Eric Ni, Eric Noland, Geng Yan, George Tucker, George-Christian Muraru, Grigory Rozhdestvenskiy, Henryk Michalewski, Ian Tenney, Ivan Grishchenko, Jacob Austin, James Keeling, Jane Labanowski, Jean-Baptiste Lespiau, Jeff Stanway, Jenny Brennan, Jeremy Chen, Johan Ferret, Justin Chiu, Justin Mao-Jones, Katherine Lee, Kathy Yu, Katie Millican, Lars Lowe Sjoesund, Lisa Lee, Lucas Dixon, Rachel Reid, Maciej Mikula, Mateo Wirth, Michael Sharman, Nikolai Chinaev, Nithum Thain, Olivier Bachem, Oscar Chang, Oscar Wahltinez, Paige Bailey, Paul Michel, Petko Yotov, Rahma Chaabouni, Ramona Comanescu, Reena Jana, Rohan Anil, Ross McIlroy, Ruibo Liu, Ryan Mullins, Samuel L Smith, Sebastian Borgeaud, Sertan Girgin, Sholto Douglas, Shree Pandya, Siamak Shakeri, Soham De, Ted Klimenko, Tom Hennigan, Vlad Feinberg, Wojciech Stokowiec, Yu hui Chen, Zafarali Ahmed, Zhitao Gong, Tris Warkentin, Ludovic Peran, Minh Giang, Clément Farabet, Oriol Vinyals, Jeff Dean, Koray Kavukcuoglu, Demis Hassabis, Zoubin Ghahramani, Douglas Eck, Joelle Barral, Fernando Pereira, Eli Collins, Armand Joulin, Noah Fiedel, Evan Senter, Alek Andreev, and Kathleen Kenealy. Gemma: Open models based on gemini research and technology, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.776, + 0.482, + 0.844 + ], + "angle": 0, + "content": "[80] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[81] Elena Voita, David Talbot, Fedor Moiseev, Rico Sennrich, and Ivan Titov. Analyzing multi-head self-attention: Specialized heads do the heavy lifting, the rest can be pruned. In Proceedings of the 57th Annual Meeting of the Asso" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.554, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "citation for Computational Linguistics, pages 5797-5808, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.122, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[82] Zhong W., Feng X., Zhao L., Li Q., Huang L., Gu Y., Ma W., Xu Y., and Qin B. Investigating and mitigating the multimodal hallucination snowballing in large vision-language models. arXiv preprint arXiv:2407.00569, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.179, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[83] Guankun Wang, Long Bai, Wan Jun Nah, Jie Wang, Zhaoxi Zhang, Zhen Chen, Jinlin Wu, Mobarakol Islam, Hongbin Liu, and Hongliang Ren. Surgical-lvIm: Learning to adapt large vision-language model for grounded visual question answering in robotic surgery. arXiv preprint arXiv:2405.10948, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.263, + 0.904, + 0.318 + ], + "angle": 0, + "content": "[84] P Wang, S Bai, S Tan, S Wang, Z Fan, J Bai, K Chen, X Liu, J Wang, W Ge, and Y Fan. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.321, + 0.904, + 0.36 + ], + "angle": 0, + "content": "[85] Xintong Wang, Jingheng Pan, Liang Ding, and Chris Biemann. Mitigating hallucinations in large vision-language models with instruction contrastive decoding, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.363, + 0.904, + 0.417 + ], + "angle": 0, + "content": "[86] Sangmin Woo, Jaehyuk Jang, Donguk Kim, Yubin Choi, and Changick Kim. Ritual: Random image transformations as a universal anti-hallucination lever in lvlms, 2024. http://arxiv.org/abs/2405.17821.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.42, + 0.904, + 0.473 + ], + "angle": 0, + "content": "[87] Sangmin Woo, Donguk Kim, Jaehyuk Jang, Yubin Choi, and Changick Kim. Don't miss the forest for the trees: Attentional vision calibration for large vision language models, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.477, + 0.904, + 0.531 + ], + "angle": 0, + "content": "[88] Junfei Wu, Qiang Liu, Ding Wang, Jinghao Zhang, Shu Wu, Liang Wang, and Tieniu Tan. Logical closed loop: Uncovering object hallucinations in large vision-language models, 2024. http://arxiv.org/abs/2402.11622.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.533, + 0.904, + 0.601 + ], + "angle": 0, + "content": "[89] Kai Wu, Boyuan Jiang, Zhengkai Jiang, Qingdong He, Donghao Luo, Shengzhi Wang, Qingwen Liu, and Chengjie Wang. Noiseboost: Alleviating hallucination with noise perturbation for multimodal large language models, 2024. http://arxiv.org/abs/2405.20081.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.604, + 0.904, + 0.644 + ], + "angle": 0, + "content": "[90] Xin Xiao, Bohong Wu, Jiacong Wang, Chunyuan Li, Xun Zhou, and Haoyuan Guo. Seeing the image: Prioritizing visual correlation by contrastive alignment, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.647, + 0.904, + 0.701 + ], + "angle": 0, + "content": "[91] Xiaohan Xu, Chongyang Tao, Tao Shen, Can Xu, Hongbo Xu, Guodong Long, and Jian guang Lou. Re-reading improves reasoning in large language models, 2024. http://arxiv.org/abs/2309.06275.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.704, + 0.905, + 0.77 + ], + "angle": 0, + "content": "[92] L Xue, M Shu, A Awadalla, J Wang, A Yan, S Purushwalkam, H Zhou, V Prabhu, Y Dai, M S Ryoo, and S Kendre. xgen-mm (blip-3): A family of open large multimodal models. arXiv preprint arXiv:2408.08872, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.774, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[93] Onoe Y., Zhang M.J., Choi E., and Durrett G. Entity cloze by date: What lms know about unseen entities. arXiv preprint arXiv:2205.02832, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.817, + 0.904, + 0.871 + ], + "angle": 0, + "content": "[94] Park Y., Lee D., Choe J., and Chang B. Convis: Contrastive decoding with hallucination visualization for mitigating hallucinations in multimodal large language models. arXiv preprint arXiv:2408.13906, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.874, + 0.904, + 0.901 + ], + "angle": 0, + "content": "[95] Zhenjie Yang, Xiaosong Jia, Hongyang Li, and Junchi Yan. Llm4drive: A survey of large language models for au" + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4220" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.133, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "tonomous driving. In NeurIPS 2024 Workshop on OpenWorld Agents, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.122, + 0.483, + 0.178 + ], + "angle": 0, + "content": "[96] Shukang Yin, Chaoyou Fu, Sirui Zhao, Tong Xu, Hao Wang, Dianbo Sui, Yunhang Shen, Ke Li, Xing Sun, and Enhong Chen. Woodpecker: Hallucination correction for multimodal large language models. arXiv preprint, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.179, + 0.483, + 0.261 + ], + "angle": 0, + "content": "[97] Qifan Yu, Juncheng Li, Longhui Wei, Liang Pang, Wentao Ye, Bosheng Qin, Siliang Tang, Qi Tian, and Yueting Zhuang. Hallucidctor: Mitigating hallucinatory toxicity in visual instruction data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12944–12953, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.264, + 0.483, + 0.346 + ], + "angle": 0, + "content": "[98] Tianyu Yu, Yuan Yao, Haoye Zhang, Taiwen He, Yifeng Han, Ganqu Cui, Jinyi Hu, et al. Rlhf-v: Towards trustworthy mllms via behavior alignment from fine-grained correctional human feedback. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13807-13816, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.349, + 0.483, + 0.458 + ], + "angle": 0, + "content": "[99] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, Cong Wei, Botao Yu, Ruibin Yuan, Renliang Sun, Ming Yin, Boyuan Zheng, Zhenzhu Yang, Yibo Liu, Wenhao Huang, Huan Sun, Yu Su, and Wenhu Chen. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.461, + 0.482, + 0.514 + ], + "angle": 0, + "content": "[100] Zihao Yue, Liang Zhang, and Qin Jin. Less is more: Mitigating multimodal hallucination from an eos decision perspective, 2024. http://arxiv.org/abs/2402.14545.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.517, + 0.482, + 0.558 + ], + "angle": 0, + "content": "[101] Bai Z., Wang P., Xiao T., He T., Han Z., Zhang Z., and Shou MZ. Hallucination of multimodal large language models: A survey. arXiv preprint arXiv:2404.18930, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.56, + 0.482, + 0.613 + ], + "angle": 0, + "content": "[102] Han Z., Bai Z., Mei H., Xu Q., Zhang C., and Shou M.Z. Skip: A simple method to reduce hallucination in large vision-language models. arXiv preprint arXiv:2402.01345, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.617, + 0.482, + 0.657 + ], + "angle": 0, + "content": "[103] Shaolei Zhang, Tian Yu, and Yang Feng. Truthx: Alleviating hallucinations by editing large language models in truthful space, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.66, + 0.482, + 0.713 + ], + "angle": 0, + "content": "[104] Yi-Fan Zhang, Weichen Yu, Qingsong Wen, Xue Wang, Zhang Zhang, Liang Wang, Rong Jin, and Tieniu Tan. Debiasing multimodal large language models. arXiv preprint, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.717, + 0.482, + 0.758 + ], + "angle": 0, + "content": "[105] L Zhao, Y Deng, W Zhang, and Q Gu. Mitigating object hallucination in large vision-language models via classifier-free guidance. 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.76, + 0.482, + 0.827 + ], + "angle": 0, + "content": "[106] Zhiyuan Zhao, Bin Wang, Linke Ouyang, Xiaoyi Dong, Jiaqi Wang, and Conghui He. Beyond hallucinations: Enhancing lvlms through hallucination-aware direct preference optimization. arXiv preprint arXiv:2311.16839, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.831, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[107] Zheng Zhao, Emilio Monti, Jens Lehmann, and Haytham Assem. Enhancing contextual understanding in large language models through contrastive decoding, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[108] Lanyun Zhu, Deyi Ji, Tianrun Chen, Peng Xu, Jieping Ye, and Jun Liu. Ibid: Alleviating hallucinations in large" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.555, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "vision-language models via image-biased decoding. arXiv preprint, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[109] Wenhao Zhu, Sizhe Liu, Shujian Huang, Shuaijie She, Chris Wendler, and Jiajun Chen. Multilingual contrastive decoding via language-agnostic layers skipping, 2024. http://arxiv.org/abs/2407.10795.2,3" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.093, + 0.905, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.956 + ], + "angle": 0, + "content": "4221" + } + ] +] \ No newline at end of file diff --git a/2025/ICT_ Image-Object Cross-Level Trusted Intervention for Mitigating Object Hallucination in Large Vision-Language Models/a90828a4-8b1b-4072-b5af-bed317d47267_origin.pdf b/2025/ICT_ Image-Object Cross-Level Trusted Intervention for Mitigating Object Hallucination in Large Vision-Language Models/a90828a4-8b1b-4072-b5af-bed317d47267_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..37f427db367c9893bbee36c84a3db55e8aaf64d5 --- /dev/null +++ b/2025/ICT_ Image-Object Cross-Level Trusted Intervention for Mitigating Object Hallucination in Large Vision-Language Models/a90828a4-8b1b-4072-b5af-bed317d47267_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5e57f7fbaec5522ac35cfa502697197fb98fd7fd0fbc9cf958aef698f16c9a4a +size 4871906 diff --git a/2025/ICT_ Image-Object Cross-Level Trusted Intervention for Mitigating Object Hallucination in Large Vision-Language Models/full.md b/2025/ICT_ Image-Object Cross-Level Trusted Intervention for Mitigating Object Hallucination in Large Vision-Language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..63d76eb0091b6ac0b033b0bf843dd4d334473dbb --- /dev/null +++ b/2025/ICT_ Image-Object Cross-Level Trusted Intervention for Mitigating Object Hallucination in Large Vision-Language Models/full.md @@ -0,0 +1,361 @@ +# ICT: Image-Object Cross-Level Trusted Intervention for Mitigating Object Hallucination in Large Vision-Language Models + +Junzhe Chen $^{1,2*}$ , Tianshu Zhang $^{1*}$ , Shiyu Huang $^{3}$ , Yuwei Niu $^{4}$ , Linfeng Zhang $^{5}$ , Lijie Wen $^{1\dagger}$ , Xuming Hu $^{2\dagger}$ + +$^{1}$ Tsinghua University, $^{2}$ The Hong Kong University of Science and Technology (Guangzhou), $^{3}$ Zhipu AI, $^{4}$ Chongqing University, $^{5}$ Shanghai Jiao Tong University. + +chenjz24@mails.tsinghua.edu.cn, + +wenlj@tsinghua.edu.cn, xuminghu97@gmail.com, + +# Abstract + +Despite the recent breakthroughs achieved by Large Vision Language Models (LVLMs) in understanding and responding to complex visual-textual contexts, their inherent hallucination tendencies limit their practical application in real-world scenarios that demand high levels of precision. Existing methods typically either fine-tune the LVLMs using additional data, which incurs extra costs in manual annotation and computational resources or perform comparisons at the decoding stage, which may eliminate useful language priors for reasoning while introducing inference time overhead. Therefore, we propose ICT, a lightweight, training-free method that calculates an intervention direction to shift the model's focus towards different levels of visual information, enhancing its attention to high-level and fine-grained visual details. During the forward pass stage, the intervention is applied to the attention heads that encode the overall image information and the fine-grained object details, effectively mitigating the phenomenon of overly language priors, and thereby alleviating hallucinations. Extensive experiments demonstrate that ICT achieves strong performance with a small amount of data and generalizes well across different datasets and models. Our codes are publicly available at:https://github.com/THU-BPM/ICT/. + +# 1. Introduction + +The recent rapid advancements in Large Vision Language Models (LVLMs) have demonstrated significant potential to tackle complex real-world tasks [5, 27, 29, 52, 56]. However, LVLMs frequently generate text outputs that are inconsistent with the visual input, such as incorrectly determining the presence of objects in an image or inaccurately + +![](images/d5ea88bbeb089521af97a2e2f50af36678eb1f491b5330addf3d1535fd0d75e6.jpg) +Figure 1. Comparison between Contrastive Decoding (top) and our proposed ICT (bottom). In the top example, Contrastive Decoding indiscriminately removes both beneficial and detrimental language priors, leading to hallucinations. In contrast, our approach enhances the model's attention to visual details while preserving useful language priors, allowing it to correctly identify and describe objects in the image. + +describing the attributes of those objects [55, 73]. This issue is commonly referred to as the object hallucination phenomenon [55], which significantly limits the applicability of LVLMs in high-stakes scenarios that demand precise accuracy, such as autonomous driving [95] and medical surgery [35, 83]. Previous research suggested that the primary reasons for hallucinations in LVLMs are as follows: 1) Large language models (LLMs), such as Vicuna [12], possess significantly stronger parameters and capabilities compared to visual encoders, such as CLIP [68]. This results in an excessively strong language prior, causing the model to overly rely on language cues while neglecting visual input [24, 25, 40, 43, 55]. 2) Current visual decoders often struggle to accurately capture fine-grained visual semantics, resulting in errors in detailed object attributes in an image, such as color and quantity [4, 9, 13, 55]. + +Based on the two issues outlined above, previous ef + +forts to mitigate the hallucination phenomenon in LVLMs can be categorized into three main approaches: 1) Finetuning with additional data: This approach involves introducing high-quality annotated data to better align the model's behavior with human interpretations, effectively teaching the model to focus more on visual information [23, 46, 52, 106]. However, this method not only requires costly human annotation but also involves updating model parameters, which demands substantial computational resources, thereby limiting its scalability. 2) Perceptual enhancement: This method incorporates additional information, such as depth maps and segmentation maps [34, 38, 106], as auxiliary inputs to assist the visual encoder in capturing more detailed visual features, thereby reducing hallucinations. However, it often requires manual selection of auxiliary features, which limits its generalizability across different tasks. 3) Contrastive decoding: This method alleviates hallucinations without requiring additional training. It induces hallucinations by applying transformations such as blurring, rotation, or cropping to the original visual input. During the decoding stage, tokens associated with these induced hallucinations are penalized, thus mitigating the influence of language priors [10, 14, 36, 43, 64, 70, 82, 87, 109]. However, methods such as VCD often indiscriminately eliminate all language priors, including those that may be beneficial. As shown in Figure 1, the original model recognizes Curry holding a golf club, while utilizing the language prior that identifies Curry as a basketball player. This language prior is valuable for achieving an accurate interpretation. However, through contrastive decoding, this useful prior is also removed, which can inadvertently result in hallucinations. + +To address the challenges of mitigating hallucinations in LVLMs, we propose Image-Object Cross-Level Trusted Intervention (ICT), a training-free, plug-and-play method applicable during the forward pass. Unlike contrastive decoding, our approach does not eliminate language priors to reduce the model's over-reliance on text modality. Instead, it intervenes during the forward pass to enhance the model's focus on both comprehensive visual information and fine-grained object details. We explored the activation patterns of attention heads when the model produces correct versus hallucinated responses. This analysis allowed us to identify activation value deviations that can shift the model from being "untrustworthy" to "trustworthy". According to prior research [8, 19, 81], which has shown that different heads in the multi-head attention mechanism encode information at varying levels of granularity, we train binary classifiers for each head to determine which heads encode overall visual information and which capture detailed visual features. During the forward pass, we adjust the activation values of these heads based on their identified levels, thereby enhancing the model's attention to the relevant visual features and + +reducing the likelihood of hallucinations. As illustrated in Figure 1, after applying ICT, the model is able to focus more on the details within the image, such as identifying the man as Curry, while simultaneously utilizing beneficial language priors (e.g., Curry is a basketball player) to infer and arrive at the correct answer. Since the intervention shift vectors are pre-computed, ICT does not introduce additional delays during the forward pass. + +Our experiments demonstrate that, for both LLaVAv1.5 [52] and Qwen-VL [5], applying ICT results in an average improvement of $6.27\%$ on POPE benchmark and 67.37 points on MME benchmark. Furthermore, ICT exhibits cross-dataset generalization and model-agnostic generalizability. Our contributions can be summarized as: + +- We propose Image-Object Cross-Level Trusted Intervention (ICT), a novel, training-free, plug-and-play method that effectively reduces hallucinations in LVLMs by enhancing focus on both overall visual information and fine-grained object details during the forward pass, without eliminating beneficial language priors. + +- Unlike existing contrastive decoding approaches, we introduce an intervention mechanism that operates during the forward pass rather than the decoding stage, allowing ICT to be orthogonal and complementary to existing decoding strategies while not introducing any additional latency. + +- Extensive experiments on LLaVA-v1.5 and Qwen-VL demonstrate that ICT significantly improves performance on both the POPE and MME benchmarks while maintaining cross-dataset and model-agnostic generalizability. + +# 2. Related Work + +# 2.1. Large Vision-Language Models + +Following the success of LLM [6, 12, 61, 78, 80], researchers have been exploring the multimodal field. Taking advantage of the remarkable capabilities of LLMs, Large Vision-Language Models incorporate visual encoders and feature projectors into powerful LLMs, enabling them to understand and generate content based on visual and text input. Most of these models undergo two training steps, namely the pre-training stage and the fine-tuning stage. The main purpose of the pre-training stage is to align textual and visual features, while the goal of the fine-tuning stage is to bridge further the modality gap between vision and language, as well as lifting their instruction following capabilities and performance on specific downstream tasks. Early attempts, such as FLamingo [2], Gemini [79], and BLIP-2 [44], have already shown promising results. Recent works like LLaVA-v1.5 [54], Qwen2-VL [84], and xgenmm [92] have advanced this field, dramatically improving the abilities of these models. To enhance the consistency between vision representation and language representation, many efforts have been made, including utilizing higher res + +olution vision encoders, transitioning to larger and more capable LLMs, and adapting reinforcement learning techniques such as RLHF [62, 98], among others. Despite the remarkable progress that has been made, LVLMs still suffer from severe hallucinations, which restricts their potential for large-scale applications in real-world scenarios. + +# 2.2. Mitigating Hallucinations in LVLMs + +Many efforts have been made to understand the causes of hallucination [1, 16, 30, 39, 50, 50, 53, 63, 65, 71, 72, 76, 93, 101]. Existing approaches to mitigate hallucinations can be roughly divided into two categories based on the stage they take place. The first category focuses on the training phase. Most of the work falling into this category introduces additional or curated datasets [21, 22, 51, 97], these works usually design a dataset specifically for hallucination-related tasks or improve data cleaning methods [69, 75], or introduce novel training objectives [7, 41, 77]. While effective, these methods typically require extensive training, which is both time-consuming and resource-intensive. The second category focuses on the inferencing phase, usually involves CD-based [47] novel decoding strategies [10, 14, 36, 59, 64, 70, 82, 87, 94, 107, 109]. Another approach is to detect potential hallucinations while generating and correct them [18, 37, 60], among others [45, 86, 100, 103]. In addition, some researchers tackle hallucinations by manipulating attention weight assigned to the image [33, 89, 90, 104, 108], or the prompt-related part of the image [3]. Other works include prompt-based methods [42, 67, 85, 88, 91, 102], utilizing external tools [11, 96, 105] or external knowledge [17, 28, 66] and so on. + +# 3. Task Formulation + +Given an LVLM parameterized by $\theta$ , the model processes a textual input $\pmb{X} = \{x_{1}, x_{2}, \ldots, x_{N_{x}}\}$ and a visual input $\pmb{V} = \{v_{1}, v_{2}, \ldots, v_{N_{v}}\}$ , where $N_{x}$ and $N_{v}$ denote the sequence lengths of the text and visual inputs, respectively. The model then concatenates the textual and visual sequences to form a unified input $\pmb{H} = \text{concat}(\pmb{V}, \pmb{X})$ , which is subsequently passed through $L$ layers of a transformer architecture. At each layer, the concatenated input undergoes multi-head attention, which is computed as: + +$$ +\boldsymbol {H} ^ {(l + 1)} = \boldsymbol {H} ^ {(l)} + \sum_ {n = 1} ^ {N} A t t n _ {n} ^ {(l)} \left(\boldsymbol {H} ^ {(l)}\right) \cdot W _ {o} ^ {(l)}, \tag {1} +$$ + +where $Attn_{n}^{(l)}$ denotes the attention operation of the $n$ -th head at the $l$ -th layer, and $W_{o}^{(l)} \in \mathbb{R}^{d \times Nd}$ is the output projection matrix, with $d$ representing the dimensionality of each attention head and $N$ the number of heads. Subsequently, the model autoregressively predicts the next token + +based on the output from the $L$ -th layer: + +$$ +p \left(y _ {t} \mid y _ {< t}\right) = \operatorname {S o f t m a x} \left(f c \left(\boldsymbol {H} _ {t} ^ {(L)}\right)\right), \tag {2} +$$ + +where $f c(\cdot)$ is an affine layer mapping the $Hd$ -dimensional vector to a probability distribution of vocabulary size $M$ . + +# 4. Methodology + +In this section, we introduce two levels of fine-grained intervention modules. The first module enables the LVLM to focus on the image, thereby reducing over-reliance on language priors. The second module encourages the LVLM to attend more closely to image objects, helping to mitigate the omission of critical objects and reduce hallucinations. + +# 4.1. Intervention at Image-Level + +This module aims to identify the attention heads associated with overall image information and to apply targeted interventions to these heads. This approach enhances the model's focus on visual input while diminishing the influence of language priors. + +As illustrated in Figure 2, consider a batch of image-question pairs, denoted as $\{(X_i,V_i)\}_{i = 1}^B$ , each labeled with the answer "Yes" from the POPE dataset. Here, $X_{i}$ takes the form of "Is there a/an [object] in the image?" For each question, we extract the specified object $O_{i}$ and reformulate the question to $q =$ "What is it in the image?" Then, for each image $V_{i}$ , we progressively add Gaussian noise following the forward diffusion process [26] to obtain the final blurred image $V_{i}^{\prime}$ : + +$$ +\begin{array}{l} q (V _ {i} ^ {(t)} \mid V _ {i} ^ {(t - 1)}) = \mathcal {N} (V _ {i} ^ {(t)}; \sqrt {1 - \beta_ {t}} V _ {i} ^ {(t - 1)}, \beta_ {t} \mathbf {I}), \\ q \left(V _ {i} ^ {\prime} \mid V _ {i}\right) = \prod_ {t = 1} ^ {T} q \left(V _ {i} ^ {(t)} \mid V _ {i} ^ {(t - 1)}\right), \tag {3} \\ \end{array} +$$ + +where $\beta_{t}$ is the noise variance at step $t$ . Finally, we construct a modified dataset $\{(q,V_i,V_i',O_i)\}_{i=1}^B$ to obtain Image-Level intervention vectors. We then concatenate $O_i$ with the question $q$ to form the answer, resulting in data pairs of trusted and untrusted data for the same sample: $(q + O_i,V_i)$ and $(q + O_i,V_i')$ . For each sample, we treat the representation of the last token as the fused representation of multimodal data and extract the attention activation from a total of $N\times L$ heads across the $L$ layers for both trusted and untrusted data. These are denoted as $A_{i} = \{Attn_{n}^{(l)}H_{s}^{(l)}\}_{n=1,l=1}^{N,L}$ and $A_{i}' = \{Attn_{n}^{(l)}H_{s}^{(l)}\}_{n=1,l=1}^{N,L}$ , where $s = N_x + N_v + 1$ . According to the trusted activation $\{A_{i}\}_{i=1}^{B}$ and the untrusted activation $\{A_{i}'\}_{i=1}^{B}$ obtained from all sample pairs, we can calculate the following activation shift vector $\{S_{n}^{(l)}\}_{n=1,l=1}^{N,L}$ that encourages the model to pay more attention to visual information, as follows: + +$$ +\boldsymbol {S} _ {n} ^ {(l)} = \frac {1}{B} \sum_ {i = 1} ^ {B} \left(\boldsymbol {A} _ {i, n} ^ {(l)} - \boldsymbol {A} _ {i, n} ^ {\prime (l)}\right). \tag {4} +$$ + +![](images/609bed94047a25603363ccc606374ecd88d35da5bdb7fac0915277db876cde1b.jpg) +Figure 2. Overview of our proposed ICT method. ICT consists of two levels of intervention: Image-Level and Object-Level. The Image-Level module enhances the model's focus on the overall scene to reduce reliance on language priors, while the Object-Level module focuses on fine-grained details to mitigate hallucinations. We apply targeted activation shifts to selected attention heads identified through binary classifiers trained to distinguish trusted and untrusted data pairs. + +Next, we train a binary classifier $f_{n}^{(l)}(\cdot)$ using $B$ sample pairs for each head to detect which heads encode Image-Level information, specifically those that can better distinguish the differences between pairs of trusted and untrusted samples. We then apply activation interventions to these selected heads: + +$$ +A t t n _ {\mathrm {i m g}} = \left\{A t t n _ {n} ^ {(l)} \mid A t t n _ {n} ^ {(l)} \in T o p K (\text {A c c u r a c y} (f _ {n} ^ {(l)} (\cdot))) \right\}, +$$ + +$$ +\boldsymbol {H} ^ {(l + 1)} = \boldsymbol {H} ^ {(l)} + \sum_ {n = 1} ^ {N} \left(A t t n _ {n} ^ {(l)} \left(\boldsymbol {H} ^ {(l)}\right) + \mathbb {I} _ {\operatorname {i m g}, n} ^ {(l)} \alpha \boldsymbol {S} _ {n} ^ {(l)}\right) \cdot W _ {o} ^ {(l)}, \tag {5} +$$ + +where $K$ is the number of selected intervention heads, $\mathbb{I}_{\mathrm{img},n}^{(l)}$ is an indicator function, which is 1 if $Attn_{n}^{(l)} \in Attn_{\mathrm{img}}$ and 0 otherwise, and $\alpha$ denotes the intensity of the intervention. After using $\{S_n^{(l)}\}_{n=1,l=1}^{N,L}$ to perform Image-Level interventions on heads that encode image information, the model enhances the trustworthiness of the visual level, places greater attention on visual information, thus mitigates the impact of overly strong language priors. + +# 4.2. Intervention at Object-Level + +After enhancing the model's trustworthiness at the Image-Level, a more fine-grained, Object-Level intervention becomes necessary to increase the model's attention to image details, thereby reducing hallucinations caused by the omission of fine details. As shown in Figure 2, similar to Section 4.1, we construct a dataset $\{(q,V_i,V_i^{\prime \prime},O_i)\}_{i = 1}^{B}$ to obtain Object-Level intervention vectors. We use Grounding DINO [58] to identify the area of object $O_{i}$ in image $V_{i}$ . Gaussian noise is then added selectively to this object region, based on Eq. 3, producing a locally blurred image $V_{i}^{\prime \prime}$ . Using both the original image $V_{i}$ and the partially blurred image $V_{i}^{\prime \prime}$ , we construct trusted and untrusted data pairs for each sample: $(q + O_i,V_i)$ and $(q + O_i,V_i^{\prime \prime})$ and analyze + +the attention activation values across $N\times L$ heads, allowing us to compute an Object-Level activation shift vector $\{S_{n,\mathrm{obj}}^{(l)}\}_{n = 1,l = 1}^{N,L}$ + +$$ +\boldsymbol {S} _ {n, \mathrm {o b j}} ^ {(l)} = \frac {1}{B} \sum_ {i = 1} ^ {B} \left(\boldsymbol {A} _ {i, n} ^ {(l)} - \boldsymbol {A} _ {i, n} ^ {\prime \prime (l)}\right). \tag {6} +$$ + +A binary classifier $f_{n,obj}^{(l)}(\cdot)$ is then trained to identify heads that effectively distinguish trusted from untrusted object-focused samples. Interventions are subsequently applied to the selected heads as follows: + +$$ +A t t n _ {\mathrm {o b j}} = \left\{A t t n _ {n} ^ {(l)} \mid A t t n _ {n} ^ {(l)} \in T o p K (\text {A c c u r a c y} (f _ {n, o b j} ^ {(l)} (\cdot))) \right\}, +$$ + +$$ +\boldsymbol {H} ^ {(l + 1)} = \boldsymbol {H} ^ {(l)} + \sum_ {n = 1} ^ {N} \left(A t t n _ {n} ^ {(l)} \left(\boldsymbol {H} ^ {(l)}\right) + \mathbb {I} _ {\mathrm {o b j}, n} ^ {(l)} \beta \boldsymbol {S} _ {\mathrm {o b j}, n} ^ {(l)}\right) \cdot W _ {o} ^ {(l)}, \tag {7} +$$ + +where $K$ is the number of intervention heads, $\mathbb{I}_{\mathrm{obj},n}^{(l)}$ is an indicator function set to 1 if $Attn_{n}^{(l)} \in Attn_{\mathrm{obj}}$ and 0 otherwise, and $\beta$ controls the intervention intensity. + +Finally, we integrate the Image-Level and Object-Level intervention modules to create a unified approach that reinforces the model's focus on both the overall visual context and finer object-specific details as follows: + +$$ +\begin{array}{l} \boldsymbol {H} ^ {(l + 1)} = \boldsymbol {H} ^ {(l)} + \sum_ {n = 1} ^ {N} \left(A t t n _ {n} ^ {(l)} \left(\boldsymbol {H} ^ {(l)}\right) \right. \tag {8} \\ \left. \right.\left. + \mathbb {I} _ {\operatorname {i m g}, n} ^ {(l)} \alpha \boldsymbol {S} _ {n} ^ {(l)} + \mathbb {I} _ {\operatorname {o b j}, n} ^ {(l)} \beta \boldsymbol {S} _ {\operatorname {o b j}, n} ^ {(l)}\right) \cdot W _ {o} ^ {(l)}. \\ \end{array} +$$ + +By applying these interventions in tandem, the model gains a balanced attention mechanism that mitigates excessive reliance on language priors while enhancing sensitivity to essential visual cues across different levels of granularity, thereby alleviating the occurrence of hallucinations. + +# 5. Experiments + +# 5.1. Experimental Setup + +# Datasets and Metrics. + +POPE [48] is a benchmark designed to assess how effectively LVLMs recognize the presence of specific objects in images, thus identifying Object-Level hallucinations. It uses Yes/No questions based on object annotations, with metrics including Accuracy, Precision, Recall, and F1 score. The dataset is balanced, with $50\%$ of queries for existing objects and $50\%$ for non-existing ones, and employs three sampling strategies: random, popular, and adversarial. Drawing from MSCOCO [49], A-OKVQA [74], and GQA [32] datasets, POPE evaluates 27,000 query-answer pairs to gauge model performance. + +MME [20] benchmark is designed to comprehensively evaluate the performance of LVLMs across various dimensions. It includes ten perception-focused tasks and four cognition-related tasks. MME benchmark specifically addresses Object-Level hallucinations through subsets focusing on object existence and count, while attribute-level hallucinations are assessed via subsets related to object position and color. The evaluation metric is Accuracy. + +Baselines. We adopt the widely used LLaVA-v1.5 [52] and Qwen-VL [5] models as our baseline LVLMs. We compared two baselines that eliminate language priors of the LVLMs in the decoding stage to alleviate hallucinations: VCD [43] and Opera [31]. + +Implementation Details. In our experiments, we utilized 1,500 QA pairs with "Yes" responses from the COCO Random subset of the POPE dataset to train the intervention shift vector. Subsequently, we evaluated the ICT approach on two datasets with significant distributional differences: POPE, and MME. This evaluation was conducted to assess the generalizability and robustness of ICT across diverse data distributions. For each attention head, we employed Support Vector Machines (SVMs) [15] as the classifier and performed 2-fold cross-validation to evaluate the classification accuracy. In the experiment, we set $\alpha = \beta$ and determined the optimal values of $\alpha$ , $\beta$ , and $K$ through a grid search. Detailed hyperparameter configurations are provided in Appendix A. All experiments were conducted on a system equipped with $8 \times$ H800 GPUs. + +# 5.2. Main Results + +Results on POPE. Table 1 presents the results of LLaVA-v1.5 and Qwen-VL on nine subsets of the POPE dataset. From comparing these methods, we can derive the following conclusions: 1) Applying ICT results in an average performance improvement on the F1 score for LLaVA-v1.5 and Qwen-VL of $7.09\%$ and $5.44\%$ on 9 subsets respectively. This improvement can be attributed to the fact that ICT does not eliminate language priors, which may provide + +useful information. Instead, it enhances the model's attention to various levels of visual information, thereby reducing the model's tendency to overly rely on language priors and mitigating the occurrence of hallucinations. 2) Individually intervening at both the image level and object level, with average gains of $5.76\%$ and $5.47\%$ in F1 score, respectively. This demonstrates that enhancing LVLMs' attention to image information at various levels can effectively mitigate hallucinations. Moreover, since Object-Level intervention also implicitly directs the model's focus toward broader image information, it achieves a comparatively higher performance gain. 3) The intervention shift vector trained using 1,500 samples from the MSCOCO random subset achieves an average improvement on F1 score of $7.67\%$ on MSCOCO random and an average improvement of $6.09\%$ across the other eight subsets with different distributions. This indicates that ICT generalizes effectively, and the resulting intervention vector captures a general direction towards trustworthiness rather than fitting a specific dataset. + +Results on MME. Figure 3 presents the experimental results on the MME benchmark, where we followed Leng et al. [43] evaluation setup, focusing on the hallucination subset of MME. Compared to the unmodified model, ICT achieved a score improvement of 80.51 on LLaVA-v1.5 and 34.67 on Qwen-VL, surpassing the performance of VCD. Additionally, on non-hallucination tasks such as commonsense QA, ICT demonstrated a 10-point improvement on LLaVA-v1.5 and a 9.5-point improvement on Qwen-VL. These results indicate that ICT not only effectively reduces hallucinations but also enhances general reasoning capabilities across different models. Moreover, since the objects present in the images of the MME dataset are generally prominent, the variant ICT $w/o$ image provides only limited improvements compared to ICT $w/o$ object. This demonstrates that the absence of Image-Level guidance restricts the model's ability to fully leverage global visual context, thereby reducing its effectiveness in scenarios where the dominant objects play a crucial role. + +![](images/83db3db20d370ddf3feed3811612747a8a58a4c6ea5e3e8026b7fa1600971d08.jpg) +Figure 3. Comparison of ICT with baseline methods (Vanilla and VCD) on the MME benchmark. The radar chart illustrates improvements across various evaluation categories, including existence, position, count, color, and commonsense QA (CSQA). + +
DatasetSettingMethodLLaVA-v1.5Qwen-VL
AccuracyF1 ScoreAccuracyF1 Score
COCORandomRegular83.2981.3384.3782.67
VCD87.7387.1688.6387.81
OPERIA89.2088.8187.3186.92
ICT w/o image89.70 (+6.41)89.93 (+8.60)88.76(+4.39)87.84(+5.17)
ICT w/o object88.50 (+5.21)88.81 (+7.48)89.13(+4.76)88.20(+5.63)
ICT90.11 (+6.82)90.03 (+8.70)89.46(+5.09)89.20(+6.63)
PopularRegular81.8880.0684.1382.06
VCD85.3885.0687.1286.40
OPERIA86.6486.6287.4486.68
ICT w/o image87.43 (+5.55)87.09 (+7.03)87.63(+3.50)86.78(+4.72)
ICT w/o object87.30 (+5.42)86.65 (+6.59)87.10(+2.97)86.28(+4.22)
ICT87.50 (+5.62)87.60 (+7.54)88.16(+4.03)87.33(+5.27)
AdversarialRegular78.9677.5782.2680.37
VCD80.8881.3384.2683.90
OPERIA81.2481.3884.7883.45
ICT w/o image84.13 (+5.17)83.31 (+5.74)85.13(+2.87)84.35(+3.98)
ICT w/o object83.63 (+4.67)82.13 (+4.56)85.23(+2.97)84.62(+4.25)
ICT84.43 (+5.47)83.74 (+6.17)84.96(+2.70)84.42(+4.05)
AOKVQARandomRegular83.4582.5686.6785.59
VCD86.1586.3489.2289.01
OPERIA88.0284.5988.1988.43
ICT w/o image89.60 (+6.15)89.75 (+7.19)89.46(+2.79)89.28(+3.69)
ICT w/o object89.00 (+5.55)88.91 (+6.35)89.26(+2.59)88.95(+3.36)
ICT89.20 (+5.75)89.41 (+6.85)89.46(+2.79)89.03(+3.44)
PopularRegular79.9079.5985.5684.63
VCD81.8582.8287.8587.81
OPERIA83.2284.6787.9187.13
ICT w/o image85.36 (+5.46)85.34 (+5.75)87.70(+2.14)87.68(+3.05)
ICT w/o object84.65 (+4.75)84.32 (+4.73)87.90(+2.34)87.48(+2.85)
ICT85.73 (+5.83)85.34 (+5.75)88.13(+3.57)87.83(+3.20)
AdversarialRegular74.0475.1579.5779.50
VCD74.9777.7381.2782.38
OPERIA73.8277.9180.8281.54
ICT w/o image79.26 (+5.22)80.14 (+4.99)81.50(+1.93)82.10(+2.60)
ICT w/o object77.79 (+3.75)79.53 (+4.38)81.40(+1.83)82.18(+2.68)
ICT79.60 (+5.56)80.43 (+5.28)81.94(+2.37)82.44(+2.94)
GQARandomRegular83.7382.9580.9779.01
VCD86.6586.9985.5985.33
OPERIA88.1388.9186.0285.29
ICT w/o image89.03 (+5.30)88.99 (+6.04)85.20(+4.23)85.90(+6.89)
ICT w/o object88.97 (+5.24)89.30 (+6.35)87.20(+6.23)86.80(+7.79)
ICT89.60 (+5.87)89.44 (+6.49)86.38(+5.41)86.96(+7.95)
PopularRegular78.1778.3775.9974.84
VCD80.7382.2481.8382.23
OPERIA79.2782.1181.9782.12
ICT w/o image84.23 (+6.06)83.92 (+5.55)81.43(+5.44)80.86(+6.02)
ICT w/o object84.70 (+6.53)84.45 (+6.08)83.93(+7.94)82.95(+8.11)
ICT84.70 (+6.53)84.78 (+6.41)82.63(+6.64)82.22(+7.38)
AdversarialRegular75.0876.0675.4674.33
VCD76.0978.7880.0180.75
OPERIA75.0078.7180.2480.64
ICT w/o image81.92 (+6.84)82.11 (+6.05)80.03(+4.57)79.90(+5.47)
ICT w/o object81.65 (+6.57)82.00 (+5.94)81.60(+6.14)81.53(+7.10)
ICT81.50 (+6.42)82.27 (+6.21)80.83(+5.37)80.60(+6.17)
+ +Table 1. Main results on POPE tasks. We evaluate the accuracy of various LVLMs on the POPE task across the MSCOCO, A-OKVQA, and GQA datasets. Regular represents the setting where direct sampling is applied. ICT $w/o$ image and ICT $w/o$ object correspond to the exclusion of image-level and object-level interventions, respectively. The bold and underlined values indicate the highest and second-highest metrics under each setting, respectively. + +# 6. Analysis and Discussions + +# 6.1. Inference Latency across Different Methods + +In Table 2, we compare the inference efficiency of two different representative approaches with our proposed method. Notably, while the CD-based methods result in a significant increase in inference time (2.1 to 3.6 times slower), our ICT incurs virtually no impact on inference efficiency. We attribute this to the fact that both VCD and OPERA either require multiple inference runs or involve substantial additional computations during inference. In contrast, ICT introduces no extra computational cost at inference, further highlighting the advantages of our method. + +
Method20-Token Len50-Token Len80-Token Len
LLaVA-v1.5405.3 ↑×1.0934.6 ↑×1.01440.0 ↑×1.0
+ VCD988.9 ↑ × 2.42031.7 ↑ × 2.23077.5 ↑ × 2.1
+ OPERA1371.7 ↑ × 3.43294.1 ↑ × 3.55717.3 ↑ × 3.6
+ ICT415.9 ↑ × 1.0931.9 ↑ × 1.01485.5 ↑ × 1.0
+ +Table 2. Comparison of the efficiency of different methods in generating tokens of varying lengths on an NVIDIA H800 GPU. Inference times are recorded in milliseconds. + +# 6.2. Generalizability between Different Models + +To verify that the derived activation shift vectors to some extent represent the model's alignment with authenticity, we explored whether these shifts are transferable to other models. Specifically, we applied the activation shift vectors obtained from LLaVA-v1.5 on the COCO random subset to intervene in Qwen-VL. We then evaluated the generalizability of this intervention on the GQA dataset, which has a distribution significantly different from COCO. The results in Table 3 demonstrate that ICT-LLaVA-v1.5 achieves an average improvement of $4.62\%$ in F1 score and $4.64\%$ in accuracy compared to the unmodified model. Additionally, we employed t-SNE to visualize the offset vectors across different layers for LLaVA-v1.5 and Qwen-VL. Figure 4 shows the Object-Level and Image-Level shift vectors of multi-head attention in the 16th and 18th layers for both models. From the figure, we can observe that in layer 16, the Image-Level vectors (blue and yellow) are relatively close to each other, indicating a degree of similarity in encoding Image-Level information between the models. However, the Object-Level vectors (red and green) are more distinct, reflecting model-specific characteristics in encoding fine-grained object details. In layer 18, the Image-Level and Object-Level vectors for both models exhibit high similarity, indicating that the resulting activation shift vectors represent universal shift vectors that guide the model to pay more attention to visual information. + +# 6.3. Impact of Hyperparameters on Performance + +Our proposed ICT primarily relies on two key hyperparameters: the intervention intensity $\alpha$ and the number of heads $K$ involved in the intervention. To investigate the + +![](images/12ba79e155dd7a4d0e372b6e33b40098d7e630c396520c34f5c5c0ab79bfac0c.jpg) +Figure 4. t-SNE visualization of Object-Level and Image-Level offset vectors for LLaVA-v1.5 and Qwen-VL at layers 16 and 18. Impact of these hyperparameters on performance, we fixed one of the parameters and calculated the average F1 score for LLaVA-v1.5 and Qwen-VL on the POPE COCO random subset. The results are presented in Figure 6, we can observe that when $\alpha$ is negative—indicating that reverse intervention reduces the model's attention to visual information—the model's performance declines. When $\alpha$ is too small, the model does not receive sufficient intervention, resulting in suboptimal outcomes. Conversely, when $\alpha$ is excessively large, the intervention becomes too strong, disrupting the model's foundational capabilities and leading to a decrease in performance. For the hyperparameter $K$ , we find that when $K$ is too small, certain attention heads that encode relevant visual information are not adequately intervened upon, resulting in suboptimal performance. Conversely, when $K$ is too large, attention heads that encode irrelevant information are unnecessarily interfered with, which leads to performance degradation. + +# 6.4. Analysis of Attention Heads for Visual Information Encoding + +In Figure 7, we present the classification accuracy of the binary classifiers for Image-Level and Object-Level features across the 1,024 multi-head attention heads in the 32 layers of LLaVA-v1.5. This analysis helps identify which heads encode overall visual information and which heads capture fine-grained visual details. From the figure, we can observe + +
SettingMethodAccuracyF1 Score
RandomRegular80.9779.01
VCD85.5985.33
ICT86.3886.96
ICT-LLaVA-v1.585.1083.27
PopularRegular75.9974.84
VCD81.8382.23
ICT82.6382.22
ICT-LLaVA-v1.581.5080.10
AdversarialRegular75.4674.33
VCD80.0180.75
ICT80.8380.60
ICT-LLaVA-v1.579.7378.68
+ +Table 3. Cross-model generalization of ICT. Performance of Qwen-VL on GQA after applying activation shift from LLaVA-v1.5. ICT-LLaVA-v1.5 refers to the results obtained by applying LLaVA-v1.5's activation shift vectors to intervene in Qwen-VL. + +![](images/061fcca9a3ee2c52e18822df4fdfafda5ac52b8f646caf5be60c6a3657ab698f.jpg) +VCD: Yes, there is. +ICT: No, there isn't. +Question: How many uncut fruits are in the image? + +![](images/0a69dc51e3541fb53d1e1ad5e1efc87d97a10cfbe2cf9560213e69c063dd809a.jpg) +Question: Is there a horse in the image? +POPE + +![](images/44463fb48a92b5fe147d9e585a798d28a936dfd3835b536e7f985137d1e2c41c.jpg) +VCD: There are two uncut fruits in the image. +ICT: There are four uncut fruits in the image. +Figure 5. Case Study and Error Analysis of ICT. The vertical axis in the right section of the figure represents the attention ratio of each question text token to the visual tokens. + +![](images/d00f3d20228a6c7d87d26059a11fbdfab60dd6780df7a954d3f4ac14926768cc.jpg) +LLaVA-Bench + +![](images/b1db38f81147a4e3fbbc8512ae0b974d3abe3653d47b72ee36fedf81b1ae1c8b.jpg) +Figure 6. Effect of hyperparameters $\alpha$ (left) and $K$ (right) on the F1 score for LLaVA-v1.5 and Qwen-VL models on the POPE COCO random subset. + +![](images/8158a5900d2bc9350a72d94b857f4a48acf733e6a778f887c09e0b05cc429b16.jpg) + +that the attention heads encoding overall visual information are predominantly located in the earlier layers of the model, such as layers 4 and 10. In contrast, the attention heads that capture fine-grained visual details are more concentrated in the later layers, such as layers 20 and 22. + +# 6.5. Case Study and Error Analysis + +In Figure 5, we present a case study on the POPE and LLaVA-Bench datasets, visualizing the attention ratio of each question text token to the visual tokens. As illustrated, after applying ICT, the model allocates a higher proportion of attention to visual tokens, especially to object tokens relevant to the question (e.g., "horse" and "fruits"). By prior- + +![](images/7dbdda3e10270cab46ddf2f6629e41aba07c1ce1f8ba4279fa920d77be7fcb7d.jpg) +Figure 7. Heatmap of classification accuracy for Image-Level (left) and Object-Level (right) features of LLaVA-v1.5. + +itizing visual information, ICT correctly identifies the absence of a horse in the image, whereas VCD erroneously concludes that a horse is present due to insufficient attention to visual cues. However, when asked, "How many uncut fruits are in the picture?", VCD incorrectly answered "two" due to a lack of focus on visual details. Although ICT correctly identifies that there are a total of four fruits in the image, this question requires not only attention to visual content but also reasoning within the text modality. The model needs to not only recognize the overall count of fruits but also focus on the attribute "uncut". Due to its failure to incorporate this information, ICT outputs a wrong answer. + +Furthermore, when the text becomes more complex, such as in "How many uncut fruits are there in the image?", the increased linguistic complexity demands greater attention from the model. This, in turn, limits the extent to which attention can be allocated to the visual domain. In contrast, function words like "is" and "there" are less complex and require minimal processing, allowing the model to shift more focus toward visual tokens. As a result, this redistribution of attention can lead to a more noticeable shift in how the model processes visual and textual information. + +# 7. Conclusion and Limitations + +In this paper, we mitigate the issue of object hallucinations in LVLMs by proposing Image-Object Cross-Level Trusted Intervention (ICT), a training-free, plug-and-play method that enhances the model's focus on both Image-Level and Object-Level visual information during the forward pass. Our experiments show that ICT significantly reduces hallucinations while also improving general reasoning capabilities. The results confirm that ICT effectively mitigates excessive reliance on language priors, enhancing both accuracy and robustness in diverse visual contexts. + +# Acknowledgement + +This work was supported by the National Key Research and Development Program of China (Grant No. 2024YFB3309702); Guangdong Provincial Department of Education Project (Grant No.2024KQNCX028); CAAI-Ant Group Research Fund; Scientific Research Projects for the Higher-educational Institutions (Grant No.2024312096), Education Bureau of Guangzhou Municipality; Guangzhou-HKUST(GZ) Joint Funding Program (Grant No.2025A03J3957), Education Bureau of Guangzhou Municipality. + +# References + +[1] Renat Aksitov, Chung-Ching Chang, David Reitter, Siamak Shakeri, and Yunhsuan Sung. Characterizing attribution and fluency tradeoffs for retrieval-augmented large language models, 2023. 3 +[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: A visual language model for few-shot learning. In Advances in Neural Information Processing Systems, pages 23716-23736, 2022. 2 +[3] Wenbin An, Feng Tian, Sicong Leng, Jiahao Nie, Haonan Lin, QianYing Wang, Guang Dai, Ping Chen, and Shijian Lu. Agla: Mitigating object hallucinations in large vision-language models with assembly of global and local attention. arXiv preprint, 2024. 3 +[4] Wenbin An, Feng Tian, Sicong Leng, Jiahao Nie, Haonan Lin, QianYing Wang, Guang Dai, Ping Chen, and Shijian Lu. Agla: Mitigating object hallucinations in large vision-language models with assembly of global and local attention. arXiv preprint arXiv:2406.12718, 2024. 1 +[5] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. 2023. 1, 2, 5 +[6] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D. Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In Advances in Neural Information Processing Systems, pages 1877-1901, 2020. 2 +[7] Beitao Chen, Xinyu Lyu, Lianli Gao, Jingkuan Song, and Heng Tao Shen. Alleviating hallucinations in large vision-language models through hallucination-induced optimization, 2024. 3 +[8] Haozhe Chen, Carl Vondrick, and Chengzhi Mao. Selfie: Self-interpretation of large language model embeddings, 2024. 2 +[9] Zhaorun Chen, Zhuokai Zhao, Hongyin Luo, Huaxiu Yao, Bo Li, and Jiawei Zhou. Halc: Object hallucination reduction via adaptive focal-contrast decoding. In *Forty-first International Conference on Machine Learning*. 1 + +[10] Zhaorun Chen, Zhuokai Zhao, Hongyin Luo, Huaxiu Yao, Bo Li, and Jiawei Zhou. Halc: Object hallucination reduction via adaptive focal-contrast decoding, 2024. http://arxiv.org/abs/2403.00425.2,3 +[11] I.-Chun Chern, Steffi Chern, Shiqi Chen, Weizhe Yuan, Kehua Feng, Chunting Zhou, Junxian He, Graham Neubig, and Pengfei Liu. Factool: Factuality detection in generative ai - a tool augmented framework for multi-task and multi-domain scenarios, 2023. http://arxiv.org/abs/2307.13528.3 +[12] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with $90\%$ * chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2(3):6, 2023. 1, 2 +[13] Jaemin Cho, Seunghyun Yoon, Ajinkya Kale, Franck Dernoncourt, Trung Bui, and Mohit Bansal. Fine-grained image captioning with clip reward. In Findings of the Association for Computational Linguistics: NAACL 2022, pages 517-527, 2022. 1 +[14] Yung-Sung Chuang, Yujia Xie, Hongyin Luo, Yoon Kim, James Glass, and Pengcheng He. Dola: Decoding by contrasting layers improves factuality in large language models, 2024. http://arxiv.org/abs/2309.03883.2, 3 +[15] Corinna Cortes. Support-vector networks. Machine Learning, 1995. 5 +[16] Li D., Rawat A.S., Zaheer M., Wang X., Lukasik M., Veit A., Yu F., and Kumar S. Large language models with controllable working memory. arXiv preprint arXiv:2211.05110, 2022. 3 +[17] Hanxing Ding, Liang Pang, Zihao Wei, Huawei Shen, and Xueqi Cheng. Retrieve only when it needs: Adaptive retrieval augmentation for hallucination mitigation in large language models, 2024. http://arxiv.org/abs/2402.10612.3 +[18] Sebastian Farquhar, Jannik Kossen, Lorenz Kuhn, and Yarin Gal. Detecting hallucinations in large language models using semantic entropy. Nature, 630(8017):625-630, 2024. 3 +[19] Javier Ferrando, Gabriele Sarti, Arianna Bisazza, and Marta R. Costa-jussà. A primer on the inner workings of transformer-based language models, 2024. 2 +[20] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023.5 +[21] Suriya Gunasekar, Yi Zhang, Jyoti Aneja, Cao César Teodoro Mendes, Allie Del Giorno, Sivakanth Gopi, Mojan Javaheripi, Piero Kauffmann, Gustavo de Rosa, Olli Saarikivi, Adil Salim, Shital Shah, Harkirat Singh Behl, Xin Wang, Sébastien Bubeck, Ronen Eldan, Adam Tauman Kalai, Yin Tat Lee, and Yuanzhi Li. Textbooks are all you need, 2023. 3 +[22] Anisha Gunjal, Jihan Yin, and Erhan Bas. Detecting and preventing hallucinations in large vision language models. + +In Proceedings of the AAAI Conference on Artificial Intelligence, pages 18135-18143, 2024. 3 +[23] Anisha Gunjal, Jihan Yin, and Erhan Bas. Detecting and preventing hallucinations in large vision language models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 18135-18143, 2024. 2 +[24] Vipul Gupta, Zhuowan Li, Adam Kortylewski, Chenyu Zhang, Yingwei Li, and Alan Yuille. Swapmix: Diagnosing and regularizing the over-reliance on visual context in visual question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5078-5088, 2022. 1 +[25] Yudong Han, Liqiang Nie, Jianhua Yin, Jianlong Wu, and Yan Yan. Visual perturbation-aware collaborative learning for overcoming the language prior problem. arXiv preprint arXiv:2207.11850, 2022. 1 +[26] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3 +[27] Wenyi Hong, Weihan Wang, Ming Ding, Wenmeng Yu, Qingsong Lv, Yan Wang, Yean Cheng, Shiyu Huang, Junhui Ji, Zhao Xue, et al. Cogvlm2: Visual language models for image and video understanding. arXiv preprint arXiv:2408.16500, 2024. 1 +[28] Haichuan Hu, Yuhan Sun, and Quanjun Zhang. Lrp4rag: Detecting hallucinations in retrieval-augmented generation via layer-wise relevance propagation, 2024. http:// arxiv.org/abs/2408.15533.3 +[29] Yingdong Hu, Fanqi Lin, Tong Zhang, Li Yi, and Yang Gao. Look before you leap: Unveiling the power of gpt-4v in robotic vision-language planning. arXiv preprint arXiv:2311.17842, 2023. 1 +[30] Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, and Ting Liu. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions, 2023. 3 +[31] Qidong Huang, Xiaoyi Dong, Pan Zhang, Bin Wang, Conghui He, Jiaqi Wang, Dahua Lin, Weiming Zhang, and Nenghai Yu. Opera: Alleviating hallucination in multimodal large language models via over-trust penalty and retrospection-allocation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13418-13427, 2024. 5 +[32] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 5 +[33] Fushuo Huo, Wenchao Xu, Zhong Zhang, Haozhao Wang, Zhicheng Chen, and Peilin Zhao. Self-introspective decoding: Alleviating hallucinations for large vision-language models, 2024. 3 +[34] Jitesh Jain, Jianwei Yang, and Humphrey Shi. Vcoder: Versatile vision encoders for multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27992-28002, 2024. 2 + +[35] Juseong Jin and Chang Wook Jeong. Surgical-llava: Toward surgical scenario understanding via large language and vision models. arXiv preprint arXiv:2410.09750, 2024. 1 +[36] Junho Kim, Hyunjun Kim, Yeonju Kim, and Yong Man Ro. Code: Contrasting self-generated description to combat hallucination in large multi-modal models, 2024. 2, 3 +[37] Lorenz Kuhn, Yarin Gal, and Sebastian Farquhar. Semantic uncertainty: Linguistic invariances for uncertainty estimation in natural language generation. arXiv, 2023. 3 +[38] Junlin Lee, Yequan Wang, Jing Li, and Min Zhang. Multimodal reasoning with multimodal knowledge graph. arXiv preprint arXiv:2406.02030, 2024. 2 +[39] Katherine Lee, Daphne Ippolito, Andrew Nystrom, Chiyuan Zhang, Douglas Eck, Chris Callison-Burch, and Nicholas Carlini. Deduplicating training data makes language models better. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8424-8445, 2022. 3 +[40] Kang-il Lee, Minbeom Kim, Seunghyun Yoon, Minsung Kim, Dongryeol Lee, Hyukhun Koh, and Kyomin Jung. Vlind-bench: Measuring language priors in large vision-language models. arXiv preprint arXiv:2406.08702, 2024. 1 +[41] Nayeon Lee, Wei Ping, Peng Xu, Mostofa Patwary, Pascale N Fung, Mohammad Shoeybi, and Bryan Catanzaro. Factuality enhanced language models for open-ended text generation. In Advances in Neural Information Processing Systems, pages 34586-34599, 2022. 3 +[42] Seongyun Lee, Sue Hyun Park, Yongrae Jo, and Minjoon Seo. Volcano: Mitigating multimodal hallucination through self-feedback guided revision. arXiv, 2024. 3 +[43] Sicong Leng, Hang Zhang, Guanzheng Chen, Xin Li, Shijian Lu, Chunyan Miao, and Lidong Bing. Mitigating object hallucinations in large vision-language models through visual contrastive decoding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13872-13882, 2024. 1, 2, 5 +[44] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In Proceedings of the International Conference on Machine Learning, pages 19730-19742. PMLR, 2023. 2 +[45] Kenneth Li, Oam Patel, Fernanda Viégas, Hanspeter Pfister, and Martin Wattenberg. Inference-time intervention: Eliciting truthful answers from a language model. Advances in Neural Information Processing Systems, 36, 2024. 3 +[46] Lei Li, Zhihui Xie, Mukai Li, Shunian Chen, Peiyi Wang, Liang Chen, Yazheng Yang, Benyou Wang, Lingpeng Kong, and Qi Liu. Vlfeedback: A large-scale ai feedback dataset for large vision-language models alignment. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 6227-6246, 2024. 2 +[47] Xiang Lisa Li, Ari Holtzman, Daniel Fried, Percy Liang, Jason Eisner, Tatsunori Hashimoto, Luke Zettlemoyer, and Mike Lewis. Contrastive decoding: Open-ended text generation as optimization. arXiv preprint, 2023. 3 + +[48] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. In The 2023 Conference on Empirical Methods in Natural Language Processing. 5 +[49] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5 +[50] Bingbin Liu, Jordan T. Ash, Surbhi Goel, Akshay Krishnamurthy, and Cyril Zhang. Exposing attention glitches with flip-flop language modeling, 2023. 3 +[51] Fuxiao Liu, Kevin Lin, Linjie Li, Jianfeng Wang, Yaser Ya-coob, and Lijuan Wang. Mitigating hallucination in large multi-modal models via robust instruction tuning. In The Twelfth International Conference on Learning Representations, 2023. 3 +[52] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 1, 2, 5 +[53] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in Neural Information Processing Systems, 36, 2024. 3 +[54] Hanchao Liu, Wenyuan Xue, Yifei Chen, Dapeng Chen, Xiutian Zhao, Ke Wang, Liping Hou, Rongjun Li, and Wei Peng. A survey on hallucination in large vision-language models. arXiv preprint, 2024. 2 +[55] Hanchao Liu, Wenyuan Xue, Yifei Chen, Dapeng Chen, Xiutian Zhao, Ke Wang, Liping Hou, Rongjun Li, and Wei Peng. A survey on hallucination in large vision-language models. arXiv preprint arXiv:2402.00253, 2024. 1 +[56] Jiajun Liu, Yibing Wang, Hanghang Ma, Xiaoping Wu, Xiaqi Ma, Xiaoming Wei, Jianbin Jiao, Enhua Wu, and Jie Hu. Kangaroo: A powerful video-language model supporting long-context video input. arXiv preprint arXiv:2408.15542, 2024. 1 +[57] Jiazhen Liu, Yuhan Fu, Ruobing Xie, Runquan Xie, Xingwu Sun, Fengzong Lian, Zhanhui Kang, and Xirong Li. PhD: A chatgpt-prompted visual hallucination evaluation dataset, 2025. 2 +[58] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 4 +[59] Shi Liu, Kecheng Zheng, and Wei Chen. Paying more attention to image: A training-free method for alleviating hallucination in lvlms, 2024. 3 +[60] Alexander Nikitin, Jannik Kossen, Yarin Gal, and Pekka Marttinen. Kernel language entropy: Fine-grained uncertainty quantification for llms from semantic similarities. arXiv, 2024. 3 +[61] OpenAI. GPT-4. https://openai.com/gpt-4, 2023.2 +[62] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, et al. + +Training language models to follow instructions with human feedback. In Advances in Neural Information Processing Systems, pages 27730-27744, 2022. 3 +[63] Amandalynne Paullada, Inioluwa Deborah Raji, Emily M. Bender, Emily Denton, and Alex Hanna. Data and its (dis)contents: A survey of dataset development and use in machine learning research. *Patterns*, 2(11):100336, 2021. 3 +[64] Phuc Phan, Hieu Tran, and Long Phan. Distillation contrastive decoding: Improving llms reasoning with contrastive decoding and distillation, 2024. 2, 3 +[65] Venkit P.N., Gautam S., Panchanadikar R., Huang T.H., and Wilson S. Nationality bias in text generation. arXiv preprint arXiv:2302.02463, 2023. 3 +[66] Xiaoye Qu, Qiyuan Chen, Wei Wei, Jishuo Sun, and Jianfeng Dong. Alleviating hallucination in large vision-language models with active retrieval augmentation, 2024. http://arxiv.org/abs/2408.00555.3 +[67] Xiaoye Qu, Jiashuo Sun, Wei Wei, and Yu Cheng. Look, compare, decide: Alleviating hallucination in large vision-language models via multi-view multi-path reasoning, 2024. http://arxiv.org/abs/2408.17150.3 +[68] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1 +[69] Vikas Raunak, Arul Menezes, and Marcin Junczys-Dowmunt. The curious case of hallucinations in neural machine translation. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1172-1183, 2021. 3 +[70] Kim S., Cho B., Bae S., Ahn S., and Yun S.Y. Vacode: Visual augmented contrastive decoding. arXiv preprint arXiv:2408.05337, 2024. 2, 3 +[71] Lin S., Hilton J., and Evans O. Truthfulqa: Measuring how models mimic human falsehoods. arXiv preprint arXiv:2109.07958, 2021. 3 +[72] Yin S., Fu C., Zhao S., Li K., Sun X., Xu T., and Chen E. A survey on multimodal large language models. arXiv preprint arXiv:2306.13549, 2023. 3 +[73] Pranab Sahoo, Prakash Meharia, Akash Ghosh, Sriparna Saha, Vinija Jain, and Aman Chadha. A comprehensive survey of hallucination in large language, image, video and audio foundation models. Findings of the Association for Computational Linguistics: EMNLP 2024, pages 11709-11724, 2024. 1 +[74] Dustin Schwenk, Apoorv Khandelwal, Christopher Clark, Kenneth Marino, and Roozbeh Mottaghi. A-okvqa: A benchmark for visual question answering using world knowledge. In European conference on computer vision, pages 146–162. Springer, 2022. 5 +[75] Lei Shen, Haolan Zhan, Xin Shen, Hongshen Chen, Xiaofang Zhao, and Xiaodan Zhu. Identifying untrustworthy samples: Data filtering for open-domain dialogues with + +bayesian optimization. In Proceedings of the 30th ACM International Conference on Information & Knowledge Management, pages 1598-1608, 2021. 3 +[76] Weijia Shi, Xiaochuang Han, Mike Lewis, Yulia Tsvetkov, Luke Zettlemoyer, and Scott Wen tau Yih. Trusting your evidence: Hallucinate less with context-aware decoding, 2023. 3 +[77] Weijia Shi, Sewon Min, Maria Lomeli, Chunting Zhou, Margaret Li, Rich James, Xi Victoria Lin, Noah A. Smith, Luke Zettlemoyer, Scott Yih, and Mike Lewis. In-context pretraining: Language modeling beyond document boundaries, 2023. 3 +[78] GLM T., Zeng A., Xu B., Wang B., Zhang C., Yin D., Rojas D., Feng G., Zhao H., Lai H., and Yu H. Chatglm: A family of large language models from glm-130b to glm-4 all tools. arXiv preprint arXiv:2406.12793, 2024. 2 +[79] Gemma Team, Thomas Mesnard, Cassidy Hardin, Robert Dadashi, Surya Bhupatiraju, Shreya Pathak, Laurent Sifre, Morgane Rivière, Mihir Sanjay Kale, Juliette Love, Pouya Tafti, Léonard Hussenot, Pier Giuseppe Sessa, Aakanksha Chowdhery, Adam Roberts, Aditya Barua, Alex Botev, Alex Castro-Ros, Ambrose Slone, Amélie Héliou, Andrea Tacchetti, Anna Bulanova, Antonia Paterson, Beth Tsai, Bobak Shahriari, Charline Le Lan, Christopher A. Choquette-Choo, Clément Crepy, Daniel Cer, Daphne Ippolito, David Reid, Elena Buchatskaya, Eric Ni, Eric Noland, Geng Yan, George Tucker, George-Christian Muraru, Grigory Rozhdestvenskiy, Henryk Michalewski, Ian Tenney, Ivan Grishchenko, Jacob Austin, James Keeling, Jane Labanowski, Jean-Baptiste Lespiau, Jeff Stanway, Jenny Brennan, Jeremy Chen, Johan Ferret, Justin Chiu, Justin Mao-Jones, Katherine Lee, Kathy Yu, Katie Millican, Lars Lowe Sjoesund, Lisa Lee, Lucas Dixon, Rachel Reid, Maciej Mikula, Mateo Wirth, Michael Sharman, Nikolai Chinaev, Nithum Thain, Olivier Bachem, Oscar Chang, Oscar Wahltinez, Paige Bailey, Paul Michel, Petko Yotov, Rahma Chaabouni, Ramona Comanescu, Reena Jana, Rohan Anil, Ross McIlroy, Ruibo Liu, Ryan Mullins, Samuel L Smith, Sebastian Borgeaud, Sertan Girgin, Sholto Douglas, Shree Pandya, Siamak Shakeri, Soham De, Ted Klimenko, Tom Hennigan, Vlad Feinberg, Wojciech Stokowiec, Yu hui Chen, Zafarali Ahmed, Zhitao Gong, Tris Warkentin, Ludovic Peran, Minh Giang, Clément Farabet, Oriol Vinyals, Jeff Dean, Koray Kavukcuoglu, Demis Hassabis, Zoubin Ghahramani, Douglas Eck, Joelle Barral, Fernando Pereira, Eli Collins, Armand Joulin, Noah Fiedel, Evan Senter, Alek Andreev, and Kathleen Kenealy. Gemma: Open models based on gemini research and technology, 2024. 2 +[80] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint, 2023. 2 +[81] Elena Voita, David Talbot, Fedor Moiseev, Rico Sennrich, and Ivan Titov. Analyzing multi-head self-attention: Specialized heads do the heavy lifting, the rest can be pruned. In Proceedings of the 57th Annual Meeting of the Asso + +citation for Computational Linguistics, pages 5797-5808, 2019. 2 +[82] Zhong W., Feng X., Zhao L., Li Q., Huang L., Gu Y., Ma W., Xu Y., and Qin B. Investigating and mitigating the multimodal hallucination snowballing in large vision-language models. arXiv preprint arXiv:2407.00569, 2024. 2, 3 +[83] Guankun Wang, Long Bai, Wan Jun Nah, Jie Wang, Zhaoxi Zhang, Zhen Chen, Jinlin Wu, Mobarakol Islam, Hongbin Liu, and Hongliang Ren. Surgical-lvIm: Learning to adapt large vision-language model for grounded visual question answering in robotic surgery. arXiv preprint arXiv:2405.10948, 2024. 1 +[84] P Wang, S Bai, S Tan, S Wang, Z Fan, J Bai, K Chen, X Liu, J Wang, W Ge, and Y Fan. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 2 +[85] Xintong Wang, Jingheng Pan, Liang Ding, and Chris Biemann. Mitigating hallucinations in large vision-language models with instruction contrastive decoding, 2024. 3 +[86] Sangmin Woo, Jaehyuk Jang, Donguk Kim, Yubin Choi, and Changick Kim. Ritual: Random image transformations as a universal anti-hallucination lever in lvlms, 2024. http://arxiv.org/abs/2405.17821.3 +[87] Sangmin Woo, Donguk Kim, Jaehyuk Jang, Yubin Choi, and Changick Kim. Don't miss the forest for the trees: Attentional vision calibration for large vision language models, 2024. 2, 3 +[88] Junfei Wu, Qiang Liu, Ding Wang, Jinghao Zhang, Shu Wu, Liang Wang, and Tieniu Tan. Logical closed loop: Uncovering object hallucinations in large vision-language models, 2024. http://arxiv.org/abs/2402.11622.3 +[89] Kai Wu, Boyuan Jiang, Zhengkai Jiang, Qingdong He, Donghao Luo, Shengzhi Wang, Qingwen Liu, and Chengjie Wang. Noiseboost: Alleviating hallucination with noise perturbation for multimodal large language models, 2024. http://arxiv.org/abs/2405.20081.3 +[90] Xin Xiao, Bohong Wu, Jiacong Wang, Chunyuan Li, Xun Zhou, and Haoyuan Guo. Seeing the image: Prioritizing visual correlation by contrastive alignment, 2024. 3 +[91] Xiaohan Xu, Chongyang Tao, Tao Shen, Can Xu, Hongbo Xu, Guodong Long, and Jian guang Lou. Re-reading improves reasoning in large language models, 2024. http://arxiv.org/abs/2309.06275.3 +[92] L Xue, M Shu, A Awadalla, J Wang, A Yan, S Purushwalkam, H Zhou, V Prabhu, Y Dai, M S Ryoo, and S Kendre. xgen-mm (blip-3): A family of open large multimodal models. arXiv preprint arXiv:2408.08872, 2024. 2 +[93] Onoe Y., Zhang M.J., Choi E., and Durrett G. Entity cloze by date: What lms know about unseen entities. arXiv preprint arXiv:2205.02832, 2022. 3 +[94] Park Y., Lee D., Choe J., and Chang B. Convis: Contrastive decoding with hallucination visualization for mitigating hallucinations in multimodal large language models. arXiv preprint arXiv:2408.13906, 2024. 3 +[95] Zhenjie Yang, Xiaosong Jia, Hongyang Li, and Junchi Yan. Llm4drive: A survey of large language models for au + +tonomous driving. In NeurIPS 2024 Workshop on OpenWorld Agents, 2023. 1 +[96] Shukang Yin, Chaoyou Fu, Sirui Zhao, Tong Xu, Hao Wang, Dianbo Sui, Yunhang Shen, Ke Li, Xing Sun, and Enhong Chen. Woodpecker: Hallucination correction for multimodal large language models. arXiv preprint, 2023. 3 +[97] Qifan Yu, Juncheng Li, Longhui Wei, Liang Pang, Wentao Ye, Bosheng Qin, Siliang Tang, Qi Tian, and Yueting Zhuang. Hallucidctor: Mitigating hallucinatory toxicity in visual instruction data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12944–12953, 2024. 3 +[98] Tianyu Yu, Yuan Yao, Haoye Zhang, Taiwen He, Yifeng Han, Ganqu Cui, Jinyi Hu, et al. Rlhf-v: Towards trustworthy mllms via behavior alignment from fine-grained correctional human feedback. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13807-13816, 2024. 3 +[99] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, Cong Wei, Botao Yu, Ruibin Yuan, Renliang Sun, Ming Yin, Boyuan Zheng, Zhenzhu Yang, Yibo Liu, Wenhao Huang, Huan Sun, Yu Su, and Wenhu Chen. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of CVPR, 2024. 2 +[100] Zihao Yue, Liang Zhang, and Qin Jin. Less is more: Mitigating multimodal hallucination from an eos decision perspective, 2024. http://arxiv.org/abs/2402.14545.3 +[101] Bai Z., Wang P., Xiao T., He T., Han Z., Zhang Z., and Shou MZ. Hallucination of multimodal large language models: A survey. arXiv preprint arXiv:2404.18930, 2024. 3 +[102] Han Z., Bai Z., Mei H., Xu Q., Zhang C., and Shou M.Z. Skip: A simple method to reduce hallucination in large vision-language models. arXiv preprint arXiv:2402.01345, 2024. 3 +[103] Shaolei Zhang, Tian Yu, and Yang Feng. Truthx: Alleviating hallucinations by editing large language models in truthful space, 2024. 3 +[104] Yi-Fan Zhang, Weichen Yu, Qingsong Wen, Xue Wang, Zhang Zhang, Liang Wang, Rong Jin, and Tieniu Tan. Debiasing multimodal large language models. arXiv preprint, 2024. 3 +[105] L Zhao, Y Deng, W Zhang, and Q Gu. Mitigating object hallucination in large vision-language models via classifier-free guidance. 2024. 3 +[106] Zhiyuan Zhao, Bin Wang, Linke Ouyang, Xiaoyi Dong, Jiaqi Wang, and Conghui He. Beyond hallucinations: Enhancing lvlms through hallucination-aware direct preference optimization. arXiv preprint arXiv:2311.16839, 2023. 2 +[107] Zheng Zhao, Emilio Monti, Jens Lehmann, and Haytham Assem. Enhancing contextual understanding in large language models through contrastive decoding, 2024. 3 +[108] Lanyun Zhu, Deyi Ji, Tianrun Chen, Peng Xu, Jieping Ye, and Jun Liu. Ibid: Alleviating hallucinations in large + +vision-language models via image-biased decoding. arXiv preprint, 2024. 3 +[109] Wenhao Zhu, Sizhe Liu, Shujian Huang, Shuaijie She, Chris Wendler, and Jiajun Chen. Multilingual contrastive decoding via language-agnostic layers skipping, 2024. http://arxiv.org/abs/2407.10795.2,3 \ No newline at end of file diff --git a/2025/ICT_ Image-Object Cross-Level Trusted Intervention for Mitigating Object Hallucination in Large Vision-Language Models/images.zip b/2025/ICT_ Image-Object Cross-Level Trusted Intervention for Mitigating Object Hallucination in Large Vision-Language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c75437abb71f9e259ff434318eb3bafdd48b5a78 --- /dev/null +++ b/2025/ICT_ Image-Object Cross-Level Trusted Intervention for Mitigating Object Hallucination in Large Vision-Language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:acef8e9d56ea4a18561ce09d681e4a4636ce8ea80ac8b60670872927c666ae53 +size 708402 diff --git a/2025/ICT_ Image-Object Cross-Level Trusted Intervention for Mitigating Object Hallucination in Large Vision-Language Models/layout.json b/2025/ICT_ Image-Object Cross-Level Trusted Intervention for Mitigating Object Hallucination in Large Vision-Language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f19db3654433ace966dde24199ff67ad7519452b --- /dev/null +++ b/2025/ICT_ Image-Object Cross-Level Trusted Intervention for Mitigating Object Hallucination in Large Vision-Language Models/layout.json @@ -0,0 +1,10684 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 73, + 102, + 538, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 102, + 538, + 140 + ], + "spans": [ + { + "bbox": [ + 73, + 102, + 538, + 140 + ], + "type": "text", + "content": "ICT: Image-Object Cross-Level Trusted Intervention for Mitigating Object Hallucination in Large Vision-Language Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 143, + 160, + 465, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 160, + 465, + 189 + ], + "spans": [ + { + "bbox": [ + 143, + 160, + 465, + 189 + ], + "type": "text", + "content": "Junzhe Chen" + }, + { + "bbox": [ + 143, + 160, + 465, + 189 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 143, + 160, + 465, + 189 + ], + "type": "text", + "content": ", Tianshu Zhang" + }, + { + "bbox": [ + 143, + 160, + 465, + 189 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 143, + 160, + 465, + 189 + ], + "type": "text", + "content": ", Shiyu Huang" + }, + { + "bbox": [ + 143, + 160, + 465, + 189 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 143, + 160, + 465, + 189 + ], + "type": "text", + "content": ", Yuwei Niu" + }, + { + "bbox": [ + 143, + 160, + 465, + 189 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 143, + 160, + 465, + 189 + ], + "type": "text", + "content": ", Linfeng Zhang" + }, + { + "bbox": [ + 143, + 160, + 465, + 189 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 143, + 160, + 465, + 189 + ], + "type": "text", + "content": ", Lijie Wen" + }, + { + "bbox": [ + 143, + 160, + 465, + 189 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 143, + 160, + 465, + 189 + ], + "type": "text", + "content": ", Xuming Hu" + }, + { + "bbox": [ + 143, + 160, + 465, + 189 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 82, + 190, + 527, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 190, + 527, + 217 + ], + "spans": [ + { + "bbox": [ + 82, + 190, + 527, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 82, + 190, + 527, + 217 + ], + "type": "text", + "content": "Tsinghua University, " + }, + { + "bbox": [ + 82, + 190, + 527, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 82, + 190, + 527, + 217 + ], + "type": "text", + "content": "The Hong Kong University of Science and Technology (Guangzhou), " + }, + { + "bbox": [ + 82, + 190, + 527, + 217 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 82, + 190, + 527, + 217 + ], + "type": "text", + "content": "Zhipu AI, " + }, + { + "bbox": [ + 82, + 190, + 527, + 217 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 82, + 190, + 527, + 217 + ], + "type": "text", + "content": "Chongqing University, " + }, + { + "bbox": [ + 82, + 190, + 527, + 217 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 82, + 190, + 527, + 217 + ], + "type": "text", + "content": "Shanghai Jiao Tong University." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 192, + 218, + 414, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 218, + 414, + 231 + ], + "spans": [ + { + "bbox": [ + 192, + 218, + 414, + 231 + ], + "type": "text", + "content": "chenjz24@mails.tsinghua.edu.cn," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 232, + 460, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 232, + 460, + 245 + ], + "spans": [ + { + "bbox": [ + 143, + 232, + 460, + 245 + ], + "type": "text", + "content": "wenlj@tsinghua.edu.cn, xuminghu97@gmail.com," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "spans": [ + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 297, + 296, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 297, + 296, + 574 + ], + "spans": [ + { + "bbox": [ + 54, + 297, + 296, + 574 + ], + "type": "text", + "content": "Despite the recent breakthroughs achieved by Large Vision Language Models (LVLMs) in understanding and responding to complex visual-textual contexts, their inherent hallucination tendencies limit their practical application in real-world scenarios that demand high levels of precision. Existing methods typically either fine-tune the LVLMs using additional data, which incurs extra costs in manual annotation and computational resources or perform comparisons at the decoding stage, which may eliminate useful language priors for reasoning while introducing inference time overhead. Therefore, we propose ICT, a lightweight, training-free method that calculates an intervention direction to shift the model's focus towards different levels of visual information, enhancing its attention to high-level and fine-grained visual details. During the forward pass stage, the intervention is applied to the attention heads that encode the overall image information and the fine-grained object details, effectively mitigating the phenomenon of overly language priors, and thereby alleviating hallucinations. Extensive experiments demonstrate that ICT achieves strong performance with a small amount of data and generalizes well across different datasets and models. Our codes are publicly available at:https://github.com/THU-BPM/ICT/." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 594, + 135, + 607 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 135, + 607 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 135, + 607 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 614, + 295, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 614, + 295, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 614, + 295, + 687 + ], + "type": "text", + "content": "The recent rapid advancements in Large Vision Language Models (LVLMs) have demonstrated significant potential to tackle complex real-world tasks [5, 27, 29, 52, 56]. However, LVLMs frequently generate text outputs that are inconsistent with the visual input, such as incorrectly determining the presence of objects in an image or inaccurately" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 273, + 553, + 427 + ], + "blocks": [ + { + "bbox": [ + 317, + 273, + 553, + 427 + ], + "lines": [ + { + "bbox": [ + 317, + 273, + 553, + 427 + ], + "spans": [ + { + "bbox": [ + 317, + 273, + 553, + 427 + ], + "type": "image", + "image_path": "d5ea88bbeb089521af97a2e2f50af36678eb1f491b5330addf3d1535fd0d75e6.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 431, + 555, + 508 + ], + "lines": [ + { + "bbox": [ + 313, + 431, + 555, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 555, + 508 + ], + "type": "text", + "content": "Figure 1. Comparison between Contrastive Decoding (top) and our proposed ICT (bottom). In the top example, Contrastive Decoding indiscriminately removes both beneficial and detrimental language priors, leading to hallucinations. In contrast, our approach enhances the model's attention to visual details while preserving useful language priors, allowing it to correctly identify and describe objects in the image." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 509, + 555, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 555, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 555, + 700 + ], + "type": "text", + "content": "describing the attributes of those objects [55, 73]. This issue is commonly referred to as the object hallucination phenomenon [55], which significantly limits the applicability of LVLMs in high-stakes scenarios that demand precise accuracy, such as autonomous driving [95] and medical surgery [35, 83]. Previous research suggested that the primary reasons for hallucinations in LVLMs are as follows: 1) Large language models (LLMs), such as Vicuna [12], possess significantly stronger parameters and capabilities compared to visual encoders, such as CLIP [68]. This results in an excessively strong language prior, causing the model to overly rely on language cues while neglecting visual input [24, 25, 40, 43, 55]. 2) Current visual decoders often struggle to accurately capture fine-grained visual semantics, resulting in errors in detailed object attributes in an image, such as color and quantity [4, 9, 13, 55]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 325, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 325, + 701, + 553, + 714 + ], + "type": "text", + "content": "Based on the two issues outlined above, previous ef" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 693, + 142, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 693, + 142, + 703 + ], + "spans": [ + { + "bbox": [ + 70, + 693, + 142, + 703 + ], + "type": "text", + "content": "* Equal Contribution." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 70, + 703, + 153, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 703, + 153, + 712 + ], + "spans": [ + { + "bbox": [ + 70, + 703, + 153, + 712 + ], + "type": "text", + "content": "† Corresponding authors." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4209" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 456 + ], + "type": "text", + "content": "forts to mitigate the hallucination phenomenon in LVLMs can be categorized into three main approaches: 1) Finetuning with additional data: This approach involves introducing high-quality annotated data to better align the model's behavior with human interpretations, effectively teaching the model to focus more on visual information [23, 46, 52, 106]. However, this method not only requires costly human annotation but also involves updating model parameters, which demands substantial computational resources, thereby limiting its scalability. 2) Perceptual enhancement: This method incorporates additional information, such as depth maps and segmentation maps [34, 38, 106], as auxiliary inputs to assist the visual encoder in capturing more detailed visual features, thereby reducing hallucinations. However, it often requires manual selection of auxiliary features, which limits its generalizability across different tasks. 3) Contrastive decoding: This method alleviates hallucinations without requiring additional training. It induces hallucinations by applying transformations such as blurring, rotation, or cropping to the original visual input. During the decoding stage, tokens associated with these induced hallucinations are penalized, thus mitigating the influence of language priors [10, 14, 36, 43, 64, 70, 82, 87, 109]. However, methods such as VCD often indiscriminately eliminate all language priors, including those that may be beneficial. As shown in Figure 1, the original model recognizes Curry holding a golf club, while utilizing the language prior that identifies Curry as a basketball player. This language prior is valuable for achieving an accurate interpretation. However, through contrastive decoding, this useful prior is also removed, which can inadvertently result in hallucinations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 462, + 298, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 298, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 298, + 714 + ], + "type": "text", + "content": "To address the challenges of mitigating hallucinations in LVLMs, we propose Image-Object Cross-Level Trusted Intervention (ICT), a training-free, plug-and-play method applicable during the forward pass. Unlike contrastive decoding, our approach does not eliminate language priors to reduce the model's over-reliance on text modality. Instead, it intervenes during the forward pass to enhance the model's focus on both comprehensive visual information and fine-grained object details. We explored the activation patterns of attention heads when the model produces correct versus hallucinated responses. This analysis allowed us to identify activation value deviations that can shift the model from being \"untrustworthy\" to \"trustworthy\". According to prior research [8, 19, 81], which has shown that different heads in the multi-head attention mechanism encode information at varying levels of granularity, we train binary classifiers for each head to determine which heads encode overall visual information and which capture detailed visual features. During the forward pass, we adjust the activation values of these heads based on their identified levels, thereby enhancing the model's attention to the relevant visual features and" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "text", + "content": "reducing the likelihood of hallucinations. As illustrated in Figure 1, after applying ICT, the model is able to focus more on the details within the image, such as identifying the man as Curry, while simultaneously utilizing beneficial language priors (e.g., Curry is a basketball player) to infer and arrive at the correct answer. Since the intervention shift vectors are pre-computed, ICT does not introduce additional delays during the forward pass." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 169, + 556, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 169, + 556, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 169, + 556, + 240 + ], + "type": "text", + "content": "Our experiments demonstrate that, for both LLaVAv1.5 [52] and Qwen-VL [5], applying ICT results in an average improvement of " + }, + { + "bbox": [ + 313, + 169, + 556, + 240 + ], + "type": "inline_equation", + "content": "6.27\\%" + }, + { + "bbox": [ + 313, + 169, + 556, + 240 + ], + "type": "text", + "content": " on POPE benchmark and 67.37 points on MME benchmark. Furthermore, ICT exhibits cross-dataset generalization and model-agnostic generalizability. Our contributions can be summarized as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 241, + 555, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 241, + 555, + 313 + ], + "spans": [ + { + "bbox": [ + 313, + 241, + 555, + 313 + ], + "type": "text", + "content": "- We propose Image-Object Cross-Level Trusted Intervention (ICT), a novel, training-free, plug-and-play method that effectively reduces hallucinations in LVLMs by enhancing focus on both overall visual information and fine-grained object details during the forward pass, without eliminating beneficial language priors." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 314, + 555, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 314, + 555, + 374 + ], + "spans": [ + { + "bbox": [ + 313, + 314, + 555, + 374 + ], + "type": "text", + "content": "- Unlike existing contrastive decoding approaches, we introduce an intervention mechanism that operates during the forward pass rather than the decoding stage, allowing ICT to be orthogonal and complementary to existing decoding strategies while not introducing any additional latency." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 375, + 554, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 375, + 554, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 375, + 554, + 422 + ], + "type": "text", + "content": "- Extensive experiments on LLaVA-v1.5 and Qwen-VL demonstrate that ICT significantly improves performance on both the POPE and MME benchmarks while maintaining cross-dataset and model-agnostic generalizability." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 435, + 402, + 448 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 435, + 402, + 448 + ], + "spans": [ + { + "bbox": [ + 313, + 435, + 402, + 448 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 456, + 485, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 485, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 485, + 468 + ], + "type": "text", + "content": "2.1. Large Vision-Language Models" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "content": "Following the success of LLM [6, 12, 61, 78, 80], researchers have been exploring the multimodal field. Taking advantage of the remarkable capabilities of LLMs, Large Vision-Language Models incorporate visual encoders and feature projectors into powerful LLMs, enabling them to understand and generate content based on visual and text input. Most of these models undergo two training steps, namely the pre-training stage and the fine-tuning stage. The main purpose of the pre-training stage is to align textual and visual features, while the goal of the fine-tuning stage is to bridge further the modality gap between vision and language, as well as lifting their instruction following capabilities and performance on specific downstream tasks. Early attempts, such as FLamingo [2], Gemini [79], and BLIP-2 [44], have already shown promising results. Recent works like LLaVA-v1.5 [54], Qwen2-VL [84], and xgenmm [92] have advanced this field, dramatically improving the abilities of these models. To enhance the consistency between vision representation and language representation, many efforts have been made, including utilizing higher res" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 317, + 757 + ], + "type": "text", + "content": "4210" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "text", + "content": "olution vision encoders, transitioning to larger and more capable LLMs, and adapting reinforcement learning techniques such as RLHF [62, 98], among others. Despite the remarkable progress that has been made, LVLMs still suffer from severe hallucinations, which restricts their potential for large-scale applications in real-world scenarios." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 155, + 250, + 168 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 155, + 250, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 155, + 250, + 168 + ], + "type": "text", + "content": "2.2. Mitigating Hallucinations in LVLMs" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 174, + 296, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 296, + 450 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 296, + 450 + ], + "type": "text", + "content": "Many efforts have been made to understand the causes of hallucination [1, 16, 30, 39, 50, 50, 53, 63, 65, 71, 72, 76, 93, 101]. Existing approaches to mitigate hallucinations can be roughly divided into two categories based on the stage they take place. The first category focuses on the training phase. Most of the work falling into this category introduces additional or curated datasets [21, 22, 51, 97], these works usually design a dataset specifically for hallucination-related tasks or improve data cleaning methods [69, 75], or introduce novel training objectives [7, 41, 77]. While effective, these methods typically require extensive training, which is both time-consuming and resource-intensive. The second category focuses on the inferencing phase, usually involves CD-based [47] novel decoding strategies [10, 14, 36, 59, 64, 70, 82, 87, 94, 107, 109]. Another approach is to detect potential hallucinations while generating and correct them [18, 37, 60], among others [45, 86, 100, 103]. In addition, some researchers tackle hallucinations by manipulating attention weight assigned to the image [33, 89, 90, 104, 108], or the prompt-related part of the image [3]. Other works include prompt-based methods [42, 67, 85, 88, 91, 102], utilizing external tools [11, 96, 105] or external knowledge [17, 28, 66] and so on." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 464, + 162, + 476 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 464, + 162, + 476 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 162, + 476 + ], + "type": "text", + "content": "3. Task Formulation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 485, + 296, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 485, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 485, + 296, + 594 + ], + "type": "text", + "content": "Given an LVLM parameterized by " + }, + { + "bbox": [ + 55, + 485, + 296, + 594 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 485, + 296, + 594 + ], + "type": "text", + "content": ", the model processes a textual input " + }, + { + "bbox": [ + 55, + 485, + 296, + 594 + ], + "type": "inline_equation", + "content": "\\pmb{X} = \\{x_{1}, x_{2}, \\ldots, x_{N_{x}}\\}" + }, + { + "bbox": [ + 55, + 485, + 296, + 594 + ], + "type": "text", + "content": " and a visual input " + }, + { + "bbox": [ + 55, + 485, + 296, + 594 + ], + "type": "inline_equation", + "content": "\\pmb{V} = \\{v_{1}, v_{2}, \\ldots, v_{N_{v}}\\}" + }, + { + "bbox": [ + 55, + 485, + 296, + 594 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 485, + 296, + 594 + ], + "type": "inline_equation", + "content": "N_{x}" + }, + { + "bbox": [ + 55, + 485, + 296, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 485, + 296, + 594 + ], + "type": "inline_equation", + "content": "N_{v}" + }, + { + "bbox": [ + 55, + 485, + 296, + 594 + ], + "type": "text", + "content": " denote the sequence lengths of the text and visual inputs, respectively. The model then concatenates the textual and visual sequences to form a unified input " + }, + { + "bbox": [ + 55, + 485, + 296, + 594 + ], + "type": "inline_equation", + "content": "\\pmb{H} = \\text{concat}(\\pmb{V}, \\pmb{X})" + }, + { + "bbox": [ + 55, + 485, + 296, + 594 + ], + "type": "text", + "content": ", which is subsequently passed through " + }, + { + "bbox": [ + 55, + 485, + 296, + 594 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 485, + 296, + 594 + ], + "type": "text", + "content": " layers of a transformer architecture. At each layer, the concatenated input undergoes multi-head attention, which is computed as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 84, + 605, + 295, + 638 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 605, + 295, + 638 + ], + "spans": [ + { + "bbox": [ + 84, + 605, + 295, + 638 + ], + "type": "interline_equation", + "content": "\\boldsymbol {H} ^ {(l + 1)} = \\boldsymbol {H} ^ {(l)} + \\sum_ {n = 1} ^ {N} A t t n _ {n} ^ {(l)} \\left(\\boldsymbol {H} ^ {(l)}\\right) \\cdot W _ {o} ^ {(l)}, \\tag {1}", + "image_path": "eeb6cff04e92cdb30a05c1d6d91aaa40eb952803f4bd2e182c2523072ac5bda0.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 651, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 651, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 651, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 651, + 296, + 714 + ], + "type": "inline_equation", + "content": "Attn_{n}^{(l)}" + }, + { + "bbox": [ + 55, + 651, + 296, + 714 + ], + "type": "text", + "content": " denotes the attention operation of the " + }, + { + "bbox": [ + 55, + 651, + 296, + 714 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 651, + 296, + 714 + ], + "type": "text", + "content": "-th head at the " + }, + { + "bbox": [ + 55, + 651, + 296, + 714 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 651, + 296, + 714 + ], + "type": "text", + "content": "-th layer, and " + }, + { + "bbox": [ + 55, + 651, + 296, + 714 + ], + "type": "inline_equation", + "content": "W_{o}^{(l)} \\in \\mathbb{R}^{d \\times Nd}" + }, + { + "bbox": [ + 55, + 651, + 296, + 714 + ], + "type": "text", + "content": " is the output projection matrix, with " + }, + { + "bbox": [ + 55, + 651, + 296, + 714 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 651, + 296, + 714 + ], + "type": "text", + "content": " representing the dimensionality of each attention head and " + }, + { + "bbox": [ + 55, + 651, + 296, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 651, + 296, + 714 + ], + "type": "text", + "content": " the number of heads. Subsequently, the model autoregressively predicts the next token" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 478, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 478, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 478, + 84 + ], + "type": "text", + "content": "based on the output from the " + }, + { + "bbox": [ + 313, + 72, + 478, + 84 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 72, + 478, + 84 + ], + "type": "text", + "content": "-th layer:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 357, + 90, + 555, + 105 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 90, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 357, + 90, + 555, + 105 + ], + "type": "interline_equation", + "content": "p \\left(y _ {t} \\mid y _ {< t}\\right) = \\operatorname {S o f t m a x} \\left(f c \\left(\\boldsymbol {H} _ {t} ^ {(L)}\\right)\\right), \\tag {2}", + "image_path": "97ef8d2cc5d67d0e7938f7e8e255af31b837f01ca82ca290b0e3b66746fc2e25.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 110, + 555, + 135 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 110, + 555, + 135 + ], + "spans": [ + { + "bbox": [ + 313, + 110, + 555, + 135 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 110, + 555, + 135 + ], + "type": "inline_equation", + "content": "f c(\\cdot)" + }, + { + "bbox": [ + 313, + 110, + 555, + 135 + ], + "type": "text", + "content": " is an affine layer mapping the " + }, + { + "bbox": [ + 313, + 110, + 555, + 135 + ], + "type": "inline_equation", + "content": "Hd" + }, + { + "bbox": [ + 313, + 110, + 555, + 135 + ], + "type": "text", + "content": "-dimensional vector to a probability distribution of vocabulary size " + }, + { + "bbox": [ + 313, + 110, + 555, + 135 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 110, + 555, + 135 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 144, + 397, + 157 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 397, + 157 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 397, + 157 + ], + "type": "text", + "content": "4. Methodology" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 163, + 554, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 163, + 554, + 236 + ], + "spans": [ + { + "bbox": [ + 313, + 163, + 554, + 236 + ], + "type": "text", + "content": "In this section, we introduce two levels of fine-grained intervention modules. The first module enables the LVLM to focus on the image, thereby reducing over-reliance on language priors. The second module encourages the LVLM to attend more closely to image objects, helping to mitigate the omission of critical objects and reduce hallucinations." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 242, + 468, + 254 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 242, + 468, + 254 + ], + "spans": [ + { + "bbox": [ + 313, + 242, + 468, + 254 + ], + "type": "text", + "content": "4.1. Intervention at Image-Level" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 258, + 554, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 554, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 554, + 319 + ], + "type": "text", + "content": "This module aims to identify the attention heads associated with overall image information and to apply targeted interventions to these heads. This approach enhances the model's focus on visual input while diminishing the influence of language priors." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 319, + 554, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 554, + 427 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 554, + 427 + ], + "type": "text", + "content": "As illustrated in Figure 2, consider a batch of image-question pairs, denoted as " + }, + { + "bbox": [ + 313, + 319, + 554, + 427 + ], + "type": "inline_equation", + "content": "\\{(X_i,V_i)\\}_{i = 1}^B" + }, + { + "bbox": [ + 313, + 319, + 554, + 427 + ], + "type": "text", + "content": ", each labeled with the answer \"Yes\" from the POPE dataset. Here, " + }, + { + "bbox": [ + 313, + 319, + 554, + 427 + ], + "type": "inline_equation", + "content": "X_{i}" + }, + { + "bbox": [ + 313, + 319, + 554, + 427 + ], + "type": "text", + "content": " takes the form of \"Is there a/an [object] in the image?\" For each question, we extract the specified object " + }, + { + "bbox": [ + 313, + 319, + 554, + 427 + ], + "type": "inline_equation", + "content": "O_{i}" + }, + { + "bbox": [ + 313, + 319, + 554, + 427 + ], + "type": "text", + "content": " and reformulate the question to " + }, + { + "bbox": [ + 313, + 319, + 554, + 427 + ], + "type": "inline_equation", + "content": "q =" + }, + { + "bbox": [ + 313, + 319, + 554, + 427 + ], + "type": "text", + "content": " \"What is it in the image?\" Then, for each image " + }, + { + "bbox": [ + 313, + 319, + 554, + 427 + ], + "type": "inline_equation", + "content": "V_{i}" + }, + { + "bbox": [ + 313, + 319, + 554, + 427 + ], + "type": "text", + "content": ", we progressively add Gaussian noise following the forward diffusion process [26] to obtain the final blurred image " + }, + { + "bbox": [ + 313, + 319, + 554, + 427 + ], + "type": "inline_equation", + "content": "V_{i}^{\\prime}" + }, + { + "bbox": [ + 313, + 319, + 554, + 427 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 324, + 431, + 553, + 482 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 431, + 553, + 482 + ], + "spans": [ + { + "bbox": [ + 324, + 431, + 553, + 482 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} q (V _ {i} ^ {(t)} \\mid V _ {i} ^ {(t - 1)}) = \\mathcal {N} (V _ {i} ^ {(t)}; \\sqrt {1 - \\beta_ {t}} V _ {i} ^ {(t - 1)}, \\beta_ {t} \\mathbf {I}), \\\\ q \\left(V _ {i} ^ {\\prime} \\mid V _ {i}\\right) = \\prod_ {t = 1} ^ {T} q \\left(V _ {i} ^ {(t)} \\mid V _ {i} ^ {(t - 1)}\\right), \\tag {3} \\\\ \\end{array}", + "image_path": "f715f417c9b008bf8c1de18b7d860388a5c3e63cc39b3503a9ea4fffd876da22.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 487, + 555, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 555, + 679 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 555, + 679 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 487, + 555, + 679 + ], + "type": "inline_equation", + "content": "\\beta_{t}" + }, + { + "bbox": [ + 313, + 487, + 555, + 679 + ], + "type": "text", + "content": " is the noise variance at step " + }, + { + "bbox": [ + 313, + 487, + 555, + 679 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 487, + 555, + 679 + ], + "type": "text", + "content": ". Finally, we construct a modified dataset " + }, + { + "bbox": [ + 313, + 487, + 555, + 679 + ], + "type": "inline_equation", + "content": "\\{(q,V_i,V_i',O_i)\\}_{i=1}^B" + }, + { + "bbox": [ + 313, + 487, + 555, + 679 + ], + "type": "text", + "content": " to obtain Image-Level intervention vectors. We then concatenate " + }, + { + "bbox": [ + 313, + 487, + 555, + 679 + ], + "type": "inline_equation", + "content": "O_i" + }, + { + "bbox": [ + 313, + 487, + 555, + 679 + ], + "type": "text", + "content": " with the question " + }, + { + "bbox": [ + 313, + 487, + 555, + 679 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 313, + 487, + 555, + 679 + ], + "type": "text", + "content": " to form the answer, resulting in data pairs of trusted and untrusted data for the same sample: " + }, + { + "bbox": [ + 313, + 487, + 555, + 679 + ], + "type": "inline_equation", + "content": "(q + O_i,V_i)" + }, + { + "bbox": [ + 313, + 487, + 555, + 679 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 487, + 555, + 679 + ], + "type": "inline_equation", + "content": "(q + O_i,V_i')" + }, + { + "bbox": [ + 313, + 487, + 555, + 679 + ], + "type": "text", + "content": ". For each sample, we treat the representation of the last token as the fused representation of multimodal data and extract the attention activation from a total of " + }, + { + "bbox": [ + 313, + 487, + 555, + 679 + ], + "type": "inline_equation", + "content": "N\\times L" + }, + { + "bbox": [ + 313, + 487, + 555, + 679 + ], + "type": "text", + "content": " heads across the " + }, + { + "bbox": [ + 313, + 487, + 555, + 679 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 487, + 555, + 679 + ], + "type": "text", + "content": " layers for both trusted and untrusted data. These are denoted as " + }, + { + "bbox": [ + 313, + 487, + 555, + 679 + ], + "type": "inline_equation", + "content": "A_{i} = \\{Attn_{n}^{(l)}H_{s}^{(l)}\\}_{n=1,l=1}^{N,L}" + }, + { + "bbox": [ + 313, + 487, + 555, + 679 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 487, + 555, + 679 + ], + "type": "inline_equation", + "content": "A_{i}' = \\{Attn_{n}^{(l)}H_{s}^{(l)}\\}_{n=1,l=1}^{N,L}" + }, + { + "bbox": [ + 313, + 487, + 555, + 679 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 487, + 555, + 679 + ], + "type": "inline_equation", + "content": "s = N_x + N_v + 1" + }, + { + "bbox": [ + 313, + 487, + 555, + 679 + ], + "type": "text", + "content": ". According to the trusted activation " + }, + { + "bbox": [ + 313, + 487, + 555, + 679 + ], + "type": "inline_equation", + "content": "\\{A_{i}\\}_{i=1}^{B}" + }, + { + "bbox": [ + 313, + 487, + 555, + 679 + ], + "type": "text", + "content": " and the untrusted activation " + }, + { + "bbox": [ + 313, + 487, + 555, + 679 + ], + "type": "inline_equation", + "content": "\\{A_{i}'\\}_{i=1}^{B}" + }, + { + "bbox": [ + 313, + 487, + 555, + 679 + ], + "type": "text", + "content": " obtained from all sample pairs, we can calculate the following activation shift vector " + }, + { + "bbox": [ + 313, + 487, + 555, + 679 + ], + "type": "inline_equation", + "content": "\\{S_{n}^{(l)}\\}_{n=1,l=1}^{N,L}" + }, + { + "bbox": [ + 313, + 487, + 555, + 679 + ], + "type": "text", + "content": " that encourages the model to pay more attention to visual information, as follows:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 369, + 685, + 555, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 685, + 555, + 717 + ], + "spans": [ + { + "bbox": [ + 369, + 685, + 555, + 717 + ], + "type": "interline_equation", + "content": "\\boldsymbol {S} _ {n} ^ {(l)} = \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\left(\\boldsymbol {A} _ {i, n} ^ {(l)} - \\boldsymbol {A} _ {i, n} ^ {\\prime (l)}\\right). \\tag {4}", + "image_path": "a738c2e06dbac64d92712b1b9feede8ae8a740a4aea3cb337cfa2714c15ce6d5.jpg" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "text", + "content": "4211" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 556, + 230 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 556, + 230 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 556, + 230 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 556, + 230 + ], + "type": "image", + "image_path": "609bed94047a25603363ccc606374ecd88d35da5bdb7fac0915277db876cde1b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 239, + 555, + 283 + ], + "lines": [ + { + "bbox": [ + 55, + 239, + 555, + 283 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 555, + 283 + ], + "type": "text", + "content": "Figure 2. Overview of our proposed ICT method. ICT consists of two levels of intervention: Image-Level and Object-Level. The Image-Level module enhances the model's focus on the overall scene to reduce reliance on language priors, while the Object-Level module focuses on fine-grained details to mitigate hallucinations. We apply targeted activation shifts to selected attention heads identified through binary classifiers trained to distinguish trusted and untrusted data pairs." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 289, + 296, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 296, + 361 + ], + "type": "text", + "content": "Next, we train a binary classifier " + }, + { + "bbox": [ + 55, + 289, + 296, + 361 + ], + "type": "inline_equation", + "content": "f_{n}^{(l)}(\\cdot)" + }, + { + "bbox": [ + 55, + 289, + 296, + 361 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 55, + 289, + 296, + 361 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 289, + 296, + 361 + ], + "type": "text", + "content": " sample pairs for each head to detect which heads encode Image-Level information, specifically those that can better distinguish the differences between pairs of trusted and untrusted samples. We then apply activation interventions to these selected heads:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 369, + 296, + 386 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 369, + 296, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 369, + 296, + 386 + ], + "type": "interline_equation", + "content": "A t t n _ {\\mathrm {i m g}} = \\left\\{A t t n _ {n} ^ {(l)} \\mid A t t n _ {n} ^ {(l)} \\in T o p K (\\text {A c c u r a c y} (f _ {n} ^ {(l)} (\\cdot))) \\right\\},", + "image_path": "65976bb8a20f21fb7ef0ed18df5f409d9994ecca37e51e7690b1a7ae7593c428.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 388, + 296, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 388, + 296, + 430 + ], + "spans": [ + { + "bbox": [ + 58, + 388, + 296, + 430 + ], + "type": "interline_equation", + "content": "\\boldsymbol {H} ^ {(l + 1)} = \\boldsymbol {H} ^ {(l)} + \\sum_ {n = 1} ^ {N} \\left(A t t n _ {n} ^ {(l)} \\left(\\boldsymbol {H} ^ {(l)}\\right) + \\mathbb {I} _ {\\operatorname {i m g}, n} ^ {(l)} \\alpha \\boldsymbol {S} _ {n} ^ {(l)}\\right) \\cdot W _ {o} ^ {(l)}, \\tag {5}", + "image_path": "ba549e4f8c37c93b0e6e7814a0340ddb69b8886d060fc1596d1d93adba9fa1c5.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 430, + 296, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 430, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 296, + 533 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 430, + 296, + 533 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 430, + 296, + 533 + ], + "type": "text", + "content": " is the number of selected intervention heads, " + }, + { + "bbox": [ + 55, + 430, + 296, + 533 + ], + "type": "inline_equation", + "content": "\\mathbb{I}_{\\mathrm{img},n}^{(l)}" + }, + { + "bbox": [ + 55, + 430, + 296, + 533 + ], + "type": "text", + "content": " is an indicator function, which is 1 if " + }, + { + "bbox": [ + 55, + 430, + 296, + 533 + ], + "type": "inline_equation", + "content": "Attn_{n}^{(l)} \\in Attn_{\\mathrm{img}}" + }, + { + "bbox": [ + 55, + 430, + 296, + 533 + ], + "type": "text", + "content": " and 0 otherwise, and " + }, + { + "bbox": [ + 55, + 430, + 296, + 533 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 430, + 296, + 533 + ], + "type": "text", + "content": " denotes the intensity of the intervention. After using " + }, + { + "bbox": [ + 55, + 430, + 296, + 533 + ], + "type": "inline_equation", + "content": "\\{S_n^{(l)}\\}_{n=1,l=1}^{N,L}" + }, + { + "bbox": [ + 55, + 430, + 296, + 533 + ], + "type": "text", + "content": " to perform Image-Level interventions on heads that encode image information, the model enhances the trustworthiness of the visual level, places greater attention on visual information, thus mitigates the impact of overly strong language priors." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 540, + 212, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 212, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 212, + 553 + ], + "type": "text", + "content": "4.2. Intervention at Object-Level" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 558, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 715 + ], + "type": "text", + "content": "After enhancing the model's trustworthiness at the Image-Level, a more fine-grained, Object-Level intervention becomes necessary to increase the model's attention to image details, thereby reducing hallucinations caused by the omission of fine details. As shown in Figure 2, similar to Section 4.1, we construct a dataset " + }, + { + "bbox": [ + 55, + 558, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\{(q,V_i,V_i^{\\prime \\prime},O_i)\\}_{i = 1}^{B}" + }, + { + "bbox": [ + 55, + 558, + 296, + 715 + ], + "type": "text", + "content": " to obtain Object-Level intervention vectors. We use Grounding DINO [58] to identify the area of object " + }, + { + "bbox": [ + 55, + 558, + 296, + 715 + ], + "type": "inline_equation", + "content": "O_{i}" + }, + { + "bbox": [ + 55, + 558, + 296, + 715 + ], + "type": "text", + "content": " in image " + }, + { + "bbox": [ + 55, + 558, + 296, + 715 + ], + "type": "inline_equation", + "content": "V_{i}" + }, + { + "bbox": [ + 55, + 558, + 296, + 715 + ], + "type": "text", + "content": ". Gaussian noise is then added selectively to this object region, based on Eq. 3, producing a locally blurred image " + }, + { + "bbox": [ + 55, + 558, + 296, + 715 + ], + "type": "inline_equation", + "content": "V_{i}^{\\prime \\prime}" + }, + { + "bbox": [ + 55, + 558, + 296, + 715 + ], + "type": "text", + "content": ". Using both the original image " + }, + { + "bbox": [ + 55, + 558, + 296, + 715 + ], + "type": "inline_equation", + "content": "V_{i}" + }, + { + "bbox": [ + 55, + 558, + 296, + 715 + ], + "type": "text", + "content": " and the partially blurred image " + }, + { + "bbox": [ + 55, + 558, + 296, + 715 + ], + "type": "inline_equation", + "content": "V_{i}^{\\prime \\prime}" + }, + { + "bbox": [ + 55, + 558, + 296, + 715 + ], + "type": "text", + "content": ", we construct trusted and untrusted data pairs for each sample: " + }, + { + "bbox": [ + 55, + 558, + 296, + 715 + ], + "type": "inline_equation", + "content": "(q + O_i,V_i)" + }, + { + "bbox": [ + 55, + 558, + 296, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 558, + 296, + 715 + ], + "type": "inline_equation", + "content": "(q + O_i,V_i^{\\prime \\prime})" + }, + { + "bbox": [ + 55, + 558, + 296, + 715 + ], + "type": "text", + "content": " and analyze" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 290, + 555, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 290, + 555, + 329 + ], + "spans": [ + { + "bbox": [ + 313, + 290, + 555, + 329 + ], + "type": "text", + "content": "the attention activation values across " + }, + { + "bbox": [ + 313, + 290, + 555, + 329 + ], + "type": "inline_equation", + "content": "N\\times L" + }, + { + "bbox": [ + 313, + 290, + 555, + 329 + ], + "type": "text", + "content": " heads, allowing us to compute an Object-Level activation shift vector " + }, + { + "bbox": [ + 313, + 290, + 555, + 329 + ], + "type": "inline_equation", + "content": "\\{S_{n,\\mathrm{obj}}^{(l)}\\}_{n = 1,l = 1}^{N,L}" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 364, + 331, + 555, + 363 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 331, + 555, + 363 + ], + "spans": [ + { + "bbox": [ + 364, + 331, + 555, + 363 + ], + "type": "interline_equation", + "content": "\\boldsymbol {S} _ {n, \\mathrm {o b j}} ^ {(l)} = \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\left(\\boldsymbol {A} _ {i, n} ^ {(l)} - \\boldsymbol {A} _ {i, n} ^ {\\prime \\prime (l)}\\right). \\tag {6}", + "image_path": "d02d8f2a541ad50e445c74498e7e9d393f1feab889e59a4533e04c4511a10edc.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 371, + 555, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 371, + 555, + 420 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 555, + 420 + ], + "type": "text", + "content": "A binary classifier " + }, + { + "bbox": [ + 313, + 371, + 555, + 420 + ], + "type": "inline_equation", + "content": "f_{n,obj}^{(l)}(\\cdot)" + }, + { + "bbox": [ + 313, + 371, + 555, + 420 + ], + "type": "text", + "content": " is then trained to identify heads that effectively distinguish trusted from untrusted object-focused samples. Interventions are subsequently applied to the selected heads as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 429, + 558, + 447 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 429, + 558, + 447 + ], + "spans": [ + { + "bbox": [ + 314, + 429, + 558, + 447 + ], + "type": "interline_equation", + "content": "A t t n _ {\\mathrm {o b j}} = \\left\\{A t t n _ {n} ^ {(l)} \\mid A t t n _ {n} ^ {(l)} \\in T o p K (\\text {A c c u r a c y} (f _ {n, o b j} ^ {(l)} (\\cdot))) \\right\\},", + "image_path": "b571da9c9b355017b06c58e22c49c393892412200a54d014185f03290b1bf2da.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 449, + 557, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 449, + 557, + 491 + ], + "spans": [ + { + "bbox": [ + 315, + 449, + 557, + 491 + ], + "type": "interline_equation", + "content": "\\boldsymbol {H} ^ {(l + 1)} = \\boldsymbol {H} ^ {(l)} + \\sum_ {n = 1} ^ {N} \\left(A t t n _ {n} ^ {(l)} \\left(\\boldsymbol {H} ^ {(l)}\\right) + \\mathbb {I} _ {\\mathrm {o b j}, n} ^ {(l)} \\beta \\boldsymbol {S} _ {\\mathrm {o b j}, n} ^ {(l)}\\right) \\cdot W _ {o} ^ {(l)}, \\tag {7}", + "image_path": "b56abddd1531adfb604069bf32766894b9e595b12af4d1328d37949763d8e52e.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 491, + 554, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 554, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 554, + 532 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 491, + 554, + 532 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 491, + 554, + 532 + ], + "type": "text", + "content": " is the number of intervention heads, " + }, + { + "bbox": [ + 313, + 491, + 554, + 532 + ], + "type": "inline_equation", + "content": "\\mathbb{I}_{\\mathrm{obj},n}^{(l)}" + }, + { + "bbox": [ + 313, + 491, + 554, + 532 + ], + "type": "text", + "content": " is an indicator function set to 1 if " + }, + { + "bbox": [ + 313, + 491, + 554, + 532 + ], + "type": "inline_equation", + "content": "Attn_{n}^{(l)} \\in Attn_{\\mathrm{obj}}" + }, + { + "bbox": [ + 313, + 491, + 554, + 532 + ], + "type": "text", + "content": " and 0 otherwise, and " + }, + { + "bbox": [ + 313, + 491, + 554, + 532 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 491, + 554, + 532 + ], + "type": "text", + "content": " controls the intervention intensity." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 533, + 555, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 555, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 555, + 581 + ], + "type": "text", + "content": "Finally, we integrate the Image-Level and Object-Level intervention modules to create a unified approach that reinforces the model's focus on both the overall visual context and finer object-specific details as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 339, + 590, + 553, + 644 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 590, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 339, + 590, + 553, + 644 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\boldsymbol {H} ^ {(l + 1)} = \\boldsymbol {H} ^ {(l)} + \\sum_ {n = 1} ^ {N} \\left(A t t n _ {n} ^ {(l)} \\left(\\boldsymbol {H} ^ {(l)}\\right) \\right. \\tag {8} \\\\ \\left. \\right.\\left. + \\mathbb {I} _ {\\operatorname {i m g}, n} ^ {(l)} \\alpha \\boldsymbol {S} _ {n} ^ {(l)} + \\mathbb {I} _ {\\operatorname {o b j}, n} ^ {(l)} \\beta \\boldsymbol {S} _ {\\operatorname {o b j}, n} ^ {(l)}\\right) \\cdot W _ {o} ^ {(l)}. \\\\ \\end{array}", + "image_path": "b15e9a232fd5aed2dd6d1a45a54f9d56c49e59dc296b5aab81640a415844dc8a.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "content": "By applying these interventions in tandem, the model gains a balanced attention mechanism that mitigates excessive reliance on language priors while enhancing sensitivity to essential visual cues across different levels of granularity, thereby alleviating the occurrence of hallucinations." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4212" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 137, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 137, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 137, + 85 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 171, + 105 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 171, + 105 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 171, + 105 + ], + "type": "text", + "content": "5.1. Experimental Setup" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 110, + 151, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 110, + 151, + 121 + ], + "spans": [ + { + "bbox": [ + 55, + 110, + 151, + 121 + ], + "type": "text", + "content": "Datasets and Metrics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 121, + 295, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 121, + 295, + 253 + ], + "spans": [ + { + "bbox": [ + 54, + 121, + 295, + 253 + ], + "type": "text", + "content": "POPE [48] is a benchmark designed to assess how effectively LVLMs recognize the presence of specific objects in images, thus identifying Object-Level hallucinations. It uses Yes/No questions based on object annotations, with metrics including Accuracy, Precision, Recall, and F1 score. The dataset is balanced, with " + }, + { + "bbox": [ + 54, + 121, + 295, + 253 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 54, + 121, + 295, + 253 + ], + "type": "text", + "content": " of queries for existing objects and " + }, + { + "bbox": [ + 54, + 121, + 295, + 253 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 54, + 121, + 295, + 253 + ], + "type": "text", + "content": " for non-existing ones, and employs three sampling strategies: random, popular, and adversarial. Drawing from MSCOCO [49], A-OKVQA [74], and GQA [32] datasets, POPE evaluates 27,000 query-answer pairs to gauge model performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 254, + 296, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 254, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 254, + 296, + 350 + ], + "type": "text", + "content": "MME [20] benchmark is designed to comprehensively evaluate the performance of LVLMs across various dimensions. It includes ten perception-focused tasks and four cognition-related tasks. MME benchmark specifically addresses Object-Level hallucinations through subsets focusing on object existence and count, while attribute-level hallucinations are assessed via subsets related to object position and color. The evaluation metric is Accuracy." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 350, + 296, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 350, + 296, + 409 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 296, + 409 + ], + "type": "text", + "content": "Baselines. We adopt the widely used LLaVA-v1.5 [52] and Qwen-VL [5] models as our baseline LVLMs. We compared two baselines that eliminate language priors of the LVLMs in the decoding stage to alleviate hallucinations: VCD [43] and Opera [31]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 411, + 296, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 296, + 590 + ], + "type": "text", + "content": "Implementation Details. In our experiments, we utilized 1,500 QA pairs with \"Yes\" responses from the COCO Random subset of the POPE dataset to train the intervention shift vector. Subsequently, we evaluated the ICT approach on two datasets with significant distributional differences: POPE, and MME. This evaluation was conducted to assess the generalizability and robustness of ICT across diverse data distributions. For each attention head, we employed Support Vector Machines (SVMs) [15] as the classifier and performed 2-fold cross-validation to evaluate the classification accuracy. In the experiment, we set " + }, + { + "bbox": [ + 55, + 411, + 296, + 590 + ], + "type": "inline_equation", + "content": "\\alpha = \\beta" + }, + { + "bbox": [ + 55, + 411, + 296, + 590 + ], + "type": "text", + "content": " and determined the optimal values of " + }, + { + "bbox": [ + 55, + 411, + 296, + 590 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 411, + 296, + 590 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 411, + 296, + 590 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 411, + 296, + 590 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 411, + 296, + 590 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 411, + 296, + 590 + ], + "type": "text", + "content": " through a grid search. Detailed hyperparameter configurations are provided in Appendix A. All experiments were conducted on a system equipped with " + }, + { + "bbox": [ + 55, + 411, + 296, + 590 + ], + "type": "inline_equation", + "content": "8 \\times" + }, + { + "bbox": [ + 55, + 411, + 296, + 590 + ], + "type": "text", + "content": " H800 GPUs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 599, + 141, + 610 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 599, + 141, + 610 + ], + "spans": [ + { + "bbox": [ + 55, + 599, + 141, + 610 + ], + "type": "text", + "content": "5.2. Main Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 616, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 295, + 715 + ], + "type": "text", + "content": "Results on POPE. Table 1 presents the results of LLaVA-v1.5 and Qwen-VL on nine subsets of the POPE dataset. From comparing these methods, we can derive the following conclusions: 1) Applying ICT results in an average performance improvement on the F1 score for LLaVA-v1.5 and Qwen-VL of " + }, + { + "bbox": [ + 55, + 616, + 295, + 715 + ], + "type": "inline_equation", + "content": "7.09\\%" + }, + { + "bbox": [ + 55, + 616, + 295, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 616, + 295, + 715 + ], + "type": "inline_equation", + "content": "5.44\\%" + }, + { + "bbox": [ + 55, + 616, + 295, + 715 + ], + "type": "text", + "content": " on 9 subsets respectively. This improvement can be attributed to the fact that ICT does not eliminate language priors, which may provide" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 555, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 301 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 301 + ], + "type": "text", + "content": "useful information. Instead, it enhances the model's attention to various levels of visual information, thereby reducing the model's tendency to overly rely on language priors and mitigating the occurrence of hallucinations. 2) Individually intervening at both the image level and object level, with average gains of " + }, + { + "bbox": [ + 313, + 72, + 555, + 301 + ], + "type": "inline_equation", + "content": "5.76\\%" + }, + { + "bbox": [ + 313, + 72, + 555, + 301 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 555, + 301 + ], + "type": "inline_equation", + "content": "5.47\\%" + }, + { + "bbox": [ + 313, + 72, + 555, + 301 + ], + "type": "text", + "content": " in F1 score, respectively. This demonstrates that enhancing LVLMs' attention to image information at various levels can effectively mitigate hallucinations. Moreover, since Object-Level intervention also implicitly directs the model's focus toward broader image information, it achieves a comparatively higher performance gain. 3) The intervention shift vector trained using 1,500 samples from the MSCOCO random subset achieves an average improvement on F1 score of " + }, + { + "bbox": [ + 313, + 72, + 555, + 301 + ], + "type": "inline_equation", + "content": "7.67\\%" + }, + { + "bbox": [ + 313, + 72, + 555, + 301 + ], + "type": "text", + "content": " on MSCOCO random and an average improvement of " + }, + { + "bbox": [ + 313, + 72, + 555, + 301 + ], + "type": "inline_equation", + "content": "6.09\\%" + }, + { + "bbox": [ + 313, + 72, + 555, + 301 + ], + "type": "text", + "content": " across the other eight subsets with different distributions. This indicates that ICT generalizes effectively, and the resulting intervention vector captures a general direction towards trustworthiness rather than fitting a specific dataset." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 320, + 556, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 320, + 556, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 320, + 556, + 548 + ], + "type": "text", + "content": "Results on MME. Figure 3 presents the experimental results on the MME benchmark, where we followed Leng et al. [43] evaluation setup, focusing on the hallucination subset of MME. Compared to the unmodified model, ICT achieved a score improvement of 80.51 on LLaVA-v1.5 and 34.67 on Qwen-VL, surpassing the performance of VCD. Additionally, on non-hallucination tasks such as commonsense QA, ICT demonstrated a 10-point improvement on LLaVA-v1.5 and a 9.5-point improvement on Qwen-VL. These results indicate that ICT not only effectively reduces hallucinations but also enhances general reasoning capabilities across different models. Moreover, since the objects present in the images of the MME dataset are generally prominent, the variant ICT " + }, + { + "bbox": [ + 313, + 320, + 556, + 548 + ], + "type": "inline_equation", + "content": "w/o" + }, + { + "bbox": [ + 313, + 320, + 556, + 548 + ], + "type": "text", + "content": " image provides only limited improvements compared to ICT " + }, + { + "bbox": [ + 313, + 320, + 556, + 548 + ], + "type": "inline_equation", + "content": "w/o" + }, + { + "bbox": [ + 313, + 320, + 556, + 548 + ], + "type": "text", + "content": " object. This demonstrates that the absence of Image-Level guidance restricts the model's ability to fully leverage global visual context, thereby reducing its effectiveness in scenarios where the dominant objects play a crucial role." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 316, + 567, + 553, + 680 + ], + "blocks": [ + { + "bbox": [ + 316, + 567, + 553, + 680 + ], + "lines": [ + { + "bbox": [ + 316, + 567, + 553, + 680 + ], + "spans": [ + { + "bbox": [ + 316, + 567, + 553, + 680 + ], + "type": "image", + "image_path": "83db3db20d370ddf3feed3811612747a8a58a4c6ea5e3e8026b7fa1600971d08.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 683, + 556, + 727 + ], + "lines": [ + { + "bbox": [ + 313, + 683, + 556, + 727 + ], + "spans": [ + { + "bbox": [ + 313, + 683, + 556, + 727 + ], + "type": "text", + "content": "Figure 3. Comparison of ICT with baseline methods (Vanilla and VCD) on the MME benchmark. The radar chart illustrates improvements across various evaluation categories, including existence, position, count, color, and commonsense QA (CSQA)." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4213" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 93, + 69, + 519, + 651 + ], + "blocks": [ + { + "bbox": [ + 93, + 69, + 519, + 651 + ], + "lines": [ + { + "bbox": [ + 93, + 69, + 519, + 651 + ], + "spans": [ + { + "bbox": [ + 93, + 69, + 519, + 651 + ], + "type": "table", + "html": "
DatasetSettingMethodLLaVA-v1.5Qwen-VL
AccuracyF1 ScoreAccuracyF1 Score
COCORandomRegular83.2981.3384.3782.67
VCD87.7387.1688.6387.81
OPERIA89.2088.8187.3186.92
ICT w/o image89.70 (+6.41)89.93 (+8.60)88.76(+4.39)87.84(+5.17)
ICT w/o object88.50 (+5.21)88.81 (+7.48)89.13(+4.76)88.20(+5.63)
ICT90.11 (+6.82)90.03 (+8.70)89.46(+5.09)89.20(+6.63)
PopularRegular81.8880.0684.1382.06
VCD85.3885.0687.1286.40
OPERIA86.6486.6287.4486.68
ICT w/o image87.43 (+5.55)87.09 (+7.03)87.63(+3.50)86.78(+4.72)
ICT w/o object87.30 (+5.42)86.65 (+6.59)87.10(+2.97)86.28(+4.22)
ICT87.50 (+5.62)87.60 (+7.54)88.16(+4.03)87.33(+5.27)
AdversarialRegular78.9677.5782.2680.37
VCD80.8881.3384.2683.90
OPERIA81.2481.3884.7883.45
ICT w/o image84.13 (+5.17)83.31 (+5.74)85.13(+2.87)84.35(+3.98)
ICT w/o object83.63 (+4.67)82.13 (+4.56)85.23(+2.97)84.62(+4.25)
ICT84.43 (+5.47)83.74 (+6.17)84.96(+2.70)84.42(+4.05)
AOKVQARandomRegular83.4582.5686.6785.59
VCD86.1586.3489.2289.01
OPERIA88.0284.5988.1988.43
ICT w/o image89.60 (+6.15)89.75 (+7.19)89.46(+2.79)89.28(+3.69)
ICT w/o object89.00 (+5.55)88.91 (+6.35)89.26(+2.59)88.95(+3.36)
ICT89.20 (+5.75)89.41 (+6.85)89.46(+2.79)89.03(+3.44)
PopularRegular79.9079.5985.5684.63
VCD81.8582.8287.8587.81
OPERIA83.2284.6787.9187.13
ICT w/o image85.36 (+5.46)85.34 (+5.75)87.70(+2.14)87.68(+3.05)
ICT w/o object84.65 (+4.75)84.32 (+4.73)87.90(+2.34)87.48(+2.85)
ICT85.73 (+5.83)85.34 (+5.75)88.13(+3.57)87.83(+3.20)
AdversarialRegular74.0475.1579.5779.50
VCD74.9777.7381.2782.38
OPERIA73.8277.9180.8281.54
ICT w/o image79.26 (+5.22)80.14 (+4.99)81.50(+1.93)82.10(+2.60)
ICT w/o object77.79 (+3.75)79.53 (+4.38)81.40(+1.83)82.18(+2.68)
ICT79.60 (+5.56)80.43 (+5.28)81.94(+2.37)82.44(+2.94)
GQARandomRegular83.7382.9580.9779.01
VCD86.6586.9985.5985.33
OPERIA88.1388.9186.0285.29
ICT w/o image89.03 (+5.30)88.99 (+6.04)85.20(+4.23)85.90(+6.89)
ICT w/o object88.97 (+5.24)89.30 (+6.35)87.20(+6.23)86.80(+7.79)
ICT89.60 (+5.87)89.44 (+6.49)86.38(+5.41)86.96(+7.95)
PopularRegular78.1778.3775.9974.84
VCD80.7382.2481.8382.23
OPERIA79.2782.1181.9782.12
ICT w/o image84.23 (+6.06)83.92 (+5.55)81.43(+5.44)80.86(+6.02)
ICT w/o object84.70 (+6.53)84.45 (+6.08)83.93(+7.94)82.95(+8.11)
ICT84.70 (+6.53)84.78 (+6.41)82.63(+6.64)82.22(+7.38)
AdversarialRegular75.0876.0675.4674.33
VCD76.0978.7880.0180.75
OPERIA75.0078.7180.2480.64
ICT w/o image81.92 (+6.84)82.11 (+6.05)80.03(+4.57)79.90(+5.47)
ICT w/o object81.65 (+6.57)82.00 (+5.94)81.60(+6.14)81.53(+7.10)
ICT81.50 (+6.42)82.27 (+6.21)80.83(+5.37)80.60(+6.17)
", + "image_path": "07bf9695ea90362d813b7e1a9618e879f909cb9861e7b977c62fb3787f4dbee2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 657, + 555, + 703 + ], + "lines": [ + { + "bbox": [ + 55, + 657, + 555, + 703 + ], + "spans": [ + { + "bbox": [ + 55, + 657, + 555, + 703 + ], + "type": "text", + "content": "Table 1. Main results on POPE tasks. We evaluate the accuracy of various LVLMs on the POPE task across the MSCOCO, A-OKVQA, and GQA datasets. Regular represents the setting where direct sampling is applied. ICT " + }, + { + "bbox": [ + 55, + 657, + 555, + 703 + ], + "type": "inline_equation", + "content": "w/o" + }, + { + "bbox": [ + 55, + 657, + 555, + 703 + ], + "type": "text", + "content": " image and ICT " + }, + { + "bbox": [ + 55, + 657, + 555, + 703 + ], + "type": "inline_equation", + "content": "w/o" + }, + { + "bbox": [ + 55, + 657, + 555, + 703 + ], + "type": "text", + "content": " object correspond to the exclusion of image-level and object-level interventions, respectively. The bold and underlined values indicate the highest and second-highest metrics under each setting, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4214" + } + ] + } + ], + "index": 2 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 198, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 198, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 198, + 85 + ], + "type": "text", + "content": "6. Analysis and Discussions" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 284, + 104 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 284, + 104 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 284, + 104 + ], + "type": "text", + "content": "6.1. Inference Latency across Different Methods" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 108, + 295, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 108, + 295, + 228 + ], + "spans": [ + { + "bbox": [ + 54, + 108, + 295, + 228 + ], + "type": "text", + "content": "In Table 2, we compare the inference efficiency of two different representative approaches with our proposed method. Notably, while the CD-based methods result in a significant increase in inference time (2.1 to 3.6 times slower), our ICT incurs virtually no impact on inference efficiency. We attribute this to the fact that both VCD and OPERA either require multiple inference runs or involve substantial additional computations during inference. In contrast, ICT introduces no extra computational cost at inference, further highlighting the advantages of our method." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 59, + 230, + 291, + 294 + ], + "blocks": [ + { + "bbox": [ + 59, + 230, + 291, + 294 + ], + "lines": [ + { + "bbox": [ + 59, + 230, + 291, + 294 + ], + "spans": [ + { + "bbox": [ + 59, + 230, + 291, + 294 + ], + "type": "table", + "html": "
Method20-Token Len50-Token Len80-Token Len
LLaVA-v1.5405.3 ↑×1.0934.6 ↑×1.01440.0 ↑×1.0
+ VCD988.9 ↑ × 2.42031.7 ↑ × 2.23077.5 ↑ × 2.1
+ OPERA1371.7 ↑ × 3.43294.1 ↑ × 3.55717.3 ↑ × 3.6
+ ICT415.9 ↑ × 1.0931.9 ↑ × 1.01485.5 ↑ × 1.0
", + "image_path": "536e03754b19c21243e41041bd5458be0c00f6d7983b662c8db62dab90bfca8a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 295, + 295, + 327 + ], + "lines": [ + { + "bbox": [ + 55, + 295, + 295, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 295, + 295, + 327 + ], + "type": "text", + "content": "Table 2. Comparison of the efficiency of different methods in generating tokens of varying lengths on an NVIDIA H800 GPU. Inference times are recorded in milliseconds." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 334, + 277, + 347 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 334, + 277, + 347 + ], + "spans": [ + { + "bbox": [ + 55, + 334, + 277, + 347 + ], + "type": "text", + "content": "6.2. Generalizability between Different Models" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 352, + 295, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 352, + 295, + 661 + ], + "spans": [ + { + "bbox": [ + 56, + 352, + 295, + 661 + ], + "type": "text", + "content": "To verify that the derived activation shift vectors to some extent represent the model's alignment with authenticity, we explored whether these shifts are transferable to other models. Specifically, we applied the activation shift vectors obtained from LLaVA-v1.5 on the COCO random subset to intervene in Qwen-VL. We then evaluated the generalizability of this intervention on the GQA dataset, which has a distribution significantly different from COCO. The results in Table 3 demonstrate that ICT-LLaVA-v1.5 achieves an average improvement of " + }, + { + "bbox": [ + 56, + 352, + 295, + 661 + ], + "type": "inline_equation", + "content": "4.62\\%" + }, + { + "bbox": [ + 56, + 352, + 295, + 661 + ], + "type": "text", + "content": " in F1 score and " + }, + { + "bbox": [ + 56, + 352, + 295, + 661 + ], + "type": "inline_equation", + "content": "4.64\\%" + }, + { + "bbox": [ + 56, + 352, + 295, + 661 + ], + "type": "text", + "content": " in accuracy compared to the unmodified model. Additionally, we employed t-SNE to visualize the offset vectors across different layers for LLaVA-v1.5 and Qwen-VL. Figure 4 shows the Object-Level and Image-Level shift vectors of multi-head attention in the 16th and 18th layers for both models. From the figure, we can observe that in layer 16, the Image-Level vectors (blue and yellow) are relatively close to each other, indicating a degree of similarity in encoding Image-Level information between the models. However, the Object-Level vectors (red and green) are more distinct, reflecting model-specific characteristics in encoding fine-grained object details. In layer 18, the Image-Level and Object-Level vectors for both models exhibit high similarity, indicating that the resulting activation shift vectors represent universal shift vectors that guide the model to pay more attention to visual information." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 661, + 288, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 661, + 288, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 661, + 288, + 673 + ], + "type": "text", + "content": "6.3. Impact of Hyperparameters on Performance" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "Our proposed ICT primarily relies on two key hyperparameters: the intervention intensity " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": " and the number of heads " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": " involved in the intervention. To investigate the" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 71, + 553, + 160 + ], + "blocks": [ + { + "bbox": [ + 316, + 71, + 553, + 160 + ], + "lines": [ + { + "bbox": [ + 316, + 71, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 316, + 71, + 553, + 160 + ], + "type": "image", + "image_path": "12ba79e155dd7a4d0e372b6e33b40098d7e630c396520c34f5c5c0ab79bfac0c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 163, + 555, + 402 + ], + "lines": [ + { + "bbox": [ + 313, + 163, + 555, + 402 + ], + "spans": [ + { + "bbox": [ + 313, + 163, + 555, + 402 + ], + "type": "text", + "content": "Figure 4. t-SNE visualization of Object-Level and Image-Level offset vectors for LLaVA-v1.5 and Qwen-VL at layers 16 and 18. Impact of these hyperparameters on performance, we fixed one of the parameters and calculated the average F1 score for LLaVA-v1.5 and Qwen-VL on the POPE COCO random subset. The results are presented in Figure 6, we can observe that when " + }, + { + "bbox": [ + 313, + 163, + 555, + 402 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 163, + 555, + 402 + ], + "type": "text", + "content": " is negative—indicating that reverse intervention reduces the model's attention to visual information—the model's performance declines. When " + }, + { + "bbox": [ + 313, + 163, + 555, + 402 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 163, + 555, + 402 + ], + "type": "text", + "content": " is too small, the model does not receive sufficient intervention, resulting in suboptimal outcomes. Conversely, when " + }, + { + "bbox": [ + 313, + 163, + 555, + 402 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 163, + 555, + 402 + ], + "type": "text", + "content": " is excessively large, the intervention becomes too strong, disrupting the model's foundational capabilities and leading to a decrease in performance. For the hyperparameter " + }, + { + "bbox": [ + 313, + 163, + 555, + 402 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 163, + 555, + 402 + ], + "type": "text", + "content": ", we find that when " + }, + { + "bbox": [ + 313, + 163, + 555, + 402 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 163, + 555, + 402 + ], + "type": "text", + "content": " is too small, certain attention heads that encode relevant visual information are not adequately intervened upon, resulting in suboptimal performance. Conversely, when " + }, + { + "bbox": [ + 313, + 163, + 555, + 402 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 163, + 555, + 402 + ], + "type": "text", + "content": " is too large, attention heads that encode irrelevant information are unnecessarily interfered with, which leads to performance degradation." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 413, + 555, + 440 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 555, + 440 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 555, + 440 + ], + "type": "text", + "content": "6.4. Analysis of Attention Heads for Visual Information Encoding" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 443, + 555, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 443, + 555, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 443, + 555, + 515 + ], + "type": "text", + "content": "In Figure 7, we present the classification accuracy of the binary classifiers for Image-Level and Object-Level features across the 1,024 multi-head attention heads in the 32 layers of LLaVA-v1.5. This analysis helps identify which heads encode overall visual information and which heads capture fine-grained visual details. From the figure, we can observe" + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 332, + 521, + 537, + 672 + ], + "blocks": [ + { + "bbox": [ + 332, + 521, + 537, + 672 + ], + "lines": [ + { + "bbox": [ + 332, + 521, + 537, + 672 + ], + "spans": [ + { + "bbox": [ + 332, + 521, + 537, + 672 + ], + "type": "table", + "html": "
SettingMethodAccuracyF1 Score
RandomRegular80.9779.01
VCD85.5985.33
ICT86.3886.96
ICT-LLaVA-v1.585.1083.27
PopularRegular75.9974.84
VCD81.8382.23
ICT82.6382.22
ICT-LLaVA-v1.581.5080.10
AdversarialRegular75.4674.33
VCD80.0180.75
ICT80.8380.60
ICT-LLaVA-v1.579.7378.68
", + "image_path": "f8beef3d8e631cb91400e7143775abb381658953d1dfb35fee4da509b3839fae.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 680, + 556, + 724 + ], + "lines": [ + { + "bbox": [ + 313, + 680, + 556, + 724 + ], + "spans": [ + { + "bbox": [ + 313, + 680, + 556, + 724 + ], + "type": "text", + "content": "Table 3. Cross-model generalization of ICT. Performance of Qwen-VL on GQA after applying activation shift from LLaVA-v1.5. ICT-LLaVA-v1.5 refers to the results obtained by applying LLaVA-v1.5's activation shift vectors to intervene in Qwen-VL." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4215" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 115, + 86, + 217, + 163 + ], + "blocks": [ + { + "bbox": [ + 115, + 86, + 217, + 163 + ], + "lines": [ + { + "bbox": [ + 115, + 86, + 217, + 163 + ], + "spans": [ + { + "bbox": [ + 115, + 86, + 217, + 163 + ], + "type": "image", + "image_path": "061fcca9a3ee2c52e18822df4fdfafda5ac52b8f646caf5be60c6a3657ab698f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 238, + 120, + 313, + 129 + ], + "lines": [ + { + "bbox": [ + 238, + 120, + 313, + 129 + ], + "spans": [ + { + "bbox": [ + 238, + 120, + 313, + 129 + ], + "type": "text", + "content": "VCD: Yes, there is." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 232, + 147, + 319, + 156 + ], + "lines": [ + { + "bbox": [ + 232, + 147, + 319, + 156 + ], + "spans": [ + { + "bbox": [ + 232, + 147, + 319, + 156 + ], + "type": "text", + "content": "ICT: No, there isn't." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 216, + 176, + 323, + 196 + ], + "lines": [ + { + "bbox": [ + 216, + 176, + 323, + 196 + ], + "spans": [ + { + "bbox": [ + 216, + 176, + 323, + 196 + ], + "type": "text", + "content": "Question: How many uncut fruits are in the image?" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 340, + 95, + 477, + 163 + ], + "blocks": [ + { + "bbox": [ + 231, + 89, + 317, + 110 + ], + "lines": [ + { + "bbox": [ + 231, + 89, + 317, + 110 + ], + "spans": [ + { + "bbox": [ + 231, + 89, + 317, + 110 + ], + "type": "text", + "content": "Question: Is there a horse in the image?" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 396, + 84, + 419, + 93 + ], + "lines": [ + { + "bbox": [ + 396, + 84, + 419, + 93 + ], + "spans": [ + { + "bbox": [ + 396, + 84, + 419, + 93 + ], + "type": "text", + "content": "POPE" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 340, + 95, + 477, + 163 + ], + "lines": [ + { + "bbox": [ + 340, + 95, + 477, + 163 + ], + "spans": [ + { + "bbox": [ + 340, + 95, + 477, + 163 + ], + "type": "image", + "image_path": "0a69dc51e3541fb53d1e1ad5e1efc87d97a10cfbe2cf9560213e69c063dd809a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 115, + 171, + 204, + 260 + ], + "blocks": [ + { + "bbox": [ + 115, + 171, + 204, + 260 + ], + "lines": [ + { + "bbox": [ + 115, + 171, + 204, + 260 + ], + "spans": [ + { + "bbox": [ + 115, + 171, + 204, + 260 + ], + "type": "image", + "image_path": "44463fb48a92b5fe147d9e585a798d28a936dfd3835b536e7f985137d1e2c41c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 216, + 205, + 323, + 224 + ], + "lines": [ + { + "bbox": [ + 216, + 205, + 323, + 224 + ], + "spans": [ + { + "bbox": [ + 216, + 205, + 323, + 224 + ], + "type": "text", + "content": "VCD: There are two uncut fruits in the image." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 217, + 235, + 325, + 255 + ], + "lines": [ + { + "bbox": [ + 217, + 235, + 325, + 255 + ], + "spans": [ + { + "bbox": [ + 217, + 235, + 325, + 255 + ], + "type": "text", + "content": "ICT: There are four uncut fruits in the image." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 261, + 555, + 283 + ], + "lines": [ + { + "bbox": [ + 55, + 261, + 555, + 283 + ], + "spans": [ + { + "bbox": [ + 55, + 261, + 555, + 283 + ], + "type": "text", + "content": "Figure 5. Case Study and Error Analysis of ICT. The vertical axis in the right section of the figure represents the attention ratio of each question text token to the visual tokens." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 340, + 194, + 485, + 256 + ], + "blocks": [ + { + "bbox": [ + 384, + 180, + 442, + 190 + ], + "lines": [ + { + "bbox": [ + 384, + 180, + 442, + 190 + ], + "spans": [ + { + "bbox": [ + 384, + 180, + 442, + 190 + ], + "type": "text", + "content": "LLaVA-Bench" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 340, + 194, + 485, + 256 + ], + "lines": [ + { + "bbox": [ + 340, + 194, + 485, + 256 + ], + "spans": [ + { + "bbox": [ + 340, + 194, + 485, + 256 + ], + "type": "image", + "image_path": "d00f3d20228a6c7d87d26059a11fbdfab60dd6780df7a954d3f4ac14926768cc.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 56, + 286, + 173, + 380 + ], + "blocks": [ + { + "bbox": [ + 56, + 286, + 173, + 380 + ], + "lines": [ + { + "bbox": [ + 56, + 286, + 173, + 380 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 173, + 380 + ], + "type": "image", + "image_path": "b1db38f81147a4e3fbbc8512ae0b974d3abe3653d47b72ee36fedf81b1ae1c8b.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 385, + 295, + 417 + ], + "lines": [ + { + "bbox": [ + 55, + 385, + 295, + 417 + ], + "spans": [ + { + "bbox": [ + 55, + 385, + 295, + 417 + ], + "type": "text", + "content": "Figure 6. Effect of hyperparameters " + }, + { + "bbox": [ + 55, + 385, + 295, + 417 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 385, + 295, + 417 + ], + "type": "text", + "content": " (left) and " + }, + { + "bbox": [ + 55, + 385, + 295, + 417 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 385, + 295, + 417 + ], + "type": "text", + "content": " (right) on the F1 score for LLaVA-v1.5 and Qwen-VL models on the POPE COCO random subset." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 173, + 286, + 294, + 380 + ], + "blocks": [ + { + "bbox": [ + 173, + 286, + 294, + 380 + ], + "lines": [ + { + "bbox": [ + 173, + 286, + 294, + 380 + ], + "spans": [ + { + "bbox": [ + 173, + 286, + 294, + 380 + ], + "type": "image", + "image_path": "8158a5900d2bc9350a72d94b857f4a48acf733e6a778f887c09e0b05cc429b16.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 419, + 296, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 419, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 419, + 296, + 479 + ], + "type": "text", + "content": "that the attention heads encoding overall visual information are predominantly located in the earlier layers of the model, such as layers 4 and 10. In contrast, the attention heads that capture fine-grained visual details are more concentrated in the later layers, such as layers 20 and 22." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 490, + 224, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 490, + 224, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 490, + 224, + 502 + ], + "type": "text", + "content": "6.5. Case Study and Error Analysis" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 508, + 296, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 508, + 296, + 580 + ], + "spans": [ + { + "bbox": [ + 55, + 508, + 296, + 580 + ], + "type": "text", + "content": "In Figure 5, we present a case study on the POPE and LLaVA-Bench datasets, visualizing the attention ratio of each question text token to the visual tokens. As illustrated, after applying ICT, the model allocates a higher proportion of attention to visual tokens, especially to object tokens relevant to the question (e.g., \"horse\" and \"fruits\"). By prior-" + } + ] + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 56, + 599, + 294, + 706 + ], + "blocks": [ + { + "bbox": [ + 56, + 599, + 294, + 706 + ], + "lines": [ + { + "bbox": [ + 56, + 599, + 294, + 706 + ], + "spans": [ + { + "bbox": [ + 56, + 599, + 294, + 706 + ], + "type": "image", + "image_path": "7dbdda3e10270cab46ddf2f6629e41aba07c1ce1f8ba4279fa920d77be7fcb7d.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 711, + 295, + 733 + ], + "lines": [ + { + "bbox": [ + 55, + 711, + 295, + 733 + ], + "spans": [ + { + "bbox": [ + 55, + 711, + 295, + 733 + ], + "type": "text", + "content": "Figure 7. Heatmap of classification accuracy for Image-Level (left) and Object-Level (right) features of LLaVA-v1.5." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 288, + 555, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 288, + 555, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 555, + 431 + ], + "type": "text", + "content": "itizing visual information, ICT correctly identifies the absence of a horse in the image, whereas VCD erroneously concludes that a horse is present due to insufficient attention to visual cues. However, when asked, \"How many uncut fruits are in the picture?\", VCD incorrectly answered \"two\" due to a lack of focus on visual details. Although ICT correctly identifies that there are a total of four fruits in the image, this question requires not only attention to visual content but also reasoning within the text modality. The model needs to not only recognize the overall count of fruits but also focus on the attribute \"uncut\". Due to its failure to incorporate this information, ICT outputs a wrong answer." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 433, + 556, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 433, + 556, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 433, + 556, + 552 + ], + "type": "text", + "content": "Furthermore, when the text becomes more complex, such as in \"How many uncut fruits are there in the image?\", the increased linguistic complexity demands greater attention from the model. This, in turn, limits the extent to which attention can be allocated to the visual domain. In contrast, function words like \"is\" and \"there\" are less complex and require minimal processing, allowing the model to shift more focus toward visual tokens. As a result, this redistribution of attention can lead to a more noticeable shift in how the model processes visual and textual information." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 314, + 572, + 471, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 572, + 471, + 585 + ], + "spans": [ + { + "bbox": [ + 314, + 572, + 471, + 585 + ], + "type": "text", + "content": "7. Conclusion and Limitations" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "text", + "content": "In this paper, we mitigate the issue of object hallucinations in LVLMs by proposing Image-Object Cross-Level Trusted Intervention (ICT), a training-free, plug-and-play method that enhances the model's focus on both Image-Level and Object-Level visual information during the forward pass. Our experiments show that ICT significantly reduces hallucinations while also improving general reasoning capabilities. The results confirm that ICT effectively mitigates excessive reliance on language priors, enhancing both accuracy and robustness in diverse visual contexts." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4216" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 297, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 297, + 202 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 297, + 202 + ], + "type": "text", + "content": "This work was supported by the National Key Research and Development Program of China (Grant No. 2024YFB3309702); Guangdong Provincial Department of Education Project (Grant No.2024KQNCX028); CAAI-Ant Group Research Fund; Scientific Research Projects for the Higher-educational Institutions (Grant No.2024312096), Education Bureau of Guangzhou Municipality; Guangzhou-HKUST(GZ) Joint Funding Program (Grant No.2025A03J3957), Education Bureau of Guangzhou Municipality." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 223, + 115, + 236 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 223, + 115, + 236 + ], + "spans": [ + { + "bbox": [ + 56, + 223, + 115, + 236 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 243, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 66, + 243, + 294, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 243, + 294, + 287 + ], + "spans": [ + { + "bbox": [ + 66, + 243, + 294, + 287 + ], + "type": "text", + "content": "[1] Renat Aksitov, Chung-Ching Chang, David Reitter, Siamak Shakeri, and Yunhsuan Sung. Characterizing attribution and fluency tradeoffs for retrieval-augmented large language models, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 289, + 295, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 289, + 295, + 354 + ], + "spans": [ + { + "bbox": [ + 66, + 289, + 295, + 354 + ], + "type": "text", + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: A visual language model for few-shot learning. In Advances in Neural Information Processing Systems, pages 23716-23736, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 66, + 355, + 295, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 355, + 295, + 409 + ], + "spans": [ + { + "bbox": [ + 66, + 355, + 295, + 409 + ], + "type": "text", + "content": "[3] Wenbin An, Feng Tian, Sicong Leng, Jiahao Nie, Haonan Lin, QianYing Wang, Guang Dai, Ping Chen, and Shijian Lu. Agla: Mitigating object hallucinations in large vision-language models with assembly of global and local attention. arXiv preprint, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 66, + 411, + 294, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 411, + 294, + 465 + ], + "spans": [ + { + "bbox": [ + 66, + 411, + 294, + 465 + ], + "type": "text", + "content": "[4] Wenbin An, Feng Tian, Sicong Leng, Jiahao Nie, Haonan Lin, QianYing Wang, Guang Dai, Ping Chen, and Shijian Lu. Agla: Mitigating object hallucinations in large vision-language models with assembly of global and local attention. arXiv preprint arXiv:2406.12718, 2024. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 66, + 468, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 468, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 66, + 468, + 294, + 521 + ], + "type": "text", + "content": "[5] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. 2023. 1, 2, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 66, + 523, + 294, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 523, + 294, + 588 + ], + "spans": [ + { + "bbox": [ + 66, + 523, + 294, + 588 + ], + "type": "text", + "content": "[6] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D. Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In Advances in Neural Information Processing Systems, pages 1877-1901, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 66, + 590, + 294, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 590, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 66, + 590, + 294, + 633 + ], + "type": "text", + "content": "[7] Beitao Chen, Xinyu Lyu, Lianli Gao, Jingkuan Song, and Heng Tao Shen. Alleviating hallucinations in large vision-language models through hallucination-induced optimization, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 66, + 635, + 294, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 635, + 294, + 666 + ], + "spans": [ + { + "bbox": [ + 66, + 635, + 294, + 666 + ], + "type": "text", + "content": "[8] Haozhe Chen, Carl Vondrick, and Chengzhi Mao. Selfie: Self-interpretation of large language model embeddings, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 66, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 670, + 294, + 713 + ], + "type": "text", + "content": "[9] Zhaorun Chen, Zhuokai Zhao, Hongyin Luo, Huaxiu Yao, Bo Li, and Jiawei Zhou. Halc: Object hallucination reduction via adaptive focal-contrast decoding. In *Forty-first International Conference on Machine Learning*. 1" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 320, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 553, + 116 + ], + "type": "text", + "content": "[10] Zhaorun Chen, Zhuokai Zhao, Hongyin Luo, Huaxiu Yao, Bo Li, and Jiawei Zhou. Halc: Object hallucination reduction via adaptive focal-contrast decoding, 2024. http://arxiv.org/abs/2403.00425.2,3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 320, + 118, + 553, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 118, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 320, + 118, + 553, + 182 + ], + "type": "text", + "content": "[11] I.-Chun Chern, Steffi Chern, Shiqi Chen, Weizhe Yuan, Kehua Feng, Chunting Zhou, Junxian He, Graham Neubig, and Pengfei Liu. Factool: Factuality detection in generative ai - a tool augmented framework for multi-task and multi-domain scenarios, 2023. http://arxiv.org/abs/2307.13528.3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 320, + 184, + 553, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 184, + 553, + 248 + ], + "spans": [ + { + "bbox": [ + 320, + 184, + 553, + 248 + ], + "type": "text", + "content": "[12] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with " + }, + { + "bbox": [ + 320, + 184, + 553, + 248 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 320, + 184, + 553, + 248 + ], + "type": "text", + "content": " * chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2(3):6, 2023. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 250, + 553, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 250, + 553, + 303 + ], + "spans": [ + { + "bbox": [ + 320, + 250, + 553, + 303 + ], + "type": "text", + "content": "[13] Jaemin Cho, Seunghyun Yoon, Ajinkya Kale, Franck Dernoncourt, Trung Bui, and Mohit Bansal. Fine-grained image captioning with clip reward. In Findings of the Association for Computational Linguistics: NAACL 2022, pages 517-527, 2022. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 304, + 553, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 304, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 320, + 304, + 553, + 358 + ], + "type": "text", + "content": "[14] Yung-Sung Chuang, Yujia Xie, Hongyin Luo, Yoon Kim, James Glass, and Pengcheng He. Dola: Decoding by contrasting layers improves factuality in large language models, 2024. http://arxiv.org/abs/2309.03883.2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 360, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 360, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 320, + 360, + 553, + 381 + ], + "type": "text", + "content": "[15] Corinna Cortes. Support-vector networks. Machine Learning, 1995. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 383, + 553, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 383, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 320, + 383, + 553, + 425 + ], + "type": "text", + "content": "[16] Li D., Rawat A.S., Zaheer M., Wang X., Lukasik M., Veit A., Yu F., and Kumar S. Large language models with controllable working memory. arXiv preprint arXiv:2211.05110, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 426, + 553, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 426, + 553, + 479 + ], + "spans": [ + { + "bbox": [ + 320, + 426, + 553, + 479 + ], + "type": "text", + "content": "[17] Hanxing Ding, Liang Pang, Zihao Wei, Huawei Shen, and Xueqi Cheng. Retrieve only when it needs: Adaptive retrieval augmentation for hallucination mitigation in large language models, 2024. http://arxiv.org/abs/2402.10612.3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 482, + 553, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 482, + 553, + 523 + ], + "spans": [ + { + "bbox": [ + 320, + 482, + 553, + 523 + ], + "type": "text", + "content": "[18] Sebastian Farquhar, Jannik Kossen, Lorenz Kuhn, and Yarin Gal. Detecting hallucinations in large language models using semantic entropy. Nature, 630(8017):625-630, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 525, + 553, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 525, + 553, + 558 + ], + "spans": [ + { + "bbox": [ + 320, + 525, + 553, + 558 + ], + "type": "text", + "content": "[19] Javier Ferrando, Gabriele Sarti, Arianna Bisazza, and Marta R. Costa-jussà. A primer on the inner workings of transformer-based language models, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 559, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 559, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 320, + 559, + 553, + 613 + ], + "type": "text", + "content": "[20] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023.5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 614, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 614, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 320, + 614, + 553, + 690 + ], + "type": "text", + "content": "[21] Suriya Gunasekar, Yi Zhang, Jyoti Aneja, Cao César Teodoro Mendes, Allie Del Giorno, Sivakanth Gopi, Mojan Javaheripi, Piero Kauffmann, Gustavo de Rosa, Olli Saarikivi, Adil Salim, Shital Shah, Harkirat Singh Behl, Xin Wang, Sébastien Bubeck, Ronen Eldan, Adam Tauman Kalai, Yin Tat Lee, and Yuanzhi Li. Textbooks are all you need, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 320, + 692, + 553, + 713 + ], + "type": "text", + "content": "[22] Anisha Gunjal, Jihan Yin, and Erhan Bas. Detecting and preventing hallucinations in large vision language models." + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4217" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 81, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 81, + 73, + 294, + 95 + ], + "type": "text", + "content": "In Proceedings of the AAAI Conference on Artificial Intelligence, pages 18135-18143, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 96, + 295, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 96, + 295, + 139 + ], + "spans": [ + { + "bbox": [ + 61, + 96, + 295, + 139 + ], + "type": "text", + "content": "[23] Anisha Gunjal, Jihan Yin, and Erhan Bas. Detecting and preventing hallucinations in large vision language models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 18135-18143, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 140, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 140, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 61, + 140, + 294, + 205 + ], + "type": "text", + "content": "[24] Vipul Gupta, Zhuowan Li, Adam Kortylewski, Chenyu Zhang, Yingwei Li, and Alan Yuille. Swapmix: Diagnosing and regularizing the over-reliance on visual context in visual question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5078-5088, 2022. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 206, + 295, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 206, + 295, + 249 + ], + "spans": [ + { + "bbox": [ + 62, + 206, + 295, + 249 + ], + "type": "text", + "content": "[25] Yudong Han, Liqiang Nie, Jianhua Yin, Jianlong Wu, and Yan Yan. Visual perturbation-aware collaborative learning for overcoming the language prior problem. arXiv preprint arXiv:2207.11850, 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 251, + 295, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 251, + 295, + 282 + ], + "spans": [ + { + "bbox": [ + 62, + 251, + 295, + 282 + ], + "type": "text", + "content": "[26] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 284, + 294, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 284, + 294, + 337 + ], + "spans": [ + { + "bbox": [ + 62, + 284, + 294, + 337 + ], + "type": "text", + "content": "[27] Wenyi Hong, Weihan Wang, Ming Ding, Wenmeng Yu, Qingsong Lv, Yan Wang, Yean Cheng, Shiyu Huang, Junhui Ji, Zhao Xue, et al. Cogvlm2: Visual language models for image and video understanding. arXiv preprint arXiv:2408.16500, 2024. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 338, + 294, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 338, + 294, + 381 + ], + "spans": [ + { + "bbox": [ + 62, + 338, + 294, + 381 + ], + "type": "text", + "content": "[28] Haichuan Hu, Yuhan Sun, and Quanjun Zhang. Lrp4rag: Detecting hallucinations in retrieval-augmented generation via layer-wise relevance propagation, 2024. http:// arxiv.org/abs/2408.15533.3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 383, + 294, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 383, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 62, + 383, + 294, + 426 + ], + "type": "text", + "content": "[29] Yingdong Hu, Fanqi Lin, Tong Zhang, Li Yi, and Yang Gao. Look before you leap: Unveiling the power of gpt-4v in robotic vision-language planning. arXiv preprint arXiv:2311.17842, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 427, + 294, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 427, + 294, + 482 + ], + "spans": [ + { + "bbox": [ + 62, + 427, + 294, + 482 + ], + "type": "text", + "content": "[30] Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, and Ting Liu. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 483, + 294, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 483, + 294, + 559 + ], + "spans": [ + { + "bbox": [ + 62, + 483, + 294, + 559 + ], + "type": "text", + "content": "[31] Qidong Huang, Xiaoyi Dong, Pan Zhang, Bin Wang, Conghui He, Jiaqi Wang, Dahua Lin, Weiming Zhang, and Nenghai Yu. Opera: Alleviating hallucination in multimodal large language models via over-trust penalty and retrospection-allocation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13418-13427, 2024. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 559, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 559, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 62, + 559, + 294, + 612 + ], + "type": "text", + "content": "[32] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 614, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 614, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 62, + 614, + 294, + 657 + ], + "type": "text", + "content": "[33] Fushuo Huo, Wenchao Xu, Zhong Zhang, Haozhao Wang, Zhicheng Chen, and Peilin Zhao. Self-introspective decoding: Alleviating hallucinations for large vision-language models, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 658, + 294, + 712 + ], + "type": "text", + "content": "[34] Jitesh Jain, Jianwei Yang, and Humphrey Shi. Vcoder: Versatile vision encoders for multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27992-28002, 2024. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 320, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 553, + 116 + ], + "type": "text", + "content": "[35] Juseong Jin and Chang Wook Jeong. Surgical-llava: Toward surgical scenario understanding via large language and vision models. arXiv preprint arXiv:2410.09750, 2024. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 320, + 118, + 553, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 118, + 553, + 151 + ], + "spans": [ + { + "bbox": [ + 320, + 118, + 553, + 151 + ], + "type": "text", + "content": "[36] Junho Kim, Hyunjun Kim, Yeonju Kim, and Yong Man Ro. Code: Contrasting self-generated description to combat hallucination in large multi-modal models, 2024. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 152, + 553, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 152, + 553, + 186 + ], + "spans": [ + { + "bbox": [ + 320, + 152, + 553, + 186 + ], + "type": "text", + "content": "[37] Lorenz Kuhn, Yarin Gal, and Sebastian Farquhar. Semantic uncertainty: Linguistic invariances for uncertainty estimation in natural language generation. arXiv, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 186, + 553, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 186, + 553, + 219 + ], + "spans": [ + { + "bbox": [ + 320, + 186, + 553, + 219 + ], + "type": "text", + "content": "[38] Junlin Lee, Yequan Wang, Jing Li, and Min Zhang. Multimodal reasoning with multimodal knowledge graph. arXiv preprint arXiv:2406.02030, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 221, + 553, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 221, + 553, + 286 + ], + "spans": [ + { + "bbox": [ + 320, + 221, + 553, + 286 + ], + "type": "text", + "content": "[39] Katherine Lee, Daphne Ippolito, Andrew Nystrom, Chiyuan Zhang, Douglas Eck, Chris Callison-Burch, and Nicholas Carlini. Deduplicating training data makes language models better. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8424-8445, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 289, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 289, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 320, + 289, + 553, + 342 + ], + "type": "text", + "content": "[40] Kang-il Lee, Minbeom Kim, Seunghyun Yoon, Minsung Kim, Dongryeol Lee, Hyukhun Koh, and Kyomin Jung. Vlind-bench: Measuring language priors in large vision-language models. arXiv preprint arXiv:2406.08702, 2024. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 344, + 553, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 344, + 553, + 399 + ], + "spans": [ + { + "bbox": [ + 320, + 344, + 553, + 399 + ], + "type": "text", + "content": "[41] Nayeon Lee, Wei Ping, Peng Xu, Mostofa Patwary, Pascale N Fung, Mohammad Shoeybi, and Bryan Catanzaro. Factuality enhanced language models for open-ended text generation. In Advances in Neural Information Processing Systems, pages 34586-34599, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 400, + 553, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 400, + 553, + 433 + ], + "spans": [ + { + "bbox": [ + 320, + 400, + 553, + 433 + ], + "type": "text", + "content": "[42] Seongyun Lee, Sue Hyun Park, Yongrae Jo, and Minjoon Seo. Volcano: Mitigating multimodal hallucination through self-feedback guided revision. arXiv, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 434, + 553, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 434, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 320, + 434, + 553, + 499 + ], + "type": "text", + "content": "[43] Sicong Leng, Hang Zhang, Guanzheng Chen, Xin Li, Shijian Lu, Chunyan Miao, and Lidong Bing. Mitigating object hallucinations in large vision-language models through visual contrastive decoding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13872-13882, 2024. 1, 2, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 501, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 501, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 320, + 501, + 553, + 555 + ], + "type": "text", + "content": "[44] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In Proceedings of the International Conference on Machine Learning, pages 19730-19742. PMLR, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 557, + 553, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 557, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 320, + 557, + 553, + 601 + ], + "type": "text", + "content": "[45] Kenneth Li, Oam Patel, Fernanda Viégas, Hanspeter Pfister, and Martin Wattenberg. Inference-time intervention: Eliciting truthful answers from a language model. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 602, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 602, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 320, + 602, + 553, + 668 + ], + "type": "text", + "content": "[46] Lei Li, Zhihui Xie, Mukai Li, Shunian Chen, Peiyi Wang, Liang Chen, Yazheng Yang, Benyou Wang, Lingpeng Kong, and Qi Liu. Vlfeedback: A large-scale ai feedback dataset for large vision-language models alignment. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 6227-6246, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 320, + 670, + 553, + 713 + ], + "type": "text", + "content": "[47] Xiang Lisa Li, Ari Holtzman, Daniel Fried, Percy Liang, Jason Eisner, Tatsunori Hashimoto, Luke Zettlemoyer, and Mike Lewis. Contrastive decoding: Open-ended text generation as optimization. arXiv preprint, 2023. 3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4218" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 72, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 72, + 297, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 72, + 297, + 117 + ], + "spans": [ + { + "bbox": [ + 61, + 72, + 297, + 117 + ], + "type": "text", + "content": "[48] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. In The 2023 Conference on Empirical Methods in Natural Language Processing. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 118, + 295, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 118, + 295, + 184 + ], + "spans": [ + { + "bbox": [ + 61, + 118, + 295, + 184 + ], + "type": "text", + "content": "[49] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 185, + 295, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 185, + 295, + 218 + ], + "spans": [ + { + "bbox": [ + 62, + 185, + 295, + 218 + ], + "type": "text", + "content": "[50] Bingbin Liu, Jordan T. Ash, Surbhi Goel, Akshay Krishnamurthy, and Cyril Zhang. Exposing attention glitches with flip-flop language modeling, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 219, + 295, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 219, + 295, + 273 + ], + "spans": [ + { + "bbox": [ + 62, + 219, + 295, + 273 + ], + "type": "text", + "content": "[51] Fuxiao Liu, Kevin Lin, Linjie Li, Jianfeng Wang, Yaser Ya-coob, and Lijuan Wang. Mitigating hallucination in large multi-modal models via robust instruction tuning. In The Twelfth International Conference on Learning Representations, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 274, + 295, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 274, + 295, + 307 + ], + "spans": [ + { + "bbox": [ + 62, + 274, + 295, + 307 + ], + "type": "text", + "content": "[52] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 1, 2, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 308, + 295, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 308, + 295, + 342 + ], + "spans": [ + { + "bbox": [ + 62, + 308, + 295, + 342 + ], + "type": "text", + "content": "[53] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 342, + 295, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 342, + 295, + 387 + ], + "spans": [ + { + "bbox": [ + 62, + 342, + 295, + 387 + ], + "type": "text", + "content": "[54] Hanchao Liu, Wenyuan Xue, Yifei Chen, Dapeng Chen, Xiutian Zhao, Ke Wang, Liping Hou, Rongjun Li, and Wei Peng. A survey on hallucination in large vision-language models. arXiv preprint, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 388, + 295, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 388, + 295, + 431 + ], + "spans": [ + { + "bbox": [ + 62, + 388, + 295, + 431 + ], + "type": "text", + "content": "[55] Hanchao Liu, Wenyuan Xue, Yifei Chen, Dapeng Chen, Xiutian Zhao, Ke Wang, Liping Hou, Rongjun Li, and Wei Peng. A survey on hallucination in large vision-language models. arXiv preprint arXiv:2402.00253, 2024. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 433, + 295, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 433, + 295, + 487 + ], + "spans": [ + { + "bbox": [ + 62, + 433, + 295, + 487 + ], + "type": "text", + "content": "[56] Jiajun Liu, Yibing Wang, Hanghang Ma, Xiaoping Wu, Xiaqi Ma, Xiaoming Wei, Jianbin Jiao, Enhua Wu, and Jie Hu. Kangaroo: A powerful video-language model supporting long-context video input. arXiv preprint arXiv:2408.15542, 2024. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 488, + 295, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 488, + 295, + 532 + ], + "spans": [ + { + "bbox": [ + 62, + 488, + 295, + 532 + ], + "type": "text", + "content": "[57] Jiazhen Liu, Yuhan Fu, Ruobing Xie, Runquan Xie, Xingwu Sun, Fengzong Lian, Zhanhui Kang, and Xirong Li. PhD: A chatgpt-prompted visual hallucination evaluation dataset, 2025. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 533, + 295, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 533, + 295, + 588 + ], + "spans": [ + { + "bbox": [ + 62, + 533, + 295, + 588 + ], + "type": "text", + "content": "[58] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 590, + 295, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 590, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 62, + 590, + 295, + 622 + ], + "type": "text", + "content": "[59] Shi Liu, Kecheng Zheng, and Wei Chen. Paying more attention to image: A training-free method for alleviating hallucination in lvlms, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 624, + 295, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 624, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 62, + 624, + 295, + 666 + ], + "type": "text", + "content": "[60] Alexander Nikitin, Jannik Kossen, Yarin Gal, and Pekka Marttinen. Kernel language entropy: Fine-grained uncertainty quantification for llms from semantic similarities. arXiv, 2024. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 669, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 669, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 62, + 669, + 295, + 689 + ], + "type": "text", + "content": "[61] OpenAI. GPT-4. https://openai.com/gpt-4, 2023.2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 62, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 691, + 295, + 713 + ], + "type": "text", + "content": "[62] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, et al." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 321, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 338, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 338, + 73, + 553, + 106 + ], + "type": "text", + "content": "Training language models to follow instructions with human feedback. In Advances in Neural Information Processing Systems, pages 27730-27744, 2022. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 107, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 107, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 321, + 107, + 553, + 162 + ], + "type": "text", + "content": "[63] Amandalynne Paullada, Inioluwa Deborah Raji, Emily M. Bender, Emily Denton, and Alex Hanna. Data and its (dis)contents: A survey of dataset development and use in machine learning research. *Patterns*, 2(11):100336, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 163, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 163, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 321, + 163, + 553, + 196 + ], + "type": "text", + "content": "[64] Phuc Phan, Hieu Tran, and Long Phan. Distillation contrastive decoding: Improving llms reasoning with contrastive decoding and distillation, 2024. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 197, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 197, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 321, + 197, + 553, + 229 + ], + "type": "text", + "content": "[65] Venkit P.N., Gautam S., Panchanadikar R., Huang T.H., and Wilson S. Nationality bias in text generation. arXiv preprint arXiv:2302.02463, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 231, + 553, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 231, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 321, + 231, + 553, + 275 + ], + "type": "text", + "content": "[66] Xiaoye Qu, Qiyuan Chen, Wei Wei, Jishuo Sun, and Jianfeng Dong. Alleviating hallucination in large vision-language models with active retrieval augmentation, 2024. http://arxiv.org/abs/2408.00555.3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 276, + 553, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 276, + 553, + 320 + ], + "spans": [ + { + "bbox": [ + 321, + 276, + 553, + 320 + ], + "type": "text", + "content": "[67] Xiaoye Qu, Jiashuo Sun, Wei Wei, and Yu Cheng. Look, compare, decide: Alleviating hallucination in large vision-language models via multi-view multi-path reasoning, 2024. http://arxiv.org/abs/2408.17150.3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 322, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 322, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 321, + 322, + 553, + 387 + ], + "type": "text", + "content": "[68] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 321, + 388, + 555, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 388, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 321, + 388, + 555, + 453 + ], + "type": "text", + "content": "[69] Vikas Raunak, Arul Menezes, and Marcin Junczys-Dowmunt. The curious case of hallucinations in neural machine translation. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1172-1183, 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 321, + 455, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 455, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 321, + 455, + 553, + 487 + ], + "type": "text", + "content": "[70] Kim S., Cho B., Bae S., Ahn S., and Yun S.Y. Vacode: Visual augmented contrastive decoding. arXiv preprint arXiv:2408.05337, 2024. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 321, + 489, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 489, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 321, + 489, + 553, + 521 + ], + "type": "text", + "content": "[71] Lin S., Hilton J., and Evans O. Truthfulqa: Measuring how models mimic human falsehoods. arXiv preprint arXiv:2109.07958, 2021. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 321, + 523, + 553, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 523, + 553, + 556 + ], + "spans": [ + { + "bbox": [ + 321, + 523, + 553, + 556 + ], + "type": "text", + "content": "[72] Yin S., Fu C., Zhao S., Li K., Sun X., Xu T., and Chen E. A survey on multimodal large language models. arXiv preprint arXiv:2306.13549, 2023. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 321, + 557, + 553, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 557, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 321, + 557, + 553, + 622 + ], + "type": "text", + "content": "[73] Pranab Sahoo, Prakash Meharia, Akash Ghosh, Sriparna Saha, Vinija Jain, and Aman Chadha. A comprehensive survey of hallucination in large language, image, video and audio foundation models. Findings of the Association for Computational Linguistics: EMNLP 2024, pages 11709-11724, 2024. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 321, + 624, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 624, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 321, + 624, + 553, + 679 + ], + "type": "text", + "content": "[74] Dustin Schwenk, Apoorv Khandelwal, Christopher Clark, Kenneth Marino, and Roozbeh Mottaghi. A-okvqa: A benchmark for visual question answering using world knowledge. In European conference on computer vision, pages 146–162. Springer, 2022. 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 321, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 321, + 681, + 553, + 713 + ], + "type": "text", + "content": "[75] Lei Shen, Haolan Zhan, Xin Shen, Hongshen Chen, Xiaofang Zhao, and Xiaodan Zhu. Identifying untrustworthy samples: Data filtering for open-domain dialogues with" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4219" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 80, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 80, + 72, + 294, + 106 + ], + "type": "text", + "content": "bayesian optimization. In Proceedings of the 30th ACM International Conference on Information & Knowledge Management, pages 1598-1608, 2021. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 107, + 295, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 107, + 295, + 149 + ], + "spans": [ + { + "bbox": [ + 61, + 107, + 295, + 149 + ], + "type": "text", + "content": "[76] Weijia Shi, Xiaochuang Han, Mike Lewis, Yulia Tsvetkov, Luke Zettlemoyer, and Scott Wen tau Yih. Trusting your evidence: Hallucinate less with context-aware decoding, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 152, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 152, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 61, + 152, + 294, + 205 + ], + "type": "text", + "content": "[77] Weijia Shi, Sewon Min, Maria Lomeli, Chunting Zhou, Margaret Li, Rich James, Xi Victoria Lin, Noah A. Smith, Luke Zettlemoyer, Scott Yih, and Mike Lewis. In-context pretraining: Language modeling beyond document boundaries, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 207, + 294, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 207, + 294, + 250 + ], + "spans": [ + { + "bbox": [ + 62, + 207, + 294, + 250 + ], + "type": "text", + "content": "[78] GLM T., Zeng A., Xu B., Wang B., Zhang C., Yin D., Rojas D., Feng G., Zhao H., Lai H., and Yu H. Chatglm: A family of large language models from glm-130b to glm-4 all tools. arXiv preprint arXiv:2406.12793, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 251, + 295, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 251, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 62, + 251, + 295, + 612 + ], + "type": "text", + "content": "[79] Gemma Team, Thomas Mesnard, Cassidy Hardin, Robert Dadashi, Surya Bhupatiraju, Shreya Pathak, Laurent Sifre, Morgane Rivière, Mihir Sanjay Kale, Juliette Love, Pouya Tafti, Léonard Hussenot, Pier Giuseppe Sessa, Aakanksha Chowdhery, Adam Roberts, Aditya Barua, Alex Botev, Alex Castro-Ros, Ambrose Slone, Amélie Héliou, Andrea Tacchetti, Anna Bulanova, Antonia Paterson, Beth Tsai, Bobak Shahriari, Charline Le Lan, Christopher A. Choquette-Choo, Clément Crepy, Daniel Cer, Daphne Ippolito, David Reid, Elena Buchatskaya, Eric Ni, Eric Noland, Geng Yan, George Tucker, George-Christian Muraru, Grigory Rozhdestvenskiy, Henryk Michalewski, Ian Tenney, Ivan Grishchenko, Jacob Austin, James Keeling, Jane Labanowski, Jean-Baptiste Lespiau, Jeff Stanway, Jenny Brennan, Jeremy Chen, Johan Ferret, Justin Chiu, Justin Mao-Jones, Katherine Lee, Kathy Yu, Katie Millican, Lars Lowe Sjoesund, Lisa Lee, Lucas Dixon, Rachel Reid, Maciej Mikula, Mateo Wirth, Michael Sharman, Nikolai Chinaev, Nithum Thain, Olivier Bachem, Oscar Chang, Oscar Wahltinez, Paige Bailey, Paul Michel, Petko Yotov, Rahma Chaabouni, Ramona Comanescu, Reena Jana, Rohan Anil, Ross McIlroy, Ruibo Liu, Ryan Mullins, Samuel L Smith, Sebastian Borgeaud, Sertan Girgin, Sholto Douglas, Shree Pandya, Siamak Shakeri, Soham De, Ted Klimenko, Tom Hennigan, Vlad Feinberg, Wojciech Stokowiec, Yu hui Chen, Zafarali Ahmed, Zhitao Gong, Tris Warkentin, Ludovic Peran, Minh Giang, Clément Farabet, Oriol Vinyals, Jeff Dean, Koray Kavukcuoglu, Demis Hassabis, Zoubin Ghahramani, Douglas Eck, Joelle Barral, Fernando Pereira, Eli Collins, Armand Joulin, Noah Fiedel, Evan Senter, Alek Andreev, and Kathleen Kenealy. Gemma: Open models based on gemini research and technology, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 614, + 294, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 614, + 294, + 668 + ], + "spans": [ + { + "bbox": [ + 62, + 614, + 294, + 668 + ], + "type": "text", + "content": "[80] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 670, + 294, + 712 + ], + "type": "text", + "content": "[81] Elena Voita, David Talbot, Fedor Moiseev, Rico Sennrich, and Ivan Titov. Analyzing multi-head self-attention: Specialized heads do the heavy lifting, the rest can be pruned. In Proceedings of the 57th Annual Meeting of the Asso" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 339, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 339, + 73, + 553, + 94 + ], + "type": "text", + "content": "citation for Computational Linguistics, pages 5797-5808, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 321, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 321, + 96, + 553, + 140 + ], + "type": "text", + "content": "[82] Zhong W., Feng X., Zhao L., Li Q., Huang L., Gu Y., Ma W., Xu Y., and Qin B. Investigating and mitigating the multimodal hallucination snowballing in large vision-language models. arXiv preprint arXiv:2407.00569, 2024. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 320, + 141, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 141, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 320, + 141, + 553, + 206 + ], + "type": "text", + "content": "[83] Guankun Wang, Long Bai, Wan Jun Nah, Jie Wang, Zhaoxi Zhang, Zhen Chen, Jinlin Wu, Mobarakol Islam, Hongbin Liu, and Hongliang Ren. Surgical-lvIm: Learning to adapt large vision-language model for grounded visual question answering in robotic surgery. arXiv preprint arXiv:2405.10948, 2024. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 320, + 208, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 208, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 320, + 208, + 553, + 251 + ], + "type": "text", + "content": "[84] P Wang, S Bai, S Tan, S Wang, Z Fan, J Bai, K Chen, X Liu, J Wang, W Ge, and Y Fan. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 321, + 254, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 254, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 321, + 254, + 553, + 285 + ], + "type": "text", + "content": "[85] Xintong Wang, Jingheng Pan, Liang Ding, and Chris Biemann. Mitigating hallucinations in large vision-language models with instruction contrastive decoding, 2024. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 321, + 287, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 287, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 321, + 287, + 553, + 330 + ], + "type": "text", + "content": "[86] Sangmin Woo, Jaehyuk Jang, Donguk Kim, Yubin Choi, and Changick Kim. Ritual: Random image transformations as a universal anti-hallucination lever in lvlms, 2024. http://arxiv.org/abs/2405.17821.3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 321, + 332, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 332, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 321, + 332, + 553, + 374 + ], + "type": "text", + "content": "[87] Sangmin Woo, Donguk Kim, Jaehyuk Jang, Yubin Choi, and Changick Kim. Don't miss the forest for the trees: Attentional vision calibration for large vision language models, 2024. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 321, + 377, + 553, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 377, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 321, + 377, + 553, + 420 + ], + "type": "text", + "content": "[88] Junfei Wu, Qiang Liu, Ding Wang, Jinghao Zhang, Shu Wu, Liang Wang, and Tieniu Tan. Logical closed loop: Uncovering object hallucinations in large vision-language models, 2024. http://arxiv.org/abs/2402.11622.3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 422, + 553, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 422, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 321, + 422, + 553, + 475 + ], + "type": "text", + "content": "[89] Kai Wu, Boyuan Jiang, Zhengkai Jiang, Qingdong He, Donghao Luo, Shengzhi Wang, Qingwen Liu, and Chengjie Wang. Noiseboost: Alleviating hallucination with noise perturbation for multimodal large language models, 2024. http://arxiv.org/abs/2405.20081.3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 478, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 478, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 321, + 478, + 553, + 510 + ], + "type": "text", + "content": "[90] Xin Xiao, Bohong Wu, Jiacong Wang, Chunyuan Li, Xun Zhou, and Haoyuan Guo. Seeing the image: Prioritizing visual correlation by contrastive alignment, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 512, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 512, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 321, + 512, + 553, + 555 + ], + "type": "text", + "content": "[91] Xiaohan Xu, Chongyang Tao, Tao Shen, Can Xu, Hongbo Xu, Guodong Long, and Jian guang Lou. Re-reading improves reasoning in large language models, 2024. http://arxiv.org/abs/2309.06275.3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 557, + 553, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 557, + 553, + 609 + ], + "spans": [ + { + "bbox": [ + 321, + 557, + 553, + 609 + ], + "type": "text", + "content": "[92] L Xue, M Shu, A Awadalla, J Wang, A Yan, S Purushwalkam, H Zhou, V Prabhu, Y Dai, M S Ryoo, and S Kendre. xgen-mm (blip-3): A family of open large multimodal models. arXiv preprint arXiv:2408.08872, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 613, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 613, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 321, + 613, + 553, + 644 + ], + "type": "text", + "content": "[93] Onoe Y., Zhang M.J., Choi E., and Durrett G. Entity cloze by date: What lms know about unseen entities. arXiv preprint arXiv:2205.02832, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 321, + 647, + 553, + 689 + ], + "type": "text", + "content": "[94] Park Y., Lee D., Choe J., and Chang B. Convis: Contrastive decoding with hallucination visualization for mitigating hallucinations in multimodal large language models. arXiv preprint arXiv:2408.13906, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 321, + 692, + 553, + 713 + ], + "type": "text", + "content": "[95] Zhenjie Yang, Xiaosong Jia, Hongyang Li, and Junchi Yan. Llm4drive: A survey of large language models for au" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4220" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 81, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 81, + 73, + 294, + 95 + ], + "type": "text", + "content": "tonomous driving. In NeurIPS 2024 Workshop on OpenWorld Agents, 2023. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 62, + 96, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 96, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 62, + 96, + 295, + 140 + ], + "type": "text", + "content": "[96] Shukang Yin, Chaoyou Fu, Sirui Zhao, Tong Xu, Hao Wang, Dianbo Sui, Yunhang Shen, Ke Li, Xing Sun, and Enhong Chen. Woodpecker: Hallucination correction for multimodal large language models. arXiv preprint, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 141, + 295, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 141, + 295, + 206 + ], + "spans": [ + { + "bbox": [ + 61, + 141, + 295, + 206 + ], + "type": "text", + "content": "[97] Qifan Yu, Juncheng Li, Longhui Wei, Liang Pang, Wentao Ye, Bosheng Qin, Siliang Tang, Qi Tian, and Yueting Zhuang. Hallucidctor: Mitigating hallucinatory toxicity in visual instruction data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12944–12953, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 209, + 295, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 209, + 295, + 274 + ], + "spans": [ + { + "bbox": [ + 62, + 209, + 295, + 274 + ], + "type": "text", + "content": "[98] Tianyu Yu, Yuan Yao, Haoye Zhang, Taiwen He, Yifeng Han, Ganqu Cui, Jinyi Hu, et al. Rlhf-v: Towards trustworthy mllms via behavior alignment from fine-grained correctional human feedback. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13807-13816, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 276, + 295, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 276, + 295, + 362 + ], + "spans": [ + { + "bbox": [ + 62, + 276, + 295, + 362 + ], + "type": "text", + "content": "[99] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, Cong Wei, Botao Yu, Ruibin Yuan, Renliang Sun, Ming Yin, Boyuan Zheng, Zhenzhu Yang, Yibo Liu, Wenhao Huang, Huan Sun, Yu Su, and Wenhu Chen. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of CVPR, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 365, + 294, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 365, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 57, + 365, + 294, + 407 + ], + "type": "text", + "content": "[100] Zihao Yue, Liang Zhang, and Qin Jin. Less is more: Mitigating multimodal hallucination from an eos decision perspective, 2024. http://arxiv.org/abs/2402.14545.3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 409, + 294, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 409, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 57, + 409, + 294, + 441 + ], + "type": "text", + "content": "[101] Bai Z., Wang P., Xiao T., He T., Han Z., Zhang Z., and Shou MZ. Hallucination of multimodal large language models: A survey. arXiv preprint arXiv:2404.18930, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 443, + 294, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 443, + 294, + 485 + ], + "spans": [ + { + "bbox": [ + 57, + 443, + 294, + 485 + ], + "type": "text", + "content": "[102] Han Z., Bai Z., Mei H., Xu Q., Zhang C., and Shou M.Z. Skip: A simple method to reduce hallucination in large vision-language models. arXiv preprint arXiv:2402.01345, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 488, + 294, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 488, + 294, + 520 + ], + "spans": [ + { + "bbox": [ + 57, + 488, + 294, + 520 + ], + "type": "text", + "content": "[103] Shaolei Zhang, Tian Yu, and Yang Feng. Truthx: Alleviating hallucinations by editing large language models in truthful space, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 522, + 294, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 522, + 294, + 564 + ], + "spans": [ + { + "bbox": [ + 57, + 522, + 294, + 564 + ], + "type": "text", + "content": "[104] Yi-Fan Zhang, Weichen Yu, Qingsong Wen, Xue Wang, Zhang Zhang, Liang Wang, Rong Jin, and Tieniu Tan. Debiasing multimodal large language models. arXiv preprint, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 567, + 294, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 567, + 294, + 600 + ], + "spans": [ + { + "bbox": [ + 57, + 567, + 294, + 600 + ], + "type": "text", + "content": "[105] L Zhao, Y Deng, W Zhang, and Q Gu. Mitigating object hallucination in large vision-language models via classifier-free guidance. 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 601, + 294, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 601, + 294, + 654 + ], + "spans": [ + { + "bbox": [ + 57, + 601, + 294, + 654 + ], + "type": "text", + "content": "[106] Zhiyuan Zhao, Bin Wang, Linke Ouyang, Xiaoyi Dong, Jiaqi Wang, and Conghui He. Beyond hallucinations: Enhancing lvlms through hallucination-aware direct preference optimization. arXiv preprint arXiv:2311.16839, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 658, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 294, + 689 + ], + "type": "text", + "content": "[107] Zheng Zhao, Emilio Monti, Jens Lehmann, and Haytham Assem. Enhancing contextual understanding in large language models through contrastive decoding, 2024. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "type": "text", + "content": "[108] Lanyun Zhu, Deyi Ji, Tianrun Chen, Peng Xu, Jieping Ye, and Jun Liu. Ibid: Alleviating hallucinations in large" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 73, + 553, + 140 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "type": "text", + "content": "vision-language models via image-biased decoding. arXiv preprint, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 553, + 140 + ], + "type": "text", + "content": "[109] Wenhao Zhu, Sizhe Liu, Shujian Huang, Shuaijie She, Chris Wendler, and Jiajun Chen. Multilingual contrastive decoding via language-agnostic layers skipping, 2024. http://arxiv.org/abs/2407.10795.2,3" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "4221" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/ID-Patch_ Robust ID Association for Group Photo Personalization/a37d3ef3-36c1-46f2-886c-58bf6589bcc4_content_list.json b/2025/ID-Patch_ Robust ID Association for Group Photo Personalization/a37d3ef3-36c1-46f2-886c-58bf6589bcc4_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..cdb25c2263966f0dbcf49280b61403c4b8aac572 --- /dev/null +++ b/2025/ID-Patch_ Robust ID Association for Group Photo Personalization/a37d3ef3-36c1-46f2-886c-58bf6589bcc4_content_list.json @@ -0,0 +1,1703 @@ +[ + { + "type": "text", + "text": "ID-Patch: Robust ID Association for Group Photo Personalization", + "text_level": 1, + "bbox": [ + 165, + 130, + 833, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yimeng Zhang $^{1,2,*}$ , Tiancheng Zhi $^{1}$ , Jing Liu $^{1}$ , Shen Sang $^{1}$ , Liming Jiang $^{1}$ , Qing Yan $^{1}$ , Sijia Liu $^{2}$ , Linjie Luo $^{1}$", + "bbox": [ + 261, + 179, + 736, + 218 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 ByteDance Inc., 2 Michigan State University", + "bbox": [ + 310, + 232, + 679, + 252 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d64583b412951698b75d34769dd212fc91c0de9fd0546b8d0ec9f60eba4a078b.jpg", + "image_caption": [ + "Pose Condition & ID Location", + "\"An indoor celebration scene featuring three women...\"", + "Figure 1. Comparison with state-of-the-art multi-identity generation methods. From left to right: the condition inputs followed by results generated using OMG [19] (with InstantID [43]), InstantFamily [17], and our proposed ID-Patch approach. Red dashed boxes highlight failures. OMG fails to preserve the hairstyles of the middle person, and creates artifacts for the right woman's hand, possibly because of the inconsistency between its generation stages (explained in Fig. 2). InstantFamily suffers from ID leakage, resulting the incorrect ID of the middle person. Our approach preserves the detailed identity of each person. In addition, our approach is 7 times faster than OMG and has less computational overhead than InstantFamily." + ], + "image_footnote": [], + "bbox": [ + 98, + 321, + 292, + 470 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4debe4e158bfe18c9c369aab5f88a45b9bba5969529a1cdcd99797b26ad68166.jpg", + "image_caption": [ + "OMG + InstantID", + "Generation Time: 71.1s" + ], + "image_footnote": [], + "bbox": [ + 300, + 321, + 495, + 470 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e7448aa925b865f1f7ed44fbfdf753faee8352c4cb3cf97b50cd71f20f655d06.jpg", + "image_caption": [ + "InstantFamily", + "Generation Time: 9.6s" + ], + "image_footnote": [], + "bbox": [ + 503, + 321, + 697, + 470 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5d9a202bf8b9bc0524d4461d6ae840f929a6599916cadd4a0adb0673632e912b.jpg", + "image_caption": [ + "ID-Patch (Ours)", + "Generation Time: 9.4s" + ], + "image_footnote": [], + "bbox": [ + 705, + 321, + 900, + 470 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 579, + 326, + 595 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The ability to synthesize personalized group photos and specify the positions of each identity offers immense creative potential. While such imagery can be visually appealing, it presents significant challenges for existing technologies. A persistent issue is identity (ID) leakage, where injected facial features interfere with one another, resulting in low face resemblance, incorrect positioning, and visual artifacts. Existing methods suffer from limitations such as the reliance on segmentation models, increased runtime, or a high probability of ID leakage. To address these challenges, we propose ID-Patch, a novel method that provides robust association between identities and 2D positions. Our approach generates an ID patch and ID embeddings from the same facial features: the ID patch is positioned on the conditional image for precise spatial control, while the ID embeddings integrate with text embeddings to ensure high resemblance. Experimental results demon", + "bbox": [ + 88, + 614, + 485, + 872 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "strate that ID-Patch surpasses baseline methods across metrics, such as face ID resemblance, ID-position association accuracy, and generation efficiency. Project Page is: https://byteaigc.github.io/ID-Patch/", + "bbox": [ + 511, + 580, + 906, + 642 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 676, + 645, + 691 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Personalized group photo generation has wide-ranging applications in entertainment, advertising, social media, and virtual reality. For example, users might want to create a group photo of friends at a virtual party, even if some could not attend in person. This technology enables the creation of engaging and memorable content that reflects real-life social interactions and relationships. However, this task poses significant challenges, particularly the risk of identity (ID) leakage, where the identity information of one individual inadvertently affects the representation of another at a different position. Such blending compromises the distinctiveness and accuracy of each person's features. Besides, being able to specify the position of each identity in the image is", + "bbox": [ + 509, + 704, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work done during internship at ByteDance.", + "bbox": [ + 114, + 886, + 354, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2986", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "an important function when users want to create a specific layout of people. Incorrect positioning of identities may lead to undesired results.", + "bbox": [ + 89, + 90, + 482, + 135 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "State-of-the-art group photo personalization methods largely rely on diffusion models [3, 6, 16, 26, 37-39, 41, 48, 49, 54, 56], which deliver superior visual quality when paired with effective prompt engineering [4, 8, 25, 27, 34, 55]. To address the issue of ID leakage, OMG [19] first generates an image without ID injection. Then, it estimates segmentation masks for the individuals in the image and subsequently injects the IDs into each segmented region. However, it depends on a segmentation model, which might fail in challenging cases. Besides, the content generated in the two stages may conflict with each other, leading to reduced visual fidelity. See Fig.1 and Fig.2 for example. Moreover, OMG requires running the denoising process separately for each identity, resulting in a runtime that increases linearly with the number of people. Alternatively, InstantFamily [17] employs a single-pass generation strategy by manipulating cross-attention masks. This approach uses approximate head region masks for each ID to direct pixel attention exclusively to specific ID embeddings derived from facial features. However, it struggles to prevent ID leakage due to two key issues: (1) imprecise masks or close proximity of faces, which can cause overlaps, and (2) unintended information propagation through self-attention and convolutional layers. Please see Fig.1 and Fig.3 for more details.", + "bbox": [ + 91, + 138, + 483, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In response to these limitations, we propose a novel method named ID-Patch, designed for efficient and robust ID-position association in group photo generation. Our idea is to let the model learn to associate input IDs with their spatially designated locations. Specifically, we project facial features onto a small RGB image patch called ID patch and token embeddings called ID embeddings. The ID patches are placed onto the conditioning image of a ControlNet [53] according to face locations to ensure precise position control of IDs within the generated images. ID embeddings are integrated with text embeddings to enhance facial details. Our ID-Patch method seamlessly integrates with various types of spatial conditions, such as poses, canny edges or depth map, enhancing the robustness and flexibility of our method.", + "bbox": [ + 89, + 518, + 483, + 743 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We summarize our key contributions as follows:", + "bbox": [ + 89, + 747, + 416, + 762 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- ID-Patch Method: We introduce a novel approach for multi-ID image generation that directly links face ID features with their locations using visual patches, ensuring accurate resemblance and position control.", + "- Efficiency Improvement: Our method enhances the efficiency of multi-ID image generation by simplifying the process to adding ID patches and concatenating ID embeddings, reducing computational overhead.", + "- Simplified Control: We eliminate the need for auxiliary" + ], + "bbox": [ + 89, + 763, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9a9242358b947183b0116abda6bcf658424be76bc062b0f0fb1fa029e1eb432d.jpg", + "image_caption": [ + "(a) OMG first stage" + ], + "image_footnote": [], + "bbox": [ + 517, + 89, + 642, + 185 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0b75236d224893c93a9df368cdade44119f60df4de1c4fee8794b9f6b5564b24.jpg", + "image_caption": [ + "(b) Human mask" + ], + "image_footnote": [], + "bbox": [ + 643, + 89, + 769, + 186 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/be8c39d8f2fd7623878a30f02d4c55d21f4f345d4886d704ecd42672f0b72194.jpg", + "image_caption": [ + "(c) OMG final result" + ], + "image_footnote": [], + "bbox": [ + 772, + 89, + 898, + 186 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f426f777d41f1f522a8d88580de0c331e5c28eeac4cfb4d0503105746da9ae63.jpg", + "image_caption": [ + "Figure 2. Why does OMG fail in Fig. 1? OMG relies on a first stage result (no ID awareness) and its human segmentation to control layout. The second stage paints injected IDs within corresponding masks. However, the second pass may conflict with the first stage semantics, leading to deteriorated ID resemblance (long hair in the middle of image (c)) or artifacts (hand in the middle).", + "(a) ID", + "Figure 3. ID leakage of InstantFamily [17]. The text prompt describes a man and a woman standing together, and ID (a) is only injected into the left man." + ], + "image_footnote": [], + "bbox": [ + 517, + 292, + 640, + 419 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6cf9aead00171fe6526fd123fe4ae035538091c65b13c96ce1c9a3b02e21265c.jpg", + "image_caption": [ + "(b) InstantFamily result" + ], + "image_footnote": [], + "bbox": [ + 643, + 292, + 771, + 419 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/84e3aa550abef8c2ee884f0ddd08a91d769d292c9790914f313689485c913f51.jpg", + "image_caption": [ + "(c) ID-Patch result" + ], + "image_footnote": [], + "bbox": [ + 772, + 292, + 898, + 419 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "segmentation models, requiring only a single point for ID position control, thereby increasing robustness and reducing complexity.", + "bbox": [ + 526, + 500, + 903, + 544 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Superior Performance: Our approach achieves higher fidelity in ID resemblance and better position control, especially in complex images with multiple identities, as demonstrated by our experimental results.", + "bbox": [ + 513, + 544, + 903, + 604 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 618, + 653, + 633 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Single-ID Personalization. The personalization of diffusion models has emerged as a popular research topic. Finetuning or optimization based methods [9, 10, 32, 33, 46] have the potential to reach high quality results in terms of ID resemblance, but are time-consuming. Inference-only methods [15, 23, 28, 43, 45, 50, 51, 57] eliminate the finetuning process usually by encoding the concept into token embeddings and injecting them into the diffusion models via cross-attention. However, most of these works focus on single concept injection rather than group photo generation. Multi-ID Personalization. The main challenge of personalized group photo generation is ID blending, i.e., the identity information of person A may affect the generation of person B at another position. To resolve this issue, OMG [19] performs separate denoising processes. First, a noncustomized image is generated and masks of each person are extracted. Second, masked regions are painted using", + "bbox": [ + 511, + 643, + 906, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2987", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3424c2213de91929ba6e557856107180c698ce40c1fb3ff61960ff2edb0bddc0.jpg", + "image_caption": [ + "Figure 4. Our pipeline. Given a text prompt (e.g., two people shaking hands), $N$ face images and locations, we generate an image with $N$ IDs. We extract face features for each ID and project them into ID patches and ID embeddings. ID patches are rendered on a black canvas (or added on top of a pose image) according to face locations and sent into a ControlNet to control the positions of generated faces. ID embeddings are appended to text embeddings to provide detailed face information to the diffusion model and ControlNet via cross-attention." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 903, + 268 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "separate customized models and stitched together. This approach is slow and conflicts between stages may cause identity loss. Additionally, it relies on a segmentation module, which may fail in challenging cases. To avoid running separate customized models, recent works [14, 42, 47] let the network to learn to attend to specific concept tokens by supervising the attention mask using segmentation or ID routing. However, these methods do not allow the specification of the accurate locations of each ID. Although they can be combined with pose conditions, describing the assignment of IDs to positions in text may be hard. InstantFamily [17] provides an opportunity to specify the approximate location of each ID by manipulating the attention masks during both training and inference. However, no constraints have been imposed to avoid the ID leakage through self-attention and convolutional layers, leading to ID blending, omitting or swapping.", + "bbox": [ + 88, + 357, + 485, + 613 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multiple General Concepts. Besides human generation, there are series of work focusing on the customization of multiple general concepts [1, 2, 7, 12, 13, 20, 24, 40, 47] and grounded image generation [2, 22, 44] associating objects with positions. GLIGEN [22] integrates grounding tokens, which align textual prompts with specific image regions. SpaText [2] and InstanceDiffusion [44] allows users to define the placement and attributes of objects. However, these concepts are usually semantically distinct while it is hard to describe the human identities in text and subtle details matter.", + "bbox": [ + 89, + 614, + 483, + 781 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 797, + 181, + 814 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first outline some preliminaries about text-to-image diffusion models and ControlNets. Subsequently, we detail the problem formulation for multi-ID image generation and the design of our modules, including the ID patch and ID embeddings, as well as the training and", + "bbox": [ + 89, + 825, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "inference schemes.", + "bbox": [ + 513, + 357, + 643, + 371 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 513, + 380, + 653, + 396 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Latent diffusion models (LDMs). LDMs [31] are image generation models that can incorporate conditioning signals into the generation process. Specifically, during training, an encoder $\\mathcal{E}$ of variational autoencoder (VAE) [18] encodes an image $x$ into a latent space, and a neural network $\\epsilon_{\\theta}(z_t,t,c)$ performs the denoising process within this latent space. It attempts to predict the original Gaussian noise $\\epsilon$ from a noisy latent variable $z_{t}$ , where $t$ represents the diffusion timestep, and $c$ is the conditioning signal, injected via cross-attention. The loss function for an LDM is defined as:", + "bbox": [ + 511, + 404, + 906, + 555 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {L D M}} = \\mathbb {E} _ {\\mathcal {E} (x), \\epsilon \\sim \\mathcal {N} (0, 1), t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (z _ {t}, t, c) \\| _ {2} ^ {2} \\right] \\quad (1)\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 566, + 906, + 587 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ControlNet. The incorporation of ControlNet [53] into diffusion models marks a significant advancement in controlled image synthesis. ControlNet is designed to steer the image generation process by applying spatial conditioning signals such as rasterized pose images, depth maps, or canny edges. It introduces a control network that processes these signals and adds their features to the base model's intermediate results. The control network is a trainable copy of the base model's encoder layers and is initialized with zero convolution. This effective integration achieves high quality controlled generation.", + "bbox": [ + 511, + 595, + 908, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Multi-ID Image Generation", + "text_level": 1, + "bbox": [ + 511, + 771, + 764, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This paper focuses on efficiently and effectively addressing the problem of multi-ID image generation utilizing conditioning signals that include text prompts, identity features, and face locations. The inputs to our model comprise text prompt embeddings $c_{t}$ , facial features $c_{f} = (f_{0}, f_{1}, \\ldots, f_{N-1})$ for $N$ distinct identities, and their corresponding facial locations $c_{l} = (l_{0}, l_{1}, \\ldots, l_{N-1})$ . The", + "bbox": [ + 511, + 794, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2988", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/75b1ba4caaae833986a82dbf80ce97f0c559b36ea6c6fba660732df60e0771ea.jpg", + "image_caption": [ + "(a) ID" + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 187, + 228 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/284eb7c683953fa9f26211122794f9c8e5276c1c5ebd9bf835d58f377fa2861e.jpg", + "image_caption": [ + "(b) w/o ID embeddings" + ], + "image_footnote": [], + "bbox": [ + 189, + 88, + 333, + 229 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fa0e881be04ff7854c1192da7234ec1832dc71e9eb4c267f6238466226442436.jpg", + "image_caption": [ + "(c) with ID embeddings", + "Figure 5. Effectiveness of ID embedding. Without ID embeddings, one can distinguish between the two people but the resemblance is low. Incorporating ID embeddings significantly improves face resemblance." + ], + "image_footnote": [], + "bbox": [ + 333, + 88, + 478, + 229 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "facial location $l_{i} = (x_{i},y_{i})$ identifies the xy-coordinates of the nose tip for each identity within the intended image. The goal is to synthesize an image following the text description with each given face located in the designated place.", + "bbox": [ + 89, + 335, + 483, + 397 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1 ID Patches and ID Embeddings", + "text_level": 1, + "bbox": [ + 89, + 419, + 362, + 434 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As discussed in Sec. 3.1, ControlNet effectively integrates spatial conditional signals into diffusion models. Building on this capability, we introduce a method to encode ID location information using condition images, subsequently processed by ControlNet to guide ID localization within the generated image. Our approach features ID conditions manifesting as both patches and embeddings. Specifically, ID patches, precisely aligned with facial locations, are integrated via ControlNet to ensure accurate identity localization, while ID embeddings are utilized via cross-attention mechanisms in the LDM to enhance the detailed representation of facial features, as illustrated in Fig. 4.", + "bbox": [ + 88, + 445, + 482, + 626 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Both ID patches and ID embeddings are derived from facial features, enabling a synergistic association between them, with both projection processes facilitated by the Perceiver Resampler [50]. For ID patches, we convert the facial feature $f_{i}$ into an identity-specific patch $p_i \\in \\mathbb{R}^{P \\times P \\times 3}$ , functioning similarly to a QR code, uniquely identifying each person and easily distinguishable from others. These ID patches are placed on a black canvas at specified facial locations to create an image $I$ , which then serves as the conditioning input for ControlNet, thereby providing precise identity localization information.", + "bbox": [ + 88, + 627, + 482, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As depicted in Fig. 5, while ID patches alone are generally sufficient for distinguishing between identities, they may not always achieve optimal results in generating images with high ID resemblance. To address this limitation, and inspired by InstantFamily [17], we also incorporate ID embeddings. These embeddings are designed to inject detailed facial appearance information into the dif", + "bbox": [ + 89, + 795, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/27e1ad580f5f8ddd1fa4a85fe3db5e48cf59290f0fe1c8cb12dcfd11fed7c274.jpg", + "image_caption": [ + "(a) ID (b) Result w/o pose", + "Figure 6. ID-Patch combined with pose conditions. Provided with user ID images in (a), our method can generate results with only the specifications of nose tip locations as shown in (b). Merging ID patches with pose images (c) enhances spatial control over the generated results as seen in (d), without incurring any computational overhead." + ], + "image_footnote": [], + "bbox": [ + 514, + 88, + 671, + 218 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/14e7f4b15a207bd8700592e1a9dcfd309dda13dbd6b064d9b06571c572985fcc.jpg", + "image_caption": [ + "(c) ID patch on pose" + ], + "image_footnote": [], + "bbox": [ + 674, + 88, + 787, + 218 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1103147907d956261f40fa38d51ab9e346cabc62a20c10fb31bd794471f6772b.jpg", + "image_caption": [ + "(d) Result with pose" + ], + "image_footnote": [], + "bbox": [ + 790, + 88, + 903, + 218 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "fusion model, thus enhancing the resemblance of the generated identities to their real-life counterparts. Specifically, ID embeddings $w_{i} \\in \\mathbb{R}^{2048 \\times M}$ , share the same token dimension 2048 as the SDXL text embeddings. As shown in Fig.4, these embeddings are appended to text embeddings $c_{t}$ to form an extended text embedding $c_{t}' = [c_{t}, w_{0}, w_{1}, \\ldots, w_{N-1}]$ , which is then fed into both the UNet and ControlNet through cross-attention mechanisms.", + "bbox": [ + 511, + 345, + 905, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since human pose conditions are often utilized to enhance group photo generation, we can combine our ID-Patch approach with the pose conditions for multi-id image generation. As illustrated in Fig. 6, the ID patches can be overlaid onto a pose-conditioned image, creating a composite condition image that encodes both identity positioning and pose information. It is important to mention that while our method can incorporate pose information, it does not inherently require it; pose is just an optional enhancement, one can also choose other conditions like canny edge, depth map for specific applications. The adaptability of our ID-Patch approach presents enhanced control over generated results yet without incurring the extra overhead of an additional ControlNet branch.", + "bbox": [ + 511, + 467, + 905, + 676 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2 Two-Stage Training", + "text_level": 1, + "bbox": [ + 511, + 698, + 705, + 714 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The training of our model employs a conventional diffusion loss function, defined as follows:", + "bbox": [ + 511, + 723, + 903, + 752 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL = \\mathbb {E} _ {\\mathcal {E} (x), \\epsilon \\sim \\mathcal {N} (0, 1), t} [ \\| \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}, t, I, c _ {t} ^ {\\prime}\\right) \\| _ {2} ^ {2} ] \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 763, + 903, + 782 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Initially, one might consider training the models end-to-end from scratch. However, our experiments reveal that such a method does not consistently ensure accurate ID localization, as depicted in Fig.7. This shortcoming likely stems from the network's tendency to preferentially inject ID information into ID embeddings, rather than ID patches, resulting in non-distinctive ID patches that could potentially", + "bbox": [ + 511, + 794, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2989", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0e013eb6c9e2891b58771e3ad72e4f8114c525c700a6f443fb02a806d6509c4e.jpg", + "image_caption": [ + "(a) Pose + ID" + ], + "image_footnote": [], + "bbox": [ + 96, + 112, + 222, + 253 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/70787222617e083dc453dc28bdd54cbe210743c2d5ee1e84e96bd1db19bb09ba.jpg", + "image_caption": [ + "(b) Single-stage train" + ], + "image_footnote": [], + "bbox": [ + 222, + 88, + 349, + 253 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/88bc0f3cafe9c6baa7b0459f2bcee5bb4a63c1b874aa48f6c8c54baecc10e238.jpg", + "image_caption": [ + "(c) Two-stage train", + "Figure 7. Two-stage training to improve positioning robustness. Given pose and ID conditions in (a), single-stage training cannot fully prevent the incorrect face positioning issue. For example, in (b), the man is mistakenly placed at the central bottom position, generating an inharmonious result. (c) Two-stage training is introduced to solve this problem. As can be seen on the top row of this figure, the ID patches from our two-stage training approach are visually more distinguishable from each other, compared to the ones from single-stage training. Our experimental results prove that this is critical in solving the ID leak issues." + ], + "image_footnote": [], + "bbox": [ + 351, + 88, + 477, + 253 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "confuse ControlNet. In the single-stage model, as shown in Fig. 7, the ID patches exhibit only subtle differences between different individuals. To resolve this, we introduce a two-stage training approach. In the initial stage, ID embeddings are omitted, compelling the network to prioritize the incorporation of ID information into ID patches, thereby enhancing their distinctiveness. In the subsequent stage, ID embeddings are introduced and the model is fine-tuned to further improve the resemblance of the generated images.", + "bbox": [ + 89, + 431, + 483, + 568 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 584, + 220, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Settings", + "text_level": 1, + "bbox": [ + 89, + 609, + 187, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details. Facial features $f_{i}$ we use are 512-dimensional vectors derived from ArcFace [5]. During training, we utilize the SDXL diffusion model [29]. During the inference phase, we employ the further finetuned SDXL, namely Juggernaut-X-v10*, as base model to enhance portrait quality. The inference process involves 50 timesteps. We set the ID token injection ratio to 0.8 to regulate the influence of identity features, where the initial $20\\%$ of timesteps operate without ID token embeddings to focus on the general style, followed by the remaining $80\\%$ timesteps which incorporate ID token embeddings to refine the identity-specific aspects of the images. The default size of ID-patch is $P \\times P = 64 \\times 64$ and the default length of ID embeddings is $M = 16$ . The network is trained using 8 NVIDIA H100 GPUs with batch size=4, learning rate=1e-05, weight decay = 0.01, using AdaFactor optimizer [36].", + "bbox": [ + 89, + 633, + 482, + 875 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/c6c1054bc4419bb2928700fd54b7c6042b25cd3f4e6e4ba71afbb50414829d90.jpg", + "table_caption": [ + "Table 1. Performance comparison across four dimensions: identity resemblance, ID-position association accuracy, text alignment, and generation time. For qualitative metrics to evaluate similarity or accuracy, higher is better. Our approach significantly outperforms baseline methods in terms of resemblance, association, and generation time, with similar text alignment scores." + ], + "table_footnote": [], + "table_body": "
MethodsID ↑Association ↑Text ↑Time (s) ↓
OMG + InstantID0.6440.9260.27184.83
InstantFamily0.6830.8620.27810.02
ID-Patch (Ours)0.7510.9580.2739.69
", + "bbox": [ + 526, + 174, + 888, + 241 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We train it for 720k iterations for stage 1, and stage 2 usually converges around 1 million iterations. We select the best model based on a 20 people validation set.", + "bbox": [ + 511, + 261, + 903, + 306 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset. Our training set consists of 17 million single-person and 1.95 million multi-person images. These images are obtained from purchased data and publicly available sources. For evaluation, we utilize two specialized datasets: 1) Style Dataset comprises 40 styles with pose and text prompt definitions, ranging from one to eight faces. 2) Identity Dataset includes 600 unique identities representing a variety of races and genders. Each style is paired with 20 randomly sampled unique identity combinations for quantitative evaluation.", + "bbox": [ + 511, + 308, + 905, + 458 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. Our analysis includes two state-of-the-art multi-ID generation methods as baselines: 1) OMG [19]: This method incorporates multiple identity features into separate segmented regions through latent stitching. It can be paired with a single ID injection method InstantID [43] for multi-ID generation. 2) InstantFamily [17]: This approach employs coarse attention masks within cross-attention layers, a mandatory process that ensures distinct identity features are injected into different image instances effectively. Because there is no official publicly available code for InstantFamily [17], for fairness, we implemented it based on the SDXL base model [29], and trained it using the same training data as our model.", + "bbox": [ + 511, + 460, + 905, + 655 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics. We evaluate the methods in a pose-conditioned setting, because pose is a requirement of InstantFamily, and without pose condition, OMG is not able to generate faces at specified locations for its first stage. Besides, pose condition makes the head size consistent, which is easier for resemblance evaluation and reduces the impact of head sizes on generation quality. For comprehensive evaluation, we consider four critical dimensions: 1) Identity resemblance: we measure the identity resemblance by calculating the cosine similarity (CosSim) between the facial features of faces in the generated images and the reference faces. For fairness, we use FaceNet [35], a face model different from the ArcFace model used for training. 2) Association accuracy: we assess the accuracy of identity-position association in generated images by ensuring that the generated faces are most similar to their assigned coun", + "bbox": [ + 511, + 657, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "\\*https://huggingface.co/RunDiffusion/Juggernaut-X-v10", + "bbox": [ + 109, + 886, + 410, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2990", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6ad51e41ac9fd5e146601be20ebe6ee7d216be296facaeac83003acf660a81dd.jpg", + "image_caption": [ + "(a) Identity Resemblance" + ], + "image_footnote": [], + "bbox": [ + 119, + 75, + 354, + 234 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/74ba8deec9c02f2b28235561317fc0094418a060acfd19f6c5fd07d65792ebcb.jpg", + "image_caption": [ + "(b) Association Accuracy" + ], + "image_footnote": [], + "bbox": [ + 383, + 75, + 619, + 234 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/18f05bfa9f534f0cd24dff0c791aa88a5fab82002624d88200566a7f57a1db94.jpg", + "image_caption": [ + "(c) Generation Time" + ], + "image_footnote": [], + "bbox": [ + 645, + 74, + 880, + 234 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f6b54bb931d500df58634fba0dee751e2cb819c2a805d611f8d98d7b7c95847c.jpg", + "image_caption": [ + "Figure 8. Performance evaluation of different model generations from three different aspects: identity resemblance, location accuracy, and generation time. As face number increases, the ID and location metric scores of OMG and Instantfamily drop much more significantly than our approach. Our approach also achieves near-constant generation time, in contrast to OMG's linearly increasing running time.", + "Figure 9. Comparison with baselines on pose-conditioned generation, where red dashed boxes highlight instances with low identity resemblance. In row 1, OMG fails to preserve the face shape of the third woman (from left to right) because its first stage result conflicts with this ID's face shape. InstantFamily incorrectly generates the second person because of the ID leakage from the third woman. In row 2, OMG does not generate correct hair styles and accurate facial features for the two people in the red boxes, while InstantFamily generates the wrong ID in the red box." + ], + "image_footnote": [], + "bbox": [ + 99, + 313, + 897, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "terparts, compared to all other input faces used in the generation process. In other words, we calculate the similarity between the $i$ -th generated face and the $j$ -th input face. Let $s(i) = \\arg \\max_{0 \\leq j \\leq N-1} \\mathrm{CosSim}(\\tilde{f}_i^{\\mathrm{gen}}, \\tilde{f}_j)$ be the face index among all input faces that is most similar to the $i$ -th generated face, where $\\tilde{f}_i^{\\mathrm{gen}}$ , $\\tilde{f}_j$ denote the FaceNet features of the $i$ -th generated face and the $j$ -th input face respec", + "bbox": [ + 89, + 784, + 485, + 891 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tively, and $N$ is the number of faces in this image. When $s(i) = i$ , the $i$ -th generated face has the correct association. The identity-position association accuracy is defined as:", + "bbox": [ + 511, + 784, + 906, + 830 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {1}{N} \\sum_ {i = 0} ^ {N - 1} \\mathbf {1} \\{i = s (i) \\} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 640, + 837, + 906, + 878 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This measurement assesses the correct placement of each", + "bbox": [ + 513, + 885, + 906, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2991", + "bbox": [ + 482, + 944, + 513, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5bbcc4fa4961d5db7cb5cb2e5f8972ddf2a453361cb6344b4ea8f4eac01c4ece.jpg", + "table_caption": [ + "Table 2. Ablation study results for our proposed ID-Patch variants, designed to assess the impact of different modules (patch projection and token projection) and training/inferencing schemes on the performance of multi-ID image generation. The full method achieves the best ID resemblance and ID-position association, indicating the necessity of each module." + ], + "table_footnote": [], + "table_body": "
VariantsID ↑Association ↑Text ↑
No patch projection0.3680.3450.271
No token projection0.5650.8950.284
Single stage training0.6600.7840.271
ID token injection ratio = 1.00.7500.9570.270
FULL0.7510.9580.273
", + "bbox": [ + 109, + 174, + 460, + 270 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "identity within the image. 3) Text alignment: this metric evaluates the alignment between the textual context of the image generation prompt and the visual output. Cosine similarity is used to compare textual features extracted by the CLIP text encoder with visual features extracted by the CLIP image encoder from the generated images [30]. 4) Generation time: we measure the efficiency by recording the time (in seconds) required to generate an image on a NVIDIA A100 GPU, excluding the time taken for model loading and image I/O.", + "bbox": [ + 89, + 295, + 483, + 446 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Comparison with Baselines", + "text_level": 1, + "bbox": [ + 89, + 472, + 334, + 489 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As listed in Tab. 1, ID-Patch excels in identity and association metrics with scores of 0.751 and 0.958, respectively, showcasing its robust capability to maintain identity consistency and achieve precise placement in generated images. There is no significant difference in text alignment scores between compared methods. Notably, ID-Patch achieves the fastest image generation time.", + "bbox": [ + 89, + 501, + 480, + 608 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Further details on the performance with varying face numbers are shown in Fig. 8. As we observed no deterioration in text alignment with an increase in the number of faces across all methods, we have relegated the corresponding figure to the appendix. As the number of faces increases, both OMG and InstantFamily experience significant performance drops due to ID leakage. Although ID-Patch also shows a decline, it is less severe, mainly because the increase in face numbers reduces the average face area, adversely affecting the quality of smaller faces produced by the SDXL models. Moreover, the presence of more faces slightly extends the image generation time due to the increased demand for facial feature extraction. InstantFamily is slightly slower than our approach due to the overhead of masked attention. For OMG, injecting identity features into designated segmented areas causes a substantial increase in generation time; it takes about 2 minutes to generate an image with eight people, whereas ID-Patch requires only about 10 seconds. See Fig. 9 for visual comparisons.", + "bbox": [ + 89, + 613, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 513, + 90, + 666, + 107 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct an ablation study on variants of the proposed method in Tab. 2. The performance notably degrades when patch projection is omitted, leading to poor location alignment as the algorithm attempts to infer ID locations based on indirect cues such as gender and pose. Similarly, omitting token projection reduces resemblance by eliminating essential facial details, quantitatively validates the finding of Fig. 5. Single-stage training results in indistinguishable ID patches (as shown in Fig. 7), which impairs effective positioning. The default setting with an ID token injection ratio of 0.8 improves text alignment compared to injecting the ID token at every timestep, while maintaining the same level of identity resemblance. This improvement may arise because ID embeddings have little effect on facial detail during the early denoising steps, but become more influential in the later stages. It is important to note that even without the inference trick (i.e., using the ID token injection ratio), our proposed method still outperforms all baseline approaches.", + "bbox": [ + 511, + 112, + 903, + 385 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Pose-Free Generation", + "text_level": 1, + "bbox": [ + 511, + 393, + 715, + 407 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our ID-Patch model can be operated effectively without the need for pose conditioning, a flexibility that permits its integration with other conditions for image generation. This capability enables the use of ID-Patch in versatile configurations, whether plug-and-play or co-trained, which we will explore in subsequent subsections.", + "bbox": [ + 511, + 416, + 903, + 506 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the pose-free mode, the ControlNet uses a conditioning image that solely features ID patches against a black background. The placement of each ID patch at the tip of the nose ensures accurate preservation of facial locations during image synthesis, while head sizes and poses are variably generated based on the initial noise. As illustrated in Fig. 10, this methodology allows our model to produce aesthetically pleasing group photographs of different number of individuals across a variety of settings, demonstrating the robustness and adaptability of our approach.", + "bbox": [ + 511, + 507, + 905, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The pose-free version of ID-Patch is designed to be compatible with other ControlNets (plug-and-play). As illustrated in Fig. 11, it is possible to combine our ID-Patch ControlNet the Canny edge ControlNet to achieve more precise control over the final image outcomes.", + "bbox": [ + 511, + 657, + 903, + 734 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Limitations", + "text_level": 1, + "bbox": [ + 511, + 741, + 638, + 756 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As illustrated in Fig. 12, our method encounters two primary limitations. First, our implementation is based on the SDXL model, which occasionally produces inaccuracies in body and finger anatomy. Transitioning to a more advanced base model could potentially rectify this issue. Second, facial features in the current model may incorporate lighting and expression details that are not fully separated from identity information, potentially compromising the generation quality. Enriching our training dataset with multiple images", + "bbox": [ + 511, + 763, + 905, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2992", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6b74927a748a75fdf6b6a21127cb53612200650f4d5b1d8368a9a6611aea5a70.jpg", + "image_caption": [ + "ID + ID Location" + ], + "image_footnote": [], + "bbox": [ + 150, + 85, + 284, + 218 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2feb6fae2b3fde4fb4d6e8a59c347a70428465b0c0481a2f81914a92975d2511.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 308, + 85, + 480, + 218 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1bea2d862a140506b34a33356db25ba2e001fa0030e325abcdd9ca538404a9b8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 85, + 696, + 218 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4c1b876c595d0ca45c26eeda1e652aff7f59c209fb9303ca88b051fcf3ac462e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 717, + 85, + 888, + 218 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c26e393fd68cc9ae27e522abbf4ca149cddc532b02d8e1e77f18e6f35243cc2b.jpg", + "image_caption": [ + "Generation Result" + ], + "image_footnote": [], + "bbox": [ + 151, + 227, + 282, + 359 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/091862fcd47e0f47c4dcb8fa0a785ed6d57b1120e3555066af192d3961fcab68.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 308, + 227, + 480, + 359 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6780fe24621dea2c6ee8c9c2a50f09f13b1dcfca30df6b009578bee328a10814.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 227, + 696, + 359 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4145dbd7bf16d7cf16d17a23a0866652151c813191bccb66851df10f9366ebb6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 715, + 227, + 888, + 359 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c2bf392d9d0cb766ad432ff907109b8c37d19eed8d35bf1aa6a2d8b5bde139e6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 93, + 491, + 171, + 551 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bd194ab43cba5bbb98d920cf533990037813ed35980ee480f6dc40d93d8e0379.jpg", + "image_caption": [ + "(a) ID" + ], + "image_footnote": [], + "bbox": [ + 93, + 551, + 169, + 613 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/67b880b750a22302c775ef7630ce3d6dc21b52647a5f9ee8fa8bd549b14c1acf.jpg", + "image_caption": [ + "Figure 10. ID-Patch pose-free generation. Here the condition image is only ID patches rendered onto a black canvas without any pose condition. The generation is conditioned on individual face locations while corresponding head sizes and body poses are inferred implicitly. Our method can accommodate a large number of people with diverse ethnic backgrounds, yet generating visually appealing group photo results. See \"Evaluation Metrics\" for reasons why we compare with baseline approaches only in pose-conditioned settings.", + "(b) Canny edge condition", + "Figure 11. Combination of pose-free ID-Patch ControlNet and pretrained Canny edge ControlNet. We can generate background and clothing details, with high quality ID injection." + ], + "image_footnote": [], + "bbox": [ + 174, + 441, + 326, + 613 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cd76eb79b1d113e2e0821f7eba9c42e75589c1f7cb797af3760b44ff870d46e4.jpg", + "image_caption": [ + "(c) ID-Patch result" + ], + "image_footnote": [], + "bbox": [ + 328, + 441, + 480, + 613 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "of the same identities, showcasing varied expressions and lighting conditions, might help alleviate this problem.", + "bbox": [ + 89, + 696, + 482, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Although the resemblance has been improved with our approach, there still exists a performance gap between generating a single person and multiple people (Fig. 8(a)). This means generating personalized group photo remains a challenging problem and requires further study.", + "bbox": [ + 89, + 726, + 482, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 89, + 814, + 209, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In conclusion, our novel method, ID-Patch, markedly enhances identity resemblance and positioning in group photo generation. By embedding each identity feature within a distinct patch and utilizing ControlNet to accurately place", + "bbox": [ + 89, + 839, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/43b1cc7aa18c5f64ece6e1b58d165bd95396cd810575f3106d46a0c427374afa.jpg", + "image_caption": [ + "(a) Abnormal anatomy" + ], + "image_footnote": [], + "bbox": [ + 516, + 441, + 679, + 568 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9a097e59cbff5246aca59814e4a35c6d2ef07eac98f6733396221a3ec0728a30.jpg", + "image_caption": [ + "(b) Over-fitting input", + "Figure 12. Limitations. (a) Abnormal anatomy can be found in complex scenes, e.g. the forearms in the red box. (b) When input faces contain strong shadows, our method struggles to fully disentangle user's ID in generated result." + ], + "image_footnote": [], + "bbox": [ + 681, + 441, + 736, + 569 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5a5123a3951aa4aded6e60d8afc6fc00f799a2dfa8faf1931d51b1ba1520fea0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 735, + 441, + 900, + 569 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "each identity at the designated spatial locations, we successfully reduce ID leakage. Our approach seamlessly integrates with additional conditioning signals such as pose control. Experimental results underscore the superiority of ID-Patch, particularly in complex scenarios involving more than three identities. This work paves the way for future explorations in multi-ID image generation. Potential future research directions include leveraging multiple images of the same individual from different angles to further enhance identity resemblance and simultaneously controlling location and facial expressions using the patch technique.", + "bbox": [ + 511, + 656, + 906, + 823 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Societal Impact: Our work studies ID injection from a technical perspective and is not intended for malicious use. However, inappropriate usage might generate undesired fake images. Users should be aware of the ethical implications and use it responsibly.", + "bbox": [ + 511, + 824, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2993", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Omri Avrahami, Kfir Aberman, Ohad Fried, Daniel Cohen-Or, and Dani Lischinski. Break-a-scene: Extracting multiple concepts from a single image. In SIGGRAPH Asia 2023 Conference Papers, pages 1–12, 2023. 3", + "[2] Omri Avrahami, Thomas Hayes, Oran Gafni, Sonal Gupta, Yaniv Taigman, Devi Parikh, Dani Lischinski, Ohad Fried, and Xi Yin. Spatext: Spatio-textual representation for controllable image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18370-18380, 2023. 3", + "[3] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Ji-aming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, et al. ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 2", + "[4] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 2", + "[5] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 5", + "[6] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2", + "[7] Ganggui Ding, Canyu Zhao, Wen Wang, Zhen Yang, Zide Liu, Hao Chen, and Chunhua Shen. Freecustom: Tuning-free customized image generation for multi-concept composition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9089-9098, 2024. 3", + "[8] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis, march 2024. URL http://arxiv.org/abs/2403.03206. 2", + "[9] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2", + "[10] Rinon Gal, Moab Arar, Yuval Atzmon, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Encoder-based domain tuning for fast personalization of text-to-image models. ACM Transactions on Graphics (TOG), 42(4):1-13, 2023. 2", + "[11] Zigang Geng, Ke Sun, Bin Xiao, Zhaoxiang Zhang, and Jingdong Wang. Bottom-up human pose estimation via disentangled keypoint regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14676-14686, 2021. 1", + "[12] Yuchao Gu, Xintao Wang, Jay Zhangjie Wu, Yujun Shi, Yunpeng Chen, Zihan Fan, Wuyou Xiao, Rui Zhao, Shuning" + ], + "bbox": [ + 93, + 114, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Chang, Weijia Wu, et al. Mix-of-show: Decentralized low-rank adaptation for multi-concept customization of diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[13] Ligong Han, Yinxiao Li, Han Zhang, Peyman Milanfar, Dimitris Metaxas, and Feng Yang. Svdiff: Compact parameter space for diffusion fine-tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7323-7334, 2023. 3", + "[14] Junjie He, Yifeng Geng, and Liefeng Bo. Uniportrait: A unified framework for identity-preserving single- and multi-human image personalization. arXiv preprint arXiv:2408.05939, 2024. 3", + "[15] Zecheng He, Bo Sun, Felix Juefei-Xu, Haoyu Ma, Ankit Ramchandani, Vincent Cheung, Siddharth Shah, Anmol Kalia, Harihar Subramanyam, Alireza Zareian, et al. Imagine yourself: Tuning-free personalized image generation. arXiv preprint arXiv:2409.13346, 2024. 2", + "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2", + "[17] Chanran Kim, Jeongin Lee, Shichang Joung, Bongmo Kim, and Yeul-Min Baek. Instantfamily: Masked attention for zero-shot multi-id image generation. arXiv preprint arXiv:2404.19427, 2024. 1, 2, 3, 4, 5", + "[18] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3", + "[19] Zhe Kong, Yong Zhang, Tianyu Yang, Tao Wang, Kaihao Zhang, Bizhu Wu, Guanying Chen, Wei Liu, and Wenhan Luo. Omg: Occlusion-friendly personalized multi-concept generation in diffusion models. In ECCV, 2024. 1, 2, 5", + "[20] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 3", + "[21] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 1", + "[22] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22511-22521, 2023. 3", + "[23] Zhen Li, Mingdeng Cao, Xintao Wang, Zhongang Qi, MingMing Cheng, and Ying Shan. Photomaker: Customizing realistic human photos via stacked id embedding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8640-8650, 2024. 2", + "[24] Zhiheng Liu, Ruili Feng, Kai Zhu, Yifei Zhang, Kecheng Zheng, Yu Liu, Deli Zhao, Jingren Zhou, and Yang Cao. Cones: Concept neurons in diffusion models for customized generation. arXiv preprint arXiv:2303.05125, 2023. 3", + "[25] Nanye Ma, Mark Goldstein, Michael S Albergo, Nicholas M Boffi, Eric Vanden-Eijnden, and Saining Xie. Sit: Exploring flow and diffusion-based generative models with scalable" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2994", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "interpolant transformers. arXiv preprint arXiv:2401.08740, 2024. 2", + "[26] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2", + "[27] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 2", + "[28] Xu Peng, Junwei Zhu, Boyuan Jiang, Ying Tai, Donghao Luo, Jiangning Zhang, Wei Lin, Taisong Jin, Chengjie Wang, and Rongrong Ji. Portraitbooth: A versatile portrait model for fast identity-preserved personalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27080-27090, 2024. 2", + "[29] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. In The Twelfth International Conference on Learning Representations, 2024. 5", + "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 7", + "[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3", + "[32] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 2", + "[33] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdreambooth: Hypernetworks for fast personalization of text-to-image models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6527-6536, 2024. 2", + "[34] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 2", + "[35] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 815-823, 2015. 5", + "[36] Noam Shazeer and Mitchell Stern. Adafactor: Adaptive learning rates with sublinear memory cost. In International" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference on Machine Learning, pages 4596-4604. PMLR, 2018. 5", + "[37] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 2", + "[38] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.", + "[39] Yang Sui, Huy Phan, Jinqi Xiao, Tianfang Zhang, Zijie Tang, Cong Shi, Yan Wang, Yingying Chen, and Bo Yuan. Disdet: Exploring detectability of backdoor attack on diffusion models. arXiv preprint arXiv:2402.02739, 2024. 2", + "[40] Yoad Tewel, Rinon Gal, Gal Chechik, and Yuval Atzmon. Key-locked rank one editing for text-to-image personalization. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 3", + "[41] Hanhui Wang, Yihua Zhang, Ruizheng Bai, Yue Zhao, Sijia Liu, and Zhengzhong Tu. Edit away and my face will not stay: Personal biometric defense against malicious generative editing. arXiv preprint arXiv:2411.16832, 2024. 2", + "[42] Kuan-Chieh Wang, Daniil Ostashev, Yuwei Fang, Sergey Tulyakov, and Kfir Aberman. Moa: Mixture-of-attention for subject-context disentanglement in personalized image generation. arXiv preprint arXiv:2404.11565, 2024. 3", + "[43] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, and Anthony Chen. Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519, 2024. 1, 2, 5", + "[44] Xudong Wang, Trevor Darrell, Sai Saketh Rambhatla, Rohit Girdhar, and Ishan Misra. Instancediffusion: Instance-level control for image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6232-6242, 2024. 3", + "[45] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15943-15953, 2023. 2", + "[46] Yuxin Wen, Neel Jain, John Kirchenbauer, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Hard prompts made easy: Gradient-based discrete optimization for prompt tuning and discovery. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[47] Guangxuan Xiao, Tianwei Yin, William T Freeman, Frédo Durand, and Song Han. Fastcomposer: Tuning-free multisubject image generation with localized attention. International Journal of Computer Vision, pages 1-20, 2024. 3", + "[48] Jinqi Xiao, Miao Yin, Yu Gong, Xiao Zang, Jian Ren, and Bo Yuan. COMCAT: Towards efficient compression and customization of attention-based vision models. In Proceedings of the 40th International Conference on Machine Learning, pages 38125-38136. PMLR, 2023. 2", + "[49] Jinqi Xiao, Shen Sang, Tiancheng Zhi, Jing Liu, Qing Yan, Linjie Luo, and Bo Yuan. Coap: Memory-efficient training" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2995", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "with correlation-aware gradient projection. arXiv preprint arXiv:2412.00071, 2024. 2", + "[50] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 2, 4", + "[51] Ge Yuan, Xiaodong Cun, Yong Zhang, Maomao Li, Chenyang Qi, Xintao Wang, Ying Shan, and Huicheng Zheng. Inserting anybody in diffusion models via celeb basis. arXiv preprint arXiv:2306.00926, 2023. 2", + "[52] Kaipeng Zhang, Zhanpeng Zhang, Zhifeng Li, and Yu Qiao. Joint face detection and alignment using multitask cascaded convolutional networks. IEEE signal processing letters, 23 (10):1499-1503, 2016. 1", + "[53] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 3", + "[54] Yimeng Zhang, Xin Chen, Jinghan Jia, Yihua Zhang, Chongyu Fan, Jiancheng Liu, Mingyi Hong, Ke Ding, and Sijia Liu. Defensive unlearning with adversarial training for robust concept erasure in diffusion models. arXiv preprint arXiv:2405.15234, 2024. 2", + "[55] Yimeng Zhang, Jinghan Jia, Xin Chen, Aochuan Chen, Yihua Zhang, Jiancheng Liu, Ke Ding, and Sijia Liu. To generate or not? safety-driven unlearned diffusion models are still easy to generate unsafe images... for now. In European Conference on Computer Vision, pages 385-403. Springer, 2024. 2", + "[56] Yihua Zhang, Yimeng Zhang, Yuguang Yao, Jinghan Jia, Jiancheng Liu, Xiaoming Liu, and Sijia Liu. Unlearncanvas: A stylized image dataset to benchmark machine unlearning for diffusion models. arXiv preprint arXiv:2402.11846, 2024. 2", + "[57] Yufan Zhou, Ruiyi Zhang, Tong Sun, and Jinhui Xu. Enhancing detail preservation for customized text-to-image generation: A regularization-free approach. arXiv preprint arXiv:2305.13579, 2023. 2" + ], + "bbox": [ + 91, + 92, + 483, + 627 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "2996", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/ID-Patch_ Robust ID Association for Group Photo Personalization/a37d3ef3-36c1-46f2-886c-58bf6589bcc4_model.json b/2025/ID-Patch_ Robust ID Association for Group Photo Personalization/a37d3ef3-36c1-46f2-886c-58bf6589bcc4_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d912878cd19728864c55e250dbb12e093e320867 --- /dev/null +++ b/2025/ID-Patch_ Robust ID Association for Group Photo Personalization/a37d3ef3-36c1-46f2-886c-58bf6589bcc4_model.json @@ -0,0 +1,2686 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.166, + 0.131, + 0.834, + 0.154 + ], + "angle": 0, + "content": "ID-Patch: Robust ID Association for Group Photo Personalization" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.18, + 0.737, + 0.219 + ], + "angle": 0, + "content": "Yimeng Zhang\\(^{1,2,*}\\), Tiancheng Zhi\\(^{1}\\), Jing Liu\\(^{1}\\), Shen Sang\\(^{1}\\), Liming Jiang\\(^{1}\\), Qing Yan\\(^{1}\\), Sijia Liu\\(^{2}\\), Linjie Luo\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.312, + 0.233, + 0.68, + 0.253 + ], + "angle": 0, + "content": "1 ByteDance Inc., 2 Michigan State University" + }, + { + "type": "image_caption", + "bbox": [ + 0.107, + 0.286, + 0.286, + 0.297 + ], + "angle": 0, + "content": "Pose Condition & ID Location" + }, + { + "type": "image_caption", + "bbox": [ + 0.126, + 0.301, + 0.267, + 0.32 + ], + "angle": 0, + "content": "\"An indoor celebration scene featuring three women...\"" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.322, + 0.293, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.344, + 0.287, + 0.445, + 0.297 + ], + "angle": 0, + "content": "OMG + InstantID" + }, + { + "type": "image_caption", + "bbox": [ + 0.323, + 0.304, + 0.466, + 0.315 + ], + "angle": 0, + "content": "Generation Time: 71.1s" + }, + { + "type": "image", + "bbox": [ + 0.301, + 0.322, + 0.496, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.554, + 0.287, + 0.639, + 0.299 + ], + "angle": 0, + "content": "InstantFamily" + }, + { + "type": "image_caption", + "bbox": [ + 0.533, + 0.304, + 0.669, + 0.315 + ], + "angle": 0, + "content": "Generation Time: 9.6s" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.322, + 0.699, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.755, + 0.287, + 0.85, + 0.299 + ], + "angle": 0, + "content": "ID-Patch (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.734, + 0.305, + 0.869, + 0.315 + ], + "angle": 0, + "content": "Generation Time: 9.4s" + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.322, + 0.901, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.484, + 0.907, + 0.568 + ], + "angle": 0, + "content": "Figure 1. Comparison with state-of-the-art multi-identity generation methods. From left to right: the condition inputs followed by results generated using OMG [19] (with InstantID [43]), InstantFamily [17], and our proposed ID-Patch approach. Red dashed boxes highlight failures. OMG fails to preserve the hairstyles of the middle person, and creates artifacts for the right woman's hand, possibly because of the inconsistency between its generation stages (explained in Fig. 2). InstantFamily suffers from ID leakage, resulting the incorrect ID of the middle person. Our approach preserves the detailed identity of each person. In addition, our approach is 7 times faster than OMG and has less computational overhead than InstantFamily." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.58, + 0.327, + 0.596 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.615, + 0.486, + 0.873 + ], + "angle": 0, + "content": "The ability to synthesize personalized group photos and specify the positions of each identity offers immense creative potential. While such imagery can be visually appealing, it presents significant challenges for existing technologies. A persistent issue is identity (ID) leakage, where injected facial features interfere with one another, resulting in low face resemblance, incorrect positioning, and visual artifacts. Existing methods suffer from limitations such as the reliance on segmentation models, increased runtime, or a high probability of ID leakage. To address these challenges, we propose ID-Patch, a novel method that provides robust association between identities and 2D positions. Our approach generates an ID patch and ID embeddings from the same facial features: the ID patch is positioned on the conditional image for precise spatial control, while the ID embeddings integrate with text embeddings to ensure high resemblance. Experimental results demon" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.582, + 0.907, + 0.643 + ], + "angle": 0, + "content": "strate that ID-Patch surpasses baseline methods across metrics, such as face ID resemblance, ID-position association accuracy, and generation efficiency. Project Page is: https://byteaigc.github.io/ID-Patch/" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.677, + 0.646, + 0.693 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.705, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Personalized group photo generation has wide-ranging applications in entertainment, advertising, social media, and virtual reality. For example, users might want to create a group photo of friends at a virtual party, even if some could not attend in person. This technology enables the creation of engaging and memorable content that reflects real-life social interactions and relationships. However, this task poses significant challenges, particularly the risk of identity (ID) leakage, where the identity information of one individual inadvertently affects the representation of another at a different position. Such blending compromises the distinctiveness and accuracy of each person's features. Besides, being able to specify the position of each identity in the image is" + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.887, + 0.355, + 0.901 + ], + "angle": 0, + "content": "*Work done during internship at ByteDance." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2986" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.136 + ], + "angle": 0, + "content": "an important function when users want to create a specific layout of people. Incorrect positioning of identities may lead to undesired results." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.14, + 0.484, + 0.515 + ], + "angle": 0, + "content": "State-of-the-art group photo personalization methods largely rely on diffusion models [3, 6, 16, 26, 37-39, 41, 48, 49, 54, 56], which deliver superior visual quality when paired with effective prompt engineering [4, 8, 25, 27, 34, 55]. To address the issue of ID leakage, OMG [19] first generates an image without ID injection. Then, it estimates segmentation masks for the individuals in the image and subsequently injects the IDs into each segmented region. However, it depends on a segmentation model, which might fail in challenging cases. Besides, the content generated in the two stages may conflict with each other, leading to reduced visual fidelity. See Fig.1 and Fig.2 for example. Moreover, OMG requires running the denoising process separately for each identity, resulting in a runtime that increases linearly with the number of people. Alternatively, InstantFamily [17] employs a single-pass generation strategy by manipulating cross-attention masks. This approach uses approximate head region masks for each ID to direct pixel attention exclusively to specific ID embeddings derived from facial features. However, it struggles to prevent ID leakage due to two key issues: (1) imprecise masks or close proximity of faces, which can cause overlaps, and (2) unintended information propagation through self-attention and convolutional layers. Please see Fig.1 and Fig.3 for more details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.52, + 0.484, + 0.744 + ], + "angle": 0, + "content": "In response to these limitations, we propose a novel method named ID-Patch, designed for efficient and robust ID-position association in group photo generation. Our idea is to let the model learn to associate input IDs with their spatially designated locations. Specifically, we project facial features onto a small RGB image patch called ID patch and token embeddings called ID embeddings. The ID patches are placed onto the conditioning image of a ControlNet [53] according to face locations to ensure precise position control of IDs within the generated images. ID embeddings are integrated with text embeddings to enhance facial details. Our ID-Patch method seamlessly integrates with various types of spatial conditions, such as poses, canny edges or depth map, enhancing the robustness and flexibility of our method." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.748, + 0.418, + 0.763 + ], + "angle": 0, + "content": "We summarize our key contributions as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.765, + 0.483, + 0.825 + ], + "angle": 0, + "content": "- ID-Patch Method: We introduce a novel approach for multi-ID image generation that directly links face ID features with their locations using visual patches, ensuring accurate resemblance and position control." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.826, + 0.483, + 0.886 + ], + "angle": 0, + "content": "- Efficiency Improvement: Our method enhances the efficiency of multi-ID image generation by simplifying the process to adding ID patches and concatenating ID embeddings, reducing computational overhead." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "- Simplified Control: We eliminate the need for auxiliary" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.09, + 0.643, + 0.186 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.529, + 0.189, + 0.633, + 0.201 + ], + "angle": 0, + "content": "(a) OMG first stage" + }, + { + "type": "image", + "bbox": [ + 0.645, + 0.09, + 0.771, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.664, + 0.19, + 0.753, + 0.2 + ], + "angle": 0, + "content": "(b) Human mask" + }, + { + "type": "image", + "bbox": [ + 0.774, + 0.09, + 0.899, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.782, + 0.189, + 0.891, + 0.2 + ], + "angle": 0, + "content": "(c) OMG final result" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.203, + 0.905, + 0.286 + ], + "angle": 0, + "content": "Figure 2. Why does OMG fail in Fig. 1? OMG relies on a first stage result (no ID awareness) and its human segmentation to control layout. The second stage paints injected IDs within corresponding masks. However, the second pass may conflict with the first stage semantics, leading to deteriorated ID resemblance (long hair in the middle of image (c)) or artifacts (hand in the middle)." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.294, + 0.642, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.564, + 0.422, + 0.598, + 0.434 + ], + "angle": 0, + "content": "(a) ID" + }, + { + "type": "image", + "bbox": [ + 0.645, + 0.294, + 0.772, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.646, + 0.422, + 0.772, + 0.435 + ], + "angle": 0, + "content": "(b) InstantFamily result" + }, + { + "type": "image", + "bbox": [ + 0.774, + 0.294, + 0.899, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.786, + 0.422, + 0.886, + 0.434 + ], + "angle": 0, + "content": "(c) ID-Patch result" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.436, + 0.905, + 0.478 + ], + "angle": 0, + "content": "Figure 3. ID leakage of InstantFamily [17]. The text prompt describes a man and a woman standing together, and ID (a) is only injected into the left man." + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.5, + 0.905, + 0.545 + ], + "angle": 0, + "content": "segmentation models, requiring only a single point for ID position control, thereby increasing robustness and reducing complexity." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.545, + 0.905, + 0.605 + ], + "angle": 0, + "content": "- Superior Performance: Our approach achieves higher fidelity in ID resemblance and better position control, especially in complex images with multiple identities, as demonstrated by our experimental results." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.619, + 0.655, + 0.634 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Single-ID Personalization. The personalization of diffusion models has emerged as a popular research topic. Finetuning or optimization based methods [9, 10, 32, 33, 46] have the potential to reach high quality results in terms of ID resemblance, but are time-consuming. Inference-only methods [15, 23, 28, 43, 45, 50, 51, 57] eliminate the finetuning process usually by encoding the concept into token embeddings and injecting them into the diffusion models via cross-attention. However, most of these works focus on single concept injection rather than group photo generation. Multi-ID Personalization. The main challenge of personalized group photo generation is ID blending, i.e., the identity information of person A may affect the generation of person B at another position. To resolve this issue, OMG [19] performs separate denoising processes. First, a noncustomized image is generated and masks of each person are extracted. Second, masked regions are painted using" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2987" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.905, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.272, + 0.907, + 0.342 + ], + "angle": 0, + "content": "Figure 4. Our pipeline. Given a text prompt (e.g., two people shaking hands), \\(N\\) face images and locations, we generate an image with \\(N\\) IDs. We extract face features for each ID and project them into ID patches and ID embeddings. ID patches are rendered on a black canvas (or added on top of a pose image) according to face locations and sent into a ControlNet to control the positions of generated faces. ID embeddings are appended to text embeddings to provide detailed face information to the diffusion model and ControlNet via cross-attention." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.358, + 0.486, + 0.614 + ], + "angle": 0, + "content": "separate customized models and stitched together. This approach is slow and conflicts between stages may cause identity loss. Additionally, it relies on a segmentation module, which may fail in challenging cases. To avoid running separate customized models, recent works [14, 42, 47] let the network to learn to attend to specific concept tokens by supervising the attention mask using segmentation or ID routing. However, these methods do not allow the specification of the accurate locations of each ID. Although they can be combined with pose conditions, describing the assignment of IDs to positions in text may be hard. InstantFamily [17] provides an opportunity to specify the approximate location of each ID by manipulating the attention masks during both training and inference. However, no constraints have been imposed to avoid the ID leakage through self-attention and convolutional layers, leading to ID blending, omitting or swapping." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.615, + 0.485, + 0.782 + ], + "angle": 0, + "content": "Multiple General Concepts. Besides human generation, there are series of work focusing on the customization of multiple general concepts [1, 2, 7, 12, 13, 20, 24, 40, 47] and grounded image generation [2, 22, 44] associating objects with positions. GLIGEN [22] integrates grounding tokens, which align textual prompts with specific image regions. SpaText [2] and InstanceDiffusion [44] allows users to define the placement and attributes of objects. However, these concepts are usually semantically distinct while it is hard to describe the human identities in text and subtle details matter." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.799, + 0.182, + 0.815 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.485, + 0.903 + ], + "angle": 0, + "content": "In this section, we first outline some preliminaries about text-to-image diffusion models and ControlNets. Subsequently, we detail the problem formulation for multi-ID image generation and the design of our modules, including the ID patch and ID embeddings, as well as the training and" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.358, + 0.644, + 0.372 + ], + "angle": 0, + "content": "inference schemes." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.381, + 0.655, + 0.397 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.405, + 0.907, + 0.556 + ], + "angle": 0, + "content": "Latent diffusion models (LDMs). LDMs [31] are image generation models that can incorporate conditioning signals into the generation process. Specifically, during training, an encoder \\(\\mathcal{E}\\) of variational autoencoder (VAE) [18] encodes an image \\(x\\) into a latent space, and a neural network \\(\\epsilon_{\\theta}(z_t,t,c)\\) performs the denoising process within this latent space. It attempts to predict the original Gaussian noise \\(\\epsilon\\) from a noisy latent variable \\(z_{t}\\), where \\(t\\) represents the diffusion timestep, and \\(c\\) is the conditioning signal, injected via cross-attention. The loss function for an LDM is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.556, + 0.567, + 0.907, + 0.588 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {L D M}} = \\mathbb {E} _ {\\mathcal {E} (x), \\epsilon \\sim \\mathcal {N} (0, 1), t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (z _ {t}, t, c) \\| _ {2} ^ {2} \\right] \\quad (1)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.596, + 0.909, + 0.764 + ], + "angle": 0, + "content": "ControlNet. The incorporation of ControlNet [53] into diffusion models marks a significant advancement in controlled image synthesis. ControlNet is designed to steer the image generation process by applying spatial conditioning signals such as rasterized pose images, depth maps, or canny edges. It introduces a control network that processes these signals and adds their features to the base model's intermediate results. The control network is a trainable copy of the base model's encoder layers and is initialized with zero convolution. This effective integration achieves high quality controlled generation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.772, + 0.766, + 0.789 + ], + "angle": 0, + "content": "3.2. Multi-ID Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.903 + ], + "angle": 0, + "content": "This paper focuses on efficiently and effectively addressing the problem of multi-ID image generation utilizing conditioning signals that include text prompts, identity features, and face locations. The inputs to our model comprise text prompt embeddings \\( c_{t} \\), facial features \\( c_{f} = (f_{0}, f_{1}, \\ldots, f_{N-1}) \\) for \\( N \\) distinct identities, and their corresponding facial locations \\( c_{l} = (l_{0}, l_{1}, \\ldots, l_{N-1}) \\). The" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2988" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.089, + 0.188, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.128, + 0.233, + 0.16, + 0.244 + ], + "angle": 0, + "content": "(a) ID" + }, + { + "type": "image", + "bbox": [ + 0.19, + 0.089, + 0.334, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.201, + 0.233, + 0.325, + 0.245 + ], + "angle": 0, + "content": "(b) w/o ID embeddings" + }, + { + "type": "image", + "bbox": [ + 0.334, + 0.089, + 0.479, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.343, + 0.233, + 0.469, + 0.245 + ], + "angle": 0, + "content": "(c) with ID embeddings" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.256, + 0.483, + 0.312 + ], + "angle": 0, + "content": "Figure 5. Effectiveness of ID embedding. Without ID embeddings, one can distinguish between the two people but the resemblance is low. Incorporating ID embeddings significantly improves face resemblance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.336, + 0.484, + 0.398 + ], + "angle": 0, + "content": "facial location \\( l_{i} = (x_{i},y_{i}) \\) identifies the xy-coordinates of the nose tip for each identity within the intended image. The goal is to synthesize an image following the text description with each given face located in the designated place." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.42, + 0.364, + 0.435 + ], + "angle": 0, + "content": "3.2.1 ID Patches and ID Embeddings" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.446, + 0.483, + 0.627 + ], + "angle": 0, + "content": "As discussed in Sec. 3.1, ControlNet effectively integrates spatial conditional signals into diffusion models. Building on this capability, we introduce a method to encode ID location information using condition images, subsequently processed by ControlNet to guide ID localization within the generated image. Our approach features ID conditions manifesting as both patches and embeddings. Specifically, ID patches, precisely aligned with facial locations, are integrated via ControlNet to ensure accurate identity localization, while ID embeddings are utilized via cross-attention mechanisms in the LDM to enhance the detailed representation of facial features, as illustrated in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.628, + 0.483, + 0.794 + ], + "angle": 0, + "content": "Both ID patches and ID embeddings are derived from facial features, enabling a synergistic association between them, with both projection processes facilitated by the Perceiver Resampler [50]. For ID patches, we convert the facial feature \\( f_{i} \\) into an identity-specific patch \\( p_i \\in \\mathbb{R}^{P \\times P \\times 3} \\), functioning similarly to a QR code, uniquely identifying each person and easily distinguishable from others. These ID patches are placed on a black canvas at specified facial locations to create an image \\( I \\), which then serves as the conditioning input for ControlNet, thereby providing precise identity localization information." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.901 + ], + "angle": 0, + "content": "As depicted in Fig. 5, while ID patches alone are generally sufficient for distinguishing between identities, they may not always achieve optimal results in generating images with high ID resemblance. To address this limitation, and inspired by InstantFamily [17], we also incorporate ID embeddings. These embeddings are designed to inject detailed facial appearance information into the dif" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.089, + 0.673, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.519, + 0.22, + 0.668, + 0.233 + ], + "angle": 0, + "content": "(a) ID (b) Result w/o pose" + }, + { + "type": "image", + "bbox": [ + 0.675, + 0.089, + 0.789, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.677, + 0.22, + 0.786, + 0.233 + ], + "angle": 0, + "content": "(c) ID patch on pose" + }, + { + "type": "image", + "bbox": [ + 0.792, + 0.089, + 0.905, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.795, + 0.22, + 0.901, + 0.233 + ], + "angle": 0, + "content": "(d) Result with pose" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.245, + 0.905, + 0.327 + ], + "angle": 0, + "content": "Figure 6. ID-Patch combined with pose conditions. Provided with user ID images in (a), our method can generate results with only the specifications of nose tip locations as shown in (b). Merging ID patches with pose images (c) enhances spatial control over the generated results as seen in (d), without incurring any computational overhead." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.346, + 0.906, + 0.467 + ], + "angle": 0, + "content": "fusion model, thus enhancing the resemblance of the generated identities to their real-life counterparts. Specifically, ID embeddings \\( w_{i} \\in \\mathbb{R}^{2048 \\times M} \\), share the same token dimension 2048 as the SDXL text embeddings. As shown in Fig.4, these embeddings are appended to text embeddings \\( c_{t} \\) to form an extended text embedding \\( c_{t}' = [c_{t}, w_{0}, w_{1}, \\ldots, w_{N-1}] \\), which is then fed into both the UNet and ControlNet through cross-attention mechanisms." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.468, + 0.906, + 0.678 + ], + "angle": 0, + "content": "Since human pose conditions are often utilized to enhance group photo generation, we can combine our ID-Patch approach with the pose conditions for multi-id image generation. As illustrated in Fig. 6, the ID patches can be overlaid onto a pose-conditioned image, creating a composite condition image that encodes both identity positioning and pose information. It is important to mention that while our method can incorporate pose information, it does not inherently require it; pose is just an optional enhancement, one can also choose other conditions like canny edge, depth map for specific applications. The adaptability of our ID-Patch approach presents enhanced control over generated results yet without incurring the extra overhead of an additional ControlNet branch." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.699, + 0.706, + 0.715 + ], + "angle": 0, + "content": "3.2.2 Two-Stage Training" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.724, + 0.905, + 0.753 + ], + "angle": 0, + "content": "The training of our model employs a conventional diffusion loss function, defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.564, + 0.765, + 0.905, + 0.784 + ], + "angle": 0, + "content": "\\[\nL = \\mathbb {E} _ {\\mathcal {E} (x), \\epsilon \\sim \\mathcal {N} (0, 1), t} [ \\| \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}, t, I, c _ {t} ^ {\\prime}\\right) \\| _ {2} ^ {2} ] \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Initially, one might consider training the models end-to-end from scratch. However, our experiments reveal that such a method does not consistently ensure accurate ID localization, as depicted in Fig.7. This shortcoming likely stems from the network's tendency to preferentially inject ID information into ID embeddings, rather than ID patches, resulting in non-distinctive ID patches that could potentially" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2989" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.113, + 0.223, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.124, + 0.257, + 0.197, + 0.268 + ], + "angle": 0, + "content": "(a) Pose + ID" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.089, + 0.351, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.231, + 0.257, + 0.345, + 0.269 + ], + "angle": 0, + "content": "(b) Single-stage train" + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.089, + 0.478, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.365, + 0.257, + 0.466, + 0.269 + ], + "angle": 0, + "content": "(c) Two-stage train" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.281, + 0.483, + 0.42 + ], + "angle": 0, + "content": "Figure 7. Two-stage training to improve positioning robustness. Given pose and ID conditions in (a), single-stage training cannot fully prevent the incorrect face positioning issue. For example, in (b), the man is mistakenly placed at the central bottom position, generating an inharmonious result. (c) Two-stage training is introduced to solve this problem. As can be seen on the top row of this figure, the ID patches from our two-stage training approach are visually more distinguishable from each other, compared to the ones from single-stage training. Our experimental results prove that this is critical in solving the ID leak issues." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.433, + 0.484, + 0.569 + ], + "angle": 0, + "content": "confuse ControlNet. In the single-stage model, as shown in Fig. 7, the ID patches exhibit only subtle differences between different individuals. To resolve this, we introduce a two-stage training approach. In the initial stage, ID embeddings are omitted, compelling the network to prioritize the incorporation of ID information into ID patches, thereby enhancing their distinctiveness. In the subsequent stage, ID embeddings are introduced and the model is fine-tuned to further improve the resemblance of the generated images." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.585, + 0.222, + 0.601 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.61, + 0.189, + 0.626 + ], + "angle": 0, + "content": "4.1. Settings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.634, + 0.483, + 0.876 + ], + "angle": 0, + "content": "Implementation details. Facial features \\( f_{i} \\) we use are 512-dimensional vectors derived from ArcFace [5]. During training, we utilize the SDXL diffusion model [29]. During the inference phase, we employ the further finetuned SDXL, namely Juggernaut-X-v10*, as base model to enhance portrait quality. The inference process involves 50 timesteps. We set the ID token injection ratio to 0.8 to regulate the influence of identity features, where the initial \\( 20\\% \\) of timesteps operate without ID token embeddings to focus on the general style, followed by the remaining \\( 80\\% \\) timesteps which incorporate ID token embeddings to refine the identity-specific aspects of the images. The default size of ID-patch is \\( P \\times P = 64 \\times 64 \\) and the default length of ID embeddings is \\( M = 16 \\). The network is trained using 8 NVIDIA H100 GPUs with batch size=4, learning rate=1e-05, weight decay = 0.01, using AdaFactor optimizer [36]." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.905, + 0.173 + ], + "angle": 0, + "content": "Table 1. Performance comparison across four dimensions: identity resemblance, ID-position association accuracy, text alignment, and generation time. For qualitative metrics to evaluate similarity or accuracy, higher is better. Our approach significantly outperforms baseline methods in terms of resemblance, association, and generation time, with similar text alignment scores." + }, + { + "type": "table", + "bbox": [ + 0.527, + 0.175, + 0.89, + 0.242 + ], + "angle": 0, + "content": "
MethodsID ↑Association ↑Text ↑Time (s) ↓
OMG + InstantID0.6440.9260.27184.83
InstantFamily0.6830.8620.27810.02
ID-Patch (Ours)0.7510.9580.2739.69
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.262, + 0.905, + 0.308 + ], + "angle": 0, + "content": "We train it for 720k iterations for stage 1, and stage 2 usually converges around 1 million iterations. We select the best model based on a 20 people validation set." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.309, + 0.906, + 0.459 + ], + "angle": 0, + "content": "Dataset. Our training set consists of 17 million single-person and 1.95 million multi-person images. These images are obtained from purchased data and publicly available sources. For evaluation, we utilize two specialized datasets: 1) Style Dataset comprises 40 styles with pose and text prompt definitions, ranging from one to eight faces. 2) Identity Dataset includes 600 unique identities representing a variety of races and genders. Each style is paired with 20 randomly sampled unique identity combinations for quantitative evaluation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.462, + 0.906, + 0.656 + ], + "angle": 0, + "content": "Baselines. Our analysis includes two state-of-the-art multi-ID generation methods as baselines: 1) OMG [19]: This method incorporates multiple identity features into separate segmented regions through latent stitching. It can be paired with a single ID injection method InstantID [43] for multi-ID generation. 2) InstantFamily [17]: This approach employs coarse attention masks within cross-attention layers, a mandatory process that ensures distinct identity features are injected into different image instances effectively. Because there is no official publicly available code for InstantFamily [17], for fairness, we implemented it based on the SDXL base model [29], and trained it using the same training data as our model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Evaluation Metrics. We evaluate the methods in a pose-conditioned setting, because pose is a requirement of InstantFamily, and without pose condition, OMG is not able to generate faces at specified locations for its first stage. Besides, pose condition makes the head size consistent, which is easier for resemblance evaluation and reduces the impact of head sizes on generation quality. For comprehensive evaluation, we consider four critical dimensions: 1) Identity resemblance: we measure the identity resemblance by calculating the cosine similarity (CosSim) between the facial features of faces in the generated images and the reference faces. For fairness, we use FaceNet [35], a face model different from the ArcFace model used for training. 2) Association accuracy: we assess the accuracy of identity-position association in generated images by ensuring that the generated faces are most similar to their assigned coun" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.887, + 0.411, + 0.901 + ], + "angle": 0, + "content": "\\*https://huggingface.co/RunDiffusion/Juggernaut-X-v10" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2990" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.12, + 0.077, + 0.355, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.169, + 0.239, + 0.304, + 0.251 + ], + "angle": 0, + "content": "(a) Identity Resemblance" + }, + { + "type": "image", + "bbox": [ + 0.384, + 0.077, + 0.62, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.432, + 0.239, + 0.567, + 0.251 + ], + "angle": 0, + "content": "(b) Association Accuracy" + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.075, + 0.882, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.708, + 0.239, + 0.817, + 0.25 + ], + "angle": 0, + "content": "(c) Generation Time" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.263, + 0.908, + 0.306 + ], + "angle": 0, + "content": "Figure 8. Performance evaluation of different model generations from three different aspects: identity resemblance, location accuracy, and generation time. As face number increases, the ID and location metric scores of OMG and Instantfamily drop much more significantly than our approach. Our approach also achieves near-constant generation time, in contrast to OMG's linearly increasing running time." + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.314, + 0.898, + 0.688 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.699, + 0.908, + 0.77 + ], + "angle": 0, + "content": "Figure 9. Comparison with baselines on pose-conditioned generation, where red dashed boxes highlight instances with low identity resemblance. In row 1, OMG fails to preserve the face shape of the third woman (from left to right) because its first stage result conflicts with this ID's face shape. InstantFamily incorrectly generates the second person because of the ID leakage from the third woman. In row 2, OMG does not generate correct hair styles and accurate facial features for the two people in the red boxes, while InstantFamily generates the wrong ID in the red box." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.785, + 0.486, + 0.892 + ], + "angle": 0, + "content": "terparts, compared to all other input faces used in the generation process. In other words, we calculate the similarity between the \\(i\\)-th generated face and the \\(j\\)-th input face. Let \\(s(i) = \\arg \\max_{0 \\leq j \\leq N-1} \\mathrm{CosSim}(\\tilde{f}_i^{\\mathrm{gen}}, \\tilde{f}_j)\\) be the face index among all input faces that is most similar to the \\(i\\)-th generated face, where \\(\\tilde{f}_i^{\\mathrm{gen}}\\), \\(\\tilde{f}_j\\) denote the FaceNet features of the \\(i\\)-th generated face and the \\(j\\)-th input face respec" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.785, + 0.907, + 0.832 + ], + "angle": 0, + "content": "tively, and \\(N\\) is the number of faces in this image. When \\(s(i) = i\\), the \\(i\\)-th generated face has the correct association. The identity-position association accuracy is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.642, + 0.838, + 0.907, + 0.88 + ], + "angle": 0, + "content": "\\[\n\\frac {1}{N} \\sum_ {i = 0} ^ {N - 1} \\mathbf {1} \\{i = s (i) \\} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.886, + 0.907, + 0.902 + ], + "angle": 0, + "content": "This measurement assesses the correct placement of each" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.958 + ], + "angle": 0, + "content": "2991" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.482, + 0.173 + ], + "angle": 0, + "content": "Table 2. Ablation study results for our proposed ID-Patch variants, designed to assess the impact of different modules (patch projection and token projection) and training/inferencing schemes on the performance of multi-ID image generation. The full method achieves the best ID resemblance and ID-position association, indicating the necessity of each module." + }, + { + "type": "table", + "bbox": [ + 0.11, + 0.175, + 0.462, + 0.271 + ], + "angle": 0, + "content": "
VariantsID ↑Association ↑Text ↑
No patch projection0.3680.3450.271
No token projection0.5650.8950.284
Single stage training0.6600.7840.271
ID token injection ratio = 1.00.7500.9570.270
FULL0.7510.9580.273
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.296, + 0.484, + 0.448 + ], + "angle": 0, + "content": "identity within the image. 3) Text alignment: this metric evaluates the alignment between the textual context of the image generation prompt and the visual output. Cosine similarity is used to compare textual features extracted by the CLIP text encoder with visual features extracted by the CLIP image encoder from the generated images [30]. 4) Generation time: we measure the efficiency by recording the time (in seconds) required to generate an image on a NVIDIA A100 GPU, excluding the time taken for model loading and image I/O." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.473, + 0.336, + 0.49 + ], + "angle": 0, + "content": "4.2. Comparison with Baselines" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.502, + 0.482, + 0.609 + ], + "angle": 0, + "content": "As listed in Tab. 1, ID-Patch excels in identity and association metrics with scores of 0.751 and 0.958, respectively, showcasing its robust capability to maintain identity consistency and achieve precise placement in generated images. There is no significant difference in text alignment scores between compared methods. Notably, ID-Patch achieves the fastest image generation time." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Further details on the performance with varying face numbers are shown in Fig. 8. As we observed no deterioration in text alignment with an increase in the number of faces across all methods, we have relegated the corresponding figure to the appendix. As the number of faces increases, both OMG and InstantFamily experience significant performance drops due to ID leakage. Although ID-Patch also shows a decline, it is less severe, mainly because the increase in face numbers reduces the average face area, adversely affecting the quality of smaller faces produced by the SDXL models. Moreover, the presence of more faces slightly extends the image generation time due to the increased demand for facial feature extraction. InstantFamily is slightly slower than our approach due to the overhead of masked attention. For OMG, injecting identity features into designated segmented areas causes a substantial increase in generation time; it takes about 2 minutes to generate an image with eight people, whereas ID-Patch requires only about 10 seconds. See Fig. 9 for visual comparisons." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.667, + 0.108 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.113, + 0.905, + 0.386 + ], + "angle": 0, + "content": "We conduct an ablation study on variants of the proposed method in Tab. 2. The performance notably degrades when patch projection is omitted, leading to poor location alignment as the algorithm attempts to infer ID locations based on indirect cues such as gender and pose. Similarly, omitting token projection reduces resemblance by eliminating essential facial details, quantitatively validates the finding of Fig. 5. Single-stage training results in indistinguishable ID patches (as shown in Fig. 7), which impairs effective positioning. The default setting with an ID token injection ratio of 0.8 improves text alignment compared to injecting the ID token at every timestep, while maintaining the same level of identity resemblance. This improvement may arise because ID embeddings have little effect on facial detail during the early denoising steps, but become more influential in the later stages. It is important to note that even without the inference trick (i.e., using the ID token injection ratio), our proposed method still outperforms all baseline approaches." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.394, + 0.717, + 0.409 + ], + "angle": 0, + "content": "4.4. Pose-Free Generation" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.417, + 0.905, + 0.507 + ], + "angle": 0, + "content": "Our ID-Patch model can be operated effectively without the need for pose conditioning, a flexibility that permits its integration with other conditions for image generation. This capability enables the use of ID-Patch in versatile configurations, whether plug-and-play or co-trained, which we will explore in subsequent subsections." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.508, + 0.906, + 0.658 + ], + "angle": 0, + "content": "In the pose-free mode, the ControlNet uses a conditioning image that solely features ID patches against a black background. The placement of each ID patch at the tip of the nose ensures accurate preservation of facial locations during image synthesis, while head sizes and poses are variably generated based on the initial noise. As illustrated in Fig. 10, this methodology allows our model to produce aesthetically pleasing group photographs of different number of individuals across a variety of settings, demonstrating the robustness and adaptability of our approach." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.659, + 0.905, + 0.735 + ], + "angle": 0, + "content": "The pose-free version of ID-Patch is designed to be compatible with other ControlNets (plug-and-play). As illustrated in Fig. 11, it is possible to combine our ID-Patch ControlNet the Canny edge ControlNet to achieve more precise control over the final image outcomes." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.742, + 0.64, + 0.757 + ], + "angle": 0, + "content": "4.5. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.903 + ], + "angle": 0, + "content": "As illustrated in Fig. 12, our method encounters two primary limitations. First, our implementation is based on the SDXL model, which occasionally produces inaccuracies in body and finger anatomy. Transitioning to a more advanced base model could potentially rectify this issue. Second, facial features in the current model may incorporate lighting and expression details that are not fully separated from identity information, potentially compromising the generation quality. Enriching our training dataset with multiple images" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2992" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.113, + 0.127, + 0.188 + ], + "angle": 270, + "content": "ID + ID Location" + }, + { + "type": "image", + "bbox": [ + 0.151, + 0.086, + 0.285, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.31, + 0.086, + 0.481, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.086, + 0.697, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.718, + 0.086, + 0.889, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.255, + 0.128, + 0.335 + ], + "angle": 270, + "content": "Generation Result" + }, + { + "type": "image", + "bbox": [ + 0.152, + 0.228, + 0.284, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.31, + 0.228, + 0.481, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.228, + 0.697, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.717, + 0.228, + 0.889, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.38, + 0.907, + 0.437 + ], + "angle": 0, + "content": "Figure 10. ID-Patch pose-free generation. Here the condition image is only ID patches rendered onto a black canvas without any pose condition. The generation is conditioned on individual face locations while corresponding head sizes and body poses are inferred implicitly. Our method can accommodate a large number of people with diverse ethnic backgrounds, yet generating visually appealing group photo results. See \"Evaluation Metrics\" for reasons why we compare with baseline approaches only in pose-conditioned settings." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.492, + 0.172, + 0.553 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.553, + 0.171, + 0.614 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.617, + 0.15, + 0.628 + ], + "angle": 0, + "content": "(a) ID" + }, + { + "type": "image", + "bbox": [ + 0.176, + 0.443, + 0.327, + 0.614 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.184, + 0.617, + 0.318, + 0.628 + ], + "angle": 0, + "content": "(b) Canny edge condition" + }, + { + "type": "image", + "bbox": [ + 0.33, + 0.443, + 0.481, + 0.614 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.356, + 0.617, + 0.455, + 0.628 + ], + "angle": 0, + "content": "(c) ID-Patch result" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.641, + 0.483, + 0.684 + ], + "angle": 0, + "content": "Figure 11. Combination of pose-free ID-Patch ControlNet and pretrained Canny edge ControlNet. We can generate background and clothing details, with high quality ID injection." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.697, + 0.483, + 0.725 + ], + "angle": 0, + "content": "of the same identities, showcasing varied expressions and lighting conditions, might help alleviate this problem." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.727, + 0.483, + 0.803 + ], + "angle": 0, + "content": "Although the resemblance has been improved with our approach, there still exists a performance gap between generating a single person and multiple people (Fig. 8(a)). This means generating personalized group photo remains a challenging problem and requires further study." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.815, + 0.21, + 0.831 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In conclusion, our novel method, ID-Patch, markedly enhances identity resemblance and positioning in group photo generation. By embedding each identity feature within a distinct patch and utilizing ControlNet to accurately place" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.443, + 0.681, + 0.569 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.538, + 0.573, + 0.659, + 0.584 + ], + "angle": 0, + "content": "(a) Abnormal anatomy" + }, + { + "type": "image", + "bbox": [ + 0.683, + 0.443, + 0.738, + 0.57 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.736, + 0.443, + 0.901, + 0.57 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.736, + 0.573, + 0.847, + 0.585 + ], + "angle": 0, + "content": "(b) Over-fitting input" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.586, + 0.906, + 0.642 + ], + "angle": 0, + "content": "Figure 12. Limitations. (a) Abnormal anatomy can be found in complex scenes, e.g. the forearms in the red box. (b) When input faces contain strong shadows, our method struggles to fully disentangle user's ID in generated result." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.657, + 0.907, + 0.824 + ], + "angle": 0, + "content": "each identity at the designated spatial locations, we successfully reduce ID leakage. Our approach seamlessly integrates with additional conditioning signals such as pose control. Experimental results underscore the superiority of ID-Patch, particularly in complex scenarios involving more than three identities. This work paves the way for future explorations in multi-ID image generation. Potential future research directions include leveraging multiple images of the same individual from different angles to further enhance identity resemblance and simultaneously controlling location and facial expressions using the patch technique." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Societal Impact: Our work studies ID injection from a technical perspective and is not intended for malicious use. However, inappropriate usage might generate undesired fake images. Users should be aware of the ethical implications and use it responsibly." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2993" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.171 + ], + "angle": 0, + "content": "[1] Omri Avrahami, Kfir Aberman, Ohad Fried, Daniel Cohen-Or, and Dani Lischinski. Break-a-scene: Extracting multiple concepts from a single image. In SIGGRAPH Asia 2023 Conference Papers, pages 1–12, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.173, + 0.484, + 0.257 + ], + "angle": 0, + "content": "[2] Omri Avrahami, Thomas Hayes, Oran Gafni, Sonal Gupta, Yaniv Taigman, Devi Parikh, Dani Lischinski, Ohad Fried, and Xi Yin. Spatext: Spatio-textual representation for controllable image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18370-18380, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.259, + 0.483, + 0.327 + ], + "angle": 0, + "content": "[3] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Ji-aming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, et al. ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.33, + 0.483, + 0.399 + ], + "angle": 0, + "content": "[4] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.402, + 0.483, + 0.469 + ], + "angle": 0, + "content": "[5] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.473, + 0.482, + 0.514 + ], + "angle": 0, + "content": "[6] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.517, + 0.483, + 0.598 + ], + "angle": 0, + "content": "[7] Ganggui Ding, Canyu Zhao, Wen Wang, Zhen Yang, Zide Liu, Hao Chen, and Chunhua Shen. Freecustom: Tuning-free customized image generation for multi-concept composition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9089-9098, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.602, + 0.483, + 0.671 + ], + "angle": 0, + "content": "[8] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis, march 2024. URL http://arxiv.org/abs/2403.03206. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.673, + 0.483, + 0.741 + ], + "angle": 0, + "content": "[9] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.745, + 0.483, + 0.8 + ], + "angle": 0, + "content": "[10] Rinon Gal, Moab Arar, Yuval Atzmon, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Encoder-based domain tuning for fast personalization of text-to-image models. ACM Transactions on Graphics (TOG), 42(4):1-13, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.802, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[11] Zigang Geng, Ke Sun, Bin Xiao, Zhaoxiang Zhang, and Jingdong Wang. Bottom-up human pose estimation via disentangled keypoint regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14676-14686, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[12] Yuchao Gu, Xintao Wang, Jay Zhangjie Wu, Yujun Shi, Yunpeng Chen, Zihan Fan, Wuyou Xiao, Rui Zhao, Shuning" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "Chang, Weijia Wu, et al. Mix-of-show: Decentralized low-rank adaptation for multi-concept customization of diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.217 + ], + "angle": 0, + "content": "[13] Ligong Han, Yinxiao Li, Han Zhang, Peyman Milanfar, Dimitris Metaxas, and Feng Yang. Svdiff: Compact parameter space for diffusion fine-tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7323-7334, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.219, + 0.905, + 0.272 + ], + "angle": 0, + "content": "[14] Junjie He, Yifeng Geng, and Liefeng Bo. Uniportrait: A unified framework for identity-preserving single- and multi-human image personalization. arXiv preprint arXiv:2408.05939, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[15] Zecheng He, Bo Sun, Felix Juefei-Xu, Haoyu Ma, Ankit Ramchandani, Vincent Cheung, Siddharth Shah, Anmol Kalia, Harihar Subramanyam, Alireza Zareian, et al. Imagine yourself: Tuning-free personalized image generation. arXiv preprint arXiv:2409.13346, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.905, + 0.385 + ], + "angle": 0, + "content": "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.386, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[17] Chanran Kim, Jeongin Lee, Shichang Joung, Bongmo Kim, and Yeul-Min Baek. Instantfamily: Masked attention for zero-shot multi-id image generation. arXiv preprint arXiv:2404.19427, 2024. 1, 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.469 + ], + "angle": 0, + "content": "[18] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.905, + 0.525 + ], + "angle": 0, + "content": "[19] Zhe Kong, Yong Zhang, Tianyu Yang, Tao Wang, Kaihao Zhang, Bizhu Wu, Guanying Chen, Wei Liu, and Wenhan Luo. Omg: Occlusion-friendly personalized multi-concept generation in diffusion models. In ECCV, 2024. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.905, + 0.594 + ], + "angle": 0, + "content": "[20] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.596, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[21] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.665, + 0.905, + 0.734 + ], + "angle": 0, + "content": "[22] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22511-22521, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.735, + 0.905, + 0.803 + ], + "angle": 0, + "content": "[23] Zhen Li, Mingdeng Cao, Xintao Wang, Zhongang Qi, MingMing Cheng, and Ying Shan. Photomaker: Customizing realistic human photos via stacked id embedding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8640-8650, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.804, + 0.905, + 0.859 + ], + "angle": 0, + "content": "[24] Zhiheng Liu, Ruili Feng, Kai Zhu, Yifei Zhang, Kecheng Zheng, Yu Liu, Deli Zhao, Jingren Zhou, and Yang Cao. Cones: Concept neurons in diffusion models for customized generation. arXiv preprint arXiv:2303.05125, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[25] Nanye Ma, Mark Goldstein, Michael S Albergo, Nicholas M Boffi, Eric Vanden-Eijnden, and Saining Xie. Sit: Exploring flow and diffusion-based generative models with scalable" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2994" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "interpolant transformers. arXiv preprint arXiv:2401.08740, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.191 + ], + "angle": 0, + "content": "[26] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.483, + 0.248 + ], + "angle": 0, + "content": "[27] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.483, + 0.333 + ], + "angle": 0, + "content": "[28] Xu Peng, Junwei Zhu, Boyuan Jiang, Ying Tai, Donghao Luo, Jiangning Zhang, Wei Lin, Taisong Jin, Chengjie Wang, and Rongrong Ji. Portraitbooth: A versatile portrait model for fast identity-preserved personalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27080-27090, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.335, + 0.483, + 0.405 + ], + "angle": 0, + "content": "[29] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. In The Twelfth International Conference on Learning Representations, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.483, + 0.489 + ], + "angle": 0, + "content": "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.483, + 0.56 + ], + "angle": 0, + "content": "[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.483, + 0.644 + ], + "angle": 0, + "content": "[32] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.483, + 0.73 + ], + "angle": 0, + "content": "[33] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdreambooth: Hypernetworks for fast personalization of text-to-image models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6527-6536, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.483, + 0.814 + ], + "angle": 0, + "content": "[34] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.483, + 0.872 + ], + "angle": 0, + "content": "[35] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 815-823, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[36] Noam Shazeer and Mitchell Stern. Adafactor: Adaptive learning rates with sublinear memory cost. In International" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "Conference on Machine Learning, pages 4596-4604. PMLR, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[37] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[38] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.236, + 0.905, + 0.291 + ], + "angle": 0, + "content": "[39] Yang Sui, Huy Phan, Jinqi Xiao, Tianfang Zhang, Zijie Tang, Cong Shi, Yan Wang, Yingying Chen, and Bo Yuan. Disdet: Exploring detectability of backdoor attack on diffusion models. arXiv preprint arXiv:2402.02739, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.293, + 0.905, + 0.348 + ], + "angle": 0, + "content": "[40] Yoad Tewel, Rinon Gal, Gal Chechik, and Yuval Atzmon. Key-locked rank one editing for text-to-image personalization. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.35, + 0.905, + 0.405 + ], + "angle": 0, + "content": "[41] Hanhui Wang, Yihua Zhang, Ruizheng Bai, Yue Zhao, Sijia Liu, and Zhengzhong Tu. Edit away and my face will not stay: Personal biometric defense against malicious generative editing. arXiv preprint arXiv:2411.16832, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.407, + 0.905, + 0.462 + ], + "angle": 0, + "content": "[42] Kuan-Chieh Wang, Daniil Ostashev, Yuwei Fang, Sergey Tulyakov, and Kfir Aberman. Moa: Mixture-of-attention for subject-context disentanglement in personalized image generation. arXiv preprint arXiv:2404.11565, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.464, + 0.905, + 0.518 + ], + "angle": 0, + "content": "[43] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, and Anthony Chen. Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519, 2024. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.521, + 0.905, + 0.589 + ], + "angle": 0, + "content": "[44] Xudong Wang, Trevor Darrell, Sai Saketh Rambhatla, Rohit Girdhar, and Ishan Misra. Instancediffusion: Instance-level control for image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6232-6242, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.591, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[45] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15943-15953, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.676, + 0.905, + 0.744 + ], + "angle": 0, + "content": "[46] Yuxin Wen, Neel Jain, John Kirchenbauer, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Hard prompts made easy: Gradient-based discrete optimization for prompt tuning and discovery. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[47] Guangxuan Xiao, Tianwei Yin, William T Freeman, Frédo Durand, and Song Han. Fastcomposer: Tuning-free multisubject image generation with localized attention. International Journal of Computer Vision, pages 1-20, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[48] Jinqi Xiao, Miao Yin, Yu Gong, Xiao Zang, Jian Ren, and Bo Yuan. COMCAT: Towards efficient compression and customization of attention-based vision models. In Proceedings of the 40th International Conference on Machine Learning, pages 38125-38136. PMLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[49] Jinqi Xiao, Shen Sang, Tiancheng Zhi, Jing Liu, Qing Yan, Linjie Luo, and Bo Yuan. Coap: Memory-efficient training" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.957 + ], + "angle": 0, + "content": "2995" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.484, + 0.12 + ], + "angle": 0, + "content": "with correlation-aware gradient projection. arXiv preprint arXiv:2412.00071, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.176 + ], + "angle": 0, + "content": "[50] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.483, + 0.233 + ], + "angle": 0, + "content": "[51] Ge Yuan, Xiaodong Cun, Yong Zhang, Maomao Li, Chenyang Qi, Xintao Wang, Ying Shan, and Huicheng Zheng. Inserting anybody in diffusion models via celeb basis. arXiv preprint arXiv:2306.00926, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.483, + 0.29 + ], + "angle": 0, + "content": "[52] Kaipeng Zhang, Zhanpeng Zhang, Zhifeng Li, and Yu Qiao. Joint face detection and alignment using multitask cascaded convolutional networks. IEEE signal processing letters, 23 (10):1499-1503, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.292, + 0.483, + 0.347 + ], + "angle": 0, + "content": "[53] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.483, + 0.416 + ], + "angle": 0, + "content": "[54] Yimeng Zhang, Xin Chen, Jinghan Jia, Yihua Zhang, Chongyu Fan, Jiancheng Liu, Mingyi Hong, Ke Ding, and Sijia Liu. Defensive unlearning with adversarial training for robust concept erasure in diffusion models. arXiv preprint arXiv:2405.15234, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.419, + 0.483, + 0.5 + ], + "angle": 0, + "content": "[55] Yimeng Zhang, Jinghan Jia, Xin Chen, Aochuan Chen, Yihua Zhang, Jiancheng Liu, Ke Ding, and Sijia Liu. To generate or not? safety-driven unlearned diffusion models are still easy to generate unsafe images... for now. In European Conference on Computer Vision, pages 385-403. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.503, + 0.483, + 0.571 + ], + "angle": 0, + "content": "[56] Yihua Zhang, Yimeng Zhang, Yuguang Yao, Jinghan Jia, Jiancheng Liu, Xiaoming Liu, and Sijia Liu. Unlearncanvas: A stylized image dataset to benchmark machine unlearning for diffusion models. arXiv preprint arXiv:2402.11846, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.574, + 0.483, + 0.628 + ], + "angle": 0, + "content": "[57] Yufan Zhou, Ruiyi Zhang, Tong Sun, and Jinhui Xu. Enhancing detail preservation for customized text-to-image generation: A regularization-free approach. arXiv preprint arXiv:2305.13579, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.628 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "2996" + } + ] +] \ No newline at end of file diff --git a/2025/ID-Patch_ Robust ID Association for Group Photo Personalization/a37d3ef3-36c1-46f2-886c-58bf6589bcc4_origin.pdf b/2025/ID-Patch_ Robust ID Association for Group Photo Personalization/a37d3ef3-36c1-46f2-886c-58bf6589bcc4_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8328bc80341853deea5b50210423e73e0108d996 --- /dev/null +++ b/2025/ID-Patch_ Robust ID Association for Group Photo Personalization/a37d3ef3-36c1-46f2-886c-58bf6589bcc4_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:935abb3b4d3475ca732596d5884bff11039baef96d06eeb9a6fda265d8fb5674 +size 5233404 diff --git a/2025/ID-Patch_ Robust ID Association for Group Photo Personalization/full.md b/2025/ID-Patch_ Robust ID Association for Group Photo Personalization/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0f56f9d17ed7a7f8b286844aa64181bbd2c85ae4 --- /dev/null +++ b/2025/ID-Patch_ Robust ID Association for Group Photo Personalization/full.md @@ -0,0 +1,344 @@ +# ID-Patch: Robust ID Association for Group Photo Personalization + +Yimeng Zhang $^{1,2,*}$ , Tiancheng Zhi $^{1}$ , Jing Liu $^{1}$ , Shen Sang $^{1}$ , Liming Jiang $^{1}$ , Qing Yan $^{1}$ , Sijia Liu $^{2}$ , Linjie Luo $^{1}$ + +1 ByteDance Inc., 2 Michigan State University + +![](images/d64583b412951698b75d34769dd212fc91c0de9fd0546b8d0ec9f60eba4a078b.jpg) +Pose Condition & ID Location +"An indoor celebration scene featuring three women..." +Figure 1. Comparison with state-of-the-art multi-identity generation methods. From left to right: the condition inputs followed by results generated using OMG [19] (with InstantID [43]), InstantFamily [17], and our proposed ID-Patch approach. Red dashed boxes highlight failures. OMG fails to preserve the hairstyles of the middle person, and creates artifacts for the right woman's hand, possibly because of the inconsistency between its generation stages (explained in Fig. 2). InstantFamily suffers from ID leakage, resulting the incorrect ID of the middle person. Our approach preserves the detailed identity of each person. In addition, our approach is 7 times faster than OMG and has less computational overhead than InstantFamily. + +![](images/4debe4e158bfe18c9c369aab5f88a45b9bba5969529a1cdcd99797b26ad68166.jpg) +OMG + InstantID +Generation Time: 71.1s + +![](images/e7448aa925b865f1f7ed44fbfdf753faee8352c4cb3cf97b50cd71f20f655d06.jpg) +InstantFamily +Generation Time: 9.6s + +![](images/5d9a202bf8b9bc0524d4461d6ae840f929a6599916cadd4a0adb0673632e912b.jpg) +ID-Patch (Ours) +Generation Time: 9.4s + +# Abstract + +The ability to synthesize personalized group photos and specify the positions of each identity offers immense creative potential. While such imagery can be visually appealing, it presents significant challenges for existing technologies. A persistent issue is identity (ID) leakage, where injected facial features interfere with one another, resulting in low face resemblance, incorrect positioning, and visual artifacts. Existing methods suffer from limitations such as the reliance on segmentation models, increased runtime, or a high probability of ID leakage. To address these challenges, we propose ID-Patch, a novel method that provides robust association between identities and 2D positions. Our approach generates an ID patch and ID embeddings from the same facial features: the ID patch is positioned on the conditional image for precise spatial control, while the ID embeddings integrate with text embeddings to ensure high resemblance. Experimental results demon + +strate that ID-Patch surpasses baseline methods across metrics, such as face ID resemblance, ID-position association accuracy, and generation efficiency. Project Page is: https://byteaigc.github.io/ID-Patch/ + +# 1. Introduction + +Personalized group photo generation has wide-ranging applications in entertainment, advertising, social media, and virtual reality. For example, users might want to create a group photo of friends at a virtual party, even if some could not attend in person. This technology enables the creation of engaging and memorable content that reflects real-life social interactions and relationships. However, this task poses significant challenges, particularly the risk of identity (ID) leakage, where the identity information of one individual inadvertently affects the representation of another at a different position. Such blending compromises the distinctiveness and accuracy of each person's features. Besides, being able to specify the position of each identity in the image is + +an important function when users want to create a specific layout of people. Incorrect positioning of identities may lead to undesired results. + +State-of-the-art group photo personalization methods largely rely on diffusion models [3, 6, 16, 26, 37-39, 41, 48, 49, 54, 56], which deliver superior visual quality when paired with effective prompt engineering [4, 8, 25, 27, 34, 55]. To address the issue of ID leakage, OMG [19] first generates an image without ID injection. Then, it estimates segmentation masks for the individuals in the image and subsequently injects the IDs into each segmented region. However, it depends on a segmentation model, which might fail in challenging cases. Besides, the content generated in the two stages may conflict with each other, leading to reduced visual fidelity. See Fig.1 and Fig.2 for example. Moreover, OMG requires running the denoising process separately for each identity, resulting in a runtime that increases linearly with the number of people. Alternatively, InstantFamily [17] employs a single-pass generation strategy by manipulating cross-attention masks. This approach uses approximate head region masks for each ID to direct pixel attention exclusively to specific ID embeddings derived from facial features. However, it struggles to prevent ID leakage due to two key issues: (1) imprecise masks or close proximity of faces, which can cause overlaps, and (2) unintended information propagation through self-attention and convolutional layers. Please see Fig.1 and Fig.3 for more details. + +In response to these limitations, we propose a novel method named ID-Patch, designed for efficient and robust ID-position association in group photo generation. Our idea is to let the model learn to associate input IDs with their spatially designated locations. Specifically, we project facial features onto a small RGB image patch called ID patch and token embeddings called ID embeddings. The ID patches are placed onto the conditioning image of a ControlNet [53] according to face locations to ensure precise position control of IDs within the generated images. ID embeddings are integrated with text embeddings to enhance facial details. Our ID-Patch method seamlessly integrates with various types of spatial conditions, such as poses, canny edges or depth map, enhancing the robustness and flexibility of our method. + +We summarize our key contributions as follows: + +- ID-Patch Method: We introduce a novel approach for multi-ID image generation that directly links face ID features with their locations using visual patches, ensuring accurate resemblance and position control. +- Efficiency Improvement: Our method enhances the efficiency of multi-ID image generation by simplifying the process to adding ID patches and concatenating ID embeddings, reducing computational overhead. +- Simplified Control: We eliminate the need for auxiliary + +![](images/9a9242358b947183b0116abda6bcf658424be76bc062b0f0fb1fa029e1eb432d.jpg) +(a) OMG first stage + +![](images/0b75236d224893c93a9df368cdade44119f60df4de1c4fee8794b9f6b5564b24.jpg) +(b) Human mask + +![](images/be8c39d8f2fd7623878a30f02d4c55d21f4f345d4886d704ecd42672f0b72194.jpg) +(c) OMG final result + +![](images/f426f777d41f1f522a8d88580de0c331e5c28eeac4cfb4d0503105746da9ae63.jpg) +Figure 2. Why does OMG fail in Fig. 1? OMG relies on a first stage result (no ID awareness) and its human segmentation to control layout. The second stage paints injected IDs within corresponding masks. However, the second pass may conflict with the first stage semantics, leading to deteriorated ID resemblance (long hair in the middle of image (c)) or artifacts (hand in the middle). +(a) ID +Figure 3. ID leakage of InstantFamily [17]. The text prompt describes a man and a woman standing together, and ID (a) is only injected into the left man. + +![](images/6cf9aead00171fe6526fd123fe4ae035538091c65b13c96ce1c9a3b02e21265c.jpg) +(b) InstantFamily result + +![](images/84e3aa550abef8c2ee884f0ddd08a91d769d292c9790914f313689485c913f51.jpg) +(c) ID-Patch result + +segmentation models, requiring only a single point for ID position control, thereby increasing robustness and reducing complexity. + +- Superior Performance: Our approach achieves higher fidelity in ID resemblance and better position control, especially in complex images with multiple identities, as demonstrated by our experimental results. + +# 2. Related Work + +Single-ID Personalization. The personalization of diffusion models has emerged as a popular research topic. Finetuning or optimization based methods [9, 10, 32, 33, 46] have the potential to reach high quality results in terms of ID resemblance, but are time-consuming. Inference-only methods [15, 23, 28, 43, 45, 50, 51, 57] eliminate the finetuning process usually by encoding the concept into token embeddings and injecting them into the diffusion models via cross-attention. However, most of these works focus on single concept injection rather than group photo generation. Multi-ID Personalization. The main challenge of personalized group photo generation is ID blending, i.e., the identity information of person A may affect the generation of person B at another position. To resolve this issue, OMG [19] performs separate denoising processes. First, a noncustomized image is generated and masks of each person are extracted. Second, masked regions are painted using + +![](images/3424c2213de91929ba6e557856107180c698ce40c1fb3ff61960ff2edb0bddc0.jpg) +Figure 4. Our pipeline. Given a text prompt (e.g., two people shaking hands), $N$ face images and locations, we generate an image with $N$ IDs. We extract face features for each ID and project them into ID patches and ID embeddings. ID patches are rendered on a black canvas (or added on top of a pose image) according to face locations and sent into a ControlNet to control the positions of generated faces. ID embeddings are appended to text embeddings to provide detailed face information to the diffusion model and ControlNet via cross-attention. + +separate customized models and stitched together. This approach is slow and conflicts between stages may cause identity loss. Additionally, it relies on a segmentation module, which may fail in challenging cases. To avoid running separate customized models, recent works [14, 42, 47] let the network to learn to attend to specific concept tokens by supervising the attention mask using segmentation or ID routing. However, these methods do not allow the specification of the accurate locations of each ID. Although they can be combined with pose conditions, describing the assignment of IDs to positions in text may be hard. InstantFamily [17] provides an opportunity to specify the approximate location of each ID by manipulating the attention masks during both training and inference. However, no constraints have been imposed to avoid the ID leakage through self-attention and convolutional layers, leading to ID blending, omitting or swapping. + +Multiple General Concepts. Besides human generation, there are series of work focusing on the customization of multiple general concepts [1, 2, 7, 12, 13, 20, 24, 40, 47] and grounded image generation [2, 22, 44] associating objects with positions. GLIGEN [22] integrates grounding tokens, which align textual prompts with specific image regions. SpaText [2] and InstanceDiffusion [44] allows users to define the placement and attributes of objects. However, these concepts are usually semantically distinct while it is hard to describe the human identities in text and subtle details matter. + +# 3. Method + +In this section, we first outline some preliminaries about text-to-image diffusion models and ControlNets. Subsequently, we detail the problem formulation for multi-ID image generation and the design of our modules, including the ID patch and ID embeddings, as well as the training and + +inference schemes. + +# 3.1. Preliminaries + +Latent diffusion models (LDMs). LDMs [31] are image generation models that can incorporate conditioning signals into the generation process. Specifically, during training, an encoder $\mathcal{E}$ of variational autoencoder (VAE) [18] encodes an image $x$ into a latent space, and a neural network $\epsilon_{\theta}(z_t,t,c)$ performs the denoising process within this latent space. It attempts to predict the original Gaussian noise $\epsilon$ from a noisy latent variable $z_{t}$ , where $t$ represents the diffusion timestep, and $c$ is the conditioning signal, injected via cross-attention. The loss function for an LDM is defined as: + +$$ +L _ {\mathrm {L D M}} = \mathbb {E} _ {\mathcal {E} (x), \epsilon \sim \mathcal {N} (0, 1), t} \left[ \| \epsilon - \epsilon_ {\theta} (z _ {t}, t, c) \| _ {2} ^ {2} \right] \quad (1) +$$ + +ControlNet. The incorporation of ControlNet [53] into diffusion models marks a significant advancement in controlled image synthesis. ControlNet is designed to steer the image generation process by applying spatial conditioning signals such as rasterized pose images, depth maps, or canny edges. It introduces a control network that processes these signals and adds their features to the base model's intermediate results. The control network is a trainable copy of the base model's encoder layers and is initialized with zero convolution. This effective integration achieves high quality controlled generation. + +# 3.2. Multi-ID Image Generation + +This paper focuses on efficiently and effectively addressing the problem of multi-ID image generation utilizing conditioning signals that include text prompts, identity features, and face locations. The inputs to our model comprise text prompt embeddings $c_{t}$ , facial features $c_{f} = (f_{0}, f_{1}, \ldots, f_{N-1})$ for $N$ distinct identities, and their corresponding facial locations $c_{l} = (l_{0}, l_{1}, \ldots, l_{N-1})$ . The + +![](images/75b1ba4caaae833986a82dbf80ce97f0c559b36ea6c6fba660732df60e0771ea.jpg) +(a) ID + +![](images/284eb7c683953fa9f26211122794f9c8e5276c1c5ebd9bf835d58f377fa2861e.jpg) +(b) w/o ID embeddings + +![](images/fa0e881be04ff7854c1192da7234ec1832dc71e9eb4c267f6238466226442436.jpg) +(c) with ID embeddings +Figure 5. Effectiveness of ID embedding. Without ID embeddings, one can distinguish between the two people but the resemblance is low. Incorporating ID embeddings significantly improves face resemblance. + +facial location $l_{i} = (x_{i},y_{i})$ identifies the xy-coordinates of the nose tip for each identity within the intended image. The goal is to synthesize an image following the text description with each given face located in the designated place. + +# 3.2.1 ID Patches and ID Embeddings + +As discussed in Sec. 3.1, ControlNet effectively integrates spatial conditional signals into diffusion models. Building on this capability, we introduce a method to encode ID location information using condition images, subsequently processed by ControlNet to guide ID localization within the generated image. Our approach features ID conditions manifesting as both patches and embeddings. Specifically, ID patches, precisely aligned with facial locations, are integrated via ControlNet to ensure accurate identity localization, while ID embeddings are utilized via cross-attention mechanisms in the LDM to enhance the detailed representation of facial features, as illustrated in Fig. 4. + +Both ID patches and ID embeddings are derived from facial features, enabling a synergistic association between them, with both projection processes facilitated by the Perceiver Resampler [50]. For ID patches, we convert the facial feature $f_{i}$ into an identity-specific patch $p_i \in \mathbb{R}^{P \times P \times 3}$ , functioning similarly to a QR code, uniquely identifying each person and easily distinguishable from others. These ID patches are placed on a black canvas at specified facial locations to create an image $I$ , which then serves as the conditioning input for ControlNet, thereby providing precise identity localization information. + +As depicted in Fig. 5, while ID patches alone are generally sufficient for distinguishing between identities, they may not always achieve optimal results in generating images with high ID resemblance. To address this limitation, and inspired by InstantFamily [17], we also incorporate ID embeddings. These embeddings are designed to inject detailed facial appearance information into the dif + +![](images/27e1ad580f5f8ddd1fa4a85fe3db5e48cf59290f0fe1c8cb12dcfd11fed7c274.jpg) +(a) ID (b) Result w/o pose +Figure 6. ID-Patch combined with pose conditions. Provided with user ID images in (a), our method can generate results with only the specifications of nose tip locations as shown in (b). Merging ID patches with pose images (c) enhances spatial control over the generated results as seen in (d), without incurring any computational overhead. + +![](images/14e7f4b15a207bd8700592e1a9dcfd309dda13dbd6b064d9b06571c572985fcc.jpg) +(c) ID patch on pose + +![](images/1103147907d956261f40fa38d51ab9e346cabc62a20c10fb31bd794471f6772b.jpg) +(d) Result with pose + +fusion model, thus enhancing the resemblance of the generated identities to their real-life counterparts. Specifically, ID embeddings $w_{i} \in \mathbb{R}^{2048 \times M}$ , share the same token dimension 2048 as the SDXL text embeddings. As shown in Fig.4, these embeddings are appended to text embeddings $c_{t}$ to form an extended text embedding $c_{t}' = [c_{t}, w_{0}, w_{1}, \ldots, w_{N-1}]$ , which is then fed into both the UNet and ControlNet through cross-attention mechanisms. + +Since human pose conditions are often utilized to enhance group photo generation, we can combine our ID-Patch approach with the pose conditions for multi-id image generation. As illustrated in Fig. 6, the ID patches can be overlaid onto a pose-conditioned image, creating a composite condition image that encodes both identity positioning and pose information. It is important to mention that while our method can incorporate pose information, it does not inherently require it; pose is just an optional enhancement, one can also choose other conditions like canny edge, depth map for specific applications. The adaptability of our ID-Patch approach presents enhanced control over generated results yet without incurring the extra overhead of an additional ControlNet branch. + +# 3.2.2 Two-Stage Training + +The training of our model employs a conventional diffusion loss function, defined as follows: + +$$ +L = \mathbb {E} _ {\mathcal {E} (x), \epsilon \sim \mathcal {N} (0, 1), t} [ \| \epsilon - \epsilon_ {\theta} \left(z _ {t}, t, I, c _ {t} ^ {\prime}\right) \| _ {2} ^ {2} ] \tag {2} +$$ + +Initially, one might consider training the models end-to-end from scratch. However, our experiments reveal that such a method does not consistently ensure accurate ID localization, as depicted in Fig.7. This shortcoming likely stems from the network's tendency to preferentially inject ID information into ID embeddings, rather than ID patches, resulting in non-distinctive ID patches that could potentially + +![](images/0e013eb6c9e2891b58771e3ad72e4f8114c525c700a6f443fb02a806d6509c4e.jpg) +(a) Pose + ID + +![](images/70787222617e083dc453dc28bdd54cbe210743c2d5ee1e84e96bd1db19bb09ba.jpg) +(b) Single-stage train + +![](images/88bc0f3cafe9c6baa7b0459f2bcee5bb4a63c1b874aa48f6c8c54baecc10e238.jpg) +(c) Two-stage train +Figure 7. Two-stage training to improve positioning robustness. Given pose and ID conditions in (a), single-stage training cannot fully prevent the incorrect face positioning issue. For example, in (b), the man is mistakenly placed at the central bottom position, generating an inharmonious result. (c) Two-stage training is introduced to solve this problem. As can be seen on the top row of this figure, the ID patches from our two-stage training approach are visually more distinguishable from each other, compared to the ones from single-stage training. Our experimental results prove that this is critical in solving the ID leak issues. + +confuse ControlNet. In the single-stage model, as shown in Fig. 7, the ID patches exhibit only subtle differences between different individuals. To resolve this, we introduce a two-stage training approach. In the initial stage, ID embeddings are omitted, compelling the network to prioritize the incorporation of ID information into ID patches, thereby enhancing their distinctiveness. In the subsequent stage, ID embeddings are introduced and the model is fine-tuned to further improve the resemblance of the generated images. + +# 4. Experiments + +# 4.1. Settings + +Implementation details. Facial features $f_{i}$ we use are 512-dimensional vectors derived from ArcFace [5]. During training, we utilize the SDXL diffusion model [29]. During the inference phase, we employ the further finetuned SDXL, namely Juggernaut-X-v10*, as base model to enhance portrait quality. The inference process involves 50 timesteps. We set the ID token injection ratio to 0.8 to regulate the influence of identity features, where the initial $20\%$ of timesteps operate without ID token embeddings to focus on the general style, followed by the remaining $80\%$ timesteps which incorporate ID token embeddings to refine the identity-specific aspects of the images. The default size of ID-patch is $P \times P = 64 \times 64$ and the default length of ID embeddings is $M = 16$ . The network is trained using 8 NVIDIA H100 GPUs with batch size=4, learning rate=1e-05, weight decay = 0.01, using AdaFactor optimizer [36]. + +Table 1. Performance comparison across four dimensions: identity resemblance, ID-position association accuracy, text alignment, and generation time. For qualitative metrics to evaluate similarity or accuracy, higher is better. Our approach significantly outperforms baseline methods in terms of resemblance, association, and generation time, with similar text alignment scores. + +
MethodsID ↑Association ↑Text ↑Time (s) ↓
OMG + InstantID0.6440.9260.27184.83
InstantFamily0.6830.8620.27810.02
ID-Patch (Ours)0.7510.9580.2739.69
+ +We train it for 720k iterations for stage 1, and stage 2 usually converges around 1 million iterations. We select the best model based on a 20 people validation set. + +Dataset. Our training set consists of 17 million single-person and 1.95 million multi-person images. These images are obtained from purchased data and publicly available sources. For evaluation, we utilize two specialized datasets: 1) Style Dataset comprises 40 styles with pose and text prompt definitions, ranging from one to eight faces. 2) Identity Dataset includes 600 unique identities representing a variety of races and genders. Each style is paired with 20 randomly sampled unique identity combinations for quantitative evaluation. + +Baselines. Our analysis includes two state-of-the-art multi-ID generation methods as baselines: 1) OMG [19]: This method incorporates multiple identity features into separate segmented regions through latent stitching. It can be paired with a single ID injection method InstantID [43] for multi-ID generation. 2) InstantFamily [17]: This approach employs coarse attention masks within cross-attention layers, a mandatory process that ensures distinct identity features are injected into different image instances effectively. Because there is no official publicly available code for InstantFamily [17], for fairness, we implemented it based on the SDXL base model [29], and trained it using the same training data as our model. + +Evaluation Metrics. We evaluate the methods in a pose-conditioned setting, because pose is a requirement of InstantFamily, and without pose condition, OMG is not able to generate faces at specified locations for its first stage. Besides, pose condition makes the head size consistent, which is easier for resemblance evaluation and reduces the impact of head sizes on generation quality. For comprehensive evaluation, we consider four critical dimensions: 1) Identity resemblance: we measure the identity resemblance by calculating the cosine similarity (CosSim) between the facial features of faces in the generated images and the reference faces. For fairness, we use FaceNet [35], a face model different from the ArcFace model used for training. 2) Association accuracy: we assess the accuracy of identity-position association in generated images by ensuring that the generated faces are most similar to their assigned coun + +![](images/6ad51e41ac9fd5e146601be20ebe6ee7d216be296facaeac83003acf660a81dd.jpg) +(a) Identity Resemblance + +![](images/74ba8deec9c02f2b28235561317fc0094418a060acfd19f6c5fd07d65792ebcb.jpg) +(b) Association Accuracy + +![](images/18f05bfa9f534f0cd24dff0c791aa88a5fab82002624d88200566a7f57a1db94.jpg) +(c) Generation Time + +![](images/f6b54bb931d500df58634fba0dee751e2cb819c2a805d611f8d98d7b7c95847c.jpg) +Figure 8. Performance evaluation of different model generations from three different aspects: identity resemblance, location accuracy, and generation time. As face number increases, the ID and location metric scores of OMG and Instantfamily drop much more significantly than our approach. Our approach also achieves near-constant generation time, in contrast to OMG's linearly increasing running time. +Figure 9. Comparison with baselines on pose-conditioned generation, where red dashed boxes highlight instances with low identity resemblance. In row 1, OMG fails to preserve the face shape of the third woman (from left to right) because its first stage result conflicts with this ID's face shape. InstantFamily incorrectly generates the second person because of the ID leakage from the third woman. In row 2, OMG does not generate correct hair styles and accurate facial features for the two people in the red boxes, while InstantFamily generates the wrong ID in the red box. + +terparts, compared to all other input faces used in the generation process. In other words, we calculate the similarity between the $i$ -th generated face and the $j$ -th input face. Let $s(i) = \arg \max_{0 \leq j \leq N-1} \mathrm{CosSim}(\tilde{f}_i^{\mathrm{gen}}, \tilde{f}_j)$ be the face index among all input faces that is most similar to the $i$ -th generated face, where $\tilde{f}_i^{\mathrm{gen}}$ , $\tilde{f}_j$ denote the FaceNet features of the $i$ -th generated face and the $j$ -th input face respec + +tively, and $N$ is the number of faces in this image. When $s(i) = i$ , the $i$ -th generated face has the correct association. The identity-position association accuracy is defined as: + +$$ +\frac {1}{N} \sum_ {i = 0} ^ {N - 1} \mathbf {1} \{i = s (i) \} \tag {3} +$$ + +This measurement assesses the correct placement of each + +Table 2. Ablation study results for our proposed ID-Patch variants, designed to assess the impact of different modules (patch projection and token projection) and training/inferencing schemes on the performance of multi-ID image generation. The full method achieves the best ID resemblance and ID-position association, indicating the necessity of each module. + +
VariantsID ↑Association ↑Text ↑
No patch projection0.3680.3450.271
No token projection0.5650.8950.284
Single stage training0.6600.7840.271
ID token injection ratio = 1.00.7500.9570.270
FULL0.7510.9580.273
+ +identity within the image. 3) Text alignment: this metric evaluates the alignment between the textual context of the image generation prompt and the visual output. Cosine similarity is used to compare textual features extracted by the CLIP text encoder with visual features extracted by the CLIP image encoder from the generated images [30]. 4) Generation time: we measure the efficiency by recording the time (in seconds) required to generate an image on a NVIDIA A100 GPU, excluding the time taken for model loading and image I/O. + +# 4.2. Comparison with Baselines + +As listed in Tab. 1, ID-Patch excels in identity and association metrics with scores of 0.751 and 0.958, respectively, showcasing its robust capability to maintain identity consistency and achieve precise placement in generated images. There is no significant difference in text alignment scores between compared methods. Notably, ID-Patch achieves the fastest image generation time. + +Further details on the performance with varying face numbers are shown in Fig. 8. As we observed no deterioration in text alignment with an increase in the number of faces across all methods, we have relegated the corresponding figure to the appendix. As the number of faces increases, both OMG and InstantFamily experience significant performance drops due to ID leakage. Although ID-Patch also shows a decline, it is less severe, mainly because the increase in face numbers reduces the average face area, adversely affecting the quality of smaller faces produced by the SDXL models. Moreover, the presence of more faces slightly extends the image generation time due to the increased demand for facial feature extraction. InstantFamily is slightly slower than our approach due to the overhead of masked attention. For OMG, injecting identity features into designated segmented areas causes a substantial increase in generation time; it takes about 2 minutes to generate an image with eight people, whereas ID-Patch requires only about 10 seconds. See Fig. 9 for visual comparisons. + +# 4.3. Ablation Study + +We conduct an ablation study on variants of the proposed method in Tab. 2. The performance notably degrades when patch projection is omitted, leading to poor location alignment as the algorithm attempts to infer ID locations based on indirect cues such as gender and pose. Similarly, omitting token projection reduces resemblance by eliminating essential facial details, quantitatively validates the finding of Fig. 5. Single-stage training results in indistinguishable ID patches (as shown in Fig. 7), which impairs effective positioning. The default setting with an ID token injection ratio of 0.8 improves text alignment compared to injecting the ID token at every timestep, while maintaining the same level of identity resemblance. This improvement may arise because ID embeddings have little effect on facial detail during the early denoising steps, but become more influential in the later stages. It is important to note that even without the inference trick (i.e., using the ID token injection ratio), our proposed method still outperforms all baseline approaches. + +# 4.4. Pose-Free Generation + +Our ID-Patch model can be operated effectively without the need for pose conditioning, a flexibility that permits its integration with other conditions for image generation. This capability enables the use of ID-Patch in versatile configurations, whether plug-and-play or co-trained, which we will explore in subsequent subsections. + +In the pose-free mode, the ControlNet uses a conditioning image that solely features ID patches against a black background. The placement of each ID patch at the tip of the nose ensures accurate preservation of facial locations during image synthesis, while head sizes and poses are variably generated based on the initial noise. As illustrated in Fig. 10, this methodology allows our model to produce aesthetically pleasing group photographs of different number of individuals across a variety of settings, demonstrating the robustness and adaptability of our approach. + +The pose-free version of ID-Patch is designed to be compatible with other ControlNets (plug-and-play). As illustrated in Fig. 11, it is possible to combine our ID-Patch ControlNet the Canny edge ControlNet to achieve more precise control over the final image outcomes. + +# 4.5. Limitations + +As illustrated in Fig. 12, our method encounters two primary limitations. First, our implementation is based on the SDXL model, which occasionally produces inaccuracies in body and finger anatomy. Transitioning to a more advanced base model could potentially rectify this issue. Second, facial features in the current model may incorporate lighting and expression details that are not fully separated from identity information, potentially compromising the generation quality. Enriching our training dataset with multiple images + +![](images/6b74927a748a75fdf6b6a21127cb53612200650f4d5b1d8368a9a6611aea5a70.jpg) +ID + ID Location + +![](images/2feb6fae2b3fde4fb4d6e8a59c347a70428465b0c0481a2f81914a92975d2511.jpg) + +![](images/1bea2d862a140506b34a33356db25ba2e001fa0030e325abcdd9ca538404a9b8.jpg) + +![](images/4c1b876c595d0ca45c26eeda1e652aff7f59c209fb9303ca88b051fcf3ac462e.jpg) + +![](images/c26e393fd68cc9ae27e522abbf4ca149cddc532b02d8e1e77f18e6f35243cc2b.jpg) +Generation Result + +![](images/091862fcd47e0f47c4dcb8fa0a785ed6d57b1120e3555066af192d3961fcab68.jpg) + +![](images/6780fe24621dea2c6ee8c9c2a50f09f13b1dcfca30df6b009578bee328a10814.jpg) + +![](images/4145dbd7bf16d7cf16d17a23a0866652151c813191bccb66851df10f9366ebb6.jpg) + +![](images/c2bf392d9d0cb766ad432ff907109b8c37d19eed8d35bf1aa6a2d8b5bde139e6.jpg) + +![](images/bd194ab43cba5bbb98d920cf533990037813ed35980ee480f6dc40d93d8e0379.jpg) +(a) ID + +![](images/67b880b750a22302c775ef7630ce3d6dc21b52647a5f9ee8fa8bd549b14c1acf.jpg) +Figure 10. ID-Patch pose-free generation. Here the condition image is only ID patches rendered onto a black canvas without any pose condition. The generation is conditioned on individual face locations while corresponding head sizes and body poses are inferred implicitly. Our method can accommodate a large number of people with diverse ethnic backgrounds, yet generating visually appealing group photo results. See "Evaluation Metrics" for reasons why we compare with baseline approaches only in pose-conditioned settings. +(b) Canny edge condition +Figure 11. Combination of pose-free ID-Patch ControlNet and pretrained Canny edge ControlNet. We can generate background and clothing details, with high quality ID injection. + +![](images/cd76eb79b1d113e2e0821f7eba9c42e75589c1f7cb797af3760b44ff870d46e4.jpg) +(c) ID-Patch result + +of the same identities, showcasing varied expressions and lighting conditions, might help alleviate this problem. + +Although the resemblance has been improved with our approach, there still exists a performance gap between generating a single person and multiple people (Fig. 8(a)). This means generating personalized group photo remains a challenging problem and requires further study. + +# 5. Conclusion + +In conclusion, our novel method, ID-Patch, markedly enhances identity resemblance and positioning in group photo generation. By embedding each identity feature within a distinct patch and utilizing ControlNet to accurately place + +![](images/43b1cc7aa18c5f64ece6e1b58d165bd95396cd810575f3106d46a0c427374afa.jpg) +(a) Abnormal anatomy + +![](images/9a097e59cbff5246aca59814e4a35c6d2ef07eac98f6733396221a3ec0728a30.jpg) +(b) Over-fitting input +Figure 12. Limitations. (a) Abnormal anatomy can be found in complex scenes, e.g. the forearms in the red box. (b) When input faces contain strong shadows, our method struggles to fully disentangle user's ID in generated result. + +![](images/5a5123a3951aa4aded6e60d8afc6fc00f799a2dfa8faf1931d51b1ba1520fea0.jpg) + +each identity at the designated spatial locations, we successfully reduce ID leakage. Our approach seamlessly integrates with additional conditioning signals such as pose control. Experimental results underscore the superiority of ID-Patch, particularly in complex scenarios involving more than three identities. This work paves the way for future explorations in multi-ID image generation. Potential future research directions include leveraging multiple images of the same individual from different angles to further enhance identity resemblance and simultaneously controlling location and facial expressions using the patch technique. + +Societal Impact: Our work studies ID injection from a technical perspective and is not intended for malicious use. However, inappropriate usage might generate undesired fake images. Users should be aware of the ethical implications and use it responsibly. + +# References + +[1] Omri Avrahami, Kfir Aberman, Ohad Fried, Daniel Cohen-Or, and Dani Lischinski. Break-a-scene: Extracting multiple concepts from a single image. In SIGGRAPH Asia 2023 Conference Papers, pages 1–12, 2023. 3 +[2] Omri Avrahami, Thomas Hayes, Oran Gafni, Sonal Gupta, Yaniv Taigman, Devi Parikh, Dani Lischinski, Ohad Fried, and Xi Yin. Spatext: Spatio-textual representation for controllable image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18370-18380, 2023. 3 +[3] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Ji-aming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, et al. ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 2 +[4] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 2 +[5] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 5 +[6] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2 +[7] Ganggui Ding, Canyu Zhao, Wen Wang, Zhen Yang, Zide Liu, Hao Chen, and Chunhua Shen. Freecustom: Tuning-free customized image generation for multi-concept composition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9089-9098, 2024. 3 +[8] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis, march 2024. URL http://arxiv.org/abs/2403.03206. 2 +[9] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2 +[10] Rinon Gal, Moab Arar, Yuval Atzmon, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Encoder-based domain tuning for fast personalization of text-to-image models. ACM Transactions on Graphics (TOG), 42(4):1-13, 2023. 2 +[11] Zigang Geng, Ke Sun, Bin Xiao, Zhaoxiang Zhang, and Jingdong Wang. Bottom-up human pose estimation via disentangled keypoint regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14676-14686, 2021. 1 +[12] Yuchao Gu, Xintao Wang, Jay Zhangjie Wu, Yujun Shi, Yunpeng Chen, Zihan Fan, Wuyou Xiao, Rui Zhao, Shuning + +Chang, Weijia Wu, et al. Mix-of-show: Decentralized low-rank adaptation for multi-concept customization of diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 3 +[13] Ligong Han, Yinxiao Li, Han Zhang, Peyman Milanfar, Dimitris Metaxas, and Feng Yang. Svdiff: Compact parameter space for diffusion fine-tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7323-7334, 2023. 3 +[14] Junjie He, Yifeng Geng, and Liefeng Bo. Uniportrait: A unified framework for identity-preserving single- and multi-human image personalization. arXiv preprint arXiv:2408.05939, 2024. 3 +[15] Zecheng He, Bo Sun, Felix Juefei-Xu, Haoyu Ma, Ankit Ramchandani, Vincent Cheung, Siddharth Shah, Anmol Kalia, Harihar Subramanyam, Alireza Zareian, et al. Imagine yourself: Tuning-free personalized image generation. arXiv preprint arXiv:2409.13346, 2024. 2 +[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2 +[17] Chanran Kim, Jeongin Lee, Shichang Joung, Bongmo Kim, and Yeul-Min Baek. Instantfamily: Masked attention for zero-shot multi-id image generation. arXiv preprint arXiv:2404.19427, 2024. 1, 2, 3, 4, 5 +[18] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3 +[19] Zhe Kong, Yong Zhang, Tianyu Yang, Tao Wang, Kaihao Zhang, Bizhu Wu, Guanying Chen, Wei Liu, and Wenhan Luo. Omg: Occlusion-friendly personalized multi-concept generation in diffusion models. In ECCV, 2024. 1, 2, 5 +[20] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 3 +[21] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 1 +[22] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22511-22521, 2023. 3 +[23] Zhen Li, Mingdeng Cao, Xintao Wang, Zhongang Qi, MingMing Cheng, and Ying Shan. Photomaker: Customizing realistic human photos via stacked id embedding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8640-8650, 2024. 2 +[24] Zhiheng Liu, Ruili Feng, Kai Zhu, Yifei Zhang, Kecheng Zheng, Yu Liu, Deli Zhao, Jingren Zhou, and Yang Cao. Cones: Concept neurons in diffusion models for customized generation. arXiv preprint arXiv:2303.05125, 2023. 3 +[25] Nanye Ma, Mark Goldstein, Michael S Albergo, Nicholas M Boffi, Eric Vanden-Eijnden, and Saining Xie. Sit: Exploring flow and diffusion-based generative models with scalable + +interpolant transformers. arXiv preprint arXiv:2401.08740, 2024. 2 +[26] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2 +[27] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 2 +[28] Xu Peng, Junwei Zhu, Boyuan Jiang, Ying Tai, Donghao Luo, Jiangning Zhang, Wei Lin, Taisong Jin, Chengjie Wang, and Rongrong Ji. Portraitbooth: A versatile portrait model for fast identity-preserved personalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27080-27090, 2024. 2 +[29] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. In The Twelfth International Conference on Learning Representations, 2024. 5 +[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 7 +[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3 +[32] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 2 +[33] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdreambooth: Hypernetworks for fast personalization of text-to-image models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6527-6536, 2024. 2 +[34] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 2 +[35] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 815-823, 2015. 5 +[36] Noam Shazeer and Mitchell Stern. Adafactor: Adaptive learning rates with sublinear memory cost. In International + +Conference on Machine Learning, pages 4596-4604. PMLR, 2018. 5 +[37] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 2 +[38] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. +[39] Yang Sui, Huy Phan, Jinqi Xiao, Tianfang Zhang, Zijie Tang, Cong Shi, Yan Wang, Yingying Chen, and Bo Yuan. Disdet: Exploring detectability of backdoor attack on diffusion models. arXiv preprint arXiv:2402.02739, 2024. 2 +[40] Yoad Tewel, Rinon Gal, Gal Chechik, and Yuval Atzmon. Key-locked rank one editing for text-to-image personalization. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 3 +[41] Hanhui Wang, Yihua Zhang, Ruizheng Bai, Yue Zhao, Sijia Liu, and Zhengzhong Tu. Edit away and my face will not stay: Personal biometric defense against malicious generative editing. arXiv preprint arXiv:2411.16832, 2024. 2 +[42] Kuan-Chieh Wang, Daniil Ostashev, Yuwei Fang, Sergey Tulyakov, and Kfir Aberman. Moa: Mixture-of-attention for subject-context disentanglement in personalized image generation. arXiv preprint arXiv:2404.11565, 2024. 3 +[43] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, and Anthony Chen. Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519, 2024. 1, 2, 5 +[44] Xudong Wang, Trevor Darrell, Sai Saketh Rambhatla, Rohit Girdhar, and Ishan Misra. Instancediffusion: Instance-level control for image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6232-6242, 2024. 3 +[45] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15943-15953, 2023. 2 +[46] Yuxin Wen, Neel Jain, John Kirchenbauer, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Hard prompts made easy: Gradient-based discrete optimization for prompt tuning and discovery. Advances in Neural Information Processing Systems, 36, 2024. 2 +[47] Guangxuan Xiao, Tianwei Yin, William T Freeman, Frédo Durand, and Song Han. Fastcomposer: Tuning-free multisubject image generation with localized attention. International Journal of Computer Vision, pages 1-20, 2024. 3 +[48] Jinqi Xiao, Miao Yin, Yu Gong, Xiao Zang, Jian Ren, and Bo Yuan. COMCAT: Towards efficient compression and customization of attention-based vision models. In Proceedings of the 40th International Conference on Machine Learning, pages 38125-38136. PMLR, 2023. 2 +[49] Jinqi Xiao, Shen Sang, Tiancheng Zhi, Jing Liu, Qing Yan, Linjie Luo, and Bo Yuan. Coap: Memory-efficient training + +with correlation-aware gradient projection. arXiv preprint arXiv:2412.00071, 2024. 2 +[50] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 2, 4 +[51] Ge Yuan, Xiaodong Cun, Yong Zhang, Maomao Li, Chenyang Qi, Xintao Wang, Ying Shan, and Huicheng Zheng. Inserting anybody in diffusion models via celeb basis. arXiv preprint arXiv:2306.00926, 2023. 2 +[52] Kaipeng Zhang, Zhanpeng Zhang, Zhifeng Li, and Yu Qiao. Joint face detection and alignment using multitask cascaded convolutional networks. IEEE signal processing letters, 23 (10):1499-1503, 2016. 1 +[53] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 3 +[54] Yimeng Zhang, Xin Chen, Jinghan Jia, Yihua Zhang, Chongyu Fan, Jiancheng Liu, Mingyi Hong, Ke Ding, and Sijia Liu. Defensive unlearning with adversarial training for robust concept erasure in diffusion models. arXiv preprint arXiv:2405.15234, 2024. 2 +[55] Yimeng Zhang, Jinghan Jia, Xin Chen, Aochuan Chen, Yihua Zhang, Jiancheng Liu, Ke Ding, and Sijia Liu. To generate or not? safety-driven unlearned diffusion models are still easy to generate unsafe images... for now. In European Conference on Computer Vision, pages 385-403. Springer, 2024. 2 +[56] Yihua Zhang, Yimeng Zhang, Yuguang Yao, Jinghan Jia, Jiancheng Liu, Xiaoming Liu, and Sijia Liu. Unlearncanvas: A stylized image dataset to benchmark machine unlearning for diffusion models. arXiv preprint arXiv:2402.11846, 2024. 2 +[57] Yufan Zhou, Ruiyi Zhang, Tong Sun, and Jinhui Xu. Enhancing detail preservation for customized text-to-image generation: A regularization-free approach. arXiv preprint arXiv:2305.13579, 2023. 2 \ No newline at end of file diff --git a/2025/ID-Patch_ Robust ID Association for Group Photo Personalization/images.zip b/2025/ID-Patch_ Robust ID Association for Group Photo Personalization/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5073d2da1cfcb9c8eca903eca1012e7ac648d288 --- /dev/null +++ b/2025/ID-Patch_ Robust ID Association for Group Photo Personalization/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8c21f3b56ac226c8ec75397dee9205b7408526bbac9308662c2a53e9f0a837a5 +size 878575 diff --git a/2025/ID-Patch_ Robust ID Association for Group Photo Personalization/layout.json b/2025/ID-Patch_ Robust ID Association for Group Photo Personalization/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ce5d24c2ea3b4df11d8fd9083dd5ce2c6ffb0b50 --- /dev/null +++ b/2025/ID-Patch_ Robust ID Association for Group Photo Personalization/layout.json @@ -0,0 +1,9366 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 101, + 103, + 510, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 103, + 510, + 121 + ], + "spans": [ + { + "bbox": [ + 101, + 103, + 510, + 121 + ], + "type": "text", + "content": "ID-Patch: Robust ID Association for Group Photo Personalization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 142, + 451, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 142, + 451, + 173 + ], + "spans": [ + { + "bbox": [ + 160, + 142, + 451, + 173 + ], + "type": "text", + "content": "Yimeng Zhang" + }, + { + "bbox": [ + 160, + 142, + 451, + 173 + ], + "type": "inline_equation", + "content": "^{1,2,*}" + }, + { + "bbox": [ + 160, + 142, + 451, + 173 + ], + "type": "text", + "content": ", Tiancheng Zhi" + }, + { + "bbox": [ + 160, + 142, + 451, + 173 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 160, + 142, + 451, + 173 + ], + "type": "text", + "content": ", Jing Liu" + }, + { + "bbox": [ + 160, + 142, + 451, + 173 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 160, + 142, + 451, + 173 + ], + "type": "text", + "content": ", Shen Sang" + }, + { + "bbox": [ + 160, + 142, + 451, + 173 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 160, + 142, + 451, + 173 + ], + "type": "text", + "content": ", Liming Jiang" + }, + { + "bbox": [ + 160, + 142, + 451, + 173 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 160, + 142, + 451, + 173 + ], + "type": "text", + "content": ", Qing Yan" + }, + { + "bbox": [ + 160, + 142, + 451, + 173 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 160, + 142, + 451, + 173 + ], + "type": "text", + "content": ", Sijia Liu" + }, + { + "bbox": [ + 160, + 142, + 451, + 173 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 160, + 142, + 451, + 173 + ], + "type": "text", + "content": ", Linjie Luo" + }, + { + "bbox": [ + 160, + 142, + 451, + 173 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 190, + 184, + 416, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 184, + 416, + 200 + ], + "spans": [ + { + "bbox": [ + 190, + 184, + 416, + 200 + ], + "type": "text", + "content": "1 ByteDance Inc., 2 Michigan State University" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 60, + 255, + 179, + 373 + ], + "blocks": [ + { + "bbox": [ + 65, + 226, + 175, + 235 + ], + "lines": [ + { + "bbox": [ + 65, + 226, + 175, + 235 + ], + "spans": [ + { + "bbox": [ + 65, + 226, + 175, + 235 + ], + "type": "text", + "content": "Pose Condition & ID Location" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 77, + 238, + 163, + 253 + ], + "lines": [ + { + "bbox": [ + 77, + 238, + 163, + 253 + ], + "spans": [ + { + "bbox": [ + 77, + 238, + 163, + 253 + ], + "type": "text", + "content": "\"An indoor celebration scene featuring three women...\"" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 60, + 255, + 179, + 373 + ], + "lines": [ + { + "bbox": [ + 60, + 255, + 179, + 373 + ], + "spans": [ + { + "bbox": [ + 60, + 255, + 179, + 373 + ], + "type": "image", + "image_path": "d64583b412951698b75d34769dd212fc91c0de9fd0546b8d0ec9f60eba4a078b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 383, + 555, + 449 + ], + "lines": [ + { + "bbox": [ + 54, + 383, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 54, + 383, + 555, + 449 + ], + "type": "text", + "content": "Figure 1. Comparison with state-of-the-art multi-identity generation methods. From left to right: the condition inputs followed by results generated using OMG [19] (with InstantID [43]), InstantFamily [17], and our proposed ID-Patch approach. Red dashed boxes highlight failures. OMG fails to preserve the hairstyles of the middle person, and creates artifacts for the right woman's hand, possibly because of the inconsistency between its generation stages (explained in Fig. 2). InstantFamily suffers from ID leakage, resulting the incorrect ID of the middle person. Our approach preserves the detailed identity of each person. In addition, our approach is 7 times faster than OMG and has less computational overhead than InstantFamily." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 184, + 255, + 303, + 373 + ], + "blocks": [ + { + "bbox": [ + 210, + 227, + 272, + 235 + ], + "lines": [ + { + "bbox": [ + 210, + 227, + 272, + 235 + ], + "spans": [ + { + "bbox": [ + 210, + 227, + 272, + 235 + ], + "type": "text", + "content": "OMG + InstantID" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 197, + 240, + 285, + 249 + ], + "lines": [ + { + "bbox": [ + 197, + 240, + 285, + 249 + ], + "spans": [ + { + "bbox": [ + 197, + 240, + 285, + 249 + ], + "type": "text", + "content": "Generation Time: 71.1s" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 184, + 255, + 303, + 373 + ], + "lines": [ + { + "bbox": [ + 184, + 255, + 303, + 373 + ], + "spans": [ + { + "bbox": [ + 184, + 255, + 303, + 373 + ], + "type": "image", + "image_path": "4debe4e158bfe18c9c369aab5f88a45b9bba5969529a1cdcd99797b26ad68166.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 308, + 255, + 427, + 373 + ], + "blocks": [ + { + "bbox": [ + 339, + 227, + 391, + 236 + ], + "lines": [ + { + "bbox": [ + 339, + 227, + 391, + 236 + ], + "spans": [ + { + "bbox": [ + 339, + 227, + 391, + 236 + ], + "type": "text", + "content": "InstantFamily" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 326, + 240, + 409, + 249 + ], + "lines": [ + { + "bbox": [ + 326, + 240, + 409, + 249 + ], + "spans": [ + { + "bbox": [ + 326, + 240, + 409, + 249 + ], + "type": "text", + "content": "Generation Time: 9.6s" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 308, + 255, + 427, + 373 + ], + "lines": [ + { + "bbox": [ + 308, + 255, + 427, + 373 + ], + "spans": [ + { + "bbox": [ + 308, + 255, + 427, + 373 + ], + "type": "image", + "image_path": "e7448aa925b865f1f7ed44fbfdf753faee8352c4cb3cf97b50cd71f20f655d06.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 432, + 255, + 551, + 373 + ], + "blocks": [ + { + "bbox": [ + 462, + 227, + 520, + 236 + ], + "lines": [ + { + "bbox": [ + 462, + 227, + 520, + 236 + ], + "spans": [ + { + "bbox": [ + 462, + 227, + 520, + 236 + ], + "type": "text", + "content": "ID-Patch (Ours)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 449, + 241, + 531, + 249 + ], + "lines": [ + { + "bbox": [ + 449, + 241, + 531, + 249 + ], + "spans": [ + { + "bbox": [ + 449, + 241, + 531, + 249 + ], + "type": "text", + "content": "Generation Time: 9.4s" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 432, + 255, + 551, + 373 + ], + "lines": [ + { + "bbox": [ + 432, + 255, + 551, + 373 + ], + "spans": [ + { + "bbox": [ + 432, + 255, + 551, + 373 + ], + "type": "image", + "image_path": "5d9a202bf8b9bc0524d4461d6ae840f929a6599916cadd4a0adb0673632e912b.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 151, + 459, + 200, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 459, + 200, + 472 + ], + "spans": [ + { + "bbox": [ + 151, + 459, + 200, + 472 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 54, + 487, + 297, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 487, + 297, + 691 + ], + "spans": [ + { + "bbox": [ + 54, + 487, + 297, + 691 + ], + "type": "text", + "content": "The ability to synthesize personalized group photos and specify the positions of each identity offers immense creative potential. While such imagery can be visually appealing, it presents significant challenges for existing technologies. A persistent issue is identity (ID) leakage, where injected facial features interfere with one another, resulting in low face resemblance, incorrect positioning, and visual artifacts. Existing methods suffer from limitations such as the reliance on segmentation models, increased runtime, or a high probability of ID leakage. To address these challenges, we propose ID-Patch, a novel method that provides robust association between identities and 2D positions. Our approach generates an ID patch and ID embeddings from the same facial features: the ID patch is positioned on the conditional image for precise spatial control, while the ID embeddings integrate with text embeddings to ensure high resemblance. Experimental results demon" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 460, + 555, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 555, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 555, + 509 + ], + "type": "text", + "content": "strate that ID-Patch surpasses baseline methods across metrics, such as face ID resemblance, ID-position association accuracy, and generation efficiency. Project Page is: https://byteaigc.github.io/ID-Patch/" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 314, + 536, + 395, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 536, + 395, + 548 + ], + "spans": [ + { + "bbox": [ + 314, + 536, + 395, + 548 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 558, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 558, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 558, + 555, + 715 + ], + "type": "text", + "content": "Personalized group photo generation has wide-ranging applications in entertainment, advertising, social media, and virtual reality. For example, users might want to create a group photo of friends at a virtual party, even if some could not attend in person. This technology enables the creation of engaging and memorable content that reflects real-life social interactions and relationships. However, this task poses significant challenges, particularly the risk of identity (ID) leakage, where the identity information of one individual inadvertently affects the representation of another at a different position. Such blending compromises the distinctiveness and accuracy of each person's features. Besides, being able to specify the position of each identity in the image is" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 702, + 217, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 702, + 217, + 713 + ], + "spans": [ + { + "bbox": [ + 70, + 702, + 217, + 713 + ], + "type": "text", + "content": "*Work done during internship at ByteDance." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2986" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 107 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 107 + ], + "type": "text", + "content": "an important function when users want to create a specific layout of people. Incorrect positioning of identities may lead to undesired results." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 110, + 296, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 110, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 110, + 296, + 407 + ], + "type": "text", + "content": "State-of-the-art group photo personalization methods largely rely on diffusion models [3, 6, 16, 26, 37-39, 41, 48, 49, 54, 56], which deliver superior visual quality when paired with effective prompt engineering [4, 8, 25, 27, 34, 55]. To address the issue of ID leakage, OMG [19] first generates an image without ID injection. Then, it estimates segmentation masks for the individuals in the image and subsequently injects the IDs into each segmented region. However, it depends on a segmentation model, which might fail in challenging cases. Besides, the content generated in the two stages may conflict with each other, leading to reduced visual fidelity. See Fig.1 and Fig.2 for example. Moreover, OMG requires running the denoising process separately for each identity, resulting in a runtime that increases linearly with the number of people. Alternatively, InstantFamily [17] employs a single-pass generation strategy by manipulating cross-attention masks. This approach uses approximate head region masks for each ID to direct pixel attention exclusively to specific ID embeddings derived from facial features. However, it struggles to prevent ID leakage due to two key issues: (1) imprecise masks or close proximity of faces, which can cause overlaps, and (2) unintended information propagation through self-attention and convolutional layers. Please see Fig.1 and Fig.3 for more details." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 411, + 296, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 296, + 589 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 296, + 589 + ], + "type": "text", + "content": "In response to these limitations, we propose a novel method named ID-Patch, designed for efficient and robust ID-position association in group photo generation. Our idea is to let the model learn to associate input IDs with their spatially designated locations. Specifically, we project facial features onto a small RGB image patch called ID patch and token embeddings called ID embeddings. The ID patches are placed onto the conditioning image of a ControlNet [53] according to face locations to ensure precise position control of IDs within the generated images. ID embeddings are integrated with text embeddings to enhance facial details. Our ID-Patch method seamlessly integrates with various types of spatial conditions, such as poses, canny edges or depth map, enhancing the robustness and flexibility of our method." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 592, + 255, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 592, + 255, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 592, + 255, + 604 + ], + "type": "text", + "content": "We summarize our key contributions as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 605, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 653 + ], + "type": "text", + "content": "- ID-Patch Method: We introduce a novel approach for multi-ID image generation that directly links face ID features with their locations using visual patches, ensuring accurate resemblance and position control." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 654, + 295, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 701 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 701 + ], + "type": "text", + "content": "- Efficiency Improvement: Our method enhances the efficiency of multi-ID image generation by simplifying the process to adding ID patches and concatenating ID embeddings, reducing computational overhead." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "content": "- Simplified Control: We eliminate the need for auxiliary" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 317, + 71, + 393, + 147 + ], + "blocks": [ + { + "bbox": [ + 317, + 71, + 393, + 147 + ], + "lines": [ + { + "bbox": [ + 317, + 71, + 393, + 147 + ], + "spans": [ + { + "bbox": [ + 317, + 71, + 393, + 147 + ], + "type": "image", + "image_path": "9a9242358b947183b0116abda6bcf658424be76bc062b0f0fb1fa029e1eb432d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 149, + 387, + 159 + ], + "lines": [ + { + "bbox": [ + 323, + 149, + 387, + 159 + ], + "spans": [ + { + "bbox": [ + 323, + 149, + 387, + 159 + ], + "type": "text", + "content": "(a) OMG first stage" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 394, + 71, + 471, + 148 + ], + "blocks": [ + { + "bbox": [ + 394, + 71, + 471, + 148 + ], + "lines": [ + { + "bbox": [ + 394, + 71, + 471, + 148 + ], + "spans": [ + { + "bbox": [ + 394, + 71, + 471, + 148 + ], + "type": "image", + "image_path": "0b75236d224893c93a9df368cdade44119f60df4de1c4fee8794b9f6b5564b24.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 150, + 460, + 158 + ], + "lines": [ + { + "bbox": [ + 406, + 150, + 460, + 158 + ], + "spans": [ + { + "bbox": [ + 406, + 150, + 460, + 158 + ], + "type": "text", + "content": "(b) Human mask" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 473, + 71, + 550, + 148 + ], + "blocks": [ + { + "bbox": [ + 473, + 71, + 550, + 148 + ], + "lines": [ + { + "bbox": [ + 473, + 71, + 550, + 148 + ], + "spans": [ + { + "bbox": [ + 473, + 71, + 550, + 148 + ], + "type": "image", + "image_path": "be8c39d8f2fd7623878a30f02d4c55d21f4f345d4886d704ecd42672f0b72194.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 478, + 149, + 545, + 158 + ], + "lines": [ + { + "bbox": [ + 478, + 149, + 545, + 158 + ], + "spans": [ + { + "bbox": [ + 478, + 149, + 545, + 158 + ], + "type": "text", + "content": "(c) OMG final result" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 317, + 232, + 392, + 332 + ], + "blocks": [ + { + "bbox": [ + 313, + 160, + 553, + 226 + ], + "lines": [ + { + "bbox": [ + 313, + 160, + 553, + 226 + ], + "spans": [ + { + "bbox": [ + 313, + 160, + 553, + 226 + ], + "type": "text", + "content": "Figure 2. Why does OMG fail in Fig. 1? OMG relies on a first stage result (no ID awareness) and its human segmentation to control layout. The second stage paints injected IDs within corresponding masks. However, the second pass may conflict with the first stage semantics, leading to deteriorated ID resemblance (long hair in the middle of image (c)) or artifacts (hand in the middle)." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 317, + 232, + 392, + 332 + ], + "lines": [ + { + "bbox": [ + 317, + 232, + 392, + 332 + ], + "spans": [ + { + "bbox": [ + 317, + 232, + 392, + 332 + ], + "type": "image", + "image_path": "f426f777d41f1f522a8d88580de0c331e5c28eeac4cfb4d0503105746da9ae63.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 334, + 365, + 343 + ], + "lines": [ + { + "bbox": [ + 345, + 334, + 365, + 343 + ], + "spans": [ + { + "bbox": [ + 345, + 334, + 365, + 343 + ], + "type": "text", + "content": "(a) ID" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 345, + 553, + 378 + ], + "lines": [ + { + "bbox": [ + 313, + 345, + 553, + 378 + ], + "spans": [ + { + "bbox": [ + 313, + 345, + 553, + 378 + ], + "type": "text", + "content": "Figure 3. ID leakage of InstantFamily [17]. The text prompt describes a man and a woman standing together, and ID (a) is only injected into the left man." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 394, + 232, + 472, + 332 + ], + "blocks": [ + { + "bbox": [ + 394, + 232, + 472, + 332 + ], + "lines": [ + { + "bbox": [ + 394, + 232, + 472, + 332 + ], + "spans": [ + { + "bbox": [ + 394, + 232, + 472, + 332 + ], + "type": "image", + "image_path": "6cf9aead00171fe6526fd123fe4ae035538091c65b13c96ce1c9a3b02e21265c.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 395, + 334, + 472, + 344 + ], + "lines": [ + { + "bbox": [ + 395, + 334, + 472, + 344 + ], + "spans": [ + { + "bbox": [ + 395, + 334, + 472, + 344 + ], + "type": "text", + "content": "(b) InstantFamily result" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 473, + 232, + 550, + 332 + ], + "blocks": [ + { + "bbox": [ + 473, + 232, + 550, + 332 + ], + "lines": [ + { + "bbox": [ + 473, + 232, + 550, + 332 + ], + "spans": [ + { + "bbox": [ + 473, + 232, + 550, + 332 + ], + "type": "image", + "image_path": "84e3aa550abef8c2ee884f0ddd08a91d769d292c9790914f313689485c913f51.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 481, + 334, + 542, + 343 + ], + "lines": [ + { + "bbox": [ + 481, + 334, + 542, + 343 + ], + "spans": [ + { + "bbox": [ + 481, + 334, + 542, + 343 + ], + "type": "text", + "content": "(c) ID-Patch result" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 322, + 396, + 553, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 396, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 322, + 396, + 553, + 431 + ], + "type": "text", + "content": "segmentation models, requiring only a single point for ID position control, thereby increasing robustness and reducing complexity." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 314, + 431, + 553, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 431, + 553, + 479 + ], + "spans": [ + { + "bbox": [ + 314, + 431, + 553, + 479 + ], + "type": "text", + "content": "- Superior Performance: Our approach achieves higher fidelity in ID resemblance and better position control, especially in complex images with multiple identities, as demonstrated by our experimental results." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 314, + 490, + 400, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 490, + 400, + 502 + ], + "spans": [ + { + "bbox": [ + 314, + 490, + 400, + 502 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 510, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 714 + ], + "type": "text", + "content": "Single-ID Personalization. The personalization of diffusion models has emerged as a popular research topic. Finetuning or optimization based methods [9, 10, 32, 33, 46] have the potential to reach high quality results in terms of ID resemblance, but are time-consuming. Inference-only methods [15, 23, 28, 43, 45, 50, 51, 57] eliminate the finetuning process usually by encoding the concept into token embeddings and injecting them into the diffusion models via cross-attention. However, most of these works focus on single concept injection rather than group photo generation. Multi-ID Personalization. The main challenge of personalized group photo generation is ID blending, i.e., the identity information of person A may affect the generation of person B at another position. To resolve this issue, OMG [19] performs separate denoising processes. First, a noncustomized image is generated and masks of each person are extracted. Second, masked regions are painted using" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2987" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 553, + 213 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 553, + 213 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 553, + 213 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 553, + 213 + ], + "type": "image", + "image_path": "3424c2213de91929ba6e557856107180c698ce40c1fb3ff61960ff2edb0bddc0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 215, + 555, + 270 + ], + "lines": [ + { + "bbox": [ + 55, + 215, + 555, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 215, + 555, + 270 + ], + "type": "text", + "content": "Figure 4. Our pipeline. Given a text prompt (e.g., two people shaking hands), " + }, + { + "bbox": [ + 55, + 215, + 555, + 270 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 215, + 555, + 270 + ], + "type": "text", + "content": " face images and locations, we generate an image with " + }, + { + "bbox": [ + 55, + 215, + 555, + 270 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 215, + 555, + 270 + ], + "type": "text", + "content": " IDs. We extract face features for each ID and project them into ID patches and ID embeddings. ID patches are rendered on a black canvas (or added on top of a pose image) according to face locations and sent into a ControlNet to control the positions of generated faces. ID embeddings are appended to text embeddings to provide detailed face information to the diffusion model and ControlNet via cross-attention." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 283, + 297, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 283, + 297, + 486 + ], + "spans": [ + { + "bbox": [ + 54, + 283, + 297, + 486 + ], + "type": "text", + "content": "separate customized models and stitched together. This approach is slow and conflicts between stages may cause identity loss. Additionally, it relies on a segmentation module, which may fail in challenging cases. To avoid running separate customized models, recent works [14, 42, 47] let the network to learn to attend to specific concept tokens by supervising the attention mask using segmentation or ID routing. However, these methods do not allow the specification of the accurate locations of each ID. Although they can be combined with pose conditions, describing the assignment of IDs to positions in text may be hard. InstantFamily [17] provides an opportunity to specify the approximate location of each ID by manipulating the attention masks during both training and inference. However, no constraints have been imposed to avoid the ID leakage through self-attention and convolutional layers, leading to ID blending, omitting or swapping." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 487, + 296, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 296, + 619 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 296, + 619 + ], + "type": "text", + "content": "Multiple General Concepts. Besides human generation, there are series of work focusing on the customization of multiple general concepts [1, 2, 7, 12, 13, 20, 24, 40, 47] and grounded image generation [2, 22, 44] associating objects with positions. GLIGEN [22] integrates grounding tokens, which align textual prompts with specific image regions. SpaText [2] and InstanceDiffusion [44] allows users to define the placement and attributes of objects. However, these concepts are usually semantically distinct while it is hard to describe the human identities in text and subtle details matter." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 632, + 111, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 632, + 111, + 645 + ], + "spans": [ + { + "bbox": [ + 55, + 632, + 111, + 645 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "content": "In this section, we first outline some preliminaries about text-to-image diffusion models and ControlNets. Subsequently, we detail the problem formulation for multi-ID image generation and the design of our modules, including the ID patch and ID embeddings, as well as the training and" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 283, + 394, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 283, + 394, + 294 + ], + "spans": [ + { + "bbox": [ + 314, + 283, + 394, + 294 + ], + "type": "text", + "content": "inference schemes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 301, + 400, + 314 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 301, + 400, + 314 + ], + "spans": [ + { + "bbox": [ + 314, + 301, + 400, + 314 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 320, + 555, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 320, + 555, + 440 + ], + "spans": [ + { + "bbox": [ + 313, + 320, + 555, + 440 + ], + "type": "text", + "content": "Latent diffusion models (LDMs). LDMs [31] are image generation models that can incorporate conditioning signals into the generation process. Specifically, during training, an encoder " + }, + { + "bbox": [ + 313, + 320, + 555, + 440 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 313, + 320, + 555, + 440 + ], + "type": "text", + "content": " of variational autoencoder (VAE) [18] encodes an image " + }, + { + "bbox": [ + 313, + 320, + 555, + 440 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 320, + 555, + 440 + ], + "type": "text", + "content": " into a latent space, and a neural network " + }, + { + "bbox": [ + 313, + 320, + 555, + 440 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(z_t,t,c)" + }, + { + "bbox": [ + 313, + 320, + 555, + 440 + ], + "type": "text", + "content": " performs the denoising process within this latent space. It attempts to predict the original Gaussian noise " + }, + { + "bbox": [ + 313, + 320, + 555, + 440 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 320, + 555, + 440 + ], + "type": "text", + "content": " from a noisy latent variable " + }, + { + "bbox": [ + 313, + 320, + 555, + 440 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 313, + 320, + 555, + 440 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 320, + 555, + 440 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 320, + 555, + 440 + ], + "type": "text", + "content": " represents the diffusion timestep, and " + }, + { + "bbox": [ + 313, + 320, + 555, + 440 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 320, + 555, + 440 + ], + "type": "text", + "content": " is the conditioning signal, injected via cross-attention. The loss function for an LDM is defined as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 340, + 449, + 555, + 465 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 449, + 555, + 465 + ], + "spans": [ + { + "bbox": [ + 340, + 449, + 555, + 465 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {L D M}} = \\mathbb {E} _ {\\mathcal {E} (x), \\epsilon \\sim \\mathcal {N} (0, 1), t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (z _ {t}, t, c) \\| _ {2} ^ {2} \\right] \\quad (1)", + "image_path": "bd62bbc624a326cd0953695a9bb371be3e637ef16c65c5fae45e404be0abfab9.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 472, + 556, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 556, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 556, + 605 + ], + "type": "text", + "content": "ControlNet. The incorporation of ControlNet [53] into diffusion models marks a significant advancement in controlled image synthesis. ControlNet is designed to steer the image generation process by applying spatial conditioning signals such as rasterized pose images, depth maps, or canny edges. It introduces a control network that processes these signals and adds their features to the base model's intermediate results. The control network is a trainable copy of the base model's encoder layers and is initialized with zero convolution. This effective integration achieves high quality controlled generation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 611, + 468, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 611, + 468, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 611, + 468, + 624 + ], + "type": "text", + "content": "3.2. Multi-ID Image Generation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": "This paper focuses on efficiently and effectively addressing the problem of multi-ID image generation utilizing conditioning signals that include text prompts, identity features, and face locations. The inputs to our model comprise text prompt embeddings " + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "c_{t}" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": ", facial features " + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "c_{f} = (f_{0}, f_{1}, \\ldots, f_{N-1})" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": " distinct identities, and their corresponding facial locations " + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "c_{l} = (l_{0}, l_{1}, \\ldots, l_{N-1})" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": ". The" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2988" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 115, + 181 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 115, + 181 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 115, + 181 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 115, + 181 + ], + "type": "image", + "image_path": "75b1ba4caaae833986a82dbf80ce97f0c559b36ea6c6fba660732df60e0771ea.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 184, + 97, + 193 + ], + "lines": [ + { + "bbox": [ + 78, + 184, + 97, + 193 + ], + "spans": [ + { + "bbox": [ + 78, + 184, + 97, + 193 + ], + "type": "text", + "content": "(a) ID" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 116, + 70, + 204, + 182 + ], + "blocks": [ + { + "bbox": [ + 116, + 70, + 204, + 182 + ], + "lines": [ + { + "bbox": [ + 116, + 70, + 204, + 182 + ], + "spans": [ + { + "bbox": [ + 116, + 70, + 204, + 182 + ], + "type": "image", + "image_path": "284eb7c683953fa9f26211122794f9c8e5276c1c5ebd9bf835d58f377fa2861e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 123, + 184, + 198, + 194 + ], + "lines": [ + { + "bbox": [ + 123, + 184, + 198, + 194 + ], + "spans": [ + { + "bbox": [ + 123, + 184, + 198, + 194 + ], + "type": "text", + "content": "(b) w/o ID embeddings" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 204, + 70, + 293, + 182 + ], + "blocks": [ + { + "bbox": [ + 204, + 70, + 293, + 182 + ], + "lines": [ + { + "bbox": [ + 204, + 70, + 293, + 182 + ], + "spans": [ + { + "bbox": [ + 204, + 70, + 293, + 182 + ], + "type": "image", + "image_path": "fa0e881be04ff7854c1192da7234ec1832dc71e9eb4c267f6238466226442436.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 209, + 184, + 287, + 194 + ], + "lines": [ + { + "bbox": [ + 209, + 184, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 209, + 184, + 287, + 194 + ], + "type": "text", + "content": "(c) with ID embeddings" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 202, + 295, + 247 + ], + "lines": [ + { + "bbox": [ + 55, + 202, + 295, + 247 + ], + "spans": [ + { + "bbox": [ + 55, + 202, + 295, + 247 + ], + "type": "text", + "content": "Figure 5. Effectiveness of ID embedding. Without ID embeddings, one can distinguish between the two people but the resemblance is low. Incorporating ID embeddings significantly improves face resemblance." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 266, + 296, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 296, + 315 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 296, + 315 + ], + "type": "text", + "content": "facial location " + }, + { + "bbox": [ + 55, + 266, + 296, + 315 + ], + "type": "inline_equation", + "content": "l_{i} = (x_{i},y_{i})" + }, + { + "bbox": [ + 55, + 266, + 296, + 315 + ], + "type": "text", + "content": " identifies the xy-coordinates of the nose tip for each identity within the intended image. The goal is to synthesize an image following the text description with each given face located in the designated place." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 332, + 222, + 344 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 332, + 222, + 344 + ], + "spans": [ + { + "bbox": [ + 55, + 332, + 222, + 344 + ], + "type": "text", + "content": "3.2.1 ID Patches and ID Embeddings" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 353, + 295, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 353, + 295, + 496 + ], + "spans": [ + { + "bbox": [ + 54, + 353, + 295, + 496 + ], + "type": "text", + "content": "As discussed in Sec. 3.1, ControlNet effectively integrates spatial conditional signals into diffusion models. Building on this capability, we introduce a method to encode ID location information using condition images, subsequently processed by ControlNet to guide ID localization within the generated image. Our approach features ID conditions manifesting as both patches and embeddings. Specifically, ID patches, precisely aligned with facial locations, are integrated via ControlNet to ensure accurate identity localization, while ID embeddings are utilized via cross-attention mechanisms in the LDM to enhance the detailed representation of facial features, as illustrated in Fig. 4." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 497, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 497, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 54, + 497, + 295, + 628 + ], + "type": "text", + "content": "Both ID patches and ID embeddings are derived from facial features, enabling a synergistic association between them, with both projection processes facilitated by the Perceiver Resampler [50]. For ID patches, we convert the facial feature " + }, + { + "bbox": [ + 54, + 497, + 295, + 628 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 54, + 497, + 295, + 628 + ], + "type": "text", + "content": " into an identity-specific patch " + }, + { + "bbox": [ + 54, + 497, + 295, + 628 + ], + "type": "inline_equation", + "content": "p_i \\in \\mathbb{R}^{P \\times P \\times 3}" + }, + { + "bbox": [ + 54, + 497, + 295, + 628 + ], + "type": "text", + "content": ", functioning similarly to a QR code, uniquely identifying each person and easily distinguishable from others. These ID patches are placed on a black canvas at specified facial locations to create an image " + }, + { + "bbox": [ + 54, + 497, + 295, + 628 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 54, + 497, + 295, + 628 + ], + "type": "text", + "content": ", which then serves as the conditioning input for ControlNet, thereby providing precise identity localization information." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "content": "As depicted in Fig. 5, while ID patches alone are generally sufficient for distinguishing between identities, they may not always achieve optimal results in generating images with high ID resemblance. To address this limitation, and inspired by InstantFamily [17], we also incorporate ID embeddings. These embeddings are designed to inject detailed facial appearance information into the dif" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 315, + 70, + 411, + 173 + ], + "blocks": [ + { + "bbox": [ + 315, + 70, + 411, + 173 + ], + "lines": [ + { + "bbox": [ + 315, + 70, + 411, + 173 + ], + "spans": [ + { + "bbox": [ + 315, + 70, + 411, + 173 + ], + "type": "image", + "image_path": "27e1ad580f5f8ddd1fa4a85fe3db5e48cf59290f0fe1c8cb12dcfd11fed7c274.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 317, + 174, + 408, + 184 + ], + "lines": [ + { + "bbox": [ + 317, + 174, + 408, + 184 + ], + "spans": [ + { + "bbox": [ + 317, + 174, + 408, + 184 + ], + "type": "text", + "content": "(a) ID (b) Result w/o pose" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 194, + 553, + 258 + ], + "lines": [ + { + "bbox": [ + 313, + 194, + 553, + 258 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 553, + 258 + ], + "type": "text", + "content": "Figure 6. ID-Patch combined with pose conditions. Provided with user ID images in (a), our method can generate results with only the specifications of nose tip locations as shown in (b). Merging ID patches with pose images (c) enhances spatial control over the generated results as seen in (d), without incurring any computational overhead." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 413, + 70, + 482, + 173 + ], + "blocks": [ + { + "bbox": [ + 413, + 70, + 482, + 173 + ], + "lines": [ + { + "bbox": [ + 413, + 70, + 482, + 173 + ], + "spans": [ + { + "bbox": [ + 413, + 70, + 482, + 173 + ], + "type": "image", + "image_path": "14e7f4b15a207bd8700592e1a9dcfd309dda13dbd6b064d9b06571c572985fcc.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 414, + 174, + 481, + 184 + ], + "lines": [ + { + "bbox": [ + 414, + 174, + 481, + 184 + ], + "spans": [ + { + "bbox": [ + 414, + 174, + 481, + 184 + ], + "type": "text", + "content": "(c) ID patch on pose" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 484, + 70, + 553, + 173 + ], + "blocks": [ + { + "bbox": [ + 484, + 70, + 553, + 173 + ], + "lines": [ + { + "bbox": [ + 484, + 70, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 484, + 70, + 553, + 173 + ], + "type": "image", + "image_path": "1103147907d956261f40fa38d51ab9e346cabc62a20c10fb31bd794471f6772b.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 486, + 174, + 551, + 184 + ], + "lines": [ + { + "bbox": [ + 486, + 174, + 551, + 184 + ], + "spans": [ + { + "bbox": [ + 486, + 174, + 551, + 184 + ], + "type": "text", + "content": "(d) Result with pose" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 274, + 554, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 274, + 554, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 554, + 369 + ], + "type": "text", + "content": "fusion model, thus enhancing the resemblance of the generated identities to their real-life counterparts. Specifically, ID embeddings " + }, + { + "bbox": [ + 313, + 274, + 554, + 369 + ], + "type": "inline_equation", + "content": "w_{i} \\in \\mathbb{R}^{2048 \\times M}" + }, + { + "bbox": [ + 313, + 274, + 554, + 369 + ], + "type": "text", + "content": ", share the same token dimension 2048 as the SDXL text embeddings. As shown in Fig.4, these embeddings are appended to text embeddings " + }, + { + "bbox": [ + 313, + 274, + 554, + 369 + ], + "type": "inline_equation", + "content": "c_{t}" + }, + { + "bbox": [ + 313, + 274, + 554, + 369 + ], + "type": "text", + "content": " to form an extended text embedding " + }, + { + "bbox": [ + 313, + 274, + 554, + 369 + ], + "type": "inline_equation", + "content": "c_{t}' = [c_{t}, w_{0}, w_{1}, \\ldots, w_{N-1}]" + }, + { + "bbox": [ + 313, + 274, + 554, + 369 + ], + "type": "text", + "content": ", which is then fed into both the UNet and ControlNet through cross-attention mechanisms." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 370, + 554, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 370, + 554, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 370, + 554, + 536 + ], + "type": "text", + "content": "Since human pose conditions are often utilized to enhance group photo generation, we can combine our ID-Patch approach with the pose conditions for multi-id image generation. As illustrated in Fig. 6, the ID patches can be overlaid onto a pose-conditioned image, creating a composite condition image that encodes both identity positioning and pose information. It is important to mention that while our method can incorporate pose information, it does not inherently require it; pose is just an optional enhancement, one can also choose other conditions like canny edge, depth map for specific applications. The adaptability of our ID-Patch approach presents enhanced control over generated results yet without incurring the extra overhead of an additional ControlNet branch." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 553, + 432, + 566 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 553, + 432, + 566 + ], + "spans": [ + { + "bbox": [ + 313, + 553, + 432, + 566 + ], + "type": "text", + "content": "3.2.2 Two-Stage Training" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 573, + 553, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 573, + 553, + 596 + ], + "spans": [ + { + "bbox": [ + 313, + 573, + 553, + 596 + ], + "type": "text", + "content": "The training of our model employs a conventional diffusion loss function, defined as follows:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 345, + 605, + 553, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 605, + 553, + 620 + ], + "spans": [ + { + "bbox": [ + 345, + 605, + 553, + 620 + ], + "type": "interline_equation", + "content": "L = \\mathbb {E} _ {\\mathcal {E} (x), \\epsilon \\sim \\mathcal {N} (0, 1), t} [ \\| \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}, t, I, c _ {t} ^ {\\prime}\\right) \\| _ {2} ^ {2} ] \\tag {2}", + "image_path": "9f165850b4a5ce9de2fcdd7f3441f2d5976a780c02617aadee3e4ca2960dc07a.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "content": "Initially, one might consider training the models end-to-end from scratch. However, our experiments reveal that such a method does not consistently ensure accurate ID localization, as depicted in Fig.7. This shortcoming likely stems from the network's tendency to preferentially inject ID information into ID embeddings, rather than ID patches, resulting in non-distinctive ID patches that could potentially" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2989" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 89, + 136, + 201 + ], + "blocks": [ + { + "bbox": [ + 59, + 89, + 136, + 201 + ], + "lines": [ + { + "bbox": [ + 59, + 89, + 136, + 201 + ], + "spans": [ + { + "bbox": [ + 59, + 89, + 136, + 201 + ], + "type": "image", + "image_path": "0e013eb6c9e2891b58771e3ad72e4f8114c525c700a6f443fb02a806d6509c4e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 75, + 203, + 120, + 212 + ], + "lines": [ + { + "bbox": [ + 75, + 203, + 120, + 212 + ], + "spans": [ + { + "bbox": [ + 75, + 203, + 120, + 212 + ], + "type": "text", + "content": "(a) Pose + ID" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 136, + 70, + 214, + 201 + ], + "blocks": [ + { + "bbox": [ + 136, + 70, + 214, + 201 + ], + "lines": [ + { + "bbox": [ + 136, + 70, + 214, + 201 + ], + "spans": [ + { + "bbox": [ + 136, + 70, + 214, + 201 + ], + "type": "image", + "image_path": "70787222617e083dc453dc28bdd54cbe210743c2d5ee1e84e96bd1db19bb09ba.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 141, + 203, + 211, + 213 + ], + "lines": [ + { + "bbox": [ + 141, + 203, + 211, + 213 + ], + "spans": [ + { + "bbox": [ + 141, + 203, + 211, + 213 + ], + "type": "text", + "content": "(b) Single-stage train" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 215, + 70, + 292, + 201 + ], + "blocks": [ + { + "bbox": [ + 215, + 70, + 292, + 201 + ], + "lines": [ + { + "bbox": [ + 215, + 70, + 292, + 201 + ], + "spans": [ + { + "bbox": [ + 215, + 70, + 292, + 201 + ], + "type": "image", + "image_path": "88bc0f3cafe9c6baa7b0459f2bcee5bb4a63c1b874aa48f6c8c54baecc10e238.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 223, + 203, + 285, + 213 + ], + "lines": [ + { + "bbox": [ + 223, + 203, + 285, + 213 + ], + "spans": [ + { + "bbox": [ + 223, + 203, + 285, + 213 + ], + "type": "text", + "content": "(c) Two-stage train" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 222, + 295, + 332 + ], + "lines": [ + { + "bbox": [ + 55, + 222, + 295, + 332 + ], + "spans": [ + { + "bbox": [ + 55, + 222, + 295, + 332 + ], + "type": "text", + "content": "Figure 7. Two-stage training to improve positioning robustness. Given pose and ID conditions in (a), single-stage training cannot fully prevent the incorrect face positioning issue. For example, in (b), the man is mistakenly placed at the central bottom position, generating an inharmonious result. (c) Two-stage training is introduced to solve this problem. As can be seen on the top row of this figure, the ID patches from our two-stage training approach are visually more distinguishable from each other, compared to the ones from single-stage training. Our experimental results prove that this is critical in solving the ID leak issues." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 342, + 296, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 296, + 450 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 296, + 450 + ], + "type": "text", + "content": "confuse ControlNet. In the single-stage model, as shown in Fig. 7, the ID patches exhibit only subtle differences between different individuals. To resolve this, we introduce a two-stage training approach. In the initial stage, ID embeddings are omitted, compelling the network to prioritize the incorporation of ID information into ID patches, thereby enhancing their distinctiveness. In the subsequent stage, ID embeddings are introduced and the model is fine-tuned to further improve the resemblance of the generated images." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 463, + 135, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 463, + 135, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 463, + 135, + 475 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 483, + 115, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 115, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 115, + 495 + ], + "type": "text", + "content": "4.1. Settings" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 502, + 295, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 295, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 295, + 693 + ], + "type": "text", + "content": "Implementation details. Facial features " + }, + { + "bbox": [ + 55, + 502, + 295, + 693 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 55, + 502, + 295, + 693 + ], + "type": "text", + "content": " we use are 512-dimensional vectors derived from ArcFace [5]. During training, we utilize the SDXL diffusion model [29]. During the inference phase, we employ the further finetuned SDXL, namely Juggernaut-X-v10*, as base model to enhance portrait quality. The inference process involves 50 timesteps. We set the ID token injection ratio to 0.8 to regulate the influence of identity features, where the initial " + }, + { + "bbox": [ + 55, + 502, + 295, + 693 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 55, + 502, + 295, + 693 + ], + "type": "text", + "content": " of timesteps operate without ID token embeddings to focus on the general style, followed by the remaining " + }, + { + "bbox": [ + 55, + 502, + 295, + 693 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 55, + 502, + 295, + 693 + ], + "type": "text", + "content": " timesteps which incorporate ID token embeddings to refine the identity-specific aspects of the images. The default size of ID-patch is " + }, + { + "bbox": [ + 55, + 502, + 295, + 693 + ], + "type": "inline_equation", + "content": "P \\times P = 64 \\times 64" + }, + { + "bbox": [ + 55, + 502, + 295, + 693 + ], + "type": "text", + "content": " and the default length of ID embeddings is " + }, + { + "bbox": [ + 55, + 502, + 295, + 693 + ], + "type": "inline_equation", + "content": "M = 16" + }, + { + "bbox": [ + 55, + 502, + 295, + 693 + ], + "type": "text", + "content": ". The network is trained using 8 NVIDIA H100 GPUs with batch size=4, learning rate=1e-05, weight decay = 0.01, using AdaFactor optimizer [36]." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 322, + 138, + 544, + 191 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 553, + 137 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 553, + 137 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 553, + 137 + ], + "type": "text", + "content": "Table 1. Performance comparison across four dimensions: identity resemblance, ID-position association accuracy, text alignment, and generation time. For qualitative metrics to evaluate similarity or accuracy, higher is better. Our approach significantly outperforms baseline methods in terms of resemblance, association, and generation time, with similar text alignment scores." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 322, + 138, + 544, + 191 + ], + "lines": [ + { + "bbox": [ + 322, + 138, + 544, + 191 + ], + "spans": [ + { + "bbox": [ + 322, + 138, + 544, + 191 + ], + "type": "table", + "html": "
MethodsID ↑Association ↑Text ↑Time (s) ↓
OMG + InstantID0.6440.9260.27184.83
InstantFamily0.6830.8620.27810.02
ID-Patch (Ours)0.7510.9580.2739.69
", + "image_path": "c6c1054bc4419bb2928700fd54b7c6042b25cd3f4e6e4ba71afbb50414829d90.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 207, + 553, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 207, + 553, + 243 + ], + "spans": [ + { + "bbox": [ + 313, + 207, + 553, + 243 + ], + "type": "text", + "content": "We train it for 720k iterations for stage 1, and stage 2 usually converges around 1 million iterations. We select the best model based on a 20 people validation set." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 244, + 554, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 244, + 554, + 363 + ], + "spans": [ + { + "bbox": [ + 313, + 244, + 554, + 363 + ], + "type": "text", + "content": "Dataset. Our training set consists of 17 million single-person and 1.95 million multi-person images. These images are obtained from purchased data and publicly available sources. For evaluation, we utilize two specialized datasets: 1) Style Dataset comprises 40 styles with pose and text prompt definitions, ranging from one to eight faces. 2) Identity Dataset includes 600 unique identities representing a variety of races and genders. Each style is paired with 20 randomly sampled unique identity combinations for quantitative evaluation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 365, + 554, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 554, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 554, + 519 + ], + "type": "text", + "content": "Baselines. Our analysis includes two state-of-the-art multi-ID generation methods as baselines: 1) OMG [19]: This method incorporates multiple identity features into separate segmented regions through latent stitching. It can be paired with a single ID injection method InstantID [43] for multi-ID generation. 2) InstantFamily [17]: This approach employs coarse attention masks within cross-attention layers, a mandatory process that ensures distinct identity features are injected into different image instances effectively. Because there is no official publicly available code for InstantFamily [17], for fairness, we implemented it based on the SDXL base model [29], and trained it using the same training data as our model." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": "Evaluation Metrics. We evaluate the methods in a pose-conditioned setting, because pose is a requirement of InstantFamily, and without pose condition, OMG is not able to generate faces at specified locations for its first stage. Besides, pose condition makes the head size consistent, which is easier for resemblance evaluation and reduces the impact of head sizes on generation quality. For comprehensive evaluation, we consider four critical dimensions: 1) Identity resemblance: we measure the identity resemblance by calculating the cosine similarity (CosSim) between the facial features of faces in the generated images and the reference faces. For fairness, we use FaceNet [35], a face model different from the ArcFace model used for training. 2) Association accuracy: we assess the accuracy of identity-position association in generated images by ensuring that the generated faces are most similar to their assigned coun" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 67, + 702, + 251, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 251, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 251, + 713 + ], + "type": "text", + "content": "\\*https://huggingface.co/RunDiffusion/Juggernaut-X-v10" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2990" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 73, + 60, + 217, + 186 + ], + "blocks": [ + { + "bbox": [ + 73, + 60, + 217, + 186 + ], + "lines": [ + { + "bbox": [ + 73, + 60, + 217, + 186 + ], + "spans": [ + { + "bbox": [ + 73, + 60, + 217, + 186 + ], + "type": "image", + "image_path": "6ad51e41ac9fd5e146601be20ebe6ee7d216be296facaeac83003acf660a81dd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 103, + 189, + 186, + 198 + ], + "lines": [ + { + "bbox": [ + 103, + 189, + 186, + 198 + ], + "spans": [ + { + "bbox": [ + 103, + 189, + 186, + 198 + ], + "type": "text", + "content": "(a) Identity Resemblance" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 235, + 60, + 379, + 186 + ], + "blocks": [ + { + "bbox": [ + 235, + 60, + 379, + 186 + ], + "lines": [ + { + "bbox": [ + 235, + 60, + 379, + 186 + ], + "spans": [ + { + "bbox": [ + 235, + 60, + 379, + 186 + ], + "type": "image", + "image_path": "74ba8deec9c02f2b28235561317fc0094418a060acfd19f6c5fd07d65792ebcb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 264, + 189, + 347, + 198 + ], + "lines": [ + { + "bbox": [ + 264, + 189, + 347, + 198 + ], + "spans": [ + { + "bbox": [ + 264, + 189, + 347, + 198 + ], + "type": "text", + "content": "(b) Association Accuracy" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 395, + 59, + 539, + 186 + ], + "blocks": [ + { + "bbox": [ + 395, + 59, + 539, + 186 + ], + "lines": [ + { + "bbox": [ + 395, + 59, + 539, + 186 + ], + "spans": [ + { + "bbox": [ + 395, + 59, + 539, + 186 + ], + "type": "image", + "image_path": "18f05bfa9f534f0cd24dff0c791aa88a5fab82002624d88200566a7f57a1db94.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 433, + 189, + 500, + 198 + ], + "lines": [ + { + "bbox": [ + 433, + 189, + 500, + 198 + ], + "spans": [ + { + "bbox": [ + 433, + 189, + 500, + 198 + ], + "type": "text", + "content": "(c) Generation Time" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 61, + 248, + 549, + 544 + ], + "blocks": [ + { + "bbox": [ + 55, + 208, + 555, + 242 + ], + "lines": [ + { + "bbox": [ + 55, + 208, + 555, + 242 + ], + "spans": [ + { + "bbox": [ + 55, + 208, + 555, + 242 + ], + "type": "text", + "content": "Figure 8. Performance evaluation of different model generations from three different aspects: identity resemblance, location accuracy, and generation time. As face number increases, the ID and location metric scores of OMG and Instantfamily drop much more significantly than our approach. Our approach also achieves near-constant generation time, in contrast to OMG's linearly increasing running time." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 61, + 248, + 549, + 544 + ], + "lines": [ + { + "bbox": [ + 61, + 248, + 549, + 544 + ], + "spans": [ + { + "bbox": [ + 61, + 248, + 549, + 544 + ], + "type": "image", + "image_path": "f6b54bb931d500df58634fba0dee751e2cb819c2a805d611f8d98d7b7c95847c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 553, + 555, + 609 + ], + "lines": [ + { + "bbox": [ + 55, + 553, + 555, + 609 + ], + "spans": [ + { + "bbox": [ + 55, + 553, + 555, + 609 + ], + "type": "text", + "content": "Figure 9. Comparison with baselines on pose-conditioned generation, where red dashed boxes highlight instances with low identity resemblance. In row 1, OMG fails to preserve the face shape of the third woman (from left to right) because its first stage result conflicts with this ID's face shape. InstantFamily incorrectly generates the second person because of the ID leakage from the third woman. In row 2, OMG does not generate correct hair styles and accurate facial features for the two people in the red boxes, while InstantFamily generates the wrong ID in the red box." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 621, + 297, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 621, + 297, + 706 + ], + "spans": [ + { + "bbox": [ + 55, + 621, + 297, + 706 + ], + "type": "text", + "content": "terparts, compared to all other input faces used in the generation process. In other words, we calculate the similarity between the " + }, + { + "bbox": [ + 55, + 621, + 297, + 706 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 621, + 297, + 706 + ], + "type": "text", + "content": "-th generated face and the " + }, + { + "bbox": [ + 55, + 621, + 297, + 706 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 621, + 297, + 706 + ], + "type": "text", + "content": "-th input face. Let " + }, + { + "bbox": [ + 55, + 621, + 297, + 706 + ], + "type": "inline_equation", + "content": "s(i) = \\arg \\max_{0 \\leq j \\leq N-1} \\mathrm{CosSim}(\\tilde{f}_i^{\\mathrm{gen}}, \\tilde{f}_j)" + }, + { + "bbox": [ + 55, + 621, + 297, + 706 + ], + "type": "text", + "content": " be the face index among all input faces that is most similar to the " + }, + { + "bbox": [ + 55, + 621, + 297, + 706 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 621, + 297, + 706 + ], + "type": "text", + "content": "-th generated face, where " + }, + { + "bbox": [ + 55, + 621, + 297, + 706 + ], + "type": "inline_equation", + "content": "\\tilde{f}_i^{\\mathrm{gen}}" + }, + { + "bbox": [ + 55, + 621, + 297, + 706 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 621, + 297, + 706 + ], + "type": "inline_equation", + "content": "\\tilde{f}_j" + }, + { + "bbox": [ + 55, + 621, + 297, + 706 + ], + "type": "text", + "content": " denote the FaceNet features of the " + }, + { + "bbox": [ + 55, + 621, + 297, + 706 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 621, + 297, + 706 + ], + "type": "text", + "content": "-th generated face and the " + }, + { + "bbox": [ + 55, + 621, + 297, + 706 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 621, + 297, + 706 + ], + "type": "text", + "content": "-th input face respec" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 621, + 555, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 621, + 555, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 621, + 555, + 658 + ], + "type": "text", + "content": "tively, and " + }, + { + "bbox": [ + 313, + 621, + 555, + 658 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 621, + 555, + 658 + ], + "type": "text", + "content": " is the number of faces in this image. When " + }, + { + "bbox": [ + 313, + 621, + 555, + 658 + ], + "type": "inline_equation", + "content": "s(i) = i" + }, + { + "bbox": [ + 313, + 621, + 555, + 658 + ], + "type": "text", + "content": ", the " + }, + { + "bbox": [ + 313, + 621, + 555, + 658 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 621, + 555, + 658 + ], + "type": "text", + "content": "-th generated face has the correct association. The identity-position association accuracy is defined as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 392, + 663, + 555, + 696 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 663, + 555, + 696 + ], + "spans": [ + { + "bbox": [ + 392, + 663, + 555, + 696 + ], + "type": "interline_equation", + "content": "\\frac {1}{N} \\sum_ {i = 0} ^ {N - 1} \\mathbf {1} \\{i = s (i) \\} \\tag {3}", + "image_path": "af1e8347494252635206629987d6fbe945ddb7032d0e4028ffb050a37beb3cd3.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 701, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 701, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 314, + 701, + 555, + 714 + ], + "type": "text", + "content": "This measurement assesses the correct placement of each" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "type": "text", + "content": "2991" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 67, + 138, + 282, + 214 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 294, + 137 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 294, + 137 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 294, + 137 + ], + "type": "text", + "content": "Table 2. Ablation study results for our proposed ID-Patch variants, designed to assess the impact of different modules (patch projection and token projection) and training/inferencing schemes on the performance of multi-ID image generation. The full method achieves the best ID resemblance and ID-position association, indicating the necessity of each module." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 138, + 282, + 214 + ], + "lines": [ + { + "bbox": [ + 67, + 138, + 282, + 214 + ], + "spans": [ + { + "bbox": [ + 67, + 138, + 282, + 214 + ], + "type": "table", + "html": "
VariantsID ↑Association ↑Text ↑
No patch projection0.3680.3450.271
No token projection0.5650.8950.284
Single stage training0.6600.7840.271
ID token injection ratio = 1.00.7500.9570.270
FULL0.7510.9580.273
", + "image_path": "5bbcc4fa4961d5db7cb5cb2e5f8972ddf2a453361cb6344b4ea8f4eac01c4ece.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 234, + 296, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 234, + 296, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 234, + 296, + 354 + ], + "type": "text", + "content": "identity within the image. 3) Text alignment: this metric evaluates the alignment between the textual context of the image generation prompt and the visual output. Cosine similarity is used to compare textual features extracted by the CLIP text encoder with visual features extracted by the CLIP image encoder from the generated images [30]. 4) Generation time: we measure the efficiency by recording the time (in seconds) required to generate an image on a NVIDIA A100 GPU, excluding the time taken for model loading and image I/O." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 374, + 205, + 388 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 374, + 205, + 388 + ], + "spans": [ + { + "bbox": [ + 55, + 374, + 205, + 388 + ], + "type": "text", + "content": "4.2. Comparison with Baselines" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 397, + 294, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 397, + 294, + 482 + ], + "spans": [ + { + "bbox": [ + 55, + 397, + 294, + 482 + ], + "type": "text", + "content": "As listed in Tab. 1, ID-Patch excels in identity and association metrics with scores of 0.751 and 0.958, respectively, showcasing its robust capability to maintain identity consistency and achieve precise placement in generated images. There is no significant difference in text alignment scores between compared methods. Notably, ID-Patch achieves the fastest image generation time." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "type": "text", + "content": "Further details on the performance with varying face numbers are shown in Fig. 8. As we observed no deterioration in text alignment with an increase in the number of faces across all methods, we have relegated the corresponding figure to the appendix. As the number of faces increases, both OMG and InstantFamily experience significant performance drops due to ID leakage. Although ID-Patch also shows a decline, it is less severe, mainly because the increase in face numbers reduces the average face area, adversely affecting the quality of smaller faces produced by the SDXL models. Moreover, the presence of more faces slightly extends the image generation time due to the increased demand for facial feature extraction. InstantFamily is slightly slower than our approach due to the overhead of masked attention. For OMG, injecting identity features into designated segmented areas causes a substantial increase in generation time; it takes about 2 minutes to generate an image with eight people, whereas ID-Patch requires only about 10 seconds. See Fig. 9 for visual comparisons." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 72, + 408, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 408, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 408, + 85 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 89, + 553, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 553, + 305 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 553, + 305 + ], + "type": "text", + "content": "We conduct an ablation study on variants of the proposed method in Tab. 2. The performance notably degrades when patch projection is omitted, leading to poor location alignment as the algorithm attempts to infer ID locations based on indirect cues such as gender and pose. Similarly, omitting token projection reduces resemblance by eliminating essential facial details, quantitatively validates the finding of Fig. 5. Single-stage training results in indistinguishable ID patches (as shown in Fig. 7), which impairs effective positioning. The default setting with an ID token injection ratio of 0.8 improves text alignment compared to injecting the ID token at every timestep, while maintaining the same level of identity resemblance. This improvement may arise because ID embeddings have little effect on facial detail during the early denoising steps, but become more influential in the later stages. It is important to note that even without the inference trick (i.e., using the ID token injection ratio), our proposed method still outperforms all baseline approaches." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 312, + 438, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 438, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 438, + 323 + ], + "type": "text", + "content": "4.4. Pose-Free Generation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 330, + 553, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 330, + 553, + 401 + ], + "spans": [ + { + "bbox": [ + 313, + 330, + 553, + 401 + ], + "type": "text", + "content": "Our ID-Patch model can be operated effectively without the need for pose conditioning, a flexibility that permits its integration with other conditions for image generation. This capability enables the use of ID-Patch in versatile configurations, whether plug-and-play or co-trained, which we will explore in subsequent subsections." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 402, + 554, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 554, + 521 + ], + "type": "text", + "content": "In the pose-free mode, the ControlNet uses a conditioning image that solely features ID patches against a black background. The placement of each ID patch at the tip of the nose ensures accurate preservation of facial locations during image synthesis, while head sizes and poses are variably generated based on the initial noise. As illustrated in Fig. 10, this methodology allows our model to produce aesthetically pleasing group photographs of different number of individuals across a variety of settings, demonstrating the robustness and adaptability of our approach." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 521, + 553, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 553, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 553, + 582 + ], + "type": "text", + "content": "The pose-free version of ID-Patch is designed to be compatible with other ControlNets (plug-and-play). As illustrated in Fig. 11, it is possible to combine our ID-Patch ControlNet the Canny edge ControlNet to achieve more precise control over the final image outcomes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 587, + 391, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 587, + 391, + 599 + ], + "spans": [ + { + "bbox": [ + 313, + 587, + 391, + 599 + ], + "type": "text", + "content": "4.5. Limitations" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 605, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 715 + ], + "type": "text", + "content": "As illustrated in Fig. 12, our method encounters two primary limitations. First, our implementation is based on the SDXL model, which occasionally produces inaccuracies in body and finger anatomy. Transitioning to a more advanced base model could potentially rectify this issue. Second, facial features in the current model may incorporate lighting and expression details that are not fully separated from identity information, potentially compromising the generation quality. Enriching our training dataset with multiple images" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2992" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 92, + 68, + 174, + 173 + ], + "blocks": [ + { + "bbox": [ + 69, + 89, + 77, + 148 + ], + "lines": [ + { + "bbox": [ + 69, + 89, + 77, + 148 + ], + "spans": [ + { + "bbox": [ + 69, + 89, + 77, + 148 + ], + "type": "text", + "content": "ID + ID Location" + } + ] + } + ], + "index": 0, + "angle": 270, + "type": "image_caption" + }, + { + "bbox": [ + 92, + 68, + 174, + 173 + ], + "lines": [ + { + "bbox": [ + 92, + 68, + 174, + 173 + ], + "spans": [ + { + "bbox": [ + 92, + 68, + 174, + 173 + ], + "type": "image", + "image_path": "6b74927a748a75fdf6b6a21127cb53612200650f4d5b1d8368a9a6611aea5a70.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 189, + 68, + 294, + 173 + ], + "blocks": [ + { + "bbox": [ + 189, + 68, + 294, + 173 + ], + "lines": [ + { + "bbox": [ + 189, + 68, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 189, + 68, + 294, + 173 + ], + "type": "image", + "image_path": "2feb6fae2b3fde4fb4d6e8a59c347a70428465b0c0481a2f81914a92975d2511.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 307, + 68, + 426, + 173 + ], + "blocks": [ + { + "bbox": [ + 307, + 68, + 426, + 173 + ], + "lines": [ + { + "bbox": [ + 307, + 68, + 426, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 68, + 426, + 173 + ], + "type": "image", + "image_path": "1bea2d862a140506b34a33356db25ba2e001fa0030e325abcdd9ca538404a9b8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 439, + 68, + 544, + 173 + ], + "blocks": [ + { + "bbox": [ + 439, + 68, + 544, + 173 + ], + "lines": [ + { + "bbox": [ + 439, + 68, + 544, + 173 + ], + "spans": [ + { + "bbox": [ + 439, + 68, + 544, + 173 + ], + "type": "image", + "image_path": "4c1b876c595d0ca45c26eeda1e652aff7f59c209fb9303ca88b051fcf3ac462e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 93, + 180, + 173, + 285 + ], + "blocks": [ + { + "bbox": [ + 69, + 201, + 78, + 265 + ], + "lines": [ + { + "bbox": [ + 69, + 201, + 78, + 265 + ], + "spans": [ + { + "bbox": [ + 69, + 201, + 78, + 265 + ], + "type": "text", + "content": "Generation Result" + } + ] + } + ], + "index": 5, + "angle": 270, + "type": "image_caption" + }, + { + "bbox": [ + 93, + 180, + 173, + 285 + ], + "lines": [ + { + "bbox": [ + 93, + 180, + 173, + 285 + ], + "spans": [ + { + "bbox": [ + 93, + 180, + 173, + 285 + ], + "type": "image", + "image_path": "c26e393fd68cc9ae27e522abbf4ca149cddc532b02d8e1e77f18e6f35243cc2b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 189, + 180, + 294, + 285 + ], + "blocks": [ + { + "bbox": [ + 189, + 180, + 294, + 285 + ], + "lines": [ + { + "bbox": [ + 189, + 180, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 189, + 180, + 294, + 285 + ], + "type": "image", + "image_path": "091862fcd47e0f47c4dcb8fa0a785ed6d57b1120e3555066af192d3961fcab68.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 180, + 426, + 285 + ], + "blocks": [ + { + "bbox": [ + 307, + 180, + 426, + 285 + ], + "lines": [ + { + "bbox": [ + 307, + 180, + 426, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 180, + 426, + 285 + ], + "type": "image", + "image_path": "6780fe24621dea2c6ee8c9c2a50f09f13b1dcfca30df6b009578bee328a10814.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 438, + 180, + 544, + 285 + ], + "blocks": [ + { + "bbox": [ + 438, + 180, + 544, + 285 + ], + "lines": [ + { + "bbox": [ + 438, + 180, + 544, + 285 + ], + "spans": [ + { + "bbox": [ + 438, + 180, + 544, + 285 + ], + "type": "image", + "image_path": "4145dbd7bf16d7cf16d17a23a0866652151c813191bccb66851df10f9366ebb6.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 57, + 389, + 105, + 437 + ], + "blocks": [ + { + "bbox": [ + 57, + 389, + 105, + 437 + ], + "lines": [ + { + "bbox": [ + 57, + 389, + 105, + 437 + ], + "spans": [ + { + "bbox": [ + 57, + 389, + 105, + 437 + ], + "type": "image", + "image_path": "c2bf392d9d0cb766ad432ff907109b8c37d19eed8d35bf1aa6a2d8b5bde139e6.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 57, + 437, + 104, + 486 + ], + "blocks": [ + { + "bbox": [ + 57, + 437, + 104, + 486 + ], + "lines": [ + { + "bbox": [ + 57, + 437, + 104, + 486 + ], + "spans": [ + { + "bbox": [ + 57, + 437, + 104, + 486 + ], + "type": "image", + "image_path": "bd194ab43cba5bbb98d920cf533990037813ed35980ee480f6dc40d93d8e0379.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 71, + 488, + 91, + 497 + ], + "lines": [ + { + "bbox": [ + 71, + 488, + 91, + 497 + ], + "spans": [ + { + "bbox": [ + 71, + 488, + 91, + 497 + ], + "type": "text", + "content": "(a) ID" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 107, + 350, + 200, + 486 + ], + "blocks": [ + { + "bbox": [ + 55, + 300, + 555, + 346 + ], + "lines": [ + { + "bbox": [ + 55, + 300, + 555, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 555, + 346 + ], + "type": "text", + "content": "Figure 10. ID-Patch pose-free generation. Here the condition image is only ID patches rendered onto a black canvas without any pose condition. The generation is conditioned on individual face locations while corresponding head sizes and body poses are inferred implicitly. Our method can accommodate a large number of people with diverse ethnic backgrounds, yet generating visually appealing group photo results. See \"Evaluation Metrics\" for reasons why we compare with baseline approaches only in pose-conditioned settings." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 107, + 350, + 200, + 486 + ], + "lines": [ + { + "bbox": [ + 107, + 350, + 200, + 486 + ], + "spans": [ + { + "bbox": [ + 107, + 350, + 200, + 486 + ], + "type": "image", + "image_path": "67b880b750a22302c775ef7630ce3d6dc21b52647a5f9ee8fa8bd549b14c1acf.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 112, + 488, + 194, + 497 + ], + "lines": [ + { + "bbox": [ + 112, + 488, + 194, + 497 + ], + "spans": [ + { + "bbox": [ + 112, + 488, + 194, + 497 + ], + "type": "text", + "content": "(b) Canny edge condition" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 507, + 295, + 541 + ], + "lines": [ + { + "bbox": [ + 55, + 507, + 295, + 541 + ], + "spans": [ + { + "bbox": [ + 55, + 507, + 295, + 541 + ], + "type": "text", + "content": "Figure 11. Combination of pose-free ID-Patch ControlNet and pretrained Canny edge ControlNet. We can generate background and clothing details, with high quality ID injection." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 201, + 350, + 294, + 486 + ], + "blocks": [ + { + "bbox": [ + 201, + 350, + 294, + 486 + ], + "lines": [ + { + "bbox": [ + 201, + 350, + 294, + 486 + ], + "spans": [ + { + "bbox": [ + 201, + 350, + 294, + 486 + ], + "type": "image", + "image_path": "cd76eb79b1d113e2e0821f7eba9c42e75589c1f7cb797af3760b44ff870d46e4.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 217, + 488, + 278, + 497 + ], + "lines": [ + { + "bbox": [ + 217, + 488, + 278, + 497 + ], + "spans": [ + { + "bbox": [ + 217, + 488, + 278, + 497 + ], + "type": "text", + "content": "(c) ID-Patch result" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 552, + 295, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 295, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 295, + 574 + ], + "type": "text", + "content": "of the same identities, showcasing varied expressions and lighting conditions, might help alleviate this problem." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 575, + 295, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 575, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 575, + 295, + 635 + ], + "type": "text", + "content": "Although the resemblance has been improved with our approach, there still exists a performance gap between generating a single person and multiple people (Fig. 8(a)). This means generating personalized group photo remains a challenging problem and requires further study." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 645, + 128, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 645, + 128, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 645, + 128, + 658 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "content": "In conclusion, our novel method, ID-Patch, markedly enhances identity resemblance and positioning in group photo generation. By embedding each identity feature within a distinct patch and utilizing ControlNet to accurately place" + } + ] + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 316, + 350, + 416, + 450 + ], + "blocks": [ + { + "bbox": [ + 316, + 350, + 416, + 450 + ], + "lines": [ + { + "bbox": [ + 316, + 350, + 416, + 450 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 416, + 450 + ], + "type": "image", + "image_path": "43b1cc7aa18c5f64ece6e1b58d165bd95396cd810575f3106d46a0c427374afa.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 453, + 403, + 462 + ], + "lines": [ + { + "bbox": [ + 329, + 453, + 403, + 462 + ], + "spans": [ + { + "bbox": [ + 329, + 453, + 403, + 462 + ], + "type": "text", + "content": "(a) Abnormal anatomy" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 417, + 350, + 451, + 451 + ], + "blocks": [ + { + "bbox": [ + 417, + 350, + 451, + 451 + ], + "lines": [ + { + "bbox": [ + 417, + 350, + 451, + 451 + ], + "spans": [ + { + "bbox": [ + 417, + 350, + 451, + 451 + ], + "type": "image", + "image_path": "9a097e59cbff5246aca59814e4a35c6d2ef07eac98f6733396221a3ec0728a30.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 450, + 453, + 518, + 463 + ], + "lines": [ + { + "bbox": [ + 450, + 453, + 518, + 463 + ], + "spans": [ + { + "bbox": [ + 450, + 453, + 518, + 463 + ], + "type": "text", + "content": "(b) Over-fitting input" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 464, + 554, + 508 + ], + "lines": [ + { + "bbox": [ + 313, + 464, + 554, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 554, + 508 + ], + "type": "text", + "content": "Figure 12. Limitations. (a) Abnormal anatomy can be found in complex scenes, e.g. the forearms in the red box. (b) When input faces contain strong shadows, our method struggles to fully disentangle user's ID in generated result." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 450, + 350, + 551, + 451 + ], + "blocks": [ + { + "bbox": [ + 450, + 350, + 551, + 451 + ], + "lines": [ + { + "bbox": [ + 450, + 350, + 551, + 451 + ], + "spans": [ + { + "bbox": [ + 450, + 350, + 551, + 451 + ], + "type": "image", + "image_path": "5a5123a3951aa4aded6e60d8afc6fc00f799a2dfa8faf1931d51b1ba1520fea0.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 520, + 555, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 520, + 555, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 520, + 555, + 652 + ], + "type": "text", + "content": "each identity at the designated spatial locations, we successfully reduce ID leakage. Our approach seamlessly integrates with additional conditioning signals such as pose control. Experimental results underscore the superiority of ID-Patch, particularly in complex scenarios involving more than three identities. This work paves the way for future explorations in multi-ID image generation. Potential future research directions include leveraging multiple images of the same individual from different angles to further enhance identity resemblance and simultaneously controlling location and facial expressions using the patch technique." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "content": "Societal Impact: Our work studies ID injection from a technical perspective and is not intended for malicious use. However, inappropriate usage might generate undesired fake images. Users should be aware of the ethical implications and use it responsibly." + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2993" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "text", + "content": "[1] Omri Avrahami, Kfir Aberman, Ohad Fried, Daniel Cohen-Or, and Dani Lischinski. Break-a-scene: Extracting multiple concepts from a single image. In SIGGRAPH Asia 2023 Conference Papers, pages 1–12, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 137, + 296, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 296, + 203 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 296, + 203 + ], + "type": "text", + "content": "[2] Omri Avrahami, Thomas Hayes, Oran Gafni, Sonal Gupta, Yaniv Taigman, Devi Parikh, Dani Lischinski, Ohad Fried, and Xi Yin. Spatext: Spatio-textual representation for controllable image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18370-18380, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 205, + 295, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 205, + 295, + 258 + ], + "spans": [ + { + "bbox": [ + 62, + 205, + 295, + 258 + ], + "type": "text", + "content": "[3] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Ji-aming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, et al. ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 261, + 295, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 261, + 295, + 316 + ], + "spans": [ + { + "bbox": [ + 62, + 261, + 295, + 316 + ], + "type": "text", + "content": "[4] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 318, + 295, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 318, + 295, + 371 + ], + "spans": [ + { + "bbox": [ + 62, + 318, + 295, + 371 + ], + "type": "text", + "content": "[5] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 374, + 294, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 374, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 62, + 374, + 294, + 407 + ], + "type": "text", + "content": "[6] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 409, + 295, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 409, + 295, + 473 + ], + "spans": [ + { + "bbox": [ + 62, + 409, + 295, + 473 + ], + "type": "text", + "content": "[7] Ganggui Ding, Canyu Zhao, Wen Wang, Zhen Yang, Zide Liu, Hao Chen, and Chunhua Shen. Freecustom: Tuning-free customized image generation for multi-concept composition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9089-9098, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 476, + 295, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 476, + 295, + 531 + ], + "spans": [ + { + "bbox": [ + 62, + 476, + 295, + 531 + ], + "type": "text", + "content": "[8] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis, march 2024. URL http://arxiv.org/abs/2403.03206. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 533, + 295, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 533, + 295, + 586 + ], + "spans": [ + { + "bbox": [ + 62, + 533, + 295, + 586 + ], + "type": "text", + "content": "[9] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 590, + 295, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 590, + 295, + 633 + ], + "spans": [ + { + "bbox": [ + 57, + 590, + 295, + 633 + ], + "type": "text", + "content": "[10] Rinon Gal, Moab Arar, Yuval Atzmon, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Encoder-based domain tuning for fast personalization of text-to-image models. ACM Transactions on Graphics (TOG), 42(4):1-13, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "type": "text", + "content": "[11] Zigang Geng, Ke Sun, Bin Xiao, Zhaoxiang Zhang, and Jingdong Wang. Bottom-up human pose estimation via disentangled keypoint regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14676-14686, 2021. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 692, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 295, + 714 + ], + "type": "text", + "content": "[12] Yuchao Gu, Xintao Wang, Jay Zhangjie Wu, Yujun Shi, Yunpeng Chen, Zihan Fan, Wuyou Xiao, Rui Zhao, Shuning" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 116 + ], + "type": "text", + "content": "Chang, Weijia Wu, et al. Mix-of-show: Decentralized low-rank adaptation for multi-concept customization of diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 171 + ], + "type": "text", + "content": "[13] Ligong Han, Yinxiao Li, Han Zhang, Peyman Milanfar, Dimitris Metaxas, and Feng Yang. Svdiff: Compact parameter space for diffusion fine-tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7323-7334, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 173, + 553, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 173, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 553, + 215 + ], + "type": "text", + "content": "[14] Junjie He, Yifeng Geng, and Liefeng Bo. Uniportrait: A unified framework for identity-preserving single- and multi-human image personalization. arXiv preprint arXiv:2408.05939, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 217, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 271 + ], + "type": "text", + "content": "[15] Zecheng He, Bo Sun, Felix Juefei-Xu, Haoyu Ma, Ankit Ramchandani, Vincent Cheung, Siddharth Shah, Anmol Kalia, Harihar Subramanyam, Alireza Zareian, et al. Imagine yourself: Tuning-free personalized image generation. arXiv preprint arXiv:2409.13346, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 272, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 553, + 304 + ], + "type": "text", + "content": "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 305, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 305, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 305, + 553, + 348 + ], + "type": "text", + "content": "[17] Chanran Kim, Jeongin Lee, Shichang Joung, Bongmo Kim, and Yeul-Min Baek. Instantfamily: Masked attention for zero-shot multi-id image generation. arXiv preprint arXiv:2404.19427, 2024. 1, 2, 3, 4, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 350, + 553, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 371 + ], + "type": "text", + "content": "[18] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 372, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 553, + 415 + ], + "type": "text", + "content": "[19] Zhe Kong, Yong Zhang, Tianyu Yang, Tao Wang, Kaihao Zhang, Bizhu Wu, Guanying Chen, Wei Liu, and Wenhan Luo. Omg: Occlusion-friendly personalized multi-concept generation in diffusion models. In ECCV, 2024. 1, 2, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 416, + 553, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 553, + 470 + ], + "type": "text", + "content": "[20] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 472, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 472, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 317, + 472, + 553, + 525 + ], + "type": "text", + "content": "[21] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 526, + 553, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 526, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 317, + 526, + 553, + 581 + ], + "type": "text", + "content": "[22] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22511-22521, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 582, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 582, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 317, + 582, + 553, + 635 + ], + "type": "text", + "content": "[23] Zhen Li, Mingdeng Cao, Xintao Wang, Zhongang Qi, MingMing Cheng, and Ying Shan. Photomaker: Customizing realistic human photos via stacked id embedding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8640-8650, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 636, + 553, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 636, + 553, + 680 + ], + "spans": [ + { + "bbox": [ + 317, + 636, + 553, + 680 + ], + "type": "text", + "content": "[24] Zhiheng Liu, Ruili Feng, Kai Zhu, Yifei Zhang, Kecheng Zheng, Yu Liu, Deli Zhao, Jingren Zhou, and Yang Cao. Cones: Concept neurons in diffusion models for customized generation. arXiv preprint arXiv:2303.05125, 2023. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[25] Nanye Ma, Mark Goldstein, Michael S Albergo, Nicholas M Boffi, Eric Vanden-Eijnden, and Saining Xie. Sit: Exploring flow and diffusion-based generative models with scalable" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2994" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "text", + "content": "interpolant transformers. arXiv preprint arXiv:2401.08740, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "type": "text", + "content": "[26] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 295, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 295, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 295, + 196 + ], + "type": "text", + "content": "[27] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 295, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 295, + 263 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 295, + 263 + ], + "type": "text", + "content": "[28] Xu Peng, Junwei Zhu, Boyuan Jiang, Ying Tai, Donghao Luo, Jiangning Zhang, Wei Lin, Taisong Jin, Chengjie Wang, and Rongrong Ji. Portraitbooth: A versatile portrait model for fast identity-preserved personalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27080-27090, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 265, + 295, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 265, + 295, + 320 + ], + "spans": [ + { + "bbox": [ + 56, + 265, + 295, + 320 + ], + "type": "text", + "content": "[29] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. In The Twelfth International Conference on Learning Representations, 2024. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 321, + 295, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 295, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 295, + 387 + ], + "type": "text", + "content": "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 388, + 295, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 295, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 295, + 443 + ], + "type": "text", + "content": "[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 445, + 295, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 295, + 510 + ], + "type": "text", + "content": "[32] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 512, + 295, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 295, + 578 + ], + "type": "text", + "content": "[33] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdreambooth: Hypernetworks for fast personalization of text-to-image models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6527-6536, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 579, + 295, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 295, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 295, + 644 + ], + "type": "text", + "content": "[34] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 647, + 295, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 295, + 690 + ], + "type": "text", + "content": "[35] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 815-823, 2015. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 692, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 295, + 713 + ], + "type": "text", + "content": "[36] Noam Shazeer and Mitchell Stern. Adafactor: Adaptive learning rates with sublinear memory cost. In International" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "text", + "content": "Conference on Machine Learning, pages 4596-4604. PMLR, 2018. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "text", + "content": "[37] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 152, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 185 + ], + "type": "text", + "content": "[38] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 186, + 553, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 230 + ], + "type": "text", + "content": "[39] Yang Sui, Huy Phan, Jinqi Xiao, Tianfang Zhang, Zijie Tang, Cong Shi, Yan Wang, Yingying Chen, and Bo Yuan. Disdet: Exploring detectability of backdoor attack on diffusion models. arXiv preprint arXiv:2402.02739, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 232, + 553, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 232, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 316, + 232, + 553, + 275 + ], + "type": "text", + "content": "[40] Yoad Tewel, Rinon Gal, Gal Chechik, and Yuval Atzmon. Key-locked rank one editing for text-to-image personalization. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 277, + 553, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 277, + 553, + 320 + ], + "spans": [ + { + "bbox": [ + 316, + 277, + 553, + 320 + ], + "type": "text", + "content": "[41] Hanhui Wang, Yihua Zhang, Ruizheng Bai, Yue Zhao, Sijia Liu, and Zhengzhong Tu. Edit away and my face will not stay: Personal biometric defense against malicious generative editing. arXiv preprint arXiv:2411.16832, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 322, + 553, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 322, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 316, + 322, + 553, + 365 + ], + "type": "text", + "content": "[42] Kuan-Chieh Wang, Daniil Ostashev, Yuwei Fang, Sergey Tulyakov, and Kfir Aberman. Moa: Mixture-of-attention for subject-context disentanglement in personalized image generation. arXiv preprint arXiv:2404.11565, 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 367, + 553, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 367, + 553, + 410 + ], + "spans": [ + { + "bbox": [ + 316, + 367, + 553, + 410 + ], + "type": "text", + "content": "[43] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, and Anthony Chen. Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519, 2024. 1, 2, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 412, + 553, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 412, + 553, + 466 + ], + "spans": [ + { + "bbox": [ + 316, + 412, + 553, + 466 + ], + "type": "text", + "content": "[44] Xudong Wang, Trevor Darrell, Sai Saketh Rambhatla, Rohit Girdhar, and Ishan Misra. Instancediffusion: Instance-level control for image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6232-6242, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 468, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 468, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 468, + 553, + 533 + ], + "type": "text", + "content": "[45] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15943-15953, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 535, + 553, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 535, + 553, + 589 + ], + "spans": [ + { + "bbox": [ + 316, + 535, + 553, + 589 + ], + "type": "text", + "content": "[46] Yuxin Wen, Neel Jain, John Kirchenbauer, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Hard prompts made easy: Gradient-based discrete optimization for prompt tuning and discovery. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 590, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 553, + 634 + ], + "type": "text", + "content": "[47] Guangxuan Xiao, Tianwei Yin, William T Freeman, Frédo Durand, and Song Han. Fastcomposer: Tuning-free multisubject image generation with localized attention. International Journal of Computer Vision, pages 1-20, 2024. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 635, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 690 + ], + "type": "text", + "content": "[48] Jinqi Xiao, Miao Yin, Yu Gong, Xiao Zang, Jian Ren, and Bo Yuan. COMCAT: Towards efficient compression and customization of attention-based vision models. In Proceedings of the 40th International Conference on Machine Learning, pages 38125-38136. PMLR, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[49] Jinqi Xiao, Shen Sang, Tiancheng Zhi, Jing Liu, Qing Yan, Linjie Luo, and Bo Yuan. Coap: Memory-efficient training" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "2995" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 497 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 76, + 73, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 296, + 95 + ], + "type": "text", + "content": "with correlation-aware gradient projection. arXiv preprint arXiv:2412.00071, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 139 + ], + "type": "text", + "content": "[50] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 2, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 295, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 295, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 295, + 184 + ], + "type": "text", + "content": "[51] Ge Yuan, Xiaodong Cun, Yong Zhang, Maomao Li, Chenyang Qi, Xintao Wang, Ying Shan, and Huicheng Zheng. Inserting anybody in diffusion models via celeb basis. arXiv preprint arXiv:2306.00926, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 295, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 295, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 295, + 229 + ], + "type": "text", + "content": "[52] Kaipeng Zhang, Zhanpeng Zhang, Zhifeng Li, and Yu Qiao. Joint face detection and alignment using multitask cascaded convolutional networks. IEEE signal processing letters, 23 (10):1499-1503, 2016. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 231, + 295, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 295, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 295, + 274 + ], + "type": "text", + "content": "[53] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 276, + 295, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 295, + 329 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 295, + 329 + ], + "type": "text", + "content": "[54] Yimeng Zhang, Xin Chen, Jinghan Jia, Yihua Zhang, Chongyu Fan, Jiancheng Liu, Mingyi Hong, Ke Ding, and Sijia Liu. Defensive unlearning with adversarial training for robust concept erasure in diffusion models. arXiv preprint arXiv:2405.15234, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 331, + 295, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 295, + 396 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 295, + 396 + ], + "type": "text", + "content": "[55] Yimeng Zhang, Jinghan Jia, Xin Chen, Aochuan Chen, Yihua Zhang, Jiancheng Liu, Ke Ding, and Sijia Liu. To generate or not? safety-driven unlearned diffusion models are still easy to generate unsafe images... for now. In European Conference on Computer Vision, pages 385-403. Springer, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 398, + 295, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 398, + 295, + 452 + ], + "spans": [ + { + "bbox": [ + 56, + 398, + 295, + 452 + ], + "type": "text", + "content": "[56] Yihua Zhang, Yimeng Zhang, Yuguang Yao, Jinghan Jia, Jiancheng Liu, Xiaoming Liu, and Sijia Liu. Unlearncanvas: A stylized image dataset to benchmark machine unlearning for diffusion models. arXiv preprint arXiv:2402.11846, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 454, + 295, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 454, + 295, + 497 + ], + "spans": [ + { + "bbox": [ + 56, + 454, + 295, + 497 + ], + "type": "text", + "content": "[57] Yufan Zhou, Ruiyi Zhang, Tong Sun, and Jinhui Xu. Enhancing detail preservation for customized text-to-image generation: A regularization-free approach. arXiv preprint arXiv:2305.13579, 2023. 2" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "2996" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/IDEA-Bench_ How Far are Generative Models from Professional Designing_/970924c3-54e6-4027-b252-42ee55ce3b43_content_list.json b/2025/IDEA-Bench_ How Far are Generative Models from Professional Designing_/970924c3-54e6-4027-b252-42ee55ce3b43_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d9d25814037a0bd51fbef0b3c0a2b77a6f48327b --- /dev/null +++ b/2025/IDEA-Bench_ How Far are Generative Models from Professional Designing_/970924c3-54e6-4027-b252-42ee55ce3b43_content_list.json @@ -0,0 +1,1526 @@ +[ + { + "type": "text", + "text": "IDEA-Bench: How Far are Generative Models from Professional Designing?", + "text_level": 1, + "bbox": [ + 112, + 130, + 883, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chen Liang $^{1,2}$", + "bbox": [ + 120, + 178, + 235, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhi-Fan $\\mathbf{W}\\mathbf{u}^{3}$", + "bbox": [ + 161, + 205, + 267, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lianghua Huang3", + "bbox": [ + 272, + 179, + 416, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yupeng Shi", + "bbox": [ + 315, + 205, + 415, + 223 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jingwu Fang", + "bbox": [ + 454, + 180, + 565, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Junge Zhang1,2", + "bbox": [ + 460, + 205, + 581, + 223 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Huanzhang Dou", + "bbox": [ + 604, + 180, + 741, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xin Zhao*", + "bbox": [ + 627, + 205, + 725, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wei Wang3", + "bbox": [ + 782, + 180, + 872, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yu Liu", + "bbox": [ + 764, + 205, + 833, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Institute of Automation, Chinese Academy of Sciences", + "bbox": [ + 277, + 239, + 720, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ School of Artificial Intelligence, University of Chinese Academy of Sciences", + "bbox": [ + 187, + 258, + 807, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Tongyi Lab, Alibaba Group", + "bbox": [ + 383, + 277, + 612, + 294 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4University of Science and Technology Beijing", + "bbox": [ + 313, + 295, + 683, + 311 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "5Zhejiang University", + "bbox": [ + 416, + 311, + 581, + 330 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 364, + 326, + 380 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advancements in image generation models enable the creation of high-quality images and targeted modifications based on textual instructions. Some models even support multimodal complex guidance and demonstrate robust task generalization capabilities. However, they still fall short of meeting the nuanced, professional demands of designers. To bridge this gap, we introduce IDEA-Bench, a comprehensive benchmark designed to advance image generation models toward applications with robust task generalization. IDEA-Bench comprises 100 professional image generation tasks and 275 specific cases, categorized into five major types based on the current capabilities of existing models. Furthermore, we provide a representative subset of 18 tasks with enhanced evaluation criteria to facilitate more nuanced and reliable evaluations using Multimodal Large Language Models (MLLMs). By assessing models' ability to comprehend and execute novel, complex tasks, IDEA-Bench paves the way toward the development of generative models with autonomous and versatile visual generation capabilities.", + "bbox": [ + 88, + 396, + 485, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 728, + 220, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advancements in Text-to-Image (T2I) models [3, 4, 7, 12, 13, 24, 36, 39–41, 44] have significantly enhanced the ability to generate high-quality images from textual descriptions. Building on these successes, models such as ControlNet [61] and T2I-Adapter [32] integrate additional networks into text-to-image diffusion frameworks to incorporate visual conditioning. Similarly, InstructPix2Pix [5] and Emu-Edit [45] are specifically trained on datasets tailored", + "bbox": [ + 89, + 753, + 483, + 876 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "for complex image editing tasks. Despite the widespread popularity of models like DALL-E 3 [40] and FLUX-1 [24], which attract millions of visits daily, professional users often rely on established workflows and specialized software for image creation. This reliance highlights a substantial gap between the capabilities of current image generation models and the demanding requirements of professional-grade image design. When existing image generation models focus on academic task research, the rapid evolution of large language models (LLMs) [6, 11, 34, 37, 38, 50, 51, 62] and multimodal language models (MLLMs) [28, 33, 49, 52] indicate that future image generation models will increasingly aim to achieve both professional-grade quality and task unification, enabling them to handle a diverse range of complex applications seamlessly.", + "bbox": [ + 511, + 364, + 906, + 592 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advances in image generation have demonstrated robust task unification and generalization through large-scale pre-training and arbitrary task learning [8, 16, 22, 46, 47, 53]. These models handle a wide range of academic tasks with a unified input-output format and can generalize to unseen tasks, showcasing in-context learning abilities. However, existing benchmarks [9, 17, 20, 23, 43, 45] are often narrow in scope, focusing primarily on isolated academic tasks and lacking comprehensive criteria to evaluate the multifaceted demands of professional image design. Consequently, there is a critical need for more robust and versatile evaluation frameworks to effectively assess the diverse and sophisticated capabilities of contemporary generative models.", + "bbox": [ + 511, + 595, + 908, + 805 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To bridge the gap between generative models and professional-grade design, we introduce IDEA-Bench (Intelligent Design Evaluation and Assessment Benchmark). We conduct a thorough review of real-world design and art tasks from diverse platforms, distilling 100 representative professional image generation tasks", + "bbox": [ + 511, + 809, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18541", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2a6d8d3483ddd97c88488c1e0488295dd43e9043aefa917882efdda10a5b2870.jpg", + "image_caption": [ + "Figure 1. Overview of IDEA-Bench. IDEA-Bench comprises 5 categories, encompassing a total of 100 professional-level subtasks, 275 cases, and 1,650 hierarchical evaluation questions. Each category provides subtask examples, quantitative statistics, and showcases a leaderboard of mainstream models." + ], + "image_footnote": [], + "bbox": [ + 133, + 89, + 867, + 375 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and 275 cases that comprehensively span the essential capabilities and effects for creating all forms of artwork. These tasks are systematically categorized into five distinct types based on required capabilities: text-to-image, image-to-image, images-to-image, text-to-images, and image(s)-to-images, as illustrated in Fig. 1. Currently, only a few models [8, 53, 55, 58] can handle the most intricate category, which demands MLLM-level understanding of both image and text and the ability to produce variable-length image sequences. To evaluate generated images, we leverage MLLM to transform the evaluation into an VQA task. This approach surpasses traditional metrics like FID [10] and CLIPScore [18], which fail to capture nuances in aesthetic quality, contextual relevance, and multimodal integration. IDEA-Bench includes six detailed evaluation questions per case, totaling 1,650 binary scoring items, and a representative subset of 18 tasks for more nuanced and reliable assessments, ensuring precise and objective evaluations aligned with professional design standards.", + "bbox": [ + 91, + 434, + 483, + 720 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows:", + "bbox": [ + 109, + 724, + 434, + 739 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce IDEA-Bench to bridge the gap between current generative model capabilities and the stringent demands of professional-grade image design, which consists of 100 carefully selected tasks.", + "- We develop a task categorization of five distinct categories based on complexity and modality requirements, with a detailed evaluation framework comprising 1,650 binary scoring items, enabling precise and objective assessment.", + "- We leverage MLLM-evaluation on 18 representative" + ], + "bbox": [ + 89, + 750, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tasks, demonstrating that MLLMs possess the capability for objective image assessment.", + "bbox": [ + 526, + 434, + 906, + 465 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 476, + 655, + 491 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Image Synthesis Models", + "text_level": 1, + "bbox": [ + 513, + 500, + 735, + 517 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, diffusion models such as Stable Diffusion [14, 36], DALL-E 3 [40], and FLUX-1 [25] have gained significant popularity for generating photorealistic images and offering improved training stability. However, state-of-the-art models still struggle with intricate prompts involving multiple visual concepts and extensive textual information, similar to image editing models [5, 45, 60], which find it challenging to interpret detailed commands requiring nuanced modifications for professional-grade tasks. To enhance generative models' capabilities, IDEA-Bench assesses existing image editing models against the rigorous demands of professional image design, aiming to emulate the creative and analytical processes of human designers. Additionally, image customization [15, 19, 42] is crucial for professionals needing varied personalized images. While techniques like Textual Inversion [15], LoRA [19], and DreamBooth [42] improve text-to-image models through fine-tuning, they lack adaptability to unseen subjects. IDEA-Bench aims at comprehensively evaluating models' customization capabilities across multi-image generation tasks, assessing consistency in content, style, identity, and conceptual aspects.", + "bbox": [ + 511, + 523, + 906, + 843 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Universal Generative Model", + "text_level": 1, + "bbox": [ + 511, + 848, + 767, + 863 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Universal generative models have become pivotal in advancing AI's ability to perform a diverse array of tasks", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18542", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/2afc00cbd263a33b0056a7d94d3802e37e8cd1ee081852b8cd1c36e3e1c6be89.jpg", + "table_caption": [ + "Table 1. Comparison with other image generation benchmarks. IDEA-Bench offers a broader range of task categories, multiple levels of model capabilities for evaluation, and longer and more complex prompts." + ], + "table_footnote": [], + "table_body": "
BenchmarkTasksEvaluation CategoryAvg. LengthMLLM Eval.
T2IT2IsI2IIs2II(s)2Is
PaintSkills [9]3✓XXXX10.44X
GenEval [17]6✓XXXX7.61X
EmuEdit [45]12XX✓XX8.57X
TIFA [20]12✓XXXX10.46GPT
DreamBooth [43]30XX✓XX7.58X
ImagenHub [23]7✓X✓XX8.61X
IDEA-Bench100✓✓✓✓✓138.68Gemini
", + "bbox": [ + 91, + 143, + 480, + 237 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "across both language and vision domains. LLMs [6, 11, 34, 37, 38, 50, 51, 62] have demonstrated remarkable versatility, excelling in tasks such as question answering and summarization through extensive pretraining on varied datasets. Building on this success, recent academic efforts [47, 48, 53, 57, 67] aim to adopt a more cohesive paradigm that addresses a wide range of visual generation tasks, including text-to-image and autoregressive image synthesis. Despite these advancements, there remains a significant gap in establishing a unified benchmark that comprehensively evaluates the multifaceted capabilities of these generative models. IDEA-Bench establishes a unified evaluation standard to assesses the strengths and limitations of existing generative foundation models, while guiding the development towards handling realistic and complex real-world image generation tasks at the same time.", + "bbox": [ + 93, + 243, + 480, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Benchmarks for Generative Models", + "text_level": 1, + "bbox": [ + 93, + 493, + 398, + 507 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The development of benchmarks typically keeps pace with advancements in text-to-image (T2I) synthesis [3, 4, 7, 12, 13, 24, 36, 39-41, 44]. DrawBench was initially introduced by Imagen [44], followed by DALL-EVAL [9], which proposed PaintSkills to evaluate visual reasoning and social bias capabilities. Recently, an increasing number of benchmarks have emerged [2, 21, 23, 26, 35]. While these benchmarks primarily focus on assessing image quality and alignment, DEsignBench [30] shares similarities with our approach by emphasizing scenarios within authentic design contexts. Differently, IDEA-Bench extends beyond this by addressing additional open challenges in professional creativity. Furthermore, IDEA-Bench is more closely aligned with the latest model capabilities, enabling the use of multi-image and complex instructions as diverse forms of guidance. As shown in Tab. 1, IDEA-Bench greatly expand the prompt length and adde two evaluation dimensions: multi image input and multi image generation. This allows for the generation of both single and multiple images, thereby better mimicking the workflows of professional designers.", + "bbox": [ + 93, + 515, + 480, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. IDEA-Bench", + "text_level": 1, + "bbox": [ + 93, + 830, + 220, + 844 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We present IDEA-Bench, a comprehensive image generation benchmark that encompasses a wide range of professional tasks in image generation and rigorously challenges", + "bbox": [ + 93, + 856, + 480, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "models across multiple dimensions of capability. In this section, we first introduce IDEA-Bench's data collection over four different model level in Sec. 3.1, as well as the manual annotation pipeline in Section Sec. 3.2, followed by a detailed explanation of both human evaluation and automated assessment methodologies, along with the specific metrics used, in Sec. 3.3.", + "bbox": [ + 516, + 90, + 903, + 195 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Data Collection", + "text_level": 1, + "bbox": [ + 516, + 209, + 666, + 223 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To ensure that the tasks evaluated in IDEA-Bench closely mirror real-world and professional scenarios, we source all task directives and data from the internet and professional designers. Leveraging the knowledge and capabilities of GPT-4o [33], we clearly define and classify the tasks, creating task variants aligned with existing model capabilities. The tasks in IDEA-Bench are categorized as follows:", + "bbox": [ + 516, + 231, + 903, + 335 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text-only Guided Image Generation While typical T2I users provide brief descriptions and rely on the model's discretion, professional users meticulously control every visual detail, often iterating multiple prompts to achieve the desired outcome. Current T2I models [3, 4, 7, 12, 13, 24, 36, 39-41, 44] struggle with complex prompts that include numerous visual concepts and lengthy text. To address this, IDEA-Bench includes 11 tasks designed to evaluate models' ability to handle extensive visual components and incorporate text within images, essential for tasks like poster generation and business card generation.", + "bbox": [ + 516, + 358, + 903, + 523 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image / Multi-image Guided Image Generation Unlike ImagenHub [23], which categorizes conditional image generation tasks under specific definitions, we unify these into image / multi-image guided image generation. This consolidation promotes the development of models with enhanced task unification and generalization, similar to advancements in LLMs. IDEA-Bench includes numerous design tasks guided by both image and text inputs in real-world contexts, such as brand merchandise generation, package rendering, and image retouching. With the advent of models like GDT [22], Emu2 [47], and OmniGen [57], we anticipate future models will exhibit strong multimodal understanding and generative capabilities.", + "bbox": [ + 516, + 546, + 903, + 742 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text & (Multi-)image Guided Multi-image Generation Distinct from previous benchmarks [9, 17, 20, 23, 43, 45], IDEA-Bench introduces tasks that require generating multiple related images simultaneously. These tasks cater to professional demands such as creating series images around a subject, multi-view image generation, and illustrating storybooks. Unlike video generation [59, 65], these image tasks involve significant visual differences while maintaining specific correlations. We further subdivide multi-image", + "bbox": [ + 516, + 763, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18543", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/41f2b111011b1cd15e697054f395155926c177b2d013f66e19b52e15a655dd50.jpg", + "image_caption": [ + "Figure 2. Dataset construction process of IDEA-Bench. We categorize the task data from professional design websites and designers based on generative model capabilities and assign capability keywords to each category. For each specific task, we design image generation prompts and hierarchical evaluation questions. Evaluators then refine these evaluation questions on a representative subset." + ], + "image_footnote": [], + "bbox": [ + 151, + 87, + 846, + 369 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "generation into text-guided and (multi-)image-guided categories to address models' limitations in handling diverse input modalities, thereby testing their ability to maintain detail, style, and content consistency.", + "bbox": [ + 89, + 431, + 483, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Data Annotation", + "text_level": 1, + "bbox": [ + 89, + 501, + 256, + 515 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Prompt Generation After collecting all tasks and corresponding images, annotators provide clear task definitions for each specific case, which, along with input images, are fed to GPT-4o [33] to generate image prompts tailored to each case. To ensure consistent evaluation across tasks, we also use GPT-4o to create a standardized set of six evaluation questions per task, ranging from basic to advanced difficulty levels. Specifically, recognizing the inherent subjectivity in both MLLM-based and human evaluations, we adopt a binary scoring system (0 or 1) for each question. This approach draws from methodologies in recent multimodal benchmarks [1, 27, 31, 64], where multiple-choice formats are used to mitigate interpretive bias.", + "bbox": [ + 88, + 523, + 482, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To further enhance objectivity and reduce variability, we modify the typical multi-level scoring scale found in existing datasets and models, converting it into a set of six binary (true/false) questions. Each question is accompanied by clearly defined scoring criteria, with 0 indicating unmet standards and 1 representing success. For deeper granularity, questions 1-2 assess the model's understanding of basic task requirements, questions 3-4 evaluate the quality of task completion, and questions 5-6 examine the model's attention to detail and aesthetic quality in the generated images. This structured assessment methodology ensures a rigorous and objective evaluation framework, aligning with", + "bbox": [ + 88, + 719, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "professional standards and reducing potential sources of instability in scoring.", + "bbox": [ + 511, + 431, + 906, + 463 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Prompt Refinement However, despite our efforts to define evaluation questions as objectively as possible, multimodal large language models (MLLMs) [28, 33, 49, 52] still struggle to match human evaluators on complex, real-world tasks, limited by current models' gaps in image comprehension relative to human capabilities. To address this, we select a subset of tasks and refine the evaluation questions based on specific examples, ensuring that MLLMs can yield intuitively reasonable results.", + "bbox": [ + 511, + 489, + 906, + 626 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "One challenge in using MLLMs as evaluators is the sensitivity of models like GPT-4V/GPT-4o to the order of image presentation [54, 56, 63, 66], which particularly makes these models less reliable for consistent comparative evaluation. MLLMs also fail to identify the sequence of multiple input images, whether it is input separately for multiple images or input as a collage. To mitigate this, we manually fine-tune evaluation questions for each case within the task subset, ensuring alignment with human intuition.", + "bbox": [ + 511, + 627, + 905, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For example, in the task of generating a children's storybook with defined character images, we refine each evaluation angle by randomly sampling pairs of input and output images to test consistency. A typical question might ask if the primary character in the first input image appears with consistent features in the third output image. For testing, we use the Gemini 1.5 Pro [49] to score model outputs and compared these results against human ratings. In cases where discrepancies arose, annotators iteratively refine the", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18544", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4d164516efce2cb7db21f9fa56239a46dc430819a5ebc91c7b1adb05cf62e940.jpg", + "image_caption": [ + "Figure 3. Visualization of model generated images on IDEA-Bench. We present the generation results in representative task examples, including text to image, image to image, and multi-image to image. In the case of no input image or no generated image, use $\\nabla$ instead." + ], + "image_footnote": [], + "bbox": [ + 109, + 89, + 888, + 424 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/267475ba11f132851f2c7f9b0772beab75e6309921a4efeb46b5cfd74b73e5a1.jpg", + "table_caption": [ + "Table 2. Experimental results on all categories of IDEA-Bench. Each task category is averaged across all its subtasks, with the top-ranked model scores for each task type highlighted in bold. Task types that a model cannot support are marked with “-” and are treated as 0 points in the average score calculation. “+” represents the use of MLLM for prompt rephrasing." + ], + "table_footnote": [], + "table_body": "
MethodParam.Scores on All CategoriesAvg. Score
T2II2IIs2IT2IsI(s)2Is
FLUX-1† [25]12B46.0612.134.8920.1529.1722.48
DALL-E 3† [40]12B24.346.955.2714.3614.4413.07
SD3† [14]2B24.0410.794.6921.5913.0614.83
Pixart† [7]0.6B14.447.753.4817.4621.3912.90
InstructPix2Pix [5]1B-17.58---3.52
MagicBrush [60]1B-19.07---3.81
Anole [8]7B0.000.640.001.740.000.48
Emu2 [47]37B17.987.058.98--6.81
OmniGen [57]3.8B21.418.172.77--6.47
Emu2† [47]37B17.987.058.9815.5312.7812.46
OmniGen† [57]3.8B21.418.172.7723.5221.3915.45
", + "bbox": [ + 94, + 549, + 483, + 695 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "language of the evaluation questions until the MLLM evaluations align with human judgment. This approach enhances the robustness of our benchmark, ensuring that evaluations reflect both task-specific accuracy and consistency across diverse visual requirements.", + "bbox": [ + 89, + 705, + 483, + 782 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Evaluation", + "text_level": 1, + "bbox": [ + 89, + 795, + 210, + 810 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.1. Human Evaluation", + "text_level": 1, + "bbox": [ + 89, + 819, + 269, + 833 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metric Each subtask in IDEA-Bench comprises 2 to 5 specific cases, with all cases within a subtask being evaluated using the same set of six binary (0 or 1) judgment questions. These evaluation questions are orga", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "nized into three hierarchical levels, reflecting the priority of task completion over mere aesthetic quality. For each case, scoring progresses sequentially through these levels. If a model does not achieve a perfect score at a lower level, the scores for all higher levels are automatically set to 0. This hierarchical approach ensures that while a model may generate high-quality and visually appealing images, it must also fully meet the task requirements to receive a high overall score. This methodology aligns the evaluation criteria with human design standards, emphasizing the importance of fulfilling task objectives.", + "bbox": [ + 511, + 467, + 906, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Formally, let the dataset be denoted as $\\mathcal{D} = \\{\\mathcal{D}_c = \\{\\mathcal{T}_t\\}_{t=1}^{T_c}\\}_{c=1}^C$ . Each subtask $\\mathcal{T}_t$ consists of 2 to 5 specific cases, all evaluated using the same set of six binary judgment questions. The score for each subtask is calculated by averaging the binary scores of its cases hierarchically. Subsequently, each major category $\\mathcal{D}_c$ receives its score by averaging the scores of its subtasks and converting the result to a percentage. The overall score for the entire dataset $\\mathcal{D}$ is then determined by averaging the scores of all $C$ major categories.", + "bbox": [ + 511, + 637, + 908, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Importantly, if a model fails to complete a specific task, it is assigned a score of 0 for that task, ensuring that incomplete performances are accurately reflected in the overall evaluation. This hierarchical scoring framework ensures a precise and objective assessment of model performance, closely aligning with the rigorous standards expected by professional designers.", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18545", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/0f84e40e79f2a4732f4560e1ef5e1bf284856127a1ecbb9da9b27a66ec90f7ba.jpg", + "table_caption": [ + "Table 3. Experimental results on Text-to-Image. Each task category is averaged across all its subtasks, with the top-ranked model scores for each subtask highlighted in bold." + ], + "table_footnote": [], + "table_body": "
MethodParam.Subtasks ScoreAvg. Score
Arch. StyleBus. CardGame UIInf. ChartInt.Paint.Sculp.TicketLand.LOGOPoster
FLUX-1 [25]12B100.0038.895.560.0066.6761.1116.6716.6783.3361.1156.6746.06
DALL-E 3 [40]12B22.220.000.000.0011.11100.0011.110.0038.8961.1123.3324.34
Stable Diffusion 3 [14]2B38.890.005.560.0050.005.5616.670.0050.0061.1136.6724.04
Pixart [7]0.6B5.560.000.000.0022.2233.3316.670.0022.2222.2236.6714.44
Anole [8]7B0.000.000.000.000.000.000.000.000.000.000.000.00
Emu2 [47]37B22.220.000.000.0038.8938.8916.670.0044.4416.6720.0017.98
OmniGen [57]3.8B38.890.000.000.0027.7850.0016.670.0038.8933.3330.0021.41
", + "bbox": [ + 151, + 117, + 841, + 213 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/66d640facf2efaeb75b10d287a8aeba5db42ab807859c74d6945a05658542590.jpg", + "table_caption": [ + "Table 4. Experimental results on Image-to-Image. Each task category is averaged across all its subtasks, with the top-ranked model scores for each subtask highlighted in bold. ${}^{\\prime \\prime }$ \" represent the use of MLLM for prompt rephrasing." + ], + "table_footnote": [], + "table_body": "
MethodParam.Subtasks ScoreAvg. Score
Anim. Attr.Coup. IconID PhotoBlurCompl.Ret.Str.Light. EditEnl.Obj. EditPack. Rend.Style EditBrand Merch.
FLUX-1† [25]12B0.0016.670.002.784.170.0012.500.00100.006.250.004.1711.1112.13
DALL-E 3‡ [40]12B5.5611.110.002.780.000.004.170.0038.894.178.334.1711.116.95
Stable Diffusion 3† [14]2B8.3327.780.002.784.170.008.330.0055.564.178.334.1716.6710.79
Pixart† [7]0.6B5.5655.564.172.784.170.000.000.0011.114.170.002.0811.117.75
InstructPix2Pix [5]1B44.4416.6716.6713.8925.0054.170.000.000.0010.4225.0016.675.5617.58
MagicBrush [60]1B33.3316.6712.5025.0025.0041.670.000.000.0035.428.3333.3316.6719.07
Anole [8]7B0.000.000.000.008.330.000.000.000.000.000.000.000.000.64
Emu2 [47]37B11.110.004.175.560.000.0020.830.0016.674.178.334.1716.677.05
OmniGen [57]3.8B16.670.004.175.560.004.174.170.000.0016.670.0010.4244.448.17
", + "bbox": [ + 91, + 248, + 903, + 358 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3.2. Automated Evaluation", + "text_level": 1, + "bbox": [ + 89, + 366, + 294, + 378 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Automated evaluation is performed on a subset of IDEA-Bench comprising 18 tasks, named IDEA-Bench-mini. During the prompt refinement phase, annotators craft unique evaluation questions for each case, in contrast to the human evaluation setup where all cases within a subtask share the same 6 evaluation questions. This tailored approach allows for more precise and context-specific assessments. Scoring is conducted using Gemini 1.5 Pro [49], adhering to the same hierarchical scoring methodology outlined in Sec. 3.3.1. To enhance result stability and reliability, each case is evaluated three times independently, and the final score is computed as the average of these three evaluations. This repetition mitigates potential variability in automated assessments and ensures consistent and robust scoring outcomes.", + "bbox": [ + 88, + 383, + 482, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "By integrating both human and automated evaluation, IDEA-Bench provides a comprehensive framework for assessing the multifaceted capabilities of generative models. This dual approach not only leverages the nuanced judgment of human evaluators but also benefits from the scalability and consistency of automated tools, thereby offering a balanced and thorough evaluation of model performance.", + "bbox": [ + 89, + 611, + 482, + 717 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 729, + 223, + 746 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Details", + "text_level": 1, + "bbox": [ + 89, + 753, + 290, + 768 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.1. Prompts Rephrasing", + "text_level": 1, + "bbox": [ + 89, + 776, + 282, + 791 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Currently, only a few models [8, 53, 55, 58] are capable of generating multiple related images simultaneously, makes it difficult to analyse a model's ability to generate multiple images potentially. However, as DALL-E 3 [40] could perform multi-image generation owing to its MLLM GPT-4o [33], which could summarize multimodal input and then generate the corresponding prompt for each image. To empower", + "bbox": [ + 89, + 794, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "other leading T2I models [7, 13, 25], we feed all input images and text prompt to GPT-4o and rephrase prompts for each image to be generated. By combining MLLM's multimodal understanding ability, we can approximately achieve the multi-image generation tasks with basic T2I models.", + "bbox": [ + 511, + 364, + 906, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.2. Model Reimplementation", + "text_level": 1, + "bbox": [ + 511, + 452, + 736, + 467 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate the capabilities of four categories of models in our study. For leading text-to-image (T2I) models, we select FLUX-1 [25], Stable Diffusion 3 (SD3) [13], Pixart [7], and DALL-E 3 [40]. For image editing models, we choose InstructPix2Pix [5] and MagicBrush [60]. To assess models with unified generation capabilities, we select Emu2 [47] and OmniGen [57]. For models that support interleaved text and image generation, we include Anole [8]. Notably, among the selected models, Anole [8] is the only one capable of handling unrestricted number of input and output images. We endeavor to utilize the officially released model parameters and inference settings wherever possible; further details are provided in the supplementary material.", + "bbox": [ + 511, + 472, + 906, + 669 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Results on IDEA-Bench", + "text_level": 1, + "bbox": [ + 511, + 681, + 733, + 696 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first present the ranking of all models based on their scores across the entire dataset, as detailed in Tab. 2. Leveraging GPT-4o's [33] multimodal understanding and task translation capabilities, FLUX-1 [25] achieve the highest score of 22.46, significantly outperforming other models. Specialized image editing models [5, 60] lead in the image-to-image tasks, while the general-purpose generation model Emu2 [47] excels in the images-to-image tasks. In contrast, Anole [8], which is capable of generating multiple images simultaneously, do not achieve an ideal final score. Furthermore, no model surpass a score of 10 without utilizing the comprehension capabilities of MLLMs, indicating that general-purpose image generation models still", + "bbox": [ + 511, + 703, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18546", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/077f6e123bc061338038760e2e427ab8d694f2743b59932a101225655ef83a74.jpg", + "table_caption": [ + "Table 5. Experimental results on Images-to-Image. Each task category is averaged across all its subtasks, with the top-ranked model scores for each subtask highlighted in bold. ${}^{\\prime \\prime } \\dagger$ \" represents the use of MLLM for prompt rephrasing." + ], + "table_footnote": [], + "table_body": "
MethodParam.Subtasks ScoreAvg. Score
Human Attr.Image Trans.Pan.Text Edit.3D. EffectImage Blend.Spec. Effect
FLUX-1† [25]3.976.670.0010.420.0012.500.005.564.89
DALL-E 3† [40]4.763.338.336.250.008.335.565.565.27
SD3† [14]4.765.000.008.330.008.335.565.564.69
Pixart† [7]4.765.008.330.000.004.170.005.563.48
Anole [8]0.000.000.000.000.000.000.000.000.00
Emu2 [47]6.350.0050.002.080.004.170.009.268.98
OmniGen [57]1.590.000.002.080.000.0011.117.412.77
", + "bbox": [ + 210, + 117, + 781, + 212 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7b1d3a7708790b2e018d0c5d724668fdde848056017da4218cb13f70483a7488.jpg", + "table_caption": [ + "Table 6. Experimental results on Text-to-Images. Each task category is averaged across all its subtasks, with the top-ranked model scores for each subtask highlighted in bold. ${}^{\\prime \\prime } \\dagger$ \" represents the use of MLLM for prompt rephrasing." + ], + "table_footnote": [], + "table_body": "
MethodParam.Subtasks ScoreAvg. Score
Anim. Grow.Child. BookDraw. Proc.Hist. Narr.Movie ShotsPhys. LawsPlant Grow.Style GroupConc. Vis.Coup. IconDyn. Char.
FLUX-1† [25]12B25.004.178.3370.834.174.178.337.0816.6750.0022.9220.15
DALL-E 3† [40]12B0.004.170.004.174.170.0041.6732.9212.5041.6716.6714.36
Stable Diffusion 3† [14]2B8.334.1725.0037.5012.508.3345.8318.7516.6741.6718.7521.59
Pixart† [7]0.6B37.504.174.1716.670.000.0029.1721.2516.6745.8316.6717.46
Anole [8]7B0.000.000.000.000.000.000.000.428.338.332.081.74
", + "bbox": [ + 91, + 246, + 906, + 319 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "face substantial challenges in achieving professional-grade performance. This underscores the need for comprehensive evaluation standards to guide the development of more specialized and capable generative models. The following section provides a detailed analysis of each category.", + "bbox": [ + 89, + 325, + 483, + 402 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Analysis Fig. 3 showcases the generated images across 7 models. For ease of comparison, we select several cases from three categories: text-to-image, image-to-image, and multi-image-to-image. In fundamental text-to-image tasks, FLUX-1 [25] demonstrates a clear advantage over other models, effectively comprehending the visual elements and professional-level requirements outlined in the prompts. Regarding image editing tasks, models that are not specifically designed for image editing struggle to preserve essential information from the original image. Emu2 [47] is the only model that roughly restores the shapes and compositions of objects from the original image. OmniGen [57] excels in maintaining the identity of objects. However, for more complex tasks such as special effect synthesis, no current model sufficiently understands the intricate demands of these professional-level requirements.", + "bbox": [ + 88, + 419, + 483, + 661 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.1. Results on Text-to-Image", + "text_level": 1, + "bbox": [ + 89, + 667, + 312, + 683 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "All models' results on text to image tasks are presented in Tab. 3. Notably, the FLUX-1 [25] model significantly outperforms the others, effectively accomplishing the task objectives in the majority of tasks. Based on the task settings and each model's performance, we observe the following:", + "bbox": [ + 89, + 686, + 482, + 762 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Text Generation Challenges: Text generation tasks are a common pain point for all models. Even though models can generate specific visual elements with high quality and understand style instructions specified in the prompts, they struggle to incorporate text into images as seamlessly as human designers. For example, in the business card generation task, only FLUX-1 [25] received scores.", + "- Performance of Basic vs. General Models: Basic text-to-image models generally demonstrate higher task" + ], + "bbox": [ + 89, + 763, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "completion rates compared to general models like Emu2 [47] and OmniGen [57]. This is expected, as general models prioritize enhancing task generalization and multimodal information processing capabilities, which inevitably compromises some aspects of image generation quality.", + "bbox": [ + 524, + 325, + 906, + 416 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "- Poor Performance on Information Charts: All models perform poorly on the information chart generation task. The primary reason is that generative models struggle to accurately transfer large amounts of text from the input to the output image, which is a basic ability to human designers. Additionally, many models do not support ultra-long text inputs.", + "bbox": [ + 513, + 417, + 906, + 523 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.2. Results on Image-to-Image & Images-to-Image", + "text_level": 1, + "bbox": [ + 511, + 531, + 885, + 546 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The image-to-image tasks represent the strengths of image editing models. In Tab. 4, MagicBrush [60] and Instruct-Pix2Pix [5] achieve the first and second highest evaluation scores, respectively. The images-to-image tasks involve more complex and diverse guidance or conditions, with Emu2 [47] attaining the highest score of 8.94, as shown in Tab. 5. From these results, we derive the following key insights:", + "bbox": [ + 511, + 550, + 905, + 671 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Pros and Cons of Image Editing Models: In local editing tasks such as image blur and image retouching, image editing models excel at preserving the original features of the image. However, these models are typically trained or fine-tuned on specific image editing datasets, which diminishes their ability to perform subject-driven tasks, such as branded merchandise generation.", + "- In-Context Learning Capabilities: Emu2 [47] and OmniGen [57] exhibit in-context learning abilities, as evidenced by OmniGen's significantly higher scores in tasks like branded merchandise generation and special effect synthesis, achieving ID consistency between input and output images. Emu2 demonstrates the ability to comprehend and partially complete the image straighten task, indicating that the model possesses an understanding of" + ], + "bbox": [ + 513, + 674, + 908, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18547", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/02b1d2754283fbbd04e281210646664c0d52058fae69d5ff77ffa88800f626f9.jpg", + "table_caption": [ + "Table 7. Experimental results on Image(s)-to-Images. Each task category is averaged across all its subtasks, with the top-ranked model scores for each subtask highlighted in bold. ${}^{\\prime \\prime }$ represents the use of MLLM for prompt rephrasing." + ], + "table_footnote": [], + "table_body": "
MethodParamsSubtasks ScoreAvg. Score
Anim. Grow.Creat. Trans.Layer Decomp.Light. EffectMovie ShotsMulti-app.Multi-dec.Multi-view Trans.
FLUX-1† [25]12B25.0066.6775.008.330.000.0016.67100.0029.17
DALL-E 3† [40]12B25.0025.008.330.004.170.0016.678.3314.44
Stable Diffusion 3† [14]2B33.3350.000.000.000.000.008.330.0013.06
Pixart† [7]0.6B33.3350.000.0058.330.000.0058.330.0021.39
Anole [8]7B0.000.000.000.000.000.000.000.000.00
", + "bbox": [ + 91, + 116, + 906, + 202 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6bfa36b7b4520f60674036ef8f31f8d0d9e2fa4b375b0f1de7e2a27aea7a2cf7.jpg", + "table_caption": [ + "Table 8. Experimental results on the subset IDEA-Bench-mini. \"G.\" represents evaluate using Gemini 1.5 pro [49], \"H.\" represents human evaluation. \"†\" represents the use of MLLM for prompt rephrasing." + ], + "table_footnote": [], + "table_body": "
MethodCLIPScoreEvaluation CategoryAvg. Score
T2II2IIs2IT2IsI(s)2Is
G.H.G.H.G.H.G.H.G.H.G.H.
FLUX-1† [25]0.343283.3367.0428.715.564.866.2532.2928.1337.5038.3337.4829.06
DALL-E 3† [40]0.346255.5641.1127.785.5620.375.5629.1721.8829.1733.3332.4121.49
SD3† [14]0.348356.3039.2623.618.345.555.5632.2923.9631.2529.8029.8019.76
Pixart† [7]0.343526.3027.0430.555.5629.862.0935.4223.9631.2530.0030.6817.73
InstructPix2Pix [5]0.1194--27.782.78------5.560.56
MagicBrush [60]0.1212--9.728.34------1.941.67
Emu2 [47]0.182135.8027.0443.528.3434.722.09----22.817.49
OmniGen [57]0.179946.3034.0740.2711.1125.460.00----22.416.81
", + "bbox": [ + 191, + 238, + 803, + 362 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the semantic aspects of visual elements. Recent advances in Multimodal Visualization-of-Thought (MVoT) [29] suggest a promising direction for enhancing generative models' ability to interpret complex instructions, potentially improving their generalization to professional design tasks.", + "bbox": [ + 102, + 375, + 482, + 464 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "- Challenges for Generative Models: General-purpose models perform poorly in tasks such as light condition editing. Although these tasks are frequently encountered by designers, the models require a certain level of understanding of real-world physical models, presenting a challenge for future generative models.", + "bbox": [ + 89, + 465, + 482, + 555 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2.3. Results on Text-to-Images & Image(s)-to-Images", + "text_level": 1, + "bbox": [ + 89, + 565, + 473, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Tab. 6 and Tab. 7 present the evaluation results for text-to-images and select image(s)-to-images tasks, respectively. In both categories, SD3 [13] and FLUX-1 [25] achieve the highest scores. For tasks such as historical event generation, MLLMs can effectively convey detailed style requirements during the prompt rephrasing phase, resulting in relatively consistent style outcomes. However, for more complex tasks like children's book generation that involve sophisticated capabilities such as consistency and preservation over ID and style, existing models struggle to perform adequately even when assisted by MLLMs. Nonetheless, generating multiple related images holds significant value; for example, the generated movie shots can serve as guiding conditions for multi-shot video generation models. We aim for the task types designed in our benchmark to align generative models with the capabilities of human designers, thereby further advancing toward general AI.", + "bbox": [ + 89, + 584, + 482, + 840 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2.4. Results on IDEA-Bench-mini", + "text_level": 1, + "bbox": [ + 89, + 851, + 339, + 864 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Tab. 8 presents the evaluation results for a subset of 18 tasks within our benchmark, where we calculate CLIPScore [18],", + "bbox": [ + 89, + 869, + 482, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Gemini 1.5 Pro [49] scores, and human evaluation scores. As shown in Tab. 8, FLUX-1 [25], leveraging MLLM's prompt rephrasing capabilities, achieve the highest scores in both MLLM-based and human evaluations within this subset. In contrast, when excluding the use of MLLMs, Emu2 [47] secure the top score. These experimental results are consistent with the overall benchmark dataset findings. Additionally, compared to CLIPScore, MLLM-based evaluations more closely align with human preferences in terms of professional evaluation data and assessment criteria. This suggests that as MLLM capabilities continue to evolve, utilizing MLLMs for image generation evaluation will become increasingly stable and reliable.", + "bbox": [ + 511, + 375, + 906, + 570 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 598, + 633, + 614 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce IDEA-Bench, a comprehensive benchmark designed to bridge the gap between current generative model capabilities and the stringent demands of professional-grade image design. IDEA-Bench encompasses 100 professional image generation tasks, utilizing a detailed evaluation framework to ensure precise and objective assessments. Additionally, we build a representative subset IDEA-Bench-mini that facilitates automated evaluation using MLLMs. Our evaluations reveal significant gaps in existing models, particularly in handling intricate instructions and maintaining consistency in multi-image generation, underscoring the need for more advanced and versatile models. By aligning evaluation standards with the nuanced requirements of human designers, IDEA-Bench serves as a pivotal tool for guiding the development of generative models toward achieving professional-grade performance and advancing toward autonomous and sophisticated visual generation capabilities.", + "bbox": [ + 511, + 628, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18548", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgments", + "text_level": 1, + "bbox": [ + 89, + 90, + 269, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is supported by the Youth Innovation Promotion Association CAS, the Beijing Natural Science Foundation (No. JQ24024), and the National Key Laboratory of Science and Technology on Space-Born Intelligent Information Processing Project (No. TJ-04-23-02).", + "bbox": [ + 89, + 113, + 483, + 185 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 210, + 187, + 227 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Shuai Bai, Shusheng Yang, Jinze Bai, Peng Wang, Xingxuan Zhang, Junyang Lin, Xinggang Wang, Chang Zhou, and Jingren Zhou. Touchstone: Evaluating vision-language models by language models. arXiv preprint arXiv:2308.16890, 2023. 4", + "[2] Eslam Mohamed Bakr, Pengzhan Sun, Xiaogian Shen, Faizan Farooq Khan, Li Erran Li, and Mohamed Elhoseiny. Hrs-bench: Holistic, reliable and scalable benchmark for text-to-image models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20041-20053, 2023. 3", + "[3] Jason Baldridge, Jakob Bauer, Mukul Bhutani, Nicole Brichtova, Andrew Bunner, Kelvin Chan, Yichang Chen, Sander Dieleman, Yuqing Du, Zach Eaton-Rosen, et al. Imagen 3. arXiv preprint arXiv:2408.07009, 2024. 1, 3", + "[4] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.openai.com/papers/dall-e-3.pdf, 2(3):8, 2023. 1, 3", + "[5] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 1, 2, 5, 6, 7, 8, 3", + "[6] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 1, 3", + "[7] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-alpha: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 1, 3, 5, 6, 7, 8", + "[8] Ethan Chern, Jiadi Su, Yan Ma, and Pengfei Liu. Anole: An open, autoregressive, native large multimodal models for interleaved image-text generation. arXiv preprint arXiv:2407.06135, 2024. 1, 2, 5, 6, 7, 8, 3", + "[9] Jaemin Cho, Abhay Zala, and Mohit Bansal. Dall-eval: Probing the reasoning skills and social biases of text-to-image generation models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3043-3054, 2023. 1, 3", + "[10] Min Jin Chong and David Forsyth. Effectively unbiased fid and inception score and where to find them. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6070-6079, 2020. 2", + "[11] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil" + ], + "bbox": [ + 93, + 236, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 1, 3", + "[12] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 1, 3", + "[13] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 1, 3, 6, 8", + "[14] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, Kyle Lacey, Alex Goodwin, Yannik Marek, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis, 2024. 2, 5, 6, 7, 8, 3", + "[15] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2", + "[16] Yuying Ge, Sijie Zhao, Ziyun Zeng, Yixiao Ge, Chen Li, Xintao Wang, and Ying Shan. Making llama see and draw with seed tokenizer. arXiv preprint arXiv:2310.01218, 2023. 1", + "[17] Dhruba Ghosh, Hannaneh Hajishirzi, and Ludwig Schmidt. Geneval: An object-focused framework for evaluating text-to-image alignment. Advances in Neural Information Processing Systems, 36, 2024. 1, 3", + "[18] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 2, 8", + "[19] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2", + "[20] Yushi Hu, Benlin Liu, Jungo Kasai, Yizhong Wang, Mari Ostendorf, Ranjay Krishna, and Noah A Smith. Tifa: Accurate and interpretable text-to-image faithfulness evaluation with question answering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20406-20417, 2023. 1, 3", + "[21] Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. Advances in Neural Information Processing Systems, 36:78723-78747, 2023. 3", + "[22] Lianghua Huang, Wei Wang, Zhi-Fan Wu, Huanzhang Dou, Yupeng Shi, Yutong Feng, Chen Liang, Yu Liu, and Jingren Zhou. Group diffusion transformers are unsupervised multitask learners. arXiv preprint arXiv:2410.15027, 2024. 1, 3, 2" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18549", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[23] Max Ku, Tianle Li, Kai Zhang, Yujie Lu, Xingyu Fu, Wenwen Zhuang, and Wenhu Chen. Imagenhub: Standardizing the evaluation of conditional image generation models. arXiv preprint arXiv:2310.01596, 2023. 1, 3", + "[24] Black Forest Labs. Flux: Inference repository. https://github.com/black-forest-labs/flux, 2024. Accessed: 2024-10-25. 1, 3", + "[25] Black Forest Labs. Flux.1, 2024. 2, 5, 6, 7, 8, 1, 3", + "[26] Tony Lee, Michihiro Yasunaga, Chenlin Meng, Yifan Mai, Joon Sung Park, Agrim Gupta, Yunzhi Zhang, Deepak Narayanan, Hannah Teufel, Marco Bellagente, et al. Holistic evaluation of text-to-image models. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[27] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023. 4", + "[28] Bo Li, Kaichen Zhang, Hao Zhang, Dong Guo, Renrui Zhang, Feng Li, Yuanhan Zhang, Ziwei Liu, and Chunyuan Li. Llava-last: Stronger llms supercharge multimodal capabilities in the wild, 2024. 1, 4", + "[29] Chengzhu Li, Wenshan Wu, Huanyu Zhang, Yan Xia, Shaoguang Mao, Li Dong, Ivan Vulic, and Furu Wei. Imagine while reasoning in space: Multimodal visualization-of-thought. arXiv preprint arXiv:2501.07542, 2025. 8", + "[30] Kevin Lin, Zhengyuan Yang, Linjie Li, Jianfeng Wang, and Lijuan Wang. Designbench: Exploring and benchmarking dall-e 3 for imagining visual design. arXiv preprint arXiv:2310.15144, 2023. 3", + "[31] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? In European Conference on Computer Vision, pages 216-233. Springer, 2025. 4", + "[32] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4296-4304, 2024. 1", + "[33] OpenAI. Gpt-4o, 2024. 1, 3, 4, 6, 2, 8", + "[34] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730-27744, 2022. 1, 3", + "[35] Vitali Petsiuk, Alexander E Siemann, Saisamrit Surbehera, Zad Chin, Keith Tyser, Gregory Hunter, Arvind Raghavan, Yann Hicke, Bryan A Plummer, Ori Kerret, et al. Human evaluation of text-to-image models on a multi-task benchmark. arXiv preprint arXiv:2211.12112, 2022. 3", + "[36] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 1, 2, 3" + ], + "bbox": [ + 91, + 90, + 483, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 1, 3", + "[38] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 1, 3", + "[39] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International conference on machine learning, pages 8821-8831. Pmlr, 2021. 1, 3", + "[40] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1, 2, 5, 6, 7, 8, 3", + "[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 3", + "[42] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv, 2022. 2", + "[43] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 1, 3", + "[44] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 1, 3", + "[45] Shelly Sheynin, Adam Polyak, Uriel Singer, Yuval Kirstain, Amit Zohar, Oron Ashual, Devi Parikh, and Yaniv Taigman. Emu edit: Precise image editing via recognition and generation tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8871-8879, 2024. 1, 2, 3", + "[46] Quan Sun, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, Yueze Wang, Hongcheng Gao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Emu: Generative pretraining in multimodality. In The Twelfth International Conference on Learning Representations, 2023. 1", + "[47] Quan Sun, Yufeng Cui, Xiaosong Zhang, Fan Zhang, Qiying Yu, Yueze Wang, Yongming Rao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative multimodal models are in-context learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14398-14409, 2024. 1, 3, 5, 6, 7, 8", + "[48] Chameleon Team. Chameleon: Mixed-modal early-fusion" + ], + "bbox": [ + 516, + 90, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18550", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "foundation models. arXiv preprint arXiv:2405.09818, 2024.3", + "[49] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 1, 4, 6, 8, 2, 3", + "[50] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1, 3", + "[51] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 1, 3", + "[52] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1, 4", + "[53] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 1, 2, 3, 6", + "[54] Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A Smith, Daniel Khashabi, and Hannaneh Hajishirzi. Self-instruct: Aligning language models with self-generated instructions. arXiv preprint arXiv:2212.10560, 2022. 4", + "[55] Shengqiong Wu, Hao Fei, Leigang Qu, Wei Ji, and Tat-Seng Chua. Next-gpt: Any-to-any multimodal llm. In Proceedings of the International Conference on Machine Learning, pages 53366-53397, 2024. 2, 6", + "[56] Tong Wu, Guandao Yang, Zhibing Li, Kai Zhang, Ziwei Liu, Leonidas Guibas, Dahua Lin, and Gordon Wetzstein. Gpt-4v (ision) is a human-aligned evaluator for text-to-3d generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22227-22238, 2024. 4", + "[57] Shitao Xiao, Yueze Wang, Junjie Zhou, Huaying Yuan, Xingrun Xing, Ruiran Yan, Shuting Wang, Tiejun Huang, and Zheng Liu. Omnigen: Unified image generation. arXiv preprint arXiv:2409.11340, 2024. 3, 5, 6, 7, 8, 1", + "[58] Shuai Yang, Yuying Ge, Yang Li, Yukang Chen, Yixiao Ge, Ying Shan, and Yingcong Chen. Seed-story: Multimodal long story generation with large language model. arXiv preprint arXiv:2407.08683, 2024. 2, 6", + "[59] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024.3", + "[60] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. Advances in Neural Information Processing Systems, 36, 2024. 2, 5, 6, 7, 8, 1, 3" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[61] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1", + "[62] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 1, 3", + "[63] Xinlu Zhang, Yujie Lu, Weizhi Wang, An Yan, Jun Yan, Lianke Qin, Heng Wang, Xifeng Yan, William Yang Wang, and Linda Ruth Petzold. Gpt-4v (ision) as a generalist evaluator for vision-language tasks. arXiv preprint arXiv:2311.01361, 2023. 4", + "[64] Yi-Fan Zhang, Huanyu Zhang, Haochen Tian, Chaoyou Fu, Shuangqing Zhang, Junfei Wu, Feng Li, Kun Wang, Qing-song Wen, Zhang Zhang, et al. Mme-realworld: Could your multimodal llm challenge high-resolution real-world scenarios that are difficult for humans? arXiv preprint arXiv:2408.13257, 2024. 4", + "[65] Xuanlei Zhao, Xiaolong Jin, Kai Wang, and Yang You. Real-time video generation with pyramid attention broadcast. arXiv preprint arXiv:2408.12588, 2024. 3", + "[66] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in Neural Information Processing Systems, 36:46595-46623, 2023. 4", + "[67] Chunting Zhou, Lili Yu, Arun Babu, Kushal Tirumala, Michihiro Yasunaga, Leonid Shamis, Jacob Kahn, Xuezhe Ma, Luke Zettlemoyer, and Omer Levy. Transfusion: Predict the next token and diffuse images with one multi-modal model, 2024. 3" + ], + "bbox": [ + 516, + 92, + 903, + 554 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "18551", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/IDEA-Bench_ How Far are Generative Models from Professional Designing_/970924c3-54e6-4027-b252-42ee55ce3b43_model.json b/2025/IDEA-Bench_ How Far are Generative Models from Professional Designing_/970924c3-54e6-4027-b252-42ee55ce3b43_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4f801003e3f77f702610a18bdc62242948606a60 --- /dev/null +++ b/2025/IDEA-Bench_ How Far are Generative Models from Professional Designing_/970924c3-54e6-4027-b252-42ee55ce3b43_model.json @@ -0,0 +1,2334 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.131, + 0.884, + 0.154 + ], + "angle": 0, + "content": "IDEA-Bench: How Far are Generative Models from Professional Designing?" + }, + { + "type": "text", + "bbox": [ + 0.122, + 0.179, + 0.236, + 0.2 + ], + "angle": 0, + "content": "Chen Liang\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.163, + 0.206, + 0.269, + 0.223 + ], + "angle": 0, + "content": "Zhi-Fan \\(\\mathbf{W}\\mathbf{u}^{3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.274, + 0.18, + 0.417, + 0.2 + ], + "angle": 0, + "content": "Lianghua Huang3" + }, + { + "type": "text", + "bbox": [ + 0.316, + 0.206, + 0.416, + 0.224 + ], + "angle": 0, + "content": "Yupeng Shi" + }, + { + "type": "text", + "bbox": [ + 0.455, + 0.181, + 0.566, + 0.199 + ], + "angle": 0, + "content": "Jingwu Fang" + }, + { + "type": "text", + "bbox": [ + 0.461, + 0.206, + 0.583, + 0.224 + ], + "angle": 0, + "content": "Junge Zhang1,2" + }, + { + "type": "text", + "bbox": [ + 0.606, + 0.181, + 0.743, + 0.199 + ], + "angle": 0, + "content": "Huanzhang Dou" + }, + { + "type": "text", + "bbox": [ + 0.629, + 0.206, + 0.726, + 0.222 + ], + "angle": 0, + "content": "Xin Zhao*" + }, + { + "type": "text", + "bbox": [ + 0.783, + 0.181, + 0.874, + 0.199 + ], + "angle": 0, + "content": "Wei Wang3" + }, + { + "type": "text", + "bbox": [ + 0.766, + 0.206, + 0.834, + 0.222 + ], + "angle": 0, + "content": "Yu Liu" + }, + { + "type": "text", + "bbox": [ + 0.278, + 0.241, + 0.721, + 0.259 + ], + "angle": 0, + "content": "\\(^{1}\\)Institute of Automation, Chinese Academy of Sciences" + }, + { + "type": "text", + "bbox": [ + 0.189, + 0.26, + 0.808, + 0.277 + ], + "angle": 0, + "content": "\\(^{2}\\)School of Artificial Intelligence, University of Chinese Academy of Sciences" + }, + { + "type": "text", + "bbox": [ + 0.385, + 0.278, + 0.613, + 0.295 + ], + "angle": 0, + "content": "\\(^{3}\\)Tongyi Lab, Alibaba Group" + }, + { + "type": "text", + "bbox": [ + 0.315, + 0.296, + 0.684, + 0.312 + ], + "angle": 0, + "content": "4University of Science and Technology Beijing" + }, + { + "type": "text", + "bbox": [ + 0.417, + 0.313, + 0.582, + 0.331 + ], + "angle": 0, + "content": "5Zhejiang University" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.365, + 0.327, + 0.381 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.397, + 0.486, + 0.701 + ], + "angle": 0, + "content": "Recent advancements in image generation models enable the creation of high-quality images and targeted modifications based on textual instructions. Some models even support multimodal complex guidance and demonstrate robust task generalization capabilities. However, they still fall short of meeting the nuanced, professional demands of designers. To bridge this gap, we introduce IDEA-Bench, a comprehensive benchmark designed to advance image generation models toward applications with robust task generalization. IDEA-Bench comprises 100 professional image generation tasks and 275 specific cases, categorized into five major types based on the current capabilities of existing models. Furthermore, we provide a representative subset of 18 tasks with enhanced evaluation criteria to facilitate more nuanced and reliable evaluations using Multimodal Large Language Models (MLLMs). By assessing models' ability to comprehend and execute novel, complex tasks, IDEA-Bench paves the way toward the development of generative models with autonomous and versatile visual generation capabilities." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.729, + 0.222, + 0.745 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.755, + 0.485, + 0.877 + ], + "angle": 0, + "content": "Recent advancements in Text-to-Image (T2I) models [3, 4, 7, 12, 13, 24, 36, 39–41, 44] have significantly enhanced the ability to generate high-quality images from textual descriptions. Building on these successes, models such as ControlNet [61] and T2I-Adapter [32] integrate additional networks into text-to-image diffusion frameworks to incorporate visual conditioning. Similarly, InstructPix2Pix [5] and Emu-Edit [45] are specifically trained on datasets tailored" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.366, + 0.908, + 0.593 + ], + "angle": 0, + "content": "for complex image editing tasks. Despite the widespread popularity of models like DALL-E 3 [40] and FLUX-1 [24], which attract millions of visits daily, professional users often rely on established workflows and specialized software for image creation. This reliance highlights a substantial gap between the capabilities of current image generation models and the demanding requirements of professional-grade image design. When existing image generation models focus on academic task research, the rapid evolution of large language models (LLMs) [6, 11, 34, 37, 38, 50, 51, 62] and multimodal language models (MLLMs) [28, 33, 49, 52] indicate that future image generation models will increasingly aim to achieve both professional-grade quality and task unification, enabling them to handle a diverse range of complex applications seamlessly." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.596, + 0.909, + 0.806 + ], + "angle": 0, + "content": "Recent advances in image generation have demonstrated robust task unification and generalization through large-scale pre-training and arbitrary task learning [8, 16, 22, 46, 47, 53]. These models handle a wide range of academic tasks with a unified input-output format and can generalize to unseen tasks, showcasing in-context learning abilities. However, existing benchmarks [9, 17, 20, 23, 43, 45] are often narrow in scope, focusing primarily on isolated academic tasks and lacking comprehensive criteria to evaluate the multifaceted demands of professional image design. Consequently, there is a critical need for more robust and versatile evaluation frameworks to effectively assess the diverse and sophisticated capabilities of contemporary generative models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.902 + ], + "angle": 0, + "content": "To bridge the gap between generative models and professional-grade design, we introduce IDEA-Bench (Intelligent Design Evaluation and Assessment Benchmark). We conduct a thorough review of real-world design and art tasks from diverse platforms, distilling 100 representative professional image generation tasks" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.236, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18541" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.09, + 0.868, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.379, + 0.908, + 0.421 + ], + "angle": 0, + "content": "Figure 1. Overview of IDEA-Bench. IDEA-Bench comprises 5 categories, encompassing a total of 100 professional-level subtasks, 275 cases, and 1,650 hierarchical evaluation questions. Each category provides subtask examples, quantitative statistics, and showcases a leaderboard of mainstream models." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.435, + 0.485, + 0.722 + ], + "angle": 0, + "content": "and 275 cases that comprehensively span the essential capabilities and effects for creating all forms of artwork. These tasks are systematically categorized into five distinct types based on required capabilities: text-to-image, image-to-image, images-to-image, text-to-images, and image(s)-to-images, as illustrated in Fig. 1. Currently, only a few models [8, 53, 55, 58] can handle the most intricate category, which demands MLLM-level understanding of both image and text and the ability to produce variable-length image sequences. To evaluate generated images, we leverage MLLM to transform the evaluation into an VQA task. This approach surpasses traditional metrics like FID [10] and CLIPScore [18], which fail to capture nuances in aesthetic quality, contextual relevance, and multimodal integration. IDEA-Bench includes six detailed evaluation questions per case, totaling 1,650 binary scoring items, and a representative subset of 18 tasks for more nuanced and reliable assessments, ensuring precise and objective evaluations aligned with professional design standards." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.725, + 0.435, + 0.74 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.75, + 0.483, + 0.81 + ], + "angle": 0, + "content": "- We introduce IDEA-Bench to bridge the gap between current generative model capabilities and the stringent demands of professional-grade image design, which consists of 100 carefully selected tasks." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.811, + 0.483, + 0.884 + ], + "angle": 0, + "content": "- We develop a task categorization of five distinct categories based on complexity and modality requirements, with a detailed evaluation framework comprising 1,650 binary scoring items, enabling precise and objective assessment." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "- We leverage MLLM-evaluation on 18 representative" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.75, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.435, + 0.907, + 0.466 + ], + "angle": 0, + "content": "tasks, demonstrating that MLLMs possess the capability for objective image assessment." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.477, + 0.656, + 0.492 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.501, + 0.736, + 0.518 + ], + "angle": 0, + "content": "2.1. Image Synthesis Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.907, + 0.844 + ], + "angle": 0, + "content": "Recently, diffusion models such as Stable Diffusion [14, 36], DALL-E 3 [40], and FLUX-1 [25] have gained significant popularity for generating photorealistic images and offering improved training stability. However, state-of-the-art models still struggle with intricate prompts involving multiple visual concepts and extensive textual information, similar to image editing models [5, 45, 60], which find it challenging to interpret detailed commands requiring nuanced modifications for professional-grade tasks. To enhance generative models' capabilities, IDEA-Bench assesses existing image editing models against the rigorous demands of professional image design, aiming to emulate the creative and analytical processes of human designers. Additionally, image customization [15, 19, 42] is crucial for professionals needing varied personalized images. While techniques like Textual Inversion [15], LoRA [19], and DreamBooth [42] improve text-to-image models through fine-tuning, they lack adaptability to unseen subjects. IDEA-Bench aims at comprehensively evaluating models' customization capabilities across multi-image generation tasks, assessing consistency in content, style, identity, and conceptual aspects." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.768, + 0.864 + ], + "angle": 0, + "content": "2.2. Universal Generative Model" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Universal generative models have become pivotal in advancing AI's ability to perform a diverse array of tasks" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18542" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.092, + 0.09, + 0.483, + 0.146 + ], + "angle": 0, + "content": "Table 1. Comparison with other image generation benchmarks. IDEA-Bench offers a broader range of task categories, multiple levels of model capabilities for evaluation, and longer and more complex prompts." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.145, + 0.482, + 0.238 + ], + "angle": 0, + "content": "
BenchmarkTasksEvaluation CategoryAvg. LengthMLLM Eval.
T2IT2IsI2IIs2II(s)2Is
PaintSkills [9]3✓XXXX10.44X
GenEval [17]6✓XXXX7.61X
EmuEdit [45]12XX✓XX8.57X
TIFA [20]12✓XXXX10.46GPT
DreamBooth [43]30XX✓XX7.58X
ImagenHub [23]7✓X✓XX8.61X
IDEA-Bench100✓✓✓✓✓138.68Gemini
" + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.244, + 0.482, + 0.485 + ], + "angle": 0, + "content": "across both language and vision domains. LLMs [6, 11, 34, 37, 38, 50, 51, 62] have demonstrated remarkable versatility, excelling in tasks such as question answering and summarization through extensive pretraining on varied datasets. Building on this success, recent academic efforts [47, 48, 53, 57, 67] aim to adopt a more cohesive paradigm that addresses a wide range of visual generation tasks, including text-to-image and autoregressive image synthesis. Despite these advancements, there remains a significant gap in establishing a unified benchmark that comprehensively evaluates the multifaceted capabilities of these generative models. IDEA-Bench establishes a unified evaluation standard to assesses the strengths and limitations of existing generative foundation models, while guiding the development towards handling realistic and complex real-world image generation tasks at the same time." + }, + { + "type": "title", + "bbox": [ + 0.094, + 0.494, + 0.4, + 0.508 + ], + "angle": 0, + "content": "2.3. Benchmarks for Generative Models" + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.516, + 0.482, + 0.817 + ], + "angle": 0, + "content": "The development of benchmarks typically keeps pace with advancements in text-to-image (T2I) synthesis [3, 4, 7, 12, 13, 24, 36, 39-41, 44]. DrawBench was initially introduced by Imagen [44], followed by DALL-EVAL [9], which proposed PaintSkills to evaluate visual reasoning and social bias capabilities. Recently, an increasing number of benchmarks have emerged [2, 21, 23, 26, 35]. While these benchmarks primarily focus on assessing image quality and alignment, DEsignBench [30] shares similarities with our approach by emphasizing scenarios within authentic design contexts. Differently, IDEA-Bench extends beyond this by addressing additional open challenges in professional creativity. Furthermore, IDEA-Bench is more closely aligned with the latest model capabilities, enabling the use of multi-image and complex instructions as diverse forms of guidance. As shown in Tab. 1, IDEA-Bench greatly expand the prompt length and adde two evaluation dimensions: multi image input and multi image generation. This allows for the generation of both single and multiple images, thereby better mimicking the workflows of professional designers." + }, + { + "type": "title", + "bbox": [ + 0.094, + 0.831, + 0.222, + 0.845 + ], + "angle": 0, + "content": "3. IDEA-Bench" + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.857, + 0.482, + 0.901 + ], + "angle": 0, + "content": "We present IDEA-Bench, a comprehensive image generation benchmark that encompasses a wide range of professional tasks in image generation and rigorously challenges" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.196 + ], + "angle": 0, + "content": "models across multiple dimensions of capability. In this section, we first introduce IDEA-Bench's data collection over four different model level in Sec. 3.1, as well as the manual annotation pipeline in Section Sec. 3.2, followed by a detailed explanation of both human evaluation and automated assessment methodologies, along with the specific metrics used, in Sec. 3.3." + }, + { + "type": "title", + "bbox": [ + 0.517, + 0.21, + 0.668, + 0.224 + ], + "angle": 0, + "content": "3.1. Data Collection" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.232, + 0.905, + 0.336 + ], + "angle": 0, + "content": "To ensure that the tasks evaluated in IDEA-Bench closely mirror real-world and professional scenarios, we source all task directives and data from the internet and professional designers. Leveraging the knowledge and capabilities of GPT-4o [33], we clearly define and classify the tasks, creating task variants aligned with existing model capabilities. The tasks in IDEA-Bench are categorized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.359, + 0.905, + 0.525 + ], + "angle": 0, + "content": "Text-only Guided Image Generation While typical T2I users provide brief descriptions and rely on the model's discretion, professional users meticulously control every visual detail, often iterating multiple prompts to achieve the desired outcome. Current T2I models [3, 4, 7, 12, 13, 24, 36, 39-41, 44] struggle with complex prompts that include numerous visual concepts and lengthy text. To address this, IDEA-Bench includes 11 tasks designed to evaluate models' ability to handle extensive visual components and incorporate text within images, essential for tasks like poster generation and business card generation." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.547, + 0.905, + 0.743 + ], + "angle": 0, + "content": "Image / Multi-image Guided Image Generation Unlike ImagenHub [23], which categorizes conditional image generation tasks under specific definitions, we unify these into image / multi-image guided image generation. This consolidation promotes the development of models with enhanced task unification and generalization, similar to advancements in LLMs. IDEA-Bench includes numerous design tasks guided by both image and text inputs in real-world contexts, such as brand merchandise generation, package rendering, and image retouching. With the advent of models like GDT [22], Emu2 [47], and OmniGen [57], we anticipate future models will exhibit strong multimodal understanding and generative capabilities." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.765, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Text & (Multi-)image Guided Multi-image Generation Distinct from previous benchmarks [9, 17, 20, 23, 43, 45], IDEA-Bench introduces tasks that require generating multiple related images simultaneously. These tasks cater to professional demands such as creating series images around a subject, multi-view image generation, and illustrating storybooks. Unlike video generation [59, 65], these image tasks involve significant visual differences while maintaining specific correlations. We further subdivide multi-image" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18543" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.152, + 0.088, + 0.848, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.377, + 0.908, + 0.421 + ], + "angle": 0, + "content": "Figure 2. Dataset construction process of IDEA-Bench. We categorize the task data from professional design websites and designers based on generative model capabilities and assign capability keywords to each category. For each specific task, we design image generation prompts and hierarchical evaluation questions. Evaluators then refine these evaluation questions on a representative subset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.433, + 0.484, + 0.495 + ], + "angle": 0, + "content": "generation into text-guided and (multi-)image-guided categories to address models' limitations in handling diverse input modalities, thereby testing their ability to maintain detail, style, and content consistency." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.502, + 0.258, + 0.516 + ], + "angle": 0, + "content": "3.2. Data Annotation" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.524, + 0.483, + 0.719 + ], + "angle": 0, + "content": "Prompt Generation After collecting all tasks and corresponding images, annotators provide clear task definitions for each specific case, which, along with input images, are fed to GPT-4o [33] to generate image prompts tailored to each case. To ensure consistent evaluation across tasks, we also use GPT-4o to create a standardized set of six evaluation questions per task, ranging from basic to advanced difficulty levels. Specifically, recognizing the inherent subjectivity in both MLLM-based and human evaluations, we adopt a binary scoring system (0 or 1) for each question. This approach draws from methodologies in recent multimodal benchmarks [1, 27, 31, 64], where multiple-choice formats are used to mitigate interpretive bias." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.72, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To further enhance objectivity and reduce variability, we modify the typical multi-level scoring scale found in existing datasets and models, converting it into a set of six binary (true/false) questions. Each question is accompanied by clearly defined scoring criteria, with 0 indicating unmet standards and 1 representing success. For deeper granularity, questions 1-2 assess the model's understanding of basic task requirements, questions 3-4 evaluate the quality of task completion, and questions 5-6 examine the model's attention to detail and aesthetic quality in the generated images. This structured assessment methodology ensures a rigorous and objective evaluation framework, aligning with" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.433, + 0.907, + 0.464 + ], + "angle": 0, + "content": "professional standards and reducing potential sources of instability in scoring." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.49, + 0.907, + 0.627 + ], + "angle": 0, + "content": "Prompt Refinement However, despite our efforts to define evaluation questions as objectively as possible, multimodal large language models (MLLMs) [28, 33, 49, 52] still struggle to match human evaluators on complex, real-world tasks, limited by current models' gaps in image comprehension relative to human capabilities. To address this, we select a subset of tasks and refine the evaluation questions based on specific examples, ensuring that MLLMs can yield intuitively reasonable results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.906, + 0.764 + ], + "angle": 0, + "content": "One challenge in using MLLMs as evaluators is the sensitivity of models like GPT-4V/GPT-4o to the order of image presentation [54, 56, 63, 66], which particularly makes these models less reliable for consistent comparative evaluation. MLLMs also fail to identify the sequence of multiple input images, whether it is input separately for multiple images or input as a collage. To mitigate this, we manually fine-tune evaluation questions for each case within the task subset, ensuring alignment with human intuition." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "For example, in the task of generating a children's storybook with defined character images, we refine each evaluation angle by randomly sampling pairs of input and output images to test consistency. A typical question might ask if the primary character in the first input image appears with consistent features in the third output image. For testing, we use the Gemini 1.5 Pro [49] to score model outputs and compared these results against human ratings. In cases where discrepancies arose, annotators iteratively refine the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18544" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.09, + 0.889, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.427, + 0.907, + 0.457 + ], + "angle": 0, + "content": "Figure 3. Visualization of model generated images on IDEA-Bench. We present the generation results in representative task examples, including text to image, image to image, and multi-image to image. In the case of no input image or no generated image, use \\(\\nabla\\) instead." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.466, + 0.483, + 0.549 + ], + "angle": 0, + "content": "Table 2. Experimental results on all categories of IDEA-Bench. Each task category is averaged across all its subtasks, with the top-ranked model scores for each task type highlighted in bold. Task types that a model cannot support are marked with “-” and are treated as 0 points in the average score calculation. “+” represents the use of MLLM for prompt rephrasing." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.55, + 0.484, + 0.696 + ], + "angle": 0, + "content": "
MethodParam.Scores on All CategoriesAvg. Score
T2II2IIs2IT2IsI(s)2Is
FLUX-1† [25]12B46.0612.134.8920.1529.1722.48
DALL-E 3† [40]12B24.346.955.2714.3614.4413.07
SD3† [14]2B24.0410.794.6921.5913.0614.83
Pixart† [7]0.6B14.447.753.4817.4621.3912.90
InstructPix2Pix [5]1B-17.58---3.52
MagicBrush [60]1B-19.07---3.81
Anole [8]7B0.000.640.001.740.000.48
Emu2 [47]37B17.987.058.98--6.81
OmniGen [57]3.8B21.418.172.77--6.47
Emu2† [47]37B17.987.058.9815.5312.7812.46
OmniGen† [57]3.8B21.418.172.7723.5221.3915.45
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.706, + 0.484, + 0.783 + ], + "angle": 0, + "content": "language of the evaluation questions until the MLLM evaluations align with human judgment. This approach enhances the robustness of our benchmark, ensuring that evaluations reflect both task-specific accuracy and consistency across diverse visual requirements." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.796, + 0.212, + 0.811 + ], + "angle": 0, + "content": "3.3. Evaluation" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.82, + 0.271, + 0.834 + ], + "angle": 0, + "content": "3.3.1. Human Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Evaluation Metric Each subtask in IDEA-Bench comprises 2 to 5 specific cases, with all cases within a subtask being evaluated using the same set of six binary (0 or 1) judgment questions. These evaluation questions are orga" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.468, + 0.907, + 0.635 + ], + "angle": 0, + "content": "nized into three hierarchical levels, reflecting the priority of task completion over mere aesthetic quality. For each case, scoring progresses sequentially through these levels. If a model does not achieve a perfect score at a lower level, the scores for all higher levels are automatically set to 0. This hierarchical approach ensures that while a model may generate high-quality and visually appealing images, it must also fully meet the task requirements to receive a high overall score. This methodology aligns the evaluation criteria with human design standards, emphasizing the importance of fulfilling task objectives." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.638, + 0.909, + 0.793 + ], + "angle": 0, + "content": "Formally, let the dataset be denoted as \\(\\mathcal{D} = \\{\\mathcal{D}_c = \\{\\mathcal{T}_t\\}_{t=1}^{T_c}\\}_{c=1}^C\\). Each subtask \\(\\mathcal{T}_t\\) consists of 2 to 5 specific cases, all evaluated using the same set of six binary judgment questions. The score for each subtask is calculated by averaging the binary scores of its cases hierarchically. Subsequently, each major category \\(\\mathcal{D}_c\\) receives its score by averaging the scores of its subtasks and converting the result to a percentage. The overall score for the entire dataset \\(\\mathcal{D}\\) is then determined by averaging the scores of all \\(C\\) major categories." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Importantly, if a model fails to complete a specific task, it is assigned a score of 0 for that task, ensuring that incomplete performances are accurately reflected in the overall evaluation. This hierarchical scoring framework ensures a precise and objective assessment of model performance, closely aligning with the rigorous standards expected by professional designers." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18545" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.09, + 0.908, + 0.117 + ], + "angle": 0, + "content": "Table 3. Experimental results on Text-to-Image. Each task category is averaged across all its subtasks, with the top-ranked model scores for each subtask highlighted in bold." + }, + { + "type": "table", + "bbox": [ + 0.152, + 0.118, + 0.842, + 0.214 + ], + "angle": 0, + "content": "
MethodParam.Subtasks ScoreAvg. Score
Arch. StyleBus. CardGame UIInf. ChartInt.Paint.Sculp.TicketLand.LOGOPoster
FLUX-1 [25]12B100.0038.895.560.0066.6761.1116.6716.6783.3361.1156.6746.06
DALL-E 3 [40]12B22.220.000.000.0011.11100.0011.110.0038.8961.1123.3324.34
Stable Diffusion 3 [14]2B38.890.005.560.0050.005.5616.670.0050.0061.1136.6724.04
Pixart [7]0.6B5.560.000.000.0022.2233.3316.670.0022.2222.2236.6714.44
Anole [8]7B0.000.000.000.000.000.000.000.000.000.000.000.00
Emu2 [47]37B22.220.000.000.0038.8938.8916.670.0044.4416.6720.0017.98
OmniGen [57]3.8B38.890.000.000.0027.7850.0016.670.0038.8933.3330.0021.41
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.222, + 0.907, + 0.248 + ], + "angle": 0, + "content": "Table 4. Experimental results on Image-to-Image. Each task category is averaged across all its subtasks, with the top-ranked model scores for each subtask highlighted in bold. \\( {}^{\\prime \\prime } \\) \" represent the use of MLLM for prompt rephrasing." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.249, + 0.905, + 0.359 + ], + "angle": 0, + "content": "
MethodParam.Subtasks ScoreAvg. Score
Anim. Attr.Coup. IconID PhotoBlurCompl.Ret.Str.Light. EditEnl.Obj. EditPack. Rend.Style EditBrand Merch.
FLUX-1† [25]12B0.0016.670.002.784.170.0012.500.00100.006.250.004.1711.1112.13
DALL-E 3‡ [40]12B5.5611.110.002.780.000.004.170.0038.894.178.334.1711.116.95
Stable Diffusion 3† [14]2B8.3327.780.002.784.170.008.330.0055.564.178.334.1716.6710.79
Pixart† [7]0.6B5.5655.564.172.784.170.000.000.0011.114.170.002.0811.117.75
InstructPix2Pix [5]1B44.4416.6716.6713.8925.0054.170.000.000.0010.4225.0016.675.5617.58
MagicBrush [60]1B33.3316.6712.5025.0025.0041.670.000.000.0035.428.3333.3316.6719.07
Anole [8]7B0.000.000.000.008.330.000.000.000.000.000.000.000.000.64
Emu2 [47]37B11.110.004.175.560.000.0020.830.0016.674.178.334.1716.677.05
OmniGen [57]3.8B16.670.004.175.560.004.174.170.000.0016.670.0010.4244.448.17
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.367, + 0.295, + 0.38 + ], + "angle": 0, + "content": "3.3.2. Automated Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.385, + 0.483, + 0.61 + ], + "angle": 0, + "content": "Automated evaluation is performed on a subset of IDEA-Bench comprising 18 tasks, named IDEA-Bench-mini. During the prompt refinement phase, annotators craft unique evaluation questions for each case, in contrast to the human evaluation setup where all cases within a subtask share the same 6 evaluation questions. This tailored approach allows for more precise and context-specific assessments. Scoring is conducted using Gemini 1.5 Pro [49], adhering to the same hierarchical scoring methodology outlined in Sec. 3.3.1. To enhance result stability and reliability, each case is evaluated three times independently, and the final score is computed as the average of these three evaluations. This repetition mitigates potential variability in automated assessments and ensures consistent and robust scoring outcomes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.612, + 0.483, + 0.718 + ], + "angle": 0, + "content": "By integrating both human and automated evaluation, IDEA-Bench provides a comprehensive framework for assessing the multifaceted capabilities of generative models. This dual approach not only leverages the nuanced judgment of human evaluators but also benefits from the scalability and consistency of automated tools, thereby offering a balanced and thorough evaluation of model performance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.73, + 0.224, + 0.747 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.754, + 0.291, + 0.77 + ], + "angle": 0, + "content": "4.1. Experimental Details" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.777, + 0.283, + 0.792 + ], + "angle": 0, + "content": "4.1.1. Prompts Rephrasing" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Currently, only a few models [8, 53, 55, 58] are capable of generating multiple related images simultaneously, makes it difficult to analyse a model's ability to generate multiple images potentially. However, as DALL-E 3 [40] could perform multi-image generation owing to its MLLM GPT-4o [33], which could summarize multimodal input and then generate the corresponding prompt for each image. To empower" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.366, + 0.907, + 0.442 + ], + "angle": 0, + "content": "other leading T2I models [7, 13, 25], we feed all input images and text prompt to GPT-4o and rephrase prompts for each image to be generated. By combining MLLM's multimodal understanding ability, we can approximately achieve the multi-image generation tasks with basic T2I models." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.453, + 0.738, + 0.468 + ], + "angle": 0, + "content": "4.1.2. Model Reimplementation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.473, + 0.907, + 0.67 + ], + "angle": 0, + "content": "We evaluate the capabilities of four categories of models in our study. For leading text-to-image (T2I) models, we select FLUX-1 [25], Stable Diffusion 3 (SD3) [13], Pixart [7], and DALL-E 3 [40]. For image editing models, we choose InstructPix2Pix [5] and MagicBrush [60]. To assess models with unified generation capabilities, we select Emu2 [47] and OmniGen [57]. For models that support interleaved text and image generation, we include Anole [8]. Notably, among the selected models, Anole [8] is the only one capable of handling unrestricted number of input and output images. We endeavor to utilize the officially released model parameters and inference settings wherever possible; further details are provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.682, + 0.734, + 0.697 + ], + "angle": 0, + "content": "4.2. Results on IDEA-Bench" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.909, + 0.901 + ], + "angle": 0, + "content": "We first present the ranking of all models based on their scores across the entire dataset, as detailed in Tab. 2. Leveraging GPT-4o's [33] multimodal understanding and task translation capabilities, FLUX-1 [25] achieve the highest score of 22.46, significantly outperforming other models. Specialized image editing models [5, 60] lead in the image-to-image tasks, while the general-purpose generation model Emu2 [47] excels in the images-to-image tasks. In contrast, Anole [8], which is capable of generating multiple images simultaneously, do not achieve an ideal final score. Furthermore, no model surpass a score of 10 without utilizing the comprehension capabilities of MLLMs, indicating that general-purpose image generation models still" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18546" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.117 + ], + "angle": 0, + "content": "Table 5. Experimental results on Images-to-Image. Each task category is averaged across all its subtasks, with the top-ranked model scores for each subtask highlighted in bold. \\( {}^{\\prime \\prime } \\dagger \\) \" represents the use of MLLM for prompt rephrasing." + }, + { + "type": "table", + "bbox": [ + 0.212, + 0.118, + 0.782, + 0.213 + ], + "angle": 0, + "content": "
MethodParam.Subtasks ScoreAvg. Score
Human Attr.Image Trans.Pan.Text Edit.3D. EffectImage Blend.Spec. Effect
FLUX-1† [25]3.976.670.0010.420.0012.500.005.564.89
DALL-E 3† [40]4.763.338.336.250.008.335.565.565.27
SD3† [14]4.765.000.008.330.008.335.565.564.69
Pixart† [7]4.765.008.330.000.004.170.005.563.48
Anole [8]0.000.000.000.000.000.000.000.000.00
Emu2 [47]6.350.0050.002.080.004.170.009.268.98
OmniGen [57]1.590.000.002.080.000.0011.117.412.77
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.219, + 0.908, + 0.247 + ], + "angle": 0, + "content": "Table 6. Experimental results on Text-to-Images. Each task category is averaged across all its subtasks, with the top-ranked model scores for each subtask highlighted in bold. \\( {}^{\\prime \\prime } \\dagger \\) \" represents the use of MLLM for prompt rephrasing." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.247, + 0.907, + 0.32 + ], + "angle": 0, + "content": "
MethodParam.Subtasks ScoreAvg. Score
Anim. Grow.Child. BookDraw. Proc.Hist. Narr.Movie ShotsPhys. LawsPlant Grow.Style GroupConc. Vis.Coup. IconDyn. Char.
FLUX-1† [25]12B25.004.178.3370.834.174.178.337.0816.6750.0022.9220.15
DALL-E 3† [40]12B0.004.170.004.174.170.0041.6732.9212.5041.6716.6714.36
Stable Diffusion 3† [14]2B8.334.1725.0037.5012.508.3345.8318.7516.6741.6718.7521.59
Pixart† [7]0.6B37.504.174.1716.670.000.0029.1721.2516.6745.8316.6717.46
Anole [8]7B0.000.000.000.000.000.000.000.428.338.332.081.74
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.327, + 0.484, + 0.404 + ], + "angle": 0, + "content": "face substantial challenges in achieving professional-grade performance. This underscores the need for comprehensive evaluation standards to guide the development of more specialized and capable generative models. The following section provides a detailed analysis of each category." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.42, + 0.484, + 0.662 + ], + "angle": 0, + "content": "Qualitative Analysis Fig. 3 showcases the generated images across 7 models. For ease of comparison, we select several cases from three categories: text-to-image, image-to-image, and multi-image-to-image. In fundamental text-to-image tasks, FLUX-1 [25] demonstrates a clear advantage over other models, effectively comprehending the visual elements and professional-level requirements outlined in the prompts. Regarding image editing tasks, models that are not specifically designed for image editing struggle to preserve essential information from the original image. Emu2 [47] is the only model that roughly restores the shapes and compositions of objects from the original image. OmniGen [57] excels in maintaining the identity of objects. However, for more complex tasks such as special effect synthesis, no current model sufficiently understands the intricate demands of these professional-level requirements." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.669, + 0.313, + 0.684 + ], + "angle": 0, + "content": "4.2.1. Results on Text-to-Image" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.688, + 0.483, + 0.763 + ], + "angle": 0, + "content": "All models' results on text to image tasks are presented in Tab. 3. Notably, the FLUX-1 [25] model significantly outperforms the others, effectively accomplishing the task objectives in the majority of tasks. Based on the task settings and each model's performance, we observe the following:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.765, + 0.483, + 0.87 + ], + "angle": 0, + "content": "- Text Generation Challenges: Text generation tasks are a common pain point for all models. Even though models can generate specific visual elements with high quality and understand style instructions specified in the prompts, they struggle to incorporate text into images as seamlessly as human designers. For example, in the business card generation task, only FLUX-1 [25] received scores." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "- Performance of Basic vs. General Models: Basic text-to-image models generally demonstrate higher task" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.765, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.327, + 0.907, + 0.417 + ], + "angle": 0, + "content": "completion rates compared to general models like Emu2 [47] and OmniGen [57]. This is expected, as general models prioritize enhancing task generalization and multimodal information processing capabilities, which inevitably compromises some aspects of image generation quality." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.418, + 0.907, + 0.525 + ], + "angle": 0, + "content": "- Poor Performance on Information Charts: All models perform poorly on the information chart generation task. The primary reason is that generative models struggle to accurately transfer large amounts of text from the input to the output image, which is a basic ability to human designers. Additionally, many models do not support ultra-long text inputs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.532, + 0.887, + 0.547 + ], + "angle": 0, + "content": "4.2.2. Results on Image-to-Image & Images-to-Image" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.551, + 0.906, + 0.672 + ], + "angle": 0, + "content": "The image-to-image tasks represent the strengths of image editing models. In Tab. 4, MagicBrush [60] and Instruct-Pix2Pix [5] achieve the first and second highest evaluation scores, respectively. The images-to-image tasks involve more complex and diverse guidance or conditions, with Emu2 [47] attaining the highest score of 8.94, as shown in Tab. 5. From these results, we derive the following key insights:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.675, + 0.907, + 0.78 + ], + "angle": 0, + "content": "- Pros and Cons of Image Editing Models: In local editing tasks such as image blur and image retouching, image editing models excel at preserving the original features of the image. However, these models are typically trained or fine-tuned on specific image editing datasets, which diminishes their ability to perform subject-driven tasks, such as branded merchandise generation." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.781, + 0.909, + 0.902 + ], + "angle": 0, + "content": "- In-Context Learning Capabilities: Emu2 [47] and OmniGen [57] exhibit in-context learning abilities, as evidenced by OmniGen's significantly higher scores in tasks like branded merchandise generation and special effect synthesis, achieving ID consistency between input and output images. Emu2 demonstrates the ability to comprehend and partially complete the image straighten task, indicating that the model possesses an understanding of" + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.675, + 0.909, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18547" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.118 + ], + "angle": 0, + "content": "Table 7. Experimental results on Image(s)-to-Images. Each task category is averaged across all its subtasks, with the top-ranked model scores for each subtask highlighted in bold. \\( {}^{\\prime \\prime } \\) represents the use of MLLM for prompt rephrasing." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.117, + 0.907, + 0.203 + ], + "angle": 0, + "content": "
MethodParamsSubtasks ScoreAvg. Score
Anim. Grow.Creat. Trans.Layer Decomp.Light. EffectMovie ShotsMulti-app.Multi-dec.Multi-view Trans.
FLUX-1† [25]12B25.0066.6775.008.330.000.0016.67100.0029.17
DALL-E 3† [40]12B25.0025.008.330.004.170.0016.678.3314.44
Stable Diffusion 3† [14]2B33.3350.000.000.000.000.008.330.0013.06
Pixart† [7]0.6B33.3350.000.0058.330.000.0058.330.0021.39
Anole [8]7B0.000.000.000.000.000.000.000.000.00
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.21, + 0.908, + 0.239 + ], + "angle": 0, + "content": "Table 8. Experimental results on the subset IDEA-Bench-mini. \"G.\" represents evaluate using Gemini 1.5 pro [49], \"H.\" represents human evaluation. \"†\" represents the use of MLLM for prompt rephrasing." + }, + { + "type": "table", + "bbox": [ + 0.192, + 0.239, + 0.805, + 0.363 + ], + "angle": 0, + "content": "
MethodCLIPScoreEvaluation CategoryAvg. Score
T2II2IIs2IT2IsI(s)2Is
G.H.G.H.G.H.G.H.G.H.G.H.
FLUX-1† [25]0.343283.3367.0428.715.564.866.2532.2928.1337.5038.3337.4829.06
DALL-E 3† [40]0.346255.5641.1127.785.5620.375.5629.1721.8829.1733.3332.4121.49
SD3† [14]0.348356.3039.2623.618.345.555.5632.2923.9631.2529.8029.8019.76
Pixart† [7]0.343526.3027.0430.555.5629.862.0935.4223.9631.2530.0030.6817.73
InstructPix2Pix [5]0.1194--27.782.78------5.560.56
MagicBrush [60]0.1212--9.728.34------1.941.67
Emu2 [47]0.182135.8027.0443.528.3434.722.09----22.817.49
OmniGen [57]0.179946.3034.0740.2711.1125.460.00----22.416.81
" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.375, + 0.483, + 0.465 + ], + "angle": 0, + "content": "the semantic aspects of visual elements. Recent advances in Multimodal Visualization-of-Thought (MVoT) [29] suggest a promising direction for enhancing generative models' ability to interpret complex instructions, potentially improving their generalization to professional design tasks." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.466, + 0.483, + 0.556 + ], + "angle": 0, + "content": "- Challenges for Generative Models: General-purpose models perform poorly in tasks such as light condition editing. Although these tasks are frequently encountered by designers, the models require a certain level of understanding of real-world physical models, presenting a challenge for future generative models." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.566, + 0.475, + 0.581 + ], + "angle": 0, + "content": "4.2.3. Results on Text-to-Images & Image(s)-to-Images" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.585, + 0.483, + 0.842 + ], + "angle": 0, + "content": "Tab. 6 and Tab. 7 present the evaluation results for text-to-images and select image(s)-to-images tasks, respectively. In both categories, SD3 [13] and FLUX-1 [25] achieve the highest scores. For tasks such as historical event generation, MLLMs can effectively convey detailed style requirements during the prompt rephrasing phase, resulting in relatively consistent style outcomes. However, for more complex tasks like children's book generation that involve sophisticated capabilities such as consistency and preservation over ID and style, existing models struggle to perform adequately even when assisted by MLLMs. Nonetheless, generating multiple related images holds significant value; for example, the generated movie shots can serve as guiding conditions for multi-shot video generation models. We aim for the task types designed in our benchmark to align generative models with the capabilities of human designers, thereby further advancing toward general AI." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.852, + 0.341, + 0.866 + ], + "angle": 0, + "content": "4.2.4. Results on IDEA-Bench-mini" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Tab. 8 presents the evaluation results for a subset of 18 tasks within our benchmark, where we calculate CLIPScore [18]," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.375, + 0.907, + 0.571 + ], + "angle": 0, + "content": "Gemini 1.5 Pro [49] scores, and human evaluation scores. As shown in Tab. 8, FLUX-1 [25], leveraging MLLM's prompt rephrasing capabilities, achieve the highest scores in both MLLM-based and human evaluations within this subset. In contrast, when excluding the use of MLLMs, Emu2 [47] secure the top score. These experimental results are consistent with the overall benchmark dataset findings. Additionally, compared to CLIPScore, MLLM-based evaluations more closely align with human preferences in terms of professional evaluation data and assessment criteria. This suggests that as MLLM capabilities continue to evolve, utilizing MLLMs for image generation evaluation will become increasingly stable and reliable." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.599, + 0.634, + 0.615 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We introduce IDEA-Bench, a comprehensive benchmark designed to bridge the gap between current generative model capabilities and the stringent demands of professional-grade image design. IDEA-Bench encompasses 100 professional image generation tasks, utilizing a detailed evaluation framework to ensure precise and objective assessments. Additionally, we build a representative subset IDEA-Bench-mini that facilitates automated evaluation using MLLMs. Our evaluations reveal significant gaps in existing models, particularly in handling intricate instructions and maintaining consistency in multi-image generation, underscoring the need for more advanced and versatile models. By aligning evaluation standards with the nuanced requirements of human designers, IDEA-Bench serves as a pivotal tool for guiding the development of generative models toward achieving professional-grade performance and advancing toward autonomous and sophisticated visual generation capabilities." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18548" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.27, + 0.108 + ], + "angle": 0, + "content": "6. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.484, + 0.186 + ], + "angle": 0, + "content": "This work is supported by the Youth Innovation Promotion Association CAS, the Beijing Natural Science Foundation (No. JQ24024), and the National Key Laboratory of Science and Technology on Space-Born Intelligent Information Processing Project (No. TJ-04-23-02)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.212, + 0.188, + 0.228 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.237, + 0.484, + 0.306 + ], + "angle": 0, + "content": "[1] Shuai Bai, Shusheng Yang, Jinze Bai, Peng Wang, Xingxuan Zhang, Junyang Lin, Xinggang Wang, Chang Zhou, and Jingren Zhou. Touchstone: Evaluating vision-language models by language models. arXiv preprint arXiv:2308.16890, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.308, + 0.484, + 0.39 + ], + "angle": 0, + "content": "[2] Eslam Mohamed Bakr, Pengzhan Sun, Xiaogian Shen, Faizan Farooq Khan, Li Erran Li, and Mohamed Elhoseiny. Hrs-bench: Holistic, reliable and scalable benchmark for text-to-image models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20041-20053, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.392, + 0.483, + 0.447 + ], + "angle": 0, + "content": "[3] Jason Baldridge, Jakob Bauer, Mukul Bhutani, Nicole Brichtova, Andrew Bunner, Kelvin Chan, Yichang Chen, Sander Dieleman, Yuqing Du, Zach Eaton-Rosen, et al. Imagen 3. arXiv preprint arXiv:2408.07009, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.449, + 0.482, + 0.517 + ], + "angle": 0, + "content": "[4] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.openai.com/papers/dall-e-3.pdf, 2(3):8, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.519, + 0.484, + 0.588 + ], + "angle": 0, + "content": "[5] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 1, 2, 5, 6, 7, 8, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.59, + 0.482, + 0.618 + ], + "angle": 0, + "content": "[6] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.62, + 0.483, + 0.688 + ], + "angle": 0, + "content": "[7] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-alpha: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 1, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.689, + 0.483, + 0.744 + ], + "angle": 0, + "content": "[8] Ethan Chern, Jiadi Su, Yan Ma, and Pengfei Liu. Anole: An open, autoregressive, native large multimodal models for interleaved image-text generation. arXiv preprint arXiv:2407.06135, 2024. 1, 2, 5, 6, 7, 8, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.746, + 0.483, + 0.814 + ], + "angle": 0, + "content": "[9] Jaemin Cho, Abhay Zala, and Mohit Bansal. Dall-eval: Probing the reasoning skills and social biases of text-to-image generation models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3043-3054, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.817, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[10] Min Jin Chong and David Forsyth. Effectively unbiased fid and inception score and where to find them. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6070-6079, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[11] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.237, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.192 + ], + "angle": 0, + "content": "[12] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.194, + 0.905, + 0.276 + ], + "angle": 0, + "content": "[13] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 1, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.905, + 0.374 + ], + "angle": 0, + "content": "[14] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, Kyle Lacey, Alex Goodwin, Yannik Marek, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis, 2024. 2, 5, 6, 7, 8, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.376, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[15] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.447, + 0.904, + 0.502 + ], + "angle": 0, + "content": "[16] Yuying Ge, Sijie Zhao, Ziyun Zeng, Yixiao Ge, Chen Li, Xintao Wang, and Ying Shan. Making llama see and draw with seed tokenizer. arXiv preprint arXiv:2310.01218, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.904, + 0.56 + ], + "angle": 0, + "content": "[17] Dhruba Ghosh, Hannaneh Hajishirzi, and Ludwig Schmidt. Geneval: An object-focused framework for evaluating text-to-image alignment. Advances in Neural Information Processing Systems, 36, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.562, + 0.905, + 0.616 + ], + "angle": 0, + "content": "[18] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.619, + 0.905, + 0.674 + ], + "angle": 0, + "content": "[19] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.676, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[20] Yushi Hu, Benlin Liu, Jungo Kasai, Yizhong Wang, Mari Ostendorf, Ranjay Krishna, and Noah A Smith. Tifa: Accurate and interpretable text-to-image faithfulness evaluation with question answering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20406-20417, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.761, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[21] Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. Advances in Neural Information Processing Systems, 36:78723-78747, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[22] Lianghua Huang, Wei Wang, Zhi-Fan Wu, Huanzhang Dou, Yupeng Shi, Yutong Feng, Chen Liang, Yu Liu, and Jingren Zhou. Group diffusion transformers are unsupervised multitask learners. arXiv preprint arXiv:2410.15027, 2024. 1, 3, 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18549" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.148 + ], + "angle": 0, + "content": "[23] Max Ku, Tianle Li, Kai Zhang, Yujie Lu, Xingyu Fu, Wenwen Zhuang, and Wenhu Chen. Imagenhub: Standardizing the evaluation of conditional image generation models. arXiv preprint arXiv:2310.01596, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.151, + 0.482, + 0.192 + ], + "angle": 0, + "content": "[24] Black Forest Labs. Flux: Inference repository. https://github.com/black-forest-labs/flux, 2024. Accessed: 2024-10-25. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.194, + 0.425, + 0.209 + ], + "angle": 0, + "content": "[25] Black Forest Labs. Flux.1, 2024. 2, 5, 6, 7, 8, 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.211, + 0.484, + 0.28 + ], + "angle": 0, + "content": "[26] Tony Lee, Michihiro Yasunaga, Chenlin Meng, Yifan Mai, Joon Sung Park, Agrim Gupta, Yunzhi Zhang, Deepak Narayanan, Hannah Teufel, Marco Bellagente, et al. Holistic evaluation of text-to-image models. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.282, + 0.482, + 0.337 + ], + "angle": 0, + "content": "[27] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.34, + 0.482, + 0.396 + ], + "angle": 0, + "content": "[28] Bo Li, Kaichen Zhang, Hao Zhang, Dong Guo, Renrui Zhang, Feng Li, Yuanhan Zhang, Ziwei Liu, and Chunyuan Li. Llava-last: Stronger llms supercharge multimodal capabilities in the wild, 2024. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.399, + 0.482, + 0.454 + ], + "angle": 0, + "content": "[29] Chengzhu Li, Wenshan Wu, Huanyu Zhang, Yan Xia, Shaoguang Mao, Li Dong, Ivan Vulic, and Furu Wei. Imagine while reasoning in space: Multimodal visualization-of-thought. arXiv preprint arXiv:2501.07542, 2025. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.457, + 0.482, + 0.512 + ], + "angle": 0, + "content": "[30] Kevin Lin, Zhengyuan Yang, Linjie Li, Jianfeng Wang, and Lijuan Wang. Designbench: Exploring and benchmarking dall-e 3 for imagining visual design. arXiv preprint arXiv:2310.15144, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.514, + 0.482, + 0.584 + ], + "angle": 0, + "content": "[31] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? In European Conference on Computer Vision, pages 216-233. Springer, 2025. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.586, + 0.482, + 0.655 + ], + "angle": 0, + "content": "[32] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4296-4304, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.657, + 0.353, + 0.672 + ], + "angle": 0, + "content": "[33] OpenAI. Gpt-4o, 2024. 1, 3, 4, 6, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.674, + 0.482, + 0.757 + ], + "angle": 0, + "content": "[34] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730-27744, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.759, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[35] Vitali Petsiuk, Alexander E Siemann, Saisamrit Surbehera, Zad Chin, Keith Tyser, Gregory Hunter, Arvind Raghavan, Yann Hicke, Bryan A Plummer, Ori Kerret, et al. Human evaluation of text-to-image models on a multi-task benchmark. arXiv preprint arXiv:2211.12112, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[36] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 1, 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.146 + ], + "angle": 0, + "content": "[37] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[38] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.288 + ], + "angle": 0, + "content": "[39] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International conference on machine learning, pages 8821-8831. Pmlr, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[40] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1, 2, 5, 6, 7, 8, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.907, + 0.418 + ], + "angle": 0, + "content": "[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.905, + 0.475 + ], + "angle": 0, + "content": "[42] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.477, + 0.905, + 0.559 + ], + "angle": 0, + "content": "[43] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.562, + 0.905, + 0.645 + ], + "angle": 0, + "content": "[44] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.647, + 0.905, + 0.729 + ], + "angle": 0, + "content": "[45] Shelly Sheynin, Adam Polyak, Uriel Singer, Yuval Kirstain, Amit Zohar, Oron Ashual, Devi Parikh, and Yaniv Taigman. Emu edit: Precise image editing via recognition and generation tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8871-8879, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.731, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[46] Quan Sun, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, Yueze Wang, Hongcheng Gao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Emu: Generative pretraining in multimodality. In The Twelfth International Conference on Learning Representations, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.885 + ], + "angle": 0, + "content": "[47] Quan Sun, Yufeng Cui, Xiaosong Zhang, Fan Zhang, Qiying Yu, Yueze Wang, Yongming Rao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative multimodal models are in-context learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14398-14409, 2024. 1, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.887, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[48] Chameleon Team. Chameleon: Mixed-modal early-fusion" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18550" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "foundation models. arXiv preprint arXiv:2405.09818, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.189 + ], + "angle": 0, + "content": "[49] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 1, 4, 6, 8, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.191, + 0.482, + 0.259 + ], + "angle": 0, + "content": "[50] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.261, + 0.482, + 0.328 + ], + "angle": 0, + "content": "[51] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.482, + 0.397 + ], + "angle": 0, + "content": "[52] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.482, + 0.466 + ], + "angle": 0, + "content": "[53] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.47, + 0.482, + 0.523 + ], + "angle": 0, + "content": "[54] Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A Smith, Daniel Khashabi, and Hannaneh Hajishirzi. Self-instruct: Aligning language models with self-generated instructions. arXiv preprint arXiv:2212.10560, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.525, + 0.482, + 0.579 + ], + "angle": 0, + "content": "[55] Shengqiong Wu, Hao Fei, Leigang Qu, Wei Ji, and Tat-Seng Chua. Next-gpt: Any-to-any multimodal llm. In Proceedings of the International Conference on Machine Learning, pages 53366-53397, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.482, + 0.661 + ], + "angle": 0, + "content": "[56] Tong Wu, Guandao Yang, Zhibing Li, Kai Zhang, Ziwei Liu, Leonidas Guibas, Dahua Lin, and Gordon Wetzstein. Gpt-4v (ision) is a human-aligned evaluator for text-to-3d generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22227-22238, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.665, + 0.482, + 0.719 + ], + "angle": 0, + "content": "[57] Shitao Xiao, Yueze Wang, Junjie Zhou, Huaying Yuan, Xingrun Xing, Ruiran Yan, Shuting Wang, Tiejun Huang, and Zheng Liu. Omnigen: Unified image generation. arXiv preprint arXiv:2409.11340, 2024. 3, 5, 6, 7, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.721, + 0.482, + 0.774 + ], + "angle": 0, + "content": "[58] Shuai Yang, Yuying Ge, Yang Li, Yukang Chen, Yixiao Ge, Ying Shan, and Yingcong Chen. Seed-story: Multimodal long story generation with large language model. arXiv preprint arXiv:2407.08683, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[59] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[60] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. Advances in Neural Information Processing Systems, 36, 2024. 2, 5, 6, 7, 8, 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[61] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.217 + ], + "angle": 0, + "content": "[62] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.905, + 0.288 + ], + "angle": 0, + "content": "[63] Xinlu Zhang, Yujie Lu, Weizhi Wang, An Yan, Jun Yan, Lianke Qin, Heng Wang, Xifeng Yan, William Yang Wang, and Linda Ruth Petzold. Gpt-4v (ision) as a generalist evaluator for vision-language tasks. arXiv preprint arXiv:2311.01361, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.905, + 0.371 + ], + "angle": 0, + "content": "[64] Yi-Fan Zhang, Huanyu Zhang, Haochen Tian, Chaoyou Fu, Shuangqing Zhang, Junfei Wu, Feng Li, Kun Wang, Qing-song Wen, Zhang Zhang, et al. Mme-realworld: Could your multimodal llm challenge high-resolution real-world scenarios that are difficult for humans? arXiv preprint arXiv:2408.13257, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.375, + 0.904, + 0.415 + ], + "angle": 0, + "content": "[65] Xuanlei Zhao, Xiaolong Jin, Kai Wang, and Yang You. Real-time video generation with pyramid attention broadcast. arXiv preprint arXiv:2408.12588, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.418, + 0.905, + 0.485 + ], + "angle": 0, + "content": "[66] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in Neural Information Processing Systems, 36:46595-46623, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.488, + 0.905, + 0.555 + ], + "angle": 0, + "content": "[67] Chunting Zhou, Lili Yu, Arun Babu, Kushal Tirumala, Michihiro Yasunaga, Leonid Shamis, Jacob Kahn, Xuezhe Ma, Luke Zettlemoyer, and Omer Levy. Transfusion: Predict the next token and diffuse images with one multi-modal model, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18551" + } + ] +] \ No newline at end of file diff --git a/2025/IDEA-Bench_ How Far are Generative Models from Professional Designing_/970924c3-54e6-4027-b252-42ee55ce3b43_origin.pdf b/2025/IDEA-Bench_ How Far are Generative Models from Professional Designing_/970924c3-54e6-4027-b252-42ee55ce3b43_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2a11b10f4685ac18a18be1bc6bd66272f85d3b43 --- /dev/null +++ b/2025/IDEA-Bench_ How Far are Generative Models from Professional Designing_/970924c3-54e6-4027-b252-42ee55ce3b43_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f4790b76c9554c7d9c86341b8649dd1da863bd7852a86c0f1dd9bbd58ccb971a +size 9568528 diff --git a/2025/IDEA-Bench_ How Far are Generative Models from Professional Designing_/full.md b/2025/IDEA-Bench_ How Far are Generative Models from Professional Designing_/full.md new file mode 100644 index 0000000000000000000000000000000000000000..672a8d896fa51aebbcbd4df6f237e7aa22684b73 --- /dev/null +++ b/2025/IDEA-Bench_ How Far are Generative Models from Professional Designing_/full.md @@ -0,0 +1,303 @@ +# IDEA-Bench: How Far are Generative Models from Professional Designing? + +Chen Liang $^{1,2}$ + +Zhi-Fan $\mathbf{W}\mathbf{u}^{3}$ + +Lianghua Huang3 + +Yupeng Shi + +Jingwu Fang + +Junge Zhang1,2 + +Huanzhang Dou + +Xin Zhao* + +Wei Wang3 + +Yu Liu + +$^{1}$ Institute of Automation, Chinese Academy of Sciences + +$^{2}$ School of Artificial Intelligence, University of Chinese Academy of Sciences + +$^{3}$ Tongyi Lab, Alibaba Group + +4University of Science and Technology Beijing + +5Zhejiang University + +# Abstract + +Recent advancements in image generation models enable the creation of high-quality images and targeted modifications based on textual instructions. Some models even support multimodal complex guidance and demonstrate robust task generalization capabilities. However, they still fall short of meeting the nuanced, professional demands of designers. To bridge this gap, we introduce IDEA-Bench, a comprehensive benchmark designed to advance image generation models toward applications with robust task generalization. IDEA-Bench comprises 100 professional image generation tasks and 275 specific cases, categorized into five major types based on the current capabilities of existing models. Furthermore, we provide a representative subset of 18 tasks with enhanced evaluation criteria to facilitate more nuanced and reliable evaluations using Multimodal Large Language Models (MLLMs). By assessing models' ability to comprehend and execute novel, complex tasks, IDEA-Bench paves the way toward the development of generative models with autonomous and versatile visual generation capabilities. + +# 1. Introduction + +Recent advancements in Text-to-Image (T2I) models [3, 4, 7, 12, 13, 24, 36, 39–41, 44] have significantly enhanced the ability to generate high-quality images from textual descriptions. Building on these successes, models such as ControlNet [61] and T2I-Adapter [32] integrate additional networks into text-to-image diffusion frameworks to incorporate visual conditioning. Similarly, InstructPix2Pix [5] and Emu-Edit [45] are specifically trained on datasets tailored + +for complex image editing tasks. Despite the widespread popularity of models like DALL-E 3 [40] and FLUX-1 [24], which attract millions of visits daily, professional users often rely on established workflows and specialized software for image creation. This reliance highlights a substantial gap between the capabilities of current image generation models and the demanding requirements of professional-grade image design. When existing image generation models focus on academic task research, the rapid evolution of large language models (LLMs) [6, 11, 34, 37, 38, 50, 51, 62] and multimodal language models (MLLMs) [28, 33, 49, 52] indicate that future image generation models will increasingly aim to achieve both professional-grade quality and task unification, enabling them to handle a diverse range of complex applications seamlessly. + +Recent advances in image generation have demonstrated robust task unification and generalization through large-scale pre-training and arbitrary task learning [8, 16, 22, 46, 47, 53]. These models handle a wide range of academic tasks with a unified input-output format and can generalize to unseen tasks, showcasing in-context learning abilities. However, existing benchmarks [9, 17, 20, 23, 43, 45] are often narrow in scope, focusing primarily on isolated academic tasks and lacking comprehensive criteria to evaluate the multifaceted demands of professional image design. Consequently, there is a critical need for more robust and versatile evaluation frameworks to effectively assess the diverse and sophisticated capabilities of contemporary generative models. + +To bridge the gap between generative models and professional-grade design, we introduce IDEA-Bench (Intelligent Design Evaluation and Assessment Benchmark). We conduct a thorough review of real-world design and art tasks from diverse platforms, distilling 100 representative professional image generation tasks + +![](images/2a6d8d3483ddd97c88488c1e0488295dd43e9043aefa917882efdda10a5b2870.jpg) +Figure 1. Overview of IDEA-Bench. IDEA-Bench comprises 5 categories, encompassing a total of 100 professional-level subtasks, 275 cases, and 1,650 hierarchical evaluation questions. Each category provides subtask examples, quantitative statistics, and showcases a leaderboard of mainstream models. + +and 275 cases that comprehensively span the essential capabilities and effects for creating all forms of artwork. These tasks are systematically categorized into five distinct types based on required capabilities: text-to-image, image-to-image, images-to-image, text-to-images, and image(s)-to-images, as illustrated in Fig. 1. Currently, only a few models [8, 53, 55, 58] can handle the most intricate category, which demands MLLM-level understanding of both image and text and the ability to produce variable-length image sequences. To evaluate generated images, we leverage MLLM to transform the evaluation into an VQA task. This approach surpasses traditional metrics like FID [10] and CLIPScore [18], which fail to capture nuances in aesthetic quality, contextual relevance, and multimodal integration. IDEA-Bench includes six detailed evaluation questions per case, totaling 1,650 binary scoring items, and a representative subset of 18 tasks for more nuanced and reliable assessments, ensuring precise and objective evaluations aligned with professional design standards. + +Our contributions can be summarized as follows: + +- We introduce IDEA-Bench to bridge the gap between current generative model capabilities and the stringent demands of professional-grade image design, which consists of 100 carefully selected tasks. +- We develop a task categorization of five distinct categories based on complexity and modality requirements, with a detailed evaluation framework comprising 1,650 binary scoring items, enabling precise and objective assessment. +- We leverage MLLM-evaluation on 18 representative + +tasks, demonstrating that MLLMs possess the capability for objective image assessment. + +# 2. Related Work + +# 2.1. Image Synthesis Models + +Recently, diffusion models such as Stable Diffusion [14, 36], DALL-E 3 [40], and FLUX-1 [25] have gained significant popularity for generating photorealistic images and offering improved training stability. However, state-of-the-art models still struggle with intricate prompts involving multiple visual concepts and extensive textual information, similar to image editing models [5, 45, 60], which find it challenging to interpret detailed commands requiring nuanced modifications for professional-grade tasks. To enhance generative models' capabilities, IDEA-Bench assesses existing image editing models against the rigorous demands of professional image design, aiming to emulate the creative and analytical processes of human designers. Additionally, image customization [15, 19, 42] is crucial for professionals needing varied personalized images. While techniques like Textual Inversion [15], LoRA [19], and DreamBooth [42] improve text-to-image models through fine-tuning, they lack adaptability to unseen subjects. IDEA-Bench aims at comprehensively evaluating models' customization capabilities across multi-image generation tasks, assessing consistency in content, style, identity, and conceptual aspects. + +# 2.2. Universal Generative Model + +Universal generative models have become pivotal in advancing AI's ability to perform a diverse array of tasks + +Table 1. Comparison with other image generation benchmarks. IDEA-Bench offers a broader range of task categories, multiple levels of model capabilities for evaluation, and longer and more complex prompts. + +
BenchmarkTasksEvaluation CategoryAvg. LengthMLLM Eval.
T2IT2IsI2IIs2II(s)2Is
PaintSkills [9]3✓XXXX10.44X
GenEval [17]6✓XXXX7.61X
EmuEdit [45]12XX✓XX8.57X
TIFA [20]12✓XXXX10.46GPT
DreamBooth [43]30XX✓XX7.58X
ImagenHub [23]7✓X✓XX8.61X
IDEA-Bench100✓✓✓✓✓138.68Gemini
+ +across both language and vision domains. LLMs [6, 11, 34, 37, 38, 50, 51, 62] have demonstrated remarkable versatility, excelling in tasks such as question answering and summarization through extensive pretraining on varied datasets. Building on this success, recent academic efforts [47, 48, 53, 57, 67] aim to adopt a more cohesive paradigm that addresses a wide range of visual generation tasks, including text-to-image and autoregressive image synthesis. Despite these advancements, there remains a significant gap in establishing a unified benchmark that comprehensively evaluates the multifaceted capabilities of these generative models. IDEA-Bench establishes a unified evaluation standard to assesses the strengths and limitations of existing generative foundation models, while guiding the development towards handling realistic and complex real-world image generation tasks at the same time. + +# 2.3. Benchmarks for Generative Models + +The development of benchmarks typically keeps pace with advancements in text-to-image (T2I) synthesis [3, 4, 7, 12, 13, 24, 36, 39-41, 44]. DrawBench was initially introduced by Imagen [44], followed by DALL-EVAL [9], which proposed PaintSkills to evaluate visual reasoning and social bias capabilities. Recently, an increasing number of benchmarks have emerged [2, 21, 23, 26, 35]. While these benchmarks primarily focus on assessing image quality and alignment, DEsignBench [30] shares similarities with our approach by emphasizing scenarios within authentic design contexts. Differently, IDEA-Bench extends beyond this by addressing additional open challenges in professional creativity. Furthermore, IDEA-Bench is more closely aligned with the latest model capabilities, enabling the use of multi-image and complex instructions as diverse forms of guidance. As shown in Tab. 1, IDEA-Bench greatly expand the prompt length and adde two evaluation dimensions: multi image input and multi image generation. This allows for the generation of both single and multiple images, thereby better mimicking the workflows of professional designers. + +# 3. IDEA-Bench + +We present IDEA-Bench, a comprehensive image generation benchmark that encompasses a wide range of professional tasks in image generation and rigorously challenges + +models across multiple dimensions of capability. In this section, we first introduce IDEA-Bench's data collection over four different model level in Sec. 3.1, as well as the manual annotation pipeline in Section Sec. 3.2, followed by a detailed explanation of both human evaluation and automated assessment methodologies, along with the specific metrics used, in Sec. 3.3. + +# 3.1. Data Collection + +To ensure that the tasks evaluated in IDEA-Bench closely mirror real-world and professional scenarios, we source all task directives and data from the internet and professional designers. Leveraging the knowledge and capabilities of GPT-4o [33], we clearly define and classify the tasks, creating task variants aligned with existing model capabilities. The tasks in IDEA-Bench are categorized as follows: + +Text-only Guided Image Generation While typical T2I users provide brief descriptions and rely on the model's discretion, professional users meticulously control every visual detail, often iterating multiple prompts to achieve the desired outcome. Current T2I models [3, 4, 7, 12, 13, 24, 36, 39-41, 44] struggle with complex prompts that include numerous visual concepts and lengthy text. To address this, IDEA-Bench includes 11 tasks designed to evaluate models' ability to handle extensive visual components and incorporate text within images, essential for tasks like poster generation and business card generation. + +Image / Multi-image Guided Image Generation Unlike ImagenHub [23], which categorizes conditional image generation tasks under specific definitions, we unify these into image / multi-image guided image generation. This consolidation promotes the development of models with enhanced task unification and generalization, similar to advancements in LLMs. IDEA-Bench includes numerous design tasks guided by both image and text inputs in real-world contexts, such as brand merchandise generation, package rendering, and image retouching. With the advent of models like GDT [22], Emu2 [47], and OmniGen [57], we anticipate future models will exhibit strong multimodal understanding and generative capabilities. + +Text & (Multi-)image Guided Multi-image Generation Distinct from previous benchmarks [9, 17, 20, 23, 43, 45], IDEA-Bench introduces tasks that require generating multiple related images simultaneously. These tasks cater to professional demands such as creating series images around a subject, multi-view image generation, and illustrating storybooks. Unlike video generation [59, 65], these image tasks involve significant visual differences while maintaining specific correlations. We further subdivide multi-image + +![](images/41f2b111011b1cd15e697054f395155926c177b2d013f66e19b52e15a655dd50.jpg) +Figure 2. Dataset construction process of IDEA-Bench. We categorize the task data from professional design websites and designers based on generative model capabilities and assign capability keywords to each category. For each specific task, we design image generation prompts and hierarchical evaluation questions. Evaluators then refine these evaluation questions on a representative subset. + +generation into text-guided and (multi-)image-guided categories to address models' limitations in handling diverse input modalities, thereby testing their ability to maintain detail, style, and content consistency. + +# 3.2. Data Annotation + +Prompt Generation After collecting all tasks and corresponding images, annotators provide clear task definitions for each specific case, which, along with input images, are fed to GPT-4o [33] to generate image prompts tailored to each case. To ensure consistent evaluation across tasks, we also use GPT-4o to create a standardized set of six evaluation questions per task, ranging from basic to advanced difficulty levels. Specifically, recognizing the inherent subjectivity in both MLLM-based and human evaluations, we adopt a binary scoring system (0 or 1) for each question. This approach draws from methodologies in recent multimodal benchmarks [1, 27, 31, 64], where multiple-choice formats are used to mitigate interpretive bias. + +To further enhance objectivity and reduce variability, we modify the typical multi-level scoring scale found in existing datasets and models, converting it into a set of six binary (true/false) questions. Each question is accompanied by clearly defined scoring criteria, with 0 indicating unmet standards and 1 representing success. For deeper granularity, questions 1-2 assess the model's understanding of basic task requirements, questions 3-4 evaluate the quality of task completion, and questions 5-6 examine the model's attention to detail and aesthetic quality in the generated images. This structured assessment methodology ensures a rigorous and objective evaluation framework, aligning with + +professional standards and reducing potential sources of instability in scoring. + +Prompt Refinement However, despite our efforts to define evaluation questions as objectively as possible, multimodal large language models (MLLMs) [28, 33, 49, 52] still struggle to match human evaluators on complex, real-world tasks, limited by current models' gaps in image comprehension relative to human capabilities. To address this, we select a subset of tasks and refine the evaluation questions based on specific examples, ensuring that MLLMs can yield intuitively reasonable results. + +One challenge in using MLLMs as evaluators is the sensitivity of models like GPT-4V/GPT-4o to the order of image presentation [54, 56, 63, 66], which particularly makes these models less reliable for consistent comparative evaluation. MLLMs also fail to identify the sequence of multiple input images, whether it is input separately for multiple images or input as a collage. To mitigate this, we manually fine-tune evaluation questions for each case within the task subset, ensuring alignment with human intuition. + +For example, in the task of generating a children's storybook with defined character images, we refine each evaluation angle by randomly sampling pairs of input and output images to test consistency. A typical question might ask if the primary character in the first input image appears with consistent features in the third output image. For testing, we use the Gemini 1.5 Pro [49] to score model outputs and compared these results against human ratings. In cases where discrepancies arose, annotators iteratively refine the + +![](images/4d164516efce2cb7db21f9fa56239a46dc430819a5ebc91c7b1adb05cf62e940.jpg) +Figure 3. Visualization of model generated images on IDEA-Bench. We present the generation results in representative task examples, including text to image, image to image, and multi-image to image. In the case of no input image or no generated image, use $\nabla$ instead. + +Table 2. Experimental results on all categories of IDEA-Bench. Each task category is averaged across all its subtasks, with the top-ranked model scores for each task type highlighted in bold. Task types that a model cannot support are marked with “-” and are treated as 0 points in the average score calculation. “+” represents the use of MLLM for prompt rephrasing. + +
MethodParam.Scores on All CategoriesAvg. Score
T2II2IIs2IT2IsI(s)2Is
FLUX-1† [25]12B46.0612.134.8920.1529.1722.48
DALL-E 3† [40]12B24.346.955.2714.3614.4413.07
SD3† [14]2B24.0410.794.6921.5913.0614.83
Pixart† [7]0.6B14.447.753.4817.4621.3912.90
InstructPix2Pix [5]1B-17.58---3.52
MagicBrush [60]1B-19.07---3.81
Anole [8]7B0.000.640.001.740.000.48
Emu2 [47]37B17.987.058.98--6.81
OmniGen [57]3.8B21.418.172.77--6.47
Emu2† [47]37B17.987.058.9815.5312.7812.46
OmniGen† [57]3.8B21.418.172.7723.5221.3915.45
+ +language of the evaluation questions until the MLLM evaluations align with human judgment. This approach enhances the robustness of our benchmark, ensuring that evaluations reflect both task-specific accuracy and consistency across diverse visual requirements. + +# 3.3. Evaluation + +# 3.3.1. Human Evaluation + +Evaluation Metric Each subtask in IDEA-Bench comprises 2 to 5 specific cases, with all cases within a subtask being evaluated using the same set of six binary (0 or 1) judgment questions. These evaluation questions are orga + +nized into three hierarchical levels, reflecting the priority of task completion over mere aesthetic quality. For each case, scoring progresses sequentially through these levels. If a model does not achieve a perfect score at a lower level, the scores for all higher levels are automatically set to 0. This hierarchical approach ensures that while a model may generate high-quality and visually appealing images, it must also fully meet the task requirements to receive a high overall score. This methodology aligns the evaluation criteria with human design standards, emphasizing the importance of fulfilling task objectives. + +Formally, let the dataset be denoted as $\mathcal{D} = \{\mathcal{D}_c = \{\mathcal{T}_t\}_{t=1}^{T_c}\}_{c=1}^C$ . Each subtask $\mathcal{T}_t$ consists of 2 to 5 specific cases, all evaluated using the same set of six binary judgment questions. The score for each subtask is calculated by averaging the binary scores of its cases hierarchically. Subsequently, each major category $\mathcal{D}_c$ receives its score by averaging the scores of its subtasks and converting the result to a percentage. The overall score for the entire dataset $\mathcal{D}$ is then determined by averaging the scores of all $C$ major categories. + +Importantly, if a model fails to complete a specific task, it is assigned a score of 0 for that task, ensuring that incomplete performances are accurately reflected in the overall evaluation. This hierarchical scoring framework ensures a precise and objective assessment of model performance, closely aligning with the rigorous standards expected by professional designers. + +Table 3. Experimental results on Text-to-Image. Each task category is averaged across all its subtasks, with the top-ranked model scores for each subtask highlighted in bold. + +
MethodParam.Subtasks ScoreAvg. Score
Arch. StyleBus. CardGame UIInf. ChartInt.Paint.Sculp.TicketLand.LOGOPoster
FLUX-1 [25]12B100.0038.895.560.0066.6761.1116.6716.6783.3361.1156.6746.06
DALL-E 3 [40]12B22.220.000.000.0011.11100.0011.110.0038.8961.1123.3324.34
Stable Diffusion 3 [14]2B38.890.005.560.0050.005.5616.670.0050.0061.1136.6724.04
Pixart [7]0.6B5.560.000.000.0022.2233.3316.670.0022.2222.2236.6714.44
Anole [8]7B0.000.000.000.000.000.000.000.000.000.000.000.00
Emu2 [47]37B22.220.000.000.0038.8938.8916.670.0044.4416.6720.0017.98
OmniGen [57]3.8B38.890.000.000.0027.7850.0016.670.0038.8933.3330.0021.41
+ +Table 4. Experimental results on Image-to-Image. Each task category is averaged across all its subtasks, with the top-ranked model scores for each subtask highlighted in bold. ${}^{\prime \prime }$ " represent the use of MLLM for prompt rephrasing. + +
MethodParam.Subtasks ScoreAvg. Score
Anim. Attr.Coup. IconID PhotoBlurCompl.Ret.Str.Light. EditEnl.Obj. EditPack. Rend.Style EditBrand Merch.
FLUX-1† [25]12B0.0016.670.002.784.170.0012.500.00100.006.250.004.1711.1112.13
DALL-E 3‡ [40]12B5.5611.110.002.780.000.004.170.0038.894.178.334.1711.116.95
Stable Diffusion 3† [14]2B8.3327.780.002.784.170.008.330.0055.564.178.334.1716.6710.79
Pixart† [7]0.6B5.5655.564.172.784.170.000.000.0011.114.170.002.0811.117.75
InstructPix2Pix [5]1B44.4416.6716.6713.8925.0054.170.000.000.0010.4225.0016.675.5617.58
MagicBrush [60]1B33.3316.6712.5025.0025.0041.670.000.000.0035.428.3333.3316.6719.07
Anole [8]7B0.000.000.000.008.330.000.000.000.000.000.000.000.000.64
Emu2 [47]37B11.110.004.175.560.000.0020.830.0016.674.178.334.1716.677.05
OmniGen [57]3.8B16.670.004.175.560.004.174.170.000.0016.670.0010.4244.448.17
+ +# 3.3.2. Automated Evaluation + +Automated evaluation is performed on a subset of IDEA-Bench comprising 18 tasks, named IDEA-Bench-mini. During the prompt refinement phase, annotators craft unique evaluation questions for each case, in contrast to the human evaluation setup where all cases within a subtask share the same 6 evaluation questions. This tailored approach allows for more precise and context-specific assessments. Scoring is conducted using Gemini 1.5 Pro [49], adhering to the same hierarchical scoring methodology outlined in Sec. 3.3.1. To enhance result stability and reliability, each case is evaluated three times independently, and the final score is computed as the average of these three evaluations. This repetition mitigates potential variability in automated assessments and ensures consistent and robust scoring outcomes. + +By integrating both human and automated evaluation, IDEA-Bench provides a comprehensive framework for assessing the multifaceted capabilities of generative models. This dual approach not only leverages the nuanced judgment of human evaluators but also benefits from the scalability and consistency of automated tools, thereby offering a balanced and thorough evaluation of model performance. + +# 4. Experiments + +# 4.1. Experimental Details + +# 4.1.1. Prompts Rephrasing + +Currently, only a few models [8, 53, 55, 58] are capable of generating multiple related images simultaneously, makes it difficult to analyse a model's ability to generate multiple images potentially. However, as DALL-E 3 [40] could perform multi-image generation owing to its MLLM GPT-4o [33], which could summarize multimodal input and then generate the corresponding prompt for each image. To empower + +other leading T2I models [7, 13, 25], we feed all input images and text prompt to GPT-4o and rephrase prompts for each image to be generated. By combining MLLM's multimodal understanding ability, we can approximately achieve the multi-image generation tasks with basic T2I models. + +# 4.1.2. Model Reimplementation + +We evaluate the capabilities of four categories of models in our study. For leading text-to-image (T2I) models, we select FLUX-1 [25], Stable Diffusion 3 (SD3) [13], Pixart [7], and DALL-E 3 [40]. For image editing models, we choose InstructPix2Pix [5] and MagicBrush [60]. To assess models with unified generation capabilities, we select Emu2 [47] and OmniGen [57]. For models that support interleaved text and image generation, we include Anole [8]. Notably, among the selected models, Anole [8] is the only one capable of handling unrestricted number of input and output images. We endeavor to utilize the officially released model parameters and inference settings wherever possible; further details are provided in the supplementary material. + +# 4.2. Results on IDEA-Bench + +We first present the ranking of all models based on their scores across the entire dataset, as detailed in Tab. 2. Leveraging GPT-4o's [33] multimodal understanding and task translation capabilities, FLUX-1 [25] achieve the highest score of 22.46, significantly outperforming other models. Specialized image editing models [5, 60] lead in the image-to-image tasks, while the general-purpose generation model Emu2 [47] excels in the images-to-image tasks. In contrast, Anole [8], which is capable of generating multiple images simultaneously, do not achieve an ideal final score. Furthermore, no model surpass a score of 10 without utilizing the comprehension capabilities of MLLMs, indicating that general-purpose image generation models still + +Table 5. Experimental results on Images-to-Image. Each task category is averaged across all its subtasks, with the top-ranked model scores for each subtask highlighted in bold. ${}^{\prime \prime } \dagger$ " represents the use of MLLM for prompt rephrasing. + +
MethodParam.Subtasks ScoreAvg. Score
Human Attr.Image Trans.Pan.Text Edit.3D. EffectImage Blend.Spec. Effect
FLUX-1† [25]3.976.670.0010.420.0012.500.005.564.89
DALL-E 3† [40]4.763.338.336.250.008.335.565.565.27
SD3† [14]4.765.000.008.330.008.335.565.564.69
Pixart† [7]4.765.008.330.000.004.170.005.563.48
Anole [8]0.000.000.000.000.000.000.000.000.00
Emu2 [47]6.350.0050.002.080.004.170.009.268.98
OmniGen [57]1.590.000.002.080.000.0011.117.412.77
+ +Table 6. Experimental results on Text-to-Images. Each task category is averaged across all its subtasks, with the top-ranked model scores for each subtask highlighted in bold. ${}^{\prime \prime } \dagger$ " represents the use of MLLM for prompt rephrasing. + +
MethodParam.Subtasks ScoreAvg. Score
Anim. Grow.Child. BookDraw. Proc.Hist. Narr.Movie ShotsPhys. LawsPlant Grow.Style GroupConc. Vis.Coup. IconDyn. Char.
FLUX-1† [25]12B25.004.178.3370.834.174.178.337.0816.6750.0022.9220.15
DALL-E 3† [40]12B0.004.170.004.174.170.0041.6732.9212.5041.6716.6714.36
Stable Diffusion 3† [14]2B8.334.1725.0037.5012.508.3345.8318.7516.6741.6718.7521.59
Pixart† [7]0.6B37.504.174.1716.670.000.0029.1721.2516.6745.8316.6717.46
Anole [8]7B0.000.000.000.000.000.000.000.428.338.332.081.74
+ +face substantial challenges in achieving professional-grade performance. This underscores the need for comprehensive evaluation standards to guide the development of more specialized and capable generative models. The following section provides a detailed analysis of each category. + +Qualitative Analysis Fig. 3 showcases the generated images across 7 models. For ease of comparison, we select several cases from three categories: text-to-image, image-to-image, and multi-image-to-image. In fundamental text-to-image tasks, FLUX-1 [25] demonstrates a clear advantage over other models, effectively comprehending the visual elements and professional-level requirements outlined in the prompts. Regarding image editing tasks, models that are not specifically designed for image editing struggle to preserve essential information from the original image. Emu2 [47] is the only model that roughly restores the shapes and compositions of objects from the original image. OmniGen [57] excels in maintaining the identity of objects. However, for more complex tasks such as special effect synthesis, no current model sufficiently understands the intricate demands of these professional-level requirements. + +# 4.2.1. Results on Text-to-Image + +All models' results on text to image tasks are presented in Tab. 3. Notably, the FLUX-1 [25] model significantly outperforms the others, effectively accomplishing the task objectives in the majority of tasks. Based on the task settings and each model's performance, we observe the following: + +- Text Generation Challenges: Text generation tasks are a common pain point for all models. Even though models can generate specific visual elements with high quality and understand style instructions specified in the prompts, they struggle to incorporate text into images as seamlessly as human designers. For example, in the business card generation task, only FLUX-1 [25] received scores. +- Performance of Basic vs. General Models: Basic text-to-image models generally demonstrate higher task + +completion rates compared to general models like Emu2 [47] and OmniGen [57]. This is expected, as general models prioritize enhancing task generalization and multimodal information processing capabilities, which inevitably compromises some aspects of image generation quality. + +- Poor Performance on Information Charts: All models perform poorly on the information chart generation task. The primary reason is that generative models struggle to accurately transfer large amounts of text from the input to the output image, which is a basic ability to human designers. Additionally, many models do not support ultra-long text inputs. + +# 4.2.2. Results on Image-to-Image & Images-to-Image + +The image-to-image tasks represent the strengths of image editing models. In Tab. 4, MagicBrush [60] and Instruct-Pix2Pix [5] achieve the first and second highest evaluation scores, respectively. The images-to-image tasks involve more complex and diverse guidance or conditions, with Emu2 [47] attaining the highest score of 8.94, as shown in Tab. 5. From these results, we derive the following key insights: + +- Pros and Cons of Image Editing Models: In local editing tasks such as image blur and image retouching, image editing models excel at preserving the original features of the image. However, these models are typically trained or fine-tuned on specific image editing datasets, which diminishes their ability to perform subject-driven tasks, such as branded merchandise generation. +- In-Context Learning Capabilities: Emu2 [47] and OmniGen [57] exhibit in-context learning abilities, as evidenced by OmniGen's significantly higher scores in tasks like branded merchandise generation and special effect synthesis, achieving ID consistency between input and output images. Emu2 demonstrates the ability to comprehend and partially complete the image straighten task, indicating that the model possesses an understanding of + +Table 7. Experimental results on Image(s)-to-Images. Each task category is averaged across all its subtasks, with the top-ranked model scores for each subtask highlighted in bold. ${}^{\prime \prime }$ represents the use of MLLM for prompt rephrasing. + +
MethodParamsSubtasks ScoreAvg. Score
Anim. Grow.Creat. Trans.Layer Decomp.Light. EffectMovie ShotsMulti-app.Multi-dec.Multi-view Trans.
FLUX-1† [25]12B25.0066.6775.008.330.000.0016.67100.0029.17
DALL-E 3† [40]12B25.0025.008.330.004.170.0016.678.3314.44
Stable Diffusion 3† [14]2B33.3350.000.000.000.000.008.330.0013.06
Pixart† [7]0.6B33.3350.000.0058.330.000.0058.330.0021.39
Anole [8]7B0.000.000.000.000.000.000.000.000.00
+ +Table 8. Experimental results on the subset IDEA-Bench-mini. "G." represents evaluate using Gemini 1.5 pro [49], "H." represents human evaluation. "†" represents the use of MLLM for prompt rephrasing. + +
MethodCLIPScoreEvaluation CategoryAvg. Score
T2II2IIs2IT2IsI(s)2Is
G.H.G.H.G.H.G.H.G.H.G.H.
FLUX-1† [25]0.343283.3367.0428.715.564.866.2532.2928.1337.5038.3337.4829.06
DALL-E 3† [40]0.346255.5641.1127.785.5620.375.5629.1721.8829.1733.3332.4121.49
SD3† [14]0.348356.3039.2623.618.345.555.5632.2923.9631.2529.8029.8019.76
Pixart† [7]0.343526.3027.0430.555.5629.862.0935.4223.9631.2530.0030.6817.73
InstructPix2Pix [5]0.1194--27.782.78------5.560.56
MagicBrush [60]0.1212--9.728.34------1.941.67
Emu2 [47]0.182135.8027.0443.528.3434.722.09----22.817.49
OmniGen [57]0.179946.3034.0740.2711.1125.460.00----22.416.81
+ +the semantic aspects of visual elements. Recent advances in Multimodal Visualization-of-Thought (MVoT) [29] suggest a promising direction for enhancing generative models' ability to interpret complex instructions, potentially improving their generalization to professional design tasks. + +- Challenges for Generative Models: General-purpose models perform poorly in tasks such as light condition editing. Although these tasks are frequently encountered by designers, the models require a certain level of understanding of real-world physical models, presenting a challenge for future generative models. + +# 4.2.3. Results on Text-to-Images & Image(s)-to-Images + +Tab. 6 and Tab. 7 present the evaluation results for text-to-images and select image(s)-to-images tasks, respectively. In both categories, SD3 [13] and FLUX-1 [25] achieve the highest scores. For tasks such as historical event generation, MLLMs can effectively convey detailed style requirements during the prompt rephrasing phase, resulting in relatively consistent style outcomes. However, for more complex tasks like children's book generation that involve sophisticated capabilities such as consistency and preservation over ID and style, existing models struggle to perform adequately even when assisted by MLLMs. Nonetheless, generating multiple related images holds significant value; for example, the generated movie shots can serve as guiding conditions for multi-shot video generation models. We aim for the task types designed in our benchmark to align generative models with the capabilities of human designers, thereby further advancing toward general AI. + +# 4.2.4. Results on IDEA-Bench-mini + +Tab. 8 presents the evaluation results for a subset of 18 tasks within our benchmark, where we calculate CLIPScore [18], + +Gemini 1.5 Pro [49] scores, and human evaluation scores. As shown in Tab. 8, FLUX-1 [25], leveraging MLLM's prompt rephrasing capabilities, achieve the highest scores in both MLLM-based and human evaluations within this subset. In contrast, when excluding the use of MLLMs, Emu2 [47] secure the top score. These experimental results are consistent with the overall benchmark dataset findings. Additionally, compared to CLIPScore, MLLM-based evaluations more closely align with human preferences in terms of professional evaluation data and assessment criteria. This suggests that as MLLM capabilities continue to evolve, utilizing MLLMs for image generation evaluation will become increasingly stable and reliable. + +# 5. Conclusion + +We introduce IDEA-Bench, a comprehensive benchmark designed to bridge the gap between current generative model capabilities and the stringent demands of professional-grade image design. IDEA-Bench encompasses 100 professional image generation tasks, utilizing a detailed evaluation framework to ensure precise and objective assessments. Additionally, we build a representative subset IDEA-Bench-mini that facilitates automated evaluation using MLLMs. Our evaluations reveal significant gaps in existing models, particularly in handling intricate instructions and maintaining consistency in multi-image generation, underscoring the need for more advanced and versatile models. By aligning evaluation standards with the nuanced requirements of human designers, IDEA-Bench serves as a pivotal tool for guiding the development of generative models toward achieving professional-grade performance and advancing toward autonomous and sophisticated visual generation capabilities. + +# 6. Acknowledgments + +This work is supported by the Youth Innovation Promotion Association CAS, the Beijing Natural Science Foundation (No. JQ24024), and the National Key Laboratory of Science and Technology on Space-Born Intelligent Information Processing Project (No. TJ-04-23-02). + +# References + +[1] Shuai Bai, Shusheng Yang, Jinze Bai, Peng Wang, Xingxuan Zhang, Junyang Lin, Xinggang Wang, Chang Zhou, and Jingren Zhou. Touchstone: Evaluating vision-language models by language models. arXiv preprint arXiv:2308.16890, 2023. 4 +[2] Eslam Mohamed Bakr, Pengzhan Sun, Xiaogian Shen, Faizan Farooq Khan, Li Erran Li, and Mohamed Elhoseiny. Hrs-bench: Holistic, reliable and scalable benchmark for text-to-image models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20041-20053, 2023. 3 +[3] Jason Baldridge, Jakob Bauer, Mukul Bhutani, Nicole Brichtova, Andrew Bunner, Kelvin Chan, Yichang Chen, Sander Dieleman, Yuqing Du, Zach Eaton-Rosen, et al. Imagen 3. arXiv preprint arXiv:2408.07009, 2024. 1, 3 +[4] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.openai.com/papers/dall-e-3.pdf, 2(3):8, 2023. 1, 3 +[5] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 1, 2, 5, 6, 7, 8, 3 +[6] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 1, 3 +[7] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-alpha: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 1, 3, 5, 6, 7, 8 +[8] Ethan Chern, Jiadi Su, Yan Ma, and Pengfei Liu. Anole: An open, autoregressive, native large multimodal models for interleaved image-text generation. arXiv preprint arXiv:2407.06135, 2024. 1, 2, 5, 6, 7, 8, 3 +[9] Jaemin Cho, Abhay Zala, and Mohit Bansal. Dall-eval: Probing the reasoning skills and social biases of text-to-image generation models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3043-3054, 2023. 1, 3 +[10] Min Jin Chong and David Forsyth. Effectively unbiased fid and inception score and where to find them. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6070-6079, 2020. 2 +[11] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil + +Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 1, 3 +[12] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 1, 3 +[13] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 1, 3, 6, 8 +[14] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, Kyle Lacey, Alex Goodwin, Yannik Marek, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis, 2024. 2, 5, 6, 7, 8, 3 +[15] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2 +[16] Yuying Ge, Sijie Zhao, Ziyun Zeng, Yixiao Ge, Chen Li, Xintao Wang, and Ying Shan. Making llama see and draw with seed tokenizer. arXiv preprint arXiv:2310.01218, 2023. 1 +[17] Dhruba Ghosh, Hannaneh Hajishirzi, and Ludwig Schmidt. Geneval: An object-focused framework for evaluating text-to-image alignment. Advances in Neural Information Processing Systems, 36, 2024. 1, 3 +[18] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 2, 8 +[19] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2 +[20] Yushi Hu, Benlin Liu, Jungo Kasai, Yizhong Wang, Mari Ostendorf, Ranjay Krishna, and Noah A Smith. Tifa: Accurate and interpretable text-to-image faithfulness evaluation with question answering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20406-20417, 2023. 1, 3 +[21] Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. Advances in Neural Information Processing Systems, 36:78723-78747, 2023. 3 +[22] Lianghua Huang, Wei Wang, Zhi-Fan Wu, Huanzhang Dou, Yupeng Shi, Yutong Feng, Chen Liang, Yu Liu, and Jingren Zhou. Group diffusion transformers are unsupervised multitask learners. arXiv preprint arXiv:2410.15027, 2024. 1, 3, 2 + +[23] Max Ku, Tianle Li, Kai Zhang, Yujie Lu, Xingyu Fu, Wenwen Zhuang, and Wenhu Chen. Imagenhub: Standardizing the evaluation of conditional image generation models. arXiv preprint arXiv:2310.01596, 2023. 1, 3 +[24] Black Forest Labs. Flux: Inference repository. https://github.com/black-forest-labs/flux, 2024. Accessed: 2024-10-25. 1, 3 +[25] Black Forest Labs. Flux.1, 2024. 2, 5, 6, 7, 8, 1, 3 +[26] Tony Lee, Michihiro Yasunaga, Chenlin Meng, Yifan Mai, Joon Sung Park, Agrim Gupta, Yunzhi Zhang, Deepak Narayanan, Hannah Teufel, Marco Bellagente, et al. Holistic evaluation of text-to-image models. Advances in Neural Information Processing Systems, 36, 2024. 3 +[27] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023. 4 +[28] Bo Li, Kaichen Zhang, Hao Zhang, Dong Guo, Renrui Zhang, Feng Li, Yuanhan Zhang, Ziwei Liu, and Chunyuan Li. Llava-last: Stronger llms supercharge multimodal capabilities in the wild, 2024. 1, 4 +[29] Chengzhu Li, Wenshan Wu, Huanyu Zhang, Yan Xia, Shaoguang Mao, Li Dong, Ivan Vulic, and Furu Wei. Imagine while reasoning in space: Multimodal visualization-of-thought. arXiv preprint arXiv:2501.07542, 2025. 8 +[30] Kevin Lin, Zhengyuan Yang, Linjie Li, Jianfeng Wang, and Lijuan Wang. Designbench: Exploring and benchmarking dall-e 3 for imagining visual design. arXiv preprint arXiv:2310.15144, 2023. 3 +[31] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? In European Conference on Computer Vision, pages 216-233. Springer, 2025. 4 +[32] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4296-4304, 2024. 1 +[33] OpenAI. Gpt-4o, 2024. 1, 3, 4, 6, 2, 8 +[34] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730-27744, 2022. 1, 3 +[35] Vitali Petsiuk, Alexander E Siemann, Saisamrit Surbehera, Zad Chin, Keith Tyser, Gregory Hunter, Arvind Raghavan, Yann Hicke, Bryan A Plummer, Ori Kerret, et al. Human evaluation of text-to-image models on a multi-task benchmark. arXiv preprint arXiv:2211.12112, 2022. 3 +[36] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 1, 2, 3 + +[37] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 1, 3 +[38] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 1, 3 +[39] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International conference on machine learning, pages 8821-8831. Pmlr, 2021. 1, 3 +[40] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1, 2, 5, 6, 7, 8, 3 +[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 3 +[42] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv, 2022. 2 +[43] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 1, 3 +[44] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 1, 3 +[45] Shelly Sheynin, Adam Polyak, Uriel Singer, Yuval Kirstain, Amit Zohar, Oron Ashual, Devi Parikh, and Yaniv Taigman. Emu edit: Precise image editing via recognition and generation tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8871-8879, 2024. 1, 2, 3 +[46] Quan Sun, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, Yueze Wang, Hongcheng Gao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Emu: Generative pretraining in multimodality. In The Twelfth International Conference on Learning Representations, 2023. 1 +[47] Quan Sun, Yufeng Cui, Xiaosong Zhang, Fan Zhang, Qiying Yu, Yueze Wang, Yongming Rao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative multimodal models are in-context learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14398-14409, 2024. 1, 3, 5, 6, 7, 8 +[48] Chameleon Team. Chameleon: Mixed-modal early-fusion + +foundation models. arXiv preprint arXiv:2405.09818, 2024.3 +[49] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 1, 4, 6, 8, 2, 3 +[50] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1, 3 +[51] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 1, 3 +[52] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1, 4 +[53] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 1, 2, 3, 6 +[54] Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A Smith, Daniel Khashabi, and Hannaneh Hajishirzi. Self-instruct: Aligning language models with self-generated instructions. arXiv preprint arXiv:2212.10560, 2022. 4 +[55] Shengqiong Wu, Hao Fei, Leigang Qu, Wei Ji, and Tat-Seng Chua. Next-gpt: Any-to-any multimodal llm. In Proceedings of the International Conference on Machine Learning, pages 53366-53397, 2024. 2, 6 +[56] Tong Wu, Guandao Yang, Zhibing Li, Kai Zhang, Ziwei Liu, Leonidas Guibas, Dahua Lin, and Gordon Wetzstein. Gpt-4v (ision) is a human-aligned evaluator for text-to-3d generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22227-22238, 2024. 4 +[57] Shitao Xiao, Yueze Wang, Junjie Zhou, Huaying Yuan, Xingrun Xing, Ruiran Yan, Shuting Wang, Tiejun Huang, and Zheng Liu. Omnigen: Unified image generation. arXiv preprint arXiv:2409.11340, 2024. 3, 5, 6, 7, 8, 1 +[58] Shuai Yang, Yuying Ge, Yang Li, Yukang Chen, Yixiao Ge, Ying Shan, and Yingcong Chen. Seed-story: Multimodal long story generation with large language model. arXiv preprint arXiv:2407.08683, 2024. 2, 6 +[59] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024.3 +[60] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. Advances in Neural Information Processing Systems, 36, 2024. 2, 5, 6, 7, 8, 1, 3 + +[61] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1 +[62] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 1, 3 +[63] Xinlu Zhang, Yujie Lu, Weizhi Wang, An Yan, Jun Yan, Lianke Qin, Heng Wang, Xifeng Yan, William Yang Wang, and Linda Ruth Petzold. Gpt-4v (ision) as a generalist evaluator for vision-language tasks. arXiv preprint arXiv:2311.01361, 2023. 4 +[64] Yi-Fan Zhang, Huanyu Zhang, Haochen Tian, Chaoyou Fu, Shuangqing Zhang, Junfei Wu, Feng Li, Kun Wang, Qing-song Wen, Zhang Zhang, et al. Mme-realworld: Could your multimodal llm challenge high-resolution real-world scenarios that are difficult for humans? arXiv preprint arXiv:2408.13257, 2024. 4 +[65] Xuanlei Zhao, Xiaolong Jin, Kai Wang, and Yang You. Real-time video generation with pyramid attention broadcast. arXiv preprint arXiv:2408.12588, 2024. 3 +[66] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in Neural Information Processing Systems, 36:46595-46623, 2023. 4 +[67] Chunting Zhou, Lili Yu, Arun Babu, Kushal Tirumala, Michihiro Yasunaga, Leonid Shamis, Jacob Kahn, Xuezhe Ma, Luke Zettlemoyer, and Omer Levy. Transfusion: Predict the next token and diffuse images with one multi-modal model, 2024. 3 \ No newline at end of file diff --git a/2025/IDEA-Bench_ How Far are Generative Models from Professional Designing_/images.zip b/2025/IDEA-Bench_ How Far are Generative Models from Professional Designing_/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..29af526541496e5635d1588b87b0bbde7da87114 --- /dev/null +++ b/2025/IDEA-Bench_ How Far are Generative Models from Professional Designing_/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:61c6232380922c6c68758ec7ba4b75a302b63ecfb3bf7a07f5539416e428db26 +size 781932 diff --git a/2025/IDEA-Bench_ How Far are Generative Models from Professional Designing_/layout.json b/2025/IDEA-Bench_ How Far are Generative Models from Professional Designing_/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3add2a8bad83a2488c5258146ec450e322c38464 --- /dev/null +++ b/2025/IDEA-Bench_ How Far are Generative Models from Professional Designing_/layout.json @@ -0,0 +1,7276 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 69, + 103, + 541, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 103, + 541, + 121 + ], + "spans": [ + { + "bbox": [ + 69, + 103, + 541, + 121 + ], + "type": "text", + "content": "IDEA-Bench: How Far are Generative Models from Professional Designing?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 74, + 141, + 144, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 141, + 144, + 158 + ], + "spans": [ + { + "bbox": [ + 74, + 141, + 144, + 158 + ], + "type": "text", + "content": "Chen Liang" + }, + { + "bbox": [ + 74, + 141, + 144, + 158 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 99, + 163, + 164, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 163, + 164, + 176 + ], + "spans": [ + { + "bbox": [ + 99, + 163, + 164, + 176 + ], + "type": "text", + "content": "Zhi-Fan " + }, + { + "bbox": [ + 99, + 163, + 164, + 176 + ], + "type": "inline_equation", + "content": "\\mathbf{W}\\mathbf{u}^{3}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 167, + 142, + 255, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 142, + 255, + 158 + ], + "spans": [ + { + "bbox": [ + 167, + 142, + 255, + 158 + ], + "type": "text", + "content": "Lianghua Huang3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 193, + 163, + 254, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 193, + 163, + 254, + 177 + ], + "spans": [ + { + "bbox": [ + 193, + 163, + 254, + 177 + ], + "type": "text", + "content": "Yupeng Shi" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 278, + 143, + 346, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 278, + 143, + 346, + 157 + ], + "spans": [ + { + "bbox": [ + 278, + 143, + 346, + 157 + ], + "type": "text", + "content": "Jingwu Fang" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 282, + 163, + 356, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 282, + 163, + 356, + 177 + ], + "spans": [ + { + "bbox": [ + 282, + 163, + 356, + 177 + ], + "type": "text", + "content": "Junge Zhang1,2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 370, + 143, + 454, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 143, + 454, + 157 + ], + "spans": [ + { + "bbox": [ + 370, + 143, + 454, + 157 + ], + "type": "text", + "content": "Huanzhang Dou" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 384, + 163, + 444, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 163, + 444, + 175 + ], + "spans": [ + { + "bbox": [ + 384, + 163, + 444, + 175 + ], + "type": "text", + "content": "Xin Zhao*" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 479, + 143, + 534, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 479, + 143, + 534, + 157 + ], + "spans": [ + { + "bbox": [ + 479, + 143, + 534, + 157 + ], + "type": "text", + "content": "Wei Wang3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 468, + 163, + 510, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 468, + 163, + 510, + 175 + ], + "spans": [ + { + "bbox": [ + 468, + 163, + 510, + 175 + ], + "type": "text", + "content": "Yu Liu" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 170, + 190, + 441, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 190, + 441, + 205 + ], + "spans": [ + { + "bbox": [ + 170, + 190, + 441, + 205 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 170, + 190, + 441, + 205 + ], + "type": "text", + "content": "Institute of Automation, Chinese Academy of Sciences" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 205, + 494, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 205, + 494, + 219 + ], + "spans": [ + { + "bbox": [ + 115, + 205, + 494, + 219 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 115, + 205, + 494, + 219 + ], + "type": "text", + "content": "School of Artificial Intelligence, University of Chinese Academy of Sciences" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 235, + 220, + 375, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 220, + 375, + 233 + ], + "spans": [ + { + "bbox": [ + 235, + 220, + 375, + 233 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 235, + 220, + 375, + 233 + ], + "type": "text", + "content": "Tongyi Lab, Alibaba Group" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 192, + 234, + 418, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 234, + 418, + 247 + ], + "spans": [ + { + "bbox": [ + 192, + 234, + 418, + 247 + ], + "type": "text", + "content": "4University of Science and Technology Beijing" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 255, + 247, + 356, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 247, + 356, + 262 + ], + "spans": [ + { + "bbox": [ + 255, + 247, + 356, + 262 + ], + "type": "text", + "content": "5Zhejiang University" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 152, + 289, + 200, + 301 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 289, + 200, + 301 + ], + "spans": [ + { + "bbox": [ + 152, + 289, + 200, + 301 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 54, + 314, + 297, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 314, + 297, + 555 + ], + "spans": [ + { + "bbox": [ + 54, + 314, + 297, + 555 + ], + "type": "text", + "content": "Recent advancements in image generation models enable the creation of high-quality images and targeted modifications based on textual instructions. Some models even support multimodal complex guidance and demonstrate robust task generalization capabilities. However, they still fall short of meeting the nuanced, professional demands of designers. To bridge this gap, we introduce IDEA-Bench, a comprehensive benchmark designed to advance image generation models toward applications with robust task generalization. IDEA-Bench comprises 100 professional image generation tasks and 275 specific cases, categorized into five major types based on the current capabilities of existing models. Furthermore, we provide a representative subset of 18 tasks with enhanced evaluation criteria to facilitate more nuanced and reliable evaluations using Multimodal Large Language Models (MLLMs). By assessing models' ability to comprehend and execute novel, complex tasks, IDEA-Bench paves the way toward the development of generative models with autonomous and versatile visual generation capabilities." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 577, + 135, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 577, + 135, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 577, + 135, + 590 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 597, + 296, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 296, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 296, + 694 + ], + "type": "text", + "content": "Recent advancements in Text-to-Image (T2I) models [3, 4, 7, 12, 13, 24, 36, 39–41, 44] have significantly enhanced the ability to generate high-quality images from textual descriptions. Building on these successes, models such as ControlNet [61] and T2I-Adapter [32] integrate additional networks into text-to-image diffusion frameworks to incorporate visual conditioning. Similarly, InstructPix2Pix [5] and Emu-Edit [45] are specifically trained on datasets tailored" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 289, + 555, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 555, + 469 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 555, + 469 + ], + "type": "text", + "content": "for complex image editing tasks. Despite the widespread popularity of models like DALL-E 3 [40] and FLUX-1 [24], which attract millions of visits daily, professional users often rely on established workflows and specialized software for image creation. This reliance highlights a substantial gap between the capabilities of current image generation models and the demanding requirements of professional-grade image design. When existing image generation models focus on academic task research, the rapid evolution of large language models (LLMs) [6, 11, 34, 37, 38, 50, 51, 62] and multimodal language models (MLLMs) [28, 33, 49, 52] indicate that future image generation models will increasingly aim to achieve both professional-grade quality and task unification, enabling them to handle a diverse range of complex applications seamlessly." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 472, + 556, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 556, + 638 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 556, + 638 + ], + "type": "text", + "content": "Recent advances in image generation have demonstrated robust task unification and generalization through large-scale pre-training and arbitrary task learning [8, 16, 22, 46, 47, 53]. These models handle a wide range of academic tasks with a unified input-output format and can generalize to unseen tasks, showcasing in-context learning abilities. However, existing benchmarks [9, 17, 20, 23, 43, 45] are often narrow in scope, focusing primarily on isolated academic tasks and lacking comprehensive criteria to evaluate the multifaceted demands of professional image design. Consequently, there is a critical need for more robust and versatile evaluation frameworks to effectively assess the diverse and sophisticated capabilities of contemporary generative models." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "content": "To bridge the gap between generative models and professional-grade design, we introduce IDEA-Bench (Intelligent Design Evaluation and Assessment Benchmark). We conduct a thorough review of real-world design and art tasks from diverse platforms, distilling 100 representative professional image generation tasks" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18541" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 71, + 531, + 297 + ], + "blocks": [ + { + "bbox": [ + 82, + 71, + 531, + 297 + ], + "lines": [ + { + "bbox": [ + 82, + 71, + 531, + 297 + ], + "spans": [ + { + "bbox": [ + 82, + 71, + 531, + 297 + ], + "type": "image", + "image_path": "2a6d8d3483ddd97c88488c1e0488295dd43e9043aefa917882efdda10a5b2870.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 300, + 555, + 333 + ], + "lines": [ + { + "bbox": [ + 55, + 300, + 555, + 333 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 555, + 333 + ], + "type": "text", + "content": "Figure 1. Overview of IDEA-Bench. IDEA-Bench comprises 5 categories, encompassing a total of 100 professional-level subtasks, 275 cases, and 1,650 hierarchical evaluation questions. Each category provides subtask examples, quantitative statistics, and showcases a leaderboard of mainstream models." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 344, + 296, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 344, + 296, + 571 + ], + "spans": [ + { + "bbox": [ + 56, + 344, + 296, + 571 + ], + "type": "text", + "content": "and 275 cases that comprehensively span the essential capabilities and effects for creating all forms of artwork. These tasks are systematically categorized into five distinct types based on required capabilities: text-to-image, image-to-image, images-to-image, text-to-images, and image(s)-to-images, as illustrated in Fig. 1. Currently, only a few models [8, 53, 55, 58] can handle the most intricate category, which demands MLLM-level understanding of both image and text and the ability to produce variable-length image sequences. To evaluate generated images, we leverage MLLM to transform the evaluation into an VQA task. This approach surpasses traditional metrics like FID [10] and CLIPScore [18], which fail to capture nuances in aesthetic quality, contextual relevance, and multimodal integration. IDEA-Bench includes six detailed evaluation questions per case, totaling 1,650 binary scoring items, and a representative subset of 18 tasks for more nuanced and reliable assessments, ensuring precise and objective evaluations aligned with professional design standards." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 574, + 266, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 574, + 266, + 586 + ], + "spans": [ + { + "bbox": [ + 67, + 574, + 266, + 586 + ], + "type": "text", + "content": "Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 594, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 641 + ], + "type": "text", + "content": "- We introduce IDEA-Bench to bridge the gap between current generative model capabilities and the stringent demands of professional-grade image design, which consists of 100 carefully selected tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 642, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 700 + ], + "type": "text", + "content": "- We develop a task categorization of five distinct categories based on complexity and modality requirements, with a detailed evaluation framework comprising 1,650 binary scoring items, enabling precise and objective assessment." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "content": "- We leverage MLLM-evaluation on 18 representative" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 322, + 344, + 555, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 344, + 555, + 369 + ], + "spans": [ + { + "bbox": [ + 322, + 344, + 555, + 369 + ], + "type": "text", + "content": "tasks, demonstrating that MLLMs possess the capability for objective image assessment." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 377, + 401, + 389 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 377, + 401, + 389 + ], + "spans": [ + { + "bbox": [ + 314, + 377, + 401, + 389 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 396, + 450, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 396, + 450, + 410 + ], + "spans": [ + { + "bbox": [ + 314, + 396, + 450, + 410 + ], + "type": "text", + "content": "2.1. Image Synthesis Models" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 415, + 555, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 555, + 668 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 555, + 668 + ], + "type": "text", + "content": "Recently, diffusion models such as Stable Diffusion [14, 36], DALL-E 3 [40], and FLUX-1 [25] have gained significant popularity for generating photorealistic images and offering improved training stability. However, state-of-the-art models still struggle with intricate prompts involving multiple visual concepts and extensive textual information, similar to image editing models [5, 45, 60], which find it challenging to interpret detailed commands requiring nuanced modifications for professional-grade tasks. To enhance generative models' capabilities, IDEA-Bench assesses existing image editing models against the rigorous demands of professional image design, aiming to emulate the creative and analytical processes of human designers. Additionally, image customization [15, 19, 42] is crucial for professionals needing varied personalized images. While techniques like Textual Inversion [15], LoRA [19], and DreamBooth [42] improve text-to-image models through fine-tuning, they lack adaptability to unseen subjects. IDEA-Bench aims at comprehensively evaluating models' customization capabilities across multi-image generation tasks, assessing consistency in content, style, identity, and conceptual aspects." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 672, + 470, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 470, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 470, + 684 + ], + "type": "text", + "content": "2.2. Universal Generative Model" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Universal generative models have become pivotal in advancing AI's ability to perform a diverse array of tasks" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18542" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 114, + 294, + 188 + ], + "blocks": [ + { + "bbox": [ + 56, + 71, + 295, + 115 + ], + "lines": [ + { + "bbox": [ + 56, + 71, + 295, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 295, + 115 + ], + "type": "text", + "content": "Table 1. Comparison with other image generation benchmarks. IDEA-Bench offers a broader range of task categories, multiple levels of model capabilities for evaluation, and longer and more complex prompts." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 114, + 294, + 188 + ], + "lines": [ + { + "bbox": [ + 56, + 114, + 294, + 188 + ], + "spans": [ + { + "bbox": [ + 56, + 114, + 294, + 188 + ], + "type": "table", + "html": "
BenchmarkTasksEvaluation CategoryAvg. LengthMLLM Eval.
T2IT2IsI2IIs2II(s)2Is
PaintSkills [9]3✓XXXX10.44X
GenEval [17]6✓XXXX7.61X
EmuEdit [45]12XX✓XX8.57X
TIFA [20]12✓XXXX10.46GPT
DreamBooth [43]30XX✓XX7.58X
ImagenHub [23]7✓X✓XX8.61X
IDEA-Bench100✓✓✓✓✓138.68Gemini
", + "image_path": "2afc00cbd263a33b0056a7d94d3802e37e8cd1ee081852b8cd1c36e3e1c6be89.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 193, + 294, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 193, + 294, + 384 + ], + "spans": [ + { + "bbox": [ + 57, + 193, + 294, + 384 + ], + "type": "text", + "content": "across both language and vision domains. LLMs [6, 11, 34, 37, 38, 50, 51, 62] have demonstrated remarkable versatility, excelling in tasks such as question answering and summarization through extensive pretraining on varied datasets. Building on this success, recent academic efforts [47, 48, 53, 57, 67] aim to adopt a more cohesive paradigm that addresses a wide range of visual generation tasks, including text-to-image and autoregressive image synthesis. Despite these advancements, there remains a significant gap in establishing a unified benchmark that comprehensively evaluates the multifaceted capabilities of these generative models. IDEA-Bench establishes a unified evaluation standard to assesses the strengths and limitations of existing generative foundation models, while guiding the development towards handling realistic and complex real-world image generation tasks at the same time." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 391, + 244, + 402 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 391, + 244, + 402 + ], + "spans": [ + { + "bbox": [ + 57, + 391, + 244, + 402 + ], + "type": "text", + "content": "2.3. Benchmarks for Generative Models" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 408, + 294, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 408, + 294, + 647 + ], + "spans": [ + { + "bbox": [ + 57, + 408, + 294, + 647 + ], + "type": "text", + "content": "The development of benchmarks typically keeps pace with advancements in text-to-image (T2I) synthesis [3, 4, 7, 12, 13, 24, 36, 39-41, 44]. DrawBench was initially introduced by Imagen [44], followed by DALL-EVAL [9], which proposed PaintSkills to evaluate visual reasoning and social bias capabilities. Recently, an increasing number of benchmarks have emerged [2, 21, 23, 26, 35]. While these benchmarks primarily focus on assessing image quality and alignment, DEsignBench [30] shares similarities with our approach by emphasizing scenarios within authentic design contexts. Differently, IDEA-Bench extends beyond this by addressing additional open challenges in professional creativity. Furthermore, IDEA-Bench is more closely aligned with the latest model capabilities, enabling the use of multi-image and complex instructions as diverse forms of guidance. As shown in Tab. 1, IDEA-Bench greatly expand the prompt length and adde two evaluation dimensions: multi image input and multi image generation. This allows for the generation of both single and multiple images, thereby better mimicking the workflows of professional designers." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 658, + 135, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 135, + 669 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 135, + 669 + ], + "type": "text", + "content": "3. IDEA-Bench" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 678, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 678, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 678, + 294, + 713 + ], + "type": "text", + "content": "We present IDEA-Bench, a comprehensive image generation benchmark that encompasses a wide range of professional tasks in image generation and rigorously challenges" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 316, + 72, + 553, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 155 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 155 + ], + "type": "text", + "content": "models across multiple dimensions of capability. In this section, we first introduce IDEA-Bench's data collection over four different model level in Sec. 3.1, as well as the manual annotation pipeline in Section Sec. 3.2, followed by a detailed explanation of both human evaluation and automated assessment methodologies, along with the specific metrics used, in Sec. 3.3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 316, + 166, + 408, + 177 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 166, + 408, + 177 + ], + "spans": [ + { + "bbox": [ + 316, + 166, + 408, + 177 + ], + "type": "text", + "content": "3.1. Data Collection" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 316, + 183, + 553, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 183, + 553, + 266 + ], + "spans": [ + { + "bbox": [ + 316, + 183, + 553, + 266 + ], + "type": "text", + "content": "To ensure that the tasks evaluated in IDEA-Bench closely mirror real-world and professional scenarios, we source all task directives and data from the internet and professional designers. Leveraging the knowledge and capabilities of GPT-4o [33], we clearly define and classify the tasks, creating task variants aligned with existing model capabilities. The tasks in IDEA-Bench are categorized as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 316, + 284, + 553, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 284, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 284, + 553, + 415 + ], + "type": "text", + "content": "Text-only Guided Image Generation While typical T2I users provide brief descriptions and rely on the model's discretion, professional users meticulously control every visual detail, often iterating multiple prompts to achieve the desired outcome. Current T2I models [3, 4, 7, 12, 13, 24, 36, 39-41, 44] struggle with complex prompts that include numerous visual concepts and lengthy text. To address this, IDEA-Bench includes 11 tasks designed to evaluate models' ability to handle extensive visual components and incorporate text within images, essential for tasks like poster generation and business card generation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 316, + 433, + 553, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 433, + 553, + 588 + ], + "spans": [ + { + "bbox": [ + 316, + 433, + 553, + 588 + ], + "type": "text", + "content": "Image / Multi-image Guided Image Generation Unlike ImagenHub [23], which categorizes conditional image generation tasks under specific definitions, we unify these into image / multi-image guided image generation. This consolidation promotes the development of models with enhanced task unification and generalization, similar to advancements in LLMs. IDEA-Bench includes numerous design tasks guided by both image and text inputs in real-world contexts, such as brand merchandise generation, package rendering, and image retouching. With the advent of models like GDT [22], Emu2 [47], and OmniGen [57], we anticipate future models will exhibit strong multimodal understanding and generative capabilities." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 316, + 605, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 605, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 605, + 553, + 713 + ], + "type": "text", + "content": "Text & (Multi-)image Guided Multi-image Generation Distinct from previous benchmarks [9, 17, 20, 23, 43, 45], IDEA-Bench introduces tasks that require generating multiple related images simultaneously. These tasks cater to professional demands such as creating series images around a subject, multi-view image generation, and illustrating storybooks. Unlike video generation [59, 65], these image tasks involve significant visual differences while maintaining specific correlations. We further subdivide multi-image" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18543" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 93, + 69, + 518, + 293 + ], + "blocks": [ + { + "bbox": [ + 93, + 69, + 518, + 293 + ], + "lines": [ + { + "bbox": [ + 93, + 69, + 518, + 293 + ], + "spans": [ + { + "bbox": [ + 93, + 69, + 518, + 293 + ], + "type": "image", + "image_path": "41f2b111011b1cd15e697054f395155926c177b2d013f66e19b52e15a655dd50.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 298, + 555, + 333 + ], + "lines": [ + { + "bbox": [ + 54, + 298, + 555, + 333 + ], + "spans": [ + { + "bbox": [ + 54, + 298, + 555, + 333 + ], + "type": "text", + "content": "Figure 2. Dataset construction process of IDEA-Bench. We categorize the task data from professional design websites and designers based on generative model capabilities and assign capability keywords to each category. For each specific task, we design image generation prompts and hierarchical evaluation questions. Evaluators then refine these evaluation questions on a representative subset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 342, + 296, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 296, + 392 + ], + "type": "text", + "content": "generation into text-guided and (multi-)image-guided categories to address models' limitations in handling diverse input modalities, thereby testing their ability to maintain detail, style, and content consistency." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 397, + 157, + 408 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 397, + 157, + 408 + ], + "spans": [ + { + "bbox": [ + 55, + 397, + 157, + 408 + ], + "type": "text", + "content": "3.2. Data Annotation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 415, + 295, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 415, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 54, + 415, + 295, + 569 + ], + "type": "text", + "content": "Prompt Generation After collecting all tasks and corresponding images, annotators provide clear task definitions for each specific case, which, along with input images, are fed to GPT-4o [33] to generate image prompts tailored to each case. To ensure consistent evaluation across tasks, we also use GPT-4o to create a standardized set of six evaluation questions per task, ranging from basic to advanced difficulty levels. Specifically, recognizing the inherent subjectivity in both MLLM-based and human evaluations, we adopt a binary scoring system (0 or 1) for each question. This approach draws from methodologies in recent multimodal benchmarks [1, 27, 31, 64], where multiple-choice formats are used to mitigate interpretive bias." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 570, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 570, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 570, + 296, + 714 + ], + "type": "text", + "content": "To further enhance objectivity and reduce variability, we modify the typical multi-level scoring scale found in existing datasets and models, converting it into a set of six binary (true/false) questions. Each question is accompanied by clearly defined scoring criteria, with 0 indicating unmet standards and 1 representing success. For deeper granularity, questions 1-2 assess the model's understanding of basic task requirements, questions 3-4 evaluate the quality of task completion, and questions 5-6 examine the model's attention to detail and aesthetic quality in the generated images. This structured assessment methodology ensures a rigorous and objective evaluation framework, aligning with" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 342, + 555, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 555, + 367 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 555, + 367 + ], + "type": "text", + "content": "professional standards and reducing potential sources of instability in scoring." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 388, + 555, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 555, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 555, + 496 + ], + "type": "text", + "content": "Prompt Refinement However, despite our efforts to define evaluation questions as objectively as possible, multimodal large language models (MLLMs) [28, 33, 49, 52] still struggle to match human evaluators on complex, real-world tasks, limited by current models' gaps in image comprehension relative to human capabilities. To address this, we select a subset of tasks and refine the evaluation questions based on specific examples, ensuring that MLLMs can yield intuitively reasonable results." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 497, + 554, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 554, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 554, + 605 + ], + "type": "text", + "content": "One challenge in using MLLMs as evaluators is the sensitivity of models like GPT-4V/GPT-4o to the order of image presentation [54, 56, 63, 66], which particularly makes these models less reliable for consistent comparative evaluation. MLLMs also fail to identify the sequence of multiple input images, whether it is input separately for multiple images or input as a collage. To mitigate this, we manually fine-tune evaluation questions for each case within the task subset, ensuring alignment with human intuition." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "For example, in the task of generating a children's storybook with defined character images, we refine each evaluation angle by randomly sampling pairs of input and output images to test consistency. A typical question might ask if the primary character in the first input image appears with consistent features in the third output image. For testing, we use the Gemini 1.5 Pro [49] to score model outputs and compared these results against human ratings. In cases where discrepancies arose, annotators iteratively refine the" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18544" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 71, + 544, + 336 + ], + "blocks": [ + { + "bbox": [ + 67, + 71, + 544, + 336 + ], + "lines": [ + { + "bbox": [ + 67, + 71, + 544, + 336 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 544, + 336 + ], + "type": "image", + "image_path": "4d164516efce2cb7db21f9fa56239a46dc430819a5ebc91c7b1adb05cf62e940.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 338, + 555, + 361 + ], + "lines": [ + { + "bbox": [ + 55, + 338, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 555, + 361 + ], + "type": "text", + "content": "Figure 3. Visualization of model generated images on IDEA-Bench. We present the generation results in representative task examples, including text to image, image to image, and multi-image to image. In the case of no input image or no generated image, use " + }, + { + "bbox": [ + 55, + 338, + 555, + 361 + ], + "type": "inline_equation", + "content": "\\nabla" + }, + { + "bbox": [ + 55, + 338, + 555, + 361 + ], + "type": "text", + "content": " instead." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 435, + 296, + 551 + ], + "blocks": [ + { + "bbox": [ + 55, + 369, + 295, + 434 + ], + "lines": [ + { + "bbox": [ + 55, + 369, + 295, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 295, + 434 + ], + "type": "text", + "content": "Table 2. Experimental results on all categories of IDEA-Bench. Each task category is averaged across all its subtasks, with the top-ranked model scores for each task type highlighted in bold. Task types that a model cannot support are marked with “-” and are treated as 0 points in the average score calculation. “+” represents the use of MLLM for prompt rephrasing." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 435, + 296, + 551 + ], + "lines": [ + { + "bbox": [ + 58, + 435, + 296, + 551 + ], + "spans": [ + { + "bbox": [ + 58, + 435, + 296, + 551 + ], + "type": "table", + "html": "
MethodParam.Scores on All CategoriesAvg. Score
T2II2IIs2IT2IsI(s)2Is
FLUX-1† [25]12B46.0612.134.8920.1529.1722.48
DALL-E 3† [40]12B24.346.955.2714.3614.4413.07
SD3† [14]2B24.0410.794.6921.5913.0614.83
Pixart† [7]0.6B14.447.753.4817.4621.3912.90
InstructPix2Pix [5]1B-17.58---3.52
MagicBrush [60]1B-19.07---3.81
Anole [8]7B0.000.640.001.740.000.48
Emu2 [47]37B17.987.058.98--6.81
OmniGen [57]3.8B21.418.172.77--6.47
Emu2† [47]37B17.987.058.9815.5312.7812.46
OmniGen† [57]3.8B21.418.172.7723.5221.3915.45
", + "image_path": "267475ba11f132851f2c7f9b0772beab75e6309921a4efeb46b5cfd74b73e5a1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 559, + 296, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 559, + 296, + 620 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 296, + 620 + ], + "type": "text", + "content": "language of the evaluation questions until the MLLM evaluations align with human judgment. This approach enhances the robustness of our benchmark, ensuring that evaluations reflect both task-specific accuracy and consistency across diverse visual requirements." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 630, + 129, + 642 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 129, + 642 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 129, + 642 + ], + "type": "text", + "content": "3.3. Evaluation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 649, + 165, + 660 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 649, + 165, + 660 + ], + "spans": [ + { + "bbox": [ + 55, + 649, + 165, + 660 + ], + "type": "text", + "content": "3.3.1. Human Evaluation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "Evaluation Metric Each subtask in IDEA-Bench comprises 2 to 5 specific cases, with all cases within a subtask being evaluated using the same set of six binary (0 or 1) judgment questions. These evaluation questions are orga" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 370, + 555, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 370, + 555, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 370, + 555, + 502 + ], + "type": "text", + "content": "nized into three hierarchical levels, reflecting the priority of task completion over mere aesthetic quality. For each case, scoring progresses sequentially through these levels. If a model does not achieve a perfect score at a lower level, the scores for all higher levels are automatically set to 0. This hierarchical approach ensures that while a model may generate high-quality and visually appealing images, it must also fully meet the task requirements to receive a high overall score. This methodology aligns the evaluation criteria with human design standards, emphasizing the importance of fulfilling task objectives." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 505, + 556, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 505, + 556, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 505, + 556, + 628 + ], + "type": "text", + "content": "Formally, let the dataset be denoted as " + }, + { + "bbox": [ + 313, + 505, + 556, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{\\mathcal{D}_c = \\{\\mathcal{T}_t\\}_{t=1}^{T_c}\\}_{c=1}^C" + }, + { + "bbox": [ + 313, + 505, + 556, + 628 + ], + "type": "text", + "content": ". Each subtask " + }, + { + "bbox": [ + 313, + 505, + 556, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_t" + }, + { + "bbox": [ + 313, + 505, + 556, + 628 + ], + "type": "text", + "content": " consists of 2 to 5 specific cases, all evaluated using the same set of six binary judgment questions. The score for each subtask is calculated by averaging the binary scores of its cases hierarchically. Subsequently, each major category " + }, + { + "bbox": [ + 313, + 505, + 556, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_c" + }, + { + "bbox": [ + 313, + 505, + 556, + 628 + ], + "type": "text", + "content": " receives its score by averaging the scores of its subtasks and converting the result to a percentage. The overall score for the entire dataset " + }, + { + "bbox": [ + 313, + 505, + 556, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 505, + 556, + 628 + ], + "type": "text", + "content": " is then determined by averaging the scores of all " + }, + { + "bbox": [ + 313, + 505, + 556, + 628 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 505, + 556, + 628 + ], + "type": "text", + "content": " major categories." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "Importantly, if a model fails to complete a specific task, it is assigned a score of 0 for that task, ensuring that incomplete performances are accurately reflected in the overall evaluation. This hierarchical scoring framework ensures a precise and objective assessment of model performance, closely aligning with the rigorous standards expected by professional designers." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18545" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 93, + 93, + 515, + 169 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 555, + 92 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 555, + 92 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 555, + 92 + ], + "type": "text", + "content": "Table 3. Experimental results on Text-to-Image. Each task category is averaged across all its subtasks, with the top-ranked model scores for each subtask highlighted in bold." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 93, + 93, + 515, + 169 + ], + "lines": [ + { + "bbox": [ + 93, + 93, + 515, + 169 + ], + "spans": [ + { + "bbox": [ + 93, + 93, + 515, + 169 + ], + "type": "table", + "html": "
MethodParam.Subtasks ScoreAvg. Score
Arch. StyleBus. CardGame UIInf. ChartInt.Paint.Sculp.TicketLand.LOGOPoster
FLUX-1 [25]12B100.0038.895.560.0066.6761.1116.6716.6783.3361.1156.6746.06
DALL-E 3 [40]12B22.220.000.000.0011.11100.0011.110.0038.8961.1123.3324.34
Stable Diffusion 3 [14]2B38.890.005.560.0050.005.5616.670.0050.0061.1136.6724.04
Pixart [7]0.6B5.560.000.000.0022.2233.3316.670.0022.2222.2236.6714.44
Anole [8]7B0.000.000.000.000.000.000.000.000.000.000.000.00
Emu2 [47]37B22.220.000.000.0038.8938.8916.670.0044.4416.6720.0017.98
OmniGen [57]3.8B38.890.000.000.0027.7850.0016.670.0038.8933.3330.0021.41
", + "image_path": "0f84e40e79f2a4732f4560e1ef5e1bf284856127a1ecbb9da9b27a66ec90f7ba.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 56, + 197, + 553, + 284 + ], + "blocks": [ + { + "bbox": [ + 55, + 175, + 555, + 196 + ], + "lines": [ + { + "bbox": [ + 55, + 175, + 555, + 196 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 555, + 196 + ], + "type": "text", + "content": "Table 4. Experimental results on Image-to-Image. Each task category is averaged across all its subtasks, with the top-ranked model scores for each subtask highlighted in bold. " + }, + { + "bbox": [ + 55, + 175, + 555, + 196 + ], + "type": "inline_equation", + "content": "{}^{\\prime \\prime }" + }, + { + "bbox": [ + 55, + 175, + 555, + 196 + ], + "type": "text", + "content": " \" represent the use of MLLM for prompt rephrasing." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 197, + 553, + 284 + ], + "lines": [ + { + "bbox": [ + 56, + 197, + 553, + 284 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 553, + 284 + ], + "type": "table", + "html": "
MethodParam.Subtasks ScoreAvg. Score
Anim. Attr.Coup. IconID PhotoBlurCompl.Ret.Str.Light. EditEnl.Obj. EditPack. Rend.Style EditBrand Merch.
FLUX-1† [25]12B0.0016.670.002.784.170.0012.500.00100.006.250.004.1711.1112.13
DALL-E 3‡ [40]12B5.5611.110.002.780.000.004.170.0038.894.178.334.1711.116.95
Stable Diffusion 3† [14]2B8.3327.780.002.784.170.008.330.0055.564.178.334.1716.6710.79
Pixart† [7]0.6B5.5655.564.172.784.170.000.000.0011.114.170.002.0811.117.75
InstructPix2Pix [5]1B44.4416.6716.6713.8925.0054.170.000.000.0010.4225.0016.675.5617.58
MagicBrush [60]1B33.3316.6712.5025.0025.0041.670.000.000.0035.428.3333.3316.6719.07
Anole [8]7B0.000.000.000.008.330.000.000.000.000.000.000.000.000.64
Emu2 [47]37B11.110.004.175.560.000.0020.830.0016.674.178.334.1716.677.05
OmniGen [57]3.8B16.670.004.175.560.004.174.170.000.0016.670.0010.4244.448.17
", + "image_path": "66d640facf2efaeb75b10d287a8aeba5db42ab807859c74d6945a05658542590.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 290, + 180, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 290, + 180, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 290, + 180, + 300 + ], + "type": "text", + "content": "3.3.2. Automated Evaluation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 304, + 295, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 304, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 54, + 304, + 295, + 483 + ], + "type": "text", + "content": "Automated evaluation is performed on a subset of IDEA-Bench comprising 18 tasks, named IDEA-Bench-mini. During the prompt refinement phase, annotators craft unique evaluation questions for each case, in contrast to the human evaluation setup where all cases within a subtask share the same 6 evaluation questions. This tailored approach allows for more precise and context-specific assessments. Scoring is conducted using Gemini 1.5 Pro [49], adhering to the same hierarchical scoring methodology outlined in Sec. 3.3.1. To enhance result stability and reliability, each case is evaluated three times independently, and the final score is computed as the average of these three evaluations. This repetition mitigates potential variability in automated assessments and ensures consistent and robust scoring outcomes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 484, + 295, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 484, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 484, + 295, + 568 + ], + "type": "text", + "content": "By integrating both human and automated evaluation, IDEA-Bench provides a comprehensive framework for assessing the multifaceted capabilities of generative models. This dual approach not only leverages the nuanced judgment of human evaluators but also benefits from the scalability and consistency of automated tools, thereby offering a balanced and thorough evaluation of model performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 578, + 137, + 591 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 578, + 137, + 591 + ], + "spans": [ + { + "bbox": [ + 55, + 578, + 137, + 591 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 597, + 178, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 178, + 609 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 178, + 609 + ], + "type": "text", + "content": "4.1. Experimental Details" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 615, + 173, + 627 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 615, + 173, + 627 + ], + "spans": [ + { + "bbox": [ + 55, + 615, + 173, + 627 + ], + "type": "text", + "content": "4.1.1. Prompts Rephrasing" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "content": "Currently, only a few models [8, 53, 55, 58] are capable of generating multiple related images simultaneously, makes it difficult to analyse a model's ability to generate multiple images potentially. However, as DALL-E 3 [40] could perform multi-image generation owing to its MLLM GPT-4o [33], which could summarize multimodal input and then generate the corresponding prompt for each image. To empower" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 289, + 555, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 555, + 350 + ], + "type": "text", + "content": "other leading T2I models [7, 13, 25], we feed all input images and text prompt to GPT-4o and rephrase prompts for each image to be generated. By combining MLLM's multimodal understanding ability, we can approximately achieve the multi-image generation tasks with basic T2I models." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 358, + 451, + 370 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 358, + 451, + 370 + ], + "spans": [ + { + "bbox": [ + 313, + 358, + 451, + 370 + ], + "type": "text", + "content": "4.1.2. Model Reimplementation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 374, + 555, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 374, + 555, + 530 + ], + "spans": [ + { + "bbox": [ + 313, + 374, + 555, + 530 + ], + "type": "text", + "content": "We evaluate the capabilities of four categories of models in our study. For leading text-to-image (T2I) models, we select FLUX-1 [25], Stable Diffusion 3 (SD3) [13], Pixart [7], and DALL-E 3 [40]. For image editing models, we choose InstructPix2Pix [5] and MagicBrush [60]. To assess models with unified generation capabilities, we select Emu2 [47] and OmniGen [57]. For models that support interleaved text and image generation, we include Anole [8]. Notably, among the selected models, Anole [8] is the only one capable of handling unrestricted number of input and output images. We endeavor to utilize the officially released model parameters and inference settings wherever possible; further details are provided in the supplementary material." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 540, + 449, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 449, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 449, + 552 + ], + "type": "text", + "content": "4.2. Results on IDEA-Bench" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "type": "text", + "content": "We first present the ranking of all models based on their scores across the entire dataset, as detailed in Tab. 2. Leveraging GPT-4o's [33] multimodal understanding and task translation capabilities, FLUX-1 [25] achieve the highest score of 22.46, significantly outperforming other models. Specialized image editing models [5, 60] lead in the image-to-image tasks, while the general-purpose generation model Emu2 [47] excels in the images-to-image tasks. In contrast, Anole [8], which is capable of generating multiple images simultaneously, do not achieve an ideal final score. Furthermore, no model surpass a score of 10 without utilizing the comprehension capabilities of MLLMs, indicating that general-purpose image generation models still" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18546" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 129, + 93, + 478, + 168 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 92 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 92 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 92 + ], + "type": "text", + "content": "Table 5. Experimental results on Images-to-Image. Each task category is averaged across all its subtasks, with the top-ranked model scores for each subtask highlighted in bold. " + }, + { + "bbox": [ + 55, + 70, + 555, + 92 + ], + "type": "inline_equation", + "content": "{}^{\\prime \\prime } \\dagger" + }, + { + "bbox": [ + 55, + 70, + 555, + 92 + ], + "type": "text", + "content": " \" represents the use of MLLM for prompt rephrasing." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 129, + 93, + 478, + 168 + ], + "lines": [ + { + "bbox": [ + 129, + 93, + 478, + 168 + ], + "spans": [ + { + "bbox": [ + 129, + 93, + 478, + 168 + ], + "type": "table", + "html": "
MethodParam.Subtasks ScoreAvg. Score
Human Attr.Image Trans.Pan.Text Edit.3D. EffectImage Blend.Spec. Effect
FLUX-1† [25]3.976.670.0010.420.0012.500.005.564.89
DALL-E 3† [40]4.763.338.336.250.008.335.565.565.27
SD3† [14]4.765.000.008.330.008.335.565.564.69
Pixart† [7]4.765.008.330.000.004.170.005.563.48
Anole [8]0.000.000.000.000.000.000.000.000.00
Emu2 [47]6.350.0050.002.080.004.170.009.268.98
OmniGen [57]1.590.000.002.080.000.0011.117.412.77
", + "image_path": "077f6e123bc061338038760e2e427ab8d694f2743b59932a101225655ef83a74.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 56, + 195, + 555, + 253 + ], + "blocks": [ + { + "bbox": [ + 55, + 173, + 555, + 195 + ], + "lines": [ + { + "bbox": [ + 55, + 173, + 555, + 195 + ], + "spans": [ + { + "bbox": [ + 55, + 173, + 555, + 195 + ], + "type": "text", + "content": "Table 6. Experimental results on Text-to-Images. Each task category is averaged across all its subtasks, with the top-ranked model scores for each subtask highlighted in bold. " + }, + { + "bbox": [ + 55, + 173, + 555, + 195 + ], + "type": "inline_equation", + "content": "{}^{\\prime \\prime } \\dagger" + }, + { + "bbox": [ + 55, + 173, + 555, + 195 + ], + "type": "text", + "content": " \" represents the use of MLLM for prompt rephrasing." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 195, + 555, + 253 + ], + "lines": [ + { + "bbox": [ + 56, + 195, + 555, + 253 + ], + "spans": [ + { + "bbox": [ + 56, + 195, + 555, + 253 + ], + "type": "table", + "html": "
MethodParam.Subtasks ScoreAvg. Score
Anim. Grow.Child. BookDraw. Proc.Hist. Narr.Movie ShotsPhys. LawsPlant Grow.Style GroupConc. Vis.Coup. IconDyn. Char.
FLUX-1† [25]12B25.004.178.3370.834.174.178.337.0816.6750.0022.9220.15
DALL-E 3† [40]12B0.004.170.004.174.170.0041.6732.9212.5041.6716.6714.36
Stable Diffusion 3† [14]2B8.334.1725.0037.5012.508.3345.8318.7516.6741.6718.7521.59
Pixart† [7]0.6B37.504.174.1716.670.000.0029.1721.2516.6745.8316.6717.46
Anole [8]7B0.000.000.000.000.000.000.000.428.338.332.081.74
", + "image_path": "7b1d3a7708790b2e018d0c5d724668fdde848056017da4218cb13f70483a7488.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 258, + 296, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 296, + 319 + ], + "type": "text", + "content": "face substantial challenges in achieving professional-grade performance. This underscores the need for comprehensive evaluation standards to guide the development of more specialized and capable generative models. The following section provides a detailed analysis of each category." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 332, + 296, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 332, + 296, + 524 + ], + "spans": [ + { + "bbox": [ + 54, + 332, + 296, + 524 + ], + "type": "text", + "content": "Qualitative Analysis Fig. 3 showcases the generated images across 7 models. For ease of comparison, we select several cases from three categories: text-to-image, image-to-image, and multi-image-to-image. In fundamental text-to-image tasks, FLUX-1 [25] demonstrates a clear advantage over other models, effectively comprehending the visual elements and professional-level requirements outlined in the prompts. Regarding image editing tasks, models that are not specifically designed for image editing struggle to preserve essential information from the original image. Emu2 [47] is the only model that roughly restores the shapes and compositions of objects from the original image. OmniGen [57] excels in maintaining the identity of objects. However, for more complex tasks such as special effect synthesis, no current model sufficiently understands the intricate demands of these professional-level requirements." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 529, + 191, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 191, + 541 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 191, + 541 + ], + "type": "text", + "content": "4.2.1. Results on Text-to-Image" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 544, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 604 + ], + "type": "text", + "content": "All models' results on text to image tasks are presented in Tab. 3. Notably, the FLUX-1 [25] model significantly outperforms the others, effectively accomplishing the task objectives in the majority of tasks. Based on the task settings and each model's performance, we observe the following:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "type": "text", + "content": "- Text Generation Challenges: Text generation tasks are a common pain point for all models. Even though models can generate specific visual elements with high quality and understand style instructions specified in the prompts, they struggle to incorporate text into images as seamlessly as human designers. For example, in the business card generation task, only FLUX-1 [25] received scores." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "- Performance of Basic vs. General Models: Basic text-to-image models generally demonstrate higher task" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 321, + 258, + 555, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 258, + 555, + 330 + ], + "spans": [ + { + "bbox": [ + 321, + 258, + 555, + 330 + ], + "type": "text", + "content": "completion rates compared to general models like Emu2 [47] and OmniGen [57]. This is expected, as general models prioritize enhancing task generalization and multimodal information processing capabilities, which inevitably compromises some aspects of image generation quality." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 331, + 555, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 331, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 314, + 331, + 555, + 415 + ], + "type": "text", + "content": "- Poor Performance on Information Charts: All models perform poorly on the information chart generation task. The primary reason is that generative models struggle to accurately transfer large amounts of text from the input to the output image, which is a basic ability to human designers. Additionally, many models do not support ultra-long text inputs." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 421, + 542, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 421, + 542, + 433 + ], + "spans": [ + { + "bbox": [ + 313, + 421, + 542, + 433 + ], + "type": "text", + "content": "4.2.2. Results on Image-to-Image & Images-to-Image" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 436, + 554, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 436, + 554, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 436, + 554, + 532 + ], + "type": "text", + "content": "The image-to-image tasks represent the strengths of image editing models. In Tab. 4, MagicBrush [60] and Instruct-Pix2Pix [5] achieve the first and second highest evaluation scores, respectively. The images-to-image tasks involve more complex and diverse guidance or conditions, with Emu2 [47] attaining the highest score of 8.94, as shown in Tab. 5. From these results, we derive the following key insights:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 534, + 556, + 714 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 314, + 534, + 555, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 534, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 314, + 534, + 555, + 617 + ], + "type": "text", + "content": "- Pros and Cons of Image Editing Models: In local editing tasks such as image blur and image retouching, image editing models excel at preserving the original features of the image. However, these models are typically trained or fine-tuned on specific image editing datasets, which diminishes their ability to perform subject-driven tasks, such as branded merchandise generation." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 618, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 618, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 314, + 618, + 556, + 714 + ], + "type": "text", + "content": "- In-Context Learning Capabilities: Emu2 [47] and OmniGen [57] exhibit in-context learning abilities, as evidenced by OmniGen's significantly higher scores in tasks like branded merchandise generation and special effect synthesis, achieving ID consistency between input and output images. Emu2 demonstrates the ability to comprehend and partially complete the image straighten task, indicating that the model possesses an understanding of" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18547" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 92, + 555, + 160 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 93 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 93 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 93 + ], + "type": "text", + "content": "Table 7. Experimental results on Image(s)-to-Images. Each task category is averaged across all its subtasks, with the top-ranked model scores for each subtask highlighted in bold. " + }, + { + "bbox": [ + 55, + 70, + 555, + 93 + ], + "type": "inline_equation", + "content": "{}^{\\prime \\prime }" + }, + { + "bbox": [ + 55, + 70, + 555, + 93 + ], + "type": "text", + "content": " represents the use of MLLM for prompt rephrasing." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 92, + 555, + 160 + ], + "lines": [ + { + "bbox": [ + 56, + 92, + 555, + 160 + ], + "spans": [ + { + "bbox": [ + 56, + 92, + 555, + 160 + ], + "type": "table", + "html": "
MethodParamsSubtasks ScoreAvg. Score
Anim. Grow.Creat. Trans.Layer Decomp.Light. EffectMovie ShotsMulti-app.Multi-dec.Multi-view Trans.
FLUX-1† [25]12B25.0066.6775.008.330.000.0016.67100.0029.17
DALL-E 3† [40]12B25.0025.008.330.004.170.0016.678.3314.44
Stable Diffusion 3† [14]2B33.3350.000.000.000.000.008.330.0013.06
Pixart† [7]0.6B33.3350.000.0058.330.000.0058.330.0021.39
Anole [8]7B0.000.000.000.000.000.000.000.000.00
", + "image_path": "02b1d2754283fbbd04e281210646664c0d52058fae69d5ff77ffa88800f626f9.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 117, + 189, + 492, + 287 + ], + "blocks": [ + { + "bbox": [ + 55, + 166, + 555, + 189 + ], + "lines": [ + { + "bbox": [ + 55, + 166, + 555, + 189 + ], + "spans": [ + { + "bbox": [ + 55, + 166, + 555, + 189 + ], + "type": "text", + "content": "Table 8. Experimental results on the subset IDEA-Bench-mini. \"G.\" represents evaluate using Gemini 1.5 pro [49], \"H.\" represents human evaluation. \"†\" represents the use of MLLM for prompt rephrasing." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 117, + 189, + 492, + 287 + ], + "lines": [ + { + "bbox": [ + 117, + 189, + 492, + 287 + ], + "spans": [ + { + "bbox": [ + 117, + 189, + 492, + 287 + ], + "type": "table", + "html": "
MethodCLIPScoreEvaluation CategoryAvg. Score
T2II2IIs2IT2IsI(s)2Is
G.H.G.H.G.H.G.H.G.H.G.H.
FLUX-1† [25]0.343283.3367.0428.715.564.866.2532.2928.1337.5038.3337.4829.06
DALL-E 3† [40]0.346255.5641.1127.785.5620.375.5629.1721.8829.1733.3332.4121.49
SD3† [14]0.348356.3039.2623.618.345.555.5632.2923.9631.2529.8029.8019.76
Pixart† [7]0.343526.3027.0430.555.5629.862.0935.4223.9631.2530.0030.6817.73
InstructPix2Pix [5]0.1194--27.782.78------5.560.56
MagicBrush [60]0.1212--9.728.34------1.941.67
Emu2 [47]0.182135.8027.0443.528.3434.722.09----22.817.49
OmniGen [57]0.179946.3034.0740.2711.1125.460.00----22.416.81
", + "image_path": "6bfa36b7b4520f60674036ef8f31f8d0d9e2fa4b375b0f1de7e2a27aea7a2cf7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 63, + 297, + 295, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 297, + 295, + 368 + ], + "spans": [ + { + "bbox": [ + 63, + 297, + 295, + 368 + ], + "type": "text", + "content": "the semantic aspects of visual elements. Recent advances in Multimodal Visualization-of-Thought (MVoT) [29] suggest a promising direction for enhancing generative models' ability to interpret complex instructions, potentially improving their generalization to professional design tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 369, + 295, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 369, + 295, + 440 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 295, + 440 + ], + "type": "text", + "content": "- Challenges for Generative Models: General-purpose models perform poorly in tasks such as light condition editing. Although these tasks are frequently encountered by designers, the models require a certain level of understanding of real-world physical models, presenting a challenge for future generative models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 448, + 290, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 448, + 290, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 448, + 290, + 460 + ], + "type": "text", + "content": "4.2.3. Results on Text-to-Images & Image(s)-to-Images" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 463, + 295, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 463, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 463, + 295, + 666 + ], + "type": "text", + "content": "Tab. 6 and Tab. 7 present the evaluation results for text-to-images and select image(s)-to-images tasks, respectively. In both categories, SD3 [13] and FLUX-1 [25] achieve the highest scores. For tasks such as historical event generation, MLLMs can effectively convey detailed style requirements during the prompt rephrasing phase, resulting in relatively consistent style outcomes. However, for more complex tasks like children's book generation that involve sophisticated capabilities such as consistency and preservation over ID and style, existing models struggle to perform adequately even when assisted by MLLMs. Nonetheless, generating multiple related images holds significant value; for example, the generated movie shots can serve as guiding conditions for multi-shot video generation models. We aim for the task types designed in our benchmark to align generative models with the capabilities of human designers, thereby further advancing toward general AI." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 674, + 208, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 674, + 208, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 674, + 208, + 685 + ], + "type": "text", + "content": "4.2.4. Results on IDEA-Bench-mini" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": "Tab. 8 presents the evaluation results for a subset of 18 tasks within our benchmark, where we calculate CLIPScore [18]," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 297, + 555, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 555, + 452 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 555, + 452 + ], + "type": "text", + "content": "Gemini 1.5 Pro [49] scores, and human evaluation scores. As shown in Tab. 8, FLUX-1 [25], leveraging MLLM's prompt rephrasing capabilities, achieve the highest scores in both MLLM-based and human evaluations within this subset. In contrast, when excluding the use of MLLMs, Emu2 [47] secure the top score. These experimental results are consistent with the overall benchmark dataset findings. Additionally, compared to CLIPScore, MLLM-based evaluations more closely align with human preferences in terms of professional evaluation data and assessment criteria. This suggests that as MLLM capabilities continue to evolve, utilizing MLLMs for image generation evaluation will become increasingly stable and reliable." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 474, + 388, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 388, + 487 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 388, + 487 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "content": "We introduce IDEA-Bench, a comprehensive benchmark designed to bridge the gap between current generative model capabilities and the stringent demands of professional-grade image design. IDEA-Bench encompasses 100 professional image generation tasks, utilizing a detailed evaluation framework to ensure precise and objective assessments. Additionally, we build a representative subset IDEA-Bench-mini that facilitates automated evaluation using MLLMs. Our evaluations reveal significant gaps in existing models, particularly in handling intricate instructions and maintaining consistency in multi-image generation, underscoring the need for more advanced and versatile models. By aligning evaluation standards with the nuanced requirements of human designers, IDEA-Bench serves as a pivotal tool for guiding the development of generative models toward achieving professional-grade performance and advancing toward autonomous and sophisticated visual generation capabilities." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18548" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 165, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 165, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 165, + 85 + ], + "type": "text", + "content": "6. Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 296, + 147 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 296, + 147 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 296, + 147 + ], + "type": "text", + "content": "This work is supported by the Youth Innovation Promotion Association CAS, the Beijing Natural Science Foundation (No. JQ24024), and the National Key Laboratory of Science and Technology on Space-Born Intelligent Information Processing Project (No. TJ-04-23-02)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 167, + 115, + 180 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 167, + 115, + 180 + ], + "spans": [ + { + "bbox": [ + 56, + 167, + 115, + 180 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 187, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 187, + 296, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 187, + 296, + 242 + ], + "spans": [ + { + "bbox": [ + 61, + 187, + 296, + 242 + ], + "type": "text", + "content": "[1] Shuai Bai, Shusheng Yang, Jinze Bai, Peng Wang, Xingxuan Zhang, Junyang Lin, Xinggang Wang, Chang Zhou, and Jingren Zhou. Touchstone: Evaluating vision-language models by language models. arXiv preprint arXiv:2308.16890, 2023. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 243, + 296, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 243, + 296, + 308 + ], + "spans": [ + { + "bbox": [ + 61, + 243, + 296, + 308 + ], + "type": "text", + "content": "[2] Eslam Mohamed Bakr, Pengzhan Sun, Xiaogian Shen, Faizan Farooq Khan, Li Erran Li, and Mohamed Elhoseiny. Hrs-bench: Holistic, reliable and scalable benchmark for text-to-image models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20041-20053, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 310, + 295, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 310, + 295, + 354 + ], + "spans": [ + { + "bbox": [ + 61, + 310, + 295, + 354 + ], + "type": "text", + "content": "[3] Jason Baldridge, Jakob Bauer, Mukul Bhutani, Nicole Brichtova, Andrew Bunner, Kelvin Chan, Yichang Chen, Sander Dieleman, Yuqing Du, Zach Eaton-Rosen, et al. Imagen 3. arXiv preprint arXiv:2408.07009, 2024. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 355, + 294, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 355, + 294, + 409 + ], + "spans": [ + { + "bbox": [ + 61, + 355, + 294, + 409 + ], + "type": "text", + "content": "[4] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.openai.com/papers/dall-e-3.pdf, 2(3):8, 2023. 1, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 411, + 296, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 411, + 296, + 465 + ], + "spans": [ + { + "bbox": [ + 62, + 411, + 296, + 465 + ], + "type": "text", + "content": "[5] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 1, 2, 5, 6, 7, 8, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 467, + 294, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 467, + 294, + 489 + ], + "spans": [ + { + "bbox": [ + 62, + 467, + 294, + 489 + ], + "type": "text", + "content": "[6] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 491, + 295, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 491, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 62, + 491, + 295, + 544 + ], + "type": "text", + "content": "[7] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-alpha: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 1, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 545, + 295, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 545, + 295, + 589 + ], + "spans": [ + { + "bbox": [ + 62, + 545, + 295, + 589 + ], + "type": "text", + "content": "[8] Ethan Chern, Jiadi Su, Yan Ma, and Pengfei Liu. Anole: An open, autoregressive, native large multimodal models for interleaved image-text generation. arXiv preprint arXiv:2407.06135, 2024. 1, 2, 5, 6, 7, 8, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 590, + 295, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 590, + 295, + 644 + ], + "spans": [ + { + "bbox": [ + 62, + 590, + 295, + 644 + ], + "type": "text", + "content": "[9] Jaemin Cho, Abhay Zala, and Mohit Bansal. Dall-eval: Probing the reasoning skills and social biases of text-to-image generation models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3043-3054, 2023. 1, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 58, + 647, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 647, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 58, + 647, + 295, + 689 + ], + "type": "text", + "content": "[10] Min Jin Chong and David Forsyth. Effectively unbiased fid and inception score and where to find them. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6070-6079, 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "text", + "content": "[11] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "text", + "content": "Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 1, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 107, + 553, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 152 + ], + "type": "text", + "content": "[12] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 1, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 153, + 553, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 153, + 553, + 218 + ], + "spans": [ + { + "bbox": [ + 316, + 153, + 553, + 218 + ], + "type": "text", + "content": "[13] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 1, 3, 6, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 220, + 553, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 296 + ], + "type": "text", + "content": "[14] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, Kyle Lacey, Alex Goodwin, Yannik Marek, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis, 2024. 2, 5, 6, 7, 8, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 297, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 352 + ], + "type": "text", + "content": "[15] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 354, + 553, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 397 + ], + "type": "text", + "content": "[16] Yuying Ge, Sijie Zhao, Ziyun Zeng, Yixiao Ge, Chen Li, Xintao Wang, and Ying Shan. Making llama see and draw with seed tokenizer. arXiv preprint arXiv:2310.01218, 2023. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 399, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 443 + ], + "type": "text", + "content": "[17] Dhruba Ghosh, Hannaneh Hajishirzi, and Ludwig Schmidt. Geneval: An object-focused framework for evaluating text-to-image alignment. Advances in Neural Information Processing Systems, 36, 2024. 1, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 445, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 553, + 487 + ], + "type": "text", + "content": "[18] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 2, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 490, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 490, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 317, + 490, + 553, + 533 + ], + "type": "text", + "content": "[19] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 535, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 535, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 317, + 535, + 553, + 600 + ], + "type": "text", + "content": "[20] Yushi Hu, Benlin Liu, Jungo Kasai, Yizhong Wang, Mari Ostendorf, Ranjay Krishna, and Noah A Smith. Tifa: Accurate and interpretable text-to-image faithfulness evaluation with question answering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20406-20417, 2023. 1, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 602, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 602, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 602, + 553, + 656 + ], + "type": "text", + "content": "[21] Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. Advances in Neural Information Processing Systems, 36:78723-78747, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "text", + "content": "[22] Lianghua Huang, Wei Wang, Zhi-Fan Wu, Huanzhang Dou, Yupeng Shi, Yutong Feng, Chen Liang, Yu Liu, and Jingren Zhou. Group diffusion transformers are unsupervised multitask learners. arXiv preprint arXiv:2410.15027, 2024. 1, 3, 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18549" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 117 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 117 + ], + "type": "text", + "content": "[23] Max Ku, Tianle Li, Kai Zhang, Yujie Lu, Xingyu Fu, Wenwen Zhuang, and Wenhu Chen. Imagenhub: Standardizing the evaluation of conditional image generation models. arXiv preprint arXiv:2310.01596, 2023. 1, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 119, + 294, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 119, + 294, + 152 + ], + "spans": [ + { + "bbox": [ + 56, + 119, + 294, + 152 + ], + "type": "text", + "content": "[24] Black Forest Labs. Flux: Inference repository. https://github.com/black-forest-labs/flux, 2024. Accessed: 2024-10-25. 1, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 153, + 260, + 165 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 153, + 260, + 165 + ], + "spans": [ + { + "bbox": [ + 56, + 153, + 260, + 165 + ], + "type": "text", + "content": "[25] Black Forest Labs. Flux.1, 2024. 2, 5, 6, 7, 8, 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 167, + 296, + 221 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 167, + 296, + 221 + ], + "spans": [ + { + "bbox": [ + 56, + 167, + 296, + 221 + ], + "type": "text", + "content": "[26] Tony Lee, Michihiro Yasunaga, Chenlin Meng, Yifan Mai, Joon Sung Park, Agrim Gupta, Yunzhi Zhang, Deepak Narayanan, Hannah Teufel, Marco Bellagente, et al. Holistic evaluation of text-to-image models. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 223, + 294, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 223, + 294, + 266 + ], + "spans": [ + { + "bbox": [ + 56, + 223, + 294, + 266 + ], + "type": "text", + "content": "[27] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 269, + 294, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 269, + 294, + 313 + ], + "spans": [ + { + "bbox": [ + 56, + 269, + 294, + 313 + ], + "type": "text", + "content": "[28] Bo Li, Kaichen Zhang, Hao Zhang, Dong Guo, Renrui Zhang, Feng Li, Yuanhan Zhang, Ziwei Liu, and Chunyuan Li. Llava-last: Stronger llms supercharge multimodal capabilities in the wild, 2024. 1, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 316, + 294, + 359 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 294, + 359 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 294, + 359 + ], + "type": "text", + "content": "[29] Chengzhu Li, Wenshan Wu, Huanyu Zhang, Yan Xia, Shaoguang Mao, Li Dong, Ivan Vulic, and Furu Wei. Imagine while reasoning in space: Multimodal visualization-of-thought. arXiv preprint arXiv:2501.07542, 2025. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 361, + 294, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 294, + 405 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 294, + 405 + ], + "type": "text", + "content": "[30] Kevin Lin, Zhengyuan Yang, Linjie Li, Jianfeng Wang, and Lijuan Wang. Designbench: Exploring and benchmarking dall-e 3 for imagining visual design. arXiv preprint arXiv:2310.15144, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 407, + 294, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 407, + 294, + 462 + ], + "spans": [ + { + "bbox": [ + 56, + 407, + 294, + 462 + ], + "type": "text", + "content": "[31] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? In European Conference on Computer Vision, pages 216-233. Springer, 2025. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 464, + 294, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 464, + 294, + 518 + ], + "spans": [ + { + "bbox": [ + 56, + 464, + 294, + 518 + ], + "type": "text", + "content": "[32] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4296-4304, 2024. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 520, + 216, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 520, + 216, + 532 + ], + "spans": [ + { + "bbox": [ + 56, + 520, + 216, + 532 + ], + "type": "text", + "content": "[33] OpenAI. Gpt-4o, 2024. 1, 3, 4, 6, 2, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 533, + 294, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 533, + 294, + 599 + ], + "spans": [ + { + "bbox": [ + 56, + 533, + 294, + 599 + ], + "type": "text", + "content": "[34] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730-27744, 2022. 1, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 601, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 294, + 656 + ], + "type": "text", + "content": "[35] Vitali Petsiuk, Alexander E Siemann, Saisamrit Surbehera, Zad Chin, Keith Tyser, Gregory Hunter, Arvind Raghavan, Yann Hicke, Bryan A Plummer, Ori Kerret, et al. Human evaluation of text-to-image models on a multi-task benchmark. arXiv preprint arXiv:2211.12112, 2022. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[36] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 1, 2, 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 115 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 115 + ], + "type": "text", + "content": "[37] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 1, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 173 + ], + "type": "text", + "content": "[38] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 1, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 175, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 228 + ], + "type": "text", + "content": "[39] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International conference on machine learning, pages 8821-8831. Pmlr, 2021. 1, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 230, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 553, + 274 + ], + "type": "text", + "content": "[40] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1, 2, 5, 6, 7, 8, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 276, + 555, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 555, + 331 + ], + "type": "text", + "content": "[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 332, + 553, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 553, + 376 + ], + "type": "text", + "content": "[42] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 377, + 553, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 377, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 316, + 377, + 553, + 442 + ], + "type": "text", + "content": "[43] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 1, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 445, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 445, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 445, + 553, + 510 + ], + "type": "text", + "content": "[44] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 1, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 512, + 553, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 512, + 553, + 577 + ], + "spans": [ + { + "bbox": [ + 316, + 512, + 553, + 577 + ], + "type": "text", + "content": "[45] Shelly Sheynin, Adam Polyak, Uriel Singer, Yuval Kirstain, Amit Zohar, Oron Ashual, Devi Parikh, and Yaniv Taigman. Emu edit: Precise image editing via recognition and generation tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8871-8879, 2024. 1, 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 578, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 578, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 578, + 553, + 633 + ], + "type": "text", + "content": "[46] Quan Sun, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, Yueze Wang, Hongcheng Gao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Emu: Generative pretraining in multimodality. In The Twelfth International Conference on Learning Representations, 2023. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 635, + 553, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 700 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 700 + ], + "type": "text", + "content": "[47] Quan Sun, Yufeng Cui, Xiaosong Zhang, Fan Zhang, Qiying Yu, Yueze Wang, Yongming Rao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative multimodal models are in-context learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14398-14409, 2024. 1, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 702, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 702, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 702, + 553, + 713 + ], + "type": "text", + "content": "[48] Chameleon Team. Chameleon: Mixed-modal early-fusion" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18550" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "text", + "content": "foundation models. arXiv preprint arXiv:2405.09818, 2024.3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 149 + ], + "type": "text", + "content": "[49] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 1, 4, 6, 8, 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 151, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 151, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 151, + 294, + 205 + ], + "type": "text", + "content": "[50] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 206, + 294, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 294, + 259 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 294, + 259 + ], + "type": "text", + "content": "[51] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 261, + 294, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 294, + 314 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 294, + 314 + ], + "type": "text", + "content": "[52] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 316, + 294, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 294, + 369 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 294, + 369 + ], + "type": "text", + "content": "[53] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 1, 2, 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 372, + 294, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 372, + 294, + 414 + ], + "spans": [ + { + "bbox": [ + 56, + 372, + 294, + 414 + ], + "type": "text", + "content": "[54] Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A Smith, Daniel Khashabi, and Hannaneh Hajishirzi. Self-instruct: Aligning language models with self-generated instructions. arXiv preprint arXiv:2212.10560, 2022. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 415, + 294, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 294, + 458 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 294, + 458 + ], + "type": "text", + "content": "[55] Shengqiong Wu, Hao Fei, Leigang Qu, Wei Ji, and Tat-Seng Chua. Next-gpt: Any-to-any multimodal llm. In Proceedings of the International Conference on Machine Learning, pages 53366-53397, 2024. 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 460, + 294, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 294, + 523 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 294, + 523 + ], + "type": "text", + "content": "[56] Tong Wu, Guandao Yang, Zhibing Li, Kai Zhang, Ziwei Liu, Leonidas Guibas, Dahua Lin, and Gordon Wetzstein. Gpt-4v (ision) is a human-aligned evaluator for text-to-3d generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22227-22238, 2024. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 526, + 294, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 526, + 294, + 569 + ], + "spans": [ + { + "bbox": [ + 56, + 526, + 294, + 569 + ], + "type": "text", + "content": "[57] Shitao Xiao, Yueze Wang, Junjie Zhou, Huaying Yuan, Xingrun Xing, Ruiran Yan, Shuting Wang, Tiejun Huang, and Zheng Liu. Omnigen: Unified image generation. arXiv preprint arXiv:2409.11340, 2024. 3, 5, 6, 7, 8, 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 571, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 571, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 571, + 294, + 613 + ], + "type": "text", + "content": "[58] Shuai Yang, Yuying Ge, Yang Li, Yukang Chen, Yixiao Ge, Ying Shan, and Yingcong Chen. Seed-story: Multimodal long story generation with large language model. arXiv preprint arXiv:2407.08683, 2024. 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 614, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 294, + 667 + ], + "type": "text", + "content": "[59] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024.3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[60] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. Advances in Neural Information Processing Systems, 36, 2024. 2, 5, 6, 7, 8, 1, 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 439 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[61] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 171 + ], + "type": "text", + "content": "[62] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 1, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 174, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 553, + 228 + ], + "type": "text", + "content": "[63] Xinlu Zhang, Yujie Lu, Weizhi Wang, An Yan, Jun Yan, Lianke Qin, Heng Wang, Xifeng Yan, William Yang Wang, and Linda Ruth Petzold. Gpt-4v (ision) as a generalist evaluator for vision-language tasks. arXiv preprint arXiv:2311.01361, 2023. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 230, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 553, + 293 + ], + "type": "text", + "content": "[64] Yi-Fan Zhang, Huanyu Zhang, Haochen Tian, Chaoyou Fu, Shuangqing Zhang, Junfei Wu, Feng Li, Kun Wang, Qing-song Wen, Zhang Zhang, et al. Mme-realworld: Could your multimodal llm challenge high-resolution real-world scenarios that are difficult for humans? arXiv preprint arXiv:2408.13257, 2024. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 297, + 553, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 328 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 328 + ], + "type": "text", + "content": "[65] Xuanlei Zhao, Xiaolong Jin, Kai Wang, and Yang You. Real-time video generation with pyramid attention broadcast. arXiv preprint arXiv:2408.12588, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 331, + 553, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 384 + ], + "type": "text", + "content": "[66] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in Neural Information Processing Systems, 36:46595-46623, 2023. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 386, + 553, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 386, + 553, + 439 + ], + "spans": [ + { + "bbox": [ + 316, + 386, + 553, + 439 + ], + "type": "text", + "content": "[67] Chunting Zhou, Lili Yu, Arun Babu, Kushal Tirumala, Michihiro Yasunaga, Leonid Shamis, Jacob Kahn, Xuezhe Ma, Luke Zettlemoyer, and Omer Levy. Transfusion: Predict the next token and diffuse images with one multi-modal model, 2024. 3" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18551" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/IDEA_ Inverted Text with Cooperative Deformable Aggregation for Multi-modal Object Re-Identification/b7176393-d8d0-4ced-ba0c-43acd09bc650_content_list.json b/2025/IDEA_ Inverted Text with Cooperative Deformable Aggregation for Multi-modal Object Re-Identification/b7176393-d8d0-4ced-ba0c-43acd09bc650_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6c9f4c55751bdac48c0f9ae85491eaffcc136c1f --- /dev/null +++ b/2025/IDEA_ Inverted Text with Cooperative Deformable Aggregation for Multi-modal Object Re-Identification/b7176393-d8d0-4ced-ba0c-43acd09bc650_content_list.json @@ -0,0 +1,1675 @@ +[ + { + "type": "text", + "text": "IDEA: Inverted Text with Cooperative Deformable Aggregation for Multi-modal Object Re-Identification", + "text_level": 1, + "bbox": [ + 94, + 128, + 903, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuhao Wang, Yongfeng Lv, Pingping Zhang* Huchuan Lu", + "bbox": [ + 267, + 203, + 728, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "School of Future Technology, School of Artificial Intelligence, Dalian University of Technology", + "bbox": [ + 119, + 220, + 875, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{924973292,1vyf}@mail.dlut.edu.cn，{zhpp,lhchuan}@dlut.edu.cn", + "bbox": [ + 228, + 241, + 764, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multi-modal object Re-Identification (ReID) aims to retrieve specific objects by utilizing complementary information from various modalities. However, existing methods focus on fusing heterogeneous visual features, neglecting the potential benefits of text-based semantic information. To address this issue, we first construct three text-enhanced multi-modal object ReID benchmarks. To be specific, we propose a standardized multi-modal caption generation pipeline for structured and concise text annotations with Multi-modal Large Language Models (MLLMs). Besides, current methods often directly aggregate multi-modal information without selecting representative local features, leading to redundancy and high complexity. To address the above issues, we introduce IDEA, a novel feature learning framework comprising the Inverted Multi-modal Feature Extractor (IMFE) and Cooperative Deformable Aggregation (CDA). The IMFE utilizes Modal Prefixes and an InverseNet to integrate multi-modal information with semantic guidance from inverted text. The CDA adaptively generates sampling positions, enabling the model to focus on the interplay between global features and discriminative local features. With the constructed benchmarks and the proposed modules, our framework can generate more robust multi-modal features under complex scenarios. Extensive experiments on three multi-modal object ReID benchmarks demonstrate the effectiveness of our proposed method.", + "bbox": [ + 89, + 323, + 483, + 717 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 744, + 223, + 760 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Object Re-Identification (ReID) focuses on retrieving the same object across different camera views. While significant progress has been made with RGB images [9, 13, 23–25, 30, 35, 39, 40, 45, 47, 53], existing methods are limited under environmental challenges like dark and glare, reducing their robustness in real-world applications. Multi-modal imaging, which combines data from multiple spectra such", + "bbox": [ + 89, + 771, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bc9351cc02ca02e4551f39b1f49c788ff2b4e5e4e7c0b26d8a84a1c573598573.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 287, + 903, + 364 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/65fe39a059a8a2886de5639e383a213b6b0371f50791fd952292d1c02bd591eb.jpg", + "image_caption": [ + "Figure 1. Overall illustration of our motivations and proposed framework. (a) Our multi-modal caption generation pipeline. (b) Limitations of existing MLLM-generated captions. (c) Comparison between previous methods and our proposed IDEA." + ], + "image_footnote": [], + "bbox": [ + 516, + 368, + 900, + 429 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "as RGB, Near Infrared (NIR) and Thermal Infrared (TIR), offers a promising solution. By leveraging complementary information, multi-modal ReID methods [4, 26, 37, 54] improve feature robustness under challenging conditions.", + "bbox": [ + 511, + 507, + 905, + 568 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Beyond visual information fusion for different spectra, integrating language with vision can enhance ReID performance [11, 27, 34]. However, due to the absence of text annotations for images, most existing methods rely solely on visual features, overlooking the benefits of text-based semantic information. To address this limitation, some studies augment RGB datasets with manual text annotations [6, 20]. While this strategy boosts performance, it is both time-consuming and labor-intensive. With the advent of Multimodal Large Language Models (MLLMs), image caption has made significant progress. Recently, researchers start to use MLLMs to generate textual descriptions for RGB images [11, 14]. However, these methods face two main challenges. (1) As shown in Fig. 1 (b), the randomness in caption generation often leads to redundancy, resulting in long and complex sentences. Additionally, the text structure varies across images. In ReID tasks, such captions may exceed input length limits of a text encoder [29, 55], with redundant content diluting the essential information. (2) Current methods primarily focus on RGB images, where the generated captions lack sufficient details. However, multimodal imaging can capture crucial information in complex", + "bbox": [ + 511, + 569, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 109, + 887, + 235, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "29701", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "environments, providing enough information for MLLMs to generate informative captions. Despite this, no existing methods provide text annotations for multi-modal images. To bridge this gap, we construct three text-enhanced multi-modal object ReID benchmarks. Meanwhile, another challenge is the aggregation of multi-modal information. As shown in Fig. 1 (c), previous methods directly aggregate the heterogeneous information, leading to high computational complexity and noise interference [54]. To address above issues, we propose a multi-modal feature learning framework named IDEA to introduce Inverted text with cooperative DEformable Aggregation for multi-modal object ReID.", + "bbox": [ + 89, + 90, + 480, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Technically, we first develop a standardized pipeline for multi-modal caption generation. It extends existing datasets with text annotations across spectral modalities. Specifically, as shown in Fig. 1 (a), our caption generation pipeline consists of two steps. In the first step, we use image-template pairs in conjunction with MLLMs to generate informative captions. Then, we leverage the powerful keyword extraction capabilities of MLLMs to extract predefined attributes from the generated captions to produce structured and concise sentences. Building upon the captions, we propose the IDEA framework, which consists of two key components: the Inverted Multi-modal Feature Extractor (IMFE) and Cooperative Deformable Aggregation (CDA). First, we utilize the IMFE to extract integrated multi-modal features using Modal Prefixes and an InverseNet. Specifically, Modal Prefixes are designed to distinguish different modalities, reducing the impact of conflicting information across modalities. The InverseNet further exploits the semantic information from inverted text to enhance feature robustness. Furthermore, we propose CDA to adaptively aggregate discriminative local information. Specifically, based on the aggregated multi-modal information, we adaptively generate sampling positions to enable the model to focus on the interplay between global features and discriminative local information. Through these modules, our proposed framework effectively utilizes semantic guidance from texts while adaptively aggregating discriminative multi-modal information. Extensive experiments on three benchmark datasets demonstrate the effectiveness of our approach. In summary, our contributions are as follows:", + "bbox": [ + 91, + 275, + 483, + 729 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We construct three text-enhanced multi-modal object ReID benchmarks, providing a structured caption generation pipeline across multiple spectral modalities.", + "- We introduce a novel feature learning framework named IDEA, which includes the Inverted Multi-modal Feature Extractor (IMFE) and Cooperative Deformable Aggregation (CDA). The IMFE integrates multi-modal features using Modal Prefixes and an InverseNet, while the CDA adaptively aggregates discriminative local information.", + "- Extensive experiments on three benchmark datasets validate the effectiveness of our proposed method." + ], + "bbox": [ + 89, + 732, + 482, + 896 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 89, + 653, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Vision-language Learning in Re-Identification", + "text_level": 1, + "bbox": [ + 511, + 114, + 900, + 131 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision-language learning plays a pivotal role in multimodal applications. As the field advances [5, 48, 49], there is a growing focus on exploring the interaction between visual and textual information in ReID tasks. Existing methods can be broadly divided into cross-modal and multimodal ReID. In cross-modal ReID, Text-to-Image ReID [6] focuses on matching text queries with image galleries. Current methods primarily emphasize feature alignment [17] and pre-training [32]. To expand application scenarios, Zhai et al. [52] combine text and sketch modalities in the query, while Chen et al. [3] propose an uncertain Query-to-RGB retrieval model to handle missing modalities. More recently, Li et al. [19] introduce flexible modality combinations and unify various uncertainties in retrieval. However, these methods focus on cross-modal alignment. In contrast, multi-modal ReID aims to leverage complementary information from multiple modalities, with both the query and gallery containing multi-modal data. The rise of vision-language models, particularly CLIP [29], has significantly advanced this area by facilitating image-text interactions. Li et al. [21] pioneer CLIP-ReID to learn from text templates for ReID tasks. However, the templates are not real descriptions of images, potentially restricting performance. To address this, Han et al. [11] leverage MLLMs to generate descriptions for RGB images. Despite improvements, existing methods fail to tackle the challenges of generating text for multi-spectral modalities and ensuring structural consistency in the generated descriptions. To address the above issues, we propose a structured multi-modal caption generation pipeline, enhancing the consistency of texts and providing informative annotations for multi-modal object ReID.", + "bbox": [ + 511, + 138, + 906, + 608 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Multi-modal Object Re-Identification", + "text_level": 1, + "bbox": [ + 513, + 619, + 836, + 636 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-modal object ReID gains great attention due to its robustness in real-world applications. Research primarily focuses on leveraging complementary information from different modalities. Early works emphasize effective fusion strategies [8, 56]. To enhance modality-specific learning, Wang et al. [41] propose an interact-embed-enlarge framework to facilitate knowledge expansion. To address missing modal data, Zheng et al. [57] introduce pixel reconstruction to handle data inconsistency. Li et al. [18] use a coherence loss to guide feature fusion. Some methods further improve feature robustness with modality generation, graph learning and instance sampling [10, 12, 58]. Notably, Wang et al. [42] introduce a test-time training framework to mine inter-modal interactions. Wang et al. [37] develop TOP-ReID, which employs token permutation to guide modality fusion. Recently, Zhang et al. [54] propose diverse token selections to mitigate background noises. Despite promis", + "bbox": [ + 511, + 643, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "29702", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/63e21ece53b6cec570bc2a86cf9a1f3d839d0d538c8ce2077e1f37bc0e963d71.jpg", + "image_caption": [ + "Figure 2. The upper row presents the template used to annotate our dataset. The lower row provides detailed information about the annotated dataset. (a) Category statistics for our annotated person ReID dataset. (b) Example images and captions from the RGBNT201 dataset. (c) Category statistics for our annotated vehicle ReID dataset. (d) Example images and captions from the MSVR310 dataset." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 491, + 314 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a562de2e9af9bafa90e69e31f4a0edf4ed3a095704512f089f31924c5f357681.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 503, + 88, + 903, + 314 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ing results, they often overlook the semantic guidance of informative text annotations. Thus, we propose the IDEA framework, which incorporates generated text annotations to leverage semantic guidance. This framework adaptively aggregates discriminative local information, enhancing feature robustness in complex scenarios.", + "bbox": [ + 89, + 378, + 483, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Method", + "text_level": 1, + "bbox": [ + 89, + 482, + 264, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce the multi-modal caption generation pipeline and the proposed feature learning framework. Fig. 2 illustrates the details of caption generation, including attribute definitions and examples. Fig. 3 presents the key modules of our proposed IDEA: the Inverted Multi-modal Feature Extractor (IMFE) and the Cooperative Deformable Aggregation (CDA). Details are described as follows.", + "bbox": [ + 89, + 507, + 483, + 613 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Multi-modal Caption Generation", + "text_level": 1, + "bbox": [ + 89, + 622, + 382, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To bridge the gap in multi-modal text annotation, we propose the Multi-modal Caption Generation pipeline. To be specific, our pipeline consists of two steps. The first step generates informative descriptions, while the second step reuses the MLLMs to extract the predefined key attributes.", + "bbox": [ + 89, + 643, + 482, + 718 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Caption Generation. Taking our person annotation as an example, we define 8 attribute categories in Fig. 2 (a), including gender, belongings and so on. These attributes are mapped into the template shown in the top part of Fig. 2 (a). To tailor the description to each modality, we add specific prefixes: \"Write a comprehensive description of the person's overall appearance based on the [RGB/NIR/TIR] image, ...\" Besides, we incorporate commands to prevent the hallucination of the MLLMs [32], forming a complete prompt. This prompt, along with the corresponding image, is fed into the MLLMs to generate the textual description.", + "bbox": [ + 89, + 719, + 483, + 885 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Attribute Extraction. After obtaining an initial descrip", + "bbox": [ + 89, + 886, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tion, we feed it back into the MLLMs to extract predefined attributes. After that, we populate these attributes into the same template we used for caption generation, forming a structured and concise description. As shown in Fig. 2 (b), the descriptions for each modality can align well with the corresponding images. For instance, the \"white t-shirt\" and \"purse\" in the RGB image are accurately captured. For the vehicle dataset, we follow similar pipelines. As illustrated in Fig. 2 (c), the text of the RGB image successfully captures the license plate number \"KQQ819\". Meanwhile, the car logo is more prominent in the TIR image, with \"Hyundai\" being identified. This approach allows us to establish a unified pipeline for multi-modal text annotation, ensuring structural descriptions across different modalities.", + "bbox": [ + 511, + 378, + 906, + 592 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. IDEA Framework", + "text_level": 1, + "bbox": [ + 513, + 602, + 689, + 616 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2.1. Inverted Multi-modal Feature Extractor", + "text_level": 1, + "bbox": [ + 511, + 625, + 841, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To leverage the semantic guidance from texts, we propose the Inverted Multi-modal Feature Extractor (IMFE). Unlike previous methods [37, 54] that focus on fusing multi-modal images, we incorporate multi-modal texts into the fusion. However, as indicated by the blue keywords in Fig. 2 (b) and (c), conflicts can arise due to the nature of multi-modal images. Directly aggregating contradictory information can lead to model confusion. Thus, IMFE employs Modal Prefixes and an InverseNet to address these issues.", + "bbox": [ + 511, + 643, + 906, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Modal Prefixes. To enhance the model's awareness of different modal texts and mitigate the impact of conflicting information, we propose a simple method called Modal Prefixes. Taking the text input for the RGB branch as an example, as shown in the top left corner of Fig. 3, $T_{RGB}$ represents the text annotation of the RGB image, while $P_{RGB}$ is the modal prefix for RGB modality. $P_{RGB}$ consists of two parts: a fixed text describing the characteristics of the", + "bbox": [ + 511, + 779, + 908, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "29703", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8b3f4150cacc93f1549feb3273c5b7a40558ff630451663822c023988f5c7eb2.jpg", + "image_caption": [ + "Figure 3. Illustration of the proposed IDEA framework. The upper part depicts the Inverted Multi-modal Feature Extractor (IMFE). It employs modal prefixes and an InverseNet to incorporate semantic text guidance for feature discriminability. The lower part highlights the Cooperative Deformable Aggregation (CDA), which adaptively integrates discriminative local information with global features. With the integration of IMFE and CDA, IDEA effectively extracts discriminative multi-modal features for object ReID." + ], + "image_footnote": [], + "bbox": [ + 89, + 85, + 903, + 340 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "RGB image and learnable tokens for fine-tuning. Specifically, $P_{RGB}$ is structured as: \"An image of a XXXX person in the visible spectrum, capturing natural colors and fine details: \", where \"XXXX\" is replaced by an equal number of learnable tokens [22]. Here, we denote the number of learnable tokens as $N_P$ . Then, we concatenate $P_{RGB}$ and $T_{RGB}$ to form the text input $T_R$ for the RGB branch:", + "bbox": [ + 88, + 426, + 483, + 532 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nT _ {R} = \\left[ P _ {R G B}, T _ {R G B} \\right]. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 537, + 480, + 556 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Other modalities follow the same structure to guide the model in distinguishing between different modal texts. With this simple yet effective method, the model can better understand the characteristics of each modality and reduce conflicts during multi-modal fusion.", + "bbox": [ + 88, + 561, + 482, + 635 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "InverseNet. To fully leverage the semantic information from texts, we propose the InverseNet. Previous methods [2, 11] often reverse global image information into pseudo-text tokens and use texts as the primary modality for downstream tasks. However, due to the limitations of MLLMs and the complexity of multi-modal image annotations, the generated texts may contain some errors and conflicts. To address this, we take a different direction, reversing global text information into pseudo-image tokens. In this reversal, we also incorporate rich information from the learnable tokens in Modal Prefixes. This approach enables the interaction between semantic information and image details, resulting in a discriminative feature representation.", + "bbox": [ + 88, + 636, + 482, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, as shown in the upper part of Fig. 3, the text $T_{m}$ is first fed into the CLIP's text encoder $\\mathcal{T}$ for extracting the semantic feature $\\hat{f}_m^t\\in \\mathbb{R}^C$ as follows:", + "bbox": [ + 89, + 832, + 483, + 878 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {f} _ {m} ^ {t} = \\mathcal {T} (T _ {m}). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 236, + 883, + 482, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $m \\in \\{R, N, T\\}$ represents the RGB, NIR and TIR modalities, respectively. $C$ denotes the embedding dimension. If learnable tokens exist in the modal prefixes, $\\hat{f}_m^t$ will be updated by the average of learnable tokens. Then, we feed $\\hat{f}_m^t$ into the InverseNet $\\mathcal{I}$ to generate the pseudo token $f_m^t \\in \\mathbb{R}^{C'}$ with the following equation:", + "bbox": [ + 511, + 426, + 906, + 518 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {m} ^ {t} = \\mathcal {I} \\left(\\hat {f} _ {m} ^ {t}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 661, + 527, + 903, + 545 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $C'$ is the dimension of the pseudo token while $\\mathcal{I}$ is essentially a simple layer of MLP. Meanwhile, for the image input $I_m$ , we first patchify it into $N_l$ patches. Then, we concatenate the pseudo token $f_m^t$ , the patch tokens and a learnable token $f_m^v$ to form the image feature $\\bar{F}_m \\in \\mathbb{R}^{(N_l + 2) \\times C'}$ . Finally, we feed $\\bar{F}_m$ into the CLIP's vision encoder $\\mathcal{V}$ to obtain the integrated feature $F_m \\in \\mathbb{R}^{(N_l + 2) \\times C}$ as follows:", + "bbox": [ + 511, + 553, + 906, + 659 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF _ {m} = \\mathcal {V} (\\bar {F} _ {m}). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 658, + 667, + 903, + 685 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By incorporating the InverseNet, we can effectively leverage the semantic information from texts and enhance the discriminability of the feature representation.", + "bbox": [ + 511, + 694, + 906, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2. Cooperative Deformable Aggregation", + "text_level": 1, + "bbox": [ + 511, + 746, + 821, + 761 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To effectively extract discriminative local information and adaptively aggregate complementary multi-modal features, we propose Cooperative Deformable Aggregation (CDA). CDA addresses the noise interference by adaptively generating sampling positions of key local regions. This module enables interactions between global features and discriminative local features, ensuring efficient feature aggregation.", + "bbox": [ + 511, + 763, + 905, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Technically, as illustrated in the lower part of Fig. 3, we first extract patch tokens from the integrated feature $F_{m}$ .", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "29704", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "These patches are then reshaped to form $\\hat{F}_m\\in \\mathbb{R}^{H\\times W\\times C}$ where $H$ and $W$ represent the height and width of the local feature map, respectively. Next, we concatenate these features along the channel dimension and pass them into the Local Mixer $\\mathcal{M}$ for interaction. Specifically, $\\mathcal{M}$ comprises point-wise convolution $\\mathcal{P}$ , the GELU activation function [15] $\\delta$ and depth-wise convolution $\\mathcal{D}$ . This process yields the aggregated local features $\\hat{F}_A\\in \\mathbb{R}^{H_S\\times W_S\\times C}$ :", + "bbox": [ + 89, + 89, + 483, + 212 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {F} _ {A} = \\mathcal {M} ([ \\hat {F} _ {R}, \\hat {F} _ {N}, \\hat {F} _ {T} ]), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 215, + 482, + 234 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {M} (\\mathcal {X}) = \\delta (\\mathcal {D} (\\delta (\\mathcal {P} (\\mathcal {X})))) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 253, + 482, + 270 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $[\\cdot]$ means concatenation. $H_{S}$ and $W_{S}$ represent the number of sampled regions in the height and width dimensions, respectively. Since the convolution $\\mathcal{D}$ aggregates local information in the receptive field of the kernel, we define the center of each convolution operation as the reference points $P\\in \\mathbb{R}^{H_S\\times W_S\\times 2}$ . Based on these reference points, the aggregated local information is then used to determine the direction in which the model should shift within the local region. Thus, we send $\\hat{F}_A$ into the Offset Generator $\\mathcal{G}$ to predict the offsets $\\Delta P\\in \\mathbb{R}^{H_S\\times W_S\\times 2}$ as follows:", + "bbox": [ + 89, + 273, + 483, + 424 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta P = \\mathcal {G} (\\hat {F} _ {A}) * k, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 223, + 429, + 480, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $k$ scales the offset magnitude [44] and $\\mathcal{G}$ denotes a linear layer. Finally, we apply the offsets $\\Delta P$ to the reference points $P$ to get the sampling locations $\\hat{P} \\in \\mathbb{R}^{H_S \\times W_S \\times 2}$ , which are shared across modalities [54]:", + "bbox": [ + 89, + 452, + 483, + 513 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {P} = P + \\Delta P. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 236, + 517, + 480, + 534 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Next, we sample the local features $\\hat{F}_m$ at the positions $\\hat{P}$ using the bilinear interpolation to obtain the discriminative local features. The features are concatenated to form the sampled feature $F_S \\in \\mathbb{R}^{3N_S \\times C}$ , where $N_S = H_S \\times W_S$ represents the number of sampled regions. At this stage, we utilize the cooperative information from multiple modalities to generate deformable local discriminative features. As illustrated by the sampling points visualization in Fig. 3, the adaptive offsets effectively concentrate on critical semantic regions of the human body, enhancing the discrimination.", + "bbox": [ + 89, + 539, + 483, + 691 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To promote the interaction between global features and discriminative local features, we incorporate cross attention [7]. Specifically, we extract the visual class token and pseudo token from the integrated feature $F_{m}$ to construct the global feature $F_{G} \\in \\mathbb{R}^{6 \\times C}$ . Then, the global feature $F_{G}$ is used as the query, while the sampled local feature $F_{S}$ serves as the key and value to facilitate interaction:", + "bbox": [ + 89, + 691, + 483, + 796 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {C D A} = F _ {G} + \\mathcal {C A} \\left(F _ {G}, F _ {S}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 803, + 482, + 820 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $F_{CDA} \\in \\mathbb{R}^{6 \\times C}$ represents the final feature while $\\mathcal{CA}$ denotes the cross attention block. By incorporating the CDA, we can effectively extract discriminative local information and adaptively aggregate complementary multimodal features, enhancing the model's discriminability.", + "bbox": [ + 89, + 824, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/fdd1d6aa4988e86a56e8c3399c79a24b790186aaf758e12c067b9a456ed165a1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsRGBNT201
mAPR-1R-5R-10
Multi-modalHAMNet [18]27.726.341.551.7
PFNet [56]38.538.952.058.4
IEEE [41]47.544.457.163.6
DENet [57]42.442.255.364.5
LRMM [43]52.353.464.673.2
UniCat* [4]57.055.7--
HTT* [42]71.173.483.187.3
TOP-ReID* [37]72.376.684.789.4
EDITOR* [54]66.568.381.188.2
RSCNet* [50]68.272.5--
WTSF-ReID* [51]67.972.283.489.7
MambaPro† [36]78.983.489.891.9
DeMo† [38]79.082.388.892.0
IDEA†80.282.190.093.3
", + "bbox": [ + 540, + 88, + 877, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Performance comparison on RGBNT201. Best results are in bold, the second bests are underlined. † denotes CLIP-based methods, * indicates ViT-based while others are CNN-based ones.", + "bbox": [ + 511, + 369, + 906, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Objective Function", + "text_level": 1, + "bbox": [ + 513, + 422, + 697, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Fig. 3, we optimize IDEA through losses applied to multiple features. To maintain consistency with prior works [37, 54], we separately extract image and text features from $F_{G}$ and $F_{CDA}$ , resulting in $F_{G}^{v}, F_{G}^{t}$ and $F_{CDA}^{v}, F_{CDA}^{t} \\in \\mathbb{R}^{3C}$ . For each feature, we apply label smoothing cross-entropy loss [31] and triplet loss [16]:", + "bbox": [ + 511, + 445, + 906, + 537 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {g} (\\mathcal {F}) = \\mathcal {L} _ {C E} (\\mathcal {F}) + \\mathcal {L} _ {T r i} (\\mathcal {F}), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 547, + 906, + 564 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{F}$ denotes the input feature. Here, $\\mathcal{L}_{CE}$ is the label smoothing cross-entropy loss while $\\mathcal{L}_{Tri}$ means the triplet loss. The overall objective function is then formulated as:", + "bbox": [ + 511, + 574, + 905, + 619 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {g} \\left(F _ {G} ^ {v}\\right) + \\mathcal {L} _ {g} \\left(F _ {C D A} ^ {v}\\right) + \\mathcal {L} _ {g} \\left(F _ {G} ^ {t}\\right) + \\mathcal {L} _ {g} \\left(F _ {C D A} ^ {t}\\right). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 628, + 903, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 513, + 657, + 645, + 674 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets and Evaluation Protocols", + "text_level": 1, + "bbox": [ + 511, + 681, + 812, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We evaluate the proposed method on three multimodal object ReID benchmarks. To efficiently annotate multi-spectral images without hardware constraints, we employ the API-based Qwen-VL [1] for automated textual description generation. To be specific, RGBNT201 [56] comprises 4,787 image triplets with 14,361 annotations, each averaging 35.47 characters and encompassing 8 distinct attributes. MSVR310 [58] contains 2,087 image triplets and 6,261 annotations, with an average length of 56.06 characters and 6 attributes. RGBNT100 [18] includes the largest number of triplets at 17,250 and 51,750 annotations with an average length of 56.25 characters and 6 attributes.", + "bbox": [ + 511, + 703, + 906, + 883 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Protocols. The performance is evaluated using", + "bbox": [ + 511, + 885, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "29705", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/c96cf1069b6916d51e45f29fd981af3a652c9f444b783a5448955cdb68b06e50.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsRGBNT100MSVR310
mAPR-1mAPR-1
Multi-modalHAMNet [18]74.593.327.142.3
PFNet [56]68.194.123.537.4
GAFNet [10]74.493.4--
GPFNet [12]75.094.5--
CCNet [58]77.296.336.455.2
LRMM [43]78.696.736.749.7
GraFT* [46]76.694.3--
UniCat* [4]79.496.2--
PHT* [28]79.992.7--
HTT* [42]75.792.6--
TOP-ReID* [37]81.296.435.944.6
EDITOR* [54]82.196.439.049.3
FACENet* [59]81.596.936.254.1
RSCNet* [50]82.396.639.549.6
WTSF-ReID* [51]82.296.539.249.1
MambaPro† [36]83.994.747.056.5
DeMo† [38]86.297.649.259.8
IDEA†87.296.547.062.4
", + "bbox": [ + 117, + 88, + 455, + 433 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "mean Average Precision (mAP) and Cumulative Matching Characteristics (CMC) at Rank-K $(K = 1,5,10)$ .", + "bbox": [ + 89, + 462, + 482, + 493 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 500, + 307, + 516 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our model is implemented using PyTorch and trained on an NVIDIA A800 GPU. The pre-trained model CLIP [29] is used for both the vision and text encoders. For the datasets, images triples are resized to $256 \\times 128$ for RGBNT201 and $128 \\times 256$ for MSVR310 and RGBNT100. Data augmentation includes random horizontal flipping, cropping and random erasing [60]. For RGBNT201 and MSVR310, the mini-batch size is set to 64, with 8 images sampled per identity. For RGBNT100, the mini-batch size is set to 128, with 16 images sampled per identity. We fine-tune the learnable modules using an initial learning rate of $3.5 \\times 10^{-6}$ , which decays to $3.5 \\times 10^{-7}$ during training. For the RGBNT201 dataset, the values of $N_{p}$ and $k$ are set to 2 and 5, respectively. Other details are provided in the supplementary material. Code and annotations are available at IDEA.", + "bbox": [ + 89, + 522, + 483, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Comparison with State-of-the-Art Methods", + "text_level": 1, + "bbox": [ + 89, + 757, + 459, + 773 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Multi-modal Person ReID. In Tab. 1, we compare our method $\\mathrm{IDEA}^{\\dagger}$ with existing multi-modal approaches on the RGBNT201 dataset. Leveraging complementary information from different modalities, multi-modal methods exhibit superior performance. Specifically, our proposed $\\mathrm{IDEA}^{\\dagger}$ achieves $80.2\\%$ mAP and $82.1\\%$ Rank-1 accuracy, surpassing TOP-ReID* by $7.9\\%$ in mAP and $5.5\\%$ in Rank-1. Compared with other leading methods like $\\mathrm{HTT}^*$ and", + "bbox": [ + 89, + 780, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/dc968a66c9b664266380eed0c575ec6559ed42a3450542dd4c54c5fce15e85f8.jpg", + "table_caption": [ + "Table 2. Performance on RGBNT100 and MSVR310." + ], + "table_footnote": [], + "table_body": "
IndexModulesMetrics
TextIMFECDAmAPRank-1
A×××70.372.1
B✓××73.475.8
C✓✓×77.281.1
D✓✓✓80.282.1
", + "bbox": [ + 568, + 88, + 848, + 176 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b59fb3e12ee78756981078e3e8ec73983fd1bdb9c7848c8a31bc2475f60a176e.jpg", + "table_caption": [ + "Table 3. Performance comparison with different modules." + ], + "table_footnote": [], + "table_body": "
IndexIMFEMetrics
InverseNetPrefixesPromptmAPRank-1
A××-72.675.1
B×✓×73.475.8
C✓×-73.777.3
D✓✓×75.478.6
E✓✓✓77.281.1
", + "bbox": [ + 539, + 205, + 877, + 304 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4. Comparison with different components in IMFE.", + "bbox": [ + 537, + 308, + 879, + 321 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "EDITOR*, IDEA† shows a superior adaptability to complex scenarios, confirming the robustness. Besides, IDEA† outperforms the CLIP-based methods MambaPro† and DeMo† by 1.3% and 1.2% in mAP, respectively. These results highlight IDEA's ability to utilize semantic information from textual guidance for improved feature discrimination.", + "bbox": [ + 511, + 335, + 903, + 426 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Multi-modal Vehicle ReID. In Tab. 2, we compare $\\mathsf{IDEA}^{\\dagger}$ with state-of-the-art methods on the RGBNT100 and MSVR310 datasets. Our $\\mathsf{IDEA}^{\\dagger}$ achieves an mAP of $87.2\\%$ , surpassing EDITOR* by $5.1\\%$ in mAP. Especially on the challenging MSVR310 dataset, $\\mathsf{IDEA}^{\\dagger}$ achieves an mAP of $47.0\\%$ and a Rank-1 accuracy of $62.4\\%$ , outperforming EDITOR* by $8.0\\%$ in mAP and $13.1\\%$ in Rank-1 accuracy. These results verify IDEA's generalization ability.", + "bbox": [ + 511, + 426, + 905, + 549 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 560, + 678, + 574 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate the effectiveness of the proposed modules on the RGBNT201 dataset. To be specific, our baseline comprises a three-branch vision encoder, which performs retrieval by concatenating the class tokens from three modalities. Upon incorporating the IMFE, we utilize $F_{G}^{t}$ for retrieval and when the CDA is included, $F_{CDA}^{t}$ is employed.", + "bbox": [ + 511, + 583, + 906, + 674 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effects of Key Modules. Tab. 3 presents the performance of various combinations of our proposed modules. Model A serves as the baseline, achieving an mAP of $70.3\\%$ and Rank-1 accuracy of $72.1\\%$ . Model B incorporates text information through parallel text encoders, improving the mAP to $73.4\\%$ and Rank-1 accuracy to $75.8\\%$ . With the inverted structure in IMFE, Model C further enhances performance, reaching an mAP of $77.2\\%$ and Rank-1 of $81.1\\%$ . Finally, Model D integrates CDA, delivering the best performance with an mAP of $80.2\\%$ . These results fully validate the effectiveness and efficiency of the proposed modules.", + "bbox": [ + 511, + 674, + 905, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effects of Key Components in IMFE. Tab. 4 shows the performance of different components in the IMFE. Model A represents the parallel text encoders without the inverted structure or prefixes, achieving an mAP of $72.6\\%$ and Rank-", + "bbox": [ + 511, + 839, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "29706", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/798abd2787957e6efe423fcf70cbc5bfb19d545292dfd153dd52d5986463281d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
IndexCDAMetrics
SampleCross AttnShared OffsetmAPRank-1
A××-76.378.7
B×✓-77.079.8
C✓××76.878.9
D✓×✓77.680.4
E✓✓×79.581.7
F✓✓✓80.282.1
", + "bbox": [ + 116, + 88, + 452, + 189 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/15e1afe55a93e5820810a97b0f5067a77ff6fbddcf1308b39a16250af60bb73f.jpg", + "table_caption": [ + "Table 5. Comparison with different components in CDA." + ], + "table_footnote": [], + "table_body": "
ModelmAPRank-1Rank-5Rank-10
IDEA w/o Text74.575.084.888.8
IDEA80.282.190.093.3
", + "bbox": [ + 117, + 215, + 452, + 263 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d22864016430797a1e246b4c41d1f1da09ee57f734354ab0f567866450c8dd90.jpg", + "table_caption": [ + "Table 6. Comparison of IDEA with and without text." + ], + "table_footnote": [], + "table_body": "
ModelmAPRank-1Rank-5Rank-10
IDEA w/o Offset78.481.389.792.2
IDEA80.282.190.093.3
", + "bbox": [ + 117, + 287, + 452, + 333 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 7. Comparison of IDEA with and without offset.", + "bbox": [ + 122, + 338, + 447, + 352 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "1 accuracy of $75.1\\%$ . With the Modal Prefixes in Model B, the mAP increases to $73.4\\%$ , demonstrating the effectiveness of our prefixes mechanism. Model C introduces the inverted structure, achieving an mAP of $73.7\\%$ . Model D combines the inverted structure and prefixes, further improving the mAP to $75.4\\%$ . Finally, Model E integrates the learnable tokens in Modal Prefixes, achieving the best performance with an mAP of $77.2\\%$ . Overall, inverted structure performs better than the parallel structure, achieving better performance. These results demonstrate the effectiveness of different components in our proposed IMFE.", + "bbox": [ + 89, + 371, + 482, + 537 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effects of Key Components in CDA. Based on the IMFE, Tab. 5 shows the performance of various components within CDA. When sampling is excluded, all local patches in $\\hat{F}_m$ are processed. Without cross attention, multi-modal interaction uses only local features, omitting global context, with pooled $\\hat{F}_m$ for retrieval. If shared offset is excluded, each modality independently generates offsets. Regarding numerical results, Model A, which does not incorporate these components, achieves an mAP of $76.3\\%$ and Rank-1 accuracy of $78.7\\%$ . Model B introduces cross attention, leading to a slight improvement in mAP to $77.0\\%$ . Model C adds sampling, resulting in a $0.5\\%$ improvement in mAP over Model A by reducing noise in the local features. Model D incorporates shared offset, yielding a $0.8\\%$ improvement in mAP and a $1.5\\%$ increase in Rank-1 accuracy compared to Model C. Model E combines cross attention and sampling, achieving an mAP of $79.5\\%$ . Finally, Model F delivers the best performance with an mAP of $80.2\\%$ . These results validate the effectiveness of each component within CDA.", + "bbox": [ + 89, + 537, + 482, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effects of Text Guidance in IDEA. Tab. 6 compares IDEA's performance with and without text input. Model A, lacking text input, achieves an mAP of $74.5\\%$ . With text guidance, Model B improves to $80.2\\%$ mAP, demonstrating the effectiveness of text in enhancing feature robustness.", + "bbox": [ + 89, + 824, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/42f1415c7c9a58be752553db781581ac718e5da4ca738c7775c2b18dc4dc0c46.jpg", + "image_caption": [ + "Figure 4. Comparison with different hyper-parameters." + ], + "image_footnote": [], + "bbox": [ + 517, + 87, + 901, + 244 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fca242f2f5a4e5487dad847ce72d4a9894f5aad349ab7f3d556eec9e62a3135b.jpg", + "image_caption": [ + "Figure 5. Visualization of the cosine similarity distribution." + ], + "image_footnote": [], + "bbox": [ + 514, + 270, + 903, + 368 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effects of Offset in IDEA. Tab. 7 compares the performance of IDEA with and without the offset mechanism. Model A uses aggregated local features with convolution to interact with global features, achieving an mAP of $78.4\\%$ and Rank-1 accuracy of $81.3\\%$ . Model B introduces the offset mechanism, improving the mAP to $80.2\\%$ and Rank-1 accuracy to $82.1\\%$ . This demonstrates that the offset mechanism helps the model focus on more important regions.", + "bbox": [ + 511, + 401, + 903, + 521 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effects of Prompt Lengths and Offset Factors. Fig. 4 illustrates the performance of IDEA with varying prompt lengths and offset factors. The results show that a prompt length of 2 and an offset factor of 5 yield the best performance. As the prompt length increases, the model's performance improves, but the gains become smaller as the prompt length grows, likely due to the text truncation. Regarding the offset factor, the model remains stable.", + "bbox": [ + 511, + 522, + 905, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Visualization Analysis", + "text_level": 1, + "bbox": [ + 511, + 652, + 718, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Cosine Similarity Distributions. Fig. 5 presents the distributions of cosine similarities among test features. The results verify that incorporating text enhances feature discrimination, while adding offset mechanism further amplifies the distinction between positive and negative samples, confirming the capture of more discriminative features.", + "bbox": [ + 511, + 674, + 903, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Multi-modal Feature Distributions. Fig. 6 visualizes the feature distributions of different modules. Comparing Fig. 6 (a) and (b), the introduction with text guidance improves feature discrimination, as instances of the same ID become more compact. In Fig. 6 (c), with IMFE, the feature distributions become more discriminative than the parallel structure in Fig. 6 (b). Finally, in Fig. 6 (d), CDA further enhances feature discrimination. These visualizations demonstrate the effectiveness of our proposed modules.", + "bbox": [ + 511, + 763, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "29707", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9f32aabe2868c593ac94fadff875e1fb100f6de5f1fc6612b6a9d47b4b1cac81.jpg", + "image_caption": [ + "(a) Baseline" + ], + "image_footnote": [], + "bbox": [ + 109, + 87, + 300, + 232 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/eef9f1ac28e78e6a38b541afc67a93819827b37f3bc122c5b7c94b7f7d18bedc.jpg", + "image_caption": [ + "(b) Baseline + Text" + ], + "image_footnote": [], + "bbox": [ + 303, + 88, + 493, + 232 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0f31d63e674c283aa75fe1c86f2e430489d4c1c2a4319543efdcdcc5d34c7edb.jpg", + "image_caption": [ + "(c) Baseline + IMFE" + ], + "image_footnote": [], + "bbox": [ + 500, + 88, + 687, + 232 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a87a0b7294c0ce3b2bcc556bdcd2c763e077467476c9e966990f3e1010401015.jpg", + "image_caption": [ + "(d) Baseline + IMFE + CDA" + ], + "image_footnote": [], + "bbox": [ + 694, + 88, + 882, + 232 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/22d5b7e53415d97b3606b49761366d21d2336aad9eb40db37adeff43076112cb.jpg", + "image_caption": [ + "Figure 6. Visualization of the feature distributions with t-SNE [33]. Different colors represent different identities." + ], + "image_footnote": [], + "bbox": [ + 106, + 270, + 465, + 449 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4924e2091fc02253bb9a74d4c8a61b1bbf946e23402da9f61b4c517d2ce0aaca.jpg", + "image_caption": [ + "Figure 7. Visualization of the generated offsets.", + "Figure 8. Rank list comparison between the baseline and IDEA." + ], + "image_footnote": [], + "bbox": [ + 99, + 470, + 468, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualization of the Generated Offsets. Fig. 7 visualizes the offsets generated by CDA. Each arrow starts at the reference point and ends at the feature sampling point. We map these offsets onto the original image to identify the regions where the model shifts its attention. As shown, the offsets highlight discriminative areas, such as the head, bag and shoes. It demonstrates how the offset mechanism guides the model to focus on the most discriminative regions, effectively reducing the influence of noisy information.", + "bbox": [ + 89, + 703, + 482, + 839 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Rank List Comparison. Fig. 8 compares the cross-camera rank lists from the baseline and IDEA. IDEA produces more accurate rankings, while the baseline yields noisier results. These visualizations validate the effectiveness of IDEA.", + "bbox": [ + 89, + 839, + 482, + 898 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5a3011b1e7c2d19c6e5e0dd3b388f366503ea0e245456c32d8628bca3bc6029e.jpg", + "image_caption": [ + "The female is wearing a vague orange blouse with vague long white skirt and white shoes. The female has short black hair and appears to be young adult. The female is carrying earrings." + ], + "image_footnote": [], + "bbox": [ + 526, + 270, + 705, + 352 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dd5afab41ddcec815592e5bdca3dd541ee3c7af658a14b9815a0cb1ff5a5ccfd.jpg", + "image_caption": [ + "The female is wearing a white black stripe shirt with blue jeans. The female has short dark hair and appears to be around 30 years old. The female is carrying beige purse." + ], + "image_footnote": [], + "bbox": [ + 710, + 270, + 890, + 352 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1fb28e64d3525493c829ebf05254dfd274b60d22d2bdd261f92435761116df85.jpg", + "image_caption": [ + "Figure 9. Visualization of the channel activation maps. Each set includes the original image, baseline and IDEA map, respectively." + ], + "image_footnote": [], + "bbox": [ + 527, + 395, + 705, + 522 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b5a87759c50629564a72d9a47b413eee466ba479fa99936f24fde31b11d5c22d.jpg", + "image_caption": [ + "The female is wearing a pink blouse with black stripes and white sneakers. The female has long dark hair and appears to be young. The female is carrying white purse." + ], + "image_footnote": [], + "bbox": [ + 710, + 395, + 890, + 477 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualization of Channel Activation Maps. Fig. 9 compares the channel activation maps of our baseline model and IDEA. The incorporation of textual guidance helps the model focus on more discriminative regions, enhancing feature robustness and improving interpretability.", + "bbox": [ + 511, + 565, + 906, + 642 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 654, + 633, + 669 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose IDEA, a novel feature learning framework for multi-modal object ReID. We first construct three text-enhanced multi-modal object ReID benchmarks using MLLMs, providing a structured caption generation pipeline. With the generated text, the Inverted Multi-modal Feature Extractor (IMFE) leverages semantic guidance from inverted texts while reducing fusion conflicts. Additionally, the Cooperative Deformable Aggregation (CDA) adaptively aggregates discriminative local features with global information. Experiments on three public ReID benchmarks demonstrate the effectiveness of our method.", + "bbox": [ + 511, + 678, + 906, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work was supported in part by the National Natural Science Foundation of China (No. 62101092, 62476044, 62388101) and Open Project of Anhui Provincial Key Laboratory of Multimodal Cognitive Computation, Anhui University (No. MMC202102, MMC202407).", + "bbox": [ + 511, + 830, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "29708", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 1(2):3, 2023.5", + "[2] Alberto Baldrati, Lorenzo Agnolucci, Marco Bertini, and Alberto Del Bimbo. Zero-shot composed image retrieval with textual inversion. In ICCV, pages 15338-15347, 2023. 4", + "[3] Cuiqun Chen, Mang Ye, and Ding Jiang. Towards modality-agnostic person re-identification with descriptive query. In CVPR, pages 15128-15137, 2023. 2", + "[4] Jennifer Crawford, Haoli Yin, Luke McDermott, and Daniel Cummings. Unicat: Crafting a stronger fusion baseline for multimodal re-identification. arXiv preprint arXiv:2310.18812, 2023. 1, 5, 6", + "[5] Haiwen Diao, Yufeng Cui, Xiaotong Li, Yueze Wang, Huchuan Lu, and Xinlong Wang. Unveiling encoder-free vision-language models. arXiv preprint arXiv:2406.11832, 2024. 2", + "[6] Zefeng Ding, Changxing Ding, Zhiyin Shao, and Dacheng Tao. Semantically self-aligned network for text-to-image part-aware person re-identification. arXiv preprint arXiv:2107.12666, 2021. 1, 2", + "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 5", + "[8] Yunpeng Gong, Liqing Huang, and Lifei Chen. Eliminate deviation with deviation for data augmentation and a general multi-modal data learning method. arXiv preprint arXiv:2101.08533, 2021. 2", + "[9] Yunpeng Gong, Zhun Zhong, Yansong Qu, Zhiming Luo, Rongrong Ji, and Min Jiang. Cross-modality perturbation synergy attack for person re-identification. In NeurIPS, 2024. 1", + "[10] Jinbo Guo, Xiaojing Zhang, Zhengyi Liu, and Yuan Wang. Generative and attentive fusion for multi-spectral vehicle re-identification. In ICSP, pages 1565–1572, 2022. 2, 6", + "[11] Qianru Han, Xinwei He, Zhi Liu, Sannyuya Liu, Ying Zhang, and Jinhai Xiang. Clip-scgi: Synthesized caption-guided inversion for person re-identification. arXiv preprint arXiv:2410.09382, 2024. 1, 2, 4", + "[12] Qiaolin He, Zefeng Lu, Zihan Wang, and Haifeng Hu. Graph-based progressive fusion network for multi-modality vehicle re-identification. TITS, pages 1-17, 2023. 2, 6", + "[13] Shuting He, Hao Luo, Pichao Wang, Fan Wang, Hao Li, and Wei Jiang. Transreid: Transformer-based object re-identification. In ICCV, pages 15013–15022, 2021. 1", + "[14] Weizhen He, Yiheng Deng, Shixiang Tang, Qihao Chen, Qingsong Xie, Yizhou Wang, Lei Bai, Feng Zhu, Rui Zhao, Wanli Ouyang, et al. Instruct-reid: A multi-purpose person re-identification task with instructions. In CVPR, pages 17521–17531, 2024. 1" + ], + "bbox": [ + 93, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 5", + "[16] Alexander Hermans, Lucas Beyer, and Bastian Leibe. In defense of the triplet loss for person re-identification. arXiv preprint arXiv:1703.07737, 2017. 5", + "[17] Ding Jiang and Mang Ye. Cross-modal implicit relation reasoning and aligning for text-to-image person retrieval. In CVPR, pages 2787-2797, 2023. 2", + "[18] Hongchao Li, Chenglong Li, Xianpeng Zhu, Aihua Zheng, and Bin Luo. Multi-spectral vehicle re-identification: A challenge. In AAAI, pages 11345–11353, 2020. 2, 5, 6", + "[19] He Li, Mang Ye, Ming Zhang, and Bo Du. All in one framework for multimodal re-identification in the wild. In CVPR, pages 17459–17469, 2024. 2", + "[20] Shuang Li, Tong Xiao, Hongsheng Li, Bolei Zhou, Dayu Yue, and Xiaogang Wang. Person search with natural language description. In CVPR, pages 1970-1979, 2017. 1", + "[21] Siyuan Li, Li Sun, and Qingli Li. Clip-reid: exploiting vision-language model for image re-identification without concrete text labels. In AAAI, pages 1405–1413, 2023. 2", + "[22] Yaowei Li, Zimo Liu, Wenming Yang, Yaowei Wang, Qingmin Liao, et al. Clip-based synergistic knowledge transfer for text-based person retrieval. arXiv preprint arXiv:2309.09496, 2023. 4", + "[23] Xuehu Liu, Pingping Zhang, Chenyang Yu, Huchuan Lu, and Xiaoyun Yang. Watching you: Global-guided reciprocal learning for video-based person re-identification. In CVPR, pages 13334–13343, 2021. 1", + "[24] Xuehu Liu, Chenyang Yu, Pingping Zhang, and Huchuan Lu. Deeply coupled convolution-transformer with spatial-temporal complementary learning for video-based person re-identification. TNNLS, 2023.", + "[25] Xuehu Liu, Pingping Zhang, Chenyang Yu, Xuesheng Qian, Xiaoyun Yang, and Huchuan Lu. A video is worth three views: Trigeminal transformers for video-based person re-identification. TITS, 2024. 1", + "[26] Hu Lu, Xuezhang Zou, and Pingping Zhang. Learning progressive modality-shared transformers for effective visible-infrared person re-identification. In AAAI, pages 1835–1843, 2023. 1", + "[27] Ke Niu, Haiyang Yu, Mengyang Zhao, Teng Fu, Siyang Yi, Wei Lu, Bin Li, Xuelin Qian, and Xiangyang Xue. Chatreid: Open-ended interactive person retrieval via hierarchical progressive tuning for vision language models. arXiv preprint arXiv:2502.19958, 2025. 1", + "[28] Wenjie Pan, Linhan Huang, Jianbao Liang, Lan Hong, and Jianqing Zhu. Progressively hybrid transformer for multimodal vehicle re-identification. Sensors, 23(9):4206, 2023. 6", + "[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 1, 2, 6", + "[30] Jiangming Shi, Xiangbo Yin, Yachao Zhang, Yuan Xie, Yanyun Qu, et al. Learning commonality, divergence and variety for unsupervised visible-infrared person re-identification. NeurIPS, 37:99715–99734, 2024. 1" + ], + "bbox": [ + 516, + 92, + 905, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "29709", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, pages 2818-2826, 2016. 5", + "[32] Wentan Tan, Changxing Ding, Jiayu Jiang, Fei Wang, Yibing Zhan, and Dapeng Tao. Harnessing the power of mllms for transferable text-to-image person reid. In CVPR, pages 17127-17137, 2024. 2, 3", + "[33] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. JMLR, 9(11), 2008. 8", + "[34] Qizao Wang, Bin Li, and Xiangyang Xue. When large vision-language models meet person re-identification. arXiv preprint arXiv:2411.18111, 2024. 1", + "[35] Yingquan Wang, Pingping Zhang, Shang Gao, Xia Geng, Hu Lu, and Dong Wang. Pyramid spatial-temporal aggregation for video-based person re-identification. In ICCV, pages 12026–12035, 2021. 1", + "[36] Yuhao Wang, Xuehu Liu, Tianyu Yan, Yang Liu, Aihua Zheng, Pingping Zhang, and Huchuan Lu. Mambapro: Multi-modal object re-identification with mamba aggregation and synergistic prompt. arXiv preprint arXiv:2412.10707, 2024.5, 6", + "[37] Yuhao Wang, Xuehu Liu, Pingping Zhang, Hu Lu, Zhengzheng Tu, and Huchuan Lu. Top-reid: Multi-spectral object re-identification with token permutation. In AAAI, pages 5758-5766, 2024. 1, 2, 3, 5, 6", + "[38] Yuhao Wang, Yang Liu, Aihua Zheng, and Pingping Zhang. Decoupled feature-based mixture of experts for multi-modal object re-identification. arXiv preprint arXiv:2412.10650, 2024. 5, 6", + "[39] Yingquan Wang, Pingping Zhang, Dong Wang, and Hutchuan Lu. Other tokens matter: Exploring global and local features of vision transformers for object re-identification. CVIU, 244:104030, 2024. 1", + "[40] Yuhao Wang, Pingping Zhang, Xuehu Liu, Zhengzheng Tu, and Huchuan Lu. Unity is strength: Unifying convolutional and transformeral features for better person re-identification. TITS, 2025. 1", + "[41] Zi Wang, Chenglong Li, Aihua Zheng, Ran He, and Jin Tang. Interact, embed, and enlarge: Boosting modality-specific representations for multi-modal person re-identification. In AAAI, pages 2633–2641, 2022. 2, 5", + "[42] Zi Wang, Huaibo Huang, Aihua Zheng, and Ran He. Heterogeneous test-time training for multi-modal person re-identification. In AAAI, pages 5850–5858, 2024. 2, 5, 6", + "[43] Di Wu, Zhihui Liu, Zihan Chen, Shenglong Gan, Kaiwen Tan, Qin Wan, and Yaonan Wang. Lrrm: Low rank multiscale multi-modal fusion for person re-identification based on rgb-ni-ti. ESWA, 263:125716, 2025. 5, 6", + "[44] Zhuofan Xia, Xuran Pan, Shiji Song, Li Erran Li, and Gao Huang. Vision transformer with deformable attention. In CVPR, pages 4794-4803, 2022. 5", + "[45] Bin Yang, Jun Chen, and Mang Ye. Shallow-deep collaborative learning for unsupervised visible-infrared person re-identification. In CVPR, pages 16870–16879, 2024. 1", + "[46] Haoli Yin, Jiayao Li, Eva Schiller, Luke McDermott, and Daniel Cummings. Graft: Gradual fusion trans" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "former for multimodal re-identification. arXiv preprint arXiv:2310.16856, 2023. 6", + "[47] Chenyang Yu, Xuehu Liu, Yingquan Wang, Pingping Zhang, and Huchuan Lu. Tf-clip: Learning text-free clip for video-based person re-identification. In AAAI, pages 6764–6772, 2024. 1", + "[48] Jiazuo Yu, Haomiao Xiong, Lu Zhang, Haiwen Diao, Yunzhi Zhuge, Lanqing Hong, Dong Wang, Huchuan Lu, You He, and Long Chen. Llms can evolve continually on modality for x-modal reasoning. arXiv preprint arXiv:2410.20178, 2024. 2", + "[49] Jiazuo Yu, Yunzhi Zhuge, Lu Zhang, Ping Hu, Dong Wang, Huchuan Lu, and You He. Boosting continual learning of vision-language models via mixture-of-experts adapters. In CVPR, pages 23219-23230, 2024. 2", + "[50] Zhi Yu, Zhiyong Huang, Mingyang Hou, Jiaming Pei, Yan Yan, Yushi Liu, and Daming Sun. Representation selective coupling via token sparsification for multi-spectral object re-identification. TCSVT, 2024. 5, 6", + "[51] Zhi Yu, Zhiyong Huang, Mingyang Hou, Yan Yan, and Yushi Liu. Wtsf-reid: Depth-driven window-oriented token selection and fusion for multi-modality vehicle re-identification with knowledge consistency constraint. ESWA, page 126921, 2025. 5, 6", + "[52] Yajing Zhai, Yawen Zeng, Da Cao, and Shaofei Lu. Trireid: Towards multi-modal person re-identification via descriptive fusion model. In ICMR, pages 63–71, 2022. 2", + "[53] Guowen Zhang, Pingping Zhang, Jinqing Qi, and Huchuan Lu. Hat: Hierarchical aggregation transformers for person re-identification. In ACM MM, pages 516-525, 2021. 1", + "[54] Pingping Zhang, Yuhao Wang, Yang Liu, Zhengzheng Tu, and Huchuan Lu. Magic tokens: Select diverse tokens for multi-modal object re-identification. In CVPR, pages 17117–17126, 2024. 1, 2, 3, 5, 6", + "[55] Qingyang Zhang, Yake Wei, Zongbo Han, Huazhu Fu, Xi Peng, Cheng Deng, Qinghua Hu, Cai Xu, Jie Wen, Di Hu, et al. Multimodal fusion on low-quality data: A comprehensive survey. arXiv preprint arXiv:2404.18947, 2024. 1", + "[56] Aihua Zheng, Zi Wang, Zihan Chen, Chenglong Li, and Jin Tang. Robust multi-modality person re-identification. In AAAI, pages 3529–3537, 2021. 2, 5, 6", + "[57] Aihua Zheng, Ziling He, Zi Wang, Chenglong Li, and Jin Tang. Dynamic enhancement network for partial multi-modality person re-identification. arXiv preprint arXiv:2305.15762, 2023. 2, 5", + "[58] Aihua Zheng, Xianpeng Zhu, Zhiqi Ma, Chenglong Li, Jin Tang, and Jixin Ma. Cross-directional consistency network with adaptive layer normalization for multi-spectral vehicle re-identification and a high-quality benchmark. Information Fusion, 100:101901, 2023. 2, 5, 6", + "[59] Aihua Zheng, Zhiqi Ma, Yongqi Sun, Zi Wang, Chenglong Li, and Jin Tang. Flare-aware cross-modal enhancement network for multi-spectral vehicle re-identification. Information Fusion, 116:102800, 2025. 6", + "[60] Zhun Zhong, Liang Zheng, Guoliang Kang, Shaozi Li, and Yi Yang. Random erasing data augmentation. In AAAI, pages 13001-13008, 2020. 6" + ], + "bbox": [ + 516, + 92, + 903, + 896 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "29710", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/IDEA_ Inverted Text with Cooperative Deformable Aggregation for Multi-modal Object Re-Identification/b7176393-d8d0-4ced-ba0c-43acd09bc650_model.json b/2025/IDEA_ Inverted Text with Cooperative Deformable Aggregation for Multi-modal Object Re-Identification/b7176393-d8d0-4ced-ba0c-43acd09bc650_model.json new file mode 100644 index 0000000000000000000000000000000000000000..71b8c5ba1ac5c1f6b0f907dbf77d64cf33278c01 --- /dev/null +++ b/2025/IDEA_ Inverted Text with Cooperative Deformable Aggregation for Multi-modal Object Re-Identification/b7176393-d8d0-4ced-ba0c-43acd09bc650_model.json @@ -0,0 +1,2420 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.095, + 0.13, + 0.905, + 0.178 + ], + "angle": 0, + "content": "IDEA: Inverted Text with Cooperative Deformable Aggregation for Multi-modal Object Re-Identification" + }, + { + "type": "text", + "bbox": [ + 0.269, + 0.204, + 0.729, + 0.222 + ], + "angle": 0, + "content": "Yuhao Wang, Yongfeng Lv, Pingping Zhang* Huchuan Lu" + }, + { + "type": "text", + "bbox": [ + 0.12, + 0.222, + 0.877, + 0.24 + ], + "angle": 0, + "content": "School of Future Technology, School of Artificial Intelligence, Dalian University of Technology" + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.242, + 0.765, + 0.257 + ], + "angle": 0, + "content": "{924973292,1vyf}@mail.dlut.edu.cn，{zhpp,lhchuan}@dlut.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.324, + 0.485, + 0.718 + ], + "angle": 0, + "content": "Multi-modal object Re-Identification (ReID) aims to retrieve specific objects by utilizing complementary information from various modalities. However, existing methods focus on fusing heterogeneous visual features, neglecting the potential benefits of text-based semantic information. To address this issue, we first construct three text-enhanced multi-modal object ReID benchmarks. To be specific, we propose a standardized multi-modal caption generation pipeline for structured and concise text annotations with Multi-modal Large Language Models (MLLMs). Besides, current methods often directly aggregate multi-modal information without selecting representative local features, leading to redundancy and high complexity. To address the above issues, we introduce IDEA, a novel feature learning framework comprising the Inverted Multi-modal Feature Extractor (IMFE) and Cooperative Deformable Aggregation (CDA). The IMFE utilizes Modal Prefixes and an InverseNet to integrate multi-modal information with semantic guidance from inverted text. The CDA adaptively generates sampling positions, enabling the model to focus on the interplay between global features and discriminative local features. With the constructed benchmarks and the proposed modules, our framework can generate more robust multi-modal features under complex scenarios. Extensive experiments on three multi-modal object ReID benchmarks demonstrate the effectiveness of our proposed method." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.746, + 0.224, + 0.761 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.772, + 0.484, + 0.879 + ], + "angle": 0, + "content": "Object Re-Identification (ReID) focuses on retrieving the same object across different camera views. While significant progress has been made with RGB images [9, 13, 23–25, 30, 35, 39, 40, 45, 47, 53], existing methods are limited under environmental challenges like dark and glare, reducing their robustness in real-world applications. Multi-modal imaging, which combines data from multiple spectra such" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.289, + 0.905, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.369, + 0.901, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.434, + 0.907, + 0.492 + ], + "angle": 0, + "content": "Figure 1. Overall illustration of our motivations and proposed framework. (a) Our multi-modal caption generation pipeline. (b) Limitations of existing MLLM-generated captions. (c) Comparison between previous methods and our proposed IDEA." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.906, + 0.569 + ], + "angle": 0, + "content": "as RGB, Near Infrared (NIR) and Thermal Infrared (TIR), offers a promising solution. By leveraging complementary information, multi-modal ReID methods [4, 26, 37, 54] improve feature robustness under challenging conditions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.57, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Beyond visual information fusion for different spectra, integrating language with vision can enhance ReID performance [11, 27, 34]. However, due to the absence of text annotations for images, most existing methods rely solely on visual features, overlooking the benefits of text-based semantic information. To address this limitation, some studies augment RGB datasets with manual text annotations [6, 20]. While this strategy boosts performance, it is both time-consuming and labor-intensive. With the advent of Multimodal Large Language Models (MLLMs), image caption has made significant progress. Recently, researchers start to use MLLMs to generate textual descriptions for RGB images [11, 14]. However, these methods face two main challenges. (1) As shown in Fig. 1 (b), the randomness in caption generation often leads to redundancy, resulting in long and complex sentences. Additionally, the text structure varies across images. In ReID tasks, such captions may exceed input length limits of a text encoder [29, 55], with redundant content diluting the essential information. (2) Current methods primarily focus on RGB images, where the generated captions lack sufficient details. However, multimodal imaging can capture crucial information in complex" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.236, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "29701" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.273 + ], + "angle": 0, + "content": "environments, providing enough information for MLLMs to generate informative captions. Despite this, no existing methods provide text annotations for multi-modal images. To bridge this gap, we construct three text-enhanced multi-modal object ReID benchmarks. Meanwhile, another challenge is the aggregation of multi-modal information. As shown in Fig. 1 (c), previous methods directly aggregate the heterogeneous information, leading to high computational complexity and noise interference [54]. To address above issues, we propose a multi-modal feature learning framework named IDEA to introduce Inverted text with cooperative DEformable Aggregation for multi-modal object ReID." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.276, + 0.485, + 0.73 + ], + "angle": 0, + "content": "Technically, we first develop a standardized pipeline for multi-modal caption generation. It extends existing datasets with text annotations across spectral modalities. Specifically, as shown in Fig. 1 (a), our caption generation pipeline consists of two steps. In the first step, we use image-template pairs in conjunction with MLLMs to generate informative captions. Then, we leverage the powerful keyword extraction capabilities of MLLMs to extract predefined attributes from the generated captions to produce structured and concise sentences. Building upon the captions, we propose the IDEA framework, which consists of two key components: the Inverted Multi-modal Feature Extractor (IMFE) and Cooperative Deformable Aggregation (CDA). First, we utilize the IMFE to extract integrated multi-modal features using Modal Prefixes and an InverseNet. Specifically, Modal Prefixes are designed to distinguish different modalities, reducing the impact of conflicting information across modalities. The InverseNet further exploits the semantic information from inverted text to enhance feature robustness. Furthermore, we propose CDA to adaptively aggregate discriminative local information. Specifically, based on the aggregated multi-modal information, we adaptively generate sampling positions to enable the model to focus on the interplay between global features and discriminative local information. Through these modules, our proposed framework effectively utilizes semantic guidance from texts while adaptively aggregating discriminative multi-modal information. Extensive experiments on three benchmark datasets demonstrate the effectiveness of our approach. In summary, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.733, + 0.483, + 0.775 + ], + "angle": 0, + "content": "- We construct three text-enhanced multi-modal object ReID benchmarks, providing a structured caption generation pipeline across multiple spectral modalities." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.778, + 0.481, + 0.866 + ], + "angle": 0, + "content": "- We introduce a novel feature learning framework named IDEA, which includes the Inverted Multi-modal Feature Extractor (IMFE) and Cooperative Deformable Aggregation (CDA). The IMFE integrates multi-modal features using Modal Prefixes and an InverseNet, while the CDA adaptively aggregates discriminative local information." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.868, + 0.481, + 0.897 + ], + "angle": 0, + "content": "- Extensive experiments on three benchmark datasets validate the effectiveness of our proposed method." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.733, + 0.483, + 0.897 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.655, + 0.105 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.116, + 0.901, + 0.132 + ], + "angle": 0, + "content": "2.1. Vision-language Learning in Re-Identification" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.139, + 0.908, + 0.609 + ], + "angle": 0, + "content": "Vision-language learning plays a pivotal role in multimodal applications. As the field advances [5, 48, 49], there is a growing focus on exploring the interaction between visual and textual information in ReID tasks. Existing methods can be broadly divided into cross-modal and multimodal ReID. In cross-modal ReID, Text-to-Image ReID [6] focuses on matching text queries with image galleries. Current methods primarily emphasize feature alignment [17] and pre-training [32]. To expand application scenarios, Zhai et al. [52] combine text and sketch modalities in the query, while Chen et al. [3] propose an uncertain Query-to-RGB retrieval model to handle missing modalities. More recently, Li et al. [19] introduce flexible modality combinations and unify various uncertainties in retrieval. However, these methods focus on cross-modal alignment. In contrast, multi-modal ReID aims to leverage complementary information from multiple modalities, with both the query and gallery containing multi-modal data. The rise of vision-language models, particularly CLIP [29], has significantly advanced this area by facilitating image-text interactions. Li et al. [21] pioneer CLIP-ReID to learn from text templates for ReID tasks. However, the templates are not real descriptions of images, potentially restricting performance. To address this, Han et al. [11] leverage MLLMs to generate descriptions for RGB images. Despite improvements, existing methods fail to tackle the challenges of generating text for multi-spectral modalities and ensuring structural consistency in the generated descriptions. To address the above issues, we propose a structured multi-modal caption generation pipeline, enhancing the consistency of texts and providing informative annotations for multi-modal object ReID." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.621, + 0.837, + 0.637 + ], + "angle": 0, + "content": "2.2. Multi-modal Object Re-Identification" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Multi-modal object ReID gains great attention due to its robustness in real-world applications. Research primarily focuses on leveraging complementary information from different modalities. Early works emphasize effective fusion strategies [8, 56]. To enhance modality-specific learning, Wang et al. [41] propose an interact-embed-enlarge framework to facilitate knowledge expansion. To address missing modal data, Zheng et al. [57] introduce pixel reconstruction to handle data inconsistency. Li et al. [18] use a coherence loss to guide feature fusion. Some methods further improve feature robustness with modality generation, graph learning and instance sampling [10, 12, 58]. Notably, Wang et al. [42] introduce a test-time training framework to mine inter-modal interactions. Wang et al. [37] develop TOP-ReID, which employs token permutation to guide modality fusion. Recently, Zhang et al. [54] propose diverse token selections to mitigate background noises. Despite promis" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29702" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.492, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.089, + 0.905, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.319, + 0.908, + 0.361 + ], + "angle": 0, + "content": "Figure 2. The upper row presents the template used to annotate our dataset. The lower row provides detailed information about the annotated dataset. (a) Category statistics for our annotated person ReID dataset. (b) Example images and captions from the RGBNT201 dataset. (c) Category statistics for our annotated vehicle ReID dataset. (d) Example images and captions from the MSVR310 dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.38, + 0.484, + 0.471 + ], + "angle": 0, + "content": "ing results, they often overlook the semantic guidance of informative text annotations. Thus, we propose the IDEA framework, which incorporates generated text annotations to leverage semantic guidance. This framework adaptively aggregates discriminative local information, enhancing feature robustness in complex scenarios." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.483, + 0.265, + 0.5 + ], + "angle": 0, + "content": "3. Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.508, + 0.484, + 0.614 + ], + "angle": 0, + "content": "In this section, we introduce the multi-modal caption generation pipeline and the proposed feature learning framework. Fig. 2 illustrates the details of caption generation, including attribute definitions and examples. Fig. 3 presents the key modules of our proposed IDEA: the Inverted Multi-modal Feature Extractor (IMFE) and the Cooperative Deformable Aggregation (CDA). Details are described as follows." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.623, + 0.383, + 0.639 + ], + "angle": 0, + "content": "3.1. Multi-modal Caption Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.483, + 0.719 + ], + "angle": 0, + "content": "To bridge the gap in multi-modal text annotation, we propose the Multi-modal Caption Generation pipeline. To be specific, our pipeline consists of two steps. The first step generates informative descriptions, while the second step reuses the MLLMs to extract the predefined key attributes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.886 + ], + "angle": 0, + "content": "Caption Generation. Taking our person annotation as an example, we define 8 attribute categories in Fig. 2 (a), including gender, belongings and so on. These attributes are mapped into the template shown in the top part of Fig. 2 (a). To tailor the description to each modality, we add specific prefixes: \"Write a comprehensive description of the person's overall appearance based on the [RGB/NIR/TIR] image, ...\" Besides, we incorporate commands to prevent the hallucination of the MLLMs [32], forming a complete prompt. This prompt, along with the corresponding image, is fed into the MLLMs to generate the textual description." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.887, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Attribute Extraction. After obtaining an initial descrip" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.38, + 0.907, + 0.593 + ], + "angle": 0, + "content": "tion, we feed it back into the MLLMs to extract predefined attributes. After that, we populate these attributes into the same template we used for caption generation, forming a structured and concise description. As shown in Fig. 2 (b), the descriptions for each modality can align well with the corresponding images. For instance, the \"white t-shirt\" and \"purse\" in the RGB image are accurately captured. For the vehicle dataset, we follow similar pipelines. As illustrated in Fig. 2 (c), the text of the RGB image successfully captures the license plate number \"KQQ819\". Meanwhile, the car logo is more prominent in the TIR image, with \"Hyundai\" being identified. This approach allows us to establish a unified pipeline for multi-modal text annotation, ensuring structural descriptions across different modalities." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.603, + 0.691, + 0.617 + ], + "angle": 0, + "content": "3.2. IDEA Framework" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.625, + 0.842, + 0.64 + ], + "angle": 0, + "content": "3.2.1. Inverted Multi-modal Feature Extractor" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.907, + 0.779 + ], + "angle": 0, + "content": "To leverage the semantic guidance from texts, we propose the Inverted Multi-modal Feature Extractor (IMFE). Unlike previous methods [37, 54] that focus on fusing multi-modal images, we incorporate multi-modal texts into the fusion. However, as indicated by the blue keywords in Fig. 2 (b) and (c), conflicts can arise due to the nature of multi-modal images. Directly aggregating contradictory information can lead to model confusion. Thus, IMFE employs Modal Prefixes and an InverseNet to address these issues." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Modal Prefixes. To enhance the model's awareness of different modal texts and mitigate the impact of conflicting information, we propose a simple method called Modal Prefixes. Taking the text input for the RGB branch as an example, as shown in the top left corner of Fig. 3, \\( T_{RGB} \\) represents the text annotation of the RGB image, while \\( P_{RGB} \\) is the modal prefix for RGB modality. \\( P_{RGB} \\) consists of two parts: a fixed text describing the characteristics of the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29703" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.087, + 0.905, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.352, + 0.908, + 0.409 + ], + "angle": 0, + "content": "Figure 3. Illustration of the proposed IDEA framework. The upper part depicts the Inverted Multi-modal Feature Extractor (IMFE). It employs modal prefixes and an InverseNet to incorporate semantic text guidance for feature discriminability. The lower part highlights the Cooperative Deformable Aggregation (CDA), which adaptively integrates discriminative local information with global features. With the integration of IMFE and CDA, IDEA effectively extracts discriminative multi-modal features for object ReID." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.428, + 0.484, + 0.534 + ], + "angle": 0, + "content": "RGB image and learnable tokens for fine-tuning. Specifically, \\( P_{RGB} \\) is structured as: \"An image of a XXXX person in the visible spectrum, capturing natural colors and fine details: \", where \"XXXX\" is replaced by an equal number of learnable tokens [22]. Here, we denote the number of learnable tokens as \\( N_P \\). Then, we concatenate \\( P_{RGB} \\) and \\( T_{RGB} \\) to form the text input \\( T_R \\) for the RGB branch:" + }, + { + "type": "equation", + "bbox": [ + 0.211, + 0.539, + 0.482, + 0.557 + ], + "angle": 0, + "content": "\\[\nT _ {R} = \\left[ P _ {R G B}, T _ {R G B} \\right]. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.562, + 0.483, + 0.636 + ], + "angle": 0, + "content": "Other modalities follow the same structure to guide the model in distinguishing between different modal texts. With this simple yet effective method, the model can better understand the characteristics of each modality and reduce conflicts during multi-modal fusion." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.637, + 0.483, + 0.832 + ], + "angle": 0, + "content": "InverseNet. To fully leverage the semantic information from texts, we propose the InverseNet. Previous methods [2, 11] often reverse global image information into pseudo-text tokens and use texts as the primary modality for downstream tasks. However, due to the limitations of MLLMs and the complexity of multi-modal image annotations, the generated texts may contain some errors and conflicts. To address this, we take a different direction, reversing global text information into pseudo-image tokens. In this reversal, we also incorporate rich information from the learnable tokens in Modal Prefixes. This approach enables the interaction between semantic information and image details, resulting in a discriminative feature representation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.833, + 0.484, + 0.88 + ], + "angle": 0, + "content": "Specifically, as shown in the upper part of Fig. 3, the text \\( T_{m} \\) is first fed into the CLIP's text encoder \\( \\mathcal{T} \\) for extracting the semantic feature \\( \\hat{f}_m^t\\in \\mathbb{R}^C \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.237, + 0.884, + 0.483, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\hat {f} _ {m} ^ {t} = \\mathcal {T} (T _ {m}). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.428, + 0.907, + 0.52 + ], + "angle": 0, + "content": "Here, \\(m \\in \\{R, N, T\\}\\) represents the RGB, NIR and TIR modalities, respectively. \\(C\\) denotes the embedding dimension. If learnable tokens exist in the modal prefixes, \\(\\hat{f}_m^t\\) will be updated by the average of learnable tokens. Then, we feed \\(\\hat{f}_m^t\\) into the InverseNet \\(\\mathcal{I}\\) to generate the pseudo token \\(f_m^t \\in \\mathbb{R}^{C'}\\) with the following equation:" + }, + { + "type": "equation", + "bbox": [ + 0.662, + 0.528, + 0.905, + 0.546 + ], + "angle": 0, + "content": "\\[\nf _ {m} ^ {t} = \\mathcal {I} \\left(\\hat {f} _ {m} ^ {t}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.554, + 0.907, + 0.66 + ], + "angle": 0, + "content": "Here, \\(C'\\) is the dimension of the pseudo token while \\(\\mathcal{I}\\) is essentially a simple layer of MLP. Meanwhile, for the image input \\(I_m\\), we first patchify it into \\(N_l\\) patches. Then, we concatenate the pseudo token \\(f_m^t\\), the patch tokens and a learnable token \\(f_m^v\\) to form the image feature \\(\\bar{F}_m \\in \\mathbb{R}^{(N_l + 2) \\times C'}\\). Finally, we feed \\(\\bar{F}_m\\) into the CLIP's vision encoder \\(\\mathcal{V}\\) to obtain the integrated feature \\(F_m \\in \\mathbb{R}^{(N_l + 2) \\times C}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.659, + 0.668, + 0.905, + 0.686 + ], + "angle": 0, + "content": "\\[\nF _ {m} = \\mathcal {V} (\\bar {F} _ {m}). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.695, + 0.907, + 0.74 + ], + "angle": 0, + "content": "By incorporating the InverseNet, we can effectively leverage the semantic information from texts and enhance the discriminability of the feature representation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.747, + 0.822, + 0.762 + ], + "angle": 0, + "content": "3.2.2. Cooperative Deformable Aggregation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.871 + ], + "angle": 0, + "content": "To effectively extract discriminative local information and adaptively aggregate complementary multi-modal features, we propose Cooperative Deformable Aggregation (CDA). CDA addresses the noise interference by adaptively generating sampling positions of key local regions. This module enables interactions between global features and discriminative local features, ensuring efficient feature aggregation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Technically, as illustrated in the lower part of Fig. 3, we first extract patch tokens from the integrated feature \\( F_{m} \\)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "29704" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.09, + 0.485, + 0.213 + ], + "angle": 0, + "content": "These patches are then reshaped to form \\(\\hat{F}_m\\in \\mathbb{R}^{H\\times W\\times C}\\) where \\(H\\) and \\(W\\) represent the height and width of the local feature map, respectively. Next, we concatenate these features along the channel dimension and pass them into the Local Mixer \\(\\mathcal{M}\\) for interaction. Specifically, \\(\\mathcal{M}\\) comprises point-wise convolution \\(\\mathcal{P}\\), the GELU activation function [15] \\(\\delta\\) and depth-wise convolution \\(\\mathcal{D}\\). This process yields the aggregated local features \\(\\hat{F}_A\\in \\mathbb{R}^{H_S\\times W_S\\times C}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.201, + 0.216, + 0.483, + 0.235 + ], + "angle": 0, + "content": "\\[\n\\hat {F} _ {A} = \\mathcal {M} ([ \\hat {F} _ {R}, \\hat {F} _ {N}, \\hat {F} _ {T} ]), \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.195, + 0.255, + 0.483, + 0.271 + ], + "angle": 0, + "content": "\\[\n\\mathcal {M} (\\mathcal {X}) = \\delta (\\mathcal {D} (\\delta (\\mathcal {P} (\\mathcal {X})))) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.274, + 0.484, + 0.425 + ], + "angle": 0, + "content": "where \\([\\cdot]\\) means concatenation. \\(H_{S}\\) and \\(W_{S}\\) represent the number of sampled regions in the height and width dimensions, respectively. Since the convolution \\(\\mathcal{D}\\) aggregates local information in the receptive field of the kernel, we define the center of each convolution operation as the reference points \\(P\\in \\mathbb{R}^{H_S\\times W_S\\times 2}\\). Based on these reference points, the aggregated local information is then used to determine the direction in which the model should shift within the local region. Thus, we send \\(\\hat{F}_A\\) into the Offset Generator \\(\\mathcal{G}\\) to predict the offsets \\(\\Delta P\\in \\mathbb{R}^{H_S\\times W_S\\times 2}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.224, + 0.43, + 0.482, + 0.448 + ], + "angle": 0, + "content": "\\[\n\\Delta P = \\mathcal {G} (\\hat {F} _ {A}) * k, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.453, + 0.484, + 0.514 + ], + "angle": 0, + "content": "where \\(k\\) scales the offset magnitude [44] and \\(\\mathcal{G}\\) denotes a linear layer. Finally, we apply the offsets \\(\\Delta P\\) to the reference points \\(P\\) to get the sampling locations \\(\\hat{P} \\in \\mathbb{R}^{H_S \\times W_S \\times 2}\\), which are shared across modalities [54]:" + }, + { + "type": "equation", + "bbox": [ + 0.237, + 0.518, + 0.482, + 0.535 + ], + "angle": 0, + "content": "\\[\n\\hat {P} = P + \\Delta P. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.54, + 0.484, + 0.692 + ], + "angle": 0, + "content": "Next, we sample the local features \\(\\hat{F}_m\\) at the positions \\(\\hat{P}\\) using the bilinear interpolation to obtain the discriminative local features. The features are concatenated to form the sampled feature \\(F_S \\in \\mathbb{R}^{3N_S \\times C}\\), where \\(N_S = H_S \\times W_S\\) represents the number of sampled regions. At this stage, we utilize the cooperative information from multiple modalities to generate deformable local discriminative features. As illustrated by the sampling points visualization in Fig. 3, the adaptive offsets effectively concentrate on critical semantic regions of the human body, enhancing the discrimination." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.692, + 0.484, + 0.797 + ], + "angle": 0, + "content": "To promote the interaction between global features and discriminative local features, we incorporate cross attention [7]. Specifically, we extract the visual class token and pseudo token from the integrated feature \\(F_{m}\\) to construct the global feature \\(F_{G} \\in \\mathbb{R}^{6 \\times C}\\). Then, the global feature \\(F_{G}\\) is used as the query, while the sampled local feature \\(F_{S}\\) serves as the key and value to facilitate interaction:" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.804, + 0.483, + 0.821 + ], + "angle": 0, + "content": "\\[\nF _ {C D A} = F _ {G} + \\mathcal {C A} \\left(F _ {G}, F _ {S}\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(F_{CDA} \\in \\mathbb{R}^{6 \\times C}\\) represents the final feature while \\(\\mathcal{CA}\\) denotes the cross attention block. By incorporating the CDA, we can effectively extract discriminative local information and adaptively aggregate complementary multimodal features, enhancing the model's discriminability." + }, + { + "type": "table", + "bbox": [ + 0.541, + 0.089, + 0.878, + 0.366 + ], + "angle": 0, + "content": "
MethodsRGBNT201
mAPR-1R-5R-10
Multi-modalHAMNet [18]27.726.341.551.7
PFNet [56]38.538.952.058.4
IEEE [41]47.544.457.163.6
DENet [57]42.442.255.364.5
LRMM [43]52.353.464.673.2
UniCat* [4]57.055.7--
HTT* [42]71.173.483.187.3
TOP-ReID* [37]72.376.684.789.4
EDITOR* [54]66.568.381.188.2
RSCNet* [50]68.272.5--
WTSF-ReID* [51]67.972.283.489.7
MambaPro† [36]78.983.489.891.9
DeMo† [38]79.082.388.892.0
IDEA†80.282.190.093.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.371, + 0.907, + 0.413 + ], + "angle": 0, + "content": "Table 1. Performance comparison on RGBNT201. Best results are in bold, the second bests are underlined. † denotes CLIP-based methods, * indicates ViT-based while others are CNN-based ones." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.424, + 0.698, + 0.44 + ], + "angle": 0, + "content": "3.3. Objective Function" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.446, + 0.907, + 0.538 + ], + "angle": 0, + "content": "As shown in Fig. 3, we optimize IDEA through losses applied to multiple features. To maintain consistency with prior works [37, 54], we separately extract image and text features from \\( F_{G} \\) and \\( F_{CDA} \\), resulting in \\( F_{G}^{v}, F_{G}^{t} \\) and \\( F_{CDA}^{v}, F_{CDA}^{t} \\in \\mathbb{R}^{3C} \\). For each feature, we apply label smoothing cross-entropy loss [31] and triplet loss [16]:" + }, + { + "type": "equation", + "bbox": [ + 0.603, + 0.548, + 0.907, + 0.565 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {g} (\\mathcal {F}) = \\mathcal {L} _ {C E} (\\mathcal {F}) + \\mathcal {L} _ {T r i} (\\mathcal {F}), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.575, + 0.906, + 0.62 + ], + "angle": 0, + "content": "where \\(\\mathcal{F}\\) denotes the input feature. Here, \\(\\mathcal{L}_{CE}\\) is the label smoothing cross-entropy loss while \\(\\mathcal{L}_{Tri}\\) means the triplet loss. The overall objective function is then formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.63, + 0.905, + 0.65 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {g} \\left(F _ {G} ^ {v}\\right) + \\mathcal {L} _ {g} \\left(F _ {C D A} ^ {v}\\right) + \\mathcal {L} _ {g} \\left(F _ {G} ^ {t}\\right) + \\mathcal {L} _ {g} \\left(F _ {C D A} ^ {t}\\right). \\tag {11}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.658, + 0.646, + 0.675 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.683, + 0.813, + 0.698 + ], + "angle": 0, + "content": "4.1. Datasets and Evaluation Protocols" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.907, + 0.885 + ], + "angle": 0, + "content": "Datasets. We evaluate the proposed method on three multimodal object ReID benchmarks. To efficiently annotate multi-spectral images without hardware constraints, we employ the API-based Qwen-VL [1] for automated textual description generation. To be specific, RGBNT201 [56] comprises 4,787 image triplets with 14,361 annotations, each averaging 35.47 characters and encompassing 8 distinct attributes. MSVR310 [58] contains 2,087 image triplets and 6,261 annotations, with an average length of 56.06 characters and 6 attributes. RGBNT100 [18] includes the largest number of triplets at 17,250 and 51,750 annotations with an average length of 56.25 characters and 6 attributes." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Evaluation Protocols. The performance is evaluated using" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29705" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.118, + 0.089, + 0.457, + 0.434 + ], + "angle": 0, + "content": "
MethodsRGBNT100MSVR310
mAPR-1mAPR-1
Multi-modalHAMNet [18]74.593.327.142.3
PFNet [56]68.194.123.537.4
GAFNet [10]74.493.4--
GPFNet [12]75.094.5--
CCNet [58]77.296.336.455.2
LRMM [43]78.696.736.749.7
GraFT* [46]76.694.3--
UniCat* [4]79.496.2--
PHT* [28]79.992.7--
HTT* [42]75.792.6--
TOP-ReID* [37]81.296.435.944.6
EDITOR* [54]82.196.439.049.3
FACENet* [59]81.596.936.254.1
RSCNet* [50]82.396.639.549.6
WTSF-ReID* [51]82.296.539.249.1
MambaPro† [36]83.994.747.056.5
DeMo† [38]86.297.649.259.8
IDEA†87.296.547.062.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.127, + 0.438, + 0.447, + 0.451 + ], + "angle": 0, + "content": "Table 2. Performance on RGBNT100 and MSVR310." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.463, + 0.483, + 0.494 + ], + "angle": 0, + "content": "mean Average Precision (mAP) and Cumulative Matching Characteristics (CMC) at Rank-K \\((K = 1,5,10)\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.501, + 0.308, + 0.517 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.523, + 0.484, + 0.75 + ], + "angle": 0, + "content": "Our model is implemented using PyTorch and trained on an NVIDIA A800 GPU. The pre-trained model CLIP [29] is used for both the vision and text encoders. For the datasets, images triples are resized to \\(256 \\times 128\\) for RGBNT201 and \\(128 \\times 256\\) for MSVR310 and RGBNT100. Data augmentation includes random horizontal flipping, cropping and random erasing [60]. For RGBNT201 and MSVR310, the mini-batch size is set to 64, with 8 images sampled per identity. For RGBNT100, the mini-batch size is set to 128, with 16 images sampled per identity. We fine-tune the learnable modules using an initial learning rate of \\(3.5 \\times 10^{-6}\\), which decays to \\(3.5 \\times 10^{-7}\\) during training. For the RGBNT201 dataset, the values of \\(N_{p}\\) and \\(k\\) are set to 2 and 5, respectively. Other details are provided in the supplementary material. Code and annotations are available at IDEA." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.758, + 0.46, + 0.774 + ], + "angle": 0, + "content": "4.3. Comparison with State-of-the-Art Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Multi-modal Person ReID. In Tab. 1, we compare our method \\(\\mathrm{IDEA}^{\\dagger}\\) with existing multi-modal approaches on the RGBNT201 dataset. Leveraging complementary information from different modalities, multi-modal methods exhibit superior performance. Specifically, our proposed \\(\\mathrm{IDEA}^{\\dagger}\\) achieves \\(80.2\\%\\) mAP and \\(82.1\\%\\) Rank-1 accuracy, surpassing TOP-ReID* by \\(7.9\\%\\) in mAP and \\(5.5\\%\\) in Rank-1. Compared with other leading methods like \\(\\mathrm{HTT}^*\\) and" + }, + { + "type": "table", + "bbox": [ + 0.569, + 0.089, + 0.849, + 0.178 + ], + "angle": 0, + "content": "
IndexModulesMetrics
TextIMFECDAmAPRank-1
A×××70.372.1
B✓××73.475.8
C✓✓×77.281.1
D✓✓✓80.282.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.538, + 0.182, + 0.88, + 0.197 + ], + "angle": 0, + "content": "Table 3. Performance comparison with different modules." + }, + { + "type": "table", + "bbox": [ + 0.54, + 0.206, + 0.878, + 0.305 + ], + "angle": 0, + "content": "
IndexIMFEMetrics
InverseNetPrefixesPromptmAPRank-1
A××-72.675.1
B×✓×73.475.8
C✓×-73.777.3
D✓✓×75.478.6
E✓✓✓77.281.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.538, + 0.309, + 0.88, + 0.323 + ], + "angle": 0, + "content": "Table 4. Comparison with different components in IMFE." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.337, + 0.905, + 0.428 + ], + "angle": 0, + "content": "EDITOR*, IDEA† shows a superior adaptability to complex scenarios, confirming the robustness. Besides, IDEA† outperforms the CLIP-based methods MambaPro† and DeMo† by 1.3% and 1.2% in mAP, respectively. These results highlight IDEA's ability to utilize semantic information from textual guidance for improved feature discrimination." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.428, + 0.906, + 0.55 + ], + "angle": 0, + "content": "Multi-modal Vehicle ReID. In Tab. 2, we compare \\(\\mathsf{IDEA}^{\\dagger}\\) with state-of-the-art methods on the RGBNT100 and MSVR310 datasets. Our \\(\\mathsf{IDEA}^{\\dagger}\\) achieves an mAP of \\(87.2\\%\\), surpassing EDITOR* by \\(5.1\\%\\) in mAP. Especially on the challenging MSVR310 dataset, \\(\\mathsf{IDEA}^{\\dagger}\\) achieves an mAP of \\(47.0\\%\\) and a Rank-1 accuracy of \\(62.4\\%\\), outperforming EDITOR* by \\(8.0\\%\\) in mAP and \\(13.1\\%\\) in Rank-1 accuracy. These results verify IDEA's generalization ability." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.561, + 0.679, + 0.575 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.907, + 0.675 + ], + "angle": 0, + "content": "We evaluate the effectiveness of the proposed modules on the RGBNT201 dataset. To be specific, our baseline comprises a three-branch vision encoder, which performs retrieval by concatenating the class tokens from three modalities. Upon incorporating the IMFE, we utilize \\( F_{G}^{t} \\) for retrieval and when the CDA is included, \\( F_{CDA}^{t} \\) is employed." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.906, + 0.84 + ], + "angle": 0, + "content": "Effects of Key Modules. Tab. 3 presents the performance of various combinations of our proposed modules. Model A serves as the baseline, achieving an mAP of \\(70.3\\%\\) and Rank-1 accuracy of \\(72.1\\%\\). Model B incorporates text information through parallel text encoders, improving the mAP to \\(73.4\\%\\) and Rank-1 accuracy to \\(75.8\\%\\). With the inverted structure in IMFE, Model C further enhances performance, reaching an mAP of \\(77.2\\%\\) and Rank-1 of \\(81.1\\%\\). Finally, Model D integrates CDA, delivering the best performance with an mAP of \\(80.2\\%\\). These results fully validate the effectiveness and efficiency of the proposed modules." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Effects of Key Components in IMFE. Tab. 4 shows the performance of different components in the IMFE. Model A represents the parallel text encoders without the inverted structure or prefixes, achieving an mAP of \\(72.6\\%\\) and Rank-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29706" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.117, + 0.089, + 0.454, + 0.19 + ], + "angle": 0, + "content": "
IndexCDAMetrics
SampleCross AttnShared OffsetmAPRank-1
A××-76.378.7
B×✓-77.079.8
C✓××76.878.9
D✓×✓77.680.4
E✓✓×79.581.7
F✓✓✓80.282.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.118, + 0.196, + 0.455, + 0.21 + ], + "angle": 0, + "content": "Table 5. Comparison with different components in CDA." + }, + { + "type": "table", + "bbox": [ + 0.118, + 0.217, + 0.454, + 0.264 + ], + "angle": 0, + "content": "
ModelmAPRank-1Rank-5Rank-10
IDEA w/o Text74.575.084.888.8
IDEA80.282.190.093.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.13, + 0.268, + 0.444, + 0.282 + ], + "angle": 0, + "content": "Table 6. Comparison of IDEA with and without text." + }, + { + "type": "table", + "bbox": [ + 0.118, + 0.289, + 0.454, + 0.334 + ], + "angle": 0, + "content": "
ModelmAPRank-1Rank-5Rank-10
IDEA w/o Offset78.481.389.792.2
IDEA80.282.190.093.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.124, + 0.339, + 0.449, + 0.353 + ], + "angle": 0, + "content": "Table 7. Comparison of IDEA with and without offset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.372, + 0.483, + 0.538 + ], + "angle": 0, + "content": "1 accuracy of \\(75.1\\%\\). With the Modal Prefixes in Model B, the mAP increases to \\(73.4\\%\\), demonstrating the effectiveness of our prefixes mechanism. Model C introduces the inverted structure, achieving an mAP of \\(73.7\\%\\). Model D combines the inverted structure and prefixes, further improving the mAP to \\(75.4\\%\\). Finally, Model E integrates the learnable tokens in Modal Prefixes, achieving the best performance with an mAP of \\(77.2\\%\\). Overall, inverted structure performs better than the parallel structure, achieving better performance. These results demonstrate the effectiveness of different components in our proposed IMFE." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.538, + 0.483, + 0.825 + ], + "angle": 0, + "content": "Effects of Key Components in CDA. Based on the IMFE, Tab. 5 shows the performance of various components within CDA. When sampling is excluded, all local patches in \\(\\hat{F}_m\\) are processed. Without cross attention, multi-modal interaction uses only local features, omitting global context, with pooled \\(\\hat{F}_m\\) for retrieval. If shared offset is excluded, each modality independently generates offsets. Regarding numerical results, Model A, which does not incorporate these components, achieves an mAP of \\(76.3\\%\\) and Rank-1 accuracy of \\(78.7\\%\\). Model B introduces cross attention, leading to a slight improvement in mAP to \\(77.0\\%\\). Model C adds sampling, resulting in a \\(0.5\\%\\) improvement in mAP over Model A by reducing noise in the local features. Model D incorporates shared offset, yielding a \\(0.8\\%\\) improvement in mAP and a \\(1.5\\%\\) increase in Rank-1 accuracy compared to Model C. Model E combines cross attention and sampling, achieving an mAP of \\(79.5\\%\\). Finally, Model F delivers the best performance with an mAP of \\(80.2\\%\\). These results validate the effectiveness of each component within CDA." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Effects of Text Guidance in IDEA. Tab. 6 compares IDEA's performance with and without text input. Model A, lacking text input, achieves an mAP of \\(74.5\\%\\). With text guidance, Model B improves to \\(80.2\\%\\) mAP, demonstrating the effectiveness of text in enhancing feature robustness." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.088, + 0.903, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.251, + 0.874, + 0.266 + ], + "angle": 0, + "content": "Figure 4. Comparison with different hyper-parameters." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.271, + 0.905, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.532, + 0.376, + 0.885, + 0.39 + ], + "angle": 0, + "content": "Figure 5. Visualization of the cosine similarity distribution." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.402, + 0.905, + 0.522 + ], + "angle": 0, + "content": "Effects of Offset in IDEA. Tab. 7 compares the performance of IDEA with and without the offset mechanism. Model A uses aggregated local features with convolution to interact with global features, achieving an mAP of \\(78.4\\%\\) and Rank-1 accuracy of \\(81.3\\%\\). Model B introduces the offset mechanism, improving the mAP to \\(80.2\\%\\) and Rank-1 accuracy to \\(82.1\\%\\). This demonstrates that the offset mechanism helps the model focus on more important regions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.523, + 0.906, + 0.644 + ], + "angle": 0, + "content": "Effects of Prompt Lengths and Offset Factors. Fig. 4 illustrates the performance of IDEA with varying prompt lengths and offset factors. The results show that a prompt length of 2 and an offset factor of 5 yield the best performance. As the prompt length increases, the model's performance improves, but the gains become smaller as the prompt length grows, likely due to the text truncation. Regarding the offset factor, the model remains stable." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.653, + 0.72, + 0.669 + ], + "angle": 0, + "content": "4.5. Visualization Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.905, + 0.764 + ], + "angle": 0, + "content": "Cosine Similarity Distributions. Fig. 5 presents the distributions of cosine similarities among test features. The results verify that incorporating text enhances feature discrimination, while adding offset mechanism further amplifies the distinction between positive and negative samples, confirming the capture of more discriminative features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Multi-modal Feature Distributions. Fig. 6 visualizes the feature distributions of different modules. Comparing Fig. 6 (a) and (b), the introduction with text guidance improves feature discrimination, as instances of the same ID become more compact. In Fig. 6 (c), with IMFE, the feature distributions become more discriminative than the parallel structure in Fig. 6 (b). Finally, in Fig. 6 (d), CDA further enhances feature discrimination. These visualizations demonstrate the effectiveness of our proposed modules." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29707" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.088, + 0.301, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.168, + 0.235, + 0.236, + 0.245 + ], + "angle": 0, + "content": "(a) Baseline" + }, + { + "type": "image", + "bbox": [ + 0.305, + 0.089, + 0.495, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.345, + 0.235, + 0.451, + 0.246 + ], + "angle": 0, + "content": "(b) Baseline + Text" + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.089, + 0.689, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.536, + 0.235, + 0.651, + 0.246 + ], + "angle": 0, + "content": "(c) Baseline + IMFE" + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.089, + 0.883, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.711, + 0.235, + 0.869, + 0.246 + ], + "angle": 0, + "content": "(d) Baseline + IMFE + CDA" + }, + { + "type": "image_caption", + "bbox": [ + 0.163, + 0.25, + 0.832, + 0.264 + ], + "angle": 0, + "content": "Figure 6. Visualization of the feature distributions with t-SNE [33]. Different colors represent different identities." + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.271, + 0.466, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.145, + 0.453, + 0.428, + 0.467 + ], + "angle": 0, + "content": "Figure 7. Visualization of the generated offsets." + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.472, + 0.469, + 0.673 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.097, + 0.677, + 0.477, + 0.69 + ], + "angle": 0, + "content": "Figure 8. Rank list comparison between the baseline and IDEA." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.483, + 0.84 + ], + "angle": 0, + "content": "Visualization of the Generated Offsets. Fig. 7 visualizes the offsets generated by CDA. Each arrow starts at the reference point and ends at the feature sampling point. We map these offsets onto the original image to identify the regions where the model shifts its attention. As shown, the offsets highlight discriminative areas, such as the head, bag and shoes. It demonstrates how the offset mechanism guides the model to focus on the most discriminative regions, effectively reducing the influence of noisy information." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.9 + ], + "angle": 0, + "content": "Rank List Comparison. Fig. 8 compares the cross-camera rank lists from the baseline and IDEA. IDEA produces more accurate rankings, while the baseline yields noisier results. These visualizations validate the effectiveness of IDEA." + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.271, + 0.707, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.528, + 0.353, + 0.704, + 0.396 + ], + "angle": 0, + "content": "The female is wearing a vague orange blouse with vague long white skirt and white shoes. The female has short black hair and appears to be young adult. The female is carrying earrings." + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.271, + 0.891, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.715, + 0.353, + 0.882, + 0.396 + ], + "angle": 0, + "content": "The female is wearing a white black stripe shirt with blue jeans. The female has short dark hair and appears to be around 30 years old. The female is carrying beige purse." + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.396, + 0.707, + 0.523 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.396, + 0.891, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.716, + 0.478, + 0.882, + 0.522 + ], + "angle": 0, + "content": "The female is wearing a pink blouse with black stripes and white sneakers. The female has long dark hair and appears to be young. The female is carrying white purse." + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.53, + 0.907, + 0.558 + ], + "angle": 0, + "content": "Figure 9. Visualization of the channel activation maps. Each set includes the original image, baseline and IDEA map, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.566, + 0.907, + 0.643 + ], + "angle": 0, + "content": "Visualization of Channel Activation Maps. Fig. 9 compares the channel activation maps of our baseline model and IDEA. The incorporation of textual guidance helps the model focus on more discriminative regions, enhancing feature robustness and improving interpretability." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.655, + 0.634, + 0.67 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.679, + 0.907, + 0.831 + ], + "angle": 0, + "content": "In this work, we propose IDEA, a novel feature learning framework for multi-modal object ReID. We first construct three text-enhanced multi-modal object ReID benchmarks using MLLMs, providing a structured caption generation pipeline. With the generated text, the Inverted Multi-modal Feature Extractor (IMFE) leverages semantic guidance from inverted texts while reducing fusion conflicts. Additionally, the Cooperative Deformable Aggregation (CDA) adaptively aggregates discriminative local features with global information. Experiments on three public ReID benchmarks demonstrate the effectiveness of our method." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.832, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Acknowledgements. This work was supported in part by the National Natural Science Foundation of China (No. 62101092, 62476044, 62388101) and Open Project of Anhui Provincial Key Laboratory of Multimodal Cognitive Computation, Anhui University (No. MMC202102, MMC202407)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29708" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 1(2):3, 2023.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.187, + 0.484, + 0.228 + ], + "angle": 0, + "content": "[2] Alberto Baldrati, Lorenzo Agnolucci, Marco Bertini, and Alberto Del Bimbo. Zero-shot composed image retrieval with textual inversion. In ICCV, pages 15338-15347, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.231, + 0.483, + 0.271 + ], + "angle": 0, + "content": "[3] Cuiqun Chen, Mang Ye, and Ding Jiang. Towards modality-agnostic person re-identification with descriptive query. In CVPR, pages 15128-15137, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.274, + 0.484, + 0.327 + ], + "angle": 0, + "content": "[4] Jennifer Crawford, Haoli Yin, Luke McDermott, and Daniel Cummings. Unicat: Crafting a stronger fusion baseline for multimodal re-identification. arXiv preprint arXiv:2310.18812, 2023. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.331, + 0.483, + 0.384 + ], + "angle": 0, + "content": "[5] Haiwen Diao, Yufeng Cui, Xiaotong Li, Yueze Wang, Huchuan Lu, and Xinlong Wang. Unveiling encoder-free vision-language models. arXiv preprint arXiv:2406.11832, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.388, + 0.483, + 0.442 + ], + "angle": 0, + "content": "[6] Zefeng Ding, Changxing Ding, Zhiyin Shao, and Dacheng Tao. Semantically self-aligned network for text-to-image part-aware person re-identification. arXiv preprint arXiv:2107.12666, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.445, + 0.483, + 0.527 + ], + "angle": 0, + "content": "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.53, + 0.483, + 0.585 + ], + "angle": 0, + "content": "[8] Yunpeng Gong, Liqing Huang, and Lifei Chen. Eliminate deviation with deviation for data augmentation and a general multi-modal data learning method. arXiv preprint arXiv:2101.08533, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.588, + 0.483, + 0.641 + ], + "angle": 0, + "content": "[9] Yunpeng Gong, Zhun Zhong, Yansong Qu, Zhiming Luo, Rongrong Ji, and Min Jiang. Cross-modality perturbation synergy attack for person re-identification. In NeurIPS, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.645, + 0.483, + 0.685 + ], + "angle": 0, + "content": "[10] Jinbo Guo, Xiaojing Zhang, Zhengyi Liu, and Yuan Wang. Generative and attentive fusion for multi-spectral vehicle re-identification. In ICSP, pages 1565–1572, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.688, + 0.483, + 0.742 + ], + "angle": 0, + "content": "[11] Qianru Han, Xinwei He, Zhi Liu, Sannyuya Liu, Ying Zhang, and Jinhai Xiang. Clip-scgi: Synthesized caption-guided inversion for person re-identification. arXiv preprint arXiv:2410.09382, 2024. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.746, + 0.483, + 0.787 + ], + "angle": 0, + "content": "[12] Qiaolin He, Zefeng Lu, Zihan Wang, and Haifeng Hu. Graph-based progressive fusion network for multi-modality vehicle re-identification. TITS, pages 1-17, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.483, + 0.83 + ], + "angle": 0, + "content": "[13] Shuting He, Hao Luo, Pichao Wang, Fan Wang, Hao Li, and Wei Jiang. Transreid: Transformer-based object re-identification. In ICCV, pages 15013–15022, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[14] Weizhen He, Yiheng Deng, Shixiang Tang, Qihao Chen, Qingsong Xie, Yizhou Wang, Lei Bai, Feng Zhu, Rui Zhao, Wanli Ouyang, et al. Instruct-reid: A multi-purpose person re-identification task with instructions. In CVPR, pages 17521–17531, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.12 + ], + "angle": 0, + "content": "[15] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.906, + 0.162 + ], + "angle": 0, + "content": "[16] Alexander Hermans, Lucas Beyer, and Bastian Leibe. In defense of the triplet loss for person re-identification. arXiv preprint arXiv:1703.07737, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.906, + 0.204 + ], + "angle": 0, + "content": "[17] Ding Jiang and Mang Ye. Cross-modal implicit relation reasoning and aligning for text-to-image person retrieval. In CVPR, pages 2787-2797, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.206, + 0.906, + 0.245 + ], + "angle": 0, + "content": "[18] Hongchao Li, Chenglong Li, Xianpeng Zhu, Aihua Zheng, and Bin Luo. Multi-spectral vehicle re-identification: A challenge. In AAAI, pages 11345–11353, 2020. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.248, + 0.906, + 0.287 + ], + "angle": 0, + "content": "[19] He Li, Mang Ye, Ming Zhang, and Bo Du. All in one framework for multimodal re-identification in the wild. In CVPR, pages 17459–17469, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.289, + 0.906, + 0.329 + ], + "angle": 0, + "content": "[20] Shuang Li, Tong Xiao, Hongsheng Li, Bolei Zhou, Dayu Yue, and Xiaogang Wang. Person search with natural language description. In CVPR, pages 1970-1979, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.331, + 0.906, + 0.371 + ], + "angle": 0, + "content": "[21] Siyuan Li, Li Sun, and Qingli Li. Clip-reid: exploiting vision-language model for image re-identification without concrete text labels. In AAAI, pages 1405–1413, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.373, + 0.906, + 0.426 + ], + "angle": 0, + "content": "[22] Yaowei Li, Zimo Liu, Wenming Yang, Yaowei Wang, Qingmin Liao, et al. Clip-based synergistic knowledge transfer for text-based person retrieval. arXiv preprint arXiv:2309.09496, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.429, + 0.906, + 0.482 + ], + "angle": 0, + "content": "[23] Xuehu Liu, Pingping Zhang, Chenyang Yu, Huchuan Lu, and Xiaoyun Yang. Watching you: Global-guided reciprocal learning for video-based person re-identification. In CVPR, pages 13334–13343, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.485, + 0.906, + 0.537 + ], + "angle": 0, + "content": "[24] Xuehu Liu, Chenyang Yu, Pingping Zhang, and Huchuan Lu. Deeply coupled convolution-transformer with spatial-temporal complementary learning for video-based person re-identification. TNNLS, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.54, + 0.906, + 0.593 + ], + "angle": 0, + "content": "[25] Xuehu Liu, Pingping Zhang, Chenyang Yu, Xuesheng Qian, Xiaoyun Yang, and Huchuan Lu. A video is worth three views: Trigeminal transformers for video-based person re-identification. TITS, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.596, + 0.906, + 0.649 + ], + "angle": 0, + "content": "[26] Hu Lu, Xuezhang Zou, and Pingping Zhang. Learning progressive modality-shared transformers for effective visible-infrared person re-identification. In AAAI, pages 1835–1843, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.652, + 0.906, + 0.718 + ], + "angle": 0, + "content": "[27] Ke Niu, Haiyang Yu, Mengyang Zhao, Teng Fu, Siyang Yi, Wei Lu, Bin Li, Xuelin Qian, and Xiangyang Xue. Chatreid: Open-ended interactive person retrieval via hierarchical progressive tuning for vision language models. arXiv preprint arXiv:2502.19958, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.721, + 0.906, + 0.773 + ], + "angle": 0, + "content": "[28] Wenjie Pan, Linhan Huang, Jianbao Liang, Lan Hong, and Jianqing Zhu. Progressively hybrid transformer for multimodal vehicle re-identification. Sensors, 23(9):4206, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.777, + 0.906, + 0.844 + ], + "angle": 0, + "content": "[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.847, + 0.906, + 0.9 + ], + "angle": 0, + "content": "[30] Jiangming Shi, Xiangbo Yin, Yachao Zhang, Yuan Xie, Yanyun Qu, et al. Learning commonality, divergence and variety for unsupervised visible-infrared person re-identification. NeurIPS, 37:99715–99734, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "29709" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.146 + ], + "angle": 0, + "content": "[31] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, pages 2818-2826, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.204 + ], + "angle": 0, + "content": "[32] Wentan Tan, Changxing Ding, Jiayu Jiang, Fei Wang, Yibing Zhan, and Dapeng Tao. Harnessing the power of mllms for transferable text-to-image person reid. In CVPR, pages 17127-17137, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.206, + 0.482, + 0.234 + ], + "angle": 0, + "content": "[33] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. JMLR, 9(11), 2008. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.482, + 0.276 + ], + "angle": 0, + "content": "[34] Qizao Wang, Bin Li, and Xiangyang Xue. When large vision-language models meet person re-identification. arXiv preprint arXiv:2411.18111, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[35] Yingquan Wang, Pingping Zhang, Shang Gao, Xia Geng, Hu Lu, and Dong Wang. Pyramid spatial-temporal aggregation for video-based person re-identification. In ICCV, pages 12026–12035, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.403 + ], + "angle": 0, + "content": "[36] Yuhao Wang, Xuehu Liu, Tianyu Yan, Yang Liu, Aihua Zheng, Pingping Zhang, and Huchuan Lu. Mambapro: Multi-modal object re-identification with mamba aggregation and synergistic prompt. arXiv preprint arXiv:2412.10707, 2024.5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.482, + 0.461 + ], + "angle": 0, + "content": "[37] Yuhao Wang, Xuehu Liu, Pingping Zhang, Hu Lu, Zhengzheng Tu, and Huchuan Lu. Top-reid: Multi-spectral object re-identification with token permutation. In AAAI, pages 5758-5766, 2024. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.482, + 0.516 + ], + "angle": 0, + "content": "[38] Yuhao Wang, Yang Liu, Aihua Zheng, and Pingping Zhang. Decoupled feature-based mixture of experts for multi-modal object re-identification. arXiv preprint arXiv:2412.10650, 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.518, + 0.482, + 0.572 + ], + "angle": 0, + "content": "[39] Yingquan Wang, Pingping Zhang, Dong Wang, and Hutchuan Lu. Other tokens matter: Exploring global and local features of vision transformers for object re-identification. CVIU, 244:104030, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.575, + 0.482, + 0.629 + ], + "angle": 0, + "content": "[40] Yuhao Wang, Pingping Zhang, Xuehu Liu, Zhengzheng Tu, and Huchuan Lu. Unity is strength: Unifying convolutional and transformeral features for better person re-identification. TITS, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.632, + 0.482, + 0.687 + ], + "angle": 0, + "content": "[41] Zi Wang, Chenglong Li, Aihua Zheng, Ran He, and Jin Tang. Interact, embed, and enlarge: Boosting modality-specific representations for multi-modal person re-identification. In AAAI, pages 2633–2641, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.688, + 0.482, + 0.73 + ], + "angle": 0, + "content": "[42] Zi Wang, Huaibo Huang, Aihua Zheng, and Ran He. Heterogeneous test-time training for multi-modal person re-identification. In AAAI, pages 5850–5858, 2024. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.482, + 0.786 + ], + "angle": 0, + "content": "[43] Di Wu, Zhihui Liu, Zihan Chen, Shenglong Gan, Kaiwen Tan, Qin Wan, and Yaonan Wang. Lrrm: Low rank multiscale multi-modal fusion for person re-identification based on rgb-ni-ti. ESWA, 263:125716, 2025. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.788, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[44] Zhuofan Xia, Xuran Pan, Shiji Song, Li Erran Li, and Gao Huang. Vision transformer with deformable attention. In CVPR, pages 4794-4803, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.831, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[45] Bin Yang, Jun Chen, and Mang Ye. Shallow-deep collaborative learning for unsupervised visible-infrared person re-identification. In CVPR, pages 16870–16879, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[46] Haoli Yin, Jiayao Li, Eva Schiller, Luke McDermott, and Daniel Cummings. Graft: Gradual fusion trans" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "former for multimodal re-identification. arXiv preprint arXiv:2310.16856, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[47] Chenyang Yu, Xuehu Liu, Yingquan Wang, Pingping Zhang, and Huchuan Lu. Tf-clip: Learning text-free clip for video-based person re-identification. In AAAI, pages 6764–6772, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[48] Jiazuo Yu, Haomiao Xiong, Lu Zhang, Haiwen Diao, Yunzhi Zhuge, Lanqing Hong, Dong Wang, Huchuan Lu, You He, and Long Chen. Llms can evolve continually on modality for x-modal reasoning. arXiv preprint arXiv:2410.20178, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[49] Jiazuo Yu, Yunzhi Zhuge, Lu Zhang, Ping Hu, Dong Wang, Huchuan Lu, and You He. Boosting continual learning of vision-language models via mixture-of-experts adapters. In CVPR, pages 23219-23230, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.359 + ], + "angle": 0, + "content": "[50] Zhi Yu, Zhiyong Huang, Mingyang Hou, Jiaming Pei, Yan Yan, Yushi Liu, and Daming Sun. Representation selective coupling via token sparsification for multi-spectral object re-identification. TCSVT, 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.43 + ], + "angle": 0, + "content": "[51] Zhi Yu, Zhiyong Huang, Mingyang Hou, Yan Yan, and Yushi Liu. Wtsf-reid: Depth-driven window-oriented token selection and fusion for multi-modality vehicle re-identification with knowledge consistency constraint. ESWA, page 126921, 2025. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.433, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[52] Yajing Zhai, Yawen Zeng, Da Cao, and Shaofei Lu. Trireid: Towards multi-modal person re-identification via descriptive fusion model. In ICMR, pages 63–71, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.905, + 0.516 + ], + "angle": 0, + "content": "[53] Guowen Zhang, Pingping Zhang, Jinqing Qi, and Huchuan Lu. Hat: Hierarchical aggregation transformers for person re-identification. In ACM MM, pages 516-525, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.518, + 0.905, + 0.572 + ], + "angle": 0, + "content": "[54] Pingping Zhang, Yuhao Wang, Yang Liu, Zhengzheng Tu, and Huchuan Lu. Magic tokens: Select diverse tokens for multi-modal object re-identification. In CVPR, pages 17117–17126, 2024. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.575, + 0.905, + 0.63 + ], + "angle": 0, + "content": "[55] Qingyang Zhang, Yake Wei, Zongbo Han, Huazhu Fu, Xi Peng, Cheng Deng, Qinghua Hu, Cai Xu, Jie Wen, Di Hu, et al. Multimodal fusion on low-quality data: A comprehensive survey. arXiv preprint arXiv:2404.18947, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.632, + 0.905, + 0.672 + ], + "angle": 0, + "content": "[56] Aihua Zheng, Zi Wang, Zihan Chen, Chenglong Li, and Jin Tang. Robust multi-modality person re-identification. In AAAI, pages 3529–3537, 2021. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.674, + 0.905, + 0.729 + ], + "angle": 0, + "content": "[57] Aihua Zheng, Ziling He, Zi Wang, Chenglong Li, and Jin Tang. Dynamic enhancement network for partial multi-modality person re-identification. arXiv preprint arXiv:2305.15762, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.731, + 0.905, + 0.799 + ], + "angle": 0, + "content": "[58] Aihua Zheng, Xianpeng Zhu, Zhiqi Ma, Chenglong Li, Jin Tang, and Jixin Ma. Cross-directional consistency network with adaptive layer normalization for multi-spectral vehicle re-identification and a high-quality benchmark. Information Fusion, 100:101901, 2023. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.855 + ], + "angle": 0, + "content": "[59] Aihua Zheng, Zhiqi Ma, Yongqi Sun, Zi Wang, Chenglong Li, and Jin Tang. Flare-aware cross-modal enhancement network for multi-spectral vehicle re-identification. Information Fusion, 116:102800, 2025. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.858, + 0.905, + 0.897 + ], + "angle": 0, + "content": "[60] Zhun Zhong, Liang Zheng, Guoliang Kang, Shaozi Li, and Yi Yang. Random erasing data augmentation. In AAAI, pages 13001-13008, 2020. 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.897 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "29710" + } + ] +] \ No newline at end of file diff --git a/2025/IDEA_ Inverted Text with Cooperative Deformable Aggregation for Multi-modal Object Re-Identification/b7176393-d8d0-4ced-ba0c-43acd09bc650_origin.pdf b/2025/IDEA_ Inverted Text with Cooperative Deformable Aggregation for Multi-modal Object Re-Identification/b7176393-d8d0-4ced-ba0c-43acd09bc650_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d55719c6cff6cd1cea08df38c0b7f56a0f05d7f5 --- /dev/null +++ b/2025/IDEA_ Inverted Text with Cooperative Deformable Aggregation for Multi-modal Object Re-Identification/b7176393-d8d0-4ced-ba0c-43acd09bc650_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:796942c489ebd1cd3185ff70fdf9c3db27f3a7e7e8646eb8a3467985f27ca77c +size 7700467 diff --git a/2025/IDEA_ Inverted Text with Cooperative Deformable Aggregation for Multi-modal Object Re-Identification/full.md b/2025/IDEA_ Inverted Text with Cooperative Deformable Aggregation for Multi-modal Object Re-Identification/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3165c5bb96eebff5553f0bbfc53dfed3f32aeb2e --- /dev/null +++ b/2025/IDEA_ Inverted Text with Cooperative Deformable Aggregation for Multi-modal Object Re-Identification/full.md @@ -0,0 +1,346 @@ +# IDEA: Inverted Text with Cooperative Deformable Aggregation for Multi-modal Object Re-Identification + +Yuhao Wang, Yongfeng Lv, Pingping Zhang* Huchuan Lu + +School of Future Technology, School of Artificial Intelligence, Dalian University of Technology + +{924973292,1vyf}@mail.dlut.edu.cn，{zhpp,lhchuan}@dlut.edu.cn + +# Abstract + +Multi-modal object Re-Identification (ReID) aims to retrieve specific objects by utilizing complementary information from various modalities. However, existing methods focus on fusing heterogeneous visual features, neglecting the potential benefits of text-based semantic information. To address this issue, we first construct three text-enhanced multi-modal object ReID benchmarks. To be specific, we propose a standardized multi-modal caption generation pipeline for structured and concise text annotations with Multi-modal Large Language Models (MLLMs). Besides, current methods often directly aggregate multi-modal information without selecting representative local features, leading to redundancy and high complexity. To address the above issues, we introduce IDEA, a novel feature learning framework comprising the Inverted Multi-modal Feature Extractor (IMFE) and Cooperative Deformable Aggregation (CDA). The IMFE utilizes Modal Prefixes and an InverseNet to integrate multi-modal information with semantic guidance from inverted text. The CDA adaptively generates sampling positions, enabling the model to focus on the interplay between global features and discriminative local features. With the constructed benchmarks and the proposed modules, our framework can generate more robust multi-modal features under complex scenarios. Extensive experiments on three multi-modal object ReID benchmarks demonstrate the effectiveness of our proposed method. + +# 1. Introduction + +Object Re-Identification (ReID) focuses on retrieving the same object across different camera views. While significant progress has been made with RGB images [9, 13, 23–25, 30, 35, 39, 40, 45, 47, 53], existing methods are limited under environmental challenges like dark and glare, reducing their robustness in real-world applications. Multi-modal imaging, which combines data from multiple spectra such + +![](images/bc9351cc02ca02e4551f39b1f49c788ff2b4e5e4e7c0b26d8a84a1c573598573.jpg) + +![](images/65fe39a059a8a2886de5639e383a213b6b0371f50791fd952292d1c02bd591eb.jpg) +Figure 1. Overall illustration of our motivations and proposed framework. (a) Our multi-modal caption generation pipeline. (b) Limitations of existing MLLM-generated captions. (c) Comparison between previous methods and our proposed IDEA. + +as RGB, Near Infrared (NIR) and Thermal Infrared (TIR), offers a promising solution. By leveraging complementary information, multi-modal ReID methods [4, 26, 37, 54] improve feature robustness under challenging conditions. + +Beyond visual information fusion for different spectra, integrating language with vision can enhance ReID performance [11, 27, 34]. However, due to the absence of text annotations for images, most existing methods rely solely on visual features, overlooking the benefits of text-based semantic information. To address this limitation, some studies augment RGB datasets with manual text annotations [6, 20]. While this strategy boosts performance, it is both time-consuming and labor-intensive. With the advent of Multimodal Large Language Models (MLLMs), image caption has made significant progress. Recently, researchers start to use MLLMs to generate textual descriptions for RGB images [11, 14]. However, these methods face two main challenges. (1) As shown in Fig. 1 (b), the randomness in caption generation often leads to redundancy, resulting in long and complex sentences. Additionally, the text structure varies across images. In ReID tasks, such captions may exceed input length limits of a text encoder [29, 55], with redundant content diluting the essential information. (2) Current methods primarily focus on RGB images, where the generated captions lack sufficient details. However, multimodal imaging can capture crucial information in complex + +environments, providing enough information for MLLMs to generate informative captions. Despite this, no existing methods provide text annotations for multi-modal images. To bridge this gap, we construct three text-enhanced multi-modal object ReID benchmarks. Meanwhile, another challenge is the aggregation of multi-modal information. As shown in Fig. 1 (c), previous methods directly aggregate the heterogeneous information, leading to high computational complexity and noise interference [54]. To address above issues, we propose a multi-modal feature learning framework named IDEA to introduce Inverted text with cooperative DEformable Aggregation for multi-modal object ReID. + +Technically, we first develop a standardized pipeline for multi-modal caption generation. It extends existing datasets with text annotations across spectral modalities. Specifically, as shown in Fig. 1 (a), our caption generation pipeline consists of two steps. In the first step, we use image-template pairs in conjunction with MLLMs to generate informative captions. Then, we leverage the powerful keyword extraction capabilities of MLLMs to extract predefined attributes from the generated captions to produce structured and concise sentences. Building upon the captions, we propose the IDEA framework, which consists of two key components: the Inverted Multi-modal Feature Extractor (IMFE) and Cooperative Deformable Aggregation (CDA). First, we utilize the IMFE to extract integrated multi-modal features using Modal Prefixes and an InverseNet. Specifically, Modal Prefixes are designed to distinguish different modalities, reducing the impact of conflicting information across modalities. The InverseNet further exploits the semantic information from inverted text to enhance feature robustness. Furthermore, we propose CDA to adaptively aggregate discriminative local information. Specifically, based on the aggregated multi-modal information, we adaptively generate sampling positions to enable the model to focus on the interplay between global features and discriminative local information. Through these modules, our proposed framework effectively utilizes semantic guidance from texts while adaptively aggregating discriminative multi-modal information. Extensive experiments on three benchmark datasets demonstrate the effectiveness of our approach. In summary, our contributions are as follows: + +- We construct three text-enhanced multi-modal object ReID benchmarks, providing a structured caption generation pipeline across multiple spectral modalities. +- We introduce a novel feature learning framework named IDEA, which includes the Inverted Multi-modal Feature Extractor (IMFE) and Cooperative Deformable Aggregation (CDA). The IMFE integrates multi-modal features using Modal Prefixes and an InverseNet, while the CDA adaptively aggregates discriminative local information. +- Extensive experiments on three benchmark datasets validate the effectiveness of our proposed method. + +# 2. Related Work + +# 2.1. Vision-language Learning in Re-Identification + +Vision-language learning plays a pivotal role in multimodal applications. As the field advances [5, 48, 49], there is a growing focus on exploring the interaction between visual and textual information in ReID tasks. Existing methods can be broadly divided into cross-modal and multimodal ReID. In cross-modal ReID, Text-to-Image ReID [6] focuses on matching text queries with image galleries. Current methods primarily emphasize feature alignment [17] and pre-training [32]. To expand application scenarios, Zhai et al. [52] combine text and sketch modalities in the query, while Chen et al. [3] propose an uncertain Query-to-RGB retrieval model to handle missing modalities. More recently, Li et al. [19] introduce flexible modality combinations and unify various uncertainties in retrieval. However, these methods focus on cross-modal alignment. In contrast, multi-modal ReID aims to leverage complementary information from multiple modalities, with both the query and gallery containing multi-modal data. The rise of vision-language models, particularly CLIP [29], has significantly advanced this area by facilitating image-text interactions. Li et al. [21] pioneer CLIP-ReID to learn from text templates for ReID tasks. However, the templates are not real descriptions of images, potentially restricting performance. To address this, Han et al. [11] leverage MLLMs to generate descriptions for RGB images. Despite improvements, existing methods fail to tackle the challenges of generating text for multi-spectral modalities and ensuring structural consistency in the generated descriptions. To address the above issues, we propose a structured multi-modal caption generation pipeline, enhancing the consistency of texts and providing informative annotations for multi-modal object ReID. + +# 2.2. Multi-modal Object Re-Identification + +Multi-modal object ReID gains great attention due to its robustness in real-world applications. Research primarily focuses on leveraging complementary information from different modalities. Early works emphasize effective fusion strategies [8, 56]. To enhance modality-specific learning, Wang et al. [41] propose an interact-embed-enlarge framework to facilitate knowledge expansion. To address missing modal data, Zheng et al. [57] introduce pixel reconstruction to handle data inconsistency. Li et al. [18] use a coherence loss to guide feature fusion. Some methods further improve feature robustness with modality generation, graph learning and instance sampling [10, 12, 58]. Notably, Wang et al. [42] introduce a test-time training framework to mine inter-modal interactions. Wang et al. [37] develop TOP-ReID, which employs token permutation to guide modality fusion. Recently, Zhang et al. [54] propose diverse token selections to mitigate background noises. Despite promis + +![](images/63e21ece53b6cec570bc2a86cf9a1f3d839d0d538c8ce2077e1f37bc0e963d71.jpg) +Figure 2. The upper row presents the template used to annotate our dataset. The lower row provides detailed information about the annotated dataset. (a) Category statistics for our annotated person ReID dataset. (b) Example images and captions from the RGBNT201 dataset. (c) Category statistics for our annotated vehicle ReID dataset. (d) Example images and captions from the MSVR310 dataset. + +![](images/a562de2e9af9bafa90e69e31f4a0edf4ed3a095704512f089f31924c5f357681.jpg) + +ing results, they often overlook the semantic guidance of informative text annotations. Thus, we propose the IDEA framework, which incorporates generated text annotations to leverage semantic guidance. This framework adaptively aggregates discriminative local information, enhancing feature robustness in complex scenarios. + +# 3. Proposed Method + +In this section, we introduce the multi-modal caption generation pipeline and the proposed feature learning framework. Fig. 2 illustrates the details of caption generation, including attribute definitions and examples. Fig. 3 presents the key modules of our proposed IDEA: the Inverted Multi-modal Feature Extractor (IMFE) and the Cooperative Deformable Aggregation (CDA). Details are described as follows. + +# 3.1. Multi-modal Caption Generation + +To bridge the gap in multi-modal text annotation, we propose the Multi-modal Caption Generation pipeline. To be specific, our pipeline consists of two steps. The first step generates informative descriptions, while the second step reuses the MLLMs to extract the predefined key attributes. + +Caption Generation. Taking our person annotation as an example, we define 8 attribute categories in Fig. 2 (a), including gender, belongings and so on. These attributes are mapped into the template shown in the top part of Fig. 2 (a). To tailor the description to each modality, we add specific prefixes: "Write a comprehensive description of the person's overall appearance based on the [RGB/NIR/TIR] image, ..." Besides, we incorporate commands to prevent the hallucination of the MLLMs [32], forming a complete prompt. This prompt, along with the corresponding image, is fed into the MLLMs to generate the textual description. + +Attribute Extraction. After obtaining an initial descrip + +tion, we feed it back into the MLLMs to extract predefined attributes. After that, we populate these attributes into the same template we used for caption generation, forming a structured and concise description. As shown in Fig. 2 (b), the descriptions for each modality can align well with the corresponding images. For instance, the "white t-shirt" and "purse" in the RGB image are accurately captured. For the vehicle dataset, we follow similar pipelines. As illustrated in Fig. 2 (c), the text of the RGB image successfully captures the license plate number "KQQ819". Meanwhile, the car logo is more prominent in the TIR image, with "Hyundai" being identified. This approach allows us to establish a unified pipeline for multi-modal text annotation, ensuring structural descriptions across different modalities. + +# 3.2. IDEA Framework + +# 3.2.1. Inverted Multi-modal Feature Extractor + +To leverage the semantic guidance from texts, we propose the Inverted Multi-modal Feature Extractor (IMFE). Unlike previous methods [37, 54] that focus on fusing multi-modal images, we incorporate multi-modal texts into the fusion. However, as indicated by the blue keywords in Fig. 2 (b) and (c), conflicts can arise due to the nature of multi-modal images. Directly aggregating contradictory information can lead to model confusion. Thus, IMFE employs Modal Prefixes and an InverseNet to address these issues. + +Modal Prefixes. To enhance the model's awareness of different modal texts and mitigate the impact of conflicting information, we propose a simple method called Modal Prefixes. Taking the text input for the RGB branch as an example, as shown in the top left corner of Fig. 3, $T_{RGB}$ represents the text annotation of the RGB image, while $P_{RGB}$ is the modal prefix for RGB modality. $P_{RGB}$ consists of two parts: a fixed text describing the characteristics of the + +![](images/8b3f4150cacc93f1549feb3273c5b7a40558ff630451663822c023988f5c7eb2.jpg) +Figure 3. Illustration of the proposed IDEA framework. The upper part depicts the Inverted Multi-modal Feature Extractor (IMFE). It employs modal prefixes and an InverseNet to incorporate semantic text guidance for feature discriminability. The lower part highlights the Cooperative Deformable Aggregation (CDA), which adaptively integrates discriminative local information with global features. With the integration of IMFE and CDA, IDEA effectively extracts discriminative multi-modal features for object ReID. + +RGB image and learnable tokens for fine-tuning. Specifically, $P_{RGB}$ is structured as: "An image of a XXXX person in the visible spectrum, capturing natural colors and fine details: ", where "XXXX" is replaced by an equal number of learnable tokens [22]. Here, we denote the number of learnable tokens as $N_P$ . Then, we concatenate $P_{RGB}$ and $T_{RGB}$ to form the text input $T_R$ for the RGB branch: + +$$ +T _ {R} = \left[ P _ {R G B}, T _ {R G B} \right]. \tag {1} +$$ + +Other modalities follow the same structure to guide the model in distinguishing between different modal texts. With this simple yet effective method, the model can better understand the characteristics of each modality and reduce conflicts during multi-modal fusion. + +InverseNet. To fully leverage the semantic information from texts, we propose the InverseNet. Previous methods [2, 11] often reverse global image information into pseudo-text tokens and use texts as the primary modality for downstream tasks. However, due to the limitations of MLLMs and the complexity of multi-modal image annotations, the generated texts may contain some errors and conflicts. To address this, we take a different direction, reversing global text information into pseudo-image tokens. In this reversal, we also incorporate rich information from the learnable tokens in Modal Prefixes. This approach enables the interaction between semantic information and image details, resulting in a discriminative feature representation. + +Specifically, as shown in the upper part of Fig. 3, the text $T_{m}$ is first fed into the CLIP's text encoder $\mathcal{T}$ for extracting the semantic feature $\hat{f}_m^t\in \mathbb{R}^C$ as follows: + +$$ +\hat {f} _ {m} ^ {t} = \mathcal {T} (T _ {m}). \tag {2} +$$ + +Here, $m \in \{R, N, T\}$ represents the RGB, NIR and TIR modalities, respectively. $C$ denotes the embedding dimension. If learnable tokens exist in the modal prefixes, $\hat{f}_m^t$ will be updated by the average of learnable tokens. Then, we feed $\hat{f}_m^t$ into the InverseNet $\mathcal{I}$ to generate the pseudo token $f_m^t \in \mathbb{R}^{C'}$ with the following equation: + +$$ +f _ {m} ^ {t} = \mathcal {I} \left(\hat {f} _ {m} ^ {t}\right). \tag {3} +$$ + +Here, $C'$ is the dimension of the pseudo token while $\mathcal{I}$ is essentially a simple layer of MLP. Meanwhile, for the image input $I_m$ , we first patchify it into $N_l$ patches. Then, we concatenate the pseudo token $f_m^t$ , the patch tokens and a learnable token $f_m^v$ to form the image feature $\bar{F}_m \in \mathbb{R}^{(N_l + 2) \times C'}$ . Finally, we feed $\bar{F}_m$ into the CLIP's vision encoder $\mathcal{V}$ to obtain the integrated feature $F_m \in \mathbb{R}^{(N_l + 2) \times C}$ as follows: + +$$ +F _ {m} = \mathcal {V} (\bar {F} _ {m}). \tag {4} +$$ + +By incorporating the InverseNet, we can effectively leverage the semantic information from texts and enhance the discriminability of the feature representation. + +# 3.2.2. Cooperative Deformable Aggregation + +To effectively extract discriminative local information and adaptively aggregate complementary multi-modal features, we propose Cooperative Deformable Aggregation (CDA). CDA addresses the noise interference by adaptively generating sampling positions of key local regions. This module enables interactions between global features and discriminative local features, ensuring efficient feature aggregation. + +Technically, as illustrated in the lower part of Fig. 3, we first extract patch tokens from the integrated feature $F_{m}$ . + +These patches are then reshaped to form $\hat{F}_m\in \mathbb{R}^{H\times W\times C}$ where $H$ and $W$ represent the height and width of the local feature map, respectively. Next, we concatenate these features along the channel dimension and pass them into the Local Mixer $\mathcal{M}$ for interaction. Specifically, $\mathcal{M}$ comprises point-wise convolution $\mathcal{P}$ , the GELU activation function [15] $\delta$ and depth-wise convolution $\mathcal{D}$ . This process yields the aggregated local features $\hat{F}_A\in \mathbb{R}^{H_S\times W_S\times C}$ : + +$$ +\hat {F} _ {A} = \mathcal {M} ([ \hat {F} _ {R}, \hat {F} _ {N}, \hat {F} _ {T} ]), \tag {5} +$$ + +$$ +\mathcal {M} (\mathcal {X}) = \delta (\mathcal {D} (\delta (\mathcal {P} (\mathcal {X})))) \tag {6} +$$ + +where $[\cdot]$ means concatenation. $H_{S}$ and $W_{S}$ represent the number of sampled regions in the height and width dimensions, respectively. Since the convolution $\mathcal{D}$ aggregates local information in the receptive field of the kernel, we define the center of each convolution operation as the reference points $P\in \mathbb{R}^{H_S\times W_S\times 2}$ . Based on these reference points, the aggregated local information is then used to determine the direction in which the model should shift within the local region. Thus, we send $\hat{F}_A$ into the Offset Generator $\mathcal{G}$ to predict the offsets $\Delta P\in \mathbb{R}^{H_S\times W_S\times 2}$ as follows: + +$$ +\Delta P = \mathcal {G} (\hat {F} _ {A}) * k, \tag {7} +$$ + +where $k$ scales the offset magnitude [44] and $\mathcal{G}$ denotes a linear layer. Finally, we apply the offsets $\Delta P$ to the reference points $P$ to get the sampling locations $\hat{P} \in \mathbb{R}^{H_S \times W_S \times 2}$ , which are shared across modalities [54]: + +$$ +\hat {P} = P + \Delta P. \tag {8} +$$ + +Next, we sample the local features $\hat{F}_m$ at the positions $\hat{P}$ using the bilinear interpolation to obtain the discriminative local features. The features are concatenated to form the sampled feature $F_S \in \mathbb{R}^{3N_S \times C}$ , where $N_S = H_S \times W_S$ represents the number of sampled regions. At this stage, we utilize the cooperative information from multiple modalities to generate deformable local discriminative features. As illustrated by the sampling points visualization in Fig. 3, the adaptive offsets effectively concentrate on critical semantic regions of the human body, enhancing the discrimination. + +To promote the interaction between global features and discriminative local features, we incorporate cross attention [7]. Specifically, we extract the visual class token and pseudo token from the integrated feature $F_{m}$ to construct the global feature $F_{G} \in \mathbb{R}^{6 \times C}$ . Then, the global feature $F_{G}$ is used as the query, while the sampled local feature $F_{S}$ serves as the key and value to facilitate interaction: + +$$ +F _ {C D A} = F _ {G} + \mathcal {C A} \left(F _ {G}, F _ {S}\right), \tag {9} +$$ + +where $F_{CDA} \in \mathbb{R}^{6 \times C}$ represents the final feature while $\mathcal{CA}$ denotes the cross attention block. By incorporating the CDA, we can effectively extract discriminative local information and adaptively aggregate complementary multimodal features, enhancing the model's discriminability. + +
MethodsRGBNT201
mAPR-1R-5R-10
Multi-modalHAMNet [18]27.726.341.551.7
PFNet [56]38.538.952.058.4
IEEE [41]47.544.457.163.6
DENet [57]42.442.255.364.5
LRMM [43]52.353.464.673.2
UniCat* [4]57.055.7--
HTT* [42]71.173.483.187.3
TOP-ReID* [37]72.376.684.789.4
EDITOR* [54]66.568.381.188.2
RSCNet* [50]68.272.5--
WTSF-ReID* [51]67.972.283.489.7
MambaPro† [36]78.983.489.891.9
DeMo† [38]79.082.388.892.0
IDEA†80.282.190.093.3
+ +Table 1. Performance comparison on RGBNT201. Best results are in bold, the second bests are underlined. † denotes CLIP-based methods, * indicates ViT-based while others are CNN-based ones. + +# 3.3. Objective Function + +As shown in Fig. 3, we optimize IDEA through losses applied to multiple features. To maintain consistency with prior works [37, 54], we separately extract image and text features from $F_{G}$ and $F_{CDA}$ , resulting in $F_{G}^{v}, F_{G}^{t}$ and $F_{CDA}^{v}, F_{CDA}^{t} \in \mathbb{R}^{3C}$ . For each feature, we apply label smoothing cross-entropy loss [31] and triplet loss [16]: + +$$ +\mathcal {L} _ {g} (\mathcal {F}) = \mathcal {L} _ {C E} (\mathcal {F}) + \mathcal {L} _ {T r i} (\mathcal {F}), \tag {10} +$$ + +where $\mathcal{F}$ denotes the input feature. Here, $\mathcal{L}_{CE}$ is the label smoothing cross-entropy loss while $\mathcal{L}_{Tri}$ means the triplet loss. The overall objective function is then formulated as: + +$$ +\mathcal {L} = \mathcal {L} _ {g} \left(F _ {G} ^ {v}\right) + \mathcal {L} _ {g} \left(F _ {C D A} ^ {v}\right) + \mathcal {L} _ {g} \left(F _ {G} ^ {t}\right) + \mathcal {L} _ {g} \left(F _ {C D A} ^ {t}\right). \tag {11} +$$ + +# 4. Experiments + +# 4.1. Datasets and Evaluation Protocols + +Datasets. We evaluate the proposed method on three multimodal object ReID benchmarks. To efficiently annotate multi-spectral images without hardware constraints, we employ the API-based Qwen-VL [1] for automated textual description generation. To be specific, RGBNT201 [56] comprises 4,787 image triplets with 14,361 annotations, each averaging 35.47 characters and encompassing 8 distinct attributes. MSVR310 [58] contains 2,087 image triplets and 6,261 annotations, with an average length of 56.06 characters and 6 attributes. RGBNT100 [18] includes the largest number of triplets at 17,250 and 51,750 annotations with an average length of 56.25 characters and 6 attributes. + +Evaluation Protocols. The performance is evaluated using + +
MethodsRGBNT100MSVR310
mAPR-1mAPR-1
Multi-modalHAMNet [18]74.593.327.142.3
PFNet [56]68.194.123.537.4
GAFNet [10]74.493.4--
GPFNet [12]75.094.5--
CCNet [58]77.296.336.455.2
LRMM [43]78.696.736.749.7
GraFT* [46]76.694.3--
UniCat* [4]79.496.2--
PHT* [28]79.992.7--
HTT* [42]75.792.6--
TOP-ReID* [37]81.296.435.944.6
EDITOR* [54]82.196.439.049.3
FACENet* [59]81.596.936.254.1
RSCNet* [50]82.396.639.549.6
WTSF-ReID* [51]82.296.539.249.1
MambaPro† [36]83.994.747.056.5
DeMo† [38]86.297.649.259.8
IDEA†87.296.547.062.4
+ +mean Average Precision (mAP) and Cumulative Matching Characteristics (CMC) at Rank-K $(K = 1,5,10)$ . + +# 4.2. Implementation Details + +Our model is implemented using PyTorch and trained on an NVIDIA A800 GPU. The pre-trained model CLIP [29] is used for both the vision and text encoders. For the datasets, images triples are resized to $256 \times 128$ for RGBNT201 and $128 \times 256$ for MSVR310 and RGBNT100. Data augmentation includes random horizontal flipping, cropping and random erasing [60]. For RGBNT201 and MSVR310, the mini-batch size is set to 64, with 8 images sampled per identity. For RGBNT100, the mini-batch size is set to 128, with 16 images sampled per identity. We fine-tune the learnable modules using an initial learning rate of $3.5 \times 10^{-6}$ , which decays to $3.5 \times 10^{-7}$ during training. For the RGBNT201 dataset, the values of $N_{p}$ and $k$ are set to 2 and 5, respectively. Other details are provided in the supplementary material. Code and annotations are available at IDEA. + +# 4.3. Comparison with State-of-the-Art Methods + +Multi-modal Person ReID. In Tab. 1, we compare our method $\mathrm{IDEA}^{\dagger}$ with existing multi-modal approaches on the RGBNT201 dataset. Leveraging complementary information from different modalities, multi-modal methods exhibit superior performance. Specifically, our proposed $\mathrm{IDEA}^{\dagger}$ achieves $80.2\%$ mAP and $82.1\%$ Rank-1 accuracy, surpassing TOP-ReID* by $7.9\%$ in mAP and $5.5\%$ in Rank-1. Compared with other leading methods like $\mathrm{HTT}^*$ and + +Table 2. Performance on RGBNT100 and MSVR310. + +
IndexModulesMetrics
TextIMFECDAmAPRank-1
A×××70.372.1
B✓××73.475.8
C✓✓×77.281.1
D✓✓✓80.282.1
+ +Table 3. Performance comparison with different modules. + +
IndexIMFEMetrics
InverseNetPrefixesPromptmAPRank-1
A××-72.675.1
B×✓×73.475.8
C✓×-73.777.3
D✓✓×75.478.6
E✓✓✓77.281.1
+ +Table 4. Comparison with different components in IMFE. + +EDITOR*, IDEA† shows a superior adaptability to complex scenarios, confirming the robustness. Besides, IDEA† outperforms the CLIP-based methods MambaPro† and DeMo† by 1.3% and 1.2% in mAP, respectively. These results highlight IDEA's ability to utilize semantic information from textual guidance for improved feature discrimination. + +Multi-modal Vehicle ReID. In Tab. 2, we compare $\mathsf{IDEA}^{\dagger}$ with state-of-the-art methods on the RGBNT100 and MSVR310 datasets. Our $\mathsf{IDEA}^{\dagger}$ achieves an mAP of $87.2\%$ , surpassing EDITOR* by $5.1\%$ in mAP. Especially on the challenging MSVR310 dataset, $\mathsf{IDEA}^{\dagger}$ achieves an mAP of $47.0\%$ and a Rank-1 accuracy of $62.4\%$ , outperforming EDITOR* by $8.0\%$ in mAP and $13.1\%$ in Rank-1 accuracy. These results verify IDEA's generalization ability. + +# 4.4. Ablation Studies + +We evaluate the effectiveness of the proposed modules on the RGBNT201 dataset. To be specific, our baseline comprises a three-branch vision encoder, which performs retrieval by concatenating the class tokens from three modalities. Upon incorporating the IMFE, we utilize $F_{G}^{t}$ for retrieval and when the CDA is included, $F_{CDA}^{t}$ is employed. + +Effects of Key Modules. Tab. 3 presents the performance of various combinations of our proposed modules. Model A serves as the baseline, achieving an mAP of $70.3\%$ and Rank-1 accuracy of $72.1\%$ . Model B incorporates text information through parallel text encoders, improving the mAP to $73.4\%$ and Rank-1 accuracy to $75.8\%$ . With the inverted structure in IMFE, Model C further enhances performance, reaching an mAP of $77.2\%$ and Rank-1 of $81.1\%$ . Finally, Model D integrates CDA, delivering the best performance with an mAP of $80.2\%$ . These results fully validate the effectiveness and efficiency of the proposed modules. + +Effects of Key Components in IMFE. Tab. 4 shows the performance of different components in the IMFE. Model A represents the parallel text encoders without the inverted structure or prefixes, achieving an mAP of $72.6\%$ and Rank- + +
IndexCDAMetrics
SampleCross AttnShared OffsetmAPRank-1
A××-76.378.7
B×✓-77.079.8
C✓××76.878.9
D✓×✓77.680.4
E✓✓×79.581.7
F✓✓✓80.282.1
+ +Table 5. Comparison with different components in CDA. + +
ModelmAPRank-1Rank-5Rank-10
IDEA w/o Text74.575.084.888.8
IDEA80.282.190.093.3
+ +Table 6. Comparison of IDEA with and without text. + +
ModelmAPRank-1Rank-5Rank-10
IDEA w/o Offset78.481.389.792.2
IDEA80.282.190.093.3
+ +Table 7. Comparison of IDEA with and without offset. + +1 accuracy of $75.1\%$ . With the Modal Prefixes in Model B, the mAP increases to $73.4\%$ , demonstrating the effectiveness of our prefixes mechanism. Model C introduces the inverted structure, achieving an mAP of $73.7\%$ . Model D combines the inverted structure and prefixes, further improving the mAP to $75.4\%$ . Finally, Model E integrates the learnable tokens in Modal Prefixes, achieving the best performance with an mAP of $77.2\%$ . Overall, inverted structure performs better than the parallel structure, achieving better performance. These results demonstrate the effectiveness of different components in our proposed IMFE. + +Effects of Key Components in CDA. Based on the IMFE, Tab. 5 shows the performance of various components within CDA. When sampling is excluded, all local patches in $\hat{F}_m$ are processed. Without cross attention, multi-modal interaction uses only local features, omitting global context, with pooled $\hat{F}_m$ for retrieval. If shared offset is excluded, each modality independently generates offsets. Regarding numerical results, Model A, which does not incorporate these components, achieves an mAP of $76.3\%$ and Rank-1 accuracy of $78.7\%$ . Model B introduces cross attention, leading to a slight improvement in mAP to $77.0\%$ . Model C adds sampling, resulting in a $0.5\%$ improvement in mAP over Model A by reducing noise in the local features. Model D incorporates shared offset, yielding a $0.8\%$ improvement in mAP and a $1.5\%$ increase in Rank-1 accuracy compared to Model C. Model E combines cross attention and sampling, achieving an mAP of $79.5\%$ . Finally, Model F delivers the best performance with an mAP of $80.2\%$ . These results validate the effectiveness of each component within CDA. + +Effects of Text Guidance in IDEA. Tab. 6 compares IDEA's performance with and without text input. Model A, lacking text input, achieves an mAP of $74.5\%$ . With text guidance, Model B improves to $80.2\%$ mAP, demonstrating the effectiveness of text in enhancing feature robustness. + +![](images/42f1415c7c9a58be752553db781581ac718e5da4ca738c7775c2b18dc4dc0c46.jpg) +Figure 4. Comparison with different hyper-parameters. + +![](images/fca242f2f5a4e5487dad847ce72d4a9894f5aad349ab7f3d556eec9e62a3135b.jpg) +Figure 5. Visualization of the cosine similarity distribution. + +Effects of Offset in IDEA. Tab. 7 compares the performance of IDEA with and without the offset mechanism. Model A uses aggregated local features with convolution to interact with global features, achieving an mAP of $78.4\%$ and Rank-1 accuracy of $81.3\%$ . Model B introduces the offset mechanism, improving the mAP to $80.2\%$ and Rank-1 accuracy to $82.1\%$ . This demonstrates that the offset mechanism helps the model focus on more important regions. + +Effects of Prompt Lengths and Offset Factors. Fig. 4 illustrates the performance of IDEA with varying prompt lengths and offset factors. The results show that a prompt length of 2 and an offset factor of 5 yield the best performance. As the prompt length increases, the model's performance improves, but the gains become smaller as the prompt length grows, likely due to the text truncation. Regarding the offset factor, the model remains stable. + +# 4.5. Visualization Analysis + +Cosine Similarity Distributions. Fig. 5 presents the distributions of cosine similarities among test features. The results verify that incorporating text enhances feature discrimination, while adding offset mechanism further amplifies the distinction between positive and negative samples, confirming the capture of more discriminative features. + +Multi-modal Feature Distributions. Fig. 6 visualizes the feature distributions of different modules. Comparing Fig. 6 (a) and (b), the introduction with text guidance improves feature discrimination, as instances of the same ID become more compact. In Fig. 6 (c), with IMFE, the feature distributions become more discriminative than the parallel structure in Fig. 6 (b). Finally, in Fig. 6 (d), CDA further enhances feature discrimination. These visualizations demonstrate the effectiveness of our proposed modules. + +![](images/9f32aabe2868c593ac94fadff875e1fb100f6de5f1fc6612b6a9d47b4b1cac81.jpg) +(a) Baseline + +![](images/eef9f1ac28e78e6a38b541afc67a93819827b37f3bc122c5b7c94b7f7d18bedc.jpg) +(b) Baseline + Text + +![](images/0f31d63e674c283aa75fe1c86f2e430489d4c1c2a4319543efdcdcc5d34c7edb.jpg) +(c) Baseline + IMFE + +![](images/a87a0b7294c0ce3b2bcc556bdcd2c763e077467476c9e966990f3e1010401015.jpg) +(d) Baseline + IMFE + CDA + +![](images/22d5b7e53415d97b3606b49761366d21d2336aad9eb40db37adeff43076112cb.jpg) +Figure 6. Visualization of the feature distributions with t-SNE [33]. Different colors represent different identities. + +![](images/4924e2091fc02253bb9a74d4c8a61b1bbf946e23402da9f61b4c517d2ce0aaca.jpg) +Figure 7. Visualization of the generated offsets. +Figure 8. Rank list comparison between the baseline and IDEA. + +Visualization of the Generated Offsets. Fig. 7 visualizes the offsets generated by CDA. Each arrow starts at the reference point and ends at the feature sampling point. We map these offsets onto the original image to identify the regions where the model shifts its attention. As shown, the offsets highlight discriminative areas, such as the head, bag and shoes. It demonstrates how the offset mechanism guides the model to focus on the most discriminative regions, effectively reducing the influence of noisy information. + +Rank List Comparison. Fig. 8 compares the cross-camera rank lists from the baseline and IDEA. IDEA produces more accurate rankings, while the baseline yields noisier results. These visualizations validate the effectiveness of IDEA. + +![](images/5a3011b1e7c2d19c6e5e0dd3b388f366503ea0e245456c32d8628bca3bc6029e.jpg) +The female is wearing a vague orange blouse with vague long white skirt and white shoes. The female has short black hair and appears to be young adult. The female is carrying earrings. + +![](images/dd5afab41ddcec815592e5bdca3dd541ee3c7af658a14b9815a0cb1ff5a5ccfd.jpg) +The female is wearing a white black stripe shirt with blue jeans. The female has short dark hair and appears to be around 30 years old. The female is carrying beige purse. + +![](images/1fb28e64d3525493c829ebf05254dfd274b60d22d2bdd261f92435761116df85.jpg) +Figure 9. Visualization of the channel activation maps. Each set includes the original image, baseline and IDEA map, respectively. + +![](images/b5a87759c50629564a72d9a47b413eee466ba479fa99936f24fde31b11d5c22d.jpg) +The female is wearing a pink blouse with black stripes and white sneakers. The female has long dark hair and appears to be young. The female is carrying white purse. + +Visualization of Channel Activation Maps. Fig. 9 compares the channel activation maps of our baseline model and IDEA. The incorporation of textual guidance helps the model focus on more discriminative regions, enhancing feature robustness and improving interpretability. + +# 5. Conclusion + +In this work, we propose IDEA, a novel feature learning framework for multi-modal object ReID. We first construct three text-enhanced multi-modal object ReID benchmarks using MLLMs, providing a structured caption generation pipeline. With the generated text, the Inverted Multi-modal Feature Extractor (IMFE) leverages semantic guidance from inverted texts while reducing fusion conflicts. Additionally, the Cooperative Deformable Aggregation (CDA) adaptively aggregates discriminative local features with global information. Experiments on three public ReID benchmarks demonstrate the effectiveness of our method. + +Acknowledgements. This work was supported in part by the National Natural Science Foundation of China (No. 62101092, 62476044, 62388101) and Open Project of Anhui Provincial Key Laboratory of Multimodal Cognitive Computation, Anhui University (No. MMC202102, MMC202407). + +# References + +[1] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 1(2):3, 2023.5 +[2] Alberto Baldrati, Lorenzo Agnolucci, Marco Bertini, and Alberto Del Bimbo. Zero-shot composed image retrieval with textual inversion. In ICCV, pages 15338-15347, 2023. 4 +[3] Cuiqun Chen, Mang Ye, and Ding Jiang. Towards modality-agnostic person re-identification with descriptive query. In CVPR, pages 15128-15137, 2023. 2 +[4] Jennifer Crawford, Haoli Yin, Luke McDermott, and Daniel Cummings. Unicat: Crafting a stronger fusion baseline for multimodal re-identification. arXiv preprint arXiv:2310.18812, 2023. 1, 5, 6 +[5] Haiwen Diao, Yufeng Cui, Xiaotong Li, Yueze Wang, Huchuan Lu, and Xinlong Wang. Unveiling encoder-free vision-language models. arXiv preprint arXiv:2406.11832, 2024. 2 +[6] Zefeng Ding, Changxing Ding, Zhiyin Shao, and Dacheng Tao. Semantically self-aligned network for text-to-image part-aware person re-identification. arXiv preprint arXiv:2107.12666, 2021. 1, 2 +[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 5 +[8] Yunpeng Gong, Liqing Huang, and Lifei Chen. Eliminate deviation with deviation for data augmentation and a general multi-modal data learning method. arXiv preprint arXiv:2101.08533, 2021. 2 +[9] Yunpeng Gong, Zhun Zhong, Yansong Qu, Zhiming Luo, Rongrong Ji, and Min Jiang. Cross-modality perturbation synergy attack for person re-identification. In NeurIPS, 2024. 1 +[10] Jinbo Guo, Xiaojing Zhang, Zhengyi Liu, and Yuan Wang. Generative and attentive fusion for multi-spectral vehicle re-identification. In ICSP, pages 1565–1572, 2022. 2, 6 +[11] Qianru Han, Xinwei He, Zhi Liu, Sannyuya Liu, Ying Zhang, and Jinhai Xiang. Clip-scgi: Synthesized caption-guided inversion for person re-identification. arXiv preprint arXiv:2410.09382, 2024. 1, 2, 4 +[12] Qiaolin He, Zefeng Lu, Zihan Wang, and Haifeng Hu. Graph-based progressive fusion network for multi-modality vehicle re-identification. TITS, pages 1-17, 2023. 2, 6 +[13] Shuting He, Hao Luo, Pichao Wang, Fan Wang, Hao Li, and Wei Jiang. Transreid: Transformer-based object re-identification. In ICCV, pages 15013–15022, 2021. 1 +[14] Weizhen He, Yiheng Deng, Shixiang Tang, Qihao Chen, Qingsong Xie, Yizhou Wang, Lei Bai, Feng Zhu, Rui Zhao, Wanli Ouyang, et al. Instruct-reid: A multi-purpose person re-identification task with instructions. In CVPR, pages 17521–17531, 2024. 1 + +[15] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 5 +[16] Alexander Hermans, Lucas Beyer, and Bastian Leibe. In defense of the triplet loss for person re-identification. arXiv preprint arXiv:1703.07737, 2017. 5 +[17] Ding Jiang and Mang Ye. Cross-modal implicit relation reasoning and aligning for text-to-image person retrieval. In CVPR, pages 2787-2797, 2023. 2 +[18] Hongchao Li, Chenglong Li, Xianpeng Zhu, Aihua Zheng, and Bin Luo. Multi-spectral vehicle re-identification: A challenge. In AAAI, pages 11345–11353, 2020. 2, 5, 6 +[19] He Li, Mang Ye, Ming Zhang, and Bo Du. All in one framework for multimodal re-identification in the wild. In CVPR, pages 17459–17469, 2024. 2 +[20] Shuang Li, Tong Xiao, Hongsheng Li, Bolei Zhou, Dayu Yue, and Xiaogang Wang. Person search with natural language description. In CVPR, pages 1970-1979, 2017. 1 +[21] Siyuan Li, Li Sun, and Qingli Li. Clip-reid: exploiting vision-language model for image re-identification without concrete text labels. In AAAI, pages 1405–1413, 2023. 2 +[22] Yaowei Li, Zimo Liu, Wenming Yang, Yaowei Wang, Qingmin Liao, et al. Clip-based synergistic knowledge transfer for text-based person retrieval. arXiv preprint arXiv:2309.09496, 2023. 4 +[23] Xuehu Liu, Pingping Zhang, Chenyang Yu, Huchuan Lu, and Xiaoyun Yang. Watching you: Global-guided reciprocal learning for video-based person re-identification. In CVPR, pages 13334–13343, 2021. 1 +[24] Xuehu Liu, Chenyang Yu, Pingping Zhang, and Huchuan Lu. Deeply coupled convolution-transformer with spatial-temporal complementary learning for video-based person re-identification. TNNLS, 2023. +[25] Xuehu Liu, Pingping Zhang, Chenyang Yu, Xuesheng Qian, Xiaoyun Yang, and Huchuan Lu. A video is worth three views: Trigeminal transformers for video-based person re-identification. TITS, 2024. 1 +[26] Hu Lu, Xuezhang Zou, and Pingping Zhang. Learning progressive modality-shared transformers for effective visible-infrared person re-identification. In AAAI, pages 1835–1843, 2023. 1 +[27] Ke Niu, Haiyang Yu, Mengyang Zhao, Teng Fu, Siyang Yi, Wei Lu, Bin Li, Xuelin Qian, and Xiangyang Xue. Chatreid: Open-ended interactive person retrieval via hierarchical progressive tuning for vision language models. arXiv preprint arXiv:2502.19958, 2025. 1 +[28] Wenjie Pan, Linhan Huang, Jianbao Liang, Lan Hong, and Jianqing Zhu. Progressively hybrid transformer for multimodal vehicle re-identification. Sensors, 23(9):4206, 2023. 6 +[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 1, 2, 6 +[30] Jiangming Shi, Xiangbo Yin, Yachao Zhang, Yuan Xie, Yanyun Qu, et al. Learning commonality, divergence and variety for unsupervised visible-infrared person re-identification. NeurIPS, 37:99715–99734, 2024. 1 + +[31] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, pages 2818-2826, 2016. 5 +[32] Wentan Tan, Changxing Ding, Jiayu Jiang, Fei Wang, Yibing Zhan, and Dapeng Tao. Harnessing the power of mllms for transferable text-to-image person reid. In CVPR, pages 17127-17137, 2024. 2, 3 +[33] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. JMLR, 9(11), 2008. 8 +[34] Qizao Wang, Bin Li, and Xiangyang Xue. When large vision-language models meet person re-identification. arXiv preprint arXiv:2411.18111, 2024. 1 +[35] Yingquan Wang, Pingping Zhang, Shang Gao, Xia Geng, Hu Lu, and Dong Wang. Pyramid spatial-temporal aggregation for video-based person re-identification. In ICCV, pages 12026–12035, 2021. 1 +[36] Yuhao Wang, Xuehu Liu, Tianyu Yan, Yang Liu, Aihua Zheng, Pingping Zhang, and Huchuan Lu. Mambapro: Multi-modal object re-identification with mamba aggregation and synergistic prompt. arXiv preprint arXiv:2412.10707, 2024.5, 6 +[37] Yuhao Wang, Xuehu Liu, Pingping Zhang, Hu Lu, Zhengzheng Tu, and Huchuan Lu. Top-reid: Multi-spectral object re-identification with token permutation. In AAAI, pages 5758-5766, 2024. 1, 2, 3, 5, 6 +[38] Yuhao Wang, Yang Liu, Aihua Zheng, and Pingping Zhang. Decoupled feature-based mixture of experts for multi-modal object re-identification. arXiv preprint arXiv:2412.10650, 2024. 5, 6 +[39] Yingquan Wang, Pingping Zhang, Dong Wang, and Hutchuan Lu. Other tokens matter: Exploring global and local features of vision transformers for object re-identification. CVIU, 244:104030, 2024. 1 +[40] Yuhao Wang, Pingping Zhang, Xuehu Liu, Zhengzheng Tu, and Huchuan Lu. Unity is strength: Unifying convolutional and transformeral features for better person re-identification. TITS, 2025. 1 +[41] Zi Wang, Chenglong Li, Aihua Zheng, Ran He, and Jin Tang. Interact, embed, and enlarge: Boosting modality-specific representations for multi-modal person re-identification. In AAAI, pages 2633–2641, 2022. 2, 5 +[42] Zi Wang, Huaibo Huang, Aihua Zheng, and Ran He. Heterogeneous test-time training for multi-modal person re-identification. In AAAI, pages 5850–5858, 2024. 2, 5, 6 +[43] Di Wu, Zhihui Liu, Zihan Chen, Shenglong Gan, Kaiwen Tan, Qin Wan, and Yaonan Wang. Lrrm: Low rank multiscale multi-modal fusion for person re-identification based on rgb-ni-ti. ESWA, 263:125716, 2025. 5, 6 +[44] Zhuofan Xia, Xuran Pan, Shiji Song, Li Erran Li, and Gao Huang. Vision transformer with deformable attention. In CVPR, pages 4794-4803, 2022. 5 +[45] Bin Yang, Jun Chen, and Mang Ye. Shallow-deep collaborative learning for unsupervised visible-infrared person re-identification. In CVPR, pages 16870–16879, 2024. 1 +[46] Haoli Yin, Jiayao Li, Eva Schiller, Luke McDermott, and Daniel Cummings. Graft: Gradual fusion trans + +former for multimodal re-identification. arXiv preprint arXiv:2310.16856, 2023. 6 +[47] Chenyang Yu, Xuehu Liu, Yingquan Wang, Pingping Zhang, and Huchuan Lu. Tf-clip: Learning text-free clip for video-based person re-identification. In AAAI, pages 6764–6772, 2024. 1 +[48] Jiazuo Yu, Haomiao Xiong, Lu Zhang, Haiwen Diao, Yunzhi Zhuge, Lanqing Hong, Dong Wang, Huchuan Lu, You He, and Long Chen. Llms can evolve continually on modality for x-modal reasoning. arXiv preprint arXiv:2410.20178, 2024. 2 +[49] Jiazuo Yu, Yunzhi Zhuge, Lu Zhang, Ping Hu, Dong Wang, Huchuan Lu, and You He. Boosting continual learning of vision-language models via mixture-of-experts adapters. In CVPR, pages 23219-23230, 2024. 2 +[50] Zhi Yu, Zhiyong Huang, Mingyang Hou, Jiaming Pei, Yan Yan, Yushi Liu, and Daming Sun. Representation selective coupling via token sparsification for multi-spectral object re-identification. TCSVT, 2024. 5, 6 +[51] Zhi Yu, Zhiyong Huang, Mingyang Hou, Yan Yan, and Yushi Liu. Wtsf-reid: Depth-driven window-oriented token selection and fusion for multi-modality vehicle re-identification with knowledge consistency constraint. ESWA, page 126921, 2025. 5, 6 +[52] Yajing Zhai, Yawen Zeng, Da Cao, and Shaofei Lu. Trireid: Towards multi-modal person re-identification via descriptive fusion model. In ICMR, pages 63–71, 2022. 2 +[53] Guowen Zhang, Pingping Zhang, Jinqing Qi, and Huchuan Lu. Hat: Hierarchical aggregation transformers for person re-identification. In ACM MM, pages 516-525, 2021. 1 +[54] Pingping Zhang, Yuhao Wang, Yang Liu, Zhengzheng Tu, and Huchuan Lu. Magic tokens: Select diverse tokens for multi-modal object re-identification. In CVPR, pages 17117–17126, 2024. 1, 2, 3, 5, 6 +[55] Qingyang Zhang, Yake Wei, Zongbo Han, Huazhu Fu, Xi Peng, Cheng Deng, Qinghua Hu, Cai Xu, Jie Wen, Di Hu, et al. Multimodal fusion on low-quality data: A comprehensive survey. arXiv preprint arXiv:2404.18947, 2024. 1 +[56] Aihua Zheng, Zi Wang, Zihan Chen, Chenglong Li, and Jin Tang. Robust multi-modality person re-identification. In AAAI, pages 3529–3537, 2021. 2, 5, 6 +[57] Aihua Zheng, Ziling He, Zi Wang, Chenglong Li, and Jin Tang. Dynamic enhancement network for partial multi-modality person re-identification. arXiv preprint arXiv:2305.15762, 2023. 2, 5 +[58] Aihua Zheng, Xianpeng Zhu, Zhiqi Ma, Chenglong Li, Jin Tang, and Jixin Ma. Cross-directional consistency network with adaptive layer normalization for multi-spectral vehicle re-identification and a high-quality benchmark. Information Fusion, 100:101901, 2023. 2, 5, 6 +[59] Aihua Zheng, Zhiqi Ma, Yongqi Sun, Zi Wang, Chenglong Li, and Jin Tang. Flare-aware cross-modal enhancement network for multi-spectral vehicle re-identification. Information Fusion, 116:102800, 2025. 6 +[60] Zhun Zhong, Liang Zheng, Guoliang Kang, Shaozi Li, and Yi Yang. Random erasing data augmentation. In AAAI, pages 13001-13008, 2020. 6 \ No newline at end of file diff --git a/2025/IDEA_ Inverted Text with Cooperative Deformable Aggregation for Multi-modal Object Re-Identification/images.zip b/2025/IDEA_ Inverted Text with Cooperative Deformable Aggregation for Multi-modal Object Re-Identification/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2140712f6212f77a97dfbfafd952531732b59ba4 --- /dev/null +++ b/2025/IDEA_ Inverted Text with Cooperative Deformable Aggregation for Multi-modal Object Re-Identification/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:34c18ded51efa959b56f21a0522b7bf1f0afc7eedff413226e960ab2d1e4d762 +size 890173 diff --git a/2025/IDEA_ Inverted Text with Cooperative Deformable Aggregation for Multi-modal Object Re-Identification/layout.json b/2025/IDEA_ Inverted Text with Cooperative Deformable Aggregation for Multi-modal Object Re-Identification/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0b044a3f1da8383fed1a46d3a0f61f7ffebf9c18 --- /dev/null +++ b/2025/IDEA_ Inverted Text with Cooperative Deformable Aggregation for Multi-modal Object Re-Identification/layout.json @@ -0,0 +1,9941 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 58, + 102, + 553, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 102, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 58, + 102, + 553, + 140 + ], + "type": "text", + "content": "IDEA: Inverted Text with Cooperative Deformable Aggregation for Multi-modal Object Re-Identification" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 164, + 161, + 446, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 161, + 446, + 175 + ], + "spans": [ + { + "bbox": [ + 164, + 161, + 446, + 175 + ], + "type": "text", + "content": "Yuhao Wang, Yongfeng Lv, Pingping Zhang* Huchuan Lu" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 73, + 175, + 536, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 175, + 536, + 190 + ], + "spans": [ + { + "bbox": [ + 73, + 175, + 536, + 190 + ], + "type": "text", + "content": "School of Future Technology, School of Artificial Intelligence, Dalian University of Technology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 140, + 191, + 468, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 191, + 468, + 203 + ], + "spans": [ + { + "bbox": [ + 140, + 191, + 468, + 203 + ], + "type": "text", + "content": "{924973292,1vyf}@mail.dlut.edu.cn，{zhpp,lhchuan}@dlut.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 256, + 296, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 256, + 296, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 256, + 296, + 568 + ], + "type": "text", + "content": "Multi-modal object Re-Identification (ReID) aims to retrieve specific objects by utilizing complementary information from various modalities. However, existing methods focus on fusing heterogeneous visual features, neglecting the potential benefits of text-based semantic information. To address this issue, we first construct three text-enhanced multi-modal object ReID benchmarks. To be specific, we propose a standardized multi-modal caption generation pipeline for structured and concise text annotations with Multi-modal Large Language Models (MLLMs). Besides, current methods often directly aggregate multi-modal information without selecting representative local features, leading to redundancy and high complexity. To address the above issues, we introduce IDEA, a novel feature learning framework comprising the Inverted Multi-modal Feature Extractor (IMFE) and Cooperative Deformable Aggregation (CDA). The IMFE utilizes Modal Prefixes and an InverseNet to integrate multi-modal information with semantic guidance from inverted text. The CDA adaptively generates sampling positions, enabling the model to focus on the interplay between global features and discriminative local features. With the constructed benchmarks and the proposed modules, our framework can generate more robust multi-modal features under complex scenarios. Extensive experiments on three multi-modal object ReID benchmarks demonstrate the effectiveness of our proposed method." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 590, + 137, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 137, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 137, + 602 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 611, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 296, + 696 + ], + "type": "text", + "content": "Object Re-Identification (ReID) focuses on retrieving the same object across different camera views. While significant progress has been made with RGB images [9, 13, 23–25, 30, 35, 39, 40, 45, 47, 53], existing methods are limited under environmental challenges like dark and glare, reducing their robustness in real-world applications. Multi-modal imaging, which combines data from multiple spectra such" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 228, + 553, + 289 + ], + "blocks": [ + { + "bbox": [ + 316, + 228, + 553, + 289 + ], + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 289 + ], + "type": "image", + "image_path": "bc9351cc02ca02e4551f39b1f49c788ff2b4e5e4e7c0b26d8a84a1c573598573.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 316, + 292, + 551, + 340 + ], + "blocks": [ + { + "bbox": [ + 316, + 292, + 551, + 340 + ], + "lines": [ + { + "bbox": [ + 316, + 292, + 551, + 340 + ], + "spans": [ + { + "bbox": [ + 316, + 292, + 551, + 340 + ], + "type": "image", + "image_path": "65fe39a059a8a2886de5639e383a213b6b0371f50791fd952292d1c02bd591eb.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 343, + 555, + 389 + ], + "lines": [ + { + "bbox": [ + 313, + 343, + 555, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 343, + 555, + 389 + ], + "type": "text", + "content": "Figure 1. Overall illustration of our motivations and proposed framework. (a) Our multi-modal caption generation pipeline. (b) Limitations of existing MLLM-generated captions. (c) Comparison between previous methods and our proposed IDEA." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 402, + 554, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 554, + 450 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 554, + 450 + ], + "type": "text", + "content": "as RGB, Near Infrared (NIR) and Thermal Infrared (TIR), offers a promising solution. By leveraging complementary information, multi-modal ReID methods [4, 26, 37, 54] improve feature robustness under challenging conditions." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 451, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 451, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 451, + 555, + 715 + ], + "type": "text", + "content": "Beyond visual information fusion for different spectra, integrating language with vision can enhance ReID performance [11, 27, 34]. However, due to the absence of text annotations for images, most existing methods rely solely on visual features, overlooking the benefits of text-based semantic information. To address this limitation, some studies augment RGB datasets with manual text annotations [6, 20]. While this strategy boosts performance, it is both time-consuming and labor-intensive. With the advent of Multimodal Large Language Models (MLLMs), image caption has made significant progress. Recently, researchers start to use MLLMs to generate textual descriptions for RGB images [11, 14]. However, these methods face two main challenges. (1) As shown in Fig. 1 (b), the randomness in caption generation often leads to redundancy, resulting in long and complex sentences. Additionally, the text structure varies across images. In ReID tasks, such captions may exceed input length limits of a text encoder [29, 55], with redundant content diluting the essential information. (2) Current methods primarily focus on RGB images, where the generated captions lack sufficient details. However, multimodal imaging can capture crucial information in complex" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29701" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "text", + "content": "environments, providing enough information for MLLMs to generate informative captions. Despite this, no existing methods provide text annotations for multi-modal images. To bridge this gap, we construct three text-enhanced multi-modal object ReID benchmarks. Meanwhile, another challenge is the aggregation of multi-modal information. As shown in Fig. 1 (c), previous methods directly aggregate the heterogeneous information, leading to high computational complexity and noise interference [54]. To address above issues, we propose a multi-modal feature learning framework named IDEA to introduce Inverted text with cooperative DEformable Aggregation for multi-modal object ReID." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 218, + 296, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 218, + 296, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 218, + 296, + 578 + ], + "type": "text", + "content": "Technically, we first develop a standardized pipeline for multi-modal caption generation. It extends existing datasets with text annotations across spectral modalities. Specifically, as shown in Fig. 1 (a), our caption generation pipeline consists of two steps. In the first step, we use image-template pairs in conjunction with MLLMs to generate informative captions. Then, we leverage the powerful keyword extraction capabilities of MLLMs to extract predefined attributes from the generated captions to produce structured and concise sentences. Building upon the captions, we propose the IDEA framework, which consists of two key components: the Inverted Multi-modal Feature Extractor (IMFE) and Cooperative Deformable Aggregation (CDA). First, we utilize the IMFE to extract integrated multi-modal features using Modal Prefixes and an InverseNet. Specifically, Modal Prefixes are designed to distinguish different modalities, reducing the impact of conflicting information across modalities. The InverseNet further exploits the semantic information from inverted text to enhance feature robustness. Furthermore, we propose CDA to adaptively aggregate discriminative local information. Specifically, based on the aggregated multi-modal information, we adaptively generate sampling positions to enable the model to focus on the interplay between global features and discriminative local information. Through these modules, our proposed framework effectively utilizes semantic guidance from texts while adaptively aggregating discriminative multi-modal information. Extensive experiments on three benchmark datasets demonstrate the effectiveness of our approach. In summary, our contributions are as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 580, + 295, + 710 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 580, + 295, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 580, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 580, + 295, + 613 + ], + "type": "text", + "content": "- We construct three text-enhanced multi-modal object ReID benchmarks, providing a structured caption generation pipeline across multiple spectral modalities." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 616, + 294, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 294, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 294, + 685 + ], + "type": "text", + "content": "- We introduce a novel feature learning framework named IDEA, which includes the Inverted Multi-modal Feature Extractor (IMFE) and Cooperative Deformable Aggregation (CDA). The IMFE integrates multi-modal features using Modal Prefixes and an InverseNet, while the CDA adaptively aggregates discriminative local information." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 687, + 294, + 710 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 687, + 294, + 710 + ], + "spans": [ + { + "bbox": [ + 55, + 687, + 294, + 710 + ], + "type": "text", + "content": "- Extensive experiments on three benchmark datasets validate the effectiveness of our proposed method." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 91, + 551, + 104 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 551, + 104 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 551, + 104 + ], + "type": "text", + "content": "2.1. Vision-language Learning in Re-Identification" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 110, + 555, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 110, + 555, + 482 + ], + "spans": [ + { + "bbox": [ + 313, + 110, + 555, + 482 + ], + "type": "text", + "content": "Vision-language learning plays a pivotal role in multimodal applications. As the field advances [5, 48, 49], there is a growing focus on exploring the interaction between visual and textual information in ReID tasks. Existing methods can be broadly divided into cross-modal and multimodal ReID. In cross-modal ReID, Text-to-Image ReID [6] focuses on matching text queries with image galleries. Current methods primarily emphasize feature alignment [17] and pre-training [32]. To expand application scenarios, Zhai et al. [52] combine text and sketch modalities in the query, while Chen et al. [3] propose an uncertain Query-to-RGB retrieval model to handle missing modalities. More recently, Li et al. [19] introduce flexible modality combinations and unify various uncertainties in retrieval. However, these methods focus on cross-modal alignment. In contrast, multi-modal ReID aims to leverage complementary information from multiple modalities, with both the query and gallery containing multi-modal data. The rise of vision-language models, particularly CLIP [29], has significantly advanced this area by facilitating image-text interactions. Li et al. [21] pioneer CLIP-ReID to learn from text templates for ReID tasks. However, the templates are not real descriptions of images, potentially restricting performance. To address this, Han et al. [11] leverage MLLMs to generate descriptions for RGB images. Despite improvements, existing methods fail to tackle the challenges of generating text for multi-spectral modalities and ensuring structural consistency in the generated descriptions. To address the above issues, we propose a structured multi-modal caption generation pipeline, enhancing the consistency of texts and providing informative annotations for multi-modal object ReID." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 491, + 512, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 491, + 512, + 504 + ], + "spans": [ + { + "bbox": [ + 314, + 491, + 512, + 504 + ], + "type": "text", + "content": "2.2. Multi-modal Object Re-Identification" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 510, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 715 + ], + "type": "text", + "content": "Multi-modal object ReID gains great attention due to its robustness in real-world applications. Research primarily focuses on leveraging complementary information from different modalities. Early works emphasize effective fusion strategies [8, 56]. To enhance modality-specific learning, Wang et al. [41] propose an interact-embed-enlarge framework to facilitate knowledge expansion. To address missing modal data, Zheng et al. [57] introduce pixel reconstruction to handle data inconsistency. Li et al. [18] use a coherence loss to guide feature fusion. Some methods further improve feature robustness with modality generation, graph learning and instance sampling [10, 12, 58]. Notably, Wang et al. [42] introduce a test-time training framework to mine inter-modal interactions. Wang et al. [37] develop TOP-ReID, which employs token permutation to guide modality fusion. Recently, Zhang et al. [54] propose diverse token selections to mitigate background noises. Despite promis" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29702" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 301, + 249 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 301, + 249 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 301, + 249 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 301, + 249 + ], + "type": "image", + "image_path": "63e21ece53b6cec570bc2a86cf9a1f3d839d0d538c8ce2077e1f37bc0e963d71.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 252, + 555, + 285 + ], + "lines": [ + { + "bbox": [ + 55, + 252, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 252, + 555, + 285 + ], + "type": "text", + "content": "Figure 2. The upper row presents the template used to annotate our dataset. The lower row provides detailed information about the annotated dataset. (a) Category statistics for our annotated person ReID dataset. (b) Example images and captions from the RGBNT201 dataset. (c) Category statistics for our annotated vehicle ReID dataset. (d) Example images and captions from the MSVR310 dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 308, + 70, + 553, + 249 + ], + "blocks": [ + { + "bbox": [ + 308, + 70, + 553, + 249 + ], + "lines": [ + { + "bbox": [ + 308, + 70, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 308, + 70, + 553, + 249 + ], + "type": "image", + "image_path": "a562de2e9af9bafa90e69e31f4a0edf4ed3a095704512f089f31924c5f357681.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 300, + 296, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 296, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 296, + 373 + ], + "type": "text", + "content": "ing results, they often overlook the semantic guidance of informative text annotations. Thus, we propose the IDEA framework, which incorporates generated text annotations to leverage semantic guidance. This framework adaptively aggregates discriminative local information, enhancing feature robustness in complex scenarios." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 382, + 162, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 382, + 162, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 382, + 162, + 396 + ], + "type": "text", + "content": "3. Proposed Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 402, + 296, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 402, + 296, + 486 + ], + "spans": [ + { + "bbox": [ + 55, + 402, + 296, + 486 + ], + "type": "text", + "content": "In this section, we introduce the multi-modal caption generation pipeline and the proposed feature learning framework. Fig. 2 illustrates the details of caption generation, including attribute definitions and examples. Fig. 3 presents the key modules of our proposed IDEA: the Inverted Multi-modal Feature Extractor (IMFE) and the Cooperative Deformable Aggregation (CDA). Details are described as follows." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 493, + 234, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 493, + 234, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 493, + 234, + 506 + ], + "type": "text", + "content": "3.1. Multi-modal Caption Generation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 510, + 295, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 569 + ], + "type": "text", + "content": "To bridge the gap in multi-modal text annotation, we propose the Multi-modal Caption Generation pipeline. To be specific, our pipeline consists of two steps. The first step generates informative descriptions, while the second step reuses the MLLMs to extract the predefined key attributes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 570, + 296, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 701 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 701 + ], + "type": "text", + "content": "Caption Generation. Taking our person annotation as an example, we define 8 attribute categories in Fig. 2 (a), including gender, belongings and so on. These attributes are mapped into the template shown in the top part of Fig. 2 (a). To tailor the description to each modality, we add specific prefixes: \"Write a comprehensive description of the person's overall appearance based on the [RGB/NIR/TIR] image, ...\" Besides, we incorporate commands to prevent the hallucination of the MLLMs [32], forming a complete prompt. This prompt, along with the corresponding image, is fed into the MLLMs to generate the textual description." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 702, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 702, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 702, + 295, + 714 + ], + "type": "text", + "content": "Attribute Extraction. After obtaining an initial descrip" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 300, + 555, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 555, + 469 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 555, + 469 + ], + "type": "text", + "content": "tion, we feed it back into the MLLMs to extract predefined attributes. After that, we populate these attributes into the same template we used for caption generation, forming a structured and concise description. As shown in Fig. 2 (b), the descriptions for each modality can align well with the corresponding images. For instance, the \"white t-shirt\" and \"purse\" in the RGB image are accurately captured. For the vehicle dataset, we follow similar pipelines. As illustrated in Fig. 2 (c), the text of the RGB image successfully captures the license plate number \"KQQ819\". Meanwhile, the car logo is more prominent in the TIR image, with \"Hyundai\" being identified. This approach allows us to establish a unified pipeline for multi-modal text annotation, ensuring structural descriptions across different modalities." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 477, + 422, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 477, + 422, + 488 + ], + "spans": [ + { + "bbox": [ + 314, + 477, + 422, + 488 + ], + "type": "text", + "content": "3.2. IDEA Framework" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 495, + 515, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 515, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 515, + 506 + ], + "type": "text", + "content": "3.2.1. Inverted Multi-modal Feature Extractor" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 510, + 555, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 616 + ], + "type": "text", + "content": "To leverage the semantic guidance from texts, we propose the Inverted Multi-modal Feature Extractor (IMFE). Unlike previous methods [37, 54] that focus on fusing multi-modal images, we incorporate multi-modal texts into the fusion. However, as indicated by the blue keywords in Fig. 2 (b) and (c), conflicts can arise due to the nature of multi-modal images. Directly aggregating contradictory information can lead to model confusion. Thus, IMFE employs Modal Prefixes and an InverseNet to address these issues." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "content": "Modal Prefixes. To enhance the model's awareness of different modal texts and mitigate the impact of conflicting information, we propose a simple method called Modal Prefixes. Taking the text input for the RGB branch as an example, as shown in the top left corner of Fig. 3, " + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "inline_equation", + "content": "T_{RGB}" + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "content": " represents the text annotation of the RGB image, while " + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "inline_equation", + "content": "P_{RGB}" + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "content": " is the modal prefix for RGB modality. " + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "inline_equation", + "content": "P_{RGB}" + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "content": " consists of two parts: a fixed text describing the characteristics of the" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29703" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 68, + 553, + 270 + ], + "blocks": [ + { + "bbox": [ + 55, + 68, + 553, + 270 + ], + "lines": [ + { + "bbox": [ + 55, + 68, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 68, + 553, + 270 + ], + "type": "image", + "image_path": "8b3f4150cacc93f1549feb3273c5b7a40558ff630451663822c023988f5c7eb2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 278, + 555, + 323 + ], + "lines": [ + { + "bbox": [ + 54, + 278, + 555, + 323 + ], + "spans": [ + { + "bbox": [ + 54, + 278, + 555, + 323 + ], + "type": "text", + "content": "Figure 3. Illustration of the proposed IDEA framework. The upper part depicts the Inverted Multi-modal Feature Extractor (IMFE). It employs modal prefixes and an InverseNet to incorporate semantic text guidance for feature discriminability. The lower part highlights the Cooperative Deformable Aggregation (CDA), which adaptively integrates discriminative local information with global features. With the integration of IMFE and CDA, IDEA effectively extracts discriminative multi-modal features for object ReID." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 338, + 296, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 338, + 296, + 422 + ], + "spans": [ + { + "bbox": [ + 54, + 338, + 296, + 422 + ], + "type": "text", + "content": "RGB image and learnable tokens for fine-tuning. Specifically, " + }, + { + "bbox": [ + 54, + 338, + 296, + 422 + ], + "type": "inline_equation", + "content": "P_{RGB}" + }, + { + "bbox": [ + 54, + 338, + 296, + 422 + ], + "type": "text", + "content": " is structured as: \"An image of a XXXX person in the visible spectrum, capturing natural colors and fine details: \", where \"XXXX\" is replaced by an equal number of learnable tokens [22]. Here, we denote the number of learnable tokens as " + }, + { + "bbox": [ + 54, + 338, + 296, + 422 + ], + "type": "inline_equation", + "content": "N_P" + }, + { + "bbox": [ + 54, + 338, + 296, + 422 + ], + "type": "text", + "content": ". Then, we concatenate " + }, + { + "bbox": [ + 54, + 338, + 296, + 422 + ], + "type": "inline_equation", + "content": "P_{RGB}" + }, + { + "bbox": [ + 54, + 338, + 296, + 422 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 338, + 296, + 422 + ], + "type": "inline_equation", + "content": "T_{RGB}" + }, + { + "bbox": [ + 54, + 338, + 296, + 422 + ], + "type": "text", + "content": " to form the text input " + }, + { + "bbox": [ + 54, + 338, + 296, + 422 + ], + "type": "inline_equation", + "content": "T_R" + }, + { + "bbox": [ + 54, + 338, + 296, + 422 + ], + "type": "text", + "content": " for the RGB branch:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 129, + 426, + 294, + 441 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 426, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 129, + 426, + 294, + 441 + ], + "type": "interline_equation", + "content": "T _ {R} = \\left[ P _ {R G B}, T _ {R G B} \\right]. \\tag {1}", + "image_path": "f2489aa782f6a8fb3e41b235d5c722eafec130f6998e00f7a862af11e553f116.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 445, + 295, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 445, + 295, + 503 + ], + "spans": [ + { + "bbox": [ + 54, + 445, + 295, + 503 + ], + "type": "text", + "content": "Other modalities follow the same structure to guide the model in distinguishing between different modal texts. With this simple yet effective method, the model can better understand the characteristics of each modality and reduce conflicts during multi-modal fusion." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 504, + 295, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 504, + 295, + 658 + ], + "spans": [ + { + "bbox": [ + 54, + 504, + 295, + 658 + ], + "type": "text", + "content": "InverseNet. To fully leverage the semantic information from texts, we propose the InverseNet. Previous methods [2, 11] often reverse global image information into pseudo-text tokens and use texts as the primary modality for downstream tasks. However, due to the limitations of MLLMs and the complexity of multi-modal image annotations, the generated texts may contain some errors and conflicts. To address this, we take a different direction, reversing global text information into pseudo-image tokens. In this reversal, we also incorporate rich information from the learnable tokens in Modal Prefixes. This approach enables the interaction between semantic information and image details, resulting in a discriminative feature representation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 659, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 296, + 696 + ], + "type": "text", + "content": "Specifically, as shown in the upper part of Fig. 3, the text " + }, + { + "bbox": [ + 55, + 659, + 296, + 696 + ], + "type": "inline_equation", + "content": "T_{m}" + }, + { + "bbox": [ + 55, + 659, + 296, + 696 + ], + "type": "text", + "content": " is first fed into the CLIP's text encoder " + }, + { + "bbox": [ + 55, + 659, + 296, + 696 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 55, + 659, + 296, + 696 + ], + "type": "text", + "content": " for extracting the semantic feature " + }, + { + "bbox": [ + 55, + 659, + 296, + 696 + ], + "type": "inline_equation", + "content": "\\hat{f}_m^t\\in \\mathbb{R}^C" + }, + { + "bbox": [ + 55, + 659, + 296, + 696 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 145, + 700, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 700, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 145, + 700, + 295, + 715 + ], + "type": "interline_equation", + "content": "\\hat {f} _ {m} ^ {t} = \\mathcal {T} (T _ {m}). \\tag {2}", + "image_path": "6bda83a31460835d8cb077c75f32e81dab5651550c48e53300706623d8b0ece9.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 338, + 555, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 555, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 555, + 411 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 338, + 555, + 411 + ], + "type": "inline_equation", + "content": "m \\in \\{R, N, T\\}" + }, + { + "bbox": [ + 313, + 338, + 555, + 411 + ], + "type": "text", + "content": " represents the RGB, NIR and TIR modalities, respectively. " + }, + { + "bbox": [ + 313, + 338, + 555, + 411 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 338, + 555, + 411 + ], + "type": "text", + "content": " denotes the embedding dimension. If learnable tokens exist in the modal prefixes, " + }, + { + "bbox": [ + 313, + 338, + 555, + 411 + ], + "type": "inline_equation", + "content": "\\hat{f}_m^t" + }, + { + "bbox": [ + 313, + 338, + 555, + 411 + ], + "type": "text", + "content": " will be updated by the average of learnable tokens. Then, we feed " + }, + { + "bbox": [ + 313, + 338, + 555, + 411 + ], + "type": "inline_equation", + "content": "\\hat{f}_m^t" + }, + { + "bbox": [ + 313, + 338, + 555, + 411 + ], + "type": "text", + "content": " into the InverseNet " + }, + { + "bbox": [ + 313, + 338, + 555, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 313, + 338, + 555, + 411 + ], + "type": "text", + "content": " to generate the pseudo token " + }, + { + "bbox": [ + 313, + 338, + 555, + 411 + ], + "type": "inline_equation", + "content": "f_m^t \\in \\mathbb{R}^{C'}" + }, + { + "bbox": [ + 313, + 338, + 555, + 411 + ], + "type": "text", + "content": " with the following equation:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 405, + 418, + 553, + 432 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 418, + 553, + 432 + ], + "spans": [ + { + "bbox": [ + 405, + 418, + 553, + 432 + ], + "type": "interline_equation", + "content": "f _ {m} ^ {t} = \\mathcal {I} \\left(\\hat {f} _ {m} ^ {t}\\right). \\tag {3}", + "image_path": "330adc1771e0c02bb62ccbcdab46a65cacb43c3a2465b7138d19efdc697d0e80.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 438, + 555, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 555, + 522 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 555, + 522 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 438, + 555, + 522 + ], + "type": "inline_equation", + "content": "C'" + }, + { + "bbox": [ + 313, + 438, + 555, + 522 + ], + "type": "text", + "content": " is the dimension of the pseudo token while " + }, + { + "bbox": [ + 313, + 438, + 555, + 522 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 313, + 438, + 555, + 522 + ], + "type": "text", + "content": " is essentially a simple layer of MLP. Meanwhile, for the image input " + }, + { + "bbox": [ + 313, + 438, + 555, + 522 + ], + "type": "inline_equation", + "content": "I_m" + }, + { + "bbox": [ + 313, + 438, + 555, + 522 + ], + "type": "text", + "content": ", we first patchify it into " + }, + { + "bbox": [ + 313, + 438, + 555, + 522 + ], + "type": "inline_equation", + "content": "N_l" + }, + { + "bbox": [ + 313, + 438, + 555, + 522 + ], + "type": "text", + "content": " patches. Then, we concatenate the pseudo token " + }, + { + "bbox": [ + 313, + 438, + 555, + 522 + ], + "type": "inline_equation", + "content": "f_m^t" + }, + { + "bbox": [ + 313, + 438, + 555, + 522 + ], + "type": "text", + "content": ", the patch tokens and a learnable token " + }, + { + "bbox": [ + 313, + 438, + 555, + 522 + ], + "type": "inline_equation", + "content": "f_m^v" + }, + { + "bbox": [ + 313, + 438, + 555, + 522 + ], + "type": "text", + "content": " to form the image feature " + }, + { + "bbox": [ + 313, + 438, + 555, + 522 + ], + "type": "inline_equation", + "content": "\\bar{F}_m \\in \\mathbb{R}^{(N_l + 2) \\times C'}" + }, + { + "bbox": [ + 313, + 438, + 555, + 522 + ], + "type": "text", + "content": ". Finally, we feed " + }, + { + "bbox": [ + 313, + 438, + 555, + 522 + ], + "type": "inline_equation", + "content": "\\bar{F}_m" + }, + { + "bbox": [ + 313, + 438, + 555, + 522 + ], + "type": "text", + "content": " into the CLIP's vision encoder " + }, + { + "bbox": [ + 313, + 438, + 555, + 522 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 313, + 438, + 555, + 522 + ], + "type": "text", + "content": " to obtain the integrated feature " + }, + { + "bbox": [ + 313, + 438, + 555, + 522 + ], + "type": "inline_equation", + "content": "F_m \\in \\mathbb{R}^{(N_l + 2) \\times C}" + }, + { + "bbox": [ + 313, + 438, + 555, + 522 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 403, + 529, + 553, + 543 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 529, + 553, + 543 + ], + "spans": [ + { + "bbox": [ + 403, + 529, + 553, + 543 + ], + "type": "interline_equation", + "content": "F _ {m} = \\mathcal {V} (\\bar {F} _ {m}). \\tag {4}", + "image_path": "aff51da6fe9f5c02e5bc7e2f807a9d110b90d259607c7b5bc87f7bc64daea52a.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 550, + 555, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 550, + 555, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 550, + 555, + 586 + ], + "type": "text", + "content": "By incorporating the InverseNet, we can effectively leverage the semantic information from texts and enhance the discriminability of the feature representation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 591, + 503, + 603 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 591, + 503, + 603 + ], + "spans": [ + { + "bbox": [ + 313, + 591, + 503, + 603 + ], + "type": "text", + "content": "3.2.2. Cooperative Deformable Aggregation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 605, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 689 + ], + "type": "text", + "content": "To effectively extract discriminative local information and adaptively aggregate complementary multi-modal features, we propose Cooperative Deformable Aggregation (CDA). CDA addresses the noise interference by adaptively generating sampling positions of key local regions. This module enables interactions between global features and discriminative local features, ensuring efficient feature aggregation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "Technically, as illustrated in the lower part of Fig. 3, we first extract patch tokens from the integrated feature " + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "inline_equation", + "content": "F_{m}" + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "29704" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 296, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 296, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 296, + 168 + ], + "type": "text", + "content": "These patches are then reshaped to form " + }, + { + "bbox": [ + 55, + 71, + 296, + 168 + ], + "type": "inline_equation", + "content": "\\hat{F}_m\\in \\mathbb{R}^{H\\times W\\times C}" + }, + { + "bbox": [ + 55, + 71, + 296, + 168 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 71, + 296, + 168 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 71, + 296, + 168 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 71, + 296, + 168 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 55, + 71, + 296, + 168 + ], + "type": "text", + "content": " represent the height and width of the local feature map, respectively. Next, we concatenate these features along the channel dimension and pass them into the Local Mixer " + }, + { + "bbox": [ + 55, + 71, + 296, + 168 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 71, + 296, + 168 + ], + "type": "text", + "content": " for interaction. Specifically, " + }, + { + "bbox": [ + 55, + 71, + 296, + 168 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 71, + 296, + 168 + ], + "type": "text", + "content": " comprises point-wise convolution " + }, + { + "bbox": [ + 55, + 71, + 296, + 168 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 55, + 71, + 296, + 168 + ], + "type": "text", + "content": ", the GELU activation function [15] " + }, + { + "bbox": [ + 55, + 71, + 296, + 168 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 71, + 296, + 168 + ], + "type": "text", + "content": " and depth-wise convolution " + }, + { + "bbox": [ + 55, + 71, + 296, + 168 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 71, + 296, + 168 + ], + "type": "text", + "content": ". This process yields the aggregated local features " + }, + { + "bbox": [ + 55, + 71, + 296, + 168 + ], + "type": "inline_equation", + "content": "\\hat{F}_A\\in \\mathbb{R}^{H_S\\times W_S\\times C}" + }, + { + "bbox": [ + 55, + 71, + 296, + 168 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 123, + 171, + 295, + 186 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 171, + 295, + 186 + ], + "spans": [ + { + "bbox": [ + 123, + 171, + 295, + 186 + ], + "type": "interline_equation", + "content": "\\hat {F} _ {A} = \\mathcal {M} ([ \\hat {F} _ {R}, \\hat {F} _ {N}, \\hat {F} _ {T} ]), \\tag {5}", + "image_path": "caeb7d67abb488c9dc13ba1b5b8f468c9cbe3ef0a448785c48453bec5f6f94c9.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 119, + 201, + 295, + 214 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 201, + 295, + 214 + ], + "spans": [ + { + "bbox": [ + 119, + 201, + 295, + 214 + ], + "type": "interline_equation", + "content": "\\mathcal {M} (\\mathcal {X}) = \\delta (\\mathcal {D} (\\delta (\\mathcal {P} (\\mathcal {X})))) \\tag {6}", + "image_path": "33d2ad597fc628d6c7ede18575ede893132720a64552e241013c2fe049d0bf8d.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 217, + 296, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 217, + 296, + 336 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 296, + 336 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 217, + 296, + 336 + ], + "type": "inline_equation", + "content": "[\\cdot]" + }, + { + "bbox": [ + 55, + 217, + 296, + 336 + ], + "type": "text", + "content": " means concatenation. " + }, + { + "bbox": [ + 55, + 217, + 296, + 336 + ], + "type": "inline_equation", + "content": "H_{S}" + }, + { + "bbox": [ + 55, + 217, + 296, + 336 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 217, + 296, + 336 + ], + "type": "inline_equation", + "content": "W_{S}" + }, + { + "bbox": [ + 55, + 217, + 296, + 336 + ], + "type": "text", + "content": " represent the number of sampled regions in the height and width dimensions, respectively. Since the convolution " + }, + { + "bbox": [ + 55, + 217, + 296, + 336 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 217, + 296, + 336 + ], + "type": "text", + "content": " aggregates local information in the receptive field of the kernel, we define the center of each convolution operation as the reference points " + }, + { + "bbox": [ + 55, + 217, + 296, + 336 + ], + "type": "inline_equation", + "content": "P\\in \\mathbb{R}^{H_S\\times W_S\\times 2}" + }, + { + "bbox": [ + 55, + 217, + 296, + 336 + ], + "type": "text", + "content": ". Based on these reference points, the aggregated local information is then used to determine the direction in which the model should shift within the local region. Thus, we send " + }, + { + "bbox": [ + 55, + 217, + 296, + 336 + ], + "type": "inline_equation", + "content": "\\hat{F}_A" + }, + { + "bbox": [ + 55, + 217, + 296, + 336 + ], + "type": "text", + "content": " into the Offset Generator " + }, + { + "bbox": [ + 55, + 217, + 296, + 336 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 55, + 217, + 296, + 336 + ], + "type": "text", + "content": " to predict the offsets " + }, + { + "bbox": [ + 55, + 217, + 296, + 336 + ], + "type": "inline_equation", + "content": "\\Delta P\\in \\mathbb{R}^{H_S\\times W_S\\times 2}" + }, + { + "bbox": [ + 55, + 217, + 296, + 336 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 137, + 340, + 294, + 354 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 340, + 294, + 354 + ], + "spans": [ + { + "bbox": [ + 137, + 340, + 294, + 354 + ], + "type": "interline_equation", + "content": "\\Delta P = \\mathcal {G} (\\hat {F} _ {A}) * k, \\tag {7}", + "image_path": "f28db7a46aad605a374a8269c7d17efba02e1fffb180bfe17f095f995bc29cbe.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 358, + 296, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 358, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 358, + 296, + 407 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 358, + 296, + 407 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 358, + 296, + 407 + ], + "type": "text", + "content": " scales the offset magnitude [44] and " + }, + { + "bbox": [ + 55, + 358, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 55, + 358, + 296, + 407 + ], + "type": "text", + "content": " denotes a linear layer. Finally, we apply the offsets " + }, + { + "bbox": [ + 55, + 358, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\Delta P" + }, + { + "bbox": [ + 55, + 358, + 296, + 407 + ], + "type": "text", + "content": " to the reference points " + }, + { + "bbox": [ + 55, + 358, + 296, + 407 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 358, + 296, + 407 + ], + "type": "text", + "content": " to get the sampling locations " + }, + { + "bbox": [ + 55, + 358, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\hat{P} \\in \\mathbb{R}^{H_S \\times W_S \\times 2}" + }, + { + "bbox": [ + 55, + 358, + 296, + 407 + ], + "type": "text", + "content": ", which are shared across modalities [54]:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 145, + 410, + 294, + 423 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 410, + 294, + 423 + ], + "spans": [ + { + "bbox": [ + 145, + 410, + 294, + 423 + ], + "type": "interline_equation", + "content": "\\hat {P} = P + \\Delta P. \\tag {8}", + "image_path": "bc45d564bfeb1c5bbc77aa0706bb7db1bd64650250f11f2fdf4338dc0fb16e2a.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 427, + 296, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 427, + 296, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 427, + 296, + 548 + ], + "type": "text", + "content": "Next, we sample the local features " + }, + { + "bbox": [ + 55, + 427, + 296, + 548 + ], + "type": "inline_equation", + "content": "\\hat{F}_m" + }, + { + "bbox": [ + 55, + 427, + 296, + 548 + ], + "type": "text", + "content": " at the positions " + }, + { + "bbox": [ + 55, + 427, + 296, + 548 + ], + "type": "inline_equation", + "content": "\\hat{P}" + }, + { + "bbox": [ + 55, + 427, + 296, + 548 + ], + "type": "text", + "content": " using the bilinear interpolation to obtain the discriminative local features. The features are concatenated to form the sampled feature " + }, + { + "bbox": [ + 55, + 427, + 296, + 548 + ], + "type": "inline_equation", + "content": "F_S \\in \\mathbb{R}^{3N_S \\times C}" + }, + { + "bbox": [ + 55, + 427, + 296, + 548 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 427, + 296, + 548 + ], + "type": "inline_equation", + "content": "N_S = H_S \\times W_S" + }, + { + "bbox": [ + 55, + 427, + 296, + 548 + ], + "type": "text", + "content": " represents the number of sampled regions. At this stage, we utilize the cooperative information from multiple modalities to generate deformable local discriminative features. As illustrated by the sampling points visualization in Fig. 3, the adaptive offsets effectively concentrate on critical semantic regions of the human body, enhancing the discrimination." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 548, + 296, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 548, + 296, + 631 + ], + "spans": [ + { + "bbox": [ + 55, + 548, + 296, + 631 + ], + "type": "text", + "content": "To promote the interaction between global features and discriminative local features, we incorporate cross attention [7]. Specifically, we extract the visual class token and pseudo token from the integrated feature " + }, + { + "bbox": [ + 55, + 548, + 296, + 631 + ], + "type": "inline_equation", + "content": "F_{m}" + }, + { + "bbox": [ + 55, + 548, + 296, + 631 + ], + "type": "text", + "content": " to construct the global feature " + }, + { + "bbox": [ + 55, + 548, + 296, + 631 + ], + "type": "inline_equation", + "content": "F_{G} \\in \\mathbb{R}^{6 \\times C}" + }, + { + "bbox": [ + 55, + 548, + 296, + 631 + ], + "type": "text", + "content": ". Then, the global feature " + }, + { + "bbox": [ + 55, + 548, + 296, + 631 + ], + "type": "inline_equation", + "content": "F_{G}" + }, + { + "bbox": [ + 55, + 548, + 296, + 631 + ], + "type": "text", + "content": " is used as the query, while the sampled local feature " + }, + { + "bbox": [ + 55, + 548, + 296, + 631 + ], + "type": "inline_equation", + "content": "F_{S}" + }, + { + "bbox": [ + 55, + 548, + 296, + 631 + ], + "type": "text", + "content": " serves as the key and value to facilitate interaction:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 636, + 295, + 650 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 636, + 295, + 650 + ], + "spans": [ + { + "bbox": [ + 115, + 636, + 295, + 650 + ], + "type": "interline_equation", + "content": "F _ {C D A} = F _ {G} + \\mathcal {C A} \\left(F _ {G}, F _ {S}\\right), \\tag {9}", + "image_path": "2628689e001fbfc457f4ac0c09273661ccc4f859c55e1ffcf9c70a79dc631581.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "F_{CDA} \\in \\mathbb{R}^{6 \\times C}" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": " represents the final feature while " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{CA}" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": " denotes the cross attention block. By incorporating the CDA, we can effectively extract discriminative local information and adaptively aggregate complementary multimodal features, enhancing the model's discriminability." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 331, + 70, + 537, + 289 + ], + "blocks": [ + { + "bbox": [ + 331, + 70, + 537, + 289 + ], + "lines": [ + { + "bbox": [ + 331, + 70, + 537, + 289 + ], + "spans": [ + { + "bbox": [ + 331, + 70, + 537, + 289 + ], + "type": "table", + "html": "
MethodsRGBNT201
mAPR-1R-5R-10
Multi-modalHAMNet [18]27.726.341.551.7
PFNet [56]38.538.952.058.4
IEEE [41]47.544.457.163.6
DENet [57]42.442.255.364.5
LRMM [43]52.353.464.673.2
UniCat* [4]57.055.7--
HTT* [42]71.173.483.187.3
TOP-ReID* [37]72.376.684.789.4
EDITOR* [54]66.568.381.188.2
RSCNet* [50]68.272.5--
WTSF-ReID* [51]67.972.283.489.7
MambaPro† [36]78.983.489.891.9
DeMo† [38]79.082.388.892.0
IDEA†80.282.190.093.3
", + "image_path": "fdd1d6aa4988e86a56e8c3399c79a24b790186aaf758e12c067b9a456ed165a1.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 293, + 555, + 327 + ], + "lines": [ + { + "bbox": [ + 313, + 293, + 555, + 327 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 555, + 327 + ], + "type": "text", + "content": "Table 1. Performance comparison on RGBNT201. Best results are in bold, the second bests are underlined. † denotes CLIP-based methods, * indicates ViT-based while others are CNN-based ones." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 314, + 335, + 427, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 335, + 427, + 348 + ], + "spans": [ + { + "bbox": [ + 314, + 335, + 427, + 348 + ], + "type": "text", + "content": "3.3. Objective Function" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 353, + 555, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 353, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 353, + 555, + 426 + ], + "type": "text", + "content": "As shown in Fig. 3, we optimize IDEA through losses applied to multiple features. To maintain consistency with prior works [37, 54], we separately extract image and text features from " + }, + { + "bbox": [ + 313, + 353, + 555, + 426 + ], + "type": "inline_equation", + "content": "F_{G}" + }, + { + "bbox": [ + 313, + 353, + 555, + 426 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 353, + 555, + 426 + ], + "type": "inline_equation", + "content": "F_{CDA}" + }, + { + "bbox": [ + 313, + 353, + 555, + 426 + ], + "type": "text", + "content": ", resulting in " + }, + { + "bbox": [ + 313, + 353, + 555, + 426 + ], + "type": "inline_equation", + "content": "F_{G}^{v}, F_{G}^{t}" + }, + { + "bbox": [ + 313, + 353, + 555, + 426 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 353, + 555, + 426 + ], + "type": "inline_equation", + "content": "F_{CDA}^{v}, F_{CDA}^{t} \\in \\mathbb{R}^{3C}" + }, + { + "bbox": [ + 313, + 353, + 555, + 426 + ], + "type": "text", + "content": ". For each feature, we apply label smoothing cross-entropy loss [31] and triplet loss [16]:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 369, + 434, + 555, + 447 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 434, + 555, + 447 + ], + "spans": [ + { + "bbox": [ + 369, + 434, + 555, + 447 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {g} (\\mathcal {F}) = \\mathcal {L} _ {C E} (\\mathcal {F}) + \\mathcal {L} _ {T r i} (\\mathcal {F}), \\tag {10}", + "image_path": "c56f268e0ea516c3fa7e487f63ef87f8bb9965c9060b6b4fab0b43ee7df275ab.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 455, + 554, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 455, + 554, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 455, + 554, + 491 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 455, + 554, + 491 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 313, + 455, + 554, + 491 + ], + "type": "text", + "content": " denotes the input feature. Here, " + }, + { + "bbox": [ + 313, + 455, + 554, + 491 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CE}" + }, + { + "bbox": [ + 313, + 455, + 554, + 491 + ], + "type": "text", + "content": " is the label smoothing cross-entropy loss while " + }, + { + "bbox": [ + 313, + 455, + 554, + 491 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{Tri}" + }, + { + "bbox": [ + 313, + 455, + 554, + 491 + ], + "type": "text", + "content": " means the triplet loss. The overall objective function is then formulated as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 318, + 498, + 553, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 498, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 318, + 498, + 553, + 514 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {g} \\left(F _ {G} ^ {v}\\right) + \\mathcal {L} _ {g} \\left(F _ {C D A} ^ {v}\\right) + \\mathcal {L} _ {g} \\left(F _ {G} ^ {t}\\right) + \\mathcal {L} _ {g} \\left(F _ {C D A} ^ {t}\\right). \\tag {11}", + "image_path": "b4d57b4b410ec4172dc28d36e8cb1acebbf33269fa5f50da4b31cf48847c8f01.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 521, + 395, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 521, + 395, + 534 + ], + "spans": [ + { + "bbox": [ + 314, + 521, + 395, + 534 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 540, + 497, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 497, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 497, + 552 + ], + "type": "text", + "content": "4.1. Datasets and Evaluation Protocols" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 557, + 555, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 555, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 555, + 700 + ], + "type": "text", + "content": "Datasets. We evaluate the proposed method on three multimodal object ReID benchmarks. To efficiently annotate multi-spectral images without hardware constraints, we employ the API-based Qwen-VL [1] for automated textual description generation. To be specific, RGBNT201 [56] comprises 4,787 image triplets with 14,361 annotations, each averaging 35.47 characters and encompassing 8 distinct attributes. MSVR310 [58] contains 2,087 image triplets and 6,261 annotations, with an average length of 56.06 characters and 6 attributes. RGBNT100 [18] includes the largest number of triplets at 17,250 and 51,750 annotations with an average length of 56.25 characters and 6 attributes." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 701, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 554, + 714 + ], + "type": "text", + "content": "Evaluation Protocols. The performance is evaluated using" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29705" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 72, + 70, + 279, + 343 + ], + "blocks": [ + { + "bbox": [ + 72, + 70, + 279, + 343 + ], + "lines": [ + { + "bbox": [ + 72, + 70, + 279, + 343 + ], + "spans": [ + { + "bbox": [ + 72, + 70, + 279, + 343 + ], + "type": "table", + "html": "
MethodsRGBNT100MSVR310
mAPR-1mAPR-1
Multi-modalHAMNet [18]74.593.327.142.3
PFNet [56]68.194.123.537.4
GAFNet [10]74.493.4--
GPFNet [12]75.094.5--
CCNet [58]77.296.336.455.2
LRMM [43]78.696.736.749.7
GraFT* [46]76.694.3--
UniCat* [4]79.496.2--
PHT* [28]79.992.7--
HTT* [42]75.792.6--
TOP-ReID* [37]81.296.435.944.6
EDITOR* [54]82.196.439.049.3
FACENet* [59]81.596.936.254.1
RSCNet* [50]82.396.639.549.6
WTSF-ReID* [51]82.296.539.249.1
MambaPro† [36]83.994.747.056.5
DeMo† [38]86.297.649.259.8
IDEA†87.296.547.062.4
", + "image_path": "c96cf1069b6916d51e45f29fd981af3a652c9f444b783a5448955cdb68b06e50.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 366, + 295, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 366, + 295, + 391 + ], + "spans": [ + { + "bbox": [ + 55, + 366, + 295, + 391 + ], + "type": "text", + "content": "mean Average Precision (mAP) and Cumulative Matching Characteristics (CMC) at Rank-K " + }, + { + "bbox": [ + 55, + 366, + 295, + 391 + ], + "type": "inline_equation", + "content": "(K = 1,5,10)" + }, + { + "bbox": [ + 55, + 366, + 295, + 391 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 396, + 188, + 409 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 188, + 409 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 188, + 409 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 414, + 296, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 414, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 414, + 296, + 594 + ], + "type": "text", + "content": "Our model is implemented using PyTorch and trained on an NVIDIA A800 GPU. The pre-trained model CLIP [29] is used for both the vision and text encoders. For the datasets, images triples are resized to " + }, + { + "bbox": [ + 55, + 414, + 296, + 594 + ], + "type": "inline_equation", + "content": "256 \\times 128" + }, + { + "bbox": [ + 55, + 414, + 296, + 594 + ], + "type": "text", + "content": " for RGBNT201 and " + }, + { + "bbox": [ + 55, + 414, + 296, + 594 + ], + "type": "inline_equation", + "content": "128 \\times 256" + }, + { + "bbox": [ + 55, + 414, + 296, + 594 + ], + "type": "text", + "content": " for MSVR310 and RGBNT100. Data augmentation includes random horizontal flipping, cropping and random erasing [60]. For RGBNT201 and MSVR310, the mini-batch size is set to 64, with 8 images sampled per identity. For RGBNT100, the mini-batch size is set to 128, with 16 images sampled per identity. We fine-tune the learnable modules using an initial learning rate of " + }, + { + "bbox": [ + 55, + 414, + 296, + 594 + ], + "type": "inline_equation", + "content": "3.5 \\times 10^{-6}" + }, + { + "bbox": [ + 55, + 414, + 296, + 594 + ], + "type": "text", + "content": ", which decays to " + }, + { + "bbox": [ + 55, + 414, + 296, + 594 + ], + "type": "inline_equation", + "content": "3.5 \\times 10^{-7}" + }, + { + "bbox": [ + 55, + 414, + 296, + 594 + ], + "type": "text", + "content": " during training. For the RGBNT201 dataset, the values of " + }, + { + "bbox": [ + 55, + 414, + 296, + 594 + ], + "type": "inline_equation", + "content": "N_{p}" + }, + { + "bbox": [ + 55, + 414, + 296, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 414, + 296, + 594 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 414, + 296, + 594 + ], + "type": "text", + "content": " are set to 2 and 5, respectively. Other details are provided in the supplementary material. Code and annotations are available at IDEA." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 600, + 281, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 281, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 281, + 613 + ], + "type": "text", + "content": "4.3. Comparison with State-of-the-Art Methods" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "content": "Multi-modal Person ReID. In Tab. 1, we compare our method " + }, + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{IDEA}^{\\dagger}" + }, + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "content": " with existing multi-modal approaches on the RGBNT201 dataset. Leveraging complementary information from different modalities, multi-modal methods exhibit superior performance. Specifically, our proposed " + }, + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{IDEA}^{\\dagger}" + }, + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "content": " achieves " + }, + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "inline_equation", + "content": "80.2\\%" + }, + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "content": " mAP and " + }, + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "inline_equation", + "content": "82.1\\%" + }, + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "content": " Rank-1 accuracy, surpassing TOP-ReID* by " + }, + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "inline_equation", + "content": "7.9\\%" + }, + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "content": " in mAP and " + }, + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "inline_equation", + "content": "5.5\\%" + }, + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "content": " in Rank-1. Compared with other leading methods like " + }, + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{HTT}^*" + }, + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "content": " and" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 348, + 70, + 519, + 140 + ], + "blocks": [ + { + "bbox": [ + 77, + 346, + 273, + 357 + ], + "lines": [ + { + "bbox": [ + 77, + 346, + 273, + 357 + ], + "spans": [ + { + "bbox": [ + 77, + 346, + 273, + 357 + ], + "type": "text", + "content": "Table 2. Performance on RGBNT100 and MSVR310." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 348, + 70, + 519, + 140 + ], + "lines": [ + { + "bbox": [ + 348, + 70, + 519, + 140 + ], + "spans": [ + { + "bbox": [ + 348, + 70, + 519, + 140 + ], + "type": "table", + "html": "
IndexModulesMetrics
TextIMFECDAmAPRank-1
A×××70.372.1
B✓××73.475.8
C✓✓×77.281.1
D✓✓✓80.282.1
", + "image_path": "dc968a66c9b664266380eed0c575ec6559ed42a3450542dd4c54c5fce15e85f8.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 330, + 163, + 537, + 241 + ], + "blocks": [ + { + "bbox": [ + 329, + 144, + 538, + 156 + ], + "lines": [ + { + "bbox": [ + 329, + 144, + 538, + 156 + ], + "spans": [ + { + "bbox": [ + 329, + 144, + 538, + 156 + ], + "type": "text", + "content": "Table 3. Performance comparison with different modules." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 330, + 163, + 537, + 241 + ], + "lines": [ + { + "bbox": [ + 330, + 163, + 537, + 241 + ], + "spans": [ + { + "bbox": [ + 330, + 163, + 537, + 241 + ], + "type": "table", + "html": "
IndexIMFEMetrics
InverseNetPrefixesPromptmAPRank-1
A××-72.675.1
B×✓×73.475.8
C✓×-73.777.3
D✓✓×75.478.6
E✓✓✓77.281.1
", + "image_path": "b59fb3e12ee78756981078e3e8ec73983fd1bdb9c7848c8a31bc2475f60a176e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 329, + 244, + 538, + 255 + ], + "lines": [ + { + "bbox": [ + 329, + 244, + 538, + 255 + ], + "spans": [ + { + "bbox": [ + 329, + 244, + 538, + 255 + ], + "type": "text", + "content": "Table 4. Comparison with different components in IMFE." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 266, + 553, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 553, + 338 + ], + "type": "text", + "content": "EDITOR*, IDEA† shows a superior adaptability to complex scenarios, confirming the robustness. Besides, IDEA† outperforms the CLIP-based methods MambaPro† and DeMo† by 1.3% and 1.2% in mAP, respectively. These results highlight IDEA's ability to utilize semantic information from textual guidance for improved feature discrimination." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 338, + 554, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 554, + 435 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 554, + 435 + ], + "type": "text", + "content": "Multi-modal Vehicle ReID. In Tab. 2, we compare " + }, + { + "bbox": [ + 313, + 338, + 554, + 435 + ], + "type": "inline_equation", + "content": "\\mathsf{IDEA}^{\\dagger}" + }, + { + "bbox": [ + 313, + 338, + 554, + 435 + ], + "type": "text", + "content": " with state-of-the-art methods on the RGBNT100 and MSVR310 datasets. Our " + }, + { + "bbox": [ + 313, + 338, + 554, + 435 + ], + "type": "inline_equation", + "content": "\\mathsf{IDEA}^{\\dagger}" + }, + { + "bbox": [ + 313, + 338, + 554, + 435 + ], + "type": "text", + "content": " achieves an mAP of " + }, + { + "bbox": [ + 313, + 338, + 554, + 435 + ], + "type": "inline_equation", + "content": "87.2\\%" + }, + { + "bbox": [ + 313, + 338, + 554, + 435 + ], + "type": "text", + "content": ", surpassing EDITOR* by " + }, + { + "bbox": [ + 313, + 338, + 554, + 435 + ], + "type": "inline_equation", + "content": "5.1\\%" + }, + { + "bbox": [ + 313, + 338, + 554, + 435 + ], + "type": "text", + "content": " in mAP. Especially on the challenging MSVR310 dataset, " + }, + { + "bbox": [ + 313, + 338, + 554, + 435 + ], + "type": "inline_equation", + "content": "\\mathsf{IDEA}^{\\dagger}" + }, + { + "bbox": [ + 313, + 338, + 554, + 435 + ], + "type": "text", + "content": " achieves an mAP of " + }, + { + "bbox": [ + 313, + 338, + 554, + 435 + ], + "type": "inline_equation", + "content": "47.0\\%" + }, + { + "bbox": [ + 313, + 338, + 554, + 435 + ], + "type": "text", + "content": " and a Rank-1 accuracy of " + }, + { + "bbox": [ + 313, + 338, + 554, + 435 + ], + "type": "inline_equation", + "content": "62.4\\%" + }, + { + "bbox": [ + 313, + 338, + 554, + 435 + ], + "type": "text", + "content": ", outperforming EDITOR* by " + }, + { + "bbox": [ + 313, + 338, + 554, + 435 + ], + "type": "inline_equation", + "content": "8.0\\%" + }, + { + "bbox": [ + 313, + 338, + 554, + 435 + ], + "type": "text", + "content": " in mAP and " + }, + { + "bbox": [ + 313, + 338, + 554, + 435 + ], + "type": "inline_equation", + "content": "13.1\\%" + }, + { + "bbox": [ + 313, + 338, + 554, + 435 + ], + "type": "text", + "content": " in Rank-1 accuracy. These results verify IDEA's generalization ability." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 444, + 415, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 444, + 415, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 444, + 415, + 455 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 462, + 555, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 555, + 534 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 555, + 534 + ], + "type": "text", + "content": "We evaluate the effectiveness of the proposed modules on the RGBNT201 dataset. To be specific, our baseline comprises a three-branch vision encoder, which performs retrieval by concatenating the class tokens from three modalities. Upon incorporating the IMFE, we utilize " + }, + { + "bbox": [ + 313, + 462, + 555, + 534 + ], + "type": "inline_equation", + "content": "F_{G}^{t}" + }, + { + "bbox": [ + 313, + 462, + 555, + 534 + ], + "type": "text", + "content": " for retrieval and when the CDA is included, " + }, + { + "bbox": [ + 313, + 462, + 555, + 534 + ], + "type": "inline_equation", + "content": "F_{CDA}^{t}" + }, + { + "bbox": [ + 313, + 462, + 555, + 534 + ], + "type": "text", + "content": " is employed." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 534, + 554, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 554, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 554, + 665 + ], + "type": "text", + "content": "Effects of Key Modules. Tab. 3 presents the performance of various combinations of our proposed modules. Model A serves as the baseline, achieving an mAP of " + }, + { + "bbox": [ + 313, + 534, + 554, + 665 + ], + "type": "inline_equation", + "content": "70.3\\%" + }, + { + "bbox": [ + 313, + 534, + 554, + 665 + ], + "type": "text", + "content": " and Rank-1 accuracy of " + }, + { + "bbox": [ + 313, + 534, + 554, + 665 + ], + "type": "inline_equation", + "content": "72.1\\%" + }, + { + "bbox": [ + 313, + 534, + 554, + 665 + ], + "type": "text", + "content": ". Model B incorporates text information through parallel text encoders, improving the mAP to " + }, + { + "bbox": [ + 313, + 534, + 554, + 665 + ], + "type": "inline_equation", + "content": "73.4\\%" + }, + { + "bbox": [ + 313, + 534, + 554, + 665 + ], + "type": "text", + "content": " and Rank-1 accuracy to " + }, + { + "bbox": [ + 313, + 534, + 554, + 665 + ], + "type": "inline_equation", + "content": "75.8\\%" + }, + { + "bbox": [ + 313, + 534, + 554, + 665 + ], + "type": "text", + "content": ". With the inverted structure in IMFE, Model C further enhances performance, reaching an mAP of " + }, + { + "bbox": [ + 313, + 534, + 554, + 665 + ], + "type": "inline_equation", + "content": "77.2\\%" + }, + { + "bbox": [ + 313, + 534, + 554, + 665 + ], + "type": "text", + "content": " and Rank-1 of " + }, + { + "bbox": [ + 313, + 534, + 554, + 665 + ], + "type": "inline_equation", + "content": "81.1\\%" + }, + { + "bbox": [ + 313, + 534, + 554, + 665 + ], + "type": "text", + "content": ". Finally, Model D integrates CDA, delivering the best performance with an mAP of " + }, + { + "bbox": [ + 313, + 534, + 554, + 665 + ], + "type": "inline_equation", + "content": "80.2\\%" + }, + { + "bbox": [ + 313, + 534, + 554, + 665 + ], + "type": "text", + "content": ". These results fully validate the effectiveness and efficiency of the proposed modules." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": "Effects of Key Components in IMFE. Tab. 4 shows the performance of different components in the IMFE. Model A represents the parallel text encoders without the inverted structure or prefixes, achieving an mAP of " + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "inline_equation", + "content": "72.6\\%" + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": " and Rank-" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29706" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 71, + 70, + 277, + 150 + ], + "blocks": [ + { + "bbox": [ + 71, + 70, + 277, + 150 + ], + "lines": [ + { + "bbox": [ + 71, + 70, + 277, + 150 + ], + "spans": [ + { + "bbox": [ + 71, + 70, + 277, + 150 + ], + "type": "table", + "html": "
IndexCDAMetrics
SampleCross AttnShared OffsetmAPRank-1
A××-76.378.7
B×✓-77.079.8
C✓××76.878.9
D✓×✓77.680.4
E✓✓×79.581.7
F✓✓✓80.282.1
", + "image_path": "798abd2787957e6efe423fcf70cbc5bfb19d545292dfd153dd52d5986463281d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 72, + 171, + 277, + 209 + ], + "blocks": [ + { + "bbox": [ + 72, + 155, + 278, + 166 + ], + "lines": [ + { + "bbox": [ + 72, + 155, + 278, + 166 + ], + "spans": [ + { + "bbox": [ + 72, + 155, + 278, + 166 + ], + "type": "text", + "content": "Table 5. Comparison with different components in CDA." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 72, + 171, + 277, + 209 + ], + "lines": [ + { + "bbox": [ + 72, + 171, + 277, + 209 + ], + "spans": [ + { + "bbox": [ + 72, + 171, + 277, + 209 + ], + "type": "table", + "html": "
ModelmAPRank-1Rank-5Rank-10
IDEA w/o Text74.575.084.888.8
IDEA80.282.190.093.3
", + "image_path": "15e1afe55a93e5820810a97b0f5067a77ff6fbddcf1308b39a16250af60bb73f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 72, + 228, + 277, + 264 + ], + "blocks": [ + { + "bbox": [ + 79, + 212, + 271, + 223 + ], + "lines": [ + { + "bbox": [ + 79, + 212, + 271, + 223 + ], + "spans": [ + { + "bbox": [ + 79, + 212, + 271, + 223 + ], + "type": "text", + "content": "Table 6. Comparison of IDEA with and without text." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 72, + 228, + 277, + 264 + ], + "lines": [ + { + "bbox": [ + 72, + 228, + 277, + 264 + ], + "spans": [ + { + "bbox": [ + 72, + 228, + 277, + 264 + ], + "type": "table", + "html": "
ModelmAPRank-1Rank-5Rank-10
IDEA w/o Offset78.481.389.792.2
IDEA80.282.190.093.3
", + "image_path": "d22864016430797a1e246b4c41d1f1da09ee57f734354ab0f567866450c8dd90.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 75, + 268, + 274, + 279 + ], + "lines": [ + { + "bbox": [ + 75, + 268, + 274, + 279 + ], + "spans": [ + { + "bbox": [ + 75, + 268, + 274, + 279 + ], + "type": "text", + "content": "Table 7. Comparison of IDEA with and without offset." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 294, + 295, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 294, + 295, + 426 + ], + "spans": [ + { + "bbox": [ + 55, + 294, + 295, + 426 + ], + "type": "text", + "content": "1 accuracy of " + }, + { + "bbox": [ + 55, + 294, + 295, + 426 + ], + "type": "inline_equation", + "content": "75.1\\%" + }, + { + "bbox": [ + 55, + 294, + 295, + 426 + ], + "type": "text", + "content": ". With the Modal Prefixes in Model B, the mAP increases to " + }, + { + "bbox": [ + 55, + 294, + 295, + 426 + ], + "type": "inline_equation", + "content": "73.4\\%" + }, + { + "bbox": [ + 55, + 294, + 295, + 426 + ], + "type": "text", + "content": ", demonstrating the effectiveness of our prefixes mechanism. Model C introduces the inverted structure, achieving an mAP of " + }, + { + "bbox": [ + 55, + 294, + 295, + 426 + ], + "type": "inline_equation", + "content": "73.7\\%" + }, + { + "bbox": [ + 55, + 294, + 295, + 426 + ], + "type": "text", + "content": ". Model D combines the inverted structure and prefixes, further improving the mAP to " + }, + { + "bbox": [ + 55, + 294, + 295, + 426 + ], + "type": "inline_equation", + "content": "75.4\\%" + }, + { + "bbox": [ + 55, + 294, + 295, + 426 + ], + "type": "text", + "content": ". Finally, Model E integrates the learnable tokens in Modal Prefixes, achieving the best performance with an mAP of " + }, + { + "bbox": [ + 55, + 294, + 295, + 426 + ], + "type": "inline_equation", + "content": "77.2\\%" + }, + { + "bbox": [ + 55, + 294, + 295, + 426 + ], + "type": "text", + "content": ". Overall, inverted structure performs better than the parallel structure, achieving better performance. These results demonstrate the effectiveness of different components in our proposed IMFE." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 426, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 295, + 653 + ], + "type": "text", + "content": "Effects of Key Components in CDA. Based on the IMFE, Tab. 5 shows the performance of various components within CDA. When sampling is excluded, all local patches in " + }, + { + "bbox": [ + 55, + 426, + 295, + 653 + ], + "type": "inline_equation", + "content": "\\hat{F}_m" + }, + { + "bbox": [ + 55, + 426, + 295, + 653 + ], + "type": "text", + "content": " are processed. Without cross attention, multi-modal interaction uses only local features, omitting global context, with pooled " + }, + { + "bbox": [ + 55, + 426, + 295, + 653 + ], + "type": "inline_equation", + "content": "\\hat{F}_m" + }, + { + "bbox": [ + 55, + 426, + 295, + 653 + ], + "type": "text", + "content": " for retrieval. If shared offset is excluded, each modality independently generates offsets. Regarding numerical results, Model A, which does not incorporate these components, achieves an mAP of " + }, + { + "bbox": [ + 55, + 426, + 295, + 653 + ], + "type": "inline_equation", + "content": "76.3\\%" + }, + { + "bbox": [ + 55, + 426, + 295, + 653 + ], + "type": "text", + "content": " and Rank-1 accuracy of " + }, + { + "bbox": [ + 55, + 426, + 295, + 653 + ], + "type": "inline_equation", + "content": "78.7\\%" + }, + { + "bbox": [ + 55, + 426, + 295, + 653 + ], + "type": "text", + "content": ". Model B introduces cross attention, leading to a slight improvement in mAP to " + }, + { + "bbox": [ + 55, + 426, + 295, + 653 + ], + "type": "inline_equation", + "content": "77.0\\%" + }, + { + "bbox": [ + 55, + 426, + 295, + 653 + ], + "type": "text", + "content": ". Model C adds sampling, resulting in a " + }, + { + "bbox": [ + 55, + 426, + 295, + 653 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 55, + 426, + 295, + 653 + ], + "type": "text", + "content": " improvement in mAP over Model A by reducing noise in the local features. Model D incorporates shared offset, yielding a " + }, + { + "bbox": [ + 55, + 426, + 295, + 653 + ], + "type": "inline_equation", + "content": "0.8\\%" + }, + { + "bbox": [ + 55, + 426, + 295, + 653 + ], + "type": "text", + "content": " improvement in mAP and a " + }, + { + "bbox": [ + 55, + 426, + 295, + 653 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 55, + 426, + 295, + 653 + ], + "type": "text", + "content": " increase in Rank-1 accuracy compared to Model C. Model E combines cross attention and sampling, achieving an mAP of " + }, + { + "bbox": [ + 55, + 426, + 295, + 653 + ], + "type": "inline_equation", + "content": "79.5\\%" + }, + { + "bbox": [ + 55, + 426, + 295, + 653 + ], + "type": "text", + "content": ". Finally, Model F delivers the best performance with an mAP of " + }, + { + "bbox": [ + 55, + 426, + 295, + 653 + ], + "type": "inline_equation", + "content": "80.2\\%" + }, + { + "bbox": [ + 55, + 426, + 295, + 653 + ], + "type": "text", + "content": ". These results validate the effectiveness of each component within CDA." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "content": "Effects of Text Guidance in IDEA. Tab. 6 compares IDEA's performance with and without text input. Model A, lacking text input, achieves an mAP of " + }, + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "inline_equation", + "content": "74.5\\%" + }, + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "content": ". With text guidance, Model B improves to " + }, + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "inline_equation", + "content": "80.2\\%" + }, + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "content": " mAP, demonstrating the effectiveness of text in enhancing feature robustness." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 69, + 552, + 194 + ], + "blocks": [ + { + "bbox": [ + 317, + 69, + 552, + 194 + ], + "lines": [ + { + "bbox": [ + 317, + 69, + 552, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 69, + 552, + 194 + ], + "type": "image", + "image_path": "42f1415c7c9a58be752553db781581ac718e5da4ca738c7775c2b18dc4dc0c46.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 333, + 198, + 534, + 210 + ], + "lines": [ + { + "bbox": [ + 333, + 198, + 534, + 210 + ], + "spans": [ + { + "bbox": [ + 333, + 198, + 534, + 210 + ], + "type": "text", + "content": "Figure 4. Comparison with different hyper-parameters." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 315, + 214, + 553, + 292 + ], + "blocks": [ + { + "bbox": [ + 315, + 214, + 553, + 292 + ], + "lines": [ + { + "bbox": [ + 315, + 214, + 553, + 292 + ], + "spans": [ + { + "bbox": [ + 315, + 214, + 553, + 292 + ], + "type": "image", + "image_path": "fca242f2f5a4e5487dad847ce72d4a9894f5aad349ab7f3d556eec9e62a3135b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 325, + 297, + 541, + 308 + ], + "lines": [ + { + "bbox": [ + 325, + 297, + 541, + 308 + ], + "spans": [ + { + "bbox": [ + 325, + 297, + 541, + 308 + ], + "type": "text", + "content": "Figure 5. Visualization of the cosine similarity distribution." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 318, + 553, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 318, + 553, + 413 + ], + "spans": [ + { + "bbox": [ + 313, + 318, + 553, + 413 + ], + "type": "text", + "content": "Effects of Offset in IDEA. Tab. 7 compares the performance of IDEA with and without the offset mechanism. Model A uses aggregated local features with convolution to interact with global features, achieving an mAP of " + }, + { + "bbox": [ + 313, + 318, + 553, + 413 + ], + "type": "inline_equation", + "content": "78.4\\%" + }, + { + "bbox": [ + 313, + 318, + 553, + 413 + ], + "type": "text", + "content": " and Rank-1 accuracy of " + }, + { + "bbox": [ + 313, + 318, + 553, + 413 + ], + "type": "inline_equation", + "content": "81.3\\%" + }, + { + "bbox": [ + 313, + 318, + 553, + 413 + ], + "type": "text", + "content": ". Model B introduces the offset mechanism, improving the mAP to " + }, + { + "bbox": [ + 313, + 318, + 553, + 413 + ], + "type": "inline_equation", + "content": "80.2\\%" + }, + { + "bbox": [ + 313, + 318, + 553, + 413 + ], + "type": "text", + "content": " and Rank-1 accuracy to " + }, + { + "bbox": [ + 313, + 318, + 553, + 413 + ], + "type": "inline_equation", + "content": "82.1\\%" + }, + { + "bbox": [ + 313, + 318, + 553, + 413 + ], + "type": "text", + "content": ". This demonstrates that the offset mechanism helps the model focus on more important regions." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 414, + 554, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 414, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 414, + 554, + 510 + ], + "type": "text", + "content": "Effects of Prompt Lengths and Offset Factors. Fig. 4 illustrates the performance of IDEA with varying prompt lengths and offset factors. The results show that a prompt length of 2 and an offset factor of 5 yield the best performance. As the prompt length increases, the model's performance improves, but the gains become smaller as the prompt length grows, likely due to the text truncation. Regarding the offset factor, the model remains stable." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 517, + 440, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 440, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 440, + 529 + ], + "type": "text", + "content": "4.5. Visualization Analysis" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 534, + 553, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 553, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 553, + 605 + ], + "type": "text", + "content": "Cosine Similarity Distributions. Fig. 5 presents the distributions of cosine similarities among test features. The results verify that incorporating text enhances feature discrimination, while adding offset mechanism further amplifies the distinction between positive and negative samples, confirming the capture of more discriminative features." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": "Multi-modal Feature Distributions. Fig. 6 visualizes the feature distributions of different modules. Comparing Fig. 6 (a) and (b), the introduction with text guidance improves feature discrimination, as instances of the same ID become more compact. In Fig. 6 (c), with IMFE, the feature distributions become more discriminative than the parallel structure in Fig. 6 (b). Finally, in Fig. 6 (d), CDA further enhances feature discrimination. These visualizations demonstrate the effectiveness of our proposed modules." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29707" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 69, + 184, + 184 + ], + "blocks": [ + { + "bbox": [ + 67, + 69, + 184, + 184 + ], + "lines": [ + { + "bbox": [ + 67, + 69, + 184, + 184 + ], + "spans": [ + { + "bbox": [ + 67, + 69, + 184, + 184 + ], + "type": "image", + "image_path": "9f32aabe2868c593ac94fadff875e1fb100f6de5f1fc6612b6a9d47b4b1cac81.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 102, + 186, + 144, + 194 + ], + "lines": [ + { + "bbox": [ + 102, + 186, + 144, + 194 + ], + "spans": [ + { + "bbox": [ + 102, + 186, + 144, + 194 + ], + "type": "text", + "content": "(a) Baseline" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 186, + 70, + 302, + 184 + ], + "blocks": [ + { + "bbox": [ + 186, + 70, + 302, + 184 + ], + "lines": [ + { + "bbox": [ + 186, + 70, + 302, + 184 + ], + "spans": [ + { + "bbox": [ + 186, + 70, + 302, + 184 + ], + "type": "image", + "image_path": "eef9f1ac28e78e6a38b541afc67a93819827b37f3bc122c5b7c94b7f7d18bedc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 211, + 186, + 276, + 194 + ], + "lines": [ + { + "bbox": [ + 211, + 186, + 276, + 194 + ], + "spans": [ + { + "bbox": [ + 211, + 186, + 276, + 194 + ], + "type": "text", + "content": "(b) Baseline + Text" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 306, + 70, + 421, + 184 + ], + "blocks": [ + { + "bbox": [ + 306, + 70, + 421, + 184 + ], + "lines": [ + { + "bbox": [ + 306, + 70, + 421, + 184 + ], + "spans": [ + { + "bbox": [ + 306, + 70, + 421, + 184 + ], + "type": "image", + "image_path": "0f31d63e674c283aa75fe1c86f2e430489d4c1c2a4319543efdcdcc5d34c7edb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 328, + 186, + 398, + 194 + ], + "lines": [ + { + "bbox": [ + 328, + 186, + 398, + 194 + ], + "spans": [ + { + "bbox": [ + 328, + 186, + 398, + 194 + ], + "type": "text", + "content": "(c) Baseline + IMFE" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 425, + 70, + 540, + 184 + ], + "blocks": [ + { + "bbox": [ + 425, + 70, + 540, + 184 + ], + "lines": [ + { + "bbox": [ + 425, + 70, + 540, + 184 + ], + "spans": [ + { + "bbox": [ + 425, + 70, + 540, + 184 + ], + "type": "image", + "image_path": "a87a0b7294c0ce3b2bcc556bdcd2c763e077467476c9e966990f3e1010401015.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 435, + 186, + 531, + 194 + ], + "lines": [ + { + "bbox": [ + 435, + 186, + 531, + 194 + ], + "spans": [ + { + "bbox": [ + 435, + 186, + 531, + 194 + ], + "type": "text", + "content": "(d) Baseline + IMFE + CDA" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 65, + 214, + 285, + 356 + ], + "blocks": [ + { + "bbox": [ + 99, + 198, + 509, + 209 + ], + "lines": [ + { + "bbox": [ + 99, + 198, + 509, + 209 + ], + "spans": [ + { + "bbox": [ + 99, + 198, + 509, + 209 + ], + "type": "text", + "content": "Figure 6. Visualization of the feature distributions with t-SNE [33]. Different colors represent different identities." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 65, + 214, + 285, + 356 + ], + "lines": [ + { + "bbox": [ + 65, + 214, + 285, + 356 + ], + "spans": [ + { + "bbox": [ + 65, + 214, + 285, + 356 + ], + "type": "image", + "image_path": "22d5b7e53415d97b3606b49761366d21d2336aad9eb40db37adeff43076112cb.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 61, + 373, + 287, + 533 + ], + "blocks": [ + { + "bbox": [ + 88, + 358, + 261, + 369 + ], + "lines": [ + { + "bbox": [ + 88, + 358, + 261, + 369 + ], + "spans": [ + { + "bbox": [ + 88, + 358, + 261, + 369 + ], + "type": "text", + "content": "Figure 7. Visualization of the generated offsets." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 61, + 373, + 287, + 533 + ], + "lines": [ + { + "bbox": [ + 61, + 373, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 61, + 373, + 287, + 533 + ], + "type": "image", + "image_path": "4924e2091fc02253bb9a74d4c8a61b1bbf946e23402da9f61b4c517d2ce0aaca.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 536, + 291, + 546 + ], + "lines": [ + { + "bbox": [ + 59, + 536, + 291, + 546 + ], + "spans": [ + { + "bbox": [ + 59, + 536, + 291, + 546 + ], + "type": "text", + "content": "Figure 8. Rank list comparison between the baseline and IDEA." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 557, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 295, + 665 + ], + "type": "text", + "content": "Visualization of the Generated Offsets. Fig. 7 visualizes the offsets generated by CDA. Each arrow starts at the reference point and ends at the feature sampling point. We map these offsets onto the original image to identify the regions where the model shifts its attention. As shown, the offsets highlight discriminative areas, such as the head, bag and shoes. It demonstrates how the offset mechanism guides the model to focus on the most discriminative regions, effectively reducing the influence of noisy information." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 665, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 712 + ], + "type": "text", + "content": "Rank List Comparison. Fig. 8 compares the cross-camera rank lists from the baseline and IDEA. IDEA produces more accurate rankings, while the baseline yields noisier results. These visualizations validate the effectiveness of IDEA." + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 322, + 214, + 432, + 279 + ], + "blocks": [ + { + "bbox": [ + 322, + 214, + 432, + 279 + ], + "lines": [ + { + "bbox": [ + 322, + 214, + 432, + 279 + ], + "spans": [ + { + "bbox": [ + 322, + 214, + 432, + 279 + ], + "type": "image", + "image_path": "5a3011b1e7c2d19c6e5e0dd3b388f366503ea0e245456c32d8628bca3bc6029e.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 279, + 430, + 313 + ], + "lines": [ + { + "bbox": [ + 323, + 279, + 430, + 313 + ], + "spans": [ + { + "bbox": [ + 323, + 279, + 430, + 313 + ], + "type": "text", + "content": "The female is wearing a vague orange blouse with vague long white skirt and white shoes. The female has short black hair and appears to be young adult. The female is carrying earrings." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 435, + 214, + 545, + 279 + ], + "blocks": [ + { + "bbox": [ + 435, + 214, + 545, + 279 + ], + "lines": [ + { + "bbox": [ + 435, + 214, + 545, + 279 + ], + "spans": [ + { + "bbox": [ + 435, + 214, + 545, + 279 + ], + "type": "image", + "image_path": "dd5afab41ddcec815592e5bdca3dd541ee3c7af658a14b9815a0cb1ff5a5ccfd.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 437, + 279, + 539, + 313 + ], + "lines": [ + { + "bbox": [ + 437, + 279, + 539, + 313 + ], + "spans": [ + { + "bbox": [ + 437, + 279, + 539, + 313 + ], + "type": "text", + "content": "The female is wearing a white black stripe shirt with blue jeans. The female has short dark hair and appears to be around 30 years old. The female is carrying beige purse." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 323, + 313, + 432, + 414 + ], + "blocks": [ + { + "bbox": [ + 323, + 313, + 432, + 414 + ], + "lines": [ + { + "bbox": [ + 323, + 313, + 432, + 414 + ], + "spans": [ + { + "bbox": [ + 323, + 313, + 432, + 414 + ], + "type": "image", + "image_path": "1fb28e64d3525493c829ebf05254dfd274b60d22d2bdd261f92435761116df85.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 419, + 555, + 441 + ], + "lines": [ + { + "bbox": [ + 313, + 419, + 555, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 419, + 555, + 441 + ], + "type": "text", + "content": "Figure 9. Visualization of the channel activation maps. Each set includes the original image, baseline and IDEA map, respectively." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 435, + 313, + 545, + 378 + ], + "blocks": [ + { + "bbox": [ + 435, + 313, + 545, + 378 + ], + "lines": [ + { + "bbox": [ + 435, + 313, + 545, + 378 + ], + "spans": [ + { + "bbox": [ + 435, + 313, + 545, + 378 + ], + "type": "image", + "image_path": "b5a87759c50629564a72d9a47b413eee466ba479fa99936f24fde31b11d5c22d.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 438, + 378, + 539, + 413 + ], + "lines": [ + { + "bbox": [ + 438, + 378, + 539, + 413 + ], + "spans": [ + { + "bbox": [ + 438, + 378, + 539, + 413 + ], + "type": "text", + "content": "The female is wearing a pink blouse with black stripes and white sneakers. The female has long dark hair and appears to be young. The female is carrying white purse." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 448, + 555, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 448, + 555, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 448, + 555, + 509 + ], + "type": "text", + "content": "Visualization of Channel Activation Maps. Fig. 9 compares the channel activation maps of our baseline model and IDEA. The incorporation of textual guidance helps the model focus on more discriminative regions, enhancing feature robustness and improving interpretability." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 314, + 518, + 388, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 518, + 388, + 530 + ], + "spans": [ + { + "bbox": [ + 314, + 518, + 388, + 530 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 537, + 555, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 537, + 555, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 537, + 555, + 658 + ], + "type": "text", + "content": "In this work, we propose IDEA, a novel feature learning framework for multi-modal object ReID. We first construct three text-enhanced multi-modal object ReID benchmarks using MLLMs, providing a structured caption generation pipeline. With the generated text, the Inverted Multi-modal Feature Extractor (IMFE) leverages semantic guidance from inverted texts while reducing fusion conflicts. Additionally, the Cooperative Deformable Aggregation (CDA) adaptively aggregates discriminative local features with global information. Experiments on three public ReID benchmarks demonstrate the effectiveness of our method." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 658, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 658, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 658, + 555, + 715 + ], + "type": "text", + "content": "Acknowledgements. This work was supported in part by the National Natural Science Foundation of China (No. 62101092, 62476044, 62388101) and Open Project of Anhui Provincial Key Laboratory of Multimodal Cognitive Computation, Anhui University (No. MMC202102, MMC202407)." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29708" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 1(2):3, 2023.5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 148, + 296, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 148, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 61, + 148, + 296, + 180 + ], + "type": "text", + "content": "[2] Alberto Baldrati, Lorenzo Agnolucci, Marco Bertini, and Alberto Del Bimbo. Zero-shot composed image retrieval with textual inversion. In ICCV, pages 15338-15347, 2023. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 182, + 295, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 182, + 295, + 214 + ], + "spans": [ + { + "bbox": [ + 62, + 182, + 295, + 214 + ], + "type": "text", + "content": "[3] Cuiqun Chen, Mang Ye, and Ding Jiang. Towards modality-agnostic person re-identification with descriptive query. In CVPR, pages 15128-15137, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 217, + 296, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 217, + 296, + 258 + ], + "spans": [ + { + "bbox": [ + 62, + 217, + 296, + 258 + ], + "type": "text", + "content": "[4] Jennifer Crawford, Haoli Yin, Luke McDermott, and Daniel Cummings. Unicat: Crafting a stronger fusion baseline for multimodal re-identification. arXiv preprint arXiv:2310.18812, 2023. 1, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 262, + 295, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 262, + 295, + 304 + ], + "spans": [ + { + "bbox": [ + 62, + 262, + 295, + 304 + ], + "type": "text", + "content": "[5] Haiwen Diao, Yufeng Cui, Xiaotong Li, Yueze Wang, Huchuan Lu, and Xinlong Wang. Unveiling encoder-free vision-language models. arXiv preprint arXiv:2406.11832, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 307, + 295, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 307, + 295, + 350 + ], + "spans": [ + { + "bbox": [ + 62, + 307, + 295, + 350 + ], + "type": "text", + "content": "[6] Zefeng Ding, Changxing Ding, Zhiyin Shao, and Dacheng Tao. Semantically self-aligned network for text-to-image part-aware person re-identification. arXiv preprint arXiv:2107.12666, 2021. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 352, + 295, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 352, + 295, + 417 + ], + "spans": [ + { + "bbox": [ + 62, + 352, + 295, + 417 + ], + "type": "text", + "content": "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 419, + 295, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 419, + 295, + 463 + ], + "spans": [ + { + "bbox": [ + 62, + 419, + 295, + 463 + ], + "type": "text", + "content": "[8] Yunpeng Gong, Liqing Huang, and Lifei Chen. Eliminate deviation with deviation for data augmentation and a general multi-modal data learning method. arXiv preprint arXiv:2101.08533, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 465, + 295, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 465, + 295, + 507 + ], + "spans": [ + { + "bbox": [ + 62, + 465, + 295, + 507 + ], + "type": "text", + "content": "[9] Yunpeng Gong, Zhun Zhong, Yansong Qu, Zhiming Luo, Rongrong Ji, and Min Jiang. Cross-modality perturbation synergy attack for person re-identification. In NeurIPS, 2024. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 510, + 295, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 510, + 295, + 542 + ], + "spans": [ + { + "bbox": [ + 57, + 510, + 295, + 542 + ], + "type": "text", + "content": "[10] Jinbo Guo, Xiaojing Zhang, Zhengyi Liu, and Yuan Wang. Generative and attentive fusion for multi-spectral vehicle re-identification. In ICSP, pages 1565–1572, 2022. 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 544, + 295, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 544, + 295, + 587 + ], + "spans": [ + { + "bbox": [ + 57, + 544, + 295, + 587 + ], + "type": "text", + "content": "[11] Qianru Han, Xinwei He, Zhi Liu, Sannyuya Liu, Ying Zhang, and Jinhai Xiang. Clip-scgi: Synthesized caption-guided inversion for person re-identification. arXiv preprint arXiv:2410.09382, 2024. 1, 2, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 590, + 295, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 590, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 57, + 590, + 295, + 623 + ], + "type": "text", + "content": "[12] Qiaolin He, Zefeng Lu, Zihan Wang, and Haifeng Hu. Graph-based progressive fusion network for multi-modality vehicle re-identification. TITS, pages 1-17, 2023. 2, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 624, + 295, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 295, + 657 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 295, + 657 + ], + "type": "text", + "content": "[13] Shuting He, Hao Luo, Pichao Wang, Fan Wang, Hao Li, and Wei Jiang. Transreid: Transformer-based object re-identification. In ICCV, pages 15013–15022, 2021. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 658, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 295, + 712 + ], + "type": "text", + "content": "[14] Weizhen He, Yiheng Deng, Shixiang Tang, Qihao Chen, Qingsong Xie, Yizhou Wang, Lei Bai, Feng Zhu, Rui Zhao, Wanli Ouyang, et al. Instruct-reid: A multi-purpose person re-identification task with instructions. In CVPR, pages 17521–17531, 2024. 1" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 712 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 95 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 95 + ], + "type": "text", + "content": "[15] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 96, + 554, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 554, + 128 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 554, + 128 + ], + "type": "text", + "content": "[16] Alexander Hermans, Lucas Beyer, and Bastian Leibe. In defense of the triplet loss for person re-identification. arXiv preprint arXiv:1703.07737, 2017. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 129, + 554, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 554, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 554, + 161 + ], + "type": "text", + "content": "[17] Ding Jiang and Mang Ye. Cross-modal implicit relation reasoning and aligning for text-to-image person retrieval. In CVPR, pages 2787-2797, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 163, + 554, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 554, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 554, + 194 + ], + "type": "text", + "content": "[18] Hongchao Li, Chenglong Li, Xianpeng Zhu, Aihua Zheng, and Bin Luo. Multi-spectral vehicle re-identification: A challenge. In AAAI, pages 11345–11353, 2020. 2, 5, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 196, + 554, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 196, + 554, + 227 + ], + "spans": [ + { + "bbox": [ + 317, + 196, + 554, + 227 + ], + "type": "text", + "content": "[19] He Li, Mang Ye, Ming Zhang, and Bo Du. All in one framework for multimodal re-identification in the wild. In CVPR, pages 17459–17469, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 228, + 554, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 228, + 554, + 260 + ], + "spans": [ + { + "bbox": [ + 317, + 228, + 554, + 260 + ], + "type": "text", + "content": "[20] Shuang Li, Tong Xiao, Hongsheng Li, Bolei Zhou, Dayu Yue, and Xiaogang Wang. Person search with natural language description. In CVPR, pages 1970-1979, 2017. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 262, + 554, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 262, + 554, + 293 + ], + "spans": [ + { + "bbox": [ + 317, + 262, + 554, + 293 + ], + "type": "text", + "content": "[21] Siyuan Li, Li Sun, and Qingli Li. Clip-reid: exploiting vision-language model for image re-identification without concrete text labels. In AAAI, pages 1405–1413, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 295, + 554, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 295, + 554, + 337 + ], + "spans": [ + { + "bbox": [ + 317, + 295, + 554, + 337 + ], + "type": "text", + "content": "[22] Yaowei Li, Zimo Liu, Wenming Yang, Yaowei Wang, Qingmin Liao, et al. Clip-based synergistic knowledge transfer for text-based person retrieval. arXiv preprint arXiv:2309.09496, 2023. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 339, + 554, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 339, + 554, + 381 + ], + "spans": [ + { + "bbox": [ + 317, + 339, + 554, + 381 + ], + "type": "text", + "content": "[23] Xuehu Liu, Pingping Zhang, Chenyang Yu, Huchuan Lu, and Xiaoyun Yang. Watching you: Global-guided reciprocal learning for video-based person re-identification. In CVPR, pages 13334–13343, 2021. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 384, + 554, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 384, + 554, + 425 + ], + "spans": [ + { + "bbox": [ + 317, + 384, + 554, + 425 + ], + "type": "text", + "content": "[24] Xuehu Liu, Chenyang Yu, Pingping Zhang, and Huchuan Lu. Deeply coupled convolution-transformer with spatial-temporal complementary learning for video-based person re-identification. TNNLS, 2023." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 427, + 554, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 427, + 554, + 469 + ], + "spans": [ + { + "bbox": [ + 317, + 427, + 554, + 469 + ], + "type": "text", + "content": "[25] Xuehu Liu, Pingping Zhang, Chenyang Yu, Xuesheng Qian, Xiaoyun Yang, and Huchuan Lu. A video is worth three views: Trigeminal transformers for video-based person re-identification. TITS, 2024. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 472, + 554, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 472, + 554, + 514 + ], + "spans": [ + { + "bbox": [ + 317, + 472, + 554, + 514 + ], + "type": "text", + "content": "[26] Hu Lu, Xuezhang Zou, and Pingping Zhang. Learning progressive modality-shared transformers for effective visible-infrared person re-identification. In AAAI, pages 1835–1843, 2023. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 516, + 554, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 516, + 554, + 568 + ], + "spans": [ + { + "bbox": [ + 317, + 516, + 554, + 568 + ], + "type": "text", + "content": "[27] Ke Niu, Haiyang Yu, Mengyang Zhao, Teng Fu, Siyang Yi, Wei Lu, Bin Li, Xuelin Qian, and Xiangyang Xue. Chatreid: Open-ended interactive person retrieval via hierarchical progressive tuning for vision language models. arXiv preprint arXiv:2502.19958, 2025. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 571, + 554, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 571, + 554, + 612 + ], + "spans": [ + { + "bbox": [ + 317, + 571, + 554, + 612 + ], + "type": "text", + "content": "[28] Wenjie Pan, Linhan Huang, Jianbao Liang, Lan Hong, and Jianqing Zhu. Progressively hybrid transformer for multimodal vehicle re-identification. Sensors, 23(9):4206, 2023. 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 615, + 554, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 615, + 554, + 668 + ], + "spans": [ + { + "bbox": [ + 317, + 615, + 554, + 668 + ], + "type": "text", + "content": "[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 1, 2, 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 670, + 554, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 670, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 670, + 554, + 712 + ], + "type": "text", + "content": "[30] Jiangming Shi, Xiangbo Yin, Yachao Zhang, Yuan Xie, Yanyun Qu, et al. Learning commonality, divergence and variety for unsupervised visible-infrared person re-identification. NeurIPS, 37:99715–99734, 2024. 1" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "29709" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "type": "text", + "content": "[31] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, pages 2818-2826, 2016. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "type": "text", + "content": "[32] Wentan Tan, Changxing Ding, Jiayu Jiang, Fei Wang, Yibing Zhan, and Dapeng Tao. Harnessing the power of mllms for transferable text-to-image person reid. In CVPR, pages 17127-17137, 2024. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 185 + ], + "type": "text", + "content": "[33] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. JMLR, 9(11), 2008. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 294, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 218 + ], + "type": "text", + "content": "[34] Qizao Wang, Bin Li, and Xiangyang Xue. When large vision-language models meet person re-identification. arXiv preprint arXiv:2411.18111, 2024. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 220, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 262 + ], + "type": "text", + "content": "[35] Yingquan Wang, Pingping Zhang, Shang Gao, Xia Geng, Hu Lu, and Dong Wang. Pyramid spatial-temporal aggregation for video-based person re-identification. In ICCV, pages 12026–12035, 2021. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 264, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 319 + ], + "type": "text", + "content": "[36] Yuhao Wang, Xuehu Liu, Tianyu Yan, Yang Liu, Aihua Zheng, Pingping Zhang, and Huchuan Lu. Mambapro: Multi-modal object re-identification with mamba aggregation and synergistic prompt. arXiv preprint arXiv:2412.10707, 2024.5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 320, + 294, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 294, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 294, + 365 + ], + "type": "text", + "content": "[37] Yuhao Wang, Xuehu Liu, Pingping Zhang, Hu Lu, Zhengzheng Tu, and Huchuan Lu. Top-reid: Multi-spectral object re-identification with token permutation. In AAAI, pages 5758-5766, 2024. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 365, + 294, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 408 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 408 + ], + "type": "text", + "content": "[38] Yuhao Wang, Yang Liu, Aihua Zheng, and Pingping Zhang. Decoupled feature-based mixture of experts for multi-modal object re-identification. arXiv preprint arXiv:2412.10650, 2024. 5, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 410, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 410, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 410, + 294, + 453 + ], + "type": "text", + "content": "[39] Yingquan Wang, Pingping Zhang, Dong Wang, and Hutchuan Lu. Other tokens matter: Exploring global and local features of vision transformers for object re-identification. CVIU, 244:104030, 2024. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 455, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 455, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 455, + 294, + 498 + ], + "type": "text", + "content": "[40] Yuhao Wang, Pingping Zhang, Xuehu Liu, Zhengzheng Tu, and Huchuan Lu. Unity is strength: Unifying convolutional and transformeral features for better person re-identification. TITS, 2025. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 500, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 500, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 500, + 294, + 544 + ], + "type": "text", + "content": "[41] Zi Wang, Chenglong Li, Aihua Zheng, Ran He, and Jin Tang. Interact, embed, and enlarge: Boosting modality-specific representations for multi-modal person re-identification. In AAAI, pages 2633–2641, 2022. 2, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 544, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 544, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 294, + 578 + ], + "type": "text", + "content": "[42] Zi Wang, Huaibo Huang, Aihua Zheng, and Ran He. Heterogeneous test-time training for multi-modal person re-identification. In AAAI, pages 5850–5858, 2024. 2, 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 579, + 294, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 294, + 622 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 294, + 622 + ], + "type": "text", + "content": "[43] Di Wu, Zhihui Liu, Zihan Chen, Shenglong Gan, Kaiwen Tan, Qin Wan, and Yaonan Wang. Lrrm: Low rank multiscale multi-modal fusion for person re-identification based on rgb-ni-ti. ESWA, 263:125716, 2025. 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 624, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 656 + ], + "type": "text", + "content": "[44] Zhuofan Xia, Xuran Pan, Shiji Song, Li Erran Li, and Gao Huang. Vision transformer with deformable attention. In CVPR, pages 4794-4803, 2022. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 658, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 689 + ], + "type": "text", + "content": "[45] Bin Yang, Jun Chen, and Mang Ye. Shallow-deep collaborative learning for unsupervised visible-infrared person re-identification. In CVPR, pages 16870–16879, 2024. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[46] Haoli Yin, Jiayao Li, Eva Schiller, Luke McDermott, and Daniel Cummings. Graft: Gradual fusion trans" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 710 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "former for multimodal re-identification. arXiv preprint arXiv:2310.16856, 2023. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "text", + "content": "[47] Chenyang Yu, Xuehu Liu, Yingquan Wang, Pingping Zhang, and Huchuan Lu. Tf-clip: Learning text-free clip for video-based person re-identification. In AAAI, pages 6764–6772, 2024. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 141, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 194 + ], + "type": "text", + "content": "[48] Jiazuo Yu, Haomiao Xiong, Lu Zhang, Haiwen Diao, Yunzhi Zhuge, Lanqing Hong, Dong Wang, Huchuan Lu, You He, and Long Chen. Llms can evolve continually on modality for x-modal reasoning. arXiv preprint arXiv:2410.20178, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 197, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 239 + ], + "type": "text", + "content": "[49] Jiazuo Yu, Yunzhi Zhuge, Lu Zhang, Ping Hu, Dong Wang, Huchuan Lu, and You He. Boosting continual learning of vision-language models via mixture-of-experts adapters. In CVPR, pages 23219-23230, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 241, + 553, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 284 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 284 + ], + "type": "text", + "content": "[50] Zhi Yu, Zhiyong Huang, Mingyang Hou, Jiaming Pei, Yan Yan, Yushi Liu, and Daming Sun. Representation selective coupling via token sparsification for multi-spectral object re-identification. TCSVT, 2024. 5, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 286, + 553, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 340 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 340 + ], + "type": "text", + "content": "[51] Zhi Yu, Zhiyong Huang, Mingyang Hou, Yan Yan, and Yushi Liu. Wtsf-reid: Depth-driven window-oriented token selection and fusion for multi-modality vehicle re-identification with knowledge consistency constraint. ESWA, page 126921, 2025. 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 342, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 553, + 374 + ], + "type": "text", + "content": "[52] Yajing Zhai, Yawen Zeng, Da Cao, and Shaofei Lu. Trireid: Towards multi-modal person re-identification via descriptive fusion model. In ICMR, pages 63–71, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 376, + 553, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 408 + ], + "type": "text", + "content": "[53] Guowen Zhang, Pingping Zhang, Jinqing Qi, and Huchuan Lu. Hat: Hierarchical aggregation transformers for person re-identification. In ACM MM, pages 516-525, 2021. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 410, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 410, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 410, + 553, + 453 + ], + "type": "text", + "content": "[54] Pingping Zhang, Yuhao Wang, Yang Liu, Zhengzheng Tu, and Huchuan Lu. Magic tokens: Select diverse tokens for multi-modal object re-identification. In CVPR, pages 17117–17126, 2024. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 455, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 455, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 455, + 553, + 498 + ], + "type": "text", + "content": "[55] Qingyang Zhang, Yake Wei, Zongbo Han, Huazhu Fu, Xi Peng, Cheng Deng, Qinghua Hu, Cai Xu, Jie Wen, Di Hu, et al. Multimodal fusion on low-quality data: A comprehensive survey. arXiv preprint arXiv:2404.18947, 2024. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 500, + 553, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 500, + 553, + 532 + ], + "spans": [ + { + "bbox": [ + 316, + 500, + 553, + 532 + ], + "type": "text", + "content": "[56] Aihua Zheng, Zi Wang, Zihan Chen, Chenglong Li, and Jin Tang. Robust multi-modality person re-identification. In AAAI, pages 3529–3537, 2021. 2, 5, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 533, + 553, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 533, + 553, + 577 + ], + "spans": [ + { + "bbox": [ + 316, + 533, + 553, + 577 + ], + "type": "text", + "content": "[57] Aihua Zheng, Ziling He, Zi Wang, Chenglong Li, and Jin Tang. Dynamic enhancement network for partial multi-modality person re-identification. arXiv preprint arXiv:2305.15762, 2023. 2, 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 578, + 553, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 578, + 553, + 632 + ], + "spans": [ + { + "bbox": [ + 316, + 578, + 553, + 632 + ], + "type": "text", + "content": "[58] Aihua Zheng, Xianpeng Zhu, Zhiqi Ma, Chenglong Li, Jin Tang, and Jixin Ma. Cross-directional consistency network with adaptive layer normalization for multi-spectral vehicle re-identification and a high-quality benchmark. Information Fusion, 100:101901, 2023. 2, 5, 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 635, + 553, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 677 + ], + "type": "text", + "content": "[59] Aihua Zheng, Zhiqi Ma, Yongqi Sun, Zi Wang, Chenglong Li, and Jin Tang. Flare-aware cross-modal enhancement network for multi-spectral vehicle re-identification. Information Fusion, 116:102800, 2025. 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 679, + 553, + 710 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 679, + 553, + 710 + ], + "spans": [ + { + "bbox": [ + 316, + 679, + 553, + 710 + ], + "type": "text", + "content": "[60] Zhun Zhong, Liang Zheng, Guoliang Kang, Shaozi Li, and Yi Yang. Random erasing data augmentation. In AAAI, pages 13001-13008, 2020. 6" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "29710" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/IDOL_ Instant Photorealistic 3D Human Creation from a Single Image/c28e8ffb-3109-4ced-b9e1-553dbeace189_content_list.json b/2025/IDOL_ Instant Photorealistic 3D Human Creation from a Single Image/c28e8ffb-3109-4ced-b9e1-553dbeace189_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..345f0bd2c35737c80af057b071c193b93031e20e --- /dev/null +++ b/2025/IDOL_ Instant Photorealistic 3D Human Creation from a Single Image/c28e8ffb-3109-4ced-b9e1-553dbeace189_content_list.json @@ -0,0 +1,1487 @@ +[ + { + "type": "text", + "text": "IDOL: Instant Photorealistic 3D Human Creation from a Single Image", + "text_level": 1, + "bbox": [ + 142, + 130, + 854, + 152 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6f12e6f0b256d25f38b2c0b2d88f71f6531892c93090f89911bd4113a582a57a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 133, + 178, + 864, + 267 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ab4b622cbe0ce6d245a6a7add80b6aa698da8bd180fcf5f33ee7c8bcf2b3c61c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 148, + 268, + 397, + 420 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/319ff23135b5cb69b47d31d4e1377b0d9cd97bc2182700ac5a93d7d1ea3c1e24.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 401, + 268, + 841, + 420 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a5afe8c7f64f696cf13ea57d102b7db378bc0ccb311228de9aacaee30d73431f.jpg", + "image_caption": [ + "Figure 1. This work introduces (a) IDOL, a feed-forward, single-image human reconstruction framework that is fast, high-fidelity, and generalizable; (b) Utilizing the proposed Large Generated Human Multi-View Dataset consisting of $100K$ multi-view subjects, our method exhibits exceptional generalizability in handling diverse human shapes, cross-domain data, severe viewpoints, and occlusions; (c) With a uniform structured representation, the avatars can be directly animatable and easily editable." + ], + "image_footnote": [], + "bbox": [ + 148, + 422, + 397, + 573 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8c58ee58915cc674be7d147ab614c31b5a845056ff29f23e7478b2d9ebe13f1a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 401, + 422, + 841, + 573 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 642, + 326, + 657 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Creating a high-fidelity, animatable 3D full-body avatar from a single image is a challenging task due to the diverse appearance and poses of humans and the limited availability of high-quality training data. To achieve fast and high-quality human reconstruction, this work rethinks the task from the perspectives of dataset, model, and representation. First, we introduce a large-scale HUmancentric GEnerated dataset, HuGe100K, consisting of 100K diverse, photorealistic sets of human images. Each set contains 24-view frames in specific human poses, generated using a pose-controllable image-to-multi-view model.", + "bbox": [ + 88, + 674, + 485, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Next, leveraging the diversity in views, poses, and appearances within HuGe100K, we develop a scalable feedforward transformer model to predict a 3D human Gaussian representation in a uniform space from a given human image. This model is trained to disentangle human pose, body shape, clothing geometry, and texture. The estimated Gaussians can be animated without post-processing. We conduct comprehensive experiments to validate the effectiveness of the proposed dataset and method. Our model demonstrates the ability to efficiently reconstruct photorealistic humans at 1K resolution from a single input image using a single GPU instantly. Additionally, it seamlessly supports various applications, as well as shape and texture editing tasks.", + "bbox": [ + 511, + 643, + 908, + 840 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Work done during the internship at Tencent by Yiyu Zhuang, Jiaxi Lv, and Wenhao.", + "bbox": [ + 89, + 849, + 482, + 875 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contributions", + "bbox": [ + 116, + 875, + 230, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Corresponding authors", + "bbox": [ + 116, + 887, + 243, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "26308", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 89, + 222, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Reconstructing 3D clothed avatars from a single image is crucial in user-friendly virtual reality, gaming, and 3D content creation. This task involves mapping 2D images to 3D models, a highly ill-posed and challenging problem due to the complexity of clothing and the diversity of human poses. Learning-based methods trained on publicly available 3D human models [17, 25, 26, 63, 79] have improved reconstruction quality, but their generalization and quality remain limited [52, 53, 64, 65, 72, 78, 80]. Techniques that integrate parametric body estimation [31, 64, 80], loop optimization [64, 65, 78], and diffusion model priors [1, 37, 60] enhance performance but often result in slower and fragile reconstructions. Recently, large generic image-to-3D reconstruction and generation models [22, 36, 57, 58, 61, 66], leveraging large-scale datasets [10, 11, 43] or pre-trained diffusion models [2, 50, 51], boost capabilities but still struggle with real-life human reconstruction due to the scarcity of photorealistic 3D human data.", + "bbox": [ + 89, + 114, + 485, + 385 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Back to the first principle, we rethink existing data, models, and representations. First, current data acquisition is limited by the photographing or scanning of individual subjects, making them inherently unscalable. For instance, the largest publicly available human dataset, MVHumanNet [63], contains only thousands of subjects and clothing variations, far from the dataset scale required to achieve robust model generalization across diverse input images. Second, in the realm of animatable reconstruction models, existing approaches predominantly rely on either reconstruct- and-rig methods that necessitate manual post-processing or entangle human parametric models (e.g., SMPL(-X) [42, 46]) for optimization. However, disentangling human pose and body shape, clothed geometry, and texture could simplify the learning process, avoid error accumulation in parameter estimation, and improve efficiency. Furthermore, when focusing solely on human reconstruction, the incorporation of multi-view image generation or refinement models warrants careful consideration [48, 60, 66]. While these models can introduce finer details, they may also lead to inconsistencies across different views. Third, the community of 3D vision develops various representations tailored to application-specific requirements and technological advancements [33, 44]. For human-centric applications, ideal representations should be well-structured and expressive to facilitate easy rigging and editing [28, 39], as well as capable of high-resolution and fast rendering [33] to enhance both functionality and realism.", + "bbox": [ + 89, + 387, + 483, + 809 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we introduce a scalable pipeline for training a simple yet efficient feed-forward model for instant photorealistic human reconstruction. We present HuGe100K, a large-scale dataset comprising over 2.4M high-resolution $(896\\times 640)$ multi-view images of 100K diverse subjects. To facilitate comprehensive 3D human reconstruction, we de", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "velop a scalable data creation pipeline that integrates synthetic and real-world data, ensuring a wide range of attributes such as age, shape, clothing, race, and gender. Building upon this dataset, we introduce a novel feedforward transformer model IDOL that efficiently predicts 3D human avatars with photorealistic textures and accurate geometry. Our approach leverages a pretrained encoder [34] and a transformer-based backbone for feature extraction and fusion, enabling instant reconstruction (within 1 second on an A100 GPU) without relying on generative models [78] for refinement. By integrating a uniform representation for the 3D human, our model achieves enhanced texture completion and ensures that the reconstructed humans are naturally animatable. Extensive evaluations demonstrate that our method excels in diverse and challenging scenarios, offering superior consistency and quality compared to existing techniques. Additionally, the efficient representation of our model enables seamless applications in animation, editing, and other downstream tasks.", + "bbox": [ + 511, + 90, + 906, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows:", + "bbox": [ + 531, + 378, + 857, + 393 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We rethink single-view 3D human reconstruction from the perspectives of data, model, and representation. We introduce a scalable pipeline for training a simple yet efficient feed-forward model, named IDOL, for instant photorealistic human reconstruction.", + "- We develop a data generation pipeline and present HuGe100K, a large-scale multi-view human dataset featuring diverse attributes, high-fidelity, high-resolution appearances, and a well-aligned SMPL-X model.", + "- Our comprehensive study demonstrates that leveraging large-scale generated training data significantly enhances model performance and generalizability, paving the way for further scaling up 3D human reconstruction models." + ], + "bbox": [ + 513, + 398, + 906, + 594 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 611, + 653, + 627 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent advancements in 3D avatar reconstruction [12-14, 20, 69, 82] and representation learning [15, 62, 83, 84] have greatly improved human modeling. Most relevant to our work are recent efforts on 3D human datasets and single-image human reconstruction.", + "bbox": [ + 511, + 637, + 906, + 713 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. 3D Human Datasets", + "text_level": 1, + "bbox": [ + 511, + 724, + 702, + 739 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "High-precision 3D human models typically rely on scanning or multi-view camera acquisition systems. Scan-based datasets, such as THuman2.0 [68], Twindom [26], and 2K2K [17], provide high-fidelity 3D geometries but are limited by a small number of subjects, simple standing poses, and non-human artifacts (e.g., 2K2K). In contrast, multiview acquisition systems [3, 7, 8, 21, 32, 38, 40, 49, 63, 71, 85] facilitate the collection of larger datasets with more flexible actions. However, these datasets often face challenges such as biased indoor lighting, fixed viewpoints, and limited", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "26309", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "scale. Large-scale synthetic and real datasets for generic objects, including Objaverse and MVImgNet [10, 11, 70], address open-world reconstruction but lack specificity for human models. To overcome these limitations, the proposed HuGe100K significantly scales up dataset size, increasing subject diversity by over 100 times compared to previous datasets. It emphasizes diversity in pose, shape, viewpoint, and clothing, paving the way for large-scale model training. We compare existing datasets with ours in Tab. 1.", + "bbox": [ + 89, + 90, + 480, + 226 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/a32e0df455487a09ef1f8601d74b80450735b788db95a8b2941990424e74c4d2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TypeDataset#FramesIDsSMPL(-X)
3D ScansTHuman [79]-200✓
THuman2.1 [68]-2500✓
2K2K [17]-2050✗
X-Avatar [55]35.5K20✓
Multi-view ImagesZJU-MoCap [49]180K10✓
Neural Actor [40]250K8✓
HUMBI [71]26M772✓
AIST++ [38]10.1M30✓
HuMMan [3]60M1000✓
GeneBody [7]2.95M50✓
ActorsHQ [32]40K8✗
DNA-Rendering [8]67.5M500✓
MVHumanNet [63]645.1M4500✓
OursHuGe100K>2.4M100K✓
", + "bbox": [ + 106, + 239, + 468, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1. Comparisons of related datasets. HuGe100K is a large-scale generated multi-view human dataset with $100K$ diverse high-fidelity humans.", + "bbox": [ + 89, + 460, + 482, + 503 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Single-Image Human Reconstruction", + "text_level": 1, + "bbox": [ + 89, + 530, + 410, + 546 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For clothed 3D reconstruction methods, implicit representation methods such as PIFU [52], PIFU-HD [53], ARCH [19, 31], and PaMIR [80] have been widely adopted. To enhance reconstruction quality and generalizability, loop optimization techniques integrate implicit representations with explicit or hybrid human priors for improved robustness [1, 64, 65, 76]. Similarly, GTA [77] and SIFU [78] entangle SMPL priors and side-view conditioned features to enhance feature representation. However, these approaches rely heavily on the accuracy of SMPL estimates and are computationally expensive, typically requiring several minutes for inference. Moreover, the resulting 3D representations are not drivable. Recent advancements focus on recovering animatable 3D humans from single images [9, 29, 30]. Some methods benefit from the pre-trained diffusion models or large reconstruction models [48, 60]. Nevertheless, all methods are constrained by the limitations of training datasets, resulting in suboptimal texture detail and limited generalizability. The latest work, E3Gen [74], combines UV maps and Gaussian splatting [33] to directly generate Gaussian attribute maps in UV space from images, which does not support arbitrary image inputs.", + "bbox": [ + 89, + 551, + 482, + 885 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To generate consistent human images or videos given a", + "bbox": [ + 109, + 886, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "single image, recent diffusion video models animate a static image with pose-controllable video conditions[5, 5, 27, 59, 59, 67, 81]. Specifically, Champ [81] incorporate several rendered maps obtained from SMPL sequences, alongside skeleton-based motion guidance, to enrich the conditions to the latent diffusion model. However, they all fail to generate a 360-degree video and animate precise expressive SMPL-X-based humans due to insufficient training data.", + "bbox": [ + 511, + 90, + 903, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Dataset Creation", + "text_level": 1, + "bbox": [ + 511, + 224, + 679, + 239 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Drawing inspiration from the latest advancements in generalizable large models, our core insight is to develop a large-scale reconstruction model with exceptional generalization capabilities. The crux of this endeavor lies in creating a comprehensive and high-quality digital human dataset. In this section, we introduce the creation of a large-scale human-centric generated dataset comprising over 2.4M high-resolution $(896 \\times 640)$ multi-view images from more than 100K diverse subjects.", + "bbox": [ + 511, + 250, + 903, + 383 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Fig. 2, the data generation pipeline consists of two stages. Firstly, diverse text prompts are generated by large language models incorporating various human-centric attributes, and high-quality images are synthesized using text-to-image generation models (Sec. 3.1). Secondly, we train a multi-view video generation model, MVChamp, conditioned on rendered full-body motions. With this, a large-scale multi-view images dataset is established by feeding both the synthesized and captured images (Sec. 3.2). Lastly, the statistics and characteristics of the HuGe100K are demonstrated in Sec. 3.3.", + "bbox": [ + 511, + 385, + 905, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Image Collection and Generation", + "text_level": 1, + "bbox": [ + 511, + 559, + 805, + 574 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given the limited amount of existing datasets [3, 49, 63, 68] and legal concerns regarding portrait and copyright issues, we propose leveraging an image generation model to construct a large-scale, diverse, and high-fidelity dataset of full-body human images. Using the latest text-to-image model Flux [23], we design descriptive prompts based on the required attributes for human figures, ensuring diversity across area, clothing, body shape, age, and gender shown in Fig. 2(i). To avoid long-tail attribute distributions and ensure uniform coverage of traits, we apply uniform sampling in attribute selection for the image generation process. This approach generated over 100K images in total. However, due to issues of visual non-compliance and high similarity among some images, we conducted manual filtering and retained a final set of 90K high-quality images. Additionally, we combine 90K synthetic images and 10K real-life images from DeepFashion [41].", + "bbox": [ + 511, + 582, + 903, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Multi-view Image Animation and Generation", + "text_level": 1, + "bbox": [ + 511, + 847, + 897, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image-based animating models [5, 27, 67, 75, 81] struggle to achieve generation of $360^{\\circ}$ -consistent and diverse hu", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "26310", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b63c0190a8160edd8b52664efcf9fd9d1a08ff4ad460f476108674f28e460969.jpg", + "image_caption": [ + "Figure 2. Pipeline for constructing our HuGe100K. Diverse attribute combinations from GPT-4 templates create text prompts, generating synthetic images via FLUX, combined with real images from DeepFashion. SMPL-X fitting produces multi-view pose sequences with 360-degree rotations and diverse animatable motions. MVChamp then converts these sequences into multi-view images, ensuring 3D consistency in the dataset." + ], + "image_footnote": [], + "bbox": [ + 89, + 85, + 903, + 325 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "man videos conditioned on pose sequences (e.g., 2D poses and 3D SMPL(-X) conditions). To address this limitation, we re-train a state-of-the-art video generative model, Champ [81], to obtain a multi-view consistent generative model, MVChamp. Firstly, We collect a curated dataset of over 100K single-person videos with various motions (e.g., dancing), including approximately 20K videos involving the action of turning around, to fine-tune the model and enhance its generalizability for diverse inputs and motions. Secondly, to further improve 3D consistency[16], we utilize scanned models from THuman 2.1 to generate uniformly distributed views via Blender. From these, we select a subset of 24 views that are evenly spaced around the full $360^{\\circ}$ rotation to fine-tune MVChamp's temporal layers using diffusion loss. Lastly, to support high-quality whole-body animation (i.e., body and hands), we introduce precise SMPL-X estimation using NLF [54] for body shape and pose estimation. Additionally, we employ HaMeR [47] for hand estimation, rendering the corresponding hand template into depth maps as additional control signals.", + "bbox": [ + 88, + 410, + 485, + 712 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After these training processes, MVChamp generates multi-view images conditioned on SMPL-X parametric model. The pose condition is set to an \"A-Pose\" $80\\%$ of the time and a random pose from dance videos (via SMPLer-X [4]) $20\\%$ of the time, balancing pose variability. To address discrepancies between the first and last frames, we propose a Temporal Shift Denoising Strategy: during the denoising steps, we shift latent inputs and pose signals temporally, moving the last frame to the first. This improves continuity between frames without additional inference costs. Finally, FaceFusion [24] is utilized to enhance facial details.", + "bbox": [ + 89, + 734, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2a67f02d46c67091a06a7612b4d661065ab4fc4612a13f4de0af8ea336f3701a.jpg", + "image_caption": [ + "Figure 3. A paired example from the proposed HuGe100K Dataset. For each reference image, we generate a set of multiview images using an estimated shape and a specific pose. The figure shows the pose is well-aligned." + ], + "image_footnote": [], + "bbox": [ + 535, + 406, + 885, + 460 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Data Statistics and Characteristics", + "text_level": 1, + "bbox": [ + 511, + 556, + 815, + 571 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Tab. 1 compares our dataset HuGe100K to other 3D human datasets regarding scale, diversity, and consistency. With over 100K human identities and 20K poses, HuGe100K offers balanced diversity across dimensions such as area, clothing, body shape, age, and gender (see Fig. 1). Each sample includes a reference image, SMPL-X estimates, 24 uniformly sampled multi-view images, camera parameters, and SMPL-X data (see Fig. 3).", + "bbox": [ + 511, + 580, + 906, + 702 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Large Human Reconstruction Model", + "text_level": 1, + "bbox": [ + 511, + 717, + 846, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we present the large-scale human reconstruction model, named IDOL. The overview pipeline is shown in Fig. 4. In Sec. 4.1, we describe the animatable human representation. Sec. 4.2 details the network architecture of IDOL, while Sec. 4.3 explains how we trained the model in an end-to-end manner using the multi-view image dataset.", + "bbox": [ + 511, + 743, + 906, + 835 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Animatable Human Representation", + "text_level": 1, + "bbox": [ + 511, + 845, + 823, + 863 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Similarly to previous work [28, 35, 74], IDOL leverages 3D Gaussian Splatting [33] in conjunction with SMPL-X", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "26311", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5316acc26691809aea53c9dfca58977e8b50f661e586da5b397cbf6d510c720d.jpg", + "image_caption": [ + "Figure 4. The architecture of IDOL, a full-differentiable transformer-based framework for reconstructing animatable 3D human from a single image. The model integrates a high-resolution $(1024\\times 1024)$ encoder [34] and fuses image tokens with learnable UV tokens through the UV-Alignment Transformer. A UV Decoder predicts Gaussian attribute maps as intermediate representations, capturing the human's geometry and appearance in a structured 2D UV space defined by the SMPL-X model. These maps, in conjunction with the SMPL-X model, represent a 3D human avatar in a canonical space, which can be animated using linear blend skinning (LBS). The model is optimized using multi-view images with diverse poses and identities, learning to disentangle pose, appearance, and shape." + ], + "image_footnote": [], + "bbox": [ + 114, + 87, + 885, + 232 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "for 3D human representation, aiming to address the challenges of real-time rendering and accurate animation of human avatars. Specifically, each Gaussian primitive $\\mathcal{G}_k$ is characterized by: $\\mathcal{G}_k = \\{\\mu_k,\\alpha_k,\\mathbf{r}_k,\\mathbf{s}_k,\\mathbf{c}_k\\}$ , where $\\mu_{k}$ is the 3D position of the Gaussian, $\\alpha_{k}$ is opacity, $\\mathbf{r}_k$ is the rotation, $\\mathbf{s}_k$ is the scale, and $\\mathbf{c}_k$ is the color.", + "bbox": [ + 88, + 349, + 480, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3D Gaussian Human. Directly predicting all 3D Gaussian primitives is computationally intensive. Instead, IDOL leverages the predefined 2D UV space of the SMPL-X model to transform the 3D representation task into a more manageable 2D problem. Initially, IDOL predicts Gaussian attribute maps that encode the offset values $\\{\\delta_{\\mu_k}, \\delta_{\\mathbf{r}_k}, \\delta \\mathbf{s}_k\\}$ , as well as the color $\\mathbf{c}_k$ and opacity $\\alpha_k$ for each Gaussian primitive $\\mathcal{G}_k$ . Similarly to $E^3 Gen$ [74], the position $\\mu_k$ , scale $\\mathbf{s}_k$ , and rotation $\\mathbf{r}_k$ of each Gaussian primitive $\\mathcal{G}_k$ is modeled relative to its SMPL-X vertex as follows: $\\mu_k = \\hat{\\mu}_k + \\delta_{\\mu_k}$ , $\\mathbf{s}_k = \\hat{\\mathbf{s}}_k \\cdot \\delta_{\\mathbf{s}_k}$ , $\\mathbf{r}_k = \\hat{\\mathbf{r}}_k \\cdot \\delta_{\\mathbf{r}_k}$ , where $\\hat{\\mu}_k, \\hat{\\mathbf{s}}_k$ , and $\\hat{\\mathbf{r}}_k$ are the initial values based on the SMPL-X model. The color $\\mathbf{c}_k$ is assigned using RGB values, and the opacity $\\alpha_k$ is set to 1, indicating full opacity. This unified UV space significantly reduces computational complexity and fully exploits the geometric and semantic priors provided by the SMPL-X model. By modeling Gaussian primitives relative to SMPL-X vertices, IDOL ensures semantic consistency across corresponding body parts of diverse avatars, thereby enhancing the model's generalization capability across various reference images.", + "bbox": [ + 91, + 459, + 483, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Animation and Rendering. Given a target pose, we calculate the transformation of each human joint using predefined kinematic relationships. The transformation of each Gaussian primitive is performed using a forward skinning technique based on LBS. Specifically, the position of each Gaussian primitive $\\mathcal{G}_k$ is transformed as follows: $\\mu_k' = \\sum_{i=1}^{n_b} w_i \\mathbf{B}_i \\mu_k$ . Additionally, the rotation matrix $\\mathbf{R}_k$ is up-", + "bbox": [ + 89, + 794, + 482, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "dated by: $\\mathbf{R}_k^{\\prime} = \\mathbf{T}_k^{1:3,1:3}\\mathbf{R}_k$ , where $\\mathbf{T}_k = \\sum_{i=1}^{n_b} w_i \\mathbf{B}_i$ , and $\\mathbf{R}_k$ is the rotation matrix representation of the rotation angle $\\mathbf{r}_k$ . Here, $n_b$ is the number of joints, $\\mathbf{B}_i$ is the transformation matrix for each joint, and $w_i$ represents the skinning weights, indicating the influence of each joint's motion on the Gaussian primitive's position $\\mu_k$ .", + "bbox": [ + 511, + 349, + 906, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To estimate the skinning weights for each $\\mathcal{G}_k$ , we first compute a body-part skinning field using a pre-defined low-resolution volume [6]. For smaller regions like the hands and face, which undergo minimal topological variation, weights are interpolated from the SMPL-X template via barycentric coordinates. This strategy enables modeling large topology changes across identities (e.g., from clothing) while stabilizing convergence in less variable areas such as fingers and facial regions.", + "bbox": [ + 511, + 441, + 908, + 578 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Network Structure", + "text_level": 1, + "bbox": [ + 511, + 587, + 694, + 602 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As illustrated in Fig. 4, IDOL is a full-differentiable framework for reconstructing animatable 3D human.", + "bbox": [ + 511, + 609, + 906, + 641 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "High-resolution Image Encoder. Higher resolution of the input image results in a high-quality reconstruction. However, previous works suffer from low-resolution ViT-based encoders, such as DINOv2 [45], which support the $448 \\times 448$ resolution. To fully leverage the resolution of the HuGe100K dataset, we further adopt a high-resolution human foundation model, Sapiens [34], to encode the $1024 \\times 1024$ resolution images into patch-wise feature tokens, formulated as: $\\mathbf{F} = \\{\\mathbf{f}_i\\}_{i=1}^n \\in \\mathbb{R}^{d_E}$ , where $i$ denotes the $i$ -th image patch, and $d_E$ is the channel length of each token. The Sapiens model is pretrained on 300 million in-the-wild human images using the Masked Autoencoder (MAE) framework [18], enabling it to excel in preserving fine-grained details and capturing diverse human poses and appearances, making it highly effective for high-resolution human image feature extraction.", + "bbox": [ + 511, + 657, + 908, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "26312", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "UV-Alignment Transformer. To map irregular and diverse reference images onto regular UV feature maps, a UV-Alignment Transformer is employed to align learnable spatial-positional UV tokens $\\mathbf{Q}^0$ with reference image features $\\mathbf{F}$ . Specifically, the UV-Alignment Transformer concatenates $\\mathbf{F}$ from Spaiens with $\\mathbf{Q}^0$ , and aggregates and refines features through $D$ transformer blocks, producing an enhanced representation $\\mathbf{Q}^D$ . Each transformer block utilizes self-attention mechanisms, enabling the model to capture complex relationships among the input tokens and impute missing parts using correlated tokens.", + "bbox": [ + 89, + 90, + 480, + 257 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "UV Decoder. The UV tokens $\\mathbf{Q}^D$ are reshaped and decoded into Gaussian attribute maps, capturing the human's geometry and appearance within the structured 2D UV space as shown in Sec. 4.1. To preserve and enhance fine details in the decoded Gaussian attribute maps, a Convolutional Neural Network is employed for spatial up-sampling.", + "bbox": [ + 89, + 273, + 480, + 367 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Training Objectives", + "text_level": 1, + "bbox": [ + 89, + 375, + 277, + 391 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "IDOL reconstructs 3D human by predicting Gaussian attribute UV maps in a single forward pass, offering significant advantages in inference speed and enabling end-to-end training with multi-view images. For each sample, we select a front view as the reference image $\\mathbf{I}_{\\mathrm{ref}}$ along with a random set of generated multi-view images $\\{\\mathbf{I}_{\\mathrm{gt},i}\\}_{i = 1}^{N}$ and their corresponding SMPL-X parameters and camera for supervision. It takes $\\mathbf{I}_{\\mathrm{ref}}$ as input to generate the 3D human, using differentiable rendering to produce the multi-view images $\\{\\mathbf{I}_{\\mathrm{pred},i}\\}_{i = 1}^{N}$ . The loss function is defined as follows:", + "bbox": [ + 89, + 398, + 482, + 551 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\sum_ {i = 1} ^ {N} \\left(\\left\\| \\mathbf {I} _ {\\mathrm {g t}, i} - \\mathbf {I} _ {\\text {p r e d}, i} \\right\\| ^ {2} + \\lambda L _ {v g g} \\left(\\mathbf {I} _ {\\mathrm {g t}, i}, \\mathbf {I} _ {\\text {p r e d}, i}\\right)\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 561, + 482, + 603 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\lambda$ controls the balance between the mean square error loss and the perceptual loss $L_{vgg}$ .", + "bbox": [ + 89, + 612, + 482, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 656, + 223, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experiments on Dataset Creation", + "text_level": 1, + "bbox": [ + 89, + 681, + 382, + 696 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fig. 5 (left) presents the qualitative results of the ablation study on our data generation method. Specifically, Fig. 5 (a) and Fig. 5 (b) demonstrate how hand refinement and face swapping effectively enhance the quality of the generated hands and faces. Fig. 5 (c) highlights the importance of fine-tuning the 3D dataset Thuman 2.1 for improving the 3D consistency of MVChamp, while Fig. 5 (d) illustrates that the Temporal Shift Denoising strategy, which involves cycling the first and last frames, improves the consistency between the first and last frames.", + "bbox": [ + 89, + 703, + 482, + 853 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fig. 5 (right) compares our MVChamp's generation quality against previous human image animation models (e.g., Champ, MimicMotion [75]) and general multi-view", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "video generation model (e.g., SV3D [58]). The issues of these methods are highlighted in the figure. Champ struggles to generate plausible hands and heads in multi-view scenarios. MimicMotion not only fails to produce realistic shoes but also has difficulty preserving the identity, especially for the anime image. SV3D produces low-quality multi-view human generation due to the lack of 3D prior knowledge of the human body. In contrast, our model generates consistent, high-quality results across views.", + "bbox": [ + 511, + 90, + 903, + 227 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Comparison of Reconstruction Model", + "text_level": 1, + "bbox": [ + 511, + 236, + 836, + 252 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. For the encoder, we employ the pre-trained Sapiens-1B model [34] and keep its weights frozen during training. We then define a transformer-based framework with a parameter size of $0.5B$ to perform feature fusion. Additionally, we densify the SMPL-X vertex set by sampling approximately 200,000 vertices to map attributes and represent the 3D human model effectively.", + "bbox": [ + 511, + 258, + 903, + 364 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset and Metrics. We train IDOL on a dataset consisting of generated multi-view images from HuGe100K and rendered images from THuman 2.1 [79]. For THuman 2.1, we render 72 view images at a resolution of $896 \\times 640$ for each scan, with rendering views uniformly distributed. To evaluate performance quantitatively, we reserve the last 50 cases from both HuGe100K and THuman 2.1 as the testing set. We use the most frontal image as the reference image and the remaining images as ground-truth data for evaluation. The ground-truth camera parameters and SMPL-X parameters are provided for all methods. The renderings of the reconstructed model are then compared to their corresponding ground truth images with several metrics, including Mean Squared Error (MSE), Learned Perceptual Image Patch Similarity (LPIPS) [73], Peak Signal-to-Noise Ratio (PSNR). More details on implementation can be found in the supplementary material.", + "bbox": [ + 511, + 383, + 906, + 641 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We compare IDOL with three baseline categories in the Single-Image Human Reconstruction task. The first category comprises methods based on loop optimization or pixel-alignment modules, including GTA [77] and SIFU [78]. These methods focus on refining human reconstruction through iterative optimization or predicting 3D geometry from pixel-aligned features.", + "bbox": [ + 511, + 657, + 903, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The second category encompasses large-scale generic reconstruction networks, represented by LGM [57]. These models are known for their ability to handle large datasets and their scalability, offering advantages in terms of fast reference speeds and large output resolutions.", + "bbox": [ + 511, + 765, + 903, + 840 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The third category involves optimization-based 3D generation methods using score distillation sampling (SDS), exemplified by DreamGaussian [56]. These methods leverage priors from 2D diffusion models to distill 3D objects.", + "bbox": [ + 511, + 840, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "26313", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1d4ebec419a5575533de3618f4338060fb86aa35099fc797dde03cd381261967.jpg", + "image_caption": [ + "(c) THuman Finetune", + "(d) Temporal Shift Denoising Strategy" + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 436, + 349 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a9518cfb5bb51df4681f3d9221f0396de4c46e0bb5e3635d484e4b7207615455.jpg", + "image_caption": [ + "Reference (e) MVChamp (ours)", + "(f) Champ (g) MimicMotion (h) SV3D" + ], + "image_footnote": [], + "bbox": [ + 450, + 87, + 903, + 349 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/16eace9fe2735034876bb1c22d34d06f18ee4bde603478020e53d7fc2f2d2dd7.jpg", + "image_caption": [ + "Figure 5. Qualitative results of our MVChamp ablation study (left) and comparison experiment (right).", + "Figure 6. Comparisons on (a) the upper: novel-view synthesis given a single image, and (b) the lower: our animated results." + ], + "image_footnote": [], + "bbox": [ + 96, + 402, + 906, + 705 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "DreamGaussian accelerates convergence by progressively densifying 3D Gaussians, significantly reducing the reconstruction time. However, it still takes approximately two minutes to reconstruct a single object.", + "bbox": [ + 89, + 756, + 480, + 818 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative comparison. As shown in Tab 2, our method outperforms all baselines in all metrics. We attribute this superior performance to the large-scale dataset HuGe100K and the design of our large-scale reconstruc", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tion model IDOL, which allows for more effective training and improved synthesis of appearance results. Despite this, we note that SIFU and GTA report a lower metric than what we expected in our test settings. While we provide accurate ground-truth SMPL-X parameters and camera settings, the ideal orthographic projection required by the pixel-alignment modules in SIFU and GTA is not well-suited to our perspective projection model, where the camera focus ranges from 35 to 80. This mismatch leads to a", + "bbox": [ + 511, + 756, + 906, + 893 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "26314", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/78908a92893a3b509ddc2549e7dea07fbf477cda146a0b8381c5da1ac2016a30.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodMSE ↓PSNR ↑LPIPS ↓
SIFU [78]0.04214.2041.612
GTA [77]0.04114.2821.629
Ours-w/o HuGe100K0.01719.2251.326
Ours-full0.00821.6731.138
", + "bbox": [ + 127, + 88, + 452, + 167 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2. Evaluation of Comparison and Ablation Experiments.", + "bbox": [ + 99, + 178, + 472, + 191 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "misalignment between the rendered images and the ground truth, adversely affecting their performance metrics. In fact, many real-life photographs are taken with medium to short focal lengths and cannot be approximated using orthographic projection, which is a less noticeable drawback of SIFU and GTA. Additionally, SIFU and GTA, trained with images at a resolution of $512 \\times 512$ , struggle to synthesize detailed textures, particularly in the invisible areas of the reference images. This limitation is primarily due to the lack of comprehensiveness and diversity in their training datasets, which restricts their performance in generating invisible aspects of the images.", + "bbox": [ + 89, + 220, + 483, + 402 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative comparison. We perform a qualitative evaluation on an in-the-wild dataset with methods from [56, 57, 77, 78], with the results presented in Fig. 6. Our evaluation includes a subject in a complex outfit featuring a baseball cap and textured clothing, demonstrating our method's ability to synthesize intricate textures and handle loose outfits across different views. Additional tests include out-of-domain cartoon data and large-angle side views and assessing model adaptability and viewpoint handling. IDOL consistently outperforms the baselines, which struggle with detail reproduction and texture consistency under these various conditions.", + "bbox": [ + 89, + 422, + 483, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Ablation Study on Reconstruction Model", + "text_level": 1, + "bbox": [ + 89, + 614, + 442, + 630 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We assess the impact of proposed components by removing them individually from IDOL. As shown in Fig. 7, replacing the Sapiens [34] encoder with DINO v2 [45] (w/o Sapiens) reduces detail quality, resulting in less realistic textures and folds. Excluding the HuGe100K dataset (w/o HuGe100K) causes significant distortion, including blurred details and color bleeding. The complete model, with all components, produces the most realistic and detailed results. These findings highlight the critical roles of both the Sapiens encoder and the HuGe100K dataset in achieving high-quality avatar generation.", + "bbox": [ + 89, + 638, + 483, + 806 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Downstream Applications", + "text_level": 1, + "bbox": [ + 89, + 816, + 326, + 832 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "IDOL reconstructs an avatar by combining Gaussian attribute maps with the SMPL-X model, enabling users to modify the avatar's appearance by editing UV texture maps and to control body shape by adjusting SMPL-X shape pa", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9d1dabd4ba7c55ce5f9bce4176e87bea8116872ba262e5598207f50a43537c1c.jpg", + "image_caption": [ + "Figure 7. Qualitative Results of Ablation Study of IDOL." + ], + "image_footnote": [], + "bbox": [ + 535, + 87, + 888, + 282 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0b0b6ed44969dd72760d24f32e2fdbbe9ce1ec8c04af4208fba8f74e6049a4bb.jpg", + "image_caption": [ + "Figure 8. Controllable Avatar Editing: (a) texture editing; (b) body shape editing." + ], + "image_footnote": [], + "bbox": [ + 514, + 325, + 906, + 407 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "rameters. Fig. 8(a) illustrates the effect of texture editing on clothing patterns, while Fig. 8(b) demonstrates shape editing to adjust the avatar's body size. This approach provides a high degree of controllability over both the avatar's appearance and body shape.", + "bbox": [ + 511, + 476, + 906, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusions and Limitations", + "text_level": 1, + "bbox": [ + 511, + 570, + 777, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In conclusion, our work has made significant strides in creating an animatable 3D human from a single image. We introduced a scalable pipeline for training a simple yet efficient feed-forward model, incorporating a dataset generation framework, a large-scale dataset $HuGe100K$ , and a scalable reconstruction transformer model, IDOL. This model efficiently reconstructs photorealistic 3D humans in under a second and demonstrates versatility across various applications. This model can efficiently reconstruct photorealistic humans in less than a second and is versatile enough to support various applications.", + "bbox": [ + 511, + 597, + 906, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "However, there are limitations. Due to the constraints of the video model used, we can only synthesize single-frame images from fixed viewpoints. Future work could consider generating longer motion sequences. The focus on body reconstruction and animation leaves facial optimization secondary, and the architecture lacks specific design for facial identity or expression. Additionally, handling half-body inputs remains challenging, and improvements in data generation strategies are needed to enhance performance.", + "bbox": [ + 511, + 763, + 908, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "26315", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was partly supported by the National Key Research and Development Program of China (2022YFF0902400), the Shenzhen Science and Technology Program (JCYJ20220818101014030), and the research fund of Tsinghua University - Tencent Joint Laboratory for Internet Innovation Technology.", + "bbox": [ + 89, + 114, + 483, + 205 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 222, + 187, + 238 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Badour AlBahar, Shunsuke Saito, Hung-Yu Tseng, Changil Kim, Johannes Kopf, and Jia-Bin Huang. Single-image 3d human digitization with shape-guided diffusion. In SIGGRAPH Asia 2023 Conference Papers, 2023. 2, 3", + "[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2", + "[3] Zhongang Cai, Daxuan Ren, Ailing Zeng, Zhengyu Lin, Tao Yu, Wenjia Wang, Xiangyu Fan, Yang Gao, Yifan Yu, Liang Pan, et al. Humman: Multi-modal 4d human dataset for versatile sensing and modeling. In European Conference on Computer Vision, pages 557-577. Springer, 2022. 2, 3", + "[4] Zhongang Cai, Wanqi Yin, Ailing Zeng, Chen Wei, Qingping Sun, Wang Yanjun, Hui En Pang, Haiyi Mei, Mingyuan Zhang, Lei Zhang, et al. Smpler-x: Scaling up expressive human pose and shape estimation. Advances in Neural Information Processing Systems, 36, 2024. 4", + "[5] Di Chang, Yichun Shi, Quankai Gao, Hongyi Xu, Jessica Fu, Guoxian Song, Qing Yan, Yizhe Zhu, Xiao Yang, and Mohammad Soleymani. Magicpose: Realistic human poses and facial expressions retargeting with identity-aware diffusion. In Forty-first International Conference on Machine Learning, 2023. 3", + "[6] Xu Chen, Tianjian Jiang, Jie Song, Max Rietmann, Andreas Geiger, Michael J Black, and Otmar Hilliges. Fast-snarf: A fast deformer for articulated neural fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(10): 11796-11809, 2023. 5", + "[7] Wei Cheng, Su Xu, Jingtan Piao, Chen Qian, Wayne Wu, Kwan-Yee Lin, and Hongsheng Li. Generalizable neural performer: Learning robust radiance fields for human novel view synthesis. arXiv preprint arXiv:2204.11798, 2022. 2, 3", + "[8] Wei Cheng, Ruixiang Chen, Siming Fan, Wanqi Yin, Keyu Chen, Zhongang Cai, Jingbo Wang, Yang Gao, Zhengming Yu, Zhengyu Lin, et al. Dna-rendering: A diverse neural actor repository for high-fidelity human-centric rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19982-19993, 2023. 2, 3", + "[9] Enric Corona, Mihai Zanfir, Thiemo Alldieck, Eduard Gabriel Bazavan, Andrei Zanfir, and Cristian Sminchisescu. Structured 3d features for reconstructing controllable avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16954-16964, 2023. 3" + ], + "bbox": [ + 99, + 247, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[10] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. 2, 3", + "[11] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-xl: A universe of $10\\mathrm{m} + 3\\mathrm{d}$ objects. Advances in Neural Information Processing Systems, 36, 2024. 2, 3", + "[12] Qiao Feng, Yebin Liu, Yu-Kun Lai, Jingyu Yang, and Kun Li. Fof: Learning fourier occupancy field for monocular real-time human reconstruction. In NeurIPS, 2022. 2", + "[13] Xiangjun Gao, Jiaolong Yang, Jongyoo Kim, Sida Peng, Zicheng Liu, and Xin Tong. Mps-nerf: Generalizable 3d human rendering from multiview images. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022.", + "[14] Xiangjun Gao, Xiaoyu Li, Chaopeng Zhang, Qi Zhang, Yanpei Cao, Ying Shan, and Long Quan. Context-human: Free-view rendering of human from a single image with texture-consistent synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10084-10094, 2024. 2", + "[15] Xiangjun Gao, Xiaoyu Li, Yiyu Zhuang, Qi Zhang, Wenbo Hu, Chaopeng Zhang, Yao Yao, Ying Shan, and Long Quan. Mani-gs: Gaussian splatting manipulation with triangular mesh. arXiv preprint arXiv:2405.17811, 2024. 2", + "[16] Junlin Han, Filippos Kokkinos, and Philip Torr. Vfusion3d: Learning scalable 3d generative models from video diffusion models. In European Conference on Computer Vision, pages 333-350. Springer, 2025. 4", + "[17] Sang-Hun Han, Min-Gyu Park, Ju Hong Yoon, Ju-Mi Kang, Young-Jae Park, and Hae-Gon Jeon. High-fidelity 3d human digitization from single 2k resolution images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12869-12879, 2023. 2, 3", + "[18] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 5", + "[19] Tong He, Yuanlu Xu, Shunsuke Saito, Stefano Soatto, and Tony Tung. Arch++: Animation-ready clothed human reconstruction revisited. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11046-11056, 2021. 3", + "[20] Xu He, Xiaoyu Li, Di Kang, Jiangnan Ye, Chaopeng Zhang, Liyang Chen, Xiangjun Gao, Han Zhang, Zhiyong Wu, and Haolin Zhuang. Magicman: Generative novel view synthesis of humans with 3d-aware diffusion and iterative refinement. arXiv preprint arXiv:2408.14211, 2024. 2", + "[21] Yuxiao He, Yiyu Zhuang, Yanwen Wang, Yao Yao, Siyu Zhu, Xiaoyu Li, Qi Zhang, Xun Cao, and Hao Zhu. Head360: Learning a parametric 3d full-head for free-view synthesis in 360. In European Conference on Computer Vision, pages 254-272. Springer, 2024. 2" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "26316", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[22] Yicong Hong, Kai Zhang, Juixiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400, 2023. 2", + "[23] https://github.com/black-forest labs/flux. Flux latent rectified flow transformers, 2024. 3", + "[24] https://github.com/facefusion/facefusion. Facefusion, 2024. 4", + "[25] https://renderpeople.com/3d people. Renderpeople dataset, 2015. 2", + "[26] https://web.twindom.com/. Twindom dataset, 2020. 2", + "[27] Li Hu. Animate anyone: Consistent and controllable image-to-video synthesis for character animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8153-8163, 2024. 3", + "[28] Liangxiao Hu, Hongwen Zhang, Yuxiang Zhang, Boyao Zhou, Boning Liu, Shengping Zhang, and Liqiang Nie. Gaussianavatar: Towards realistic human avatar modeling from a single video via animatable 3d gaussians. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 634-644, 2024. 2, 4", + "[29] Shoukang Hu, Fangzhou Hong, Liang Pan, Haiyi Mei, Lei Yang, and Ziwei Liu. Sherf: Generalizable human nerf from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 3", + "[30] Yangyi Huang, Hongwei Yi, Weiyang Liu, Haofan Wang, Boxi Wu, Wenxiao Wang, Binbin Lin, Debing Zhang, and Deng Cai. One-shot implicit animatable avatars with model-based priors. In IEEE Conference on Computer Vision (ICCV), 2023. 3", + "[31] Zeng Huang, Yuanlu Xu, Christoph Lassner, Hao Li, and Tony Tung. Arch: Animatable reconstruction of clothed humans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3093-3102, 2020. 2, 3", + "[32] Mustafa Isik, Martin Rünz, Markos Georgopoulos, Taras Khakhulin, Jonathan Starck, Lourdes Agapito, and Matthias Nießner. Humanrf: High-fidelity neural radiance fields for humans in motion. ACM Transactions on Graphics (TOG), 42(4):1-12, 2023. 2, 3", + "[33] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 2, 3, 4", + "[34] Rawal Khirodkar, Timur Bagautdinov, Julieta Martinez, Su Zhaoen, Austin James, Peter Selednik, Stuart Anderson, and Shunsuke Saito. Sapiens: Foundation for human vision models. In European Conference on Computer Vision, pages 206-228. Springer, 2025. 2, 5, 6, 8", + "[35] Tobias Kirschstein, Simon Giebenhain, Jiapeng Tang, Markos Georgopoulos, and Matthias Nießner. Gghead: Fast and generalizable 3d gaussian heads. arXiv preprint arXiv:2406.09377, 2024. 4", + "[36] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model." + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In International Conference on Learning Representations (ICLR), 2024. 2", + "[37] Peng Li, Wangguandong Zheng, Yuan Liu, Tao Yu, Yangguang Li, Xingqun Qi, Mengfei Li, Xiaowei Chi, Siyu Xia, Wei Xue, et al. Pshuman: Photorealistic single-view human reconstruction using cross-scale diffusion. arXiv preprint arXiv:2409.10141, 2024. 2", + "[38] Ruilong Li, Shan Yang, David A Ross, and Angjoo Kanazawa. Ai choreographer: Music conditioned 3d dance generation with aist++. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13401-13412, 2021. 2, 3", + "[39] Zhe Li, Zerong Zheng, Lizhen Wang, and Yebin Liu. Animatable gaussians: Learning pose-dependent gaussian maps for high-fidelity human avatar modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19711-19722, 2024. 2", + "[40] Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural actor: Neural free-view synthesis of human actors with pose control. ACM transactions on graphics (TOG), 40(6):1-16, 2021. 2, 3", + "[41] Ziwei Liu, Ping Luo, Shi Qiu, Xiaogang Wang, and Xiaou Tang. Deepfashion: Powering robust clothes recognition and retrieval with rich annotations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1096-1104, 2016. 3", + "[42] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. ACM Transactions on Graphics, 34(6), 2015. 2", + "[43] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[44] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2", + "[45] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 5, 8", + "[46] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10975-10985, 2019. 2", + "[47] Georgios Pavlakos, Dandan Shan, Ilija Radosavovic, Angjoo Kanazawa, David Fouhey, and Jitendra Malik. Reconstructing hands in 3d with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9826-9836, 2024. 4", + "[48] Hao-Yang Peng, Jia-Peng Zhang, Meng-Hao Guo, Yan-Pei Cao, and Shi-Min Hu. Charactergen: Efficient 3d character" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "26317", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "generation from single images with multi-view pose canonicalization. ACM Transactions on Graphics (TOG), 2024. 2, 3", + "[49] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9054-9063, 2021. 2, 3", + "[50] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2", + "[51] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2022. 2", + "[52] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2304-2314, 2019. 2, 3", + "[53] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 84-93, 2020. 2, 3", + "[54] István Sárándi and Gerard Pons-Moll. Neural localizer fields for continuous 3d human pose and shape estimation. arXiv preprint arXiv:2407.07532, 2024. 4", + "[55] Kaiyue Shen, Chen Guo, Manuel Kaufmann, Juan Jose Zarate, Julien Valentin, Jie Song, and Otmar Hilliges. X-avatar: Expressive human avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16911-16921, 2023. 3", + "[56] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023. 6, 8", + "[57] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. In European Conference on Computer Vision, pages 1-18. Springer, 2025. 2, 6, 8", + "[58] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent video diffusion. In European Conference on Computer Vision, pages 439-457. Springer, 2025. 2, 6", + "[59] Qilin Wang, Zhengkai Jiang, Chengming Xu, Jiangning Zhang, Yabiao Wang, Xinyi Zhang, Yun Cao, Weijian Cao, Chengjie Wang, and Yanwei Fu. Vividpose: Advancing stable video diffusion for realistic human image animation. arXiv preprint arXiv:2405.18156, 2024. 3", + "[60] Zhenzhen Weng, Jingyuan Liu, Hao Tan, Zhan Xu, Yang Zhou, Serena Yeung-Levy, and Jimei Yang. Template-free" + ], + "bbox": [ + 91, + 92, + 482, + 901 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "single-view 3d human digitalization with diffusion-guided lrm. arXiv preprint arXiv:2401.12175, 2024. 2, 3", + "[61] Kailu Wu, Fangfu Liu, Zhihan Cai, Runjie Yan, Hanyang Wang, Yating Hu, Yueqi Duan, and Kaisheng Ma. Unique3d: High-quality and efficient 3d mesh generation from a single image. arXiv preprint arXiv:2405.20343, 2024. 2", + "[62] Menghua Wu, Hao Zhu, Linjia Huang, Yiyu Zhuang, Yuanxun Lu, and Xun Cao. High-fidelity 3d face generation from natural language descriptions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4521-4530, 2023. 2", + "[63] Zhangyang Xiong, Chenghong Li, Kenkun Liu, Hongjie Liao, Jianqiao Hu, Junyi Zhu, Shuliang Ning, Lingteng Qiu, Chongjie Wang, Shijie Wang, et al. Mvhumannet: A large-scale dataset of multi-view daily dressing human captures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19801-19811, 2024. 2, 3", + "[64] Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J Black. Icon: Implicit clothed humans obtained from normals. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13286-13296. IEEE, 2022. 2, 3", + "[65] Yuliang Xiu, Jinlong Yang, Xu Cao, Dimitrios Tzionas, and Michael J. Black. ECON: Explicit Clothed humans Optimized via Normal integration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3", + "[66] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191, 2024. 2", + "[67] Zhongcong Xu, Jianfeng Zhang, Jun Hao Liew, Hanshu Yan, Jia-Wei Liu, Chenxu Zhang, Jiashi Feng, and Mike Zheng Shou. Magicanimate: Temporally consistent human image animation using diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1481–1490, 2024. 3", + "[68] Tao Yu, Zerong Zheng, Kaiwen Guo, Pengpeng Liu, Qionghai Dai, and Yebin Liu. Function4d: Real-time human volumetric capture from very sparse consumer rgbd sensors. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5746-5756, 2021. 2, 3", + "[69] Wangbo Yu, Li Yuan, Yan-Pei Cao, Xiangjun Gao, Xiaoyu Li, Wenbo Hu, Long Quan, Ying Shan, and Yonghong Tian. Hifi-123: Towards high-fidelity one image to 3d content generation. In European Conference on Computer Vision, pages 258–274. Springer, 2024. 2", + "[70] Xianggang Yu, Mutian Xu, Yidan Zhang, Haolin Liu, Chongjie Ye, Yushuang Wu, Zizheng Yan, Chenming Zhu, Zhangyang Xiong, Tianyou Liang, et al. Mvmgnet: A large-scale dataset of multi-view images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9150-9161, 2023. 3", + "[71] Zhixuan Yu, Jae Shin Yoon, In Kyu Lee, Prashanth Venkatesh, Jaesik Park, Jihun Yu, and Hyun Soo Park." + ], + "bbox": [ + 516, + 92, + 906, + 901 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "26318", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Humbi: A large multiview dataset of human body expressions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2990-3000, 2020. 2, 3", + "[72] Jingbo Zhang, Xiaoyu Li, Qi Zhang, Yanpei Cao, Ying Shan, and Jing Liao. Humanref: Single image to 3d human generation via reference-guided diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1844–1854, 2024. 2", + "[73] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6", + "[74] Weitian Zhang, Yichao Yan, Yunhui Liu, Xingdong Sheng, and Xiaokang Yang. E3gen: Efficient, expressive and editable avatars generation. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 6860-6869, 2024. 3, 4, 5", + "[75] Yang Zhang, Jiaxi Gu, Li-Wen Wang, Han Wang, Junqi Cheng, Yuefeng Zhu, and Fangyuan Zou. Mimic motion: High-quality human motion video generation with confidence-aware pose guidance. arXiv preprint arXiv:2406.19680, 2024.3, 6", + "[76] Zechuan Zhang, Li Sun, Zongxin Yang, Ling Chen, and Yi Yang. Global-correlated 3d-decoupling transformer for clothed avatar reconstruction. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 3", + "[77] Zechuan Zhang, Li Sun, Zongxin Yang, Ling Chen, and Yi Yang. Global-correlated 3d-decoupling transformer for clothed avatar reconstruction. Advances in Neural Information Processing Systems, 36, 2024. 3, 6, 8", + "[78] Zechuan Zhang, Zongxin Yang, and Yi Yang. Sifu: Side-view conditioned implicit function for real-world usable clothed human reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9936-9947, 2024. 2, 3, 6, 8", + "[79] Zerong Zheng, Tao Yu, Yixuan Wei, Qionghai Dai, and Yebin Liu. Deephuman: 3d human reconstruction from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7739-7749, 2019. 2, 3, 6", + "[80] Zerong Zheng, Tao Yu, Yebin Liu, and Qionghai Dai. Pamir: Parametric model-conditioned implicit representation for image-based human reconstruction. IEEE transactions on pattern analysis and machine intelligence, 44(6): 3170-3184, 2021. 2, 3", + "[81] Shenhao Zhu, Junming Leo Chen, Zuozhuo Dai, Yinghui Xu, Xun Cao, Yao Yao, Hao Zhu, and Siyu Zhu. Champ: Controllable and consistent human image animation with 3d parametric guidance. In European Conference on Computer Vision (ECCV), 2024. 3, 4", + "[82] Yiyu Zhuang, Hao Zhu, Xusen Sun, and Xun Cao. Mofanerf: Morphable facial neural radiance field. In European conference on computer vision, pages 268-285. Springer, 2022. 2", + "[83] Yiyu Zhuang, Qi Zhang, Ying Feng, Hao Zhu, Yao Yao, Xiaoyu Li, Yan-Pei Cao, Ying Shan, and Xun Cao. Anti-aliased" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "neural implicit surfaces with encoding level of detail. In SIGGRAPH Asia 2023 Conference Papers, pages 1-10, 2023. 2", + "[84] Yiyu Zhuang, Qi Zhang, Xuan Wang, Hao Zhu, Ying Feng, Xiaoyu Li, Ying Shan, and Xun Cao. Neai: A pre-convoluted representation for plug-and-play neural ambient illumination. arXiv preprint arXiv:2304.08757, 2023. 2", + "[85] Yiyu Zhuang, Yuxiao He, Jiawei Zhang, Yanwen Wang, Ji-ahe Zhu, Yao Yao, Siyu Zhu, Xun Cao, and Hao Zhu. Towards native generative model for 3d head avatar. arXiv preprint arXiv:2410.01226, 2024. 2" + ], + "bbox": [ + 516, + 90, + 903, + 233 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "26319", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/IDOL_ Instant Photorealistic 3D Human Creation from a Single Image/c28e8ffb-3109-4ced-b9e1-553dbeace189_model.json b/2025/IDOL_ Instant Photorealistic 3D Human Creation from a Single Image/c28e8ffb-3109-4ced-b9e1-553dbeace189_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f117b2c171ac706e2d2d828169b378ce65f3900e --- /dev/null +++ b/2025/IDOL_ Instant Photorealistic 3D Human Creation from a Single Image/c28e8ffb-3109-4ced-b9e1-553dbeace189_model.json @@ -0,0 +1,2490 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.143, + 0.131, + 0.856, + 0.154 + ], + "angle": 0, + "content": "IDOL: Instant Photorealistic 3D Human Creation from a Single Image" + }, + { + "type": "image", + "bbox": [ + 0.134, + 0.179, + 0.865, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.149, + 0.27, + 0.398, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.403, + 0.27, + 0.842, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.15, + 0.423, + 0.398, + 0.574 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.403, + 0.423, + 0.842, + 0.574 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.575, + 0.908, + 0.631 + ], + "angle": 0, + "content": "Figure 1. This work introduces (a) IDOL, a feed-forward, single-image human reconstruction framework that is fast, high-fidelity, and generalizable; (b) Utilizing the proposed Large Generated Human Multi-View Dataset consisting of \\(100K\\) multi-view subjects, our method exhibits exceptional generalizability in handling diverse human shapes, cross-domain data, severe viewpoints, and occlusions; (c) With a uniform structured representation, the avatars can be directly animatable and easily editable." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.643, + 0.327, + 0.658 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.675, + 0.486, + 0.842 + ], + "angle": 0, + "content": "Creating a high-fidelity, animatable 3D full-body avatar from a single image is a challenging task due to the diverse appearance and poses of humans and the limited availability of high-quality training data. To achieve fast and high-quality human reconstruction, this work rethinks the task from the perspectives of dataset, model, and representation. First, we introduce a large-scale HUmancentric GEnerated dataset, HuGe100K, consisting of 100K diverse, photorealistic sets of human images. Each set contains 24-view frames in specific human poses, generated using a pose-controllable image-to-multi-view model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.909, + 0.842 + ], + "angle": 0, + "content": "Next, leveraging the diversity in views, poses, and appearances within HuGe100K, we develop a scalable feedforward transformer model to predict a 3D human Gaussian representation in a uniform space from a given human image. This model is trained to disentangle human pose, body shape, clothing geometry, and texture. The estimated Gaussians can be animated without post-processing. We conduct comprehensive experiments to validate the effectiveness of the proposed dataset and method. Our model demonstrates the ability to efficiently reconstruct photorealistic humans at 1K resolution from a single input image using a single GPU instantly. Additionally, it seamlessly supports various applications, as well as shape and texture editing tasks." + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.851, + 0.483, + 0.875 + ], + "angle": 0, + "content": "Work done during the internship at Tencent by Yiyu Zhuang, Jiaxi Lv, and Wenhao." + }, + { + "type": "page_footnote", + "bbox": [ + 0.117, + 0.876, + 0.231, + 0.888 + ], + "angle": 0, + "content": "*Equal contributions" + }, + { + "type": "page_footnote", + "bbox": [ + 0.117, + 0.888, + 0.245, + 0.9 + ], + "angle": 0, + "content": "†Corresponding authors" + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.851, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "26308" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.223, + 0.107 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.486, + 0.386 + ], + "angle": 0, + "content": "Reconstructing 3D clothed avatars from a single image is crucial in user-friendly virtual reality, gaming, and 3D content creation. This task involves mapping 2D images to 3D models, a highly ill-posed and challenging problem due to the complexity of clothing and the diversity of human poses. Learning-based methods trained on publicly available 3D human models [17, 25, 26, 63, 79] have improved reconstruction quality, but their generalization and quality remain limited [52, 53, 64, 65, 72, 78, 80]. Techniques that integrate parametric body estimation [31, 64, 80], loop optimization [64, 65, 78], and diffusion model priors [1, 37, 60] enhance performance but often result in slower and fragile reconstructions. Recently, large generic image-to-3D reconstruction and generation models [22, 36, 57, 58, 61, 66], leveraging large-scale datasets [10, 11, 43] or pre-trained diffusion models [2, 50, 51], boost capabilities but still struggle with real-life human reconstruction due to the scarcity of photorealistic 3D human data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.388, + 0.485, + 0.81 + ], + "angle": 0, + "content": "Back to the first principle, we rethink existing data, models, and representations. First, current data acquisition is limited by the photographing or scanning of individual subjects, making them inherently unscalable. For instance, the largest publicly available human dataset, MVHumanNet [63], contains only thousands of subjects and clothing variations, far from the dataset scale required to achieve robust model generalization across diverse input images. Second, in the realm of animatable reconstruction models, existing approaches predominantly rely on either reconstruct- and-rig methods that necessitate manual post-processing or entangle human parametric models (e.g., SMPL(-X) [42, 46]) for optimization. However, disentangling human pose and body shape, clothed geometry, and texture could simplify the learning process, avoid error accumulation in parameter estimation, and improve efficiency. Furthermore, when focusing solely on human reconstruction, the incorporation of multi-view image generation or refinement models warrants careful consideration [48, 60, 66]. While these models can introduce finer details, they may also lead to inconsistencies across different views. Third, the community of 3D vision develops various representations tailored to application-specific requirements and technological advancements [33, 44]. For human-centric applications, ideal representations should be well-structured and expressive to facilitate easy rigging and editing [28, 39], as well as capable of high-resolution and fast rendering [33] to enhance both functionality and realism." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.485, + 0.902 + ], + "angle": 0, + "content": "In this work, we introduce a scalable pipeline for training a simple yet efficient feed-forward model for instant photorealistic human reconstruction. We present HuGe100K, a large-scale dataset comprising over 2.4M high-resolution \\((896\\times 640)\\) multi-view images of 100K diverse subjects. To facilitate comprehensive 3D human reconstruction, we de" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.379 + ], + "angle": 0, + "content": "velop a scalable data creation pipeline that integrates synthetic and real-world data, ensuring a wide range of attributes such as age, shape, clothing, race, and gender. Building upon this dataset, we introduce a novel feedforward transformer model IDOL that efficiently predicts 3D human avatars with photorealistic textures and accurate geometry. Our approach leverages a pretrained encoder [34] and a transformer-based backbone for feature extraction and fusion, enabling instant reconstruction (within 1 second on an A100 GPU) without relying on generative models [78] for refinement. By integrating a uniform representation for the 3D human, our model achieves enhanced texture completion and ensures that the reconstructed humans are naturally animatable. Extensive evaluations demonstrate that our method excels in diverse and challenging scenarios, offering superior consistency and quality compared to existing techniques. Additionally, the efficient representation of our model enables seamless applications in animation, editing, and other downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.38, + 0.859, + 0.394 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.399, + 0.907, + 0.473 + ], + "angle": 0, + "content": "- We rethink single-view 3D human reconstruction from the perspectives of data, model, and representation. We introduce a scalable pipeline for training a simple yet efficient feed-forward model, named IDOL, for instant photorealistic human reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.475, + 0.906, + 0.534 + ], + "angle": 0, + "content": "- We develop a data generation pipeline and present HuGe100K, a large-scale multi-view human dataset featuring diverse attributes, high-fidelity, high-resolution appearances, and a well-aligned SMPL-X model." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.535, + 0.906, + 0.595 + ], + "angle": 0, + "content": "- Our comprehensive study demonstrates that leveraging large-scale generated training data significantly enhances model performance and generalizability, paving the way for further scaling up 3D human reconstruction models." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.399, + 0.907, + 0.595 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.612, + 0.655, + 0.628 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.638, + 0.908, + 0.714 + ], + "angle": 0, + "content": "Recent advancements in 3D avatar reconstruction [12-14, 20, 69, 82] and representation learning [15, 62, 83, 84] have greatly improved human modeling. Most relevant to our work are recent efforts on 3D human datasets and single-image human reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.726, + 0.704, + 0.74 + ], + "angle": 0, + "content": "2.1. 3D Human Datasets" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "High-precision 3D human models typically rely on scanning or multi-view camera acquisition systems. Scan-based datasets, such as THuman2.0 [68], Twindom [26], and 2K2K [17], provide high-fidelity 3D geometries but are limited by a small number of subjects, simple standing poses, and non-human artifacts (e.g., 2K2K). In contrast, multiview acquisition systems [3, 7, 8, 21, 32, 38, 40, 49, 63, 71, 85] facilitate the collection of larger datasets with more flexible actions. However, these datasets often face challenges such as biased indoor lighting, fixed viewpoints, and limited" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26309" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.227 + ], + "angle": 0, + "content": "scale. Large-scale synthetic and real datasets for generic objects, including Objaverse and MVImgNet [10, 11, 70], address open-world reconstruction but lack specificity for human models. To overcome these limitations, the proposed HuGe100K significantly scales up dataset size, increasing subject diversity by over 100 times compared to previous datasets. It emphasizes diversity in pose, shape, viewpoint, and clothing, paving the way for large-scale model training. We compare existing datasets with ours in Tab. 1." + }, + { + "type": "table", + "bbox": [ + 0.107, + 0.24, + 0.47, + 0.452 + ], + "angle": 0, + "content": "
TypeDataset#FramesIDsSMPL(-X)
3D ScansTHuman [79]-200✓
THuman2.1 [68]-2500✓
2K2K [17]-2050✗
X-Avatar [55]35.5K20✓
Multi-view ImagesZJU-MoCap [49]180K10✓
Neural Actor [40]250K8✓
HUMBI [71]26M772✓
AIST++ [38]10.1M30✓
HuMMan [3]60M1000✓
GeneBody [7]2.95M50✓
ActorsHQ [32]40K8✗
DNA-Rendering [8]67.5M500✓
MVHumanNet [63]645.1M4500✓
OursHuGe100K>2.4M100K✓
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.462, + 0.483, + 0.505 + ], + "angle": 0, + "content": "Table 1. Comparisons of related datasets. HuGe100K is a large-scale generated multi-view human dataset with \\(100K\\) diverse high-fidelity humans." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.531, + 0.411, + 0.547 + ], + "angle": 0, + "content": "2.2. Single-Image Human Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.553, + 0.483, + 0.886 + ], + "angle": 0, + "content": "For clothed 3D reconstruction methods, implicit representation methods such as PIFU [52], PIFU-HD [53], ARCH [19, 31], and PaMIR [80] have been widely adopted. To enhance reconstruction quality and generalizability, loop optimization techniques integrate implicit representations with explicit or hybrid human priors for improved robustness [1, 64, 65, 76]. Similarly, GTA [77] and SIFU [78] entangle SMPL priors and side-view conditioned features to enhance feature representation. However, these approaches rely heavily on the accuracy of SMPL estimates and are computationally expensive, typically requiring several minutes for inference. Moreover, the resulting 3D representations are not drivable. Recent advancements focus on recovering animatable 3D humans from single images [9, 29, 30]. Some methods benefit from the pre-trained diffusion models or large reconstruction models [48, 60]. Nevertheless, all methods are constrained by the limitations of training datasets, resulting in suboptimal texture detail and limited generalizability. The latest work, E3Gen [74], combines UV maps and Gaussian splatting [33] to directly generate Gaussian attribute maps in UV space from images, which does not support arbitrary image inputs." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.887, + 0.483, + 0.902 + ], + "angle": 0, + "content": "To generate consistent human images or videos given a" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.213 + ], + "angle": 0, + "content": "single image, recent diffusion video models animate a static image with pose-controllable video conditions[5, 5, 27, 59, 59, 67, 81]. Specifically, Champ [81] incorporate several rendered maps obtained from SMPL sequences, alongside skeleton-based motion guidance, to enrich the conditions to the latent diffusion model. However, they all fail to generate a 360-degree video and animate precise expressive SMPL-X-based humans due to insufficient training data." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.225, + 0.681, + 0.24 + ], + "angle": 0, + "content": "3. Dataset Creation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.25, + 0.905, + 0.385 + ], + "angle": 0, + "content": "Drawing inspiration from the latest advancements in generalizable large models, our core insight is to develop a large-scale reconstruction model with exceptional generalization capabilities. The crux of this endeavor lies in creating a comprehensive and high-quality digital human dataset. In this section, we introduce the creation of a large-scale human-centric generated dataset comprising over 2.4M high-resolution \\((896 \\times 640)\\) multi-view images from more than 100K diverse subjects." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.386, + 0.906, + 0.552 + ], + "angle": 0, + "content": "As shown in Fig. 2, the data generation pipeline consists of two stages. Firstly, diverse text prompts are generated by large language models incorporating various human-centric attributes, and high-quality images are synthesized using text-to-image generation models (Sec. 3.1). Secondly, we train a multi-view video generation model, MVChamp, conditioned on rendered full-body motions. With this, a large-scale multi-view images dataset is established by feeding both the synthesized and captured images (Sec. 3.2). Lastly, the statistics and characteristics of the HuGe100K are demonstrated in Sec. 3.3." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.56, + 0.807, + 0.575 + ], + "angle": 0, + "content": "3.1. Image Collection and Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.905, + 0.84 + ], + "angle": 0, + "content": "Given the limited amount of existing datasets [3, 49, 63, 68] and legal concerns regarding portrait and copyright issues, we propose leveraging an image generation model to construct a large-scale, diverse, and high-fidelity dataset of full-body human images. Using the latest text-to-image model Flux [23], we design descriptive prompts based on the required attributes for human figures, ensuring diversity across area, clothing, body shape, age, and gender shown in Fig. 2(i). To avoid long-tail attribute distributions and ensure uniform coverage of traits, we apply uniform sampling in attribute selection for the image generation process. This approach generated over 100K images in total. However, due to issues of visual non-compliance and high similarity among some images, we conducted manual filtering and retained a final set of 90K high-quality images. Additionally, we combine 90K synthetic images and 10K real-life images from DeepFashion [41]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.898, + 0.864 + ], + "angle": 0, + "content": "3.2. Multi-view Image Animation and Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Image-based animating models [5, 27, 67, 75, 81] struggle to achieve generation of \\(360^{\\circ}\\)-consistent and diverse hu" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "26310" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.087, + 0.905, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.328, + 0.907, + 0.385 + ], + "angle": 0, + "content": "Figure 2. Pipeline for constructing our HuGe100K. Diverse attribute combinations from GPT-4 templates create text prompts, generating synthetic images via FLUX, combined with real images from DeepFashion. SMPL-X fitting produces multi-view pose sequences with 360-degree rotations and diverse animatable motions. MVChamp then converts these sequences into multi-view images, ensuring 3D consistency in the dataset." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.411, + 0.486, + 0.713 + ], + "angle": 0, + "content": "man videos conditioned on pose sequences (e.g., 2D poses and 3D SMPL(-X) conditions). To address this limitation, we re-train a state-of-the-art video generative model, Champ [81], to obtain a multi-view consistent generative model, MVChamp. Firstly, We collect a curated dataset of over 100K single-person videos with various motions (e.g., dancing), including approximately 20K videos involving the action of turning around, to fine-tune the model and enhance its generalizability for diverse inputs and motions. Secondly, to further improve 3D consistency[16], we utilize scanned models from THuman 2.1 to generate uniformly distributed views via Blender. From these, we select a subset of 24 views that are evenly spaced around the full \\(360^{\\circ}\\) rotation to fine-tune MVChamp's temporal layers using diffusion loss. Lastly, to support high-quality whole-body animation (i.e., body and hands), we introduce precise SMPL-X estimation using NLF [54] for body shape and pose estimation. Additionally, we employ HaMeR [47] for hand estimation, rendering the corresponding hand template into depth maps as additional control signals." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.485, + 0.901 + ], + "angle": 0, + "content": "After these training processes, MVChamp generates multi-view images conditioned on SMPL-X parametric model. The pose condition is set to an \"A-Pose\" \\(80\\%\\) of the time and a random pose from dance videos (via SMPLer-X [4]) \\(20\\%\\) of the time, balancing pose variability. To address discrepancies between the first and last frames, we propose a Temporal Shift Denoising Strategy: during the denoising steps, we shift latent inputs and pose signals temporally, moving the last frame to the first. This improves continuity between frames without additional inference costs. Finally, FaceFusion [24] is utilized to enhance facial details." + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.407, + 0.887, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.473, + 0.907, + 0.53 + ], + "angle": 0, + "content": "Figure 3. A paired example from the proposed HuGe100K Dataset. For each reference image, we generate a set of multiview images using an estimated shape and a specific pose. The figure shows the pose is well-aligned." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.557, + 0.816, + 0.572 + ], + "angle": 0, + "content": "3.3. Data Statistics and Characteristics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.581, + 0.907, + 0.703 + ], + "angle": 0, + "content": "Tab. 1 compares our dataset HuGe100K to other 3D human datasets regarding scale, diversity, and consistency. With over 100K human identities and 20K poses, HuGe100K offers balanced diversity across dimensions such as area, clothing, body shape, age, and gender (see Fig. 1). Each sample includes a reference image, SMPL-X estimates, 24 uniformly sampled multi-view images, camera parameters, and SMPL-X data (see Fig. 3)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.718, + 0.848, + 0.735 + ], + "angle": 0, + "content": "4. Large Human Reconstruction Model" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.744, + 0.907, + 0.837 + ], + "angle": 0, + "content": "In this section, we present the large-scale human reconstruction model, named IDOL. The overview pipeline is shown in Fig. 4. In Sec. 4.1, we describe the animatable human representation. Sec. 4.2 details the network architecture of IDOL, while Sec. 4.3 explains how we trained the model in an end-to-end manner using the multi-view image dataset." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.847, + 0.825, + 0.864 + ], + "angle": 0, + "content": "4.1. Animatable Human Representation" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Similarly to previous work [28, 35, 74], IDOL leverages 3D Gaussian Splatting [33] in conjunction with SMPL-X" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26311" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.115, + 0.088, + 0.887, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.241, + 0.908, + 0.327 + ], + "angle": 0, + "content": "Figure 4. The architecture of IDOL, a full-differentiable transformer-based framework for reconstructing animatable 3D human from a single image. The model integrates a high-resolution \\((1024\\times 1024)\\) encoder [34] and fuses image tokens with learnable UV tokens through the UV-Alignment Transformer. A UV Decoder predicts Gaussian attribute maps as intermediate representations, capturing the human's geometry and appearance in a structured 2D UV space defined by the SMPL-X model. These maps, in conjunction with the SMPL-X model, represent a 3D human avatar in a canonical space, which can be animated using linear blend skinning (LBS). The model is optimized using multi-view images with diverse poses and identities, learning to disentangle pose, appearance, and shape." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.351, + 0.482, + 0.442 + ], + "angle": 0, + "content": "for 3D human representation, aiming to address the challenges of real-time rendering and accurate animation of human avatars. Specifically, each Gaussian primitive \\(\\mathcal{G}_k\\) is characterized by: \\(\\mathcal{G}_k = \\{\\mu_k,\\alpha_k,\\mathbf{r}_k,\\mathbf{s}_k,\\mathbf{c}_k\\}\\), where \\(\\mu_{k}\\) is the 3D position of the Gaussian, \\(\\alpha_{k}\\) is opacity, \\(\\mathbf{r}_k\\) is the rotation, \\(\\mathbf{s}_k\\) is the scale, and \\(\\mathbf{c}_k\\) is the color." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.46, + 0.485, + 0.779 + ], + "angle": 0, + "content": "3D Gaussian Human. Directly predicting all 3D Gaussian primitives is computationally intensive. Instead, IDOL leverages the predefined 2D UV space of the SMPL-X model to transform the 3D representation task into a more manageable 2D problem. Initially, IDOL predicts Gaussian attribute maps that encode the offset values \\(\\{\\delta_{\\mu_k}, \\delta_{\\mathbf{r}_k}, \\delta \\mathbf{s}_k\\}\\), as well as the color \\(\\mathbf{c}_k\\) and opacity \\(\\alpha_k\\) for each Gaussian primitive \\(\\mathcal{G}_k\\). Similarly to \\(E^3 Gen\\) [74], the position \\(\\mu_k\\), scale \\(\\mathbf{s}_k\\), and rotation \\(\\mathbf{r}_k\\) of each Gaussian primitive \\(\\mathcal{G}_k\\) is modeled relative to its SMPL-X vertex as follows: \\(\\mu_k = \\hat{\\mu}_k + \\delta_{\\mu_k}\\), \\(\\mathbf{s}_k = \\hat{\\mathbf{s}}_k \\cdot \\delta_{\\mathbf{s}_k}\\), \\(\\mathbf{r}_k = \\hat{\\mathbf{r}}_k \\cdot \\delta_{\\mathbf{r}_k}\\), where \\(\\hat{\\mu}_k, \\hat{\\mathbf{s}}_k\\), and \\(\\hat{\\mathbf{r}}_k\\) are the initial values based on the SMPL-X model. The color \\(\\mathbf{c}_k\\) is assigned using RGB values, and the opacity \\(\\alpha_k\\) is set to 1, indicating full opacity. This unified UV space significantly reduces computational complexity and fully exploits the geometric and semantic priors provided by the SMPL-X model. By modeling Gaussian primitives relative to SMPL-X vertices, IDOL ensures semantic consistency across corresponding body parts of diverse avatars, thereby enhancing the model's generalization capability across various reference images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.904 + ], + "angle": 0, + "content": "Animation and Rendering. Given a target pose, we calculate the transformation of each human joint using predefined kinematic relationships. The transformation of each Gaussian primitive is performed using a forward skinning technique based on LBS. Specifically, the position of each Gaussian primitive \\(\\mathcal{G}_k\\) is transformed as follows: \\(\\mu_k' = \\sum_{i=1}^{n_b} w_i \\mathbf{B}_i \\mu_k\\). Additionally, the rotation matrix \\(\\mathbf{R}_k\\) is up-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.35, + 0.907, + 0.442 + ], + "angle": 0, + "content": "dated by: \\(\\mathbf{R}_k^{\\prime} = \\mathbf{T}_k^{1:3,1:3}\\mathbf{R}_k\\), where \\(\\mathbf{T}_k = \\sum_{i=1}^{n_b} w_i \\mathbf{B}_i\\), and \\(\\mathbf{R}_k\\) is the rotation matrix representation of the rotation angle \\(\\mathbf{r}_k\\). Here, \\(n_b\\) is the number of joints, \\(\\mathbf{B}_i\\) is the transformation matrix for each joint, and \\(w_i\\) represents the skinning weights, indicating the influence of each joint's motion on the Gaussian primitive's position \\(\\mu_k\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.442, + 0.909, + 0.579 + ], + "angle": 0, + "content": "To estimate the skinning weights for each \\(\\mathcal{G}_k\\), we first compute a body-part skinning field using a pre-defined low-resolution volume [6]. For smaller regions like the hands and face, which undergo minimal topological variation, weights are interpolated from the SMPL-X template via barycentric coordinates. This strategy enables modeling large topology changes across identities (e.g., from clothing) while stabilizing convergence in less variable areas such as fingers and facial regions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.588, + 0.696, + 0.603 + ], + "angle": 0, + "content": "4.2. Network Structure" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.61, + 0.907, + 0.642 + ], + "angle": 0, + "content": "As illustrated in Fig. 4, IDOL is a full-differentiable framework for reconstructing animatable 3D human." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.909, + 0.902 + ], + "angle": 0, + "content": "High-resolution Image Encoder. Higher resolution of the input image results in a high-quality reconstruction. However, previous works suffer from low-resolution ViT-based encoders, such as DINOv2 [45], which support the \\(448 \\times 448\\) resolution. To fully leverage the resolution of the HuGe100K dataset, we further adopt a high-resolution human foundation model, Sapiens [34], to encode the \\(1024 \\times 1024\\) resolution images into patch-wise feature tokens, formulated as: \\(\\mathbf{F} = \\{\\mathbf{f}_i\\}_{i=1}^n \\in \\mathbb{R}^{d_E}\\), where \\(i\\) denotes the \\(i\\)-th image patch, and \\(d_E\\) is the channel length of each token. The Sapiens model is pretrained on 300 million in-the-wild human images using the Masked Autoencoder (MAE) framework [18], enabling it to excel in preserving fine-grained details and capturing diverse human poses and appearances, making it highly effective for high-resolution human image feature extraction." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26312" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.258 + ], + "angle": 0, + "content": "UV-Alignment Transformer. To map irregular and diverse reference images onto regular UV feature maps, a UV-Alignment Transformer is employed to align learnable spatial-positional UV tokens \\(\\mathbf{Q}^0\\) with reference image features \\(\\mathbf{F}\\). Specifically, the UV-Alignment Transformer concatenates \\(\\mathbf{F}\\) from Spaiens with \\(\\mathbf{Q}^0\\), and aggregates and refines features through \\(D\\) transformer blocks, producing an enhanced representation \\(\\mathbf{Q}^D\\). Each transformer block utilizes self-attention mechanisms, enabling the model to capture complex relationships among the input tokens and impute missing parts using correlated tokens." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.275, + 0.482, + 0.368 + ], + "angle": 0, + "content": "UV Decoder. The UV tokens \\(\\mathbf{Q}^D\\) are reshaped and decoded into Gaussian attribute maps, capturing the human's geometry and appearance within the structured 2D UV space as shown in Sec. 4.1. To preserve and enhance fine details in the decoded Gaussian attribute maps, a Convolutional Neural Network is employed for spatial up-sampling." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.376, + 0.279, + 0.392 + ], + "angle": 0, + "content": "4.3. Training Objectives" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.399, + 0.483, + 0.552 + ], + "angle": 0, + "content": "IDOL reconstructs 3D human by predicting Gaussian attribute UV maps in a single forward pass, offering significant advantages in inference speed and enabling end-to-end training with multi-view images. For each sample, we select a front view as the reference image \\(\\mathbf{I}_{\\mathrm{ref}}\\) along with a random set of generated multi-view images \\(\\{\\mathbf{I}_{\\mathrm{gt},i}\\}_{i = 1}^{N}\\) and their corresponding SMPL-X parameters and camera for supervision. It takes \\(\\mathbf{I}_{\\mathrm{ref}}\\) as input to generate the 3D human, using differentiable rendering to produce the multi-view images \\(\\{\\mathbf{I}_{\\mathrm{pred},i}\\}_{i = 1}^{N}\\). The loss function is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.562, + 0.483, + 0.604 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\sum_ {i = 1} ^ {N} \\left(\\left\\| \\mathbf {I} _ {\\mathrm {g t}, i} - \\mathbf {I} _ {\\text {p r e d}, i} \\right\\| ^ {2} + \\lambda L _ {v g g} \\left(\\mathbf {I} _ {\\mathrm {g t}, i}, \\mathbf {I} _ {\\text {p r e d}, i}\\right)\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.613, + 0.483, + 0.645 + ], + "angle": 0, + "content": "where \\(\\lambda\\) controls the balance between the mean square error loss and the perceptual loss \\(L_{vgg}\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.657, + 0.224, + 0.673 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.682, + 0.383, + 0.697 + ], + "angle": 0, + "content": "5.1. Experiments on Dataset Creation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.483, + 0.854 + ], + "angle": 0, + "content": "Fig. 5 (left) presents the qualitative results of the ablation study on our data generation method. Specifically, Fig. 5 (a) and Fig. 5 (b) demonstrate how hand refinement and face swapping effectively enhance the quality of the generated hands and faces. Fig. 5 (c) highlights the importance of fine-tuning the 3D dataset Thuman 2.1 for improving the 3D consistency of MVChamp, while Fig. 5 (d) illustrates that the Temporal Shift Denoising strategy, which involves cycling the first and last frames, improves the consistency between the first and last frames." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Fig. 5 (right) compares our MVChamp's generation quality against previous human image animation models (e.g., Champ, MimicMotion [75]) and general multi-view" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.228 + ], + "angle": 0, + "content": "video generation model (e.g., SV3D [58]). The issues of these methods are highlighted in the figure. Champ struggles to generate plausible hands and heads in multi-view scenarios. MimicMotion not only fails to produce realistic shoes but also has difficulty preserving the identity, especially for the anime image. SV3D produces low-quality multi-view human generation due to the lack of 3D prior knowledge of the human body. In contrast, our model generates consistent, high-quality results across views." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.237, + 0.838, + 0.253 + ], + "angle": 0, + "content": "5.2. Comparison of Reconstruction Model" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.26, + 0.905, + 0.366 + ], + "angle": 0, + "content": "Implementation Details. For the encoder, we employ the pre-trained Sapiens-1B model [34] and keep its weights frozen during training. We then define a transformer-based framework with a parameter size of \\(0.5B\\) to perform feature fusion. Additionally, we densify the SMPL-X vertex set by sampling approximately 200,000 vertices to map attributes and represent the 3D human model effectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.385, + 0.907, + 0.642 + ], + "angle": 0, + "content": "Dataset and Metrics. We train IDOL on a dataset consisting of generated multi-view images from HuGe100K and rendered images from THuman 2.1 [79]. For THuman 2.1, we render 72 view images at a resolution of \\(896 \\times 640\\) for each scan, with rendering views uniformly distributed. To evaluate performance quantitatively, we reserve the last 50 cases from both HuGe100K and THuman 2.1 as the testing set. We use the most frontal image as the reference image and the remaining images as ground-truth data for evaluation. The ground-truth camera parameters and SMPL-X parameters are provided for all methods. The renderings of the reconstructed model are then compared to their corresponding ground truth images with several metrics, including Mean Squared Error (MSE), Learned Perceptual Image Patch Similarity (LPIPS) [73], Peak Signal-to-Noise Ratio (PSNR). More details on implementation can be found in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.905, + 0.765 + ], + "angle": 0, + "content": "Baselines. We compare IDOL with three baseline categories in the Single-Image Human Reconstruction task. The first category comprises methods based on loop optimization or pixel-alignment modules, including GTA [77] and SIFU [78]. These methods focus on refining human reconstruction through iterative optimization or predicting 3D geometry from pixel-aligned features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.905, + 0.841 + ], + "angle": 0, + "content": "The second category encompasses large-scale generic reconstruction networks, represented by LGM [57]. These models are known for their ability to handle large datasets and their scalability, offering advantages in terms of fast reference speeds and large output resolutions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.905, + 0.901 + ], + "angle": 0, + "content": "The third category involves optimization-based 3D generation methods using score distillation sampling (SDS), exemplified by DreamGaussian [56]. These methods leverage priors from 2D diffusion models to distill 3D objects." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26313" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.089, + 0.437, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.119, + 0.351, + 0.227, + 0.36 + ], + "angle": 0, + "content": "(c) THuman Finetune" + }, + { + "type": "image_caption", + "bbox": [ + 0.252, + 0.351, + 0.442, + 0.361 + ], + "angle": 0, + "content": "(d) Temporal Shift Denoising Strategy" + }, + { + "type": "image", + "bbox": [ + 0.452, + 0.088, + 0.905, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.457, + 0.351, + 0.612, + 0.361 + ], + "angle": 0, + "content": "Reference (e) MVChamp (ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.634, + 0.351, + 0.877, + 0.361 + ], + "angle": 0, + "content": "(f) Champ (g) MimicMotion (h) SV3D" + }, + { + "type": "image_caption", + "bbox": [ + 0.194, + 0.373, + 0.802, + 0.387 + ], + "angle": 0, + "content": "Figure 5. Qualitative results of our MVChamp ablation study (left) and comparison experiment (right)." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.404, + 0.907, + 0.706 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.132, + 0.716, + 0.863, + 0.731 + ], + "angle": 0, + "content": "Figure 6. Comparisons on (a) the upper: novel-view synthesis given a single image, and (b) the lower: our animated results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.757, + 0.482, + 0.819 + ], + "angle": 0, + "content": "DreamGaussian accelerates convergence by progressively densifying 3D Gaussians, significantly reducing the reconstruction time. However, it still takes approximately two minutes to reconstruct a single object." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.485, + 0.901 + ], + "angle": 0, + "content": "Quantitative comparison. As shown in Tab 2, our method outperforms all baselines in all metrics. We attribute this superior performance to the large-scale dataset HuGe100K and the design of our large-scale reconstruc" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.757, + 0.907, + 0.894 + ], + "angle": 0, + "content": "tion model IDOL, which allows for more effective training and improved synthesis of appearance results. Despite this, we note that SIFU and GTA report a lower metric than what we expected in our test settings. While we provide accurate ground-truth SMPL-X parameters and camera settings, the ideal orthographic projection required by the pixel-alignment modules in SIFU and GTA is not well-suited to our perspective projection model, where the camera focus ranges from 35 to 80. This mismatch leads to a" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26314" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.128, + 0.089, + 0.454, + 0.169 + ], + "angle": 0, + "content": "
MethodMSE ↓PSNR ↑LPIPS ↓
SIFU [78]0.04214.2041.612
GTA [77]0.04114.2821.629
Ours-w/o HuGe100K0.01719.2251.326
Ours-full0.00821.6731.138
" + }, + { + "type": "table_caption", + "bbox": [ + 0.101, + 0.179, + 0.473, + 0.193 + ], + "angle": 0, + "content": "Table 2. Evaluation of Comparison and Ablation Experiments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.221, + 0.485, + 0.403 + ], + "angle": 0, + "content": "misalignment between the rendered images and the ground truth, adversely affecting their performance metrics. In fact, many real-life photographs are taken with medium to short focal lengths and cannot be approximated using orthographic projection, which is a less noticeable drawback of SIFU and GTA. Additionally, SIFU and GTA, trained with images at a resolution of \\(512 \\times 512\\), struggle to synthesize detailed textures, particularly in the invisible areas of the reference images. This limitation is primarily due to the lack of comprehensiveness and diversity in their training datasets, which restricts their performance in generating invisible aspects of the images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.424, + 0.484, + 0.605 + ], + "angle": 0, + "content": "Qualitative comparison. We perform a qualitative evaluation on an in-the-wild dataset with methods from [56, 57, 77, 78], with the results presented in Fig. 6. Our evaluation includes a subject in a complex outfit featuring a baseball cap and textured clothing, demonstrating our method's ability to synthesize intricate textures and handle loose outfits across different views. Additional tests include out-of-domain cartoon data and large-angle side views and assessing model adaptability and viewpoint handling. IDOL consistently outperforms the baselines, which struggle with detail reproduction and texture consistency under these various conditions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.616, + 0.443, + 0.631 + ], + "angle": 0, + "content": "5.3. Ablation Study on Reconstruction Model" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.64, + 0.484, + 0.807 + ], + "angle": 0, + "content": "We assess the impact of proposed components by removing them individually from IDOL. As shown in Fig. 7, replacing the Sapiens [34] encoder with DINO v2 [45] (w/o Sapiens) reduces detail quality, resulting in less realistic textures and folds. Excluding the HuGe100K dataset (w/o HuGe100K) causes significant distortion, including blurred details and color bleeding. The complete model, with all components, produces the most realistic and detailed results. These findings highlight the critical roles of both the Sapiens encoder and the HuGe100K dataset in achieving high-quality avatar generation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.817, + 0.327, + 0.833 + ], + "angle": 0, + "content": "5.4. Downstream Applications" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "IDOL reconstructs an avatar by combining Gaussian attribute maps with the SMPL-X model, enabling users to modify the avatar's appearance by editing UV texture maps and to control body shape by adjusting SMPL-X shape pa" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.088, + 0.889, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.539, + 0.295, + 0.88, + 0.31 + ], + "angle": 0, + "content": "Figure 7. Qualitative Results of Ablation Study of IDOL." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.326, + 0.907, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.419, + 0.907, + 0.448 + ], + "angle": 0, + "content": "Figure 8. Controllable Avatar Editing: (a) texture editing; (b) body shape editing." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.477, + 0.907, + 0.553 + ], + "angle": 0, + "content": "rameters. Fig. 8(a) illustrates the effect of texture editing on clothing patterns, while Fig. 8(b) demonstrates shape editing to adjust the avatar's body size. This approach provides a high degree of controllability over both the avatar's appearance and body shape." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.571, + 0.779, + 0.587 + ], + "angle": 0, + "content": "6. Conclusions and Limitations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.907, + 0.764 + ], + "angle": 0, + "content": "In conclusion, our work has made significant strides in creating an animatable 3D human from a single image. We introduced a scalable pipeline for training a simple yet efficient feed-forward model, incorporating a dataset generation framework, a large-scale dataset \\( HuGe100K \\), and a scalable reconstruction transformer model, IDOL. This model efficiently reconstructs photorealistic 3D humans in under a second and demonstrates versatility across various applications. This model can efficiently reconstruct photorealistic humans in less than a second and is versatile enough to support various applications." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.902 + ], + "angle": 0, + "content": "However, there are limitations. Due to the constraints of the video model used, we can only synthesize single-frame images from fixed viewpoints. Future work could consider generating longer motion sequences. The focus on body reconstruction and animation leaves facial optimization secondary, and the architecture lacks specific design for facial identity or expression. Additionally, handling half-body inputs remains challenging, and improvements in data generation strategies are needed to enhance performance." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26315" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.484, + 0.207 + ], + "angle": 0, + "content": "This work was partly supported by the National Key Research and Development Program of China (2022YFF0902400), the Shenzhen Science and Technology Program (JCYJ20220818101014030), and the research fund of Tsinghua University - Tencent Joint Laboratory for Internet Innovation Technology." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.223, + 0.188, + 0.239 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.248, + 0.484, + 0.304 + ], + "angle": 0, + "content": "[1] Badour AlBahar, Shunsuke Saito, Hung-Yu Tseng, Changil Kim, Johannes Kopf, and Jia-Bin Huang. Single-image 3d human digitization with shape-guided diffusion. In SIGGRAPH Asia 2023 Conference Papers, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.306, + 0.484, + 0.375 + ], + "angle": 0, + "content": "[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.377, + 0.484, + 0.447 + ], + "angle": 0, + "content": "[3] Zhongang Cai, Daxuan Ren, Ailing Zeng, Zhengyu Lin, Tao Yu, Wenjia Wang, Xiangyu Fan, Yang Gao, Yifan Yu, Liang Pan, et al. Humman: Multi-modal 4d human dataset for versatile sensing and modeling. In European Conference on Computer Vision, pages 557-577. Springer, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.448, + 0.484, + 0.517 + ], + "angle": 0, + "content": "[4] Zhongang Cai, Wanqi Yin, Ailing Zeng, Chen Wei, Qingping Sun, Wang Yanjun, Hui En Pang, Haiyi Mei, Mingyuan Zhang, Lei Zhang, et al. Smpler-x: Scaling up expressive human pose and shape estimation. Advances in Neural Information Processing Systems, 36, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.519, + 0.484, + 0.603 + ], + "angle": 0, + "content": "[5] Di Chang, Yichun Shi, Quankai Gao, Hongyi Xu, Jessica Fu, Guoxian Song, Qing Yan, Yizhe Zhu, Xiao Yang, and Mohammad Soleymani. Magicpose: Realistic human poses and facial expressions retargeting with identity-aware diffusion. In Forty-first International Conference on Machine Learning, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.605, + 0.484, + 0.673 + ], + "angle": 0, + "content": "[6] Xu Chen, Tianjian Jiang, Jie Song, Max Rietmann, Andreas Geiger, Michael J Black, and Otmar Hilliges. Fast-snarf: A fast deformer for articulated neural fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(10): 11796-11809, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.676, + 0.484, + 0.732 + ], + "angle": 0, + "content": "[7] Wei Cheng, Su Xu, Jingtan Piao, Chen Qian, Wayne Wu, Kwan-Yee Lin, and Hongsheng Li. Generalizable neural performer: Learning robust radiance fields for human novel view synthesis. arXiv preprint arXiv:2204.11798, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.734, + 0.484, + 0.816 + ], + "angle": 0, + "content": "[8] Wei Cheng, Ruixiang Chen, Siming Fan, Wanqi Yin, Keyu Chen, Zhongang Cai, Jingbo Wang, Yang Gao, Zhengming Yu, Zhengyu Lin, et al. Dna-rendering: A diverse neural actor repository for high-fidelity human-centric rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19982-19993, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.818, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[9] Enric Corona, Mihai Zanfir, Thiemo Alldieck, Eduard Gabriel Bazavan, Andrei Zanfir, and Cristian Sminchisescu. Structured 3d features for reconstructing controllable avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16954-16964, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.248, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.176 + ], + "angle": 0, + "content": "[10] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.907, + 0.247 + ], + "angle": 0, + "content": "[11] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-xl: A universe of \\(10\\mathrm{m} + 3\\mathrm{d}\\) objects. Advances in Neural Information Processing Systems, 36, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.907, + 0.289 + ], + "angle": 0, + "content": "[12] Qiao Feng, Yebin Liu, Yu-Kun Lai, Jingyu Yang, and Kun Li. Fof: Learning fourier occupancy field for monocular real-time human reconstruction. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.291, + 0.906, + 0.348 + ], + "angle": 0, + "content": "[13] Xiangjun Gao, Jiaolong Yang, Jongyoo Kim, Sida Peng, Zicheng Liu, and Xin Tong. Mps-nerf: Generalizable 3d human rendering from multiview images. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.35, + 0.906, + 0.431 + ], + "angle": 0, + "content": "[14] Xiangjun Gao, Xiaoyu Li, Chaopeng Zhang, Qi Zhang, Yanpei Cao, Ying Shan, and Long Quan. Context-human: Free-view rendering of human from a single image with texture-consistent synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10084-10094, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.434, + 0.906, + 0.489 + ], + "angle": 0, + "content": "[15] Xiangjun Gao, Xiaoyu Li, Yiyu Zhuang, Qi Zhang, Wenbo Hu, Chaopeng Zhang, Yao Yao, Ying Shan, and Long Quan. Mani-gs: Gaussian splatting manipulation with triangular mesh. arXiv preprint arXiv:2405.17811, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.491, + 0.906, + 0.547 + ], + "angle": 0, + "content": "[16] Junlin Han, Filippos Kokkinos, and Philip Torr. Vfusion3d: Learning scalable 3d generative models from video diffusion models. In European Conference on Computer Vision, pages 333-350. Springer, 2025. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.549, + 0.906, + 0.617 + ], + "angle": 0, + "content": "[17] Sang-Hun Han, Min-Gyu Park, Ju Hong Yoon, Ju-Mi Kang, Young-Jae Park, and Hae-Gon Jeon. High-fidelity 3d human digitization from single 2k resolution images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12869-12879, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.62, + 0.906, + 0.687 + ], + "angle": 0, + "content": "[18] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.69, + 0.906, + 0.758 + ], + "angle": 0, + "content": "[19] Tong He, Yuanlu Xu, Shunsuke Saito, Stefano Soatto, and Tony Tung. Arch++: Animation-ready clothed human reconstruction revisited. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11046-11056, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.761, + 0.906, + 0.83 + ], + "angle": 0, + "content": "[20] Xu He, Xiaoyu Li, Di Kang, Jiangnan Ye, Chaopeng Zhang, Liyang Chen, Xiangjun Gao, Han Zhang, Zhiyong Wu, and Haolin Zhuang. Magicman: Generative novel view synthesis of humans with 3d-aware diffusion and iterative refinement. arXiv preprint arXiv:2408.14211, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[21] Yuxiao He, Yiyu Zhuang, Yanwen Wang, Yao Yao, Siyu Zhu, Xiaoyu Li, Qi Zhang, Xun Cao, and Hao Zhu. Head360: Learning a parametric 3d full-head for free-view synthesis in 360. In European Conference on Computer Vision, pages 254-272. Springer, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "26316" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[22] Yicong Hong, Kai Zhang, Juixiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.175 + ], + "angle": 0, + "content": "[23] https://github.com/black-forest labs/flux. Flux latent rectified flow transformers, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.179, + 0.482, + 0.204 + ], + "angle": 0, + "content": "[24] https://github.com/facefusion/facefusion. Facefusion, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.208, + 0.482, + 0.234 + ], + "angle": 0, + "content": "[25] https://renderpeople.com/3d people. Renderpeople dataset, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.237, + 0.445, + 0.25 + ], + "angle": 0, + "content": "[26] https://web.twindom.com/. Twindom dataset, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.253, + 0.484, + 0.307 + ], + "angle": 0, + "content": "[27] Li Hu. Animate anyone: Consistent and controllable image-to-video synthesis for character animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8153-8163, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.309, + 0.482, + 0.391 + ], + "angle": 0, + "content": "[28] Liangxiao Hu, Hongwen Zhang, Yuxiang Zhang, Boyao Zhou, Boning Liu, Shengping Zhang, and Liqiang Nie. Gaussianavatar: Towards realistic human avatar modeling from a single video via animatable 3d gaussians. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 634-644, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.393, + 0.482, + 0.448 + ], + "angle": 0, + "content": "[29] Shoukang Hu, Fangzhou Hong, Liang Pan, Haiyi Mei, Lei Yang, and Ziwei Liu. Sherf: Generalizable human nerf from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.45, + 0.482, + 0.518 + ], + "angle": 0, + "content": "[30] Yangyi Huang, Hongwei Yi, Weiyang Liu, Haofan Wang, Boxi Wu, Wenxiao Wang, Binbin Lin, Debing Zhang, and Deng Cai. One-shot implicit animatable avatars with model-based priors. In IEEE Conference on Computer Vision (ICCV), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.521, + 0.482, + 0.588 + ], + "angle": 0, + "content": "[31] Zeng Huang, Yuanlu Xu, Christoph Lassner, Hao Li, and Tony Tung. Arch: Animatable reconstruction of clothed humans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3093-3102, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.591, + 0.482, + 0.659 + ], + "angle": 0, + "content": "[32] Mustafa Isik, Martin Rünz, Markos Georgopoulos, Taras Khakhulin, Jonathan Starck, Lourdes Agapito, and Matthias Nießner. Humanrf: High-fidelity neural radiance fields for humans in motion. ACM Transactions on Graphics (TOG), 42(4):1-12, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.662, + 0.482, + 0.716 + ], + "angle": 0, + "content": "[33] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.719, + 0.482, + 0.787 + ], + "angle": 0, + "content": "[34] Rawal Khirodkar, Timur Bagautdinov, Julieta Martinez, Su Zhaoen, Austin James, Peter Selednik, Stuart Anderson, and Shunsuke Saito. Sapiens: Foundation for human vision models. In European Conference on Computer Vision, pages 206-228. Springer, 2025. 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[35] Tobias Kirschstein, Simon Giebenhain, Jiapeng Tang, Markos Georgopoulos, and Matthias Nießner. Gghead: Fast and generalizable 3d gaussian heads. arXiv preprint arXiv:2406.09377, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[36] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "In International Conference on Learning Representations (ICLR), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.188 + ], + "angle": 0, + "content": "[37] Peng Li, Wangguandong Zheng, Yuan Liu, Tao Yu, Yangguang Li, Xingqun Qi, Mengfei Li, Xiaowei Chi, Siyu Xia, Wei Xue, et al. Pshuman: Photorealistic single-view human reconstruction using cross-scale diffusion. arXiv preprint arXiv:2409.10141, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.19, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[38] Ruilong Li, Shan Yang, David A Ross, and Angjoo Kanazawa. Ai choreographer: Music conditioned 3d dance generation with aist++. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13401-13412, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.329 + ], + "angle": 0, + "content": "[39] Zhe Li, Zerong Zheng, Lizhen Wang, and Yebin Liu. Animatable gaussians: Learning pose-dependent gaussian maps for high-fidelity human avatar modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19711-19722, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.905, + 0.398 + ], + "angle": 0, + "content": "[40] Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural actor: Neural free-view synthesis of human actors with pose control. ACM transactions on graphics (TOG), 40(6):1-16, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.905, + 0.468 + ], + "angle": 0, + "content": "[41] Ziwei Liu, Ping Luo, Shi Qiu, Xiaogang Wang, and Xiaou Tang. Deepfashion: Powering robust clothes recognition and retrieval with rich annotations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1096-1104, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.905, + 0.523 + ], + "angle": 0, + "content": "[42] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. ACM Transactions on Graphics, 34(6), 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.905, + 0.579 + ], + "angle": 0, + "content": "[43] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.581, + 0.905, + 0.648 + ], + "angle": 0, + "content": "[44] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.65, + 0.905, + 0.719 + ], + "angle": 0, + "content": "[45] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.721, + 0.905, + 0.802 + ], + "angle": 0, + "content": "[46] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10975-10985, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[47] Georgios Pavlakos, Dandan Shan, Ilija Radosavovic, Angjoo Kanazawa, David Fouhey, and Jitendra Malik. Reconstructing hands in 3d with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9826-9836, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[48] Hao-Yang Peng, Jia-Peng Zhang, Meng-Hao Guo, Yan-Pei Cao, and Shi-Min Hu. Charactergen: Efficient 3d character" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26317" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.482, + 0.133 + ], + "angle": 0, + "content": "generation from single images with multi-view pose canonicalization. ACM Transactions on Graphics (TOG), 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.219 + ], + "angle": 0, + "content": "[49] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9054-9063, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.289 + ], + "angle": 0, + "content": "[50] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[51] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.404 + ], + "angle": 0, + "content": "[52] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2304-2314, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.483, + 0.475 + ], + "angle": 0, + "content": "[53] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 84-93, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.482, + 0.518 + ], + "angle": 0, + "content": "[54] István Sárándi and Gerard Pons-Moll. Neural localizer fields for continuous 3d human pose and shape estimation. arXiv preprint arXiv:2407.07532, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.482, + 0.589 + ], + "angle": 0, + "content": "[55] Kaiyue Shen, Chen Guo, Manuel Kaufmann, Juan Jose Zarate, Julien Valentin, Jie Song, and Otmar Hilliges. X-avatar: Expressive human avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16911-16921, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.482, + 0.645 + ], + "angle": 0, + "content": "[56] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.482, + 0.716 + ], + "angle": 0, + "content": "[57] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. In European Conference on Computer Vision, pages 1-18. Springer, 2025. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.482, + 0.802 + ], + "angle": 0, + "content": "[58] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent video diffusion. In European Conference on Computer Vision, pages 439-457. Springer, 2025. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[59] Qilin Wang, Zhengkai Jiang, Chengming Xu, Jiangning Zhang, Yabiao Wang, Xinyi Zhang, Yun Cao, Weijian Cao, Chengjie Wang, and Yanwei Fu. Vividpose: Advancing stable video diffusion for realistic human image animation. arXiv preprint arXiv:2405.18156, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[60] Zhenzhen Weng, Jingyuan Liu, Hao Tan, Zhan Xu, Yang Zhou, Serena Yeung-Levy, and Jimei Yang. Template-free" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "single-view 3d human digitalization with diffusion-guided lrm. arXiv preprint arXiv:2401.12175, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[61] Kailu Wu, Fangfu Liu, Zhihan Cai, Runjie Yan, Hanyang Wang, Yating Hu, Yueqi Duan, and Kaisheng Ma. Unique3d: High-quality and efficient 3d mesh generation from a single image. arXiv preprint arXiv:2405.20343, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.907, + 0.248 + ], + "angle": 0, + "content": "[62] Menghua Wu, Hao Zhu, Linjia Huang, Yiyu Zhuang, Yuanxun Lu, and Xun Cao. High-fidelity 3d face generation from natural language descriptions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4521-4530, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.346 + ], + "angle": 0, + "content": "[63] Zhangyang Xiong, Chenghong Li, Kenkun Liu, Hongjie Liao, Jianqiao Hu, Junyi Zhu, Shuliang Ning, Lingteng Qiu, Chongjie Wang, Shijie Wang, et al. Mvhumannet: A large-scale dataset of multi-view daily dressing human captures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19801-19811, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[64] Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J Black. Icon: Implicit clothed humans obtained from normals. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13286-13296. IEEE, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.905, + 0.489 + ], + "angle": 0, + "content": "[65] Yuliang Xiu, Jinlong Yang, Xu Cao, Dimitrios Tzionas, and Michael J. Black. ECON: Explicit Clothed humans Optimized via Normal integration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.491, + 0.905, + 0.559 + ], + "angle": 0, + "content": "[66] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.562, + 0.905, + 0.645 + ], + "angle": 0, + "content": "[67] Zhongcong Xu, Jianfeng Zhang, Jun Hao Liew, Hanshu Yan, Jia-Wei Liu, Chenxu Zhang, Jiashi Feng, and Mike Zheng Shou. Magicanimate: Temporally consistent human image animation using diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1481–1490, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.647, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[68] Tao Yu, Zerong Zheng, Kaiwen Guo, Pengpeng Liu, Qionghai Dai, and Yebin Liu. Function4d: Real-time human volumetric capture from very sparse consumer rgbd sensors. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5746-5756, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.905, + 0.787 + ], + "angle": 0, + "content": "[69] Wangbo Yu, Li Yuan, Yan-Pei Cao, Xiangjun Gao, Xiaoyu Li, Wenbo Hu, Long Quan, Ying Shan, and Yonghong Tian. Hifi-123: Towards high-fidelity one image to 3d content generation. In European Conference on Computer Vision, pages 258–274. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[70] Xianggang Yu, Mutian Xu, Yidan Zhang, Haolin Liu, Chongjie Ye, Yushuang Wu, Zizheng Yan, Chenming Zhu, Zhangyang Xiong, Tianyou Liang, et al. Mvmgnet: A large-scale dataset of multi-view images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9150-9161, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[71] Zhixuan Yu, Jae Shin Yoon, In Kyu Lee, Prashanth Venkatesh, Jaesik Park, Jihun Yu, and Hyun Soo Park." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26318" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.147 + ], + "angle": 0, + "content": "Humbi: A large multiview dataset of human body expressions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2990-3000, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.484, + 0.219 + ], + "angle": 0, + "content": "[72] Jingbo Zhang, Xiaoyu Li, Qi Zhang, Yanpei Cao, Ying Shan, and Jing Liao. Humanref: Single image to 3d human generation via reference-guided diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1844–1854, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.483, + 0.29 + ], + "angle": 0, + "content": "[73] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.483, + 0.359 + ], + "angle": 0, + "content": "[74] Weitian Zhang, Yichao Yan, Yunhui Liu, Xingdong Sheng, and Xiaokang Yang. E3gen: Efficient, expressive and editable avatars generation. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 6860-6869, 2024. 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.361, + 0.483, + 0.431 + ], + "angle": 0, + "content": "[75] Yang Zhang, Jiaxi Gu, Li-Wen Wang, Han Wang, Junqi Cheng, Yuefeng Zhu, and Fangyuan Zou. Mimic motion: High-quality human motion video generation with confidence-aware pose guidance. arXiv preprint arXiv:2406.19680, 2024.3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.433, + 0.483, + 0.488 + ], + "angle": 0, + "content": "[76] Zechuan Zhang, Li Sun, Zongxin Yang, Ling Chen, and Yi Yang. Global-correlated 3d-decoupling transformer for clothed avatar reconstruction. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.483, + 0.545 + ], + "angle": 0, + "content": "[77] Zechuan Zhang, Li Sun, Zongxin Yang, Ling Chen, and Yi Yang. Global-correlated 3d-decoupling transformer for clothed avatar reconstruction. Advances in Neural Information Processing Systems, 36, 2024. 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.547, + 0.483, + 0.616 + ], + "angle": 0, + "content": "[78] Zechuan Zhang, Zongxin Yang, and Yi Yang. Sifu: Side-view conditioned implicit function for real-world usable clothed human reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9936-9947, 2024. 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.617, + 0.483, + 0.686 + ], + "angle": 0, + "content": "[79] Zerong Zheng, Tao Yu, Yixuan Wei, Qionghai Dai, and Yebin Liu. Deephuman: 3d human reconstruction from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7739-7749, 2019. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.688, + 0.483, + 0.757 + ], + "angle": 0, + "content": "[80] Zerong Zheng, Tao Yu, Yebin Liu, and Qionghai Dai. Pamir: Parametric model-conditioned implicit representation for image-based human reconstruction. IEEE transactions on pattern analysis and machine intelligence, 44(6): 3170-3184, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.759, + 0.483, + 0.828 + ], + "angle": 0, + "content": "[81] Shenhao Zhu, Junming Leo Chen, Zuozhuo Dai, Yinghui Xu, Xun Cao, Yao Yao, Hao Zhu, and Siyu Zhu. Champ: Controllable and consistent human image animation with 3d parametric guidance. In European Conference on Computer Vision (ECCV), 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.83, + 0.483, + 0.872 + ], + "angle": 0, + "content": "[82] Yiyu Zhuang, Hao Zhu, Xusen Sun, and Xun Cao. Mofanerf: Morphable facial neural radiance field. In European conference on computer vision, pages 268-285. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[83] Yiyu Zhuang, Qi Zhang, Ying Feng, Hao Zhu, Yao Yao, Xiaoyu Li, Yan-Pei Cao, Ying Shan, and Xun Cao. Anti-aliased" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "neural implicit surfaces with encoding level of detail. In SIGGRAPH Asia 2023 Conference Papers, pages 1-10, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[84] Yiyu Zhuang, Qi Zhang, Xuan Wang, Hao Zhu, Ying Feng, Xiaoyu Li, Ying Shan, and Xun Cao. Neai: A pre-convoluted representation for plug-and-play neural ambient illumination. arXiv preprint arXiv:2304.08757, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[85] Yiyu Zhuang, Yuxiao He, Jiawei Zhang, Yanwen Wang, Ji-ahe Zhu, Yao Yao, Siyu Zhu, Xun Cao, and Hao Zhu. Towards native generative model for 3d head avatar. arXiv preprint arXiv:2410.01226, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26319" + } + ] +] \ No newline at end of file diff --git a/2025/IDOL_ Instant Photorealistic 3D Human Creation from a Single Image/c28e8ffb-3109-4ced-b9e1-553dbeace189_origin.pdf b/2025/IDOL_ Instant Photorealistic 3D Human Creation from a Single Image/c28e8ffb-3109-4ced-b9e1-553dbeace189_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ad457ca68b01c3d640ff1cd94f59c555a6f8baff --- /dev/null +++ b/2025/IDOL_ Instant Photorealistic 3D Human Creation from a Single Image/c28e8ffb-3109-4ced-b9e1-553dbeace189_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:622b0f7ceef49cb8319187e1035c71ee4bbea306cac7154bc4c34c58c83a81cf +size 8223194 diff --git a/2025/IDOL_ Instant Photorealistic 3D Human Creation from a Single Image/full.md b/2025/IDOL_ Instant Photorealistic 3D Human Creation from a Single Image/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a469088bc59c94b0c807a7dc2916d5e3b88a1f7c --- /dev/null +++ b/2025/IDOL_ Instant Photorealistic 3D Human Creation from a Single Image/full.md @@ -0,0 +1,299 @@ +# IDOL: Instant Photorealistic 3D Human Creation from a Single Image + +![](images/6f12e6f0b256d25f38b2c0b2d88f71f6531892c93090f89911bd4113a582a57a.jpg) + +![](images/ab4b622cbe0ce6d245a6a7add80b6aa698da8bd180fcf5f33ee7c8bcf2b3c61c.jpg) + +![](images/319ff23135b5cb69b47d31d4e1377b0d9cd97bc2182700ac5a93d7d1ea3c1e24.jpg) + +![](images/a5afe8c7f64f696cf13ea57d102b7db378bc0ccb311228de9aacaee30d73431f.jpg) +Figure 1. This work introduces (a) IDOL, a feed-forward, single-image human reconstruction framework that is fast, high-fidelity, and generalizable; (b) Utilizing the proposed Large Generated Human Multi-View Dataset consisting of $100K$ multi-view subjects, our method exhibits exceptional generalizability in handling diverse human shapes, cross-domain data, severe viewpoints, and occlusions; (c) With a uniform structured representation, the avatars can be directly animatable and easily editable. + +![](images/8c58ee58915cc674be7d147ab614c31b5a845056ff29f23e7478b2d9ebe13f1a.jpg) + +# Abstract + +Creating a high-fidelity, animatable 3D full-body avatar from a single image is a challenging task due to the diverse appearance and poses of humans and the limited availability of high-quality training data. To achieve fast and high-quality human reconstruction, this work rethinks the task from the perspectives of dataset, model, and representation. First, we introduce a large-scale HUmancentric GEnerated dataset, HuGe100K, consisting of 100K diverse, photorealistic sets of human images. Each set contains 24-view frames in specific human poses, generated using a pose-controllable image-to-multi-view model. + +Next, leveraging the diversity in views, poses, and appearances within HuGe100K, we develop a scalable feedforward transformer model to predict a 3D human Gaussian representation in a uniform space from a given human image. This model is trained to disentangle human pose, body shape, clothing geometry, and texture. The estimated Gaussians can be animated without post-processing. We conduct comprehensive experiments to validate the effectiveness of the proposed dataset and method. Our model demonstrates the ability to efficiently reconstruct photorealistic humans at 1K resolution from a single input image using a single GPU instantly. Additionally, it seamlessly supports various applications, as well as shape and texture editing tasks. + +# 1. Introduction + +Reconstructing 3D clothed avatars from a single image is crucial in user-friendly virtual reality, gaming, and 3D content creation. This task involves mapping 2D images to 3D models, a highly ill-posed and challenging problem due to the complexity of clothing and the diversity of human poses. Learning-based methods trained on publicly available 3D human models [17, 25, 26, 63, 79] have improved reconstruction quality, but their generalization and quality remain limited [52, 53, 64, 65, 72, 78, 80]. Techniques that integrate parametric body estimation [31, 64, 80], loop optimization [64, 65, 78], and diffusion model priors [1, 37, 60] enhance performance but often result in slower and fragile reconstructions. Recently, large generic image-to-3D reconstruction and generation models [22, 36, 57, 58, 61, 66], leveraging large-scale datasets [10, 11, 43] or pre-trained diffusion models [2, 50, 51], boost capabilities but still struggle with real-life human reconstruction due to the scarcity of photorealistic 3D human data. + +Back to the first principle, we rethink existing data, models, and representations. First, current data acquisition is limited by the photographing or scanning of individual subjects, making them inherently unscalable. For instance, the largest publicly available human dataset, MVHumanNet [63], contains only thousands of subjects and clothing variations, far from the dataset scale required to achieve robust model generalization across diverse input images. Second, in the realm of animatable reconstruction models, existing approaches predominantly rely on either reconstruct- and-rig methods that necessitate manual post-processing or entangle human parametric models (e.g., SMPL(-X) [42, 46]) for optimization. However, disentangling human pose and body shape, clothed geometry, and texture could simplify the learning process, avoid error accumulation in parameter estimation, and improve efficiency. Furthermore, when focusing solely on human reconstruction, the incorporation of multi-view image generation or refinement models warrants careful consideration [48, 60, 66]. While these models can introduce finer details, they may also lead to inconsistencies across different views. Third, the community of 3D vision develops various representations tailored to application-specific requirements and technological advancements [33, 44]. For human-centric applications, ideal representations should be well-structured and expressive to facilitate easy rigging and editing [28, 39], as well as capable of high-resolution and fast rendering [33] to enhance both functionality and realism. + +In this work, we introduce a scalable pipeline for training a simple yet efficient feed-forward model for instant photorealistic human reconstruction. We present HuGe100K, a large-scale dataset comprising over 2.4M high-resolution $(896\times 640)$ multi-view images of 100K diverse subjects. To facilitate comprehensive 3D human reconstruction, we de + +velop a scalable data creation pipeline that integrates synthetic and real-world data, ensuring a wide range of attributes such as age, shape, clothing, race, and gender. Building upon this dataset, we introduce a novel feedforward transformer model IDOL that efficiently predicts 3D human avatars with photorealistic textures and accurate geometry. Our approach leverages a pretrained encoder [34] and a transformer-based backbone for feature extraction and fusion, enabling instant reconstruction (within 1 second on an A100 GPU) without relying on generative models [78] for refinement. By integrating a uniform representation for the 3D human, our model achieves enhanced texture completion and ensures that the reconstructed humans are naturally animatable. Extensive evaluations demonstrate that our method excels in diverse and challenging scenarios, offering superior consistency and quality compared to existing techniques. Additionally, the efficient representation of our model enables seamless applications in animation, editing, and other downstream tasks. + +Our contributions can be summarized as follows: + +- We rethink single-view 3D human reconstruction from the perspectives of data, model, and representation. We introduce a scalable pipeline for training a simple yet efficient feed-forward model, named IDOL, for instant photorealistic human reconstruction. +- We develop a data generation pipeline and present HuGe100K, a large-scale multi-view human dataset featuring diverse attributes, high-fidelity, high-resolution appearances, and a well-aligned SMPL-X model. +- Our comprehensive study demonstrates that leveraging large-scale generated training data significantly enhances model performance and generalizability, paving the way for further scaling up 3D human reconstruction models. + +# 2. Related Work + +Recent advancements in 3D avatar reconstruction [12-14, 20, 69, 82] and representation learning [15, 62, 83, 84] have greatly improved human modeling. Most relevant to our work are recent efforts on 3D human datasets and single-image human reconstruction. + +# 2.1. 3D Human Datasets + +High-precision 3D human models typically rely on scanning or multi-view camera acquisition systems. Scan-based datasets, such as THuman2.0 [68], Twindom [26], and 2K2K [17], provide high-fidelity 3D geometries but are limited by a small number of subjects, simple standing poses, and non-human artifacts (e.g., 2K2K). In contrast, multiview acquisition systems [3, 7, 8, 21, 32, 38, 40, 49, 63, 71, 85] facilitate the collection of larger datasets with more flexible actions. However, these datasets often face challenges such as biased indoor lighting, fixed viewpoints, and limited + +scale. Large-scale synthetic and real datasets for generic objects, including Objaverse and MVImgNet [10, 11, 70], address open-world reconstruction but lack specificity for human models. To overcome these limitations, the proposed HuGe100K significantly scales up dataset size, increasing subject diversity by over 100 times compared to previous datasets. It emphasizes diversity in pose, shape, viewpoint, and clothing, paving the way for large-scale model training. We compare existing datasets with ours in Tab. 1. + +
TypeDataset#FramesIDsSMPL(-X)
3D ScansTHuman [79]-200✓
THuman2.1 [68]-2500✓
2K2K [17]-2050✗
X-Avatar [55]35.5K20✓
Multi-view ImagesZJU-MoCap [49]180K10✓
Neural Actor [40]250K8✓
HUMBI [71]26M772✓
AIST++ [38]10.1M30✓
HuMMan [3]60M1000✓
GeneBody [7]2.95M50✓
ActorsHQ [32]40K8✗
DNA-Rendering [8]67.5M500✓
MVHumanNet [63]645.1M4500✓
OursHuGe100K>2.4M100K✓
+ +Table 1. Comparisons of related datasets. HuGe100K is a large-scale generated multi-view human dataset with $100K$ diverse high-fidelity humans. + +# 2.2. Single-Image Human Reconstruction + +For clothed 3D reconstruction methods, implicit representation methods such as PIFU [52], PIFU-HD [53], ARCH [19, 31], and PaMIR [80] have been widely adopted. To enhance reconstruction quality and generalizability, loop optimization techniques integrate implicit representations with explicit or hybrid human priors for improved robustness [1, 64, 65, 76]. Similarly, GTA [77] and SIFU [78] entangle SMPL priors and side-view conditioned features to enhance feature representation. However, these approaches rely heavily on the accuracy of SMPL estimates and are computationally expensive, typically requiring several minutes for inference. Moreover, the resulting 3D representations are not drivable. Recent advancements focus on recovering animatable 3D humans from single images [9, 29, 30]. Some methods benefit from the pre-trained diffusion models or large reconstruction models [48, 60]. Nevertheless, all methods are constrained by the limitations of training datasets, resulting in suboptimal texture detail and limited generalizability. The latest work, E3Gen [74], combines UV maps and Gaussian splatting [33] to directly generate Gaussian attribute maps in UV space from images, which does not support arbitrary image inputs. + +To generate consistent human images or videos given a + +single image, recent diffusion video models animate a static image with pose-controllable video conditions[5, 5, 27, 59, 59, 67, 81]. Specifically, Champ [81] incorporate several rendered maps obtained from SMPL sequences, alongside skeleton-based motion guidance, to enrich the conditions to the latent diffusion model. However, they all fail to generate a 360-degree video and animate precise expressive SMPL-X-based humans due to insufficient training data. + +# 3. Dataset Creation + +Drawing inspiration from the latest advancements in generalizable large models, our core insight is to develop a large-scale reconstruction model with exceptional generalization capabilities. The crux of this endeavor lies in creating a comprehensive and high-quality digital human dataset. In this section, we introduce the creation of a large-scale human-centric generated dataset comprising over 2.4M high-resolution $(896 \times 640)$ multi-view images from more than 100K diverse subjects. + +As shown in Fig. 2, the data generation pipeline consists of two stages. Firstly, diverse text prompts are generated by large language models incorporating various human-centric attributes, and high-quality images are synthesized using text-to-image generation models (Sec. 3.1). Secondly, we train a multi-view video generation model, MVChamp, conditioned on rendered full-body motions. With this, a large-scale multi-view images dataset is established by feeding both the synthesized and captured images (Sec. 3.2). Lastly, the statistics and characteristics of the HuGe100K are demonstrated in Sec. 3.3. + +# 3.1. Image Collection and Generation + +Given the limited amount of existing datasets [3, 49, 63, 68] and legal concerns regarding portrait and copyright issues, we propose leveraging an image generation model to construct a large-scale, diverse, and high-fidelity dataset of full-body human images. Using the latest text-to-image model Flux [23], we design descriptive prompts based on the required attributes for human figures, ensuring diversity across area, clothing, body shape, age, and gender shown in Fig. 2(i). To avoid long-tail attribute distributions and ensure uniform coverage of traits, we apply uniform sampling in attribute selection for the image generation process. This approach generated over 100K images in total. However, due to issues of visual non-compliance and high similarity among some images, we conducted manual filtering and retained a final set of 90K high-quality images. Additionally, we combine 90K synthetic images and 10K real-life images from DeepFashion [41]. + +# 3.2. Multi-view Image Animation and Generation + +Image-based animating models [5, 27, 67, 75, 81] struggle to achieve generation of $360^{\circ}$ -consistent and diverse hu + +![](images/b63c0190a8160edd8b52664efcf9fd9d1a08ff4ad460f476108674f28e460969.jpg) +Figure 2. Pipeline for constructing our HuGe100K. Diverse attribute combinations from GPT-4 templates create text prompts, generating synthetic images via FLUX, combined with real images from DeepFashion. SMPL-X fitting produces multi-view pose sequences with 360-degree rotations and diverse animatable motions. MVChamp then converts these sequences into multi-view images, ensuring 3D consistency in the dataset. + +man videos conditioned on pose sequences (e.g., 2D poses and 3D SMPL(-X) conditions). To address this limitation, we re-train a state-of-the-art video generative model, Champ [81], to obtain a multi-view consistent generative model, MVChamp. Firstly, We collect a curated dataset of over 100K single-person videos with various motions (e.g., dancing), including approximately 20K videos involving the action of turning around, to fine-tune the model and enhance its generalizability for diverse inputs and motions. Secondly, to further improve 3D consistency[16], we utilize scanned models from THuman 2.1 to generate uniformly distributed views via Blender. From these, we select a subset of 24 views that are evenly spaced around the full $360^{\circ}$ rotation to fine-tune MVChamp's temporal layers using diffusion loss. Lastly, to support high-quality whole-body animation (i.e., body and hands), we introduce precise SMPL-X estimation using NLF [54] for body shape and pose estimation. Additionally, we employ HaMeR [47] for hand estimation, rendering the corresponding hand template into depth maps as additional control signals. + +After these training processes, MVChamp generates multi-view images conditioned on SMPL-X parametric model. The pose condition is set to an "A-Pose" $80\%$ of the time and a random pose from dance videos (via SMPLer-X [4]) $20\%$ of the time, balancing pose variability. To address discrepancies between the first and last frames, we propose a Temporal Shift Denoising Strategy: during the denoising steps, we shift latent inputs and pose signals temporally, moving the last frame to the first. This improves continuity between frames without additional inference costs. Finally, FaceFusion [24] is utilized to enhance facial details. + +![](images/2a67f02d46c67091a06a7612b4d661065ab4fc4612a13f4de0af8ea336f3701a.jpg) +Figure 3. A paired example from the proposed HuGe100K Dataset. For each reference image, we generate a set of multiview images using an estimated shape and a specific pose. The figure shows the pose is well-aligned. + +# 3.3. Data Statistics and Characteristics + +Tab. 1 compares our dataset HuGe100K to other 3D human datasets regarding scale, diversity, and consistency. With over 100K human identities and 20K poses, HuGe100K offers balanced diversity across dimensions such as area, clothing, body shape, age, and gender (see Fig. 1). Each sample includes a reference image, SMPL-X estimates, 24 uniformly sampled multi-view images, camera parameters, and SMPL-X data (see Fig. 3). + +# 4. Large Human Reconstruction Model + +In this section, we present the large-scale human reconstruction model, named IDOL. The overview pipeline is shown in Fig. 4. In Sec. 4.1, we describe the animatable human representation. Sec. 4.2 details the network architecture of IDOL, while Sec. 4.3 explains how we trained the model in an end-to-end manner using the multi-view image dataset. + +# 4.1. Animatable Human Representation + +Similarly to previous work [28, 35, 74], IDOL leverages 3D Gaussian Splatting [33] in conjunction with SMPL-X + +![](images/5316acc26691809aea53c9dfca58977e8b50f661e586da5b397cbf6d510c720d.jpg) +Figure 4. The architecture of IDOL, a full-differentiable transformer-based framework for reconstructing animatable 3D human from a single image. The model integrates a high-resolution $(1024\times 1024)$ encoder [34] and fuses image tokens with learnable UV tokens through the UV-Alignment Transformer. A UV Decoder predicts Gaussian attribute maps as intermediate representations, capturing the human's geometry and appearance in a structured 2D UV space defined by the SMPL-X model. These maps, in conjunction with the SMPL-X model, represent a 3D human avatar in a canonical space, which can be animated using linear blend skinning (LBS). The model is optimized using multi-view images with diverse poses and identities, learning to disentangle pose, appearance, and shape. + +for 3D human representation, aiming to address the challenges of real-time rendering and accurate animation of human avatars. Specifically, each Gaussian primitive $\mathcal{G}_k$ is characterized by: $\mathcal{G}_k = \{\mu_k,\alpha_k,\mathbf{r}_k,\mathbf{s}_k,\mathbf{c}_k\}$ , where $\mu_{k}$ is the 3D position of the Gaussian, $\alpha_{k}$ is opacity, $\mathbf{r}_k$ is the rotation, $\mathbf{s}_k$ is the scale, and $\mathbf{c}_k$ is the color. + +3D Gaussian Human. Directly predicting all 3D Gaussian primitives is computationally intensive. Instead, IDOL leverages the predefined 2D UV space of the SMPL-X model to transform the 3D representation task into a more manageable 2D problem. Initially, IDOL predicts Gaussian attribute maps that encode the offset values $\{\delta_{\mu_k}, \delta_{\mathbf{r}_k}, \delta \mathbf{s}_k\}$ , as well as the color $\mathbf{c}_k$ and opacity $\alpha_k$ for each Gaussian primitive $\mathcal{G}_k$ . Similarly to $E^3 Gen$ [74], the position $\mu_k$ , scale $\mathbf{s}_k$ , and rotation $\mathbf{r}_k$ of each Gaussian primitive $\mathcal{G}_k$ is modeled relative to its SMPL-X vertex as follows: $\mu_k = \hat{\mu}_k + \delta_{\mu_k}$ , $\mathbf{s}_k = \hat{\mathbf{s}}_k \cdot \delta_{\mathbf{s}_k}$ , $\mathbf{r}_k = \hat{\mathbf{r}}_k \cdot \delta_{\mathbf{r}_k}$ , where $\hat{\mu}_k, \hat{\mathbf{s}}_k$ , and $\hat{\mathbf{r}}_k$ are the initial values based on the SMPL-X model. The color $\mathbf{c}_k$ is assigned using RGB values, and the opacity $\alpha_k$ is set to 1, indicating full opacity. This unified UV space significantly reduces computational complexity and fully exploits the geometric and semantic priors provided by the SMPL-X model. By modeling Gaussian primitives relative to SMPL-X vertices, IDOL ensures semantic consistency across corresponding body parts of diverse avatars, thereby enhancing the model's generalization capability across various reference images. + +Animation and Rendering. Given a target pose, we calculate the transformation of each human joint using predefined kinematic relationships. The transformation of each Gaussian primitive is performed using a forward skinning technique based on LBS. Specifically, the position of each Gaussian primitive $\mathcal{G}_k$ is transformed as follows: $\mu_k' = \sum_{i=1}^{n_b} w_i \mathbf{B}_i \mu_k$ . Additionally, the rotation matrix $\mathbf{R}_k$ is up- + +dated by: $\mathbf{R}_k^{\prime} = \mathbf{T}_k^{1:3,1:3}\mathbf{R}_k$ , where $\mathbf{T}_k = \sum_{i=1}^{n_b} w_i \mathbf{B}_i$ , and $\mathbf{R}_k$ is the rotation matrix representation of the rotation angle $\mathbf{r}_k$ . Here, $n_b$ is the number of joints, $\mathbf{B}_i$ is the transformation matrix for each joint, and $w_i$ represents the skinning weights, indicating the influence of each joint's motion on the Gaussian primitive's position $\mu_k$ . + +To estimate the skinning weights for each $\mathcal{G}_k$ , we first compute a body-part skinning field using a pre-defined low-resolution volume [6]. For smaller regions like the hands and face, which undergo minimal topological variation, weights are interpolated from the SMPL-X template via barycentric coordinates. This strategy enables modeling large topology changes across identities (e.g., from clothing) while stabilizing convergence in less variable areas such as fingers and facial regions. + +# 4.2. Network Structure + +As illustrated in Fig. 4, IDOL is a full-differentiable framework for reconstructing animatable 3D human. + +High-resolution Image Encoder. Higher resolution of the input image results in a high-quality reconstruction. However, previous works suffer from low-resolution ViT-based encoders, such as DINOv2 [45], which support the $448 \times 448$ resolution. To fully leverage the resolution of the HuGe100K dataset, we further adopt a high-resolution human foundation model, Sapiens [34], to encode the $1024 \times 1024$ resolution images into patch-wise feature tokens, formulated as: $\mathbf{F} = \{\mathbf{f}_i\}_{i=1}^n \in \mathbb{R}^{d_E}$ , where $i$ denotes the $i$ -th image patch, and $d_E$ is the channel length of each token. The Sapiens model is pretrained on 300 million in-the-wild human images using the Masked Autoencoder (MAE) framework [18], enabling it to excel in preserving fine-grained details and capturing diverse human poses and appearances, making it highly effective for high-resolution human image feature extraction. + +UV-Alignment Transformer. To map irregular and diverse reference images onto regular UV feature maps, a UV-Alignment Transformer is employed to align learnable spatial-positional UV tokens $\mathbf{Q}^0$ with reference image features $\mathbf{F}$ . Specifically, the UV-Alignment Transformer concatenates $\mathbf{F}$ from Spaiens with $\mathbf{Q}^0$ , and aggregates and refines features through $D$ transformer blocks, producing an enhanced representation $\mathbf{Q}^D$ . Each transformer block utilizes self-attention mechanisms, enabling the model to capture complex relationships among the input tokens and impute missing parts using correlated tokens. + +UV Decoder. The UV tokens $\mathbf{Q}^D$ are reshaped and decoded into Gaussian attribute maps, capturing the human's geometry and appearance within the structured 2D UV space as shown in Sec. 4.1. To preserve and enhance fine details in the decoded Gaussian attribute maps, a Convolutional Neural Network is employed for spatial up-sampling. + +# 4.3. Training Objectives + +IDOL reconstructs 3D human by predicting Gaussian attribute UV maps in a single forward pass, offering significant advantages in inference speed and enabling end-to-end training with multi-view images. For each sample, we select a front view as the reference image $\mathbf{I}_{\mathrm{ref}}$ along with a random set of generated multi-view images $\{\mathbf{I}_{\mathrm{gt},i}\}_{i = 1}^{N}$ and their corresponding SMPL-X parameters and camera for supervision. It takes $\mathbf{I}_{\mathrm{ref}}$ as input to generate the 3D human, using differentiable rendering to produce the multi-view images $\{\mathbf{I}_{\mathrm{pred},i}\}_{i = 1}^{N}$ . The loss function is defined as follows: + +$$ +\mathcal {L} = \sum_ {i = 1} ^ {N} \left(\left\| \mathbf {I} _ {\mathrm {g t}, i} - \mathbf {I} _ {\text {p r e d}, i} \right\| ^ {2} + \lambda L _ {v g g} \left(\mathbf {I} _ {\mathrm {g t}, i}, \mathbf {I} _ {\text {p r e d}, i}\right)\right), \tag {1} +$$ + +where $\lambda$ controls the balance between the mean square error loss and the perceptual loss $L_{vgg}$ . + +# 5. Experiments + +# 5.1. Experiments on Dataset Creation + +Fig. 5 (left) presents the qualitative results of the ablation study on our data generation method. Specifically, Fig. 5 (a) and Fig. 5 (b) demonstrate how hand refinement and face swapping effectively enhance the quality of the generated hands and faces. Fig. 5 (c) highlights the importance of fine-tuning the 3D dataset Thuman 2.1 for improving the 3D consistency of MVChamp, while Fig. 5 (d) illustrates that the Temporal Shift Denoising strategy, which involves cycling the first and last frames, improves the consistency between the first and last frames. + +Fig. 5 (right) compares our MVChamp's generation quality against previous human image animation models (e.g., Champ, MimicMotion [75]) and general multi-view + +video generation model (e.g., SV3D [58]). The issues of these methods are highlighted in the figure. Champ struggles to generate plausible hands and heads in multi-view scenarios. MimicMotion not only fails to produce realistic shoes but also has difficulty preserving the identity, especially for the anime image. SV3D produces low-quality multi-view human generation due to the lack of 3D prior knowledge of the human body. In contrast, our model generates consistent, high-quality results across views. + +# 5.2. Comparison of Reconstruction Model + +Implementation Details. For the encoder, we employ the pre-trained Sapiens-1B model [34] and keep its weights frozen during training. We then define a transformer-based framework with a parameter size of $0.5B$ to perform feature fusion. Additionally, we densify the SMPL-X vertex set by sampling approximately 200,000 vertices to map attributes and represent the 3D human model effectively. + +Dataset and Metrics. We train IDOL on a dataset consisting of generated multi-view images from HuGe100K and rendered images from THuman 2.1 [79]. For THuman 2.1, we render 72 view images at a resolution of $896 \times 640$ for each scan, with rendering views uniformly distributed. To evaluate performance quantitatively, we reserve the last 50 cases from both HuGe100K and THuman 2.1 as the testing set. We use the most frontal image as the reference image and the remaining images as ground-truth data for evaluation. The ground-truth camera parameters and SMPL-X parameters are provided for all methods. The renderings of the reconstructed model are then compared to their corresponding ground truth images with several metrics, including Mean Squared Error (MSE), Learned Perceptual Image Patch Similarity (LPIPS) [73], Peak Signal-to-Noise Ratio (PSNR). More details on implementation can be found in the supplementary material. + +Baselines. We compare IDOL with three baseline categories in the Single-Image Human Reconstruction task. The first category comprises methods based on loop optimization or pixel-alignment modules, including GTA [77] and SIFU [78]. These methods focus on refining human reconstruction through iterative optimization or predicting 3D geometry from pixel-aligned features. + +The second category encompasses large-scale generic reconstruction networks, represented by LGM [57]. These models are known for their ability to handle large datasets and their scalability, offering advantages in terms of fast reference speeds and large output resolutions. + +The third category involves optimization-based 3D generation methods using score distillation sampling (SDS), exemplified by DreamGaussian [56]. These methods leverage priors from 2D diffusion models to distill 3D objects. + +![](images/1d4ebec419a5575533de3618f4338060fb86aa35099fc797dde03cd381261967.jpg) +(c) THuman Finetune +(d) Temporal Shift Denoising Strategy + +![](images/a9518cfb5bb51df4681f3d9221f0396de4c46e0bb5e3635d484e4b7207615455.jpg) +Reference (e) MVChamp (ours) +(f) Champ (g) MimicMotion (h) SV3D + +![](images/16eace9fe2735034876bb1c22d34d06f18ee4bde603478020e53d7fc2f2d2dd7.jpg) +Figure 5. Qualitative results of our MVChamp ablation study (left) and comparison experiment (right). +Figure 6. Comparisons on (a) the upper: novel-view synthesis given a single image, and (b) the lower: our animated results. + +DreamGaussian accelerates convergence by progressively densifying 3D Gaussians, significantly reducing the reconstruction time. However, it still takes approximately two minutes to reconstruct a single object. + +Quantitative comparison. As shown in Tab 2, our method outperforms all baselines in all metrics. We attribute this superior performance to the large-scale dataset HuGe100K and the design of our large-scale reconstruc + +tion model IDOL, which allows for more effective training and improved synthesis of appearance results. Despite this, we note that SIFU and GTA report a lower metric than what we expected in our test settings. While we provide accurate ground-truth SMPL-X parameters and camera settings, the ideal orthographic projection required by the pixel-alignment modules in SIFU and GTA is not well-suited to our perspective projection model, where the camera focus ranges from 35 to 80. This mismatch leads to a + +
MethodMSE ↓PSNR ↑LPIPS ↓
SIFU [78]0.04214.2041.612
GTA [77]0.04114.2821.629
Ours-w/o HuGe100K0.01719.2251.326
Ours-full0.00821.6731.138
+ +Table 2. Evaluation of Comparison and Ablation Experiments. + +misalignment between the rendered images and the ground truth, adversely affecting their performance metrics. In fact, many real-life photographs are taken with medium to short focal lengths and cannot be approximated using orthographic projection, which is a less noticeable drawback of SIFU and GTA. Additionally, SIFU and GTA, trained with images at a resolution of $512 \times 512$ , struggle to synthesize detailed textures, particularly in the invisible areas of the reference images. This limitation is primarily due to the lack of comprehensiveness and diversity in their training datasets, which restricts their performance in generating invisible aspects of the images. + +Qualitative comparison. We perform a qualitative evaluation on an in-the-wild dataset with methods from [56, 57, 77, 78], with the results presented in Fig. 6. Our evaluation includes a subject in a complex outfit featuring a baseball cap and textured clothing, demonstrating our method's ability to synthesize intricate textures and handle loose outfits across different views. Additional tests include out-of-domain cartoon data and large-angle side views and assessing model adaptability and viewpoint handling. IDOL consistently outperforms the baselines, which struggle with detail reproduction and texture consistency under these various conditions. + +# 5.3. Ablation Study on Reconstruction Model + +We assess the impact of proposed components by removing them individually from IDOL. As shown in Fig. 7, replacing the Sapiens [34] encoder with DINO v2 [45] (w/o Sapiens) reduces detail quality, resulting in less realistic textures and folds. Excluding the HuGe100K dataset (w/o HuGe100K) causes significant distortion, including blurred details and color bleeding. The complete model, with all components, produces the most realistic and detailed results. These findings highlight the critical roles of both the Sapiens encoder and the HuGe100K dataset in achieving high-quality avatar generation. + +# 5.4. Downstream Applications + +IDOL reconstructs an avatar by combining Gaussian attribute maps with the SMPL-X model, enabling users to modify the avatar's appearance by editing UV texture maps and to control body shape by adjusting SMPL-X shape pa + +![](images/9d1dabd4ba7c55ce5f9bce4176e87bea8116872ba262e5598207f50a43537c1c.jpg) +Figure 7. Qualitative Results of Ablation Study of IDOL. + +![](images/0b0b6ed44969dd72760d24f32e2fdbbe9ce1ec8c04af4208fba8f74e6049a4bb.jpg) +Figure 8. Controllable Avatar Editing: (a) texture editing; (b) body shape editing. + +rameters. Fig. 8(a) illustrates the effect of texture editing on clothing patterns, while Fig. 8(b) demonstrates shape editing to adjust the avatar's body size. This approach provides a high degree of controllability over both the avatar's appearance and body shape. + +# 6. Conclusions and Limitations + +In conclusion, our work has made significant strides in creating an animatable 3D human from a single image. We introduced a scalable pipeline for training a simple yet efficient feed-forward model, incorporating a dataset generation framework, a large-scale dataset $HuGe100K$ , and a scalable reconstruction transformer model, IDOL. This model efficiently reconstructs photorealistic 3D humans in under a second and demonstrates versatility across various applications. This model can efficiently reconstruct photorealistic humans in less than a second and is versatile enough to support various applications. + +However, there are limitations. Due to the constraints of the video model used, we can only synthesize single-frame images from fixed viewpoints. Future work could consider generating longer motion sequences. The focus on body reconstruction and animation leaves facial optimization secondary, and the architecture lacks specific design for facial identity or expression. Additionally, handling half-body inputs remains challenging, and improvements in data generation strategies are needed to enhance performance. + +# Acknowledgments + +This work was partly supported by the National Key Research and Development Program of China (2022YFF0902400), the Shenzhen Science and Technology Program (JCYJ20220818101014030), and the research fund of Tsinghua University - Tencent Joint Laboratory for Internet Innovation Technology. + +# References + +[1] Badour AlBahar, Shunsuke Saito, Hung-Yu Tseng, Changil Kim, Johannes Kopf, and Jia-Bin Huang. Single-image 3d human digitization with shape-guided diffusion. In SIGGRAPH Asia 2023 Conference Papers, 2023. 2, 3 +[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2 +[3] Zhongang Cai, Daxuan Ren, Ailing Zeng, Zhengyu Lin, Tao Yu, Wenjia Wang, Xiangyu Fan, Yang Gao, Yifan Yu, Liang Pan, et al. Humman: Multi-modal 4d human dataset for versatile sensing and modeling. In European Conference on Computer Vision, pages 557-577. Springer, 2022. 2, 3 +[4] Zhongang Cai, Wanqi Yin, Ailing Zeng, Chen Wei, Qingping Sun, Wang Yanjun, Hui En Pang, Haiyi Mei, Mingyuan Zhang, Lei Zhang, et al. Smpler-x: Scaling up expressive human pose and shape estimation. Advances in Neural Information Processing Systems, 36, 2024. 4 +[5] Di Chang, Yichun Shi, Quankai Gao, Hongyi Xu, Jessica Fu, Guoxian Song, Qing Yan, Yizhe Zhu, Xiao Yang, and Mohammad Soleymani. Magicpose: Realistic human poses and facial expressions retargeting with identity-aware diffusion. In Forty-first International Conference on Machine Learning, 2023. 3 +[6] Xu Chen, Tianjian Jiang, Jie Song, Max Rietmann, Andreas Geiger, Michael J Black, and Otmar Hilliges. Fast-snarf: A fast deformer for articulated neural fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(10): 11796-11809, 2023. 5 +[7] Wei Cheng, Su Xu, Jingtan Piao, Chen Qian, Wayne Wu, Kwan-Yee Lin, and Hongsheng Li. Generalizable neural performer: Learning robust radiance fields for human novel view synthesis. arXiv preprint arXiv:2204.11798, 2022. 2, 3 +[8] Wei Cheng, Ruixiang Chen, Siming Fan, Wanqi Yin, Keyu Chen, Zhongang Cai, Jingbo Wang, Yang Gao, Zhengming Yu, Zhengyu Lin, et al. Dna-rendering: A diverse neural actor repository for high-fidelity human-centric rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19982-19993, 2023. 2, 3 +[9] Enric Corona, Mihai Zanfir, Thiemo Alldieck, Eduard Gabriel Bazavan, Andrei Zanfir, and Cristian Sminchisescu. Structured 3d features for reconstructing controllable avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16954-16964, 2023. 3 + +[10] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. 2, 3 +[11] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-xl: A universe of $10\mathrm{m} + 3\mathrm{d}$ objects. Advances in Neural Information Processing Systems, 36, 2024. 2, 3 +[12] Qiao Feng, Yebin Liu, Yu-Kun Lai, Jingyu Yang, and Kun Li. Fof: Learning fourier occupancy field for monocular real-time human reconstruction. In NeurIPS, 2022. 2 +[13] Xiangjun Gao, Jiaolong Yang, Jongyoo Kim, Sida Peng, Zicheng Liu, and Xin Tong. Mps-nerf: Generalizable 3d human rendering from multiview images. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. +[14] Xiangjun Gao, Xiaoyu Li, Chaopeng Zhang, Qi Zhang, Yanpei Cao, Ying Shan, and Long Quan. Context-human: Free-view rendering of human from a single image with texture-consistent synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10084-10094, 2024. 2 +[15] Xiangjun Gao, Xiaoyu Li, Yiyu Zhuang, Qi Zhang, Wenbo Hu, Chaopeng Zhang, Yao Yao, Ying Shan, and Long Quan. Mani-gs: Gaussian splatting manipulation with triangular mesh. arXiv preprint arXiv:2405.17811, 2024. 2 +[16] Junlin Han, Filippos Kokkinos, and Philip Torr. Vfusion3d: Learning scalable 3d generative models from video diffusion models. In European Conference on Computer Vision, pages 333-350. Springer, 2025. 4 +[17] Sang-Hun Han, Min-Gyu Park, Ju Hong Yoon, Ju-Mi Kang, Young-Jae Park, and Hae-Gon Jeon. High-fidelity 3d human digitization from single 2k resolution images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12869-12879, 2023. 2, 3 +[18] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 5 +[19] Tong He, Yuanlu Xu, Shunsuke Saito, Stefano Soatto, and Tony Tung. Arch++: Animation-ready clothed human reconstruction revisited. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11046-11056, 2021. 3 +[20] Xu He, Xiaoyu Li, Di Kang, Jiangnan Ye, Chaopeng Zhang, Liyang Chen, Xiangjun Gao, Han Zhang, Zhiyong Wu, and Haolin Zhuang. Magicman: Generative novel view synthesis of humans with 3d-aware diffusion and iterative refinement. arXiv preprint arXiv:2408.14211, 2024. 2 +[21] Yuxiao He, Yiyu Zhuang, Yanwen Wang, Yao Yao, Siyu Zhu, Xiaoyu Li, Qi Zhang, Xun Cao, and Hao Zhu. Head360: Learning a parametric 3d full-head for free-view synthesis in 360. In European Conference on Computer Vision, pages 254-272. Springer, 2024. 2 + +[22] Yicong Hong, Kai Zhang, Juixiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400, 2023. 2 +[23] https://github.com/black-forest labs/flux. Flux latent rectified flow transformers, 2024. 3 +[24] https://github.com/facefusion/facefusion. Facefusion, 2024. 4 +[25] https://renderpeople.com/3d people. Renderpeople dataset, 2015. 2 +[26] https://web.twindom.com/. Twindom dataset, 2020. 2 +[27] Li Hu. Animate anyone: Consistent and controllable image-to-video synthesis for character animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8153-8163, 2024. 3 +[28] Liangxiao Hu, Hongwen Zhang, Yuxiang Zhang, Boyao Zhou, Boning Liu, Shengping Zhang, and Liqiang Nie. Gaussianavatar: Towards realistic human avatar modeling from a single video via animatable 3d gaussians. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 634-644, 2024. 2, 4 +[29] Shoukang Hu, Fangzhou Hong, Liang Pan, Haiyi Mei, Lei Yang, and Ziwei Liu. Sherf: Generalizable human nerf from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 3 +[30] Yangyi Huang, Hongwei Yi, Weiyang Liu, Haofan Wang, Boxi Wu, Wenxiao Wang, Binbin Lin, Debing Zhang, and Deng Cai. One-shot implicit animatable avatars with model-based priors. In IEEE Conference on Computer Vision (ICCV), 2023. 3 +[31] Zeng Huang, Yuanlu Xu, Christoph Lassner, Hao Li, and Tony Tung. Arch: Animatable reconstruction of clothed humans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3093-3102, 2020. 2, 3 +[32] Mustafa Isik, Martin Rünz, Markos Georgopoulos, Taras Khakhulin, Jonathan Starck, Lourdes Agapito, and Matthias Nießner. Humanrf: High-fidelity neural radiance fields for humans in motion. ACM Transactions on Graphics (TOG), 42(4):1-12, 2023. 2, 3 +[33] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 2, 3, 4 +[34] Rawal Khirodkar, Timur Bagautdinov, Julieta Martinez, Su Zhaoen, Austin James, Peter Selednik, Stuart Anderson, and Shunsuke Saito. Sapiens: Foundation for human vision models. In European Conference on Computer Vision, pages 206-228. Springer, 2025. 2, 5, 6, 8 +[35] Tobias Kirschstein, Simon Giebenhain, Jiapeng Tang, Markos Georgopoulos, and Matthias Nießner. Gghead: Fast and generalizable 3d gaussian heads. arXiv preprint arXiv:2406.09377, 2024. 4 +[36] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model. + +In International Conference on Learning Representations (ICLR), 2024. 2 +[37] Peng Li, Wangguandong Zheng, Yuan Liu, Tao Yu, Yangguang Li, Xingqun Qi, Mengfei Li, Xiaowei Chi, Siyu Xia, Wei Xue, et al. Pshuman: Photorealistic single-view human reconstruction using cross-scale diffusion. arXiv preprint arXiv:2409.10141, 2024. 2 +[38] Ruilong Li, Shan Yang, David A Ross, and Angjoo Kanazawa. Ai choreographer: Music conditioned 3d dance generation with aist++. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13401-13412, 2021. 2, 3 +[39] Zhe Li, Zerong Zheng, Lizhen Wang, and Yebin Liu. Animatable gaussians: Learning pose-dependent gaussian maps for high-fidelity human avatar modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19711-19722, 2024. 2 +[40] Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural actor: Neural free-view synthesis of human actors with pose control. ACM transactions on graphics (TOG), 40(6):1-16, 2021. 2, 3 +[41] Ziwei Liu, Ping Luo, Shi Qiu, Xiaogang Wang, and Xiaou Tang. Deepfashion: Powering robust clothes recognition and retrieval with rich annotations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1096-1104, 2016. 3 +[42] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. ACM Transactions on Graphics, 34(6), 2015. 2 +[43] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models. Advances in Neural Information Processing Systems, 36, 2024. 2 +[44] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2 +[45] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 5, 8 +[46] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10975-10985, 2019. 2 +[47] Georgios Pavlakos, Dandan Shan, Ilija Radosavovic, Angjoo Kanazawa, David Fouhey, and Jitendra Malik. Reconstructing hands in 3d with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9826-9836, 2024. 4 +[48] Hao-Yang Peng, Jia-Peng Zhang, Meng-Hao Guo, Yan-Pei Cao, and Shi-Min Hu. Charactergen: Efficient 3d character + +generation from single images with multi-view pose canonicalization. ACM Transactions on Graphics (TOG), 2024. 2, 3 +[49] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9054-9063, 2021. 2, 3 +[50] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2 +[51] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2022. 2 +[52] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2304-2314, 2019. 2, 3 +[53] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 84-93, 2020. 2, 3 +[54] István Sárándi and Gerard Pons-Moll. Neural localizer fields for continuous 3d human pose and shape estimation. arXiv preprint arXiv:2407.07532, 2024. 4 +[55] Kaiyue Shen, Chen Guo, Manuel Kaufmann, Juan Jose Zarate, Julien Valentin, Jie Song, and Otmar Hilliges. X-avatar: Expressive human avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16911-16921, 2023. 3 +[56] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023. 6, 8 +[57] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. In European Conference on Computer Vision, pages 1-18. Springer, 2025. 2, 6, 8 +[58] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent video diffusion. In European Conference on Computer Vision, pages 439-457. Springer, 2025. 2, 6 +[59] Qilin Wang, Zhengkai Jiang, Chengming Xu, Jiangning Zhang, Yabiao Wang, Xinyi Zhang, Yun Cao, Weijian Cao, Chengjie Wang, and Yanwei Fu. Vividpose: Advancing stable video diffusion for realistic human image animation. arXiv preprint arXiv:2405.18156, 2024. 3 +[60] Zhenzhen Weng, Jingyuan Liu, Hao Tan, Zhan Xu, Yang Zhou, Serena Yeung-Levy, and Jimei Yang. Template-free + +single-view 3d human digitalization with diffusion-guided lrm. arXiv preprint arXiv:2401.12175, 2024. 2, 3 +[61] Kailu Wu, Fangfu Liu, Zhihan Cai, Runjie Yan, Hanyang Wang, Yating Hu, Yueqi Duan, and Kaisheng Ma. Unique3d: High-quality and efficient 3d mesh generation from a single image. arXiv preprint arXiv:2405.20343, 2024. 2 +[62] Menghua Wu, Hao Zhu, Linjia Huang, Yiyu Zhuang, Yuanxun Lu, and Xun Cao. High-fidelity 3d face generation from natural language descriptions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4521-4530, 2023. 2 +[63] Zhangyang Xiong, Chenghong Li, Kenkun Liu, Hongjie Liao, Jianqiao Hu, Junyi Zhu, Shuliang Ning, Lingteng Qiu, Chongjie Wang, Shijie Wang, et al. Mvhumannet: A large-scale dataset of multi-view daily dressing human captures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19801-19811, 2024. 2, 3 +[64] Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J Black. Icon: Implicit clothed humans obtained from normals. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13286-13296. IEEE, 2022. 2, 3 +[65] Yuliang Xiu, Jinlong Yang, Xu Cao, Dimitrios Tzionas, and Michael J. Black. ECON: Explicit Clothed humans Optimized via Normal integration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3 +[66] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191, 2024. 2 +[67] Zhongcong Xu, Jianfeng Zhang, Jun Hao Liew, Hanshu Yan, Jia-Wei Liu, Chenxu Zhang, Jiashi Feng, and Mike Zheng Shou. Magicanimate: Temporally consistent human image animation using diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1481–1490, 2024. 3 +[68] Tao Yu, Zerong Zheng, Kaiwen Guo, Pengpeng Liu, Qionghai Dai, and Yebin Liu. Function4d: Real-time human volumetric capture from very sparse consumer rgbd sensors. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5746-5756, 2021. 2, 3 +[69] Wangbo Yu, Li Yuan, Yan-Pei Cao, Xiangjun Gao, Xiaoyu Li, Wenbo Hu, Long Quan, Ying Shan, and Yonghong Tian. Hifi-123: Towards high-fidelity one image to 3d content generation. In European Conference on Computer Vision, pages 258–274. Springer, 2024. 2 +[70] Xianggang Yu, Mutian Xu, Yidan Zhang, Haolin Liu, Chongjie Ye, Yushuang Wu, Zizheng Yan, Chenming Zhu, Zhangyang Xiong, Tianyou Liang, et al. Mvmgnet: A large-scale dataset of multi-view images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9150-9161, 2023. 3 +[71] Zhixuan Yu, Jae Shin Yoon, In Kyu Lee, Prashanth Venkatesh, Jaesik Park, Jihun Yu, and Hyun Soo Park. + +Humbi: A large multiview dataset of human body expressions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2990-3000, 2020. 2, 3 +[72] Jingbo Zhang, Xiaoyu Li, Qi Zhang, Yanpei Cao, Ying Shan, and Jing Liao. Humanref: Single image to 3d human generation via reference-guided diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1844–1854, 2024. 2 +[73] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6 +[74] Weitian Zhang, Yichao Yan, Yunhui Liu, Xingdong Sheng, and Xiaokang Yang. E3gen: Efficient, expressive and editable avatars generation. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 6860-6869, 2024. 3, 4, 5 +[75] Yang Zhang, Jiaxi Gu, Li-Wen Wang, Han Wang, Junqi Cheng, Yuefeng Zhu, and Fangyuan Zou. Mimic motion: High-quality human motion video generation with confidence-aware pose guidance. arXiv preprint arXiv:2406.19680, 2024.3, 6 +[76] Zechuan Zhang, Li Sun, Zongxin Yang, Ling Chen, and Yi Yang. Global-correlated 3d-decoupling transformer for clothed avatar reconstruction. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 3 +[77] Zechuan Zhang, Li Sun, Zongxin Yang, Ling Chen, and Yi Yang. Global-correlated 3d-decoupling transformer for clothed avatar reconstruction. Advances in Neural Information Processing Systems, 36, 2024. 3, 6, 8 +[78] Zechuan Zhang, Zongxin Yang, and Yi Yang. Sifu: Side-view conditioned implicit function for real-world usable clothed human reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9936-9947, 2024. 2, 3, 6, 8 +[79] Zerong Zheng, Tao Yu, Yixuan Wei, Qionghai Dai, and Yebin Liu. Deephuman: 3d human reconstruction from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7739-7749, 2019. 2, 3, 6 +[80] Zerong Zheng, Tao Yu, Yebin Liu, and Qionghai Dai. Pamir: Parametric model-conditioned implicit representation for image-based human reconstruction. IEEE transactions on pattern analysis and machine intelligence, 44(6): 3170-3184, 2021. 2, 3 +[81] Shenhao Zhu, Junming Leo Chen, Zuozhuo Dai, Yinghui Xu, Xun Cao, Yao Yao, Hao Zhu, and Siyu Zhu. Champ: Controllable and consistent human image animation with 3d parametric guidance. In European Conference on Computer Vision (ECCV), 2024. 3, 4 +[82] Yiyu Zhuang, Hao Zhu, Xusen Sun, and Xun Cao. Mofanerf: Morphable facial neural radiance field. In European conference on computer vision, pages 268-285. Springer, 2022. 2 +[83] Yiyu Zhuang, Qi Zhang, Ying Feng, Hao Zhu, Yao Yao, Xiaoyu Li, Yan-Pei Cao, Ying Shan, and Xun Cao. Anti-aliased + +neural implicit surfaces with encoding level of detail. In SIGGRAPH Asia 2023 Conference Papers, pages 1-10, 2023. 2 +[84] Yiyu Zhuang, Qi Zhang, Xuan Wang, Hao Zhu, Ying Feng, Xiaoyu Li, Ying Shan, and Xun Cao. Neai: A pre-convoluted representation for plug-and-play neural ambient illumination. arXiv preprint arXiv:2304.08757, 2023. 2 +[85] Yiyu Zhuang, Yuxiao He, Jiawei Zhang, Yanwen Wang, Ji-ahe Zhu, Yao Yao, Siyu Zhu, Xun Cao, and Hao Zhu. Towards native generative model for 3d head avatar. arXiv preprint arXiv:2410.01226, 2024. 2 \ No newline at end of file diff --git a/2025/IDOL_ Instant Photorealistic 3D Human Creation from a Single Image/images.zip b/2025/IDOL_ Instant Photorealistic 3D Human Creation from a Single Image/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d75bde537b9e74bc1858b4a2f115d21aee3a95e0 --- /dev/null +++ b/2025/IDOL_ Instant Photorealistic 3D Human Creation from a Single Image/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b9a75b8211eaf1150f3519a3c66389eda361926be02d0f530f3e7174c83d7ef4 +size 805257 diff --git a/2025/IDOL_ Instant Photorealistic 3D Human Creation from a Single Image/layout.json b/2025/IDOL_ Instant Photorealistic 3D Human Creation from a Single Image/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..96a3de41bf569d50b1389855bfe06bcf6928c052 --- /dev/null +++ b/2025/IDOL_ Instant Photorealistic 3D Human Creation from a Single Image/layout.json @@ -0,0 +1,8869 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 87, + 103, + 523, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 103, + 523, + 121 + ], + "spans": [ + { + "bbox": [ + 87, + 103, + 523, + 121 + ], + "type": "text", + "content": "IDOL: Instant Photorealistic 3D Human Creation from a Single Image" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 82, + 141, + 529, + 212 + ], + "blocks": [ + { + "bbox": [ + 82, + 141, + 529, + 212 + ], + "lines": [ + { + "bbox": [ + 82, + 141, + 529, + 212 + ], + "spans": [ + { + "bbox": [ + 82, + 141, + 529, + 212 + ], + "type": "image", + "image_path": "6f12e6f0b256d25f38b2c0b2d88f71f6531892c93090f89911bd4113a582a57a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 91, + 213, + 243, + 333 + ], + "blocks": [ + { + "bbox": [ + 91, + 213, + 243, + 333 + ], + "lines": [ + { + "bbox": [ + 91, + 213, + 243, + 333 + ], + "spans": [ + { + "bbox": [ + 91, + 213, + 243, + 333 + ], + "type": "image", + "image_path": "ab4b622cbe0ce6d245a6a7add80b6aa698da8bd180fcf5f33ee7c8bcf2b3c61c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 246, + 213, + 515, + 333 + ], + "blocks": [ + { + "bbox": [ + 246, + 213, + 515, + 333 + ], + "lines": [ + { + "bbox": [ + 246, + 213, + 515, + 333 + ], + "spans": [ + { + "bbox": [ + 246, + 213, + 515, + 333 + ], + "type": "image", + "image_path": "319ff23135b5cb69b47d31d4e1377b0d9cd97bc2182700ac5a93d7d1ea3c1e24.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 91, + 335, + 243, + 454 + ], + "blocks": [ + { + "bbox": [ + 91, + 335, + 243, + 454 + ], + "lines": [ + { + "bbox": [ + 91, + 335, + 243, + 454 + ], + "spans": [ + { + "bbox": [ + 91, + 335, + 243, + 454 + ], + "type": "image", + "image_path": "a5afe8c7f64f696cf13ea57d102b7db378bc0ccb311228de9aacaee30d73431f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 455, + 555, + 499 + ], + "lines": [ + { + "bbox": [ + 54, + 455, + 555, + 499 + ], + "spans": [ + { + "bbox": [ + 54, + 455, + 555, + 499 + ], + "type": "text", + "content": "Figure 1. This work introduces (a) IDOL, a feed-forward, single-image human reconstruction framework that is fast, high-fidelity, and generalizable; (b) Utilizing the proposed Large Generated Human Multi-View Dataset consisting of " + }, + { + "bbox": [ + 54, + 455, + 555, + 499 + ], + "type": "inline_equation", + "content": "100K" + }, + { + "bbox": [ + 54, + 455, + 555, + 499 + ], + "type": "text", + "content": " multi-view subjects, our method exhibits exceptional generalizability in handling diverse human shapes, cross-domain data, severe viewpoints, and occlusions; (c) With a uniform structured representation, the avatars can be directly animatable and easily editable." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 246, + 335, + 515, + 454 + ], + "blocks": [ + { + "bbox": [ + 246, + 335, + 515, + 454 + ], + "lines": [ + { + "bbox": [ + 246, + 335, + 515, + 454 + ], + "spans": [ + { + "bbox": [ + 246, + 335, + 515, + 454 + ], + "type": "image", + "image_path": "8c58ee58915cc674be7d147ab614c31b5a845056ff29f23e7478b2d9ebe13f1a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 151, + 509, + 200, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 509, + 200, + 521 + ], + "spans": [ + { + "bbox": [ + 151, + 509, + 200, + 521 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 534, + 297, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 534, + 297, + 666 + ], + "spans": [ + { + "bbox": [ + 54, + 534, + 297, + 666 + ], + "type": "text", + "content": "Creating a high-fidelity, animatable 3D full-body avatar from a single image is a challenging task due to the diverse appearance and poses of humans and the limited availability of high-quality training data. To achieve fast and high-quality human reconstruction, this work rethinks the task from the perspectives of dataset, model, and representation. First, we introduce a large-scale HUmancentric GEnerated dataset, HuGe100K, consisting of 100K diverse, photorealistic sets of human images. Each set contains 24-view frames in specific human poses, generated using a pose-controllable image-to-multi-view model." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 510, + 556, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 666 + ], + "type": "text", + "content": "Next, leveraging the diversity in views, poses, and appearances within HuGe100K, we develop a scalable feedforward transformer model to predict a 3D human Gaussian representation in a uniform space from a given human image. This model is trained to disentangle human pose, body shape, clothing geometry, and texture. The estimated Gaussians can be animated without post-processing. We conduct comprehensive experiments to validate the effectiveness of the proposed dataset and method. Our model demonstrates the ability to efficiently reconstruct photorealistic humans at 1K resolution from a single input image using a single GPU instantly. Additionally, it seamlessly supports various applications, as well as shape and texture editing tasks." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 673, + 295, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 673, + 295, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 673, + 295, + 693 + ], + "type": "text", + "content": "Work done during the internship at Tencent by Yiyu Zhuang, Jiaxi Lv, and Wenhao." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 71, + 693, + 141, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 693, + 141, + 703 + ], + "spans": [ + { + "bbox": [ + 71, + 693, + 141, + 703 + ], + "type": "text", + "content": "*Equal contributions" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 71, + 703, + 149, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 703, + 149, + 712 + ], + "spans": [ + { + "bbox": [ + 71, + 703, + 149, + 712 + ], + "type": "text", + "content": "†Corresponding authors" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26308" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 136, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 136, + 84 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 136, + 84 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 305 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 305 + ], + "type": "text", + "content": "Reconstructing 3D clothed avatars from a single image is crucial in user-friendly virtual reality, gaming, and 3D content creation. This task involves mapping 2D images to 3D models, a highly ill-posed and challenging problem due to the complexity of clothing and the diversity of human poses. Learning-based methods trained on publicly available 3D human models [17, 25, 26, 63, 79] have improved reconstruction quality, but their generalization and quality remain limited [52, 53, 64, 65, 72, 78, 80]. Techniques that integrate parametric body estimation [31, 64, 80], loop optimization [64, 65, 78], and diffusion model priors [1, 37, 60] enhance performance but often result in slower and fragile reconstructions. Recently, large generic image-to-3D reconstruction and generation models [22, 36, 57, 58, 61, 66], leveraging large-scale datasets [10, 11, 43] or pre-trained diffusion models [2, 50, 51], boost capabilities but still struggle with real-life human reconstruction due to the scarcity of photorealistic 3D human data." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 307, + 296, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 307, + 296, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 307, + 296, + 641 + ], + "type": "text", + "content": "Back to the first principle, we rethink existing data, models, and representations. First, current data acquisition is limited by the photographing or scanning of individual subjects, making them inherently unscalable. For instance, the largest publicly available human dataset, MVHumanNet [63], contains only thousands of subjects and clothing variations, far from the dataset scale required to achieve robust model generalization across diverse input images. Second, in the realm of animatable reconstruction models, existing approaches predominantly rely on either reconstruct- and-rig methods that necessitate manual post-processing or entangle human parametric models (e.g., SMPL(-X) [42, 46]) for optimization. However, disentangling human pose and body shape, clothed geometry, and texture could simplify the learning process, avoid error accumulation in parameter estimation, and improve efficiency. Furthermore, when focusing solely on human reconstruction, the incorporation of multi-view image generation or refinement models warrants careful consideration [48, 60, 66]. While these models can introduce finer details, they may also lead to inconsistencies across different views. Third, the community of 3D vision develops various representations tailored to application-specific requirements and technological advancements [33, 44]. For human-centric applications, ideal representations should be well-structured and expressive to facilitate easy rigging and editing [28, 39], as well as capable of high-resolution and fast rendering [33] to enhance both functionality and realism." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "In this work, we introduce a scalable pipeline for training a simple yet efficient feed-forward model for instant photorealistic human reconstruction. We present HuGe100K, a large-scale dataset comprising over 2.4M high-resolution " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "(896\\times 640)" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": " multi-view images of 100K diverse subjects. To facilitate comprehensive 3D human reconstruction, we de" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "type": "text", + "content": "velop a scalable data creation pipeline that integrates synthetic and real-world data, ensuring a wide range of attributes such as age, shape, clothing, race, and gender. Building upon this dataset, we introduce a novel feedforward transformer model IDOL that efficiently predicts 3D human avatars with photorealistic textures and accurate geometry. Our approach leverages a pretrained encoder [34] and a transformer-based backbone for feature extraction and fusion, enabling instant reconstruction (within 1 second on an A100 GPU) without relying on generative models [78] for refinement. By integrating a uniform representation for the 3D human, our model achieves enhanced texture completion and ensures that the reconstructed humans are naturally animatable. Extensive evaluations demonstrate that our method excels in diverse and challenging scenarios, offering superior consistency and quality compared to existing techniques. Additionally, the efficient representation of our model enables seamless applications in animation, editing, and other downstream tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 325, + 300, + 525, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 300, + 525, + 312 + ], + "spans": [ + { + "bbox": [ + 325, + 300, + 525, + 312 + ], + "type": "text", + "content": "Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 316, + 555, + 471 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 316, + 555, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 316, + 555, + 374 + ], + "spans": [ + { + "bbox": [ + 314, + 316, + 555, + 374 + ], + "type": "text", + "content": "- We rethink single-view 3D human reconstruction from the perspectives of data, model, and representation. We introduce a scalable pipeline for training a simple yet efficient feed-forward model, named IDOL, for instant photorealistic human reconstruction." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 376, + 554, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 376, + 554, + 422 + ], + "spans": [ + { + "bbox": [ + 314, + 376, + 554, + 422 + ], + "type": "text", + "content": "- We develop a data generation pipeline and present HuGe100K, a large-scale multi-view human dataset featuring diverse attributes, high-fidelity, high-resolution appearances, and a well-aligned SMPL-X model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 423, + 554, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 423, + 554, + 471 + ], + "spans": [ + { + "bbox": [ + 314, + 423, + 554, + 471 + ], + "type": "text", + "content": "- Our comprehensive study demonstrates that leveraging large-scale generated training data significantly enhances model performance and generalizability, paving the way for further scaling up 3D human reconstruction models." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 484, + 400, + 497 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 484, + 400, + 497 + ], + "spans": [ + { + "bbox": [ + 314, + 484, + 400, + 497 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 505, + 555, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 505, + 555, + 565 + ], + "spans": [ + { + "bbox": [ + 313, + 505, + 555, + 565 + ], + "type": "text", + "content": "Recent advancements in 3D avatar reconstruction [12-14, 20, 69, 82] and representation learning [15, 62, 83, 84] have greatly improved human modeling. Most relevant to our work are recent efforts on 3D human datasets and single-image human reconstruction." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 574, + 430, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 574, + 430, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 574, + 430, + 586 + ], + "type": "text", + "content": "2.1. 3D Human Datasets" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "High-precision 3D human models typically rely on scanning or multi-view camera acquisition systems. Scan-based datasets, such as THuman2.0 [68], Twindom [26], and 2K2K [17], provide high-fidelity 3D geometries but are limited by a small number of subjects, simple standing poses, and non-human artifacts (e.g., 2K2K). In contrast, multiview acquisition systems [3, 7, 8, 21, 32, 38, 40, 49, 63, 71, 85] facilitate the collection of larger datasets with more flexible actions. However, these datasets often face challenges such as biased indoor lighting, fixed viewpoints, and limited" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26309" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": "scale. Large-scale synthetic and real datasets for generic objects, including Objaverse and MVImgNet [10, 11, 70], address open-world reconstruction but lack specificity for human models. To overcome these limitations, the proposed HuGe100K significantly scales up dataset size, increasing subject diversity by over 100 times compared to previous datasets. It emphasizes diversity in pose, shape, viewpoint, and clothing, paving the way for large-scale model training. We compare existing datasets with ours in Tab. 1." + } + ] + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 65, + 190, + 287, + 357 + ], + "blocks": [ + { + "bbox": [ + 65, + 190, + 287, + 357 + ], + "lines": [ + { + "bbox": [ + 65, + 190, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 65, + 190, + 287, + 357 + ], + "type": "table", + "html": "
TypeDataset#FramesIDsSMPL(-X)
3D ScansTHuman [79]-200✓
THuman2.1 [68]-2500✓
2K2K [17]-2050✗
X-Avatar [55]35.5K20✓
Multi-view ImagesZJU-MoCap [49]180K10✓
Neural Actor [40]250K8✓
HUMBI [71]26M772✓
AIST++ [38]10.1M30✓
HuMMan [3]60M1000✓
GeneBody [7]2.95M50✓
ActorsHQ [32]40K8✗
DNA-Rendering [8]67.5M500✓
MVHumanNet [63]645.1M4500✓
OursHuGe100K>2.4M100K✓
", + "image_path": "a32e0df455487a09ef1f8601d74b80450735b788db95a8b2941990424e74c4d2.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 365, + 295, + 399 + ], + "lines": [ + { + "bbox": [ + 55, + 365, + 295, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 295, + 399 + ], + "type": "text", + "content": "Table 1. Comparisons of related datasets. HuGe100K is a large-scale generated multi-view human dataset with " + }, + { + "bbox": [ + 55, + 365, + 295, + 399 + ], + "type": "inline_equation", + "content": "100K" + }, + { + "bbox": [ + 55, + 365, + 295, + 399 + ], + "type": "text", + "content": " diverse high-fidelity humans." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 420, + 251, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 420, + 251, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 420, + 251, + 433 + ], + "type": "text", + "content": "2.2. Single-Image Human Reconstruction" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 437, + 295, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 295, + 701 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 295, + 701 + ], + "type": "text", + "content": "For clothed 3D reconstruction methods, implicit representation methods such as PIFU [52], PIFU-HD [53], ARCH [19, 31], and PaMIR [80] have been widely adopted. To enhance reconstruction quality and generalizability, loop optimization techniques integrate implicit representations with explicit or hybrid human priors for improved robustness [1, 64, 65, 76]. Similarly, GTA [77] and SIFU [78] entangle SMPL priors and side-view conditioned features to enhance feature representation. However, these approaches rely heavily on the accuracy of SMPL estimates and are computationally expensive, typically requiring several minutes for inference. Moreover, the resulting 3D representations are not drivable. Recent advancements focus on recovering animatable 3D humans from single images [9, 29, 30]. Some methods benefit from the pre-trained diffusion models or large reconstruction models [48, 60]. Nevertheless, all methods are constrained by the limitations of training datasets, resulting in suboptimal texture detail and limited generalizability. The latest work, E3Gen [74], combines UV maps and Gaussian splatting [33] to directly generate Gaussian attribute maps in UV space from images, which does not support arbitrary image inputs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 702, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 295, + 714 + ], + "type": "text", + "content": "To generate consistent human images or videos given a" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "text", + "content": "single image, recent diffusion video models animate a static image with pose-controllable video conditions[5, 5, 27, 59, 59, 67, 81]. Specifically, Champ [81] incorporate several rendered maps obtained from SMPL sequences, alongside skeleton-based motion guidance, to enrich the conditions to the latent diffusion model. However, they all fail to generate a 360-degree video and animate precise expressive SMPL-X-based humans due to insufficient training data." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 178, + 416, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 178, + 416, + 190 + ], + "spans": [ + { + "bbox": [ + 313, + 178, + 416, + 190 + ], + "type": "text", + "content": "3. Dataset Creation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 198, + 553, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 198, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 553, + 304 + ], + "type": "text", + "content": "Drawing inspiration from the latest advancements in generalizable large models, our core insight is to develop a large-scale reconstruction model with exceptional generalization capabilities. The crux of this endeavor lies in creating a comprehensive and high-quality digital human dataset. In this section, we introduce the creation of a large-scale human-centric generated dataset comprising over 2.4M high-resolution " + }, + { + "bbox": [ + 313, + 198, + 553, + 304 + ], + "type": "inline_equation", + "content": "(896 \\times 640)" + }, + { + "bbox": [ + 313, + 198, + 553, + 304 + ], + "type": "text", + "content": " multi-view images from more than 100K diverse subjects." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 305, + 554, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 305, + 554, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 305, + 554, + 437 + ], + "type": "text", + "content": "As shown in Fig. 2, the data generation pipeline consists of two stages. Firstly, diverse text prompts are generated by large language models incorporating various human-centric attributes, and high-quality images are synthesized using text-to-image generation models (Sec. 3.1). Secondly, we train a multi-view video generation model, MVChamp, conditioned on rendered full-body motions. With this, a large-scale multi-view images dataset is established by feeding both the synthesized and captured images (Sec. 3.2). Lastly, the statistics and characteristics of the HuGe100K are demonstrated in Sec. 3.3." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 443, + 493, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 443, + 493, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 443, + 493, + 455 + ], + "type": "text", + "content": "3.1. Image Collection and Generation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 461, + 553, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 553, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 553, + 665 + ], + "type": "text", + "content": "Given the limited amount of existing datasets [3, 49, 63, 68] and legal concerns regarding portrait and copyright issues, we propose leveraging an image generation model to construct a large-scale, diverse, and high-fidelity dataset of full-body human images. Using the latest text-to-image model Flux [23], we design descriptive prompts based on the required attributes for human figures, ensuring diversity across area, clothing, body shape, age, and gender shown in Fig. 2(i). To avoid long-tail attribute distributions and ensure uniform coverage of traits, we apply uniform sampling in attribute selection for the image generation process. This approach generated over 100K images in total. However, due to issues of visual non-compliance and high similarity among some images, we conducted manual filtering and retained a final set of 90K high-quality images. Additionally, we combine 90K synthetic images and 10K real-life images from DeepFashion [41]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 671, + 549, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 549, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 549, + 684 + ], + "type": "text", + "content": "3.2. Multi-view Image Animation and Generation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "Image-based animating models [5, 27, 67, 75, 81] struggle to achieve generation of " + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "-consistent and diverse hu" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26310" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 68, + 553, + 258 + ], + "blocks": [ + { + "bbox": [ + 55, + 68, + 553, + 258 + ], + "lines": [ + { + "bbox": [ + 55, + 68, + 553, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 68, + 553, + 258 + ], + "type": "image", + "image_path": "b63c0190a8160edd8b52664efcf9fd9d1a08ff4ad460f476108674f28e460969.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 259, + 555, + 304 + ], + "lines": [ + { + "bbox": [ + 55, + 259, + 555, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 259, + 555, + 304 + ], + "type": "text", + "content": "Figure 2. Pipeline for constructing our HuGe100K. Diverse attribute combinations from GPT-4 templates create text prompts, generating synthetic images via FLUX, combined with real images from DeepFashion. SMPL-X fitting produces multi-view pose sequences with 360-degree rotations and diverse animatable motions. MVChamp then converts these sequences into multi-view images, ensuring 3D consistency in the dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 325, + 297, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 325, + 297, + 564 + ], + "spans": [ + { + "bbox": [ + 54, + 325, + 297, + 564 + ], + "type": "text", + "content": "man videos conditioned on pose sequences (e.g., 2D poses and 3D SMPL(-X) conditions). To address this limitation, we re-train a state-of-the-art video generative model, Champ [81], to obtain a multi-view consistent generative model, MVChamp. Firstly, We collect a curated dataset of over 100K single-person videos with various motions (e.g., dancing), including approximately 20K videos involving the action of turning around, to fine-tune the model and enhance its generalizability for diverse inputs and motions. Secondly, to further improve 3D consistency[16], we utilize scanned models from THuman 2.1 to generate uniformly distributed views via Blender. From these, we select a subset of 24 views that are evenly spaced around the full " + }, + { + "bbox": [ + 54, + 325, + 297, + 564 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 54, + 325, + 297, + 564 + ], + "type": "text", + "content": " rotation to fine-tune MVChamp's temporal layers using diffusion loss. Lastly, to support high-quality whole-body animation (i.e., body and hands), we introduce precise SMPL-X estimation using NLF [54] for body shape and pose estimation. Additionally, we employ HaMeR [47] for hand estimation, rendering the corresponding hand template into depth maps as additional control signals." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": "After these training processes, MVChamp generates multi-view images conditioned on SMPL-X parametric model. The pose condition is set to an \"A-Pose\" " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": " of the time and a random pose from dance videos (via SMPLer-X [4]) " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": " of the time, balancing pose variability. To address discrepancies between the first and last frames, we propose a Temporal Shift Denoising Strategy: during the denoising steps, we shift latent inputs and pose signals temporally, moving the last frame to the first. This improves continuity between frames without additional inference costs. Finally, FaceFusion [24] is utilized to enhance facial details." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 328, + 322, + 542, + 365 + ], + "blocks": [ + { + "bbox": [ + 328, + 322, + 542, + 365 + ], + "lines": [ + { + "bbox": [ + 328, + 322, + 542, + 365 + ], + "spans": [ + { + "bbox": [ + 328, + 322, + 542, + 365 + ], + "type": "image", + "image_path": "2a67f02d46c67091a06a7612b4d661065ab4fc4612a13f4de0af8ea336f3701a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 374, + 555, + 419 + ], + "lines": [ + { + "bbox": [ + 313, + 374, + 555, + 419 + ], + "spans": [ + { + "bbox": [ + 313, + 374, + 555, + 419 + ], + "type": "text", + "content": "Figure 3. A paired example from the proposed HuGe100K Dataset. For each reference image, we generate a set of multiview images using an estimated shape and a specific pose. The figure shows the pose is well-aligned." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 441, + 499, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 499, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 499, + 453 + ], + "type": "text", + "content": "3.3. Data Statistics and Characteristics" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 460, + 555, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 555, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 555, + 556 + ], + "type": "text", + "content": "Tab. 1 compares our dataset HuGe100K to other 3D human datasets regarding scale, diversity, and consistency. With over 100K human identities and 20K poses, HuGe100K offers balanced diversity across dimensions such as area, clothing, body shape, age, and gender (see Fig. 1). Each sample includes a reference image, SMPL-X estimates, 24 uniformly sampled multi-view images, camera parameters, and SMPL-X data (see Fig. 3)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 568, + 518, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 568, + 518, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 568, + 518, + 582 + ], + "type": "text", + "content": "4. Large Human Reconstruction Model" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 589, + 555, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 589, + 555, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 589, + 555, + 662 + ], + "type": "text", + "content": "In this section, we present the large-scale human reconstruction model, named IDOL. The overview pipeline is shown in Fig. 4. In Sec. 4.1, we describe the animatable human representation. Sec. 4.2 details the network architecture of IDOL, while Sec. 4.3 explains how we trained the model in an end-to-end manner using the multi-view image dataset." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 670, + 504, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 670, + 504, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 670, + 504, + 684 + ], + "type": "text", + "content": "4.1. Animatable Human Representation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "Similarly to previous work [28, 35, 74], IDOL leverages 3D Gaussian Splatting [33] in conjunction with SMPL-X" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26311" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 69, + 542, + 184 + ], + "blocks": [ + { + "bbox": [ + 70, + 69, + 542, + 184 + ], + "lines": [ + { + "bbox": [ + 70, + 69, + 542, + 184 + ], + "spans": [ + { + "bbox": [ + 70, + 69, + 542, + 184 + ], + "type": "image", + "image_path": "5316acc26691809aea53c9dfca58977e8b50f661e586da5b397cbf6d510c720d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 190, + 555, + 258 + ], + "lines": [ + { + "bbox": [ + 54, + 190, + 555, + 258 + ], + "spans": [ + { + "bbox": [ + 54, + 190, + 555, + 258 + ], + "type": "text", + "content": "Figure 4. The architecture of IDOL, a full-differentiable transformer-based framework for reconstructing animatable 3D human from a single image. The model integrates a high-resolution " + }, + { + "bbox": [ + 54, + 190, + 555, + 258 + ], + "type": "inline_equation", + "content": "(1024\\times 1024)" + }, + { + "bbox": [ + 54, + 190, + 555, + 258 + ], + "type": "text", + "content": " encoder [34] and fuses image tokens with learnable UV tokens through the UV-Alignment Transformer. A UV Decoder predicts Gaussian attribute maps as intermediate representations, capturing the human's geometry and appearance in a structured 2D UV space defined by the SMPL-X model. These maps, in conjunction with the SMPL-X model, represent a 3D human avatar in a canonical space, which can be animated using linear blend skinning (LBS). The model is optimized using multi-view images with diverse poses and identities, learning to disentangle pose, appearance, and shape." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 277, + 294, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 277, + 294, + 350 + ], + "spans": [ + { + "bbox": [ + 54, + 277, + 294, + 350 + ], + "type": "text", + "content": "for 3D human representation, aiming to address the challenges of real-time rendering and accurate animation of human avatars. Specifically, each Gaussian primitive " + }, + { + "bbox": [ + 54, + 277, + 294, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_k" + }, + { + "bbox": [ + 54, + 277, + 294, + 350 + ], + "type": "text", + "content": " is characterized by: " + }, + { + "bbox": [ + 54, + 277, + 294, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_k = \\{\\mu_k,\\alpha_k,\\mathbf{r}_k,\\mathbf{s}_k,\\mathbf{c}_k\\}" + }, + { + "bbox": [ + 54, + 277, + 294, + 350 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 54, + 277, + 294, + 350 + ], + "type": "inline_equation", + "content": "\\mu_{k}" + }, + { + "bbox": [ + 54, + 277, + 294, + 350 + ], + "type": "text", + "content": " is the 3D position of the Gaussian, " + }, + { + "bbox": [ + 54, + 277, + 294, + 350 + ], + "type": "inline_equation", + "content": "\\alpha_{k}" + }, + { + "bbox": [ + 54, + 277, + 294, + 350 + ], + "type": "text", + "content": " is opacity, " + }, + { + "bbox": [ + 54, + 277, + 294, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_k" + }, + { + "bbox": [ + 54, + 277, + 294, + 350 + ], + "type": "text", + "content": " is the rotation, " + }, + { + "bbox": [ + 54, + 277, + 294, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_k" + }, + { + "bbox": [ + 54, + 277, + 294, + 350 + ], + "type": "text", + "content": " is the scale, and " + }, + { + "bbox": [ + 54, + 277, + 294, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_k" + }, + { + "bbox": [ + 54, + 277, + 294, + 350 + ], + "type": "text", + "content": " is the color." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 364, + 296, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 364, + 296, + 616 + ], + "spans": [ + { + "bbox": [ + 56, + 364, + 296, + 616 + ], + "type": "text", + "content": "3D Gaussian Human. Directly predicting all 3D Gaussian primitives is computationally intensive. Instead, IDOL leverages the predefined 2D UV space of the SMPL-X model to transform the 3D representation task into a more manageable 2D problem. Initially, IDOL predicts Gaussian attribute maps that encode the offset values " + }, + { + "bbox": [ + 56, + 364, + 296, + 616 + ], + "type": "inline_equation", + "content": "\\{\\delta_{\\mu_k}, \\delta_{\\mathbf{r}_k}, \\delta \\mathbf{s}_k\\}" + }, + { + "bbox": [ + 56, + 364, + 296, + 616 + ], + "type": "text", + "content": ", as well as the color " + }, + { + "bbox": [ + 56, + 364, + 296, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_k" + }, + { + "bbox": [ + 56, + 364, + 296, + 616 + ], + "type": "text", + "content": " and opacity " + }, + { + "bbox": [ + 56, + 364, + 296, + 616 + ], + "type": "inline_equation", + "content": "\\alpha_k" + }, + { + "bbox": [ + 56, + 364, + 296, + 616 + ], + "type": "text", + "content": " for each Gaussian primitive " + }, + { + "bbox": [ + 56, + 364, + 296, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_k" + }, + { + "bbox": [ + 56, + 364, + 296, + 616 + ], + "type": "text", + "content": ". Similarly to " + }, + { + "bbox": [ + 56, + 364, + 296, + 616 + ], + "type": "inline_equation", + "content": "E^3 Gen" + }, + { + "bbox": [ + 56, + 364, + 296, + 616 + ], + "type": "text", + "content": " [74], the position " + }, + { + "bbox": [ + 56, + 364, + 296, + 616 + ], + "type": "inline_equation", + "content": "\\mu_k" + }, + { + "bbox": [ + 56, + 364, + 296, + 616 + ], + "type": "text", + "content": ", scale " + }, + { + "bbox": [ + 56, + 364, + 296, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_k" + }, + { + "bbox": [ + 56, + 364, + 296, + 616 + ], + "type": "text", + "content": ", and rotation " + }, + { + "bbox": [ + 56, + 364, + 296, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_k" + }, + { + "bbox": [ + 56, + 364, + 296, + 616 + ], + "type": "text", + "content": " of each Gaussian primitive " + }, + { + "bbox": [ + 56, + 364, + 296, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_k" + }, + { + "bbox": [ + 56, + 364, + 296, + 616 + ], + "type": "text", + "content": " is modeled relative to its SMPL-X vertex as follows: " + }, + { + "bbox": [ + 56, + 364, + 296, + 616 + ], + "type": "inline_equation", + "content": "\\mu_k = \\hat{\\mu}_k + \\delta_{\\mu_k}" + }, + { + "bbox": [ + 56, + 364, + 296, + 616 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 56, + 364, + 296, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_k = \\hat{\\mathbf{s}}_k \\cdot \\delta_{\\mathbf{s}_k}" + }, + { + "bbox": [ + 56, + 364, + 296, + 616 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 56, + 364, + 296, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_k = \\hat{\\mathbf{r}}_k \\cdot \\delta_{\\mathbf{r}_k}" + }, + { + "bbox": [ + 56, + 364, + 296, + 616 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 56, + 364, + 296, + 616 + ], + "type": "inline_equation", + "content": "\\hat{\\mu}_k, \\hat{\\mathbf{s}}_k" + }, + { + "bbox": [ + 56, + 364, + 296, + 616 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 56, + 364, + 296, + 616 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{r}}_k" + }, + { + "bbox": [ + 56, + 364, + 296, + 616 + ], + "type": "text", + "content": " are the initial values based on the SMPL-X model. The color " + }, + { + "bbox": [ + 56, + 364, + 296, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_k" + }, + { + "bbox": [ + 56, + 364, + 296, + 616 + ], + "type": "text", + "content": " is assigned using RGB values, and the opacity " + }, + { + "bbox": [ + 56, + 364, + 296, + 616 + ], + "type": "inline_equation", + "content": "\\alpha_k" + }, + { + "bbox": [ + 56, + 364, + 296, + 616 + ], + "type": "text", + "content": " is set to 1, indicating full opacity. This unified UV space significantly reduces computational complexity and fully exploits the geometric and semantic priors provided by the SMPL-X model. By modeling Gaussian primitives relative to SMPL-X vertices, IDOL ensures semantic consistency across corresponding body parts of diverse avatars, thereby enhancing the model's generalization capability across various reference images." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "text", + "content": "Animation and Rendering. Given a target pose, we calculate the transformation of each human joint using predefined kinematic relationships. The transformation of each Gaussian primitive is performed using a forward skinning technique based on LBS. Specifically, the position of each Gaussian primitive " + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_k" + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "text", + "content": " is transformed as follows: " + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\mu_k' = \\sum_{i=1}^{n_b} w_i \\mathbf{B}_i \\mu_k" + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "text", + "content": ". Additionally, the rotation matrix " + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_k" + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "text", + "content": " is up-" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 277, + 555, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 555, + 350 + ], + "type": "text", + "content": "dated by: " + }, + { + "bbox": [ + 313, + 277, + 555, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_k^{\\prime} = \\mathbf{T}_k^{1:3,1:3}\\mathbf{R}_k" + }, + { + "bbox": [ + 313, + 277, + 555, + 350 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 277, + 555, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_k = \\sum_{i=1}^{n_b} w_i \\mathbf{B}_i" + }, + { + "bbox": [ + 313, + 277, + 555, + 350 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 277, + 555, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_k" + }, + { + "bbox": [ + 313, + 277, + 555, + 350 + ], + "type": "text", + "content": " is the rotation matrix representation of the rotation angle " + }, + { + "bbox": [ + 313, + 277, + 555, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_k" + }, + { + "bbox": [ + 313, + 277, + 555, + 350 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 313, + 277, + 555, + 350 + ], + "type": "inline_equation", + "content": "n_b" + }, + { + "bbox": [ + 313, + 277, + 555, + 350 + ], + "type": "text", + "content": " is the number of joints, " + }, + { + "bbox": [ + 313, + 277, + 555, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_i" + }, + { + "bbox": [ + 313, + 277, + 555, + 350 + ], + "type": "text", + "content": " is the transformation matrix for each joint, and " + }, + { + "bbox": [ + 313, + 277, + 555, + 350 + ], + "type": "inline_equation", + "content": "w_i" + }, + { + "bbox": [ + 313, + 277, + 555, + 350 + ], + "type": "text", + "content": " represents the skinning weights, indicating the influence of each joint's motion on the Gaussian primitive's position " + }, + { + "bbox": [ + 313, + 277, + 555, + 350 + ], + "type": "inline_equation", + "content": "\\mu_k" + }, + { + "bbox": [ + 313, + 277, + 555, + 350 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 350, + 556, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 556, + 458 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 556, + 458 + ], + "type": "text", + "content": "To estimate the skinning weights for each " + }, + { + "bbox": [ + 313, + 350, + 556, + 458 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_k" + }, + { + "bbox": [ + 313, + 350, + 556, + 458 + ], + "type": "text", + "content": ", we first compute a body-part skinning field using a pre-defined low-resolution volume [6]. For smaller regions like the hands and face, which undergo minimal topological variation, weights are interpolated from the SMPL-X template via barycentric coordinates. This strategy enables modeling large topology changes across identities (e.g., from clothing) while stabilizing convergence in less variable areas such as fingers and facial regions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 465, + 425, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 465, + 425, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 465, + 425, + 477 + ], + "type": "text", + "content": "4.2. Network Structure" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 483, + 555, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 555, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 555, + 508 + ], + "type": "text", + "content": "As illustrated in Fig. 4, IDOL is a full-differentiable framework for reconstructing animatable 3D human." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 521, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 556, + 714 + ], + "type": "text", + "content": "High-resolution Image Encoder. Higher resolution of the input image results in a high-quality reconstruction. However, previous works suffer from low-resolution ViT-based encoders, such as DINOv2 [45], which support the " + }, + { + "bbox": [ + 313, + 521, + 556, + 714 + ], + "type": "inline_equation", + "content": "448 \\times 448" + }, + { + "bbox": [ + 313, + 521, + 556, + 714 + ], + "type": "text", + "content": " resolution. To fully leverage the resolution of the HuGe100K dataset, we further adopt a high-resolution human foundation model, Sapiens [34], to encode the " + }, + { + "bbox": [ + 313, + 521, + 556, + 714 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 313, + 521, + 556, + 714 + ], + "type": "text", + "content": " resolution images into patch-wise feature tokens, formulated as: " + }, + { + "bbox": [ + 313, + 521, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{F} = \\{\\mathbf{f}_i\\}_{i=1}^n \\in \\mathbb{R}^{d_E}" + }, + { + "bbox": [ + 313, + 521, + 556, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 521, + 556, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 521, + 556, + 714 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 313, + 521, + 556, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 521, + 556, + 714 + ], + "type": "text", + "content": "-th image patch, and " + }, + { + "bbox": [ + 313, + 521, + 556, + 714 + ], + "type": "inline_equation", + "content": "d_E" + }, + { + "bbox": [ + 313, + 521, + 556, + 714 + ], + "type": "text", + "content": " is the channel length of each token. The Sapiens model is pretrained on 300 million in-the-wild human images using the Masked Autoencoder (MAE) framework [18], enabling it to excel in preserving fine-grained details and capturing diverse human poses and appearances, making it highly effective for high-resolution human image feature extraction." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26312" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "content": "UV-Alignment Transformer. To map irregular and diverse reference images onto regular UV feature maps, a UV-Alignment Transformer is employed to align learnable spatial-positional UV tokens " + }, + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^0" + }, + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "content": " with reference image features " + }, + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "content": ". Specifically, the UV-Alignment Transformer concatenates " + }, + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "content": " from Spaiens with " + }, + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^0" + }, + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "content": ", and aggregates and refines features through " + }, + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "content": " transformer blocks, producing an enhanced representation " + }, + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^D" + }, + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "content": ". Each transformer block utilizes self-attention mechanisms, enabling the model to capture complex relationships among the input tokens and impute missing parts using correlated tokens." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 217, + 294, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 217, + 294, + 291 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 294, + 291 + ], + "type": "text", + "content": "UV Decoder. The UV tokens " + }, + { + "bbox": [ + 55, + 217, + 294, + 291 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^D" + }, + { + "bbox": [ + 55, + 217, + 294, + 291 + ], + "type": "text", + "content": " are reshaped and decoded into Gaussian attribute maps, capturing the human's geometry and appearance within the structured 2D UV space as shown in Sec. 4.1. To preserve and enhance fine details in the decoded Gaussian attribute maps, a Convolutional Neural Network is employed for spatial up-sampling." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 297, + 170, + 310 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 297, + 170, + 310 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 170, + 310 + ], + "type": "text", + "content": "4.3. Training Objectives" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 316, + 295, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 295, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 295, + 437 + ], + "type": "text", + "content": "IDOL reconstructs 3D human by predicting Gaussian attribute UV maps in a single forward pass, offering significant advantages in inference speed and enabling end-to-end training with multi-view images. For each sample, we select a front view as the reference image " + }, + { + "bbox": [ + 55, + 316, + 295, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{ref}}" + }, + { + "bbox": [ + 55, + 316, + 295, + 437 + ], + "type": "text", + "content": " along with a random set of generated multi-view images " + }, + { + "bbox": [ + 55, + 316, + 295, + 437 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{I}_{\\mathrm{gt},i}\\}_{i = 1}^{N}" + }, + { + "bbox": [ + 55, + 316, + 295, + 437 + ], + "type": "text", + "content": " and their corresponding SMPL-X parameters and camera for supervision. It takes " + }, + { + "bbox": [ + 55, + 316, + 295, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{ref}}" + }, + { + "bbox": [ + 55, + 316, + 295, + 437 + ], + "type": "text", + "content": " as input to generate the 3D human, using differentiable rendering to produce the multi-view images " + }, + { + "bbox": [ + 55, + 316, + 295, + 437 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{I}_{\\mathrm{pred},i}\\}_{i = 1}^{N}" + }, + { + "bbox": [ + 55, + 316, + 295, + 437 + ], + "type": "text", + "content": ". The loss function is defined as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 65, + 445, + 295, + 478 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 445, + 295, + 478 + ], + "spans": [ + { + "bbox": [ + 65, + 445, + 295, + 478 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\sum_ {i = 1} ^ {N} \\left(\\left\\| \\mathbf {I} _ {\\mathrm {g t}, i} - \\mathbf {I} _ {\\text {p r e d}, i} \\right\\| ^ {2} + \\lambda L _ {v g g} \\left(\\mathbf {I} _ {\\mathrm {g t}, i}, \\mathbf {I} _ {\\text {p r e d}, i}\\right)\\right), \\tag {1}", + "image_path": "8704f9a9b3b6f7e3114566f596c5071c4a9e63aa2dc89899a43c508bc8a882ee.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 485, + 295, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 485, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 485, + 295, + 510 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 485, + 295, + 510 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 485, + 295, + 510 + ], + "type": "text", + "content": " controls the balance between the mean square error loss and the perceptual loss " + }, + { + "bbox": [ + 55, + 485, + 295, + 510 + ], + "type": "inline_equation", + "content": "L_{vgg}" + }, + { + "bbox": [ + 55, + 485, + 295, + 510 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 520, + 137, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 520, + 137, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 520, + 137, + 533 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 540, + 234, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 234, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 234, + 552 + ], + "type": "text", + "content": "5.1. Experiments on Dataset Creation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 557, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 295, + 676 + ], + "type": "text", + "content": "Fig. 5 (left) presents the qualitative results of the ablation study on our data generation method. Specifically, Fig. 5 (a) and Fig. 5 (b) demonstrate how hand refinement and face swapping effectively enhance the quality of the generated hands and faces. Fig. 5 (c) highlights the importance of fine-tuning the 3D dataset Thuman 2.1 for improving the 3D consistency of MVChamp, while Fig. 5 (d) illustrates that the Temporal Shift Denoising strategy, which involves cycling the first and last frames, improves the consistency between the first and last frames." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "Fig. 5 (right) compares our MVChamp's generation quality against previous human image animation models (e.g., Champ, MimicMotion [75]) and general multi-view" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": "video generation model (e.g., SV3D [58]). The issues of these methods are highlighted in the figure. Champ struggles to generate plausible hands and heads in multi-view scenarios. MimicMotion not only fails to produce realistic shoes but also has difficulty preserving the identity, especially for the anime image. SV3D produces low-quality multi-view human generation due to the lack of 3D prior knowledge of the human body. In contrast, our model generates consistent, high-quality results across views." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 187, + 512, + 200 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 187, + 512, + 200 + ], + "spans": [ + { + "bbox": [ + 313, + 187, + 512, + 200 + ], + "type": "text", + "content": "5.2. Comparison of Reconstruction Model" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 205, + 553, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 553, + 289 + ], + "type": "text", + "content": "Implementation Details. For the encoder, we employ the pre-trained Sapiens-1B model [34] and keep its weights frozen during training. We then define a transformer-based framework with a parameter size of " + }, + { + "bbox": [ + 313, + 205, + 553, + 289 + ], + "type": "inline_equation", + "content": "0.5B" + }, + { + "bbox": [ + 313, + 205, + 553, + 289 + ], + "type": "text", + "content": " to perform feature fusion. Additionally, we densify the SMPL-X vertex set by sampling approximately 200,000 vertices to map attributes and represent the 3D human model effectively." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 304, + 555, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 555, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 555, + 508 + ], + "type": "text", + "content": "Dataset and Metrics. We train IDOL on a dataset consisting of generated multi-view images from HuGe100K and rendered images from THuman 2.1 [79]. For THuman 2.1, we render 72 view images at a resolution of " + }, + { + "bbox": [ + 313, + 304, + 555, + 508 + ], + "type": "inline_equation", + "content": "896 \\times 640" + }, + { + "bbox": [ + 313, + 304, + 555, + 508 + ], + "type": "text", + "content": " for each scan, with rendering views uniformly distributed. To evaluate performance quantitatively, we reserve the last 50 cases from both HuGe100K and THuman 2.1 as the testing set. We use the most frontal image as the reference image and the remaining images as ground-truth data for evaluation. The ground-truth camera parameters and SMPL-X parameters are provided for all methods. The renderings of the reconstructed model are then compared to their corresponding ground truth images with several metrics, including Mean Squared Error (MSE), Learned Perceptual Image Patch Similarity (LPIPS) [73], Peak Signal-to-Noise Ratio (PSNR). More details on implementation can be found in the supplementary material." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 521, + 553, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 553, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 553, + 605 + ], + "type": "text", + "content": "Baselines. We compare IDOL with three baseline categories in the Single-Image Human Reconstruction task. The first category comprises methods based on loop optimization or pixel-alignment modules, including GTA [77] and SIFU [78]. These methods focus on refining human reconstruction through iterative optimization or predicting 3D geometry from pixel-aligned features." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 606, + 553, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 553, + 666 + ], + "type": "text", + "content": "The second category encompasses large-scale generic reconstruction networks, represented by LGM [57]. These models are known for their ability to handle large datasets and their scalability, offering advantages in terms of fast reference speeds and large output resolutions." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "type": "text", + "content": "The third category involves optimization-based 3D generation methods using score distillation sampling (SDS), exemplified by DreamGaussian [56]. These methods leverage priors from 2D diffusion models to distill 3D objects." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26313" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 267, + 277 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 267, + 277 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 267, + 277 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 267, + 277 + ], + "type": "image", + "image_path": "1d4ebec419a5575533de3618f4338060fb86aa35099fc797dde03cd381261967.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 72, + 277, + 138, + 285 + ], + "lines": [ + { + "bbox": [ + 72, + 277, + 138, + 285 + ], + "spans": [ + { + "bbox": [ + 72, + 277, + 138, + 285 + ], + "type": "text", + "content": "(c) THuman Finetune" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 154, + 277, + 270, + 285 + ], + "lines": [ + { + "bbox": [ + 154, + 277, + 270, + 285 + ], + "spans": [ + { + "bbox": [ + 154, + 277, + 270, + 285 + ], + "type": "text", + "content": "(d) Temporal Shift Denoising Strategy" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 276, + 69, + 553, + 277 + ], + "blocks": [ + { + "bbox": [ + 276, + 69, + 553, + 277 + ], + "lines": [ + { + "bbox": [ + 276, + 69, + 553, + 277 + ], + "spans": [ + { + "bbox": [ + 276, + 69, + 553, + 277 + ], + "type": "image", + "image_path": "a9518cfb5bb51df4681f3d9221f0396de4c46e0bb5e3635d484e4b7207615455.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 279, + 277, + 374, + 285 + ], + "lines": [ + { + "bbox": [ + 279, + 277, + 374, + 285 + ], + "spans": [ + { + "bbox": [ + 279, + 277, + 374, + 285 + ], + "type": "text", + "content": "Reference (e) MVChamp (ours)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 388, + 277, + 536, + 285 + ], + "lines": [ + { + "bbox": [ + 388, + 277, + 536, + 285 + ], + "spans": [ + { + "bbox": [ + 388, + 277, + 536, + 285 + ], + "type": "text", + "content": "(f) Champ (g) MimicMotion (h) SV3D" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 59, + 319, + 555, + 559 + ], + "blocks": [ + { + "bbox": [ + 118, + 295, + 490, + 306 + ], + "lines": [ + { + "bbox": [ + 118, + 295, + 490, + 306 + ], + "spans": [ + { + "bbox": [ + 118, + 295, + 490, + 306 + ], + "type": "text", + "content": "Figure 5. Qualitative results of our MVChamp ablation study (left) and comparison experiment (right)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 319, + 555, + 559 + ], + "lines": [ + { + "bbox": [ + 59, + 319, + 555, + 559 + ], + "spans": [ + { + "bbox": [ + 59, + 319, + 555, + 559 + ], + "type": "image", + "image_path": "16eace9fe2735034876bb1c22d34d06f18ee4bde603478020e53d7fc2f2d2dd7.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 80, + 567, + 528, + 578 + ], + "lines": [ + { + "bbox": [ + 80, + 567, + 528, + 578 + ], + "spans": [ + { + "bbox": [ + 80, + 567, + 528, + 578 + ], + "type": "text", + "content": "Figure 6. Comparisons on (a) the upper: novel-view synthesis given a single image, and (b) the lower: our animated results." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 599, + 294, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 599, + 294, + 648 + ], + "spans": [ + { + "bbox": [ + 55, + 599, + 294, + 648 + ], + "type": "text", + "content": "DreamGaussian accelerates convergence by progressively densifying 3D Gaussians, significantly reducing the reconstruction time. However, it still takes approximately two minutes to reconstruct a single object." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "Quantitative comparison. As shown in Tab 2, our method outperforms all baselines in all metrics. We attribute this superior performance to the large-scale dataset HuGe100K and the design of our large-scale reconstruc" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 599, + 555, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 599, + 555, + 708 + ], + "spans": [ + { + "bbox": [ + 313, + 599, + 555, + 708 + ], + "type": "text", + "content": "tion model IDOL, which allows for more effective training and improved synthesis of appearance results. Despite this, we note that SIFU and GTA report a lower metric than what we expected in our test settings. While we provide accurate ground-truth SMPL-X parameters and camera settings, the ideal orthographic projection required by the pixel-alignment modules in SIFU and GTA is not well-suited to our perspective projection model, where the camera focus ranges from 35 to 80. This mismatch leads to a" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26314" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 78, + 70, + 277, + 133 + ], + "blocks": [ + { + "bbox": [ + 78, + 70, + 277, + 133 + ], + "lines": [ + { + "bbox": [ + 78, + 70, + 277, + 133 + ], + "spans": [ + { + "bbox": [ + 78, + 70, + 277, + 133 + ], + "type": "table", + "html": "
MethodMSE ↓PSNR ↑LPIPS ↓
SIFU [78]0.04214.2041.612
GTA [77]0.04114.2821.629
Ours-w/o HuGe100K0.01719.2251.326
Ours-full0.00821.6731.138
", + "image_path": "78908a92893a3b509ddc2549e7dea07fbf477cda146a0b8381c5da1ac2016a30.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 141, + 289, + 152 + ], + "lines": [ + { + "bbox": [ + 61, + 141, + 289, + 152 + ], + "spans": [ + { + "bbox": [ + 61, + 141, + 289, + 152 + ], + "type": "text", + "content": "Table 2. Evaluation of Comparison and Ablation Experiments." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 175, + 296, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 175, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 296, + 319 + ], + "type": "text", + "content": "misalignment between the rendered images and the ground truth, adversely affecting their performance metrics. In fact, many real-life photographs are taken with medium to short focal lengths and cannot be approximated using orthographic projection, which is a less noticeable drawback of SIFU and GTA. Additionally, SIFU and GTA, trained with images at a resolution of " + }, + { + "bbox": [ + 55, + 175, + 296, + 319 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 55, + 175, + 296, + 319 + ], + "type": "text", + "content": ", struggle to synthesize detailed textures, particularly in the invisible areas of the reference images. This limitation is primarily due to the lack of comprehensiveness and diversity in their training datasets, which restricts their performance in generating invisible aspects of the images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 335, + 296, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 335, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 296, + 479 + ], + "type": "text", + "content": "Qualitative comparison. We perform a qualitative evaluation on an in-the-wild dataset with methods from [56, 57, 77, 78], with the results presented in Fig. 6. Our evaluation includes a subject in a complex outfit featuring a baseball cap and textured clothing, demonstrating our method's ability to synthesize intricate textures and handle loose outfits across different views. Additional tests include out-of-domain cartoon data and large-angle side views and assessing model adaptability and viewpoint handling. IDOL consistently outperforms the baselines, which struggle with detail reproduction and texture consistency under these various conditions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 487, + 271, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 271, + 499 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 271, + 499 + ], + "type": "text", + "content": "5.3. Ablation Study on Reconstruction Model" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 506, + 296, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 506, + 296, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 296, + 639 + ], + "type": "text", + "content": "We assess the impact of proposed components by removing them individually from IDOL. As shown in Fig. 7, replacing the Sapiens [34] encoder with DINO v2 [45] (w/o Sapiens) reduces detail quality, resulting in less realistic textures and folds. Excluding the HuGe100K dataset (w/o HuGe100K) causes significant distortion, including blurred details and color bleeding. The complete model, with all components, produces the most realistic and detailed results. These findings highlight the critical roles of both the Sapiens encoder and the HuGe100K dataset in achieving high-quality avatar generation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 647, + 200, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 200, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 200, + 659 + ], + "type": "text", + "content": "5.4. Downstream Applications" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "IDOL reconstructs an avatar by combining Gaussian attribute maps with the SMPL-X model, enabling users to modify the avatar's appearance by editing UV texture maps and to control body shape by adjusting SMPL-X shape pa" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 328, + 69, + 544, + 224 + ], + "blocks": [ + { + "bbox": [ + 328, + 69, + 544, + 224 + ], + "lines": [ + { + "bbox": [ + 328, + 69, + 544, + 224 + ], + "spans": [ + { + "bbox": [ + 328, + 69, + 544, + 224 + ], + "type": "image", + "image_path": "9d1dabd4ba7c55ce5f9bce4176e87bea8116872ba262e5598207f50a43537c1c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 233, + 538, + 245 + ], + "lines": [ + { + "bbox": [ + 329, + 233, + 538, + 245 + ], + "spans": [ + { + "bbox": [ + 329, + 233, + 538, + 245 + ], + "type": "text", + "content": "Figure 7. Qualitative Results of Ablation Study of IDOL." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 315, + 258, + 555, + 323 + ], + "blocks": [ + { + "bbox": [ + 315, + 258, + 555, + 323 + ], + "lines": [ + { + "bbox": [ + 315, + 258, + 555, + 323 + ], + "spans": [ + { + "bbox": [ + 315, + 258, + 555, + 323 + ], + "type": "image", + "image_path": "0b0b6ed44969dd72760d24f32e2fdbbe9ce1ec8c04af4208fba8f74e6049a4bb.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 331, + 555, + 354 + ], + "lines": [ + { + "bbox": [ + 313, + 331, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 555, + 354 + ], + "type": "text", + "content": "Figure 8. Controllable Avatar Editing: (a) texture editing; (b) body shape editing." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 377, + 555, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 377, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 555, + 437 + ], + "type": "text", + "content": "rameters. Fig. 8(a) illustrates the effect of texture editing on clothing patterns, while Fig. 8(b) demonstrates shape editing to adjust the avatar's body size. This approach provides a high degree of controllability over both the avatar's appearance and body shape." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 452, + 476, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 452, + 476, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 452, + 476, + 464 + ], + "type": "text", + "content": "6. Conclusions and Limitations" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 473, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 555, + 605 + ], + "type": "text", + "content": "In conclusion, our work has made significant strides in creating an animatable 3D human from a single image. We introduced a scalable pipeline for training a simple yet efficient feed-forward model, incorporating a dataset generation framework, a large-scale dataset " + }, + { + "bbox": [ + 313, + 473, + 555, + 605 + ], + "type": "inline_equation", + "content": "HuGe100K" + }, + { + "bbox": [ + 313, + 473, + 555, + 605 + ], + "type": "text", + "content": ", and a scalable reconstruction transformer model, IDOL. This model efficiently reconstructs photorealistic 3D humans in under a second and demonstrates versatility across various applications. This model can efficiently reconstruct photorealistic humans in less than a second and is versatile enough to support various applications." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "text", + "content": "However, there are limitations. Due to the constraints of the video model used, we can only synthesize single-frame images from fixed viewpoints. Future work could consider generating longer motion sequences. The focus on body reconstruction and animation leaves facial optimization secondary, and the architecture lacks specific design for facial identity or expression. Additionally, handling half-body inputs remains challenging, and improvements in data generation strategies are needed to enhance performance." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26315" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "content": "This work was partly supported by the National Key Research and Development Program of China (2022YFF0902400), the Shenzhen Science and Technology Program (JCYJ20220818101014030), and the research fund of Tsinghua University - Tencent Joint Laboratory for Internet Innovation Technology." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 176, + 115, + 189 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 176, + 115, + 189 + ], + "spans": [ + { + "bbox": [ + 56, + 176, + 115, + 189 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 196, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 196, + 296, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 196, + 296, + 240 + ], + "spans": [ + { + "bbox": [ + 61, + 196, + 296, + 240 + ], + "type": "text", + "content": "[1] Badour AlBahar, Shunsuke Saito, Hung-Yu Tseng, Changil Kim, Johannes Kopf, and Jia-Bin Huang. Single-image 3d human digitization with shape-guided diffusion. In SIGGRAPH Asia 2023 Conference Papers, 2023. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 242, + 296, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 242, + 296, + 297 + ], + "spans": [ + { + "bbox": [ + 61, + 242, + 296, + 297 + ], + "type": "text", + "content": "[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 298, + 296, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 298, + 296, + 354 + ], + "spans": [ + { + "bbox": [ + 62, + 298, + 296, + 354 + ], + "type": "text", + "content": "[3] Zhongang Cai, Daxuan Ren, Ailing Zeng, Zhengyu Lin, Tao Yu, Wenjia Wang, Xiangyu Fan, Yang Gao, Yifan Yu, Liang Pan, et al. Humman: Multi-modal 4d human dataset for versatile sensing and modeling. In European Conference on Computer Vision, pages 557-577. Springer, 2022. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 354, + 296, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 354, + 296, + 409 + ], + "spans": [ + { + "bbox": [ + 62, + 354, + 296, + 409 + ], + "type": "text", + "content": "[4] Zhongang Cai, Wanqi Yin, Ailing Zeng, Chen Wei, Qingping Sun, Wang Yanjun, Hui En Pang, Haiyi Mei, Mingyuan Zhang, Lei Zhang, et al. Smpler-x: Scaling up expressive human pose and shape estimation. Advances in Neural Information Processing Systems, 36, 2024. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 411, + 296, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 411, + 296, + 477 + ], + "spans": [ + { + "bbox": [ + 62, + 411, + 296, + 477 + ], + "type": "text", + "content": "[5] Di Chang, Yichun Shi, Quankai Gao, Hongyi Xu, Jessica Fu, Guoxian Song, Qing Yan, Yizhe Zhu, Xiao Yang, and Mohammad Soleymani. Magicpose: Realistic human poses and facial expressions retargeting with identity-aware diffusion. In Forty-first International Conference on Machine Learning, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 479, + 296, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 479, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 62, + 479, + 296, + 533 + ], + "type": "text", + "content": "[6] Xu Chen, Tianjian Jiang, Jie Song, Max Rietmann, Andreas Geiger, Michael J Black, and Otmar Hilliges. Fast-snarf: A fast deformer for articulated neural fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(10): 11796-11809, 2023. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 535, + 296, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 535, + 296, + 579 + ], + "spans": [ + { + "bbox": [ + 62, + 535, + 296, + 579 + ], + "type": "text", + "content": "[7] Wei Cheng, Su Xu, Jingtan Piao, Chen Qian, Wayne Wu, Kwan-Yee Lin, and Hongsheng Li. Generalizable neural performer: Learning robust radiance fields for human novel view synthesis. arXiv preprint arXiv:2204.11798, 2022. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 581, + 296, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 581, + 296, + 646 + ], + "spans": [ + { + "bbox": [ + 62, + 581, + 296, + 646 + ], + "type": "text", + "content": "[8] Wei Cheng, Ruixiang Chen, Siming Fan, Wanqi Yin, Keyu Chen, Zhongang Cai, Jingbo Wang, Yang Gao, Zhengming Yu, Zhengyu Lin, et al. Dna-rendering: A diverse neural actor repository for high-fidelity human-centric rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19982-19993, 2023. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 647, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 647, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 647, + 296, + 712 + ], + "type": "text", + "content": "[9] Enric Corona, Mihai Zanfir, Thiemo Alldieck, Eduard Gabriel Bazavan, Andrei Zanfir, and Cristian Sminchisescu. Structured 3d features for reconstructing controllable avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16954-16964, 2023. 3" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 139 + ], + "type": "text", + "content": "[10] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 141, + 555, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 555, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 555, + 195 + ], + "type": "text", + "content": "[11] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-xl: A universe of " + }, + { + "bbox": [ + 316, + 141, + 555, + 195 + ], + "type": "inline_equation", + "content": "10\\mathrm{m} + 3\\mathrm{d}" + }, + { + "bbox": [ + 316, + 141, + 555, + 195 + ], + "type": "text", + "content": " objects. Advances in Neural Information Processing Systems, 36, 2024. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 197, + 555, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 555, + 228 + ], + "type": "text", + "content": "[12] Qiao Feng, Yebin Liu, Yu-Kun Lai, Jingyu Yang, and Kun Li. Fof: Learning fourier occupancy field for monocular real-time human reconstruction. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 230, + 554, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 230, + 554, + 275 + ], + "spans": [ + { + "bbox": [ + 317, + 230, + 554, + 275 + ], + "type": "text", + "content": "[13] Xiangjun Gao, Jiaolong Yang, Jongyoo Kim, Sida Peng, Zicheng Liu, and Xin Tong. Mps-nerf: Generalizable 3d human rendering from multiview images. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 277, + 554, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 277, + 554, + 341 + ], + "spans": [ + { + "bbox": [ + 317, + 277, + 554, + 341 + ], + "type": "text", + "content": "[14] Xiangjun Gao, Xiaoyu Li, Chaopeng Zhang, Qi Zhang, Yanpei Cao, Ying Shan, and Long Quan. Context-human: Free-view rendering of human from a single image with texture-consistent synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10084-10094, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 343, + 554, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 343, + 554, + 387 + ], + "spans": [ + { + "bbox": [ + 317, + 343, + 554, + 387 + ], + "type": "text", + "content": "[15] Xiangjun Gao, Xiaoyu Li, Yiyu Zhuang, Qi Zhang, Wenbo Hu, Chaopeng Zhang, Yao Yao, Ying Shan, and Long Quan. Mani-gs: Gaussian splatting manipulation with triangular mesh. arXiv preprint arXiv:2405.17811, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 388, + 554, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 388, + 554, + 433 + ], + "spans": [ + { + "bbox": [ + 317, + 388, + 554, + 433 + ], + "type": "text", + "content": "[16] Junlin Han, Filippos Kokkinos, and Philip Torr. Vfusion3d: Learning scalable 3d generative models from video diffusion models. In European Conference on Computer Vision, pages 333-350. Springer, 2025. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 434, + 554, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 434, + 554, + 488 + ], + "spans": [ + { + "bbox": [ + 317, + 434, + 554, + 488 + ], + "type": "text", + "content": "[17] Sang-Hun Han, Min-Gyu Park, Ju Hong Yoon, Ju-Mi Kang, Young-Jae Park, and Hae-Gon Jeon. High-fidelity 3d human digitization from single 2k resolution images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12869-12879, 2023. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 491, + 554, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 491, + 554, + 544 + ], + "spans": [ + { + "bbox": [ + 317, + 491, + 554, + 544 + ], + "type": "text", + "content": "[18] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 546, + 554, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 546, + 554, + 600 + ], + "spans": [ + { + "bbox": [ + 317, + 546, + 554, + 600 + ], + "type": "text", + "content": "[19] Tong He, Yuanlu Xu, Shunsuke Saito, Stefano Soatto, and Tony Tung. Arch++: Animation-ready clothed human reconstruction revisited. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11046-11056, 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 602, + 554, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 602, + 554, + 657 + ], + "spans": [ + { + "bbox": [ + 317, + 602, + 554, + 657 + ], + "type": "text", + "content": "[20] Xu He, Xiaoyu Li, Di Kang, Jiangnan Ye, Chaopeng Zhang, Liyang Chen, Xiangjun Gao, Han Zhang, Zhiyong Wu, and Haolin Zhuang. Magicman: Generative novel view synthesis of humans with 3d-aware diffusion and iterative refinement. arXiv preprint arXiv:2408.14211, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 658, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 554, + 713 + ], + "type": "text", + "content": "[21] Yuxiao He, Yiyu Zhuang, Yanwen Wang, Yao Yao, Siyu Zhu, Xiaoyu Li, Qi Zhang, Xun Cao, and Hao Zhu. Head360: Learning a parametric 3d full-head for free-view synthesis in 360. In European Conference on Computer Vision, pages 254-272. Springer, 2024. 2" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26316" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[22] Yicong Hong, Kai Zhang, Juixiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 138 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 138 + ], + "type": "text", + "content": "[23] https://github.com/black-forest labs/flux. Flux latent rectified flow transformers, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 141, + 294, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 141, + 294, + 161 + ], + "spans": [ + { + "bbox": [ + 57, + 141, + 294, + 161 + ], + "type": "text", + "content": "[24] https://github.com/facefusion/facefusion. Facefusion, 2024. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 164, + 294, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 164, + 294, + 185 + ], + "spans": [ + { + "bbox": [ + 57, + 164, + 294, + 185 + ], + "type": "text", + "content": "[25] https://renderpeople.com/3d people. Renderpeople dataset, 2015. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 187, + 272, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 187, + 272, + 198 + ], + "spans": [ + { + "bbox": [ + 57, + 187, + 272, + 198 + ], + "type": "text", + "content": "[26] https://web.twindom.com/. Twindom dataset, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 200, + 296, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 200, + 296, + 243 + ], + "spans": [ + { + "bbox": [ + 56, + 200, + 296, + 243 + ], + "type": "text", + "content": "[27] Li Hu. Animate anyone: Consistent and controllable image-to-video synthesis for character animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8153-8163, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 244, + 294, + 309 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 244, + 294, + 309 + ], + "spans": [ + { + "bbox": [ + 56, + 244, + 294, + 309 + ], + "type": "text", + "content": "[28] Liangxiao Hu, Hongwen Zhang, Yuxiang Zhang, Boyao Zhou, Boning Liu, Shengping Zhang, and Liqiang Nie. Gaussianavatar: Towards realistic human avatar modeling from a single video via animatable 3d gaussians. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 634-644, 2024. 2, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 311, + 294, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 311, + 294, + 354 + ], + "spans": [ + { + "bbox": [ + 56, + 311, + 294, + 354 + ], + "type": "text", + "content": "[29] Shoukang Hu, Fangzhou Hong, Liang Pan, Haiyi Mei, Lei Yang, and Ziwei Liu. Sherf: Generalizable human nerf from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 356, + 294, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 356, + 294, + 410 + ], + "spans": [ + { + "bbox": [ + 56, + 356, + 294, + 410 + ], + "type": "text", + "content": "[30] Yangyi Huang, Hongwei Yi, Weiyang Liu, Haofan Wang, Boxi Wu, Wenxiao Wang, Binbin Lin, Debing Zhang, and Deng Cai. One-shot implicit animatable avatars with model-based priors. In IEEE Conference on Computer Vision (ICCV), 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 412, + 294, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 412, + 294, + 465 + ], + "spans": [ + { + "bbox": [ + 56, + 412, + 294, + 465 + ], + "type": "text", + "content": "[31] Zeng Huang, Yuanlu Xu, Christoph Lassner, Hao Li, and Tony Tung. Arch: Animatable reconstruction of clothed humans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3093-3102, 2020. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 468, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 468, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 468, + 294, + 521 + ], + "type": "text", + "content": "[32] Mustafa Isik, Martin Rünz, Markos Georgopoulos, Taras Khakhulin, Jonathan Starck, Lourdes Agapito, and Matthias Nießner. Humanrf: High-fidelity neural radiance fields for humans in motion. ACM Transactions on Graphics (TOG), 42(4):1-12, 2023. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 524, + 294, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 524, + 294, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 524, + 294, + 567 + ], + "type": "text", + "content": "[33] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 2, 3, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 569, + 294, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 569, + 294, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 569, + 294, + 623 + ], + "type": "text", + "content": "[34] Rawal Khirodkar, Timur Bagautdinov, Julieta Martinez, Su Zhaoen, Austin James, Peter Selednik, Stuart Anderson, and Shunsuke Saito. Sapiens: Foundation for human vision models. In European Conference on Computer Vision, pages 206-228. Springer, 2025. 2, 5, 6, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 624, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 667 + ], + "type": "text", + "content": "[35] Tobias Kirschstein, Simon Giebenhain, Jiapeng Tang, Markos Georgopoulos, and Matthias Nießner. Gghead: Fast and generalizable 3d gaussian heads. arXiv preprint arXiv:2406.09377, 2024. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "text", + "content": "[36] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "In International Conference on Learning Representations (ICLR), 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 96, + 553, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 148 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 148 + ], + "type": "text", + "content": "[37] Peng Li, Wangguandong Zheng, Yuan Liu, Tao Yu, Yangguang Li, Xingqun Qi, Mengfei Li, Xiaowei Chi, Siyu Xia, Wei Xue, et al. Pshuman: Photorealistic single-view human reconstruction using cross-scale diffusion. arXiv preprint arXiv:2409.10141, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 150, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 150, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 150, + 553, + 205 + ], + "type": "text", + "content": "[38] Ruilong Li, Shan Yang, David A Ross, and Angjoo Kanazawa. Ai choreographer: Music conditioned 3d dance generation with aist++. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13401-13412, 2021. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 206, + 553, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 260 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 260 + ], + "type": "text", + "content": "[39] Zhe Li, Zerong Zheng, Lizhen Wang, and Yebin Liu. Animatable gaussians: Learning pose-dependent gaussian maps for high-fidelity human avatar modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19711-19722, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 261, + 553, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 553, + 315 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 553, + 315 + ], + "type": "text", + "content": "[40] Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural actor: Neural free-view synthesis of human actors with pose control. ACM transactions on graphics (TOG), 40(6):1-16, 2021. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 316, + 553, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 553, + 370 + ], + "type": "text", + "content": "[41] Ziwei Liu, Ping Luo, Shi Qiu, Xiaogang Wang, and Xiaou Tang. Deepfashion: Powering robust clothes recognition and retrieval with rich annotations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1096-1104, 2016. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 372, + 553, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 553, + 414 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 553, + 414 + ], + "type": "text", + "content": "[42] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. ACM Transactions on Graphics, 34(6), 2015. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 415, + 553, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 553, + 458 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 553, + 458 + ], + "type": "text", + "content": "[43] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 460, + 553, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 553, + 513 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 553, + 513 + ], + "type": "text", + "content": "[44] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 514, + 553, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 514, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 316, + 514, + 553, + 569 + ], + "type": "text", + "content": "[45] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 5, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 571, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 571, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 571, + 553, + 635 + ], + "type": "text", + "content": "[46] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10975-10985, 2019. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 636, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 553, + 690 + ], + "type": "text", + "content": "[47] Georgios Pavlakos, Dandan Shan, Ilija Radosavovic, Angjoo Kanazawa, David Fouhey, and Jitendra Malik. Reconstructing hands in 3d with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9826-9836, 2024. 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[48] Hao-Yang Peng, Jia-Peng Zhang, Meng-Hao Guo, Yan-Pei Cao, and Shi-Min Hu. Charactergen: Efficient 3d character" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26317" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 77, + 73, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 294, + 105 + ], + "type": "text", + "content": "generation from single images with multi-view pose canonicalization. ACM Transactions on Graphics (TOG), 2024. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 173 + ], + "type": "text", + "content": "[49] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9054-9063, 2021. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "type": "text", + "content": "[50] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 294, + 262 + ], + "type": "text", + "content": "[51] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 319 + ], + "type": "text", + "content": "[52] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2304-2314, 2019. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 320, + 295, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 295, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 295, + 376 + ], + "type": "text", + "content": "[53] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 84-93, 2020. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 376, + 294, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 294, + 410 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 294, + 410 + ], + "type": "text", + "content": "[54] István Sárándi and Gerard Pons-Moll. Neural localizer fields for continuous 3d human pose and shape estimation. arXiv preprint arXiv:2407.07532, 2024. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 411, + 294, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 466 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 466 + ], + "type": "text", + "content": "[55] Kaiyue Shen, Chen Guo, Manuel Kaufmann, Juan Jose Zarate, Julien Valentin, Jie Song, and Otmar Hilliges. X-avatar: Expressive human avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16911-16921, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 467, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 294, + 510 + ], + "type": "text", + "content": "[56] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023. 6, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 512, + 294, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 294, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 294, + 567 + ], + "type": "text", + "content": "[57] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. In European Conference on Computer Vision, pages 1-18. Springer, 2025. 2, 6, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 568, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 294, + 635 + ], + "type": "text", + "content": "[58] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent video diffusion. In European Conference on Computer Vision, pages 439-457. Springer, 2025. 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "text", + "content": "[59] Qilin Wang, Zhengkai Jiang, Chengming Xu, Jiangning Zhang, Yabiao Wang, Xinyi Zhang, Yun Cao, Weijian Cao, Chengjie Wang, and Yanwei Fu. Vividpose: Advancing stable video diffusion for realistic human image animation. arXiv preprint arXiv:2405.18156, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "type": "text", + "content": "[60] Zhenzhen Weng, Jingyuan Liu, Hao Tan, Zhan Xu, Yang Zhou, Serena Yeung-Levy, and Jimei Yang. Template-free" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "single-view 3d human digitalization with diffusion-guided lrm. arXiv preprint arXiv:2401.12175, 2024. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "text", + "content": "[61] Kailu Wu, Fangfu Liu, Zhihan Cai, Runjie Yan, Hanyang Wang, Yating Hu, Yueqi Duan, and Kaisheng Ma. Unique3d: High-quality and efficient 3d mesh generation from a single image. arXiv preprint arXiv:2405.20343, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 141, + 555, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 555, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 555, + 196 + ], + "type": "text", + "content": "[62] Menghua Wu, Hao Zhu, Linjia Huang, Yiyu Zhuang, Yuanxun Lu, and Xun Cao. High-fidelity 3d face generation from natural language descriptions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4521-4530, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 198, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 274 + ], + "type": "text", + "content": "[63] Zhangyang Xiong, Chenghong Li, Kenkun Liu, Hongjie Liao, Jianqiao Hu, Junyi Zhu, Shuliang Ning, Lingteng Qiu, Chongjie Wang, Shijie Wang, et al. Mvhumannet: A large-scale dataset of multi-view daily dressing human captures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19801-19811, 2024. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 276, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 553, + 330 + ], + "type": "text", + "content": "[64] Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J Black. Icon: Implicit clothed humans obtained from normals. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13286-13296. IEEE, 2022. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 332, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 553, + 387 + ], + "type": "text", + "content": "[65] Yuliang Xiu, Jinlong Yang, Xu Cao, Dimitrios Tzionas, and Michael J. Black. ECON: Explicit Clothed humans Optimized via Normal integration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 388, + 553, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 442 + ], + "type": "text", + "content": "[66] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 445, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 445, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 445, + 553, + 510 + ], + "type": "text", + "content": "[67] Zhongcong Xu, Jianfeng Zhang, Jun Hao Liew, Hanshu Yan, Jia-Wei Liu, Chenxu Zhang, Jiashi Feng, and Mike Zheng Shou. Magicanimate: Temporally consistent human image animation using diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1481–1490, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 512, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 512, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 512, + 553, + 567 + ], + "type": "text", + "content": "[68] Tao Yu, Zerong Zheng, Kaiwen Guo, Pengpeng Liu, Qionghai Dai, and Yebin Liu. Function4d: Real-time human volumetric capture from very sparse consumer rgbd sensors. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5746-5756, 2021. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 568, + 553, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 553, + 623 + ], + "type": "text", + "content": "[69] Wangbo Yu, Li Yuan, Yan-Pei Cao, Xiangjun Gao, Xiaoyu Li, Wenbo Hu, Long Quan, Ying Shan, and Yonghong Tian. Hifi-123: Towards high-fidelity one image to 3d content generation. In European Conference on Computer Vision, pages 258–274. Springer, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 624, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 689 + ], + "type": "text", + "content": "[70] Xianggang Yu, Mutian Xu, Yidan Zhang, Haolin Liu, Chongjie Ye, Yushuang Wu, Zizheng Yan, Chenming Zhu, Zhangyang Xiong, Tianyou Liang, et al. Mvmgnet: A large-scale dataset of multi-view images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9150-9161, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "text", + "content": "[71] Zhixuan Yu, Jae Shin Yoon, In Kyu Lee, Prashanth Venkatesh, Jaesik Park, Jihun Yu, and Hyun Soo Park." + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26318" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "type": "text", + "content": "Humbi: A large multiview dataset of human body expressions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2990-3000, 2020. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 296, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 296, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 296, + 173 + ], + "type": "text", + "content": "[72] Jingbo Zhang, Xiaoyu Li, Qi Zhang, Yanpei Cao, Ying Shan, and Jing Liao. Humanref: Single image to 3d human generation via reference-guided diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1844–1854, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 295, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 295, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 295, + 229 + ], + "type": "text", + "content": "[73] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 295, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 295, + 284 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 295, + 284 + ], + "type": "text", + "content": "[74] Weitian Zhang, Yichao Yan, Yunhui Liu, Xingdong Sheng, and Xiaokang Yang. E3gen: Efficient, expressive and editable avatars generation. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 6860-6869, 2024. 3, 4, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 285, + 295, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 285, + 295, + 341 + ], + "spans": [ + { + "bbox": [ + 56, + 285, + 295, + 341 + ], + "type": "text", + "content": "[75] Yang Zhang, Jiaxi Gu, Li-Wen Wang, Han Wang, Junqi Cheng, Yuefeng Zhu, and Fangyuan Zou. Mimic motion: High-quality human motion video generation with confidence-aware pose guidance. arXiv preprint arXiv:2406.19680, 2024.3, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 342, + 295, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 342, + 295, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 342, + 295, + 386 + ], + "type": "text", + "content": "[76] Zechuan Zhang, Li Sun, Zongxin Yang, Ling Chen, and Yi Yang. Global-correlated 3d-decoupling transformer for clothed avatar reconstruction. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 388, + 295, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 295, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 295, + 431 + ], + "type": "text", + "content": "[77] Zechuan Zhang, Li Sun, Zongxin Yang, Ling Chen, and Yi Yang. Global-correlated 3d-decoupling transformer for clothed avatar reconstruction. Advances in Neural Information Processing Systems, 36, 2024. 3, 6, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 433, + 295, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 433, + 295, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 433, + 295, + 487 + ], + "type": "text", + "content": "[78] Zechuan Zhang, Zongxin Yang, and Yi Yang. Sifu: Side-view conditioned implicit function for real-world usable clothed human reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9936-9947, 2024. 2, 3, 6, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 488, + 295, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 488, + 295, + 543 + ], + "spans": [ + { + "bbox": [ + 56, + 488, + 295, + 543 + ], + "type": "text", + "content": "[79] Zerong Zheng, Tao Yu, Yixuan Wei, Qionghai Dai, and Yebin Liu. Deephuman: 3d human reconstruction from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7739-7749, 2019. 2, 3, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 544, + 295, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 544, + 295, + 599 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 295, + 599 + ], + "type": "text", + "content": "[80] Zerong Zheng, Tao Yu, Yebin Liu, and Qionghai Dai. Pamir: Parametric model-conditioned implicit representation for image-based human reconstruction. IEEE transactions on pattern analysis and machine intelligence, 44(6): 3170-3184, 2021. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 601, + 295, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 295, + 655 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 295, + 655 + ], + "type": "text", + "content": "[81] Shenhao Zhu, Junming Leo Chen, Zuozhuo Dai, Yinghui Xu, Xun Cao, Yao Yao, Hao Zhu, and Siyu Zhu. Champ: Controllable and consistent human image animation with 3d parametric guidance. In European Conference on Computer Vision (ECCV), 2024. 3, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 657, + 295, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 657, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 657, + 295, + 690 + ], + "type": "text", + "content": "[82] Yiyu Zhuang, Hao Zhu, Xusen Sun, and Xun Cao. Mofanerf: Morphable facial neural radiance field. In European conference on computer vision, pages 268-285. Springer, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "type": "text", + "content": "[83] Yiyu Zhuang, Qi Zhang, Ying Feng, Hao Zhu, Yao Yao, Xiaoyu Li, Yan-Pei Cao, Ying Shan, and Xun Cao. Anti-aliased" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 185 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 335, + 72, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 72, + 553, + 95 + ], + "type": "text", + "content": "neural implicit surfaces with encoding level of detail. In SIGGRAPH Asia 2023 Conference Papers, pages 1-10, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "text", + "content": "[84] Yiyu Zhuang, Qi Zhang, Xuan Wang, Hao Zhu, Ying Feng, Xiaoyu Li, Ying Shan, and Xun Cao. Neai: A pre-convoluted representation for plug-and-play neural ambient illumination. arXiv preprint arXiv:2304.08757, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "type": "text", + "content": "[85] Yiyu Zhuang, Yuxiao He, Jiawei Zhang, Yanwen Wang, Ji-ahe Zhu, Yao Yao, Siyu Zhu, Xun Cao, and Hao Zhu. Towards native generative model for 3d head avatar. arXiv preprint arXiv:2410.01226, 2024. 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "26319" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/IDProtector_ An Adversarial Noise Encoder to Protect Against ID-Preserving Image Generation/cbe88f8c-13ab-4200-b1b9-bb453b62ad1d_content_list.json b/2025/IDProtector_ An Adversarial Noise Encoder to Protect Against ID-Preserving Image Generation/cbe88f8c-13ab-4200-b1b9-bb453b62ad1d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1f44d5ec3c068251b31b40bdd9ddb583f7d93cfa --- /dev/null +++ b/2025/IDProtector_ An Adversarial Noise Encoder to Protect Against ID-Preserving Image Generation/cbe88f8c-13ab-4200-b1b9-bb453b62ad1d_content_list.json @@ -0,0 +1,1359 @@ +[ + { + "type": "text", + "text": "IDProtector: An Adversarial Noise Encoder to Protect Against ID-Preserving Image Generation", + "text_level": 1, + "bbox": [ + 109, + 128, + 888, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yiren Song* Pei Yang* Hai Ci† Mike Zheng Shou† Show Lab, National University of Singapore", + "bbox": [ + 282, + 202, + 720, + 241 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ac85b335f5891fd7e0110dce65aa9951d2ad253be68a637e5b245d66912b69e2.jpg", + "image_caption": [ + "Figure 1. Using only a single reference image, Encoder-based ID-preserving methods can generate realistic portraits, posing serious risks for malicious uses such as generation of malicious images. To prevent unauthorized ID-preserving generation, we propose IDProtector, which adds small perturbations to images. The protected images could disrupt ID-preserving generation by misleading the customization models to generate a dissimilar face, thereby achieving protection." + ], + "image_footnote": [], + "bbox": [ + 96, + 308, + 890, + 505 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 585, + 326, + 602 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, zero-shot methods like InstantID have revolutionized identity-preserving generation. Unlike multi-image finetuning approaches such as DreamBooth, these zero-shot methods leverage powerful facial encoders to extract identity information from a single portrait photo, enabling efficient identity-preserving generation through a single inference pass. However, this convenience introduces new threats to the facial identity protection. This paper aims to safeguard portrait photos from unauthorized encoder-based customization. We introduce IDProtector, an adversarial noise encoder that applies imperceptible adversarial noise to portrait photos in a single forward pass. Our approach offers universal protection for portraits against multiple state-of-the-art encoder-based methods, including InstantID, IP-Adapter, and PhotoMaker, while ensuring robustness to common image transformations such as JPEG", + "bbox": [ + 86, + 618, + 485, + 861 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "compression, resizing, and affine transformations. Experiments across diverse portrait datasets and generative models reveal that IDProtector generalizes effectively to unseen data and even closed-source proprietary models. Project page: https://github.com/showlab/IDProtector.", + "bbox": [ + 511, + 587, + 908, + 662 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 694, + 645, + 709 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advancements in diffusion models have opened new avenues for high-fidelity ID-preserving generation, widely used in applications such as personalized avatars. These methods can be categorized into tuning-based and encoder-based approaches. While tuning-based methods like Dreambooth[17] and LoRA[7] are effective, they require finetuning across multiple photos, which can be inconvenient. Encoder-based techniques, such as InstantID[26] and IP-Adapter[27], leveraging a face encoder to extract facial identity features in a zero-shot way, eliminating the need for finetuning and significantly simplifying the customization process. However, these advancements also", + "bbox": [ + 511, + 720, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 107, + 876, + 220, + 888 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 109, + 888, + 235, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3019", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "intensify threats to facial privacy and security, making the protection of personal identity information in portrait photos increasingly critical. In previous research, AntiDreamBooth[24] utilized PGD[12] optimization against finetuning-based methods like LoRA and Dreambooth. Yet, there remains a gap in the literature regarding attacks on encoder-based ID-preserving generation.", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Encoder-based ID preservation technology evolves from domain-agnostic encoders to domain-specific ones. Initially, methods like Face0[23] and IP-Adapter[27] used a general CLIP[16] image encoder to extract facial features, yet they struggled to maintain the identity fidelity. IP-Adapter-Plus[27] addressed this issue by utilizing patch features before the projection layer. InstantID[26] introduced the use of a pre-trained Arcface[2] encoder, significantly enhancing identity consistency and influencing subsequent works like Photomaker [9] and PULID [4]. A natural progression is to attack both the CLIP image encoder and Arcface encoder simultaneously, targeting all current Encoder-based ID preservation technologies.", + "bbox": [ + 89, + 199, + 480, + 396 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "There are four major challenges in ID protection: 1. Universality. Given the numerous existing encoder-based ID-preserving methods, we believe that universal protection is necessary, as attackers can easily switch to other methods to bypass protection. 2. Efficiency. Previous work typically relies on the Projected Gradient Descent (PGD) [12] to generate adversarial noise. However, PGD requires per-image optimization, taking several minutes per image, which significantly limits its practicality for protecting large volumes of portrait photos on social media in real world. 3. Robustness. The adversarial noise must be robust enough to withstand common image transformations, such as cropping, resizing, and JPEG compression, while also countering the inherent pre-processing operations of identity-preserving methods, such as face alignment. 4. Imperceptibility. The adversarial noise added for ID protection should be imperceptible to human eyes.", + "bbox": [ + 89, + 398, + 482, + 655 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In response to these challenges, this paper presents ID-Protector, the first feed-forward method to protect portrait photos from state-of-the-art encoder-based identity preservation techniques. Our approach incorporates several key innovations. For efficiency, we propose a ViT-based encoder that directly predicts adversarial noise for a given target portrait. To ensure universality, we carefully examine the attack surfaces and define a composite loss function tailored to four mainstream encoder-based generation methods, including InstantID [26], IP-Adapter [27], IP-Adapter Plus [27], and PhotoMaker [9]. Additionally, we employ an L1-based image regularization to minimize the visual impact of adversarial noise on image quality. For robustness against common transformations, a typical approach involves a combination of post-hoc differentiable augmentations like differentiable JPEG and cropping during train", + "bbox": [ + 89, + 659, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ing. However, we find that this method introduces a substantial learning burden. To address this, we propose a surrogate method that leverages the face preprocessing pipeline of existing identity-preserving methods. Specifically, we introduce small random noise into the affine transformation matrix during face alignment stage to achieve data augmentation. This not only enhances the robustness of adversarial noise against common image transformations but also enables it to withstand image preprocessing operations during identity-preserving generation.", + "bbox": [ + 511, + 90, + 903, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We train and extensively evaluate IDProtector on the CelebA dataset [11]. We further test it on unseen data VGG Face [15] and proprietary models like Midjourney [13] and Jing Gou, demonstrating its strong generalization to unseen data and closed-source models. By advancing adversarial attack techniques from theoretical constructs to practically deployable solutions, IDProtector sets a new benchmark in facial privacy protection. We summarize our main contributions as follows:", + "bbox": [ + 511, + 242, + 903, + 377 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce IDProtector, the first feed-forward method to protect portrait photos from SOTA encoder-based identity preservation techniques.", + "- We present several innovative techniques including a ViT-based noise encoder, a joint learning framework with carefully designed objectives, and affine-based data augmentation to simultaneously achieve the EURI standard (Efficiency, Universality, Robustness, and Imperceptibility)." + ], + "bbox": [ + 511, + 378, + 903, + 515 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 531, + 661, + 546 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Tuning-based Customized Generation", + "text_level": 1, + "bbox": [ + 511, + 556, + 841, + 571 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With the rapid advancement of diffusion models, customized generation methods have begun to emerge. Techniques like DreamBooth [17], Custom Diffusion [8], Low-Rank Adaptation (LoRA) [7] and Textual Inversion [3] facilitate the adaptation of text-to-image diffusion models through fine-tuning strategies that range from adjusting all parameters to optimizing specific word vectors. Despite their impressive effects, these methods require extensive computational resources and time due to their need for multi-image fine-tuning. This can make them impractical for real-time or large-scale applications.", + "bbox": [ + 511, + 579, + 903, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Encoder-based ID-Preserving Generation", + "text_level": 1, + "bbox": [ + 511, + 757, + 867, + 772 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Encoder-based Customized generation methods[23, 27, 29, 30] utilize pre-trained encoders to extract appearance feature, facilitating zero-shot generation in real-time. These methods evolved from domain-agnostic encoders to domain-specific encoders. Initially, CLIP[16] image encoder was used to extract facial features, such as with Face0[23] and IP-Adapter[27]; however, their maintenance of facial detail fidelity was not satisfactory. IP-Adapter", + "bbox": [ + 511, + 779, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3020", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Plus addressed this issue by utilizing patch features before the projection layer. InstantID[26] was the first to propose using a pre-trained Arcface[2] Encoder as a condition, significantly enhancing identity consistency and influencing subsequent works like Photomaker[9] and PULID[4]. StoryMaker[32] employs both the Arcface image encoder and the CLIP image encoder to encode facial features and broader regions, respectively, preserving not only facial identities but also clothing, hairstyles, and body features. Additionally, encoder-based ID preservation technology is widely applied in multi-human image personalization[5, 14] and video tasks[6].", + "bbox": [ + 89, + 90, + 480, + 272 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Defensive Perturbations for Diffusion Models.", + "text_level": 1, + "bbox": [ + 89, + 281, + 478, + 296 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The focus on safeguarding personal artworks and facial images from unauthorized usage in the fine-tuning of diffusion models has led to the development of subtle yet effective perturbations. As summarized in Tab. 1, distinct approaches have been tailored to address this issue: Glaze [20] strategically adjusts the feature-space distances to misdirect the style learning process in Stable Diffusion, effectively preventing style appropriation. Conversely, AdvDM [10] implements adversarial disruptions by maximizing the Mean Squared Error during the optimization phase to shield personal visuals. Meanwhile, Anti-DreamBooth [24] integrates adversarial interference directly within the DreamBooth fine-tuning cycle, applying a complex optimization strategy to formulate defensive perturbations. Further investigations [21, 25, 28, 31] into adversarial protection have demonstrated their effectiveness in obstructing finetuning-based customization methods such as DreamBooth [17], LoRA [7], and textual inversion [3]. However, how to protect personal identity from the recently popular encoder-based methods remains unexplored. This paper endeavors to bridge the current gaps by proposing advanced methodologies to shield images from encoder-based exploitations, enhancing the security measures for personal digital assets.", + "bbox": [ + 89, + 303, + 483, + 650 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/d95480a08f9e1977e0e7c4f33c2b3dbc14564883e5e299e0fa20a84b3227fb65.jpg", + "table_caption": [ + "Table 1. Overview of encoder-based methods for ID-preserving generation." + ], + "table_footnote": [], + "table_body": "
EncoderID-Preserving Methods
CLIP [16]IP-Adapter [27], IP-Adapter Plus [27], PhotoMaker [9], Face0 [23]
ArcFace [2]InstantID [26], PhotoMaker [9], PuLID [4], IP-Adapter-FaceID [27]
CLIP & ArcFaceIDAdapter [1], Storymaker [32]
", + "bbox": [ + 93, + 691, + 480, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Problem Definition", + "text_level": 1, + "bbox": [ + 89, + 768, + 274, + 784 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We consider two participants in a portrait identity protection scenario.: \"image protector\" Alice and \"image exploiter\" Bob. Bob seeks to use an encoder-based, identity-preserving approach to alter the portrait photos Alice shares on social media for malicious purposes. While Alice aims to protect her portrait by adding adversarial noise to thwart Bob's methods, inducing distortions or significant decreases", + "bbox": [ + 89, + 794, + 480, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "in ID similarity in the generated content. Specifically, we explain the workflow of the two parties as follows:", + "bbox": [ + 511, + 90, + 903, + 119 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Image Protector Alice: The image protector aims to provide protection for images against exploitation by encoder-based ID-preserving methods. In this context, the chosen protection method involves adding imperceptible protective perturbations to the images, with the goal of providing protection while minimizing modifications to the original image. In real-world scenarios, image protectors encounter several challenges. For instance, implementing universal protection to guard against various encoder-based, identity-preserving methods is difficult. Additionally, the protection should be robust to natural distortions such as resizing, JPEG compression, and blurring.", + "2. Image Exploiter Bob: Bob aims to use portrait photos for ID-preserving customization. Bob can choose any open-source or proprietary encoder-based models, which poses significant challenges on the universality of the protection method." + ], + "bbox": [ + 511, + 121, + 903, + 376 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Objective of the Study: Given an image $x$ containing a face portrait, our objective is to generate adversarial perturbations $\\delta$ that protect against ID-preserving generation. We aim to maximize the difference in identity consistency between the unperturbed generated image $g(x)$ and the perturbed generated image $g(x + \\delta)$ , where $g$ is an ID-preserving generator. To unify the protection against different models, we can find a $\\delta$ that simultaneously disrupts several feature spaces, lowering the sum of the similarity values. Thus, we train a noise encoder network $E_{\\theta}$ , solving", + "bbox": [ + 511, + 378, + 906, + 529 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} S \\left\\{g _ {i} (x), g _ {i} (x + \\delta) \\right\\}, \\quad \\delta = E _ {\\theta} (x) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 540, + 903, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "subject to the constraint that $\\| \\delta \\|_{\\infty} < \\epsilon$ . Here, $S$ denotes an identity similarity metric, which we evaluate using cosine similarity in feature spaces of ArcFace [2] and CLIP [16].", + "bbox": [ + 511, + 573, + 903, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 511, + 632, + 604, + 647 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The core of our method is an adversarial noise encoder model that can add adversarial noise to facial images for protection. We will first introduce the complete workflow in Sec. 4.1, followed by the design of the adversarial noise encoder in Sec. 4.2. Sec. 4.3 and 4.4 introduce ID protection objectives and loss functions. Sec. 4.5 introduces the data augmentation approach to enhance the robustness.", + "bbox": [ + 511, + 657, + 903, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Overall Method", + "text_level": 1, + "bbox": [ + 511, + 772, + 671, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fig. 2 details the pipeline of our method. For an arbitrary-sized input image $x$ containing a face, it is resized to $224 \\times 224$ and sent into the adversarial noise encoder model, whose output is an additive adversarial noise that can be resized back and perturbed onto the original image for protection. During adversarial noise encoder training, this protected image is further passed through face feature extrac", + "bbox": [ + 511, + 794, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3021", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ee7bfd173de19bfe5dea26574c7db9c30d07cbe1054171e65fd06d3498026cb6.jpg", + "image_caption": [ + "Figure 2. Overall schematics of our method. Our method's key design includes the noise encoder, loss functions, and the gradient optimization path that allows backpropagation. During the protection phase, an image is first resized to $224 \\times 224$ and sent to an adversarial noise encoder model. The model outputs a perturbation that can be resized back and added to the image, achieving an adversarial attack effect – the protected image prevents subsequent pipelines from correctly extracting facial feature information, thereby hindering face similarity in face generation tasks. This achieves the goal of ID protection." + ], + "image_footnote": [], + "bbox": [ + 101, + 92, + 897, + 284 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tion pipelines to obtain the loss functions that aim at creating maximum differences in the feature space.", + "bbox": [ + 89, + 385, + 482, + 417 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Noise Encoder", + "text_level": 1, + "bbox": [ + 89, + 424, + 238, + 439 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The network takes a $224 \\times 224$ RGB image as input and outputs three-channel adversarial perturbations. The perturbations are projected to a range of $[-1, 1]$ , then denormalized $[- \\epsilon, \\epsilon]$ , resized to the original image dimensions, and added to the image. We utilize a Vision Transformer (ViT) to generate the adversarial noise, with both input and output dimensions set to $224 \\times 224$ . Empirically, we find that adding an additional prior mask channel, indicating the location of the face, aids in training. This face localization mask is generated by the InsightFace pipeline based on facial landmark points, which specify the region of the face within the image. The mask is concatenated to the image as the fourth input channel, eliminating the need for the network to learn face localization ability from scratch, thus reducing training difficulty.", + "bbox": [ + 89, + 446, + 483, + 674 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. Adversarial ID Protection", + "text_level": 1, + "bbox": [ + 89, + 681, + 326, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since generative networks' only knowledge of faces comes from feature embeddings, to achieve ID protection, we can make the extracted features from protected images significantly different from the originals, thereby preventing generation models from accessing correct facial features. Based on this principle, we analyze the specific pipelines of InstantID, IP-Adapter, IP-Adapter-Plus, and PhotoMaker to design targeted attack objectives and loss functions.", + "bbox": [ + 89, + 703, + 482, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "InstantID obtains facial features in two steps, as shown in Fig. 2. It first aligns the face to a pre-defined location, before feeding the aligned face into an ArcFace feature extractor to obtain facial information. We choose to fool the ArcFace model, minimizing the cosine similarity between", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the ArcFace features of the original v.s. protected images.", + "bbox": [ + 511, + 385, + 893, + 400 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "IP-Adapter, IP-Adapter-Plus, and PhotoMaker all rely on facial features extracted by the CLIP vision encoder, albeit using different versions of CLIP vision. IP-Adapter and PhotoMaker utilize CLIP Vision's output, while IP-Adapter-Plus takes features before CLIP Vision's second-to-last layer. As illustrated in Fig. 2, embeddings from different layers offer various attack surfaces. We select victim embeddings for attack according to three principles: (1) blocking all potential pathways, ensuring that any information flow from left to right passes through at least one victim embedding, ensuring all information flow would be disrupted; (2) choosing features as early as possible in the network (towards the left in Fig. 2), to reduce the backpropagation path length and thus simplify optimization; and (3) targeting embeddings with dense semantic information for more effective manipulation.", + "bbox": [ + 511, + 401, + 903, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "An example of an unsuitable target is the last hidden layer feature of CLIP Vision. While modifying this feature could influence all three models, its huge dimension $(257\\times 1280)$ and lack of semantic density make it highly challenging for optimization to converge. Ultimately, we selected the embeddings marked red in Fig. 2 as the primary attack targets. The objective, similar to that of InstantID, is to maximize the cosine similarity between the perturbed embedding and the original embedding, effectively aligning them post-attack. Therefore, the final adversarial loss is a weighted average of all losses, given by", + "bbox": [ + 511, + 645, + 906, + 811 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {a d v}} = \\sum_ {\\text {C u t s o m M o d e l} i} \\alpha_ {i} \\cdot \\operatorname {c o s s i m} \\left(e _ {i}, e _ {i} ^ {\\prime}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 824, + 903, + 858 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $e_i$ and $e_i'$ denote the face embeddings of the clean and protected images, respectively.", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3022", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.4. Imperceptibility", + "text_level": 1, + "bbox": [ + 89, + 90, + 250, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To minimize the visual impact on the image quality, we enforce an $\\ell_1$ regularization on the predicted adversarial noise $\\delta$ . We further introduce an auxiliary penalty on any noise values exceeding the $\\epsilon$ -ball boundary. Together, these terms constitute our regularization loss, given by:", + "bbox": [ + 89, + 112, + 483, + 189 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {r e g}} = \\beta_ {1} \\cdot \\| \\delta \\| _ {1} + \\beta_ {2} \\cdot \\| \\delta - \\operatorname {c l i p} _ {\\pm \\epsilon} (\\delta) \\| _ {1}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 200, + 482, + 219 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The final loss is thus the sum of adversarial loss and regularization, given by:", + "bbox": [ + 89, + 229, + 482, + 261 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL = L _ {\\mathrm {a d v}} + L _ {\\mathrm {r e g}}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 227, + 273, + 482, + 292 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.5. Robustness", + "text_level": 1, + "bbox": [ + 89, + 303, + 215, + 318 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Fig. 2, the InstantID pipeline involves a step of facial alignment aiming to align any faces to a predefined standard face for better feature extraction. The alignment process includes detecting the face and key points $\\tilde{p}$ , calculating the affine matrix $A$ with standard face key points $\\tilde{p}'$ , and aligning them. This process is computationally expensive. To reduce computational cost and simplify training, we pre-compute $A$ for each image and keep it fixed throughout training. However, during inference, perturbations carried by protected images may slightly alter the precomputed affine matrix $A$ , leading to minor misalignment that weakens the effectiveness the adversarial perturbations.", + "bbox": [ + 89, + 325, + 483, + 507 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To adapt to this misalignment as well as being computationally efficient, we leverage the existing affine transformation operation and introduce small gaussian noise to $A$ during training to simulate face transformations. The noise level is empirically set to $\\sigma = 0.003$ . The face alignment transformation is therefore given by", + "bbox": [ + 89, + 508, + 483, + 599 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {p} ^ {\\prime} = (A + N (0, \\sigma I)) \\tilde {p}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 611, + 482, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We incorporate this augmentation only in the InstantID branch. Our experiments reveal that it also enhances the robustness of other branches, being resilient to various common image transformations.", + "bbox": [ + 89, + 641, + 483, + 702 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 715, + 223, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 741, + 281, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Data. For practicality, we selected 190,000 uncropped images from the CelebA dataset [11] as the training set. These images have arbitrary sizes and aspect ratios. The test set includes two subsets: 50 unseen images from CelebA and 50 images from VGG Face [15]. To ensure compatibility with baseline methods, these images were resized to $512 \\times 512$ and $224 \\times 224$ , respectively.", + "bbox": [ + 89, + 763, + 482, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Details. We trained a ViT-S/8 adversarial noise encoder from scratch using Eq. 4 as the loss function. We", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/8f707b0926e987183a8e611484f9870d285499e382d7547c32998531c62ff952.jpg", + "table_caption": [ + "Table 2. Curriculum learning stages with corresponding hyperparameters. $\\eta$ represents learning rate. ${\\alpha }_{1}$ to ${\\alpha }_{4}$ are weights for IP-Adapter,IP-Adapter-Plus,PhotoMaker,InstantID,respectively." + ], + "table_footnote": [], + "table_body": "
StageEpochsα1α2α3α4β1β2εη
11202/3714/3720/371/379e-31e-30.051e-2
2204/162/169/162/161.8e-32e-40.042e-5
3202/142/149/141/144.5e-45e-50.0352e-5
", + "bbox": [ + 516, + 132, + 903, + 189 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "employed a three-stage curriculum learning approach, progressively shrinking the $\\epsilon$ -ball and balancing the different losses by adjusting the hyperparameters (presented in Table 2). Training began with a 2,500-step linear warm-up, gradually increasing the learning rate from zero. Gradient clipping was applied during training to cap the $\\ell_2$ norm of gradients at a maximum of 10. The model was trained with a batch size of 112 on four 80GB NVIDIA H100 GPUs for 10 days.", + "bbox": [ + 511, + 215, + 906, + 351 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation process and metrics. To evaluate the protection performance of different methods, for each face-containing image in the dataset (denoted as NP ref image), we apply different protection methods (including our ID-Protector) to generate its protected counterpart (denoted as P ref image). Throughout all experiments, we constrain the maximum pixel perturbation to 9/255 (an $\\epsilon$ -ball of 0.035), following common practice in adversarial machine learning [18]. Subsequently, both P and NP ref images are fed into customization generation models (such as InstantID) to obtain generated images (denoted as P/NP gen images).", + "bbox": [ + 511, + 352, + 908, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate the identify shift caused by different protection methods through Identity Score Matching (ISM), which is calculated as the cosine similarity between facial features of NP images and generated images extracted by ArcFace [2]. Lower ISM values indicate better protection as they suggest a greater identity shift between original (NP ref) and generated (P gen) faces.", + "bbox": [ + 511, + 518, + 908, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since protected images may result in no detectable faces being generated, we also report the Face Detection Rate (FDR) using InsightFace detector, as ISM can only be computed when faces are detected in both images. Additionally, to ensure comprehensive evaluation, we measure the perceptual quality of each generated face using SER-FIQ [22].", + "bbox": [ + 511, + 625, + 908, + 717 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Comparison with Baselines", + "text_level": 1, + "bbox": [ + 511, + 727, + 759, + 743 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Protection Performance. We compare our method with the following baseline methods: ACE [31], AdvDM [10], Anti-DreamBooth [24], PhotoGuard [19], and SimAC [25]. Most baselines aim to reduce the quality of generated images or hinder the diffusion denoising process, but we aim to reduce the similarity of generated identities so as to directly disrupt identity-preserving generation. To compare our method with baselines, we first applied each protection method to generate protected versions of test images. We then used both the original images (NP) and the protected", + "bbox": [ + 511, + 750, + 908, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3023", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/425ffa070c84b8fa4527206248c51af6132b7951f62e8a74bb4a3af856ef2bd6.jpg", + "image_caption": [ + "Figure 3. Qualitative comparison with baseline. Images protected by our method cannot be used to generate similar faces. Compared to baseline protection methods, under the same perturbation bound of $9/255$ ( $\\epsilon = 3.5\\%$ ), ours causes more significant facial differences." + ], + "image_footnote": [], + "bbox": [ + 98, + 89, + 898, + 597 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "versions as references in customized generation tasks with four models: IP-Adapter, IP-Adapter-Plus, PhotoMaker, and InstantID. Tab. 3 shows that, for the first three methods, while baseline methods reduced ISM by an average of 0.02, with a maximum reduction of 0.06, our method achieved an ISM reduction between 0.1 and 0.14, demonstrating a larger identity shift. On InstantID, our IDProtector demonstrated even more pronounced protection: when baseline methods reduced facial similarity by less than 0.1, our method reduced similarity by over 0.4 across both datasets. Additionally, FDR results validate the reliability of ISM as FDR reflects the percentage of images with a face detected and contributes to ISM calculations. Overall, these results underscore the superior identity protection performance of our method compared to existing approaches. Together, these highlight the superior identity protection performance of", + "bbox": [ + 88, + 655, + 486, + 898 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "our approach.", + "text_level": 1, + "bbox": [ + 511, + 656, + 607, + 670 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following previous study [24], we evaluate SER-FIQ as a quality metric of generated (P/NP gen) images. Although our method has a noticeable impact on image generation quality (as shown in Fig. 3), we observed inconsistencies between the SER-FIQ scores and qualitative assessments. This discrepancy suggests that a more robust and reliable quality metric may be required to accurately measure the quality of generated images in future research.", + "bbox": [ + 511, + 672, + 906, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also adapted our method to a per-image tuning variant, employing PGD [12] as the optimization technique. We observed that PGD achieves a more substantial ISM reduction across most methods, which aligns with our expectations since per-image tuning can better leverage the specific characteristics of each individual image. However, a notable drawback of PGD is the prolonged optimization time,", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3024", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/25ccbf6300aa7b4ad2eb081878c462bb049d1bc1ab83d878ef1314cf1789c23e.jpg", + "table_caption": [ + "Table 3. Comparison of identity protection performance against baseline methods. Protection performance is evaluated against IP-Adapter, IP-Adapter-Plus, PhotoMaker, and InstantID. \"No Protect\" denotes direct use of the original (NP ref) images for identity-preserving generation. Experiments were conducted on five prompts per dataset, and average results are reported." + ], + "table_footnote": [], + "table_body": "
DatasetProtection MethodIP-AdapterIP-Adapter-PlusPhotoMakerInstantID
FDR ↓ISM ↓SER-FIQ ↓FDR ↓ISM ↓SER-FIQ ↓FDR ↓ISM ↓SER-FIQ ↓FDR ↓ISM ↓SER-FIQ ↓
CelebANo Protect1.0000.1760.6411.0000.2400.6630.6640.2650.6050.9960.7750.774
ACE1.0000.1590.6401.0000.2310.6810.6360.2680.5941.0000.7450.758
AdvDM0.9880.1570.6301.0000.2260.6660.6400.2550.5901.0000.7570.757
Anti-DB1.0000.1550.6151.0000.2220.6760.6400.2550.5991.0000.7530.752
PhotoGuard1.0000.1520.6211.0000.2230.6700.6520.2590.5921.0000.7580.767
SimAC1.0000.1620.6271.0000.2170.6830.6520.2530.5991.0000.7520.755
Ours (PGD)0.8440.0240.5440.9120.0720.7180.1120.0940.6390.996-0.1660.721
Ours0.9520.0600.6880.9960.1140.6440.6720.1230.6221.0000.2310.613
VGG FaceNo Protect1.0000.2010.6721.0000.2360.6980.6320.2540.6000.9960.7640.828
ACE0.9880.1540.6171.0000.2090.6831.0000.2090.6831.0000.7030.829
AdvDM1.0000.1470.5971.0000.2150.6931.0000.2150.6931.0000.7060.820
Anti-DB1.0000.1380.5971.0000.2100.6780.6240.2380.5991.0000.7070.811
PhotoGuard1.0000.1590.6431.0000.2170.6860.6240.2290.6120.9960.6910.799
SimAC1.0000.1500.6051.0000.2020.6770.6240.2280.6441.0000.7070.826
Ours (PGD)0.9680.0430.6091.0000.2020.7400.6240.2280.6301.0000.7070.769
Ours1.0000.1050.7371.0000.1380.6870.6320.1260.6141.0000.2320.713
", + "bbox": [ + 91, + 131, + 906, + 359 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4728a2d2a721ad63d3a1496b09cd8ed8a7ebab09c6ffc07b68a9877e083ed7cb.jpg", + "table_caption": [ + "Table 4. Comparison of time cost and quality impact with an epsilon-budget of 9/255 (±3.5%). PSNR and SSIM are calculated between clean (NP ref) and perturbed (P ref) image pairs." + ], + "table_footnote": [], + "table_body": "
ACEAdvDMAnti-DBPhoto-GuardSimACOurs PGDOurs
Time (s)87.051.926123.332340890.173
PSNR (dB)31.0631.7031.7031.1032.3732.9432.15
SSIM (dB)0.7410.7660.7650.7570.7910.8450.809
", + "bbox": [ + 91, + 422, + 483, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "which we will address in the following section.", + "bbox": [ + 89, + 521, + 403, + 536 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The protection effects are visualized in Fig. 3. The baseline methods can protect against some of the generation methods. For example, AdvDM and Anti-DreamBooth cause IP-Adapter to produce faces with color block artifacts, while several outputs from IP-Adapter-Plus exhibit unnatural patterns and shadows. However, the baselines are ineffective in protecting against PhotoMaker and InstantID. With IDProtector's protection, faces generated by all four methods appear significantly altered. In some cases, even the gender of the person is unidentifiable. This fulfills our protection goal: protected images could not be used for identity-preserving generation.", + "bbox": [ + 88, + 537, + 482, + 718 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Time and Image Quality. Our IDProtector is fast to protect an image with less degradation to image quality. As compared in Table 4, the PSNR and SSIM of our method are higher than the baselines, indicating less visual impact. The baseline methods require an average of 23.3 to 323 seconds per image for protection. Using PGD with our optimization objective would take over an hour on average. In contrast, our ViT-based IDProtector achieves an average protection time of only 0.173 seconds per image – less than $1\\%$ of the fastest baseline. Note that the forward-feeding ViT supports better parallelization compared with PGD, the time cost could further decrease with mini-batch inference.", + "bbox": [ + 88, + 719, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6b7fa9b246726fc0429b01ae5ba2881438458361e14219e35618ddd5dbcc65f0.jpg", + "table_caption": [ + "Table 5. Generalization of ID protection by our IDProtector on unseen generators. IPA-FID+/P/XL denotes IP-Adapter-FaceID/Plus/Portrait/SDXL, with NP and P indicating no and with protection, respectively." + ], + "table_footnote": [], + "table_body": "
IPA-FIDIPA-FID+IPA-FIDPIPA-FIDXLFlux-IPAStory-MakerMid-journeyJingGou
ISM ↓NP0.5650.6440.1290.5230.0540.5720.1990.672
P0.1530.2170.0520.1320.0290.1930.0980.131
SER-NP0.7060.7510.7200.4020.7280.6990.5680.791
FIQ ↓P0.8030.7300.6580.6670.6730.7340.6270.805
", + "bbox": [ + 514, + 436, + 905, + 518 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8d3a17b59eb2baa69c8daf3c6d0e744ae0b33a82b77c500f6b1ee1f4b3685d83.jpg", + "image_caption": [ + "Figure 4. ID protection performance on unseen generators." + ], + "image_footnote": [], + "bbox": [ + 516, + 537, + 901, + 741 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Generalization to Unseen Data and Models", + "text_level": 1, + "bbox": [ + 511, + 786, + 879, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Generalization to unseen datasets. We evaluated our method on the VGG Face dataset that was unseen during training. All following experiments are conducted on the CelebA test set unless otherwise specified. As shown in Tab. 3, the ISM reduction achieved by our method shows minimal variation across datasets. For example, our IDPro", + "bbox": [ + 511, + 810, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3025", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/594b3bab479fbd02006c074c73f6d4e09375ed1ccffd0f8c7200886c151bd8ff.jpg", + "table_caption": [ + "Table 6. Robustness evaluation of IDProtector under image distortions. NP denotes non-protected images for identity-preserving generation, while P represents protected but non-distorted images, providing an ablation. Distortion averages exclude NP and P." + ], + "table_footnote": [], + "table_body": "
IDProtector (ours, ViT)Adaptive PGD
IP-AdapterIP-Adapter+PhotoMakerInstantIDIP-AdapterIP-Adapter+PhotoMakerInstantID
FDR ↓ISM ↓FDR ↓ISM ↓FDR ↓ISM ↓FDR ↓ISM ↓FDR ↓ISM ↓FDR ↓ISM ↓FDR ↓ISM ↓FDR ↓ISM ↓
NP1.0000.1761.0000.2400.6640.2650.9960.7750.8440.0240.9120.0720.1120.0940.996-0.166
P0.9520.0600.9960.1140.6720.1231.0000.2310.8440.0240.9120.0720.1120.0940.996-0.166
P+Affine1.0000.0681.0000.1060.6560.1351.0000.2160.8440.0640.9760.0810.112-0.0091.000-0.147
P+JPEG0.9640.0691.0000.0970.6600.1381.0000.2410.7840.0150.9480.0620.108-0.0121.000-0.161
P+Crop1.0000.0721.0000.1080.6520.1321.0000.2160.7520.0580.9400.0700.1080.0201.000-0.167
P+Noisy0.9520.0690.9840.1140.6560.1321.0000.2520.8320.0230.9320.0720.112-0.0101.000-0.130
Avg (P+)0.9790.0700.9960.1060.6560.1341.0000.2310.8030.0400.9490.0710.110-0.0031.000-0.151
", + "bbox": [ + 91, + 118, + 906, + 247 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8aa4b299e8c0047648a0d6a052476a39670289cf352eebd33d7d2302bbdf0968.jpg", + "image_caption": [ + "Figure 5. Robustness evaluation of IDProtector under various distortions. Despite geometric distortions like affine transformations, IDProtector effectively misleads all four models into generating dissimilar faces or even artifacts." + ], + "image_footnote": [], + "bbox": [ + 94, + 273, + 478, + 517 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tector reduces the ISM of IP-Adapter by approximately 0.1 and InstantID by around 0.55 on both datasets. This result suggests that our method generalises effectively beyond the CelebA dataset, demonstrating protection capabilities that are not restricted to in-domain faces.", + "bbox": [ + 89, + 607, + 483, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Generalization to unseen customization models. Beyond the four customization models seen during training, we tested protected images on a wider range of customization models. Across these models, our protection approach achieves an average ISM reduction of 0.31, with a maximum reduction of 0.54. Notably, our method remains effective against proprietary commercial models such as Midjourney (0.10 reduction) and Jing Gou (0.54 reduction), underscoring its ability to generalise to unseen customization models.", + "bbox": [ + 88, + 686, + 482, + 835 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "These protection effects are visualised in Fig. 4. As shown, protected images used in these generators cause noticeable quality degradation, including altered facial shapes (e.g., IP-Adapter-FaceID-XL), reduced prompt-following", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ability, and image contamination (e.g., Midjourney). The common result is a decreased resemblance between the generated face and the reference identity.", + "bbox": [ + 511, + 272, + 906, + 318 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Robustness Evaluation", + "text_level": 1, + "bbox": [ + 511, + 325, + 723, + 340 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our method demonstrates robustness against various distortions. For protected (P) images, we applied several distortions, including JPEG compression (quality=85), random cropping that removes up to $20\\%$ of the region, noise perturbation with a variance of 100, and an affine transformation. The affine transformation was implemented as $\\tilde{q} = (A + \\mathcal{N}(0,0.05^2))\\tilde{p}$ , where the matrix $A$ is an identity matrix. As shown in Tab. 6, the ISM values under these distortions remain comparable to those obtained without distortion, indicating that our protection remains effective under distortion. Fig. 5 provides a visual comparison, where even under various distortions, InstantID fails to reconstruct the correct female face, generating a male portrait instead. This demonstrates that the common distortions do not compromise the efficacy of our protection method.", + "bbox": [ + 511, + 348, + 906, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion and Limitation", + "text_level": 1, + "bbox": [ + 511, + 588, + 764, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce IDProtector, a method designed to protect facial identities from encoder-based identity-preserving generation. By training a noise encoder to add adversarial perturbations onto images, IDProtector can protect an image within 0.2s. Our method achieves significantly lower face similarity in generated images than existing protection methods. IDProtector also demonstrates strong practical applicability. Its protection remains effective under distortions such as affine transformations and JPEG compression. Furthermore, IDProtector not only shows effectiveness against closed-source models, such as Midjourney, but also unifies protection against multiple ID-preserving models, setting a new standard for privacy protection against generative AI.", + "bbox": [ + 511, + 613, + 906, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A limitation of our method is that the adversarial noise is not fully invisible, necessitating a trade-off between attack effectiveness and Imperceptibility. Future efforts will focus on enhancing the Imperceptibility of the noise while maintaining its robustness.", + "bbox": [ + 511, + 825, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3026", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 277, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 172, + 187, + 188 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Siying Cui, Jia Guo, Xiang An, Jiankang Deng, Yongle Zhao, Xinyu Wei, and Ziyong Feng. Idadapter: Learning mixed features for tuning-free personalization of text-to-image models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 950-959, 2024.", + "[2] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019.", + "[3] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022.", + "[4] Zinan Guo, Yanze Wu, Zhuowei Chen, Lang Chen, and Qian He. Pulid: Pure and lightning id customization via contrastive alignment. arXiv preprint arXiv:2404.16022, 2024.", + "[5] Junjie He, Yifeng Geng, and Liefeng Bo. Uniportrait: A unified framework for identity-preserving single- and multi-human image personalization. arXiv preprint arXiv:2408.05939, 2024.", + "[6] Xuanhua He, Quande Liu, Shengju Qian, Xin Wang, Tao Hu, Ke Cao, Keyu Yan, Man Zhou, and Jie Zhang. Id-animator: Zero-shot identity-preserving human video generation. arXiv preprint arXiv:2404.15275, 2024.", + "[7] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022.", + "[8] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023.", + "[9] Zhen Li, Mingdeng Cao, Xintao Wang, Zhongang Qi, Ming-Ming Cheng, and Ying Shan. Photomaker: Customizing realistic human photos via stacked id embedding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8640-8650, 2024.", + "[10] Chumeng Liang, Xiaoyu Wu, Yang Hua, Jiaru Zhang, Yiming Xue, Tao Song, Zhengui Xue, Ruhui Ma, and Haibing Guan. Adversarial example does good: Preventing painting imitation from diffusion models via adversarial examples. arXiv preprint arXiv:2302.04578, 2023.", + "[11] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Large-scale celebfaces attributes (celeba) dataset. Retrieved August, 15(2018):11, 2018." + ], + "bbox": [ + 93, + 199, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "This research is supported by the Ministry of Education, Singapore, under its Academic Research Fund Tier 2 (Award No: MOE-T2EP20124-0012).", + "[12] Aleksander Madry. Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083, 2017.", + "[13] Midjourney.com. Midjourney. https://www.midjourney.com/home.", + "[14] Daniil Ostashev, Yuwei Fang, Sergey Tulyakov, Kfir Aberman, et al. Moa: Mixture-of-attention for subject-context disentanglement in personalized image generation. arXiv preprint arXiv:2404.11565, 2024.", + "[15] Omkar Parkhi, Andrea Vedaldi, and Andrew Zisserman. Deep face recognition. In BMVC 2015-Proceedings of the British Machine Vision Conference 2015. British Machine Vision Association, 2015.", + "[16] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021.", + "[17] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023.", + "[18] Mehrdad Saberi, Vinu Sankar Sadasivan, Keivan Rezaei, Aounon Kumar, Atoosa Chegini, Wenxiao Wang, and Soheil Feizi. Robustness of ai-image detectors: Fundamental limits and practical attacks. arXiv preprint arXiv:2310.00076, 2023.", + "[19] Hadi Salman, Alaa Khaddaj, Guillaume Leclerc, Andrew Ilyas, and Aleksander Madry. Raising the cost of malicious ai-powered image editing. arXiv preprint arXiv:2302.06588, 2023.", + "[20] Shawn Shan, Jenna Cryan, Emily Wenger, Haitao Zheng, Rana Hanocka, and Ben Y Zhao. Glaze: Protecting artists from style mimicry by {Text-to-Image} models. In 32nd USENIX Security Symposium (USENIX Security 23), pages 2187-2204, 2023.", + "[21] Yiren Song, Shengtao Lou, Xiaokang Liu, Hai Ci, Pei Yang, Jiaming Liu, and Mike Zheng Shou. Anti-reference: Universal and immediate defense against reference-based generation. arXiv preprint arXiv:2412.05980, 2024.", + "[22] Philipp Terhorst, Jan Niklas Kolf, Naser Damer, Florian Kirchbuchner, and Arjan Kuijper. Ser-fiq: Unsupervised estimation of face image quality based on stochastic embedding robustness. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5651-5660, 2020.", + "[23] Dani Valevski, Danny Wasserman, Yossi Matias, and Yaniv Leviathan. Face0: Instantaneously conditioning a text-to-image model on a face. arXiv preprint arXiv:2306.06638, 2023.", + "[24] Thanh Van Le, Hao Phung, Thuan Hoang Nguyen, Quan Dao, Ngoc N Tran, and Anh Tran. Anti-dreambooth: Protecting users from personalized text-to-image synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2116-2127, 2023." + ], + "bbox": [ + 93, + 92, + 903, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3027", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Feifei Wang, Zhentao Tan, Tianyi Wei, Yue Wu, and Qidong Huang. Simac: A simple anti-customization method for protecting face privacy against text-to-image synthesis of diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12047-12056, 2024.", + "[26] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, and Anthony Chen. Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519, 2024.", + "[27] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023.", + "[28] Xiaoyu Ye, Hao Huang, Jiaqi An, and Yongtao Wang. Duaw: Data-free universal adversarial watermark against stable diffusion customization. arXiv preprint arXiv:2308.09889, 2023.", + "[29] Yuxuan Zhang, Yiren Song, Jiaming Liu, Rui Wang, Jinpeng Yu, Hao Tang, Huaxia Li, Xu Tang, Yao Hu, Han Pan, et al. Ssr-encoder: Encoding selective subject representation for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8069–8078, 2024.", + "[30] Yuxuan Zhang, Yiren Song, Jinpeng Yu, Han Pan, and Zhongliang Jing. Fast personalized text to image synthesis with attention injection. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 6195-6199. IEEE, 2024.", + "[31] Boyang Zheng, Chumeng Liang, Xiaoyu Wu, and Yan Liu. Understanding and improving adversarial attacks on latent diffusion model. arXiv preprint arXiv:2310.04687, 2023.", + "[32] Zhengguang Zhou, Jing Li, Huaxia Li, Nemo Chen, and Xu Tang. Storymaker: Towards holistic consistent characters in text-to-image generation. arXiv preprint arXiv:2409.12576, 2024." + ], + "bbox": [ + 91, + 92, + 480, + 583 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3028", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/IDProtector_ An Adversarial Noise Encoder to Protect Against ID-Preserving Image Generation/cbe88f8c-13ab-4200-b1b9-bb453b62ad1d_model.json b/2025/IDProtector_ An Adversarial Noise Encoder to Protect Against ID-Preserving Image Generation/cbe88f8c-13ab-4200-b1b9-bb453b62ad1d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7800d624f78ee56adac91691d5acf6ff9d7936d2 --- /dev/null +++ b/2025/IDProtector_ An Adversarial Noise Encoder to Protect Against ID-Preserving Image Generation/cbe88f8c-13ab-4200-b1b9-bb453b62ad1d_model.json @@ -0,0 +1,1793 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.11, + 0.13, + 0.89, + 0.177 + ], + "angle": 0, + "content": "IDProtector: An Adversarial Noise Encoder to Protect Against ID-Preserving Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.284, + 0.203, + 0.721, + 0.242 + ], + "angle": 0, + "content": "Yiren Song* Pei Yang* Hai Ci† Mike Zheng Shou† Show Lab, National University of Singapore" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.309, + 0.891, + 0.506 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.511, + 0.908, + 0.568 + ], + "angle": 0, + "content": "Figure 1. Using only a single reference image, Encoder-based ID-preserving methods can generate realistic portraits, posing serious risks for malicious uses such as generation of malicious images. To prevent unauthorized ID-preserving generation, we propose IDProtector, which adds small perturbations to images. The protected images could disrupt ID-preserving generation by misleading the customization models to generate a dissimilar face, thereby achieving protection." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.587, + 0.327, + 0.603 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.619, + 0.486, + 0.862 + ], + "angle": 0, + "content": "Recently, zero-shot methods like InstantID have revolutionized identity-preserving generation. Unlike multi-image finetuning approaches such as DreamBooth, these zero-shot methods leverage powerful facial encoders to extract identity information from a single portrait photo, enabling efficient identity-preserving generation through a single inference pass. However, this convenience introduces new threats to the facial identity protection. This paper aims to safeguard portrait photos from unauthorized encoder-based customization. We introduce IDProtector, an adversarial noise encoder that applies imperceptible adversarial noise to portrait photos in a single forward pass. Our approach offers universal protection for portraits against multiple state-of-the-art encoder-based methods, including InstantID, IP-Adapter, and PhotoMaker, while ensuring robustness to common image transformations such as JPEG" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.588, + 0.909, + 0.664 + ], + "angle": 0, + "content": "compression, resizing, and affine transformations. Experiments across diverse portrait datasets and generative models reveal that IDProtector generalizes effectively to unseen data and even closed-source proprietary models. Project page: https://github.com/showlab/IDProtector." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.695, + 0.646, + 0.71 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.721, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Recent advancements in diffusion models have opened new avenues for high-fidelity ID-preserving generation, widely used in applications such as personalized avatars. These methods can be categorized into tuning-based and encoder-based approaches. While tuning-based methods like Dreambooth[17] and LoRA[7] are effective, they require finetuning across multiple photos, which can be inconvenient. Encoder-based techniques, such as InstantID[26] and IP-Adapter[27], leveraging a face encoder to extract facial identity features in a zero-shot way, eliminating the need for finetuning and significantly simplifying the customization process. However, these advancements also" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.877, + 0.221, + 0.889 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.889, + 0.236, + 0.902 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.877, + 0.236, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3019" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "intensify threats to facial privacy and security, making the protection of personal identity information in portrait photos increasingly critical. In previous research, AntiDreamBooth[24] utilized PGD[12] optimization against finetuning-based methods like LoRA and Dreambooth. Yet, there remains a gap in the literature regarding attacks on encoder-based ID-preserving generation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.2, + 0.482, + 0.397 + ], + "angle": 0, + "content": "Encoder-based ID preservation technology evolves from domain-agnostic encoders to domain-specific ones. Initially, methods like Face0[23] and IP-Adapter[27] used a general CLIP[16] image encoder to extract facial features, yet they struggled to maintain the identity fidelity. IP-Adapter-Plus[27] addressed this issue by utilizing patch features before the projection layer. InstantID[26] introduced the use of a pre-trained Arcface[2] encoder, significantly enhancing identity consistency and influencing subsequent works like Photomaker [9] and PULID [4]. A natural progression is to attack both the CLIP image encoder and Arcface encoder simultaneously, targeting all current Encoder-based ID preservation technologies." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.4, + 0.483, + 0.656 + ], + "angle": 0, + "content": "There are four major challenges in ID protection: 1. Universality. Given the numerous existing encoder-based ID-preserving methods, we believe that universal protection is necessary, as attackers can easily switch to other methods to bypass protection. 2. Efficiency. Previous work typically relies on the Projected Gradient Descent (PGD) [12] to generate adversarial noise. However, PGD requires per-image optimization, taking several minutes per image, which significantly limits its practicality for protecting large volumes of portrait photos on social media in real world. 3. Robustness. The adversarial noise must be robust enough to withstand common image transformations, such as cropping, resizing, and JPEG compression, while also countering the inherent pre-processing operations of identity-preserving methods, such as face alignment. 4. Imperceptibility. The adversarial noise added for ID protection should be imperceptible to human eyes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In response to these challenges, this paper presents ID-Protector, the first feed-forward method to protect portrait photos from state-of-the-art encoder-based identity preservation techniques. Our approach incorporates several key innovations. For efficiency, we propose a ViT-based encoder that directly predicts adversarial noise for a given target portrait. To ensure universality, we carefully examine the attack surfaces and define a composite loss function tailored to four mainstream encoder-based generation methods, including InstantID [26], IP-Adapter [27], IP-Adapter Plus [27], and PhotoMaker [9]. Additionally, we employ an L1-based image regularization to minimize the visual impact of adversarial noise on image quality. For robustness against common transformations, a typical approach involves a combination of post-hoc differentiable augmentations like differentiable JPEG and cropping during train" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.242 + ], + "angle": 0, + "content": "ing. However, we find that this method introduces a substantial learning burden. To address this, we propose a surrogate method that leverages the face preprocessing pipeline of existing identity-preserving methods. Specifically, we introduce small random noise into the affine transformation matrix during face alignment stage to achieve data augmentation. This not only enhances the robustness of adversarial noise against common image transformations but also enables it to withstand image preprocessing operations during identity-preserving generation." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.243, + 0.905, + 0.378 + ], + "angle": 0, + "content": "We train and extensively evaluate IDProtector on the CelebA dataset [11]. We further test it on unseen data VGG Face [15] and proprietary models like Midjourney [13] and Jing Gou, demonstrating its strong generalization to unseen data and closed-source models. By advancing adversarial attack techniques from theoretical constructs to practically deployable solutions, IDProtector sets a new benchmark in facial privacy protection. We summarize our main contributions as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.38, + 0.904, + 0.424 + ], + "angle": 0, + "content": "- We introduce IDProtector, the first feed-forward method to protect portrait photos from SOTA encoder-based identity preservation techniques." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.425, + 0.904, + 0.516 + ], + "angle": 0, + "content": "- We present several innovative techniques including a ViT-based noise encoder, a joint learning framework with carefully designed objectives, and affine-based data augmentation to simultaneously achieve the EURI standard (Efficiency, Universality, Robustness, and Imperceptibility)." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.38, + 0.904, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.532, + 0.663, + 0.547 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.557, + 0.842, + 0.573 + ], + "angle": 0, + "content": "2.1. Tuning-based Customized Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.58, + 0.905, + 0.746 + ], + "angle": 0, + "content": "With the rapid advancement of diffusion models, customized generation methods have begun to emerge. Techniques like DreamBooth [17], Custom Diffusion [8], Low-Rank Adaptation (LoRA) [7] and Textual Inversion [3] facilitate the adaptation of text-to-image diffusion models through fine-tuning strategies that range from adjusting all parameters to optimizing specific word vectors. Despite their impressive effects, these methods require extensive computational resources and time due to their need for multi-image fine-tuning. This can make them impractical for real-time or large-scale applications." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.758, + 0.868, + 0.773 + ], + "angle": 0, + "content": "2.2. Encoder-based ID-Preserving Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Encoder-based Customized generation methods[23, 27, 29, 30] utilize pre-trained encoders to extract appearance feature, facilitating zero-shot generation in real-time. These methods evolved from domain-agnostic encoders to domain-specific encoders. Initially, CLIP[16] image encoder was used to extract facial features, such as with Face0[23] and IP-Adapter[27]; however, their maintenance of facial detail fidelity was not satisfactory. IP-Adapter" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3020" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.273 + ], + "angle": 0, + "content": "Plus addressed this issue by utilizing patch features before the projection layer. InstantID[26] was the first to propose using a pre-trained Arcface[2] Encoder as a condition, significantly enhancing identity consistency and influencing subsequent works like Photomaker[9] and PULID[4]. StoryMaker[32] employs both the Arcface image encoder and the CLIP image encoder to encode facial features and broader regions, respectively, preserving not only facial identities but also clothing, hairstyles, and body features. Additionally, encoder-based ID preservation technology is widely applied in multi-human image personalization[5, 14] and video tasks[6]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.282, + 0.479, + 0.297 + ], + "angle": 0, + "content": "2.3. Defensive Perturbations for Diffusion Models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.304, + 0.485, + 0.651 + ], + "angle": 0, + "content": "The focus on safeguarding personal artworks and facial images from unauthorized usage in the fine-tuning of diffusion models has led to the development of subtle yet effective perturbations. As summarized in Tab. 1, distinct approaches have been tailored to address this issue: Glaze [20] strategically adjusts the feature-space distances to misdirect the style learning process in Stable Diffusion, effectively preventing style appropriation. Conversely, AdvDM [10] implements adversarial disruptions by maximizing the Mean Squared Error during the optimization phase to shield personal visuals. Meanwhile, Anti-DreamBooth [24] integrates adversarial interference directly within the DreamBooth fine-tuning cycle, applying a complex optimization strategy to formulate defensive perturbations. Further investigations [21, 25, 28, 31] into adversarial protection have demonstrated their effectiveness in obstructing finetuning-based customization methods such as DreamBooth [17], LoRA [7], and textual inversion [3]. However, how to protect personal identity from the recently popular encoder-based methods remains unexplored. This paper endeavors to bridge the current gaps by proposing advanced methodologies to shield images from encoder-based exploitations, enhancing the security measures for personal digital assets." + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.664, + 0.483, + 0.69 + ], + "angle": 0, + "content": "Table 1. Overview of encoder-based methods for ID-preserving generation." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.692, + 0.481, + 0.741 + ], + "angle": 0, + "content": "
EncoderID-Preserving Methods
CLIP [16]IP-Adapter [27], IP-Adapter Plus [27], PhotoMaker [9], Face0 [23]
ArcFace [2]InstantID [26], PhotoMaker [9], PuLID [4], IP-Adapter-FaceID [27]
CLIP & ArcFaceIDAdapter [1], Storymaker [32]
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.77, + 0.275, + 0.785 + ], + "angle": 0, + "content": "3. Problem Definition" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.482, + 0.901 + ], + "angle": 0, + "content": "We consider two participants in a portrait identity protection scenario.: \"image protector\" Alice and \"image exploiter\" Bob. Bob seeks to use an encoder-based, identity-preserving approach to alter the portrait photos Alice shares on social media for malicious purposes. While Alice aims to protect her portrait by adding adversarial noise to thwart Bob's methods, inducing distortions or significant decreases" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "in ID similarity in the generated content. Specifically, we explain the workflow of the two parties as follows:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.122, + 0.905, + 0.302 + ], + "angle": 0, + "content": "1. Image Protector Alice: The image protector aims to provide protection for images against exploitation by encoder-based ID-preserving methods. In this context, the chosen protection method involves adding imperceptible protective perturbations to the images, with the goal of providing protection while minimizing modifications to the original image. In real-world scenarios, image protectors encounter several challenges. For instance, implementing universal protection to guard against various encoder-based, identity-preserving methods is difficult. Additionally, the protection should be robust to natural distortions such as resizing, JPEG compression, and blurring." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.303, + 0.905, + 0.377 + ], + "angle": 0, + "content": "2. Image Exploiter Bob: Bob aims to use portrait photos for ID-preserving customization. Bob can choose any open-source or proprietary encoder-based models, which poses significant challenges on the universality of the protection method." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.122, + 0.905, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.379, + 0.907, + 0.53 + ], + "angle": 0, + "content": "Objective of the Study: Given an image \\( x \\) containing a face portrait, our objective is to generate adversarial perturbations \\( \\delta \\) that protect against ID-preserving generation. We aim to maximize the difference in identity consistency between the unperturbed generated image \\( g(x) \\) and the perturbed generated image \\( g(x + \\delta) \\), where \\( g \\) is an ID-preserving generator. To unify the protection against different models, we can find a \\( \\delta \\) that simultaneously disrupts several feature spaces, lowering the sum of the similarity values. Thus, we train a noise encoder network \\( E_{\\theta} \\), solving" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.541, + 0.905, + 0.563 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} S \\left\\{g _ {i} (x), g _ {i} (x + \\delta) \\right\\}, \\quad \\delta = E _ {\\theta} (x) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.574, + 0.905, + 0.62 + ], + "angle": 0, + "content": "subject to the constraint that \\(\\| \\delta \\|_{\\infty} < \\epsilon\\). Here, \\(S\\) denotes an identity similarity metric, which we evaluate using cosine similarity in feature spaces of ArcFace [2] and CLIP [16]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.633, + 0.605, + 0.648 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.658, + 0.905, + 0.765 + ], + "angle": 0, + "content": "The core of our method is an adversarial noise encoder model that can add adversarial noise to facial images for protection. We will first introduce the complete workflow in Sec. 4.1, followed by the design of the adversarial noise encoder in Sec. 4.2. Sec. 4.3 and 4.4 introduce ID protection objectives and loss functions. Sec. 4.5 introduces the data augmentation approach to enhance the robustness." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.773, + 0.673, + 0.788 + ], + "angle": 0, + "content": "4.1. Overall Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Fig. 2 details the pipeline of our method. For an arbitrary-sized input image \\( x \\) containing a face, it is resized to \\( 224 \\times 224 \\) and sent into the adversarial noise encoder model, whose output is an additive adversarial noise that can be resized back and perturbed onto the original image for protection. During adversarial noise encoder training, this protected image is further passed through face feature extrac" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "3021" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.093, + 0.898, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.289, + 0.908, + 0.361 + ], + "angle": 0, + "content": "Figure 2. Overall schematics of our method. Our method's key design includes the noise encoder, loss functions, and the gradient optimization path that allows backpropagation. During the protection phase, an image is first resized to \\(224 \\times 224\\) and sent to an adversarial noise encoder model. The model outputs a perturbation that can be resized back and added to the image, achieving an adversarial attack effect – the protected image prevents subsequent pipelines from correctly extracting facial feature information, thereby hindering face similarity in face generation tasks. This achieves the goal of ID protection." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.386, + 0.483, + 0.418 + ], + "angle": 0, + "content": "tion pipelines to obtain the loss functions that aim at creating maximum differences in the feature space." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.425, + 0.24, + 0.44 + ], + "angle": 0, + "content": "4.2. Noise Encoder" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.447, + 0.484, + 0.675 + ], + "angle": 0, + "content": "The network takes a \\(224 \\times 224\\) RGB image as input and outputs three-channel adversarial perturbations. The perturbations are projected to a range of \\([-1, 1]\\), then denormalized \\([- \\epsilon, \\epsilon]\\), resized to the original image dimensions, and added to the image. We utilize a Vision Transformer (ViT) to generate the adversarial noise, with both input and output dimensions set to \\(224 \\times 224\\). Empirically, we find that adding an additional prior mask channel, indicating the location of the face, aids in training. This face localization mask is generated by the InsightFace pipeline based on facial landmark points, which specify the region of the face within the image. The mask is concatenated to the image as the fourth input channel, eliminating the need for the network to learn face localization ability from scratch, thus reducing training difficulty." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.682, + 0.327, + 0.698 + ], + "angle": 0, + "content": "4.3. Adversarial ID Protection" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.483, + 0.825 + ], + "angle": 0, + "content": "Since generative networks' only knowledge of faces comes from feature embeddings, to achieve ID protection, we can make the extracted features from protected images significantly different from the originals, thereby preventing generation models from accessing correct facial features. Based on this principle, we analyze the specific pipelines of InstantID, IP-Adapter, IP-Adapter-Plus, and PhotoMaker to design targeted attack objectives and loss functions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "InstantID obtains facial features in two steps, as shown in Fig. 2. It first aligns the face to a pre-defined location, before feeding the aligned face into an ArcFace feature extractor to obtain facial information. We choose to fool the ArcFace model, minimizing the cosine similarity between" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.386, + 0.895, + 0.401 + ], + "angle": 0, + "content": "the ArcFace features of the original v.s. protected images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.402, + 0.905, + 0.644 + ], + "angle": 0, + "content": "IP-Adapter, IP-Adapter-Plus, and PhotoMaker all rely on facial features extracted by the CLIP vision encoder, albeit using different versions of CLIP vision. IP-Adapter and PhotoMaker utilize CLIP Vision's output, while IP-Adapter-Plus takes features before CLIP Vision's second-to-last layer. As illustrated in Fig. 2, embeddings from different layers offer various attack surfaces. We select victim embeddings for attack according to three principles: (1) blocking all potential pathways, ensuring that any information flow from left to right passes through at least one victim embedding, ensuring all information flow would be disrupted; (2) choosing features as early as possible in the network (towards the left in Fig. 2), to reduce the backpropagation path length and thus simplify optimization; and (3) targeting embeddings with dense semantic information for more effective manipulation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.646, + 0.907, + 0.812 + ], + "angle": 0, + "content": "An example of an unsuitable target is the last hidden layer feature of CLIP Vision. While modifying this feature could influence all three models, its huge dimension \\((257\\times 1280)\\) and lack of semantic density make it highly challenging for optimization to converge. Ultimately, we selected the embeddings marked red in Fig. 2 as the primary attack targets. The objective, similar to that of InstantID, is to maximize the cosine similarity between the perturbed embedding and the original embedding, effectively aligning them post-attack. Therefore, the final adversarial loss is a weighted average of all losses, given by" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.825, + 0.905, + 0.859 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {a d v}} = \\sum_ {\\text {C u t s o m M o d e l} i} \\alpha_ {i} \\cdot \\operatorname {c o s s i m} \\left(e _ {i}, e _ {i} ^ {\\prime}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.908, + 0.903 + ], + "angle": 0, + "content": "where \\( e_i \\) and \\( e_i' \\) denote the face embeddings of the clean and protected images, respectively." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3022" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "4.4. Imperceptibility" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.484, + 0.19 + ], + "angle": 0, + "content": "To minimize the visual impact on the image quality, we enforce an \\(\\ell_1\\) regularization on the predicted adversarial noise \\(\\delta\\). We further introduce an auxiliary penalty on any noise values exceeding the \\(\\epsilon\\)-ball boundary. Together, these terms constitute our regularization loss, given by:" + }, + { + "type": "equation", + "bbox": [ + 0.147, + 0.202, + 0.483, + 0.22 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {r e g}} = \\beta_ {1} \\cdot \\| \\delta \\| _ {1} + \\beta_ {2} \\cdot \\| \\delta - \\operatorname {c l i p} _ {\\pm \\epsilon} (\\delta) \\| _ {1}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.231, + 0.483, + 0.262 + ], + "angle": 0, + "content": "The final loss is thus the sum of adversarial loss and regularization, given by:" + }, + { + "type": "equation", + "bbox": [ + 0.228, + 0.275, + 0.483, + 0.293 + ], + "angle": 0, + "content": "\\[\nL = L _ {\\mathrm {a d v}} + L _ {\\mathrm {r e g}}. \\tag {4}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.304, + 0.216, + 0.319 + ], + "angle": 0, + "content": "4.5. Robustness" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.326, + 0.484, + 0.508 + ], + "angle": 0, + "content": "As shown in Fig. 2, the InstantID pipeline involves a step of facial alignment aiming to align any faces to a predefined standard face for better feature extraction. The alignment process includes detecting the face and key points \\(\\tilde{p}\\), calculating the affine matrix \\(A\\) with standard face key points \\(\\tilde{p}'\\), and aligning them. This process is computationally expensive. To reduce computational cost and simplify training, we pre-compute \\(A\\) for each image and keep it fixed throughout training. However, during inference, perturbations carried by protected images may slightly alter the precomputed affine matrix \\(A\\), leading to minor misalignment that weakens the effectiveness the adversarial perturbations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.509, + 0.484, + 0.6 + ], + "angle": 0, + "content": "To adapt to this misalignment as well as being computationally efficient, we leverage the existing affine transformation operation and introduce small gaussian noise to \\( A \\) during training to simulate face transformations. The noise level is empirically set to \\( \\sigma = 0.003 \\). The face alignment transformation is therefore given by" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.612, + 0.483, + 0.629 + ], + "angle": 0, + "content": "\\[\n\\tilde {p} ^ {\\prime} = (A + N (0, \\sigma I)) \\tilde {p}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.642, + 0.484, + 0.703 + ], + "angle": 0, + "content": "We incorporate this augmentation only in the InstantID branch. Our experiments reveal that it also enhances the robustness of other branches, being resilient to various common image transformations." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.717, + 0.224, + 0.734 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.742, + 0.282, + 0.759 + ], + "angle": 0, + "content": "5.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.87 + ], + "angle": 0, + "content": "Training Data. For practicality, we selected 190,000 uncropped images from the CelebA dataset [11] as the training set. These images have arbitrary sizes and aspect ratios. The test set includes two subsets: 50 unseen images from CelebA and 50 images from VGG Face [15]. To ensure compatibility with baseline methods, these images were resized to \\(512 \\times 512\\) and \\(224 \\times 224\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Training Details. We trained a ViT-S/8 adversarial noise encoder from scratch using Eq. 4 as the loss function. We" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.907, + 0.131 + ], + "angle": 0, + "content": "Table 2. Curriculum learning stages with corresponding hyperparameters. \\( \\eta \\) represents learning rate. \\( {\\alpha }_{1} \\) to \\( {\\alpha }_{4} \\) are weights for IP-Adapter,IP-Adapter-Plus,PhotoMaker,InstantID,respectively." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.133, + 0.905, + 0.19 + ], + "angle": 0, + "content": "
StageEpochsα1α2α3α4β1β2εη
11202/3714/3720/371/379e-31e-30.051e-2
2204/162/169/162/161.8e-32e-40.042e-5
3202/142/149/141/144.5e-45e-50.0352e-5
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.217, + 0.907, + 0.352 + ], + "angle": 0, + "content": "employed a three-stage curriculum learning approach, progressively shrinking the \\(\\epsilon\\)-ball and balancing the different losses by adjusting the hyperparameters (presented in Table 2). Training began with a 2,500-step linear warm-up, gradually increasing the learning rate from zero. Gradient clipping was applied during training to cap the \\(\\ell_2\\) norm of gradients at a maximum of 10. The model was trained with a batch size of 112 on four 80GB NVIDIA H100 GPUs for 10 days." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.353, + 0.909, + 0.519 + ], + "angle": 0, + "content": "Evaluation process and metrics. To evaluate the protection performance of different methods, for each face-containing image in the dataset (denoted as NP ref image), we apply different protection methods (including our ID-Protector) to generate its protected counterpart (denoted as P ref image). Throughout all experiments, we constrain the maximum pixel perturbation to 9/255 (an \\(\\epsilon\\)-ball of 0.035), following common practice in adversarial machine learning [18]. Subsequently, both P and NP ref images are fed into customization generation models (such as InstantID) to obtain generated images (denoted as P/NP gen images)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.52, + 0.909, + 0.626 + ], + "angle": 0, + "content": "We evaluate the identify shift caused by different protection methods through Identity Score Matching (ISM), which is calculated as the cosine similarity between facial features of NP images and generated images extracted by ArcFace [2]. Lower ISM values indicate better protection as they suggest a greater identity shift between original (NP ref) and generated (P gen) faces." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.626, + 0.909, + 0.718 + ], + "angle": 0, + "content": "Since protected images may result in no detectable faces being generated, we also report the Face Detection Rate (FDR) using InsightFace detector, as ISM can only be computed when faces are detected in both images. Additionally, to ensure comprehensive evaluation, we measure the perceptual quality of each generated face using SER-FIQ [22]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.728, + 0.76, + 0.744 + ], + "angle": 0, + "content": "5.2. Comparison with Baselines" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Protection Performance. We compare our method with the following baseline methods: ACE [31], AdvDM [10], Anti-DreamBooth [24], PhotoGuard [19], and SimAC [25]. Most baselines aim to reduce the quality of generated images or hinder the diffusion denoising process, but we aim to reduce the similarity of generated identities so as to directly disrupt identity-preserving generation. To compare our method with baselines, we first applied each protection method to generate protected versions of test images. We then used both the original images (NP) and the protected" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3023" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.09, + 0.9, + 0.598 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.601, + 0.907, + 0.631 + ], + "angle": 0, + "content": "Figure 3. Qualitative comparison with baseline. Images protected by our method cannot be used to generate similar faces. Compared to baseline protection methods, under the same perturbation bound of \\(9/255\\) (\\(\\epsilon = 3.5\\%\\)), ours causes more significant facial differences." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.656, + 0.487, + 0.899 + ], + "angle": 0, + "content": "versions as references in customized generation tasks with four models: IP-Adapter, IP-Adapter-Plus, PhotoMaker, and InstantID. Tab. 3 shows that, for the first three methods, while baseline methods reduced ISM by an average of 0.02, with a maximum reduction of 0.06, our method achieved an ISM reduction between 0.1 and 0.14, demonstrating a larger identity shift. On InstantID, our IDProtector demonstrated even more pronounced protection: when baseline methods reduced facial similarity by less than 0.1, our method reduced similarity by over 0.4 across both datasets. Additionally, FDR results validate the reliability of ISM as FDR reflects the percentage of images with a face detected and contributes to ISM calculations. Overall, these results underscore the superior identity protection performance of our method compared to existing approaches. Together, these highlight the superior identity protection performance of" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.657, + 0.608, + 0.671 + ], + "angle": 0, + "content": "our approach." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.673, + 0.907, + 0.794 + ], + "angle": 0, + "content": "Following previous study [24], we evaluate SER-FIQ as a quality metric of generated (P/NP gen) images. Although our method has a noticeable impact on image generation quality (as shown in Fig. 3), we observed inconsistencies between the SER-FIQ scores and qualitative assessments. This discrepancy suggests that a more robust and reliable quality metric may be required to accurately measure the quality of generated images in future research." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We also adapted our method to a per-image tuning variant, employing PGD [12] as the optimization technique. We observed that PGD achieves a more substantial ISM reduction across most methods, which aligns with our expectations since per-image tuning can better leverage the specific characteristics of each individual image. However, a notable drawback of PGD is the prolonged optimization time," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3024" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.131 + ], + "angle": 0, + "content": "Table 3. Comparison of identity protection performance against baseline methods. Protection performance is evaluated against IP-Adapter, IP-Adapter-Plus, PhotoMaker, and InstantID. \"No Protect\" denotes direct use of the original (NP ref) images for identity-preserving generation. Experiments were conducted on five prompts per dataset, and average results are reported." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.132, + 0.907, + 0.361 + ], + "angle": 0, + "content": "
DatasetProtection MethodIP-AdapterIP-Adapter-PlusPhotoMakerInstantID
FDR ↓ISM ↓SER-FIQ ↓FDR ↓ISM ↓SER-FIQ ↓FDR ↓ISM ↓SER-FIQ ↓FDR ↓ISM ↓SER-FIQ ↓
CelebANo Protect1.0000.1760.6411.0000.2400.6630.6640.2650.6050.9960.7750.774
ACE1.0000.1590.6401.0000.2310.6810.6360.2680.5941.0000.7450.758
AdvDM0.9880.1570.6301.0000.2260.6660.6400.2550.5901.0000.7570.757
Anti-DB1.0000.1550.6151.0000.2220.6760.6400.2550.5991.0000.7530.752
PhotoGuard1.0000.1520.6211.0000.2230.6700.6520.2590.5921.0000.7580.767
SimAC1.0000.1620.6271.0000.2170.6830.6520.2530.5991.0000.7520.755
Ours (PGD)0.8440.0240.5440.9120.0720.7180.1120.0940.6390.996-0.1660.721
Ours0.9520.0600.6880.9960.1140.6440.6720.1230.6221.0000.2310.613
VGG FaceNo Protect1.0000.2010.6721.0000.2360.6980.6320.2540.6000.9960.7640.828
ACE0.9880.1540.6171.0000.2090.6831.0000.2090.6831.0000.7030.829
AdvDM1.0000.1470.5971.0000.2150.6931.0000.2150.6931.0000.7060.820
Anti-DB1.0000.1380.5971.0000.2100.6780.6240.2380.5991.0000.7070.811
PhotoGuard1.0000.1590.6431.0000.2170.6860.6240.2290.6120.9960.6910.799
SimAC1.0000.1500.6051.0000.2020.6770.6240.2280.6441.0000.7070.826
Ours (PGD)0.9680.0430.6091.0000.2020.7400.6240.2280.6301.0000.7070.769
Ours1.0000.1050.7371.0000.1380.6870.6320.1260.6141.0000.2320.713
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.382, + 0.483, + 0.424 + ], + "angle": 0, + "content": "Table 4. Comparison of time cost and quality impact with an epsilon-budget of 9/255 (±3.5%). PSNR and SSIM are calculated between clean (NP ref) and perturbed (P ref) image pairs." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.424, + 0.485, + 0.496 + ], + "angle": 0, + "content": "
ACEAdvDMAnti-DBPhoto-GuardSimACOurs PGDOurs
Time (s)87.051.926123.332340890.173
PSNR (dB)31.0631.7031.7031.1032.3732.9432.15
SSIM (dB)0.7410.7660.7650.7570.7910.8450.809
" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.522, + 0.405, + 0.537 + ], + "angle": 0, + "content": "which we will address in the following section." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.538, + 0.483, + 0.719 + ], + "angle": 0, + "content": "The protection effects are visualized in Fig. 3. The baseline methods can protect against some of the generation methods. For example, AdvDM and Anti-DreamBooth cause IP-Adapter to produce faces with color block artifacts, while several outputs from IP-Adapter-Plus exhibit unnatural patterns and shadows. However, the baselines are ineffective in protecting against PhotoMaker and InstantID. With IDProtector's protection, faces generated by all four methods appear significantly altered. In some cases, even the gender of the person is unidentifiable. This fulfills our protection goal: protected images could not be used for identity-preserving generation." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.72, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Time and Image Quality. Our IDProtector is fast to protect an image with less degradation to image quality. As compared in Table 4, the PSNR and SSIM of our method are higher than the baselines, indicating less visual impact. The baseline methods require an average of 23.3 to 323 seconds per image for protection. Using PGD with our optimization objective would take over an hour on average. In contrast, our ViT-based IDProtector achieves an average protection time of only 0.173 seconds per image – less than \\(1\\%\\) of the fastest baseline. Note that the forward-feeding ViT supports better parallelization compared with PGD, the time cost could further decrease with mini-batch inference." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.382, + 0.907, + 0.437 + ], + "angle": 0, + "content": "Table 5. Generalization of ID protection by our IDProtector on unseen generators. IPA-FID+/P/XL denotes IP-Adapter-FaceID/Plus/Portrait/SDXL, with NP and P indicating no and with protection, respectively." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.438, + 0.906, + 0.519 + ], + "angle": 0, + "content": "
IPA-FIDIPA-FID+IPA-FIDPIPA-FIDXLFlux-IPAStory-MakerMid-journeyJingGou
ISM ↓NP0.5650.6440.1290.5230.0540.5720.1990.672
P0.1530.2170.0520.1320.0290.1930.0980.131
SER-NP0.7060.7510.7200.4020.7280.6990.5680.791
FIQ ↓P0.8030.7300.6580.6670.6730.7340.6270.805
" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.538, + 0.902, + 0.742 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.534, + 0.743, + 0.885, + 0.757 + ], + "angle": 0, + "content": "Figure 4. ID protection performance on unseen generators." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.787, + 0.88, + 0.802 + ], + "angle": 0, + "content": "5.3. Generalization to Unseen Data and Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Generalization to unseen datasets. We evaluated our method on the VGG Face dataset that was unseen during training. All following experiments are conducted on the CelebA test set unless otherwise specified. As shown in Tab. 3, the ISM reduction achieved by our method shows minimal variation across datasets. For example, our IDPro" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3025" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.118 + ], + "angle": 0, + "content": "Table 6. Robustness evaluation of IDProtector under image distortions. NP denotes non-protected images for identity-preserving generation, while P represents protected but non-distorted images, providing an ablation. Distortion averages exclude NP and P." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.119, + 0.907, + 0.248 + ], + "angle": 0, + "content": "
IDProtector (ours, ViT)Adaptive PGD
IP-AdapterIP-Adapter+PhotoMakerInstantIDIP-AdapterIP-Adapter+PhotoMakerInstantID
FDR ↓ISM ↓FDR ↓ISM ↓FDR ↓ISM ↓FDR ↓ISM ↓FDR ↓ISM ↓FDR ↓ISM ↓FDR ↓ISM ↓FDR ↓ISM ↓
NP1.0000.1761.0000.2400.6640.2650.9960.7750.8440.0240.9120.0720.1120.0940.996-0.166
P0.9520.0600.9960.1140.6720.1231.0000.2310.8440.0240.9120.0720.1120.0940.996-0.166
P+Affine1.0000.0681.0000.1060.6560.1351.0000.2160.8440.0640.9760.0810.112-0.0091.000-0.147
P+JPEG0.9640.0691.0000.0970.6600.1381.0000.2410.7840.0150.9480.0620.108-0.0121.000-0.161
P+Crop1.0000.0721.0000.1080.6520.1321.0000.2160.7520.0580.9400.0700.1080.0201.000-0.167
P+Noisy0.9520.0690.9840.1140.6560.1321.0000.2520.8320.0230.9320.0720.112-0.0101.000-0.130
Avg (P+)0.9790.0700.9960.1060.6560.1341.0000.2310.8030.0400.9490.0710.110-0.0031.000-0.151
" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.274, + 0.48, + 0.518 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.52, + 0.483, + 0.575 + ], + "angle": 0, + "content": "Figure 5. Robustness evaluation of IDProtector under various distortions. Despite geometric distortions like affine transformations, IDProtector effectively misleads all four models into generating dissimilar faces or even artifacts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.608, + 0.484, + 0.682 + ], + "angle": 0, + "content": "tector reduces the ISM of IP-Adapter by approximately 0.1 and InstantID by around 0.55 on both datasets. This result suggests that our method generalises effectively beyond the CelebA dataset, demonstrating protection capabilities that are not restricted to in-domain faces." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.687, + 0.483, + 0.836 + ], + "angle": 0, + "content": "Generalization to unseen customization models. Beyond the four customization models seen during training, we tested protected images on a wider range of customization models. Across these models, our protection approach achieves an average ISM reduction of 0.31, with a maximum reduction of 0.54. Notably, our method remains effective against proprietary commercial models such as Midjourney (0.10 reduction) and Jing Gou (0.54 reduction), underscoring its ability to generalise to unseen customization models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.903 + ], + "angle": 0, + "content": "These protection effects are visualised in Fig. 4. As shown, protected images used in these generators cause noticeable quality degradation, including altered facial shapes (e.g., IP-Adapter-FaceID-XL), reduced prompt-following" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.273, + 0.907, + 0.319 + ], + "angle": 0, + "content": "ability, and image contamination (e.g., Midjourney). The common result is a decreased resemblance between the generated face and the reference identity." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.327, + 0.725, + 0.342 + ], + "angle": 0, + "content": "5.4. Robustness Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.349, + 0.907, + 0.576 + ], + "angle": 0, + "content": "Our method demonstrates robustness against various distortions. For protected (P) images, we applied several distortions, including JPEG compression (quality=85), random cropping that removes up to \\(20\\%\\) of the region, noise perturbation with a variance of 100, and an affine transformation. The affine transformation was implemented as \\(\\tilde{q} = (A + \\mathcal{N}(0,0.05^2))\\tilde{p}\\), where the matrix \\(A\\) is an identity matrix. As shown in Tab. 6, the ISM values under these distortions remain comparable to those obtained without distortion, indicating that our protection remains effective under distortion. Fig. 5 provides a visual comparison, where even under various distortions, InstantID fails to reconstruct the correct female face, generating a male portrait instead. This demonstrates that the common distortions do not compromise the efficacy of our protection method." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.589, + 0.765, + 0.605 + ], + "angle": 0, + "content": "6. Conclusion and Limitation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.907, + 0.825 + ], + "angle": 0, + "content": "In this paper, we introduce IDProtector, a method designed to protect facial identities from encoder-based identity-preserving generation. By training a noise encoder to add adversarial perturbations onto images, IDProtector can protect an image within 0.2s. Our method achieves significantly lower face similarity in generated images than existing protection methods. IDProtector also demonstrates strong practical applicability. Its protection remains effective under distortions such as affine transformations and JPEG compression. Furthermore, IDProtector not only shows effectiveness against closed-source models, such as Midjourney, but also unifies protection against multiple ID-preserving models, setting a new standard for privacy protection against generative AI." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.907, + 0.901 + ], + "angle": 0, + "content": "A limitation of our method is that the adversarial noise is not fully invisible, necessitating a trade-off between attack effectiveness and Imperceptibility. Future efforts will focus on enhancing the Imperceptibility of the noise while maintaining its robustness." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3026" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.279, + 0.108 + ], + "angle": 0, + "content": "7. Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.484, + 0.161 + ], + "angle": 0, + "content": "This research is supported by the Ministry of Education, Singapore, under its Academic Research Fund Tier 2 (Award No: MOE-T2EP20124-0012)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.174, + 0.188, + 0.189 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.2, + 0.484, + 0.282 + ], + "angle": 0, + "content": "[1] Siying Cui, Jia Guo, Xiang An, Jiankang Deng, Yongle Zhao, Xinyu Wei, and Ziyong Feng. Idadapter: Learning mixed features for tuning-free personalization of text-to-image models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 950-959, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.284, + 0.484, + 0.351 + ], + "angle": 0, + "content": "[2] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.354, + 0.484, + 0.422 + ], + "angle": 0, + "content": "[3] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.424, + 0.483, + 0.465 + ], + "angle": 0, + "content": "[4] Zinan Guo, Yanze Wu, Zhuowei Chen, Lang Chen, and Qian He. Pulid: Pure and lightning id customization via contrastive alignment. arXiv preprint arXiv:2404.16022, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.467, + 0.484, + 0.521 + ], + "angle": 0, + "content": "[5] Junjie He, Yifeng Geng, and Liefeng Bo. Uniportrait: A unified framework for identity-preserving single- and multi-human image personalization. arXiv preprint arXiv:2408.05939, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.523, + 0.484, + 0.578 + ], + "angle": 0, + "content": "[6] Xuanhua He, Quande Liu, Shengju Qian, Xin Wang, Tao Hu, Ke Cao, Keyu Yan, Man Zhou, and Jie Zhang. Id-animator: Zero-shot identity-preserving human video generation. arXiv preprint arXiv:2404.15275, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.58, + 0.484, + 0.647 + ], + "angle": 0, + "content": "[7] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.649, + 0.484, + 0.718 + ], + "angle": 0, + "content": "[8] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.72, + 0.484, + 0.789 + ], + "angle": 0, + "content": "[9] Zhen Li, Mingdeng Cao, Xintao Wang, Zhongang Qi, Ming-Ming Cheng, and Ying Shan. Photomaker: Customizing realistic human photos via stacked id embedding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8640-8650, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.79, + 0.483, + 0.858 + ], + "angle": 0, + "content": "[10] Chumeng Liang, Xiaoyu Wu, Yang Hua, Jiaru Zhang, Yiming Xue, Tao Song, Zhengui Xue, Ruhui Ma, and Haibing Guan. Adversarial example does good: Preventing painting imitation from diffusion models via adversarial examples. arXiv preprint arXiv:2302.04578, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[11] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Large-scale celebfaces attributes (celeba) dataset. Retrieved August, 15(2018):11, 2018." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.2, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "[12] Aleksander Madry. Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.904, + 0.164 + ], + "angle": 0, + "content": "[13] Midjourney.com. Midjourney. https://www.midjourney.com/home." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.166, + 0.905, + 0.221 + ], + "angle": 0, + "content": "[14] Daniil Ostashev, Yuwei Fang, Sergey Tulyakov, Kfir Aberman, et al. Moa: Mixture-of-attention for subject-context disentanglement in personalized image generation. arXiv preprint arXiv:2404.11565, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.223, + 0.905, + 0.277 + ], + "angle": 0, + "content": "[15] Omkar Parkhi, Andrea Vedaldi, and Andrew Zisserman. Deep face recognition. In BMVC 2015-Proceedings of the British Machine Vision Conference 2015. British Machine Vision Association, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.28, + 0.905, + 0.348 + ], + "angle": 0, + "content": "[16] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.35, + 0.905, + 0.432 + ], + "angle": 0, + "content": "[17] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.435, + 0.905, + 0.503 + ], + "angle": 0, + "content": "[18] Mehrdad Saberi, Vinu Sankar Sadasivan, Keivan Rezaei, Aounon Kumar, Atoosa Chegini, Wenxiao Wang, and Soheil Feizi. Robustness of ai-image detectors: Fundamental limits and practical attacks. arXiv preprint arXiv:2310.00076, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.505, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[19] Hadi Salman, Alaa Khaddaj, Guillaume Leclerc, Andrew Ilyas, and Aleksander Madry. Raising the cost of malicious ai-powered image editing. arXiv preprint arXiv:2302.06588, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.563, + 0.905, + 0.631 + ], + "angle": 0, + "content": "[20] Shawn Shan, Jenna Cryan, Emily Wenger, Haitao Zheng, Rana Hanocka, and Ben Y Zhao. Glaze: Protecting artists from style mimicry by {Text-to-Image} models. In 32nd USENIX Security Symposium (USENIX Security 23), pages 2187-2204, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.633, + 0.905, + 0.689 + ], + "angle": 0, + "content": "[21] Yiren Song, Shengtao Lou, Xiaokang Liu, Hai Ci, Pei Yang, Jiaming Liu, and Mike Zheng Shou. Anti-reference: Universal and immediate defense against reference-based generation. arXiv preprint arXiv:2412.05980, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.691, + 0.905, + 0.772 + ], + "angle": 0, + "content": "[22] Philipp Terhorst, Jan Niklas Kolf, Naser Damer, Florian Kirchbuchner, and Arjan Kuijper. Ser-fiq: Unsupervised estimation of face image quality based on stochastic embedding robustness. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5651-5660, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.775, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[23] Dani Valevski, Danny Wasserman, Yossi Matias, and Yaniv Leviathan. Face0: Instantaneously conditioning a text-to-image model on a face. arXiv preprint arXiv:2306.06638, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[24] Thanh Van Le, Hao Phung, Thuan Hoang Nguyen, Quan Dao, Ngoc N Tran, and Anh Tran. Anti-dreambooth: Protecting users from personalized text-to-image synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2116-2127, 2023." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3027" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.174 + ], + "angle": 0, + "content": "[25] Feifei Wang, Zhentao Tan, Tianyi Wei, Yue Wu, and Qidong Huang. Simac: A simple anti-customization method for protecting face privacy against text-to-image synthesis of diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12047-12056, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.178, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[26] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, and Anthony Chen. Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.274 + ], + "angle": 0, + "content": "[27] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.482, + 0.33 + ], + "angle": 0, + "content": "[28] Xiaoyu Ye, Hao Huang, Jiaqi An, and Yongtao Wang. Duaw: Data-free universal adversarial watermark against stable diffusion customization. arXiv preprint arXiv:2308.09889, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.333, + 0.482, + 0.415 + ], + "angle": 0, + "content": "[29] Yuxuan Zhang, Yiren Song, Jiaming Liu, Rui Wang, Jinpeng Yu, Hao Tang, Huaxia Li, Xu Tang, Yao Hu, Han Pan, et al. Ssr-encoder: Encoding selective subject representation for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8069–8078, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.418, + 0.482, + 0.486 + ], + "angle": 0, + "content": "[30] Yuxuan Zhang, Yiren Song, Jinpeng Yu, Han Pan, and Zhongliang Jing. Fast personalized text to image synthesis with attention injection. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 6195-6199. IEEE, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.488, + 0.482, + 0.53 + ], + "angle": 0, + "content": "[31] Boyang Zheng, Chumeng Liang, Xiaoyu Wu, and Yan Liu. Understanding and improving adversarial attacks on latent diffusion model. arXiv preprint arXiv:2310.04687, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.531, + 0.482, + 0.584 + ], + "angle": 0, + "content": "[32] Zhengguang Zhou, Jing Li, Huaxia Li, Nemo Chen, and Xu Tang. Storymaker: Towards holistic consistent characters in text-to-image generation. arXiv preprint arXiv:2409.12576, 2024." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.584 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3028" + } + ] +] \ No newline at end of file diff --git a/2025/IDProtector_ An Adversarial Noise Encoder to Protect Against ID-Preserving Image Generation/cbe88f8c-13ab-4200-b1b9-bb453b62ad1d_origin.pdf b/2025/IDProtector_ An Adversarial Noise Encoder to Protect Against ID-Preserving Image Generation/cbe88f8c-13ab-4200-b1b9-bb453b62ad1d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..04f983335c5bd6d9e2bc89ed596a0653444ca04e --- /dev/null +++ b/2025/IDProtector_ An Adversarial Noise Encoder to Protect Against ID-Preserving Image Generation/cbe88f8c-13ab-4200-b1b9-bb453b62ad1d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cad2c47f79201e5b26a40e3c49062bda21679cc4aecd31cd235da374b809c4aa +size 2189000 diff --git a/2025/IDProtector_ An Adversarial Noise Encoder to Protect Against ID-Preserving Image Generation/full.md b/2025/IDProtector_ An Adversarial Noise Encoder to Protect Against ID-Preserving Image Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..610fb481117f404445c6c2cd4305402b8d26af6f --- /dev/null +++ b/2025/IDProtector_ An Adversarial Noise Encoder to Protect Against ID-Preserving Image Generation/full.md @@ -0,0 +1,254 @@ +# IDProtector: An Adversarial Noise Encoder to Protect Against ID-Preserving Image Generation + +Yiren Song* Pei Yang* Hai Ci† Mike Zheng Shou† Show Lab, National University of Singapore + +![](images/ac85b335f5891fd7e0110dce65aa9951d2ad253be68a637e5b245d66912b69e2.jpg) +Figure 1. Using only a single reference image, Encoder-based ID-preserving methods can generate realistic portraits, posing serious risks for malicious uses such as generation of malicious images. To prevent unauthorized ID-preserving generation, we propose IDProtector, which adds small perturbations to images. The protected images could disrupt ID-preserving generation by misleading the customization models to generate a dissimilar face, thereby achieving protection. + +# Abstract + +Recently, zero-shot methods like InstantID have revolutionized identity-preserving generation. Unlike multi-image finetuning approaches such as DreamBooth, these zero-shot methods leverage powerful facial encoders to extract identity information from a single portrait photo, enabling efficient identity-preserving generation through a single inference pass. However, this convenience introduces new threats to the facial identity protection. This paper aims to safeguard portrait photos from unauthorized encoder-based customization. We introduce IDProtector, an adversarial noise encoder that applies imperceptible adversarial noise to portrait photos in a single forward pass. Our approach offers universal protection for portraits against multiple state-of-the-art encoder-based methods, including InstantID, IP-Adapter, and PhotoMaker, while ensuring robustness to common image transformations such as JPEG + +compression, resizing, and affine transformations. Experiments across diverse portrait datasets and generative models reveal that IDProtector generalizes effectively to unseen data and even closed-source proprietary models. Project page: https://github.com/showlab/IDProtector. + +# 1. Introduction + +Recent advancements in diffusion models have opened new avenues for high-fidelity ID-preserving generation, widely used in applications such as personalized avatars. These methods can be categorized into tuning-based and encoder-based approaches. While tuning-based methods like Dreambooth[17] and LoRA[7] are effective, they require finetuning across multiple photos, which can be inconvenient. Encoder-based techniques, such as InstantID[26] and IP-Adapter[27], leveraging a face encoder to extract facial identity features in a zero-shot way, eliminating the need for finetuning and significantly simplifying the customization process. However, these advancements also + +intensify threats to facial privacy and security, making the protection of personal identity information in portrait photos increasingly critical. In previous research, AntiDreamBooth[24] utilized PGD[12] optimization against finetuning-based methods like LoRA and Dreambooth. Yet, there remains a gap in the literature regarding attacks on encoder-based ID-preserving generation. + +Encoder-based ID preservation technology evolves from domain-agnostic encoders to domain-specific ones. Initially, methods like Face0[23] and IP-Adapter[27] used a general CLIP[16] image encoder to extract facial features, yet they struggled to maintain the identity fidelity. IP-Adapter-Plus[27] addressed this issue by utilizing patch features before the projection layer. InstantID[26] introduced the use of a pre-trained Arcface[2] encoder, significantly enhancing identity consistency and influencing subsequent works like Photomaker [9] and PULID [4]. A natural progression is to attack both the CLIP image encoder and Arcface encoder simultaneously, targeting all current Encoder-based ID preservation technologies. + +There are four major challenges in ID protection: 1. Universality. Given the numerous existing encoder-based ID-preserving methods, we believe that universal protection is necessary, as attackers can easily switch to other methods to bypass protection. 2. Efficiency. Previous work typically relies on the Projected Gradient Descent (PGD) [12] to generate adversarial noise. However, PGD requires per-image optimization, taking several minutes per image, which significantly limits its practicality for protecting large volumes of portrait photos on social media in real world. 3. Robustness. The adversarial noise must be robust enough to withstand common image transformations, such as cropping, resizing, and JPEG compression, while also countering the inherent pre-processing operations of identity-preserving methods, such as face alignment. 4. Imperceptibility. The adversarial noise added for ID protection should be imperceptible to human eyes. + +In response to these challenges, this paper presents ID-Protector, the first feed-forward method to protect portrait photos from state-of-the-art encoder-based identity preservation techniques. Our approach incorporates several key innovations. For efficiency, we propose a ViT-based encoder that directly predicts adversarial noise for a given target portrait. To ensure universality, we carefully examine the attack surfaces and define a composite loss function tailored to four mainstream encoder-based generation methods, including InstantID [26], IP-Adapter [27], IP-Adapter Plus [27], and PhotoMaker [9]. Additionally, we employ an L1-based image regularization to minimize the visual impact of adversarial noise on image quality. For robustness against common transformations, a typical approach involves a combination of post-hoc differentiable augmentations like differentiable JPEG and cropping during train + +ing. However, we find that this method introduces a substantial learning burden. To address this, we propose a surrogate method that leverages the face preprocessing pipeline of existing identity-preserving methods. Specifically, we introduce small random noise into the affine transformation matrix during face alignment stage to achieve data augmentation. This not only enhances the robustness of adversarial noise against common image transformations but also enables it to withstand image preprocessing operations during identity-preserving generation. + +We train and extensively evaluate IDProtector on the CelebA dataset [11]. We further test it on unseen data VGG Face [15] and proprietary models like Midjourney [13] and Jing Gou, demonstrating its strong generalization to unseen data and closed-source models. By advancing adversarial attack techniques from theoretical constructs to practically deployable solutions, IDProtector sets a new benchmark in facial privacy protection. We summarize our main contributions as follows: + +- We introduce IDProtector, the first feed-forward method to protect portrait photos from SOTA encoder-based identity preservation techniques. +- We present several innovative techniques including a ViT-based noise encoder, a joint learning framework with carefully designed objectives, and affine-based data augmentation to simultaneously achieve the EURI standard (Efficiency, Universality, Robustness, and Imperceptibility). + +# 2. Related Works + +# 2.1. Tuning-based Customized Generation + +With the rapid advancement of diffusion models, customized generation methods have begun to emerge. Techniques like DreamBooth [17], Custom Diffusion [8], Low-Rank Adaptation (LoRA) [7] and Textual Inversion [3] facilitate the adaptation of text-to-image diffusion models through fine-tuning strategies that range from adjusting all parameters to optimizing specific word vectors. Despite their impressive effects, these methods require extensive computational resources and time due to their need for multi-image fine-tuning. This can make them impractical for real-time or large-scale applications. + +# 2.2. Encoder-based ID-Preserving Generation + +Encoder-based Customized generation methods[23, 27, 29, 30] utilize pre-trained encoders to extract appearance feature, facilitating zero-shot generation in real-time. These methods evolved from domain-agnostic encoders to domain-specific encoders. Initially, CLIP[16] image encoder was used to extract facial features, such as with Face0[23] and IP-Adapter[27]; however, their maintenance of facial detail fidelity was not satisfactory. IP-Adapter + +Plus addressed this issue by utilizing patch features before the projection layer. InstantID[26] was the first to propose using a pre-trained Arcface[2] Encoder as a condition, significantly enhancing identity consistency and influencing subsequent works like Photomaker[9] and PULID[4]. StoryMaker[32] employs both the Arcface image encoder and the CLIP image encoder to encode facial features and broader regions, respectively, preserving not only facial identities but also clothing, hairstyles, and body features. Additionally, encoder-based ID preservation technology is widely applied in multi-human image personalization[5, 14] and video tasks[6]. + +# 2.3. Defensive Perturbations for Diffusion Models. + +The focus on safeguarding personal artworks and facial images from unauthorized usage in the fine-tuning of diffusion models has led to the development of subtle yet effective perturbations. As summarized in Tab. 1, distinct approaches have been tailored to address this issue: Glaze [20] strategically adjusts the feature-space distances to misdirect the style learning process in Stable Diffusion, effectively preventing style appropriation. Conversely, AdvDM [10] implements adversarial disruptions by maximizing the Mean Squared Error during the optimization phase to shield personal visuals. Meanwhile, Anti-DreamBooth [24] integrates adversarial interference directly within the DreamBooth fine-tuning cycle, applying a complex optimization strategy to formulate defensive perturbations. Further investigations [21, 25, 28, 31] into adversarial protection have demonstrated their effectiveness in obstructing finetuning-based customization methods such as DreamBooth [17], LoRA [7], and textual inversion [3]. However, how to protect personal identity from the recently popular encoder-based methods remains unexplored. This paper endeavors to bridge the current gaps by proposing advanced methodologies to shield images from encoder-based exploitations, enhancing the security measures for personal digital assets. + +Table 1. Overview of encoder-based methods for ID-preserving generation. + +
EncoderID-Preserving Methods
CLIP [16]IP-Adapter [27], IP-Adapter Plus [27], PhotoMaker [9], Face0 [23]
ArcFace [2]InstantID [26], PhotoMaker [9], PuLID [4], IP-Adapter-FaceID [27]
CLIP & ArcFaceIDAdapter [1], Storymaker [32]
+ +# 3. Problem Definition + +We consider two participants in a portrait identity protection scenario.: "image protector" Alice and "image exploiter" Bob. Bob seeks to use an encoder-based, identity-preserving approach to alter the portrait photos Alice shares on social media for malicious purposes. While Alice aims to protect her portrait by adding adversarial noise to thwart Bob's methods, inducing distortions or significant decreases + +in ID similarity in the generated content. Specifically, we explain the workflow of the two parties as follows: + +1. Image Protector Alice: The image protector aims to provide protection for images against exploitation by encoder-based ID-preserving methods. In this context, the chosen protection method involves adding imperceptible protective perturbations to the images, with the goal of providing protection while minimizing modifications to the original image. In real-world scenarios, image protectors encounter several challenges. For instance, implementing universal protection to guard against various encoder-based, identity-preserving methods is difficult. Additionally, the protection should be robust to natural distortions such as resizing, JPEG compression, and blurring. +2. Image Exploiter Bob: Bob aims to use portrait photos for ID-preserving customization. Bob can choose any open-source or proprietary encoder-based models, which poses significant challenges on the universality of the protection method. + +Objective of the Study: Given an image $x$ containing a face portrait, our objective is to generate adversarial perturbations $\delta$ that protect against ID-preserving generation. We aim to maximize the difference in identity consistency between the unperturbed generated image $g(x)$ and the perturbed generated image $g(x + \delta)$ , where $g$ is an ID-preserving generator. To unify the protection against different models, we can find a $\delta$ that simultaneously disrupts several feature spaces, lowering the sum of the similarity values. Thus, we train a noise encoder network $E_{\theta}$ , solving + +$$ +\min _ {\theta} S \left\{g _ {i} (x), g _ {i} (x + \delta) \right\}, \quad \delta = E _ {\theta} (x) \tag {1} +$$ + +subject to the constraint that $\| \delta \|_{\infty} < \epsilon$ . Here, $S$ denotes an identity similarity metric, which we evaluate using cosine similarity in feature spaces of ArcFace [2] and CLIP [16]. + +# 4. Method + +The core of our method is an adversarial noise encoder model that can add adversarial noise to facial images for protection. We will first introduce the complete workflow in Sec. 4.1, followed by the design of the adversarial noise encoder in Sec. 4.2. Sec. 4.3 and 4.4 introduce ID protection objectives and loss functions. Sec. 4.5 introduces the data augmentation approach to enhance the robustness. + +# 4.1. Overall Method + +Fig. 2 details the pipeline of our method. For an arbitrary-sized input image $x$ containing a face, it is resized to $224 \times 224$ and sent into the adversarial noise encoder model, whose output is an additive adversarial noise that can be resized back and perturbed onto the original image for protection. During adversarial noise encoder training, this protected image is further passed through face feature extrac + +![](images/ee7bfd173de19bfe5dea26574c7db9c30d07cbe1054171e65fd06d3498026cb6.jpg) +Figure 2. Overall schematics of our method. Our method's key design includes the noise encoder, loss functions, and the gradient optimization path that allows backpropagation. During the protection phase, an image is first resized to $224 \times 224$ and sent to an adversarial noise encoder model. The model outputs a perturbation that can be resized back and added to the image, achieving an adversarial attack effect – the protected image prevents subsequent pipelines from correctly extracting facial feature information, thereby hindering face similarity in face generation tasks. This achieves the goal of ID protection. + +tion pipelines to obtain the loss functions that aim at creating maximum differences in the feature space. + +# 4.2. Noise Encoder + +The network takes a $224 \times 224$ RGB image as input and outputs three-channel adversarial perturbations. The perturbations are projected to a range of $[-1, 1]$ , then denormalized $[- \epsilon, \epsilon]$ , resized to the original image dimensions, and added to the image. We utilize a Vision Transformer (ViT) to generate the adversarial noise, with both input and output dimensions set to $224 \times 224$ . Empirically, we find that adding an additional prior mask channel, indicating the location of the face, aids in training. This face localization mask is generated by the InsightFace pipeline based on facial landmark points, which specify the region of the face within the image. The mask is concatenated to the image as the fourth input channel, eliminating the need for the network to learn face localization ability from scratch, thus reducing training difficulty. + +# 4.3. Adversarial ID Protection + +Since generative networks' only knowledge of faces comes from feature embeddings, to achieve ID protection, we can make the extracted features from protected images significantly different from the originals, thereby preventing generation models from accessing correct facial features. Based on this principle, we analyze the specific pipelines of InstantID, IP-Adapter, IP-Adapter-Plus, and PhotoMaker to design targeted attack objectives and loss functions. + +InstantID obtains facial features in two steps, as shown in Fig. 2. It first aligns the face to a pre-defined location, before feeding the aligned face into an ArcFace feature extractor to obtain facial information. We choose to fool the ArcFace model, minimizing the cosine similarity between + +the ArcFace features of the original v.s. protected images. + +IP-Adapter, IP-Adapter-Plus, and PhotoMaker all rely on facial features extracted by the CLIP vision encoder, albeit using different versions of CLIP vision. IP-Adapter and PhotoMaker utilize CLIP Vision's output, while IP-Adapter-Plus takes features before CLIP Vision's second-to-last layer. As illustrated in Fig. 2, embeddings from different layers offer various attack surfaces. We select victim embeddings for attack according to three principles: (1) blocking all potential pathways, ensuring that any information flow from left to right passes through at least one victim embedding, ensuring all information flow would be disrupted; (2) choosing features as early as possible in the network (towards the left in Fig. 2), to reduce the backpropagation path length and thus simplify optimization; and (3) targeting embeddings with dense semantic information for more effective manipulation. + +An example of an unsuitable target is the last hidden layer feature of CLIP Vision. While modifying this feature could influence all three models, its huge dimension $(257\times 1280)$ and lack of semantic density make it highly challenging for optimization to converge. Ultimately, we selected the embeddings marked red in Fig. 2 as the primary attack targets. The objective, similar to that of InstantID, is to maximize the cosine similarity between the perturbed embedding and the original embedding, effectively aligning them post-attack. Therefore, the final adversarial loss is a weighted average of all losses, given by + +$$ +L _ {\mathrm {a d v}} = \sum_ {\text {C u t s o m M o d e l} i} \alpha_ {i} \cdot \operatorname {c o s s i m} \left(e _ {i}, e _ {i} ^ {\prime}\right), \tag {2} +$$ + +where $e_i$ and $e_i'$ denote the face embeddings of the clean and protected images, respectively. + +# 4.4. Imperceptibility + +To minimize the visual impact on the image quality, we enforce an $\ell_1$ regularization on the predicted adversarial noise $\delta$ . We further introduce an auxiliary penalty on any noise values exceeding the $\epsilon$ -ball boundary. Together, these terms constitute our regularization loss, given by: + +$$ +L _ {\text {r e g}} = \beta_ {1} \cdot \| \delta \| _ {1} + \beta_ {2} \cdot \| \delta - \operatorname {c l i p} _ {\pm \epsilon} (\delta) \| _ {1}. \tag {3} +$$ + +The final loss is thus the sum of adversarial loss and regularization, given by: + +$$ +L = L _ {\mathrm {a d v}} + L _ {\mathrm {r e g}}. \tag {4} +$$ + +# 4.5. Robustness + +As shown in Fig. 2, the InstantID pipeline involves a step of facial alignment aiming to align any faces to a predefined standard face for better feature extraction. The alignment process includes detecting the face and key points $\tilde{p}$ , calculating the affine matrix $A$ with standard face key points $\tilde{p}'$ , and aligning them. This process is computationally expensive. To reduce computational cost and simplify training, we pre-compute $A$ for each image and keep it fixed throughout training. However, during inference, perturbations carried by protected images may slightly alter the precomputed affine matrix $A$ , leading to minor misalignment that weakens the effectiveness the adversarial perturbations. + +To adapt to this misalignment as well as being computationally efficient, we leverage the existing affine transformation operation and introduce small gaussian noise to $A$ during training to simulate face transformations. The noise level is empirically set to $\sigma = 0.003$ . The face alignment transformation is therefore given by + +$$ +\tilde {p} ^ {\prime} = (A + N (0, \sigma I)) \tilde {p}. \tag {5} +$$ + +We incorporate this augmentation only in the InstantID branch. Our experiments reveal that it also enhances the robustness of other branches, being resilient to various common image transformations. + +# 5. Experiments + +# 5.1. Experimental Setup + +Training Data. For practicality, we selected 190,000 uncropped images from the CelebA dataset [11] as the training set. These images have arbitrary sizes and aspect ratios. The test set includes two subsets: 50 unseen images from CelebA and 50 images from VGG Face [15]. To ensure compatibility with baseline methods, these images were resized to $512 \times 512$ and $224 \times 224$ , respectively. + +Training Details. We trained a ViT-S/8 adversarial noise encoder from scratch using Eq. 4 as the loss function. We + +Table 2. Curriculum learning stages with corresponding hyperparameters. $\eta$ represents learning rate. ${\alpha }_{1}$ to ${\alpha }_{4}$ are weights for IP-Adapter,IP-Adapter-Plus,PhotoMaker,InstantID,respectively. + +
StageEpochsα1α2α3α4β1β2εη
11202/3714/3720/371/379e-31e-30.051e-2
2204/162/169/162/161.8e-32e-40.042e-5
3202/142/149/141/144.5e-45e-50.0352e-5
+ +employed a three-stage curriculum learning approach, progressively shrinking the $\epsilon$ -ball and balancing the different losses by adjusting the hyperparameters (presented in Table 2). Training began with a 2,500-step linear warm-up, gradually increasing the learning rate from zero. Gradient clipping was applied during training to cap the $\ell_2$ norm of gradients at a maximum of 10. The model was trained with a batch size of 112 on four 80GB NVIDIA H100 GPUs for 10 days. + +Evaluation process and metrics. To evaluate the protection performance of different methods, for each face-containing image in the dataset (denoted as NP ref image), we apply different protection methods (including our ID-Protector) to generate its protected counterpart (denoted as P ref image). Throughout all experiments, we constrain the maximum pixel perturbation to 9/255 (an $\epsilon$ -ball of 0.035), following common practice in adversarial machine learning [18]. Subsequently, both P and NP ref images are fed into customization generation models (such as InstantID) to obtain generated images (denoted as P/NP gen images). + +We evaluate the identify shift caused by different protection methods through Identity Score Matching (ISM), which is calculated as the cosine similarity between facial features of NP images and generated images extracted by ArcFace [2]. Lower ISM values indicate better protection as they suggest a greater identity shift between original (NP ref) and generated (P gen) faces. + +Since protected images may result in no detectable faces being generated, we also report the Face Detection Rate (FDR) using InsightFace detector, as ISM can only be computed when faces are detected in both images. Additionally, to ensure comprehensive evaluation, we measure the perceptual quality of each generated face using SER-FIQ [22]. + +# 5.2. Comparison with Baselines + +Protection Performance. We compare our method with the following baseline methods: ACE [31], AdvDM [10], Anti-DreamBooth [24], PhotoGuard [19], and SimAC [25]. Most baselines aim to reduce the quality of generated images or hinder the diffusion denoising process, but we aim to reduce the similarity of generated identities so as to directly disrupt identity-preserving generation. To compare our method with baselines, we first applied each protection method to generate protected versions of test images. We then used both the original images (NP) and the protected + +![](images/425ffa070c84b8fa4527206248c51af6132b7951f62e8a74bb4a3af856ef2bd6.jpg) +Figure 3. Qualitative comparison with baseline. Images protected by our method cannot be used to generate similar faces. Compared to baseline protection methods, under the same perturbation bound of $9/255$ ( $\epsilon = 3.5\%$ ), ours causes more significant facial differences. + +versions as references in customized generation tasks with four models: IP-Adapter, IP-Adapter-Plus, PhotoMaker, and InstantID. Tab. 3 shows that, for the first three methods, while baseline methods reduced ISM by an average of 0.02, with a maximum reduction of 0.06, our method achieved an ISM reduction between 0.1 and 0.14, demonstrating a larger identity shift. On InstantID, our IDProtector demonstrated even more pronounced protection: when baseline methods reduced facial similarity by less than 0.1, our method reduced similarity by over 0.4 across both datasets. Additionally, FDR results validate the reliability of ISM as FDR reflects the percentage of images with a face detected and contributes to ISM calculations. Overall, these results underscore the superior identity protection performance of our method compared to existing approaches. Together, these highlight the superior identity protection performance of + +# our approach. + +Following previous study [24], we evaluate SER-FIQ as a quality metric of generated (P/NP gen) images. Although our method has a noticeable impact on image generation quality (as shown in Fig. 3), we observed inconsistencies between the SER-FIQ scores and qualitative assessments. This discrepancy suggests that a more robust and reliable quality metric may be required to accurately measure the quality of generated images in future research. + +We also adapted our method to a per-image tuning variant, employing PGD [12] as the optimization technique. We observed that PGD achieves a more substantial ISM reduction across most methods, which aligns with our expectations since per-image tuning can better leverage the specific characteristics of each individual image. However, a notable drawback of PGD is the prolonged optimization time, + +Table 3. Comparison of identity protection performance against baseline methods. Protection performance is evaluated against IP-Adapter, IP-Adapter-Plus, PhotoMaker, and InstantID. "No Protect" denotes direct use of the original (NP ref) images for identity-preserving generation. Experiments were conducted on five prompts per dataset, and average results are reported. + +
DatasetProtection MethodIP-AdapterIP-Adapter-PlusPhotoMakerInstantID
FDR ↓ISM ↓SER-FIQ ↓FDR ↓ISM ↓SER-FIQ ↓FDR ↓ISM ↓SER-FIQ ↓FDR ↓ISM ↓SER-FIQ ↓
CelebANo Protect1.0000.1760.6411.0000.2400.6630.6640.2650.6050.9960.7750.774
ACE1.0000.1590.6401.0000.2310.6810.6360.2680.5941.0000.7450.758
AdvDM0.9880.1570.6301.0000.2260.6660.6400.2550.5901.0000.7570.757
Anti-DB1.0000.1550.6151.0000.2220.6760.6400.2550.5991.0000.7530.752
PhotoGuard1.0000.1520.6211.0000.2230.6700.6520.2590.5921.0000.7580.767
SimAC1.0000.1620.6271.0000.2170.6830.6520.2530.5991.0000.7520.755
Ours (PGD)0.8440.0240.5440.9120.0720.7180.1120.0940.6390.996-0.1660.721
Ours0.9520.0600.6880.9960.1140.6440.6720.1230.6221.0000.2310.613
VGG FaceNo Protect1.0000.2010.6721.0000.2360.6980.6320.2540.6000.9960.7640.828
ACE0.9880.1540.6171.0000.2090.6831.0000.2090.6831.0000.7030.829
AdvDM1.0000.1470.5971.0000.2150.6931.0000.2150.6931.0000.7060.820
Anti-DB1.0000.1380.5971.0000.2100.6780.6240.2380.5991.0000.7070.811
PhotoGuard1.0000.1590.6431.0000.2170.6860.6240.2290.6120.9960.6910.799
SimAC1.0000.1500.6051.0000.2020.6770.6240.2280.6441.0000.7070.826
Ours (PGD)0.9680.0430.6091.0000.2020.7400.6240.2280.6301.0000.7070.769
Ours1.0000.1050.7371.0000.1380.6870.6320.1260.6141.0000.2320.713
+ +Table 4. Comparison of time cost and quality impact with an epsilon-budget of 9/255 (±3.5%). PSNR and SSIM are calculated between clean (NP ref) and perturbed (P ref) image pairs. + +
ACEAdvDMAnti-DBPhoto-GuardSimACOurs PGDOurs
Time (s)87.051.926123.332340890.173
PSNR (dB)31.0631.7031.7031.1032.3732.9432.15
SSIM (dB)0.7410.7660.7650.7570.7910.8450.809
+ +which we will address in the following section. + +The protection effects are visualized in Fig. 3. The baseline methods can protect against some of the generation methods. For example, AdvDM and Anti-DreamBooth cause IP-Adapter to produce faces with color block artifacts, while several outputs from IP-Adapter-Plus exhibit unnatural patterns and shadows. However, the baselines are ineffective in protecting against PhotoMaker and InstantID. With IDProtector's protection, faces generated by all four methods appear significantly altered. In some cases, even the gender of the person is unidentifiable. This fulfills our protection goal: protected images could not be used for identity-preserving generation. + +Time and Image Quality. Our IDProtector is fast to protect an image with less degradation to image quality. As compared in Table 4, the PSNR and SSIM of our method are higher than the baselines, indicating less visual impact. The baseline methods require an average of 23.3 to 323 seconds per image for protection. Using PGD with our optimization objective would take over an hour on average. In contrast, our ViT-based IDProtector achieves an average protection time of only 0.173 seconds per image – less than $1\%$ of the fastest baseline. Note that the forward-feeding ViT supports better parallelization compared with PGD, the time cost could further decrease with mini-batch inference. + +Table 5. Generalization of ID protection by our IDProtector on unseen generators. IPA-FID+/P/XL denotes IP-Adapter-FaceID/Plus/Portrait/SDXL, with NP and P indicating no and with protection, respectively. + +
IPA-FIDIPA-FID+IPA-FIDPIPA-FIDXLFlux-IPAStory-MakerMid-journeyJingGou
ISM ↓NP0.5650.6440.1290.5230.0540.5720.1990.672
P0.1530.2170.0520.1320.0290.1930.0980.131
SER-NP0.7060.7510.7200.4020.7280.6990.5680.791
FIQ ↓P0.8030.7300.6580.6670.6730.7340.6270.805
+ +![](images/8d3a17b59eb2baa69c8daf3c6d0e744ae0b33a82b77c500f6b1ee1f4b3685d83.jpg) +Figure 4. ID protection performance on unseen generators. + +# 5.3. Generalization to Unseen Data and Models + +Generalization to unseen datasets. We evaluated our method on the VGG Face dataset that was unseen during training. All following experiments are conducted on the CelebA test set unless otherwise specified. As shown in Tab. 3, the ISM reduction achieved by our method shows minimal variation across datasets. For example, our IDPro + +Table 6. Robustness evaluation of IDProtector under image distortions. NP denotes non-protected images for identity-preserving generation, while P represents protected but non-distorted images, providing an ablation. Distortion averages exclude NP and P. + +
IDProtector (ours, ViT)Adaptive PGD
IP-AdapterIP-Adapter+PhotoMakerInstantIDIP-AdapterIP-Adapter+PhotoMakerInstantID
FDR ↓ISM ↓FDR ↓ISM ↓FDR ↓ISM ↓FDR ↓ISM ↓FDR ↓ISM ↓FDR ↓ISM ↓FDR ↓ISM ↓FDR ↓ISM ↓
NP1.0000.1761.0000.2400.6640.2650.9960.7750.8440.0240.9120.0720.1120.0940.996-0.166
P0.9520.0600.9960.1140.6720.1231.0000.2310.8440.0240.9120.0720.1120.0940.996-0.166
P+Affine1.0000.0681.0000.1060.6560.1351.0000.2160.8440.0640.9760.0810.112-0.0091.000-0.147
P+JPEG0.9640.0691.0000.0970.6600.1381.0000.2410.7840.0150.9480.0620.108-0.0121.000-0.161
P+Crop1.0000.0721.0000.1080.6520.1321.0000.2160.7520.0580.9400.0700.1080.0201.000-0.167
P+Noisy0.9520.0690.9840.1140.6560.1321.0000.2520.8320.0230.9320.0720.112-0.0101.000-0.130
Avg (P+)0.9790.0700.9960.1060.6560.1341.0000.2310.8030.0400.9490.0710.110-0.0031.000-0.151
+ +![](images/8aa4b299e8c0047648a0d6a052476a39670289cf352eebd33d7d2302bbdf0968.jpg) +Figure 5. Robustness evaluation of IDProtector under various distortions. Despite geometric distortions like affine transformations, IDProtector effectively misleads all four models into generating dissimilar faces or even artifacts. + +tector reduces the ISM of IP-Adapter by approximately 0.1 and InstantID by around 0.55 on both datasets. This result suggests that our method generalises effectively beyond the CelebA dataset, demonstrating protection capabilities that are not restricted to in-domain faces. + +Generalization to unseen customization models. Beyond the four customization models seen during training, we tested protected images on a wider range of customization models. Across these models, our protection approach achieves an average ISM reduction of 0.31, with a maximum reduction of 0.54. Notably, our method remains effective against proprietary commercial models such as Midjourney (0.10 reduction) and Jing Gou (0.54 reduction), underscoring its ability to generalise to unseen customization models. + +These protection effects are visualised in Fig. 4. As shown, protected images used in these generators cause noticeable quality degradation, including altered facial shapes (e.g., IP-Adapter-FaceID-XL), reduced prompt-following + +ability, and image contamination (e.g., Midjourney). The common result is a decreased resemblance between the generated face and the reference identity. + +# 5.4. Robustness Evaluation + +Our method demonstrates robustness against various distortions. For protected (P) images, we applied several distortions, including JPEG compression (quality=85), random cropping that removes up to $20\%$ of the region, noise perturbation with a variance of 100, and an affine transformation. The affine transformation was implemented as $\tilde{q} = (A + \mathcal{N}(0,0.05^2))\tilde{p}$ , where the matrix $A$ is an identity matrix. As shown in Tab. 6, the ISM values under these distortions remain comparable to those obtained without distortion, indicating that our protection remains effective under distortion. Fig. 5 provides a visual comparison, where even under various distortions, InstantID fails to reconstruct the correct female face, generating a male portrait instead. This demonstrates that the common distortions do not compromise the efficacy of our protection method. + +# 6. Conclusion and Limitation + +In this paper, we introduce IDProtector, a method designed to protect facial identities from encoder-based identity-preserving generation. By training a noise encoder to add adversarial perturbations onto images, IDProtector can protect an image within 0.2s. Our method achieves significantly lower face similarity in generated images than existing protection methods. IDProtector also demonstrates strong practical applicability. Its protection remains effective under distortions such as affine transformations and JPEG compression. Furthermore, IDProtector not only shows effectiveness against closed-source models, such as Midjourney, but also unifies protection against multiple ID-preserving models, setting a new standard for privacy protection against generative AI. + +A limitation of our method is that the adversarial noise is not fully invisible, necessitating a trade-off between attack effectiveness and Imperceptibility. Future efforts will focus on enhancing the Imperceptibility of the noise while maintaining its robustness. + +# 7. Acknowledgements + +# References + +[1] Siying Cui, Jia Guo, Xiang An, Jiankang Deng, Yongle Zhao, Xinyu Wei, and Ziyong Feng. Idadapter: Learning mixed features for tuning-free personalization of text-to-image models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 950-959, 2024. +[2] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. +[3] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. +[4] Zinan Guo, Yanze Wu, Zhuowei Chen, Lang Chen, and Qian He. Pulid: Pure and lightning id customization via contrastive alignment. arXiv preprint arXiv:2404.16022, 2024. +[5] Junjie He, Yifeng Geng, and Liefeng Bo. Uniportrait: A unified framework for identity-preserving single- and multi-human image personalization. arXiv preprint arXiv:2408.05939, 2024. +[6] Xuanhua He, Quande Liu, Shengju Qian, Xin Wang, Tao Hu, Ke Cao, Keyu Yan, Man Zhou, and Jie Zhang. Id-animator: Zero-shot identity-preserving human video generation. arXiv preprint arXiv:2404.15275, 2024. +[7] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. +[8] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. +[9] Zhen Li, Mingdeng Cao, Xintao Wang, Zhongang Qi, Ming-Ming Cheng, and Ying Shan. Photomaker: Customizing realistic human photos via stacked id embedding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8640-8650, 2024. +[10] Chumeng Liang, Xiaoyu Wu, Yang Hua, Jiaru Zhang, Yiming Xue, Tao Song, Zhengui Xue, Ruhui Ma, and Haibing Guan. Adversarial example does good: Preventing painting imitation from diffusion models via adversarial examples. arXiv preprint arXiv:2302.04578, 2023. +[11] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Large-scale celebfaces attributes (celeba) dataset. Retrieved August, 15(2018):11, 2018. + +This research is supported by the Ministry of Education, Singapore, under its Academic Research Fund Tier 2 (Award No: MOE-T2EP20124-0012). +[12] Aleksander Madry. Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083, 2017. +[13] Midjourney.com. Midjourney. https://www.midjourney.com/home. +[14] Daniil Ostashev, Yuwei Fang, Sergey Tulyakov, Kfir Aberman, et al. Moa: Mixture-of-attention for subject-context disentanglement in personalized image generation. arXiv preprint arXiv:2404.11565, 2024. +[15] Omkar Parkhi, Andrea Vedaldi, and Andrew Zisserman. Deep face recognition. In BMVC 2015-Proceedings of the British Machine Vision Conference 2015. British Machine Vision Association, 2015. +[16] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. +[17] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. +[18] Mehrdad Saberi, Vinu Sankar Sadasivan, Keivan Rezaei, Aounon Kumar, Atoosa Chegini, Wenxiao Wang, and Soheil Feizi. Robustness of ai-image detectors: Fundamental limits and practical attacks. arXiv preprint arXiv:2310.00076, 2023. +[19] Hadi Salman, Alaa Khaddaj, Guillaume Leclerc, Andrew Ilyas, and Aleksander Madry. Raising the cost of malicious ai-powered image editing. arXiv preprint arXiv:2302.06588, 2023. +[20] Shawn Shan, Jenna Cryan, Emily Wenger, Haitao Zheng, Rana Hanocka, and Ben Y Zhao. Glaze: Protecting artists from style mimicry by {Text-to-Image} models. In 32nd USENIX Security Symposium (USENIX Security 23), pages 2187-2204, 2023. +[21] Yiren Song, Shengtao Lou, Xiaokang Liu, Hai Ci, Pei Yang, Jiaming Liu, and Mike Zheng Shou. Anti-reference: Universal and immediate defense against reference-based generation. arXiv preprint arXiv:2412.05980, 2024. +[22] Philipp Terhorst, Jan Niklas Kolf, Naser Damer, Florian Kirchbuchner, and Arjan Kuijper. Ser-fiq: Unsupervised estimation of face image quality based on stochastic embedding robustness. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5651-5660, 2020. +[23] Dani Valevski, Danny Wasserman, Yossi Matias, and Yaniv Leviathan. Face0: Instantaneously conditioning a text-to-image model on a face. arXiv preprint arXiv:2306.06638, 2023. +[24] Thanh Van Le, Hao Phung, Thuan Hoang Nguyen, Quan Dao, Ngoc N Tran, and Anh Tran. Anti-dreambooth: Protecting users from personalized text-to-image synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2116-2127, 2023. + +[25] Feifei Wang, Zhentao Tan, Tianyi Wei, Yue Wu, and Qidong Huang. Simac: A simple anti-customization method for protecting face privacy against text-to-image synthesis of diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12047-12056, 2024. +[26] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, and Anthony Chen. Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519, 2024. +[27] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. +[28] Xiaoyu Ye, Hao Huang, Jiaqi An, and Yongtao Wang. Duaw: Data-free universal adversarial watermark against stable diffusion customization. arXiv preprint arXiv:2308.09889, 2023. +[29] Yuxuan Zhang, Yiren Song, Jiaming Liu, Rui Wang, Jinpeng Yu, Hao Tang, Huaxia Li, Xu Tang, Yao Hu, Han Pan, et al. Ssr-encoder: Encoding selective subject representation for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8069–8078, 2024. +[30] Yuxuan Zhang, Yiren Song, Jinpeng Yu, Han Pan, and Zhongliang Jing. Fast personalized text to image synthesis with attention injection. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 6195-6199. IEEE, 2024. +[31] Boyang Zheng, Chumeng Liang, Xiaoyu Wu, and Yan Liu. Understanding and improving adversarial attacks on latent diffusion model. arXiv preprint arXiv:2310.04687, 2023. +[32] Zhengguang Zhou, Jing Li, Huaxia Li, Nemo Chen, and Xu Tang. Storymaker: Towards holistic consistent characters in text-to-image generation. arXiv preprint arXiv:2409.12576, 2024. \ No newline at end of file diff --git a/2025/IDProtector_ An Adversarial Noise Encoder to Protect Against ID-Preserving Image Generation/images.zip b/2025/IDProtector_ An Adversarial Noise Encoder to Protect Against ID-Preserving Image Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6d2b81b827e8280ec553ee9dc146804ee779ac42 --- /dev/null +++ b/2025/IDProtector_ An Adversarial Noise Encoder to Protect Against ID-Preserving Image Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1fa4e5eb1281b23d4b9af300a9eea355f165c463875cc9d340b44c6dd882812f +size 990816 diff --git a/2025/IDProtector_ An Adversarial Noise Encoder to Protect Against ID-Preserving Image Generation/layout.json b/2025/IDProtector_ An Adversarial Noise Encoder to Protect Against ID-Preserving Image Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..487852d15898634358f1077a9ea679d02b6f507b --- /dev/null +++ b/2025/IDProtector_ An Adversarial Noise Encoder to Protect Against ID-Preserving Image Generation/layout.json @@ -0,0 +1,6304 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 67, + 102, + 544, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 102, + 544, + 140 + ], + "spans": [ + { + "bbox": [ + 67, + 102, + 544, + 140 + ], + "type": "text", + "content": "IDProtector: An Adversarial Noise Encoder to Protect Against ID-Preserving Image Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 173, + 160, + 441, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 160, + 441, + 191 + ], + "spans": [ + { + "bbox": [ + 173, + 160, + 441, + 191 + ], + "type": "text", + "content": "Yiren Song* Pei Yang* Hai Ci† Mike Zheng Shou† Show Lab, National University of Singapore" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 59, + 244, + 545, + 400 + ], + "blocks": [ + { + "bbox": [ + 59, + 244, + 545, + 400 + ], + "lines": [ + { + "bbox": [ + 59, + 244, + 545, + 400 + ], + "spans": [ + { + "bbox": [ + 59, + 244, + 545, + 400 + ], + "type": "image", + "image_path": "ac85b335f5891fd7e0110dce65aa9951d2ad253be68a637e5b245d66912b69e2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 404, + 555, + 449 + ], + "lines": [ + { + "bbox": [ + 55, + 404, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 404, + 555, + 449 + ], + "type": "text", + "content": "Figure 1. Using only a single reference image, Encoder-based ID-preserving methods can generate realistic portraits, posing serious risks for malicious uses such as generation of malicious images. To prevent unauthorized ID-preserving generation, we propose IDProtector, which adds small perturbations to images. The protected images could disrupt ID-preserving generation by misleading the customization models to generate a dissimilar face, thereby achieving protection." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 464, + 200, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 464, + 200, + 477 + ], + "spans": [ + { + "bbox": [ + 151, + 464, + 200, + 477 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 490, + 297, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 490, + 297, + 682 + ], + "spans": [ + { + "bbox": [ + 53, + 490, + 297, + 682 + ], + "type": "text", + "content": "Recently, zero-shot methods like InstantID have revolutionized identity-preserving generation. Unlike multi-image finetuning approaches such as DreamBooth, these zero-shot methods leverage powerful facial encoders to extract identity information from a single portrait photo, enabling efficient identity-preserving generation through a single inference pass. However, this convenience introduces new threats to the facial identity protection. This paper aims to safeguard portrait photos from unauthorized encoder-based customization. We introduce IDProtector, an adversarial noise encoder that applies imperceptible adversarial noise to portrait photos in a single forward pass. Our approach offers universal protection for portraits against multiple state-of-the-art encoder-based methods, including InstantID, IP-Adapter, and PhotoMaker, while ensuring robustness to common image transformations such as JPEG" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 465, + 556, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 465, + 556, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 465, + 556, + 525 + ], + "type": "text", + "content": "compression, resizing, and affine transformations. Experiments across diverse portrait datasets and generative models reveal that IDProtector generalizes effectively to unseen data and even closed-source proprietary models. Project page: https://github.com/showlab/IDProtector." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 550, + 395, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 550, + 395, + 562 + ], + "spans": [ + { + "bbox": [ + 314, + 550, + 395, + 562 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 571, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 571, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 571, + 555, + 715 + ], + "type": "text", + "content": "Recent advancements in diffusion models have opened new avenues for high-fidelity ID-preserving generation, widely used in applications such as personalized avatars. These methods can be categorized into tuning-based and encoder-based approaches. While tuning-based methods like Dreambooth[17] and LoRA[7] are effective, they require finetuning across multiple photos, which can be inconvenient. Encoder-based techniques, such as InstantID[26] and IP-Adapter[27], leveraging a face encoder to extract facial identity features in a zero-shot way, eliminating the need for finetuning and significantly simplifying the customization process. However, these advancements also" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 694, + 135, + 704 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 694, + 135, + 704 + ], + "spans": [ + { + "bbox": [ + 66, + 694, + 135, + 704 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 67, + 704, + 144, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 704, + 144, + 714 + ], + "spans": [ + { + "bbox": [ + 67, + 704, + 144, + 714 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3019" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "intensify threats to facial privacy and security, making the protection of personal identity information in portrait photos increasingly critical. In previous research, AntiDreamBooth[24] utilized PGD[12] optimization against finetuning-based methods like LoRA and Dreambooth. Yet, there remains a gap in the literature regarding attacks on encoder-based ID-preserving generation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 158, + 294, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 158, + 294, + 314 + ], + "spans": [ + { + "bbox": [ + 55, + 158, + 294, + 314 + ], + "type": "text", + "content": "Encoder-based ID preservation technology evolves from domain-agnostic encoders to domain-specific ones. Initially, methods like Face0[23] and IP-Adapter[27] used a general CLIP[16] image encoder to extract facial features, yet they struggled to maintain the identity fidelity. IP-Adapter-Plus[27] addressed this issue by utilizing patch features before the projection layer. InstantID[26] introduced the use of a pre-trained Arcface[2] encoder, significantly enhancing identity consistency and influencing subsequent works like Photomaker [9] and PULID [4]. A natural progression is to attack both the CLIP image encoder and Arcface encoder simultaneously, targeting all current Encoder-based ID preservation technologies." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 316, + 295, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 295, + 519 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 295, + 519 + ], + "type": "text", + "content": "There are four major challenges in ID protection: 1. Universality. Given the numerous existing encoder-based ID-preserving methods, we believe that universal protection is necessary, as attackers can easily switch to other methods to bypass protection. 2. Efficiency. Previous work typically relies on the Projected Gradient Descent (PGD) [12] to generate adversarial noise. However, PGD requires per-image optimization, taking several minutes per image, which significantly limits its practicality for protecting large volumes of portrait photos on social media in real world. 3. Robustness. The adversarial noise must be robust enough to withstand common image transformations, such as cropping, resizing, and JPEG compression, while also countering the inherent pre-processing operations of identity-preserving methods, such as face alignment. 4. Imperceptibility. The adversarial noise added for ID protection should be imperceptible to human eyes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "content": "In response to these challenges, this paper presents ID-Protector, the first feed-forward method to protect portrait photos from state-of-the-art encoder-based identity preservation techniques. Our approach incorporates several key innovations. For efficiency, we propose a ViT-based encoder that directly predicts adversarial noise for a given target portrait. To ensure universality, we carefully examine the attack surfaces and define a composite loss function tailored to four mainstream encoder-based generation methods, including InstantID [26], IP-Adapter [27], IP-Adapter Plus [27], and PhotoMaker [9]. Additionally, we employ an L1-based image regularization to minimize the visual impact of adversarial noise on image quality. For robustness against common transformations, a typical approach involves a combination of post-hoc differentiable augmentations like differentiable JPEG and cropping during train" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "type": "text", + "content": "ing. However, we find that this method introduces a substantial learning burden. To address this, we propose a surrogate method that leverages the face preprocessing pipeline of existing identity-preserving methods. Specifically, we introduce small random noise into the affine transformation matrix during face alignment stage to achieve data augmentation. This not only enhances the robustness of adversarial noise against common image transformations but also enables it to withstand image preprocessing operations during identity-preserving generation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 192, + 553, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 192, + 553, + 299 + ], + "spans": [ + { + "bbox": [ + 313, + 192, + 553, + 299 + ], + "type": "text", + "content": "We train and extensively evaluate IDProtector on the CelebA dataset [11]. We further test it on unseen data VGG Face [15] and proprietary models like Midjourney [13] and Jing Gou, demonstrating its strong generalization to unseen data and closed-source models. By advancing adversarial attack techniques from theoretical constructs to practically deployable solutions, IDProtector sets a new benchmark in facial privacy protection. We summarize our main contributions as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 300, + 553, + 408 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 314, + 300, + 553, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 300, + 553, + 335 + ], + "spans": [ + { + "bbox": [ + 314, + 300, + 553, + 335 + ], + "type": "text", + "content": "- We introduce IDProtector, the first feed-forward method to protect portrait photos from SOTA encoder-based identity preservation techniques." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 336, + 553, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 336, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 313, + 336, + 553, + 408 + ], + "type": "text", + "content": "- We present several innovative techniques including a ViT-based noise encoder, a joint learning framework with carefully designed objectives, and affine-based data augmentation to simultaneously achieve the EURI standard (Efficiency, Universality, Robustness, and Imperceptibility)." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 421, + 405, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 421, + 405, + 433 + ], + "spans": [ + { + "bbox": [ + 314, + 421, + 405, + 433 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 441, + 515, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 515, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 515, + 453 + ], + "type": "text", + "content": "2.1. Tuning-based Customized Generation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 459, + 553, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 459, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 459, + 553, + 590 + ], + "type": "text", + "content": "With the rapid advancement of diffusion models, customized generation methods have begun to emerge. Techniques like DreamBooth [17], Custom Diffusion [8], Low-Rank Adaptation (LoRA) [7] and Textual Inversion [3] facilitate the adaptation of text-to-image diffusion models through fine-tuning strategies that range from adjusting all parameters to optimizing specific word vectors. Despite their impressive effects, these methods require extensive computational resources and time due to their need for multi-image fine-tuning. This can make them impractical for real-time or large-scale applications." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 600, + 531, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 531, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 531, + 612 + ], + "type": "text", + "content": "2.2. Encoder-based ID-Preserving Generation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "type": "text", + "content": "Encoder-based Customized generation methods[23, 27, 29, 30] utilize pre-trained encoders to extract appearance feature, facilitating zero-shot generation in real-time. These methods evolved from domain-agnostic encoders to domain-specific encoders. Initially, CLIP[16] image encoder was used to extract facial features, such as with Face0[23] and IP-Adapter[27]; however, their maintenance of facial detail fidelity was not satisfactory. IP-Adapter" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3020" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "text", + "content": "Plus addressed this issue by utilizing patch features before the projection layer. InstantID[26] was the first to propose using a pre-trained Arcface[2] Encoder as a condition, significantly enhancing identity consistency and influencing subsequent works like Photomaker[9] and PULID[4]. StoryMaker[32] employs both the Arcface image encoder and the CLIP image encoder to encode facial features and broader regions, respectively, preserving not only facial identities but also clothing, hairstyles, and body features. Additionally, encoder-based ID preservation technology is widely applied in multi-human image personalization[5, 14] and video tasks[6]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 223, + 293, + 235 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 223, + 293, + 235 + ], + "spans": [ + { + "bbox": [ + 55, + 223, + 293, + 235 + ], + "type": "text", + "content": "2.3. Defensive Perturbations for Diffusion Models." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 240, + 296, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 240, + 296, + 515 + ], + "spans": [ + { + "bbox": [ + 55, + 240, + 296, + 515 + ], + "type": "text", + "content": "The focus on safeguarding personal artworks and facial images from unauthorized usage in the fine-tuning of diffusion models has led to the development of subtle yet effective perturbations. As summarized in Tab. 1, distinct approaches have been tailored to address this issue: Glaze [20] strategically adjusts the feature-space distances to misdirect the style learning process in Stable Diffusion, effectively preventing style appropriation. Conversely, AdvDM [10] implements adversarial disruptions by maximizing the Mean Squared Error during the optimization phase to shield personal visuals. Meanwhile, Anti-DreamBooth [24] integrates adversarial interference directly within the DreamBooth fine-tuning cycle, applying a complex optimization strategy to formulate defensive perturbations. Further investigations [21, 25, 28, 31] into adversarial protection have demonstrated their effectiveness in obstructing finetuning-based customization methods such as DreamBooth [17], LoRA [7], and textual inversion [3]. However, how to protect personal identity from the recently popular encoder-based methods remains unexplored. This paper endeavors to bridge the current gaps by proposing advanced methodologies to shield images from encoder-based exploitations, enhancing the security measures for personal digital assets." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 57, + 548, + 294, + 586 + ], + "blocks": [ + { + "bbox": [ + 55, + 525, + 295, + 546 + ], + "lines": [ + { + "bbox": [ + 55, + 525, + 295, + 546 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 295, + 546 + ], + "type": "text", + "content": "Table 1. Overview of encoder-based methods for ID-preserving generation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 548, + 294, + 586 + ], + "lines": [ + { + "bbox": [ + 57, + 548, + 294, + 586 + ], + "spans": [ + { + "bbox": [ + 57, + 548, + 294, + 586 + ], + "type": "table", + "html": "
EncoderID-Preserving Methods
CLIP [16]IP-Adapter [27], IP-Adapter Plus [27], PhotoMaker [9], Face0 [23]
ArcFace [2]InstantID [26], PhotoMaker [9], PuLID [4], IP-Adapter-FaceID [27]
CLIP & ArcFaceIDAdapter [1], Storymaker [32]
", + "image_path": "d95480a08f9e1977e0e7c4f33c2b3dbc14564883e5e299e0fa20a84b3227fb65.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 609, + 168, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 168, + 621 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 168, + 621 + ], + "type": "text", + "content": "3. Problem Definition" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 629, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 294, + 713 + ], + "type": "text", + "content": "We consider two participants in a portrait identity protection scenario.: \"image protector\" Alice and \"image exploiter\" Bob. Bob seeks to use an encoder-based, identity-preserving approach to alter the portrait photos Alice shares on social media for malicious purposes. While Alice aims to protect her portrait by adding adversarial noise to thwart Bob's methods, inducing distortions or significant decreases" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "content": "in ID similarity in the generated content. Specifically, we explain the workflow of the two parties as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 96, + 553, + 298 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 313, + 96, + 553, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 96, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 96, + 553, + 239 + ], + "type": "text", + "content": "1. Image Protector Alice: The image protector aims to provide protection for images against exploitation by encoder-based ID-preserving methods. In this context, the chosen protection method involves adding imperceptible protective perturbations to the images, with the goal of providing protection while minimizing modifications to the original image. In real-world scenarios, image protectors encounter several challenges. For instance, implementing universal protection to guard against various encoder-based, identity-preserving methods is difficult. Additionally, the protection should be robust to natural distortions such as resizing, JPEG compression, and blurring." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 239, + 553, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 553, + 298 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 553, + 298 + ], + "type": "text", + "content": "2. Image Exploiter Bob: Bob aims to use portrait photos for ID-preserving customization. Bob can choose any open-source or proprietary encoder-based models, which poses significant challenges on the universality of the protection method." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 300, + 555, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 555, + 419 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 555, + 419 + ], + "type": "text", + "content": "Objective of the Study: Given an image " + }, + { + "bbox": [ + 313, + 300, + 555, + 419 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 300, + 555, + 419 + ], + "type": "text", + "content": " containing a face portrait, our objective is to generate adversarial perturbations " + }, + { + "bbox": [ + 313, + 300, + 555, + 419 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 300, + 555, + 419 + ], + "type": "text", + "content": " that protect against ID-preserving generation. We aim to maximize the difference in identity consistency between the unperturbed generated image " + }, + { + "bbox": [ + 313, + 300, + 555, + 419 + ], + "type": "inline_equation", + "content": "g(x)" + }, + { + "bbox": [ + 313, + 300, + 555, + 419 + ], + "type": "text", + "content": " and the perturbed generated image " + }, + { + "bbox": [ + 313, + 300, + 555, + 419 + ], + "type": "inline_equation", + "content": "g(x + \\delta)" + }, + { + "bbox": [ + 313, + 300, + 555, + 419 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 300, + 555, + 419 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 313, + 300, + 555, + 419 + ], + "type": "text", + "content": " is an ID-preserving generator. To unify the protection against different models, we can find a " + }, + { + "bbox": [ + 313, + 300, + 555, + 419 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 300, + 555, + 419 + ], + "type": "text", + "content": " that simultaneously disrupts several feature spaces, lowering the sum of the similarity values. Thus, we train a noise encoder network " + }, + { + "bbox": [ + 313, + 300, + 555, + 419 + ], + "type": "inline_equation", + "content": "E_{\\theta}" + }, + { + "bbox": [ + 313, + 300, + 555, + 419 + ], + "type": "text", + "content": ", solving" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 351, + 428, + 553, + 445 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 428, + 553, + 445 + ], + "spans": [ + { + "bbox": [ + 351, + 428, + 553, + 445 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} S \\left\\{g _ {i} (x), g _ {i} (x + \\delta) \\right\\}, \\quad \\delta = E _ {\\theta} (x) \\tag {1}", + "image_path": "b353daf3352dc013ec2ad9b8a04e439fb6541f1e6a0e613077b61bfec25caeba.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 454, + 553, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 454, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 553, + 491 + ], + "type": "text", + "content": "subject to the constraint that " + }, + { + "bbox": [ + 313, + 454, + 553, + 491 + ], + "type": "inline_equation", + "content": "\\| \\delta \\|_{\\infty} < \\epsilon" + }, + { + "bbox": [ + 313, + 454, + 553, + 491 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 313, + 454, + 553, + 491 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 454, + 553, + 491 + ], + "type": "text", + "content": " denotes an identity similarity metric, which we evaluate using cosine similarity in feature spaces of ArcFace [2] and CLIP [16]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 501, + 370, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 501, + 370, + 513 + ], + "spans": [ + { + "bbox": [ + 313, + 501, + 370, + 513 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 521, + 553, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 553, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 553, + 605 + ], + "type": "text", + "content": "The core of our method is an adversarial noise encoder model that can add adversarial noise to facial images for protection. We will first introduce the complete workflow in Sec. 4.1, followed by the design of the adversarial noise encoder in Sec. 4.2. Sec. 4.3 and 4.4 introduce ID protection objectives and loss functions. Sec. 4.5 introduces the data augmentation approach to enhance the robustness." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 612, + 411, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 612, + 411, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 612, + 411, + 624 + ], + "type": "text", + "content": "4.1. Overall Method" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "text", + "content": "Fig. 2 details the pipeline of our method. For an arbitrary-sized input image " + }, + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "text", + "content": " containing a face, it is resized to " + }, + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "text", + "content": " and sent into the adversarial noise encoder model, whose output is an additive adversarial noise that can be resized back and perturbed onto the original image for protection. During adversarial noise encoder training, this protected image is further passed through face feature extrac" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "3021" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 73, + 549, + 225 + ], + "blocks": [ + { + "bbox": [ + 62, + 73, + 549, + 225 + ], + "lines": [ + { + "bbox": [ + 62, + 73, + 549, + 225 + ], + "spans": [ + { + "bbox": [ + 62, + 73, + 549, + 225 + ], + "type": "image", + "image_path": "ee7bfd173de19bfe5dea26574c7db9c30d07cbe1054171e65fd06d3498026cb6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 228, + 555, + 285 + ], + "lines": [ + { + "bbox": [ + 54, + 228, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 54, + 228, + 555, + 285 + ], + "type": "text", + "content": "Figure 2. Overall schematics of our method. Our method's key design includes the noise encoder, loss functions, and the gradient optimization path that allows backpropagation. During the protection phase, an image is first resized to " + }, + { + "bbox": [ + 54, + 228, + 555, + 285 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 54, + 228, + 555, + 285 + ], + "type": "text", + "content": " and sent to an adversarial noise encoder model. The model outputs a perturbation that can be resized back and added to the image, achieving an adversarial attack effect – the protected image prevents subsequent pipelines from correctly extracting facial feature information, thereby hindering face similarity in face generation tasks. This achieves the goal of ID protection." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 305, + 295, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 305, + 295, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 305, + 295, + 331 + ], + "type": "text", + "content": "tion pipelines to obtain the loss functions that aim at creating maximum differences in the feature space." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 336, + 146, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 336, + 146, + 348 + ], + "spans": [ + { + "bbox": [ + 55, + 336, + 146, + 348 + ], + "type": "text", + "content": "4.2. Noise Encoder" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 354, + 296, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 296, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 296, + 534 + ], + "type": "text", + "content": "The network takes a " + }, + { + "bbox": [ + 55, + 354, + 296, + 534 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 55, + 354, + 296, + 534 + ], + "type": "text", + "content": " RGB image as input and outputs three-channel adversarial perturbations. The perturbations are projected to a range of " + }, + { + "bbox": [ + 55, + 354, + 296, + 534 + ], + "type": "inline_equation", + "content": "[-1, 1]" + }, + { + "bbox": [ + 55, + 354, + 296, + 534 + ], + "type": "text", + "content": ", then denormalized " + }, + { + "bbox": [ + 55, + 354, + 296, + 534 + ], + "type": "inline_equation", + "content": "[- \\epsilon, \\epsilon]" + }, + { + "bbox": [ + 55, + 354, + 296, + 534 + ], + "type": "text", + "content": ", resized to the original image dimensions, and added to the image. We utilize a Vision Transformer (ViT) to generate the adversarial noise, with both input and output dimensions set to " + }, + { + "bbox": [ + 55, + 354, + 296, + 534 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 55, + 354, + 296, + 534 + ], + "type": "text", + "content": ". Empirically, we find that adding an additional prior mask channel, indicating the location of the face, aids in training. This face localization mask is generated by the InsightFace pipeline based on facial landmark points, which specify the region of the face within the image. The mask is concatenated to the image as the fourth input channel, eliminating the need for the network to learn face localization ability from scratch, thus reducing training difficulty." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 540, + 200, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 200, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 200, + 552 + ], + "type": "text", + "content": "4.3. Adversarial ID Protection" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 557, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 295, + 653 + ], + "type": "text", + "content": "Since generative networks' only knowledge of faces comes from feature embeddings, to achieve ID protection, we can make the extracted features from protected images significantly different from the originals, thereby preventing generation models from accessing correct facial features. Based on this principle, we analyze the specific pipelines of InstantID, IP-Adapter, IP-Adapter-Plus, and PhotoMaker to design targeted attack objectives and loss functions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "InstantID obtains facial features in two steps, as shown in Fig. 2. It first aligns the face to a pre-defined location, before feeding the aligned face into an ArcFace feature extractor to obtain facial information. We choose to fool the ArcFace model, minimizing the cosine similarity between" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 305, + 547, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 305, + 547, + 317 + ], + "spans": [ + { + "bbox": [ + 313, + 305, + 547, + 317 + ], + "type": "text", + "content": "the ArcFace features of the original v.s. protected images." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 318, + 553, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 318, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 318, + 553, + 510 + ], + "type": "text", + "content": "IP-Adapter, IP-Adapter-Plus, and PhotoMaker all rely on facial features extracted by the CLIP vision encoder, albeit using different versions of CLIP vision. IP-Adapter and PhotoMaker utilize CLIP Vision's output, while IP-Adapter-Plus takes features before CLIP Vision's second-to-last layer. As illustrated in Fig. 2, embeddings from different layers offer various attack surfaces. We select victim embeddings for attack according to three principles: (1) blocking all potential pathways, ensuring that any information flow from left to right passes through at least one victim embedding, ensuring all information flow would be disrupted; (2) choosing features as early as possible in the network (towards the left in Fig. 2), to reduce the backpropagation path length and thus simplify optimization; and (3) targeting embeddings with dense semantic information for more effective manipulation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 511, + 555, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 511, + 555, + 643 + ], + "spans": [ + { + "bbox": [ + 313, + 511, + 555, + 643 + ], + "type": "text", + "content": "An example of an unsuitable target is the last hidden layer feature of CLIP Vision. While modifying this feature could influence all three models, its huge dimension " + }, + { + "bbox": [ + 313, + 511, + 555, + 643 + ], + "type": "inline_equation", + "content": "(257\\times 1280)" + }, + { + "bbox": [ + 313, + 511, + 555, + 643 + ], + "type": "text", + "content": " and lack of semantic density make it highly challenging for optimization to converge. Ultimately, we selected the embeddings marked red in Fig. 2 as the primary attack targets. The objective, similar to that of InstantID, is to maximize the cosine similarity between the perturbed embedding and the original embedding, effectively aligning them post-attack. Therefore, the final adversarial loss is a weighted average of all losses, given by" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 355, + 653, + 553, + 680 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 653, + 553, + 680 + ], + "spans": [ + { + "bbox": [ + 355, + 653, + 553, + 680 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {a d v}} = \\sum_ {\\text {C u t s o m M o d e l} i} \\alpha_ {i} \\cdot \\operatorname {c o s s i m} \\left(e _ {i}, e _ {i} ^ {\\prime}\\right), \\tag {2}", + "image_path": "b6a9ac478e34d9c11fa48c0a059bf839c268fa3792a9cc98720de5c59e50d5ea.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "inline_equation", + "content": "e_i" + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "inline_equation", + "content": "e_i'" + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": " denote the face embeddings of the clean and protected images, respectively." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3022" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 153, + 85 + ], + "type": "text", + "content": "4.4. Imperceptibility" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 296, + 150 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 296, + 150 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 296, + 150 + ], + "type": "text", + "content": "To minimize the visual impact on the image quality, we enforce an " + }, + { + "bbox": [ + 55, + 89, + 296, + 150 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 55, + 89, + 296, + 150 + ], + "type": "text", + "content": " regularization on the predicted adversarial noise " + }, + { + "bbox": [ + 55, + 89, + 296, + 150 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 89, + 296, + 150 + ], + "type": "text", + "content": ". We further introduce an auxiliary penalty on any noise values exceeding the " + }, + { + "bbox": [ + 55, + 89, + 296, + 150 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 89, + 296, + 150 + ], + "type": "text", + "content": "-ball boundary. Together, these terms constitute our regularization loss, given by:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 89, + 159, + 295, + 174 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 159, + 295, + 174 + ], + "spans": [ + { + "bbox": [ + 89, + 159, + 295, + 174 + ], + "type": "interline_equation", + "content": "L _ {\\text {r e g}} = \\beta_ {1} \\cdot \\| \\delta \\| _ {1} + \\beta_ {2} \\cdot \\| \\delta - \\operatorname {c l i p} _ {\\pm \\epsilon} (\\delta) \\| _ {1}. \\tag {3}", + "image_path": "0ec9dffb0dd36e2c00b17faa48977c9a0f90a233b467dcf3524c332c9b6341ec.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 182, + 295, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 182, + 295, + 207 + ], + "spans": [ + { + "bbox": [ + 55, + 182, + 295, + 207 + ], + "type": "text", + "content": "The final loss is thus the sum of adversarial loss and regularization, given by:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 139, + 217, + 295, + 232 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 217, + 295, + 232 + ], + "spans": [ + { + "bbox": [ + 139, + 217, + 295, + 232 + ], + "type": "interline_equation", + "content": "L = L _ {\\mathrm {a d v}} + L _ {\\mathrm {r e g}}. \\tag {4}", + "image_path": "2d1619516fe1cbc4eed60d8d040f54be729265a8614d4c5823375e0895688811.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 240, + 132, + 252 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 240, + 132, + 252 + ], + "spans": [ + { + "bbox": [ + 55, + 240, + 132, + 252 + ], + "type": "text", + "content": "4.5. Robustness" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 258, + 296, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 296, + 402 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 296, + 402 + ], + "type": "text", + "content": "As shown in Fig. 2, the InstantID pipeline involves a step of facial alignment aiming to align any faces to a predefined standard face for better feature extraction. The alignment process includes detecting the face and key points " + }, + { + "bbox": [ + 55, + 258, + 296, + 402 + ], + "type": "inline_equation", + "content": "\\tilde{p}" + }, + { + "bbox": [ + 55, + 258, + 296, + 402 + ], + "type": "text", + "content": ", calculating the affine matrix " + }, + { + "bbox": [ + 55, + 258, + 296, + 402 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 55, + 258, + 296, + 402 + ], + "type": "text", + "content": " with standard face key points " + }, + { + "bbox": [ + 55, + 258, + 296, + 402 + ], + "type": "inline_equation", + "content": "\\tilde{p}'" + }, + { + "bbox": [ + 55, + 258, + 296, + 402 + ], + "type": "text", + "content": ", and aligning them. This process is computationally expensive. To reduce computational cost and simplify training, we pre-compute " + }, + { + "bbox": [ + 55, + 258, + 296, + 402 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 55, + 258, + 296, + 402 + ], + "type": "text", + "content": " for each image and keep it fixed throughout training. However, during inference, perturbations carried by protected images may slightly alter the precomputed affine matrix " + }, + { + "bbox": [ + 55, + 258, + 296, + 402 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 55, + 258, + 296, + 402 + ], + "type": "text", + "content": ", leading to minor misalignment that weakens the effectiveness the adversarial perturbations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 403, + 296, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 296, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 296, + 475 + ], + "type": "text", + "content": "To adapt to this misalignment as well as being computationally efficient, we leverage the existing affine transformation operation and introduce small gaussian noise to " + }, + { + "bbox": [ + 55, + 403, + 296, + 475 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 55, + 403, + 296, + 475 + ], + "type": "text", + "content": " during training to simulate face transformations. The noise level is empirically set to " + }, + { + "bbox": [ + 55, + 403, + 296, + 475 + ], + "type": "inline_equation", + "content": "\\sigma = 0.003" + }, + { + "bbox": [ + 55, + 403, + 296, + 475 + ], + "type": "text", + "content": ". The face alignment transformation is therefore given by" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 125, + 484, + 295, + 498 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 484, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 125, + 484, + 295, + 498 + ], + "type": "interline_equation", + "content": "\\tilde {p} ^ {\\prime} = (A + N (0, \\sigma I)) \\tilde {p}. \\tag {5}", + "image_path": "425e207ad15f8ac10bda08393ee9c12406c833c6caf8d0d8ea3da9c50bf2e993.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 508, + 296, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 508, + 296, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 508, + 296, + 556 + ], + "type": "text", + "content": "We incorporate this augmentation only in the InstantID branch. Our experiments reveal that it also enhances the robustness of other branches, being resilient to various common image transformations." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 567, + 137, + 581 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 567, + 137, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 137, + 581 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 587, + 172, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 587, + 172, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 587, + 172, + 601 + ], + "type": "text", + "content": "5.1. Experimental Setup" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "type": "text", + "content": "Training Data. For practicality, we selected 190,000 uncropped images from the CelebA dataset [11] as the training set. These images have arbitrary sizes and aspect ratios. The test set includes two subsets: 50 unseen images from CelebA and 50 images from VGG Face [15]. To ensure compatibility with baseline methods, these images were resized to " + }, + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Training Details. We trained a ViT-S/8 adversarial noise encoder from scratch using Eq. 4 as the loss function. We" + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 316, + 105, + 553, + 150 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 555, + 103 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 555, + 103 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 555, + 103 + ], + "type": "text", + "content": "Table 2. Curriculum learning stages with corresponding hyperparameters. " + }, + { + "bbox": [ + 313, + 71, + 555, + 103 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 71, + 555, + 103 + ], + "type": "text", + "content": " represents learning rate. " + }, + { + "bbox": [ + 313, + 71, + 555, + 103 + ], + "type": "inline_equation", + "content": "{\\alpha }_{1}" + }, + { + "bbox": [ + 313, + 71, + 555, + 103 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 71, + 555, + 103 + ], + "type": "inline_equation", + "content": "{\\alpha }_{4}" + }, + { + "bbox": [ + 313, + 71, + 555, + 103 + ], + "type": "text", + "content": " are weights for IP-Adapter,IP-Adapter-Plus,PhotoMaker,InstantID,respectively." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 105, + 553, + 150 + ], + "lines": [ + { + "bbox": [ + 316, + 105, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 105, + 553, + 150 + ], + "type": "table", + "html": "
StageEpochsα1α2α3α4β1β2εη
11202/3714/3720/371/379e-31e-30.051e-2
2204/162/169/162/161.8e-32e-40.042e-5
3202/142/149/141/144.5e-45e-50.0352e-5
", + "image_path": "8f707b0926e987183a8e611484f9870d285499e382d7547c32998531c62ff952.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 171, + 555, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 171, + 555, + 278 + ], + "spans": [ + { + "bbox": [ + 313, + 171, + 555, + 278 + ], + "type": "text", + "content": "employed a three-stage curriculum learning approach, progressively shrinking the " + }, + { + "bbox": [ + 313, + 171, + 555, + 278 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 171, + 555, + 278 + ], + "type": "text", + "content": "-ball and balancing the different losses by adjusting the hyperparameters (presented in Table 2). Training began with a 2,500-step linear warm-up, gradually increasing the learning rate from zero. Gradient clipping was applied during training to cap the " + }, + { + "bbox": [ + 313, + 171, + 555, + 278 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 313, + 171, + 555, + 278 + ], + "type": "text", + "content": " norm of gradients at a maximum of 10. The model was trained with a batch size of 112 on four 80GB NVIDIA H100 GPUs for 10 days." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 279, + 556, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 279, + 556, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 279, + 556, + 411 + ], + "type": "text", + "content": "Evaluation process and metrics. To evaluate the protection performance of different methods, for each face-containing image in the dataset (denoted as NP ref image), we apply different protection methods (including our ID-Protector) to generate its protected counterpart (denoted as P ref image). Throughout all experiments, we constrain the maximum pixel perturbation to 9/255 (an " + }, + { + "bbox": [ + 313, + 279, + 556, + 411 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 279, + 556, + 411 + ], + "type": "text", + "content": "-ball of 0.035), following common practice in adversarial machine learning [18]. Subsequently, both P and NP ref images are fed into customization generation models (such as InstantID) to obtain generated images (denoted as P/NP gen images)." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 411, + 556, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 556, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 556, + 495 + ], + "type": "text", + "content": "We evaluate the identify shift caused by different protection methods through Identity Score Matching (ISM), which is calculated as the cosine similarity between facial features of NP images and generated images extracted by ArcFace [2]. Lower ISM values indicate better protection as they suggest a greater identity shift between original (NP ref) and generated (P gen) faces." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 495, + 556, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 556, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 556, + 568 + ], + "type": "text", + "content": "Since protected images may result in no detectable faces being generated, we also report the Face Detection Rate (FDR) using InsightFace detector, as ISM can only be computed when faces are detected in both images. Additionally, to ensure comprehensive evaluation, we measure the perceptual quality of each generated face using SER-FIQ [22]." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 576, + 465, + 589 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 576, + 465, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 576, + 465, + 589 + ], + "type": "text", + "content": "5.2. Comparison with Baselines" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "text", + "content": "Protection Performance. We compare our method with the following baseline methods: ACE [31], AdvDM [10], Anti-DreamBooth [24], PhotoGuard [19], and SimAC [25]. Most baselines aim to reduce the quality of generated images or hinder the diffusion denoising process, but we aim to reduce the similarity of generated identities so as to directly disrupt identity-preserving generation. To compare our method with baselines, we first applied each protection method to generate protected versions of test images. We then used both the original images (NP) and the protected" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3023" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 71, + 550, + 473 + ], + "blocks": [ + { + "bbox": [ + 60, + 71, + 550, + 473 + ], + "lines": [ + { + "bbox": [ + 60, + 71, + 550, + 473 + ], + "spans": [ + { + "bbox": [ + 60, + 71, + 550, + 473 + ], + "type": "image", + "image_path": "425ffa070c84b8fa4527206248c51af6132b7951f62e8a74bb4a3af856ef2bd6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 475, + 555, + 499 + ], + "lines": [ + { + "bbox": [ + 55, + 475, + 555, + 499 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 555, + 499 + ], + "type": "text", + "content": "Figure 3. Qualitative comparison with baseline. Images protected by our method cannot be used to generate similar faces. Compared to baseline protection methods, under the same perturbation bound of " + }, + { + "bbox": [ + 55, + 475, + 555, + 499 + ], + "type": "inline_equation", + "content": "9/255" + }, + { + "bbox": [ + 55, + 475, + 555, + 499 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 55, + 475, + 555, + 499 + ], + "type": "inline_equation", + "content": "\\epsilon = 3.5\\%" + }, + { + "bbox": [ + 55, + 475, + 555, + 499 + ], + "type": "text", + "content": "), ours causes more significant facial differences." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 519, + 298, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 519, + 298, + 712 + ], + "spans": [ + { + "bbox": [ + 54, + 519, + 298, + 712 + ], + "type": "text", + "content": "versions as references in customized generation tasks with four models: IP-Adapter, IP-Adapter-Plus, PhotoMaker, and InstantID. Tab. 3 shows that, for the first three methods, while baseline methods reduced ISM by an average of 0.02, with a maximum reduction of 0.06, our method achieved an ISM reduction between 0.1 and 0.14, demonstrating a larger identity shift. On InstantID, our IDProtector demonstrated even more pronounced protection: when baseline methods reduced facial similarity by less than 0.1, our method reduced similarity by over 0.4 across both datasets. Additionally, FDR results validate the reliability of ISM as FDR reflects the percentage of images with a face detected and contributes to ISM calculations. Overall, these results underscore the superior identity protection performance of our method compared to existing approaches. Together, these highlight the superior identity protection performance of" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 520, + 372, + 531 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 520, + 372, + 531 + ], + "spans": [ + { + "bbox": [ + 313, + 520, + 372, + 531 + ], + "type": "text", + "content": "our approach." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "type": "text", + "content": "Following previous study [24], we evaluate SER-FIQ as a quality metric of generated (P/NP gen) images. Although our method has a noticeable impact on image generation quality (as shown in Fig. 3), we observed inconsistencies between the SER-FIQ scores and qualitative assessments. This discrepancy suggests that a more robust and reliable quality metric may be required to accurately measure the quality of generated images in future research." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "We also adapted our method to a per-image tuning variant, employing PGD [12] as the optimization technique. We observed that PGD achieves a more substantial ISM reduction across most methods, which aligns with our expectations since per-image tuning can better leverage the specific characteristics of each individual image. However, a notable drawback of PGD is the prolonged optimization time," + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3024" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 104, + 555, + 285 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "type": "text", + "content": "Table 3. Comparison of identity protection performance against baseline methods. Protection performance is evaluated against IP-Adapter, IP-Adapter-Plus, PhotoMaker, and InstantID. \"No Protect\" denotes direct use of the original (NP ref) images for identity-preserving generation. Experiments were conducted on five prompts per dataset, and average results are reported." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 104, + 555, + 285 + ], + "lines": [ + { + "bbox": [ + 56, + 104, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 104, + 555, + 285 + ], + "type": "table", + "html": "
DatasetProtection MethodIP-AdapterIP-Adapter-PlusPhotoMakerInstantID
FDR ↓ISM ↓SER-FIQ ↓FDR ↓ISM ↓SER-FIQ ↓FDR ↓ISM ↓SER-FIQ ↓FDR ↓ISM ↓SER-FIQ ↓
CelebANo Protect1.0000.1760.6411.0000.2400.6630.6640.2650.6050.9960.7750.774
ACE1.0000.1590.6401.0000.2310.6810.6360.2680.5941.0000.7450.758
AdvDM0.9880.1570.6301.0000.2260.6660.6400.2550.5901.0000.7570.757
Anti-DB1.0000.1550.6151.0000.2220.6760.6400.2550.5991.0000.7530.752
PhotoGuard1.0000.1520.6211.0000.2230.6700.6520.2590.5921.0000.7580.767
SimAC1.0000.1620.6271.0000.2170.6830.6520.2530.5991.0000.7520.755
Ours (PGD)0.8440.0240.5440.9120.0720.7180.1120.0940.6390.996-0.1660.721
Ours0.9520.0600.6880.9960.1140.6440.6720.1230.6221.0000.2310.613
VGG FaceNo Protect1.0000.2010.6721.0000.2360.6980.6320.2540.6000.9960.7640.828
ACE0.9880.1540.6171.0000.2090.6831.0000.2090.6831.0000.7030.829
AdvDM1.0000.1470.5971.0000.2150.6931.0000.2150.6931.0000.7060.820
Anti-DB1.0000.1380.5971.0000.2100.6780.6240.2380.5991.0000.7070.811
PhotoGuard1.0000.1590.6431.0000.2170.6860.6240.2290.6120.9960.6910.799
SimAC1.0000.1500.6051.0000.2020.6770.6240.2280.6441.0000.7070.826
Ours (PGD)0.9680.0430.6091.0000.2020.7400.6240.2280.6301.0000.7070.769
Ours1.0000.1050.7371.0000.1380.6870.6320.1260.6141.0000.2320.713
", + "image_path": "25ccbf6300aa7b4ad2eb081878c462bb049d1bc1ab83d878ef1314cf1789c23e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 56, + 335, + 296, + 392 + ], + "blocks": [ + { + "bbox": [ + 55, + 302, + 295, + 335 + ], + "lines": [ + { + "bbox": [ + 55, + 302, + 295, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 302, + 295, + 335 + ], + "type": "text", + "content": "Table 4. Comparison of time cost and quality impact with an epsilon-budget of 9/255 (±3.5%). PSNR and SSIM are calculated between clean (NP ref) and perturbed (P ref) image pairs." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 335, + 296, + 392 + ], + "lines": [ + { + "bbox": [ + 56, + 335, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 335, + 296, + 392 + ], + "type": "table", + "html": "
ACEAdvDMAnti-DBPhoto-GuardSimACOurs PGDOurs
Time (s)87.051.926123.332340890.173
PSNR (dB)31.0631.7031.7031.1032.3732.9432.15
SSIM (dB)0.7410.7660.7650.7570.7910.8450.809
", + "image_path": "4728a2d2a721ad63d3a1496b09cd8ed8a7ebab09c6ffc07b68a9877e083ed7cb.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 413, + 247, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 413, + 247, + 425 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 247, + 425 + ], + "type": "text", + "content": "which we will address in the following section." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 426, + 295, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 426, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 54, + 426, + 295, + 569 + ], + "type": "text", + "content": "The protection effects are visualized in Fig. 3. The baseline methods can protect against some of the generation methods. For example, AdvDM and Anti-DreamBooth cause IP-Adapter to produce faces with color block artifacts, while several outputs from IP-Adapter-Plus exhibit unnatural patterns and shadows. However, the baselines are ineffective in protecting against PhotoMaker and InstantID. With IDProtector's protection, faces generated by all four methods appear significantly altered. In some cases, even the gender of the person is unidentifiable. This fulfills our protection goal: protected images could not be used for identity-preserving generation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 570, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 570, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 570, + 295, + 713 + ], + "type": "text", + "content": "Time and Image Quality. Our IDProtector is fast to protect an image with less degradation to image quality. As compared in Table 4, the PSNR and SSIM of our method are higher than the baselines, indicating less visual impact. The baseline methods require an average of 23.3 to 323 seconds per image for protection. Using PGD with our optimization objective would take over an hour on average. In contrast, our ViT-based IDProtector achieves an average protection time of only 0.173 seconds per image – less than " + }, + { + "bbox": [ + 54, + 570, + 295, + 713 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 54, + 570, + 295, + 713 + ], + "type": "text", + "content": " of the fastest baseline. Note that the forward-feeding ViT supports better parallelization compared with PGD, the time cost could further decrease with mini-batch inference." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 315, + 346, + 554, + 411 + ], + "blocks": [ + { + "bbox": [ + 313, + 302, + 555, + 346 + ], + "lines": [ + { + "bbox": [ + 313, + 302, + 555, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 302, + 555, + 346 + ], + "type": "text", + "content": "Table 5. Generalization of ID protection by our IDProtector on unseen generators. IPA-FID+/P/XL denotes IP-Adapter-FaceID/Plus/Portrait/SDXL, with NP and P indicating no and with protection, respectively." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 346, + 554, + 411 + ], + "lines": [ + { + "bbox": [ + 315, + 346, + 554, + 411 + ], + "spans": [ + { + "bbox": [ + 315, + 346, + 554, + 411 + ], + "type": "table", + "html": "
IPA-FIDIPA-FID+IPA-FIDPIPA-FIDXLFlux-IPAStory-MakerMid-journeyJingGou
ISM ↓NP0.5650.6440.1290.5230.0540.5720.1990.672
P0.1530.2170.0520.1320.0290.1930.0980.131
SER-NP0.7060.7510.7200.4020.7280.6990.5680.791
FIQ ↓P0.8030.7300.6580.6670.6730.7340.6270.805
", + "image_path": "6b7fa9b246726fc0429b01ae5ba2881438458361e14219e35618ddd5dbcc65f0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 426, + 552, + 587 + ], + "blocks": [ + { + "bbox": [ + 316, + 426, + 552, + 587 + ], + "lines": [ + { + "bbox": [ + 316, + 426, + 552, + 587 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 552, + 587 + ], + "type": "image", + "image_path": "8d3a17b59eb2baa69c8daf3c6d0e744ae0b33a82b77c500f6b1ee1f4b3685d83.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 588, + 541, + 599 + ], + "lines": [ + { + "bbox": [ + 326, + 588, + 541, + 599 + ], + "spans": [ + { + "bbox": [ + 326, + 588, + 541, + 599 + ], + "type": "text", + "content": "Figure 4. ID protection performance on unseen generators." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 623, + 538, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 623, + 538, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 623, + 538, + 635 + ], + "type": "text", + "content": "5.3. Generalization to Unseen Data and Models" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": "Generalization to unseen datasets. We evaluated our method on the VGG Face dataset that was unseen during training. All following experiments are conducted on the CelebA test set unless otherwise specified. As shown in Tab. 3, the ISM reduction achieved by our method shows minimal variation across datasets. For example, our IDPro" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3025" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 94, + 555, + 196 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 93 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 93 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 93 + ], + "type": "text", + "content": "Table 6. Robustness evaluation of IDProtector under image distortions. NP denotes non-protected images for identity-preserving generation, while P represents protected but non-distorted images, providing an ablation. Distortion averages exclude NP and P." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 94, + 555, + 196 + ], + "lines": [ + { + "bbox": [ + 56, + 94, + 555, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 94, + 555, + 196 + ], + "type": "table", + "html": "
IDProtector (ours, ViT)Adaptive PGD
IP-AdapterIP-Adapter+PhotoMakerInstantIDIP-AdapterIP-Adapter+PhotoMakerInstantID
FDR ↓ISM ↓FDR ↓ISM ↓FDR ↓ISM ↓FDR ↓ISM ↓FDR ↓ISM ↓FDR ↓ISM ↓FDR ↓ISM ↓FDR ↓ISM ↓
NP1.0000.1761.0000.2400.6640.2650.9960.7750.8440.0240.9120.0720.1120.0940.996-0.166
P0.9520.0600.9960.1140.6720.1231.0000.2310.8440.0240.9120.0720.1120.0940.996-0.166
P+Affine1.0000.0681.0000.1060.6560.1351.0000.2160.8440.0640.9760.0810.112-0.0091.000-0.147
P+JPEG0.9640.0691.0000.0970.6600.1381.0000.2410.7840.0150.9480.0620.108-0.0121.000-0.161
P+Crop1.0000.0721.0000.1080.6520.1321.0000.2160.7520.0580.9400.0700.1080.0201.000-0.167
P+Noisy0.9520.0690.9840.1140.6560.1321.0000.2520.8320.0230.9320.0720.112-0.0101.000-0.130
Avg (P+)0.9790.0700.9960.1060.6560.1341.0000.2310.8030.0400.9490.0710.110-0.0031.000-0.151
", + "image_path": "594b3bab479fbd02006c074c73f6d4e09375ed1ccffd0f8c7200886c151bd8ff.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 58, + 217, + 293, + 410 + ], + "blocks": [ + { + "bbox": [ + 58, + 217, + 293, + 410 + ], + "lines": [ + { + "bbox": [ + 58, + 217, + 293, + 410 + ], + "spans": [ + { + "bbox": [ + 58, + 217, + 293, + 410 + ], + "type": "image", + "image_path": "8aa4b299e8c0047648a0d6a052476a39670289cf352eebd33d7d2302bbdf0968.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 411, + 295, + 455 + ], + "lines": [ + { + "bbox": [ + 55, + 411, + 295, + 455 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 295, + 455 + ], + "type": "text", + "content": "Figure 5. Robustness evaluation of IDProtector under various distortions. Despite geometric distortions like affine transformations, IDProtector effectively misleads all four models into generating dissimilar faces or even artifacts." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 481, + 296, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 481, + 296, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 481, + 296, + 540 + ], + "type": "text", + "content": "tector reduces the ISM of IP-Adapter by approximately 0.1 and InstantID by around 0.55 on both datasets. This result suggests that our method generalises effectively beyond the CelebA dataset, demonstrating protection capabilities that are not restricted to in-domain faces." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 544, + 295, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 544, + 295, + 662 + ], + "spans": [ + { + "bbox": [ + 54, + 544, + 295, + 662 + ], + "type": "text", + "content": "Generalization to unseen customization models. Beyond the four customization models seen during training, we tested protected images on a wider range of customization models. Across these models, our protection approach achieves an average ISM reduction of 0.31, with a maximum reduction of 0.54. Notably, our method remains effective against proprietary commercial models such as Midjourney (0.10 reduction) and Jing Gou (0.54 reduction), underscoring its ability to generalise to unseen customization models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "These protection effects are visualised in Fig. 4. As shown, protected images used in these generators cause noticeable quality degradation, including altered facial shapes (e.g., IP-Adapter-FaceID-XL), reduced prompt-following" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 216, + 555, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 216, + 555, + 252 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 555, + 252 + ], + "type": "text", + "content": "ability, and image contamination (e.g., Midjourney). The common result is a decreased resemblance between the generated face and the reference identity." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 258, + 443, + 270 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 443, + 270 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 443, + 270 + ], + "type": "text", + "content": "5.4. Robustness Evaluation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 276, + 555, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 276, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 276, + 555, + 456 + ], + "type": "text", + "content": "Our method demonstrates robustness against various distortions. For protected (P) images, we applied several distortions, including JPEG compression (quality=85), random cropping that removes up to " + }, + { + "bbox": [ + 313, + 276, + 555, + 456 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 313, + 276, + 555, + 456 + ], + "type": "text", + "content": " of the region, noise perturbation with a variance of 100, and an affine transformation. The affine transformation was implemented as " + }, + { + "bbox": [ + 313, + 276, + 555, + 456 + ], + "type": "inline_equation", + "content": "\\tilde{q} = (A + \\mathcal{N}(0,0.05^2))\\tilde{p}" + }, + { + "bbox": [ + 313, + 276, + 555, + 456 + ], + "type": "text", + "content": ", where the matrix " + }, + { + "bbox": [ + 313, + 276, + 555, + 456 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 276, + 555, + 456 + ], + "type": "text", + "content": " is an identity matrix. As shown in Tab. 6, the ISM values under these distortions remain comparable to those obtained without distortion, indicating that our protection remains effective under distortion. Fig. 5 provides a visual comparison, where even under various distortions, InstantID fails to reconstruct the correct female face, generating a male portrait instead. This demonstrates that the common distortions do not compromise the efficacy of our protection method." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 466, + 468, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 466, + 468, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 466, + 468, + 479 + ], + "type": "text", + "content": "6. Conclusion and Limitation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 486, + 555, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 555, + 653 + ], + "type": "text", + "content": "In this paper, we introduce IDProtector, a method designed to protect facial identities from encoder-based identity-preserving generation. By training a noise encoder to add adversarial perturbations onto images, IDProtector can protect an image within 0.2s. Our method achieves significantly lower face similarity in generated images than existing protection methods. IDProtector also demonstrates strong practical applicability. Its protection remains effective under distortions such as affine transformations and JPEG compression. Furthermore, IDProtector not only shows effectiveness against closed-source models, such as Midjourney, but also unifies protection against multiple ID-preserving models, setting a new standard for privacy protection against generative AI." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "content": "A limitation of our method is that the adversarial noise is not fully invisible, necessitating a trade-off between attack effectiveness and Imperceptibility. Future efforts will focus on enhancing the Imperceptibility of the noise while maintaining its robustness." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3026" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 170, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 170, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 170, + 85 + ], + "type": "text", + "content": "7. Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 137, + 115, + 149 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 137, + 115, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 137, + 115, + 149 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 158, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 158, + 296, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 158, + 296, + 223 + ], + "spans": [ + { + "bbox": [ + 61, + 158, + 296, + 223 + ], + "type": "text", + "content": "[1] Siying Cui, Jia Guo, Xiang An, Jiankang Deng, Yongle Zhao, Xinyu Wei, and Ziyong Feng. Idadapter: Learning mixed features for tuning-free personalization of text-to-image models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 950-959, 2024." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 224, + 296, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 224, + 296, + 277 + ], + "spans": [ + { + "bbox": [ + 61, + 224, + 296, + 277 + ], + "type": "text", + "content": "[2] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 280, + 296, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 280, + 296, + 334 + ], + "spans": [ + { + "bbox": [ + 62, + 280, + 296, + 334 + ], + "type": "text", + "content": "[3] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 335, + 295, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 335, + 295, + 368 + ], + "spans": [ + { + "bbox": [ + 62, + 335, + 295, + 368 + ], + "type": "text", + "content": "[4] Zinan Guo, Yanze Wu, Zhuowei Chen, Lang Chen, and Qian He. Pulid: Pure and lightning id customization via contrastive alignment. arXiv preprint arXiv:2404.16022, 2024." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 369, + 296, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 369, + 296, + 412 + ], + "spans": [ + { + "bbox": [ + 62, + 369, + 296, + 412 + ], + "type": "text", + "content": "[5] Junjie He, Yifeng Geng, and Liefeng Bo. Uniportrait: A unified framework for identity-preserving single- and multi-human image personalization. arXiv preprint arXiv:2408.05939, 2024." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 414, + 296, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 414, + 296, + 457 + ], + "spans": [ + { + "bbox": [ + 62, + 414, + 296, + 457 + ], + "type": "text", + "content": "[6] Xuanhua He, Quande Liu, Shengju Qian, Xin Wang, Tao Hu, Ke Cao, Keyu Yan, Man Zhou, and Jie Zhang. Id-animator: Zero-shot identity-preserving human video generation. arXiv preprint arXiv:2404.15275, 2024." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 459, + 296, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 459, + 296, + 512 + ], + "spans": [ + { + "bbox": [ + 62, + 459, + 296, + 512 + ], + "type": "text", + "content": "[7] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 514, + 296, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 514, + 296, + 568 + ], + "spans": [ + { + "bbox": [ + 62, + 514, + 296, + 568 + ], + "type": "text", + "content": "[8] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 570, + 296, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 570, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 62, + 570, + 296, + 624 + ], + "type": "text", + "content": "[9] Zhen Li, Mingdeng Cao, Xintao Wang, Zhongang Qi, Ming-Ming Cheng, and Ying Shan. Photomaker: Customizing realistic human photos via stacked id embedding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8640-8650, 2024." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 625, + 295, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 625, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 57, + 625, + 295, + 679 + ], + "type": "text", + "content": "[10] Chumeng Liang, Xiaoyu Wu, Yang Hua, Jiaru Zhang, Yiming Xue, Tao Song, Zhengui Xue, Ruhui Ma, and Haibing Guan. Adversarial example does good: Preventing painting imitation from diffusion models via adversarial examples. arXiv preprint arXiv:2302.04578, 2023." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 681, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 714 + ], + "type": "text", + "content": "[11] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Large-scale celebfaces attributes (celeba) dataset. Retrieved August, 15(2018):11, 2018." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 57, + 73, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 55, + 91, + 296, + 127 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 127 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 127 + ], + "type": "text", + "content": "This research is supported by the Ministry of Education, Singapore, under its Academic Research Fund Tier 2 (Award No: MOE-T2EP20124-0012)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "type": "text", + "content": "[12] Aleksander Madry. Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083, 2017." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 107, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 129 + ], + "type": "text", + "content": "[13] Midjourney.com. Midjourney. https://www.midjourney.com/home." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 131, + 553, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 131, + 553, + 175 + ], + "spans": [ + { + "bbox": [ + 316, + 131, + 553, + 175 + ], + "type": "text", + "content": "[14] Daniil Ostashev, Yuwei Fang, Sergey Tulyakov, Kfir Aberman, et al. Moa: Mixture-of-attention for subject-context disentanglement in personalized image generation. arXiv preprint arXiv:2404.11565, 2024." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 176, + 553, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 176, + 553, + 219 + ], + "spans": [ + { + "bbox": [ + 316, + 176, + 553, + 219 + ], + "type": "text", + "content": "[15] Omkar Parkhi, Andrea Vedaldi, and Andrew Zisserman. Deep face recognition. In BMVC 2015-Proceedings of the British Machine Vision Conference 2015. British Machine Vision Association, 2015." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 221, + 553, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 221, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 316, + 221, + 553, + 275 + ], + "type": "text", + "content": "[16] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 277, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 277, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 277, + 553, + 342 + ], + "type": "text", + "content": "[17] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 344, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 344, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 344, + 553, + 398 + ], + "type": "text", + "content": "[18] Mehrdad Saberi, Vinu Sankar Sadasivan, Keivan Rezaei, Aounon Kumar, Atoosa Chegini, Wenxiao Wang, and Soheil Feizi. Robustness of ai-image detectors: Fundamental limits and practical attacks. arXiv preprint arXiv:2310.00076, 2023." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 399, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 443 + ], + "type": "text", + "content": "[19] Hadi Salman, Alaa Khaddaj, Guillaume Leclerc, Andrew Ilyas, and Aleksander Madry. Raising the cost of malicious ai-powered image editing. arXiv preprint arXiv:2302.06588, 2023." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 445, + 553, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 553, + 499 + ], + "type": "text", + "content": "[20] Shawn Shan, Jenna Cryan, Emily Wenger, Haitao Zheng, Rana Hanocka, and Ben Y Zhao. Glaze: Protecting artists from style mimicry by {Text-to-Image} models. In 32nd USENIX Security Symposium (USENIX Security 23), pages 2187-2204, 2023." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 501, + 553, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 501, + 553, + 545 + ], + "spans": [ + { + "bbox": [ + 316, + 501, + 553, + 545 + ], + "type": "text", + "content": "[21] Yiren Song, Shengtao Lou, Xiaokang Liu, Hai Ci, Pei Yang, Jiaming Liu, and Mike Zheng Shou. Anti-reference: Universal and immediate defense against reference-based generation. arXiv preprint arXiv:2412.05980, 2024." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 547, + 553, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 547, + 553, + 611 + ], + "spans": [ + { + "bbox": [ + 317, + 547, + 553, + 611 + ], + "type": "text", + "content": "[22] Philipp Terhorst, Jan Niklas Kolf, Naser Damer, Florian Kirchbuchner, and Arjan Kuijper. Ser-fiq: Unsupervised estimation of face image quality based on stochastic embedding robustness. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5651-5660, 2020." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 613, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 656 + ], + "type": "text", + "content": "[23] Dani Valevski, Danny Wasserman, Yossi Matias, and Yaniv Leviathan. Face0: Instantaneously conditioning a text-to-image model on a face. arXiv preprint arXiv:2306.06638, 2023." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 658, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 714 + ], + "type": "text", + "content": "[24] Thanh Van Le, Hao Phung, Thuan Hoang Nguyen, Quan Dao, Ngoc N Tran, and Anh Tran. Anti-dreambooth: Protecting users from personalized text-to-image synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2116-2127, 2023." + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3027" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 462 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 137 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 137 + ], + "type": "text", + "content": "[25] Feifei Wang, Zhentao Tan, Tianyi Wei, Yue Wu, and Qidong Huang. Simac: A simple anti-customization method for protecting face privacy against text-to-image synthesis of diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12047-12056, 2024." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 173 + ], + "type": "text", + "content": "[26] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, and Anthony Chen. Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519, 2024." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 217 + ], + "type": "text", + "content": "[27] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 294, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 294, + 261 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 294, + 261 + ], + "type": "text", + "content": "[28] Xiaoyu Ye, Hao Huang, Jiaqi An, and Yongtao Wang. Duaw: Data-free universal adversarial watermark against stable diffusion customization. arXiv preprint arXiv:2308.09889, 2023." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 263, + 294, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 263, + 294, + 328 + ], + "spans": [ + { + "bbox": [ + 56, + 263, + 294, + 328 + ], + "type": "text", + "content": "[29] Yuxuan Zhang, Yiren Song, Jiaming Liu, Rui Wang, Jinpeng Yu, Hao Tang, Huaxia Li, Xu Tang, Yao Hu, Han Pan, et al. Ssr-encoder: Encoding selective subject representation for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8069–8078, 2024." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 331, + 294, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 294, + 384 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 294, + 384 + ], + "type": "text", + "content": "[30] Yuxuan Zhang, Yiren Song, Jinpeng Yu, Han Pan, and Zhongliang Jing. Fast personalized text to image synthesis with attention injection. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 6195-6199. IEEE, 2024." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 386, + 294, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 386, + 294, + 419 + ], + "spans": [ + { + "bbox": [ + 56, + 386, + 294, + 419 + ], + "type": "text", + "content": "[31] Boyang Zheng, Chumeng Liang, Xiaoyu Wu, and Yan Liu. Understanding and improving adversarial attacks on latent diffusion model. arXiv preprint arXiv:2310.04687, 2023." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 420, + 294, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 420, + 294, + 462 + ], + "spans": [ + { + "bbox": [ + 56, + 420, + 294, + 462 + ], + "type": "text", + "content": "[32] Zhengguang Zhou, Jing Li, Huaxia Li, Nemo Chen, and Xu Tang. Storymaker: Towards holistic consistent characters in text-to-image generation. arXiv preprint arXiv:2409.12576, 2024." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3028" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/ILIAS_ Instance-Level Image retrieval At Scale/bd262cec-535c-47b6-b08f-5430c3f91bd5_content_list.json b/2025/ILIAS_ Instance-Level Image retrieval At Scale/bd262cec-535c-47b6-b08f-5430c3f91bd5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..481434d94a6d7befd51af1854c2c1ff6f7082830 --- /dev/null +++ b/2025/ILIAS_ Instance-Level Image retrieval At Scale/bd262cec-535c-47b6-b08f-5430c3f91bd5_content_list.json @@ -0,0 +1,1350 @@ +[ + { + "type": "text", + "text": "ILIAS: Instance-Level Image retrieval At Scale", + "text_level": 1, + "bbox": [ + 259, + 130, + 740, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Giorgos Kordopatis-Zilos Vladan Stojnic Anna Manko Pavel Šuma Nikolaos-Antonios Ypsilantis Nikos Efthymiadis Zakaria Laskar Jiří Matas Ondřej Chum Giorgos Tolias", + "bbox": [ + 202, + 179, + 792, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "VRG, FEE, Czech Technical University in Prague", + "bbox": [ + 297, + 251, + 694, + 270 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 295, + 326, + 311 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This work introduces ILIAS, a new test dataset for Instance-Level Image retrieval At Scale. It is designed to evaluate the ability of current and future foundation models and retrieval techniques to recognize particular objects. The key benefits over existing datasets include large scale, domain diversity, accurate ground truth, and a performance that is far from saturated. ILIAS includes query and positive images for 1,000 object instances, manually collected to capture challenging conditions and diverse domains. Large-scale retrieval is conducted against 100 million distractor images from YFCC100M. To avoid false negatives without extra annotation effort, we include only query objects confirmed to have emerged after 2014, i.e. the compilation date of YFCC100M. An extensive benchmarking is performed with the following observations: i) models fine-tuned on specific domains, such as landmarks or products, excel in that domain but fail on ILIAS ii) learning a linear adaptation layer using multi-domain class supervision results in performance improvements, especially for vision-language models iii) local descriptors in retrieval re-ranking are still a key ingredient, especially in the presence of severe background clutter iv) the text-to-image performance of the vision-language foundation models is surprisingly close to the corresponding image-to-image case.", + "bbox": [ + 86, + 328, + 485, + 690 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "website: https://vrg.fel(cvut.cz/ilias/", + "bbox": [ + 89, + 691, + 450, + 705 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 722, + 220, + 738 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The ability to recognize and differentiate every unique object instance in the physical world represents one of the ultimate goals for foundation representation models [5, 34, 41, 75]. This work aims to assess this capability through the lens of instance-level image retrieval at a very large scale. Instance-level image retrieval corresponds to searching for images of particular objects within large collections. All images of a particular object form their own instance-level class. This is an important information retrieval task due to its numerous real-world applications in robotics [30, 45], e", + "bbox": [ + 88, + 750, + 482, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3e869c247b7b01b1f9361e6e73886bb942599fafb8737dd642fe9cbc714c3a52.jpg", + "image_caption": [ + "Figure 1. Performance timeline on ILIAS. Curves indicate best performance in chronological order for image-to-image and text-to-image retrieval, showing a significant boost with the release of foundation models. Representations are linearly adapted via multi-domain learning on UnED [72]. Re-ranking with local descriptors achieves the best results by a significant margin." + ], + "image_footnote": [], + "bbox": [ + 514, + 292, + 906, + 474 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "commerce [74, 76], and cultural heritage [16, 50], to name just a few. The task faces challenges because of the substantial variations among positive examples, such as illumination/viewpoint [21, 62] changes and background clutter [4, 31]. An additional difficulty is the high similarity among negatives, which is driven by the extremely fine granularity in the class definitions. It becomes even more challenging at a real-world scale, where searching through millions or even billions of images requires handling an open-world setup with countless unseen objects spanning diverse and complex domains.", + "bbox": [ + 509, + 566, + 908, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Benchmarking instance-level retrieval under real-world challenges is currently limited by the lack of suitable datasets. Constructing a dataset with instance-level class definitions necessitates huge development effort, reflected by the many shortcomings of existing datasets. Shortcomings exist in several key aspects, such as dataset size [66], domain diversity [39, 53], and ground-truth accuracy [67], which suffers from both false positives and false negatives. Popular datasets are typically limited to landmarks [39], and as dataset scale increases, ground-truth quality tends to decline [52, 67]. This is a consequence of automating the", + "bbox": [ + 509, + 734, + 910, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14777", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ground-truth creation process to facilitate scaling up. To address such limitations, we introduce the Instance-Level Image retrieval At Scale (ILIAS) evaluation dataset.", + "bbox": [ + 89, + 90, + 480, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The creation of our dataset has two key elements. First, query and positive images are manually captured to ensure challenging variations, covering 1,000 objects across diverse domains. Second, to expand the dataset size without ground-truth errors or additional annotation effort, we leverage a key technique: distractor images, collected in 2014 from YFCC100M, are combined with query objects verified not to have publicly existed until after 2014. This distractor set includes 100 million images, two orders of magnitude larger than the largest existing dataset [39]. Notably, all images have a permissive license, allowing us to ensure long-term online availability to the full extent.", + "bbox": [ + 88, + 138, + 482, + 318 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ILIAS includes both image and text queries. The latter is in the form of detailed descriptions of objects and their distinctive features. The dataset is designed to support future research in image-to-image and text-to-image retrieval for particular objects, and additionally serves as a large-scale benchmark for evaluating representations of foundation vision and language models (VLM) [41, 75]. To facilitate faster experimentation, we provide a mini, but challenging, version (5M) of the distractor set.", + "bbox": [ + 89, + 320, + 482, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We perform an extensive evaluation comparison, including many foundation image-to-image and text-to-image models, and establish a comprehensive testbed that enables future comparisons. The provided evaluation includes retrieval with global image representation but also re-ranking techniques that use local representations [37, 43, 56] and query expansion [10, 40]. We observe the following:", + "bbox": [ + 89, + 458, + 482, + 565 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Performance of standard 10-year-old models, measured by mean Average Precision, is as low as $1.3\\%$ , while the best-performing model achieves $31.3\\%$ , as shown in Fig. 1. This points out the vast progress of representation models and the high challenging factors of ILIAS.", + "- VLMs are the top-performing models.", + "- Smaller (ViT-B) models trained/tested on large resolution (512/724) outperform larger models (ViT-L) trained/tested on small resolution (256/384).", + "- Using Universal Embedding Dataset (UnED) [72] to learn a linear adaptation layer on top of frozen models improves performance of most models, making it a candidate training set to couple with ILIAS. Notably, VLMs demonstrate the largest benefits, presumably because their training stage does not optimize image-to-image relations.", + "- In contrast to the current belief [48], local representation is a key ingredient, while global representation, despite being efficient and compact, performs much lower.", + "- The performance gap between image-to-image and text-to-image models is surprisingly small. Therefore, detailed text queries are a reasonable proxy in the absence of image queries, even at the instance level." + ], + "bbox": [ + 89, + 566, + 482, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 513, + 90, + 651, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we review the related work in terms of existing datasets and benchmarks in the literature.", + "bbox": [ + 511, + 114, + 903, + 143 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Datasets. Tab. 1 presents the datasets from the image retrieval literature related to ILIAS. The datasets can be compared based on five main axes: (i) Class definition adopted. Many datasets [20, 32, 33, 39, 67] adopt a strict definition very similar to ours, satisfying instance-level requirements. Others [2, 53, 66, 76] adopt a more relaxed definition, where some minor variations are permitted, e.g. color changes in objects of the same class. Even more relaxed are the fine-grained definitions [74], where the object of the very same type is considered related, e.g. same product with different variant. (ii) Domain of the dataset. Most datasets are tailored for a specific domain. Landmarks are among the most popular domains [2, 20, 33, 39, 67]. Other domains include products [32, 36, 76] and fashion [28, 53]. Some datasets cover multiple domains, either being standalone [66] or bundle of repurposed datasets [47, 72]. (iii) Scale of database. Most of the datasets are small-scale, counting a few thousand images [28, 66, 76]. Larger ones [39, 72, 74] expend slightly above a million. None satisfies large-scale requirements. (iv) Noise in ground truth. Most datasets consist of clean annotations, except for a few cases that contain inaccuracies, including false positives [67], i.e. images wrongly annotated as relevant, false negatives [32, 39], i.e. relevant images that have not been annotated as positives, or the possibility of false positives [39]. (v) Availability. Most datasets are publicly available with permissive licenses, with few exceptions of partial [2, 20] or no [33, 74] availability. To this end, no publicly available dataset fits the strict instance-level definition, contains objects from multiple domains, ensures error-free labeling and is large scale. This gap is filled with ILIAS satisfying all the aforementioned requirements.", + "bbox": [ + 511, + 145, + 906, + 628 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Evaluation benchmarks. Benchmarking [51] tracks the progress in the field, which is even more necessary with the emergence of foundation models. Several benchmarks papers [24, 25, 77] exist in the instance-level retrieval literature, investigating the impact of learning scheme, post-processing, model ensembling, query expansion, and whitening. The most relevant benchmark to ILIAS is UnED [72] that combines existing datasets to create a union that evaluates models performance across various domains. Due to its wide variety, UnED serves as the training dataset for linear adaptation.", + "bbox": [ + 511, + 628, + 905, + 794 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Regarding the evaluation of foundation models, the most common practice [14, 34, 57] is measuring classification performance on top of frozen models on ImageNet [12]. This is performed either with or without the training of a classifier via linear probing or k-NN search. Furthermore, models are usually evaluated on dense prediction tasks [34] and several multiple-downstream single-domain tasks [1].", + "bbox": [ + 511, + 795, + 903, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14778", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/bd1613d8e9182d7c1e32abac76db6d91abdd4dddc940b275138132fdc55b619e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
datasetsyearobjectsquerypositivesdatabasegtclass def.domainbboxonlinelicense
UKB [32]20062.5K10K10K10KFNILproductXFullyN/A
Holidays [20]20085005009911MCleanILlandmarkXPartiallyCC
Sculptures [2]201110703.1K3.1KCleanPartial ILsculptureXPartiallyFlickr TC
INSTRE [66]2015200125027.3K27.3KCleanPartial ILmulti✓FullyFlickr TC
SOP [53]201511.3K60.5K60.5K60.5KCleanPartial ILproductXFullyMIT License
InShop [28]20163.9K14.2K12.6K12.6KCleanPartial ILfashion✓FullyN/A
R-Oxford [39]201811705K1MFN?ILlandmarkXFullyFlickr TC, CC
R-Paris [39]201811706.3K1MFN?ILlandmarkXFullyFlickr TC, CC
GLDv1 [33]201830KN/AN/A1.1MCleanILlandmarkXPartiallyMultiple
GLDv2 [67]20203181.1K3.1K762KFPILlandmarkXFullyCC/ Public-domain
Product1M [76]20213926.5K40K40KCleanPartial ILproductXPartiallyN/A
RP2K [36]20211.2K10.9K10.9K10.9KCleanPartial ILproduct✓FullyN/A
GPR1200 [47]20211.2K12K12K12KMixIL+FGmultiXFullyMultiple
eProduct [74]202120610KN/A1.1MN/AFGproductXNoN/A
FORB [69]2023N/A13.9K4.5K49.8KCleanILplanarXFullySnap Inc.
UnED [72]202321K241K244K1.4MMixIL+FGmultiXFullyMultiple
ILIAS20251,0001,2324,715100MCleanILmulti✓FullyCC
", + "bbox": [ + 130, + 75, + 861, + 310 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1. Comparison with other instance-level datasets. Datasets are compared based on their size (object, query, positives, database), the accuracy of the ground truth (gt), type of class definition, domain, supplementary annotations (bbox) and accessibility (online, license). N/A: not available. FP/FN: false positives/negatives. FN?: possibility of false negatives. Mix: combination of clean and noisy datasets. IL: instance-level. FG: fine-grained. Partial IL: instance-level with subtle variations among same class objects. CC: Creative Commons.", + "bbox": [ + 89, + 314, + 906, + 369 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For VLMs, zero-shot classification and retrieval serve as the primary benchmarks [41, 70, 75], utilizing class text labels. In this work, we provide similar evaluation protocols tailored for instance-level retrieval. One can test the raw model capabilities or adapt for the instance-level task via linear adaptation on UnED. Text-to-image retrieval is also facilitated for the evaluation of VLMs.", + "bbox": [ + 89, + 378, + 480, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. ILIAS dataset", + "text_level": 1, + "bbox": [ + 89, + 502, + 235, + 517 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Composition and collection", + "text_level": 1, + "bbox": [ + 89, + 527, + 334, + 544 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Instance-level class definition. Following an instance-level class definition [39, 71], we consider all indistinguishable object instances of the real world to form their own class. Nevertheless, we add a restriction to consider a pair of images as relevant to each other only if there is a view overlap. Other cases are explicitly not included in the dataset, contrasting the existing work [28, 53, 67]. Therefore, models should mostly rely on estimating the visual similarity and less on shortcuts through semantics.", + "bbox": [ + 89, + 551, + 482, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overview. ILIAS supports both image-to-image (i2i) and text-to-image (t2i) retrieval and follows the standard setup for retrieval datasets, consisting of two main parts: (i) query images and text, and (ii) database (db) images. The objective is to rank positives - db images relevant to the query - at the top ranks. The collected objects cover a wide range of categories and are not restricted to specific domains. An overview of some collected objects is provided in Fig. 2.", + "bbox": [ + 89, + 686, + 482, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Queries and positives are created/collected by a group of collectors that are well-informed about the task objectives. In addition to positives, in the database, we include numerous distractors - irrelevant (negative) images to the queries - that make retrieval more challenging. Following previous work [39], adding a large, uncurated set of ran", + "bbox": [ + 89, + 810, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "dom images achieves this. The larger the set, the higher the chances of hard negatives – images with similar appearance or semantics to the queries. To this end, we select the YFCC100M [61] dataset to serve as the source of distractors due to its size and permissive license.", + "bbox": [ + 511, + 378, + 906, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Selected objects. Ensuring that distractor images include no false negatives cannot be performed in a scalable or accurate way if one relies on human annotation or metadata. Instead, we take advantage of the fact that YFCC100M was crawled from Flickr in 2014. Hence, an object qualifies in ILIAS if it could not have appeared on Flickr before 2014. To verify this, we rely either on publicly available information, e.g. objects known to be created/manufactured after 2014, or on the collector's knowledge about the object not being publicly available. Additionally, we opt for objects with distinctive and unique features that set them apart from others within the same category. For example, we avoid recent smartphones that look like plain black screens or new objects with distinctive parts closely resembling older ones.", + "bbox": [ + 511, + 458, + 908, + 671 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Queries and positives. Query images depict the instance on a clean or uniform background. When this is not feasible (e.g. buildings or statues), background blurring or cropping is applied. This is performed to avoid including background objects in the query that do not have corresponding positives in our ground truth information. Positives are images featuring the query object in challenging conditions, such as clutter, scale changes, occlusions, and partial views. Prior work [39] reveals that easy positives dominate performance metrics. Thus, we specifically opt for challenging cases that cannot be easily retrieved by the models. To avoid taking advantage of camera identification, most query and positive images are captured with at least two different camera models to introduce diversity. We also incorporate older camera models that are used in YFCC100M.", + "bbox": [ + 511, + 674, + 908, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14779", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b80f76ca47b653aad9f0f85ba98c45ce40709240a8b2314ae451ca72300b8782.jpg", + "image_caption": [ + "Figure 2. Examples of query, positive and hard negatives within the distractor set. Average Precision per query and rank of the negatives and positives is reported using SigLIP [75] model. Gray: queries. Green: positives. Red: distractors." + ], + "image_footnote": [], + "bbox": [ + 151, + 87, + 836, + 319 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Each text query consists of a detailed and fine-grained textual description of an object. Descriptions are initially created by a large language model prompted to provide highly detailed depictions of the object shown in query images. Generated descriptions are manually edited to fix errors, insufficient descriptions, or nuances of the model.", + "bbox": [ + 89, + 354, + 482, + 445 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Distractors. The YFCC100M dataset was chosen for the distractor set due to its large scale and diverse range of concepts. It consists of 100 million Flickr images, collected without specific filtering, aside from being shared under a permissive CC-BY license.", + "bbox": [ + 89, + 445, + 483, + 521 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Bounding box annotation. We include supplementary bounding boxes that specify the precise location of objects in query and positive images. They provide statistics about the position and size of object areas, assist our analysis of the dataset challenges, and support future research in instance-level localization.", + "bbox": [ + 89, + 521, + 483, + 611 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Evaluation metric. Retrieval performance is evaluated via mean Average Precision (mAP), a widely used metric in instance-level image retrieval [37, 38, 40]. Specifically, we adopt mAP@1k [67], which assesses the ranking of the top-1k nearest neighbors for each query, treating any positive not ranked among the top-1k as not retrieved. We estimate the area under the curve using rectangles and not trapezoids.", + "bbox": [ + 89, + 612, + 483, + 719 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Statistics", + "text_level": 1, + "bbox": [ + 89, + 727, + 196, + 742 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Dataset size. The final ILIAS dataset includes 1,000 object instances captured in 5,947 images, of which 1,232 are queries and 4,715 are positives. Fig. 3a shows the distribution of positives per object. Also, 99,144,315 images from YFCC100M are downloaded. All images (queries, positives, distractors) are transferred through Flickr to ensure the same pre-processing.", + "bbox": [ + 89, + 750, + 482, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Taxonomy. A hierarchical 3-level taxonomy is composed for ILIAS. All instances are assigned across one to three categories of different granularity levels. The taxonomy", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "consists of 8 categories on the coarser level, 42 on the mid level, and 38 on the finer level. The categories are derived through manual labeling of the objects based on their semantic content. To form the coarser-level categories, we use domain definitions borrowed from prior work [28, 53, 67] to align with the literature, i.e. art, landmarks, products, fashion. We also define novel categories based on the objects that do not fit into any existing domain. The distribution of objects across categories is uneven, e.g. ranging from 168 and 162 for art and landmarks to 83 for products. Each mid- and finer-level category contains at least 4 instances. Note that taxonomy is given to provide statistics about the domains of objects and assist our analysis instead of being leveraged as ground truth. The distribution of taxonomy categories can be inferred by Fig. 5, and a detailed figure is provided in the supplementary material.", + "bbox": [ + 511, + 354, + 906, + 597 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Bounding box analysis. A total number of 6,117 bounding boxes are annotated for both queries and positives. Note that positives may display multiple objects of near identical appearance to the query; in such cases, bounding boxes are drawn on all indistinguishable objects. There are 235 images with more than one bounding box. Based on the annotated bounding boxes, we compute the area covered in the image by the object instances to derive its relative scale. Fig. 3c shows the distribution of the scale ratio for queries and positives. Most objects in queries cover the largest area of the images; while in the vast majority of positives, the object covers a small area of less than half the image. It is a result of the severe scale changes and partial views. Moreover, we use the Segment Anything Model (SAM) [23, 42] to extract object segments from positives. The number of detected segments outside the query object's bounding boxes is computed. This indicates clutter from other items in the positives. Fig. 3b shows the segment number distribution, with most images containing multiple segments due to clutter.", + "bbox": [ + 511, + 598, + 908, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14780", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/682c2ec5fdb3ff2516d4dcacbea8a3efb6b4bd7ac55b67b3b003a252a840c08e.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 91, + 92, + 222, + 202 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7d23c8a6223101fd7452578cac9c2d0217afbe8dc317f95a94a3584b8d8355ba.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 222, + 92, + 349, + 202 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/87fe4f25dbb74f5839ab093adcaadd6bb9fd1fc582a1e8708babd9a956df2551.jpg", + "image_caption": [ + "(c)", + "Figure 3. ILIAS statistics. (a) number of positives per object, (b) positive distribution by the SAM segments outside the bounding box, (c) image distribution by the relative bounding box area." + ], + "image_footnote": [], + "bbox": [ + 349, + 92, + 475, + 202 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. mini-ILIAS", + "text_level": 1, + "bbox": [ + 89, + 263, + 210, + 277 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We provide a small version of ILIAS, called mini-ILIAS, to facilitate quick experimentations. It consists of the query and positive images collected for ILIAS, and a subset of the YFCC100M distractors. Instead of randomly subsampling YFCC100M, we construct a challenging subset with the help of VLMs. We aim at selecting distractors displaying objects of similar categories as the query objects. We use the text category labels of the taxonomy as text queries. We also extend them with standard templates used for zero-shot recognition [41], which resulted in several thousands. T2i similarity between each text query and each distractor image is estimated. A similarity score for each distractor is derived based on its maximum similarity over the text queries. We ensemble the scores of 3 VLMs to rank images. The top-5M ranked distractors compose the final mini-ILIAS. Our experiments indicate that this subset is significantly more challenging than a random subset of the same size.", + "bbox": [ + 89, + 285, + 483, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Benchmark methods", + "text_level": 1, + "bbox": [ + 89, + 556, + 287, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We describe the methods and foundation models we evaluate on ILIAS, which are grouped according to their type of representations used for retrieval in global (i2i) representations, re-ranking with global (i2i) representations, re-ranking with local (i2i) representations, and text-to-image. A detailed list of models, their performance, and implementation details are in the supplementary materials.", + "bbox": [ + 89, + 583, + 482, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Image-to-image retrieval with global representations. Global representation methods use image encoders to map images to global descriptors and rank db images based on cosine similarity. We evaluate legacy and recent foundation models, varying in architecture, descriptor dimensionality, training scheme, training data, and input resolution. Foundation models [5] are the models trained with a training set on the scale of a hundred million. Particularly, 23 CNN [17, 26, 29, 49, 58, 59, 78] and 45 ViT [13, 14] models, trained with supervision [22, 54, 64, 68], self-supervision [7, 8, 18, 34], distillation [1, 46, 64], or visual-language alignment [9, 41, 57, 65, 70, 75] are benchmarked. Most of the non-foundation models are trained on ImageNet [12]. There are models trained on single specific do", + "bbox": [ + 89, + 689, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "mains [27, 35, 48], i.e. landmarks or products on GLDv2 or SOP. Universal models [1, 46, 72, 73] trained on multi-domains or multi-task schemes are included. The full list of models and results is provided in supplementary materials.", + "bbox": [ + 511, + 90, + 905, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To mitigate the differences in training resolution, we use three widely-used resolutions, i.e. 384, 512, and 724 and resize images so that their larger dimension matches one of the three. The test resolution is defined to be one resolution above the one used for training, e.g. a network trained with 224 or 384 is tested with 384 or 512, respectively. The vast majority of models achieve best performance under this rule. Similar behavior is observed in the literature [55, 63].", + "bbox": [ + 511, + 152, + 906, + 273 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Linear adaptation for i2i retrieval. Pre-trained foundation models, as well as legacy models, are trained to extract representations that are applicable to various tasks; not all encoded features are directly relevant to instance-level retrieval. To adapt the representation to the task at hand, we propose to train a single linear layer (projection) on top of frozen backbones. The recently introduced Universal Embeddings (UnED) dataset [72] is used for learning the linear adaptation. UnED contains images from 8 different domains with fine-grained and/or instance-level class annotation. In our experiments, the linear layer that converts the backbone output to a 512D descriptor is trained on a uniformly sampled subset of 1M images from UnED. The linear adaptation layer is trained with the UJCDs [72] method.", + "bbox": [ + 511, + 275, + 905, + 487 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Text-to-image retrieval. Text-to-image retrieval is performed using Vision-Language Models (VLMs) trained to align the two modalities. Retrieval is performed based on cosine similarity between the text query and db image descriptors that are extracted using the textual and visual encoder, respectively. We evaluate 17 VLM models.", + "bbox": [ + 511, + 489, + 905, + 579 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Re-ranking with global representations. Such methods rely on global descriptors for exhaustive search during the initial ranking, but also for a second refinement stage that issues a new query. We experiment with $\\alpha$ QE [40], the adaptive variant of average Query Expansion [10]. After the initial ranking, the descriptors of the top-ranked images are aggregated with the query via weighted average pooling. The weights are derived from the similarity to the query in the power of $\\alpha$ . We don't have a validation set; hence, we use a fixed value $\\alpha = 1$ .", + "bbox": [ + 511, + 580, + 905, + 731 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Re-ranking with local representations. These re-ranking methods rely on global descriptors for exhaustive search during the initial ranking but estimate query-to-db image similarity based on local descriptors for a second refinement stage of the ranked list of images. We experiment with three methods: (i) Chamfer Similarity (CS) [3, 44] on the similarity matrix between local descriptors across the image pair. We use the asymmetric variant of CS with max over db descriptors and sum over query descriptors. (ii) Spatial Verification (SP) [6, 15, 37], a common re-ranking method where point correspondences are processed with a RANSAC-like", + "bbox": [ + 511, + 734, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14781", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4ce71404f66387c321e50c94fc99e008a90563dca08aad2b6d38d93f372d00f2.jpg", + "image_caption": [ + "Figure 4. Comparison with other instance-level retrieval datasets via reporting mAP@1k. INSTRE: 27.3K db size, multi-domain. GLDv2: 762K db size, single-domain. SOP: 60.5K db size, single-domain. Different network types are color-coded. For GLDv2 and SOP, models fine-tuned on these domains with the corresponding training sets are highlighted. No linear adaptation is used." + ], + "image_footnote": [], + "bbox": [ + 186, + 73, + 406, + 229 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e7c5074ef4c2bcc30f91ca2cf784aef44481d5c23fe4d25dd109ec16d8c3e356.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 418, + 74, + 612, + 229 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/40a3c2b659d85d707ea1775a60ba6be002b095d2cb8ed37f71020b047fb89ede.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 74, + 826, + 229 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "process and the number of inliers is used for re-ranking. (iii) AMES [56], a recent transformer-based network to estimate the similarity between sets of local descriptors. Due to the scale of ILIAS database, we use only 100 binary local descriptors for each database image and 600 for the query image. Local descriptors are extracted using the base variant of DINOv2 with registers [11, 34] and selected based the local descriptor detector used in AMES [56]. Top-1k retrieved images are re-ranked.", + "bbox": [ + 88, + 287, + 485, + 422 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 435, + 225, + 453 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate all the above models and methods, extracting useful insights regarding the factors that boost retrieval performance. ILIAS is compared with other existing datasets for instance-level image retrieval. We analyze the performance of selected models1 to break down the impact of different ILIAS attributes, such as domains, clutter, and scale. Unless stated otherwise, we use the large model variants with the largest resolution available, e.g. in our analysis we use SigLIP ViT-L trained with 384 resolution.", + "bbox": [ + 89, + 460, + 483, + 597 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Comparison with other instance-level datasets", + "text_level": 1, + "bbox": [ + 89, + 604, + 480, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Fig. 4, ILIAS is compared with other instance-level retrieval datasets based on evaluation of the same models. Linear adaptation is not applied as parts of GLDv2 and SOP are included in the UnED dataset. Only for the sake of this comparison, and for no other experiment in this work, we use models fine-tuned on specific domains (in-domain models), i.e. on the training sets of SOP and GLDv2. INSTRE, which is also multi-domain, shows a correlation to ILIAS, but its performance is saturated due to its small size. For single-domain datasets, in-domain models outperform others by a large margin, with few exceptions, i.e. DINOv2, which includes the trainset of GLDv2 in its training data. Several multi-domain models perform well on SOP since their training set is usually included in the training data. However, in-domain and multi-domain models face challenges on ILIAS, highlighting the diversity of our dataset.", + "bbox": [ + 89, + 626, + 482, + 867 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Method comparison", + "text_level": 1, + "bbox": [ + 511, + 287, + 705, + 303 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Image-to-image retrieval with global representations. Tab. 2 presents the performance of global descriptor models on ILIAS. Selected models are presented to highlight useful comparisons, while many other models are included in the supplementary material. The main factors that improve performance are the size of the training set, training resolution, and model architecture, which aligns with the literature. The impact of dataset size is apparent in various model combinations, e.g. CLIP with openai and laion2b. This is also pronounced by the dominance of foundation models. Training with large resolution brings significant gains and consistently improves mAP@1k. In some cases of SigLIP, smaller models trained with large resolutions outperform larger ones trained with small resolutions. For the models of the same resolution, it is a common trend for larger model variants to bring corresponding performance gains. In general, VLMs perform the best. From non-VLMs, only DINOv2 and Unicom achieve competitive performance. Masked Image Modeling (MIM) and supervised models are not performing well. Our linear adaptation scheme is very effective, improving most models. The boost is more pronounced in the case of VLMs. A possible explanation for such improvements is that image-to-image relations are not optimized during the training of VLMs.", + "bbox": [ + 511, + 309, + 906, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Text-to-image retrieval. Following results in Tab. 2, similar conclusions are derived for the t2i case. Retrieval performance improves with the scaling of the training data. The larger model achieves significantly better results, i.e. compare the base with large variants. Finally, it is noteworthy that the best performance achieved by SigLIP2 is very close to the i2i performance when no adaptation is used. Note that t2i includes 1k text queries in total, with one query per object, while i2i 1,232 image queries.", + "bbox": [ + 511, + 672, + 908, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation of mini-ILIAS selection. Tab. 3 shows performance on mini-ILIAS for five models with linear adaptation. The selected subset is significantly more challenging than a random selection of 5M images. More precisely, a set of approximately $\\sim 26\\mathrm{M}$ random images matches the performance of mini-ILIAS.", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "1 Although the very recent SigLIP2 is the best-performing model, we conduct most experiments with SigLIP.", + "bbox": [ + 89, + 875, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14782", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/010d658bd9aca2244660ed55227da0f526577ce8ffccc0b13ed6bd466bd154a5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
modelarchtraindatasetdata sizetrain restest resimage-to-imagetext-to-image
5M†100M†100M100M5M
ResNet50 [17]R50supin1k1M2243842.51.81.7--
DINO [8]R50sslin1k1M2243844.12.92.9--
ConvNext [29]CN-Lsupin1k1M2883844.22.92.2--
OAI-CLIP [17]R50vlaopenai400M2243848.56.03.21.52.3
OpenCLIP [19, 29]CN-Bvlalaion2b2B25638418.114.07.94.67.0
OpenCLIP [19, 29]CN-Lvlalaion2b2B32051222.918.39.68.111.5
ViT [13, 54]ViT-Bsupin1k1M2243841.91.31.0--
EVA-MIM [14]ViT-Bsslin22k142M2243844.73.22.1--
ViT [13, 54]ViT-Bsupin21k14M2243846.24.43.0--
DINO [8]ViT-Bsslin1k1M2243846.64.83.7--
UDON-CLIP [73]ViT-Bsupuned2.8M2243849.26.75.9--
OAI-CLIP [41]ViT-Bvlaopenai400M22438410.77.94.21.62.7
EVA-CLIP [57]ViT-Bvlamerged2b2B22438411.78.75.92.54.4
MetaCLIP [70]ViT-Bvla2pt5b2.5B22438412.79.46.64.97.6
DINOv2 [34]ViT-Bssllvd142m142M51872415.012.111.5--
SigLIP [75]ViT-Bvlawebli10B25638420.616.711.57.510.3
SigLIP [75]ViT-Bvlawebli10B38451226.221.515.611.014.4
SigLIP [75]ViT-Bvlawebli10B51272427.523.016.611.114.6
SigLIP2 [65]ViT-Bvlawebli10B51272428.623.515.410.414.6
EVA-MIM [14]ViT-Lsslin22k14M2243843.92.71.5--
ViT [13, 54]ViT-Lsupin21k14M2243847.35.34.6--
EVA-MIM [14]ViT-Lsslmerged38m38B2243848.86.14.7--
OAI-CLIP [41]ViT-Lvlaopenai400M22438415.811.97.04.66.7
OpenCLIP [9, 19]ViT-Lvlalaion2b2B22438417.513.79.47.09.4
Unicom [1]ViT-Ldistlaion400m400M33651218.614.613.9--
OAI-CLIP [41]ViT-Lvlaopenai400M33651219.915.29.45.88.4
DINOv2 [34]ViT-Lssllvd142m142M51872418.815.315.3--
EVA-CLIP [57]ViT-Lvlamerged2b2B33651220.916.010.97.210.6
MetaCLIP [70]ViT-Lvla2pt5b2.5B22438421.716.911.79.213.1
SigLIP [75]ViT-Lvlawebli10B25638426.321.815.212.816.4
SigLIP [75]ViT-Lvlawebli10B38451234.328.919.618.122.2
SigLIP2 [65]ViT-Lvlawebli10B51272437.331.320.819.824.7
", + "bbox": [ + 171, + 75, + 831, + 458 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1ab422f096486bf278655cbc0eac4c71bf2b41bf54ed15abb733c155cb371b4b.jpg", + "table_caption": [ + "Table 2. Performance comparison using mAP@1k on ILIAS and mini-ILIAS for global representation models for i2i and t2i. Comparison of model architecture (arch), training scheme (train), training data, and train/test resolution. $\\dagger$ indicates results with the linear adaptation. 5M and 100M correspond to the mini and full versions of the dataset, respectively. sup, ssl, dist, vla: supervised learning, self-supervised learning, distillation and vision-language alignment. R50, CN: ResNet50 and ConvNext." + ], + "table_footnote": [], + "table_body": "
model100M5M-mini5M-rand
DINOv2† [34]15.318.822.7±0.2
EVA-CLIP† [57]16.020.928.8±0.2
MetaCLIP† [70]16.921.729.2±0.1
OpenCLIP† [19, 29]18.322.930.9±0.2
SigLIP† [75]28.934.341.8±0.1
", + "bbox": [ + 122, + 520, + 444, + 606 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/085a1528a507639b0c63d18cba9d38f092f99136207b4cbaf9b663aa32408c2c.jpg", + "table_caption": [ + "Table 3. A challenging distractor subset for mini-ILIAS. mAP@1k evaluated for different distractor sets, 100M: the full dataset, 5M-mini: mini-ILIAS subset, 5M-rand: random subset. We report the mean and std of 3 randomly sampled subsets. $\\dagger$ indicates results with the linear adaptation." + ], + "table_footnote": [], + "table_body": "
rerankingSigLIP [75]SigLIP† [75]
mAP@1koraclemAP@1koracle
global19.648.728.956.0
αQE1 [10, 40]22.144.733.756.9
αQE2 [10, 40]20.440.831.554.4
αQE5 [10, 40]14.334.923.549.3
CS [44]22.948.732.556.0
SP [37]21.848.730.556.0
AMES [56]26.448.735.656.0
", + "bbox": [ + 119, + 679, + 450, + 815 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Performance comparison for re-ranking methods. Oracle represents the performance of perfect re-ranking at the top-1k images. Top: query expansion with global descriptors. Bottom: re-ranking with local descriptors. †: results with linear adaptation.", + "bbox": [ + 89, + 816, + 482, + 875 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Retrieval with re-ranking. Tab. 4 shows the performance of re-ranking methods applied on top of SigLIP with and without linear adaptation on ILIAS. Complementary to $\\mathrm{mAP}@1\\mathrm{k}$ , an oracle-based top-1k re-ranking metric is reported as the upper bound of a re-ranking method that processes the top 1k images. Local similarity estimated by a learned model proves to be very effective for re-ranking. Nevertheless, the oracle re-ranking performance indicates that there is a lot more space for improvements. Re-ranking with QE is useful when the number of aggregated neighbors is low and drops below the baseline when the number of neighbors is increased. Notably, global re-ranking affects and, interestingly, decreases oracle performance since the whole db is re-ranked; while local re-ranking does not affect it since it is performed only on a shortlist of images.", + "bbox": [ + 511, + 523, + 906, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Analysis", + "text_level": 1, + "bbox": [ + 511, + 757, + 614, + 773 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Performance per domain. Fig. 5 shows the performance per taxonomy categories. The taxonomy annotations allow a fine-grained view of the results, which can possibly allow us to capture imbalanced improvements in future work. For example, DINOV2, despite being overall inferior to SigLIP, is outperforming it in categories like architecture and sculptures or is quite similar in categories like public art and paper art. This is possibly attributed to the curation and com", + "bbox": [ + 511, + 779, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14783", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0317cd1cc7d997e82893f6d6434525ce3d95ec7803efde0fd5563d2649867c53.jpg", + "image_caption": [ + "Figure 5. Performance comparison per category. mAP@1k averaged over objects in the same mid-level taxonomy category, organized by their primary-level category size, with sorting within each group by SigLIP† + AMES performance. Comparison between SigLIP with and without adaptation, SigLIP combined with AMES reranking, SigLIP t2i, and DINOv2. † indicates results with the linear adaptation." + ], + "image_footnote": [], + "bbox": [ + 112, + 78, + 869, + 239 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5397c4df56f4103ac2489ba3f99c0326cd4a35faeceb0dabb7baf6a9e023db0d.jpg", + "image_caption": [ + "Figure 6. Performance comparison reporting AP per query for different approaches with SigLIP. Pearson correlation reported in parenthesis. $\\dagger$ indicates results with the linear adaptation." + ], + "image_footnote": [], + "bbox": [ + 94, + 282, + 460, + 563 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "position of the DINOV2 training set, which includes artwork and landmark datasets. Also, some categories are hurt by reranking with AMES, with some demonstrating big drops, i.e. sport, gaming, perfume. These categories deviate significantly from the domain AMES is trained, i.e. landmarks, which could potentially justify such drops.", + "bbox": [ + 89, + 613, + 482, + 704 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Per query comparisons. Fig. 6 shows the AP per query for various methods. Linear adaptation boosts most queries, i.e. performance drop only for 192 queries. Image- and text-based retrieval are not strongly correlated despite performing similarly, which is good evidence [60] for the effectiveness of model ensembles. Indeed, ensembling i2i and t2i by averaging similarities brings $+6.1$ improvement over i2i retrieval. Query expansion improves the queries with at least some positives at top positions, i.e. AP greater than 20. However, it harms many low-performing queries by aggregating descriptors irrelevant to the query. AMES improves the majority of the queries; however, many are harmed, indicating that there is plenty of room for improvement.", + "bbox": [ + 89, + 704, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bdf3609aaa66228e22da5ab294907ec2faef93ce98c4dc4d2e9dbf3bf7e0345f.jpg", + "image_caption": [ + "Figure 7. Performance evaluation (mAP@1k) across different amounts of object area coverage and background clutter. Positives across all queries are jointly ranked based on coverage or clutter and split into 4 equal size groups. Queries with no positive in the corresponding group are discarded. No. of queries per group is in parentheses. $\\dagger$ indicates results with the linear adaptation." + ], + "image_footnote": [], + "bbox": [ + 516, + 284, + 903, + 441 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of clutter and scale in positives. To quantify the impact of background clutter and scale changes, Fig. 7 presents the performance for different groups of positives. Dealing with small objects and multi-object scenes form major weaknesses of existing models. Notably, t2i beats i2i without adaptation in small-scale groups.", + "bbox": [ + 511, + 551, + 906, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusions", + "text_level": 1, + "bbox": [ + 513, + 661, + 640, + 678 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce ILIAS and conduct an extensive evaluation of current foundational models and retrieval methods, highlighting that instance-level retrieval remains an unsolved problem. Our results indicate that off-the-shelf application of foundational models leaves considerable room for improvement, particularly in handling small objects and complex backgrounds. While specialized retrieval methods leveraging local descriptors are effective in these cases, their high memory and computational costs become impractical at the scale of ILIAS or beyond. ILIAS is designed to become a standard benchmark for evaluating foundational representation models and retrieval methods, accommodating both global and local representations, and advancing the field of instance-level recognition.", + "bbox": [ + 511, + 688, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14784", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Funding acknowledgement. This work was supported by the Junior Star GACR (grant no. GM 21-28830M), Horizon MSCA-PF (grant no. 101154126), Programme Johannes Amos Comenius (grant no. CZ.02.01.01/00/22.010/0003405), CTU in Prague (grant no. SGS23/173/OHK3/3T/13), the CTU institutional support (Future fund). JM was supported by Ministry of the Interior of the Czech Republic (project no. VJ02010041). We acknowledge VSB - Technical University of Ostrava, IT4Innovations National Supercomputing Center, Czech Republic, for awarding this project access to the LUMI supercomputer, owned by the EuroHPC Joint Undertaking, hosted by CSC (Finland) and the LUMI consortium through the Ministry of Education, Youth and Sports of the Czech Republic through the e-INFRA CZ (grant ID: 90254).", + "bbox": [ + 89, + 114, + 485, + 343 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Contributors acknowledgments. We want to thank Larysa Ivashechkina for her work on the annotation of object bounding boxes and masks and the initial filtering of text queries. We also want to thank all contributors for the collection of ILIAS dataset. Additionally to all authors, the list of external contributor in alphabetical order: Aggeliki Tserota, Anna Nisyraiou, Celeste Abreu, Charalambos Tzamos, Christina Tserota, Dimitris Karageorgiou, Dmytro Mishkin, Eleni Karantali, Eleni Papadopoulou, Eva Tsiliakou, Kelly Kordopati-Zilou, Markos Zampoglou, Noa Garcia, Panagiotis Tassis, Paraskevas Kordopatis, Pavlos Alexantonakis, Ruslan Rozumnyi, Snežana Čurguz, Tereza Nejedlá, Tomáš Jelinek, Yannis Kalantidis, Vasilis Alexiadis, Yankun Wu.", + "bbox": [ + 89, + 349, + 483, + 561 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 575, + 187, + 590 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Xiang An, Jiankang Deng, Kaicheng Yang, Jaiwei Li, Ziyong Feng, Jia Guo, Jing Yang, and Tongliang Liu. Unicom: Universal and compact representation learning for image retrieval. In ICLR, 2023. 2, 5, 7", + "[2] Relja Arandjelović and Andrew Zisserman. Smooth object retrieval using a bag of boundaries. In ICCV, 2011. 2, 3", + "[3] Harry G Barrow, Jay M Tenenbaum, Robert C Bolles, and Helen C Wolf. Parametric correspondence and chamfer matching: Two new techniques for image matching. In *IJ-CAI*, 1977. 5", + "[4] Wissam Bejjani, Wisdom C Agboh, Mehmet R Dogar, and Matteo Leonetti. Occlusion-aware search for object retrieval in clutter. In IROS, 2021. 1", + "[5] Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselat, Emma Brunskill, et al. On the opportunities and risks of foundation models. In arXiv, 2021. 1, 5", + "[6] Bingyi Cao, André Araujo, and Jack Sim. Unifying deep local and global features for image search. In ECCV, 2020. 5" + ], + "bbox": [ + 99, + 601, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[7] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, 2020. 5", + "[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 5, 7", + "[9] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In CVPR, 2023. 5, 7", + "[10] Ondrej Chum, James Philbin, Josef Sivic, Michael Isard, and Andrew Zisserman. Total recall: Automatic query expansion with a generative feature model for object retrieval. In ICCV, 2007. 2, 5, 7", + "[11] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. In ICLR, 2024. 6", + "[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 2, 5", + "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICML, 2021. 5, 7", + "[14] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. EVA: Exploring the limits of masked visual representation learning at scale. In CVPR, 2023. 2, 5, 7", + "[15] Martin A. Fischler and Robert C. Bolles. Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 1981. 5", + "[16] Christopher Flagg and Ophir Frieder. Reconstruction of artifacts from digital image repositories. JOCCH, 2022. 1", + "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 5, 7", + "[18] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. 5", + "[19] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. 7", + "[20] Hervé Jégou, Matthijs Douze, and Cordelia Schmid. Hamming embedding and weak geometric consistency for large scale image search. In ECCV, 2008. 2, 3", + "[21] Tomas Jenicek and Ondrej Chum. No fear of the dark: Image retrieval under varying illumination conditions. In CVPR, 2019. 1" + ], + "bbox": [ + 514, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14785", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[22] Sungyeon Kim, Boseung Jeong, and Suha Kwak. HIER: Metric learning beyond class labels via hierarchical regularization. In CVPR, 2023. 5", + "[23] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In CVPR, 2023. 4", + "[24] Byungsoo Ko, Minchul Shin, Geonmo Gu, HeeJae Jun, Tae Kwan Lee, and Youngjoon Kim. A benchmark on tricks for large-scale image retrieval. In arXiv, 2019. 2", + "[25] Tarun Krishna, Kevin McGuinness, and Noel O'Connor. Evaluating contrastive models for instance-based image retrieval. In ICMR, 2021. 2", + "[26] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In NeurIPS, 2012. 5", + "[27] Seongwon Lee, Hongje Seong, Suhyeon Lee, and Euntai Kim. Correlation verification for image retrieval. In CVPR, 2022. 5", + "[28] Ziwei Liu, Ping Luo, Shi Qiu, Xiaogang Wang, and Xiaou Tang. Deepfashion: Powering robust clothes recognition and retrieval with rich annotations. In CVPR, 2016. 2, 3, 4", + "[29] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. CVPR, 2022. 5, 7", + "[30] Shiyang Lu, Haonan Chang, Eric Pu Jing, Abdeslam Boularias, and Kostas Bekris. OVIR-3D: Open-vocabulary 3d instance retrieval without training on 3d data. In CoRL, 2023. 1", + "[31] Claudio Michaelis, Matthias Bethge, and Alexander Ecker. One-shot segmentation in clutter. In ICML, 2018. 1", + "[32] David Nister and Henrik Stewenius. Scalable recognition with a vocabulary tree. In CVPR, 2006. 2, 3", + "[33] Hyeonwoo Noh, Andre Araujo, Jack Sim, Tobias Weyand, and Bohyung Han. Large-scale image retrieval with attentive deep local features. In ICCV, 2017. 2, 3", + "[34] Maxime Oquab, Timothee Darct, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINOv2: Learning robust visual features without supervision. TMLR, 2024. 1, 2, 5, 6, 7", + "[35] Yash Patel, Giorgos Tolias, and Jiří Matas. Recall@k surrogate loss with large batches and similarity mixup. In CVPR, 2022. 5", + "[36] Jingtian Peng, Chang Xiao, and Yifan Li. RP2K: A large-scale retail product dataset for fine-grained image classification. In arXiv, 2020. 2, 3", + "[37] James Philbin, Ondrej Chum, Michael Isard, Josef Sivic, and Andrew Zisserman. Object retrieval with large vocabularies and fast spatial matching. In CVPR, 2007. 2, 4, 5, 7", + "[38] James Philbin, Ondrej Chum, Michael Isard, Josef Sivic, and Andrew Zisserman. Lost in quantization: Improving particular object retrieval in large scale image databases. In CVPR, 2008. 4" + ], + "bbox": [ + 91, + 92, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Filip Radenović, Ahmet Iscen, Giorgos Tolias, Yannis Avrithis, and Ondrej Chum. Revisiting oxford and paris: Large-scale image retrieval benchmarking. In CVPR, 2018. 1, 2, 3", + "[40] Filip Radenovic, Giorgos Tolias, and Ondrej Chum. Finetuning cnn image retrieval with no human annotation. PAMI, 2019. 2, 4, 5, 7", + "[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 3, 5, 7", + "[42] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, et al. SAM 2: Segment anything in images and videos. In ICLR, 2025. 4", + "[43] Ali Sharif Razavian, Hossein Azizpour, Josephine Sullivan, and Stefan Carlsson. CNN features off-the-shelf: An astounding baseline for recognition. In CVPRW, 2014. 2", + "[44] Ali S Razavian, Josephine Sullivan, Stefan Carlsson, and Atsuto Maki. Visual instance retrieval with deep convolutional networks. ITE Trans. on Media Technology and Applications, 2016. 5, 7", + "[45] Taichi Sakaguchi, Akira Taniguchi, Yoshinobu Hagiwara, Lotfi El Hafi, Shoichi Hasegawa, and Tadahiro Taniguchi. Object instance retrieval in assistive robotics: Leveraging fine-tuned simsiam with multi-view images based on 3d semantic map. In IROS, 2024. 1", + "[46] Mert Bulent Sariyildiz, Philippe Weinzaepfel, Thomas Lucas, Diane Larlus, and Yannis Kalantidis. UNIC: Universal classification models via multi-teacher distillation. In ECCV, 2024. 5", + "[47] Konstantin Schall, Kai Uwe Barthel, Nico Hezel, and Klaus Jung. GPR1200: a benchmark for general-purpose content-based image retrieval. In MMM, 2022. 2, 3", + "[48] Shihao Shao, Kaifeng Chen, Arjun Karpur, Qinghua Cui, André Araujo, and Bingyi Cao. Global features are all you need for image retrieval and reranking. In ICCV, 2023. 2, 5", + "[49] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In ICLR, 2015. 5", + "[50] Ivan Sipiran, Patrick Lazo, Cristian Lopez, Milagritos Jimenez, Nihar Bagewadi, Benjamin Bustos, Hieu Dao, Shankar Gangisetty, Martin Hanik, Ngoc-Phuong Ho-Thi, et al. SHREC 2021: Retrieval of cultural heritage objects. Computers & Graphics, 2021. 1", + "[51] Arnold WM Smeulders, Marcel Worring, Simone Santini, Amarnath Gupta, and Ramesh Jain. Content-based image retrieval at the end of the early years. PAMI, 2000. 2", + "[52] Chull Hwan Song, Jooyoung Yoon, Taebaek Hwang, Shunghyun Choi, Yeong Hyeon Gu, and Yannis Avrithis. On train-test class overlap and detection for image retrieval. In CVPR, 2024. 1", + "[53] Hyun Oh Song, Yu Xiang, Stefanie Jegelka, and Silvio Savarese. Deep metric learning via lifted structured feature embedding. In CVPR, 2016. 1, 2, 3, 4" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14786", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[54] Andreas Steiner, Alexander Kolesnikov, Xiaohua Zhai, Ross Wightman, Jakob Uszkoreit, and Lucas Beyer. How to train your vit? data, augmentation, and regularization in vision transformers. TMLR, 2021. 5, 7", + "[55] Pavel Suma and Giorgos Tolias. Large-to-small image resolution asymmetry in deep metric learning. In WACV, 2023. 5", + "[56] Pavel Suma, Giorgos Kordopatis-Zilos, Ahmet Iscen, and Giorgos Tolias. AMES: Asymmetric and memory-efficient similarity estimation for instance-level retrieval. In ECCV, 2024. 2, 6, 7", + "[57] Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. EVA-CLIP: Improved training techniques for clip at scale. In arXiv, 2023. 2, 5, 7", + "[58] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In CVPR, 2015. 5", + "[59] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In ICML, 2019. 5", + "[60] Ryan Theisen, Hyunsuk Kim, Yaoqing Yang, Liam Hodgkinson, and Michael W Mahoney. When are ensembles really effective? In NeurIPS, 2024. 8", + "[61] Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. YFCC100M: The new data in multimedia research. Communications of the ACM, 2016. 3", + "[62] Akihiko Torii, Relja Arandjelovic, Josef Sivic, Masatoshi Okutomi, and Tomas Pajdla. 24/7 place recognition by view synthesis. In CVPR, 2015. 1", + "[63] Hugo Touvron, Andrea Vedaldi, Matthijs Douze, and Herve Jegou. Fixing the train-test resolution discrepancy. In NeurIPS, 2019. 5", + "[64] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, 2021. 5", + "[65] Michael Tschannen, Alexey Gritsenko, Xiao Wang, Muhammad Ferjad Naeem, Ibrahim Alabdulmohsin, Nikhil Parthasarathy, Talfan Evans, Lucas Beyer, Ye Xia, Basil Mustafa, et al. SigLIP 2: Multilingual vision-language encoders with improved semantic understanding, localization, and dense features. In arXiv, 2025. 5, 7", + "[66] Shuang Wang and Shuqiang Jiang. INSTRE: A new benchmark for instance-level object retrieval and recognition. TOMM, 2015. 1, 2, 3", + "[67] Tobias Weyand, André Araujo, Bingyi Cao, and Jack Sim. Google Landmarks Dataset v2 - A large-scale benchmark for instance-level recognition and retrieval. In CVPR, 2020. 1, 2, 3, 4", + "[68] Ross Wightman, Hugo Touvron, and Hervé Jégou. Resnet strikes back: An improved training procedure in timm. In NeurIPSW, 2021. 5", + "[69] Pengxiang Wu, Siman Wang, Kevin Dela Rosa, and Derek Hu. FORB: A flat object retrieval benchmark for universal image embedding. In NeurIPS, 2024. 3" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[70] Hu Xu, Saining Xie, Xiaqing Ellen Tan, Po-Yao Huang, Russell Howes, Vasu Sharma, Shang-Wen Li, Gargi Ghosh, Luke Zettlemoyer, and Christoph Feichtenhofer. Demystifying clip data. In ICLR, 2024. 3, 5, 7", + "[71] Nikolaos-Antonios Ypsilantis, Noa Garcia, Guangxing Han, Sarah Ibrahimi, Nanne Van Noord, and Giorgos Tolias. The MET dataset: Instance-level recognition for artworks. In NeurIPS, 2021. 3", + "[72] Nikolaos-Antonios Ypsilantis, Kaifeng Chen, Bingyi Cao, Mario Lipovsky, Pelin Dogan-Schönberger, Grzegorz Makosa, Boris Bluntschli, Mojtaba Seyedhosseini, Ondrej Chum, and André Araujo. Towards universal image embeddings: A large-scale dataset and challenge for generic image representations. In ICCV, 2023. 1, 2, 3, 5", + "[73] Nikolaos-Antonios Ypsilantis, Kaifeng Chen, Andre Araujo, and Ondrej Chum. UDON: Universal dynamic online distillation for generic image representations. In NeurIPS, 2024. 5, 7", + "[74] Jiangbo Yuan, An-Ti Chiang, Wen Tang, and Antonio Haro. eProduct: A million-scale visual search benchmark to address product recognition challenges. In CVPRW, 2021. 1, 2, 3", + "[75] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In ICCV, 2023. 1, 2, 3, 4, 5, 7", + "[76] Xunlin Zhan, Yangxin Wu, Xiao Dong, Yunchao Wei, Minlong Lu, Yichi Zhang, Hang Xu, and Xiaodan Liang. Product: Towards weakly supervised instance-level product retrieval via cross-modal pretraining. In ICCV, 2021. 1, 2, 3", + "[77] Liang Zheng, Yi Yang, and Qi Tian. Sift meets cnn: A decade survey of instance retrieval. PAMI, 2017. 2", + "[78] Barret Zoph, Vijay Vasudevan, Jonathon Shlens, and Quoc V Le. Learning transferable architectures for scalable image recognition. In CVPR, 2018. 5" + ], + "bbox": [ + 516, + 92, + 903, + 571 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "14787", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/ILIAS_ Instance-Level Image retrieval At Scale/bd262cec-535c-47b6-b08f-5430c3f91bd5_model.json b/2025/ILIAS_ Instance-Level Image retrieval At Scale/bd262cec-535c-47b6-b08f-5430c3f91bd5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..373161ae3b1f1515cda269bae63768a599bd72e1 --- /dev/null +++ b/2025/ILIAS_ Instance-Level Image retrieval At Scale/bd262cec-535c-47b6-b08f-5430c3f91bd5_model.json @@ -0,0 +1,2257 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.261, + 0.131, + 0.741, + 0.154 + ], + "angle": 0, + "content": "ILIAS: Instance-Level Image retrieval At Scale" + }, + { + "type": "text", + "bbox": [ + 0.204, + 0.18, + 0.794, + 0.235 + ], + "angle": 0, + "content": "Giorgos Kordopatis-Zilos Vladan Stojnic Anna Manko Pavel Šuma Nikolaos-Antonios Ypsilantis Nikos Efthymiadis Zakaria Laskar Jiří Matas Ondřej Chum Giorgos Tolias" + }, + { + "type": "text", + "bbox": [ + 0.298, + 0.252, + 0.696, + 0.271 + ], + "angle": 0, + "content": "VRG, FEE, Czech Technical University in Prague" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.296, + 0.327, + 0.312 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.329, + 0.486, + 0.691 + ], + "angle": 0, + "content": "This work introduces ILIAS, a new test dataset for Instance-Level Image retrieval At Scale. It is designed to evaluate the ability of current and future foundation models and retrieval techniques to recognize particular objects. The key benefits over existing datasets include large scale, domain diversity, accurate ground truth, and a performance that is far from saturated. ILIAS includes query and positive images for 1,000 object instances, manually collected to capture challenging conditions and diverse domains. Large-scale retrieval is conducted against 100 million distractor images from YFCC100M. To avoid false negatives without extra annotation effort, we include only query objects confirmed to have emerged after 2014, i.e. the compilation date of YFCC100M. An extensive benchmarking is performed with the following observations: i) models fine-tuned on specific domains, such as landmarks or products, excel in that domain but fail on ILIAS ii) learning a linear adaptation layer using multi-domain class supervision results in performance improvements, especially for vision-language models iii) local descriptors in retrieval re-ranking are still a key ingredient, especially in the presence of severe background clutter iv) the text-to-image performance of the vision-language foundation models is surprisingly close to the corresponding image-to-image case." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.692, + 0.451, + 0.707 + ], + "angle": 0, + "content": "website: https://vrg.fel(cvut.cz/ilias/" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.723, + 0.222, + 0.739 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "The ability to recognize and differentiate every unique object instance in the physical world represents one of the ultimate goals for foundation representation models [5, 34, 41, 75]. This work aims to assess this capability through the lens of instance-level image retrieval at a very large scale. Instance-level image retrieval corresponds to searching for images of particular objects within large collections. All images of a particular object form their own instance-level class. This is an important information retrieval task due to its numerous real-world applications in robotics [30, 45], e" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.294, + 0.907, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.477, + 0.907, + 0.561 + ], + "angle": 0, + "content": "Figure 1. Performance timeline on ILIAS. Curves indicate best performance in chronological order for image-to-image and text-to-image retrieval, showing a significant boost with the release of foundation models. Representations are linearly adapted via multi-domain learning on UnED [72]. Re-ranking with local descriptors achieves the best results by a significant margin." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.568, + 0.909, + 0.735 + ], + "angle": 0, + "content": "commerce [74, 76], and cultural heritage [16, 50], to name just a few. The task faces challenges because of the substantial variations among positive examples, such as illumination/viewpoint [21, 62] changes and background clutter [4, 31]. An additional difficulty is the high similarity among negatives, which is driven by the extremely fine granularity in the class definitions. It becomes even more challenging at a real-world scale, where searching through millions or even billions of images requires handling an open-world setup with countless unseen objects spanning diverse and complex domains." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.735, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Benchmarking instance-level retrieval under real-world challenges is currently limited by the lack of suitable datasets. Constructing a dataset with instance-level class definitions necessitates huge development effort, reflected by the many shortcomings of existing datasets. Shortcomings exist in several key aspects, such as dataset size [66], domain diversity [39, 53], and ground-truth accuracy [67], which suffers from both false positives and false negatives. Popular datasets are typically limited to landmarks [39], and as dataset scale increases, ground-truth quality tends to decline [52, 67]. This is a consequence of automating the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14777" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.137 + ], + "angle": 0, + "content": "ground-truth creation process to facilitate scaling up. To address such limitations, we introduce the Instance-Level Image retrieval At Scale (ILIAS) evaluation dataset." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.139, + 0.483, + 0.319 + ], + "angle": 0, + "content": "The creation of our dataset has two key elements. First, query and positive images are manually captured to ensure challenging variations, covering 1,000 objects across diverse domains. Second, to expand the dataset size without ground-truth errors or additional annotation effort, we leverage a key technique: distractor images, collected in 2014 from YFCC100M, are combined with query objects verified not to have publicly existed until after 2014. This distractor set includes 100 million images, two orders of magnitude larger than the largest existing dataset [39]. Notably, all images have a permissive license, allowing us to ensure long-term online availability to the full extent." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.321, + 0.483, + 0.457 + ], + "angle": 0, + "content": "ILIAS includes both image and text queries. The latter is in the form of detailed descriptions of objects and their distinctive features. The dataset is designed to support future research in image-to-image and text-to-image retrieval for particular objects, and additionally serves as a large-scale benchmark for evaluating representations of foundation vision and language models (VLM) [41, 75]. To facilitate faster experimentation, we provide a mini, but challenging, version (5M) of the distractor set." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.459, + 0.483, + 0.566 + ], + "angle": 0, + "content": "We perform an extensive evaluation comparison, including many foundation image-to-image and text-to-image models, and establish a comprehensive testbed that enables future comparisons. The provided evaluation includes retrieval with global image representation but also re-ranking techniques that use local representations [37, 43, 56] and query expansion [10, 40]. We observe the following:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.567, + 0.483, + 0.642 + ], + "angle": 0, + "content": "- Performance of standard 10-year-old models, measured by mean Average Precision, is as low as \\(1.3\\%\\), while the best-performing model achieves \\(31.3\\%\\), as shown in Fig. 1. This points out the vast progress of representation models and the high challenging factors of ILIAS." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.642, + 0.363, + 0.657 + ], + "angle": 0, + "content": "- VLMs are the top-performing models." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.657, + 0.483, + 0.702 + ], + "angle": 0, + "content": "- Smaller (ViT-B) models trained/tested on large resolution (512/724) outperform larger models (ViT-L) trained/tested on small resolution (256/384)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.702, + 0.483, + 0.793 + ], + "angle": 0, + "content": "- Using Universal Embedding Dataset (UnED) [72] to learn a linear adaptation layer on top of frozen models improves performance of most models, making it a candidate training set to couple with ILIAS. Notably, VLMs demonstrate the largest benefits, presumably because their training stage does not optimize image-to-image relations." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.794, + 0.483, + 0.838 + ], + "angle": 0, + "content": "- In contrast to the current belief [48], local representation is a key ingredient, while global representation, despite being efficient and compact, performs much lower." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.838, + 0.483, + 0.899 + ], + "angle": 0, + "content": "- The performance gap between image-to-image and text-to-image models is surprisingly small. Therefore, detailed text queries are a reasonable proxy in the absence of image queries, even at the instance level." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.567, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.652, + 0.106 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.116, + 0.905, + 0.145 + ], + "angle": 0, + "content": "In this section, we review the related work in terms of existing datasets and benchmarks in the literature." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.146, + 0.907, + 0.629 + ], + "angle": 0, + "content": "Datasets. Tab. 1 presents the datasets from the image retrieval literature related to ILIAS. The datasets can be compared based on five main axes: (i) Class definition adopted. Many datasets [20, 32, 33, 39, 67] adopt a strict definition very similar to ours, satisfying instance-level requirements. Others [2, 53, 66, 76] adopt a more relaxed definition, where some minor variations are permitted, e.g. color changes in objects of the same class. Even more relaxed are the fine-grained definitions [74], where the object of the very same type is considered related, e.g. same product with different variant. (ii) Domain of the dataset. Most datasets are tailored for a specific domain. Landmarks are among the most popular domains [2, 20, 33, 39, 67]. Other domains include products [32, 36, 76] and fashion [28, 53]. Some datasets cover multiple domains, either being standalone [66] or bundle of repurposed datasets [47, 72]. (iii) Scale of database. Most of the datasets are small-scale, counting a few thousand images [28, 66, 76]. Larger ones [39, 72, 74] expend slightly above a million. None satisfies large-scale requirements. (iv) Noise in ground truth. Most datasets consist of clean annotations, except for a few cases that contain inaccuracies, including false positives [67], i.e. images wrongly annotated as relevant, false negatives [32, 39], i.e. relevant images that have not been annotated as positives, or the possibility of false positives [39]. (v) Availability. Most datasets are publicly available with permissive licenses, with few exceptions of partial [2, 20] or no [33, 74] availability. To this end, no publicly available dataset fits the strict instance-level definition, contains objects from multiple domains, ensures error-free labeling and is large scale. This gap is filled with ILIAS satisfying all the aforementioned requirements." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.63, + 0.906, + 0.795 + ], + "angle": 0, + "content": "Evaluation benchmarks. Benchmarking [51] tracks the progress in the field, which is even more necessary with the emergence of foundation models. Several benchmarks papers [24, 25, 77] exist in the instance-level retrieval literature, investigating the impact of learning scheme, post-processing, model ensembling, query expansion, and whitening. The most relevant benchmark to ILIAS is UnED [72] that combines existing datasets to create a union that evaluates models performance across various domains. Due to its wide variety, UnED serves as the training dataset for linear adaptation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Regarding the evaluation of foundation models, the most common practice [14, 34, 57] is measuring classification performance on top of frozen models on ImageNet [12]. This is performed either with or without the training of a classifier via linear probing or k-NN search. Furthermore, models are usually evaluated on dense prediction tasks [34] and several multiple-downstream single-domain tasks [1]." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14778" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.132, + 0.077, + 0.862, + 0.311 + ], + "angle": 0, + "content": "
datasetsyearobjectsquerypositivesdatabasegtclass def.domainbboxonlinelicense
UKB [32]20062.5K10K10K10KFNILproductXFullyN/A
Holidays [20]20085005009911MCleanILlandmarkXPartiallyCC
Sculptures [2]201110703.1K3.1KCleanPartial ILsculptureXPartiallyFlickr TC
INSTRE [66]2015200125027.3K27.3KCleanPartial ILmulti✓FullyFlickr TC
SOP [53]201511.3K60.5K60.5K60.5KCleanPartial ILproductXFullyMIT License
InShop [28]20163.9K14.2K12.6K12.6KCleanPartial ILfashion✓FullyN/A
R-Oxford [39]201811705K1MFN?ILlandmarkXFullyFlickr TC, CC
R-Paris [39]201811706.3K1MFN?ILlandmarkXFullyFlickr TC, CC
GLDv1 [33]201830KN/AN/A1.1MCleanILlandmarkXPartiallyMultiple
GLDv2 [67]20203181.1K3.1K762KFPILlandmarkXFullyCC/ Public-domain
Product1M [76]20213926.5K40K40KCleanPartial ILproductXPartiallyN/A
RP2K [36]20211.2K10.9K10.9K10.9KCleanPartial ILproduct✓FullyN/A
GPR1200 [47]20211.2K12K12K12KMixIL+FGmultiXFullyMultiple
eProduct [74]202120610KN/A1.1MN/AFGproductXNoN/A
FORB [69]2023N/A13.9K4.5K49.8KCleanILplanarXFullySnap Inc.
UnED [72]202321K241K244K1.4MMixIL+FGmultiXFullyMultiple
ILIAS20251,0001,2324,715100MCleanILmulti✓FullyCC
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.315, + 0.907, + 0.371 + ], + "angle": 0, + "content": "Table 1. Comparison with other instance-level datasets. Datasets are compared based on their size (object, query, positives, database), the accuracy of the ground truth (gt), type of class definition, domain, supplementary annotations (bbox) and accessibility (online, license). N/A: not available. FP/FN: false positives/negatives. FN?: possibility of false negatives. Mix: combination of clean and noisy datasets. IL: instance-level. FG: fine-grained. Partial IL: instance-level with subtle variations among same class objects. CC: Creative Commons." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.38, + 0.482, + 0.485 + ], + "angle": 0, + "content": "For VLMs, zero-shot classification and retrieval serve as the primary benchmarks [41, 70, 75], utilizing class text labels. In this work, we provide similar evaluation protocols tailored for instance-level retrieval. One can test the raw model capabilities or adapt for the instance-level task via linear adaptation on UnED. Text-to-image retrieval is also facilitated for the evaluation of VLMs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.503, + 0.236, + 0.518 + ], + "angle": 0, + "content": "3. ILIAS dataset" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.529, + 0.336, + 0.545 + ], + "angle": 0, + "content": "3.1. Composition and collection" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.552, + 0.483, + 0.687 + ], + "angle": 0, + "content": "Instance-level class definition. Following an instance-level class definition [39, 71], we consider all indistinguishable object instances of the real world to form their own class. Nevertheless, we add a restriction to consider a pair of images as relevant to each other only if there is a view overlap. Other cases are explicitly not included in the dataset, contrasting the existing work [28, 53, 67]. Therefore, models should mostly rely on estimating the visual similarity and less on shortcuts through semantics." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.688, + 0.483, + 0.81 + ], + "angle": 0, + "content": "Overview. ILIAS supports both image-to-image (i2i) and text-to-image (t2i) retrieval and follows the standard setup for retrieval datasets, consisting of two main parts: (i) query images and text, and (ii) database (db) images. The objective is to rank positives - db images relevant to the query - at the top ranks. The collected objects cover a wide range of categories and are not restricted to specific domains. An overview of some collected objects is provided in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Queries and positives are created/collected by a group of collectors that are well-informed about the task objectives. In addition to positives, in the database, we include numerous distractors - irrelevant (negative) images to the queries - that make retrieval more challenging. Following previous work [39], adding a large, uncurated set of ran" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.38, + 0.907, + 0.456 + ], + "angle": 0, + "content": "dom images achieves this. The larger the set, the higher the chances of hard negatives – images with similar appearance or semantics to the queries. To this end, we select the YFCC100M [61] dataset to serve as the source of distractors due to its size and permissive license." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.459, + 0.909, + 0.672 + ], + "angle": 0, + "content": "Selected objects. Ensuring that distractor images include no false negatives cannot be performed in a scalable or accurate way if one relies on human annotation or metadata. Instead, we take advantage of the fact that YFCC100M was crawled from Flickr in 2014. Hence, an object qualifies in ILIAS if it could not have appeared on Flickr before 2014. To verify this, we rely either on publicly available information, e.g. objects known to be created/manufactured after 2014, or on the collector's knowledge about the object not being publicly available. Additionally, we opt for objects with distinctive and unique features that set them apart from others within the same category. For example, we avoid recent smartphones that look like plain black screens or new objects with distinctive parts closely resembling older ones." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Queries and positives. Query images depict the instance on a clean or uniform background. When this is not feasible (e.g. buildings or statues), background blurring or cropping is applied. This is performed to avoid including background objects in the query that do not have corresponding positives in our ground truth information. Positives are images featuring the query object in challenging conditions, such as clutter, scale changes, occlusions, and partial views. Prior work [39] reveals that easy positives dominate performance metrics. Thus, we specifically opt for challenging cases that cannot be easily retrieved by the models. To avoid taking advantage of camera identification, most query and positive images are captured with at least two different camera models to introduce diversity. We also incorporate older camera models that are used in YFCC100M." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14779" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.153, + 0.088, + 0.837, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.322, + 0.907, + 0.351 + ], + "angle": 0, + "content": "Figure 2. Examples of query, positive and hard negatives within the distractor set. Average Precision per query and rank of the negatives and positives is reported using SigLIP [75] model. Gray: queries. Green: positives. Red: distractors." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.356, + 0.483, + 0.446 + ], + "angle": 0, + "content": "Each text query consists of a detailed and fine-grained textual description of an object. Descriptions are initially created by a large language model prompted to provide highly detailed depictions of the object shown in query images. Generated descriptions are manually edited to fix errors, insufficient descriptions, or nuances of the model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.446, + 0.484, + 0.522 + ], + "angle": 0, + "content": "Distractors. The YFCC100M dataset was chosen for the distractor set due to its large scale and diverse range of concepts. It consists of 100 million Flickr images, collected without specific filtering, aside from being shared under a permissive CC-BY license." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.522, + 0.484, + 0.612 + ], + "angle": 0, + "content": "Bounding box annotation. We include supplementary bounding boxes that specify the precise location of objects in query and positive images. They provide statistics about the position and size of object areas, assist our analysis of the dataset challenges, and support future research in instance-level localization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.613, + 0.484, + 0.72 + ], + "angle": 0, + "content": "Evaluation metric. Retrieval performance is evaluated via mean Average Precision (mAP), a widely used metric in instance-level image retrieval [37, 38, 40]. Specifically, we adopt mAP@1k [67], which assesses the ranking of the top-1k nearest neighbors for each query, treating any positive not ranked among the top-1k as not retrieved. We estimate the area under the curve using rectangles and not trapezoids." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.728, + 0.197, + 0.743 + ], + "angle": 0, + "content": "3.2. Statistics" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.856 + ], + "angle": 0, + "content": "Dataset size. The final ILIAS dataset includes 1,000 object instances captured in 5,947 images, of which 1,232 are queries and 4,715 are positives. Fig. 3a shows the distribution of positives per object. Also, 99,144,315 images from YFCC100M are downloaded. All images (queries, positives, distractors) are transferred through Flickr to ensure the same pre-processing." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Taxonomy. A hierarchical 3-level taxonomy is composed for ILIAS. All instances are assigned across one to three categories of different granularity levels. The taxonomy" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.356, + 0.907, + 0.598 + ], + "angle": 0, + "content": "consists of 8 categories on the coarser level, 42 on the mid level, and 38 on the finer level. The categories are derived through manual labeling of the objects based on their semantic content. To form the coarser-level categories, we use domain definitions borrowed from prior work [28, 53, 67] to align with the literature, i.e. art, landmarks, products, fashion. We also define novel categories based on the objects that do not fit into any existing domain. The distribution of objects across categories is uneven, e.g. ranging from 168 and 162 for art and landmarks to 83 for products. Each mid- and finer-level category contains at least 4 instances. Note that taxonomy is given to provide statistics about the domains of objects and assist our analysis instead of being leveraged as ground truth. The distribution of taxonomy categories can be inferred by Fig. 5, and a detailed figure is provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Bounding box analysis. A total number of 6,117 bounding boxes are annotated for both queries and positives. Note that positives may display multiple objects of near identical appearance to the query; in such cases, bounding boxes are drawn on all indistinguishable objects. There are 235 images with more than one bounding box. Based on the annotated bounding boxes, we compute the area covered in the image by the object instances to derive its relative scale. Fig. 3c shows the distribution of the scale ratio for queries and positives. Most objects in queries cover the largest area of the images; while in the vast majority of positives, the object covers a small area of less than half the image. It is a result of the severe scale changes and partial views. Moreover, we use the Segment Anything Model (SAM) [23, 42] to extract object segments from positives. The number of detected segments outside the query object's bounding boxes is computed. This indicates clutter from other items in the positives. Fig. 3b shows the segment number distribution, with most images containing multiple segments due to clutter." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14780" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.093, + 0.223, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.151, + 0.204, + 0.167, + 0.214 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.093, + 0.351, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.282, + 0.204, + 0.3, + 0.214 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.351, + 0.093, + 0.477, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.408, + 0.204, + 0.424, + 0.214 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.215, + 0.483, + 0.256 + ], + "angle": 0, + "content": "Figure 3. ILIAS statistics. (a) number of positives per object, (b) positive distribution by the SAM segments outside the bounding box, (c) image distribution by the relative bounding box area." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.264, + 0.212, + 0.278 + ], + "angle": 0, + "content": "3.3. mini-ILIAS" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.286, + 0.484, + 0.543 + ], + "angle": 0, + "content": "We provide a small version of ILIAS, called mini-ILIAS, to facilitate quick experimentations. It consists of the query and positive images collected for ILIAS, and a subset of the YFCC100M distractors. Instead of randomly subsampling YFCC100M, we construct a challenging subset with the help of VLMs. We aim at selecting distractors displaying objects of similar categories as the query objects. We use the text category labels of the taxonomy as text queries. We also extend them with standard templates used for zero-shot recognition [41], which resulted in several thousands. T2i similarity between each text query and each distractor image is estimated. A similarity score for each distractor is derived based on its maximum similarity over the text queries. We ensemble the scores of 3 VLMs to rank images. The top-5M ranked distractors compose the final mini-ILIAS. Our experiments indicate that this subset is significantly more challenging than a random subset of the same size." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.558, + 0.289, + 0.573 + ], + "angle": 0, + "content": "4. Benchmark methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.483, + 0.689 + ], + "angle": 0, + "content": "We describe the methods and foundation models we evaluate on ILIAS, which are grouped according to their type of representations used for retrieval in global (i2i) representations, re-ranking with global (i2i) representations, re-ranking with local (i2i) representations, and text-to-image. A detailed list of models, their performance, and implementation details are in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Image-to-image retrieval with global representations. Global representation methods use image encoders to map images to global descriptors and rank db images based on cosine similarity. We evaluate legacy and recent foundation models, varying in architecture, descriptor dimensionality, training scheme, training data, and input resolution. Foundation models [5] are the models trained with a training set on the scale of a hundred million. Particularly, 23 CNN [17, 26, 29, 49, 58, 59, 78] and 45 ViT [13, 14] models, trained with supervision [22, 54, 64, 68], self-supervision [7, 8, 18, 34], distillation [1, 46, 64], or visual-language alignment [9, 41, 57, 65, 70, 75] are benchmarked. Most of the non-foundation models are trained on ImageNet [12]. There are models trained on single specific do" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.152 + ], + "angle": 0, + "content": "mains [27, 35, 48], i.e. landmarks or products on GLDv2 or SOP. Universal models [1, 46, 72, 73] trained on multi-domains or multi-task schemes are included. The full list of models and results is provided in supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.154, + 0.907, + 0.274 + ], + "angle": 0, + "content": "To mitigate the differences in training resolution, we use three widely-used resolutions, i.e. 384, 512, and 724 and resize images so that their larger dimension matches one of the three. The test resolution is defined to be one resolution above the one used for training, e.g. a network trained with 224 or 384 is tested with 384 or 512, respectively. The vast majority of models achieve best performance under this rule. Similar behavior is observed in the literature [55, 63]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.276, + 0.906, + 0.488 + ], + "angle": 0, + "content": "Linear adaptation for i2i retrieval. Pre-trained foundation models, as well as legacy models, are trained to extract representations that are applicable to various tasks; not all encoded features are directly relevant to instance-level retrieval. To adapt the representation to the task at hand, we propose to train a single linear layer (projection) on top of frozen backbones. The recently introduced Universal Embeddings (UnED) dataset [72] is used for learning the linear adaptation. UnED contains images from 8 different domains with fine-grained and/or instance-level class annotation. In our experiments, the linear layer that converts the backbone output to a 512D descriptor is trained on a uniformly sampled subset of 1M images from UnED. The linear adaptation layer is trained with the UJCDs [72] method." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.49, + 0.906, + 0.58 + ], + "angle": 0, + "content": "Text-to-image retrieval. Text-to-image retrieval is performed using Vision-Language Models (VLMs) trained to align the two modalities. Retrieval is performed based on cosine similarity between the text query and db image descriptors that are extracted using the textual and visual encoder, respectively. We evaluate 17 VLM models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.582, + 0.906, + 0.732 + ], + "angle": 0, + "content": "Re-ranking with global representations. Such methods rely on global descriptors for exhaustive search during the initial ranking, but also for a second refinement stage that issues a new query. We experiment with \\(\\alpha\\)QE [40], the adaptive variant of average Query Expansion [10]. After the initial ranking, the descriptors of the top-ranked images are aggregated with the query via weighted average pooling. The weights are derived from the similarity to the query in the power of \\(\\alpha\\). We don't have a validation set; hence, we use a fixed value \\(\\alpha = 1\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Re-ranking with local representations. These re-ranking methods rely on global descriptors for exhaustive search during the initial ranking but estimate query-to-db image similarity based on local descriptors for a second refinement stage of the ranked list of images. We experiment with three methods: (i) Chamfer Similarity (CS) [3, 44] on the similarity matrix between local descriptors across the image pair. We use the asymmetric variant of CS with max over db descriptors and sum over query descriptors. (ii) Spatial Verification (SP) [6, 15, 37], a common re-ranking method where point correspondences are processed with a RANSAC-like" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "14781" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.187, + 0.074, + 0.408, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.075, + 0.614, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.075, + 0.828, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.234, + 0.907, + 0.278 + ], + "angle": 0, + "content": "Figure 4. Comparison with other instance-level retrieval datasets via reporting mAP@1k. INSTRE: 27.3K db size, multi-domain. GLDv2: 762K db size, single-domain. SOP: 60.5K db size, single-domain. Different network types are color-coded. For GLDv2 and SOP, models fine-tuned on these domains with the corresponding training sets are highlighted. No linear adaptation is used." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.288, + 0.486, + 0.424 + ], + "angle": 0, + "content": "process and the number of inliers is used for re-ranking. (iii) AMES [56], a recent transformer-based network to estimate the similarity between sets of local descriptors. Due to the scale of ILIAS database, we use only 100 binary local descriptors for each database image and 600 for the query image. Local descriptors are extracted using the base variant of DINOv2 with registers [11, 34] and selected based the local descriptor detector used in AMES [56]. Top-1k retrieved images are re-ranked." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.436, + 0.226, + 0.454 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.461, + 0.484, + 0.598 + ], + "angle": 0, + "content": "We evaluate all the above models and methods, extracting useful insights regarding the factors that boost retrieval performance. ILIAS is compared with other existing datasets for instance-level image retrieval. We analyze the performance of selected models1 to break down the impact of different ILIAS attributes, such as domains, clutter, and scale. Unless stated otherwise, we use the large model variants with the largest resolution available, e.g. in our analysis we use SigLIP ViT-L trained with 384 resolution." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.605, + 0.481, + 0.621 + ], + "angle": 0, + "content": "5.1. Comparison with other instance-level datasets" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.627, + 0.483, + 0.868 + ], + "angle": 0, + "content": "In Fig. 4, ILIAS is compared with other instance-level retrieval datasets based on evaluation of the same models. Linear adaptation is not applied as parts of GLDv2 and SOP are included in the UnED dataset. Only for the sake of this comparison, and for no other experiment in this work, we use models fine-tuned on specific domains (in-domain models), i.e. on the training sets of SOP and GLDv2. INSTRE, which is also multi-domain, shows a correlation to ILIAS, but its performance is saturated due to its small size. For single-domain datasets, in-domain models outperform others by a large margin, with few exceptions, i.e. DINOv2, which includes the trainset of GLDv2 in its training data. Several multi-domain models perform well on SOP since their training set is usually included in the training data. However, in-domain and multi-domain models face challenges on ILIAS, highlighting the diversity of our dataset." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.288, + 0.707, + 0.304 + ], + "angle": 0, + "content": "5.2. Method comparison" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.31, + 0.907, + 0.673 + ], + "angle": 0, + "content": "Image-to-image retrieval with global representations. Tab. 2 presents the performance of global descriptor models on ILIAS. Selected models are presented to highlight useful comparisons, while many other models are included in the supplementary material. The main factors that improve performance are the size of the training set, training resolution, and model architecture, which aligns with the literature. The impact of dataset size is apparent in various model combinations, e.g. CLIP with openai and laion2b. This is also pronounced by the dominance of foundation models. Training with large resolution brings significant gains and consistently improves mAP@1k. In some cases of SigLIP, smaller models trained with large resolutions outperform larger ones trained with small resolutions. For the models of the same resolution, it is a common trend for larger model variants to bring corresponding performance gains. In general, VLMs perform the best. From non-VLMs, only DINOv2 and Unicom achieve competitive performance. Masked Image Modeling (MIM) and supervised models are not performing well. Our linear adaptation scheme is very effective, improving most models. The boost is more pronounced in the case of VLMs. A possible explanation for such improvements is that image-to-image relations are not optimized during the training of VLMs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.673, + 0.909, + 0.809 + ], + "angle": 0, + "content": "Text-to-image retrieval. Following results in Tab. 2, similar conclusions are derived for the t2i case. Retrieval performance improves with the scaling of the training data. The larger model achieves significantly better results, i.e. compare the base with large variants. Finally, it is noteworthy that the best performance achieved by SigLIP2 is very close to the i2i performance when no adaptation is used. Note that t2i includes 1k text queries in total, with one query per object, while i2i 1,232 image queries." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Evaluation of mini-ILIAS selection. Tab. 3 shows performance on mini-ILIAS for five models with linear adaptation. The selected subset is significantly more challenging than a random selection of 5M images. More precisely, a set of approximately \\(\\sim 26\\mathrm{M}\\) random images matches the performance of mini-ILIAS." + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.875, + 0.483, + 0.901 + ], + "angle": 0, + "content": "1 Although the very recent SigLIP2 is the best-performing model, we conduct most experiments with SigLIP." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14782" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.173, + 0.076, + 0.832, + 0.459 + ], + "angle": 0, + "content": "
modelarchtraindatasetdata sizetrain restest resimage-to-imagetext-to-image
5M†100M†100M100M5M
ResNet50 [17]R50supin1k1M2243842.51.81.7--
DINO [8]R50sslin1k1M2243844.12.92.9--
ConvNext [29]CN-Lsupin1k1M2883844.22.92.2--
OAI-CLIP [17]R50vlaopenai400M2243848.56.03.21.52.3
OpenCLIP [19, 29]CN-Bvlalaion2b2B25638418.114.07.94.67.0
OpenCLIP [19, 29]CN-Lvlalaion2b2B32051222.918.39.68.111.5
ViT [13, 54]ViT-Bsupin1k1M2243841.91.31.0--
EVA-MIM [14]ViT-Bsslin22k142M2243844.73.22.1--
ViT [13, 54]ViT-Bsupin21k14M2243846.24.43.0--
DINO [8]ViT-Bsslin1k1M2243846.64.83.7--
UDON-CLIP [73]ViT-Bsupuned2.8M2243849.26.75.9--
OAI-CLIP [41]ViT-Bvlaopenai400M22438410.77.94.21.62.7
EVA-CLIP [57]ViT-Bvlamerged2b2B22438411.78.75.92.54.4
MetaCLIP [70]ViT-Bvla2pt5b2.5B22438412.79.46.64.97.6
DINOv2 [34]ViT-Bssllvd142m142M51872415.012.111.5--
SigLIP [75]ViT-Bvlawebli10B25638420.616.711.57.510.3
SigLIP [75]ViT-Bvlawebli10B38451226.221.515.611.014.4
SigLIP [75]ViT-Bvlawebli10B51272427.523.016.611.114.6
SigLIP2 [65]ViT-Bvlawebli10B51272428.623.515.410.414.6
EVA-MIM [14]ViT-Lsslin22k14M2243843.92.71.5--
ViT [13, 54]ViT-Lsupin21k14M2243847.35.34.6--
EVA-MIM [14]ViT-Lsslmerged38m38B2243848.86.14.7--
OAI-CLIP [41]ViT-Lvlaopenai400M22438415.811.97.04.66.7
OpenCLIP [9, 19]ViT-Lvlalaion2b2B22438417.513.79.47.09.4
Unicom [1]ViT-Ldistlaion400m400M33651218.614.613.9--
OAI-CLIP [41]ViT-Lvlaopenai400M33651219.915.29.45.88.4
DINOv2 [34]ViT-Lssllvd142m142M51872418.815.315.3--
EVA-CLIP [57]ViT-Lvlamerged2b2B33651220.916.010.97.210.6
MetaCLIP [70]ViT-Lvla2pt5b2.5B22438421.716.911.79.213.1
SigLIP [75]ViT-Lvlawebli10B25638426.321.815.212.816.4
SigLIP [75]ViT-Lvlawebli10B38451234.328.919.618.122.2
SigLIP2 [65]ViT-Lvlawebli10B51272437.331.320.819.824.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.461, + 0.907, + 0.517 + ], + "angle": 0, + "content": "Table 2. Performance comparison using mAP@1k on ILIAS and mini-ILIAS for global representation models for i2i and t2i. Comparison of model architecture (arch), training scheme (train), training data, and train/test resolution. \\(\\dagger\\) indicates results with the linear adaptation. 5M and 100M correspond to the mini and full versions of the dataset, respectively. sup, ssl, dist, vla: supervised learning, self-supervised learning, distillation and vision-language alignment. R50, CN: ResNet50 and ConvNext." + }, + { + "type": "table", + "bbox": [ + 0.123, + 0.521, + 0.446, + 0.607 + ], + "angle": 0, + "content": "
model100M5M-mini5M-rand
DINOv2† [34]15.318.822.7±0.2
EVA-CLIP† [57]16.020.928.8±0.2
MetaCLIP† [70]16.921.729.2±0.1
OpenCLIP† [19, 29]18.322.930.9±0.2
SigLIP† [75]28.934.341.8±0.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.61, + 0.483, + 0.679 + ], + "angle": 0, + "content": "Table 3. A challenging distractor subset for mini-ILIAS. mAP@1k evaluated for different distractor sets, 100M: the full dataset, 5M-mini: mini-ILIAS subset, 5M-rand: random subset. We report the mean and std of 3 randomly sampled subsets. \\(\\dagger\\) indicates results with the linear adaptation." + }, + { + "type": "table", + "bbox": [ + 0.12, + 0.68, + 0.451, + 0.816 + ], + "angle": 0, + "content": "
rerankingSigLIP [75]SigLIP† [75]
mAP@1koraclemAP@1koracle
global19.648.728.956.0
αQE1 [10, 40]22.144.733.756.9
αQE2 [10, 40]20.440.831.554.4
αQE5 [10, 40]14.334.923.549.3
CS [44]22.948.732.556.0
SP [37]21.848.730.556.0
AMES [56]26.448.735.656.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.818, + 0.483, + 0.875 + ], + "angle": 0, + "content": "Table 4. Performance comparison for re-ranking methods. Oracle represents the performance of perfect re-ranking at the top-1k images. Top: query expansion with global descriptors. Bottom: re-ranking with local descriptors. †: results with linear adaptation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.907, + 0.751 + ], + "angle": 0, + "content": "Retrieval with re-ranking. Tab. 4 shows the performance of re-ranking methods applied on top of SigLIP with and without linear adaptation on ILIAS. Complementary to \\( \\mathrm{mAP}@1\\mathrm{k} \\), an oracle-based top-1k re-ranking metric is reported as the upper bound of a re-ranking method that processes the top 1k images. Local similarity estimated by a learned model proves to be very effective for re-ranking. Nevertheless, the oracle re-ranking performance indicates that there is a lot more space for improvements. Re-ranking with QE is useful when the number of aggregated neighbors is low and drops below the baseline when the number of neighbors is increased. Notably, global re-ranking affects and, interestingly, decreases oracle performance since the whole db is re-ranked; while local re-ranking does not affect it since it is performed only on a shortlist of images." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.758, + 0.616, + 0.774 + ], + "angle": 0, + "content": "5.3. Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Performance per domain. Fig. 5 shows the performance per taxonomy categories. The taxonomy annotations allow a fine-grained view of the results, which can possibly allow us to capture imbalanced improvements in future work. For example, DINOV2, despite being overall inferior to SigLIP, is outperforming it in categories like architecture and sculptures or is quite similar in categories like public art and paper art. This is possibly attributed to the curation and com" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14783" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.114, + 0.079, + 0.87, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.241, + 0.908, + 0.282 + ], + "angle": 0, + "content": "Figure 5. Performance comparison per category. mAP@1k averaged over objects in the same mid-level taxonomy category, organized by their primary-level category size, with sorting within each group by SigLIP† + AMES performance. Comparison between SigLIP with and without adaptation, SigLIP combined with AMES reranking, SigLIP t2i, and DINOv2. † indicates results with the linear adaptation." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.284, + 0.462, + 0.564 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.565, + 0.483, + 0.608 + ], + "angle": 0, + "content": "Figure 6. Performance comparison reporting AP per query for different approaches with SigLIP. Pearson correlation reported in parenthesis. \\(\\dagger\\) indicates results with the linear adaptation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.483, + 0.705 + ], + "angle": 0, + "content": "position of the DINOV2 training set, which includes artwork and landmark datasets. Also, some categories are hurt by reranking with AMES, with some demonstrating big drops, i.e. sport, gaming, perfume. These categories deviate significantly from the domain AMES is trained, i.e. landmarks, which could potentially justify such drops." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Per query comparisons. Fig. 6 shows the AP per query for various methods. Linear adaptation boosts most queries, i.e. performance drop only for 192 queries. Image- and text-based retrieval are not strongly correlated despite performing similarly, which is good evidence [60] for the effectiveness of model ensembles. Indeed, ensembling i2i and t2i by averaging similarities brings \\(+6.1\\) improvement over i2i retrieval. Query expansion improves the queries with at least some positives at top positions, i.e. AP greater than 20. However, it harms many low-performing queries by aggregating descriptors irrelevant to the query. AMES improves the majority of the queries; however, many are harmed, indicating that there is plenty of room for improvement." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.285, + 0.905, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.448, + 0.908, + 0.532 + ], + "angle": 0, + "content": "Figure 7. Performance evaluation (mAP@1k) across different amounts of object area coverage and background clutter. Positives across all queries are jointly ranked based on coverage or clutter and split into 4 equal size groups. Queries with no positive in the corresponding group are discarded. No. of queries per group is in parentheses. \\(\\dagger\\) indicates results with the linear adaptation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.552, + 0.907, + 0.645 + ], + "angle": 0, + "content": "Impact of clutter and scale in positives. To quantify the impact of background clutter and scale changes, Fig. 7 presents the performance for different groups of positives. Dealing with small objects and multi-object scenes form major weaknesses of existing models. Notably, t2i beats i2i without adaptation in small-scale groups." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.662, + 0.642, + 0.679 + ], + "angle": 0, + "content": "6. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We introduce ILIAS and conduct an extensive evaluation of current foundational models and retrieval methods, highlighting that instance-level retrieval remains an unsolved problem. Our results indicate that off-the-shelf application of foundational models leaves considerable room for improvement, particularly in handling small objects and complex backgrounds. While specialized retrieval methods leveraging local descriptors are effective in these cases, their high memory and computational costs become impractical at the scale of ILIAS or beyond. ILIAS is designed to become a standard benchmark for evaluating foundational representation models and retrieval methods, accommodating both global and local representations, and advancing the field of instance-level recognition." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14784" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.486, + 0.344 + ], + "angle": 0, + "content": "Funding acknowledgement. This work was supported by the Junior Star GACR (grant no. GM 21-28830M), Horizon MSCA-PF (grant no. 101154126), Programme Johannes Amos Comenius (grant no. CZ.02.01.01/00/22.010/0003405), CTU in Prague (grant no. SGS23/173/OHK3/3T/13), the CTU institutional support (Future fund). JM was supported by Ministry of the Interior of the Czech Republic (project no. VJ02010041). We acknowledge VSB - Technical University of Ostrava, IT4Innovations National Supercomputing Center, Czech Republic, for awarding this project access to the LUMI supercomputer, owned by the EuroHPC Joint Undertaking, hosted by CSC (Finland) and the LUMI consortium through the Ministry of Education, Youth and Sports of the Czech Republic through the e-INFRA CZ (grant ID: 90254)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.351, + 0.485, + 0.562 + ], + "angle": 0, + "content": "Contributors acknowledgments. We want to thank Larysa Ivashechkina for her work on the annotation of object bounding boxes and masks and the initial filtering of text queries. We also want to thank all contributors for the collection of ILIAS dataset. Additionally to all authors, the list of external contributor in alphabetical order: Aggeliki Tserota, Anna Nisyraiou, Celeste Abreu, Charalambos Tzamos, Christina Tserota, Dimitris Karageorgiou, Dmytro Mishkin, Eleni Karantali, Eleni Papadopoulou, Eva Tsiliakou, Kelly Kordopati-Zilou, Markos Zampoglou, Noa Garcia, Panagiotis Tassis, Paraskevas Kordopatis, Pavlos Alexantonakis, Ruslan Rozumnyi, Snežana Čurguz, Tereza Nejedlá, Tomáš Jelinek, Yannis Kalantidis, Vasilis Alexiadis, Yankun Wu." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.577, + 0.188, + 0.592 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.602, + 0.483, + 0.657 + ], + "angle": 0, + "content": "[1] Xiang An, Jiankang Deng, Kaicheng Yang, Jaiwei Li, Ziyong Feng, Jia Guo, Jing Yang, and Tongliang Liu. Unicom: Universal and compact representation learning for image retrieval. In ICLR, 2023. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.659, + 0.483, + 0.687 + ], + "angle": 0, + "content": "[2] Relja Arandjelović and Andrew Zisserman. Smooth object retrieval using a bag of boundaries. In ICCV, 2011. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.689, + 0.484, + 0.743 + ], + "angle": 0, + "content": "[3] Harry G Barrow, Jay M Tenenbaum, Robert C Bolles, and Helen C Wolf. Parametric correspondence and chamfer matching: Two new techniques for image matching. In *IJ-CAI*, 1977. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.746, + 0.483, + 0.786 + ], + "angle": 0, + "content": "[4] Wissam Bejjani, Wisdom C Agboh, Mehmet R Dogar, and Matteo Leonetti. Occlusion-aware search for object retrieval in clutter. In IROS, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.789, + 0.484, + 0.857 + ], + "angle": 0, + "content": "[5] Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselat, Emma Brunskill, et al. On the opportunities and risks of foundation models. In arXiv, 2021. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.859, + 0.484, + 0.899 + ], + "angle": 0, + "content": "[6] Bingyi Cao, André Araujo, and Jack Sim. Unifying deep local and global features for image search. In ECCV, 2020. 5" + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.602, + 0.484, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.093, + 0.907, + 0.147 + ], + "angle": 0, + "content": "[7] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.151, + 0.906, + 0.205 + ], + "angle": 0, + "content": "[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.208, + 0.906, + 0.276 + ], + "angle": 0, + "content": "[9] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In CVPR, 2023. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.28, + 0.905, + 0.333 + ], + "angle": 0, + "content": "[10] Ondrej Chum, James Philbin, Josef Sivic, Michael Isard, and Andrew Zisserman. Total recall: Automatic query expansion with a generative feature model for object retrieval. In ICCV, 2007. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.337, + 0.905, + 0.378 + ], + "angle": 0, + "content": "[11] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. In ICLR, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.381, + 0.905, + 0.422 + ], + "angle": 0, + "content": "[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.425, + 0.905, + 0.508 + ], + "angle": 0, + "content": "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICML, 2021. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.511, + 0.905, + 0.566 + ], + "angle": 0, + "content": "[14] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. EVA: Exploring the limits of masked visual representation learning at scale. In CVPR, 2023. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.568, + 0.905, + 0.623 + ], + "angle": 0, + "content": "[15] Martin A. Fischler and Robert C. Bolles. Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 1981. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.626, + 0.905, + 0.654 + ], + "angle": 0, + "content": "[16] Christopher Flagg and Ophir Frieder. Reconstruction of artifacts from digital image repositories. JOCCH, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.656, + 0.905, + 0.696 + ], + "angle": 0, + "content": "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.7, + 0.905, + 0.741 + ], + "angle": 0, + "content": "[18] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.744, + 0.905, + 0.813 + ], + "angle": 0, + "content": "[19] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.816, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[20] Hervé Jégou, Matthijs Douze, and Cordelia Schmid. Hamming embedding and weak geometric consistency for large scale image search. In ECCV, 2008. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.859, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[21] Tomas Jenicek and Ondrej Chum. No fear of the dark: Image retrieval under varying illumination conditions. In CVPR, 2019. 1" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.093, + 0.907, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14785" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[22] Sungyeon Kim, Boseung Jeong, and Suha Kwak. HIER: Metric learning beyond class labels via hierarchical regularization. In CVPR, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.482, + 0.189 + ], + "angle": 0, + "content": "[23] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In CVPR, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[24] Byungsoo Ko, Minchul Shin, Geonmo Gu, HeeJae Jun, Tae Kwan Lee, and Youngjoon Kim. A benchmark on tricks for large-scale image retrieval. In arXiv, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.232, + 0.482, + 0.273 + ], + "angle": 0, + "content": "[25] Tarun Krishna, Kevin McGuinness, and Noel O'Connor. Evaluating contrastive models for instance-based image retrieval. In ICMR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.273, + 0.482, + 0.315 + ], + "angle": 0, + "content": "[26] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In NeurIPS, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.482, + 0.356 + ], + "angle": 0, + "content": "[27] Seongwon Lee, Hongje Seong, Suhyeon Lee, and Euntai Kim. Correlation verification for image retrieval. In CVPR, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.358, + 0.482, + 0.399 + ], + "angle": 0, + "content": "[28] Ziwei Liu, Ping Luo, Shi Qiu, Xiaogang Wang, and Xiaou Tang. Deepfashion: Powering robust clothes recognition and retrieval with rich annotations. In CVPR, 2016. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.482, + 0.441 + ], + "angle": 0, + "content": "[29] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. CVPR, 2022. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.482, + 0.495 + ], + "angle": 0, + "content": "[30] Shiyang Lu, Haonan Chang, Eric Pu Jing, Abdeslam Boularias, and Kostas Bekris. OVIR-3D: Open-vocabulary 3d instance retrieval without training on 3d data. In CoRL, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.497, + 0.482, + 0.525 + ], + "angle": 0, + "content": "[31] Claudio Michaelis, Matthias Bethge, and Alexander Ecker. One-shot segmentation in clutter. In ICML, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.525, + 0.482, + 0.553 + ], + "angle": 0, + "content": "[32] David Nister and Henrik Stewenius. Scalable recognition with a vocabulary tree. In CVPR, 2006. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.554, + 0.482, + 0.594 + ], + "angle": 0, + "content": "[33] Hyeonwoo Noh, Andre Araujo, Jack Sim, Tobias Weyand, and Bohyung Han. Large-scale image retrieval with attentive deep local features. In ICCV, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.482, + 0.719 + ], + "angle": 0, + "content": "[34] Maxime Oquab, Timothee Darct, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINOv2: Learning robust visual features without supervision. TMLR, 2024. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.482, + 0.761 + ], + "angle": 0, + "content": "[35] Yash Patel, Giorgos Tolias, and Jiří Matas. Recall@k surrogate loss with large batches and similarity mixup. In CVPR, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.482, + 0.803 + ], + "angle": 0, + "content": "[36] Jingtian Peng, Chang Xiao, and Yifan Li. RP2K: A large-scale retail product dataset for fine-grained image classification. In arXiv, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.482, + 0.845 + ], + "angle": 0, + "content": "[37] James Philbin, Ondrej Chum, Michael Isard, Josef Sivic, and Andrew Zisserman. Object retrieval with large vocabularies and fast spatial matching. In CVPR, 2007. 2, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[38] James Philbin, Ondrej Chum, Michael Isard, Josef Sivic, and Andrew Zisserman. Lost in quantization: Improving particular object retrieval in large scale image databases. In CVPR, 2008. 4" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[39] Filip Radenović, Ahmet Iscen, Giorgos Tolias, Yannis Avrithis, and Ondrej Chum. Revisiting oxford and paris: Large-scale image retrieval benchmarking. In CVPR, 2018. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[40] Filip Radenovic, Giorgos Tolias, and Ondrej Chum. Finetuning cnn image retrieval with no human annotation. PAMI, 2019. 2, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.274 + ], + "angle": 0, + "content": "[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.276, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[42] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, et al. SAM 2: Segment anything in images and videos. In ICLR, 2025. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[43] Ali Sharif Razavian, Hossein Azizpour, Josephine Sullivan, and Stefan Carlsson. CNN features off-the-shelf: An astounding baseline for recognition. In CVPRW, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.376, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[44] Ali S Razavian, Josephine Sullivan, Stefan Carlsson, and Atsuto Maki. Visual instance retrieval with deep convolutional networks. ITE Trans. on Media Technology and Applications, 2016. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.433, + 0.905, + 0.502 + ], + "angle": 0, + "content": "[45] Taichi Sakaguchi, Akira Taniguchi, Yoshinobu Hagiwara, Lotfi El Hafi, Shoichi Hasegawa, and Tadahiro Taniguchi. Object instance retrieval in assistive robotics: Leveraging fine-tuned simsiam with multi-view images based on 3d semantic map. In IROS, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.905, + 0.558 + ], + "angle": 0, + "content": "[46] Mert Bulent Sariyildiz, Philippe Weinzaepfel, Thomas Lucas, Diane Larlus, and Yannis Kalantidis. UNIC: Universal classification models via multi-teacher distillation. In ECCV, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.56, + 0.905, + 0.602 + ], + "angle": 0, + "content": "[47] Konstantin Schall, Kai Uwe Barthel, Nico Hezel, and Klaus Jung. GPR1200: a benchmark for general-purpose content-based image retrieval. In MMM, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.603, + 0.905, + 0.645 + ], + "angle": 0, + "content": "[48] Shihao Shao, Kaifeng Chen, Arjun Karpur, Qinghua Cui, André Araujo, and Bingyi Cao. Global features are all you need for image retrieval and reranking. In ICCV, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.646, + 0.905, + 0.686 + ], + "angle": 0, + "content": "[49] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In ICLR, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[50] Ivan Sipiran, Patrick Lazo, Cristian Lopez, Milagritos Jimenez, Nihar Bagewadi, Benjamin Bustos, Hieu Dao, Shankar Gangisetty, Martin Hanik, Ngoc-Phuong Ho-Thi, et al. SHREC 2021: Retrieval of cultural heritage objects. Computers & Graphics, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.759, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[51] Arnold WM Smeulders, Marcel Worring, Simone Santini, Amarnath Gupta, and Ramesh Jain. Content-based image retrieval at the end of the early years. PAMI, 2000. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[52] Chull Hwan Song, Jooyoung Yoon, Taebaek Hwang, Shunghyun Choi, Yeong Hyeon Gu, and Yannis Avrithis. On train-test class overlap and detection for image retrieval. In CVPR, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[53] Hyun Oh Song, Yu Xiang, Stefanie Jegelka, and Silvio Savarese. Deep metric learning via lifted structured feature embedding. In CVPR, 2016. 1, 2, 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14786" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[54] Andreas Steiner, Alexander Kolesnikov, Xiaohua Zhai, Ross Wightman, Jakob Uszkoreit, and Lucas Beyer. How to train your vit? data, augmentation, and regularization in vision transformers. TMLR, 2021. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.15, + 0.482, + 0.191 + ], + "angle": 0, + "content": "[55] Pavel Suma and Giorgos Tolias. Large-to-small image resolution asymmetry in deep metric learning. In WACV, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.194, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[56] Pavel Suma, Giorgos Kordopatis-Zilos, Ahmet Iscen, and Giorgos Tolias. AMES: Asymmetric and memory-efficient similarity estimation for instance-level retrieval. In ECCV, 2024. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.252, + 0.482, + 0.292 + ], + "angle": 0, + "content": "[57] Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. EVA-CLIP: Improved training techniques for clip at scale. In arXiv, 2023. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.296, + 0.482, + 0.35 + ], + "angle": 0, + "content": "[58] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In CVPR, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.353, + 0.482, + 0.381 + ], + "angle": 0, + "content": "[59] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In ICML, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.383, + 0.482, + 0.424 + ], + "angle": 0, + "content": "[60] Ryan Theisen, Hyunsuk Kim, Yaoqing Yang, Liam Hodgkinson, and Michael W Mahoney. When are ensembles really effective? In NeurIPS, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.427, + 0.482, + 0.481 + ], + "angle": 0, + "content": "[61] Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. YFCC100M: The new data in multimedia research. Communications of the ACM, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.482, + 0.525 + ], + "angle": 0, + "content": "[62] Akihiko Torii, Relja Arandjelovic, Josef Sivic, Masatoshi Okutomi, and Tomas Pajdla. 24/7 place recognition by view synthesis. In CVPR, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.528, + 0.482, + 0.569 + ], + "angle": 0, + "content": "[63] Hugo Touvron, Andrea Vedaldi, Matthijs Douze, and Herve Jegou. Fixing the train-test resolution discrepancy. In NeurIPS, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.572, + 0.482, + 0.626 + ], + "angle": 0, + "content": "[64] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.63, + 0.482, + 0.712 + ], + "angle": 0, + "content": "[65] Michael Tschannen, Alexey Gritsenko, Xiao Wang, Muhammad Ferjad Naeem, Ibrahim Alabdulmohsin, Nikhil Parthasarathy, Talfan Evans, Lucas Beyer, Ye Xia, Basil Mustafa, et al. SigLIP 2: Multilingual vision-language encoders with improved semantic understanding, localization, and dense features. In arXiv, 2025. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.715, + 0.482, + 0.755 + ], + "angle": 0, + "content": "[66] Shuang Wang and Shuqiang Jiang. INSTRE: A new benchmark for instance-level object retrieval and recognition. TOMM, 2015. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.758, + 0.482, + 0.813 + ], + "angle": 0, + "content": "[67] Tobias Weyand, André Araujo, Bingyi Cao, and Jack Sim. Google Landmarks Dataset v2 - A large-scale benchmark for instance-level recognition and retrieval. In CVPR, 2020. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.816, + 0.482, + 0.856 + ], + "angle": 0, + "content": "[68] Ross Wightman, Hugo Touvron, and Hervé Jégou. Resnet strikes back: An improved training procedure in timm. In NeurIPSW, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[69] Pengxiang Wu, Siman Wang, Kevin Dela Rosa, and Derek Hu. FORB: A flat object retrieval benchmark for universal image embedding. In NeurIPS, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[70] Hu Xu, Saining Xie, Xiaqing Ellen Tan, Po-Yao Huang, Russell Howes, Vasu Sharma, Shang-Wen Li, Gargi Ghosh, Luke Zettlemoyer, and Christoph Feichtenhofer. Demystifying clip data. In ICLR, 2024. 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[71] Nikolaos-Antonios Ypsilantis, Noa Garcia, Guangxing Han, Sarah Ibrahimi, Nanne Van Noord, and Giorgos Tolias. The MET dataset: Instance-level recognition for artworks. In NeurIPS, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[72] Nikolaos-Antonios Ypsilantis, Kaifeng Chen, Bingyi Cao, Mario Lipovsky, Pelin Dogan-Schönberger, Grzegorz Makosa, Boris Bluntschli, Mojtaba Seyedhosseini, Ondrej Chum, and André Araujo. Towards universal image embeddings: A large-scale dataset and challenge for generic image representations. In ICCV, 2023. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.905, + 0.344 + ], + "angle": 0, + "content": "[73] Nikolaos-Antonios Ypsilantis, Kaifeng Chen, Andre Araujo, and Ondrej Chum. UDON: Universal dynamic online distillation for generic image representations. In NeurIPS, 2024. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.347, + 0.905, + 0.401 + ], + "angle": 0, + "content": "[74] Jiangbo Yuan, An-Ti Chiang, Wen Tang, and Antonio Haro. eProduct: A million-scale visual search benchmark to address product recognition challenges. In CVPRW, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.404, + 0.905, + 0.444 + ], + "angle": 0, + "content": "[75] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In ICCV, 2023. 1, 2, 3, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.447, + 0.905, + 0.502 + ], + "angle": 0, + "content": "[76] Xunlin Zhan, Yangxin Wu, Xiao Dong, Yunchao Wei, Minlong Lu, Yichi Zhang, Hang Xu, and Xiaodan Liang. Product: Towards weakly supervised instance-level product retrieval via cross-modal pretraining. In ICCV, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.905, + 0.53 + ], + "angle": 0, + "content": "[77] Liang Zheng, Yi Yang, and Qi Tian. Sift meets cnn: A decade survey of instance retrieval. PAMI, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.532, + 0.905, + 0.573 + ], + "angle": 0, + "content": "[78] Barret Zoph, Vijay Vasudevan, Jonathon Shlens, and Quoc V Le. Learning transferable architectures for scalable image recognition. In CVPR, 2018. 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.573 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "14787" + } + ] +] \ No newline at end of file diff --git a/2025/ILIAS_ Instance-Level Image retrieval At Scale/bd262cec-535c-47b6-b08f-5430c3f91bd5_origin.pdf b/2025/ILIAS_ Instance-Level Image retrieval At Scale/bd262cec-535c-47b6-b08f-5430c3f91bd5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..15cc29391416e4cf1ddb9da4352bd41673438b9b --- /dev/null +++ b/2025/ILIAS_ Instance-Level Image retrieval At Scale/bd262cec-535c-47b6-b08f-5430c3f91bd5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0bd7dedf5525ad83ebb24c4b33a52d20f684c1d0a3a5a79d7c9fae9fe51145e9 +size 2260227 diff --git a/2025/ILIAS_ Instance-Level Image retrieval At Scale/full.md b/2025/ILIAS_ Instance-Level Image retrieval At Scale/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c34761457d052fa44eed956d25ad023c1759d456 --- /dev/null +++ b/2025/ILIAS_ Instance-Level Image retrieval At Scale/full.md @@ -0,0 +1,276 @@ +# ILIAS: Instance-Level Image retrieval At Scale + +Giorgos Kordopatis-Zilos Vladan Stojnic Anna Manko Pavel Šuma Nikolaos-Antonios Ypsilantis Nikos Efthymiadis Zakaria Laskar Jiří Matas Ondřej Chum Giorgos Tolias + +VRG, FEE, Czech Technical University in Prague + +# Abstract + +This work introduces ILIAS, a new test dataset for Instance-Level Image retrieval At Scale. It is designed to evaluate the ability of current and future foundation models and retrieval techniques to recognize particular objects. The key benefits over existing datasets include large scale, domain diversity, accurate ground truth, and a performance that is far from saturated. ILIAS includes query and positive images for 1,000 object instances, manually collected to capture challenging conditions and diverse domains. Large-scale retrieval is conducted against 100 million distractor images from YFCC100M. To avoid false negatives without extra annotation effort, we include only query objects confirmed to have emerged after 2014, i.e. the compilation date of YFCC100M. An extensive benchmarking is performed with the following observations: i) models fine-tuned on specific domains, such as landmarks or products, excel in that domain but fail on ILIAS ii) learning a linear adaptation layer using multi-domain class supervision results in performance improvements, especially for vision-language models iii) local descriptors in retrieval re-ranking are still a key ingredient, especially in the presence of severe background clutter iv) the text-to-image performance of the vision-language foundation models is surprisingly close to the corresponding image-to-image case. + +website: https://vrg.fel(cvut.cz/ilias/ + +# 1. Introduction + +The ability to recognize and differentiate every unique object instance in the physical world represents one of the ultimate goals for foundation representation models [5, 34, 41, 75]. This work aims to assess this capability through the lens of instance-level image retrieval at a very large scale. Instance-level image retrieval corresponds to searching for images of particular objects within large collections. All images of a particular object form their own instance-level class. This is an important information retrieval task due to its numerous real-world applications in robotics [30, 45], e + +![](images/3e869c247b7b01b1f9361e6e73886bb942599fafb8737dd642fe9cbc714c3a52.jpg) +Figure 1. Performance timeline on ILIAS. Curves indicate best performance in chronological order for image-to-image and text-to-image retrieval, showing a significant boost with the release of foundation models. Representations are linearly adapted via multi-domain learning on UnED [72]. Re-ranking with local descriptors achieves the best results by a significant margin. + +commerce [74, 76], and cultural heritage [16, 50], to name just a few. The task faces challenges because of the substantial variations among positive examples, such as illumination/viewpoint [21, 62] changes and background clutter [4, 31]. An additional difficulty is the high similarity among negatives, which is driven by the extremely fine granularity in the class definitions. It becomes even more challenging at a real-world scale, where searching through millions or even billions of images requires handling an open-world setup with countless unseen objects spanning diverse and complex domains. + +Benchmarking instance-level retrieval under real-world challenges is currently limited by the lack of suitable datasets. Constructing a dataset with instance-level class definitions necessitates huge development effort, reflected by the many shortcomings of existing datasets. Shortcomings exist in several key aspects, such as dataset size [66], domain diversity [39, 53], and ground-truth accuracy [67], which suffers from both false positives and false negatives. Popular datasets are typically limited to landmarks [39], and as dataset scale increases, ground-truth quality tends to decline [52, 67]. This is a consequence of automating the + +ground-truth creation process to facilitate scaling up. To address such limitations, we introduce the Instance-Level Image retrieval At Scale (ILIAS) evaluation dataset. + +The creation of our dataset has two key elements. First, query and positive images are manually captured to ensure challenging variations, covering 1,000 objects across diverse domains. Second, to expand the dataset size without ground-truth errors or additional annotation effort, we leverage a key technique: distractor images, collected in 2014 from YFCC100M, are combined with query objects verified not to have publicly existed until after 2014. This distractor set includes 100 million images, two orders of magnitude larger than the largest existing dataset [39]. Notably, all images have a permissive license, allowing us to ensure long-term online availability to the full extent. + +ILIAS includes both image and text queries. The latter is in the form of detailed descriptions of objects and their distinctive features. The dataset is designed to support future research in image-to-image and text-to-image retrieval for particular objects, and additionally serves as a large-scale benchmark for evaluating representations of foundation vision and language models (VLM) [41, 75]. To facilitate faster experimentation, we provide a mini, but challenging, version (5M) of the distractor set. + +We perform an extensive evaluation comparison, including many foundation image-to-image and text-to-image models, and establish a comprehensive testbed that enables future comparisons. The provided evaluation includes retrieval with global image representation but also re-ranking techniques that use local representations [37, 43, 56] and query expansion [10, 40]. We observe the following: + +- Performance of standard 10-year-old models, measured by mean Average Precision, is as low as $1.3\%$ , while the best-performing model achieves $31.3\%$ , as shown in Fig. 1. This points out the vast progress of representation models and the high challenging factors of ILIAS. +- VLMs are the top-performing models. +- Smaller (ViT-B) models trained/tested on large resolution (512/724) outperform larger models (ViT-L) trained/tested on small resolution (256/384). +- Using Universal Embedding Dataset (UnED) [72] to learn a linear adaptation layer on top of frozen models improves performance of most models, making it a candidate training set to couple with ILIAS. Notably, VLMs demonstrate the largest benefits, presumably because their training stage does not optimize image-to-image relations. +- In contrast to the current belief [48], local representation is a key ingredient, while global representation, despite being efficient and compact, performs much lower. +- The performance gap between image-to-image and text-to-image models is surprisingly small. Therefore, detailed text queries are a reasonable proxy in the absence of image queries, even at the instance level. + +# 2. Related work + +In this section, we review the related work in terms of existing datasets and benchmarks in the literature. + +Datasets. Tab. 1 presents the datasets from the image retrieval literature related to ILIAS. The datasets can be compared based on five main axes: (i) Class definition adopted. Many datasets [20, 32, 33, 39, 67] adopt a strict definition very similar to ours, satisfying instance-level requirements. Others [2, 53, 66, 76] adopt a more relaxed definition, where some minor variations are permitted, e.g. color changes in objects of the same class. Even more relaxed are the fine-grained definitions [74], where the object of the very same type is considered related, e.g. same product with different variant. (ii) Domain of the dataset. Most datasets are tailored for a specific domain. Landmarks are among the most popular domains [2, 20, 33, 39, 67]. Other domains include products [32, 36, 76] and fashion [28, 53]. Some datasets cover multiple domains, either being standalone [66] or bundle of repurposed datasets [47, 72]. (iii) Scale of database. Most of the datasets are small-scale, counting a few thousand images [28, 66, 76]. Larger ones [39, 72, 74] expend slightly above a million. None satisfies large-scale requirements. (iv) Noise in ground truth. Most datasets consist of clean annotations, except for a few cases that contain inaccuracies, including false positives [67], i.e. images wrongly annotated as relevant, false negatives [32, 39], i.e. relevant images that have not been annotated as positives, or the possibility of false positives [39]. (v) Availability. Most datasets are publicly available with permissive licenses, with few exceptions of partial [2, 20] or no [33, 74] availability. To this end, no publicly available dataset fits the strict instance-level definition, contains objects from multiple domains, ensures error-free labeling and is large scale. This gap is filled with ILIAS satisfying all the aforementioned requirements. + +Evaluation benchmarks. Benchmarking [51] tracks the progress in the field, which is even more necessary with the emergence of foundation models. Several benchmarks papers [24, 25, 77] exist in the instance-level retrieval literature, investigating the impact of learning scheme, post-processing, model ensembling, query expansion, and whitening. The most relevant benchmark to ILIAS is UnED [72] that combines existing datasets to create a union that evaluates models performance across various domains. Due to its wide variety, UnED serves as the training dataset for linear adaptation. + +Regarding the evaluation of foundation models, the most common practice [14, 34, 57] is measuring classification performance on top of frozen models on ImageNet [12]. This is performed either with or without the training of a classifier via linear probing or k-NN search. Furthermore, models are usually evaluated on dense prediction tasks [34] and several multiple-downstream single-domain tasks [1]. + +
datasetsyearobjectsquerypositivesdatabasegtclass def.domainbboxonlinelicense
UKB [32]20062.5K10K10K10KFNILproductXFullyN/A
Holidays [20]20085005009911MCleanILlandmarkXPartiallyCC
Sculptures [2]201110703.1K3.1KCleanPartial ILsculptureXPartiallyFlickr TC
INSTRE [66]2015200125027.3K27.3KCleanPartial ILmulti✓FullyFlickr TC
SOP [53]201511.3K60.5K60.5K60.5KCleanPartial ILproductXFullyMIT License
InShop [28]20163.9K14.2K12.6K12.6KCleanPartial ILfashion✓FullyN/A
R-Oxford [39]201811705K1MFN?ILlandmarkXFullyFlickr TC, CC
R-Paris [39]201811706.3K1MFN?ILlandmarkXFullyFlickr TC, CC
GLDv1 [33]201830KN/AN/A1.1MCleanILlandmarkXPartiallyMultiple
GLDv2 [67]20203181.1K3.1K762KFPILlandmarkXFullyCC/ Public-domain
Product1M [76]20213926.5K40K40KCleanPartial ILproductXPartiallyN/A
RP2K [36]20211.2K10.9K10.9K10.9KCleanPartial ILproduct✓FullyN/A
GPR1200 [47]20211.2K12K12K12KMixIL+FGmultiXFullyMultiple
eProduct [74]202120610KN/A1.1MN/AFGproductXNoN/A
FORB [69]2023N/A13.9K4.5K49.8KCleanILplanarXFullySnap Inc.
UnED [72]202321K241K244K1.4MMixIL+FGmultiXFullyMultiple
ILIAS20251,0001,2324,715100MCleanILmulti✓FullyCC
+ +Table 1. Comparison with other instance-level datasets. Datasets are compared based on their size (object, query, positives, database), the accuracy of the ground truth (gt), type of class definition, domain, supplementary annotations (bbox) and accessibility (online, license). N/A: not available. FP/FN: false positives/negatives. FN?: possibility of false negatives. Mix: combination of clean and noisy datasets. IL: instance-level. FG: fine-grained. Partial IL: instance-level with subtle variations among same class objects. CC: Creative Commons. + +For VLMs, zero-shot classification and retrieval serve as the primary benchmarks [41, 70, 75], utilizing class text labels. In this work, we provide similar evaluation protocols tailored for instance-level retrieval. One can test the raw model capabilities or adapt for the instance-level task via linear adaptation on UnED. Text-to-image retrieval is also facilitated for the evaluation of VLMs. + +# 3. ILIAS dataset + +# 3.1. Composition and collection + +Instance-level class definition. Following an instance-level class definition [39, 71], we consider all indistinguishable object instances of the real world to form their own class. Nevertheless, we add a restriction to consider a pair of images as relevant to each other only if there is a view overlap. Other cases are explicitly not included in the dataset, contrasting the existing work [28, 53, 67]. Therefore, models should mostly rely on estimating the visual similarity and less on shortcuts through semantics. + +Overview. ILIAS supports both image-to-image (i2i) and text-to-image (t2i) retrieval and follows the standard setup for retrieval datasets, consisting of two main parts: (i) query images and text, and (ii) database (db) images. The objective is to rank positives - db images relevant to the query - at the top ranks. The collected objects cover a wide range of categories and are not restricted to specific domains. An overview of some collected objects is provided in Fig. 2. + +Queries and positives are created/collected by a group of collectors that are well-informed about the task objectives. In addition to positives, in the database, we include numerous distractors - irrelevant (negative) images to the queries - that make retrieval more challenging. Following previous work [39], adding a large, uncurated set of ran + +dom images achieves this. The larger the set, the higher the chances of hard negatives – images with similar appearance or semantics to the queries. To this end, we select the YFCC100M [61] dataset to serve as the source of distractors due to its size and permissive license. + +Selected objects. Ensuring that distractor images include no false negatives cannot be performed in a scalable or accurate way if one relies on human annotation or metadata. Instead, we take advantage of the fact that YFCC100M was crawled from Flickr in 2014. Hence, an object qualifies in ILIAS if it could not have appeared on Flickr before 2014. To verify this, we rely either on publicly available information, e.g. objects known to be created/manufactured after 2014, or on the collector's knowledge about the object not being publicly available. Additionally, we opt for objects with distinctive and unique features that set them apart from others within the same category. For example, we avoid recent smartphones that look like plain black screens or new objects with distinctive parts closely resembling older ones. + +Queries and positives. Query images depict the instance on a clean or uniform background. When this is not feasible (e.g. buildings or statues), background blurring or cropping is applied. This is performed to avoid including background objects in the query that do not have corresponding positives in our ground truth information. Positives are images featuring the query object in challenging conditions, such as clutter, scale changes, occlusions, and partial views. Prior work [39] reveals that easy positives dominate performance metrics. Thus, we specifically opt for challenging cases that cannot be easily retrieved by the models. To avoid taking advantage of camera identification, most query and positive images are captured with at least two different camera models to introduce diversity. We also incorporate older camera models that are used in YFCC100M. + +![](images/b80f76ca47b653aad9f0f85ba98c45ce40709240a8b2314ae451ca72300b8782.jpg) +Figure 2. Examples of query, positive and hard negatives within the distractor set. Average Precision per query and rank of the negatives and positives is reported using SigLIP [75] model. Gray: queries. Green: positives. Red: distractors. + +Each text query consists of a detailed and fine-grained textual description of an object. Descriptions are initially created by a large language model prompted to provide highly detailed depictions of the object shown in query images. Generated descriptions are manually edited to fix errors, insufficient descriptions, or nuances of the model. + +Distractors. The YFCC100M dataset was chosen for the distractor set due to its large scale and diverse range of concepts. It consists of 100 million Flickr images, collected without specific filtering, aside from being shared under a permissive CC-BY license. + +Bounding box annotation. We include supplementary bounding boxes that specify the precise location of objects in query and positive images. They provide statistics about the position and size of object areas, assist our analysis of the dataset challenges, and support future research in instance-level localization. + +Evaluation metric. Retrieval performance is evaluated via mean Average Precision (mAP), a widely used metric in instance-level image retrieval [37, 38, 40]. Specifically, we adopt mAP@1k [67], which assesses the ranking of the top-1k nearest neighbors for each query, treating any positive not ranked among the top-1k as not retrieved. We estimate the area under the curve using rectangles and not trapezoids. + +# 3.2. Statistics + +Dataset size. The final ILIAS dataset includes 1,000 object instances captured in 5,947 images, of which 1,232 are queries and 4,715 are positives. Fig. 3a shows the distribution of positives per object. Also, 99,144,315 images from YFCC100M are downloaded. All images (queries, positives, distractors) are transferred through Flickr to ensure the same pre-processing. + +Taxonomy. A hierarchical 3-level taxonomy is composed for ILIAS. All instances are assigned across one to three categories of different granularity levels. The taxonomy + +consists of 8 categories on the coarser level, 42 on the mid level, and 38 on the finer level. The categories are derived through manual labeling of the objects based on their semantic content. To form the coarser-level categories, we use domain definitions borrowed from prior work [28, 53, 67] to align with the literature, i.e. art, landmarks, products, fashion. We also define novel categories based on the objects that do not fit into any existing domain. The distribution of objects across categories is uneven, e.g. ranging from 168 and 162 for art and landmarks to 83 for products. Each mid- and finer-level category contains at least 4 instances. Note that taxonomy is given to provide statistics about the domains of objects and assist our analysis instead of being leveraged as ground truth. The distribution of taxonomy categories can be inferred by Fig. 5, and a detailed figure is provided in the supplementary material. + +Bounding box analysis. A total number of 6,117 bounding boxes are annotated for both queries and positives. Note that positives may display multiple objects of near identical appearance to the query; in such cases, bounding boxes are drawn on all indistinguishable objects. There are 235 images with more than one bounding box. Based on the annotated bounding boxes, we compute the area covered in the image by the object instances to derive its relative scale. Fig. 3c shows the distribution of the scale ratio for queries and positives. Most objects in queries cover the largest area of the images; while in the vast majority of positives, the object covers a small area of less than half the image. It is a result of the severe scale changes and partial views. Moreover, we use the Segment Anything Model (SAM) [23, 42] to extract object segments from positives. The number of detected segments outside the query object's bounding boxes is computed. This indicates clutter from other items in the positives. Fig. 3b shows the segment number distribution, with most images containing multiple segments due to clutter. + +![](images/682c2ec5fdb3ff2516d4dcacbea8a3efb6b4bd7ac55b67b3b003a252a840c08e.jpg) +(a) + +![](images/7d23c8a6223101fd7452578cac9c2d0217afbe8dc317f95a94a3584b8d8355ba.jpg) +(b) + +![](images/87fe4f25dbb74f5839ab093adcaadd6bb9fd1fc582a1e8708babd9a956df2551.jpg) +(c) +Figure 3. ILIAS statistics. (a) number of positives per object, (b) positive distribution by the SAM segments outside the bounding box, (c) image distribution by the relative bounding box area. + +# 3.3. mini-ILIAS + +We provide a small version of ILIAS, called mini-ILIAS, to facilitate quick experimentations. It consists of the query and positive images collected for ILIAS, and a subset of the YFCC100M distractors. Instead of randomly subsampling YFCC100M, we construct a challenging subset with the help of VLMs. We aim at selecting distractors displaying objects of similar categories as the query objects. We use the text category labels of the taxonomy as text queries. We also extend them with standard templates used for zero-shot recognition [41], which resulted in several thousands. T2i similarity between each text query and each distractor image is estimated. A similarity score for each distractor is derived based on its maximum similarity over the text queries. We ensemble the scores of 3 VLMs to rank images. The top-5M ranked distractors compose the final mini-ILIAS. Our experiments indicate that this subset is significantly more challenging than a random subset of the same size. + +# 4. Benchmark methods + +We describe the methods and foundation models we evaluate on ILIAS, which are grouped according to their type of representations used for retrieval in global (i2i) representations, re-ranking with global (i2i) representations, re-ranking with local (i2i) representations, and text-to-image. A detailed list of models, their performance, and implementation details are in the supplementary materials. + +Image-to-image retrieval with global representations. Global representation methods use image encoders to map images to global descriptors and rank db images based on cosine similarity. We evaluate legacy and recent foundation models, varying in architecture, descriptor dimensionality, training scheme, training data, and input resolution. Foundation models [5] are the models trained with a training set on the scale of a hundred million. Particularly, 23 CNN [17, 26, 29, 49, 58, 59, 78] and 45 ViT [13, 14] models, trained with supervision [22, 54, 64, 68], self-supervision [7, 8, 18, 34], distillation [1, 46, 64], or visual-language alignment [9, 41, 57, 65, 70, 75] are benchmarked. Most of the non-foundation models are trained on ImageNet [12]. There are models trained on single specific do + +mains [27, 35, 48], i.e. landmarks or products on GLDv2 or SOP. Universal models [1, 46, 72, 73] trained on multi-domains or multi-task schemes are included. The full list of models and results is provided in supplementary materials. + +To mitigate the differences in training resolution, we use three widely-used resolutions, i.e. 384, 512, and 724 and resize images so that their larger dimension matches one of the three. The test resolution is defined to be one resolution above the one used for training, e.g. a network trained with 224 or 384 is tested with 384 or 512, respectively. The vast majority of models achieve best performance under this rule. Similar behavior is observed in the literature [55, 63]. + +Linear adaptation for i2i retrieval. Pre-trained foundation models, as well as legacy models, are trained to extract representations that are applicable to various tasks; not all encoded features are directly relevant to instance-level retrieval. To adapt the representation to the task at hand, we propose to train a single linear layer (projection) on top of frozen backbones. The recently introduced Universal Embeddings (UnED) dataset [72] is used for learning the linear adaptation. UnED contains images from 8 different domains with fine-grained and/or instance-level class annotation. In our experiments, the linear layer that converts the backbone output to a 512D descriptor is trained on a uniformly sampled subset of 1M images from UnED. The linear adaptation layer is trained with the UJCDs [72] method. + +Text-to-image retrieval. Text-to-image retrieval is performed using Vision-Language Models (VLMs) trained to align the two modalities. Retrieval is performed based on cosine similarity between the text query and db image descriptors that are extracted using the textual and visual encoder, respectively. We evaluate 17 VLM models. + +Re-ranking with global representations. Such methods rely on global descriptors for exhaustive search during the initial ranking, but also for a second refinement stage that issues a new query. We experiment with $\alpha$ QE [40], the adaptive variant of average Query Expansion [10]. After the initial ranking, the descriptors of the top-ranked images are aggregated with the query via weighted average pooling. The weights are derived from the similarity to the query in the power of $\alpha$ . We don't have a validation set; hence, we use a fixed value $\alpha = 1$ . + +Re-ranking with local representations. These re-ranking methods rely on global descriptors for exhaustive search during the initial ranking but estimate query-to-db image similarity based on local descriptors for a second refinement stage of the ranked list of images. We experiment with three methods: (i) Chamfer Similarity (CS) [3, 44] on the similarity matrix between local descriptors across the image pair. We use the asymmetric variant of CS with max over db descriptors and sum over query descriptors. (ii) Spatial Verification (SP) [6, 15, 37], a common re-ranking method where point correspondences are processed with a RANSAC-like + +![](images/4ce71404f66387c321e50c94fc99e008a90563dca08aad2b6d38d93f372d00f2.jpg) +Figure 4. Comparison with other instance-level retrieval datasets via reporting mAP@1k. INSTRE: 27.3K db size, multi-domain. GLDv2: 762K db size, single-domain. SOP: 60.5K db size, single-domain. Different network types are color-coded. For GLDv2 and SOP, models fine-tuned on these domains with the corresponding training sets are highlighted. No linear adaptation is used. + +![](images/e7c5074ef4c2bcc30f91ca2cf784aef44481d5c23fe4d25dd109ec16d8c3e356.jpg) + +![](images/40a3c2b659d85d707ea1775a60ba6be002b095d2cb8ed37f71020b047fb89ede.jpg) + +process and the number of inliers is used for re-ranking. (iii) AMES [56], a recent transformer-based network to estimate the similarity between sets of local descriptors. Due to the scale of ILIAS database, we use only 100 binary local descriptors for each database image and 600 for the query image. Local descriptors are extracted using the base variant of DINOv2 with registers [11, 34] and selected based the local descriptor detector used in AMES [56]. Top-1k retrieved images are re-ranked. + +# 5. Experiments + +We evaluate all the above models and methods, extracting useful insights regarding the factors that boost retrieval performance. ILIAS is compared with other existing datasets for instance-level image retrieval. We analyze the performance of selected models1 to break down the impact of different ILIAS attributes, such as domains, clutter, and scale. Unless stated otherwise, we use the large model variants with the largest resolution available, e.g. in our analysis we use SigLIP ViT-L trained with 384 resolution. + +# 5.1. Comparison with other instance-level datasets + +In Fig. 4, ILIAS is compared with other instance-level retrieval datasets based on evaluation of the same models. Linear adaptation is not applied as parts of GLDv2 and SOP are included in the UnED dataset. Only for the sake of this comparison, and for no other experiment in this work, we use models fine-tuned on specific domains (in-domain models), i.e. on the training sets of SOP and GLDv2. INSTRE, which is also multi-domain, shows a correlation to ILIAS, but its performance is saturated due to its small size. For single-domain datasets, in-domain models outperform others by a large margin, with few exceptions, i.e. DINOv2, which includes the trainset of GLDv2 in its training data. Several multi-domain models perform well on SOP since their training set is usually included in the training data. However, in-domain and multi-domain models face challenges on ILIAS, highlighting the diversity of our dataset. + +# 5.2. Method comparison + +Image-to-image retrieval with global representations. Tab. 2 presents the performance of global descriptor models on ILIAS. Selected models are presented to highlight useful comparisons, while many other models are included in the supplementary material. The main factors that improve performance are the size of the training set, training resolution, and model architecture, which aligns with the literature. The impact of dataset size is apparent in various model combinations, e.g. CLIP with openai and laion2b. This is also pronounced by the dominance of foundation models. Training with large resolution brings significant gains and consistently improves mAP@1k. In some cases of SigLIP, smaller models trained with large resolutions outperform larger ones trained with small resolutions. For the models of the same resolution, it is a common trend for larger model variants to bring corresponding performance gains. In general, VLMs perform the best. From non-VLMs, only DINOv2 and Unicom achieve competitive performance. Masked Image Modeling (MIM) and supervised models are not performing well. Our linear adaptation scheme is very effective, improving most models. The boost is more pronounced in the case of VLMs. A possible explanation for such improvements is that image-to-image relations are not optimized during the training of VLMs. + +Text-to-image retrieval. Following results in Tab. 2, similar conclusions are derived for the t2i case. Retrieval performance improves with the scaling of the training data. The larger model achieves significantly better results, i.e. compare the base with large variants. Finally, it is noteworthy that the best performance achieved by SigLIP2 is very close to the i2i performance when no adaptation is used. Note that t2i includes 1k text queries in total, with one query per object, while i2i 1,232 image queries. + +Evaluation of mini-ILIAS selection. Tab. 3 shows performance on mini-ILIAS for five models with linear adaptation. The selected subset is significantly more challenging than a random selection of 5M images. More precisely, a set of approximately $\sim 26\mathrm{M}$ random images matches the performance of mini-ILIAS. + +
modelarchtraindatasetdata sizetrain restest resimage-to-imagetext-to-image
5M†100M†100M100M5M
ResNet50 [17]R50supin1k1M2243842.51.81.7--
DINO [8]R50sslin1k1M2243844.12.92.9--
ConvNext [29]CN-Lsupin1k1M2883844.22.92.2--
OAI-CLIP [17]R50vlaopenai400M2243848.56.03.21.52.3
OpenCLIP [19, 29]CN-Bvlalaion2b2B25638418.114.07.94.67.0
OpenCLIP [19, 29]CN-Lvlalaion2b2B32051222.918.39.68.111.5
ViT [13, 54]ViT-Bsupin1k1M2243841.91.31.0--
EVA-MIM [14]ViT-Bsslin22k142M2243844.73.22.1--
ViT [13, 54]ViT-Bsupin21k14M2243846.24.43.0--
DINO [8]ViT-Bsslin1k1M2243846.64.83.7--
UDON-CLIP [73]ViT-Bsupuned2.8M2243849.26.75.9--
OAI-CLIP [41]ViT-Bvlaopenai400M22438410.77.94.21.62.7
EVA-CLIP [57]ViT-Bvlamerged2b2B22438411.78.75.92.54.4
MetaCLIP [70]ViT-Bvla2pt5b2.5B22438412.79.46.64.97.6
DINOv2 [34]ViT-Bssllvd142m142M51872415.012.111.5--
SigLIP [75]ViT-Bvlawebli10B25638420.616.711.57.510.3
SigLIP [75]ViT-Bvlawebli10B38451226.221.515.611.014.4
SigLIP [75]ViT-Bvlawebli10B51272427.523.016.611.114.6
SigLIP2 [65]ViT-Bvlawebli10B51272428.623.515.410.414.6
EVA-MIM [14]ViT-Lsslin22k14M2243843.92.71.5--
ViT [13, 54]ViT-Lsupin21k14M2243847.35.34.6--
EVA-MIM [14]ViT-Lsslmerged38m38B2243848.86.14.7--
OAI-CLIP [41]ViT-Lvlaopenai400M22438415.811.97.04.66.7
OpenCLIP [9, 19]ViT-Lvlalaion2b2B22438417.513.79.47.09.4
Unicom [1]ViT-Ldistlaion400m400M33651218.614.613.9--
OAI-CLIP [41]ViT-Lvlaopenai400M33651219.915.29.45.88.4
DINOv2 [34]ViT-Lssllvd142m142M51872418.815.315.3--
EVA-CLIP [57]ViT-Lvlamerged2b2B33651220.916.010.97.210.6
MetaCLIP [70]ViT-Lvla2pt5b2.5B22438421.716.911.79.213.1
SigLIP [75]ViT-Lvlawebli10B25638426.321.815.212.816.4
SigLIP [75]ViT-Lvlawebli10B38451234.328.919.618.122.2
SigLIP2 [65]ViT-Lvlawebli10B51272437.331.320.819.824.7
+ +Table 2. Performance comparison using mAP@1k on ILIAS and mini-ILIAS for global representation models for i2i and t2i. Comparison of model architecture (arch), training scheme (train), training data, and train/test resolution. $\dagger$ indicates results with the linear adaptation. 5M and 100M correspond to the mini and full versions of the dataset, respectively. sup, ssl, dist, vla: supervised learning, self-supervised learning, distillation and vision-language alignment. R50, CN: ResNet50 and ConvNext. + +
model100M5M-mini5M-rand
DINOv2† [34]15.318.822.7±0.2
EVA-CLIP† [57]16.020.928.8±0.2
MetaCLIP† [70]16.921.729.2±0.1
OpenCLIP† [19, 29]18.322.930.9±0.2
SigLIP† [75]28.934.341.8±0.1
+ +Table 3. A challenging distractor subset for mini-ILIAS. mAP@1k evaluated for different distractor sets, 100M: the full dataset, 5M-mini: mini-ILIAS subset, 5M-rand: random subset. We report the mean and std of 3 randomly sampled subsets. $\dagger$ indicates results with the linear adaptation. + +
rerankingSigLIP [75]SigLIP† [75]
mAP@1koraclemAP@1koracle
global19.648.728.956.0
αQE1 [10, 40]22.144.733.756.9
αQE2 [10, 40]20.440.831.554.4
αQE5 [10, 40]14.334.923.549.3
CS [44]22.948.732.556.0
SP [37]21.848.730.556.0
AMES [56]26.448.735.656.0
+ +Table 4. Performance comparison for re-ranking methods. Oracle represents the performance of perfect re-ranking at the top-1k images. Top: query expansion with global descriptors. Bottom: re-ranking with local descriptors. †: results with linear adaptation. + +Retrieval with re-ranking. Tab. 4 shows the performance of re-ranking methods applied on top of SigLIP with and without linear adaptation on ILIAS. Complementary to $\mathrm{mAP}@1\mathrm{k}$ , an oracle-based top-1k re-ranking metric is reported as the upper bound of a re-ranking method that processes the top 1k images. Local similarity estimated by a learned model proves to be very effective for re-ranking. Nevertheless, the oracle re-ranking performance indicates that there is a lot more space for improvements. Re-ranking with QE is useful when the number of aggregated neighbors is low and drops below the baseline when the number of neighbors is increased. Notably, global re-ranking affects and, interestingly, decreases oracle performance since the whole db is re-ranked; while local re-ranking does not affect it since it is performed only on a shortlist of images. + +# 5.3. Analysis + +Performance per domain. Fig. 5 shows the performance per taxonomy categories. The taxonomy annotations allow a fine-grained view of the results, which can possibly allow us to capture imbalanced improvements in future work. For example, DINOV2, despite being overall inferior to SigLIP, is outperforming it in categories like architecture and sculptures or is quite similar in categories like public art and paper art. This is possibly attributed to the curation and com + +![](images/0317cd1cc7d997e82893f6d6434525ce3d95ec7803efde0fd5563d2649867c53.jpg) +Figure 5. Performance comparison per category. mAP@1k averaged over objects in the same mid-level taxonomy category, organized by their primary-level category size, with sorting within each group by SigLIP† + AMES performance. Comparison between SigLIP with and without adaptation, SigLIP combined with AMES reranking, SigLIP t2i, and DINOv2. † indicates results with the linear adaptation. + +![](images/5397c4df56f4103ac2489ba3f99c0326cd4a35faeceb0dabb7baf6a9e023db0d.jpg) +Figure 6. Performance comparison reporting AP per query for different approaches with SigLIP. Pearson correlation reported in parenthesis. $\dagger$ indicates results with the linear adaptation. + +position of the DINOV2 training set, which includes artwork and landmark datasets. Also, some categories are hurt by reranking with AMES, with some demonstrating big drops, i.e. sport, gaming, perfume. These categories deviate significantly from the domain AMES is trained, i.e. landmarks, which could potentially justify such drops. + +Per query comparisons. Fig. 6 shows the AP per query for various methods. Linear adaptation boosts most queries, i.e. performance drop only for 192 queries. Image- and text-based retrieval are not strongly correlated despite performing similarly, which is good evidence [60] for the effectiveness of model ensembles. Indeed, ensembling i2i and t2i by averaging similarities brings $+6.1$ improvement over i2i retrieval. Query expansion improves the queries with at least some positives at top positions, i.e. AP greater than 20. However, it harms many low-performing queries by aggregating descriptors irrelevant to the query. AMES improves the majority of the queries; however, many are harmed, indicating that there is plenty of room for improvement. + +![](images/bdf3609aaa66228e22da5ab294907ec2faef93ce98c4dc4d2e9dbf3bf7e0345f.jpg) +Figure 7. Performance evaluation (mAP@1k) across different amounts of object area coverage and background clutter. Positives across all queries are jointly ranked based on coverage or clutter and split into 4 equal size groups. Queries with no positive in the corresponding group are discarded. No. of queries per group is in parentheses. $\dagger$ indicates results with the linear adaptation. + +Impact of clutter and scale in positives. To quantify the impact of background clutter and scale changes, Fig. 7 presents the performance for different groups of positives. Dealing with small objects and multi-object scenes form major weaknesses of existing models. Notably, t2i beats i2i without adaptation in small-scale groups. + +# 6. Conclusions + +We introduce ILIAS and conduct an extensive evaluation of current foundational models and retrieval methods, highlighting that instance-level retrieval remains an unsolved problem. Our results indicate that off-the-shelf application of foundational models leaves considerable room for improvement, particularly in handling small objects and complex backgrounds. While specialized retrieval methods leveraging local descriptors are effective in these cases, their high memory and computational costs become impractical at the scale of ILIAS or beyond. ILIAS is designed to become a standard benchmark for evaluating foundational representation models and retrieval methods, accommodating both global and local representations, and advancing the field of instance-level recognition. + +# Acknowledgments + +Funding acknowledgement. This work was supported by the Junior Star GACR (grant no. GM 21-28830M), Horizon MSCA-PF (grant no. 101154126), Programme Johannes Amos Comenius (grant no. CZ.02.01.01/00/22.010/0003405), CTU in Prague (grant no. SGS23/173/OHK3/3T/13), the CTU institutional support (Future fund). JM was supported by Ministry of the Interior of the Czech Republic (project no. VJ02010041). We acknowledge VSB - Technical University of Ostrava, IT4Innovations National Supercomputing Center, Czech Republic, for awarding this project access to the LUMI supercomputer, owned by the EuroHPC Joint Undertaking, hosted by CSC (Finland) and the LUMI consortium through the Ministry of Education, Youth and Sports of the Czech Republic through the e-INFRA CZ (grant ID: 90254). + +Contributors acknowledgments. We want to thank Larysa Ivashechkina for her work on the annotation of object bounding boxes and masks and the initial filtering of text queries. We also want to thank all contributors for the collection of ILIAS dataset. Additionally to all authors, the list of external contributor in alphabetical order: Aggeliki Tserota, Anna Nisyraiou, Celeste Abreu, Charalambos Tzamos, Christina Tserota, Dimitris Karageorgiou, Dmytro Mishkin, Eleni Karantali, Eleni Papadopoulou, Eva Tsiliakou, Kelly Kordopati-Zilou, Markos Zampoglou, Noa Garcia, Panagiotis Tassis, Paraskevas Kordopatis, Pavlos Alexantonakis, Ruslan Rozumnyi, Snežana Čurguz, Tereza Nejedlá, Tomáš Jelinek, Yannis Kalantidis, Vasilis Alexiadis, Yankun Wu. + +# References + +[1] Xiang An, Jiankang Deng, Kaicheng Yang, Jaiwei Li, Ziyong Feng, Jia Guo, Jing Yang, and Tongliang Liu. Unicom: Universal and compact representation learning for image retrieval. In ICLR, 2023. 2, 5, 7 +[2] Relja Arandjelović and Andrew Zisserman. Smooth object retrieval using a bag of boundaries. In ICCV, 2011. 2, 3 +[3] Harry G Barrow, Jay M Tenenbaum, Robert C Bolles, and Helen C Wolf. Parametric correspondence and chamfer matching: Two new techniques for image matching. In *IJ-CAI*, 1977. 5 +[4] Wissam Bejjani, Wisdom C Agboh, Mehmet R Dogar, and Matteo Leonetti. Occlusion-aware search for object retrieval in clutter. In IROS, 2021. 1 +[5] Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselat, Emma Brunskill, et al. On the opportunities and risks of foundation models. In arXiv, 2021. 1, 5 +[6] Bingyi Cao, André Araujo, and Jack Sim. Unifying deep local and global features for image search. In ECCV, 2020. 5 + +[7] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, 2020. 5 +[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 5, 7 +[9] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In CVPR, 2023. 5, 7 +[10] Ondrej Chum, James Philbin, Josef Sivic, Michael Isard, and Andrew Zisserman. Total recall: Automatic query expansion with a generative feature model for object retrieval. In ICCV, 2007. 2, 5, 7 +[11] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. In ICLR, 2024. 6 +[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 2, 5 +[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICML, 2021. 5, 7 +[14] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. EVA: Exploring the limits of masked visual representation learning at scale. In CVPR, 2023. 2, 5, 7 +[15] Martin A. Fischler and Robert C. Bolles. Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 1981. 5 +[16] Christopher Flagg and Ophir Frieder. Reconstruction of artifacts from digital image repositories. JOCCH, 2022. 1 +[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 5, 7 +[18] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. 5 +[19] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. 7 +[20] Hervé Jégou, Matthijs Douze, and Cordelia Schmid. Hamming embedding and weak geometric consistency for large scale image search. In ECCV, 2008. 2, 3 +[21] Tomas Jenicek and Ondrej Chum. No fear of the dark: Image retrieval under varying illumination conditions. In CVPR, 2019. 1 + +[22] Sungyeon Kim, Boseung Jeong, and Suha Kwak. HIER: Metric learning beyond class labels via hierarchical regularization. In CVPR, 2023. 5 +[23] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In CVPR, 2023. 4 +[24] Byungsoo Ko, Minchul Shin, Geonmo Gu, HeeJae Jun, Tae Kwan Lee, and Youngjoon Kim. A benchmark on tricks for large-scale image retrieval. In arXiv, 2019. 2 +[25] Tarun Krishna, Kevin McGuinness, and Noel O'Connor. Evaluating contrastive models for instance-based image retrieval. In ICMR, 2021. 2 +[26] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In NeurIPS, 2012. 5 +[27] Seongwon Lee, Hongje Seong, Suhyeon Lee, and Euntai Kim. Correlation verification for image retrieval. In CVPR, 2022. 5 +[28] Ziwei Liu, Ping Luo, Shi Qiu, Xiaogang Wang, and Xiaou Tang. Deepfashion: Powering robust clothes recognition and retrieval with rich annotations. In CVPR, 2016. 2, 3, 4 +[29] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. CVPR, 2022. 5, 7 +[30] Shiyang Lu, Haonan Chang, Eric Pu Jing, Abdeslam Boularias, and Kostas Bekris. OVIR-3D: Open-vocabulary 3d instance retrieval without training on 3d data. In CoRL, 2023. 1 +[31] Claudio Michaelis, Matthias Bethge, and Alexander Ecker. One-shot segmentation in clutter. In ICML, 2018. 1 +[32] David Nister and Henrik Stewenius. Scalable recognition with a vocabulary tree. In CVPR, 2006. 2, 3 +[33] Hyeonwoo Noh, Andre Araujo, Jack Sim, Tobias Weyand, and Bohyung Han. Large-scale image retrieval with attentive deep local features. In ICCV, 2017. 2, 3 +[34] Maxime Oquab, Timothee Darct, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINOv2: Learning robust visual features without supervision. TMLR, 2024. 1, 2, 5, 6, 7 +[35] Yash Patel, Giorgos Tolias, and Jiří Matas. Recall@k surrogate loss with large batches and similarity mixup. In CVPR, 2022. 5 +[36] Jingtian Peng, Chang Xiao, and Yifan Li. RP2K: A large-scale retail product dataset for fine-grained image classification. In arXiv, 2020. 2, 3 +[37] James Philbin, Ondrej Chum, Michael Isard, Josef Sivic, and Andrew Zisserman. Object retrieval with large vocabularies and fast spatial matching. In CVPR, 2007. 2, 4, 5, 7 +[38] James Philbin, Ondrej Chum, Michael Isard, Josef Sivic, and Andrew Zisserman. Lost in quantization: Improving particular object retrieval in large scale image databases. In CVPR, 2008. 4 + +[39] Filip Radenović, Ahmet Iscen, Giorgos Tolias, Yannis Avrithis, and Ondrej Chum. Revisiting oxford and paris: Large-scale image retrieval benchmarking. In CVPR, 2018. 1, 2, 3 +[40] Filip Radenovic, Giorgos Tolias, and Ondrej Chum. Finetuning cnn image retrieval with no human annotation. PAMI, 2019. 2, 4, 5, 7 +[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 3, 5, 7 +[42] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, et al. SAM 2: Segment anything in images and videos. In ICLR, 2025. 4 +[43] Ali Sharif Razavian, Hossein Azizpour, Josephine Sullivan, and Stefan Carlsson. CNN features off-the-shelf: An astounding baseline for recognition. In CVPRW, 2014. 2 +[44] Ali S Razavian, Josephine Sullivan, Stefan Carlsson, and Atsuto Maki. Visual instance retrieval with deep convolutional networks. ITE Trans. on Media Technology and Applications, 2016. 5, 7 +[45] Taichi Sakaguchi, Akira Taniguchi, Yoshinobu Hagiwara, Lotfi El Hafi, Shoichi Hasegawa, and Tadahiro Taniguchi. Object instance retrieval in assistive robotics: Leveraging fine-tuned simsiam with multi-view images based on 3d semantic map. In IROS, 2024. 1 +[46] Mert Bulent Sariyildiz, Philippe Weinzaepfel, Thomas Lucas, Diane Larlus, and Yannis Kalantidis. UNIC: Universal classification models via multi-teacher distillation. In ECCV, 2024. 5 +[47] Konstantin Schall, Kai Uwe Barthel, Nico Hezel, and Klaus Jung. GPR1200: a benchmark for general-purpose content-based image retrieval. In MMM, 2022. 2, 3 +[48] Shihao Shao, Kaifeng Chen, Arjun Karpur, Qinghua Cui, André Araujo, and Bingyi Cao. Global features are all you need for image retrieval and reranking. In ICCV, 2023. 2, 5 +[49] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In ICLR, 2015. 5 +[50] Ivan Sipiran, Patrick Lazo, Cristian Lopez, Milagritos Jimenez, Nihar Bagewadi, Benjamin Bustos, Hieu Dao, Shankar Gangisetty, Martin Hanik, Ngoc-Phuong Ho-Thi, et al. SHREC 2021: Retrieval of cultural heritage objects. Computers & Graphics, 2021. 1 +[51] Arnold WM Smeulders, Marcel Worring, Simone Santini, Amarnath Gupta, and Ramesh Jain. Content-based image retrieval at the end of the early years. PAMI, 2000. 2 +[52] Chull Hwan Song, Jooyoung Yoon, Taebaek Hwang, Shunghyun Choi, Yeong Hyeon Gu, and Yannis Avrithis. On train-test class overlap and detection for image retrieval. In CVPR, 2024. 1 +[53] Hyun Oh Song, Yu Xiang, Stefanie Jegelka, and Silvio Savarese. Deep metric learning via lifted structured feature embedding. In CVPR, 2016. 1, 2, 3, 4 + +[54] Andreas Steiner, Alexander Kolesnikov, Xiaohua Zhai, Ross Wightman, Jakob Uszkoreit, and Lucas Beyer. How to train your vit? data, augmentation, and regularization in vision transformers. TMLR, 2021. 5, 7 +[55] Pavel Suma and Giorgos Tolias. Large-to-small image resolution asymmetry in deep metric learning. In WACV, 2023. 5 +[56] Pavel Suma, Giorgos Kordopatis-Zilos, Ahmet Iscen, and Giorgos Tolias. AMES: Asymmetric and memory-efficient similarity estimation for instance-level retrieval. In ECCV, 2024. 2, 6, 7 +[57] Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. EVA-CLIP: Improved training techniques for clip at scale. In arXiv, 2023. 2, 5, 7 +[58] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In CVPR, 2015. 5 +[59] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In ICML, 2019. 5 +[60] Ryan Theisen, Hyunsuk Kim, Yaoqing Yang, Liam Hodgkinson, and Michael W Mahoney. When are ensembles really effective? In NeurIPS, 2024. 8 +[61] Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. YFCC100M: The new data in multimedia research. Communications of the ACM, 2016. 3 +[62] Akihiko Torii, Relja Arandjelovic, Josef Sivic, Masatoshi Okutomi, and Tomas Pajdla. 24/7 place recognition by view synthesis. In CVPR, 2015. 1 +[63] Hugo Touvron, Andrea Vedaldi, Matthijs Douze, and Herve Jegou. Fixing the train-test resolution discrepancy. In NeurIPS, 2019. 5 +[64] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, 2021. 5 +[65] Michael Tschannen, Alexey Gritsenko, Xiao Wang, Muhammad Ferjad Naeem, Ibrahim Alabdulmohsin, Nikhil Parthasarathy, Talfan Evans, Lucas Beyer, Ye Xia, Basil Mustafa, et al. SigLIP 2: Multilingual vision-language encoders with improved semantic understanding, localization, and dense features. In arXiv, 2025. 5, 7 +[66] Shuang Wang and Shuqiang Jiang. INSTRE: A new benchmark for instance-level object retrieval and recognition. TOMM, 2015. 1, 2, 3 +[67] Tobias Weyand, André Araujo, Bingyi Cao, and Jack Sim. Google Landmarks Dataset v2 - A large-scale benchmark for instance-level recognition and retrieval. In CVPR, 2020. 1, 2, 3, 4 +[68] Ross Wightman, Hugo Touvron, and Hervé Jégou. Resnet strikes back: An improved training procedure in timm. In NeurIPSW, 2021. 5 +[69] Pengxiang Wu, Siman Wang, Kevin Dela Rosa, and Derek Hu. FORB: A flat object retrieval benchmark for universal image embedding. In NeurIPS, 2024. 3 + +[70] Hu Xu, Saining Xie, Xiaqing Ellen Tan, Po-Yao Huang, Russell Howes, Vasu Sharma, Shang-Wen Li, Gargi Ghosh, Luke Zettlemoyer, and Christoph Feichtenhofer. Demystifying clip data. In ICLR, 2024. 3, 5, 7 +[71] Nikolaos-Antonios Ypsilantis, Noa Garcia, Guangxing Han, Sarah Ibrahimi, Nanne Van Noord, and Giorgos Tolias. The MET dataset: Instance-level recognition for artworks. In NeurIPS, 2021. 3 +[72] Nikolaos-Antonios Ypsilantis, Kaifeng Chen, Bingyi Cao, Mario Lipovsky, Pelin Dogan-Schönberger, Grzegorz Makosa, Boris Bluntschli, Mojtaba Seyedhosseini, Ondrej Chum, and André Araujo. Towards universal image embeddings: A large-scale dataset and challenge for generic image representations. In ICCV, 2023. 1, 2, 3, 5 +[73] Nikolaos-Antonios Ypsilantis, Kaifeng Chen, Andre Araujo, and Ondrej Chum. UDON: Universal dynamic online distillation for generic image representations. In NeurIPS, 2024. 5, 7 +[74] Jiangbo Yuan, An-Ti Chiang, Wen Tang, and Antonio Haro. eProduct: A million-scale visual search benchmark to address product recognition challenges. In CVPRW, 2021. 1, 2, 3 +[75] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In ICCV, 2023. 1, 2, 3, 4, 5, 7 +[76] Xunlin Zhan, Yangxin Wu, Xiao Dong, Yunchao Wei, Minlong Lu, Yichi Zhang, Hang Xu, and Xiaodan Liang. Product: Towards weakly supervised instance-level product retrieval via cross-modal pretraining. In ICCV, 2021. 1, 2, 3 +[77] Liang Zheng, Yi Yang, and Qi Tian. Sift meets cnn: A decade survey of instance retrieval. PAMI, 2017. 2 +[78] Barret Zoph, Vijay Vasudevan, Jonathon Shlens, and Quoc V Le. Learning transferable architectures for scalable image recognition. In CVPR, 2018. 5 \ No newline at end of file diff --git a/2025/ILIAS_ Instance-Level Image retrieval At Scale/images.zip b/2025/ILIAS_ Instance-Level Image retrieval At Scale/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..dc0d334620816993d1d219866dd86582d9cd74cf --- /dev/null +++ b/2025/ILIAS_ Instance-Level Image retrieval At Scale/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b04aea3e9ae0272c0e33654b4424e973e5d8299764b515a63156ad05ff6629fa +size 799745 diff --git a/2025/ILIAS_ Instance-Level Image retrieval At Scale/layout.json b/2025/ILIAS_ Instance-Level Image retrieval At Scale/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..82e5159ef92de30ccd6ca52a47a17c479d240f23 --- /dev/null +++ b/2025/ILIAS_ Instance-Level Image retrieval At Scale/layout.json @@ -0,0 +1,7117 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 159, + 103, + 453, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 103, + 453, + 121 + ], + "spans": [ + { + "bbox": [ + 159, + 103, + 453, + 121 + ], + "type": "text", + "content": "ILIAS: Instance-Level Image retrieval At Scale" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 124, + 142, + 485, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 142, + 485, + 186 + ], + "spans": [ + { + "bbox": [ + 124, + 142, + 485, + 186 + ], + "type": "text", + "content": "Giorgos Kordopatis-Zilos Vladan Stojnic Anna Manko Pavel Šuma Nikolaos-Antonios Ypsilantis Nikos Efthymiadis Zakaria Laskar Jiří Matas Ondřej Chum Giorgos Tolias" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 182, + 199, + 425, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 199, + 425, + 214 + ], + "spans": [ + { + "bbox": [ + 182, + 199, + 425, + 214 + ], + "type": "text", + "content": "VRG, FEE, Czech Technical University in Prague" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 234, + 200, + 247 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 234, + 200, + 247 + ], + "spans": [ + { + "bbox": [ + 151, + 234, + 200, + 247 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 260, + 297, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 260, + 297, + 547 + ], + "spans": [ + { + "bbox": [ + 53, + 260, + 297, + 547 + ], + "type": "text", + "content": "This work introduces ILIAS, a new test dataset for Instance-Level Image retrieval At Scale. It is designed to evaluate the ability of current and future foundation models and retrieval techniques to recognize particular objects. The key benefits over existing datasets include large scale, domain diversity, accurate ground truth, and a performance that is far from saturated. ILIAS includes query and positive images for 1,000 object instances, manually collected to capture challenging conditions and diverse domains. Large-scale retrieval is conducted against 100 million distractor images from YFCC100M. To avoid false negatives without extra annotation effort, we include only query objects confirmed to have emerged after 2014, i.e. the compilation date of YFCC100M. An extensive benchmarking is performed with the following observations: i) models fine-tuned on specific domains, such as landmarks or products, excel in that domain but fail on ILIAS ii) learning a linear adaptation layer using multi-domain class supervision results in performance improvements, especially for vision-language models iii) local descriptors in retrieval re-ranking are still a key ingredient, especially in the presence of severe background clutter iv) the text-to-image performance of the vision-language foundation models is surprisingly close to the corresponding image-to-image case." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 548, + 276, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 548, + 276, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 548, + 276, + 559 + ], + "type": "text", + "content": "website: https://vrg.fel(cvut.cz/ilias/" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 572, + 135, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 572, + 135, + 585 + ], + "spans": [ + { + "bbox": [ + 55, + 572, + 135, + 585 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "content": "The ability to recognize and differentiate every unique object instance in the physical world represents one of the ultimate goals for foundation representation models [5, 34, 41, 75]. This work aims to assess this capability through the lens of instance-level image retrieval at a very large scale. Instance-level image retrieval corresponds to searching for images of particular objects within large collections. All images of a particular object form their own instance-level class. This is an important information retrieval task due to its numerous real-world applications in robotics [30, 45], e" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 315, + 232, + 555, + 376 + ], + "blocks": [ + { + "bbox": [ + 315, + 232, + 555, + 376 + ], + "lines": [ + { + "bbox": [ + 315, + 232, + 555, + 376 + ], + "spans": [ + { + "bbox": [ + 315, + 232, + 555, + 376 + ], + "type": "image", + "image_path": "3e869c247b7b01b1f9361e6e73886bb942599fafb8737dd642fe9cbc714c3a52.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 377, + 555, + 444 + ], + "lines": [ + { + "bbox": [ + 313, + 377, + 555, + 444 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 555, + 444 + ], + "type": "text", + "content": "Figure 1. Performance timeline on ILIAS. Curves indicate best performance in chronological order for image-to-image and text-to-image retrieval, showing a significant boost with the release of foundation models. Representations are linearly adapted via multi-domain learning on UnED [72]. Re-ranking with local descriptors achieves the best results by a significant margin." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 449, + 556, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 449, + 556, + 582 + ], + "spans": [ + { + "bbox": [ + 312, + 449, + 556, + 582 + ], + "type": "text", + "content": "commerce [74, 76], and cultural heritage [16, 50], to name just a few. The task faces challenges because of the substantial variations among positive examples, such as illumination/viewpoint [21, 62] changes and background clutter [4, 31]. An additional difficulty is the high similarity among negatives, which is driven by the extremely fine granularity in the class definitions. It becomes even more challenging at a real-world scale, where searching through millions or even billions of images requires handling an open-world setup with countless unseen objects spanning diverse and complex domains." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 582, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 582, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 582, + 557, + 715 + ], + "type": "text", + "content": "Benchmarking instance-level retrieval under real-world challenges is currently limited by the lack of suitable datasets. Constructing a dataset with instance-level class definitions necessitates huge development effort, reflected by the many shortcomings of existing datasets. Shortcomings exist in several key aspects, such as dataset size [66], domain diversity [39, 53], and ground-truth accuracy [67], which suffers from both false positives and false negatives. Popular datasets are typically limited to landmarks [39], and as dataset scale increases, ground-truth quality tends to decline [52, 67]. This is a consequence of automating the" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14777" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": "ground-truth creation process to facilitate scaling up. To address such limitations, we introduce the Instance-Level Image retrieval At Scale (ILIAS) evaluation dataset." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 110, + 295, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 110, + 295, + 252 + ], + "spans": [ + { + "bbox": [ + 54, + 110, + 295, + 252 + ], + "type": "text", + "content": "The creation of our dataset has two key elements. First, query and positive images are manually captured to ensure challenging variations, covering 1,000 objects across diverse domains. Second, to expand the dataset size without ground-truth errors or additional annotation effort, we leverage a key technique: distractor images, collected in 2014 from YFCC100M, are combined with query objects verified not to have publicly existed until after 2014. This distractor set includes 100 million images, two orders of magnitude larger than the largest existing dataset [39]. Notably, all images have a permissive license, allowing us to ensure long-term online availability to the full extent." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 254, + 295, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 254, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 254, + 295, + 361 + ], + "type": "text", + "content": "ILIAS includes both image and text queries. The latter is in the form of detailed descriptions of objects and their distinctive features. The dataset is designed to support future research in image-to-image and text-to-image retrieval for particular objects, and additionally serves as a large-scale benchmark for evaluating representations of foundation vision and language models (VLM) [41, 75]. To facilitate faster experimentation, we provide a mini, but challenging, version (5M) of the distractor set." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 363, + 295, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 363, + 295, + 448 + ], + "spans": [ + { + "bbox": [ + 55, + 363, + 295, + 448 + ], + "type": "text", + "content": "We perform an extensive evaluation comparison, including many foundation image-to-image and text-to-image models, and establish a comprehensive testbed that enables future comparisons. The provided evaluation includes retrieval with global image representation but also re-ranking techniques that use local representations [37, 43, 56] and query expansion [10, 40]. We observe the following:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 449, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 55, + 449, + 295, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 295, + 508 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 295, + 508 + ], + "type": "text", + "content": "- Performance of standard 10-year-old models, measured by mean Average Precision, is as low as " + }, + { + "bbox": [ + 55, + 449, + 295, + 508 + ], + "type": "inline_equation", + "content": "1.3\\%" + }, + { + "bbox": [ + 55, + 449, + 295, + 508 + ], + "type": "text", + "content": ", while the best-performing model achieves " + }, + { + "bbox": [ + 55, + 449, + 295, + 508 + ], + "type": "inline_equation", + "content": "31.3\\%" + }, + { + "bbox": [ + 55, + 449, + 295, + 508 + ], + "type": "text", + "content": ", as shown in Fig. 1. This points out the vast progress of representation models and the high challenging factors of ILIAS." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 508, + 222, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 508, + 222, + 520 + ], + "spans": [ + { + "bbox": [ + 55, + 508, + 222, + 520 + ], + "type": "text", + "content": "- VLMs are the top-performing models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 520, + 295, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 520, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 520, + 295, + 555 + ], + "type": "text", + "content": "- Smaller (ViT-B) models trained/tested on large resolution (512/724) outperform larger models (ViT-L) trained/tested on small resolution (256/384)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 555, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 295, + 628 + ], + "type": "text", + "content": "- Using Universal Embedding Dataset (UnED) [72] to learn a linear adaptation layer on top of frozen models improves performance of most models, making it a candidate training set to couple with ILIAS. Notably, VLMs demonstrate the largest benefits, presumably because their training stage does not optimize image-to-image relations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 628, + 295, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 628, + 295, + 663 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 295, + 663 + ], + "type": "text", + "content": "- In contrast to the current belief [48], local representation is a key ingredient, while global representation, despite being efficient and compact, performs much lower." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 663, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 663, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 663, + 295, + 712 + ], + "type": "text", + "content": "- The performance gap between image-to-image and text-to-image models is surprisingly small. Therefore, detailed text queries are a reasonable proxy in the absence of image queries, even at the instance level." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 72, + 399, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 399, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 399, + 83 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 91, + 553, + 114 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 553, + 114 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 553, + 114 + ], + "type": "text", + "content": "In this section, we review the related work in terms of existing datasets and benchmarks in the literature." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 115, + 555, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 115, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 115, + 555, + 498 + ], + "type": "text", + "content": "Datasets. Tab. 1 presents the datasets from the image retrieval literature related to ILIAS. The datasets can be compared based on five main axes: (i) Class definition adopted. Many datasets [20, 32, 33, 39, 67] adopt a strict definition very similar to ours, satisfying instance-level requirements. Others [2, 53, 66, 76] adopt a more relaxed definition, where some minor variations are permitted, e.g. color changes in objects of the same class. Even more relaxed are the fine-grained definitions [74], where the object of the very same type is considered related, e.g. same product with different variant. (ii) Domain of the dataset. Most datasets are tailored for a specific domain. Landmarks are among the most popular domains [2, 20, 33, 39, 67]. Other domains include products [32, 36, 76] and fashion [28, 53]. Some datasets cover multiple domains, either being standalone [66] or bundle of repurposed datasets [47, 72]. (iii) Scale of database. Most of the datasets are small-scale, counting a few thousand images [28, 66, 76]. Larger ones [39, 72, 74] expend slightly above a million. None satisfies large-scale requirements. (iv) Noise in ground truth. Most datasets consist of clean annotations, except for a few cases that contain inaccuracies, including false positives [67], i.e. images wrongly annotated as relevant, false negatives [32, 39], i.e. relevant images that have not been annotated as positives, or the possibility of false positives [39]. (v) Availability. Most datasets are publicly available with permissive licenses, with few exceptions of partial [2, 20] or no [33, 74] availability. To this end, no publicly available dataset fits the strict instance-level definition, contains objects from multiple domains, ensures error-free labeling and is large scale. This gap is filled with ILIAS satisfying all the aforementioned requirements." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 498, + 554, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 554, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 554, + 629 + ], + "type": "text", + "content": "Evaluation benchmarks. Benchmarking [51] tracks the progress in the field, which is even more necessary with the emergence of foundation models. Several benchmarks papers [24, 25, 77] exist in the instance-level retrieval literature, investigating the impact of learning scheme, post-processing, model ensembling, query expansion, and whitening. The most relevant benchmark to ILIAS is UnED [72] that combines existing datasets to create a union that evaluates models performance across various domains. Due to its wide variety, UnED serves as the training dataset for linear adaptation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 630, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 553, + 714 + ], + "type": "text", + "content": "Regarding the evaluation of foundation models, the most common practice [14, 34, 57] is measuring classification performance on top of frozen models on ImageNet [12]. This is performed either with or without the training of a classifier via linear probing or k-NN search. Furthermore, models are usually evaluated on dense prediction tasks [34] and several multiple-downstream single-domain tasks [1]." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14778" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 80, + 60, + 527, + 246 + ], + "blocks": [ + { + "bbox": [ + 80, + 60, + 527, + 246 + ], + "lines": [ + { + "bbox": [ + 80, + 60, + 527, + 246 + ], + "spans": [ + { + "bbox": [ + 80, + 60, + 527, + 246 + ], + "type": "table", + "html": "
datasetsyearobjectsquerypositivesdatabasegtclass def.domainbboxonlinelicense
UKB [32]20062.5K10K10K10KFNILproductXFullyN/A
Holidays [20]20085005009911MCleanILlandmarkXPartiallyCC
Sculptures [2]201110703.1K3.1KCleanPartial ILsculptureXPartiallyFlickr TC
INSTRE [66]2015200125027.3K27.3KCleanPartial ILmulti✓FullyFlickr TC
SOP [53]201511.3K60.5K60.5K60.5KCleanPartial ILproductXFullyMIT License
InShop [28]20163.9K14.2K12.6K12.6KCleanPartial ILfashion✓FullyN/A
R-Oxford [39]201811705K1MFN?ILlandmarkXFullyFlickr TC, CC
R-Paris [39]201811706.3K1MFN?ILlandmarkXFullyFlickr TC, CC
GLDv1 [33]201830KN/AN/A1.1MCleanILlandmarkXPartiallyMultiple
GLDv2 [67]20203181.1K3.1K762KFPILlandmarkXFullyCC/ Public-domain
Product1M [76]20213926.5K40K40KCleanPartial ILproductXPartiallyN/A
RP2K [36]20211.2K10.9K10.9K10.9KCleanPartial ILproduct✓FullyN/A
GPR1200 [47]20211.2K12K12K12KMixIL+FGmultiXFullyMultiple
eProduct [74]202120610KN/A1.1MN/AFGproductXNoN/A
FORB [69]2023N/A13.9K4.5K49.8KCleanILplanarXFullySnap Inc.
UnED [72]202321K241K244K1.4MMixIL+FGmultiXFullyMultiple
ILIAS20251,0001,2324,715100MCleanILmulti✓FullyCC
", + "image_path": "bd1613d8e9182d7c1e32abac76db6d91abdd4dddc940b275138132fdc55b619e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 249, + 555, + 293 + ], + "lines": [ + { + "bbox": [ + 55, + 249, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 249, + 555, + 293 + ], + "type": "text", + "content": "Table 1. Comparison with other instance-level datasets. Datasets are compared based on their size (object, query, positives, database), the accuracy of the ground truth (gt), type of class definition, domain, supplementary annotations (bbox) and accessibility (online, license). N/A: not available. FP/FN: false positives/negatives. FN?: possibility of false negatives. Mix: combination of clean and noisy datasets. IL: instance-level. FG: fine-grained. Partial IL: instance-level with subtle variations among same class objects. CC: Creative Commons." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 300, + 294, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 294, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 294, + 384 + ], + "type": "text", + "content": "For VLMs, zero-shot classification and retrieval serve as the primary benchmarks [41, 70, 75], utilizing class text labels. In this work, we provide similar evaluation protocols tailored for instance-level retrieval. One can test the raw model capabilities or adapt for the instance-level task via linear adaptation on UnED. Text-to-image retrieval is also facilitated for the evaluation of VLMs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 398, + 144, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 398, + 144, + 410 + ], + "spans": [ + { + "bbox": [ + 55, + 398, + 144, + 410 + ], + "type": "text", + "content": "3. ILIAS dataset" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 418, + 205, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 205, + 431 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 205, + 431 + ], + "type": "text", + "content": "3.1. Composition and collection" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 437, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 295, + 544 + ], + "type": "text", + "content": "Instance-level class definition. Following an instance-level class definition [39, 71], we consider all indistinguishable object instances of the real world to form their own class. Nevertheless, we add a restriction to consider a pair of images as relevant to each other only if there is a view overlap. Other cases are explicitly not included in the dataset, contrasting the existing work [28, 53, 67]. Therefore, models should mostly rely on estimating the visual similarity and less on shortcuts through semantics." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 544, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 641 + ], + "type": "text", + "content": "Overview. ILIAS supports both image-to-image (i2i) and text-to-image (t2i) retrieval and follows the standard setup for retrieval datasets, consisting of two main parts: (i) query images and text, and (ii) database (db) images. The objective is to rank positives - db images relevant to the query - at the top ranks. The collected objects cover a wide range of categories and are not restricted to specific domains. An overview of some collected objects is provided in Fig. 2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": "Queries and positives are created/collected by a group of collectors that are well-informed about the task objectives. In addition to positives, in the database, we include numerous distractors - irrelevant (negative) images to the queries - that make retrieval more challenging. Following previous work [39], adding a large, uncurated set of ran" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 300, + 555, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 555, + 361 + ], + "type": "text", + "content": "dom images achieves this. The larger the set, the higher the chances of hard negatives – images with similar appearance or semantics to the queries. To this end, we select the YFCC100M [61] dataset to serve as the source of distractors due to its size and permissive license." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 363, + 556, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 363, + 556, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 556, + 532 + ], + "type": "text", + "content": "Selected objects. Ensuring that distractor images include no false negatives cannot be performed in a scalable or accurate way if one relies on human annotation or metadata. Instead, we take advantage of the fact that YFCC100M was crawled from Flickr in 2014. Hence, an object qualifies in ILIAS if it could not have appeared on Flickr before 2014. To verify this, we rely either on publicly available information, e.g. objects known to be created/manufactured after 2014, or on the collector's knowledge about the object not being publicly available. Additionally, we opt for objects with distinctive and unique features that set them apart from others within the same category. For example, we avoid recent smartphones that look like plain black screens or new objects with distinctive parts closely resembling older ones." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "text", + "content": "Queries and positives. Query images depict the instance on a clean or uniform background. When this is not feasible (e.g. buildings or statues), background blurring or cropping is applied. This is performed to avoid including background objects in the query that do not have corresponding positives in our ground truth information. Positives are images featuring the query object in challenging conditions, such as clutter, scale changes, occlusions, and partial views. Prior work [39] reveals that easy positives dominate performance metrics. Thus, we specifically opt for challenging cases that cannot be easily retrieved by the models. To avoid taking advantage of camera identification, most query and positive images are captured with at least two different camera models to introduce diversity. We also incorporate older camera models that are used in YFCC100M." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14779" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 93, + 69, + 512, + 253 + ], + "blocks": [ + { + "bbox": [ + 93, + 69, + 512, + 253 + ], + "lines": [ + { + "bbox": [ + 93, + 69, + 512, + 253 + ], + "spans": [ + { + "bbox": [ + 93, + 69, + 512, + 253 + ], + "type": "image", + "image_path": "b80f76ca47b653aad9f0f85ba98c45ce40709240a8b2314ae451ca72300b8782.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 255, + 555, + 277 + ], + "lines": [ + { + "bbox": [ + 55, + 255, + 555, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 555, + 277 + ], + "type": "text", + "content": "Figure 2. Examples of query, positive and hard negatives within the distractor set. Average Precision per query and rank of the negatives and positives is reported using SigLIP [75] model. Gray: queries. Green: positives. Red: distractors." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 281, + 295, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 281, + 295, + 353 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 295, + 353 + ], + "type": "text", + "content": "Each text query consists of a detailed and fine-grained textual description of an object. Descriptions are initially created by a large language model prompted to provide highly detailed depictions of the object shown in query images. Generated descriptions are manually edited to fix errors, insufficient descriptions, or nuances of the model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 353, + 296, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 353, + 296, + 413 + ], + "spans": [ + { + "bbox": [ + 55, + 353, + 296, + 413 + ], + "type": "text", + "content": "Distractors. The YFCC100M dataset was chosen for the distractor set due to its large scale and diverse range of concepts. It consists of 100 million Flickr images, collected without specific filtering, aside from being shared under a permissive CC-BY license." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 413, + 296, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 413, + 296, + 484 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 296, + 484 + ], + "type": "text", + "content": "Bounding box annotation. We include supplementary bounding boxes that specify the precise location of objects in query and positive images. They provide statistics about the position and size of object areas, assist our analysis of the dataset challenges, and support future research in instance-level localization." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 485, + 296, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 485, + 296, + 570 + ], + "spans": [ + { + "bbox": [ + 55, + 485, + 296, + 570 + ], + "type": "text", + "content": "Evaluation metric. Retrieval performance is evaluated via mean Average Precision (mAP), a widely used metric in instance-level image retrieval [37, 38, 40]. Specifically, we adopt mAP@1k [67], which assesses the ranking of the top-1k nearest neighbors for each query, treating any positive not ranked among the top-1k as not retrieved. We estimate the area under the curve using rectangles and not trapezoids." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 576, + 120, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 120, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 120, + 588 + ], + "type": "text", + "content": "3.2. Statistics" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 594, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 677 + ], + "type": "text", + "content": "Dataset size. The final ILIAS dataset includes 1,000 object instances captured in 5,947 images, of which 1,232 are queries and 4,715 are positives. Fig. 3a shows the distribution of positives per object. Also, 99,144,315 images from YFCC100M are downloaded. All images (queries, positives, distractors) are transferred through Flickr to ensure the same pre-processing." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "Taxonomy. A hierarchical 3-level taxonomy is composed for ILIAS. All instances are assigned across one to three categories of different granularity levels. The taxonomy" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 281, + 555, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 555, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 555, + 473 + ], + "type": "text", + "content": "consists of 8 categories on the coarser level, 42 on the mid level, and 38 on the finer level. The categories are derived through manual labeling of the objects based on their semantic content. To form the coarser-level categories, we use domain definitions borrowed from prior work [28, 53, 67] to align with the literature, i.e. art, landmarks, products, fashion. We also define novel categories based on the objects that do not fit into any existing domain. The distribution of objects across categories is uneven, e.g. ranging from 168 and 162 for art and landmarks to 83 for products. Each mid- and finer-level category contains at least 4 instances. Note that taxonomy is given to provide statistics about the domains of objects and assist our analysis instead of being leveraged as ground truth. The distribution of taxonomy categories can be inferred by Fig. 5, and a detailed figure is provided in the supplementary material." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 474, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 556, + 714 + ], + "type": "text", + "content": "Bounding box analysis. A total number of 6,117 bounding boxes are annotated for both queries and positives. Note that positives may display multiple objects of near identical appearance to the query; in such cases, bounding boxes are drawn on all indistinguishable objects. There are 235 images with more than one bounding box. Based on the annotated bounding boxes, we compute the area covered in the image by the object instances to derive its relative scale. Fig. 3c shows the distribution of the scale ratio for queries and positives. Most objects in queries cover the largest area of the images; while in the vast majority of positives, the object covers a small area of less than half the image. It is a result of the severe scale changes and partial views. Moreover, we use the Segment Anything Model (SAM) [23, 42] to extract object segments from positives. The number of detected segments outside the query object's bounding boxes is computed. This indicates clutter from other items in the positives. Fig. 3b shows the segment number distribution, with most images containing multiple segments due to clutter." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14780" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 73, + 136, + 160 + ], + "blocks": [ + { + "bbox": [ + 56, + 73, + 136, + 160 + ], + "lines": [ + { + "bbox": [ + 56, + 73, + 136, + 160 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 136, + 160 + ], + "type": "image", + "image_path": "682c2ec5fdb3ff2516d4dcacbea8a3efb6b4bd7ac55b67b3b003a252a840c08e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 92, + 161, + 102, + 169 + ], + "lines": [ + { + "bbox": [ + 92, + 161, + 102, + 169 + ], + "spans": [ + { + "bbox": [ + 92, + 161, + 102, + 169 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 136, + 73, + 214, + 160 + ], + "blocks": [ + { + "bbox": [ + 136, + 73, + 214, + 160 + ], + "lines": [ + { + "bbox": [ + 136, + 73, + 214, + 160 + ], + "spans": [ + { + "bbox": [ + 136, + 73, + 214, + 160 + ], + "type": "image", + "image_path": "7d23c8a6223101fd7452578cac9c2d0217afbe8dc317f95a94a3584b8d8355ba.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 172, + 161, + 183, + 169 + ], + "lines": [ + { + "bbox": [ + 172, + 161, + 183, + 169 + ], + "spans": [ + { + "bbox": [ + 172, + 161, + 183, + 169 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 214, + 73, + 291, + 160 + ], + "blocks": [ + { + "bbox": [ + 214, + 73, + 291, + 160 + ], + "lines": [ + { + "bbox": [ + 214, + 73, + 291, + 160 + ], + "spans": [ + { + "bbox": [ + 214, + 73, + 291, + 160 + ], + "type": "image", + "image_path": "87fe4f25dbb74f5839ab093adcaadd6bb9fd1fc582a1e8708babd9a956df2551.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 249, + 161, + 259, + 169 + ], + "lines": [ + { + "bbox": [ + 249, + 161, + 259, + 169 + ], + "spans": [ + { + "bbox": [ + 249, + 161, + 259, + 169 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 170, + 295, + 202 + ], + "lines": [ + { + "bbox": [ + 55, + 170, + 295, + 202 + ], + "spans": [ + { + "bbox": [ + 55, + 170, + 295, + 202 + ], + "type": "text", + "content": "Figure 3. ILIAS statistics. (a) number of positives per object, (b) positive distribution by the SAM segments outside the bounding box, (c) image distribution by the relative bounding box area." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 209, + 129, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 209, + 129, + 220 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 129, + 220 + ], + "type": "text", + "content": "3.3. mini-ILIAS" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 226, + 296, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 226, + 296, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 226, + 296, + 430 + ], + "type": "text", + "content": "We provide a small version of ILIAS, called mini-ILIAS, to facilitate quick experimentations. It consists of the query and positive images collected for ILIAS, and a subset of the YFCC100M distractors. Instead of randomly subsampling YFCC100M, we construct a challenging subset with the help of VLMs. We aim at selecting distractors displaying objects of similar categories as the query objects. We use the text category labels of the taxonomy as text queries. We also extend them with standard templates used for zero-shot recognition [41], which resulted in several thousands. T2i similarity between each text query and each distractor image is estimated. A similarity score for each distractor is derived based on its maximum similarity over the text queries. We ensemble the scores of 3 VLMs to rank images. The top-5M ranked distractors compose the final mini-ILIAS. Our experiments indicate that this subset is significantly more challenging than a random subset of the same size." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 441, + 176, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 176, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 176, + 453 + ], + "type": "text", + "content": "4. Benchmark methods" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 462, + 295, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 295, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 295, + 545 + ], + "type": "text", + "content": "We describe the methods and foundation models we evaluate on ILIAS, which are grouped according to their type of representations used for retrieval in global (i2i) representations, re-ranking with global (i2i) representations, re-ranking with local (i2i) representations, and text-to-image. A detailed list of models, their performance, and implementation details are in the supplementary materials." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "text", + "content": "Image-to-image retrieval with global representations. Global representation methods use image encoders to map images to global descriptors and rank db images based on cosine similarity. We evaluate legacy and recent foundation models, varying in architecture, descriptor dimensionality, training scheme, training data, and input resolution. Foundation models [5] are the models trained with a training set on the scale of a hundred million. Particularly, 23 CNN [17, 26, 29, 49, 58, 59, 78] and 45 ViT [13, 14] models, trained with supervision [22, 54, 64, 68], self-supervision [7, 8, 18, 34], distillation [1, 46, 64], or visual-language alignment [9, 41, 57, 65, 70, 75] are benchmarked. Most of the non-foundation models are trained on ImageNet [12]. There are models trained on single specific do" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": "mains [27, 35, 48], i.e. landmarks or products on GLDv2 or SOP. Universal models [1, 46, 72, 73] trained on multi-domains or multi-task schemes are included. The full list of models and results is provided in supplementary materials." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 121, + 555, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 555, + 217 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 555, + 217 + ], + "type": "text", + "content": "To mitigate the differences in training resolution, we use three widely-used resolutions, i.e. 384, 512, and 724 and resize images so that their larger dimension matches one of the three. The test resolution is defined to be one resolution above the one used for training, e.g. a network trained with 224 or 384 is tested with 384 or 512, respectively. The vast majority of models achieve best performance under this rule. Similar behavior is observed in the literature [55, 63]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 218, + 554, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 218, + 554, + 386 + ], + "spans": [ + { + "bbox": [ + 313, + 218, + 554, + 386 + ], + "type": "text", + "content": "Linear adaptation for i2i retrieval. Pre-trained foundation models, as well as legacy models, are trained to extract representations that are applicable to various tasks; not all encoded features are directly relevant to instance-level retrieval. To adapt the representation to the task at hand, we propose to train a single linear layer (projection) on top of frozen backbones. The recently introduced Universal Embeddings (UnED) dataset [72] is used for learning the linear adaptation. UnED contains images from 8 different domains with fine-grained and/or instance-level class annotation. In our experiments, the linear layer that converts the backbone output to a 512D descriptor is trained on a uniformly sampled subset of 1M images from UnED. The linear adaptation layer is trained with the UJCDs [72] method." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 388, + 554, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 554, + 459 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 554, + 459 + ], + "type": "text", + "content": "Text-to-image retrieval. Text-to-image retrieval is performed using Vision-Language Models (VLMs) trained to align the two modalities. Retrieval is performed based on cosine similarity between the text query and db image descriptors that are extracted using the textual and visual encoder, respectively. We evaluate 17 VLM models." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 460, + 554, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 554, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 554, + 579 + ], + "type": "text", + "content": "Re-ranking with global representations. Such methods rely on global descriptors for exhaustive search during the initial ranking, but also for a second refinement stage that issues a new query. We experiment with " + }, + { + "bbox": [ + 313, + 460, + 554, + 579 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 460, + 554, + 579 + ], + "type": "text", + "content": "QE [40], the adaptive variant of average Query Expansion [10]. After the initial ranking, the descriptors of the top-ranked images are aggregated with the query via weighted average pooling. The weights are derived from the similarity to the query in the power of " + }, + { + "bbox": [ + 313, + 460, + 554, + 579 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 460, + 554, + 579 + ], + "type": "text", + "content": ". We don't have a validation set; hence, we use a fixed value " + }, + { + "bbox": [ + 313, + 460, + 554, + 579 + ], + "type": "inline_equation", + "content": "\\alpha = 1" + }, + { + "bbox": [ + 313, + 460, + 554, + 579 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "text", + "content": "Re-ranking with local representations. These re-ranking methods rely on global descriptors for exhaustive search during the initial ranking but estimate query-to-db image similarity based on local descriptors for a second refinement stage of the ranked list of images. We experiment with three methods: (i) Chamfer Similarity (CS) [3, 44] on the similarity matrix between local descriptors across the image pair. We use the asymmetric variant of CS with max over db descriptors and sum over query descriptors. (ii) Spatial Verification (SP) [6, 15, 37], a common re-ranking method where point correspondences are processed with a RANSAC-like" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14781" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 114, + 58, + 249, + 182 + ], + "blocks": [ + { + "bbox": [ + 114, + 58, + 249, + 182 + ], + "lines": [ + { + "bbox": [ + 114, + 58, + 249, + 182 + ], + "spans": [ + { + "bbox": [ + 114, + 58, + 249, + 182 + ], + "type": "image", + "image_path": "4ce71404f66387c321e50c94fc99e008a90563dca08aad2b6d38d93f372d00f2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 185, + 555, + 220 + ], + "lines": [ + { + "bbox": [ + 55, + 185, + 555, + 220 + ], + "spans": [ + { + "bbox": [ + 55, + 185, + 555, + 220 + ], + "type": "text", + "content": "Figure 4. Comparison with other instance-level retrieval datasets via reporting mAP@1k. INSTRE: 27.3K db size, multi-domain. GLDv2: 762K db size, single-domain. SOP: 60.5K db size, single-domain. Different network types are color-coded. For GLDv2 and SOP, models fine-tuned on these domains with the corresponding training sets are highlighted. No linear adaptation is used." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 256, + 59, + 375, + 182 + ], + "blocks": [ + { + "bbox": [ + 256, + 59, + 375, + 182 + ], + "lines": [ + { + "bbox": [ + 256, + 59, + 375, + 182 + ], + "spans": [ + { + "bbox": [ + 256, + 59, + 375, + 182 + ], + "type": "image", + "image_path": "e7c5074ef4c2bcc30f91ca2cf784aef44481d5c23fe4d25dd109ec16d8c3e356.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 389, + 59, + 506, + 182 + ], + "blocks": [ + { + "bbox": [ + 389, + 59, + 506, + 182 + ], + "lines": [ + { + "bbox": [ + 389, + 59, + 506, + 182 + ], + "spans": [ + { + "bbox": [ + 389, + 59, + 506, + 182 + ], + "type": "image", + "image_path": "40a3c2b659d85d707ea1775a60ba6be002b095d2cb8ed37f71020b047fb89ede.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 228, + 297, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 228, + 297, + 335 + ], + "spans": [ + { + "bbox": [ + 54, + 228, + 297, + 335 + ], + "type": "text", + "content": "process and the number of inliers is used for re-ranking. (iii) AMES [56], a recent transformer-based network to estimate the similarity between sets of local descriptors. Due to the scale of ILIAS database, we use only 100 binary local descriptors for each database image and 600 for the query image. Local descriptors are extracted using the base variant of DINOv2 with registers [11, 34] and selected based the local descriptor detector used in AMES [56]. Top-1k retrieved images are re-ranked." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 345, + 138, + 359 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 345, + 138, + 359 + ], + "spans": [ + { + "bbox": [ + 55, + 345, + 138, + 359 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 365, + 296, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 296, + 473 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 296, + 473 + ], + "type": "text", + "content": "We evaluate all the above models and methods, extracting useful insights regarding the factors that boost retrieval performance. ILIAS is compared with other existing datasets for instance-level image retrieval. We analyze the performance of selected models1 to break down the impact of different ILIAS attributes, such as domains, clutter, and scale. Unless stated otherwise, we use the large model variants with the largest resolution available, e.g. in our analysis we use SigLIP ViT-L trained with 384 resolution." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 479, + 294, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 294, + 491 + ], + "type": "text", + "content": "5.1. Comparison with other instance-level datasets" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 496, + 295, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 496, + 295, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 496, + 295, + 687 + ], + "type": "text", + "content": "In Fig. 4, ILIAS is compared with other instance-level retrieval datasets based on evaluation of the same models. Linear adaptation is not applied as parts of GLDv2 and SOP are included in the UnED dataset. Only for the sake of this comparison, and for no other experiment in this work, we use models fine-tuned on specific domains (in-domain models), i.e. on the training sets of SOP and GLDv2. INSTRE, which is also multi-domain, shows a correlation to ILIAS, but its performance is saturated due to its small size. For single-domain datasets, in-domain models outperform others by a large margin, with few exceptions, i.e. DINOv2, which includes the trainset of GLDv2 in its training data. Several multi-domain models perform well on SOP since their training set is usually included in the training data. However, in-domain and multi-domain models face challenges on ILIAS, highlighting the diversity of our dataset." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 228, + 432, + 240 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 432, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 432, + 240 + ], + "type": "text", + "content": "5.2. Method comparison" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 245, + 555, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 245, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 245, + 555, + 533 + ], + "type": "text", + "content": "Image-to-image retrieval with global representations. Tab. 2 presents the performance of global descriptor models on ILIAS. Selected models are presented to highlight useful comparisons, while many other models are included in the supplementary material. The main factors that improve performance are the size of the training set, training resolution, and model architecture, which aligns with the literature. The impact of dataset size is apparent in various model combinations, e.g. CLIP with openai and laion2b. This is also pronounced by the dominance of foundation models. Training with large resolution brings significant gains and consistently improves mAP@1k. In some cases of SigLIP, smaller models trained with large resolutions outperform larger ones trained with small resolutions. For the models of the same resolution, it is a common trend for larger model variants to bring corresponding performance gains. In general, VLMs perform the best. From non-VLMs, only DINOv2 and Unicom achieve competitive performance. Masked Image Modeling (MIM) and supervised models are not performing well. Our linear adaptation scheme is very effective, improving most models. The boost is more pronounced in the case of VLMs. A possible explanation for such improvements is that image-to-image relations are not optimized during the training of VLMs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 533, + 556, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 556, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 556, + 640 + ], + "type": "text", + "content": "Text-to-image retrieval. Following results in Tab. 2, similar conclusions are derived for the t2i case. Retrieval performance improves with the scaling of the training data. The larger model achieves significantly better results, i.e. compare the base with large variants. Finally, it is noteworthy that the best performance achieved by SigLIP2 is very close to the i2i performance when no adaptation is used. Note that t2i includes 1k text queries in total, with one query per object, while i2i 1,232 image queries." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "Evaluation of mini-ILIAS selection. Tab. 3 shows performance on mini-ILIAS for five models with linear adaptation. The selected subset is significantly more challenging than a random selection of 5M images. More precisely, a set of approximately " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\sim 26\\mathrm{M}" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": " random images matches the performance of mini-ILIAS." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "text", + "content": "1 Although the very recent SigLIP2 is the best-performing model, we conduct most experiments with SigLIP." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14782" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 105, + 60, + 509, + 363 + ], + "blocks": [ + { + "bbox": [ + 105, + 60, + 509, + 363 + ], + "lines": [ + { + "bbox": [ + 105, + 60, + 509, + 363 + ], + "spans": [ + { + "bbox": [ + 105, + 60, + 509, + 363 + ], + "type": "table", + "html": "
modelarchtraindatasetdata sizetrain restest resimage-to-imagetext-to-image
5M†100M†100M100M5M
ResNet50 [17]R50supin1k1M2243842.51.81.7--
DINO [8]R50sslin1k1M2243844.12.92.9--
ConvNext [29]CN-Lsupin1k1M2883844.22.92.2--
OAI-CLIP [17]R50vlaopenai400M2243848.56.03.21.52.3
OpenCLIP [19, 29]CN-Bvlalaion2b2B25638418.114.07.94.67.0
OpenCLIP [19, 29]CN-Lvlalaion2b2B32051222.918.39.68.111.5
ViT [13, 54]ViT-Bsupin1k1M2243841.91.31.0--
EVA-MIM [14]ViT-Bsslin22k142M2243844.73.22.1--
ViT [13, 54]ViT-Bsupin21k14M2243846.24.43.0--
DINO [8]ViT-Bsslin1k1M2243846.64.83.7--
UDON-CLIP [73]ViT-Bsupuned2.8M2243849.26.75.9--
OAI-CLIP [41]ViT-Bvlaopenai400M22438410.77.94.21.62.7
EVA-CLIP [57]ViT-Bvlamerged2b2B22438411.78.75.92.54.4
MetaCLIP [70]ViT-Bvla2pt5b2.5B22438412.79.46.64.97.6
DINOv2 [34]ViT-Bssllvd142m142M51872415.012.111.5--
SigLIP [75]ViT-Bvlawebli10B25638420.616.711.57.510.3
SigLIP [75]ViT-Bvlawebli10B38451226.221.515.611.014.4
SigLIP [75]ViT-Bvlawebli10B51272427.523.016.611.114.6
SigLIP2 [65]ViT-Bvlawebli10B51272428.623.515.410.414.6
EVA-MIM [14]ViT-Lsslin22k14M2243843.92.71.5--
ViT [13, 54]ViT-Lsupin21k14M2243847.35.34.6--
EVA-MIM [14]ViT-Lsslmerged38m38B2243848.86.14.7--
OAI-CLIP [41]ViT-Lvlaopenai400M22438415.811.97.04.66.7
OpenCLIP [9, 19]ViT-Lvlalaion2b2B22438417.513.79.47.09.4
Unicom [1]ViT-Ldistlaion400m400M33651218.614.613.9--
OAI-CLIP [41]ViT-Lvlaopenai400M33651219.915.29.45.88.4
DINOv2 [34]ViT-Lssllvd142m142M51872418.815.315.3--
EVA-CLIP [57]ViT-Lvlamerged2b2B33651220.916.010.97.210.6
MetaCLIP [70]ViT-Lvla2pt5b2.5B22438421.716.911.79.213.1
SigLIP [75]ViT-Lvlawebli10B25638426.321.815.212.816.4
SigLIP [75]ViT-Lvlawebli10B38451234.328.919.618.122.2
SigLIP2 [65]ViT-Lvlawebli10B51272437.331.320.819.824.7
", + "image_path": "010d658bd9aca2244660ed55227da0f526577ce8ffccc0b13ed6bd466bd154a5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 75, + 412, + 272, + 480 + ], + "blocks": [ + { + "bbox": [ + 55, + 365, + 555, + 409 + ], + "lines": [ + { + "bbox": [ + 55, + 365, + 555, + 409 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 555, + 409 + ], + "type": "text", + "content": "Table 2. Performance comparison using mAP@1k on ILIAS and mini-ILIAS for global representation models for i2i and t2i. Comparison of model architecture (arch), training scheme (train), training data, and train/test resolution. " + }, + { + "bbox": [ + 55, + 365, + 555, + 409 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 55, + 365, + 555, + 409 + ], + "type": "text", + "content": " indicates results with the linear adaptation. 5M and 100M correspond to the mini and full versions of the dataset, respectively. sup, ssl, dist, vla: supervised learning, self-supervised learning, distillation and vision-language alignment. R50, CN: ResNet50 and ConvNext." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 75, + 412, + 272, + 480 + ], + "lines": [ + { + "bbox": [ + 75, + 412, + 272, + 480 + ], + "spans": [ + { + "bbox": [ + 75, + 412, + 272, + 480 + ], + "type": "table", + "html": "
model100M5M-mini5M-rand
DINOv2† [34]15.318.822.7±0.2
EVA-CLIP† [57]16.020.928.8±0.2
MetaCLIP† [70]16.921.729.2±0.1
OpenCLIP† [19, 29]18.322.930.9±0.2
SigLIP† [75]28.934.341.8±0.1
", + "image_path": "1ab422f096486bf278655cbc0eac4c71bf2b41bf54ed15abb733c155cb371b4b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 73, + 538, + 276, + 646 + ], + "blocks": [ + { + "bbox": [ + 55, + 483, + 295, + 537 + ], + "lines": [ + { + "bbox": [ + 55, + 483, + 295, + 537 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 295, + 537 + ], + "type": "text", + "content": "Table 3. A challenging distractor subset for mini-ILIAS. mAP@1k evaluated for different distractor sets, 100M: the full dataset, 5M-mini: mini-ILIAS subset, 5M-rand: random subset. We report the mean and std of 3 randomly sampled subsets. " + }, + { + "bbox": [ + 55, + 483, + 295, + 537 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 55, + 483, + 295, + 537 + ], + "type": "text", + "content": " indicates results with the linear adaptation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 73, + 538, + 276, + 646 + ], + "lines": [ + { + "bbox": [ + 73, + 538, + 276, + 646 + ], + "spans": [ + { + "bbox": [ + 73, + 538, + 276, + 646 + ], + "type": "table", + "html": "
rerankingSigLIP [75]SigLIP† [75]
mAP@1koraclemAP@1koracle
global19.648.728.956.0
αQE1 [10, 40]22.144.733.756.9
αQE2 [10, 40]20.440.831.554.4
αQE5 [10, 40]14.334.923.549.3
CS [44]22.948.732.556.0
SP [37]21.848.730.556.0
AMES [56]26.448.735.656.0
", + "image_path": "085a1528a507639b0c63d18cba9d38f092f99136207b4cbaf9b663aa32408c2c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 647, + 295, + 693 + ], + "lines": [ + { + "bbox": [ + 55, + 647, + 295, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 295, + 693 + ], + "type": "text", + "content": "Table 4. Performance comparison for re-ranking methods. Oracle represents the performance of perfect re-ranking at the top-1k images. Top: query expansion with global descriptors. Bottom: re-ranking with local descriptors. †: results with linear adaptation." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 415, + 555, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 555, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 555, + 594 + ], + "type": "text", + "content": "Retrieval with re-ranking. Tab. 4 shows the performance of re-ranking methods applied on top of SigLIP with and without linear adaptation on ILIAS. Complementary to " + }, + { + "bbox": [ + 313, + 415, + 555, + 594 + ], + "type": "inline_equation", + "content": "\\mathrm{mAP}@1\\mathrm{k}" + }, + { + "bbox": [ + 313, + 415, + 555, + 594 + ], + "type": "text", + "content": ", an oracle-based top-1k re-ranking metric is reported as the upper bound of a re-ranking method that processes the top 1k images. Local similarity estimated by a learned model proves to be very effective for re-ranking. Nevertheless, the oracle re-ranking performance indicates that there is a lot more space for improvements. Re-ranking with QE is useful when the number of aggregated neighbors is low and drops below the baseline when the number of neighbors is increased. Notably, global re-ranking affects and, interestingly, decreases oracle performance since the whole db is re-ranked; while local re-ranking does not affect it since it is performed only on a shortlist of images." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 600, + 376, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 376, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 376, + 613 + ], + "type": "text", + "content": "5.3. Analysis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "content": "Performance per domain. Fig. 5 shows the performance per taxonomy categories. The taxonomy annotations allow a fine-grained view of the results, which can possibly allow us to capture imbalanced improvements in future work. For example, DINOV2, despite being overall inferior to SigLIP, is outperforming it in categories like architecture and sculptures or is quite similar in categories like public art and paper art. This is possibly attributed to the curation and com" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14783" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 62, + 532, + 190 + ], + "blocks": [ + { + "bbox": [ + 69, + 62, + 532, + 190 + ], + "lines": [ + { + "bbox": [ + 69, + 62, + 532, + 190 + ], + "spans": [ + { + "bbox": [ + 69, + 62, + 532, + 190 + ], + "type": "image", + "image_path": "0317cd1cc7d997e82893f6d6434525ce3d95ec7803efde0fd5563d2649867c53.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 190, + 555, + 223 + ], + "lines": [ + { + "bbox": [ + 55, + 190, + 555, + 223 + ], + "spans": [ + { + "bbox": [ + 55, + 190, + 555, + 223 + ], + "type": "text", + "content": "Figure 5. Performance comparison per category. mAP@1k averaged over objects in the same mid-level taxonomy category, organized by their primary-level category size, with sorting within each group by SigLIP† + AMES performance. Comparison between SigLIP with and without adaptation, SigLIP combined with AMES reranking, SigLIP t2i, and DINOv2. † indicates results with the linear adaptation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 224, + 282, + 446 + ], + "blocks": [ + { + "bbox": [ + 58, + 224, + 282, + 446 + ], + "lines": [ + { + "bbox": [ + 58, + 224, + 282, + 446 + ], + "spans": [ + { + "bbox": [ + 58, + 224, + 282, + 446 + ], + "type": "image", + "image_path": "5397c4df56f4103ac2489ba3f99c0326cd4a35faeceb0dabb7baf6a9e023db0d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 447, + 295, + 481 + ], + "lines": [ + { + "bbox": [ + 55, + 447, + 295, + 481 + ], + "spans": [ + { + "bbox": [ + 55, + 447, + 295, + 481 + ], + "type": "text", + "content": "Figure 6. Performance comparison reporting AP per query for different approaches with SigLIP. Pearson correlation reported in parenthesis. " + }, + { + "bbox": [ + 55, + 447, + 295, + 481 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 55, + 447, + 295, + 481 + ], + "type": "text", + "content": " indicates results with the linear adaptation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 486, + 295, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 295, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 295, + 558 + ], + "type": "text", + "content": "position of the DINOV2 training set, which includes artwork and landmark datasets. Also, some categories are hurt by reranking with AMES, with some demonstrating big drops, i.e. sport, gaming, perfume. These categories deviate significantly from the domain AMES is trained, i.e. landmarks, which could potentially justify such drops." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "text", + "content": "Per query comparisons. Fig. 6 shows the AP per query for various methods. Linear adaptation boosts most queries, i.e. performance drop only for 192 queries. Image- and text-based retrieval are not strongly correlated despite performing similarly, which is good evidence [60] for the effectiveness of model ensembles. Indeed, ensembling i2i and t2i by averaging similarities brings " + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "inline_equation", + "content": "+6.1" + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "text", + "content": " improvement over i2i retrieval. Query expansion improves the queries with at least some positives at top positions, i.e. AP greater than 20. However, it harms many low-performing queries by aggregating descriptors irrelevant to the query. AMES improves the majority of the queries; however, many are harmed, indicating that there is plenty of room for improvement." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 316, + 225, + 553, + 350 + ], + "blocks": [ + { + "bbox": [ + 316, + 225, + 553, + 350 + ], + "lines": [ + { + "bbox": [ + 316, + 225, + 553, + 350 + ], + "spans": [ + { + "bbox": [ + 316, + 225, + 553, + 350 + ], + "type": "image", + "image_path": "bdf3609aaa66228e22da5ab294907ec2faef93ce98c4dc4d2e9dbf3bf7e0345f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 354, + 555, + 421 + ], + "lines": [ + { + "bbox": [ + 313, + 354, + 555, + 421 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 555, + 421 + ], + "type": "text", + "content": "Figure 7. Performance evaluation (mAP@1k) across different amounts of object area coverage and background clutter. Positives across all queries are jointly ranked based on coverage or clutter and split into 4 equal size groups. Queries with no positive in the corresponding group are discarded. No. of queries per group is in parentheses. " + }, + { + "bbox": [ + 313, + 354, + 555, + 421 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 313, + 354, + 555, + 421 + ], + "type": "text", + "content": " indicates results with the linear adaptation." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 437, + 555, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 555, + 510 + ], + "type": "text", + "content": "Impact of clutter and scale in positives. To quantify the impact of background clutter and scale changes, Fig. 7 presents the performance for different groups of positives. Dealing with small objects and multi-object scenes form major weaknesses of existing models. Notably, t2i beats i2i without adaptation in small-scale groups." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 524, + 392, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 524, + 392, + 537 + ], + "spans": [ + { + "bbox": [ + 314, + 524, + 392, + 537 + ], + "type": "text", + "content": "6. Conclusions" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "content": "We introduce ILIAS and conduct an extensive evaluation of current foundational models and retrieval methods, highlighting that instance-level retrieval remains an unsolved problem. Our results indicate that off-the-shelf application of foundational models leaves considerable room for improvement, particularly in handling small objects and complex backgrounds. While specialized retrieval methods leveraging local descriptors are effective in these cases, their high memory and computational costs become impractical at the scale of ILIAS or beyond. ILIAS is designed to become a standard benchmark for evaluating foundational representation models and retrieval methods, accommodating both global and local representations, and advancing the field of instance-level recognition." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14784" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 272 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 272 + ], + "type": "text", + "content": "Funding acknowledgement. This work was supported by the Junior Star GACR (grant no. GM 21-28830M), Horizon MSCA-PF (grant no. 101154126), Programme Johannes Amos Comenius (grant no. CZ.02.01.01/00/22.010/0003405), CTU in Prague (grant no. SGS23/173/OHK3/3T/13), the CTU institutional support (Future fund). JM was supported by Ministry of the Interior of the Czech Republic (project no. VJ02010041). We acknowledge VSB - Technical University of Ostrava, IT4Innovations National Supercomputing Center, Czech Republic, for awarding this project access to the LUMI supercomputer, owned by the EuroHPC Joint Undertaking, hosted by CSC (Finland) and the LUMI consortium through the Ministry of Education, Youth and Sports of the Czech Republic through the e-INFRA CZ (grant ID: 90254)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 277, + 296, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 296, + 445 + ], + "type": "text", + "content": "Contributors acknowledgments. We want to thank Larysa Ivashechkina for her work on the annotation of object bounding boxes and masks and the initial filtering of text queries. We also want to thank all contributors for the collection of ILIAS dataset. Additionally to all authors, the list of external contributor in alphabetical order: Aggeliki Tserota, Anna Nisyraiou, Celeste Abreu, Charalambos Tzamos, Christina Tserota, Dimitris Karageorgiou, Dmytro Mishkin, Eleni Karantali, Eleni Papadopoulou, Eva Tsiliakou, Kelly Kordopati-Zilou, Markos Zampoglou, Noa Garcia, Panagiotis Tassis, Paraskevas Kordopatis, Pavlos Alexantonakis, Ruslan Rozumnyi, Snežana Čurguz, Tereza Nejedlá, Tomáš Jelinek, Yannis Kalantidis, Vasilis Alexiadis, Yankun Wu." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 456, + 115, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 115, + 468 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 115, + 468 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 476, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 61, + 476, + 295, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 476, + 295, + 520 + ], + "spans": [ + { + "bbox": [ + 61, + 476, + 295, + 520 + ], + "type": "text", + "content": "[1] Xiang An, Jiankang Deng, Kaicheng Yang, Jaiwei Li, Ziyong Feng, Jia Guo, Jing Yang, and Tongliang Liu. Unicom: Universal and compact representation learning for image retrieval. In ICLR, 2023. 2, 5, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 521, + 295, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 521, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 61, + 521, + 295, + 544 + ], + "type": "text", + "content": "[2] Relja Arandjelović and Andrew Zisserman. Smooth object retrieval using a bag of boundaries. In ICCV, 2011. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 545, + 296, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 545, + 296, + 588 + ], + "spans": [ + { + "bbox": [ + 61, + 545, + 296, + 588 + ], + "type": "text", + "content": "[3] Harry G Barrow, Jay M Tenenbaum, Robert C Bolles, and Helen C Wolf. Parametric correspondence and chamfer matching: Two new techniques for image matching. In *IJ-CAI*, 1977. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 590, + 295, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 590, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 61, + 590, + 295, + 622 + ], + "type": "text", + "content": "[4] Wissam Bejjani, Wisdom C Agboh, Mehmet R Dogar, and Matteo Leonetti. Occlusion-aware search for object retrieval in clutter. In IROS, 2021. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 624, + 296, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 624, + 296, + 678 + ], + "spans": [ + { + "bbox": [ + 62, + 624, + 296, + 678 + ], + "type": "text", + "content": "[5] Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselat, Emma Brunskill, et al. On the opportunities and risks of foundation models. In arXiv, 2021. 1, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 680, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 680, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 680, + 296, + 712 + ], + "type": "text", + "content": "[6] Bingyi Cao, André Araujo, and Jack Sim. Unifying deep local and global features for image search. In ECCV, 2020. 5" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 320, + 73, + 555, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 555, + 116 + ], + "type": "text", + "content": "[7] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, 2020. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 320, + 119, + 554, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 119, + 554, + 162 + ], + "spans": [ + { + "bbox": [ + 320, + 119, + 554, + 162 + ], + "type": "text", + "content": "[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 5, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 320, + 164, + 554, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 164, + 554, + 218 + ], + "spans": [ + { + "bbox": [ + 320, + 164, + 554, + 218 + ], + "type": "text", + "content": "[9] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In CVPR, 2023. 5, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 315, + 221, + 553, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 221, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 315, + 221, + 553, + 263 + ], + "type": "text", + "content": "[10] Ondrej Chum, James Philbin, Josef Sivic, Michael Isard, and Andrew Zisserman. Total recall: Automatic query expansion with a generative feature model for object retrieval. In ICCV, 2007. 2, 5, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 315, + 266, + 553, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 266, + 553, + 299 + ], + "spans": [ + { + "bbox": [ + 315, + 266, + 553, + 299 + ], + "type": "text", + "content": "[11] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. In ICLR, 2024. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 301, + 553, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 301, + 553, + 334 + ], + "spans": [ + { + "bbox": [ + 315, + 301, + 553, + 334 + ], + "type": "text", + "content": "[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 2, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 336, + 553, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 336, + 553, + 402 + ], + "spans": [ + { + "bbox": [ + 315, + 336, + 553, + 402 + ], + "type": "text", + "content": "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICML, 2021. 5, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 404, + 553, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 404, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 315, + 404, + 553, + 448 + ], + "type": "text", + "content": "[14] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. EVA: Exploring the limits of masked visual representation learning at scale. In CVPR, 2023. 2, 5, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 449, + 553, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 493 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 493 + ], + "type": "text", + "content": "[15] Martin A. Fischler and Robert C. Bolles. Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 1981. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 495, + 553, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 495, + 553, + 517 + ], + "spans": [ + { + "bbox": [ + 315, + 495, + 553, + 517 + ], + "type": "text", + "content": "[16] Christopher Flagg and Ophir Frieder. Reconstruction of artifacts from digital image repositories. JOCCH, 2022. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 519, + 553, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 519, + 553, + 551 + ], + "spans": [ + { + "bbox": [ + 315, + 519, + 553, + 551 + ], + "type": "text", + "content": "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 5, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 315, + 554, + 553, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 554, + 553, + 586 + ], + "spans": [ + { + "bbox": [ + 315, + 554, + 553, + 586 + ], + "type": "text", + "content": "[18] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 589, + 553, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 589, + 553, + 643 + ], + "spans": [ + { + "bbox": [ + 316, + 589, + 553, + 643 + ], + "type": "text", + "content": "[19] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 315, + 646, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 646, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 315, + 646, + 553, + 678 + ], + "type": "text", + "content": "[20] Hervé Jégou, Matthijs Douze, and Cordelia Schmid. Hamming embedding and weak geometric consistency for large scale image search. In ECCV, 2008. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 315, + 680, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 680, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 315, + 680, + 553, + 712 + ], + "type": "text", + "content": "[21] Tomas Jenicek and Ondrej Chum. No fear of the dark: Image retrieval under varying illumination conditions. In CVPR, 2019. 1" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14785" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[22] Sungyeon Kim, Boseung Jeong, and Suha Kwak. HIER: Metric learning beyond class labels via hierarchical regularization. In CVPR, 2023. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 294, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 294, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 294, + 149 + ], + "type": "text", + "content": "[23] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In CVPR, 2023. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 294, + 183 + ], + "type": "text", + "content": "[24] Byungsoo Ko, Minchul Shin, Geonmo Gu, HeeJae Jun, Tae Kwan Lee, and Youngjoon Kim. A benchmark on tricks for large-scale image retrieval. In arXiv, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 183, + 294, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 183, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 183, + 294, + 216 + ], + "type": "text", + "content": "[25] Tarun Krishna, Kevin McGuinness, and Noel O'Connor. Evaluating contrastive models for instance-based image retrieval. In ICMR, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 216, + 294, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 216, + 294, + 249 + ], + "spans": [ + { + "bbox": [ + 56, + 216, + 294, + 249 + ], + "type": "text", + "content": "[26] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In NeurIPS, 2012. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 250, + 294, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 294, + 281 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 294, + 281 + ], + "type": "text", + "content": "[27] Seongwon Lee, Hongje Seong, Suhyeon Lee, and Euntai Kim. Correlation verification for image retrieval. In CVPR, 2022. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 283, + 294, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 283, + 294, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 283, + 294, + 316 + ], + "type": "text", + "content": "[28] Ziwei Liu, Ping Luo, Shi Qiu, Xiaogang Wang, and Xiaou Tang. Deepfashion: Powering robust clothes recognition and retrieval with rich annotations. In CVPR, 2016. 2, 3, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 316, + 294, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 294, + 349 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 294, + 349 + ], + "type": "text", + "content": "[29] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. CVPR, 2022. 5, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 350, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 294, + 392 + ], + "type": "text", + "content": "[30] Shiyang Lu, Haonan Chang, Eric Pu Jing, Abdeslam Boularias, and Kostas Bekris. OVIR-3D: Open-vocabulary 3d instance retrieval without training on 3d data. In CoRL, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 393, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 393, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 393, + 294, + 415 + ], + "type": "text", + "content": "[31] Claudio Michaelis, Matthias Bethge, and Alexander Ecker. One-shot segmentation in clutter. In ICML, 2018. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 415, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 294, + 437 + ], + "type": "text", + "content": "[32] David Nister and Henrik Stewenius. Scalable recognition with a vocabulary tree. In CVPR, 2006. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 438, + 294, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 438, + 294, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 438, + 294, + 470 + ], + "type": "text", + "content": "[33] Hyeonwoo Noh, Andre Araujo, Jack Sim, Tobias Weyand, and Bohyung Han. Large-scale image retrieval with attentive deep local features. In ICCV, 2017. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 471, + 294, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 294, + 569 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 294, + 569 + ], + "type": "text", + "content": "[34] Maxime Oquab, Timothee Darct, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINOv2: Learning robust visual features without supervision. TMLR, 2024. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 570, + 294, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 294, + 602 + ], + "type": "text", + "content": "[35] Yash Patel, Giorgos Tolias, and Jiří Matas. Recall@k surrogate loss with large batches and similarity mixup. In CVPR, 2022. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 602, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 294, + 635 + ], + "type": "text", + "content": "[36] Jingtian Peng, Chang Xiao, and Yifan Li. RP2K: A large-scale retail product dataset for fine-grained image classification. In arXiv, 2020. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 636, + 294, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 294, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 294, + 669 + ], + "type": "text", + "content": "[37] James Philbin, Ondrej Chum, Michael Isard, Josef Sivic, and Andrew Zisserman. Object retrieval with large vocabularies and fast spatial matching. In CVPR, 2007. 2, 4, 5, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[38] James Philbin, Ondrej Chum, Michael Isard, Josef Sivic, and Andrew Zisserman. Lost in quantization: Improving particular object retrieval in large scale image databases. In CVPR, 2008. 4" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[39] Filip Radenović, Ahmet Iscen, Giorgos Tolias, Yannis Avrithis, and Ondrej Chum. Revisiting oxford and paris: Large-scale image retrieval benchmarking. In CVPR, 2018. 1, 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 118, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 150 + ], + "type": "text", + "content": "[40] Filip Radenovic, Giorgos Tolias, and Ondrej Chum. Finetuning cnn image retrieval with no human annotation. PAMI, 2019. 2, 4, 5, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 152, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 217 + ], + "type": "text", + "content": "[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 3, 5, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 218, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 218, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 218, + 553, + 262 + ], + "type": "text", + "content": "[42] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, et al. SAM 2: Segment anything in images and videos. In ICLR, 2025. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 264, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 297 + ], + "type": "text", + "content": "[43] Ali Sharif Razavian, Hossein Azizpour, Josephine Sullivan, and Stefan Carlsson. CNN features off-the-shelf: An astounding baseline for recognition. In CVPRW, 2014. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 297, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 341 + ], + "type": "text", + "content": "[44] Ali S Razavian, Josephine Sullivan, Stefan Carlsson, and Atsuto Maki. Visual instance retrieval with deep convolutional networks. ITE Trans. on Media Technology and Applications, 2016. 5, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 342, + 553, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 553, + 397 + ], + "type": "text", + "content": "[45] Taichi Sakaguchi, Akira Taniguchi, Yoshinobu Hagiwara, Lotfi El Hafi, Shoichi Hasegawa, and Tadahiro Taniguchi. Object instance retrieval in assistive robotics: Leveraging fine-tuned simsiam with multi-view images based on 3d semantic map. In IROS, 2024. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 399, + 553, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 441 + ], + "type": "text", + "content": "[46] Mert Bulent Sariyildiz, Philippe Weinzaepfel, Thomas Lucas, Diane Larlus, and Yannis Kalantidis. UNIC: Universal classification models via multi-teacher distillation. In ECCV, 2024. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 443, + 553, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 443, + 553, + 476 + ], + "spans": [ + { + "bbox": [ + 316, + 443, + 553, + 476 + ], + "type": "text", + "content": "[47] Konstantin Schall, Kai Uwe Barthel, Nico Hezel, and Klaus Jung. GPR1200: a benchmark for general-purpose content-based image retrieval. In MMM, 2022. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 477, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 477, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 477, + 553, + 510 + ], + "type": "text", + "content": "[48] Shihao Shao, Kaifeng Chen, Arjun Karpur, Qinghua Cui, André Araujo, and Bingyi Cao. Global features are all you need for image retrieval and reranking. In ICCV, 2023. 2, 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 511, + 553, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 511, + 553, + 543 + ], + "spans": [ + { + "bbox": [ + 316, + 511, + 553, + 543 + ], + "type": "text", + "content": "[49] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In ICLR, 2015. 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 545, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 553, + 600 + ], + "type": "text", + "content": "[50] Ivan Sipiran, Patrick Lazo, Cristian Lopez, Milagritos Jimenez, Nihar Bagewadi, Benjamin Bustos, Hieu Dao, Shankar Gangisetty, Martin Hanik, Ngoc-Phuong Ho-Thi, et al. SHREC 2021: Retrieval of cultural heritage objects. Computers & Graphics, 2021. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 601, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 634 + ], + "type": "text", + "content": "[51] Arnold WM Smeulders, Marcel Worring, Simone Santini, Amarnath Gupta, and Ramesh Jain. Content-based image retrieval at the end of the early years. PAMI, 2000. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 635, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 678 + ], + "type": "text", + "content": "[52] Chull Hwan Song, Jooyoung Yoon, Taebaek Hwang, Shunghyun Choi, Yeong Hyeon Gu, and Yannis Avrithis. On train-test class overlap and detection for image retrieval. In CVPR, 2024. 1" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "text", + "content": "[53] Hyun Oh Song, Yu Xiang, Stefanie Jegelka, and Silvio Savarese. Deep metric learning via lifted structured feature embedding. In CVPR, 2016. 1, 2, 3, 4" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14786" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[54] Andreas Steiner, Alexander Kolesnikov, Xiaohua Zhai, Ross Wightman, Jakob Uszkoreit, and Lucas Beyer. How to train your vit? data, augmentation, and regularization in vision transformers. TMLR, 2021. 5, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 151 + ], + "type": "text", + "content": "[55] Pavel Suma and Giorgos Tolias. Large-to-small image resolution asymmetry in deep metric learning. In WACV, 2023. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 153, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 153, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 153, + 294, + 196 + ], + "type": "text", + "content": "[56] Pavel Suma, Giorgos Kordopatis-Zilos, Ahmet Iscen, and Giorgos Tolias. AMES: Asymmetric and memory-efficient similarity estimation for instance-level retrieval. In ECCV, 2024. 2, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 199, + 294, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 199, + 294, + 231 + ], + "spans": [ + { + "bbox": [ + 56, + 199, + 294, + 231 + ], + "type": "text", + "content": "[57] Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. EVA-CLIP: Improved training techniques for clip at scale. In arXiv, 2023. 2, 5, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 234, + 294, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 234, + 294, + 277 + ], + "spans": [ + { + "bbox": [ + 56, + 234, + 294, + 277 + ], + "type": "text", + "content": "[58] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In CVPR, 2015. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 279, + 294, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 279, + 294, + 301 + ], + "spans": [ + { + "bbox": [ + 56, + 279, + 294, + 301 + ], + "type": "text", + "content": "[59] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In ICML, 2019. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 303, + 294, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 303, + 294, + 335 + ], + "spans": [ + { + "bbox": [ + 56, + 303, + 294, + 335 + ], + "type": "text", + "content": "[60] Ryan Theisen, Hyunsuk Kim, Yaoqing Yang, Liam Hodgkinson, and Michael W Mahoney. When are ensembles really effective? In NeurIPS, 2024. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 338, + 294, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 380 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 380 + ], + "type": "text", + "content": "[61] Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. YFCC100M: The new data in multimedia research. Communications of the ACM, 2016. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 383, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 294, + 415 + ], + "type": "text", + "content": "[62] Akihiko Torii, Relja Arandjelovic, Josef Sivic, Masatoshi Okutomi, and Tomas Pajdla. 24/7 place recognition by view synthesis. In CVPR, 2015. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 418, + 294, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 418, + 294, + 450 + ], + "spans": [ + { + "bbox": [ + 56, + 418, + 294, + 450 + ], + "type": "text", + "content": "[63] Hugo Touvron, Andrea Vedaldi, Matthijs Douze, and Herve Jegou. Fixing the train-test resolution discrepancy. In NeurIPS, 2019. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 453, + 294, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 453, + 294, + 495 + ], + "spans": [ + { + "bbox": [ + 56, + 453, + 294, + 495 + ], + "type": "text", + "content": "[64] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, 2021. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 498, + 294, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 498, + 294, + 563 + ], + "spans": [ + { + "bbox": [ + 56, + 498, + 294, + 563 + ], + "type": "text", + "content": "[65] Michael Tschannen, Alexey Gritsenko, Xiao Wang, Muhammad Ferjad Naeem, Ibrahim Alabdulmohsin, Nikhil Parthasarathy, Talfan Evans, Lucas Beyer, Ye Xia, Basil Mustafa, et al. SigLIP 2: Multilingual vision-language encoders with improved semantic understanding, localization, and dense features. In arXiv, 2025. 5, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 566, + 294, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 566, + 294, + 597 + ], + "spans": [ + { + "bbox": [ + 56, + 566, + 294, + 597 + ], + "type": "text", + "content": "[66] Shuang Wang and Shuqiang Jiang. INSTRE: A new benchmark for instance-level object retrieval and recognition. TOMM, 2015. 1, 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 600, + 294, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 600, + 294, + 643 + ], + "spans": [ + { + "bbox": [ + 56, + 600, + 294, + 643 + ], + "type": "text", + "content": "[67] Tobias Weyand, André Araujo, Bingyi Cao, and Jack Sim. Google Landmarks Dataset v2 - A large-scale benchmark for instance-level recognition and retrieval. In CVPR, 2020. 1, 2, 3, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 646, + 294, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 646, + 294, + 677 + ], + "spans": [ + { + "bbox": [ + 56, + 646, + 294, + 677 + ], + "type": "text", + "content": "[68] Ross Wightman, Hugo Touvron, and Hervé Jégou. Resnet strikes back: An improved training procedure in timm. In NeurIPSW, 2021. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "text", + "content": "[69] Pengxiang Wu, Siman Wang, Kevin Dela Rosa, and Derek Hu. FORB: A flat object retrieval benchmark for universal image embedding. In NeurIPS, 2024. 3" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 453 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[70] Hu Xu, Saining Xie, Xiaqing Ellen Tan, Po-Yao Huang, Russell Howes, Vasu Sharma, Shang-Wen Li, Gargi Ghosh, Luke Zettlemoyer, and Christoph Feichtenhofer. Demystifying clip data. In ICLR, 2024. 3, 5, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "type": "text", + "content": "[71] Nikolaos-Antonios Ypsilantis, Noa Garcia, Guangxing Han, Sarah Ibrahimi, Nanne Van Noord, and Giorgos Tolias. The MET dataset: Instance-level recognition for artworks. In NeurIPS, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 163, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 228 + ], + "type": "text", + "content": "[72] Nikolaos-Antonios Ypsilantis, Kaifeng Chen, Bingyi Cao, Mario Lipovsky, Pelin Dogan-Schönberger, Grzegorz Makosa, Boris Bluntschli, Mojtaba Seyedhosseini, Ondrej Chum, and André Araujo. Towards universal image embeddings: A large-scale dataset and challenge for generic image representations. In ICCV, 2023. 1, 2, 3, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 230, + 553, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 553, + 272 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 553, + 272 + ], + "type": "text", + "content": "[73] Nikolaos-Antonios Ypsilantis, Kaifeng Chen, Andre Araujo, and Ondrej Chum. UDON: Universal dynamic online distillation for generic image representations. In NeurIPS, 2024. 5, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 274, + 553, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 274, + 553, + 317 + ], + "spans": [ + { + "bbox": [ + 316, + 274, + 553, + 317 + ], + "type": "text", + "content": "[74] Jiangbo Yuan, An-Ti Chiang, Wen Tang, and Antonio Haro. eProduct: A million-scale visual search benchmark to address product recognition challenges. In CVPRW, 2021. 1, 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 319, + 553, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 319, + 553, + 351 + ], + "spans": [ + { + "bbox": [ + 316, + 319, + 553, + 351 + ], + "type": "text", + "content": "[75] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In ICCV, 2023. 1, 2, 3, 4, 5, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 354, + 553, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 397 + ], + "type": "text", + "content": "[76] Xunlin Zhan, Yangxin Wu, Xiao Dong, Yunchao Wei, Minlong Lu, Yichi Zhang, Hang Xu, and Xiaodan Liang. Product: Towards weakly supervised instance-level product retrieval via cross-modal pretraining. In ICCV, 2021. 1, 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 399, + 553, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 419 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 419 + ], + "type": "text", + "content": "[77] Liang Zheng, Yi Yang, and Qi Tian. Sift meets cnn: A decade survey of instance retrieval. PAMI, 2017. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 421, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 421, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 421, + 553, + 453 + ], + "type": "text", + "content": "[78] Barret Zoph, Vijay Vasudevan, Jonathon Shlens, and Quoc V Le. Learning transferable architectures for scalable image recognition. In CVPR, 2018. 5" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14787" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/IM-Portrait_ Learning 3D-aware Video Diffusion for Photorealistic Talking Heads from Monocular VideosC/d89a7508-cd12-421c-9424-93ad28adceb7_content_list.json b/2025/IM-Portrait_ Learning 3D-aware Video Diffusion for Photorealistic Talking Heads from Monocular VideosC/d89a7508-cd12-421c-9424-93ad28adceb7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..35cc5e08491356be1afdc7cf25dcef810c3983b7 --- /dev/null +++ b/2025/IM-Portrait_ Learning 3D-aware Video Diffusion for Photorealistic Talking Heads from Monocular VideosC/d89a7508-cd12-421c-9424-93ad28adceb7_content_list.json @@ -0,0 +1,2121 @@ +[ + { + "type": "text", + "text": "IM-Portrait: Learning 3D-aware Video Diffusion for Photorealistic Talking Heads from Monocular Videos", + "text_level": 1, + "bbox": [ + 161, + 130, + 836, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuan Li $^{1,2*§}$ Ziqian Bai $^{2*}$ Feitong Tan $^{2}$ Zhaopeng Cui $^{1\\dagger}$ Sean Fanello $^{2}$ Yinda Zhang $^{2\\dagger}$ $^{1}$ State Key Lab of CAD&CG, Zhejiang University $^{2}$ Google", + "bbox": [ + 189, + 200, + 802, + 258 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ec9265c354faa8646c8f3e3609394ad1ca6cea16252a845c6c2c582dff18c73d.jpg", + "image_caption": [ + "Inputs" + ], + "image_footnote": [], + "bbox": [ + 98, + 286, + 230, + 426 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3fd86c5b7b4dce66736362fbb47662b14ff889b097dcad8e5c5c104b04acc529.jpg", + "image_caption": [ + "Generated Results" + ], + "image_footnote": [], + "bbox": [ + 236, + 286, + 362, + 425 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/73b69f0a3b8f6b1e1ad6d356ed210561ba0f665808b37483e3b9a9f668211473.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 286, + 493, + 425 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/de5f90519439ddd28220066c889d6301b4846905698e0679a5142c716876dead.jpg", + "image_caption": [ + "Figure 1. We propose a 3D-aware video diffusion model for talking head synthesis. Given an image as identity and a sequence of tracking signals (as shown on the left for each example), our model directly generates videos in Multiplane Images (MPIs) in a single denoising process, which is ready for efficient novel-view rendering. This enables immersive viewing experience, e.g. rendering binocular stereo or perspective distortion in VR headset." + ], + "image_footnote": [], + "bbox": [ + 98, + 433, + 228, + 574 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/39e679271344c55ad39128792cff5a576f415bda6d20f1f565510c27728a9724.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 236, + 433, + 362, + 571 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0f2026fbf02b6b9aa72d77aca96cf9106890065906adf574188bbaf19fc45f1e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 433, + 495, + 571 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/34992577c8be39c707c626f21a9f0b9075dd446c73a5213cc2a4c8c5ba9f0fe4.jpg", + "image_caption": [ + "Inputs" + ], + "image_footnote": [], + "bbox": [ + 509, + 285, + 637, + 424 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1ae3402671efe9adefba86ad704f97f2dd666fc44d01090e3cfde99218ecca1b.jpg", + "image_caption": [ + "Generated Results" + ], + "image_footnote": [], + "bbox": [ + 645, + 285, + 771, + 424 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7bf921429c92fbd024546c9c194049301803dacf37b21fdcf8d517b8a99778f9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 777, + 286, + 903, + 424 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9990816515274e38c39a4ae9c4269d30110b18c5ac120a7ad0e1446885558faf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 433, + 637, + 571 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6fe0a1cc798d76d9ccc3c5f90355f6d00aeed7dda8d6af7a0075322e5a5f1bff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 433, + 771, + 571 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/22e0927c20c3ff7ceec2a2c2b0b571116889d2f7b4c05f9f899c44590fb2f781.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 777, + 433, + 903, + 571 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 647, + 326, + 662 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We propose a novel 3D-aware diffusion-based method for generating photorealistic talking head videos directly from a single identity image and explicit control signals (e.g., expressions). Our method generates Multiplane Images (MPIs) that ensure geometric consistency, making them ideal for immersive viewing experiences like binocular videos for VR headsets. Unlike existing methods that often require a separate stage or joint optimization to reconstruct a 3D representation (such as NeRF or 3D Gaussians), our approach directly generates the final output through a single denoising process, eliminating the need for post", + "bbox": [ + 89, + 681, + 483, + 848 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "processing steps to render novel views efficiently. To effectively learn from monocular videos, we introduce a training mechanism that reconstructs the output MPI randomly in either the target or the reference camera space. This approach enables the model to simultaneously learn sharp image details and underlying 3D information. Extensive experiments demonstrate the effectiveness of our method, which achieves competitive avatar quality and novel-view rendering capabilities, even without explicit 3D reconstruction or high-quality multi-view training data.", + "bbox": [ + 511, + 648, + 906, + 801 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 829, + 645, + 845 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Synthesizing realistic talking heads, has been a long-lasting challenging problem in computer vision for decades. The recent emergence of diffusion models within the generative", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Authors contributed equally.", + "bbox": [ + 107, + 862, + 267, + 875 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding authors.", + "bbox": [ + 109, + 875, + 241, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{\\S}$ Work was conducted while Yuan Li was an intern at Google.", + "bbox": [ + 109, + 887, + 434, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21107", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "AI domain offers a promising solution to this problem. Significant advancements have been made in modeling talking heads using image or video diffusion models conditioned on tracking signals [13, 28, 29, 42], leveraging the remarkable generative capabilities of diffusion models.", + "bbox": [ + 89, + 90, + 480, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Most diffusion-based approaches, however, operate in 2D image space, and the extension of these methods to 3D rendering remains largely unexplored. One potential approach involves applying 3D transformations to the tracking signal, such as parametric models or landmarks. However, this necessitates running the reverse diffusion process for each new camera view, leading to significant computational overhead. Alternatively, some diffusion-based approaches generate multi-view images, which often lack geometric consistency. While off-the-shelf methods like NeRF [12] or 3D Gaussians can be used to construct a 3D representation for efficient rendering, the inherent inconsistencies in the input multi-view images limit the quality of the final 3D reconstruction, leading to potential issues in multi-view and temporal consistency.", + "bbox": [ + 89, + 169, + 482, + 396 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we present, for the first time, a diffusion based approach that produces 3D talking heads that can be rendered efficiently in varying camera poses. We adopt Multiplane Image (MPI) as the 3D representation due to its remarkable rendering efficiency and quality, and the affinity to 2D images such that we can make good use of existing 2D diffusion models. Essentially, we formulate the problem of synthesizing talking heads images as a denoising diffusion process, which is conditioned on a reference image (i.e. the subject identity) and a sequence of control signals from parametric models (i.e. expressions and head poses). In a single denoising process, our model generates MPI videos that can be readily rendered in varying camera pose and support spatial viewing experience.", + "bbox": [ + 89, + 398, + 482, + 611 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our model is learned directly from a large scale monocular talking head videos in-the-wild. Due to the missing of multi-view datasets, like previous works, the 3D consistency needs to be learned from the head pose variability between video frames. To achieve this, it is required to construct MPs in a camera that looks at the head from a different viewpoint than the target images. Specifically, we use the viewpoint of the reference image to construct MPs. The estimated target images can then be rendered from MPs by moving the camera according to the relative head pose differences between the reference and target images. In this way, however, the frontal views of MPs will be misaligned with the input noisy images (in our case the target ground-truth images), which will break the training of the diffusion model. To tackle the problem, we propose a novel training framework to bootstrap our model to synthesize pseudo noisy input images that align with the MPs in reference cameras. With this, the MPs are randomly generated from either the reference camera viewpoint or the", + "bbox": [ + 89, + 613, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "target camera viewpoint during the training, which allows the diffusion model to learn both 3D information and fine details.", + "bbox": [ + 511, + 90, + 903, + 135 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are as follows.", + "bbox": [ + 532, + 138, + 750, + 151 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a diffusion-based approach for 3D-aware talking heads generation.", + "- Our diffusion model directly generates 3D renderable talking heads in a single denoising process, eliminating the need for post-processing steps.", + "- We propose a novel training framework that enables our model to learn from monocular videos, and show how randomly selecting the camera viewpoint between reference and target images allows for better convergence and sharper results.", + "- Extensive experiments demonstrate that our method achieves comparable or superior image and video quality to existing diffusion-based 2D talking head methods, while also offering efficient and consistent multiview rendering on par with specialized (i.e. non-diffusion based) 3D approaches." + ], + "bbox": [ + 513, + 154, + 903, + 396 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 412, + 663, + 428 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2D Talking Heads. In recent year, there has been a great advancement on synthesizing 2D talking heads with neural networks. A recent line of works [7, 14-16, 34, 40, 43, 44] propose to leverage neural networks to encode the appearance and motion from input images [15, 16, 34, 40] (or explicit control signals [14, 43, 44]) to some latent features or manually defined representations (e.g. key-points, warping fields, texture maps), which are injected into 2D CNN generators with various architectures for talking head synthesis. More recently, many works [13, 28, 29, 42] started to adopt 2D diffusion models, leveraging their impressive image quality, for controllable talking head synthesis. However, these approaches usually use simplified camera models (e.g. orthogonal projection) or they do not have any explicit camera control, making them unsuitable for spatial viewing experiences such as binocular stereo and perspective geometry. In addition, these methods usually do not produce explicit 3D geometry, which also limits their downstream applications. While most of prior work learns talking head models from monocular videos which are available in a large scale in-the-wild, multi-view datasets, e.g. from lightstage [21], could be leveraged to further improve fidelity, however such data requires intensive collection labor and there are no large scale multi-view datasets that are publicly available.", + "bbox": [ + 511, + 439, + 906, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Talking Heads. The problem of generating 3D talking heads has been a hot topic in the community. People attempted to model their talking heads with various 3D representations, such as textured mesh [9, 20], Neural Radiance Fields (NeRFs) [30] or other implicit func", + "bbox": [ + 511, + 824, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21108", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tions [2-4, 11, 12, 18, 19, 48], and, more recently, Gaussian Splatting [10, 32, 33]. While most of these approaches are person-specific, which requires to retrain a new model for every new person, some methods [11, 12] also explore the potential of person-agnostic models with explicit 3D representations. Despite having good 3D properties, these approaches cannot easily leverage the recent advancement on 2D diffusion, which lower the upper bound of their image quality.", + "bbox": [ + 89, + 90, + 480, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion-based 3D Generative Model. Another line of work, investigates how to apply diffusion models on 3D representations. The most straightforward way is applying the standard diffusion process directly on 3D data [38, 41]. However, this requires large scale 3D datasets, which are generally not available. One typical workaround is to train the diffusion model with 2D images. Some works [17, 24, 25, 31, 39] firstly learn a latent space of the 3D samples with multiview 2D images, then apply diffusion on the learned latent space. However, this results in an additional latent space learning, which is non-trivial thus increasing the pipeline complexity. Some other works [1, 35] directly model 3D scene distributions with image space diffusion by integrating the 3D-to-2D rendering into the reverse process. However, these methods are mostly proposed for static scenes where multiview datasets are available. In contrast to prior research, our method handles dynamic 3D talking heads and is trained exclusively on monocular RGB videos from real-world sources.", + "bbox": [ + 91, + 234, + 483, + 521 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 537, + 181, + 554 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our goal is to generate 3D talking head videos represented as Multiplane Images (MPIs), conditioned on a reference image (the subject's identity) and a sequence of control signals from a parametric model to determine expressions and head poses. The generated videos must exhibit high image quality, controllability, and efficient rendering for VR scenarios, including binocular consistency.", + "bbox": [ + 89, + 565, + 482, + 671 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given the scarcity of large-scale multi-view video datasets, we design our training pipeline to operate solely on 2D monocular videos from the wild. To leverage the power of 2D diffusion models, we formulate our model as a video diffusion process operating in the image space. However, this raises two key challenges: 1) How to generate 3D scenes from a diffusion denoising process defined on 2D observations, without access to ground-truth MPs? 2) How to learn 3D information from monocular 2D videos?", + "bbox": [ + 89, + 671, + 482, + 806 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address the first challenge, we design a video diffusion pipeline integrated with a differentiable forward rendering model [35] (Sec. 3.2) to directly model 3D scene distributions from 2D monocular data.", + "bbox": [ + 89, + 809, + 482, + 868 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address the second challenge, a possible approach involves constructing MPs of a 3D talking head in the refer", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ence image's camera, then moving the target camera to render the head in novel views during training. This allows us to aggregate multi-view observations of the head from different frames into a consistent 3D space to learn 3D shapes. To this end, we propose an alternating training strategy that switches between using the reference and target views as the frontal camera for MPI construction and compute the loss on target views accordingly (Sec. 3.3). This approach encourages the network to learn to synthesize high-quality, sharp images when building the MPIs in target views and to learn multiview consistency when building it in the reference view. When MPIs are constructed in the reference camera, the target ground-truth images cannot be used as input of the diffusion model, as they are not pixel-aligned with the MPIs. To tackle this, we bootstrap our model to generate the pseudo ground truth under the reference camera (Sec. 3.3). During inference, our model generate 3D scene representations in a single denoising diffusion process without any post-processing or optimization (Sec. 3.4).", + "bbox": [ + 511, + 90, + 903, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 511, + 383, + 653, + 398 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion Formulation. Denoising diffusion probabilistic models (DDPM) [23] are one kind of generative models that learn to capture a data distribution $p_{\\theta}(\\mathbf{x})$ . The method first constructs a forward Markovian process $q(\\mathbf{x}_{0:T})$ to iteratively add noise to data as:", + "bbox": [ + 511, + 407, + 906, + 482 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq (\\mathbf {x} _ {t} | \\mathbf {x} _ {t - 1}) = \\mathcal {N} (\\mathbf {x} _ {t}; \\sqrt {1 - \\beta_ {t}} \\mathbf {x} _ {t - 1}, \\beta_ {t} \\mathbf {I}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 488, + 906, + 507 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\beta_{t}$ is scheduled to control the amount of noise being added to data in each step. A denoising diffusion model $g_{\\theta}$ with parameters $\\theta$ is trained to iteratively perform the reverse process $p_{\\theta}(\\mathbf{x}_{t - 1}|\\mathbf{x}_t)$ to generate a data sample $\\mathbf{x}_0$ from a random Gaussian noise $\\mathbf{x}_T$ . If we parameterize the model $g_{\\theta}$ to predict the estimated clean data sample $\\hat{\\mathbf{x}}_{t - 1} = g_{\\theta}(\\mathbf{x}_t,t)\\approx \\mathbf{x}_0$ , then the reverse process can be written as:", + "bbox": [ + 511, + 511, + 905, + 619 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} p _ {\\theta} \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}\\right) = p _ {\\theta} \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}, \\hat {\\mathbf {x}} _ {t - 1}\\right) \\\\ = \\mathcal {N} \\left(\\mathbf {x} _ {t - 1}; \\boldsymbol {\\mu} _ {t} \\left(\\mathbf {x} _ {t}, \\hat {\\mathbf {x}} _ {t - 1}\\right), \\sigma_ {t} ^ {2} \\mathbf {I}\\right) \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 623, + 903, + 660 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Please find the detailed formulation of $\\pmb{\\mu}_t$ and $\\sigma_t^2$ in Eq. (7) of DDPM [23].", + "bbox": [ + 511, + 666, + 903, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multiplane Images. We leverage Multiplane Images (MPIs) to represent a 3D scene S (e.g. a 3D talking head). An MPI representation consists of $D$ RGBA image planes placed in a reference 3D space defined by an MPI frontal camera $\\phi^{\\mathrm{mpi}}$ . These planes are parallel to the image plane of the frontal camera and equally spaced in disparity between near $d_{n}$ and far $d_{f}$ depths from the camera. During rendering, the planes are firstly warped from the frontal camera $\\phi^{\\mathrm{mpi}}$ to a target camera $\\phi^{\\mathrm{trgt}}$ by a homography. The warped planes are then composited via alpha blending. We also follow Tucker et al. [36] to use only two RGB images, one as frontal, the other as residual, during MPI rendering to set a proper rendering capability of our model.", + "bbox": [ + 511, + 704, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21109", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e7f9121d82338d66dd3c1d1a3cea917b4c07119bc2826f61f8eb5c20888dc5d7.jpg", + "image_caption": [ + "Figure 2. Inference pipeline. Our model is built on the architecture of Lumiere [5], which takes an identity image, 2D noise video, the sequence of expressions rendered from 3DMM model and first frame image as input and outputs MPI video sequences. During inference, the network is conditioned on a reference portrait and takes the last frame of previously generated clip as the first frame condition. We separate the network into a color branch and a geometry branch. The two branches share information by zero convolution [45]." + ], + "image_footnote": [], + "bbox": [ + 112, + 88, + 883, + 344 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. 3D Talking Head Video Diffusion", + "text_level": 1, + "bbox": [ + 89, + 412, + 382, + 429 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Unlike previous diffusion-based 2D talking head models [29, 42], our model aims to generate a sequence of 3D head representations in the form of an MPI video with $N$ frames $\\{\\mathbf{S}\\}_{N}$ . This is achieved by conditioning the model on a reference identity image $\\mathbf{O}^{\\mathrm{ref}}$ and a sequence of control signals $\\{\\mathbf{C}\\}_{N}$ , which are the parameters of a mesh-based parametric head model [6, 26], controlling facial expressions and head pose. Inspired by recent 3D diffusion models [1, 35], we propose a video diffusion framework that generates 3D content (an MPI video) directly from 2D observations by integrating the MPI rendering into the denoising process. More specifically, we formulate the forward process in our avatar diffusion as", + "bbox": [ + 89, + 434, + 483, + 630 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(\\mathbf {O} _ {t} ^ {\\mathrm {m p i}} \\mid \\mathbf {O} _ {t - 1} ^ {\\mathrm {m p i}}\\right) = \\mathcal {N} \\left(\\mathbf {O} _ {t} ^ {\\mathrm {m p i}}; \\sqrt {1 - \\beta_ {t}} \\mathbf {O} _ {t - 1} ^ {\\mathrm {m p i}}, \\beta_ {t} \\mathbf {I}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 637, + 482, + 656 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{O}^{\\mathrm{mpi}}$ is the 2D video observed by the MPI frontal camera $\\phi^{\\mathrm{mpi}}$ defined in Sec. 3.1. Note that we intentionally formulate the diffusion process on the image space of MPI frontal camera since it benefits the output quality. Also, we omit the $\\{.\\}_{N}$ in $\\{\\mathbf{O}^{\\mathrm{mpi}}\\}_{N}$ for simplicity and do the same for $\\{\\phi^{\\mathrm{mpi}}\\}_{N}$ , $\\{\\mathbf{S}\\}_{N}$ and $\\{\\mathbf{C}\\}_{N}$ . For the reverse process, instead of directly predicting the estimated clean 2D video $\\hat{\\mathbf{O}}_{t-1}^{\\mathrm{mpi}}$ , we let the denoising model $g_{\\theta}$ to predict the MPI video $\\mathbf{S}_{t-1}$ , followed by the MPI rendering to obtain $\\hat{\\mathbf{O}}_{t-1}^{\\mathrm{mpi}}$ . Formally, we have", + "bbox": [ + 89, + 662, + 483, + 816 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {S} _ {t - 1} = g _ {\\theta} \\left(\\mathbf {O} ^ {\\mathrm {r e f}}, \\mathbf {O} _ {t} ^ {\\mathrm {m p i}}; t, \\mathbf {C} ^ {\\mathrm {m p i}}\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 183, + 823, + 482, + 840 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {O}} _ {t - 1} ^ {\\mathrm {m p i}} = \\mathcal {R} \\left(\\mathbf {S} _ {t - 1}, \\phi^ {\\mathrm {m p i}}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 844, + 482, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{R}$ is the MPI rendering function. Here, we rasterize the control signals $\\mathbf{C}$ into UV coordinate maps under the", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "MPI frontal camera $\\phi^{\\mathrm{mpi}}$ to provide a pixel-aligned condition $\\mathbf{C}^{\\mathrm{mpi}}$ . We also convert the meshes of the parametric model under $\\mathbf{C}$ into alpha planes and feed them into the model. In practice, we also concatenate the identity image $\\mathbf{O}^{\\mathrm{ref}}$ with its corresponding UV coordinate map rasterized from the reference camera as the model input. Finally, we plug in the estimated clean 2D video $\\hat{\\mathbf{O}}_{t - 1}^{\\mathrm{mpi}}$ to Eq. (2) to perform the reverse process, which is repeated $T$ times to sample the final MPI video output $\\mathbf{S}_0$ . Given a novel side view camera (or trajectory) $\\phi^{\\mathrm{side}}$ , we can simply run the MPI rendering to obtain the desired video $\\mathbf{O}_0^{\\mathrm{side}} = \\mathcal{R}(\\mathbf{S}_0,\\phi^{\\mathrm{side}})$ .", + "bbox": [ + 511, + 412, + 906, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Reference-Target Alternating Training", + "text_level": 1, + "bbox": [ + 511, + 587, + 848, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following a similar analysis as Tewari et al. [35], a straightforward way to train our model in Sec. 3.2 needs two losses on videos rendered from two different cameras $\\phi^{\\mathrm{MPI}}$ (the MPI frontal camera) and $\\phi^{\\mathrm{side}}$ (a novel side view camera):", + "bbox": [ + 511, + 609, + 905, + 671 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {O}} _ {t - 1} ^ {\\text {s i d e}} = \\mathcal {R} \\left(\\mathbf {S} _ {t - 1}, \\phi^ {\\text {s i d e}}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 676, + 687, + 695 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\theta} ^ {\\mathrm {m p i}} = \\mathbb {E} _ {\\mathbf {O} ^ {\\mathrm {r e f}}, \\mathbf {O} ^ {\\mathrm {m p i}}, \\mathbf {C} ^ {\\mathrm {m p i}}, t} [ \\| \\mathbf {O} ^ {\\mathrm {m p i}} - \\hat {\\mathbf {O}} _ {t - 1} ^ {\\mathrm {m p i}} \\| ^ {2} ] \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 698, + 903, + 718 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\theta} ^ {\\text {s i d e}} = \\mathbb {E} _ {\\mathbf {O} ^ {\\text {r e f}}, \\mathbf {O} ^ {\\text {m p i}}, \\mathbf {C} ^ {\\text {m p i}}, t, \\mathbf {O} ^ {\\text {s i d e}}, \\phi^ {\\text {s i d e}}} [ \\| \\mathbf {O} ^ {\\text {s i d e}} - \\hat {\\mathbf {O}} _ {t - 1} ^ {\\text {s i d e}} \\| ^ {2} ]. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 720, + 903, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Intuitively, the first loss $\\mathcal{L}_{\\theta}^{\\mathrm{mpi}}$ encourages the model to give good frontal views of the MPs, while the second loss $\\mathcal{L}_{\\theta}^{\\mathrm{side}}$ encourages the model to render good side views from the MPs, which would be the only supervision for learning 3D shapes. However, such approach would require a large scale multiview video dataset, which is not generally available.", + "bbox": [ + 511, + 750, + 905, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In order to enable multi-view supervision when training purely on monocular 2D videos, a possible approach is to place MPs in a canonical space with a fixed head pose, then move the rendering camera to simulate head pose changes.", + "bbox": [ + 511, + 840, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21110", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/84147aac9bd68af2b8e6b72b8d893ccd8c41c306d238900bbd66b7194036ac12.jpg", + "image_caption": [ + "Figure 3. Illustration of Reference-Target Alternating Training. On the top branch, we construct MPs in the target camera and render the frontal view, where we directly use the ground truth target image to compute loss Eq. (6) to learn sharp renderings and bootstrap the model. On the bottom branch, we construct MPs in the reference camera. We first rasterize the target control signal (expression) in the reference camera, then use the bootstrapped model to generate pseudo ground truth by rendering the frontal view. The pseudo ground truth is added with noise, serving as the model input to compute loss Eq. (7) for learning 3D shapes." + ], + "image_footnote": [], + "bbox": [ + 117, + 90, + 880, + 306 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this way, the image data representing different head poses will act as a proxy ground truth for novel camera views to provide multi-view supervision. Therefore, we could separately compute loss $\\mathcal{L}_{\\theta}^{\\mathrm{mpi}}$ Eq. (6) and loss $\\mathcal{L}_{\\theta}^{\\mathrm{side}}$ Eq. (7) with only monocular 2D videos by using two different MPI setups as shown in Fig. 3. More specifically, when computing the first loss $\\mathcal{L}_{\\theta}^{\\mathrm{mpi}}$ , we would set the MPI frontal cameras $\\phi^{\\mathrm{mpi}}$ as the target image camera and use the target image as the frontal view ground truth $\\mathbf{O}^{\\mathrm{mpi}}$ . Similarly, when computing the second loss $\\mathcal{L}_{\\theta}^{\\mathrm{side}}$ , we would set the MPI frontal cameras $\\phi^{\\mathrm{mpi}}$ to the reference image camera and use the target image as the side view ground truth $\\mathbf{O}^{\\mathrm{side}}$ .", + "bbox": [ + 88, + 383, + 483, + 565 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, the above strategy only satisfies the ground truth requirements of the two losses, but does not consider the model input. Indeed, when computing the second loss $\\mathcal{L}_{\\theta}^{\\mathrm{side}}$ Eq. (7), where the MPI frontal cameras $\\phi^{\\mathrm{mpi}}$ are set to the reference camera, the training video frames are used as the ground truth to supervise the side view of MPs. However these frames cannot be used as input of the diffusion model, as they are no longer pixel-aligned with the MPs, which are constructed using the reference camera in this setting. To tackle this problem, we propose a bootstrapping training strategy that allows our model to generate the pseudo ground truth under the reference camera by itself.", + "bbox": [ + 89, + 569, + 483, + 751 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Bootstrapping. In order to bootstrap our model for generating pseudo ground truth under the reference camera, we firstly train our model $g_{\\theta}(.)$ with only the loss $\\mathcal{L}_{\\theta}^{\\mathrm{mpi}}$ Eq. (6) (i.e. setting the MPI frontal cameras $\\phi^{\\mathrm{mpi}}$ to the target camera). In this way, the model learns to generate plausible frontal views rendered from MPs.", + "bbox": [ + 89, + 758, + 483, + 851 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With this bootstrapping, we can then enable the second loss $\\mathcal{L}_{\\theta}^{\\mathrm{side}}$ Eq. (7) to learn 3D structures of the MPs. More specifically, we first set the MPI frontal cameras $\\phi^{\\mathrm{mpi}}$ to the", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "reference camera, then run the diffusion sampling with $k$ steps to generate the output MPI video $\\mathbf{S}_k$ , and render the frontal view to get images $\\hat{\\mathbf{O}}_k^{\\mathrm{mpi}}$ . Since the model has been trained to give plausible frontal views, the rendered images $\\hat{\\mathbf{O}}_k^{\\mathrm{mpi}}$ will have a reasonable quality, acting as the pseudo ground truth under the reference camera. Thus we can add noise to $\\hat{\\mathbf{O}}_k^{\\mathrm{mpi}}$ to obtain the model input $\\mathbf{O}_t^{\\mathrm{mpi}}$ when computing the second loss $\\mathcal{L}_{\\theta}^{\\mathrm{side}}$ Eq. (7) to train our model for learning 3D shapes inside MPsIs.", + "bbox": [ + 511, + 383, + 906, + 521 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Late-stage Noise Sampling. When training with the second loss $\\mathcal{L}_{\\theta}^{\\mathrm{side}}$ Eq. (7), we compute the model input from the pseudo ground truth $\\hat{\\mathbf{O}}_k^{\\mathrm{mpi}}$ , which cannot perfectly match the real image distribution. Such distribution mismatch could harm the model learning and subsequently lead to blurry results. To mitigate this negative influence, when training with loss $\\mathcal{L}_{\\theta}^{\\mathrm{side}}$ Eq. (7), we only sample $t$ from the second half of the forward Markovian process $[T/2, T]$ so that the added noise is relatively large thus has a better chance to hide the distribution mismatch.", + "bbox": [ + 511, + 527, + 908, + 680 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Loss Functions. During training, we compose the MPs and a background image calculated from the reference portrait $\\mathbf{O}^{\\mathrm{ref}}$ with alpha blending to get the foreground renderings and the full image renderings. Using pre-computed foreground masks of the data videos, we calculate the standard L2 losses Eq. (6) and Eq. (7) on both the foreground renderings and the full image renderings. Following Tewari et al. [35], we additionally adopt several regularization losses besides the standard L2 losses Eq. (6) and Eq. (7). The regularization losses include: 1) A LPIPS perceptual loss $\\mathcal{L}_{\\mathrm{lpips}}$ for foreground renderings; 2) A mask loss $\\mathcal{L}_{\\mathrm{mask}}$ to supervise the rendered alpha mask; 3) An edge-aware depth smoothing loss and a disparity loss following Tucker et al. [36] to further supervise the underlying geometry,", + "bbox": [ + 511, + 689, + 908, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21111", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/dd36d299f27cae28dddc731a48352e05b7a2cc3c11876bac43d667aab7cce981.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. Quantitative comparison and ablation study. We compare our method with previous work on both HDTF and Talkinghead1kh datasets. Our method achieves comparable image quality (L1, LPIPS, FID) and the best video quality (FVD) and novel-view rendering efficiency overall. Compared to ablation cases, the averaged metrics are close, and we highlight the actual huge gaps in perceived image, geometry, and side view rendering quality in Fig. 7 and Fig. 6. All the render speeds are tested on a single Nvidia A100 GPU. *We report the generation speeds of diffusion models for Follow-your-emoji [29] and X-Portrait [42], as they do not support explicit camera control." + ], + "table_body": "
MethodHDTFTalkinghead1khNovel View Render Speed
L1 ↓LPIPS ↓FID ↓FVD ↓L1 ↓LPIPS ↓FID ↓FVD ↓
Face-V2V [40]0.0290.13220.5778.110.0400.18630.52141.0222 FPS in 2562
EmoPortrait [16]0.0390.18924.77140.510.0580.25241.08280.227 FPS in 5122
Portrait4d-v2 [12]--25.74139.11--37.38227.3210 FPS in 5122
Follow-your-emoji [29]0.0350.11516.09117.540.0470.15919.37209.910.5FPS* in 5122
X-Portrait [42]0.0340.11915.66132.900.0450.16419.86183.050.1FPS* in 5122
Ours0.0310.11815.9369.040.0450.18027.83135.12109 FPS in 5122
Ours w/o residual0.0300.11614.7665.750.0440.17428.19127.11109 FPS in 5122
Ours w/o late-stage0.0310.12316.7070.690.0450.18029.86136.07109 FPS in 5122
", + "bbox": [ + 125, + 88, + 870, + 238 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where we use the disparity maps rendered from the parametric model as pesudo ground truth for disparity loss.", + "bbox": [ + 89, + 313, + 480, + 343 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Run-time Inference", + "text_level": 1, + "bbox": [ + 89, + 351, + 277, + 366 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "When generating a 3D talking head from a driving video, we set the MPI frontal cameras $\\phi^{\\mathrm{MPI}}$ to the driving video. In order to generate long video sequence and smooth the discontinuity between adjacent video clips, we input ground truth first frame to the model in a similar way as Lumiere [5] with random drop during training. During inference, we take the last frame from the previous clip as the first frame of next clip. The generated first frame is then dropped due to overlapping when concatenating two adjacent clips.", + "bbox": [ + 89, + 373, + 482, + 508 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 521, + 222, + 537 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we evaluate our method by comparing with prior arts on talking heads image quality, temporal quality, multi-view consistency, and computational efficiency. We also conduct qualitative comparisons and ablation studies.", + "bbox": [ + 89, + 546, + 482, + 606 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. We train our model on a self-collected dataset consisting of 35k in-the-wild talking head monocular videos. Each video has a length between 4 to 6 seconds, and we crop the face region in $512 \\times 512$ for training. Our model generates MPI video with 32 frames in a single network forwarding, and each MPI consists of 24 planes. We dynamically set the near and far planes based on the distance between the 3DMM head joint and the camera position. We optimize our model end-to-end on 8 NVIDIA H100 GPUs with a batch size of 8. During inference, we use DDIM sampling and do classifier-free guidance on first-frame and reference portrait with two scales [8]. The guidance scale for the reference portrait is set to 1.5, while the guidance scale for the first frame is linearly reduced from 1 to 0 between frames 1 and 16. For the remaining 16 frames, the guidance scale for the first frame is fixed at 0.", + "bbox": [ + 89, + 609, + 482, + 851 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Benchmark. We perform our qualitative comparison on two widely used public talking-head datasets, Talking-head1kh [40] and HDTF [47] datasets. For HDTF dataset,", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "we randomly sample video clips from 100 identities and each video clip consists of 180 frames. For Talkinghead1kh dataset, we randomly sample two 180-frame video clip from all evaluation identities. During testing, the first frame of each video clip is chosen as reference portrait $\\mathbf{O}^{\\mathrm{src}}$ and all the animations are performed in a self-reenactment way. Cross-identity driving results are provided in the supp.", + "bbox": [ + 511, + 313, + 906, + 419 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We compare our method with existing one-shot photoreal talking heads approaches trained from monocular videos in the wild. Specifically, we compare with representatives for three families of approaches: 1) Feed-forward models (e.g. Face-V2V [40] and EmoPortrait [16]). These methods do not use diffusion model but directly build 3D implicit feature volumes (as appearance) driven by warping fields (as expression). These methods are often optimized for the controllability in 2D space, but not the generation quality or 3D consistency. 2) 3D NeRF-based one-shot talking heads (e.g. Portrait4D-v2 [11]), which establish the standard for 3D consistency due to the use of native 3D representation. 3) Video diffusion based methods (e.g. follow-your-emoji [29] and xportrait [42]), which tend to be particularly strong on the quality of the generated images or videos. We perform all the evaluations on $512 \\times 512$ resolution since most of the baselines outputs results at it, except Face-V2V for which we upsample its outputs to the desired resolution.", + "bbox": [ + 511, + 422, + 908, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Appearance and Motion Quality", + "text_level": 1, + "bbox": [ + 511, + 724, + 797, + 739 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following prior works, we compare image quality with L1, LPIPS [46] and FID score [22] and video quality, a combination of images and motions, using FVD score [37]. Note that L1 and LPIPS are not available for Portrait4D-v2 due to misalignments between its outputs and ground truth.", + "bbox": [ + 511, + 748, + 905, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative results are shown in Tab. 1. While the ranking based on different metrics varies, our method remains competitive among the state-of-the-art. Particularly, our model consistently achieves the best FVD on both datasets indicating our overall best video quality, taking into con", + "bbox": [ + 511, + 825, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21112", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c642acab0bdae3620bcbc091ad2797a5d80b15f06d73bba686a3cc8c1aaeb3eb.jpg", + "image_caption": [ + "Ref." + ], + "image_footnote": [], + "bbox": [ + 96, + 103, + 196, + 180 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/89d8c590740daa691c9164a8d1259823ed4ae3bc426434c384cfa5e6e4400ac2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 181, + 196, + 258 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b83c9d19b347d4453050684c07302dd0e0b90d24ba62ac468b56d7d760164420.jpg", + "image_caption": [ + "Face-v2v" + ], + "image_footnote": [], + "bbox": [ + 209, + 103, + 307, + 180 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f4d5f62e09ef39015552657f33bed2260f98f06afdd0bc6b6461a55b08d2fd4c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 209, + 181, + 307, + 258 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5a334b689c72cdeecab37b656c19d053afa8834bd64d7b37a116f458a440bafb.jpg", + "image_caption": [ + "EmoPortrait" + ], + "image_footnote": [], + "bbox": [ + 308, + 103, + 406, + 180 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8867b80361146200bd439127ec451ae97fe61928cb97f6db43e320712bc07da9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 308, + 181, + 406, + 258 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3290d83bc362a81ab2166fc022a5702389e065d768a7227d79a47ceeb15ace4d.jpg", + "image_caption": [ + "Portrait4D-v2" + ], + "image_footnote": [], + "bbox": [ + 408, + 103, + 504, + 180 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e06bdd21c9c69a581b3515d9ee45d5f360e2ad0834d6411fcf4a243796cee6c2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 406, + 181, + 504, + 258 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bf16d72210ca759e0af88f6bedaa2e36764141e0b8149cdc081b8a143429d3b3.jpg", + "image_caption": [ + "Follow-your-emoji" + ], + "image_footnote": [], + "bbox": [ + 504, + 103, + 602, + 180 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e26af30bea92d49287fafb895755257f60cc0ed8b17faa2f15bf604059e6d859.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 181, + 602, + 258 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1d62fa7534ab4fdadf1a2526177418b6ce516889c0219eb2482ed771aebd879d.jpg", + "image_caption": [ + "X-Portrait" + ], + "image_footnote": [], + "bbox": [ + 604, + 103, + 702, + 180 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c11f77920000cbb00879dd7af84f18dbb75b756b22d41c3d2fdb402f10161ed4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 604, + 181, + 702, + 258 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/819ee2f58c22cdd2813a13c851fdebb7c98e8715291a0e4283fde975dcdc8a42.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 702, + 103, + 800, + 180 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a6a96bb2f37ab399be0d1ce40d8efd7af30a48215e39c6a11d76081493698387.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 702, + 181, + 800, + 258 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9d1455dc7359830834edeffff48465822150ba616f41e64e8b7f867a199aadd3.jpg", + "image_caption": [ + "G.T." + ], + "image_footnote": [], + "bbox": [ + 802, + 103, + 901, + 180 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9c34e39dcffc4fb317309105cfb4bd12883a051394dd6d19fb15af1d5a0d02d5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 802, + 181, + 901, + 258 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8018ac97d1feb9051449f7bae0cf8d596aac1e0b583f0eb578465243a7ac1335.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 261, + 196, + 337 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/082ef2c299f773c6a8c95823377446841f38d2387ae9d48e6fa6fb5ae93a1970.jpg", + "image_caption": [ + "Figure 4. Talking head results. We show results from previous work and our method. Our method generates results with sharp appearance and closely aligned expression with the ground truth. Note Portrait4d-v2 uses a customized camera space which is non-trivial to render GT camera aligned images." + ], + "image_footnote": [], + "bbox": [ + 96, + 339, + 196, + 415 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/62202f3ee1b0a73439096147aa888cb2af8f74743e780b1d8b97c33a7a1e2d59.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 209, + 260, + 307, + 337 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4b3d4bfcd46a86e9ce83ad3279312c5330806981d53a613c5ed8c1a8155444b5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 209, + 339, + 307, + 414 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fd1c3105532f0c6ef3831eaeb5f3abff33b9dc4c6f8b246af675f641b129cca5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 308, + 260, + 406, + 337 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f09e6d859d68de1c1f35ec30ff3a903b5c2d2ae188271eabc34e315fbc8acb79.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 308, + 339, + 406, + 414 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/14c3b381bddb55639298d9d9ed44faaacbaaf50321c41ef58b281a8335951386.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 408, + 260, + 504, + 337 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0d6e38aa068c015d03add82d7db30276cb1635c26b6f7248592b44cb2c27e84d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 408, + 339, + 504, + 414 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3270e05dad7b509d8839c76b065adcb5f12ec1faecdec2fc6dc2511bef5aa715.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 260, + 602, + 337 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/74b2d41c0208482947589e9817352b004e185e0a72df7246498240873719f934.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 339, + 602, + 414 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9b3debf6ba0ffd62bcfa5ce98b55734ecc1da9b30855356c79a96c676fe95d99.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 604, + 260, + 702, + 337 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/72b284bbe4bcfc591cc3121a2769582a283f07228b53a006c83f3cf2ed1e119e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 604, + 339, + 702, + 414 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/053a66fc000a54e5cc345000c96162b2c5a6ae88c65651b93e847095cbcb1227.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 702, + 260, + 800, + 337 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/27dfd1e2dd4a220d94eede647ae90873a55063beb59eb237e004346675ddd35f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 702, + 339, + 800, + 414 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7f9e61f023e45e723af21658a2fff36d64603374f16d5dc5e1930f961ff22088.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 802, + 260, + 901, + 337 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cda0833a21bf6c6b007c20c10de2245250045513c9cf7028c4163f14848c47c8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 802, + 339, + 901, + 414 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sideration both image quality and motion naturalness. This also indicates that a native video diffusion model is stronger than a finetuned image diffusion with added temporal component (like X-Portrait [42] and Follow-your-emoji [29]).", + "bbox": [ + 89, + 463, + 482, + 523 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also show qualitative results in Fig. 4. FaceV2V [40] gives stable results and follows the input identity, but produces less details due to its low output resolution. EmoPortrait [16] gives sharper results, but suffers from larger identity shifts. Portrait4D-v2 [12] produces good sharpness and less identity shifts, however relatively stiff expressions. The diffusion-based methods Follow-your-emoji [29] and X-portrait [42] render good high-frequency details, but do not precisely follow the input head poses or expressions. In contrast, our method gives good image quality, faithful identity, while follows the input control signal.", + "bbox": [ + 89, + 523, + 482, + 690 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Effectiveness for 3D Rendering", + "text_level": 1, + "bbox": [ + 89, + 696, + 366, + 714 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate the effectiveness on 3D of different methods from the following aspects. First, we consider the task of binocular stereo rendering, which is one of the foundational use cases for VR/AR applications. Also, we compare how well these approaches perform on perspective geometry by moving the camera closer.", + "bbox": [ + 89, + 719, + 482, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Note that Face-V2V [40] and EmoPortrait [16] use orthogonal projection as a simplified camera model, thus they cannot produce any reasonable renderings for stereo pairs (i.e. two views will have the same disparity across the whole image) and moving camera closer (i.e. getting identical renderings). Follow-your-emoji [29] and X-xportrait [42] only", + "bbox": [ + 89, + 810, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "use detected landmarks and images as the driving signal but do not allow any explicit control on the camera, thus they cannot perform these two 3D tasks neither. As a result, we only show the renderings from Portrait4D-v2 [12] and ours, where truth 3D rendering is supported.", + "bbox": [ + 511, + 463, + 906, + 539 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Binocular Stereo Rendering. We render stereo pairs with a 5cm baseline looking from 1m in front of the talking head, and run an off-the-shell stereo matching method [27] to get the disparity map. In Fig. 5, ours gives stereo pairs with reasonable estimated disparities, and is competitive with the NeRF-based Portrait4D-v2 [12], which acts as a strong baseline on 3D effects, while being one order of magnitude faster in novel view rendering (Tab. 1).", + "bbox": [ + 511, + 539, + 906, + 659 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Perspective Geometry. To demonstrate the perspective effect, we render a close view by moving the camera closer while reducing camera focal length to keep the whole head in view. As in Fig. 5, ours produces good perspective distortions that is comparable to the NeRF-based Portrait4Dv2 [12], while being more faithful to the reference identity shown in Fig. 4 second row.", + "bbox": [ + 511, + 660, + 905, + 765 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation", + "text_level": 1, + "bbox": [ + 511, + 773, + 617, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "MPI as representation. Our diffusion model generates MPI frames, where each of them has a frontal, a residual, and a set of planes with alpha. While the frontal contains most of the appearance, we hereby evaluate the importance of the residual image. We train our model producing a MPI with only a frontal image, and reuse it for all the planes for alpha composition. While the metrics on frontal view", + "bbox": [ + 511, + 794, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21113", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6194596807da3763c55474691f13be5355903e48a9f6c9faa49e9689229dab22.jpg", + "image_caption": [ + "Portrait4d", + "original" + ], + "image_footnote": [], + "bbox": [ + 99, + 99, + 194, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2f077548f90113150eef8303a6fb466fae0942714cbde3314480a59f18ca3302.jpg", + "image_caption": [ + "stereo" + ], + "image_footnote": [], + "bbox": [ + 196, + 99, + 292, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7271bb85012a4af6de2a3a62bcfbf31b92bd667b80f115870cc59e34b74bfbbf.jpg", + "image_caption": [ + "ours", + "original" + ], + "image_footnote": [], + "bbox": [ + 299, + 99, + 395, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8689c874ca36f49b83a75295bb652f024cfa28b1e390dacf8497298847e41bc4.jpg", + "image_caption": [ + "stereo" + ], + "image_footnote": [], + "bbox": [ + 398, + 99, + 496, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f7e2513c9298989e9ff4f8e54b4f3f6c2cd78d749a124f34523585a24994a9e2.jpg", + "image_caption": [ + "close view" + ], + "image_footnote": [], + "bbox": [ + 99, + 186, + 194, + 260 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d77810a3249f6029a9324027de75663dea15eff25c1ad28161d17a3469ba3b0b.jpg", + "image_caption": [ + "disparity" + ], + "image_footnote": [], + "bbox": [ + 196, + 186, + 292, + 260 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4be5362c1939e65175a089c41581c72f1590ff05e9ce3db2597b23f1a790888e.jpg", + "image_caption": [ + "close view" + ], + "image_footnote": [], + "bbox": [ + 299, + 186, + 395, + 261 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d92ce221da6afc2556b3e2e514747bf2b0b4c6d2285333e7f2d19490e7ab0b7c.jpg", + "image_caption": [ + "disparity" + ], + "image_footnote": [], + "bbox": [ + 398, + 186, + 496, + 261 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5fc468d1ae9ea3a518f9e5b2836af06a23cfc45a8699ffe48d5ffec0bea91f1e.jpg", + "image_caption": [ + "Portrait4d", + "original" + ], + "image_footnote": [], + "bbox": [ + 504, + 99, + 599, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4ab73827595ad5c269d7aa175ad5ff213671a620d86511d38c4c720798cb6e82.jpg", + "image_caption": [ + "stereo" + ], + "image_footnote": [], + "bbox": [ + 601, + 99, + 697, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a706d7d8f16f9252adcb13dd0a3aca2fcb2689deac8606b0ef1035f071f4596a.jpg", + "image_caption": [ + "ours", + "original" + ], + "image_footnote": [], + "bbox": [ + 705, + 99, + 802, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/06c43bfda5347cd84dde2db9ab0725cafe168af9adff40f58aac97e7c1aa96c0.jpg", + "image_caption": [ + "stereo" + ], + "image_footnote": [], + "bbox": [ + 805, + 99, + 898, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/817280bcaf7387ae19f5f793e5d4ec39250fb286b1f2d1abca04c533f34359f1.jpg", + "image_caption": [ + "close view" + ], + "image_footnote": [], + "bbox": [ + 504, + 186, + 601, + 260 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/20134c948b0e954bc6f27bab96dd28ef6158162b9cdc6bd7d1ff1443f9e91e0a.jpg", + "image_caption": [ + "disparity" + ], + "image_footnote": [], + "bbox": [ + 602, + 186, + 697, + 261 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cede3950382761297e365b735eda01711768badaf49c25639628209f8d98a4e0.jpg", + "image_caption": [ + "close view" + ], + "image_footnote": [], + "bbox": [ + 705, + 186, + 802, + 261 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/db0e4f4da297a193b7040f5cdab4b8cc9ace940530aa815e1182d929033db986.jpg", + "image_caption": [ + "disparity" + ], + "image_footnote": [], + "bbox": [ + 803, + 186, + 898, + 261 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f912fddce4e25d30ea339c9a9eb7fa0d1a4db027fb55e56a7378d05cd57af512.jpg", + "image_caption": [ + "Figure 5. Binocular stereo and perspective effects. Our method generates MPI videos that enables efficient spatial rendering, e.g. binocular stereo and perspective effects. We render stereo pairs and calculate a disparity using RAFT-Stereo [27] to visualize the perceived geometry. To demonstrate perspective effects, we render a view with camera moving closer to the subject. Our method renders visually plausible and comparable 3D effects with NeRF based approach like Portrait4d-v2, while being orders of magnitude faster.", + "Figure 6. Effect of residual RGB. We retrain our model with only a single frontal image in the MPI. We show both the rendered side view images and the disparity converted from the MPI alpha in frontal view. We found this significantly reduces the capacity of MPI, which leads to noisy and irregular geometry and blatant blurriness near the occlusion boundaries." + ], + "image_footnote": [], + "bbox": [ + 96, + 335, + 478, + 555 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "become even stronger compared to the full model (1 (Ours w/o residual)), the model produces much worse geometry (Fig. 6), presumable due to the missing of capacity of handling view dependent effects. It also leads to obvious blurriness in the boundary when rendering from the side view.", + "bbox": [ + 89, + 655, + 482, + 731 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A further degenerated case is to use a RGB image and depth for novel-view rendering, and please check our supp for more comparison.", + "bbox": [ + 89, + 732, + 482, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Late-stage noise sampling. Late-stage noise sampling is also important to learn sharp images since that significantly reduces the negative impact from the imperfect pseudo ground truth. To verify this, we train a model with full noise sampling, and show the quantitative and qualitative comparison in Tab. 1 and Fig. 7. Without large noise sampling, the diffusion model still produces reasonable structure but missing details, e.g. wrinkles, mold, hair strand.", + "bbox": [ + 89, + 779, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ae234078cb16a5a801a65afdbac80c7ebd1ec1c52c1370d9ed343f6fec06426e.jpg", + "image_caption": [ + "Figure 7. Effect of Late-stage noise sampling. We show a comparison between our model trained with and without the Late-stage noise sampling. Training with regular noise sampling leads to excessive blurriness and lost of details." + ], + "image_footnote": [], + "bbox": [ + 519, + 335, + 901, + 518 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 597, + 632, + 612 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present a method to learn a 3D-aware video diffusion model, which generates photorealistic talking head multi-plane images given an identity reference image and a sequence of tracking signals. We show such a model can be learned directly from large scale monocular videos without multi-view datasets. With appropriate learning strategies, both 3D information and high quality image generation can be learned concurrently. Extensive experiments show that, our method generates not only talking head video with quality comparable and even better than existing arts but also 3D rendering efficiently that support spatial viewing experience, like in VR headset. As limitations, the generated talking head video does not support rendering from excessive camera viewpoint changes, which is a common issue of the MPI representation. Extending our framework with other efficient 3D representations or enhancing training with multi-view dataset could be a promising direction to improve both the image and 3D quality.", + "bbox": [ + 511, + 628, + 906, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21114", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Titas Anciukevicius, Zexiang Xu, Matthew Fisher, Paul Henderson, Hakan Bilen, Niloy J Mitra, and Paul Guerrero. Renderdiffusion: Image diffusion for 3d reconstruction, inpainting and generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12608-12618, 2023. 3, 4", + "[2] ShahRukh Athar, Zexiang Xu, Kalyan Sunkavalli, Eli Shechtman, and Zhixin Shu. Rignerf: Fully controllable neural 3d portraits. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 20364-20373, 2022. 3", + "[3] Ziqian Bai, Feitong Tan, Zeng Huang, Kripasindhu Sarkar, Danhang Tang, Di Qiu, Abhimitra Meka, Ruofei Du, Mingsong Dou, Sergio Orts-Escalano, et al. Learning personalized high quality volumetric head avatars from monocular rgb videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16890-16900, 2023.", + "[4] Ziqian Bai, Feitong Tan, Sean Fanello, Rohit Pandey, Mingsong Dou, Shichen Liu, Ping Tan, and Yinda Zhang. Efficient 3d implicit head avatar with mesh-anchored hash table blendshapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1975–1984, 2024. 3", + "[5] Omer Bar-Tal, Hila Chefer, Omer Tov, Charles Herrmann, Roni Paiss, Shiran Zada, Ariel Ephrat, Junhwa Hur, Guanghui Liu, Amit Raj, et al. Lumiere: A spacetime diffusion model for video generation. arXiv preprint arXiv:2401.12945, 2024. 4, 6", + "[6] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 157-164. 2023. 4", + "[7] Stella Bounareli, Christos Tzelepis, Vasileios Argyriou, Ioannis Patras, and Georgios Tzimiropoulos. Hyperreenact: One-shot reenactment via jointly learning to refine and retarget faces. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2", + "[8] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 6", + "[9] Bindita Chaudhuri, Noranart Vesdapunt, Linda Shapiro, and Baoyuan Wang. Personalized face modeling for improved face reconstruction and motion retargeting. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part V 16, pages 142-160. Springer, 2020. 2", + "[10] Yufan Chen, Lizhen Wang, Qijing Li, Hongjiang Xiao, Shengping Zhang, Hongxun Yao, and Yebin Liu. Monogaussianavatar: Monocular gaussian point-based head avatar. In ACM SIGGRAPH 2024 Conference Papers, pages 1-9, 2024. 3", + "[11] Yu Deng, Duomin Wang, and Baoyuan Wang. Portrait4d-v2: Pseudo multi-view data creates better 4d head synthesizer. arXiv preprint arXiv:2403.13570, 2024. 3, 6" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Yu Deng, Duomin Wang, and baoyuan Wang. Portrait4d-v2: Pseudo multi-view data creates better 4d head synthesizer. arXiv, 2024. 2, 3, 6, 7", + "[13] Zheng Ding, Xuaner Zhang, Zhihao Xia, Lars Jebe, Zhuowen Tu, and Xiuming Zhang. Diffusionrig: Learning personalized priors for facial appearance editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12736-12746, 2023. 2", + "[14] Michail Christos Doukas, Stefanos Zafeiriou, and Viktoriia Sharmanska. Headgan: One-shot neural head synthesis and editing. In Proceedings of the IEEE/CVF International conference on Computer Vision, pages 14398-14407, 2021. 2", + "[15] Nikita Drobyshev, Jenya Chelishev, Taras Khakhulin, Aleksei Ivakhnenko, Victor Lempitsky, and Egor Zakharov. Megaportraits: One-shot megapixel neural head avatars. In Proceedings of the 30th ACM International Conference on Multimedia, pages 2663–2671, 2022. 2", + "[16] Nikita Drobyshev, Antoni Bigata Casademunt, Konstantinos Vougioukas, Zoe Landgraf, Stavros Petridis, and Maja Pantic. Emoportraits: Emotion-enhanced multimodal one-shot head avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8498-8507, 2024. 2, 6, 7", + "[17] Emilien Dupont, Hyunjik Kim, SM Eslami, Danilo Rezende, and Dan Rosenbaum. From data to functa: Your data point is a function and you can treat it like one. arXiv preprint arXiv:2201.12204, 2022. 3", + "[18] Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8649-8658, 2021. 3", + "[19] Xuan Gao, Chenglai Zhong, Jun Xiang, Yang Hong, Yudong Guo, and Juyong Zhang. Reconstructing personalized semantic facial nerf models from monocular video. ACM Transactions on Graphics (TOG), 41(6):1-12, 2022. 3", + "[20] Philip-William Grassal, Malte Prinzler, Titus Leistner, Carsten Rother, Matthias Nießner, and Justus Thies. Neural head avatars from monocular rgb videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18653–18664, 2022. 2", + "[21] Jianzhu Guo, Dingyun Zhang, Xiaogiang Liu, Zhizhou Zhong, Yuan Zhang, Pengfei Wan, and Di Zhang. Liveportrait: Efficient portrait animation with stitching and retargeting control. arXiv preprint arXiv:2407.03168, 2024. 2", + "[22] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6", + "[23] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3", + "[24] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023. 3", + "[25] Yushi Lan, Fangzhou Hong, Shuai Yang, Shangchen Zhou, Xuyi Meng, Bo Dai, Xingang Pan, and Chen Change Loy." + ], + "bbox": [ + 516, + 92, + 906, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21115", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Ln3diff: Scalable latent neural fields diffusion for speedy 3d generation. In European Conference on Computer Vision, pages 112-130. Springer, 2025. 3", + "[26] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph., 36(6):194-1, 2017. 4", + "[27] Lahav Lipson, Zachary Teed, and Jia Deng. Raft-stereo: Multilevel recurrent field transforms for stereo matching. In 2021 International Conference on 3D Vision (3DV), pages 218-227. IEEE, 2021. 7, 8", + "[28] Tao Liu, Chenpeng Du, Shuai Fan, Feilong Chen, and Kai Yu. Diffhub: Person-generic visual dubbing using inpainting renderer with diffusion auto-encoder. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 3630-3634. IEEE, 2024. 2", + "[29] Yue Ma, Hongyu Liu, Hongfa Wang, Heng Pan, Yingqing He, Junkun Yuan, Ailing Zeng, Chengfei Cai, Heung-Yeung Shum, Wei Liu, et al. Follow-your-emoji: Fine-controllable and expressive freestyle portrait animation. arXiv preprint arXiv:2406.01900, 2024. 2, 4, 6, 7", + "[30] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2", + "[31] Norman Müller, Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Bulo, Peter Kontschieder, and Matthias Nießner. Diffrf: Rendering-guided 3d radiance field diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4328-4338, 2023. 3", + "[32] Shenhan Qian, Tobias Kirschstein, Liam Schoneveld, Davide Davoli, Simon Giebenhain, and Matthias Nießner. Gaussianavatars: Photorealistic head avatars with rigged 3d gaus-sians. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20299-20309, 2024. 3", + "[33] Shunsuke Saito, Gabriel Schwartz, Tomas Simon, Junxuan Li, and Giljoo Nam. Relightable gaussian codec avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 130-141, 2024. 3", + "[34] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. Advances in neural information processing systems, 32, 2019. 2", + "[35] Ayush Tewari, Tianwei Yin, George Cazenavette, Semon Rezhikov, Josh Tenenbaum, Frédo Durand, Bill Freeman, and Vincent Sitzmann. Diffusion with forward models: Solving stochastic inverse problems without direct supervision. Advances in Neural Information Processing Systems, 36: 12349-12362, 2023. 3, 4, 5", + "[36] Richard Tucker and Noah Snavely. Single-view view synthesis with multiplane images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 551-560, 2020. 3, 5", + "[37] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. To" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "wards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 6", + "[38] Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, Karsten Kreis, et al. Lion: Latent point diffusion models for 3d shape generation. Advances in Neural Information Processing Systems, 35:10021-10039, 2022. 3", + "[39] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, et al. Rodin: A generative model for sculpting 3d digital avatars using diffusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4563-4573, 2023. 3", + "[40] Ting-Chun Wang, Arun Mallya, and Ming-Yu Liu. One-shot free-view neural talking-head synthesis for video conferencing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10039-10049, 2021. 2, 6, 7", + "[41] Zhen Wang, Qiangeng Xu, Feitong Tan, Mengei Chai, Shichen Liu, Rohit Pandey, Sean Fanello, Achuta Kadambi, and Yinda Zhang. Mvdd: Multi-view depth diffusion models. In European Conference on Computer Vision, pages 236-253. Springer, 2025. 3", + "[42] You Xie, Hongyi Xu, Guoxian Song, Chao Wang, Yichun Shi, and Linjie Luo. X-portrait: Expressive portrait animation with hierarchical motion attention. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2, 4, 6, 7", + "[43] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. Styleheat: One-shot high-resolution editable talking face generation via pre-trained stylegan. In European conference on computer vision, pages 85-101. Springer, 2022. 2", + "[44] Egor Zakharov, Aleksei Ivakhnenko, Aliaksandra Shysheya, and Victor Lempitsky. Fast bi-layer neural synthesis of one-shot realistic head avatars. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XII 16, pages 524-540. Springer, 2020. 2", + "[45] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 4", + "[46] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6", + "[47] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661–3670, 2021. 6", + "[48] Wojciech Zielonka, Timo Bolkart, and Justus Thies. Instant volumetric head avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4574-4584, 2023. 3" + ], + "bbox": [ + 516, + 92, + 903, + 880 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21116", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/IM-Portrait_ Learning 3D-aware Video Diffusion for Photorealistic Talking Heads from Monocular VideosC/d89a7508-cd12-421c-9424-93ad28adceb7_model.json b/2025/IM-Portrait_ Learning 3D-aware Video Diffusion for Photorealistic Talking Heads from Monocular VideosC/d89a7508-cd12-421c-9424-93ad28adceb7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d22d5bb892ee04bca63c5eeb928cb9d6abdb553a --- /dev/null +++ b/2025/IM-Portrait_ Learning 3D-aware Video Diffusion for Photorealistic Talking Heads from Monocular VideosC/d89a7508-cd12-421c-9424-93ad28adceb7_model.json @@ -0,0 +1,2893 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.163, + 0.131, + 0.838, + 0.177 + ], + "angle": 0, + "content": "IM-Portrait: Learning 3D-aware Video Diffusion for Photorealistic Talking Heads from Monocular Videos" + }, + { + "type": "text", + "bbox": [ + 0.191, + 0.202, + 0.803, + 0.26 + ], + "angle": 0, + "content": "Yuan Li\\(^{1,2*§}\\) Ziqian Bai\\(^{2*}\\) Feitong Tan\\(^{2}\\) Zhaopeng Cui\\(^{1\\dagger}\\) Sean Fanello\\(^{2}\\) Yinda Zhang\\(^{2\\dagger}\\) \n\\(^{1}\\)State Key Lab of CAD&CG, Zhejiang University \\(^{2}\\)Google" + }, + { + "type": "image_caption", + "bbox": [ + 0.135, + 0.271, + 0.187, + 0.285 + ], + "angle": 0, + "content": "Inputs" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.287, + 0.231, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.289, + 0.271, + 0.431, + 0.284 + ], + "angle": 0, + "content": "Generated Results" + }, + { + "type": "image", + "bbox": [ + 0.238, + 0.287, + 0.363, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.287, + 0.495, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.434, + 0.229, + 0.575 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.238, + 0.434, + 0.363, + 0.573 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.434, + 0.496, + 0.572 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.546, + 0.271, + 0.596, + 0.285 + ], + "angle": 0, + "content": "Inputs" + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.286, + 0.638, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.699, + 0.271, + 0.841, + 0.284 + ], + "angle": 0, + "content": "Generated Results" + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.286, + 0.772, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.778, + 0.287, + 0.904, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.434, + 0.638, + 0.573 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.646, + 0.434, + 0.772, + 0.573 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.778, + 0.434, + 0.904, + 0.573 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.581, + 0.908, + 0.637 + ], + "angle": 0, + "content": "Figure 1. We propose a 3D-aware video diffusion model for talking head synthesis. Given an image as identity and a sequence of tracking signals (as shown on the left for each example), our model directly generates videos in Multiplane Images (MPIs) in a single denoising process, which is ready for efficient novel-view rendering. This enables immersive viewing experience, e.g. rendering binocular stereo or perspective distortion in VR headset." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.648, + 0.327, + 0.664 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.682, + 0.484, + 0.849 + ], + "angle": 0, + "content": "We propose a novel 3D-aware diffusion-based method for generating photorealistic talking head videos directly from a single identity image and explicit control signals (e.g., expressions). Our method generates Multiplane Images (MPIs) that ensure geometric consistency, making them ideal for immersive viewing experiences like binocular videos for VR headsets. Unlike existing methods that often require a separate stage or joint optimization to reconstruct a 3D representation (such as NeRF or 3D Gaussians), our approach directly generates the final output through a single denoising process, eliminating the need for post" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.65, + 0.907, + 0.803 + ], + "angle": 0, + "content": "processing steps to render novel views efficiently. To effectively learn from monocular videos, we introduce a training mechanism that reconstructs the output MPI randomly in either the target or the reference camera space. This approach enables the model to simultaneously learn sharp image details and underlying 3D information. Extensive experiments demonstrate the effectiveness of our method, which achieves competitive avatar quality and novel-view rendering capabilities, even without explicit 3D reconstruction or high-quality multi-view training data." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.83, + 0.646, + 0.846 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Synthesizing realistic talking heads, has been a long-lasting challenging problem in computer vision for decades. The recent emergence of diffusion models within the generative" + }, + { + "type": "page_footnote", + "bbox": [ + 0.108, + 0.863, + 0.269, + 0.876 + ], + "angle": 0, + "content": "*Authors contributed equally." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.876, + 0.242, + 0.888 + ], + "angle": 0, + "content": "† Corresponding authors." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.436, + 0.9 + ], + "angle": 0, + "content": "\\(^{\\S}\\)Work was conducted while Yuan Li was an intern at Google." + }, + { + "type": "list", + "bbox": [ + 0.108, + 0.863, + 0.436, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21107" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.168 + ], + "angle": 0, + "content": "AI domain offers a promising solution to this problem. Significant advancements have been made in modeling talking heads using image or video diffusion models conditioned on tracking signals [13, 28, 29, 42], leveraging the remarkable generative capabilities of diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.17, + 0.483, + 0.397 + ], + "angle": 0, + "content": "Most diffusion-based approaches, however, operate in 2D image space, and the extension of these methods to 3D rendering remains largely unexplored. One potential approach involves applying 3D transformations to the tracking signal, such as parametric models or landmarks. However, this necessitates running the reverse diffusion process for each new camera view, leading to significant computational overhead. Alternatively, some diffusion-based approaches generate multi-view images, which often lack geometric consistency. While off-the-shelf methods like NeRF [12] or 3D Gaussians can be used to construct a 3D representation for efficient rendering, the inherent inconsistencies in the input multi-view images limit the quality of the final 3D reconstruction, leading to potential issues in multi-view and temporal consistency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.4, + 0.483, + 0.612 + ], + "angle": 0, + "content": "In this work, we present, for the first time, a diffusion based approach that produces 3D talking heads that can be rendered efficiently in varying camera poses. We adopt Multiplane Image (MPI) as the 3D representation due to its remarkable rendering efficiency and quality, and the affinity to 2D images such that we can make good use of existing 2D diffusion models. Essentially, we formulate the problem of synthesizing talking heads images as a denoising diffusion process, which is conditioned on a reference image (i.e. the subject identity) and a sequence of control signals from parametric models (i.e. expressions and head poses). In a single denoising process, our model generates MPI videos that can be readily rendered in varying camera pose and support spatial viewing experience." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Our model is learned directly from a large scale monocular talking head videos in-the-wild. Due to the missing of multi-view datasets, like previous works, the 3D consistency needs to be learned from the head pose variability between video frames. To achieve this, it is required to construct MPs in a camera that looks at the head from a different viewpoint than the target images. Specifically, we use the viewpoint of the reference image to construct MPs. The estimated target images can then be rendered from MPs by moving the camera according to the relative head pose differences between the reference and target images. In this way, however, the frontal views of MPs will be misaligned with the input noisy images (in our case the target ground-truth images), which will break the training of the diffusion model. To tackle the problem, we propose a novel training framework to bootstrap our model to synthesize pseudo noisy input images that align with the MPs in reference cameras. With this, the MPs are randomly generated from either the reference camera viewpoint or the" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.136 + ], + "angle": 0, + "content": "target camera viewpoint during the training, which allows the diffusion model to learn both 3D information and fine details." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.139, + 0.751, + 0.152 + ], + "angle": 0, + "content": "Our contributions are as follows." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.155, + 0.905, + 0.184 + ], + "angle": 0, + "content": "- We propose a diffusion-based approach for 3D-aware talking heads generation." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.185, + 0.905, + 0.23 + ], + "angle": 0, + "content": "- Our diffusion model directly generates 3D renderable talking heads in a single denoising process, eliminating the need for post-processing steps." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.231, + 0.905, + 0.305 + ], + "angle": 0, + "content": "- We propose a novel training framework that enables our model to learn from monocular videos, and show how randomly selecting the camera viewpoint between reference and target images allows for better convergence and sharper results." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.306, + 0.905, + 0.397 + ], + "angle": 0, + "content": "- Extensive experiments demonstrate that our method achieves comparable or superior image and video quality to existing diffusion-based 2D talking head methods, while also offering efficient and consistent multiview rendering on par with specialized (i.e. non-diffusion based) 3D approaches." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.155, + 0.905, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.413, + 0.664, + 0.429 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.44, + 0.907, + 0.818 + ], + "angle": 0, + "content": "2D Talking Heads. In recent year, there has been a great advancement on synthesizing 2D talking heads with neural networks. A recent line of works [7, 14-16, 34, 40, 43, 44] propose to leverage neural networks to encode the appearance and motion from input images [15, 16, 34, 40] (or explicit control signals [14, 43, 44]) to some latent features or manually defined representations (e.g. key-points, warping fields, texture maps), which are injected into 2D CNN generators with various architectures for talking head synthesis. More recently, many works [13, 28, 29, 42] started to adopt 2D diffusion models, leveraging their impressive image quality, for controllable talking head synthesis. However, these approaches usually use simplified camera models (e.g. orthogonal projection) or they do not have any explicit camera control, making them unsuitable for spatial viewing experiences such as binocular stereo and perspective geometry. In addition, these methods usually do not produce explicit 3D geometry, which also limits their downstream applications. While most of prior work learns talking head models from monocular videos which are available in a large scale in-the-wild, multi-view datasets, e.g. from lightstage [21], could be leveraged to further improve fidelity, however such data requires intensive collection labor and there are no large scale multi-view datasets that are publicly available." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.825, + 0.905, + 0.901 + ], + "angle": 0, + "content": "3D Talking Heads. The problem of generating 3D talking heads has been a hot topic in the community. People attempted to model their talking heads with various 3D representations, such as textured mesh [9, 20], Neural Radiance Fields (NeRFs) [30] or other implicit func" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21108" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.228 + ], + "angle": 0, + "content": "tions [2-4, 11, 12, 18, 19, 48], and, more recently, Gaussian Splatting [10, 32, 33]. While most of these approaches are person-specific, which requires to retrain a new model for every new person, some methods [11, 12] also explore the potential of person-agnostic models with explicit 3D representations. Despite having good 3D properties, these approaches cannot easily leverage the recent advancement on 2D diffusion, which lower the upper bound of their image quality." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.236, + 0.485, + 0.522 + ], + "angle": 0, + "content": "Diffusion-based 3D Generative Model. Another line of work, investigates how to apply diffusion models on 3D representations. The most straightforward way is applying the standard diffusion process directly on 3D data [38, 41]. However, this requires large scale 3D datasets, which are generally not available. One typical workaround is to train the diffusion model with 2D images. Some works [17, 24, 25, 31, 39] firstly learn a latent space of the 3D samples with multiview 2D images, then apply diffusion on the learned latent space. However, this results in an additional latent space learning, which is non-trivial thus increasing the pipeline complexity. Some other works [1, 35] directly model 3D scene distributions with image space diffusion by integrating the 3D-to-2D rendering into the reverse process. However, these methods are mostly proposed for static scenes where multiview datasets are available. In contrast to prior research, our method handles dynamic 3D talking heads and is trained exclusively on monocular RGB videos from real-world sources." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.539, + 0.182, + 0.555 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.566, + 0.483, + 0.672 + ], + "angle": 0, + "content": "Our goal is to generate 3D talking head videos represented as Multiplane Images (MPIs), conditioned on a reference image (the subject's identity) and a sequence of control signals from a parametric model to determine expressions and head poses. The generated videos must exhibit high image quality, controllability, and efficient rendering for VR scenarios, including binocular consistency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.672, + 0.483, + 0.808 + ], + "angle": 0, + "content": "Given the scarcity of large-scale multi-view video datasets, we design our training pipeline to operate solely on 2D monocular videos from the wild. To leverage the power of 2D diffusion models, we formulate our model as a video diffusion process operating in the image space. However, this raises two key challenges: 1) How to generate 3D scenes from a diffusion denoising process defined on 2D observations, without access to ground-truth MPs? 2) How to learn 3D information from monocular 2D videos?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.483, + 0.869 + ], + "angle": 0, + "content": "To address the first challenge, we design a video diffusion pipeline integrated with a differentiable forward rendering model [35] (Sec. 3.2) to directly model 3D scene distributions from 2D monocular data." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To address the second challenge, a possible approach involves constructing MPs of a 3D talking head in the refer" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.38 + ], + "angle": 0, + "content": "ence image's camera, then moving the target camera to render the head in novel views during training. This allows us to aggregate multi-view observations of the head from different frames into a consistent 3D space to learn 3D shapes. To this end, we propose an alternating training strategy that switches between using the reference and target views as the frontal camera for MPI construction and compute the loss on target views accordingly (Sec. 3.3). This approach encourages the network to learn to synthesize high-quality, sharp images when building the MPIs in target views and to learn multiview consistency when building it in the reference view. When MPIs are constructed in the reference camera, the target ground-truth images cannot be used as input of the diffusion model, as they are not pixel-aligned with the MPIs. To tackle this, we bootstrap our model to generate the pseudo ground truth under the reference camera (Sec. 3.3). During inference, our model generate 3D scene representations in a single denoising diffusion process without any post-processing or optimization (Sec. 3.4)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.385, + 0.655, + 0.4 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.408, + 0.907, + 0.483 + ], + "angle": 0, + "content": "Diffusion Formulation. Denoising diffusion probabilistic models (DDPM) [23] are one kind of generative models that learn to capture a data distribution \\( p_{\\theta}(\\mathbf{x}) \\). The method first constructs a forward Markovian process \\( q(\\mathbf{x}_{0:T}) \\) to iteratively add noise to data as:" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.489, + 0.907, + 0.508 + ], + "angle": 0, + "content": "\\[\nq (\\mathbf {x} _ {t} | \\mathbf {x} _ {t - 1}) = \\mathcal {N} (\\mathbf {x} _ {t}; \\sqrt {1 - \\beta_ {t}} \\mathbf {x} _ {t - 1}, \\beta_ {t} \\mathbf {I}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.512, + 0.906, + 0.62 + ], + "angle": 0, + "content": "where \\(\\beta_{t}\\) is scheduled to control the amount of noise being added to data in each step. A denoising diffusion model \\(g_{\\theta}\\) with parameters \\(\\theta\\) is trained to iteratively perform the reverse process \\(p_{\\theta}(\\mathbf{x}_{t - 1}|\\mathbf{x}_t)\\) to generate a data sample \\(\\mathbf{x}_0\\) from a random Gaussian noise \\(\\mathbf{x}_T\\). If we parameterize the model \\(g_{\\theta}\\) to predict the estimated clean data sample \\(\\hat{\\mathbf{x}}_{t - 1} = g_{\\theta}(\\mathbf{x}_t,t)\\approx \\mathbf{x}_0\\), then the reverse process can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.624, + 0.905, + 0.661 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} p _ {\\theta} \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}\\right) = p _ {\\theta} \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}, \\hat {\\mathbf {x}} _ {t - 1}\\right) \\\\ = \\mathcal {N} \\left(\\mathbf {x} _ {t - 1}; \\boldsymbol {\\mu} _ {t} \\left(\\mathbf {x} _ {t}, \\hat {\\mathbf {x}} _ {t - 1}\\right), \\sigma_ {t} ^ {2} \\mathbf {I}\\right) \\tag {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.667, + 0.905, + 0.698 + ], + "angle": 0, + "content": "Please find the detailed formulation of \\(\\pmb{\\mu}_t\\) and \\(\\sigma_t^2\\) in Eq. (7) of DDPM [23]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Multiplane Images. We leverage Multiplane Images (MPIs) to represent a 3D scene S (e.g. a 3D talking head). An MPI representation consists of \\( D \\) RGBA image planes placed in a reference 3D space defined by an MPI frontal camera \\( \\phi^{\\mathrm{mpi}} \\). These planes are parallel to the image plane of the frontal camera and equally spaced in disparity between near \\( d_{n} \\) and far \\( d_{f} \\) depths from the camera. During rendering, the planes are firstly warped from the frontal camera \\( \\phi^{\\mathrm{mpi}} \\) to a target camera \\( \\phi^{\\mathrm{trgt}} \\) by a homography. The warped planes are then composited via alpha blending. We also follow Tucker et al. [36] to use only two RGB images, one as frontal, the other as residual, during MPI rendering to set a proper rendering capability of our model." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21109" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.089, + 0.885, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.35, + 0.908, + 0.407 + ], + "angle": 0, + "content": "Figure 2. Inference pipeline. Our model is built on the architecture of Lumiere [5], which takes an identity image, 2D noise video, the sequence of expressions rendered from 3DMM model and first frame image as input and outputs MPI video sequences. During inference, the network is conditioned on a reference portrait and takes the last frame of previously generated clip as the first frame condition. We separate the network into a color branch and a geometry branch. The two branches share information by zero convolution [45]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.413, + 0.383, + 0.43 + ], + "angle": 0, + "content": "3.2. 3D Talking Head Video Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.435, + 0.484, + 0.631 + ], + "angle": 0, + "content": "Unlike previous diffusion-based 2D talking head models [29, 42], our model aims to generate a sequence of 3D head representations in the form of an MPI video with \\(N\\) frames \\(\\{\\mathbf{S}\\}_{N}\\). This is achieved by conditioning the model on a reference identity image \\(\\mathbf{O}^{\\mathrm{ref}}\\) and a sequence of control signals \\(\\{\\mathbf{C}\\}_{N}\\), which are the parameters of a mesh-based parametric head model [6, 26], controlling facial expressions and head pose. Inspired by recent 3D diffusion models [1, 35], we propose a video diffusion framework that generates 3D content (an MPI video) directly from 2D observations by integrating the MPI rendering into the denoising process. More specifically, we formulate the forward process in our avatar diffusion as" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.638, + 0.483, + 0.657 + ], + "angle": 0, + "content": "\\[\nq \\left(\\mathbf {O} _ {t} ^ {\\mathrm {m p i}} \\mid \\mathbf {O} _ {t - 1} ^ {\\mathrm {m p i}}\\right) = \\mathcal {N} \\left(\\mathbf {O} _ {t} ^ {\\mathrm {m p i}}; \\sqrt {1 - \\beta_ {t}} \\mathbf {O} _ {t - 1} ^ {\\mathrm {m p i}}, \\beta_ {t} \\mathbf {I}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.663, + 0.484, + 0.817 + ], + "angle": 0, + "content": "where \\(\\mathbf{O}^{\\mathrm{mpi}}\\) is the 2D video observed by the MPI frontal camera \\(\\phi^{\\mathrm{mpi}}\\) defined in Sec. 3.1. Note that we intentionally formulate the diffusion process on the image space of MPI frontal camera since it benefits the output quality. Also, we omit the \\(\\{.\\}_{N}\\) in \\(\\{\\mathbf{O}^{\\mathrm{mpi}}\\}_{N}\\) for simplicity and do the same for \\(\\{\\phi^{\\mathrm{mpi}}\\}_{N}\\), \\(\\{\\mathbf{S}\\}_{N}\\) and \\(\\{\\mathbf{C}\\}_{N}\\). For the reverse process, instead of directly predicting the estimated clean 2D video \\(\\hat{\\mathbf{O}}_{t-1}^{\\mathrm{mpi}}\\), we let the denoising model \\(g_{\\theta}\\) to predict the MPI video \\(\\mathbf{S}_{t-1}\\), followed by the MPI rendering to obtain \\(\\hat{\\mathbf{O}}_{t-1}^{\\mathrm{mpi}}\\). Formally, we have" + }, + { + "type": "equation", + "bbox": [ + 0.184, + 0.824, + 0.483, + 0.842 + ], + "angle": 0, + "content": "\\[\n\\mathbf {S} _ {t - 1} = g _ {\\theta} \\left(\\mathbf {O} ^ {\\mathrm {r e f}}, \\mathbf {O} _ {t} ^ {\\mathrm {m p i}}; t, \\mathbf {C} ^ {\\mathrm {m p i}}\\right) \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.845, + 0.483, + 0.865 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {O}} _ {t - 1} ^ {\\mathrm {m p i}} = \\mathcal {R} \\left(\\mathbf {S} _ {t - 1}, \\phi^ {\\mathrm {m p i}}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathcal{R}\\) is the MPI rendering function. Here, we rasterize the control signals \\(\\mathbf{C}\\) into UV coordinate maps under the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.414, + 0.907, + 0.582 + ], + "angle": 0, + "content": "MPI frontal camera \\(\\phi^{\\mathrm{mpi}}\\) to provide a pixel-aligned condition \\(\\mathbf{C}^{\\mathrm{mpi}}\\). We also convert the meshes of the parametric model under \\(\\mathbf{C}\\) into alpha planes and feed them into the model. In practice, we also concatenate the identity image \\(\\mathbf{O}^{\\mathrm{ref}}\\) with its corresponding UV coordinate map rasterized from the reference camera as the model input. Finally, we plug in the estimated clean 2D video \\(\\hat{\\mathbf{O}}_{t - 1}^{\\mathrm{mpi}}\\) to Eq. (2) to perform the reverse process, which is repeated \\(T\\) times to sample the final MPI video output \\(\\mathbf{S}_0\\). Given a novel side view camera (or trajectory) \\(\\phi^{\\mathrm{side}}\\), we can simply run the MPI rendering to obtain the desired video \\(\\mathbf{O}_0^{\\mathrm{side}} = \\mathcal{R}(\\mathbf{S}_0,\\phi^{\\mathrm{side}})\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.588, + 0.849, + 0.605 + ], + "angle": 0, + "content": "3.3. Reference-Target Alternating Training" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.61, + 0.906, + 0.672 + ], + "angle": 0, + "content": "Following a similar analysis as Tewari et al. [35], a straightforward way to train our model in Sec. 3.2 needs two losses on videos rendered from two different cameras \\(\\phi^{\\mathrm{MPI}}\\) (the MPI frontal camera) and \\(\\phi^{\\mathrm{side}}\\) (a novel side view camera):" + }, + { + "type": "equation", + "bbox": [ + 0.529, + 0.678, + 0.689, + 0.696 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {O}} _ {t - 1} ^ {\\text {s i d e}} = \\mathcal {R} \\left(\\mathbf {S} _ {t - 1}, \\phi^ {\\text {s i d e}}\\right)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.699, + 0.905, + 0.719 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\theta} ^ {\\mathrm {m p i}} = \\mathbb {E} _ {\\mathbf {O} ^ {\\mathrm {r e f}}, \\mathbf {O} ^ {\\mathrm {m p i}}, \\mathbf {C} ^ {\\mathrm {m p i}}, t} [ \\| \\mathbf {O} ^ {\\mathrm {m p i}} - \\hat {\\mathbf {O}} _ {t - 1} ^ {\\mathrm {m p i}} \\| ^ {2} ] \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.721, + 0.905, + 0.74 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\theta} ^ {\\text {s i d e}} = \\mathbb {E} _ {\\mathbf {O} ^ {\\text {r e f}}, \\mathbf {O} ^ {\\text {m p i}}, \\mathbf {C} ^ {\\text {m p i}}, t, \\mathbf {O} ^ {\\text {s i d e}}, \\phi^ {\\text {s i d e}}} [ \\| \\mathbf {O} ^ {\\text {s i d e}} - \\hat {\\mathbf {O}} _ {t - 1} ^ {\\text {s i d e}} \\| ^ {2} ]. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.906, + 0.84 + ], + "angle": 0, + "content": "Intuitively, the first loss \\(\\mathcal{L}_{\\theta}^{\\mathrm{mpi}}\\) encourages the model to give good frontal views of the MPs, while the second loss \\(\\mathcal{L}_{\\theta}^{\\mathrm{side}}\\) encourages the model to render good side views from the MPs, which would be the only supervision for learning 3D shapes. However, such approach would require a large scale multiview video dataset, which is not generally available." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.906, + 0.902 + ], + "angle": 0, + "content": "In order to enable multi-view supervision when training purely on monocular 2D videos, a possible approach is to place MPs in a canonical space with a fixed head pose, then move the rendering camera to simulate head pose changes." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21110" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.092, + 0.881, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.311, + 0.907, + 0.382 + ], + "angle": 0, + "content": "Figure 3. Illustration of Reference-Target Alternating Training. On the top branch, we construct MPs in the target camera and render the frontal view, where we directly use the ground truth target image to compute loss Eq. (6) to learn sharp renderings and bootstrap the model. On the bottom branch, we construct MPs in the reference camera. We first rasterize the target control signal (expression) in the reference camera, then use the bootstrapped model to generate pseudo ground truth by rendering the frontal view. The pseudo ground truth is added with noise, serving as the model input to compute loss Eq. (7) for learning 3D shapes." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.385, + 0.484, + 0.566 + ], + "angle": 0, + "content": "In this way, the image data representing different head poses will act as a proxy ground truth for novel camera views to provide multi-view supervision. Therefore, we could separately compute loss \\(\\mathcal{L}_{\\theta}^{\\mathrm{mpi}}\\) Eq. (6) and loss \\(\\mathcal{L}_{\\theta}^{\\mathrm{side}}\\) Eq. (7) with only monocular 2D videos by using two different MPI setups as shown in Fig. 3. More specifically, when computing the first loss \\(\\mathcal{L}_{\\theta}^{\\mathrm{mpi}}\\), we would set the MPI frontal cameras \\(\\phi^{\\mathrm{mpi}}\\) as the target image camera and use the target image as the frontal view ground truth \\(\\mathbf{O}^{\\mathrm{mpi}}\\). Similarly, when computing the second loss \\(\\mathcal{L}_{\\theta}^{\\mathrm{side}}\\), we would set the MPI frontal cameras \\(\\phi^{\\mathrm{mpi}}\\) to the reference image camera and use the target image as the side view ground truth \\(\\mathbf{O}^{\\mathrm{side}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.57, + 0.484, + 0.752 + ], + "angle": 0, + "content": "However, the above strategy only satisfies the ground truth requirements of the two losses, but does not consider the model input. Indeed, when computing the second loss \\(\\mathcal{L}_{\\theta}^{\\mathrm{side}}\\) Eq. (7), where the MPI frontal cameras \\(\\phi^{\\mathrm{mpi}}\\) are set to the reference camera, the training video frames are used as the ground truth to supervise the side view of MPs. However these frames cannot be used as input of the diffusion model, as they are no longer pixel-aligned with the MPs, which are constructed using the reference camera in this setting. To tackle this problem, we propose a bootstrapping training strategy that allows our model to generate the pseudo ground truth under the reference camera by itself." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.76, + 0.484, + 0.852 + ], + "angle": 0, + "content": "Bootstrapping. In order to bootstrap our model for generating pseudo ground truth under the reference camera, we firstly train our model \\( g_{\\theta}(.) \\) with only the loss \\( \\mathcal{L}_{\\theta}^{\\mathrm{mpi}} \\) Eq. (6) (i.e. setting the MPI frontal cameras \\( \\phi^{\\mathrm{mpi}} \\) to the target camera). In this way, the model learns to generate plausible frontal views rendered from MPs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "With this bootstrapping, we can then enable the second loss \\(\\mathcal{L}_{\\theta}^{\\mathrm{side}}\\) Eq. (7) to learn 3D structures of the MPs. More specifically, we first set the MPI frontal cameras \\(\\phi^{\\mathrm{mpi}}\\) to the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.385, + 0.907, + 0.522 + ], + "angle": 0, + "content": "reference camera, then run the diffusion sampling with \\( k \\) steps to generate the output MPI video \\( \\mathbf{S}_k \\), and render the frontal view to get images \\( \\hat{\\mathbf{O}}_k^{\\mathrm{mpi}} \\). Since the model has been trained to give plausible frontal views, the rendered images \\( \\hat{\\mathbf{O}}_k^{\\mathrm{mpi}} \\) will have a reasonable quality, acting as the pseudo ground truth under the reference camera. Thus we can add noise to \\( \\hat{\\mathbf{O}}_k^{\\mathrm{mpi}} \\) to obtain the model input \\( \\mathbf{O}_t^{\\mathrm{mpi}} \\) when computing the second loss \\( \\mathcal{L}_{\\theta}^{\\mathrm{side}} \\) Eq. (7) to train our model for learning 3D shapes inside MPsIs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.529, + 0.909, + 0.681 + ], + "angle": 0, + "content": "Late-stage Noise Sampling. When training with the second loss \\(\\mathcal{L}_{\\theta}^{\\mathrm{side}}\\) Eq. (7), we compute the model input from the pseudo ground truth \\(\\hat{\\mathbf{O}}_k^{\\mathrm{mpi}}\\), which cannot perfectly match the real image distribution. Such distribution mismatch could harm the model learning and subsequently lead to blurry results. To mitigate this negative influence, when training with loss \\(\\mathcal{L}_{\\theta}^{\\mathrm{side}}\\) Eq. (7), we only sample \\(t\\) from the second half of the forward Markovian process \\([T/2, T]\\) so that the added noise is relatively large thus has a better chance to hide the distribution mismatch." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Loss Functions. During training, we compose the MPs and a background image calculated from the reference portrait \\(\\mathbf{O}^{\\mathrm{ref}}\\) with alpha blending to get the foreground renderings and the full image renderings. Using pre-computed foreground masks of the data videos, we calculate the standard L2 losses Eq. (6) and Eq. (7) on both the foreground renderings and the full image renderings. Following Tewari et al. [35], we additionally adopt several regularization losses besides the standard L2 losses Eq. (6) and Eq. (7). The regularization losses include: 1) A LPIPS perceptual loss \\(\\mathcal{L}_{\\mathrm{lpips}}\\) for foreground renderings; 2) A mask loss \\(\\mathcal{L}_{\\mathrm{mask}}\\) to supervise the rendered alpha mask; 3) An edge-aware depth smoothing loss and a disparity loss following Tucker et al. [36] to further supervise the underlying geometry," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21111" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.126, + 0.089, + 0.872, + 0.239 + ], + "angle": 0, + "content": "
MethodHDTFTalkinghead1khNovel View Render Speed
L1 ↓LPIPS ↓FID ↓FVD ↓L1 ↓LPIPS ↓FID ↓FVD ↓
Face-V2V [40]0.0290.13220.5778.110.0400.18630.52141.0222 FPS in 2562
EmoPortrait [16]0.0390.18924.77140.510.0580.25241.08280.227 FPS in 5122
Portrait4d-v2 [12]--25.74139.11--37.38227.3210 FPS in 5122
Follow-your-emoji [29]0.0350.11516.09117.540.0470.15919.37209.910.5FPS* in 5122
X-Portrait [42]0.0340.11915.66132.900.0450.16419.86183.050.1FPS* in 5122
Ours0.0310.11815.9369.040.0450.18027.83135.12109 FPS in 5122
Ours w/o residual0.0300.11614.7665.750.0440.17428.19127.11109 FPS in 5122
Ours w/o late-stage0.0310.12316.7070.690.0450.18029.86136.07109 FPS in 5122
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.24, + 0.907, + 0.311 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison and ablation study. We compare our method with previous work on both HDTF and Talkinghead1kh datasets. Our method achieves comparable image quality (L1, LPIPS, FID) and the best video quality (FVD) and novel-view rendering efficiency overall. Compared to ablation cases, the averaged metrics are close, and we highlight the actual huge gaps in perceived image, geometry, and side view rendering quality in Fig. 7 and Fig. 6. All the render speeds are tested on a single Nvidia A100 GPU. *We report the generation speeds of diffusion models for Follow-your-emoji [29] and X-Portrait [42], as they do not support explicit camera control." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.314, + 0.482, + 0.344 + ], + "angle": 0, + "content": "where we use the disparity maps rendered from the parametric model as pesudo ground truth for disparity loss." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.352, + 0.279, + 0.367 + ], + "angle": 0, + "content": "3.4. Run-time Inference" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.374, + 0.483, + 0.51 + ], + "angle": 0, + "content": "When generating a 3D talking head from a driving video, we set the MPI frontal cameras \\(\\phi^{\\mathrm{MPI}}\\) to the driving video. In order to generate long video sequence and smooth the discontinuity between adjacent video clips, we input ground truth first frame to the model in a similar way as Lumiere [5] with random drop during training. During inference, we take the last frame from the previous clip as the first frame of next clip. The generated first frame is then dropped due to overlapping when concatenating two adjacent clips." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.522, + 0.223, + 0.538 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.547, + 0.483, + 0.607 + ], + "angle": 0, + "content": "In this section, we evaluate our method by comparing with prior arts on talking heads image quality, temporal quality, multi-view consistency, and computational efficiency. We also conduct qualitative comparisons and ablation studies." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.611, + 0.483, + 0.852 + ], + "angle": 0, + "content": "Implementation Details. We train our model on a self-collected dataset consisting of 35k in-the-wild talking head monocular videos. Each video has a length between 4 to 6 seconds, and we crop the face region in \\(512 \\times 512\\) for training. Our model generates MPI video with 32 frames in a single network forwarding, and each MPI consists of 24 planes. We dynamically set the near and far planes based on the distance between the 3DMM head joint and the camera position. We optimize our model end-to-end on 8 NVIDIA H100 GPUs with a batch size of 8. During inference, we use DDIM sampling and do classifier-free guidance on first-frame and reference portrait with two scales [8]. The guidance scale for the reference portrait is set to 1.5, while the guidance scale for the first frame is linearly reduced from 1 to 0 between frames 1 and 16. For the remaining 16 frames, the guidance scale for the first frame is fixed at 0." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Benchmark. We perform our qualitative comparison on two widely used public talking-head datasets, Talking-head1kh [40] and HDTF [47] datasets. For HDTF dataset," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.314, + 0.907, + 0.42 + ], + "angle": 0, + "content": "we randomly sample video clips from 100 identities and each video clip consists of 180 frames. For Talkinghead1kh dataset, we randomly sample two 180-frame video clip from all evaluation identities. During testing, the first frame of each video clip is chosen as reference portrait \\(\\mathbf{O}^{\\mathrm{src}}\\) and all the animations are performed in a self-reenactment way. Cross-identity driving results are provided in the supp." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.424, + 0.909, + 0.711 + ], + "angle": 0, + "content": "Baselines. We compare our method with existing one-shot photoreal talking heads approaches trained from monocular videos in the wild. Specifically, we compare with representatives for three families of approaches: 1) Feed-forward models (e.g. Face-V2V [40] and EmoPortrait [16]). These methods do not use diffusion model but directly build 3D implicit feature volumes (as appearance) driven by warping fields (as expression). These methods are often optimized for the controllability in 2D space, but not the generation quality or 3D consistency. 2) 3D NeRF-based one-shot talking heads (e.g. Portrait4D-v2 [11]), which establish the standard for 3D consistency due to the use of native 3D representation. 3) Video diffusion based methods (e.g. follow-your-emoji [29] and xportrait [42]), which tend to be particularly strong on the quality of the generated images or videos. We perform all the evaluations on \\(512 \\times 512\\) resolution since most of the baselines outputs results at it, except Face-V2V for which we upsample its outputs to the desired resolution." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.726, + 0.798, + 0.741 + ], + "angle": 0, + "content": "4.1. Appearance and Motion Quality" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.749, + 0.906, + 0.825 + ], + "angle": 0, + "content": "Following prior works, we compare image quality with L1, LPIPS [46] and FID score [22] and video quality, a combination of images and motions, using FVD score [37]. Note that L1 and LPIPS are not available for Portrait4D-v2 due to misalignments between its outputs and ground truth." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Quantitative results are shown in Tab. 1. While the ranking based on different metrics varies, our method remains competitive among the state-of-the-art. Particularly, our model consistently achieves the best FVD on both datasets indicating our overall best video quality, taking into con" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21112" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.137, + 0.089, + 0.157, + 0.098 + ], + "angle": 0, + "content": "Ref." + }, + { + "type": "image_caption", + "bbox": [ + 0.235, + 0.089, + 0.284, + 0.098 + ], + "angle": 0, + "content": "Face-v2v" + }, + { + "type": "image_caption", + "bbox": [ + 0.327, + 0.089, + 0.391, + 0.098 + ], + "angle": 0, + "content": "EmoPortrait" + }, + { + "type": "image_caption", + "bbox": [ + 0.421, + 0.089, + 0.493, + 0.098 + ], + "angle": 0, + "content": "Portrait4D-v2" + }, + { + "type": "image_caption", + "bbox": [ + 0.527, + 0.089, + 0.582, + 0.098 + ], + "angle": 0, + "content": "X-Portrait" + }, + { + "type": "image_caption", + "bbox": [ + 0.607, + 0.089, + 0.703, + 0.099 + ], + "angle": 0, + "content": "Follow-your-emoji" + }, + { + "type": "image_caption", + "bbox": [ + 0.739, + 0.089, + 0.766, + 0.098 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.842, + 0.089, + 0.862, + 0.098 + ], + "angle": 0, + "content": "G.T." + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.104, + 0.197, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.183, + 0.197, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.21, + 0.104, + 0.308, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.21, + 0.183, + 0.308, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.31, + 0.104, + 0.408, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.31, + 0.183, + 0.408, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.409, + 0.104, + 0.506, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.183, + 0.506, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.104, + 0.604, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.183, + 0.604, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.606, + 0.104, + 0.703, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.606, + 0.183, + 0.704, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.704, + 0.104, + 0.802, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.704, + 0.183, + 0.802, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.803, + 0.104, + 0.902, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.803, + 0.183, + 0.902, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.262, + 0.197, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.34, + 0.197, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.21, + 0.261, + 0.308, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.21, + 0.34, + 0.308, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.31, + 0.261, + 0.408, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.31, + 0.34, + 0.408, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.41, + 0.261, + 0.506, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.41, + 0.34, + 0.506, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.261, + 0.604, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.34, + 0.604, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.606, + 0.261, + 0.704, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.606, + 0.34, + 0.704, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.704, + 0.261, + 0.802, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.704, + 0.34, + 0.802, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.803, + 0.261, + 0.902, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.803, + 0.34, + 0.902, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.417, + 0.907, + 0.459 + ], + "angle": 0, + "content": "Figure 4. Talking head results. We show results from previous work and our method. Our method generates results with sharp appearance and closely aligned expression with the ground truth. Note Portrait4d-v2 uses a customized camera space which is non-trivial to render GT camera aligned images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.464, + 0.483, + 0.525 + ], + "angle": 0, + "content": "sideration both image quality and motion naturalness. This also indicates that a native video diffusion model is stronger than a finetuned image diffusion with added temporal component (like X-Portrait [42] and Follow-your-emoji [29])." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.525, + 0.483, + 0.691 + ], + "angle": 0, + "content": "We also show qualitative results in Fig. 4. FaceV2V [40] gives stable results and follows the input identity, but produces less details due to its low output resolution. EmoPortrait [16] gives sharper results, but suffers from larger identity shifts. Portrait4D-v2 [12] produces good sharpness and less identity shifts, however relatively stiff expressions. The diffusion-based methods Follow-your-emoji [29] and X-portrait [42] render good high-frequency details, but do not precisely follow the input head poses or expressions. In contrast, our method gives good image quality, faithful identity, while follows the input control signal." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.698, + 0.367, + 0.715 + ], + "angle": 0, + "content": "4.2. Effectiveness for 3D Rendering" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.81 + ], + "angle": 0, + "content": "We evaluate the effectiveness on 3D of different methods from the following aspects. First, we consider the task of binocular stereo rendering, which is one of the foundational use cases for VR/AR applications. Also, we compare how well these approaches perform on perspective geometry by moving the camera closer." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Note that Face-V2V [40] and EmoPortrait [16] use orthogonal projection as a simplified camera model, thus they cannot produce any reasonable renderings for stereo pairs (i.e. two views will have the same disparity across the whole image) and moving camera closer (i.e. getting identical renderings). Follow-your-emoji [29] and X-xportrait [42] only" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.464, + 0.907, + 0.54 + ], + "angle": 0, + "content": "use detected landmarks and images as the driving signal but do not allow any explicit control on the camera, thus they cannot perform these two 3D tasks neither. As a result, we only show the renderings from Portrait4D-v2 [12] and ours, where truth 3D rendering is supported." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.54, + 0.907, + 0.66 + ], + "angle": 0, + "content": "Binocular Stereo Rendering. We render stereo pairs with a 5cm baseline looking from 1m in front of the talking head, and run an off-the-shell stereo matching method [27] to get the disparity map. In Fig. 5, ours gives stereo pairs with reasonable estimated disparities, and is competitive with the NeRF-based Portrait4D-v2 [12], which acts as a strong baseline on 3D effects, while being one order of magnitude faster in novel view rendering (Tab. 1)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.661, + 0.906, + 0.766 + ], + "angle": 0, + "content": "Perspective Geometry. To demonstrate the perspective effect, we render a close view by moving the camera closer while reducing camera focal length to keep the whole head in view. As in Fig. 5, ours produces good perspective distortions that is comparable to the NeRF-based Portrait4Dv2 [12], while being more faithful to the reference identity shown in Fig. 4 second row." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.774, + 0.618, + 0.788 + ], + "angle": 0, + "content": "4.3. Ablation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.901 + ], + "angle": 0, + "content": "MPI as representation. Our diffusion model generates MPI frames, where each of them has a frontal, a residual, and a set of planes with alpha. While the frontal contains most of the appearance, we hereby evaluate the importance of the residual image. We train our model producing a MPI with only a frontal image, and reuse it for all the planes for alpha composition. While the metrics on frontal view" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21113" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.166, + 0.09, + 0.228, + 0.1 + ], + "angle": 0, + "content": "Portrait4d" + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.101, + 0.196, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.125, + 0.174, + 0.171, + 0.185 + ], + "angle": 0, + "content": "original" + }, + { + "type": "image", + "bbox": [ + 0.197, + 0.101, + 0.294, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.226, + 0.174, + 0.265, + 0.184 + ], + "angle": 0, + "content": "stereo" + }, + { + "type": "image_caption", + "bbox": [ + 0.384, + 0.091, + 0.411, + 0.1 + ], + "angle": 0, + "content": "ours" + }, + { + "type": "image", + "bbox": [ + 0.3, + 0.101, + 0.396, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.324, + 0.174, + 0.371, + 0.184 + ], + "angle": 0, + "content": "original" + }, + { + "type": "image", + "bbox": [ + 0.399, + 0.101, + 0.497, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.426, + 0.174, + 0.464, + 0.184 + ], + "angle": 0, + "content": "stereo" + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.187, + 0.196, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.262, + 0.18, + 0.272 + ], + "angle": 0, + "content": "close view" + }, + { + "type": "image", + "bbox": [ + 0.197, + 0.187, + 0.294, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.219, + 0.262, + 0.273, + 0.273 + ], + "angle": 0, + "content": "disparity" + }, + { + "type": "image", + "bbox": [ + 0.3, + 0.187, + 0.397, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.316, + 0.262, + 0.376, + 0.272 + ], + "angle": 0, + "content": "close view" + }, + { + "type": "image", + "bbox": [ + 0.399, + 0.187, + 0.497, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.418, + 0.262, + 0.473, + 0.273 + ], + "angle": 0, + "content": "disparity" + }, + { + "type": "image_caption", + "bbox": [ + 0.567, + 0.09, + 0.629, + 0.1 + ], + "angle": 0, + "content": "Portrait4d" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.101, + 0.601, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.174, + 0.573, + 0.184 + ], + "angle": 0, + "content": "original" + }, + { + "type": "image", + "bbox": [ + 0.602, + 0.101, + 0.699, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.628, + 0.174, + 0.666, + 0.184 + ], + "angle": 0, + "content": "stereo" + }, + { + "type": "image_caption", + "bbox": [ + 0.785, + 0.091, + 0.813, + 0.1 + ], + "angle": 0, + "content": "ours" + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.101, + 0.803, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.726, + 0.174, + 0.772, + 0.185 + ], + "angle": 0, + "content": "original" + }, + { + "type": "image", + "bbox": [ + 0.806, + 0.101, + 0.899, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.828, + 0.174, + 0.866, + 0.184 + ], + "angle": 0, + "content": "stereo" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.187, + 0.602, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.523, + 0.262, + 0.585, + 0.272 + ], + "angle": 0, + "content": "close view" + }, + { + "type": "image", + "bbox": [ + 0.603, + 0.187, + 0.699, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.621, + 0.262, + 0.675, + 0.273 + ], + "angle": 0, + "content": "disparity" + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.187, + 0.803, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.722, + 0.262, + 0.783, + 0.272 + ], + "angle": 0, + "content": "close view" + }, + { + "type": "image", + "bbox": [ + 0.805, + 0.187, + 0.899, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.821, + 0.262, + 0.874, + 0.273 + ], + "angle": 0, + "content": "disparity" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.275, + 0.907, + 0.331 + ], + "angle": 0, + "content": "Figure 5. Binocular stereo and perspective effects. Our method generates MPI videos that enables efficient spatial rendering, e.g. binocular stereo and perspective effects. We render stereo pairs and calculate a disparity using RAFT-Stereo [27] to visualize the perceived geometry. To demonstrate perspective effects, we render a view with camera moving closer to the subject. Our method renders visually plausible and comparable 3D effects with NeRF based approach like Portrait4d-v2, while being orders of magnitude faster." + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.336, + 0.479, + 0.556 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.561, + 0.483, + 0.645 + ], + "angle": 0, + "content": "Figure 6. Effect of residual RGB. We retrain our model with only a single frontal image in the MPI. We show both the rendered side view images and the disparity converted from the MPI alpha in frontal view. We found this significantly reduces the capacity of MPI, which leads to noisy and irregular geometry and blatant blurriness near the occlusion boundaries." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.656, + 0.483, + 0.732 + ], + "angle": 0, + "content": "become even stronger compared to the full model (1 (Ours w/o residual)), the model produces much worse geometry (Fig. 6), presumable due to the missing of capacity of handling view dependent effects. It also leads to obvious blurriness in the boundary when rendering from the side view." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.733, + 0.483, + 0.778 + ], + "angle": 0, + "content": "A further degenerated case is to use a RGB image and depth for novel-view rendering, and please check our supp for more comparison." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Late-stage noise sampling. Late-stage noise sampling is also important to learn sharp images since that significantly reduces the negative impact from the imperfect pseudo ground truth. To verify this, we train a model with full noise sampling, and show the quantitative and qualitative comparison in Tab. 1 and Fig. 7. Without large noise sampling, the diffusion model still produces reasonable structure but missing details, e.g. wrinkles, mold, hair strand." + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.336, + 0.903, + 0.52 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.525, + 0.907, + 0.581 + ], + "angle": 0, + "content": "Figure 7. Effect of Late-stage noise sampling. We show a comparison between our model trained with and without the Late-stage noise sampling. Training with regular noise sampling leads to excessive blurriness and lost of details." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.598, + 0.633, + 0.613 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.907, + 0.902 + ], + "angle": 0, + "content": "We present a method to learn a 3D-aware video diffusion model, which generates photorealistic talking head multi-plane images given an identity reference image and a sequence of tracking signals. We show such a model can be learned directly from large scale monocular videos without multi-view datasets. With appropriate learning strategies, both 3D information and high quality image generation can be learned concurrently. Extensive experiments show that, our method generates not only talking head video with quality comparable and even better than existing arts but also 3D rendering efficiently that support spatial viewing experience, like in VR headset. As limitations, the generated talking head video does not support rendering from excessive camera viewpoint changes, which is a common issue of the MPI representation. Extending our framework with other efficient 3D representations or enhancing training with multi-view dataset could be a promising direction to improve both the image and 3D quality." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21114" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.198 + ], + "angle": 0, + "content": "[1] Titas Anciukevicius, Zexiang Xu, Matthew Fisher, Paul Henderson, Hakan Bilen, Niloy J Mitra, and Paul Guerrero. Renderdiffusion: Image diffusion for 3d reconstruction, inpainting and generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12608-12618, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.2, + 0.484, + 0.268 + ], + "angle": 0, + "content": "[2] ShahRukh Athar, Zexiang Xu, Kalyan Sunkavalli, Eli Shechtman, and Zhixin Shu. Rignerf: Fully controllable neural 3d portraits. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 20364-20373, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.27, + 0.484, + 0.366 + ], + "angle": 0, + "content": "[3] Ziqian Bai, Feitong Tan, Zeng Huang, Kripasindhu Sarkar, Danhang Tang, Di Qiu, Abhimitra Meka, Ruofei Du, Mingsong Dou, Sergio Orts-Escalano, et al. Learning personalized high quality volumetric head avatars from monocular rgb videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16890-16900, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.368, + 0.483, + 0.45 + ], + "angle": 0, + "content": "[4] Ziqian Bai, Feitong Tan, Sean Fanello, Rohit Pandey, Mingsong Dou, Shichen Liu, Ping Tan, and Yinda Zhang. Efficient 3d implicit head avatar with mesh-anchored hash table blendshapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1975–1984, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.452, + 0.483, + 0.521 + ], + "angle": 0, + "content": "[5] Omer Bar-Tal, Hila Chefer, Omer Tov, Charles Herrmann, Roni Paiss, Shiran Zada, Ariel Ephrat, Junhwa Hur, Guanghui Liu, Amit Raj, et al. Lumiere: A spacetime diffusion model for video generation. arXiv preprint arXiv:2401.12945, 2024. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.523, + 0.483, + 0.564 + ], + "angle": 0, + "content": "[6] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 157-164. 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.565, + 0.483, + 0.634 + ], + "angle": 0, + "content": "[7] Stella Bounareli, Christos Tzelepis, Vasileios Argyriou, Ioannis Patras, and Georgios Tzimiropoulos. Hyperreenact: One-shot reenactment via jointly learning to refine and retarget faces. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.635, + 0.483, + 0.703 + ], + "angle": 0, + "content": "[8] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.705, + 0.483, + 0.788 + ], + "angle": 0, + "content": "[9] Bindita Chaudhuri, Noranart Vesdapunt, Linda Shapiro, and Baoyuan Wang. Personalized face modeling for improved face reconstruction and motion retargeting. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part V 16, pages 142-160. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.79, + 0.483, + 0.856 + ], + "angle": 0, + "content": "[10] Yufan Chen, Lizhen Wang, Qijing Li, Hongjiang Xiao, Shengping Zhang, Hongxun Yao, and Yebin Liu. Monogaussianavatar: Monocular gaussian point-based head avatar. In ACM SIGGRAPH 2024 Conference Papers, pages 1-9, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[11] Yu Deng, Duomin Wang, and Baoyuan Wang. Portrait4d-v2: Pseudo multi-view data creates better 4d head synthesizer. arXiv preprint arXiv:2403.13570, 2024. 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[12] Yu Deng, Duomin Wang, and baoyuan Wang. Portrait4d-v2: Pseudo multi-view data creates better 4d head synthesizer. arXiv, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[13] Zheng Ding, Xuaner Zhang, Zhihao Xia, Lars Jebe, Zhuowen Tu, and Xiuming Zhang. Diffusionrig: Learning personalized priors for facial appearance editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12736-12746, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.205, + 0.905, + 0.26 + ], + "angle": 0, + "content": "[14] Michail Christos Doukas, Stefanos Zafeiriou, and Viktoriia Sharmanska. Headgan: One-shot neural head synthesis and editing. In Proceedings of the IEEE/CVF International conference on Computer Vision, pages 14398-14407, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.261, + 0.905, + 0.329 + ], + "angle": 0, + "content": "[15] Nikita Drobyshev, Jenya Chelishev, Taras Khakhulin, Aleksei Ivakhnenko, Victor Lempitsky, and Egor Zakharov. Megaportraits: One-shot megapixel neural head avatars. In Proceedings of the 30th ACM International Conference on Multimedia, pages 2663–2671, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.905, + 0.413 + ], + "angle": 0, + "content": "[16] Nikita Drobyshev, Antoni Bigata Casademunt, Konstantinos Vougioukas, Zoe Landgraf, Stavros Petridis, and Maja Pantic. Emoportraits: Emotion-enhanced multimodal one-shot head avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8498-8507, 2024. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.905, + 0.468 + ], + "angle": 0, + "content": "[17] Emilien Dupont, Hyunjik Kim, SM Eslami, Danilo Rezende, and Dan Rosenbaum. From data to functa: Your data point is a function and you can treat it like one. arXiv preprint arXiv:2201.12204, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.469, + 0.905, + 0.538 + ], + "angle": 0, + "content": "[18] Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8649-8658, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.905, + 0.594 + ], + "angle": 0, + "content": "[19] Xuan Gao, Chenglai Zhong, Jun Xiang, Yang Hong, Yudong Guo, and Juyong Zhang. Reconstructing personalized semantic facial nerf models from monocular video. ACM Transactions on Graphics (TOG), 41(6):1-12, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.595, + 0.907, + 0.663 + ], + "angle": 0, + "content": "[20] Philip-William Grassal, Malte Prinzler, Titus Leistner, Carsten Rother, Matthias Nießner, and Justus Thies. Neural head avatars from monocular rgb videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18653–18664, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.664, + 0.905, + 0.72 + ], + "angle": 0, + "content": "[21] Jianzhu Guo, Dingyun Zhang, Xiaogiang Liu, Zhizhou Zhong, Yuan Zhang, Pengfei Wan, and Di Zhang. Liveportrait: Efficient portrait animation with stitching and retargeting control. arXiv preprint arXiv:2407.03168, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.721, + 0.905, + 0.788 + ], + "angle": 0, + "content": "[22] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[23] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[24] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[25] Yushi Lan, Fangzhou Hong, Shuai Yang, Shangchen Zhou, Xuyi Meng, Bo Dai, Xingang Pan, and Chen Change Loy." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21115" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "Ln3diff: Scalable latent neural fields diffusion for speedy 3d generation. In European Conference on Computer Vision, pages 112-130. Springer, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.135, + 0.483, + 0.176 + ], + "angle": 0, + "content": "[26] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph., 36(6):194-1, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[27] Lahav Lipson, Zachary Teed, and Jia Deng. Raft-stereo: Multilevel recurrent field transforms for stereo matching. In 2021 International Conference on 3D Vision (3DV), pages 218-227. IEEE, 2021. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.482, + 0.313 + ], + "angle": 0, + "content": "[28] Tao Liu, Chenpeng Du, Shuai Fan, Feilong Chen, and Kai Yu. Diffhub: Person-generic visual dubbing using inpainting renderer with diffusion auto-encoder. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 3630-3634. IEEE, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.482, + 0.384 + ], + "angle": 0, + "content": "[29] Yue Ma, Hongyu Liu, Hongfa Wang, Heng Pan, Yingqing He, Junkun Yuan, Ailing Zeng, Chengfei Cai, Heung-Yeung Shum, Wei Liu, et al. Follow-your-emoji: Fine-controllable and expressive freestyle portrait animation. arXiv preprint arXiv:2406.01900, 2024. 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.385, + 0.482, + 0.453 + ], + "angle": 0, + "content": "[30] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.455, + 0.482, + 0.537 + ], + "angle": 0, + "content": "[31] Norman Müller, Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Bulo, Peter Kontschieder, and Matthias Nießner. Diffrf: Rendering-guided 3d radiance field diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4328-4338, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.482, + 0.62 + ], + "angle": 0, + "content": "[32] Shenhan Qian, Tobias Kirschstein, Liam Schoneveld, Davide Davoli, Simon Giebenhain, and Matthias Nießner. Gaussianavatars: Photorealistic head avatars with rigged 3d gaus-sians. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20299-20309, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.622, + 0.482, + 0.678 + ], + "angle": 0, + "content": "[33] Shunsuke Saito, Gabriel Schwartz, Tomas Simon, Junxuan Li, and Giljoo Nam. Relightable gaussian codec avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 130-141, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.482, + 0.734 + ], + "angle": 0, + "content": "[34] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. Advances in neural information processing systems, 32, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.482, + 0.816 + ], + "angle": 0, + "content": "[35] Ayush Tewari, Tianwei Yin, George Cazenavette, Semon Rezhikov, Josh Tenenbaum, Frédo Durand, Bill Freeman, and Vincent Sitzmann. Diffusion with forward models: Solving stochastic inverse problems without direct supervision. Advances in Neural Information Processing Systems, 36: 12349-12362, 2023. 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.873 + ], + "angle": 0, + "content": "[36] Richard Tucker and Noah Snavely. Single-view view synthesis with multiplane images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 551-560, 2020. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[37] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. To" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "wards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[38] Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, Karsten Kreis, et al. Lion: Latent point diffusion models for 3d shape generation. Advances in Neural Information Processing Systems, 35:10021-10039, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[39] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, et al. Rodin: A generative model for sculpting 3d digital avatars using diffusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4563-4573, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.331 + ], + "angle": 0, + "content": "[40] Ting-Chun Wang, Arun Mallya, and Ming-Yu Liu. One-shot free-view neural talking-head synthesis for video conferencing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10039-10049, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.333, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[41] Zhen Wang, Qiangeng Xu, Feitong Tan, Mengei Chai, Shichen Liu, Rohit Pandey, Sean Fanello, Achuta Kadambi, and Yinda Zhang. Mvdd: Multi-view depth diffusion models. In European Conference on Computer Vision, pages 236-253. Springer, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.404, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[42] You Xie, Hongyi Xu, Guoxian Song, Chao Wang, Yichun Shi, and Linjie Luo. X-portrait: Expressive portrait animation with hierarchical motion attention. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.461, + 0.905, + 0.543 + ], + "angle": 0, + "content": "[43] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. Styleheat: One-shot high-resolution editable talking face generation via pre-trained stylegan. In European conference on computer vision, pages 85-101. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.545, + 0.905, + 0.626 + ], + "angle": 0, + "content": "[44] Egor Zakharov, Aleksei Ivakhnenko, Aliaksandra Shysheya, and Victor Lempitsky. Fast bi-layer neural synthesis of one-shot realistic head avatars. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XII 16, pages 524-540. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.629, + 0.905, + 0.684 + ], + "angle": 0, + "content": "[45] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.686, + 0.905, + 0.755 + ], + "angle": 0, + "content": "[46] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.757, + 0.905, + 0.825 + ], + "angle": 0, + "content": "[47] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661–3670, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.827, + 0.905, + 0.881 + ], + "angle": 0, + "content": "[48] Wojciech Zielonka, Timo Bolkart, and Justus Thies. Instant volumetric head avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4574-4584, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.881 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21116" + } + ] +] \ No newline at end of file diff --git a/2025/IM-Portrait_ Learning 3D-aware Video Diffusion for Photorealistic Talking Heads from Monocular VideosC/d89a7508-cd12-421c-9424-93ad28adceb7_origin.pdf b/2025/IM-Portrait_ Learning 3D-aware Video Diffusion for Photorealistic Talking Heads from Monocular VideosC/d89a7508-cd12-421c-9424-93ad28adceb7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..94846cf02a8cb4a741043d701ae6bf7b46934cd8 --- /dev/null +++ b/2025/IM-Portrait_ Learning 3D-aware Video Diffusion for Photorealistic Talking Heads from Monocular VideosC/d89a7508-cd12-421c-9424-93ad28adceb7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:500e70f850cc2b68ba906fc72e28b176fe98137ffc3eed31955d59160ccff93e +size 3498615 diff --git a/2025/IM-Portrait_ Learning 3D-aware Video Diffusion for Photorealistic Talking Heads from Monocular VideosC/full.md b/2025/IM-Portrait_ Learning 3D-aware Video Diffusion for Photorealistic Talking Heads from Monocular VideosC/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3fc06c8b9b1e4df8a1b5cb6d019b494a3e31d0fb --- /dev/null +++ b/2025/IM-Portrait_ Learning 3D-aware Video Diffusion for Photorealistic Talking Heads from Monocular VideosC/full.md @@ -0,0 +1,407 @@ +# IM-Portrait: Learning 3D-aware Video Diffusion for Photorealistic Talking Heads from Monocular Videos + +Yuan Li $^{1,2*§}$ Ziqian Bai $^{2*}$ Feitong Tan $^{2}$ Zhaopeng Cui $^{1\dagger}$ Sean Fanello $^{2}$ Yinda Zhang $^{2\dagger}$ $^{1}$ State Key Lab of CAD&CG, Zhejiang University $^{2}$ Google + +![](images/ec9265c354faa8646c8f3e3609394ad1ca6cea16252a845c6c2c582dff18c73d.jpg) +Inputs + +![](images/3fd86c5b7b4dce66736362fbb47662b14ff889b097dcad8e5c5c104b04acc529.jpg) +Generated Results + +![](images/73b69f0a3b8f6b1e1ad6d356ed210561ba0f665808b37483e3b9a9f668211473.jpg) + +![](images/de5f90519439ddd28220066c889d6301b4846905698e0679a5142c716876dead.jpg) +Figure 1. We propose a 3D-aware video diffusion model for talking head synthesis. Given an image as identity and a sequence of tracking signals (as shown on the left for each example), our model directly generates videos in Multiplane Images (MPIs) in a single denoising process, which is ready for efficient novel-view rendering. This enables immersive viewing experience, e.g. rendering binocular stereo or perspective distortion in VR headset. + +![](images/39e679271344c55ad39128792cff5a576f415bda6d20f1f565510c27728a9724.jpg) + +![](images/0f2026fbf02b6b9aa72d77aca96cf9106890065906adf574188bbaf19fc45f1e.jpg) + +![](images/34992577c8be39c707c626f21a9f0b9075dd446c73a5213cc2a4c8c5ba9f0fe4.jpg) +Inputs + +![](images/1ae3402671efe9adefba86ad704f97f2dd666fc44d01090e3cfde99218ecca1b.jpg) +Generated Results + +![](images/7bf921429c92fbd024546c9c194049301803dacf37b21fdcf8d517b8a99778f9.jpg) + +![](images/9990816515274e38c39a4ae9c4269d30110b18c5ac120a7ad0e1446885558faf.jpg) + +![](images/6fe0a1cc798d76d9ccc3c5f90355f6d00aeed7dda8d6af7a0075322e5a5f1bff.jpg) + +![](images/22e0927c20c3ff7ceec2a2c2b0b571116889d2f7b4c05f9f899c44590fb2f781.jpg) + +# Abstract + +We propose a novel 3D-aware diffusion-based method for generating photorealistic talking head videos directly from a single identity image and explicit control signals (e.g., expressions). Our method generates Multiplane Images (MPIs) that ensure geometric consistency, making them ideal for immersive viewing experiences like binocular videos for VR headsets. Unlike existing methods that often require a separate stage or joint optimization to reconstruct a 3D representation (such as NeRF or 3D Gaussians), our approach directly generates the final output through a single denoising process, eliminating the need for post + +processing steps to render novel views efficiently. To effectively learn from monocular videos, we introduce a training mechanism that reconstructs the output MPI randomly in either the target or the reference camera space. This approach enables the model to simultaneously learn sharp image details and underlying 3D information. Extensive experiments demonstrate the effectiveness of our method, which achieves competitive avatar quality and novel-view rendering capabilities, even without explicit 3D reconstruction or high-quality multi-view training data. + +# 1. Introduction + +Synthesizing realistic talking heads, has been a long-lasting challenging problem in computer vision for decades. The recent emergence of diffusion models within the generative + +AI domain offers a promising solution to this problem. Significant advancements have been made in modeling talking heads using image or video diffusion models conditioned on tracking signals [13, 28, 29, 42], leveraging the remarkable generative capabilities of diffusion models. + +Most diffusion-based approaches, however, operate in 2D image space, and the extension of these methods to 3D rendering remains largely unexplored. One potential approach involves applying 3D transformations to the tracking signal, such as parametric models or landmarks. However, this necessitates running the reverse diffusion process for each new camera view, leading to significant computational overhead. Alternatively, some diffusion-based approaches generate multi-view images, which often lack geometric consistency. While off-the-shelf methods like NeRF [12] or 3D Gaussians can be used to construct a 3D representation for efficient rendering, the inherent inconsistencies in the input multi-view images limit the quality of the final 3D reconstruction, leading to potential issues in multi-view and temporal consistency. + +In this work, we present, for the first time, a diffusion based approach that produces 3D talking heads that can be rendered efficiently in varying camera poses. We adopt Multiplane Image (MPI) as the 3D representation due to its remarkable rendering efficiency and quality, and the affinity to 2D images such that we can make good use of existing 2D diffusion models. Essentially, we formulate the problem of synthesizing talking heads images as a denoising diffusion process, which is conditioned on a reference image (i.e. the subject identity) and a sequence of control signals from parametric models (i.e. expressions and head poses). In a single denoising process, our model generates MPI videos that can be readily rendered in varying camera pose and support spatial viewing experience. + +Our model is learned directly from a large scale monocular talking head videos in-the-wild. Due to the missing of multi-view datasets, like previous works, the 3D consistency needs to be learned from the head pose variability between video frames. To achieve this, it is required to construct MPs in a camera that looks at the head from a different viewpoint than the target images. Specifically, we use the viewpoint of the reference image to construct MPs. The estimated target images can then be rendered from MPs by moving the camera according to the relative head pose differences between the reference and target images. In this way, however, the frontal views of MPs will be misaligned with the input noisy images (in our case the target ground-truth images), which will break the training of the diffusion model. To tackle the problem, we propose a novel training framework to bootstrap our model to synthesize pseudo noisy input images that align with the MPs in reference cameras. With this, the MPs are randomly generated from either the reference camera viewpoint or the + +target camera viewpoint during the training, which allows the diffusion model to learn both 3D information and fine details. + +Our contributions are as follows. + +- We propose a diffusion-based approach for 3D-aware talking heads generation. +- Our diffusion model directly generates 3D renderable talking heads in a single denoising process, eliminating the need for post-processing steps. +- We propose a novel training framework that enables our model to learn from monocular videos, and show how randomly selecting the camera viewpoint between reference and target images allows for better convergence and sharper results. +- Extensive experiments demonstrate that our method achieves comparable or superior image and video quality to existing diffusion-based 2D talking head methods, while also offering efficient and consistent multiview rendering on par with specialized (i.e. non-diffusion based) 3D approaches. + +# 2. Related Works + +2D Talking Heads. In recent year, there has been a great advancement on synthesizing 2D talking heads with neural networks. A recent line of works [7, 14-16, 34, 40, 43, 44] propose to leverage neural networks to encode the appearance and motion from input images [15, 16, 34, 40] (or explicit control signals [14, 43, 44]) to some latent features or manually defined representations (e.g. key-points, warping fields, texture maps), which are injected into 2D CNN generators with various architectures for talking head synthesis. More recently, many works [13, 28, 29, 42] started to adopt 2D diffusion models, leveraging their impressive image quality, for controllable talking head synthesis. However, these approaches usually use simplified camera models (e.g. orthogonal projection) or they do not have any explicit camera control, making them unsuitable for spatial viewing experiences such as binocular stereo and perspective geometry. In addition, these methods usually do not produce explicit 3D geometry, which also limits their downstream applications. While most of prior work learns talking head models from monocular videos which are available in a large scale in-the-wild, multi-view datasets, e.g. from lightstage [21], could be leveraged to further improve fidelity, however such data requires intensive collection labor and there are no large scale multi-view datasets that are publicly available. + +3D Talking Heads. The problem of generating 3D talking heads has been a hot topic in the community. People attempted to model their talking heads with various 3D representations, such as textured mesh [9, 20], Neural Radiance Fields (NeRFs) [30] or other implicit func + +tions [2-4, 11, 12, 18, 19, 48], and, more recently, Gaussian Splatting [10, 32, 33]. While most of these approaches are person-specific, which requires to retrain a new model for every new person, some methods [11, 12] also explore the potential of person-agnostic models with explicit 3D representations. Despite having good 3D properties, these approaches cannot easily leverage the recent advancement on 2D diffusion, which lower the upper bound of their image quality. + +Diffusion-based 3D Generative Model. Another line of work, investigates how to apply diffusion models on 3D representations. The most straightforward way is applying the standard diffusion process directly on 3D data [38, 41]. However, this requires large scale 3D datasets, which are generally not available. One typical workaround is to train the diffusion model with 2D images. Some works [17, 24, 25, 31, 39] firstly learn a latent space of the 3D samples with multiview 2D images, then apply diffusion on the learned latent space. However, this results in an additional latent space learning, which is non-trivial thus increasing the pipeline complexity. Some other works [1, 35] directly model 3D scene distributions with image space diffusion by integrating the 3D-to-2D rendering into the reverse process. However, these methods are mostly proposed for static scenes where multiview datasets are available. In contrast to prior research, our method handles dynamic 3D talking heads and is trained exclusively on monocular RGB videos from real-world sources. + +# 3. Method + +Our goal is to generate 3D talking head videos represented as Multiplane Images (MPIs), conditioned on a reference image (the subject's identity) and a sequence of control signals from a parametric model to determine expressions and head poses. The generated videos must exhibit high image quality, controllability, and efficient rendering for VR scenarios, including binocular consistency. + +Given the scarcity of large-scale multi-view video datasets, we design our training pipeline to operate solely on 2D monocular videos from the wild. To leverage the power of 2D diffusion models, we formulate our model as a video diffusion process operating in the image space. However, this raises two key challenges: 1) How to generate 3D scenes from a diffusion denoising process defined on 2D observations, without access to ground-truth MPs? 2) How to learn 3D information from monocular 2D videos? + +To address the first challenge, we design a video diffusion pipeline integrated with a differentiable forward rendering model [35] (Sec. 3.2) to directly model 3D scene distributions from 2D monocular data. + +To address the second challenge, a possible approach involves constructing MPs of a 3D talking head in the refer + +ence image's camera, then moving the target camera to render the head in novel views during training. This allows us to aggregate multi-view observations of the head from different frames into a consistent 3D space to learn 3D shapes. To this end, we propose an alternating training strategy that switches between using the reference and target views as the frontal camera for MPI construction and compute the loss on target views accordingly (Sec. 3.3). This approach encourages the network to learn to synthesize high-quality, sharp images when building the MPIs in target views and to learn multiview consistency when building it in the reference view. When MPIs are constructed in the reference camera, the target ground-truth images cannot be used as input of the diffusion model, as they are not pixel-aligned with the MPIs. To tackle this, we bootstrap our model to generate the pseudo ground truth under the reference camera (Sec. 3.3). During inference, our model generate 3D scene representations in a single denoising diffusion process without any post-processing or optimization (Sec. 3.4). + +# 3.1. Preliminaries + +Diffusion Formulation. Denoising diffusion probabilistic models (DDPM) [23] are one kind of generative models that learn to capture a data distribution $p_{\theta}(\mathbf{x})$ . The method first constructs a forward Markovian process $q(\mathbf{x}_{0:T})$ to iteratively add noise to data as: + +$$ +q (\mathbf {x} _ {t} | \mathbf {x} _ {t - 1}) = \mathcal {N} (\mathbf {x} _ {t}; \sqrt {1 - \beta_ {t}} \mathbf {x} _ {t - 1}, \beta_ {t} \mathbf {I}), \tag {1} +$$ + +where $\beta_{t}$ is scheduled to control the amount of noise being added to data in each step. A denoising diffusion model $g_{\theta}$ with parameters $\theta$ is trained to iteratively perform the reverse process $p_{\theta}(\mathbf{x}_{t - 1}|\mathbf{x}_t)$ to generate a data sample $\mathbf{x}_0$ from a random Gaussian noise $\mathbf{x}_T$ . If we parameterize the model $g_{\theta}$ to predict the estimated clean data sample $\hat{\mathbf{x}}_{t - 1} = g_{\theta}(\mathbf{x}_t,t)\approx \mathbf{x}_0$ , then the reverse process can be written as: + +$$ +\begin{array}{l} p _ {\theta} \left(\mathbf {x} _ {t - 1} \mid \mathbf {x} _ {t}\right) = p _ {\theta} \left(\mathbf {x} _ {t - 1} \mid \mathbf {x} _ {t}, \hat {\mathbf {x}} _ {t - 1}\right) \\ = \mathcal {N} \left(\mathbf {x} _ {t - 1}; \boldsymbol {\mu} _ {t} \left(\mathbf {x} _ {t}, \hat {\mathbf {x}} _ {t - 1}\right), \sigma_ {t} ^ {2} \mathbf {I}\right) \tag {2} \\ \end{array} +$$ + +Please find the detailed formulation of $\pmb{\mu}_t$ and $\sigma_t^2$ in Eq. (7) of DDPM [23]. + +Multiplane Images. We leverage Multiplane Images (MPIs) to represent a 3D scene S (e.g. a 3D talking head). An MPI representation consists of $D$ RGBA image planes placed in a reference 3D space defined by an MPI frontal camera $\phi^{\mathrm{mpi}}$ . These planes are parallel to the image plane of the frontal camera and equally spaced in disparity between near $d_{n}$ and far $d_{f}$ depths from the camera. During rendering, the planes are firstly warped from the frontal camera $\phi^{\mathrm{mpi}}$ to a target camera $\phi^{\mathrm{trgt}}$ by a homography. The warped planes are then composited via alpha blending. We also follow Tucker et al. [36] to use only two RGB images, one as frontal, the other as residual, during MPI rendering to set a proper rendering capability of our model. + +![](images/e7f9121d82338d66dd3c1d1a3cea917b4c07119bc2826f61f8eb5c20888dc5d7.jpg) +Figure 2. Inference pipeline. Our model is built on the architecture of Lumiere [5], which takes an identity image, 2D noise video, the sequence of expressions rendered from 3DMM model and first frame image as input and outputs MPI video sequences. During inference, the network is conditioned on a reference portrait and takes the last frame of previously generated clip as the first frame condition. We separate the network into a color branch and a geometry branch. The two branches share information by zero convolution [45]. + +# 3.2. 3D Talking Head Video Diffusion + +Unlike previous diffusion-based 2D talking head models [29, 42], our model aims to generate a sequence of 3D head representations in the form of an MPI video with $N$ frames $\{\mathbf{S}\}_{N}$ . This is achieved by conditioning the model on a reference identity image $\mathbf{O}^{\mathrm{ref}}$ and a sequence of control signals $\{\mathbf{C}\}_{N}$ , which are the parameters of a mesh-based parametric head model [6, 26], controlling facial expressions and head pose. Inspired by recent 3D diffusion models [1, 35], we propose a video diffusion framework that generates 3D content (an MPI video) directly from 2D observations by integrating the MPI rendering into the denoising process. More specifically, we formulate the forward process in our avatar diffusion as + +$$ +q \left(\mathbf {O} _ {t} ^ {\mathrm {m p i}} \mid \mathbf {O} _ {t - 1} ^ {\mathrm {m p i}}\right) = \mathcal {N} \left(\mathbf {O} _ {t} ^ {\mathrm {m p i}}; \sqrt {1 - \beta_ {t}} \mathbf {O} _ {t - 1} ^ {\mathrm {m p i}}, \beta_ {t} \mathbf {I}\right), \tag {3} +$$ + +where $\mathbf{O}^{\mathrm{mpi}}$ is the 2D video observed by the MPI frontal camera $\phi^{\mathrm{mpi}}$ defined in Sec. 3.1. Note that we intentionally formulate the diffusion process on the image space of MPI frontal camera since it benefits the output quality. Also, we omit the $\{.\}_{N}$ in $\{\mathbf{O}^{\mathrm{mpi}}\}_{N}$ for simplicity and do the same for $\{\phi^{\mathrm{mpi}}\}_{N}$ , $\{\mathbf{S}\}_{N}$ and $\{\mathbf{C}\}_{N}$ . For the reverse process, instead of directly predicting the estimated clean 2D video $\hat{\mathbf{O}}_{t-1}^{\mathrm{mpi}}$ , we let the denoising model $g_{\theta}$ to predict the MPI video $\mathbf{S}_{t-1}$ , followed by the MPI rendering to obtain $\hat{\mathbf{O}}_{t-1}^{\mathrm{mpi}}$ . Formally, we have + +$$ +\mathbf {S} _ {t - 1} = g _ {\theta} \left(\mathbf {O} ^ {\mathrm {r e f}}, \mathbf {O} _ {t} ^ {\mathrm {m p i}}; t, \mathbf {C} ^ {\mathrm {m p i}}\right) \tag {4} +$$ + +$$ +\hat {\mathbf {O}} _ {t - 1} ^ {\mathrm {m p i}} = \mathcal {R} \left(\mathbf {S} _ {t - 1}, \phi^ {\mathrm {m p i}}\right), \tag {5} +$$ + +where $\mathcal{R}$ is the MPI rendering function. Here, we rasterize the control signals $\mathbf{C}$ into UV coordinate maps under the + +MPI frontal camera $\phi^{\mathrm{mpi}}$ to provide a pixel-aligned condition $\mathbf{C}^{\mathrm{mpi}}$ . We also convert the meshes of the parametric model under $\mathbf{C}$ into alpha planes and feed them into the model. In practice, we also concatenate the identity image $\mathbf{O}^{\mathrm{ref}}$ with its corresponding UV coordinate map rasterized from the reference camera as the model input. Finally, we plug in the estimated clean 2D video $\hat{\mathbf{O}}_{t - 1}^{\mathrm{mpi}}$ to Eq. (2) to perform the reverse process, which is repeated $T$ times to sample the final MPI video output $\mathbf{S}_0$ . Given a novel side view camera (or trajectory) $\phi^{\mathrm{side}}$ , we can simply run the MPI rendering to obtain the desired video $\mathbf{O}_0^{\mathrm{side}} = \mathcal{R}(\mathbf{S}_0,\phi^{\mathrm{side}})$ . + +# 3.3. Reference-Target Alternating Training + +Following a similar analysis as Tewari et al. [35], a straightforward way to train our model in Sec. 3.2 needs two losses on videos rendered from two different cameras $\phi^{\mathrm{MPI}}$ (the MPI frontal camera) and $\phi^{\mathrm{side}}$ (a novel side view camera): + +$$ +\hat {\mathbf {O}} _ {t - 1} ^ {\text {s i d e}} = \mathcal {R} \left(\mathbf {S} _ {t - 1}, \phi^ {\text {s i d e}}\right) +$$ + +$$ +\mathcal {L} _ {\theta} ^ {\mathrm {m p i}} = \mathbb {E} _ {\mathbf {O} ^ {\mathrm {r e f}}, \mathbf {O} ^ {\mathrm {m p i}}, \mathbf {C} ^ {\mathrm {m p i}}, t} [ \| \mathbf {O} ^ {\mathrm {m p i}} - \hat {\mathbf {O}} _ {t - 1} ^ {\mathrm {m p i}} \| ^ {2} ] \tag {6} +$$ + +$$ +\mathcal {L} _ {\theta} ^ {\text {s i d e}} = \mathbb {E} _ {\mathbf {O} ^ {\text {r e f}}, \mathbf {O} ^ {\text {m p i}}, \mathbf {C} ^ {\text {m p i}}, t, \mathbf {O} ^ {\text {s i d e}}, \phi^ {\text {s i d e}}} [ \| \mathbf {O} ^ {\text {s i d e}} - \hat {\mathbf {O}} _ {t - 1} ^ {\text {s i d e}} \| ^ {2} ]. \tag {7} +$$ + +Intuitively, the first loss $\mathcal{L}_{\theta}^{\mathrm{mpi}}$ encourages the model to give good frontal views of the MPs, while the second loss $\mathcal{L}_{\theta}^{\mathrm{side}}$ encourages the model to render good side views from the MPs, which would be the only supervision for learning 3D shapes. However, such approach would require a large scale multiview video dataset, which is not generally available. + +In order to enable multi-view supervision when training purely on monocular 2D videos, a possible approach is to place MPs in a canonical space with a fixed head pose, then move the rendering camera to simulate head pose changes. + +![](images/84147aac9bd68af2b8e6b72b8d893ccd8c41c306d238900bbd66b7194036ac12.jpg) +Figure 3. Illustration of Reference-Target Alternating Training. On the top branch, we construct MPs in the target camera and render the frontal view, where we directly use the ground truth target image to compute loss Eq. (6) to learn sharp renderings and bootstrap the model. On the bottom branch, we construct MPs in the reference camera. We first rasterize the target control signal (expression) in the reference camera, then use the bootstrapped model to generate pseudo ground truth by rendering the frontal view. The pseudo ground truth is added with noise, serving as the model input to compute loss Eq. (7) for learning 3D shapes. + +In this way, the image data representing different head poses will act as a proxy ground truth for novel camera views to provide multi-view supervision. Therefore, we could separately compute loss $\mathcal{L}_{\theta}^{\mathrm{mpi}}$ Eq. (6) and loss $\mathcal{L}_{\theta}^{\mathrm{side}}$ Eq. (7) with only monocular 2D videos by using two different MPI setups as shown in Fig. 3. More specifically, when computing the first loss $\mathcal{L}_{\theta}^{\mathrm{mpi}}$ , we would set the MPI frontal cameras $\phi^{\mathrm{mpi}}$ as the target image camera and use the target image as the frontal view ground truth $\mathbf{O}^{\mathrm{mpi}}$ . Similarly, when computing the second loss $\mathcal{L}_{\theta}^{\mathrm{side}}$ , we would set the MPI frontal cameras $\phi^{\mathrm{mpi}}$ to the reference image camera and use the target image as the side view ground truth $\mathbf{O}^{\mathrm{side}}$ . + +However, the above strategy only satisfies the ground truth requirements of the two losses, but does not consider the model input. Indeed, when computing the second loss $\mathcal{L}_{\theta}^{\mathrm{side}}$ Eq. (7), where the MPI frontal cameras $\phi^{\mathrm{mpi}}$ are set to the reference camera, the training video frames are used as the ground truth to supervise the side view of MPs. However these frames cannot be used as input of the diffusion model, as they are no longer pixel-aligned with the MPs, which are constructed using the reference camera in this setting. To tackle this problem, we propose a bootstrapping training strategy that allows our model to generate the pseudo ground truth under the reference camera by itself. + +Bootstrapping. In order to bootstrap our model for generating pseudo ground truth under the reference camera, we firstly train our model $g_{\theta}(.)$ with only the loss $\mathcal{L}_{\theta}^{\mathrm{mpi}}$ Eq. (6) (i.e. setting the MPI frontal cameras $\phi^{\mathrm{mpi}}$ to the target camera). In this way, the model learns to generate plausible frontal views rendered from MPs. + +With this bootstrapping, we can then enable the second loss $\mathcal{L}_{\theta}^{\mathrm{side}}$ Eq. (7) to learn 3D structures of the MPs. More specifically, we first set the MPI frontal cameras $\phi^{\mathrm{mpi}}$ to the + +reference camera, then run the diffusion sampling with $k$ steps to generate the output MPI video $\mathbf{S}_k$ , and render the frontal view to get images $\hat{\mathbf{O}}_k^{\mathrm{mpi}}$ . Since the model has been trained to give plausible frontal views, the rendered images $\hat{\mathbf{O}}_k^{\mathrm{mpi}}$ will have a reasonable quality, acting as the pseudo ground truth under the reference camera. Thus we can add noise to $\hat{\mathbf{O}}_k^{\mathrm{mpi}}$ to obtain the model input $\mathbf{O}_t^{\mathrm{mpi}}$ when computing the second loss $\mathcal{L}_{\theta}^{\mathrm{side}}$ Eq. (7) to train our model for learning 3D shapes inside MPsIs. + +Late-stage Noise Sampling. When training with the second loss $\mathcal{L}_{\theta}^{\mathrm{side}}$ Eq. (7), we compute the model input from the pseudo ground truth $\hat{\mathbf{O}}_k^{\mathrm{mpi}}$ , which cannot perfectly match the real image distribution. Such distribution mismatch could harm the model learning and subsequently lead to blurry results. To mitigate this negative influence, when training with loss $\mathcal{L}_{\theta}^{\mathrm{side}}$ Eq. (7), we only sample $t$ from the second half of the forward Markovian process $[T/2, T]$ so that the added noise is relatively large thus has a better chance to hide the distribution mismatch. + +Loss Functions. During training, we compose the MPs and a background image calculated from the reference portrait $\mathbf{O}^{\mathrm{ref}}$ with alpha blending to get the foreground renderings and the full image renderings. Using pre-computed foreground masks of the data videos, we calculate the standard L2 losses Eq. (6) and Eq. (7) on both the foreground renderings and the full image renderings. Following Tewari et al. [35], we additionally adopt several regularization losses besides the standard L2 losses Eq. (6) and Eq. (7). The regularization losses include: 1) A LPIPS perceptual loss $\mathcal{L}_{\mathrm{lpips}}$ for foreground renderings; 2) A mask loss $\mathcal{L}_{\mathrm{mask}}$ to supervise the rendered alpha mask; 3) An edge-aware depth smoothing loss and a disparity loss following Tucker et al. [36] to further supervise the underlying geometry, + +
MethodHDTFTalkinghead1khNovel View Render Speed
L1 ↓LPIPS ↓FID ↓FVD ↓L1 ↓LPIPS ↓FID ↓FVD ↓
Face-V2V [40]0.0290.13220.5778.110.0400.18630.52141.0222 FPS in 2562
EmoPortrait [16]0.0390.18924.77140.510.0580.25241.08280.227 FPS in 5122
Portrait4d-v2 [12]--25.74139.11--37.38227.3210 FPS in 5122
Follow-your-emoji [29]0.0350.11516.09117.540.0470.15919.37209.910.5FPS* in 5122
X-Portrait [42]0.0340.11915.66132.900.0450.16419.86183.050.1FPS* in 5122
Ours0.0310.11815.9369.040.0450.18027.83135.12109 FPS in 5122
Ours w/o residual0.0300.11614.7665.750.0440.17428.19127.11109 FPS in 5122
Ours w/o late-stage0.0310.12316.7070.690.0450.18029.86136.07109 FPS in 5122
+ +Table 1. Quantitative comparison and ablation study. We compare our method with previous work on both HDTF and Talkinghead1kh datasets. Our method achieves comparable image quality (L1, LPIPS, FID) and the best video quality (FVD) and novel-view rendering efficiency overall. Compared to ablation cases, the averaged metrics are close, and we highlight the actual huge gaps in perceived image, geometry, and side view rendering quality in Fig. 7 and Fig. 6. All the render speeds are tested on a single Nvidia A100 GPU. *We report the generation speeds of diffusion models for Follow-your-emoji [29] and X-Portrait [42], as they do not support explicit camera control. + +where we use the disparity maps rendered from the parametric model as pesudo ground truth for disparity loss. + +# 3.4. Run-time Inference + +When generating a 3D talking head from a driving video, we set the MPI frontal cameras $\phi^{\mathrm{MPI}}$ to the driving video. In order to generate long video sequence and smooth the discontinuity between adjacent video clips, we input ground truth first frame to the model in a similar way as Lumiere [5] with random drop during training. During inference, we take the last frame from the previous clip as the first frame of next clip. The generated first frame is then dropped due to overlapping when concatenating two adjacent clips. + +# 4. Experiments + +In this section, we evaluate our method by comparing with prior arts on talking heads image quality, temporal quality, multi-view consistency, and computational efficiency. We also conduct qualitative comparisons and ablation studies. + +Implementation Details. We train our model on a self-collected dataset consisting of 35k in-the-wild talking head monocular videos. Each video has a length between 4 to 6 seconds, and we crop the face region in $512 \times 512$ for training. Our model generates MPI video with 32 frames in a single network forwarding, and each MPI consists of 24 planes. We dynamically set the near and far planes based on the distance between the 3DMM head joint and the camera position. We optimize our model end-to-end on 8 NVIDIA H100 GPUs with a batch size of 8. During inference, we use DDIM sampling and do classifier-free guidance on first-frame and reference portrait with two scales [8]. The guidance scale for the reference portrait is set to 1.5, while the guidance scale for the first frame is linearly reduced from 1 to 0 between frames 1 and 16. For the remaining 16 frames, the guidance scale for the first frame is fixed at 0. + +Benchmark. We perform our qualitative comparison on two widely used public talking-head datasets, Talking-head1kh [40] and HDTF [47] datasets. For HDTF dataset, + +we randomly sample video clips from 100 identities and each video clip consists of 180 frames. For Talkinghead1kh dataset, we randomly sample two 180-frame video clip from all evaluation identities. During testing, the first frame of each video clip is chosen as reference portrait $\mathbf{O}^{\mathrm{src}}$ and all the animations are performed in a self-reenactment way. Cross-identity driving results are provided in the supp. + +Baselines. We compare our method with existing one-shot photoreal talking heads approaches trained from monocular videos in the wild. Specifically, we compare with representatives for three families of approaches: 1) Feed-forward models (e.g. Face-V2V [40] and EmoPortrait [16]). These methods do not use diffusion model but directly build 3D implicit feature volumes (as appearance) driven by warping fields (as expression). These methods are often optimized for the controllability in 2D space, but not the generation quality or 3D consistency. 2) 3D NeRF-based one-shot talking heads (e.g. Portrait4D-v2 [11]), which establish the standard for 3D consistency due to the use of native 3D representation. 3) Video diffusion based methods (e.g. follow-your-emoji [29] and xportrait [42]), which tend to be particularly strong on the quality of the generated images or videos. We perform all the evaluations on $512 \times 512$ resolution since most of the baselines outputs results at it, except Face-V2V for which we upsample its outputs to the desired resolution. + +# 4.1. Appearance and Motion Quality + +Following prior works, we compare image quality with L1, LPIPS [46] and FID score [22] and video quality, a combination of images and motions, using FVD score [37]. Note that L1 and LPIPS are not available for Portrait4D-v2 due to misalignments between its outputs and ground truth. + +Quantitative results are shown in Tab. 1. While the ranking based on different metrics varies, our method remains competitive among the state-of-the-art. Particularly, our model consistently achieves the best FVD on both datasets indicating our overall best video quality, taking into con + +![](images/c642acab0bdae3620bcbc091ad2797a5d80b15f06d73bba686a3cc8c1aaeb3eb.jpg) +Ref. + +![](images/89d8c590740daa691c9164a8d1259823ed4ae3bc426434c384cfa5e6e4400ac2.jpg) + +![](images/b83c9d19b347d4453050684c07302dd0e0b90d24ba62ac468b56d7d760164420.jpg) +Face-v2v + +![](images/f4d5f62e09ef39015552657f33bed2260f98f06afdd0bc6b6461a55b08d2fd4c.jpg) + +![](images/5a334b689c72cdeecab37b656c19d053afa8834bd64d7b37a116f458a440bafb.jpg) +EmoPortrait + +![](images/8867b80361146200bd439127ec451ae97fe61928cb97f6db43e320712bc07da9.jpg) + +![](images/3290d83bc362a81ab2166fc022a5702389e065d768a7227d79a47ceeb15ace4d.jpg) +Portrait4D-v2 + +![](images/e06bdd21c9c69a581b3515d9ee45d5f360e2ad0834d6411fcf4a243796cee6c2.jpg) + +![](images/bf16d72210ca759e0af88f6bedaa2e36764141e0b8149cdc081b8a143429d3b3.jpg) +Follow-your-emoji + +![](images/e26af30bea92d49287fafb895755257f60cc0ed8b17faa2f15bf604059e6d859.jpg) + +![](images/1d62fa7534ab4fdadf1a2526177418b6ce516889c0219eb2482ed771aebd879d.jpg) +X-Portrait + +![](images/c11f77920000cbb00879dd7af84f18dbb75b756b22d41c3d2fdb402f10161ed4.jpg) + +![](images/819ee2f58c22cdd2813a13c851fdebb7c98e8715291a0e4283fde975dcdc8a42.jpg) +Ours + +![](images/a6a96bb2f37ab399be0d1ce40d8efd7af30a48215e39c6a11d76081493698387.jpg) + +![](images/9d1455dc7359830834edeffff48465822150ba616f41e64e8b7f867a199aadd3.jpg) +G.T. + +![](images/9c34e39dcffc4fb317309105cfb4bd12883a051394dd6d19fb15af1d5a0d02d5.jpg) + +![](images/8018ac97d1feb9051449f7bae0cf8d596aac1e0b583f0eb578465243a7ac1335.jpg) + +![](images/082ef2c299f773c6a8c95823377446841f38d2387ae9d48e6fa6fb5ae93a1970.jpg) +Figure 4. Talking head results. We show results from previous work and our method. Our method generates results with sharp appearance and closely aligned expression with the ground truth. Note Portrait4d-v2 uses a customized camera space which is non-trivial to render GT camera aligned images. + +![](images/62202f3ee1b0a73439096147aa888cb2af8f74743e780b1d8b97c33a7a1e2d59.jpg) + +![](images/4b3d4bfcd46a86e9ce83ad3279312c5330806981d53a613c5ed8c1a8155444b5.jpg) + +![](images/fd1c3105532f0c6ef3831eaeb5f3abff33b9dc4c6f8b246af675f641b129cca5.jpg) + +![](images/f09e6d859d68de1c1f35ec30ff3a903b5c2d2ae188271eabc34e315fbc8acb79.jpg) + +![](images/14c3b381bddb55639298d9d9ed44faaacbaaf50321c41ef58b281a8335951386.jpg) + +![](images/0d6e38aa068c015d03add82d7db30276cb1635c26b6f7248592b44cb2c27e84d.jpg) + +![](images/3270e05dad7b509d8839c76b065adcb5f12ec1faecdec2fc6dc2511bef5aa715.jpg) + +![](images/74b2d41c0208482947589e9817352b004e185e0a72df7246498240873719f934.jpg) + +![](images/9b3debf6ba0ffd62bcfa5ce98b55734ecc1da9b30855356c79a96c676fe95d99.jpg) + +![](images/72b284bbe4bcfc591cc3121a2769582a283f07228b53a006c83f3cf2ed1e119e.jpg) + +![](images/053a66fc000a54e5cc345000c96162b2c5a6ae88c65651b93e847095cbcb1227.jpg) + +![](images/27dfd1e2dd4a220d94eede647ae90873a55063beb59eb237e004346675ddd35f.jpg) + +![](images/7f9e61f023e45e723af21658a2fff36d64603374f16d5dc5e1930f961ff22088.jpg) + +![](images/cda0833a21bf6c6b007c20c10de2245250045513c9cf7028c4163f14848c47c8.jpg) + +sideration both image quality and motion naturalness. This also indicates that a native video diffusion model is stronger than a finetuned image diffusion with added temporal component (like X-Portrait [42] and Follow-your-emoji [29]). + +We also show qualitative results in Fig. 4. FaceV2V [40] gives stable results and follows the input identity, but produces less details due to its low output resolution. EmoPortrait [16] gives sharper results, but suffers from larger identity shifts. Portrait4D-v2 [12] produces good sharpness and less identity shifts, however relatively stiff expressions. The diffusion-based methods Follow-your-emoji [29] and X-portrait [42] render good high-frequency details, but do not precisely follow the input head poses or expressions. In contrast, our method gives good image quality, faithful identity, while follows the input control signal. + +# 4.2. Effectiveness for 3D Rendering + +We evaluate the effectiveness on 3D of different methods from the following aspects. First, we consider the task of binocular stereo rendering, which is one of the foundational use cases for VR/AR applications. Also, we compare how well these approaches perform on perspective geometry by moving the camera closer. + +Note that Face-V2V [40] and EmoPortrait [16] use orthogonal projection as a simplified camera model, thus they cannot produce any reasonable renderings for stereo pairs (i.e. two views will have the same disparity across the whole image) and moving camera closer (i.e. getting identical renderings). Follow-your-emoji [29] and X-xportrait [42] only + +use detected landmarks and images as the driving signal but do not allow any explicit control on the camera, thus they cannot perform these two 3D tasks neither. As a result, we only show the renderings from Portrait4D-v2 [12] and ours, where truth 3D rendering is supported. + +Binocular Stereo Rendering. We render stereo pairs with a 5cm baseline looking from 1m in front of the talking head, and run an off-the-shell stereo matching method [27] to get the disparity map. In Fig. 5, ours gives stereo pairs with reasonable estimated disparities, and is competitive with the NeRF-based Portrait4D-v2 [12], which acts as a strong baseline on 3D effects, while being one order of magnitude faster in novel view rendering (Tab. 1). + +Perspective Geometry. To demonstrate the perspective effect, we render a close view by moving the camera closer while reducing camera focal length to keep the whole head in view. As in Fig. 5, ours produces good perspective distortions that is comparable to the NeRF-based Portrait4Dv2 [12], while being more faithful to the reference identity shown in Fig. 4 second row. + +# 4.3. Ablation + +MPI as representation. Our diffusion model generates MPI frames, where each of them has a frontal, a residual, and a set of planes with alpha. While the frontal contains most of the appearance, we hereby evaluate the importance of the residual image. We train our model producing a MPI with only a frontal image, and reuse it for all the planes for alpha composition. While the metrics on frontal view + +![](images/6194596807da3763c55474691f13be5355903e48a9f6c9faa49e9689229dab22.jpg) +Portrait4d +original + +![](images/2f077548f90113150eef8303a6fb466fae0942714cbde3314480a59f18ca3302.jpg) +stereo + +![](images/7271bb85012a4af6de2a3a62bcfbf31b92bd667b80f115870cc59e34b74bfbbf.jpg) +ours +original + +![](images/8689c874ca36f49b83a75295bb652f024cfa28b1e390dacf8497298847e41bc4.jpg) +stereo + +![](images/f7e2513c9298989e9ff4f8e54b4f3f6c2cd78d749a124f34523585a24994a9e2.jpg) +close view + +![](images/d77810a3249f6029a9324027de75663dea15eff25c1ad28161d17a3469ba3b0b.jpg) +disparity + +![](images/4be5362c1939e65175a089c41581c72f1590ff05e9ce3db2597b23f1a790888e.jpg) +close view + +![](images/d92ce221da6afc2556b3e2e514747bf2b0b4c6d2285333e7f2d19490e7ab0b7c.jpg) +disparity + +![](images/5fc468d1ae9ea3a518f9e5b2836af06a23cfc45a8699ffe48d5ffec0bea91f1e.jpg) +Portrait4d +original + +![](images/4ab73827595ad5c269d7aa175ad5ff213671a620d86511d38c4c720798cb6e82.jpg) +stereo + +![](images/a706d7d8f16f9252adcb13dd0a3aca2fcb2689deac8606b0ef1035f071f4596a.jpg) +ours +original + +![](images/06c43bfda5347cd84dde2db9ab0725cafe168af9adff40f58aac97e7c1aa96c0.jpg) +stereo + +![](images/817280bcaf7387ae19f5f793e5d4ec39250fb286b1f2d1abca04c533f34359f1.jpg) +close view + +![](images/20134c948b0e954bc6f27bab96dd28ef6158162b9cdc6bd7d1ff1443f9e91e0a.jpg) +disparity + +![](images/cede3950382761297e365b735eda01711768badaf49c25639628209f8d98a4e0.jpg) +close view + +![](images/db0e4f4da297a193b7040f5cdab4b8cc9ace940530aa815e1182d929033db986.jpg) +disparity + +![](images/f912fddce4e25d30ea339c9a9eb7fa0d1a4db027fb55e56a7378d05cd57af512.jpg) +Figure 5. Binocular stereo and perspective effects. Our method generates MPI videos that enables efficient spatial rendering, e.g. binocular stereo and perspective effects. We render stereo pairs and calculate a disparity using RAFT-Stereo [27] to visualize the perceived geometry. To demonstrate perspective effects, we render a view with camera moving closer to the subject. Our method renders visually plausible and comparable 3D effects with NeRF based approach like Portrait4d-v2, while being orders of magnitude faster. +Figure 6. Effect of residual RGB. We retrain our model with only a single frontal image in the MPI. We show both the rendered side view images and the disparity converted from the MPI alpha in frontal view. We found this significantly reduces the capacity of MPI, which leads to noisy and irregular geometry and blatant blurriness near the occlusion boundaries. + +become even stronger compared to the full model (1 (Ours w/o residual)), the model produces much worse geometry (Fig. 6), presumable due to the missing of capacity of handling view dependent effects. It also leads to obvious blurriness in the boundary when rendering from the side view. + +A further degenerated case is to use a RGB image and depth for novel-view rendering, and please check our supp for more comparison. + +Late-stage noise sampling. Late-stage noise sampling is also important to learn sharp images since that significantly reduces the negative impact from the imperfect pseudo ground truth. To verify this, we train a model with full noise sampling, and show the quantitative and qualitative comparison in Tab. 1 and Fig. 7. Without large noise sampling, the diffusion model still produces reasonable structure but missing details, e.g. wrinkles, mold, hair strand. + +![](images/ae234078cb16a5a801a65afdbac80c7ebd1ec1c52c1370d9ed343f6fec06426e.jpg) +Figure 7. Effect of Late-stage noise sampling. We show a comparison between our model trained with and without the Late-stage noise sampling. Training with regular noise sampling leads to excessive blurriness and lost of details. + +# 5. Conclusion + +We present a method to learn a 3D-aware video diffusion model, which generates photorealistic talking head multi-plane images given an identity reference image and a sequence of tracking signals. We show such a model can be learned directly from large scale monocular videos without multi-view datasets. With appropriate learning strategies, both 3D information and high quality image generation can be learned concurrently. Extensive experiments show that, our method generates not only talking head video with quality comparable and even better than existing arts but also 3D rendering efficiently that support spatial viewing experience, like in VR headset. As limitations, the generated talking head video does not support rendering from excessive camera viewpoint changes, which is a common issue of the MPI representation. Extending our framework with other efficient 3D representations or enhancing training with multi-view dataset could be a promising direction to improve both the image and 3D quality. + +# References + +[1] Titas Anciukevicius, Zexiang Xu, Matthew Fisher, Paul Henderson, Hakan Bilen, Niloy J Mitra, and Paul Guerrero. Renderdiffusion: Image diffusion for 3d reconstruction, inpainting and generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12608-12618, 2023. 3, 4 +[2] ShahRukh Athar, Zexiang Xu, Kalyan Sunkavalli, Eli Shechtman, and Zhixin Shu. Rignerf: Fully controllable neural 3d portraits. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 20364-20373, 2022. 3 +[3] Ziqian Bai, Feitong Tan, Zeng Huang, Kripasindhu Sarkar, Danhang Tang, Di Qiu, Abhimitra Meka, Ruofei Du, Mingsong Dou, Sergio Orts-Escalano, et al. Learning personalized high quality volumetric head avatars from monocular rgb videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16890-16900, 2023. +[4] Ziqian Bai, Feitong Tan, Sean Fanello, Rohit Pandey, Mingsong Dou, Shichen Liu, Ping Tan, and Yinda Zhang. Efficient 3d implicit head avatar with mesh-anchored hash table blendshapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1975–1984, 2024. 3 +[5] Omer Bar-Tal, Hila Chefer, Omer Tov, Charles Herrmann, Roni Paiss, Shiran Zada, Ariel Ephrat, Junhwa Hur, Guanghui Liu, Amit Raj, et al. Lumiere: A spacetime diffusion model for video generation. arXiv preprint arXiv:2401.12945, 2024. 4, 6 +[6] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 157-164. 2023. 4 +[7] Stella Bounareli, Christos Tzelepis, Vasileios Argyriou, Ioannis Patras, and Georgios Tzimiropoulos. Hyperreenact: One-shot reenactment via jointly learning to refine and retarget faces. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2 +[8] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 6 +[9] Bindita Chaudhuri, Noranart Vesdapunt, Linda Shapiro, and Baoyuan Wang. Personalized face modeling for improved face reconstruction and motion retargeting. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part V 16, pages 142-160. Springer, 2020. 2 +[10] Yufan Chen, Lizhen Wang, Qijing Li, Hongjiang Xiao, Shengping Zhang, Hongxun Yao, and Yebin Liu. Monogaussianavatar: Monocular gaussian point-based head avatar. In ACM SIGGRAPH 2024 Conference Papers, pages 1-9, 2024. 3 +[11] Yu Deng, Duomin Wang, and Baoyuan Wang. Portrait4d-v2: Pseudo multi-view data creates better 4d head synthesizer. arXiv preprint arXiv:2403.13570, 2024. 3, 6 + +[12] Yu Deng, Duomin Wang, and baoyuan Wang. Portrait4d-v2: Pseudo multi-view data creates better 4d head synthesizer. arXiv, 2024. 2, 3, 6, 7 +[13] Zheng Ding, Xuaner Zhang, Zhihao Xia, Lars Jebe, Zhuowen Tu, and Xiuming Zhang. Diffusionrig: Learning personalized priors for facial appearance editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12736-12746, 2023. 2 +[14] Michail Christos Doukas, Stefanos Zafeiriou, and Viktoriia Sharmanska. Headgan: One-shot neural head synthesis and editing. In Proceedings of the IEEE/CVF International conference on Computer Vision, pages 14398-14407, 2021. 2 +[15] Nikita Drobyshev, Jenya Chelishev, Taras Khakhulin, Aleksei Ivakhnenko, Victor Lempitsky, and Egor Zakharov. Megaportraits: One-shot megapixel neural head avatars. In Proceedings of the 30th ACM International Conference on Multimedia, pages 2663–2671, 2022. 2 +[16] Nikita Drobyshev, Antoni Bigata Casademunt, Konstantinos Vougioukas, Zoe Landgraf, Stavros Petridis, and Maja Pantic. Emoportraits: Emotion-enhanced multimodal one-shot head avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8498-8507, 2024. 2, 6, 7 +[17] Emilien Dupont, Hyunjik Kim, SM Eslami, Danilo Rezende, and Dan Rosenbaum. From data to functa: Your data point is a function and you can treat it like one. arXiv preprint arXiv:2201.12204, 2022. 3 +[18] Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8649-8658, 2021. 3 +[19] Xuan Gao, Chenglai Zhong, Jun Xiang, Yang Hong, Yudong Guo, and Juyong Zhang. Reconstructing personalized semantic facial nerf models from monocular video. ACM Transactions on Graphics (TOG), 41(6):1-12, 2022. 3 +[20] Philip-William Grassal, Malte Prinzler, Titus Leistner, Carsten Rother, Matthias Nießner, and Justus Thies. Neural head avatars from monocular rgb videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18653–18664, 2022. 2 +[21] Jianzhu Guo, Dingyun Zhang, Xiaogiang Liu, Zhizhou Zhong, Yuan Zhang, Pengfei Wan, and Di Zhang. Liveportrait: Efficient portrait animation with stitching and retargeting control. arXiv preprint arXiv:2407.03168, 2024. 2 +[22] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6 +[23] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3 +[24] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023. 3 +[25] Yushi Lan, Fangzhou Hong, Shuai Yang, Shangchen Zhou, Xuyi Meng, Bo Dai, Xingang Pan, and Chen Change Loy. + +Ln3diff: Scalable latent neural fields diffusion for speedy 3d generation. In European Conference on Computer Vision, pages 112-130. Springer, 2025. 3 +[26] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph., 36(6):194-1, 2017. 4 +[27] Lahav Lipson, Zachary Teed, and Jia Deng. Raft-stereo: Multilevel recurrent field transforms for stereo matching. In 2021 International Conference on 3D Vision (3DV), pages 218-227. IEEE, 2021. 7, 8 +[28] Tao Liu, Chenpeng Du, Shuai Fan, Feilong Chen, and Kai Yu. Diffhub: Person-generic visual dubbing using inpainting renderer with diffusion auto-encoder. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 3630-3634. IEEE, 2024. 2 +[29] Yue Ma, Hongyu Liu, Hongfa Wang, Heng Pan, Yingqing He, Junkun Yuan, Ailing Zeng, Chengfei Cai, Heung-Yeung Shum, Wei Liu, et al. Follow-your-emoji: Fine-controllable and expressive freestyle portrait animation. arXiv preprint arXiv:2406.01900, 2024. 2, 4, 6, 7 +[30] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2 +[31] Norman Müller, Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Bulo, Peter Kontschieder, and Matthias Nießner. Diffrf: Rendering-guided 3d radiance field diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4328-4338, 2023. 3 +[32] Shenhan Qian, Tobias Kirschstein, Liam Schoneveld, Davide Davoli, Simon Giebenhain, and Matthias Nießner. Gaussianavatars: Photorealistic head avatars with rigged 3d gaus-sians. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20299-20309, 2024. 3 +[33] Shunsuke Saito, Gabriel Schwartz, Tomas Simon, Junxuan Li, and Giljoo Nam. Relightable gaussian codec avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 130-141, 2024. 3 +[34] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. Advances in neural information processing systems, 32, 2019. 2 +[35] Ayush Tewari, Tianwei Yin, George Cazenavette, Semon Rezhikov, Josh Tenenbaum, Frédo Durand, Bill Freeman, and Vincent Sitzmann. Diffusion with forward models: Solving stochastic inverse problems without direct supervision. Advances in Neural Information Processing Systems, 36: 12349-12362, 2023. 3, 4, 5 +[36] Richard Tucker and Noah Snavely. Single-view view synthesis with multiplane images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 551-560, 2020. 3, 5 +[37] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. To + +wards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 6 +[38] Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, Karsten Kreis, et al. Lion: Latent point diffusion models for 3d shape generation. Advances in Neural Information Processing Systems, 35:10021-10039, 2022. 3 +[39] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, et al. Rodin: A generative model for sculpting 3d digital avatars using diffusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4563-4573, 2023. 3 +[40] Ting-Chun Wang, Arun Mallya, and Ming-Yu Liu. One-shot free-view neural talking-head synthesis for video conferencing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10039-10049, 2021. 2, 6, 7 +[41] Zhen Wang, Qiangeng Xu, Feitong Tan, Mengei Chai, Shichen Liu, Rohit Pandey, Sean Fanello, Achuta Kadambi, and Yinda Zhang. Mvdd: Multi-view depth diffusion models. In European Conference on Computer Vision, pages 236-253. Springer, 2025. 3 +[42] You Xie, Hongyi Xu, Guoxian Song, Chao Wang, Yichun Shi, and Linjie Luo. X-portrait: Expressive portrait animation with hierarchical motion attention. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2, 4, 6, 7 +[43] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. Styleheat: One-shot high-resolution editable talking face generation via pre-trained stylegan. In European conference on computer vision, pages 85-101. Springer, 2022. 2 +[44] Egor Zakharov, Aleksei Ivakhnenko, Aliaksandra Shysheya, and Victor Lempitsky. Fast bi-layer neural synthesis of one-shot realistic head avatars. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XII 16, pages 524-540. Springer, 2020. 2 +[45] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 4 +[46] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6 +[47] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661–3670, 2021. 6 +[48] Wojciech Zielonka, Timo Bolkart, and Justus Thies. Instant volumetric head avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4574-4584, 2023. 3 \ No newline at end of file diff --git a/2025/IM-Portrait_ Learning 3D-aware Video Diffusion for Photorealistic Talking Heads from Monocular VideosC/images.zip b/2025/IM-Portrait_ Learning 3D-aware Video Diffusion for Photorealistic Talking Heads from Monocular VideosC/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9496b1c5dade888d285a362b34050a6dfa94e441 --- /dev/null +++ b/2025/IM-Portrait_ Learning 3D-aware Video Diffusion for Photorealistic Talking Heads from Monocular VideosC/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d4f2814f5f05b4bdfa5cff66f474fa45683e81346c1d4d47e6d52f52c20dd292 +size 781825 diff --git a/2025/IM-Portrait_ Learning 3D-aware Video Diffusion for Photorealistic Talking Heads from Monocular VideosC/layout.json b/2025/IM-Portrait_ Learning 3D-aware Video Diffusion for Photorealistic Talking Heads from Monocular VideosC/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..397cea3fc29f410590f1ed26515c1ba9b898b367 --- /dev/null +++ b/2025/IM-Portrait_ Learning 3D-aware Video Diffusion for Photorealistic Talking Heads from Monocular VideosC/layout.json @@ -0,0 +1,11190 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 99, + 103, + 512, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 103, + 512, + 140 + ], + "spans": [ + { + "bbox": [ + 99, + 103, + 512, + 140 + ], + "type": "text", + "content": "IM-Portrait: Learning 3D-aware Video Diffusion for Photorealistic Talking Heads from Monocular Videos" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 159, + 491, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 159, + 491, + 205 + ], + "spans": [ + { + "bbox": [ + 116, + 159, + 491, + 205 + ], + "type": "text", + "content": "Yuan Li" + }, + { + "bbox": [ + 116, + 159, + 491, + 205 + ], + "type": "inline_equation", + "content": "^{1,2*§}" + }, + { + "bbox": [ + 116, + 159, + 491, + 205 + ], + "type": "text", + "content": " Ziqian Bai" + }, + { + "bbox": [ + 116, + 159, + 491, + 205 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 116, + 159, + 491, + 205 + ], + "type": "text", + "content": " Feitong Tan" + }, + { + "bbox": [ + 116, + 159, + 491, + 205 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 116, + 159, + 491, + 205 + ], + "type": "text", + "content": " Zhaopeng Cui" + }, + { + "bbox": [ + 116, + 159, + 491, + 205 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 116, + 159, + 491, + 205 + ], + "type": "text", + "content": " Sean Fanello" + }, + { + "bbox": [ + 116, + 159, + 491, + 205 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 116, + 159, + 491, + 205 + ], + "type": "text", + "content": " Yinda Zhang" + }, + { + "bbox": [ + 116, + 159, + 491, + 205 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 116, + 159, + 491, + 205 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 116, + 159, + 491, + 205 + ], + "type": "text", + "content": "State Key Lab of CAD&CG, Zhejiang University " + }, + { + "bbox": [ + 116, + 159, + 491, + 205 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 116, + 159, + 491, + 205 + ], + "type": "text", + "content": "Google" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 60, + 227, + 141, + 338 + ], + "blocks": [ + { + "bbox": [ + 82, + 214, + 114, + 225 + ], + "lines": [ + { + "bbox": [ + 82, + 214, + 114, + 225 + ], + "spans": [ + { + "bbox": [ + 82, + 214, + 114, + 225 + ], + "type": "text", + "content": "Inputs" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 60, + 227, + 141, + 338 + ], + "lines": [ + { + "bbox": [ + 60, + 227, + 141, + 338 + ], + "spans": [ + { + "bbox": [ + 60, + 227, + 141, + 338 + ], + "type": "image", + "image_path": "ec9265c354faa8646c8f3e3609394ad1ca6cea16252a845c6c2c582dff18c73d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 145, + 227, + 222, + 337 + ], + "blocks": [ + { + "bbox": [ + 176, + 214, + 263, + 224 + ], + "lines": [ + { + "bbox": [ + 176, + 214, + 263, + 224 + ], + "spans": [ + { + "bbox": [ + 176, + 214, + 263, + 224 + ], + "type": "text", + "content": "Generated Results" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 145, + 227, + 222, + 337 + ], + "lines": [ + { + "bbox": [ + 145, + 227, + 222, + 337 + ], + "spans": [ + { + "bbox": [ + 145, + 227, + 222, + 337 + ], + "type": "image", + "image_path": "3fd86c5b7b4dce66736362fbb47662b14ff889b097dcad8e5c5c104b04acc529.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 225, + 227, + 302, + 337 + ], + "blocks": [ + { + "bbox": [ + 225, + 227, + 302, + 337 + ], + "lines": [ + { + "bbox": [ + 225, + 227, + 302, + 337 + ], + "spans": [ + { + "bbox": [ + 225, + 227, + 302, + 337 + ], + "type": "image", + "image_path": "73b69f0a3b8f6b1e1ad6d356ed210561ba0f665808b37483e3b9a9f668211473.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 60, + 343, + 140, + 455 + ], + "blocks": [ + { + "bbox": [ + 60, + 343, + 140, + 455 + ], + "lines": [ + { + "bbox": [ + 60, + 343, + 140, + 455 + ], + "spans": [ + { + "bbox": [ + 60, + 343, + 140, + 455 + ], + "type": "image", + "image_path": "de5f90519439ddd28220066c889d6301b4846905698e0679a5142c716876dead.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 460, + 555, + 504 + ], + "lines": [ + { + "bbox": [ + 55, + 460, + 555, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 555, + 504 + ], + "type": "text", + "content": "Figure 1. We propose a 3D-aware video diffusion model for talking head synthesis. Given an image as identity and a sequence of tracking signals (as shown on the left for each example), our model directly generates videos in Multiplane Images (MPIs) in a single denoising process, which is ready for efficient novel-view rendering. This enables immersive viewing experience, e.g. rendering binocular stereo or perspective distortion in VR headset." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 145, + 343, + 222, + 453 + ], + "blocks": [ + { + "bbox": [ + 145, + 343, + 222, + 453 + ], + "lines": [ + { + "bbox": [ + 145, + 343, + 222, + 453 + ], + "spans": [ + { + "bbox": [ + 145, + 343, + 222, + 453 + ], + "type": "image", + "image_path": "39e679271344c55ad39128792cff5a576f415bda6d20f1f565510c27728a9724.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 225, + 343, + 303, + 453 + ], + "blocks": [ + { + "bbox": [ + 225, + 343, + 303, + 453 + ], + "lines": [ + { + "bbox": [ + 225, + 343, + 303, + 453 + ], + "spans": [ + { + "bbox": [ + 225, + 343, + 303, + 453 + ], + "type": "image", + "image_path": "0f2026fbf02b6b9aa72d77aca96cf9106890065906adf574188bbaf19fc45f1e.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 312, + 226, + 390, + 336 + ], + "blocks": [ + { + "bbox": [ + 334, + 214, + 364, + 225 + ], + "lines": [ + { + "bbox": [ + 334, + 214, + 364, + 225 + ], + "spans": [ + { + "bbox": [ + 334, + 214, + 364, + 225 + ], + "type": "text", + "content": "Inputs" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 312, + 226, + 390, + 336 + ], + "lines": [ + { + "bbox": [ + 312, + 226, + 390, + 336 + ], + "spans": [ + { + "bbox": [ + 312, + 226, + 390, + 336 + ], + "type": "image", + "image_path": "34992577c8be39c707c626f21a9f0b9075dd446c73a5213cc2a4c8c5ba9f0fe4.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 395, + 226, + 472, + 336 + ], + "blocks": [ + { + "bbox": [ + 427, + 214, + 514, + 224 + ], + "lines": [ + { + "bbox": [ + 427, + 214, + 514, + 224 + ], + "spans": [ + { + "bbox": [ + 427, + 214, + 514, + 224 + ], + "type": "text", + "content": "Generated Results" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 395, + 226, + 472, + 336 + ], + "lines": [ + { + "bbox": [ + 395, + 226, + 472, + 336 + ], + "spans": [ + { + "bbox": [ + 395, + 226, + 472, + 336 + ], + "type": "image", + "image_path": "1ae3402671efe9adefba86ad704f97f2dd666fc44d01090e3cfde99218ecca1b.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 476, + 227, + 553, + 336 + ], + "blocks": [ + { + "bbox": [ + 476, + 227, + 553, + 336 + ], + "lines": [ + { + "bbox": [ + 476, + 227, + 553, + 336 + ], + "spans": [ + { + "bbox": [ + 476, + 227, + 553, + 336 + ], + "type": "image", + "image_path": "7bf921429c92fbd024546c9c194049301803dacf37b21fdcf8d517b8a99778f9.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 312, + 343, + 390, + 453 + ], + "blocks": [ + { + "bbox": [ + 312, + 343, + 390, + 453 + ], + "lines": [ + { + "bbox": [ + 312, + 343, + 390, + 453 + ], + "spans": [ + { + "bbox": [ + 312, + 343, + 390, + 453 + ], + "type": "image", + "image_path": "9990816515274e38c39a4ae9c4269d30110b18c5ac120a7ad0e1446885558faf.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 395, + 343, + 472, + 453 + ], + "blocks": [ + { + "bbox": [ + 395, + 343, + 472, + 453 + ], + "lines": [ + { + "bbox": [ + 395, + 343, + 472, + 453 + ], + "spans": [ + { + "bbox": [ + 395, + 343, + 472, + 453 + ], + "type": "image", + "image_path": "6fe0a1cc798d76d9ccc3c5f90355f6d00aeed7dda8d6af7a0075322e5a5f1bff.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 476, + 343, + 553, + 453 + ], + "blocks": [ + { + "bbox": [ + 476, + 343, + 553, + 453 + ], + "lines": [ + { + "bbox": [ + 476, + 343, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 476, + 343, + 553, + 453 + ], + "type": "image", + "image_path": "22e0927c20c3ff7ceec2a2c2b0b571116889d2f7b4c05f9f899c44590fb2f781.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 152, + 513, + 200, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 513, + 200, + 525 + ], + "spans": [ + { + "bbox": [ + 152, + 513, + 200, + 525 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 55, + 540, + 296, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 296, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 296, + 672 + ], + "type": "text", + "content": "We propose a novel 3D-aware diffusion-based method for generating photorealistic talking head videos directly from a single identity image and explicit control signals (e.g., expressions). Our method generates Multiplane Images (MPIs) that ensure geometric consistency, making them ideal for immersive viewing experiences like binocular videos for VR headsets. Unlike existing methods that often require a separate stage or joint optimization to reconstruct a 3D representation (such as NeRF or 3D Gaussians), our approach directly generates the final output through a single denoising process, eliminating the need for post" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 514, + 555, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 555, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 555, + 635 + ], + "type": "text", + "content": "processing steps to render novel views efficiently. To effectively learn from monocular videos, we introduce a training mechanism that reconstructs the output MPI randomly in either the target or the reference camera space. This approach enables the model to simultaneously learn sharp image details and underlying 3D information. Extensive experiments demonstrate the effectiveness of our method, which achieves competitive avatar quality and novel-view rendering capabilities, even without explicit 3D reconstruction or high-quality multi-view training data." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 314, + 657, + 395, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 657, + 395, + 670 + ], + "spans": [ + { + "bbox": [ + 314, + 657, + 395, + 670 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "Synthesizing realistic talking heads, has been a long-lasting challenging problem in computer vision for decades. The recent emergence of diffusion models within the generative" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 683, + 164, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 683, + 164, + 693 + ], + "spans": [ + { + "bbox": [ + 66, + 683, + 164, + 693 + ], + "type": "text", + "content": "*Authors contributed equally." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 67, + 693, + 148, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 693, + 148, + 703 + ], + "spans": [ + { + "bbox": [ + 67, + 693, + 148, + 703 + ], + "type": "text", + "content": "† Corresponding authors." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 67, + 703, + 266, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 266, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 266, + 712 + ], + "type": "inline_equation", + "content": "^{\\S}" + }, + { + "bbox": [ + 67, + 703, + 266, + 712 + ], + "type": "text", + "content": "Work was conducted while Yuan Li was an intern at Google." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21107" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": "AI domain offers a promising solution to this problem. Significant advancements have been made in modeling talking heads using image or video diffusion models conditioned on tracking signals [13, 28, 29, 42], leveraging the remarkable generative capabilities of diffusion models." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 134, + 295, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 134, + 295, + 314 + ], + "spans": [ + { + "bbox": [ + 55, + 134, + 295, + 314 + ], + "type": "text", + "content": "Most diffusion-based approaches, however, operate in 2D image space, and the extension of these methods to 3D rendering remains largely unexplored. One potential approach involves applying 3D transformations to the tracking signal, such as parametric models or landmarks. However, this necessitates running the reverse diffusion process for each new camera view, leading to significant computational overhead. Alternatively, some diffusion-based approaches generate multi-view images, which often lack geometric consistency. While off-the-shelf methods like NeRF [12] or 3D Gaussians can be used to construct a 3D representation for efficient rendering, the inherent inconsistencies in the input multi-view images limit the quality of the final 3D reconstruction, leading to potential issues in multi-view and temporal consistency." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 316, + 295, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 295, + 484 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 295, + 484 + ], + "type": "text", + "content": "In this work, we present, for the first time, a diffusion based approach that produces 3D talking heads that can be rendered efficiently in varying camera poses. We adopt Multiplane Image (MPI) as the 3D representation due to its remarkable rendering efficiency and quality, and the affinity to 2D images such that we can make good use of existing 2D diffusion models. Essentially, we formulate the problem of synthesizing talking heads images as a denoising diffusion process, which is conditioned on a reference image (i.e. the subject identity) and a sequence of control signals from parametric models (i.e. expressions and head poses). In a single denoising process, our model generates MPI videos that can be readily rendered in varying camera pose and support spatial viewing experience." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "type": "text", + "content": "Our model is learned directly from a large scale monocular talking head videos in-the-wild. Due to the missing of multi-view datasets, like previous works, the 3D consistency needs to be learned from the head pose variability between video frames. To achieve this, it is required to construct MPs in a camera that looks at the head from a different viewpoint than the target images. Specifically, we use the viewpoint of the reference image to construct MPs. The estimated target images can then be rendered from MPs by moving the camera according to the relative head pose differences between the reference and target images. In this way, however, the frontal views of MPs will be misaligned with the input noisy images (in our case the target ground-truth images), which will break the training of the diffusion model. To tackle the problem, we propose a novel training framework to bootstrap our model to synthesize pseudo noisy input images that align with the MPs in reference cameras. With this, the MPs are randomly generated from either the reference camera viewpoint or the" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "text", + "content": "target camera viewpoint during the training, which allows the diffusion model to learn both 3D information and fine details." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 326, + 110, + 459, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 110, + 459, + 120 + ], + "spans": [ + { + "bbox": [ + 326, + 110, + 459, + 120 + ], + "type": "text", + "content": "Our contributions are as follows." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 122, + 553, + 314 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 314, + 122, + 553, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 122, + 553, + 145 + ], + "spans": [ + { + "bbox": [ + 314, + 122, + 553, + 145 + ], + "type": "text", + "content": "- We propose a diffusion-based approach for 3D-aware talking heads generation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 146, + 553, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 146, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 314, + 146, + 553, + 182 + ], + "type": "text", + "content": "- Our diffusion model directly generates 3D renderable talking heads in a single denoising process, eliminating the need for post-processing steps." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 182, + 553, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 182, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 314, + 182, + 553, + 241 + ], + "type": "text", + "content": "- We propose a novel training framework that enables our model to learn from monocular videos, and show how randomly selecting the camera viewpoint between reference and target images allows for better convergence and sharper results." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 242, + 553, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 242, + 553, + 314 + ], + "spans": [ + { + "bbox": [ + 314, + 242, + 553, + 314 + ], + "type": "text", + "content": "- Extensive experiments demonstrate that our method achieves comparable or superior image and video quality to existing diffusion-based 2D talking head methods, while also offering efficient and consistent multiview rendering on par with specialized (i.e. non-diffusion based) 3D approaches." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 327, + 406, + 339 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 327, + 406, + 339 + ], + "spans": [ + { + "bbox": [ + 314, + 327, + 406, + 339 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 348, + 555, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 348, + 555, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 348, + 555, + 647 + ], + "type": "text", + "content": "2D Talking Heads. In recent year, there has been a great advancement on synthesizing 2D talking heads with neural networks. A recent line of works [7, 14-16, 34, 40, 43, 44] propose to leverage neural networks to encode the appearance and motion from input images [15, 16, 34, 40] (or explicit control signals [14, 43, 44]) to some latent features or manually defined representations (e.g. key-points, warping fields, texture maps), which are injected into 2D CNN generators with various architectures for talking head synthesis. More recently, many works [13, 28, 29, 42] started to adopt 2D diffusion models, leveraging their impressive image quality, for controllable talking head synthesis. However, these approaches usually use simplified camera models (e.g. orthogonal projection) or they do not have any explicit camera control, making them unsuitable for spatial viewing experiences such as binocular stereo and perspective geometry. In addition, these methods usually do not produce explicit 3D geometry, which also limits their downstream applications. While most of prior work learns talking head models from monocular videos which are available in a large scale in-the-wild, multi-view datasets, e.g. from lightstage [21], could be leveraged to further improve fidelity, however such data requires intensive collection labor and there are no large scale multi-view datasets that are publicly available." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 653, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 553, + 713 + ], + "type": "text", + "content": "3D Talking Heads. The problem of generating 3D talking heads has been a hot topic in the community. People attempted to model their talking heads with various 3D representations, such as textured mesh [9, 20], Neural Radiance Fields (NeRFs) [30] or other implicit func" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21108" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "content": "tions [2-4, 11, 12, 18, 19, 48], and, more recently, Gaussian Splatting [10, 32, 33]. While most of these approaches are person-specific, which requires to retrain a new model for every new person, some methods [11, 12] also explore the potential of person-agnostic models with explicit 3D representations. Despite having good 3D properties, these approaches cannot easily leverage the recent advancement on 2D diffusion, which lower the upper bound of their image quality." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 186, + 296, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 296, + 413 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 296, + 413 + ], + "type": "text", + "content": "Diffusion-based 3D Generative Model. Another line of work, investigates how to apply diffusion models on 3D representations. The most straightforward way is applying the standard diffusion process directly on 3D data [38, 41]. However, this requires large scale 3D datasets, which are generally not available. One typical workaround is to train the diffusion model with 2D images. Some works [17, 24, 25, 31, 39] firstly learn a latent space of the 3D samples with multiview 2D images, then apply diffusion on the learned latent space. However, this results in an additional latent space learning, which is non-trivial thus increasing the pipeline complexity. Some other works [1, 35] directly model 3D scene distributions with image space diffusion by integrating the 3D-to-2D rendering into the reverse process. However, these methods are mostly proposed for static scenes where multiview datasets are available. In contrast to prior research, our method handles dynamic 3D talking heads and is trained exclusively on monocular RGB videos from real-world sources." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 426, + 111, + 439 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 111, + 439 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 111, + 439 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 448, + 295, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 448, + 295, + 532 + ], + "spans": [ + { + "bbox": [ + 55, + 448, + 295, + 532 + ], + "type": "text", + "content": "Our goal is to generate 3D talking head videos represented as Multiplane Images (MPIs), conditioned on a reference image (the subject's identity) and a sequence of control signals from a parametric model to determine expressions and head poses. The generated videos must exhibit high image quality, controllability, and efficient rendering for VR scenarios, including binocular consistency." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 532, + 295, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 532, + 295, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 532, + 295, + 639 + ], + "type": "text", + "content": "Given the scarcity of large-scale multi-view video datasets, we design our training pipeline to operate solely on 2D monocular videos from the wild. To leverage the power of 2D diffusion models, we formulate our model as a video diffusion process operating in the image space. However, this raises two key challenges: 1) How to generate 3D scenes from a diffusion denoising process defined on 2D observations, without access to ground-truth MPs? 2) How to learn 3D information from monocular 2D videos?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 641, + 295, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 295, + 688 + ], + "type": "text", + "content": "To address the first challenge, we design a video diffusion pipeline integrated with a differentiable forward rendering model [35] (Sec. 3.2) to directly model 3D scene distributions from 2D monocular data." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "To address the second challenge, a possible approach involves constructing MPs of a 3D talking head in the refer" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 300 + ], + "type": "text", + "content": "ence image's camera, then moving the target camera to render the head in novel views during training. This allows us to aggregate multi-view observations of the head from different frames into a consistent 3D space to learn 3D shapes. To this end, we propose an alternating training strategy that switches between using the reference and target views as the frontal camera for MPI construction and compute the loss on target views accordingly (Sec. 3.3). This approach encourages the network to learn to synthesize high-quality, sharp images when building the MPIs in target views and to learn multiview consistency when building it in the reference view. When MPIs are constructed in the reference camera, the target ground-truth images cannot be used as input of the diffusion model, as they are not pixel-aligned with the MPIs. To tackle this, we bootstrap our model to generate the pseudo ground truth under the reference camera (Sec. 3.3). During inference, our model generate 3D scene representations in a single denoising diffusion process without any post-processing or optimization (Sec. 3.4)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 304, + 400, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 400, + 316 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 400, + 316 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 323, + 555, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 555, + 382 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 555, + 382 + ], + "type": "text", + "content": "Diffusion Formulation. Denoising diffusion probabilistic models (DDPM) [23] are one kind of generative models that learn to capture a data distribution " + }, + { + "bbox": [ + 313, + 323, + 555, + 382 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 323, + 555, + 382 + ], + "type": "text", + "content": ". The method first constructs a forward Markovian process " + }, + { + "bbox": [ + 313, + 323, + 555, + 382 + ], + "type": "inline_equation", + "content": "q(\\mathbf{x}_{0:T})" + }, + { + "bbox": [ + 313, + 323, + 555, + 382 + ], + "type": "text", + "content": " to iteratively add noise to data as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 349, + 387, + 555, + 402 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 387, + 555, + 402 + ], + "spans": [ + { + "bbox": [ + 349, + 387, + 555, + 402 + ], + "type": "interline_equation", + "content": "q (\\mathbf {x} _ {t} | \\mathbf {x} _ {t - 1}) = \\mathcal {N} (\\mathbf {x} _ {t}; \\sqrt {1 - \\beta_ {t}} \\mathbf {x} _ {t - 1}, \\beta_ {t} \\mathbf {I}), \\tag {1}", + "image_path": "5761610af616440504ea113e6d7d1523a1623f0ad48a6bbdfd3ab00c828fac71.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 405, + 554, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 405, + 554, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 405, + 554, + 491 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 405, + 554, + 491 + ], + "type": "inline_equation", + "content": "\\beta_{t}" + }, + { + "bbox": [ + 313, + 405, + 554, + 491 + ], + "type": "text", + "content": " is scheduled to control the amount of noise being added to data in each step. A denoising diffusion model " + }, + { + "bbox": [ + 313, + 405, + 554, + 491 + ], + "type": "inline_equation", + "content": "g_{\\theta}" + }, + { + "bbox": [ + 313, + 405, + 554, + 491 + ], + "type": "text", + "content": " with parameters " + }, + { + "bbox": [ + 313, + 405, + 554, + 491 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 405, + 554, + 491 + ], + "type": "text", + "content": " is trained to iteratively perform the reverse process " + }, + { + "bbox": [ + 313, + 405, + 554, + 491 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\mathbf{x}_{t - 1}|\\mathbf{x}_t)" + }, + { + "bbox": [ + 313, + 405, + 554, + 491 + ], + "type": "text", + "content": " to generate a data sample " + }, + { + "bbox": [ + 313, + 405, + 554, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 313, + 405, + 554, + 491 + ], + "type": "text", + "content": " from a random Gaussian noise " + }, + { + "bbox": [ + 313, + 405, + 554, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T" + }, + { + "bbox": [ + 313, + 405, + 554, + 491 + ], + "type": "text", + "content": ". If we parameterize the model " + }, + { + "bbox": [ + 313, + 405, + 554, + 491 + ], + "type": "inline_equation", + "content": "g_{\\theta}" + }, + { + "bbox": [ + 313, + 405, + 554, + 491 + ], + "type": "text", + "content": " to predict the estimated clean data sample " + }, + { + "bbox": [ + 313, + 405, + 554, + 491 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_{t - 1} = g_{\\theta}(\\mathbf{x}_t,t)\\approx \\mathbf{x}_0" + }, + { + "bbox": [ + 313, + 405, + 554, + 491 + ], + "type": "text", + "content": ", then the reverse process can be written as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 344, + 494, + 553, + 523 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 494, + 553, + 523 + ], + "spans": [ + { + "bbox": [ + 344, + 494, + 553, + 523 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} p _ {\\theta} \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}\\right) = p _ {\\theta} \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}, \\hat {\\mathbf {x}} _ {t - 1}\\right) \\\\ = \\mathcal {N} \\left(\\mathbf {x} _ {t - 1}; \\boldsymbol {\\mu} _ {t} \\left(\\mathbf {x} _ {t}, \\hat {\\mathbf {x}} _ {t - 1}\\right), \\sigma_ {t} ^ {2} \\mathbf {I}\\right) \\tag {2} \\\\ \\end{array}", + "image_path": "d48dbe0a1ce4026c1e64618319d42a73ed390d2a8cbb7ba1dde10111a743bda2.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 528, + 553, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 528, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 528, + 553, + 552 + ], + "type": "text", + "content": "Please find the detailed formulation of " + }, + { + "bbox": [ + 313, + 528, + 553, + 552 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}_t" + }, + { + "bbox": [ + 313, + 528, + 553, + 552 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 528, + 553, + 552 + ], + "type": "inline_equation", + "content": "\\sigma_t^2" + }, + { + "bbox": [ + 313, + 528, + 553, + 552 + ], + "type": "text", + "content": " in Eq. (7) of DDPM [23]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 558, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 715 + ], + "type": "text", + "content": "Multiplane Images. We leverage Multiplane Images (MPIs) to represent a 3D scene S (e.g. a 3D talking head). An MPI representation consists of " + }, + { + "bbox": [ + 313, + 558, + 555, + 715 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 558, + 555, + 715 + ], + "type": "text", + "content": " RGBA image planes placed in a reference 3D space defined by an MPI frontal camera " + }, + { + "bbox": [ + 313, + 558, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\phi^{\\mathrm{mpi}}" + }, + { + "bbox": [ + 313, + 558, + 555, + 715 + ], + "type": "text", + "content": ". These planes are parallel to the image plane of the frontal camera and equally spaced in disparity between near " + }, + { + "bbox": [ + 313, + 558, + 555, + 715 + ], + "type": "inline_equation", + "content": "d_{n}" + }, + { + "bbox": [ + 313, + 558, + 555, + 715 + ], + "type": "text", + "content": " and far " + }, + { + "bbox": [ + 313, + 558, + 555, + 715 + ], + "type": "inline_equation", + "content": "d_{f}" + }, + { + "bbox": [ + 313, + 558, + 555, + 715 + ], + "type": "text", + "content": " depths from the camera. During rendering, the planes are firstly warped from the frontal camera " + }, + { + "bbox": [ + 313, + 558, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\phi^{\\mathrm{mpi}}" + }, + { + "bbox": [ + 313, + 558, + 555, + 715 + ], + "type": "text", + "content": " to a target camera " + }, + { + "bbox": [ + 313, + 558, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\phi^{\\mathrm{trgt}}" + }, + { + "bbox": [ + 313, + 558, + 555, + 715 + ], + "type": "text", + "content": " by a homography. The warped planes are then composited via alpha blending. We also follow Tucker et al. [36] to use only two RGB images, one as frontal, the other as residual, during MPI rendering to set a proper rendering capability of our model." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21109" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 70, + 541, + 273 + ], + "blocks": [ + { + "bbox": [ + 69, + 70, + 541, + 273 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 541, + 273 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 541, + 273 + ], + "type": "image", + "image_path": "e7f9121d82338d66dd3c1d1a3cea917b4c07119bc2826f61f8eb5c20888dc5d7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 277, + 555, + 322 + ], + "lines": [ + { + "bbox": [ + 54, + 277, + 555, + 322 + ], + "spans": [ + { + "bbox": [ + 54, + 277, + 555, + 322 + ], + "type": "text", + "content": "Figure 2. Inference pipeline. Our model is built on the architecture of Lumiere [5], which takes an identity image, 2D noise video, the sequence of expressions rendered from 3DMM model and first frame image as input and outputs MPI video sequences. During inference, the network is conditioned on a reference portrait and takes the last frame of previously generated clip as the first frame condition. We separate the network into a color branch and a geometry branch. The two branches share information by zero convolution [45]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 327, + 234, + 340 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 234, + 340 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 234, + 340 + ], + "type": "text", + "content": "3.2. 3D Talking Head Video Diffusion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 344, + 296, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 344, + 296, + 499 + ], + "spans": [ + { + "bbox": [ + 55, + 344, + 296, + 499 + ], + "type": "text", + "content": "Unlike previous diffusion-based 2D talking head models [29, 42], our model aims to generate a sequence of 3D head representations in the form of an MPI video with " + }, + { + "bbox": [ + 55, + 344, + 296, + 499 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 344, + 296, + 499 + ], + "type": "text", + "content": " frames " + }, + { + "bbox": [ + 55, + 344, + 296, + 499 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{S}\\}_{N}" + }, + { + "bbox": [ + 55, + 344, + 296, + 499 + ], + "type": "text", + "content": ". This is achieved by conditioning the model on a reference identity image " + }, + { + "bbox": [ + 55, + 344, + 296, + 499 + ], + "type": "inline_equation", + "content": "\\mathbf{O}^{\\mathrm{ref}}" + }, + { + "bbox": [ + 55, + 344, + 296, + 499 + ], + "type": "text", + "content": " and a sequence of control signals " + }, + { + "bbox": [ + 55, + 344, + 296, + 499 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{C}\\}_{N}" + }, + { + "bbox": [ + 55, + 344, + 296, + 499 + ], + "type": "text", + "content": ", which are the parameters of a mesh-based parametric head model [6, 26], controlling facial expressions and head pose. Inspired by recent 3D diffusion models [1, 35], we propose a video diffusion framework that generates 3D content (an MPI video) directly from 2D observations by integrating the MPI rendering into the denoising process. More specifically, we formulate the forward process in our avatar diffusion as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 71, + 505, + 295, + 520 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 505, + 295, + 520 + ], + "spans": [ + { + "bbox": [ + 71, + 505, + 295, + 520 + ], + "type": "interline_equation", + "content": "q \\left(\\mathbf {O} _ {t} ^ {\\mathrm {m p i}} \\mid \\mathbf {O} _ {t - 1} ^ {\\mathrm {m p i}}\\right) = \\mathcal {N} \\left(\\mathbf {O} _ {t} ^ {\\mathrm {m p i}}; \\sqrt {1 - \\beta_ {t}} \\mathbf {O} _ {t - 1} ^ {\\mathrm {m p i}}, \\beta_ {t} \\mathbf {I}\\right), \\tag {3}", + "image_path": "b2064a93beaa016986854008938282cdc4d3059908f08cbf4d968ce45099cde1.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 525, + 296, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 296, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 296, + 647 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 525, + 296, + 647 + ], + "type": "inline_equation", + "content": "\\mathbf{O}^{\\mathrm{mpi}}" + }, + { + "bbox": [ + 55, + 525, + 296, + 647 + ], + "type": "text", + "content": " is the 2D video observed by the MPI frontal camera " + }, + { + "bbox": [ + 55, + 525, + 296, + 647 + ], + "type": "inline_equation", + "content": "\\phi^{\\mathrm{mpi}}" + }, + { + "bbox": [ + 55, + 525, + 296, + 647 + ], + "type": "text", + "content": " defined in Sec. 3.1. Note that we intentionally formulate the diffusion process on the image space of MPI frontal camera since it benefits the output quality. Also, we omit the " + }, + { + "bbox": [ + 55, + 525, + 296, + 647 + ], + "type": "inline_equation", + "content": "\\{.\\}_{N}" + }, + { + "bbox": [ + 55, + 525, + 296, + 647 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 55, + 525, + 296, + 647 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{O}^{\\mathrm{mpi}}\\}_{N}" + }, + { + "bbox": [ + 55, + 525, + 296, + 647 + ], + "type": "text", + "content": " for simplicity and do the same for " + }, + { + "bbox": [ + 55, + 525, + 296, + 647 + ], + "type": "inline_equation", + "content": "\\{\\phi^{\\mathrm{mpi}}\\}_{N}" + }, + { + "bbox": [ + 55, + 525, + 296, + 647 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 525, + 296, + 647 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{S}\\}_{N}" + }, + { + "bbox": [ + 55, + 525, + 296, + 647 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 525, + 296, + 647 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{C}\\}_{N}" + }, + { + "bbox": [ + 55, + 525, + 296, + 647 + ], + "type": "text", + "content": ". For the reverse process, instead of directly predicting the estimated clean 2D video " + }, + { + "bbox": [ + 55, + 525, + 296, + 647 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{O}}_{t-1}^{\\mathrm{mpi}}" + }, + { + "bbox": [ + 55, + 525, + 296, + 647 + ], + "type": "text", + "content": ", we let the denoising model " + }, + { + "bbox": [ + 55, + 525, + 296, + 647 + ], + "type": "inline_equation", + "content": "g_{\\theta}" + }, + { + "bbox": [ + 55, + 525, + 296, + 647 + ], + "type": "text", + "content": " to predict the MPI video " + }, + { + "bbox": [ + 55, + 525, + 296, + 647 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{t-1}" + }, + { + "bbox": [ + 55, + 525, + 296, + 647 + ], + "type": "text", + "content": ", followed by the MPI rendering to obtain " + }, + { + "bbox": [ + 55, + 525, + 296, + 647 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{O}}_{t-1}^{\\mathrm{mpi}}" + }, + { + "bbox": [ + 55, + 525, + 296, + 647 + ], + "type": "text", + "content": ". Formally, we have" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 112, + 652, + 295, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 652, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 112, + 652, + 295, + 666 + ], + "type": "interline_equation", + "content": "\\mathbf {S} _ {t - 1} = g _ {\\theta} \\left(\\mathbf {O} ^ {\\mathrm {r e f}}, \\mathbf {O} _ {t} ^ {\\mathrm {m p i}}; t, \\mathbf {C} ^ {\\mathrm {m p i}}\\right) \\tag {4}", + "image_path": "e2e6b62a9117ca0d760dce8c65a5c9f84461f7ee1ea52eef584b9053c319a4d5.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 111, + 669, + 295, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 669, + 295, + 685 + ], + "spans": [ + { + "bbox": [ + 111, + 669, + 295, + 685 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {O}} _ {t - 1} ^ {\\mathrm {m p i}} = \\mathcal {R} \\left(\\mathbf {S} _ {t - 1}, \\phi^ {\\mathrm {m p i}}\\right), \\tag {5}", + "image_path": "72f9143438b897fa2c3887975d7fe58738afe992ffda094e357dd190e84e9a6b.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " is the MPI rendering function. Here, we rasterize the control signals " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " into UV coordinate maps under the" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 327, + 555, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 555, + 460 + ], + "type": "text", + "content": "MPI frontal camera " + }, + { + "bbox": [ + 313, + 327, + 555, + 460 + ], + "type": "inline_equation", + "content": "\\phi^{\\mathrm{mpi}}" + }, + { + "bbox": [ + 313, + 327, + 555, + 460 + ], + "type": "text", + "content": " to provide a pixel-aligned condition " + }, + { + "bbox": [ + 313, + 327, + 555, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{C}^{\\mathrm{mpi}}" + }, + { + "bbox": [ + 313, + 327, + 555, + 460 + ], + "type": "text", + "content": ". We also convert the meshes of the parametric model under " + }, + { + "bbox": [ + 313, + 327, + 555, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 313, + 327, + 555, + 460 + ], + "type": "text", + "content": " into alpha planes and feed them into the model. In practice, we also concatenate the identity image " + }, + { + "bbox": [ + 313, + 327, + 555, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{O}^{\\mathrm{ref}}" + }, + { + "bbox": [ + 313, + 327, + 555, + 460 + ], + "type": "text", + "content": " with its corresponding UV coordinate map rasterized from the reference camera as the model input. Finally, we plug in the estimated clean 2D video " + }, + { + "bbox": [ + 313, + 327, + 555, + 460 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{O}}_{t - 1}^{\\mathrm{mpi}}" + }, + { + "bbox": [ + 313, + 327, + 555, + 460 + ], + "type": "text", + "content": " to Eq. (2) to perform the reverse process, which is repeated " + }, + { + "bbox": [ + 313, + 327, + 555, + 460 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 327, + 555, + 460 + ], + "type": "text", + "content": " times to sample the final MPI video output " + }, + { + "bbox": [ + 313, + 327, + 555, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_0" + }, + { + "bbox": [ + 313, + 327, + 555, + 460 + ], + "type": "text", + "content": ". Given a novel side view camera (or trajectory) " + }, + { + "bbox": [ + 313, + 327, + 555, + 460 + ], + "type": "inline_equation", + "content": "\\phi^{\\mathrm{side}}" + }, + { + "bbox": [ + 313, + 327, + 555, + 460 + ], + "type": "text", + "content": ", we can simply run the MPI rendering to obtain the desired video " + }, + { + "bbox": [ + 313, + 327, + 555, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_0^{\\mathrm{side}} = \\mathcal{R}(\\mathbf{S}_0,\\phi^{\\mathrm{side}})" + }, + { + "bbox": [ + 313, + 327, + 555, + 460 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 465, + 519, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 465, + 519, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 465, + 519, + 479 + ], + "type": "text", + "content": "3.3. Reference-Target Alternating Training" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 483, + 554, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 554, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 554, + 532 + ], + "type": "text", + "content": "Following a similar analysis as Tewari et al. [35], a straightforward way to train our model in Sec. 3.2 needs two losses on videos rendered from two different cameras " + }, + { + "bbox": [ + 313, + 483, + 554, + 532 + ], + "type": "inline_equation", + "content": "\\phi^{\\mathrm{MPI}}" + }, + { + "bbox": [ + 313, + 483, + 554, + 532 + ], + "type": "text", + "content": " (the MPI frontal camera) and " + }, + { + "bbox": [ + 313, + 483, + 554, + 532 + ], + "type": "inline_equation", + "content": "\\phi^{\\mathrm{side}}" + }, + { + "bbox": [ + 313, + 483, + 554, + 532 + ], + "type": "text", + "content": " (a novel side view camera):" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 323, + 536, + 421, + 551 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 536, + 421, + 551 + ], + "spans": [ + { + "bbox": [ + 323, + 536, + 421, + 551 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {O}} _ {t - 1} ^ {\\text {s i d e}} = \\mathcal {R} \\left(\\mathbf {S} _ {t - 1}, \\phi^ {\\text {s i d e}}\\right)", + "image_path": "2465890a626f007e738fc8c2210ab4afb3bbe1e9fc5a77a015ddbedc91c7e1f7.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 328, + 553, + 553, + 569 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 553, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 328, + 553, + 553, + 569 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\theta} ^ {\\mathrm {m p i}} = \\mathbb {E} _ {\\mathbf {O} ^ {\\mathrm {r e f}}, \\mathbf {O} ^ {\\mathrm {m p i}}, \\mathbf {C} ^ {\\mathrm {m p i}}, t} [ \\| \\mathbf {O} ^ {\\mathrm {m p i}} - \\hat {\\mathbf {O}} _ {t - 1} ^ {\\mathrm {m p i}} \\| ^ {2} ] \\tag {6}", + "image_path": "1f0437672c508e491936452a52a2bf7b90130ec85b3821cbd0e2b1ed770292e8.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 328, + 571, + 553, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 571, + 553, + 586 + ], + "spans": [ + { + "bbox": [ + 328, + 571, + 553, + 586 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\theta} ^ {\\text {s i d e}} = \\mathbb {E} _ {\\mathbf {O} ^ {\\text {r e f}}, \\mathbf {O} ^ {\\text {m p i}}, \\mathbf {C} ^ {\\text {m p i}}, t, \\mathbf {O} ^ {\\text {s i d e}}, \\phi^ {\\text {s i d e}}} [ \\| \\mathbf {O} ^ {\\text {s i d e}} - \\hat {\\mathbf {O}} _ {t - 1} ^ {\\text {s i d e}} \\| ^ {2} ]. \\tag {7}", + "image_path": "668735c2415a8028b01364d1b906776866f5efd569936b34e005fd1c29405e51.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 594, + 554, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 554, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 554, + 665 + ], + "type": "text", + "content": "Intuitively, the first loss " + }, + { + "bbox": [ + 313, + 594, + 554, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\theta}^{\\mathrm{mpi}}" + }, + { + "bbox": [ + 313, + 594, + 554, + 665 + ], + "type": "text", + "content": " encourages the model to give good frontal views of the MPs, while the second loss " + }, + { + "bbox": [ + 313, + 594, + 554, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\theta}^{\\mathrm{side}}" + }, + { + "bbox": [ + 313, + 594, + 554, + 665 + ], + "type": "text", + "content": " encourages the model to render good side views from the MPs, which would be the only supervision for learning 3D shapes. However, such approach would require a large scale multiview video dataset, which is not generally available." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 666, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 554, + 714 + ], + "type": "text", + "content": "In order to enable multi-view supervision when training purely on monocular 2D videos, a possible approach is to place MPs in a canonical space with a fixed head pose, then move the rendering camera to simulate head pose changes." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21110" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 72, + 539, + 243 + ], + "blocks": [ + { + "bbox": [ + 72, + 72, + 539, + 243 + ], + "lines": [ + { + "bbox": [ + 72, + 72, + 539, + 243 + ], + "spans": [ + { + "bbox": [ + 72, + 72, + 539, + 243 + ], + "type": "image", + "image_path": "84147aac9bd68af2b8e6b72b8d893ccd8c41c306d238900bbd66b7194036ac12.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 246, + 555, + 302 + ], + "lines": [ + { + "bbox": [ + 54, + 246, + 555, + 302 + ], + "spans": [ + { + "bbox": [ + 54, + 246, + 555, + 302 + ], + "type": "text", + "content": "Figure 3. Illustration of Reference-Target Alternating Training. On the top branch, we construct MPs in the target camera and render the frontal view, where we directly use the ground truth target image to compute loss Eq. (6) to learn sharp renderings and bootstrap the model. On the bottom branch, we construct MPs in the reference camera. We first rasterize the target control signal (expression) in the reference camera, then use the bootstrapped model to generate pseudo ground truth by rendering the frontal view. The pseudo ground truth is added with noise, serving as the model input to compute loss Eq. (7) for learning 3D shapes." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 304, + 296, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 304, + 296, + 448 + ], + "spans": [ + { + "bbox": [ + 54, + 304, + 296, + 448 + ], + "type": "text", + "content": "In this way, the image data representing different head poses will act as a proxy ground truth for novel camera views to provide multi-view supervision. Therefore, we could separately compute loss " + }, + { + "bbox": [ + 54, + 304, + 296, + 448 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\theta}^{\\mathrm{mpi}}" + }, + { + "bbox": [ + 54, + 304, + 296, + 448 + ], + "type": "text", + "content": " Eq. (6) and loss " + }, + { + "bbox": [ + 54, + 304, + 296, + 448 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\theta}^{\\mathrm{side}}" + }, + { + "bbox": [ + 54, + 304, + 296, + 448 + ], + "type": "text", + "content": " Eq. (7) with only monocular 2D videos by using two different MPI setups as shown in Fig. 3. More specifically, when computing the first loss " + }, + { + "bbox": [ + 54, + 304, + 296, + 448 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\theta}^{\\mathrm{mpi}}" + }, + { + "bbox": [ + 54, + 304, + 296, + 448 + ], + "type": "text", + "content": ", we would set the MPI frontal cameras " + }, + { + "bbox": [ + 54, + 304, + 296, + 448 + ], + "type": "inline_equation", + "content": "\\phi^{\\mathrm{mpi}}" + }, + { + "bbox": [ + 54, + 304, + 296, + 448 + ], + "type": "text", + "content": " as the target image camera and use the target image as the frontal view ground truth " + }, + { + "bbox": [ + 54, + 304, + 296, + 448 + ], + "type": "inline_equation", + "content": "\\mathbf{O}^{\\mathrm{mpi}}" + }, + { + "bbox": [ + 54, + 304, + 296, + 448 + ], + "type": "text", + "content": ". Similarly, when computing the second loss " + }, + { + "bbox": [ + 54, + 304, + 296, + 448 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\theta}^{\\mathrm{side}}" + }, + { + "bbox": [ + 54, + 304, + 296, + 448 + ], + "type": "text", + "content": ", we would set the MPI frontal cameras " + }, + { + "bbox": [ + 54, + 304, + 296, + 448 + ], + "type": "inline_equation", + "content": "\\phi^{\\mathrm{mpi}}" + }, + { + "bbox": [ + 54, + 304, + 296, + 448 + ], + "type": "text", + "content": " to the reference image camera and use the target image as the side view ground truth " + }, + { + "bbox": [ + 54, + 304, + 296, + 448 + ], + "type": "inline_equation", + "content": "\\mathbf{O}^{\\mathrm{side}}" + }, + { + "bbox": [ + 54, + 304, + 296, + 448 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 451, + 296, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 451, + 296, + 595 + ], + "spans": [ + { + "bbox": [ + 55, + 451, + 296, + 595 + ], + "type": "text", + "content": "However, the above strategy only satisfies the ground truth requirements of the two losses, but does not consider the model input. Indeed, when computing the second loss " + }, + { + "bbox": [ + 55, + 451, + 296, + 595 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\theta}^{\\mathrm{side}}" + }, + { + "bbox": [ + 55, + 451, + 296, + 595 + ], + "type": "text", + "content": " Eq. (7), where the MPI frontal cameras " + }, + { + "bbox": [ + 55, + 451, + 296, + 595 + ], + "type": "inline_equation", + "content": "\\phi^{\\mathrm{mpi}}" + }, + { + "bbox": [ + 55, + 451, + 296, + 595 + ], + "type": "text", + "content": " are set to the reference camera, the training video frames are used as the ground truth to supervise the side view of MPs. However these frames cannot be used as input of the diffusion model, as they are no longer pixel-aligned with the MPs, which are constructed using the reference camera in this setting. To tackle this problem, we propose a bootstrapping training strategy that allows our model to generate the pseudo ground truth under the reference camera by itself." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 601, + 296, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 601, + 296, + 674 + ], + "spans": [ + { + "bbox": [ + 55, + 601, + 296, + 674 + ], + "type": "text", + "content": "Bootstrapping. In order to bootstrap our model for generating pseudo ground truth under the reference camera, we firstly train our model " + }, + { + "bbox": [ + 55, + 601, + 296, + 674 + ], + "type": "inline_equation", + "content": "g_{\\theta}(.)" + }, + { + "bbox": [ + 55, + 601, + 296, + 674 + ], + "type": "text", + "content": " with only the loss " + }, + { + "bbox": [ + 55, + 601, + 296, + 674 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\theta}^{\\mathrm{mpi}}" + }, + { + "bbox": [ + 55, + 601, + 296, + 674 + ], + "type": "text", + "content": " Eq. (6) (i.e. setting the MPI frontal cameras " + }, + { + "bbox": [ + 55, + 601, + 296, + 674 + ], + "type": "inline_equation", + "content": "\\phi^{\\mathrm{mpi}}" + }, + { + "bbox": [ + 55, + 601, + 296, + 674 + ], + "type": "text", + "content": " to the target camera). In this way, the model learns to generate plausible frontal views rendered from MPs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "With this bootstrapping, we can then enable the second loss " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\theta}^{\\mathrm{side}}" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": " Eq. (7) to learn 3D structures of the MPs. More specifically, we first set the MPI frontal cameras " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\phi^{\\mathrm{mpi}}" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": " to the" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 304, + 555, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 555, + 413 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 555, + 413 + ], + "type": "text", + "content": "reference camera, then run the diffusion sampling with " + }, + { + "bbox": [ + 313, + 304, + 555, + 413 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 304, + 555, + 413 + ], + "type": "text", + "content": " steps to generate the output MPI video " + }, + { + "bbox": [ + 313, + 304, + 555, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_k" + }, + { + "bbox": [ + 313, + 304, + 555, + 413 + ], + "type": "text", + "content": ", and render the frontal view to get images " + }, + { + "bbox": [ + 313, + 304, + 555, + 413 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{O}}_k^{\\mathrm{mpi}}" + }, + { + "bbox": [ + 313, + 304, + 555, + 413 + ], + "type": "text", + "content": ". Since the model has been trained to give plausible frontal views, the rendered images " + }, + { + "bbox": [ + 313, + 304, + 555, + 413 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{O}}_k^{\\mathrm{mpi}}" + }, + { + "bbox": [ + 313, + 304, + 555, + 413 + ], + "type": "text", + "content": " will have a reasonable quality, acting as the pseudo ground truth under the reference camera. Thus we can add noise to " + }, + { + "bbox": [ + 313, + 304, + 555, + 413 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{O}}_k^{\\mathrm{mpi}}" + }, + { + "bbox": [ + 313, + 304, + 555, + 413 + ], + "type": "text", + "content": " to obtain the model input " + }, + { + "bbox": [ + 313, + 304, + 555, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_t^{\\mathrm{mpi}}" + }, + { + "bbox": [ + 313, + 304, + 555, + 413 + ], + "type": "text", + "content": " when computing the second loss " + }, + { + "bbox": [ + 313, + 304, + 555, + 413 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\theta}^{\\mathrm{side}}" + }, + { + "bbox": [ + 313, + 304, + 555, + 413 + ], + "type": "text", + "content": " Eq. (7) to train our model for learning 3D shapes inside MPsIs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 418, + 556, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 556, + 539 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 556, + 539 + ], + "type": "text", + "content": "Late-stage Noise Sampling. When training with the second loss " + }, + { + "bbox": [ + 313, + 418, + 556, + 539 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\theta}^{\\mathrm{side}}" + }, + { + "bbox": [ + 313, + 418, + 556, + 539 + ], + "type": "text", + "content": " Eq. (7), we compute the model input from the pseudo ground truth " + }, + { + "bbox": [ + 313, + 418, + 556, + 539 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{O}}_k^{\\mathrm{mpi}}" + }, + { + "bbox": [ + 313, + 418, + 556, + 539 + ], + "type": "text", + "content": ", which cannot perfectly match the real image distribution. Such distribution mismatch could harm the model learning and subsequently lead to blurry results. To mitigate this negative influence, when training with loss " + }, + { + "bbox": [ + 313, + 418, + 556, + 539 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\theta}^{\\mathrm{side}}" + }, + { + "bbox": [ + 313, + 418, + 556, + 539 + ], + "type": "text", + "content": " Eq. (7), we only sample " + }, + { + "bbox": [ + 313, + 418, + 556, + 539 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 418, + 556, + 539 + ], + "type": "text", + "content": " from the second half of the forward Markovian process " + }, + { + "bbox": [ + 313, + 418, + 556, + 539 + ], + "type": "inline_equation", + "content": "[T/2, T]" + }, + { + "bbox": [ + 313, + 418, + 556, + 539 + ], + "type": "text", + "content": " so that the added noise is relatively large thus has a better chance to hide the distribution mismatch." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "type": "text", + "content": "Loss Functions. During training, we compose the MPs and a background image calculated from the reference portrait " + }, + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{O}^{\\mathrm{ref}}" + }, + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "type": "text", + "content": " with alpha blending to get the foreground renderings and the full image renderings. Using pre-computed foreground masks of the data videos, we calculate the standard L2 losses Eq. (6) and Eq. (7) on both the foreground renderings and the full image renderings. Following Tewari et al. [35], we additionally adopt several regularization losses besides the standard L2 losses Eq. (6) and Eq. (7). The regularization losses include: 1) A LPIPS perceptual loss " + }, + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{lpips}}" + }, + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "type": "text", + "content": " for foreground renderings; 2) A mask loss " + }, + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{mask}}" + }, + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "type": "text", + "content": " to supervise the rendered alpha mask; 3) An edge-aware depth smoothing loss and a disparity loss following Tucker et al. [36] to further supervise the underlying geometry," + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21111" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 77, + 70, + 533, + 189 + ], + "blocks": [ + { + "bbox": [ + 77, + 70, + 533, + 189 + ], + "lines": [ + { + "bbox": [ + 77, + 70, + 533, + 189 + ], + "spans": [ + { + "bbox": [ + 77, + 70, + 533, + 189 + ], + "type": "table", + "html": "
MethodHDTFTalkinghead1khNovel View Render Speed
L1 ↓LPIPS ↓FID ↓FVD ↓L1 ↓LPIPS ↓FID ↓FVD ↓
Face-V2V [40]0.0290.13220.5778.110.0400.18630.52141.0222 FPS in 2562
EmoPortrait [16]0.0390.18924.77140.510.0580.25241.08280.227 FPS in 5122
Portrait4d-v2 [12]--25.74139.11--37.38227.3210 FPS in 5122
Follow-your-emoji [29]0.0350.11516.09117.540.0470.15919.37209.910.5FPS* in 5122
X-Portrait [42]0.0340.11915.66132.900.0450.16419.86183.050.1FPS* in 5122
Ours0.0310.11815.9369.040.0450.18027.83135.12109 FPS in 5122
Ours w/o residual0.0300.11614.7665.750.0440.17428.19127.11109 FPS in 5122
Ours w/o late-stage0.0310.12316.7070.690.0450.18029.86136.07109 FPS in 5122
", + "image_path": "dd36d299f27cae28dddc731a48352e05b7a2cc3c11876bac43d667aab7cce981.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 190, + 555, + 246 + ], + "lines": [ + { + "bbox": [ + 55, + 190, + 555, + 246 + ], + "spans": [ + { + "bbox": [ + 55, + 190, + 555, + 246 + ], + "type": "text", + "content": "Table 1. Quantitative comparison and ablation study. We compare our method with previous work on both HDTF and Talkinghead1kh datasets. Our method achieves comparable image quality (L1, LPIPS, FID) and the best video quality (FVD) and novel-view rendering efficiency overall. Compared to ablation cases, the averaged metrics are close, and we highlight the actual huge gaps in perceived image, geometry, and side view rendering quality in Fig. 7 and Fig. 6. All the render speeds are tested on a single Nvidia A100 GPU. *We report the generation speeds of diffusion models for Follow-your-emoji [29] and X-Portrait [42], as they do not support explicit camera control." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 248, + 294, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 248, + 294, + 272 + ], + "spans": [ + { + "bbox": [ + 55, + 248, + 294, + 272 + ], + "type": "text", + "content": "where we use the disparity maps rendered from the parametric model as pesudo ground truth for disparity loss." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 278, + 170, + 290 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 278, + 170, + 290 + ], + "spans": [ + { + "bbox": [ + 55, + 278, + 170, + 290 + ], + "type": "text", + "content": "3.4. Run-time Inference" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 296, + 295, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 296, + 295, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 296, + 295, + 403 + ], + "type": "text", + "content": "When generating a 3D talking head from a driving video, we set the MPI frontal cameras " + }, + { + "bbox": [ + 55, + 296, + 295, + 403 + ], + "type": "inline_equation", + "content": "\\phi^{\\mathrm{MPI}}" + }, + { + "bbox": [ + 55, + 296, + 295, + 403 + ], + "type": "text", + "content": " to the driving video. In order to generate long video sequence and smooth the discontinuity between adjacent video clips, we input ground truth first frame to the model in a similar way as Lumiere [5] with random drop during training. During inference, we take the last frame from the previous clip as the first frame of next clip. The generated first frame is then dropped due to overlapping when concatenating two adjacent clips." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 413, + 136, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 413, + 136, + 426 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 136, + 426 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 433, + 295, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 433, + 295, + 480 + ], + "spans": [ + { + "bbox": [ + 55, + 433, + 295, + 480 + ], + "type": "text", + "content": "In this section, we evaluate our method by comparing with prior arts on talking heads image quality, temporal quality, multi-view consistency, and computational efficiency. We also conduct qualitative comparisons and ablation studies." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 483, + 295, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 295, + 674 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 295, + 674 + ], + "type": "text", + "content": "Implementation Details. We train our model on a self-collected dataset consisting of 35k in-the-wild talking head monocular videos. Each video has a length between 4 to 6 seconds, and we crop the face region in " + }, + { + "bbox": [ + 55, + 483, + 295, + 674 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 55, + 483, + 295, + 674 + ], + "type": "text", + "content": " for training. Our model generates MPI video with 32 frames in a single network forwarding, and each MPI consists of 24 planes. We dynamically set the near and far planes based on the distance between the 3DMM head joint and the camera position. We optimize our model end-to-end on 8 NVIDIA H100 GPUs with a batch size of 8. During inference, we use DDIM sampling and do classifier-free guidance on first-frame and reference portrait with two scales [8]. The guidance scale for the reference portrait is set to 1.5, while the guidance scale for the first frame is linearly reduced from 1 to 0 between frames 1 and 16. For the remaining 16 frames, the guidance scale for the first frame is fixed at 0." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "Benchmark. We perform our qualitative comparison on two widely used public talking-head datasets, Talking-head1kh [40] and HDTF [47] datasets. For HDTF dataset," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 248, + 555, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 248, + 555, + 332 + ], + "spans": [ + { + "bbox": [ + 313, + 248, + 555, + 332 + ], + "type": "text", + "content": "we randomly sample video clips from 100 identities and each video clip consists of 180 frames. For Talkinghead1kh dataset, we randomly sample two 180-frame video clip from all evaluation identities. During testing, the first frame of each video clip is chosen as reference portrait " + }, + { + "bbox": [ + 313, + 248, + 555, + 332 + ], + "type": "inline_equation", + "content": "\\mathbf{O}^{\\mathrm{src}}" + }, + { + "bbox": [ + 313, + 248, + 555, + 332 + ], + "type": "text", + "content": " and all the animations are performed in a self-reenactment way. Cross-identity driving results are provided in the supp." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 335, + 556, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 556, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 556, + 563 + ], + "type": "text", + "content": "Baselines. We compare our method with existing one-shot photoreal talking heads approaches trained from monocular videos in the wild. Specifically, we compare with representatives for three families of approaches: 1) Feed-forward models (e.g. Face-V2V [40] and EmoPortrait [16]). These methods do not use diffusion model but directly build 3D implicit feature volumes (as appearance) driven by warping fields (as expression). These methods are often optimized for the controllability in 2D space, but not the generation quality or 3D consistency. 2) 3D NeRF-based one-shot talking heads (e.g. Portrait4D-v2 [11]), which establish the standard for 3D consistency due to the use of native 3D representation. 3) Video diffusion based methods (e.g. follow-your-emoji [29] and xportrait [42]), which tend to be particularly strong on the quality of the generated images or videos. We perform all the evaluations on " + }, + { + "bbox": [ + 313, + 335, + 556, + 563 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 313, + 335, + 556, + 563 + ], + "type": "text", + "content": " resolution since most of the baselines outputs results at it, except Face-V2V for which we upsample its outputs to the desired resolution." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 574, + 488, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 574, + 488, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 574, + 488, + 586 + ], + "type": "text", + "content": "4.1. Appearance and Motion Quality" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 593, + 554, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 593, + 554, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 593, + 554, + 653 + ], + "type": "text", + "content": "Following prior works, we compare image quality with L1, LPIPS [46] and FID score [22] and video quality, a combination of images and motions, using FVD score [37]. Note that L1 and LPIPS are not available for Portrait4D-v2 due to misalignments between its outputs and ground truth." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": "Quantitative results are shown in Tab. 1. While the ranking based on different metrics varies, our method remains competitive among the state-of-the-art. Particularly, our model consistently achieves the best FVD on both datasets indicating our overall best video quality, taking into con" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21112" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 82, + 120, + 143 + ], + "blocks": [ + { + "bbox": [ + 83, + 70, + 96, + 77 + ], + "lines": [ + { + "bbox": [ + 83, + 70, + 96, + 77 + ], + "spans": [ + { + "bbox": [ + 83, + 70, + 96, + 77 + ], + "type": "text", + "content": "Ref." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 82, + 120, + 143 + ], + "lines": [ + { + "bbox": [ + 59, + 82, + 120, + 143 + ], + "spans": [ + { + "bbox": [ + 59, + 82, + 120, + 143 + ], + "type": "image", + "image_path": "c642acab0bdae3620bcbc091ad2797a5d80b15f06d73bba686a3cc8c1aaeb3eb.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 59, + 144, + 120, + 205 + ], + "blocks": [ + { + "bbox": [ + 59, + 144, + 120, + 205 + ], + "lines": [ + { + "bbox": [ + 59, + 144, + 120, + 205 + ], + "spans": [ + { + "bbox": [ + 59, + 144, + 120, + 205 + ], + "type": "image", + "image_path": "89d8c590740daa691c9164a8d1259823ed4ae3bc426434c384cfa5e6e4400ac2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 128, + 82, + 188, + 143 + ], + "blocks": [ + { + "bbox": [ + 143, + 70, + 173, + 77 + ], + "lines": [ + { + "bbox": [ + 143, + 70, + 173, + 77 + ], + "spans": [ + { + "bbox": [ + 143, + 70, + 173, + 77 + ], + "type": "text", + "content": "Face-v2v" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 128, + 82, + 188, + 143 + ], + "lines": [ + { + "bbox": [ + 128, + 82, + 188, + 143 + ], + "spans": [ + { + "bbox": [ + 128, + 82, + 188, + 143 + ], + "type": "image", + "image_path": "b83c9d19b347d4453050684c07302dd0e0b90d24ba62ac468b56d7d760164420.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 128, + 144, + 188, + 205 + ], + "blocks": [ + { + "bbox": [ + 128, + 144, + 188, + 205 + ], + "lines": [ + { + "bbox": [ + 128, + 144, + 188, + 205 + ], + "spans": [ + { + "bbox": [ + 128, + 144, + 188, + 205 + ], + "type": "image", + "image_path": "f4d5f62e09ef39015552657f33bed2260f98f06afdd0bc6b6461a55b08d2fd4c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 189, + 82, + 249, + 143 + ], + "blocks": [ + { + "bbox": [ + 200, + 70, + 239, + 77 + ], + "lines": [ + { + "bbox": [ + 200, + 70, + 239, + 77 + ], + "spans": [ + { + "bbox": [ + 200, + 70, + 239, + 77 + ], + "type": "text", + "content": "EmoPortrait" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 189, + 82, + 249, + 143 + ], + "lines": [ + { + "bbox": [ + 189, + 82, + 249, + 143 + ], + "spans": [ + { + "bbox": [ + 189, + 82, + 249, + 143 + ], + "type": "image", + "image_path": "5a334b689c72cdeecab37b656c19d053afa8834bd64d7b37a116f458a440bafb.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 189, + 144, + 249, + 205 + ], + "blocks": [ + { + "bbox": [ + 189, + 144, + 249, + 205 + ], + "lines": [ + { + "bbox": [ + 189, + 144, + 249, + 205 + ], + "spans": [ + { + "bbox": [ + 189, + 144, + 249, + 205 + ], + "type": "image", + "image_path": "8867b80361146200bd439127ec451ae97fe61928cb97f6db43e320712bc07da9.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 250, + 82, + 309, + 143 + ], + "blocks": [ + { + "bbox": [ + 257, + 70, + 301, + 77 + ], + "lines": [ + { + "bbox": [ + 257, + 70, + 301, + 77 + ], + "spans": [ + { + "bbox": [ + 257, + 70, + 301, + 77 + ], + "type": "text", + "content": "Portrait4D-v2" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 250, + 82, + 309, + 143 + ], + "lines": [ + { + "bbox": [ + 250, + 82, + 309, + 143 + ], + "spans": [ + { + "bbox": [ + 250, + 82, + 309, + 143 + ], + "type": "image", + "image_path": "3290d83bc362a81ab2166fc022a5702389e065d768a7227d79a47ceeb15ace4d.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 249, + 144, + 309, + 205 + ], + "blocks": [ + { + "bbox": [ + 249, + 144, + 309, + 205 + ], + "lines": [ + { + "bbox": [ + 249, + 144, + 309, + 205 + ], + "spans": [ + { + "bbox": [ + 249, + 144, + 309, + 205 + ], + "type": "image", + "image_path": "e06bdd21c9c69a581b3515d9ee45d5f360e2ad0834d6411fcf4a243796cee6c2.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 309, + 82, + 369, + 143 + ], + "blocks": [ + { + "bbox": [ + 371, + 70, + 430, + 78 + ], + "lines": [ + { + "bbox": [ + 371, + 70, + 430, + 78 + ], + "spans": [ + { + "bbox": [ + 371, + 70, + 430, + 78 + ], + "type": "text", + "content": "Follow-your-emoji" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 309, + 82, + 369, + 143 + ], + "lines": [ + { + "bbox": [ + 309, + 82, + 369, + 143 + ], + "spans": [ + { + "bbox": [ + 309, + 82, + 369, + 143 + ], + "type": "image", + "image_path": "bf16d72210ca759e0af88f6bedaa2e36764141e0b8149cdc081b8a143429d3b3.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 309, + 144, + 369, + 205 + ], + "blocks": [ + { + "bbox": [ + 309, + 144, + 369, + 205 + ], + "lines": [ + { + "bbox": [ + 309, + 144, + 369, + 205 + ], + "spans": [ + { + "bbox": [ + 309, + 144, + 369, + 205 + ], + "type": "image", + "image_path": "e26af30bea92d49287fafb895755257f60cc0ed8b17faa2f15bf604059e6d859.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 370, + 82, + 430, + 143 + ], + "blocks": [ + { + "bbox": [ + 322, + 70, + 356, + 77 + ], + "lines": [ + { + "bbox": [ + 322, + 70, + 356, + 77 + ], + "spans": [ + { + "bbox": [ + 322, + 70, + 356, + 77 + ], + "type": "text", + "content": "X-Portrait" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 370, + 82, + 430, + 143 + ], + "lines": [ + { + "bbox": [ + 370, + 82, + 430, + 143 + ], + "spans": [ + { + "bbox": [ + 370, + 82, + 430, + 143 + ], + "type": "image", + "image_path": "1d62fa7534ab4fdadf1a2526177418b6ce516889c0219eb2482ed771aebd879d.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 370, + 144, + 430, + 205 + ], + "blocks": [ + { + "bbox": [ + 370, + 144, + 430, + 205 + ], + "lines": [ + { + "bbox": [ + 370, + 144, + 430, + 205 + ], + "spans": [ + { + "bbox": [ + 370, + 144, + 430, + 205 + ], + "type": "image", + "image_path": "c11f77920000cbb00879dd7af84f18dbb75b756b22d41c3d2fdb402f10161ed4.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 430, + 82, + 490, + 143 + ], + "blocks": [ + { + "bbox": [ + 452, + 70, + 468, + 77 + ], + "lines": [ + { + "bbox": [ + 452, + 70, + 468, + 77 + ], + "spans": [ + { + "bbox": [ + 452, + 70, + 468, + 77 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 430, + 82, + 490, + 143 + ], + "lines": [ + { + "bbox": [ + 430, + 82, + 490, + 143 + ], + "spans": [ + { + "bbox": [ + 430, + 82, + 490, + 143 + ], + "type": "image", + "image_path": "819ee2f58c22cdd2813a13c851fdebb7c98e8715291a0e4283fde975dcdc8a42.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 430, + 144, + 490, + 205 + ], + "blocks": [ + { + "bbox": [ + 430, + 144, + 490, + 205 + ], + "lines": [ + { + "bbox": [ + 430, + 144, + 490, + 205 + ], + "spans": [ + { + "bbox": [ + 430, + 144, + 490, + 205 + ], + "type": "image", + "image_path": "a6a96bb2f37ab399be0d1ce40d8efd7af30a48215e39c6a11d76081493698387.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 491, + 82, + 552, + 143 + ], + "blocks": [ + { + "bbox": [ + 515, + 70, + 527, + 77 + ], + "lines": [ + { + "bbox": [ + 515, + 70, + 527, + 77 + ], + "spans": [ + { + "bbox": [ + 515, + 70, + 527, + 77 + ], + "type": "text", + "content": "G.T." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 491, + 82, + 552, + 143 + ], + "lines": [ + { + "bbox": [ + 491, + 82, + 552, + 143 + ], + "spans": [ + { + "bbox": [ + 491, + 82, + 552, + 143 + ], + "type": "image", + "image_path": "9d1455dc7359830834edeffff48465822150ba616f41e64e8b7f867a199aadd3.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 491, + 144, + 552, + 205 + ], + "blocks": [ + { + "bbox": [ + 491, + 144, + 552, + 205 + ], + "lines": [ + { + "bbox": [ + 491, + 144, + 552, + 205 + ], + "spans": [ + { + "bbox": [ + 491, + 144, + 552, + 205 + ], + "type": "image", + "image_path": "9c34e39dcffc4fb317309105cfb4bd12883a051394dd6d19fb15af1d5a0d02d5.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 59, + 207, + 120, + 267 + ], + "blocks": [ + { + "bbox": [ + 59, + 207, + 120, + 267 + ], + "lines": [ + { + "bbox": [ + 59, + 207, + 120, + 267 + ], + "spans": [ + { + "bbox": [ + 59, + 207, + 120, + 267 + ], + "type": "image", + "image_path": "8018ac97d1feb9051449f7bae0cf8d596aac1e0b583f0eb578465243a7ac1335.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 59, + 269, + 120, + 329 + ], + "blocks": [ + { + "bbox": [ + 59, + 269, + 120, + 329 + ], + "lines": [ + { + "bbox": [ + 59, + 269, + 120, + 329 + ], + "spans": [ + { + "bbox": [ + 59, + 269, + 120, + 329 + ], + "type": "image", + "image_path": "082ef2c299f773c6a8c95823377446841f38d2387ae9d48e6fa6fb5ae93a1970.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 330, + 555, + 363 + ], + "lines": [ + { + "bbox": [ + 55, + 330, + 555, + 363 + ], + "spans": [ + { + "bbox": [ + 55, + 330, + 555, + 363 + ], + "type": "text", + "content": "Figure 4. Talking head results. We show results from previous work and our method. Our method generates results with sharp appearance and closely aligned expression with the ground truth. Note Portrait4d-v2 uses a customized camera space which is non-trivial to render GT camera aligned images." + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 128, + 206, + 188, + 267 + ], + "blocks": [ + { + "bbox": [ + 128, + 206, + 188, + 267 + ], + "lines": [ + { + "bbox": [ + 128, + 206, + 188, + 267 + ], + "spans": [ + { + "bbox": [ + 128, + 206, + 188, + 267 + ], + "type": "image", + "image_path": "62202f3ee1b0a73439096147aa888cb2af8f74743e780b1d8b97c33a7a1e2d59.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 128, + 269, + 188, + 328 + ], + "blocks": [ + { + "bbox": [ + 128, + 269, + 188, + 328 + ], + "lines": [ + { + "bbox": [ + 128, + 269, + 188, + 328 + ], + "spans": [ + { + "bbox": [ + 128, + 269, + 188, + 328 + ], + "type": "image", + "image_path": "4b3d4bfcd46a86e9ce83ad3279312c5330806981d53a613c5ed8c1a8155444b5.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 189, + 206, + 249, + 267 + ], + "blocks": [ + { + "bbox": [ + 189, + 206, + 249, + 267 + ], + "lines": [ + { + "bbox": [ + 189, + 206, + 249, + 267 + ], + "spans": [ + { + "bbox": [ + 189, + 206, + 249, + 267 + ], + "type": "image", + "image_path": "fd1c3105532f0c6ef3831eaeb5f3abff33b9dc4c6f8b246af675f641b129cca5.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 189, + 269, + 249, + 328 + ], + "blocks": [ + { + "bbox": [ + 189, + 269, + 249, + 328 + ], + "lines": [ + { + "bbox": [ + 189, + 269, + 249, + 328 + ], + "spans": [ + { + "bbox": [ + 189, + 269, + 249, + 328 + ], + "type": "image", + "image_path": "f09e6d859d68de1c1f35ec30ff3a903b5c2d2ae188271eabc34e315fbc8acb79.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 250, + 206, + 309, + 267 + ], + "blocks": [ + { + "bbox": [ + 250, + 206, + 309, + 267 + ], + "lines": [ + { + "bbox": [ + 250, + 206, + 309, + 267 + ], + "spans": [ + { + "bbox": [ + 250, + 206, + 309, + 267 + ], + "type": "image", + "image_path": "14c3b381bddb55639298d9d9ed44faaacbaaf50321c41ef58b281a8335951386.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 250, + 269, + 309, + 328 + ], + "blocks": [ + { + "bbox": [ + 250, + 269, + 309, + 328 + ], + "lines": [ + { + "bbox": [ + 250, + 269, + 309, + 328 + ], + "spans": [ + { + "bbox": [ + 250, + 269, + 309, + 328 + ], + "type": "image", + "image_path": "0d6e38aa068c015d03add82d7db30276cb1635c26b6f7248592b44cb2c27e84d.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 309, + 206, + 369, + 267 + ], + "blocks": [ + { + "bbox": [ + 309, + 206, + 369, + 267 + ], + "lines": [ + { + "bbox": [ + 309, + 206, + 369, + 267 + ], + "spans": [ + { + "bbox": [ + 309, + 206, + 369, + 267 + ], + "type": "image", + "image_path": "3270e05dad7b509d8839c76b065adcb5f12ec1faecdec2fc6dc2511bef5aa715.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 309, + 269, + 369, + 328 + ], + "blocks": [ + { + "bbox": [ + 309, + 269, + 369, + 328 + ], + "lines": [ + { + "bbox": [ + 309, + 269, + 369, + 328 + ], + "spans": [ + { + "bbox": [ + 309, + 269, + 369, + 328 + ], + "type": "image", + "image_path": "74b2d41c0208482947589e9817352b004e185e0a72df7246498240873719f934.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 370, + 206, + 430, + 267 + ], + "blocks": [ + { + "bbox": [ + 370, + 206, + 430, + 267 + ], + "lines": [ + { + "bbox": [ + 370, + 206, + 430, + 267 + ], + "spans": [ + { + "bbox": [ + 370, + 206, + 430, + 267 + ], + "type": "image", + "image_path": "9b3debf6ba0ffd62bcfa5ce98b55734ecc1da9b30855356c79a96c676fe95d99.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 370, + 269, + 430, + 328 + ], + "blocks": [ + { + "bbox": [ + 370, + 269, + 430, + 328 + ], + "lines": [ + { + "bbox": [ + 370, + 269, + 430, + 328 + ], + "spans": [ + { + "bbox": [ + 370, + 269, + 430, + 328 + ], + "type": "image", + "image_path": "72b284bbe4bcfc591cc3121a2769582a283f07228b53a006c83f3cf2ed1e119e.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 430, + 206, + 490, + 267 + ], + "blocks": [ + { + "bbox": [ + 430, + 206, + 490, + 267 + ], + "lines": [ + { + "bbox": [ + 430, + 206, + 490, + 267 + ], + "spans": [ + { + "bbox": [ + 430, + 206, + 490, + 267 + ], + "type": "image", + "image_path": "053a66fc000a54e5cc345000c96162b2c5a6ae88c65651b93e847095cbcb1227.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 430, + 269, + 490, + 328 + ], + "blocks": [ + { + "bbox": [ + 430, + 269, + 490, + 328 + ], + "lines": [ + { + "bbox": [ + 430, + 269, + 490, + 328 + ], + "spans": [ + { + "bbox": [ + 430, + 269, + 490, + 328 + ], + "type": "image", + "image_path": "27dfd1e2dd4a220d94eede647ae90873a55063beb59eb237e004346675ddd35f.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 491, + 206, + 552, + 267 + ], + "blocks": [ + { + "bbox": [ + 491, + 206, + 552, + 267 + ], + "lines": [ + { + "bbox": [ + 491, + 206, + 552, + 267 + ], + "spans": [ + { + "bbox": [ + 491, + 206, + 552, + 267 + ], + "type": "image", + "image_path": "7f9e61f023e45e723af21658a2fff36d64603374f16d5dc5e1930f961ff22088.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 491, + 269, + 552, + 328 + ], + "blocks": [ + { + "bbox": [ + 491, + 269, + 552, + 328 + ], + "lines": [ + { + "bbox": [ + 491, + 269, + 552, + 328 + ], + "spans": [ + { + "bbox": [ + 491, + 269, + 552, + 328 + ], + "type": "image", + "image_path": "cda0833a21bf6c6b007c20c10de2245250045513c9cf7028c4163f14848c47c8.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "bbox": [ + 55, + 367, + 295, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 367, + 295, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 367, + 295, + 415 + ], + "type": "text", + "content": "sideration both image quality and motion naturalness. This also indicates that a native video diffusion model is stronger than a finetuned image diffusion with added temporal component (like X-Portrait [42] and Follow-your-emoji [29])." + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 55, + 415, + 295, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 295, + 547 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 295, + 547 + ], + "type": "text", + "content": "We also show qualitative results in Fig. 4. FaceV2V [40] gives stable results and follows the input identity, but produces less details due to its low output resolution. EmoPortrait [16] gives sharper results, but suffers from larger identity shifts. Portrait4D-v2 [12] produces good sharpness and less identity shifts, however relatively stiff expressions. The diffusion-based methods Follow-your-emoji [29] and X-portrait [42] render good high-frequency details, but do not precisely follow the input head poses or expressions. In contrast, our method gives good image quality, faithful identity, while follows the input control signal." + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 55, + 552, + 224, + 566 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 224, + 566 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 224, + 566 + ], + "type": "text", + "content": "4.2. Effectiveness for 3D Rendering" + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 55, + 570, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 641 + ], + "type": "text", + "content": "We evaluate the effectiveness on 3D of different methods from the following aspects. First, we consider the task of binocular stereo rendering, which is one of the foundational use cases for VR/AR applications. Also, we compare how well these approaches perform on perspective geometry by moving the camera closer." + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": "Note that Face-V2V [40] and EmoPortrait [16] use orthogonal projection as a simplified camera model, thus they cannot produce any reasonable renderings for stereo pairs (i.e. two views will have the same disparity across the whole image) and moving camera closer (i.e. getting identical renderings). Follow-your-emoji [29] and X-xportrait [42] only" + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 313, + 367, + 555, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 367, + 555, + 427 + ], + "spans": [ + { + "bbox": [ + 313, + 367, + 555, + 427 + ], + "type": "text", + "content": "use detected landmarks and images as the driving signal but do not allow any explicit control on the camera, thus they cannot perform these two 3D tasks neither. As a result, we only show the renderings from Portrait4D-v2 [12] and ours, where truth 3D rendering is supported." + } + ] + } + ], + "index": 46 + }, + { + "bbox": [ + 313, + 427, + 555, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 427, + 555, + 522 + ], + "spans": [ + { + "bbox": [ + 313, + 427, + 555, + 522 + ], + "type": "text", + "content": "Binocular Stereo Rendering. We render stereo pairs with a 5cm baseline looking from 1m in front of the talking head, and run an off-the-shell stereo matching method [27] to get the disparity map. In Fig. 5, ours gives stereo pairs with reasonable estimated disparities, and is competitive with the NeRF-based Portrait4D-v2 [12], which acts as a strong baseline on 3D effects, while being one order of magnitude faster in novel view rendering (Tab. 1)." + } + ] + } + ], + "index": 47 + }, + { + "bbox": [ + 313, + 523, + 554, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 523, + 554, + 606 + ], + "spans": [ + { + "bbox": [ + 313, + 523, + 554, + 606 + ], + "type": "text", + "content": "Perspective Geometry. To demonstrate the perspective effect, we render a close view by moving the camera closer while reducing camera focal length to keep the whole head in view. As in Fig. 5, ours produces good perspective distortions that is comparable to the NeRF-based Portrait4Dv2 [12], while being more faithful to the reference identity shown in Fig. 4 second row." + } + ] + } + ], + "index": 48 + }, + { + "bbox": [ + 313, + 613, + 378, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 613, + 378, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 613, + 378, + 624 + ], + "type": "text", + "content": "4.3. Ablation" + } + ] + } + ], + "index": 49 + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "content": "MPI as representation. Our diffusion model generates MPI frames, where each of them has a frontal, a residual, and a set of planes with alpha. While the frontal contains most of the appearance, we hereby evaluate the importance of the residual image. We train our model producing a MPI with only a frontal image, and reuse it for all the planes for alpha composition. While the metrics on frontal view" + } + ] + } + ], + "index": 50 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21113" + } + ] + } + ], + "index": 51 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 79, + 119, + 137 + ], + "blocks": [ + { + "bbox": [ + 101, + 71, + 139, + 79 + ], + "lines": [ + { + "bbox": [ + 101, + 71, + 139, + 79 + ], + "spans": [ + { + "bbox": [ + 101, + 71, + 139, + 79 + ], + "type": "text", + "content": "Portrait4d" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 61, + 79, + 119, + 137 + ], + "lines": [ + { + "bbox": [ + 61, + 79, + 119, + 137 + ], + "spans": [ + { + "bbox": [ + 61, + 79, + 119, + 137 + ], + "type": "image", + "image_path": "6194596807da3763c55474691f13be5355903e48a9f6c9faa49e9689229dab22.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 76, + 137, + 104, + 146 + ], + "lines": [ + { + "bbox": [ + 76, + 137, + 104, + 146 + ], + "spans": [ + { + "bbox": [ + 76, + 137, + 104, + 146 + ], + "type": "text", + "content": "original" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 120, + 79, + 179, + 137 + ], + "blocks": [ + { + "bbox": [ + 120, + 79, + 179, + 137 + ], + "lines": [ + { + "bbox": [ + 120, + 79, + 179, + 137 + ], + "spans": [ + { + "bbox": [ + 120, + 79, + 179, + 137 + ], + "type": "image", + "image_path": "2f077548f90113150eef8303a6fb466fae0942714cbde3314480a59f18ca3302.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 138, + 137, + 162, + 145 + ], + "lines": [ + { + "bbox": [ + 138, + 137, + 162, + 145 + ], + "spans": [ + { + "bbox": [ + 138, + 137, + 162, + 145 + ], + "type": "text", + "content": "stereo" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 183, + 79, + 242, + 137 + ], + "blocks": [ + { + "bbox": [ + 235, + 72, + 251, + 79 + ], + "lines": [ + { + "bbox": [ + 235, + 72, + 251, + 79 + ], + "spans": [ + { + "bbox": [ + 235, + 72, + 251, + 79 + ], + "type": "text", + "content": "ours" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 183, + 79, + 242, + 137 + ], + "lines": [ + { + "bbox": [ + 183, + 79, + 242, + 137 + ], + "spans": [ + { + "bbox": [ + 183, + 79, + 242, + 137 + ], + "type": "image", + "image_path": "7271bb85012a4af6de2a3a62bcfbf31b92bd667b80f115870cc59e34b74bfbbf.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 198, + 137, + 227, + 145 + ], + "lines": [ + { + "bbox": [ + 198, + 137, + 227, + 145 + ], + "spans": [ + { + "bbox": [ + 198, + 137, + 227, + 145 + ], + "type": "text", + "content": "original" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 244, + 79, + 304, + 137 + ], + "blocks": [ + { + "bbox": [ + 244, + 79, + 304, + 137 + ], + "lines": [ + { + "bbox": [ + 244, + 79, + 304, + 137 + ], + "spans": [ + { + "bbox": [ + 244, + 79, + 304, + 137 + ], + "type": "image", + "image_path": "8689c874ca36f49b83a75295bb652f024cfa28b1e390dacf8497298847e41bc4.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 260, + 137, + 283, + 145 + ], + "lines": [ + { + "bbox": [ + 260, + 137, + 283, + 145 + ], + "spans": [ + { + "bbox": [ + 260, + 137, + 283, + 145 + ], + "type": "text", + "content": "stereo" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 61, + 148, + 119, + 206 + ], + "blocks": [ + { + "bbox": [ + 61, + 148, + 119, + 206 + ], + "lines": [ + { + "bbox": [ + 61, + 148, + 119, + 206 + ], + "spans": [ + { + "bbox": [ + 61, + 148, + 119, + 206 + ], + "type": "image", + "image_path": "f7e2513c9298989e9ff4f8e54b4f3f6c2cd78d749a124f34523585a24994a9e2.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 71, + 207, + 110, + 215 + ], + "lines": [ + { + "bbox": [ + 71, + 207, + 110, + 215 + ], + "spans": [ + { + "bbox": [ + 71, + 207, + 110, + 215 + ], + "type": "text", + "content": "close view" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 120, + 148, + 179, + 206 + ], + "blocks": [ + { + "bbox": [ + 120, + 148, + 179, + 206 + ], + "lines": [ + { + "bbox": [ + 120, + 148, + 179, + 206 + ], + "spans": [ + { + "bbox": [ + 120, + 148, + 179, + 206 + ], + "type": "image", + "image_path": "d77810a3249f6029a9324027de75663dea15eff25c1ad28161d17a3469ba3b0b.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 134, + 207, + 167, + 216 + ], + "lines": [ + { + "bbox": [ + 134, + 207, + 167, + 216 + ], + "spans": [ + { + "bbox": [ + 134, + 207, + 167, + 216 + ], + "type": "text", + "content": "disparity" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 183, + 148, + 242, + 207 + ], + "blocks": [ + { + "bbox": [ + 183, + 148, + 242, + 207 + ], + "lines": [ + { + "bbox": [ + 183, + 148, + 242, + 207 + ], + "spans": [ + { + "bbox": [ + 183, + 148, + 242, + 207 + ], + "type": "image", + "image_path": "4be5362c1939e65175a089c41581c72f1590ff05e9ce3db2597b23f1a790888e.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 193, + 207, + 230, + 215 + ], + "lines": [ + { + "bbox": [ + 193, + 207, + 230, + 215 + ], + "spans": [ + { + "bbox": [ + 193, + 207, + 230, + 215 + ], + "type": "text", + "content": "close view" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 244, + 148, + 304, + 207 + ], + "blocks": [ + { + "bbox": [ + 244, + 148, + 304, + 207 + ], + "lines": [ + { + "bbox": [ + 244, + 148, + 304, + 207 + ], + "spans": [ + { + "bbox": [ + 244, + 148, + 304, + 207 + ], + "type": "image", + "image_path": "d92ce221da6afc2556b3e2e514747bf2b0b4c6d2285333e7f2d19490e7ab0b7c.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 255, + 207, + 289, + 216 + ], + "lines": [ + { + "bbox": [ + 255, + 207, + 289, + 216 + ], + "spans": [ + { + "bbox": [ + 255, + 207, + 289, + 216 + ], + "type": "text", + "content": "disparity" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 309, + 79, + 367, + 137 + ], + "blocks": [ + { + "bbox": [ + 347, + 71, + 384, + 79 + ], + "lines": [ + { + "bbox": [ + 347, + 71, + 384, + 79 + ], + "spans": [ + { + "bbox": [ + 347, + 71, + 384, + 79 + ], + "type": "text", + "content": "Portrait4d" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 309, + 79, + 367, + 137 + ], + "lines": [ + { + "bbox": [ + 309, + 79, + 367, + 137 + ], + "spans": [ + { + "bbox": [ + 309, + 79, + 367, + 137 + ], + "type": "image", + "image_path": "5fc468d1ae9ea3a518f9e5b2836af06a23cfc45a8699ffe48d5ffec0bea91f1e.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 137, + 350, + 145 + ], + "lines": [ + { + "bbox": [ + 321, + 137, + 350, + 145 + ], + "spans": [ + { + "bbox": [ + 321, + 137, + 350, + 145 + ], + "type": "text", + "content": "original" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 368, + 79, + 427, + 137 + ], + "blocks": [ + { + "bbox": [ + 368, + 79, + 427, + 137 + ], + "lines": [ + { + "bbox": [ + 368, + 79, + 427, + 137 + ], + "spans": [ + { + "bbox": [ + 368, + 79, + 427, + 137 + ], + "type": "image", + "image_path": "4ab73827595ad5c269d7aa175ad5ff213671a620d86511d38c4c720798cb6e82.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 384, + 137, + 407, + 145 + ], + "lines": [ + { + "bbox": [ + 384, + 137, + 407, + 145 + ], + "spans": [ + { + "bbox": [ + 384, + 137, + 407, + 145 + ], + "type": "text", + "content": "stereo" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 432, + 79, + 491, + 137 + ], + "blocks": [ + { + "bbox": [ + 480, + 72, + 497, + 79 + ], + "lines": [ + { + "bbox": [ + 480, + 72, + 497, + 79 + ], + "spans": [ + { + "bbox": [ + 480, + 72, + 497, + 79 + ], + "type": "text", + "content": "ours" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 432, + 79, + 491, + 137 + ], + "lines": [ + { + "bbox": [ + 432, + 79, + 491, + 137 + ], + "spans": [ + { + "bbox": [ + 432, + 79, + 491, + 137 + ], + "type": "image", + "image_path": "a706d7d8f16f9252adcb13dd0a3aca2fcb2689deac8606b0ef1035f071f4596a.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 444, + 137, + 472, + 146 + ], + "lines": [ + { + "bbox": [ + 444, + 137, + 472, + 146 + ], + "spans": [ + { + "bbox": [ + 444, + 137, + 472, + 146 + ], + "type": "text", + "content": "original" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 493, + 79, + 550, + 137 + ], + "blocks": [ + { + "bbox": [ + 493, + 79, + 550, + 137 + ], + "lines": [ + { + "bbox": [ + 493, + 79, + 550, + 137 + ], + "spans": [ + { + "bbox": [ + 493, + 79, + 550, + 137 + ], + "type": "image", + "image_path": "06c43bfda5347cd84dde2db9ab0725cafe168af9adff40f58aac97e7c1aa96c0.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 506, + 137, + 529, + 145 + ], + "lines": [ + { + "bbox": [ + 506, + 137, + 529, + 145 + ], + "spans": [ + { + "bbox": [ + 506, + 137, + 529, + 145 + ], + "type": "text", + "content": "stereo" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 309, + 148, + 368, + 206 + ], + "blocks": [ + { + "bbox": [ + 309, + 148, + 368, + 206 + ], + "lines": [ + { + "bbox": [ + 309, + 148, + 368, + 206 + ], + "spans": [ + { + "bbox": [ + 309, + 148, + 368, + 206 + ], + "type": "image", + "image_path": "817280bcaf7387ae19f5f793e5d4ec39250fb286b1f2d1abca04c533f34359f1.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 207, + 358, + 215 + ], + "lines": [ + { + "bbox": [ + 320, + 207, + 358, + 215 + ], + "spans": [ + { + "bbox": [ + 320, + 207, + 358, + 215 + ], + "type": "text", + "content": "close view" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 369, + 148, + 427, + 207 + ], + "blocks": [ + { + "bbox": [ + 369, + 148, + 427, + 207 + ], + "lines": [ + { + "bbox": [ + 369, + 148, + 427, + 207 + ], + "spans": [ + { + "bbox": [ + 369, + 148, + 427, + 207 + ], + "type": "image", + "image_path": "20134c948b0e954bc6f27bab96dd28ef6158162b9cdc6bd7d1ff1443f9e91e0a.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 380, + 207, + 413, + 216 + ], + "lines": [ + { + "bbox": [ + 380, + 207, + 413, + 216 + ], + "spans": [ + { + "bbox": [ + 380, + 207, + 413, + 216 + ], + "type": "text", + "content": "disparity" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 432, + 148, + 491, + 207 + ], + "blocks": [ + { + "bbox": [ + 432, + 148, + 491, + 207 + ], + "lines": [ + { + "bbox": [ + 432, + 148, + 491, + 207 + ], + "spans": [ + { + "bbox": [ + 432, + 148, + 491, + 207 + ], + "type": "image", + "image_path": "cede3950382761297e365b735eda01711768badaf49c25639628209f8d98a4e0.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 441, + 207, + 479, + 215 + ], + "lines": [ + { + "bbox": [ + 441, + 207, + 479, + 215 + ], + "spans": [ + { + "bbox": [ + 441, + 207, + 479, + 215 + ], + "type": "text", + "content": "close view" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 492, + 148, + 550, + 207 + ], + "blocks": [ + { + "bbox": [ + 492, + 148, + 550, + 207 + ], + "lines": [ + { + "bbox": [ + 492, + 148, + 550, + 207 + ], + "spans": [ + { + "bbox": [ + 492, + 148, + 550, + 207 + ], + "type": "image", + "image_path": "db0e4f4da297a193b7040f5cdab4b8cc9ace940530aa815e1182d929033db986.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 502, + 207, + 534, + 216 + ], + "lines": [ + { + "bbox": [ + 502, + 207, + 534, + 216 + ], + "spans": [ + { + "bbox": [ + 502, + 207, + 534, + 216 + ], + "type": "text", + "content": "disparity" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 59, + 266, + 293, + 440 + ], + "blocks": [ + { + "bbox": [ + 55, + 217, + 555, + 262 + ], + "lines": [ + { + "bbox": [ + 55, + 217, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 555, + 262 + ], + "type": "text", + "content": "Figure 5. Binocular stereo and perspective effects. Our method generates MPI videos that enables efficient spatial rendering, e.g. binocular stereo and perspective effects. We render stereo pairs and calculate a disparity using RAFT-Stereo [27] to visualize the perceived geometry. To demonstrate perspective effects, we render a view with camera moving closer to the subject. Our method renders visually plausible and comparable 3D effects with NeRF based approach like Portrait4d-v2, while being orders of magnitude faster." + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 266, + 293, + 440 + ], + "lines": [ + { + "bbox": [ + 59, + 266, + 293, + 440 + ], + "spans": [ + { + "bbox": [ + 59, + 266, + 293, + 440 + ], + "type": "image", + "image_path": "f912fddce4e25d30ea339c9a9eb7fa0d1a4db027fb55e56a7378d05cd57af512.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 444, + 295, + 510 + ], + "lines": [ + { + "bbox": [ + 55, + 444, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 444, + 295, + 510 + ], + "type": "text", + "content": "Figure 6. Effect of residual RGB. We retrain our model with only a single frontal image in the MPI. We show both the rendered side view images and the disparity converted from the MPI alpha in frontal view. We found this significantly reduces the capacity of MPI, which leads to noisy and irregular geometry and blatant blurriness near the occlusion boundaries." + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + } + ], + "index": 37 + }, + { + "bbox": [ + 55, + 519, + 295, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 519, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 519, + 295, + 579 + ], + "type": "text", + "content": "become even stronger compared to the full model (1 (Ours w/o residual)), the model produces much worse geometry (Fig. 6), presumable due to the missing of capacity of handling view dependent effects. It also leads to obvious blurriness in the boundary when rendering from the side view." + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 55, + 580, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 580, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 580, + 295, + 616 + ], + "type": "text", + "content": "A further degenerated case is to use a RGB image and depth for novel-view rendering, and please check our supp for more comparison." + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": "Late-stage noise sampling. Late-stage noise sampling is also important to learn sharp images since that significantly reduces the negative impact from the imperfect pseudo ground truth. To verify this, we train a model with full noise sampling, and show the quantitative and qualitative comparison in Tab. 1 and Fig. 7. Without large noise sampling, the diffusion model still produces reasonable structure but missing details, e.g. wrinkles, mold, hair strand." + } + ] + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 318, + 266, + 552, + 411 + ], + "blocks": [ + { + "bbox": [ + 318, + 266, + 552, + 411 + ], + "lines": [ + { + "bbox": [ + 318, + 266, + 552, + 411 + ], + "spans": [ + { + "bbox": [ + 318, + 266, + 552, + 411 + ], + "type": "image", + "image_path": "ae234078cb16a5a801a65afdbac80c7ebd1ec1c52c1370d9ed343f6fec06426e.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 415, + 555, + 460 + ], + "lines": [ + { + "bbox": [ + 313, + 415, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 555, + 460 + ], + "type": "text", + "content": "Figure 7. Effect of Late-stage noise sampling. We show a comparison between our model trained with and without the Late-stage noise sampling. Training with regular noise sampling leads to excessive blurriness and lost of details." + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + } + ], + "index": 42 + }, + { + "bbox": [ + 314, + 473, + 387, + 485 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 473, + 387, + 485 + ], + "spans": [ + { + "bbox": [ + 314, + 473, + 387, + 485 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 313, + 498, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 714 + ], + "type": "text", + "content": "We present a method to learn a 3D-aware video diffusion model, which generates photorealistic talking head multi-plane images given an identity reference image and a sequence of tracking signals. We show such a model can be learned directly from large scale monocular videos without multi-view datasets. With appropriate learning strategies, both 3D information and high quality image generation can be learned concurrently. Extensive experiments show that, our method generates not only talking head video with quality comparable and even better than existing arts but also 3D rendering efficiently that support spatial viewing experience, like in VR headset. As limitations, the generated talking head video does not support rendering from excessive camera viewpoint changes, which is a common issue of the MPI representation. Extending our framework with other efficient 3D representations or enhancing training with multi-view dataset could be a promising direction to improve both the image and 3D quality." + } + ] + } + ], + "index": 45 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21114" + } + ] + } + ], + "index": 46 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "type": "text", + "content": "[1] Titas Anciukevicius, Zexiang Xu, Matthew Fisher, Paul Henderson, Hakan Bilen, Niloy J Mitra, and Paul Guerrero. Renderdiffusion: Image diffusion for 3d reconstruction, inpainting and generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12608-12618, 2023. 3, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 158, + 296, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 158, + 296, + 212 + ], + "spans": [ + { + "bbox": [ + 61, + 158, + 296, + 212 + ], + "type": "text", + "content": "[2] ShahRukh Athar, Zexiang Xu, Kalyan Sunkavalli, Eli Shechtman, and Zhixin Shu. Rignerf: Fully controllable neural 3d portraits. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 20364-20373, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 213, + 296, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 213, + 296, + 289 + ], + "spans": [ + { + "bbox": [ + 62, + 213, + 296, + 289 + ], + "type": "text", + "content": "[3] Ziqian Bai, Feitong Tan, Zeng Huang, Kripasindhu Sarkar, Danhang Tang, Di Qiu, Abhimitra Meka, Ruofei Du, Mingsong Dou, Sergio Orts-Escalano, et al. Learning personalized high quality volumetric head avatars from monocular rgb videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16890-16900, 2023." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 291, + 295, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 291, + 295, + 356 + ], + "spans": [ + { + "bbox": [ + 62, + 291, + 295, + 356 + ], + "type": "text", + "content": "[4] Ziqian Bai, Feitong Tan, Sean Fanello, Rohit Pandey, Mingsong Dou, Shichen Liu, Ping Tan, and Yinda Zhang. Efficient 3d implicit head avatar with mesh-anchored hash table blendshapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1975–1984, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 357, + 295, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 357, + 295, + 412 + ], + "spans": [ + { + "bbox": [ + 62, + 357, + 295, + 412 + ], + "type": "text", + "content": "[5] Omer Bar-Tal, Hila Chefer, Omer Tov, Charles Herrmann, Roni Paiss, Shiran Zada, Ariel Ephrat, Junhwa Hur, Guanghui Liu, Amit Raj, et al. Lumiere: A spacetime diffusion model for video generation. arXiv preprint arXiv:2401.12945, 2024. 4, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 414, + 295, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 414, + 295, + 446 + ], + "spans": [ + { + "bbox": [ + 62, + 414, + 295, + 446 + ], + "type": "text", + "content": "[6] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 157-164. 2023. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 447, + 295, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 447, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 62, + 447, + 295, + 502 + ], + "type": "text", + "content": "[7] Stella Bounareli, Christos Tzelepis, Vasileios Argyriou, Ioannis Patras, and Georgios Tzimiropoulos. Hyperreenact: One-shot reenactment via jointly learning to refine and retarget faces. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 502, + 295, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 502, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 62, + 502, + 295, + 556 + ], + "type": "text", + "content": "[8] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 558, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 558, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 62, + 558, + 295, + 624 + ], + "type": "text", + "content": "[9] Bindita Chaudhuri, Noranart Vesdapunt, Linda Shapiro, and Baoyuan Wang. Personalized face modeling for improved face reconstruction and motion retargeting. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part V 16, pages 142-160. Springer, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 625, + 295, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 625, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 57, + 625, + 295, + 677 + ], + "type": "text", + "content": "[10] Yufan Chen, Lizhen Wang, Qijing Li, Hongjiang Xiao, Shengping Zhang, Hongxun Yao, and Yebin Liu. Monogaussianavatar: Monocular gaussian point-based head avatar. In ACM SIGGRAPH 2024 Conference Papers, pages 1-9, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 680, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 295, + 713 + ], + "type": "text", + "content": "[11] Yu Deng, Duomin Wang, and Baoyuan Wang. Portrait4d-v2: Pseudo multi-view data creates better 4d head synthesizer. arXiv preprint arXiv:2403.13570, 2024. 3, 6" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[12] Yu Deng, Duomin Wang, and baoyuan Wang. Portrait4d-v2: Pseudo multi-view data creates better 4d head synthesizer. arXiv, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 107, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 161 + ], + "type": "text", + "content": "[13] Zheng Ding, Xuaner Zhang, Zhihao Xia, Lars Jebe, Zhuowen Tu, and Xiuming Zhang. Diffusionrig: Learning personalized priors for facial appearance editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12736-12746, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 162, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 162, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 317, + 162, + 553, + 205 + ], + "type": "text", + "content": "[14] Michail Christos Doukas, Stefanos Zafeiriou, and Viktoriia Sharmanska. Headgan: One-shot neural head synthesis and editing. In Proceedings of the IEEE/CVF International conference on Computer Vision, pages 14398-14407, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 206, + 553, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 206, + 553, + 260 + ], + "spans": [ + { + "bbox": [ + 317, + 206, + 553, + 260 + ], + "type": "text", + "content": "[15] Nikita Drobyshev, Jenya Chelishev, Taras Khakhulin, Aleksei Ivakhnenko, Victor Lempitsky, and Egor Zakharov. Megaportraits: One-shot megapixel neural head avatars. In Proceedings of the 30th ACM International Conference on Multimedia, pages 2663–2671, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 261, + 553, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 553, + 327 + ], + "type": "text", + "content": "[16] Nikita Drobyshev, Antoni Bigata Casademunt, Konstantinos Vougioukas, Zoe Landgraf, Stavros Petridis, and Maja Pantic. Emoportraits: Emotion-enhanced multimodal one-shot head avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8498-8507, 2024. 2, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 327, + 553, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 370 + ], + "type": "text", + "content": "[17] Emilien Dupont, Hyunjik Kim, SM Eslami, Danilo Rezende, and Dan Rosenbaum. From data to functa: Your data point is a function and you can treat it like one. arXiv preprint arXiv:2201.12204, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 371, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 371, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 371, + 553, + 426 + ], + "type": "text", + "content": "[18] Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8649-8658, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 426, + 553, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 553, + 470 + ], + "type": "text", + "content": "[19] Xuan Gao, Chenglai Zhong, Jun Xiang, Yang Hong, Yudong Guo, and Juyong Zhang. Reconstructing personalized semantic facial nerf models from monocular video. ACM Transactions on Graphics (TOG), 41(6):1-12, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 471, + 555, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 471, + 555, + 525 + ], + "spans": [ + { + "bbox": [ + 317, + 471, + 555, + 525 + ], + "type": "text", + "content": "[20] Philip-William Grassal, Malte Prinzler, Titus Leistner, Carsten Rother, Matthias Nießner, and Justus Thies. Neural head avatars from monocular rgb videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18653–18664, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 525, + 553, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 525, + 553, + 570 + ], + "spans": [ + { + "bbox": [ + 317, + 525, + 553, + 570 + ], + "type": "text", + "content": "[21] Jianzhu Guo, Dingyun Zhang, Xiaogiang Liu, Zhizhou Zhong, Yuan Zhang, Pengfei Wan, and Di Zhang. Liveportrait: Efficient portrait animation with stitching and retargeting control. arXiv preprint arXiv:2407.03168, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 571, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 571, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 317, + 571, + 553, + 624 + ], + "type": "text", + "content": "[22] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 624, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 658 + ], + "type": "text", + "content": "[23] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 658, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 690 + ], + "type": "text", + "content": "[24] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "text", + "content": "[25] Yushi Lan, Fangzhou Hong, Shuai Yang, Shangchen Zhou, Xuyi Meng, Bo Dai, Xingang Pan, and Chen Change Loy." + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21115" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "text", + "content": "Ln3diff: Scalable latent neural fields diffusion for speedy 3d generation. In European Conference on Computer Vision, pages 112-130. Springer, 2025. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 295, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 295, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 295, + 139 + ], + "type": "text", + "content": "[26] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph., 36(6):194-1, 2017. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "type": "text", + "content": "[27] Lahav Lipson, Zachary Teed, and Jia Deng. Raft-stereo: Multilevel recurrent field transforms for stereo matching. In 2021 International Conference on 3D Vision (3DV), pages 218-227. IEEE, 2021. 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 184, + 294, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 294, + 247 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 294, + 247 + ], + "type": "text", + "content": "[28] Tao Liu, Chenpeng Du, Shuai Fan, Feilong Chen, and Kai Yu. Diffhub: Person-generic visual dubbing using inpainting renderer with diffusion auto-encoder. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 3630-3634. IEEE, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 250, + 294, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 294, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 294, + 304 + ], + "type": "text", + "content": "[29] Yue Ma, Hongyu Liu, Hongfa Wang, Heng Pan, Yingqing He, Junkun Yuan, Ailing Zeng, Chengfei Cai, Heung-Yeung Shum, Wei Liu, et al. Follow-your-emoji: Fine-controllable and expressive freestyle portrait animation. arXiv preprint arXiv:2406.01900, 2024. 2, 4, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 304, + 294, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 304, + 294, + 358 + ], + "spans": [ + { + "bbox": [ + 56, + 304, + 294, + 358 + ], + "type": "text", + "content": "[30] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 360, + 294, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 360, + 294, + 425 + ], + "spans": [ + { + "bbox": [ + 56, + 360, + 294, + 425 + ], + "type": "text", + "content": "[31] Norman Müller, Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Bulo, Peter Kontschieder, and Matthias Nießner. Diffrf: Rendering-guided 3d radiance field diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4328-4338, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 426, + 294, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 294, + 491 + ], + "type": "text", + "content": "[32] Shenhan Qian, Tobias Kirschstein, Liam Schoneveld, Davide Davoli, Simon Giebenhain, and Matthias Nießner. Gaussianavatars: Photorealistic head avatars with rigged 3d gaus-sians. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20299-20309, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 492, + 294, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 492, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 492, + 294, + 536 + ], + "type": "text", + "content": "[33] Shunsuke Saito, Gabriel Schwartz, Tomas Simon, Junxuan Li, and Giljoo Nam. Relightable gaussian codec avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 130-141, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 537, + 294, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 294, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 294, + 581 + ], + "type": "text", + "content": "[34] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. Advances in neural information processing systems, 32, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 582, + 294, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 294, + 646 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 294, + 646 + ], + "type": "text", + "content": "[35] Ayush Tewari, Tianwei Yin, George Cazenavette, Semon Rezhikov, Josh Tenenbaum, Frédo Durand, Bill Freeman, and Vincent Sitzmann. Diffusion with forward models: Solving stochastic inverse problems without direct supervision. Advances in Neural Information Processing Systems, 36: 12349-12362, 2023. 3, 4, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 647, + 294, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 691 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 691 + ], + "type": "text", + "content": "[36] Richard Tucker and Noah Snavely. Single-view view synthesis with multiplane images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 551-560, 2020. 3, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "text", + "content": "[37] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. To" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 697 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "wards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "text", + "content": "[38] Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, Karsten Kreis, et al. Lion: Latent point diffusion models for 3d shape generation. Advances in Neural Information Processing Systems, 35:10021-10039, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 141, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 206 + ], + "type": "text", + "content": "[39] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, et al. Rodin: A generative model for sculpting 3d digital avatars using diffusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4563-4573, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 208, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 262 + ], + "type": "text", + "content": "[40] Ting-Chun Wang, Arun Mallya, and Ming-Yu Liu. One-shot free-view neural talking-head synthesis for video conferencing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10039-10049, 2021. 2, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 263, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 263, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 263, + 553, + 319 + ], + "type": "text", + "content": "[41] Zhen Wang, Qiangeng Xu, Feitong Tan, Mengei Chai, Shichen Liu, Rohit Pandey, Sean Fanello, Achuta Kadambi, and Yinda Zhang. Mvdd: Multi-view depth diffusion models. In European Conference on Computer Vision, pages 236-253. Springer, 2025. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 319, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 319, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 319, + 553, + 363 + ], + "type": "text", + "content": "[42] You Xie, Hongyi Xu, Guoxian Song, Chao Wang, Yichun Shi, and Linjie Luo. X-portrait: Expressive portrait animation with hierarchical motion attention. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2, 4, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 365, + 553, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 430 + ], + "type": "text", + "content": "[43] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. Styleheat: One-shot high-resolution editable talking face generation via pre-trained stylegan. In European conference on computer vision, pages 85-101. Springer, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 431, + 553, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 431, + 553, + 495 + ], + "spans": [ + { + "bbox": [ + 316, + 431, + 553, + 495 + ], + "type": "text", + "content": "[44] Egor Zakharov, Aleksei Ivakhnenko, Aliaksandra Shysheya, and Victor Lempitsky. Fast bi-layer neural synthesis of one-shot realistic head avatars. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XII 16, pages 524-540. Springer, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 498, + 553, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 498, + 553, + 541 + ], + "spans": [ + { + "bbox": [ + 316, + 498, + 553, + 541 + ], + "type": "text", + "content": "[45] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 543, + 553, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 543, + 553, + 597 + ], + "spans": [ + { + "bbox": [ + 316, + 543, + 553, + 597 + ], + "type": "text", + "content": "[46] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 599, + 553, + 653 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 599, + 553, + 653 + ], + "spans": [ + { + "bbox": [ + 316, + 599, + 553, + 653 + ], + "type": "text", + "content": "[47] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661–3670, 2021. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 654, + 553, + 697 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 654, + 553, + 697 + ], + "spans": [ + { + "bbox": [ + 316, + 654, + 553, + 697 + ], + "type": "text", + "content": "[48] Wojciech Zielonka, Timo Bolkart, and Justus Thies. Instant volumetric head avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4574-4584, 2023. 3" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "21116" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/IM-Zero_ Instance-level Motion Controllable Video Generation in a Zero-shot Manner/a669fc0a-17c9-4001-97e0-7737b0fc24de_content_list.json b/2025/IM-Zero_ Instance-level Motion Controllable Video Generation in a Zero-shot Manner/a669fc0a-17c9-4001-97e0-7737b0fc24de_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9b78c36ac19cc4a7bd774384e4d832971ad9d0c0 --- /dev/null +++ b/2025/IM-Zero_ Instance-level Motion Controllable Video Generation in a Zero-shot Manner/a669fc0a-17c9-4001-97e0-7737b0fc24de_content_list.json @@ -0,0 +1,1960 @@ +[ + { + "type": "text", + "text": "IM-Zero: Instance-level Motion Controllable Video Generation in a Zero-shot Manner", + "text_level": 1, + "bbox": [ + 179, + 130, + 818, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuyang Huang $^{1*}$ , Yabo Chen $^{1*}$ , Li Ding $^{1}$ , Xiaopeng Zhang $^{2\\dagger}$ , Wenrui Dai $^{1\\dagger}$ , Junni Zou $^{1\\dagger}$ , Hongkai Xiong $^{1}$ , Qi Tian $^{2}$ $^{1}$ Shanghai Jiao Tong University, Shanghai, China $^{2}$ Huawei Inc., Shenzhen, China \n{huangyuyang, chenyabo, wangzhongren, daiwenrui, zoujunni, xionghongkai}@sjtu.edu.cn \nzxphistory@gmail.com, tian.qil@huawei.com", + "bbox": [ + 116, + 202, + 883, + 292 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ddf1302253eba6ebfb247c2bd97298b3e8d639c64388c2685df0852a75223205.jpg", + "image_caption": [ + "(a) A person riding a snowmobile down a snowy mountain." + ], + "image_footnote": [], + "bbox": [ + 138, + 301, + 220, + 366 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8720deab3e64e069231ca4b8ac39f369ff2d3bc4db5d1ce641d85042454a9adb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 230, + 320, + 271, + 344 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c19e69c270bd3dd827a7ad24217b440c964ba5972c41ae2d997cfd00739a40af.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 279, + 303, + 362, + 367 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f9c922b45beea473a69c469cce46743cf4cd03b3e395510f4d5b8340718b1a7a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 303, + 444, + 367 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/94861a0f6fbd7d8d9f210d344e24bdb4f48b5ccadddb6b92d56cd71be05bf072.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 447, + 303, + 527, + 367 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4add69d4bcbcf5355b1167a5620c4ed2e0c6277cee9dba226dc972508958ba0c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 529, + 303, + 612, + 367 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c3c320c402799d3ca87ba4f23b2455f88a44baea6db7b6c2ae28aa18323ee175.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 614, + 303, + 694, + 367 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cfa9cfeade01f1bf9b8b3793dc427b22403aeff2f25ee3024b74293740eab982.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 303, + 777, + 367 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/994cd997a8ce11e979216a9bb83571de2b21ad98f7b49e6367c6231f63367379.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 779, + 303, + 862, + 367 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/751983bbda54f3545057a0fe6f37bcda25d2012fa5b99865d0baac58aa5a3af0.jpg", + "image_caption": [ + "(b) A chef with a white hat preparing a gourmet meal." + ], + "image_footnote": [], + "bbox": [ + 138, + 380, + 222, + 445 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d0f42cdbfd9cac8254a7da54da51b0b2eeb091ed5ea3d6cfe48824531add0b71.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 230, + 400, + 271, + 424 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/eec11408165b12e44f08c8405468b482b7cecbcd12a2fdaab0b871d9260dde5d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 279, + 380, + 362, + 443 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d3b730321a091d650ea5372a859eb766bc5d3b562cf9f478aa106b0ffe925abe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 380, + 444, + 443 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8023b8eaa4ced1db4f86bdc0c490bdf103eb093f853227682274d01b00fc2a11.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 447, + 380, + 527, + 443 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/df09c6f36ec505e9d54a7a50f838df7d5a68a0b856cbba7dde0a387b2a8b0f37.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 529, + 380, + 612, + 443 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4e108b5dd1091674a527545dc0da32d3d0556b2ebcbedfe87ec18e6efa74d444.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 614, + 380, + 694, + 443 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f2db678dfc32d0a8603d3547093bec5d13df6c9a8c0e35427d47b81ba10d7208.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 380, + 777, + 443 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/56f713d1351b73f7e499e3ce9fc2d2200489a40d199bbb831756b51fb4456ee1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 779, + 380, + 862, + 443 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/26b41eca45b63bd8c695130f674f01b8d3ea93facc29603a94ddc3daf7f9c720.jpg", + "image_caption": [ + "(c) A snow leopard is walking on the side of a cliff." + ], + "image_footnote": [], + "bbox": [ + 138, + 459, + 220, + 522 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b8b14fab5ed06744e00a6b10345d3c82f6f257f194537c5dff405277aa69a734.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 230, + 481, + 271, + 503 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ab704978e737c61b7242c82a3e201f987a4d4125c050fb0f8aab2231efa59f2d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 279, + 459, + 362, + 522 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b2d0a23fe3c75f4a3b2a1e2c916ad27de42e92de3b8f93c27b01faea61cd9eee.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 459, + 444, + 522 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/13069143961c353f004a316d264531ca0efdaab5df16af5c6c0f37b08bca01fb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 447, + 459, + 527, + 522 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5c047608058f83ba1d346c30e6cdebaf97f82dc4c2a035ff67d3e80484b5d678.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 529, + 459, + 612, + 522 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ae4eec9ce99811c00b66bb2298a5d2d7e3749885df9f0edf73766ccaff47f5bf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 614, + 459, + 694, + 522 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/56b5c5aec26818581508635bec94422b854f827ad3669c3ce6462aa61c91a700.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 459, + 777, + 522 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0f747f8280de2d9dd608531482e32f6f9ddf71ceddc543b6a24ab2a8b7bed826.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 779, + 459, + 862, + 522 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a7c5c0224f9fc4cea4f27532b909aa22f9ba1ba6ac247c47da9693cb3b3e2343.jpg", + "image_caption": [ + "(d) A lemur walking on a xylophone, in a tropical rainforest." + ], + "image_footnote": [], + "bbox": [ + 138, + 536, + 220, + 599 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7ccc0f4ca6dac6e3014b4898e91026ea62a0e883d622bbb6fc81e8ee139d5f21.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 222, + 536, + 303, + 599 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/15627c4a157921f2733760f2014733f111a777a81eae020a3b862c695ddfab69.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 305, + 536, + 387, + 599 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a318956bf9177ef3fd2b5aed094ce1c3b5cb804deac25da1fb6a29d75b93ef7d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 388, + 536, + 470, + 599 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/61fa9e32ac0c0dbf5ce834cbbe3c598b6242cb1c69409092295d66bef9e942fb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 477, + 554, + 517, + 577 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/70c75a3b854464b13d5d2394cec35b8b447a06981cc275350815ae1ac66bca1d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 527, + 536, + 611, + 599 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/181b8d1051f05bacb30f0967b30341b7fc0a3510eae90c0fe6e04a9699e0697e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 614, + 536, + 694, + 599 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7c38996f49cb0016440a54684837cb9765c5fc8e6c26df167d1e6f630871548e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 536, + 777, + 599 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b0254591e78a5349e840a51972a571ef830b059b7d73f37410fad318fa7fe790.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 779, + 536, + 862, + 599 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c6e4913e12352c0c54f7e5b3b19fb17149ed05d06ca61c3c52ea7f1864a7ef4c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 145, + 627, + 202, + 669 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/550195189daa5648d4bba597da08af01c98116d3cad69f35bf35880fb5e96e7a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 230, + 638, + 271, + 660 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/94fb323760576ac2860c9b357aeba66cde2f3a5150b78451f62597c429f11a7d.jpg", + "image_caption": [ + "(e) A snowman in a forest caught in a burst of flames.", + "Figure 1. IM-Zero introduces a zero-shot framework for instance-level motion controllable video generation. IM-Zero allows users to: (a) Control overall instance motion. (b) Control instance subpart motion. (c) Control instance motion and specify instance shape via mask. (d) Perform motion transfer by reference video. (e) Perform high-quality text-to-video generation." + ], + "image_footnote": [], + "bbox": [ + 279, + 626, + 362, + 679 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d05828622d257352665a2b9033966df61fb65749578e567143e73490968b13b2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 619, + 444, + 679 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d44c0c7d45a6d1bdd8e8bfabdd7bbe88df104a1b213e8598c99c23f32a3596dd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 447, + 617, + 527, + 679 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fea318659d0bd51f124ad1fb421dce5436313f2111de241a2425f631a13c7aaf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 529, + 617, + 612, + 679 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/461c6585ef807732b233b444866892bc336966b622e5d3897c09936ddc9266f3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 614, + 617, + 694, + 679 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/31bda9be05576e5ce87ddbb0ae85dcdfdeb4ed3ef8f7b1c7b8975f813a8e582e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 617, + 777, + 679 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a096790f957ad8d8cde071b5e09dcb9e215445ea05b42377c3314b3fde046092.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 779, + 617, + 862, + 679 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 750, + 326, + 765 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Controllability of video generation has been recently concerned in addition to the quality of generated videos. The main challenge to controllable video generation is to synthesize videos based on user-specified instance spatial locations and movement trajectories. However, existing methods suffer from a dilemma between the resource con", + "bbox": [ + 89, + 770, + 482, + 861 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sumption, generation quality, and user controllability. As an efficient alternative to prohibitive training-based video generation, existing zero-shot video generation methods cannot generate high-quality and motion-consistent videos under the control of layouts and movement trajectories. In this paper, we propose a novel zero-shot method named IM-Zero that ameliorates instance-level motion controllable video generation with enhanced control accuracy, motion consistency, and richness of details to address this prob", + "bbox": [ + 511, + 752, + 906, + 888 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*These authors contributed equally to this work.", + "bbox": [ + 107, + 875, + 364, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding authors: Xiaopeng Zhang; Wenrui Dai; Junni Zou.", + "bbox": [ + 109, + 887, + 462, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7265", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "lem. Specifically, we first present a motion generation stage that extracts motion and textural guidance from keyframe candidates from pre-trained grounded text-to-image model to generate the desired coarse motion video. Subsequently, we develop a video refinement stage that injects the motion priors of pre-trained text-to-video models and detail priors of pre-trained text-to-image models into the latents of coarse motion videos to further enhance video motion consistency and richness of details. To our best knowledge, IM-Zero is the first to simultaneously achieve high-quality video generation and allow to control both layouts and movement trajectories in a zero-shot manner. Extensive experiments demonstrate that IM-Zero outperforms existing methods in terms of video quality, inter-frame consistency, and the alignment of location and trajectory. Furthermore, compared with existing methods, IM-Zero enjoys extra advantages of versatility in video generation, including motion control of subparts within instances, finer control of specifying instance shapes via masks, and more difficult tasks of motion transfer for customizing fine-grained motion patterns through reference videos and high-quality text-to-video generation.", + "bbox": [ + 89, + 90, + 485, + 424 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 452, + 222, + 468 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Demands for enhanced controllability of text-to-video (T2V) generation are booming with the increasing quality of generated videos [3, 6, 7, 13-15, 22, 37, 38, 40, 62]. One of the critical problems lie in the precise control over the spatial layout and trajectories of instances within the generated videos. Existing methods fail to achieve fine-grained control by solely relying on text prompts to steer pre-trained T2V models, even employing fundamental T2V models with large-scale pre-training like Sora [4], KLING, and Doubao-PixelDance [53].", + "bbox": [ + 89, + 477, + 482, + 628 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing methods usually extend ControlNet [55] for controllable video generation with frame-by-frame fine-grained control signals, including Canny, depth, and HED signals [9, 16, 36, 56]. To obtain dense control signals, these methods impose high demands on the input end and are limited in practical applications. One approach is to extract control signals from reference videos but it is hard to obtain reference videos perfectly matching the desired motion patterns. The other alternative requires manually crafting the dense control signals. However, it is also non-trivial to craft frame-by-frame control signals without expertise.", + "bbox": [ + 89, + 628, + 482, + 794 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent attempts [20, 29, 32] use spatial locations (i.e., bounding boxes) and motion trajectories of instances, along with corresponding prompts, to create desired videos. The locations and motion trajectories of instances are extracted from real videos and are encoded as additional control conditions for joint training with the video generation models. However, existing training-based methods can only control", + "bbox": [ + 89, + 795, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the movement trajectory [11, 41, 44, 52, 58, 65], since substantial data and training resources are required to simultaneously learn controlling layouts and movement trajectories for video generation.", + "bbox": [ + 511, + 90, + 903, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we develop a novel training-free approach to address this problem. Existing training-free methods are restricted in obtaining accurate location and movement effects by manipulating spatial attention maps and temporal attention maps to generate instances at specified locations [20, 29, 32]. These methods convert location signals in the pixel space provided by users to a latent space with reduced dimension for manipulating attention maps, and inevitably compromise the precision of location control and obtain inaccurate instance placement. Excessive constraints on the attention maps also disrupt the motion prior of video diffusion. Moreover, frame interaction is neglected during the manipulation, which also causes unreasonable motion effects and unnatural generated videos. Additionally, these methods also suffer from generating low-quality videos with choppy motion and insufficient details.", + "bbox": [ + 511, + 154, + 906, + 396 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We find that both the inaccurate instance positioning and the unnatural motion stem from the limitations of attention map manipulation, which is a compromise solution to mitigate the lack of a clear understanding of instance concepts in vanilla latent diffusion models. It introduces constraints to enforce the instances to appear at designated positions. Contrary to attention map manipulation for video generation, instance-level knowledge is usually considered for image generation. Existing grounded T2I models like GLIGEN [25] and InstanceDiffusion [39] encapsulate extensive instance-level knowledge, but they are confined to generating images rather than consistent videos. Thus, we propose to generate coarse motion videos aligning with the user input and exhibit natural motion using instance-level priors in grounded T2I models, and further refine coarse motion videos by harnessing the motion and detail priors of T2V models for motion and detail injection in a zero-shot manner to boost the motion consistency and richness of details.", + "bbox": [ + 511, + 398, + 905, + 671 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we propose a zero-shot and plug-and-play method named IM-Zero that allows to simultaneously control instance layouts and movement trajectories for video generation. IM-Zero consists of two stages, i.e., a motion generation stage that acquires a coarse motion video, and a video refinement stage to enhance the quality of the coarse motion video. In the motion generation stage, spatial location signals (e.g., bounding boxes and masks) are first augmented according to the input trajectory to obtain a sequence of layout control signals. A batch of keyframe candidates with texture consistency is then obtained by inflating and transforming the pre-trained grounded T2I model, and is used to extract motion guidance and textural guidance for generating the desired coarse motion video. In the video refinement stage, we encode the coarse motion video", + "bbox": [ + 511, + 674, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7266", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "into video latents, and inject the motion and detail priors extracted from the pre-trained T2V models into the video latents to improve motion consistency and enhance the richness of details, respectively.", + "bbox": [ + 89, + 90, + 480, + 150 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our contributions are summarized as below.", + "bbox": [ + 109, + 152, + 401, + 165 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel framework named IM-Zero for zero-shot motion controllable video generation. IM-Zero supports various types of layout inputs and enables fine-grained motion control of the instances and their subparts.", + "- We present motion generation that generates keyframe candidates with the layout control capability of grounded T2I models to extract motion and textural guidance for coarse motion video generation.", + "- We develop zero-shot video refinement that injects the motion and detail priors of the pre-trained text-to-video model and text-to-image model into the latents of the coarse motion video to further enhance video motion consistency and richness of details.", + "- We further achieve motion transfer to customize complex motion patterns using reference videos and support high-quality text-to-video generation with only text input beyond instance layout and movement trajectory controls." + ], + "bbox": [ + 91, + 167, + 482, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Extensive experiments demonstrate that IM-Zero outperforms existing methods in terms of video quality, interframe consistency, and the alignment of location and trajectory. Compared with existing methods, IM-Zero enjoys extra advantages of versatility in video generation, including motion control of subparts within instances and finer control with instance shapes specified by masked inputs. Furthermore, it can be adapted to motion transfer tasks to support the customized fine-grained motion patterns through reference videos and can also be adapted to high-quality text-to-video generation.", + "bbox": [ + 89, + 424, + 482, + 589 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 603, + 232, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Grounded Text-to-image Generation. Grounded text-to-image (T2I) models aim to generate images that align with input layout information. Some works specialize in modeling layout information of the bounding box type. Notable examples include GLIGEN [25], InstanceDiffusion [39], BoxDiff [46], MultiDiffusion [1], ReCo [49], Layout-Guidance [8], MIGC [63, 64], and TraDiffusion [43]. We believe that some training-based methods such as InstanceDiffusion [39] have enabled models to grasp the concept of instances. Their capability aligns to some extent with our objective of achieving location control and trajectory control in video generation. However, these methods are limited to generating single images and cannot produce consistent videos. Therefore, we aim to leverage Grounded T2I models to benefit controllable video generation tasks.", + "bbox": [ + 89, + 628, + 482, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Location and Trajectory Control in Video Generation. Diffusion models have demonstrated exceptional capabilities in image generation [45, 61], video generation [56, 57],", + "bbox": [ + 89, + 854, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D generation [10, 27], and other visual tasks [18, 54, 60]. With the improvement in video generation quality [3, 6, 7, 13-15, 22, 37, 38, 40, 47, 62], there is a growing interest in achieving generation controllability, with a particular focus on controlling the spatial location and movement trajectories of instances. Currently, training-based methods can only achieve limited control over trajectory guidance, at the expense of requiring substantial amounts of data and computational resources, with typical examples being DragNUWA [52], DragAnything [44], DragVideo [11], Tora [58], TrackGo [65] and MotionCtrl [41]. In contrast, some training-free methods can concurrently control both spatial location and movement trajectories. For instance, Peekaboo [20] and TrailBlazer [29] enforce constraints on attention maps to position instances at specified locations. However, this approach can lead to issues such as inaccuracies in location and poor motion effects. FreeTraj [32] follows a similar attention map manipulation approach but introduces noise flow control. In addition to the aforementioned challenges, it also introduces difficulties in handling bounding boxes that undergo size changes. Therefore, we propose a zero-shot method that surpasses attention map manipulation to address the aforementioned challenges and enhance video quality and motion consistency.", + "bbox": [ + 511, + 90, + 906, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 486, + 604, + 502 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 513, + 517, + 653, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Grounded Text-to-image Diffusion Models. Grounded text-to-image (T2I) diffusion models [25, 39, 63, 64] are capable of generating instances at user-specified locations by incorporating the layout input as conditions for joint training. InstanceDiffusion [39] supports four types of layout input: bounding box, mask, point, and scribble. It inserts learnable UniFusion blocks between the self-attention and cross-attention layers of the T2I model. UniFusion sample the instance layout as a 2D points $\\mathbf{p}$ sequence and convert it to layout token by $\\mathbf{g} = \\mathrm{MLP}\\left([ \\tau_{\\theta}(\\mathbf{c}), \\gamma(\\mathbf{p}) \\right]$ , where $\\gamma(\\cdot)$ is a Fourier mapping and $\\mathbf{c}$ is the text prompt. Then the layout token is fused with the backbone visual tokens via a learnable masked self-attention layer.", + "bbox": [ + 511, + 546, + 906, + 743 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text-to-video Diffusion Models. Significant progress has also been made in Text-to-Video (T2V) diffusion models. While some T2V models are trained from scratch, the majority of them are trained based on existing T2I models. A typical example isAnimateDiff [13], which utilizes pre-trained T2I models as the spatial layer and trains additional motion modeling modules to reconstruct N frames clean video latents $z^{1:N}$ from i.i.d Gaussian noise conditioned on the text prompt $c$ by $\\mathcal{L}(\\theta) = \\mathbb{E}_{z_0^{1:N},\\epsilon ,c_t,t}\\left[\\| \\epsilon -\\hat{\\epsilon}_\\theta \\left(z_t^{1:N},c_t,t\\right)\\right]$ .", + "bbox": [ + 511, + 750, + 908, + 904 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7267", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c19118d955a0bf730caca8b242170a566cbfeed7a179503c644b0de4dea17533.jpg", + "image_caption": [ + "Figure 2. The Motion Generation Stage consists of two parts: Keyframe Candidates Generation and Coarse Motion Video Generation. In the first part, we augment the user input to obtain Layout, then generate keyframe candidates Frame with an inflated grounded T2I model. In the second part, we extract motion guidance and textural guidance from Frame to generate a coarse motion video." + ], + "image_footnote": [], + "bbox": [ + 133, + 68, + 867, + 262 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Overall Framework", + "text_level": 1, + "bbox": [ + 89, + 334, + 279, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "IM-Zero framework consists of two critical stages: Motion Generation Stage and Video Refinement Stage. In Section 3.1, we first briefly introduce the fundamentals of grounded T2I and T2V diffusion models. Then, we introduce the Motion Generation Stage in Section 3.3 as shown in Figure 2. This stage generates keyframe candidates and extracts guidance from them to generate a coarse motion video. Section 3.4 introduces the Video Refinement Stage as shown in Figure 3. This stage utilizes motion prior of T2V models to perform Motion Injection and utilizes the detail-generation capacity of T2I models to perform Detail Injection, further enhancing the visual quality of coarse motion video. Lastly, we also offer an option for motion transfer via reference video input, and we will explain how to easily modify IM-Zero in this case. As illustrated in Figure 1, IM-Zero achieves motion control for overall instances or subparts and allows users to further specify instance shapes as the grounded T2I model can generate keyframe candidates via masks. IM-Zero can perform motion transfer. Besides, IM-Zero allows high-quality T2V generation by refining any video via the Video Refinement Stage.", + "bbox": [ + 89, + 357, + 483, + 675 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Motion Generation Stage", + "text_level": 1, + "bbox": [ + 89, + 681, + 321, + 698 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The core task involves presenting suitable motion patterns in the video, including two key aspects: the user-desired instance spatial location and the user-desired instance movement. Previous training-free methods achieved this task by manipulating attention maps but encountered issues such as inaccurate location, poor motion effects, and low video quality, as discussed in Section 1. IM-Zero aims to achieve the decoupling of instance spatial location and instance movement to enhance the performance of both aspects. For the former, we utilize a grounded T2I model [39] to generate instances at user-specified locations. For the latter, we depart from the previous approach of imposing excessive constraints on the attention map and instead efficiently", + "bbox": [ + 89, + 704, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "leverage the motion priors of the T2V model.", + "bbox": [ + 511, + 335, + 813, + 351 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Keyframe Candidates Generation. Considering the user input $S$ layout signals and $L$ trajectory segments, We denote the layout signals as $\\mathbf{Layout} = \\{s_i|i\\sim [1,2,\\dots S]\\}$ (i.e., bounding boxes) and the movement trajectory as Traj = $\\{l_i|i\\sim [1,2,\\dots ,L]\\}$ . We ensure that each $l_{i}$ segment has a layout signal at both endpoints, for the endpoint without a layout signal, we select the closest $s_i$ from Layout and duplicate it here, obtaining the augmented layout signals $\\mathbf{Layout} = \\{s_i|i\\sim [1,2,\\dots,S']\\}$ . These signals are then input into the grounded T2I model to generate keyframe candidates $\\mathbf{Frame} = \\{f_i|i\\sim [1,2,\\dots,S']\\}$ . Since generating each image separately can lead to significant inconsistencies, following [22], we inflate the grounded T2I model and replace each self-attention in UNet with cross-frame attention (CFA). The CFA for the $i$ -th keyframe candidate $f_{i}$ is defined as CFA $(Q^{i},K^{1:N_{\\mathrm{S^{\\prime}}}},V^{1:N_{\\mathrm{S^{\\prime}}}}) =$ Softmax $(Q^{i}(K^{1})^{T} / \\sqrt{c})V^{1}$ where $Q,K$ and $V$ denote the queries, keys and values. The first candidate provides key and value for all following ones, which can be interpreted from an image editing perspective as querying relevant content and textures from the first candidate to ensure consistency in textures [5, 22].", + "bbox": [ + 511, + 354, + 906, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Coarse Motion Video Generation. After obtaining Frame, we utilize these candidates to generate a coarse motion video reflecting the user's input. To ensure smooth and natural instance motion, we have avoided the previous approach of directly enforcing excessive constraints on the attention map, which could disrupt the T2V motion priors. Instead, we extract two types of guidance from Frame to guide the T2V model. (1) Motion Guidance. To ensure the fidelity of the video to the user's input, encompassing spatial location and motion patterns, we extract foreground instances from Frame and convert them into Canny edge images. Subsequently, we employ ControlNet [12, 16, 55] to guide the generation of the coarse motion video. The rationale behind our approach is elucidated as follows: firstly,", + "bbox": [ + 511, + 689, + 908, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7268", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f7b89de37622acb7658d1156de3b5f92d9fdc196fdcae0f3ffba79510d6691b8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 133, + 69, + 598, + 145 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/537bb9c31e78acb36f5242d41bb91e058f816e55c24b1e19f1ffbd6f2998e43f.jpg", + "image_caption": [ + "Figure 3. In Video Refinement Stage, we diffuse coarser motion video latents and perform motion injection and detail injection in the reverse process. In motion injection, we first obtain $z_{t-1}$ by $z_t$ , then denoise $z_{t-1}$ for further $M$ steps using a T2V model to $\\tilde{z}_{t-1-M}'$ and combine them. We visualize the motion injection effect using CoTracker [21]. In detail injection, we first transform $\\tilde{z}_{t-1}$ from T2V noise distribution to T2I noise distribution, then denoise it for further $N$ steps with an inflated T2I model. The result is reversed to T2V noise distribution and combined with $\\tilde{z}_{t-1}$ . We visualize the detail injection effect using FFT analysis." + ], + "image_footnote": [], + "bbox": [ + 133, + 148, + 598, + 229 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b32e7d313bff07f481ff53c8c3d8f9463787037d7f103cde51ea2e41194272bc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 607, + 69, + 864, + 233 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the separation of foreground from the background is essential because instances described in the user's input manifest as foreground instances in Frame [25, 39]. Although we apply operations to enhance consistency by CFA, background distortions and abrupt changes can still occur [22]. Secondly, transforming images into Canny edge representations not only aligns with the requirements of ControlNet but also further mitigates potential inconsistencies in instance textures within Frame. (2) Textural Guidance. Use Motion Guidance only may cause textural inconsistency, hence we further perform Textural Guidance. We diffuse the first keyframe candidate $f_{1}$ to form the initial noise of the T2V model. To help the guidance transmit from $f_{1}$ to subsequent frames, we devise the following strategies: firstly, we employ an IP-Adapter [51] to utilize $f_{1}$ as an image prompt. Moreover, we also apply CFA to the spatial layers of the T2V model.", + "bbox": [ + 91, + 339, + 483, + 597 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Video Refinement Stage", + "text_level": 1, + "bbox": [ + 89, + 614, + 310, + 631 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We have observed that while coarse motion videos may align with user-desired inputs, unsmooth motion issues and suboptimal visual quality issues such as watermark and flickering exist. To attack these issues, we introduce a zero-shot method by more efficiently leveraging the motion priors inherent in existing T2V models and the detail-generation capacity of T2I models. As shown in Figure 3, we integrate a T2V model into the reverse process of video generation. We leverage T2V motion priors to inject motion into video latents, ensuring smoother and more natural motion. Subsequently, we also insert a T2I model, leveraging its detail generation capacity to enrich video detail richness.", + "bbox": [ + 89, + 641, + 482, + 821 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given a low-quality coarse motion video as described in Section 3.3, we start by encoding it using the encoder $\\mathcal{E}$ of a vanilla T2V model (i.e.,AnimateDiff [13]) and adding Gaussian noise by timestep $T$ to obtain noisy video latents $z_{T}$ . For each denoising time step $t$ , we first denoise $z_{t}$ using", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the denoiser of the T2V model to obtain $z_{t - 1}$ by:", + "bbox": [ + 511, + 339, + 839, + 354 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(z _ {t - 1} \\mid z _ {t}\\right) = \\mathcal {N} \\left(z _ {t - 1}; z _ {t} - \\hat {\\epsilon} _ {\\theta} \\left(z _ {t}; c, t\\right), \\left(1 - \\alpha_ {t} ^ {2}\\right) \\mathbf {I}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 363, + 903, + 395 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $c$ is the text prompt embedding. Then we sequentially perform motion injection and detail injection on $z_{t - 1}$", + "bbox": [ + 511, + 396, + 903, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Motion Injection. We employ a vanilla T2V model (i.e.,AnimateDiff [13]) for motion injection as shown in Figure 3. Specifically, we denoise $z_{t - 1}$ using its denoiser for further $M$ steps to acquire $\\tilde{z}_{t - 1 - M}^{\\prime}$ , where each denoising step follows Equation (1). We observed a significant improvement in the temporal consistency of $\\tilde{z}_{t - 1 - M}^{\\prime}$ , resulting in very smooth motion. However, this enhancement comes at the cost of losing many detailed features. Therefore, we combine $\\tilde{z}_{t - 1 - M}^{\\prime}$ with $z_{t - 1}$ to enhance motion smoothness while preserving details by:", + "bbox": [ + 511, + 426, + 905, + 578 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {z} _ {t - 1} = \\lambda_ {1} * z _ {t - 1} + (1 - \\lambda_ {1}) * \\tilde {z} _ {t - 1 - M} ^ {\\prime} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 587, + 903, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{1}$ is a coefficient controlling the fusion scale.", + "bbox": [ + 511, + 614, + 862, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Detail Injection. We then employ an inflated T2I model for detail injection as shown in Figure 3. Due to the different noise distributions in the T2I and T2V models, a noise transformation is required. To convert video latents into image latents, we directly estimate clean latents [48] $\\hat{z}_0$ by:", + "bbox": [ + 511, + 628, + 905, + 705 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {z} _ {0} = \\frac {\\tilde {z} _ {t} - \\sqrt {1 - \\alpha_ {t}} \\epsilon_ {\\theta} (\\tilde {z} _ {t} , c , t)}{\\sqrt {\\alpha_ {t}}} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 714, + 903, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and we then conduct DDIM inversion [30] to derive the corresponding noisy image latents $\\tilde{z}_{t - 1}^{I}$ , the DDIM inversion process is defined by:", + "bbox": [ + 511, + 757, + 906, + 803 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {z} _ {t} ^ {I} = \\sqrt {\\alpha_ {t}} \\hat {z} _ {0} + \\sqrt {1 - \\alpha_ {t}} \\cdot \\epsilon_ {\\theta} (\\tilde {z} _ {t - 1} ^ {I}, c, t) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 811, + 903, + 830 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Subsequently, we denoise the noisy image latents with the inflated T2I model for further $D$ steps to obtain $\\tilde{z}_{t-1-D}^{I}$ . The denoising process resembles Equation (1). As the T2I model excels in detail generation compared to the T2V", + "bbox": [ + 511, + 840, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7269", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/af8934e9b64973ed90739bcec9e2fb12bab3a5e4c2413e8971a8c7ef4cdee6dc.jpg", + "image_caption": [ + "Figure 4. Qualitative Results of IM-Zero compared with other methods. Videos generated by IM-Zero outperform others in control accuracy, motion effects, and video quality." + ], + "image_footnote": [], + "bbox": [ + 143, + 66, + 866, + 478 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "model, we revert $\\tilde{z}_{t - 1 - D}^{I}$ back to the noise distribution of the T2V model, combining the transformed latents with $\\tilde{z}_{t - 1}$ obtained from Equation (2) to integrate details. Specifically, in the reversed noise transformation process, we predict clean noise from $\\tilde{z}_{t - 1}$ using Equation (3) and utilize the DDPM forward process to obtain the corresponding noisy video latents $\\tilde{z}_{t - 1 - D}^{V}$ . Then we fuse them by:", + "bbox": [ + 89, + 535, + 483, + 643 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nz _ {t - 1} = \\lambda_ {2} * \\tilde {z} _ {t - 1} + (1 - \\lambda_ {2}) * \\tilde {z} _ {t - 1 - D} ^ {V} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 655, + 482, + 674 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\lambda_{2}$ controls the fusion scale and $z_{t - 1}$ is used for the next denoising timestep.", + "bbox": [ + 89, + 685, + 482, + 717 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.5. Adapt IM-Zero for Motion Transfer", + "text_level": 1, + "bbox": [ + 89, + 726, + 405, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For a more customized motion pattern provided by reference videos, IM-Zero can also be easily adapted for motion transfer tasks. Given a reference video and the target text prompt, we extract control signals (e.g., depth maps) from each frame of the reference video. Subsequently, we use these dense control signals and the target text prompt to guide the T2V model in generating a coarse motion video via ControlNet. Then, the coarse motion video undergoes video refinement stage as detailed in Section 3.4 to enhance motion consistency and video frame quality.", + "bbox": [ + 89, + 750, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 511, + 534, + 638, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Qualitative Results", + "text_level": 1, + "bbox": [ + 511, + 560, + 697, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Regarding the comparative methods, we primarily selected Peekaboo [20], TrailBlazer [29], and MotionBooth [42], which generate videos by specifying boxes and trajectories. Furthermore, to further illustrate the effectiveness of our approach, we also included T2V-Zero [22], which offers rough guidance by specifying translation vectors. To ensure a fair comparison, following the experimental settings of [29], all methods utilize the same layout input, trajectory input, and text prompt input. Additional results of controllable video generation are available in the supplementary material. We also compared with FreeTraj [32], but as it additionally requires inputting boxes with fixed sizes across frames to make noise flow, we place the comparison in the supplementary material.", + "bbox": [ + 509, + 582, + 906, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Figure 4, our method significantly outperforms other approaches in terms of alignment with the input, motion effects, and video quality. In terms of alignment with the input, T2V-Zero exhibits almost no resemblance to the input, while the two methods based on attention map manipulation, Peekaboo, and TrailBlazer, sometimes roughly match the input location and trajectory but", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7270", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4a2e730bd470dfde2cc192aa8bbbceb492b5208533dfac3c82b8451a19c26fe1.jpg", + "table_caption": [ + "Table 1. Quantitative results of IM-Zero compared with other methods on DAVIS-17 [31] dataset and GOT10k [17] dataset. IM-Zero achieves SOTA in all metrics, including metrics measuring video quality, frame consistency, and control accuracy." + ], + "table_footnote": [], + "table_body": "
MethodDAVIS-17GOT10k
FVD (↓)KID (↓)CLIPSim (↑)mIoU (↑)CD (↓)FVD (↓)KID (↓)CLIPSim (↑)mIoU (↑)CD (↓)
T2V-Zero [22]5004.0030.3795.120.190.344335.6924.7295.410.120.32
Peekaboo [20]4264.7026.4897.330.190.464960.4919.0598.680.110.34
TrailBlazer [29]4593.3726.0498.100.210.294695.1816.1897.800.130.40
MotionBooth [42]4192.1926.8596.640.220.364311.7116.0597.470.140.42
IM-Zero (Ours)4070.0125.7598.910.270.264187.2715.9299.040.210.27
", + "bbox": [ + 91, + 109, + 906, + 237 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/65216db07423447a9641d987edd2958730d24491713be4fda04faeeb044d068e.jpg", + "table_caption": [ + "Table 2. Ablation on different components in Motion Generation Stage." + ], + "table_footnote": [], + "table_body": "
MethodFVD (↓)KID (↓)CLIPSim (↑)mIoU (↑)CD (↓)
w/o 1st CFA4320.0726.8498.840.260.29
w/o 2nd CFA4528.0928.0898.780.210.31
w/o init noise4786.1726.6998.820.260.31
w/o IP-Adapter4549.9427.7598.870.260.28
IM-Zero (Ours)4070.0125.7598.910.270.26
", + "bbox": [ + 168, + 262, + 823, + 369 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "still lack precision. Regarding motion effects, Peekaboo and TrailBlazer generate videos with unnatural motion, resulting in strange discontinuities between frames. Concerning video quality, Peekaboo and TrailBlazer exhibit low image quality, lack detail, and introduce artifacts. We also provide a video version of Figure 4 in the supplementary material.", + "bbox": [ + 88, + 387, + 480, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Quantitative Results", + "text_level": 1, + "bbox": [ + 89, + 500, + 284, + 517 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Experiment Settings. For datasets, we utilize the training set and validation set of the DAVIS-17 dataset [31] and the test set of the GOT10k dataset [17]. We use BLIP-2 [24] to generate text prompts and GroundingDINO [28] to extract bounding boxes from the first and last frame as inputs of all the methods. After eliminating videos where no bounding boxes are extracted, we randomly select 50 videos and 70 videos from the rest videos of DAVIS-17 [31] and GOT10k [17]. For our method, in the first stage we use InstanceDiffusion [39] as grounded T2I model and SparseCtrl [12] as ControlNet, in the second stage we use Stable Diffusion V1.5 [34] as T2I model, on both stage we useAnimateDiff [13] as T2V model. For the evaluation metrics, we employ mean Intersection over Union (mIoU) and Centroid Distance (CD) to assess control alignment, Fréchet Video Distance (FVD) [35] and Kernel Inception Distance (KID) [2] to evaluate video quality, and CLIP similarities (CLIPSim) [33] to assess frame-to-frame consistency. Detailed settings are available in the supplementary material.", + "bbox": [ + 91, + 523, + 482, + 810 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Metrics Evaluation Results. We conducted evaluations on the aforementioned metrics on the DAVIS-17 dataset [31] and the GOT10k dataset [17]. As shown in Table 1, our method achieved state-of-the-art (SOTA) performance compared to several other methods. Firstly, in terms of the alignment between the generated videos and user inputs,", + "bbox": [ + 89, + 810, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2e8334a14956e47173a7c32e0f02d9043905eb46b07527dfc4d78fa690a29fbc.jpg", + "table_caption": [ + "Table 3. Motion transfer results compared with other methods." + ], + "table_footnote": [], + "table_body": "
MethodMotion Fidelity (↑)Imaging Quality (↑)
MotionClone [26]74.50.72
MOFT [23]51.20.71
MotionDirector [59]75.50.68
IM-Zero (Ours)63.70.73
", + "bbox": [ + 513, + 411, + 908, + 500 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9b1b257a1d9acfdfccf5b1bd37bf741db8044da6f804bc434fa173afbd88049a.jpg", + "table_caption": [ + "Table 4. T2V results compared with UNet baseline." + ], + "table_footnote": [], + "table_body": "
MethodImaging Quality (↑)CLIPSim (↑)
AnimateDiff [13]0.6799.05
IM-Zero (Ours)0.7499.73
", + "bbox": [ + 529, + 537, + 890, + 597 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "our method demonstrates a significant lead. Compared to the previous best methods, our approach achieved an improvement of $28.6\\%$ and $61.5\\%$ in mIoU, and reductions of $10.4\\%$ and $15.6\\%$ in CD on the two datasets, indicating that our method offers more accurate location control and motion trajectory control compared to previous methods based on attention map manipulation. Secondly, regarding frame-to-frame consistency in the videos, thanks to our enhanced consistency techniques and motion injection methods, our approach also reaches the SOTA performance. Moreover, in terms of video quality, our method significantly outperforms others, with FVD reduced to 4070.01 and 4187.27, and KID reduced to 25.75 and 15.92, respectively.", + "bbox": [ + 511, + 625, + 906, + 821 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Versatile Capacity Analysis", + "text_level": 1, + "bbox": [ + 511, + 832, + 759, + 849 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our method possesses versatile capacity as shown in Figure 5. First, our method can control the subparts of instances, achieving effects such as head turning and hand", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7271", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/54d5cbbc0ed4713fd4710482d519fe249ff17f62fdd1a5cd9ae15d12c13078d3.jpg", + "image_caption": [ + "#", + "Figure 5. Besides box-level overall instance motion control, IM-Zero can: (a) Perform instance subpart control. (b) Specify instance shape by mask. (c) Perform motion transfer from the reference video(left) to the target video(right). (d) Perform high-quality T2V generation." + ], + "image_footnote": [], + "bbox": [ + 163, + 63, + 866, + 335 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "movement. Second, besides bounding boxes, we also support specifying instance shapes more precisely using masks. Third, for cases where users provide a reference video, our method can perform motion transfer. Finally, users can directly provide text inputs for high-quality T2V generation. For motion transfer and T2V generation, we compare IM-Zero with other methods as follows.", + "bbox": [ + 88, + 401, + 480, + 505 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Motion Transfer. We compare with MOFT [23], MotionClone [26], and MotionDirector [59] using the open-sourced data from MotionClone [26]. For quantitative results, we use Motion Fidelity [50] and Imaging Quality [19] as metrics following MOFT [23]. Table 3 shows that our method is competitive in Motion Fidelity [50] and is the best in Imaging Quality [19]. Qualitative results and other details are provided in the supplementary material.", + "bbox": [ + 88, + 507, + 482, + 627 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "T2V Generation. As we use UNet-based models, we compare T2V with UNet baselineAnimateDiff [13]. For quantitative results, we use Imaging Quality [19] and CLIP-Sim [33] as metrics. As there are no ground-truth videos, we generate videos using 50 prompts randomly generated by GPT. Table 4 shows that our method is better in Imaging Quality [19] and CLIPSim [33]. Refer to the supplementary material for qualitative results and other details.", + "bbox": [ + 88, + 628, + 482, + 748 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Other Capacities. The proposed method also supports multi-instance motion control, and the decomposition of our pipeline allows autoregressive usage of the second stage to enhance temporal consistency, as elaborated in the supplementary material.", + "bbox": [ + 88, + 750, + 482, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 833, + 256, + 848 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation on different components in Motion Generation", + "text_level": 1, + "bbox": [ + 89, + 854, + 482, + 869 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Stage. We ablate on different components in Section 3.3. As depicted in Table 2, these methods improve the quality", + "bbox": [ + 89, + 869, + 480, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "of the final videos and also enhance the alignment with user inputs. Refer to the supplementary material for details.", + "bbox": [ + 511, + 401, + 906, + 431 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation on different components in Video Refinement Stage. We also conduct ablation studies on Video Refinement Stage. Refer to the supplementary material for details.", + "bbox": [ + 511, + 435, + 906, + 481 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion and Limitation", + "text_level": 1, + "bbox": [ + 511, + 505, + 764, + 522 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce IM-Zero, a novel training-free framework for instance-level motion controllable video generation using diffusion models. We leverage grounded text-to-image models to generate keyframe candidates and extract motion guidance and textural guidance from them to generate videos that align with user inputs. To further enhance video motion and quality, we also introduce a zero-shot method for motion injection and detail injection. Additionally, our method can control the subpart motion of instances, use masks to specify instance shapes, perform motion transfer from reference videos, and directly support high-quality text-to-video generation.", + "bbox": [ + 511, + 534, + 906, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitation. The zero-shot features impose certain limitations on our method. Firstly, as zero-shot methods depend on pre-trained model weights, the videos generated by our method are inherently constrained by the pre-trained models employed. For example, the alignment with layout input is influenced by the grounded T2I model and video ControlNet, the video quality is affected by the T2V model, which can be improved by using more advanced pre-trained models. Secondly, the temporal consistency of videos generated by zero-shot methods is inherently weaker compared to training-based methods, which can be improved by incorporating some tuning components to our method.", + "bbox": [ + 511, + 719, + 908, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7272", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by the National Natural Science Foundation of China under Grant 62125109, Grant 62431017, Grant U24A20251, Grant 62320106003, Grant 62371288, Grant 62401357, Grant 62401366, Grant 62301299, Grant 62120106007, and in part by the Program of Shanghai Science and Technology Innovation Project under Grant 24BC3200800.", + "bbox": [ + 89, + 114, + 483, + 220 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 232, + 187, + 248 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. MultiDiffusion: Fusing diffusion paths for controlled image generation. In Proceedings of the 40th International Conference on Machine Learning, pages 1737-1752, 2023. 3", + "[2] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying MMD GANs. In *The Sixth International Conference on Learning Representations*, 2018. 7", + "[3] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 2, 3", + "[4] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. OpenAI Blog, 2024. 2", + "[5] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. MasaCtrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22560–22570, 2023. 4", + "[6] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, Chao Weng, and Ying Shan. VideoCrafter1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512, 2023. 2, 3", + "[7] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. VideoCrafter2: Overcoming data limitations for high-quality video diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7310-7320, 2024. 2, 3", + "[8] Minghao Chen, Iro Laina, and Andrea Vedaldi. Training-free layout control with cross-attention guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5343–5353, 2024. 3", + "[9] Weifeng Chen, Yatai Ji, Jie Wu, Hefeng Wu, Pan Xie, Jiashi Li, Xin Xia, Xuefeng Xiao, and Liang Lin. Control-A-Video: Controllable text-to-video generation with diffusion models. arXiv preprint arXiv:2305.13840, 2023. 2", + "[10] Yabo Chen, Jiemin Fang, Yuyang Huang, Taoran Yi, Xiaopeng Zhang, Lingxi Xie, Xinggang Wang, Wenrui Dai," + ], + "bbox": [ + 93, + 257, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Hongkai Xiong, and Qi Tian. Cascade-Zero123: One image to highly consistent 3D with self-prompted nearby views. In Proceedings of the 18th European Conference on Computer Vision, pages 311-330, 2024. 3", + "[11] Yufan Deng, Ruida Wang, Yuhao Zhang, Yu-Wing Tai, and Chi-Keung Tang. DragVideo: Interactive drag-style video editing. In Proceedings of the 18th European Conference on Computer Vision, pages 183–199, 2024. 2, 3", + "[12] Yuwei Guo, Ceyuan Yang, Anyi Rao, Maneesh Agrawala, Dahua Lin, and Bo Dai. SparseCtrl: Adding sparse controls to text-to-video diffusion models. In Proceedings of the 18th European Conference on Computer Vision, pages 330-348, 2024. 4, 7", + "[13] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. AnimateDiff:Animate your personalized text-to-image diffusion models without specific tuning. In The Twelfth International Conference on Learning Representations, 2024. 2, 3, 5, 7, 8", + "[14] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P. Kingma, Ben Poole, Mohammad Norouzi, David J. Fleet, and Tim Salimans. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022.", + "[15] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. In Advances in Neural Information Processing Systems 35, pages 8633-8646, 2022. 2, 3", + "[16] Zhihao Hu and Dong Xu. VideoControlNet: A motion-guided video-to-video translation framework by using diffusion model with ControlNet. arXiv preprint arXiv:2307.14073, 2023. 2, 4", + "[17] Lianghua Huang, Xin Zhao, and Kaiqi Huang. GOT-10k: A large high-diversity benchmark for generic object tracking in the wild. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(5):1562-1577, 2019. 7", + "[18] Yuyang Huang, Yabo Chen, Yuchen Liu, Xiaopeng Zhang, Wenrui Dai, Hongkai Xiong, and Qi Tian. DomainFusion: Generalizing to unseen domains with latent diffusion models. In Proceedings of the 18th European Conference on Computer Vision, pages 480-498, 2024. 3", + "[19] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, Yaohui Wang, Xinyuan Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. VBench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024. 8", + "[20] Yash Jain, Anshul Nasery, Vibhav Vineet, and Harkirat Behl. PEEKABOO: Interactive video generation via masked-diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8079-8088, 2024. 2, 3, 6, 7", + "[21] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. CoTracker: It is better to track together. In Proceedings of the 18th European Conference on Computer Vision, pages 18-35, 2024. 5" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7273", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[22] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2Video-Zero: Text-to-image diffusion models are zero-shot video generators. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15954-15964, 2023. 2, 3, 4, 5, 6, 7", + "[23] Karlo Koledić, Igor Cvišić, Ivan Marković, and Ivan Petrović. MOFT: Monocular odometry based on deep depth and careful feature selection and tracking. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 6175-6181. IEEE, 2023. 7, 8", + "[24] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In Proceedings of the 40th International Conference on Machine Learning, pages 19730–19742, 2023. 7", + "[25] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. GLIGEN: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22511-22521, 2023. 2, 3, 5", + "[26] Pengyang Ling, Jiazi Bu, Pan Zhang, Xiaoyi Dong, Yuhang Zang, Tong Wu, Huaian Chen, Jiaqi Wang, and Yi Jin. MotionClone: Training-free motion cloning for controllable video generation. In The Thirteenth International Conference on Learning Representations, 2025. 7, 8", + "[27] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9298–9309, 2023. 3", + "[28] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, and Lei Zhang. Grounding DINO: Marrying DINO with grounded pre-training for open-set object detection. In Proceedings of the 18th European Conference on Computer Vision, pages 38-55, 2024. 7", + "[29] Wan-Duo Kurt Ma, John P Lewis, and W Bastiaan Kleijn. TrailBlazer: Trajectory control for diffusion-based video generation. In SIGGRAPH Asia 2024, number 97, 2024. 2, 3, 6, 7", + "[30] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 5", + "[31] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 DAVIS challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 7", + "[32] Haonan Qiu, Zhaoxi Chen, Zhouxia Wang, Yingqing He, Menghan Xia, and Ziwei Liu. FreeTraj: Tuning-free trajectory control in video diffusion models. arXiv preprint arXiv:2406.16863, 2024. 2, 3, 6", + "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry," + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763, 2021. 7, 8", + "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 7", + "[35] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 7", + "[36] Cong Wang, Jiaxi Gu, Panwen Hu, Haoyu Zhao, Yuanfan Guo, Jianhua Han, Hang Xu, and Xiaodan Liang. EasyControl: Transfer ControlNet to video diffusion for controllable generation and interpolation. arXiv preprint arXiv:2408.13005, 2024. 2", + "[37] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. ModelScope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 2, 3", + "[38] Weimin Wang, Jiawei Liu, Zhijie Lin, Jiangqiao Yan, Shuo Chen, Chetwin Low, Tuyen Hoang, Jie Wu, Jun Hao Liew, Hanshu Yan, Daquan Zhou, and Jiashi Feng. MagicVideoV2: Multi-stage high-aesthetic video generation. arXiv preprint arXiv:2401.04468, 2024. 2, 3", + "[39] Xudong Wang, Trevor Darrell, Sai Saketh Rambhatla, Rohit Girdhar, and Ishan Misra. InstanceDiffusion: Instance-level control for image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6232-6242, 2024. 2, 3, 4, 5, 7", + "[40] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, Yuwei Guo, Tianxing Wu, Chenyang Si, Yuming Jiang, Cunjian Chen, Chen Change Loy, Bo Dai, Dahua Lin, Yu Qiao, and Ziwei Liu. LaVie: High-quality video generation with cascaded latent diffusion models. International Journal of Computer Vision, 2024. 2, 3", + "[41] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. MotionCtrl: A unified and flexible motion controller for video generation. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2, 3", + "[42] Jianzong Wu, Xiangtai Li, Yanhong Zeng, Jiangning Zhang, Qianyu Zhou, Yining Li, Yunhai Tong, and Kai Chen. MotionBooth: Motion-aware customized text-to-video generation. In Advances in Neural Information Processing Systems 37, pages 34322-34348, 2024. 6, 7", + "[43] Mingrui Wu, Oucheng Huang, Jiayi Ji, Jiale Li, Xinyue Cai, Huafeng Kuang, Jianzhuang Liu, Xiaoshuai Sun, and Rongrong Ji. TraDiffusion: Trajectory-based training-free image generation. arXiv preprint arXiv:2408.09739, 2024. 3", + "[44] Weijia Wu, Zhuang Li, Yuchao Gu, Rui Zhao, Yefei He, David Junhao Zhang, Mike Zheng Shou, Yan Li, Tingting Gao, and Di Zhang. DragAnything: Motion control for anything using entity representation. In Proceedings of the 18th" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7274", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "European Conference on Computer Vision, pages 331-348, 2024. 2, 3", + "[45] Enze Xie, Junsong Chen, Junyu Chen, Han Cai, Haotian Tang, Yujun Lin, Zhekai Zhang, Muyang Li, Ligeng Zhu, Yao Lu, and Song Han. SANA: Efficient high-resolution image synthesis with linear diffusion transformers. In The Thirteenth International Conference on Learning Representations, 2025. 3", + "[46] Jinheng Xie, Yuexiang Li, Yawen Huang, Haozhe Liu, Wentian Zhang, Yefeng Zheng, and Mike Zheng Shou. BoxDiff: Text-to-image synthesis with training-free box-constrained diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7452-7461, 2023. 3", + "[47] Yichao Yan, Zanwei Zhou, Zi Wang, Jingnan Gao, and Xiaokang Yang. DialogueNeRF: Towards realistic avatar face-to-face conversation video generation. Visual Intelligence, 2 (1):24, 2024. 3", + "[48] Serin Yang, Hyunmin Hwang, and Jong Chul Ye. Zero-shot contrastive loss for text-guided diffusion image style transfer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22873-22882, 2023. 5", + "[49] Zhengyuan Yang, Jianfeng Wang, Zhe Gan, Linjie Li, Kevin Lin, Chenfei Wu, Nan Duan, Zicheng Liu, Ce Liu, Michael Zeng, and Lijuan Wang. ReCo: Region-controlled text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14246-14255, 2023. 3", + "[50] Danah Yatim, Rafail Fridman, Omer Bar-Tal, Yoni Kasten, and Tali Dekel. Space-time diffusion features for zero-shot text-driven motion transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8466-8476, 2024. 8", + "[51] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. IP-Adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 5", + "[52] Shengming Yin, Chenfei Wu, Jian Liang, Jie Shi, Houqiang Li, Gong Ming, and Nan Duan. DragNUWA: Fine-grained control in video generation by integrating text, image, and trajectory. arXiv preprint arXiv:2308.08089, 2023. 2, 3", + "[53] Yan Zeng, Guoqiang Wei, Jiani Zheng, Jiaxin Zou, Yang Wei, Yuchen Zhang, and Hang Li. Make pixels dance: High-dynamic video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8850-8860, 2024. 2", + "[54] Junyi Zhang, Charles Herrmann, Junhwa Hur, Luisa Polania Cabrera, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. A tale of two features: Stable diffusion complements DINO for zero-shot semantic correspondence. In Advances in Neural Information Processing Systems 36, pages 45533-45547, 2023. 3", + "[55] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 4", + "[56] Yabo Zhang, Yuxiang Wei, Dongsheng Jiang, Xiaopeng Zhang, Wangmeng Zuo, and Qi Tian. ControlVideo:" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Training-free controllable text-to-video generation. In The Twelfth International Conference on Learning Representations, 2024. 2, 3", + "[57] Yabo Zhang, Yuxiang Wei, Xianhui Lin, Zheng Hui, Peiran Ren, Xuansong Xie, Xiangyang Ji, and Wangmeng Zuo. VideoElevator: elevating video generation quality with versatile text-to-image diffusion models. arXiv preprint arXiv:2403.05438, 2024. 3", + "[58] Zhenghao Zhang, Junchao Liao, Menghao Li, Long Qin, and Weizhi Wang. Tora: Trajectory-oriented diffusion transformer for video generation. arXiv preprint arXiv:2407.21705, 2024. 2, 3", + "[59] Rui Zhao, Yuchao Gu, Jay Zhangjie Wu, David Junhao Zhang, Jia-Wei Liu, Weijia Wu, Jussi Keppo, and Mike Zheng Shou. MotionDirector: Motion customization of text-to-video diffusion models. In Proceedings of the 18th European Conference on Computer Vision, pages 273–290, 2024. 7, 8", + "[60] Wenliang Zhao, Yongming Rao, Zuyan Liu, Benlin Liu, Jie Zhou, and Jiwen Lu. Unleashing text-to-image diffusion models for visual perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5706-5716, 2023. 3", + "[61] Wendi Zheng, Jiayan Teng, Zhuoyi Yang, Weihan Wang, Ji-dong Chen, Xiaotao Gu, Yuxiao Dong, Ming Ding, and Jie Tang. CogView3: Finer and faster text-to-image generation via relay diffusion. In Proceedings of the 18th European Conference on Computer Vision, pages 1-22, 2024. 3", + "[62] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. MagicVideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022. 2, 3", + "[63] Dewei Zhou, You Li, Fan Ma, Xiaoting Zhang, and Yi Yang. MIGC: Multi-instance generation controller for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6818-6828, 2024. 3", + "[64] Dewei Zhou, You Li, Fan Ma, Zongxin Yang, and Yi Yang. MIGC++: Advanced multi-instance generation controller for image synthesis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 47(3):1714-1728, 2025. 3", + "[65] Haitao Zhou, Chuang Wang, Rui Nie, Jinxiao Lin, Dongdong Yu, Qian Yu, and Changhu Wang. TrackGo: A flexible and efficient method for controllable video generation. arXiv preprint arXiv:2408.11475, 2024. 2, 3" + ], + "bbox": [ + 516, + 92, + 903, + 726 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "7275", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/IM-Zero_ Instance-level Motion Controllable Video Generation in a Zero-shot Manner/a669fc0a-17c9-4001-97e0-7737b0fc24de_model.json b/2025/IM-Zero_ Instance-level Motion Controllable Video Generation in a Zero-shot Manner/a669fc0a-17c9-4001-97e0-7737b0fc24de_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a563da7f3449413e39ec3650913cd9bc2f139825 --- /dev/null +++ b/2025/IM-Zero_ Instance-level Motion Controllable Video Generation in a Zero-shot Manner/a669fc0a-17c9-4001-97e0-7737b0fc24de_model.json @@ -0,0 +1,2708 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.18, + 0.131, + 0.82, + 0.175 + ], + "angle": 0, + "content": "IM-Zero: Instance-level Motion Controllable Video Generation in a Zero-shot Manner" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.203, + 0.884, + 0.293 + ], + "angle": 0, + "content": "Yuyang Huang\\(^{1*}\\), Yabo Chen\\(^{1*}\\), Li Ding\\(^{1}\\), Xiaopeng Zhang\\(^{2\\dagger}\\), Wenrui Dai\\(^{1\\dagger}\\), Junni Zou\\(^{1\\dagger}\\), Hongkai Xiong\\(^{1}\\), Qi Tian\\(^{2}\\) \n\\(^{1}\\)Shanghai Jiao Tong University, Shanghai, China \\(^{2}\\)Huawei Inc., Shenzhen, China \n{huangyuyang, chenyabo, wangzhongren, daiwenrui, zoujunni, xionghongkai}@sjtu.edu.cn \nzxphistory@gmail.com, tian.qil@huawei.com" + }, + { + "type": "image", + "bbox": [ + 0.139, + 0.302, + 0.222, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.321, + 0.272, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.304, + 0.363, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.304, + 0.446, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.448, + 0.304, + 0.529, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.531, + 0.304, + 0.613, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.615, + 0.304, + 0.696, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.304, + 0.779, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.781, + 0.304, + 0.863, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.139, + 0.369, + 0.454, + 0.38 + ], + "angle": 0, + "content": "(a) A person riding a snowmobile down a snowy mountain." + }, + { + "type": "image", + "bbox": [ + 0.139, + 0.381, + 0.223, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.401, + 0.272, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.281, + 0.381, + 0.363, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.381, + 0.446, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.448, + 0.381, + 0.529, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.531, + 0.381, + 0.613, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.615, + 0.381, + 0.696, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.381, + 0.779, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.781, + 0.381, + 0.863, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.14, + 0.447, + 0.439, + 0.459 + ], + "angle": 0, + "content": "(b) A chef with a white hat preparing a gourmet meal." + }, + { + "type": "image", + "bbox": [ + 0.139, + 0.46, + 0.222, + 0.523 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.482, + 0.272, + 0.505 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.281, + 0.46, + 0.363, + 0.523 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.46, + 0.446, + 0.523 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.448, + 0.46, + 0.529, + 0.523 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.531, + 0.46, + 0.613, + 0.523 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.615, + 0.46, + 0.696, + 0.523 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.46, + 0.779, + 0.523 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.781, + 0.46, + 0.863, + 0.523 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.14, + 0.523, + 0.422, + 0.535 + ], + "angle": 0, + "content": "(c) A snow leopard is walking on the side of a cliff." + }, + { + "type": "image", + "bbox": [ + 0.139, + 0.537, + 0.222, + 0.601 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.537, + 0.304, + 0.6 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.306, + 0.537, + 0.388, + 0.6 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.39, + 0.537, + 0.472, + 0.6 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.478, + 0.555, + 0.518, + 0.578 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.537, + 0.612, + 0.6 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.615, + 0.537, + 0.696, + 0.6 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.537, + 0.779, + 0.6 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.781, + 0.537, + 0.863, + 0.6 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.139, + 0.603, + 0.468, + 0.616 + ], + "angle": 0, + "content": "(d) A lemur walking on a xylophone, in a tropical rainforest." + }, + { + "type": "image", + "bbox": [ + 0.147, + 0.628, + 0.203, + 0.67 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.639, + 0.272, + 0.661 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.627, + 0.363, + 0.68 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.62, + 0.446, + 0.68 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.448, + 0.618, + 0.529, + 0.68 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.531, + 0.618, + 0.613, + 0.68 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.615, + 0.618, + 0.696, + 0.68 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.618, + 0.779, + 0.68 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.781, + 0.618, + 0.863, + 0.68 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.138, + 0.682, + 0.441, + 0.695 + ], + "angle": 0, + "content": "(e) A snowman in a forest caught in a burst of flames." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.697, + 0.907, + 0.739 + ], + "angle": 0, + "content": "Figure 1. IM-Zero introduces a zero-shot framework for instance-level motion controllable video generation. IM-Zero allows users to: (a) Control overall instance motion. (b) Control instance subpart motion. (c) Control instance motion and specify instance shape via mask. (d) Perform motion transfer by reference video. (e) Perform high-quality text-to-video generation." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.751, + 0.327, + 0.766 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.771, + 0.483, + 0.862 + ], + "angle": 0, + "content": "Controllability of video generation has been recently concerned in addition to the quality of generated videos. The main challenge to controllable video generation is to synthesize videos based on user-specified instance spatial locations and movement trajectories. However, existing methods suffer from a dilemma between the resource con" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.753, + 0.907, + 0.89 + ], + "angle": 0, + "content": "sumption, generation quality, and user controllability. As an efficient alternative to prohibitive training-based video generation, existing zero-shot video generation methods cannot generate high-quality and motion-consistent videos under the control of layouts and movement trajectories. In this paper, we propose a novel zero-shot method named IM-Zero that ameliorates instance-level motion controllable video generation with enhanced control accuracy, motion consistency, and richness of details to address this prob" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.365, + 0.888 + ], + "angle": 0, + "content": "*These authors contributed equally to this work." + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.463, + 0.9 + ], + "angle": 0, + "content": "† Corresponding authors: Xiaopeng Zhang; Wenrui Dai; Junni Zou." + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.463, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7265" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.425 + ], + "angle": 0, + "content": "lem. Specifically, we first present a motion generation stage that extracts motion and textural guidance from keyframe candidates from pre-trained grounded text-to-image model to generate the desired coarse motion video. Subsequently, we develop a video refinement stage that injects the motion priors of pre-trained text-to-video models and detail priors of pre-trained text-to-image models into the latents of coarse motion videos to further enhance video motion consistency and richness of details. To our best knowledge, IM-Zero is the first to simultaneously achieve high-quality video generation and allow to control both layouts and movement trajectories in a zero-shot manner. Extensive experiments demonstrate that IM-Zero outperforms existing methods in terms of video quality, inter-frame consistency, and the alignment of location and trajectory. Furthermore, compared with existing methods, IM-Zero enjoys extra advantages of versatility in video generation, including motion control of subparts within instances, finer control of specifying instance shapes via masks, and more difficult tasks of motion transfer for customizing fine-grained motion patterns through reference videos and high-quality text-to-video generation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.453, + 0.223, + 0.469 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.478, + 0.483, + 0.629 + ], + "angle": 0, + "content": "Demands for enhanced controllability of text-to-video (T2V) generation are booming with the increasing quality of generated videos [3, 6, 7, 13-15, 22, 37, 38, 40, 62]. One of the critical problems lie in the precise control over the spatial layout and trajectories of instances within the generated videos. Existing methods fail to achieve fine-grained control by solely relying on text prompts to steer pre-trained T2V models, even employing fundamental T2V models with large-scale pre-training like Sora [4], KLING, and Doubao-PixelDance [53]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.63, + 0.483, + 0.795 + ], + "angle": 0, + "content": "Existing methods usually extend ControlNet [55] for controllable video generation with frame-by-frame fine-grained control signals, including Canny, depth, and HED signals [9, 16, 36, 56]. To obtain dense control signals, these methods impose high demands on the input end and are limited in practical applications. One approach is to extract control signals from reference videos but it is hard to obtain reference videos perfectly matching the desired motion patterns. The other alternative requires manually crafting the dense control signals. However, it is also non-trivial to craft frame-by-frame control signals without expertise." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Recent attempts [20, 29, 32] use spatial locations (i.e., bounding boxes) and motion trajectories of instances, along with corresponding prompts, to create desired videos. The locations and motion trajectories of instances are extracted from real videos and are encoded as additional control conditions for joint training with the video generation models. However, existing training-based methods can only control" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "the movement trajectory [11, 41, 44, 52, 58, 65], since substantial data and training resources are required to simultaneously learn controlling layouts and movement trajectories for video generation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.155, + 0.907, + 0.397 + ], + "angle": 0, + "content": "In this paper, we develop a novel training-free approach to address this problem. Existing training-free methods are restricted in obtaining accurate location and movement effects by manipulating spatial attention maps and temporal attention maps to generate instances at specified locations [20, 29, 32]. These methods convert location signals in the pixel space provided by users to a latent space with reduced dimension for manipulating attention maps, and inevitably compromise the precision of location control and obtain inaccurate instance placement. Excessive constraints on the attention maps also disrupt the motion prior of video diffusion. Moreover, frame interaction is neglected during the manipulation, which also causes unreasonable motion effects and unnatural generated videos. Additionally, these methods also suffer from generating low-quality videos with choppy motion and insufficient details." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.399, + 0.906, + 0.672 + ], + "angle": 0, + "content": "We find that both the inaccurate instance positioning and the unnatural motion stem from the limitations of attention map manipulation, which is a compromise solution to mitigate the lack of a clear understanding of instance concepts in vanilla latent diffusion models. It introduces constraints to enforce the instances to appear at designated positions. Contrary to attention map manipulation for video generation, instance-level knowledge is usually considered for image generation. Existing grounded T2I models like GLIGEN [25] and InstanceDiffusion [39] encapsulate extensive instance-level knowledge, but they are confined to generating images rather than consistent videos. Thus, we propose to generate coarse motion videos aligning with the user input and exhibit natural motion using instance-level priors in grounded T2I models, and further refine coarse motion videos by harnessing the motion and detail priors of T2V models for motion and detail injection in a zero-shot manner to boost the motion consistency and richness of details." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.907, + 0.901 + ], + "angle": 0, + "content": "To this end, we propose a zero-shot and plug-and-play method named IM-Zero that allows to simultaneously control instance layouts and movement trajectories for video generation. IM-Zero consists of two stages, i.e., a motion generation stage that acquires a coarse motion video, and a video refinement stage to enhance the quality of the coarse motion video. In the motion generation stage, spatial location signals (e.g., bounding boxes and masks) are first augmented according to the input trajectory to obtain a sequence of layout control signals. A batch of keyframe candidates with texture consistency is then obtained by inflating and transforming the pre-trained grounded T2I model, and is used to extract motion guidance and textural guidance for generating the desired coarse motion video. In the video refinement stage, we encode the coarse motion video" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7266" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.151 + ], + "angle": 0, + "content": "into video latents, and inject the motion and detail priors extracted from the pre-trained T2V models into the video latents to improve motion consistency and enhance the richness of details, respectively." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.153, + 0.402, + 0.166 + ], + "angle": 0, + "content": "Our contributions are summarized as below." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.168, + 0.483, + 0.228 + ], + "angle": 0, + "content": "- We propose a novel framework named IM-Zero for zero-shot motion controllable video generation. IM-Zero supports various types of layout inputs and enables fine-grained motion control of the instances and their subparts." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.229, + 0.483, + 0.288 + ], + "angle": 0, + "content": "- We present motion generation that generates keyframe candidates with the layout control capability of grounded T2I models to extract motion and textural guidance for coarse motion video generation." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.289, + 0.483, + 0.363 + ], + "angle": 0, + "content": "- We develop zero-shot video refinement that injects the motion and detail priors of the pre-trained text-to-video model and text-to-image model into the latents of the coarse motion video to further enhance video motion consistency and richness of details." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.364, + 0.483, + 0.424 + ], + "angle": 0, + "content": "- We further achieve motion transfer to customize complex motion patterns using reference videos and support high-quality text-to-video generation with only text input beyond instance layout and movement trajectory controls." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.168, + 0.483, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.425, + 0.483, + 0.59 + ], + "angle": 0, + "content": "Extensive experiments demonstrate that IM-Zero outperforms existing methods in terms of video quality, interframe consistency, and the alignment of location and trajectory. Compared with existing methods, IM-Zero enjoys extra advantages of versatility in video generation, including motion control of subparts within instances and finer control with instance shapes specified by masked inputs. Furthermore, it can be adapted to motion transfer tasks to support the customized fine-grained motion patterns through reference videos and can also be adapted to high-quality text-to-video generation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.604, + 0.233, + 0.62 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.856 + ], + "angle": 0, + "content": "Grounded Text-to-image Generation. Grounded text-to-image (T2I) models aim to generate images that align with input layout information. Some works specialize in modeling layout information of the bounding box type. Notable examples include GLIGEN [25], InstanceDiffusion [39], BoxDiff [46], MultiDiffusion [1], ReCo [49], Layout-Guidance [8], MIGC [63, 64], and TraDiffusion [43]. We believe that some training-based methods such as InstanceDiffusion [39] have enabled models to grasp the concept of instances. Their capability aligns to some extent with our objective of achieving location control and trajectory control in video generation. However, these methods are limited to generating single images and cannot produce consistent videos. Therefore, we aim to leverage Grounded T2I models to benefit controllable video generation tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Location and Trajectory Control in Video Generation. Diffusion models have demonstrated exceptional capabilities in image generation [45, 61], video generation [56, 57]," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.456 + ], + "angle": 0, + "content": "3D generation [10, 27], and other visual tasks [18, 54, 60]. With the improvement in video generation quality [3, 6, 7, 13-15, 22, 37, 38, 40, 47, 62], there is a growing interest in achieving generation controllability, with a particular focus on controlling the spatial location and movement trajectories of instances. Currently, training-based methods can only achieve limited control over trajectory guidance, at the expense of requiring substantial amounts of data and computational resources, with typical examples being DragNUWA [52], DragAnything [44], DragVideo [11], Tora [58], TrackGo [65] and MotionCtrl [41]. In contrast, some training-free methods can concurrently control both spatial location and movement trajectories. For instance, Peekaboo [20] and TrailBlazer [29] enforce constraints on attention maps to position instances at specified locations. However, this approach can lead to issues such as inaccuracies in location and poor motion effects. FreeTraj [32] follows a similar attention map manipulation approach but introduces noise flow control. In addition to the aforementioned challenges, it also introduces difficulties in handling bounding boxes that undergo size changes. Therefore, we propose a zero-shot method that surpasses attention map manipulation to address the aforementioned challenges and enhance video quality and motion consistency." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.487, + 0.605, + 0.503 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.518, + 0.655, + 0.533 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.547, + 0.907, + 0.744 + ], + "angle": 0, + "content": "Grounded Text-to-image Diffusion Models. Grounded text-to-image (T2I) diffusion models [25, 39, 63, 64] are capable of generating instances at user-specified locations by incorporating the layout input as conditions for joint training. InstanceDiffusion [39] supports four types of layout input: bounding box, mask, point, and scribble. It inserts learnable UniFusion blocks between the self-attention and cross-attention layers of the T2I model. UniFusion sample the instance layout as a 2D points \\(\\mathbf{p}\\) sequence and convert it to layout token by \\(\\mathbf{g} = \\mathrm{MLP}\\left([ \\tau_{\\theta}(\\mathbf{c}), \\gamma(\\mathbf{p}) \\right]\\), where \\(\\gamma(\\cdot)\\) is a Fourier mapping and \\(\\mathbf{c}\\) is the text prompt. Then the layout token is fused with the backbone visual tokens via a learnable masked self-attention layer." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.905 + ], + "angle": 0, + "content": "Text-to-video Diffusion Models. Significant progress has also been made in Text-to-Video (T2V) diffusion models. While some T2V models are trained from scratch, the majority of them are trained based on existing T2I models. A typical example isAnimateDiff [13], which utilizes pre-trained T2I models as the spatial layer and trains additional motion modeling modules to reconstruct N frames clean video latents \\( z^{1:N} \\) from i.i.d Gaussian noise conditioned on the text prompt \\( c \\) by \\( \\mathcal{L}(\\theta) = \\mathbb{E}_{z_0^{1:N},\\epsilon ,c_t,t}\\left[\\| \\epsilon -\\hat{\\epsilon}_\\theta \\left(z_t^{1:N},c_t,t\\right)\\right] \\)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7267" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.069, + 0.868, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.274, + 0.907, + 0.318 + ], + "angle": 0, + "content": "Figure 2. The Motion Generation Stage consists of two parts: Keyframe Candidates Generation and Coarse Motion Video Generation. In the first part, we augment the user input to obtain Layout, then generate keyframe candidates Frame with an inflated grounded T2I model. In the second part, we extract motion guidance and textural guidance from Frame to generate a coarse motion video." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.335, + 0.281, + 0.35 + ], + "angle": 0, + "content": "3.2. Overall Framework" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.358, + 0.484, + 0.676 + ], + "angle": 0, + "content": "IM-Zero framework consists of two critical stages: Motion Generation Stage and Video Refinement Stage. In Section 3.1, we first briefly introduce the fundamentals of grounded T2I and T2V diffusion models. Then, we introduce the Motion Generation Stage in Section 3.3 as shown in Figure 2. This stage generates keyframe candidates and extracts guidance from them to generate a coarse motion video. Section 3.4 introduces the Video Refinement Stage as shown in Figure 3. This stage utilizes motion prior of T2V models to perform Motion Injection and utilizes the detail-generation capacity of T2I models to perform Detail Injection, further enhancing the visual quality of coarse motion video. Lastly, we also offer an option for motion transfer via reference video input, and we will explain how to easily modify IM-Zero in this case. As illustrated in Figure 1, IM-Zero achieves motion control for overall instances or subparts and allows users to further specify instance shapes as the grounded T2I model can generate keyframe candidates via masks. IM-Zero can perform motion transfer. Besides, IM-Zero allows high-quality T2V generation by refining any video via the Video Refinement Stage." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.683, + 0.322, + 0.699 + ], + "angle": 0, + "content": "3.3. Motion Generation Stage" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.902 + ], + "angle": 0, + "content": "The core task involves presenting suitable motion patterns in the video, including two key aspects: the user-desired instance spatial location and the user-desired instance movement. Previous training-free methods achieved this task by manipulating attention maps but encountered issues such as inaccurate location, poor motion effects, and low video quality, as discussed in Section 1. IM-Zero aims to achieve the decoupling of instance spatial location and instance movement to enhance the performance of both aspects. For the former, we utilize a grounded T2I model [39] to generate instances at user-specified locations. For the latter, we depart from the previous approach of imposing excessive constraints on the attention map and instead efficiently" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.337, + 0.814, + 0.352 + ], + "angle": 0, + "content": "leverage the motion priors of the T2V model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.355, + 0.907, + 0.687 + ], + "angle": 0, + "content": "Keyframe Candidates Generation. Considering the user input \\(S\\) layout signals and \\(L\\) trajectory segments, We denote the layout signals as \\(\\mathbf{Layout} = \\{s_i|i\\sim [1,2,\\dots S]\\}\\) (i.e., bounding boxes) and the movement trajectory as Traj = \\(\\{l_i|i\\sim [1,2,\\dots ,L]\\}\\). We ensure that each \\(l_{i}\\) segment has a layout signal at both endpoints, for the endpoint without a layout signal, we select the closest \\(s_i\\) from Layout and duplicate it here, obtaining the augmented layout signals \\(\\mathbf{Layout} = \\{s_i|i\\sim [1,2,\\dots,S']\\}\\). These signals are then input into the grounded T2I model to generate keyframe candidates \\(\\mathbf{Frame} = \\{f_i|i\\sim [1,2,\\dots,S']\\}\\). Since generating each image separately can lead to significant inconsistencies, following [22], we inflate the grounded T2I model and replace each self-attention in UNet with cross-frame attention (CFA). The CFA for the \\(i\\)-th keyframe candidate \\(f_{i}\\) is defined as CFA \\((Q^{i},K^{1:N_{\\mathrm{S^{\\prime}}}},V^{1:N_{\\mathrm{S^{\\prime}}}}) =\\) Softmax \\((Q^{i}(K^{1})^{T} / \\sqrt{c})V^{1}\\) where \\(Q,K\\) and \\(V\\) denote the queries, keys and values. The first candidate provides key and value for all following ones, which can be interpreted from an image editing perspective as querying relevant content and textures from the first candidate to ensure consistency in textures [5, 22]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Coarse Motion Video Generation. After obtaining Frame, we utilize these candidates to generate a coarse motion video reflecting the user's input. To ensure smooth and natural instance motion, we have avoided the previous approach of directly enforcing excessive constraints on the attention map, which could disrupt the T2V motion priors. Instead, we extract two types of guidance from Frame to guide the T2V model. (1) Motion Guidance. To ensure the fidelity of the video to the user's input, encompassing spatial location and motion patterns, we extract foreground instances from Frame and convert them into Canny edge images. Subsequently, we employ ControlNet [12, 16, 55] to guide the generation of the coarse motion video. The rationale behind our approach is elucidated as follows: firstly," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7268" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.135, + 0.07, + 0.599, + 0.146 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.134, + 0.15, + 0.599, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.609, + 0.07, + 0.865, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.25, + 0.908, + 0.321 + ], + "angle": 0, + "content": "Figure 3. In Video Refinement Stage, we diffuse coarser motion video latents and perform motion injection and detail injection in the reverse process. In motion injection, we first obtain \\( z_{t-1} \\) by \\( z_t \\), then denoise \\( z_{t-1} \\) for further \\( M \\) steps using a T2V model to \\( \\tilde{z}_{t-1-M}' \\) and combine them. We visualize the motion injection effect using CoTracker [21]. In detail injection, we first transform \\( \\tilde{z}_{t-1} \\) from T2V noise distribution to T2I noise distribution, then denoise it for further \\( N \\) steps with an inflated T2I model. The result is reversed to T2V noise distribution and combined with \\( \\tilde{z}_{t-1} \\). We visualize the detail injection effect using FFT analysis." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.34, + 0.485, + 0.598 + ], + "angle": 0, + "content": "the separation of foreground from the background is essential because instances described in the user's input manifest as foreground instances in Frame [25, 39]. Although we apply operations to enhance consistency by CFA, background distortions and abrupt changes can still occur [22]. Secondly, transforming images into Canny edge representations not only aligns with the requirements of ControlNet but also further mitigates potential inconsistencies in instance textures within Frame. (2) Textural Guidance. Use Motion Guidance only may cause textural inconsistency, hence we further perform Textural Guidance. We diffuse the first keyframe candidate \\( f_{1} \\) to form the initial noise of the T2V model. To help the guidance transmit from \\( f_{1} \\) to subsequent frames, we devise the following strategies: firstly, we employ an IP-Adapter [51] to utilize \\( f_{1} \\) as an image prompt. Moreover, we also apply CFA to the spatial layers of the T2V model." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.616, + 0.312, + 0.632 + ], + "angle": 0, + "content": "3.4. Video Refinement Stage" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.642, + 0.483, + 0.822 + ], + "angle": 0, + "content": "We have observed that while coarse motion videos may align with user-desired inputs, unsmooth motion issues and suboptimal visual quality issues such as watermark and flickering exist. To attack these issues, we introduce a zero-shot method by more efficiently leveraging the motion priors inherent in existing T2V models and the detail-generation capacity of T2I models. As shown in Figure 3, we integrate a T2V model into the reverse process of video generation. We leverage T2V motion priors to inject motion into video latents, ensuring smoother and more natural motion. Subsequently, we also insert a T2I model, leveraging its detail generation capacity to enrich video detail richness." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Given a low-quality coarse motion video as described in Section 3.3, we start by encoding it using the encoder \\(\\mathcal{E}\\) of a vanilla T2V model (i.e.,AnimateDiff [13]) and adding Gaussian noise by timestep \\(T\\) to obtain noisy video latents \\(z_{T}\\). For each denoising time step \\(t\\), we first denoise \\(z_{t}\\) using" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.34, + 0.84, + 0.355 + ], + "angle": 0, + "content": "the denoiser of the T2V model to obtain \\( z_{t - 1} \\) by:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.364, + 0.905, + 0.396 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(z _ {t - 1} \\mid z _ {t}\\right) = \\mathcal {N} \\left(z _ {t - 1}; z _ {t} - \\hat {\\epsilon} _ {\\theta} \\left(z _ {t}; c, t\\right), \\left(1 - \\alpha_ {t} ^ {2}\\right) \\mathbf {I}\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.397, + 0.905, + 0.427 + ], + "angle": 0, + "content": "where \\(c\\) is the text prompt embedding. Then we sequentially perform motion injection and detail injection on \\(z_{t - 1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.427, + 0.906, + 0.579 + ], + "angle": 0, + "content": "Motion Injection. We employ a vanilla T2V model (i.e.,AnimateDiff [13]) for motion injection as shown in Figure 3. Specifically, we denoise \\( z_{t - 1} \\) using its denoiser for further \\( M \\) steps to acquire \\( \\tilde{z}_{t - 1 - M}^{\\prime} \\), where each denoising step follows Equation (1). We observed a significant improvement in the temporal consistency of \\( \\tilde{z}_{t - 1 - M}^{\\prime} \\), resulting in very smooth motion. However, this enhancement comes at the cost of losing many detailed features. Therefore, we combine \\( \\tilde{z}_{t - 1 - M}^{\\prime} \\) with \\( z_{t - 1} \\) to enhance motion smoothness while preserving details by:" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.588, + 0.905, + 0.606 + ], + "angle": 0, + "content": "\\[\n\\tilde {z} _ {t - 1} = \\lambda_ {1} * z _ {t - 1} + (1 - \\lambda_ {1}) * \\tilde {z} _ {t - 1 - M} ^ {\\prime} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.615, + 0.864, + 0.629 + ], + "angle": 0, + "content": "where \\(\\lambda_{1}\\) is a coefficient controlling the fusion scale." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.63, + 0.906, + 0.707 + ], + "angle": 0, + "content": "Detail Injection. We then employ an inflated T2I model for detail injection as shown in Figure 3. Due to the different noise distributions in the T2I and T2V models, a noise transformation is required. To convert video latents into image latents, we directly estimate clean latents [48] \\(\\hat{z}_0\\) by:" + }, + { + "type": "equation", + "bbox": [ + 0.606, + 0.715, + 0.905, + 0.75 + ], + "angle": 0, + "content": "\\[\n\\hat {z} _ {0} = \\frac {\\tilde {z} _ {t} - \\sqrt {1 - \\alpha_ {t}} \\epsilon_ {\\theta} (\\tilde {z} _ {t} , c , t)}{\\sqrt {\\alpha_ {t}}} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.758, + 0.908, + 0.804 + ], + "angle": 0, + "content": "and we then conduct DDIM inversion [30] to derive the corresponding noisy image latents \\(\\tilde{z}_{t - 1}^{I}\\), the DDIM inversion process is defined by:" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.812, + 0.905, + 0.832 + ], + "angle": 0, + "content": "\\[\n\\tilde {z} _ {t} ^ {I} = \\sqrt {\\alpha_ {t}} \\hat {z} _ {0} + \\sqrt {1 - \\alpha_ {t}} \\cdot \\epsilon_ {\\theta} (\\tilde {z} _ {t - 1} ^ {I}, c, t) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.908, + 0.902 + ], + "angle": 0, + "content": "Subsequently, we denoise the noisy image latents with the inflated T2I model for further \\(D\\) steps to obtain \\(\\tilde{z}_{t-1-D}^{I}\\). The denoising process resembles Equation (1). As the T2I model excels in detail generation compared to the T2V" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7269" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.145, + 0.068, + 0.867, + 0.479 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.488, + 0.908, + 0.517 + ], + "angle": 0, + "content": "Figure 4. Qualitative Results of IM-Zero compared with other methods. Videos generated by IM-Zero outperform others in control accuracy, motion effects, and video quality." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.536, + 0.484, + 0.645 + ], + "angle": 0, + "content": "model, we revert \\(\\tilde{z}_{t - 1 - D}^{I}\\) back to the noise distribution of the T2V model, combining the transformed latents with \\(\\tilde{z}_{t - 1}\\) obtained from Equation (2) to integrate details. Specifically, in the reversed noise transformation process, we predict clean noise from \\(\\tilde{z}_{t - 1}\\) using Equation (3) and utilize the DDPM forward process to obtain the corresponding noisy video latents \\(\\tilde{z}_{t - 1 - D}^{V}\\). Then we fuse them by:" + }, + { + "type": "equation", + "bbox": [ + 0.155, + 0.656, + 0.483, + 0.675 + ], + "angle": 0, + "content": "\\[\nz _ {t - 1} = \\lambda_ {2} * \\tilde {z} _ {t - 1} + (1 - \\lambda_ {2}) * \\tilde {z} _ {t - 1 - D} ^ {V} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.686, + 0.483, + 0.718 + ], + "angle": 0, + "content": "where \\(\\lambda_{2}\\) controls the fusion scale and \\(z_{t - 1}\\) is used for the next denoising timestep." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.727, + 0.406, + 0.743 + ], + "angle": 0, + "content": "3.5. Adapt IM-Zero for Motion Transfer" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.903 + ], + "angle": 0, + "content": "For a more customized motion pattern provided by reference videos, IM-Zero can also be easily adapted for motion transfer tasks. Given a reference video and the target text prompt, we extract control signals (e.g., depth maps) from each frame of the reference video. Subsequently, we use these dense control signals and the target text prompt to guide the T2V model in generating a coarse motion video via ControlNet. Then, the coarse motion video undergoes video refinement stage as detailed in Section 3.4 to enhance motion consistency and video frame quality." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.535, + 0.64, + 0.553 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.561, + 0.698, + 0.577 + ], + "angle": 0, + "content": "4.1. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.583, + 0.907, + 0.794 + ], + "angle": 0, + "content": "Regarding the comparative methods, we primarily selected Peekaboo [20], TrailBlazer [29], and MotionBooth [42], which generate videos by specifying boxes and trajectories. Furthermore, to further illustrate the effectiveness of our approach, we also included T2V-Zero [22], which offers rough guidance by specifying translation vectors. To ensure a fair comparison, following the experimental settings of [29], all methods utilize the same layout input, trajectory input, and text prompt input. Additional results of controllable video generation are available in the supplementary material. We also compared with FreeTraj [32], but as it additionally requires inputting boxes with fixed sizes across frames to make noise flow, we place the comparison in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "As shown in Figure 4, our method significantly outperforms other approaches in terms of alignment with the input, motion effects, and video quality. In terms of alignment with the input, T2V-Zero exhibits almost no resemblance to the input, while the two methods based on attention map manipulation, Peekaboo, and TrailBlazer, sometimes roughly match the input location and trajectory but" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7270" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.077, + 0.908, + 0.107 + ], + "angle": 0, + "content": "Table 1. Quantitative results of IM-Zero compared with other methods on DAVIS-17 [31] dataset and GOT10k [17] dataset. IM-Zero achieves SOTA in all metrics, including metrics measuring video quality, frame consistency, and control accuracy." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.11, + 0.908, + 0.238 + ], + "angle": 0, + "content": "
MethodDAVIS-17GOT10k
FVD (↓)KID (↓)CLIPSim (↑)mIoU (↑)CD (↓)FVD (↓)KID (↓)CLIPSim (↑)mIoU (↑)CD (↓)
T2V-Zero [22]5004.0030.3795.120.190.344335.6924.7295.410.120.32
Peekaboo [20]4264.7026.4897.330.190.464960.4919.0598.680.110.34
TrailBlazer [29]4593.3726.0498.100.210.294695.1816.1897.800.130.40
MotionBooth [42]4192.1926.8596.640.220.364311.7116.0597.470.140.42
IM-Zero (Ours)4070.0125.7598.910.270.264187.2715.9299.040.210.27
" + }, + { + "type": "table_caption", + "bbox": [ + 0.286, + 0.244, + 0.712, + 0.26 + ], + "angle": 0, + "content": "Table 2. Ablation on different components in Motion Generation Stage." + }, + { + "type": "table", + "bbox": [ + 0.169, + 0.263, + 0.825, + 0.371 + ], + "angle": 0, + "content": "
MethodFVD (↓)KID (↓)CLIPSim (↑)mIoU (↑)CD (↓)
w/o 1st CFA4320.0726.8498.840.260.29
w/o 2nd CFA4528.0928.0898.780.210.31
w/o init noise4786.1726.6998.820.260.31
w/o IP-Adapter4549.9427.7598.870.260.28
IM-Zero (Ours)4070.0125.7598.910.270.26
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.388, + 0.482, + 0.495 + ], + "angle": 0, + "content": "still lack precision. Regarding motion effects, Peekaboo and TrailBlazer generate videos with unnatural motion, resulting in strange discontinuities between frames. Concerning video quality, Peekaboo and TrailBlazer exhibit low image quality, lack detail, and introduce artifacts. We also provide a video version of Figure 4 in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.501, + 0.285, + 0.518 + ], + "angle": 0, + "content": "4.2. Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.524, + 0.483, + 0.811 + ], + "angle": 0, + "content": "Experiment Settings. For datasets, we utilize the training set and validation set of the DAVIS-17 dataset [31] and the test set of the GOT10k dataset [17]. We use BLIP-2 [24] to generate text prompts and GroundingDINO [28] to extract bounding boxes from the first and last frame as inputs of all the methods. After eliminating videos where no bounding boxes are extracted, we randomly select 50 videos and 70 videos from the rest videos of DAVIS-17 [31] and GOT10k [17]. For our method, in the first stage we use InstanceDiffusion [39] as grounded T2I model and SparseCtrl [12] as ControlNet, in the second stage we use Stable Diffusion V1.5 [34] as T2I model, on both stage we useAnimateDiff [13] as T2V model. For the evaluation metrics, we employ mean Intersection over Union (mIoU) and Centroid Distance (CD) to assess control alignment, Fréchet Video Distance (FVD) [35] and Kernel Inception Distance (KID) [2] to evaluate video quality, and CLIP similarities (CLIPSim) [33] to assess frame-to-frame consistency. Detailed settings are available in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Metrics Evaluation Results. We conducted evaluations on the aforementioned metrics on the DAVIS-17 dataset [31] and the GOT10k dataset [17]. As shown in Table 1, our method achieved state-of-the-art (SOTA) performance compared to several other methods. Firstly, in terms of the alignment between the generated videos and user inputs," + }, + { + "type": "table_caption", + "bbox": [ + 0.523, + 0.387, + 0.895, + 0.402 + ], + "angle": 0, + "content": "Table 3. Motion transfer results compared with other methods." + }, + { + "type": "table", + "bbox": [ + 0.514, + 0.412, + 0.91, + 0.5 + ], + "angle": 0, + "content": "
MethodMotion Fidelity (↑)Imaging Quality (↑)
MotionClone [26]74.50.72
MOFT [23]51.20.71
MotionDirector [59]75.50.68
IM-Zero (Ours)63.70.73
" + }, + { + "type": "table_caption", + "bbox": [ + 0.556, + 0.513, + 0.862, + 0.528 + ], + "angle": 0, + "content": "Table 4. T2V results compared with UNet baseline." + }, + { + "type": "table", + "bbox": [ + 0.53, + 0.538, + 0.891, + 0.598 + ], + "angle": 0, + "content": "
MethodImaging Quality (↑)CLIPSim (↑)
AnimateDiff [13]0.6799.05
IM-Zero (Ours)0.7499.73
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.625, + 0.907, + 0.822 + ], + "angle": 0, + "content": "our method demonstrates a significant lead. Compared to the previous best methods, our approach achieved an improvement of \\(28.6\\%\\) and \\(61.5\\%\\) in mIoU, and reductions of \\(10.4\\%\\) and \\(15.6\\%\\) in CD on the two datasets, indicating that our method offers more accurate location control and motion trajectory control compared to previous methods based on attention map manipulation. Secondly, regarding frame-to-frame consistency in the videos, thanks to our enhanced consistency techniques and motion injection methods, our approach also reaches the SOTA performance. Moreover, in terms of video quality, our method significantly outperforms others, with FVD reduced to 4070.01 and 4187.27, and KID reduced to 25.75 and 15.92, respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.833, + 0.761, + 0.85 + ], + "angle": 0, + "content": "4.3. Versatile Capacity Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Our method possesses versatile capacity as shown in Figure 5. First, our method can control the subparts of instances, achieving effects such as head turning and hand" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7271" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.138, + 0.093, + 0.158, + 0.109 + ], + "angle": 0, + "content": "#" + }, + { + "type": "image", + "bbox": [ + 0.164, + 0.064, + 0.867, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.347, + 0.908, + 0.377 + ], + "angle": 0, + "content": "Figure 5. Besides box-level overall instance motion control, IM-Zero can: (a) Perform instance subpart control. (b) Specify instance shape by mask. (c) Perform motion transfer from the reference video(left) to the target video(right). (d) Perform high-quality T2V generation." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.402, + 0.482, + 0.506 + ], + "angle": 0, + "content": "movement. Second, besides bounding boxes, we also support specifying instance shapes more precisely using masks. Third, for cases where users provide a reference video, our method can perform motion transfer. Finally, users can directly provide text inputs for high-quality T2V generation. For motion transfer and T2V generation, we compare IM-Zero with other methods as follows." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.508, + 0.483, + 0.628 + ], + "angle": 0, + "content": "Motion Transfer. We compare with MOFT [23], MotionClone [26], and MotionDirector [59] using the open-sourced data from MotionClone [26]. For quantitative results, we use Motion Fidelity [50] and Imaging Quality [19] as metrics following MOFT [23]. Table 3 shows that our method is competitive in Motion Fidelity [50] and is the best in Imaging Quality [19]. Qualitative results and other details are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.629, + 0.483, + 0.749 + ], + "angle": 0, + "content": "T2V Generation. As we use UNet-based models, we compare T2V with UNet baselineAnimateDiff [13]. For quantitative results, we use Imaging Quality [19] and CLIP-Sim [33] as metrics. As there are no ground-truth videos, we generate videos using 50 prompts randomly generated by GPT. Table 4 shows that our method is better in Imaging Quality [19] and CLIPSim [33]. Refer to the supplementary material for qualitative results and other details." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.75, + 0.483, + 0.825 + ], + "angle": 0, + "content": "Other Capacities. The proposed method also supports multi-instance motion control, and the decomposition of our pipeline allows autoregressive usage of the second stage to enhance temporal consistency, as elaborated in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.834, + 0.258, + 0.849 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.871 + ], + "angle": 0, + "content": "Ablation on different components in Motion Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.482, + 0.902 + ], + "angle": 0, + "content": "Stage. We ablate on different components in Section 3.3. As depicted in Table 2, these methods improve the quality" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.402, + 0.907, + 0.432 + ], + "angle": 0, + "content": "of the final videos and also enhance the alignment with user inputs. Refer to the supplementary material for details." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.436, + 0.907, + 0.482 + ], + "angle": 0, + "content": "Ablation on different components in Video Refinement Stage. We also conduct ablation studies on Video Refinement Stage. Refer to the supplementary material for details." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.506, + 0.766, + 0.523 + ], + "angle": 0, + "content": "5. Conclusion and Limitation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.535, + 0.907, + 0.717 + ], + "angle": 0, + "content": "We introduce IM-Zero, a novel training-free framework for instance-level motion controllable video generation using diffusion models. We leverage grounded text-to-image models to generate keyframe candidates and extract motion guidance and textural guidance from them to generate videos that align with user inputs. To further enhance video motion and quality, we also introduce a zero-shot method for motion injection and detail injection. Additionally, our method can control the subpart motion of instances, use masks to specify instance shapes, perform motion transfer from reference videos, and directly support high-quality text-to-video generation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Limitation. The zero-shot features impose certain limitations on our method. Firstly, as zero-shot methods depend on pre-trained model weights, the videos generated by our method are inherently constrained by the pre-trained models employed. For example, the alignment with layout input is influenced by the grounded T2I model and video ControlNet, the video quality is affected by the T2V model, which can be improved by using more advanced pre-trained models. Secondly, the temporal consistency of videos generated by zero-shot methods is inherently weaker compared to training-based methods, which can be improved by incorporating some tuning components to our method." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7272" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.221 + ], + "angle": 0, + "content": "This work was supported in part by the National Natural Science Foundation of China under Grant 62125109, Grant 62431017, Grant U24A20251, Grant 62320106003, Grant 62371288, Grant 62401357, Grant 62401366, Grant 62301299, Grant 62120106007, and in part by the Program of Shanghai Science and Technology Innovation Project under Grant 24BC3200800." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.233, + 0.188, + 0.249 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.258, + 0.483, + 0.314 + ], + "angle": 0, + "content": "[1] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. MultiDiffusion: Fusing diffusion paths for controlled image generation. In Proceedings of the 40th International Conference on Machine Learning, pages 1737-1752, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.316, + 0.484, + 0.369 + ], + "angle": 0, + "content": "[2] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying MMD GANs. In *The Sixth International Conference on Learning Representations*, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.371, + 0.483, + 0.453 + ], + "angle": 0, + "content": "[3] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.455, + 0.483, + 0.523 + ], + "angle": 0, + "content": "[4] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. OpenAI Blog, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.525, + 0.483, + 0.594 + ], + "angle": 0, + "content": "[5] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. MasaCtrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22560–22570, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.595, + 0.483, + 0.676 + ], + "angle": 0, + "content": "[6] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, Chao Weng, and Ying Shan. VideoCrafter1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.678, + 0.483, + 0.76 + ], + "angle": 0, + "content": "[7] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. VideoCrafter2: Overcoming data limitations for high-quality video diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7310-7320, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.761, + 0.484, + 0.817 + ], + "angle": 0, + "content": "[8] Minghao Chen, Iro Laina, and Andrea Vedaldi. Training-free layout control with cross-attention guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5343–5353, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.818, + 0.483, + 0.873 + ], + "angle": 0, + "content": "[9] Weifeng Chen, Yatai Ji, Jie Wu, Hefeng Wu, Pan Xie, Jiashi Li, Xin Xia, Xuefeng Xiao, and Liang Lin. Control-A-Video: Controllable text-to-video generation with diffusion models. arXiv preprint arXiv:2305.13840, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[10] Yabo Chen, Jiemin Fang, Yuyang Huang, Taoran Yi, Xiaopeng Zhang, Lingxi Xie, Xinggang Wang, Wenrui Dai," + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.258, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.148 + ], + "angle": 0, + "content": "Hongkai Xiong, and Qi Tian. Cascade-Zero123: One image to highly consistent 3D with self-prompted nearby views. In Proceedings of the 18th European Conference on Computer Vision, pages 311-330, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.906, + 0.204 + ], + "angle": 0, + "content": "[11] Yufan Deng, Ruida Wang, Yuhao Zhang, Yu-Wing Tai, and Chi-Keung Tang. DragVideo: Interactive drag-style video editing. In Proceedings of the 18th European Conference on Computer Vision, pages 183–199, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.205, + 0.906, + 0.272 + ], + "angle": 0, + "content": "[12] Yuwei Guo, Ceyuan Yang, Anyi Rao, Maneesh Agrawala, Dahua Lin, and Bo Dai. SparseCtrl: Adding sparse controls to text-to-video diffusion models. In Proceedings of the 18th European Conference on Computer Vision, pages 330-348, 2024. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.905, + 0.356 + ], + "angle": 0, + "content": "[13] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. AnimateDiff:Animate your personalized text-to-image diffusion models without specific tuning. In The Twelfth International Conference on Learning Representations, 2024. 2, 3, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.358, + 0.905, + 0.427 + ], + "angle": 0, + "content": "[14] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P. Kingma, Ben Poole, Mohammad Norouzi, David J. Fleet, and Tim Salimans. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.905, + 0.483 + ], + "angle": 0, + "content": "[15] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. In Advances in Neural Information Processing Systems 35, pages 8633-8646, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.905, + 0.538 + ], + "angle": 0, + "content": "[16] Zhihao Hu and Dong Xu. VideoControlNet: A motion-guided video-to-video translation framework by using diffusion model with ControlNet. arXiv preprint arXiv:2307.14073, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.54, + 0.905, + 0.594 + ], + "angle": 0, + "content": "[17] Lianghua Huang, Xin Zhao, and Kaiqi Huang. GOT-10k: A large high-diversity benchmark for generic object tracking in the wild. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(5):1562-1577, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.596, + 0.905, + 0.664 + ], + "angle": 0, + "content": "[18] Yuyang Huang, Yabo Chen, Yuchen Liu, Xiaopeng Zhang, Wenrui Dai, Hongkai Xiong, and Qi Tian. DomainFusion: Generalizing to unseen domains with latent diffusion models. In Proceedings of the 18th European Conference on Computer Vision, pages 480-498, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.665, + 0.905, + 0.761 + ], + "angle": 0, + "content": "[19] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, Yaohui Wang, Xinyuan Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. VBench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.762, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[20] Yash Jain, Anshul Nasery, Vibhav Vineet, and Harkirat Behl. PEEKABOO: Interactive video generation via masked-diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8079-8088, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[21] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. CoTracker: It is better to track together. In Proceedings of the 18th European Conference on Computer Vision, pages 18-35, 2024. 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7273" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.176 + ], + "angle": 0, + "content": "[22] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2Video-Zero: Text-to-image diffusion models are zero-shot video generators. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15954-15964, 2023. 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.178, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[23] Karlo Koledić, Igor Cvišić, Ivan Marković, and Ivan Petrović. MOFT: Monocular odometry based on deep depth and careful feature selection and tracking. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 6175-6181. IEEE, 2023. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.318 + ], + "angle": 0, + "content": "[24] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In Proceedings of the 40th International Conference on Machine Learning, pages 19730–19742, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.482, + 0.401 + ], + "angle": 0, + "content": "[25] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. GLIGEN: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22511-22521, 2023. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.404, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[26] Pengyang Ling, Jiazi Bu, Pan Zhang, Xiaoyi Dong, Yuhang Zang, Tong Wu, Huaian Chen, Jiaqi Wang, and Yi Jin. MotionClone: Training-free motion cloning for controllable video generation. In The Thirteenth International Conference on Learning Representations, 2025. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.482, + 0.545 + ], + "angle": 0, + "content": "[27] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9298–9309, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.547, + 0.482, + 0.629 + ], + "angle": 0, + "content": "[28] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, and Lei Zhang. Grounding DINO: Marrying DINO with grounded pre-training for open-set object detection. In Proceedings of the 18th European Conference on Computer Vision, pages 38-55, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.631, + 0.482, + 0.685 + ], + "angle": 0, + "content": "[29] Wan-Duo Kurt Ma, John P Lewis, and W Bastiaan Kleijn. TrailBlazer: Trajectory control for diffusion-based video generation. In SIGGRAPH Asia 2024, number 97, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.688, + 0.482, + 0.757 + ], + "angle": 0, + "content": "[30] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.759, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[31] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 DAVIS challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[32] Haonan Qiu, Zhaoxi Chen, Zhouxia Wang, Yingqing He, Menghan Xia, and Ziwei Liu. FreeTraj: Tuning-free trajectory control in video diffusion models. arXiv preprint arXiv:2406.16863, 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry," + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.162 + ], + "angle": 0, + "content": "Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763, 2021. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.907, + 0.232 + ], + "angle": 0, + "content": "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.233, + 0.905, + 0.288 + ], + "angle": 0, + "content": "[35] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.289, + 0.905, + 0.356 + ], + "angle": 0, + "content": "[36] Cong Wang, Jiaxi Gu, Panwen Hu, Haoyu Zhao, Yuanfan Guo, Jianhua Han, Hang Xu, and Xiaodan Liang. EasyControl: Transfer ControlNet to video diffusion for controllable generation and interpolation. arXiv preprint arXiv:2408.13005, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.358, + 0.905, + 0.411 + ], + "angle": 0, + "content": "[37] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. ModelScope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.413, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[38] Weimin Wang, Jiawei Liu, Zhijie Lin, Jiangqiao Yan, Shuo Chen, Chetwin Low, Tuyen Hoang, Jie Wu, Jun Hao Liew, Hanshu Yan, Daquan Zhou, and Jiashi Feng. MagicVideoV2: Multi-stage high-aesthetic video generation. arXiv preprint arXiv:2401.04468, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.905, + 0.552 + ], + "angle": 0, + "content": "[39] Xudong Wang, Trevor Darrell, Sai Saketh Rambhatla, Rohit Girdhar, and Ishan Misra. InstanceDiffusion: Instance-level control for image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6232-6242, 2024. 2, 3, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.554, + 0.905, + 0.649 + ], + "angle": 0, + "content": "[40] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, Yuwei Guo, Tianxing Wu, Chenyang Si, Yuming Jiang, Cunjian Chen, Chen Change Loy, Bo Dai, Dahua Lin, Yu Qiao, and Ziwei Liu. LaVie: High-quality video generation with cascaded latent diffusion models. International Journal of Computer Vision, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.651, + 0.905, + 0.719 + ], + "angle": 0, + "content": "[41] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. MotionCtrl: A unified and flexible motion controller for video generation. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.721, + 0.905, + 0.788 + ], + "angle": 0, + "content": "[42] Jianzong Wu, Xiangtai Li, Yanhong Zeng, Jiangning Zhang, Qianyu Zhou, Yining Li, Yunhai Tong, and Kai Chen. MotionBooth: Motion-aware customized text-to-video generation. In Advances in Neural Information Processing Systems 37, pages 34322-34348, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[43] Mingrui Wu, Oucheng Huang, Jiayi Ji, Jiale Li, Xinyue Cai, Huafeng Kuang, Jianzhuang Liu, Xiaoshuai Sun, and Rongrong Ji. TraDiffusion: Trajectory-based training-free image generation. arXiv preprint arXiv:2408.09739, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.847, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[44] Weijia Wu, Zhuang Li, Yuchao Gu, Rui Zhao, Yefei He, David Junhao Zhang, Mike Zheng Shou, Yan Li, Tingting Gao, and Di Zhang. DragAnything: Motion control for anything using entity representation. In Proceedings of the 18th" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7274" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "European Conference on Computer Vision, pages 331-348, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[45] Enze Xie, Junsong Chen, Junyu Chen, Han Cai, Haotian Tang, Yujun Lin, Zhekai Zhang, Muyang Li, Ligeng Zhu, Yao Lu, and Song Han. SANA: Efficient high-resolution image synthesis with linear diffusion transformers. In The Thirteenth International Conference on Learning Representations, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.483, + 0.276 + ], + "angle": 0, + "content": "[46] Jinheng Xie, Yuexiang Li, Yawen Huang, Haozhe Liu, Wentian Zhang, Yefeng Zheng, and Mike Zheng Shou. BoxDiff: Text-to-image synthesis with training-free box-constrained diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7452-7461, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[47] Yichao Yan, Zanwei Zhou, Zi Wang, Jingnan Gao, and Xiaokang Yang. DialogueNeRF: Towards realistic avatar face-to-face conversation video generation. Visual Intelligence, 2 (1):24, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.483, + 0.39 + ], + "angle": 0, + "content": "[48] Serin Yang, Hyunmin Hwang, and Jong Chul Ye. Zero-shot contrastive loss for text-guided diffusion image style transfer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22873-22882, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.392, + 0.483, + 0.473 + ], + "angle": 0, + "content": "[49] Zhengyuan Yang, Jianfeng Wang, Zhe Gan, Linjie Li, Kevin Lin, Chenfei Wu, Nan Duan, Zicheng Liu, Ce Liu, Michael Zeng, and Lijuan Wang. ReCo: Region-controlled text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14246-14255, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.483, + 0.546 + ], + "angle": 0, + "content": "[50] Danah Yatim, Rafail Fridman, Omer Bar-Tal, Yoni Kasten, and Tali Dekel. Space-time diffusion features for zero-shot text-driven motion transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8466-8476, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.483, + 0.601 + ], + "angle": 0, + "content": "[51] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. IP-Adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.483, + 0.659 + ], + "angle": 0, + "content": "[52] Shengming Yin, Chenfei Wu, Jian Liang, Jie Shi, Houqiang Li, Gong Ming, and Nan Duan. DragNUWA: Fine-grained control in video generation by integrating text, image, and trajectory. arXiv preprint arXiv:2308.08089, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.483, + 0.731 + ], + "angle": 0, + "content": "[53] Yan Zeng, Guoqiang Wei, Jiani Zheng, Jiaxin Zou, Yang Wei, Yuchen Zhang, and Hang Li. Make pixels dance: High-dynamic video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8850-8860, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.483, + 0.814 + ], + "angle": 0, + "content": "[54] Junyi Zhang, Charles Herrmann, Junhwa Hur, Luisa Polania Cabrera, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. A tale of two features: Stable diffusion complements DINO for zero-shot semantic correspondence. In Advances in Neural Information Processing Systems 36, pages 45533-45547, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.483, + 0.872 + ], + "angle": 0, + "content": "[55] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[56] Yabo Zhang, Yuxiang Wei, Dongsheng Jiang, Xiaopeng Zhang, Wangmeng Zuo, and Qi Tian. ControlVideo:" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "Training-free controllable text-to-video generation. In The Twelfth International Conference on Learning Representations, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[57] Yabo Zhang, Yuxiang Wei, Xianhui Lin, Zheng Hui, Peiran Ren, Xuansong Xie, Xiangyang Ji, and Wangmeng Zuo. VideoElevator: elevating video generation quality with versatile text-to-image diffusion models. arXiv preprint arXiv:2403.05438, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.206, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[58] Zhenghao Zhang, Junchao Liao, Menghao Li, Long Qin, and Weizhi Wang. Tora: Trajectory-oriented diffusion transformer for video generation. arXiv preprint arXiv:2407.21705, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.344 + ], + "angle": 0, + "content": "[59] Rui Zhao, Yuchao Gu, Jay Zhangjie Wu, David Junhao Zhang, Jia-Wei Liu, Weijia Wu, Jussi Keppo, and Mike Zheng Shou. MotionDirector: Motion customization of text-to-video diffusion models. In Proceedings of the 18th European Conference on Computer Vision, pages 273–290, 2024. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.347, + 0.905, + 0.416 + ], + "angle": 0, + "content": "[60] Wenliang Zhao, Yongming Rao, Zuyan Liu, Benlin Liu, Jie Zhou, and Jiwen Lu. Unleashing text-to-image diffusion models for visual perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5706-5716, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.418, + 0.905, + 0.487 + ], + "angle": 0, + "content": "[61] Wendi Zheng, Jiayan Teng, Zhuoyi Yang, Weihan Wang, Ji-dong Chen, Xiaotao Gu, Yuxiao Dong, Ming Ding, and Jie Tang. CogView3: Finer and faster text-to-image generation via relay diffusion. In Proceedings of the 18th European Conference on Computer Vision, pages 1-22, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.488, + 0.905, + 0.543 + ], + "angle": 0, + "content": "[62] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. MagicVideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.545, + 0.905, + 0.613 + ], + "angle": 0, + "content": "[63] Dewei Zhou, You Li, Fan Ma, Xiaoting Zhang, and Yi Yang. MIGC: Multi-instance generation controller for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6818-6828, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.615, + 0.905, + 0.67 + ], + "angle": 0, + "content": "[64] Dewei Zhou, You Li, Fan Ma, Zongxin Yang, and Yi Yang. MIGC++: Advanced multi-instance generation controller for image synthesis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 47(3):1714-1728, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.672, + 0.905, + 0.727 + ], + "angle": 0, + "content": "[65] Haitao Zhou, Chuang Wang, Rui Nie, Jinxiao Lin, Dongdong Yu, Qian Yu, and Changhu Wang. TrackGo: A flexible and efficient method for controllable video generation. arXiv preprint arXiv:2408.11475, 2024. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.727 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7275" + } + ] +] \ No newline at end of file diff --git a/2025/IM-Zero_ Instance-level Motion Controllable Video Generation in a Zero-shot Manner/a669fc0a-17c9-4001-97e0-7737b0fc24de_origin.pdf b/2025/IM-Zero_ Instance-level Motion Controllable Video Generation in a Zero-shot Manner/a669fc0a-17c9-4001-97e0-7737b0fc24de_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c4c09798fc6a7aeb2aceb02ed77606f4a53b2334 --- /dev/null +++ b/2025/IM-Zero_ Instance-level Motion Controllable Video Generation in a Zero-shot Manner/a669fc0a-17c9-4001-97e0-7737b0fc24de_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1e24878f38495d12c5f2c625b4efb021ca504bc0280bbb4d99b16bc4cf9f67ae +size 7416424 diff --git a/2025/IM-Zero_ Instance-level Motion Controllable Video Generation in a Zero-shot Manner/full.md b/2025/IM-Zero_ Instance-level Motion Controllable Video Generation in a Zero-shot Manner/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5da9f6d30d9088753a61cf97dfcbf0ed3efa12c8 --- /dev/null +++ b/2025/IM-Zero_ Instance-level Motion Controllable Video Generation in a Zero-shot Manner/full.md @@ -0,0 +1,376 @@ +# IM-Zero: Instance-level Motion Controllable Video Generation in a Zero-shot Manner + +Yuyang Huang $^{1*}$ , Yabo Chen $^{1*}$ , Li Ding $^{1}$ , Xiaopeng Zhang $^{2\dagger}$ , Wenrui Dai $^{1\dagger}$ , Junni Zou $^{1\dagger}$ , Hongkai Xiong $^{1}$ , Qi Tian $^{2}$ $^{1}$ Shanghai Jiao Tong University, Shanghai, China $^{2}$ Huawei Inc., Shenzhen, China +{huangyuyang, chenyabo, wangzhongren, daiwenrui, zoujunni, xionghongkai}@sjtu.edu.cn +zxphistory@gmail.com, tian.qil@huawei.com + +![](images/ddf1302253eba6ebfb247c2bd97298b3e8d639c64388c2685df0852a75223205.jpg) +(a) A person riding a snowmobile down a snowy mountain. + +![](images/8720deab3e64e069231ca4b8ac39f369ff2d3bc4db5d1ce641d85042454a9adb.jpg) + +![](images/c19e69c270bd3dd827a7ad24217b440c964ba5972c41ae2d997cfd00739a40af.jpg) + +![](images/f9c922b45beea473a69c469cce46743cf4cd03b3e395510f4d5b8340718b1a7a.jpg) + +![](images/94861a0f6fbd7d8d9f210d344e24bdb4f48b5ccadddb6b92d56cd71be05bf072.jpg) + +![](images/4add69d4bcbcf5355b1167a5620c4ed2e0c6277cee9dba226dc972508958ba0c.jpg) + +![](images/c3c320c402799d3ca87ba4f23b2455f88a44baea6db7b6c2ae28aa18323ee175.jpg) + +![](images/cfa9cfeade01f1bf9b8b3793dc427b22403aeff2f25ee3024b74293740eab982.jpg) + +![](images/994cd997a8ce11e979216a9bb83571de2b21ad98f7b49e6367c6231f63367379.jpg) + +![](images/751983bbda54f3545057a0fe6f37bcda25d2012fa5b99865d0baac58aa5a3af0.jpg) +(b) A chef with a white hat preparing a gourmet meal. + +![](images/d0f42cdbfd9cac8254a7da54da51b0b2eeb091ed5ea3d6cfe48824531add0b71.jpg) + +![](images/eec11408165b12e44f08c8405468b482b7cecbcd12a2fdaab0b871d9260dde5d.jpg) + +![](images/d3b730321a091d650ea5372a859eb766bc5d3b562cf9f478aa106b0ffe925abe.jpg) + +![](images/8023b8eaa4ced1db4f86bdc0c490bdf103eb093f853227682274d01b00fc2a11.jpg) + +![](images/df09c6f36ec505e9d54a7a50f838df7d5a68a0b856cbba7dde0a387b2a8b0f37.jpg) + +![](images/4e108b5dd1091674a527545dc0da32d3d0556b2ebcbedfe87ec18e6efa74d444.jpg) + +![](images/f2db678dfc32d0a8603d3547093bec5d13df6c9a8c0e35427d47b81ba10d7208.jpg) + +![](images/56f713d1351b73f7e499e3ce9fc2d2200489a40d199bbb831756b51fb4456ee1.jpg) + +![](images/26b41eca45b63bd8c695130f674f01b8d3ea93facc29603a94ddc3daf7f9c720.jpg) +(c) A snow leopard is walking on the side of a cliff. + +![](images/b8b14fab5ed06744e00a6b10345d3c82f6f257f194537c5dff405277aa69a734.jpg) + +![](images/ab704978e737c61b7242c82a3e201f987a4d4125c050fb0f8aab2231efa59f2d.jpg) + +![](images/b2d0a23fe3c75f4a3b2a1e2c916ad27de42e92de3b8f93c27b01faea61cd9eee.jpg) + +![](images/13069143961c353f004a316d264531ca0efdaab5df16af5c6c0f37b08bca01fb.jpg) + +![](images/5c047608058f83ba1d346c30e6cdebaf97f82dc4c2a035ff67d3e80484b5d678.jpg) + +![](images/ae4eec9ce99811c00b66bb2298a5d2d7e3749885df9f0edf73766ccaff47f5bf.jpg) + +![](images/56b5c5aec26818581508635bec94422b854f827ad3669c3ce6462aa61c91a700.jpg) + +![](images/0f747f8280de2d9dd608531482e32f6f9ddf71ceddc543b6a24ab2a8b7bed826.jpg) + +![](images/a7c5c0224f9fc4cea4f27532b909aa22f9ba1ba6ac247c47da9693cb3b3e2343.jpg) +(d) A lemur walking on a xylophone, in a tropical rainforest. + +![](images/7ccc0f4ca6dac6e3014b4898e91026ea62a0e883d622bbb6fc81e8ee139d5f21.jpg) + +![](images/15627c4a157921f2733760f2014733f111a777a81eae020a3b862c695ddfab69.jpg) + +![](images/a318956bf9177ef3fd2b5aed094ce1c3b5cb804deac25da1fb6a29d75b93ef7d.jpg) + +![](images/61fa9e32ac0c0dbf5ce834cbbe3c598b6242cb1c69409092295d66bef9e942fb.jpg) + +![](images/70c75a3b854464b13d5d2394cec35b8b447a06981cc275350815ae1ac66bca1d.jpg) + +![](images/181b8d1051f05bacb30f0967b30341b7fc0a3510eae90c0fe6e04a9699e0697e.jpg) + +![](images/7c38996f49cb0016440a54684837cb9765c5fc8e6c26df167d1e6f630871548e.jpg) + +![](images/b0254591e78a5349e840a51972a571ef830b059b7d73f37410fad318fa7fe790.jpg) + +![](images/c6e4913e12352c0c54f7e5b3b19fb17149ed05d06ca61c3c52ea7f1864a7ef4c.jpg) + +![](images/550195189daa5648d4bba597da08af01c98116d3cad69f35bf35880fb5e96e7a.jpg) + +![](images/94fb323760576ac2860c9b357aeba66cde2f3a5150b78451f62597c429f11a7d.jpg) +(e) A snowman in a forest caught in a burst of flames. +Figure 1. IM-Zero introduces a zero-shot framework for instance-level motion controllable video generation. IM-Zero allows users to: (a) Control overall instance motion. (b) Control instance subpart motion. (c) Control instance motion and specify instance shape via mask. (d) Perform motion transfer by reference video. (e) Perform high-quality text-to-video generation. + +![](images/d05828622d257352665a2b9033966df61fb65749578e567143e73490968b13b2.jpg) + +![](images/d44c0c7d45a6d1bdd8e8bfabdd7bbe88df104a1b213e8598c99c23f32a3596dd.jpg) + +![](images/fea318659d0bd51f124ad1fb421dce5436313f2111de241a2425f631a13c7aaf.jpg) + +![](images/461c6585ef807732b233b444866892bc336966b622e5d3897c09936ddc9266f3.jpg) + +![](images/31bda9be05576e5ce87ddbb0ae85dcdfdeb4ed3ef8f7b1c7b8975f813a8e582e.jpg) + +![](images/a096790f957ad8d8cde071b5e09dcb9e215445ea05b42377c3314b3fde046092.jpg) + +# Abstract + +Controllability of video generation has been recently concerned in addition to the quality of generated videos. The main challenge to controllable video generation is to synthesize videos based on user-specified instance spatial locations and movement trajectories. However, existing methods suffer from a dilemma between the resource con + +sumption, generation quality, and user controllability. As an efficient alternative to prohibitive training-based video generation, existing zero-shot video generation methods cannot generate high-quality and motion-consistent videos under the control of layouts and movement trajectories. In this paper, we propose a novel zero-shot method named IM-Zero that ameliorates instance-level motion controllable video generation with enhanced control accuracy, motion consistency, and richness of details to address this prob + +lem. Specifically, we first present a motion generation stage that extracts motion and textural guidance from keyframe candidates from pre-trained grounded text-to-image model to generate the desired coarse motion video. Subsequently, we develop a video refinement stage that injects the motion priors of pre-trained text-to-video models and detail priors of pre-trained text-to-image models into the latents of coarse motion videos to further enhance video motion consistency and richness of details. To our best knowledge, IM-Zero is the first to simultaneously achieve high-quality video generation and allow to control both layouts and movement trajectories in a zero-shot manner. Extensive experiments demonstrate that IM-Zero outperforms existing methods in terms of video quality, inter-frame consistency, and the alignment of location and trajectory. Furthermore, compared with existing methods, IM-Zero enjoys extra advantages of versatility in video generation, including motion control of subparts within instances, finer control of specifying instance shapes via masks, and more difficult tasks of motion transfer for customizing fine-grained motion patterns through reference videos and high-quality text-to-video generation. + +# 1. Introduction + +Demands for enhanced controllability of text-to-video (T2V) generation are booming with the increasing quality of generated videos [3, 6, 7, 13-15, 22, 37, 38, 40, 62]. One of the critical problems lie in the precise control over the spatial layout and trajectories of instances within the generated videos. Existing methods fail to achieve fine-grained control by solely relying on text prompts to steer pre-trained T2V models, even employing fundamental T2V models with large-scale pre-training like Sora [4], KLING, and Doubao-PixelDance [53]. + +Existing methods usually extend ControlNet [55] for controllable video generation with frame-by-frame fine-grained control signals, including Canny, depth, and HED signals [9, 16, 36, 56]. To obtain dense control signals, these methods impose high demands on the input end and are limited in practical applications. One approach is to extract control signals from reference videos but it is hard to obtain reference videos perfectly matching the desired motion patterns. The other alternative requires manually crafting the dense control signals. However, it is also non-trivial to craft frame-by-frame control signals without expertise. + +Recent attempts [20, 29, 32] use spatial locations (i.e., bounding boxes) and motion trajectories of instances, along with corresponding prompts, to create desired videos. The locations and motion trajectories of instances are extracted from real videos and are encoded as additional control conditions for joint training with the video generation models. However, existing training-based methods can only control + +the movement trajectory [11, 41, 44, 52, 58, 65], since substantial data and training resources are required to simultaneously learn controlling layouts and movement trajectories for video generation. + +In this paper, we develop a novel training-free approach to address this problem. Existing training-free methods are restricted in obtaining accurate location and movement effects by manipulating spatial attention maps and temporal attention maps to generate instances at specified locations [20, 29, 32]. These methods convert location signals in the pixel space provided by users to a latent space with reduced dimension for manipulating attention maps, and inevitably compromise the precision of location control and obtain inaccurate instance placement. Excessive constraints on the attention maps also disrupt the motion prior of video diffusion. Moreover, frame interaction is neglected during the manipulation, which also causes unreasonable motion effects and unnatural generated videos. Additionally, these methods also suffer from generating low-quality videos with choppy motion and insufficient details. + +We find that both the inaccurate instance positioning and the unnatural motion stem from the limitations of attention map manipulation, which is a compromise solution to mitigate the lack of a clear understanding of instance concepts in vanilla latent diffusion models. It introduces constraints to enforce the instances to appear at designated positions. Contrary to attention map manipulation for video generation, instance-level knowledge is usually considered for image generation. Existing grounded T2I models like GLIGEN [25] and InstanceDiffusion [39] encapsulate extensive instance-level knowledge, but they are confined to generating images rather than consistent videos. Thus, we propose to generate coarse motion videos aligning with the user input and exhibit natural motion using instance-level priors in grounded T2I models, and further refine coarse motion videos by harnessing the motion and detail priors of T2V models for motion and detail injection in a zero-shot manner to boost the motion consistency and richness of details. + +To this end, we propose a zero-shot and plug-and-play method named IM-Zero that allows to simultaneously control instance layouts and movement trajectories for video generation. IM-Zero consists of two stages, i.e., a motion generation stage that acquires a coarse motion video, and a video refinement stage to enhance the quality of the coarse motion video. In the motion generation stage, spatial location signals (e.g., bounding boxes and masks) are first augmented according to the input trajectory to obtain a sequence of layout control signals. A batch of keyframe candidates with texture consistency is then obtained by inflating and transforming the pre-trained grounded T2I model, and is used to extract motion guidance and textural guidance for generating the desired coarse motion video. In the video refinement stage, we encode the coarse motion video + +into video latents, and inject the motion and detail priors extracted from the pre-trained T2V models into the video latents to improve motion consistency and enhance the richness of details, respectively. + +Our contributions are summarized as below. + +- We propose a novel framework named IM-Zero for zero-shot motion controllable video generation. IM-Zero supports various types of layout inputs and enables fine-grained motion control of the instances and their subparts. +- We present motion generation that generates keyframe candidates with the layout control capability of grounded T2I models to extract motion and textural guidance for coarse motion video generation. +- We develop zero-shot video refinement that injects the motion and detail priors of the pre-trained text-to-video model and text-to-image model into the latents of the coarse motion video to further enhance video motion consistency and richness of details. +- We further achieve motion transfer to customize complex motion patterns using reference videos and support high-quality text-to-video generation with only text input beyond instance layout and movement trajectory controls. + +Extensive experiments demonstrate that IM-Zero outperforms existing methods in terms of video quality, interframe consistency, and the alignment of location and trajectory. Compared with existing methods, IM-Zero enjoys extra advantages of versatility in video generation, including motion control of subparts within instances and finer control with instance shapes specified by masked inputs. Furthermore, it can be adapted to motion transfer tasks to support the customized fine-grained motion patterns through reference videos and can also be adapted to high-quality text-to-video generation. + +# 2. Related Work + +Grounded Text-to-image Generation. Grounded text-to-image (T2I) models aim to generate images that align with input layout information. Some works specialize in modeling layout information of the bounding box type. Notable examples include GLIGEN [25], InstanceDiffusion [39], BoxDiff [46], MultiDiffusion [1], ReCo [49], Layout-Guidance [8], MIGC [63, 64], and TraDiffusion [43]. We believe that some training-based methods such as InstanceDiffusion [39] have enabled models to grasp the concept of instances. Their capability aligns to some extent with our objective of achieving location control and trajectory control in video generation. However, these methods are limited to generating single images and cannot produce consistent videos. Therefore, we aim to leverage Grounded T2I models to benefit controllable video generation tasks. + +Location and Trajectory Control in Video Generation. Diffusion models have demonstrated exceptional capabilities in image generation [45, 61], video generation [56, 57], + +3D generation [10, 27], and other visual tasks [18, 54, 60]. With the improvement in video generation quality [3, 6, 7, 13-15, 22, 37, 38, 40, 47, 62], there is a growing interest in achieving generation controllability, with a particular focus on controlling the spatial location and movement trajectories of instances. Currently, training-based methods can only achieve limited control over trajectory guidance, at the expense of requiring substantial amounts of data and computational resources, with typical examples being DragNUWA [52], DragAnything [44], DragVideo [11], Tora [58], TrackGo [65] and MotionCtrl [41]. In contrast, some training-free methods can concurrently control both spatial location and movement trajectories. For instance, Peekaboo [20] and TrailBlazer [29] enforce constraints on attention maps to position instances at specified locations. However, this approach can lead to issues such as inaccuracies in location and poor motion effects. FreeTraj [32] follows a similar attention map manipulation approach but introduces noise flow control. In addition to the aforementioned challenges, it also introduces difficulties in handling bounding boxes that undergo size changes. Therefore, we propose a zero-shot method that surpasses attention map manipulation to address the aforementioned challenges and enhance video quality and motion consistency. + +# 3. Method + +# 3.1. Preliminaries + +Grounded Text-to-image Diffusion Models. Grounded text-to-image (T2I) diffusion models [25, 39, 63, 64] are capable of generating instances at user-specified locations by incorporating the layout input as conditions for joint training. InstanceDiffusion [39] supports four types of layout input: bounding box, mask, point, and scribble. It inserts learnable UniFusion blocks between the self-attention and cross-attention layers of the T2I model. UniFusion sample the instance layout as a 2D points $\mathbf{p}$ sequence and convert it to layout token by $\mathbf{g} = \mathrm{MLP}\left([ \tau_{\theta}(\mathbf{c}), \gamma(\mathbf{p}) \right]$ , where $\gamma(\cdot)$ is a Fourier mapping and $\mathbf{c}$ is the text prompt. Then the layout token is fused with the backbone visual tokens via a learnable masked self-attention layer. + +Text-to-video Diffusion Models. Significant progress has also been made in Text-to-Video (T2V) diffusion models. While some T2V models are trained from scratch, the majority of them are trained based on existing T2I models. A typical example isAnimateDiff [13], which utilizes pre-trained T2I models as the spatial layer and trains additional motion modeling modules to reconstruct N frames clean video latents $z^{1:N}$ from i.i.d Gaussian noise conditioned on the text prompt $c$ by $\mathcal{L}(\theta) = \mathbb{E}_{z_0^{1:N},\epsilon ,c_t,t}\left[\| \epsilon -\hat{\epsilon}_\theta \left(z_t^{1:N},c_t,t\right)\right]$ . + +![](images/c19118d955a0bf730caca8b242170a566cbfeed7a179503c644b0de4dea17533.jpg) +Figure 2. The Motion Generation Stage consists of two parts: Keyframe Candidates Generation and Coarse Motion Video Generation. In the first part, we augment the user input to obtain Layout, then generate keyframe candidates Frame with an inflated grounded T2I model. In the second part, we extract motion guidance and textural guidance from Frame to generate a coarse motion video. + +# 3.2. Overall Framework + +IM-Zero framework consists of two critical stages: Motion Generation Stage and Video Refinement Stage. In Section 3.1, we first briefly introduce the fundamentals of grounded T2I and T2V diffusion models. Then, we introduce the Motion Generation Stage in Section 3.3 as shown in Figure 2. This stage generates keyframe candidates and extracts guidance from them to generate a coarse motion video. Section 3.4 introduces the Video Refinement Stage as shown in Figure 3. This stage utilizes motion prior of T2V models to perform Motion Injection and utilizes the detail-generation capacity of T2I models to perform Detail Injection, further enhancing the visual quality of coarse motion video. Lastly, we also offer an option for motion transfer via reference video input, and we will explain how to easily modify IM-Zero in this case. As illustrated in Figure 1, IM-Zero achieves motion control for overall instances or subparts and allows users to further specify instance shapes as the grounded T2I model can generate keyframe candidates via masks. IM-Zero can perform motion transfer. Besides, IM-Zero allows high-quality T2V generation by refining any video via the Video Refinement Stage. + +# 3.3. Motion Generation Stage + +The core task involves presenting suitable motion patterns in the video, including two key aspects: the user-desired instance spatial location and the user-desired instance movement. Previous training-free methods achieved this task by manipulating attention maps but encountered issues such as inaccurate location, poor motion effects, and low video quality, as discussed in Section 1. IM-Zero aims to achieve the decoupling of instance spatial location and instance movement to enhance the performance of both aspects. For the former, we utilize a grounded T2I model [39] to generate instances at user-specified locations. For the latter, we depart from the previous approach of imposing excessive constraints on the attention map and instead efficiently + +leverage the motion priors of the T2V model. + +Keyframe Candidates Generation. Considering the user input $S$ layout signals and $L$ trajectory segments, We denote the layout signals as $\mathbf{Layout} = \{s_i|i\sim [1,2,\dots S]\}$ (i.e., bounding boxes) and the movement trajectory as Traj = $\{l_i|i\sim [1,2,\dots ,L]\}$ . We ensure that each $l_{i}$ segment has a layout signal at both endpoints, for the endpoint without a layout signal, we select the closest $s_i$ from Layout and duplicate it here, obtaining the augmented layout signals $\mathbf{Layout} = \{s_i|i\sim [1,2,\dots,S']\}$ . These signals are then input into the grounded T2I model to generate keyframe candidates $\mathbf{Frame} = \{f_i|i\sim [1,2,\dots,S']\}$ . Since generating each image separately can lead to significant inconsistencies, following [22], we inflate the grounded T2I model and replace each self-attention in UNet with cross-frame attention (CFA). The CFA for the $i$ -th keyframe candidate $f_{i}$ is defined as CFA $(Q^{i},K^{1:N_{\mathrm{S^{\prime}}}},V^{1:N_{\mathrm{S^{\prime}}}}) =$ Softmax $(Q^{i}(K^{1})^{T} / \sqrt{c})V^{1}$ where $Q,K$ and $V$ denote the queries, keys and values. The first candidate provides key and value for all following ones, which can be interpreted from an image editing perspective as querying relevant content and textures from the first candidate to ensure consistency in textures [5, 22]. + +Coarse Motion Video Generation. After obtaining Frame, we utilize these candidates to generate a coarse motion video reflecting the user's input. To ensure smooth and natural instance motion, we have avoided the previous approach of directly enforcing excessive constraints on the attention map, which could disrupt the T2V motion priors. Instead, we extract two types of guidance from Frame to guide the T2V model. (1) Motion Guidance. To ensure the fidelity of the video to the user's input, encompassing spatial location and motion patterns, we extract foreground instances from Frame and convert them into Canny edge images. Subsequently, we employ ControlNet [12, 16, 55] to guide the generation of the coarse motion video. The rationale behind our approach is elucidated as follows: firstly, + +![](images/f7b89de37622acb7658d1156de3b5f92d9fdc196fdcae0f3ffba79510d6691b8.jpg) + +![](images/537bb9c31e78acb36f5242d41bb91e058f816e55c24b1e19f1ffbd6f2998e43f.jpg) +Figure 3. In Video Refinement Stage, we diffuse coarser motion video latents and perform motion injection and detail injection in the reverse process. In motion injection, we first obtain $z_{t-1}$ by $z_t$ , then denoise $z_{t-1}$ for further $M$ steps using a T2V model to $\tilde{z}_{t-1-M}'$ and combine them. We visualize the motion injection effect using CoTracker [21]. In detail injection, we first transform $\tilde{z}_{t-1}$ from T2V noise distribution to T2I noise distribution, then denoise it for further $N$ steps with an inflated T2I model. The result is reversed to T2V noise distribution and combined with $\tilde{z}_{t-1}$ . We visualize the detail injection effect using FFT analysis. + +![](images/b32e7d313bff07f481ff53c8c3d8f9463787037d7f103cde51ea2e41194272bc.jpg) + +the separation of foreground from the background is essential because instances described in the user's input manifest as foreground instances in Frame [25, 39]. Although we apply operations to enhance consistency by CFA, background distortions and abrupt changes can still occur [22]. Secondly, transforming images into Canny edge representations not only aligns with the requirements of ControlNet but also further mitigates potential inconsistencies in instance textures within Frame. (2) Textural Guidance. Use Motion Guidance only may cause textural inconsistency, hence we further perform Textural Guidance. We diffuse the first keyframe candidate $f_{1}$ to form the initial noise of the T2V model. To help the guidance transmit from $f_{1}$ to subsequent frames, we devise the following strategies: firstly, we employ an IP-Adapter [51] to utilize $f_{1}$ as an image prompt. Moreover, we also apply CFA to the spatial layers of the T2V model. + +# 3.4. Video Refinement Stage + +We have observed that while coarse motion videos may align with user-desired inputs, unsmooth motion issues and suboptimal visual quality issues such as watermark and flickering exist. To attack these issues, we introduce a zero-shot method by more efficiently leveraging the motion priors inherent in existing T2V models and the detail-generation capacity of T2I models. As shown in Figure 3, we integrate a T2V model into the reverse process of video generation. We leverage T2V motion priors to inject motion into video latents, ensuring smoother and more natural motion. Subsequently, we also insert a T2I model, leveraging its detail generation capacity to enrich video detail richness. + +Given a low-quality coarse motion video as described in Section 3.3, we start by encoding it using the encoder $\mathcal{E}$ of a vanilla T2V model (i.e.,AnimateDiff [13]) and adding Gaussian noise by timestep $T$ to obtain noisy video latents $z_{T}$ . For each denoising time step $t$ , we first denoise $z_{t}$ using + +the denoiser of the T2V model to obtain $z_{t - 1}$ by: + +$$ +p _ {\theta} \left(z _ {t - 1} \mid z _ {t}\right) = \mathcal {N} \left(z _ {t - 1}; z _ {t} - \hat {\epsilon} _ {\theta} \left(z _ {t}; c, t\right), \left(1 - \alpha_ {t} ^ {2}\right) \mathbf {I}\right) \tag {1} +$$ + +where $c$ is the text prompt embedding. Then we sequentially perform motion injection and detail injection on $z_{t - 1}$ + +Motion Injection. We employ a vanilla T2V model (i.e.,AnimateDiff [13]) for motion injection as shown in Figure 3. Specifically, we denoise $z_{t - 1}$ using its denoiser for further $M$ steps to acquire $\tilde{z}_{t - 1 - M}^{\prime}$ , where each denoising step follows Equation (1). We observed a significant improvement in the temporal consistency of $\tilde{z}_{t - 1 - M}^{\prime}$ , resulting in very smooth motion. However, this enhancement comes at the cost of losing many detailed features. Therefore, we combine $\tilde{z}_{t - 1 - M}^{\prime}$ with $z_{t - 1}$ to enhance motion smoothness while preserving details by: + +$$ +\tilde {z} _ {t - 1} = \lambda_ {1} * z _ {t - 1} + (1 - \lambda_ {1}) * \tilde {z} _ {t - 1 - M} ^ {\prime} \tag {2} +$$ + +where $\lambda_{1}$ is a coefficient controlling the fusion scale. + +Detail Injection. We then employ an inflated T2I model for detail injection as shown in Figure 3. Due to the different noise distributions in the T2I and T2V models, a noise transformation is required. To convert video latents into image latents, we directly estimate clean latents [48] $\hat{z}_0$ by: + +$$ +\hat {z} _ {0} = \frac {\tilde {z} _ {t} - \sqrt {1 - \alpha_ {t}} \epsilon_ {\theta} (\tilde {z} _ {t} , c , t)}{\sqrt {\alpha_ {t}}} \tag {3} +$$ + +and we then conduct DDIM inversion [30] to derive the corresponding noisy image latents $\tilde{z}_{t - 1}^{I}$ , the DDIM inversion process is defined by: + +$$ +\tilde {z} _ {t} ^ {I} = \sqrt {\alpha_ {t}} \hat {z} _ {0} + \sqrt {1 - \alpha_ {t}} \cdot \epsilon_ {\theta} (\tilde {z} _ {t - 1} ^ {I}, c, t) \tag {4} +$$ + +Subsequently, we denoise the noisy image latents with the inflated T2I model for further $D$ steps to obtain $\tilde{z}_{t-1-D}^{I}$ . The denoising process resembles Equation (1). As the T2I model excels in detail generation compared to the T2V + +![](images/af8934e9b64973ed90739bcec9e2fb12bab3a5e4c2413e8971a8c7ef4cdee6dc.jpg) +Figure 4. Qualitative Results of IM-Zero compared with other methods. Videos generated by IM-Zero outperform others in control accuracy, motion effects, and video quality. + +model, we revert $\tilde{z}_{t - 1 - D}^{I}$ back to the noise distribution of the T2V model, combining the transformed latents with $\tilde{z}_{t - 1}$ obtained from Equation (2) to integrate details. Specifically, in the reversed noise transformation process, we predict clean noise from $\tilde{z}_{t - 1}$ using Equation (3) and utilize the DDPM forward process to obtain the corresponding noisy video latents $\tilde{z}_{t - 1 - D}^{V}$ . Then we fuse them by: + +$$ +z _ {t - 1} = \lambda_ {2} * \tilde {z} _ {t - 1} + (1 - \lambda_ {2}) * \tilde {z} _ {t - 1 - D} ^ {V} \tag {5} +$$ + +where $\lambda_{2}$ controls the fusion scale and $z_{t - 1}$ is used for the next denoising timestep. + +# 3.5. Adapt IM-Zero for Motion Transfer + +For a more customized motion pattern provided by reference videos, IM-Zero can also be easily adapted for motion transfer tasks. Given a reference video and the target text prompt, we extract control signals (e.g., depth maps) from each frame of the reference video. Subsequently, we use these dense control signals and the target text prompt to guide the T2V model in generating a coarse motion video via ControlNet. Then, the coarse motion video undergoes video refinement stage as detailed in Section 3.4 to enhance motion consistency and video frame quality. + +# 4. Experiment + +# 4.1. Qualitative Results + +Regarding the comparative methods, we primarily selected Peekaboo [20], TrailBlazer [29], and MotionBooth [42], which generate videos by specifying boxes and trajectories. Furthermore, to further illustrate the effectiveness of our approach, we also included T2V-Zero [22], which offers rough guidance by specifying translation vectors. To ensure a fair comparison, following the experimental settings of [29], all methods utilize the same layout input, trajectory input, and text prompt input. Additional results of controllable video generation are available in the supplementary material. We also compared with FreeTraj [32], but as it additionally requires inputting boxes with fixed sizes across frames to make noise flow, we place the comparison in the supplementary material. + +As shown in Figure 4, our method significantly outperforms other approaches in terms of alignment with the input, motion effects, and video quality. In terms of alignment with the input, T2V-Zero exhibits almost no resemblance to the input, while the two methods based on attention map manipulation, Peekaboo, and TrailBlazer, sometimes roughly match the input location and trajectory but + +Table 1. Quantitative results of IM-Zero compared with other methods on DAVIS-17 [31] dataset and GOT10k [17] dataset. IM-Zero achieves SOTA in all metrics, including metrics measuring video quality, frame consistency, and control accuracy. + +
MethodDAVIS-17GOT10k
FVD (↓)KID (↓)CLIPSim (↑)mIoU (↑)CD (↓)FVD (↓)KID (↓)CLIPSim (↑)mIoU (↑)CD (↓)
T2V-Zero [22]5004.0030.3795.120.190.344335.6924.7295.410.120.32
Peekaboo [20]4264.7026.4897.330.190.464960.4919.0598.680.110.34
TrailBlazer [29]4593.3726.0498.100.210.294695.1816.1897.800.130.40
MotionBooth [42]4192.1926.8596.640.220.364311.7116.0597.470.140.42
IM-Zero (Ours)4070.0125.7598.910.270.264187.2715.9299.040.210.27
+ +Table 2. Ablation on different components in Motion Generation Stage. + +
MethodFVD (↓)KID (↓)CLIPSim (↑)mIoU (↑)CD (↓)
w/o 1st CFA4320.0726.8498.840.260.29
w/o 2nd CFA4528.0928.0898.780.210.31
w/o init noise4786.1726.6998.820.260.31
w/o IP-Adapter4549.9427.7598.870.260.28
IM-Zero (Ours)4070.0125.7598.910.270.26
+ +still lack precision. Regarding motion effects, Peekaboo and TrailBlazer generate videos with unnatural motion, resulting in strange discontinuities between frames. Concerning video quality, Peekaboo and TrailBlazer exhibit low image quality, lack detail, and introduce artifacts. We also provide a video version of Figure 4 in the supplementary material. + +# 4.2. Quantitative Results + +Experiment Settings. For datasets, we utilize the training set and validation set of the DAVIS-17 dataset [31] and the test set of the GOT10k dataset [17]. We use BLIP-2 [24] to generate text prompts and GroundingDINO [28] to extract bounding boxes from the first and last frame as inputs of all the methods. After eliminating videos where no bounding boxes are extracted, we randomly select 50 videos and 70 videos from the rest videos of DAVIS-17 [31] and GOT10k [17]. For our method, in the first stage we use InstanceDiffusion [39] as grounded T2I model and SparseCtrl [12] as ControlNet, in the second stage we use Stable Diffusion V1.5 [34] as T2I model, on both stage we useAnimateDiff [13] as T2V model. For the evaluation metrics, we employ mean Intersection over Union (mIoU) and Centroid Distance (CD) to assess control alignment, Fréchet Video Distance (FVD) [35] and Kernel Inception Distance (KID) [2] to evaluate video quality, and CLIP similarities (CLIPSim) [33] to assess frame-to-frame consistency. Detailed settings are available in the supplementary material. + +Metrics Evaluation Results. We conducted evaluations on the aforementioned metrics on the DAVIS-17 dataset [31] and the GOT10k dataset [17]. As shown in Table 1, our method achieved state-of-the-art (SOTA) performance compared to several other methods. Firstly, in terms of the alignment between the generated videos and user inputs, + +Table 3. Motion transfer results compared with other methods. + +
MethodMotion Fidelity (↑)Imaging Quality (↑)
MotionClone [26]74.50.72
MOFT [23]51.20.71
MotionDirector [59]75.50.68
IM-Zero (Ours)63.70.73
+ +Table 4. T2V results compared with UNet baseline. + +
MethodImaging Quality (↑)CLIPSim (↑)
AnimateDiff [13]0.6799.05
IM-Zero (Ours)0.7499.73
+ +our method demonstrates a significant lead. Compared to the previous best methods, our approach achieved an improvement of $28.6\%$ and $61.5\%$ in mIoU, and reductions of $10.4\%$ and $15.6\%$ in CD on the two datasets, indicating that our method offers more accurate location control and motion trajectory control compared to previous methods based on attention map manipulation. Secondly, regarding frame-to-frame consistency in the videos, thanks to our enhanced consistency techniques and motion injection methods, our approach also reaches the SOTA performance. Moreover, in terms of video quality, our method significantly outperforms others, with FVD reduced to 4070.01 and 4187.27, and KID reduced to 25.75 and 15.92, respectively. + +# 4.3. Versatile Capacity Analysis + +Our method possesses versatile capacity as shown in Figure 5. First, our method can control the subparts of instances, achieving effects such as head turning and hand + +![](images/54d5cbbc0ed4713fd4710482d519fe249ff17f62fdd1a5cd9ae15d12c13078d3.jpg) +# +Figure 5. Besides box-level overall instance motion control, IM-Zero can: (a) Perform instance subpart control. (b) Specify instance shape by mask. (c) Perform motion transfer from the reference video(left) to the target video(right). (d) Perform high-quality T2V generation. + +movement. Second, besides bounding boxes, we also support specifying instance shapes more precisely using masks. Third, for cases where users provide a reference video, our method can perform motion transfer. Finally, users can directly provide text inputs for high-quality T2V generation. For motion transfer and T2V generation, we compare IM-Zero with other methods as follows. + +Motion Transfer. We compare with MOFT [23], MotionClone [26], and MotionDirector [59] using the open-sourced data from MotionClone [26]. For quantitative results, we use Motion Fidelity [50] and Imaging Quality [19] as metrics following MOFT [23]. Table 3 shows that our method is competitive in Motion Fidelity [50] and is the best in Imaging Quality [19]. Qualitative results and other details are provided in the supplementary material. + +T2V Generation. As we use UNet-based models, we compare T2V with UNet baselineAnimateDiff [13]. For quantitative results, we use Imaging Quality [19] and CLIP-Sim [33] as metrics. As there are no ground-truth videos, we generate videos using 50 prompts randomly generated by GPT. Table 4 shows that our method is better in Imaging Quality [19] and CLIPSim [33]. Refer to the supplementary material for qualitative results and other details. + +Other Capacities. The proposed method also supports multi-instance motion control, and the decomposition of our pipeline allows autoregressive usage of the second stage to enhance temporal consistency, as elaborated in the supplementary material. + +# 4.4. Ablation Studies + +# Ablation on different components in Motion Generation + +Stage. We ablate on different components in Section 3.3. As depicted in Table 2, these methods improve the quality + +of the final videos and also enhance the alignment with user inputs. Refer to the supplementary material for details. + +Ablation on different components in Video Refinement Stage. We also conduct ablation studies on Video Refinement Stage. Refer to the supplementary material for details. + +# 5. Conclusion and Limitation + +We introduce IM-Zero, a novel training-free framework for instance-level motion controllable video generation using diffusion models. We leverage grounded text-to-image models to generate keyframe candidates and extract motion guidance and textural guidance from them to generate videos that align with user inputs. To further enhance video motion and quality, we also introduce a zero-shot method for motion injection and detail injection. Additionally, our method can control the subpart motion of instances, use masks to specify instance shapes, perform motion transfer from reference videos, and directly support high-quality text-to-video generation. + +Limitation. The zero-shot features impose certain limitations on our method. Firstly, as zero-shot methods depend on pre-trained model weights, the videos generated by our method are inherently constrained by the pre-trained models employed. For example, the alignment with layout input is influenced by the grounded T2I model and video ControlNet, the video quality is affected by the T2V model, which can be improved by using more advanced pre-trained models. Secondly, the temporal consistency of videos generated by zero-shot methods is inherently weaker compared to training-based methods, which can be improved by incorporating some tuning components to our method. + +# Acknowledgements + +This work was supported in part by the National Natural Science Foundation of China under Grant 62125109, Grant 62431017, Grant U24A20251, Grant 62320106003, Grant 62371288, Grant 62401357, Grant 62401366, Grant 62301299, Grant 62120106007, and in part by the Program of Shanghai Science and Technology Innovation Project under Grant 24BC3200800. + +# References + +[1] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. MultiDiffusion: Fusing diffusion paths for controlled image generation. In Proceedings of the 40th International Conference on Machine Learning, pages 1737-1752, 2023. 3 +[2] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying MMD GANs. In *The Sixth International Conference on Learning Representations*, 2018. 7 +[3] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 2, 3 +[4] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. OpenAI Blog, 2024. 2 +[5] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. MasaCtrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22560–22570, 2023. 4 +[6] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, Chao Weng, and Ying Shan. VideoCrafter1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512, 2023. 2, 3 +[7] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. VideoCrafter2: Overcoming data limitations for high-quality video diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7310-7320, 2024. 2, 3 +[8] Minghao Chen, Iro Laina, and Andrea Vedaldi. Training-free layout control with cross-attention guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5343–5353, 2024. 3 +[9] Weifeng Chen, Yatai Ji, Jie Wu, Hefeng Wu, Pan Xie, Jiashi Li, Xin Xia, Xuefeng Xiao, and Liang Lin. Control-A-Video: Controllable text-to-video generation with diffusion models. arXiv preprint arXiv:2305.13840, 2023. 2 +[10] Yabo Chen, Jiemin Fang, Yuyang Huang, Taoran Yi, Xiaopeng Zhang, Lingxi Xie, Xinggang Wang, Wenrui Dai, + +Hongkai Xiong, and Qi Tian. Cascade-Zero123: One image to highly consistent 3D with self-prompted nearby views. In Proceedings of the 18th European Conference on Computer Vision, pages 311-330, 2024. 3 +[11] Yufan Deng, Ruida Wang, Yuhao Zhang, Yu-Wing Tai, and Chi-Keung Tang. DragVideo: Interactive drag-style video editing. In Proceedings of the 18th European Conference on Computer Vision, pages 183–199, 2024. 2, 3 +[12] Yuwei Guo, Ceyuan Yang, Anyi Rao, Maneesh Agrawala, Dahua Lin, and Bo Dai. SparseCtrl: Adding sparse controls to text-to-video diffusion models. In Proceedings of the 18th European Conference on Computer Vision, pages 330-348, 2024. 4, 7 +[13] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. AnimateDiff:Animate your personalized text-to-image diffusion models without specific tuning. In The Twelfth International Conference on Learning Representations, 2024. 2, 3, 5, 7, 8 +[14] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P. Kingma, Ben Poole, Mohammad Norouzi, David J. Fleet, and Tim Salimans. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. +[15] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. In Advances in Neural Information Processing Systems 35, pages 8633-8646, 2022. 2, 3 +[16] Zhihao Hu and Dong Xu. VideoControlNet: A motion-guided video-to-video translation framework by using diffusion model with ControlNet. arXiv preprint arXiv:2307.14073, 2023. 2, 4 +[17] Lianghua Huang, Xin Zhao, and Kaiqi Huang. GOT-10k: A large high-diversity benchmark for generic object tracking in the wild. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(5):1562-1577, 2019. 7 +[18] Yuyang Huang, Yabo Chen, Yuchen Liu, Xiaopeng Zhang, Wenrui Dai, Hongkai Xiong, and Qi Tian. DomainFusion: Generalizing to unseen domains with latent diffusion models. In Proceedings of the 18th European Conference on Computer Vision, pages 480-498, 2024. 3 +[19] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, Yaohui Wang, Xinyuan Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. VBench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024. 8 +[20] Yash Jain, Anshul Nasery, Vibhav Vineet, and Harkirat Behl. PEEKABOO: Interactive video generation via masked-diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8079-8088, 2024. 2, 3, 6, 7 +[21] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. CoTracker: It is better to track together. In Proceedings of the 18th European Conference on Computer Vision, pages 18-35, 2024. 5 + +[22] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2Video-Zero: Text-to-image diffusion models are zero-shot video generators. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15954-15964, 2023. 2, 3, 4, 5, 6, 7 +[23] Karlo Koledić, Igor Cvišić, Ivan Marković, and Ivan Petrović. MOFT: Monocular odometry based on deep depth and careful feature selection and tracking. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 6175-6181. IEEE, 2023. 7, 8 +[24] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In Proceedings of the 40th International Conference on Machine Learning, pages 19730–19742, 2023. 7 +[25] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. GLIGEN: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22511-22521, 2023. 2, 3, 5 +[26] Pengyang Ling, Jiazi Bu, Pan Zhang, Xiaoyi Dong, Yuhang Zang, Tong Wu, Huaian Chen, Jiaqi Wang, and Yi Jin. MotionClone: Training-free motion cloning for controllable video generation. In The Thirteenth International Conference on Learning Representations, 2025. 7, 8 +[27] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9298–9309, 2023. 3 +[28] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, and Lei Zhang. Grounding DINO: Marrying DINO with grounded pre-training for open-set object detection. In Proceedings of the 18th European Conference on Computer Vision, pages 38-55, 2024. 7 +[29] Wan-Duo Kurt Ma, John P Lewis, and W Bastiaan Kleijn. TrailBlazer: Trajectory control for diffusion-based video generation. In SIGGRAPH Asia 2024, number 97, 2024. 2, 3, 6, 7 +[30] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 5 +[31] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 DAVIS challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 7 +[32] Haonan Qiu, Zhaoxi Chen, Zhouxia Wang, Yingqing He, Menghan Xia, and Ziwei Liu. FreeTraj: Tuning-free trajectory control in video diffusion models. arXiv preprint arXiv:2406.16863, 2024. 2, 3, 6 +[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, + +Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763, 2021. 7, 8 +[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 7 +[35] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 7 +[36] Cong Wang, Jiaxi Gu, Panwen Hu, Haoyu Zhao, Yuanfan Guo, Jianhua Han, Hang Xu, and Xiaodan Liang. EasyControl: Transfer ControlNet to video diffusion for controllable generation and interpolation. arXiv preprint arXiv:2408.13005, 2024. 2 +[37] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. ModelScope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 2, 3 +[38] Weimin Wang, Jiawei Liu, Zhijie Lin, Jiangqiao Yan, Shuo Chen, Chetwin Low, Tuyen Hoang, Jie Wu, Jun Hao Liew, Hanshu Yan, Daquan Zhou, and Jiashi Feng. MagicVideoV2: Multi-stage high-aesthetic video generation. arXiv preprint arXiv:2401.04468, 2024. 2, 3 +[39] Xudong Wang, Trevor Darrell, Sai Saketh Rambhatla, Rohit Girdhar, and Ishan Misra. InstanceDiffusion: Instance-level control for image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6232-6242, 2024. 2, 3, 4, 5, 7 +[40] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, Yuwei Guo, Tianxing Wu, Chenyang Si, Yuming Jiang, Cunjian Chen, Chen Change Loy, Bo Dai, Dahua Lin, Yu Qiao, and Ziwei Liu. LaVie: High-quality video generation with cascaded latent diffusion models. International Journal of Computer Vision, 2024. 2, 3 +[41] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. MotionCtrl: A unified and flexible motion controller for video generation. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2, 3 +[42] Jianzong Wu, Xiangtai Li, Yanhong Zeng, Jiangning Zhang, Qianyu Zhou, Yining Li, Yunhai Tong, and Kai Chen. MotionBooth: Motion-aware customized text-to-video generation. In Advances in Neural Information Processing Systems 37, pages 34322-34348, 2024. 6, 7 +[43] Mingrui Wu, Oucheng Huang, Jiayi Ji, Jiale Li, Xinyue Cai, Huafeng Kuang, Jianzhuang Liu, Xiaoshuai Sun, and Rongrong Ji. TraDiffusion: Trajectory-based training-free image generation. arXiv preprint arXiv:2408.09739, 2024. 3 +[44] Weijia Wu, Zhuang Li, Yuchao Gu, Rui Zhao, Yefei He, David Junhao Zhang, Mike Zheng Shou, Yan Li, Tingting Gao, and Di Zhang. DragAnything: Motion control for anything using entity representation. In Proceedings of the 18th + +European Conference on Computer Vision, pages 331-348, 2024. 2, 3 +[45] Enze Xie, Junsong Chen, Junyu Chen, Han Cai, Haotian Tang, Yujun Lin, Zhekai Zhang, Muyang Li, Ligeng Zhu, Yao Lu, and Song Han. SANA: Efficient high-resolution image synthesis with linear diffusion transformers. In The Thirteenth International Conference on Learning Representations, 2025. 3 +[46] Jinheng Xie, Yuexiang Li, Yawen Huang, Haozhe Liu, Wentian Zhang, Yefeng Zheng, and Mike Zheng Shou. BoxDiff: Text-to-image synthesis with training-free box-constrained diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7452-7461, 2023. 3 +[47] Yichao Yan, Zanwei Zhou, Zi Wang, Jingnan Gao, and Xiaokang Yang. DialogueNeRF: Towards realistic avatar face-to-face conversation video generation. Visual Intelligence, 2 (1):24, 2024. 3 +[48] Serin Yang, Hyunmin Hwang, and Jong Chul Ye. Zero-shot contrastive loss for text-guided diffusion image style transfer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22873-22882, 2023. 5 +[49] Zhengyuan Yang, Jianfeng Wang, Zhe Gan, Linjie Li, Kevin Lin, Chenfei Wu, Nan Duan, Zicheng Liu, Ce Liu, Michael Zeng, and Lijuan Wang. ReCo: Region-controlled text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14246-14255, 2023. 3 +[50] Danah Yatim, Rafail Fridman, Omer Bar-Tal, Yoni Kasten, and Tali Dekel. Space-time diffusion features for zero-shot text-driven motion transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8466-8476, 2024. 8 +[51] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. IP-Adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 5 +[52] Shengming Yin, Chenfei Wu, Jian Liang, Jie Shi, Houqiang Li, Gong Ming, and Nan Duan. DragNUWA: Fine-grained control in video generation by integrating text, image, and trajectory. arXiv preprint arXiv:2308.08089, 2023. 2, 3 +[53] Yan Zeng, Guoqiang Wei, Jiani Zheng, Jiaxin Zou, Yang Wei, Yuchen Zhang, and Hang Li. Make pixels dance: High-dynamic video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8850-8860, 2024. 2 +[54] Junyi Zhang, Charles Herrmann, Junhwa Hur, Luisa Polania Cabrera, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. A tale of two features: Stable diffusion complements DINO for zero-shot semantic correspondence. In Advances in Neural Information Processing Systems 36, pages 45533-45547, 2023. 3 +[55] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 4 +[56] Yabo Zhang, Yuxiang Wei, Dongsheng Jiang, Xiaopeng Zhang, Wangmeng Zuo, and Qi Tian. ControlVideo: + +Training-free controllable text-to-video generation. In The Twelfth International Conference on Learning Representations, 2024. 2, 3 +[57] Yabo Zhang, Yuxiang Wei, Xianhui Lin, Zheng Hui, Peiran Ren, Xuansong Xie, Xiangyang Ji, and Wangmeng Zuo. VideoElevator: elevating video generation quality with versatile text-to-image diffusion models. arXiv preprint arXiv:2403.05438, 2024. 3 +[58] Zhenghao Zhang, Junchao Liao, Menghao Li, Long Qin, and Weizhi Wang. Tora: Trajectory-oriented diffusion transformer for video generation. arXiv preprint arXiv:2407.21705, 2024. 2, 3 +[59] Rui Zhao, Yuchao Gu, Jay Zhangjie Wu, David Junhao Zhang, Jia-Wei Liu, Weijia Wu, Jussi Keppo, and Mike Zheng Shou. MotionDirector: Motion customization of text-to-video diffusion models. In Proceedings of the 18th European Conference on Computer Vision, pages 273–290, 2024. 7, 8 +[60] Wenliang Zhao, Yongming Rao, Zuyan Liu, Benlin Liu, Jie Zhou, and Jiwen Lu. Unleashing text-to-image diffusion models for visual perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5706-5716, 2023. 3 +[61] Wendi Zheng, Jiayan Teng, Zhuoyi Yang, Weihan Wang, Ji-dong Chen, Xiaotao Gu, Yuxiao Dong, Ming Ding, and Jie Tang. CogView3: Finer and faster text-to-image generation via relay diffusion. In Proceedings of the 18th European Conference on Computer Vision, pages 1-22, 2024. 3 +[62] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. MagicVideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022. 2, 3 +[63] Dewei Zhou, You Li, Fan Ma, Xiaoting Zhang, and Yi Yang. MIGC: Multi-instance generation controller for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6818-6828, 2024. 3 +[64] Dewei Zhou, You Li, Fan Ma, Zongxin Yang, and Yi Yang. MIGC++: Advanced multi-instance generation controller for image synthesis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 47(3):1714-1728, 2025. 3 +[65] Haitao Zhou, Chuang Wang, Rui Nie, Jinxiao Lin, Dongdong Yu, Qian Yu, and Changhu Wang. TrackGo: A flexible and efficient method for controllable video generation. arXiv preprint arXiv:2408.11475, 2024. 2, 3 \ No newline at end of file diff --git a/2025/IM-Zero_ Instance-level Motion Controllable Video Generation in a Zero-shot Manner/images.zip b/2025/IM-Zero_ Instance-level Motion Controllable Video Generation in a Zero-shot Manner/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1a72ee568f62d9f25f8d56b7613bc67195b2767a --- /dev/null +++ b/2025/IM-Zero_ Instance-level Motion Controllable Video Generation in a Zero-shot Manner/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5b4dadc8b5154a60a519249debe96f1957ebf2a79fcf79988b7ca764b7ada1c0 +size 846639 diff --git a/2025/IM-Zero_ Instance-level Motion Controllable Video Generation in a Zero-shot Manner/layout.json b/2025/IM-Zero_ Instance-level Motion Controllable Video Generation in a Zero-shot Manner/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..85d60251fe503df1166c6fad10c64ed222d6cdfb --- /dev/null +++ b/2025/IM-Zero_ Instance-level Motion Controllable Video Generation in a Zero-shot Manner/layout.json @@ -0,0 +1,10082 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 110, + 103, + 501, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 103, + 501, + 138 + ], + "spans": [ + { + "bbox": [ + 110, + 103, + 501, + 138 + ], + "type": "text", + "content": "IM-Zero: Instance-level Motion Controllable Video Generation in a Zero-shot Manner" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 71, + 160, + 541, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 160, + 541, + 232 + ], + "spans": [ + { + "bbox": [ + 71, + 160, + 541, + 232 + ], + "type": "text", + "content": "Yuyang Huang" + }, + { + "bbox": [ + 71, + 160, + 541, + 232 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 71, + 160, + 541, + 232 + ], + "type": "text", + "content": ", Yabo Chen" + }, + { + "bbox": [ + 71, + 160, + 541, + 232 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 71, + 160, + 541, + 232 + ], + "type": "text", + "content": ", Li Ding" + }, + { + "bbox": [ + 71, + 160, + 541, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 71, + 160, + 541, + 232 + ], + "type": "text", + "content": ", Xiaopeng Zhang" + }, + { + "bbox": [ + 71, + 160, + 541, + 232 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 71, + 160, + 541, + 232 + ], + "type": "text", + "content": ", Wenrui Dai" + }, + { + "bbox": [ + 71, + 160, + 541, + 232 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 71, + 160, + 541, + 232 + ], + "type": "text", + "content": ", Junni Zou" + }, + { + "bbox": [ + 71, + 160, + 541, + 232 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 71, + 160, + 541, + 232 + ], + "type": "text", + "content": ", Hongkai Xiong" + }, + { + "bbox": [ + 71, + 160, + 541, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 71, + 160, + 541, + 232 + ], + "type": "text", + "content": ", Qi Tian" + }, + { + "bbox": [ + 71, + 160, + 541, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 71, + 160, + 541, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 71, + 160, + 541, + 232 + ], + "type": "text", + "content": "Shanghai Jiao Tong University, Shanghai, China " + }, + { + "bbox": [ + 71, + 160, + 541, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 71, + 160, + 541, + 232 + ], + "type": "text", + "content": "Huawei Inc., Shenzhen, China \n{huangyuyang, chenyabo, wangzhongren, daiwenrui, zoujunni, xionghongkai}@sjtu.edu.cn \nzxphistory@gmail.com, tian.qil@huawei.com" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 85, + 239, + 135, + 290 + ], + "blocks": [ + { + "bbox": [ + 85, + 239, + 135, + 290 + ], + "lines": [ + { + "bbox": [ + 85, + 239, + 135, + 290 + ], + "spans": [ + { + "bbox": [ + 85, + 239, + 135, + 290 + ], + "type": "image", + "image_path": "ddf1302253eba6ebfb247c2bd97298b3e8d639c64388c2685df0852a75223205.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 85, + 292, + 277, + 300 + ], + "lines": [ + { + "bbox": [ + 85, + 292, + 277, + 300 + ], + "spans": [ + { + "bbox": [ + 85, + 292, + 277, + 300 + ], + "type": "text", + "content": "(a) A person riding a snowmobile down a snowy mountain." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 141, + 254, + 166, + 273 + ], + "blocks": [ + { + "bbox": [ + 141, + 254, + 166, + 273 + ], + "lines": [ + { + "bbox": [ + 141, + 254, + 166, + 273 + ], + "spans": [ + { + "bbox": [ + 141, + 254, + 166, + 273 + ], + "type": "image", + "image_path": "8720deab3e64e069231ca4b8ac39f369ff2d3bc4db5d1ce641d85042454a9adb.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 171, + 240, + 222, + 291 + ], + "blocks": [ + { + "bbox": [ + 171, + 240, + 222, + 291 + ], + "lines": [ + { + "bbox": [ + 171, + 240, + 222, + 291 + ], + "spans": [ + { + "bbox": [ + 171, + 240, + 222, + 291 + ], + "type": "image", + "image_path": "c19e69c270bd3dd827a7ad24217b440c964ba5972c41ae2d997cfd00739a40af.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 223, + 240, + 272, + 291 + ], + "blocks": [ + { + "bbox": [ + 223, + 240, + 272, + 291 + ], + "lines": [ + { + "bbox": [ + 223, + 240, + 272, + 291 + ], + "spans": [ + { + "bbox": [ + 223, + 240, + 272, + 291 + ], + "type": "image", + "image_path": "f9c922b45beea473a69c469cce46743cf4cd03b3e395510f4d5b8340718b1a7a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 274, + 240, + 323, + 291 + ], + "blocks": [ + { + "bbox": [ + 274, + 240, + 323, + 291 + ], + "lines": [ + { + "bbox": [ + 274, + 240, + 323, + 291 + ], + "spans": [ + { + "bbox": [ + 274, + 240, + 323, + 291 + ], + "type": "image", + "image_path": "94861a0f6fbd7d8d9f210d344e24bdb4f48b5ccadddb6b92d56cd71be05bf072.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 324, + 240, + 375, + 291 + ], + "blocks": [ + { + "bbox": [ + 324, + 240, + 375, + 291 + ], + "lines": [ + { + "bbox": [ + 324, + 240, + 375, + 291 + ], + "spans": [ + { + "bbox": [ + 324, + 240, + 375, + 291 + ], + "type": "image", + "image_path": "4add69d4bcbcf5355b1167a5620c4ed2e0c6277cee9dba226dc972508958ba0c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 376, + 240, + 425, + 291 + ], + "blocks": [ + { + "bbox": [ + 376, + 240, + 425, + 291 + ], + "lines": [ + { + "bbox": [ + 376, + 240, + 425, + 291 + ], + "spans": [ + { + "bbox": [ + 376, + 240, + 425, + 291 + ], + "type": "image", + "image_path": "c3c320c402799d3ca87ba4f23b2455f88a44baea6db7b6c2ae28aa18323ee175.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 427, + 240, + 476, + 291 + ], + "blocks": [ + { + "bbox": [ + 427, + 240, + 476, + 291 + ], + "lines": [ + { + "bbox": [ + 427, + 240, + 476, + 291 + ], + "spans": [ + { + "bbox": [ + 427, + 240, + 476, + 291 + ], + "type": "image", + "image_path": "cfa9cfeade01f1bf9b8b3793dc427b22403aeff2f25ee3024b74293740eab982.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 477, + 240, + 528, + 291 + ], + "blocks": [ + { + "bbox": [ + 477, + 240, + 528, + 291 + ], + "lines": [ + { + "bbox": [ + 477, + 240, + 528, + 291 + ], + "spans": [ + { + "bbox": [ + 477, + 240, + 528, + 291 + ], + "type": "image", + "image_path": "994cd997a8ce11e979216a9bb83571de2b21ad98f7b49e6367c6231f63367379.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 85, + 301, + 136, + 353 + ], + "blocks": [ + { + "bbox": [ + 85, + 301, + 136, + 353 + ], + "lines": [ + { + "bbox": [ + 85, + 301, + 136, + 353 + ], + "spans": [ + { + "bbox": [ + 85, + 301, + 136, + 353 + ], + "type": "image", + "image_path": "751983bbda54f3545057a0fe6f37bcda25d2012fa5b99865d0baac58aa5a3af0.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 85, + 354, + 268, + 363 + ], + "lines": [ + { + "bbox": [ + 85, + 354, + 268, + 363 + ], + "spans": [ + { + "bbox": [ + 85, + 354, + 268, + 363 + ], + "type": "text", + "content": "(b) A chef with a white hat preparing a gourmet meal." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 141, + 317, + 166, + 336 + ], + "blocks": [ + { + "bbox": [ + 141, + 317, + 166, + 336 + ], + "lines": [ + { + "bbox": [ + 141, + 317, + 166, + 336 + ], + "spans": [ + { + "bbox": [ + 141, + 317, + 166, + 336 + ], + "type": "image", + "image_path": "d0f42cdbfd9cac8254a7da54da51b0b2eeb091ed5ea3d6cfe48824531add0b71.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 171, + 301, + 222, + 351 + ], + "blocks": [ + { + "bbox": [ + 171, + 301, + 222, + 351 + ], + "lines": [ + { + "bbox": [ + 171, + 301, + 222, + 351 + ], + "spans": [ + { + "bbox": [ + 171, + 301, + 222, + 351 + ], + "type": "image", + "image_path": "eec11408165b12e44f08c8405468b482b7cecbcd12a2fdaab0b871d9260dde5d.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 223, + 301, + 272, + 351 + ], + "blocks": [ + { + "bbox": [ + 223, + 301, + 272, + 351 + ], + "lines": [ + { + "bbox": [ + 223, + 301, + 272, + 351 + ], + "spans": [ + { + "bbox": [ + 223, + 301, + 272, + 351 + ], + "type": "image", + "image_path": "d3b730321a091d650ea5372a859eb766bc5d3b562cf9f478aa106b0ffe925abe.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 274, + 301, + 323, + 351 + ], + "blocks": [ + { + "bbox": [ + 274, + 301, + 323, + 351 + ], + "lines": [ + { + "bbox": [ + 274, + 301, + 323, + 351 + ], + "spans": [ + { + "bbox": [ + 274, + 301, + 323, + 351 + ], + "type": "image", + "image_path": "8023b8eaa4ced1db4f86bdc0c490bdf103eb093f853227682274d01b00fc2a11.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 324, + 301, + 375, + 351 + ], + "blocks": [ + { + "bbox": [ + 324, + 301, + 375, + 351 + ], + "lines": [ + { + "bbox": [ + 324, + 301, + 375, + 351 + ], + "spans": [ + { + "bbox": [ + 324, + 301, + 375, + 351 + ], + "type": "image", + "image_path": "df09c6f36ec505e9d54a7a50f838df7d5a68a0b856cbba7dde0a387b2a8b0f37.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 376, + 301, + 425, + 351 + ], + "blocks": [ + { + "bbox": [ + 376, + 301, + 425, + 351 + ], + "lines": [ + { + "bbox": [ + 376, + 301, + 425, + 351 + ], + "spans": [ + { + "bbox": [ + 376, + 301, + 425, + 351 + ], + "type": "image", + "image_path": "4e108b5dd1091674a527545dc0da32d3d0556b2ebcbedfe87ec18e6efa74d444.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 427, + 301, + 476, + 351 + ], + "blocks": [ + { + "bbox": [ + 427, + 301, + 476, + 351 + ], + "lines": [ + { + "bbox": [ + 427, + 301, + 476, + 351 + ], + "spans": [ + { + "bbox": [ + 427, + 301, + 476, + 351 + ], + "type": "image", + "image_path": "f2db678dfc32d0a8603d3547093bec5d13df6c9a8c0e35427d47b81ba10d7208.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 477, + 301, + 528, + 351 + ], + "blocks": [ + { + "bbox": [ + 477, + 301, + 528, + 351 + ], + "lines": [ + { + "bbox": [ + 477, + 301, + 528, + 351 + ], + "spans": [ + { + "bbox": [ + 477, + 301, + 528, + 351 + ], + "type": "image", + "image_path": "56f713d1351b73f7e499e3ce9fc2d2200489a40d199bbb831756b51fb4456ee1.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 85, + 364, + 135, + 414 + ], + "blocks": [ + { + "bbox": [ + 85, + 364, + 135, + 414 + ], + "lines": [ + { + "bbox": [ + 85, + 364, + 135, + 414 + ], + "spans": [ + { + "bbox": [ + 85, + 364, + 135, + 414 + ], + "type": "image", + "image_path": "26b41eca45b63bd8c695130f674f01b8d3ea93facc29603a94ddc3daf7f9c720.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 85, + 414, + 258, + 423 + ], + "lines": [ + { + "bbox": [ + 85, + 414, + 258, + 423 + ], + "spans": [ + { + "bbox": [ + 85, + 414, + 258, + 423 + ], + "type": "text", + "content": "(c) A snow leopard is walking on the side of a cliff." + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 141, + 381, + 166, + 399 + ], + "blocks": [ + { + "bbox": [ + 141, + 381, + 166, + 399 + ], + "lines": [ + { + "bbox": [ + 141, + 381, + 166, + 399 + ], + "spans": [ + { + "bbox": [ + 141, + 381, + 166, + 399 + ], + "type": "image", + "image_path": "b8b14fab5ed06744e00a6b10345d3c82f6f257f194537c5dff405277aa69a734.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 171, + 364, + 222, + 414 + ], + "blocks": [ + { + "bbox": [ + 171, + 364, + 222, + 414 + ], + "lines": [ + { + "bbox": [ + 171, + 364, + 222, + 414 + ], + "spans": [ + { + "bbox": [ + 171, + 364, + 222, + 414 + ], + "type": "image", + "image_path": "ab704978e737c61b7242c82a3e201f987a4d4125c050fb0f8aab2231efa59f2d.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 223, + 364, + 272, + 414 + ], + "blocks": [ + { + "bbox": [ + 223, + 364, + 272, + 414 + ], + "lines": [ + { + "bbox": [ + 223, + 364, + 272, + 414 + ], + "spans": [ + { + "bbox": [ + 223, + 364, + 272, + 414 + ], + "type": "image", + "image_path": "b2d0a23fe3c75f4a3b2a1e2c916ad27de42e92de3b8f93c27b01faea61cd9eee.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 274, + 364, + 323, + 414 + ], + "blocks": [ + { + "bbox": [ + 274, + 364, + 323, + 414 + ], + "lines": [ + { + "bbox": [ + 274, + 364, + 323, + 414 + ], + "spans": [ + { + "bbox": [ + 274, + 364, + 323, + 414 + ], + "type": "image", + "image_path": "13069143961c353f004a316d264531ca0efdaab5df16af5c6c0f37b08bca01fb.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 324, + 364, + 375, + 414 + ], + "blocks": [ + { + "bbox": [ + 324, + 364, + 375, + 414 + ], + "lines": [ + { + "bbox": [ + 324, + 364, + 375, + 414 + ], + "spans": [ + { + "bbox": [ + 324, + 364, + 375, + 414 + ], + "type": "image", + "image_path": "5c047608058f83ba1d346c30e6cdebaf97f82dc4c2a035ff67d3e80484b5d678.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 376, + 364, + 425, + 414 + ], + "blocks": [ + { + "bbox": [ + 376, + 364, + 425, + 414 + ], + "lines": [ + { + "bbox": [ + 376, + 364, + 425, + 414 + ], + "spans": [ + { + "bbox": [ + 376, + 364, + 425, + 414 + ], + "type": "image", + "image_path": "ae4eec9ce99811c00b66bb2298a5d2d7e3749885df9f0edf73766ccaff47f5bf.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 427, + 364, + 476, + 414 + ], + "blocks": [ + { + "bbox": [ + 427, + 364, + 476, + 414 + ], + "lines": [ + { + "bbox": [ + 427, + 364, + 476, + 414 + ], + "spans": [ + { + "bbox": [ + 427, + 364, + 476, + 414 + ], + "type": "image", + "image_path": "56b5c5aec26818581508635bec94422b854f827ad3669c3ce6462aa61c91a700.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 477, + 364, + 528, + 414 + ], + "blocks": [ + { + "bbox": [ + 477, + 364, + 528, + 414 + ], + "lines": [ + { + "bbox": [ + 477, + 364, + 528, + 414 + ], + "spans": [ + { + "bbox": [ + 477, + 364, + 528, + 414 + ], + "type": "image", + "image_path": "0f747f8280de2d9dd608531482e32f6f9ddf71ceddc543b6a24ab2a8b7bed826.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 85, + 425, + 135, + 475 + ], + "blocks": [ + { + "bbox": [ + 85, + 425, + 135, + 475 + ], + "lines": [ + { + "bbox": [ + 85, + 425, + 135, + 475 + ], + "spans": [ + { + "bbox": [ + 85, + 425, + 135, + 475 + ], + "type": "image", + "image_path": "a7c5c0224f9fc4cea4f27532b909aa22f9ba1ba6ac247c47da9693cb3b3e2343.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 85, + 477, + 286, + 487 + ], + "lines": [ + { + "bbox": [ + 85, + 477, + 286, + 487 + ], + "spans": [ + { + "bbox": [ + 85, + 477, + 286, + 487 + ], + "type": "text", + "content": "(d) A lemur walking on a xylophone, in a tropical rainforest." + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 136, + 425, + 186, + 475 + ], + "blocks": [ + { + "bbox": [ + 136, + 425, + 186, + 475 + ], + "lines": [ + { + "bbox": [ + 136, + 425, + 186, + 475 + ], + "spans": [ + { + "bbox": [ + 136, + 425, + 186, + 475 + ], + "type": "image", + "image_path": "7ccc0f4ca6dac6e3014b4898e91026ea62a0e883d622bbb6fc81e8ee139d5f21.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 187, + 425, + 237, + 475 + ], + "blocks": [ + { + "bbox": [ + 187, + 425, + 237, + 475 + ], + "lines": [ + { + "bbox": [ + 187, + 425, + 237, + 475 + ], + "spans": [ + { + "bbox": [ + 187, + 425, + 237, + 475 + ], + "type": "image", + "image_path": "15627c4a157921f2733760f2014733f111a777a81eae020a3b862c695ddfab69.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 238, + 425, + 288, + 475 + ], + "blocks": [ + { + "bbox": [ + 238, + 425, + 288, + 475 + ], + "lines": [ + { + "bbox": [ + 238, + 425, + 288, + 475 + ], + "spans": [ + { + "bbox": [ + 238, + 425, + 288, + 475 + ], + "type": "image", + "image_path": "a318956bf9177ef3fd2b5aed094ce1c3b5cb804deac25da1fb6a29d75b93ef7d.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 292, + 439, + 317, + 457 + ], + "blocks": [ + { + "bbox": [ + 292, + 439, + 317, + 457 + ], + "lines": [ + { + "bbox": [ + 292, + 439, + 317, + 457 + ], + "spans": [ + { + "bbox": [ + 292, + 439, + 317, + 457 + ], + "type": "image", + "image_path": "61fa9e32ac0c0dbf5ce834cbbe3c598b6242cb1c69409092295d66bef9e942fb.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 323, + 425, + 374, + 475 + ], + "blocks": [ + { + "bbox": [ + 323, + 425, + 374, + 475 + ], + "lines": [ + { + "bbox": [ + 323, + 425, + 374, + 475 + ], + "spans": [ + { + "bbox": [ + 323, + 425, + 374, + 475 + ], + "type": "image", + "image_path": "70c75a3b854464b13d5d2394cec35b8b447a06981cc275350815ae1ac66bca1d.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 376, + 425, + 425, + 475 + ], + "blocks": [ + { + "bbox": [ + 376, + 425, + 425, + 475 + ], + "lines": [ + { + "bbox": [ + 376, + 425, + 425, + 475 + ], + "spans": [ + { + "bbox": [ + 376, + 425, + 425, + 475 + ], + "type": "image", + "image_path": "181b8d1051f05bacb30f0967b30341b7fc0a3510eae90c0fe6e04a9699e0697e.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 427, + 425, + 476, + 475 + ], + "blocks": [ + { + "bbox": [ + 427, + 425, + 476, + 475 + ], + "lines": [ + { + "bbox": [ + 427, + 425, + 476, + 475 + ], + "spans": [ + { + "bbox": [ + 427, + 425, + 476, + 475 + ], + "type": "image", + "image_path": "7c38996f49cb0016440a54684837cb9765c5fc8e6c26df167d1e6f630871548e.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 477, + 425, + 528, + 475 + ], + "blocks": [ + { + "bbox": [ + 477, + 425, + 528, + 475 + ], + "lines": [ + { + "bbox": [ + 477, + 425, + 528, + 475 + ], + "spans": [ + { + "bbox": [ + 477, + 425, + 528, + 475 + ], + "type": "image", + "image_path": "b0254591e78a5349e840a51972a571ef830b059b7d73f37410fad318fa7fe790.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 89, + 497, + 124, + 530 + ], + "blocks": [ + { + "bbox": [ + 89, + 497, + 124, + 530 + ], + "lines": [ + { + "bbox": [ + 89, + 497, + 124, + 530 + ], + "spans": [ + { + "bbox": [ + 89, + 497, + 124, + 530 + ], + "type": "image", + "image_path": "c6e4913e12352c0c54f7e5b3b19fb17149ed05d06ca61c3c52ea7f1864a7ef4c.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 141, + 506, + 166, + 523 + ], + "blocks": [ + { + "bbox": [ + 141, + 506, + 166, + 523 + ], + "lines": [ + { + "bbox": [ + 141, + 506, + 166, + 523 + ], + "spans": [ + { + "bbox": [ + 141, + 506, + 166, + 523 + ], + "type": "image", + "image_path": "550195189daa5648d4bba597da08af01c98116d3cad69f35bf35880fb5e96e7a.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 171, + 496, + 222, + 538 + ], + "blocks": [ + { + "bbox": [ + 171, + 496, + 222, + 538 + ], + "lines": [ + { + "bbox": [ + 171, + 496, + 222, + 538 + ], + "spans": [ + { + "bbox": [ + 171, + 496, + 222, + 538 + ], + "type": "image", + "image_path": "94fb323760576ac2860c9b357aeba66cde2f3a5150b78451f62597c429f11a7d.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 84, + 540, + 269, + 550 + ], + "lines": [ + { + "bbox": [ + 84, + 540, + 269, + 550 + ], + "spans": [ + { + "bbox": [ + 84, + 540, + 269, + 550 + ], + "type": "text", + "content": "(e) A snowman in a forest caught in a burst of flames." + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 552, + 555, + 585 + ], + "lines": [ + { + "bbox": [ + 55, + 552, + 555, + 585 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 555, + 585 + ], + "type": "text", + "content": "Figure 1. IM-Zero introduces a zero-shot framework for instance-level motion controllable video generation. IM-Zero allows users to: (a) Control overall instance motion. (b) Control instance subpart motion. (c) Control instance motion and specify instance shape via mask. (d) Perform motion transfer by reference video. (e) Perform high-quality text-to-video generation." + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_caption" + } + ], + "index": 46 + }, + { + "type": "image", + "bbox": [ + 223, + 491, + 272, + 538 + ], + "blocks": [ + { + "bbox": [ + 223, + 491, + 272, + 538 + ], + "lines": [ + { + "bbox": [ + 223, + 491, + 272, + 538 + ], + "spans": [ + { + "bbox": [ + 223, + 491, + 272, + 538 + ], + "type": "image", + "image_path": "d05828622d257352665a2b9033966df61fb65749578e567143e73490968b13b2.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + } + ], + "index": 47 + }, + { + "type": "image", + "bbox": [ + 274, + 489, + 323, + 538 + ], + "blocks": [ + { + "bbox": [ + 274, + 489, + 323, + 538 + ], + "lines": [ + { + "bbox": [ + 274, + 489, + 323, + 538 + ], + "spans": [ + { + "bbox": [ + 274, + 489, + 323, + 538 + ], + "type": "image", + "image_path": "d44c0c7d45a6d1bdd8e8bfabdd7bbe88df104a1b213e8598c99c23f32a3596dd.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + } + ], + "index": 48 + }, + { + "type": "image", + "bbox": [ + 324, + 489, + 375, + 538 + ], + "blocks": [ + { + "bbox": [ + 324, + 489, + 375, + 538 + ], + "lines": [ + { + "bbox": [ + 324, + 489, + 375, + 538 + ], + "spans": [ + { + "bbox": [ + 324, + 489, + 375, + 538 + ], + "type": "image", + "image_path": "fea318659d0bd51f124ad1fb421dce5436313f2111de241a2425f631a13c7aaf.jpg" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_body" + } + ], + "index": 49 + }, + { + "type": "image", + "bbox": [ + 376, + 489, + 425, + 538 + ], + "blocks": [ + { + "bbox": [ + 376, + 489, + 425, + 538 + ], + "lines": [ + { + "bbox": [ + 376, + 489, + 425, + 538 + ], + "spans": [ + { + "bbox": [ + 376, + 489, + 425, + 538 + ], + "type": "image", + "image_path": "461c6585ef807732b233b444866892bc336966b622e5d3897c09936ddc9266f3.jpg" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_body" + } + ], + "index": 50 + }, + { + "type": "image", + "bbox": [ + 427, + 489, + 476, + 538 + ], + "blocks": [ + { + "bbox": [ + 427, + 489, + 476, + 538 + ], + "lines": [ + { + "bbox": [ + 427, + 489, + 476, + 538 + ], + "spans": [ + { + "bbox": [ + 427, + 489, + 476, + 538 + ], + "type": "image", + "image_path": "31bda9be05576e5ce87ddbb0ae85dcdfdeb4ed3ef8f7b1c7b8975f813a8e582e.jpg" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_body" + } + ], + "index": 51 + }, + { + "type": "image", + "bbox": [ + 477, + 489, + 528, + 538 + ], + "blocks": [ + { + "bbox": [ + 477, + 489, + 528, + 538 + ], + "lines": [ + { + "bbox": [ + 477, + 489, + 528, + 538 + ], + "spans": [ + { + "bbox": [ + 477, + 489, + 528, + 538 + ], + "type": "image", + "image_path": "a096790f957ad8d8cde071b5e09dcb9e215445ea05b42377c3314b3fde046092.jpg" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_body" + } + ], + "index": 52 + }, + { + "bbox": [ + 151, + 594, + 200, + 606 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 594, + 200, + 606 + ], + "spans": [ + { + "bbox": [ + 151, + 594, + 200, + 606 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 55 + }, + { + "bbox": [ + 55, + 610, + 295, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 610, + 295, + 682 + ], + "spans": [ + { + "bbox": [ + 55, + 610, + 295, + 682 + ], + "type": "text", + "content": "Controllability of video generation has been recently concerned in addition to the quality of generated videos. The main challenge to controllable video generation is to synthesize videos based on user-specified instance spatial locations and movement trajectories. However, existing methods suffer from a dilemma between the resource con" + } + ] + } + ], + "index": 56 + }, + { + "bbox": [ + 313, + 596, + 555, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 596, + 555, + 704 + ], + "spans": [ + { + "bbox": [ + 313, + 596, + 555, + 704 + ], + "type": "text", + "content": "sumption, generation quality, and user controllability. As an efficient alternative to prohibitive training-based video generation, existing zero-shot video generation methods cannot generate high-quality and motion-consistent videos under the control of layouts and movement trajectories. In this paper, we propose a novel zero-shot method named IM-Zero that ameliorates instance-level motion controllable video generation with enhanced control accuracy, motion consistency, and richness of details to address this prob" + } + ] + } + ], + "index": 57 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 223, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 223, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 223, + 703 + ], + "type": "text", + "content": "*These authors contributed equally to this work." + } + ] + } + ], + "index": 58 + }, + { + "bbox": [ + 67, + 703, + 283, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 283, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 283, + 712 + ], + "type": "text", + "content": "† Corresponding authors: Xiaopeng Zhang; Wenrui Dai; Junni Zou." + } + ] + } + ], + "index": 59 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7265" + } + ] + } + ], + "index": 61 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 336 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 336 + ], + "type": "text", + "content": "lem. Specifically, we first present a motion generation stage that extracts motion and textural guidance from keyframe candidates from pre-trained grounded text-to-image model to generate the desired coarse motion video. Subsequently, we develop a video refinement stage that injects the motion priors of pre-trained text-to-video models and detail priors of pre-trained text-to-image models into the latents of coarse motion videos to further enhance video motion consistency and richness of details. To our best knowledge, IM-Zero is the first to simultaneously achieve high-quality video generation and allow to control both layouts and movement trajectories in a zero-shot manner. Extensive experiments demonstrate that IM-Zero outperforms existing methods in terms of video quality, inter-frame consistency, and the alignment of location and trajectory. Furthermore, compared with existing methods, IM-Zero enjoys extra advantages of versatility in video generation, including motion control of subparts within instances, finer control of specifying instance shapes via masks, and more difficult tasks of motion transfer for customizing fine-grained motion patterns through reference videos and high-quality text-to-video generation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 358, + 136, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 358, + 136, + 371 + ], + "spans": [ + { + "bbox": [ + 55, + 358, + 136, + 371 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 378, + 295, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 378, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 378, + 295, + 498 + ], + "type": "text", + "content": "Demands for enhanced controllability of text-to-video (T2V) generation are booming with the increasing quality of generated videos [3, 6, 7, 13-15, 22, 37, 38, 40, 62]. One of the critical problems lie in the precise control over the spatial layout and trajectories of instances within the generated videos. Existing methods fail to achieve fine-grained control by solely relying on text prompts to steer pre-trained T2V models, even employing fundamental T2V models with large-scale pre-training like Sora [4], KLING, and Doubao-PixelDance [53]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 498, + 295, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 629 + ], + "type": "text", + "content": "Existing methods usually extend ControlNet [55] for controllable video generation with frame-by-frame fine-grained control signals, including Canny, depth, and HED signals [9, 16, 36, 56]. To obtain dense control signals, these methods impose high demands on the input end and are limited in practical applications. One approach is to extract control signals from reference videos but it is hard to obtain reference videos perfectly matching the desired motion patterns. The other alternative requires manually crafting the dense control signals. However, it is also non-trivial to craft frame-by-frame control signals without expertise." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "content": "Recent attempts [20, 29, 32] use spatial locations (i.e., bounding boxes) and motion trajectories of instances, along with corresponding prompts, to create desired videos. The locations and motion trajectories of instances are extracted from real videos and are encoded as additional control conditions for joint training with the video generation models. However, existing training-based methods can only control" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": "the movement trajectory [11, 41, 44, 52, 58, 65], since substantial data and training resources are required to simultaneously learn controlling layouts and movement trajectories for video generation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 122, + 555, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 122, + 555, + 314 + ], + "spans": [ + { + "bbox": [ + 313, + 122, + 555, + 314 + ], + "type": "text", + "content": "In this paper, we develop a novel training-free approach to address this problem. Existing training-free methods are restricted in obtaining accurate location and movement effects by manipulating spatial attention maps and temporal attention maps to generate instances at specified locations [20, 29, 32]. These methods convert location signals in the pixel space provided by users to a latent space with reduced dimension for manipulating attention maps, and inevitably compromise the precision of location control and obtain inaccurate instance placement. Excessive constraints on the attention maps also disrupt the motion prior of video diffusion. Moreover, frame interaction is neglected during the manipulation, which also causes unreasonable motion effects and unnatural generated videos. Additionally, these methods also suffer from generating low-quality videos with choppy motion and insufficient details." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 316, + 554, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 554, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 554, + 532 + ], + "type": "text", + "content": "We find that both the inaccurate instance positioning and the unnatural motion stem from the limitations of attention map manipulation, which is a compromise solution to mitigate the lack of a clear understanding of instance concepts in vanilla latent diffusion models. It introduces constraints to enforce the instances to appear at designated positions. Contrary to attention map manipulation for video generation, instance-level knowledge is usually considered for image generation. Existing grounded T2I models like GLIGEN [25] and InstanceDiffusion [39] encapsulate extensive instance-level knowledge, but they are confined to generating images rather than consistent videos. Thus, we propose to generate coarse motion videos aligning with the user input and exhibit natural motion using instance-level priors in grounded T2I models, and further refine coarse motion videos by harnessing the motion and detail priors of T2V models for motion and detail injection in a zero-shot manner to boost the motion consistency and richness of details." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": "To this end, we propose a zero-shot and plug-and-play method named IM-Zero that allows to simultaneously control instance layouts and movement trajectories for video generation. IM-Zero consists of two stages, i.e., a motion generation stage that acquires a coarse motion video, and a video refinement stage to enhance the quality of the coarse motion video. In the motion generation stage, spatial location signals (e.g., bounding boxes and masks) are first augmented according to the input trajectory to obtain a sequence of layout control signals. A batch of keyframe candidates with texture consistency is then obtained by inflating and transforming the pre-trained grounded T2I model, and is used to extract motion guidance and textural guidance for generating the desired coarse motion video. In the video refinement stage, we encode the coarse motion video" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7266" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "text", + "content": "into video latents, and inject the motion and detail priors extracted from the pre-trained T2V models into the video latents to improve motion consistency and enhance the richness of details, respectively." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 121, + 246, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 121, + 246, + 131 + ], + "spans": [ + { + "bbox": [ + 67, + 121, + 246, + 131 + ], + "type": "text", + "content": "Our contributions are summarized as below." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 133, + 295, + 335 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 56, + 133, + 295, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 133, + 295, + 180 + ], + "spans": [ + { + "bbox": [ + 56, + 133, + 295, + 180 + ], + "type": "text", + "content": "- We propose a novel framework named IM-Zero for zero-shot motion controllable video generation. IM-Zero supports various types of layout inputs and enables fine-grained motion control of the instances and their subparts." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 181, + 295, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 181, + 295, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 181, + 295, + 228 + ], + "type": "text", + "content": "- We present motion generation that generates keyframe candidates with the layout control capability of grounded T2I models to extract motion and textural guidance for coarse motion video generation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 228, + 295, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 295, + 287 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 295, + 287 + ], + "type": "text", + "content": "- We develop zero-shot video refinement that injects the motion and detail priors of the pre-trained text-to-video model and text-to-image model into the latents of the coarse motion video to further enhance video motion consistency and richness of details." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 288, + 295, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 288, + 295, + 335 + ], + "spans": [ + { + "bbox": [ + 56, + 288, + 295, + 335 + ], + "type": "text", + "content": "- We further achieve motion transfer to customize complex motion patterns using reference videos and support high-quality text-to-video generation with only text input beyond instance layout and movement trajectory controls." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 336, + 295, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 336, + 295, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 336, + 295, + 467 + ], + "type": "text", + "content": "Extensive experiments demonstrate that IM-Zero outperforms existing methods in terms of video quality, interframe consistency, and the alignment of location and trajectory. Compared with existing methods, IM-Zero enjoys extra advantages of versatility in video generation, including motion control of subparts within instances and finer control with instance shapes specified by masked inputs. Furthermore, it can be adapted to motion transfer tasks to support the customized fine-grained motion patterns through reference videos and can also be adapted to high-quality text-to-video generation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 478, + 142, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 478, + 142, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 478, + 142, + 491 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 498, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 677 + ], + "type": "text", + "content": "Grounded Text-to-image Generation. Grounded text-to-image (T2I) models aim to generate images that align with input layout information. Some works specialize in modeling layout information of the bounding box type. Notable examples include GLIGEN [25], InstanceDiffusion [39], BoxDiff [46], MultiDiffusion [1], ReCo [49], Layout-Guidance [8], MIGC [63, 64], and TraDiffusion [43]. We believe that some training-based methods such as InstanceDiffusion [39] have enabled models to grasp the concept of instances. Their capability aligns to some extent with our objective of achieving location control and trajectory control in video generation. However, these methods are limited to generating single images and cannot produce consistent videos. Therefore, we aim to leverage Grounded T2I models to benefit controllable video generation tasks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "type": "text", + "content": "Location and Trajectory Control in Video Generation. Diffusion models have demonstrated exceptional capabilities in image generation [45, 61], video generation [56, 57]," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 555, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 361 + ], + "type": "text", + "content": "3D generation [10, 27], and other visual tasks [18, 54, 60]. With the improvement in video generation quality [3, 6, 7, 13-15, 22, 37, 38, 40, 47, 62], there is a growing interest in achieving generation controllability, with a particular focus on controlling the spatial location and movement trajectories of instances. Currently, training-based methods can only achieve limited control over trajectory guidance, at the expense of requiring substantial amounts of data and computational resources, with typical examples being DragNUWA [52], DragAnything [44], DragVideo [11], Tora [58], TrackGo [65] and MotionCtrl [41]. In contrast, some training-free methods can concurrently control both spatial location and movement trajectories. For instance, Peekaboo [20] and TrailBlazer [29] enforce constraints on attention maps to position instances at specified locations. However, this approach can lead to issues such as inaccuracies in location and poor motion effects. FreeTraj [32] follows a similar attention map manipulation approach but introduces noise flow control. In addition to the aforementioned challenges, it also introduces difficulties in handling bounding boxes that undergo size changes. Therefore, we propose a zero-shot method that surpasses attention map manipulation to address the aforementioned challenges and enhance video quality and motion consistency." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 385, + 370, + 398 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 385, + 370, + 398 + ], + "spans": [ + { + "bbox": [ + 314, + 385, + 370, + 398 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 410, + 400, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 410, + 400, + 422 + ], + "spans": [ + { + "bbox": [ + 314, + 410, + 400, + 422 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 433, + 555, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 433, + 555, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 433, + 555, + 589 + ], + "type": "text", + "content": "Grounded Text-to-image Diffusion Models. Grounded text-to-image (T2I) diffusion models [25, 39, 63, 64] are capable of generating instances at user-specified locations by incorporating the layout input as conditions for joint training. InstanceDiffusion [39] supports four types of layout input: bounding box, mask, point, and scribble. It inserts learnable UniFusion blocks between the self-attention and cross-attention layers of the T2I model. UniFusion sample the instance layout as a 2D points " + }, + { + "bbox": [ + 313, + 433, + 555, + 589 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 313, + 433, + 555, + 589 + ], + "type": "text", + "content": " sequence and convert it to layout token by " + }, + { + "bbox": [ + 313, + 433, + 555, + 589 + ], + "type": "inline_equation", + "content": "\\mathbf{g} = \\mathrm{MLP}\\left([ \\tau_{\\theta}(\\mathbf{c}), \\gamma(\\mathbf{p}) \\right]" + }, + { + "bbox": [ + 313, + 433, + 555, + 589 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 433, + 555, + 589 + ], + "type": "inline_equation", + "content": "\\gamma(\\cdot)" + }, + { + "bbox": [ + 313, + 433, + 555, + 589 + ], + "type": "text", + "content": " is a Fourier mapping and " + }, + { + "bbox": [ + 313, + 433, + 555, + 589 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 313, + 433, + 555, + 589 + ], + "type": "text", + "content": " is the text prompt. Then the layout token is fused with the backbone visual tokens via a learnable masked self-attention layer." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 594, + 556, + 716 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 716 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 716 + ], + "type": "text", + "content": "Text-to-video Diffusion Models. Significant progress has also been made in Text-to-Video (T2V) diffusion models. While some T2V models are trained from scratch, the majority of them are trained based on existing T2I models. A typical example isAnimateDiff [13], which utilizes pre-trained T2I models as the spatial layer and trains additional motion modeling modules to reconstruct N frames clean video latents " + }, + { + "bbox": [ + 313, + 594, + 556, + 716 + ], + "type": "inline_equation", + "content": "z^{1:N}" + }, + { + "bbox": [ + 313, + 594, + 556, + 716 + ], + "type": "text", + "content": " from i.i.d Gaussian noise conditioned on the text prompt " + }, + { + "bbox": [ + 313, + 594, + 556, + 716 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 594, + 556, + 716 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 313, + 594, + 556, + 716 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\theta) = \\mathbb{E}_{z_0^{1:N},\\epsilon ,c_t,t}\\left[\\| \\epsilon -\\hat{\\epsilon}_\\theta \\left(z_t^{1:N},c_t,t\\right)\\right]" + }, + { + "bbox": [ + 313, + 594, + 556, + 716 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7267" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 54, + 531, + 208 + ], + "blocks": [ + { + "bbox": [ + 82, + 54, + 531, + 208 + ], + "lines": [ + { + "bbox": [ + 82, + 54, + 531, + 208 + ], + "spans": [ + { + "bbox": [ + 82, + 54, + 531, + 208 + ], + "type": "image", + "image_path": "c19118d955a0bf730caca8b242170a566cbfeed7a179503c644b0de4dea17533.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 217, + 555, + 251 + ], + "lines": [ + { + "bbox": [ + 55, + 217, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 555, + 251 + ], + "type": "text", + "content": "Figure 2. The Motion Generation Stage consists of two parts: Keyframe Candidates Generation and Coarse Motion Video Generation. In the first part, we augment the user input to obtain Layout, then generate keyframe candidates Frame with an inflated grounded T2I model. In the second part, we extract motion guidance and textural guidance from Frame to generate a coarse motion video." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 265, + 171, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 265, + 171, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 265, + 171, + 277 + ], + "type": "text", + "content": "3.2. Overall Framework" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 283, + 296, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 283, + 296, + 535 + ], + "spans": [ + { + "bbox": [ + 55, + 283, + 296, + 535 + ], + "type": "text", + "content": "IM-Zero framework consists of two critical stages: Motion Generation Stage and Video Refinement Stage. In Section 3.1, we first briefly introduce the fundamentals of grounded T2I and T2V diffusion models. Then, we introduce the Motion Generation Stage in Section 3.3 as shown in Figure 2. This stage generates keyframe candidates and extracts guidance from them to generate a coarse motion video. Section 3.4 introduces the Video Refinement Stage as shown in Figure 3. This stage utilizes motion prior of T2V models to perform Motion Injection and utilizes the detail-generation capacity of T2I models to perform Detail Injection, further enhancing the visual quality of coarse motion video. Lastly, we also offer an option for motion transfer via reference video input, and we will explain how to easily modify IM-Zero in this case. As illustrated in Figure 1, IM-Zero achieves motion control for overall instances or subparts and allows users to further specify instance shapes as the grounded T2I model can generate keyframe candidates via masks. IM-Zero can perform motion transfer. Besides, IM-Zero allows high-quality T2V generation by refining any video via the Video Refinement Stage." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 540, + 197, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 197, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 197, + 553 + ], + "type": "text", + "content": "3.3. Motion Generation Stage" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": "The core task involves presenting suitable motion patterns in the video, including two key aspects: the user-desired instance spatial location and the user-desired instance movement. Previous training-free methods achieved this task by manipulating attention maps but encountered issues such as inaccurate location, poor motion effects, and low video quality, as discussed in Section 1. IM-Zero aims to achieve the decoupling of instance spatial location and instance movement to enhance the performance of both aspects. For the former, we utilize a grounded T2I model [39] to generate instances at user-specified locations. For the latter, we depart from the previous approach of imposing excessive constraints on the attention map and instead efficiently" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 266, + 498, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 498, + 278 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 498, + 278 + ], + "type": "text", + "content": "leverage the motion priors of the T2V model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 281, + 555, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 555, + 544 + ], + "type": "text", + "content": "Keyframe Candidates Generation. Considering the user input " + }, + { + "bbox": [ + 313, + 281, + 555, + 544 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 281, + 555, + 544 + ], + "type": "text", + "content": " layout signals and " + }, + { + "bbox": [ + 313, + 281, + 555, + 544 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 281, + 555, + 544 + ], + "type": "text", + "content": " trajectory segments, We denote the layout signals as " + }, + { + "bbox": [ + 313, + 281, + 555, + 544 + ], + "type": "inline_equation", + "content": "\\mathbf{Layout} = \\{s_i|i\\sim [1,2,\\dots S]\\}" + }, + { + "bbox": [ + 313, + 281, + 555, + 544 + ], + "type": "text", + "content": " (i.e., bounding boxes) and the movement trajectory as Traj = " + }, + { + "bbox": [ + 313, + 281, + 555, + 544 + ], + "type": "inline_equation", + "content": "\\{l_i|i\\sim [1,2,\\dots ,L]\\}" + }, + { + "bbox": [ + 313, + 281, + 555, + 544 + ], + "type": "text", + "content": ". We ensure that each " + }, + { + "bbox": [ + 313, + 281, + 555, + 544 + ], + "type": "inline_equation", + "content": "l_{i}" + }, + { + "bbox": [ + 313, + 281, + 555, + 544 + ], + "type": "text", + "content": " segment has a layout signal at both endpoints, for the endpoint without a layout signal, we select the closest " + }, + { + "bbox": [ + 313, + 281, + 555, + 544 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 313, + 281, + 555, + 544 + ], + "type": "text", + "content": " from Layout and duplicate it here, obtaining the augmented layout signals " + }, + { + "bbox": [ + 313, + 281, + 555, + 544 + ], + "type": "inline_equation", + "content": "\\mathbf{Layout} = \\{s_i|i\\sim [1,2,\\dots,S']\\}" + }, + { + "bbox": [ + 313, + 281, + 555, + 544 + ], + "type": "text", + "content": ". These signals are then input into the grounded T2I model to generate keyframe candidates " + }, + { + "bbox": [ + 313, + 281, + 555, + 544 + ], + "type": "inline_equation", + "content": "\\mathbf{Frame} = \\{f_i|i\\sim [1,2,\\dots,S']\\}" + }, + { + "bbox": [ + 313, + 281, + 555, + 544 + ], + "type": "text", + "content": ". Since generating each image separately can lead to significant inconsistencies, following [22], we inflate the grounded T2I model and replace each self-attention in UNet with cross-frame attention (CFA). The CFA for the " + }, + { + "bbox": [ + 313, + 281, + 555, + 544 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 281, + 555, + 544 + ], + "type": "text", + "content": "-th keyframe candidate " + }, + { + "bbox": [ + 313, + 281, + 555, + 544 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 313, + 281, + 555, + 544 + ], + "type": "text", + "content": " is defined as CFA " + }, + { + "bbox": [ + 313, + 281, + 555, + 544 + ], + "type": "inline_equation", + "content": "(Q^{i},K^{1:N_{\\mathrm{S^{\\prime}}}},V^{1:N_{\\mathrm{S^{\\prime}}}}) =" + }, + { + "bbox": [ + 313, + 281, + 555, + 544 + ], + "type": "text", + "content": " Softmax " + }, + { + "bbox": [ + 313, + 281, + 555, + 544 + ], + "type": "inline_equation", + "content": "(Q^{i}(K^{1})^{T} / \\sqrt{c})V^{1}" + }, + { + "bbox": [ + 313, + 281, + 555, + 544 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 281, + 555, + 544 + ], + "type": "inline_equation", + "content": "Q,K" + }, + { + "bbox": [ + 313, + 281, + 555, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 281, + 555, + 544 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 281, + 555, + 544 + ], + "type": "text", + "content": " denote the queries, keys and values. The first candidate provides key and value for all following ones, which can be interpreted from an image editing perspective as querying relevant content and textures from the first candidate to ensure consistency in textures [5, 22]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 546, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 556, + 714 + ], + "type": "text", + "content": "Coarse Motion Video Generation. After obtaining Frame, we utilize these candidates to generate a coarse motion video reflecting the user's input. To ensure smooth and natural instance motion, we have avoided the previous approach of directly enforcing excessive constraints on the attention map, which could disrupt the T2V motion priors. Instead, we extract two types of guidance from Frame to guide the T2V model. (1) Motion Guidance. To ensure the fidelity of the video to the user's input, encompassing spatial location and motion patterns, we extract foreground instances from Frame and convert them into Canny edge images. Subsequently, we employ ControlNet [12, 16, 55] to guide the generation of the coarse motion video. The rationale behind our approach is elucidated as follows: firstly," + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7268" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 55, + 366, + 115 + ], + "blocks": [ + { + "bbox": [ + 82, + 55, + 366, + 115 + ], + "lines": [ + { + "bbox": [ + 82, + 55, + 366, + 115 + ], + "spans": [ + { + "bbox": [ + 82, + 55, + 366, + 115 + ], + "type": "image", + "image_path": "f7b89de37622acb7658d1156de3b5f92d9fdc196fdcae0f3ffba79510d6691b8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 82, + 118, + 366, + 182 + ], + "blocks": [ + { + "bbox": [ + 82, + 118, + 366, + 182 + ], + "lines": [ + { + "bbox": [ + 82, + 118, + 366, + 182 + ], + "spans": [ + { + "bbox": [ + 82, + 118, + 366, + 182 + ], + "type": "image", + "image_path": "537bb9c31e78acb36f5242d41bb91e058f816e55c24b1e19f1ffbd6f2998e43f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 198, + 555, + 254 + ], + "lines": [ + { + "bbox": [ + 54, + 198, + 555, + 254 + ], + "spans": [ + { + "bbox": [ + 54, + 198, + 555, + 254 + ], + "type": "text", + "content": "Figure 3. In Video Refinement Stage, we diffuse coarser motion video latents and perform motion injection and detail injection in the reverse process. In motion injection, we first obtain " + }, + { + "bbox": [ + 54, + 198, + 555, + 254 + ], + "type": "inline_equation", + "content": "z_{t-1}" + }, + { + "bbox": [ + 54, + 198, + 555, + 254 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 54, + 198, + 555, + 254 + ], + "type": "inline_equation", + "content": "z_t" + }, + { + "bbox": [ + 54, + 198, + 555, + 254 + ], + "type": "text", + "content": ", then denoise " + }, + { + "bbox": [ + 54, + 198, + 555, + 254 + ], + "type": "inline_equation", + "content": "z_{t-1}" + }, + { + "bbox": [ + 54, + 198, + 555, + 254 + ], + "type": "text", + "content": " for further " + }, + { + "bbox": [ + 54, + 198, + 555, + 254 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 54, + 198, + 555, + 254 + ], + "type": "text", + "content": " steps using a T2V model to " + }, + { + "bbox": [ + 54, + 198, + 555, + 254 + ], + "type": "inline_equation", + "content": "\\tilde{z}_{t-1-M}'" + }, + { + "bbox": [ + 54, + 198, + 555, + 254 + ], + "type": "text", + "content": " and combine them. We visualize the motion injection effect using CoTracker [21]. In detail injection, we first transform " + }, + { + "bbox": [ + 54, + 198, + 555, + 254 + ], + "type": "inline_equation", + "content": "\\tilde{z}_{t-1}" + }, + { + "bbox": [ + 54, + 198, + 555, + 254 + ], + "type": "text", + "content": " from T2V noise distribution to T2I noise distribution, then denoise it for further " + }, + { + "bbox": [ + 54, + 198, + 555, + 254 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 54, + 198, + 555, + 254 + ], + "type": "text", + "content": " steps with an inflated T2I model. The result is reversed to T2V noise distribution and combined with " + }, + { + "bbox": [ + 54, + 198, + 555, + 254 + ], + "type": "inline_equation", + "content": "\\tilde{z}_{t-1}" + }, + { + "bbox": [ + 54, + 198, + 555, + 254 + ], + "type": "text", + "content": ". We visualize the detail injection effect using FFT analysis." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 372, + 55, + 529, + 185 + ], + "blocks": [ + { + "bbox": [ + 372, + 55, + 529, + 185 + ], + "lines": [ + { + "bbox": [ + 372, + 55, + 529, + 185 + ], + "spans": [ + { + "bbox": [ + 372, + 55, + 529, + 185 + ], + "type": "image", + "image_path": "b32e7d313bff07f481ff53c8c3d8f9463787037d7f103cde51ea2e41194272bc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 269, + 296, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 269, + 296, + 473 + ], + "spans": [ + { + "bbox": [ + 56, + 269, + 296, + 473 + ], + "type": "text", + "content": "the separation of foreground from the background is essential because instances described in the user's input manifest as foreground instances in Frame [25, 39]. Although we apply operations to enhance consistency by CFA, background distortions and abrupt changes can still occur [22]. Secondly, transforming images into Canny edge representations not only aligns with the requirements of ControlNet but also further mitigates potential inconsistencies in instance textures within Frame. (2) Textural Guidance. Use Motion Guidance only may cause textural inconsistency, hence we further perform Textural Guidance. We diffuse the first keyframe candidate " + }, + { + "bbox": [ + 56, + 269, + 296, + 473 + ], + "type": "inline_equation", + "content": "f_{1}" + }, + { + "bbox": [ + 56, + 269, + 296, + 473 + ], + "type": "text", + "content": " to form the initial noise of the T2V model. To help the guidance transmit from " + }, + { + "bbox": [ + 56, + 269, + 296, + 473 + ], + "type": "inline_equation", + "content": "f_{1}" + }, + { + "bbox": [ + 56, + 269, + 296, + 473 + ], + "type": "text", + "content": " to subsequent frames, we devise the following strategies: firstly, we employ an IP-Adapter [51] to utilize " + }, + { + "bbox": [ + 56, + 269, + 296, + 473 + ], + "type": "inline_equation", + "content": "f_{1}" + }, + { + "bbox": [ + 56, + 269, + 296, + 473 + ], + "type": "text", + "content": " as an image prompt. Moreover, we also apply CFA to the spatial layers of the T2V model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 487, + 190, + 500 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 190, + 500 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 190, + 500 + ], + "type": "text", + "content": "3.4. Video Refinement Stage" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 508, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 508, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 508, + 295, + 651 + ], + "type": "text", + "content": "We have observed that while coarse motion videos may align with user-desired inputs, unsmooth motion issues and suboptimal visual quality issues such as watermark and flickering exist. To attack these issues, we introduce a zero-shot method by more efficiently leveraging the motion priors inherent in existing T2V models and the detail-generation capacity of T2I models. As shown in Figure 3, we integrate a T2V model into the reverse process of video generation. We leverage T2V motion priors to inject motion into video latents, ensuring smoother and more natural motion. Subsequently, we also insert a T2I model, leveraging its detail generation capacity to enrich video detail richness." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "Given a low-quality coarse motion video as described in Section 3.3, we start by encoding it using the encoder " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": " of a vanilla T2V model (i.e.,AnimateDiff [13]) and adding Gaussian noise by timestep " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": " to obtain noisy video latents " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "z_{T}" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": ". For each denoising time step " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": ", we first denoise " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": " using" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 269, + 514, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 269, + 514, + 281 + ], + "spans": [ + { + "bbox": [ + 313, + 269, + 514, + 281 + ], + "type": "text", + "content": "the denoiser of the T2V model to obtain " + }, + { + "bbox": [ + 313, + 269, + 514, + 281 + ], + "type": "inline_equation", + "content": "z_{t - 1}" + }, + { + "bbox": [ + 313, + 269, + 514, + 281 + ], + "type": "text", + "content": " by:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 321, + 288, + 553, + 313 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 288, + 553, + 313 + ], + "spans": [ + { + "bbox": [ + 321, + 288, + 553, + 313 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(z _ {t - 1} \\mid z _ {t}\\right) = \\mathcal {N} \\left(z _ {t - 1}; z _ {t} - \\hat {\\epsilon} _ {\\theta} \\left(z _ {t}; c, t\\right), \\left(1 - \\alpha_ {t} ^ {2}\\right) \\mathbf {I}\\right) \\tag {1}", + "image_path": "5ca46e55df92cf63d4662aa71635429a18fbc2b3e9eb299f72ea7a8ecf2111f9.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 314, + 553, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 314, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 314, + 553, + 338 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 314, + 553, + 338 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 314, + 553, + 338 + ], + "type": "text", + "content": " is the text prompt embedding. Then we sequentially perform motion injection and detail injection on " + }, + { + "bbox": [ + 313, + 314, + 553, + 338 + ], + "type": "inline_equation", + "content": "z_{t - 1}" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 338, + 554, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 554, + 458 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 554, + 458 + ], + "type": "text", + "content": "Motion Injection. We employ a vanilla T2V model (i.e.,AnimateDiff [13]) for motion injection as shown in Figure 3. Specifically, we denoise " + }, + { + "bbox": [ + 313, + 338, + 554, + 458 + ], + "type": "inline_equation", + "content": "z_{t - 1}" + }, + { + "bbox": [ + 313, + 338, + 554, + 458 + ], + "type": "text", + "content": " using its denoiser for further " + }, + { + "bbox": [ + 313, + 338, + 554, + 458 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 338, + 554, + 458 + ], + "type": "text", + "content": " steps to acquire " + }, + { + "bbox": [ + 313, + 338, + 554, + 458 + ], + "type": "inline_equation", + "content": "\\tilde{z}_{t - 1 - M}^{\\prime}" + }, + { + "bbox": [ + 313, + 338, + 554, + 458 + ], + "type": "text", + "content": ", where each denoising step follows Equation (1). We observed a significant improvement in the temporal consistency of " + }, + { + "bbox": [ + 313, + 338, + 554, + 458 + ], + "type": "inline_equation", + "content": "\\tilde{z}_{t - 1 - M}^{\\prime}" + }, + { + "bbox": [ + 313, + 338, + 554, + 458 + ], + "type": "text", + "content": ", resulting in very smooth motion. However, this enhancement comes at the cost of losing many detailed features. Therefore, we combine " + }, + { + "bbox": [ + 313, + 338, + 554, + 458 + ], + "type": "inline_equation", + "content": "\\tilde{z}_{t - 1 - M}^{\\prime}" + }, + { + "bbox": [ + 313, + 338, + 554, + 458 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 338, + 554, + 458 + ], + "type": "inline_equation", + "content": "z_{t - 1}" + }, + { + "bbox": [ + 313, + 338, + 554, + 458 + ], + "type": "text", + "content": " to enhance motion smoothness while preserving details by:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 352, + 465, + 553, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 465, + 553, + 479 + ], + "spans": [ + { + "bbox": [ + 352, + 465, + 553, + 479 + ], + "type": "interline_equation", + "content": "\\tilde {z} _ {t - 1} = \\lambda_ {1} * z _ {t - 1} + (1 - \\lambda_ {1}) * \\tilde {z} _ {t - 1 - M} ^ {\\prime} \\tag {2}", + "image_path": "424153a9ae170a675cdace10dc1d160f26da59149d254c09562e273fb27c5ce8.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 487, + 528, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 528, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 528, + 498 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 487, + 528, + 498 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 313, + 487, + 528, + 498 + ], + "type": "text", + "content": " is a coefficient controlling the fusion scale." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 498, + 554, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 554, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 554, + 559 + ], + "type": "text", + "content": "Detail Injection. We then employ an inflated T2I model for detail injection as shown in Figure 3. Due to the different noise distributions in the T2I and T2V models, a noise transformation is required. To convert video latents into image latents, we directly estimate clean latents [48] " + }, + { + "bbox": [ + 313, + 498, + 554, + 559 + ], + "type": "inline_equation", + "content": "\\hat{z}_0" + }, + { + "bbox": [ + 313, + 498, + 554, + 559 + ], + "type": "text", + "content": " by:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 370, + 566, + 553, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 566, + 553, + 594 + ], + "spans": [ + { + "bbox": [ + 370, + 566, + 553, + 594 + ], + "type": "interline_equation", + "content": "\\hat {z} _ {0} = \\frac {\\tilde {z} _ {t} - \\sqrt {1 - \\alpha_ {t}} \\epsilon_ {\\theta} (\\tilde {z} _ {t} , c , t)}{\\sqrt {\\alpha_ {t}}} \\tag {3}", + "image_path": "9d2a19687752d835f78619d67cfc2f7b800c59da0ef77753fb4b31b79e141a36.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 600, + 555, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 555, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 555, + 636 + ], + "type": "text", + "content": "and we then conduct DDIM inversion [30] to derive the corresponding noisy image latents " + }, + { + "bbox": [ + 313, + 600, + 555, + 636 + ], + "type": "inline_equation", + "content": "\\tilde{z}_{t - 1}^{I}" + }, + { + "bbox": [ + 313, + 600, + 555, + 636 + ], + "type": "text", + "content": ", the DDIM inversion process is defined by:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 353, + 643, + 553, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 643, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 353, + 643, + 553, + 658 + ], + "type": "interline_equation", + "content": "\\tilde {z} _ {t} ^ {I} = \\sqrt {\\alpha_ {t}} \\hat {z} _ {0} + \\sqrt {1 - \\alpha_ {t}} \\cdot \\epsilon_ {\\theta} (\\tilde {z} _ {t - 1} ^ {I}, c, t) \\tag {4}", + "image_path": "7bb790e765104ee707dbb6db3a5ba69a3faa18df576e9355771b5cc5e4d17a9c.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "text", + "content": "Subsequently, we denoise the noisy image latents with the inflated T2I model for further " + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "text", + "content": " steps to obtain " + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\tilde{z}_{t-1-D}^{I}" + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "text", + "content": ". The denoising process resembles Equation (1). As the T2I model excels in detail generation compared to the T2V" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7269" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 88, + 53, + 530, + 379 + ], + "blocks": [ + { + "bbox": [ + 88, + 53, + 530, + 379 + ], + "lines": [ + { + "bbox": [ + 88, + 53, + 530, + 379 + ], + "spans": [ + { + "bbox": [ + 88, + 53, + 530, + 379 + ], + "type": "image", + "image_path": "af8934e9b64973ed90739bcec9e2fb12bab3a5e4c2413e8971a8c7ef4cdee6dc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 386, + 555, + 409 + ], + "lines": [ + { + "bbox": [ + 54, + 386, + 555, + 409 + ], + "spans": [ + { + "bbox": [ + 54, + 386, + 555, + 409 + ], + "type": "text", + "content": "Figure 4. Qualitative Results of IM-Zero compared with other methods. Videos generated by IM-Zero outperform others in control accuracy, motion effects, and video quality." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 424, + 296, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 424, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 424, + 296, + 510 + ], + "type": "text", + "content": "model, we revert " + }, + { + "bbox": [ + 55, + 424, + 296, + 510 + ], + "type": "inline_equation", + "content": "\\tilde{z}_{t - 1 - D}^{I}" + }, + { + "bbox": [ + 55, + 424, + 296, + 510 + ], + "type": "text", + "content": " back to the noise distribution of the T2V model, combining the transformed latents with " + }, + { + "bbox": [ + 55, + 424, + 296, + 510 + ], + "type": "inline_equation", + "content": "\\tilde{z}_{t - 1}" + }, + { + "bbox": [ + 55, + 424, + 296, + 510 + ], + "type": "text", + "content": " obtained from Equation (2) to integrate details. Specifically, in the reversed noise transformation process, we predict clean noise from " + }, + { + "bbox": [ + 55, + 424, + 296, + 510 + ], + "type": "inline_equation", + "content": "\\tilde{z}_{t - 1}" + }, + { + "bbox": [ + 55, + 424, + 296, + 510 + ], + "type": "text", + "content": " using Equation (3) and utilize the DDPM forward process to obtain the corresponding noisy video latents " + }, + { + "bbox": [ + 55, + 424, + 296, + 510 + ], + "type": "inline_equation", + "content": "\\tilde{z}_{t - 1 - D}^{V}" + }, + { + "bbox": [ + 55, + 424, + 296, + 510 + ], + "type": "text", + "content": ". Then we fuse them by:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 94, + 519, + 295, + 534 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 519, + 295, + 534 + ], + "spans": [ + { + "bbox": [ + 94, + 519, + 295, + 534 + ], + "type": "interline_equation", + "content": "z _ {t - 1} = \\lambda_ {2} * \\tilde {z} _ {t - 1} + (1 - \\lambda_ {2}) * \\tilde {z} _ {t - 1 - D} ^ {V} \\tag {5}", + "image_path": "c3e2d103d10b896cda524b134d973b932b513cf5a226911e6b03350cc6f06379.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 543, + 295, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 543, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 543, + 295, + 568 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 543, + 295, + 568 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 55, + 543, + 295, + 568 + ], + "type": "text", + "content": " controls the fusion scale and " + }, + { + "bbox": [ + 55, + 543, + 295, + 568 + ], + "type": "inline_equation", + "content": "z_{t - 1}" + }, + { + "bbox": [ + 55, + 543, + 295, + 568 + ], + "type": "text", + "content": " is used for the next denoising timestep." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 575, + 248, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 575, + 248, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 575, + 248, + 588 + ], + "type": "text", + "content": "3.5. Adapt IM-Zero for Motion Transfer" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "text", + "content": "For a more customized motion pattern provided by reference videos, IM-Zero can also be easily adapted for motion transfer tasks. Given a reference video and the target text prompt, we extract control signals (e.g., depth maps) from each frame of the reference video. Subsequently, we use these dense control signals and the target text prompt to guide the T2V model in generating a coarse motion video via ControlNet. Then, the coarse motion video undergoes video refinement stage as detailed in Section 3.4 to enhance motion consistency and video frame quality." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 423, + 391, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 391, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 391, + 437 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 444, + 427, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 444, + 427, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 444, + 427, + 456 + ], + "type": "text", + "content": "4.1. Qualitative Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 312, + 461, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 461, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 312, + 461, + 555, + 628 + ], + "type": "text", + "content": "Regarding the comparative methods, we primarily selected Peekaboo [20], TrailBlazer [29], and MotionBooth [42], which generate videos by specifying boxes and trajectories. Furthermore, to further illustrate the effectiveness of our approach, we also included T2V-Zero [22], which offers rough guidance by specifying translation vectors. To ensure a fair comparison, following the experimental settings of [29], all methods utilize the same layout input, trajectory input, and text prompt input. Additional results of controllable video generation are available in the supplementary material. We also compared with FreeTraj [32], but as it additionally requires inputting boxes with fixed sizes across frames to make noise flow, we place the comparison in the supplementary material." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "As shown in Figure 4, our method significantly outperforms other approaches in terms of alignment with the input, motion effects, and video quality. In terms of alignment with the input, T2V-Zero exhibits almost no resemblance to the input, while the two methods based on attention map manipulation, Peekaboo, and TrailBlazer, sometimes roughly match the input location and trajectory but" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7270" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 87, + 555, + 188 + ], + "blocks": [ + { + "bbox": [ + 55, + 60, + 555, + 84 + ], + "lines": [ + { + "bbox": [ + 55, + 60, + 555, + 84 + ], + "spans": [ + { + "bbox": [ + 55, + 60, + 555, + 84 + ], + "type": "text", + "content": "Table 1. Quantitative results of IM-Zero compared with other methods on DAVIS-17 [31] dataset and GOT10k [17] dataset. IM-Zero achieves SOTA in all metrics, including metrics measuring video quality, frame consistency, and control accuracy." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 87, + 555, + 188 + ], + "lines": [ + { + "bbox": [ + 56, + 87, + 555, + 188 + ], + "spans": [ + { + "bbox": [ + 56, + 87, + 555, + 188 + ], + "type": "table", + "html": "
MethodDAVIS-17GOT10k
FVD (↓)KID (↓)CLIPSim (↑)mIoU (↑)CD (↓)FVD (↓)KID (↓)CLIPSim (↑)mIoU (↑)CD (↓)
T2V-Zero [22]5004.0030.3795.120.190.344335.6924.7295.410.120.32
Peekaboo [20]4264.7026.4897.330.190.464960.4919.0598.680.110.34
TrailBlazer [29]4593.3726.0498.100.210.294695.1816.1897.800.130.40
MotionBooth [42]4192.1926.8596.640.220.364311.7116.0597.470.140.42
IM-Zero (Ours)4070.0125.7598.910.270.264187.2715.9299.040.210.27
", + "image_path": "4a2e730bd470dfde2cc192aa8bbbceb492b5208533dfac3c82b8451a19c26fe1.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 103, + 208, + 504, + 293 + ], + "blocks": [ + { + "bbox": [ + 175, + 193, + 435, + 205 + ], + "lines": [ + { + "bbox": [ + 175, + 193, + 435, + 205 + ], + "spans": [ + { + "bbox": [ + 175, + 193, + 435, + 205 + ], + "type": "text", + "content": "Table 2. Ablation on different components in Motion Generation Stage." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 103, + 208, + 504, + 293 + ], + "lines": [ + { + "bbox": [ + 103, + 208, + 504, + 293 + ], + "spans": [ + { + "bbox": [ + 103, + 208, + 504, + 293 + ], + "type": "table", + "html": "
MethodFVD (↓)KID (↓)CLIPSim (↑)mIoU (↑)CD (↓)
w/o 1st CFA4320.0726.8498.840.260.29
w/o 2nd CFA4528.0928.0898.780.210.31
w/o init noise4786.1726.6998.820.260.31
w/o IP-Adapter4549.9427.7598.870.260.28
IM-Zero (Ours)4070.0125.7598.910.270.26
", + "image_path": "65216db07423447a9641d987edd2958730d24491713be4fda04faeeb044d068e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 307, + 294, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 307, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 54, + 307, + 294, + 392 + ], + "type": "text", + "content": "still lack precision. Regarding motion effects, Peekaboo and TrailBlazer generate videos with unnatural motion, resulting in strange discontinuities between frames. Concerning video quality, Peekaboo and TrailBlazer exhibit low image quality, lack detail, and introduce artifacts. We also provide a video version of Figure 4 in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 396, + 174, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 174, + 410 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 174, + 410 + ], + "type": "text", + "content": "4.2. Quantitative Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 415, + 295, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 295, + 642 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 295, + 642 + ], + "type": "text", + "content": "Experiment Settings. For datasets, we utilize the training set and validation set of the DAVIS-17 dataset [31] and the test set of the GOT10k dataset [17]. We use BLIP-2 [24] to generate text prompts and GroundingDINO [28] to extract bounding boxes from the first and last frame as inputs of all the methods. After eliminating videos where no bounding boxes are extracted, we randomly select 50 videos and 70 videos from the rest videos of DAVIS-17 [31] and GOT10k [17]. For our method, in the first stage we use InstanceDiffusion [39] as grounded T2I model and SparseCtrl [12] as ControlNet, in the second stage we use Stable Diffusion V1.5 [34] as T2I model, on both stage we useAnimateDiff [13] as T2V model. For the evaluation metrics, we employ mean Intersection over Union (mIoU) and Centroid Distance (CD) to assess control alignment, Fréchet Video Distance (FVD) [35] and Kernel Inception Distance (KID) [2] to evaluate video quality, and CLIP similarities (CLIPSim) [33] to assess frame-to-frame consistency. Detailed settings are available in the supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": "Metrics Evaluation Results. We conducted evaluations on the aforementioned metrics on the DAVIS-17 dataset [31] and the GOT10k dataset [17]. As shown in Table 1, our method achieved state-of-the-art (SOTA) performance compared to several other methods. Firstly, in terms of the alignment between the generated videos and user inputs," + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 314, + 326, + 556, + 396 + ], + "blocks": [ + { + "bbox": [ + 320, + 306, + 547, + 318 + ], + "lines": [ + { + "bbox": [ + 320, + 306, + 547, + 318 + ], + "spans": [ + { + "bbox": [ + 320, + 306, + 547, + 318 + ], + "type": "text", + "content": "Table 3. Motion transfer results compared with other methods." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 314, + 326, + 556, + 396 + ], + "lines": [ + { + "bbox": [ + 314, + 326, + 556, + 396 + ], + "spans": [ + { + "bbox": [ + 314, + 326, + 556, + 396 + ], + "type": "table", + "html": "
MethodMotion Fidelity (↑)Imaging Quality (↑)
MotionClone [26]74.50.72
MOFT [23]51.20.71
MotionDirector [59]75.50.68
IM-Zero (Ours)63.70.73
", + "image_path": "2e8334a14956e47173a7c32e0f02d9043905eb46b07527dfc4d78fa690a29fbc.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 324, + 426, + 545, + 473 + ], + "blocks": [ + { + "bbox": [ + 340, + 406, + 527, + 418 + ], + "lines": [ + { + "bbox": [ + 340, + 406, + 527, + 418 + ], + "spans": [ + { + "bbox": [ + 340, + 406, + 527, + 418 + ], + "type": "text", + "content": "Table 4. T2V results compared with UNet baseline." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 324, + 426, + 545, + 473 + ], + "lines": [ + { + "bbox": [ + 324, + 426, + 545, + 473 + ], + "spans": [ + { + "bbox": [ + 324, + 426, + 545, + 473 + ], + "type": "table", + "html": "
MethodImaging Quality (↑)CLIPSim (↑)
AnimateDiff [13]0.6799.05
IM-Zero (Ours)0.7499.73
", + "image_path": "9b1b257a1d9acfdfccf5b1bd37bf741db8044da6f804bc434fa173afbd88049a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 495, + 555, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 555, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 555, + 651 + ], + "type": "text", + "content": "our method demonstrates a significant lead. Compared to the previous best methods, our approach achieved an improvement of " + }, + { + "bbox": [ + 313, + 495, + 555, + 651 + ], + "type": "inline_equation", + "content": "28.6\\%" + }, + { + "bbox": [ + 313, + 495, + 555, + 651 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 495, + 555, + 651 + ], + "type": "inline_equation", + "content": "61.5\\%" + }, + { + "bbox": [ + 313, + 495, + 555, + 651 + ], + "type": "text", + "content": " in mIoU, and reductions of " + }, + { + "bbox": [ + 313, + 495, + 555, + 651 + ], + "type": "inline_equation", + "content": "10.4\\%" + }, + { + "bbox": [ + 313, + 495, + 555, + 651 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 495, + 555, + 651 + ], + "type": "inline_equation", + "content": "15.6\\%" + }, + { + "bbox": [ + 313, + 495, + 555, + 651 + ], + "type": "text", + "content": " in CD on the two datasets, indicating that our method offers more accurate location control and motion trajectory control compared to previous methods based on attention map manipulation. Secondly, regarding frame-to-frame consistency in the videos, thanks to our enhanced consistency techniques and motion injection methods, our approach also reaches the SOTA performance. Moreover, in terms of video quality, our method significantly outperforms others, with FVD reduced to 4070.01 and 4187.27, and KID reduced to 25.75 and 15.92, respectively." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 659, + 465, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 659, + 465, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 659, + 465, + 673 + ], + "type": "text", + "content": "4.3. Versatile Capacity Analysis" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "Our method possesses versatile capacity as shown in Figure 5. First, our method can control the subparts of instances, achieving effects such as head turning and hand" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7271" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 100, + 50, + 530, + 266 + ], + "blocks": [ + { + "bbox": [ + 84, + 73, + 96, + 86 + ], + "lines": [ + { + "bbox": [ + 84, + 73, + 96, + 86 + ], + "spans": [ + { + "bbox": [ + 84, + 73, + 96, + 86 + ], + "type": "text", + "content": "#" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 100, + 50, + 530, + 266 + ], + "lines": [ + { + "bbox": [ + 100, + 50, + 530, + 266 + ], + "spans": [ + { + "bbox": [ + 100, + 50, + 530, + 266 + ], + "type": "image", + "image_path": "54d5cbbc0ed4713fd4710482d519fe249ff17f62fdd1a5cd9ae15d12c13078d3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 274, + 555, + 298 + ], + "lines": [ + { + "bbox": [ + 55, + 274, + 555, + 298 + ], + "spans": [ + { + "bbox": [ + 55, + 274, + 555, + 298 + ], + "type": "text", + "content": "Figure 5. Besides box-level overall instance motion control, IM-Zero can: (a) Perform instance subpart control. (b) Specify instance shape by mask. (c) Perform motion transfer from the reference video(left) to the target video(right). (d) Perform high-quality T2V generation." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 318, + 294, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 318, + 294, + 400 + ], + "spans": [ + { + "bbox": [ + 54, + 318, + 294, + 400 + ], + "type": "text", + "content": "movement. Second, besides bounding boxes, we also support specifying instance shapes more precisely using masks. Third, for cases where users provide a reference video, our method can perform motion transfer. Finally, users can directly provide text inputs for high-quality T2V generation. For motion transfer and T2V generation, we compare IM-Zero with other methods as follows." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 402, + 295, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 402, + 295, + 497 + ], + "spans": [ + { + "bbox": [ + 54, + 402, + 295, + 497 + ], + "type": "text", + "content": "Motion Transfer. We compare with MOFT [23], MotionClone [26], and MotionDirector [59] using the open-sourced data from MotionClone [26]. For quantitative results, we use Motion Fidelity [50] and Imaging Quality [19] as metrics following MOFT [23]. Table 3 shows that our method is competitive in Motion Fidelity [50] and is the best in Imaging Quality [19]. Qualitative results and other details are provided in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 498, + 295, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 498, + 295, + 593 + ], + "spans": [ + { + "bbox": [ + 54, + 498, + 295, + 593 + ], + "type": "text", + "content": "T2V Generation. As we use UNet-based models, we compare T2V with UNet baselineAnimateDiff [13]. For quantitative results, we use Imaging Quality [19] and CLIP-Sim [33] as metrics. As there are no ground-truth videos, we generate videos using 50 prompts randomly generated by GPT. Table 4 shows that our method is better in Imaging Quality [19] and CLIPSim [33]. Refer to the supplementary material for qualitative results and other details." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 594, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 594, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 54, + 594, + 295, + 653 + ], + "type": "text", + "content": "Other Capacities. The proposed method also supports multi-instance motion control, and the decomposition of our pipeline allows autoregressive usage of the second stage to enhance temporal consistency, as elaborated in the supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 660, + 157, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 660, + 157, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 660, + 157, + 672 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 295, + 689 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 689 + ], + "type": "text", + "content": "Ablation on different components in Motion Generation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 294, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 294, + 714 + ], + "type": "text", + "content": "Stage. We ablate on different components in Section 3.3. As depicted in Table 2, these methods improve the quality" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 318, + 555, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 318, + 555, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 318, + 555, + 342 + ], + "type": "text", + "content": "of the final videos and also enhance the alignment with user inputs. Refer to the supplementary material for details." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 345, + 555, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 345, + 555, + 381 + ], + "spans": [ + { + "bbox": [ + 313, + 345, + 555, + 381 + ], + "type": "text", + "content": "Ablation on different components in Video Refinement Stage. We also conduct ablation studies on Video Refinement Stage. Refer to the supplementary material for details." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 400, + 468, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 400, + 468, + 414 + ], + "spans": [ + { + "bbox": [ + 313, + 400, + 468, + 414 + ], + "type": "text", + "content": "5. Conclusion and Limitation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 423, + 555, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 555, + 567 + ], + "type": "text", + "content": "We introduce IM-Zero, a novel training-free framework for instance-level motion controllable video generation using diffusion models. We leverage grounded text-to-image models to generate keyframe candidates and extract motion guidance and textural guidance from them to generate videos that align with user inputs. To further enhance video motion and quality, we also introduce a zero-shot method for motion injection and detail injection. Additionally, our method can control the subpart motion of instances, use masks to specify instance shapes, perform motion transfer from reference videos, and directly support high-quality text-to-video generation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": "Limitation. The zero-shot features impose certain limitations on our method. Firstly, as zero-shot methods depend on pre-trained model weights, the videos generated by our method are inherently constrained by the pre-trained models employed. For example, the alignment with layout input is influenced by the grounded T2I model and video ControlNet, the video quality is affected by the T2V model, which can be improved by using more advanced pre-trained models. Secondly, the temporal consistency of videos generated by zero-shot methods is inherently weaker compared to training-based methods, which can be improved by incorporating some tuning components to our method." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7272" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 175 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 175 + ], + "type": "text", + "content": "This work was supported in part by the National Natural Science Foundation of China under Grant 62125109, Grant 62431017, Grant U24A20251, Grant 62320106003, Grant 62371288, Grant 62401357, Grant 62401366, Grant 62301299, Grant 62120106007, and in part by the Program of Shanghai Science and Technology Innovation Project under Grant 24BC3200800." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 184, + 115, + 197 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 115, + 197 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 115, + 197 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 204, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 204, + 295, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 204, + 295, + 248 + ], + "spans": [ + { + "bbox": [ + 61, + 204, + 295, + 248 + ], + "type": "text", + "content": "[1] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. MultiDiffusion: Fusing diffusion paths for controlled image generation. In Proceedings of the 40th International Conference on Machine Learning, pages 1737-1752, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 250, + 296, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 250, + 296, + 292 + ], + "spans": [ + { + "bbox": [ + 61, + 250, + 296, + 292 + ], + "type": "text", + "content": "[2] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying MMD GANs. In *The Sixth International Conference on Learning Representations*, 2018. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 293, + 295, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 293, + 295, + 358 + ], + "spans": [ + { + "bbox": [ + 62, + 293, + 295, + 358 + ], + "type": "text", + "content": "[3] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 360, + 295, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 360, + 295, + 414 + ], + "spans": [ + { + "bbox": [ + 62, + 360, + 295, + 414 + ], + "type": "text", + "content": "[4] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. OpenAI Blog, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 415, + 295, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 415, + 295, + 470 + ], + "spans": [ + { + "bbox": [ + 62, + 415, + 295, + 470 + ], + "type": "text", + "content": "[5] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. MasaCtrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22560–22570, 2023. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 471, + 295, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 471, + 295, + 535 + ], + "spans": [ + { + "bbox": [ + 62, + 471, + 295, + 535 + ], + "type": "text", + "content": "[6] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, Chao Weng, and Ying Shan. VideoCrafter1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512, 2023. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 536, + 295, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 536, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 62, + 536, + 295, + 601 + ], + "type": "text", + "content": "[7] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. VideoCrafter2: Overcoming data limitations for high-quality video diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7310-7320, 2024. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 602, + 296, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 602, + 296, + 647 + ], + "spans": [ + { + "bbox": [ + 62, + 602, + 296, + 647 + ], + "type": "text", + "content": "[8] Minghao Chen, Iro Laina, and Andrea Vedaldi. Training-free layout control with cross-attention guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5343–5353, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 647, + 295, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 647, + 295, + 691 + ], + "spans": [ + { + "bbox": [ + 62, + 647, + 295, + 691 + ], + "type": "text", + "content": "[9] Weifeng Chen, Yatai Ji, Jie Wu, Hefeng Wu, Pan Xie, Jiashi Li, Xin Xia, Xuefeng Xiao, and Liang Lin. Control-A-Video: Controllable text-to-video generation with diffusion models. arXiv preprint arXiv:2305.13840, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "type": "text", + "content": "[10] Yabo Chen, Jiemin Fang, Yuyang Huang, Taoran Yi, Xiaopeng Zhang, Lingxi Xie, Xinggang Wang, Wenrui Dai," + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 117 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 117 + ], + "type": "text", + "content": "Hongkai Xiong, and Qi Tian. Cascade-Zero123: One image to highly consistent 3D with self-prompted nearby views. In Proceedings of the 18th European Conference on Computer Vision, pages 311-330, 2024. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 554, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 554, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 554, + 161 + ], + "type": "text", + "content": "[11] Yufan Deng, Ruida Wang, Yuhao Zhang, Yu-Wing Tai, and Chi-Keung Tang. DragVideo: Interactive drag-style video editing. In Proceedings of the 18th European Conference on Computer Vision, pages 183–199, 2024. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 162, + 554, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 162, + 554, + 215 + ], + "spans": [ + { + "bbox": [ + 317, + 162, + 554, + 215 + ], + "type": "text", + "content": "[12] Yuwei Guo, Ceyuan Yang, Anyi Rao, Maneesh Agrawala, Dahua Lin, and Bo Dai. SparseCtrl: Adding sparse controls to text-to-video diffusion models. In Proceedings of the 18th European Conference on Computer Vision, pages 330-348, 2024. 4, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 217, + 553, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 281 + ], + "type": "text", + "content": "[13] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. AnimateDiff:Animate your personalized text-to-image diffusion models without specific tuning. In The Twelfth International Conference on Learning Representations, 2024. 2, 3, 5, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 283, + 553, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 283, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 316, + 283, + 553, + 338 + ], + "type": "text", + "content": "[14] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P. Kingma, Ben Poole, Mohammad Norouzi, David J. Fleet, and Tim Salimans. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 338, + 553, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 553, + 382 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 553, + 382 + ], + "type": "text", + "content": "[15] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. In Advances in Neural Information Processing Systems 35, pages 8633-8646, 2022. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 383, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 553, + 426 + ], + "type": "text", + "content": "[16] Zhihao Hu and Dong Xu. VideoControlNet: A motion-guided video-to-video translation framework by using diffusion model with ControlNet. arXiv preprint arXiv:2307.14073, 2023. 2, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 427, + 553, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 427, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 316, + 427, + 553, + 470 + ], + "type": "text", + "content": "[17] Lianghua Huang, Xin Zhao, and Kaiqi Huang. GOT-10k: A large high-diversity benchmark for generic object tracking in the wild. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(5):1562-1577, 2019. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 472, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 472, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 317, + 472, + 553, + 525 + ], + "type": "text", + "content": "[18] Yuyang Huang, Yabo Chen, Yuchen Liu, Xiaopeng Zhang, Wenrui Dai, Hongkai Xiong, and Qi Tian. DomainFusion: Generalizing to unseen domains with latent diffusion models. In Proceedings of the 18th European Conference on Computer Vision, pages 480-498, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 526, + 553, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 526, + 553, + 602 + ], + "spans": [ + { + "bbox": [ + 317, + 526, + 553, + 602 + ], + "type": "text", + "content": "[19] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, Yaohui Wang, Xinyuan Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. VBench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 603, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 603, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 603, + 553, + 658 + ], + "type": "text", + "content": "[20] Yash Jain, Anshul Nasery, Vibhav Vineet, and Harkirat Behl. PEEKABOO: Interactive video generation via masked-diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8079-8088, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "text", + "content": "[21] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. CoTracker: It is better to track together. In Proceedings of the 18th European Conference on Computer Vision, pages 18-35, 2024. 5" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7273" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 139 + ], + "type": "text", + "content": "[22] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2Video-Zero: Text-to-image diffusion models are zero-shot video generators. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15954-15964, 2023. 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 195 + ], + "type": "text", + "content": "[23] Karlo Koledić, Igor Cvišić, Ivan Marković, and Ivan Petrović. MOFT: Monocular odometry based on deep depth and careful feature selection and tracking. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 6175-6181. IEEE, 2023. 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "type": "text", + "content": "[24] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In Proceedings of the 40th International Conference on Machine Learning, pages 19730–19742, 2023. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 253, + 294, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 294, + 317 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 294, + 317 + ], + "type": "text", + "content": "[25] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. GLIGEN: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22511-22521, 2023. 2, 3, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 319, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 319, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 319, + 294, + 374 + ], + "type": "text", + "content": "[26] Pengyang Ling, Jiazi Bu, Pan Zhang, Xiaoyi Dong, Yuhang Zang, Tong Wu, Huaian Chen, Jiaqi Wang, and Yi Jin. MotionClone: Training-free motion cloning for controllable video generation. In The Thirteenth International Conference on Learning Representations, 2025. 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 376, + 294, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 294, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 294, + 431 + ], + "type": "text", + "content": "[27] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9298–9309, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 433, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 433, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 433, + 294, + 498 + ], + "type": "text", + "content": "[28] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, and Lei Zhang. Grounding DINO: Marrying DINO with grounded pre-training for open-set object detection. In Proceedings of the 18th European Conference on Computer Vision, pages 38-55, 2024. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 499, + 294, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 499, + 294, + 542 + ], + "spans": [ + { + "bbox": [ + 56, + 499, + 294, + 542 + ], + "type": "text", + "content": "[29] Wan-Duo Kurt Ma, John P Lewis, and W Bastiaan Kleijn. TrailBlazer: Trajectory control for diffusion-based video generation. In SIGGRAPH Asia 2024, number 97, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 544, + 294, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 544, + 294, + 599 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 294, + 599 + ], + "type": "text", + "content": "[30] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 601, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 294, + 644 + ], + "type": "text", + "content": "[31] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 DAVIS challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "text", + "content": "[32] Haonan Qiu, Zhaoxi Chen, Zhouxia Wang, Yingqing He, Menghan Xia, and Ziwei Liu. FreeTraj: Tuning-free trajectory control in video diffusion models. arXiv preprint arXiv:2406.16863, 2024. 2, 3, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry," + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 128 + ], + "type": "text", + "content": "Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763, 2021. 7, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 129, + 555, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 555, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 555, + 183 + ], + "type": "text", + "content": "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 184, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 184, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 184, + 553, + 228 + ], + "type": "text", + "content": "[35] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 228, + 553, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 281 + ], + "type": "text", + "content": "[36] Cong Wang, Jiaxi Gu, Panwen Hu, Haoyu Zhao, Yuanfan Guo, Jianhua Han, Hang Xu, and Xiaodan Liang. EasyControl: Transfer ControlNet to video diffusion for controllable generation and interpolation. arXiv preprint arXiv:2408.13005, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 283, + 553, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 283, + 553, + 325 + ], + "spans": [ + { + "bbox": [ + 316, + 283, + 553, + 325 + ], + "type": "text", + "content": "[37] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. ModelScope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 327, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 381 + ], + "type": "text", + "content": "[38] Weimin Wang, Jiawei Liu, Zhijie Lin, Jiangqiao Yan, Shuo Chen, Chetwin Low, Tuyen Hoang, Jie Wu, Jun Hao Liew, Hanshu Yan, Daquan Zhou, and Jiashi Feng. MagicVideoV2: Multi-stage high-aesthetic video generation. arXiv preprint arXiv:2401.04468, 2024. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 383, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 553, + 437 + ], + "type": "text", + "content": "[39] Xudong Wang, Trevor Darrell, Sai Saketh Rambhatla, Rohit Girdhar, and Ishan Misra. InstanceDiffusion: Instance-level control for image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6232-6242, 2024. 2, 3, 4, 5, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 438, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 438, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 438, + 553, + 514 + ], + "type": "text", + "content": "[40] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, Yuwei Guo, Tianxing Wu, Chenyang Si, Yuming Jiang, Cunjian Chen, Chen Change Loy, Bo Dai, Dahua Lin, Yu Qiao, and Ziwei Liu. LaVie: High-quality video generation with cascaded latent diffusion models. International Journal of Computer Vision, 2024. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 515, + 553, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 515, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 316, + 515, + 553, + 569 + ], + "type": "text", + "content": "[41] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. MotionCtrl: A unified and flexible motion controller for video generation. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 571, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 571, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 571, + 553, + 624 + ], + "type": "text", + "content": "[42] Jianzong Wu, Xiangtai Li, Yanhong Zeng, Jiangning Zhang, Qianyu Zhou, Yining Li, Yunhai Tong, and Kai Chen. MotionBooth: Motion-aware customized text-to-video generation. In Advances in Neural Information Processing Systems 37, pages 34322-34348, 2024. 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 625, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 669 + ], + "type": "text", + "content": "[43] Mingrui Wu, Oucheng Huang, Jiayi Ji, Jiale Li, Xinyue Cai, Huafeng Kuang, Jianzhuang Liu, Xiaoshuai Sun, and Rongrong Ji. TraDiffusion: Trajectory-based training-free image generation. arXiv preprint arXiv:2408.09739, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[44] Weijia Wu, Zhuang Li, Yuchao Gu, Rui Zhao, Yefei He, David Junhao Zhang, Mike Zheng Shou, Yan Li, Tingting Gao, and Di Zhang. DragAnything: Motion control for anything using entity representation. In Proceedings of the 18th" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7274" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "text", + "content": "European Conference on Computer Vision, pages 331-348, 2024. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 162 + ], + "type": "text", + "content": "[45] Enze Xie, Junsong Chen, Junyu Chen, Han Cai, Haotian Tang, Yujun Lin, Zhekai Zhang, Muyang Li, Ligeng Zhu, Yao Lu, and Song Han. SANA: Efficient high-resolution image synthesis with linear diffusion transformers. In The Thirteenth International Conference on Learning Representations, 2025. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 295, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 295, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 295, + 218 + ], + "type": "text", + "content": "[46] Jinheng Xie, Yuexiang Li, Yawen Huang, Haozhe Liu, Wentian Zhang, Yefeng Zheng, and Mike Zheng Shou. BoxDiff: Text-to-image synthesis with training-free box-constrained diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7452-7461, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 262 + ], + "type": "text", + "content": "[47] Yichao Yan, Zanwei Zhou, Zi Wang, Jingnan Gao, and Xiaokang Yang. DialogueNeRF: Towards realistic avatar face-to-face conversation video generation. Visual Intelligence, 2 (1):24, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 295, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 295, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 295, + 308 + ], + "type": "text", + "content": "[48] Serin Yang, Hyunmin Hwang, and Jong Chul Ye. Zero-shot contrastive loss for text-guided diffusion image style transfer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22873-22882, 2023. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 310, + 295, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 310, + 295, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 310, + 295, + 374 + ], + "type": "text", + "content": "[49] Zhengyuan Yang, Jianfeng Wang, Zhe Gan, Linjie Li, Kevin Lin, Chenfei Wu, Nan Duan, Zicheng Liu, Ce Liu, Michael Zeng, and Lijuan Wang. ReCo: Region-controlled text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14246-14255, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 376, + 295, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 295, + 432 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 295, + 432 + ], + "type": "text", + "content": "[50] Danah Yatim, Rafail Fridman, Omer Bar-Tal, Yoni Kasten, and Tali Dekel. Space-time diffusion features for zero-shot text-driven motion transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8466-8476, 2024. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 434, + 295, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 295, + 475 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 295, + 475 + ], + "type": "text", + "content": "[51] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. IP-Adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 478, + 295, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 295, + 521 + ], + "type": "text", + "content": "[52] Shengming Yin, Chenfei Wu, Jian Liang, Jie Shi, Houqiang Li, Gong Ming, and Nan Duan. DragNUWA: Fine-grained control in video generation by integrating text, image, and trajectory. arXiv preprint arXiv:2308.08089, 2023. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 523, + 295, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 295, + 578 + ], + "type": "text", + "content": "[53] Yan Zeng, Guoqiang Wei, Jiani Zheng, Jiaxin Zou, Yang Wei, Yuchen Zhang, and Hang Li. Make pixels dance: High-dynamic video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8850-8860, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 579, + 295, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 295, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 295, + 644 + ], + "type": "text", + "content": "[54] Junyi Zhang, Charles Herrmann, Junhwa Hur, Luisa Polania Cabrera, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. A tale of two features: Stable diffusion complements DINO for zero-shot semantic correspondence. In Advances in Neural Information Processing Systems 36, pages 45533-45547, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 647, + 295, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 295, + 690 + ], + "type": "text", + "content": "[55] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 692, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 295, + 713 + ], + "type": "text", + "content": "[56] Yabo Zhang, Yuxiang Wei, Dongsheng Jiang, Xiaopeng Zhang, Wangmeng Zuo, and Qi Tian. ControlVideo:" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 575 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "Training-free controllable text-to-video generation. In The Twelfth International Conference on Learning Representations, 2024. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 161 + ], + "type": "text", + "content": "[57] Yabo Zhang, Yuxiang Wei, Xianhui Lin, Zheng Hui, Peiran Ren, Xuansong Xie, Xiangyang Ji, and Wangmeng Zuo. VideoElevator: elevating video generation quality with versatile text-to-image diffusion models. arXiv preprint arXiv:2403.05438, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "type": "text", + "content": "[58] Zhenghao Zhang, Junchao Liao, Menghao Li, Long Qin, and Weizhi Wang. Tora: Trajectory-oriented diffusion transformer for video generation. arXiv preprint arXiv:2407.21705, 2024. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 208, + 553, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 272 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 272 + ], + "type": "text", + "content": "[59] Rui Zhao, Yuchao Gu, Jay Zhangjie Wu, David Junhao Zhang, Jia-Wei Liu, Weijia Wu, Jussi Keppo, and Mike Zheng Shou. MotionDirector: Motion customization of text-to-video diffusion models. In Proceedings of the 18th European Conference on Computer Vision, pages 273–290, 2024. 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 274, + 553, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 274, + 553, + 329 + ], + "spans": [ + { + "bbox": [ + 316, + 274, + 553, + 329 + ], + "type": "text", + "content": "[60] Wenliang Zhao, Yongming Rao, Zuyan Liu, Benlin Liu, Jie Zhou, and Jiwen Lu. Unleashing text-to-image diffusion models for visual perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5706-5716, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 331, + 553, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 385 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 385 + ], + "type": "text", + "content": "[61] Wendi Zheng, Jiayan Teng, Zhuoyi Yang, Weihan Wang, Ji-dong Chen, Xiaotao Gu, Yuxiao Dong, Ming Ding, and Jie Tang. CogView3: Finer and faster text-to-image generation via relay diffusion. In Proceedings of the 18th European Conference on Computer Vision, pages 1-22, 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 386, + 553, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 386, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 386, + 553, + 430 + ], + "type": "text", + "content": "[62] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. MagicVideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 431, + 553, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 431, + 553, + 485 + ], + "spans": [ + { + "bbox": [ + 316, + 431, + 553, + 485 + ], + "type": "text", + "content": "[63] Dewei Zhou, You Li, Fan Ma, Xiaoting Zhang, and Yi Yang. MIGC: Multi-instance generation controller for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6818-6828, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 487, + 553, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 487, + 553, + 530 + ], + "spans": [ + { + "bbox": [ + 316, + 487, + 553, + 530 + ], + "type": "text", + "content": "[64] Dewei Zhou, You Li, Fan Ma, Zongxin Yang, and Yi Yang. MIGC++: Advanced multi-instance generation controller for image synthesis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 47(3):1714-1728, 2025. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 532, + 553, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 532, + 553, + 575 + ], + "spans": [ + { + "bbox": [ + 316, + 532, + 553, + 575 + ], + "type": "text", + "content": "[65] Haitao Zhou, Chuang Wang, Rui Nie, Jinxiao Lin, Dongdong Yu, Qian Yu, and Changhu Wang. TrackGo: A flexible and efficient method for controllable video generation. arXiv preprint arXiv:2408.11475, 2024. 2, 3" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7275" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/IMFine_ 3D Inpainting via Geometry-guided Multi-view Refinement/0becb03e-65d1-4a50-a09c-58afb744258c_content_list.json b/2025/IMFine_ 3D Inpainting via Geometry-guided Multi-view Refinement/0becb03e-65d1-4a50-a09c-58afb744258c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1536015cd5bf9263d3d348de42de0e16d3253dea --- /dev/null +++ b/2025/IMFine_ 3D Inpainting via Geometry-guided Multi-view Refinement/0becb03e-65d1-4a50-a09c-58afb744258c_content_list.json @@ -0,0 +1,1980 @@ +[ + { + "type": "text", + "text": "IMFine: 3D Inpainting via Geometry-guided Multi-view Refinement", + "text_level": 1, + "bbox": [ + 153, + 130, + 843, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhihao Shi\\*1", + "bbox": [ + 209, + 179, + 316, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dong Huo\\*o2", + "bbox": [ + 361, + 180, + 468, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Juwei Lu1", + "bbox": [ + 375, + 199, + 457, + 214 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuhongze Zhou", + "bbox": [ + 514, + 180, + 647, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yan Min", + "bbox": [ + 700, + 180, + 785, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xinxin Zuo4", + "bbox": [ + 519, + 198, + 620, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Huawei Canada Research Institute", + "bbox": [ + 246, + 215, + 527, + 231 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ University of Alberta", + "bbox": [ + 573, + 215, + 750, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ McMaster University", + "bbox": [ + 297, + 232, + 475, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4Concordia University", + "bbox": [ + 521, + 233, + 699, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zhihaoshi2022@gmail.com, dhuo@ualberta.ca, yuhongze.zhou@mail.mcgill.com", + "bbox": [ + 184, + 253, + 821, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "miny13@mcmaster.ca, juwei.lu@huawei.com, xinxin.zuo@concordia.ca", + "bbox": [ + 215, + 272, + 784, + 286 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/802c94cab733739ffc55572c05715245fe5039bb40248878ee56001ddc0327ca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 91, + 321, + 225, + 422 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5dc2cf05ddd57a12ecd97ea59658a2da28c9dd06ede82f459e9d57adae0d1c06.jpg", + "image_caption": [ + "Figure 1. We present a novel 3D inpainting method capable of removing target objects from the 3D scenes and seamlessly inpainting the occluded regions across various scenarios. The proposed method ensures the generation of plausible geometry and consistent textures. It is highly recommended to see video results in our project page." + ], + "image_footnote": [], + "bbox": [ + 91, + 422, + 222, + 523 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/80f8bb6e32cd2512f1ead2a7f4eef8a3cf4c3e2ff5e3d4ff4f6641a4a633f502.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 225, + 357, + 261, + 387 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/330b90eb00feaa3260b89887b0a9660ffa4e74064dc2096c7d4b52475c9b3dc7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 261, + 321, + 516, + 421 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bfa34b0b2843481996d6b67ca21be45f598f982c8be62bb92a0713fca16e2ad7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 321, + 776, + 421 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b6ede422d3009909ed3ed87dcd79711bdecf1fea85e748ac4fd19604344ae172.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 777, + 321, + 903, + 421 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/88ca5647b1541e54ab9c3c16950ce39224e3ffe9a3690105d30a17862dee6e85.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 261, + 422, + 516, + 523 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d2eb3d80fddd61e745966c05b441e0a16e913afa23d6b5959af0edf9451e2cf3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 422, + 776, + 523 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 590, + 326, + 606 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Current 3D inpainting and object removal methods are largely limited to front-facing scenes, facing substantial challenges when applied to diverse, \"unconstrained\" scenes where the camera orientation and trajectory are unrestricted. To bridge this gap, we introduce a novel approach that produces inpainted 3D scenes with consistent visual quality and coherent underlying geometry across both front-facing and unconstrained scenes. Specifically, we propose a robust 3D inpainting pipeline that incorporates geometric priors and a multi-view refinement network trained via test-time adaptation, building on a pretrained image inpainting model. Additionally, we develop a novel inpainting mask detection technique to derive targeted inpainting masks from object masks, boosting the performance in handling unconstrained scenes. To validate the efficacy of our approach, we create a challenging", + "bbox": [ + 88, + 622, + 485, + 864 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and diverse benchmark that spans a wide range of scenes. Comprehensive experiments demonstrate that our proposed method substantially outperforms existing state-of-the-art approaches.", + "bbox": [ + 511, + 592, + 906, + 654 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 679, + 643, + 694 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D editing is an active research area with significant potential for various applications, including video games, the Metaverse, virtual reality, and augmented reality. For example, after a 3D scene is created or reconstructed, a user might want to remove certain objects from the scene. This paper mainly focuses on the challenging task of inpainting in 3D scenes to fill the occluded regions to enable seamless object removal.", + "bbox": [ + 509, + 703, + 906, + 824 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Compared to image inpainting [22, 28], 3D inpainting models lag significantly behind due to the lack of large-scale 3D datasets and efficient 3D data structures. Therefore, existing 3D inpainting methods mainly focus on lifting 2D inpainting priors to 3D. Although the current diffusion", + "bbox": [ + 511, + 825, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* denotes equal contribution.", + "bbox": [ + 89, + 875, + 248, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Work done during an internship at Huawei Canada Research Institute", + "bbox": [ + 91, + 888, + 468, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "26694", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "based image inpainting models can generate images with vivid textures and high-fidelity details, they are trained on single images without considering any 3D consistency. It poses great challenges in maintaining view consistency and preserving 3D coherence when lifting those 2D priors to 3D.", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the inconsistency issue, there are approaches that implicitly lift the 2D models into 3D via score distillation sampling (SDS) loss [3, 18], which was proposed initially in text-to-3D generation [20], to harness the inherent coherence of 2D diffusion models under the same prompt, complemented by the iterative updating scheme during optimization to ensure view consistency. However, they are constrained by the inherent limitation of SDS loss, which results in over-saturation and over-smoothing in the optimized images, even with different variants of SDS losses. Alternatively, explicit methods [2, 5, 16, 17, 30] first generate one or few images with the image inpainting models and try to preserve the 3D consistency by enforcing geometry constraints or minimizing perceptual losses between rendered images and the inpainted 2D images. Although they have demonstrated good performance, especially for front-facing scenes with very limited view changes, the inpainted 3D scenes have artifacts such as floaters and blurry textures in unconstrained scenarios with complex geometry and large view variation. The reasons are that first, 3D geometry is rather difficult to get perfectly restored; second, perceptual loss can only alleviate the inconsistency by minimizing overall similarity. Inconsistency still exists between those inpainted images, leading to blurry textures for the inpainted 3D scenes. More recently, [2] proposed to train a multi-view inpainting network to generate consistent multiview inpainted images. However, due to the limited diversity of current multi-view datasets, they cannot generalize to unconstrained 3D scenes (e.g., $360^{\\circ}$ cases).", + "bbox": [ + 91, + 186, + 486, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To enhance the ability to generalize to various scenes for 3D inpainting, we utilize geometry priors and develop a multi-view refinement network fine-tuned at test time to produce consistent images in a coarse-to-fine manner. Specifically, we start with a single reference view inpainted with image inpainting models, followed by coarse geometry restoration and a warping operation to warp the inpainted content from the reference view to others. However, due to the imperfect reconstructed geometry and view-dependent effects, the warped images tend to have artifacts, such as distorted textures and mismatches between the warped and neighboring contents. To refine the warped contents, we present a multi-view refinement model to generate consistent images with high fidelity, taking the warped images as a condition. The key idea is that instead of training a model with extensive datasets, we take the pre-trained image inpainting model, add space-time attention layers, and fine-tune them with the current test scenes. We take the", + "bbox": [ + 89, + 628, + 486, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "originally captured multi-view images and implement various augmentations to synthesize paired data to fine-tune the multi-view refinement model.", + "bbox": [ + 511, + 90, + 903, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Furthermore, previous works usually take the entire object masks as inpainted regions. However, some regions inside the object masks might have already been exposed in other views, especially for 3D object removal and inpainting in $360^{\\circ}$ scenes (see Fig. 4). Therefore, we believe it is critical to differentiate between object masks and inpainting masks, fully exploiting the available information of the reconstructed 3D scenes. To this end, a novel inpainting mask detection approach is proposed to extract masks accurately and exclusively for real occluded regions.", + "bbox": [ + 511, + 137, + 906, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Finally, considering that the existing benchmarks mostly cover front-facing scenes or planar surfaces, we introduce a new dataset encompassing 20 diverse scenes, from indoor to outdoor environments, covering large view variations and complex interfacial surfaces, to better evaluate and advance 3D inpainting methods.", + "bbox": [ + 511, + 289, + 908, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are as follows:", + "bbox": [ + 513, + 381, + 813, + 395 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. A multi-view consistent inpainting pipeline is proposed. We leverage geometric priors and a multi-view refinement model that is fine-tuned from image inpainting models via test-time adaptation to generate consistent multi-view images.", + "2. We further develop an inpainting mask detection technique that accurately distinguishes occluded regions, enabling robust processing of complex scenes, such as $360^{\\circ}$ environments.", + "3. The proposed inpainting method has achieved state-of-the-art performance across various benchmarks, particularly excelling in challenging, unconstrained scenarios.", + "4. We contribute a new and challenging 3D scene inpainting benchmark, offering diverse scenes to facilitate broader evaluation and advancement in 3D inpainting research." + ], + "bbox": [ + 514, + 397, + 903, + 637 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 654, + 656, + 669 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. 3D Inpainting", + "text_level": 1, + "bbox": [ + 513, + 680, + 658, + 696 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing 3D inpainting methods [3, 5, 16-18, 30, 32] are implemented by lifting 2D inpainting priors to 3D with various strategies designed to ensure 3D consistency. One direction, inspired by score distillation-based approaches [20] in text-to-3D generation, formulates 3D inpainting as a continuous distillation process [3, 18]. While these approaches achieve reasonably consistent inpainting results, they face inherent limitations of score distillation, such as producing over-saturated or overly smooth outputs. Another direction explicitly enforces 3D consistency by independently inpainting one or multiple 2D views and then consolidating multi-view inconsistencies in 3D. Reference-guided NeRF [16] addresses this challenge by using a single view to", + "bbox": [ + 511, + 703, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "26695", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "guide geometry inpainting to avoid inconsistencies. However, it struggles with accurate geometry capture in complex $360^{\\circ}$ scenes using just a single view. SPINeRF [17], and InNeRF360 [30] utilize multiple inpainted images to guide 3D inpainting and incorporate perceptual loss to reduce inconsistencies resulting from independently inpainted images. Nevertheless, while perceptual loss can maintain overall consistency, it performs poorly on finer details, often leading to blurry 3D results. Our method lies in the latter direction. It addresses these issues through a geometry-guided refinement network that generates a set of multi-view consistent images, reconstructing clearer and consistent 3D inpainted scenes.", + "bbox": [ + 89, + 90, + 480, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Video and Multi-View Diffusion", + "text_level": 1, + "bbox": [ + 89, + 297, + 372, + 313 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The diffusion model for text-to-image (T2I) generation [7, 8, 22, 27] has achieved significant success in recent years, known for its ability to produce high-quality, diverse, and controllable outputs. Building on this, extensive studies have extended image diffusion models to the video generation domain [9, 10, 26, 40] by incorporating temporal layers within each transformer block to perform space-time attention, which captures correlations across frames and shows promising results in generating realistic videos.", + "bbox": [ + 89, + 321, + 483, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inspired by these advances in video generation, researchers have begun exploring the adaptation of T2I models into multi-view diffusion models for implicit 3D reasoning [15, 24, 25, 31]. For example, methods such as MVDream [25] and Zero $123++$ [24] leverage this idea and derive multi-view diffusion models, capable of generating consistent multiple views simultaneously for 3D generation. However, the full space-time attention mechanism used in these methods can lead to prohibitive memory and computational costs, making it impractical for handling large numbers of views. To address this, recent work [35] introduces a sparse space-time attention mechanism, which reduces memory requirements by attending only to the initial and previous frames. However, the sparse attention approach assumes input frames are from a continuous video sequence and remains under-explored in sparse multi-view scenarios.", + "bbox": [ + 89, + 458, + 483, + 714 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Diffusion Model Fine-tuning", + "text_level": 1, + "bbox": [ + 89, + 724, + 346, + 742 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Adapting large pre-trained diffusion models has proven to be an effective and efficient approach for completing user-specified tasks, such as synthesizing specific subjects in new contexts [4, 23] and inpainting images with instance matching reference images [29, 36]. Adaptation methods include injecting extra layers [19, 36, 38, 39], searching appropriate unique prompt [23], updating partial or full model weights [14, 18, 29], or combinations of these techniques. Similarly, in this paper, instead of relying on extensive multiview datasets, we propose adapting the pre-trained image", + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "diffusion model into a multi-view refinement model for 3D inpainting purposes.", + "bbox": [ + 511, + 90, + 906, + 121 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminary", + "text_level": 1, + "bbox": [ + 513, + 133, + 638, + 152 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Gaussian splating (GS) methods [11, 12] represent a scene as a collection of multivariate Gaussian primitives: $\\mathcal{G} = \\{(\\mu_m,\\Sigma_m,\\Gamma_m,\\mathrm{SH}_m)\\}_{m = 1}^M$ , where $\\mu_{m},\\Sigma_{m},\\Gamma_{m}$ and $\\mathrm{SH}_m$ denote the mean vector, covariance matrix, opacity, and spherical harmonic coefficients of a multivariate Gaussian, respectively. View synthesis is performed by standard point-based $\\alpha^\\prime$ -rendering, where a single pixel color is computed by sorting $\\mathcal{G}$ (or their 2D projection) in their depth order and alpha-blending each contributed primitive:", + "bbox": [ + 511, + 160, + 906, + 297 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nC = \\sum_ {m} c _ {m} \\alpha_ {m} \\prod_ {j = 1} ^ {m - 1} (1 - \\alpha_ {j}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 308, + 906, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $C$ is the output pixel color, $c_{m}$ is the pre-computed color of the $m$ -th primitive, and $\\alpha_{m}$ is a calculated opacity based on distance from the Gaussian center to the interaction point. Due to the excellent performance in geometry reconstruction, in this work, we choose 2D-GS [11] as the 3D representation.", + "bbox": [ + 511, + 359, + 905, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 513, + 464, + 604, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a GS scene $\\mathcal{G}$ , along with corresponding training images $\\{I_n\\}_{n=1}^N$ , associated camera poses $\\{\\Pi_n\\}_{n=1}^N$ , and a set of object masks $\\{M_n\\}_{n=1}^N$ identifying objects to be removed, our goal is to generate an inpainted scene $\\mathcal{G}^*$ where the target objects are removed and occluded regions are filled seamlessly with the background. The entire pipeline contains two stages, namely object removal and 3D scene inpainting. We first remove object primitives (Sec. 4.1), exposing never-before-seen (NBS) regions which are the areas that truly need inpainting. We then either use the object masks as inpainting masks when they are identical or detect specific inpainting masks (Sec. 4.3), ensuring that the inpainting mask exclusively covers the NBS regions in each view, denoted as $\\{P_n\\}_{n=1}^N$ . Given the pruned GS scene and inpainting masks, we propose a novel 3D inpainting pipeline to inpaint $\\mathcal{G}$ as $\\mathcal{G}^*$ (Sec. 4.2).", + "bbox": [ + 511, + 489, + 906, + 732 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Object Removal", + "text_level": 1, + "bbox": [ + 513, + 739, + 671, + 757 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To remove the objects from the reconstructed GS scene, we append an additional learnable parameter $\\{p_m\\}_{m=1}^M \\in [0,1]$ to each Gaussian primitive, indicating whether it belongs to the object to be removed, which is initialized as 0.5. To compute the value of $p_m$ for each Gaussian, we map the object masks back to the Gaussian primitives. Specifically, we render a set of object masks $\\{\\hat{M}\\}_{n=1}^N$ using $\\alpha'$ rendering (similar to color rendering) at $\\{\\Pi_n\\}_{n=1}^N$ but replacing $c_m$ with $p_m$ in Eq. 1, followed by $K_1$ ( $K_1 = 1000$ ) steps of the", + "bbox": [ + 511, + 763, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "26696", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7968f66a025678ba9933d82a59a7b787e212e73a9c0c17d9b2de96c033e38049.jpg", + "image_caption": [ + "(a) Multi-view Inpainting Pipeline" + ], + "image_footnote": [], + "bbox": [ + 91, + 107, + 529, + 303 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d83ad3298341204ba8fc517ea7f5d381d85159f0f684b99db962b341f96a7a50.jpg", + "image_caption": [ + "(b) Test-time Adaptation", + "Figure 2. Overview of our proposed approach. (a) shows the proposed 3D inpainting pipeline. Specifically, a reference view is first inpainted, which will be used to reconstruct complete geometry, along with rendered multi-view depth maps. The inpainted reference image is then warped to other views, which are taken as input conditions for the multi-view refinement network to generate a set of view-consistent images. Afterward, we fine-tune the 3D scene to get the inpainted scene. (b) illustrates the detailed structure and training of the multi-view refinement network. Basically, we take the pre-trained image inpainting model, incorporate additional space-time attention layers, and exploit the test-time adaptation to fine-tune it into a multi-view refinement model. To generate a fine-tuning dataset, we synthesize warped images under the original captured scene, which has the ground-truth captured images. In detail, reference views are selected randomly, and we set random masks over the images, apply geometry jittering, and warp the reference images in other views." + ], + "image_footnote": [], + "bbox": [ + 535, + 109, + 903, + 304 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "optimization to update $p_m$ with respect to $\\mathcal{L}_1$ loss between $\\{\\hat{M}\\}_{n=1}^N$ and $\\{M\\}_{n=1}^N$ . Notably, other parameters remain unchanged during this procedure. Finally, a pruning operation is performed to produce the pruned scene $\\mathcal{G}_{\\mathrm{pruned}}$ by thresholding $p_m$ with a threshold $\\tau$ ( $\\tau = 0.4$ ).", + "bbox": [ + 89, + 436, + 483, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Proposed 3D Inpainting Pipeline", + "text_level": 1, + "bbox": [ + 89, + 523, + 377, + 540 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the pruned GS scene $\\mathcal{G}_{\\mathrm{pruned}}$ as well as inpainting masks $\\{P_n\\}_{n=1}^N$ indicating the regions to be inpainted for each view, we propose a novel 3D inpainting pipeline shown in Fig 2. We first render color images $\\{I_n'\\}_{n=1}^N$ from $\\mathcal{G}_{\\mathrm{pruned}}$ for each view and select a reference view with image $I_r' \\in \\{I_n'\\}_{n=1}^N$ , which is then inpainted using a pre-trained inpainting model [22] to reference image $I_r^*$ . We conduct geometry reconstruction to fill out the hole in 3D, which will be utilized to warp the inpainted contents in the reference image into the remaining views. We employ a multi-view refinement model to further refine the warped images, ensuring high fidelity and consistency, and then update the 3D scene. The model is fine-tuned from the pre-trained image inpainting model via a test-time adaptation strategy. The geometry reconstruction and multi-view refinement model are described as follows.", + "bbox": [ + 89, + 547, + 483, + 790 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2.1 Geometry Reconstruction", + "text_level": 1, + "bbox": [ + 89, + 815, + 323, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following [11], we render a set of depth maps $\\{D_n\\}_{n=1}^N$ from the prune GS scene $\\mathcal{G}_{\\mathrm{pruned}}$ and adopt a truncated signed distance function (TSDF) [6] technique to reconstruct the 3D mesh. However, $\\{D_n\\}_{n=1}^N$ only contributes", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "to the reconstruction of the background region, as the depth values within their inpainting mask areas are invalid. To fill the masked region, we aim to generate an all-valid depth map $D_r^*$ of the reference view by filling the masked areas in $D_r$ based on information from $I_r^*$ . The filled depth map should meet two key requirements: 1) it must be consistent with the reference image $I_r^*$ ; 2) it should seamlessly connect with the depth values outside the masked region.", + "bbox": [ + 511, + 436, + 906, + 556 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We begin by exploiting a monocular depth estimator [37] to estimate depth $\\bar{D}_r$ from $I_r^*$ , followed by a linear transformation to convert $\\bar{D}_r$ to scene scale $\\tilde{D}_r$ : $\\tilde{D}_r \\colon \\tilde{D}_r = a * \\bar{D}_r + b$ , where the parameters $a, b$ are computed in the neighborhood of $P_r$ to minimize the difference between $\\bar{D}_r$ and $D_r$ in advance, following [16]. We then develop a filling scheme according to $\\tilde{D}_r$ . Formally, the filled values of $D_r^*$ are computed as follows:", + "bbox": [ + 511, + 565, + 908, + 686 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} D _ {r} ^ {*} (x) = \\underset {D _ {r} ^ {*} (x)} {\\arg \\min } \\sum_ {x \\in P _ {r}} | | \\nabla D _ {r} ^ {*} (x) - \\nabla \\tilde {D} _ {r} (x) | | ^ {2} + \\lambda \\nabla D _ {r} ^ {*} (x) \\tag {2} \\\\ \\begin{array}{l} \\text {s . t .} D _ {r} ^ {*} (x) = D _ {r} (x), \\text {w h e r e} P _ {r} (x) = 0, \\end{array} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 713, + 908, + 781 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\nabla$ denotes the Laplacian calculation and $x$ represents the pixel coordinate. Intuitively, the first term encourages $D_r^*$ to maintain a similar changing trend as $\\tilde{D}_r$ , while the second term regularizes $D_r^*$ to be smooth and consistent with the depths outside the mask. As such, $D_r^*$ can be used for the reconstruction of the NBS region.", + "bbox": [ + 511, + 810, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "26697", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/88981aea9c4662bcea365b0224b5d7bbf5df389647a3280871185b14de82bfdb.jpg", + "image_caption": [ + "(a) Reference Image" + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 222, + 186 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3ec1a30d6659afcffad3ba55f2073a41cdb4aa20caecd5365eb545408158abbd.jpg", + "image_caption": [ + "(b) Warped Image" + ], + "image_footnote": [], + "bbox": [ + 222, + 88, + 346, + 186 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7378e1bf8d1d89f23736300e8914ce815fa40c5ab0ffec7fb51bd432026cd2a2.jpg", + "image_caption": [ + "(c) Refined Image", + "Figure 3. Illustration of the importance of multi-view refinement. The multi-view refinement model is proposed to tackle the warping artifacts with the example shown in (b), generating high-fidelity multi-view consistent images with one sampled view shown in (c)." + ], + "image_footnote": [], + "bbox": [ + 348, + 88, + 473, + 186 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.2 Test-Time Adaptation for Multi-view Refinement", + "text_level": 1, + "bbox": [ + 89, + 282, + 482, + 297 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We warp the inpainted contents from the reference view $I_{r}$ to other views with the above restored geometry. Due to the imperfect geometry and the neglect of view-dependent effects, those warped contents contain artifacts, such as texture distortion and inconsistent appearance with the neighboring regions, especially when there are large view angle changes with respect to the reference view as shown in Fig 3. On the other hand, these warped contents have coarse-level consistency across views. To this end, we propose a multi-view refinement model that produces high-fidelity, consistent images conditioning on the warped images.", + "bbox": [ + 89, + 309, + 483, + 489 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Network: Our model is built upon a pre-trained diffusion-based image inpainting model [22], which takes a masked image and its associated mask as the input, corresponding to the warped image and inpainting mask in our scenarios, respectively. To ensure multi-view consistency, we replace the self-attention layer in each transformer block with space-time attention to explore correlations across multiple views. Considering the quadratic complexity of global attention and the large number of warped images, space-time attention is implemented as the sparse space-time attention [10] that only attends to neighboring images. The attention layer is also allowed to attend to the reference image for alignment.", + "bbox": [ + 89, + 491, + 483, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training: Instead of relying on large datasets and extensive training, we exploit the test-time adaptation strategy for each scene to train the multi-view refinement network. Paired training sets are created for model adaptation. Specifically, given the original images $\\{I_n\\}_{n=1}^N$ , at each training step, $L + 1$ images are randomly selected as ground truth. To simulate realistic warping scenarios, we randomly choose one image as the reference, generate a random mask, and warp both the reference image and mask to other views with geometry guidance from $\\mathcal{G}$ . In addition, to encourage cross-view learning as in [2], we also synthesize another set of paired images by applying image augmentations independently on each captured image, i.e., irregularly shaped masks that are randomly generated on", + "bbox": [ + 89, + 689, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b0432dd7b66401940f44d0310e20a9b8bfb2459010e1664b471a952a376df89c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 521, + 88, + 709, + 186 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7d5206d2cdc47c5069a14666acfbe0d2d3bd6a29a334ecf1653616312322ea5d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 88, + 900, + 186 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/eda211d0b1eb916ea42106ff58661e6affc2b6af5654d671649e248da47c942c.jpg", + "image_caption": [ + "(a) Object mask and dilated mask", + "(c) Mask-remapping interpretation", + "Figure 4. Inpainting mask detection. (a) shows the object mask (orange) along with the dilated mask (pink), which includes additional neighboring pixels; Each dilated mask consists of two parts: the actual inpainting mask and out-of-interest areas, marked as regions $①$ and $②$ in (b). These regions are mapped in 3D to the NBS and background areas, respectively, as shown in regions $①$ and $②$ in (c). During multi-view mapping optimization, the central NBS region is progressively refined, while background areas are suppressed by each other; (d) shows the rendered neighbor mask after GS scene optimization." + ], + "image_footnote": [], + "bbox": [ + 517, + 200, + 705, + 299 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/67965b8e942fe8f3b2ea1f02e734edc57d6164253a69e28ba82e7ef84834dd19.jpg", + "image_caption": [ + "(b) Object Mask Separation", + "(d) Render mask" + ], + "image_footnote": [], + "bbox": [ + 709, + 200, + 898, + 299 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "each image. Image-based augmentations—such as elastic transformations and color jittering, are then applied to the masked regions to simulate warping artifacts. We train the multi-view refinement model with those paired images, employing a simplified variational bound objective [8].", + "bbox": [ + 511, + 470, + 905, + 546 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inference: At inference, similar to [33], we randomly shuffle the order of the $L$ warped images at each denosing step to break the sequential constraint of sparse attention, further enhancing cross-view consistency.", + "bbox": [ + 511, + 547, + 905, + 607 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.3 Fine-tune GS Scene", + "text_level": 1, + "bbox": [ + 511, + 628, + 705, + 642 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With the generated multi-view consistent images, we fine-tune $\\mathcal{G}_{\\mathrm{pruned}}$ for $K_{2}(K_{2} = 7000)$ steps with an $L_{1}$ loss and a SSIM loss as straightforward pixel-wise supervision. It is worth noting that other robust losses (e.g., perceptual loss) are functionally orthogonal to our approach and can be used to improve it further.", + "bbox": [ + 511, + 654, + 905, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Inpainting Mask Detection", + "text_level": 1, + "bbox": [ + 511, + 755, + 754, + 771 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the overall pipeline of 3D object removal and scene inpainting, identifying the occluded or NBS regions for inpainting is also critical, especially for $360^{\\circ}$ cases. In those cases, the object mask often covers a much larger area than the regions that require inpainting (as illustrated in Fig. 4). Therefore, a novel inpainting mask detection approach is proposed to identify the NBS region for each view.", + "bbox": [ + 511, + 779, + 905, + 883 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "An intuitive idea is to detect the NBS region in the", + "bbox": [ + 532, + 885, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "26698", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "pruned GS scene and project it back into each view to obtain inpainting masks. However, since, by definition, the NBS region is occluded by the object, it usually lacks good geometry and meaningful Gaussians, making it challenging to detect in 3D. Here we propose to detect the inpainting masks by leveraging neighbors of NBS region. In detail, we first dilate the object masks to include neighboring pixels (Fig. 4(a)), and then map the dilated masks onto the pruned scene $\\mathcal{G}_{\\mathrm{pruned}}$ . Mapping is implemented as updating the learnable parameter $\\{p_m\\}_{m=1}^M$ of each Gaussian with respect to $\\mathcal{L}_1$ loss between rendered masks of $\\mathcal{G}_{\\mathrm{pruned}}$ and the dilated masks. The optimization process is the same as object removal (Sec. 4.1) but with the object mask replaced with dilated masks, and the optimization is performed over the pruned scene. We illustrate the interpretation of this process in Fig. 4(b-c). After the mapping process, we are able to get the neighbor mask by rendering $\\{p_m\\}_{m=1}^M$ from $\\mathcal{G}_{\\mathrm{pruned}}$ as shown in Fig. 4(d). Finally, we extract the bounding box for the above rendered mask, which will be taken as input prompt for segment anything model [13] to get the final inpainting mask.", + "bbox": [ + 89, + 90, + 480, + 409 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 424, + 222, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets: Most existing publicly available 3D inpainting datasets [17, 34] are limited to front-facing cases, which cannot be used to evaluate methods in more diverse or realistic situations. To address this, we construct a dataset that contains 20 diverse cases, ranging from indoor to outdoor, covering various camera trajectories-including front-facing, $90^{\\circ}$ , $180^{\\circ}$ , and $360^{\\circ}$ , and containing various interfacial surfaces such as single plane, multiple planes, curvature surfaces, and irregular surfaces. Each scene is recorded using a DJI Pocket 3 at 60fps with a resolution of $1920 \\times 1080$ , from which we evenly extract 175 images and generate the corresponding camera poses and object masks. The first 50 images are used for evaluation purpose with the object removed, while the remaining 125 images are object-present and are used to reconstruct the original GS scene. Previews of our collected dataset are shown in supplementary materials. We use half resolution in the following experiments. In addition, we evaluate our method on SPINeRF dataset [17] to compare with state-of-the-art.", + "bbox": [ + 89, + 450, + 482, + 736 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details: StableDiffusionInpainting-Pipeline is utilized for image inpainting and our multi-view refinement model initialization. During training, we randomly select $L = 5$ images from training images and crop $512 \\times 512$ patches to fine-tune the model over 1000 iterations with the learning rate and batch size set as $3 \\times 10^{-5}$ and 1, individually. At inference, directly refining all warped images, i.e., $L = 125$ would result in prohibitive memory usage and computational costs. To", + "bbox": [ + 89, + 738, + 482, + 875 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5108cb807a4550bc3965e1a60d6873f7fc10edea040f0beeed2c73dffb72769a.jpg", + "table_caption": [ + "Table 1. Quantitative evaluation on the SPINeRF dataset and our dataset. Numbers in bold indicate the best performance, and underscored numbers indicate the second best." + ], + "table_footnote": [], + "table_body": "
MethodOursSPINeRF
PSNR (↑)LPIPS (↓)FID (↓)PSNR (↑)LPIPS (↓)FID (↓)
GaussianEditor15.710.6163375.0314.410.6247343.16
GScream17.180.4431290.6316.960.3931154.71
SPInNeRF18.750.3519206.4317.470.5740239.99
MVPIneRF18.630.4332278.9917.670.5070255.51
Ours19.670.2685149.5217.580.4513154.34
", + "bbox": [ + 514, + 131, + 903, + 213 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "address this, we cluster all views into $L$ groups and select the center view of each group as the anchor. Multi-view refinement is applied only to these $L$ anchor views, while the remaining views in each group are warped from their respective anchors, as artifacts caused by warping between closely aligned views are negligible. $L$ is set to 12 for $360^{\\circ}$ cases and linearly scaled down for other scenarios, such as 4 for $90^{\\circ}$ cases. We directly use object masks when evaluating on front-facing scenes and apply inpainting mask detection on others. All experiments are conducted on a single NVIDIA V100 GPU, with the inpainting of a 3D scene taking approximately one hour.", + "bbox": [ + 511, + 229, + 906, + 412 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Evaluation against State-of-the-Art Methods", + "text_level": 1, + "bbox": [ + 511, + 422, + 890, + 439 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate our method against the state-of-the-art methods on our dataset and SPINeRF dataset. The following state-of-the-art methods are reproduced on our datasets, including GaussianEditor [5], GScream [32], MVIPNeRF [3] and SPInNeRF [17] using their official implementations. In particular, since the naive NeRF models used in MVIPNeRF and SPInNeRF cannot handle complex unbounded scenes, we replace them with MeRF [21] considering both memory usage and reconstruction performance. On our dataset, we utilize the evaluation protocol and official implementation from [3] to compare methods. Specifically, we render color images at corresponding test views from the inpainted scenes and evaluate them against the ground truth using three metrics, namely PSNR, LPIPS, and FID. These metrics are calculated only within the bounding box defined by the object mask. For the SPINeRF dataset, we copy values from [3] for SPInNeRF and MVIPNeRF, and we reproduce GaussianEditor and GScream to align with the metric computation protocol in [3].", + "bbox": [ + 511, + 446, + 906, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As can be seen from Tab. 1, the proposed method consistently achieves the best performance on our dataset, with a particularly large margin that demonstrates its superiority in unconstrained realistic scenarios. On the SPINeRF dataset, our method achieves state-of-the-art results comparable to other approaches, given that this dataset is limited to front-facing cases and challenging to distinguish methods in handling complex cases.", + "bbox": [ + 511, + 734, + 905, + 853 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In addition, we provide qualitative comparisons in Fig. 5, where our method successfully removes objects and seamlessly inpaints with appropriate content. Moreover, the in", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "https://huggingface.co/stabilityai/stable-diffusion-2-inpainting", + "bbox": [ + 91, + 886, + 421, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "26699", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9fd45b9fab78c3e9a2527e9fd5205b01a9011690108551aff11c6a29a9b58681.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 106, + 89, + 328, + 262 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/01996cb03495c23b237d7263cf289df809e7e37ed041bcc0ea67549d142cb25c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 331, + 90, + 441, + 261 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d368c73cf0a32cd2779b81ac836b1c506b37e07ec5402fa14be8bac51375afed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 442, + 90, + 553, + 261 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6dc19e7aa41c2ce6438f5942a191f0594835e39db08764b3172189bef8454922.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 555, + 90, + 665, + 261 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/29ef3461c4289f94df28cadce23062c5e7ed20c7975b04a2cc326109b2cb55bb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 666, + 90, + 777, + 261 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/393a64c93ad923a356538319f047929bf7b23d99bfdea56b7190f941e4b060bb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 779, + 90, + 890, + 261 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/23b4d95936c79134a3ee59bae087a35ffbded0d3fb10a43fa17bb26dc2d8b3d4.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 106, + 263, + 328, + 433 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/064e37c5ac318744af35fac67b8967bc2c999c883a1dd6c0c3435861cfdf6272.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 331, + 263, + 441, + 347 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e7e636c3576c2b4cd7ba19bd0ab1853f0199428da4111d1c7df7e29c8fc29e79.jpg", + "image_caption": [ + "GaussianEditor" + ], + "image_footnote": [], + "bbox": [ + 331, + 348, + 441, + 433 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b314c43523702de0c838c532eddedc7f8a75ff56a0a2971881d8efd640fe1726.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 444, + 263, + 553, + 347 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7a0245a2d170bdb0759f41c5ea56c32bf28e4fb17832cc083355337da73668d9.jpg", + "image_caption": [ + "GSream", + "Figure 5. Qualitative evaluation. We show the visual comparisons against the state-of-the-art 3D inpainting algorithms, with two views displayed at the top and bottom for each method. Our method generates consistent, higher-quality results with rich texture details and fewer artifacts." + ], + "image_footnote": [], + "bbox": [ + 444, + 348, + 553, + 433 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/18057c39ebeec521382308ebb62748e9c88aa5cce12cb7fff284d36dd2d3b0a5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 555, + 263, + 665, + 348 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/23c613235a9a7650ebcec634059754b0e36b1b7eb92a1a457601de563c784ad0.jpg", + "image_caption": [ + "SPInNeRF" + ], + "image_footnote": [], + "bbox": [ + 555, + 348, + 665, + 433 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/76577a197a90ffb1f2c874fe28a04038d84e28882547323ff888a62e4d5627ae.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 666, + 263, + 777, + 348 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/aadc38e749d5d184ec8b39ccb4ea2f87f13adfe6bc75682c7cfc609f808ac06a.jpg", + "image_caption": [ + "MVIPNeRF" + ], + "image_footnote": [], + "bbox": [ + 666, + 348, + 777, + 433 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d1bcbcbe4d8838ccf3ed6c1bc363d868c93e33e11983ca4fc616ab8629507eac.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 779, + 263, + 890, + 348 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f79a365c1e4dd18df55890efd61162f1add6aa610e4cb234eb03889fdcbfbcc3.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 779, + 348, + 890, + 433 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "painted scene maintains 3D consistency, yielding visually more pleasing outcomes and a coherent geometric structure. In contrast, other methods often produce severe artifacts and fail to inpaint objects effectively in complex, realistic scenarios. We highly recommend viewing our video visualizations for a more comprehensive comparison.", + "bbox": [ + 89, + 508, + 483, + 599 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Detection of Inpainting Masks", + "text_level": 1, + "bbox": [ + 89, + 616, + 359, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We then demonstrate the importance of doing inpainting mask detection and differentiating them from object masks. For our collected dataset, we manually annotate inpainting masks from rendered images of pruned scenes and compare them with the original object masks. Since our dataset is derived from continuous capture videos, we annotate one out of every 10 images in each scene, resulting in a total of 260 annotated masks. Our analysis shows that the inpainting masks, on average, occupy only $50.78\\%$ of the area covered by the object masks. This finding highlights that adopting entire object masks requires inpainting nearly twice the necessary area, complicating the task and leading to suboptimal results.", + "bbox": [ + 89, + 641, + 482, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also compare our detected inpainting masks with the manually labeled masks using the mean Intersection over Union (mIoU) metric. Our method achieves an $81.12\\%$ IoU score, validating the accuracy and robustness of the pro", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a8815df0819bbef22fda360c24a5094f9c989ce7146efcc4c83261e70d62cecf.jpg", + "table_caption": [ + "Table 2. Quantitative results of ablation studies." + ], + "table_footnote": [], + "table_body": "
MethodPSNR (↑)LPIPS (↓)FID (↓)
w/o Warping17.850.3215198.24
w/o Refinement18.900.3069206.96
General Refinement19.080.2719165.80
Single-View Refinement19.460.2725154.33
Multi-View Refinement (Ours)19.670.2685149.52
", + "bbox": [ + 516, + 522, + 903, + 614 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "posed inpainting mask detection strategy. In contrast, the method proposed in [5] achieves only $42.55\\%$ , which is significantly inferior to our approach.", + "bbox": [ + 511, + 626, + 906, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 680, + 666, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our dataset is chosen for the ablation study due to its diversity and complexity.", + "bbox": [ + 511, + 704, + 903, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of geometry-guided warping: With image rendering of the pruned scene and inpainting masks, we adopt a warping-refinement scheme rather than directly inpainting the masked areas. To demonstrate the effectiveness of our approach, we disable geometry reconstruction and warping operations, repurposing the multi-view refinement model as a multi-view inpainting model. Specifically, we zero out the masked regions in the input images and attempt to consistently inpaint the missing areas. The inpainting model is also fine-tuned for each given scene, with all other settings remaining unchanged.", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "26700", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare images generated with and without the warping process in Fig. 6. Despite the strength of multiview architecture, direct inpainting fails to generate consistent output with the reference image in large view-change cases, which aligns with the observation in [2]. In contrast, our warping-refinement scheme provides a significant advantage, as the warped content offers strong guidance and regularizes the model to generate consistent results. We then use the generated images to fine-tune 3D scene. As shown in Tab. 2, the fine-tuned scene indeed shows inferior results to our complete method.", + "bbox": [ + 89, + 90, + 483, + 257 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/31be7770edf282eb21ab19ed4c49a286db472285d9d90a9d146ff167fa8fc12c.jpg", + "image_caption": [ + "(a) Reference Image" + ], + "image_footnote": [], + "bbox": [ + 96, + 270, + 222, + 367 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/786e7dac24072b7b9e9074c2f48bc737797f3ead5c09d3167dd3ac1a133333f2.jpg", + "image_caption": [ + "(b) W/o warping" + ], + "image_footnote": [], + "bbox": [ + 222, + 270, + 346, + 367 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fb0dd3e8044c742661fe54d5d481ec128a757f18579a0a7d0120ede43df40e9a.jpg", + "image_caption": [ + "(c) W/ warping", + "Figure 6. Multi-view refinement with and without conditioning on the warped image. (a) shows the inpainted reference with the originally captured image shown in the top right corner. (c) and (b) are the images from an alternate view generated with and without having the warped image as a condition, respectively." + ], + "image_footnote": [], + "bbox": [ + 346, + 270, + 473, + 367 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ce6aae3bb56c4389b91fd14d9a46bdd521928b7792f7aa49fcdcb6019b69f5a9.jpg", + "image_caption": [ + "(a) Reference Image" + ], + "image_footnote": [], + "bbox": [ + 99, + 483, + 223, + 578 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/814bec8e5c19ce142250ffa0aab31261da8dc5b3dc45571a1dcd9b60190fe916.jpg", + "image_caption": [ + "(b) W/o refine" + ], + "image_footnote": [], + "bbox": [ + 225, + 483, + 348, + 578 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cbc5d993772d0903e028951f23ad5d32877d26246c8252bc227f1bcdad516912.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 349, + 483, + 473, + 578 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/023687d7fe1395b1155c79244c6b5d27823535b42f1100a1f0fb33cc6b56721d.jpg", + "image_caption": [ + "(d) Reference image", + "Figure 7. Ablation on multi-view refinement. (a) and (d) show the inpainted reference images with the originally captured images shown in the top right corner; (b) and (c) are the images from an alternate view generated without and with the refinement network, respectively; (e) and (f) are the images generated by the single-view and multi-view refinement model, respectively." + ], + "image_footnote": [], + "bbox": [ + 96, + 593, + 223, + 688 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/83eb1c12036272c5fd582894d9047e73830a5a494243daea5d29337264102b98.jpg", + "image_caption": [ + "(e) Single-view refine" + ], + "image_footnote": [], + "bbox": [ + 225, + 593, + 346, + 688 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5784a93a682cefca5911ade79761263de7a46da5608ad1199fafc847b5350a34.jpg", + "image_caption": [ + "(c) Multi-view refine", + "(f) Multi-view refine" + ], + "image_footnote": [], + "bbox": [ + 349, + 593, + 470, + 688 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of multi-view refinement model: We design the following two variants to illustrate the importance of the multi-view refinement model. a) w/o refinement: the refinement model is removed, and the warped images are directly used to fine-tune the pruned scene. b) single view: we replace the space-time attention with the original self-attention layer, performing refinement independently for", + "bbox": [ + 89, + 794, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "each view without leveraging correlations across different views. When fine-tuning the single-view model for a given scene, only random image pairs are used.", + "bbox": [ + 511, + 90, + 905, + 136 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Similarly, we compare images generated by these variants in Fig. 7. As seen in (a), (b), and (c), warping indeed introduces artifacts and texture mismatches, while the multiview refinement model effectively addresses them. In (d), (e), and (f), the single-view refinement model fails to generate results consistent with the reference image, whereas the multi-view model achieves consistency. We further fine-tune the 3D scene with corresponding images. As illustrated in Tab. 2, our completed method achieves the best result.", + "bbox": [ + 511, + 137, + 905, + 271 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Necessity of test-time adaption: Our pipeline requires a per-scene fine-tuned refinement model to generate images. To demonstrate its necessity, we also train a generalized refinement model on the DI3DV dataset, a scene-level multiview dataset, where we randomly selected two views as the reference and target, synthesized warping artifacts on one, and trained the model to restore. However, as shown in Tab. 2, its performance is inferior to our proposed method due to the limited size of the training set and domain gaps between the training set and real-world cases, such as differences in the severity of artifacts/mismatch. In addition, we use an existing inpainting method, LeftRefill [1], to directly inpaint images based on the reference image. However, as shown in Fig.8, LeftRefill can only ensures overall visual similarity but lacks pixel-level consistency.", + "bbox": [ + 511, + 273, + 906, + 500 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/870c58f54a8262529f9431a97220e7520dad64a5d07280ab999bae359cf02f09.jpg", + "image_caption": [ + "(a) Reference", + "Figure 8. Inpainting results by LeftRefill." + ], + "image_footnote": [], + "bbox": [ + 522, + 513, + 645, + 569 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4a4cd5a62f73ec830ab653a961ee92f71c33b28911ef2781319a1231865d1d72.jpg", + "image_caption": [ + "(b) 1st View" + ], + "image_footnote": [], + "bbox": [ + 648, + 513, + 772, + 569 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/330676a2728fdf714710c5f254163437544d0ae9e578ceeab4b84dd2bd8e803a.jpg", + "image_caption": [ + "(c) 2nd View" + ], + "image_footnote": [], + "bbox": [ + 772, + 513, + 897, + 569 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion and Limitation", + "text_level": 1, + "bbox": [ + 513, + 612, + 763, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present a novel 3D inpainting pipeline that achieves state-of-the-art results across both front-facing and unconstrained benchmarks, with particularly strong performance on the latter. We achieve this by utilizing geometry priors and a multi-view refinement network to generate multi-view consistent inpainted images. Additionally, a novel inpainting mask detection technique is developed to boost performance in unconstrained scenes. The effectiveness of our method is demonstrated through extensive experiments and ablation studies.", + "bbox": [ + 511, + 638, + 905, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations: The multi-view refinement network requires customization for each scene, which may take considerable time. Future advancements in parameter-efficient fine-tuning and the personalization of diffusion models may help to accelerate and further optimize our approach. In addition, our framework relies on monocular depth estimation and image segmentation modules. The accuracy and robustness of these modules can affect the overall performance.", + "bbox": [ + 511, + 789, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "26701", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Chenjie Cao, Yunuo Cai, Qiaole Dong, Yikai Wang, and Yanwei Fu. Leftrightfill: Filling right canvas based on left reference through generalized text-to-image diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7705-7715, 2024.", + "[2] Chenjie Cao, Chaohui Yu, Yanwei Fu, Fan Wang, and Xi-angyang Xue. Mvinpainter: Learning multi-view consistent inpainting to bridge 2d and 3d editing. arXiv preprint arXiv:2408.08000, 2024.", + "[3] Honghua Chen, Chen Change Loy, and Xingang Pan. Mvipnerf: Multi-view 3d inpainting on nerf scenes via diffusion prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5344-5353, 2024.", + "[4] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. Anydoor: Zero-shot object-level image customization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6593-6602, 2024.", + "[5] Yiwen Chen, Zilong Chen, Chi Zhang, Feng Wang, Xiaofeng Yang, Yikai Wang, Zhongang Cai, Lei Yang, Huaping Liu, and Guosheng Lin. Gaussianeditor: Swift and controllable 3d editing with gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21476-21485, 2024.", + "[6] Brian Curless and Marc Levoy. A volumetric method for building complex models from range images. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 303-312, 1996.", + "[7] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022.", + "[8] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020.", + "[9] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022.", + "[10] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022.", + "[11] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024.", + "[12] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023.", + "[13] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023." + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Chieh Hubert Lin, Changil Kim, Jia-Bin Huang, Qinbo Li, Chih-Yao Ma, Johannes Kopf, Ming-Hsuan Yang, and Hung-Yu Tseng. Taming latent diffusion model for neural radiance field inpainting. In European Conference on Computer Vision, pages 149-165. Springer, 2025.", + "[15] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453, 2023.", + "[16] Ashkan Mirzaei, Tristan Aumentado-Armstrong, Marcus A Brubaker, Jonathan Kelly, Alex Levinshtein, Konstantinos G Derpanis, and Igor Gilitschenski. Reference-guided controllable inpainting of neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17815-17825, 2023.", + "[17] Ashkan Mirzaei, Tristan Aumentado-Armstrong, Konstantinos G Derpanis, Jonathan Kelly, Marcus A Brubaker, Igor Gilitschenski, and Alex Levinshtein. Spin-nerf: Multiview segmentation and perceptual inpainting with neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20669–20679, 2023.", + "[18] Ashkan Mirzaei, Riccardo De Lutio, Seung Wook Kim, David Acuna, Jonathan Kelly, Sanja Fidler, Igor Gilitschenski, and Zan Gojcic. Refusion: Reference adapted diffusion models for 3d scene inpainting. arXiv preprint arXiv:2404.10765, 2024.", + "[19] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4296–4304, 2024.", + "[20] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022.", + "[21] Christian Reiser, Rick Szeliski, Dor Verbin, Pratul Srinivasan, Ben Mildenhall, Andreas Geiger, Jon Barron, and Peter Hedman. Merf: Memory-efficient radiance fields for real-time view synthesis in unbounded scenes. ACM Transactions on Graphics (TOG), 42(4):1-12, 2023.", + "[22] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022.", + "[23] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023.", + "[24] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123: a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110, 2023.", + "[25] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023." + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "26702", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022.", + "[27] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.", + "[28] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2149–2159, 2022.", + "[29] Luming Tang, Nataniel Ruiz, Qinghao Chu, Yuanzhen Li, Aleksander Holynski, David E Jacobs, Bharath Hariharan, Yael Pritch, Neal Wadhwa, Kfir Aberman, et al. Realfill: Reference-driven generation for authentic image completion. ACM Transactions on Graphics (TOG), 43(4):1-12, 2024.", + "[30] Dongqing Wang, Tong Zhang, Alaa Abboud, and Sabine Susstrunk. Innerf360: Text-guided 3d-consistent object inpainting on 360-degree neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12677-12686, 2024.", + "[31] Peng Wang and Yichun Shi. Imagedream: Image-prompt multi-view diffusion for 3d generation. arXiv preprint arXiv:2312.02201, 2023.", + "[32] Yuxin Wang, Qianyi Wu, Guofeng Zhang, and Dan Xu. Gscream: Learning 3d geometry and feature consistent gaussian splatting for object removal. arXiv preprint arXiv:2404.13679, 2024.", + "[33] Ethan Weber, Aleksander Holynski, Varun Jampani, Saurabh Saxena, Noah Snavely, Abhishek Kar, and Angjoo Kanazawa. Nerfiller: Completing scenes via generative 3d inpainting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20731-20741, 2024.", + "[34] Silvan Weder, Guillermo Garcia-Hernando, Aron Monsz-part, Marc Pollefeys, Gabriel J Brostow, Michael Firman, and Sara Vicente. Removing objects from neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16528-16538, 2023.", + "[35] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7623-7633, 2023.", + "[36] Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen, and Fang Wen. Paint by example: Exemplar-based image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18381-18391, 2023.", + "[37] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In CVPR, 2024.", + "[38] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023.", + "[39] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023.", + "[40] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022." + ], + "bbox": [ + 516, + 92, + 903, + 287 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "26703", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/IMFine_ 3D Inpainting via Geometry-guided Multi-view Refinement/0becb03e-65d1-4a50-a09c-58afb744258c_model.json b/2025/IMFine_ 3D Inpainting via Geometry-guided Multi-view Refinement/0becb03e-65d1-4a50-a09c-58afb744258c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..dd68f6dd1991a390e46f553084259d3e7cc1a9ae --- /dev/null +++ b/2025/IMFine_ 3D Inpainting via Geometry-guided Multi-view Refinement/0becb03e-65d1-4a50-a09c-58afb744258c_model.json @@ -0,0 +1,2673 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.155, + 0.131, + 0.844, + 0.154 + ], + "angle": 0, + "content": "IMFine: 3D Inpainting via Geometry-guided Multi-view Refinement" + }, + { + "type": "text", + "bbox": [ + 0.21, + 0.18, + 0.317, + 0.198 + ], + "angle": 0, + "content": "Zhihao Shi\\*1" + }, + { + "type": "text", + "bbox": [ + 0.362, + 0.181, + 0.47, + 0.199 + ], + "angle": 0, + "content": "Dong Huo\\*o2" + }, + { + "type": "text", + "bbox": [ + 0.376, + 0.2, + 0.458, + 0.215 + ], + "angle": 0, + "content": "Juwei Lu1" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.181, + 0.648, + 0.199 + ], + "angle": 0, + "content": "Yuhongze Zhou" + }, + { + "type": "text", + "bbox": [ + 0.701, + 0.181, + 0.786, + 0.198 + ], + "angle": 0, + "content": "Yan Min" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.199, + 0.621, + 0.216 + ], + "angle": 0, + "content": "Xinxin Zuo4" + }, + { + "type": "text", + "bbox": [ + 0.248, + 0.217, + 0.528, + 0.232 + ], + "angle": 0, + "content": "1Huawei Canada Research Institute" + }, + { + "type": "text", + "bbox": [ + 0.574, + 0.217, + 0.75, + 0.234 + ], + "angle": 0, + "content": "\\(^{2}\\)University of Alberta" + }, + { + "type": "text", + "bbox": [ + 0.298, + 0.233, + 0.476, + 0.253 + ], + "angle": 0, + "content": "\\(^{3}\\)McMaster University" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.234, + 0.7, + 0.253 + ], + "angle": 0, + "content": "4Concordia University" + }, + { + "type": "text", + "bbox": [ + 0.185, + 0.255, + 0.822, + 0.269 + ], + "angle": 0, + "content": "zhihaoshi2022@gmail.com, dhuo@ualberta.ca, yuhongze.zhou@mail.mcgill.com" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.273, + 0.785, + 0.287 + ], + "angle": 0, + "content": "miny13@mcmaster.ca, juwei.lu@huawei.com, xinxin.zuo@concordia.ca" + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.322, + 0.226, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.424, + 0.223, + 0.524 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.358, + 0.262, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.263, + 0.323, + 0.517, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.323, + 0.777, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.779, + 0.323, + 0.905, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.263, + 0.423, + 0.517, + 0.524 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.423, + 0.777, + 0.524 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.535, + 0.908, + 0.579 + ], + "angle": 0, + "content": "Figure 1. We present a novel 3D inpainting method capable of removing target objects from the 3D scenes and seamlessly inpainting the occluded regions across various scenarios. The proposed method ensures the generation of plausible geometry and consistent textures. It is highly recommended to see video results in our project page." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.591, + 0.327, + 0.607 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.623, + 0.486, + 0.866 + ], + "angle": 0, + "content": "Current 3D inpainting and object removal methods are largely limited to front-facing scenes, facing substantial challenges when applied to diverse, \"unconstrained\" scenes where the camera orientation and trajectory are unrestricted. To bridge this gap, we introduce a novel approach that produces inpainted 3D scenes with consistent visual quality and coherent underlying geometry across both front-facing and unconstrained scenes. Specifically, we propose a robust 3D inpainting pipeline that incorporates geometric priors and a multi-view refinement network trained via test-time adaptation, building on a pretrained image inpainting model. Additionally, we develop a novel inpainting mask detection technique to derive targeted inpainting masks from object masks, boosting the performance in handling unconstrained scenes. To validate the efficacy of our approach, we create a challenging" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.593, + 0.907, + 0.655 + ], + "angle": 0, + "content": "and diverse benchmark that spans a wide range of scenes. Comprehensive experiments demonstrate that our proposed method substantially outperforms existing state-of-the-art approaches." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.68, + 0.645, + 0.695 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.704, + 0.907, + 0.825 + ], + "angle": 0, + "content": "3D editing is an active research area with significant potential for various applications, including video games, the Metaverse, virtual reality, and augmented reality. For example, after a 3D scene is created or reconstructed, a user might want to remove certain objects from the scene. This paper mainly focuses on the challenging task of inpainting in 3D scenes to fill the occluded regions to enable seamless object removal." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Compared to image inpainting [22, 28], 3D inpainting models lag significantly behind due to the lack of large-scale 3D datasets and efficient 3D data structures. Therefore, existing 3D inpainting methods mainly focus on lifting 2D inpainting priors to 3D. Although the current diffusion" + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.876, + 0.249, + 0.888 + ], + "angle": 0, + "content": "* denotes equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.093, + 0.889, + 0.469, + 0.9 + ], + "angle": 0, + "content": "Work done during an internship at Huawei Canada Research Institute" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.876, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26694" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "based image inpainting models can generate images with vivid textures and high-fidelity details, they are trained on single images without considering any 3D consistency. It poses great challenges in maintaining view consistency and preserving 3D coherence when lifting those 2D priors to 3D." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.187, + 0.487, + 0.626 + ], + "angle": 0, + "content": "To address the inconsistency issue, there are approaches that implicitly lift the 2D models into 3D via score distillation sampling (SDS) loss [3, 18], which was proposed initially in text-to-3D generation [20], to harness the inherent coherence of 2D diffusion models under the same prompt, complemented by the iterative updating scheme during optimization to ensure view consistency. However, they are constrained by the inherent limitation of SDS loss, which results in over-saturation and over-smoothing in the optimized images, even with different variants of SDS losses. Alternatively, explicit methods [2, 5, 16, 17, 30] first generate one or few images with the image inpainting models and try to preserve the 3D consistency by enforcing geometry constraints or minimizing perceptual losses between rendered images and the inpainted 2D images. Although they have demonstrated good performance, especially for front-facing scenes with very limited view changes, the inpainted 3D scenes have artifacts such as floaters and blurry textures in unconstrained scenarios with complex geometry and large view variation. The reasons are that first, 3D geometry is rather difficult to get perfectly restored; second, perceptual loss can only alleviate the inconsistency by minimizing overall similarity. Inconsistency still exists between those inpainted images, leading to blurry textures for the inpainted 3D scenes. More recently, [2] proposed to train a multi-view inpainting network to generate consistent multiview inpainted images. However, due to the limited diversity of current multi-view datasets, they cannot generalize to unconstrained 3D scenes (e.g., \\(360^{\\circ}\\) cases)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.487, + 0.901 + ], + "angle": 0, + "content": "To enhance the ability to generalize to various scenes for 3D inpainting, we utilize geometry priors and develop a multi-view refinement network fine-tuned at test time to produce consistent images in a coarse-to-fine manner. Specifically, we start with a single reference view inpainted with image inpainting models, followed by coarse geometry restoration and a warping operation to warp the inpainted content from the reference view to others. However, due to the imperfect reconstructed geometry and view-dependent effects, the warped images tend to have artifacts, such as distorted textures and mismatches between the warped and neighboring contents. To refine the warped contents, we present a multi-view refinement model to generate consistent images with high fidelity, taking the warped images as a condition. The key idea is that instead of training a model with extensive datasets, we take the pre-trained image inpainting model, add space-time attention layers, and fine-tune them with the current test scenes. We take the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "originally captured multi-view images and implement various augmentations to synthesize paired data to fine-tune the multi-view refinement model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.138, + 0.907, + 0.289 + ], + "angle": 0, + "content": "Furthermore, previous works usually take the entire object masks as inpainted regions. However, some regions inside the object masks might have already been exposed in other views, especially for 3D object removal and inpainting in \\(360^{\\circ}\\) scenes (see Fig. 4). Therefore, we believe it is critical to differentiate between object masks and inpainting masks, fully exploiting the available information of the reconstructed 3D scenes. To this end, a novel inpainting mask detection approach is proposed to extract masks accurately and exclusively for real occluded regions." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.29, + 0.909, + 0.38 + ], + "angle": 0, + "content": "Finally, considering that the existing benchmarks mostly cover front-facing scenes or planar surfaces, we introduce a new dataset encompassing 20 diverse scenes, from indoor to outdoor environments, covering large view variations and complex interfacial surfaces, to better evaluate and advance 3D inpainting methods." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.382, + 0.815, + 0.396 + ], + "angle": 0, + "content": "In summary, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.398, + 0.905, + 0.473 + ], + "angle": 0, + "content": "1. A multi-view consistent inpainting pipeline is proposed. We leverage geometric priors and a multi-view refinement model that is fine-tuned from image inpainting models via test-time adaptation to generate consistent multi-view images." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.473, + 0.905, + 0.532 + ], + "angle": 0, + "content": "2. We further develop an inpainting mask detection technique that accurately distinguishes occluded regions, enabling robust processing of complex scenes, such as \\(360^{\\circ}\\) environments." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.533, + 0.905, + 0.579 + ], + "angle": 0, + "content": "3. The proposed inpainting method has achieved state-of-the-art performance across various benchmarks, particularly excelling in challenging, unconstrained scenarios." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.579, + 0.905, + 0.638 + ], + "angle": 0, + "content": "4. We contribute a new and challenging 3D scene inpainting benchmark, offering diverse scenes to facilitate broader evaluation and advancement in 3D inpainting research." + }, + { + "type": "list", + "bbox": [ + 0.515, + 0.398, + 0.905, + 0.638 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.655, + 0.657, + 0.67 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.681, + 0.66, + 0.697 + ], + "angle": 0, + "content": "2.1. 3D Inpainting" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Existing 3D inpainting methods [3, 5, 16-18, 30, 32] are implemented by lifting 2D inpainting priors to 3D with various strategies designed to ensure 3D consistency. One direction, inspired by score distillation-based approaches [20] in text-to-3D generation, formulates 3D inpainting as a continuous distillation process [3, 18]. While these approaches achieve reasonably consistent inpainting results, they face inherent limitations of score distillation, such as producing over-saturated or overly smooth outputs. Another direction explicitly enforces 3D consistency by independently inpainting one or multiple 2D views and then consolidating multi-view inconsistencies in 3D. Reference-guided NeRF [16] addresses this challenge by using a single view to" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26695" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.288 + ], + "angle": 0, + "content": "guide geometry inpainting to avoid inconsistencies. However, it struggles with accurate geometry capture in complex \\(360^{\\circ}\\) scenes using just a single view. SPINeRF [17], and InNeRF360 [30] utilize multiple inpainted images to guide 3D inpainting and incorporate perceptual loss to reduce inconsistencies resulting from independently inpainted images. Nevertheless, while perceptual loss can maintain overall consistency, it performs poorly on finer details, often leading to blurry 3D results. Our method lies in the latter direction. It addresses these issues through a geometry-guided refinement network that generates a set of multi-view consistent images, reconstructing clearer and consistent 3D inpainted scenes." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.299, + 0.373, + 0.314 + ], + "angle": 0, + "content": "2.2. Video and Multi-View Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.323, + 0.484, + 0.457 + ], + "angle": 0, + "content": "The diffusion model for text-to-image (T2I) generation [7, 8, 22, 27] has achieved significant success in recent years, known for its ability to produce high-quality, diverse, and controllable outputs. Building on this, extensive studies have extended image diffusion models to the video generation domain [9, 10, 26, 40] by incorporating temporal layers within each transformer block to perform space-time attention, which captures correlations across frames and shows promising results in generating realistic videos." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.459, + 0.484, + 0.715 + ], + "angle": 0, + "content": "Inspired by these advances in video generation, researchers have begun exploring the adaptation of T2I models into multi-view diffusion models for implicit 3D reasoning [15, 24, 25, 31]. For example, methods such as MVDream [25] and Zero \\(123++\\) [24] leverage this idea and derive multi-view diffusion models, capable of generating consistent multiple views simultaneously for 3D generation. However, the full space-time attention mechanism used in these methods can lead to prohibitive memory and computational costs, making it impractical for handling large numbers of views. To address this, recent work [35] introduces a sparse space-time attention mechanism, which reduces memory requirements by attending only to the initial and previous frames. However, the sparse attention approach assumes input frames are from a continuous video sequence and remains under-explored in sparse multi-view scenarios." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.726, + 0.347, + 0.743 + ], + "angle": 0, + "content": "2.3. Diffusion Model Fine-tuning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Adapting large pre-trained diffusion models has proven to be an effective and efficient approach for completing user-specified tasks, such as synthesizing specific subjects in new contexts [4, 23] and inpainting images with instance matching reference images [29, 36]. Adaptation methods include injecting extra layers [19, 36, 38, 39], searching appropriate unique prompt [23], updating partial or full model weights [14, 18, 29], or combinations of these techniques. Similarly, in this paper, instead of relying on extensive multiview datasets, we propose adapting the pre-trained image" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.122 + ], + "angle": 0, + "content": "diffusion model into a multi-view refinement model for 3D inpainting purposes." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.135, + 0.64, + 0.153 + ], + "angle": 0, + "content": "3. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.161, + 0.907, + 0.298 + ], + "angle": 0, + "content": "Gaussian splating (GS) methods [11, 12] represent a scene as a collection of multivariate Gaussian primitives: \\(\\mathcal{G} = \\{(\\mu_m,\\Sigma_m,\\Gamma_m,\\mathrm{SH}_m)\\}_{m = 1}^M\\), where \\(\\mu_{m},\\Sigma_{m},\\Gamma_{m}\\) and \\(\\mathrm{SH}_m\\) denote the mean vector, covariance matrix, opacity, and spherical harmonic coefficients of a multivariate Gaussian, respectively. View synthesis is performed by standard point-based \\(\\alpha^\\prime\\)-rendering, where a single pixel color is computed by sorting \\(\\mathcal{G}\\) (or their 2D projection) in their depth order and alpha-blending each contributed primitive:" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.309, + 0.907, + 0.351 + ], + "angle": 0, + "content": "\\[\nC = \\sum_ {m} c _ {m} \\alpha_ {m} \\prod_ {j = 1} ^ {m - 1} (1 - \\alpha_ {j}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.361, + 0.906, + 0.452 + ], + "angle": 0, + "content": "where \\(C\\) is the output pixel color, \\(c_{m}\\) is the pre-computed color of the \\(m\\)-th primitive, and \\(\\alpha_{m}\\) is a calculated opacity based on distance from the Gaussian center to the interaction point. Due to the excellent performance in geometry reconstruction, in this work, we choose 2D-GS [11] as the 3D representation." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.465, + 0.605, + 0.481 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.491, + 0.907, + 0.733 + ], + "angle": 0, + "content": "Given a GS scene \\(\\mathcal{G}\\), along with corresponding training images \\(\\{I_n\\}_{n=1}^N\\), associated camera poses \\(\\{\\Pi_n\\}_{n=1}^N\\), and a set of object masks \\(\\{M_n\\}_{n=1}^N\\) identifying objects to be removed, our goal is to generate an inpainted scene \\(\\mathcal{G}^*\\) where the target objects are removed and occluded regions are filled seamlessly with the background. The entire pipeline contains two stages, namely object removal and 3D scene inpainting. We first remove object primitives (Sec. 4.1), exposing never-before-seen (NBS) regions which are the areas that truly need inpainting. We then either use the object masks as inpainting masks when they are identical or detect specific inpainting masks (Sec. 4.3), ensuring that the inpainting mask exclusively covers the NBS regions in each view, denoted as \\(\\{P_n\\}_{n=1}^N\\). Given the pruned GS scene and inpainting masks, we propose a novel 3D inpainting pipeline to inpaint \\(\\mathcal{G}\\) as \\(\\mathcal{G}^*\\) (Sec. 4.2)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.741, + 0.673, + 0.758 + ], + "angle": 0, + "content": "4.1. Object Removal" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.903 + ], + "angle": 0, + "content": "To remove the objects from the reconstructed GS scene, we append an additional learnable parameter \\(\\{p_m\\}_{m=1}^M \\in [0,1]\\) to each Gaussian primitive, indicating whether it belongs to the object to be removed, which is initialized as 0.5. To compute the value of \\(p_m\\) for each Gaussian, we map the object masks back to the Gaussian primitives. Specifically, we render a set of object masks \\(\\{\\hat{M}\\}_{n=1}^N\\) using \\(\\alpha'\\) rendering (similar to color rendering) at \\(\\{\\Pi_n\\}_{n=1}^N\\) but replacing \\(c_m\\) with \\(p_m\\) in Eq. 1, followed by \\(K_1\\) (\\(K_1 = 1000\\)) steps of the" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26696" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.206, + 0.089, + 0.437, + 0.105 + ], + "angle": 0, + "content": "(a) Multi-view Inpainting Pipeline" + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.108, + 0.531, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.638, + 0.089, + 0.804, + 0.105 + ], + "angle": 0, + "content": "(b) Test-time Adaptation" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.11, + 0.905, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.317, + 0.907, + 0.429 + ], + "angle": 0, + "content": "Figure 2. Overview of our proposed approach. (a) shows the proposed 3D inpainting pipeline. Specifically, a reference view is first inpainted, which will be used to reconstruct complete geometry, along with rendered multi-view depth maps. The inpainted reference image is then warped to other views, which are taken as input conditions for the multi-view refinement network to generate a set of view-consistent images. Afterward, we fine-tune the 3D scene to get the inpainted scene. (b) illustrates the detailed structure and training of the multi-view refinement network. Basically, we take the pre-trained image inpainting model, incorporate additional space-time attention layers, and exploit the test-time adaptation to fine-tune it into a multi-view refinement model. To generate a fine-tuning dataset, we synthesize warped images under the original captured scene, which has the ground-truth captured images. In detail, reference views are selected randomly, and we set random masks over the images, apply geometry jittering, and warp the reference images in other views." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.437, + 0.484, + 0.514 + ], + "angle": 0, + "content": "optimization to update \\( p_m \\) with respect to \\( \\mathcal{L}_1 \\) loss between \\( \\{\\hat{M}\\}_{n=1}^N \\) and \\( \\{M\\}_{n=1}^N \\). Notably, other parameters remain unchanged during this procedure. Finally, a pruning operation is performed to produce the pruned scene \\( \\mathcal{G}_{\\mathrm{pruned}} \\) by thresholding \\( p_m \\) with a threshold \\( \\tau \\) (\\( \\tau = 0.4 \\))." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.524, + 0.378, + 0.541 + ], + "angle": 0, + "content": "4.2. Proposed 3D Inpainting Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.549, + 0.484, + 0.791 + ], + "angle": 0, + "content": "Given the pruned GS scene \\(\\mathcal{G}_{\\mathrm{pruned}}\\) as well as inpainting masks \\(\\{P_n\\}_{n=1}^N\\) indicating the regions to be inpainted for each view, we propose a novel 3D inpainting pipeline shown in Fig 2. We first render color images \\(\\{I_n'\\}_{n=1}^N\\) from \\(\\mathcal{G}_{\\mathrm{pruned}}\\) for each view and select a reference view with image \\(I_r' \\in \\{I_n'\\}_{n=1}^N\\), which is then inpainted using a pre-trained inpainting model [22] to reference image \\(I_r^*\\). We conduct geometry reconstruction to fill out the hole in 3D, which will be utilized to warp the inpainted contents in the reference image into the remaining views. We employ a multi-view refinement model to further refine the warped images, ensuring high fidelity and consistency, and then update the 3D scene. The model is fine-tuned from the pre-trained image inpainting model via a test-time adaptation strategy. The geometry reconstruction and multi-view refinement model are described as follows." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.816, + 0.325, + 0.831 + ], + "angle": 0, + "content": "4.2.1 Geometry Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Following [11], we render a set of depth maps \\(\\{D_n\\}_{n=1}^N\\) from the prune GS scene \\(\\mathcal{G}_{\\mathrm{pruned}}\\) and adopt a truncated signed distance function (TSDF) [6] technique to reconstruct the 3D mesh. However, \\(\\{D_n\\}_{n=1}^N\\) only contributes" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.437, + 0.907, + 0.558 + ], + "angle": 0, + "content": "to the reconstruction of the background region, as the depth values within their inpainting mask areas are invalid. To fill the masked region, we aim to generate an all-valid depth map \\( D_r^* \\) of the reference view by filling the masked areas in \\( D_r \\) based on information from \\( I_r^* \\). The filled depth map should meet two key requirements: 1) it must be consistent with the reference image \\( I_r^* \\); 2) it should seamlessly connect with the depth values outside the masked region." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.566, + 0.909, + 0.687 + ], + "angle": 0, + "content": "We begin by exploiting a monocular depth estimator [37] to estimate depth \\(\\bar{D}_r\\) from \\(I_r^*\\), followed by a linear transformation to convert \\(\\bar{D}_r\\) to scene scale \\(\\tilde{D}_r\\): \\(\\tilde{D}_r \\colon \\tilde{D}_r = a * \\bar{D}_r + b\\), where the parameters \\(a, b\\) are computed in the neighborhood of \\(P_r\\) to minimize the difference between \\(\\bar{D}_r\\) and \\(D_r\\) in advance, following [16]. We then develop a filling scheme according to \\(\\tilde{D}_r\\). Formally, the filled values of \\(D_r^*\\) are computed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.714, + 0.91, + 0.782 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} D _ {r} ^ {*} (x) = \\underset {D _ {r} ^ {*} (x)} {\\arg \\min } \\sum_ {x \\in P _ {r}} | | \\nabla D _ {r} ^ {*} (x) - \\nabla \\tilde {D} _ {r} (x) | | ^ {2} + \\lambda \\nabla D _ {r} ^ {*} (x) \\tag {2} \\\\ \\begin{array}{l} \\text {s . t .} D _ {r} ^ {*} (x) = D _ {r} (x), \\text {w h e r e} P _ {r} (x) = 0, \\end{array} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.907, + 0.901 + ], + "angle": 0, + "content": "where \\(\\nabla\\) denotes the Laplacian calculation and \\(x\\) represents the pixel coordinate. Intuitively, the first term encourages \\(D_r^*\\) to maintain a similar changing trend as \\(\\tilde{D}_r\\), while the second term regularizes \\(D_r^*\\) to be smooth and consistent with the depths outside the mask. As such, \\(D_r^*\\) can be used for the reconstruction of the NBS region." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26697" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.089, + 0.223, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.106, + 0.189, + 0.215, + 0.201 + ], + "angle": 0, + "content": "(a) Reference Image" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.089, + 0.348, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.238, + 0.189, + 0.334, + 0.201 + ], + "angle": 0, + "content": "(b) Warped Image" + }, + { + "type": "image", + "bbox": [ + 0.349, + 0.089, + 0.474, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.363, + 0.189, + 0.46, + 0.201 + ], + "angle": 0, + "content": "(c) Refined Image" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.202, + 0.483, + 0.271 + ], + "angle": 0, + "content": "Figure 3. Illustration of the importance of multi-view refinement. The multi-view refinement model is proposed to tackle the warping artifacts with the example shown in (b), generating high-fidelity multi-view consistent images with one sampled view shown in (c)." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.284, + 0.483, + 0.299 + ], + "angle": 0, + "content": "4.2.2 Test-Time Adaptation for Multi-view Refinement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.31, + 0.484, + 0.49 + ], + "angle": 0, + "content": "We warp the inpainted contents from the reference view \\( I_{r} \\) to other views with the above restored geometry. Due to the imperfect geometry and the neglect of view-dependent effects, those warped contents contain artifacts, such as texture distortion and inconsistent appearance with the neighboring regions, especially when there are large view angle changes with respect to the reference view as shown in Fig 3. On the other hand, these warped contents have coarse-level consistency across views. To this end, we propose a multi-view refinement model that produces high-fidelity, consistent images conditioning on the warped images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.492, + 0.484, + 0.688 + ], + "angle": 0, + "content": "Network: Our model is built upon a pre-trained diffusion-based image inpainting model [22], which takes a masked image and its associated mask as the input, corresponding to the warped image and inpainting mask in our scenarios, respectively. To ensure multi-view consistency, we replace the self-attention layer in each transformer block with space-time attention to explore correlations across multiple views. Considering the quadratic complexity of global attention and the large number of warped images, space-time attention is implemented as the sparse space-time attention [10] that only attends to neighboring images. The attention layer is also allowed to attend to the reference image for alignment." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Training: Instead of relying on large datasets and extensive training, we exploit the test-time adaptation strategy for each scene to train the multi-view refinement network. Paired training sets are created for model adaptation. Specifically, given the original images \\(\\{I_n\\}_{n=1}^N\\), at each training step, \\(L + 1\\) images are randomly selected as ground truth. To simulate realistic warping scenarios, we randomly choose one image as the reference, generate a random mask, and warp both the reference image and mask to other views with geometry guidance from \\(\\mathcal{G}\\). In addition, to encourage cross-view learning as in [2], we also synthesize another set of paired images by applying image augmentations independently on each captured image, i.e., irregularly shaped masks that are randomly generated on" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.089, + 0.71, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.527, + 0.19, + 0.704, + 0.201 + ], + "angle": 0, + "content": "(a) Object mask and dilated mask" + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.089, + 0.901, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.733, + 0.19, + 0.879, + 0.201 + ], + "angle": 0, + "content": "(b) Object Mask Separation" + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.202, + 0.707, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.522, + 0.303, + 0.705, + 0.314 + ], + "angle": 0, + "content": "(c) Mask-remapping interpretation" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.202, + 0.899, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.756, + 0.303, + 0.852, + 0.313 + ], + "angle": 0, + "content": "(d) Render mask" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.317, + 0.907, + 0.456 + ], + "angle": 0, + "content": "Figure 4. Inpainting mask detection. (a) shows the object mask (orange) along with the dilated mask (pink), which includes additional neighboring pixels; Each dilated mask consists of two parts: the actual inpainting mask and out-of-interest areas, marked as regions \\(①\\) and \\(②\\) in (b). These regions are mapped in 3D to the NBS and background areas, respectively, as shown in regions \\(①\\) and \\(②\\) in (c). During multi-view mapping optimization, the central NBS region is progressively refined, while background areas are suppressed by each other; (d) shows the rendered neighbor mask after GS scene optimization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.472, + 0.906, + 0.547 + ], + "angle": 0, + "content": "each image. Image-based augmentations—such as elastic transformations and color jittering, are then applied to the masked regions to simulate warping artifacts. We train the multi-view refinement model with those paired images, employing a simplified variational bound objective [8]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.548, + 0.906, + 0.608 + ], + "angle": 0, + "content": "Inference: At inference, similar to [33], we randomly shuffle the order of the \\( L \\) warped images at each denosing step to break the sequential constraint of sparse attention, further enhancing cross-view consistency." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.63, + 0.706, + 0.643 + ], + "angle": 0, + "content": "4.2.3 Fine-tune GS Scene" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.655, + 0.906, + 0.745 + ], + "angle": 0, + "content": "With the generated multi-view consistent images, we fine-tune \\(\\mathcal{G}_{\\mathrm{pruned}}\\) for \\(K_{2}(K_{2} = 7000)\\) steps with an \\(L_{1}\\) loss and a SSIM loss as straightforward pixel-wise supervision. It is worth noting that other robust losses (e.g., perceptual loss) are functionally orthogonal to our approach and can be used to improve it further." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.756, + 0.756, + 0.772 + ], + "angle": 0, + "content": "4.3. Inpainting Mask Detection" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.906, + 0.885 + ], + "angle": 0, + "content": "In the overall pipeline of 3D object removal and scene inpainting, identifying the occluded or NBS regions for inpainting is also critical, especially for \\(360^{\\circ}\\) cases. In those cases, the object mask often covers a much larger area than the regions that require inpainting (as illustrated in Fig. 4). Therefore, a novel inpainting mask detection approach is proposed to identify the NBS region for each view." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.886, + 0.906, + 0.901 + ], + "angle": 0, + "content": "An intuitive idea is to detect the NBS region in the" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26698" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.41 + ], + "angle": 0, + "content": "pruned GS scene and project it back into each view to obtain inpainting masks. However, since, by definition, the NBS region is occluded by the object, it usually lacks good geometry and meaningful Gaussians, making it challenging to detect in 3D. Here we propose to detect the inpainting masks by leveraging neighbors of NBS region. In detail, we first dilate the object masks to include neighboring pixels (Fig. 4(a)), and then map the dilated masks onto the pruned scene \\(\\mathcal{G}_{\\mathrm{pruned}}\\). Mapping is implemented as updating the learnable parameter \\(\\{p_m\\}_{m=1}^M\\) of each Gaussian with respect to \\(\\mathcal{L}_1\\) loss between rendered masks of \\(\\mathcal{G}_{\\mathrm{pruned}}\\) and the dilated masks. The optimization process is the same as object removal (Sec. 4.1) but with the object mask replaced with dilated masks, and the optimization is performed over the pruned scene. We illustrate the interpretation of this process in Fig. 4(b-c). After the mapping process, we are able to get the neighbor mask by rendering \\(\\{p_m\\}_{m=1}^M\\) from \\(\\mathcal{G}_{\\mathrm{pruned}}\\) as shown in Fig. 4(d). Finally, we extract the bounding box for the above rendered mask, which will be taken as input prompt for segment anything model [13] to get the final inpainting mask." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.425, + 0.223, + 0.442 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.451, + 0.483, + 0.737 + ], + "angle": 0, + "content": "Datasets: Most existing publicly available 3D inpainting datasets [17, 34] are limited to front-facing cases, which cannot be used to evaluate methods in more diverse or realistic situations. To address this, we construct a dataset that contains 20 diverse cases, ranging from indoor to outdoor, covering various camera trajectories-including front-facing, \\(90^{\\circ}\\), \\(180^{\\circ}\\), and \\(360^{\\circ}\\), and containing various interfacial surfaces such as single plane, multiple planes, curvature surfaces, and irregular surfaces. Each scene is recorded using a DJI Pocket 3 at 60fps with a resolution of \\(1920 \\times 1080\\), from which we evenly extract 175 images and generate the corresponding camera poses and object masks. The first 50 images are used for evaluation purpose with the object removed, while the remaining 125 images are object-present and are used to reconstruct the original GS scene. Previews of our collected dataset are shown in supplementary materials. We use half resolution in the following experiments. In addition, we evaluate our method on SPINeRF dataset [17] to compare with state-of-the-art." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.739, + 0.483, + 0.875 + ], + "angle": 0, + "content": "Implementation Details: StableDiffusionInpainting-Pipeline is utilized for image inpainting and our multi-view refinement model initialization. During training, we randomly select \\( L = 5 \\) images from training images and crop \\( 512 \\times 512 \\) patches to fine-tune the model over 1000 iterations with the learning rate and batch size set as \\( 3 \\times 10^{-5} \\) and 1, individually. At inference, directly refining all warped images, i.e., \\( L = 125 \\) would result in prohibitive memory usage and computational costs. To" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.907, + 0.131 + ], + "angle": 0, + "content": "Table 1. Quantitative evaluation on the SPINeRF dataset and our dataset. Numbers in bold indicate the best performance, and underscored numbers indicate the second best." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.132, + 0.905, + 0.214 + ], + "angle": 0, + "content": "
MethodOursSPINeRF
PSNR (↑)LPIPS (↓)FID (↓)PSNR (↑)LPIPS (↓)FID (↓)
GaussianEditor15.710.6163375.0314.410.6247343.16
GScream17.180.4431290.6316.960.3931154.71
SPInNeRF18.750.3519206.4317.470.5740239.99
MVPIneRF18.630.4332278.9917.670.5070255.51
Ours19.670.2685149.5217.580.4513154.34
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.231, + 0.907, + 0.413 + ], + "angle": 0, + "content": "address this, we cluster all views into \\( L \\) groups and select the center view of each group as the anchor. Multi-view refinement is applied only to these \\( L \\) anchor views, while the remaining views in each group are warped from their respective anchors, as artifacts caused by warping between closely aligned views are negligible. \\( L \\) is set to 12 for \\( 360^{\\circ} \\) cases and linearly scaled down for other scenarios, such as 4 for \\( 90^{\\circ} \\) cases. We directly use object masks when evaluating on front-facing scenes and apply inpainting mask detection on others. All experiments are conducted on a single NVIDIA V100 GPU, with the inpainting of a 3D scene taking approximately one hour." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.424, + 0.892, + 0.44 + ], + "angle": 0, + "content": "5.1. Evaluation against State-of-the-Art Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.447, + 0.907, + 0.734 + ], + "angle": 0, + "content": "We evaluate our method against the state-of-the-art methods on our dataset and SPINeRF dataset. The following state-of-the-art methods are reproduced on our datasets, including GaussianEditor [5], GScream [32], MVIPNeRF [3] and SPInNeRF [17] using their official implementations. In particular, since the naive NeRF models used in MVIPNeRF and SPInNeRF cannot handle complex unbounded scenes, we replace them with MeRF [21] considering both memory usage and reconstruction performance. On our dataset, we utilize the evaluation protocol and official implementation from [3] to compare methods. Specifically, we render color images at corresponding test views from the inpainted scenes and evaluate them against the ground truth using three metrics, namely PSNR, LPIPS, and FID. These metrics are calculated only within the bounding box defined by the object mask. For the SPINeRF dataset, we copy values from [3] for SPInNeRF and MVIPNeRF, and we reproduce GaussianEditor and GScream to align with the metric computation protocol in [3]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.906, + 0.854 + ], + "angle": 0, + "content": "As can be seen from Tab. 1, the proposed method consistently achieves the best performance on our dataset, with a particularly large margin that demonstrates its superiority in unconstrained realistic scenarios. On the SPINeRF dataset, our method achieves state-of-the-art results comparable to other approaches, given that this dataset is limited to front-facing cases and challenging to distinguish methods in handling complex cases." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "In addition, we provide qualitative comparisons in Fig. 5, where our method successfully removes objects and seamlessly inpaints with appropriate content. Moreover, the in" + }, + { + "type": "page_footnote", + "bbox": [ + 0.092, + 0.887, + 0.422, + 0.901 + ], + "angle": 0, + "content": "https://huggingface.co/stabilityai/stable-diffusion-2-inpainting" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26699" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.09, + 0.33, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.332, + 0.091, + 0.442, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.444, + 0.091, + 0.555, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.091, + 0.666, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.091, + 0.779, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.781, + 0.091, + 0.892, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.264, + 0.33, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.205, + 0.436, + 0.232, + 0.447 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.332, + 0.264, + 0.442, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.333, + 0.349, + 0.442, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.352, + 0.436, + 0.424, + 0.446 + ], + "angle": 0, + "content": "GaussianEditor" + }, + { + "type": "image", + "bbox": [ + 0.445, + 0.264, + 0.554, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.445, + 0.349, + 0.555, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.48, + 0.436, + 0.519, + 0.445 + ], + "angle": 0, + "content": "GSream" + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.264, + 0.666, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.349, + 0.666, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.587, + 0.436, + 0.638, + 0.445 + ], + "angle": 0, + "content": "SPInNeRF" + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.264, + 0.779, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.349, + 0.779, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.697, + 0.436, + 0.753, + 0.445 + ], + "angle": 0, + "content": "MVIPNeRF" + }, + { + "type": "image", + "bbox": [ + 0.781, + 0.264, + 0.892, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.781, + 0.349, + 0.892, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.822, + 0.436, + 0.849, + 0.445 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.458, + 0.908, + 0.501 + ], + "angle": 0, + "content": "Figure 5. Qualitative evaluation. We show the visual comparisons against the state-of-the-art 3D inpainting algorithms, with two views displayed at the top and bottom for each method. Our method generates consistent, higher-quality results with rich texture details and fewer artifacts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.51, + 0.484, + 0.601 + ], + "angle": 0, + "content": "painted scene maintains 3D consistency, yielding visually more pleasing outcomes and a coherent geometric structure. In contrast, other methods often produce severe artifacts and fail to inpaint objects effectively in complex, realistic scenarios. We highly recommend viewing our video visualizations for a more comprehensive comparison." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.617, + 0.36, + 0.634 + ], + "angle": 0, + "content": "5.2. Detection of Inpainting Masks" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.642, + 0.483, + 0.837 + ], + "angle": 0, + "content": "We then demonstrate the importance of doing inpainting mask detection and differentiating them from object masks. For our collected dataset, we manually annotate inpainting masks from rendered images of pruned scenes and compare them with the original object masks. Since our dataset is derived from continuous capture videos, we annotate one out of every 10 images in each scene, resulting in a total of 260 annotated masks. Our analysis shows that the inpainting masks, on average, occupy only \\(50.78\\%\\) of the area covered by the object masks. This finding highlights that adopting entire object masks requires inpainting nearly twice the necessary area, complicating the task and leading to suboptimal results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We also compare our detected inpainting masks with the manually labeled masks using the mean Intersection over Union (mIoU) metric. Our method achieves an \\(81.12\\%\\) IoU score, validating the accuracy and robustness of the pro" + }, + { + "type": "table_caption", + "bbox": [ + 0.559, + 0.508, + 0.86, + 0.521 + ], + "angle": 0, + "content": "Table 2. Quantitative results of ablation studies." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.523, + 0.905, + 0.615 + ], + "angle": 0, + "content": "
MethodPSNR (↑)LPIPS (↓)FID (↓)
w/o Warping17.850.3215198.24
w/o Refinement18.900.3069206.96
General Refinement19.080.2719165.80
Single-View Refinement19.460.2725154.33
Multi-View Refinement (Ours)19.670.2685149.52
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.627, + 0.907, + 0.673 + ], + "angle": 0, + "content": "posed inpainting mask detection strategy. In contrast, the method proposed in [5] achieves only \\(42.55\\%\\), which is significantly inferior to our approach." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.681, + 0.667, + 0.698 + ], + "angle": 0, + "content": "5.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.905, + 0.734 + ], + "angle": 0, + "content": "Our dataset is chosen for the ablation study due to its diversity and complexity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Effectiveness of geometry-guided warping: With image rendering of the pruned scene and inpainting masks, we adopt a warping-refinement scheme rather than directly inpainting the masked areas. To demonstrate the effectiveness of our approach, we disable geometry reconstruction and warping operations, repurposing the multi-view refinement model as a multi-view inpainting model. Specifically, we zero out the masked regions in the input images and attempt to consistently inpaint the missing areas. The inpainting model is also fine-tuned for each given scene, with all other settings remaining unchanged." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "26700" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.258 + ], + "angle": 0, + "content": "We compare images generated with and without the warping process in Fig. 6. Despite the strength of multiview architecture, direct inpainting fails to generate consistent output with the reference image in large view-change cases, which aligns with the observation in [2]. In contrast, our warping-refinement scheme provides a significant advantage, as the warped content offers strong guidance and regularizes the model to generate consistent results. We then use the generated images to fine-tune 3D scene. As shown in Tab. 2, the fine-tuned scene indeed shows inferior results to our complete method." + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.271, + 0.223, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.106, + 0.37, + 0.216, + 0.381 + ], + "angle": 0, + "content": "(a) Reference Image" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.271, + 0.348, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.241, + 0.37, + 0.331, + 0.381 + ], + "angle": 0, + "content": "(b) W/o warping" + }, + { + "type": "image", + "bbox": [ + 0.348, + 0.271, + 0.474, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.371, + 0.37, + 0.453, + 0.381 + ], + "angle": 0, + "content": "(c) W/ warping" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.383, + 0.484, + 0.453 + ], + "angle": 0, + "content": "Figure 6. Multi-view refinement with and without conditioning on the warped image. (a) shows the inpainted reference with the originally captured image shown in the top right corner. (c) and (b) are the images from an alternate view generated with and without having the warped image as a condition, respectively." + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.484, + 0.225, + 0.579 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.108, + 0.582, + 0.217, + 0.593 + ], + "angle": 0, + "content": "(a) Reference Image" + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.484, + 0.349, + 0.579 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.25, + 0.582, + 0.326, + 0.593 + ], + "angle": 0, + "content": "(b) W/o refine" + }, + { + "type": "image", + "bbox": [ + 0.351, + 0.484, + 0.475, + 0.579 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.358, + 0.582, + 0.469, + 0.593 + ], + "angle": 0, + "content": "(c) Multi-view refine" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.594, + 0.225, + 0.689 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.106, + 0.692, + 0.215, + 0.703 + ], + "angle": 0, + "content": "(d) Reference image" + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.594, + 0.348, + 0.689 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.228, + 0.692, + 0.343, + 0.703 + ], + "angle": 0, + "content": "(e) Single-view refine" + }, + { + "type": "image", + "bbox": [ + 0.35, + 0.594, + 0.472, + 0.689 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.357, + 0.692, + 0.466, + 0.703 + ], + "angle": 0, + "content": "(f) Multi-view refine" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.789 + ], + "angle": 0, + "content": "Figure 7. Ablation on multi-view refinement. (a) and (d) show the inpainted reference images with the originally captured images shown in the top right corner; (b) and (c) are the images from an alternate view generated without and with the refinement network, respectively; (e) and (f) are the images generated by the single-view and multi-view refinement model, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Effectiveness of multi-view refinement model: We design the following two variants to illustrate the importance of the multi-view refinement model. a) w/o refinement: the refinement model is removed, and the warped images are directly used to fine-tune the pruned scene. b) single view: we replace the space-time attention with the original self-attention layer, performing refinement independently for" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.137 + ], + "angle": 0, + "content": "each view without leveraging correlations across different views. When fine-tuning the single-view model for a given scene, only random image pairs are used." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.138, + 0.906, + 0.272 + ], + "angle": 0, + "content": "Similarly, we compare images generated by these variants in Fig. 7. As seen in (a), (b), and (c), warping indeed introduces artifacts and texture mismatches, while the multiview refinement model effectively addresses them. In (d), (e), and (f), the single-view refinement model fails to generate results consistent with the reference image, whereas the multi-view model achieves consistency. We further fine-tune the 3D scene with corresponding images. As illustrated in Tab. 2, our completed method achieves the best result." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.274, + 0.907, + 0.501 + ], + "angle": 0, + "content": "Necessity of test-time adaption: Our pipeline requires a per-scene fine-tuned refinement model to generate images. To demonstrate its necessity, we also train a generalized refinement model on the DI3DV dataset, a scene-level multiview dataset, where we randomly selected two views as the reference and target, synthesized warping artifacts on one, and trained the model to restore. However, as shown in Tab. 2, its performance is inferior to our proposed method due to the limited size of the training set and domain gaps between the training set and real-world cases, such as differences in the severity of artifacts/mismatch. In addition, we use an existing inpainting method, LeftRefill [1], to directly inpaint images based on the reference image. However, as shown in Fig.8, LeftRefill can only ensures overall visual similarity but lacks pixel-level consistency." + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.514, + 0.647, + 0.57 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.549, + 0.572, + 0.621, + 0.582 + ], + "angle": 0, + "content": "(a) Reference" + }, + { + "type": "image", + "bbox": [ + 0.649, + 0.514, + 0.773, + 0.57 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.677, + 0.572, + 0.743, + 0.582 + ], + "angle": 0, + "content": "(b) 1st View" + }, + { + "type": "image", + "bbox": [ + 0.774, + 0.514, + 0.898, + 0.57 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.801, + 0.572, + 0.871, + 0.582 + ], + "angle": 0, + "content": "(c) 2nd View" + }, + { + "type": "image_caption", + "bbox": [ + 0.586, + 0.585, + 0.833, + 0.599 + ], + "angle": 0, + "content": "Figure 8. Inpainting results by LeftRefill." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.613, + 0.764, + 0.629 + ], + "angle": 0, + "content": "6. Conclusion and Limitation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.639, + 0.906, + 0.789 + ], + "angle": 0, + "content": "We present a novel 3D inpainting pipeline that achieves state-of-the-art results across both front-facing and unconstrained benchmarks, with particularly strong performance on the latter. We achieve this by utilizing geometry priors and a multi-view refinement network to generate multi-view consistent inpainted images. Additionally, a novel inpainting mask detection technique is developed to boost performance in unconstrained scenes. The effectiveness of our method is demonstrated through extensive experiments and ablation studies." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.79, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Limitations: The multi-view refinement network requires customization for each scene, which may take considerable time. Future advancements in parameter-efficient fine-tuning and the personalization of diffusion models may help to accelerate and further optimize our approach. In addition, our framework relies on monocular depth estimation and image segmentation modules. The accuracy and robustness of these modules can affect the overall performance." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26701" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Chenjie Cao, Yunuo Cai, Qiaole Dong, Yikai Wang, and Yanwei Fu. Leftrightfill: Filling right canvas based on left reference through generalized text-to-image diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7705-7715, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.185, + 0.484, + 0.24 + ], + "angle": 0, + "content": "[2] Chenjie Cao, Chaohui Yu, Yanwei Fu, Fan Wang, and Xi-angyang Xue. Mvinpainter: Learning multi-view consistent inpainting to bridge 2d and 3d editing. arXiv preprint arXiv:2408.08000, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.242, + 0.483, + 0.31 + ], + "angle": 0, + "content": "[3] Honghua Chen, Chen Change Loy, and Xingang Pan. Mvipnerf: Multi-view 3d inpainting on nerf scenes via diffusion prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5344-5353, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.311, + 0.483, + 0.38 + ], + "angle": 0, + "content": "[4] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. Anydoor: Zero-shot object-level image customization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6593-6602, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.381, + 0.484, + 0.464 + ], + "angle": 0, + "content": "[5] Yiwen Chen, Zilong Chen, Chi Zhang, Feng Wang, Xiaofeng Yang, Yikai Wang, Zhongang Cai, Lei Yang, Huaping Liu, and Guosheng Lin. Gaussianeditor: Swift and controllable 3d editing with gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21476-21485, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.466, + 0.483, + 0.522 + ], + "angle": 0, + "content": "[6] Brian Curless and Marc Levoy. A volumetric method for building complex models from range images. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 303-312, 1996." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.522, + 0.482, + 0.55 + ], + "angle": 0, + "content": "[7] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.55, + 0.482, + 0.592 + ], + "angle": 0, + "content": "[8] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.593, + 0.482, + 0.662 + ], + "angle": 0, + "content": "[9] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.663, + 0.483, + 0.717 + ], + "angle": 0, + "content": "[10] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.719, + 0.483, + 0.775 + ], + "angle": 0, + "content": "[11] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.776, + 0.483, + 0.829 + ], + "angle": 0, + "content": "[12] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.831, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[13] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[14] Chieh Hubert Lin, Changil Kim, Jia-Bin Huang, Qinbo Li, Chih-Yao Ma, Johannes Kopf, Ming-Hsuan Yang, and Hung-Yu Tseng. Taming latent diffusion model for neural radiance field inpainting. In European Conference on Computer Vision, pages 149-165. Springer, 2025." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.162, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[15] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.218, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[16] Ashkan Mirzaei, Tristan Aumentado-Armstrong, Marcus A Brubaker, Jonathan Kelly, Alex Levinshtein, Konstantinos G Derpanis, and Igor Gilitschenski. Reference-guided controllable inpainting of neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17815-17825, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.905, + 0.397 + ], + "angle": 0, + "content": "[17] Ashkan Mirzaei, Tristan Aumentado-Armstrong, Konstantinos G Derpanis, Jonathan Kelly, Marcus A Brubaker, Igor Gilitschenski, and Alex Levinshtein. Spin-nerf: Multiview segmentation and perceptual inpainting with neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20669–20679, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.399, + 0.905, + 0.468 + ], + "angle": 0, + "content": "[18] Ashkan Mirzaei, Riccardo De Lutio, Seung Wook Kim, David Acuna, Jonathan Kelly, Sanja Fidler, Igor Gilitschenski, and Zan Gojcic. Refusion: Reference adapted diffusion models for 3d scene inpainting. arXiv preprint arXiv:2404.10765, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.469, + 0.905, + 0.538 + ], + "angle": 0, + "content": "[19] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4296–4304, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.905, + 0.58 + ], + "angle": 0, + "content": "[20] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.58, + 0.905, + 0.649 + ], + "angle": 0, + "content": "[21] Christian Reiser, Rick Szeliski, Dor Verbin, Pratul Srinivasan, Ben Mildenhall, Andreas Geiger, Jon Barron, and Peter Hedman. Merf: Memory-efficient radiance fields for real-time view synthesis in unbounded scenes. ACM Transactions on Graphics (TOG), 42(4):1-12, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.65, + 0.905, + 0.719 + ], + "angle": 0, + "content": "[22] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.719, + 0.905, + 0.802 + ], + "angle": 0, + "content": "[23] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.905, + 0.859 + ], + "angle": 0, + "content": "[24] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123: a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[25] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26702" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.16 + ], + "angle": 0, + "content": "[26] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.165, + 0.482, + 0.204 + ], + "angle": 0, + "content": "[27] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.208, + 0.482, + 0.302 + ], + "angle": 0, + "content": "[28] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2149–2159, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.307, + 0.482, + 0.374 + ], + "angle": 0, + "content": "[29] Luming Tang, Nataniel Ruiz, Qinghao Chu, Yuanzhen Li, Aleksander Holynski, David E Jacobs, Bharath Hariharan, Yael Pritch, Neal Wadhwa, Kfir Aberman, et al. Realfill: Reference-driven generation for authentic image completion. ACM Transactions on Graphics (TOG), 43(4):1-12, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.377, + 0.482, + 0.445 + ], + "angle": 0, + "content": "[30] Dongqing Wang, Tong Zhang, Alaa Abboud, and Sabine Susstrunk. Innerf360: Text-guided 3d-consistent object inpainting on 360-degree neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12677-12686, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.448, + 0.482, + 0.488 + ], + "angle": 0, + "content": "[31] Peng Wang and Yichun Shi. Imagedream: Image-prompt multi-view diffusion for 3d generation. arXiv preprint arXiv:2312.02201, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.492, + 0.482, + 0.545 + ], + "angle": 0, + "content": "[32] Yuxin Wang, Qianyi Wu, Guofeng Zhang, and Dan Xu. Gscream: Learning 3d geometry and feature consistent gaussian splatting for object removal. arXiv preprint arXiv:2404.13679, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.549, + 0.482, + 0.63 + ], + "angle": 0, + "content": "[33] Ethan Weber, Aleksander Holynski, Varun Jampani, Saurabh Saxena, Noah Snavely, Abhishek Kar, and Angjoo Kanazawa. Nerfiller: Completing scenes via generative 3d inpainting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20731-20741, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.482, + 0.714 + ], + "angle": 0, + "content": "[34] Silvan Weder, Guillermo Garcia-Hernando, Aron Monsz-part, Marc Pollefeys, Gabriel J Brostow, Michael Firman, and Sara Vicente. Removing objects from neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16528-16538, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.482, + 0.8 + ], + "angle": 0, + "content": "[35] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7623-7633, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.883 + ], + "angle": 0, + "content": "[36] Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen, and Fang Wen. Paint by example: Exemplar-based image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18381-18391, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.888, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[37] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In CVPR, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[38] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[39] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[40] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "26703" + } + ] +] \ No newline at end of file diff --git a/2025/IMFine_ 3D Inpainting via Geometry-guided Multi-view Refinement/0becb03e-65d1-4a50-a09c-58afb744258c_origin.pdf b/2025/IMFine_ 3D Inpainting via Geometry-guided Multi-view Refinement/0becb03e-65d1-4a50-a09c-58afb744258c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..14ed749b5ac90076d18b47d74850df808258e3c1 --- /dev/null +++ b/2025/IMFine_ 3D Inpainting via Geometry-guided Multi-view Refinement/0becb03e-65d1-4a50-a09c-58afb744258c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6021fdb29a1fa445ab88f399bc0367f5abdfb0ac15581b10873637d9d770f291 +size 4555112 diff --git a/2025/IMFine_ 3D Inpainting via Geometry-guided Multi-view Refinement/full.md b/2025/IMFine_ 3D Inpainting via Geometry-guided Multi-view Refinement/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f21a654e0630c52bc41e1fb632e8c8cafe819be8 --- /dev/null +++ b/2025/IMFine_ 3D Inpainting via Geometry-guided Multi-view Refinement/full.md @@ -0,0 +1,368 @@ +# IMFine: 3D Inpainting via Geometry-guided Multi-view Refinement + +Zhihao Shi\*1 + +Dong Huo\*o2 + +Juwei Lu1 + +Yuhongze Zhou + +Yan Min + +Xinxin Zuo4 + +1Huawei Canada Research Institute + +$^{2}$ University of Alberta + +$^{3}$ McMaster University + +4Concordia University + +zhihaoshi2022@gmail.com, dhuo@ualberta.ca, yuhongze.zhou@mail.mcgill.com + +miny13@mcmaster.ca, juwei.lu@huawei.com, xinxin.zuo@concordia.ca + +![](images/802c94cab733739ffc55572c05715245fe5039bb40248878ee56001ddc0327ca.jpg) + +![](images/5dc2cf05ddd57a12ecd97ea59658a2da28c9dd06ede82f459e9d57adae0d1c06.jpg) +Figure 1. We present a novel 3D inpainting method capable of removing target objects from the 3D scenes and seamlessly inpainting the occluded regions across various scenarios. The proposed method ensures the generation of plausible geometry and consistent textures. It is highly recommended to see video results in our project page. + +![](images/80f8bb6e32cd2512f1ead2a7f4eef8a3cf4c3e2ff5e3d4ff4f6641a4a633f502.jpg) + +![](images/330b90eb00feaa3260b89887b0a9660ffa4e74064dc2096c7d4b52475c9b3dc7.jpg) + +![](images/bfa34b0b2843481996d6b67ca21be45f598f982c8be62bb92a0713fca16e2ad7.jpg) + +![](images/b6ede422d3009909ed3ed87dcd79711bdecf1fea85e748ac4fd19604344ae172.jpg) + +![](images/88ca5647b1541e54ab9c3c16950ce39224e3ffe9a3690105d30a17862dee6e85.jpg) + +![](images/d2eb3d80fddd61e745966c05b441e0a16e913afa23d6b5959af0edf9451e2cf3.jpg) + +# Abstract + +Current 3D inpainting and object removal methods are largely limited to front-facing scenes, facing substantial challenges when applied to diverse, "unconstrained" scenes where the camera orientation and trajectory are unrestricted. To bridge this gap, we introduce a novel approach that produces inpainted 3D scenes with consistent visual quality and coherent underlying geometry across both front-facing and unconstrained scenes. Specifically, we propose a robust 3D inpainting pipeline that incorporates geometric priors and a multi-view refinement network trained via test-time adaptation, building on a pretrained image inpainting model. Additionally, we develop a novel inpainting mask detection technique to derive targeted inpainting masks from object masks, boosting the performance in handling unconstrained scenes. To validate the efficacy of our approach, we create a challenging + +and diverse benchmark that spans a wide range of scenes. Comprehensive experiments demonstrate that our proposed method substantially outperforms existing state-of-the-art approaches. + +# 1. Introduction + +3D editing is an active research area with significant potential for various applications, including video games, the Metaverse, virtual reality, and augmented reality. For example, after a 3D scene is created or reconstructed, a user might want to remove certain objects from the scene. This paper mainly focuses on the challenging task of inpainting in 3D scenes to fill the occluded regions to enable seamless object removal. + +Compared to image inpainting [22, 28], 3D inpainting models lag significantly behind due to the lack of large-scale 3D datasets and efficient 3D data structures. Therefore, existing 3D inpainting methods mainly focus on lifting 2D inpainting priors to 3D. Although the current diffusion + +based image inpainting models can generate images with vivid textures and high-fidelity details, they are trained on single images without considering any 3D consistency. It poses great challenges in maintaining view consistency and preserving 3D coherence when lifting those 2D priors to 3D. + +To address the inconsistency issue, there are approaches that implicitly lift the 2D models into 3D via score distillation sampling (SDS) loss [3, 18], which was proposed initially in text-to-3D generation [20], to harness the inherent coherence of 2D diffusion models under the same prompt, complemented by the iterative updating scheme during optimization to ensure view consistency. However, they are constrained by the inherent limitation of SDS loss, which results in over-saturation and over-smoothing in the optimized images, even with different variants of SDS losses. Alternatively, explicit methods [2, 5, 16, 17, 30] first generate one or few images with the image inpainting models and try to preserve the 3D consistency by enforcing geometry constraints or minimizing perceptual losses between rendered images and the inpainted 2D images. Although they have demonstrated good performance, especially for front-facing scenes with very limited view changes, the inpainted 3D scenes have artifacts such as floaters and blurry textures in unconstrained scenarios with complex geometry and large view variation. The reasons are that first, 3D geometry is rather difficult to get perfectly restored; second, perceptual loss can only alleviate the inconsistency by minimizing overall similarity. Inconsistency still exists between those inpainted images, leading to blurry textures for the inpainted 3D scenes. More recently, [2] proposed to train a multi-view inpainting network to generate consistent multiview inpainted images. However, due to the limited diversity of current multi-view datasets, they cannot generalize to unconstrained 3D scenes (e.g., $360^{\circ}$ cases). + +To enhance the ability to generalize to various scenes for 3D inpainting, we utilize geometry priors and develop a multi-view refinement network fine-tuned at test time to produce consistent images in a coarse-to-fine manner. Specifically, we start with a single reference view inpainted with image inpainting models, followed by coarse geometry restoration and a warping operation to warp the inpainted content from the reference view to others. However, due to the imperfect reconstructed geometry and view-dependent effects, the warped images tend to have artifacts, such as distorted textures and mismatches between the warped and neighboring contents. To refine the warped contents, we present a multi-view refinement model to generate consistent images with high fidelity, taking the warped images as a condition. The key idea is that instead of training a model with extensive datasets, we take the pre-trained image inpainting model, add space-time attention layers, and fine-tune them with the current test scenes. We take the + +originally captured multi-view images and implement various augmentations to synthesize paired data to fine-tune the multi-view refinement model. + +Furthermore, previous works usually take the entire object masks as inpainted regions. However, some regions inside the object masks might have already been exposed in other views, especially for 3D object removal and inpainting in $360^{\circ}$ scenes (see Fig. 4). Therefore, we believe it is critical to differentiate between object masks and inpainting masks, fully exploiting the available information of the reconstructed 3D scenes. To this end, a novel inpainting mask detection approach is proposed to extract masks accurately and exclusively for real occluded regions. + +Finally, considering that the existing benchmarks mostly cover front-facing scenes or planar surfaces, we introduce a new dataset encompassing 20 diverse scenes, from indoor to outdoor environments, covering large view variations and complex interfacial surfaces, to better evaluate and advance 3D inpainting methods. + +In summary, our contributions are as follows: + +1. A multi-view consistent inpainting pipeline is proposed. We leverage geometric priors and a multi-view refinement model that is fine-tuned from image inpainting models via test-time adaptation to generate consistent multi-view images. +2. We further develop an inpainting mask detection technique that accurately distinguishes occluded regions, enabling robust processing of complex scenes, such as $360^{\circ}$ environments. +3. The proposed inpainting method has achieved state-of-the-art performance across various benchmarks, particularly excelling in challenging, unconstrained scenarios. +4. We contribute a new and challenging 3D scene inpainting benchmark, offering diverse scenes to facilitate broader evaluation and advancement in 3D inpainting research. + +# 2. Related Work + +# 2.1. 3D Inpainting + +Existing 3D inpainting methods [3, 5, 16-18, 30, 32] are implemented by lifting 2D inpainting priors to 3D with various strategies designed to ensure 3D consistency. One direction, inspired by score distillation-based approaches [20] in text-to-3D generation, formulates 3D inpainting as a continuous distillation process [3, 18]. While these approaches achieve reasonably consistent inpainting results, they face inherent limitations of score distillation, such as producing over-saturated or overly smooth outputs. Another direction explicitly enforces 3D consistency by independently inpainting one or multiple 2D views and then consolidating multi-view inconsistencies in 3D. Reference-guided NeRF [16] addresses this challenge by using a single view to + +guide geometry inpainting to avoid inconsistencies. However, it struggles with accurate geometry capture in complex $360^{\circ}$ scenes using just a single view. SPINeRF [17], and InNeRF360 [30] utilize multiple inpainted images to guide 3D inpainting and incorporate perceptual loss to reduce inconsistencies resulting from independently inpainted images. Nevertheless, while perceptual loss can maintain overall consistency, it performs poorly on finer details, often leading to blurry 3D results. Our method lies in the latter direction. It addresses these issues through a geometry-guided refinement network that generates a set of multi-view consistent images, reconstructing clearer and consistent 3D inpainted scenes. + +# 2.2. Video and Multi-View Diffusion + +The diffusion model for text-to-image (T2I) generation [7, 8, 22, 27] has achieved significant success in recent years, known for its ability to produce high-quality, diverse, and controllable outputs. Building on this, extensive studies have extended image diffusion models to the video generation domain [9, 10, 26, 40] by incorporating temporal layers within each transformer block to perform space-time attention, which captures correlations across frames and shows promising results in generating realistic videos. + +Inspired by these advances in video generation, researchers have begun exploring the adaptation of T2I models into multi-view diffusion models for implicit 3D reasoning [15, 24, 25, 31]. For example, methods such as MVDream [25] and Zero $123++$ [24] leverage this idea and derive multi-view diffusion models, capable of generating consistent multiple views simultaneously for 3D generation. However, the full space-time attention mechanism used in these methods can lead to prohibitive memory and computational costs, making it impractical for handling large numbers of views. To address this, recent work [35] introduces a sparse space-time attention mechanism, which reduces memory requirements by attending only to the initial and previous frames. However, the sparse attention approach assumes input frames are from a continuous video sequence and remains under-explored in sparse multi-view scenarios. + +# 2.3. Diffusion Model Fine-tuning + +Adapting large pre-trained diffusion models has proven to be an effective and efficient approach for completing user-specified tasks, such as synthesizing specific subjects in new contexts [4, 23] and inpainting images with instance matching reference images [29, 36]. Adaptation methods include injecting extra layers [19, 36, 38, 39], searching appropriate unique prompt [23], updating partial or full model weights [14, 18, 29], or combinations of these techniques. Similarly, in this paper, instead of relying on extensive multiview datasets, we propose adapting the pre-trained image + +diffusion model into a multi-view refinement model for 3D inpainting purposes. + +# 3. Preliminary + +Gaussian splating (GS) methods [11, 12] represent a scene as a collection of multivariate Gaussian primitives: $\mathcal{G} = \{(\mu_m,\Sigma_m,\Gamma_m,\mathrm{SH}_m)\}_{m = 1}^M$ , where $\mu_{m},\Sigma_{m},\Gamma_{m}$ and $\mathrm{SH}_m$ denote the mean vector, covariance matrix, opacity, and spherical harmonic coefficients of a multivariate Gaussian, respectively. View synthesis is performed by standard point-based $\alpha^\prime$ -rendering, where a single pixel color is computed by sorting $\mathcal{G}$ (or their 2D projection) in their depth order and alpha-blending each contributed primitive: + +$$ +C = \sum_ {m} c _ {m} \alpha_ {m} \prod_ {j = 1} ^ {m - 1} (1 - \alpha_ {j}), \tag {1} +$$ + +where $C$ is the output pixel color, $c_{m}$ is the pre-computed color of the $m$ -th primitive, and $\alpha_{m}$ is a calculated opacity based on distance from the Gaussian center to the interaction point. Due to the excellent performance in geometry reconstruction, in this work, we choose 2D-GS [11] as the 3D representation. + +# 4. Method + +Given a GS scene $\mathcal{G}$ , along with corresponding training images $\{I_n\}_{n=1}^N$ , associated camera poses $\{\Pi_n\}_{n=1}^N$ , and a set of object masks $\{M_n\}_{n=1}^N$ identifying objects to be removed, our goal is to generate an inpainted scene $\mathcal{G}^*$ where the target objects are removed and occluded regions are filled seamlessly with the background. The entire pipeline contains two stages, namely object removal and 3D scene inpainting. We first remove object primitives (Sec. 4.1), exposing never-before-seen (NBS) regions which are the areas that truly need inpainting. We then either use the object masks as inpainting masks when they are identical or detect specific inpainting masks (Sec. 4.3), ensuring that the inpainting mask exclusively covers the NBS regions in each view, denoted as $\{P_n\}_{n=1}^N$ . Given the pruned GS scene and inpainting masks, we propose a novel 3D inpainting pipeline to inpaint $\mathcal{G}$ as $\mathcal{G}^*$ (Sec. 4.2). + +# 4.1. Object Removal + +To remove the objects from the reconstructed GS scene, we append an additional learnable parameter $\{p_m\}_{m=1}^M \in [0,1]$ to each Gaussian primitive, indicating whether it belongs to the object to be removed, which is initialized as 0.5. To compute the value of $p_m$ for each Gaussian, we map the object masks back to the Gaussian primitives. Specifically, we render a set of object masks $\{\hat{M}\}_{n=1}^N$ using $\alpha'$ rendering (similar to color rendering) at $\{\Pi_n\}_{n=1}^N$ but replacing $c_m$ with $p_m$ in Eq. 1, followed by $K_1$ ( $K_1 = 1000$ ) steps of the + +![](images/7968f66a025678ba9933d82a59a7b787e212e73a9c0c17d9b2de96c033e38049.jpg) +(a) Multi-view Inpainting Pipeline + +![](images/d83ad3298341204ba8fc517ea7f5d381d85159f0f684b99db962b341f96a7a50.jpg) +(b) Test-time Adaptation +Figure 2. Overview of our proposed approach. (a) shows the proposed 3D inpainting pipeline. Specifically, a reference view is first inpainted, which will be used to reconstruct complete geometry, along with rendered multi-view depth maps. The inpainted reference image is then warped to other views, which are taken as input conditions for the multi-view refinement network to generate a set of view-consistent images. Afterward, we fine-tune the 3D scene to get the inpainted scene. (b) illustrates the detailed structure and training of the multi-view refinement network. Basically, we take the pre-trained image inpainting model, incorporate additional space-time attention layers, and exploit the test-time adaptation to fine-tune it into a multi-view refinement model. To generate a fine-tuning dataset, we synthesize warped images under the original captured scene, which has the ground-truth captured images. In detail, reference views are selected randomly, and we set random masks over the images, apply geometry jittering, and warp the reference images in other views. + +optimization to update $p_m$ with respect to $\mathcal{L}_1$ loss between $\{\hat{M}\}_{n=1}^N$ and $\{M\}_{n=1}^N$ . Notably, other parameters remain unchanged during this procedure. Finally, a pruning operation is performed to produce the pruned scene $\mathcal{G}_{\mathrm{pruned}}$ by thresholding $p_m$ with a threshold $\tau$ ( $\tau = 0.4$ ). + +# 4.2. Proposed 3D Inpainting Pipeline + +Given the pruned GS scene $\mathcal{G}_{\mathrm{pruned}}$ as well as inpainting masks $\{P_n\}_{n=1}^N$ indicating the regions to be inpainted for each view, we propose a novel 3D inpainting pipeline shown in Fig 2. We first render color images $\{I_n'\}_{n=1}^N$ from $\mathcal{G}_{\mathrm{pruned}}$ for each view and select a reference view with image $I_r' \in \{I_n'\}_{n=1}^N$ , which is then inpainted using a pre-trained inpainting model [22] to reference image $I_r^*$ . We conduct geometry reconstruction to fill out the hole in 3D, which will be utilized to warp the inpainted contents in the reference image into the remaining views. We employ a multi-view refinement model to further refine the warped images, ensuring high fidelity and consistency, and then update the 3D scene. The model is fine-tuned from the pre-trained image inpainting model via a test-time adaptation strategy. The geometry reconstruction and multi-view refinement model are described as follows. + +# 4.2.1 Geometry Reconstruction + +Following [11], we render a set of depth maps $\{D_n\}_{n=1}^N$ from the prune GS scene $\mathcal{G}_{\mathrm{pruned}}$ and adopt a truncated signed distance function (TSDF) [6] technique to reconstruct the 3D mesh. However, $\{D_n\}_{n=1}^N$ only contributes + +to the reconstruction of the background region, as the depth values within their inpainting mask areas are invalid. To fill the masked region, we aim to generate an all-valid depth map $D_r^*$ of the reference view by filling the masked areas in $D_r$ based on information from $I_r^*$ . The filled depth map should meet two key requirements: 1) it must be consistent with the reference image $I_r^*$ ; 2) it should seamlessly connect with the depth values outside the masked region. + +We begin by exploiting a monocular depth estimator [37] to estimate depth $\bar{D}_r$ from $I_r^*$ , followed by a linear transformation to convert $\bar{D}_r$ to scene scale $\tilde{D}_r$ : $\tilde{D}_r \colon \tilde{D}_r = a * \bar{D}_r + b$ , where the parameters $a, b$ are computed in the neighborhood of $P_r$ to minimize the difference between $\bar{D}_r$ and $D_r$ in advance, following [16]. We then develop a filling scheme according to $\tilde{D}_r$ . Formally, the filled values of $D_r^*$ are computed as follows: + +$$ +\begin{array}{l} D _ {r} ^ {*} (x) = \underset {D _ {r} ^ {*} (x)} {\arg \min } \sum_ {x \in P _ {r}} | | \nabla D _ {r} ^ {*} (x) - \nabla \tilde {D} _ {r} (x) | | ^ {2} + \lambda \nabla D _ {r} ^ {*} (x) \tag {2} \\ \begin{array}{l} \text {s . t .} D _ {r} ^ {*} (x) = D _ {r} (x), \text {w h e r e} P _ {r} (x) = 0, \end{array} \\ \end{array} +$$ + +where $\nabla$ denotes the Laplacian calculation and $x$ represents the pixel coordinate. Intuitively, the first term encourages $D_r^*$ to maintain a similar changing trend as $\tilde{D}_r$ , while the second term regularizes $D_r^*$ to be smooth and consistent with the depths outside the mask. As such, $D_r^*$ can be used for the reconstruction of the NBS region. + +![](images/88981aea9c4662bcea365b0224b5d7bbf5df389647a3280871185b14de82bfdb.jpg) +(a) Reference Image + +![](images/3ec1a30d6659afcffad3ba55f2073a41cdb4aa20caecd5365eb545408158abbd.jpg) +(b) Warped Image + +![](images/7378e1bf8d1d89f23736300e8914ce815fa40c5ab0ffec7fb51bd432026cd2a2.jpg) +(c) Refined Image +Figure 3. Illustration of the importance of multi-view refinement. The multi-view refinement model is proposed to tackle the warping artifacts with the example shown in (b), generating high-fidelity multi-view consistent images with one sampled view shown in (c). + +# 4.2.2 Test-Time Adaptation for Multi-view Refinement + +We warp the inpainted contents from the reference view $I_{r}$ to other views with the above restored geometry. Due to the imperfect geometry and the neglect of view-dependent effects, those warped contents contain artifacts, such as texture distortion and inconsistent appearance with the neighboring regions, especially when there are large view angle changes with respect to the reference view as shown in Fig 3. On the other hand, these warped contents have coarse-level consistency across views. To this end, we propose a multi-view refinement model that produces high-fidelity, consistent images conditioning on the warped images. + +Network: Our model is built upon a pre-trained diffusion-based image inpainting model [22], which takes a masked image and its associated mask as the input, corresponding to the warped image and inpainting mask in our scenarios, respectively. To ensure multi-view consistency, we replace the self-attention layer in each transformer block with space-time attention to explore correlations across multiple views. Considering the quadratic complexity of global attention and the large number of warped images, space-time attention is implemented as the sparse space-time attention [10] that only attends to neighboring images. The attention layer is also allowed to attend to the reference image for alignment. + +Training: Instead of relying on large datasets and extensive training, we exploit the test-time adaptation strategy for each scene to train the multi-view refinement network. Paired training sets are created for model adaptation. Specifically, given the original images $\{I_n\}_{n=1}^N$ , at each training step, $L + 1$ images are randomly selected as ground truth. To simulate realistic warping scenarios, we randomly choose one image as the reference, generate a random mask, and warp both the reference image and mask to other views with geometry guidance from $\mathcal{G}$ . In addition, to encourage cross-view learning as in [2], we also synthesize another set of paired images by applying image augmentations independently on each captured image, i.e., irregularly shaped masks that are randomly generated on + +![](images/b0432dd7b66401940f44d0310e20a9b8bfb2459010e1664b471a952a376df89c.jpg) + +![](images/7d5206d2cdc47c5069a14666acfbe0d2d3bd6a29a334ecf1653616312322ea5d.jpg) + +![](images/eda211d0b1eb916ea42106ff58661e6affc2b6af5654d671649e248da47c942c.jpg) +(a) Object mask and dilated mask +(c) Mask-remapping interpretation +Figure 4. Inpainting mask detection. (a) shows the object mask (orange) along with the dilated mask (pink), which includes additional neighboring pixels; Each dilated mask consists of two parts: the actual inpainting mask and out-of-interest areas, marked as regions $①$ and $②$ in (b). These regions are mapped in 3D to the NBS and background areas, respectively, as shown in regions $①$ and $②$ in (c). During multi-view mapping optimization, the central NBS region is progressively refined, while background areas are suppressed by each other; (d) shows the rendered neighbor mask after GS scene optimization. + +![](images/67965b8e942fe8f3b2ea1f02e734edc57d6164253a69e28ba82e7ef84834dd19.jpg) +(b) Object Mask Separation +(d) Render mask + +each image. Image-based augmentations—such as elastic transformations and color jittering, are then applied to the masked regions to simulate warping artifacts. We train the multi-view refinement model with those paired images, employing a simplified variational bound objective [8]. + +Inference: At inference, similar to [33], we randomly shuffle the order of the $L$ warped images at each denosing step to break the sequential constraint of sparse attention, further enhancing cross-view consistency. + +# 4.2.3 Fine-tune GS Scene + +With the generated multi-view consistent images, we fine-tune $\mathcal{G}_{\mathrm{pruned}}$ for $K_{2}(K_{2} = 7000)$ steps with an $L_{1}$ loss and a SSIM loss as straightforward pixel-wise supervision. It is worth noting that other robust losses (e.g., perceptual loss) are functionally orthogonal to our approach and can be used to improve it further. + +# 4.3. Inpainting Mask Detection + +In the overall pipeline of 3D object removal and scene inpainting, identifying the occluded or NBS regions for inpainting is also critical, especially for $360^{\circ}$ cases. In those cases, the object mask often covers a much larger area than the regions that require inpainting (as illustrated in Fig. 4). Therefore, a novel inpainting mask detection approach is proposed to identify the NBS region for each view. + +An intuitive idea is to detect the NBS region in the + +pruned GS scene and project it back into each view to obtain inpainting masks. However, since, by definition, the NBS region is occluded by the object, it usually lacks good geometry and meaningful Gaussians, making it challenging to detect in 3D. Here we propose to detect the inpainting masks by leveraging neighbors of NBS region. In detail, we first dilate the object masks to include neighboring pixels (Fig. 4(a)), and then map the dilated masks onto the pruned scene $\mathcal{G}_{\mathrm{pruned}}$ . Mapping is implemented as updating the learnable parameter $\{p_m\}_{m=1}^M$ of each Gaussian with respect to $\mathcal{L}_1$ loss between rendered masks of $\mathcal{G}_{\mathrm{pruned}}$ and the dilated masks. The optimization process is the same as object removal (Sec. 4.1) but with the object mask replaced with dilated masks, and the optimization is performed over the pruned scene. We illustrate the interpretation of this process in Fig. 4(b-c). After the mapping process, we are able to get the neighbor mask by rendering $\{p_m\}_{m=1}^M$ from $\mathcal{G}_{\mathrm{pruned}}$ as shown in Fig. 4(d). Finally, we extract the bounding box for the above rendered mask, which will be taken as input prompt for segment anything model [13] to get the final inpainting mask. + +# 5. Experiments + +Datasets: Most existing publicly available 3D inpainting datasets [17, 34] are limited to front-facing cases, which cannot be used to evaluate methods in more diverse or realistic situations. To address this, we construct a dataset that contains 20 diverse cases, ranging from indoor to outdoor, covering various camera trajectories-including front-facing, $90^{\circ}$ , $180^{\circ}$ , and $360^{\circ}$ , and containing various interfacial surfaces such as single plane, multiple planes, curvature surfaces, and irregular surfaces. Each scene is recorded using a DJI Pocket 3 at 60fps with a resolution of $1920 \times 1080$ , from which we evenly extract 175 images and generate the corresponding camera poses and object masks. The first 50 images are used for evaluation purpose with the object removed, while the remaining 125 images are object-present and are used to reconstruct the original GS scene. Previews of our collected dataset are shown in supplementary materials. We use half resolution in the following experiments. In addition, we evaluate our method on SPINeRF dataset [17] to compare with state-of-the-art. + +Implementation Details: StableDiffusionInpainting-Pipeline is utilized for image inpainting and our multi-view refinement model initialization. During training, we randomly select $L = 5$ images from training images and crop $512 \times 512$ patches to fine-tune the model over 1000 iterations with the learning rate and batch size set as $3 \times 10^{-5}$ and 1, individually. At inference, directly refining all warped images, i.e., $L = 125$ would result in prohibitive memory usage and computational costs. To + +Table 1. Quantitative evaluation on the SPINeRF dataset and our dataset. Numbers in bold indicate the best performance, and underscored numbers indicate the second best. + +
MethodOursSPINeRF
PSNR (↑)LPIPS (↓)FID (↓)PSNR (↑)LPIPS (↓)FID (↓)
GaussianEditor15.710.6163375.0314.410.6247343.16
GScream17.180.4431290.6316.960.3931154.71
SPInNeRF18.750.3519206.4317.470.5740239.99
MVPIneRF18.630.4332278.9917.670.5070255.51
Ours19.670.2685149.5217.580.4513154.34
+ +address this, we cluster all views into $L$ groups and select the center view of each group as the anchor. Multi-view refinement is applied only to these $L$ anchor views, while the remaining views in each group are warped from their respective anchors, as artifacts caused by warping between closely aligned views are negligible. $L$ is set to 12 for $360^{\circ}$ cases and linearly scaled down for other scenarios, such as 4 for $90^{\circ}$ cases. We directly use object masks when evaluating on front-facing scenes and apply inpainting mask detection on others. All experiments are conducted on a single NVIDIA V100 GPU, with the inpainting of a 3D scene taking approximately one hour. + +# 5.1. Evaluation against State-of-the-Art Methods + +We evaluate our method against the state-of-the-art methods on our dataset and SPINeRF dataset. The following state-of-the-art methods are reproduced on our datasets, including GaussianEditor [5], GScream [32], MVIPNeRF [3] and SPInNeRF [17] using their official implementations. In particular, since the naive NeRF models used in MVIPNeRF and SPInNeRF cannot handle complex unbounded scenes, we replace them with MeRF [21] considering both memory usage and reconstruction performance. On our dataset, we utilize the evaluation protocol and official implementation from [3] to compare methods. Specifically, we render color images at corresponding test views from the inpainted scenes and evaluate them against the ground truth using three metrics, namely PSNR, LPIPS, and FID. These metrics are calculated only within the bounding box defined by the object mask. For the SPINeRF dataset, we copy values from [3] for SPInNeRF and MVIPNeRF, and we reproduce GaussianEditor and GScream to align with the metric computation protocol in [3]. + +As can be seen from Tab. 1, the proposed method consistently achieves the best performance on our dataset, with a particularly large margin that demonstrates its superiority in unconstrained realistic scenarios. On the SPINeRF dataset, our method achieves state-of-the-art results comparable to other approaches, given that this dataset is limited to front-facing cases and challenging to distinguish methods in handling complex cases. + +In addition, we provide qualitative comparisons in Fig. 5, where our method successfully removes objects and seamlessly inpaints with appropriate content. Moreover, the in + +![](images/9fd45b9fab78c3e9a2527e9fd5205b01a9011690108551aff11c6a29a9b58681.jpg) + +![](images/01996cb03495c23b237d7263cf289df809e7e37ed041bcc0ea67549d142cb25c.jpg) + +![](images/d368c73cf0a32cd2779b81ac836b1c506b37e07ec5402fa14be8bac51375afed.jpg) + +![](images/6dc19e7aa41c2ce6438f5942a191f0594835e39db08764b3172189bef8454922.jpg) + +![](images/29ef3461c4289f94df28cadce23062c5e7ed20c7975b04a2cc326109b2cb55bb.jpg) + +![](images/393a64c93ad923a356538319f047929bf7b23d99bfdea56b7190f941e4b060bb.jpg) + +![](images/23b4d95936c79134a3ee59bae087a35ffbded0d3fb10a43fa17bb26dc2d8b3d4.jpg) +Input + +![](images/064e37c5ac318744af35fac67b8967bc2c999c883a1dd6c0c3435861cfdf6272.jpg) + +![](images/e7e636c3576c2b4cd7ba19bd0ab1853f0199428da4111d1c7df7e29c8fc29e79.jpg) +GaussianEditor + +![](images/b314c43523702de0c838c532eddedc7f8a75ff56a0a2971881d8efd640fe1726.jpg) + +![](images/7a0245a2d170bdb0759f41c5ea56c32bf28e4fb17832cc083355337da73668d9.jpg) +GSream +Figure 5. Qualitative evaluation. We show the visual comparisons against the state-of-the-art 3D inpainting algorithms, with two views displayed at the top and bottom for each method. Our method generates consistent, higher-quality results with rich texture details and fewer artifacts. + +![](images/18057c39ebeec521382308ebb62748e9c88aa5cce12cb7fff284d36dd2d3b0a5.jpg) + +![](images/23c613235a9a7650ebcec634059754b0e36b1b7eb92a1a457601de563c784ad0.jpg) +SPInNeRF + +![](images/76577a197a90ffb1f2c874fe28a04038d84e28882547323ff888a62e4d5627ae.jpg) + +![](images/aadc38e749d5d184ec8b39ccb4ea2f87f13adfe6bc75682c7cfc609f808ac06a.jpg) +MVIPNeRF + +![](images/d1bcbcbe4d8838ccf3ed6c1bc363d868c93e33e11983ca4fc616ab8629507eac.jpg) + +![](images/f79a365c1e4dd18df55890efd61162f1add6aa610e4cb234eb03889fdcbfbcc3.jpg) +Ours + +painted scene maintains 3D consistency, yielding visually more pleasing outcomes and a coherent geometric structure. In contrast, other methods often produce severe artifacts and fail to inpaint objects effectively in complex, realistic scenarios. We highly recommend viewing our video visualizations for a more comprehensive comparison. + +# 5.2. Detection of Inpainting Masks + +We then demonstrate the importance of doing inpainting mask detection and differentiating them from object masks. For our collected dataset, we manually annotate inpainting masks from rendered images of pruned scenes and compare them with the original object masks. Since our dataset is derived from continuous capture videos, we annotate one out of every 10 images in each scene, resulting in a total of 260 annotated masks. Our analysis shows that the inpainting masks, on average, occupy only $50.78\%$ of the area covered by the object masks. This finding highlights that adopting entire object masks requires inpainting nearly twice the necessary area, complicating the task and leading to suboptimal results. + +We also compare our detected inpainting masks with the manually labeled masks using the mean Intersection over Union (mIoU) metric. Our method achieves an $81.12\%$ IoU score, validating the accuracy and robustness of the pro + +Table 2. Quantitative results of ablation studies. + +
MethodPSNR (↑)LPIPS (↓)FID (↓)
w/o Warping17.850.3215198.24
w/o Refinement18.900.3069206.96
General Refinement19.080.2719165.80
Single-View Refinement19.460.2725154.33
Multi-View Refinement (Ours)19.670.2685149.52
+ +posed inpainting mask detection strategy. In contrast, the method proposed in [5] achieves only $42.55\%$ , which is significantly inferior to our approach. + +# 5.3. Ablation Study + +Our dataset is chosen for the ablation study due to its diversity and complexity. + +Effectiveness of geometry-guided warping: With image rendering of the pruned scene and inpainting masks, we adopt a warping-refinement scheme rather than directly inpainting the masked areas. To demonstrate the effectiveness of our approach, we disable geometry reconstruction and warping operations, repurposing the multi-view refinement model as a multi-view inpainting model. Specifically, we zero out the masked regions in the input images and attempt to consistently inpaint the missing areas. The inpainting model is also fine-tuned for each given scene, with all other settings remaining unchanged. + +We compare images generated with and without the warping process in Fig. 6. Despite the strength of multiview architecture, direct inpainting fails to generate consistent output with the reference image in large view-change cases, which aligns with the observation in [2]. In contrast, our warping-refinement scheme provides a significant advantage, as the warped content offers strong guidance and regularizes the model to generate consistent results. We then use the generated images to fine-tune 3D scene. As shown in Tab. 2, the fine-tuned scene indeed shows inferior results to our complete method. + +![](images/31be7770edf282eb21ab19ed4c49a286db472285d9d90a9d146ff167fa8fc12c.jpg) +(a) Reference Image + +![](images/786e7dac24072b7b9e9074c2f48bc737797f3ead5c09d3167dd3ac1a133333f2.jpg) +(b) W/o warping + +![](images/fb0dd3e8044c742661fe54d5d481ec128a757f18579a0a7d0120ede43df40e9a.jpg) +(c) W/ warping +Figure 6. Multi-view refinement with and without conditioning on the warped image. (a) shows the inpainted reference with the originally captured image shown in the top right corner. (c) and (b) are the images from an alternate view generated with and without having the warped image as a condition, respectively. + +![](images/ce6aae3bb56c4389b91fd14d9a46bdd521928b7792f7aa49fcdcb6019b69f5a9.jpg) +(a) Reference Image + +![](images/814bec8e5c19ce142250ffa0aab31261da8dc5b3dc45571a1dcd9b60190fe916.jpg) +(b) W/o refine + +![](images/cbc5d993772d0903e028951f23ad5d32877d26246c8252bc227f1bcdad516912.jpg) + +![](images/023687d7fe1395b1155c79244c6b5d27823535b42f1100a1f0fb33cc6b56721d.jpg) +(d) Reference image +Figure 7. Ablation on multi-view refinement. (a) and (d) show the inpainted reference images with the originally captured images shown in the top right corner; (b) and (c) are the images from an alternate view generated without and with the refinement network, respectively; (e) and (f) are the images generated by the single-view and multi-view refinement model, respectively. + +![](images/83eb1c12036272c5fd582894d9047e73830a5a494243daea5d29337264102b98.jpg) +(e) Single-view refine + +![](images/5784a93a682cefca5911ade79761263de7a46da5608ad1199fafc847b5350a34.jpg) +(c) Multi-view refine +(f) Multi-view refine + +Effectiveness of multi-view refinement model: We design the following two variants to illustrate the importance of the multi-view refinement model. a) w/o refinement: the refinement model is removed, and the warped images are directly used to fine-tune the pruned scene. b) single view: we replace the space-time attention with the original self-attention layer, performing refinement independently for + +each view without leveraging correlations across different views. When fine-tuning the single-view model for a given scene, only random image pairs are used. + +Similarly, we compare images generated by these variants in Fig. 7. As seen in (a), (b), and (c), warping indeed introduces artifacts and texture mismatches, while the multiview refinement model effectively addresses them. In (d), (e), and (f), the single-view refinement model fails to generate results consistent with the reference image, whereas the multi-view model achieves consistency. We further fine-tune the 3D scene with corresponding images. As illustrated in Tab. 2, our completed method achieves the best result. + +Necessity of test-time adaption: Our pipeline requires a per-scene fine-tuned refinement model to generate images. To demonstrate its necessity, we also train a generalized refinement model on the DI3DV dataset, a scene-level multiview dataset, where we randomly selected two views as the reference and target, synthesized warping artifacts on one, and trained the model to restore. However, as shown in Tab. 2, its performance is inferior to our proposed method due to the limited size of the training set and domain gaps between the training set and real-world cases, such as differences in the severity of artifacts/mismatch. In addition, we use an existing inpainting method, LeftRefill [1], to directly inpaint images based on the reference image. However, as shown in Fig.8, LeftRefill can only ensures overall visual similarity but lacks pixel-level consistency. + +![](images/870c58f54a8262529f9431a97220e7520dad64a5d07280ab999bae359cf02f09.jpg) +(a) Reference +Figure 8. Inpainting results by LeftRefill. + +![](images/4a4cd5a62f73ec830ab653a961ee92f71c33b28911ef2781319a1231865d1d72.jpg) +(b) 1st View + +![](images/330676a2728fdf714710c5f254163437544d0ae9e578ceeab4b84dd2bd8e803a.jpg) +(c) 2nd View + +# 6. Conclusion and Limitation + +We present a novel 3D inpainting pipeline that achieves state-of-the-art results across both front-facing and unconstrained benchmarks, with particularly strong performance on the latter. We achieve this by utilizing geometry priors and a multi-view refinement network to generate multi-view consistent inpainted images. Additionally, a novel inpainting mask detection technique is developed to boost performance in unconstrained scenes. The effectiveness of our method is demonstrated through extensive experiments and ablation studies. + +Limitations: The multi-view refinement network requires customization for each scene, which may take considerable time. Future advancements in parameter-efficient fine-tuning and the personalization of diffusion models may help to accelerate and further optimize our approach. In addition, our framework relies on monocular depth estimation and image segmentation modules. The accuracy and robustness of these modules can affect the overall performance. + +# References + +[1] Chenjie Cao, Yunuo Cai, Qiaole Dong, Yikai Wang, and Yanwei Fu. Leftrightfill: Filling right canvas based on left reference through generalized text-to-image diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7705-7715, 2024. +[2] Chenjie Cao, Chaohui Yu, Yanwei Fu, Fan Wang, and Xi-angyang Xue. Mvinpainter: Learning multi-view consistent inpainting to bridge 2d and 3d editing. arXiv preprint arXiv:2408.08000, 2024. +[3] Honghua Chen, Chen Change Loy, and Xingang Pan. Mvipnerf: Multi-view 3d inpainting on nerf scenes via diffusion prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5344-5353, 2024. +[4] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. Anydoor: Zero-shot object-level image customization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6593-6602, 2024. +[5] Yiwen Chen, Zilong Chen, Chi Zhang, Feng Wang, Xiaofeng Yang, Yikai Wang, Zhongang Cai, Lei Yang, Huaping Liu, and Guosheng Lin. Gaussianeditor: Swift and controllable 3d editing with gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21476-21485, 2024. +[6] Brian Curless and Marc Levoy. A volumetric method for building complex models from range images. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 303-312, 1996. +[7] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. +[8] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. +[9] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. +[10] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. +[11] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. +[12] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. +[13] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. + +[14] Chieh Hubert Lin, Changil Kim, Jia-Bin Huang, Qinbo Li, Chih-Yao Ma, Johannes Kopf, Ming-Hsuan Yang, and Hung-Yu Tseng. Taming latent diffusion model for neural radiance field inpainting. In European Conference on Computer Vision, pages 149-165. Springer, 2025. +[15] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453, 2023. +[16] Ashkan Mirzaei, Tristan Aumentado-Armstrong, Marcus A Brubaker, Jonathan Kelly, Alex Levinshtein, Konstantinos G Derpanis, and Igor Gilitschenski. Reference-guided controllable inpainting of neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17815-17825, 2023. +[17] Ashkan Mirzaei, Tristan Aumentado-Armstrong, Konstantinos G Derpanis, Jonathan Kelly, Marcus A Brubaker, Igor Gilitschenski, and Alex Levinshtein. Spin-nerf: Multiview segmentation and perceptual inpainting with neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20669–20679, 2023. +[18] Ashkan Mirzaei, Riccardo De Lutio, Seung Wook Kim, David Acuna, Jonathan Kelly, Sanja Fidler, Igor Gilitschenski, and Zan Gojcic. Refusion: Reference adapted diffusion models for 3d scene inpainting. arXiv preprint arXiv:2404.10765, 2024. +[19] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4296–4304, 2024. +[20] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. +[21] Christian Reiser, Rick Szeliski, Dor Verbin, Pratul Srinivasan, Ben Mildenhall, Andreas Geiger, Jon Barron, and Peter Hedman. Merf: Memory-efficient radiance fields for real-time view synthesis in unbounded scenes. ACM Transactions on Graphics (TOG), 42(4):1-12, 2023. +[22] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. +[23] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. +[24] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123: a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110, 2023. +[25] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. + +[26] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. +[27] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. +[28] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2149–2159, 2022. +[29] Luming Tang, Nataniel Ruiz, Qinghao Chu, Yuanzhen Li, Aleksander Holynski, David E Jacobs, Bharath Hariharan, Yael Pritch, Neal Wadhwa, Kfir Aberman, et al. Realfill: Reference-driven generation for authentic image completion. ACM Transactions on Graphics (TOG), 43(4):1-12, 2024. +[30] Dongqing Wang, Tong Zhang, Alaa Abboud, and Sabine Susstrunk. Innerf360: Text-guided 3d-consistent object inpainting on 360-degree neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12677-12686, 2024. +[31] Peng Wang and Yichun Shi. Imagedream: Image-prompt multi-view diffusion for 3d generation. arXiv preprint arXiv:2312.02201, 2023. +[32] Yuxin Wang, Qianyi Wu, Guofeng Zhang, and Dan Xu. Gscream: Learning 3d geometry and feature consistent gaussian splatting for object removal. arXiv preprint arXiv:2404.13679, 2024. +[33] Ethan Weber, Aleksander Holynski, Varun Jampani, Saurabh Saxena, Noah Snavely, Abhishek Kar, and Angjoo Kanazawa. Nerfiller: Completing scenes via generative 3d inpainting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20731-20741, 2024. +[34] Silvan Weder, Guillermo Garcia-Hernando, Aron Monsz-part, Marc Pollefeys, Gabriel J Brostow, Michael Firman, and Sara Vicente. Removing objects from neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16528-16538, 2023. +[35] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7623-7633, 2023. +[36] Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen, and Fang Wen. Paint by example: Exemplar-based image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18381-18391, 2023. +[37] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi + +Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In CVPR, 2024. +[38] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. +[39] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. +[40] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022. \ No newline at end of file diff --git a/2025/IMFine_ 3D Inpainting via Geometry-guided Multi-view Refinement/images.zip b/2025/IMFine_ 3D Inpainting via Geometry-guided Multi-view Refinement/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7c2896ea866e7c85570eaf91e78ec34101ebd892 --- /dev/null +++ b/2025/IMFine_ 3D Inpainting via Geometry-guided Multi-view Refinement/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:137e59a57b765d0aba2c5d6201caf55bd946299bd792cb48e6cf0b3a7e4e0f0b +size 630891 diff --git a/2025/IMFine_ 3D Inpainting via Geometry-guided Multi-view Refinement/layout.json b/2025/IMFine_ 3D Inpainting via Geometry-guided Multi-view Refinement/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d0ee4330dc1f4be6be6707c77f67f4990ad57604 --- /dev/null +++ b/2025/IMFine_ 3D Inpainting via Geometry-guided Multi-view Refinement/layout.json @@ -0,0 +1,10621 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 94, + 103, + 516, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 103, + 516, + 121 + ], + "spans": [ + { + "bbox": [ + 94, + 103, + 516, + 121 + ], + "type": "text", + "content": "IMFine: 3D Inpainting via Geometry-guided Multi-view Refinement" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 128, + 142, + 194, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 142, + 194, + 156 + ], + "spans": [ + { + "bbox": [ + 128, + 142, + 194, + 156 + ], + "type": "text", + "content": "Zhihao Shi\\*1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 221, + 143, + 287, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 221, + 143, + 287, + 157 + ], + "spans": [ + { + "bbox": [ + 221, + 143, + 287, + 157 + ], + "type": "text", + "content": "Dong Huo\\*o2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 230, + 158, + 280, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 158, + 280, + 170 + ], + "spans": [ + { + "bbox": [ + 230, + 158, + 280, + 170 + ], + "type": "text", + "content": "Juwei Lu1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 315, + 143, + 396, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 143, + 396, + 157 + ], + "spans": [ + { + "bbox": [ + 315, + 143, + 396, + 157 + ], + "type": "text", + "content": "Yuhongze Zhou" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 429, + 143, + 481, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 429, + 143, + 481, + 156 + ], + "spans": [ + { + "bbox": [ + 429, + 143, + 481, + 156 + ], + "type": "text", + "content": "Yan Min" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 318, + 157, + 380, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 157, + 380, + 171 + ], + "spans": [ + { + "bbox": [ + 318, + 157, + 380, + 171 + ], + "type": "text", + "content": "Xinxin Zuo4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 151, + 171, + 323, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 171, + 323, + 183 + ], + "spans": [ + { + "bbox": [ + 151, + 171, + 323, + 183 + ], + "type": "text", + "content": "1Huawei Canada Research Institute" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 351, + 171, + 459, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 171, + 459, + 185 + ], + "spans": [ + { + "bbox": [ + 351, + 171, + 459, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 351, + 171, + 459, + 185 + ], + "type": "text", + "content": "University of Alberta" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 182, + 184, + 291, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 184, + 291, + 200 + ], + "spans": [ + { + "bbox": [ + 182, + 184, + 291, + 200 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 182, + 184, + 291, + 200 + ], + "type": "text", + "content": "McMaster University" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 319, + 185, + 428, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 185, + 428, + 200 + ], + "spans": [ + { + "bbox": [ + 319, + 185, + 428, + 200 + ], + "type": "text", + "content": "4Concordia University" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 113, + 201, + 503, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 201, + 503, + 213 + ], + "spans": [ + { + "bbox": [ + 113, + 201, + 503, + 213 + ], + "type": "text", + "content": "zhihaoshi2022@gmail.com, dhuo@ualberta.ca, yuhongze.zhou@mail.mcgill.com" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 216, + 480, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 227 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 227 + ], + "type": "text", + "content": "miny13@mcmaster.ca, juwei.lu@huawei.com, xinxin.zuo@concordia.ca" + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 56, + 255, + 138, + 335 + ], + "blocks": [ + { + "bbox": [ + 56, + 255, + 138, + 335 + ], + "lines": [ + { + "bbox": [ + 56, + 255, + 138, + 335 + ], + "spans": [ + { + "bbox": [ + 56, + 255, + 138, + 335 + ], + "type": "image", + "image_path": "802c94cab733739ffc55572c05715245fe5039bb40248878ee56001ddc0327ca.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 56, + 335, + 136, + 415 + ], + "blocks": [ + { + "bbox": [ + 56, + 335, + 136, + 415 + ], + "lines": [ + { + "bbox": [ + 56, + 335, + 136, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 335, + 136, + 415 + ], + "type": "image", + "image_path": "5dc2cf05ddd57a12ecd97ea59658a2da28c9dd06ede82f459e9d57adae0d1c06.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 423, + 555, + 458 + ], + "lines": [ + { + "bbox": [ + 54, + 423, + 555, + 458 + ], + "spans": [ + { + "bbox": [ + 54, + 423, + 555, + 458 + ], + "type": "text", + "content": "Figure 1. We present a novel 3D inpainting method capable of removing target objects from the 3D scenes and seamlessly inpainting the occluded regions across various scenarios. The proposed method ensures the generation of plausible geometry and consistent textures. It is highly recommended to see video results in our project page." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 138, + 283, + 160, + 307 + ], + "blocks": [ + { + "bbox": [ + 138, + 283, + 160, + 307 + ], + "lines": [ + { + "bbox": [ + 138, + 283, + 160, + 307 + ], + "spans": [ + { + "bbox": [ + 138, + 283, + 160, + 307 + ], + "type": "image", + "image_path": "80f8bb6e32cd2512f1ead2a7f4eef8a3cf4c3e2ff5e3d4ff4f6641a4a633f502.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 160, + 255, + 316, + 334 + ], + "blocks": [ + { + "bbox": [ + 160, + 255, + 316, + 334 + ], + "lines": [ + { + "bbox": [ + 160, + 255, + 316, + 334 + ], + "spans": [ + { + "bbox": [ + 160, + 255, + 316, + 334 + ], + "type": "image", + "image_path": "330b90eb00feaa3260b89887b0a9660ffa4e74064dc2096c7d4b52475c9b3dc7.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 316, + 255, + 475, + 334 + ], + "blocks": [ + { + "bbox": [ + 316, + 255, + 475, + 334 + ], + "lines": [ + { + "bbox": [ + 316, + 255, + 475, + 334 + ], + "spans": [ + { + "bbox": [ + 316, + 255, + 475, + 334 + ], + "type": "image", + "image_path": "bfa34b0b2843481996d6b67ca21be45f598f982c8be62bb92a0713fca16e2ad7.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 476, + 255, + 553, + 334 + ], + "blocks": [ + { + "bbox": [ + 476, + 255, + 553, + 334 + ], + "lines": [ + { + "bbox": [ + 476, + 255, + 553, + 334 + ], + "spans": [ + { + "bbox": [ + 476, + 255, + 553, + 334 + ], + "type": "image", + "image_path": "b6ede422d3009909ed3ed87dcd79711bdecf1fea85e748ac4fd19604344ae172.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 160, + 335, + 316, + 415 + ], + "blocks": [ + { + "bbox": [ + 160, + 335, + 316, + 415 + ], + "lines": [ + { + "bbox": [ + 160, + 335, + 316, + 415 + ], + "spans": [ + { + "bbox": [ + 160, + 335, + 316, + 415 + ], + "type": "image", + "image_path": "88ca5647b1541e54ab9c3c16950ce39224e3ffe9a3690105d30a17862dee6e85.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 316, + 335, + 475, + 415 + ], + "blocks": [ + { + "bbox": [ + 316, + 335, + 475, + 415 + ], + "lines": [ + { + "bbox": [ + 316, + 335, + 475, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 335, + 475, + 415 + ], + "type": "image", + "image_path": "d2eb3d80fddd61e745966c05b441e0a16e913afa23d6b5959af0edf9451e2cf3.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "bbox": [ + 151, + 468, + 200, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 468, + 200, + 480 + ], + "spans": [ + { + "bbox": [ + 151, + 468, + 200, + 480 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 54, + 493, + 297, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 493, + 297, + 685 + ], + "spans": [ + { + "bbox": [ + 54, + 493, + 297, + 685 + ], + "type": "text", + "content": "Current 3D inpainting and object removal methods are largely limited to front-facing scenes, facing substantial challenges when applied to diverse, \"unconstrained\" scenes where the camera orientation and trajectory are unrestricted. To bridge this gap, we introduce a novel approach that produces inpainted 3D scenes with consistent visual quality and coherent underlying geometry across both front-facing and unconstrained scenes. Specifically, we propose a robust 3D inpainting pipeline that incorporates geometric priors and a multi-view refinement network trained via test-time adaptation, building on a pretrained image inpainting model. Additionally, we develop a novel inpainting mask detection technique to derive targeted inpainting masks from object masks, boosting the performance in handling unconstrained scenes. To validate the efficacy of our approach, we create a challenging" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 469, + 555, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 469, + 555, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 469, + 555, + 518 + ], + "type": "text", + "content": "and diverse benchmark that spans a wide range of scenes. Comprehensive experiments demonstrate that our proposed method substantially outperforms existing state-of-the-art approaches." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 314, + 538, + 394, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 538, + 394, + 550 + ], + "spans": [ + { + "bbox": [ + 314, + 538, + 394, + 550 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 557, + 555, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 557, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 312, + 557, + 555, + 653 + ], + "type": "text", + "content": "3D editing is an active research area with significant potential for various applications, including video games, the Metaverse, virtual reality, and augmented reality. For example, after a 3D scene is created or reconstructed, a user might want to remove certain objects from the scene. This paper mainly focuses on the challenging task of inpainting in 3D scenes to fill the occluded regions to enable seamless object removal." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": "Compared to image inpainting [22, 28], 3D inpainting models lag significantly behind due to the lack of large-scale 3D datasets and efficient 3D data structures. Therefore, existing 3D inpainting methods mainly focus on lifting 2D inpainting priors to 3D. Although the current diffusion" + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 693, + 152, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 152, + 703 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 152, + 703 + ], + "type": "text", + "content": "* denotes equal contribution." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 56, + 704, + 287, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 704, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 704, + 287, + 712 + ], + "type": "text", + "content": "Work done during an internship at Huawei Canada Research Institute" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26694" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "based image inpainting models can generate images with vivid textures and high-fidelity details, they are trained on single images without considering any 3D consistency. It poses great challenges in maintaining view consistency and preserving 3D coherence when lifting those 2D priors to 3D." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 148, + 298, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 148, + 298, + 495 + ], + "spans": [ + { + "bbox": [ + 56, + 148, + 298, + 495 + ], + "type": "text", + "content": "To address the inconsistency issue, there are approaches that implicitly lift the 2D models into 3D via score distillation sampling (SDS) loss [3, 18], which was proposed initially in text-to-3D generation [20], to harness the inherent coherence of 2D diffusion models under the same prompt, complemented by the iterative updating scheme during optimization to ensure view consistency. However, they are constrained by the inherent limitation of SDS loss, which results in over-saturation and over-smoothing in the optimized images, even with different variants of SDS losses. Alternatively, explicit methods [2, 5, 16, 17, 30] first generate one or few images with the image inpainting models and try to preserve the 3D consistency by enforcing geometry constraints or minimizing perceptual losses between rendered images and the inpainted 2D images. Although they have demonstrated good performance, especially for front-facing scenes with very limited view changes, the inpainted 3D scenes have artifacts such as floaters and blurry textures in unconstrained scenarios with complex geometry and large view variation. The reasons are that first, 3D geometry is rather difficult to get perfectly restored; second, perceptual loss can only alleviate the inconsistency by minimizing overall similarity. Inconsistency still exists between those inpainted images, leading to blurry textures for the inpainted 3D scenes. More recently, [2] proposed to train a multi-view inpainting network to generate consistent multiview inpainted images. However, due to the limited diversity of current multi-view datasets, they cannot generalize to unconstrained 3D scenes (e.g., " + }, + { + "bbox": [ + 56, + 148, + 298, + 495 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 56, + 148, + 298, + 495 + ], + "type": "text", + "content": " cases)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 498, + 298, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 298, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 298, + 713 + ], + "type": "text", + "content": "To enhance the ability to generalize to various scenes for 3D inpainting, we utilize geometry priors and develop a multi-view refinement network fine-tuned at test time to produce consistent images in a coarse-to-fine manner. Specifically, we start with a single reference view inpainted with image inpainting models, followed by coarse geometry restoration and a warping operation to warp the inpainted content from the reference view to others. However, due to the imperfect reconstructed geometry and view-dependent effects, the warped images tend to have artifacts, such as distorted textures and mismatches between the warped and neighboring contents. To refine the warped contents, we present a multi-view refinement model to generate consistent images with high fidelity, taking the warped images as a condition. The key idea is that instead of training a model with extensive datasets, we take the pre-trained image inpainting model, add space-time attention layers, and fine-tune them with the current test scenes. We take the" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": "originally captured multi-view images and implement various augmentations to synthesize paired data to fine-tune the multi-view refinement model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 109, + 555, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 109, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 109, + 555, + 228 + ], + "type": "text", + "content": "Furthermore, previous works usually take the entire object masks as inpainted regions. However, some regions inside the object masks might have already been exposed in other views, especially for 3D object removal and inpainting in " + }, + { + "bbox": [ + 313, + 109, + 555, + 228 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 313, + 109, + 555, + 228 + ], + "type": "text", + "content": " scenes (see Fig. 4). Therefore, we believe it is critical to differentiate between object masks and inpainting masks, fully exploiting the available information of the reconstructed 3D scenes. To this end, a novel inpainting mask detection approach is proposed to extract masks accurately and exclusively for real occluded regions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 229, + 556, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 229, + 556, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 229, + 556, + 300 + ], + "type": "text", + "content": "Finally, considering that the existing benchmarks mostly cover front-facing scenes or planar surfaces, we introduce a new dataset encompassing 20 diverse scenes, from indoor to outdoor environments, covering large view variations and complex interfacial surfaces, to better evaluate and advance 3D inpainting methods." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 302, + 498, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 302, + 498, + 313 + ], + "spans": [ + { + "bbox": [ + 314, + 302, + 498, + 313 + ], + "type": "text", + "content": "In summary, our contributions are as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 315, + 315, + 553, + 505 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 315, + 315, + 553, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 315, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 315, + 315, + 553, + 374 + ], + "type": "text", + "content": "1. A multi-view consistent inpainting pipeline is proposed. We leverage geometric priors and a multi-view refinement model that is fine-tuned from image inpainting models via test-time adaptation to generate consistent multi-view images." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 315, + 374, + 553, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 374, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 315, + 374, + 553, + 421 + ], + "type": "text", + "content": "2. We further develop an inpainting mask detection technique that accurately distinguishes occluded regions, enabling robust processing of complex scenes, such as " + }, + { + "bbox": [ + 315, + 374, + 553, + 421 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 315, + 374, + 553, + 421 + ], + "type": "text", + "content": " environments." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 315, + 422, + 553, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 422, + 553, + 458 + ], + "spans": [ + { + "bbox": [ + 315, + 422, + 553, + 458 + ], + "type": "text", + "content": "3. The proposed inpainting method has achieved state-of-the-art performance across various benchmarks, particularly excelling in challenging, unconstrained scenarios." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 315, + 458, + 553, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 458, + 553, + 505 + ], + "spans": [ + { + "bbox": [ + 315, + 458, + 553, + 505 + ], + "type": "text", + "content": "4. We contribute a new and challenging 3D scene inpainting benchmark, offering diverse scenes to facilitate broader evaluation and advancement in 3D inpainting research." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 518, + 402, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 518, + 402, + 530 + ], + "spans": [ + { + "bbox": [ + 314, + 518, + 402, + 530 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 539, + 403, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 539, + 403, + 552 + ], + "spans": [ + { + "bbox": [ + 314, + 539, + 403, + 552 + ], + "type": "text", + "content": "2.1. 3D Inpainting" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 557, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 556, + 715 + ], + "type": "text", + "content": "Existing 3D inpainting methods [3, 5, 16-18, 30, 32] are implemented by lifting 2D inpainting priors to 3D with various strategies designed to ensure 3D consistency. One direction, inspired by score distillation-based approaches [20] in text-to-3D generation, formulates 3D inpainting as a continuous distillation process [3, 18]. While these approaches achieve reasonably consistent inpainting results, they face inherent limitations of score distillation, such as producing over-saturated or overly smooth outputs. Another direction explicitly enforces 3D consistency by independently inpainting one or multiple 2D views and then consolidating multi-view inconsistencies in 3D. Reference-guided NeRF [16] addresses this challenge by using a single view to" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26695" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "content": "guide geometry inpainting to avoid inconsistencies. However, it struggles with accurate geometry capture in complex " + }, + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "content": " scenes using just a single view. SPINeRF [17], and InNeRF360 [30] utilize multiple inpainted images to guide 3D inpainting and incorporate perceptual loss to reduce inconsistencies resulting from independently inpainted images. Nevertheless, while perceptual loss can maintain overall consistency, it performs poorly on finer details, often leading to blurry 3D results. Our method lies in the latter direction. It addresses these issues through a geometry-guided refinement network that generates a set of multi-view consistent images, reconstructing clearer and consistent 3D inpainted scenes." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 236, + 228, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 236, + 228, + 248 + ], + "spans": [ + { + "bbox": [ + 55, + 236, + 228, + 248 + ], + "type": "text", + "content": "2.2. Video and Multi-View Diffusion" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 255, + 296, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 296, + 361 + ], + "type": "text", + "content": "The diffusion model for text-to-image (T2I) generation [7, 8, 22, 27] has achieved significant success in recent years, known for its ability to produce high-quality, diverse, and controllable outputs. Building on this, extensive studies have extended image diffusion models to the video generation domain [9, 10, 26, 40] by incorporating temporal layers within each transformer block to perform space-time attention, which captures correlations across frames and shows promising results in generating realistic videos." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 363, + 296, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 363, + 296, + 566 + ], + "spans": [ + { + "bbox": [ + 55, + 363, + 296, + 566 + ], + "type": "text", + "content": "Inspired by these advances in video generation, researchers have begun exploring the adaptation of T2I models into multi-view diffusion models for implicit 3D reasoning [15, 24, 25, 31]. For example, methods such as MVDream [25] and Zero " + }, + { + "bbox": [ + 55, + 363, + 296, + 566 + ], + "type": "inline_equation", + "content": "123++" + }, + { + "bbox": [ + 55, + 363, + 296, + 566 + ], + "type": "text", + "content": " [24] leverage this idea and derive multi-view diffusion models, capable of generating consistent multiple views simultaneously for 3D generation. However, the full space-time attention mechanism used in these methods can lead to prohibitive memory and computational costs, making it impractical for handling large numbers of views. To address this, recent work [35] introduces a sparse space-time attention mechanism, which reduces memory requirements by attending only to the initial and previous frames. However, the sparse attention approach assumes input frames are from a continuous video sequence and remains under-explored in sparse multi-view scenarios." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 574, + 212, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 574, + 212, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 212, + 588 + ], + "type": "text", + "content": "2.3. Diffusion Model Fine-tuning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": "Adapting large pre-trained diffusion models has proven to be an effective and efficient approach for completing user-specified tasks, such as synthesizing specific subjects in new contexts [4, 23] and inpainting images with instance matching reference images [29, 36]. Adaptation methods include injecting extra layers [19, 36, 38, 39], searching appropriate unique prompt [23], updating partial or full model weights [14, 18, 29], or combinations of these techniques. Similarly, in this paper, instead of relying on extensive multiview datasets, we propose adapting the pre-trained image" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": "diffusion model into a multi-view refinement model for 3D inpainting purposes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 106, + 391, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 106, + 391, + 121 + ], + "spans": [ + { + "bbox": [ + 314, + 106, + 391, + 121 + ], + "type": "text", + "content": "3. Preliminary" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 127, + 555, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 127, + 555, + 236 + ], + "spans": [ + { + "bbox": [ + 313, + 127, + 555, + 236 + ], + "type": "text", + "content": "Gaussian splating (GS) methods [11, 12] represent a scene as a collection of multivariate Gaussian primitives: " + }, + { + "bbox": [ + 313, + 127, + 555, + 236 + ], + "type": "inline_equation", + "content": "\\mathcal{G} = \\{(\\mu_m,\\Sigma_m,\\Gamma_m,\\mathrm{SH}_m)\\}_{m = 1}^M" + }, + { + "bbox": [ + 313, + 127, + 555, + 236 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 127, + 555, + 236 + ], + "type": "inline_equation", + "content": "\\mu_{m},\\Sigma_{m},\\Gamma_{m}" + }, + { + "bbox": [ + 313, + 127, + 555, + 236 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 127, + 555, + 236 + ], + "type": "inline_equation", + "content": "\\mathrm{SH}_m" + }, + { + "bbox": [ + 313, + 127, + 555, + 236 + ], + "type": "text", + "content": " denote the mean vector, covariance matrix, opacity, and spherical harmonic coefficients of a multivariate Gaussian, respectively. View synthesis is performed by standard point-based " + }, + { + "bbox": [ + 313, + 127, + 555, + 236 + ], + "type": "inline_equation", + "content": "\\alpha^\\prime" + }, + { + "bbox": [ + 313, + 127, + 555, + 236 + ], + "type": "text", + "content": "-rendering, where a single pixel color is computed by sorting " + }, + { + "bbox": [ + 313, + 127, + 555, + 236 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 313, + 127, + 555, + 236 + ], + "type": "text", + "content": " (or their 2D projection) in their depth order and alpha-blending each contributed primitive:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 372, + 244, + 555, + 277 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 244, + 555, + 277 + ], + "spans": [ + { + "bbox": [ + 372, + 244, + 555, + 277 + ], + "type": "interline_equation", + "content": "C = \\sum_ {m} c _ {m} \\alpha_ {m} \\prod_ {j = 1} ^ {m - 1} (1 - \\alpha_ {j}), \\tag {1}", + "image_path": "a31953e2f52ffb2e60076bb3fdd3ffce6d394ccfb0d03c6c01d1987d306a14cc.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 285, + 554, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 554, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 554, + 357 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 285, + 554, + 357 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 285, + 554, + 357 + ], + "type": "text", + "content": " is the output pixel color, " + }, + { + "bbox": [ + 313, + 285, + 554, + 357 + ], + "type": "inline_equation", + "content": "c_{m}" + }, + { + "bbox": [ + 313, + 285, + 554, + 357 + ], + "type": "text", + "content": " is the pre-computed color of the " + }, + { + "bbox": [ + 313, + 285, + 554, + 357 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 285, + 554, + 357 + ], + "type": "text", + "content": "-th primitive, and " + }, + { + "bbox": [ + 313, + 285, + 554, + 357 + ], + "type": "inline_equation", + "content": "\\alpha_{m}" + }, + { + "bbox": [ + 313, + 285, + 554, + 357 + ], + "type": "text", + "content": " is a calculated opacity based on distance from the Gaussian center to the interaction point. Due to the excellent performance in geometry reconstruction, in this work, we choose 2D-GS [11] as the 3D representation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 368, + 370, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 368, + 370, + 380 + ], + "spans": [ + { + "bbox": [ + 314, + 368, + 370, + 380 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 388, + 555, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 555, + 580 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 555, + 580 + ], + "type": "text", + "content": "Given a GS scene " + }, + { + "bbox": [ + 313, + 388, + 555, + 580 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 313, + 388, + 555, + 580 + ], + "type": "text", + "content": ", along with corresponding training images " + }, + { + "bbox": [ + 313, + 388, + 555, + 580 + ], + "type": "inline_equation", + "content": "\\{I_n\\}_{n=1}^N" + }, + { + "bbox": [ + 313, + 388, + 555, + 580 + ], + "type": "text", + "content": ", associated camera poses " + }, + { + "bbox": [ + 313, + 388, + 555, + 580 + ], + "type": "inline_equation", + "content": "\\{\\Pi_n\\}_{n=1}^N" + }, + { + "bbox": [ + 313, + 388, + 555, + 580 + ], + "type": "text", + "content": ", and a set of object masks " + }, + { + "bbox": [ + 313, + 388, + 555, + 580 + ], + "type": "inline_equation", + "content": "\\{M_n\\}_{n=1}^N" + }, + { + "bbox": [ + 313, + 388, + 555, + 580 + ], + "type": "text", + "content": " identifying objects to be removed, our goal is to generate an inpainted scene " + }, + { + "bbox": [ + 313, + 388, + 555, + 580 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^*" + }, + { + "bbox": [ + 313, + 388, + 555, + 580 + ], + "type": "text", + "content": " where the target objects are removed and occluded regions are filled seamlessly with the background. The entire pipeline contains two stages, namely object removal and 3D scene inpainting. We first remove object primitives (Sec. 4.1), exposing never-before-seen (NBS) regions which are the areas that truly need inpainting. We then either use the object masks as inpainting masks when they are identical or detect specific inpainting masks (Sec. 4.3), ensuring that the inpainting mask exclusively covers the NBS regions in each view, denoted as " + }, + { + "bbox": [ + 313, + 388, + 555, + 580 + ], + "type": "inline_equation", + "content": "\\{P_n\\}_{n=1}^N" + }, + { + "bbox": [ + 313, + 388, + 555, + 580 + ], + "type": "text", + "content": ". Given the pruned GS scene and inpainting masks, we propose a novel 3D inpainting pipeline to inpaint " + }, + { + "bbox": [ + 313, + 388, + 555, + 580 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 313, + 388, + 555, + 580 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 388, + 555, + 580 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^*" + }, + { + "bbox": [ + 313, + 388, + 555, + 580 + ], + "type": "text", + "content": " (Sec. 4.2)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 586, + 411, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 586, + 411, + 600 + ], + "spans": [ + { + "bbox": [ + 314, + 586, + 411, + 600 + ], + "type": "text", + "content": "4.1. Object Removal" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": "To remove the objects from the reconstructed GS scene, we append an additional learnable parameter " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\{p_m\\}_{m=1}^M \\in [0,1]" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": " to each Gaussian primitive, indicating whether it belongs to the object to be removed, which is initialized as 0.5. To compute the value of " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "p_m" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": " for each Gaussian, we map the object masks back to the Gaussian primitives. Specifically, we render a set of object masks " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\{\\hat{M}\\}_{n=1}^N" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\alpha'" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": " rendering (similar to color rendering) at " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\{\\Pi_n\\}_{n=1}^N" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": " but replacing " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "c_m" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "p_m" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": " in Eq. 1, followed by " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "K_1" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "K_1 = 1000" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": ") steps of the" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "26696" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 85, + 324, + 240 + ], + "blocks": [ + { + "bbox": [ + 126, + 70, + 267, + 83 + ], + "lines": [ + { + "bbox": [ + 126, + 70, + 267, + 83 + ], + "spans": [ + { + "bbox": [ + 126, + 70, + 267, + 83 + ], + "type": "text", + "content": "(a) Multi-view Inpainting Pipeline" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 56, + 85, + 324, + 240 + ], + "lines": [ + { + "bbox": [ + 56, + 85, + 324, + 240 + ], + "spans": [ + { + "bbox": [ + 56, + 85, + 324, + 240 + ], + "type": "image", + "image_path": "7968f66a025678ba9933d82a59a7b787e212e73a9c0c17d9b2de96c033e38049.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 328, + 87, + 553, + 241 + ], + "blocks": [ + { + "bbox": [ + 390, + 70, + 492, + 83 + ], + "lines": [ + { + "bbox": [ + 390, + 70, + 492, + 83 + ], + "spans": [ + { + "bbox": [ + 390, + 70, + 492, + 83 + ], + "type": "text", + "content": "(b) Test-time Adaptation" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 328, + 87, + 553, + 241 + ], + "lines": [ + { + "bbox": [ + 328, + 87, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 328, + 87, + 553, + 241 + ], + "type": "image", + "image_path": "d83ad3298341204ba8fc517ea7f5d381d85159f0f684b99db962b341f96a7a50.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 251, + 555, + 339 + ], + "lines": [ + { + "bbox": [ + 54, + 251, + 555, + 339 + ], + "spans": [ + { + "bbox": [ + 54, + 251, + 555, + 339 + ], + "type": "text", + "content": "Figure 2. Overview of our proposed approach. (a) shows the proposed 3D inpainting pipeline. Specifically, a reference view is first inpainted, which will be used to reconstruct complete geometry, along with rendered multi-view depth maps. The inpainted reference image is then warped to other views, which are taken as input conditions for the multi-view refinement network to generate a set of view-consistent images. Afterward, we fine-tune the 3D scene to get the inpainted scene. (b) illustrates the detailed structure and training of the multi-view refinement network. Basically, we take the pre-trained image inpainting model, incorporate additional space-time attention layers, and exploit the test-time adaptation to fine-tune it into a multi-view refinement model. To generate a fine-tuning dataset, we synthesize warped images under the original captured scene, which has the ground-truth captured images. In detail, reference views are selected randomly, and we set random masks over the images, apply geometry jittering, and warp the reference images in other views." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 346, + 296, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 346, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 296, + 407 + ], + "type": "text", + "content": "optimization to update " + }, + { + "bbox": [ + 55, + 346, + 296, + 407 + ], + "type": "inline_equation", + "content": "p_m" + }, + { + "bbox": [ + 55, + 346, + 296, + 407 + ], + "type": "text", + "content": " with respect to " + }, + { + "bbox": [ + 55, + 346, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 55, + 346, + 296, + 407 + ], + "type": "text", + "content": " loss between " + }, + { + "bbox": [ + 55, + 346, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\{\\hat{M}\\}_{n=1}^N" + }, + { + "bbox": [ + 55, + 346, + 296, + 407 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 346, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\{M\\}_{n=1}^N" + }, + { + "bbox": [ + 55, + 346, + 296, + 407 + ], + "type": "text", + "content": ". Notably, other parameters remain unchanged during this procedure. Finally, a pruning operation is performed to produce the pruned scene " + }, + { + "bbox": [ + 55, + 346, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\mathrm{pruned}}" + }, + { + "bbox": [ + 55, + 346, + 296, + 407 + ], + "type": "text", + "content": " by thresholding " + }, + { + "bbox": [ + 55, + 346, + 296, + 407 + ], + "type": "inline_equation", + "content": "p_m" + }, + { + "bbox": [ + 55, + 346, + 296, + 407 + ], + "type": "text", + "content": " with a threshold " + }, + { + "bbox": [ + 55, + 346, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 346, + 296, + 407 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 55, + 346, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\tau = 0.4" + }, + { + "bbox": [ + 55, + 346, + 296, + 407 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 415, + 231, + 428 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 231, + 428 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 231, + 428 + ], + "type": "text", + "content": "4.2. Proposed 3D Inpainting Pipeline" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 434, + 296, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 296, + 626 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 296, + 626 + ], + "type": "text", + "content": "Given the pruned GS scene " + }, + { + "bbox": [ + 55, + 434, + 296, + 626 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\mathrm{pruned}}" + }, + { + "bbox": [ + 55, + 434, + 296, + 626 + ], + "type": "text", + "content": " as well as inpainting masks " + }, + { + "bbox": [ + 55, + 434, + 296, + 626 + ], + "type": "inline_equation", + "content": "\\{P_n\\}_{n=1}^N" + }, + { + "bbox": [ + 55, + 434, + 296, + 626 + ], + "type": "text", + "content": " indicating the regions to be inpainted for each view, we propose a novel 3D inpainting pipeline shown in Fig 2. We first render color images " + }, + { + "bbox": [ + 55, + 434, + 296, + 626 + ], + "type": "inline_equation", + "content": "\\{I_n'\\}_{n=1}^N" + }, + { + "bbox": [ + 55, + 434, + 296, + 626 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 434, + 296, + 626 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\mathrm{pruned}}" + }, + { + "bbox": [ + 55, + 434, + 296, + 626 + ], + "type": "text", + "content": " for each view and select a reference view with image " + }, + { + "bbox": [ + 55, + 434, + 296, + 626 + ], + "type": "inline_equation", + "content": "I_r' \\in \\{I_n'\\}_{n=1}^N" + }, + { + "bbox": [ + 55, + 434, + 296, + 626 + ], + "type": "text", + "content": ", which is then inpainted using a pre-trained inpainting model [22] to reference image " + }, + { + "bbox": [ + 55, + 434, + 296, + 626 + ], + "type": "inline_equation", + "content": "I_r^*" + }, + { + "bbox": [ + 55, + 434, + 296, + 626 + ], + "type": "text", + "content": ". We conduct geometry reconstruction to fill out the hole in 3D, which will be utilized to warp the inpainted contents in the reference image into the remaining views. We employ a multi-view refinement model to further refine the warped images, ensuring high fidelity and consistency, and then update the 3D scene. The model is fine-tuned from the pre-trained image inpainting model via a test-time adaptation strategy. The geometry reconstruction and multi-view refinement model are described as follows." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 646, + 198, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 646, + 198, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 646, + 198, + 658 + ], + "type": "text", + "content": "4.2.1 Geometry Reconstruction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "Following [11], we render a set of depth maps " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\{D_n\\}_{n=1}^N" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": " from the prune GS scene " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\mathrm{pruned}}" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": " and adopt a truncated signed distance function (TSDF) [6] technique to reconstruct the 3D mesh. However, " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\{D_n\\}_{n=1}^N" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": " only contributes" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 346, + 555, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 555, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 555, + 441 + ], + "type": "text", + "content": "to the reconstruction of the background region, as the depth values within their inpainting mask areas are invalid. To fill the masked region, we aim to generate an all-valid depth map " + }, + { + "bbox": [ + 313, + 346, + 555, + 441 + ], + "type": "inline_equation", + "content": "D_r^*" + }, + { + "bbox": [ + 313, + 346, + 555, + 441 + ], + "type": "text", + "content": " of the reference view by filling the masked areas in " + }, + { + "bbox": [ + 313, + 346, + 555, + 441 + ], + "type": "inline_equation", + "content": "D_r" + }, + { + "bbox": [ + 313, + 346, + 555, + 441 + ], + "type": "text", + "content": " based on information from " + }, + { + "bbox": [ + 313, + 346, + 555, + 441 + ], + "type": "inline_equation", + "content": "I_r^*" + }, + { + "bbox": [ + 313, + 346, + 555, + 441 + ], + "type": "text", + "content": ". The filled depth map should meet two key requirements: 1) it must be consistent with the reference image " + }, + { + "bbox": [ + 313, + 346, + 555, + 441 + ], + "type": "inline_equation", + "content": "I_r^*" + }, + { + "bbox": [ + 313, + 346, + 555, + 441 + ], + "type": "text", + "content": "; 2) it should seamlessly connect with the depth values outside the masked region." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 448, + 556, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 448, + 556, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 448, + 556, + 544 + ], + "type": "text", + "content": "We begin by exploiting a monocular depth estimator [37] to estimate depth " + }, + { + "bbox": [ + 313, + 448, + 556, + 544 + ], + "type": "inline_equation", + "content": "\\bar{D}_r" + }, + { + "bbox": [ + 313, + 448, + 556, + 544 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 448, + 556, + 544 + ], + "type": "inline_equation", + "content": "I_r^*" + }, + { + "bbox": [ + 313, + 448, + 556, + 544 + ], + "type": "text", + "content": ", followed by a linear transformation to convert " + }, + { + "bbox": [ + 313, + 448, + 556, + 544 + ], + "type": "inline_equation", + "content": "\\bar{D}_r" + }, + { + "bbox": [ + 313, + 448, + 556, + 544 + ], + "type": "text", + "content": " to scene scale " + }, + { + "bbox": [ + 313, + 448, + 556, + 544 + ], + "type": "inline_equation", + "content": "\\tilde{D}_r" + }, + { + "bbox": [ + 313, + 448, + 556, + 544 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 313, + 448, + 556, + 544 + ], + "type": "inline_equation", + "content": "\\tilde{D}_r \\colon \\tilde{D}_r = a * \\bar{D}_r + b" + }, + { + "bbox": [ + 313, + 448, + 556, + 544 + ], + "type": "text", + "content": ", where the parameters " + }, + { + "bbox": [ + 313, + 448, + 556, + 544 + ], + "type": "inline_equation", + "content": "a, b" + }, + { + "bbox": [ + 313, + 448, + 556, + 544 + ], + "type": "text", + "content": " are computed in the neighborhood of " + }, + { + "bbox": [ + 313, + 448, + 556, + 544 + ], + "type": "inline_equation", + "content": "P_r" + }, + { + "bbox": [ + 313, + 448, + 556, + 544 + ], + "type": "text", + "content": " to minimize the difference between " + }, + { + "bbox": [ + 313, + 448, + 556, + 544 + ], + "type": "inline_equation", + "content": "\\bar{D}_r" + }, + { + "bbox": [ + 313, + 448, + 556, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 448, + 556, + 544 + ], + "type": "inline_equation", + "content": "D_r" + }, + { + "bbox": [ + 313, + 448, + 556, + 544 + ], + "type": "text", + "content": " in advance, following [16]. We then develop a filling scheme according to " + }, + { + "bbox": [ + 313, + 448, + 556, + 544 + ], + "type": "inline_equation", + "content": "\\tilde{D}_r" + }, + { + "bbox": [ + 313, + 448, + 556, + 544 + ], + "type": "text", + "content": ". Formally, the filled values of " + }, + { + "bbox": [ + 313, + 448, + 556, + 544 + ], + "type": "inline_equation", + "content": "D_r^*" + }, + { + "bbox": [ + 313, + 448, + 556, + 544 + ], + "type": "text", + "content": " are computed as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 565, + 556, + 619 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 565, + 556, + 619 + ], + "spans": [ + { + "bbox": [ + 314, + 565, + 556, + 619 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} D _ {r} ^ {*} (x) = \\underset {D _ {r} ^ {*} (x)} {\\arg \\min } \\sum_ {x \\in P _ {r}} | | \\nabla D _ {r} ^ {*} (x) - \\nabla \\tilde {D} _ {r} (x) | | ^ {2} + \\lambda \\nabla D _ {r} ^ {*} (x) \\tag {2} \\\\ \\begin{array}{l} \\text {s . t .} D _ {r} ^ {*} (x) = D _ {r} (x), \\text {w h e r e} P _ {r} (x) = 0, \\end{array} \\\\ \\end{array}", + "image_path": "8290398a0e52e81a12b792dbafc87a5a4e1a2974f4225748efa0a8c6c8f39fec.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\nabla" + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": " denotes the Laplacian calculation and " + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": " represents the pixel coordinate. Intuitively, the first term encourages " + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "inline_equation", + "content": "D_r^*" + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": " to maintain a similar changing trend as " + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\tilde{D}_r" + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": ", while the second term regularizes " + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "inline_equation", + "content": "D_r^*" + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": " to be smooth and consistent with the depths outside the mask. As such, " + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "inline_equation", + "content": "D_r^*" + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": " can be used for the reconstruction of the NBS region." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26697" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 136, + 148 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 136, + 148 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 136, + 148 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 136, + 148 + ], + "type": "image", + "image_path": "88981aea9c4662bcea365b0224b5d7bbf5df389647a3280871185b14de82bfdb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 149, + 131, + 159 + ], + "lines": [ + { + "bbox": [ + 64, + 149, + 131, + 159 + ], + "spans": [ + { + "bbox": [ + 64, + 149, + 131, + 159 + ], + "type": "text", + "content": "(a) Reference Image" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 136, + 70, + 212, + 148 + ], + "blocks": [ + { + "bbox": [ + 136, + 70, + 212, + 148 + ], + "lines": [ + { + "bbox": [ + 136, + 70, + 212, + 148 + ], + "spans": [ + { + "bbox": [ + 136, + 70, + 212, + 148 + ], + "type": "image", + "image_path": "3ec1a30d6659afcffad3ba55f2073a41cdb4aa20caecd5365eb545408158abbd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 145, + 149, + 204, + 159 + ], + "lines": [ + { + "bbox": [ + 145, + 149, + 204, + 159 + ], + "spans": [ + { + "bbox": [ + 145, + 149, + 204, + 159 + ], + "type": "text", + "content": "(b) Warped Image" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 213, + 70, + 290, + 148 + ], + "blocks": [ + { + "bbox": [ + 213, + 70, + 290, + 148 + ], + "lines": [ + { + "bbox": [ + 213, + 70, + 290, + 148 + ], + "spans": [ + { + "bbox": [ + 213, + 70, + 290, + 148 + ], + "type": "image", + "image_path": "7378e1bf8d1d89f23736300e8914ce815fa40c5ab0ffec7fb51bd432026cd2a2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 222, + 149, + 281, + 159 + ], + "lines": [ + { + "bbox": [ + 222, + 149, + 281, + 159 + ], + "spans": [ + { + "bbox": [ + 222, + 149, + 281, + 159 + ], + "type": "text", + "content": "(c) Refined Image" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 159, + 295, + 214 + ], + "lines": [ + { + "bbox": [ + 55, + 159, + 295, + 214 + ], + "spans": [ + { + "bbox": [ + 55, + 159, + 295, + 214 + ], + "type": "text", + "content": "Figure 3. Illustration of the importance of multi-view refinement. The multi-view refinement model is proposed to tackle the warping artifacts with the example shown in (b), generating high-fidelity multi-view consistent images with one sampled view shown in (c)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 224, + 295, + 236 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 224, + 295, + 236 + ], + "spans": [ + { + "bbox": [ + 55, + 224, + 295, + 236 + ], + "type": "text", + "content": "4.2.2 Test-Time Adaptation for Multi-view Refinement" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 245, + 296, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 245, + 296, + 388 + ], + "spans": [ + { + "bbox": [ + 55, + 245, + 296, + 388 + ], + "type": "text", + "content": "We warp the inpainted contents from the reference view " + }, + { + "bbox": [ + 55, + 245, + 296, + 388 + ], + "type": "inline_equation", + "content": "I_{r}" + }, + { + "bbox": [ + 55, + 245, + 296, + 388 + ], + "type": "text", + "content": " to other views with the above restored geometry. Due to the imperfect geometry and the neglect of view-dependent effects, those warped contents contain artifacts, such as texture distortion and inconsistent appearance with the neighboring regions, especially when there are large view angle changes with respect to the reference view as shown in Fig 3. On the other hand, these warped contents have coarse-level consistency across views. To this end, we propose a multi-view refinement model that produces high-fidelity, consistent images conditioning on the warped images." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 389, + 296, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 389, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 296, + 544 + ], + "type": "text", + "content": "Network: Our model is built upon a pre-trained diffusion-based image inpainting model [22], which takes a masked image and its associated mask as the input, corresponding to the warped image and inpainting mask in our scenarios, respectively. To ensure multi-view consistency, we replace the self-attention layer in each transformer block with space-time attention to explore correlations across multiple views. Considering the quadratic complexity of global attention and the large number of warped images, space-time attention is implemented as the sparse space-time attention [10] that only attends to neighboring images. The attention layer is also allowed to attend to the reference image for alignment." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "text", + "content": "Training: Instead of relying on large datasets and extensive training, we exploit the test-time adaptation strategy for each scene to train the multi-view refinement network. Paired training sets are created for model adaptation. Specifically, given the original images " + }, + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\{I_n\\}_{n=1}^N" + }, + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "text", + "content": ", at each training step, " + }, + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "inline_equation", + "content": "L + 1" + }, + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "text", + "content": " images are randomly selected as ground truth. To simulate realistic warping scenarios, we randomly choose one image as the reference, generate a random mask, and warp both the reference image and mask to other views with geometry guidance from " + }, + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "text", + "content": ". In addition, to encourage cross-view learning as in [2], we also synthesize another set of paired images by applying image augmentations independently on each captured image, i.e., irregularly shaped masks that are randomly generated on" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 319, + 70, + 434, + 148 + ], + "blocks": [ + { + "bbox": [ + 319, + 70, + 434, + 148 + ], + "lines": [ + { + "bbox": [ + 319, + 70, + 434, + 148 + ], + "spans": [ + { + "bbox": [ + 319, + 70, + 434, + 148 + ], + "type": "image", + "image_path": "b0432dd7b66401940f44d0310e20a9b8bfb2459010e1664b471a952a376df89c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 435, + 70, + 551, + 148 + ], + "blocks": [ + { + "bbox": [ + 435, + 70, + 551, + 148 + ], + "lines": [ + { + "bbox": [ + 435, + 70, + 551, + 148 + ], + "spans": [ + { + "bbox": [ + 435, + 70, + 551, + 148 + ], + "type": "image", + "image_path": "7d5206d2cdc47c5069a14666acfbe0d2d3bd6a29a334ecf1653616312322ea5d.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 317, + 159, + 432, + 237 + ], + "blocks": [ + { + "bbox": [ + 322, + 150, + 430, + 159 + ], + "lines": [ + { + "bbox": [ + 322, + 150, + 430, + 159 + ], + "spans": [ + { + "bbox": [ + 322, + 150, + 430, + 159 + ], + "type": "text", + "content": "(a) Object mask and dilated mask" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 317, + 159, + 432, + 237 + ], + "lines": [ + { + "bbox": [ + 317, + 159, + 432, + 237 + ], + "spans": [ + { + "bbox": [ + 317, + 159, + 432, + 237 + ], + "type": "image", + "image_path": "eda211d0b1eb916ea42106ff58661e6affc2b6af5654d671649e248da47c942c.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 319, + 239, + 431, + 248 + ], + "lines": [ + { + "bbox": [ + 319, + 239, + 431, + 248 + ], + "spans": [ + { + "bbox": [ + 319, + 239, + 431, + 248 + ], + "type": "text", + "content": "(c) Mask-remapping interpretation" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 251, + 555, + 361 + ], + "lines": [ + { + "bbox": [ + 313, + 251, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 555, + 361 + ], + "type": "text", + "content": "Figure 4. Inpainting mask detection. (a) shows the object mask (orange) along with the dilated mask (pink), which includes additional neighboring pixels; Each dilated mask consists of two parts: the actual inpainting mask and out-of-interest areas, marked as regions " + }, + { + "bbox": [ + 313, + 251, + 555, + 361 + ], + "type": "inline_equation", + "content": "①" + }, + { + "bbox": [ + 313, + 251, + 555, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 251, + 555, + 361 + ], + "type": "inline_equation", + "content": "②" + }, + { + "bbox": [ + 313, + 251, + 555, + 361 + ], + "type": "text", + "content": " in (b). These regions are mapped in 3D to the NBS and background areas, respectively, as shown in regions " + }, + { + "bbox": [ + 313, + 251, + 555, + 361 + ], + "type": "inline_equation", + "content": "①" + }, + { + "bbox": [ + 313, + 251, + 555, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 251, + 555, + 361 + ], + "type": "inline_equation", + "content": "②" + }, + { + "bbox": [ + 313, + 251, + 555, + 361 + ], + "type": "text", + "content": " in (c). During multi-view mapping optimization, the central NBS region is progressively refined, while background areas are suppressed by each other; (d) shows the rendered neighbor mask after GS scene optimization." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 434, + 159, + 550, + 237 + ], + "blocks": [ + { + "bbox": [ + 448, + 150, + 537, + 159 + ], + "lines": [ + { + "bbox": [ + 448, + 150, + 537, + 159 + ], + "spans": [ + { + "bbox": [ + 448, + 150, + 537, + 159 + ], + "type": "text", + "content": "(b) Object Mask Separation" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 434, + 159, + 550, + 237 + ], + "lines": [ + { + "bbox": [ + 434, + 159, + 550, + 237 + ], + "spans": [ + { + "bbox": [ + 434, + 159, + 550, + 237 + ], + "type": "image", + "image_path": "67965b8e942fe8f3b2ea1f02e734edc57d6164253a69e28ba82e7ef84834dd19.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 462, + 239, + 521, + 247 + ], + "lines": [ + { + "bbox": [ + 462, + 239, + 521, + 247 + ], + "spans": [ + { + "bbox": [ + 462, + 239, + 521, + 247 + ], + "type": "text", + "content": "(d) Render mask" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 373, + 554, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 554, + 433 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 554, + 433 + ], + "type": "text", + "content": "each image. Image-based augmentations—such as elastic transformations and color jittering, are then applied to the masked regions to simulate warping artifacts. We train the multi-view refinement model with those paired images, employing a simplified variational bound objective [8]." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 434, + 554, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 554, + 481 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 554, + 481 + ], + "type": "text", + "content": "Inference: At inference, similar to [33], we randomly shuffle the order of the " + }, + { + "bbox": [ + 313, + 434, + 554, + 481 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 434, + 554, + 481 + ], + "type": "text", + "content": " warped images at each denosing step to break the sequential constraint of sparse attention, further enhancing cross-view consistency." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 498, + 432, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 432, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 432, + 509 + ], + "type": "text", + "content": "4.2.3 Fine-tune GS Scene" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 518, + 554, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 518, + 554, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 518, + 554, + 590 + ], + "type": "text", + "content": "With the generated multi-view consistent images, we fine-tune " + }, + { + "bbox": [ + 313, + 518, + 554, + 590 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\mathrm{pruned}}" + }, + { + "bbox": [ + 313, + 518, + 554, + 590 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 313, + 518, + 554, + 590 + ], + "type": "inline_equation", + "content": "K_{2}(K_{2} = 7000)" + }, + { + "bbox": [ + 313, + 518, + 554, + 590 + ], + "type": "text", + "content": " steps with an " + }, + { + "bbox": [ + 313, + 518, + 554, + 590 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 313, + 518, + 554, + 590 + ], + "type": "text", + "content": " loss and a SSIM loss as straightforward pixel-wise supervision. It is worth noting that other robust losses (e.g., perceptual loss) are functionally orthogonal to our approach and can be used to improve it further." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 598, + 462, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 598, + 462, + 611 + ], + "spans": [ + { + "bbox": [ + 313, + 598, + 462, + 611 + ], + "type": "text", + "content": "4.3. Inpainting Mask Detection" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 617, + 554, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 554, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 554, + 700 + ], + "type": "text", + "content": "In the overall pipeline of 3D object removal and scene inpainting, identifying the occluded or NBS regions for inpainting is also critical, especially for " + }, + { + "bbox": [ + 313, + 617, + 554, + 700 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 313, + 617, + 554, + 700 + ], + "type": "text", + "content": " cases. In those cases, the object mask often covers a much larger area than the regions that require inpainting (as illustrated in Fig. 4). Therefore, a novel inpainting mask detection approach is proposed to identify the NBS region for each view." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 326, + 701, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 554, + 713 + ], + "type": "text", + "content": "An intuitive idea is to detect the NBS region in the" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "26698" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 324 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 324 + ], + "type": "text", + "content": "pruned GS scene and project it back into each view to obtain inpainting masks. However, since, by definition, the NBS region is occluded by the object, it usually lacks good geometry and meaningful Gaussians, making it challenging to detect in 3D. Here we propose to detect the inpainting masks by leveraging neighbors of NBS region. In detail, we first dilate the object masks to include neighboring pixels (Fig. 4(a)), and then map the dilated masks onto the pruned scene " + }, + { + "bbox": [ + 55, + 72, + 294, + 324 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\mathrm{pruned}}" + }, + { + "bbox": [ + 55, + 72, + 294, + 324 + ], + "type": "text", + "content": ". Mapping is implemented as updating the learnable parameter " + }, + { + "bbox": [ + 55, + 72, + 294, + 324 + ], + "type": "inline_equation", + "content": "\\{p_m\\}_{m=1}^M" + }, + { + "bbox": [ + 55, + 72, + 294, + 324 + ], + "type": "text", + "content": " of each Gaussian with respect to " + }, + { + "bbox": [ + 55, + 72, + 294, + 324 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 55, + 72, + 294, + 324 + ], + "type": "text", + "content": " loss between rendered masks of " + }, + { + "bbox": [ + 55, + 72, + 294, + 324 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\mathrm{pruned}}" + }, + { + "bbox": [ + 55, + 72, + 294, + 324 + ], + "type": "text", + "content": " and the dilated masks. The optimization process is the same as object removal (Sec. 4.1) but with the object mask replaced with dilated masks, and the optimization is performed over the pruned scene. We illustrate the interpretation of this process in Fig. 4(b-c). After the mapping process, we are able to get the neighbor mask by rendering " + }, + { + "bbox": [ + 55, + 72, + 294, + 324 + ], + "type": "inline_equation", + "content": "\\{p_m\\}_{m=1}^M" + }, + { + "bbox": [ + 55, + 72, + 294, + 324 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 72, + 294, + 324 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\mathrm{pruned}}" + }, + { + "bbox": [ + 55, + 72, + 294, + 324 + ], + "type": "text", + "content": " as shown in Fig. 4(d). Finally, we extract the bounding box for the above rendered mask, which will be taken as input prompt for segment anything model [13] to get the final inpainting mask." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 336, + 136, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 336, + 136, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 336, + 136, + 350 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 357, + 295, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 357, + 295, + 583 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 295, + 583 + ], + "type": "text", + "content": "Datasets: Most existing publicly available 3D inpainting datasets [17, 34] are limited to front-facing cases, which cannot be used to evaluate methods in more diverse or realistic situations. To address this, we construct a dataset that contains 20 diverse cases, ranging from indoor to outdoor, covering various camera trajectories-including front-facing, " + }, + { + "bbox": [ + 55, + 357, + 295, + 583 + ], + "type": "inline_equation", + "content": "90^{\\circ}" + }, + { + "bbox": [ + 55, + 357, + 295, + 583 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 357, + 295, + 583 + ], + "type": "inline_equation", + "content": "180^{\\circ}" + }, + { + "bbox": [ + 55, + 357, + 295, + 583 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 357, + 295, + 583 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 55, + 357, + 295, + 583 + ], + "type": "text", + "content": ", and containing various interfacial surfaces such as single plane, multiple planes, curvature surfaces, and irregular surfaces. Each scene is recorded using a DJI Pocket 3 at 60fps with a resolution of " + }, + { + "bbox": [ + 55, + 357, + 295, + 583 + ], + "type": "inline_equation", + "content": "1920 \\times 1080" + }, + { + "bbox": [ + 55, + 357, + 295, + 583 + ], + "type": "text", + "content": ", from which we evenly extract 175 images and generate the corresponding camera poses and object masks. The first 50 images are used for evaluation purpose with the object removed, while the remaining 125 images are object-present and are used to reconstruct the original GS scene. Previews of our collected dataset are shown in supplementary materials. We use half resolution in the following experiments. In addition, we evaluate our method on SPINeRF dataset [17] to compare with state-of-the-art." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 585, + 295, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 585, + 295, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 585, + 295, + 693 + ], + "type": "text", + "content": "Implementation Details: StableDiffusionInpainting-Pipeline is utilized for image inpainting and our multi-view refinement model initialization. During training, we randomly select " + }, + { + "bbox": [ + 55, + 585, + 295, + 693 + ], + "type": "inline_equation", + "content": "L = 5" + }, + { + "bbox": [ + 55, + 585, + 295, + 693 + ], + "type": "text", + "content": " images from training images and crop " + }, + { + "bbox": [ + 55, + 585, + 295, + 693 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 55, + 585, + 295, + 693 + ], + "type": "text", + "content": " patches to fine-tune the model over 1000 iterations with the learning rate and batch size set as " + }, + { + "bbox": [ + 55, + 585, + 295, + 693 + ], + "type": "inline_equation", + "content": "3 \\times 10^{-5}" + }, + { + "bbox": [ + 55, + 585, + 295, + 693 + ], + "type": "text", + "content": " and 1, individually. At inference, directly refining all warped images, i.e., " + }, + { + "bbox": [ + 55, + 585, + 295, + 693 + ], + "type": "inline_equation", + "content": "L = 125" + }, + { + "bbox": [ + 55, + 585, + 295, + 693 + ], + "type": "text", + "content": " would result in prohibitive memory usage and computational costs. To" + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 315, + 104, + 553, + 169 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 555, + 103 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 555, + 103 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 555, + 103 + ], + "type": "text", + "content": "Table 1. Quantitative evaluation on the SPINeRF dataset and our dataset. Numbers in bold indicate the best performance, and underscored numbers indicate the second best." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 104, + 553, + 169 + ], + "lines": [ + { + "bbox": [ + 315, + 104, + 553, + 169 + ], + "spans": [ + { + "bbox": [ + 315, + 104, + 553, + 169 + ], + "type": "table", + "html": "
MethodOursSPINeRF
PSNR (↑)LPIPS (↓)FID (↓)PSNR (↑)LPIPS (↓)FID (↓)
GaussianEditor15.710.6163375.0314.410.6247343.16
GScream17.180.4431290.6316.960.3931154.71
SPInNeRF18.750.3519206.4317.470.5740239.99
MVPIneRF18.630.4332278.9917.670.5070255.51
Ours19.670.2685149.5217.580.4513154.34
", + "image_path": "5108cb807a4550bc3965e1a60d6873f7fc10edea040f0beeed2c73dffb72769a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 182, + 555, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 182, + 555, + 327 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 555, + 327 + ], + "type": "text", + "content": "address this, we cluster all views into " + }, + { + "bbox": [ + 313, + 182, + 555, + 327 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 182, + 555, + 327 + ], + "type": "text", + "content": " groups and select the center view of each group as the anchor. Multi-view refinement is applied only to these " + }, + { + "bbox": [ + 313, + 182, + 555, + 327 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 182, + 555, + 327 + ], + "type": "text", + "content": " anchor views, while the remaining views in each group are warped from their respective anchors, as artifacts caused by warping between closely aligned views are negligible. " + }, + { + "bbox": [ + 313, + 182, + 555, + 327 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 182, + 555, + 327 + ], + "type": "text", + "content": " is set to 12 for " + }, + { + "bbox": [ + 313, + 182, + 555, + 327 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 313, + 182, + 555, + 327 + ], + "type": "text", + "content": " cases and linearly scaled down for other scenarios, such as 4 for " + }, + { + "bbox": [ + 313, + 182, + 555, + 327 + ], + "type": "inline_equation", + "content": "90^{\\circ}" + }, + { + "bbox": [ + 313, + 182, + 555, + 327 + ], + "type": "text", + "content": " cases. We directly use object masks when evaluating on front-facing scenes and apply inpainting mask detection on others. All experiments are conducted on a single NVIDIA V100 GPU, with the inpainting of a 3D scene taking approximately one hour." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 335, + 545, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 545, + 348 + ], + "type": "text", + "content": "5.1. Evaluation against State-of-the-Art Methods" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 354, + 555, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 555, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 555, + 581 + ], + "type": "text", + "content": "We evaluate our method against the state-of-the-art methods on our dataset and SPINeRF dataset. The following state-of-the-art methods are reproduced on our datasets, including GaussianEditor [5], GScream [32], MVIPNeRF [3] and SPInNeRF [17] using their official implementations. In particular, since the naive NeRF models used in MVIPNeRF and SPInNeRF cannot handle complex unbounded scenes, we replace them with MeRF [21] considering both memory usage and reconstruction performance. On our dataset, we utilize the evaluation protocol and official implementation from [3] to compare methods. Specifically, we render color images at corresponding test views from the inpainted scenes and evaluate them against the ground truth using three metrics, namely PSNR, LPIPS, and FID. These metrics are calculated only within the bounding box defined by the object mask. For the SPINeRF dataset, we copy values from [3] for SPInNeRF and MVIPNeRF, and we reproduce GaussianEditor and GScream to align with the metric computation protocol in [3]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 582, + 554, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 554, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 554, + 676 + ], + "type": "text", + "content": "As can be seen from Tab. 1, the proposed method consistently achieves the best performance on our dataset, with a particularly large margin that demonstrates its superiority in unconstrained realistic scenarios. On the SPINeRF dataset, our method achieves state-of-the-art results comparable to other approaches, given that this dataset is limited to front-facing cases and challenging to distinguish methods in handling complex cases." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "In addition, we provide qualitative comparisons in Fig. 5, where our method successfully removes objects and seamlessly inpaints with appropriate content. Moreover, the in" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 56, + 702, + 258, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 702, + 258, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 702, + 258, + 713 + ], + "type": "text", + "content": "https://huggingface.co/stabilityai/stable-diffusion-2-inpainting" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26699" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 71, + 201, + 208 + ], + "blocks": [ + { + "bbox": [ + 65, + 71, + 201, + 208 + ], + "lines": [ + { + "bbox": [ + 65, + 71, + 201, + 208 + ], + "spans": [ + { + "bbox": [ + 65, + 71, + 201, + 208 + ], + "type": "image", + "image_path": "9fd45b9fab78c3e9a2527e9fd5205b01a9011690108551aff11c6a29a9b58681.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 203, + 72, + 270, + 207 + ], + "blocks": [ + { + "bbox": [ + 203, + 72, + 270, + 207 + ], + "lines": [ + { + "bbox": [ + 203, + 72, + 270, + 207 + ], + "spans": [ + { + "bbox": [ + 203, + 72, + 270, + 207 + ], + "type": "image", + "image_path": "01996cb03495c23b237d7263cf289df809e7e37ed041bcc0ea67549d142cb25c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 271, + 72, + 339, + 207 + ], + "blocks": [ + { + "bbox": [ + 271, + 72, + 339, + 207 + ], + "lines": [ + { + "bbox": [ + 271, + 72, + 339, + 207 + ], + "spans": [ + { + "bbox": [ + 271, + 72, + 339, + 207 + ], + "type": "image", + "image_path": "d368c73cf0a32cd2779b81ac836b1c506b37e07ec5402fa14be8bac51375afed.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 340, + 72, + 407, + 207 + ], + "blocks": [ + { + "bbox": [ + 340, + 72, + 407, + 207 + ], + "lines": [ + { + "bbox": [ + 340, + 72, + 407, + 207 + ], + "spans": [ + { + "bbox": [ + 340, + 72, + 407, + 207 + ], + "type": "image", + "image_path": "6dc19e7aa41c2ce6438f5942a191f0594835e39db08764b3172189bef8454922.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 408, + 72, + 476, + 207 + ], + "blocks": [ + { + "bbox": [ + 408, + 72, + 476, + 207 + ], + "lines": [ + { + "bbox": [ + 408, + 72, + 476, + 207 + ], + "spans": [ + { + "bbox": [ + 408, + 72, + 476, + 207 + ], + "type": "image", + "image_path": "29ef3461c4289f94df28cadce23062c5e7ed20c7975b04a2cc326109b2cb55bb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 477, + 72, + 545, + 207 + ], + "blocks": [ + { + "bbox": [ + 477, + 72, + 545, + 207 + ], + "lines": [ + { + "bbox": [ + 477, + 72, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 477, + 72, + 545, + 207 + ], + "type": "image", + "image_path": "393a64c93ad923a356538319f047929bf7b23d99bfdea56b7190f941e4b060bb.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 65, + 209, + 201, + 343 + ], + "blocks": [ + { + "bbox": [ + 65, + 209, + 201, + 343 + ], + "lines": [ + { + "bbox": [ + 65, + 209, + 201, + 343 + ], + "spans": [ + { + "bbox": [ + 65, + 209, + 201, + 343 + ], + "type": "image", + "image_path": "23b4d95936c79134a3ee59bae087a35ffbded0d3fb10a43fa17bb26dc2d8b3d4.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 125, + 345, + 141, + 354 + ], + "lines": [ + { + "bbox": [ + 125, + 345, + 141, + 354 + ], + "spans": [ + { + "bbox": [ + 125, + 345, + 141, + 354 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 203, + 209, + 270, + 275 + ], + "blocks": [ + { + "bbox": [ + 203, + 209, + 270, + 275 + ], + "lines": [ + { + "bbox": [ + 203, + 209, + 270, + 275 + ], + "spans": [ + { + "bbox": [ + 203, + 209, + 270, + 275 + ], + "type": "image", + "image_path": "064e37c5ac318744af35fac67b8967bc2c999c883a1dd6c0c3435861cfdf6272.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 203, + 276, + 270, + 343 + ], + "blocks": [ + { + "bbox": [ + 203, + 276, + 270, + 343 + ], + "lines": [ + { + "bbox": [ + 203, + 276, + 270, + 343 + ], + "spans": [ + { + "bbox": [ + 203, + 276, + 270, + 343 + ], + "type": "image", + "image_path": "e7e636c3576c2b4cd7ba19bd0ab1853f0199428da4111d1c7df7e29c8fc29e79.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 215, + 345, + 259, + 353 + ], + "lines": [ + { + "bbox": [ + 215, + 345, + 259, + 353 + ], + "spans": [ + { + "bbox": [ + 215, + 345, + 259, + 353 + ], + "type": "text", + "content": "GaussianEditor" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 272, + 209, + 339, + 275 + ], + "blocks": [ + { + "bbox": [ + 272, + 209, + 339, + 275 + ], + "lines": [ + { + "bbox": [ + 272, + 209, + 339, + 275 + ], + "spans": [ + { + "bbox": [ + 272, + 209, + 339, + 275 + ], + "type": "image", + "image_path": "b314c43523702de0c838c532eddedc7f8a75ff56a0a2971881d8efd640fe1726.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 272, + 276, + 339, + 343 + ], + "blocks": [ + { + "bbox": [ + 272, + 276, + 339, + 343 + ], + "lines": [ + { + "bbox": [ + 272, + 276, + 339, + 343 + ], + "spans": [ + { + "bbox": [ + 272, + 276, + 339, + 343 + ], + "type": "image", + "image_path": "7a0245a2d170bdb0759f41c5ea56c32bf28e4fb17832cc083355337da73668d9.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 293, + 345, + 317, + 352 + ], + "lines": [ + { + "bbox": [ + 293, + 345, + 317, + 352 + ], + "spans": [ + { + "bbox": [ + 293, + 345, + 317, + 352 + ], + "type": "text", + "content": "GSream" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 362, + 555, + 396 + ], + "lines": [ + { + "bbox": [ + 55, + 362, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 362, + 555, + 396 + ], + "type": "text", + "content": "Figure 5. Qualitative evaluation. We show the visual comparisons against the state-of-the-art 3D inpainting algorithms, with two views displayed at the top and bottom for each method. Our method generates consistent, higher-quality results with rich texture details and fewer artifacts." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 340, + 209, + 407, + 276 + ], + "blocks": [ + { + "bbox": [ + 340, + 209, + 407, + 276 + ], + "lines": [ + { + "bbox": [ + 340, + 209, + 407, + 276 + ], + "spans": [ + { + "bbox": [ + 340, + 209, + 407, + 276 + ], + "type": "image", + "image_path": "18057c39ebeec521382308ebb62748e9c88aa5cce12cb7fff284d36dd2d3b0a5.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 340, + 276, + 407, + 343 + ], + "blocks": [ + { + "bbox": [ + 340, + 276, + 407, + 343 + ], + "lines": [ + { + "bbox": [ + 340, + 276, + 407, + 343 + ], + "spans": [ + { + "bbox": [ + 340, + 276, + 407, + 343 + ], + "type": "image", + "image_path": "23c613235a9a7650ebcec634059754b0e36b1b7eb92a1a457601de563c784ad0.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 359, + 345, + 390, + 352 + ], + "lines": [ + { + "bbox": [ + 359, + 345, + 390, + 352 + ], + "spans": [ + { + "bbox": [ + 359, + 345, + 390, + 352 + ], + "type": "text", + "content": "SPInNeRF" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 408, + 209, + 476, + 276 + ], + "blocks": [ + { + "bbox": [ + 408, + 209, + 476, + 276 + ], + "lines": [ + { + "bbox": [ + 408, + 209, + 476, + 276 + ], + "spans": [ + { + "bbox": [ + 408, + 209, + 476, + 276 + ], + "type": "image", + "image_path": "76577a197a90ffb1f2c874fe28a04038d84e28882547323ff888a62e4d5627ae.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 408, + 276, + 476, + 343 + ], + "blocks": [ + { + "bbox": [ + 408, + 276, + 476, + 343 + ], + "lines": [ + { + "bbox": [ + 408, + 276, + 476, + 343 + ], + "spans": [ + { + "bbox": [ + 408, + 276, + 476, + 343 + ], + "type": "image", + "image_path": "aadc38e749d5d184ec8b39ccb4ea2f87f13adfe6bc75682c7cfc609f808ac06a.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 426, + 345, + 460, + 352 + ], + "lines": [ + { + "bbox": [ + 426, + 345, + 460, + 352 + ], + "spans": [ + { + "bbox": [ + 426, + 345, + 460, + 352 + ], + "type": "text", + "content": "MVIPNeRF" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 477, + 209, + 545, + 276 + ], + "blocks": [ + { + "bbox": [ + 477, + 209, + 545, + 276 + ], + "lines": [ + { + "bbox": [ + 477, + 209, + 545, + 276 + ], + "spans": [ + { + "bbox": [ + 477, + 209, + 545, + 276 + ], + "type": "image", + "image_path": "d1bcbcbe4d8838ccf3ed6c1bc363d868c93e33e11983ca4fc616ab8629507eac.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 477, + 276, + 545, + 343 + ], + "blocks": [ + { + "bbox": [ + 477, + 276, + 545, + 343 + ], + "lines": [ + { + "bbox": [ + 477, + 276, + 545, + 343 + ], + "spans": [ + { + "bbox": [ + 477, + 276, + 545, + 343 + ], + "type": "image", + "image_path": "f79a365c1e4dd18df55890efd61162f1add6aa610e4cb234eb03889fdcbfbcc3.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 503, + 345, + 519, + 352 + ], + "lines": [ + { + "bbox": [ + 503, + 345, + 519, + 352 + ], + "spans": [ + { + "bbox": [ + 503, + 345, + 519, + 352 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "bbox": [ + 55, + 403, + 296, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 296, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 296, + 475 + ], + "type": "text", + "content": "painted scene maintains 3D consistency, yielding visually more pleasing outcomes and a coherent geometric structure. In contrast, other methods often produce severe artifacts and fail to inpaint objects effectively in complex, realistic scenarios. We highly recommend viewing our video visualizations for a more comprehensive comparison." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 55, + 488, + 220, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 488, + 220, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 488, + 220, + 502 + ], + "type": "text", + "content": "5.2. Detection of Inpainting Masks" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 55, + 508, + 295, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 508, + 295, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 508, + 295, + 662 + ], + "type": "text", + "content": "We then demonstrate the importance of doing inpainting mask detection and differentiating them from object masks. For our collected dataset, we manually annotate inpainting masks from rendered images of pruned scenes and compare them with the original object masks. Since our dataset is derived from continuous capture videos, we annotate one out of every 10 images in each scene, resulting in a total of 260 annotated masks. Our analysis shows that the inpainting masks, on average, occupy only " + }, + { + "bbox": [ + 55, + 508, + 295, + 662 + ], + "type": "inline_equation", + "content": "50.78\\%" + }, + { + "bbox": [ + 55, + 508, + 295, + 662 + ], + "type": "text", + "content": " of the area covered by the object masks. This finding highlights that adopting entire object masks requires inpainting nearly twice the necessary area, complicating the task and leading to suboptimal results." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "We also compare our detected inpainting masks with the manually labeled masks using the mean Intersection over Union (mIoU) metric. Our method achieves an " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "81.12\\%" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " IoU score, validating the accuracy and robustness of the pro" + } + ] + } + ], + "index": 27 + }, + { + "type": "table", + "bbox": [ + 316, + 414, + 553, + 487 + ], + "blocks": [ + { + "bbox": [ + 342, + 402, + 526, + 412 + ], + "lines": [ + { + "bbox": [ + 342, + 402, + 526, + 412 + ], + "spans": [ + { + "bbox": [ + 342, + 402, + 526, + 412 + ], + "type": "text", + "content": "Table 2. Quantitative results of ablation studies." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 414, + 553, + 487 + ], + "lines": [ + { + "bbox": [ + 316, + 414, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 414, + 553, + 487 + ], + "type": "table", + "html": "
MethodPSNR (↑)LPIPS (↓)FID (↓)
w/o Warping17.850.3215198.24
w/o Refinement18.900.3069206.96
General Refinement19.080.2719165.80
Single-View Refinement19.460.2725154.33
Multi-View Refinement (Ours)19.670.2685149.52
", + "image_path": "a8815df0819bbef22fda360c24a5094f9c989ce7146efcc4c83261e70d62cecf.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "table_body" + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 496, + 555, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 496, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 496, + 555, + 533 + ], + "type": "text", + "content": "posed inpainting mask detection strategy. In contrast, the method proposed in [5] achieves only " + }, + { + "bbox": [ + 313, + 496, + 555, + 533 + ], + "type": "inline_equation", + "content": "42.55\\%" + }, + { + "bbox": [ + 313, + 496, + 555, + 533 + ], + "type": "text", + "content": ", which is significantly inferior to our approach." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 539, + 408, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 539, + 408, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 539, + 408, + 552 + ], + "type": "text", + "content": "5.3. Ablation Study" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 558, + 553, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 553, + 581 + ], + "type": "text", + "content": "Our dataset is chosen for the ablation study due to its diversity and complexity." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "Effectiveness of geometry-guided warping: With image rendering of the pruned scene and inpainting masks, we adopt a warping-refinement scheme rather than directly inpainting the masked areas. To demonstrate the effectiveness of our approach, we disable geometry reconstruction and warping operations, repurposing the multi-view refinement model as a multi-view inpainting model. Specifically, we zero out the masked regions in the input images and attempt to consistently inpaint the missing areas. The inpainting model is also fine-tuned for each given scene, with all other settings remaining unchanged." + } + ] + } + ], + "index": 33 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26700" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": "We compare images generated with and without the warping process in Fig. 6. Despite the strength of multiview architecture, direct inpainting fails to generate consistent output with the reference image in large view-change cases, which aligns with the observation in [2]. In contrast, our warping-refinement scheme provides a significant advantage, as the warped content offers strong guidance and regularizes the model to generate consistent results. We then use the generated images to fine-tune 3D scene. As shown in Tab. 2, the fine-tuned scene indeed shows inferior results to our complete method." + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 214, + 136, + 291 + ], + "blocks": [ + { + "bbox": [ + 59, + 214, + 136, + 291 + ], + "lines": [ + { + "bbox": [ + 59, + 214, + 136, + 291 + ], + "spans": [ + { + "bbox": [ + 59, + 214, + 136, + 291 + ], + "type": "image", + "image_path": "31be7770edf282eb21ab19ed4c49a286db472285d9d90a9d146ff167fa8fc12c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 293, + 132, + 301 + ], + "lines": [ + { + "bbox": [ + 64, + 293, + 132, + 301 + ], + "spans": [ + { + "bbox": [ + 64, + 293, + 132, + 301 + ], + "type": "text", + "content": "(a) Reference Image" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 136, + 214, + 212, + 291 + ], + "blocks": [ + { + "bbox": [ + 136, + 214, + 212, + 291 + ], + "lines": [ + { + "bbox": [ + 136, + 214, + 212, + 291 + ], + "spans": [ + { + "bbox": [ + 136, + 214, + 212, + 291 + ], + "type": "image", + "image_path": "786e7dac24072b7b9e9074c2f48bc737797f3ead5c09d3167dd3ac1a133333f2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 147, + 293, + 202, + 301 + ], + "lines": [ + { + "bbox": [ + 147, + 293, + 202, + 301 + ], + "spans": [ + { + "bbox": [ + 147, + 293, + 202, + 301 + ], + "type": "text", + "content": "(b) W/o warping" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 212, + 214, + 290, + 291 + ], + "blocks": [ + { + "bbox": [ + 212, + 214, + 290, + 291 + ], + "lines": [ + { + "bbox": [ + 212, + 214, + 290, + 291 + ], + "spans": [ + { + "bbox": [ + 212, + 214, + 290, + 291 + ], + "type": "image", + "image_path": "fb0dd3e8044c742661fe54d5d481ec128a757f18579a0a7d0120ede43df40e9a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 227, + 293, + 277, + 301 + ], + "lines": [ + { + "bbox": [ + 227, + 293, + 277, + 301 + ], + "spans": [ + { + "bbox": [ + 227, + 293, + 277, + 301 + ], + "type": "text", + "content": "(c) W/ warping" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 303, + 296, + 358 + ], + "lines": [ + { + "bbox": [ + 55, + 303, + 296, + 358 + ], + "spans": [ + { + "bbox": [ + 55, + 303, + 296, + 358 + ], + "type": "text", + "content": "Figure 6. Multi-view refinement with and without conditioning on the warped image. (a) shows the inpainted reference with the originally captured image shown in the top right corner. (c) and (b) are the images from an alternate view generated with and without having the warped image as a condition, respectively." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 61, + 383, + 137, + 458 + ], + "blocks": [ + { + "bbox": [ + 61, + 383, + 137, + 458 + ], + "lines": [ + { + "bbox": [ + 61, + 383, + 137, + 458 + ], + "spans": [ + { + "bbox": [ + 61, + 383, + 137, + 458 + ], + "type": "image", + "image_path": "ce6aae3bb56c4389b91fd14d9a46bdd521928b7792f7aa49fcdcb6019b69f5a9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 66, + 460, + 132, + 469 + ], + "lines": [ + { + "bbox": [ + 66, + 460, + 132, + 469 + ], + "spans": [ + { + "bbox": [ + 66, + 460, + 132, + 469 + ], + "type": "text", + "content": "(a) Reference Image" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 138, + 383, + 213, + 458 + ], + "blocks": [ + { + "bbox": [ + 138, + 383, + 213, + 458 + ], + "lines": [ + { + "bbox": [ + 138, + 383, + 213, + 458 + ], + "spans": [ + { + "bbox": [ + 138, + 383, + 213, + 458 + ], + "type": "image", + "image_path": "814bec8e5c19ce142250ffa0aab31261da8dc5b3dc45571a1dcd9b60190fe916.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 153, + 460, + 199, + 469 + ], + "lines": [ + { + "bbox": [ + 153, + 460, + 199, + 469 + ], + "spans": [ + { + "bbox": [ + 153, + 460, + 199, + 469 + ], + "type": "text", + "content": "(b) W/o refine" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 214, + 383, + 290, + 458 + ], + "blocks": [ + { + "bbox": [ + 214, + 383, + 290, + 458 + ], + "lines": [ + { + "bbox": [ + 214, + 383, + 290, + 458 + ], + "spans": [ + { + "bbox": [ + 214, + 383, + 290, + 458 + ], + "type": "image", + "image_path": "cbc5d993772d0903e028951f23ad5d32877d26246c8252bc227f1bcdad516912.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 59, + 470, + 137, + 545 + ], + "blocks": [ + { + "bbox": [ + 59, + 470, + 137, + 545 + ], + "lines": [ + { + "bbox": [ + 59, + 470, + 137, + 545 + ], + "spans": [ + { + "bbox": [ + 59, + 470, + 137, + 545 + ], + "type": "image", + "image_path": "023687d7fe1395b1155c79244c6b5d27823535b42f1100a1f0fb33cc6b56721d.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 548, + 131, + 556 + ], + "lines": [ + { + "bbox": [ + 64, + 548, + 131, + 556 + ], + "spans": [ + { + "bbox": [ + 64, + 548, + 131, + 556 + ], + "type": "text", + "content": "(d) Reference image" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 558, + 296, + 624 + ], + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 624 + ], + "type": "text", + "content": "Figure 7. Ablation on multi-view refinement. (a) and (d) show the inpainted reference images with the originally captured images shown in the top right corner; (b) and (c) are the images from an alternate view generated without and with the refinement network, respectively; (e) and (f) are the images generated by the single-view and multi-view refinement model, respectively." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 138, + 470, + 212, + 545 + ], + "blocks": [ + { + "bbox": [ + 138, + 470, + 212, + 545 + ], + "lines": [ + { + "bbox": [ + 138, + 470, + 212, + 545 + ], + "spans": [ + { + "bbox": [ + 138, + 470, + 212, + 545 + ], + "type": "image", + "image_path": "83eb1c12036272c5fd582894d9047e73830a5a494243daea5d29337264102b98.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 139, + 548, + 209, + 556 + ], + "lines": [ + { + "bbox": [ + 139, + 548, + 209, + 556 + ], + "spans": [ + { + "bbox": [ + 139, + 548, + 209, + 556 + ], + "type": "text", + "content": "(e) Single-view refine" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 214, + 470, + 288, + 545 + ], + "blocks": [ + { + "bbox": [ + 219, + 460, + 287, + 469 + ], + "lines": [ + { + "bbox": [ + 219, + 460, + 287, + 469 + ], + "spans": [ + { + "bbox": [ + 219, + 460, + 287, + 469 + ], + "type": "text", + "content": "(c) Multi-view refine" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 214, + 470, + 288, + 545 + ], + "lines": [ + { + "bbox": [ + 214, + 470, + 288, + 545 + ], + "spans": [ + { + "bbox": [ + 214, + 470, + 288, + 545 + ], + "type": "image", + "image_path": "5784a93a682cefca5911ade79761263de7a46da5608ad1199fafc847b5350a34.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 218, + 548, + 285, + 556 + ], + "lines": [ + { + "bbox": [ + 218, + 548, + 285, + 556 + ], + "spans": [ + { + "bbox": [ + 218, + 548, + 285, + 556 + ], + "type": "text", + "content": "(f) Multi-view refine" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": "Effectiveness of multi-view refinement model: We design the following two variants to illustrate the importance of the multi-view refinement model. a) w/o refinement: the refinement model is removed, and the warped images are directly used to fine-tune the pruned scene. b) single view: we replace the space-time attention with the original self-attention layer, performing refinement independently for" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "type": "text", + "content": "each view without leveraging correlations across different views. When fine-tuning the single-view model for a given scene, only random image pairs are used." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 109, + 554, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 109, + 554, + 215 + ], + "spans": [ + { + "bbox": [ + 313, + 109, + 554, + 215 + ], + "type": "text", + "content": "Similarly, we compare images generated by these variants in Fig. 7. As seen in (a), (b), and (c), warping indeed introduces artifacts and texture mismatches, while the multiview refinement model effectively addresses them. In (d), (e), and (f), the single-view refinement model fails to generate results consistent with the reference image, whereas the multi-view model achieves consistency. We further fine-tune the 3D scene with corresponding images. As illustrated in Tab. 2, our completed method achieves the best result." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 217, + 555, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 555, + 396 + ], + "type": "text", + "content": "Necessity of test-time adaption: Our pipeline requires a per-scene fine-tuned refinement model to generate images. To demonstrate its necessity, we also train a generalized refinement model on the DI3DV dataset, a scene-level multiview dataset, where we randomly selected two views as the reference and target, synthesized warping artifacts on one, and trained the model to restore. However, as shown in Tab. 2, its performance is inferior to our proposed method due to the limited size of the training set and domain gaps between the training set and real-world cases, such as differences in the severity of artifacts/mismatch. In addition, we use an existing inpainting method, LeftRefill [1], to directly inpaint images based on the reference image. However, as shown in Fig.8, LeftRefill can only ensures overall visual similarity but lacks pixel-level consistency." + } + ] + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 320, + 407, + 395, + 451 + ], + "blocks": [ + { + "bbox": [ + 320, + 407, + 395, + 451 + ], + "lines": [ + { + "bbox": [ + 320, + 407, + 395, + 451 + ], + "spans": [ + { + "bbox": [ + 320, + 407, + 395, + 451 + ], + "type": "image", + "image_path": "870c58f54a8262529f9431a97220e7520dad64a5d07280ab999bae359cf02f09.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 453, + 380, + 460 + ], + "lines": [ + { + "bbox": [ + 335, + 453, + 380, + 460 + ], + "spans": [ + { + "bbox": [ + 335, + 453, + 380, + 460 + ], + "type": "text", + "content": "(a) Reference" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 358, + 463, + 509, + 474 + ], + "lines": [ + { + "bbox": [ + 358, + 463, + 509, + 474 + ], + "spans": [ + { + "bbox": [ + 358, + 463, + 509, + 474 + ], + "type": "text", + "content": "Figure 8. Inpainting results by LeftRefill." + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 397, + 407, + 473, + 451 + ], + "blocks": [ + { + "bbox": [ + 397, + 407, + 473, + 451 + ], + "lines": [ + { + "bbox": [ + 397, + 407, + 473, + 451 + ], + "spans": [ + { + "bbox": [ + 397, + 407, + 473, + 451 + ], + "type": "image", + "image_path": "4a4cd5a62f73ec830ab653a961ee92f71c33b28911ef2781319a1231865d1d72.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 414, + 453, + 454, + 460 + ], + "lines": [ + { + "bbox": [ + 414, + 453, + 454, + 460 + ], + "spans": [ + { + "bbox": [ + 414, + 453, + 454, + 460 + ], + "type": "text", + "content": "(b) 1st View" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 473, + 407, + 549, + 451 + ], + "blocks": [ + { + "bbox": [ + 473, + 407, + 549, + 451 + ], + "lines": [ + { + "bbox": [ + 473, + 407, + 549, + 451 + ], + "spans": [ + { + "bbox": [ + 473, + 407, + 549, + 451 + ], + "type": "image", + "image_path": "330676a2728fdf714710c5f254163437544d0ae9e578ceeab4b84dd2bd8e803a.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 490, + 453, + 533, + 460 + ], + "lines": [ + { + "bbox": [ + 490, + 453, + 533, + 460 + ], + "spans": [ + { + "bbox": [ + 490, + 453, + 533, + 460 + ], + "type": "text", + "content": "(c) 2nd View" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "bbox": [ + 314, + 485, + 467, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 485, + 467, + 498 + ], + "spans": [ + { + "bbox": [ + 314, + 485, + 467, + 498 + ], + "type": "text", + "content": "6. Conclusion and Limitation" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 506, + 554, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 554, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 554, + 624 + ], + "type": "text", + "content": "We present a novel 3D inpainting pipeline that achieves state-of-the-art results across both front-facing and unconstrained benchmarks, with particularly strong performance on the latter. We achieve this by utilizing geometry priors and a multi-view refinement network to generate multi-view consistent inpainted images. Additionally, a novel inpainting mask detection technique is developed to boost performance in unconstrained scenes. The effectiveness of our method is demonstrated through extensive experiments and ablation studies." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 313, + 625, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 625, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 625, + 554, + 713 + ], + "type": "text", + "content": "Limitations: The multi-view refinement network requires customization for each scene, which may take considerable time. Future advancements in parameter-efficient fine-tuning and the personalization of diffusion models may help to accelerate and further optimize our approach. In addition, our framework relies on monocular depth estimation and image segmentation modules. The accuracy and robustness of these modules can affect the overall performance." + } + ] + } + ], + "index": 34 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26701" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Chenjie Cao, Yunuo Cai, Qiaole Dong, Yikai Wang, and Yanwei Fu. Leftrightfill: Filling right canvas based on left reference through generalized text-to-image diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7705-7715, 2024." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 146, + 296, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 146, + 296, + 190 + ], + "spans": [ + { + "bbox": [ + 61, + 146, + 296, + 190 + ], + "type": "text", + "content": "[2] Chenjie Cao, Chaohui Yu, Yanwei Fu, Fan Wang, and Xi-angyang Xue. Mvinpainter: Learning multi-view consistent inpainting to bridge 2d and 3d editing. arXiv preprint arXiv:2408.08000, 2024." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 191, + 295, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 191, + 295, + 245 + ], + "spans": [ + { + "bbox": [ + 61, + 191, + 295, + 245 + ], + "type": "text", + "content": "[3] Honghua Chen, Chen Change Loy, and Xingang Pan. Mvipnerf: Multi-view 3d inpainting on nerf scenes via diffusion prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5344-5353, 2024." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 246, + 295, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 246, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 61, + 246, + 295, + 300 + ], + "type": "text", + "content": "[4] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. Anydoor: Zero-shot object-level image customization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6593-6602, 2024." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 301, + 296, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 301, + 296, + 367 + ], + "spans": [ + { + "bbox": [ + 61, + 301, + 296, + 367 + ], + "type": "text", + "content": "[5] Yiwen Chen, Zilong Chen, Chi Zhang, Feng Wang, Xiaofeng Yang, Yikai Wang, Zhongang Cai, Lei Yang, Huaping Liu, and Guosheng Lin. Gaussianeditor: Swift and controllable 3d editing with gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21476-21485, 2024." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 369, + 295, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 369, + 295, + 413 + ], + "spans": [ + { + "bbox": [ + 61, + 369, + 295, + 413 + ], + "type": "text", + "content": "[6] Brian Curless and Marc Levoy. A volumetric method for building complex models from range images. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 303-312, 1996." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 413, + 294, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 413, + 294, + 435 + ], + "spans": [ + { + "bbox": [ + 62, + 413, + 294, + 435 + ], + "type": "text", + "content": "[7] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 435, + 294, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 435, + 294, + 468 + ], + "spans": [ + { + "bbox": [ + 62, + 435, + 294, + 468 + ], + "type": "text", + "content": "[8] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 469, + 294, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 469, + 294, + 524 + ], + "spans": [ + { + "bbox": [ + 62, + 469, + 294, + 524 + ], + "type": "text", + "content": "[9] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 525, + 295, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 525, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 57, + 525, + 295, + 567 + ], + "type": "text", + "content": "[10] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 569, + 295, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 569, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 57, + 569, + 295, + 613 + ], + "type": "text", + "content": "[11] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 614, + 295, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 614, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 57, + 614, + 295, + 656 + ], + "type": "text", + "content": "[12] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "type": "text", + "content": "[13] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "text", + "content": "[14] Chieh Hubert Lin, Changil Kim, Jia-Bin Huang, Qinbo Li, Chih-Yao Ma, Johannes Kopf, Ming-Hsuan Yang, and Hung-Yu Tseng. Taming latent diffusion model for neural radiance field inpainting. In European Conference on Computer Vision, pages 149-165. Springer, 2025." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 128, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 128, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 128, + 553, + 172 + ], + "type": "text", + "content": "[15] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453, 2023." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 172, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 172, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 172, + 553, + 239 + ], + "type": "text", + "content": "[16] Ashkan Mirzaei, Tristan Aumentado-Armstrong, Marcus A Brubaker, Jonathan Kelly, Alex Levinshtein, Konstantinos G Derpanis, and Igor Gilitschenski. Reference-guided controllable inpainting of neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17815-17825, 2023." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 239, + 553, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 314 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 314 + ], + "type": "text", + "content": "[17] Ashkan Mirzaei, Tristan Aumentado-Armstrong, Konstantinos G Derpanis, Jonathan Kelly, Marcus A Brubaker, Igor Gilitschenski, and Alex Levinshtein. Spin-nerf: Multiview segmentation and perceptual inpainting with neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20669–20679, 2023." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 316, + 553, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 553, + 370 + ], + "type": "text", + "content": "[18] Ashkan Mirzaei, Riccardo De Lutio, Seung Wook Kim, David Acuna, Jonathan Kelly, Sanja Fidler, Igor Gilitschenski, and Zan Gojcic. Refusion: Reference adapted diffusion models for 3d scene inpainting. arXiv preprint arXiv:2404.10765, 2024." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 371, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 371, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 371, + 553, + 426 + ], + "type": "text", + "content": "[19] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4296–4304, 2024." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 426, + 553, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 553, + 459 + ], + "type": "text", + "content": "[20] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 459, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 459, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 459, + 553, + 514 + ], + "type": "text", + "content": "[21] Christian Reiser, Rick Szeliski, Dor Verbin, Pratul Srinivasan, Ben Mildenhall, Andreas Geiger, Jon Barron, and Peter Hedman. Merf: Memory-efficient radiance fields for real-time view synthesis in unbounded scenes. ACM Transactions on Graphics (TOG), 42(4):1-12, 2023." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 514, + 553, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 514, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 316, + 514, + 553, + 569 + ], + "type": "text", + "content": "[22] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 569, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 569, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 569, + 553, + 635 + ], + "type": "text", + "content": "[23] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 635, + 553, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 680 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 680 + ], + "type": "text", + "content": "[24] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123: a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110, 2023." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[25] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023." + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26702" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "type": "text", + "content": "[26] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 130, + 294, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 130, + 294, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 130, + 294, + 161 + ], + "type": "text", + "content": "[27] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 164, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 164, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 294, + 239 + ], + "type": "text", + "content": "[28] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2149–2159, 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 243, + 294, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 294, + 296 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 294, + 296 + ], + "type": "text", + "content": "[29] Luming Tang, Nataniel Ruiz, Qinghao Chu, Yuanzhen Li, Aleksander Holynski, David E Jacobs, Bharath Hariharan, Yael Pritch, Neal Wadhwa, Kfir Aberman, et al. Realfill: Reference-driven generation for authentic image completion. ACM Transactions on Graphics (TOG), 43(4):1-12, 2024." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 298, + 294, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 298, + 294, + 352 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 294, + 352 + ], + "type": "text", + "content": "[30] Dongqing Wang, Tong Zhang, Alaa Abboud, and Sabine Susstrunk. Innerf360: Text-guided 3d-consistent object inpainting on 360-degree neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12677-12686, 2024." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 354, + 294, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 294, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 294, + 386 + ], + "type": "text", + "content": "[31] Peng Wang and Yichun Shi. Imagedream: Image-prompt multi-view diffusion for 3d generation. arXiv preprint arXiv:2312.02201, 2023." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 389, + 294, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 389, + 294, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 389, + 294, + 431 + ], + "type": "text", + "content": "[32] Yuxin Wang, Qianyi Wu, Guofeng Zhang, and Dan Xu. Gscream: Learning 3d geometry and feature consistent gaussian splatting for object removal. arXiv preprint arXiv:2404.13679, 2024." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 434, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 294, + 498 + ], + "type": "text", + "content": "[33] Ethan Weber, Aleksander Holynski, Varun Jampani, Saurabh Saxena, Noah Snavely, Abhishek Kar, and Angjoo Kanazawa. Nerfiller: Completing scenes via generative 3d inpainting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20731-20741, 2024." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 501, + 294, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 294, + 565 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 294, + 565 + ], + "type": "text", + "content": "[34] Silvan Weder, Guillermo Garcia-Hernando, Aron Monsz-part, Marc Pollefeys, Gabriel J Brostow, Michael Firman, and Sara Vicente. Removing objects from neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16528-16538, 2023." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 568, + 294, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 294, + 633 + ], + "type": "text", + "content": "[35] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7623-7633, 2023." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 635, + 294, + 699 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 699 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 699 + ], + "type": "text", + "content": "[36] Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen, and Fang Wen. Paint by example: Exemplar-based image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18381-18391, 2023." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 703, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 703, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 703, + 294, + 713 + ], + "type": "text", + "content": "[37] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 228 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In CVPR, 2024." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 97, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 553, + 139 + ], + "type": "text", + "content": "[38] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "type": "text", + "content": "[39] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 186, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 228 + ], + "type": "text", + "content": "[40] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "26703" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/INFP_ Audio-Driven Interactive Head Generation in Dyadic Conversations/4440ddb1-10e9-45a7-942b-205ac25fbaaa_content_list.json b/2025/INFP_ Audio-Driven Interactive Head Generation in Dyadic Conversations/4440ddb1-10e9-45a7-942b-205ac25fbaaa_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c7d42782d4329eac52c9384b2ff3ee8b308e6bec --- /dev/null +++ b/2025/INFP_ Audio-Driven Interactive Head Generation in Dyadic Conversations/4440ddb1-10e9-45a7-942b-205ac25fbaaa_content_list.json @@ -0,0 +1,1347 @@ +[ + { + "type": "text", + "text": "INFP: Audio-Driven Interactive Head Generation in Dyadic Conversations", + "text_level": 1, + "bbox": [ + 122, + 130, + 875, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yongming Zhu* Longhao Zhang* Zhengkun Rong* Tianshu Hu*† Shuang Liang Zhipeng Ge Bytedance", + "bbox": [ + 101, + 180, + 895, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zhuyongming, zhanglonghao.zlh, rongzhengkun, tianshu.hu, liangshuang. echo, zhipengge}@bytedance.com", + "bbox": [ + 245, + 219, + 746, + 251 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4e20afd929fea7b3889fc607e28cc363064719c32fbc24bc1a901cd98e39189c.jpg", + "image_caption": [ + "Figure 1. We present INFP, an audio-driven interactive head generation framework for dyadic conversations. Given the dual-track audio in dyadic conversations and a single portrait image of arbitrary agent, our framework can dynamically synthesize verbal, non-verbal and interactive agent videos with lifelike facial expressions and rhythmic head pose movements. Additionally, our framework is lightweight yet powerful, making it practical in instant communication scenarios with acceptable latency, such as the video conferencing. INFP denotes our method is Interactive, Natural, Flash and Person-generic." + ], + "image_footnote": [], + "bbox": [ + 99, + 273, + 903, + 508 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 599, + 326, + 616 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Imagine having a conversation with a socially intelligent agent. It can attentively listen to your words and offer visual and linguistic feedback promptly. This seamless interaction allows for multiple rounds of conversation to flow smoothly and naturally. In pursuit of actualizing it, we propose INFP, a novel audio-driven head generation framework for dyadic interaction. Unlike previous head generation works that only focus on single-sided communication, or require manual role assignment and explicit role switching, our model drives the agent portrait dynamically alternates between speaking and listening state, guided by the input dyadic audio. Specifically, INFP comprises a Motion-Based Head Imitation stage and an Audio-Guided Motion Generation stage. The first stage learns to project facial communicative behaviors from real-life conversation videos into a low-", + "bbox": [ + 88, + 633, + 485, + 861 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "dimensional motion latent space, and use the motion latent codes to animate a static image. The second stage learns the mapping from the input dyadic audio to motion latent codes through denoising, leading to the audio-driven head generation in interactive scenarios. To facilitate this line of research, we introduce DyConv, a large scale dataset of rich dyadic conversations collected from the Internet. Extensive experiments and visualizations demonstrate superior performance and effectiveness of our method.", + "bbox": [ + 511, + 601, + 908, + 739 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 768, + 645, + 784 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, researchers have dedicated considerable attention to audio-driven head generation [9, 18, 23, 24, 31, 37, 39, 42, 46, 56, 59], in order to build the conversational agent. However, most studies only focus on single-sided communication, such as talking or listening, ignoring the dyadic properties in human-human interaction. Talking-head generation [2, 17, 27, 36, 37, 41, 45, 46, 48, 49, 52, 56,", + "bbox": [ + 511, + 794, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 107, + 875, + 220, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding authors.", + "bbox": [ + 109, + 887, + 241, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "10667", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "60] aims at synthesizing speaker's facial animations based on their reference images and a driving audio. Although these works can produce vibrant videos with accurate lip synchronization, they solely emphasize the speaker's role and overlook the listener's feedback. Listening-head generation [16, 18, 23, 24, 31, 32, 55, 58], on the other hand, intends to respond to speaker's behaviors. However, these works restrict the listener's response solely to non-verbal facial movements, which is clearly far from real-life interaction scenarios. Some recent studies [39, 42, 59] have started exploring head generation for dyadic interaction, but they require manual role assignment between Listener and Speaker and cannot achieve smooth and natural role switching. We believe that for a conversational agent, it should not be predetermined a role, but rather be able to freely switch between listening and speaking states based on the current dialogue context. When having a conversation with the agent, you may interrupt or respond to it at any moment. At that point, it naturally transitions from a speaker to a listener, and vice versa.", + "bbox": [ + 89, + 90, + 483, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a novel audio-driven head generation framework for dyadic interaction, namely INFP. Unlike previous works, we do not divide the model into Speaker Generator and Listener Generator, or apply explicit role switching. The driven individual dynamically alternates between the speaking and listening state, guided by the dyadic audio. Specifically, INFP comprises a Motion-Based Head Imitation stage and an Audio-Guided Motion Generation stage. In the first stage, our model learns to extract communicative behavior, including non-verbal listening cues and verbal speaking patterns, from massive real-life conversational videos and encode them to a motion latent space. The motion latent codes are then used to animate a static portrait image to authentically imitate the character in the video. A well-crafted motion latent space should exhibit a high degree of disentanglement, which means that head pose, facial expression, and emotion should be decoupled from appearance. To this end, we perform facial structure discretization and facial pixel masking to the input. While in the second stage, the model is trained to map the dyadic audio to the motion latent space pretrained in stage 1 to obtain the corresponding motion latent codes. It consists of an interactive motion guider and a conditional diffusion transformer. The former takes as input the audio from both agent and its conversation partner, and retrieves verbal and non-verbal motions from learnable memory banks to construct the interactive motion feature. The latter utilizes the interactive motion feature as the condition and, together with other signals, generates the motion latent codes through denoising.", + "bbox": [ + 91, + 397, + 483, + 851 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To support our research on interactive head generation, we introduce DyConv, a large scale dataset of rich dyadic conversations collected from the Internet. It contains non", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b22d306760e5e765f73694d335d631ca5064a91d6ceaa1fe679c4df1e33bc84d.jpg", + "image_caption": [ + "Figure 2. Objective illustration. Existing interactive head generation (left) applied manual role assignment and explicit role switching. Our proposed INFP (right) is a unified framework which can dynamically and naturally adapt to various conversational states." + ], + "image_footnote": [], + "bbox": [ + 516, + 90, + 709, + 210 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2fa039f3238f285f4a7c9432007d5d12fb09c18f1ddbf2ecd825e0d3b7b172c0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 723, + 125, + 908, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "scripted video conversations between pairs of individuals from diverse backgrounds. The videos capture their discussions on a wide range of topics and emotions, portraying authentic communication in real-life settings. Compared to previous datasets like ViCo [58], ViCo-X [57] and RealTalk [8], DyConv has a larger scale, higher quality, and richer interactions.", + "bbox": [ + 511, + 321, + 906, + 426 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions are summarized as follows:", + "bbox": [ + 531, + 428, + 872, + 441 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose INFP, an audio-driven interactive head generation framework with a new paradigm. To our best knowledge, we are the first to enable the driven individual to naturally transition between listening and speaking states during multi-turn conversations, without the requirement for manual role assignment or explicit role switching.", + "- We design a novel interactive motion guider with learnable memory banks to adaptively extract interactive information from dual-track dyadic audio and construct speaking-listening mixed motions.", + "- We introduce a large-scale and high-quality dataset called DyConv. We hope it will encourage the investigation of head generation for dyadic interaction.", + "- Our framework achieves real-time and person-generic head generation. Extensive experiments and visualizations demonstrate superior performance and effectiveness of our method." + ], + "bbox": [ + 513, + 446, + 906, + 702 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 717, + 656, + 733 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Single-Sided Audio-Driven Head Generation", + "text_level": 1, + "bbox": [ + 511, + 742, + 890, + 758 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Listening Head Generation. As a pioneer, RLHG [58] contributes the ViCo dataset as a common benchmark and builds a baseline method with a sequential decoder. Following that, PCH [13] introduces the enhanced renderer to improve the visual quality of generated videos. Song et al. [32] organizes the REACT challenge to encourage the community to investigate the face reaction in offline and online setting, separately. L2L [23] leverages the VQ-VAE [40] to store head motion sequences in discrete codebooks,", + "bbox": [ + 511, + 763, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10668", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "leading to the better motion diversity. ELP [31] rearranges the latent space based on emotional priors to emphasize the emotion of generated motions. To get rid of simple emotional labels, CustomListener [18] proposes to control listener head motions via the text prior. MFR-Net [16] makes a progress of identity preservation and motion diversity. RealTalk [8] proposes a medium-scale dataset and attempts to retrieve possible videos of listener head with large language model. Ng et al. [24] presents a transfer-based approach from pretrained large language models to predict the motion sequences.", + "bbox": [ + 89, + 90, + 480, + 256 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Talking Head Generation Audio-driven talking head generation has become a prominent research area in recent years. Early GAN-based approaches, like Wav2Lip [27], PC-AVS [56], and MakeItTalk [60], directly integrate the input audio and video to produce lip-synced visuals. VASA-1 [46] takes a significant leap forward in the field by decoupling latent feature spaces, enabling the generation of more realistic video outputs. This model can operate on a single image rather than requiring a full video sequence. With the emergence of diffusion models, several works [33, 37, 41] are proposed to enhance the generation quality. EchoMimic [2] is concurrently trained using both audios and facial landmarks, and is capable of generating portrait videos not only by audios and facial landmarks individually, but also by a combination of both audios and selected facial landmarks. AniTalker [56], employing two self-supervised learning strategies, excels in capturing intricate facial movements.", + "bbox": [ + 91, + 258, + 486, + 530 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Head Generation in Dyadic Conversations", + "text_level": 1, + "bbox": [ + 89, + 544, + 450, + 559 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recently, the research community has paid increasing attention on the audio-driven head generation in the dyadic interaction. Zhou et al. [59] extend RHLG [58] by introducing ViCo-X, a multi-turn conversational dataset, and a Role Switcher to connect the Listener Generator and the Speaker Generator. However, explicit role switching can lead to unnaturalness and inconsistency, and this paradigm fails to cover all dyadic conversation states, such as simultaneous speaking. DIM [39] uses a pretraining approach to model both speakers' and listeners' motions, capturing dyadic context. However, the pretrained model requires additional fine-tuning for tasks like talking head and listening head generation, necessitating manual role assignment in dyadic conversations, which can cause inappropriate transitions. AgentAvatar [42] designs a streamlined pipeline to predict behaviors of a photo-realistic avatar agent in dyadic interactions. While the motion synthesis is conditioned on the high-level textual description, AgentAvatar fails to generate facial behaviors that precisely align with the dialogue audio, as demo videos shown in their project page. Additionally, there are some other studies in dyadic interactions [14, 25, 26, 34, 47], but all of them are person-specific and", + "bbox": [ + 89, + 568, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "suffer from the generalization ability to arbitrary individuals.", + "bbox": [ + 511, + 90, + 903, + 119 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 511, + 138, + 604, + 155 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our goal is to synthesize the agent head during a dyadic conversation. To this end, we propose a two-stage interactive head generation framework, called INFP.", + "bbox": [ + 511, + 166, + 903, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the first stage, the model learns to extract and compress multiple conversational behaviors in videos into a low-dimensional motion latent space, and use these latent codes to animate the portrait image of the agent, denoted as $I_{\\text{self}}$ . Note that, to ensure the richness and diversity of the motion latent space, we train stage 1 with plenty of real-life dialogue videos. Once the training converges, we freeze this part and start to train the second stage. The model takes as input the audio of the agent, denoted as $A_{\\text{self}}$ , and the audio of the conversation partner, denoted as $A_{\\text{other}}$ , to learn the mapping from this dual-track dialogue audio to motion latent codes through denoising. We present a comprehensive introduction to these two stages below.", + "bbox": [ + 511, + 213, + 906, + 409 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Motion-Based Head Imitation", + "text_level": 1, + "bbox": [ + 511, + 422, + 779, + 436 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Motion Encoding. For motion encoding, we employ a motion encoder $\\mathbf{E}_m$ to learn an implicit latent representation, which captures verbal and non-verbal communicative behaviors from input facial images. $\\mathbf{E}_m$ comprises a convolutional-based feature extractor followed by MLP layers, which compresses the motion feature into a latent code. Note that, we use implicit representation instead of explicit 3DMM coefficients [28] because the expressiveness of 3DMM model is limited, and the commonly used expression coefficients, predicted by SOTA 3D facial estimation methods [5, 7], entangle with face shape to a certain extent.", + "bbox": [ + 511, + 446, + 906, + 612 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A well-crafted motion latent space should demonstrate a high level of disentanglement to ensure that the extracted motions are appearance-independent. To achieve this, we apply two measures. (a) We set the motion latent code as a low-dimensional 1-D descriptor. It fosters effectively capturing the fundamental semantics of facial motion without entangling with appearance information, as explained in previous works [38, 43, 44]. (b) We design a hybrid facial representation to replace the original input image of $\\mathbf{E}_m$ . In particular, we first mask the majority of facial pixels, preserving only the eyes and lip regions, as we regard them as the most intricate and expressive components of facial expressions. This can block the interference of irrelevant information such as hair and background for expression distill. To provide facial orientation and contour information, we then use an off-the-shelf facial estimation model [20] to obtain the facial vertices, and project the contour-related points onto the masked image. Our ablation study shows the superiority of our proposed hybrid facial representation", + "bbox": [ + 511, + 613, + 908, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10669", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7cc4c46f83c049322df0e0cfa2e462981e31bf2829f50178c427e8177dec017f.jpg", + "image_caption": [ + "Figure 3. Overview of INFP. The first stage (Motion-Based Head Imitation) learns to project facial communicative behaviors from real-life conversation videos into a low-dimensional motion latent space, and use the latent codes to animate a static image. The second stage (Audio-Guided Motion Generation) learns the mapping from the input dyadic audio to motion latent codes through denoising, to achieve the audio-driven interactive head generation." + ], + "image_footnote": [], + "bbox": [ + 98, + 89, + 588, + 318 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/32f5ec0b399785cca4e30e3ce51d5c14fa59c4fefe57de140f4facab683dce5f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 589, + 89, + 898, + 319 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "compared with the other representations like the intact image [6] or 2D landmarks map [22].", + "bbox": [ + 88, + 415, + 480, + 445 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Head Imitation. In this paper, we adopt a motion flow-based generation pipeline. Concretely, we first extract source motion code $m_{self}$ from the input portrait image $I_{self}$ , and driving motion codes $m_{dri}^{1:N}$ from the driving video $V_{dri}$ . A motion flow estimation model $\\mathbf{F}$ takes these motion codes as input to predict the self $\\rightarrow$ driving flow. We utilize a face encoder $\\mathbf{E}_{face}$ to extract 3D appearance feature volume from the input image $I_{self}$ . This feature volume is then warped by the motion flow and sent into a face decoder $\\mathbf{D}_{face}$ to synthesize the final predictions $I_{pred}^{1:N}$ . The overall pipeline can be written as:", + "bbox": [ + 88, + 446, + 482, + 612 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l}\\text {F l o w} _ {s \\rightarrow d} = \\mathbf {F} (\\mathbf {E} _ {m} (I _ {\\text {s e l f}}), \\mathbf {E} _ {m} (V _ {\\text {d r i}}))\\\\\\text {I} ^ {1: N} = \\mathbf {D} _ {n} \\left(W _ {n} - \\left(I _ {n}\\right) I _ {n} + \\dots\\right)\\end{array}\\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 617, + 482, + 643 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI _ {p r e d} ^ {1: N} = \\mathbf {D} _ {f a c e} (W a r p (\\mathbf {E} _ {f a c e} (I _ {s e l f}), F l o w _ {s \\rightarrow d}))\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 638, + 452, + 656 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Model details and loss function are shown in the supplementary material.", + "bbox": [ + 89, + 659, + 482, + 690 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Audio-Guided Motion Generation", + "text_level": 1, + "bbox": [ + 89, + 696, + 387, + 712 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Previous works [34, 39, 59] tend to divide their models into a Speaker Generator and a Listener Generator. Since the Speaker Generator only receives speaking audio, while the Listener Generator solely processes listening audio, it is necessary to segment the audio of both agent and its conversation partner into speaking clips and listening clips. Hence, in a multi-turn dialogue, explicit role switching is also crucial. We argue that it can lead to unnatural transitions in communicative states, particularly in real-life interaction contexts characterized by frequent role switching. Besides, this design paradigm clearly goes against the intention of the conversational agent, as it should be able to", + "bbox": [ + 88, + 719, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "naturally present different communicative states based on the progress of the conversation.", + "bbox": [ + 511, + 415, + 903, + 445 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this paper, we present a new paradigm, in which the model takes as input the audio from both the agent and its conversation partner, eliminating the necessity for explicit role assignment or switching. The key innovations lie in extracting interactive information from dual-track dyadic audio to construct speaking-listening mixed motions, and mapping these motions to the motion latent space of stage 1 to achieve audio-driven head generation.", + "bbox": [ + 511, + 446, + 905, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Motion Feature Construction. As shown in Fig. 3, an interactive motion guider is introduced to extract motion feature from the input audio. Within it, there are two memory banks $M_v$ and $M_{nv}$ , storing verbal and nonverbal motion, respectively. Each motion memory bank consists of learnable embeddings to memorize typical motions, denoted as $e^{1:K}$ , where $e^k \\in \\mathbb{R}^d$ represents the $k$ -th motion embedding and $d$ is the dimension. We set $K$ to 64 and $d$ to 512 in our experiments. When agent is speaking, $A_{self}$ contains rich information, so we take it as $Query$ and obtain verbal motion feature from $M_v$ , which is taken as $Key$ and $Value$ , through a cross-attention layer. Since the partner is listening, $A_{other}$ carries very little information, hence no effective non-verbal motion feature can be obtained from $M_{nv}$ . Therefore, after feature fusion, which involves element-wise sum and several MLP layers, verbal motion feature dominates and drives the agent to express the intended speaking state. However, when the partner starts speaking, $A_{self}$ becomes weak and $A_{other}$ fulfills a vital role. Accordingly, non-verbal motion feature from $M_{nv}$ dominates and forces the agent to present a listening state. In this way, our model can dynamically construct", + "bbox": [ + 511, + 568, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "10670", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "interactive motion feature $f_{m}$ based on the content of the dialogue audio, allowing the agent to convey the intended state. Note that, $A_{\\text{self}}$ and $A_{\\text{other}}$ are encoded to the feature space through an off-the-shelf HuBERT audio encoder [12], before sent into the model.", + "bbox": [ + 89, + 90, + 480, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inspired by stylistic facial animation methods [21, 35], we also introduce a motion style vector $s_m$ to explicitly edit the motion embeddings of memory banks, through a style modulation layer raised by StyleGAN2[15]. $s_m$ is first extracted from motion latent codes through a motion style encoder $\\mathbf{E}_{style}$ , and then squeezed along the temporal dimension to a global descriptor. We argue that $s_m$ contains some global information like emotions and attitudes, thus enhancing the authenticity and vividness of the generated results. Note that during training, $s_m$ is sourced from an arbitrary video clip of the driven individual, while in testing, $s_m$ can be extracted from any videos or set as null.", + "bbox": [ + 89, + 174, + 482, + 354 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Motion Mapping. To map the interactive motions $f_{m}$ to the pretrained motion latent space of stage 1, we employ a conditional diffusion transformer $\\mathbf{T}$ [11]. Given the data distribution $q(m^{1:N},f_m)$ where $m^{1:N}$ represents the corresponding motion latent codes with $N$ frames, the diffusion model approximates the conditional distribution $q(m^{1:N}|f_m)$ . $\\mathbf{T}$ only has 4 blocks, which makes our framework lightweight enough to achieve real-time interaction. Each block consists of a self-attention layer, a motion-attention layer, and a temporal-attention layer. $\\mathbf{T}$ predicts the noise added to the ground-truth motion latent codes for each denoising step. Diffusion timestep is converted into the sinusoidal embedding and then concatenated to the noised motion latent codes at the time dimension. In the motion attention layer, we take the latent feature output from preceding self-attention layer as the Query, while the interactive motion feature $f_{m}$ is utilized as the Key and Value. Inspired byAnimateDiff [10], we implement a temporal layer to ensure a smooth transition between adjacent windows. In particular, we incorporate the last 10 frames of the generated motion latent codes from the previous window as the condition of the current one. Note that in the inference time, we add noise to the self motion latent codes $m_{\\text{self}}^{1:N}$ , which are extracted from $I_{\\text{self}}$ and replicated into $n$ copies.", + "bbox": [ + 91, + 364, + 482, + 741 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Strategy. In the training stage, we randomly set style vector to null with a probability of 0.3, and randomly drop interactive motion feature and previous motion latent codes with a probability of 0.5. Also note that we adopt a warm-up strategy for training stage 2. Specifically, we only select single-sided conversation clips from our training set as \"simple case\" to train the model for several epochs, which helps with the initialization of two memory banks. After that, we randomly sample multi-turn clips from entire dataset for the remaining training process.", + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Dataset", + "text_level": 1, + "bbox": [ + 513, + 90, + 602, + 104 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Existing dyadic interaction datasets have limitations in both scale and quality. ViCo [58] is a small-scale dataset with 1.6 hours and 96 IDs, and it lacks multi-turn conversation scenarios. Although ViCo-X [57] features rich multi-turn conversations, the total duration is only 0.4 hours. Besides, its background is monotonous, all of which are green screen. RealTalk [8] is a medium-scale dataset capturing dyadic conversations between pairs of individuals totaling to 115 hours of online public videos. However, it does not restrict the resolution of faces in the video, leading to low face resolution in many videos, thus impacting the quality of model generation.", + "bbox": [ + 511, + 114, + 903, + 295 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We propose DyConv, a large-scale dataset of multi-turn dyadic conversations collected from the Internet. The dataset contains video clips with a duration of over 200 hours, captures a wide gamut of emotions and expressions. We employ an off-the-shelf face detection model [51], and keep only frames where both individuals of the conversation are fully visible and their facial resolution is greater than $400 \\times 400$ . In addition, we utilize a SOTA speech separation model [54] to separate the audio of two persons in the conversation, denoted as $A_{p1}$ and $A_{p2}$ . Then we run an active speaker detection model [1] to match each audio clip to the corresponding face in the original video. We will publicly release the dataset, along with separated audios and annotations. We hope it will encourage the investigation of head generation for dyadic interaction.", + "bbox": [ + 511, + 296, + 906, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiment", + "text_level": 1, + "bbox": [ + 513, + 536, + 638, + 553 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 561, + 732, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In our experiments, we train the diffusion transformer using AdamW optimizer [19], with $\\beta_{1} = 0.9$ and $\\beta_{2} = 0.999$ , a learning rate of 1e-4, a weight decay of 1e-2, and a batch size of 32. For inference, we use a DDIM sampler [30] for 20 denoising steps. The CFG parameter of motion feature is set to 2, and the CFG parameter of previous motion latent is set to 1.5.", + "bbox": [ + 511, + 582, + 905, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Comparisons", + "text_level": 1, + "bbox": [ + 511, + 696, + 653, + 713 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although originally designed for dyadic interactive scenarios, our method can be naturally generalized to other applications, such as listening head generation and talking head generation. This allows us to fully validate the advancement of our proposed framework against some of the current SOTA methods in related areas.", + "bbox": [ + 511, + 719, + 903, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics. To quantitatively evaluate the visual quality, we utilize the Structured Similarity (SSIM), Peak Signal-to-Noise Ratio (PSNR) and Frechet Inception Distance (FID) for the generated videos. Additionally, LPIPS [50] is employed to measure the feature-level similarity between generated and ground-truth frames. To as", + "bbox": [ + 511, + 810, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "10671", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b8d49fc67c745b6073605764386dd2a295b68fe2afbc67a65ec00e528d219b02.jpg", + "image_caption": [ + "Reference Image" + ], + "image_footnote": [], + "bbox": [ + 127, + 174, + 222, + 244 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c571530f3918fed859d3aed95d251d8bbca07b5f8d2bcc876d05cf6b41820208.jpg", + "image_caption": [ + "Figure 4. Qualitative comparison of interactive head generation on DyConv. Given the input audio of dyadic conversations (the audio of the predicted agent is black and the audio of the conversation partner is blue), our framework generates reasonable and vivid facial and head pose movements as various roles in interactive scenarios." + ], + "image_footnote": [], + "bbox": [ + 232, + 93, + 875, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "sess the preservation of facial appearance, we calculate the cosine similarity (CSIM) [4] between the facial features of the reference image and the generated video frames. In term of motion diversity of generated videos, we use SI for diversity (SID) and Variance (Var) following previous studies [23, 39]. We use the confidence score of SyncNet (SyncScore) [3] to evaluate the synchronization between lip movements and audio signals of the agent.", + "bbox": [ + 89, + 689, + 482, + 811 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2.1. Interactive Head Generation", + "text_level": 1, + "bbox": [ + 89, + 820, + 336, + 834 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Since only a few works focus on the interactive head generation [42, 59] and these methods are not open-sourced, direct comparisons with them are considerably difficult. To effectively validate our method, we select the SOTA dyadic", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "modeling work DIM [39] as our baseline and utilize the officially released code to re-train the model on our datasets. Specifically, we first use the whole dataset to pretrain the VQ-VAE for the listener and speaker and then split the dataset into two subsets to train the listener generator and the speaker generator, respectively. To better visualize results of DIM, we also train the face renderer [28] to generate photo-realistic videos.", + "bbox": [ + 511, + 689, + 906, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Results. The quantitative experiments are carried out on the test set of DyConv. Please refer to Tab. 1 for the results. It can be seen that our method outperforms the baseline cross a variety of metrics. In particular, our face decoder achieves a better visual quality such as SSIM, PSNR and FID. Thanks to the disentanglement and expres", + "bbox": [ + 511, + 810, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "10672", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "siveness of our latent space, our method is superior to DIM in term of audio-lip-synchronization by a large margin as there is a much closer distance of our SyncScore to the ground truth. As for the LPIPS and CISM, our method still achieves solid performances, reflecting the better identity preservation. Of particular note, our method is far ahead of the baseline in term of SID and Var, which proves the generalization and effectiveness of our method to generate a great range of diverse motions. It benefits from the design of interactive motion guider as well as conditional denoising transformer.", + "bbox": [ + 89, + 90, + 480, + 256 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Results. Subjective evaluation is crucial in identifying the performance of generative models, particularly on videos. We strongly recommend readers to watch our supplementary video. Two examples and their comparisons with the baseline are shown in Fig. 4. We use two typical audio clips in the dyadic conversation scenario to drive reference identities randomly selected from the test set. It can be obviously observed that our method can generate more diverse head pose movements in comparison to DIM. Meanwhile, results of our method demonstrate more vivid facial expressions, which significantly improve the naturalness of generated videos. It is worth to noting that facial behaviors synthesized by our method are well aligned with the dialogue audio content. This suggests the effectiveness of our method in dyadic interactions, which is capable of capturing distinct characteristics of various states, such as listening, speaking as well as the mutual interaction, and naturally adapting its role among them based on the progress of the conversation. In contrast, DIM is unable to generate appropriate behaviors such as reasonable lip movement and response feedback.", + "bbox": [ + 91, + 257, + 482, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "User Study. We also invite 20 participants to conduct a user study for further subjective evaluation. Specifically, we apply INFP and DIM to synthesize results of 20 dyadic conversation videos randomly sampled from the test set. By adopting the commonly used Mean Opinion Scores (MOS) rating protocol, we request all the participants to provide their ratings (from 1 to 5, the higher the better) on four aspects for each generated video: naturalness, motion diversity, audio-visual alignment and visual quality. As detailed in Tab. 2, our method demonstrates superior performances across all evaluated aspects compared to DIM.", + "bbox": [ + 89, + 574, + 482, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2.2. Listening Head Generation", + "text_level": 1, + "bbox": [ + 89, + 746, + 326, + 760 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Since our method can generalize to the task of listening head generation, we directly use our model without any modification to conduct additional experiments. Specifically, we compare our framework with SOTA listening head generation methods, including L2L [23], RLHG [58] and DIM [39]. ViCo [58] is selected as the benchmark. For quantitative evaluation, we utilized the same test set and inherited both quantitative and qualitative results presented in DIM. We conducted the comparison on FD, MSE, SID", + "bbox": [ + 89, + 763, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and Var and each metric is calculated on expression and head pose, respectively. Different to other methods, our method generate photo-realistic video rather than 3DMM coefficients. To this end, we use the off-the-shelf 3D face reconstruction method [7] to extract expression and head pose coefficients from our generated videos. As shown in Tab. 3, INFP outperforms other SOTA methods in the majority of metrics. While SID and Var are performing well, our method remains highly effective in both FD and MSE. This indicates the power of our novel framework design and dataset quality, making generated facial expressions and head pose movements closer to real human behaviors. The visualization results in the supplementary materials also demonstrate that INFP achieves more expressive and diverse motions.", + "bbox": [ + 511, + 90, + 903, + 316 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2.3. Talking Head Generation", + "text_level": 1, + "bbox": [ + 511, + 328, + 736, + 343 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For talking head generation, we select SadTalker[52], AntiTalker[17], and EchoMimic[2] as comparing methods. As all of them reported evaluation results on HDTF [53], we randomly sampled 20 videos from it as the test set to conduct our experiments for a fair comparison. Results in Tab. 4 show that our model is superior in audiolip-synchronization and identity preservation. The reason behind this might be that we carefully build a disentangled and expressive motion latent space and it makes our model easier to learn the correlation between audio and lip movements without the interference of appearance features. It can also be observed from the visualization results in the supplementary materials. Our SSIM and PSNR score is slightly behind EchoMimic because EchoMimic is finetuned based on the Stable Diffusion V1.5 [29]. Despite this, our method can achieve better scores in terms of FID, meaning a good visual quality of our results.", + "bbox": [ + 511, + 348, + 906, + 604 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 618, + 666, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We ablate 3 key components of our framework, and the results are listed in Tab. 1.", + "bbox": [ + 511, + 641, + 903, + 670 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Motion Memory Bank. We remove the verbal and non-verbal motion memory banks from interactive motion guider in stage 2, and directly use the dyadic audio feature after several MLP layers as the condition for the motion-attention layer. Results show that the generated lip-related motion behaviors are clearly inconsistent with the ground-truth. In additional, there is a significant drop of both SID and Var scores. These suggest that memory banks can effectively store rich motion patterns in different conversational states. The experiment on memory bank size is shown in the supplementary material.", + "bbox": [ + 511, + 672, + 905, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Style Modulation. We remove the motion style vector $s_m$ and corresponding modulation layers. As we expected, the authenticity and vividness of the results decrease, which can also be reflected in the user study.", + "bbox": [ + 511, + 839, + 903, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "10673", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a1980df37238dd08edd195e3c4a14d544ceccc8cb460867351ee4c18066f806b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsSSIM↑PSNR↑FID↓SyncScore↑LPIPS↓CSIM↑SID↑Var↑
DIM [39]0.65120.41734.3614.7780.4850.8240.7660.825
INFP (Ours)0.83431.56215.7277.1880.2570.9042.6132.386
w/o Motion Memory0.83031.21818.3346.1030.2590.8992.1532.016
w/o Style Modulation0.83131.44216.0297.0620.2710.9042.5512.316
w/ Intact Image0.80228.48816.9906.8120.2660.8422.4702.148
w/ Landmarks Map0.82130.69316.3276.8330.2810.9012.6012.335
GT1.000N/A0.0007.2610.0000.9672.8912.435
", + "bbox": [ + 143, + 88, + 857, + 239 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b61b10dccf149a32b048c6c087de601f6c86143f0187a1c836661928cedc4a11.jpg", + "table_caption": [ + "Table 1. Quantitative comparison of interactive head generation on DyConv." + ], + "table_footnote": [], + "table_body": "
MethodsNaturalnessMotion DiversityAV-AlignVisual
DIM [39]2.712.142.653.13
INFP w/o Style4.023.984.284.11
INFP (Ours)4.384.494.334.13
", + "bbox": [ + 94, + 286, + 480, + 353 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/42b8d37ab92fcc3d7c2e90f3fb16c1460034776a861cc69710539c1f885fe46e.jpg", + "table_caption": [ + "Table 2. User study measured by Mean Opinion Scores." + ], + "table_footnote": [], + "table_body": "
MethodFD↓MSE↓SID↑Var↑
ExpPoseExpPoseExpPoseExpPose
L2L [23]33.930.060.930.012.772.660.830.02
RLHG [58]39.020.070.860.013.623.171.520.02
DIM [39]23.880.060.700.013.712.351.530.02
INFP (Ours)18.630.070.510.014.783.922.830.18
", + "bbox": [ + 94, + 390, + 480, + 486 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/be212b0b1c3e2397cfd34902fcadc4a61ee38b3fe23f0484a12e5dacf2163b83.jpg", + "table_caption": [ + "Table 3. Quantitative comparison of listening head generation on ViCo [58]." + ], + "table_footnote": [], + "table_body": "
MethodsSSIM↑PSNR↑FID↓SyncScore↑CSIM↑
SadTalker [52]0.78126.65028.5166.1250.814
AniTalker [17]0.78926.82228.7785.7380.858
EchoMimic [2]0.80728.98320.8736.2510.863
INFP (Ours)0.80428.65718.5826.6700.886
", + "bbox": [ + 96, + 537, + 477, + 623 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Quantitative comparison of talking head generation on HDTF [53].", + "bbox": [ + 89, + 632, + 482, + 660 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Hybrid Facial Representation. We change the input to $\\mathbf{E}_m$ in stage 1 from our carefully-designed hybrid facial representation to the original intact image or 2D landmarks map. We observe that there is degradation in generation quality and a leakage of appearance information.", + "bbox": [ + 89, + 686, + 482, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Visualization of Style Control", + "text_level": 1, + "bbox": [ + 89, + 772, + 354, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As illustrated in Sec. 3.2, our model supports inputting style vector $s_m$ extracted from any portrait video to globally control the emotions or attitudes of the generated results. Here we select 3 video clips with varying degrees of \"happiness\" to control emotions during generation. Results are shown in Fig. 5. It can be seen from the figure that $s_m$ effectively control the emotions of the driven individual.", + "bbox": [ + 89, + 794, + 482, + 898 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dc25b44437a90aa874b190cb2f512ca773ff5ad605d728b42492dcf5a0378246.jpg", + "image_caption": [ + "Figure 5. Visualization of style control with our style modulation." + ], + "image_footnote": [], + "bbox": [ + 535, + 287, + 887, + 468 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 521, + 633, + 536 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose INFP, a novel audio-driven interactive head generation framework for dyadic conversations, which aims to mimic facial behaviors of human-human interactions conditioning on the dyadic audio. As demonstrated, our method is a unified framework and is capable of smoothly and naturally adapting to various conversational roles without explicit assignment or switching. Meanwhile, we collect a large scale audio-visual dataset for multi-turn dyadic conversations, which could facilitate further researches of dyadic interactive head generation. Extensive experiments and ablations prove the advancement and effectiveness of our method over other SOTA methods.", + "bbox": [ + 509, + 547, + 906, + 729 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitation and Future Work. Currently, our method only takes conversation audios as the input, combining control signals from multiple modalities, such as visual and textual contents, can offer additional capacities. As our proposed framework focuses on head synthesis, extending the generation range to the upper, even whole body might be a interesting and challenging future work.", + "bbox": [ + 511, + 732, + 906, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ethical Considerations. There might be the potential misuse of our method to fabricate deceptive talks and speeches. To this end, we will restrict access to our core models exclusively to research institutions.", + "bbox": [ + 511, + 839, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "10674", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Stanislav Beliaev, Yuriy Rebryk, and Boris Ginsburg. Talknet: Fully-convolutional non-autoregressive speech synthesis model, 2020. 5", + "[2] Zhiyuan Chen, Jiajiong Cao, Zhiquan Chen, Yuming Li, and Chenguang Ma. Echomimic: Lifelike audio-driven portrait animations through editable landmark conditions. arXiv preprint arXiv:2407.08136, 2024. 1, 3, 7, 8", + "[3] J. S. Chung and A. Zisserman. Out of time: automated lip sync in the wild. In Workshop on Multi-view Lip-reading, ACCV, 2016. 6", + "[4] Jiankang Deng, Jia Guo, Xue Niannan, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In CVPR, 2019. 6", + "[5] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 3", + "[6] Nikita Drobyshev, Jenya Chelishev, Taras Khakhulin, Aleksei Ivakhnenko, Victor Lempitsky, and Egor Zakharov. Megaportraits: One-shot megapixel neural head avatars. 2022. 4", + "[7] Panagiotis P. Filntisis, George Retsinas, Foivos Paraperas-Papantoniou, Athanasios Katsamanis, Anastasios Roussos, and Petros Maragos. Visual speech-aware perceptual 3d facial expression reconstruction from videos. arXiv preprint arXiv:2207.11094, 2022. 3, 7", + "[8] Scott Geng, Revant Teotia, Purva Tendulkar, Sachit Menon, and Carl Vondrick. Affective faces for goal-driven dyadic communication. 2023. 2, 3, 5", + "[9] Jiazhi Guan, Zhanwang Zhang, Hang Zhou, Tianshu Hu, Kaisiyuan Wang, Dongliang He, Haocheng Feng, Jingtuo Liu, Errui Ding, Ziwei Liu, et al. Stylesync: High-fidelity generalized and personalized lip sync in style-based generator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1505-1515, 2023. 1", + "[10] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Imagine your personalized text-to-image diffusion models without specific tuning. International Conference on Learning Representations, 2024. 5", + "[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 5", + "[12] Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. Hubert: Self-supervised speech representation learning by masked prediction of hidden units. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 29:3451-3460, 2021. 5", + "[13] Ailin Huang, Zhewei Huang, and Shuchang Zhou. Perceptual conversational head generation with regularized driver and enhanced renderer. In Proceedings of the 30th ACM" + ], + "bbox": [ + 93, + 114, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "International Conference on Multimedia, page 7050-7054. ACM, 2022. 2", + "[14] Yinghao Huang, Leo Ho, Dafei Qin, Mingyi Shi, and Taku Komura. Interact: Capture and modelling of realistic, expressive and interactive activities between two persons in daily scenarios. 2024. 3", + "[15] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In Proc. CVPR, 2020. 5", + "[16] Jin Liu, Xi Wang, Xiaomeng Fu, Yesheng Chai, Cai Yu, Jiao Dai, and Jizhong Han. Mfr-net: Multi-faceted responsive listening head generation via denoising diffusion model. In Proceedings of the 31st ACM International Conference on Multimedia, pages 6734-6743, 2023. 2, 3", + "[17] Tao Liu, Feilong Chen, Shuai Fan, Chenpeng Du, Qi Chen, Xie Chen, and Kai Yu. Anitalker: Animate vivid and diverse talking faces through identity-decoupled facial motion encoding, 2024. 1, 7, 8", + "[18] Xi Liu, Ying Guo, Cheng Zhen, Tong Li, Yingying Ao, and Pengfei Yan. Custom listener: Text-guided responsive interaction for user-friendly listening head generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2415-2424, 2024. 1, 2, 3", + "[19] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization, 2019. 5", + "[20] Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuo-Ling Chang, Ming Yong, Juhyun Lee, Wan-Teh Chang, Wei Hua, Manfred Georg, and Matthias Grundmann. Mediapipe: A framework for perceiving and processing reality. In CVPR 2019, 2019. 3", + "[21] Yifeng Ma, Suzhen Wang, Zhipeng Hu, Changjie Fan, Tangjie Lv, Yu Ding, Zhidong Deng, and Xin Yu. Styletalk: one-shot talking head generation with controllable speaking styles. In Proceedings of the Thirty-Seventh AAAI Conference on Artificial Intelligence and Thirty-Fifth Conference on Innovative Applications of Artificial Intelligence and Thirteenth Symposium on Educational Advances in Artificial Intelligence. AAAI Press, 2023. 5", + "[22] Yue Ma, Hongyu Liu, Hongfa Wang, Heng Pan, Yingqing He, Junkun Yuan, Ailing Zeng, Chengfei Cai, Heung-Yeung Shum, Wei Liu, et al. Follow-your-emoji: Fine-controllable and expressive freestyle portrait animation. arXiv preprint arXiv:2406.01900, 2024. 4", + "[23] Evonne Ng, Hanbyul Joo, Liwen Hu, Hao Li, Trevor Darrell, Angjoo Kanazawa, and Shiry Ginosar. Learning to listen: Modeling non-deterministic dyadic facial motion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20395-20405, 2022. 1, 2, 6, 7, 8", + "[24] Evonne Ng, Sanjay Subramanian, Dan Klein, Angjoo Kanazawa, Trevor Darrell, and Shiry Ginosar. Can language models learn to listen? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10083-10093, 2023. 1, 2, 3", + "[25] Evonne Ng, Javier Romero, Timur Bagautdinov, Shaojie Bai, Trevor Darrell, Angjoo Kanazawa, and Alexander Richard." + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "10675", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "From audio to photoreal embodiment: Synthesizing humans in conversations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1001-1010, 2024. 3", + "[26] Se Jin Park, Chae Won Kim, Hyeongseop Rha, Minsu Kim, Joanna Hong, Jeong Hun Yeo, and Yong Man Ro. Let's go real talk: Spoken dialogue model for face-to-face conversation. arXiv preprint arXiv:2406.07867, 2024. 3", + "[27] K R Prajwal, Rudrabha Mukhopadhyay, Vinay P. Namboodiri, and C.V. Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM International Conference on Multimedia, page 484-492, New York, NY, USA, 2020. Association for Computing Machinery. 1, 3", + "[28] Yurui Ren, Ge Li, Yuanqi Chen, Thomas H. Li, and Shan Liu. Pirenderer: Controllable portrait image generation via semantic neural rendering, 2021. 3, 6", + "[29] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 7", + "[30] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv:2010.02502, 2020. 5", + "[31] Luchuan Song, Guojun Yin, Zhenchao Jin, Xiaoyi Dong, and Chenliang Xu. Emotional listener portrait: Realistic listener motion simulation in conversation. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 20782-20792. IEEE, 2023. 1, 2, 3", + "[32] Siyang Song, Micol Spitale, Cheng Luo, Cristina Palmero, German Barquero, Hengde Zhu, Sergio Escalera, Michel Valstar, Tobias Baur, Fabien Ringeval, Elisabeth Andre, and Hatice Gunes. React 2024: the second multiple appropriate facial reaction generation challenge, 2024. 2", + "[33] Michal Stypulkowski, Konstantinos Vougioukas, Sen He, Maciej Zieba, Stavros Petridis, and Maja Pantic. Diffused heads: Diffusion models beat gans on talking-face generation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5091-5100, 2024. 3", + "[34] Mingze Sun, Chao Xu, Xinyu Jiang, Yang Liu, Baigui Sun, and Ruqi Huang. Beyond talking—generating holistic 3d human dyadic motion for communication. arXiv preprint arXiv:2403.19467, 2024. 3, 4", + "[35] Zhiyao Sun, Tian Lv, Sheng Ye, Matthieu Lin, Jenny Sheng, Yu-Hui Wen, Minjing Yu, and Yong-jin Liu. Diffpositalk: Speech-driven stylistic 3d facial animation and head pose generation via diffusion models. ACM Transactions on Graphics (TOG), 43(4):1-9, 2024. 5", + "[36] Jiaxiang Tang, Kaisiyuan Wang, Hang Zhou, Xiaokang Chen, Dongliang He, Tianshu Hu, Jingtuo Liu, Gang Zeng, and Jingdong Wang. Real-time neural radiance talking portrait synthesis via audio-spatial decomposition. arXiv preprint arXiv:2211.12368, 2022. 1", + "[37] Linrui Tian, Qi Wang, Bang Zhang, and Liefeng Bo. Emo: Emote portrait alive-generating expressive portrait videos with audio2video diffusion model under weak conditions. arXiv preprint arXiv:2402.17485, 2024. 1, 3", + "[38] Naftali Tishby, Fernando C Pereira, and William Bialek. The information bottleneck method. arXiv preprint physics/0004057, 2000. 3" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Minh Tran, Di Chang, Maksim Siniukov, and Mohammad Soleymani. Dyadic interaction modeling for social behavior generation. arXiv preprint arXiv:2403.09069, 2024. 1, 2, 3, 4, 6, 7, 8", + "[40] Aaron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu. Neural discrete representation learning, 2018. 2", + "[41] Cong Wang, Kuan Tian, Jun Zhang, Yonghang Guan, Feng Luo, Fei Shen, Zhiwei Jiang, Qing Gu, Xiao Han, and Wei Yang. V-express: Conditional dropout for progressive training of portrait video generation. arXiv preprint arXiv:2406.02511, 2024. 1, 3", + "[42] Duomin Wang, Bin Dai, Yu Deng, and Baoyuan Wang. Disentangling planning, driving and rendering for photorealistic avatar agents. 2023. 1, 2, 3, 6", + "[43] Duomin Wang, Yu Deng, Zixin Yin, Heung-Yeung Shum, and Baoyuan Wang. Progressive disentangled representation learning for fine-grained controllable talking head synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17979-17989, 2023. 3", + "[44] Yaohui Wang, Di Yang, Francois Bremond, and Antitza Dantcheva. Latent image animator: Learning to animate images via latent space navigation. arXiv preprint arXiv:2203.09043, 2022. 3", + "[45] Mingwang Xu, Hui Li, Qingkun Su, Hanlin Shang, Liwei Zhang, Ce Liu, Jingdong Wang, Luc Van Gool, Yao Yao, and Siyu Zhu. Hallo: Hierarchical audio-driven visual synthesis for portrait image animation. arXiv preprint arXiv:2406.08801, 2024. 1", + "[46] Sicheng Xu, Guojun Chen, Yu-Xiao Guo, Jiaolong Yang, Chong Li, Zhenyu Zang, Yizhong Zhang, Xin Tong, and Baining Guo. Vasa-1: Lifelike audio-driven talking faces generated in real time. arXiv preprint arXiv:2404.10667, 2024. 1, 3", + "[47] Yichao Yan, Zanwei Zhou, Zi Wang, Jingnan Gao, and Xiaokang Yang. Dialoguenerf: Towards realistic avatar face-to-face conversation video generation, 2023. 3", + "[48] Chenxu Zhang, Chao Wang, Jianfeng Zhang, Hongyi Xu, Guoxian Song, You Xie, Linjie Luo, Yapeng Tian, Xiaohu Guo, and Jiashi Feng. Dream-talk: diffusion-based realistic emotional audio-driven method for single image talking face generation. arXiv preprint arXiv:2312.13578, 2023. 1", + "[49] Longhao Zhang, Shuang Liang, Zhipeng Ge, and Tianshu Hu. Personatack: Bring attention to your persona in visual dubbing. arXiv preprint arXiv:2409.05379, 2024. 1", + "[50] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5", + "[51] Shifeng Zhang, Xiangyu Zhu, Zhen Lei, Hailin Shi, Xiaobo Wang, and Stan Z. Li. S $^3$ fd: Single shot scale-invariant face detector, 2017. 5", + "[52] Wenxuan Zhang, Xiaodong Cun, Xuan Wang, Yong Zhang, Xi Shen, Yu Guo, Ying Shan, and Fei Wang. Sadtalker: Learning realistic 3d motion coefficients for stylized audio-driven single image talking face animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8652-8661, 2023. 1, 7, 8" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10676", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661–3670, 2021. 7, 8", + "[54] Shengkui Zhao, Yukun Ma, Chongjia Ni, Chong Zhang, Hao Wang, Trung Hieu Nguyen, Kun Zhou, Jiaqi Yip, Dianwen Ng, and Bin Ma. Mossformer2: Combining transformer and rnnprecent network for enhanced time-domain monaural speech separation, 2023. 5", + "[55] Wei Zhao, Peng Xiao, Rongju Zhang, Yijun Wang, and Jianxin Lin. Semantic-aware responsive listener head synthesis. In Proceedings of the 30th ACM International Conference on Multimedia, pages 7065-7069, 2022. 2", + "[56] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking face generation by implicitly modularized audio-visual representation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 3", + "[57] Mohan Zhou, Yalong Bai, Wei Zhang, Ting Yao, Tiejun Zhao, and Tao Mei. Vico-x: Multimodal conversation dataset. https://project.mhzhou.com/vico, 2022. Accessed: 2022-09-30. 2, 5", + "[58] Mohan Zhou, Yalong Bai, Wei Zhang, Ting Yao, Tiejun Zhao, and Tao Mei. Responsive listening head generation: A benchmark dataset and baseline. In Proceedings of the European conference on computer vision (ECCV), 2022. 2, 3, 5, 7, 8", + "[59] Mohan Zhou, Yalong Bai, Wei Zhang, Ting Yao, and Tiejun Zhao. Interactive conversational head generation. arXiv preprint arXiv:2307.02090, 2023. 1, 2, 3, 4, 6", + "[60] Yang Zhou, Xintong Han, Eli Shechtman, Jose Echevarria, Evangelos Kalogerakis, and Dingzeyu Li. Makelttalk: speaker-aware talking-head animation. ACM Transactions on Graphics, 39(6):1-15, 2020. 2, 3" + ], + "bbox": [ + 91, + 92, + 485, + 585 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "10677", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/INFP_ Audio-Driven Interactive Head Generation in Dyadic Conversations/4440ddb1-10e9-45a7-942b-205ac25fbaaa_model.json b/2025/INFP_ Audio-Driven Interactive Head Generation in Dyadic Conversations/4440ddb1-10e9-45a7-942b-205ac25fbaaa_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c20f0c623f2d4e34ba377fd6485984b5889de378 --- /dev/null +++ b/2025/INFP_ Audio-Driven Interactive Head Generation in Dyadic Conversations/4440ddb1-10e9-45a7-942b-205ac25fbaaa_model.json @@ -0,0 +1,2059 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.123, + 0.131, + 0.877, + 0.154 + ], + "angle": 0, + "content": "INFP: Audio-Driven Interactive Head Generation in Dyadic Conversations" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.181, + 0.896, + 0.217 + ], + "angle": 0, + "content": "Yongming Zhu* Longhao Zhang* Zhengkun Rong* Tianshu Hu*† Shuang Liang Zhipeng Ge Bytedance" + }, + { + "type": "text", + "bbox": [ + 0.246, + 0.22, + 0.748, + 0.252 + ], + "angle": 0, + "content": "{zhuyongming, zhanglonghao.zlh, rongzhengkun, tianshu.hu, liangshuang. echo, zhipengge}@bytedance.com" + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.274, + 0.904, + 0.509 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.518, + 0.908, + 0.59 + ], + "angle": 0, + "content": "Figure 1. We present INFP, an audio-driven interactive head generation framework for dyadic conversations. Given the dual-track audio in dyadic conversations and a single portrait image of arbitrary agent, our framework can dynamically synthesize verbal, non-verbal and interactive agent videos with lifelike facial expressions and rhythmic head pose movements. Additionally, our framework is lightweight yet powerful, making it practical in instant communication scenarios with acceptable latency, such as the video conferencing. INFP denotes our method is Interactive, Natural, Flash and Person-generic." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.601, + 0.327, + 0.617 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.635, + 0.486, + 0.862 + ], + "angle": 0, + "content": "Imagine having a conversation with a socially intelligent agent. It can attentively listen to your words and offer visual and linguistic feedback promptly. This seamless interaction allows for multiple rounds of conversation to flow smoothly and naturally. In pursuit of actualizing it, we propose INFP, a novel audio-driven head generation framework for dyadic interaction. Unlike previous head generation works that only focus on single-sided communication, or require manual role assignment and explicit role switching, our model drives the agent portrait dynamically alternates between speaking and listening state, guided by the input dyadic audio. Specifically, INFP comprises a Motion-Based Head Imitation stage and an Audio-Guided Motion Generation stage. The first stage learns to project facial communicative behaviors from real-life conversation videos into a low-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.602, + 0.909, + 0.74 + ], + "angle": 0, + "content": "dimensional motion latent space, and use the motion latent codes to animate a static image. The second stage learns the mapping from the input dyadic audio to motion latent codes through denoising, leading to the audio-driven head generation in interactive scenarios. To facilitate this line of research, we introduce DyConv, a large scale dataset of rich dyadic conversations collected from the Internet. Extensive experiments and visualizations demonstrate superior performance and effectiveness of our method." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.769, + 0.646, + 0.785 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In recent years, researchers have dedicated considerable attention to audio-driven head generation [9, 18, 23, 24, 31, 37, 39, 42, 46, 56, 59], in order to build the conversational agent. However, most studies only focus on single-sided communication, such as talking or listening, ignoring the dyadic properties in human-human interaction. Talking-head generation [2, 17, 27, 36, 37, 41, 45, 46, 48, 49, 52, 56," + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.221, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.242, + 0.9 + ], + "angle": 0, + "content": "† Corresponding authors." + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.242, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10667" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.394 + ], + "angle": 0, + "content": "60] aims at synthesizing speaker's facial animations based on their reference images and a driving audio. Although these works can produce vibrant videos with accurate lip synchronization, they solely emphasize the speaker's role and overlook the listener's feedback. Listening-head generation [16, 18, 23, 24, 31, 32, 55, 58], on the other hand, intends to respond to speaker's behaviors. However, these works restrict the listener's response solely to non-verbal facial movements, which is clearly far from real-life interaction scenarios. Some recent studies [39, 42, 59] have started exploring head generation for dyadic interaction, but they require manual role assignment between Listener and Speaker and cannot achieve smooth and natural role switching. We believe that for a conversational agent, it should not be predetermined a role, but rather be able to freely switch between listening and speaking states based on the current dialogue context. When having a conversation with the agent, you may interrupt or respond to it at any moment. At that point, it naturally transitions from a speaker to a listener, and vice versa." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.398, + 0.485, + 0.852 + ], + "angle": 0, + "content": "In this paper, we propose a novel audio-driven head generation framework for dyadic interaction, namely INFP. Unlike previous works, we do not divide the model into Speaker Generator and Listener Generator, or apply explicit role switching. The driven individual dynamically alternates between the speaking and listening state, guided by the dyadic audio. Specifically, INFP comprises a Motion-Based Head Imitation stage and an Audio-Guided Motion Generation stage. In the first stage, our model learns to extract communicative behavior, including non-verbal listening cues and verbal speaking patterns, from massive real-life conversational videos and encode them to a motion latent space. The motion latent codes are then used to animate a static portrait image to authentically imitate the character in the video. A well-crafted motion latent space should exhibit a high degree of disentanglement, which means that head pose, facial expression, and emotion should be decoupled from appearance. To this end, we perform facial structure discretization and facial pixel masking to the input. While in the second stage, the model is trained to map the dyadic audio to the motion latent space pretrained in stage 1 to obtain the corresponding motion latent codes. It consists of an interactive motion guider and a conditional diffusion transformer. The former takes as input the audio from both agent and its conversation partner, and retrieves verbal and non-verbal motions from learnable memory banks to construct the interactive motion feature. The latter utilizes the interactive motion feature as the condition and, together with other signals, generates the motion latent codes through denoising." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.485, + 0.901 + ], + "angle": 0, + "content": "To support our research on interactive head generation, we introduce DyConv, a large scale dataset of rich dyadic conversations collected from the Internet. It contains non" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.092, + 0.71, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.725, + 0.125, + 0.91, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.239, + 0.908, + 0.298 + ], + "angle": 0, + "content": "Figure 2. Objective illustration. Existing interactive head generation (left) applied manual role assignment and explicit role switching. Our proposed INFP (right) is a unified framework which can dynamically and naturally adapt to various conversational states." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.322, + 0.907, + 0.427 + ], + "angle": 0, + "content": "scripted video conversations between pairs of individuals from diverse backgrounds. The videos capture their discussions on a wide range of topics and emotions, portraying authentic communication in real-life settings. Compared to previous datasets like ViCo [58], ViCo-X [57] and RealTalk [8], DyConv has a larger scale, higher quality, and richer interactions." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.429, + 0.874, + 0.443 + ], + "angle": 0, + "content": "Our main contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.447, + 0.907, + 0.538 + ], + "angle": 0, + "content": "- We propose INFP, an audio-driven interactive head generation framework with a new paradigm. To our best knowledge, we are the first to enable the driven individual to naturally transition between listening and speaking states during multi-turn conversations, without the requirement for manual role assignment or explicit role switching." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.538, + 0.907, + 0.597 + ], + "angle": 0, + "content": "- We design a novel interactive motion guider with learnable memory banks to adaptively extract interactive information from dual-track dyadic audio and construct speaking-listening mixed motions." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.598, + 0.907, + 0.643 + ], + "angle": 0, + "content": "- We introduce a large-scale and high-quality dataset called DyConv. We hope it will encourage the investigation of head generation for dyadic interaction." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.643, + 0.906, + 0.703 + ], + "angle": 0, + "content": "- Our framework achieves real-time and person-generic head generation. Extensive experiments and visualizations demonstrate superior performance and effectiveness of our method." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.447, + 0.907, + 0.703 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.718, + 0.657, + 0.734 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.743, + 0.892, + 0.759 + ], + "angle": 0, + "content": "2.1. Single-Sided Audio-Driven Head Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Listening Head Generation. As a pioneer, RLHG [58] contributes the ViCo dataset as a common benchmark and builds a baseline method with a sequential decoder. Following that, PCH [13] introduces the enhanced renderer to improve the visual quality of generated videos. Song et al. [32] organizes the REACT challenge to encourage the community to investigate the face reaction in offline and online setting, separately. L2L [23] leverages the VQ-VAE [40] to store head motion sequences in discrete codebooks," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10668" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.257 + ], + "angle": 0, + "content": "leading to the better motion diversity. ELP [31] rearranges the latent space based on emotional priors to emphasize the emotion of generated motions. To get rid of simple emotional labels, CustomListener [18] proposes to control listener head motions via the text prior. MFR-Net [16] makes a progress of identity preservation and motion diversity. RealTalk [8] proposes a medium-scale dataset and attempts to retrieve possible videos of listener head with large language model. Ng et al. [24] presents a transfer-based approach from pretrained large language models to predict the motion sequences." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.26, + 0.487, + 0.531 + ], + "angle": 0, + "content": "Talking Head Generation Audio-driven talking head generation has become a prominent research area in recent years. Early GAN-based approaches, like Wav2Lip [27], PC-AVS [56], and MakeItTalk [60], directly integrate the input audio and video to produce lip-synced visuals. VASA-1 [46] takes a significant leap forward in the field by decoupling latent feature spaces, enabling the generation of more realistic video outputs. This model can operate on a single image rather than requiring a full video sequence. With the emergence of diffusion models, several works [33, 37, 41] are proposed to enhance the generation quality. EchoMimic [2] is concurrently trained using both audios and facial landmarks, and is capable of generating portrait videos not only by audios and facial landmarks individually, but also by a combination of both audios and selected facial landmarks. AniTalker [56], employing two self-supervised learning strategies, excels in capturing intricate facial movements." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.545, + 0.452, + 0.56 + ], + "angle": 0, + "content": "2.2. Head Generation in Dyadic Conversations" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Recently, the research community has paid increasing attention on the audio-driven head generation in the dyadic interaction. Zhou et al. [59] extend RHLG [58] by introducing ViCo-X, a multi-turn conversational dataset, and a Role Switcher to connect the Listener Generator and the Speaker Generator. However, explicit role switching can lead to unnaturalness and inconsistency, and this paradigm fails to cover all dyadic conversation states, such as simultaneous speaking. DIM [39] uses a pretraining approach to model both speakers' and listeners' motions, capturing dyadic context. However, the pretrained model requires additional fine-tuning for tasks like talking head and listening head generation, necessitating manual role assignment in dyadic conversations, which can cause inappropriate transitions. AgentAvatar [42] designs a streamlined pipeline to predict behaviors of a photo-realistic avatar agent in dyadic interactions. While the motion synthesis is conditioned on the high-level textual description, AgentAvatar fails to generate facial behaviors that precisely align with the dialogue audio, as demo videos shown in their project page. Additionally, there are some other studies in dyadic interactions [14, 25, 26, 34, 47], but all of them are person-specific and" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "suffer from the generalization ability to arbitrary individuals." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.14, + 0.605, + 0.156 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.167, + 0.905, + 0.212 + ], + "angle": 0, + "content": "Our goal is to synthesize the agent head during a dyadic conversation. To this end, we propose a two-stage interactive head generation framework, called INFP." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.214, + 0.908, + 0.41 + ], + "angle": 0, + "content": "In the first stage, the model learns to extract and compress multiple conversational behaviors in videos into a low-dimensional motion latent space, and use these latent codes to animate the portrait image of the agent, denoted as \\( I_{\\text{self}} \\). Note that, to ensure the richness and diversity of the motion latent space, we train stage 1 with plenty of real-life dialogue videos. Once the training converges, we freeze this part and start to train the second stage. The model takes as input the audio of the agent, denoted as \\( A_{\\text{self}} \\), and the audio of the conversation partner, denoted as \\( A_{\\text{other}} \\), to learn the mapping from this dual-track dialogue audio to motion latent codes through denoising. We present a comprehensive introduction to these two stages below." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.424, + 0.78, + 0.438 + ], + "angle": 0, + "content": "3.1. Motion-Based Head Imitation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.447, + 0.907, + 0.613 + ], + "angle": 0, + "content": "Motion Encoding. For motion encoding, we employ a motion encoder \\(\\mathbf{E}_m\\) to learn an implicit latent representation, which captures verbal and non-verbal communicative behaviors from input facial images. \\(\\mathbf{E}_m\\) comprises a convolutional-based feature extractor followed by MLP layers, which compresses the motion feature into a latent code. Note that, we use implicit representation instead of explicit 3DMM coefficients [28] because the expressiveness of 3DMM model is limited, and the commonly used expression coefficients, predicted by SOTA 3D facial estimation methods [5, 7], entangle with face shape to a certain extent." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.909, + 0.901 + ], + "angle": 0, + "content": "A well-crafted motion latent space should demonstrate a high level of disentanglement to ensure that the extracted motions are appearance-independent. To achieve this, we apply two measures. (a) We set the motion latent code as a low-dimensional 1-D descriptor. It fosters effectively capturing the fundamental semantics of facial motion without entangling with appearance information, as explained in previous works [38, 43, 44]. (b) We design a hybrid facial representation to replace the original input image of \\(\\mathbf{E}_m\\). In particular, we first mask the majority of facial pixels, preserving only the eyes and lip regions, as we regard them as the most intricate and expressive components of facial expressions. This can block the interference of irrelevant information such as hair and background for expression distill. To provide facial orientation and contour information, we then use an off-the-shelf facial estimation model [20] to obtain the facial vertices, and project the contour-related points onto the masked image. Our ablation study shows the superiority of our proposed hybrid facial representation" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10669" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.09, + 0.589, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.591, + 0.09, + 0.9, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.333, + 0.908, + 0.391 + ], + "angle": 0, + "content": "Figure 3. Overview of INFP. The first stage (Motion-Based Head Imitation) learns to project facial communicative behaviors from real-life conversation videos into a low-dimensional motion latent space, and use the latent codes to animate a static image. The second stage (Audio-Guided Motion Generation) learns the mapping from the input dyadic audio to motion latent codes through denoising, to achieve the audio-driven interactive head generation." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.416, + 0.482, + 0.446 + ], + "angle": 0, + "content": "compared with the other representations like the intact image [6] or 2D landmarks map [22]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.447, + 0.483, + 0.613 + ], + "angle": 0, + "content": "Head Imitation. In this paper, we adopt a motion flow-based generation pipeline. Concretely, we first extract source motion code \\( m_{self} \\) from the input portrait image \\( I_{self} \\), and driving motion codes \\( m_{dri}^{1:N} \\) from the driving video \\( V_{dri} \\). A motion flow estimation model \\( \\mathbf{F} \\) takes these motion codes as input to predict the self \\( \\rightarrow \\) driving flow. We utilize a face encoder \\( \\mathbf{E}_{face} \\) to extract 3D appearance feature volume from the input image \\( I_{self} \\). This feature volume is then warped by the motion flow and sent into a face decoder \\( \\mathbf{D}_{face} \\) to synthesize the final predictions \\( I_{pred}^{1:N} \\). The overall pipeline can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.618, + 0.483, + 0.644 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l}\\text {F l o w} _ {s \\rightarrow d} = \\mathbf {F} (\\mathbf {E} _ {m} (I _ {\\text {s e l f}}), \\mathbf {E} _ {m} (V _ {\\text {d r i}}))\\\\\\text {I} ^ {1: N} = \\mathbf {D} _ {n} \\left(W _ {n} - \\left(I _ {n}\\right) I _ {n} + \\dots\\right)\\end{array}\\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.639, + 0.453, + 0.657 + ], + "angle": 0, + "content": "\\[\nI _ {p r e d} ^ {1: N} = \\mathbf {D} _ {f a c e} (W a r p (\\mathbf {E} _ {f a c e} (I _ {s e l f}), F l o w _ {s \\rightarrow d}))\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.691 + ], + "angle": 0, + "content": "Model details and loss function are shown in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.698, + 0.388, + 0.713 + ], + "angle": 0, + "content": "3.2. Audio-Guided Motion Generation" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.72, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Previous works [34, 39, 59] tend to divide their models into a Speaker Generator and a Listener Generator. Since the Speaker Generator only receives speaking audio, while the Listener Generator solely processes listening audio, it is necessary to segment the audio of both agent and its conversation partner into speaking clips and listening clips. Hence, in a multi-turn dialogue, explicit role switching is also crucial. We argue that it can lead to unnatural transitions in communicative states, particularly in real-life interaction contexts characterized by frequent role switching. Besides, this design paradigm clearly goes against the intention of the conversational agent, as it should be able to" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.416, + 0.905, + 0.446 + ], + "angle": 0, + "content": "naturally present different communicative states based on the progress of the conversation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.448, + 0.906, + 0.568 + ], + "angle": 0, + "content": "In this paper, we present a new paradigm, in which the model takes as input the audio from both the agent and its conversation partner, eliminating the necessity for explicit role assignment or switching. The key innovations lie in extracting interactive information from dual-track dyadic audio to construct speaking-listening mixed motions, and mapping these motions to the motion latent space of stage 1 to achieve audio-driven head generation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Motion Feature Construction. As shown in Fig. 3, an interactive motion guider is introduced to extract motion feature from the input audio. Within it, there are two memory banks \\( M_v \\) and \\( M_{nv} \\), storing verbal and nonverbal motion, respectively. Each motion memory bank consists of learnable embeddings to memorize typical motions, denoted as \\( e^{1:K} \\), where \\( e^k \\in \\mathbb{R}^d \\) represents the \\( k \\)-th motion embedding and \\( d \\) is the dimension. We set \\( K \\) to 64 and \\( d \\) to 512 in our experiments. When agent is speaking, \\( A_{self} \\) contains rich information, so we take it as \\( Query \\) and obtain verbal motion feature from \\( M_v \\), which is taken as \\( Key \\) and \\( Value \\), through a cross-attention layer. Since the partner is listening, \\( A_{other} \\) carries very little information, hence no effective non-verbal motion feature can be obtained from \\( M_{nv} \\). Therefore, after feature fusion, which involves element-wise sum and several MLP layers, verbal motion feature dominates and drives the agent to express the intended speaking state. However, when the partner starts speaking, \\( A_{self} \\) becomes weak and \\( A_{other} \\) fulfills a vital role. Accordingly, non-verbal motion feature from \\( M_{nv} \\) dominates and forces the agent to present a listening state. In this way, our model can dynamically construct" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10670" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.168 + ], + "angle": 0, + "content": "interactive motion feature \\( f_{m} \\) based on the content of the dialogue audio, allowing the agent to convey the intended state. Note that, \\( A_{\\text{self}} \\) and \\( A_{\\text{other}} \\) are encoded to the feature space through an off-the-shelf HuBERT audio encoder [12], before sent into the model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.175, + 0.483, + 0.356 + ], + "angle": 0, + "content": "Inspired by stylistic facial animation methods [21, 35], we also introduce a motion style vector \\( s_m \\) to explicitly edit the motion embeddings of memory banks, through a style modulation layer raised by StyleGAN2[15]. \\( s_m \\) is first extracted from motion latent codes through a motion style encoder \\( \\mathbf{E}_{style} \\), and then squeezed along the temporal dimension to a global descriptor. We argue that \\( s_m \\) contains some global information like emotions and attitudes, thus enhancing the authenticity and vividness of the generated results. Note that during training, \\( s_m \\) is sourced from an arbitrary video clip of the driven individual, while in testing, \\( s_m \\) can be extracted from any videos or set as null." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.365, + 0.483, + 0.742 + ], + "angle": 0, + "content": "Motion Mapping. To map the interactive motions \\( f_{m} \\) to the pretrained motion latent space of stage 1, we employ a conditional diffusion transformer \\( \\mathbf{T} \\) [11]. Given the data distribution \\( q(m^{1:N},f_m) \\) where \\( m^{1:N} \\) represents the corresponding motion latent codes with \\( N \\) frames, the diffusion model approximates the conditional distribution \\( q(m^{1:N}|f_m) \\). \\( \\mathbf{T} \\) only has 4 blocks, which makes our framework lightweight enough to achieve real-time interaction. Each block consists of a self-attention layer, a motion-attention layer, and a temporal-attention layer. \\( \\mathbf{T} \\) predicts the noise added to the ground-truth motion latent codes for each denoising step. Diffusion timestep is converted into the sinusoidal embedding and then concatenated to the noised motion latent codes at the time dimension. In the motion attention layer, we take the latent feature output from preceding self-attention layer as the Query, while the interactive motion feature \\( f_{m} \\) is utilized as the Key and Value. Inspired byAnimateDiff [10], we implement a temporal layer to ensure a smooth transition between adjacent windows. In particular, we incorporate the last 10 frames of the generated motion latent codes from the previous window as the condition of the current one. Note that in the inference time, we add noise to the self motion latent codes \\( m_{\\text{self}}^{1:N} \\), which are extracted from \\( I_{\\text{self}} \\) and replicated into \\( n \\) copies." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Training Strategy. In the training stage, we randomly set style vector to null with a probability of 0.3, and randomly drop interactive motion feature and previous motion latent codes with a probability of 0.5. Also note that we adopt a warm-up strategy for training stage 2. Specifically, we only select single-sided conversation clips from our training set as \"simple case\" to train the model for several epochs, which helps with the initialization of two memory banks. After that, we randomly sample multi-turn clips from entire dataset for the remaining training process." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.603, + 0.105 + ], + "angle": 0, + "content": "4. Dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.115, + 0.905, + 0.296 + ], + "angle": 0, + "content": "Existing dyadic interaction datasets have limitations in both scale and quality. ViCo [58] is a small-scale dataset with 1.6 hours and 96 IDs, and it lacks multi-turn conversation scenarios. Although ViCo-X [57] features rich multi-turn conversations, the total duration is only 0.4 hours. Besides, its background is monotonous, all of which are green screen. RealTalk [8] is a medium-scale dataset capturing dyadic conversations between pairs of individuals totaling to 115 hours of online public videos. However, it does not restrict the resolution of faces in the video, leading to low face resolution in many videos, thus impacting the quality of model generation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.297, + 0.907, + 0.525 + ], + "angle": 0, + "content": "We propose DyConv, a large-scale dataset of multi-turn dyadic conversations collected from the Internet. The dataset contains video clips with a duration of over 200 hours, captures a wide gamut of emotions and expressions. We employ an off-the-shelf face detection model [51], and keep only frames where both individuals of the conversation are fully visible and their facial resolution is greater than \\(400 \\times 400\\). In addition, we utilize a SOTA speech separation model [54] to separate the audio of two persons in the conversation, denoted as \\(A_{p1}\\) and \\(A_{p2}\\). Then we run an active speaker detection model [1] to match each audio clip to the corresponding face in the original video. We will publicly release the dataset, along with separated audios and annotations. We hope it will encourage the investigation of head generation for dyadic interaction." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.537, + 0.64, + 0.554 + ], + "angle": 0, + "content": "5. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.562, + 0.733, + 0.577 + ], + "angle": 0, + "content": "5.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.906, + 0.688 + ], + "angle": 0, + "content": "In our experiments, we train the diffusion transformer using AdamW optimizer [19], with \\(\\beta_{1} = 0.9\\) and \\(\\beta_{2} = 0.999\\), a learning rate of 1e-4, a weight decay of 1e-2, and a batch size of 32. For inference, we use a DDIM sampler [30] for 20 denoising steps. The CFG parameter of motion feature is set to 2, and the CFG parameter of previous motion latent is set to 1.5." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.698, + 0.655, + 0.714 + ], + "angle": 0, + "content": "5.2. Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.905, + 0.809 + ], + "angle": 0, + "content": "Although originally designed for dyadic interactive scenarios, our method can be naturally generalized to other applications, such as listening head generation and talking head generation. This allows us to fully validate the advancement of our proposed framework against some of the current SOTA methods in related areas." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Evaluation Metrics. To quantitatively evaluate the visual quality, we utilize the Structured Similarity (SSIM), Peak Signal-to-Noise Ratio (PSNR) and Frechet Inception Distance (FID) for the generated videos. Additionally, LPIPS [50] is employed to measure the feature-level similarity between generated and ground-truth frames. To as" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10671" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.129, + 0.175, + 0.223, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.132, + 0.246, + 0.222, + 0.258 + ], + "angle": 0, + "content": "Reference Image" + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.094, + 0.877, + 0.611 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.621, + 0.908, + 0.664 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison of interactive head generation on DyConv. Given the input audio of dyadic conversations (the audio of the predicted agent is black and the audio of the conversation partner is blue), our framework generates reasonable and vivid facial and head pose movements as various roles in interactive scenarios." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.483, + 0.812 + ], + "angle": 0, + "content": "sess the preservation of facial appearance, we calculate the cosine similarity (CSIM) [4] between the facial features of the reference image and the generated video frames. In term of motion diversity of generated videos, we use SI for diversity (SID) and Variance (Var) following previous studies [23, 39]. We use the confidence score of SyncNet (SyncScore) [3] to evaluate the synchronization between lip movements and audio signals of the agent." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.821, + 0.338, + 0.835 + ], + "angle": 0, + "content": "5.2.1. Interactive Head Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Since only a few works focus on the interactive head generation [42, 59] and these methods are not open-sourced, direct comparisons with them are considerably difficult. To effectively validate our method, we select the SOTA dyadic" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.907, + 0.809 + ], + "angle": 0, + "content": "modeling work DIM [39] as our baseline and utilize the officially released code to re-train the model on our datasets. Specifically, we first use the whole dataset to pretrain the VQ-VAE for the listener and speaker and then split the dataset into two subsets to train the listener generator and the speaker generator, respectively. To better visualize results of DIM, we also train the face renderer [28] to generate photo-realistic videos." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Quantitative Results. The quantitative experiments are carried out on the test set of DyConv. Please refer to Tab. 1 for the results. It can be seen that our method outperforms the baseline cross a variety of metrics. In particular, our face decoder achieves a better visual quality such as SSIM, PSNR and FID. Thanks to the disentanglement and expres" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10672" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.257 + ], + "angle": 0, + "content": "siveness of our latent space, our method is superior to DIM in term of audio-lip-synchronization by a large margin as there is a much closer distance of our SyncScore to the ground truth. As for the LPIPS and CISM, our method still achieves solid performances, reflecting the better identity preservation. Of particular note, our method is far ahead of the baseline in term of SID and Var, which proves the generalization and effectiveness of our method to generate a great range of diverse motions. It benefits from the design of interactive motion guider as well as conditional denoising transformer." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.258, + 0.483, + 0.573 + ], + "angle": 0, + "content": "Qualitative Results. Subjective evaluation is crucial in identifying the performance of generative models, particularly on videos. We strongly recommend readers to watch our supplementary video. Two examples and their comparisons with the baseline are shown in Fig. 4. We use two typical audio clips in the dyadic conversation scenario to drive reference identities randomly selected from the test set. It can be obviously observed that our method can generate more diverse head pose movements in comparison to DIM. Meanwhile, results of our method demonstrate more vivid facial expressions, which significantly improve the naturalness of generated videos. It is worth to noting that facial behaviors synthesized by our method are well aligned with the dialogue audio content. This suggests the effectiveness of our method in dyadic interactions, which is capable of capturing distinct characteristics of various states, such as listening, speaking as well as the mutual interaction, and naturally adapting its role among them based on the progress of the conversation. In contrast, DIM is unable to generate appropriate behaviors such as reasonable lip movement and response feedback." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.575, + 0.483, + 0.741 + ], + "angle": 0, + "content": "User Study. We also invite 20 participants to conduct a user study for further subjective evaluation. Specifically, we apply INFP and DIM to synthesize results of 20 dyadic conversation videos randomly sampled from the test set. By adopting the commonly used Mean Opinion Scores (MOS) rating protocol, we request all the participants to provide their ratings (from 1 to 5, the higher the better) on four aspects for each generated video: naturalness, motion diversity, audio-visual alignment and visual quality. As detailed in Tab. 2, our method demonstrates superior performances across all evaluated aspects compared to DIM." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.747, + 0.327, + 0.761 + ], + "angle": 0, + "content": "5.2.2. Listening Head Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Since our method can generalize to the task of listening head generation, we directly use our model without any modification to conduct additional experiments. Specifically, we compare our framework with SOTA listening head generation methods, including L2L [23], RLHG [58] and DIM [39]. ViCo [58] is selected as the benchmark. For quantitative evaluation, we utilized the same test set and inherited both quantitative and qualitative results presented in DIM. We conducted the comparison on FD, MSE, SID" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.318 + ], + "angle": 0, + "content": "and Var and each metric is calculated on expression and head pose, respectively. Different to other methods, our method generate photo-realistic video rather than 3DMM coefficients. To this end, we use the off-the-shelf 3D face reconstruction method [7] to extract expression and head pose coefficients from our generated videos. As shown in Tab. 3, INFP outperforms other SOTA methods in the majority of metrics. While SID and Var are performing well, our method remains highly effective in both FD and MSE. This indicates the power of our novel framework design and dataset quality, making generated facial expressions and head pose movements closer to real human behaviors. The visualization results in the supplementary materials also demonstrate that INFP achieves more expressive and diverse motions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.329, + 0.738, + 0.344 + ], + "angle": 0, + "content": "5.2.3. Talking Head Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.349, + 0.907, + 0.606 + ], + "angle": 0, + "content": "For talking head generation, we select SadTalker[52], AntiTalker[17], and EchoMimic[2] as comparing methods. As all of them reported evaluation results on HDTF [53], we randomly sampled 20 videos from it as the test set to conduct our experiments for a fair comparison. Results in Tab. 4 show that our model is superior in audiolip-synchronization and identity preservation. The reason behind this might be that we carefully build a disentangled and expressive motion latent space and it makes our model easier to learn the correlation between audio and lip movements without the interference of appearance features. It can also be observed from the visualization results in the supplementary materials. Our SSIM and PSNR score is slightly behind EchoMimic because EchoMimic is finetuned based on the Stable Diffusion V1.5 [29]. Despite this, our method can achieve better scores in terms of FID, meaning a good visual quality of our results." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.619, + 0.667, + 0.635 + ], + "angle": 0, + "content": "5.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.642, + 0.905, + 0.671 + ], + "angle": 0, + "content": "We ablate 3 key components of our framework, and the results are listed in Tab. 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.673, + 0.906, + 0.839 + ], + "angle": 0, + "content": "Motion Memory Bank. We remove the verbal and non-verbal motion memory banks from interactive motion guider in stage 2, and directly use the dyadic audio feature after several MLP layers as the condition for the motion-attention layer. Results show that the generated lip-related motion behaviors are clearly inconsistent with the ground-truth. In additional, there is a significant drop of both SID and Var scores. These suggest that memory banks can effectively store rich motion patterns in different conversational states. The experiment on memory bank size is shown in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Style Modulation. We remove the motion style vector \\( s_m \\) and corresponding modulation layers. As we expected, the authenticity and vividness of the results decrease, which can also be reflected in the user study." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10673" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.144, + 0.089, + 0.858, + 0.24 + ], + "angle": 0, + "content": "
MethodsSSIM↑PSNR↑FID↓SyncScore↑LPIPS↓CSIM↑SID↑Var↑
DIM [39]0.65120.41734.3614.7780.4850.8240.7660.825
INFP (Ours)0.83431.56215.7277.1880.2570.9042.6132.386
w/o Motion Memory0.83031.21818.3346.1030.2590.8992.1532.016
w/o Style Modulation0.83131.44216.0297.0620.2710.9042.5512.316
w/ Intact Image0.80228.48816.9906.8120.2660.8422.4702.148
w/ Landmarks Map0.82130.69316.3276.8330.2810.9012.6012.335
GT1.000N/A0.0007.2610.0000.9672.8912.435
" + }, + { + "type": "table_caption", + "bbox": [ + 0.272, + 0.25, + 0.724, + 0.264 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison of interactive head generation on DyConv." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.287, + 0.481, + 0.354 + ], + "angle": 0, + "content": "
MethodsNaturalnessMotion DiversityAV-AlignVisual
DIM [39]2.712.142.653.13
INFP w/o Style4.023.984.284.11
INFP (Ours)4.384.494.334.13
" + }, + { + "type": "table_caption", + "bbox": [ + 0.12, + 0.363, + 0.454, + 0.377 + ], + "angle": 0, + "content": "Table 2. User study measured by Mean Opinion Scores." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.391, + 0.481, + 0.487 + ], + "angle": 0, + "content": "
MethodFD↓MSE↓SID↑Var↑
ExpPoseExpPoseExpPoseExpPose
L2L [23]33.930.060.930.012.772.660.830.02
RLHG [58]39.020.070.860.013.623.171.520.02
DIM [39]23.880.060.700.013.712.351.530.02
INFP (Ours)18.630.070.510.014.783.922.830.18
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.497, + 0.483, + 0.525 + ], + "angle": 0, + "content": "Table 3. Quantitative comparison of listening head generation on ViCo [58]." + }, + { + "type": "table", + "bbox": [ + 0.098, + 0.539, + 0.478, + 0.624 + ], + "angle": 0, + "content": "
MethodsSSIM↑PSNR↑FID↓SyncScore↑CSIM↑
SadTalker [52]0.78126.65028.5166.1250.814
AniTalker [17]0.78926.82228.7785.7380.858
EchoMimic [2]0.80728.98320.8736.2510.863
INFP (Ours)0.80428.65718.5826.6700.886
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.633, + 0.483, + 0.661 + ], + "angle": 0, + "content": "Table 4. Quantitative comparison of talking head generation on HDTF [53]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.688, + 0.483, + 0.765 + ], + "angle": 0, + "content": "Hybrid Facial Representation. We change the input to \\(\\mathbf{E}_m\\) in stage 1 from our carefully-designed hybrid facial representation to the original intact image or 2D landmarks map. We observe that there is degradation in generation quality and a leakage of appearance information." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.773, + 0.355, + 0.789 + ], + "angle": 0, + "content": "5.4. Visualization of Style Control" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.9 + ], + "angle": 0, + "content": "As illustrated in Sec. 3.2, our model supports inputting style vector \\( s_m \\) extracted from any portrait video to globally control the emotions or attitudes of the generated results. Here we select 3 video clips with varying degrees of \"happiness\" to control emotions during generation. Results are shown in Fig. 5. It can be seen from the figure that \\( s_m \\) effectively control the emotions of the driven individual." + }, + { + "type": "image", + "bbox": [ + 0.537, + 0.289, + 0.888, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.514, + 0.478, + 0.905, + 0.492 + ], + "angle": 0, + "content": "Figure 5. Visualization of style control with our style modulation." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.522, + 0.634, + 0.537 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.549, + 0.907, + 0.731 + ], + "angle": 0, + "content": "In this paper, we propose INFP, a novel audio-driven interactive head generation framework for dyadic conversations, which aims to mimic facial behaviors of human-human interactions conditioning on the dyadic audio. As demonstrated, our method is a unified framework and is capable of smoothly and naturally adapting to various conversational roles without explicit assignment or switching. Meanwhile, we collect a large scale audio-visual dataset for multi-turn dyadic conversations, which could facilitate further researches of dyadic interactive head generation. Extensive experiments and ablations prove the advancement and effectiveness of our method over other SOTA methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.733, + 0.907, + 0.839 + ], + "angle": 0, + "content": "Limitation and Future Work. Currently, our method only takes conversation audios as the input, combining control signals from multiple modalities, such as visual and textual contents, can offer additional capacities. As our proposed framework focuses on head synthesis, extending the generation range to the upper, even whole body might be a interesting and challenging future work." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Ethical Considerations. There might be the potential misuse of our method to fabricate deceptive talks and speeches. To this end, we will restrict access to our core models exclusively to research institutions." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10674" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.482, + 0.156 + ], + "angle": 0, + "content": "[1] Stanislav Beliaev, Yuriy Rebryk, and Boris Ginsburg. Talknet: Fully-convolutional non-autoregressive speech synthesis model, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.159, + 0.483, + 0.214 + ], + "angle": 0, + "content": "[2] Zhiyuan Chen, Jiajiong Cao, Zhiquan Chen, Yuming Li, and Chenguang Ma. Echomimic: Lifelike audio-driven portrait animations through editable landmark conditions. arXiv preprint arXiv:2407.08136, 2024. 1, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.217, + 0.482, + 0.257 + ], + "angle": 0, + "content": "[3] J. S. Chung and A. Zisserman. Out of time: automated lip sync in the wild. In Workshop on Multi-view Lip-reading, ACCV, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.26, + 0.482, + 0.302 + ], + "angle": 0, + "content": "[4] Jiankang Deng, Jia Guo, Xue Niannan, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In CVPR, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.304, + 0.482, + 0.385 + ], + "angle": 0, + "content": "[5] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.389, + 0.482, + 0.443 + ], + "angle": 0, + "content": "[6] Nikita Drobyshev, Jenya Chelishev, Taras Khakhulin, Aleksei Ivakhnenko, Victor Lempitsky, and Egor Zakharov. Megaportraits: One-shot megapixel neural head avatars. 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.446, + 0.482, + 0.514 + ], + "angle": 0, + "content": "[7] Panagiotis P. Filntisis, George Retsinas, Foivos Paraperas-Papantoniou, Athanasios Katsamanis, Anastasios Roussos, and Petros Maragos. Visual speech-aware perceptual 3d facial expression reconstruction from videos. arXiv preprint arXiv:2207.11094, 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.517, + 0.482, + 0.559 + ], + "angle": 0, + "content": "[8] Scott Geng, Revant Teotia, Purva Tendulkar, Sachit Menon, and Carl Vondrick. Affective faces for goal-driven dyadic communication. 2023. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.562, + 0.482, + 0.657 + ], + "angle": 0, + "content": "[9] Jiazhi Guan, Zhanwang Zhang, Hang Zhou, Tianshu Hu, Kaisiyuan Wang, Dongliang He, Haocheng Feng, Jingtuo Liu, Errui Ding, Ziwei Liu, et al. Stylesync: High-fidelity generalized and personalized lip sync in style-based generator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1505-1515, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.66, + 0.482, + 0.729 + ], + "angle": 0, + "content": "[10] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Imagine your personalized text-to-image diffusion models without specific tuning. International Conference on Learning Representations, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.731, + 0.482, + 0.772 + ], + "angle": 0, + "content": "[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.775, + 0.482, + 0.856 + ], + "angle": 0, + "content": "[12] Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. Hubert: Self-supervised speech representation learning by masked prediction of hidden units. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 29:3451-3460, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[13] Ailin Huang, Zhewei Huang, and Shuchang Zhou. Perceptual conversational head generation with regularized driver and enhanced renderer. In Proceedings of the 30th ACM" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "International Conference on Multimedia, page 7050-7054. ACM, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[14] Yinghao Huang, Leo Ho, Dafei Qin, Mingyi Shi, and Taku Komura. Interact: Capture and modelling of realistic, expressive and interactive activities between two persons in daily scenarios. 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.221 + ], + "angle": 0, + "content": "[15] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In Proc. CVPR, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.223, + 0.905, + 0.291 + ], + "angle": 0, + "content": "[16] Jin Liu, Xi Wang, Xiaomeng Fu, Yesheng Chai, Cai Yu, Jiao Dai, and Jizhong Han. Mfr-net: Multi-faceted responsive listening head generation via denoising diffusion model. In Proceedings of the 31st ACM International Conference on Multimedia, pages 6734-6743, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.294, + 0.905, + 0.348 + ], + "angle": 0, + "content": "[17] Tao Liu, Feilong Chen, Shuai Fan, Chenpeng Du, Qi Chen, Xie Chen, and Kai Yu. Anitalker: Animate vivid and diverse talking faces through identity-decoupled facial motion encoding, 2024. 1, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.35, + 0.905, + 0.419 + ], + "angle": 0, + "content": "[18] Xi Liu, Ying Guo, Cheng Zhen, Tong Li, Yingying Ao, and Pengfei Yan. Custom listener: Text-guided responsive interaction for user-friendly listening head generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2415-2424, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.421, + 0.905, + 0.449 + ], + "angle": 0, + "content": "[19] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.451, + 0.905, + 0.532 + ], + "angle": 0, + "content": "[20] Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuo-Ling Chang, Ming Yong, Juhyun Lee, Wan-Teh Chang, Wei Hua, Manfred Georg, and Matthias Grundmann. Mediapipe: A framework for perceiving and processing reality. In CVPR 2019, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.535, + 0.905, + 0.645 + ], + "angle": 0, + "content": "[21] Yifeng Ma, Suzhen Wang, Zhipeng Hu, Changjie Fan, Tangjie Lv, Yu Ding, Zhidong Deng, and Xin Yu. Styletalk: one-shot talking head generation with controllable speaking styles. In Proceedings of the Thirty-Seventh AAAI Conference on Artificial Intelligence and Thirty-Fifth Conference on Innovative Applications of Artificial Intelligence and Thirteenth Symposium on Educational Advances in Artificial Intelligence. AAAI Press, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.648, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[22] Yue Ma, Hongyu Liu, Hongfa Wang, Heng Pan, Yingqing He, Junkun Yuan, Ailing Zeng, Chengfei Cai, Heung-Yeung Shum, Wei Liu, et al. Follow-your-emoji: Fine-controllable and expressive freestyle portrait animation. arXiv preprint arXiv:2406.01900, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.718, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[23] Evonne Ng, Hanbyul Joo, Liwen Hu, Hao Li, Trevor Darrell, Angjoo Kanazawa, and Shiry Ginosar. Learning to listen: Modeling non-deterministic dyadic facial motion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20395-20405, 2022. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.803, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[24] Evonne Ng, Sanjay Subramanian, Dan Klein, Angjoo Kanazawa, Trevor Darrell, and Shiry Ginosar. Can language models learn to listen? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10083-10093, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[25] Evonne Ng, Javier Romero, Timur Bagautdinov, Shaojie Bai, Trevor Darrell, Angjoo Kanazawa, and Alexander Richard." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10675" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.147 + ], + "angle": 0, + "content": "From audio to photoreal embodiment: Synthesizing humans in conversations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1001-1010, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.204 + ], + "angle": 0, + "content": "[26] Se Jin Park, Chae Won Kim, Hyeongseop Rha, Minsu Kim, Joanna Hong, Jeong Hun Yeo, and Yong Man Ro. Let's go real talk: Spoken dialogue model for face-to-face conversation. arXiv preprint arXiv:2406.07867, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.205, + 0.484, + 0.287 + ], + "angle": 0, + "content": "[27] K R Prajwal, Rudrabha Mukhopadhyay, Vinay P. Namboodiri, and C.V. Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM International Conference on Multimedia, page 484-492, New York, NY, USA, 2020. Association for Computing Machinery. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.288, + 0.483, + 0.329 + ], + "angle": 0, + "content": "[28] Yurui Ren, Ge Li, Yuanqi Chen, Thomas H. Li, and Shan Liu. Pirenderer: Controllable portrait image generation via semantic neural rendering, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.33, + 0.483, + 0.371 + ], + "angle": 0, + "content": "[29] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.372, + 0.483, + 0.399 + ], + "angle": 0, + "content": "[30] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv:2010.02502, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.401, + 0.483, + 0.468 + ], + "angle": 0, + "content": "[31] Luchuan Song, Guojun Yin, Zhenchao Jin, Xiaoyi Dong, and Chenliang Xu. Emotional listener portrait: Realistic listener motion simulation in conversation. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 20782-20792. IEEE, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.47, + 0.483, + 0.538 + ], + "angle": 0, + "content": "[32] Siyang Song, Micol Spitale, Cheng Luo, Cristina Palmero, German Barquero, Hengde Zhu, Sergio Escalera, Michel Valstar, Tobias Baur, Fabien Ringeval, Elisabeth Andre, and Hatice Gunes. React 2024: the second multiple appropriate facial reaction generation challenge, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.483, + 0.609 + ], + "angle": 0, + "content": "[33] Michal Stypulkowski, Konstantinos Vougioukas, Sen He, Maciej Zieba, Stavros Petridis, and Maja Pantic. Diffused heads: Diffusion models beat gans on talking-face generation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5091-5100, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.61, + 0.483, + 0.663 + ], + "angle": 0, + "content": "[34] Mingze Sun, Chao Xu, Xinyu Jiang, Yang Liu, Baigui Sun, and Ruqi Huang. Beyond talking—generating holistic 3d human dyadic motion for communication. arXiv preprint arXiv:2403.19467, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.665, + 0.483, + 0.734 + ], + "angle": 0, + "content": "[35] Zhiyao Sun, Tian Lv, Sheng Ye, Matthieu Lin, Jenny Sheng, Yu-Hui Wen, Minjing Yu, and Yong-jin Liu. Diffpositalk: Speech-driven stylistic 3d facial animation and head pose generation via diffusion models. ACM Transactions on Graphics (TOG), 43(4):1-9, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.483, + 0.803 + ], + "angle": 0, + "content": "[36] Jiaxiang Tang, Kaisiyuan Wang, Hang Zhou, Xiaokang Chen, Dongliang He, Tianshu Hu, Jingtuo Liu, Gang Zeng, and Jingdong Wang. Real-time neural radiance talking portrait synthesis via audio-spatial decomposition. arXiv preprint arXiv:2211.12368, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.805, + 0.483, + 0.859 + ], + "angle": 0, + "content": "[37] Linrui Tian, Qi Wang, Bang Zhang, and Liefeng Bo. Emo: Emote portrait alive-generating expressive portrait videos with audio2video diffusion model under weak conditions. arXiv preprint arXiv:2402.17485, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[38] Naftali Tishby, Fernando C Pereira, and William Bialek. The information bottleneck method. arXiv preprint physics/0004057, 2000. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[39] Minh Tran, Di Chang, Maksim Siniukov, and Mohammad Soleymani. Dyadic interaction modeling for social behavior generation. arXiv preprint arXiv:2403.09069, 2024. 1, 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.188 + ], + "angle": 0, + "content": "[40] Aaron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu. Neural discrete representation learning, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.19, + 0.906, + 0.259 + ], + "angle": 0, + "content": "[41] Cong Wang, Kuan Tian, Jun Zhang, Yonghang Guan, Feng Luo, Fei Shen, Zhiwei Jiang, Qing Gu, Xiao Han, and Wei Yang. V-express: Conditional dropout for progressive training of portrait video generation. arXiv preprint arXiv:2406.02511, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[42] Duomin Wang, Bin Dai, Yu Deng, and Baoyuan Wang. Disentangling planning, driving and rendering for photorealistic avatar agents. 2023. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.303, + 0.905, + 0.383 + ], + "angle": 0, + "content": "[43] Duomin Wang, Yu Deng, Zixin Yin, Heung-Yeung Shum, and Baoyuan Wang. Progressive disentangled representation learning for fine-grained controllable talking head synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17979-17989, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.385, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[44] Yaohui Wang, Di Yang, Francois Bremond, and Antitza Dantcheva. Latent image animator: Learning to animate images via latent space navigation. arXiv preprint arXiv:2203.09043, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.509 + ], + "angle": 0, + "content": "[45] Mingwang Xu, Hui Li, Qingkun Su, Hanlin Shang, Liwei Zhang, Ce Liu, Jingdong Wang, Luc Van Gool, Yao Yao, and Siyu Zhu. Hallo: Hierarchical audio-driven visual synthesis for portrait image animation. arXiv preprint arXiv:2406.08801, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.511, + 0.905, + 0.579 + ], + "angle": 0, + "content": "[46] Sicheng Xu, Guojun Chen, Yu-Xiao Guo, Jiaolong Yang, Chong Li, Zhenyu Zang, Yizhong Zhang, Xin Tong, and Baining Guo. Vasa-1: Lifelike audio-driven talking faces generated in real time. arXiv preprint arXiv:2404.10667, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.581, + 0.905, + 0.622 + ], + "angle": 0, + "content": "[47] Yichao Yan, Zanwei Zhou, Zi Wang, Jingnan Gao, and Xiaokang Yang. Dialoguenerf: Towards realistic avatar face-to-face conversation video generation, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.623, + 0.905, + 0.692 + ], + "angle": 0, + "content": "[48] Chenxu Zhang, Chao Wang, Jianfeng Zhang, Hongyi Xu, Guoxian Song, You Xie, Linjie Luo, Yapeng Tian, Xiaohu Guo, and Jiashi Feng. Dream-talk: diffusion-based realistic emotional audio-driven method for single image talking face generation. arXiv preprint arXiv:2312.13578, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.693, + 0.905, + 0.734 + ], + "angle": 0, + "content": "[49] Longhao Zhang, Shuang Liang, Zhipeng Ge, and Tianshu Hu. Personatack: Bring attention to your persona in visual dubbing. arXiv preprint arXiv:2409.05379, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.905, + 0.775 + ], + "angle": 0, + "content": "[50] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.777, + 0.905, + 0.816 + ], + "angle": 0, + "content": "[51] Shifeng Zhang, Xiangyu Zhu, Zhen Lei, Hailin Shi, Xiaobo Wang, and Stan Z. Li. S\\(^3\\)fd: Single shot scale-invariant face detector, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[52] Wenxuan Zhang, Xiaodong Cun, Xuan Wang, Yong Zhang, Xi Shen, Yu Guo, Ying Shan, and Fei Wang. Sadtalker: Learning realistic 3d motion coefficients for stylized audio-driven single image talking face animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8652-8661, 2023. 1, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10676" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.161 + ], + "angle": 0, + "content": "[53] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661–3670, 2021. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.483, + 0.233 + ], + "angle": 0, + "content": "[54] Shengkui Zhao, Yukun Ma, Chongjia Ni, Chong Zhang, Hao Wang, Trung Hieu Nguyen, Kun Zhou, Jiaqi Yip, Dianwen Ng, and Bin Ma. Mossformer2: Combining transformer and rnnprecent network for enhanced time-domain monaural speech separation, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.483, + 0.289 + ], + "angle": 0, + "content": "[55] Wei Zhao, Peng Xiao, Rongju Zhang, Yijun Wang, and Jianxin Lin. Semantic-aware responsive listener head synthesis. In Proceedings of the 30th ACM International Conference on Multimedia, pages 7065-7069, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.483, + 0.359 + ], + "angle": 0, + "content": "[56] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking face generation by implicitly modularized audio-visual representation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.361, + 0.483, + 0.415 + ], + "angle": 0, + "content": "[57] Mohan Zhou, Yalong Bai, Wei Zhang, Ting Yao, Tiejun Zhao, and Tao Mei. Vico-x: Multimodal conversation dataset. https://project.mhzhou.com/vico, 2022. Accessed: 2022-09-30. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.418, + 0.483, + 0.485 + ], + "angle": 0, + "content": "[58] Mohan Zhou, Yalong Bai, Wei Zhang, Ting Yao, Tiejun Zhao, and Tao Mei. Responsive listening head generation: A benchmark dataset and baseline. In Proceedings of the European conference on computer vision (ECCV), 2022. 2, 3, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.488, + 0.483, + 0.53 + ], + "angle": 0, + "content": "[59] Mohan Zhou, Yalong Bai, Wei Zhang, Ting Yao, and Tiejun Zhao. Interactive conversational head generation. arXiv preprint arXiv:2307.02090, 2023. 1, 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.531, + 0.483, + 0.586 + ], + "angle": 0, + "content": "[60] Yang Zhou, Xintong Han, Eli Shechtman, Jose Echevarria, Evangelos Kalogerakis, and Dingzeyu Li. Makelttalk: speaker-aware talking-head animation. ACM Transactions on Graphics, 39(6):1-15, 2020. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.586 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "10677" + } + ] +] \ No newline at end of file diff --git a/2025/INFP_ Audio-Driven Interactive Head Generation in Dyadic Conversations/4440ddb1-10e9-45a7-942b-205ac25fbaaa_origin.pdf b/2025/INFP_ Audio-Driven Interactive Head Generation in Dyadic Conversations/4440ddb1-10e9-45a7-942b-205ac25fbaaa_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..23d0fefa101a209621e32a4b18b8c95d718df3c6 --- /dev/null +++ b/2025/INFP_ Audio-Driven Interactive Head Generation in Dyadic Conversations/4440ddb1-10e9-45a7-942b-205ac25fbaaa_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7eb3e0d0e850ec8eb358fef8f0e03441932dc7f868c7022bd6f897d427b28f49 +size 8655693 diff --git a/2025/INFP_ Audio-Driven Interactive Head Generation in Dyadic Conversations/full.md b/2025/INFP_ Audio-Driven Interactive Head Generation in Dyadic Conversations/full.md new file mode 100644 index 0000000000000000000000000000000000000000..15052112a1f1d8c29dac6acba876e6ef7742ace8 --- /dev/null +++ b/2025/INFP_ Audio-Driven Interactive Head Generation in Dyadic Conversations/full.md @@ -0,0 +1,263 @@ +# INFP: Audio-Driven Interactive Head Generation in Dyadic Conversations + +Yongming Zhu* Longhao Zhang* Zhengkun Rong* Tianshu Hu*† Shuang Liang Zhipeng Ge Bytedance + +{zhuyongming, zhanglonghao.zlh, rongzhengkun, tianshu.hu, liangshuang. echo, zhipengge}@bytedance.com + +![](images/4e20afd929fea7b3889fc607e28cc363064719c32fbc24bc1a901cd98e39189c.jpg) +Figure 1. We present INFP, an audio-driven interactive head generation framework for dyadic conversations. Given the dual-track audio in dyadic conversations and a single portrait image of arbitrary agent, our framework can dynamically synthesize verbal, non-verbal and interactive agent videos with lifelike facial expressions and rhythmic head pose movements. Additionally, our framework is lightweight yet powerful, making it practical in instant communication scenarios with acceptable latency, such as the video conferencing. INFP denotes our method is Interactive, Natural, Flash and Person-generic. + +# Abstract + +Imagine having a conversation with a socially intelligent agent. It can attentively listen to your words and offer visual and linguistic feedback promptly. This seamless interaction allows for multiple rounds of conversation to flow smoothly and naturally. In pursuit of actualizing it, we propose INFP, a novel audio-driven head generation framework for dyadic interaction. Unlike previous head generation works that only focus on single-sided communication, or require manual role assignment and explicit role switching, our model drives the agent portrait dynamically alternates between speaking and listening state, guided by the input dyadic audio. Specifically, INFP comprises a Motion-Based Head Imitation stage and an Audio-Guided Motion Generation stage. The first stage learns to project facial communicative behaviors from real-life conversation videos into a low- + +dimensional motion latent space, and use the motion latent codes to animate a static image. The second stage learns the mapping from the input dyadic audio to motion latent codes through denoising, leading to the audio-driven head generation in interactive scenarios. To facilitate this line of research, we introduce DyConv, a large scale dataset of rich dyadic conversations collected from the Internet. Extensive experiments and visualizations demonstrate superior performance and effectiveness of our method. + +# 1. Introduction + +In recent years, researchers have dedicated considerable attention to audio-driven head generation [9, 18, 23, 24, 31, 37, 39, 42, 46, 56, 59], in order to build the conversational agent. However, most studies only focus on single-sided communication, such as talking or listening, ignoring the dyadic properties in human-human interaction. Talking-head generation [2, 17, 27, 36, 37, 41, 45, 46, 48, 49, 52, 56, + +60] aims at synthesizing speaker's facial animations based on their reference images and a driving audio. Although these works can produce vibrant videos with accurate lip synchronization, they solely emphasize the speaker's role and overlook the listener's feedback. Listening-head generation [16, 18, 23, 24, 31, 32, 55, 58], on the other hand, intends to respond to speaker's behaviors. However, these works restrict the listener's response solely to non-verbal facial movements, which is clearly far from real-life interaction scenarios. Some recent studies [39, 42, 59] have started exploring head generation for dyadic interaction, but they require manual role assignment between Listener and Speaker and cannot achieve smooth and natural role switching. We believe that for a conversational agent, it should not be predetermined a role, but rather be able to freely switch between listening and speaking states based on the current dialogue context. When having a conversation with the agent, you may interrupt or respond to it at any moment. At that point, it naturally transitions from a speaker to a listener, and vice versa. + +In this paper, we propose a novel audio-driven head generation framework for dyadic interaction, namely INFP. Unlike previous works, we do not divide the model into Speaker Generator and Listener Generator, or apply explicit role switching. The driven individual dynamically alternates between the speaking and listening state, guided by the dyadic audio. Specifically, INFP comprises a Motion-Based Head Imitation stage and an Audio-Guided Motion Generation stage. In the first stage, our model learns to extract communicative behavior, including non-verbal listening cues and verbal speaking patterns, from massive real-life conversational videos and encode them to a motion latent space. The motion latent codes are then used to animate a static portrait image to authentically imitate the character in the video. A well-crafted motion latent space should exhibit a high degree of disentanglement, which means that head pose, facial expression, and emotion should be decoupled from appearance. To this end, we perform facial structure discretization and facial pixel masking to the input. While in the second stage, the model is trained to map the dyadic audio to the motion latent space pretrained in stage 1 to obtain the corresponding motion latent codes. It consists of an interactive motion guider and a conditional diffusion transformer. The former takes as input the audio from both agent and its conversation partner, and retrieves verbal and non-verbal motions from learnable memory banks to construct the interactive motion feature. The latter utilizes the interactive motion feature as the condition and, together with other signals, generates the motion latent codes through denoising. + +To support our research on interactive head generation, we introduce DyConv, a large scale dataset of rich dyadic conversations collected from the Internet. It contains non + +![](images/b22d306760e5e765f73694d335d631ca5064a91d6ceaa1fe679c4df1e33bc84d.jpg) +Figure 2. Objective illustration. Existing interactive head generation (left) applied manual role assignment and explicit role switching. Our proposed INFP (right) is a unified framework which can dynamically and naturally adapt to various conversational states. + +![](images/2fa039f3238f285f4a7c9432007d5d12fb09c18f1ddbf2ecd825e0d3b7b172c0.jpg) + +scripted video conversations between pairs of individuals from diverse backgrounds. The videos capture their discussions on a wide range of topics and emotions, portraying authentic communication in real-life settings. Compared to previous datasets like ViCo [58], ViCo-X [57] and RealTalk [8], DyConv has a larger scale, higher quality, and richer interactions. + +Our main contributions are summarized as follows: + +- We propose INFP, an audio-driven interactive head generation framework with a new paradigm. To our best knowledge, we are the first to enable the driven individual to naturally transition between listening and speaking states during multi-turn conversations, without the requirement for manual role assignment or explicit role switching. +- We design a novel interactive motion guider with learnable memory banks to adaptively extract interactive information from dual-track dyadic audio and construct speaking-listening mixed motions. +- We introduce a large-scale and high-quality dataset called DyConv. We hope it will encourage the investigation of head generation for dyadic interaction. +- Our framework achieves real-time and person-generic head generation. Extensive experiments and visualizations demonstrate superior performance and effectiveness of our method. + +# 2. Related Work + +# 2.1. Single-Sided Audio-Driven Head Generation + +Listening Head Generation. As a pioneer, RLHG [58] contributes the ViCo dataset as a common benchmark and builds a baseline method with a sequential decoder. Following that, PCH [13] introduces the enhanced renderer to improve the visual quality of generated videos. Song et al. [32] organizes the REACT challenge to encourage the community to investigate the face reaction in offline and online setting, separately. L2L [23] leverages the VQ-VAE [40] to store head motion sequences in discrete codebooks, + +leading to the better motion diversity. ELP [31] rearranges the latent space based on emotional priors to emphasize the emotion of generated motions. To get rid of simple emotional labels, CustomListener [18] proposes to control listener head motions via the text prior. MFR-Net [16] makes a progress of identity preservation and motion diversity. RealTalk [8] proposes a medium-scale dataset and attempts to retrieve possible videos of listener head with large language model. Ng et al. [24] presents a transfer-based approach from pretrained large language models to predict the motion sequences. + +Talking Head Generation Audio-driven talking head generation has become a prominent research area in recent years. Early GAN-based approaches, like Wav2Lip [27], PC-AVS [56], and MakeItTalk [60], directly integrate the input audio and video to produce lip-synced visuals. VASA-1 [46] takes a significant leap forward in the field by decoupling latent feature spaces, enabling the generation of more realistic video outputs. This model can operate on a single image rather than requiring a full video sequence. With the emergence of diffusion models, several works [33, 37, 41] are proposed to enhance the generation quality. EchoMimic [2] is concurrently trained using both audios and facial landmarks, and is capable of generating portrait videos not only by audios and facial landmarks individually, but also by a combination of both audios and selected facial landmarks. AniTalker [56], employing two self-supervised learning strategies, excels in capturing intricate facial movements. + +# 2.2. Head Generation in Dyadic Conversations + +Recently, the research community has paid increasing attention on the audio-driven head generation in the dyadic interaction. Zhou et al. [59] extend RHLG [58] by introducing ViCo-X, a multi-turn conversational dataset, and a Role Switcher to connect the Listener Generator and the Speaker Generator. However, explicit role switching can lead to unnaturalness and inconsistency, and this paradigm fails to cover all dyadic conversation states, such as simultaneous speaking. DIM [39] uses a pretraining approach to model both speakers' and listeners' motions, capturing dyadic context. However, the pretrained model requires additional fine-tuning for tasks like talking head and listening head generation, necessitating manual role assignment in dyadic conversations, which can cause inappropriate transitions. AgentAvatar [42] designs a streamlined pipeline to predict behaviors of a photo-realistic avatar agent in dyadic interactions. While the motion synthesis is conditioned on the high-level textual description, AgentAvatar fails to generate facial behaviors that precisely align with the dialogue audio, as demo videos shown in their project page. Additionally, there are some other studies in dyadic interactions [14, 25, 26, 34, 47], but all of them are person-specific and + +suffer from the generalization ability to arbitrary individuals. + +# 3. Method + +Our goal is to synthesize the agent head during a dyadic conversation. To this end, we propose a two-stage interactive head generation framework, called INFP. + +In the first stage, the model learns to extract and compress multiple conversational behaviors in videos into a low-dimensional motion latent space, and use these latent codes to animate the portrait image of the agent, denoted as $I_{\text{self}}$ . Note that, to ensure the richness and diversity of the motion latent space, we train stage 1 with plenty of real-life dialogue videos. Once the training converges, we freeze this part and start to train the second stage. The model takes as input the audio of the agent, denoted as $A_{\text{self}}$ , and the audio of the conversation partner, denoted as $A_{\text{other}}$ , to learn the mapping from this dual-track dialogue audio to motion latent codes through denoising. We present a comprehensive introduction to these two stages below. + +# 3.1. Motion-Based Head Imitation + +Motion Encoding. For motion encoding, we employ a motion encoder $\mathbf{E}_m$ to learn an implicit latent representation, which captures verbal and non-verbal communicative behaviors from input facial images. $\mathbf{E}_m$ comprises a convolutional-based feature extractor followed by MLP layers, which compresses the motion feature into a latent code. Note that, we use implicit representation instead of explicit 3DMM coefficients [28] because the expressiveness of 3DMM model is limited, and the commonly used expression coefficients, predicted by SOTA 3D facial estimation methods [5, 7], entangle with face shape to a certain extent. + +A well-crafted motion latent space should demonstrate a high level of disentanglement to ensure that the extracted motions are appearance-independent. To achieve this, we apply two measures. (a) We set the motion latent code as a low-dimensional 1-D descriptor. It fosters effectively capturing the fundamental semantics of facial motion without entangling with appearance information, as explained in previous works [38, 43, 44]. (b) We design a hybrid facial representation to replace the original input image of $\mathbf{E}_m$ . In particular, we first mask the majority of facial pixels, preserving only the eyes and lip regions, as we regard them as the most intricate and expressive components of facial expressions. This can block the interference of irrelevant information such as hair and background for expression distill. To provide facial orientation and contour information, we then use an off-the-shelf facial estimation model [20] to obtain the facial vertices, and project the contour-related points onto the masked image. Our ablation study shows the superiority of our proposed hybrid facial representation + +![](images/7cc4c46f83c049322df0e0cfa2e462981e31bf2829f50178c427e8177dec017f.jpg) +Figure 3. Overview of INFP. The first stage (Motion-Based Head Imitation) learns to project facial communicative behaviors from real-life conversation videos into a low-dimensional motion latent space, and use the latent codes to animate a static image. The second stage (Audio-Guided Motion Generation) learns the mapping from the input dyadic audio to motion latent codes through denoising, to achieve the audio-driven interactive head generation. + +![](images/32f5ec0b399785cca4e30e3ce51d5c14fa59c4fefe57de140f4facab683dce5f.jpg) + +compared with the other representations like the intact image [6] or 2D landmarks map [22]. + +Head Imitation. In this paper, we adopt a motion flow-based generation pipeline. Concretely, we first extract source motion code $m_{self}$ from the input portrait image $I_{self}$ , and driving motion codes $m_{dri}^{1:N}$ from the driving video $V_{dri}$ . A motion flow estimation model $\mathbf{F}$ takes these motion codes as input to predict the self $\rightarrow$ driving flow. We utilize a face encoder $\mathbf{E}_{face}$ to extract 3D appearance feature volume from the input image $I_{self}$ . This feature volume is then warped by the motion flow and sent into a face decoder $\mathbf{D}_{face}$ to synthesize the final predictions $I_{pred}^{1:N}$ . The overall pipeline can be written as: + +$$ +\begin{array}{l}\text {F l o w} _ {s \rightarrow d} = \mathbf {F} (\mathbf {E} _ {m} (I _ {\text {s e l f}}), \mathbf {E} _ {m} (V _ {\text {d r i}}))\\\text {I} ^ {1: N} = \mathbf {D} _ {n} \left(W _ {n} - \left(I _ {n}\right) I _ {n} + \dots\right)\end{array}\tag {1} +$$ + +$$ +I _ {p r e d} ^ {1: N} = \mathbf {D} _ {f a c e} (W a r p (\mathbf {E} _ {f a c e} (I _ {s e l f}), F l o w _ {s \rightarrow d})) +$$ + +Model details and loss function are shown in the supplementary material. + +# 3.2. Audio-Guided Motion Generation + +Previous works [34, 39, 59] tend to divide their models into a Speaker Generator and a Listener Generator. Since the Speaker Generator only receives speaking audio, while the Listener Generator solely processes listening audio, it is necessary to segment the audio of both agent and its conversation partner into speaking clips and listening clips. Hence, in a multi-turn dialogue, explicit role switching is also crucial. We argue that it can lead to unnatural transitions in communicative states, particularly in real-life interaction contexts characterized by frequent role switching. Besides, this design paradigm clearly goes against the intention of the conversational agent, as it should be able to + +naturally present different communicative states based on the progress of the conversation. + +In this paper, we present a new paradigm, in which the model takes as input the audio from both the agent and its conversation partner, eliminating the necessity for explicit role assignment or switching. The key innovations lie in extracting interactive information from dual-track dyadic audio to construct speaking-listening mixed motions, and mapping these motions to the motion latent space of stage 1 to achieve audio-driven head generation. + +Motion Feature Construction. As shown in Fig. 3, an interactive motion guider is introduced to extract motion feature from the input audio. Within it, there are two memory banks $M_v$ and $M_{nv}$ , storing verbal and nonverbal motion, respectively. Each motion memory bank consists of learnable embeddings to memorize typical motions, denoted as $e^{1:K}$ , where $e^k \in \mathbb{R}^d$ represents the $k$ -th motion embedding and $d$ is the dimension. We set $K$ to 64 and $d$ to 512 in our experiments. When agent is speaking, $A_{self}$ contains rich information, so we take it as $Query$ and obtain verbal motion feature from $M_v$ , which is taken as $Key$ and $Value$ , through a cross-attention layer. Since the partner is listening, $A_{other}$ carries very little information, hence no effective non-verbal motion feature can be obtained from $M_{nv}$ . Therefore, after feature fusion, which involves element-wise sum and several MLP layers, verbal motion feature dominates and drives the agent to express the intended speaking state. However, when the partner starts speaking, $A_{self}$ becomes weak and $A_{other}$ fulfills a vital role. Accordingly, non-verbal motion feature from $M_{nv}$ dominates and forces the agent to present a listening state. In this way, our model can dynamically construct + +interactive motion feature $f_{m}$ based on the content of the dialogue audio, allowing the agent to convey the intended state. Note that, $A_{\text{self}}$ and $A_{\text{other}}$ are encoded to the feature space through an off-the-shelf HuBERT audio encoder [12], before sent into the model. + +Inspired by stylistic facial animation methods [21, 35], we also introduce a motion style vector $s_m$ to explicitly edit the motion embeddings of memory banks, through a style modulation layer raised by StyleGAN2[15]. $s_m$ is first extracted from motion latent codes through a motion style encoder $\mathbf{E}_{style}$ , and then squeezed along the temporal dimension to a global descriptor. We argue that $s_m$ contains some global information like emotions and attitudes, thus enhancing the authenticity and vividness of the generated results. Note that during training, $s_m$ is sourced from an arbitrary video clip of the driven individual, while in testing, $s_m$ can be extracted from any videos or set as null. + +Motion Mapping. To map the interactive motions $f_{m}$ to the pretrained motion latent space of stage 1, we employ a conditional diffusion transformer $\mathbf{T}$ [11]. Given the data distribution $q(m^{1:N},f_m)$ where $m^{1:N}$ represents the corresponding motion latent codes with $N$ frames, the diffusion model approximates the conditional distribution $q(m^{1:N}|f_m)$ . $\mathbf{T}$ only has 4 blocks, which makes our framework lightweight enough to achieve real-time interaction. Each block consists of a self-attention layer, a motion-attention layer, and a temporal-attention layer. $\mathbf{T}$ predicts the noise added to the ground-truth motion latent codes for each denoising step. Diffusion timestep is converted into the sinusoidal embedding and then concatenated to the noised motion latent codes at the time dimension. In the motion attention layer, we take the latent feature output from preceding self-attention layer as the Query, while the interactive motion feature $f_{m}$ is utilized as the Key and Value. Inspired byAnimateDiff [10], we implement a temporal layer to ensure a smooth transition between adjacent windows. In particular, we incorporate the last 10 frames of the generated motion latent codes from the previous window as the condition of the current one. Note that in the inference time, we add noise to the self motion latent codes $m_{\text{self}}^{1:N}$ , which are extracted from $I_{\text{self}}$ and replicated into $n$ copies. + +Training Strategy. In the training stage, we randomly set style vector to null with a probability of 0.3, and randomly drop interactive motion feature and previous motion latent codes with a probability of 0.5. Also note that we adopt a warm-up strategy for training stage 2. Specifically, we only select single-sided conversation clips from our training set as "simple case" to train the model for several epochs, which helps with the initialization of two memory banks. After that, we randomly sample multi-turn clips from entire dataset for the remaining training process. + +# 4. Dataset + +Existing dyadic interaction datasets have limitations in both scale and quality. ViCo [58] is a small-scale dataset with 1.6 hours and 96 IDs, and it lacks multi-turn conversation scenarios. Although ViCo-X [57] features rich multi-turn conversations, the total duration is only 0.4 hours. Besides, its background is monotonous, all of which are green screen. RealTalk [8] is a medium-scale dataset capturing dyadic conversations between pairs of individuals totaling to 115 hours of online public videos. However, it does not restrict the resolution of faces in the video, leading to low face resolution in many videos, thus impacting the quality of model generation. + +We propose DyConv, a large-scale dataset of multi-turn dyadic conversations collected from the Internet. The dataset contains video clips with a duration of over 200 hours, captures a wide gamut of emotions and expressions. We employ an off-the-shelf face detection model [51], and keep only frames where both individuals of the conversation are fully visible and their facial resolution is greater than $400 \times 400$ . In addition, we utilize a SOTA speech separation model [54] to separate the audio of two persons in the conversation, denoted as $A_{p1}$ and $A_{p2}$ . Then we run an active speaker detection model [1] to match each audio clip to the corresponding face in the original video. We will publicly release the dataset, along with separated audios and annotations. We hope it will encourage the investigation of head generation for dyadic interaction. + +# 5. Experiment + +# 5.1. Implementation Details + +In our experiments, we train the diffusion transformer using AdamW optimizer [19], with $\beta_{1} = 0.9$ and $\beta_{2} = 0.999$ , a learning rate of 1e-4, a weight decay of 1e-2, and a batch size of 32. For inference, we use a DDIM sampler [30] for 20 denoising steps. The CFG parameter of motion feature is set to 2, and the CFG parameter of previous motion latent is set to 1.5. + +# 5.2. Comparisons + +Although originally designed for dyadic interactive scenarios, our method can be naturally generalized to other applications, such as listening head generation and talking head generation. This allows us to fully validate the advancement of our proposed framework against some of the current SOTA methods in related areas. + +Evaluation Metrics. To quantitatively evaluate the visual quality, we utilize the Structured Similarity (SSIM), Peak Signal-to-Noise Ratio (PSNR) and Frechet Inception Distance (FID) for the generated videos. Additionally, LPIPS [50] is employed to measure the feature-level similarity between generated and ground-truth frames. To as + +![](images/b8d49fc67c745b6073605764386dd2a295b68fe2afbc67a65ec00e528d219b02.jpg) +Reference Image + +![](images/c571530f3918fed859d3aed95d251d8bbca07b5f8d2bcc876d05cf6b41820208.jpg) +Figure 4. Qualitative comparison of interactive head generation on DyConv. Given the input audio of dyadic conversations (the audio of the predicted agent is black and the audio of the conversation partner is blue), our framework generates reasonable and vivid facial and head pose movements as various roles in interactive scenarios. + +sess the preservation of facial appearance, we calculate the cosine similarity (CSIM) [4] between the facial features of the reference image and the generated video frames. In term of motion diversity of generated videos, we use SI for diversity (SID) and Variance (Var) following previous studies [23, 39]. We use the confidence score of SyncNet (SyncScore) [3] to evaluate the synchronization between lip movements and audio signals of the agent. + +# 5.2.1. Interactive Head Generation + +Since only a few works focus on the interactive head generation [42, 59] and these methods are not open-sourced, direct comparisons with them are considerably difficult. To effectively validate our method, we select the SOTA dyadic + +modeling work DIM [39] as our baseline and utilize the officially released code to re-train the model on our datasets. Specifically, we first use the whole dataset to pretrain the VQ-VAE for the listener and speaker and then split the dataset into two subsets to train the listener generator and the speaker generator, respectively. To better visualize results of DIM, we also train the face renderer [28] to generate photo-realistic videos. + +Quantitative Results. The quantitative experiments are carried out on the test set of DyConv. Please refer to Tab. 1 for the results. It can be seen that our method outperforms the baseline cross a variety of metrics. In particular, our face decoder achieves a better visual quality such as SSIM, PSNR and FID. Thanks to the disentanglement and expres + +siveness of our latent space, our method is superior to DIM in term of audio-lip-synchronization by a large margin as there is a much closer distance of our SyncScore to the ground truth. As for the LPIPS and CISM, our method still achieves solid performances, reflecting the better identity preservation. Of particular note, our method is far ahead of the baseline in term of SID and Var, which proves the generalization and effectiveness of our method to generate a great range of diverse motions. It benefits from the design of interactive motion guider as well as conditional denoising transformer. + +Qualitative Results. Subjective evaluation is crucial in identifying the performance of generative models, particularly on videos. We strongly recommend readers to watch our supplementary video. Two examples and their comparisons with the baseline are shown in Fig. 4. We use two typical audio clips in the dyadic conversation scenario to drive reference identities randomly selected from the test set. It can be obviously observed that our method can generate more diverse head pose movements in comparison to DIM. Meanwhile, results of our method demonstrate more vivid facial expressions, which significantly improve the naturalness of generated videos. It is worth to noting that facial behaviors synthesized by our method are well aligned with the dialogue audio content. This suggests the effectiveness of our method in dyadic interactions, which is capable of capturing distinct characteristics of various states, such as listening, speaking as well as the mutual interaction, and naturally adapting its role among them based on the progress of the conversation. In contrast, DIM is unable to generate appropriate behaviors such as reasonable lip movement and response feedback. + +User Study. We also invite 20 participants to conduct a user study for further subjective evaluation. Specifically, we apply INFP and DIM to synthesize results of 20 dyadic conversation videos randomly sampled from the test set. By adopting the commonly used Mean Opinion Scores (MOS) rating protocol, we request all the participants to provide their ratings (from 1 to 5, the higher the better) on four aspects for each generated video: naturalness, motion diversity, audio-visual alignment and visual quality. As detailed in Tab. 2, our method demonstrates superior performances across all evaluated aspects compared to DIM. + +# 5.2.2. Listening Head Generation + +Since our method can generalize to the task of listening head generation, we directly use our model without any modification to conduct additional experiments. Specifically, we compare our framework with SOTA listening head generation methods, including L2L [23], RLHG [58] and DIM [39]. ViCo [58] is selected as the benchmark. For quantitative evaluation, we utilized the same test set and inherited both quantitative and qualitative results presented in DIM. We conducted the comparison on FD, MSE, SID + +and Var and each metric is calculated on expression and head pose, respectively. Different to other methods, our method generate photo-realistic video rather than 3DMM coefficients. To this end, we use the off-the-shelf 3D face reconstruction method [7] to extract expression and head pose coefficients from our generated videos. As shown in Tab. 3, INFP outperforms other SOTA methods in the majority of metrics. While SID and Var are performing well, our method remains highly effective in both FD and MSE. This indicates the power of our novel framework design and dataset quality, making generated facial expressions and head pose movements closer to real human behaviors. The visualization results in the supplementary materials also demonstrate that INFP achieves more expressive and diverse motions. + +# 5.2.3. Talking Head Generation + +For talking head generation, we select SadTalker[52], AntiTalker[17], and EchoMimic[2] as comparing methods. As all of them reported evaluation results on HDTF [53], we randomly sampled 20 videos from it as the test set to conduct our experiments for a fair comparison. Results in Tab. 4 show that our model is superior in audiolip-synchronization and identity preservation. The reason behind this might be that we carefully build a disentangled and expressive motion latent space and it makes our model easier to learn the correlation between audio and lip movements without the interference of appearance features. It can also be observed from the visualization results in the supplementary materials. Our SSIM and PSNR score is slightly behind EchoMimic because EchoMimic is finetuned based on the Stable Diffusion V1.5 [29]. Despite this, our method can achieve better scores in terms of FID, meaning a good visual quality of our results. + +# 5.3. Ablation Study + +We ablate 3 key components of our framework, and the results are listed in Tab. 1. + +Motion Memory Bank. We remove the verbal and non-verbal motion memory banks from interactive motion guider in stage 2, and directly use the dyadic audio feature after several MLP layers as the condition for the motion-attention layer. Results show that the generated lip-related motion behaviors are clearly inconsistent with the ground-truth. In additional, there is a significant drop of both SID and Var scores. These suggest that memory banks can effectively store rich motion patterns in different conversational states. The experiment on memory bank size is shown in the supplementary material. + +Style Modulation. We remove the motion style vector $s_m$ and corresponding modulation layers. As we expected, the authenticity and vividness of the results decrease, which can also be reflected in the user study. + +
MethodsSSIM↑PSNR↑FID↓SyncScore↑LPIPS↓CSIM↑SID↑Var↑
DIM [39]0.65120.41734.3614.7780.4850.8240.7660.825
INFP (Ours)0.83431.56215.7277.1880.2570.9042.6132.386
w/o Motion Memory0.83031.21818.3346.1030.2590.8992.1532.016
w/o Style Modulation0.83131.44216.0297.0620.2710.9042.5512.316
w/ Intact Image0.80228.48816.9906.8120.2660.8422.4702.148
w/ Landmarks Map0.82130.69316.3276.8330.2810.9012.6012.335
GT1.000N/A0.0007.2610.0000.9672.8912.435
+ +Table 1. Quantitative comparison of interactive head generation on DyConv. + +
MethodsNaturalnessMotion DiversityAV-AlignVisual
DIM [39]2.712.142.653.13
INFP w/o Style4.023.984.284.11
INFP (Ours)4.384.494.334.13
+ +Table 2. User study measured by Mean Opinion Scores. + +
MethodFD↓MSE↓SID↑Var↑
ExpPoseExpPoseExpPoseExpPose
L2L [23]33.930.060.930.012.772.660.830.02
RLHG [58]39.020.070.860.013.623.171.520.02
DIM [39]23.880.060.700.013.712.351.530.02
INFP (Ours)18.630.070.510.014.783.922.830.18
+ +Table 3. Quantitative comparison of listening head generation on ViCo [58]. + +
MethodsSSIM↑PSNR↑FID↓SyncScore↑CSIM↑
SadTalker [52]0.78126.65028.5166.1250.814
AniTalker [17]0.78926.82228.7785.7380.858
EchoMimic [2]0.80728.98320.8736.2510.863
INFP (Ours)0.80428.65718.5826.6700.886
+ +Table 4. Quantitative comparison of talking head generation on HDTF [53]. + +Hybrid Facial Representation. We change the input to $\mathbf{E}_m$ in stage 1 from our carefully-designed hybrid facial representation to the original intact image or 2D landmarks map. We observe that there is degradation in generation quality and a leakage of appearance information. + +# 5.4. Visualization of Style Control + +As illustrated in Sec. 3.2, our model supports inputting style vector $s_m$ extracted from any portrait video to globally control the emotions or attitudes of the generated results. Here we select 3 video clips with varying degrees of "happiness" to control emotions during generation. Results are shown in Fig. 5. It can be seen from the figure that $s_m$ effectively control the emotions of the driven individual. + +![](images/dc25b44437a90aa874b190cb2f512ca773ff5ad605d728b42492dcf5a0378246.jpg) +Figure 5. Visualization of style control with our style modulation. + +# 6. Conclusion + +In this paper, we propose INFP, a novel audio-driven interactive head generation framework for dyadic conversations, which aims to mimic facial behaviors of human-human interactions conditioning on the dyadic audio. As demonstrated, our method is a unified framework and is capable of smoothly and naturally adapting to various conversational roles without explicit assignment or switching. Meanwhile, we collect a large scale audio-visual dataset for multi-turn dyadic conversations, which could facilitate further researches of dyadic interactive head generation. Extensive experiments and ablations prove the advancement and effectiveness of our method over other SOTA methods. + +Limitation and Future Work. Currently, our method only takes conversation audios as the input, combining control signals from multiple modalities, such as visual and textual contents, can offer additional capacities. As our proposed framework focuses on head synthesis, extending the generation range to the upper, even whole body might be a interesting and challenging future work. + +Ethical Considerations. There might be the potential misuse of our method to fabricate deceptive talks and speeches. To this end, we will restrict access to our core models exclusively to research institutions. + +# References + +[1] Stanislav Beliaev, Yuriy Rebryk, and Boris Ginsburg. Talknet: Fully-convolutional non-autoregressive speech synthesis model, 2020. 5 +[2] Zhiyuan Chen, Jiajiong Cao, Zhiquan Chen, Yuming Li, and Chenguang Ma. Echomimic: Lifelike audio-driven portrait animations through editable landmark conditions. arXiv preprint arXiv:2407.08136, 2024. 1, 3, 7, 8 +[3] J. S. Chung and A. Zisserman. Out of time: automated lip sync in the wild. In Workshop on Multi-view Lip-reading, ACCV, 2016. 6 +[4] Jiankang Deng, Jia Guo, Xue Niannan, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In CVPR, 2019. 6 +[5] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 3 +[6] Nikita Drobyshev, Jenya Chelishev, Taras Khakhulin, Aleksei Ivakhnenko, Victor Lempitsky, and Egor Zakharov. Megaportraits: One-shot megapixel neural head avatars. 2022. 4 +[7] Panagiotis P. Filntisis, George Retsinas, Foivos Paraperas-Papantoniou, Athanasios Katsamanis, Anastasios Roussos, and Petros Maragos. Visual speech-aware perceptual 3d facial expression reconstruction from videos. arXiv preprint arXiv:2207.11094, 2022. 3, 7 +[8] Scott Geng, Revant Teotia, Purva Tendulkar, Sachit Menon, and Carl Vondrick. Affective faces for goal-driven dyadic communication. 2023. 2, 3, 5 +[9] Jiazhi Guan, Zhanwang Zhang, Hang Zhou, Tianshu Hu, Kaisiyuan Wang, Dongliang He, Haocheng Feng, Jingtuo Liu, Errui Ding, Ziwei Liu, et al. Stylesync: High-fidelity generalized and personalized lip sync in style-based generator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1505-1515, 2023. 1 +[10] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Imagine your personalized text-to-image diffusion models without specific tuning. International Conference on Learning Representations, 2024. 5 +[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 5 +[12] Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. Hubert: Self-supervised speech representation learning by masked prediction of hidden units. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 29:3451-3460, 2021. 5 +[13] Ailin Huang, Zhewei Huang, and Shuchang Zhou. Perceptual conversational head generation with regularized driver and enhanced renderer. In Proceedings of the 30th ACM + +International Conference on Multimedia, page 7050-7054. ACM, 2022. 2 +[14] Yinghao Huang, Leo Ho, Dafei Qin, Mingyi Shi, and Taku Komura. Interact: Capture and modelling of realistic, expressive and interactive activities between two persons in daily scenarios. 2024. 3 +[15] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In Proc. CVPR, 2020. 5 +[16] Jin Liu, Xi Wang, Xiaomeng Fu, Yesheng Chai, Cai Yu, Jiao Dai, and Jizhong Han. Mfr-net: Multi-faceted responsive listening head generation via denoising diffusion model. In Proceedings of the 31st ACM International Conference on Multimedia, pages 6734-6743, 2023. 2, 3 +[17] Tao Liu, Feilong Chen, Shuai Fan, Chenpeng Du, Qi Chen, Xie Chen, and Kai Yu. Anitalker: Animate vivid and diverse talking faces through identity-decoupled facial motion encoding, 2024. 1, 7, 8 +[18] Xi Liu, Ying Guo, Cheng Zhen, Tong Li, Yingying Ao, and Pengfei Yan. Custom listener: Text-guided responsive interaction for user-friendly listening head generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2415-2424, 2024. 1, 2, 3 +[19] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization, 2019. 5 +[20] Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuo-Ling Chang, Ming Yong, Juhyun Lee, Wan-Teh Chang, Wei Hua, Manfred Georg, and Matthias Grundmann. Mediapipe: A framework for perceiving and processing reality. In CVPR 2019, 2019. 3 +[21] Yifeng Ma, Suzhen Wang, Zhipeng Hu, Changjie Fan, Tangjie Lv, Yu Ding, Zhidong Deng, and Xin Yu. Styletalk: one-shot talking head generation with controllable speaking styles. In Proceedings of the Thirty-Seventh AAAI Conference on Artificial Intelligence and Thirty-Fifth Conference on Innovative Applications of Artificial Intelligence and Thirteenth Symposium on Educational Advances in Artificial Intelligence. AAAI Press, 2023. 5 +[22] Yue Ma, Hongyu Liu, Hongfa Wang, Heng Pan, Yingqing He, Junkun Yuan, Ailing Zeng, Chengfei Cai, Heung-Yeung Shum, Wei Liu, et al. Follow-your-emoji: Fine-controllable and expressive freestyle portrait animation. arXiv preprint arXiv:2406.01900, 2024. 4 +[23] Evonne Ng, Hanbyul Joo, Liwen Hu, Hao Li, Trevor Darrell, Angjoo Kanazawa, and Shiry Ginosar. Learning to listen: Modeling non-deterministic dyadic facial motion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20395-20405, 2022. 1, 2, 6, 7, 8 +[24] Evonne Ng, Sanjay Subramanian, Dan Klein, Angjoo Kanazawa, Trevor Darrell, and Shiry Ginosar. Can language models learn to listen? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10083-10093, 2023. 1, 2, 3 +[25] Evonne Ng, Javier Romero, Timur Bagautdinov, Shaojie Bai, Trevor Darrell, Angjoo Kanazawa, and Alexander Richard. + +From audio to photoreal embodiment: Synthesizing humans in conversations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1001-1010, 2024. 3 +[26] Se Jin Park, Chae Won Kim, Hyeongseop Rha, Minsu Kim, Joanna Hong, Jeong Hun Yeo, and Yong Man Ro. Let's go real talk: Spoken dialogue model for face-to-face conversation. arXiv preprint arXiv:2406.07867, 2024. 3 +[27] K R Prajwal, Rudrabha Mukhopadhyay, Vinay P. Namboodiri, and C.V. Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM International Conference on Multimedia, page 484-492, New York, NY, USA, 2020. Association for Computing Machinery. 1, 3 +[28] Yurui Ren, Ge Li, Yuanqi Chen, Thomas H. Li, and Shan Liu. Pirenderer: Controllable portrait image generation via semantic neural rendering, 2021. 3, 6 +[29] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 7 +[30] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv:2010.02502, 2020. 5 +[31] Luchuan Song, Guojun Yin, Zhenchao Jin, Xiaoyi Dong, and Chenliang Xu. Emotional listener portrait: Realistic listener motion simulation in conversation. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 20782-20792. IEEE, 2023. 1, 2, 3 +[32] Siyang Song, Micol Spitale, Cheng Luo, Cristina Palmero, German Barquero, Hengde Zhu, Sergio Escalera, Michel Valstar, Tobias Baur, Fabien Ringeval, Elisabeth Andre, and Hatice Gunes. React 2024: the second multiple appropriate facial reaction generation challenge, 2024. 2 +[33] Michal Stypulkowski, Konstantinos Vougioukas, Sen He, Maciej Zieba, Stavros Petridis, and Maja Pantic. Diffused heads: Diffusion models beat gans on talking-face generation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5091-5100, 2024. 3 +[34] Mingze Sun, Chao Xu, Xinyu Jiang, Yang Liu, Baigui Sun, and Ruqi Huang. Beyond talking—generating holistic 3d human dyadic motion for communication. arXiv preprint arXiv:2403.19467, 2024. 3, 4 +[35] Zhiyao Sun, Tian Lv, Sheng Ye, Matthieu Lin, Jenny Sheng, Yu-Hui Wen, Minjing Yu, and Yong-jin Liu. Diffpositalk: Speech-driven stylistic 3d facial animation and head pose generation via diffusion models. ACM Transactions on Graphics (TOG), 43(4):1-9, 2024. 5 +[36] Jiaxiang Tang, Kaisiyuan Wang, Hang Zhou, Xiaokang Chen, Dongliang He, Tianshu Hu, Jingtuo Liu, Gang Zeng, and Jingdong Wang. Real-time neural radiance talking portrait synthesis via audio-spatial decomposition. arXiv preprint arXiv:2211.12368, 2022. 1 +[37] Linrui Tian, Qi Wang, Bang Zhang, and Liefeng Bo. Emo: Emote portrait alive-generating expressive portrait videos with audio2video diffusion model under weak conditions. arXiv preprint arXiv:2402.17485, 2024. 1, 3 +[38] Naftali Tishby, Fernando C Pereira, and William Bialek. The information bottleneck method. arXiv preprint physics/0004057, 2000. 3 + +[39] Minh Tran, Di Chang, Maksim Siniukov, and Mohammad Soleymani. Dyadic interaction modeling for social behavior generation. arXiv preprint arXiv:2403.09069, 2024. 1, 2, 3, 4, 6, 7, 8 +[40] Aaron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu. Neural discrete representation learning, 2018. 2 +[41] Cong Wang, Kuan Tian, Jun Zhang, Yonghang Guan, Feng Luo, Fei Shen, Zhiwei Jiang, Qing Gu, Xiao Han, and Wei Yang. V-express: Conditional dropout for progressive training of portrait video generation. arXiv preprint arXiv:2406.02511, 2024. 1, 3 +[42] Duomin Wang, Bin Dai, Yu Deng, and Baoyuan Wang. Disentangling planning, driving and rendering for photorealistic avatar agents. 2023. 1, 2, 3, 6 +[43] Duomin Wang, Yu Deng, Zixin Yin, Heung-Yeung Shum, and Baoyuan Wang. Progressive disentangled representation learning for fine-grained controllable talking head synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17979-17989, 2023. 3 +[44] Yaohui Wang, Di Yang, Francois Bremond, and Antitza Dantcheva. Latent image animator: Learning to animate images via latent space navigation. arXiv preprint arXiv:2203.09043, 2022. 3 +[45] Mingwang Xu, Hui Li, Qingkun Su, Hanlin Shang, Liwei Zhang, Ce Liu, Jingdong Wang, Luc Van Gool, Yao Yao, and Siyu Zhu. Hallo: Hierarchical audio-driven visual synthesis for portrait image animation. arXiv preprint arXiv:2406.08801, 2024. 1 +[46] Sicheng Xu, Guojun Chen, Yu-Xiao Guo, Jiaolong Yang, Chong Li, Zhenyu Zang, Yizhong Zhang, Xin Tong, and Baining Guo. Vasa-1: Lifelike audio-driven talking faces generated in real time. arXiv preprint arXiv:2404.10667, 2024. 1, 3 +[47] Yichao Yan, Zanwei Zhou, Zi Wang, Jingnan Gao, and Xiaokang Yang. Dialoguenerf: Towards realistic avatar face-to-face conversation video generation, 2023. 3 +[48] Chenxu Zhang, Chao Wang, Jianfeng Zhang, Hongyi Xu, Guoxian Song, You Xie, Linjie Luo, Yapeng Tian, Xiaohu Guo, and Jiashi Feng. Dream-talk: diffusion-based realistic emotional audio-driven method for single image talking face generation. arXiv preprint arXiv:2312.13578, 2023. 1 +[49] Longhao Zhang, Shuang Liang, Zhipeng Ge, and Tianshu Hu. Personatack: Bring attention to your persona in visual dubbing. arXiv preprint arXiv:2409.05379, 2024. 1 +[50] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5 +[51] Shifeng Zhang, Xiangyu Zhu, Zhen Lei, Hailin Shi, Xiaobo Wang, and Stan Z. Li. S $^3$ fd: Single shot scale-invariant face detector, 2017. 5 +[52] Wenxuan Zhang, Xiaodong Cun, Xuan Wang, Yong Zhang, Xi Shen, Yu Guo, Ying Shan, and Fei Wang. Sadtalker: Learning realistic 3d motion coefficients for stylized audio-driven single image talking face animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8652-8661, 2023. 1, 7, 8 + +[53] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661–3670, 2021. 7, 8 +[54] Shengkui Zhao, Yukun Ma, Chongjia Ni, Chong Zhang, Hao Wang, Trung Hieu Nguyen, Kun Zhou, Jiaqi Yip, Dianwen Ng, and Bin Ma. Mossformer2: Combining transformer and rnnprecent network for enhanced time-domain monaural speech separation, 2023. 5 +[55] Wei Zhao, Peng Xiao, Rongju Zhang, Yijun Wang, and Jianxin Lin. Semantic-aware responsive listener head synthesis. In Proceedings of the 30th ACM International Conference on Multimedia, pages 7065-7069, 2022. 2 +[56] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking face generation by implicitly modularized audio-visual representation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 3 +[57] Mohan Zhou, Yalong Bai, Wei Zhang, Ting Yao, Tiejun Zhao, and Tao Mei. Vico-x: Multimodal conversation dataset. https://project.mhzhou.com/vico, 2022. Accessed: 2022-09-30. 2, 5 +[58] Mohan Zhou, Yalong Bai, Wei Zhang, Ting Yao, Tiejun Zhao, and Tao Mei. Responsive listening head generation: A benchmark dataset and baseline. In Proceedings of the European conference on computer vision (ECCV), 2022. 2, 3, 5, 7, 8 +[59] Mohan Zhou, Yalong Bai, Wei Zhang, Ting Yao, and Tiejun Zhao. Interactive conversational head generation. arXiv preprint arXiv:2307.02090, 2023. 1, 2, 3, 4, 6 +[60] Yang Zhou, Xintong Han, Eli Shechtman, Jose Echevarria, Evangelos Kalogerakis, and Dingzeyu Li. Makelttalk: speaker-aware talking-head animation. ACM Transactions on Graphics, 39(6):1-15, 2020. 2, 3 \ No newline at end of file diff --git a/2025/INFP_ Audio-Driven Interactive Head Generation in Dyadic Conversations/images.zip b/2025/INFP_ Audio-Driven Interactive Head Generation in Dyadic Conversations/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b62c48d159effde2c2f611f9474ba47bab5ec081 --- /dev/null +++ b/2025/INFP_ Audio-Driven Interactive Head Generation in Dyadic Conversations/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6baa0c846bad15ec217d6af02d3a60c0bc4ed254877c8f5a26f3cf4979b336e9 +size 606944 diff --git a/2025/INFP_ Audio-Driven Interactive Head Generation in Dyadic Conversations/layout.json b/2025/INFP_ Audio-Driven Interactive Head Generation in Dyadic Conversations/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c93ff3a2b7f3ace9d199395df79f060b0f3e5aa7 --- /dev/null +++ b/2025/INFP_ Audio-Driven Interactive Head Generation in Dyadic Conversations/layout.json @@ -0,0 +1,7535 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 75, + 103, + 536, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 103, + 536, + 121 + ], + "spans": [ + { + "bbox": [ + 75, + 103, + 536, + 121 + ], + "type": "text", + "content": "INFP: Audio-Driven Interactive Head Generation in Dyadic Conversations" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 143, + 548, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 143, + 548, + 171 + ], + "spans": [ + { + "bbox": [ + 62, + 143, + 548, + 171 + ], + "type": "text", + "content": "Yongming Zhu* Longhao Zhang* Zhengkun Rong* Tianshu Hu*† Shuang Liang Zhipeng Ge Bytedance" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 150, + 174, + 457, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 174, + 457, + 199 + ], + "spans": [ + { + "bbox": [ + 150, + 174, + 457, + 199 + ], + "type": "text", + "content": "{zhuyongming, zhanglonghao.zlh, rongzhengkun, tianshu.hu, liangshuang. echo, zhipengge}@bytedance.com" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 61, + 217, + 553, + 403 + ], + "blocks": [ + { + "bbox": [ + 61, + 217, + 553, + 403 + ], + "lines": [ + { + "bbox": [ + 61, + 217, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 61, + 217, + 553, + 403 + ], + "type": "image", + "image_path": "4e20afd929fea7b3889fc607e28cc363064719c32fbc24bc1a901cd98e39189c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 410, + 555, + 467 + ], + "lines": [ + { + "bbox": [ + 54, + 410, + 555, + 467 + ], + "spans": [ + { + "bbox": [ + 54, + 410, + 555, + 467 + ], + "type": "text", + "content": "Figure 1. We present INFP, an audio-driven interactive head generation framework for dyadic conversations. Given the dual-track audio in dyadic conversations and a single portrait image of arbitrary agent, our framework can dynamically synthesize verbal, non-verbal and interactive agent videos with lifelike facial expressions and rhythmic head pose movements. Additionally, our framework is lightweight yet powerful, making it practical in instant communication scenarios with acceptable latency, such as the video conferencing. INFP denotes our method is Interactive, Natural, Flash and Person-generic." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 475, + 200, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 475, + 200, + 488 + ], + "spans": [ + { + "bbox": [ + 151, + 475, + 200, + 488 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 502, + 297, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 502, + 297, + 682 + ], + "spans": [ + { + "bbox": [ + 54, + 502, + 297, + 682 + ], + "type": "text", + "content": "Imagine having a conversation with a socially intelligent agent. It can attentively listen to your words and offer visual and linguistic feedback promptly. This seamless interaction allows for multiple rounds of conversation to flow smoothly and naturally. In pursuit of actualizing it, we propose INFP, a novel audio-driven head generation framework for dyadic interaction. Unlike previous head generation works that only focus on single-sided communication, or require manual role assignment and explicit role switching, our model drives the agent portrait dynamically alternates between speaking and listening state, guided by the input dyadic audio. Specifically, INFP comprises a Motion-Based Head Imitation stage and an Audio-Guided Motion Generation stage. The first stage learns to project facial communicative behaviors from real-life conversation videos into a low-" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 476, + 556, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 476, + 556, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 476, + 556, + 586 + ], + "type": "text", + "content": "dimensional motion latent space, and use the motion latent codes to animate a static image. The second stage learns the mapping from the input dyadic audio to motion latent codes through denoising, leading to the audio-driven head generation in interactive scenarios. To facilitate this line of research, we introduce DyConv, a large scale dataset of rich dyadic conversations collected from the Internet. Extensive experiments and visualizations demonstrate superior performance and effectiveness of our method." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 609, + 395, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 609, + 395, + 621 + ], + "spans": [ + { + "bbox": [ + 314, + 609, + 395, + 621 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": "In recent years, researchers have dedicated considerable attention to audio-driven head generation [9, 18, 23, 24, 31, 37, 39, 42, 46, 56, 59], in order to build the conversational agent. However, most studies only focus on single-sided communication, such as talking or listening, ignoring the dyadic properties in human-human interaction. Talking-head generation [2, 17, 27, 36, 37, 41, 45, 46, 48, 49, 52, 56," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 67, + 703, + 148, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 148, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 148, + 712 + ], + "type": "text", + "content": "† Corresponding authors." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10667" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 312 + ], + "type": "text", + "content": "60] aims at synthesizing speaker's facial animations based on their reference images and a driving audio. Although these works can produce vibrant videos with accurate lip synchronization, they solely emphasize the speaker's role and overlook the listener's feedback. Listening-head generation [16, 18, 23, 24, 31, 32, 55, 58], on the other hand, intends to respond to speaker's behaviors. However, these works restrict the listener's response solely to non-verbal facial movements, which is clearly far from real-life interaction scenarios. Some recent studies [39, 42, 59] have started exploring head generation for dyadic interaction, but they require manual role assignment between Listener and Speaker and cannot achieve smooth and natural role switching. We believe that for a conversational agent, it should not be predetermined a role, but rather be able to freely switch between listening and speaking states based on the current dialogue context. When having a conversation with the agent, you may interrupt or respond to it at any moment. At that point, it naturally transitions from a speaker to a listener, and vice versa." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 315, + 296, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 315, + 296, + 674 + ], + "spans": [ + { + "bbox": [ + 56, + 315, + 296, + 674 + ], + "type": "text", + "content": "In this paper, we propose a novel audio-driven head generation framework for dyadic interaction, namely INFP. Unlike previous works, we do not divide the model into Speaker Generator and Listener Generator, or apply explicit role switching. The driven individual dynamically alternates between the speaking and listening state, guided by the dyadic audio. Specifically, INFP comprises a Motion-Based Head Imitation stage and an Audio-Guided Motion Generation stage. In the first stage, our model learns to extract communicative behavior, including non-verbal listening cues and verbal speaking patterns, from massive real-life conversational videos and encode them to a motion latent space. The motion latent codes are then used to animate a static portrait image to authentically imitate the character in the video. A well-crafted motion latent space should exhibit a high degree of disentanglement, which means that head pose, facial expression, and emotion should be decoupled from appearance. To this end, we perform facial structure discretization and facial pixel masking to the input. While in the second stage, the model is trained to map the dyadic audio to the motion latent space pretrained in stage 1 to obtain the corresponding motion latent codes. It consists of an interactive motion guider and a conditional diffusion transformer. The former takes as input the audio from both agent and its conversation partner, and retrieves verbal and non-verbal motions from learnable memory banks to construct the interactive motion feature. The latter utilizes the interactive motion feature as the condition and, together with other signals, generates the motion latent codes through denoising." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "To support our research on interactive head generation, we introduce DyConv, a large scale dataset of rich dyadic conversations collected from the Internet. It contains non" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 316, + 72, + 434, + 167 + ], + "blocks": [ + { + "bbox": [ + 316, + 72, + 434, + 167 + ], + "lines": [ + { + "bbox": [ + 316, + 72, + 434, + 167 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 434, + 167 + ], + "type": "image", + "image_path": "b22d306760e5e765f73694d335d631ca5064a91d6ceaa1fe679c4df1e33bc84d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 189, + 555, + 236 + ], + "lines": [ + { + "bbox": [ + 313, + 189, + 555, + 236 + ], + "spans": [ + { + "bbox": [ + 313, + 189, + 555, + 236 + ], + "type": "text", + "content": "Figure 2. Objective illustration. Existing interactive head generation (left) applied manual role assignment and explicit role switching. Our proposed INFP (right) is a unified framework which can dynamically and naturally adapt to various conversational states." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 443, + 99, + 556, + 144 + ], + "blocks": [ + { + "bbox": [ + 443, + 99, + 556, + 144 + ], + "lines": [ + { + "bbox": [ + 443, + 99, + 556, + 144 + ], + "spans": [ + { + "bbox": [ + 443, + 99, + 556, + 144 + ], + "type": "image", + "image_path": "2fa039f3238f285f4a7c9432007d5d12fb09c18f1ddbf2ecd825e0d3b7b172c0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 255, + 555, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 255, + 555, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 555, + 338 + ], + "type": "text", + "content": "scripted video conversations between pairs of individuals from diverse backgrounds. The videos capture their discussions on a wide range of topics and emotions, portraying authentic communication in real-life settings. Compared to previous datasets like ViCo [58], ViCo-X [57] and RealTalk [8], DyConv has a larger scale, higher quality, and richer interactions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 325, + 339, + 534, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 339, + 534, + 350 + ], + "spans": [ + { + "bbox": [ + 325, + 339, + 534, + 350 + ], + "type": "text", + "content": "Our main contributions are summarized as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 354, + 555, + 556 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 314, + 354, + 555, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 354, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 314, + 354, + 555, + 426 + ], + "type": "text", + "content": "- We propose INFP, an audio-driven interactive head generation framework with a new paradigm. To our best knowledge, we are the first to enable the driven individual to naturally transition between listening and speaking states during multi-turn conversations, without the requirement for manual role assignment or explicit role switching." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 426, + 555, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 426, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 314, + 426, + 555, + 472 + ], + "type": "text", + "content": "- We design a novel interactive motion guider with learnable memory banks to adaptively extract interactive information from dual-track dyadic audio and construct speaking-listening mixed motions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 473, + 555, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 473, + 555, + 509 + ], + "spans": [ + { + "bbox": [ + 314, + 473, + 555, + 509 + ], + "type": "text", + "content": "- We introduce a large-scale and high-quality dataset called DyConv. We hope it will encourage the investigation of head generation for dyadic interaction." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 509, + 554, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 509, + 554, + 556 + ], + "spans": [ + { + "bbox": [ + 314, + 509, + 554, + 556 + ], + "type": "text", + "content": "- Our framework achieves real-time and person-generic head generation. Extensive experiments and visualizations demonstrate superior performance and effectiveness of our method." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 568, + 402, + 581 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 568, + 402, + 581 + ], + "spans": [ + { + "bbox": [ + 314, + 568, + 402, + 581 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 588, + 545, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 588, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 545, + 601 + ], + "type": "text", + "content": "2.1. Single-Sided Audio-Driven Head Generation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": "Listening Head Generation. As a pioneer, RLHG [58] contributes the ViCo dataset as a common benchmark and builds a baseline method with a sequential decoder. Following that, PCH [13] introduces the enhanced renderer to improve the visual quality of generated videos. Song et al. [32] organizes the REACT challenge to encourage the community to investigate the face reaction in offline and online setting, separately. L2L [23] leverages the VQ-VAE [40] to store head motion sequences in discrete codebooks," + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10668" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 203 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 203 + ], + "type": "text", + "content": "leading to the better motion diversity. ELP [31] rearranges the latent space based on emotional priors to emphasize the emotion of generated motions. To get rid of simple emotional labels, CustomListener [18] proposes to control listener head motions via the text prior. MFR-Net [16] makes a progress of identity preservation and motion diversity. RealTalk [8] proposes a medium-scale dataset and attempts to retrieve possible videos of listener head with large language model. Ng et al. [24] presents a transfer-based approach from pretrained large language models to predict the motion sequences." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 205, + 298, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 205, + 298, + 420 + ], + "spans": [ + { + "bbox": [ + 56, + 205, + 298, + 420 + ], + "type": "text", + "content": "Talking Head Generation Audio-driven talking head generation has become a prominent research area in recent years. Early GAN-based approaches, like Wav2Lip [27], PC-AVS [56], and MakeItTalk [60], directly integrate the input audio and video to produce lip-synced visuals. VASA-1 [46] takes a significant leap forward in the field by decoupling latent feature spaces, enabling the generation of more realistic video outputs. This model can operate on a single image rather than requiring a full video sequence. With the emergence of diffusion models, several works [33, 37, 41] are proposed to enhance the generation quality. EchoMimic [2] is concurrently trained using both audios and facial landmarks, and is capable of generating portrait videos not only by audios and facial landmarks individually, but also by a combination of both audios and selected facial landmarks. AniTalker [56], employing two self-supervised learning strategies, excels in capturing intricate facial movements." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 431, + 276, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 431, + 276, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 431, + 276, + 443 + ], + "type": "text", + "content": "2.2. Head Generation in Dyadic Conversations" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 450, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 295, + 713 + ], + "type": "text", + "content": "Recently, the research community has paid increasing attention on the audio-driven head generation in the dyadic interaction. Zhou et al. [59] extend RHLG [58] by introducing ViCo-X, a multi-turn conversational dataset, and a Role Switcher to connect the Listener Generator and the Speaker Generator. However, explicit role switching can lead to unnaturalness and inconsistency, and this paradigm fails to cover all dyadic conversation states, such as simultaneous speaking. DIM [39] uses a pretraining approach to model both speakers' and listeners' motions, capturing dyadic context. However, the pretrained model requires additional fine-tuning for tasks like talking head and listening head generation, necessitating manual role assignment in dyadic conversations, which can cause inappropriate transitions. AgentAvatar [42] designs a streamlined pipeline to predict behaviors of a photo-realistic avatar agent in dyadic interactions. While the motion synthesis is conditioned on the high-level textual description, AgentAvatar fails to generate facial behaviors that precisely align with the dialogue audio, as demo videos shown in their project page. Additionally, there are some other studies in dyadic interactions [14, 25, 26, 34, 47], but all of them are person-specific and" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "content": "suffer from the generalization ability to arbitrary individuals." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 110, + 370, + 123 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 110, + 370, + 123 + ], + "spans": [ + { + "bbox": [ + 313, + 110, + 370, + 123 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 132, + 553, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 132, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 132, + 553, + 167 + ], + "type": "text", + "content": "Our goal is to synthesize the agent head during a dyadic conversation. To this end, we propose a two-stage interactive head generation framework, called INFP." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 169, + 555, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 169, + 555, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 169, + 555, + 324 + ], + "type": "text", + "content": "In the first stage, the model learns to extract and compress multiple conversational behaviors in videos into a low-dimensional motion latent space, and use these latent codes to animate the portrait image of the agent, denoted as " + }, + { + "bbox": [ + 313, + 169, + 555, + 324 + ], + "type": "inline_equation", + "content": "I_{\\text{self}}" + }, + { + "bbox": [ + 313, + 169, + 555, + 324 + ], + "type": "text", + "content": ". Note that, to ensure the richness and diversity of the motion latent space, we train stage 1 with plenty of real-life dialogue videos. Once the training converges, we freeze this part and start to train the second stage. The model takes as input the audio of the agent, denoted as " + }, + { + "bbox": [ + 313, + 169, + 555, + 324 + ], + "type": "inline_equation", + "content": "A_{\\text{self}}" + }, + { + "bbox": [ + 313, + 169, + 555, + 324 + ], + "type": "text", + "content": ", and the audio of the conversation partner, denoted as " + }, + { + "bbox": [ + 313, + 169, + 555, + 324 + ], + "type": "inline_equation", + "content": "A_{\\text{other}}" + }, + { + "bbox": [ + 313, + 169, + 555, + 324 + ], + "type": "text", + "content": ", to learn the mapping from this dual-track dialogue audio to motion latent codes through denoising. We present a comprehensive introduction to these two stages below." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 335, + 477, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 477, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 477, + 346 + ], + "type": "text", + "content": "3.1. Motion-Based Head Imitation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 354, + 555, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 555, + 485 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 555, + 485 + ], + "type": "text", + "content": "Motion Encoding. For motion encoding, we employ a motion encoder " + }, + { + "bbox": [ + 313, + 354, + 555, + 485 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_m" + }, + { + "bbox": [ + 313, + 354, + 555, + 485 + ], + "type": "text", + "content": " to learn an implicit latent representation, which captures verbal and non-verbal communicative behaviors from input facial images. " + }, + { + "bbox": [ + 313, + 354, + 555, + 485 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_m" + }, + { + "bbox": [ + 313, + 354, + 555, + 485 + ], + "type": "text", + "content": " comprises a convolutional-based feature extractor followed by MLP layers, which compresses the motion feature into a latent code. Note that, we use implicit representation instead of explicit 3DMM coefficients [28] because the expressiveness of 3DMM model is limited, and the commonly used expression coefficients, predicted by SOTA 3D facial estimation methods [5, 7], entangle with face shape to a certain extent." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 486, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 556, + 713 + ], + "type": "text", + "content": "A well-crafted motion latent space should demonstrate a high level of disentanglement to ensure that the extracted motions are appearance-independent. To achieve this, we apply two measures. (a) We set the motion latent code as a low-dimensional 1-D descriptor. It fosters effectively capturing the fundamental semantics of facial motion without entangling with appearance information, as explained in previous works [38, 43, 44]. (b) We design a hybrid facial representation to replace the original input image of " + }, + { + "bbox": [ + 313, + 486, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_m" + }, + { + "bbox": [ + 313, + 486, + 556, + 713 + ], + "type": "text", + "content": ". In particular, we first mask the majority of facial pixels, preserving only the eyes and lip regions, as we regard them as the most intricate and expressive components of facial expressions. This can block the interference of irrelevant information such as hair and background for expression distill. To provide facial orientation and contour information, we then use an off-the-shelf facial estimation model [20] to obtain the facial vertices, and project the contour-related points onto the masked image. Our ablation study shows the superiority of our proposed hybrid facial representation" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10669" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 71, + 360, + 252 + ], + "blocks": [ + { + "bbox": [ + 60, + 71, + 360, + 252 + ], + "lines": [ + { + "bbox": [ + 60, + 71, + 360, + 252 + ], + "spans": [ + { + "bbox": [ + 60, + 71, + 360, + 252 + ], + "type": "image", + "image_path": "7cc4c46f83c049322df0e0cfa2e462981e31bf2829f50178c427e8177dec017f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 263, + 555, + 309 + ], + "lines": [ + { + "bbox": [ + 54, + 263, + 555, + 309 + ], + "spans": [ + { + "bbox": [ + 54, + 263, + 555, + 309 + ], + "type": "text", + "content": "Figure 3. Overview of INFP. The first stage (Motion-Based Head Imitation) learns to project facial communicative behaviors from real-life conversation videos into a low-dimensional motion latent space, and use the latent codes to animate a static image. The second stage (Audio-Guided Motion Generation) learns the mapping from the input dyadic audio to motion latent codes through denoising, to achieve the audio-driven interactive head generation." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 361, + 71, + 550, + 253 + ], + "blocks": [ + { + "bbox": [ + 361, + 71, + 550, + 253 + ], + "lines": [ + { + "bbox": [ + 361, + 71, + 550, + 253 + ], + "spans": [ + { + "bbox": [ + 361, + 71, + 550, + 253 + ], + "type": "image", + "image_path": "32f5ec0b399785cca4e30e3ce51d5c14fa59c4fefe57de140f4facab683dce5f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 329, + 294, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 329, + 294, + 353 + ], + "spans": [ + { + "bbox": [ + 54, + 329, + 294, + 353 + ], + "type": "text", + "content": "compared with the other representations like the intact image [6] or 2D landmarks map [22]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 354, + 295, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 354, + 295, + 485 + ], + "spans": [ + { + "bbox": [ + 54, + 354, + 295, + 485 + ], + "type": "text", + "content": "Head Imitation. In this paper, we adopt a motion flow-based generation pipeline. Concretely, we first extract source motion code " + }, + { + "bbox": [ + 54, + 354, + 295, + 485 + ], + "type": "inline_equation", + "content": "m_{self}" + }, + { + "bbox": [ + 54, + 354, + 295, + 485 + ], + "type": "text", + "content": " from the input portrait image " + }, + { + "bbox": [ + 54, + 354, + 295, + 485 + ], + "type": "inline_equation", + "content": "I_{self}" + }, + { + "bbox": [ + 54, + 354, + 295, + 485 + ], + "type": "text", + "content": ", and driving motion codes " + }, + { + "bbox": [ + 54, + 354, + 295, + 485 + ], + "type": "inline_equation", + "content": "m_{dri}^{1:N}" + }, + { + "bbox": [ + 54, + 354, + 295, + 485 + ], + "type": "text", + "content": " from the driving video " + }, + { + "bbox": [ + 54, + 354, + 295, + 485 + ], + "type": "inline_equation", + "content": "V_{dri}" + }, + { + "bbox": [ + 54, + 354, + 295, + 485 + ], + "type": "text", + "content": ". A motion flow estimation model " + }, + { + "bbox": [ + 54, + 354, + 295, + 485 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 54, + 354, + 295, + 485 + ], + "type": "text", + "content": " takes these motion codes as input to predict the self " + }, + { + "bbox": [ + 54, + 354, + 295, + 485 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 54, + 354, + 295, + 485 + ], + "type": "text", + "content": " driving flow. We utilize a face encoder " + }, + { + "bbox": [ + 54, + 354, + 295, + 485 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{face}" + }, + { + "bbox": [ + 54, + 354, + 295, + 485 + ], + "type": "text", + "content": " to extract 3D appearance feature volume from the input image " + }, + { + "bbox": [ + 54, + 354, + 295, + 485 + ], + "type": "inline_equation", + "content": "I_{self}" + }, + { + "bbox": [ + 54, + 354, + 295, + 485 + ], + "type": "text", + "content": ". This feature volume is then warped by the motion flow and sent into a face decoder " + }, + { + "bbox": [ + 54, + 354, + 295, + 485 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{face}" + }, + { + "bbox": [ + 54, + 354, + 295, + 485 + ], + "type": "text", + "content": " to synthesize the final predictions " + }, + { + "bbox": [ + 54, + 354, + 295, + 485 + ], + "type": "inline_equation", + "content": "I_{pred}^{1:N}" + }, + { + "bbox": [ + 54, + 354, + 295, + 485 + ], + "type": "text", + "content": ". The overall pipeline can be written as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 489, + 295, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 489, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 62, + 489, + 295, + 510 + ], + "type": "interline_equation", + "content": "\\begin{array}{l}\\text {F l o w} _ {s \\rightarrow d} = \\mathbf {F} (\\mathbf {E} _ {m} (I _ {\\text {s e l f}}), \\mathbf {E} _ {m} (V _ {\\text {d r i}}))\\\\\\text {I} ^ {1: N} = \\mathbf {D} _ {n} \\left(W _ {n} - \\left(I _ {n}\\right) I _ {n} + \\dots\\right)\\end{array}\\tag {1}", + "image_path": "1dd3e7ba4ab0e02e83d6964a7a507f449b74ac4ed9a505368a64de2ef453298b.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 82, + 506, + 277, + 520 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 506, + 277, + 520 + ], + "spans": [ + { + "bbox": [ + 82, + 506, + 277, + 520 + ], + "type": "interline_equation", + "content": "I _ {p r e d} ^ {1: N} = \\mathbf {D} _ {f a c e} (W a r p (\\mathbf {E} _ {f a c e} (I _ {s e l f}), F l o w _ {s \\rightarrow d}))", + "image_path": "f0270669b41e3ad3c62de486b3c5f373cdcd7dc6738c45074030efbe41427c17.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 522, + 295, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 547 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 547 + ], + "type": "text", + "content": "Model details and loss function are shown in the supplementary material." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 552, + 237, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 237, + 564 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 237, + 564 + ], + "type": "text", + "content": "3.2. Audio-Guided Motion Generation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 570, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 570, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 570, + 295, + 713 + ], + "type": "text", + "content": "Previous works [34, 39, 59] tend to divide their models into a Speaker Generator and a Listener Generator. Since the Speaker Generator only receives speaking audio, while the Listener Generator solely processes listening audio, it is necessary to segment the audio of both agent and its conversation partner into speaking clips and listening clips. Hence, in a multi-turn dialogue, explicit role switching is also crucial. We argue that it can lead to unnatural transitions in communicative states, particularly in real-life interaction contexts characterized by frequent role switching. Besides, this design paradigm clearly goes against the intention of the conversational agent, as it should be able to" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 329, + 553, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 329, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 313, + 329, + 553, + 353 + ], + "type": "text", + "content": "naturally present different communicative states based on the progress of the conversation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 354, + 554, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 554, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 554, + 449 + ], + "type": "text", + "content": "In this paper, we present a new paradigm, in which the model takes as input the audio from both the agent and its conversation partner, eliminating the necessity for explicit role assignment or switching. The key innovations lie in extracting interactive information from dual-track dyadic audio to construct speaking-listening mixed motions, and mapping these motions to the motion latent space of stage 1 to achieve audio-driven head generation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 450, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 555, + 714 + ], + "type": "text", + "content": "Motion Feature Construction. As shown in Fig. 3, an interactive motion guider is introduced to extract motion feature from the input audio. Within it, there are two memory banks " + }, + { + "bbox": [ + 313, + 450, + 555, + 714 + ], + "type": "inline_equation", + "content": "M_v" + }, + { + "bbox": [ + 313, + 450, + 555, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 450, + 555, + 714 + ], + "type": "inline_equation", + "content": "M_{nv}" + }, + { + "bbox": [ + 313, + 450, + 555, + 714 + ], + "type": "text", + "content": ", storing verbal and nonverbal motion, respectively. Each motion memory bank consists of learnable embeddings to memorize typical motions, denoted as " + }, + { + "bbox": [ + 313, + 450, + 555, + 714 + ], + "type": "inline_equation", + "content": "e^{1:K}" + }, + { + "bbox": [ + 313, + 450, + 555, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 450, + 555, + 714 + ], + "type": "inline_equation", + "content": "e^k \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 313, + 450, + 555, + 714 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 313, + 450, + 555, + 714 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 450, + 555, + 714 + ], + "type": "text", + "content": "-th motion embedding and " + }, + { + "bbox": [ + 313, + 450, + 555, + 714 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 450, + 555, + 714 + ], + "type": "text", + "content": " is the dimension. We set " + }, + { + "bbox": [ + 313, + 450, + 555, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 450, + 555, + 714 + ], + "type": "text", + "content": " to 64 and " + }, + { + "bbox": [ + 313, + 450, + 555, + 714 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 450, + 555, + 714 + ], + "type": "text", + "content": " to 512 in our experiments. When agent is speaking, " + }, + { + "bbox": [ + 313, + 450, + 555, + 714 + ], + "type": "inline_equation", + "content": "A_{self}" + }, + { + "bbox": [ + 313, + 450, + 555, + 714 + ], + "type": "text", + "content": " contains rich information, so we take it as " + }, + { + "bbox": [ + 313, + 450, + 555, + 714 + ], + "type": "inline_equation", + "content": "Query" + }, + { + "bbox": [ + 313, + 450, + 555, + 714 + ], + "type": "text", + "content": " and obtain verbal motion feature from " + }, + { + "bbox": [ + 313, + 450, + 555, + 714 + ], + "type": "inline_equation", + "content": "M_v" + }, + { + "bbox": [ + 313, + 450, + 555, + 714 + ], + "type": "text", + "content": ", which is taken as " + }, + { + "bbox": [ + 313, + 450, + 555, + 714 + ], + "type": "inline_equation", + "content": "Key" + }, + { + "bbox": [ + 313, + 450, + 555, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 450, + 555, + 714 + ], + "type": "inline_equation", + "content": "Value" + }, + { + "bbox": [ + 313, + 450, + 555, + 714 + ], + "type": "text", + "content": ", through a cross-attention layer. Since the partner is listening, " + }, + { + "bbox": [ + 313, + 450, + 555, + 714 + ], + "type": "inline_equation", + "content": "A_{other}" + }, + { + "bbox": [ + 313, + 450, + 555, + 714 + ], + "type": "text", + "content": " carries very little information, hence no effective non-verbal motion feature can be obtained from " + }, + { + "bbox": [ + 313, + 450, + 555, + 714 + ], + "type": "inline_equation", + "content": "M_{nv}" + }, + { + "bbox": [ + 313, + 450, + 555, + 714 + ], + "type": "text", + "content": ". Therefore, after feature fusion, which involves element-wise sum and several MLP layers, verbal motion feature dominates and drives the agent to express the intended speaking state. However, when the partner starts speaking, " + }, + { + "bbox": [ + 313, + 450, + 555, + 714 + ], + "type": "inline_equation", + "content": "A_{self}" + }, + { + "bbox": [ + 313, + 450, + 555, + 714 + ], + "type": "text", + "content": " becomes weak and " + }, + { + "bbox": [ + 313, + 450, + 555, + 714 + ], + "type": "inline_equation", + "content": "A_{other}" + }, + { + "bbox": [ + 313, + 450, + 555, + 714 + ], + "type": "text", + "content": " fulfills a vital role. Accordingly, non-verbal motion feature from " + }, + { + "bbox": [ + 313, + 450, + 555, + 714 + ], + "type": "inline_equation", + "content": "M_{nv}" + }, + { + "bbox": [ + 313, + 450, + 555, + 714 + ], + "type": "text", + "content": " dominates and forces the agent to present a listening state. In this way, our model can dynamically construct" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10670" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": "interactive motion feature " + }, + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "inline_equation", + "content": "f_{m}" + }, + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": " based on the content of the dialogue audio, allowing the agent to convey the intended state. Note that, " + }, + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "inline_equation", + "content": "A_{\\text{self}}" + }, + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "inline_equation", + "content": "A_{\\text{other}}" + }, + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": " are encoded to the feature space through an off-the-shelf HuBERT audio encoder [12], before sent into the model." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 138, + 295, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 138, + 295, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 138, + 295, + 281 + ], + "type": "text", + "content": "Inspired by stylistic facial animation methods [21, 35], we also introduce a motion style vector " + }, + { + "bbox": [ + 55, + 138, + 295, + 281 + ], + "type": "inline_equation", + "content": "s_m" + }, + { + "bbox": [ + 55, + 138, + 295, + 281 + ], + "type": "text", + "content": " to explicitly edit the motion embeddings of memory banks, through a style modulation layer raised by StyleGAN2[15]. " + }, + { + "bbox": [ + 55, + 138, + 295, + 281 + ], + "type": "inline_equation", + "content": "s_m" + }, + { + "bbox": [ + 55, + 138, + 295, + 281 + ], + "type": "text", + "content": " is first extracted from motion latent codes through a motion style encoder " + }, + { + "bbox": [ + 55, + 138, + 295, + 281 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{style}" + }, + { + "bbox": [ + 55, + 138, + 295, + 281 + ], + "type": "text", + "content": ", and then squeezed along the temporal dimension to a global descriptor. We argue that " + }, + { + "bbox": [ + 55, + 138, + 295, + 281 + ], + "type": "inline_equation", + "content": "s_m" + }, + { + "bbox": [ + 55, + 138, + 295, + 281 + ], + "type": "text", + "content": " contains some global information like emotions and attitudes, thus enhancing the authenticity and vividness of the generated results. Note that during training, " + }, + { + "bbox": [ + 55, + 138, + 295, + 281 + ], + "type": "inline_equation", + "content": "s_m" + }, + { + "bbox": [ + 55, + 138, + 295, + 281 + ], + "type": "text", + "content": " is sourced from an arbitrary video clip of the driven individual, while in testing, " + }, + { + "bbox": [ + 55, + 138, + 295, + 281 + ], + "type": "inline_equation", + "content": "s_m" + }, + { + "bbox": [ + 55, + 138, + 295, + 281 + ], + "type": "text", + "content": " can be extracted from any videos or set as null." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 289, + 295, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 289, + 295, + 587 + ], + "spans": [ + { + "bbox": [ + 56, + 289, + 295, + 587 + ], + "type": "text", + "content": "Motion Mapping. To map the interactive motions " + }, + { + "bbox": [ + 56, + 289, + 295, + 587 + ], + "type": "inline_equation", + "content": "f_{m}" + }, + { + "bbox": [ + 56, + 289, + 295, + 587 + ], + "type": "text", + "content": " to the pretrained motion latent space of stage 1, we employ a conditional diffusion transformer " + }, + { + "bbox": [ + 56, + 289, + 295, + 587 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 56, + 289, + 295, + 587 + ], + "type": "text", + "content": " [11]. Given the data distribution " + }, + { + "bbox": [ + 56, + 289, + 295, + 587 + ], + "type": "inline_equation", + "content": "q(m^{1:N},f_m)" + }, + { + "bbox": [ + 56, + 289, + 295, + 587 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 56, + 289, + 295, + 587 + ], + "type": "inline_equation", + "content": "m^{1:N}" + }, + { + "bbox": [ + 56, + 289, + 295, + 587 + ], + "type": "text", + "content": " represents the corresponding motion latent codes with " + }, + { + "bbox": [ + 56, + 289, + 295, + 587 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 56, + 289, + 295, + 587 + ], + "type": "text", + "content": " frames, the diffusion model approximates the conditional distribution " + }, + { + "bbox": [ + 56, + 289, + 295, + 587 + ], + "type": "inline_equation", + "content": "q(m^{1:N}|f_m)" + }, + { + "bbox": [ + 56, + 289, + 295, + 587 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 56, + 289, + 295, + 587 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 56, + 289, + 295, + 587 + ], + "type": "text", + "content": " only has 4 blocks, which makes our framework lightweight enough to achieve real-time interaction. Each block consists of a self-attention layer, a motion-attention layer, and a temporal-attention layer. " + }, + { + "bbox": [ + 56, + 289, + 295, + 587 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 56, + 289, + 295, + 587 + ], + "type": "text", + "content": " predicts the noise added to the ground-truth motion latent codes for each denoising step. Diffusion timestep is converted into the sinusoidal embedding and then concatenated to the noised motion latent codes at the time dimension. In the motion attention layer, we take the latent feature output from preceding self-attention layer as the Query, while the interactive motion feature " + }, + { + "bbox": [ + 56, + 289, + 295, + 587 + ], + "type": "inline_equation", + "content": "f_{m}" + }, + { + "bbox": [ + 56, + 289, + 295, + 587 + ], + "type": "text", + "content": " is utilized as the Key and Value. Inspired byAnimateDiff [10], we implement a temporal layer to ensure a smooth transition between adjacent windows. In particular, we incorporate the last 10 frames of the generated motion latent codes from the previous window as the condition of the current one. Note that in the inference time, we add noise to the self motion latent codes " + }, + { + "bbox": [ + 56, + 289, + 295, + 587 + ], + "type": "inline_equation", + "content": "m_{\\text{self}}^{1:N}" + }, + { + "bbox": [ + 56, + 289, + 295, + 587 + ], + "type": "text", + "content": ", which are extracted from " + }, + { + "bbox": [ + 56, + 289, + 295, + 587 + ], + "type": "inline_equation", + "content": "I_{\\text{self}}" + }, + { + "bbox": [ + 56, + 289, + 295, + 587 + ], + "type": "text", + "content": " and replicated into " + }, + { + "bbox": [ + 56, + 289, + 295, + 587 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 56, + 289, + 295, + 587 + ], + "type": "text", + "content": " copies." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": "Training Strategy. In the training stage, we randomly set style vector to null with a probability of 0.3, and randomly drop interactive motion feature and previous motion latent codes with a probability of 0.5. Also note that we adopt a warm-up strategy for training stage 2. Specifically, we only select single-sided conversation clips from our training set as \"simple case\" to train the model for several epochs, which helps with the initialization of two memory banks. After that, we randomly sample multi-turn clips from entire dataset for the remaining training process." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 314, + 72, + 369, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 369, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 369, + 83 + ], + "type": "text", + "content": "4. Dataset" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 91, + 553, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 553, + 234 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 553, + 234 + ], + "type": "text", + "content": "Existing dyadic interaction datasets have limitations in both scale and quality. ViCo [58] is a small-scale dataset with 1.6 hours and 96 IDs, and it lacks multi-turn conversation scenarios. Although ViCo-X [57] features rich multi-turn conversations, the total duration is only 0.4 hours. Besides, its background is monotonous, all of which are green screen. RealTalk [8] is a medium-scale dataset capturing dyadic conversations between pairs of individuals totaling to 115 hours of online public videos. However, it does not restrict the resolution of faces in the video, leading to low face resolution in many videos, thus impacting the quality of model generation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 235, + 555, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 235, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 235, + 555, + 415 + ], + "type": "text", + "content": "We propose DyConv, a large-scale dataset of multi-turn dyadic conversations collected from the Internet. The dataset contains video clips with a duration of over 200 hours, captures a wide gamut of emotions and expressions. We employ an off-the-shelf face detection model [51], and keep only frames where both individuals of the conversation are fully visible and their facial resolution is greater than " + }, + { + "bbox": [ + 313, + 235, + 555, + 415 + ], + "type": "inline_equation", + "content": "400 \\times 400" + }, + { + "bbox": [ + 313, + 235, + 555, + 415 + ], + "type": "text", + "content": ". In addition, we utilize a SOTA speech separation model [54] to separate the audio of two persons in the conversation, denoted as " + }, + { + "bbox": [ + 313, + 235, + 555, + 415 + ], + "type": "inline_equation", + "content": "A_{p1}" + }, + { + "bbox": [ + 313, + 235, + 555, + 415 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 235, + 555, + 415 + ], + "type": "inline_equation", + "content": "A_{p2}" + }, + { + "bbox": [ + 313, + 235, + 555, + 415 + ], + "type": "text", + "content": ". Then we run an active speaker detection model [1] to match each audio clip to the corresponding face in the original video. We will publicly release the dataset, along with separated audios and annotations. We hope it will encourage the investigation of head generation for dyadic interaction." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 425, + 391, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 425, + 391, + 438 + ], + "spans": [ + { + "bbox": [ + 314, + 425, + 391, + 438 + ], + "type": "text", + "content": "5. Experiment" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 445, + 448, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 448, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 448, + 456 + ], + "type": "text", + "content": "5.1. Implementation Details" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 461, + 554, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 554, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 554, + 544 + ], + "type": "text", + "content": "In our experiments, we train the diffusion transformer using AdamW optimizer [19], with " + }, + { + "bbox": [ + 313, + 461, + 554, + 544 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.9" + }, + { + "bbox": [ + 313, + 461, + 554, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 461, + 554, + 544 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.999" + }, + { + "bbox": [ + 313, + 461, + 554, + 544 + ], + "type": "text", + "content": ", a learning rate of 1e-4, a weight decay of 1e-2, and a batch size of 32. For inference, we use a DDIM sampler [30] for 20 denoising steps. The CFG parameter of motion feature is set to 2, and the CFG parameter of previous motion latent is set to 1.5." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 552, + 400, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 400, + 565 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 400, + 565 + ], + "type": "text", + "content": "5.2. Comparisons" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 570, + 553, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 553, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 553, + 640 + ], + "type": "text", + "content": "Although originally designed for dyadic interactive scenarios, our method can be naturally generalized to other applications, such as listening head generation and talking head generation. This allows us to fully validate the advancement of our proposed framework against some of the current SOTA methods in related areas." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "text", + "content": "Evaluation Metrics. To quantitatively evaluate the visual quality, we utilize the Structured Similarity (SSIM), Peak Signal-to-Noise Ratio (PSNR) and Frechet Inception Distance (FID) for the generated videos. Additionally, LPIPS [50] is employed to measure the feature-level similarity between generated and ground-truth frames. To as" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10671" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 78, + 138, + 136, + 194 + ], + "blocks": [ + { + "bbox": [ + 78, + 138, + 136, + 194 + ], + "lines": [ + { + "bbox": [ + 78, + 138, + 136, + 194 + ], + "spans": [ + { + "bbox": [ + 78, + 138, + 136, + 194 + ], + "type": "image", + "image_path": "b8d49fc67c745b6073605764386dd2a295b68fe2afbc67a65ec00e528d219b02.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 80, + 194, + 135, + 204 + ], + "lines": [ + { + "bbox": [ + 80, + 194, + 135, + 204 + ], + "spans": [ + { + "bbox": [ + 80, + 194, + 135, + 204 + ], + "type": "text", + "content": "Reference Image" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 142, + 74, + 536, + 483 + ], + "blocks": [ + { + "bbox": [ + 142, + 74, + 536, + 483 + ], + "lines": [ + { + "bbox": [ + 142, + 74, + 536, + 483 + ], + "spans": [ + { + "bbox": [ + 142, + 74, + 536, + 483 + ], + "type": "image", + "image_path": "c571530f3918fed859d3aed95d251d8bbca07b5f8d2bcc876d05cf6b41820208.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 491, + 555, + 525 + ], + "lines": [ + { + "bbox": [ + 55, + 491, + 555, + 525 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 555, + 525 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison of interactive head generation on DyConv. Given the input audio of dyadic conversations (the audio of the predicted agent is black and the audio of the conversation partner is blue), our framework generates reasonable and vivid facial and head pose movements as various roles in interactive scenarios." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 546, + 295, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 295, + 643 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 295, + 643 + ], + "type": "text", + "content": "sess the preservation of facial appearance, we calculate the cosine similarity (CSIM) [4] between the facial features of the reference image and the generated video frames. In term of motion diversity of generated videos, we use SI for diversity (SID) and Variance (Var) following previous studies [23, 39]. We use the confidence score of SyncNet (SyncScore) [3] to evaluate the synchronization between lip movements and audio signals of the agent." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 650, + 206, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 650, + 206, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 650, + 206, + 661 + ], + "type": "text", + "content": "5.2.1. Interactive Head Generation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "Since only a few works focus on the interactive head generation [42, 59] and these methods are not open-sourced, direct comparisons with them are considerably difficult. To effectively validate our method, we select the SOTA dyadic" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 546, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 555, + 640 + ], + "type": "text", + "content": "modeling work DIM [39] as our baseline and utilize the officially released code to re-train the model on our datasets. Specifically, we first use the whole dataset to pretrain the VQ-VAE for the listener and speaker and then split the dataset into two subsets to train the listener generator and the speaker generator, respectively. To better visualize results of DIM, we also train the face renderer [28] to generate photo-realistic videos." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": "Quantitative Results. The quantitative experiments are carried out on the test set of DyConv. Please refer to Tab. 1 for the results. It can be seen that our method outperforms the baseline cross a variety of metrics. In particular, our face decoder achieves a better visual quality such as SSIM, PSNR and FID. Thanks to the disentanglement and expres" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10672" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 203 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 203 + ], + "type": "text", + "content": "siveness of our latent space, our method is superior to DIM in term of audio-lip-synchronization by a large margin as there is a much closer distance of our SyncScore to the ground truth. As for the LPIPS and CISM, our method still achieves solid performances, reflecting the better identity preservation. Of particular note, our method is far ahead of the baseline in term of SID and Var, which proves the generalization and effectiveness of our method to generate a great range of diverse motions. It benefits from the design of interactive motion guider as well as conditional denoising transformer." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 204, + 295, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 204, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 204, + 295, + 453 + ], + "type": "text", + "content": "Qualitative Results. Subjective evaluation is crucial in identifying the performance of generative models, particularly on videos. We strongly recommend readers to watch our supplementary video. Two examples and their comparisons with the baseline are shown in Fig. 4. We use two typical audio clips in the dyadic conversation scenario to drive reference identities randomly selected from the test set. It can be obviously observed that our method can generate more diverse head pose movements in comparison to DIM. Meanwhile, results of our method demonstrate more vivid facial expressions, which significantly improve the naturalness of generated videos. It is worth to noting that facial behaviors synthesized by our method are well aligned with the dialogue audio content. This suggests the effectiveness of our method in dyadic interactions, which is capable of capturing distinct characteristics of various states, such as listening, speaking as well as the mutual interaction, and naturally adapting its role among them based on the progress of the conversation. In contrast, DIM is unable to generate appropriate behaviors such as reasonable lip movement and response feedback." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 455, + 295, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 455, + 295, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 455, + 295, + 586 + ], + "type": "text", + "content": "User Study. We also invite 20 participants to conduct a user study for further subjective evaluation. Specifically, we apply INFP and DIM to synthesize results of 20 dyadic conversation videos randomly sampled from the test set. By adopting the commonly used Mean Opinion Scores (MOS) rating protocol, we request all the participants to provide their ratings (from 1 to 5, the higher the better) on four aspects for each generated video: naturalness, motion diversity, audio-visual alignment and visual quality. As detailed in Tab. 2, our method demonstrates superior performances across all evaluated aspects compared to DIM." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 591, + 200, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 591, + 200, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 591, + 200, + 602 + ], + "type": "text", + "content": "5.2.2. Listening Head Generation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": "Since our method can generalize to the task of listening head generation, we directly use our model without any modification to conduct additional experiments. Specifically, we compare our framework with SOTA listening head generation methods, including L2L [23], RLHG [58] and DIM [39]. ViCo [58] is selected as the benchmark. For quantitative evaluation, we utilized the same test set and inherited both quantitative and qualitative results presented in DIM. We conducted the comparison on FD, MSE, SID" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "text", + "content": "and Var and each metric is calculated on expression and head pose, respectively. Different to other methods, our method generate photo-realistic video rather than 3DMM coefficients. To this end, we use the off-the-shelf 3D face reconstruction method [7] to extract expression and head pose coefficients from our generated videos. As shown in Tab. 3, INFP outperforms other SOTA methods in the majority of metrics. While SID and Var are performing well, our method remains highly effective in both FD and MSE. This indicates the power of our novel framework design and dataset quality, making generated facial expressions and head pose movements closer to real human behaviors. The visualization results in the supplementary materials also demonstrate that INFP achieves more expressive and diverse motions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 260, + 451, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 260, + 451, + 272 + ], + "spans": [ + { + "bbox": [ + 313, + 260, + 451, + 272 + ], + "type": "text", + "content": "5.2.3. Talking Head Generation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 276, + 555, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 276, + 555, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 276, + 555, + 479 + ], + "type": "text", + "content": "For talking head generation, we select SadTalker[52], AntiTalker[17], and EchoMimic[2] as comparing methods. As all of them reported evaluation results on HDTF [53], we randomly sampled 20 videos from it as the test set to conduct our experiments for a fair comparison. Results in Tab. 4 show that our model is superior in audiolip-synchronization and identity preservation. The reason behind this might be that we carefully build a disentangled and expressive motion latent space and it makes our model easier to learn the correlation between audio and lip movements without the interference of appearance features. It can also be observed from the visualization results in the supplementary materials. Our SSIM and PSNR score is slightly behind EchoMimic because EchoMimic is finetuned based on the Stable Diffusion V1.5 [29]. Despite this, our method can achieve better scores in terms of FID, meaning a good visual quality of our results." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 490, + 408, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 490, + 408, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 490, + 408, + 502 + ], + "type": "text", + "content": "5.3. Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 508, + 553, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 508, + 553, + 531 + ], + "spans": [ + { + "bbox": [ + 313, + 508, + 553, + 531 + ], + "type": "text", + "content": "We ablate 3 key components of our framework, and the results are listed in Tab. 1." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 533, + 554, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 554, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 554, + 664 + ], + "type": "text", + "content": "Motion Memory Bank. We remove the verbal and non-verbal motion memory banks from interactive motion guider in stage 2, and directly use the dyadic audio feature after several MLP layers as the condition for the motion-attention layer. Results show that the generated lip-related motion behaviors are clearly inconsistent with the ground-truth. In additional, there is a significant drop of both SID and Var scores. These suggest that memory banks can effectively store rich motion patterns in different conversational states. The experiment on memory bank size is shown in the supplementary material." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": "Style Modulation. We remove the motion style vector " + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "inline_equation", + "content": "s_m" + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": " and corresponding modulation layers. As we expected, the authenticity and vividness of the results decrease, which can also be reflected in the user study." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10673" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 88, + 70, + 525, + 190 + ], + "blocks": [ + { + "bbox": [ + 88, + 70, + 525, + 190 + ], + "lines": [ + { + "bbox": [ + 88, + 70, + 525, + 190 + ], + "spans": [ + { + "bbox": [ + 88, + 70, + 525, + 190 + ], + "type": "table", + "html": "
MethodsSSIM↑PSNR↑FID↓SyncScore↑LPIPS↓CSIM↑SID↑Var↑
DIM [39]0.65120.41734.3614.7780.4850.8240.7660.825
INFP (Ours)0.83431.56215.7277.1880.2570.9042.6132.386
w/o Motion Memory0.83031.21818.3346.1030.2590.8992.1532.016
w/o Style Modulation0.83131.44216.0297.0620.2710.9042.5512.316
w/ Intact Image0.80228.48816.9906.8120.2660.8422.4702.148
w/ Landmarks Map0.82130.69316.3276.8330.2810.9012.6012.335
GT1.000N/A0.0007.2610.0000.9672.8912.435
", + "image_path": "a1980df37238dd08edd195e3c4a14d544ceccc8cb460867351ee4c18066f806b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 227, + 294, + 280 + ], + "blocks": [ + { + "bbox": [ + 166, + 198, + 443, + 209 + ], + "lines": [ + { + "bbox": [ + 166, + 198, + 443, + 209 + ], + "spans": [ + { + "bbox": [ + 166, + 198, + 443, + 209 + ], + "type": "text", + "content": "Table 1. Quantitative comparison of interactive head generation on DyConv." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 227, + 294, + 280 + ], + "lines": [ + { + "bbox": [ + 58, + 227, + 294, + 280 + ], + "spans": [ + { + "bbox": [ + 58, + 227, + 294, + 280 + ], + "type": "table", + "html": "
MethodsNaturalnessMotion DiversityAV-AlignVisual
DIM [39]2.712.142.653.13
INFP w/o Style4.023.984.284.11
INFP (Ours)4.384.494.334.13
", + "image_path": "b61b10dccf149a32b048c6c087de601f6c86143f0187a1c836661928cedc4a11.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 58, + 309, + 294, + 385 + ], + "blocks": [ + { + "bbox": [ + 73, + 287, + 277, + 298 + ], + "lines": [ + { + "bbox": [ + 73, + 287, + 277, + 298 + ], + "spans": [ + { + "bbox": [ + 73, + 287, + 277, + 298 + ], + "type": "text", + "content": "Table 2. User study measured by Mean Opinion Scores." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 309, + 294, + 385 + ], + "lines": [ + { + "bbox": [ + 58, + 309, + 294, + 385 + ], + "spans": [ + { + "bbox": [ + 58, + 309, + 294, + 385 + ], + "type": "table", + "html": "
MethodFD↓MSE↓SID↑Var↑
ExpPoseExpPoseExpPoseExpPose
L2L [23]33.930.060.930.012.772.660.830.02
RLHG [58]39.020.070.860.013.623.171.520.02
DIM [39]23.880.060.700.013.712.351.530.02
INFP (Ours)18.630.070.510.014.783.922.830.18
", + "image_path": "42b8d37ab92fcc3d7c2e90f3fb16c1460034776a861cc69710539c1f885fe46e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 59, + 426, + 292, + 494 + ], + "blocks": [ + { + "bbox": [ + 55, + 393, + 295, + 415 + ], + "lines": [ + { + "bbox": [ + 55, + 393, + 295, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 393, + 295, + 415 + ], + "type": "text", + "content": "Table 3. Quantitative comparison of listening head generation on ViCo [58]." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 426, + 292, + 494 + ], + "lines": [ + { + "bbox": [ + 59, + 426, + 292, + 494 + ], + "spans": [ + { + "bbox": [ + 59, + 426, + 292, + 494 + ], + "type": "table", + "html": "
MethodsSSIM↑PSNR↑FID↓SyncScore↑CSIM↑
SadTalker [52]0.78126.65028.5166.1250.814
AniTalker [17]0.78926.82228.7785.7380.858
EchoMimic [2]0.80728.98320.8736.2510.863
INFP (Ours)0.80428.65718.5826.6700.886
", + "image_path": "be212b0b1c3e2397cfd34902fcadc4a61ee38b3fe23f0484a12e5dacf2163b83.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 501, + 295, + 523 + ], + "lines": [ + { + "bbox": [ + 55, + 501, + 295, + 523 + ], + "spans": [ + { + "bbox": [ + 55, + 501, + 295, + 523 + ], + "type": "text", + "content": "Table 4. Quantitative comparison of talking head generation on HDTF [53]." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 544, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 605 + ], + "type": "text", + "content": "Hybrid Facial Representation. We change the input to " + }, + { + "bbox": [ + 55, + 544, + 295, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_m" + }, + { + "bbox": [ + 55, + 544, + 295, + 605 + ], + "type": "text", + "content": " in stage 1 from our carefully-designed hybrid facial representation to the original intact image or 2D landmarks map. We observe that there is degradation in generation quality and a leakage of appearance information." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 612, + 217, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 217, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 217, + 624 + ], + "type": "text", + "content": "5.4. Visualization of Style Control" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 629, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 712 + ], + "type": "text", + "content": "As illustrated in Sec. 3.2, our model supports inputting style vector " + }, + { + "bbox": [ + 55, + 629, + 295, + 712 + ], + "type": "inline_equation", + "content": "s_m" + }, + { + "bbox": [ + 55, + 629, + 295, + 712 + ], + "type": "text", + "content": " extracted from any portrait video to globally control the emotions or attitudes of the generated results. Here we select 3 video clips with varying degrees of \"happiness\" to control emotions during generation. Results are shown in Fig. 5. It can be seen from the figure that " + }, + { + "bbox": [ + 55, + 629, + 295, + 712 + ], + "type": "inline_equation", + "content": "s_m" + }, + { + "bbox": [ + 55, + 629, + 295, + 712 + ], + "type": "text", + "content": " effectively control the emotions of the driven individual." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 328, + 228, + 543, + 371 + ], + "blocks": [ + { + "bbox": [ + 328, + 228, + 543, + 371 + ], + "lines": [ + { + "bbox": [ + 328, + 228, + 543, + 371 + ], + "spans": [ + { + "bbox": [ + 328, + 228, + 543, + 371 + ], + "type": "image", + "image_path": "dc25b44437a90aa874b190cb2f512ca773ff5ad605d728b42492dcf5a0378246.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 314, + 378, + 553, + 389 + ], + "lines": [ + { + "bbox": [ + 314, + 378, + 553, + 389 + ], + "spans": [ + { + "bbox": [ + 314, + 378, + 553, + 389 + ], + "type": "text", + "content": "Figure 5. Visualization of style control with our style modulation." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 413, + 388, + 425 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 413, + 388, + 425 + ], + "spans": [ + { + "bbox": [ + 314, + 413, + 388, + 425 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 434, + 555, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 434, + 555, + 578 + ], + "spans": [ + { + "bbox": [ + 312, + 434, + 555, + 578 + ], + "type": "text", + "content": "In this paper, we propose INFP, a novel audio-driven interactive head generation framework for dyadic conversations, which aims to mimic facial behaviors of human-human interactions conditioning on the dyadic audio. As demonstrated, our method is a unified framework and is capable of smoothly and naturally adapting to various conversational roles without explicit assignment or switching. Meanwhile, we collect a large scale audio-visual dataset for multi-turn dyadic conversations, which could facilitate further researches of dyadic interactive head generation. Extensive experiments and ablations prove the advancement and effectiveness of our method over other SOTA methods." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 580, + 555, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 580, + 555, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 580, + 555, + 664 + ], + "type": "text", + "content": "Limitation and Future Work. Currently, our method only takes conversation audios as the input, combining control signals from multiple modalities, such as visual and textual contents, can offer additional capacities. As our proposed framework focuses on head synthesis, extending the generation range to the upper, even whole body might be a interesting and challenging future work." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": "Ethical Considerations. There might be the potential misuse of our method to fabricate deceptive talks and speeches. To this end, we will restrict access to our core models exclusively to research institutions." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10674" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 294, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 294, + 123 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 294, + 123 + ], + "type": "text", + "content": "[1] Stanislav Beliaev, Yuriy Rebryk, and Boris Ginsburg. Talknet: Fully-convolutional non-autoregressive speech synthesis model, 2020. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 295, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 295, + 169 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 295, + 169 + ], + "type": "text", + "content": "[2] Zhiyuan Chen, Jiajiong Cao, Zhiquan Chen, Yuming Li, and Chenguang Ma. Echomimic: Lifelike audio-driven portrait animations through editable landmark conditions. arXiv preprint arXiv:2407.08136, 2024. 1, 3, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 171, + 294, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 171, + 294, + 203 + ], + "spans": [ + { + "bbox": [ + 62, + 171, + 294, + 203 + ], + "type": "text", + "content": "[3] J. S. Chung and A. Zisserman. Out of time: automated lip sync in the wild. In Workshop on Multi-view Lip-reading, ACCV, 2016. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 205, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 205, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 62, + 205, + 294, + 239 + ], + "type": "text", + "content": "[4] Jiankang Deng, Jia Guo, Xue Niannan, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In CVPR, 2019. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 240, + 294, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 240, + 294, + 304 + ], + "spans": [ + { + "bbox": [ + 62, + 240, + 294, + 304 + ], + "type": "text", + "content": "[5] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 308, + 294, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 308, + 294, + 350 + ], + "spans": [ + { + "bbox": [ + 62, + 308, + 294, + 350 + ], + "type": "text", + "content": "[6] Nikita Drobyshev, Jenya Chelishev, Taras Khakhulin, Aleksei Ivakhnenko, Victor Lempitsky, and Egor Zakharov. Megaportraits: One-shot megapixel neural head avatars. 2022. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 353, + 294, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 353, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 62, + 353, + 294, + 407 + ], + "type": "text", + "content": "[7] Panagiotis P. Filntisis, George Retsinas, Foivos Paraperas-Papantoniou, Athanasios Katsamanis, Anastasios Roussos, and Petros Maragos. Visual speech-aware perceptual 3d facial expression reconstruction from videos. arXiv preprint arXiv:2207.11094, 2022. 3, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 409, + 294, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 409, + 294, + 442 + ], + "spans": [ + { + "bbox": [ + 62, + 409, + 294, + 442 + ], + "type": "text", + "content": "[8] Scott Geng, Revant Teotia, Purva Tendulkar, Sachit Menon, and Carl Vondrick. Affective faces for goal-driven dyadic communication. 2023. 2, 3, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 445, + 294, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 445, + 294, + 520 + ], + "spans": [ + { + "bbox": [ + 62, + 445, + 294, + 520 + ], + "type": "text", + "content": "[9] Jiazhi Guan, Zhanwang Zhang, Hang Zhou, Tianshu Hu, Kaisiyuan Wang, Dongliang He, Haocheng Feng, Jingtuo Liu, Errui Ding, Ziwei Liu, et al. Stylesync: High-fidelity generalized and personalized lip sync in style-based generator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1505-1515, 2023. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 522, + 294, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 522, + 294, + 577 + ], + "spans": [ + { + "bbox": [ + 57, + 522, + 294, + 577 + ], + "type": "text", + "content": "[10] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Imagine your personalized text-to-image diffusion models without specific tuning. International Conference on Learning Representations, 2024. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 578, + 294, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 578, + 294, + 611 + ], + "spans": [ + { + "bbox": [ + 57, + 578, + 294, + 611 + ], + "type": "text", + "content": "[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 613, + 294, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 294, + 677 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 294, + 677 + ], + "type": "text", + "content": "[12] Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. Hubert: Self-supervised speech representation learning by masked prediction of hidden units. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 29:3451-3460, 2021. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 680, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 294, + 713 + ], + "type": "text", + "content": "[13] Ailin Huang, Zhewei Huang, and Shuchang Zhou. Perceptual conversational head generation with regularized driver and enhanced renderer. In Proceedings of the 30th ACM" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "text", + "content": "International Conference on Multimedia, page 7050-7054. ACM, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "text", + "content": "[14] Yinghao Huang, Leo Ho, Dafei Qin, Mingyi Shi, and Taku Komura. Interact: Capture and modelling of realistic, expressive and interactive activities between two persons in daily scenarios. 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 141, + 553, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 175 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 175 + ], + "type": "text", + "content": "[15] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In Proc. CVPR, 2020. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 176, + 553, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 176, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 316, + 176, + 553, + 230 + ], + "type": "text", + "content": "[16] Jin Liu, Xi Wang, Xiaomeng Fu, Yesheng Chai, Cai Yu, Jiao Dai, and Jizhong Han. Mfr-net: Multi-faceted responsive listening head generation via denoising diffusion model. In Proceedings of the 31st ACM International Conference on Multimedia, pages 6734-6743, 2023. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 232, + 553, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 232, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 317, + 232, + 553, + 275 + ], + "type": "text", + "content": "[17] Tao Liu, Feilong Chen, Shuai Fan, Chenpeng Du, Qi Chen, Xie Chen, and Kai Yu. Anitalker: Animate vivid and diverse talking faces through identity-decoupled facial motion encoding, 2024. 1, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 277, + 553, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 277, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 317, + 277, + 553, + 331 + ], + "type": "text", + "content": "[18] Xi Liu, Ying Guo, Cheng Zhen, Tong Li, Yingying Ao, and Pengfei Yan. Custom listener: Text-guided responsive interaction for user-friendly listening head generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2415-2424, 2024. 1, 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 333, + 553, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 333, + 553, + 355 + ], + "spans": [ + { + "bbox": [ + 317, + 333, + 553, + 355 + ], + "type": "text", + "content": "[19] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization, 2019. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 357, + 553, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 357, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 317, + 357, + 553, + 421 + ], + "type": "text", + "content": "[20] Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuo-Ling Chang, Ming Yong, Juhyun Lee, Wan-Teh Chang, Wei Hua, Manfred Georg, and Matthias Grundmann. Mediapipe: A framework for perceiving and processing reality. In CVPR 2019, 2019. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 423, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 423, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 317, + 423, + 553, + 510 + ], + "type": "text", + "content": "[21] Yifeng Ma, Suzhen Wang, Zhipeng Hu, Changjie Fan, Tangjie Lv, Yu Ding, Zhidong Deng, and Xin Yu. Styletalk: one-shot talking head generation with controllable speaking styles. In Proceedings of the Thirty-Seventh AAAI Conference on Artificial Intelligence and Thirty-Fifth Conference on Innovative Applications of Artificial Intelligence and Thirteenth Symposium on Educational Advances in Artificial Intelligence. AAAI Press, 2023. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 513, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 513, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 317, + 513, + 553, + 567 + ], + "type": "text", + "content": "[22] Yue Ma, Hongyu Liu, Hongfa Wang, Heng Pan, Yingqing He, Junkun Yuan, Ailing Zeng, Chengfei Cai, Heung-Yeung Shum, Wei Liu, et al. Follow-your-emoji: Fine-controllable and expressive freestyle portrait animation. arXiv preprint arXiv:2406.01900, 2024. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 568, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 568, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 317, + 568, + 553, + 633 + ], + "type": "text", + "content": "[23] Evonne Ng, Hanbyul Joo, Liwen Hu, Hao Li, Trevor Darrell, Angjoo Kanazawa, and Shiry Ginosar. Learning to listen: Modeling non-deterministic dyadic facial motion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20395-20405, 2022. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 635, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 553, + 689 + ], + "type": "text", + "content": "[24] Evonne Ng, Sanjay Subramanian, Dan Klein, Angjoo Kanazawa, Trevor Darrell, and Shiry Ginosar. Can language models learn to listen? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10083-10093, 2023. 1, 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[25] Evonne Ng, Javier Romero, Timur Bagautdinov, Shaojie Bai, Trevor Darrell, Angjoo Kanazawa, and Alexander Richard." + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10675" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "type": "text", + "content": "From audio to photoreal embodiment: Synthesizing humans in conversations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1001-1010, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 161 + ], + "type": "text", + "content": "[26] Se Jin Park, Chae Won Kim, Hyeongseop Rha, Minsu Kim, Joanna Hong, Jeong Hun Yeo, and Yong Man Ro. Let's go real talk: Spoken dialogue model for face-to-face conversation. arXiv preprint arXiv:2406.07867, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 162, + 296, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 296, + 227 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 296, + 227 + ], + "type": "text", + "content": "[27] K R Prajwal, Rudrabha Mukhopadhyay, Vinay P. Namboodiri, and C.V. Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM International Conference on Multimedia, page 484-492, New York, NY, USA, 2020. Association for Computing Machinery. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 228, + 295, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 228, + 295, + 260 + ], + "spans": [ + { + "bbox": [ + 57, + 228, + 295, + 260 + ], + "type": "text", + "content": "[28] Yurui Ren, Ge Li, Yuanqi Chen, Thomas H. Li, and Shan Liu. Pirenderer: Controllable portrait image generation via semantic neural rendering, 2021. 3, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 261, + 295, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 261, + 295, + 293 + ], + "spans": [ + { + "bbox": [ + 57, + 261, + 295, + 293 + ], + "type": "text", + "content": "[29] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 294, + 295, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 294, + 295, + 316 + ], + "spans": [ + { + "bbox": [ + 57, + 294, + 295, + 316 + ], + "type": "text", + "content": "[30] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv:2010.02502, 2020. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 317, + 295, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 317, + 295, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 317, + 295, + 370 + ], + "type": "text", + "content": "[31] Luchuan Song, Guojun Yin, Zhenchao Jin, Xiaoyi Dong, and Chenliang Xu. Emotional listener portrait: Realistic listener motion simulation in conversation. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 20782-20792. IEEE, 2023. 1, 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 372, + 295, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 372, + 295, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 372, + 295, + 426 + ], + "type": "text", + "content": "[32] Siyang Song, Micol Spitale, Cheng Luo, Cristina Palmero, German Barquero, Hengde Zhu, Sergio Escalera, Michel Valstar, Tobias Baur, Fabien Ringeval, Elisabeth Andre, and Hatice Gunes. React 2024: the second multiple appropriate facial reaction generation challenge, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 426, + 295, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 295, + 482 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 295, + 482 + ], + "type": "text", + "content": "[33] Michal Stypulkowski, Konstantinos Vougioukas, Sen He, Maciej Zieba, Stavros Petridis, and Maja Pantic. Diffused heads: Diffusion models beat gans on talking-face generation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5091-5100, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 483, + 295, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 483, + 295, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 483, + 295, + 525 + ], + "type": "text", + "content": "[34] Mingze Sun, Chao Xu, Xinyu Jiang, Yang Liu, Baigui Sun, and Ruqi Huang. Beyond talking—generating holistic 3d human dyadic motion for communication. arXiv preprint arXiv:2403.19467, 2024. 3, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 526, + 295, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 526, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 526, + 295, + 581 + ], + "type": "text", + "content": "[35] Zhiyao Sun, Tian Lv, Sheng Ye, Matthieu Lin, Jenny Sheng, Yu-Hui Wen, Minjing Yu, and Yong-jin Liu. Diffpositalk: Speech-driven stylistic 3d facial animation and head pose generation via diffusion models. ACM Transactions on Graphics (TOG), 43(4):1-9, 2024. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 582, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 295, + 635 + ], + "type": "text", + "content": "[36] Jiaxiang Tang, Kaisiyuan Wang, Hang Zhou, Xiaokang Chen, Dongliang He, Tianshu Hu, Jingtuo Liu, Gang Zeng, and Jingdong Wang. Real-time neural radiance talking portrait synthesis via audio-spatial decomposition. arXiv preprint arXiv:2211.12368, 2022. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 637, + 295, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 637, + 295, + 680 + ], + "spans": [ + { + "bbox": [ + 56, + 637, + 295, + 680 + ], + "type": "text", + "content": "[37] Linrui Tian, Qi Wang, Bang Zhang, and Liefeng Bo. Emo: Emote portrait alive-generating expressive portrait videos with audio2video diffusion model under weak conditions. arXiv preprint arXiv:2402.17485, 2024. 1, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "type": "text", + "content": "[38] Naftali Tishby, Fernando C Pereira, and William Bialek. The information bottleneck method. arXiv preprint physics/0004057, 2000. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "text", + "content": "[39] Minh Tran, Di Chang, Maksim Siniukov, and Mohammad Soleymani. Dyadic interaction modeling for social behavior generation. arXiv preprint arXiv:2403.09069, 2024. 1, 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 553, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 148 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 148 + ], + "type": "text", + "content": "[40] Aaron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu. Neural discrete representation learning, 2018. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 150, + 554, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 150, + 554, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 150, + 554, + 205 + ], + "type": "text", + "content": "[41] Cong Wang, Kuan Tian, Jun Zhang, Yonghang Guan, Feng Luo, Fei Shen, Zhiwei Jiang, Qing Gu, Xiao Han, and Wei Yang. V-express: Conditional dropout for progressive training of portrait video generation. arXiv preprint arXiv:2406.02511, 2024. 1, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 206, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 239 + ], + "type": "text", + "content": "[42] Duomin Wang, Bin Dai, Yu Deng, and Baoyuan Wang. Disentangling planning, driving and rendering for photorealistic avatar agents. 2023. 1, 2, 3, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 239, + 553, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 303 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 303 + ], + "type": "text", + "content": "[43] Duomin Wang, Yu Deng, Zixin Yin, Heung-Yeung Shum, and Baoyuan Wang. Progressive disentangled representation learning for fine-grained controllable talking head synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17979-17989, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 304, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 304, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 304, + 553, + 348 + ], + "type": "text", + "content": "[44] Yaohui Wang, Di Yang, Francois Bremond, and Antitza Dantcheva. Latent image animator: Learning to animate images via latent space navigation. arXiv preprint arXiv:2203.09043, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 350, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 403 + ], + "type": "text", + "content": "[45] Mingwang Xu, Hui Li, Qingkun Su, Hanlin Shang, Liwei Zhang, Ce Liu, Jingdong Wang, Luc Van Gool, Yao Yao, and Siyu Zhu. Hallo: Hierarchical audio-driven visual synthesis for portrait image animation. arXiv preprint arXiv:2406.08801, 2024. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 404, + 553, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 404, + 553, + 458 + ], + "spans": [ + { + "bbox": [ + 316, + 404, + 553, + 458 + ], + "type": "text", + "content": "[46] Sicheng Xu, Guojun Chen, Yu-Xiao Guo, Jiaolong Yang, Chong Li, Zhenyu Zang, Yizhong Zhang, Xin Tong, and Baining Guo. Vasa-1: Lifelike audio-driven talking faces generated in real time. arXiv preprint arXiv:2404.10667, 2024. 1, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 460, + 553, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 553, + 492 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 553, + 492 + ], + "type": "text", + "content": "[47] Yichao Yan, Zanwei Zhou, Zi Wang, Jingnan Gao, and Xiaokang Yang. Dialoguenerf: Towards realistic avatar face-to-face conversation video generation, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 493, + 553, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 493, + 553, + 548 + ], + "spans": [ + { + "bbox": [ + 316, + 493, + 553, + 548 + ], + "type": "text", + "content": "[48] Chenxu Zhang, Chao Wang, Jianfeng Zhang, Hongyi Xu, Guoxian Song, You Xie, Linjie Luo, Yapeng Tian, Xiaohu Guo, and Jiashi Feng. Dream-talk: diffusion-based realistic emotional audio-driven method for single image talking face generation. arXiv preprint arXiv:2312.13578, 2023. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 548, + 553, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 553, + 581 + ], + "type": "text", + "content": "[49] Longhao Zhang, Shuang Liang, Zhipeng Ge, and Tianshu Hu. Personatack: Bring attention to your persona in visual dubbing. arXiv preprint arXiv:2409.05379, 2024. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 582, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 553, + 613 + ], + "type": "text", + "content": "[50] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 615, + 553, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 615, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 316, + 615, + 553, + 646 + ], + "type": "text", + "content": "[51] Shifeng Zhang, Xiangyu Zhu, Zhen Lei, Hailin Shi, Xiaobo Wang, and Stan Z. Li. S" + }, + { + "bbox": [ + 316, + 615, + 553, + 646 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 316, + 615, + 553, + 646 + ], + "type": "text", + "content": "fd: Single shot scale-invariant face detector, 2017. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 647, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 713 + ], + "type": "text", + "content": "[52] Wenxuan Zhang, Xiaodong Cun, Xuan Wang, Yong Zhang, Xi Shen, Yu Guo, Ying Shan, and Fei Wang. Sadtalker: Learning realistic 3d motion coefficients for stylized audio-driven single image talking face animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8652-8661, 2023. 1, 7, 8" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10676" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 464 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 297, + 127 + ], + "type": "text", + "content": "[53] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661–3670, 2021. 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 295, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 295, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 295, + 184 + ], + "type": "text", + "content": "[54] Shengkui Zhao, Yukun Ma, Chongjia Ni, Chong Zhang, Hao Wang, Trung Hieu Nguyen, Kun Zhou, Jiaqi Yip, Dianwen Ng, and Bin Ma. Mossformer2: Combining transformer and rnnprecent network for enhanced time-domain monaural speech separation, 2023. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 295, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 295, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 295, + 228 + ], + "type": "text", + "content": "[55] Wei Zhao, Peng Xiao, Rongju Zhang, Yijun Wang, and Jianxin Lin. Semantic-aware responsive listener head synthesis. In Proceedings of the 30th ACM International Conference on Multimedia, pages 7065-7069, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 295, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 295, + 284 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 295, + 284 + ], + "type": "text", + "content": "[56] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking face generation by implicitly modularized audio-visual representation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 285, + 295, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 285, + 295, + 328 + ], + "spans": [ + { + "bbox": [ + 56, + 285, + 295, + 328 + ], + "type": "text", + "content": "[57] Mohan Zhou, Yalong Bai, Wei Zhang, Ting Yao, Tiejun Zhao, and Tao Mei. Vico-x: Multimodal conversation dataset. https://project.mhzhou.com/vico, 2022. Accessed: 2022-09-30. 2, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 331, + 295, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 295, + 384 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 295, + 384 + ], + "type": "text", + "content": "[58] Mohan Zhou, Yalong Bai, Wei Zhang, Ting Yao, Tiejun Zhao, and Tao Mei. Responsive listening head generation: A benchmark dataset and baseline. In Proceedings of the European conference on computer vision (ECCV), 2022. 2, 3, 5, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 386, + 295, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 386, + 295, + 419 + ], + "spans": [ + { + "bbox": [ + 56, + 386, + 295, + 419 + ], + "type": "text", + "content": "[59] Mohan Zhou, Yalong Bai, Wei Zhang, Ting Yao, and Tiejun Zhao. Interactive conversational head generation. arXiv preprint arXiv:2307.02090, 2023. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 420, + 295, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 420, + 295, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 420, + 295, + 464 + ], + "type": "text", + "content": "[60] Yang Zhou, Xintong Han, Eli Shechtman, Jose Echevarria, Evangelos Kalogerakis, and Dingzeyu Li. Makelttalk: speaker-aware talking-head animation. ACM Transactions on Graphics, 39(6):1-15, 2020. 2, 3" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10677" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/IRGS_ Inter-Reflective Gaussian Splatting with 2D Gaussian Ray Tracing/ff8cd318-c293-446d-b3d4-b22f9fc186ef_content_list.json b/2025/IRGS_ Inter-Reflective Gaussian Splatting with 2D Gaussian Ray Tracing/ff8cd318-c293-446d-b3d4-b22f9fc186ef_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9cf50282ca5537821bc9828248ba30864a027a48 --- /dev/null +++ b/2025/IRGS_ Inter-Reflective Gaussian Splatting with 2D Gaussian Ray Tracing/ff8cd318-c293-446d-b3d4-b22f9fc186ef_content_list.json @@ -0,0 +1,1491 @@ +[ + { + "type": "text", + "text": "IRGS: Inter-Reflective Gaussian Splating with 2D Gaussian Ray Tracing", + "text_level": 1, + "bbox": [ + 127, + 130, + 869, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chun Gu $^{1,2}$ Xiaofei Wei $^{1,2}$ Zixuan Zeng $^{1}$ Yuxuan Yao $^{1}$ Li Zhang $^{1,2*}$", + "bbox": [ + 192, + 179, + 807, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Data Science, Fudan University $^{2}$ Shanghai Innovation Institute", + "bbox": [ + 192, + 198, + 797, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://fudan-zvq.github.io/IRGS", + "bbox": [ + 303, + 224, + 684, + 241 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 276, + 326, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In inverse rendering, accurately modeling visibility and indirect radiance for incident light is essential for capturing secondary effects. Due to the absence of a powerful Gaussian ray tracer, previous 3DGS-based methods have either adopted a simplified rendering equation or used learnable parameters to approximate incident light, resulting in inaccurate material and lighting estimations. To this end, we introduce inter-reflective Gaussian splatting (IRGS) for inverse rendering. To capture inter-reflection, we apply the full rendering equation without simplification and compute incident radiance on the fly using the proposed differentiable 2D Gaussian ray tracing. Additionally, we present an efficient optimization scheme to handle the computational demands of Monte Carlo sampling for rendering equation evaluation. Furthermore, we introduce a novel strategy for querying the indirect radiance of incident light when relighting the optimized scenes. Extensive experiments on multiple standard benchmarks validate the effectiveness of IRGS, demonstrating its capability to accurately model complex inter-reflection effects.", + "bbox": [ + 89, + 308, + 483, + 609 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 638, + 220, + 655 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Inverse rendering is a fundamental problem in computer vision and graphics, aiming to reconstruct geometry and estimate material and lighting from a set of posed images. The introduction of neural radiance field (NeRF) [21], which uses a neural implicit field modeled by a multi-layer perceptron (MLP) to represent 3D scenes, has inspired many NeRF-based methods [5, 15, 20, 28, 40] to address inverse rendering tasks. Leveraging NeRF's ray-based rendering technique, these methods can accurately compute the visibility and indirect radiance for incident light as required by the rendering equation. However, the exhaustive querying of a neural field during ray marching and its limited representational capacity constrain the efficiency and performance of neural implicit fields in inverse rendering.", + "bbox": [ + 89, + 664, + 482, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/543abf89ba94525d5f3e15010162a76c61de896c8ac2469bb64c796c6d54566d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 531, + 275, + 890, + 450 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/49ec55b658c17a7f302914ba24b45925ca2558bdc2ce662d38fef04af8097f57.jpg", + "image_caption": [ + "Figure 1. Global and indirect illumination in a Gaussian-based scene using our IRGS, highlighting its inter-reflective capabilities." + ], + "image_footnote": [], + "bbox": [ + 531, + 450, + 890, + 628 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, 3D Gaussian splatting (3DGS) [16] has emerged as a promising technique for 3D scene representation, offering both remarkable rendering quality and efficiency. By modeling a static 3D scene as a collection of 3D Gaussians, 3DGS achieves real-time, high-quality rendering with a tile-based rasterizer. However, despite its strengths, 3DGS has notable limitations due to its reliance on rasterization, which limits its ability to accurately simulate ray-based effects that are essential for inverse rendering. As a result, some 3DGS-based inverse rendering methods [14, 19, 27] have adopted simplified versions of the rendering equation, such as the split-sum approximation [24]. Other approaches, such as R3DG [10], use learnable parameters to approximate indirect radiance, which fails to", + "bbox": [ + 511, + 688, + 906, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author (lizhangfd@fudan.edu.cn).", + "bbox": [ + 107, + 887, + 419, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "10943", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "capture actual inter-reflection. These methods struggle to accurately model the rendering equation, limiting their effectiveness in realistic material and lighting estimation.", + "bbox": [ + 89, + 90, + 480, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our goal is to integrate the full rendering equation into the efficient Gaussian splatting for accurate inverse rendering, enabling the capture of complex interreflection effects. This requires efficient and precise computation of visibility and indirect radiance for incident light. Drawing inspiration from 3DGRT [22], which achieves efficient ray tracing on 3D Gaussian primitives via a $k$ -buffer hits-based marching technique and hardware-accelerated NVIDIA OptiX interface [25], we explore adapting this technique to model inter-reflection within the Gaussian splatting framework.", + "bbox": [ + 88, + 138, + 480, + 290 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we introduce inter-reflective Gaussian splatting (IRGS), a novel inverse rendering framework that leverages the proposed 2D Gaussian ray tracing technique to model inter-reflections within Gaussian splatting. To the best of our knowledge, this is the first method to integrate the full rendering equation without simplification into Gaussian splatting, accurately capturing inter-reflection effects by querying incident lighting through ray tracing. First, we propose 2D Gaussian ray tracing to efficiently and accurately trace across 2D Gaussian primitives, computing the accumulated opacity and color of arbitrary rays. Unlike 3DGRT [22], our approach features well-defined ray-splat intersections, enabling direct ray tracing on pretrained 2D Gaussian splatting [13] with minimal quality degradation. Next, we introduce the inter-reflective Gaussian splatting framework: starting from a pretrained 2D Gaussian splatting model, we apply the full rendering equation without simplification at intersection points determined by the depth map, while visibility and indirect radiance are computed on the fly using our 2D Gaussian ray tracing technique. Notably, our 2D Gaussian ray tracing is fully differentiable, enabling the optimization of indirect light through backpropagation. We further propose an efficient optimization scheme to manage the computational demands of the exhaustive Monte Carlo sampling for the rendering equation. Additionally, we introduce a novel strategy to query indirect radiance during relighting, overcoming limitations in previous methods that often omit this component. Extensive experiments on multiple standard benchmarks confirm the effectiveness of IRGS. In Fig. 1, we visualize the global illumination (considering both direct and indirect radiance) and the indirect illumination (considering only indirect radiance) of a Gaussian-based scene rendered using IRGS, demonstrating that IRGS is capable of modeling remarkable inter-reflection effects.", + "bbox": [ + 91, + 291, + 483, + 821 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The contributions of this paper are as follows: (i) 2D Gaussian ray tracing, a technique that enables direct ray tracing on pretrained 2D Gaussian splatting with minimal quality degradation; (ii) IRGS, an inverse rendering framework that incorporates the full rendering equation without", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "simplification, using 2D Gaussian ray tracing to compute incident radiance on the fly; (iii) An optimization scheme that manages computational demands for rendering equation evaluation; (iv) A novel strategy that enables querying of indirect radiance during relighting.", + "bbox": [ + 511, + 90, + 906, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 513, + 178, + 651, + 194 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Novel view synthesis aims to leverage a limited collection of observed images of a scene to generate new images from an unseen viewpoint. Neural radiance field (NeRF) [21] produces photo-realistic quality images by representing the scenes in terms of a volumetric radiance field encoded in a coordinate-based neural network. Subsequent advancements have focused on accelerating rendering[23, 26], enhancing image quality [2-4], and improving geometry reconstruction [18, 29, 38]. More recently, 3D Gaussian splatting (3DGS) [16] introduced the use of fuzzy, anisotropic 3D Gaussian point cloud for scene representation along with an efficient tile-based rasterizer, achieving state-of-the-art results in both rendering quality and efficiency. 3DGS has inspired numerous approaches targeting various goals, including geometry reconstruction [13, 39], dynamic scene reconstruction [34, 35], inverse rendering [10, 19], and street scene applications [8, 32]. However, the dependence on rasterization limits Gaussian splatting's ability to simulate ray-based effects accurately. To address this, 3DGRT [22] proposes a differentiable Gaussian ray tracer that efficiently performs ray tracing across 3D Gaussian primitives, allowing radiance computation along arbitrary rays. In this paper, we propose ray tracing on 2D Gaussian primitives. Unlike 3DGRT, 2D Gaussian primitives [13] have a well-defined ray-splat intersection, enabling direct ray tracing on pre-trained 2D Gaussian splatting with minimal quality degradation.", + "bbox": [ + 511, + 203, + 906, + 611 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inverse rendering focuses on recovering geometry, material, and lighting properties from captured images, presenting a complex challenge due to high-dimensional light-surface interactions. Inspired by the success of neural radiance fields, many methods [1, 5, 6, 15, 20, 28, 33, 36, 40, 43] leverage flexible ray marching techniques and neural implicit fields for inverse rendering. Recent approaches [9-11, 14, 19, 27, 31, 37] have adapted 3D Gaussian splatting (3DGS) to inverse rendering by assigning material-related properties to each Gaussian primitive. Leveraging the rapid rendering capabilities of 3DGS, these methods achieve much higher efficiency compared to NeRF-based techniques. However, due to the lack of an efficient Gaussian ray tracer, existing 3DGS-based methods either apply a simplified version of the rendering equation [19] or use learnable parameters to model incident light [10], which limits the accuracy of material and lighting estimation. For example, GS-IR [19] employs split-sum approximation to", + "bbox": [ + 511, + 628, + 908, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10944", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4d68f25c23fae156617de34e6486be3db749414b84b377e9696048edad14b68a.jpg", + "image_caption": [ + "Figure 2. Schematic illustration of the proposed IRGS. Starting from a set of 2D Gaussians equipped with material properties, we apply rasterization to generate albedo, roughness, position, and normal maps. We then evaluate the rendering equation using stratified sampling at the corresponding position, drawing geometry and material values from these feature maps. The radiance of incident light is decomposed into direct radiance $L_{\\mathrm{dir}}$ from the environment map, and indirect radiance $L_{\\mathrm{ind}}$ with visibility $V$ , obtained via 2D Gaussian ray tracing." + ], + "image_footnote": [], + "bbox": [ + 101, + 88, + 903, + 239 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "simplify the rendering equation and uses baked volumes to store occlusion and indirect illumination. R3DG [10] shades each Gaussian individually, using point-based ray tracing to bake occlusion and spherical harmonics to parameterize indirect radiance. To overcome these limitations, we propose the IRGS framework that utilizes 2D Gaussian ray tracing to accurately model inter-reflections.", + "bbox": [ + 88, + 320, + 483, + 428 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 440, + 181, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce our inter-reflective Gaussian splatting with 2D Gaussian ray tracing, designed to reconstruct geometry and estimate material and lighting from a set of posed RGB images. We begin by providing the necessary background on 3D Gaussian splatting and the rendering equation (Section 3.1). We then present the 2D Gaussian ray tracing technique, which performs ray tracing on 2D Gaussian primitives (Section 3.2). Finally, we outline our IRGS framework, which leverages the proposed 2D Gaussian ray tracing technique to query visibility and indirect radiance for incident light (Section 3.3).", + "bbox": [ + 88, + 465, + 483, + 632 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 89, + 641, + 220, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Gaussian splatting 3D Gaussian splatting (3DGS) represents a 3D scene using 3D Gaussian primitives, each defined as an ellipsoid characterized by a 3D mean $\\mu$ and a covariance matrix $\\Sigma$ . The influence of a Gaussian at a spatial position $x$ is expressed as:", + "bbox": [ + 89, + 664, + 483, + 742 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nG (\\boldsymbol {x}) = \\exp \\left(- \\frac {1}{2} (\\boldsymbol {x} - \\boldsymbol {\\mu}) ^ {\\top} \\Sigma^ {- 1} (\\boldsymbol {x} - \\boldsymbol {\\mu})\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 751, + 482, + 785 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Each Gaussian is also assigned an opacity $o$ and a view-dependent color $c$ modeled by spherical harmonics (SH). The rendering process of 3DGS starts by projecting 3D Gaussian primitives onto the 2D image plane. This is achieved by first applying a viewing transformation $W$ and then performing a perspective projection. The covariance matrix of the resulting 2D Gaussians can be approximated", + "bbox": [ + 89, + 794, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "as $\\Sigma' = JW\\Sigma W^\\top J^\\top$ , where $J$ represents the Jacobian matrix of the perspective projection. The final image is rendered by alpha-blending these projected 2D Gaussians from front to back:", + "bbox": [ + 511, + 319, + 906, + 380 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {C} = \\sum_ {i = 1} ^ {N} T _ {i} \\alpha_ {i} \\boldsymbol {c} _ {i}, \\quad T _ {i} = \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 388, + 906, + 433 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\alpha_{i}$ is calculated by multiplying opacity $o_i$ with each Gaussian's influence, determined by $\\Sigma^{\\prime}$ and the pixel coordinate. The covariance matrix $\\Sigma$ is parameterized by a unit quaternion and a scaling vector to facilitate optimization.", + "bbox": [ + 511, + 441, + 906, + 503 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Rendering equation In physically-based rendering, illumination at a given surface point $\\pmb{x}$ is computed via the rendering equation:", + "bbox": [ + 511, + 521, + 905, + 566 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {o} \\left(\\boldsymbol {\\omega} _ {o}, \\boldsymbol {x}\\right) = \\int_ {\\Omega} f \\left(\\boldsymbol {\\omega} _ {o}, \\boldsymbol {\\omega} _ {i}, \\boldsymbol {x}\\right) L _ {1} \\left(\\boldsymbol {\\omega} _ {i}, \\boldsymbol {x}\\right) \\left(\\boldsymbol {\\omega} _ {i} \\cdot \\mathbf {n}\\right) d \\boldsymbol {\\omega} _ {i}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 577, + 903, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\omega_{o}$ is the outgoing direction, $\\omega_{i}$ is the incident direction over the hemisphere defined by the surface normal $\\pmb{n}$ , $f$ is the bidirectional reflectance distribution function (BRDF), and $L_{\\mathrm{i}}(\\boldsymbol{\\omega}_i,\\boldsymbol{x})$ is the radiance of the incident light. In this work, we use a simplified Disney BRDF model [7] with only diffuse albedo $\\pmb{a}$ and roughness $r$ as parameters, and assume the material is dielectric. The BRDF $f$ in Eq. (3) can be separated into a diffuse term $f_{d} = \\frac{a}{\\pi}$ and a specular term $f_{s}$ :", + "bbox": [ + 511, + 619, + 906, + 755 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf _ {s} \\left(\\boldsymbol {\\omega} _ {i}, \\boldsymbol {\\omega} _ {o}, \\boldsymbol {x}\\right) = \\frac {D F G}{4 \\left(\\boldsymbol {\\omega} _ {i} \\cdot \\mathbf {n}\\right) \\left(\\boldsymbol {\\omega} _ {o} \\cdot \\mathbf {n}\\right)}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 763, + 903, + 799 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $D, F$ , and $G$ represent the normal distribution function, the Fresnel term, and the geometry term, respectively.", + "bbox": [ + 511, + 806, + 903, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. 2D Gaussian ray tracing", + "text_level": 1, + "bbox": [ + 511, + 847, + 738, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While 3DGS [16] has been successful in applications such as 3D reconstruction, achieving superior rendering quality", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10945", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/46528fd0fe2cedbaf808364828dc2395595e7e26f092daded8dc70456a9a8ee2.jpg", + "image_caption": [ + "Figure 3. Performance of directly applying Gaussian ray tracing on a pretrained Gaussian splatting checkpoint in 3D Gaussian and 2D Gaussian cases, respectively. Our proposed 2DGRT achieves significantly less quality degradation than 3DGRT [22] in both quantitative metrics and visual results." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 472, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and efficiency compared to NeRF-based methods, it is limited by the constraints of rasterization. These limitations are primarily related to ray-based effects, including rendering with highly distorted cameras and handling secondary ray effects. To address these challenges, 3DGRT [22] proposes performing ray tracing across 3D Gaussian primitives. This approach overcomes the inefficiencies of ray tracing on 3D Gaussian primitives by implementing a $k$ -buffer hits-based marching technique with a hardware-accelerated NVIDIA OptiX interface [25]. In this paper, as our focus is on inverse rendering, performing ray tracing on Gaussian primitives is essential to accurately compute inter-reflection effects when evaluating the rendering equation, including visibility and indirect radiance of incident light.", + "bbox": [ + 88, + 289, + 482, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3D Gaussian ray tracing [22] performs ray tracing on 3D Gaussian primitives and determines the ray-splat intersection as the point of maximum response from the corresponding 3D Gaussian along the ray. This behavior differs from that of 3D Gaussian splatting, which computes the particle response in 2D image space, leaving the position of the 3D intersection point undefined. To leverage both the efficiency of Gaussian splatting and the flexible rendering capabilities of Gaussian ray tracing, we aim to perform ray tracing directly on the Gaussian splatting model for incident radiance inference. However, as shown in Fig. 3, applying ray tracing directly on a 3DGS checkpoint results in a noticeable decline in rendering quality compared to splatting. To address this issue, we propose 2D Gaussian ray tracing (2DGRT), which performs ray tracing on 2D Gaussian primitives [13]. Since 2D Gaussian disks have a well-defined ray-splat intersection, this approach eliminates inconsistencies in intersection points present in 3D Gaussians.", + "bbox": [ + 89, + 503, + 482, + 775 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a 2D Gaussian [13] defined by its center point $\\pmb{\\mu} \\in \\mathbb{R}^3$ , an opacity parameter $o \\in [0,1]$ , two principal tangential vectors $\\pmb{t}_u \\in \\mathbb{R}^3$ and $\\pmb{t}_v \\in \\mathbb{R}^3$ , and a scaling vector $\\pmb{s} = (s_u, s_v) \\in \\mathbb{R}^2$ , we follow 3DGRT [22] to fit a bounding proxy of each 2D Gaussian using an adaptive icosahedron mesh, ensuring that a minimum influence of $\\alpha_{\\mathrm{min}}$ is enclosed. The vertices $\\pmb{v}$ of the bounding icosahedron mesh are obtained by transforming a regular icosahedron with a", + "bbox": [ + 89, + 779, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "unit inner sphere:", + "bbox": [ + 513, + 92, + 633, + 106 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {v} \\leftarrow \\sqrt {2 \\log \\left(o / \\alpha_ {\\min }\\right)} \\left(s _ {u} \\boldsymbol {t} _ {u} \\quad s _ {v} \\boldsymbol {t} _ {v} \\quad \\epsilon \\mathbf {1}\\right) \\boldsymbol {v} + \\boldsymbol {\\mu}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 114, + 903, + 132 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\epsilon$ denotes a small positive number. Each icosahedron mesh consists of 20 triangular faces, resulting in a total of $20N$ triangles for $N$ 2D Gaussian primitives. To leverage the hardware-accelerated ray-triangle intersection provided by OptiX [25], we construct a Bounding Volume Hierarchy (BVH) using these triangles. By applying the $k$ -buffer per-ray sorting algorithm [22], we obtain the exact ordering of intersected Gaussians for each ray. To ensure consistency with 2DGS [13], we analytically compute the ray-splat intersection point $p$ :", + "bbox": [ + 511, + 140, + 906, + 292 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {p} = \\boldsymbol {r} _ {o} + \\tau \\boldsymbol {r} _ {d}, \\quad \\text {w h e r e} \\quad \\tau = \\frac {\\boldsymbol {n} ^ {\\top} (\\boldsymbol {\\mu} - \\boldsymbol {r} _ {o})}{\\boldsymbol {n} ^ {\\top} \\boldsymbol {r} _ {d}}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 297, + 903, + 333 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\pmb{n} = \\pmb{t}_u\\times \\pmb{t}_v$ is the normal vector of the 2D Gaussian disk, and $\\pmb{r}_o$ and $\\pmb{r}_d$ represent the origin and direction of the given ray, respectively. Thus, the influence of the current 2D Gaussian primitive is given by:", + "bbox": [ + 511, + 338, + 905, + 398 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nG (\\boldsymbol {p}) = \\exp \\left(- \\frac {u ^ {2} + v ^ {2}}{2}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 406, + 903, + 440 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $u = \\frac{1}{s_v} (\\pmb {p} - \\pmb {\\mu})^\\top \\pmb {t}_u,v = \\frac{1}{s_v} (\\pmb {p} - \\pmb {\\mu})^\\top \\pmb {t}_v$", + "bbox": [ + 511, + 446, + 830, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Consequently, the RGB value and accumulated opacity value of the ray can be obtained by alpha-blending (Eq. (2)) the intersected 2D Gaussians from front to back:", + "bbox": [ + 511, + 464, + 905, + 508 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left(\\boldsymbol {c} _ {\\mathrm {r t}}, o _ {\\mathrm {r t}}\\right) \\leftarrow \\operatorname {T r a c e} \\left(\\boldsymbol {r} _ {o}, \\boldsymbol {r} _ {d}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 518, + 903, + 534 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the RGB value of each Gaussian is modeled by the view-dependent color $c$ as in vanilla 2D Gaussian splatting. It is worth noting that subtle differences remain between 2D Gaussian ray tracing and 2DGS due to sorting errors introduced by per-tile sorting in 2DGS. However, as shown in Fig. 3, the quality degradation from 2D Gaussian ray tracing is minimal and significantly less noticeable than that observed in 3D Gaussians.", + "bbox": [ + 511, + 542, + 906, + 662 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Inter-reflective Gaussian splatting", + "text_level": 1, + "bbox": [ + 511, + 671, + 808, + 688 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.1 Stage I: 2D Gaussian pretraining", + "text_level": 1, + "bbox": [ + 511, + 695, + 797, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Before estimating material and lighting for inverse rendering, we first pretrain a standard 2D Gaussian splatting model. This step is essential, as reliable geometry is required for inverse rendering, and evaluating the rendering equation is computationally intensive. Using a pretrained 2D Gaussian splatting model can greatly accelerate the training process. Along with the RGB rendering described in Eq. (2), we also render normalized depth and normal maps as follows:", + "bbox": [ + 511, + 718, + 905, + 854 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\mathcal {D}, \\mathcal {N} \\right\\} = \\sum_ {i = 1} ^ {N} w _ {i} \\left\\{d _ {i}, \\boldsymbol {n} _ {i} \\right\\}, \\text {w h e r e} w _ {i} = \\frac {T _ {i} \\alpha_ {i}}{\\sum_ {i = 1} ^ {N} T _ {i} \\alpha_ {i}}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 862, + 903, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "10946", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training loss Following 2DGS [13], we apply a normal consistency loss $\\mathcal{L}_{\\mathrm{n}}$ to ensure alignment between the rendered normal map and the underlying geometry, as well as a depth distortion loss [13] $\\mathcal{L}_{\\mathrm{d}}$ to encourage concentration of the 2D Gaussians. Additionally, we introduce an edge-aware smoothness loss [10, 12] $\\mathcal{L}_{\\mathrm{s,n}}$ on the normal map to enhance geometric robustness:", + "bbox": [ + 89, + 90, + 483, + 196 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {s}, \\mathrm {n}} = \\| \\nabla \\mathcal {N} \\| \\exp (- \\| \\nabla \\mathcal {C} _ {\\mathrm {g t}} \\|), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 209, + 482, + 227 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{C}_{\\mathrm{gt}}$ is the ground truth training image. To address the presence of floaters during standard 2DGS training, we also incorporate a binary cross-entropy loss [29] to constrain the geometry using the provided object mask $\\mathcal{M}$ :", + "bbox": [ + 89, + 238, + 483, + 299 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {o}} = - \\mathcal {M} \\log \\mathcal {O} - (1 - \\mathcal {M}) \\log (1 - \\mathcal {O}), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 311, + 482, + 329 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{O} = \\sum_{i=1}^{N} T_i \\alpha_i$ , denotes the accumulated opacity map. Combined with the RGB reconstruction loss $\\mathcal{L}_{\\mathrm{c}}$ from 3DGS [16], the training loss for the first stage is defined as:", + "bbox": [ + 89, + 340, + 483, + 388 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {1} = \\mathcal {L} _ {\\mathrm {c}} + \\lambda_ {\\mathrm {n}} \\mathcal {L} _ {\\mathrm {n}} + \\lambda_ {\\mathrm {d}} \\mathcal {L} _ {\\mathrm {d}} + \\lambda_ {\\mathrm {s}, \\mathrm {n}} \\mathcal {L} _ {\\mathrm {s}, \\mathrm {n}} + \\lambda_ {\\mathrm {o}} \\mathcal {L} _ {\\mathrm {o}}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 415, + 482, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.2 Stage II: inverse rendering", + "text_level": 1, + "bbox": [ + 89, + 452, + 331, + 467 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The order of shading and rasterization is crucial in physically based rendering with Gaussian splatting. Previous works, such as R3DG [10], propose performing shading on each Gaussian, taking advantage of properties modeled on each Gaussian, such as indirect radiance. However, since the rendered normal map is supervised at the pixel level after rasterization, shading on individual Gaussians can lead to inaccurate and blurred results. Consequently, we propose applying the rendering equation after rasterization.", + "bbox": [ + 89, + 476, + 483, + 611 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Rasterization As introduced in Section 3.1, a dielectric BRDF model is used for the rendering equation. We assign each 2D Gaussian primitive additional material properties: diffuse albedo $\\pmb{a}$ and roughness $r$ . Pixel-level albedo map $\\mathcal{A}$ and roughness map $\\mathcal{R}$ are obtained through alpha-blending during rasterization: $\\{\\mathcal{A},\\mathcal{R}\\} = \\sum_{i = 1}^{N}w_{i}\\{a_{i},r_{i}\\}$ .", + "bbox": [ + 89, + 613, + 483, + 704 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Previous inverse rendering methods using Gaussian splatting have lacked an efficient way to query the visibility and radiance for arbitrary incident directions. As a result, these methods typically apply simplified versions of the rendering equation [19] (e.g., split-sum approximation [24]) or use learnable parameters to model indirect light [10]. In this paper, we leverage the proposed 2D Gaussian ray tracing technique to efficiently and accurately query inter-reflection effects on the fly, including the visibility and indirect radiance of incident light. This allows us to apply the full rendering equation without simplification.", + "bbox": [ + 89, + 705, + 483, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Light parametrization We decompose the incident radiance at surface point $\\pmb{x}$ with direction $\\omega_{i}$ in Eq. (3) into", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "direct radiance from infinity and indirect radiance:", + "bbox": [ + 511, + 90, + 846, + 106 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {i}} \\left(\\boldsymbol {\\omega} _ {i}, \\boldsymbol {x}\\right) = V \\left(\\boldsymbol {\\omega} _ {i}, \\boldsymbol {x}\\right) L _ {\\mathrm {d i r}} \\left(\\boldsymbol {\\omega} _ {i}\\right) + L _ {\\mathrm {i n d}} \\left(\\boldsymbol {\\omega} _ {i}, \\boldsymbol {x}\\right), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 113, + 903, + 130 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $L_{\\mathrm{dir}}(\\omega_i)$ is modeled using an environment cubemap, and $V(\\omega_i, \\boldsymbol{x})$ along with $L_{\\mathrm{ind}}(\\omega_i, \\boldsymbol{x})$ are queried using 2D Gaussian ray tracing:", + "bbox": [ + 511, + 136, + 903, + 181 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left(L _ {\\text {i n d}} \\left(\\boldsymbol {\\omega} _ {i}, \\boldsymbol {x}\\right), 1 - V \\left(\\boldsymbol {\\omega} _ {i}, \\boldsymbol {x}\\right)\\right) \\leftarrow \\operatorname {T r a c e} \\left(\\boldsymbol {x}, \\boldsymbol {\\omega} _ {i}\\right) \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 544, + 189, + 903, + 205 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "It is important to note that the RGB values used in ray tracing correspond to the view-dependent color $c$ from the first stage. Additionally, our 2D Gaussian ray tracing approach is fully differentiable, allowing gradients to propagate through the ray tracing process to optimize the indirect radiance of the incident ray.", + "bbox": [ + 511, + 212, + 903, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Rendering From the rendered depth and normal maps, we can determine the surface position $\\mathbf{x}$ and the corresponding surface normal vector $\\mathbf{n}$ for each pixel coordinate. Monte Carlo sampling is used to evaluate the rendering equation. Since we focus on reconstructing objects with dielectric materials, we employ stratified sampling to uniformly sample $N_{\\mathrm{r}}$ incident directions across the hemisphere. The final rendering result is then computed as:", + "bbox": [ + 511, + 303, + 905, + 424 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {c} _ {\\mathrm {p b r}} = \\frac {2 \\pi}{N _ {\\mathrm {r}}} \\sum_ {i = 1} ^ {N _ {\\mathrm {r}}} \\left(f _ {\\mathrm {d}} + f _ {\\mathrm {s}}\\right) L _ {\\mathrm {i}} \\left(\\boldsymbol {\\omega} _ {i}, \\boldsymbol {x}\\right) \\left(\\boldsymbol {\\omega} _ {i} \\cdot \\mathbf {n}\\right), \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 430, + 903, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training loss In addition to the training losses from the first stage, we apply regularization to facilitate material and lighting estimation. Given the prior that diffuse light $L_{\\mathrm{diffuse}} = \\frac{1}{N_{\\mathrm{r}}} \\sum_{i=1}^{N_{\\mathrm{r}}} L(\\boldsymbol{\\omega}_i, \\boldsymbol{x})$ exhibits a natural white radiance, we minimize the difference between the RGB values [20]:", + "bbox": [ + 511, + 477, + 905, + 566 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {l i g h t}} = \\sum_ {c} \\left(L _ {\\text {d i f f u s e}} - \\frac {1}{3} \\sum_ {c} L _ {\\text {d i f f u s e}}\\right), c \\in \\{R, G, B \\}. \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 573, + 903, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Similar to Eq. (10), we also apply edge-aware smoothness regularization $\\mathcal{L}_{\\mathrm{s,a}}$ and $\\mathcal{L}_{\\mathrm{s,r}}$ to the rendered albedo and roughness map, respectively, to ensure smooth material estimation. Additionally, for training efficiency, we apply the rendering equation to only a subset of pixels in each iteration. We set a maximum of $N_{\\mathrm{rays}}$ rays per iteration, meaning only $\\lfloor N_{\\mathrm{rays}} / N_{\\mathrm{r}}\\rfloor$ pixels are randomly selected for final rendering equation evaluation. This scheme allows us to use a large $N_{\\mathrm{r}}$ when evaluating the rendering equation, significantly enhancing estimation quality. Consequently, only a L1 loss $\\mathcal{L}_1^{\\mathrm{pbr}}$ is used to supervise the final color $c_{\\mathrm{pbr}}$ . It is important to note that the rendered image $\\mathcal{C}$ from the vanilla 2DGS is still optimized, while the view-dependent color $c$ associated with each Gaussian is used to compute radiance in 2D Gaussian ray tracing. The total loss for the second stage is given by:", + "bbox": [ + 511, + 621, + 906, + 863 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {2} = \\mathcal {L} ^ {1} + \\lambda_ {1} ^ {\\mathrm {p b r}} \\mathcal {L} _ {1} ^ {\\mathrm {p b r}} + \\lambda_ {\\text {l i g h t}} \\mathcal {L} _ {\\text {l i g h t}} + \\lambda_ {\\mathrm {s}, \\mathrm {a}} \\mathcal {L} _ {\\mathrm {s}, \\mathrm {a}} + \\lambda_ {\\mathrm {s}, \\mathrm {r}} \\mathcal {L} _ {\\mathrm {s}, \\mathrm {r}}, \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 867, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "10947", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a24d95e65a3050f9c226f23018eff22a373148a8f331dc6379bc25ab0adc6d14.jpg", + "image_caption": [ + "Figure 4. Qualitative comparison of NVS, material and lighting estimation, and relighting results on the Synthetic4Relight dataset [43]." + ], + "image_footnote": [], + "bbox": [ + 99, + 88, + 898, + 406 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/21db21ba55210412e0948f1e2f233a0f12c31ea98417857c80f52f12edc93a6a.jpg", + "table_caption": [ + "Table 1. Quantitative comparison on the Synthetic4Relight dataset [43]. A higher intensity of the red color signifies a better result." + ], + "table_footnote": [], + "table_body": "
Novel view synthesisRelightingAlbedoRoughness MSE↓Time (hours)
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
NeRFactor [42]22.800.9160.15021.540.8750.17119.490.8640.206N/A>48
InvRender [43]30.740.9530.08628.670.9500.09128.280.9350.0720.00814
TensorIR [15]35.800.9780.04929.690.9510.07930.580.9460.0650.0153
GS-IR [19]33.950.9650.05725.400.9240.08319.480.8960.1170.0110.4
R3DG [10]36.800.9820.02831.000.9640.05028.310.9510.0580.0130.9
IRGS (Ours)35.480.9740.04334.900.9690.04830.810.9570.0550.0080.7
", + "bbox": [ + 96, + 469, + 901, + 590 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Relighting When performing relighting, changes in environmental lighting make the optimized radiance $c$ of each Gaussian no longer applicable for indirect radiance in 2D Gaussian ray tracing. Unlike previous methods [10, 19] that omit indirect lighting during relighting, we propose using the split-sum approximation [24] to efficiently query indirect radiance without recursive evaluation. Specifically, we apply ray tracing on each incident ray to alpha-blend the albedo, roughness, and normal values, resulting in $\\mathcal{A}_{\\mathrm{r}}$ , $\\mathcal{R}_{\\mathrm{r}}$ , and $\\mathcal{N}_{\\mathrm{r}}$ . These aggregated values are then shaded using a pre-filtered environment map, following the approach in Nvdiffrec [24]. Additionally, we apply importance sampling on the intensity of the ground truth environment map instead of stratified sampling, enabling more accurate rendering equation evaluation by concentrating samples in high-intensity directions. Please refer to the appendix for further details.", + "bbox": [ + 89, + 614, + 485, + 875 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 513, + 614, + 638, + 633 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets and metrics To validate the effectiveness of IRGS, we conduct extensive experiments on multiple datasets for the inverse rendering task, including two synthetic datasets: Synthetic4Relight dataset [43] and the TensoIR dataset [15], as well as a real-world dataset: Stanford-ORB dataset [17]. For quantitative comparisons, we follow previous works [10, 19] and employ various metrics to evaluate performance across different outputs. Specifically, we use PSNR, SSIM [30], and LPIPS [41] for assessing image quality in novel view synthesis, albedo, and relighting tasks. For normal and roughness, we use mean angular error (MAE) and mean squared error (MSE), respectively.", + "bbox": [ + 511, + 641, + 906, + 823 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details As detailed in Section 3.3, our training process consists of two stages. The first stage is trained over 40,000 iterations with the hyperparameters $\\lambda_{\\mathrm{n}}$ , $\\lambda_{\\mathrm{d}}$ , $\\lambda_{\\mathrm{s,n}}$ , and $\\lambda_{\\mathrm{o}}$ set to 0.05, 1000, 0.02, and 0.01, respectively. The second stage continues for an additional 20,000", + "bbox": [ + 511, + 824, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "10948", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/eef7f716dce66220e89a9df2cd24762582a2adc67a534649135886c79fa6988e.jpg", + "image_caption": [ + "Figure 5. Visualization of estimated components in incident light, including the averaged direct radiance $L_{\\mathrm{dir}}$ , indirect radiance $L_{\\mathrm{ind}}$ , visibility $V$ (ambient occlusion), incident radiance $L_{\\mathrm{i}}$ . Compared to R3DG [10], IRGS achieves notably more realistic results, especially in estimated indirect light, due to its accurate modeling of inter-reflections." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 903, + 250 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6e490e3a8e5d7b425dd3ab2b7e03202bdf97c329ee11b22934bead54b94e1fa9.jpg", + "image_caption": [ + "Figure 6. Qualitative comparison of relighting results on TensoIR dataset [43]." + ], + "image_footnote": [], + "bbox": [ + 107, + 303, + 888, + 462 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ae8e944288f59cde34fdc7b5aaa7351e371cd29473ab445c5602aa4eaf212f76.jpg", + "table_caption": [ + "Table 2. Quantitative comparison on TensoIR dataset [15]." + ], + "table_footnote": [], + "table_body": "
MethodNormal MAE↓NVS PSNR↑Albedo PSNR↑Relight PSNR↑
NeRFactor [42]6.31424.67925.12523.383
InvRender [43]5.07427.36727.34123.973
TensoIR [15]4.10035.08829.27528.580
GS-IR [19]4.94835.33330.28624.374
R3DG [10]5.92737.34326.19927.367
IRGS (Ours)4.11235.87633.79629.907
", + "bbox": [ + 96, + 503, + 473, + 604 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "iterations, with $\\lambda_{1}^{\\mathrm{pbr}}$ , $\\lambda_{\\mathrm{light}}$ , $\\lambda_{\\mathrm{s,a}}$ , and $\\lambda_{\\mathrm{s,r}}$ set to 1.0, 0.01, 2.0, and 2.0, respectively. We implement 2D Gaussian ray tracing in OptiX [25] via PyTorch CUDA extensions, updating the BVH at each training iteration for the varying geometry, with each update taking approximately 3 ms. In line with 3DGRT [22], we use a buffer of $k = 16$ for per-ray sorting and terminate ray tracing when transmittance falls below 0.03. We employ $N_{\\mathrm{r}} = 256$ rays for rendering equation evaluation, and each training iteration samples up to $N_{\\mathrm{rays}} = 2^{18}$ rays. At $N_{\\mathrm{r}} = 256$ , rendering a complete image takes about 1 second. For environmental lighting, we use a $32 \\times 32$ cubemap during optimization. The learning rates for albedo, roughness, and the cubemap are set to 0.005, 0.005, and 0.01, respectively, with other hyperparameters following those in 2DGS [13]. The entire training process takes around 40 minutes, with 15 minutes for the first stage and 25 minutes for the second stage. All experiments are conducted on a single NVIDIA RTX 3090 GPU.", + "bbox": [ + 89, + 627, + 485, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1. Results", + "text_level": 1, + "bbox": [ + 513, + 489, + 607, + 503 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Synthetic4Relight For experiments on Synthetic4Relight [43], we evaluate IRGS's performance in novel view synthesis (NVS), relighting, material estimation, and efficiency. In Tab. 1, our approach achieves higher quality than previous methods across relighting, and material estimation, demonstrating IRGS's ability to accurately capture material details. While R3DG [10] achieves higher performance in NVS, this is primarily due to its shading on Gaussians, which enhances NVS capability at the expense of relighting performance. IRGS also completes optimization in a relatively short time of 0.7 hours, emphasizing its efficiency in inverse rendering applications. In Fig. 4, we provide a qualitative comparison of the reconstructed \"chair\" scene against GS-based competitors [10, 19], visualizing NVS, albedo, roughness, estimated environment maps, and relighting images. Our results demonstrate high-fidelity material estimation and realistic relighting outcomes. Fig. 5 further illustrates a comparison of the estimated incident light components between IRGS and R3DG [10], covering direct radiance, indirect radiance, ambient occlusion, and incident radiance. IRGS achieves more reasonable indirect radiance and clearer ambient occlusion, enhancing the overall realism.", + "bbox": [ + 511, + 517, + 906, + 866 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "TensoIR For TensoIR dataset [15], we report metrics for NVS, relighting, and albedo estimation in Tab. 2. Com", + "bbox": [ + 511, + 869, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "10949", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/656eec5ab206eec445bcdcd4a06138af8bab875376cb8a8778f0986ac1391988.jpg", + "image_caption": [ + "Figure 7. Relighting results on Stanford-ORB dataset [17]." + ], + "image_footnote": [], + "bbox": [ + 99, + 93, + 465, + 316 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/989f5d0a6f4e3c99ed98ade289bb9bb0a6b368da7a327fa1792f9cc6bcce82d0.jpg", + "image_caption": [ + "(a) Ablation studies on indirect light modeling." + ], + "image_footnote": [], + "bbox": [ + 101, + 338, + 472, + 479 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/293ec42bb39ff67cc490758100f568f4d4227e0eab89e9e74b91651455e471be.jpg", + "image_caption": [ + "(b) Ablation studies on number of sampled rays $N_{\\mathrm{r}}$", + "Figure 8. Ablation studies on various components of IRGS." + ], + "image_footnote": [], + "bbox": [ + 102, + 500, + 468, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "pared to prior arts [10, 15, 19, 42, 43], our method consistently outperforms in relighting and albedo estimation, and achieving comparable performance in normal quality with TensoIR [15]. Similar to Synthetic4Relight, R3DG [10] achieves higher accuracy in NVs. Fig. 6 presents a qualitative comparison of relighting images generated by GS-based methods [10, 19], showing that IRGS produces more realistic relighting results, with natural shadows.", + "bbox": [ + 89, + 652, + 482, + 773 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Stanford-ORB dataset We also conduct experiments on the real-world Stanford-ORB dataset [17]. As shown in Fig. 7, IRGS produces realistic relighting results.", + "bbox": [ + 89, + 773, + 482, + 820 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2. Ablation study", + "text_level": 1, + "bbox": [ + 89, + 830, + 240, + 847 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Tab. 3, we conduct ablation studies on various components of IRGS to validate their effectiveness. We report the average PSNR on NVS, albedo estimation, and relighting", + "bbox": [ + 89, + 854, + 482, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b03eb00d3851e35587fc395313a5e7bfaf08d74a6af9653cf59283af680e6bb2.jpg", + "table_caption": [ + "Table 3. Ablation studies on various components of IRGS." + ], + "table_footnote": [], + "table_body": "
MethodNVS PSNR ↑Albedo PSNR ↑Relighting PSNR ↑
Detach indirect34.2130.2934.22
w/o indirect (train)34.0930.1033.93
w/o indirect (relight)--33.84
Nr=1634.0130.2129.46
Nr=6434.9830.6333.11
Full35.4830.8134.68
", + "bbox": [ + 517, + 103, + 901, + 186 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "across four scenes in Synthetic4Relight dataset [43].", + "bbox": [ + 511, + 209, + 859, + 224 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Indirect light We perform ablation studies to evaluate the impact of indirect light modeling. \"Detach indirect\" indicates that gradients are not backpropagated through ray tracing to optimize indirect radiance. \"w/o indirect (train)\" signifies that the indirect term in Eq. (13) is omitted during training, while \"w/o indirect (relight)\" omits the proposed strategy for computing indirect radiance during relighting. As shown in Fig. 8a, detaching the indirect term results in unrealistic indirect radiance, and training without the indirect term leads to inaccurate albedo estimation. We also demonstrate the impact of indirect radiance on relighting results of a composited scene in Fig. 1 (indirect illumination). Note that the renderings in Fig. 1 are produced by assigning an additional high metallic to the Gaussian-based \"Ground\", enhancing the visual clarity of inter-reflections.", + "bbox": [ + 511, + 228, + 906, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Number of rays We also ablate on the number of sampled rays $N_{\\mathrm{r}}$ for rendering equation. We investigated the effects of $N_{\\mathrm{r}} = 16$ , $N_{\\mathrm{r}} = 64$ , and $N_{\\mathrm{r}} = 256$ . As shown in Fig. 8b, more sampled rays lead to more accurate and smooth renderings and ambient occlusion in novel view synthesis.", + "bbox": [ + 511, + 460, + 906, + 536 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 563, + 633, + 579 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce inter-reflective Gaussian splating (IRGS), a novel framework that overcomes the limitations of previous Gaussian-based approaches in accurately modeling visibility and indirect radiance for incident light in inverse rendering. IRGS incorporates the full rendering equation without simplifications and utilizes a differentiable 2D Gaussian ray tracing technique for precise computation of incident radiance at intersection points. We further propose an optimization scheme to address the computational demands of Monte Carlo sampling in rendering equation evaluation, and a novel strategy for querying indirect radiance during relighting. Extensive experiments on multiple benchmarks validate IRGS's ability to accurately model complex inter-reflection, establishing it as an effective solution for inverse rendering.", + "bbox": [ + 511, + 594, + 906, + 820 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitation Due to the significant computational demands of rendering equation evaluation, we are unable to render at a high frame rate. However, this could be addressed by baking the incident radiance or pre-computing radiance transfer, which we leave as future work.", + "bbox": [ + 511, + 825, + 905, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "10950", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by National Natural Science Foundation of China (Grant No. 62376060).", + "bbox": [ + 89, + 114, + 483, + 146 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 157, + 187, + 174 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Benjamin Attal, Dor Verbin, Ben Mildenhall, Peter Hedman, Jonathan T Barron, Matthew O'Toole, and Pratul P Srinivasan. Flash cache: Reducing bias in radiance cache based inverse rendering. In ECCV, 2025. 2", + "[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In ICCV, 2021. 2", + "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In CVPR, 2022.", + "[4] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In ICCV, 2023. 2", + "[5] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T. Barron, Ce Liu, and Hendrik P.A. Lensch. Nerd: Neural reflectance decomposition from image collections. In ICCV, 2021. 1, 2", + "[6] Mark Boss, Andreas Engelhardt, Abhishek Kar, Yuzhen Li, Deqing Sun, Jonathan Barron, Hendrik Lensch, and Varun Jampani. *Samurai: Shape and material from unconstrained real-world arbitrary image collections*. *NeurIPS*, 2022. 2", + "[7] Brent Burley and Walt Disney Animation Studios. Physically-based shading at disney. In Acm Siggraph, 2012. 3", + "[8] Yurui Chen, Chun Gu, Junzhe Jiang, Xiatian Zhu, and Li Zhang. Periodic vibration gaussian: Dynamic urban scene reconstruction and real-time rendering. arXiv preprint, 2023. 2", + "[9] Jan-Niklas Dihlmann, Arjun Majumdar, Andreas Engelhardt, Raphael Braun, and Hendrik Lensch. Subsurface scattering for 3d gaussian splattering. arXiv preprint, 2024. 2", + "[10] Jian Gao, Chun Gu, Youtian Lin, Hao Zhu, Xun Cao, Li Zhang, and Yao Yao. Relightable 3d gaussian: Real-time point cloud relighting with brdf decomposition and ray tracing. arXiv preprint, 2023. 1, 2, 3, 5, 6, 7, 8", + "[11] Yijia Guo, Yuanxi Bai, Liwen Hu, Ziyi Guo, Mianzhi Liu, Yu Cai, Tiejun Huang, and Lei Ma. Prtgs: Precomputed radiance transfer of gaussian splats for real-time high-quality relighting. In ACMMM, 2024. 2", + "[12] Yuan-Chen Guo, Di Kang, Linchao Bao, Yu He, and Song-Hai Zhang. Nerfren: Neural radiance fields with reflections. In CVPR, 2022. 5", + "[13] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In SIGGRAPH, 2024. 2, 4, 5, 7", + "[14] Yingwenqi Jiang, Jiadong Tu, Yuan Liu, Xifeng Gao, Xiaoxiao Long, Wenping Wang, and Yuexin Ma. Gaussianshader: 3d gaussian splatting with shading functions for reflective surfaces. arXiv preprint, 2023. 1, 2" + ], + "bbox": [ + 93, + 183, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Haian Jin, Isabella Liu, Peijia Xu, Xiaoshuai Zhang, Songfang Han, Sai Bi, Xiaowei Zhou, Zexiang Xu, and Hao Su. Tensoir: Tensorial inverse rendering. In CVPR, 2023. 1, 2, 6, 7, 8", + "[16] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. TOG, 2023. 1, 2, 3, 5", + "[17] Zhengfei Kuang, Yunzhi Zhang, Hong-Xing Yu, Samir Agarwala, Elliott Wu, Jiajun Wu, et al. Stanford-orb: a real-world 3d object inverse rendering benchmark. NeurIPS, 2024. 6, 8", + "[18] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In CVPR, 2023. 2", + "[19] Zhihao Liang, Qi Zhang, Ying Feng, Ying Shan, and Kui Jia. Gs-ir: 3d gaussian splatting for inverse rendering. In CVPR, 2024. 1, 2, 5, 6, 7, 8", + "[20] Yuan Liu, Peng Wang, Cheng Lin, Xiaoxiao Long, Jiepeng Wang, Lingjie Liu, Taku Komura, and Wenping Wang. Nano: Neural geometry and brdf reconstruction of reflective objects from multiview images. In SIGGRAPH, 2023. 1, 2, 5", + "[21] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2", + "[22] Nicolas Moënne-Loccoz, Ashkan Mirzaei, Or Perel, Ricardo de Lutio, Janick Martinez Esturo, Gavriel State, Sanja Fidler, Nicholas Sharp, and Zan Gojcic. 3d gaussian ray tracing: Fast tracing of particle scenes. arXiv preprint, 2024. 2, 4, 7", + "[23] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 2022. 2", + "[24] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting triangular 3d models, materials, and lighting from images. In CVPR, 2022. 1, 5, 6", + "[25] Steven G Parker, James Bigler, Andreas Dietrich, Heiko Friedrich, Jared Hoberock, David Luebke, David McAllister, Morgan McGuire, Keith Morley, Austin Robison, et al. Optix: a general purpose ray tracing engine. TOG, 2010. 2, 4, 7", + "[26] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mllps. In ICCV, 2021. 2", + "[27] Yahao Shi, Yanmin Wu, Chenming Wu, Xing Liu, Chen Zhao, Haocheng Feng, Jingtuo Liu, Liangjun Zhang, Jian Zhang, Bin Zhou, et al. Gir: 3d gaussian inverse rendering for relightable scene factorization. arXiv preprint, 2023. 1, 2", + "[28] Pratul P Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T Barron. Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In CVPR, 2021. 1, 2", + "[29] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint, 2021. 2, 5" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "10951", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Trans. Image Process., 2004. 6", + "[31] Tong Wu, Jia-Mu Sun, Yu-Kun Lai, Yuewen Ma, Leif Kobbelt, and Lin Gao. Deferreds: Decoupled and editable gaussian splatting with deferred shading. arXiv preprint, 2024. 2", + "[32] Yunzhi Yan, Haotong Lin, Chenxu Zhou, Weijie Wang, Haiyang Sun, Kun Zhan, Xianpeng Lang, Xiaowei Zhou, and Sida Peng. Street gaussians for modeling dynamic urban scenes. arXiv preprint, 2024. 2", + "[33] Ziyi Yang, Yanzhen Chen, Xinyu Gao, Yazhen Yuan, Yu Wu, Xiaowei Zhou, and Xiaogang Jin. Sire-ir: Inverse rendering for brdf reconstruction with shadow and illumination removal in high-illuminance scenes. arXiv preprint, 2023. 2", + "[34] Zeyu Yang, Hongye Yang, Zijie Pan, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. arXiv preprint, 2023. 2", + "[35] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. In CVPR, 2024. 2", + "[36] Yao Yao, Jingyang Zhang, Jingbo Liu, Yihang Qu, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Neilf: Neural incident light field for physically-based material estimation. In ECCV, 2022. 2", + "[37] Yuxuan Yao, Zixuan Zeng, Chun Gu, Xiatian Zhu, and Li Zhang. Reflective gaussian splatting. In ICLR, 2025. 2", + "[38] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. NeurIPS, 2021. 2", + "[39] Zehao Yu, Torsten Sattler, and Andreas Geiger. Gaussian opacity fields: Efficient adaptive surface reconstruction in unbounded scenes. TOG, 2024. 2", + "[40] Jingyang Zhang, Yao Yao, Shiwei Li, Jingbo Liu, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Neilf++: Inter-reflectable light fields for geometry and material estimation. In ICCV, 2023. 1, 2", + "[41] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6", + "[42] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM Trans. Graph., 2021. 6, 7, 8", + "[43] Yuanqing Zhang, Jiaming Sun, Xingyi He, Huan Fu, Rongfei Jia, and Xiaowei Zhou. Modeling indirect illumination for inverse rendering. In CVPR, 2022. 2, 6, 7, 8" + ], + "bbox": [ + 91, + 92, + 482, + 772 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10952", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/IRGS_ Inter-Reflective Gaussian Splatting with 2D Gaussian Ray Tracing/ff8cd318-c293-446d-b3d4-b22f9fc186ef_model.json b/2025/IRGS_ Inter-Reflective Gaussian Splatting with 2D Gaussian Ray Tracing/ff8cd318-c293-446d-b3d4-b22f9fc186ef_model.json new file mode 100644 index 0000000000000000000000000000000000000000..32a3d2b5ccb650a43f06392e1d4a647f9cf1517b --- /dev/null +++ b/2025/IRGS_ Inter-Reflective Gaussian Splatting with 2D Gaussian Ray Tracing/ff8cd318-c293-446d-b3d4-b22f9fc186ef_model.json @@ -0,0 +1,1991 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.129, + 0.131, + 0.87, + 0.153 + ], + "angle": 0, + "content": "IRGS: Inter-Reflective Gaussian Splating with 2D Gaussian Ray Tracing" + }, + { + "type": "text", + "bbox": [ + 0.193, + 0.18, + 0.808, + 0.2 + ], + "angle": 0, + "content": "Chun Gu\\(^{1,2}\\) Xiaofei Wei\\(^{1,2}\\) Zixuan Zeng\\(^{1}\\) Yuxuan Yao\\(^{1}\\) Li Zhang\\(^{1,2*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.193, + 0.199, + 0.798, + 0.218 + ], + "angle": 0, + "content": "\\(^{1}\\)School of Data Science, Fudan University \\(^{2}\\)Shanghai Innovation Institute" + }, + { + "type": "text", + "bbox": [ + 0.305, + 0.225, + 0.686, + 0.242 + ], + "angle": 0, + "content": "https://fudan-zvq.github.io/IRGS" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.277, + 0.327, + 0.293 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.309, + 0.485, + 0.611 + ], + "angle": 0, + "content": "In inverse rendering, accurately modeling visibility and indirect radiance for incident light is essential for capturing secondary effects. Due to the absence of a powerful Gaussian ray tracer, previous 3DGS-based methods have either adopted a simplified rendering equation or used learnable parameters to approximate incident light, resulting in inaccurate material and lighting estimations. To this end, we introduce inter-reflective Gaussian splatting (IRGS) for inverse rendering. To capture inter-reflection, we apply the full rendering equation without simplification and compute incident radiance on the fly using the proposed differentiable 2D Gaussian ray tracing. Additionally, we present an efficient optimization scheme to handle the computational demands of Monte Carlo sampling for rendering equation evaluation. Furthermore, we introduce a novel strategy for querying the indirect radiance of incident light when relighting the optimized scenes. Extensive experiments on multiple standard benchmarks validate the effectiveness of IRGS, demonstrating its capability to accurately model complex inter-reflection effects." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.64, + 0.222, + 0.656 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.665, + 0.483, + 0.877 + ], + "angle": 0, + "content": "Inverse rendering is a fundamental problem in computer vision and graphics, aiming to reconstruct geometry and estimate material and lighting from a set of posed images. The introduction of neural radiance field (NeRF) [21], which uses a neural implicit field modeled by a multi-layer perceptron (MLP) to represent 3D scenes, has inspired many NeRF-based methods [5, 15, 20, 28, 40] to address inverse rendering tasks. Leveraging NeRF's ray-based rendering technique, these methods can accurately compute the visibility and indirect radiance for incident light as required by the rendering equation. However, the exhaustive querying of a neural field during ray marching and its limited representational capacity constrain the efficiency and performance of neural implicit fields in inverse rendering." + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.276, + 0.891, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.452, + 0.891, + 0.629 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.639, + 0.908, + 0.669 + ], + "angle": 0, + "content": "Figure 1. Global and indirect illumination in a Gaussian-based scene using our IRGS, highlighting its inter-reflective capabilities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Recently, 3D Gaussian splatting (3DGS) [16] has emerged as a promising technique for 3D scene representation, offering both remarkable rendering quality and efficiency. By modeling a static 3D scene as a collection of 3D Gaussians, 3DGS achieves real-time, high-quality rendering with a tile-based rasterizer. However, despite its strengths, 3DGS has notable limitations due to its reliance on rasterization, which limits its ability to accurately simulate ray-based effects that are essential for inverse rendering. As a result, some 3DGS-based inverse rendering methods [14, 19, 27] have adopted simplified versions of the rendering equation, such as the split-sum approximation [24]. Other approaches, such as R3DG [10], use learnable parameters to approximate indirect radiance, which fails to" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.888, + 0.421, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author (lizhangfd@fudan.edu.cn)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10943" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.137 + ], + "angle": 0, + "content": "capture actual inter-reflection. These methods struggle to accurately model the rendering equation, limiting their effectiveness in realistic material and lighting estimation." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.14, + 0.482, + 0.291 + ], + "angle": 0, + "content": "Our goal is to integrate the full rendering equation into the efficient Gaussian splatting for accurate inverse rendering, enabling the capture of complex interreflection effects. This requires efficient and precise computation of visibility and indirect radiance for incident light. Drawing inspiration from 3DGRT [22], which achieves efficient ray tracing on 3D Gaussian primitives via a \\(k\\)-buffer hits-based marching technique and hardware-accelerated NVIDIA OptiX interface [25], we explore adapting this technique to model inter-reflection within the Gaussian splatting framework." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.292, + 0.485, + 0.822 + ], + "angle": 0, + "content": "In this paper, we introduce inter-reflective Gaussian splatting (IRGS), a novel inverse rendering framework that leverages the proposed 2D Gaussian ray tracing technique to model inter-reflections within Gaussian splatting. To the best of our knowledge, this is the first method to integrate the full rendering equation without simplification into Gaussian splatting, accurately capturing inter-reflection effects by querying incident lighting through ray tracing. First, we propose 2D Gaussian ray tracing to efficiently and accurately trace across 2D Gaussian primitives, computing the accumulated opacity and color of arbitrary rays. Unlike 3DGRT [22], our approach features well-defined ray-splat intersections, enabling direct ray tracing on pretrained 2D Gaussian splatting [13] with minimal quality degradation. Next, we introduce the inter-reflective Gaussian splatting framework: starting from a pretrained 2D Gaussian splatting model, we apply the full rendering equation without simplification at intersection points determined by the depth map, while visibility and indirect radiance are computed on the fly using our 2D Gaussian ray tracing technique. Notably, our 2D Gaussian ray tracing is fully differentiable, enabling the optimization of indirect light through backpropagation. We further propose an efficient optimization scheme to manage the computational demands of the exhaustive Monte Carlo sampling for the rendering equation. Additionally, we introduce a novel strategy to query indirect radiance during relighting, overcoming limitations in previous methods that often omit this component. Extensive experiments on multiple standard benchmarks confirm the effectiveness of IRGS. In Fig. 1, we visualize the global illumination (considering both direct and indirect radiance) and the indirect illumination (considering only indirect radiance) of a Gaussian-based scene rendered using IRGS, demonstrating that IRGS is capable of modeling remarkable inter-reflection effects." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.485, + 0.903 + ], + "angle": 0, + "content": "The contributions of this paper are as follows: (i) 2D Gaussian ray tracing, a technique that enables direct ray tracing on pretrained 2D Gaussian splatting with minimal quality degradation; (ii) IRGS, an inverse rendering framework that incorporates the full rendering equation without" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.168 + ], + "angle": 0, + "content": "simplification, using 2D Gaussian ray tracing to compute incident radiance on the fly; (iii) An optimization scheme that manages computational demands for rendering equation evaluation; (iv) A novel strategy that enables querying of indirect radiance during relighting." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.179, + 0.652, + 0.195 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.204, + 0.908, + 0.612 + ], + "angle": 0, + "content": "Novel view synthesis aims to leverage a limited collection of observed images of a scene to generate new images from an unseen viewpoint. Neural radiance field (NeRF) [21] produces photo-realistic quality images by representing the scenes in terms of a volumetric radiance field encoded in a coordinate-based neural network. Subsequent advancements have focused on accelerating rendering[23, 26], enhancing image quality [2-4], and improving geometry reconstruction [18, 29, 38]. More recently, 3D Gaussian splatting (3DGS) [16] introduced the use of fuzzy, anisotropic 3D Gaussian point cloud for scene representation along with an efficient tile-based rasterizer, achieving state-of-the-art results in both rendering quality and efficiency. 3DGS has inspired numerous approaches targeting various goals, including geometry reconstruction [13, 39], dynamic scene reconstruction [34, 35], inverse rendering [10, 19], and street scene applications [8, 32]. However, the dependence on rasterization limits Gaussian splatting's ability to simulate ray-based effects accurately. To address this, 3DGRT [22] proposes a differentiable Gaussian ray tracer that efficiently performs ray tracing across 3D Gaussian primitives, allowing radiance computation along arbitrary rays. In this paper, we propose ray tracing on 2D Gaussian primitives. Unlike 3DGRT, 2D Gaussian primitives [13] have a well-defined ray-splat intersection, enabling direct ray tracing on pre-trained 2D Gaussian splatting with minimal quality degradation." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.629, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Inverse rendering focuses on recovering geometry, material, and lighting properties from captured images, presenting a complex challenge due to high-dimensional light-surface interactions. Inspired by the success of neural radiance fields, many methods [1, 5, 6, 15, 20, 28, 33, 36, 40, 43] leverage flexible ray marching techniques and neural implicit fields for inverse rendering. Recent approaches [9-11, 14, 19, 27, 31, 37] have adapted 3D Gaussian splatting (3DGS) to inverse rendering by assigning material-related properties to each Gaussian primitive. Leveraging the rapid rendering capabilities of 3DGS, these methods achieve much higher efficiency compared to NeRF-based techniques. However, due to the lack of an efficient Gaussian ray tracer, existing 3DGS-based methods either apply a simplified version of the rendering equation [19] or use learnable parameters to model incident light [10], which limits the accuracy of material and lighting estimation. For example, GS-IR [19] employs split-sum approximation to" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10944" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.089, + 0.905, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.249, + 0.908, + 0.308 + ], + "angle": 0, + "content": "Figure 2. Schematic illustration of the proposed IRGS. Starting from a set of 2D Gaussians equipped with material properties, we apply rasterization to generate albedo, roughness, position, and normal maps. We then evaluate the rendering equation using stratified sampling at the corresponding position, drawing geometry and material values from these feature maps. The radiance of incident light is decomposed into direct radiance \\( L_{\\mathrm{dir}} \\) from the environment map, and indirect radiance \\( L_{\\mathrm{ind}} \\) with visibility \\( V \\), obtained via 2D Gaussian ray tracing." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.321, + 0.485, + 0.429 + ], + "angle": 0, + "content": "simplify the rendering equation and uses baked volumes to store occlusion and indirect illumination. R3DG [10] shades each Gaussian individually, using point-based ray tracing to bake occlusion and spherical harmonics to parameterize indirect radiance. To overcome these limitations, we propose the IRGS framework that utilizes 2D Gaussian ray tracing to accurately model inter-reflections." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.441, + 0.182, + 0.456 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.466, + 0.484, + 0.633 + ], + "angle": 0, + "content": "In this section, we introduce our inter-reflective Gaussian splatting with 2D Gaussian ray tracing, designed to reconstruct geometry and estimate material and lighting from a set of posed RGB images. We begin by providing the necessary background on 3D Gaussian splatting and the rendering equation (Section 3.1). We then present the 2D Gaussian ray tracing technique, which performs ray tracing on 2D Gaussian primitives (Section 3.2). Finally, we outline our IRGS framework, which leverages the proposed 2D Gaussian ray tracing technique to query visibility and indirect radiance for incident light (Section 3.3)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.642, + 0.221, + 0.658 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.665, + 0.484, + 0.743 + ], + "angle": 0, + "content": "3D Gaussian splatting 3D Gaussian splatting (3DGS) represents a 3D scene using 3D Gaussian primitives, each defined as an ellipsoid characterized by a 3D mean \\(\\mu\\) and a covariance matrix \\(\\Sigma\\). The influence of a Gaussian at a spatial position \\(x\\) is expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.752, + 0.483, + 0.786 + ], + "angle": 0, + "content": "\\[\nG (\\boldsymbol {x}) = \\exp \\left(- \\frac {1}{2} (\\boldsymbol {x} - \\boldsymbol {\\mu}) ^ {\\top} \\Sigma^ {- 1} (\\boldsymbol {x} - \\boldsymbol {\\mu})\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Each Gaussian is also assigned an opacity \\(o\\) and a view-dependent color \\(c\\) modeled by spherical harmonics (SH). The rendering process of 3DGS starts by projecting 3D Gaussian primitives onto the 2D image plane. This is achieved by first applying a viewing transformation \\(W\\) and then performing a perspective projection. The covariance matrix of the resulting 2D Gaussians can be approximated" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.32, + 0.907, + 0.381 + ], + "angle": 0, + "content": "as \\(\\Sigma' = JW\\Sigma W^\\top J^\\top\\), where \\(J\\) represents the Jacobian matrix of the perspective projection. The final image is rendered by alpha-blending these projected 2D Gaussians from front to back:" + }, + { + "type": "equation", + "bbox": [ + 0.582, + 0.389, + 0.907, + 0.434 + ], + "angle": 0, + "content": "\\[\n\\mathcal {C} = \\sum_ {i = 1} ^ {N} T _ {i} \\alpha_ {i} \\boldsymbol {c} _ {i}, \\quad T _ {i} = \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.443, + 0.907, + 0.505 + ], + "angle": 0, + "content": "where \\(\\alpha_{i}\\) is calculated by multiplying opacity \\(o_i\\) with each Gaussian's influence, determined by \\(\\Sigma^{\\prime}\\) and the pixel coordinate. The covariance matrix \\(\\Sigma\\) is parameterized by a unit quaternion and a scaling vector to facilitate optimization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.522, + 0.906, + 0.568 + ], + "angle": 0, + "content": "Rendering equation In physically-based rendering, illumination at a given surface point \\( \\pmb{x} \\) is computed via the rendering equation:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.578, + 0.905, + 0.611 + ], + "angle": 0, + "content": "\\[\nL _ {o} \\left(\\boldsymbol {\\omega} _ {o}, \\boldsymbol {x}\\right) = \\int_ {\\Omega} f \\left(\\boldsymbol {\\omega} _ {o}, \\boldsymbol {\\omega} _ {i}, \\boldsymbol {x}\\right) L _ {1} \\left(\\boldsymbol {\\omega} _ {i}, \\boldsymbol {x}\\right) \\left(\\boldsymbol {\\omega} _ {i} \\cdot \\mathbf {n}\\right) d \\boldsymbol {\\omega} _ {i}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.62, + 0.907, + 0.756 + ], + "angle": 0, + "content": "where \\(\\omega_{o}\\) is the outgoing direction, \\(\\omega_{i}\\) is the incident direction over the hemisphere defined by the surface normal \\(\\pmb{n}\\), \\(f\\) is the bidirectional reflectance distribution function (BRDF), and \\(L_{\\mathrm{i}}(\\boldsymbol{\\omega}_i,\\boldsymbol{x})\\) is the radiance of the incident light. In this work, we use a simplified Disney BRDF model [7] with only diffuse albedo \\(\\pmb{a}\\) and roughness \\(r\\) as parameters, and assume the material is dielectric. The BRDF \\(f\\) in Eq. (3) can be separated into a diffuse term \\(f_{d} = \\frac{a}{\\pi}\\) and a specular term \\(f_{s}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.765, + 0.905, + 0.8 + ], + "angle": 0, + "content": "\\[\nf _ {s} \\left(\\boldsymbol {\\omega} _ {i}, \\boldsymbol {\\omega} _ {o}, \\boldsymbol {x}\\right) = \\frac {D F G}{4 \\left(\\boldsymbol {\\omega} _ {i} \\cdot \\mathbf {n}\\right) \\left(\\boldsymbol {\\omega} _ {o} \\cdot \\mathbf {n}\\right)}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.808, + 0.905, + 0.84 + ], + "angle": 0, + "content": "where \\(D, F\\), and \\(G\\) represent the normal distribution function, the Fresnel term, and the geometry term, respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.74, + 0.864 + ], + "angle": 0, + "content": "3.2. 2D Gaussian ray tracing" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "While 3DGS [16] has been successful in applications such as 3D reconstruction, achieving superior rendering quality" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10945" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.089, + 0.473, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.204, + 0.483, + 0.273 + ], + "angle": 0, + "content": "Figure 3. Performance of directly applying Gaussian ray tracing on a pretrained Gaussian splatting checkpoint in 3D Gaussian and 2D Gaussian cases, respectively. Our proposed 2DGRT achieves significantly less quality degradation than 3DGRT [22] in both quantitative metrics and visual results." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.29, + 0.483, + 0.501 + ], + "angle": 0, + "content": "and efficiency compared to NeRF-based methods, it is limited by the constraints of rasterization. These limitations are primarily related to ray-based effects, including rendering with highly distorted cameras and handling secondary ray effects. To address these challenges, 3DGRT [22] proposes performing ray tracing across 3D Gaussian primitives. This approach overcomes the inefficiencies of ray tracing on 3D Gaussian primitives by implementing a \\(k\\)-buffer hits-based marching technique with a hardware-accelerated NVIDIA OptiX interface [25]. In this paper, as our focus is on inverse rendering, performing ray tracing on Gaussian primitives is essential to accurately compute inter-reflection effects when evaluating the rendering equation, including visibility and indirect radiance of incident light." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.505, + 0.483, + 0.776 + ], + "angle": 0, + "content": "3D Gaussian ray tracing [22] performs ray tracing on 3D Gaussian primitives and determines the ray-splat intersection as the point of maximum response from the corresponding 3D Gaussian along the ray. This behavior differs from that of 3D Gaussian splatting, which computes the particle response in 2D image space, leaving the position of the 3D intersection point undefined. To leverage both the efficiency of Gaussian splatting and the flexible rendering capabilities of Gaussian ray tracing, we aim to perform ray tracing directly on the Gaussian splatting model for incident radiance inference. However, as shown in Fig. 3, applying ray tracing directly on a 3DGS checkpoint results in a noticeable decline in rendering quality compared to splatting. To address this issue, we propose 2D Gaussian ray tracing (2DGRT), which performs ray tracing on 2D Gaussian primitives [13]. Since 2D Gaussian disks have a well-defined ray-splat intersection, this approach eliminates inconsistencies in intersection points present in 3D Gaussians." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Given a 2D Gaussian [13] defined by its center point \\(\\pmb{\\mu} \\in \\mathbb{R}^3\\), an opacity parameter \\(o \\in [0,1]\\), two principal tangential vectors \\(\\pmb{t}_u \\in \\mathbb{R}^3\\) and \\(\\pmb{t}_v \\in \\mathbb{R}^3\\), and a scaling vector \\(\\pmb{s} = (s_u, s_v) \\in \\mathbb{R}^2\\), we follow 3DGRT [22] to fit a bounding proxy of each 2D Gaussian using an adaptive icosahedron mesh, ensuring that a minimum influence of \\(\\alpha_{\\mathrm{min}}\\) is enclosed. The vertices \\(\\pmb{v}\\) of the bounding icosahedron mesh are obtained by transforming a regular icosahedron with a" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.093, + 0.634, + 0.107 + ], + "angle": 0, + "content": "unit inner sphere:" + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.115, + 0.905, + 0.133 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {v} \\leftarrow \\sqrt {2 \\log \\left(o / \\alpha_ {\\min }\\right)} \\left(s _ {u} \\boldsymbol {t} _ {u} \\quad s _ {v} \\boldsymbol {t} _ {v} \\quad \\epsilon \\mathbf {1}\\right) \\boldsymbol {v} + \\boldsymbol {\\mu}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.141, + 0.907, + 0.293 + ], + "angle": 0, + "content": "where \\(\\epsilon\\) denotes a small positive number. Each icosahedron mesh consists of 20 triangular faces, resulting in a total of \\(20N\\) triangles for \\(N\\) 2D Gaussian primitives. To leverage the hardware-accelerated ray-triangle intersection provided by OptiX [25], we construct a Bounding Volume Hierarchy (BVH) using these triangles. By applying the \\(k\\)-buffer per-ray sorting algorithm [22], we obtain the exact ordering of intersected Gaussians for each ray. To ensure consistency with 2DGS [13], we analytically compute the ray-splat intersection point \\(p\\):" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.299, + 0.905, + 0.334 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {p} = \\boldsymbol {r} _ {o} + \\tau \\boldsymbol {r} _ {d}, \\quad \\text {w h e r e} \\quad \\tau = \\frac {\\boldsymbol {n} ^ {\\top} (\\boldsymbol {\\mu} - \\boldsymbol {r} _ {o})}{\\boldsymbol {n} ^ {\\top} \\boldsymbol {r} _ {d}}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.339, + 0.906, + 0.4 + ], + "angle": 0, + "content": "where \\(\\pmb{n} = \\pmb{t}_u\\times \\pmb{t}_v\\) is the normal vector of the 2D Gaussian disk, and \\(\\pmb{r}_o\\) and \\(\\pmb{r}_d\\) represent the origin and direction of the given ray, respectively. Thus, the influence of the current 2D Gaussian primitive is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.614, + 0.407, + 0.905, + 0.441 + ], + "angle": 0, + "content": "\\[\nG (\\boldsymbol {p}) = \\exp \\left(- \\frac {u ^ {2} + v ^ {2}}{2}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.448, + 0.831, + 0.466 + ], + "angle": 0, + "content": "where \\(u = \\frac{1}{s_v} (\\pmb {p} - \\pmb {\\mu})^\\top \\pmb {t}_u,v = \\frac{1}{s_v} (\\pmb {p} - \\pmb {\\mu})^\\top \\pmb {t}_v\\)" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.465, + 0.906, + 0.509 + ], + "angle": 0, + "content": "Consequently, the RGB value and accumulated opacity value of the ray can be obtained by alpha-blending (Eq. (2)) the intersected 2D Gaussians from front to back:" + }, + { + "type": "equation", + "bbox": [ + 0.618, + 0.519, + 0.905, + 0.535 + ], + "angle": 0, + "content": "\\[\n\\left(\\boldsymbol {c} _ {\\mathrm {r t}}, o _ {\\mathrm {r t}}\\right) \\leftarrow \\operatorname {T r a c e} \\left(\\boldsymbol {r} _ {o}, \\boldsymbol {r} _ {d}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.543, + 0.907, + 0.663 + ], + "angle": 0, + "content": "where the RGB value of each Gaussian is modeled by the view-dependent color \\( c \\) as in vanilla 2D Gaussian splatting. It is worth noting that subtle differences remain between 2D Gaussian ray tracing and 2DGS due to sorting errors introduced by per-tile sorting in 2DGS. However, as shown in Fig. 3, the quality degradation from 2D Gaussian ray tracing is minimal and significantly less noticeable than that observed in 3D Gaussians." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.672, + 0.81, + 0.689 + ], + "angle": 0, + "content": "3.3. Inter-reflective Gaussian splatting" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.696, + 0.799, + 0.711 + ], + "angle": 0, + "content": "3.3.1 Stage I: 2D Gaussian pretraining" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.719, + 0.906, + 0.855 + ], + "angle": 0, + "content": "Before estimating material and lighting for inverse rendering, we first pretrain a standard 2D Gaussian splatting model. This step is essential, as reliable geometry is required for inverse rendering, and evaluating the rendering equation is computationally intensive. Using a pretrained 2D Gaussian splatting model can greatly accelerate the training process. Along with the RGB rendering described in Eq. (2), we also render normalized depth and normal maps as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.863, + 0.905, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\mathcal {D}, \\mathcal {N} \\right\\} = \\sum_ {i = 1} ^ {N} w _ {i} \\left\\{d _ {i}, \\boldsymbol {n} _ {i} \\right\\}, \\text {w h e r e} w _ {i} = \\frac {T _ {i} \\alpha_ {i}}{\\sum_ {i = 1} ^ {N} T _ {i} \\alpha_ {i}}. \\tag {9}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10946" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.198 + ], + "angle": 0, + "content": "Training loss Following 2DGS [13], we apply a normal consistency loss \\(\\mathcal{L}_{\\mathrm{n}}\\) to ensure alignment between the rendered normal map and the underlying geometry, as well as a depth distortion loss [13] \\(\\mathcal{L}_{\\mathrm{d}}\\) to encourage concentration of the 2D Gaussians. Additionally, we introduce an edge-aware smoothness loss [10, 12] \\(\\mathcal{L}_{\\mathrm{s,n}}\\) on the normal map to enhance geometric robustness:" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.21, + 0.483, + 0.228 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {s}, \\mathrm {n}} = \\| \\nabla \\mathcal {N} \\| \\exp (- \\| \\nabla \\mathcal {C} _ {\\mathrm {g t}} \\|), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.239, + 0.484, + 0.3 + ], + "angle": 0, + "content": "where \\(\\mathcal{C}_{\\mathrm{gt}}\\) is the ground truth training image. To address the presence of floaters during standard 2DGS training, we also incorporate a binary cross-entropy loss [29] to constrain the geometry using the provided object mask \\(\\mathcal{M}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.313, + 0.483, + 0.33 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {o}} = - \\mathcal {M} \\log \\mathcal {O} - (1 - \\mathcal {M}) \\log (1 - \\mathcal {O}), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.342, + 0.484, + 0.389 + ], + "angle": 0, + "content": "where \\(\\mathcal{O} = \\sum_{i=1}^{N} T_i \\alpha_i\\), denotes the accumulated opacity map. Combined with the RGB reconstruction loss \\(\\mathcal{L}_{\\mathrm{c}}\\) from 3DGS [16], the training loss for the first stage is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.118, + 0.416, + 0.483, + 0.433 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {1} = \\mathcal {L} _ {\\mathrm {c}} + \\lambda_ {\\mathrm {n}} \\mathcal {L} _ {\\mathrm {n}} + \\lambda_ {\\mathrm {d}} \\mathcal {L} _ {\\mathrm {d}} + \\lambda_ {\\mathrm {s}, \\mathrm {n}} \\mathcal {L} _ {\\mathrm {s}, \\mathrm {n}} + \\lambda_ {\\mathrm {o}} \\mathcal {L} _ {\\mathrm {o}}. \\tag {12}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.453, + 0.332, + 0.468 + ], + "angle": 0, + "content": "3.3.2 Stage II: inverse rendering" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.477, + 0.484, + 0.612 + ], + "angle": 0, + "content": "The order of shading and rasterization is crucial in physically based rendering with Gaussian splatting. Previous works, such as R3DG [10], propose performing shading on each Gaussian, taking advantage of properties modeled on each Gaussian, such as indirect radiance. However, since the rendered normal map is supervised at the pixel level after rasterization, shading on individual Gaussians can lead to inaccurate and blurred results. Consequently, we propose applying the rendering equation after rasterization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.484, + 0.705 + ], + "angle": 0, + "content": "Rasterization As introduced in Section 3.1, a dielectric BRDF model is used for the rendering equation. We assign each 2D Gaussian primitive additional material properties: diffuse albedo \\( \\pmb{a} \\) and roughness \\( r \\). Pixel-level albedo map \\( \\mathcal{A} \\) and roughness map \\( \\mathcal{R} \\) are obtained through alpha-blending during rasterization: \\( \\{\\mathcal{A},\\mathcal{R}\\} = \\sum_{i = 1}^{N}w_{i}\\{a_{i},r_{i}\\} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.706, + 0.484, + 0.87 + ], + "angle": 0, + "content": "Previous inverse rendering methods using Gaussian splatting have lacked an efficient way to query the visibility and radiance for arbitrary incident directions. As a result, these methods typically apply simplified versions of the rendering equation [19] (e.g., split-sum approximation [24]) or use learnable parameters to model indirect light [10]. In this paper, we leverage the proposed 2D Gaussian ray tracing technique to efficiently and accurately query inter-reflection effects on the fly, including the visibility and indirect radiance of incident light. This allows us to apply the full rendering equation without simplification." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Light parametrization We decompose the incident radiance at surface point \\( \\pmb{x} \\) with direction \\( \\omega_{i} \\) in Eq. (3) into" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.848, + 0.107 + ], + "angle": 0, + "content": "direct radiance from infinity and indirect radiance:" + }, + { + "type": "equation", + "bbox": [ + 0.541, + 0.114, + 0.905, + 0.131 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {i}} \\left(\\boldsymbol {\\omega} _ {i}, \\boldsymbol {x}\\right) = V \\left(\\boldsymbol {\\omega} _ {i}, \\boldsymbol {x}\\right) L _ {\\mathrm {d i r}} \\left(\\boldsymbol {\\omega} _ {i}\\right) + L _ {\\mathrm {i n d}} \\left(\\boldsymbol {\\omega} _ {i}, \\boldsymbol {x}\\right), \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.137, + 0.905, + 0.183 + ], + "angle": 0, + "content": "where \\(L_{\\mathrm{dir}}(\\omega_i)\\) is modeled using an environment cubemap, and \\(V(\\omega_i, \\boldsymbol{x})\\) along with \\(L_{\\mathrm{ind}}(\\omega_i, \\boldsymbol{x})\\) are queried using 2D Gaussian ray tracing:" + }, + { + "type": "equation", + "bbox": [ + 0.545, + 0.19, + 0.905, + 0.207 + ], + "angle": 0, + "content": "\\[\n\\left(L _ {\\text {i n d}} \\left(\\boldsymbol {\\omega} _ {i}, \\boldsymbol {x}\\right), 1 - V \\left(\\boldsymbol {\\omega} _ {i}, \\boldsymbol {x}\\right)\\right) \\leftarrow \\operatorname {T r a c e} \\left(\\boldsymbol {x}, \\boldsymbol {\\omega} _ {i}\\right) \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.213, + 0.905, + 0.303 + ], + "angle": 0, + "content": "It is important to note that the RGB values used in ray tracing correspond to the view-dependent color \\( c \\) from the first stage. Additionally, our 2D Gaussian ray tracing approach is fully differentiable, allowing gradients to propagate through the ray tracing process to optimize the indirect radiance of the incident ray." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.304, + 0.906, + 0.425 + ], + "angle": 0, + "content": "Rendering From the rendered depth and normal maps, we can determine the surface position \\( \\mathbf{x} \\) and the corresponding surface normal vector \\( \\mathbf{n} \\) for each pixel coordinate. Monte Carlo sampling is used to evaluate the rendering equation. Since we focus on reconstructing objects with dielectric materials, we employ stratified sampling to uniformly sample \\( N_{\\mathrm{r}} \\) incident directions across the hemisphere. The final rendering result is then computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.551, + 0.431, + 0.905, + 0.472 + ], + "angle": 0, + "content": "\\[\n\\mathbf {c} _ {\\mathrm {p b r}} = \\frac {2 \\pi}{N _ {\\mathrm {r}}} \\sum_ {i = 1} ^ {N _ {\\mathrm {r}}} \\left(f _ {\\mathrm {d}} + f _ {\\mathrm {s}}\\right) L _ {\\mathrm {i}} \\left(\\boldsymbol {\\omega} _ {i}, \\boldsymbol {x}\\right) \\left(\\boldsymbol {\\omega} _ {i} \\cdot \\mathbf {n}\\right), \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.478, + 0.906, + 0.568 + ], + "angle": 0, + "content": "Training loss In addition to the training losses from the first stage, we apply regularization to facilitate material and lighting estimation. Given the prior that diffuse light \\( L_{\\mathrm{diffuse}} = \\frac{1}{N_{\\mathrm{r}}} \\sum_{i=1}^{N_{\\mathrm{r}}} L(\\boldsymbol{\\omega}_i, \\boldsymbol{x}) \\) exhibits a natural white radiance, we minimize the difference between the RGB values [20]:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.574, + 0.905, + 0.621 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {l i g h t}} = \\sum_ {c} \\left(L _ {\\text {d i f f u s e}} - \\frac {1}{3} \\sum_ {c} L _ {\\text {d i f f u s e}}\\right), c \\in \\{R, G, B \\}. \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.622, + 0.907, + 0.864 + ], + "angle": 0, + "content": "Similar to Eq. (10), we also apply edge-aware smoothness regularization \\(\\mathcal{L}_{\\mathrm{s,a}}\\) and \\(\\mathcal{L}_{\\mathrm{s,r}}\\) to the rendered albedo and roughness map, respectively, to ensure smooth material estimation. Additionally, for training efficiency, we apply the rendering equation to only a subset of pixels in each iteration. We set a maximum of \\(N_{\\mathrm{rays}}\\) rays per iteration, meaning only \\(\\lfloor N_{\\mathrm{rays}} / N_{\\mathrm{r}}\\rfloor\\) pixels are randomly selected for final rendering equation evaluation. This scheme allows us to use a large \\(N_{\\mathrm{r}}\\) when evaluating the rendering equation, significantly enhancing estimation quality. Consequently, only a L1 loss \\(\\mathcal{L}_1^{\\mathrm{pbr}}\\) is used to supervise the final color \\(c_{\\mathrm{pbr}}\\). It is important to note that the rendered image \\(\\mathcal{C}\\) from the vanilla 2DGS is still optimized, while the view-dependent color \\(c\\) associated with each Gaussian is used to compute radiance in 2D Gaussian ray tracing. The total loss for the second stage is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.868, + 0.905, + 0.901 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {2} = \\mathcal {L} ^ {1} + \\lambda_ {1} ^ {\\mathrm {p b r}} \\mathcal {L} _ {1} ^ {\\mathrm {p b r}} + \\lambda_ {\\text {l i g h t}} \\mathcal {L} _ {\\text {l i g h t}} + \\lambda_ {\\mathrm {s}, \\mathrm {a}} \\mathcal {L} _ {\\mathrm {s}, \\mathrm {a}} + \\lambda_ {\\mathrm {s}, \\mathrm {r}} \\mathcal {L} _ {\\mathrm {s}, \\mathrm {r}}, \\tag {17}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10947" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.089, + 0.899, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.098, + 0.416, + 0.899, + 0.432 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison of NVS, material and lighting estimation, and relighting results on the Synthetic4Relight dataset [43]." + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.445, + 0.884, + 0.46 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison on the Synthetic4Relight dataset [43]. A higher intensity of the red color signifies a better result." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.47, + 0.903, + 0.591 + ], + "angle": 0, + "content": "
Novel view synthesisRelightingAlbedoRoughness MSE↓Time (hours)
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
NeRFactor [42]22.800.9160.15021.540.8750.17119.490.8640.206N/A>48
InvRender [43]30.740.9530.08628.670.9500.09128.280.9350.0720.00814
TensorIR [15]35.800.9780.04929.690.9510.07930.580.9460.0650.0153
GS-IR [19]33.950.9650.05725.400.9240.08319.480.8960.1170.0110.4
R3DG [10]36.800.9820.02831.000.9640.05028.310.9510.0580.0130.9
IRGS (Ours)35.480.9740.04334.900.9690.04830.810.9570.0550.0080.7
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.616, + 0.486, + 0.875 + ], + "angle": 0, + "content": "Relighting When performing relighting, changes in environmental lighting make the optimized radiance \\( c \\) of each Gaussian no longer applicable for indirect radiance in 2D Gaussian ray tracing. Unlike previous methods [10, 19] that omit indirect lighting during relighting, we propose using the split-sum approximation [24] to efficiently query indirect radiance without recursive evaluation. Specifically, we apply ray tracing on each incident ray to alpha-blend the albedo, roughness, and normal values, resulting in \\( \\mathcal{A}_{\\mathrm{r}} \\), \\( \\mathcal{R}_{\\mathrm{r}} \\), and \\( \\mathcal{N}_{\\mathrm{r}} \\). These aggregated values are then shaded using a pre-filtered environment map, following the approach in Nvdiffrec [24]. Additionally, we apply importance sampling on the intensity of the ground truth environment map instead of stratified sampling, enabling more accurate rendering equation evaluation by concentrating samples in high-intensity directions. Please refer to the appendix for further details." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.615, + 0.64, + 0.634 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.642, + 0.907, + 0.824 + ], + "angle": 0, + "content": "Datasets and metrics To validate the effectiveness of IRGS, we conduct extensive experiments on multiple datasets for the inverse rendering task, including two synthetic datasets: Synthetic4Relight dataset [43] and the TensoIR dataset [15], as well as a real-world dataset: Stanford-ORB dataset [17]. For quantitative comparisons, we follow previous works [10, 19] and employ various metrics to evaluate performance across different outputs. Specifically, we use PSNR, SSIM [30], and LPIPS [41] for assessing image quality in novel view synthesis, albedo, and relighting tasks. For normal and roughness, we use mean angular error (MAE) and mean squared error (MSE), respectively." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.825, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Implementation details As detailed in Section 3.3, our training process consists of two stages. The first stage is trained over 40,000 iterations with the hyperparameters \\(\\lambda_{\\mathrm{n}}\\), \\(\\lambda_{\\mathrm{d}}\\), \\(\\lambda_{\\mathrm{s,n}}\\), and \\(\\lambda_{\\mathrm{o}}\\) set to 0.05, 1000, 0.02, and 0.01, respectively. The second stage continues for an additional 20,000" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10948" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.904, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.253, + 0.908, + 0.298 + ], + "angle": 0, + "content": "Figure 5. Visualization of estimated components in incident light, including the averaged direct radiance \\( L_{\\mathrm{dir}} \\), indirect radiance \\( L_{\\mathrm{ind}} \\), visibility \\( V \\) (ambient occlusion), incident radiance \\( L_{\\mathrm{i}} \\). Compared to R3DG [10], IRGS achieves notably more realistic results, especially in estimated indirect light, due to its accurate modeling of inter-reflections." + }, + { + "type": "image", + "bbox": [ + 0.108, + 0.304, + 0.889, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.265, + 0.465, + 0.731, + 0.479 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparison of relighting results on TensoIR dataset [43]." + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.489, + 0.46, + 0.503 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison on TensoIR dataset [15]." + }, + { + "type": "table", + "bbox": [ + 0.098, + 0.504, + 0.474, + 0.606 + ], + "angle": 0, + "content": "
MethodNormal MAE↓NVS PSNR↑Albedo PSNR↑Relight PSNR↑
NeRFactor [42]6.31424.67925.12523.383
InvRender [43]5.07427.36727.34123.973
TensoIR [15]4.10035.08829.27528.580
GS-IR [19]4.94835.33330.28624.374
R3DG [10]5.92737.34326.19927.367
IRGS (Ours)4.11235.87633.79629.907
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.628, + 0.486, + 0.901 + ], + "angle": 0, + "content": "iterations, with \\(\\lambda_{1}^{\\mathrm{pbr}}\\), \\(\\lambda_{\\mathrm{light}}\\), \\(\\lambda_{\\mathrm{s,a}}\\), and \\(\\lambda_{\\mathrm{s,r}}\\) set to 1.0, 0.01, 2.0, and 2.0, respectively. We implement 2D Gaussian ray tracing in OptiX [25] via PyTorch CUDA extensions, updating the BVH at each training iteration for the varying geometry, with each update taking approximately 3 ms. In line with 3DGRT [22], we use a buffer of \\(k = 16\\) for per-ray sorting and terminate ray tracing when transmittance falls below 0.03. We employ \\(N_{\\mathrm{r}} = 256\\) rays for rendering equation evaluation, and each training iteration samples up to \\(N_{\\mathrm{rays}} = 2^{18}\\) rays. At \\(N_{\\mathrm{r}} = 256\\), rendering a complete image takes about 1 second. For environmental lighting, we use a \\(32 \\times 32\\) cubemap during optimization. The learning rates for albedo, roughness, and the cubemap are set to 0.005, 0.005, and 0.01, respectively, with other hyperparameters following those in 2DGS [13]. The entire training process takes around 40 minutes, with 15 minutes for the first stage and 25 minutes for the second stage. All experiments are conducted on a single NVIDIA RTX 3090 GPU." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.49, + 0.608, + 0.505 + ], + "angle": 0, + "content": "4.1. Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.518, + 0.907, + 0.867 + ], + "angle": 0, + "content": "Synthetic4Relight For experiments on Synthetic4Relight [43], we evaluate IRGS's performance in novel view synthesis (NVS), relighting, material estimation, and efficiency. In Tab. 1, our approach achieves higher quality than previous methods across relighting, and material estimation, demonstrating IRGS's ability to accurately capture material details. While R3DG [10] achieves higher performance in NVS, this is primarily due to its shading on Gaussians, which enhances NVS capability at the expense of relighting performance. IRGS also completes optimization in a relatively short time of 0.7 hours, emphasizing its efficiency in inverse rendering applications. In Fig. 4, we provide a qualitative comparison of the reconstructed \"chair\" scene against GS-based competitors [10, 19], visualizing NVS, albedo, roughness, estimated environment maps, and relighting images. Our results demonstrate high-fidelity material estimation and realistic relighting outcomes. Fig. 5 further illustrates a comparison of the estimated incident light components between IRGS and R3DG [10], covering direct radiance, indirect radiance, ambient occlusion, and incident radiance. IRGS achieves more reasonable indirect radiance and clearer ambient occlusion, enhancing the overall realism." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.909, + 0.903 + ], + "angle": 0, + "content": "TensoIR For TensoIR dataset [15], we report metrics for NVS, relighting, and albedo estimation in Tab. 2. Com" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10949" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.094, + 0.466, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.112, + 0.32, + 0.462, + 0.334 + ], + "angle": 0, + "content": "Figure 7. Relighting results on Stanford-ORB dataset [17]." + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.339, + 0.473, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.163, + 0.484, + 0.411, + 0.497 + ], + "angle": 0, + "content": "(a) Ablation studies on indirect light modeling." + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.5, + 0.47, + 0.621 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.149, + 0.625, + 0.425, + 0.637 + ], + "angle": 0, + "content": "(b) Ablation studies on number of sampled rays \\(N_{\\mathrm{r}}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.11, + 0.64, + 0.462, + 0.653 + ], + "angle": 0, + "content": "Figure 8. Ablation studies on various components of IRGS." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.654, + 0.483, + 0.774 + ], + "angle": 0, + "content": "pared to prior arts [10, 15, 19, 42, 43], our method consistently outperforms in relighting and albedo estimation, and achieving comparable performance in normal quality with TensoIR [15]. Similar to Synthetic4Relight, R3DG [10] achieves higher accuracy in NVs. Fig. 6 presents a qualitative comparison of relighting images generated by GS-based methods [10, 19], showing that IRGS produces more realistic relighting results, with natural shadows." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.775, + 0.483, + 0.821 + ], + "angle": 0, + "content": "Stanford-ORB dataset We also conduct experiments on the real-world Stanford-ORB dataset [17]. As shown in Fig. 7, IRGS produces realistic relighting results." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.832, + 0.241, + 0.848 + ], + "angle": 0, + "content": "4.2. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.903 + ], + "angle": 0, + "content": "In Tab. 3, we conduct ablation studies on various components of IRGS to validate their effectiveness. We report the average PSNR on NVS, albedo estimation, and relighting" + }, + { + "type": "table_caption", + "bbox": [ + 0.536, + 0.09, + 0.882, + 0.103 + ], + "angle": 0, + "content": "Table 3. Ablation studies on various components of IRGS." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.104, + 0.903, + 0.188 + ], + "angle": 0, + "content": "
MethodNVS PSNR ↑Albedo PSNR ↑Relighting PSNR ↑
Detach indirect34.2130.2934.22
w/o indirect (train)34.0930.1033.93
w/o indirect (relight)--33.84
Nr=1634.0130.2129.46
Nr=6434.9830.6333.11
Full35.4830.8134.68
" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.21, + 0.86, + 0.225 + ], + "angle": 0, + "content": "across four scenes in Synthetic4Relight dataset [43]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.229, + 0.907, + 0.456 + ], + "angle": 0, + "content": "Indirect light We perform ablation studies to evaluate the impact of indirect light modeling. \"Detach indirect\" indicates that gradients are not backpropagated through ray tracing to optimize indirect radiance. \"w/o indirect (train)\" signifies that the indirect term in Eq. (13) is omitted during training, while \"w/o indirect (relight)\" omits the proposed strategy for computing indirect radiance during relighting. As shown in Fig. 8a, detaching the indirect term results in unrealistic indirect radiance, and training without the indirect term leads to inaccurate albedo estimation. We also demonstrate the impact of indirect radiance on relighting results of a composited scene in Fig. 1 (indirect illumination). Note that the renderings in Fig. 1 are produced by assigning an additional high metallic to the Gaussian-based \"Ground\", enhancing the visual clarity of inter-reflections." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.461, + 0.907, + 0.537 + ], + "angle": 0, + "content": "Number of rays We also ablate on the number of sampled rays \\( N_{\\mathrm{r}} \\) for rendering equation. We investigated the effects of \\( N_{\\mathrm{r}} = 16 \\), \\( N_{\\mathrm{r}} = 64 \\), and \\( N_{\\mathrm{r}} = 256 \\). As shown in Fig. 8b, more sampled rays lead to more accurate and smooth renderings and ambient occlusion in novel view synthesis." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.564, + 0.634, + 0.58 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.595, + 0.907, + 0.821 + ], + "angle": 0, + "content": "In this paper, we introduce inter-reflective Gaussian splating (IRGS), a novel framework that overcomes the limitations of previous Gaussian-based approaches in accurately modeling visibility and indirect radiance for incident light in inverse rendering. IRGS incorporates the full rendering equation without simplifications and utilizes a differentiable 2D Gaussian ray tracing technique for precise computation of incident radiance at intersection points. We further propose an optimization scheme to address the computational demands of Monte Carlo sampling in rendering equation evaluation, and a novel strategy for querying indirect radiance during relighting. Extensive experiments on multiple benchmarks validate IRGS's ability to accurately model complex inter-reflection, establishing it as an effective solution for inverse rendering." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.9 + ], + "angle": 0, + "content": "Limitation Due to the significant computational demands of rendering equation evaluation, we are unable to render at a high frame rate. However, this could be addressed by baking the incident radiance or pre-computing radiance transfer, which we leave as future work." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10950" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.484, + 0.147 + ], + "angle": 0, + "content": "This work was supported in part by National Natural Science Foundation of China (Grant No. 62376060)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.159, + 0.188, + 0.175 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.184, + 0.484, + 0.239 + ], + "angle": 0, + "content": "[1] Benjamin Attal, Dor Verbin, Ben Mildenhall, Peter Hedman, Jonathan T Barron, Matthew O'Toole, and Pratul P Srinivasan. Flash cache: Reducing bias in radiance cache based inverse rendering. In ECCV, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.241, + 0.484, + 0.295 + ], + "angle": 0, + "content": "[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.297, + 0.483, + 0.337 + ], + "angle": 0, + "content": "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.339, + 0.483, + 0.38 + ], + "angle": 0, + "content": "[4] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.382, + 0.483, + 0.435 + ], + "angle": 0, + "content": "[5] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T. Barron, Ce Liu, and Hendrik P.A. Lensch. Nerd: Neural reflectance decomposition from image collections. In ICCV, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.438, + 0.483, + 0.505 + ], + "angle": 0, + "content": "[6] Mark Boss, Andreas Engelhardt, Abhishek Kar, Yuzhen Li, Deqing Sun, Jonathan Barron, Hendrik Lensch, and Varun Jampani. *Samurai: Shape and material from unconstrained real-world arbitrary image collections*. *NeurIPS*, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.508, + 0.482, + 0.548 + ], + "angle": 0, + "content": "[7] Brent Burley and Walt Disney Animation Studios. Physically-based shading at disney. In Acm Siggraph, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.551, + 0.483, + 0.603 + ], + "angle": 0, + "content": "[8] Yurui Chen, Chun Gu, Junzhe Jiang, Xiatian Zhu, and Li Zhang. Periodic vibration gaussian: Dynamic urban scene reconstruction and real-time rendering. arXiv preprint, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.607, + 0.483, + 0.647 + ], + "angle": 0, + "content": "[9] Jan-Niklas Dihlmann, Arjun Majumdar, Andreas Engelhardt, Raphael Braun, and Hendrik Lensch. Subsurface scattering for 3d gaussian splattering. arXiv preprint, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.649, + 0.483, + 0.704 + ], + "angle": 0, + "content": "[10] Jian Gao, Chun Gu, Youtian Lin, Hao Zhu, Xun Cao, Li Zhang, and Yao Yao. Relightable 3d gaussian: Real-time point cloud relighting with brdf decomposition and ray tracing. arXiv preprint, 2023. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.705, + 0.483, + 0.76 + ], + "angle": 0, + "content": "[11] Yijia Guo, Yuanxi Bai, Liwen Hu, Ziyi Guo, Mianzhi Liu, Yu Cai, Tiejun Huang, and Lei Ma. Prtgs: Precomputed radiance transfer of gaussian splats for real-time high-quality relighting. In ACMMM, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.761, + 0.483, + 0.802 + ], + "angle": 0, + "content": "[12] Yuan-Chen Guo, Di Kang, Linchao Bao, Yu He, and Song-Hai Zhang. Nerfren: Neural radiance fields with reflections. In CVPR, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[13] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In SIGGRAPH, 2024. 2, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[14] Yingwenqi Jiang, Jiadong Tu, Yuan Liu, Xifeng Gao, Xiaoxiao Long, Wenping Wang, and Yuexin Ma. Gaussianshader: 3d gaussian splatting with shading functions for reflective surfaces. arXiv preprint, 2023. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.184, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[15] Haian Jin, Isabella Liu, Peijia Xu, Xiaoshuai Zhang, Songfang Han, Sai Bi, Xiaowei Zhou, Zexiang Xu, and Hao Su. Tensoir: Tensorial inverse rendering. In CVPR, 2023. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[16] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. TOG, 2023. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.192, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[17] Zhengfei Kuang, Yunzhi Zhang, Hong-Xing Yu, Samir Agarwala, Elliott Wu, Jiajun Wu, et al. Stanford-orb: a real-world 3d object inverse rendering benchmark. NeurIPS, 2024. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.234, + 0.905, + 0.287 + ], + "angle": 0, + "content": "[18] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.288, + 0.905, + 0.328 + ], + "angle": 0, + "content": "[19] Zhihao Liang, Qi Zhang, Ying Feng, Ying Shan, and Kui Jia. Gs-ir: 3d gaussian splatting for inverse rendering. In CVPR, 2024. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.331, + 0.905, + 0.385 + ], + "angle": 0, + "content": "[20] Yuan Liu, Peng Wang, Cheng Lin, Xiaoxiao Long, Jiepeng Wang, Lingjie Liu, Taku Komura, and Wenping Wang. Nano: Neural geometry and brdf reconstruction of reflective objects from multiview images. In SIGGRAPH, 2023. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.386, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[21] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.51 + ], + "angle": 0, + "content": "[22] Nicolas Moënne-Loccoz, Ashkan Mirzaei, Or Perel, Ricardo de Lutio, Janick Martinez Esturo, Gavriel State, Sanja Fidler, Nicholas Sharp, and Zan Gojcic. 3d gaussian ray tracing: Fast tracing of particle scenes. arXiv preprint, 2024. 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.511, + 0.905, + 0.553 + ], + "angle": 0, + "content": "[23] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.554, + 0.905, + 0.608 + ], + "angle": 0, + "content": "[24] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting triangular 3d models, materials, and lighting from images. In CVPR, 2022. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.61, + 0.905, + 0.677 + ], + "angle": 0, + "content": "[25] Steven G Parker, James Bigler, Andreas Dietrich, Heiko Friedrich, Jared Hoberock, David Luebke, David McAllister, Morgan McGuire, Keith Morley, Austin Robison, et al. Optix: a general purpose ray tracing engine. TOG, 2010. 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.905, + 0.719 + ], + "angle": 0, + "content": "[26] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mllps. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.721, + 0.905, + 0.787 + ], + "angle": 0, + "content": "[27] Yahao Shi, Yanmin Wu, Chenming Wu, Xing Liu, Chen Zhao, Haocheng Feng, Jingtuo Liu, Liangjun Zhang, Jian Zhang, Bin Zhou, et al. Gir: 3d gaussian inverse rendering for relightable scene factorization. arXiv preprint, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[28] Pratul P Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T Barron. Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In CVPR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[29] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint, 2021. 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10951" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[30] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Trans. Image Process., 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.189 + ], + "angle": 0, + "content": "[31] Tong Wu, Jia-Mu Sun, Yu-Kun Lai, Yuewen Ma, Leif Kobbelt, and Lin Gao. Deferreds: Decoupled and editable gaussian splatting with deferred shading. arXiv preprint, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[32] Yunzhi Yan, Haotong Lin, Chenxu Zhou, Weijie Wang, Haiyang Sun, Kun Zhan, Xianpeng Lang, Xiaowei Zhou, and Sida Peng. Street gaussians for modeling dynamic urban scenes. arXiv preprint, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.303 + ], + "angle": 0, + "content": "[33] Ziyi Yang, Yanzhen Chen, Xinyu Gao, Yazhen Yuan, Yu Wu, Xiaowei Zhou, and Xiaogang Jin. Sire-ir: Inverse rendering for brdf reconstruction with shadow and illumination removal in high-illuminance scenes. arXiv preprint, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.305, + 0.482, + 0.347 + ], + "angle": 0, + "content": "[34] Zeyu Yang, Hongye Yang, Zijie Pan, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. arXiv preprint, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.482, + 0.402 + ], + "angle": 0, + "content": "[35] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.404, + 0.482, + 0.459 + ], + "angle": 0, + "content": "[36] Yao Yao, Jingyang Zhang, Jingbo Liu, Yihang Qu, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Neilf: Neural incident light field for physically-based material estimation. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.461, + 0.482, + 0.489 + ], + "angle": 0, + "content": "[37] Yuxuan Yao, Zixuan Zeng, Chun Gu, Xiatian Zhu, and Li Zhang. Reflective gaussian splatting. In ICLR, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.482, + 0.531 + ], + "angle": 0, + "content": "[38] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.482, + 0.573 + ], + "angle": 0, + "content": "[39] Zehao Yu, Torsten Sattler, and Andreas Geiger. Gaussian opacity fields: Efficient adaptive surface reconstruction in unbounded scenes. TOG, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.576, + 0.482, + 0.63 + ], + "angle": 0, + "content": "[40] Jingyang Zhang, Yao Yao, Shiwei Li, Jingbo Liu, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Neilf++: Inter-reflectable light fields for geometry and material estimation. In ICCV, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[41] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.482, + 0.73 + ], + "angle": 0, + "content": "[42] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM Trans. Graph., 2021. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.733, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[43] Yuanqing Zhang, Jiaming Sun, Xingyi He, Huan Fu, Rongfei Jia, and Xiaowei Zhou. Modeling indirect illumination for inverse rendering. In CVPR, 2022. 2, 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.773 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "10952" + } + ] +] \ No newline at end of file diff --git a/2025/IRGS_ Inter-Reflective Gaussian Splatting with 2D Gaussian Ray Tracing/ff8cd318-c293-446d-b3d4-b22f9fc186ef_origin.pdf b/2025/IRGS_ Inter-Reflective Gaussian Splatting with 2D Gaussian Ray Tracing/ff8cd318-c293-446d-b3d4-b22f9fc186ef_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..baa1aa63fffb13431eff1a5cc06e0a9ba5d67a11 --- /dev/null +++ b/2025/IRGS_ Inter-Reflective Gaussian Splatting with 2D Gaussian Ray Tracing/ff8cd318-c293-446d-b3d4-b22f9fc186ef_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:80275d8d697e846ade57fbe1f3e02cf1bd7ce7df0da76dc962d3c626b0cc1a1d +size 6471671 diff --git a/2025/IRGS_ Inter-Reflective Gaussian Splatting with 2D Gaussian Ray Tracing/full.md b/2025/IRGS_ Inter-Reflective Gaussian Splatting with 2D Gaussian Ray Tracing/full.md new file mode 100644 index 0000000000000000000000000000000000000000..393de9f56b2bd9f504518eaad1068765571cdb4d --- /dev/null +++ b/2025/IRGS_ Inter-Reflective Gaussian Splatting with 2D Gaussian Ray Tracing/full.md @@ -0,0 +1,309 @@ +# IRGS: Inter-Reflective Gaussian Splating with 2D Gaussian Ray Tracing + +Chun Gu $^{1,2}$ Xiaofei Wei $^{1,2}$ Zixuan Zeng $^{1}$ Yuxuan Yao $^{1}$ Li Zhang $^{1,2*}$ + +$^{1}$ School of Data Science, Fudan University $^{2}$ Shanghai Innovation Institute + +https://fudan-zvq.github.io/IRGS + +# Abstract + +In inverse rendering, accurately modeling visibility and indirect radiance for incident light is essential for capturing secondary effects. Due to the absence of a powerful Gaussian ray tracer, previous 3DGS-based methods have either adopted a simplified rendering equation or used learnable parameters to approximate incident light, resulting in inaccurate material and lighting estimations. To this end, we introduce inter-reflective Gaussian splatting (IRGS) for inverse rendering. To capture inter-reflection, we apply the full rendering equation without simplification and compute incident radiance on the fly using the proposed differentiable 2D Gaussian ray tracing. Additionally, we present an efficient optimization scheme to handle the computational demands of Monte Carlo sampling for rendering equation evaluation. Furthermore, we introduce a novel strategy for querying the indirect radiance of incident light when relighting the optimized scenes. Extensive experiments on multiple standard benchmarks validate the effectiveness of IRGS, demonstrating its capability to accurately model complex inter-reflection effects. + +# 1. Introduction + +Inverse rendering is a fundamental problem in computer vision and graphics, aiming to reconstruct geometry and estimate material and lighting from a set of posed images. The introduction of neural radiance field (NeRF) [21], which uses a neural implicit field modeled by a multi-layer perceptron (MLP) to represent 3D scenes, has inspired many NeRF-based methods [5, 15, 20, 28, 40] to address inverse rendering tasks. Leveraging NeRF's ray-based rendering technique, these methods can accurately compute the visibility and indirect radiance for incident light as required by the rendering equation. However, the exhaustive querying of a neural field during ray marching and its limited representational capacity constrain the efficiency and performance of neural implicit fields in inverse rendering. + +![](images/543abf89ba94525d5f3e15010162a76c61de896c8ac2469bb64c796c6d54566d.jpg) + +![](images/49ec55b658c17a7f302914ba24b45925ca2558bdc2ce662d38fef04af8097f57.jpg) +Figure 1. Global and indirect illumination in a Gaussian-based scene using our IRGS, highlighting its inter-reflective capabilities. + +Recently, 3D Gaussian splatting (3DGS) [16] has emerged as a promising technique for 3D scene representation, offering both remarkable rendering quality and efficiency. By modeling a static 3D scene as a collection of 3D Gaussians, 3DGS achieves real-time, high-quality rendering with a tile-based rasterizer. However, despite its strengths, 3DGS has notable limitations due to its reliance on rasterization, which limits its ability to accurately simulate ray-based effects that are essential for inverse rendering. As a result, some 3DGS-based inverse rendering methods [14, 19, 27] have adopted simplified versions of the rendering equation, such as the split-sum approximation [24]. Other approaches, such as R3DG [10], use learnable parameters to approximate indirect radiance, which fails to + +capture actual inter-reflection. These methods struggle to accurately model the rendering equation, limiting their effectiveness in realistic material and lighting estimation. + +Our goal is to integrate the full rendering equation into the efficient Gaussian splatting for accurate inverse rendering, enabling the capture of complex interreflection effects. This requires efficient and precise computation of visibility and indirect radiance for incident light. Drawing inspiration from 3DGRT [22], which achieves efficient ray tracing on 3D Gaussian primitives via a $k$ -buffer hits-based marching technique and hardware-accelerated NVIDIA OptiX interface [25], we explore adapting this technique to model inter-reflection within the Gaussian splatting framework. + +In this paper, we introduce inter-reflective Gaussian splatting (IRGS), a novel inverse rendering framework that leverages the proposed 2D Gaussian ray tracing technique to model inter-reflections within Gaussian splatting. To the best of our knowledge, this is the first method to integrate the full rendering equation without simplification into Gaussian splatting, accurately capturing inter-reflection effects by querying incident lighting through ray tracing. First, we propose 2D Gaussian ray tracing to efficiently and accurately trace across 2D Gaussian primitives, computing the accumulated opacity and color of arbitrary rays. Unlike 3DGRT [22], our approach features well-defined ray-splat intersections, enabling direct ray tracing on pretrained 2D Gaussian splatting [13] with minimal quality degradation. Next, we introduce the inter-reflective Gaussian splatting framework: starting from a pretrained 2D Gaussian splatting model, we apply the full rendering equation without simplification at intersection points determined by the depth map, while visibility and indirect radiance are computed on the fly using our 2D Gaussian ray tracing technique. Notably, our 2D Gaussian ray tracing is fully differentiable, enabling the optimization of indirect light through backpropagation. We further propose an efficient optimization scheme to manage the computational demands of the exhaustive Monte Carlo sampling for the rendering equation. Additionally, we introduce a novel strategy to query indirect radiance during relighting, overcoming limitations in previous methods that often omit this component. Extensive experiments on multiple standard benchmarks confirm the effectiveness of IRGS. In Fig. 1, we visualize the global illumination (considering both direct and indirect radiance) and the indirect illumination (considering only indirect radiance) of a Gaussian-based scene rendered using IRGS, demonstrating that IRGS is capable of modeling remarkable inter-reflection effects. + +The contributions of this paper are as follows: (i) 2D Gaussian ray tracing, a technique that enables direct ray tracing on pretrained 2D Gaussian splatting with minimal quality degradation; (ii) IRGS, an inverse rendering framework that incorporates the full rendering equation without + +simplification, using 2D Gaussian ray tracing to compute incident radiance on the fly; (iii) An optimization scheme that manages computational demands for rendering equation evaluation; (iv) A novel strategy that enables querying of indirect radiance during relighting. + +# 2. Related work + +Novel view synthesis aims to leverage a limited collection of observed images of a scene to generate new images from an unseen viewpoint. Neural radiance field (NeRF) [21] produces photo-realistic quality images by representing the scenes in terms of a volumetric radiance field encoded in a coordinate-based neural network. Subsequent advancements have focused on accelerating rendering[23, 26], enhancing image quality [2-4], and improving geometry reconstruction [18, 29, 38]. More recently, 3D Gaussian splatting (3DGS) [16] introduced the use of fuzzy, anisotropic 3D Gaussian point cloud for scene representation along with an efficient tile-based rasterizer, achieving state-of-the-art results in both rendering quality and efficiency. 3DGS has inspired numerous approaches targeting various goals, including geometry reconstruction [13, 39], dynamic scene reconstruction [34, 35], inverse rendering [10, 19], and street scene applications [8, 32]. However, the dependence on rasterization limits Gaussian splatting's ability to simulate ray-based effects accurately. To address this, 3DGRT [22] proposes a differentiable Gaussian ray tracer that efficiently performs ray tracing across 3D Gaussian primitives, allowing radiance computation along arbitrary rays. In this paper, we propose ray tracing on 2D Gaussian primitives. Unlike 3DGRT, 2D Gaussian primitives [13] have a well-defined ray-splat intersection, enabling direct ray tracing on pre-trained 2D Gaussian splatting with minimal quality degradation. + +Inverse rendering focuses on recovering geometry, material, and lighting properties from captured images, presenting a complex challenge due to high-dimensional light-surface interactions. Inspired by the success of neural radiance fields, many methods [1, 5, 6, 15, 20, 28, 33, 36, 40, 43] leverage flexible ray marching techniques and neural implicit fields for inverse rendering. Recent approaches [9-11, 14, 19, 27, 31, 37] have adapted 3D Gaussian splatting (3DGS) to inverse rendering by assigning material-related properties to each Gaussian primitive. Leveraging the rapid rendering capabilities of 3DGS, these methods achieve much higher efficiency compared to NeRF-based techniques. However, due to the lack of an efficient Gaussian ray tracer, existing 3DGS-based methods either apply a simplified version of the rendering equation [19] or use learnable parameters to model incident light [10], which limits the accuracy of material and lighting estimation. For example, GS-IR [19] employs split-sum approximation to + +![](images/4d68f25c23fae156617de34e6486be3db749414b84b377e9696048edad14b68a.jpg) +Figure 2. Schematic illustration of the proposed IRGS. Starting from a set of 2D Gaussians equipped with material properties, we apply rasterization to generate albedo, roughness, position, and normal maps. We then evaluate the rendering equation using stratified sampling at the corresponding position, drawing geometry and material values from these feature maps. The radiance of incident light is decomposed into direct radiance $L_{\mathrm{dir}}$ from the environment map, and indirect radiance $L_{\mathrm{ind}}$ with visibility $V$ , obtained via 2D Gaussian ray tracing. + +simplify the rendering equation and uses baked volumes to store occlusion and indirect illumination. R3DG [10] shades each Gaussian individually, using point-based ray tracing to bake occlusion and spherical harmonics to parameterize indirect radiance. To overcome these limitations, we propose the IRGS framework that utilizes 2D Gaussian ray tracing to accurately model inter-reflections. + +# 3. Method + +In this section, we introduce our inter-reflective Gaussian splatting with 2D Gaussian ray tracing, designed to reconstruct geometry and estimate material and lighting from a set of posed RGB images. We begin by providing the necessary background on 3D Gaussian splatting and the rendering equation (Section 3.1). We then present the 2D Gaussian ray tracing technique, which performs ray tracing on 2D Gaussian primitives (Section 3.2). Finally, we outline our IRGS framework, which leverages the proposed 2D Gaussian ray tracing technique to query visibility and indirect radiance for incident light (Section 3.3). + +# 3.1. Preliminary + +3D Gaussian splatting 3D Gaussian splatting (3DGS) represents a 3D scene using 3D Gaussian primitives, each defined as an ellipsoid characterized by a 3D mean $\mu$ and a covariance matrix $\Sigma$ . The influence of a Gaussian at a spatial position $x$ is expressed as: + +$$ +G (\boldsymbol {x}) = \exp \left(- \frac {1}{2} (\boldsymbol {x} - \boldsymbol {\mu}) ^ {\top} \Sigma^ {- 1} (\boldsymbol {x} - \boldsymbol {\mu})\right). \tag {1} +$$ + +Each Gaussian is also assigned an opacity $o$ and a view-dependent color $c$ modeled by spherical harmonics (SH). The rendering process of 3DGS starts by projecting 3D Gaussian primitives onto the 2D image plane. This is achieved by first applying a viewing transformation $W$ and then performing a perspective projection. The covariance matrix of the resulting 2D Gaussians can be approximated + +as $\Sigma' = JW\Sigma W^\top J^\top$ , where $J$ represents the Jacobian matrix of the perspective projection. The final image is rendered by alpha-blending these projected 2D Gaussians from front to back: + +$$ +\mathcal {C} = \sum_ {i = 1} ^ {N} T _ {i} \alpha_ {i} \boldsymbol {c} _ {i}, \quad T _ {i} = \prod_ {j = 1} ^ {i - 1} (1 - \alpha_ {j}), \tag {2} +$$ + +where $\alpha_{i}$ is calculated by multiplying opacity $o_i$ with each Gaussian's influence, determined by $\Sigma^{\prime}$ and the pixel coordinate. The covariance matrix $\Sigma$ is parameterized by a unit quaternion and a scaling vector to facilitate optimization. + +Rendering equation In physically-based rendering, illumination at a given surface point $\pmb{x}$ is computed via the rendering equation: + +$$ +L _ {o} \left(\boldsymbol {\omega} _ {o}, \boldsymbol {x}\right) = \int_ {\Omega} f \left(\boldsymbol {\omega} _ {o}, \boldsymbol {\omega} _ {i}, \boldsymbol {x}\right) L _ {1} \left(\boldsymbol {\omega} _ {i}, \boldsymbol {x}\right) \left(\boldsymbol {\omega} _ {i} \cdot \mathbf {n}\right) d \boldsymbol {\omega} _ {i}, \tag {3} +$$ + +where $\omega_{o}$ is the outgoing direction, $\omega_{i}$ is the incident direction over the hemisphere defined by the surface normal $\pmb{n}$ , $f$ is the bidirectional reflectance distribution function (BRDF), and $L_{\mathrm{i}}(\boldsymbol{\omega}_i,\boldsymbol{x})$ is the radiance of the incident light. In this work, we use a simplified Disney BRDF model [7] with only diffuse albedo $\pmb{a}$ and roughness $r$ as parameters, and assume the material is dielectric. The BRDF $f$ in Eq. (3) can be separated into a diffuse term $f_{d} = \frac{a}{\pi}$ and a specular term $f_{s}$ : + +$$ +f _ {s} \left(\boldsymbol {\omega} _ {i}, \boldsymbol {\omega} _ {o}, \boldsymbol {x}\right) = \frac {D F G}{4 \left(\boldsymbol {\omega} _ {i} \cdot \mathbf {n}\right) \left(\boldsymbol {\omega} _ {o} \cdot \mathbf {n}\right)}, \tag {4} +$$ + +where $D, F$ , and $G$ represent the normal distribution function, the Fresnel term, and the geometry term, respectively. + +# 3.2. 2D Gaussian ray tracing + +While 3DGS [16] has been successful in applications such as 3D reconstruction, achieving superior rendering quality + +![](images/46528fd0fe2cedbaf808364828dc2395595e7e26f092daded8dc70456a9a8ee2.jpg) +Figure 3. Performance of directly applying Gaussian ray tracing on a pretrained Gaussian splatting checkpoint in 3D Gaussian and 2D Gaussian cases, respectively. Our proposed 2DGRT achieves significantly less quality degradation than 3DGRT [22] in both quantitative metrics and visual results. + +and efficiency compared to NeRF-based methods, it is limited by the constraints of rasterization. These limitations are primarily related to ray-based effects, including rendering with highly distorted cameras and handling secondary ray effects. To address these challenges, 3DGRT [22] proposes performing ray tracing across 3D Gaussian primitives. This approach overcomes the inefficiencies of ray tracing on 3D Gaussian primitives by implementing a $k$ -buffer hits-based marching technique with a hardware-accelerated NVIDIA OptiX interface [25]. In this paper, as our focus is on inverse rendering, performing ray tracing on Gaussian primitives is essential to accurately compute inter-reflection effects when evaluating the rendering equation, including visibility and indirect radiance of incident light. + +3D Gaussian ray tracing [22] performs ray tracing on 3D Gaussian primitives and determines the ray-splat intersection as the point of maximum response from the corresponding 3D Gaussian along the ray. This behavior differs from that of 3D Gaussian splatting, which computes the particle response in 2D image space, leaving the position of the 3D intersection point undefined. To leverage both the efficiency of Gaussian splatting and the flexible rendering capabilities of Gaussian ray tracing, we aim to perform ray tracing directly on the Gaussian splatting model for incident radiance inference. However, as shown in Fig. 3, applying ray tracing directly on a 3DGS checkpoint results in a noticeable decline in rendering quality compared to splatting. To address this issue, we propose 2D Gaussian ray tracing (2DGRT), which performs ray tracing on 2D Gaussian primitives [13]. Since 2D Gaussian disks have a well-defined ray-splat intersection, this approach eliminates inconsistencies in intersection points present in 3D Gaussians. + +Given a 2D Gaussian [13] defined by its center point $\pmb{\mu} \in \mathbb{R}^3$ , an opacity parameter $o \in [0,1]$ , two principal tangential vectors $\pmb{t}_u \in \mathbb{R}^3$ and $\pmb{t}_v \in \mathbb{R}^3$ , and a scaling vector $\pmb{s} = (s_u, s_v) \in \mathbb{R}^2$ , we follow 3DGRT [22] to fit a bounding proxy of each 2D Gaussian using an adaptive icosahedron mesh, ensuring that a minimum influence of $\alpha_{\mathrm{min}}$ is enclosed. The vertices $\pmb{v}$ of the bounding icosahedron mesh are obtained by transforming a regular icosahedron with a + +unit inner sphere: + +$$ +\boldsymbol {v} \leftarrow \sqrt {2 \log \left(o / \alpha_ {\min }\right)} \left(s _ {u} \boldsymbol {t} _ {u} \quad s _ {v} \boldsymbol {t} _ {v} \quad \epsilon \mathbf {1}\right) \boldsymbol {v} + \boldsymbol {\mu}, \tag {5} +$$ + +where $\epsilon$ denotes a small positive number. Each icosahedron mesh consists of 20 triangular faces, resulting in a total of $20N$ triangles for $N$ 2D Gaussian primitives. To leverage the hardware-accelerated ray-triangle intersection provided by OptiX [25], we construct a Bounding Volume Hierarchy (BVH) using these triangles. By applying the $k$ -buffer per-ray sorting algorithm [22], we obtain the exact ordering of intersected Gaussians for each ray. To ensure consistency with 2DGS [13], we analytically compute the ray-splat intersection point $p$ : + +$$ +\boldsymbol {p} = \boldsymbol {r} _ {o} + \tau \boldsymbol {r} _ {d}, \quad \text {w h e r e} \quad \tau = \frac {\boldsymbol {n} ^ {\top} (\boldsymbol {\mu} - \boldsymbol {r} _ {o})}{\boldsymbol {n} ^ {\top} \boldsymbol {r} _ {d}}. \tag {6} +$$ + +where $\pmb{n} = \pmb{t}_u\times \pmb{t}_v$ is the normal vector of the 2D Gaussian disk, and $\pmb{r}_o$ and $\pmb{r}_d$ represent the origin and direction of the given ray, respectively. Thus, the influence of the current 2D Gaussian primitive is given by: + +$$ +G (\boldsymbol {p}) = \exp \left(- \frac {u ^ {2} + v ^ {2}}{2}\right), \tag {7} +$$ + +where $u = \frac{1}{s_v} (\pmb {p} - \pmb {\mu})^\top \pmb {t}_u,v = \frac{1}{s_v} (\pmb {p} - \pmb {\mu})^\top \pmb {t}_v$ + +Consequently, the RGB value and accumulated opacity value of the ray can be obtained by alpha-blending (Eq. (2)) the intersected 2D Gaussians from front to back: + +$$ +\left(\boldsymbol {c} _ {\mathrm {r t}}, o _ {\mathrm {r t}}\right) \leftarrow \operatorname {T r a c e} \left(\boldsymbol {r} _ {o}, \boldsymbol {r} _ {d}\right), \tag {8} +$$ + +where the RGB value of each Gaussian is modeled by the view-dependent color $c$ as in vanilla 2D Gaussian splatting. It is worth noting that subtle differences remain between 2D Gaussian ray tracing and 2DGS due to sorting errors introduced by per-tile sorting in 2DGS. However, as shown in Fig. 3, the quality degradation from 2D Gaussian ray tracing is minimal and significantly less noticeable than that observed in 3D Gaussians. + +# 3.3. Inter-reflective Gaussian splatting + +# 3.3.1 Stage I: 2D Gaussian pretraining + +Before estimating material and lighting for inverse rendering, we first pretrain a standard 2D Gaussian splatting model. This step is essential, as reliable geometry is required for inverse rendering, and evaluating the rendering equation is computationally intensive. Using a pretrained 2D Gaussian splatting model can greatly accelerate the training process. Along with the RGB rendering described in Eq. (2), we also render normalized depth and normal maps as follows: + +$$ +\left\{\mathcal {D}, \mathcal {N} \right\} = \sum_ {i = 1} ^ {N} w _ {i} \left\{d _ {i}, \boldsymbol {n} _ {i} \right\}, \text {w h e r e} w _ {i} = \frac {T _ {i} \alpha_ {i}}{\sum_ {i = 1} ^ {N} T _ {i} \alpha_ {i}}. \tag {9} +$$ + +Training loss Following 2DGS [13], we apply a normal consistency loss $\mathcal{L}_{\mathrm{n}}$ to ensure alignment between the rendered normal map and the underlying geometry, as well as a depth distortion loss [13] $\mathcal{L}_{\mathrm{d}}$ to encourage concentration of the 2D Gaussians. Additionally, we introduce an edge-aware smoothness loss [10, 12] $\mathcal{L}_{\mathrm{s,n}}$ on the normal map to enhance geometric robustness: + +$$ +\mathcal {L} _ {\mathrm {s}, \mathrm {n}} = \| \nabla \mathcal {N} \| \exp (- \| \nabla \mathcal {C} _ {\mathrm {g t}} \|), \tag {10} +$$ + +where $\mathcal{C}_{\mathrm{gt}}$ is the ground truth training image. To address the presence of floaters during standard 2DGS training, we also incorporate a binary cross-entropy loss [29] to constrain the geometry using the provided object mask $\mathcal{M}$ : + +$$ +\mathcal {L} _ {\mathrm {o}} = - \mathcal {M} \log \mathcal {O} - (1 - \mathcal {M}) \log (1 - \mathcal {O}), \tag {11} +$$ + +where $\mathcal{O} = \sum_{i=1}^{N} T_i \alpha_i$ , denotes the accumulated opacity map. Combined with the RGB reconstruction loss $\mathcal{L}_{\mathrm{c}}$ from 3DGS [16], the training loss for the first stage is defined as: + +$$ +\mathcal {L} ^ {1} = \mathcal {L} _ {\mathrm {c}} + \lambda_ {\mathrm {n}} \mathcal {L} _ {\mathrm {n}} + \lambda_ {\mathrm {d}} \mathcal {L} _ {\mathrm {d}} + \lambda_ {\mathrm {s}, \mathrm {n}} \mathcal {L} _ {\mathrm {s}, \mathrm {n}} + \lambda_ {\mathrm {o}} \mathcal {L} _ {\mathrm {o}}. \tag {12} +$$ + +# 3.3.2 Stage II: inverse rendering + +The order of shading and rasterization is crucial in physically based rendering with Gaussian splatting. Previous works, such as R3DG [10], propose performing shading on each Gaussian, taking advantage of properties modeled on each Gaussian, such as indirect radiance. However, since the rendered normal map is supervised at the pixel level after rasterization, shading on individual Gaussians can lead to inaccurate and blurred results. Consequently, we propose applying the rendering equation after rasterization. + +Rasterization As introduced in Section 3.1, a dielectric BRDF model is used for the rendering equation. We assign each 2D Gaussian primitive additional material properties: diffuse albedo $\pmb{a}$ and roughness $r$ . Pixel-level albedo map $\mathcal{A}$ and roughness map $\mathcal{R}$ are obtained through alpha-blending during rasterization: $\{\mathcal{A},\mathcal{R}\} = \sum_{i = 1}^{N}w_{i}\{a_{i},r_{i}\}$ . + +Previous inverse rendering methods using Gaussian splatting have lacked an efficient way to query the visibility and radiance for arbitrary incident directions. As a result, these methods typically apply simplified versions of the rendering equation [19] (e.g., split-sum approximation [24]) or use learnable parameters to model indirect light [10]. In this paper, we leverage the proposed 2D Gaussian ray tracing technique to efficiently and accurately query inter-reflection effects on the fly, including the visibility and indirect radiance of incident light. This allows us to apply the full rendering equation without simplification. + +Light parametrization We decompose the incident radiance at surface point $\pmb{x}$ with direction $\omega_{i}$ in Eq. (3) into + +direct radiance from infinity and indirect radiance: + +$$ +L _ {\mathrm {i}} \left(\boldsymbol {\omega} _ {i}, \boldsymbol {x}\right) = V \left(\boldsymbol {\omega} _ {i}, \boldsymbol {x}\right) L _ {\mathrm {d i r}} \left(\boldsymbol {\omega} _ {i}\right) + L _ {\mathrm {i n d}} \left(\boldsymbol {\omega} _ {i}, \boldsymbol {x}\right), \tag {13} +$$ + +where $L_{\mathrm{dir}}(\omega_i)$ is modeled using an environment cubemap, and $V(\omega_i, \boldsymbol{x})$ along with $L_{\mathrm{ind}}(\omega_i, \boldsymbol{x})$ are queried using 2D Gaussian ray tracing: + +$$ +\left(L _ {\text {i n d}} \left(\boldsymbol {\omega} _ {i}, \boldsymbol {x}\right), 1 - V \left(\boldsymbol {\omega} _ {i}, \boldsymbol {x}\right)\right) \leftarrow \operatorname {T r a c e} \left(\boldsymbol {x}, \boldsymbol {\omega} _ {i}\right) \tag {14} +$$ + +It is important to note that the RGB values used in ray tracing correspond to the view-dependent color $c$ from the first stage. Additionally, our 2D Gaussian ray tracing approach is fully differentiable, allowing gradients to propagate through the ray tracing process to optimize the indirect radiance of the incident ray. + +Rendering From the rendered depth and normal maps, we can determine the surface position $\mathbf{x}$ and the corresponding surface normal vector $\mathbf{n}$ for each pixel coordinate. Monte Carlo sampling is used to evaluate the rendering equation. Since we focus on reconstructing objects with dielectric materials, we employ stratified sampling to uniformly sample $N_{\mathrm{r}}$ incident directions across the hemisphere. The final rendering result is then computed as: + +$$ +\mathbf {c} _ {\mathrm {p b r}} = \frac {2 \pi}{N _ {\mathrm {r}}} \sum_ {i = 1} ^ {N _ {\mathrm {r}}} \left(f _ {\mathrm {d}} + f _ {\mathrm {s}}\right) L _ {\mathrm {i}} \left(\boldsymbol {\omega} _ {i}, \boldsymbol {x}\right) \left(\boldsymbol {\omega} _ {i} \cdot \mathbf {n}\right), \tag {15} +$$ + +Training loss In addition to the training losses from the first stage, we apply regularization to facilitate material and lighting estimation. Given the prior that diffuse light $L_{\mathrm{diffuse}} = \frac{1}{N_{\mathrm{r}}} \sum_{i=1}^{N_{\mathrm{r}}} L(\boldsymbol{\omega}_i, \boldsymbol{x})$ exhibits a natural white radiance, we minimize the difference between the RGB values [20]: + +$$ +\mathcal {L} _ {\text {l i g h t}} = \sum_ {c} \left(L _ {\text {d i f f u s e}} - \frac {1}{3} \sum_ {c} L _ {\text {d i f f u s e}}\right), c \in \{R, G, B \}. \tag {16} +$$ + +Similar to Eq. (10), we also apply edge-aware smoothness regularization $\mathcal{L}_{\mathrm{s,a}}$ and $\mathcal{L}_{\mathrm{s,r}}$ to the rendered albedo and roughness map, respectively, to ensure smooth material estimation. Additionally, for training efficiency, we apply the rendering equation to only a subset of pixels in each iteration. We set a maximum of $N_{\mathrm{rays}}$ rays per iteration, meaning only $\lfloor N_{\mathrm{rays}} / N_{\mathrm{r}}\rfloor$ pixels are randomly selected for final rendering equation evaluation. This scheme allows us to use a large $N_{\mathrm{r}}$ when evaluating the rendering equation, significantly enhancing estimation quality. Consequently, only a L1 loss $\mathcal{L}_1^{\mathrm{pbr}}$ is used to supervise the final color $c_{\mathrm{pbr}}$ . It is important to note that the rendered image $\mathcal{C}$ from the vanilla 2DGS is still optimized, while the view-dependent color $c$ associated with each Gaussian is used to compute radiance in 2D Gaussian ray tracing. The total loss for the second stage is given by: + +$$ +\mathcal {L} ^ {2} = \mathcal {L} ^ {1} + \lambda_ {1} ^ {\mathrm {p b r}} \mathcal {L} _ {1} ^ {\mathrm {p b r}} + \lambda_ {\text {l i g h t}} \mathcal {L} _ {\text {l i g h t}} + \lambda_ {\mathrm {s}, \mathrm {a}} \mathcal {L} _ {\mathrm {s}, \mathrm {a}} + \lambda_ {\mathrm {s}, \mathrm {r}} \mathcal {L} _ {\mathrm {s}, \mathrm {r}}, \tag {17} +$$ + +![](images/a24d95e65a3050f9c226f23018eff22a373148a8f331dc6379bc25ab0adc6d14.jpg) +Figure 4. Qualitative comparison of NVS, material and lighting estimation, and relighting results on the Synthetic4Relight dataset [43]. + +Table 1. Quantitative comparison on the Synthetic4Relight dataset [43]. A higher intensity of the red color signifies a better result. + +
Novel view synthesisRelightingAlbedoRoughness MSE↓Time (hours)
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
NeRFactor [42]22.800.9160.15021.540.8750.17119.490.8640.206N/A>48
InvRender [43]30.740.9530.08628.670.9500.09128.280.9350.0720.00814
TensorIR [15]35.800.9780.04929.690.9510.07930.580.9460.0650.0153
GS-IR [19]33.950.9650.05725.400.9240.08319.480.8960.1170.0110.4
R3DG [10]36.800.9820.02831.000.9640.05028.310.9510.0580.0130.9
IRGS (Ours)35.480.9740.04334.900.9690.04830.810.9570.0550.0080.7
+ +Relighting When performing relighting, changes in environmental lighting make the optimized radiance $c$ of each Gaussian no longer applicable for indirect radiance in 2D Gaussian ray tracing. Unlike previous methods [10, 19] that omit indirect lighting during relighting, we propose using the split-sum approximation [24] to efficiently query indirect radiance without recursive evaluation. Specifically, we apply ray tracing on each incident ray to alpha-blend the albedo, roughness, and normal values, resulting in $\mathcal{A}_{\mathrm{r}}$ , $\mathcal{R}_{\mathrm{r}}$ , and $\mathcal{N}_{\mathrm{r}}$ . These aggregated values are then shaded using a pre-filtered environment map, following the approach in Nvdiffrec [24]. Additionally, we apply importance sampling on the intensity of the ground truth environment map instead of stratified sampling, enabling more accurate rendering equation evaluation by concentrating samples in high-intensity directions. Please refer to the appendix for further details. + +# 4. Experiment + +Datasets and metrics To validate the effectiveness of IRGS, we conduct extensive experiments on multiple datasets for the inverse rendering task, including two synthetic datasets: Synthetic4Relight dataset [43] and the TensoIR dataset [15], as well as a real-world dataset: Stanford-ORB dataset [17]. For quantitative comparisons, we follow previous works [10, 19] and employ various metrics to evaluate performance across different outputs. Specifically, we use PSNR, SSIM [30], and LPIPS [41] for assessing image quality in novel view synthesis, albedo, and relighting tasks. For normal and roughness, we use mean angular error (MAE) and mean squared error (MSE), respectively. + +Implementation details As detailed in Section 3.3, our training process consists of two stages. The first stage is trained over 40,000 iterations with the hyperparameters $\lambda_{\mathrm{n}}$ , $\lambda_{\mathrm{d}}$ , $\lambda_{\mathrm{s,n}}$ , and $\lambda_{\mathrm{o}}$ set to 0.05, 1000, 0.02, and 0.01, respectively. The second stage continues for an additional 20,000 + +![](images/eef7f716dce66220e89a9df2cd24762582a2adc67a534649135886c79fa6988e.jpg) +Figure 5. Visualization of estimated components in incident light, including the averaged direct radiance $L_{\mathrm{dir}}$ , indirect radiance $L_{\mathrm{ind}}$ , visibility $V$ (ambient occlusion), incident radiance $L_{\mathrm{i}}$ . Compared to R3DG [10], IRGS achieves notably more realistic results, especially in estimated indirect light, due to its accurate modeling of inter-reflections. + +![](images/6e490e3a8e5d7b425dd3ab2b7e03202bdf97c329ee11b22934bead54b94e1fa9.jpg) +Figure 6. Qualitative comparison of relighting results on TensoIR dataset [43]. + +Table 2. Quantitative comparison on TensoIR dataset [15]. + +
MethodNormal MAE↓NVS PSNR↑Albedo PSNR↑Relight PSNR↑
NeRFactor [42]6.31424.67925.12523.383
InvRender [43]5.07427.36727.34123.973
TensoIR [15]4.10035.08829.27528.580
GS-IR [19]4.94835.33330.28624.374
R3DG [10]5.92737.34326.19927.367
IRGS (Ours)4.11235.87633.79629.907
+ +iterations, with $\lambda_{1}^{\mathrm{pbr}}$ , $\lambda_{\mathrm{light}}$ , $\lambda_{\mathrm{s,a}}$ , and $\lambda_{\mathrm{s,r}}$ set to 1.0, 0.01, 2.0, and 2.0, respectively. We implement 2D Gaussian ray tracing in OptiX [25] via PyTorch CUDA extensions, updating the BVH at each training iteration for the varying geometry, with each update taking approximately 3 ms. In line with 3DGRT [22], we use a buffer of $k = 16$ for per-ray sorting and terminate ray tracing when transmittance falls below 0.03. We employ $N_{\mathrm{r}} = 256$ rays for rendering equation evaluation, and each training iteration samples up to $N_{\mathrm{rays}} = 2^{18}$ rays. At $N_{\mathrm{r}} = 256$ , rendering a complete image takes about 1 second. For environmental lighting, we use a $32 \times 32$ cubemap during optimization. The learning rates for albedo, roughness, and the cubemap are set to 0.005, 0.005, and 0.01, respectively, with other hyperparameters following those in 2DGS [13]. The entire training process takes around 40 minutes, with 15 minutes for the first stage and 25 minutes for the second stage. All experiments are conducted on a single NVIDIA RTX 3090 GPU. + +# 4.1. Results + +Synthetic4Relight For experiments on Synthetic4Relight [43], we evaluate IRGS's performance in novel view synthesis (NVS), relighting, material estimation, and efficiency. In Tab. 1, our approach achieves higher quality than previous methods across relighting, and material estimation, demonstrating IRGS's ability to accurately capture material details. While R3DG [10] achieves higher performance in NVS, this is primarily due to its shading on Gaussians, which enhances NVS capability at the expense of relighting performance. IRGS also completes optimization in a relatively short time of 0.7 hours, emphasizing its efficiency in inverse rendering applications. In Fig. 4, we provide a qualitative comparison of the reconstructed "chair" scene against GS-based competitors [10, 19], visualizing NVS, albedo, roughness, estimated environment maps, and relighting images. Our results demonstrate high-fidelity material estimation and realistic relighting outcomes. Fig. 5 further illustrates a comparison of the estimated incident light components between IRGS and R3DG [10], covering direct radiance, indirect radiance, ambient occlusion, and incident radiance. IRGS achieves more reasonable indirect radiance and clearer ambient occlusion, enhancing the overall realism. + +TensoIR For TensoIR dataset [15], we report metrics for NVS, relighting, and albedo estimation in Tab. 2. Com + +![](images/656eec5ab206eec445bcdcd4a06138af8bab875376cb8a8778f0986ac1391988.jpg) +Figure 7. Relighting results on Stanford-ORB dataset [17]. + +![](images/989f5d0a6f4e3c99ed98ade289bb9bb0a6b368da7a327fa1792f9cc6bcce82d0.jpg) +(a) Ablation studies on indirect light modeling. + +![](images/293ec42bb39ff67cc490758100f568f4d4227e0eab89e9e74b91651455e471be.jpg) +(b) Ablation studies on number of sampled rays $N_{\mathrm{r}}$ +Figure 8. Ablation studies on various components of IRGS. + +pared to prior arts [10, 15, 19, 42, 43], our method consistently outperforms in relighting and albedo estimation, and achieving comparable performance in normal quality with TensoIR [15]. Similar to Synthetic4Relight, R3DG [10] achieves higher accuracy in NVs. Fig. 6 presents a qualitative comparison of relighting images generated by GS-based methods [10, 19], showing that IRGS produces more realistic relighting results, with natural shadows. + +Stanford-ORB dataset We also conduct experiments on the real-world Stanford-ORB dataset [17]. As shown in Fig. 7, IRGS produces realistic relighting results. + +# 4.2. Ablation study + +In Tab. 3, we conduct ablation studies on various components of IRGS to validate their effectiveness. We report the average PSNR on NVS, albedo estimation, and relighting + +Table 3. Ablation studies on various components of IRGS. + +
MethodNVS PSNR ↑Albedo PSNR ↑Relighting PSNR ↑
Detach indirect34.2130.2934.22
w/o indirect (train)34.0930.1033.93
w/o indirect (relight)--33.84
Nr=1634.0130.2129.46
Nr=6434.9830.6333.11
Full35.4830.8134.68
+ +across four scenes in Synthetic4Relight dataset [43]. + +Indirect light We perform ablation studies to evaluate the impact of indirect light modeling. "Detach indirect" indicates that gradients are not backpropagated through ray tracing to optimize indirect radiance. "w/o indirect (train)" signifies that the indirect term in Eq. (13) is omitted during training, while "w/o indirect (relight)" omits the proposed strategy for computing indirect radiance during relighting. As shown in Fig. 8a, detaching the indirect term results in unrealistic indirect radiance, and training without the indirect term leads to inaccurate albedo estimation. We also demonstrate the impact of indirect radiance on relighting results of a composited scene in Fig. 1 (indirect illumination). Note that the renderings in Fig. 1 are produced by assigning an additional high metallic to the Gaussian-based "Ground", enhancing the visual clarity of inter-reflections. + +Number of rays We also ablate on the number of sampled rays $N_{\mathrm{r}}$ for rendering equation. We investigated the effects of $N_{\mathrm{r}} = 16$ , $N_{\mathrm{r}} = 64$ , and $N_{\mathrm{r}} = 256$ . As shown in Fig. 8b, more sampled rays lead to more accurate and smooth renderings and ambient occlusion in novel view synthesis. + +# 5. Conclusion + +In this paper, we introduce inter-reflective Gaussian splating (IRGS), a novel framework that overcomes the limitations of previous Gaussian-based approaches in accurately modeling visibility and indirect radiance for incident light in inverse rendering. IRGS incorporates the full rendering equation without simplifications and utilizes a differentiable 2D Gaussian ray tracing technique for precise computation of incident radiance at intersection points. We further propose an optimization scheme to address the computational demands of Monte Carlo sampling in rendering equation evaluation, and a novel strategy for querying indirect radiance during relighting. Extensive experiments on multiple benchmarks validate IRGS's ability to accurately model complex inter-reflection, establishing it as an effective solution for inverse rendering. + +Limitation Due to the significant computational demands of rendering equation evaluation, we are unable to render at a high frame rate. However, this could be addressed by baking the incident radiance or pre-computing radiance transfer, which we leave as future work. + +# Acknowledgments + +This work was supported in part by National Natural Science Foundation of China (Grant No. 62376060). + +# References + +[1] Benjamin Attal, Dor Verbin, Ben Mildenhall, Peter Hedman, Jonathan T Barron, Matthew O'Toole, and Pratul P Srinivasan. Flash cache: Reducing bias in radiance cache based inverse rendering. In ECCV, 2025. 2 +[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In ICCV, 2021. 2 +[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In CVPR, 2022. +[4] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In ICCV, 2023. 2 +[5] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T. Barron, Ce Liu, and Hendrik P.A. Lensch. Nerd: Neural reflectance decomposition from image collections. In ICCV, 2021. 1, 2 +[6] Mark Boss, Andreas Engelhardt, Abhishek Kar, Yuzhen Li, Deqing Sun, Jonathan Barron, Hendrik Lensch, and Varun Jampani. *Samurai: Shape and material from unconstrained real-world arbitrary image collections*. *NeurIPS*, 2022. 2 +[7] Brent Burley and Walt Disney Animation Studios. Physically-based shading at disney. In Acm Siggraph, 2012. 3 +[8] Yurui Chen, Chun Gu, Junzhe Jiang, Xiatian Zhu, and Li Zhang. Periodic vibration gaussian: Dynamic urban scene reconstruction and real-time rendering. arXiv preprint, 2023. 2 +[9] Jan-Niklas Dihlmann, Arjun Majumdar, Andreas Engelhardt, Raphael Braun, and Hendrik Lensch. Subsurface scattering for 3d gaussian splattering. arXiv preprint, 2024. 2 +[10] Jian Gao, Chun Gu, Youtian Lin, Hao Zhu, Xun Cao, Li Zhang, and Yao Yao. Relightable 3d gaussian: Real-time point cloud relighting with brdf decomposition and ray tracing. arXiv preprint, 2023. 1, 2, 3, 5, 6, 7, 8 +[11] Yijia Guo, Yuanxi Bai, Liwen Hu, Ziyi Guo, Mianzhi Liu, Yu Cai, Tiejun Huang, and Lei Ma. Prtgs: Precomputed radiance transfer of gaussian splats for real-time high-quality relighting. In ACMMM, 2024. 2 +[12] Yuan-Chen Guo, Di Kang, Linchao Bao, Yu He, and Song-Hai Zhang. Nerfren: Neural radiance fields with reflections. In CVPR, 2022. 5 +[13] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In SIGGRAPH, 2024. 2, 4, 5, 7 +[14] Yingwenqi Jiang, Jiadong Tu, Yuan Liu, Xifeng Gao, Xiaoxiao Long, Wenping Wang, and Yuexin Ma. Gaussianshader: 3d gaussian splatting with shading functions for reflective surfaces. arXiv preprint, 2023. 1, 2 + +[15] Haian Jin, Isabella Liu, Peijia Xu, Xiaoshuai Zhang, Songfang Han, Sai Bi, Xiaowei Zhou, Zexiang Xu, and Hao Su. Tensoir: Tensorial inverse rendering. In CVPR, 2023. 1, 2, 6, 7, 8 +[16] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. TOG, 2023. 1, 2, 3, 5 +[17] Zhengfei Kuang, Yunzhi Zhang, Hong-Xing Yu, Samir Agarwala, Elliott Wu, Jiajun Wu, et al. Stanford-orb: a real-world 3d object inverse rendering benchmark. NeurIPS, 2024. 6, 8 +[18] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In CVPR, 2023. 2 +[19] Zhihao Liang, Qi Zhang, Ying Feng, Ying Shan, and Kui Jia. Gs-ir: 3d gaussian splatting for inverse rendering. In CVPR, 2024. 1, 2, 5, 6, 7, 8 +[20] Yuan Liu, Peng Wang, Cheng Lin, Xiaoxiao Long, Jiepeng Wang, Lingjie Liu, Taku Komura, and Wenping Wang. Nano: Neural geometry and brdf reconstruction of reflective objects from multiview images. In SIGGRAPH, 2023. 1, 2, 5 +[21] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2 +[22] Nicolas Moënne-Loccoz, Ashkan Mirzaei, Or Perel, Ricardo de Lutio, Janick Martinez Esturo, Gavriel State, Sanja Fidler, Nicholas Sharp, and Zan Gojcic. 3d gaussian ray tracing: Fast tracing of particle scenes. arXiv preprint, 2024. 2, 4, 7 +[23] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 2022. 2 +[24] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting triangular 3d models, materials, and lighting from images. In CVPR, 2022. 1, 5, 6 +[25] Steven G Parker, James Bigler, Andreas Dietrich, Heiko Friedrich, Jared Hoberock, David Luebke, David McAllister, Morgan McGuire, Keith Morley, Austin Robison, et al. Optix: a general purpose ray tracing engine. TOG, 2010. 2, 4, 7 +[26] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mllps. In ICCV, 2021. 2 +[27] Yahao Shi, Yanmin Wu, Chenming Wu, Xing Liu, Chen Zhao, Haocheng Feng, Jingtuo Liu, Liangjun Zhang, Jian Zhang, Bin Zhou, et al. Gir: 3d gaussian inverse rendering for relightable scene factorization. arXiv preprint, 2023. 1, 2 +[28] Pratul P Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T Barron. Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In CVPR, 2021. 1, 2 +[29] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint, 2021. 2, 5 + +[30] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Trans. Image Process., 2004. 6 +[31] Tong Wu, Jia-Mu Sun, Yu-Kun Lai, Yuewen Ma, Leif Kobbelt, and Lin Gao. Deferreds: Decoupled and editable gaussian splatting with deferred shading. arXiv preprint, 2024. 2 +[32] Yunzhi Yan, Haotong Lin, Chenxu Zhou, Weijie Wang, Haiyang Sun, Kun Zhan, Xianpeng Lang, Xiaowei Zhou, and Sida Peng. Street gaussians for modeling dynamic urban scenes. arXiv preprint, 2024. 2 +[33] Ziyi Yang, Yanzhen Chen, Xinyu Gao, Yazhen Yuan, Yu Wu, Xiaowei Zhou, and Xiaogang Jin. Sire-ir: Inverse rendering for brdf reconstruction with shadow and illumination removal in high-illuminance scenes. arXiv preprint, 2023. 2 +[34] Zeyu Yang, Hongye Yang, Zijie Pan, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. arXiv preprint, 2023. 2 +[35] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. In CVPR, 2024. 2 +[36] Yao Yao, Jingyang Zhang, Jingbo Liu, Yihang Qu, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Neilf: Neural incident light field for physically-based material estimation. In ECCV, 2022. 2 +[37] Yuxuan Yao, Zixuan Zeng, Chun Gu, Xiatian Zhu, and Li Zhang. Reflective gaussian splatting. In ICLR, 2025. 2 +[38] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. NeurIPS, 2021. 2 +[39] Zehao Yu, Torsten Sattler, and Andreas Geiger. Gaussian opacity fields: Efficient adaptive surface reconstruction in unbounded scenes. TOG, 2024. 2 +[40] Jingyang Zhang, Yao Yao, Shiwei Li, Jingbo Liu, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Neilf++: Inter-reflectable light fields for geometry and material estimation. In ICCV, 2023. 1, 2 +[41] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6 +[42] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM Trans. Graph., 2021. 6, 7, 8 +[43] Yuanqing Zhang, Jiaming Sun, Xingyi He, Huan Fu, Rongfei Jia, and Xiaowei Zhou. Modeling indirect illumination for inverse rendering. In CVPR, 2022. 2, 6, 7, 8 \ No newline at end of file diff --git a/2025/IRGS_ Inter-Reflective Gaussian Splatting with 2D Gaussian Ray Tracing/images.zip b/2025/IRGS_ Inter-Reflective Gaussian Splatting with 2D Gaussian Ray Tracing/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d151edb055603156409cc229f2b62f576ff47c75 --- /dev/null +++ b/2025/IRGS_ Inter-Reflective Gaussian Splatting with 2D Gaussian Ray Tracing/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3db9c34bf1331f30dcf77ebb218aee1791df490024843b074a06c174b9159286 +size 676689 diff --git a/2025/IRGS_ Inter-Reflective Gaussian Splatting with 2D Gaussian Ray Tracing/layout.json b/2025/IRGS_ Inter-Reflective Gaussian Splatting with 2D Gaussian Ray Tracing/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f2aecd99f144333924ef7c319d4b710adb248f32 --- /dev/null +++ b/2025/IRGS_ Inter-Reflective Gaussian Splatting with 2D Gaussian Ray Tracing/layout.json @@ -0,0 +1,8322 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 78, + 103, + 532, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 103, + 532, + 121 + ], + "spans": [ + { + "bbox": [ + 78, + 103, + 532, + 121 + ], + "type": "text", + "content": "IRGS: Inter-Reflective Gaussian Splating with 2D Gaussian Ray Tracing" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 118, + 142, + 494, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 142, + 494, + 158 + ], + "spans": [ + { + "bbox": [ + 118, + 142, + 494, + 158 + ], + "type": "text", + "content": "Chun Gu" + }, + { + "bbox": [ + 118, + 142, + 494, + 158 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 118, + 142, + 494, + 158 + ], + "type": "text", + "content": " Xiaofei Wei" + }, + { + "bbox": [ + 118, + 142, + 494, + 158 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 118, + 142, + 494, + 158 + ], + "type": "text", + "content": " Zixuan Zeng" + }, + { + "bbox": [ + 118, + 142, + 494, + 158 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 118, + 142, + 494, + 158 + ], + "type": "text", + "content": " Yuxuan Yao" + }, + { + "bbox": [ + 118, + 142, + 494, + 158 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 118, + 142, + 494, + 158 + ], + "type": "text", + "content": " Li Zhang" + }, + { + "bbox": [ + 118, + 142, + 494, + 158 + ], + "type": "inline_equation", + "content": "^{1,2*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 118, + 157, + 488, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 157, + 488, + 172 + ], + "spans": [ + { + "bbox": [ + 118, + 157, + 488, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 118, + 157, + 488, + 172 + ], + "type": "text", + "content": "School of Data Science, Fudan University " + }, + { + "bbox": [ + 118, + 157, + 488, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 118, + 157, + 488, + 172 + ], + "type": "text", + "content": "Shanghai Innovation Institute" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 186, + 178, + 419, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 178, + 419, + 191 + ], + "spans": [ + { + "bbox": [ + 186, + 178, + 419, + 191 + ], + "type": "text", + "content": "https://fudan-zvq.github.io/IRGS" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 219, + 200, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 219, + 200, + 232 + ], + "spans": [ + { + "bbox": [ + 151, + 219, + 200, + 232 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 244, + 296, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 244, + 296, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 244, + 296, + 483 + ], + "type": "text", + "content": "In inverse rendering, accurately modeling visibility and indirect radiance for incident light is essential for capturing secondary effects. Due to the absence of a powerful Gaussian ray tracer, previous 3DGS-based methods have either adopted a simplified rendering equation or used learnable parameters to approximate incident light, resulting in inaccurate material and lighting estimations. To this end, we introduce inter-reflective Gaussian splatting (IRGS) for inverse rendering. To capture inter-reflection, we apply the full rendering equation without simplification and compute incident radiance on the fly using the proposed differentiable 2D Gaussian ray tracing. Additionally, we present an efficient optimization scheme to handle the computational demands of Monte Carlo sampling for rendering equation evaluation. Furthermore, we introduce a novel strategy for querying the indirect radiance of incident light when relighting the optimized scenes. Extensive experiments on multiple standard benchmarks validate the effectiveness of IRGS, demonstrating its capability to accurately model complex inter-reflection effects." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 506, + 135, + 519 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 506, + 135, + 519 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 135, + 519 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 526, + 295, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 526, + 295, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 526, + 295, + 694 + ], + "type": "text", + "content": "Inverse rendering is a fundamental problem in computer vision and graphics, aiming to reconstruct geometry and estimate material and lighting from a set of posed images. The introduction of neural radiance field (NeRF) [21], which uses a neural implicit field modeled by a multi-layer perceptron (MLP) to represent 3D scenes, has inspired many NeRF-based methods [5, 15, 20, 28, 40] to address inverse rendering tasks. Leveraging NeRF's ray-based rendering technique, these methods can accurately compute the visibility and indirect radiance for incident light as required by the rendering equation. However, the exhaustive querying of a neural field during ray marching and its limited representational capacity constrain the efficiency and performance of neural implicit fields in inverse rendering." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 325, + 218, + 545, + 357 + ], + "blocks": [ + { + "bbox": [ + 325, + 218, + 545, + 357 + ], + "lines": [ + { + "bbox": [ + 325, + 218, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 325, + 218, + 545, + 357 + ], + "type": "image", + "image_path": "543abf89ba94525d5f3e15010162a76c61de896c8ac2469bb64c796c6d54566d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 325, + 357, + 545, + 498 + ], + "blocks": [ + { + "bbox": [ + 325, + 357, + 545, + 498 + ], + "lines": [ + { + "bbox": [ + 325, + 357, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 325, + 357, + 545, + 498 + ], + "type": "image", + "image_path": "49ec55b658c17a7f302914ba24b45925ca2558bdc2ce662d38fef04af8097f57.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 506, + 555, + 529 + ], + "lines": [ + { + "bbox": [ + 313, + 506, + 555, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 555, + 529 + ], + "type": "text", + "content": "Figure 1. Global and indirect illumination in a Gaussian-based scene using our IRGS, highlighting its inter-reflective capabilities." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "content": "Recently, 3D Gaussian splatting (3DGS) [16] has emerged as a promising technique for 3D scene representation, offering both remarkable rendering quality and efficiency. By modeling a static 3D scene as a collection of 3D Gaussians, 3DGS achieves real-time, high-quality rendering with a tile-based rasterizer. However, despite its strengths, 3DGS has notable limitations due to its reliance on rasterization, which limits its ability to accurately simulate ray-based effects that are essential for inverse rendering. As a result, some 3DGS-based inverse rendering methods [14, 19, 27] have adopted simplified versions of the rendering equation, such as the split-sum approximation [24]. Other approaches, such as R3DG [10], use learnable parameters to approximate indirect radiance, which fails to" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 703, + 257, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 703, + 257, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 703, + 257, + 713 + ], + "type": "text", + "content": "*Corresponding author (lizhangfd@fudan.edu.cn)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10943" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": "capture actual inter-reflection. These methods struggle to accurately model the rendering equation, limiting their effectiveness in realistic material and lighting estimation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 110, + 294, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 110, + 294, + 230 + ], + "spans": [ + { + "bbox": [ + 54, + 110, + 294, + 230 + ], + "type": "text", + "content": "Our goal is to integrate the full rendering equation into the efficient Gaussian splatting for accurate inverse rendering, enabling the capture of complex interreflection effects. This requires efficient and precise computation of visibility and indirect radiance for incident light. Drawing inspiration from 3DGRT [22], which achieves efficient ray tracing on 3D Gaussian primitives via a " + }, + { + "bbox": [ + 54, + 110, + 294, + 230 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 54, + 110, + 294, + 230 + ], + "type": "text", + "content": "-buffer hits-based marching technique and hardware-accelerated NVIDIA OptiX interface [25], we explore adapting this technique to model inter-reflection within the Gaussian splatting framework." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 231, + 296, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 296, + 651 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 296, + 651 + ], + "type": "text", + "content": "In this paper, we introduce inter-reflective Gaussian splatting (IRGS), a novel inverse rendering framework that leverages the proposed 2D Gaussian ray tracing technique to model inter-reflections within Gaussian splatting. To the best of our knowledge, this is the first method to integrate the full rendering equation without simplification into Gaussian splatting, accurately capturing inter-reflection effects by querying incident lighting through ray tracing. First, we propose 2D Gaussian ray tracing to efficiently and accurately trace across 2D Gaussian primitives, computing the accumulated opacity and color of arbitrary rays. Unlike 3DGRT [22], our approach features well-defined ray-splat intersections, enabling direct ray tracing on pretrained 2D Gaussian splatting [13] with minimal quality degradation. Next, we introduce the inter-reflective Gaussian splatting framework: starting from a pretrained 2D Gaussian splatting model, we apply the full rendering equation without simplification at intersection points determined by the depth map, while visibility and indirect radiance are computed on the fly using our 2D Gaussian ray tracing technique. Notably, our 2D Gaussian ray tracing is fully differentiable, enabling the optimization of indirect light through backpropagation. We further propose an efficient optimization scheme to manage the computational demands of the exhaustive Monte Carlo sampling for the rendering equation. Additionally, we introduce a novel strategy to query indirect radiance during relighting, overcoming limitations in previous methods that often omit this component. Extensive experiments on multiple standard benchmarks confirm the effectiveness of IRGS. In Fig. 1, we visualize the global illumination (considering both direct and indirect radiance) and the indirect illumination (considering only indirect radiance) of a Gaussian-based scene rendered using IRGS, demonstrating that IRGS is capable of modeling remarkable inter-reflection effects." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "The contributions of this paper are as follows: (i) 2D Gaussian ray tracing, a technique that enables direct ray tracing on pretrained 2D Gaussian splatting with minimal quality degradation; (ii) IRGS, an inverse rendering framework that incorporates the full rendering equation without" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 555, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 133 + ], + "type": "text", + "content": "simplification, using 2D Gaussian ray tracing to compute incident radiance on the fly; (iii) An optimization scheme that manages computational demands for rendering equation evaluation; (iv) A novel strategy that enables querying of indirect radiance during relighting." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 141, + 399, + 154 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 141, + 399, + 154 + ], + "spans": [ + { + "bbox": [ + 314, + 141, + 399, + 154 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 161, + 555, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 161, + 555, + 484 + ], + "spans": [ + { + "bbox": [ + 313, + 161, + 555, + 484 + ], + "type": "text", + "content": "Novel view synthesis aims to leverage a limited collection of observed images of a scene to generate new images from an unseen viewpoint. Neural radiance field (NeRF) [21] produces photo-realistic quality images by representing the scenes in terms of a volumetric radiance field encoded in a coordinate-based neural network. Subsequent advancements have focused on accelerating rendering[23, 26], enhancing image quality [2-4], and improving geometry reconstruction [18, 29, 38]. More recently, 3D Gaussian splatting (3DGS) [16] introduced the use of fuzzy, anisotropic 3D Gaussian point cloud for scene representation along with an efficient tile-based rasterizer, achieving state-of-the-art results in both rendering quality and efficiency. 3DGS has inspired numerous approaches targeting various goals, including geometry reconstruction [13, 39], dynamic scene reconstruction [34, 35], inverse rendering [10, 19], and street scene applications [8, 32]. However, the dependence on rasterization limits Gaussian splatting's ability to simulate ray-based effects accurately. To address this, 3DGRT [22] proposes a differentiable Gaussian ray tracer that efficiently performs ray tracing across 3D Gaussian primitives, allowing radiance computation along arbitrary rays. In this paper, we propose ray tracing on 2D Gaussian primitives. Unlike 3DGRT, 2D Gaussian primitives [13] have a well-defined ray-splat intersection, enabling direct ray tracing on pre-trained 2D Gaussian splatting with minimal quality degradation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 498, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 713 + ], + "type": "text", + "content": "Inverse rendering focuses on recovering geometry, material, and lighting properties from captured images, presenting a complex challenge due to high-dimensional light-surface interactions. Inspired by the success of neural radiance fields, many methods [1, 5, 6, 15, 20, 28, 33, 36, 40, 43] leverage flexible ray marching techniques and neural implicit fields for inverse rendering. Recent approaches [9-11, 14, 19, 27, 31, 37] have adapted 3D Gaussian splatting (3DGS) to inverse rendering by assigning material-related properties to each Gaussian primitive. Leveraging the rapid rendering capabilities of 3DGS, these methods achieve much higher efficiency compared to NeRF-based techniques. However, due to the lack of an efficient Gaussian ray tracer, existing 3DGS-based methods either apply a simplified version of the rendering equation [19] or use learnable parameters to model incident light [10], which limits the accuracy of material and lighting estimation. For example, GS-IR [19] employs split-sum approximation to" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10944" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 70, + 553, + 190 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 553, + 190 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 553, + 190 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 553, + 190 + ], + "type": "image", + "image_path": "4d68f25c23fae156617de34e6486be3db749414b84b377e9696048edad14b68a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 197, + 555, + 243 + ], + "lines": [ + { + "bbox": [ + 54, + 197, + 555, + 243 + ], + "spans": [ + { + "bbox": [ + 54, + 197, + 555, + 243 + ], + "type": "text", + "content": "Figure 2. Schematic illustration of the proposed IRGS. Starting from a set of 2D Gaussians equipped with material properties, we apply rasterization to generate albedo, roughness, position, and normal maps. We then evaluate the rendering equation using stratified sampling at the corresponding position, drawing geometry and material values from these feature maps. The radiance of incident light is decomposed into direct radiance " + }, + { + "bbox": [ + 54, + 197, + 555, + 243 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{dir}}" + }, + { + "bbox": [ + 54, + 197, + 555, + 243 + ], + "type": "text", + "content": " from the environment map, and indirect radiance " + }, + { + "bbox": [ + 54, + 197, + 555, + 243 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{ind}}" + }, + { + "bbox": [ + 54, + 197, + 555, + 243 + ], + "type": "text", + "content": " with visibility " + }, + { + "bbox": [ + 54, + 197, + 555, + 243 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 54, + 197, + 555, + 243 + ], + "type": "text", + "content": ", obtained via 2D Gaussian ray tracing." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 254, + 296, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 254, + 296, + 339 + ], + "spans": [ + { + "bbox": [ + 54, + 254, + 296, + 339 + ], + "type": "text", + "content": "simplify the rendering equation and uses baked volumes to store occlusion and indirect illumination. R3DG [10] shades each Gaussian individually, using point-based ray tracing to bake occlusion and spherical harmonics to parameterize indirect radiance. To overcome these limitations, we propose the IRGS framework that utilizes 2D Gaussian ray tracing to accurately model inter-reflections." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 349, + 111, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 349, + 111, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 349, + 111, + 361 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 369, + 296, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 369, + 296, + 501 + ], + "spans": [ + { + "bbox": [ + 54, + 369, + 296, + 501 + ], + "type": "text", + "content": "In this section, we introduce our inter-reflective Gaussian splatting with 2D Gaussian ray tracing, designed to reconstruct geometry and estimate material and lighting from a set of posed RGB images. We begin by providing the necessary background on 3D Gaussian splatting and the rendering equation (Section 3.1). We then present the 2D Gaussian ray tracing technique, which performs ray tracing on 2D Gaussian primitives (Section 3.2). Finally, we outline our IRGS framework, which leverages the proposed 2D Gaussian ray tracing technique to query visibility and indirect radiance for incident light (Section 3.3)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 508, + 135, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 508, + 135, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 508, + 135, + 521 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 526, + 296, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 526, + 296, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 526, + 296, + 588 + ], + "type": "text", + "content": "3D Gaussian splatting 3D Gaussian splatting (3DGS) represents a 3D scene using 3D Gaussian primitives, each defined as an ellipsoid characterized by a 3D mean " + }, + { + "bbox": [ + 55, + 526, + 296, + 588 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 55, + 526, + 296, + 588 + ], + "type": "text", + "content": " and a covariance matrix " + }, + { + "bbox": [ + 55, + 526, + 296, + 588 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 55, + 526, + 296, + 588 + ], + "type": "text", + "content": ". The influence of a Gaussian at a spatial position " + }, + { + "bbox": [ + 55, + 526, + 296, + 588 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 526, + 296, + 588 + ], + "type": "text", + "content": " is expressed as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 84, + 595, + 295, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 595, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 84, + 595, + 295, + 622 + ], + "type": "interline_equation", + "content": "G (\\boldsymbol {x}) = \\exp \\left(- \\frac {1}{2} (\\boldsymbol {x} - \\boldsymbol {\\mu}) ^ {\\top} \\Sigma^ {- 1} (\\boldsymbol {x} - \\boldsymbol {\\mu})\\right). \\tag {1}", + "image_path": "4804f7085ac958ed02a4df61fce0603d62d34040afd72fe3183feea1fb360f7a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "text", + "content": "Each Gaussian is also assigned an opacity " + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "text", + "content": " and a view-dependent color " + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "text", + "content": " modeled by spherical harmonics (SH). The rendering process of 3DGS starts by projecting 3D Gaussian primitives onto the 2D image plane. This is achieved by first applying a viewing transformation " + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "text", + "content": " and then performing a perspective projection. The covariance matrix of the resulting 2D Gaussians can be approximated" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 253, + 555, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 253, + 555, + 301 + ], + "spans": [ + { + "bbox": [ + 313, + 253, + 555, + 301 + ], + "type": "text", + "content": "as " + }, + { + "bbox": [ + 313, + 253, + 555, + 301 + ], + "type": "inline_equation", + "content": "\\Sigma' = JW\\Sigma W^\\top J^\\top" + }, + { + "bbox": [ + 313, + 253, + 555, + 301 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 253, + 555, + 301 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 313, + 253, + 555, + 301 + ], + "type": "text", + "content": " represents the Jacobian matrix of the perspective projection. The final image is rendered by alpha-blending these projected 2D Gaussians from front to back:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 356, + 308, + 555, + 343 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 308, + 555, + 343 + ], + "spans": [ + { + "bbox": [ + 356, + 308, + 555, + 343 + ], + "type": "interline_equation", + "content": "\\mathcal {C} = \\sum_ {i = 1} ^ {N} T _ {i} \\alpha_ {i} \\boldsymbol {c} _ {i}, \\quad T _ {i} = \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), \\tag {2}", + "image_path": "b16aae5d863be9a6171ef5c976c9ebbf9b67da1149494ef923fbf9bb11bd8d03.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 350, + 555, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 555, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 555, + 399 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 350, + 555, + 399 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 313, + 350, + 555, + 399 + ], + "type": "text", + "content": " is calculated by multiplying opacity " + }, + { + "bbox": [ + 313, + 350, + 555, + 399 + ], + "type": "inline_equation", + "content": "o_i" + }, + { + "bbox": [ + 313, + 350, + 555, + 399 + ], + "type": "text", + "content": " with each Gaussian's influence, determined by " + }, + { + "bbox": [ + 313, + 350, + 555, + 399 + ], + "type": "inline_equation", + "content": "\\Sigma^{\\prime}" + }, + { + "bbox": [ + 313, + 350, + 555, + 399 + ], + "type": "text", + "content": " and the pixel coordinate. The covariance matrix " + }, + { + "bbox": [ + 313, + 350, + 555, + 399 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 313, + 350, + 555, + 399 + ], + "type": "text", + "content": " is parameterized by a unit quaternion and a scaling vector to facilitate optimization." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 413, + 554, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 554, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 554, + 449 + ], + "type": "text", + "content": "Rendering equation In physically-based rendering, illumination at a given surface point " + }, + { + "bbox": [ + 313, + 413, + 554, + 449 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 313, + 413, + 554, + 449 + ], + "type": "text", + "content": " is computed via the rendering equation:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 321, + 457, + 553, + 483 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 457, + 553, + 483 + ], + "spans": [ + { + "bbox": [ + 321, + 457, + 553, + 483 + ], + "type": "interline_equation", + "content": "L _ {o} \\left(\\boldsymbol {\\omega} _ {o}, \\boldsymbol {x}\\right) = \\int_ {\\Omega} f \\left(\\boldsymbol {\\omega} _ {o}, \\boldsymbol {\\omega} _ {i}, \\boldsymbol {x}\\right) L _ {1} \\left(\\boldsymbol {\\omega} _ {i}, \\boldsymbol {x}\\right) \\left(\\boldsymbol {\\omega} _ {i} \\cdot \\mathbf {n}\\right) d \\boldsymbol {\\omega} _ {i}, \\tag {3}", + "image_path": "a8613440677ef1d0966b149ce27cb662daa4aca3100ebe3d2fc444f299f36640.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 491, + 555, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 555, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 555, + 598 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 491, + 555, + 598 + ], + "type": "inline_equation", + "content": "\\omega_{o}" + }, + { + "bbox": [ + 313, + 491, + 555, + 598 + ], + "type": "text", + "content": " is the outgoing direction, " + }, + { + "bbox": [ + 313, + 491, + 555, + 598 + ], + "type": "inline_equation", + "content": "\\omega_{i}" + }, + { + "bbox": [ + 313, + 491, + 555, + 598 + ], + "type": "text", + "content": " is the incident direction over the hemisphere defined by the surface normal " + }, + { + "bbox": [ + 313, + 491, + 555, + 598 + ], + "type": "inline_equation", + "content": "\\pmb{n}" + }, + { + "bbox": [ + 313, + 491, + 555, + 598 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 491, + 555, + 598 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 491, + 555, + 598 + ], + "type": "text", + "content": " is the bidirectional reflectance distribution function (BRDF), and " + }, + { + "bbox": [ + 313, + 491, + 555, + 598 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{i}}(\\boldsymbol{\\omega}_i,\\boldsymbol{x})" + }, + { + "bbox": [ + 313, + 491, + 555, + 598 + ], + "type": "text", + "content": " is the radiance of the incident light. In this work, we use a simplified Disney BRDF model [7] with only diffuse albedo " + }, + { + "bbox": [ + 313, + 491, + 555, + 598 + ], + "type": "inline_equation", + "content": "\\pmb{a}" + }, + { + "bbox": [ + 313, + 491, + 555, + 598 + ], + "type": "text", + "content": " and roughness " + }, + { + "bbox": [ + 313, + 491, + 555, + 598 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 491, + 555, + 598 + ], + "type": "text", + "content": " as parameters, and assume the material is dielectric. The BRDF " + }, + { + "bbox": [ + 313, + 491, + 555, + 598 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 491, + 555, + 598 + ], + "type": "text", + "content": " in Eq. (3) can be separated into a diffuse term " + }, + { + "bbox": [ + 313, + 491, + 555, + 598 + ], + "type": "inline_equation", + "content": "f_{d} = \\frac{a}{\\pi}" + }, + { + "bbox": [ + 313, + 491, + 555, + 598 + ], + "type": "text", + "content": " and a specular term " + }, + { + "bbox": [ + 313, + 491, + 555, + 598 + ], + "type": "inline_equation", + "content": "f_{s}" + }, + { + "bbox": [ + 313, + 491, + 555, + 598 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 361, + 605, + 553, + 633 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 605, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 361, + 605, + 553, + 633 + ], + "type": "interline_equation", + "content": "f _ {s} \\left(\\boldsymbol {\\omega} _ {i}, \\boldsymbol {\\omega} _ {o}, \\boldsymbol {x}\\right) = \\frac {D F G}{4 \\left(\\boldsymbol {\\omega} _ {i} \\cdot \\mathbf {n}\\right) \\left(\\boldsymbol {\\omega} _ {o} \\cdot \\mathbf {n}\\right)}, \\tag {4}", + "image_path": "fe6037a02b636b107ffd8fe696c48da1bfe0c70a5a86f3584e31185910bc11fa.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 639, + 553, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 639, + 553, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 639, + 553, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 639, + 553, + 665 + ], + "type": "inline_equation", + "content": "D, F" + }, + { + "bbox": [ + 313, + 639, + 553, + 665 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 639, + 553, + 665 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 639, + 553, + 665 + ], + "type": "text", + "content": " represent the normal distribution function, the Fresnel term, and the geometry term, respectively." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 671, + 452, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 452, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 452, + 684 + ], + "type": "text", + "content": "3.2. 2D Gaussian ray tracing" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "While 3DGS [16] has been successful in applications such as 3D reconstruction, achieving superior rendering quality" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10945" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 289, + 152 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 289, + 152 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 289, + 152 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 289, + 152 + ], + "type": "image", + "image_path": "46528fd0fe2cedbaf808364828dc2395595e7e26f092daded8dc70456a9a8ee2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 161, + 295, + 216 + ], + "lines": [ + { + "bbox": [ + 55, + 161, + 295, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 161, + 295, + 216 + ], + "type": "text", + "content": "Figure 3. Performance of directly applying Gaussian ray tracing on a pretrained Gaussian splatting checkpoint in 3D Gaussian and 2D Gaussian cases, respectively. Our proposed 2DGRT achieves significantly less quality degradation than 3DGRT [22] in both quantitative metrics and visual results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 229, + 295, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 229, + 295, + 396 + ], + "spans": [ + { + "bbox": [ + 54, + 229, + 295, + 396 + ], + "type": "text", + "content": "and efficiency compared to NeRF-based methods, it is limited by the constraints of rasterization. These limitations are primarily related to ray-based effects, including rendering with highly distorted cameras and handling secondary ray effects. To address these challenges, 3DGRT [22] proposes performing ray tracing across 3D Gaussian primitives. This approach overcomes the inefficiencies of ray tracing on 3D Gaussian primitives by implementing a " + }, + { + "bbox": [ + 54, + 229, + 295, + 396 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 54, + 229, + 295, + 396 + ], + "type": "text", + "content": "-buffer hits-based marching technique with a hardware-accelerated NVIDIA OptiX interface [25]. In this paper, as our focus is on inverse rendering, performing ray tracing on Gaussian primitives is essential to accurately compute inter-reflection effects when evaluating the rendering equation, including visibility and indirect radiance of incident light." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 399, + 295, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 399, + 295, + 614 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 295, + 614 + ], + "type": "text", + "content": "3D Gaussian ray tracing [22] performs ray tracing on 3D Gaussian primitives and determines the ray-splat intersection as the point of maximum response from the corresponding 3D Gaussian along the ray. This behavior differs from that of 3D Gaussian splatting, which computes the particle response in 2D image space, leaving the position of the 3D intersection point undefined. To leverage both the efficiency of Gaussian splatting and the flexible rendering capabilities of Gaussian ray tracing, we aim to perform ray tracing directly on the Gaussian splatting model for incident radiance inference. However, as shown in Fig. 3, applying ray tracing directly on a 3DGS checkpoint results in a noticeable decline in rendering quality compared to splatting. To address this issue, we propose 2D Gaussian ray tracing (2DGRT), which performs ray tracing on 2D Gaussian primitives [13]. Since 2D Gaussian disks have a well-defined ray-splat intersection, this approach eliminates inconsistencies in intersection points present in 3D Gaussians." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": "Given a 2D Gaussian [13] defined by its center point " + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": ", an opacity parameter " + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "inline_equation", + "content": "o \\in [0,1]" + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": ", two principal tangential vectors " + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{t}_u \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{t}_v \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": ", and a scaling vector " + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{s} = (s_u, s_v) \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": ", we follow 3DGRT [22] to fit a bounding proxy of each 2D Gaussian using an adaptive icosahedron mesh, ensuring that a minimum influence of " + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\alpha_{\\mathrm{min}}" + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": " is enclosed. The vertices " + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{v}" + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": " of the bounding icosahedron mesh are obtained by transforming a regular icosahedron with a" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 73, + 388, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 73, + 388, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 73, + 388, + 84 + ], + "type": "text", + "content": "unit inner sphere:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 328, + 91, + 553, + 105 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 91, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 328, + 91, + 553, + 105 + ], + "type": "interline_equation", + "content": "\\boldsymbol {v} \\leftarrow \\sqrt {2 \\log \\left(o / \\alpha_ {\\min }\\right)} \\left(s _ {u} \\boldsymbol {t} _ {u} \\quad s _ {v} \\boldsymbol {t} _ {v} \\quad \\epsilon \\mathbf {1}\\right) \\boldsymbol {v} + \\boldsymbol {\\mu}, \\tag {5}", + "image_path": "a342476cff2a6c619db62d948bef41cf12cab0c80044f11dc081f87d6337e65c.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 111, + 555, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 111, + 555, + 232 + ], + "spans": [ + { + "bbox": [ + 313, + 111, + 555, + 232 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 111, + 555, + 232 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 111, + 555, + 232 + ], + "type": "text", + "content": " denotes a small positive number. Each icosahedron mesh consists of 20 triangular faces, resulting in a total of " + }, + { + "bbox": [ + 313, + 111, + 555, + 232 + ], + "type": "inline_equation", + "content": "20N" + }, + { + "bbox": [ + 313, + 111, + 555, + 232 + ], + "type": "text", + "content": " triangles for " + }, + { + "bbox": [ + 313, + 111, + 555, + 232 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 111, + 555, + 232 + ], + "type": "text", + "content": " 2D Gaussian primitives. To leverage the hardware-accelerated ray-triangle intersection provided by OptiX [25], we construct a Bounding Volume Hierarchy (BVH) using these triangles. By applying the " + }, + { + "bbox": [ + 313, + 111, + 555, + 232 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 111, + 555, + 232 + ], + "type": "text", + "content": "-buffer per-ray sorting algorithm [22], we obtain the exact ordering of intersected Gaussians for each ray. To ensure consistency with 2DGS [13], we analytically compute the ray-splat intersection point " + }, + { + "bbox": [ + 313, + 111, + 555, + 232 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 111, + 555, + 232 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 341, + 236, + 553, + 264 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 236, + 553, + 264 + ], + "spans": [ + { + "bbox": [ + 341, + 236, + 553, + 264 + ], + "type": "interline_equation", + "content": "\\boldsymbol {p} = \\boldsymbol {r} _ {o} + \\tau \\boldsymbol {r} _ {d}, \\quad \\text {w h e r e} \\quad \\tau = \\frac {\\boldsymbol {n} ^ {\\top} (\\boldsymbol {\\mu} - \\boldsymbol {r} _ {o})}{\\boldsymbol {n} ^ {\\top} \\boldsymbol {r} _ {d}}. \\tag {6}", + "image_path": "ec95e91fd7bcf36630eb391aa48689c9434baff96a77c31ad4a89072a4bed4fc.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 268, + 554, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 268, + 554, + 316 + ], + "spans": [ + { + "bbox": [ + 313, + 268, + 554, + 316 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 268, + 554, + 316 + ], + "type": "inline_equation", + "content": "\\pmb{n} = \\pmb{t}_u\\times \\pmb{t}_v" + }, + { + "bbox": [ + 313, + 268, + 554, + 316 + ], + "type": "text", + "content": " is the normal vector of the 2D Gaussian disk, and " + }, + { + "bbox": [ + 313, + 268, + 554, + 316 + ], + "type": "inline_equation", + "content": "\\pmb{r}_o" + }, + { + "bbox": [ + 313, + 268, + 554, + 316 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 268, + 554, + 316 + ], + "type": "inline_equation", + "content": "\\pmb{r}_d" + }, + { + "bbox": [ + 313, + 268, + 554, + 316 + ], + "type": "text", + "content": " represent the origin and direction of the given ray, respectively. Thus, the influence of the current 2D Gaussian primitive is given by:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 375, + 322, + 553, + 349 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 322, + 553, + 349 + ], + "spans": [ + { + "bbox": [ + 375, + 322, + 553, + 349 + ], + "type": "interline_equation", + "content": "G (\\boldsymbol {p}) = \\exp \\left(- \\frac {u ^ {2} + v ^ {2}}{2}\\right), \\tag {7}", + "image_path": "22a1936e1018299e70d0de1631cc90ff05a8794f54001ba4d143d0066fac1e63.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 354, + 508, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 508, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 508, + 369 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 354, + 508, + 369 + ], + "type": "inline_equation", + "content": "u = \\frac{1}{s_v} (\\pmb {p} - \\pmb {\\mu})^\\top \\pmb {t}_u,v = \\frac{1}{s_v} (\\pmb {p} - \\pmb {\\mu})^\\top \\pmb {t}_v" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 368, + 554, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 368, + 554, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 368, + 554, + 403 + ], + "type": "text", + "content": "Consequently, the RGB value and accumulated opacity value of the ray can be obtained by alpha-blending (Eq. (2)) the intersected 2D Gaussians from front to back:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 378, + 411, + 553, + 423 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 411, + 553, + 423 + ], + "spans": [ + { + "bbox": [ + 378, + 411, + 553, + 423 + ], + "type": "interline_equation", + "content": "\\left(\\boldsymbol {c} _ {\\mathrm {r t}}, o _ {\\mathrm {r t}}\\right) \\leftarrow \\operatorname {T r a c e} \\left(\\boldsymbol {r} _ {o}, \\boldsymbol {r} _ {d}\\right), \\tag {8}", + "image_path": "45f20784008fd97d29aeb4b14cedcd04896660e2776dffad8114039d6024294c.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 430, + 555, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 555, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 555, + 525 + ], + "type": "text", + "content": "where the RGB value of each Gaussian is modeled by the view-dependent color " + }, + { + "bbox": [ + 313, + 430, + 555, + 525 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 430, + 555, + 525 + ], + "type": "text", + "content": " as in vanilla 2D Gaussian splatting. It is worth noting that subtle differences remain between 2D Gaussian ray tracing and 2DGS due to sorting errors introduced by per-tile sorting in 2DGS. However, as shown in Fig. 3, the quality degradation from 2D Gaussian ray tracing is minimal and significantly less noticeable than that observed in 3D Gaussians." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 532, + 495, + 545 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 532, + 495, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 532, + 495, + 545 + ], + "type": "text", + "content": "3.3. Inter-reflective Gaussian splatting" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 551, + 488, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 551, + 488, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 551, + 488, + 563 + ], + "type": "text", + "content": "3.3.1 Stage I: 2D Gaussian pretraining" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 569, + 554, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 569, + 554, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 569, + 554, + 677 + ], + "type": "text", + "content": "Before estimating material and lighting for inverse rendering, we first pretrain a standard 2D Gaussian splatting model. This step is essential, as reliable geometry is required for inverse rendering, and evaluating the rendering equation is computationally intensive. Using a pretrained 2D Gaussian splatting model can greatly accelerate the training process. Along with the RGB rendering described in Eq. (2), we also render normalized depth and normal maps as follows:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 683, + 553, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 683, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 321, + 683, + 553, + 715 + ], + "type": "interline_equation", + "content": "\\left\\{\\mathcal {D}, \\mathcal {N} \\right\\} = \\sum_ {i = 1} ^ {N} w _ {i} \\left\\{d _ {i}, \\boldsymbol {n} _ {i} \\right\\}, \\text {w h e r e} w _ {i} = \\frac {T _ {i} \\alpha_ {i}}{\\sum_ {i = 1} ^ {N} T _ {i} \\alpha_ {i}}. \\tag {9}", + "image_path": "6d68ca3809775a8da56b57baf1b181b9a93ad0819ed15c45bd6d6abedaedf036.jpg" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10946" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": "Training loss Following 2DGS [13], we apply a normal consistency loss " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{n}}" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " to ensure alignment between the rendered normal map and the underlying geometry, as well as a depth distortion loss [13] " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{d}}" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " to encourage concentration of the 2D Gaussians. Additionally, we introduce an edge-aware smoothness loss [10, 12] " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{s,n}}" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " on the normal map to enhance geometric robustness:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 111, + 166, + 295, + 180 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 166, + 295, + 180 + ], + "spans": [ + { + "bbox": [ + 111, + 166, + 295, + 180 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {s}, \\mathrm {n}} = \\| \\nabla \\mathcal {N} \\| \\exp (- \\| \\nabla \\mathcal {C} _ {\\mathrm {g t}} \\|), \\tag {10}", + "image_path": "f7d1129e24c75dceab430588569fa11a17e864ba31ae127138d8c7ea51f22b50.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 189, + 296, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 189, + 296, + 237 + ], + "spans": [ + { + "bbox": [ + 55, + 189, + 296, + 237 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 189, + 296, + 237 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{\\mathrm{gt}}" + }, + { + "bbox": [ + 55, + 189, + 296, + 237 + ], + "type": "text", + "content": " is the ground truth training image. To address the presence of floaters during standard 2DGS training, we also incorporate a binary cross-entropy loss [29] to constrain the geometry using the provided object mask " + }, + { + "bbox": [ + 55, + 189, + 296, + 237 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 189, + 296, + 237 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 89, + 247, + 295, + 261 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 247, + 295, + 261 + ], + "spans": [ + { + "bbox": [ + 89, + 247, + 295, + 261 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {o}} = - \\mathcal {M} \\log \\mathcal {O} - (1 - \\mathcal {M}) \\log (1 - \\mathcal {O}), \\tag {11}", + "image_path": "2ca812d766df72d4aa4b643e7a2a63274154e38bdb89fc87ec2b76280d998741.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 270, + 296, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 296, + 308 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 296, + 308 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 270, + 296, + 308 + ], + "type": "inline_equation", + "content": "\\mathcal{O} = \\sum_{i=1}^{N} T_i \\alpha_i" + }, + { + "bbox": [ + 55, + 270, + 296, + 308 + ], + "type": "text", + "content": ", denotes the accumulated opacity map. Combined with the RGB reconstruction loss " + }, + { + "bbox": [ + 55, + 270, + 296, + 308 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{c}}" + }, + { + "bbox": [ + 55, + 270, + 296, + 308 + ], + "type": "text", + "content": " from 3DGS [16], the training loss for the first stage is defined as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 72, + 329, + 295, + 342 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 329, + 295, + 342 + ], + "spans": [ + { + "bbox": [ + 72, + 329, + 295, + 342 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {1} = \\mathcal {L} _ {\\mathrm {c}} + \\lambda_ {\\mathrm {n}} \\mathcal {L} _ {\\mathrm {n}} + \\lambda_ {\\mathrm {d}} \\mathcal {L} _ {\\mathrm {d}} + \\lambda_ {\\mathrm {s}, \\mathrm {n}} \\mathcal {L} _ {\\mathrm {s}, \\mathrm {n}} + \\lambda_ {\\mathrm {o}} \\mathcal {L} _ {\\mathrm {o}}. \\tag {12}", + "image_path": "35d4cad3a98bc33f9a276381e3ab1889c64ae6507c829c3afb87f529086d4203.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 358, + 203, + 370 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 358, + 203, + 370 + ], + "spans": [ + { + "bbox": [ + 55, + 358, + 203, + 370 + ], + "type": "text", + "content": "3.3.2 Stage II: inverse rendering" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 377, + 296, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 377, + 296, + 484 + ], + "spans": [ + { + "bbox": [ + 55, + 377, + 296, + 484 + ], + "type": "text", + "content": "The order of shading and rasterization is crucial in physically based rendering with Gaussian splatting. Previous works, such as R3DG [10], propose performing shading on each Gaussian, taking advantage of properties modeled on each Gaussian, such as indirect radiance. However, since the rendered normal map is supervised at the pixel level after rasterization, shading on individual Gaussians can lead to inaccurate and blurred results. Consequently, we propose applying the rendering equation after rasterization." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 486, + 296, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 296, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 296, + 558 + ], + "type": "text", + "content": "Rasterization As introduced in Section 3.1, a dielectric BRDF model is used for the rendering equation. We assign each 2D Gaussian primitive additional material properties: diffuse albedo " + }, + { + "bbox": [ + 55, + 486, + 296, + 558 + ], + "type": "inline_equation", + "content": "\\pmb{a}" + }, + { + "bbox": [ + 55, + 486, + 296, + 558 + ], + "type": "text", + "content": " and roughness " + }, + { + "bbox": [ + 55, + 486, + 296, + 558 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 486, + 296, + 558 + ], + "type": "text", + "content": ". Pixel-level albedo map " + }, + { + "bbox": [ + 55, + 486, + 296, + 558 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 55, + 486, + 296, + 558 + ], + "type": "text", + "content": " and roughness map " + }, + { + "bbox": [ + 55, + 486, + 296, + 558 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 55, + 486, + 296, + 558 + ], + "type": "text", + "content": " are obtained through alpha-blending during rasterization: " + }, + { + "bbox": [ + 55, + 486, + 296, + 558 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{A},\\mathcal{R}\\} = \\sum_{i = 1}^{N}w_{i}\\{a_{i},r_{i}\\}" + }, + { + "bbox": [ + 55, + 486, + 296, + 558 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 559, + 296, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 559, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 296, + 689 + ], + "type": "text", + "content": "Previous inverse rendering methods using Gaussian splatting have lacked an efficient way to query the visibility and radiance for arbitrary incident directions. As a result, these methods typically apply simplified versions of the rendering equation [19] (e.g., split-sum approximation [24]) or use learnable parameters to model indirect light [10]. In this paper, we leverage the proposed 2D Gaussian ray tracing technique to efficiently and accurately query inter-reflection effects on the fly, including the visibility and indirect radiance of incident light. This allows us to apply the full rendering equation without simplification." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Light parametrization We decompose the incident radiance at surface point " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " with direction " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\omega_{i}" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " in Eq. (3) into" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 518, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 518, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 518, + 84 + ], + "type": "text", + "content": "direct radiance from infinity and indirect radiance:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 331, + 90, + 553, + 103 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 90, + 553, + 103 + ], + "spans": [ + { + "bbox": [ + 331, + 90, + 553, + 103 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {i}} \\left(\\boldsymbol {\\omega} _ {i}, \\boldsymbol {x}\\right) = V \\left(\\boldsymbol {\\omega} _ {i}, \\boldsymbol {x}\\right) L _ {\\mathrm {d i r}} \\left(\\boldsymbol {\\omega} _ {i}\\right) + L _ {\\mathrm {i n d}} \\left(\\boldsymbol {\\omega} _ {i}, \\boldsymbol {x}\\right), \\tag {13}", + "image_path": "3dc89cf89354f0e918e7cb3706d6e9bf4d4ca8775b1cd4fb6d7a7fffe76e971b.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 108, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 108, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 108, + 553, + 144 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 108, + 553, + 144 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{dir}}(\\omega_i)" + }, + { + "bbox": [ + 313, + 108, + 553, + 144 + ], + "type": "text", + "content": " is modeled using an environment cubemap, and " + }, + { + "bbox": [ + 313, + 108, + 553, + 144 + ], + "type": "inline_equation", + "content": "V(\\omega_i, \\boldsymbol{x})" + }, + { + "bbox": [ + 313, + 108, + 553, + 144 + ], + "type": "text", + "content": " along with " + }, + { + "bbox": [ + 313, + 108, + 553, + 144 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{ind}}(\\omega_i, \\boldsymbol{x})" + }, + { + "bbox": [ + 313, + 108, + 553, + 144 + ], + "type": "text", + "content": " are queried using 2D Gaussian ray tracing:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 333, + 150, + 553, + 163 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 150, + 553, + 163 + ], + "spans": [ + { + "bbox": [ + 333, + 150, + 553, + 163 + ], + "type": "interline_equation", + "content": "\\left(L _ {\\text {i n d}} \\left(\\boldsymbol {\\omega} _ {i}, \\boldsymbol {x}\\right), 1 - V \\left(\\boldsymbol {\\omega} _ {i}, \\boldsymbol {x}\\right)\\right) \\leftarrow \\operatorname {T r a c e} \\left(\\boldsymbol {x}, \\boldsymbol {\\omega} _ {i}\\right) \\tag {14}", + "image_path": "d4eff808b0bb36d85b567870021c67fab4efccab9e2d197f7f77d51405241890.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 168, + 553, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 168, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 168, + 553, + 239 + ], + "type": "text", + "content": "It is important to note that the RGB values used in ray tracing correspond to the view-dependent color " + }, + { + "bbox": [ + 313, + 168, + 553, + 239 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 168, + 553, + 239 + ], + "type": "text", + "content": " from the first stage. Additionally, our 2D Gaussian ray tracing approach is fully differentiable, allowing gradients to propagate through the ray tracing process to optimize the indirect radiance of the incident ray." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 240, + 554, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 240, + 554, + 336 + ], + "spans": [ + { + "bbox": [ + 313, + 240, + 554, + 336 + ], + "type": "text", + "content": "Rendering From the rendered depth and normal maps, we can determine the surface position " + }, + { + "bbox": [ + 313, + 240, + 554, + 336 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 240, + 554, + 336 + ], + "type": "text", + "content": " and the corresponding surface normal vector " + }, + { + "bbox": [ + 313, + 240, + 554, + 336 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 313, + 240, + 554, + 336 + ], + "type": "text", + "content": " for each pixel coordinate. Monte Carlo sampling is used to evaluate the rendering equation. Since we focus on reconstructing objects with dielectric materials, we employ stratified sampling to uniformly sample " + }, + { + "bbox": [ + 313, + 240, + 554, + 336 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{r}}" + }, + { + "bbox": [ + 313, + 240, + 554, + 336 + ], + "type": "text", + "content": " incident directions across the hemisphere. The final rendering result is then computed as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 337, + 341, + 553, + 373 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 341, + 553, + 373 + ], + "spans": [ + { + "bbox": [ + 337, + 341, + 553, + 373 + ], + "type": "interline_equation", + "content": "\\mathbf {c} _ {\\mathrm {p b r}} = \\frac {2 \\pi}{N _ {\\mathrm {r}}} \\sum_ {i = 1} ^ {N _ {\\mathrm {r}}} \\left(f _ {\\mathrm {d}} + f _ {\\mathrm {s}}\\right) L _ {\\mathrm {i}} \\left(\\boldsymbol {\\omega} _ {i}, \\boldsymbol {x}\\right) \\left(\\boldsymbol {\\omega} _ {i} \\cdot \\mathbf {n}\\right), \\tag {15}", + "image_path": "cfb653e331c3d75184d71c96d72501533dd6b5a6501288862250156a9ce3f421.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 378, + 554, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 554, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 554, + 449 + ], + "type": "text", + "content": "Training loss In addition to the training losses from the first stage, we apply regularization to facilitate material and lighting estimation. Given the prior that diffuse light " + }, + { + "bbox": [ + 313, + 378, + 554, + 449 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{diffuse}} = \\frac{1}{N_{\\mathrm{r}}} \\sum_{i=1}^{N_{\\mathrm{r}}} L(\\boldsymbol{\\omega}_i, \\boldsymbol{x})" + }, + { + "bbox": [ + 313, + 378, + 554, + 449 + ], + "type": "text", + "content": " exhibits a natural white radiance, we minimize the difference between the RGB values [20]:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 454, + 553, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 454, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 321, + 454, + 553, + 491 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {l i g h t}} = \\sum_ {c} \\left(L _ {\\text {d i f f u s e}} - \\frac {1}{3} \\sum_ {c} L _ {\\text {d i f f u s e}}\\right), c \\in \\{R, G, B \\}. \\tag {16}", + "image_path": "d364e5996acaeeeb58b58eb9a2bdff65ae9b108b33bea77f5db95d7fa5e4f793.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 492, + 555, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 492, + 555, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 492, + 555, + 684 + ], + "type": "text", + "content": "Similar to Eq. (10), we also apply edge-aware smoothness regularization " + }, + { + "bbox": [ + 313, + 492, + 555, + 684 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{s,a}}" + }, + { + "bbox": [ + 313, + 492, + 555, + 684 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 492, + 555, + 684 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{s,r}}" + }, + { + "bbox": [ + 313, + 492, + 555, + 684 + ], + "type": "text", + "content": " to the rendered albedo and roughness map, respectively, to ensure smooth material estimation. Additionally, for training efficiency, we apply the rendering equation to only a subset of pixels in each iteration. We set a maximum of " + }, + { + "bbox": [ + 313, + 492, + 555, + 684 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{rays}}" + }, + { + "bbox": [ + 313, + 492, + 555, + 684 + ], + "type": "text", + "content": " rays per iteration, meaning only " + }, + { + "bbox": [ + 313, + 492, + 555, + 684 + ], + "type": "inline_equation", + "content": "\\lfloor N_{\\mathrm{rays}} / N_{\\mathrm{r}}\\rfloor" + }, + { + "bbox": [ + 313, + 492, + 555, + 684 + ], + "type": "text", + "content": " pixels are randomly selected for final rendering equation evaluation. This scheme allows us to use a large " + }, + { + "bbox": [ + 313, + 492, + 555, + 684 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{r}}" + }, + { + "bbox": [ + 313, + 492, + 555, + 684 + ], + "type": "text", + "content": " when evaluating the rendering equation, significantly enhancing estimation quality. Consequently, only a L1 loss " + }, + { + "bbox": [ + 313, + 492, + 555, + 684 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1^{\\mathrm{pbr}}" + }, + { + "bbox": [ + 313, + 492, + 555, + 684 + ], + "type": "text", + "content": " is used to supervise the final color " + }, + { + "bbox": [ + 313, + 492, + 555, + 684 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{pbr}}" + }, + { + "bbox": [ + 313, + 492, + 555, + 684 + ], + "type": "text", + "content": ". It is important to note that the rendered image " + }, + { + "bbox": [ + 313, + 492, + 555, + 684 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 313, + 492, + 555, + 684 + ], + "type": "text", + "content": " from the vanilla 2DGS is still optimized, while the view-dependent color " + }, + { + "bbox": [ + 313, + 492, + 555, + 684 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 492, + 555, + 684 + ], + "type": "text", + "content": " associated with each Gaussian is used to compute radiance in 2D Gaussian ray tracing. The total loss for the second stage is given by:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 687, + 553, + 713 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 687, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 315, + 687, + 553, + 713 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {2} = \\mathcal {L} ^ {1} + \\lambda_ {1} ^ {\\mathrm {p b r}} \\mathcal {L} _ {1} ^ {\\mathrm {p b r}} + \\lambda_ {\\text {l i g h t}} \\mathcal {L} _ {\\text {l i g h t}} + \\lambda_ {\\mathrm {s}, \\mathrm {a}} \\mathcal {L} _ {\\mathrm {s}, \\mathrm {a}} + \\lambda_ {\\mathrm {s}, \\mathrm {r}} \\mathcal {L} _ {\\mathrm {s}, \\mathrm {r}}, \\tag {17}", + "image_path": "ee5a4e151c32b4fbebb9264ad35f670add11a19d21756f61091d90cdf7789769.jpg" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10947" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 550, + 322 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 550, + 322 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 550, + 322 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 550, + 322 + ], + "type": "image", + "image_path": "a24d95e65a3050f9c226f23018eff22a373148a8f331dc6379bc25ab0adc6d14.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 329, + 550, + 342 + ], + "lines": [ + { + "bbox": [ + 59, + 329, + 550, + 342 + ], + "spans": [ + { + "bbox": [ + 59, + 329, + 550, + 342 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison of NVS, material and lighting estimation, and relighting results on the Synthetic4Relight dataset [43]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 59, + 372, + 552, + 468 + ], + "blocks": [ + { + "bbox": [ + 69, + 352, + 541, + 364 + ], + "lines": [ + { + "bbox": [ + 69, + 352, + 541, + 364 + ], + "spans": [ + { + "bbox": [ + 69, + 352, + 541, + 364 + ], + "type": "text", + "content": "Table 1. Quantitative comparison on the Synthetic4Relight dataset [43]. A higher intensity of the red color signifies a better result." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 372, + 552, + 468 + ], + "lines": [ + { + "bbox": [ + 59, + 372, + 552, + 468 + ], + "spans": [ + { + "bbox": [ + 59, + 372, + 552, + 468 + ], + "type": "table", + "html": "
Novel view synthesisRelightingAlbedoRoughness MSE↓Time (hours)
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
NeRFactor [42]22.800.9160.15021.540.8750.17119.490.8640.206N/A>48
InvRender [43]30.740.9530.08628.670.9500.09128.280.9350.0720.00814
TensorIR [15]35.800.9780.04929.690.9510.07930.580.9460.0650.0153
GS-IR [19]33.950.9650.05725.400.9240.08319.480.8960.1170.0110.4
R3DG [10]36.800.9820.02831.000.9640.05028.310.9510.0580.0130.9
IRGS (Ours)35.480.9740.04334.900.9690.04830.810.9570.0550.0080.7
", + "image_path": "21db21ba55210412e0948f1e2f233a0f12c31ea98417857c80f52f12edc93a6a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 487, + 297, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 297, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 297, + 693 + ], + "type": "text", + "content": "Relighting When performing relighting, changes in environmental lighting make the optimized radiance " + }, + { + "bbox": [ + 55, + 487, + 297, + 693 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 487, + 297, + 693 + ], + "type": "text", + "content": " of each Gaussian no longer applicable for indirect radiance in 2D Gaussian ray tracing. Unlike previous methods [10, 19] that omit indirect lighting during relighting, we propose using the split-sum approximation [24] to efficiently query indirect radiance without recursive evaluation. Specifically, we apply ray tracing on each incident ray to alpha-blend the albedo, roughness, and normal values, resulting in " + }, + { + "bbox": [ + 55, + 487, + 297, + 693 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\mathrm{r}}" + }, + { + "bbox": [ + 55, + 487, + 297, + 693 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 487, + 297, + 693 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{\\mathrm{r}}" + }, + { + "bbox": [ + 55, + 487, + 297, + 693 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 487, + 297, + 693 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_{\\mathrm{r}}" + }, + { + "bbox": [ + 55, + 487, + 297, + 693 + ], + "type": "text", + "content": ". These aggregated values are then shaded using a pre-filtered environment map, following the approach in Nvdiffrec [24]. Additionally, we apply importance sampling on the intensity of the ground truth environment map instead of stratified sampling, enabling more accurate rendering equation evaluation by concentrating samples in high-intensity directions. Please refer to the appendix for further details." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 487, + 391, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 487, + 391, + 502 + ], + "spans": [ + { + "bbox": [ + 314, + 487, + 391, + 502 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 508, + 555, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 508, + 555, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 508, + 555, + 652 + ], + "type": "text", + "content": "Datasets and metrics To validate the effectiveness of IRGS, we conduct extensive experiments on multiple datasets for the inverse rendering task, including two synthetic datasets: Synthetic4Relight dataset [43] and the TensoIR dataset [15], as well as a real-world dataset: Stanford-ORB dataset [17]. For quantitative comparisons, we follow previous works [10, 19] and employ various metrics to evaluate performance across different outputs. Specifically, we use PSNR, SSIM [30], and LPIPS [41] for assessing image quality in novel view synthesis, albedo, and relighting tasks. For normal and roughness, we use mean angular error (MAE) and mean squared error (MSE), respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "content": "Implementation details As detailed in Section 3.3, our training process consists of two stages. The first stage is trained over 40,000 iterations with the hyperparameters " + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{n}}" + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{d}}" + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{s,n}}" + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{o}}" + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "content": " set to 0.05, 1000, 0.02, and 0.01, respectively. The second stage continues for an additional 20,000" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10948" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 553, + 198 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 553, + 198 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 553, + 198 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 553, + 198 + ], + "type": "image", + "image_path": "eef7f716dce66220e89a9df2cd24762582a2adc67a534649135886c79fa6988e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 200, + 555, + 236 + ], + "lines": [ + { + "bbox": [ + 55, + 200, + 555, + 236 + ], + "spans": [ + { + "bbox": [ + 55, + 200, + 555, + 236 + ], + "type": "text", + "content": "Figure 5. Visualization of estimated components in incident light, including the averaged direct radiance " + }, + { + "bbox": [ + 55, + 200, + 555, + 236 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{dir}}" + }, + { + "bbox": [ + 55, + 200, + 555, + 236 + ], + "type": "text", + "content": ", indirect radiance " + }, + { + "bbox": [ + 55, + 200, + 555, + 236 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{ind}}" + }, + { + "bbox": [ + 55, + 200, + 555, + 236 + ], + "type": "text", + "content": ", visibility " + }, + { + "bbox": [ + 55, + 200, + 555, + 236 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 200, + 555, + 236 + ], + "type": "text", + "content": " (ambient occlusion), incident radiance " + }, + { + "bbox": [ + 55, + 200, + 555, + 236 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{i}}" + }, + { + "bbox": [ + 55, + 200, + 555, + 236 + ], + "type": "text", + "content": ". Compared to R3DG [10], IRGS achieves notably more realistic results, especially in estimated indirect light, due to its accurate modeling of inter-reflections." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 66, + 240, + 544, + 366 + ], + "blocks": [ + { + "bbox": [ + 66, + 240, + 544, + 366 + ], + "lines": [ + { + "bbox": [ + 66, + 240, + 544, + 366 + ], + "spans": [ + { + "bbox": [ + 66, + 240, + 544, + 366 + ], + "type": "image", + "image_path": "6e490e3a8e5d7b425dd3ab2b7e03202bdf97c329ee11b22934bead54b94e1fa9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 162, + 368, + 447, + 379 + ], + "lines": [ + { + "bbox": [ + 162, + 368, + 447, + 379 + ], + "spans": [ + { + "bbox": [ + 162, + 368, + 447, + 379 + ], + "type": "text", + "content": "Figure 6. Qualitative comparison of relighting results on TensoIR dataset [43]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 59, + 399, + 290, + 479 + ], + "blocks": [ + { + "bbox": [ + 69, + 387, + 281, + 398 + ], + "lines": [ + { + "bbox": [ + 69, + 387, + 281, + 398 + ], + "spans": [ + { + "bbox": [ + 69, + 387, + 281, + 398 + ], + "type": "text", + "content": "Table 2. Quantitative comparison on TensoIR dataset [15]." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 399, + 290, + 479 + ], + "lines": [ + { + "bbox": [ + 59, + 399, + 290, + 479 + ], + "spans": [ + { + "bbox": [ + 59, + 399, + 290, + 479 + ], + "type": "table", + "html": "
MethodNormal MAE↓NVS PSNR↑Albedo PSNR↑Relight PSNR↑
NeRFactor [42]6.31424.67925.12523.383
InvRender [43]5.07427.36727.34123.973
TensoIR [15]4.10035.08829.27528.580
GS-IR [19]4.94835.33330.28624.374
R3DG [10]5.92737.34326.19927.367
IRGS (Ours)4.11235.87633.79629.907
", + "image_path": "ae8e944288f59cde34fdc7b5aaa7351e371cd29473ab445c5602aa4eaf212f76.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 497, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 497, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 497, + 297, + 713 + ], + "type": "text", + "content": "iterations, with " + }, + { + "bbox": [ + 55, + 497, + 297, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{1}^{\\mathrm{pbr}}" + }, + { + "bbox": [ + 55, + 497, + 297, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 497, + 297, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{light}}" + }, + { + "bbox": [ + 55, + 497, + 297, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 497, + 297, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{s,a}}" + }, + { + "bbox": [ + 55, + 497, + 297, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 497, + 297, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{s,r}}" + }, + { + "bbox": [ + 55, + 497, + 297, + 713 + ], + "type": "text", + "content": " set to 1.0, 0.01, 2.0, and 2.0, respectively. We implement 2D Gaussian ray tracing in OptiX [25] via PyTorch CUDA extensions, updating the BVH at each training iteration for the varying geometry, with each update taking approximately 3 ms. In line with 3DGRT [22], we use a buffer of " + }, + { + "bbox": [ + 55, + 497, + 297, + 713 + ], + "type": "inline_equation", + "content": "k = 16" + }, + { + "bbox": [ + 55, + 497, + 297, + 713 + ], + "type": "text", + "content": " for per-ray sorting and terminate ray tracing when transmittance falls below 0.03. We employ " + }, + { + "bbox": [ + 55, + 497, + 297, + 713 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{r}} = 256" + }, + { + "bbox": [ + 55, + 497, + 297, + 713 + ], + "type": "text", + "content": " rays for rendering equation evaluation, and each training iteration samples up to " + }, + { + "bbox": [ + 55, + 497, + 297, + 713 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{rays}} = 2^{18}" + }, + { + "bbox": [ + 55, + 497, + 297, + 713 + ], + "type": "text", + "content": " rays. At " + }, + { + "bbox": [ + 55, + 497, + 297, + 713 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{r}} = 256" + }, + { + "bbox": [ + 55, + 497, + 297, + 713 + ], + "type": "text", + "content": ", rendering a complete image takes about 1 second. For environmental lighting, we use a " + }, + { + "bbox": [ + 55, + 497, + 297, + 713 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 55, + 497, + 297, + 713 + ], + "type": "text", + "content": " cubemap during optimization. The learning rates for albedo, roughness, and the cubemap are set to 0.005, 0.005, and 0.01, respectively, with other hyperparameters following those in 2DGS [13]. The entire training process takes around 40 minutes, with 15 minutes for the first stage and 25 minutes for the second stage. All experiments are conducted on a single NVIDIA RTX 3090 GPU." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 388, + 372, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 388, + 372, + 399 + ], + "spans": [ + { + "bbox": [ + 314, + 388, + 372, + 399 + ], + "type": "text", + "content": "4.1. Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 410, + 555, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 410, + 555, + 686 + ], + "spans": [ + { + "bbox": [ + 313, + 410, + 555, + 686 + ], + "type": "text", + "content": "Synthetic4Relight For experiments on Synthetic4Relight [43], we evaluate IRGS's performance in novel view synthesis (NVS), relighting, material estimation, and efficiency. In Tab. 1, our approach achieves higher quality than previous methods across relighting, and material estimation, demonstrating IRGS's ability to accurately capture material details. While R3DG [10] achieves higher performance in NVS, this is primarily due to its shading on Gaussians, which enhances NVS capability at the expense of relighting performance. IRGS also completes optimization in a relatively short time of 0.7 hours, emphasizing its efficiency in inverse rendering applications. In Fig. 4, we provide a qualitative comparison of the reconstructed \"chair\" scene against GS-based competitors [10, 19], visualizing NVS, albedo, roughness, estimated environment maps, and relighting images. Our results demonstrate high-fidelity material estimation and realistic relighting outcomes. Fig. 5 further illustrates a comparison of the estimated incident light components between IRGS and R3DG [10], covering direct radiance, indirect radiance, ambient occlusion, and incident radiance. IRGS achieves more reasonable indirect radiance and clearer ambient occlusion, enhancing the overall realism." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "content": "TensoIR For TensoIR dataset [15], we report metrics for NVS, relighting, and albedo estimation in Tab. 2. Com" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10949" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 74, + 285, + 251 + ], + "blocks": [ + { + "bbox": [ + 61, + 74, + 285, + 251 + ], + "lines": [ + { + "bbox": [ + 61, + 74, + 285, + 251 + ], + "spans": [ + { + "bbox": [ + 61, + 74, + 285, + 251 + ], + "type": "image", + "image_path": "656eec5ab206eec445bcdcd4a06138af8bab875376cb8a8778f0986ac1391988.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 68, + 253, + 282, + 264 + ], + "lines": [ + { + "bbox": [ + 68, + 253, + 282, + 264 + ], + "spans": [ + { + "bbox": [ + 68, + 253, + 282, + 264 + ], + "type": "text", + "content": "Figure 7. Relighting results on Stanford-ORB dataset [17]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 62, + 268, + 289, + 380 + ], + "blocks": [ + { + "bbox": [ + 62, + 268, + 289, + 380 + ], + "lines": [ + { + "bbox": [ + 62, + 268, + 289, + 380 + ], + "spans": [ + { + "bbox": [ + 62, + 268, + 289, + 380 + ], + "type": "image", + "image_path": "989f5d0a6f4e3c99ed98ade289bb9bb0a6b368da7a327fa1792f9cc6bcce82d0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 99, + 383, + 251, + 393 + ], + "lines": [ + { + "bbox": [ + 99, + 383, + 251, + 393 + ], + "spans": [ + { + "bbox": [ + 99, + 383, + 251, + 393 + ], + "type": "text", + "content": "(a) Ablation studies on indirect light modeling." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 63, + 396, + 287, + 491 + ], + "blocks": [ + { + "bbox": [ + 63, + 396, + 287, + 491 + ], + "lines": [ + { + "bbox": [ + 63, + 396, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 63, + 396, + 287, + 491 + ], + "type": "image", + "image_path": "293ec42bb39ff67cc490758100f568f4d4227e0eab89e9e74b91651455e471be.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 91, + 495, + 260, + 504 + ], + "lines": [ + { + "bbox": [ + 91, + 495, + 260, + 504 + ], + "spans": [ + { + "bbox": [ + 91, + 495, + 260, + 504 + ], + "type": "text", + "content": "(b) Ablation studies on number of sampled rays " + }, + { + "bbox": [ + 91, + 495, + 260, + 504 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{r}}" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 67, + 506, + 282, + 517 + ], + "lines": [ + { + "bbox": [ + 67, + 506, + 282, + 517 + ], + "spans": [ + { + "bbox": [ + 67, + 506, + 282, + 517 + ], + "type": "text", + "content": "Figure 8. Ablation studies on various components of IRGS." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 517, + 295, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 295, + 613 + ], + "type": "text", + "content": "pared to prior arts [10, 15, 19, 42, 43], our method consistently outperforms in relighting and albedo estimation, and achieving comparable performance in normal quality with TensoIR [15]. Similar to Synthetic4Relight, R3DG [10] achieves higher accuracy in NVs. Fig. 6 presents a qualitative comparison of relighting images generated by GS-based methods [10, 19], showing that IRGS produces more realistic relighting results, with natural shadows." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 613, + 295, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 613, + 295, + 650 + ], + "spans": [ + { + "bbox": [ + 55, + 613, + 295, + 650 + ], + "type": "text", + "content": "Stanford-ORB dataset We also conduct experiments on the real-world Stanford-ORB dataset [17]. As shown in Fig. 7, IRGS produces realistic relighting results." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 658, + 147, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 658, + 147, + 671 + ], + "spans": [ + { + "bbox": [ + 55, + 658, + 147, + 671 + ], + "type": "text", + "content": "4.2. Ablation study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "text", + "content": "In Tab. 3, we conduct ablation studies on various components of IRGS to validate their effectiveness. We report the average PSNR on NVS, albedo estimation, and relighting" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 317, + 82, + 552, + 148 + ], + "blocks": [ + { + "bbox": [ + 328, + 71, + 539, + 81 + ], + "lines": [ + { + "bbox": [ + 328, + 71, + 539, + 81 + ], + "spans": [ + { + "bbox": [ + 328, + 71, + 539, + 81 + ], + "type": "text", + "content": "Table 3. Ablation studies on various components of IRGS." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 82, + 552, + 148 + ], + "lines": [ + { + "bbox": [ + 317, + 82, + 552, + 148 + ], + "spans": [ + { + "bbox": [ + 317, + 82, + 552, + 148 + ], + "type": "table", + "html": "
MethodNVS PSNR ↑Albedo PSNR ↑Relighting PSNR ↑
Detach indirect34.2130.2934.22
w/o indirect (train)34.0930.1033.93
w/o indirect (relight)--33.84
Nr=1634.0130.2129.46
Nr=6434.9830.6333.11
Full35.4830.8134.68
", + "image_path": "b03eb00d3851e35587fc395313a5e7bfaf08d74a6af9653cf59283af680e6bb2.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 166, + 526, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 166, + 526, + 178 + ], + "spans": [ + { + "bbox": [ + 313, + 166, + 526, + 178 + ], + "type": "text", + "content": "across four scenes in Synthetic4Relight dataset [43]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 181, + 555, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 181, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 181, + 555, + 361 + ], + "type": "text", + "content": "Indirect light We perform ablation studies to evaluate the impact of indirect light modeling. \"Detach indirect\" indicates that gradients are not backpropagated through ray tracing to optimize indirect radiance. \"w/o indirect (train)\" signifies that the indirect term in Eq. (13) is omitted during training, while \"w/o indirect (relight)\" omits the proposed strategy for computing indirect radiance during relighting. As shown in Fig. 8a, detaching the indirect term results in unrealistic indirect radiance, and training without the indirect term leads to inaccurate albedo estimation. We also demonstrate the impact of indirect radiance on relighting results of a composited scene in Fig. 1 (indirect illumination). Note that the renderings in Fig. 1 are produced by assigning an additional high metallic to the Gaussian-based \"Ground\", enhancing the visual clarity of inter-reflections." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 365, + 555, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 555, + 425 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 555, + 425 + ], + "type": "text", + "content": "Number of rays We also ablate on the number of sampled rays " + }, + { + "bbox": [ + 313, + 365, + 555, + 425 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{r}}" + }, + { + "bbox": [ + 313, + 365, + 555, + 425 + ], + "type": "text", + "content": " for rendering equation. We investigated the effects of " + }, + { + "bbox": [ + 313, + 365, + 555, + 425 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{r}} = 16" + }, + { + "bbox": [ + 313, + 365, + 555, + 425 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 365, + 555, + 425 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{r}} = 64" + }, + { + "bbox": [ + 313, + 365, + 555, + 425 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 365, + 555, + 425 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{r}} = 256" + }, + { + "bbox": [ + 313, + 365, + 555, + 425 + ], + "type": "text", + "content": ". As shown in Fig. 8b, more sampled rays lead to more accurate and smooth renderings and ambient occlusion in novel view synthesis." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 446, + 388, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 446, + 388, + 459 + ], + "spans": [ + { + "bbox": [ + 314, + 446, + 388, + 459 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 471, + 555, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 471, + 555, + 650 + ], + "spans": [ + { + "bbox": [ + 313, + 471, + 555, + 650 + ], + "type": "text", + "content": "In this paper, we introduce inter-reflective Gaussian splating (IRGS), a novel framework that overcomes the limitations of previous Gaussian-based approaches in accurately modeling visibility and indirect radiance for incident light in inverse rendering. IRGS incorporates the full rendering equation without simplifications and utilizes a differentiable 2D Gaussian ray tracing technique for precise computation of incident radiance at intersection points. We further propose an optimization scheme to address the computational demands of Monte Carlo sampling in rendering equation evaluation, and a novel strategy for querying indirect radiance during relighting. Extensive experiments on multiple benchmarks validate IRGS's ability to accurately model complex inter-reflection, establishing it as an effective solution for inverse rendering." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 654, + 554, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 712 + ], + "type": "text", + "content": "Limitation Due to the significant computational demands of rendering equation evaluation, we are unable to render at a high frame rate. However, this could be addressed by baking the incident radiance or pre-computing radiance transfer, which we leave as future work." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10950" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 116 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 116 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 116 + ], + "type": "text", + "content": "This work was supported in part by National Natural Science Foundation of China (Grant No. 62376060)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 125, + 115, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 125, + 115, + 138 + ], + "spans": [ + { + "bbox": [ + 56, + 125, + 115, + 138 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 145, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 61, + 145, + 296, + 189 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 145, + 296, + 189 + ], + "spans": [ + { + "bbox": [ + 61, + 145, + 296, + 189 + ], + "type": "text", + "content": "[1] Benjamin Attal, Dor Verbin, Ben Mildenhall, Peter Hedman, Jonathan T Barron, Matthew O'Toole, and Pratul P Srinivasan. Flash cache: Reducing bias in radiance cache based inverse rendering. In ECCV, 2025. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 190, + 296, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 190, + 296, + 233 + ], + "spans": [ + { + "bbox": [ + 61, + 190, + 296, + 233 + ], + "type": "text", + "content": "[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In ICCV, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 235, + 295, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 235, + 295, + 266 + ], + "spans": [ + { + "bbox": [ + 62, + 235, + 295, + 266 + ], + "type": "text", + "content": "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In CVPR, 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 268, + 295, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 268, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 62, + 268, + 295, + 300 + ], + "type": "text", + "content": "[4] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In ICCV, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 302, + 295, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 302, + 295, + 344 + ], + "spans": [ + { + "bbox": [ + 62, + 302, + 295, + 344 + ], + "type": "text", + "content": "[5] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T. Barron, Ce Liu, and Hendrik P.A. Lensch. Nerd: Neural reflectance decomposition from image collections. In ICCV, 2021. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 346, + 295, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 346, + 295, + 399 + ], + "spans": [ + { + "bbox": [ + 62, + 346, + 295, + 399 + ], + "type": "text", + "content": "[6] Mark Boss, Andreas Engelhardt, Abhishek Kar, Yuzhen Li, Deqing Sun, Jonathan Barron, Hendrik Lensch, and Varun Jampani. *Samurai: Shape and material from unconstrained real-world arbitrary image collections*. *NeurIPS*, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 402, + 294, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 402, + 294, + 434 + ], + "spans": [ + { + "bbox": [ + 62, + 402, + 294, + 434 + ], + "type": "text", + "content": "[7] Brent Burley and Walt Disney Animation Studios. Physically-based shading at disney. In Acm Siggraph, 2012. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 436, + 295, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 436, + 295, + 477 + ], + "spans": [ + { + "bbox": [ + 62, + 436, + 295, + 477 + ], + "type": "text", + "content": "[8] Yurui Chen, Chun Gu, Junzhe Jiang, Xiatian Zhu, and Li Zhang. Periodic vibration gaussian: Dynamic urban scene reconstruction and real-time rendering. arXiv preprint, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 480, + 295, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 480, + 295, + 512 + ], + "spans": [ + { + "bbox": [ + 62, + 480, + 295, + 512 + ], + "type": "text", + "content": "[9] Jan-Niklas Dihlmann, Arjun Majumdar, Andreas Engelhardt, Raphael Braun, and Hendrik Lensch. Subsurface scattering for 3d gaussian splattering. arXiv preprint, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 514, + 295, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 514, + 295, + 557 + ], + "spans": [ + { + "bbox": [ + 57, + 514, + 295, + 557 + ], + "type": "text", + "content": "[10] Jian Gao, Chun Gu, Youtian Lin, Hao Zhu, Xun Cao, Li Zhang, and Yao Yao. Relightable 3d gaussian: Real-time point cloud relighting with brdf decomposition and ray tracing. arXiv preprint, 2023. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 558, + 295, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 558, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 57, + 558, + 295, + 601 + ], + "type": "text", + "content": "[11] Yijia Guo, Yuanxi Bai, Liwen Hu, Ziyi Guo, Mianzhi Liu, Yu Cai, Tiejun Huang, and Lei Ma. Prtgs: Precomputed radiance transfer of gaussian splats for real-time high-quality relighting. In ACMMM, 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 602, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 602, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 602, + 295, + 635 + ], + "type": "text", + "content": "[12] Yuan-Chen Guo, Di Kang, Linchao Bao, Yu He, and Song-Hai Zhang. Nerfren: Neural radiance fields with reflections. In CVPR, 2022. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 636, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 295, + 668 + ], + "type": "text", + "content": "[13] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In SIGGRAPH, 2024. 2, 4, 5, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "text", + "content": "[14] Yingwenqi Jiang, Jiadong Tu, Yuan Liu, Xifeng Gao, Xiaoxiao Long, Wenping Wang, and Yuexin Ma. Gaussianshader: 3d gaussian splatting with shading functions for reflective surfaces. arXiv preprint, 2023. 1, 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "text", + "content": "[15] Haian Jin, Isabella Liu, Peijia Xu, Xiaoshuai Zhang, Songfang Han, Sai Bi, Xiaowei Zhou, Zexiang Xu, and Hao Su. Tensoir: Tensorial inverse rendering. In CVPR, 2023. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 118, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 150 + ], + "type": "text", + "content": "[16] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. TOG, 2023. 1, 2, 3, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 152, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 152, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 317, + 152, + 553, + 183 + ], + "type": "text", + "content": "[17] Zhengfei Kuang, Yunzhi Zhang, Hong-Xing Yu, Samir Agarwala, Elliott Wu, Jiajun Wu, et al. Stanford-orb: a real-world 3d object inverse rendering benchmark. NeurIPS, 2024. 6, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 185, + 553, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 185, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 317, + 185, + 553, + 227 + ], + "type": "text", + "content": "[18] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In CVPR, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 228, + 553, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 228, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 317, + 228, + 553, + 259 + ], + "type": "text", + "content": "[19] Zhihao Liang, Qi Zhang, Ying Feng, Ying Shan, and Kui Jia. Gs-ir: 3d gaussian splatting for inverse rendering. In CVPR, 2024. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 262, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 262, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 262, + 553, + 304 + ], + "type": "text", + "content": "[20] Yuan Liu, Peng Wang, Cheng Lin, Xiaoxiao Long, Jiepeng Wang, Lingjie Liu, Taku Komura, and Wenping Wang. Nano: Neural geometry and brdf reconstruction of reflective objects from multiview images. In SIGGRAPH, 2023. 1, 2, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 305, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 305, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 305, + 553, + 348 + ], + "type": "text", + "content": "[21] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 350, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 403 + ], + "type": "text", + "content": "[22] Nicolas Moënne-Loccoz, Ashkan Mirzaei, Or Perel, Ricardo de Lutio, Janick Martinez Esturo, Gavriel State, Sanja Fidler, Nicholas Sharp, and Zan Gojcic. 3d gaussian ray tracing: Fast tracing of particle scenes. arXiv preprint, 2024. 2, 4, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 404, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 404, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 404, + 553, + 437 + ], + "type": "text", + "content": "[23] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 438, + 553, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 438, + 553, + 481 + ], + "spans": [ + { + "bbox": [ + 316, + 438, + 553, + 481 + ], + "type": "text", + "content": "[24] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting triangular 3d models, materials, and lighting from images. In CVPR, 2022. 1, 5, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 483, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 483, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 317, + 483, + 553, + 536 + ], + "type": "text", + "content": "[25] Steven G Parker, James Bigler, Andreas Dietrich, Heiko Friedrich, Jared Hoberock, David Luebke, David McAllister, Morgan McGuire, Keith Morley, Austin Robison, et al. Optix: a general purpose ray tracing engine. TOG, 2010. 2, 4, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 537, + 553, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 553, + 569 + ], + "type": "text", + "content": "[26] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mllps. In ICCV, 2021. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 571, + 553, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 571, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 317, + 571, + 553, + 623 + ], + "type": "text", + "content": "[27] Yahao Shi, Yanmin Wu, Chenming Wu, Xing Liu, Chen Zhao, Haocheng Feng, Jingtuo Liu, Liangjun Zhang, Jian Zhang, Bin Zhou, et al. Gir: 3d gaussian inverse rendering for relightable scene factorization. arXiv preprint, 2023. 1, 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 625, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 668 + ], + "type": "text", + "content": "[28] Pratul P Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T Barron. Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In CVPR, 2021. 1, 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[29] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint, 2021. 2, 5" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10951" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 612 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[30] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Trans. Image Process., 2004. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 149 + ], + "type": "text", + "content": "[31] Tong Wu, Jia-Mu Sun, Yu-Kun Lai, Yuewen Ma, Leif Kobbelt, and Lin Gao. Deferreds: Decoupled and editable gaussian splatting with deferred shading. arXiv preprint, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "text", + "content": "[32] Yunzhi Yan, Haotong Lin, Chenxu Zhou, Weijie Wang, Haiyang Sun, Kun Zhan, Xianpeng Lang, Xiaowei Zhou, and Sida Peng. Street gaussians for modeling dynamic urban scenes. arXiv preprint, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 239 + ], + "type": "text", + "content": "[33] Ziyi Yang, Yanzhen Chen, Xinyu Gao, Yazhen Yuan, Yu Wu, Xiaowei Zhou, and Xiaogang Jin. Sire-ir: Inverse rendering for brdf reconstruction with shadow and illumination removal in high-illuminance scenes. arXiv preprint, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 241, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 294, + 274 + ], + "type": "text", + "content": "[34] Zeyu Yang, Hongye Yang, Zijie Pan, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. arXiv preprint, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 276, + 294, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 318 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 318 + ], + "type": "text", + "content": "[35] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. In CVPR, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 319, + 294, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 319, + 294, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 319, + 294, + 363 + ], + "type": "text", + "content": "[36] Yao Yao, Jingyang Zhang, Jingbo Liu, Yihang Qu, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Neilf: Neural incident light field for physically-based material estimation. In ECCV, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 365, + 294, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 387 + ], + "type": "text", + "content": "[37] Yuxuan Yao, Zixuan Zeng, Chun Gu, Xiatian Zhu, and Li Zhang. Reflective gaussian splatting. In ICLR, 2025. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 388, + 294, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 420 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 420 + ], + "type": "text", + "content": "[38] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. NeurIPS, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 422, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 453 + ], + "type": "text", + "content": "[39] Zehao Yu, Torsten Sattler, and Andreas Geiger. Gaussian opacity fields: Efficient adaptive surface reconstruction in unbounded scenes. TOG, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 456, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 294, + 498 + ], + "type": "text", + "content": "[40] Jingyang Zhang, Yao Yao, Shiwei Li, Jingbo Liu, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Neilf++: Inter-reflectable light fields for geometry and material estimation. In ICCV, 2023. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 501, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 294, + 533 + ], + "type": "text", + "content": "[41] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 535, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 294, + 578 + ], + "type": "text", + "content": "[42] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM Trans. Graph., 2021. 6, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 580, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 580, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 580, + 294, + 612 + ], + "type": "text", + "content": "[43] Yuanqing Zhang, Jiaming Sun, Xingyi He, Huan Fu, Rongfei Jia, and Xiaowei Zhou. Modeling indirect illumination for inverse rendering. In CVPR, 2022. 2, 6, 7, 8" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10952" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/IRIS_ Inverse Rendering of Indoor Scenes from Low Dynamic Range Images/53c93036-c714-4f00-9960-2f72fce37288_content_list.json b/2025/IRIS_ Inverse Rendering of Indoor Scenes from Low Dynamic Range Images/53c93036-c714-4f00-9960-2f72fce37288_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..710a82396d60d475911420d7d24432e890c316e4 --- /dev/null +++ b/2025/IRIS_ Inverse Rendering of Indoor Scenes from Low Dynamic Range Images/53c93036-c714-4f00-9960-2f72fce37288_content_list.json @@ -0,0 +1,1703 @@ +[ + { + "type": "text", + "text": "IRIS: Inverse Rendering of Indoor Scenes from Low Dynamic Range Images", + "text_level": 1, + "bbox": [ + 112, + 130, + 883, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chih-Hao Lin1,2", + "bbox": [ + 138, + 179, + 272, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jia-Bin Huang $^{1,3}$", + "bbox": [ + 292, + 180, + 428, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhengqin Li", + "bbox": [ + 449, + 180, + 555, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhao Dong1", + "bbox": [ + 578, + 180, + 676, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Christian Richardt1", + "bbox": [ + 700, + 180, + 852, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tuotuo Li", + "bbox": [ + 156, + 199, + 245, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Michael Zollhöfer1", + "bbox": [ + 267, + 199, + 419, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Johannes Kopf1", + "bbox": [ + 442, + 199, + 566, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shenlong Wang", + "bbox": [ + 589, + 199, + 723, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Changil Kim", + "bbox": [ + 746, + 199, + 857, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Meta $^{2}$ University of Illinois Urbana-Champaign $^{3}$ University of Maryland College Park https://irisldr.github.io/", + "bbox": [ + 133, + 222, + 862, + 268 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a25f7bdb278f3a437c12f874cec517b605582adcde0b705c67620274d7cdb59c.jpg", + "image_caption": [ + "Figure 1. Physically-based inverse rendering from LDR images. Our method takes LDR input images and estimates high-quality spatially-varying HDR lighting, physically-based material properties, and a camera response function (CRF). This decomposition allows us to insert objects or new light sources, and to relight the scene in a physically accurate manner." + ], + "image_footnote": [], + "bbox": [ + 94, + 286, + 235, + 523 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/45a2c0323811594e2d8734d9a8db03655eac18ee26fb98076cda118ad6bd52c0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 246, + 287, + 475, + 523 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1f18b3f222474e7922aeede920737cf91aeaafc49426652883f4a83acc5ab410.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 488, + 287, + 691, + 523 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1d57008291cd2f4717b0bfdee3c4d4972004e91e749557bd53408542007a3a20.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 702, + 287, + 903, + 523 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 577, + 643, + 592 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 607, + 326, + 622 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Inverse rendering seeks to recover 3D geometry, surface material, and lighting from captured images, enabling advanced applications such as novel-view synthesis, relighting, and virtual object insertion. However, most existing techniques rely on high dynamic range (HDR) images as input, limiting accessibility for general users. In response, we introduce IRIS, an inverse rendering framework that recovers the physically based material, spatially-varying HDR lighting, and camera response functions from multi-view, low-dynamic-range (LDR) images. By eliminating the dependence on HDR input, we make inverse rendering technology more accessible. We evaluate our approach on real-world and synthetic scenes and compare it with state-of-the-art methods. Our results show that IRIS effectively recovers HDR lighting, accurate material, and plausible camera response functions, supporting photorealistic relighting and object insertion.", + "bbox": [ + 89, + 641, + 485, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Physically based inverse rendering enables reconstructing realistic material properties and lighting in scenes. This decomposition is valuable for various applications, including relighting, material editing, and realistic object insertion. However, most existing inverse rendering methods require input images with high dynamic range (HDR) to capture the full light transport in the scene. This poses a significant barrier to broader adoption, as HDR capture typically requires specialized hardware or merging multiple aligned LDR images through exposure bracketing [35].", + "bbox": [ + 511, + 608, + 906, + 760 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Most common imaging sensors do not capture sufficient dynamic range for scenes' light-emitting and dark regions. Moreover, cameras often convert raw sensor readings to 8-bit low-dynamic-range (LDR) images for storage and transmission. The non-linear mapping and quantization lead to the additional loss of lighting information. The complex light transport in indoor environments makes reconstructing the original HDR lighting - critical for faithful inverse rendering - even more challenging. If material and lighting estimation", + "bbox": [ + 509, + 763, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "465", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "could be achieved from a \"casual\" capture using standard devices like cameras or phones, inverse rendering would become far more accessible to a wider range of users.", + "bbox": [ + 89, + 90, + 482, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Many state-of-the-art inverse rendering methods rely on HDR inputs to accurately recover material properties and lighting [43, 48, 52]. Some approaches attempt to overcome this issue by taking LDR images as inputs and solving for the lighting. However, these methods often make simplifying assumptions, such as infinitely distant light sources [42], requiring additional inputs like emitter masks [55], or neglecting multi-bounce light transport [45, 52, 55]. As a result, these methods struggle to handle the complex light transport in indoor scenes and often fail to produce high-quality surface material and lighting.", + "bbox": [ + 89, + 138, + 483, + 304 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, we introduce IRIS, an inverse rendering method for indoor scenes, using only multi-view LDR images. IRIS models tone mapping (i.e., HDR-to-LDR conversion), allowing us to work directly with LDR inputs. IRIS automatically identifies emitters and reconstructs spatially varying HDR lighting through physically based inverse rendering, which is crucial for faithful material estimation. Jointly estimating lighting, surface material, and camera response function (CRF) leads to unstable optimization due to ambiguities. To tackle this, we design a novel optimization strategy that effectively resolves these ambiguities, enabling high-quality estimation. Our core contributions are:", + "bbox": [ + 89, + 306, + 483, + 487 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. IRIS faithfully estimates spatially-varying HDR lighting, physically-based materials, and camera response function from LDR images and outperforms several state-of-the-art inverse rendering methods.", + "2. We explicitly model LDR image formation in our (inverse) rendering pipeline so that LDR images can be used directly, broadening the accessibility of high-quality inverse rendering.", + "3. Finally, IRIS is extensively evaluated on synthetic and real-world scenes, demonstrating diverse and realistic view synthesis, relighting, and object insertion." + ], + "bbox": [ + 89, + 488, + 483, + 655 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 672, + 232, + 689 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Data-driven inverse rendering. Inverse rendering seeks to reconstruct a scene's geometry, material, and lighting from single or multiple images. This process is notoriously challenging and ill-posed due to the inherent ambiguity. Numerous methods employ deep priors learned from large-scale datasets for different tasks, including intrinsic image decomposition [20, 21], SVBRDF estimation [6, 19, 22, 50, 56], lighting estimation [8, 23, 38], lighting editing [25] and relighting [34]. These learning-based methods take a single or a few images [2, 38] as inputs, reducing capturing requirements of classical measurement-based methods [9]. However, high-quality datasets are essential for data-driven methods to achieve compelling generalization ability. Synthetic", + "bbox": [ + 89, + 704, + 483, + 900 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/623c122cc4fd7ca90e59fd26643fa3100c61d5cc29df715dd759e542f954791f.jpg", + "image_caption": [ + "Surface material" + ], + "image_footnote": [], + "bbox": [ + 553, + 88, + 712, + 162 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f4355db530dab1e2421567746725b4d6766152de6687f53fc9f26d04170ff64a.jpg", + "image_caption": [ + "Lighting" + ], + "image_footnote": [], + "bbox": [ + 722, + 88, + 864, + 162 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4f2cebbd55fa7f3e052cf7c13277e25cf9d0781996e466523c66e457c41637e3.jpg", + "image_caption": [ + "Surface material", + "Figure 2. Limitation of SOTA. A typical image formation process causes the loss of lighting information, posing challenges in inverse rendering. FIPT [43] assumes HDR input and NeILF [45] ignores multi-bounce light transport. Both methods fail to estimate accurate material (red boxes) and HDR lighting (yellow boxes). We demonstrate significantly better results (Figure 1)." + ], + "image_footnote": [], + "bbox": [ + 553, + 169, + 712, + 243 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ead907e3468f92182f191e4607671aa13abe8c1f98af1d41a665f260bbf4a536.jpg", + "image_caption": [ + "Lighting" + ], + "image_footnote": [], + "bbox": [ + 722, + 169, + 862, + 243 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "datasets are often used for training because ground-truth labeling is challenging to obtain in real environments. Despite efforts to create photorealistic synthetic datasets [20, 24, 36], the inherent diversity of real-world environments, especially in complex indoor scenes, still causes a substantial domain gap. Furthermore, these methods do not reconstruct complete 3D scenes and fail to render consistent novel views under novel illuminations. Our method utilizes prediction from a data-driven method [56] for initialization and reconstructs 3D geometry, material and lighting with a carefully designed optimization framework, demonstrating superior performance for real-world scenes.", + "bbox": [ + 511, + 349, + 908, + 531 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Optimization-based inverse rendering. Differentiable path tracing [14] enables the direct optimization of scene parameters with physically-based multi-bounce light transport. Recent methods, including MILO [48], IPT [1], and follow-up work [33] exploit this to jointly optimize BRDF and lighting emission, accounting for global light transport. FIPT [43] utilizes a factored light transport formulation for BRDF estimation and emitter detection. Nevertheless, these methods assume piecewise constant materials or are designed to take high-dynamic range images as input, and the robustness is limited in real-world scenes. Several recent NeRF-based approaches [12, 15, 26, 32, 40, 42, 53, 54] parameterize materials as a neural field and optimize with a volume rendering formulation. They tend to assume distant lighting and cannot handle spatially varying lighting in indoor environments well. NeILF [45] and NeILF++ [52] represent spatially-varying lighting as a neural field, $\\mathrm{I}^2$ -SDF [55] constructs lighting from given emitter masks. While handling more complex lighting, these works require additional assumptions (e.g., HDR input, emitter masks), and single-bounce ray tracing cannot model complex light transport. In contrast, our method automatically identifies emitters and reconstructs HDR lighting from casual (LDR) images, while achieving high-quality material and lighting estimation that", + "bbox": [ + 511, + 537, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "466", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "outperforms prior work in this practical scenario.", + "bbox": [ + 89, + 90, + 415, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "CRF and HDR estimation. The complete information about the illumination is usually lost within the photography pipeline. Estimating the HDR lighting and CRF enables us to edit photos and achieve realistic object insertion. Debevec and Malik [5] propose to take multiple photos with varying exposure levels to recover the CRF and HDR image. Some previous works [27, 28, 44, 51] predict HDR images with a data-driven approach by learning the LDR-HDR mapping from a single image. However, these methods rely on HDR data for training, and the estimation is inconsistent across views and scenes. To handle the HDR-LDR mapping, FIPT [43] and RawNeRF [30] assume a known gamma correction function; NeILF++ [52] learn a single gamma correction parameter; HDR-NeRF [13] parameterizes tone-mapping function with an MLP. However, these works are designed to take HDR images or LDR with multiple exposure levels as input. We leverage physically-based rendering to recover spatially varying BRDF, HDR illumination, and CRF from LDR images. Our method generates realistic view synthesis, relighting, and object insertion.", + "bbox": [ + 89, + 109, + 485, + 412 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Background: LDR image formation", + "text_level": 1, + "bbox": [ + 89, + 426, + 413, + 443 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Real-world radiance exhibits high dynamic range (HDR) and contains bright light sources and dark regions. To preserve the complete lighting information, HDR images are typically produced with a high-end camera, laborious exposure bracketing [5], and are stored in specialized format (16 or 32 bits per color channel). However, capturing HDR is often impractical for casual users and is typically absent in standard photography assets. Due to the limitations of standard sensors and displays, the radiance is frequently processed with multiple stages to produce low dynamic range (LDR) images stored in more lightweight formats (such as JPEG or PNG at 8 bits and RAW at 12 bits per color channel). We describe the major steps in this process below.", + "bbox": [ + 89, + 450, + 483, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dynamic range clipping. Given the scene irradiance $E$ and exposure time $\\Delta t$ , the observed radiance is clipped at a certain maximum value due to the sensor limitation, which could be formulated as: $Z_{c} = \\min (E\\Delta t,1)$ . Due to the clipping operation, the lighting information is lost for overexposed/saturated pixels.", + "bbox": [ + 89, + 651, + 483, + 743 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Non-linear mapping of CRF. To enhance the image's visual quality and match the human perception of the scene, the clipped intensity is further transformed with a non-linear camera response function (CRF) to produce LDR pixel values: $Z = \\mathrm{CRF}(Z_c) = \\mathrm{CRF}(\\min(E\\Delta t, 1))$ . Please refer to Debevec and Malik [5] for a more detailed imaging pipeline explanation.", + "bbox": [ + 89, + 744, + 485, + 851 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Challenges in inverse rendering. Real-world scenes often exhibit extreme variances in radiance, making it challenging to capture these scenes in LDR images without under", + "bbox": [ + 89, + 854, + 485, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "exposing or saturating them. Due to the clipping and nonlinear CRF mapping, critical lighting information is lost, posing challenges in previous works, illustrated in Figure 2. In this work, we propose an inverse rendering framework that jointly recovers HDR lighting and CRF from LDR images.", + "bbox": [ + 511, + 90, + 908, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Representation and rendering", + "text_level": 1, + "bbox": [ + 511, + 179, + 790, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Camera response function (CRF). The CRF is crucial for recovering original radiance from 8-bit LDR images. However, the CRF varies with each camera and is often unknown to users. Thus, deducing the CRF from LDR image observations is critical to our solution. We use Gross et al.'s [11] empirical model of response (EMoR) for CRF modeling [27]. They collected a database of 201 real-world CRFs of various devices and parameterized them as vectors. A mean curve $\\bar{\\mathbf{g}}$ and PCA basis $\\mathbf{g}_b$ are calculated from the database, and we parameterize a learnable CRF as", + "bbox": [ + 511, + 207, + 908, + 358 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {g} = \\bar {\\mathbf {g}} + \\Sigma_ {b} w _ {b} \\mathbf {g} _ {b}, \\mathbf {g} \\in \\mathbb {R} ^ {1 0 2 4}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 364, + 906, + 383 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\{w_{b}\\}$ weigh each basis curve and are learnable parameters. The non-linear function is discretized as $\\mathrm{CRF}(x) = \\mathrm{LERP}(\\{g_k\\}_k, x)$ , which linearly interpolates the uniformly sampled CRF $\\{g_k\\}_{k=0}^{1023} \\in [0,1]^{1024}$ , independently per color channel. We make this design choice because it covers most of the CRF space, and the simplicity makes it easy to enforce regularization, which brings huge advantage in the complex optimization of inverse rendering.", + "bbox": [ + 511, + 391, + 908, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Spatially-varying lighting. While environment maps handle object-centric and outdoor scenes well [15, 39, 42, 53], they cannot capture complex spatially-varying lighting of indoor scenes. To handle this, we define direct lighting on the scene mesh and measure emission as view-independent radiance: $\\mathbf{L}_{\\mathrm{e}}(\\mathbf{x}) \\in \\mathbb{R}^3$ . The parameterization of $\\mathbf{L}_{\\mathrm{e}}$ is detailed in our supplementary material. We also define an emitter mask $M_{\\mathrm{e}}(\\mathbf{x}) \\in \\{0,1\\}$ to denote whether a surface point $\\mathbf{x}$ is on an emitter. Global illumination is represented as a view-independent surface light field (SLF): $\\mathbf{L}_{\\mathrm{SLF}}(\\mathbf{x}) \\in \\mathbb{R}^3$ . The sRGB pixel colors of input LDR images are linearized with an inverse CRF mapping: $\\mathbf{I}_{\\mathrm{linear}} = \\mathrm{CRF}^{-1}(\\mathbf{I}_{\\mathrm{LDR}})$ . If a voxel contains no surface, it is set to zero; otherwise, it collects the average linearized radiances across all input images with ray casting. This greatly accelerates the approximation of global illumination without requiring multi-bounce ray tracing.", + "bbox": [ + 511, + 515, + 908, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Material BRDF. We represent spatially-varying materials using the Cook-Torrance BRDF [4]. The surface albedo $\\mathbf{a}(\\mathbf{x})\\in \\mathbb{R}^3$ , roughness $\\sigma (\\mathbf{x})\\in \\mathbb{R}$ , and metallicity $m(\\mathbf{x})\\in \\mathbb{R}$ model the diffuse reflectance $\\mathbf{k_d} = \\mathbf{a}\\cdot (1 - m)$ and specular reflectance $\\mathbf{k_s} = 0.04\\cdot (1 - m)$ , and are parameterized as a neural field $\\mathbf{f}\\colon \\mathbf{x}\\mapsto (\\mathbf{a},\\sigma ,m)$ , implemented with multi-resolution hash encoding and an MLP [31].", + "bbox": [ + 511, + 760, + 908, + 867 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Factorized light transport. To render images from scene geometry, material, and lighting information, We follow", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "467", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/390b3c22db5234603ac9224731083b54eecb757ed4d85522755cf096d9eff524.jpg", + "image_caption": [ + "Figure 3. Framework Overview. Given multi-view posed LDR images and a surface mesh, our inverse rendering pipeline is divided into two main stages. In the initialization stage, we initialize the BRDF (Section 5.1), extract a surface light field (Section 4), and estimate emitter geometry (Equation (3)). In the optimization stage, we first recover HDR radiance from the LDR input (Section 5.2), then bake shading maps (Section 5.3), and jointly optimize BRDF and CRF parameters (Section 5.4). The improved parameters are used to refine the emission again. These three steps are repeated until convergence." + ], + "image_footnote": [], + "bbox": [ + 102, + 90, + 903, + 300 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the rendering equation [16] to model physically-based light transport for realistic rendering:", + "bbox": [ + 89, + 392, + 482, + 422 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {L} _ {\\mathrm {o}} (\\mathbf {x}, \\omega_ {\\mathrm {o}}) = \\mathbf {L} _ {\\mathrm {e}} (\\mathbf {x}, \\omega_ {\\mathrm {o}}) + \\int_ {\\Omega^ {+}} \\mathbf {L} _ {\\mathrm {i}} (\\mathbf {x}, \\omega_ {\\mathrm {i}}) f (\\mathbf {x}, \\omega_ {\\mathrm {i}}, \\omega_ {\\mathrm {o}}) d \\omega_ {\\mathrm {i}}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 431, + 483, + 462 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{L}_{\\mathrm{o}}$ is the radiance observed along a ray at a 3D position $\\mathbf{x}$ in direction $\\omega_{\\mathrm{o}}$ , $\\mathbf{L}_{\\mathrm{e}}$ is the emission term, $\\mathbf{L}_{\\mathrm{r}} = \\int_{\\Omega^{+}}\\mathbf{L}_{\\mathrm{i}}(\\mathbf{x},\\omega_{\\mathrm{i}})f(\\mathbf{x},\\omega_{\\mathrm{i}},\\omega_{\\mathrm{o}})d\\omega_{\\mathrm{i}}$ is the reflectance term, and $f(\\mathbf{x},\\omega_{\\mathrm{i}},\\omega_{\\mathrm{o}})$ is the BRDF. The recursive nature of the rendering equation is approximated with Monte-Carlo path tracing [16, 18]. However, modeling multiple bounces is still computationally expensive and unstable in the inverse rendering process. To enhance the efficiency and robustness of the optimization, we follow FIPT [43] and adopt factorized light transport [17, 37, 41], which precomputes shading maps $\\mathbf{L}_{\\mathrm{d}}$ , $\\mathbf{L}_{\\mathrm{s}}^{0}$ , $\\mathbf{L}_{\\mathrm{s}}^{1}$ and linearly interpolates with roughness to obtain final rendering. Detailed notation and formulation are available in the supplemental material.", + "bbox": [ + 89, + 469, + 483, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5. Inverse Rendering from LDR Images", + "text_level": 1, + "bbox": [ + 89, + 679, + 426, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given multi-view posed LDR images $\\{\\mathbf{I}_i\\}$ captured with the same or varying exposure levels $\\{\\Delta t_i\\}$ , our goal is to estimate spatially-varying HDR lighting, surface material (as albedo, roughness, and metallic), and the camera response function (CRF) shared by all input images. We assume that the primary light sources are observed in the input images, following the recent literature [43, 48, 55].", + "bbox": [ + 89, + 703, + 482, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Previous approaches perform differentiable path tracing during the material and lighting optimization [1, 33], but the number of samples is limited due to heavy computation. This introduces high variance and unstable estimation. The factorized formulation [17, 37, 41, 43] splits light transport into BRDF parameters and shading, and enables updating", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "them alternately. While the method is shown to be more stable and produce high-quality decompositions, it requires HDR images as input to aggregate and bake shading maps, detect emitters, and estimate HDR emissions.", + "bbox": [ + 511, + 392, + 906, + 452 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address this ill-posed optimization problem, we propose a multi-stage optimization strategy alternatingly estimating spatially varying lighting, material, and CRF. Specifically, we (1) initialize the BRDF with an off-the-shelf monocular inverse rendering method (Section 5.1), (2) recover HDR lighting from LDR with physically-based rendering (Section 5.2), (3) bake the diffuse and specular shading (Section 5.3), and (4) jointly estimate BRDF and CRF (Section 5.4). We repeat Steps 2-4 in that order until all estimates are converged. Figure 3 illustrates our framework.", + "bbox": [ + 511, + 455, + 908, + 607 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Geometry reconstruction. From input images with camera poses, we reconstruct the geometry with BakedSDF [46], and use normal estimation from off-the-shelf method [7] for regulating surface geometry, following MonoSDF [49].", + "bbox": [ + 511, + 614, + 908, + 675 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5.1. BRDF Initialization", + "text_level": 1, + "bbox": [ + 511, + 693, + 702, + 709 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For each input image $\\mathbf{I}_i$ , we predict a 2D albedo map $\\mathbf{A}_i$ using an off-the-shelf single-image albedo estimation method [56] and averaging per-pixel albedo values within each semantic part. Given $\\{\\mathbf{A}_i\\}$ , we initialize albedo $\\mathbf{a}(\\mathbf{x})$ by minimizing the error between the projected albedo and the estimated albedo across all input images: $\\min_{\\mathbf{a}} \\sum_i \\| \\pi_i(\\mathbf{a}) - \\mathbf{A}_i \\|_2$ . Here, $\\pi_i$ represents the perspective projection onto the image space of $\\mathbf{I}_i$ by intersecting camera rays with the surface mesh of the input scene and reading $\\mathbf{a}$ at their first intersections. We obtain semantic category labels from part IDs in synthetic scenes or using Mask2Former [3] for real scenes. Roughness and metallicity are both initialized to 0.5.", + "bbox": [ + 511, + 719, + 908, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "468", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ec1874197a8d9d93fdee5bd1aaf6acfb9b4cf15bcd686a521be7d86c6557cc76.jpg", + "image_caption": [ + "Figure 4. Emitter geometry estimation $M_{\\mathrm{e}}(\\mathbf{x})$ . The point $\\mathbf{x}_1$ on the window is saturated across all input views and thus identified as an emitter. The point $\\mathbf{x}_2$ on the table is reflective and saturated in some views (e.g., view 2) but not in others and is NOT an emitter." + ], + "image_footnote": [], + "bbox": [ + 130, + 88, + 436, + 217 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. HDR Emission Restoration", + "text_level": 1, + "bbox": [ + 89, + 279, + 334, + 292 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given that emitters appear bright across views and produce saturated pixels in LDR images, we estimate the binary emitter geometry $M_{\\mathrm{e}}(\\mathbf{x})$ on the surface mesh as follows:", + "bbox": [ + 89, + 301, + 483, + 347 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nM _ {\\mathrm {e}} (\\mathbf {x}) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} \\frac {1}{N} \\sum_ {i} v _ {i} (\\mathbf {x}) \\mathbf {I} _ {i} \\left(\\pi_ {i} (\\mathbf {x})\\right) \\geq 0. 9 9 \\\\ 0 & \\text {o t h e r w i s e ,} \\end{array} \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 111, + 354, + 483, + 395 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $v(\\cdot)\\in \\{0,1\\}$ indicates visibility. See Figure 4.", + "bbox": [ + 89, + 402, + 442, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The HDR intensity is crucial for inverse rendering but is missing in the LDR images. To recover the HDR lighting, we perform physically-based rendering with single-bounce path tracing to generate the images (2). To estimate the radiance at the point $\\mathbf{x}$ observed from a camera ray, we trace $N = 128$ secondary rays from $\\mathbf{x}$ , sampled with the current BRDF estimates. The resulting radiance is approximated by:", + "bbox": [ + 89, + 417, + 483, + 525 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {L} _ {\\mathrm {r}} \\left(\\mathbf {x}, \\omega_ {\\mathrm {o}}\\right) = \\sum_ {i = 1} ^ {N} \\mathbf {L} _ {\\text {e n d}} \\left(\\mathbf {x} ^ {\\prime}\\right) f \\left(\\mathbf {x}, \\omega_ {\\mathrm {i}}, \\omega_ {\\mathrm {o}}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 532, + 482, + 573 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {L} _ {\\text {e n d}} (\\mathbf {x}) = \\left\\{ \\begin{array}{l l} \\mathbf {L} _ {\\mathrm {e}} (\\mathbf {x}) & \\text {i f} M _ {\\mathrm {e}} (\\mathbf {x}) = 1 \\\\ \\mathbf {L} _ {\\text {S L F}} (\\mathbf {x}) & \\text {o t h e r w i s e ,} \\end{array} \\right. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 578, + 482, + 616 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{x}' = \\mathbf{x} + d\\cdot \\boldsymbol{\\omega}_{\\mathrm{i}}$ is the intersected point of the secondary ray along direction $\\boldsymbol{\\omega}_{\\mathrm{i}}$ . Intuitively, we retrieve the radiance from the learnable emission radiance $\\mathbf{L}_{\\mathrm{e}}(\\mathbf{x})$ if the secondary ray hits an emitter directly, or otherwise from the pre-computed surface light field $\\mathbf{L}_{\\mathrm{SLF}}(\\mathbf{x})$ to approximate the global illumination. Figure 5 shows an illustration.", + "bbox": [ + 89, + 625, + 483, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To recover the HDR emitter radiance, we minimize the photometric loss with gradient descent: $\\min_{\\mathbf{L}_{\\mathrm{e}}}\\mathcal{L}_{\\mathrm{photo}}$ , where", + "bbox": [ + 89, + 715, + 482, + 747 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {p h o t o}} = \\sum_ {i} \\| \\mathrm {C R F} \\left(\\min \\left(\\mathbf {L} _ {\\mathrm {o}} \\left(\\mathbf {x}, \\omega_ {\\mathrm {o}}\\right) \\Delta t _ {i}, 1\\right)\\right) - \\mathbf {I} _ {i} \\| _ {2}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 755, + 483, + 785 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The emitter radiance $\\mathbf{L}_{\\mathrm{e}}(\\mathbf{x})$ is enhanced to match the captured images, while the BRDF and CRF are fixed in this stage. Intuitively, in order to minimize the photometric loss with physically-based rendering, the emitter intensity $\\mathbf{L}_{\\mathrm{e}}(\\mathbf{x})$ should be increased significantly and enhance the overall brightness of the rendered image. Therefore, the HDR lighting is recovered in this process.", + "bbox": [ + 89, + 794, + 485, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/afa70ad937f130915d42a5e9b0ab63b70c9583aa7aa8fbf7f494b64e91136125.jpg", + "image_caption": [ + "Figure 5. HDR emission restoration. Top: Ray sampling process. Learnable direct lighting $\\mathbf{L}_{\\mathrm{e}}$ is retrieved if a ray hits an emitter (e.g., window) or $\\mathbf{L}_{\\mathrm{SLF}}$ is retrieved otherwise (e.g., wall). Bottom: HDR restoration process. By performing differentiable physically-based rendering, the photometric loss enhances the emitter intensity $\\mathbf{L}_{\\mathrm{e}}$ ." + ], + "image_footnote": [], + "bbox": [ + 566, + 88, + 846, + 222 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.3. Shading Baking", + "text_level": 1, + "bbox": [ + 513, + 309, + 671, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After the HDR emission is recovered in the previous stage, we bake the diffuse and specular shading maps $\\mathbf{L}_{\\mathrm{d}}$ , $\\mathbf{L}_{\\mathrm{s}}^{0}$ , $\\mathbf{L}_{\\mathrm{s}}^{1}$ , following factorized light transport (details in supplementary material). Given a surface point intersected by a camera ray, multiple rays are sampled according to the current (fixed) BRDF. The incident radiance is computed with multibounce path tracing, where a ray keeps tracing if it hits a non-emissive specular surface and stops otherwise. The radiance of the endpoint follows Equation (5).", + "bbox": [ + 511, + 332, + 906, + 468 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.4. BRDF & CRF Optimization", + "text_level": 1, + "bbox": [ + 513, + 479, + 766, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With the baked shading maps $\\mathbf{L}_{\\mathrm{d}}$ , $\\mathbf{L}_{\\mathrm{s}}^{0}$ , and $\\mathbf{L}_{\\mathrm{s}}^{1}$ , the albedo, metallicity, roughness, and CRF are jointly optimized. The optimization involves the following objective function:", + "bbox": [ + 511, + 501, + 906, + 547 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\mathbf {a}, m, \\sigma , \\mathbf {g}} \\mathcal {L} _ {\\text {p h o t o}} + \\lambda_ {\\mathrm {a}} \\mathcal {L} _ {\\text {a l b e d o}} + \\lambda_ {\\mathrm {c}} \\mathcal {L} _ {\\text {C R F}} + \\lambda_ {\\mathrm {m}} \\mathcal {L} _ {\\text {m a t}}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 552, + 560, + 906, + 583 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{\\mathrm{a}} = 0.01$ $\\lambda_{\\mathrm{c}} = 0.001$ , and $\\lambda_{\\mathrm{m}} = 0.005$", + "bbox": [ + 511, + 595, + 826, + 611 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Photometric loss. We perform physically-based rendering, and the photometric loss $\\mathcal{L}_{\\mathrm{photo}}$ is computed as in Eq. 6.", + "bbox": [ + 511, + 614, + 906, + 646 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Albedo regularization. Motivated by the shift-scale-invariant loss in MonoSDF [49], we adopt a shift-invariant loss against the monocular albedo maps $\\{\\mathbf{A}_i\\}$ :", + "bbox": [ + 511, + 648, + 906, + 695 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {a l b e d o}} = \\sum_ {i} \\| \\pi_ {i} (\\mathbf {a}) - s _ {i} \\mathbf {A} _ {i} \\| _ {2}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 707, + 906, + 738 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $s_i$ is the scale used to align the two albedo maps.", + "bbox": [ + 511, + 750, + 879, + 765 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "CRF regularization. Based on the fair assumption that the learned CRF should not deviate significantly from the mean curve $\\bar{\\mathbf{g}}$ (Eq. 1), and that the curve should be monotonically increasing, we regularize the $L_{2}$ norm of the estimated coefficients of the PCA bases, and enforce the monotonicity of the sampled CRF as follows:", + "bbox": [ + 511, + 768, + 906, + 859 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {C R F}} = \\| \\mathbf {w} \\| _ {2} + \\sum_ {i} \\max \\left(g _ {i - 1} - g _ {i}, 0\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 872, + 906, + 904 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "469", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Material regularization. We apply roughness-metallicity regularization $\\mathcal{L}_{\\mathrm{mat}}$ , which enforces the consistency of surface roughness and metallic within the same semantic instances (e.g. chairs, tables). While the shading and emission are initially inaccurate and lead to erroneous BRDFs, we alternately update the BRDFs and produce better emission and shading, improving the BRDFs. Our alternating strategy yields more stable optimization and better estimation.", + "bbox": [ + 89, + 90, + 483, + 212 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Experiments", + "text_level": 1, + "bbox": [ + 89, + 229, + 225, + 247 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate our method on both synthetic and real-world data and compare it against several baseline methods. We analyze the quality of our inverse rendering, relighting, and novel-view synthesis capability. Please see our supplementary material for additional results and evaluations. We will release our code and data.", + "bbox": [ + 89, + 255, + 483, + 347 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We evaluate our method and baselines on the ScanNet++ [47] and FIPT [43] datasets. ScanNet++ contains 380 real-world indoor scenes and collects multi-view LDR images with a DSLR camera at a constant exposure level. We select four scenes that cover a variety of indoor scenes (e.g., office, bathroom). The FIPT dataset consists of four synthetic and two real-world scenes. The synthetic scenes provide ground-truth geometry and material properties for evaluation. We used the LDR images provided in the dataset, where the lighting is clipped in overly-exposed regions, compressed with a nonlinear CRF, quantized into 256 levels, and stored in 8-bit PNG format. We adopt the constant exposure setting to provide a fair comparison with baselines, as they cannot handle varying exposure. We further demonstrate that IRIS handles varying exposure levels of input LDR images and estimates CRF jointly.", + "bbox": [ + 89, + 351, + 483, + 593 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We compare our method with the following baselines. Li et al. [25] estimate BRDF, geometry (depth & normal), and lighting condition of an indoor scene from a single image using a data-driven approach. In addition to the single image, the model also takes emitter masks as input, and we provide the mask estimated by our method (3). NeILF [45] targets inverse rendering from known geometry and multiview images. The spatially-varying BRDF and illumination are parameterized as neural fields [29]. Physically-based rendering and single-bounce ray tracing are performed during the training and rendering process. The HDR-to-LDR tone mapping is modeled as a learnable gamma correction. The original FIPT [43] is designed for HDR input, and fails to estimate the emission mask from LDR images. Thus, we additionally provide the emission mask estimated by our method (3). We denote this modified version as \"FIPT*\".", + "bbox": [ + 89, + 597, + 483, + 838 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To further evaluate our performance, we compare the original FIPT with HDR input if the HDR information is available. This privileged information will make illumination estimation much more accurate.", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/be20db0f963098f42cfdca513b969458135d58b83b292b83d6b975225a58804e.jpg", + "table_caption": [ + "Table 1. BRDF-emission comparison on synthetic data. Numbers are averaged across four synthetic scenes with GT [43]. The best metrics among LDR methods are highlighted in bold." + ], + "table_footnote": [], + "table_body": "
MethodInputkda'σLe
PSNR ↑IoU ↑L2 ↓
Li et al [25]LDR16.7314.0111.300.352.29
NeILF [45]LDR16.8514.0216.96——
FIPT*LDR15.499.744.990.690.28
OursLDR22.3317.9221.380.690.12
FIPT [43]HDR29.9525.9826.370.860.03
", + "bbox": [ + 550, + 141, + 864, + 247 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.1. Inverse Rendering of Real Scenes", + "text_level": 1, + "bbox": [ + 511, + 258, + 805, + 276 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We provide a qualitative evaluation of inverse rendering in Figure 6. Conducting single-bounce ray tracing in the neural light field, NeILF [45] fails to remove the shadow in diffuse reflectance $\\mathbf{k}_{\\mathrm{d}}$ , estimates lower roughness $\\sigma$ on the walls than the mirror, and produces an incorrect emission mask. FIPT* struggles to recover surface material from LDR input, where accurate lighting information is missing. Please note that FIPT* takes emitter masks from Equation (3) as additional input, but does not recover HDR emission in the optimization procedure. On the other hand, IRIS recovers an almost shading-free diffuse field. Regarding roughness, our method effectively identifies specular surfaces with low roughness on the mirror, demonstrating better specular/rough material separation. Additionally, our proposed HDR emission estimation proves effective for real scenes, recovering HDR lighting sources (i.e., window and ceiling light). The restored HDR lighting makes the light transport more accurate and significantly enhances the performance of inverse rendering.", + "bbox": [ + 511, + 284, + 908, + 556 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.2. View Synthesis and Relighting of Real Scenes", + "text_level": 1, + "bbox": [ + 511, + 573, + 897, + 590 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare the relighting with FIPT* in Figure 7. FIPT* fails to estimate accurate surface material from LDR input and tends to recover low roughness for most surfaces, which unrealistically reflects the inserted light source and objects. To be more specific, the ground of the office (first scene) and the wall of the bathroom (second scene) look overly smooth and specular, which is not consistent with the observation and produces unrealistic reflections. In contrast, IRIS produces better relighting and object insertion quality thanks to better inverse rendering. The reflections on the specular surfaces (e.g., whiteboard, mirror) and the shading on the wall demonstrate the strength of IRIS for realistic relighting. Furthermore, the recovered HDR lighting can produce naturalistic specular reflection and shadows for inserted objects. Additionally, we insert moving light sources in the scene, shown in Figure 8, where the shadow on the ground and shading on the table change consistently. IRIS enables the photorealistic rendering of novel objects under diverse lighting and from arbitrary viewpoints, showcasing the significant potential for content creation applications.", + "bbox": [ + 511, + 598, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "470", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9f1ef19091189052fc0baf90f7b6eabf7bd4cdbacdb68c5b113ea16e503c1f19.jpg", + "image_caption": [ + "Figure 6. Decompositions of real scenes from ScanNet++ [47]. In addition to good reconstruction, IRIS produces a detailed diffuse map, identifies the mirror as a low-roughness region, and recovers HDR lighting from windows and ceiling light. Please note that the emission maps are normalized to visualize the relative intensity." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 906, + 358 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ecbc00fa12f23d8c3a13c4900109eefb7b3585b4dbdd8a5bf7df0ddf20890c7c.jpg", + "image_caption": [ + "Figure 7. Relighting and object insertion with real scenes from the ScanNet++ [47] and FIPT datasets [43]. When inserting new light sources (Relighting 1 & 2), IRIS not only changes shadows but also produces reflection of the light on the whiteboard (first scene) and mirror (second scene), enhancing realism substantially. The recovered HDR lighting also strengthens the reflection and shadow of inserted objects (\"Object Insertion\" column)." + ], + "image_footnote": [], + "bbox": [ + 89, + 407, + 906, + 743 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.3. Quantitative Evaluation", + "text_level": 1, + "bbox": [ + 89, + 813, + 312, + 829 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative evaluation of inverse rendering is difficult for real-world scenes as ground truth is not available. As a result, we evaluate IRIS and baselines on the synthetic scenes from FIPT, where ground truth of surface material, illumina-", + "bbox": [ + 89, + 839, + 485, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tion, and relighting are provided. Please note that we use the ground-truth mesh during optimization for all methods to minimize the influence of imperfect geometry and conduct a fair comparison. We compute the metrics of inverse rendering and provide the comparison in Table 1, which aggregates", + "bbox": [ + 511, + 813, + 908, + 891 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "471", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7bdac9d5775894e719933053cd92fa6664dc83509605d3211c2c508bdd9b28e2.jpg", + "image_caption": [ + "Reconstruction" + ], + "image_footnote": [], + "bbox": [ + 192, + 89, + 287, + 150 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1e0287af317380a0de0e428152b06857a3422a84bdee19af3d20194ece9e4dd8.jpg", + "image_caption": [ + "Relight 2", + "Figure 8. Relighting with a moving light ball in a real scene." + ], + "image_footnote": [], + "bbox": [ + 290, + 101, + 382, + 150 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/47f37bc9591563489e577d706a1323d34d2c5d36516a5372171188687e3a1486.jpg", + "image_caption": [ + "Relight 3" + ], + "image_footnote": [], + "bbox": [ + 383, + 101, + 475, + 150 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bfa617a85da7f2f229e9ff2a4ed943149c37afa3c93074556c5a557f6db6f6f7.jpg", + "image_caption": [ + "Figure 9. CRF estimation from real-world devices. The orange dashed line is the ground truth CRF, and the blue line our estimate." + ], + "image_footnote": [], + "bbox": [ + 96, + 176, + 187, + 256 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5739447e2d8337658fbe15bb0fd8b7090ac15092f8319022289fba06645e86a9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 192, + 176, + 284, + 257 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/977019b9a4a2a13ad98529e9ce91ccdd642e418274d9af61c74ba58c46c773aa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 290, + 176, + 380, + 257 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7cdb30f3648732ac98e9e51acd30aba19a9bc3a40bb08acbee964f6f28140033.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 385, + 176, + 475, + 257 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the results from the 4 synthetic FIPT scenes [43].", + "bbox": [ + 89, + 321, + 416, + 335 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We compare the roughness $\\sigma$ , diffuse reflectance $\\mathbf{k}_{\\mathrm{d}}$ among the diffuse surfaces, and material reflectance: $\\mathbf{a}^{\\prime}$ . We report the PSNR for these BRDF estimates. For emission, we calculate the intersection over the union (IoU) of the emission masks and compare the $L_{2}$ error of estimated emission maps. All the estimates are rendered in image space and compared with ground truth.", + "bbox": [ + 89, + 337, + 483, + 441 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 1 shows that our method surpasses all baseline methods that use LDR inputs regarding inverse rendering quality. Compared to the state-of-the-art single-view method [25], our method excels in all metrics across various scenes, indicating the superior decomposition of material properties from multi-view observations. Furthermore, even when the estimated emitter mask is provided for FIPT*, it fails to recover emission radiance, leading to compromised BRDF estimation accuracy. The $L_{2}$ error of emission maps shows that our proposed HDR restoration (Section 5.2) successfully recovers the HDR lighting from LDR observations, which leads to better light transport modeling and more accurate BRDF estimation. We also compare with FIPT taking HDR images as input (bottom row), which serves as a reference, indicating complete lighting information is crucial for accurate inverse rendering. IRIS recovers HDR and achieves the smallest metric gaps among all methods taking LDR input.", + "bbox": [ + 89, + 443, + 483, + 700 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We compare novel-view synthesis and relighting of all competing algorithms and report results in Table 2. The novel-view synthesis results show that IRIS is comparable with baseline methods [45], while exhibiting superior material and lighting estimation. Our main focus is relighting, and IRIS outperforms all baselines with LDR input, demonstrating the effectiveness of inverse rendering.", + "bbox": [ + 89, + 700, + 483, + 806 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.4. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 816, + 243, + 832 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We validate our design choices and provide an ablation study in Table 3, where we compute PSNR for material $(\\mathbf{k}_{\\mathrm{d}},\\mathbf{a}^{\\prime},\\sigma)$ and L2 for CRF curves and emission maps $\\mathbf{L}_{\\mathrm{e}}$ . The emitter mask estimation in Equation (3) identifies the emitter from", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/cb98fdb31e01ee6de5d0c06784e4948e45f850ce1c948d8877d774a060cfec54.jpg", + "table_caption": [ + "Table 2. Quantitative results of novel-view synthesis (NVS) and relighting on synthetic scenes with ground truth." + ], + "table_footnote": [], + "table_body": "
MethodInputPSNR ↑SSIM ↑LPIPS ↓
NVSNeILF [45]LDR30.0310.8990.180
FIPT*LDR15.1800.7440.437
OursLDR29.4680.9030.181
FIPT [43]HDR28.7600.9020.186
RelightingLi et al. [25]LDR22.4440.8130.398
FIPT*LDR11.6260.7050.359
OursLDR22.8610.8820.173
FIPT [43]HDR29.0120.9020.139
", + "bbox": [ + 550, + 128, + 864, + 258 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/338c4acc60e015e8b96780170cb5dd72344b9879fb92705de7d432b1f15e1080.jpg", + "table_caption": [ + "Table 3. Ablation study" + ], + "table_footnote": [], + "table_body": "
Methodkda'σLeCRF
PSNR ↑IoU ↑L2 ↓L2 ↓
- Emitter mask Me (Equation (3))16.5011.304.810.000.488.91
- CRF modeling (Equation (1))18.1813.5917.520.690.12—
- HDR restoration (Section 5.2)18.2014.407.310.690.286.23
- Lalbedo (Equation (8))21.5917.3617.430.690.152.42
- LCRF (Equation (9))14.4912.8125.590.690.139.40
Full model (Ours)22.3317.9221.380.690.122.35
", + "bbox": [ + 514, + 279, + 903, + 380 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "LDR observations and improves the emission map in the second row. CRF estimation is critical for modeling HDR-to-LDR conversion and helps inverse rendering in the third row. Nevertheless, the HDR restoration (Section 5.2) plays an important role in recovering accurate light transport from the captured images. Once the HDR light is restored by our proposed approach, the material (especially roughness $\\sigma$ ), CRF, and emission map are improved significantly. The results demonstrate the significance of each component and show that inverse rendering from casually captured LDR images benefits from our proposed framework.", + "bbox": [ + 511, + 392, + 906, + 559 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.5. CRF Estimation", + "text_level": 1, + "bbox": [ + 511, + 573, + 674, + 588 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To further validate the robustness of our method, we generate LDR images from FIPT's synthetic HDR scenes using different CRFs collected from real-world devices [10]. In this evaluation, we perform inverse rendering with IRIS and plot the CRF. The results in Figure 9 show that our method can generalize to various imaging pipelines. Please refer to the supplementary details for more analysis of CRF estimation.", + "bbox": [ + 511, + 597, + 906, + 704 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 720, + 633, + 737 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We presented IRIS, an inverse rendering framework for indoor scenes from casually captured LDR images. Our key innovations include HDR lighting restoration, CRF estimation, surface material estimation, and alternating optimization to recover all scene properties accurately. IRIS estimates accurate surface material, spatially varying HDR illumination, and CRF from real-world LDR images. As a result, IRIS enables photorealistic rendering under diverse lighting and from arbitrary viewpoints, showcasing the significant potential for content creation applications.", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "472", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Dejan Azinovic, Tzu-Mao Li, Anton Kaplanyan, and Matthias Nießner. Inverse path tracing for joint material and lighting estimation. In CVPR, 2019. 2, 4", + "[2] Sai Bi, Zexiang Xu, Pratul Srinivasan, Ben Mildenhall, Kalyan Sunkavalli, Miloš Hasan, Yannick Hold-Geoffroy, David Kriegman, and Ravi Ramamoorthi. Neural reflectance fields for appearance acquisition. arXiv:2008.03824, 2020. 2", + "[3] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2022. 4", + "[4] Robert L Cook and Kenneth E. Torrance. A reflectance model for computer graphics. ACM Transactions on Graphics (ToG), 1982. 3", + "[5] Paul E. Debevec and Jitendra Malik. Recovering high dynamic range radiance maps from photographs. In SIGGRAPH, 1997. 3", + "[6] Valentin Deschaintre, Miika Aittala, Fredo Durand, George Drettakis, and Adrien Bousseau. Single-image SVBRDF capture with a rendering-aware deep network. ACM Transactions on Graphics (ToG), 37(4):128:1-15, 2018. 2", + "[7] Ainaz Eftekhar, Alexander Sax, Jitendra Malik, and Amir Zamir. Omnidata: A scalable pipeline for making multi-task mid-level vision datasets from 3D scans. In ICCV, 2021. 4", + "[8] Marc-Andre Gardner, Kalyan Sunkavalli, Ersin Yumer, Xiaohui Shen, Emiliano Gambaretto, Christian Gagné, and Jean-François Lalonde. Learning to predict indoor illumination from a single image. ACM Transactions on Graphics, 36(6): 176:1-14, 2017. 2", + "[9] Dan B Goldman, Brian Curless, Aaron Hertzmann, and Steven M Seitz. Shape and spatially-varying BRDFs from photometric stereo. IEEE Transactions on Pattern Analysis and Machine Intelligence, 32(6):1060-1071, 2009. 2", + "[10] Michael D Grossberg and Shree K Nayar. What is the space of camera response functions? In CVPR, 2003. 8", + "[11] Michael D. Grossberg and Shree K. Nayar. Modeling the space of camera response functions. IEEE TPAMI, 26(10): 1272-1282, 2004. 3", + "[12] Jon Hasselgren, Nikolai Hofmann, and Jacob Munkberg. Shape, light & material decomposition from images using monte carlo rendering and denoising. NeurIPS, 2022. 2", + "[13] Xin Huang, Qi Zhang, Ying Feng, Hongdong Li, Xuan Wang, and Qing Wang. HDR-NeRF: High dynamic range neural radiance fields. In CVPR, 2022. 3", + "[14] Wenzel Jakob, Sébastien Speierer, Nicolas Roussel, Merlin Nimier-David, Delio Vicini, Tizian Zeltner, Baptiste Nicolet, Miguel Crespo, Vincent Leroy, and Ziyi Zhang. Mitsuba 3 renderer, 2022. https://mitsuba-renderer.org. 2", + "[15] Haian Jin, Isabella Liu, Peijia Xu, Xiaoshuai Zhang, Songfang Han, Sai Bi, Xiaowei Zhou, Zexiang Xu, and Hao Su. TensoIR: Tensorial inverse rendering. CVPR, 2023. 2, 3", + "[16] James T Kajiya. The rendering equation. In SIGGRAPH, 1986. 4", + "[17] Jaroslav Krivánek and Pascal Gautron. Practical global illumination with irradiance caching. Springer, 2022. 4" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[18] Eric Lafortune. Mathematical models and Monte Carlo algorithms for physically based rendering. PhD thesis, Katholieke Universiteit Leuven, 1996. 4", + "[19] Xiao Li, Yue Dong, Pieter Peers, and Xin Tong. Modeling surface appearance from a single photograph using self-augmented convolutional neural networks. ACM Transactions on Graphics (ToG), 36(4):45:1-11, 2017. 2", + "[20] Zhengqi Li and Noah Snavely. CGIntrinsics: Better intrinsic image decomposition through physically-based rendering. In ECCV, 2018. 2", + "[21] Zhengqi Li and Noah Snavely. Learning intrinsic image decomposition from watching the world. In CVPR, 2018. 2", + "[22] Zhengqin Li, Kalyan Sunkavalli, and Manmohan Chandraker. Materials for masses: SVBRDF acquisition with a single mobile phone image. In ECCV, 2018. 2", + "[23] Zhengqin Li, Mohammad Shafiei, Ravi Ramamoorthi, Kalyan Sunkavalli, and Manmohan Chandraker. Inverse rendering for complex indoor scenes: Shape, spatially-varying lighting and SVBRDF from a single image. In CVPR, 2020. 2", + "[24] Zhengqin Li, Ting-Wei Yu, Shen Sang, Sarah Wang, Meng Song, Yuhan Liu, Yu-Ying Yeh, Rui Zhu, Nitesh Gundavarapu, Jia Shi, et al. OpenRooms: An open framework for photorealistic indoor scene datasets. In CVPR, 2021. 2", + "[25] Zhengqin Li, Jia Shi, Sai Bi, Rui Zhu, Kalyan Sunkavalli, Miloš Hasan, Zexiang Xu, Ravi Ramamoorthi, and Manmohan Chandraker. Physically-based editing of indoor scene lighting from a single image. In ECCV, 2022. 2, 6, 8", + "[26] Zhi-Hao Lin, Bohan Liu, Yi-Ting Chen, David Forsyth, Jia-Bin Huang, Anand Bhattachad, and Shenlong Wang. UrbanIR: Large-scale urban scene inverse rendering from a single video. 3DV, 2025. 2", + "[27] Yu-Lun Liu, Wei-Sheng Lai, Yu-Sheng Chen, Yi-Lung Kao, Ming-Hsuan Yang, Yung-Yu Chuang, and Jia-Bin Huang. Single-image HDR reconstruction by learning to reverse the camera pipeline. In CVPR, 2020. 3", + "[28] Demetris Marnerides, Thomas Bashford-Rogers, Jonathan Hatchett, and Kurt Debattista. ExpandNet: A deep convolutional neural network for high dynamic range expansion from low dynamic range content. Computer Graphics Forum, 2018. 3", + "[29] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 6", + "[30] Ben Mildenhall, Peter Hedman, Ricardo Martin-Brualla, Pratul Srinivasan, and Jonathan T. Barron. NeRF in the dark: High dynamic range view synthesis from noisy raw images. In CVPR, 2022. 3", + "[31] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM TOG, 2022. 3", + "[32] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting triangular 3D models, materials, and lighting from images. In CVPR, 2022. 2", + "[33] Merlin Nimier-David, Zhao Dong, Wenzel Jakob, and Anton Kaplanyan. Material and lighting reconstruction for complex" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "473", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "indoor scenes with texture-space differentiable rendering. In Eurographics Symposium on Rendering, 2021. 2, 4", + "[34] Julien Philip, Sébastien Morgenthaler, Michael Gharbi, and George Drettakis. Free-viewpoint indoor neural relighting from multi-view stereo. ACM Trans. Graph., 40(5):194:1-18, 2021. 2", + "[35] Erik Reinhard, Greg Ward, Sumanta Pattanaik, Paul Debevec, Wolfgang Heidrich, and Karol Myszkowski. High Dynamic Range Imaging - Acquisition, Display and Image-Based Lighting. Morgan Kaufmann, 2nd edition, 2010. 1", + "[36] Mike Roberts, Jason Ramapuram, Anurag Ranjan, Atulit Kumar, Miguel Angel Bautista, Nathan Paczan, Russ Webb, and Joshua M. Susskind. Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding. In ICCV, 2021. 2", + "[37] Dario Seyb, Peter-Pike Sloan, Ari Silvennoinen, Michal Iwanicki, and Wojciech Jarosz. The design and evolution of the UberBake light baking system. ACM Transactions on Graphics (TOG), 2020. 4", + "[38] Pratul P Srinivasan, Ben Mildenhall, Matthew Tancik, Jonathan T Barron, Richard Tucker, and Noah Snavely. Lighthouse: Predicting lighting volumes for spatially-coherent illumination. In CVPR, 2020. 2", + "[39] Pratul P. Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T. Barron. NeRV: Neural reflectance and visibility fields for relighting and view synthesis. In CVPR, 2021. 3", + "[40] Cheng Sun, Guangyan Cai, Zhengqin Li, Kai Yan, Cheng Zhang, Carl Marshall, Jia-Bin Huang, Shuang Zhao, and Zhao Dong. Neural-pbir reconstruction of shape, material, and illumination. In ICCV, 2023. 2", + "[41] Zian Wang, Jonah Philion, Sanja Fidler, and Jan Kautz. Learning indoor inverse rendering with 3D spatially-varying lighting. In ICCV, 2021. 4", + "[42] Zian Wang, Tianchang Shen, Jun Gao, Shengyu Huang, Jacob Munkberg, Jon Hasselgren, Zan Gojcic, Wenzheng Chen, and Sanja Fidler. Neural fields meet explicit geometric representations for inverse rendering of urban scenes. In CVPR, 2023. 2, 3", + "[43] Liwen Wu, Rui Zhu, Mustafa B Yaldiz, Yinhao Zhu, Hong Cai, Janarbek Matai, Fatih Porikli, Tzu-Mao Li, Manmohan Chandraker, and Ravi Ramamoorthi. Factorized inverse path tracing for efficient and accurate material-lighting estimation. In ICCV, 2023. 2, 3, 4, 6, 7, 8", + "[44] Xin Yang, Ke Xu, Yibing Song, Qiang Zhang, Xiaopeng Wei, and Rynson WH Lau. Image correction via deep reciprocating HDR transformation. In CVPR, 2018. 3", + "[45] Yao Yao, Jingyang Zhang, Jingbo Liu, Yihang Qu, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. NeILF: Neural incident light field for physically-based material estimation. In ECCV, 2022. 2, 6, 7, 8", + "[46] Lior Yariv, Peter Hedman, Christian Reiser, Dor Verbin, Pratul P Srinivasan, Richard Szeliski, Jonathan T Barron, and Ben Mildenhall. Bakedsdf: Meshing neural sdfs for real-time view synthesis. In ACM SIGGRAPH computer graphics, 2023. 4" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[47] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. ScanNet++: A high-fidelity dataset of 3D indoor scenes. In ICCV, 2023. 6, 7", + "[48] Bohan Yu, Siqi Yang, Xuanning Cui, Siyan Dong, Baoquan Chen, and Boxin Shi. MILO: Multi-bounce inverse rendering for indoor scene with light-emitting objects. IEEE TPAMI, 2023. 2, 4", + "[49] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. MonoSDF: Exploring monocular geometric cues for neural implicit surface reconstruction. In NeurIPS, 2022. 4, 5", + "[50] Zheng Zeng, Valentin Deschaintre, Iliyan Georgiev, Yannick Hold-Geoffroy, Yiwei Hu, Fujun Luan, Ling-Qi Yan, and Miloš Hasan. RGB ↔ X: Image decomposition and synthesis using material- and lighting-aware diffusion models. In SIGGRAPH, 2024. 2", + "[51] Jinsong Zhang and Jean-François Lalonde. Learning high dynamic range from outdoor panoramicas. In ICCV, 2017. 3", + "[52] Jingyang Zhang, Yao Yao, Shiwei Li, Jingbo Liu, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. NeILF++: Inter-reflectable light fields for geometry and material estimation. In ICCV, 2023. 2, 3", + "[53] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snavely. PhySG: Inverse rendering with spherical Gaussians for physics-based material editing and relighting. In CVPR, 2021. 2, 3", + "[54] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. NeR-Factor: Neural factorization of shape and reflectance under an unknown illumination. ACM TOG, 2021. 2", + "[55] Jingsen Zhu, Yuchi Huo, Qi Ye, Fujun Luan, Jifan Li, Dianbing Xi, Lisha Wang, Rui Tang, Wei Hua, Hujun Bao, and Rui Wang. I²-SDF: Intrinsic indoor scene reconstruction and editing via raytracing in neural SDFs. In CVPR, 2023. 2, 4", + "[56] Rui Zhu, Zhengqin Li, Janarbek Matai, Fatih Porikli, and Manmohan Chandraker. IRISformer: Dense vision transformers for single-image inverse rendering in indoor scenes. In CVPR, 2022. 2, 4" + ], + "bbox": [ + 516, + 90, + 905, + 628 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "474", + "bbox": [ + 485, + 944, + 511, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/IRIS_ Inverse Rendering of Indoor Scenes from Low Dynamic Range Images/53c93036-c714-4f00-9960-2f72fce37288_model.json b/2025/IRIS_ Inverse Rendering of Indoor Scenes from Low Dynamic Range Images/53c93036-c714-4f00-9960-2f72fce37288_model.json new file mode 100644 index 0000000000000000000000000000000000000000..000aed878eb82b5c4d861709695cbb525624b3b3 --- /dev/null +++ b/2025/IRIS_ Inverse Rendering of Indoor Scenes from Low Dynamic Range Images/53c93036-c714-4f00-9960-2f72fce37288_model.json @@ -0,0 +1,2409 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.113, + 0.131, + 0.885, + 0.154 + ], + "angle": 0, + "content": "IRIS: Inverse Rendering of Indoor Scenes from Low Dynamic Range Images" + }, + { + "type": "text", + "bbox": [ + 0.14, + 0.18, + 0.273, + 0.198 + ], + "angle": 0, + "content": "Chih-Hao Lin1,2" + }, + { + "type": "text", + "bbox": [ + 0.294, + 0.181, + 0.429, + 0.199 + ], + "angle": 0, + "content": "Jia-Bin Huang\\(^{1,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.45, + 0.181, + 0.557, + 0.199 + ], + "angle": 0, + "content": "Zhengqin Li" + }, + { + "type": "text", + "bbox": [ + 0.579, + 0.181, + 0.678, + 0.199 + ], + "angle": 0, + "content": "Zhao Dong1" + }, + { + "type": "text", + "bbox": [ + 0.701, + 0.181, + 0.854, + 0.198 + ], + "angle": 0, + "content": "Christian Richardt1" + }, + { + "type": "text", + "bbox": [ + 0.157, + 0.2, + 0.246, + 0.216 + ], + "angle": 0, + "content": "Tuotuo Li" + }, + { + "type": "text", + "bbox": [ + 0.268, + 0.2, + 0.42, + 0.216 + ], + "angle": 0, + "content": "Michael Zollhöfer1" + }, + { + "type": "text", + "bbox": [ + 0.443, + 0.2, + 0.568, + 0.217 + ], + "angle": 0, + "content": "Johannes Kopf1" + }, + { + "type": "text", + "bbox": [ + 0.59, + 0.2, + 0.725, + 0.217 + ], + "angle": 0, + "content": "Shenlong Wang" + }, + { + "type": "text", + "bbox": [ + 0.748, + 0.2, + 0.858, + 0.217 + ], + "angle": 0, + "content": "Changil Kim" + }, + { + "type": "text", + "bbox": [ + 0.135, + 0.223, + 0.863, + 0.269 + ], + "angle": 0, + "content": "\\(^{1}\\)Meta \\(^{2}\\)University of Illinois Urbana-Champaign \\(^{3}\\)University of Maryland College Park https://irisldr.github.io/" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.287, + 0.236, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.247, + 0.288, + 0.477, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.489, + 0.288, + 0.692, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.704, + 0.288, + 0.905, + 0.524 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.532, + 0.907, + 0.575 + ], + "angle": 0, + "content": "Figure 1. Physically-based inverse rendering from LDR images. Our method takes LDR input images and estimates high-quality spatially-varying HDR lighting, physically-based material properties, and a camera response function (CRF). This decomposition allows us to insert objects or new light sources, and to relight the scene in a physically accurate manner." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.578, + 0.645, + 0.593 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.608, + 0.327, + 0.623 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.642, + 0.486, + 0.884 + ], + "angle": 0, + "content": "Inverse rendering seeks to recover 3D geometry, surface material, and lighting from captured images, enabling advanced applications such as novel-view synthesis, relighting, and virtual object insertion. However, most existing techniques rely on high dynamic range (HDR) images as input, limiting accessibility for general users. In response, we introduce IRIS, an inverse rendering framework that recovers the physically based material, spatially-varying HDR lighting, and camera response functions from multi-view, low-dynamic-range (LDR) images. By eliminating the dependence on HDR input, we make inverse rendering technology more accessible. We evaluate our approach on real-world and synthetic scenes and compare it with state-of-the-art methods. Our results show that IRIS effectively recovers HDR lighting, accurate material, and plausible camera response functions, supporting photorealistic relighting and object insertion." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.609, + 0.907, + 0.761 + ], + "angle": 0, + "content": "Physically based inverse rendering enables reconstructing realistic material properties and lighting in scenes. This decomposition is valuable for various applications, including relighting, material editing, and realistic object insertion. However, most existing inverse rendering methods require input images with high dynamic range (HDR) to capture the full light transport in the scene. This poses a significant barrier to broader adoption, as HDR capture typically requires specialized hardware or merging multiple aligned LDR images through exposure bracketing [35]." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Most common imaging sensors do not capture sufficient dynamic range for scenes' light-emitting and dark regions. Moreover, cameras often convert raw sensor readings to 8-bit low-dynamic-range (LDR) images for storage and transmission. The non-linear mapping and quantization lead to the additional loss of lighting information. The complex light transport in indoor environments makes reconstructing the original HDR lighting - critical for faithful inverse rendering - even more challenging. If material and lighting estimation" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "465" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.137 + ], + "angle": 0, + "content": "could be achieved from a \"casual\" capture using standard devices like cameras or phones, inverse rendering would become far more accessible to a wider range of users." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.139, + 0.484, + 0.305 + ], + "angle": 0, + "content": "Many state-of-the-art inverse rendering methods rely on HDR inputs to accurately recover material properties and lighting [43, 48, 52]. Some approaches attempt to overcome this issue by taking LDR images as inputs and solving for the lighting. However, these methods often make simplifying assumptions, such as infinitely distant light sources [42], requiring additional inputs like emitter masks [55], or neglecting multi-bounce light transport [45, 52, 55]. As a result, these methods struggle to handle the complex light transport in indoor scenes and often fail to produce high-quality surface material and lighting." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.307, + 0.484, + 0.488 + ], + "angle": 0, + "content": "To address these challenges, we introduce IRIS, an inverse rendering method for indoor scenes, using only multi-view LDR images. IRIS models tone mapping (i.e., HDR-to-LDR conversion), allowing us to work directly with LDR inputs. IRIS automatically identifies emitters and reconstructs spatially varying HDR lighting through physically based inverse rendering, which is crucial for faithful material estimation. Jointly estimating lighting, surface material, and camera response function (CRF) leads to unstable optimization due to ambiguities. To tackle this, we design a novel optimization strategy that effectively resolves these ambiguities, enabling high-quality estimation. Our core contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.489, + 0.484, + 0.549 + ], + "angle": 0, + "content": "1. IRIS faithfully estimates spatially-varying HDR lighting, physically-based materials, and camera response function from LDR images and outperforms several state-of-the-art inverse rendering methods." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.55, + 0.483, + 0.61 + ], + "angle": 0, + "content": "2. We explicitly model LDR image formation in our (inverse) rendering pipeline so that LDR images can be used directly, broadening the accessibility of high-quality inverse rendering." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.61, + 0.483, + 0.656 + ], + "angle": 0, + "content": "3. Finally, IRIS is extensively evaluated on synthetic and real-world scenes, demonstrating diverse and realistic view synthesis, relighting, and object insertion." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.489, + 0.484, + 0.656 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.674, + 0.233, + 0.69 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Data-driven inverse rendering. Inverse rendering seeks to reconstruct a scene's geometry, material, and lighting from single or multiple images. This process is notoriously challenging and ill-posed due to the inherent ambiguity. Numerous methods employ deep priors learned from large-scale datasets for different tasks, including intrinsic image decomposition [20, 21], SVBRDF estimation [6, 19, 22, 50, 56], lighting estimation [8, 23, 38], lighting editing [25] and relighting [34]. These learning-based methods take a single or a few images [2, 38] as inputs, reducing capturing requirements of classical measurement-based methods [9]. However, high-quality datasets are essential for data-driven methods to achieve compelling generalization ability. Synthetic" + }, + { + "type": "image", + "bbox": [ + 0.554, + 0.089, + 0.714, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.604, + 0.163, + 0.681, + 0.17 + ], + "angle": 0, + "content": "Surface material" + }, + { + "type": "image", + "bbox": [ + 0.723, + 0.089, + 0.865, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.773, + 0.163, + 0.814, + 0.171 + ], + "angle": 0, + "content": "Lighting" + }, + { + "type": "image", + "bbox": [ + 0.554, + 0.17, + 0.714, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.604, + 0.245, + 0.681, + 0.254 + ], + "angle": 0, + "content": "Surface material" + }, + { + "type": "image", + "bbox": [ + 0.723, + 0.17, + 0.864, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.773, + 0.245, + 0.814, + 0.255 + ], + "angle": 0, + "content": "Lighting" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.255, + 0.908, + 0.337 + ], + "angle": 0, + "content": "Figure 2. Limitation of SOTA. A typical image formation process causes the loss of lighting information, posing challenges in inverse rendering. FIPT [43] assumes HDR input and NeILF [45] ignores multi-bounce light transport. Both methods fail to estimate accurate material (red boxes) and HDR lighting (yellow boxes). We demonstrate significantly better results (Figure 1)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.35, + 0.909, + 0.532 + ], + "angle": 0, + "content": "datasets are often used for training because ground-truth labeling is challenging to obtain in real environments. Despite efforts to create photorealistic synthetic datasets [20, 24, 36], the inherent diversity of real-world environments, especially in complex indoor scenes, still causes a substantial domain gap. Furthermore, these methods do not reconstruct complete 3D scenes and fail to render consistent novel views under novel illuminations. Our method utilizes prediction from a data-driven method [56] for initialization and reconstructs 3D geometry, material and lighting with a carefully designed optimization framework, demonstrating superior performance for real-world scenes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Optimization-based inverse rendering. Differentiable path tracing [14] enables the direct optimization of scene parameters with physically-based multi-bounce light transport. Recent methods, including MILO [48], IPT [1], and follow-up work [33] exploit this to jointly optimize BRDF and lighting emission, accounting for global light transport. FIPT [43] utilizes a factored light transport formulation for BRDF estimation and emitter detection. Nevertheless, these methods assume piecewise constant materials or are designed to take high-dynamic range images as input, and the robustness is limited in real-world scenes. Several recent NeRF-based approaches [12, 15, 26, 32, 40, 42, 53, 54] parameterize materials as a neural field and optimize with a volume rendering formulation. They tend to assume distant lighting and cannot handle spatially varying lighting in indoor environments well. NeILF [45] and NeILF++ [52] represent spatially-varying lighting as a neural field, \\(\\mathrm{I}^2\\)-SDF [55] constructs lighting from given emitter masks. While handling more complex lighting, these works require additional assumptions (e.g., HDR input, emitter masks), and single-bounce ray tracing cannot model complex light transport. In contrast, our method automatically identifies emitters and reconstructs HDR lighting from casual (LDR) images, while achieving high-quality material and lighting estimation that" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "466" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.416, + 0.106 + ], + "angle": 0, + "content": "outperforms prior work in this practical scenario." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.11, + 0.486, + 0.414 + ], + "angle": 0, + "content": "CRF and HDR estimation. The complete information about the illumination is usually lost within the photography pipeline. Estimating the HDR lighting and CRF enables us to edit photos and achieve realistic object insertion. Debevec and Malik [5] propose to take multiple photos with varying exposure levels to recover the CRF and HDR image. Some previous works [27, 28, 44, 51] predict HDR images with a data-driven approach by learning the LDR-HDR mapping from a single image. However, these methods rely on HDR data for training, and the estimation is inconsistent across views and scenes. To handle the HDR-LDR mapping, FIPT [43] and RawNeRF [30] assume a known gamma correction function; NeILF++ [52] learn a single gamma correction parameter; HDR-NeRF [13] parameterizes tone-mapping function with an MLP. However, these works are designed to take HDR images or LDR with multiple exposure levels as input. We leverage physically-based rendering to recover spatially varying BRDF, HDR illumination, and CRF from LDR images. Our method generates realistic view synthesis, relighting, and object insertion." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.427, + 0.415, + 0.444 + ], + "angle": 0, + "content": "3. Background: LDR image formation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.452, + 0.485, + 0.649 + ], + "angle": 0, + "content": "Real-world radiance exhibits high dynamic range (HDR) and contains bright light sources and dark regions. To preserve the complete lighting information, HDR images are typically produced with a high-end camera, laborious exposure bracketing [5], and are stored in specialized format (16 or 32 bits per color channel). However, capturing HDR is often impractical for casual users and is typically absent in standard photography assets. Due to the limitations of standard sensors and displays, the radiance is frequently processed with multiple stages to produce low dynamic range (LDR) images stored in more lightweight formats (such as JPEG or PNG at 8 bits and RAW at 12 bits per color channel). We describe the major steps in this process below." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.652, + 0.485, + 0.744 + ], + "angle": 0, + "content": "Dynamic range clipping. Given the scene irradiance \\( E \\) and exposure time \\( \\Delta t \\), the observed radiance is clipped at a certain maximum value due to the sensor limitation, which could be formulated as: \\( Z_{c} = \\min (E\\Delta t,1) \\). Due to the clipping operation, the lighting information is lost for overexposed/saturated pixels." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.746, + 0.486, + 0.852 + ], + "angle": 0, + "content": "Non-linear mapping of CRF. To enhance the image's visual quality and match the human perception of the scene, the clipped intensity is further transformed with a non-linear camera response function (CRF) to produce LDR pixel values: \\( Z = \\mathrm{CRF}(Z_c) = \\mathrm{CRF}(\\min(E\\Delta t, 1)) \\). Please refer to Debevec and Malik [5] for a more detailed imaging pipeline explanation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Challenges in inverse rendering. Real-world scenes often exhibit extreme variances in radiance, making it challenging to capture these scenes in LDR images without under" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.909, + 0.169 + ], + "angle": 0, + "content": "exposing or saturating them. Due to the clipping and nonlinear CRF mapping, critical lighting information is lost, posing challenges in previous works, illustrated in Figure 2. In this work, we propose an inverse rendering framework that jointly recovers HDR lighting and CRF from LDR images." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.18, + 0.791, + 0.198 + ], + "angle": 0, + "content": "4. Representation and rendering" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.208, + 0.909, + 0.359 + ], + "angle": 0, + "content": "Camera response function (CRF). The CRF is crucial for recovering original radiance from 8-bit LDR images. However, the CRF varies with each camera and is often unknown to users. Thus, deducing the CRF from LDR image observations is critical to our solution. We use Gross et al.'s [11] empirical model of response (EMoR) for CRF modeling [27]. They collected a database of 201 real-world CRFs of various devices and parameterized them as vectors. A mean curve \\(\\bar{\\mathbf{g}}\\) and PCA basis \\(\\mathbf{g}_b\\) are calculated from the database, and we parameterize a learnable CRF as" + }, + { + "type": "equation", + "bbox": [ + 0.61, + 0.366, + 0.907, + 0.385 + ], + "angle": 0, + "content": "\\[\n\\mathbf {g} = \\bar {\\mathbf {g}} + \\Sigma_ {b} w _ {b} \\mathbf {g} _ {b}, \\mathbf {g} \\in \\mathbb {R} ^ {1 0 2 4}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.392, + 0.909, + 0.514 + ], + "angle": 0, + "content": "where \\(\\{w_{b}\\}\\) weigh each basis curve and are learnable parameters. The non-linear function is discretized as \\(\\mathrm{CRF}(x) = \\mathrm{LERP}(\\{g_k\\}_k, x)\\), which linearly interpolates the uniformly sampled CRF \\(\\{g_k\\}_{k=0}^{1023} \\in [0,1]^{1024}\\), independently per color channel. We make this design choice because it covers most of the CRF space, and the simplicity makes it easy to enforce regularization, which brings huge advantage in the complex optimization of inverse rendering." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.516, + 0.909, + 0.759 + ], + "angle": 0, + "content": "Spatially-varying lighting. While environment maps handle object-centric and outdoor scenes well [15, 39, 42, 53], they cannot capture complex spatially-varying lighting of indoor scenes. To handle this, we define direct lighting on the scene mesh and measure emission as view-independent radiance: \\(\\mathbf{L}_{\\mathrm{e}}(\\mathbf{x}) \\in \\mathbb{R}^3\\). The parameterization of \\(\\mathbf{L}_{\\mathrm{e}}\\) is detailed in our supplementary material. We also define an emitter mask \\(M_{\\mathrm{e}}(\\mathbf{x}) \\in \\{0,1\\}\\) to denote whether a surface point \\(\\mathbf{x}\\) is on an emitter. Global illumination is represented as a view-independent surface light field (SLF): \\(\\mathbf{L}_{\\mathrm{SLF}}(\\mathbf{x}) \\in \\mathbb{R}^3\\). The sRGB pixel colors of input LDR images are linearized with an inverse CRF mapping: \\(\\mathbf{I}_{\\mathrm{linear}} = \\mathrm{CRF}^{-1}(\\mathbf{I}_{\\mathrm{LDR}})\\). If a voxel contains no surface, it is set to zero; otherwise, it collects the average linearized radiances across all input images with ray casting. This greatly accelerates the approximation of global illumination without requiring multi-bounce ray tracing." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.761, + 0.909, + 0.868 + ], + "angle": 0, + "content": "Material BRDF. We represent spatially-varying materials using the Cook-Torrance BRDF [4]. The surface albedo \\(\\mathbf{a}(\\mathbf{x})\\in \\mathbb{R}^3\\), roughness \\(\\sigma (\\mathbf{x})\\in \\mathbb{R}\\), and metallicity \\(m(\\mathbf{x})\\in \\mathbb{R}\\) model the diffuse reflectance \\(\\mathbf{k_d} = \\mathbf{a}\\cdot (1 - m)\\) and specular reflectance \\(\\mathbf{k_s} = 0.04\\cdot (1 - m)\\), and are parameterized as a neural field \\(\\mathbf{f}\\colon \\mathbf{x}\\mapsto (\\mathbf{a},\\sigma ,m)\\), implemented with multi-resolution hash encoding and an MLP [31]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.902 + ], + "angle": 0, + "content": "Factorized light transport. To render images from scene geometry, material, and lighting information, We follow" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "467" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.092, + 0.904, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.311, + 0.908, + 0.381 + ], + "angle": 0, + "content": "Figure 3. Framework Overview. Given multi-view posed LDR images and a surface mesh, our inverse rendering pipeline is divided into two main stages. In the initialization stage, we initialize the BRDF (Section 5.1), extract a surface light field (Section 4), and estimate emitter geometry (Equation (3)). In the optimization stage, we first recover HDR radiance from the LDR input (Section 5.2), then bake shading maps (Section 5.3), and jointly optimize BRDF and CRF parameters (Section 5.4). The improved parameters are used to refine the emission again. These three steps are repeated until convergence." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.393, + 0.483, + 0.424 + ], + "angle": 0, + "content": "the rendering equation [16] to model physically-based light transport for realistic rendering:" + }, + { + "type": "equation", + "bbox": [ + 0.098, + 0.432, + 0.484, + 0.463 + ], + "angle": 0, + "content": "\\[\n\\mathbf {L} _ {\\mathrm {o}} (\\mathbf {x}, \\omega_ {\\mathrm {o}}) = \\mathbf {L} _ {\\mathrm {e}} (\\mathbf {x}, \\omega_ {\\mathrm {o}}) + \\int_ {\\Omega^ {+}} \\mathbf {L} _ {\\mathrm {i}} (\\mathbf {x}, \\omega_ {\\mathrm {i}}) f (\\mathbf {x}, \\omega_ {\\mathrm {i}}, \\omega_ {\\mathrm {o}}) d \\omega_ {\\mathrm {i}}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.47, + 0.484, + 0.668 + ], + "angle": 0, + "content": "where \\(\\mathbf{L}_{\\mathrm{o}}\\) is the radiance observed along a ray at a 3D position \\(\\mathbf{x}\\) in direction \\(\\omega_{\\mathrm{o}}\\), \\(\\mathbf{L}_{\\mathrm{e}}\\) is the emission term, \\(\\mathbf{L}_{\\mathrm{r}} = \\int_{\\Omega^{+}}\\mathbf{L}_{\\mathrm{i}}(\\mathbf{x},\\omega_{\\mathrm{i}})f(\\mathbf{x},\\omega_{\\mathrm{i}},\\omega_{\\mathrm{o}})d\\omega_{\\mathrm{i}}\\) is the reflectance term, and \\(f(\\mathbf{x},\\omega_{\\mathrm{i}},\\omega_{\\mathrm{o}})\\) is the BRDF. The recursive nature of the rendering equation is approximated with Monte-Carlo path tracing [16, 18]. However, modeling multiple bounces is still computationally expensive and unstable in the inverse rendering process. To enhance the efficiency and robustness of the optimization, we follow FIPT [43] and adopt factorized light transport [17, 37, 41], which precomputes shading maps \\(\\mathbf{L}_{\\mathrm{d}}\\), \\(\\mathbf{L}_{\\mathrm{s}}^{0}\\), \\(\\mathbf{L}_{\\mathrm{s}}^{1}\\) and linearly interpolates with roughness to obtain final rendering. Detailed notation and formulation are available in the supplemental material." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.68, + 0.427, + 0.697 + ], + "angle": 0, + "content": "5. Inverse Rendering from LDR Images" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.483, + 0.81 + ], + "angle": 0, + "content": "Given multi-view posed LDR images \\(\\{\\mathbf{I}_i\\}\\) captured with the same or varying exposure levels \\(\\{\\Delta t_i\\}\\), our goal is to estimate spatially-varying HDR lighting, surface material (as albedo, roughness, and metallic), and the camera response function (CRF) shared by all input images. We assume that the primary light sources are observed in the input images, following the recent literature [43, 48, 55]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Previous approaches perform differentiable path tracing during the material and lighting optimization [1, 33], but the number of samples is limited due to heavy computation. This introduces high variance and unstable estimation. The factorized formulation [17, 37, 41, 43] splits light transport into BRDF parameters and shading, and enables updating" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.393, + 0.907, + 0.453 + ], + "angle": 0, + "content": "them alternately. While the method is shown to be more stable and produce high-quality decompositions, it requires HDR images as input to aggregate and bake shading maps, detect emitters, and estimate HDR emissions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.456, + 0.909, + 0.608 + ], + "angle": 0, + "content": "To address this ill-posed optimization problem, we propose a multi-stage optimization strategy alternatingly estimating spatially varying lighting, material, and CRF. Specifically, we (1) initialize the BRDF with an off-the-shelf monocular inverse rendering method (Section 5.1), (2) recover HDR lighting from LDR with physically-based rendering (Section 5.2), (3) bake the diffuse and specular shading (Section 5.3), and (4) jointly estimate BRDF and CRF (Section 5.4). We repeat Steps 2-4 in that order until all estimates are converged. Figure 3 illustrates our framework." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.615, + 0.909, + 0.676 + ], + "angle": 0, + "content": "Geometry reconstruction. From input images with camera poses, we reconstruct the geometry with BakedSDF [46], and use normal estimation from off-the-shelf method [7] for regulating surface geometry, following MonoSDF [49]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.694, + 0.704, + 0.71 + ], + "angle": 0, + "content": "5.1. BRDF Initialization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.901 + ], + "angle": 0, + "content": "For each input image \\(\\mathbf{I}_i\\), we predict a 2D albedo map \\(\\mathbf{A}_i\\) using an off-the-shelf single-image albedo estimation method [56] and averaging per-pixel albedo values within each semantic part. Given \\(\\{\\mathbf{A}_i\\}\\), we initialize albedo \\(\\mathbf{a}(\\mathbf{x})\\) by minimizing the error between the projected albedo and the estimated albedo across all input images: \\(\\min_{\\mathbf{a}} \\sum_i \\| \\pi_i(\\mathbf{a}) - \\mathbf{A}_i \\|_2\\). Here, \\(\\pi_i\\) represents the perspective projection onto the image space of \\(\\mathbf{I}_i\\) by intersecting camera rays with the surface mesh of the input scene and reading \\(\\mathbf{a}\\) at their first intersections. We obtain semantic category labels from part IDs in synthetic scenes or using Mask2Former [3] for real scenes. Roughness and metallicity are both initialized to 0.5." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "468" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.131, + 0.089, + 0.437, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.218, + 0.485, + 0.273 + ], + "angle": 0, + "content": "Figure 4. Emitter geometry estimation \\( M_{\\mathrm{e}}(\\mathbf{x}) \\). The point \\( \\mathbf{x}_1 \\) on the window is saturated across all input views and thus identified as an emitter. The point \\( \\mathbf{x}_2 \\) on the table is reflective and saturated in some views (e.g., view 2) but not in others and is NOT an emitter." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.28, + 0.336, + 0.294 + ], + "angle": 0, + "content": "5.2. HDR Emission Restoration" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.302, + 0.484, + 0.348 + ], + "angle": 0, + "content": "Given that emitters appear bright across views and produce saturated pixels in LDR images, we estimate the binary emitter geometry \\( M_{\\mathrm{e}}(\\mathbf{x}) \\) on the surface mesh as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.112, + 0.356, + 0.484, + 0.396 + ], + "angle": 0, + "content": "\\[\nM _ {\\mathrm {e}} (\\mathbf {x}) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} \\frac {1}{N} \\sum_ {i} v _ {i} (\\mathbf {x}) \\mathbf {I} _ {i} \\left(\\pi_ {i} (\\mathbf {x})\\right) \\geq 0. 9 9 \\\\ 0 & \\text {o t h e r w i s e ,} \\end{array} \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.403, + 0.443, + 0.419 + ], + "angle": 0, + "content": "where \\(v(\\cdot)\\in \\{0,1\\}\\) indicates visibility. See Figure 4." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.419, + 0.484, + 0.526 + ], + "angle": 0, + "content": "The HDR intensity is crucial for inverse rendering but is missing in the LDR images. To recover the HDR lighting, we perform physically-based rendering with single-bounce path tracing to generate the images (2). To estimate the radiance at the point \\(\\mathbf{x}\\) observed from a camera ray, we trace \\(N = 128\\) secondary rays from \\(\\mathbf{x}\\), sampled with the current BRDF estimates. The resulting radiance is approximated by:" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.533, + 0.483, + 0.574 + ], + "angle": 0, + "content": "\\[\n\\mathbf {L} _ {\\mathrm {r}} \\left(\\mathbf {x}, \\omega_ {\\mathrm {o}}\\right) = \\sum_ {i = 1} ^ {N} \\mathbf {L} _ {\\text {e n d}} \\left(\\mathbf {x} ^ {\\prime}\\right) f \\left(\\mathbf {x}, \\omega_ {\\mathrm {i}}, \\omega_ {\\mathrm {o}}\\right), \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.579, + 0.483, + 0.617 + ], + "angle": 0, + "content": "\\[\n\\mathbf {L} _ {\\text {e n d}} (\\mathbf {x}) = \\left\\{ \\begin{array}{l l} \\mathbf {L} _ {\\mathrm {e}} (\\mathbf {x}) & \\text {i f} M _ {\\mathrm {e}} (\\mathbf {x}) = 1 \\\\ \\mathbf {L} _ {\\text {S L F}} (\\mathbf {x}) & \\text {o t h e r w i s e ,} \\end{array} \\right. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.625, + 0.484, + 0.716 + ], + "angle": 0, + "content": "where \\(\\mathbf{x}' = \\mathbf{x} + d\\cdot \\boldsymbol{\\omega}_{\\mathrm{i}}\\) is the intersected point of the secondary ray along direction \\(\\boldsymbol{\\omega}_{\\mathrm{i}}\\). Intuitively, we retrieve the radiance from the learnable emission radiance \\(\\mathbf{L}_{\\mathrm{e}}(\\mathbf{x})\\) if the secondary ray hits an emitter directly, or otherwise from the pre-computed surface light field \\(\\mathbf{L}_{\\mathrm{SLF}}(\\mathbf{x})\\) to approximate the global illumination. Figure 5 shows an illustration." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.716, + 0.483, + 0.748 + ], + "angle": 0, + "content": "To recover the HDR emitter radiance, we minimize the photometric loss with gradient descent: \\(\\min_{\\mathbf{L}_{\\mathrm{e}}}\\mathcal{L}_{\\mathrm{photo}}\\) , where" + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.756, + 0.484, + 0.786 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {p h o t o}} = \\sum_ {i} \\| \\mathrm {C R F} \\left(\\min \\left(\\mathbf {L} _ {\\mathrm {o}} \\left(\\mathbf {x}, \\omega_ {\\mathrm {o}}\\right) \\Delta t _ {i}, 1\\right)\\right) - \\mathbf {I} _ {i} \\| _ {2}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.486, + 0.902 + ], + "angle": 0, + "content": "The emitter radiance \\(\\mathbf{L}_{\\mathrm{e}}(\\mathbf{x})\\) is enhanced to match the captured images, while the BRDF and CRF are fixed in this stage. Intuitively, in order to minimize the photometric loss with physically-based rendering, the emitter intensity \\(\\mathbf{L}_{\\mathrm{e}}(\\mathbf{x})\\) should be increased significantly and enhance the overall brightness of the rendered image. Therefore, the HDR lighting is recovered in this process." + }, + { + "type": "image", + "bbox": [ + 0.568, + 0.089, + 0.848, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.225, + 0.908, + 0.294 + ], + "angle": 0, + "content": "Figure 5. HDR emission restoration. Top: Ray sampling process. Learnable direct lighting \\(\\mathbf{L}_{\\mathrm{e}}\\) is retrieved if a ray hits an emitter (e.g., window) or \\(\\mathbf{L}_{\\mathrm{SLF}}\\) is retrieved otherwise (e.g., wall). Bottom: HDR restoration process. By performing differentiable physically-based rendering, the photometric loss enhances the emitter intensity \\(\\mathbf{L}_{\\mathrm{e}}\\)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.31, + 0.673, + 0.327 + ], + "angle": 0, + "content": "5.3. Shading Baking" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.333, + 0.908, + 0.469 + ], + "angle": 0, + "content": "After the HDR emission is recovered in the previous stage, we bake the diffuse and specular shading maps \\(\\mathbf{L}_{\\mathrm{d}}\\), \\(\\mathbf{L}_{\\mathrm{s}}^{0}\\), \\(\\mathbf{L}_{\\mathrm{s}}^{1}\\), following factorized light transport (details in supplementary material). Given a surface point intersected by a camera ray, multiple rays are sampled according to the current (fixed) BRDF. The incident radiance is computed with multibounce path tracing, where a ray keeps tracing if it hits a non-emissive specular surface and stops otherwise. The radiance of the endpoint follows Equation (5)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.48, + 0.767, + 0.496 + ], + "angle": 0, + "content": "5.4. BRDF & CRF Optimization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.502, + 0.907, + 0.548 + ], + "angle": 0, + "content": "With the baked shading maps \\(\\mathbf{L}_{\\mathrm{d}}\\), \\(\\mathbf{L}_{\\mathrm{s}}^{0}\\), and \\(\\mathbf{L}_{\\mathrm{s}}^{1}\\), the albedo, metallicity, roughness, and CRF are jointly optimized. The optimization involves the following objective function:" + }, + { + "type": "equation", + "bbox": [ + 0.553, + 0.561, + 0.907, + 0.584 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\mathbf {a}, m, \\sigma , \\mathbf {g}} \\mathcal {L} _ {\\text {p h o t o}} + \\lambda_ {\\mathrm {a}} \\mathcal {L} _ {\\text {a l b e d o}} + \\lambda_ {\\mathrm {c}} \\mathcal {L} _ {\\text {C R F}} + \\lambda_ {\\mathrm {m}} \\mathcal {L} _ {\\text {m a t}}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.596, + 0.828, + 0.612 + ], + "angle": 0, + "content": "where \\(\\lambda_{\\mathrm{a}} = 0.01\\) \\(\\lambda_{\\mathrm{c}} = 0.001\\) , and \\(\\lambda_{\\mathrm{m}} = 0.005\\)" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.615, + 0.907, + 0.647 + ], + "angle": 0, + "content": "Photometric loss. We perform physically-based rendering, and the photometric loss \\(\\mathcal{L}_{\\mathrm{photo}}\\) is computed as in Eq. 6." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.65, + 0.908, + 0.696 + ], + "angle": 0, + "content": "Albedo regularization. Motivated by the shift-scale-invariant loss in MonoSDF [49], we adopt a shift-invariant loss against the monocular albedo maps \\(\\{\\mathbf{A}_i\\}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.708, + 0.907, + 0.739 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {a l b e d o}} = \\sum_ {i} \\| \\pi_ {i} (\\mathbf {a}) - s _ {i} \\mathbf {A} _ {i} \\| _ {2}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.751, + 0.88, + 0.766 + ], + "angle": 0, + "content": "where \\(s_i\\) is the scale used to align the two albedo maps." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.769, + 0.908, + 0.86 + ], + "angle": 0, + "content": "CRF regularization. Based on the fair assumption that the learned CRF should not deviate significantly from the mean curve \\(\\bar{\\mathbf{g}}\\) (Eq. 1), and that the curve should be monotonically increasing, we regularize the \\(L_{2}\\) norm of the estimated coefficients of the PCA bases, and enforce the monotonicity of the sampled CRF as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.873, + 0.907, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {C R F}} = \\| \\mathbf {w} \\| _ {2} + \\sum_ {i} \\max \\left(g _ {i - 1} - g _ {i}, 0\\right). \\tag {9}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "469" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.213 + ], + "angle": 0, + "content": "Material regularization. We apply roughness-metallicity regularization \\(\\mathcal{L}_{\\mathrm{mat}}\\), which enforces the consistency of surface roughness and metallic within the same semantic instances (e.g. chairs, tables). While the shading and emission are initially inaccurate and lead to erroneous BRDFs, we alternately update the BRDFs and produce better emission and shading, improving the BRDFs. Our alternating strategy yields more stable optimization and better estimation." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.231, + 0.226, + 0.248 + ], + "angle": 0, + "content": "6. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.256, + 0.484, + 0.348 + ], + "angle": 0, + "content": "We evaluate our method on both synthetic and real-world data and compare it against several baseline methods. We analyze the quality of our inverse rendering, relighting, and novel-view synthesis capability. Please see our supplementary material for additional results and evaluations. We will release our code and data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.352, + 0.484, + 0.594 + ], + "angle": 0, + "content": "Datasets. We evaluate our method and baselines on the ScanNet++ [47] and FIPT [43] datasets. ScanNet++ contains 380 real-world indoor scenes and collects multi-view LDR images with a DSLR camera at a constant exposure level. We select four scenes that cover a variety of indoor scenes (e.g., office, bathroom). The FIPT dataset consists of four synthetic and two real-world scenes. The synthetic scenes provide ground-truth geometry and material properties for evaluation. We used the LDR images provided in the dataset, where the lighting is clipped in overly-exposed regions, compressed with a nonlinear CRF, quantized into 256 levels, and stored in 8-bit PNG format. We adopt the constant exposure setting to provide a fair comparison with baselines, as they cannot handle varying exposure. We further demonstrate that IRIS handles varying exposure levels of input LDR images and estimates CRF jointly." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.598, + 0.484, + 0.839 + ], + "angle": 0, + "content": "Baselines. We compare our method with the following baselines. Li et al. [25] estimate BRDF, geometry (depth & normal), and lighting condition of an indoor scene from a single image using a data-driven approach. In addition to the single image, the model also takes emitter masks as input, and we provide the mask estimated by our method (3). NeILF [45] targets inverse rendering from known geometry and multiview images. The spatially-varying BRDF and illumination are parameterized as neural fields [29]. Physically-based rendering and single-bounce ray tracing are performed during the training and rendering process. The HDR-to-LDR tone mapping is modeled as a learnable gamma correction. The original FIPT [43] is designed for HDR input, and fails to estimate the emission mask from LDR images. Thus, we additionally provide the emission mask estimated by our method (3). We denote this modified version as \"FIPT*\"." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": "To further evaluate our performance, we compare the original FIPT with HDR input if the HDR information is available. This privileged information will make illumination estimation much more accurate." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.907, + 0.133 + ], + "angle": 0, + "content": "Table 1. BRDF-emission comparison on synthetic data. Numbers are averaged across four synthetic scenes with GT [43]. The best metrics among LDR methods are highlighted in bold." + }, + { + "type": "table", + "bbox": [ + 0.552, + 0.142, + 0.866, + 0.248 + ], + "angle": 0, + "content": "
MethodInputkda'σLe
PSNR ↑IoU ↑L2 ↓
Li et al [25]LDR16.7314.0111.300.352.29
NeILF [45]LDR16.8514.0216.96——
FIPT*LDR15.499.744.990.690.28
OursLDR22.3317.9221.380.690.12
FIPT [43]HDR29.9525.9826.370.860.03
" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.26, + 0.807, + 0.277 + ], + "angle": 0, + "content": "6.1. Inverse Rendering of Real Scenes" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.285, + 0.909, + 0.558 + ], + "angle": 0, + "content": "We provide a qualitative evaluation of inverse rendering in Figure 6. Conducting single-bounce ray tracing in the neural light field, NeILF [45] fails to remove the shadow in diffuse reflectance \\(\\mathbf{k}_{\\mathrm{d}}\\), estimates lower roughness \\(\\sigma\\) on the walls than the mirror, and produces an incorrect emission mask. FIPT* struggles to recover surface material from LDR input, where accurate lighting information is missing. Please note that FIPT* takes emitter masks from Equation (3) as additional input, but does not recover HDR emission in the optimization procedure. On the other hand, IRIS recovers an almost shading-free diffuse field. Regarding roughness, our method effectively identifies specular surfaces with low roughness on the mirror, demonstrating better specular/rough material separation. Additionally, our proposed HDR emission estimation proves effective for real scenes, recovering HDR lighting sources (i.e., window and ceiling light). The restored HDR lighting makes the light transport more accurate and significantly enhances the performance of inverse rendering." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.574, + 0.898, + 0.591 + ], + "angle": 0, + "content": "6.2. View Synthesis and Relighting of Real Scenes" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We compare the relighting with FIPT* in Figure 7. FIPT* fails to estimate accurate surface material from LDR input and tends to recover low roughness for most surfaces, which unrealistically reflects the inserted light source and objects. To be more specific, the ground of the office (first scene) and the wall of the bathroom (second scene) look overly smooth and specular, which is not consistent with the observation and produces unrealistic reflections. In contrast, IRIS produces better relighting and object insertion quality thanks to better inverse rendering. The reflections on the specular surfaces (e.g., whiteboard, mirror) and the shading on the wall demonstrate the strength of IRIS for realistic relighting. Furthermore, the recovered HDR lighting can produce naturalistic specular reflection and shadows for inserted objects. Additionally, we insert moving light sources in the scene, shown in Figure 8, where the shadow on the ground and shading on the table change consistently. IRIS enables the photorealistic rendering of novel objects under diverse lighting and from arbitrary viewpoints, showcasing the significant potential for content creation applications." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.513, + 0.957 + ], + "angle": 0, + "content": "470" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.087, + 0.907, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.363, + 0.907, + 0.406 + ], + "angle": 0, + "content": "Figure 6. Decompositions of real scenes from ScanNet++ [47]. In addition to good reconstruction, IRIS produces a detailed diffuse map, identifies the mirror as a low-roughness region, and recovers HDR lighting from windows and ceiling light. Please note that the emission maps are normalized to visualize the relative intensity." + }, + { + "type": "image", + "bbox": [ + 0.091, + 0.409, + 0.907, + 0.744 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.749, + 0.907, + 0.806 + ], + "angle": 0, + "content": "Figure 7. Relighting and object insertion with real scenes from the ScanNet++ [47] and FIPT datasets [43]. When inserting new light sources (Relighting 1 & 2), IRIS not only changes shadows but also produces reflection of the light on the whiteboard (first scene) and mirror (second scene), enhancing realism substantially. The recovered HDR lighting also strengthens the reflection and shadow of inserted objects (\"Object Insertion\" column)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.814, + 0.313, + 0.83 + ], + "angle": 0, + "content": "6.3. Quantitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.486, + 0.901 + ], + "angle": 0, + "content": "Quantitative evaluation of inverse rendering is difficult for real-world scenes as ground truth is not available. As a result, we evaluate IRIS and baselines on the synthetic scenes from FIPT, where ground truth of surface material, illumina-" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.814, + 0.909, + 0.892 + ], + "angle": 0, + "content": "tion, and relighting are provided. Please note that we use the ground-truth mesh during optimization for all methods to minimize the influence of imperfect geometry and conduct a fair comparison. We compute the metrics of inverse rendering and provide the comparison in Table 1, which aggregates" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "471" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.104, + 0.09, + 0.186, + 0.1 + ], + "angle": 0, + "content": "Reconstruction" + }, + { + "type": "image", + "bbox": [ + 0.194, + 0.09, + 0.289, + 0.151 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.309, + 0.09, + 0.361, + 0.101 + ], + "angle": 0, + "content": "Relight 2" + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.102, + 0.383, + 0.151 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.404, + 0.09, + 0.455, + 0.101 + ], + "angle": 0, + "content": "Relight 3" + }, + { + "type": "image", + "bbox": [ + 0.385, + 0.102, + 0.477, + 0.151 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.155, + 0.471, + 0.17 + ], + "angle": 0, + "content": "Figure 8. Relighting with a moving light ball in a real scene." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.177, + 0.188, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.194, + 0.178, + 0.285, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.178, + 0.381, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.387, + 0.178, + 0.477, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.266, + 0.483, + 0.295 + ], + "angle": 0, + "content": "Figure 9. CRF estimation from real-world devices. The orange dashed line is the ground truth CRF, and the blue line our estimate." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.322, + 0.417, + 0.337 + ], + "angle": 0, + "content": "the results from the 4 synthetic FIPT scenes [43]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.338, + 0.484, + 0.443 + ], + "angle": 0, + "content": "We compare the roughness \\(\\sigma\\), diffuse reflectance \\(\\mathbf{k}_{\\mathrm{d}}\\) among the diffuse surfaces, and material reflectance: \\(\\mathbf{a}^{\\prime}\\). We report the PSNR for these BRDF estimates. For emission, we calculate the intersection over the union (IoU) of the emission masks and compare the \\(L_{2}\\) error of estimated emission maps. All the estimates are rendered in image space and compared with ground truth." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.444, + 0.484, + 0.701 + ], + "angle": 0, + "content": "Table 1 shows that our method surpasses all baseline methods that use LDR inputs regarding inverse rendering quality. Compared to the state-of-the-art single-view method [25], our method excels in all metrics across various scenes, indicating the superior decomposition of material properties from multi-view observations. Furthermore, even when the estimated emitter mask is provided for FIPT*, it fails to recover emission radiance, leading to compromised BRDF estimation accuracy. The \\(L_{2}\\) error of emission maps shows that our proposed HDR restoration (Section 5.2) successfully recovers the HDR lighting from LDR observations, which leads to better light transport modeling and more accurate BRDF estimation. We also compare with FIPT taking HDR images as input (bottom row), which serves as a reference, indicating complete lighting information is crucial for accurate inverse rendering. IRIS recovers HDR and achieves the smallest metric gaps among all methods taking LDR input." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.702, + 0.484, + 0.807 + ], + "angle": 0, + "content": "We compare novel-view synthesis and relighting of all competing algorithms and report results in Table 2. The novel-view synthesis results show that IRIS is comparable with baseline methods [45], while exhibiting superior material and lighting estimation. Our main focus is relighting, and IRIS outperforms all baselines with LDR input, demonstrating the effectiveness of inverse rendering." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.818, + 0.245, + 0.833 + ], + "angle": 0, + "content": "6.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We validate our design choices and provide an ablation study in Table 3, where we compute PSNR for material \\((\\mathbf{k}_{\\mathrm{d}},\\mathbf{a}^{\\prime},\\sigma)\\) and L2 for CRF curves and emission maps \\(\\mathbf{L}_{\\mathrm{e}}\\). The emitter mask estimation in Equation (3) identifies the emitter from" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.907, + 0.118 + ], + "angle": 0, + "content": "Table 2. Quantitative results of novel-view synthesis (NVS) and relighting on synthetic scenes with ground truth." + }, + { + "type": "table", + "bbox": [ + 0.552, + 0.129, + 0.866, + 0.259 + ], + "angle": 0, + "content": "
MethodInputPSNR ↑SSIM ↑LPIPS ↓
NVSNeILF [45]LDR30.0310.8990.180
FIPT*LDR15.1800.7440.437
OursLDR29.4680.9030.181
FIPT [43]HDR28.7600.9020.186
RelightingLi et al. [25]LDR22.4440.8130.398
FIPT*LDR11.6260.7050.359
OursLDR22.8610.8820.173
FIPT [43]HDR29.0120.9020.139
" + }, + { + "type": "table_caption", + "bbox": [ + 0.636, + 0.265, + 0.783, + 0.278 + ], + "angle": 0, + "content": "Table 3. Ablation study" + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.28, + 0.905, + 0.381 + ], + "angle": 0, + "content": "
Methodkda'σLeCRF
PSNR ↑IoU ↑L2 ↓L2 ↓
- Emitter mask Me (Equation (3))16.5011.304.810.000.488.91
- CRF modeling (Equation (1))18.1813.5917.520.690.12—
- HDR restoration (Section 5.2)18.2014.407.310.690.286.23
- Lalbedo (Equation (8))21.5917.3617.430.690.152.42
- LCRF (Equation (9))14.4912.8125.590.690.139.40
Full model (Ours)22.3317.9221.380.690.122.35
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.393, + 0.907, + 0.56 + ], + "angle": 0, + "content": "LDR observations and improves the emission map in the second row. CRF estimation is critical for modeling HDR-to-LDR conversion and helps inverse rendering in the third row. Nevertheless, the HDR restoration (Section 5.2) plays an important role in recovering accurate light transport from the captured images. Once the HDR light is restored by our proposed approach, the material (especially roughness \\(\\sigma\\)), CRF, and emission map are improved significantly. The results demonstrate the significance of each component and show that inverse rendering from casually captured LDR images benefits from our proposed framework." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.574, + 0.676, + 0.589 + ], + "angle": 0, + "content": "6.5. CRF Estimation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.907, + 0.705 + ], + "angle": 0, + "content": "To further validate the robustness of our method, we generate LDR images from FIPT's synthetic HDR scenes using different CRFs collected from real-world devices [10]. In this evaluation, we perform inverse rendering with IRIS and plot the CRF. The results in Figure 9 show that our method can generalize to various imaging pipelines. Please refer to the supplementary details for more analysis of CRF estimation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.722, + 0.634, + 0.738 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We presented IRIS, an inverse rendering framework for indoor scenes from casually captured LDR images. Our key innovations include HDR lighting restoration, CRF estimation, surface material estimation, and alternating optimization to recover all scene properties accurately. IRIS estimates accurate surface material, spatially varying HDR illumination, and CRF from real-world LDR images. As a result, IRIS enables photorealistic rendering under diverse lighting and from arbitrary viewpoints, showcasing the significant potential for content creation applications." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "472" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.156 + ], + "angle": 0, + "content": "[1] Dejan Azinovic, Tzu-Mao Li, Anton Kaplanyan, and Matthias Nießner. Inverse path tracing for joint material and lighting estimation. In CVPR, 2019. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.159, + 0.484, + 0.213 + ], + "angle": 0, + "content": "[2] Sai Bi, Zexiang Xu, Pratul Srinivasan, Ben Mildenhall, Kalyan Sunkavalli, Miloš Hasan, Yannick Hold-Geoffroy, David Kriegman, and Ravi Ramamoorthi. Neural reflectance fields for appearance acquisition. arXiv:2008.03824, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.216, + 0.484, + 0.269 + ], + "angle": 0, + "content": "[3] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.272, + 0.484, + 0.313 + ], + "angle": 0, + "content": "[4] Robert L Cook and Kenneth E. Torrance. A reflectance model for computer graphics. ACM Transactions on Graphics (ToG), 1982. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.315, + 0.484, + 0.355 + ], + "angle": 0, + "content": "[5] Paul E. Debevec and Jitendra Malik. Recovering high dynamic range radiance maps from photographs. In SIGGRAPH, 1997. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.359, + 0.484, + 0.413 + ], + "angle": 0, + "content": "[6] Valentin Deschaintre, Miika Aittala, Fredo Durand, George Drettakis, and Adrien Bousseau. Single-image SVBRDF capture with a rendering-aware deep network. ACM Transactions on Graphics (ToG), 37(4):128:1-15, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.416, + 0.484, + 0.456 + ], + "angle": 0, + "content": "[7] Ainaz Eftekhar, Alexander Sax, Jitendra Malik, and Amir Zamir. Omnidata: A scalable pipeline for making multi-task mid-level vision datasets from 3D scans. In ICCV, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.459, + 0.484, + 0.526 + ], + "angle": 0, + "content": "[8] Marc-Andre Gardner, Kalyan Sunkavalli, Ersin Yumer, Xiaohui Shen, Emiliano Gambaretto, Christian Gagné, and Jean-François Lalonde. Learning to predict indoor illumination from a single image. ACM Transactions on Graphics, 36(6): 176:1-14, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.529, + 0.484, + 0.584 + ], + "angle": 0, + "content": "[9] Dan B Goldman, Brian Curless, Aaron Hertzmann, and Steven M Seitz. Shape and spatially-varying BRDFs from photometric stereo. IEEE Transactions on Pattern Analysis and Machine Intelligence, 32(6):1060-1071, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.586, + 0.483, + 0.613 + ], + "angle": 0, + "content": "[10] Michael D Grossberg and Shree K Nayar. What is the space of camera response functions? In CVPR, 2003. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.616, + 0.484, + 0.655 + ], + "angle": 0, + "content": "[11] Michael D. Grossberg and Shree K. Nayar. Modeling the space of camera response functions. IEEE TPAMI, 26(10): 1272-1282, 2004. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.659, + 0.483, + 0.699 + ], + "angle": 0, + "content": "[12] Jon Hasselgren, Nikolai Hofmann, and Jacob Munkberg. Shape, light & material decomposition from images using monte carlo rendering and denoising. NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.702, + 0.483, + 0.741 + ], + "angle": 0, + "content": "[13] Xin Huang, Qi Zhang, Ying Feng, Hongdong Li, Xuan Wang, and Qing Wang. HDR-NeRF: High dynamic range neural radiance fields. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.745, + 0.483, + 0.799 + ], + "angle": 0, + "content": "[14] Wenzel Jakob, Sébastien Speierer, Nicolas Roussel, Merlin Nimier-David, Delio Vicini, Tizian Zeltner, Baptiste Nicolet, Miguel Crespo, Vincent Leroy, and Ziyi Zhang. Mitsuba 3 renderer, 2022. https://mitsuba-renderer.org. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.802, + 0.483, + 0.842 + ], + "angle": 0, + "content": "[15] Haian Jin, Isabella Liu, Peijia Xu, Xiaoshuai Zhang, Songfang Han, Sai Bi, Xiaowei Zhou, Zexiang Xu, and Hao Su. TensoIR: Tensorial inverse rendering. CVPR, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.845, + 0.483, + 0.87 + ], + "angle": 0, + "content": "[16] James T Kajiya. The rendering equation. In SIGGRAPH, 1986. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[17] Jaroslav Krivánek and Pascal Gautron. Practical global illumination with irradiance caching. Springer, 2022. 4" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.134 + ], + "angle": 0, + "content": "[18] Eric Lafortune. Mathematical models and Monte Carlo algorithms for physically based rendering. PhD thesis, Katholieke Universiteit Leuven, 1996. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.908, + 0.19 + ], + "angle": 0, + "content": "[19] Xiao Li, Yue Dong, Pieter Peers, and Xin Tong. Modeling surface appearance from a single photograph using self-augmented convolutional neural networks. ACM Transactions on Graphics (ToG), 36(4):45:1-11, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.193, + 0.907, + 0.233 + ], + "angle": 0, + "content": "[20] Zhengqi Li and Noah Snavely. CGIntrinsics: Better intrinsic image decomposition through physically-based rendering. In ECCV, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.236, + 0.906, + 0.262 + ], + "angle": 0, + "content": "[21] Zhengqi Li and Noah Snavely. Learning intrinsic image decomposition from watching the world. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.265, + 0.907, + 0.305 + ], + "angle": 0, + "content": "[22] Zhengqin Li, Kalyan Sunkavalli, and Manmohan Chandraker. Materials for masses: SVBRDF acquisition with a single mobile phone image. In ECCV, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.308, + 0.906, + 0.361 + ], + "angle": 0, + "content": "[23] Zhengqin Li, Mohammad Shafiei, Ravi Ramamoorthi, Kalyan Sunkavalli, and Manmohan Chandraker. Inverse rendering for complex indoor scenes: Shape, spatially-varying lighting and SVBRDF from a single image. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.364, + 0.907, + 0.418 + ], + "angle": 0, + "content": "[24] Zhengqin Li, Ting-Wei Yu, Shen Sang, Sarah Wang, Meng Song, Yuhan Liu, Yu-Ying Yeh, Rui Zhu, Nitesh Gundavarapu, Jia Shi, et al. OpenRooms: An open framework for photorealistic indoor scene datasets. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.421, + 0.907, + 0.475 + ], + "angle": 0, + "content": "[25] Zhengqin Li, Jia Shi, Sai Bi, Rui Zhu, Kalyan Sunkavalli, Miloš Hasan, Zexiang Xu, Ravi Ramamoorthi, and Manmohan Chandraker. Physically-based editing of indoor scene lighting from a single image. In ECCV, 2022. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.477, + 0.908, + 0.531 + ], + "angle": 0, + "content": "[26] Zhi-Hao Lin, Bohan Liu, Yi-Ting Chen, David Forsyth, Jia-Bin Huang, Anand Bhattachad, and Shenlong Wang. UrbanIR: Large-scale urban scene inverse rendering from a single video. 3DV, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.534, + 0.908, + 0.588 + ], + "angle": 0, + "content": "[27] Yu-Lun Liu, Wei-Sheng Lai, Yu-Sheng Chen, Yi-Lung Kao, Ming-Hsuan Yang, Yung-Yu Chuang, and Jia-Bin Huang. Single-image HDR reconstruction by learning to reverse the camera pipeline. In CVPR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.591, + 0.908, + 0.657 + ], + "angle": 0, + "content": "[28] Demetris Marnerides, Thomas Bashford-Rogers, Jonathan Hatchett, and Kurt Debattista. ExpandNet: A deep convolutional neural network for high dynamic range expansion from low dynamic range content. Computer Graphics Forum, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.661, + 0.908, + 0.715 + ], + "angle": 0, + "content": "[29] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.718, + 0.908, + 0.771 + ], + "angle": 0, + "content": "[30] Ben Mildenhall, Peter Hedman, Ricardo Martin-Brualla, Pratul Srinivasan, and Jonathan T. Barron. NeRF in the dark: High dynamic range view synthesis from noisy raw images. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.775, + 0.908, + 0.815 + ], + "angle": 0, + "content": "[31] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM TOG, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.907, + 0.87 + ], + "angle": 0, + "content": "[32] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting triangular 3D models, materials, and lighting from images. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[33] Merlin Nimier-David, Zhao Dong, Wenzel Jakob, and Anton Kaplanyan. Material and lighting reconstruction for complex" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "473" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "indoor scenes with texture-space differentiable rendering. In Eurographics Symposium on Rendering, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.123, + 0.483, + 0.178 + ], + "angle": 0, + "content": "[34] Julien Philip, Sébastien Morgenthaler, Michael Gharbi, and George Drettakis. Free-viewpoint indoor neural relighting from multi-view stereo. ACM Trans. Graph., 40(5):194:1-18, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.181, + 0.483, + 0.236 + ], + "angle": 0, + "content": "[35] Erik Reinhard, Greg Ward, Sumanta Pattanaik, Paul Debevec, Wolfgang Heidrich, and Karol Myszkowski. High Dynamic Range Imaging - Acquisition, Display and Image-Based Lighting. Morgan Kaufmann, 2nd edition, 2010. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.239, + 0.483, + 0.307 + ], + "angle": 0, + "content": "[36] Mike Roberts, Jason Ramapuram, Anurag Ranjan, Atulit Kumar, Miguel Angel Bautista, Nathan Paczan, Russ Webb, and Joshua M. Susskind. Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.31, + 0.483, + 0.366 + ], + "angle": 0, + "content": "[37] Dario Seyb, Peter-Pike Sloan, Ari Silvennoinen, Michal Iwanicki, and Wojciech Jarosz. The design and evolution of the UberBake light baking system. ACM Transactions on Graphics (TOG), 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.368, + 0.483, + 0.422 + ], + "angle": 0, + "content": "[38] Pratul P Srinivasan, Ben Mildenhall, Matthew Tancik, Jonathan T Barron, Richard Tucker, and Noah Snavely. Lighthouse: Predicting lighting volumes for spatially-coherent illumination. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.426, + 0.483, + 0.481 + ], + "angle": 0, + "content": "[39] Pratul P. Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T. Barron. NeRV: Neural reflectance and visibility fields for relighting and view synthesis. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.483, + 0.539 + ], + "angle": 0, + "content": "[40] Cheng Sun, Guangyan Cai, Zhengqin Li, Kai Yan, Cheng Zhang, Carl Marshall, Jia-Bin Huang, Shuang Zhao, and Zhao Dong. Neural-pbir reconstruction of shape, material, and illumination. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.542, + 0.483, + 0.583 + ], + "angle": 0, + "content": "[41] Zian Wang, Jonah Philion, Sanja Fidler, and Jan Kautz. Learning indoor inverse rendering with 3D spatially-varying lighting. In ICCV, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.586, + 0.483, + 0.655 + ], + "angle": 0, + "content": "[42] Zian Wang, Tianchang Shen, Jun Gao, Shengyu Huang, Jacob Munkberg, Jon Hasselgren, Zan Gojcic, Wenzheng Chen, and Sanja Fidler. Neural fields meet explicit geometric representations for inverse rendering of urban scenes. In CVPR, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.657, + 0.483, + 0.727 + ], + "angle": 0, + "content": "[43] Liwen Wu, Rui Zhu, Mustafa B Yaldiz, Yinhao Zhu, Hong Cai, Janarbek Matai, Fatih Porikli, Tzu-Mao Li, Manmohan Chandraker, and Ravi Ramamoorthi. Factorized inverse path tracing for efficient and accurate material-lighting estimation. In ICCV, 2023. 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.729, + 0.483, + 0.771 + ], + "angle": 0, + "content": "[44] Xin Yang, Ke Xu, Yibing Song, Qiang Zhang, Xiaopeng Wei, and Rynson WH Lau. Image correction via deep reciprocating HDR transformation. In CVPR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.774, + 0.483, + 0.828 + ], + "angle": 0, + "content": "[45] Yao Yao, Jingyang Zhang, Jingbo Liu, Yihang Qu, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. NeILF: Neural incident light field for physically-based material estimation. In ECCV, 2022. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.831, + 0.483, + 0.899 + ], + "angle": 0, + "content": "[46] Lior Yariv, Peter Hedman, Christian Reiser, Dor Verbin, Pratul P Srinivasan, Richard Szeliski, Jonathan T Barron, and Ben Mildenhall. Bakedsdf: Meshing neural sdfs for real-time view synthesis. In ACM SIGGRAPH computer graphics, 2023. 4" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.134 + ], + "angle": 0, + "content": "[47] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. ScanNet++: A high-fidelity dataset of 3D indoor scenes. In ICCV, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.906, + 0.19 + ], + "angle": 0, + "content": "[48] Bohan Yu, Siqi Yang, Xuanning Cui, Siyan Dong, Baoquan Chen, and Boxin Shi. MILO: Multi-bounce inverse rendering for indoor scene with light-emitting objects. IEEE TPAMI, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.906, + 0.247 + ], + "angle": 0, + "content": "[49] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. MonoSDF: Exploring monocular geometric cues for neural implicit surface reconstruction. In NeurIPS, 2022. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.906, + 0.317 + ], + "angle": 0, + "content": "[50] Zheng Zeng, Valentin Deschaintre, Iliyan Georgiev, Yannick Hold-Geoffroy, Yiwei Hu, Fujun Luan, Ling-Qi Yan, and Miloš Hasan. RGB ↔ X: Image decomposition and synthesis using material- and lighting-aware diffusion models. In SIGGRAPH, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.319, + 0.906, + 0.347 + ], + "angle": 0, + "content": "[51] Jinsong Zhang and Jean-François Lalonde. Learning high dynamic range from outdoor panoramicas. In ICCV, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.906, + 0.403 + ], + "angle": 0, + "content": "[52] Jingyang Zhang, Yao Yao, Shiwei Li, Jingbo Liu, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. NeILF++: Inter-reflectable light fields for geometry and material estimation. In ICCV, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.405, + 0.906, + 0.459 + ], + "angle": 0, + "content": "[53] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snavely. PhySG: Inverse rendering with spherical Gaussians for physics-based material editing and relighting. In CVPR, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.461, + 0.906, + 0.516 + ], + "angle": 0, + "content": "[54] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. NeR-Factor: Neural factorization of shape and reflectance under an unknown illumination. ACM TOG, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.518, + 0.906, + 0.573 + ], + "angle": 0, + "content": "[55] Jingsen Zhu, Yuchi Huo, Qi Ye, Fujun Luan, Jifan Li, Dianbing Xi, Lisha Wang, Rui Tang, Wei Hua, Hujun Bao, and Rui Wang. I²-SDF: Intrinsic indoor scene reconstruction and editing via raytracing in neural SDFs. In CVPR, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.575, + 0.906, + 0.629 + ], + "angle": 0, + "content": "[56] Rui Zhu, Zhengqin Li, Janarbek Matai, Fatih Porikli, and Manmohan Chandraker. IRISformer: Dense vision transformers for single-image inverse rendering in indoor scenes. In CVPR, 2022. 2, 4" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.629 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.512, + 0.956 + ], + "angle": 0, + "content": "474" + } + ] +] \ No newline at end of file diff --git a/2025/IRIS_ Inverse Rendering of Indoor Scenes from Low Dynamic Range Images/53c93036-c714-4f00-9960-2f72fce37288_origin.pdf b/2025/IRIS_ Inverse Rendering of Indoor Scenes from Low Dynamic Range Images/53c93036-c714-4f00-9960-2f72fce37288_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..dc55bd88b4715dc73e151ed360b45600b9d25a53 --- /dev/null +++ b/2025/IRIS_ Inverse Rendering of Indoor Scenes from Low Dynamic Range Images/53c93036-c714-4f00-9960-2f72fce37288_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1cedae1ec2d90974c928f76bd65b6550b5ee4ebd703fc134175ed07ce4a9d950 +size 8281377 diff --git a/2025/IRIS_ Inverse Rendering of Indoor Scenes from Low Dynamic Range Images/full.md b/2025/IRIS_ Inverse Rendering of Indoor Scenes from Low Dynamic Range Images/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5e58145e7c91b9422338cddd95f45c8706cafd08 --- /dev/null +++ b/2025/IRIS_ Inverse Rendering of Indoor Scenes from Low Dynamic Range Images/full.md @@ -0,0 +1,346 @@ +# IRIS: Inverse Rendering of Indoor Scenes from Low Dynamic Range Images + +Chih-Hao Lin1,2 + +Jia-Bin Huang $^{1,3}$ + +Zhengqin Li + +Zhao Dong1 + +Christian Richardt1 + +Tuotuo Li + +Michael Zollhöfer1 + +Johannes Kopf1 + +Shenlong Wang + +Changil Kim + +$^{1}$ Meta $^{2}$ University of Illinois Urbana-Champaign $^{3}$ University of Maryland College Park https://irisldr.github.io/ + +![](images/a25f7bdb278f3a437c12f874cec517b605582adcde0b705c67620274d7cdb59c.jpg) +Figure 1. Physically-based inverse rendering from LDR images. Our method takes LDR input images and estimates high-quality spatially-varying HDR lighting, physically-based material properties, and a camera response function (CRF). This decomposition allows us to insert objects or new light sources, and to relight the scene in a physically accurate manner. + +![](images/45a2c0323811594e2d8734d9a8db03655eac18ee26fb98076cda118ad6bd52c0.jpg) + +![](images/1f18b3f222474e7922aeede920737cf91aeaafc49426652883f4a83acc5ab410.jpg) + +![](images/1d57008291cd2f4717b0bfdee3c4d4972004e91e749557bd53408542007a3a20.jpg) + +# 1. Introduction + +# Abstract + +Inverse rendering seeks to recover 3D geometry, surface material, and lighting from captured images, enabling advanced applications such as novel-view synthesis, relighting, and virtual object insertion. However, most existing techniques rely on high dynamic range (HDR) images as input, limiting accessibility for general users. In response, we introduce IRIS, an inverse rendering framework that recovers the physically based material, spatially-varying HDR lighting, and camera response functions from multi-view, low-dynamic-range (LDR) images. By eliminating the dependence on HDR input, we make inverse rendering technology more accessible. We evaluate our approach on real-world and synthetic scenes and compare it with state-of-the-art methods. Our results show that IRIS effectively recovers HDR lighting, accurate material, and plausible camera response functions, supporting photorealistic relighting and object insertion. + +Physically based inverse rendering enables reconstructing realistic material properties and lighting in scenes. This decomposition is valuable for various applications, including relighting, material editing, and realistic object insertion. However, most existing inverse rendering methods require input images with high dynamic range (HDR) to capture the full light transport in the scene. This poses a significant barrier to broader adoption, as HDR capture typically requires specialized hardware or merging multiple aligned LDR images through exposure bracketing [35]. + +Most common imaging sensors do not capture sufficient dynamic range for scenes' light-emitting and dark regions. Moreover, cameras often convert raw sensor readings to 8-bit low-dynamic-range (LDR) images for storage and transmission. The non-linear mapping and quantization lead to the additional loss of lighting information. The complex light transport in indoor environments makes reconstructing the original HDR lighting - critical for faithful inverse rendering - even more challenging. If material and lighting estimation + +could be achieved from a "casual" capture using standard devices like cameras or phones, inverse rendering would become far more accessible to a wider range of users. + +Many state-of-the-art inverse rendering methods rely on HDR inputs to accurately recover material properties and lighting [43, 48, 52]. Some approaches attempt to overcome this issue by taking LDR images as inputs and solving for the lighting. However, these methods often make simplifying assumptions, such as infinitely distant light sources [42], requiring additional inputs like emitter masks [55], or neglecting multi-bounce light transport [45, 52, 55]. As a result, these methods struggle to handle the complex light transport in indoor scenes and often fail to produce high-quality surface material and lighting. + +To address these challenges, we introduce IRIS, an inverse rendering method for indoor scenes, using only multi-view LDR images. IRIS models tone mapping (i.e., HDR-to-LDR conversion), allowing us to work directly with LDR inputs. IRIS automatically identifies emitters and reconstructs spatially varying HDR lighting through physically based inverse rendering, which is crucial for faithful material estimation. Jointly estimating lighting, surface material, and camera response function (CRF) leads to unstable optimization due to ambiguities. To tackle this, we design a novel optimization strategy that effectively resolves these ambiguities, enabling high-quality estimation. Our core contributions are: + +1. IRIS faithfully estimates spatially-varying HDR lighting, physically-based materials, and camera response function from LDR images and outperforms several state-of-the-art inverse rendering methods. +2. We explicitly model LDR image formation in our (inverse) rendering pipeline so that LDR images can be used directly, broadening the accessibility of high-quality inverse rendering. +3. Finally, IRIS is extensively evaluated on synthetic and real-world scenes, demonstrating diverse and realistic view synthesis, relighting, and object insertion. + +# 2. Related Work + +Data-driven inverse rendering. Inverse rendering seeks to reconstruct a scene's geometry, material, and lighting from single or multiple images. This process is notoriously challenging and ill-posed due to the inherent ambiguity. Numerous methods employ deep priors learned from large-scale datasets for different tasks, including intrinsic image decomposition [20, 21], SVBRDF estimation [6, 19, 22, 50, 56], lighting estimation [8, 23, 38], lighting editing [25] and relighting [34]. These learning-based methods take a single or a few images [2, 38] as inputs, reducing capturing requirements of classical measurement-based methods [9]. However, high-quality datasets are essential for data-driven methods to achieve compelling generalization ability. Synthetic + +![](images/623c122cc4fd7ca90e59fd26643fa3100c61d5cc29df715dd759e542f954791f.jpg) +Surface material + +![](images/f4355db530dab1e2421567746725b4d6766152de6687f53fc9f26d04170ff64a.jpg) +Lighting + +![](images/4f2cebbd55fa7f3e052cf7c13277e25cf9d0781996e466523c66e457c41637e3.jpg) +Surface material +Figure 2. Limitation of SOTA. A typical image formation process causes the loss of lighting information, posing challenges in inverse rendering. FIPT [43] assumes HDR input and NeILF [45] ignores multi-bounce light transport. Both methods fail to estimate accurate material (red boxes) and HDR lighting (yellow boxes). We demonstrate significantly better results (Figure 1). + +![](images/ead907e3468f92182f191e4607671aa13abe8c1f98af1d41a665f260bbf4a536.jpg) +Lighting + +datasets are often used for training because ground-truth labeling is challenging to obtain in real environments. Despite efforts to create photorealistic synthetic datasets [20, 24, 36], the inherent diversity of real-world environments, especially in complex indoor scenes, still causes a substantial domain gap. Furthermore, these methods do not reconstruct complete 3D scenes and fail to render consistent novel views under novel illuminations. Our method utilizes prediction from a data-driven method [56] for initialization and reconstructs 3D geometry, material and lighting with a carefully designed optimization framework, demonstrating superior performance for real-world scenes. + +Optimization-based inverse rendering. Differentiable path tracing [14] enables the direct optimization of scene parameters with physically-based multi-bounce light transport. Recent methods, including MILO [48], IPT [1], and follow-up work [33] exploit this to jointly optimize BRDF and lighting emission, accounting for global light transport. FIPT [43] utilizes a factored light transport formulation for BRDF estimation and emitter detection. Nevertheless, these methods assume piecewise constant materials or are designed to take high-dynamic range images as input, and the robustness is limited in real-world scenes. Several recent NeRF-based approaches [12, 15, 26, 32, 40, 42, 53, 54] parameterize materials as a neural field and optimize with a volume rendering formulation. They tend to assume distant lighting and cannot handle spatially varying lighting in indoor environments well. NeILF [45] and NeILF++ [52] represent spatially-varying lighting as a neural field, $\mathrm{I}^2$ -SDF [55] constructs lighting from given emitter masks. While handling more complex lighting, these works require additional assumptions (e.g., HDR input, emitter masks), and single-bounce ray tracing cannot model complex light transport. In contrast, our method automatically identifies emitters and reconstructs HDR lighting from casual (LDR) images, while achieving high-quality material and lighting estimation that + +outperforms prior work in this practical scenario. + +CRF and HDR estimation. The complete information about the illumination is usually lost within the photography pipeline. Estimating the HDR lighting and CRF enables us to edit photos and achieve realistic object insertion. Debevec and Malik [5] propose to take multiple photos with varying exposure levels to recover the CRF and HDR image. Some previous works [27, 28, 44, 51] predict HDR images with a data-driven approach by learning the LDR-HDR mapping from a single image. However, these methods rely on HDR data for training, and the estimation is inconsistent across views and scenes. To handle the HDR-LDR mapping, FIPT [43] and RawNeRF [30] assume a known gamma correction function; NeILF++ [52] learn a single gamma correction parameter; HDR-NeRF [13] parameterizes tone-mapping function with an MLP. However, these works are designed to take HDR images or LDR with multiple exposure levels as input. We leverage physically-based rendering to recover spatially varying BRDF, HDR illumination, and CRF from LDR images. Our method generates realistic view synthesis, relighting, and object insertion. + +# 3. Background: LDR image formation + +Real-world radiance exhibits high dynamic range (HDR) and contains bright light sources and dark regions. To preserve the complete lighting information, HDR images are typically produced with a high-end camera, laborious exposure bracketing [5], and are stored in specialized format (16 or 32 bits per color channel). However, capturing HDR is often impractical for casual users and is typically absent in standard photography assets. Due to the limitations of standard sensors and displays, the radiance is frequently processed with multiple stages to produce low dynamic range (LDR) images stored in more lightweight formats (such as JPEG or PNG at 8 bits and RAW at 12 bits per color channel). We describe the major steps in this process below. + +Dynamic range clipping. Given the scene irradiance $E$ and exposure time $\Delta t$ , the observed radiance is clipped at a certain maximum value due to the sensor limitation, which could be formulated as: $Z_{c} = \min (E\Delta t,1)$ . Due to the clipping operation, the lighting information is lost for overexposed/saturated pixels. + +Non-linear mapping of CRF. To enhance the image's visual quality and match the human perception of the scene, the clipped intensity is further transformed with a non-linear camera response function (CRF) to produce LDR pixel values: $Z = \mathrm{CRF}(Z_c) = \mathrm{CRF}(\min(E\Delta t, 1))$ . Please refer to Debevec and Malik [5] for a more detailed imaging pipeline explanation. + +Challenges in inverse rendering. Real-world scenes often exhibit extreme variances in radiance, making it challenging to capture these scenes in LDR images without under + +exposing or saturating them. Due to the clipping and nonlinear CRF mapping, critical lighting information is lost, posing challenges in previous works, illustrated in Figure 2. In this work, we propose an inverse rendering framework that jointly recovers HDR lighting and CRF from LDR images. + +# 4. Representation and rendering + +Camera response function (CRF). The CRF is crucial for recovering original radiance from 8-bit LDR images. However, the CRF varies with each camera and is often unknown to users. Thus, deducing the CRF from LDR image observations is critical to our solution. We use Gross et al.'s [11] empirical model of response (EMoR) for CRF modeling [27]. They collected a database of 201 real-world CRFs of various devices and parameterized them as vectors. A mean curve $\bar{\mathbf{g}}$ and PCA basis $\mathbf{g}_b$ are calculated from the database, and we parameterize a learnable CRF as + +$$ +\mathbf {g} = \bar {\mathbf {g}} + \Sigma_ {b} w _ {b} \mathbf {g} _ {b}, \mathbf {g} \in \mathbb {R} ^ {1 0 2 4}, \tag {1} +$$ + +where $\{w_{b}\}$ weigh each basis curve and are learnable parameters. The non-linear function is discretized as $\mathrm{CRF}(x) = \mathrm{LERP}(\{g_k\}_k, x)$ , which linearly interpolates the uniformly sampled CRF $\{g_k\}_{k=0}^{1023} \in [0,1]^{1024}$ , independently per color channel. We make this design choice because it covers most of the CRF space, and the simplicity makes it easy to enforce regularization, which brings huge advantage in the complex optimization of inverse rendering. + +Spatially-varying lighting. While environment maps handle object-centric and outdoor scenes well [15, 39, 42, 53], they cannot capture complex spatially-varying lighting of indoor scenes. To handle this, we define direct lighting on the scene mesh and measure emission as view-independent radiance: $\mathbf{L}_{\mathrm{e}}(\mathbf{x}) \in \mathbb{R}^3$ . The parameterization of $\mathbf{L}_{\mathrm{e}}$ is detailed in our supplementary material. We also define an emitter mask $M_{\mathrm{e}}(\mathbf{x}) \in \{0,1\}$ to denote whether a surface point $\mathbf{x}$ is on an emitter. Global illumination is represented as a view-independent surface light field (SLF): $\mathbf{L}_{\mathrm{SLF}}(\mathbf{x}) \in \mathbb{R}^3$ . The sRGB pixel colors of input LDR images are linearized with an inverse CRF mapping: $\mathbf{I}_{\mathrm{linear}} = \mathrm{CRF}^{-1}(\mathbf{I}_{\mathrm{LDR}})$ . If a voxel contains no surface, it is set to zero; otherwise, it collects the average linearized radiances across all input images with ray casting. This greatly accelerates the approximation of global illumination without requiring multi-bounce ray tracing. + +Material BRDF. We represent spatially-varying materials using the Cook-Torrance BRDF [4]. The surface albedo $\mathbf{a}(\mathbf{x})\in \mathbb{R}^3$ , roughness $\sigma (\mathbf{x})\in \mathbb{R}$ , and metallicity $m(\mathbf{x})\in \mathbb{R}$ model the diffuse reflectance $\mathbf{k_d} = \mathbf{a}\cdot (1 - m)$ and specular reflectance $\mathbf{k_s} = 0.04\cdot (1 - m)$ , and are parameterized as a neural field $\mathbf{f}\colon \mathbf{x}\mapsto (\mathbf{a},\sigma ,m)$ , implemented with multi-resolution hash encoding and an MLP [31]. + +Factorized light transport. To render images from scene geometry, material, and lighting information, We follow + +![](images/390b3c22db5234603ac9224731083b54eecb757ed4d85522755cf096d9eff524.jpg) +Figure 3. Framework Overview. Given multi-view posed LDR images and a surface mesh, our inverse rendering pipeline is divided into two main stages. In the initialization stage, we initialize the BRDF (Section 5.1), extract a surface light field (Section 4), and estimate emitter geometry (Equation (3)). In the optimization stage, we first recover HDR radiance from the LDR input (Section 5.2), then bake shading maps (Section 5.3), and jointly optimize BRDF and CRF parameters (Section 5.4). The improved parameters are used to refine the emission again. These three steps are repeated until convergence. + +the rendering equation [16] to model physically-based light transport for realistic rendering: + +$$ +\mathbf {L} _ {\mathrm {o}} (\mathbf {x}, \omega_ {\mathrm {o}}) = \mathbf {L} _ {\mathrm {e}} (\mathbf {x}, \omega_ {\mathrm {o}}) + \int_ {\Omega^ {+}} \mathbf {L} _ {\mathrm {i}} (\mathbf {x}, \omega_ {\mathrm {i}}) f (\mathbf {x}, \omega_ {\mathrm {i}}, \omega_ {\mathrm {o}}) d \omega_ {\mathrm {i}}, \tag {2} +$$ + +where $\mathbf{L}_{\mathrm{o}}$ is the radiance observed along a ray at a 3D position $\mathbf{x}$ in direction $\omega_{\mathrm{o}}$ , $\mathbf{L}_{\mathrm{e}}$ is the emission term, $\mathbf{L}_{\mathrm{r}} = \int_{\Omega^{+}}\mathbf{L}_{\mathrm{i}}(\mathbf{x},\omega_{\mathrm{i}})f(\mathbf{x},\omega_{\mathrm{i}},\omega_{\mathrm{o}})d\omega_{\mathrm{i}}$ is the reflectance term, and $f(\mathbf{x},\omega_{\mathrm{i}},\omega_{\mathrm{o}})$ is the BRDF. The recursive nature of the rendering equation is approximated with Monte-Carlo path tracing [16, 18]. However, modeling multiple bounces is still computationally expensive and unstable in the inverse rendering process. To enhance the efficiency and robustness of the optimization, we follow FIPT [43] and adopt factorized light transport [17, 37, 41], which precomputes shading maps $\mathbf{L}_{\mathrm{d}}$ , $\mathbf{L}_{\mathrm{s}}^{0}$ , $\mathbf{L}_{\mathrm{s}}^{1}$ and linearly interpolates with roughness to obtain final rendering. Detailed notation and formulation are available in the supplemental material. + +# 5. Inverse Rendering from LDR Images + +Given multi-view posed LDR images $\{\mathbf{I}_i\}$ captured with the same or varying exposure levels $\{\Delta t_i\}$ , our goal is to estimate spatially-varying HDR lighting, surface material (as albedo, roughness, and metallic), and the camera response function (CRF) shared by all input images. We assume that the primary light sources are observed in the input images, following the recent literature [43, 48, 55]. + +Previous approaches perform differentiable path tracing during the material and lighting optimization [1, 33], but the number of samples is limited due to heavy computation. This introduces high variance and unstable estimation. The factorized formulation [17, 37, 41, 43] splits light transport into BRDF parameters and shading, and enables updating + +them alternately. While the method is shown to be more stable and produce high-quality decompositions, it requires HDR images as input to aggregate and bake shading maps, detect emitters, and estimate HDR emissions. + +To address this ill-posed optimization problem, we propose a multi-stage optimization strategy alternatingly estimating spatially varying lighting, material, and CRF. Specifically, we (1) initialize the BRDF with an off-the-shelf monocular inverse rendering method (Section 5.1), (2) recover HDR lighting from LDR with physically-based rendering (Section 5.2), (3) bake the diffuse and specular shading (Section 5.3), and (4) jointly estimate BRDF and CRF (Section 5.4). We repeat Steps 2-4 in that order until all estimates are converged. Figure 3 illustrates our framework. + +Geometry reconstruction. From input images with camera poses, we reconstruct the geometry with BakedSDF [46], and use normal estimation from off-the-shelf method [7] for regulating surface geometry, following MonoSDF [49]. + +# 5.1. BRDF Initialization + +For each input image $\mathbf{I}_i$ , we predict a 2D albedo map $\mathbf{A}_i$ using an off-the-shelf single-image albedo estimation method [56] and averaging per-pixel albedo values within each semantic part. Given $\{\mathbf{A}_i\}$ , we initialize albedo $\mathbf{a}(\mathbf{x})$ by minimizing the error between the projected albedo and the estimated albedo across all input images: $\min_{\mathbf{a}} \sum_i \| \pi_i(\mathbf{a}) - \mathbf{A}_i \|_2$ . Here, $\pi_i$ represents the perspective projection onto the image space of $\mathbf{I}_i$ by intersecting camera rays with the surface mesh of the input scene and reading $\mathbf{a}$ at their first intersections. We obtain semantic category labels from part IDs in synthetic scenes or using Mask2Former [3] for real scenes. Roughness and metallicity are both initialized to 0.5. + +![](images/ec1874197a8d9d93fdee5bd1aaf6acfb9b4cf15bcd686a521be7d86c6557cc76.jpg) +Figure 4. Emitter geometry estimation $M_{\mathrm{e}}(\mathbf{x})$ . The point $\mathbf{x}_1$ on the window is saturated across all input views and thus identified as an emitter. The point $\mathbf{x}_2$ on the table is reflective and saturated in some views (e.g., view 2) but not in others and is NOT an emitter. + +# 5.2. HDR Emission Restoration + +Given that emitters appear bright across views and produce saturated pixels in LDR images, we estimate the binary emitter geometry $M_{\mathrm{e}}(\mathbf{x})$ on the surface mesh as follows: + +$$ +M _ {\mathrm {e}} (\mathbf {x}) = \left\{ \begin{array}{l l} 1 & \text {i f} \frac {1}{N} \sum_ {i} v _ {i} (\mathbf {x}) \mathbf {I} _ {i} \left(\pi_ {i} (\mathbf {x})\right) \geq 0. 9 9 \\ 0 & \text {o t h e r w i s e ,} \end{array} \right. \tag {3} +$$ + +where $v(\cdot)\in \{0,1\}$ indicates visibility. See Figure 4. + +The HDR intensity is crucial for inverse rendering but is missing in the LDR images. To recover the HDR lighting, we perform physically-based rendering with single-bounce path tracing to generate the images (2). To estimate the radiance at the point $\mathbf{x}$ observed from a camera ray, we trace $N = 128$ secondary rays from $\mathbf{x}$ , sampled with the current BRDF estimates. The resulting radiance is approximated by: + +$$ +\mathbf {L} _ {\mathrm {r}} \left(\mathbf {x}, \omega_ {\mathrm {o}}\right) = \sum_ {i = 1} ^ {N} \mathbf {L} _ {\text {e n d}} \left(\mathbf {x} ^ {\prime}\right) f \left(\mathbf {x}, \omega_ {\mathrm {i}}, \omega_ {\mathrm {o}}\right), \tag {4} +$$ + +$$ +\mathbf {L} _ {\text {e n d}} (\mathbf {x}) = \left\{ \begin{array}{l l} \mathbf {L} _ {\mathrm {e}} (\mathbf {x}) & \text {i f} M _ {\mathrm {e}} (\mathbf {x}) = 1 \\ \mathbf {L} _ {\text {S L F}} (\mathbf {x}) & \text {o t h e r w i s e ,} \end{array} \right. \tag {5} +$$ + +where $\mathbf{x}' = \mathbf{x} + d\cdot \boldsymbol{\omega}_{\mathrm{i}}$ is the intersected point of the secondary ray along direction $\boldsymbol{\omega}_{\mathrm{i}}$ . Intuitively, we retrieve the radiance from the learnable emission radiance $\mathbf{L}_{\mathrm{e}}(\mathbf{x})$ if the secondary ray hits an emitter directly, or otherwise from the pre-computed surface light field $\mathbf{L}_{\mathrm{SLF}}(\mathbf{x})$ to approximate the global illumination. Figure 5 shows an illustration. + +To recover the HDR emitter radiance, we minimize the photometric loss with gradient descent: $\min_{\mathbf{L}_{\mathrm{e}}}\mathcal{L}_{\mathrm{photo}}$ , where + +$$ +\mathcal {L} _ {\mathrm {p h o t o}} = \sum_ {i} \| \mathrm {C R F} \left(\min \left(\mathbf {L} _ {\mathrm {o}} \left(\mathbf {x}, \omega_ {\mathrm {o}}\right) \Delta t _ {i}, 1\right)\right) - \mathbf {I} _ {i} \| _ {2}. \tag {6} +$$ + +The emitter radiance $\mathbf{L}_{\mathrm{e}}(\mathbf{x})$ is enhanced to match the captured images, while the BRDF and CRF are fixed in this stage. Intuitively, in order to minimize the photometric loss with physically-based rendering, the emitter intensity $\mathbf{L}_{\mathrm{e}}(\mathbf{x})$ should be increased significantly and enhance the overall brightness of the rendered image. Therefore, the HDR lighting is recovered in this process. + +![](images/afa70ad937f130915d42a5e9b0ab63b70c9583aa7aa8fbf7f494b64e91136125.jpg) +Figure 5. HDR emission restoration. Top: Ray sampling process. Learnable direct lighting $\mathbf{L}_{\mathrm{e}}$ is retrieved if a ray hits an emitter (e.g., window) or $\mathbf{L}_{\mathrm{SLF}}$ is retrieved otherwise (e.g., wall). Bottom: HDR restoration process. By performing differentiable physically-based rendering, the photometric loss enhances the emitter intensity $\mathbf{L}_{\mathrm{e}}$ . + +# 5.3. Shading Baking + +After the HDR emission is recovered in the previous stage, we bake the diffuse and specular shading maps $\mathbf{L}_{\mathrm{d}}$ , $\mathbf{L}_{\mathrm{s}}^{0}$ , $\mathbf{L}_{\mathrm{s}}^{1}$ , following factorized light transport (details in supplementary material). Given a surface point intersected by a camera ray, multiple rays are sampled according to the current (fixed) BRDF. The incident radiance is computed with multibounce path tracing, where a ray keeps tracing if it hits a non-emissive specular surface and stops otherwise. The radiance of the endpoint follows Equation (5). + +# 5.4. BRDF & CRF Optimization + +With the baked shading maps $\mathbf{L}_{\mathrm{d}}$ , $\mathbf{L}_{\mathrm{s}}^{0}$ , and $\mathbf{L}_{\mathrm{s}}^{1}$ , the albedo, metallicity, roughness, and CRF are jointly optimized. The optimization involves the following objective function: + +$$ +\min _ {\mathbf {a}, m, \sigma , \mathbf {g}} \mathcal {L} _ {\text {p h o t o}} + \lambda_ {\mathrm {a}} \mathcal {L} _ {\text {a l b e d o}} + \lambda_ {\mathrm {c}} \mathcal {L} _ {\text {C R F}} + \lambda_ {\mathrm {m}} \mathcal {L} _ {\text {m a t}}, \tag {7} +$$ + +where $\lambda_{\mathrm{a}} = 0.01$ $\lambda_{\mathrm{c}} = 0.001$ , and $\lambda_{\mathrm{m}} = 0.005$ + +Photometric loss. We perform physically-based rendering, and the photometric loss $\mathcal{L}_{\mathrm{photo}}$ is computed as in Eq. 6. + +Albedo regularization. Motivated by the shift-scale-invariant loss in MonoSDF [49], we adopt a shift-invariant loss against the monocular albedo maps $\{\mathbf{A}_i\}$ : + +$$ +\mathcal {L} _ {\text {a l b e d o}} = \sum_ {i} \| \pi_ {i} (\mathbf {a}) - s _ {i} \mathbf {A} _ {i} \| _ {2}, \tag {8} +$$ + +where $s_i$ is the scale used to align the two albedo maps. + +CRF regularization. Based on the fair assumption that the learned CRF should not deviate significantly from the mean curve $\bar{\mathbf{g}}$ (Eq. 1), and that the curve should be monotonically increasing, we regularize the $L_{2}$ norm of the estimated coefficients of the PCA bases, and enforce the monotonicity of the sampled CRF as follows: + +$$ +\mathcal {L} _ {\mathrm {C R F}} = \| \mathbf {w} \| _ {2} + \sum_ {i} \max \left(g _ {i - 1} - g _ {i}, 0\right). \tag {9} +$$ + +Material regularization. We apply roughness-metallicity regularization $\mathcal{L}_{\mathrm{mat}}$ , which enforces the consistency of surface roughness and metallic within the same semantic instances (e.g. chairs, tables). While the shading and emission are initially inaccurate and lead to erroneous BRDFs, we alternately update the BRDFs and produce better emission and shading, improving the BRDFs. Our alternating strategy yields more stable optimization and better estimation. + +# 6. Experiments + +We evaluate our method on both synthetic and real-world data and compare it against several baseline methods. We analyze the quality of our inverse rendering, relighting, and novel-view synthesis capability. Please see our supplementary material for additional results and evaluations. We will release our code and data. + +Datasets. We evaluate our method and baselines on the ScanNet++ [47] and FIPT [43] datasets. ScanNet++ contains 380 real-world indoor scenes and collects multi-view LDR images with a DSLR camera at a constant exposure level. We select four scenes that cover a variety of indoor scenes (e.g., office, bathroom). The FIPT dataset consists of four synthetic and two real-world scenes. The synthetic scenes provide ground-truth geometry and material properties for evaluation. We used the LDR images provided in the dataset, where the lighting is clipped in overly-exposed regions, compressed with a nonlinear CRF, quantized into 256 levels, and stored in 8-bit PNG format. We adopt the constant exposure setting to provide a fair comparison with baselines, as they cannot handle varying exposure. We further demonstrate that IRIS handles varying exposure levels of input LDR images and estimates CRF jointly. + +Baselines. We compare our method with the following baselines. Li et al. [25] estimate BRDF, geometry (depth & normal), and lighting condition of an indoor scene from a single image using a data-driven approach. In addition to the single image, the model also takes emitter masks as input, and we provide the mask estimated by our method (3). NeILF [45] targets inverse rendering from known geometry and multiview images. The spatially-varying BRDF and illumination are parameterized as neural fields [29]. Physically-based rendering and single-bounce ray tracing are performed during the training and rendering process. The HDR-to-LDR tone mapping is modeled as a learnable gamma correction. The original FIPT [43] is designed for HDR input, and fails to estimate the emission mask from LDR images. Thus, we additionally provide the emission mask estimated by our method (3). We denote this modified version as "FIPT*". + +To further evaluate our performance, we compare the original FIPT with HDR input if the HDR information is available. This privileged information will make illumination estimation much more accurate. + +Table 1. BRDF-emission comparison on synthetic data. Numbers are averaged across four synthetic scenes with GT [43]. The best metrics among LDR methods are highlighted in bold. + +
MethodInputkda'σLe
PSNR ↑IoU ↑L2 ↓
Li et al [25]LDR16.7314.0111.300.352.29
NeILF [45]LDR16.8514.0216.96——
FIPT*LDR15.499.744.990.690.28
OursLDR22.3317.9221.380.690.12
FIPT [43]HDR29.9525.9826.370.860.03
+ +# 6.1. Inverse Rendering of Real Scenes + +We provide a qualitative evaluation of inverse rendering in Figure 6. Conducting single-bounce ray tracing in the neural light field, NeILF [45] fails to remove the shadow in diffuse reflectance $\mathbf{k}_{\mathrm{d}}$ , estimates lower roughness $\sigma$ on the walls than the mirror, and produces an incorrect emission mask. FIPT* struggles to recover surface material from LDR input, where accurate lighting information is missing. Please note that FIPT* takes emitter masks from Equation (3) as additional input, but does not recover HDR emission in the optimization procedure. On the other hand, IRIS recovers an almost shading-free diffuse field. Regarding roughness, our method effectively identifies specular surfaces with low roughness on the mirror, demonstrating better specular/rough material separation. Additionally, our proposed HDR emission estimation proves effective for real scenes, recovering HDR lighting sources (i.e., window and ceiling light). The restored HDR lighting makes the light transport more accurate and significantly enhances the performance of inverse rendering. + +# 6.2. View Synthesis and Relighting of Real Scenes + +We compare the relighting with FIPT* in Figure 7. FIPT* fails to estimate accurate surface material from LDR input and tends to recover low roughness for most surfaces, which unrealistically reflects the inserted light source and objects. To be more specific, the ground of the office (first scene) and the wall of the bathroom (second scene) look overly smooth and specular, which is not consistent with the observation and produces unrealistic reflections. In contrast, IRIS produces better relighting and object insertion quality thanks to better inverse rendering. The reflections on the specular surfaces (e.g., whiteboard, mirror) and the shading on the wall demonstrate the strength of IRIS for realistic relighting. Furthermore, the recovered HDR lighting can produce naturalistic specular reflection and shadows for inserted objects. Additionally, we insert moving light sources in the scene, shown in Figure 8, where the shadow on the ground and shading on the table change consistently. IRIS enables the photorealistic rendering of novel objects under diverse lighting and from arbitrary viewpoints, showcasing the significant potential for content creation applications. + +![](images/9f1ef19091189052fc0baf90f7b6eabf7bd4cdbacdb68c5b113ea16e503c1f19.jpg) +Figure 6. Decompositions of real scenes from ScanNet++ [47]. In addition to good reconstruction, IRIS produces a detailed diffuse map, identifies the mirror as a low-roughness region, and recovers HDR lighting from windows and ceiling light. Please note that the emission maps are normalized to visualize the relative intensity. + +![](images/ecbc00fa12f23d8c3a13c4900109eefb7b3585b4dbdd8a5bf7df0ddf20890c7c.jpg) +Figure 7. Relighting and object insertion with real scenes from the ScanNet++ [47] and FIPT datasets [43]. When inserting new light sources (Relighting 1 & 2), IRIS not only changes shadows but also produces reflection of the light on the whiteboard (first scene) and mirror (second scene), enhancing realism substantially. The recovered HDR lighting also strengthens the reflection and shadow of inserted objects ("Object Insertion" column). + +# 6.3. Quantitative Evaluation + +Quantitative evaluation of inverse rendering is difficult for real-world scenes as ground truth is not available. As a result, we evaluate IRIS and baselines on the synthetic scenes from FIPT, where ground truth of surface material, illumina- + +tion, and relighting are provided. Please note that we use the ground-truth mesh during optimization for all methods to minimize the influence of imperfect geometry and conduct a fair comparison. We compute the metrics of inverse rendering and provide the comparison in Table 1, which aggregates + +![](images/7bdac9d5775894e719933053cd92fa6664dc83509605d3211c2c508bdd9b28e2.jpg) +Reconstruction + +![](images/1e0287af317380a0de0e428152b06857a3422a84bdee19af3d20194ece9e4dd8.jpg) +Relight 2 +Figure 8. Relighting with a moving light ball in a real scene. + +![](images/47f37bc9591563489e577d706a1323d34d2c5d36516a5372171188687e3a1486.jpg) +Relight 3 + +![](images/bfa617a85da7f2f229e9ff2a4ed943149c37afa3c93074556c5a557f6db6f6f7.jpg) +Figure 9. CRF estimation from real-world devices. The orange dashed line is the ground truth CRF, and the blue line our estimate. + +![](images/5739447e2d8337658fbe15bb0fd8b7090ac15092f8319022289fba06645e86a9.jpg) + +![](images/977019b9a4a2a13ad98529e9ce91ccdd642e418274d9af61c74ba58c46c773aa.jpg) + +![](images/7cdb30f3648732ac98e9e51acd30aba19a9bc3a40bb08acbee964f6f28140033.jpg) + +the results from the 4 synthetic FIPT scenes [43]. + +We compare the roughness $\sigma$ , diffuse reflectance $\mathbf{k}_{\mathrm{d}}$ among the diffuse surfaces, and material reflectance: $\mathbf{a}^{\prime}$ . We report the PSNR for these BRDF estimates. For emission, we calculate the intersection over the union (IoU) of the emission masks and compare the $L_{2}$ error of estimated emission maps. All the estimates are rendered in image space and compared with ground truth. + +Table 1 shows that our method surpasses all baseline methods that use LDR inputs regarding inverse rendering quality. Compared to the state-of-the-art single-view method [25], our method excels in all metrics across various scenes, indicating the superior decomposition of material properties from multi-view observations. Furthermore, even when the estimated emitter mask is provided for FIPT*, it fails to recover emission radiance, leading to compromised BRDF estimation accuracy. The $L_{2}$ error of emission maps shows that our proposed HDR restoration (Section 5.2) successfully recovers the HDR lighting from LDR observations, which leads to better light transport modeling and more accurate BRDF estimation. We also compare with FIPT taking HDR images as input (bottom row), which serves as a reference, indicating complete lighting information is crucial for accurate inverse rendering. IRIS recovers HDR and achieves the smallest metric gaps among all methods taking LDR input. + +We compare novel-view synthesis and relighting of all competing algorithms and report results in Table 2. The novel-view synthesis results show that IRIS is comparable with baseline methods [45], while exhibiting superior material and lighting estimation. Our main focus is relighting, and IRIS outperforms all baselines with LDR input, demonstrating the effectiveness of inverse rendering. + +# 6.4. Ablation Study + +We validate our design choices and provide an ablation study in Table 3, where we compute PSNR for material $(\mathbf{k}_{\mathrm{d}},\mathbf{a}^{\prime},\sigma)$ and L2 for CRF curves and emission maps $\mathbf{L}_{\mathrm{e}}$ . The emitter mask estimation in Equation (3) identifies the emitter from + +Table 2. Quantitative results of novel-view synthesis (NVS) and relighting on synthetic scenes with ground truth. + +
MethodInputPSNR ↑SSIM ↑LPIPS ↓
NVSNeILF [45]LDR30.0310.8990.180
FIPT*LDR15.1800.7440.437
OursLDR29.4680.9030.181
FIPT [43]HDR28.7600.9020.186
RelightingLi et al. [25]LDR22.4440.8130.398
FIPT*LDR11.6260.7050.359
OursLDR22.8610.8820.173
FIPT [43]HDR29.0120.9020.139
+ +Table 3. Ablation study + +
Methodkda'σLeCRF
PSNR ↑IoU ↑L2 ↓L2 ↓
- Emitter mask Me (Equation (3))16.5011.304.810.000.488.91
- CRF modeling (Equation (1))18.1813.5917.520.690.12—
- HDR restoration (Section 5.2)18.2014.407.310.690.286.23
- Lalbedo (Equation (8))21.5917.3617.430.690.152.42
- LCRF (Equation (9))14.4912.8125.590.690.139.40
Full model (Ours)22.3317.9221.380.690.122.35
+ +LDR observations and improves the emission map in the second row. CRF estimation is critical for modeling HDR-to-LDR conversion and helps inverse rendering in the third row. Nevertheless, the HDR restoration (Section 5.2) plays an important role in recovering accurate light transport from the captured images. Once the HDR light is restored by our proposed approach, the material (especially roughness $\sigma$ ), CRF, and emission map are improved significantly. The results demonstrate the significance of each component and show that inverse rendering from casually captured LDR images benefits from our proposed framework. + +# 6.5. CRF Estimation + +To further validate the robustness of our method, we generate LDR images from FIPT's synthetic HDR scenes using different CRFs collected from real-world devices [10]. In this evaluation, we perform inverse rendering with IRIS and plot the CRF. The results in Figure 9 show that our method can generalize to various imaging pipelines. Please refer to the supplementary details for more analysis of CRF estimation. + +# 7. Conclusion + +We presented IRIS, an inverse rendering framework for indoor scenes from casually captured LDR images. Our key innovations include HDR lighting restoration, CRF estimation, surface material estimation, and alternating optimization to recover all scene properties accurately. IRIS estimates accurate surface material, spatially varying HDR illumination, and CRF from real-world LDR images. As a result, IRIS enables photorealistic rendering under diverse lighting and from arbitrary viewpoints, showcasing the significant potential for content creation applications. + +# References + +[1] Dejan Azinovic, Tzu-Mao Li, Anton Kaplanyan, and Matthias Nießner. Inverse path tracing for joint material and lighting estimation. In CVPR, 2019. 2, 4 +[2] Sai Bi, Zexiang Xu, Pratul Srinivasan, Ben Mildenhall, Kalyan Sunkavalli, Miloš Hasan, Yannick Hold-Geoffroy, David Kriegman, and Ravi Ramamoorthi. Neural reflectance fields for appearance acquisition. arXiv:2008.03824, 2020. 2 +[3] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2022. 4 +[4] Robert L Cook and Kenneth E. Torrance. A reflectance model for computer graphics. ACM Transactions on Graphics (ToG), 1982. 3 +[5] Paul E. Debevec and Jitendra Malik. Recovering high dynamic range radiance maps from photographs. In SIGGRAPH, 1997. 3 +[6] Valentin Deschaintre, Miika Aittala, Fredo Durand, George Drettakis, and Adrien Bousseau. Single-image SVBRDF capture with a rendering-aware deep network. ACM Transactions on Graphics (ToG), 37(4):128:1-15, 2018. 2 +[7] Ainaz Eftekhar, Alexander Sax, Jitendra Malik, and Amir Zamir. Omnidata: A scalable pipeline for making multi-task mid-level vision datasets from 3D scans. In ICCV, 2021. 4 +[8] Marc-Andre Gardner, Kalyan Sunkavalli, Ersin Yumer, Xiaohui Shen, Emiliano Gambaretto, Christian Gagné, and Jean-François Lalonde. Learning to predict indoor illumination from a single image. ACM Transactions on Graphics, 36(6): 176:1-14, 2017. 2 +[9] Dan B Goldman, Brian Curless, Aaron Hertzmann, and Steven M Seitz. Shape and spatially-varying BRDFs from photometric stereo. IEEE Transactions on Pattern Analysis and Machine Intelligence, 32(6):1060-1071, 2009. 2 +[10] Michael D Grossberg and Shree K Nayar. What is the space of camera response functions? In CVPR, 2003. 8 +[11] Michael D. Grossberg and Shree K. Nayar. Modeling the space of camera response functions. IEEE TPAMI, 26(10): 1272-1282, 2004. 3 +[12] Jon Hasselgren, Nikolai Hofmann, and Jacob Munkberg. Shape, light & material decomposition from images using monte carlo rendering and denoising. NeurIPS, 2022. 2 +[13] Xin Huang, Qi Zhang, Ying Feng, Hongdong Li, Xuan Wang, and Qing Wang. HDR-NeRF: High dynamic range neural radiance fields. In CVPR, 2022. 3 +[14] Wenzel Jakob, Sébastien Speierer, Nicolas Roussel, Merlin Nimier-David, Delio Vicini, Tizian Zeltner, Baptiste Nicolet, Miguel Crespo, Vincent Leroy, and Ziyi Zhang. Mitsuba 3 renderer, 2022. https://mitsuba-renderer.org. 2 +[15] Haian Jin, Isabella Liu, Peijia Xu, Xiaoshuai Zhang, Songfang Han, Sai Bi, Xiaowei Zhou, Zexiang Xu, and Hao Su. TensoIR: Tensorial inverse rendering. CVPR, 2023. 2, 3 +[16] James T Kajiya. The rendering equation. In SIGGRAPH, 1986. 4 +[17] Jaroslav Krivánek and Pascal Gautron. Practical global illumination with irradiance caching. Springer, 2022. 4 + +[18] Eric Lafortune. Mathematical models and Monte Carlo algorithms for physically based rendering. PhD thesis, Katholieke Universiteit Leuven, 1996. 4 +[19] Xiao Li, Yue Dong, Pieter Peers, and Xin Tong. Modeling surface appearance from a single photograph using self-augmented convolutional neural networks. ACM Transactions on Graphics (ToG), 36(4):45:1-11, 2017. 2 +[20] Zhengqi Li and Noah Snavely. CGIntrinsics: Better intrinsic image decomposition through physically-based rendering. In ECCV, 2018. 2 +[21] Zhengqi Li and Noah Snavely. Learning intrinsic image decomposition from watching the world. In CVPR, 2018. 2 +[22] Zhengqin Li, Kalyan Sunkavalli, and Manmohan Chandraker. Materials for masses: SVBRDF acquisition with a single mobile phone image. In ECCV, 2018. 2 +[23] Zhengqin Li, Mohammad Shafiei, Ravi Ramamoorthi, Kalyan Sunkavalli, and Manmohan Chandraker. Inverse rendering for complex indoor scenes: Shape, spatially-varying lighting and SVBRDF from a single image. In CVPR, 2020. 2 +[24] Zhengqin Li, Ting-Wei Yu, Shen Sang, Sarah Wang, Meng Song, Yuhan Liu, Yu-Ying Yeh, Rui Zhu, Nitesh Gundavarapu, Jia Shi, et al. OpenRooms: An open framework for photorealistic indoor scene datasets. In CVPR, 2021. 2 +[25] Zhengqin Li, Jia Shi, Sai Bi, Rui Zhu, Kalyan Sunkavalli, Miloš Hasan, Zexiang Xu, Ravi Ramamoorthi, and Manmohan Chandraker. Physically-based editing of indoor scene lighting from a single image. In ECCV, 2022. 2, 6, 8 +[26] Zhi-Hao Lin, Bohan Liu, Yi-Ting Chen, David Forsyth, Jia-Bin Huang, Anand Bhattachad, and Shenlong Wang. UrbanIR: Large-scale urban scene inverse rendering from a single video. 3DV, 2025. 2 +[27] Yu-Lun Liu, Wei-Sheng Lai, Yu-Sheng Chen, Yi-Lung Kao, Ming-Hsuan Yang, Yung-Yu Chuang, and Jia-Bin Huang. Single-image HDR reconstruction by learning to reverse the camera pipeline. In CVPR, 2020. 3 +[28] Demetris Marnerides, Thomas Bashford-Rogers, Jonathan Hatchett, and Kurt Debattista. ExpandNet: A deep convolutional neural network for high dynamic range expansion from low dynamic range content. Computer Graphics Forum, 2018. 3 +[29] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 6 +[30] Ben Mildenhall, Peter Hedman, Ricardo Martin-Brualla, Pratul Srinivasan, and Jonathan T. Barron. NeRF in the dark: High dynamic range view synthesis from noisy raw images. In CVPR, 2022. 3 +[31] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM TOG, 2022. 3 +[32] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting triangular 3D models, materials, and lighting from images. In CVPR, 2022. 2 +[33] Merlin Nimier-David, Zhao Dong, Wenzel Jakob, and Anton Kaplanyan. Material and lighting reconstruction for complex + +indoor scenes with texture-space differentiable rendering. In Eurographics Symposium on Rendering, 2021. 2, 4 +[34] Julien Philip, Sébastien Morgenthaler, Michael Gharbi, and George Drettakis. Free-viewpoint indoor neural relighting from multi-view stereo. ACM Trans. Graph., 40(5):194:1-18, 2021. 2 +[35] Erik Reinhard, Greg Ward, Sumanta Pattanaik, Paul Debevec, Wolfgang Heidrich, and Karol Myszkowski. High Dynamic Range Imaging - Acquisition, Display and Image-Based Lighting. Morgan Kaufmann, 2nd edition, 2010. 1 +[36] Mike Roberts, Jason Ramapuram, Anurag Ranjan, Atulit Kumar, Miguel Angel Bautista, Nathan Paczan, Russ Webb, and Joshua M. Susskind. Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding. In ICCV, 2021. 2 +[37] Dario Seyb, Peter-Pike Sloan, Ari Silvennoinen, Michal Iwanicki, and Wojciech Jarosz. The design and evolution of the UberBake light baking system. ACM Transactions on Graphics (TOG), 2020. 4 +[38] Pratul P Srinivasan, Ben Mildenhall, Matthew Tancik, Jonathan T Barron, Richard Tucker, and Noah Snavely. Lighthouse: Predicting lighting volumes for spatially-coherent illumination. In CVPR, 2020. 2 +[39] Pratul P. Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T. Barron. NeRV: Neural reflectance and visibility fields for relighting and view synthesis. In CVPR, 2021. 3 +[40] Cheng Sun, Guangyan Cai, Zhengqin Li, Kai Yan, Cheng Zhang, Carl Marshall, Jia-Bin Huang, Shuang Zhao, and Zhao Dong. Neural-pbir reconstruction of shape, material, and illumination. In ICCV, 2023. 2 +[41] Zian Wang, Jonah Philion, Sanja Fidler, and Jan Kautz. Learning indoor inverse rendering with 3D spatially-varying lighting. In ICCV, 2021. 4 +[42] Zian Wang, Tianchang Shen, Jun Gao, Shengyu Huang, Jacob Munkberg, Jon Hasselgren, Zan Gojcic, Wenzheng Chen, and Sanja Fidler. Neural fields meet explicit geometric representations for inverse rendering of urban scenes. In CVPR, 2023. 2, 3 +[43] Liwen Wu, Rui Zhu, Mustafa B Yaldiz, Yinhao Zhu, Hong Cai, Janarbek Matai, Fatih Porikli, Tzu-Mao Li, Manmohan Chandraker, and Ravi Ramamoorthi. Factorized inverse path tracing for efficient and accurate material-lighting estimation. In ICCV, 2023. 2, 3, 4, 6, 7, 8 +[44] Xin Yang, Ke Xu, Yibing Song, Qiang Zhang, Xiaopeng Wei, and Rynson WH Lau. Image correction via deep reciprocating HDR transformation. In CVPR, 2018. 3 +[45] Yao Yao, Jingyang Zhang, Jingbo Liu, Yihang Qu, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. NeILF: Neural incident light field for physically-based material estimation. In ECCV, 2022. 2, 6, 7, 8 +[46] Lior Yariv, Peter Hedman, Christian Reiser, Dor Verbin, Pratul P Srinivasan, Richard Szeliski, Jonathan T Barron, and Ben Mildenhall. Bakedsdf: Meshing neural sdfs for real-time view synthesis. In ACM SIGGRAPH computer graphics, 2023. 4 + +[47] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. ScanNet++: A high-fidelity dataset of 3D indoor scenes. In ICCV, 2023. 6, 7 +[48] Bohan Yu, Siqi Yang, Xuanning Cui, Siyan Dong, Baoquan Chen, and Boxin Shi. MILO: Multi-bounce inverse rendering for indoor scene with light-emitting objects. IEEE TPAMI, 2023. 2, 4 +[49] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. MonoSDF: Exploring monocular geometric cues for neural implicit surface reconstruction. In NeurIPS, 2022. 4, 5 +[50] Zheng Zeng, Valentin Deschaintre, Iliyan Georgiev, Yannick Hold-Geoffroy, Yiwei Hu, Fujun Luan, Ling-Qi Yan, and Miloš Hasan. RGB ↔ X: Image decomposition and synthesis using material- and lighting-aware diffusion models. In SIGGRAPH, 2024. 2 +[51] Jinsong Zhang and Jean-François Lalonde. Learning high dynamic range from outdoor panoramicas. In ICCV, 2017. 3 +[52] Jingyang Zhang, Yao Yao, Shiwei Li, Jingbo Liu, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. NeILF++: Inter-reflectable light fields for geometry and material estimation. In ICCV, 2023. 2, 3 +[53] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snavely. PhySG: Inverse rendering with spherical Gaussians for physics-based material editing and relighting. In CVPR, 2021. 2, 3 +[54] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. NeR-Factor: Neural factorization of shape and reflectance under an unknown illumination. ACM TOG, 2021. 2 +[55] Jingsen Zhu, Yuchi Huo, Qi Ye, Fujun Luan, Jifan Li, Dianbing Xi, Lisha Wang, Rui Tang, Wei Hua, Hujun Bao, and Rui Wang. I²-SDF: Intrinsic indoor scene reconstruction and editing via raytracing in neural SDFs. In CVPR, 2023. 2, 4 +[56] Rui Zhu, Zhengqin Li, Janarbek Matai, Fatih Porikli, and Manmohan Chandraker. IRISformer: Dense vision transformers for single-image inverse rendering in indoor scenes. In CVPR, 2022. 2, 4 \ No newline at end of file diff --git a/2025/IRIS_ Inverse Rendering of Indoor Scenes from Low Dynamic Range Images/images.zip b/2025/IRIS_ Inverse Rendering of Indoor Scenes from Low Dynamic Range Images/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d8513dc9d38f03e4ff3d0aa5189ac8c7c3490b1d --- /dev/null +++ b/2025/IRIS_ Inverse Rendering of Indoor Scenes from Low Dynamic Range Images/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ec5b6ecc3b0badb6110056e00fbd9c00f605b1299c6e44e84d7b8ff69c159121 +size 599558 diff --git a/2025/IRIS_ Inverse Rendering of Indoor Scenes from Low Dynamic Range Images/layout.json b/2025/IRIS_ Inverse Rendering of Indoor Scenes from Low Dynamic Range Images/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ee7cc8a1cdd03a8a9dbe6beeea94a06376ef7521 --- /dev/null +++ b/2025/IRIS_ Inverse Rendering of Indoor Scenes from Low Dynamic Range Images/layout.json @@ -0,0 +1,9190 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 69, + 103, + 541, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 103, + 541, + 121 + ], + "spans": [ + { + "bbox": [ + 69, + 103, + 541, + 121 + ], + "type": "text", + "content": "IRIS: Inverse Rendering of Indoor Scenes from Low Dynamic Range Images" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 85, + 142, + 167, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 142, + 167, + 156 + ], + "spans": [ + { + "bbox": [ + 85, + 142, + 167, + 156 + ], + "type": "text", + "content": "Chih-Hao Lin1,2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 179, + 143, + 262, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 143, + 262, + 157 + ], + "spans": [ + { + "bbox": [ + 179, + 143, + 262, + 157 + ], + "type": "text", + "content": "Jia-Bin Huang" + }, + { + "bbox": [ + 179, + 143, + 262, + 157 + ], + "type": "inline_equation", + "content": "^{1,3}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 275, + 143, + 340, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 275, + 143, + 340, + 157 + ], + "spans": [ + { + "bbox": [ + 275, + 143, + 340, + 157 + ], + "type": "text", + "content": "Zhengqin Li" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 354, + 143, + 414, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 143, + 414, + 157 + ], + "spans": [ + { + "bbox": [ + 354, + 143, + 414, + 157 + ], + "type": "text", + "content": "Zhao Dong1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 429, + 143, + 522, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 429, + 143, + 522, + 156 + ], + "spans": [ + { + "bbox": [ + 429, + 143, + 522, + 156 + ], + "type": "text", + "content": "Christian Richardt1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 96, + 158, + 150, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 158, + 150, + 171 + ], + "spans": [ + { + "bbox": [ + 96, + 158, + 150, + 171 + ], + "type": "text", + "content": "Tuotuo Li" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 164, + 158, + 257, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 158, + 257, + 171 + ], + "spans": [ + { + "bbox": [ + 164, + 158, + 257, + 171 + ], + "type": "text", + "content": "Michael Zollhöfer1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 271, + 158, + 347, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 271, + 158, + 347, + 171 + ], + "spans": [ + { + "bbox": [ + 271, + 158, + 347, + 171 + ], + "type": "text", + "content": "Johannes Kopf1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 361, + 158, + 443, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 158, + 443, + 171 + ], + "spans": [ + { + "bbox": [ + 361, + 158, + 443, + 171 + ], + "type": "text", + "content": "Shenlong Wang" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 457, + 158, + 525, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 457, + 158, + 525, + 171 + ], + "spans": [ + { + "bbox": [ + 457, + 158, + 525, + 171 + ], + "type": "text", + "content": "Changil Kim" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 82, + 176, + 528, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 176, + 528, + 213 + ], + "spans": [ + { + "bbox": [ + 82, + 176, + 528, + 213 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 82, + 176, + 528, + 213 + ], + "type": "text", + "content": "Meta " + }, + { + "bbox": [ + 82, + 176, + 528, + 213 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 82, + 176, + 528, + 213 + ], + "type": "text", + "content": "University of Illinois Urbana-Champaign " + }, + { + "bbox": [ + 82, + 176, + 528, + 213 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 82, + 176, + 528, + 213 + ], + "type": "text", + "content": "University of Maryland College Park https://irisldr.github.io/" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 58, + 227, + 144, + 415 + ], + "blocks": [ + { + "bbox": [ + 58, + 227, + 144, + 415 + ], + "lines": [ + { + "bbox": [ + 58, + 227, + 144, + 415 + ], + "spans": [ + { + "bbox": [ + 58, + 227, + 144, + 415 + ], + "type": "image", + "image_path": "a25f7bdb278f3a437c12f874cec517b605582adcde0b705c67620274d7cdb59c.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 421, + 555, + 455 + ], + "lines": [ + { + "bbox": [ + 55, + 421, + 555, + 455 + ], + "spans": [ + { + "bbox": [ + 55, + 421, + 555, + 455 + ], + "type": "text", + "content": "Figure 1. Physically-based inverse rendering from LDR images. Our method takes LDR input images and estimates high-quality spatially-varying HDR lighting, physically-based material properties, and a camera response function (CRF). This decomposition allows us to insert objects or new light sources, and to relight the scene in a physically accurate manner." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 151, + 228, + 291, + 415 + ], + "blocks": [ + { + "bbox": [ + 151, + 228, + 291, + 415 + ], + "lines": [ + { + "bbox": [ + 151, + 228, + 291, + 415 + ], + "spans": [ + { + "bbox": [ + 151, + 228, + 291, + 415 + ], + "type": "image", + "image_path": "45a2c0323811594e2d8734d9a8db03655eac18ee26fb98076cda118ad6bd52c0.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 299, + 228, + 423, + 415 + ], + "blocks": [ + { + "bbox": [ + 299, + 228, + 423, + 415 + ], + "lines": [ + { + "bbox": [ + 299, + 228, + 423, + 415 + ], + "spans": [ + { + "bbox": [ + 299, + 228, + 423, + 415 + ], + "type": "image", + "image_path": "1f18b3f222474e7922aeede920737cf91aeaafc49426652883f4a83acc5ab410.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 430, + 228, + 553, + 415 + ], + "blocks": [ + { + "bbox": [ + 430, + 228, + 553, + 415 + ], + "lines": [ + { + "bbox": [ + 430, + 228, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 430, + 228, + 553, + 415 + ], + "type": "image", + "image_path": "1d57008291cd2f4717b0bfdee3c4d4972004e91e749557bd53408542007a3a20.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 457, + 394, + 469 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 457, + 394, + 469 + ], + "spans": [ + { + "bbox": [ + 315, + 457, + 394, + 469 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 151, + 481, + 200, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 481, + 200, + 493 + ], + "spans": [ + { + "bbox": [ + 151, + 481, + 200, + 493 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 508, + 297, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 508, + 297, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 508, + 297, + 700 + ], + "type": "text", + "content": "Inverse rendering seeks to recover 3D geometry, surface material, and lighting from captured images, enabling advanced applications such as novel-view synthesis, relighting, and virtual object insertion. However, most existing techniques rely on high dynamic range (HDR) images as input, limiting accessibility for general users. In response, we introduce IRIS, an inverse rendering framework that recovers the physically based material, spatially-varying HDR lighting, and camera response functions from multi-view, low-dynamic-range (LDR) images. By eliminating the dependence on HDR input, we make inverse rendering technology more accessible. We evaluate our approach on real-world and synthetic scenes and compare it with state-of-the-art methods. Our results show that IRIS effectively recovers HDR lighting, accurate material, and plausible camera response functions, supporting photorealistic relighting and object insertion." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 482, + 555, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 482, + 555, + 602 + ], + "spans": [ + { + "bbox": [ + 313, + 482, + 555, + 602 + ], + "type": "text", + "content": "Physically based inverse rendering enables reconstructing realistic material properties and lighting in scenes. This decomposition is valuable for various applications, including relighting, material editing, and realistic object insertion. However, most existing inverse rendering methods require input images with high dynamic range (HDR) to capture the full light transport in the scene. This poses a significant barrier to broader adoption, as HDR capture typically requires specialized hardware or merging multiple aligned LDR images through exposure bracketing [35]." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 605, + 556, + 715 + ], + "type": "text", + "content": "Most common imaging sensors do not capture sufficient dynamic range for scenes' light-emitting and dark regions. Moreover, cameras often convert raw sensor readings to 8-bit low-dynamic-range (LDR) images for storage and transmission. The non-linear mapping and quantization lead to the additional loss of lighting information. The complex light transport in indoor environments makes reconstructing the original HDR lighting - critical for faithful inverse rendering - even more challenging. If material and lighting estimation" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "465" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "content": "could be achieved from a \"casual\" capture using standard devices like cameras or phones, inverse rendering would become far more accessible to a wider range of users." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 110, + 296, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 110, + 296, + 241 + ], + "spans": [ + { + "bbox": [ + 55, + 110, + 296, + 241 + ], + "type": "text", + "content": "Many state-of-the-art inverse rendering methods rely on HDR inputs to accurately recover material properties and lighting [43, 48, 52]. Some approaches attempt to overcome this issue by taking LDR images as inputs and solving for the lighting. However, these methods often make simplifying assumptions, such as infinitely distant light sources [42], requiring additional inputs like emitter masks [55], or neglecting multi-bounce light transport [45, 52, 55]. As a result, these methods struggle to handle the complex light transport in indoor scenes and often fail to produce high-quality surface material and lighting." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 243, + 296, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 243, + 296, + 386 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 296, + 386 + ], + "type": "text", + "content": "To address these challenges, we introduce IRIS, an inverse rendering method for indoor scenes, using only multi-view LDR images. IRIS models tone mapping (i.e., HDR-to-LDR conversion), allowing us to work directly with LDR inputs. IRIS automatically identifies emitters and reconstructs spatially varying HDR lighting through physically based inverse rendering, which is crucial for faithful material estimation. Jointly estimating lighting, surface material, and camera response function (CRF) leads to unstable optimization due to ambiguities. To tackle this, we design a novel optimization strategy that effectively resolves these ambiguities, enabling high-quality estimation. Our core contributions are:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 387, + 296, + 519 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 56, + 387, + 296, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 387, + 296, + 434 + ], + "spans": [ + { + "bbox": [ + 56, + 387, + 296, + 434 + ], + "type": "text", + "content": "1. IRIS faithfully estimates spatially-varying HDR lighting, physically-based materials, and camera response function from LDR images and outperforms several state-of-the-art inverse rendering methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 435, + 295, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 435, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 435, + 295, + 483 + ], + "type": "text", + "content": "2. We explicitly model LDR image formation in our (inverse) rendering pipeline so that LDR images can be used directly, broadening the accessibility of high-quality inverse rendering." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 483, + 295, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 295, + 519 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 295, + 519 + ], + "type": "text", + "content": "3. Finally, IRIS is extensively evaluated on synthetic and real-world scenes, demonstrating diverse and realistic view synthesis, relighting, and object insertion." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 533, + 142, + 546 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 142, + 546 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 142, + 546 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": "Data-driven inverse rendering. Inverse rendering seeks to reconstruct a scene's geometry, material, and lighting from single or multiple images. This process is notoriously challenging and ill-posed due to the inherent ambiguity. Numerous methods employ deep priors learned from large-scale datasets for different tasks, including intrinsic image decomposition [20, 21], SVBRDF estimation [6, 19, 22, 50, 56], lighting estimation [8, 23, 38], lighting editing [25] and relighting [34]. These learning-based methods take a single or a few images [2, 38] as inputs, reducing capturing requirements of classical measurement-based methods [9]. However, high-quality datasets are essential for data-driven methods to achieve compelling generalization ability. Synthetic" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 339, + 70, + 436, + 129 + ], + "blocks": [ + { + "bbox": [ + 339, + 70, + 436, + 129 + ], + "lines": [ + { + "bbox": [ + 339, + 70, + 436, + 129 + ], + "spans": [ + { + "bbox": [ + 339, + 70, + 436, + 129 + ], + "type": "image", + "image_path": "623c122cc4fd7ca90e59fd26643fa3100c61d5cc29df715dd759e542f954791f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 369, + 129, + 416, + 134 + ], + "lines": [ + { + "bbox": [ + 369, + 129, + 416, + 134 + ], + "spans": [ + { + "bbox": [ + 369, + 129, + 416, + 134 + ], + "type": "text", + "content": "Surface material" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 442, + 70, + 529, + 129 + ], + "blocks": [ + { + "bbox": [ + 442, + 70, + 529, + 129 + ], + "lines": [ + { + "bbox": [ + 442, + 70, + 529, + 129 + ], + "spans": [ + { + "bbox": [ + 442, + 70, + 529, + 129 + ], + "type": "image", + "image_path": "f4355db530dab1e2421567746725b4d6766152de6687f53fc9f26d04170ff64a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 473, + 129, + 498, + 135 + ], + "lines": [ + { + "bbox": [ + 473, + 129, + 498, + 135 + ], + "spans": [ + { + "bbox": [ + 473, + 129, + 498, + 135 + ], + "type": "text", + "content": "Lighting" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 339, + 134, + 436, + 193 + ], + "blocks": [ + { + "bbox": [ + 339, + 134, + 436, + 193 + ], + "lines": [ + { + "bbox": [ + 339, + 134, + 436, + 193 + ], + "spans": [ + { + "bbox": [ + 339, + 134, + 436, + 193 + ], + "type": "image", + "image_path": "4f2cebbd55fa7f3e052cf7c13277e25cf9d0781996e466523c66e457c41637e3.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 369, + 194, + 416, + 201 + ], + "lines": [ + { + "bbox": [ + 369, + 194, + 416, + 201 + ], + "spans": [ + { + "bbox": [ + 369, + 194, + 416, + 201 + ], + "type": "text", + "content": "Surface material" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 201, + 555, + 266 + ], + "lines": [ + { + "bbox": [ + 313, + 201, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 201, + 555, + 266 + ], + "type": "text", + "content": "Figure 2. Limitation of SOTA. A typical image formation process causes the loss of lighting information, posing challenges in inverse rendering. FIPT [43] assumes HDR input and NeILF [45] ignores multi-bounce light transport. Both methods fail to estimate accurate material (red boxes) and HDR lighting (yellow boxes). We demonstrate significantly better results (Figure 1)." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 442, + 134, + 528, + 193 + ], + "blocks": [ + { + "bbox": [ + 442, + 134, + 528, + 193 + ], + "lines": [ + { + "bbox": [ + 442, + 134, + 528, + 193 + ], + "spans": [ + { + "bbox": [ + 442, + 134, + 528, + 193 + ], + "type": "image", + "image_path": "ead907e3468f92182f191e4607671aa13abe8c1f98af1d41a665f260bbf4a536.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 473, + 194, + 498, + 201 + ], + "lines": [ + { + "bbox": [ + 473, + 194, + 498, + 201 + ], + "spans": [ + { + "bbox": [ + 473, + 194, + 498, + 201 + ], + "type": "text", + "content": "Lighting" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 277, + 556, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 556, + 421 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 556, + 421 + ], + "type": "text", + "content": "datasets are often used for training because ground-truth labeling is challenging to obtain in real environments. Despite efforts to create photorealistic synthetic datasets [20, 24, 36], the inherent diversity of real-world environments, especially in complex indoor scenes, still causes a substantial domain gap. Furthermore, these methods do not reconstruct complete 3D scenes and fail to render consistent novel views under novel illuminations. Our method utilizes prediction from a data-driven method [56] for initialization and reconstructs 3D geometry, material and lighting with a carefully designed optimization framework, demonstrating superior performance for real-world scenes." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "text", + "content": "Optimization-based inverse rendering. Differentiable path tracing [14] enables the direct optimization of scene parameters with physically-based multi-bounce light transport. Recent methods, including MILO [48], IPT [1], and follow-up work [33] exploit this to jointly optimize BRDF and lighting emission, accounting for global light transport. FIPT [43] utilizes a factored light transport formulation for BRDF estimation and emitter detection. Nevertheless, these methods assume piecewise constant materials or are designed to take high-dynamic range images as input, and the robustness is limited in real-world scenes. Several recent NeRF-based approaches [12, 15, 26, 32, 40, 42, 53, 54] parameterize materials as a neural field and optimize with a volume rendering formulation. They tend to assume distant lighting and cannot handle spatially varying lighting in indoor environments well. NeILF [45] and NeILF++ [52] represent spatially-varying lighting as a neural field, " + }, + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{I}^2" + }, + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "text", + "content": "-SDF [55] constructs lighting from given emitter masks. While handling more complex lighting, these works require additional assumptions (e.g., HDR input, emitter masks), and single-bounce ray tracing cannot model complex light transport. In contrast, our method automatically identifies emitters and reconstructs HDR lighting from casual (LDR) images, while achieving high-quality material and lighting estimation that" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "466" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 254, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 254, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 254, + 83 + ], + "type": "text", + "content": "outperforms prior work in this practical scenario." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 87, + 297, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 87, + 297, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 87, + 297, + 327 + ], + "type": "text", + "content": "CRF and HDR estimation. The complete information about the illumination is usually lost within the photography pipeline. Estimating the HDR lighting and CRF enables us to edit photos and achieve realistic object insertion. Debevec and Malik [5] propose to take multiple photos with varying exposure levels to recover the CRF and HDR image. Some previous works [27, 28, 44, 51] predict HDR images with a data-driven approach by learning the LDR-HDR mapping from a single image. However, these methods rely on HDR data for training, and the estimation is inconsistent across views and scenes. To handle the HDR-LDR mapping, FIPT [43] and RawNeRF [30] assume a known gamma correction function; NeILF++ [52] learn a single gamma correction parameter; HDR-NeRF [13] parameterizes tone-mapping function with an MLP. However, these works are designed to take HDR images or LDR with multiple exposure levels as input. We leverage physically-based rendering to recover spatially varying BRDF, HDR illumination, and CRF from LDR images. Our method generates realistic view synthesis, relighting, and object insertion." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 338, + 253, + 351 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 338, + 253, + 351 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 253, + 351 + ], + "type": "text", + "content": "3. Background: LDR image formation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 357, + 296, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 357, + 296, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 296, + 514 + ], + "type": "text", + "content": "Real-world radiance exhibits high dynamic range (HDR) and contains bright light sources and dark regions. To preserve the complete lighting information, HDR images are typically produced with a high-end camera, laborious exposure bracketing [5], and are stored in specialized format (16 or 32 bits per color channel). However, capturing HDR is often impractical for casual users and is typically absent in standard photography assets. Due to the limitations of standard sensors and displays, the radiance is frequently processed with multiple stages to produce low dynamic range (LDR) images stored in more lightweight formats (such as JPEG or PNG at 8 bits and RAW at 12 bits per color channel). We describe the major steps in this process below." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 516, + 296, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 516, + 296, + 589 + ], + "spans": [ + { + "bbox": [ + 55, + 516, + 296, + 589 + ], + "type": "text", + "content": "Dynamic range clipping. Given the scene irradiance " + }, + { + "bbox": [ + 55, + 516, + 296, + 589 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 55, + 516, + 296, + 589 + ], + "type": "text", + "content": " and exposure time " + }, + { + "bbox": [ + 55, + 516, + 296, + 589 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 55, + 516, + 296, + 589 + ], + "type": "text", + "content": ", the observed radiance is clipped at a certain maximum value due to the sensor limitation, which could be formulated as: " + }, + { + "bbox": [ + 55, + 516, + 296, + 589 + ], + "type": "inline_equation", + "content": "Z_{c} = \\min (E\\Delta t,1)" + }, + { + "bbox": [ + 55, + 516, + 296, + 589 + ], + "type": "text", + "content": ". Due to the clipping operation, the lighting information is lost for overexposed/saturated pixels." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 590, + 297, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 297, + 674 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 297, + 674 + ], + "type": "text", + "content": "Non-linear mapping of CRF. To enhance the image's visual quality and match the human perception of the scene, the clipped intensity is further transformed with a non-linear camera response function (CRF) to produce LDR pixel values: " + }, + { + "bbox": [ + 55, + 590, + 297, + 674 + ], + "type": "inline_equation", + "content": "Z = \\mathrm{CRF}(Z_c) = \\mathrm{CRF}(\\min(E\\Delta t, 1))" + }, + { + "bbox": [ + 55, + 590, + 297, + 674 + ], + "type": "text", + "content": ". Please refer to Debevec and Malik [5] for a more detailed imaging pipeline explanation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 297, + 714 + ], + "type": "text", + "content": "Challenges in inverse rendering. Real-world scenes often exhibit extreme variances in radiance, making it challenging to capture these scenes in LDR images without under" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 556, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 556, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 556, + 133 + ], + "type": "text", + "content": "exposing or saturating them. Due to the clipping and nonlinear CRF mapping, critical lighting information is lost, posing challenges in previous works, illustrated in Figure 2. In this work, we propose an inverse rendering framework that jointly recovers HDR lighting and CRF from LDR images." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 142, + 484, + 156 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 142, + 484, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 142, + 484, + 156 + ], + "type": "text", + "content": "4. Representation and rendering" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 164, + 556, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 164, + 556, + 284 + ], + "spans": [ + { + "bbox": [ + 313, + 164, + 556, + 284 + ], + "type": "text", + "content": "Camera response function (CRF). The CRF is crucial for recovering original radiance from 8-bit LDR images. However, the CRF varies with each camera and is often unknown to users. Thus, deducing the CRF from LDR image observations is critical to our solution. We use Gross et al.'s [11] empirical model of response (EMoR) for CRF modeling [27]. They collected a database of 201 real-world CRFs of various devices and parameterized them as vectors. A mean curve " + }, + { + "bbox": [ + 313, + 164, + 556, + 284 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{g}}" + }, + { + "bbox": [ + 313, + 164, + 556, + 284 + ], + "type": "text", + "content": " and PCA basis " + }, + { + "bbox": [ + 313, + 164, + 556, + 284 + ], + "type": "inline_equation", + "content": "\\mathbf{g}_b" + }, + { + "bbox": [ + 313, + 164, + 556, + 284 + ], + "type": "text", + "content": " are calculated from the database, and we parameterize a learnable CRF as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 373, + 289, + 555, + 304 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 289, + 555, + 304 + ], + "spans": [ + { + "bbox": [ + 373, + 289, + 555, + 304 + ], + "type": "interline_equation", + "content": "\\mathbf {g} = \\bar {\\mathbf {g}} + \\Sigma_ {b} w _ {b} \\mathbf {g} _ {b}, \\mathbf {g} \\in \\mathbb {R} ^ {1 0 2 4}, \\tag {1}", + "image_path": "1695f736779157d304ed5bdf004c4df81c256228b0e9d989d3bfb22c6fbbfcec.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 310, + 556, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 310, + 556, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 310, + 556, + 407 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 310, + 556, + 407 + ], + "type": "inline_equation", + "content": "\\{w_{b}\\}" + }, + { + "bbox": [ + 313, + 310, + 556, + 407 + ], + "type": "text", + "content": " weigh each basis curve and are learnable parameters. The non-linear function is discretized as " + }, + { + "bbox": [ + 313, + 310, + 556, + 407 + ], + "type": "inline_equation", + "content": "\\mathrm{CRF}(x) = \\mathrm{LERP}(\\{g_k\\}_k, x)" + }, + { + "bbox": [ + 313, + 310, + 556, + 407 + ], + "type": "text", + "content": ", which linearly interpolates the uniformly sampled CRF " + }, + { + "bbox": [ + 313, + 310, + 556, + 407 + ], + "type": "inline_equation", + "content": "\\{g_k\\}_{k=0}^{1023} \\in [0,1]^{1024}" + }, + { + "bbox": [ + 313, + 310, + 556, + 407 + ], + "type": "text", + "content": ", independently per color channel. We make this design choice because it covers most of the CRF space, and the simplicity makes it easy to enforce regularization, which brings huge advantage in the complex optimization of inverse rendering." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 408, + 556, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 408, + 556, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 408, + 556, + 601 + ], + "type": "text", + "content": "Spatially-varying lighting. While environment maps handle object-centric and outdoor scenes well [15, 39, 42, 53], they cannot capture complex spatially-varying lighting of indoor scenes. To handle this, we define direct lighting on the scene mesh and measure emission as view-independent radiance: " + }, + { + "bbox": [ + 313, + 408, + 556, + 601 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{\\mathrm{e}}(\\mathbf{x}) \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 408, + 556, + 601 + ], + "type": "text", + "content": ". The parameterization of " + }, + { + "bbox": [ + 313, + 408, + 556, + 601 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{\\mathrm{e}}" + }, + { + "bbox": [ + 313, + 408, + 556, + 601 + ], + "type": "text", + "content": " is detailed in our supplementary material. We also define an emitter mask " + }, + { + "bbox": [ + 313, + 408, + 556, + 601 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{e}}(\\mathbf{x}) \\in \\{0,1\\}" + }, + { + "bbox": [ + 313, + 408, + 556, + 601 + ], + "type": "text", + "content": " to denote whether a surface point " + }, + { + "bbox": [ + 313, + 408, + 556, + 601 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 408, + 556, + 601 + ], + "type": "text", + "content": " is on an emitter. Global illumination is represented as a view-independent surface light field (SLF): " + }, + { + "bbox": [ + 313, + 408, + 556, + 601 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{\\mathrm{SLF}}(\\mathbf{x}) \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 408, + 556, + 601 + ], + "type": "text", + "content": ". The sRGB pixel colors of input LDR images are linearized with an inverse CRF mapping: " + }, + { + "bbox": [ + 313, + 408, + 556, + 601 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{linear}} = \\mathrm{CRF}^{-1}(\\mathbf{I}_{\\mathrm{LDR}})" + }, + { + "bbox": [ + 313, + 408, + 556, + 601 + ], + "type": "text", + "content": ". If a voxel contains no surface, it is set to zero; otherwise, it collects the average linearized radiances across all input images with ray casting. This greatly accelerates the approximation of global illumination without requiring multi-bounce ray tracing." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 602, + 556, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 602, + 556, + 687 + ], + "spans": [ + { + "bbox": [ + 313, + 602, + 556, + 687 + ], + "type": "text", + "content": "Material BRDF. We represent spatially-varying materials using the Cook-Torrance BRDF [4]. The surface albedo " + }, + { + "bbox": [ + 313, + 602, + 556, + 687 + ], + "type": "inline_equation", + "content": "\\mathbf{a}(\\mathbf{x})\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 602, + 556, + 687 + ], + "type": "text", + "content": ", roughness " + }, + { + "bbox": [ + 313, + 602, + 556, + 687 + ], + "type": "inline_equation", + "content": "\\sigma (\\mathbf{x})\\in \\mathbb{R}" + }, + { + "bbox": [ + 313, + 602, + 556, + 687 + ], + "type": "text", + "content": ", and metallicity " + }, + { + "bbox": [ + 313, + 602, + 556, + 687 + ], + "type": "inline_equation", + "content": "m(\\mathbf{x})\\in \\mathbb{R}" + }, + { + "bbox": [ + 313, + 602, + 556, + 687 + ], + "type": "text", + "content": " model the diffuse reflectance " + }, + { + "bbox": [ + 313, + 602, + 556, + 687 + ], + "type": "inline_equation", + "content": "\\mathbf{k_d} = \\mathbf{a}\\cdot (1 - m)" + }, + { + "bbox": [ + 313, + 602, + 556, + 687 + ], + "type": "text", + "content": " and specular reflectance " + }, + { + "bbox": [ + 313, + 602, + 556, + 687 + ], + "type": "inline_equation", + "content": "\\mathbf{k_s} = 0.04\\cdot (1 - m)" + }, + { + "bbox": [ + 313, + 602, + 556, + 687 + ], + "type": "text", + "content": ", and are parameterized as a neural field " + }, + { + "bbox": [ + 313, + 602, + 556, + 687 + ], + "type": "inline_equation", + "content": "\\mathbf{f}\\colon \\mathbf{x}\\mapsto (\\mathbf{a},\\sigma ,m)" + }, + { + "bbox": [ + 313, + 602, + 556, + 687 + ], + "type": "text", + "content": ", implemented with multi-resolution hash encoding and an MLP [31]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Factorized light transport. To render images from scene geometry, material, and lighting information, We follow" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "467" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 72, + 553, + 238 + ], + "blocks": [ + { + "bbox": [ + 63, + 72, + 553, + 238 + ], + "lines": [ + { + "bbox": [ + 63, + 72, + 553, + 238 + ], + "spans": [ + { + "bbox": [ + 63, + 72, + 553, + 238 + ], + "type": "image", + "image_path": "390b3c22db5234603ac9224731083b54eecb757ed4d85522755cf096d9eff524.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 246, + 555, + 301 + ], + "lines": [ + { + "bbox": [ + 54, + 246, + 555, + 301 + ], + "spans": [ + { + "bbox": [ + 54, + 246, + 555, + 301 + ], + "type": "text", + "content": "Figure 3. Framework Overview. Given multi-view posed LDR images and a surface mesh, our inverse rendering pipeline is divided into two main stages. In the initialization stage, we initialize the BRDF (Section 5.1), extract a surface light field (Section 4), and estimate emitter geometry (Equation (3)). In the optimization stage, we first recover HDR radiance from the LDR input (Section 5.2), then bake shading maps (Section 5.3), and jointly optimize BRDF and CRF parameters (Section 5.4). The improved parameters are used to refine the emission again. These three steps are repeated until convergence." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 311, + 295, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 311, + 295, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 311, + 295, + 335 + ], + "type": "text", + "content": "the rendering equation [16] to model physically-based light transport for realistic rendering:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 342, + 296, + 366 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 342, + 296, + 366 + ], + "spans": [ + { + "bbox": [ + 59, + 342, + 296, + 366 + ], + "type": "interline_equation", + "content": "\\mathbf {L} _ {\\mathrm {o}} (\\mathbf {x}, \\omega_ {\\mathrm {o}}) = \\mathbf {L} _ {\\mathrm {e}} (\\mathbf {x}, \\omega_ {\\mathrm {o}}) + \\int_ {\\Omega^ {+}} \\mathbf {L} _ {\\mathrm {i}} (\\mathbf {x}, \\omega_ {\\mathrm {i}}) f (\\mathbf {x}, \\omega_ {\\mathrm {i}}, \\omega_ {\\mathrm {o}}) d \\omega_ {\\mathrm {i}}, \\tag {2}", + "image_path": "1eb3411bf24edd129b8fcbf28412695a88d0b7d2a64e071ddbf79dd9abc78b2d.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 372, + 296, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 372, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 372, + 296, + 529 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 372, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{\\mathrm{o}}" + }, + { + "bbox": [ + 55, + 372, + 296, + 529 + ], + "type": "text", + "content": " is the radiance observed along a ray at a 3D position " + }, + { + "bbox": [ + 55, + 372, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 372, + 296, + 529 + ], + "type": "text", + "content": " in direction " + }, + { + "bbox": [ + 55, + 372, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\omega_{\\mathrm{o}}" + }, + { + "bbox": [ + 55, + 372, + 296, + 529 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 372, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{\\mathrm{e}}" + }, + { + "bbox": [ + 55, + 372, + 296, + 529 + ], + "type": "text", + "content": " is the emission term, " + }, + { + "bbox": [ + 55, + 372, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{\\mathrm{r}} = \\int_{\\Omega^{+}}\\mathbf{L}_{\\mathrm{i}}(\\mathbf{x},\\omega_{\\mathrm{i}})f(\\mathbf{x},\\omega_{\\mathrm{i}},\\omega_{\\mathrm{o}})d\\omega_{\\mathrm{i}}" + }, + { + "bbox": [ + 55, + 372, + 296, + 529 + ], + "type": "text", + "content": " is the reflectance term, and " + }, + { + "bbox": [ + 55, + 372, + 296, + 529 + ], + "type": "inline_equation", + "content": "f(\\mathbf{x},\\omega_{\\mathrm{i}},\\omega_{\\mathrm{o}})" + }, + { + "bbox": [ + 55, + 372, + 296, + 529 + ], + "type": "text", + "content": " is the BRDF. The recursive nature of the rendering equation is approximated with Monte-Carlo path tracing [16, 18]. However, modeling multiple bounces is still computationally expensive and unstable in the inverse rendering process. To enhance the efficiency and robustness of the optimization, we follow FIPT [43] and adopt factorized light transport [17, 37, 41], which precomputes shading maps " + }, + { + "bbox": [ + 55, + 372, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{\\mathrm{d}}" + }, + { + "bbox": [ + 55, + 372, + 296, + 529 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 372, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{\\mathrm{s}}^{0}" + }, + { + "bbox": [ + 55, + 372, + 296, + 529 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 372, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{\\mathrm{s}}^{1}" + }, + { + "bbox": [ + 55, + 372, + 296, + 529 + ], + "type": "text", + "content": " and linearly interpolates with roughness to obtain final rendering. Detailed notation and formulation are available in the supplemental material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 538, + 261, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 538, + 261, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 538, + 261, + 552 + ], + "type": "text", + "content": "5. Inverse Rendering from LDR Images" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 557, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 295, + 641 + ], + "type": "text", + "content": "Given multi-view posed LDR images " + }, + { + "bbox": [ + 55, + 557, + 295, + 641 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{I}_i\\}" + }, + { + "bbox": [ + 55, + 557, + 295, + 641 + ], + "type": "text", + "content": " captured with the same or varying exposure levels " + }, + { + "bbox": [ + 55, + 557, + 295, + 641 + ], + "type": "inline_equation", + "content": "\\{\\Delta t_i\\}" + }, + { + "bbox": [ + 55, + 557, + 295, + 641 + ], + "type": "text", + "content": ", our goal is to estimate spatially-varying HDR lighting, surface material (as albedo, roughness, and metallic), and the camera response function (CRF) shared by all input images. We assume that the primary light sources are observed in the input images, following the recent literature [43, 48, 55]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "Previous approaches perform differentiable path tracing during the material and lighting optimization [1, 33], but the number of samples is limited due to heavy computation. This introduces high variance and unstable estimation. The factorized formulation [17, 37, 41, 43] splits light transport into BRDF parameters and shading, and enables updating" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 311, + 555, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 311, + 555, + 358 + ], + "spans": [ + { + "bbox": [ + 313, + 311, + 555, + 358 + ], + "type": "text", + "content": "them alternately. While the method is shown to be more stable and produce high-quality decompositions, it requires HDR images as input to aggregate and bake shading maps, detect emitters, and estimate HDR emissions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 361, + 556, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 556, + 481 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 556, + 481 + ], + "type": "text", + "content": "To address this ill-posed optimization problem, we propose a multi-stage optimization strategy alternatingly estimating spatially varying lighting, material, and CRF. Specifically, we (1) initialize the BRDF with an off-the-shelf monocular inverse rendering method (Section 5.1), (2) recover HDR lighting from LDR with physically-based rendering (Section 5.2), (3) bake the diffuse and specular shading (Section 5.3), and (4) jointly estimate BRDF and CRF (Section 5.4). We repeat Steps 2-4 in that order until all estimates are converged. Figure 3 illustrates our framework." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 487, + 556, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 556, + 535 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 556, + 535 + ], + "type": "text", + "content": "Geometry reconstruction. From input images with camera poses, we reconstruct the geometry with BakedSDF [46], and use normal estimation from off-the-shelf method [7] for regulating surface geometry, following MonoSDF [49]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 549, + 430, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 549, + 430, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 549, + 430, + 562 + ], + "type": "text", + "content": "5.1. BRDF Initialization" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": "For each input image " + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_i" + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": ", we predict a 2D albedo map " + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_i" + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": " using an off-the-shelf single-image albedo estimation method [56] and averaging per-pixel albedo values within each semantic part. Given " + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{A}_i\\}" + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": ", we initialize albedo " + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{a}(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": " by minimizing the error between the projected albedo and the estimated albedo across all input images: " + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\min_{\\mathbf{a}} \\sum_i \\| \\pi_i(\\mathbf{a}) - \\mathbf{A}_i \\|_2" + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\pi_i" + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": " represents the perspective projection onto the image space of " + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_i" + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": " by intersecting camera rays with the surface mesh of the input scene and reading " + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{a}" + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": " at their first intersections. We obtain semantic category labels from part IDs in synthetic scenes or using Mask2Former [3] for real scenes. Roughness and metallicity are both initialized to 0.5." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "468" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 80, + 70, + 267, + 172 + ], + "blocks": [ + { + "bbox": [ + 80, + 70, + 267, + 172 + ], + "lines": [ + { + "bbox": [ + 80, + 70, + 267, + 172 + ], + "spans": [ + { + "bbox": [ + 80, + 70, + 267, + 172 + ], + "type": "image", + "image_path": "ec1874197a8d9d93fdee5bd1aaf6acfb9b4cf15bcd686a521be7d86c6557cc76.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 172, + 296, + 216 + ], + "lines": [ + { + "bbox": [ + 55, + 172, + 296, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 172, + 296, + 216 + ], + "type": "text", + "content": "Figure 4. Emitter geometry estimation " + }, + { + "bbox": [ + 55, + 172, + 296, + 216 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{e}}(\\mathbf{x})" + }, + { + "bbox": [ + 55, + 172, + 296, + 216 + ], + "type": "text", + "content": ". The point " + }, + { + "bbox": [ + 55, + 172, + 296, + 216 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_1" + }, + { + "bbox": [ + 55, + 172, + 296, + 216 + ], + "type": "text", + "content": " on the window is saturated across all input views and thus identified as an emitter. The point " + }, + { + "bbox": [ + 55, + 172, + 296, + 216 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_2" + }, + { + "bbox": [ + 55, + 172, + 296, + 216 + ], + "type": "text", + "content": " on the table is reflective and saturated in some views (e.g., view 2) but not in others and is NOT an emitter." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 221, + 205, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 221, + 205, + 232 + ], + "spans": [ + { + "bbox": [ + 55, + 221, + 205, + 232 + ], + "type": "text", + "content": "5.2. HDR Emission Restoration" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 239, + 296, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 296, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 296, + 275 + ], + "type": "text", + "content": "Given that emitters appear bright across views and produce saturated pixels in LDR images, we estimate the binary emitter geometry " + }, + { + "bbox": [ + 55, + 239, + 296, + 275 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{e}}(\\mathbf{x})" + }, + { + "bbox": [ + 55, + 239, + 296, + 275 + ], + "type": "text", + "content": " on the surface mesh as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 281, + 296, + 313 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 281, + 296, + 313 + ], + "spans": [ + { + "bbox": [ + 68, + 281, + 296, + 313 + ], + "type": "interline_equation", + "content": "M _ {\\mathrm {e}} (\\mathbf {x}) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} \\frac {1}{N} \\sum_ {i} v _ {i} (\\mathbf {x}) \\mathbf {I} _ {i} \\left(\\pi_ {i} (\\mathbf {x})\\right) \\geq 0. 9 9 \\\\ 0 & \\text {o t h e r w i s e ,} \\end{array} \\right. \\tag {3}", + "image_path": "ca2559a9b43478cc242c0035bb3c1c71285fbfb4fef807f91bb7f01cc842f06c.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 319, + 271, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 271, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 271, + 331 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 319, + 271, + 331 + ], + "type": "inline_equation", + "content": "v(\\cdot)\\in \\{0,1\\}" + }, + { + "bbox": [ + 55, + 319, + 271, + 331 + ], + "type": "text", + "content": " indicates visibility. See Figure 4." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 331, + 296, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 331, + 296, + 416 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 296, + 416 + ], + "type": "text", + "content": "The HDR intensity is crucial for inverse rendering but is missing in the LDR images. To recover the HDR lighting, we perform physically-based rendering with single-bounce path tracing to generate the images (2). To estimate the radiance at the point " + }, + { + "bbox": [ + 55, + 331, + 296, + 416 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 331, + 296, + 416 + ], + "type": "text", + "content": " observed from a camera ray, we trace " + }, + { + "bbox": [ + 55, + 331, + 296, + 416 + ], + "type": "inline_equation", + "content": "N = 128" + }, + { + "bbox": [ + 55, + 331, + 296, + 416 + ], + "type": "text", + "content": " secondary rays from " + }, + { + "bbox": [ + 55, + 331, + 296, + 416 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 331, + 296, + 416 + ], + "type": "text", + "content": ", sampled with the current BRDF estimates. The resulting radiance is approximated by:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 88, + 422, + 295, + 454 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 422, + 295, + 454 + ], + "spans": [ + { + "bbox": [ + 88, + 422, + 295, + 454 + ], + "type": "interline_equation", + "content": "\\mathbf {L} _ {\\mathrm {r}} \\left(\\mathbf {x}, \\omega_ {\\mathrm {o}}\\right) = \\sum_ {i = 1} ^ {N} \\mathbf {L} _ {\\text {e n d}} \\left(\\mathbf {x} ^ {\\prime}\\right) f \\left(\\mathbf {x}, \\omega_ {\\mathrm {i}}, \\omega_ {\\mathrm {o}}\\right), \\tag {4}", + "image_path": "45e14ce6f0c40cfd75da7dd8ecad8c7fb7bb175be0bf77b442b2ab3ef0549e9d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 97, + 458, + 295, + 488 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 458, + 295, + 488 + ], + "spans": [ + { + "bbox": [ + 97, + 458, + 295, + 488 + ], + "type": "interline_equation", + "content": "\\mathbf {L} _ {\\text {e n d}} (\\mathbf {x}) = \\left\\{ \\begin{array}{l l} \\mathbf {L} _ {\\mathrm {e}} (\\mathbf {x}) & \\text {i f} M _ {\\mathrm {e}} (\\mathbf {x}) = 1 \\\\ \\mathbf {L} _ {\\text {S L F}} (\\mathbf {x}) & \\text {o t h e r w i s e ,} \\end{array} \\right. \\tag {5}", + "image_path": "c22fb1674a0827bea9791c36109d76bd60e3f76b0da1f738dd00f74a6e91b41f.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 495, + 296, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 296, + 567 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 495, + 296, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{x}' = \\mathbf{x} + d\\cdot \\boldsymbol{\\omega}_{\\mathrm{i}}" + }, + { + "bbox": [ + 55, + 495, + 296, + 567 + ], + "type": "text", + "content": " is the intersected point of the secondary ray along direction " + }, + { + "bbox": [ + 55, + 495, + 296, + 567 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\omega}_{\\mathrm{i}}" + }, + { + "bbox": [ + 55, + 495, + 296, + 567 + ], + "type": "text", + "content": ". Intuitively, we retrieve the radiance from the learnable emission radiance " + }, + { + "bbox": [ + 55, + 495, + 296, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{\\mathrm{e}}(\\mathbf{x})" + }, + { + "bbox": [ + 55, + 495, + 296, + 567 + ], + "type": "text", + "content": " if the secondary ray hits an emitter directly, or otherwise from the pre-computed surface light field " + }, + { + "bbox": [ + 55, + 495, + 296, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{\\mathrm{SLF}}(\\mathbf{x})" + }, + { + "bbox": [ + 55, + 495, + 296, + 567 + ], + "type": "text", + "content": " to approximate the global illumination. Figure 5 shows an illustration." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 567, + 295, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 567, + 295, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 295, + 592 + ], + "type": "text", + "content": "To recover the HDR emitter radiance, we minimize the photometric loss with gradient descent: " + }, + { + "bbox": [ + 55, + 567, + 295, + 592 + ], + "type": "inline_equation", + "content": "\\min_{\\mathbf{L}_{\\mathrm{e}}}\\mathcal{L}_{\\mathrm{photo}}" + }, + { + "bbox": [ + 55, + 567, + 295, + 592 + ], + "type": "text", + "content": " , where" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 65, + 598, + 296, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 598, + 296, + 622 + ], + "spans": [ + { + "bbox": [ + 65, + 598, + 296, + 622 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {p h o t o}} = \\sum_ {i} \\| \\mathrm {C R F} \\left(\\min \\left(\\mathbf {L} _ {\\mathrm {o}} \\left(\\mathbf {x}, \\omega_ {\\mathrm {o}}\\right) \\Delta t _ {i}, 1\\right)\\right) - \\mathbf {I} _ {i} \\| _ {2}. \\tag {6}", + "image_path": "e8e195091ec3d6e18aca03b2efd95c089a80b03a6e1d49dc1f074592b7b47fcd.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 629, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 297, + 714 + ], + "type": "text", + "content": "The emitter radiance " + }, + { + "bbox": [ + 55, + 629, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{\\mathrm{e}}(\\mathbf{x})" + }, + { + "bbox": [ + 55, + 629, + 297, + 714 + ], + "type": "text", + "content": " is enhanced to match the captured images, while the BRDF and CRF are fixed in this stage. Intuitively, in order to minimize the photometric loss with physically-based rendering, the emitter intensity " + }, + { + "bbox": [ + 55, + 629, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{\\mathrm{e}}(\\mathbf{x})" + }, + { + "bbox": [ + 55, + 629, + 297, + 714 + ], + "type": "text", + "content": " should be increased significantly and enhance the overall brightness of the rendered image. Therefore, the HDR lighting is recovered in this process." + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 347, + 70, + 518, + 176 + ], + "blocks": [ + { + "bbox": [ + 347, + 70, + 518, + 176 + ], + "lines": [ + { + "bbox": [ + 347, + 70, + 518, + 176 + ], + "spans": [ + { + "bbox": [ + 347, + 70, + 518, + 176 + ], + "type": "image", + "image_path": "afa70ad937f130915d42a5e9b0ab63b70c9583aa7aa8fbf7f494b64e91136125.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 178, + 555, + 232 + ], + "lines": [ + { + "bbox": [ + 313, + 178, + 555, + 232 + ], + "spans": [ + { + "bbox": [ + 313, + 178, + 555, + 232 + ], + "type": "text", + "content": "Figure 5. HDR emission restoration. Top: Ray sampling process. Learnable direct lighting " + }, + { + "bbox": [ + 313, + 178, + 555, + 232 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{\\mathrm{e}}" + }, + { + "bbox": [ + 313, + 178, + 555, + 232 + ], + "type": "text", + "content": " is retrieved if a ray hits an emitter (e.g., window) or " + }, + { + "bbox": [ + 313, + 178, + 555, + 232 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{\\mathrm{SLF}}" + }, + { + "bbox": [ + 313, + 178, + 555, + 232 + ], + "type": "text", + "content": " is retrieved otherwise (e.g., wall). Bottom: HDR restoration process. By performing differentiable physically-based rendering, the photometric loss enhances the emitter intensity " + }, + { + "bbox": [ + 313, + 178, + 555, + 232 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{\\mathrm{e}}" + }, + { + "bbox": [ + 313, + 178, + 555, + 232 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 245, + 411, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 245, + 411, + 258 + ], + "spans": [ + { + "bbox": [ + 314, + 245, + 411, + 258 + ], + "type": "text", + "content": "5.3. Shading Baking" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 263, + 555, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 263, + 555, + 371 + ], + "spans": [ + { + "bbox": [ + 313, + 263, + 555, + 371 + ], + "type": "text", + "content": "After the HDR emission is recovered in the previous stage, we bake the diffuse and specular shading maps " + }, + { + "bbox": [ + 313, + 263, + 555, + 371 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{\\mathrm{d}}" + }, + { + "bbox": [ + 313, + 263, + 555, + 371 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 263, + 555, + 371 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{\\mathrm{s}}^{0}" + }, + { + "bbox": [ + 313, + 263, + 555, + 371 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 263, + 555, + 371 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{\\mathrm{s}}^{1}" + }, + { + "bbox": [ + 313, + 263, + 555, + 371 + ], + "type": "text", + "content": ", following factorized light transport (details in supplementary material). Given a surface point intersected by a camera ray, multiple rays are sampled according to the current (fixed) BRDF. The incident radiance is computed with multibounce path tracing, where a ray keeps tracing if it hits a non-emissive specular surface and stops otherwise. The radiance of the endpoint follows Equation (5)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 380, + 469, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 380, + 469, + 392 + ], + "spans": [ + { + "bbox": [ + 314, + 380, + 469, + 392 + ], + "type": "text", + "content": "5.4. BRDF & CRF Optimization" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 397, + 555, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 397, + 555, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 397, + 555, + 434 + ], + "type": "text", + "content": "With the baked shading maps " + }, + { + "bbox": [ + 313, + 397, + 555, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{\\mathrm{d}}" + }, + { + "bbox": [ + 313, + 397, + 555, + 434 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 397, + 555, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{\\mathrm{s}}^{0}" + }, + { + "bbox": [ + 313, + 397, + 555, + 434 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 397, + 555, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{\\mathrm{s}}^{1}" + }, + { + "bbox": [ + 313, + 397, + 555, + 434 + ], + "type": "text", + "content": ", the albedo, metallicity, roughness, and CRF are jointly optimized. The optimization involves the following objective function:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 338, + 444, + 555, + 462 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 444, + 555, + 462 + ], + "spans": [ + { + "bbox": [ + 338, + 444, + 555, + 462 + ], + "type": "interline_equation", + "content": "\\min _ {\\mathbf {a}, m, \\sigma , \\mathbf {g}} \\mathcal {L} _ {\\text {p h o t o}} + \\lambda_ {\\mathrm {a}} \\mathcal {L} _ {\\text {a l b e d o}} + \\lambda_ {\\mathrm {c}} \\mathcal {L} _ {\\text {C R F}} + \\lambda_ {\\mathrm {m}} \\mathcal {L} _ {\\text {m a t}}, \\tag {7}", + "image_path": "561ed6be50b2166ed11b8820c478d5a6f26dadebd814bb0d7e39790839ece76e.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 472, + 506, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 506, + 484 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 506, + 484 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 472, + 506, + 484 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{a}} = 0.01" + }, + { + "bbox": [ + 313, + 472, + 506, + 484 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{c}} = 0.001" + }, + { + "bbox": [ + 313, + 472, + 506, + 484 + ], + "type": "text", + "content": " , and " + }, + { + "bbox": [ + 313, + 472, + 506, + 484 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{m}} = 0.005" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 487, + 555, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 555, + 512 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 555, + 512 + ], + "type": "text", + "content": "Photometric loss. We perform physically-based rendering, and the photometric loss " + }, + { + "bbox": [ + 313, + 487, + 555, + 512 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{photo}}" + }, + { + "bbox": [ + 313, + 487, + 555, + 512 + ], + "type": "text", + "content": " is computed as in Eq. 6." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 514, + 555, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 555, + 551 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 555, + 551 + ], + "type": "text", + "content": "Albedo regularization. Motivated by the shift-scale-invariant loss in MonoSDF [49], we adopt a shift-invariant loss against the monocular albedo maps " + }, + { + "bbox": [ + 313, + 514, + 555, + 551 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{A}_i\\}" + }, + { + "bbox": [ + 313, + 514, + 555, + 551 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 368, + 560, + 555, + 585 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 560, + 555, + 585 + ], + "spans": [ + { + "bbox": [ + 368, + 560, + 555, + 585 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {a l b e d o}} = \\sum_ {i} \\| \\pi_ {i} (\\mathbf {a}) - s _ {i} \\mathbf {A} _ {i} \\| _ {2}, \\tag {8}", + "image_path": "c0d991f8accc1e6636dafc990e7e61980e47cfc221759db016fed539111b55d0.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 594, + 538, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 538, + 606 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 538, + 606 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 594, + 538, + 606 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 313, + 594, + 538, + 606 + ], + "type": "text", + "content": " is the scale used to align the two albedo maps." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 609, + 555, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 609, + 555, + 681 + ], + "spans": [ + { + "bbox": [ + 313, + 609, + 555, + 681 + ], + "type": "text", + "content": "CRF regularization. Based on the fair assumption that the learned CRF should not deviate significantly from the mean curve " + }, + { + "bbox": [ + 313, + 609, + 555, + 681 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{g}}" + }, + { + "bbox": [ + 313, + 609, + 555, + 681 + ], + "type": "text", + "content": " (Eq. 1), and that the curve should be monotonically increasing, we regularize the " + }, + { + "bbox": [ + 313, + 609, + 555, + 681 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 313, + 609, + 555, + 681 + ], + "type": "text", + "content": " norm of the estimated coefficients of the PCA bases, and enforce the monotonicity of the sampled CRF as follows:" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 350, + 691, + 555, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 691, + 555, + 716 + ], + "spans": [ + { + "bbox": [ + 350, + 691, + 555, + 716 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {C R F}} = \\| \\mathbf {w} \\| _ {2} + \\sum_ {i} \\max \\left(g _ {i - 1} - g _ {i}, 0\\right). \\tag {9}", + "image_path": "0092d737931af6790d9d79c0cc47d8aede1306ef88e5f25bdc7f003cd624e51d.jpg" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "469" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": "Material regularization. We apply roughness-metallicity regularization " + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{mat}}" + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": ", which enforces the consistency of surface roughness and metallic within the same semantic instances (e.g. chairs, tables). While the shading and emission are initially inaccurate and lead to erroneous BRDFs, we alternately update the BRDFs and produce better emission and shading, improving the BRDFs. Our alternating strategy yields more stable optimization and better estimation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 182, + 138, + 196 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 182, + 138, + 196 + ], + "spans": [ + { + "bbox": [ + 55, + 182, + 138, + 196 + ], + "type": "text", + "content": "6. Experiments" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 202, + 296, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 202, + 296, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 202, + 296, + 275 + ], + "type": "text", + "content": "We evaluate our method on both synthetic and real-world data and compare it against several baseline methods. We analyze the quality of our inverse rendering, relighting, and novel-view synthesis capability. Please see our supplementary material for additional results and evaluations. We will release our code and data." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 278, + 296, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 278, + 296, + 470 + ], + "spans": [ + { + "bbox": [ + 55, + 278, + 296, + 470 + ], + "type": "text", + "content": "Datasets. We evaluate our method and baselines on the ScanNet++ [47] and FIPT [43] datasets. ScanNet++ contains 380 real-world indoor scenes and collects multi-view LDR images with a DSLR camera at a constant exposure level. We select four scenes that cover a variety of indoor scenes (e.g., office, bathroom). The FIPT dataset consists of four synthetic and two real-world scenes. The synthetic scenes provide ground-truth geometry and material properties for evaluation. We used the LDR images provided in the dataset, where the lighting is clipped in overly-exposed regions, compressed with a nonlinear CRF, quantized into 256 levels, and stored in 8-bit PNG format. We adopt the constant exposure setting to provide a fair comparison with baselines, as they cannot handle varying exposure. We further demonstrate that IRIS handles varying exposure levels of input LDR images and estimates CRF jointly." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 473, + 296, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 296, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 296, + 664 + ], + "type": "text", + "content": "Baselines. We compare our method with the following baselines. Li et al. [25] estimate BRDF, geometry (depth & normal), and lighting condition of an indoor scene from a single image using a data-driven approach. In addition to the single image, the model also takes emitter masks as input, and we provide the mask estimated by our method (3). NeILF [45] targets inverse rendering from known geometry and multiview images. The spatially-varying BRDF and illumination are parameterized as neural fields [29]. Physically-based rendering and single-bounce ray tracing are performed during the training and rendering process. The HDR-to-LDR tone mapping is modeled as a learnable gamma correction. The original FIPT [43] is designed for HDR input, and fails to estimate the emission mask from LDR images. Thus, we additionally provide the emission mask estimated by our method (3). We denote this modified version as \"FIPT*\"." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "To further evaluate our performance, we compare the original FIPT with HDR input if the HDR information is available. This privileged information will make illumination estimation much more accurate." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 337, + 112, + 529, + 196 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 555, + 105 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 555, + 105 + ], + "type": "text", + "content": "Table 1. BRDF-emission comparison on synthetic data. Numbers are averaged across four synthetic scenes with GT [43]. The best metrics among LDR methods are highlighted in bold." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 337, + 112, + 529, + 196 + ], + "lines": [ + { + "bbox": [ + 337, + 112, + 529, + 196 + ], + "spans": [ + { + "bbox": [ + 337, + 112, + 529, + 196 + ], + "type": "table", + "html": "
MethodInputkda'σLe
PSNR ↑IoU ↑L2 ↓
Li et al [25]LDR16.7314.0111.300.352.29
NeILF [45]LDR16.8514.0216.96——
FIPT*LDR15.499.744.990.690.28
OursLDR22.3317.9221.380.690.12
FIPT [43]HDR29.9525.9826.370.860.03
", + "image_path": "be20db0f963098f42cfdca513b969458135d58b83b292b83d6b975225a58804e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 205, + 493, + 219 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 493, + 219 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 493, + 219 + ], + "type": "text", + "content": "6.1. Inverse Rendering of Real Scenes" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 225, + 556, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 225, + 556, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 225, + 556, + 441 + ], + "type": "text", + "content": "We provide a qualitative evaluation of inverse rendering in Figure 6. Conducting single-bounce ray tracing in the neural light field, NeILF [45] fails to remove the shadow in diffuse reflectance " + }, + { + "bbox": [ + 313, + 225, + 556, + 441 + ], + "type": "inline_equation", + "content": "\\mathbf{k}_{\\mathrm{d}}" + }, + { + "bbox": [ + 313, + 225, + 556, + 441 + ], + "type": "text", + "content": ", estimates lower roughness " + }, + { + "bbox": [ + 313, + 225, + 556, + 441 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 225, + 556, + 441 + ], + "type": "text", + "content": " on the walls than the mirror, and produces an incorrect emission mask. FIPT* struggles to recover surface material from LDR input, where accurate lighting information is missing. Please note that FIPT* takes emitter masks from Equation (3) as additional input, but does not recover HDR emission in the optimization procedure. On the other hand, IRIS recovers an almost shading-free diffuse field. Regarding roughness, our method effectively identifies specular surfaces with low roughness on the mirror, demonstrating better specular/rough material separation. Additionally, our proposed HDR emission estimation proves effective for real scenes, recovering HDR lighting sources (i.e., window and ceiling light). The restored HDR lighting makes the light transport more accurate and significantly enhances the performance of inverse rendering." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 454, + 549, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 454, + 549, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 549, + 468 + ], + "type": "text", + "content": "6.2. View Synthesis and Relighting of Real Scenes" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "content": "We compare the relighting with FIPT* in Figure 7. FIPT* fails to estimate accurate surface material from LDR input and tends to recover low roughness for most surfaces, which unrealistically reflects the inserted light source and objects. To be more specific, the ground of the office (first scene) and the wall of the bathroom (second scene) look overly smooth and specular, which is not consistent with the observation and produces unrealistic reflections. In contrast, IRIS produces better relighting and object insertion quality thanks to better inverse rendering. The reflections on the specular surfaces (e.g., whiteboard, mirror) and the shading on the wall demonstrate the strength of IRIS for realistic relighting. Furthermore, the recovered HDR lighting can produce naturalistic specular reflection and shadows for inserted objects. Additionally, we insert moving light sources in the scene, shown in Figure 8, where the shadow on the ground and shading on the table change consistently. IRIS enables the photorealistic rendering of novel objects under diverse lighting and from arbitrary viewpoints, showcasing the significant potential for content creation applications." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "470" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 555, + 284 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 555, + 284 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 555, + 284 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 555, + 284 + ], + "type": "image", + "image_path": "9f1ef19091189052fc0baf90f7b6eabf7bd4cdbacdb68c5b113ea16e503c1f19.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 287, + 555, + 321 + ], + "lines": [ + { + "bbox": [ + 55, + 287, + 555, + 321 + ], + "spans": [ + { + "bbox": [ + 55, + 287, + 555, + 321 + ], + "type": "text", + "content": "Figure 6. Decompositions of real scenes from ScanNet++ [47]. In addition to good reconstruction, IRIS produces a detailed diffuse map, identifies the mirror as a low-roughness region, and recovers HDR lighting from windows and ceiling light. Please note that the emission maps are normalized to visualize the relative intensity." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 55, + 323, + 555, + 589 + ], + "blocks": [ + { + "bbox": [ + 55, + 323, + 555, + 589 + ], + "lines": [ + { + "bbox": [ + 55, + 323, + 555, + 589 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 555, + 589 + ], + "type": "image", + "image_path": "ecbc00fa12f23d8c3a13c4900109eefb7b3585b4dbdd8a5bf7df0ddf20890c7c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 593, + 555, + 638 + ], + "lines": [ + { + "bbox": [ + 55, + 593, + 555, + 638 + ], + "spans": [ + { + "bbox": [ + 55, + 593, + 555, + 638 + ], + "type": "text", + "content": "Figure 7. Relighting and object insertion with real scenes from the ScanNet++ [47] and FIPT datasets [43]. When inserting new light sources (Relighting 1 & 2), IRIS not only changes shadows but also produces reflection of the light on the whiteboard (first scene) and mirror (second scene), enhancing realism substantially. The recovered HDR lighting also strengthens the reflection and shadow of inserted objects (\"Object Insertion\" column)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 644, + 191, + 657 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 644, + 191, + 657 + ], + "spans": [ + { + "bbox": [ + 55, + 644, + 191, + 657 + ], + "type": "text", + "content": "6.3. Quantitative Evaluation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 665, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 297, + 713 + ], + "type": "text", + "content": "Quantitative evaluation of inverse rendering is difficult for real-world scenes as ground truth is not available. As a result, we evaluate IRIS and baselines on the synthetic scenes from FIPT, where ground truth of surface material, illumina-" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 644, + 556, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 644, + 556, + 706 + ], + "spans": [ + { + "bbox": [ + 313, + 644, + 556, + 706 + ], + "type": "text", + "content": "tion, and relighting are provided. Please note that we use the ground-truth mesh during optimization for all methods to minimize the influence of imperfect geometry and conduct a fair comparison. We compute the metrics of inverse rendering and provide the comparison in Table 1, which aggregates" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "471" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 118, + 71, + 176, + 119 + ], + "blocks": [ + { + "bbox": [ + 63, + 71, + 113, + 79 + ], + "lines": [ + { + "bbox": [ + 63, + 71, + 113, + 79 + ], + "spans": [ + { + "bbox": [ + 63, + 71, + 113, + 79 + ], + "type": "text", + "content": "Reconstruction" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 118, + 71, + 176, + 119 + ], + "lines": [ + { + "bbox": [ + 118, + 71, + 176, + 119 + ], + "spans": [ + { + "bbox": [ + 118, + 71, + 176, + 119 + ], + "type": "image", + "image_path": "7bdac9d5775894e719933053cd92fa6664dc83509605d3211c2c508bdd9b28e2.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 178, + 80, + 234, + 119 + ], + "blocks": [ + { + "bbox": [ + 189, + 71, + 220, + 79 + ], + "lines": [ + { + "bbox": [ + 189, + 71, + 220, + 79 + ], + "spans": [ + { + "bbox": [ + 189, + 71, + 220, + 79 + ], + "type": "text", + "content": "Relight 2" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 178, + 80, + 234, + 119 + ], + "lines": [ + { + "bbox": [ + 178, + 80, + 234, + 119 + ], + "spans": [ + { + "bbox": [ + 178, + 80, + 234, + 119 + ], + "type": "image", + "image_path": "1e0287af317380a0de0e428152b06857a3422a84bdee19af3d20194ece9e4dd8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 62, + 122, + 288, + 134 + ], + "lines": [ + { + "bbox": [ + 62, + 122, + 288, + 134 + ], + "spans": [ + { + "bbox": [ + 62, + 122, + 288, + 134 + ], + "type": "text", + "content": "Figure 8. Relighting with a moving light ball in a real scene." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 235, + 80, + 291, + 119 + ], + "blocks": [ + { + "bbox": [ + 247, + 71, + 278, + 79 + ], + "lines": [ + { + "bbox": [ + 247, + 71, + 278, + 79 + ], + "spans": [ + { + "bbox": [ + 247, + 71, + 278, + 79 + ], + "type": "text", + "content": "Relight 3" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 235, + 80, + 291, + 119 + ], + "lines": [ + { + "bbox": [ + 235, + 80, + 291, + 119 + ], + "spans": [ + { + "bbox": [ + 235, + 80, + 291, + 119 + ], + "type": "image", + "image_path": "47f37bc9591563489e577d706a1323d34d2c5d36516a5372171188687e3a1486.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 59, + 140, + 115, + 203 + ], + "blocks": [ + { + "bbox": [ + 59, + 140, + 115, + 203 + ], + "lines": [ + { + "bbox": [ + 59, + 140, + 115, + 203 + ], + "spans": [ + { + "bbox": [ + 59, + 140, + 115, + 203 + ], + "type": "image", + "image_path": "bfa617a85da7f2f229e9ff2a4ed943149c37afa3c93074556c5a557f6db6f6f7.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 210, + 295, + 233 + ], + "lines": [ + { + "bbox": [ + 55, + 210, + 295, + 233 + ], + "spans": [ + { + "bbox": [ + 55, + 210, + 295, + 233 + ], + "type": "text", + "content": "Figure 9. CRF estimation from real-world devices. The orange dashed line is the ground truth CRF, and the blue line our estimate." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 118, + 140, + 174, + 204 + ], + "blocks": [ + { + "bbox": [ + 118, + 140, + 174, + 204 + ], + "lines": [ + { + "bbox": [ + 118, + 140, + 174, + 204 + ], + "spans": [ + { + "bbox": [ + 118, + 140, + 174, + 204 + ], + "type": "image", + "image_path": "5739447e2d8337658fbe15bb0fd8b7090ac15092f8319022289fba06645e86a9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 178, + 140, + 233, + 204 + ], + "blocks": [ + { + "bbox": [ + 178, + 140, + 233, + 204 + ], + "lines": [ + { + "bbox": [ + 178, + 140, + 233, + 204 + ], + "spans": [ + { + "bbox": [ + 178, + 140, + 233, + 204 + ], + "type": "image", + "image_path": "977019b9a4a2a13ad98529e9ce91ccdd642e418274d9af61c74ba58c46c773aa.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 236, + 140, + 291, + 204 + ], + "blocks": [ + { + "bbox": [ + 236, + 140, + 291, + 204 + ], + "lines": [ + { + "bbox": [ + 236, + 140, + 291, + 204 + ], + "spans": [ + { + "bbox": [ + 236, + 140, + 291, + 204 + ], + "type": "image", + "image_path": "7cdb30f3648732ac98e9e51acd30aba19a9bc3a40bb08acbee964f6f28140033.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 255, + 255, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 255, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 255, + 266 + ], + "type": "text", + "content": "the results from the 4 synthetic FIPT scenes [43]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 267, + 296, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 267, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 267, + 296, + 350 + ], + "type": "text", + "content": "We compare the roughness " + }, + { + "bbox": [ + 55, + 267, + 296, + 350 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 55, + 267, + 296, + 350 + ], + "type": "text", + "content": ", diffuse reflectance " + }, + { + "bbox": [ + 55, + 267, + 296, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{k}_{\\mathrm{d}}" + }, + { + "bbox": [ + 55, + 267, + 296, + 350 + ], + "type": "text", + "content": " among the diffuse surfaces, and material reflectance: " + }, + { + "bbox": [ + 55, + 267, + 296, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{a}^{\\prime}" + }, + { + "bbox": [ + 55, + 267, + 296, + 350 + ], + "type": "text", + "content": ". We report the PSNR for these BRDF estimates. For emission, we calculate the intersection over the union (IoU) of the emission masks and compare the " + }, + { + "bbox": [ + 55, + 267, + 296, + 350 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 55, + 267, + 296, + 350 + ], + "type": "text", + "content": " error of estimated emission maps. All the estimates are rendered in image space and compared with ground truth." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 351, + 296, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 351, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 351, + 296, + 555 + ], + "type": "text", + "content": "Table 1 shows that our method surpasses all baseline methods that use LDR inputs regarding inverse rendering quality. Compared to the state-of-the-art single-view method [25], our method excels in all metrics across various scenes, indicating the superior decomposition of material properties from multi-view observations. Furthermore, even when the estimated emitter mask is provided for FIPT*, it fails to recover emission radiance, leading to compromised BRDF estimation accuracy. The " + }, + { + "bbox": [ + 55, + 351, + 296, + 555 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 55, + 351, + 296, + 555 + ], + "type": "text", + "content": " error of emission maps shows that our proposed HDR restoration (Section 5.2) successfully recovers the HDR lighting from LDR observations, which leads to better light transport modeling and more accurate BRDF estimation. We also compare with FIPT taking HDR images as input (bottom row), which serves as a reference, indicating complete lighting information is crucial for accurate inverse rendering. IRIS recovers HDR and achieves the smallest metric gaps among all methods taking LDR input." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 555, + 296, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 296, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 296, + 639 + ], + "type": "text", + "content": "We compare novel-view synthesis and relighting of all competing algorithms and report results in Table 2. The novel-view synthesis results show that IRIS is comparable with baseline methods [45], while exhibiting superior material and lighting estimation. Our main focus is relighting, and IRIS outperforms all baselines with LDR input, demonstrating the effectiveness of inverse rendering." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 647, + 149, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 149, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 149, + 659 + ], + "type": "text", + "content": "6.4. Ablation Study" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "We validate our design choices and provide an ablation study in Table 3, where we compute PSNR for material " + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "inline_equation", + "content": "(\\mathbf{k}_{\\mathrm{d}},\\mathbf{a}^{\\prime},\\sigma)" + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": " and L2 for CRF curves and emission maps " + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{\\mathrm{e}}" + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": ". The emitter mask estimation in Equation (3) identifies the emitter from" + } + ] + } + ], + "index": 17 + }, + { + "type": "table", + "bbox": [ + 337, + 102, + 529, + 205 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 555, + 93 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 555, + 93 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 555, + 93 + ], + "type": "text", + "content": "Table 2. Quantitative results of novel-view synthesis (NVS) and relighting on synthetic scenes with ground truth." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 337, + 102, + 529, + 205 + ], + "lines": [ + { + "bbox": [ + 337, + 102, + 529, + 205 + ], + "spans": [ + { + "bbox": [ + 337, + 102, + 529, + 205 + ], + "type": "table", + "html": "
MethodInputPSNR ↑SSIM ↑LPIPS ↓
NVSNeILF [45]LDR30.0310.8990.180
FIPT*LDR15.1800.7440.437
OursLDR29.4680.9030.181
FIPT [43]HDR28.7600.9020.186
RelightingLi et al. [25]LDR22.4440.8130.398
FIPT*LDR11.6260.7050.359
OursLDR22.8610.8820.173
FIPT [43]HDR29.0120.9020.139
", + "image_path": "cb98fdb31e01ee6de5d0c06784e4948e45f850ce1c948d8877d774a060cfec54.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_body" + } + ], + "index": 19 + }, + { + "type": "table", + "bbox": [ + 315, + 221, + 553, + 301 + ], + "blocks": [ + { + "bbox": [ + 389, + 209, + 479, + 220 + ], + "lines": [ + { + "bbox": [ + 389, + 209, + 479, + 220 + ], + "spans": [ + { + "bbox": [ + 389, + 209, + 479, + 220 + ], + "type": "text", + "content": "Table 3. Ablation study" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 221, + 553, + 301 + ], + "lines": [ + { + "bbox": [ + 315, + 221, + 553, + 301 + ], + "spans": [ + { + "bbox": [ + 315, + 221, + 553, + 301 + ], + "type": "table", + "html": "
Methodkda'σLeCRF
PSNR ↑IoU ↑L2 ↓L2 ↓
- Emitter mask Me (Equation (3))16.5011.304.810.000.488.91
- CRF modeling (Equation (1))18.1813.5917.520.690.12—
- HDR restoration (Section 5.2)18.2014.407.310.690.286.23
- Lalbedo (Equation (8))21.5917.3617.430.690.152.42
- LCRF (Equation (9))14.4912.8125.590.690.139.40
Full model (Ours)22.3317.9221.380.690.122.35
", + "image_path": "338c4acc60e015e8b96780170cb5dd72344b9879fb92705de7d432b1f15e1080.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "table_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 311, + 555, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 311, + 555, + 443 + ], + "spans": [ + { + "bbox": [ + 313, + 311, + 555, + 443 + ], + "type": "text", + "content": "LDR observations and improves the emission map in the second row. CRF estimation is critical for modeling HDR-to-LDR conversion and helps inverse rendering in the third row. Nevertheless, the HDR restoration (Section 5.2) plays an important role in recovering accurate light transport from the captured images. Once the HDR light is restored by our proposed approach, the material (especially roughness " + }, + { + "bbox": [ + 313, + 311, + 555, + 443 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 311, + 555, + 443 + ], + "type": "text", + "content": "), CRF, and emission map are improved significantly. The results demonstrate the significance of each component and show that inverse rendering from casually captured LDR images benefits from our proposed framework." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 454, + 413, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 454, + 413, + 466 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 413, + 466 + ], + "type": "text", + "content": "6.5. CRF Estimation" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 473, + 555, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 555, + 558 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 555, + 558 + ], + "type": "text", + "content": "To further validate the robustness of our method, we generate LDR images from FIPT's synthetic HDR scenes using different CRFs collected from real-world devices [10]. In this evaluation, we perform inverse rendering with IRIS and plot the CRF. The results in Figure 9 show that our method can generalize to various imaging pipelines. Please refer to the supplementary details for more analysis of CRF estimation." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 571, + 388, + 584 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 571, + 388, + 584 + ], + "spans": [ + { + "bbox": [ + 313, + 571, + 388, + 584 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "We presented IRIS, an inverse rendering framework for indoor scenes from casually captured LDR images. Our key innovations include HDR lighting restoration, CRF estimation, surface material estimation, and alternating optimization to recover all scene properties accurately. IRIS estimates accurate surface material, spatially varying HDR illumination, and CRF from real-world LDR images. As a result, IRIS enables photorealistic rendering under diverse lighting and from arbitrary viewpoints, showcasing the significant potential for content creation applications." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "472" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "type": "text", + "content": "[1] Dejan Azinovic, Tzu-Mao Li, Anton Kaplanyan, and Matthias Nießner. Inverse path tracing for joint material and lighting estimation. In CVPR, 2019. 2, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 296, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 296, + 168 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 296, + 168 + ], + "type": "text", + "content": "[2] Sai Bi, Zexiang Xu, Pratul Srinivasan, Ben Mildenhall, Kalyan Sunkavalli, Miloš Hasan, Yannick Hold-Geoffroy, David Kriegman, and Ravi Ramamoorthi. Neural reflectance fields for appearance acquisition. arXiv:2008.03824, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 171, + 296, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 171, + 296, + 213 + ], + "spans": [ + { + "bbox": [ + 62, + 171, + 296, + 213 + ], + "type": "text", + "content": "[3] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2022. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 215, + 296, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 215, + 296, + 247 + ], + "spans": [ + { + "bbox": [ + 62, + 215, + 296, + 247 + ], + "type": "text", + "content": "[4] Robert L Cook and Kenneth E. Torrance. A reflectance model for computer graphics. ACM Transactions on Graphics (ToG), 1982. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 249, + 296, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 249, + 296, + 281 + ], + "spans": [ + { + "bbox": [ + 62, + 249, + 296, + 281 + ], + "type": "text", + "content": "[5] Paul E. Debevec and Jitendra Malik. Recovering high dynamic range radiance maps from photographs. In SIGGRAPH, 1997. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 284, + 296, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 284, + 296, + 327 + ], + "spans": [ + { + "bbox": [ + 62, + 284, + 296, + 327 + ], + "type": "text", + "content": "[6] Valentin Deschaintre, Miika Aittala, Fredo Durand, George Drettakis, and Adrien Bousseau. Single-image SVBRDF capture with a rendering-aware deep network. ACM Transactions on Graphics (ToG), 37(4):128:1-15, 2018. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 329, + 296, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 329, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 62, + 329, + 296, + 361 + ], + "type": "text", + "content": "[7] Ainaz Eftekhar, Alexander Sax, Jitendra Malik, and Amir Zamir. Omnidata: A scalable pipeline for making multi-task mid-level vision datasets from 3D scans. In ICCV, 2021. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 363, + 296, + 416 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 363, + 296, + 416 + ], + "spans": [ + { + "bbox": [ + 62, + 363, + 296, + 416 + ], + "type": "text", + "content": "[8] Marc-Andre Gardner, Kalyan Sunkavalli, Ersin Yumer, Xiaohui Shen, Emiliano Gambaretto, Christian Gagné, and Jean-François Lalonde. Learning to predict indoor illumination from a single image. ACM Transactions on Graphics, 36(6): 176:1-14, 2017. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 418, + 296, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 418, + 296, + 462 + ], + "spans": [ + { + "bbox": [ + 62, + 418, + 296, + 462 + ], + "type": "text", + "content": "[9] Dan B Goldman, Brian Curless, Aaron Hertzmann, and Steven M Seitz. Shape and spatially-varying BRDFs from photometric stereo. IEEE Transactions on Pattern Analysis and Machine Intelligence, 32(6):1060-1071, 2009. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 464, + 295, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 464, + 295, + 485 + ], + "spans": [ + { + "bbox": [ + 57, + 464, + 295, + 485 + ], + "type": "text", + "content": "[10] Michael D Grossberg and Shree K Nayar. What is the space of camera response functions? In CVPR, 2003. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 487, + 296, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 487, + 296, + 518 + ], + "spans": [ + { + "bbox": [ + 57, + 487, + 296, + 518 + ], + "type": "text", + "content": "[11] Michael D. Grossberg and Shree K. Nayar. Modeling the space of camera response functions. IEEE TPAMI, 26(10): 1272-1282, 2004. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 521, + 295, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 521, + 295, + 553 + ], + "spans": [ + { + "bbox": [ + 57, + 521, + 295, + 553 + ], + "type": "text", + "content": "[12] Jon Hasselgren, Nikolai Hofmann, and Jacob Munkberg. Shape, light & material decomposition from images using monte carlo rendering and denoising. NeurIPS, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 555, + 295, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 555, + 295, + 586 + ], + "spans": [ + { + "bbox": [ + 57, + 555, + 295, + 586 + ], + "type": "text", + "content": "[13] Xin Huang, Qi Zhang, Ying Feng, Hongdong Li, Xuan Wang, and Qing Wang. HDR-NeRF: High dynamic range neural radiance fields. In CVPR, 2022. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 590, + 295, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 590, + 295, + 632 + ], + "spans": [ + { + "bbox": [ + 57, + 590, + 295, + 632 + ], + "type": "text", + "content": "[14] Wenzel Jakob, Sébastien Speierer, Nicolas Roussel, Merlin Nimier-David, Delio Vicini, Tizian Zeltner, Baptiste Nicolet, Miguel Crespo, Vincent Leroy, and Ziyi Zhang. Mitsuba 3 renderer, 2022. https://mitsuba-renderer.org. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 635, + 295, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 295, + 666 + ], + "type": "text", + "content": "[15] Haian Jin, Isabella Liu, Peijia Xu, Xiaoshuai Zhang, Songfang Han, Sai Bi, Xiaowei Zhou, Zexiang Xu, and Hao Su. TensoIR: Tensorial inverse rendering. CVPR, 2023. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 669, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 669, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 669, + 295, + 689 + ], + "type": "text", + "content": "[16] James T Kajiya. The rendering equation. In SIGGRAPH, 1986. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "text", + "content": "[17] Jaroslav Krivánek and Pascal Gautron. Practical global illumination with irradiance caching. Springer, 2022. 4" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 106 + ], + "type": "text", + "content": "[18] Eric Lafortune. Mathematical models and Monte Carlo algorithms for physically based rendering. PhD thesis, Katholieke Universiteit Leuven, 1996. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 108, + 555, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 555, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 555, + 150 + ], + "type": "text", + "content": "[19] Xiao Li, Yue Dong, Pieter Peers, and Xin Tong. Modeling surface appearance from a single photograph using self-augmented convolutional neural networks. ACM Transactions on Graphics (ToG), 36(4):45:1-11, 2017. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 152, + 555, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 152, + 555, + 184 + ], + "spans": [ + { + "bbox": [ + 317, + 152, + 555, + 184 + ], + "type": "text", + "content": "[20] Zhengqi Li and Noah Snavely. CGIntrinsics: Better intrinsic image decomposition through physically-based rendering. In ECCV, 2018. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 186, + 554, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 554, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 554, + 207 + ], + "type": "text", + "content": "[21] Zhengqi Li and Noah Snavely. Learning intrinsic image decomposition from watching the world. In CVPR, 2018. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 209, + 555, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 209, + 555, + 241 + ], + "spans": [ + { + "bbox": [ + 317, + 209, + 555, + 241 + ], + "type": "text", + "content": "[22] Zhengqin Li, Kalyan Sunkavalli, and Manmohan Chandraker. Materials for masses: SVBRDF acquisition with a single mobile phone image. In ECCV, 2018. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 243, + 554, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 243, + 554, + 285 + ], + "spans": [ + { + "bbox": [ + 317, + 243, + 554, + 285 + ], + "type": "text", + "content": "[23] Zhengqin Li, Mohammad Shafiei, Ravi Ramamoorthi, Kalyan Sunkavalli, and Manmohan Chandraker. Inverse rendering for complex indoor scenes: Shape, spatially-varying lighting and SVBRDF from a single image. In CVPR, 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 288, + 555, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 288, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 288, + 555, + 331 + ], + "type": "text", + "content": "[24] Zhengqin Li, Ting-Wei Yu, Shen Sang, Sarah Wang, Meng Song, Yuhan Liu, Yu-Ying Yeh, Rui Zhu, Nitesh Gundavarapu, Jia Shi, et al. OpenRooms: An open framework for photorealistic indoor scene datasets. In CVPR, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 333, + 555, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 333, + 555, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 333, + 555, + 376 + ], + "type": "text", + "content": "[25] Zhengqin Li, Jia Shi, Sai Bi, Rui Zhu, Kalyan Sunkavalli, Miloš Hasan, Zexiang Xu, Ravi Ramamoorthi, and Manmohan Chandraker. Physically-based editing of indoor scene lighting from a single image. In ECCV, 2022. 2, 6, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 377, + 555, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 377, + 555, + 420 + ], + "spans": [ + { + "bbox": [ + 317, + 377, + 555, + 420 + ], + "type": "text", + "content": "[26] Zhi-Hao Lin, Bohan Liu, Yi-Ting Chen, David Forsyth, Jia-Bin Huang, Anand Bhattachad, and Shenlong Wang. UrbanIR: Large-scale urban scene inverse rendering from a single video. 3DV, 2025. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 422, + 555, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 422, + 555, + 465 + ], + "spans": [ + { + "bbox": [ + 317, + 422, + 555, + 465 + ], + "type": "text", + "content": "[27] Yu-Lun Liu, Wei-Sheng Lai, Yu-Sheng Chen, Yi-Lung Kao, Ming-Hsuan Yang, Yung-Yu Chuang, and Jia-Bin Huang. Single-image HDR reconstruction by learning to reverse the camera pipeline. In CVPR, 2020. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 468, + 555, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 468, + 555, + 520 + ], + "spans": [ + { + "bbox": [ + 317, + 468, + 555, + 520 + ], + "type": "text", + "content": "[28] Demetris Marnerides, Thomas Bashford-Rogers, Jonathan Hatchett, and Kurt Debattista. ExpandNet: A deep convolutional neural network for high dynamic range expansion from low dynamic range content. Computer Graphics Forum, 2018. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 523, + 555, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 523, + 555, + 566 + ], + "spans": [ + { + "bbox": [ + 317, + 523, + 555, + 566 + ], + "type": "text", + "content": "[29] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 568, + 555, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 568, + 555, + 610 + ], + "spans": [ + { + "bbox": [ + 317, + 568, + 555, + 610 + ], + "type": "text", + "content": "[30] Ben Mildenhall, Peter Hedman, Ricardo Martin-Brualla, Pratul Srinivasan, and Jonathan T. Barron. NeRF in the dark: High dynamic range view synthesis from noisy raw images. In CVPR, 2022. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 317, + 613, + 555, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 613, + 555, + 645 + ], + "spans": [ + { + "bbox": [ + 317, + 613, + 555, + 645 + ], + "type": "text", + "content": "[31] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM TOG, 2022. 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 647, + 555, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 555, + 689 + ], + "type": "text", + "content": "[32] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting triangular 3D models, materials, and lighting from images. In CVPR, 2022. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 692, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 554, + 713 + ], + "type": "text", + "content": "[33] Merlin Nimier-David, Zhao Dong, Wenzel Jakob, and Anton Kaplanyan. Material and lighting reconstruction for complex" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "473" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "indoor scenes with texture-space differentiable rendering. In Eurographics Symposium on Rendering, 2021. 2, 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 97, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 97, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 97, + 295, + 140 + ], + "type": "text", + "content": "[34] Julien Philip, Sébastien Morgenthaler, Michael Gharbi, and George Drettakis. Free-viewpoint indoor neural relighting from multi-view stereo. ACM Trans. Graph., 40(5):194:1-18, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 143, + 295, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 143, + 295, + 186 + ], + "spans": [ + { + "bbox": [ + 56, + 143, + 295, + 186 + ], + "type": "text", + "content": "[35] Erik Reinhard, Greg Ward, Sumanta Pattanaik, Paul Debevec, Wolfgang Heidrich, and Karol Myszkowski. High Dynamic Range Imaging - Acquisition, Display and Image-Based Lighting. Morgan Kaufmann, 2nd edition, 2010. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 189, + 295, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 189, + 295, + 243 + ], + "spans": [ + { + "bbox": [ + 56, + 189, + 295, + 243 + ], + "type": "text", + "content": "[36] Mike Roberts, Jason Ramapuram, Anurag Ranjan, Atulit Kumar, Miguel Angel Bautista, Nathan Paczan, Russ Webb, and Joshua M. Susskind. Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding. In ICCV, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 245, + 295, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 245, + 295, + 289 + ], + "spans": [ + { + "bbox": [ + 56, + 245, + 295, + 289 + ], + "type": "text", + "content": "[37] Dario Seyb, Peter-Pike Sloan, Ari Silvennoinen, Michal Iwanicki, and Wojciech Jarosz. The design and evolution of the UberBake light baking system. ACM Transactions on Graphics (TOG), 2020. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 291, + 295, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 291, + 295, + 334 + ], + "spans": [ + { + "bbox": [ + 56, + 291, + 295, + 334 + ], + "type": "text", + "content": "[38] Pratul P Srinivasan, Ben Mildenhall, Matthew Tancik, Jonathan T Barron, Richard Tucker, and Noah Snavely. Lighthouse: Predicting lighting volumes for spatially-coherent illumination. In CVPR, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 337, + 295, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 337, + 295, + 380 + ], + "spans": [ + { + "bbox": [ + 56, + 337, + 295, + 380 + ], + "type": "text", + "content": "[39] Pratul P. Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T. Barron. NeRV: Neural reflectance and visibility fields for relighting and view synthesis. In CVPR, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 383, + 295, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 295, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 295, + 426 + ], + "type": "text", + "content": "[40] Cheng Sun, Guangyan Cai, Zhengqin Li, Kai Yan, Cheng Zhang, Carl Marshall, Jia-Bin Huang, Shuang Zhao, and Zhao Dong. Neural-pbir reconstruction of shape, material, and illumination. In ICCV, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 429, + 295, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 429, + 295, + 461 + ], + "spans": [ + { + "bbox": [ + 56, + 429, + 295, + 461 + ], + "type": "text", + "content": "[41] Zian Wang, Jonah Philion, Sanja Fidler, and Jan Kautz. Learning indoor inverse rendering with 3D spatially-varying lighting. In ICCV, 2021. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 464, + 295, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 464, + 295, + 518 + ], + "spans": [ + { + "bbox": [ + 56, + 464, + 295, + 518 + ], + "type": "text", + "content": "[42] Zian Wang, Tianchang Shen, Jun Gao, Shengyu Huang, Jacob Munkberg, Jon Hasselgren, Zan Gojcic, Wenzheng Chen, and Sanja Fidler. Neural fields meet explicit geometric representations for inverse rendering of urban scenes. In CVPR, 2023. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 520, + 295, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 520, + 295, + 575 + ], + "spans": [ + { + "bbox": [ + 56, + 520, + 295, + 575 + ], + "type": "text", + "content": "[43] Liwen Wu, Rui Zhu, Mustafa B Yaldiz, Yinhao Zhu, Hong Cai, Janarbek Matai, Fatih Porikli, Tzu-Mao Li, Manmohan Chandraker, and Ravi Ramamoorthi. Factorized inverse path tracing for efficient and accurate material-lighting estimation. In ICCV, 2023. 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 577, + 295, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 577, + 295, + 610 + ], + "spans": [ + { + "bbox": [ + 56, + 577, + 295, + 610 + ], + "type": "text", + "content": "[44] Xin Yang, Ke Xu, Yibing Song, Qiang Zhang, Xiaopeng Wei, and Rynson WH Lau. Image correction via deep reciprocating HDR transformation. In CVPR, 2018. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 613, + 295, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 295, + 655 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 295, + 655 + ], + "type": "text", + "content": "[45] Yao Yao, Jingyang Zhang, Jingbo Liu, Yihang Qu, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. NeILF: Neural incident light field for physically-based material estimation. In ECCV, 2022. 2, 6, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 658, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 295, + 712 + ], + "type": "text", + "content": "[46] Lior Yariv, Peter Hedman, Christian Reiser, Dor Verbin, Pratul P Srinivasan, Richard Szeliski, Jonathan T Barron, and Ben Mildenhall. Bakedsdf: Meshing neural sdfs for real-time view synthesis. In ACM SIGGRAPH computer graphics, 2023. 4" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 554, + 498 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 316, + 72, + 554, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 554, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 554, + 106 + ], + "type": "text", + "content": "[47] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. ScanNet++: A high-fidelity dataset of 3D indoor scenes. In ICCV, 2023. 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 107, + 554, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 554, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 554, + 150 + ], + "type": "text", + "content": "[48] Bohan Yu, Siqi Yang, Xuanning Cui, Siyan Dong, Baoquan Chen, and Boxin Shi. MILO: Multi-bounce inverse rendering for indoor scene with light-emitting objects. IEEE TPAMI, 2023. 2, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 152, + 554, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 554, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 554, + 195 + ], + "type": "text", + "content": "[49] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. MonoSDF: Exploring monocular geometric cues for neural implicit surface reconstruction. In NeurIPS, 2022. 4, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 197, + 554, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 554, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 554, + 251 + ], + "type": "text", + "content": "[50] Zheng Zeng, Valentin Deschaintre, Iliyan Georgiev, Yannick Hold-Geoffroy, Yiwei Hu, Fujun Luan, Ling-Qi Yan, and Miloš Hasan. RGB ↔ X: Image decomposition and synthesis using material- and lighting-aware diffusion models. In SIGGRAPH, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 252, + 554, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 252, + 554, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 252, + 554, + 274 + ], + "type": "text", + "content": "[51] Jinsong Zhang and Jean-François Lalonde. Learning high dynamic range from outdoor panoramicas. In ICCV, 2017. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 276, + 554, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 554, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 554, + 319 + ], + "type": "text", + "content": "[52] Jingyang Zhang, Yao Yao, Shiwei Li, Jingbo Liu, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. NeILF++: Inter-reflectable light fields for geometry and material estimation. In ICCV, 2023. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 320, + 554, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 320, + 554, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 320, + 554, + 363 + ], + "type": "text", + "content": "[53] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snavely. PhySG: Inverse rendering with spherical Gaussians for physics-based material editing and relighting. In CVPR, 2021. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 365, + 554, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 554, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 554, + 408 + ], + "type": "text", + "content": "[54] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. NeR-Factor: Neural factorization of shape and reflectance under an unknown illumination. ACM TOG, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 410, + 554, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 410, + 554, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 410, + 554, + 453 + ], + "type": "text", + "content": "[55] Jingsen Zhu, Yuchi Huo, Qi Ye, Fujun Luan, Jifan Li, Dianbing Xi, Lisha Wang, Rui Tang, Wei Hua, Hujun Bao, and Rui Wang. I²-SDF: Intrinsic indoor scene reconstruction and editing via raytracing in neural SDFs. In CVPR, 2023. 2, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 455, + 554, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 455, + 554, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 455, + 554, + 498 + ], + "type": "text", + "content": "[56] Rui Zhu, Zhengqin Li, Janarbek Matai, Fatih Porikli, and Manmohan Chandraker. IRISformer: Dense vision transformers for single-image inverse rendering in indoor scenes. In CVPR, 2022. 2, 4" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "type": "text", + "content": "474" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/ITA-MDT_ Image-Timestep-Adaptive Masked Diffusion Transformer Framework for Image-Based Virtual Try-On/e6b13099-74a4-4157-8bc0-2e7511ad889f_content_list.json b/2025/ITA-MDT_ Image-Timestep-Adaptive Masked Diffusion Transformer Framework for Image-Based Virtual Try-On/e6b13099-74a4-4157-8bc0-2e7511ad889f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..dee33a977be5482dd252f4944e97befa570abdf1 --- /dev/null +++ b/2025/ITA-MDT_ Image-Timestep-Adaptive Masked Diffusion Transformer Framework for Image-Based Virtual Try-On/e6b13099-74a4-4157-8bc0-2e7511ad889f_content_list.json @@ -0,0 +1,1566 @@ +[ + { + "type": "text", + "text": "ITA-MDT: Image-Timestep-Adaptive Masked Diffusion Transformer Framework for Image-Based Virtual Try-On", + "text_level": 1, + "bbox": [ + 91, + 128, + 906, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ji Woo Hong $^{1}$ Tri Ton $^{1}$ Trung X. Pham $^{1}$ Gwanhyeong Koo $^{1}$ Sunjae Yoon $^{1}$ Chang D. Yoo $^{1}$ $^{1}$ Korea Advanced Institute of Science and Technology (KAIST), South Korea", + "bbox": [ + 99, + 202, + 906, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{jiwohong93, tritth, trungpx, kookie, dbstjswo505, cd_yoo}@kaist.ac.kr", + "bbox": [ + 187, + 241, + 816, + 255 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ffcbc1c8930d7110b713b648e9c1b8d78e62c74edc0193f164ce4ac02682992b.jpg", + "image_caption": [ + "Figure 1. Image-based Virtual Try-on results generated by our ITA-MDT. From left to right: VITON-HD[1], DressCode[20] upper, DressCode lower, and DressCode dresses, respectively." + ], + "image_footnote": [], + "bbox": [ + 91, + 290, + 905, + 434 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 484, + 326, + 500 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper introduces ITA-MDT, the Image-Timestep-Adaptive Masked Diffusion Transformer Framework for Image-Based Virtual Try-On (IVTON), designed to overcome the limitations of previous approaches by leveraging the Masked Diffusion Transformer (MDT) for improved handling of both global garment context and fine-grained details. The IVTON task involves seamlessly superimposing a garment from one image onto a person in another, creating a realistic depiction of the person wearing the specified garment. Unlike conventional diffusion-based virtual try-on models that depend on large pre-trained U-Net architectures, ITA-MDT leverages a lightweight, scalable transformer-based denoising diffusion model with a mask latent modeling scheme, achieving competitive results while reducing computational overhead. A key component of ITAMDT is the Image-Timestep Adaptive Feature Aggregator (ITAFA), a dynamic feature aggregator that combines all of the features from the image encoder into a unified feature of the same size, guided by diffusion timestep and garment image complexity. This enables adaptive weighting of features, allowing the model to emphasize either global information or fine-grained details based on the requirements of the denoising stage. Additionally, the Salient Region Extractor (SRE) module is presented to identify complex region of the garment to provide high-resolution local infor", + "bbox": [ + 88, + 523, + 485, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "mation to the denoising model as an additional condition alongside the global information of the full garment image. This targeted conditioning strategy enhances detail preservation of fine details in highly salient garment regions, optimizing computational resources by avoiding unnecessarily processing entire garment image. Comparative evaluations confirms that ITA-MDT improves efficiency while maintaining strong performance, reaching state-of-the-art results in several metrics. Our project page is available at https://jiwoohong93.github.io/ita-mdt/.", + "bbox": [ + 511, + 487, + 908, + 638 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 662, + 643, + 679 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The task of Image-based Virtual Try-on (IVTON) is to generate an image where a person appears dressed in specific clothes, using provided images of both the garment and the individual. In real-world applications, IVTON holds great potential for revolutionizing online shopping, fashion design [3, 9, 14, 32], and even entertainment industries for creating digital avatars or virtual costumes [19, 27, 39]. For this task, deep learning methods [1, 6, 10, 15, 18, 36, 38, 40] employ paired datasets [1, 20] containing both standalone garment images and the images of the individual wearing these for training their models, along with additional reference images such as mask images of the garment region and DensePose [8] image.", + "bbox": [ + 509, + 688, + 908, + 885 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, diffusion models [13, 34] have shown great", + "bbox": [ + 531, + 885, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "28284", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "promise in image generation and synthesis in many different tasks [17, 25, 42-44]. Especially, large-scale pre-trained diffusion models [28, 29, 31], such as Stable Diffusion [29], are widely used to preserving high fidelity in generated images; however, their large size and slow inference speeds present practical limitations in commercial application.", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this, transformer-based diffusion models have emerged as an efficient alternative. The Diffusion Transformer (DiT) [24] based on the Visual Transformer (ViT) [4] architecture with Latent Diffusion Models [29] can achieve excellent scalability and efficiency. Extending this concept, the Masked Diffusion Transformer (MDT) [5] applied mask prediction scheme to enhance spatial representation learning, achieving state-of-the-art results in class-conditional image generation. While promising, directly applying MDT to IVTON task presents challenges, as the task requires high fidelity and consistency across multiple reference images, demanding more than a straightforward approach.", + "bbox": [ + 89, + 184, + 482, + 366 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This paper presents the ITA-MDT (Image-Timestep-Adaptive Masked Diffusion Transformer) Framework, which incorporates our modified version of MDT adapted specifically for IVTON, named MDT-IVTON, as well as two novel components: the Image-Timestep Adaptive Feature Aggregator (ITAFA) and the Salient Region Extractor (SRE). This framework is designed to preserve garment fidelity while maintaining computational efficiency. Our MDT-IVTON model leverages MDT's mask prediction scheme to efficiently learn spatial relationships between garment and individual. Additionally, we incorporate an auxiliary loss to enhance the model's focus on the inpainting objective of the task. While this architecture effectively captures global dependencies, it struggles to retain fine-grained details, which U-Net-based models handle relatively naturally through their multi-scale encoder-decoder structure that captures and reconstructs information of the image across multiple scales through progressive downsampling and upsampling. Our ITAFA and SRE bridge this gap.", + "bbox": [ + 89, + 368, + 482, + 656 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The Image-Timestep Adaptive Feature Aggregator (ITAFA) is a learnable feature aggregator designed to condition the diffusion model effectively and efficiently. It dynamically combines multiple image features from ViT image encoder, guided by the current diffusion timestep and garment image complexity score. While simply providing multiple image features from different layers can condition the model with extensive information, this naive approach significantly increases the computational load. By assigning adaptive weights to each feature, ITAFA aggregates them into a single representation. These weights are modulated based on the timestep information, which indicates how far the generation process has progressed. The weighting is further refined using the visual complexity of the garment, which reflects how much fine-grained information the condition image contains. This design enables the model to gradually shift its focus", + "bbox": [ + 89, + 659, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/38953b5a0f8d37d04c01a469fda50e5a301ce13bb0c7ea02d0e5949abba92f5f.jpg", + "image_caption": [ + "Figure 2. Comparison of our ITA-MDT with other diffusion-based methods on the VTON task using the VITON-HD dataset. For all metrics in the analysis, including FID, model parameters, and inference speed, lower values indicate better performance." + ], + "image_footnote": [], + "bbox": [ + 516, + 101, + 903, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "from global semantic features in the early stages to fine local details in the later stages of the diffusion process.", + "bbox": [ + 511, + 354, + 903, + 383 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The Salient Region Extractor (SRE) provides high-resolution local information of key garment area, such as logos or complex patterns, identified through entropy-based detection. The region extracted offers additional crucial condition guidance to the diffusion model alongside the global context from the full garment image. While simply using a high-resolution, big size, image enhances detail representation, it results in substantial computation. Notably, most regions of garments are either solid-colored or display large, simple patterns that can be captured adequately through lower resolution. Therefore, rather than uniformly processing the entire image at high resolution, SRE selectively extracts and provides additional high-resolution information only for regions where fine-grained local information is crucial. In summary, our contributions include:", + "bbox": [ + 511, + 386, + 906, + 611 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We present ITA-MDT, a framework for the image-based virtual try-on (IVTON) task that builds upon MDT-IVTON, our adaptation of the Masked Diffusion Transformer (MDT) for IVTON, and incorporates our proposed methods to effectively address the challenges of transformer-based denoising diffusion models in preserving fine-grained details.", + "2. We propose ITAFA, a learnable feature aggregator designed to dynamically combine features from the image encoder into a single feature based on the current denoising timestep and garment image complexity score.", + "3. We propose SRE to provide high-resolution local information of key garment area, identified through entropy-based detection, as additional condition guidance to the diffusion model alongside the global context from the full garment image.", + "4. ITA-MDT's combination of MDT-IVTON, ITAFA, and SRE offers a highly efficient solution for IVTON, achieving a balance between model efficiency and output fi" + ], + "bbox": [ + 514, + 613, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "28285", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "delity. Figure 2 presents the comparison of our ITA-MDT with previous diffusion-based methods. It achieves state-of-the-art results in several metrics while maintaining competitive performance across the others, demonstrating its overall effectiveness and suitability for practical virtual try-on applications.", + "bbox": [ + 109, + 90, + 486, + 183 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 196, + 232, + 213 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Image-based Virtual Try-on with Diffusion", + "text_level": 1, + "bbox": [ + 89, + 223, + 455, + 239 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The integration of diffusion models has significantly advanced the field of Image-based Virtual Try-on, as evidenced in numerous recent studies. Notably, TryOnDiffusion [47] introduced an architecture utilizing dual U-Nets, highlighting the potential of diffusion-based try-on but requiring large-scale datasets composed of image pairs of the same person in varied poses. This large-scale dataset requirement has led researchers to pivot towards leveraging large pre-trained diffusion models [28, 29, 31]. To condition these models effectively, methods such as representing garments as pseudowords in LaDI-VTON [21], applying warping networks in DCI-VTON [7], modifying attention mechanisms in StableVITON [16] and IDM-VTON [2], utilizing ControlNet for garment control and GAN-based sampling in CAT-DM [45], and concatenating masked person and garment images for direct texture transfer in TPD [41] have been introduced. However, the use of large pre-trained U-Nets comes with notable limitations. The large number of parameters in these models significantly increases memory usage and slows down inference speed, making them resource-intensive and less practical for efficient deployment. A variety of solutions could be explored to address these issues, among which transformer-based diffusion models present a promising approach.", + "bbox": [ + 89, + 246, + 485, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Transformers for Diffusion Process", + "text_level": 1, + "bbox": [ + 89, + 621, + 397, + 636 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Traditionally, The U-Net architecture [30] has been foundational for diffusion models and continues to be a prevalent choice across various diffusion-based generation tasks [13, 35]. However, the incorporation of transformers into diffusion processes also marked a significant evolution in generative modeling. The introduction of the Diffusion Transformer (DiT) [24] merged the architecture of the Vision Transformer (ViT) [4] with latent diffusion models. DiT has demonstrated outstanding scalability and compute efficiency, delivering superior generative performance compared to traditional U-Net architectures. Building on this, the Masked Diffusion Transformer (MDT) [5] has further advanced the diffusion transformer model by employing an asymmetric masking schedule to optimize the contextual representation learning, setting a new benchmark in class image generation on ImageNet. In an attempt to leverage MDT for image-to-image generation, [25] introduced an aggregation network", + "bbox": [ + 89, + 643, + 485, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "that consolidates image conditions into a single vector for guiding the diffusion process in pose-guided human image generation tasks. While effective within its domain, directly applying this approach to the IVTON task is unsuitable. IVTON involves a significantly larger set of condition images, and compressing these into a single vector leads to substantial information loss, ultimately degrading output quality. Given the need to preserve fine details from condition images, enhancing rather than compressing condition information is preferred.", + "bbox": [ + 511, + 90, + 908, + 243 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Enhancing Condition Information in Virtual Try-On task", + "text_level": 1, + "bbox": [ + 511, + 250, + 906, + 282 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To enhance the intricate details essential to IVTON, the recent method, PICTURE [22], introduces a two-stage disentanglement pipeline that separates style and texture features by employing a hierarchical CLIP feature extraction module with position encoding, enabling the representation of complex, non-stationary textures. While effective in handling intricate style and texture combinations, this method requires an in-depth analysis of features from the hidden layers of the image encoder. Specifically, this involves clustering hidden layer features and manually selecting representative features from each cluster, a process that lacks flexibility for it requires re-evaluation when a different encoder is to be used. Additionally, the selected features from each cluster are concatenated for cross-attention conditioning, substantially increasing computational demand. Together, these requirements render the approach resource-intensive and less adaptable to diverse architectures.", + "bbox": [ + 511, + 287, + 908, + 544 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 511, + 556, + 604, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The ITA-MDT framework performs Image-Based Virtual Try-on through a transformer-based denoising diffusion network, leveraging the mask latent modeling scheme of MDTv2 [5]. An overview of the architecture is shown in Figure 3, which comprises three primary components: (1) the Masked Diffusion Transformer for Image-Based Virtual Try-On (MDT-IVTON), (2) the Image-Timestep Adaptive Feature Aggregator (ITAFA), and (3) the Salient Region Extractor (SRE) module. Detailed descriptions of each component are provided below.", + "bbox": [ + 511, + 582, + 908, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Masked Diffusion Transformer for Image-Based Virtual Try-On (MDT-IVTON)", + "text_level": 1, + "bbox": [ + 511, + 742, + 906, + 772 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our ITA-MDT framework integrates a Masked Diffusion Transformer (MDT) tailored for Image-Based Virtual Try-On (IVTON), leveraging the transformer-based denoising diffusion model architecture of DiT [24]. This model operates in a latent space encoded via a pre-trained VAE following the Latent Diffusion Model (LDM) approach [29]. Specifically, the VAE encoder maps garment agnostic inputs into this latent space, including the Garment Agnostic", + "bbox": [ + 511, + 779, + 910, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "28286", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e23e4b455c2609c241fb95d6d7598db11cf16d0a32ee3dbbd0d3056454d1452d.jpg", + "image_caption": [ + "Figure 3. Overview of the ITA-MDT Framework for Image-Based Virtual Try-On (IVTON). The framework takes multiple reference images encoded into latent space as query, which includes Garment Agnostic Map $A$ , DensePose $P$ , and Garment Agnostic Mask $M_X$ . These reference latent images are concatenated to be patch embedded for masked diffusion process within our MDT-IVTON, which follows the architecture of MDTv2 [5] with integrated cross-attention blocks. The image feature of Garment $X$ is extracted with the ViT image encoder, DINOv2 [23], and are adaptively aggregated with our proposed Image-Timestep Adaptive Feature Aggregator (ITAFA) to produce Garment Feature $F_g$ . With our Salient Region Extractor (SRE), the Salient Region $X_s$ is extracted from the Garment $X$ and processed through ITAFA separately to produce Salient Feature $F_s$ . The Garment Feature $F_g$ and Salient Feature $F_s$ are concatenated to serve as conditions of MDT-IVTON. Positional embeddings are omitted in this figure for simplicity." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 903, + 343 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Map $A$ , DensePose $P$ , and Garment Agnostic Mask $M_X$ , concatenated with the noised latent representation $z_t$ .", + "bbox": [ + 88, + 489, + 482, + 520 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The MDT-IVTON model processes the input by embedding it into a patchified latent representation $L$ and sequentially passing it through $N_{1}$ Encoder Blocks followed by $N_{2}$ Decoder Blocks. Each block utilizes Layer Normalization, Multi-Head Self-Attention, Multi-Head Cross Attention, Multi-Layer Perceptron (MLP), and skip connections to refine the denoised latent representation $z_{t - 1}$ , as illustrated in the right side of Figure 3. Conditioning information derived from the concatenated garment features $F_{g}$ and $F_{s}$ , extracted from garment image $X$ and its salient region $X_{s}$ using the ViT encoder and ITAFA module, is infused through cross-attention mechanisms to guide the generation process. Detailed descriptions of ITAFA and SRE are available in Sec. 3.2 and Sec. 3.3.", + "bbox": [ + 88, + 522, + 482, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Three objective functions guide the MDT-IVTON training to optimize its generative performance:", + "bbox": [ + 89, + 737, + 482, + 767 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1.1. Denoising Objective", + "text_level": 1, + "bbox": [ + 89, + 784, + 277, + 800 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The primary objective function minimizes the mean squared error (MSE) between the predicted noise and the actual noise in the denoised latent $z_{t}$ at each timestep $t$ , following the standard diffusion objective:", + "bbox": [ + 89, + 806, + 482, + 867 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {d e n o i s e}} = \\mathcal {E} _ {y, c, \\epsilon \\sim \\mathcal {N} (0, I), t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(y _ {t}, c, t\\right) \\| ^ {2} \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 883, + 482, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $y_{t}$ denotes the latent representation at timestep $t$ , $c$ the conditioning, and $\\epsilon_{\\theta}$ the model's predicted noise.", + "bbox": [ + 511, + 489, + 903, + 520 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1.2. Mask Reconstruction Objective", + "text_level": 1, + "bbox": [ + 511, + 526, + 779, + 541 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In MDT-IVTON, the Mask Reconstruction Objective enforces spatial coherence and semantic consistency by reconstructing masked tokens within the masked latent representation $E_{m}$ . This objective ensures alignment between the reference and generated images, enhancing the model's ability to synthesize realistic outputs in masked regions.", + "bbox": [ + 511, + 544, + 906, + 635 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To achieve this, we incorporate the Side-Interpolator module, which plays a pivotal role in interpolating information across unmasked tokens to predict values for masked tokens in $E_{m}$ . The Side-Interpolator effectively fills in missing information by conditioning on neighboring unmasked regions, encouraging consistent token-level synthesis within the masked latent.", + "bbox": [ + 511, + 635, + 906, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Mask Reconstruction Objective follows an MSE formulation similar to the denoising objective:", + "bbox": [ + 511, + 741, + 906, + 771 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {m a s k}} = \\mathcal {E} _ {y, c, \\epsilon \\sim \\mathcal {N} (0, I), t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (\\text {S i d e - I n t} (E _ {m}), c, t) \\| ^ {2} \\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 777, + 906, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where Side-Int processes $E_{m}$ , leveraging spatial context to predict masked regions. The conditioning $c$ includes garment details, aiding in preserving fidelity to the garment appearance in masked areas. This objective promotes detailed reconstruction by interpolating across the context in $E_{m}$ , enhancing semantic alignment within generated images.", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "28287", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1.3. Inpainting Objective", + "text_level": 1, + "bbox": [ + 89, + 90, + 281, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The inpainting objective in MDT-IVTON is designed to enhance garment fidelity by concentrating on regions defined by the garment agnostic mask $M_X$ , encoded as $\\mathcal{E}(M_X)$ in latent space. During training, this mask serves as a selective weighting factor in the mean squared error (MSE) between the target and predicted outputs, guiding the model's focus to reconstruct details more accurately within garment-specific regions.", + "bbox": [ + 89, + 109, + 483, + 229 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To ensure numerical stability and maintain well-behaved gradients, the mask is normalized to the range [0, 1], preventing extreme values that could otherwise destabilize training. This targeted inpainting strategy enhances garment detail fidelity by directing the model's attention to areas where high-quality reconstruction is most critical.", + "bbox": [ + 89, + 231, + 483, + 320 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Latent Mask Normalization: To prevent unstable gradients, the single channel of the latent mask $\\mathcal{E}(M_X)$ that most closely resembles the binary mask of the inpainting task is chosen to be used as the mask across all channels in the loss calculation. This selected channel is normalized to values between 0 and 1 to ensure stability:", + "bbox": [ + 89, + 321, + 483, + 411 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {E} \\left(M _ {X}\\right) = \\frac {\\mathcal {E} \\left(M _ {X}\\right) - \\operatorname* {m i n} \\left(\\mathcal {E} \\left(M _ {X}\\right)\\right)}{\\operatorname* {m a x} \\left(\\mathcal {E} \\left(M _ {X}\\right)\\right) - \\operatorname* {m i n} \\left(\\mathcal {E} \\left(M _ {X}\\right)\\right) + 1 \\times 1 0 ^ {- 8}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 419, + 483, + 463 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $1 \\times 10^{-8}$ ensures numerical stability. This approach avoids unintended gradients and aligns the loss with garment-focused regions. A visualization of the latent mask selected is in the Supplementary Materials.", + "bbox": [ + 89, + 464, + 483, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Masked Loss Computation: With the rescaled single channel of $\\mathcal{E}(M_X)$ , the masked inpainting loss is calculated as a weighted mean squared error (MSE) applied uniformly across all channels in the latent space:", + "bbox": [ + 89, + 525, + 483, + 585 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {i n p a i n t}} = \\mathcal {E} _ {y, c, t} \\left[ \\| \\mathcal {E} (M _ {X}) \\cdot \\left(y _ {t} - y _ {t - 1}\\right) \\| ^ {2} \\right], \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 592, + 483, + 612 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $y_{t}$ and $y_{t-1}$ denote the latent representations at timesteps $t$ and $t-1$ , respectively. This loss term emphasizes reconstruction quality in the masked regions, promoting accurate garment-specific detail representation. Further details of MDT-IVTON are provided in Supplementary Materials.", + "bbox": [ + 89, + 619, + 483, + 695 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Image-Timestep Adaptive Feature Aggregator (ITAFA)", + "text_level": 1, + "bbox": [ + 89, + 702, + 480, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The ITAFA is a dynamic feature aggregation module designed to adaptively weight feature embeddings from a Vision Transformer (ViT) image encoder, utilizing learnable parameters and feature complexity analysis to dynamically modulate the contribution of each hidden layer output of the image encoder at different timesteps of the denoising diffusion process.", + "bbox": [ + 89, + 739, + 483, + 845 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.1. Components and Parameters", + "text_level": 1, + "bbox": [ + 89, + 851, + 339, + 866 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Timestep Embedding Projector. The diffusion timestep embedding, denoted as $T$ , represents the current timestep in", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the denoising process with dimension $\\mathcal{H}$ of the MDT-IVTON model. To align $T$ with the image feature dimension, it is projected to the dimensionality of the hidden layer of the ViT, $\\mathcal{H}_{\\varepsilon}$ . This is achieved through a linear transformation:", + "bbox": [ + 511, + 90, + 905, + 151 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nW ^ {T} = \\operatorname {L i n e a r} (\\mathcal {H}, \\mathcal {H} _ {\\mathcal {E}}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 157, + 906, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The linear layer maps the timestep embedding into a representation suitable for interacting with the ViT features.", + "bbox": [ + 511, + 183, + 906, + 214 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Image Complexity Projector. ITAFA incorporates three complexity components: feature sparsity, feature variance, and gradient magnitude. These collectively capture the complexity of garment texture and structure.", + "bbox": [ + 511, + 231, + 906, + 291 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "1. Feature Sparsity measures the proportion of near-zero activations in the feature embedding tensor $f$ , providing insight into the sparsity of learned representations. It is computed as:", + "bbox": [ + 514, + 292, + 905, + 352 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {S} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\mathbb {I} \\left(\\left| f _ {i} \\right| < \\delta\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 640, + 359, + 906, + 398 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\delta$ is a small threshold (i.e., 0.01), $n$ denotes the total number of elements in $f$ , and $\\mathbb{I}(\\cdot)$ is the indicator function.", + "bbox": [ + 531, + 405, + 905, + 449 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2. Feature Variance captures the variability of activations, reflecting structural complexity within the image:", + "bbox": [ + 511, + 450, + 905, + 481 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nV = \\operatorname {V a r} (f), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 674, + 489, + 906, + 506 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\operatorname{Var}(f)$ denotes the variance across all elements in the feature embedding.", + "bbox": [ + 531, + 513, + 905, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3. Gradient Magnitude measures local changes in the feature embedding, calculated by taking the spatial gradients along patch and embedding dimensions. The magnitude of gradients are computed, and averaging them provides a single score $G$ that captures texture complexity.", + "bbox": [ + 511, + 542, + 906, + 619 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nG = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\sqrt {\\left(\\frac {\\partial f}{\\partial x}\\right) ^ {2} + \\left(\\frac {\\partial f}{\\partial y}\\right) ^ {2}}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 630, + 906, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "These components are concatenated to form the complexity score vector $[S,V,G]$ , which is then linearly projected to match the hidden layer dimension $\\mathcal{H}_{\\mathcal{E}}$ :", + "bbox": [ + 511, + 676, + 905, + 720 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nW ^ {I} = \\operatorname {L i n e a r} (3, \\mathcal {H} _ {\\mathcal {E}}). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 728, + 906, + 746 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Learnable Weighting Parameter. The contribution of the timestep and complexity projections is balanced by a learnable parameter $\\alpha \\in [0,1]$ , initialized to 0.5 and optimized during training:", + "bbox": [ + 511, + 762, + 906, + 823 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nW ^ {c o m} = \\alpha \\cdot W ^ {T} + (1 - \\alpha) \\cdot W ^ {I}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 593, + 829, + 906, + 848 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This weighted combination provides an adaptive control over the influence of timestep and image complexity on the image feature aggregation.", + "bbox": [ + 511, + 854, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "28288", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/549b3d146ae9897987a9bc98cd0747da5e80a6044a7b7e51870a33efd1cdd583.jpg", + "image_caption": [ + "Image-Timestep Adaptive Feature Aggregator" + ], + "image_footnote": [], + "bbox": [ + 94, + 109, + 428, + 313 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7c2c18b4ff01c3d191f55a76bf5e65eb51c240fbc2873b81856d965658da3040.jpg", + "image_caption": [ + "Salient Region Extractor", + "Figure 4. Overview of the Image-Timestep Adaptive Feature Aggregator (ITAFA) and the Salient Region Extractor (SRE). Left: The ITAFA dynamically aggregates Vision Transformer (ViT) feature embeddings $f$ based on a combination of the Timestep Embedding Projector, which projects the diffusion timestep embedding $T$ to match the feature embedding dimensions, and the Image Complexity Projector, which transforms the image complexity vector $[S, V, G]$ (sparsity, variance, gradient magnitude) into a comparable dimension. The weight vectors are combined and normalized via softmax to form $W$ , which is used to adaptively aggregate the feature embeddings $\\{f_i\\}_{i=0}^h$ across hidden layers to produce the final output tensor $F$ . Garment $X$ and Salient Region $X_s$ are processed separately through ITAFA to generate the Garment Feature $F_g$ and Salient Feature $F_s$ . Right: The SRE processes the input image $X$ by computing an entropy map $X_e$ , creating a binary High-Entropy Mask $X_m$ , and circular region expansion from the entropy centroid to extract the final high-entropy region $X_s$ , ensuring preservation of detail within a consistent aspect ratio." + ], + "image_footnote": [], + "bbox": [ + 436, + 109, + 903, + 311 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2.2. Feature Aggregation", + "text_level": 1, + "bbox": [ + 89, + 477, + 279, + 491 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The aggregated weights are normalized via softmax to ensure they sum to one across hidden layers, thereby controlling the contribution of each layers.", + "bbox": [ + 89, + 494, + 482, + 541 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nW = \\operatorname {s o f t m a x} \\left(W ^ {\\text {c o m}}\\right). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 551, + 482, + 569 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "These weights are then applied to each hidden layer output from the ViT encoder, denoted by $f$ , via an element-wise multiplication and summation across layers:", + "bbox": [ + 89, + 580, + 482, + 626 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nF = \\sum_ {l = 1} ^ {\\mathcal {H}} W [ l ] \\cdot f [ l ], \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 636, + 482, + 678 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $F$ represents the final aggregated feature tensor, capturing a balanced representation based on both denoising timestep and image complexities.", + "bbox": [ + 89, + 688, + 483, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In-depth details of ITAFA are explained in Supplementary Materials.", + "bbox": [ + 89, + 733, + 482, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3. Salient Region Extractor (SRE)", + "text_level": 1, + "bbox": [ + 89, + 772, + 369, + 789 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our SRE is a simple yet effective algorithm designed to identify and extract salient, high-entropy, areas within an image. The method includes: Entropy Map Computation, High-Entropy Region Identification, Entropy Centroid Localization, Circular Region Expansion, and Region Extraction.", + "bbox": [ + 89, + 794, + 483, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "By utilizing Shannon Entropy [33] to compute an entropy map $X_{e}$ , which measures information content across pixel", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "neighborhoods. High-entropy regions are then identified by thresholding this map to create a binary High-Entropy Mask $X_{m}$ that highlights areas with substantial visual complexity. The centroid of this region is computed to serve as the initial center for region extraction. From this centroid, a circular region expansion method grows the bounding box outward in small increments, checking for low-entropy areas along each edge as stopping point to ensure meaningful content within the boundary. After the expansion, the region is expanded to follow the aspect ratio of the original image, then extracted as the final output $X_{s}$ , capturing detailed areas while maintaining minimal distortion.", + "bbox": [ + 511, + 476, + 906, + 656 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Details of SRE process and algorithm are elaborated in Supplementary Materials.", + "bbox": [ + 511, + 657, + 905, + 689 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 705, + 645, + 723 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details: We use the MDTv2 XL model [5] with RGB reference images of shape $512 \\times 512 \\times 3$ , encoded into latent $Z$ -space via the Stable Diffusion XL VAE [26]. These latent features are patch-embedded and processed within our model. For conditioning of garment features, the ViT encoder extracts features from RGB images of shape $224 \\times 224 \\times 3$ . Training was conducted on two 80GB A100 GPUs with a batch size of 6. Further details are provided in the Supplementary Materials.", + "bbox": [ + 511, + 732, + 906, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets and Evaluation Metrics: We train and evaluate our model on two public datasets: VITON-HD [1] and", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "28289", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/eb79d031d2bde39e04e0397257f637b4afdf9ef6ef7e7c67968d420d572ae3e1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodInf. Time (s)↓Param. (M)↓LPIPS↓VITON-HDDressCodeUpper
SSIM↑FID(pair/unp.)↓LPIPS↓SSIM↑FID(pair/unp.)↓
LaDI-VTON [21]4.2258650.0980.8816.587/9.2650.0620.92710.447/14.048
DCI-VTON [7]15.4088590.0850.8965.559/8.7500.0680.91810.501/12.490
StableVITON [16]13.3571,0390.0860.8727.039/9.7170.0660.91911.572/13.666
CAT-DM [45]4.9478590.1040.8557.697/10.2630.0580.92210.257/14.039
TPD [41]28.7848590.0970.88510.161/13.0710.0480.9569.500/12.904
IDM-VTON [2]12.2882,9910.0820.8816.000/9.1630.0360.9447.471/11.915
MDT-IVTON (Base)4.5966710.0990.8877.773/9.8380.0610.92110.300/13.998
Base+ITAFA4.5966710.0860.8815.717/8.8790.0340.9515.414/13.111
Base+ITAFA+HR6.2156710.0830.8845.542/8.8000.0340.9515.384/12.528
Base+ITAFA+SRE4.5966710.0830.8855.462/8.6760.0340.9515.412/10.069
", + "bbox": [ + 91, + 88, + 919, + 292 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Quantitative results of diffusion-based virtual try-on methods on VITON-HD and DressCode upper-body test sets. Inf. Time refers to the inference time, diffusion model only, measured for generating a single image on our environment, and Param. refers to the denoising diffusion model's parameter size. We report both paired and unpaired FID. Following the previous works, feature of last hidden layer of ViT image encoder is used for our MDT-IVTON without ITAFA. HR refers to use of single High-resolution $(448\\times 448\\times 3)$ garment image to formulate condition vector $c$ . The best performance for each metric is indicated in bold, while the second best is underlined.", + "bbox": [ + 88, + 301, + 906, + 376 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "DressCode [20] upper, lower, and dresses. Data augmentation techniques from StableVITON [16] are applied to enhance model robustness. To assess performance, we use standard metrics from previous VTON works: LPIPS [46], SSIM [37], and FID [11]. All evaluations follow the protocol established in LaDI-VTON [21] to ensure fair comparison with previous virtual try-on models.", + "bbox": [ + 88, + 398, + 480, + 503 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Further information about the datasets and evaluation metrics is provided in the Supplementary Materials.", + "bbox": [ + 89, + 505, + 482, + 537 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1. Qualitative and Quantitative Results", + "text_level": 1, + "bbox": [ + 89, + 544, + 408, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The quantitative Results of Table. 1 compare our proposed ITA-MDT framework to previous diffusion-based methods, as well as show effect of ITAFA, and SRE on MDT-IVTON. The values of the previous works are re-measured by implementation with their checkpoints released. As it can be observed, our method provides comparable, state-of-the-art, results with faster inference speed and smaller parameters. The qualitative results and comparison to previous methods are shown in Fig. 6, and more in Supplementary Materials. As part of the inpainting process, the generated images are blended with agnostic images.", + "bbox": [ + 88, + 566, + 482, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 742, + 243, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The qualitative results of our proposed ITA-MDT framework and effects of each components are available in the Supplementary Materials. MDT-IVTON alone captures the global structure well, including garment shapes, colors, and human pose, but falls short in fidelity of garment details. The integration of ITAFA and SRE addresses this limitation significantly, with SRE shown as effective as a high-resolution image with twice the width compared to SRE integrated condition vector $c$ .", + "bbox": [ + 88, + 763, + 482, + 898 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/00868f013a1d495a31cfa7a11db618e0e3dd5fd19d4fc9ca92f613af4ffa65fc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 397, + 702, + 508 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c6b5400ceb27fdf9dfe7f00157c034af81f05d48f713ea4120fccbe5e50cac72.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 705, + 398, + 901, + 507 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6081b1391c5c01fb49c668eb543e7a828b6fc8896d8d8d7dc41b6e7863c1be4a.jpg", + "image_caption": [ + "Figure 5. Visualization of the ITAFA's average weight distribution across ViT encoder layers of our final model. (a): use of hidden layers as denoising progresses. The dotted arrow refers to denoising direction. (b, c, d): changes in the use of hidden layers as the sparsity, variance, and gradient magnitude of garment image changes. The dotted arrow refers to the increase in these image complexity components." + ], + "image_footnote": [], + "bbox": [ + 517, + 510, + 697, + 621 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8cd65d56bb3fa52d9f682e2699dcfaf20426fec0df2b454b3d5cdca43d824c11.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 700, + 508, + 903, + 622 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 5 visualizes how ITAFA leverages features from different layers of the ViT image encoder, DINOv2, across the denoising process. In Fig. 5 (a), where it observes the use of layers with respect to timesteps, the middle layers dominate the feature aggregation for much of the diffusion; however, a distinct shift in layer usage occurs from early to later layers as the timestep progresses. In the early timestep when the input is largely noise, the encoder's first layer is prominently utilized, and at the final timestep, the use of the", + "bbox": [ + 511, + 763, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "28290", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/067efeaff30710115ae707c1c8b3e8b953aab6375882236c3bff8468ef12a788.jpg", + "image_caption": [ + "Figure 6. Qualitative comparison of ITA-MDT with prior methods on VITON-HD and DressCode Upper-body datasets. The first three rows show VITON-HD results; the next three show DressCode results." + ], + "image_footnote": [], + "bbox": [ + 89, + 85, + 906, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "last layer markedly increases. This observation supports our hypothesis that dynamically aggregating features from different layers as the diffusion progresses effectively optimizes feature utilization across denoising stages. In Fig. 5 (b, c, d), we observe a clear pattern where specific layers are consistently favored depending on the intensity of particular image complexity (e.g., feature sparsity, variance, or gradient magnitude). This indicates that our model adaptively activates layers containing information most relevant to the type of visual complexity, highlighting the value and adaptability of our ITAFA.", + "bbox": [ + 89, + 719, + 485, + 885 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 718, + 633, + 734 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper presented ITA-MDT, a framework for the IVTON task that successfully leverages MDT, along with the proposed ITAFA and SRE modules, to address the challenges of transformer-based denoising diffusion models in preserving fine-grained details efficiently. Analysis confirmed that ITAFA and SRE effectively function as intended, enabling ITA-MDT to achieve state-of-the-art performance in several metrics with reduced model parameters and faster inference speed compared to previous methods.", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "28291", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by Institute for Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No.RS-2021-II211381, Development of Causal AI through Video Understanding and Reinforcement Learning, and Its Applications to Real Environments) and partly supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No.RS-2022-II220184, Development and Study of AI Technologies to Inexpensively Conform to Evolving Policy on Ethics)", + "bbox": [ + 89, + 114, + 485, + 282 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 292, + 187, + 308 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Seunghwan Choi, Sunghyun Park, Minsoo Lee, and Jaegul Choo. Viton-hd: High-resolution virtual try-on via misalignment-aware normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14131-14140, 2021. 1, 6", + "[2] Yisol Choi, Sangkyung Kwak, Kyungmin Lee, Hyungwon Choi, and Jinwoo Shin. Improving diffusion models for virtual try-on. arXiv preprint arXiv:2403.05139, 2024. 3, 7", + "[3] Lavinia De Divitiis, Federico Becattini, Claudio Baecchi, and Alberto Del Bimbo. Disentangling features for fashion recommendation. ACM Transactions on Multimedia Computing, Communications and Applications, 19(1s):1-21, 2023. 1", + "[4] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 3", + "[5] Shanghua Gao, Pan Zhou, Ming-Ming Cheng, and Shuicheng Yan. Masked diffusion transformer is a strong image synthesizer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23164-23173, 2023. 2, 3, 4, 6", + "[6] Yuying Ge, Yibing Song, Ruimao Zhang, Chongjian Ge, Wei Liu, and Ping Luo. Parser-free virtual try-on via distilling appearance flows. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8485-8493, 2021. 1", + "[7] Junhong Gou, Siyu Sun, Jianfu Zhang, Jianlou Si, Chen Qian, and Liqing Zhang. Taming the power of diffusion models for high-quality virtual try-on with appearance flow. In Proceedings of the 31st ACM International Conference on Multimedia, pages 7599-7607, 2023. 3, 7", + "[8] Ríza Alp Güler, Natalia Neverova, and Iasonas Kokkinos. Denseseposé: Dense human pose estimation in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7297-7306, 2018. 1", + "[9] M Hadi Kiapour, Xufeng Han, Svetlana Lazebnik, Alexander C Berg, and Tamara L Berg. Where to buy it: Matching street clothing photos in online shops. In Proceedings of the IEEE international conference on computer vision, pages 3343-3351, 2015. 1" + ], + "bbox": [ + 99, + 316, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[10] Xintong Han, Zuxuan Wu, Zhe Wu, Ruichi Yu, and Larry S Davis. Viton: An image-based virtual try-on network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7543-7552, 2018. 1", + "[11] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 7", + "[12] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 3", + "[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 3", + "[14] Wei-Lin Hsiao and Kristen Grauman. Creating capsule wardrobes from fashion images. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7161-7170, 2018. 1", + "[15] Thibaut Issenhuth, Jérémie Mary, and Clément Calauzenes. Do not mask what you do not need to mask: a parser-free virtual try-on. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XX 16, pages 619–635. Springer, 2020. 1", + "[16] Jeongho Kim, Guojung Gu, Minho Park, Sunghyun Park, and Jaegul Choo. Stableviton: Learning semantic correspondence with latent diffusion model for virtual try-on. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8176-8185, 2024. 3, 7", + "[17] Gwanhyeong Koo, Sunjae Yoon, Ji Woo Hong, and Chang D Yoo. Flexied: Frequency-aware latent refinement for enhanced non-rigid editing. In European Conference on Computer Vision, pages 363-379. Springer, 2024. 2", + "[18] Sangyun Lee, Gyojung Gu, Sunghyun Park, Seunghwan Choi, and Jaegul Choo. High-resolution virtual try-on with misalignment and occlusion-handled conditions. In European Conference on Computer Vision, pages 204-219. Springer, 2022. 1", + "[19] Shugao Ma, Tomas Simon, Jason Saragih, Dawei Wang, Yuecheng Li, Fernando De La Torre, and Yaser Sheikh. Pixel codec avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 64-73, 2021. 1", + "[20] Davide Morelli, Matteo Fincato, Marcella Cornia, Federico Landi, Fabio Cesari, and Rita Cucchiara. Dress code: high-resolution multi-category virtual try-on. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2231-2235, 2022. 1, 7", + "[21] Davide Morelli, Alberto Baldrati, Giuseppe Cartella, Marcelia Cornia, Marco Bertini, and Rita Cucchiara. Ladi-vton: latent diffusion textual-inversion enhanced virtual try-on. In Proceedings of the 31st ACM International Conference on Multimedia, pages 8580-8589, 2023. 3, 7", + "[22] Shuliang Ning, Duomin Wang, Yipeng Qin, Zirong Jin, Baoyuan Wang, and Xiaoguang Han. Picture: Photorealistic virtual try-on from unconstrained designs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6976-6985, 2024. 3" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "28292", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[23] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 4", + "[24] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 2, 3, 4", + "[25] Trung X Pham, Zhang Kang, and Chang D Yoo. Cross-view masked diffusion transformers for person image synthesis. arXiv preprint arXiv:2402.01516, 2024. 2, 3", + "[26] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 6, 4", + "[27] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM international conference on multimedia, pages 484-492, 2020. 1", + "[28] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International conference on machine learning, pages 8821-8831. Pmlr, 2021. 2, 3", + "[29] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3", + "[30] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 3", + "[31] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 2, 3", + "[32] Rohan Sarkar, Navaneeth Bodla, Mariya I Vasileva, Yen-Liang Lin, Anurag Beniwal, Alan Lu, and Gerard Medioni. Outfittransformer: Learning outfit representations for fashion recommendation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3601-3609, 2023. 1", + "[33] Claude Elwood Shannon. A mathematical theory of communication. The Bell system technical journal, 27(3):379-423, 1948. 6, 2", + "[34] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1", + "[35] Yang Song and Stefano Ermon. Generative modeling by" + ], + "bbox": [ + 91, + 90, + 485, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019. 3", + "[36] Bochao Wang, Huabin Zheng, Xiaodan Liang, Yimin Chen, Liang Lin, and Meng Yang. Toward characteristic-preserving image-based virtual try-on network. In Proceedings of the European conference on computer vision (ECCV), pages 589-604, 2018. 1", + "[37] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 7", + "[38] Zhenyu Xie, Zaiyu Huang, Xin Dong, Fuwei Zhao, Haoye Dong, Xijin Zhang, Feida Zhu, and Xiaodan Liang. Gp-vton: Towards general purpose virtual try-on via collaborative localflow global-parsing learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23550-23559, 2023. 1", + "[39] Sicheng Xu, Guojun Chen, Yu-Xiao Guo, Jiaolong Yang, Chong Li, Zhenyu Zang, Yizhong Zhang, Xin Tong, and Baining Guo. Vasa-1: Lifelike audio-driven talking faces generated in real time. arXiv preprint arXiv:2404.10667, 2024. 1", + "[40] Han Yang, Ruimao Zhang, Xiaobao Guo, Wei Liu, Wangmeng Zuo, and Ping Luo. Towards photo-realistic virtual try-on by adaptively generating-preserving image content. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7850-7859, 2020. 1", + "[41] Xu Yang, Changxing Ding, Zhibin Hong, Junhao Huang, Jin Tao, and Xiangmin Xu. Texture-preserving diffusion models for high-fidelity virtual try-on. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7017-7026, 2024. 3, 7", + "[42] Sunjae Yoon, Gwanhyeong Koo, Ji Woo Hong, and Chang D Yoo. Dni: Dilutional noise initialization for diffusion video editing. In European Conference on Computer Vision, pages 180-195. Springer, 2024. 2", + "[43] Sunjae Yoon, Gwanhyeong Koo, Geonwoo Kim, and Chang D Yoo. Frag: Frequency adapting group for diffusion video editing. arXiv preprint arXiv:2406.06044, 2024.", + "[44] Sunjae Yoon, Gwanhyeong Koo, Younghwan Lee, and Chang Yoo. Tpc: Test-time procrustes calibration for diffusion-based human image animation. Advances in Neural Information Processing Systems, 37:118654-118677, 2024. 2", + "[45] Jianhao Zeng, Dan Song, Weizhi Nie, Hongshuo Tian, Tongtong Wang, and An-An Liu. Cat-dm: Controllable accelerated virtual try-on with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8372-8382, 2024. 3, 7", + "[46] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7", + "[47] Luyang Zhu, Dawei Yang, Tyler Zhu, Fitsum Reda, William Chan, Chitwan Sahara, Mohammad Norouzi, and Ira Kemelmacher-Shlizerman. Tryondiffusion: A tale of two unets. In Proceedings of the IEEE/CVF Conference on Com" + ], + "bbox": [ + 516, + 92, + 906, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "28293", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "puter Vision and Pattern Recognition, pages 4606-4615, 2023. 3", + "bbox": [ + 122, + 90, + 483, + 118 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "28294", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/ITA-MDT_ Image-Timestep-Adaptive Masked Diffusion Transformer Framework for Image-Based Virtual Try-On/e6b13099-74a4-4157-8bc0-2e7511ad889f_model.json b/2025/ITA-MDT_ Image-Timestep-Adaptive Masked Diffusion Transformer Framework for Image-Based Virtual Try-On/e6b13099-74a4-4157-8bc0-2e7511ad889f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..cc5f4076081216cabdb8d249fbfc3f755f0f14d6 --- /dev/null +++ b/2025/ITA-MDT_ Image-Timestep-Adaptive Masked Diffusion Transformer Framework for Image-Based Virtual Try-On/e6b13099-74a4-4157-8bc0-2e7511ad889f_model.json @@ -0,0 +1,2114 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.13, + 0.908, + 0.177 + ], + "angle": 0, + "content": "ITA-MDT: Image-Timestep-Adaptive Masked Diffusion Transformer Framework for Image-Based Virtual Try-On" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.203, + 0.908, + 0.24 + ], + "angle": 0, + "content": "Ji Woo Hong\\(^{1}\\) Tri Ton\\(^{1}\\) Trung X. Pham\\(^{1}\\) Gwanhyeong Koo\\(^{1}\\) Sunjae Yoon\\(^{1}\\) Chang D. Yoo\\(^{1}\\) \n\\(^{1}\\)Korea Advanced Institute of Science and Technology (KAIST), South Korea" + }, + { + "type": "text", + "bbox": [ + 0.189, + 0.242, + 0.817, + 0.256 + ], + "angle": 0, + "content": "{jiwohong93, tritth, trungpx, kookie, dbstjswo505, cd_yoo}@kaist.ac.kr" + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.291, + 0.906, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.445, + 0.908, + 0.473 + ], + "angle": 0, + "content": "Figure 1. Image-based Virtual Try-on results generated by our ITA-MDT. From left to right: VITON-HD[1], DressCode[20] upper, DressCode lower, and DressCode dresses, respectively." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.485, + 0.327, + 0.501 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.524, + 0.486, + 0.902 + ], + "angle": 0, + "content": "This paper introduces ITA-MDT, the Image-Timestep-Adaptive Masked Diffusion Transformer Framework for Image-Based Virtual Try-On (IVTON), designed to overcome the limitations of previous approaches by leveraging the Masked Diffusion Transformer (MDT) for improved handling of both global garment context and fine-grained details. The IVTON task involves seamlessly superimposing a garment from one image onto a person in another, creating a realistic depiction of the person wearing the specified garment. Unlike conventional diffusion-based virtual try-on models that depend on large pre-trained U-Net architectures, ITA-MDT leverages a lightweight, scalable transformer-based denoising diffusion model with a mask latent modeling scheme, achieving competitive results while reducing computational overhead. A key component of ITAMDT is the Image-Timestep Adaptive Feature Aggregator (ITAFA), a dynamic feature aggregator that combines all of the features from the image encoder into a unified feature of the same size, guided by diffusion timestep and garment image complexity. This enables adaptive weighting of features, allowing the model to emphasize either global information or fine-grained details based on the requirements of the denoising stage. Additionally, the Salient Region Extractor (SRE) module is presented to identify complex region of the garment to provide high-resolution local infor" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.488, + 0.909, + 0.639 + ], + "angle": 0, + "content": "mation to the denoising model as an additional condition alongside the global information of the full garment image. This targeted conditioning strategy enhances detail preservation of fine details in highly salient garment regions, optimizing computational resources by avoiding unnecessarily processing entire garment image. Comparative evaluations confirms that ITA-MDT improves efficiency while maintaining strong performance, reaching state-of-the-art results in several metrics. Our project page is available at https://jiwoohong93.github.io/ita-mdt/." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.664, + 0.645, + 0.68 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.689, + 0.909, + 0.886 + ], + "angle": 0, + "content": "The task of Image-based Virtual Try-on (IVTON) is to generate an image where a person appears dressed in specific clothes, using provided images of both the garment and the individual. In real-world applications, IVTON holds great potential for revolutionizing online shopping, fashion design [3, 9, 14, 32], and even entertainment industries for creating digital avatars or virtual costumes [19, 27, 39]. For this task, deep learning methods [1, 6, 10, 15, 18, 36, 38, 40] employ paired datasets [1, 20] containing both standalone garment images and the images of the individual wearing these for training their models, along with additional reference images such as mask images of the garment region and DensePose [8] image." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.886, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Recently, diffusion models [13, 34] have shown great" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28284" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.183 + ], + "angle": 0, + "content": "promise in image generation and synthesis in many different tasks [17, 25, 42-44]. Especially, large-scale pre-trained diffusion models [28, 29, 31], such as Stable Diffusion [29], are widely used to preserving high fidelity in generated images; however, their large size and slow inference speeds present practical limitations in commercial application." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.185, + 0.483, + 0.367 + ], + "angle": 0, + "content": "To address this, transformer-based diffusion models have emerged as an efficient alternative. The Diffusion Transformer (DiT) [24] based on the Visual Transformer (ViT) [4] architecture with Latent Diffusion Models [29] can achieve excellent scalability and efficiency. Extending this concept, the Masked Diffusion Transformer (MDT) [5] applied mask prediction scheme to enhance spatial representation learning, achieving state-of-the-art results in class-conditional image generation. While promising, directly applying MDT to IVTON task presents challenges, as the task requires high fidelity and consistency across multiple reference images, demanding more than a straightforward approach." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.369, + 0.483, + 0.657 + ], + "angle": 0, + "content": "This paper presents the ITA-MDT (Image-Timestep-Adaptive Masked Diffusion Transformer) Framework, which incorporates our modified version of MDT adapted specifically for IVTON, named MDT-IVTON, as well as two novel components: the Image-Timestep Adaptive Feature Aggregator (ITAFA) and the Salient Region Extractor (SRE). This framework is designed to preserve garment fidelity while maintaining computational efficiency. Our MDT-IVTON model leverages MDT's mask prediction scheme to efficiently learn spatial relationships between garment and individual. Additionally, we incorporate an auxiliary loss to enhance the model's focus on the inpainting objective of the task. While this architecture effectively captures global dependencies, it struggles to retain fine-grained details, which U-Net-based models handle relatively naturally through their multi-scale encoder-decoder structure that captures and reconstructs information of the image across multiple scales through progressive downsampling and upsampling. Our ITAFA and SRE bridge this gap." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.901 + ], + "angle": 0, + "content": "The Image-Timestep Adaptive Feature Aggregator (ITAFA) is a learnable feature aggregator designed to condition the diffusion model effectively and efficiently. It dynamically combines multiple image features from ViT image encoder, guided by the current diffusion timestep and garment image complexity score. While simply providing multiple image features from different layers can condition the model with extensive information, this naive approach significantly increases the computational load. By assigning adaptive weights to each feature, ITAFA aggregates them into a single representation. These weights are modulated based on the timestep information, which indicates how far the generation process has progressed. The weighting is further refined using the visual complexity of the garment, which reflects how much fine-grained information the condition image contains. This design enables the model to gradually shift its focus" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.102, + 0.905, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.27, + 0.907, + 0.327 + ], + "angle": 0, + "content": "Figure 2. Comparison of our ITA-MDT with other diffusion-based methods on the VTON task using the VITON-HD dataset. For all metrics in the analysis, including FID, model parameters, and inference speed, lower values indicate better performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.355, + 0.905, + 0.385 + ], + "angle": 0, + "content": "from global semantic features in the early stages to fine local details in the later stages of the diffusion process." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.387, + 0.908, + 0.612 + ], + "angle": 0, + "content": "The Salient Region Extractor (SRE) provides high-resolution local information of key garment area, such as logos or complex patterns, identified through entropy-based detection. The region extracted offers additional crucial condition guidance to the diffusion model alongside the global context from the full garment image. While simply using a high-resolution, big size, image enhances detail representation, it results in substantial computation. Notably, most regions of garments are either solid-colored or display large, simple patterns that can be captured adequately through lower resolution. Therefore, rather than uniformly processing the entire image at high resolution, SRE selectively extracts and provides additional high-resolution information only for regions where fine-grained local information is crucial. In summary, our contributions include:" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.614, + 0.908, + 0.72 + ], + "angle": 0, + "content": "1. We present ITA-MDT, a framework for the image-based virtual try-on (IVTON) task that builds upon MDT-IVTON, our adaptation of the Masked Diffusion Transformer (MDT) for IVTON, and incorporates our proposed methods to effectively address the challenges of transformer-based denoising diffusion models in preserving fine-grained details." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.72, + 0.908, + 0.781 + ], + "angle": 0, + "content": "2. We propose ITAFA, a learnable feature aggregator designed to dynamically combine features from the image encoder into a single feature based on the current denoising timestep and garment image complexity score." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.781, + 0.908, + 0.856 + ], + "angle": 0, + "content": "3. We propose SRE to provide high-resolution local information of key garment area, identified through entropy-based detection, as additional condition guidance to the diffusion model alongside the global context from the full garment image." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.856, + 0.908, + 0.901 + ], + "angle": 0, + "content": "4. ITA-MDT's combination of MDT-IVTON, ITAFA, and SRE offers a highly efficient solution for IVTON, achieving a balance between model efficiency and output fi" + }, + { + "type": "list", + "bbox": [ + 0.515, + 0.614, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "28285" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.11, + 0.092, + 0.488, + 0.184 + ], + "angle": 0, + "content": "delity. Figure 2 presents the comparison of our ITA-MDT with previous diffusion-based methods. It achieves state-of-the-art results in several metrics while maintaining competitive performance across the others, demonstrating its overall effectiveness and suitability for practical virtual try-on applications." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.198, + 0.233, + 0.214 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.224, + 0.456, + 0.24 + ], + "angle": 0, + "content": "2.1. Image-based Virtual Try-on with Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.247, + 0.486, + 0.61 + ], + "angle": 0, + "content": "The integration of diffusion models has significantly advanced the field of Image-based Virtual Try-on, as evidenced in numerous recent studies. Notably, TryOnDiffusion [47] introduced an architecture utilizing dual U-Nets, highlighting the potential of diffusion-based try-on but requiring large-scale datasets composed of image pairs of the same person in varied poses. This large-scale dataset requirement has led researchers to pivot towards leveraging large pre-trained diffusion models [28, 29, 31]. To condition these models effectively, methods such as representing garments as pseudowords in LaDI-VTON [21], applying warping networks in DCI-VTON [7], modifying attention mechanisms in StableVITON [16] and IDM-VTON [2], utilizing ControlNet for garment control and GAN-based sampling in CAT-DM [45], and concatenating masked person and garment images for direct texture transfer in TPD [41] have been introduced. However, the use of large pre-trained U-Nets comes with notable limitations. The large number of parameters in these models significantly increases memory usage and slows down inference speed, making them resource-intensive and less practical for efficient deployment. A variety of solutions could be explored to address these issues, among which transformer-based diffusion models present a promising approach." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.622, + 0.398, + 0.637 + ], + "angle": 0, + "content": "2.2. Transformers for Diffusion Process" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Traditionally, The U-Net architecture [30] has been foundational for diffusion models and continues to be a prevalent choice across various diffusion-based generation tasks [13, 35]. However, the incorporation of transformers into diffusion processes also marked a significant evolution in generative modeling. The introduction of the Diffusion Transformer (DiT) [24] merged the architecture of the Vision Transformer (ViT) [4] with latent diffusion models. DiT has demonstrated outstanding scalability and compute efficiency, delivering superior generative performance compared to traditional U-Net architectures. Building on this, the Masked Diffusion Transformer (MDT) [5] has further advanced the diffusion transformer model by employing an asymmetric masking schedule to optimize the contextual representation learning, setting a new benchmark in class image generation on ImageNet. In an attempt to leverage MDT for image-to-image generation, [25] introduced an aggregation network" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.909, + 0.244 + ], + "angle": 0, + "content": "that consolidates image conditions into a single vector for guiding the diffusion process in pose-guided human image generation tasks. While effective within its domain, directly applying this approach to the IVTON task is unsuitable. IVTON involves a significantly larger set of condition images, and compressing these into a single vector leads to substantial information loss, ultimately degrading output quality. Given the need to preserve fine details from condition images, enhancing rather than compressing condition information is preferred." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.251, + 0.907, + 0.283 + ], + "angle": 0, + "content": "2.3. Enhancing Condition Information in Virtual Try-On task" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.288, + 0.909, + 0.545 + ], + "angle": 0, + "content": "To enhance the intricate details essential to IVTON, the recent method, PICTURE [22], introduces a two-stage disentanglement pipeline that separates style and texture features by employing a hierarchical CLIP feature extraction module with position encoding, enabling the representation of complex, non-stationary textures. While effective in handling intricate style and texture combinations, this method requires an in-depth analysis of features from the hidden layers of the image encoder. Specifically, this involves clustering hidden layer features and manually selecting representative features from each cluster, a process that lacks flexibility for it requires re-evaluation when a different encoder is to be used. Additionally, the selected features from each cluster are concatenated for cross-attention conditioning, substantially increasing computational demand. Together, these requirements render the approach resource-intensive and less adaptable to diverse architectures." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.558, + 0.606, + 0.573 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.909, + 0.735 + ], + "angle": 0, + "content": "The ITA-MDT framework performs Image-Based Virtual Try-on through a transformer-based denoising diffusion network, leveraging the mask latent modeling scheme of MDTv2 [5]. An overview of the architecture is shown in Figure 3, which comprises three primary components: (1) the Masked Diffusion Transformer for Image-Based Virtual Try-On (MDT-IVTON), (2) the Image-Timestep Adaptive Feature Aggregator (ITAFA), and (3) the Salient Region Extractor (SRE) module. Detailed descriptions of each component are provided below." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.743, + 0.907, + 0.773 + ], + "angle": 0, + "content": "3.1. Masked Diffusion Transformer for Image-Based Virtual Try-On (MDT-IVTON)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Our ITA-MDT framework integrates a Masked Diffusion Transformer (MDT) tailored for Image-Based Virtual Try-On (IVTON), leveraging the transformer-based denoising diffusion model architecture of DiT [24]. This model operates in a latent space encoded via a pre-trained VAE following the Latent Diffusion Model (LDM) approach [29]. Specifically, the VAE encoder maps garment agnostic inputs into this latent space, including the Garment Agnostic" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28286" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.905, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.352, + 0.908, + 0.466 + ], + "angle": 0, + "content": "Figure 3. Overview of the ITA-MDT Framework for Image-Based Virtual Try-On (IVTON). The framework takes multiple reference images encoded into latent space as query, which includes Garment Agnostic Map \\( A \\), DensePose \\( P \\), and Garment Agnostic Mask \\( M_X \\). These reference latent images are concatenated to be patch embedded for masked diffusion process within our MDT-IVTON, which follows the architecture of MDTv2 [5] with integrated cross-attention blocks. The image feature of Garment \\( X \\) is extracted with the ViT image encoder, DINOv2 [23], and are adaptively aggregated with our proposed Image-Timestep Adaptive Feature Aggregator (ITAFA) to produce Garment Feature \\( F_g \\). With our Salient Region Extractor (SRE), the Salient Region \\( X_s \\) is extracted from the Garment \\( X \\) and processed through ITAFA separately to produce Salient Feature \\( F_s \\). The Garment Feature \\( F_g \\) and Salient Feature \\( F_s \\) are concatenated to serve as conditions of MDT-IVTON. Positional embeddings are omitted in this figure for simplicity." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.49, + 0.483, + 0.521 + ], + "angle": 0, + "content": "Map \\(A\\), DensePose \\(P\\), and Garment Agnostic Mask \\(M_X\\), concatenated with the noised latent representation \\(z_t\\)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.523, + 0.483, + 0.734 + ], + "angle": 0, + "content": "The MDT-IVTON model processes the input by embedding it into a patchified latent representation \\( L \\) and sequentially passing it through \\( N_{1} \\) Encoder Blocks followed by \\( N_{2} \\) Decoder Blocks. Each block utilizes Layer Normalization, Multi-Head Self-Attention, Multi-Head Cross Attention, Multi-Layer Perceptron (MLP), and skip connections to refine the denoised latent representation \\( z_{t - 1} \\), as illustrated in the right side of Figure 3. Conditioning information derived from the concatenated garment features \\( F_{g} \\) and \\( F_{s} \\), extracted from garment image \\( X \\) and its salient region \\( X_{s} \\) using the ViT encoder and ITAFA module, is infused through cross-attention mechanisms to guide the generation process. Detailed descriptions of ITAFA and SRE are available in Sec. 3.2 and Sec. 3.3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.738, + 0.483, + 0.768 + ], + "angle": 0, + "content": "Three objective functions guide the MDT-IVTON training to optimize its generative performance:" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.785, + 0.279, + 0.801 + ], + "angle": 0, + "content": "3.1.1. Denoising Objective" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.807, + 0.483, + 0.868 + ], + "angle": 0, + "content": "The primary objective function minimizes the mean squared error (MSE) between the predicted noise and the actual noise in the denoised latent \\( z_{t} \\) at each timestep \\( t \\), following the standard diffusion objective:" + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.884, + 0.483, + 0.904 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {d e n o i s e}} = \\mathcal {E} _ {y, c, \\epsilon \\sim \\mathcal {N} (0, I), t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(y _ {t}, c, t\\right) \\| ^ {2} \\right], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.491, + 0.905, + 0.521 + ], + "angle": 0, + "content": "where \\(y_{t}\\) denotes the latent representation at timestep \\(t\\), \\(c\\) the conditioning, and \\(\\epsilon_{\\theta}\\) the model's predicted noise." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.527, + 0.78, + 0.542 + ], + "angle": 0, + "content": "3.1.2. Mask Reconstruction Objective" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.545, + 0.908, + 0.636 + ], + "angle": 0, + "content": "In MDT-IVTON, the Mask Reconstruction Objective enforces spatial coherence and semantic consistency by reconstructing masked tokens within the masked latent representation \\( E_{m} \\). This objective ensures alignment between the reference and generated images, enhancing the model's ability to synthesize realistic outputs in masked regions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.636, + 0.908, + 0.74 + ], + "angle": 0, + "content": "To achieve this, we incorporate the Side-Interpolator module, which plays a pivotal role in interpolating information across unmasked tokens to predict values for masked tokens in \\( E_{m} \\). The Side-Interpolator effectively fills in missing information by conditioning on neighboring unmasked regions, encouraging consistent token-level synthesis within the masked latent." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.742, + 0.908, + 0.772 + ], + "angle": 0, + "content": "The Mask Reconstruction Objective follows an MSE formulation similar to the denoising objective:" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.778, + 0.907, + 0.81 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {m a s k}} = \\mathcal {E} _ {y, c, \\epsilon \\sim \\mathcal {N} (0, I), t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (\\text {S i d e - I n t} (E _ {m}), c, t) \\| ^ {2} \\right], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "where Side-Int processes \\( E_{m} \\), leveraging spatial context to predict masked regions. The conditioning \\( c \\) includes garment details, aiding in preserving fidelity to the garment appearance in masked areas. This objective promotes detailed reconstruction by interpolating across the context in \\( E_{m} \\), enhancing semantic alignment within generated images." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28287" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.092, + 0.282, + 0.107 + ], + "angle": 0, + "content": "3.1.3. Inpainting Objective" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.11, + 0.484, + 0.231 + ], + "angle": 0, + "content": "The inpainting objective in MDT-IVTON is designed to enhance garment fidelity by concentrating on regions defined by the garment agnostic mask \\(M_X\\), encoded as \\(\\mathcal{E}(M_X)\\) in latent space. During training, this mask serves as a selective weighting factor in the mean squared error (MSE) between the target and predicted outputs, guiding the model's focus to reconstruct details more accurately within garment-specific regions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.232, + 0.484, + 0.321 + ], + "angle": 0, + "content": "To ensure numerical stability and maintain well-behaved gradients, the mask is normalized to the range [0, 1], preventing extreme values that could otherwise destabilize training. This targeted inpainting strategy enhances garment detail fidelity by directing the model's attention to areas where high-quality reconstruction is most critical." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.322, + 0.484, + 0.412 + ], + "angle": 0, + "content": "Latent Mask Normalization: To prevent unstable gradients, the single channel of the latent mask \\(\\mathcal{E}(M_X)\\) that most closely resembles the binary mask of the inpainting task is chosen to be used as the mask across all channels in the loss calculation. This selected channel is normalized to values between 0 and 1 to ensure stability:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.42, + 0.484, + 0.464 + ], + "angle": 0, + "content": "\\[\n\\mathcal {E} \\left(M _ {X}\\right) = \\frac {\\mathcal {E} \\left(M _ {X}\\right) - \\operatorname* {m i n} \\left(\\mathcal {E} \\left(M _ {X}\\right)\\right)}{\\operatorname* {m a x} \\left(\\mathcal {E} \\left(M _ {X}\\right)\\right) - \\operatorname* {m i n} \\left(\\mathcal {E} \\left(M _ {X}\\right)\\right) + 1 \\times 1 0 ^ {- 8}}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.465, + 0.484, + 0.525 + ], + "angle": 0, + "content": "where \\(1 \\times 10^{-8}\\) ensures numerical stability. This approach avoids unintended gradients and aligns the loss with garment-focused regions. A visualization of the latent mask selected is in the Supplementary Materials." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.526, + 0.484, + 0.587 + ], + "angle": 0, + "content": "Masked Loss Computation: With the rescaled single channel of \\(\\mathcal{E}(M_X)\\), the masked inpainting loss is calculated as a weighted mean squared error (MSE) applied uniformly across all channels in the latent space:" + }, + { + "type": "equation", + "bbox": [ + 0.141, + 0.593, + 0.484, + 0.613 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {i n p a i n t}} = \\mathcal {E} _ {y, c, t} \\left[ \\| \\mathcal {E} (M _ {X}) \\cdot \\left(y _ {t} - y _ {t - 1}\\right) \\| ^ {2} \\right], \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.62, + 0.484, + 0.696 + ], + "angle": 0, + "content": "where \\( y_{t} \\) and \\( y_{t-1} \\) denote the latent representations at timesteps \\( t \\) and \\( t-1 \\), respectively. This loss term emphasizes reconstruction quality in the masked regions, promoting accurate garment-specific detail representation. Further details of MDT-IVTON are provided in Supplementary Materials." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.703, + 0.482, + 0.734 + ], + "angle": 0, + "content": "3.2. Image-Timestep Adaptive Feature Aggregator (ITAFA)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.74, + 0.484, + 0.846 + ], + "angle": 0, + "content": "The ITAFA is a dynamic feature aggregation module designed to adaptively weight feature embeddings from a Vision Transformer (ViT) image encoder, utilizing learnable parameters and feature complexity analysis to dynamically modulate the contribution of each hidden layer output of the image encoder at different timesteps of the denoising diffusion process." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.852, + 0.341, + 0.867 + ], + "angle": 0, + "content": "3.2.1. Components and Parameters" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Timestep Embedding Projector. The diffusion timestep embedding, denoted as \\( T \\), represents the current timestep in" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.152 + ], + "angle": 0, + "content": "the denoising process with dimension \\(\\mathcal{H}\\) of the MDT-IVTON model. To align \\(T\\) with the image feature dimension, it is projected to the dimensionality of the hidden layer of the ViT, \\(\\mathcal{H}_{\\varepsilon}\\). This is achieved through a linear transformation:" + }, + { + "type": "equation", + "bbox": [ + 0.631, + 0.159, + 0.907, + 0.177 + ], + "angle": 0, + "content": "\\[\nW ^ {T} = \\operatorname {L i n e a r} (\\mathcal {H}, \\mathcal {H} _ {\\mathcal {E}}), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.184, + 0.908, + 0.215 + ], + "angle": 0, + "content": "The linear layer maps the timestep embedding into a representation suitable for interacting with the ViT features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.232, + 0.907, + 0.292 + ], + "angle": 0, + "content": "Image Complexity Projector. ITAFA incorporates three complexity components: feature sparsity, feature variance, and gradient magnitude. These collectively capture the complexity of garment texture and structure." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.293, + 0.906, + 0.353 + ], + "angle": 0, + "content": "1. Feature Sparsity measures the proportion of near-zero activations in the feature embedding tensor \\( f \\), providing insight into the sparsity of learned representations. It is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.642, + 0.361, + 0.907, + 0.4 + ], + "angle": 0, + "content": "\\[\n\\mathrm {S} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\mathbb {I} \\left(\\left| f _ {i} \\right| < \\delta\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.406, + 0.906, + 0.45 + ], + "angle": 0, + "content": "where \\(\\delta\\) is a small threshold (i.e., 0.01), \\(n\\) denotes the total number of elements in \\(f\\), and \\(\\mathbb{I}(\\cdot)\\) is the indicator function." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.451, + 0.906, + 0.482 + ], + "angle": 0, + "content": "2. Feature Variance captures the variability of activations, reflecting structural complexity within the image:" + }, + { + "type": "equation", + "bbox": [ + 0.675, + 0.49, + 0.907, + 0.507 + ], + "angle": 0, + "content": "\\[\nV = \\operatorname {V a r} (f), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.514, + 0.906, + 0.544 + ], + "angle": 0, + "content": "where \\(\\operatorname{Var}(f)\\) denotes the variance across all elements in the feature embedding." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.544, + 0.907, + 0.62 + ], + "angle": 0, + "content": "3. Gradient Magnitude measures local changes in the feature embedding, calculated by taking the spatial gradients along patch and embedding dimensions. The magnitude of gradients are computed, and averaging them provides a single score \\( G \\) that captures texture complexity." + }, + { + "type": "equation", + "bbox": [ + 0.603, + 0.631, + 0.907, + 0.673 + ], + "angle": 0, + "content": "\\[\nG = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\sqrt {\\left(\\frac {\\partial f}{\\partial x}\\right) ^ {2} + \\left(\\frac {\\partial f}{\\partial y}\\right) ^ {2}}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.677, + 0.906, + 0.722 + ], + "angle": 0, + "content": "These components are concatenated to form the complexity score vector \\([S,V,G]\\), which is then linearly projected to match the hidden layer dimension \\(\\mathcal{H}_{\\mathcal{E}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.635, + 0.729, + 0.907, + 0.747 + ], + "angle": 0, + "content": "\\[\nW ^ {I} = \\operatorname {L i n e a r} (3, \\mathcal {H} _ {\\mathcal {E}}). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.763, + 0.907, + 0.824 + ], + "angle": 0, + "content": "Learnable Weighting Parameter. The contribution of the timestep and complexity projections is balanced by a learnable parameter \\(\\alpha \\in [0,1]\\), initialized to 0.5 and optimized during training:" + }, + { + "type": "equation", + "bbox": [ + 0.594, + 0.83, + 0.907, + 0.849 + ], + "angle": 0, + "content": "\\[\nW ^ {c o m} = \\alpha \\cdot W ^ {T} + (1 - \\alpha) \\cdot W ^ {I}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.902 + ], + "angle": 0, + "content": "This weighted combination provides an adaptive control over the influence of timestep and image complexity on the image feature aggregation." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28288" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.098, + 0.095, + 0.388, + 0.108 + ], + "angle": 0, + "content": "Image-Timestep Adaptive Feature Aggregator" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.111, + 0.429, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.63, + 0.095, + 0.786, + 0.108 + ], + "angle": 0, + "content": "Salient Region Extractor" + }, + { + "type": "image", + "bbox": [ + 0.437, + 0.111, + 0.904, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.325, + 0.907, + 0.451 + ], + "angle": 0, + "content": "Figure 4. Overview of the Image-Timestep Adaptive Feature Aggregator (ITAFA) and the Salient Region Extractor (SRE). Left: The ITAFA dynamically aggregates Vision Transformer (ViT) feature embeddings \\( f \\) based on a combination of the Timestep Embedding Projector, which projects the diffusion timestep embedding \\( T \\) to match the feature embedding dimensions, and the Image Complexity Projector, which transforms the image complexity vector \\( [S, V, G] \\) (sparsity, variance, gradient magnitude) into a comparable dimension. The weight vectors are combined and normalized via softmax to form \\( W \\), which is used to adaptively aggregate the feature embeddings \\( \\{f_i\\}_{i=0}^h \\) across hidden layers to produce the final output tensor \\( F \\). Garment \\( X \\) and Salient Region \\( X_s \\) are processed separately through ITAFA to generate the Garment Feature \\( F_g \\) and Salient Feature \\( F_s \\). Right: The SRE processes the input image \\( X \\) by computing an entropy map \\( X_e \\), creating a binary High-Entropy Mask \\( X_m \\), and circular region expansion from the entropy centroid to extract the final high-entropy region \\( X_s \\), ensuring preservation of detail within a consistent aspect ratio." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.478, + 0.281, + 0.492 + ], + "angle": 0, + "content": "3.2.2. Feature Aggregation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.496, + 0.483, + 0.542 + ], + "angle": 0, + "content": "The aggregated weights are normalized via softmax to ensure they sum to one across hidden layers, thereby controlling the contribution of each layers." + }, + { + "type": "equation", + "bbox": [ + 0.209, + 0.553, + 0.483, + 0.57 + ], + "angle": 0, + "content": "\\[\nW = \\operatorname {s o f t m a x} \\left(W ^ {\\text {c o m}}\\right). \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.581, + 0.483, + 0.627 + ], + "angle": 0, + "content": "These weights are then applied to each hidden layer output from the ViT encoder, denoted by \\( f \\), via an element-wise multiplication and summation across layers:" + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.637, + 0.483, + 0.679 + ], + "angle": 0, + "content": "\\[\nF = \\sum_ {l = 1} ^ {\\mathcal {H}} W [ l ] \\cdot f [ l ], \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.485, + 0.734 + ], + "angle": 0, + "content": "where \\(F\\) represents the final aggregated feature tensor, capturing a balanced representation based on both denoising timestep and image complexities." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.734, + 0.483, + 0.763 + ], + "angle": 0, + "content": "In-depth details of ITAFA are explained in Supplementary Materials." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.773, + 0.37, + 0.79 + ], + "angle": 0, + "content": "3.3. Salient Region Extractor (SRE)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.871 + ], + "angle": 0, + "content": "Our SRE is a simple yet effective algorithm designed to identify and extract salient, high-entropy, areas within an image. The method includes: Entropy Map Computation, High-Entropy Region Identification, Entropy Centroid Localization, Circular Region Expansion, and Region Extraction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "By utilizing Shannon Entropy [33] to compute an entropy map \\( X_{e} \\), which measures information content across pixel" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.477, + 0.907, + 0.657 + ], + "angle": 0, + "content": "neighborhoods. High-entropy regions are then identified by thresholding this map to create a binary High-Entropy Mask \\( X_{m} \\) that highlights areas with substantial visual complexity. The centroid of this region is computed to serve as the initial center for region extraction. From this centroid, a circular region expansion method grows the bounding box outward in small increments, checking for low-entropy areas along each edge as stopping point to ensure meaningful content within the boundary. After the expansion, the region is expanded to follow the aspect ratio of the original image, then extracted as the final output \\( X_{s} \\), capturing detailed areas while maintaining minimal distortion." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.659, + 0.906, + 0.69 + ], + "angle": 0, + "content": "Details of SRE process and algorithm are elaborated in Supplementary Materials." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.707, + 0.646, + 0.724 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.733, + 0.907, + 0.87 + ], + "angle": 0, + "content": "Implementation details: We use the MDTv2 XL model [5] with RGB reference images of shape \\(512 \\times 512 \\times 3\\), encoded into latent \\(Z\\)-space via the Stable Diffusion XL VAE [26]. These latent features are patch-embedded and processed within our model. For conditioning of garment features, the ViT encoder extracts features from RGB images of shape \\(224 \\times 224 \\times 3\\). Training was conducted on two 80GB A100 GPUs with a batch size of 6. Further details are provided in the Supplementary Materials." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Datasets and Evaluation Metrics: We train and evaluate our model on two public datasets: VITON-HD [1] and" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28289" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.921, + 0.293 + ], + "angle": 0, + "content": "
MethodInf. Time (s)↓Param. (M)↓LPIPS↓VITON-HDDressCodeUpper
SSIM↑FID(pair/unp.)↓LPIPS↓SSIM↑FID(pair/unp.)↓
LaDI-VTON [21]4.2258650.0980.8816.587/9.2650.0620.92710.447/14.048
DCI-VTON [7]15.4088590.0850.8965.559/8.7500.0680.91810.501/12.490
StableVITON [16]13.3571,0390.0860.8727.039/9.7170.0660.91911.572/13.666
CAT-DM [45]4.9478590.1040.8557.697/10.2630.0580.92210.257/14.039
TPD [41]28.7848590.0970.88510.161/13.0710.0480.9569.500/12.904
IDM-VTON [2]12.2882,9910.0820.8816.000/9.1630.0360.9447.471/11.915
MDT-IVTON (Base)4.5966710.0990.8877.773/9.8380.0610.92110.300/13.998
Base+ITAFA4.5966710.0860.8815.717/8.8790.0340.9515.414/13.111
Base+ITAFA+HR6.2156710.0830.8845.542/8.8000.0340.9515.384/12.528
Base+ITAFA+SRE4.5966710.0830.8855.462/8.6760.0340.9515.412/10.069
" + }, + { + "type": "table_caption", + "bbox": [ + 0.089, + 0.303, + 0.908, + 0.377 + ], + "angle": 0, + "content": "Table 1. Quantitative results of diffusion-based virtual try-on methods on VITON-HD and DressCode upper-body test sets. Inf. Time refers to the inference time, diffusion model only, measured for generating a single image on our environment, and Param. refers to the denoising diffusion model's parameter size. We report both paired and unpaired FID. Following the previous works, feature of last hidden layer of ViT image encoder is used for our MDT-IVTON without ITAFA. HR refers to use of single High-resolution \\((448\\times 448\\times 3)\\) garment image to formulate condition vector \\(c\\). The best performance for each metric is indicated in bold, while the second best is underlined." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.4, + 0.482, + 0.505 + ], + "angle": 0, + "content": "DressCode [20] upper, lower, and dresses. Data augmentation techniques from StableVITON [16] are applied to enhance model robustness. To assess performance, we use standard metrics from previous VTON works: LPIPS [46], SSIM [37], and FID [11]. All evaluations follow the protocol established in LaDI-VTON [21] to ensure fair comparison with previous virtual try-on models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.506, + 0.483, + 0.539 + ], + "angle": 0, + "content": "Further information about the datasets and evaluation metrics is provided in the Supplementary Materials." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.545, + 0.409, + 0.562 + ], + "angle": 0, + "content": "4.1. Qualitative and Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.568, + 0.483, + 0.734 + ], + "angle": 0, + "content": "The quantitative Results of Table. 1 compare our proposed ITA-MDT framework to previous diffusion-based methods, as well as show effect of ITAFA, and SRE on MDT-IVTON. The values of the previous works are re-measured by implementation with their checkpoints released. As it can be observed, our method provides comparable, state-of-the-art, results with faster inference speed and smaller parameters. The qualitative results and comparison to previous methods are shown in Fig. 6, and more in Supplementary Materials. As part of the inpainting process, the generated images are blended with agnostic images." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.743, + 0.245, + 0.759 + ], + "angle": 0, + "content": "4.2. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.765, + 0.483, + 0.9 + ], + "angle": 0, + "content": "The qualitative results of our proposed ITA-MDT framework and effects of each components are available in the Supplementary Materials. MDT-IVTON alone captures the global structure well, including garment shapes, colors, and human pose, but falls short in fidelity of garment details. The integration of ITAFA and SRE addresses this limitation significantly, with SRE shown as effective as a high-resolution image with twice the width compared to SRE integrated condition vector \\( c \\)." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.398, + 0.704, + 0.51 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.399, + 0.903, + 0.508 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.511, + 0.698, + 0.622 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.701, + 0.51, + 0.904, + 0.623 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.633, + 0.907, + 0.732 + ], + "angle": 0, + "content": "Figure 5. Visualization of the ITAFA's average weight distribution across ViT encoder layers of our final model. (a): use of hidden layers as denoising progresses. The dotted arrow refers to denoising direction. (b, c, d): changes in the use of hidden layers as the sparsity, variance, and gradient magnitude of garment image changes. The dotted arrow refers to the increase in these image complexity components." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Fig. 5 visualizes how ITAFA leverages features from different layers of the ViT image encoder, DINOv2, across the denoising process. In Fig. 5 (a), where it observes the use of layers with respect to timesteps, the middle layers dominate the feature aggregation for much of the diffusion; however, a distinct shift in layer usage occurs from early to later layers as the timestep progresses. In the early timestep when the input is largely noise, the encoder's first layer is prominently utilized, and at the final timestep, the use of the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28290" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.086, + 0.907, + 0.658 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.665, + 0.908, + 0.694 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparison of ITA-MDT with prior methods on VITON-HD and DressCode Upper-body datasets. The first three rows show VITON-HD results; the next three show DressCode results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.486, + 0.886 + ], + "angle": 0, + "content": "last layer markedly increases. This observation supports our hypothesis that dynamically aggregating features from different layers as the diffusion progresses effectively optimizes feature utilization across denoising stages. In Fig. 5 (b, c, d), we observe a clear pattern where specific layers are consistently favored depending on the intensity of particular image complexity (e.g., feature sparsity, variance, or gradient magnitude). This indicates that our model adaptively activates layers containing information most relevant to the type of visual complexity, highlighting the value and adaptability of our ITAFA." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.719, + 0.634, + 0.735 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "This paper presented ITA-MDT, a framework for the IVTON task that successfully leverages MDT, along with the proposed ITAFA and SRE modules, to address the challenges of transformer-based denoising diffusion models in preserving fine-grained details efficiently. Analysis confirmed that ITAFA and SRE effectively function as intended, enabling ITA-MDT to achieve state-of-the-art performance in several metrics with reduced model parameters and faster inference speed compared to previous methods." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28291" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.486, + 0.283 + ], + "angle": 0, + "content": "This work was supported by Institute for Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No.RS-2021-II211381, Development of Causal AI through Video Understanding and Reinforcement Learning, and Its Applications to Real Environments) and partly supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No.RS-2022-II220184, Development and Study of AI Technologies to Inexpensively Conform to Evolving Policy on Ethics)" + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.293, + 0.188, + 0.309 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.318, + 0.484, + 0.387 + ], + "angle": 0, + "content": "[1] Seunghwan Choi, Sunghyun Park, Minsoo Lee, and Jaegul Choo. Viton-hd: High-resolution virtual try-on via misalignment-aware normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14131-14140, 2021. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.388, + 0.484, + 0.429 + ], + "angle": 0, + "content": "[2] Yisol Choi, Sangkyung Kwak, Kyungmin Lee, Hyungwon Choi, and Jinwoo Shin. Improving diffusion models for virtual try-on. arXiv preprint arXiv:2403.05139, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.43, + 0.484, + 0.485 + ], + "angle": 0, + "content": "[3] Lavinia De Divitiis, Federico Becattini, Claudio Baecchi, and Alberto Del Bimbo. Disentangling features for fashion recommendation. ACM Transactions on Multimedia Computing, Communications and Applications, 19(1s):1-21, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.486, + 0.484, + 0.567 + ], + "angle": 0, + "content": "[4] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.568, + 0.484, + 0.636 + ], + "angle": 0, + "content": "[5] Shanghua Gao, Pan Zhou, Ming-Ming Cheng, and Shuicheng Yan. Masked diffusion transformer is a strong image synthesizer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23164-23173, 2023. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.638, + 0.484, + 0.706 + ], + "angle": 0, + "content": "[6] Yuying Ge, Yibing Song, Ruimao Zhang, Chongjian Ge, Wei Liu, and Ping Luo. Parser-free virtual try-on via distilling appearance flows. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8485-8493, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.707, + 0.484, + 0.775 + ], + "angle": 0, + "content": "[7] Junhong Gou, Siyu Sun, Jianfu Zhang, Jianlou Si, Chen Qian, and Liqing Zhang. Taming the power of diffusion models for high-quality virtual try-on with appearance flow. In Proceedings of the 31st ACM International Conference on Multimedia, pages 7599-7607, 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.777, + 0.484, + 0.831 + ], + "angle": 0, + "content": "[8] Ríza Alp Güler, Natalia Neverova, and Iasonas Kokkinos. Denseseposé: Dense human pose estimation in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7297-7306, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.832, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[9] M Hadi Kiapour, Xufeng Han, Svetlana Lazebnik, Alexander C Berg, and Tamara L Berg. Where to buy it: Matching street clothing photos in online shops. In Proceedings of the IEEE international conference on computer vision, pages 3343-3351, 2015. 1" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.318, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.148 + ], + "angle": 0, + "content": "[10] Xintong Han, Zuxuan Wu, Zhe Wu, Ruichi Yu, and Larry S Davis. Viton: An image-based virtual try-on network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7543-7552, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.907, + 0.217 + ], + "angle": 0, + "content": "[11] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[12] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.907, + 0.291 + ], + "angle": 0, + "content": "[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.293, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[14] Wei-Lin Hsiao and Kristen Grauman. Creating capsule wardrobes from fashion images. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7161-7170, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.35, + 0.907, + 0.419 + ], + "angle": 0, + "content": "[15] Thibaut Issenhuth, Jérémie Mary, and Clément Calauzenes. Do not mask what you do not need to mask: a parser-free virtual try-on. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XX 16, pages 619–635. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.421, + 0.905, + 0.49 + ], + "angle": 0, + "content": "[16] Jeongho Kim, Guojung Gu, Minho Park, Sunghyun Park, and Jaegul Choo. Stableviton: Learning semantic correspondence with latent diffusion model for virtual try-on. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8176-8185, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.492, + 0.907, + 0.547 + ], + "angle": 0, + "content": "[17] Gwanhyeong Koo, Sunjae Yoon, Ji Woo Hong, and Chang D Yoo. Flexied: Frequency-aware latent refinement for enhanced non-rigid editing. In European Conference on Computer Vision, pages 363-379. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.549, + 0.907, + 0.616 + ], + "angle": 0, + "content": "[18] Sangyun Lee, Gyojung Gu, Sunghyun Park, Seunghwan Choi, and Jaegul Choo. High-resolution virtual try-on with misalignment and occlusion-handled conditions. In European Conference on Computer Vision, pages 204-219. Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.62, + 0.907, + 0.687 + ], + "angle": 0, + "content": "[19] Shugao Ma, Tomas Simon, Jason Saragih, Dawei Wang, Yuecheng Li, Fernando De La Torre, and Yaser Sheikh. Pixel codec avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 64-73, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.69, + 0.907, + 0.759 + ], + "angle": 0, + "content": "[20] Davide Morelli, Matteo Fincato, Marcella Cornia, Federico Landi, Fabio Cesari, and Rita Cucchiara. Dress code: high-resolution multi-category virtual try-on. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2231-2235, 2022. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.761, + 0.907, + 0.83 + ], + "angle": 0, + "content": "[21] Davide Morelli, Alberto Baldrati, Giuseppe Cartella, Marcelia Cornia, Marco Bertini, and Rita Cucchiara. Ladi-vton: latent diffusion textual-inversion enhanced virtual try-on. In Proceedings of the 31st ACM International Conference on Multimedia, pages 8580-8589, 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.907, + 0.9 + ], + "angle": 0, + "content": "[22] Shuliang Ning, Duomin Wang, Yipeng Qin, Zirong Jin, Baoyuan Wang, and Xiaoguang Han. Picture: Photorealistic virtual try-on from unconstrained designs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6976-6985, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "28292" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.161 + ], + "angle": 0, + "content": "[23] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.485, + 0.219 + ], + "angle": 0, + "content": "[24] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.484, + 0.262 + ], + "angle": 0, + "content": "[25] Trung X Pham, Zhang Kang, and Chang D Yoo. Cross-view masked diffusion transformers for person image synthesis. arXiv preprint arXiv:2402.01516, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.484, + 0.331 + ], + "angle": 0, + "content": "[26] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 6, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.484, + 0.403 + ], + "angle": 0, + "content": "[27] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM international conference on multimedia, pages 484-492, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.484, + 0.473 + ], + "angle": 0, + "content": "[28] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International conference on machine learning, pages 8821-8831. Pmlr, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.484, + 0.546 + ], + "angle": 0, + "content": "[29] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.484, + 0.63 + ], + "angle": 0, + "content": "[30] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.632, + 0.484, + 0.714 + ], + "angle": 0, + "content": "[31] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.716, + 0.484, + 0.799 + ], + "angle": 0, + "content": "[32] Rohan Sarkar, Navaneeth Bodla, Mariya I Vasileva, Yen-Liang Lin, Anurag Beniwal, Alan Lu, and Gerard Medioni. Outfittransformer: Learning outfit representations for fashion recommendation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3601-3609, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.801, + 0.484, + 0.841 + ], + "angle": 0, + "content": "[33] Claude Elwood Shannon. A mathematical theory of communication. The Bell system technical journal, 27(3):379-423, 1948. 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.844, + 0.484, + 0.884 + ], + "angle": 0, + "content": "[34] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.887, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[35] Yang Song and Stefano Ermon. Generative modeling by" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.121 + ], + "angle": 0, + "content": "estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.908, + 0.19 + ], + "angle": 0, + "content": "[36] Bochao Wang, Huabin Zheng, Xiaodan Liang, Yimin Chen, Liang Lin, and Meng Yang. Toward characteristic-preserving image-based virtual try-on network. In Proceedings of the European conference on computer vision (ECCV), pages 589-604, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.907, + 0.247 + ], + "angle": 0, + "content": "[37] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.907, + 0.332 + ], + "angle": 0, + "content": "[38] Zhenyu Xie, Zaiyu Huang, Xin Dong, Fuwei Zhao, Haoye Dong, Xijin Zhang, Feida Zhu, and Xiaodan Liang. Gp-vton: Towards general purpose virtual try-on via collaborative localflow global-parsing learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23550-23559, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.907, + 0.403 + ], + "angle": 0, + "content": "[39] Sicheng Xu, Guojun Chen, Yu-Xiao Guo, Jiaolong Yang, Chong Li, Zhenyu Zang, Yizhong Zhang, Xin Tong, and Baining Guo. Vasa-1: Lifelike audio-driven talking faces generated in real time. arXiv preprint arXiv:2404.10667, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.405, + 0.907, + 0.475 + ], + "angle": 0, + "content": "[40] Han Yang, Ruimao Zhang, Xiaobao Guo, Wei Liu, Wangmeng Zuo, and Ping Luo. Towards photo-realistic virtual try-on by adaptively generating-preserving image content. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7850-7859, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.908, + 0.546 + ], + "angle": 0, + "content": "[41] Xu Yang, Changxing Ding, Zhibin Hong, Junhao Huang, Jin Tao, and Xiangmin Xu. Texture-preserving diffusion models for high-fidelity virtual try-on. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7017-7026, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.906, + 0.603 + ], + "angle": 0, + "content": "[42] Sunjae Yoon, Gwanhyeong Koo, Ji Woo Hong, and Chang D Yoo. Dni: Dilutional noise initialization for diffusion video editing. In European Conference on Computer Vision, pages 180-195. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.906, + 0.646 + ], + "angle": 0, + "content": "[43] Sunjae Yoon, Gwanhyeong Koo, Geonwoo Kim, and Chang D Yoo. Frag: Frequency adapting group for diffusion video editing. arXiv preprint arXiv:2406.06044, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.647, + 0.906, + 0.702 + ], + "angle": 0, + "content": "[44] Sunjae Yoon, Gwanhyeong Koo, Younghwan Lee, and Chang Yoo. Tpc: Test-time procrustes calibration for diffusion-based human image animation. Advances in Neural Information Processing Systems, 37:118654-118677, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.704, + 0.908, + 0.773 + ], + "angle": 0, + "content": "[45] Jianhao Zeng, Dan Song, Weizhi Nie, Hongshuo Tian, Tongtong Wang, and An-An Liu. Cat-dm: Controllable accelerated virtual try-on with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8372-8382, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.775, + 0.907, + 0.843 + ], + "angle": 0, + "content": "[46] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.907, + 0.902 + ], + "angle": 0, + "content": "[47] Luyang Zhu, Dawei Yang, Tyler Zhu, Fitsum Reda, William Chan, Chitwan Sahara, Mohammad Norouzi, and Ira Kemelmacher-Shlizerman. Tryondiffusion: A tale of two unets. In Proceedings of the IEEE/CVF Conference on Com" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28293" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.124, + 0.092, + 0.485, + 0.119 + ], + "angle": 0, + "content": "puter Vision and Pattern Recognition, pages 4606-4615, 2023. 3" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28294" + } + ] +] \ No newline at end of file diff --git a/2025/ITA-MDT_ Image-Timestep-Adaptive Masked Diffusion Transformer Framework for Image-Based Virtual Try-On/e6b13099-74a4-4157-8bc0-2e7511ad889f_origin.pdf b/2025/ITA-MDT_ Image-Timestep-Adaptive Masked Diffusion Transformer Framework for Image-Based Virtual Try-On/e6b13099-74a4-4157-8bc0-2e7511ad889f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..dac6d0fb78982e57635afce67e440ce89571caa5 --- /dev/null +++ b/2025/ITA-MDT_ Image-Timestep-Adaptive Masked Diffusion Transformer Framework for Image-Based Virtual Try-On/e6b13099-74a4-4157-8bc0-2e7511ad889f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:001e4eda3922fdad2970c5351acf2e89bc8ce3375dc1fcaa52d7d780ad8e44ec +size 5121919 diff --git a/2025/ITA-MDT_ Image-Timestep-Adaptive Masked Diffusion Transformer Framework for Image-Based Virtual Try-On/full.md b/2025/ITA-MDT_ Image-Timestep-Adaptive Masked Diffusion Transformer Framework for Image-Based Virtual Try-On/full.md new file mode 100644 index 0000000000000000000000000000000000000000..305ebc0c91cf42d57479d6e7118d82ad955f6aeb --- /dev/null +++ b/2025/ITA-MDT_ Image-Timestep-Adaptive Masked Diffusion Transformer Framework for Image-Based Virtual Try-On/full.md @@ -0,0 +1,310 @@ +# ITA-MDT: Image-Timestep-Adaptive Masked Diffusion Transformer Framework for Image-Based Virtual Try-On + +Ji Woo Hong $^{1}$ Tri Ton $^{1}$ Trung X. Pham $^{1}$ Gwanhyeong Koo $^{1}$ Sunjae Yoon $^{1}$ Chang D. Yoo $^{1}$ $^{1}$ Korea Advanced Institute of Science and Technology (KAIST), South Korea + +{jiwohong93, tritth, trungpx, kookie, dbstjswo505, cd_yoo}@kaist.ac.kr + +![](images/ffcbc1c8930d7110b713b648e9c1b8d78e62c74edc0193f164ce4ac02682992b.jpg) +Figure 1. Image-based Virtual Try-on results generated by our ITA-MDT. From left to right: VITON-HD[1], DressCode[20] upper, DressCode lower, and DressCode dresses, respectively. + +# Abstract + +This paper introduces ITA-MDT, the Image-Timestep-Adaptive Masked Diffusion Transformer Framework for Image-Based Virtual Try-On (IVTON), designed to overcome the limitations of previous approaches by leveraging the Masked Diffusion Transformer (MDT) for improved handling of both global garment context and fine-grained details. The IVTON task involves seamlessly superimposing a garment from one image onto a person in another, creating a realistic depiction of the person wearing the specified garment. Unlike conventional diffusion-based virtual try-on models that depend on large pre-trained U-Net architectures, ITA-MDT leverages a lightweight, scalable transformer-based denoising diffusion model with a mask latent modeling scheme, achieving competitive results while reducing computational overhead. A key component of ITAMDT is the Image-Timestep Adaptive Feature Aggregator (ITAFA), a dynamic feature aggregator that combines all of the features from the image encoder into a unified feature of the same size, guided by diffusion timestep and garment image complexity. This enables adaptive weighting of features, allowing the model to emphasize either global information or fine-grained details based on the requirements of the denoising stage. Additionally, the Salient Region Extractor (SRE) module is presented to identify complex region of the garment to provide high-resolution local infor + +mation to the denoising model as an additional condition alongside the global information of the full garment image. This targeted conditioning strategy enhances detail preservation of fine details in highly salient garment regions, optimizing computational resources by avoiding unnecessarily processing entire garment image. Comparative evaluations confirms that ITA-MDT improves efficiency while maintaining strong performance, reaching state-of-the-art results in several metrics. Our project page is available at https://jiwoohong93.github.io/ita-mdt/. + +# 1. Introduction + +The task of Image-based Virtual Try-on (IVTON) is to generate an image where a person appears dressed in specific clothes, using provided images of both the garment and the individual. In real-world applications, IVTON holds great potential for revolutionizing online shopping, fashion design [3, 9, 14, 32], and even entertainment industries for creating digital avatars or virtual costumes [19, 27, 39]. For this task, deep learning methods [1, 6, 10, 15, 18, 36, 38, 40] employ paired datasets [1, 20] containing both standalone garment images and the images of the individual wearing these for training their models, along with additional reference images such as mask images of the garment region and DensePose [8] image. + +Recently, diffusion models [13, 34] have shown great + +promise in image generation and synthesis in many different tasks [17, 25, 42-44]. Especially, large-scale pre-trained diffusion models [28, 29, 31], such as Stable Diffusion [29], are widely used to preserving high fidelity in generated images; however, their large size and slow inference speeds present practical limitations in commercial application. + +To address this, transformer-based diffusion models have emerged as an efficient alternative. The Diffusion Transformer (DiT) [24] based on the Visual Transformer (ViT) [4] architecture with Latent Diffusion Models [29] can achieve excellent scalability and efficiency. Extending this concept, the Masked Diffusion Transformer (MDT) [5] applied mask prediction scheme to enhance spatial representation learning, achieving state-of-the-art results in class-conditional image generation. While promising, directly applying MDT to IVTON task presents challenges, as the task requires high fidelity and consistency across multiple reference images, demanding more than a straightforward approach. + +This paper presents the ITA-MDT (Image-Timestep-Adaptive Masked Diffusion Transformer) Framework, which incorporates our modified version of MDT adapted specifically for IVTON, named MDT-IVTON, as well as two novel components: the Image-Timestep Adaptive Feature Aggregator (ITAFA) and the Salient Region Extractor (SRE). This framework is designed to preserve garment fidelity while maintaining computational efficiency. Our MDT-IVTON model leverages MDT's mask prediction scheme to efficiently learn spatial relationships between garment and individual. Additionally, we incorporate an auxiliary loss to enhance the model's focus on the inpainting objective of the task. While this architecture effectively captures global dependencies, it struggles to retain fine-grained details, which U-Net-based models handle relatively naturally through their multi-scale encoder-decoder structure that captures and reconstructs information of the image across multiple scales through progressive downsampling and upsampling. Our ITAFA and SRE bridge this gap. + +The Image-Timestep Adaptive Feature Aggregator (ITAFA) is a learnable feature aggregator designed to condition the diffusion model effectively and efficiently. It dynamically combines multiple image features from ViT image encoder, guided by the current diffusion timestep and garment image complexity score. While simply providing multiple image features from different layers can condition the model with extensive information, this naive approach significantly increases the computational load. By assigning adaptive weights to each feature, ITAFA aggregates them into a single representation. These weights are modulated based on the timestep information, which indicates how far the generation process has progressed. The weighting is further refined using the visual complexity of the garment, which reflects how much fine-grained information the condition image contains. This design enables the model to gradually shift its focus + +![](images/38953b5a0f8d37d04c01a469fda50e5a301ce13bb0c7ea02d0e5949abba92f5f.jpg) +Figure 2. Comparison of our ITA-MDT with other diffusion-based methods on the VTON task using the VITON-HD dataset. For all metrics in the analysis, including FID, model parameters, and inference speed, lower values indicate better performance. + +from global semantic features in the early stages to fine local details in the later stages of the diffusion process. + +The Salient Region Extractor (SRE) provides high-resolution local information of key garment area, such as logos or complex patterns, identified through entropy-based detection. The region extracted offers additional crucial condition guidance to the diffusion model alongside the global context from the full garment image. While simply using a high-resolution, big size, image enhances detail representation, it results in substantial computation. Notably, most regions of garments are either solid-colored or display large, simple patterns that can be captured adequately through lower resolution. Therefore, rather than uniformly processing the entire image at high resolution, SRE selectively extracts and provides additional high-resolution information only for regions where fine-grained local information is crucial. In summary, our contributions include: + +1. We present ITA-MDT, a framework for the image-based virtual try-on (IVTON) task that builds upon MDT-IVTON, our adaptation of the Masked Diffusion Transformer (MDT) for IVTON, and incorporates our proposed methods to effectively address the challenges of transformer-based denoising diffusion models in preserving fine-grained details. +2. We propose ITAFA, a learnable feature aggregator designed to dynamically combine features from the image encoder into a single feature based on the current denoising timestep and garment image complexity score. +3. We propose SRE to provide high-resolution local information of key garment area, identified through entropy-based detection, as additional condition guidance to the diffusion model alongside the global context from the full garment image. +4. ITA-MDT's combination of MDT-IVTON, ITAFA, and SRE offers a highly efficient solution for IVTON, achieving a balance between model efficiency and output fi + +delity. Figure 2 presents the comparison of our ITA-MDT with previous diffusion-based methods. It achieves state-of-the-art results in several metrics while maintaining competitive performance across the others, demonstrating its overall effectiveness and suitability for practical virtual try-on applications. + +# 2. Related Work + +# 2.1. Image-based Virtual Try-on with Diffusion + +The integration of diffusion models has significantly advanced the field of Image-based Virtual Try-on, as evidenced in numerous recent studies. Notably, TryOnDiffusion [47] introduced an architecture utilizing dual U-Nets, highlighting the potential of diffusion-based try-on but requiring large-scale datasets composed of image pairs of the same person in varied poses. This large-scale dataset requirement has led researchers to pivot towards leveraging large pre-trained diffusion models [28, 29, 31]. To condition these models effectively, methods such as representing garments as pseudowords in LaDI-VTON [21], applying warping networks in DCI-VTON [7], modifying attention mechanisms in StableVITON [16] and IDM-VTON [2], utilizing ControlNet for garment control and GAN-based sampling in CAT-DM [45], and concatenating masked person and garment images for direct texture transfer in TPD [41] have been introduced. However, the use of large pre-trained U-Nets comes with notable limitations. The large number of parameters in these models significantly increases memory usage and slows down inference speed, making them resource-intensive and less practical for efficient deployment. A variety of solutions could be explored to address these issues, among which transformer-based diffusion models present a promising approach. + +# 2.2. Transformers for Diffusion Process + +Traditionally, The U-Net architecture [30] has been foundational for diffusion models and continues to be a prevalent choice across various diffusion-based generation tasks [13, 35]. However, the incorporation of transformers into diffusion processes also marked a significant evolution in generative modeling. The introduction of the Diffusion Transformer (DiT) [24] merged the architecture of the Vision Transformer (ViT) [4] with latent diffusion models. DiT has demonstrated outstanding scalability and compute efficiency, delivering superior generative performance compared to traditional U-Net architectures. Building on this, the Masked Diffusion Transformer (MDT) [5] has further advanced the diffusion transformer model by employing an asymmetric masking schedule to optimize the contextual representation learning, setting a new benchmark in class image generation on ImageNet. In an attempt to leverage MDT for image-to-image generation, [25] introduced an aggregation network + +that consolidates image conditions into a single vector for guiding the diffusion process in pose-guided human image generation tasks. While effective within its domain, directly applying this approach to the IVTON task is unsuitable. IVTON involves a significantly larger set of condition images, and compressing these into a single vector leads to substantial information loss, ultimately degrading output quality. Given the need to preserve fine details from condition images, enhancing rather than compressing condition information is preferred. + +# 2.3. Enhancing Condition Information in Virtual Try-On task + +To enhance the intricate details essential to IVTON, the recent method, PICTURE [22], introduces a two-stage disentanglement pipeline that separates style and texture features by employing a hierarchical CLIP feature extraction module with position encoding, enabling the representation of complex, non-stationary textures. While effective in handling intricate style and texture combinations, this method requires an in-depth analysis of features from the hidden layers of the image encoder. Specifically, this involves clustering hidden layer features and manually selecting representative features from each cluster, a process that lacks flexibility for it requires re-evaluation when a different encoder is to be used. Additionally, the selected features from each cluster are concatenated for cross-attention conditioning, substantially increasing computational demand. Together, these requirements render the approach resource-intensive and less adaptable to diverse architectures. + +# 3. Method + +The ITA-MDT framework performs Image-Based Virtual Try-on through a transformer-based denoising diffusion network, leveraging the mask latent modeling scheme of MDTv2 [5]. An overview of the architecture is shown in Figure 3, which comprises three primary components: (1) the Masked Diffusion Transformer for Image-Based Virtual Try-On (MDT-IVTON), (2) the Image-Timestep Adaptive Feature Aggregator (ITAFA), and (3) the Salient Region Extractor (SRE) module. Detailed descriptions of each component are provided below. + +# 3.1. Masked Diffusion Transformer for Image-Based Virtual Try-On (MDT-IVTON) + +Our ITA-MDT framework integrates a Masked Diffusion Transformer (MDT) tailored for Image-Based Virtual Try-On (IVTON), leveraging the transformer-based denoising diffusion model architecture of DiT [24]. This model operates in a latent space encoded via a pre-trained VAE following the Latent Diffusion Model (LDM) approach [29]. Specifically, the VAE encoder maps garment agnostic inputs into this latent space, including the Garment Agnostic + +![](images/e23e4b455c2609c241fb95d6d7598db11cf16d0a32ee3dbbd0d3056454d1452d.jpg) +Figure 3. Overview of the ITA-MDT Framework for Image-Based Virtual Try-On (IVTON). The framework takes multiple reference images encoded into latent space as query, which includes Garment Agnostic Map $A$ , DensePose $P$ , and Garment Agnostic Mask $M_X$ . These reference latent images are concatenated to be patch embedded for masked diffusion process within our MDT-IVTON, which follows the architecture of MDTv2 [5] with integrated cross-attention blocks. The image feature of Garment $X$ is extracted with the ViT image encoder, DINOv2 [23], and are adaptively aggregated with our proposed Image-Timestep Adaptive Feature Aggregator (ITAFA) to produce Garment Feature $F_g$ . With our Salient Region Extractor (SRE), the Salient Region $X_s$ is extracted from the Garment $X$ and processed through ITAFA separately to produce Salient Feature $F_s$ . The Garment Feature $F_g$ and Salient Feature $F_s$ are concatenated to serve as conditions of MDT-IVTON. Positional embeddings are omitted in this figure for simplicity. + +Map $A$ , DensePose $P$ , and Garment Agnostic Mask $M_X$ , concatenated with the noised latent representation $z_t$ . + +The MDT-IVTON model processes the input by embedding it into a patchified latent representation $L$ and sequentially passing it through $N_{1}$ Encoder Blocks followed by $N_{2}$ Decoder Blocks. Each block utilizes Layer Normalization, Multi-Head Self-Attention, Multi-Head Cross Attention, Multi-Layer Perceptron (MLP), and skip connections to refine the denoised latent representation $z_{t - 1}$ , as illustrated in the right side of Figure 3. Conditioning information derived from the concatenated garment features $F_{g}$ and $F_{s}$ , extracted from garment image $X$ and its salient region $X_{s}$ using the ViT encoder and ITAFA module, is infused through cross-attention mechanisms to guide the generation process. Detailed descriptions of ITAFA and SRE are available in Sec. 3.2 and Sec. 3.3. + +Three objective functions guide the MDT-IVTON training to optimize its generative performance: + +# 3.1.1. Denoising Objective + +The primary objective function minimizes the mean squared error (MSE) between the predicted noise and the actual noise in the denoised latent $z_{t}$ at each timestep $t$ , following the standard diffusion objective: + +$$ +L _ {\text {d e n o i s e}} = \mathcal {E} _ {y, c, \epsilon \sim \mathcal {N} (0, I), t} \left[ \| \epsilon - \epsilon_ {\theta} \left(y _ {t}, c, t\right) \| ^ {2} \right], \tag {1} +$$ + +where $y_{t}$ denotes the latent representation at timestep $t$ , $c$ the conditioning, and $\epsilon_{\theta}$ the model's predicted noise. + +# 3.1.2. Mask Reconstruction Objective + +In MDT-IVTON, the Mask Reconstruction Objective enforces spatial coherence and semantic consistency by reconstructing masked tokens within the masked latent representation $E_{m}$ . This objective ensures alignment between the reference and generated images, enhancing the model's ability to synthesize realistic outputs in masked regions. + +To achieve this, we incorporate the Side-Interpolator module, which plays a pivotal role in interpolating information across unmasked tokens to predict values for masked tokens in $E_{m}$ . The Side-Interpolator effectively fills in missing information by conditioning on neighboring unmasked regions, encouraging consistent token-level synthesis within the masked latent. + +The Mask Reconstruction Objective follows an MSE formulation similar to the denoising objective: + +$$ +L _ {\text {m a s k}} = \mathcal {E} _ {y, c, \epsilon \sim \mathcal {N} (0, I), t} \left[ \| \epsilon - \epsilon_ {\theta} (\text {S i d e - I n t} (E _ {m}), c, t) \| ^ {2} \right], \tag {2} +$$ + +where Side-Int processes $E_{m}$ , leveraging spatial context to predict masked regions. The conditioning $c$ includes garment details, aiding in preserving fidelity to the garment appearance in masked areas. This objective promotes detailed reconstruction by interpolating across the context in $E_{m}$ , enhancing semantic alignment within generated images. + +# 3.1.3. Inpainting Objective + +The inpainting objective in MDT-IVTON is designed to enhance garment fidelity by concentrating on regions defined by the garment agnostic mask $M_X$ , encoded as $\mathcal{E}(M_X)$ in latent space. During training, this mask serves as a selective weighting factor in the mean squared error (MSE) between the target and predicted outputs, guiding the model's focus to reconstruct details more accurately within garment-specific regions. + +To ensure numerical stability and maintain well-behaved gradients, the mask is normalized to the range [0, 1], preventing extreme values that could otherwise destabilize training. This targeted inpainting strategy enhances garment detail fidelity by directing the model's attention to areas where high-quality reconstruction is most critical. + +Latent Mask Normalization: To prevent unstable gradients, the single channel of the latent mask $\mathcal{E}(M_X)$ that most closely resembles the binary mask of the inpainting task is chosen to be used as the mask across all channels in the loss calculation. This selected channel is normalized to values between 0 and 1 to ensure stability: + +$$ +\mathcal {E} \left(M _ {X}\right) = \frac {\mathcal {E} \left(M _ {X}\right) - \operatorname* {m i n} \left(\mathcal {E} \left(M _ {X}\right)\right)}{\operatorname* {m a x} \left(\mathcal {E} \left(M _ {X}\right)\right) - \operatorname* {m i n} \left(\mathcal {E} \left(M _ {X}\right)\right) + 1 \times 1 0 ^ {- 8}}, \tag {3} +$$ + +where $1 \times 10^{-8}$ ensures numerical stability. This approach avoids unintended gradients and aligns the loss with garment-focused regions. A visualization of the latent mask selected is in the Supplementary Materials. + +Masked Loss Computation: With the rescaled single channel of $\mathcal{E}(M_X)$ , the masked inpainting loss is calculated as a weighted mean squared error (MSE) applied uniformly across all channels in the latent space: + +$$ +L _ {\text {i n p a i n t}} = \mathcal {E} _ {y, c, t} \left[ \| \mathcal {E} (M _ {X}) \cdot \left(y _ {t} - y _ {t - 1}\right) \| ^ {2} \right], \tag {4} +$$ + +where $y_{t}$ and $y_{t-1}$ denote the latent representations at timesteps $t$ and $t-1$ , respectively. This loss term emphasizes reconstruction quality in the masked regions, promoting accurate garment-specific detail representation. Further details of MDT-IVTON are provided in Supplementary Materials. + +# 3.2. Image-Timestep Adaptive Feature Aggregator (ITAFA) + +The ITAFA is a dynamic feature aggregation module designed to adaptively weight feature embeddings from a Vision Transformer (ViT) image encoder, utilizing learnable parameters and feature complexity analysis to dynamically modulate the contribution of each hidden layer output of the image encoder at different timesteps of the denoising diffusion process. + +# 3.2.1. Components and Parameters + +Timestep Embedding Projector. The diffusion timestep embedding, denoted as $T$ , represents the current timestep in + +the denoising process with dimension $\mathcal{H}$ of the MDT-IVTON model. To align $T$ with the image feature dimension, it is projected to the dimensionality of the hidden layer of the ViT, $\mathcal{H}_{\varepsilon}$ . This is achieved through a linear transformation: + +$$ +W ^ {T} = \operatorname {L i n e a r} (\mathcal {H}, \mathcal {H} _ {\mathcal {E}}), \tag {5} +$$ + +The linear layer maps the timestep embedding into a representation suitable for interacting with the ViT features. + +Image Complexity Projector. ITAFA incorporates three complexity components: feature sparsity, feature variance, and gradient magnitude. These collectively capture the complexity of garment texture and structure. + +1. Feature Sparsity measures the proportion of near-zero activations in the feature embedding tensor $f$ , providing insight into the sparsity of learned representations. It is computed as: + +$$ +\mathrm {S} = \frac {1}{n} \sum_ {i = 1} ^ {n} \mathbb {I} \left(\left| f _ {i} \right| < \delta\right), \tag {6} +$$ + +where $\delta$ is a small threshold (i.e., 0.01), $n$ denotes the total number of elements in $f$ , and $\mathbb{I}(\cdot)$ is the indicator function. + +2. Feature Variance captures the variability of activations, reflecting structural complexity within the image: + +$$ +V = \operatorname {V a r} (f), \tag {7} +$$ + +where $\operatorname{Var}(f)$ denotes the variance across all elements in the feature embedding. + +3. Gradient Magnitude measures local changes in the feature embedding, calculated by taking the spatial gradients along patch and embedding dimensions. The magnitude of gradients are computed, and averaging them provides a single score $G$ that captures texture complexity. + +$$ +G = \frac {1}{n} \sum_ {i = 1} ^ {n} \sqrt {\left(\frac {\partial f}{\partial x}\right) ^ {2} + \left(\frac {\partial f}{\partial y}\right) ^ {2}}. \tag {8} +$$ + +These components are concatenated to form the complexity score vector $[S,V,G]$ , which is then linearly projected to match the hidden layer dimension $\mathcal{H}_{\mathcal{E}}$ : + +$$ +W ^ {I} = \operatorname {L i n e a r} (3, \mathcal {H} _ {\mathcal {E}}). \tag {9} +$$ + +Learnable Weighting Parameter. The contribution of the timestep and complexity projections is balanced by a learnable parameter $\alpha \in [0,1]$ , initialized to 0.5 and optimized during training: + +$$ +W ^ {c o m} = \alpha \cdot W ^ {T} + (1 - \alpha) \cdot W ^ {I}. \tag {10} +$$ + +This weighted combination provides an adaptive control over the influence of timestep and image complexity on the image feature aggregation. + +![](images/549b3d146ae9897987a9bc98cd0747da5e80a6044a7b7e51870a33efd1cdd583.jpg) +Image-Timestep Adaptive Feature Aggregator + +![](images/7c2c18b4ff01c3d191f55a76bf5e65eb51c240fbc2873b81856d965658da3040.jpg) +Salient Region Extractor +Figure 4. Overview of the Image-Timestep Adaptive Feature Aggregator (ITAFA) and the Salient Region Extractor (SRE). Left: The ITAFA dynamically aggregates Vision Transformer (ViT) feature embeddings $f$ based on a combination of the Timestep Embedding Projector, which projects the diffusion timestep embedding $T$ to match the feature embedding dimensions, and the Image Complexity Projector, which transforms the image complexity vector $[S, V, G]$ (sparsity, variance, gradient magnitude) into a comparable dimension. The weight vectors are combined and normalized via softmax to form $W$ , which is used to adaptively aggregate the feature embeddings $\{f_i\}_{i=0}^h$ across hidden layers to produce the final output tensor $F$ . Garment $X$ and Salient Region $X_s$ are processed separately through ITAFA to generate the Garment Feature $F_g$ and Salient Feature $F_s$ . Right: The SRE processes the input image $X$ by computing an entropy map $X_e$ , creating a binary High-Entropy Mask $X_m$ , and circular region expansion from the entropy centroid to extract the final high-entropy region $X_s$ , ensuring preservation of detail within a consistent aspect ratio. + +# 3.2.2. Feature Aggregation + +The aggregated weights are normalized via softmax to ensure they sum to one across hidden layers, thereby controlling the contribution of each layers. + +$$ +W = \operatorname {s o f t m a x} \left(W ^ {\text {c o m}}\right). \tag {11} +$$ + +These weights are then applied to each hidden layer output from the ViT encoder, denoted by $f$ , via an element-wise multiplication and summation across layers: + +$$ +F = \sum_ {l = 1} ^ {\mathcal {H}} W [ l ] \cdot f [ l ], \tag {12} +$$ + +where $F$ represents the final aggregated feature tensor, capturing a balanced representation based on both denoising timestep and image complexities. + +In-depth details of ITAFA are explained in Supplementary Materials. + +# 3.3. Salient Region Extractor (SRE) + +Our SRE is a simple yet effective algorithm designed to identify and extract salient, high-entropy, areas within an image. The method includes: Entropy Map Computation, High-Entropy Region Identification, Entropy Centroid Localization, Circular Region Expansion, and Region Extraction. + +By utilizing Shannon Entropy [33] to compute an entropy map $X_{e}$ , which measures information content across pixel + +neighborhoods. High-entropy regions are then identified by thresholding this map to create a binary High-Entropy Mask $X_{m}$ that highlights areas with substantial visual complexity. The centroid of this region is computed to serve as the initial center for region extraction. From this centroid, a circular region expansion method grows the bounding box outward in small increments, checking for low-entropy areas along each edge as stopping point to ensure meaningful content within the boundary. After the expansion, the region is expanded to follow the aspect ratio of the original image, then extracted as the final output $X_{s}$ , capturing detailed areas while maintaining minimal distortion. + +Details of SRE process and algorithm are elaborated in Supplementary Materials. + +# 4. Experiments + +Implementation details: We use the MDTv2 XL model [5] with RGB reference images of shape $512 \times 512 \times 3$ , encoded into latent $Z$ -space via the Stable Diffusion XL VAE [26]. These latent features are patch-embedded and processed within our model. For conditioning of garment features, the ViT encoder extracts features from RGB images of shape $224 \times 224 \times 3$ . Training was conducted on two 80GB A100 GPUs with a batch size of 6. Further details are provided in the Supplementary Materials. + +Datasets and Evaluation Metrics: We train and evaluate our model on two public datasets: VITON-HD [1] and + +
MethodInf. Time (s)↓Param. (M)↓LPIPS↓VITON-HDDressCodeUpper
SSIM↑FID(pair/unp.)↓LPIPS↓SSIM↑FID(pair/unp.)↓
LaDI-VTON [21]4.2258650.0980.8816.587/9.2650.0620.92710.447/14.048
DCI-VTON [7]15.4088590.0850.8965.559/8.7500.0680.91810.501/12.490
StableVITON [16]13.3571,0390.0860.8727.039/9.7170.0660.91911.572/13.666
CAT-DM [45]4.9478590.1040.8557.697/10.2630.0580.92210.257/14.039
TPD [41]28.7848590.0970.88510.161/13.0710.0480.9569.500/12.904
IDM-VTON [2]12.2882,9910.0820.8816.000/9.1630.0360.9447.471/11.915
MDT-IVTON (Base)4.5966710.0990.8877.773/9.8380.0610.92110.300/13.998
Base+ITAFA4.5966710.0860.8815.717/8.8790.0340.9515.414/13.111
Base+ITAFA+HR6.2156710.0830.8845.542/8.8000.0340.9515.384/12.528
Base+ITAFA+SRE4.5966710.0830.8855.462/8.6760.0340.9515.412/10.069
+ +Table 1. Quantitative results of diffusion-based virtual try-on methods on VITON-HD and DressCode upper-body test sets. Inf. Time refers to the inference time, diffusion model only, measured for generating a single image on our environment, and Param. refers to the denoising diffusion model's parameter size. We report both paired and unpaired FID. Following the previous works, feature of last hidden layer of ViT image encoder is used for our MDT-IVTON without ITAFA. HR refers to use of single High-resolution $(448\times 448\times 3)$ garment image to formulate condition vector $c$ . The best performance for each metric is indicated in bold, while the second best is underlined. + +DressCode [20] upper, lower, and dresses. Data augmentation techniques from StableVITON [16] are applied to enhance model robustness. To assess performance, we use standard metrics from previous VTON works: LPIPS [46], SSIM [37], and FID [11]. All evaluations follow the protocol established in LaDI-VTON [21] to ensure fair comparison with previous virtual try-on models. + +Further information about the datasets and evaluation metrics is provided in the Supplementary Materials. + +# 4.1. Qualitative and Quantitative Results + +The quantitative Results of Table. 1 compare our proposed ITA-MDT framework to previous diffusion-based methods, as well as show effect of ITAFA, and SRE on MDT-IVTON. The values of the previous works are re-measured by implementation with their checkpoints released. As it can be observed, our method provides comparable, state-of-the-art, results with faster inference speed and smaller parameters. The qualitative results and comparison to previous methods are shown in Fig. 6, and more in Supplementary Materials. As part of the inpainting process, the generated images are blended with agnostic images. + +# 4.2. Ablation Study + +The qualitative results of our proposed ITA-MDT framework and effects of each components are available in the Supplementary Materials. MDT-IVTON alone captures the global structure well, including garment shapes, colors, and human pose, but falls short in fidelity of garment details. The integration of ITAFA and SRE addresses this limitation significantly, with SRE shown as effective as a high-resolution image with twice the width compared to SRE integrated condition vector $c$ . + +![](images/00868f013a1d495a31cfa7a11db618e0e3dd5fd19d4fc9ca92f613af4ffa65fc.jpg) + +![](images/c6b5400ceb27fdf9dfe7f00157c034af81f05d48f713ea4120fccbe5e50cac72.jpg) + +![](images/6081b1391c5c01fb49c668eb543e7a828b6fc8896d8d8d7dc41b6e7863c1be4a.jpg) +Figure 5. Visualization of the ITAFA's average weight distribution across ViT encoder layers of our final model. (a): use of hidden layers as denoising progresses. The dotted arrow refers to denoising direction. (b, c, d): changes in the use of hidden layers as the sparsity, variance, and gradient magnitude of garment image changes. The dotted arrow refers to the increase in these image complexity components. + +![](images/8cd65d56bb3fa52d9f682e2699dcfaf20426fec0df2b454b3d5cdca43d824c11.jpg) + +Fig. 5 visualizes how ITAFA leverages features from different layers of the ViT image encoder, DINOv2, across the denoising process. In Fig. 5 (a), where it observes the use of layers with respect to timesteps, the middle layers dominate the feature aggregation for much of the diffusion; however, a distinct shift in layer usage occurs from early to later layers as the timestep progresses. In the early timestep when the input is largely noise, the encoder's first layer is prominently utilized, and at the final timestep, the use of the + +![](images/067efeaff30710115ae707c1c8b3e8b953aab6375882236c3bff8468ef12a788.jpg) +Figure 6. Qualitative comparison of ITA-MDT with prior methods on VITON-HD and DressCode Upper-body datasets. The first three rows show VITON-HD results; the next three show DressCode results. + +last layer markedly increases. This observation supports our hypothesis that dynamically aggregating features from different layers as the diffusion progresses effectively optimizes feature utilization across denoising stages. In Fig. 5 (b, c, d), we observe a clear pattern where specific layers are consistently favored depending on the intensity of particular image complexity (e.g., feature sparsity, variance, or gradient magnitude). This indicates that our model adaptively activates layers containing information most relevant to the type of visual complexity, highlighting the value and adaptability of our ITAFA. + +# 5. Conclusion + +This paper presented ITA-MDT, a framework for the IVTON task that successfully leverages MDT, along with the proposed ITAFA and SRE modules, to address the challenges of transformer-based denoising diffusion models in preserving fine-grained details efficiently. Analysis confirmed that ITAFA and SRE effectively function as intended, enabling ITA-MDT to achieve state-of-the-art performance in several metrics with reduced model parameters and faster inference speed compared to previous methods. + +# Acknowledgments + +This work was supported by Institute for Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No.RS-2021-II211381, Development of Causal AI through Video Understanding and Reinforcement Learning, and Its Applications to Real Environments) and partly supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No.RS-2022-II220184, Development and Study of AI Technologies to Inexpensively Conform to Evolving Policy on Ethics) + +# References + +[1] Seunghwan Choi, Sunghyun Park, Minsoo Lee, and Jaegul Choo. Viton-hd: High-resolution virtual try-on via misalignment-aware normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14131-14140, 2021. 1, 6 +[2] Yisol Choi, Sangkyung Kwak, Kyungmin Lee, Hyungwon Choi, and Jinwoo Shin. Improving diffusion models for virtual try-on. arXiv preprint arXiv:2403.05139, 2024. 3, 7 +[3] Lavinia De Divitiis, Federico Becattini, Claudio Baecchi, and Alberto Del Bimbo. Disentangling features for fashion recommendation. ACM Transactions on Multimedia Computing, Communications and Applications, 19(1s):1-21, 2023. 1 +[4] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 3 +[5] Shanghua Gao, Pan Zhou, Ming-Ming Cheng, and Shuicheng Yan. Masked diffusion transformer is a strong image synthesizer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23164-23173, 2023. 2, 3, 4, 6 +[6] Yuying Ge, Yibing Song, Ruimao Zhang, Chongjian Ge, Wei Liu, and Ping Luo. Parser-free virtual try-on via distilling appearance flows. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8485-8493, 2021. 1 +[7] Junhong Gou, Siyu Sun, Jianfu Zhang, Jianlou Si, Chen Qian, and Liqing Zhang. Taming the power of diffusion models for high-quality virtual try-on with appearance flow. In Proceedings of the 31st ACM International Conference on Multimedia, pages 7599-7607, 2023. 3, 7 +[8] Ríza Alp Güler, Natalia Neverova, and Iasonas Kokkinos. Denseseposé: Dense human pose estimation in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7297-7306, 2018. 1 +[9] M Hadi Kiapour, Xufeng Han, Svetlana Lazebnik, Alexander C Berg, and Tamara L Berg. Where to buy it: Matching street clothing photos in online shops. In Proceedings of the IEEE international conference on computer vision, pages 3343-3351, 2015. 1 + +[10] Xintong Han, Zuxuan Wu, Zhe Wu, Ruichi Yu, and Larry S Davis. Viton: An image-based virtual try-on network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7543-7552, 2018. 1 +[11] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 7 +[12] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 3 +[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 3 +[14] Wei-Lin Hsiao and Kristen Grauman. Creating capsule wardrobes from fashion images. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7161-7170, 2018. 1 +[15] Thibaut Issenhuth, Jérémie Mary, and Clément Calauzenes. Do not mask what you do not need to mask: a parser-free virtual try-on. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XX 16, pages 619–635. Springer, 2020. 1 +[16] Jeongho Kim, Guojung Gu, Minho Park, Sunghyun Park, and Jaegul Choo. Stableviton: Learning semantic correspondence with latent diffusion model for virtual try-on. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8176-8185, 2024. 3, 7 +[17] Gwanhyeong Koo, Sunjae Yoon, Ji Woo Hong, and Chang D Yoo. Flexied: Frequency-aware latent refinement for enhanced non-rigid editing. In European Conference on Computer Vision, pages 363-379. Springer, 2024. 2 +[18] Sangyun Lee, Gyojung Gu, Sunghyun Park, Seunghwan Choi, and Jaegul Choo. High-resolution virtual try-on with misalignment and occlusion-handled conditions. In European Conference on Computer Vision, pages 204-219. Springer, 2022. 1 +[19] Shugao Ma, Tomas Simon, Jason Saragih, Dawei Wang, Yuecheng Li, Fernando De La Torre, and Yaser Sheikh. Pixel codec avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 64-73, 2021. 1 +[20] Davide Morelli, Matteo Fincato, Marcella Cornia, Federico Landi, Fabio Cesari, and Rita Cucchiara. Dress code: high-resolution multi-category virtual try-on. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2231-2235, 2022. 1, 7 +[21] Davide Morelli, Alberto Baldrati, Giuseppe Cartella, Marcelia Cornia, Marco Bertini, and Rita Cucchiara. Ladi-vton: latent diffusion textual-inversion enhanced virtual try-on. In Proceedings of the 31st ACM International Conference on Multimedia, pages 8580-8589, 2023. 3, 7 +[22] Shuliang Ning, Duomin Wang, Yipeng Qin, Zirong Jin, Baoyuan Wang, and Xiaoguang Han. Picture: Photorealistic virtual try-on from unconstrained designs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6976-6985, 2024. 3 + +[23] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 4 +[24] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 2, 3, 4 +[25] Trung X Pham, Zhang Kang, and Chang D Yoo. Cross-view masked diffusion transformers for person image synthesis. arXiv preprint arXiv:2402.01516, 2024. 2, 3 +[26] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 6, 4 +[27] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM international conference on multimedia, pages 484-492, 2020. 1 +[28] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International conference on machine learning, pages 8821-8831. Pmlr, 2021. 2, 3 +[29] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3 +[30] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 3 +[31] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 2, 3 +[32] Rohan Sarkar, Navaneeth Bodla, Mariya I Vasileva, Yen-Liang Lin, Anurag Beniwal, Alan Lu, and Gerard Medioni. Outfittransformer: Learning outfit representations for fashion recommendation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3601-3609, 2023. 1 +[33] Claude Elwood Shannon. A mathematical theory of communication. The Bell system technical journal, 27(3):379-423, 1948. 6, 2 +[34] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1 +[35] Yang Song and Stefano Ermon. Generative modeling by + +estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019. 3 +[36] Bochao Wang, Huabin Zheng, Xiaodan Liang, Yimin Chen, Liang Lin, and Meng Yang. Toward characteristic-preserving image-based virtual try-on network. In Proceedings of the European conference on computer vision (ECCV), pages 589-604, 2018. 1 +[37] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 7 +[38] Zhenyu Xie, Zaiyu Huang, Xin Dong, Fuwei Zhao, Haoye Dong, Xijin Zhang, Feida Zhu, and Xiaodan Liang. Gp-vton: Towards general purpose virtual try-on via collaborative localflow global-parsing learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23550-23559, 2023. 1 +[39] Sicheng Xu, Guojun Chen, Yu-Xiao Guo, Jiaolong Yang, Chong Li, Zhenyu Zang, Yizhong Zhang, Xin Tong, and Baining Guo. Vasa-1: Lifelike audio-driven talking faces generated in real time. arXiv preprint arXiv:2404.10667, 2024. 1 +[40] Han Yang, Ruimao Zhang, Xiaobao Guo, Wei Liu, Wangmeng Zuo, and Ping Luo. Towards photo-realistic virtual try-on by adaptively generating-preserving image content. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7850-7859, 2020. 1 +[41] Xu Yang, Changxing Ding, Zhibin Hong, Junhao Huang, Jin Tao, and Xiangmin Xu. Texture-preserving diffusion models for high-fidelity virtual try-on. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7017-7026, 2024. 3, 7 +[42] Sunjae Yoon, Gwanhyeong Koo, Ji Woo Hong, and Chang D Yoo. Dni: Dilutional noise initialization for diffusion video editing. In European Conference on Computer Vision, pages 180-195. Springer, 2024. 2 +[43] Sunjae Yoon, Gwanhyeong Koo, Geonwoo Kim, and Chang D Yoo. Frag: Frequency adapting group for diffusion video editing. arXiv preprint arXiv:2406.06044, 2024. +[44] Sunjae Yoon, Gwanhyeong Koo, Younghwan Lee, and Chang Yoo. Tpc: Test-time procrustes calibration for diffusion-based human image animation. Advances in Neural Information Processing Systems, 37:118654-118677, 2024. 2 +[45] Jianhao Zeng, Dan Song, Weizhi Nie, Hongshuo Tian, Tongtong Wang, and An-An Liu. Cat-dm: Controllable accelerated virtual try-on with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8372-8382, 2024. 3, 7 +[46] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7 +[47] Luyang Zhu, Dawei Yang, Tyler Zhu, Fitsum Reda, William Chan, Chitwan Sahara, Mohammad Norouzi, and Ira Kemelmacher-Shlizerman. Tryondiffusion: A tale of two unets. In Proceedings of the IEEE/CVF Conference on Com \ No newline at end of file diff --git a/2025/ITA-MDT_ Image-Timestep-Adaptive Masked Diffusion Transformer Framework for Image-Based Virtual Try-On/images.zip b/2025/ITA-MDT_ Image-Timestep-Adaptive Masked Diffusion Transformer Framework for Image-Based Virtual Try-On/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4fbfd782f23a9ba5ec2222900441950c348e3d4f --- /dev/null +++ b/2025/ITA-MDT_ Image-Timestep-Adaptive Masked Diffusion Transformer Framework for Image-Based Virtual Try-On/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5fe073ed6ca50d34eaaa74c170515a304a5e87e39fd2fd707c52a530bb67def3 +size 743034 diff --git a/2025/ITA-MDT_ Image-Timestep-Adaptive Masked Diffusion Transformer Framework for Image-Based Virtual Try-On/layout.json b/2025/ITA-MDT_ Image-Timestep-Adaptive Masked Diffusion Transformer Framework for Image-Based Virtual Try-On/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7a88389016a78937ac14dbf067402cd45ee2b398 --- /dev/null +++ b/2025/ITA-MDT_ Image-Timestep-Adaptive Masked Diffusion Transformer Framework for Image-Based Virtual Try-On/layout.json @@ -0,0 +1,8196 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 56, + 102, + 555, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 102, + 555, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 102, + 555, + 140 + ], + "type": "text", + "content": "ITA-MDT: Image-Timestep-Adaptive Masked Diffusion Transformer Framework for Image-Based Virtual Try-On" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 160, + 555, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 160, + 555, + 190 + ], + "spans": [ + { + "bbox": [ + 61, + 160, + 555, + 190 + ], + "type": "text", + "content": "Ji Woo Hong" + }, + { + "bbox": [ + 61, + 160, + 555, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 160, + 555, + 190 + ], + "type": "text", + "content": " Tri Ton" + }, + { + "bbox": [ + 61, + 160, + 555, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 160, + 555, + 190 + ], + "type": "text", + "content": " Trung X. Pham" + }, + { + "bbox": [ + 61, + 160, + 555, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 160, + 555, + 190 + ], + "type": "text", + "content": " Gwanhyeong Koo" + }, + { + "bbox": [ + 61, + 160, + 555, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 160, + 555, + 190 + ], + "type": "text", + "content": " Sunjae Yoon" + }, + { + "bbox": [ + 61, + 160, + 555, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 160, + 555, + 190 + ], + "type": "text", + "content": " Chang D. Yoo" + }, + { + "bbox": [ + 61, + 160, + 555, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 160, + 555, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 160, + 555, + 190 + ], + "type": "text", + "content": "Korea Advanced Institute of Science and Technology (KAIST), South Korea" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 191, + 500, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 191, + 500, + 202 + ], + "spans": [ + { + "bbox": [ + 115, + 191, + 500, + 202 + ], + "type": "text", + "content": "{jiwohong93, tritth, trungpx, kookie, dbstjswo505, cd_yoo}@kaist.ac.kr" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 56, + 230, + 554, + 344 + ], + "blocks": [ + { + "bbox": [ + 56, + 230, + 554, + 344 + ], + "lines": [ + { + "bbox": [ + 56, + 230, + 554, + 344 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 554, + 344 + ], + "type": "image", + "image_path": "ffcbc1c8930d7110b713b648e9c1b8d78e62c74edc0193f164ce4ac02682992b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 352, + 555, + 374 + ], + "lines": [ + { + "bbox": [ + 55, + 352, + 555, + 374 + ], + "spans": [ + { + "bbox": [ + 55, + 352, + 555, + 374 + ], + "type": "text", + "content": "Figure 1. Image-based Virtual Try-on results generated by our ITA-MDT. From left to right: VITON-HD[1], DressCode[20] upper, DressCode lower, and DressCode dresses, respectively." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 152, + 384, + 200, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 384, + 200, + 396 + ], + "spans": [ + { + "bbox": [ + 152, + 384, + 200, + 396 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 415, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 415, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 415, + 297, + 714 + ], + "type": "text", + "content": "This paper introduces ITA-MDT, the Image-Timestep-Adaptive Masked Diffusion Transformer Framework for Image-Based Virtual Try-On (IVTON), designed to overcome the limitations of previous approaches by leveraging the Masked Diffusion Transformer (MDT) for improved handling of both global garment context and fine-grained details. The IVTON task involves seamlessly superimposing a garment from one image onto a person in another, creating a realistic depiction of the person wearing the specified garment. Unlike conventional diffusion-based virtual try-on models that depend on large pre-trained U-Net architectures, ITA-MDT leverages a lightweight, scalable transformer-based denoising diffusion model with a mask latent modeling scheme, achieving competitive results while reducing computational overhead. A key component of ITAMDT is the Image-Timestep Adaptive Feature Aggregator (ITAFA), a dynamic feature aggregator that combines all of the features from the image encoder into a unified feature of the same size, guided by diffusion timestep and garment image complexity. This enables adaptive weighting of features, allowing the model to emphasize either global information or fine-grained details based on the requirements of the denoising stage. Additionally, the Salient Region Extractor (SRE) module is presented to identify complex region of the garment to provide high-resolution local infor" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 386, + 556, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 386, + 556, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 386, + 556, + 506 + ], + "type": "text", + "content": "mation to the denoising model as an additional condition alongside the global information of the full garment image. This targeted conditioning strategy enhances detail preservation of fine details in highly salient garment regions, optimizing computational resources by avoiding unnecessarily processing entire garment image. Comparative evaluations confirms that ITA-MDT improves efficiency while maintaining strong performance, reaching state-of-the-art results in several metrics. Our project page is available at https://jiwoohong93.github.io/ita-mdt/." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 525, + 394, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 525, + 394, + 538 + ], + "spans": [ + { + "bbox": [ + 314, + 525, + 394, + 538 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 545, + 556, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 545, + 556, + 701 + ], + "spans": [ + { + "bbox": [ + 312, + 545, + 556, + 701 + ], + "type": "text", + "content": "The task of Image-based Virtual Try-on (IVTON) is to generate an image where a person appears dressed in specific clothes, using provided images of both the garment and the individual. In real-world applications, IVTON holds great potential for revolutionizing online shopping, fashion design [3, 9, 14, 32], and even entertainment industries for creating digital avatars or virtual costumes [19, 27, 39]. For this task, deep learning methods [1, 6, 10, 15, 18, 36, 38, 40] employ paired datasets [1, 20] containing both standalone garment images and the images of the individual wearing these for training their models, along with additional reference images such as mask images of the garment region and DensePose [8] image." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 325, + 701, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 701, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 325, + 701, + 555, + 714 + ], + "type": "text", + "content": "Recently, diffusion models [13, 34] have shown great" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "28284" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "promise in image generation and synthesis in many different tasks [17, 25, 42-44]. Especially, large-scale pre-trained diffusion models [28, 29, 31], such as Stable Diffusion [29], are widely used to preserving high fidelity in generated images; however, their large size and slow inference speeds present practical limitations in commercial application." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 146, + 295, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 146, + 295, + 290 + ], + "spans": [ + { + "bbox": [ + 55, + 146, + 295, + 290 + ], + "type": "text", + "content": "To address this, transformer-based diffusion models have emerged as an efficient alternative. The Diffusion Transformer (DiT) [24] based on the Visual Transformer (ViT) [4] architecture with Latent Diffusion Models [29] can achieve excellent scalability and efficiency. Extending this concept, the Masked Diffusion Transformer (MDT) [5] applied mask prediction scheme to enhance spatial representation learning, achieving state-of-the-art results in class-conditional image generation. While promising, directly applying MDT to IVTON task presents challenges, as the task requires high fidelity and consistency across multiple reference images, demanding more than a straightforward approach." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 292, + 295, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 292, + 295, + 520 + ], + "spans": [ + { + "bbox": [ + 55, + 292, + 295, + 520 + ], + "type": "text", + "content": "This paper presents the ITA-MDT (Image-Timestep-Adaptive Masked Diffusion Transformer) Framework, which incorporates our modified version of MDT adapted specifically for IVTON, named MDT-IVTON, as well as two novel components: the Image-Timestep Adaptive Feature Aggregator (ITAFA) and the Salient Region Extractor (SRE). This framework is designed to preserve garment fidelity while maintaining computational efficiency. Our MDT-IVTON model leverages MDT's mask prediction scheme to efficiently learn spatial relationships between garment and individual. Additionally, we incorporate an auxiliary loss to enhance the model's focus on the inpainting objective of the task. While this architecture effectively captures global dependencies, it struggles to retain fine-grained details, which U-Net-based models handle relatively naturally through their multi-scale encoder-decoder structure that captures and reconstructs information of the image across multiple scales through progressive downsampling and upsampling. Our ITAFA and SRE bridge this gap." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "content": "The Image-Timestep Adaptive Feature Aggregator (ITAFA) is a learnable feature aggregator designed to condition the diffusion model effectively and efficiently. It dynamically combines multiple image features from ViT image encoder, guided by the current diffusion timestep and garment image complexity score. While simply providing multiple image features from different layers can condition the model with extensive information, this naive approach significantly increases the computational load. By assigning adaptive weights to each feature, ITAFA aggregates them into a single representation. These weights are modulated based on the timestep information, which indicates how far the generation process has progressed. The weighting is further refined using the visual complexity of the garment, which reflects how much fine-grained information the condition image contains. This design enables the model to gradually shift its focus" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 316, + 80, + 553, + 191 + ], + "blocks": [ + { + "bbox": [ + 316, + 80, + 553, + 191 + ], + "lines": [ + { + "bbox": [ + 316, + 80, + 553, + 191 + ], + "spans": [ + { + "bbox": [ + 316, + 80, + 553, + 191 + ], + "type": "image", + "image_path": "38953b5a0f8d37d04c01a469fda50e5a301ce13bb0c7ea02d0e5949abba92f5f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 213, + 555, + 258 + ], + "lines": [ + { + "bbox": [ + 313, + 213, + 555, + 258 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 555, + 258 + ], + "type": "text", + "content": "Figure 2. Comparison of our ITA-MDT with other diffusion-based methods on the VTON task using the VITON-HD dataset. For all metrics in the analysis, including FID, model parameters, and inference speed, lower values indicate better performance." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 281, + 553, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 553, + 304 + ], + "type": "text", + "content": "from global semantic features in the early stages to fine local details in the later stages of the diffusion process." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 306, + 555, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 306, + 555, + 484 + ], + "spans": [ + { + "bbox": [ + 313, + 306, + 555, + 484 + ], + "type": "text", + "content": "The Salient Region Extractor (SRE) provides high-resolution local information of key garment area, such as logos or complex patterns, identified through entropy-based detection. The region extracted offers additional crucial condition guidance to the diffusion model alongside the global context from the full garment image. While simply using a high-resolution, big size, image enhances detail representation, it results in substantial computation. Notably, most regions of garments are either solid-colored or display large, simple patterns that can be captured adequately through lower resolution. Therefore, rather than uniformly processing the entire image at high resolution, SRE selectively extracts and provides additional high-resolution information only for regions where fine-grained local information is crucial. In summary, our contributions include:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 315, + 486, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 315, + 486, + 555, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 486, + 555, + 570 + ], + "spans": [ + { + "bbox": [ + 315, + 486, + 555, + 570 + ], + "type": "text", + "content": "1. We present ITA-MDT, a framework for the image-based virtual try-on (IVTON) task that builds upon MDT-IVTON, our adaptation of the Masked Diffusion Transformer (MDT) for IVTON, and incorporates our proposed methods to effectively address the challenges of transformer-based denoising diffusion models in preserving fine-grained details." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 315, + 570, + 555, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 570, + 555, + 618 + ], + "spans": [ + { + "bbox": [ + 315, + 570, + 555, + 618 + ], + "type": "text", + "content": "2. We propose ITAFA, a learnable feature aggregator designed to dynamically combine features from the image encoder into a single feature based on the current denoising timestep and garment image complexity score." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 315, + 618, + 555, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 618, + 555, + 677 + ], + "spans": [ + { + "bbox": [ + 315, + 618, + 555, + 677 + ], + "type": "text", + "content": "3. We propose SRE to provide high-resolution local information of key garment area, identified through entropy-based detection, as additional condition guidance to the diffusion model alongside the global context from the full garment image." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 315, + 677, + 555, + 713 + ], + "type": "text", + "content": "4. ITA-MDT's combination of MDT-IVTON, ITAFA, and SRE offers a highly efficient solution for IVTON, achieving a balance between model efficiency and output fi" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28285" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 72, + 298, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 298, + 145 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 298, + 145 + ], + "type": "text", + "content": "delity. Figure 2 presents the comparison of our ITA-MDT with previous diffusion-based methods. It achieves state-of-the-art results in several metrics while maintaining competitive performance across the others, demonstrating its overall effectiveness and suitability for practical virtual try-on applications." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 156, + 142, + 169 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 156, + 142, + 169 + ], + "spans": [ + { + "bbox": [ + 55, + 156, + 142, + 169 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 177, + 279, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 177, + 279, + 190 + ], + "spans": [ + { + "bbox": [ + 55, + 177, + 279, + 190 + ], + "type": "text", + "content": "2.1. Image-based Virtual Try-on with Diffusion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 195, + 297, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 195, + 297, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 195, + 297, + 483 + ], + "type": "text", + "content": "The integration of diffusion models has significantly advanced the field of Image-based Virtual Try-on, as evidenced in numerous recent studies. Notably, TryOnDiffusion [47] introduced an architecture utilizing dual U-Nets, highlighting the potential of diffusion-based try-on but requiring large-scale datasets composed of image pairs of the same person in varied poses. This large-scale dataset requirement has led researchers to pivot towards leveraging large pre-trained diffusion models [28, 29, 31]. To condition these models effectively, methods such as representing garments as pseudowords in LaDI-VTON [21], applying warping networks in DCI-VTON [7], modifying attention mechanisms in StableVITON [16] and IDM-VTON [2], utilizing ControlNet for garment control and GAN-based sampling in CAT-DM [45], and concatenating masked person and garment images for direct texture transfer in TPD [41] have been introduced. However, the use of large pre-trained U-Nets comes with notable limitations. The large number of parameters in these models significantly increases memory usage and slows down inference speed, making them resource-intensive and less practical for efficient deployment. A variety of solutions could be explored to address these issues, among which transformer-based diffusion models present a promising approach." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 492, + 243, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 492, + 243, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 492, + 243, + 504 + ], + "type": "text", + "content": "2.2. Transformers for Diffusion Process" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 510, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 297, + 714 + ], + "type": "text", + "content": "Traditionally, The U-Net architecture [30] has been foundational for diffusion models and continues to be a prevalent choice across various diffusion-based generation tasks [13, 35]. However, the incorporation of transformers into diffusion processes also marked a significant evolution in generative modeling. The introduction of the Diffusion Transformer (DiT) [24] merged the architecture of the Vision Transformer (ViT) [4] with latent diffusion models. DiT has demonstrated outstanding scalability and compute efficiency, delivering superior generative performance compared to traditional U-Net architectures. Building on this, the Masked Diffusion Transformer (MDT) [5] has further advanced the diffusion transformer model by employing an asymmetric masking schedule to optimize the contextual representation learning, setting a new benchmark in class image generation on ImageNet. In an attempt to leverage MDT for image-to-image generation, [25] introduced an aggregation network" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 556, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 556, + 193 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 556, + 193 + ], + "type": "text", + "content": "that consolidates image conditions into a single vector for guiding the diffusion process in pose-guided human image generation tasks. While effective within its domain, directly applying this approach to the IVTON task is unsuitable. IVTON involves a significantly larger set of condition images, and compressing these into a single vector leads to substantial information loss, ultimately degrading output quality. Given the need to preserve fine details from condition images, enhancing rather than compressing condition information is preferred." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 198, + 555, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 198, + 555, + 224 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 555, + 224 + ], + "type": "text", + "content": "2.3. Enhancing Condition Information in Virtual Try-On task" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 228, + 556, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 556, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 556, + 431 + ], + "type": "text", + "content": "To enhance the intricate details essential to IVTON, the recent method, PICTURE [22], introduces a two-stage disentanglement pipeline that separates style and texture features by employing a hierarchical CLIP feature extraction module with position encoding, enabling the representation of complex, non-stationary textures. While effective in handling intricate style and texture combinations, this method requires an in-depth analysis of features from the hidden layers of the image encoder. Specifically, this involves clustering hidden layer features and manually selecting representative features from each cluster, a process that lacks flexibility for it requires re-evaluation when a different encoder is to be used. Additionally, the selected features from each cluster are concatenated for cross-attention conditioning, substantially increasing computational demand. Together, these requirements render the approach resource-intensive and less adaptable to diverse architectures." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 441, + 370, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 370, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 370, + 453 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 461, + 556, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 556, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 556, + 582 + ], + "type": "text", + "content": "The ITA-MDT framework performs Image-Based Virtual Try-on through a transformer-based denoising diffusion network, leveraging the mask latent modeling scheme of MDTv2 [5]. An overview of the architecture is shown in Figure 3, which comprises three primary components: (1) the Masked Diffusion Transformer for Image-Based Virtual Try-On (MDT-IVTON), (2) the Image-Timestep Adaptive Feature Aggregator (ITAFA), and (3) the Salient Region Extractor (SRE) module. Detailed descriptions of each component are provided below." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 588, + 555, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 588, + 555, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 555, + 612 + ], + "type": "text", + "content": "3.1. Masked Diffusion Transformer for Image-Based Virtual Try-On (MDT-IVTON)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 617, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 557, + 715 + ], + "type": "text", + "content": "Our ITA-MDT framework integrates a Masked Diffusion Transformer (MDT) tailored for Image-Based Virtual Try-On (IVTON), leveraging the transformer-based denoising diffusion model architecture of DiT [24]. This model operates in a latent space encoded via a pre-trained VAE following the Latent Diffusion Model (LDM) approach [29]. Specifically, the VAE encoder maps garment agnostic inputs into this latent space, including the Garment Agnostic" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "28286" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 553, + 272 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 272 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 272 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 272 + ], + "type": "image", + "image_path": "e23e4b455c2609c241fb95d6d7598db11cf16d0a32ee3dbbd0d3056454d1452d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 278, + 555, + 369 + ], + "lines": [ + { + "bbox": [ + 54, + 278, + 555, + 369 + ], + "spans": [ + { + "bbox": [ + 54, + 278, + 555, + 369 + ], + "type": "text", + "content": "Figure 3. Overview of the ITA-MDT Framework for Image-Based Virtual Try-On (IVTON). The framework takes multiple reference images encoded into latent space as query, which includes Garment Agnostic Map " + }, + { + "bbox": [ + 54, + 278, + 555, + 369 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 54, + 278, + 555, + 369 + ], + "type": "text", + "content": ", DensePose " + }, + { + "bbox": [ + 54, + 278, + 555, + 369 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 54, + 278, + 555, + 369 + ], + "type": "text", + "content": ", and Garment Agnostic Mask " + }, + { + "bbox": [ + 54, + 278, + 555, + 369 + ], + "type": "inline_equation", + "content": "M_X" + }, + { + "bbox": [ + 54, + 278, + 555, + 369 + ], + "type": "text", + "content": ". These reference latent images are concatenated to be patch embedded for masked diffusion process within our MDT-IVTON, which follows the architecture of MDTv2 [5] with integrated cross-attention blocks. The image feature of Garment " + }, + { + "bbox": [ + 54, + 278, + 555, + 369 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 54, + 278, + 555, + 369 + ], + "type": "text", + "content": " is extracted with the ViT image encoder, DINOv2 [23], and are adaptively aggregated with our proposed Image-Timestep Adaptive Feature Aggregator (ITAFA) to produce Garment Feature " + }, + { + "bbox": [ + 54, + 278, + 555, + 369 + ], + "type": "inline_equation", + "content": "F_g" + }, + { + "bbox": [ + 54, + 278, + 555, + 369 + ], + "type": "text", + "content": ". With our Salient Region Extractor (SRE), the Salient Region " + }, + { + "bbox": [ + 54, + 278, + 555, + 369 + ], + "type": "inline_equation", + "content": "X_s" + }, + { + "bbox": [ + 54, + 278, + 555, + 369 + ], + "type": "text", + "content": " is extracted from the Garment " + }, + { + "bbox": [ + 54, + 278, + 555, + 369 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 54, + 278, + 555, + 369 + ], + "type": "text", + "content": " and processed through ITAFA separately to produce Salient Feature " + }, + { + "bbox": [ + 54, + 278, + 555, + 369 + ], + "type": "inline_equation", + "content": "F_s" + }, + { + "bbox": [ + 54, + 278, + 555, + 369 + ], + "type": "text", + "content": ". The Garment Feature " + }, + { + "bbox": [ + 54, + 278, + 555, + 369 + ], + "type": "inline_equation", + "content": "F_g" + }, + { + "bbox": [ + 54, + 278, + 555, + 369 + ], + "type": "text", + "content": " and Salient Feature " + }, + { + "bbox": [ + 54, + 278, + 555, + 369 + ], + "type": "inline_equation", + "content": "F_s" + }, + { + "bbox": [ + 54, + 278, + 555, + 369 + ], + "type": "text", + "content": " are concatenated to serve as conditions of MDT-IVTON. Positional embeddings are omitted in this figure for simplicity." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 388, + 295, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 388, + 295, + 412 + ], + "spans": [ + { + "bbox": [ + 54, + 388, + 295, + 412 + ], + "type": "text", + "content": "Map " + }, + { + "bbox": [ + 54, + 388, + 295, + 412 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 54, + 388, + 295, + 412 + ], + "type": "text", + "content": ", DensePose " + }, + { + "bbox": [ + 54, + 388, + 295, + 412 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 54, + 388, + 295, + 412 + ], + "type": "text", + "content": ", and Garment Agnostic Mask " + }, + { + "bbox": [ + 54, + 388, + 295, + 412 + ], + "type": "inline_equation", + "content": "M_X" + }, + { + "bbox": [ + 54, + 388, + 295, + 412 + ], + "type": "text", + "content": ", concatenated with the noised latent representation " + }, + { + "bbox": [ + 54, + 388, + 295, + 412 + ], + "type": "inline_equation", + "content": "z_t" + }, + { + "bbox": [ + 54, + 388, + 295, + 412 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 414, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 414, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 54, + 414, + 295, + 581 + ], + "type": "text", + "content": "The MDT-IVTON model processes the input by embedding it into a patchified latent representation " + }, + { + "bbox": [ + 54, + 414, + 295, + 581 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 54, + 414, + 295, + 581 + ], + "type": "text", + "content": " and sequentially passing it through " + }, + { + "bbox": [ + 54, + 414, + 295, + 581 + ], + "type": "inline_equation", + "content": "N_{1}" + }, + { + "bbox": [ + 54, + 414, + 295, + 581 + ], + "type": "text", + "content": " Encoder Blocks followed by " + }, + { + "bbox": [ + 54, + 414, + 295, + 581 + ], + "type": "inline_equation", + "content": "N_{2}" + }, + { + "bbox": [ + 54, + 414, + 295, + 581 + ], + "type": "text", + "content": " Decoder Blocks. Each block utilizes Layer Normalization, Multi-Head Self-Attention, Multi-Head Cross Attention, Multi-Layer Perceptron (MLP), and skip connections to refine the denoised latent representation " + }, + { + "bbox": [ + 54, + 414, + 295, + 581 + ], + "type": "inline_equation", + "content": "z_{t - 1}" + }, + { + "bbox": [ + 54, + 414, + 295, + 581 + ], + "type": "text", + "content": ", as illustrated in the right side of Figure 3. Conditioning information derived from the concatenated garment features " + }, + { + "bbox": [ + 54, + 414, + 295, + 581 + ], + "type": "inline_equation", + "content": "F_{g}" + }, + { + "bbox": [ + 54, + 414, + 295, + 581 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 414, + 295, + 581 + ], + "type": "inline_equation", + "content": "F_{s}" + }, + { + "bbox": [ + 54, + 414, + 295, + 581 + ], + "type": "text", + "content": ", extracted from garment image " + }, + { + "bbox": [ + 54, + 414, + 295, + 581 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 54, + 414, + 295, + 581 + ], + "type": "text", + "content": " and its salient region " + }, + { + "bbox": [ + 54, + 414, + 295, + 581 + ], + "type": "inline_equation", + "content": "X_{s}" + }, + { + "bbox": [ + 54, + 414, + 295, + 581 + ], + "type": "text", + "content": " using the ViT encoder and ITAFA module, is infused through cross-attention mechanisms to guide the generation process. Detailed descriptions of ITAFA and SRE are available in Sec. 3.2 and Sec. 3.3." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 584, + 295, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 584, + 295, + 608 + ], + "spans": [ + { + "bbox": [ + 55, + 584, + 295, + 608 + ], + "type": "text", + "content": "Three objective functions guide the MDT-IVTON training to optimize its generative performance:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 621, + 170, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 621, + 170, + 634 + ], + "spans": [ + { + "bbox": [ + 55, + 621, + 170, + 634 + ], + "type": "text", + "content": "3.1.1. Denoising Objective" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 639, + 295, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 639, + 295, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 639, + 295, + 687 + ], + "type": "text", + "content": "The primary objective function minimizes the mean squared error (MSE) between the predicted noise and the actual noise in the denoised latent " + }, + { + "bbox": [ + 55, + 639, + 295, + 687 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 55, + 639, + 295, + 687 + ], + "type": "text", + "content": " at each timestep " + }, + { + "bbox": [ + 55, + 639, + 295, + 687 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 639, + 295, + 687 + ], + "type": "text", + "content": ", following the standard diffusion objective:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 80, + 700, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 700, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 80, + 700, + 295, + 715 + ], + "type": "interline_equation", + "content": "L _ {\\text {d e n o i s e}} = \\mathcal {E} _ {y, c, \\epsilon \\sim \\mathcal {N} (0, I), t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(y _ {t}, c, t\\right) \\| ^ {2} \\right], \\tag {1}", + "image_path": "ea4138fca6dac13fe3be779f58ea422bbc9a92971f95d7afe041a86b70be9481.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 388, + 553, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 553, + 412 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 553, + 412 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 388, + 553, + 412 + ], + "type": "inline_equation", + "content": "y_{t}" + }, + { + "bbox": [ + 313, + 388, + 553, + 412 + ], + "type": "text", + "content": " denotes the latent representation at timestep " + }, + { + "bbox": [ + 313, + 388, + 553, + 412 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 388, + 553, + 412 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 388, + 553, + 412 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 388, + 553, + 412 + ], + "type": "text", + "content": " the conditioning, and " + }, + { + "bbox": [ + 313, + 388, + 553, + 412 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 313, + 388, + 553, + 412 + ], + "type": "text", + "content": " the model's predicted noise." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 417, + 477, + 429 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 417, + 477, + 429 + ], + "spans": [ + { + "bbox": [ + 313, + 417, + 477, + 429 + ], + "type": "text", + "content": "3.1.2. Mask Reconstruction Objective" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 431, + 555, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 555, + 503 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 555, + 503 + ], + "type": "text", + "content": "In MDT-IVTON, the Mask Reconstruction Objective enforces spatial coherence and semantic consistency by reconstructing masked tokens within the masked latent representation " + }, + { + "bbox": [ + 313, + 431, + 555, + 503 + ], + "type": "inline_equation", + "content": "E_{m}" + }, + { + "bbox": [ + 313, + 431, + 555, + 503 + ], + "type": "text", + "content": ". This objective ensures alignment between the reference and generated images, enhancing the model's ability to synthesize realistic outputs in masked regions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 503, + 555, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 503, + 555, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 503, + 555, + 586 + ], + "type": "text", + "content": "To achieve this, we incorporate the Side-Interpolator module, which plays a pivotal role in interpolating information across unmasked tokens to predict values for masked tokens in " + }, + { + "bbox": [ + 313, + 503, + 555, + 586 + ], + "type": "inline_equation", + "content": "E_{m}" + }, + { + "bbox": [ + 313, + 503, + 555, + 586 + ], + "type": "text", + "content": ". The Side-Interpolator effectively fills in missing information by conditioning on neighboring unmasked regions, encouraging consistent token-level synthesis within the masked latent." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 587, + 555, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 587, + 555, + 611 + ], + "spans": [ + { + "bbox": [ + 313, + 587, + 555, + 611 + ], + "type": "text", + "content": "The Mask Reconstruction Objective follows an MSE formulation similar to the denoising objective:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 318, + 616, + 555, + 641 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 616, + 555, + 641 + ], + "spans": [ + { + "bbox": [ + 318, + 616, + 555, + 641 + ], + "type": "interline_equation", + "content": "L _ {\\text {m a s k}} = \\mathcal {E} _ {y, c, \\epsilon \\sim \\mathcal {N} (0, I), t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (\\text {S i d e - I n t} (E _ {m}), c, t) \\| ^ {2} \\right], \\tag {2}", + "image_path": "2b9d44c9f5bb3109610b68bf7c6104478b6d645f4a27d5ca60c70f60be6062c0.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "where Side-Int processes " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "E_{m}" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": ", leveraging spatial context to predict masked regions. The conditioning " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": " includes garment details, aiding in preserving fidelity to the garment appearance in masked areas. This objective promotes detailed reconstruction by interpolating across the context in " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "E_{m}" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": ", enhancing semantic alignment within generated images." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28287" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 172, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 172, + 84 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 172, + 84 + ], + "type": "text", + "content": "3.1.3. Inpainting Objective" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 87, + 296, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 87, + 296, + 182 + ], + "spans": [ + { + "bbox": [ + 55, + 87, + 296, + 182 + ], + "type": "text", + "content": "The inpainting objective in MDT-IVTON is designed to enhance garment fidelity by concentrating on regions defined by the garment agnostic mask " + }, + { + "bbox": [ + 55, + 87, + 296, + 182 + ], + "type": "inline_equation", + "content": "M_X" + }, + { + "bbox": [ + 55, + 87, + 296, + 182 + ], + "type": "text", + "content": ", encoded as " + }, + { + "bbox": [ + 55, + 87, + 296, + 182 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(M_X)" + }, + { + "bbox": [ + 55, + 87, + 296, + 182 + ], + "type": "text", + "content": " in latent space. During training, this mask serves as a selective weighting factor in the mean squared error (MSE) between the target and predicted outputs, guiding the model's focus to reconstruct details more accurately within garment-specific regions." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 183, + 296, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 183, + 296, + 254 + ], + "spans": [ + { + "bbox": [ + 55, + 183, + 296, + 254 + ], + "type": "text", + "content": "To ensure numerical stability and maintain well-behaved gradients, the mask is normalized to the range [0, 1], preventing extreme values that could otherwise destabilize training. This targeted inpainting strategy enhances garment detail fidelity by directing the model's attention to areas where high-quality reconstruction is most critical." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 255, + 296, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 296, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 296, + 326 + ], + "type": "text", + "content": "Latent Mask Normalization: To prevent unstable gradients, the single channel of the latent mask " + }, + { + "bbox": [ + 55, + 255, + 296, + 326 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(M_X)" + }, + { + "bbox": [ + 55, + 255, + 296, + 326 + ], + "type": "text", + "content": " that most closely resembles the binary mask of the inpainting task is chosen to be used as the mask across all channels in the loss calculation. This selected channel is normalized to values between 0 and 1 to ensure stability:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 332, + 296, + 367 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 332, + 296, + 367 + ], + "spans": [ + { + "bbox": [ + 61, + 332, + 296, + 367 + ], + "type": "interline_equation", + "content": "\\mathcal {E} \\left(M _ {X}\\right) = \\frac {\\mathcal {E} \\left(M _ {X}\\right) - \\operatorname* {m i n} \\left(\\mathcal {E} \\left(M _ {X}\\right)\\right)}{\\operatorname* {m a x} \\left(\\mathcal {E} \\left(M _ {X}\\right)\\right) - \\operatorname* {m i n} \\left(\\mathcal {E} \\left(M _ {X}\\right)\\right) + 1 \\times 1 0 ^ {- 8}}, \\tag {3}", + "image_path": "9756a49296a63f8e62a8463ccac906ccd93fd62283e4566ed4d41e672db7dc1e.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 368, + 296, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 368, + 296, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 368, + 296, + 415 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 368, + 296, + 415 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-8}" + }, + { + "bbox": [ + 55, + 368, + 296, + 415 + ], + "type": "text", + "content": " ensures numerical stability. This approach avoids unintended gradients and aligns the loss with garment-focused regions. A visualization of the latent mask selected is in the Supplementary Materials." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 416, + 296, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 416, + 296, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 416, + 296, + 464 + ], + "type": "text", + "content": "Masked Loss Computation: With the rescaled single channel of " + }, + { + "bbox": [ + 55, + 416, + 296, + 464 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(M_X)" + }, + { + "bbox": [ + 55, + 416, + 296, + 464 + ], + "type": "text", + "content": ", the masked inpainting loss is calculated as a weighted mean squared error (MSE) applied uniformly across all channels in the latent space:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 86, + 469, + 296, + 485 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 469, + 296, + 485 + ], + "spans": [ + { + "bbox": [ + 86, + 469, + 296, + 485 + ], + "type": "interline_equation", + "content": "L _ {\\text {i n p a i n t}} = \\mathcal {E} _ {y, c, t} \\left[ \\| \\mathcal {E} (M _ {X}) \\cdot \\left(y _ {t} - y _ {t - 1}\\right) \\| ^ {2} \\right], \\tag {4}", + "image_path": "713bfc5455ad2695acc4f33f005d63d3df407f6a7ba71575f727305133d2dbf0.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 491, + 296, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 296, + 551 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 296, + 551 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 491, + 296, + 551 + ], + "type": "inline_equation", + "content": "y_{t}" + }, + { + "bbox": [ + 55, + 491, + 296, + 551 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 491, + 296, + 551 + ], + "type": "inline_equation", + "content": "y_{t-1}" + }, + { + "bbox": [ + 55, + 491, + 296, + 551 + ], + "type": "text", + "content": " denote the latent representations at timesteps " + }, + { + "bbox": [ + 55, + 491, + 296, + 551 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 491, + 296, + 551 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 491, + 296, + 551 + ], + "type": "inline_equation", + "content": "t-1" + }, + { + "bbox": [ + 55, + 491, + 296, + 551 + ], + "type": "text", + "content": ", respectively. This loss term emphasizes reconstruction quality in the masked regions, promoting accurate garment-specific detail representation. Further details of MDT-IVTON are provided in Supplementary Materials." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 556, + 294, + 581 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 556, + 294, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 294, + 581 + ], + "type": "text", + "content": "3.2. Image-Timestep Adaptive Feature Aggregator (ITAFA)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 586, + 296, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 296, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 296, + 670 + ], + "type": "text", + "content": "The ITAFA is a dynamic feature aggregation module designed to adaptively weight feature embeddings from a Vision Transformer (ViT) image encoder, utilizing learnable parameters and feature complexity analysis to dynamically modulate the contribution of each hidden layer output of the image encoder at different timesteps of the denoising diffusion process." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 674, + 208, + 686 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 674, + 208, + 686 + ], + "spans": [ + { + "bbox": [ + 55, + 674, + 208, + 686 + ], + "type": "text", + "content": "3.2.1. Components and Parameters" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Timestep Embedding Projector. The diffusion timestep embedding, denoted as " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": ", represents the current timestep in" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": "the denoising process with dimension " + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": " of the MDT-IVTON model. To align " + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": " with the image feature dimension, it is projected to the dimensionality of the hidden layer of the ViT, " + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\varepsilon}" + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": ". This is achieved through a linear transformation:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 386, + 125, + 555, + 140 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 125, + 555, + 140 + ], + "spans": [ + { + "bbox": [ + 386, + 125, + 555, + 140 + ], + "type": "interline_equation", + "content": "W ^ {T} = \\operatorname {L i n e a r} (\\mathcal {H}, \\mathcal {H} _ {\\mathcal {E}}), \\tag {5}", + "image_path": "bc90b19f3f61590044f4ded0d6d6f035aa0dfc2edd902ff058dbd79f261e2fdd.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 145, + 555, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 145, + 555, + 170 + ], + "spans": [ + { + "bbox": [ + 313, + 145, + 555, + 170 + ], + "type": "text", + "content": "The linear layer maps the timestep embedding into a representation suitable for interacting with the ViT features." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 183, + 555, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 183, + 555, + 231 + ], + "spans": [ + { + "bbox": [ + 313, + 183, + 555, + 231 + ], + "type": "text", + "content": "Image Complexity Projector. ITAFA incorporates three complexity components: feature sparsity, feature variance, and gradient magnitude. These collectively capture the complexity of garment texture and structure." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 232, + 554, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 232, + 554, + 279 + ], + "spans": [ + { + "bbox": [ + 315, + 232, + 554, + 279 + ], + "type": "text", + "content": "1. Feature Sparsity measures the proportion of near-zero activations in the feature embedding tensor " + }, + { + "bbox": [ + 315, + 232, + 554, + 279 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 315, + 232, + 554, + 279 + ], + "type": "text", + "content": ", providing insight into the sparsity of learned representations. It is computed as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 392, + 285, + 555, + 316 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 285, + 555, + 316 + ], + "spans": [ + { + "bbox": [ + 392, + 285, + 555, + 316 + ], + "type": "interline_equation", + "content": "\\mathrm {S} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\mathbb {I} \\left(\\left| f _ {i} \\right| < \\delta\\right), \\tag {6}", + "image_path": "deca4065c6ab0cec47a7094b1645f5d8416d087b68317cc99dd31a1e51279dde.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 325, + 321, + 554, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 321, + 554, + 356 + ], + "spans": [ + { + "bbox": [ + 325, + 321, + 554, + 356 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 325, + 321, + 554, + 356 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 325, + 321, + 554, + 356 + ], + "type": "text", + "content": " is a small threshold (i.e., 0.01), " + }, + { + "bbox": [ + 325, + 321, + 554, + 356 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 325, + 321, + 554, + 356 + ], + "type": "text", + "content": " denotes the total number of elements in " + }, + { + "bbox": [ + 325, + 321, + 554, + 356 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 325, + 321, + 554, + 356 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 325, + 321, + 554, + 356 + ], + "type": "inline_equation", + "content": "\\mathbb{I}(\\cdot)" + }, + { + "bbox": [ + 325, + 321, + 554, + 356 + ], + "type": "text", + "content": " is the indicator function." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 357, + 554, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 554, + 381 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 554, + 381 + ], + "type": "text", + "content": "2. Feature Variance captures the variability of activations, reflecting structural complexity within the image:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 413, + 388, + 555, + 401 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 413, + 388, + 555, + 401 + ], + "spans": [ + { + "bbox": [ + 413, + 388, + 555, + 401 + ], + "type": "interline_equation", + "content": "V = \\operatorname {V a r} (f), \\tag {7}", + "image_path": "6a07e9bc302543aa4c1c0601ab32e068e8b2d3b0004ae4ad6d2e90a768c830b3.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 325, + 407, + 554, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 407, + 554, + 430 + ], + "spans": [ + { + "bbox": [ + 325, + 407, + 554, + 430 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 325, + 407, + 554, + 430 + ], + "type": "inline_equation", + "content": "\\operatorname{Var}(f)" + }, + { + "bbox": [ + 325, + 407, + 554, + 430 + ], + "type": "text", + "content": " denotes the variance across all elements in the feature embedding." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 430, + 555, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 555, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 555, + 491 + ], + "type": "text", + "content": "3. Gradient Magnitude measures local changes in the feature embedding, calculated by taking the spatial gradients along patch and embedding dimensions. The magnitude of gradients are computed, and averaging them provides a single score " + }, + { + "bbox": [ + 313, + 430, + 555, + 491 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 430, + 555, + 491 + ], + "type": "text", + "content": " that captures texture complexity." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 369, + 499, + 555, + 533 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 499, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 369, + 499, + 555, + 533 + ], + "type": "interline_equation", + "content": "G = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\sqrt {\\left(\\frac {\\partial f}{\\partial x}\\right) ^ {2} + \\left(\\frac {\\partial f}{\\partial y}\\right) ^ {2}}. \\tag {8}", + "image_path": "363520666e4cf0481c3655087233a5727a7fb9e5922319f0fffd53419562c83d.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 536, + 554, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 536, + 554, + 571 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 554, + 571 + ], + "type": "text", + "content": "These components are concatenated to form the complexity score vector " + }, + { + "bbox": [ + 313, + 536, + 554, + 571 + ], + "type": "inline_equation", + "content": "[S,V,G]" + }, + { + "bbox": [ + 313, + 536, + 554, + 571 + ], + "type": "text", + "content": ", which is then linearly projected to match the hidden layer dimension " + }, + { + "bbox": [ + 313, + 536, + 554, + 571 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\mathcal{E}}" + }, + { + "bbox": [ + 313, + 536, + 554, + 571 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 388, + 577, + 555, + 591 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 577, + 555, + 591 + ], + "spans": [ + { + "bbox": [ + 388, + 577, + 555, + 591 + ], + "type": "interline_equation", + "content": "W ^ {I} = \\operatorname {L i n e a r} (3, \\mathcal {H} _ {\\mathcal {E}}). \\tag {9}", + "image_path": "8173a315d724c4eba59c61d3cc78b1b0e01aa63071df9152c60b19aac8d85caa.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 604, + 555, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 604, + 555, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 604, + 555, + 652 + ], + "type": "text", + "content": "Learnable Weighting Parameter. The contribution of the timestep and complexity projections is balanced by a learnable parameter " + }, + { + "bbox": [ + 313, + 604, + 555, + 652 + ], + "type": "inline_equation", + "content": "\\alpha \\in [0,1]" + }, + { + "bbox": [ + 313, + 604, + 555, + 652 + ], + "type": "text", + "content": ", initialized to 0.5 and optimized during training:" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 363, + 657, + 555, + 672 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 657, + 555, + 672 + ], + "spans": [ + { + "bbox": [ + 363, + 657, + 555, + 672 + ], + "type": "interline_equation", + "content": "W ^ {c o m} = \\alpha \\cdot W ^ {T} + (1 - \\alpha) \\cdot W ^ {I}. \\tag {10}", + "image_path": "604e95402ba4be10a2b984d001c913f713ff02947683322bfab0522af3cc1a53.jpg" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": "This weighted combination provides an adaptive control over the influence of timestep and image complexity on the image feature aggregation." + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28288" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 87, + 262, + 248 + ], + "blocks": [ + { + "bbox": [ + 59, + 75, + 237, + 85 + ], + "lines": [ + { + "bbox": [ + 59, + 75, + 237, + 85 + ], + "spans": [ + { + "bbox": [ + 59, + 75, + 237, + 85 + ], + "type": "text", + "content": "Image-Timestep Adaptive Feature Aggregator" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 87, + 262, + 248 + ], + "lines": [ + { + "bbox": [ + 58, + 87, + 262, + 248 + ], + "spans": [ + { + "bbox": [ + 58, + 87, + 262, + 248 + ], + "type": "image", + "image_path": "549b3d146ae9897987a9bc98cd0747da5e80a6044a7b7e51870a33efd1cdd583.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 267, + 87, + 553, + 247 + ], + "blocks": [ + { + "bbox": [ + 385, + 75, + 481, + 85 + ], + "lines": [ + { + "bbox": [ + 385, + 75, + 481, + 85 + ], + "spans": [ + { + "bbox": [ + 385, + 75, + 481, + 85 + ], + "type": "text", + "content": "Salient Region Extractor" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 267, + 87, + 553, + 247 + ], + "lines": [ + { + "bbox": [ + 267, + 87, + 553, + 247 + ], + "spans": [ + { + "bbox": [ + 267, + 87, + 553, + 247 + ], + "type": "image", + "image_path": "7c2c18b4ff01c3d191f55a76bf5e65eb51c240fbc2873b81856d965658da3040.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 257, + 555, + 357 + ], + "lines": [ + { + "bbox": [ + 54, + 257, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 54, + 257, + 555, + 357 + ], + "type": "text", + "content": "Figure 4. Overview of the Image-Timestep Adaptive Feature Aggregator (ITAFA) and the Salient Region Extractor (SRE). Left: The ITAFA dynamically aggregates Vision Transformer (ViT) feature embeddings " + }, + { + "bbox": [ + 54, + 257, + 555, + 357 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 54, + 257, + 555, + 357 + ], + "type": "text", + "content": " based on a combination of the Timestep Embedding Projector, which projects the diffusion timestep embedding " + }, + { + "bbox": [ + 54, + 257, + 555, + 357 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 54, + 257, + 555, + 357 + ], + "type": "text", + "content": " to match the feature embedding dimensions, and the Image Complexity Projector, which transforms the image complexity vector " + }, + { + "bbox": [ + 54, + 257, + 555, + 357 + ], + "type": "inline_equation", + "content": "[S, V, G]" + }, + { + "bbox": [ + 54, + 257, + 555, + 357 + ], + "type": "text", + "content": " (sparsity, variance, gradient magnitude) into a comparable dimension. The weight vectors are combined and normalized via softmax to form " + }, + { + "bbox": [ + 54, + 257, + 555, + 357 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 54, + 257, + 555, + 357 + ], + "type": "text", + "content": ", which is used to adaptively aggregate the feature embeddings " + }, + { + "bbox": [ + 54, + 257, + 555, + 357 + ], + "type": "inline_equation", + "content": "\\{f_i\\}_{i=0}^h" + }, + { + "bbox": [ + 54, + 257, + 555, + 357 + ], + "type": "text", + "content": " across hidden layers to produce the final output tensor " + }, + { + "bbox": [ + 54, + 257, + 555, + 357 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 54, + 257, + 555, + 357 + ], + "type": "text", + "content": ". Garment " + }, + { + "bbox": [ + 54, + 257, + 555, + 357 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 54, + 257, + 555, + 357 + ], + "type": "text", + "content": " and Salient Region " + }, + { + "bbox": [ + 54, + 257, + 555, + 357 + ], + "type": "inline_equation", + "content": "X_s" + }, + { + "bbox": [ + 54, + 257, + 555, + 357 + ], + "type": "text", + "content": " are processed separately through ITAFA to generate the Garment Feature " + }, + { + "bbox": [ + 54, + 257, + 555, + 357 + ], + "type": "inline_equation", + "content": "F_g" + }, + { + "bbox": [ + 54, + 257, + 555, + 357 + ], + "type": "text", + "content": " and Salient Feature " + }, + { + "bbox": [ + 54, + 257, + 555, + 357 + ], + "type": "inline_equation", + "content": "F_s" + }, + { + "bbox": [ + 54, + 257, + 555, + 357 + ], + "type": "text", + "content": ". Right: The SRE processes the input image " + }, + { + "bbox": [ + 54, + 257, + 555, + 357 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 54, + 257, + 555, + 357 + ], + "type": "text", + "content": " by computing an entropy map " + }, + { + "bbox": [ + 54, + 257, + 555, + 357 + ], + "type": "inline_equation", + "content": "X_e" + }, + { + "bbox": [ + 54, + 257, + 555, + 357 + ], + "type": "text", + "content": ", creating a binary High-Entropy Mask " + }, + { + "bbox": [ + 54, + 257, + 555, + 357 + ], + "type": "inline_equation", + "content": "X_m" + }, + { + "bbox": [ + 54, + 257, + 555, + 357 + ], + "type": "text", + "content": ", and circular region expansion from the entropy centroid to extract the final high-entropy region " + }, + { + "bbox": [ + 54, + 257, + 555, + 357 + ], + "type": "inline_equation", + "content": "X_s" + }, + { + "bbox": [ + 54, + 257, + 555, + 357 + ], + "type": "text", + "content": ", ensuring preservation of detail within a consistent aspect ratio." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 378, + 171, + 389 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 378, + 171, + 389 + ], + "spans": [ + { + "bbox": [ + 55, + 378, + 171, + 389 + ], + "type": "text", + "content": "3.2.2. Feature Aggregation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 392, + 295, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 392, + 295, + 429 + ], + "spans": [ + { + "bbox": [ + 55, + 392, + 295, + 429 + ], + "type": "text", + "content": "The aggregated weights are normalized via softmax to ensure they sum to one across hidden layers, thereby controlling the contribution of each layers." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 127, + 437, + 295, + 451 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 437, + 295, + 451 + ], + "spans": [ + { + "bbox": [ + 127, + 437, + 295, + 451 + ], + "type": "interline_equation", + "content": "W = \\operatorname {s o f t m a x} \\left(W ^ {\\text {c o m}}\\right). \\tag {11}", + "image_path": "8ae40269d36473a93ac0455d57e9c50c380db3d9fcf53c0aecd0aa7b2c18db51.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 460, + 295, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 496 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 496 + ], + "type": "text", + "content": "These weights are then applied to each hidden layer output from the ViT encoder, denoted by " + }, + { + "bbox": [ + 55, + 460, + 295, + 496 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 460, + 295, + 496 + ], + "type": "text", + "content": ", via an element-wise multiplication and summation across layers:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 504, + 295, + 537 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 504, + 295, + 537 + ], + "spans": [ + { + "bbox": [ + 133, + 504, + 295, + 537 + ], + "type": "interline_equation", + "content": "F = \\sum_ {l = 1} ^ {\\mathcal {H}} W [ l ] \\cdot f [ l ], \\tag {12}", + "image_path": "45fcf7e97d4df556776e130854cd15466a24dd96e41d05c16a87ec49d0c91e8b.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 545, + 296, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 581 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 545, + 296, + 581 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 545, + 296, + 581 + ], + "type": "text", + "content": " represents the final aggregated feature tensor, capturing a balanced representation based on both denoising timestep and image complexities." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 581, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 581, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 581, + 295, + 604 + ], + "type": "text", + "content": "In-depth details of ITAFA are explained in Supplementary Materials." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 612, + 226, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 226, + 625 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 226, + 625 + ], + "type": "text", + "content": "3.3. Salient Region Extractor (SRE)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 629, + 296, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 689 + ], + "type": "text", + "content": "Our SRE is a simple yet effective algorithm designed to identify and extract salient, high-entropy, areas within an image. The method includes: Entropy Map Computation, High-Entropy Region Identification, Entropy Centroid Localization, Circular Region Expansion, and Region Extraction." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "By utilizing Shannon Entropy [33] to compute an entropy map " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "X_{e}" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": ", which measures information content across pixel" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 377, + 555, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 377, + 555, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 555, + 520 + ], + "type": "text", + "content": "neighborhoods. High-entropy regions are then identified by thresholding this map to create a binary High-Entropy Mask " + }, + { + "bbox": [ + 313, + 377, + 555, + 520 + ], + "type": "inline_equation", + "content": "X_{m}" + }, + { + "bbox": [ + 313, + 377, + 555, + 520 + ], + "type": "text", + "content": " that highlights areas with substantial visual complexity. The centroid of this region is computed to serve as the initial center for region extraction. From this centroid, a circular region expansion method grows the bounding box outward in small increments, checking for low-entropy areas along each edge as stopping point to ensure meaningful content within the boundary. After the expansion, the region is expanded to follow the aspect ratio of the original image, then extracted as the final output " + }, + { + "bbox": [ + 313, + 377, + 555, + 520 + ], + "type": "inline_equation", + "content": "X_{s}" + }, + { + "bbox": [ + 313, + 377, + 555, + 520 + ], + "type": "text", + "content": ", capturing detailed areas while maintaining minimal distortion." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 521, + 554, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 554, + 546 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 554, + 546 + ], + "type": "text", + "content": "Details of SRE process and algorithm are elaborated in Supplementary Materials." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 559, + 395, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 559, + 395, + 573 + ], + "spans": [ + { + "bbox": [ + 313, + 559, + 395, + 573 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 580, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 580, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 580, + 555, + 689 + ], + "type": "text", + "content": "Implementation details: We use the MDTv2 XL model [5] with RGB reference images of shape " + }, + { + "bbox": [ + 313, + 580, + 555, + 689 + ], + "type": "inline_equation", + "content": "512 \\times 512 \\times 3" + }, + { + "bbox": [ + 313, + 580, + 555, + 689 + ], + "type": "text", + "content": ", encoded into latent " + }, + { + "bbox": [ + 313, + 580, + 555, + 689 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 313, + 580, + 555, + 689 + ], + "type": "text", + "content": "-space via the Stable Diffusion XL VAE [26]. These latent features are patch-embedded and processed within our model. For conditioning of garment features, the ViT encoder extracts features from RGB images of shape " + }, + { + "bbox": [ + 313, + 580, + 555, + 689 + ], + "type": "inline_equation", + "content": "224 \\times 224 \\times 3" + }, + { + "bbox": [ + 313, + 580, + 555, + 689 + ], + "type": "text", + "content": ". Training was conducted on two 80GB A100 GPUs with a batch size of 6. Further details are provided in the Supplementary Materials." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "Datasets and Evaluation Metrics: We train and evaluate our model on two public datasets: VITON-HD [1] and" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28289" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 563, + 232 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 563, + 232 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 563, + 232 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 563, + 232 + ], + "type": "table", + "html": "
MethodInf. Time (s)↓Param. (M)↓LPIPS↓VITON-HDDressCodeUpper
SSIM↑FID(pair/unp.)↓LPIPS↓SSIM↑FID(pair/unp.)↓
LaDI-VTON [21]4.2258650.0980.8816.587/9.2650.0620.92710.447/14.048
DCI-VTON [7]15.4088590.0850.8965.559/8.7500.0680.91810.501/12.490
StableVITON [16]13.3571,0390.0860.8727.039/9.7170.0660.91911.572/13.666
CAT-DM [45]4.9478590.1040.8557.697/10.2630.0580.92210.257/14.039
TPD [41]28.7848590.0970.88510.161/13.0710.0480.9569.500/12.904
IDM-VTON [2]12.2882,9910.0820.8816.000/9.1630.0360.9447.471/11.915
MDT-IVTON (Base)4.5966710.0990.8877.773/9.8380.0610.92110.300/13.998
Base+ITAFA4.5966710.0860.8815.717/8.8790.0340.9515.414/13.111
Base+ITAFA+HR6.2156710.0830.8845.542/8.8000.0340.9515.384/12.528
Base+ITAFA+SRE4.5966710.0830.8855.462/8.6760.0340.9515.412/10.069
", + "image_path": "eb79d031d2bde39e04e0397257f637b4afdf9ef6ef7e7c67968d420d572ae3e1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 239, + 555, + 298 + ], + "lines": [ + { + "bbox": [ + 54, + 239, + 555, + 298 + ], + "spans": [ + { + "bbox": [ + 54, + 239, + 555, + 298 + ], + "type": "text", + "content": "Table 1. Quantitative results of diffusion-based virtual try-on methods on VITON-HD and DressCode upper-body test sets. Inf. Time refers to the inference time, diffusion model only, measured for generating a single image on our environment, and Param. refers to the denoising diffusion model's parameter size. We report both paired and unpaired FID. Following the previous works, feature of last hidden layer of ViT image encoder is used for our MDT-IVTON without ITAFA. HR refers to use of single High-resolution " + }, + { + "bbox": [ + 54, + 239, + 555, + 298 + ], + "type": "inline_equation", + "content": "(448\\times 448\\times 3)" + }, + { + "bbox": [ + 54, + 239, + 555, + 298 + ], + "type": "text", + "content": " garment image to formulate condition vector " + }, + { + "bbox": [ + 54, + 239, + 555, + 298 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 54, + 239, + 555, + 298 + ], + "type": "text", + "content": ". The best performance for each metric is indicated in bold, while the second best is underlined." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 316, + 294, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 316, + 294, + 399 + ], + "spans": [ + { + "bbox": [ + 54, + 316, + 294, + 399 + ], + "type": "text", + "content": "DressCode [20] upper, lower, and dresses. Data augmentation techniques from StableVITON [16] are applied to enhance model robustness. To assess performance, we use standard metrics from previous VTON works: LPIPS [46], SSIM [37], and FID [11]. All evaluations follow the protocol established in LaDI-VTON [21] to ensure fair comparison with previous virtual try-on models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 400, + 295, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 400, + 295, + 426 + ], + "spans": [ + { + "bbox": [ + 55, + 400, + 295, + 426 + ], + "type": "text", + "content": "Further information about the datasets and evaluation metrics is provided in the Supplementary Materials." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 431, + 250, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 431, + 250, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 431, + 250, + 445 + ], + "type": "text", + "content": "4.1. Qualitative and Quantitative Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 449, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 449, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 54, + 449, + 295, + 581 + ], + "type": "text", + "content": "The quantitative Results of Table. 1 compare our proposed ITA-MDT framework to previous diffusion-based methods, as well as show effect of ITAFA, and SRE on MDT-IVTON. The values of the previous works are re-measured by implementation with their checkpoints released. As it can be observed, our method provides comparable, state-of-the-art, results with faster inference speed and smaller parameters. The qualitative results and comparison to previous methods are shown in Fig. 6, and more in Supplementary Materials. As part of the inpainting process, the generated images are blended with agnostic images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 588, + 149, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 149, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 149, + 601 + ], + "type": "text", + "content": "4.2. Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 605, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 605, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 54, + 605, + 295, + 712 + ], + "type": "text", + "content": "The qualitative results of our proposed ITA-MDT framework and effects of each components are available in the Supplementary Materials. MDT-IVTON alone captures the global structure well, including garment shapes, colors, and human pose, but falls short in fidelity of garment details. The integration of ITAFA and SRE addresses this limitation significantly, with SRE shown as effective as a high-resolution image with twice the width compared to SRE integrated condition vector " + }, + { + "bbox": [ + 54, + 605, + 295, + 712 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 54, + 605, + 295, + 712 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 317, + 315, + 430, + 403 + ], + "blocks": [ + { + "bbox": [ + 317, + 315, + 430, + 403 + ], + "lines": [ + { + "bbox": [ + 317, + 315, + 430, + 403 + ], + "spans": [ + { + "bbox": [ + 317, + 315, + 430, + 403 + ], + "type": "image", + "image_path": "00868f013a1d495a31cfa7a11db618e0e3dd5fd19d4fc9ca92f613af4ffa65fc.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 432, + 316, + 552, + 402 + ], + "blocks": [ + { + "bbox": [ + 432, + 316, + 552, + 402 + ], + "lines": [ + { + "bbox": [ + 432, + 316, + 552, + 402 + ], + "spans": [ + { + "bbox": [ + 432, + 316, + 552, + 402 + ], + "type": "image", + "image_path": "c6b5400ceb27fdf9dfe7f00157c034af81f05d48f713ea4120fccbe5e50cac72.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 404, + 427, + 492 + ], + "blocks": [ + { + "bbox": [ + 317, + 404, + 427, + 492 + ], + "lines": [ + { + "bbox": [ + 317, + 404, + 427, + 492 + ], + "spans": [ + { + "bbox": [ + 317, + 404, + 427, + 492 + ], + "type": "image", + "image_path": "6081b1391c5c01fb49c668eb543e7a828b6fc8896d8d8d7dc41b6e7863c1be4a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 501, + 555, + 579 + ], + "lines": [ + { + "bbox": [ + 313, + 501, + 555, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 501, + 555, + 579 + ], + "type": "text", + "content": "Figure 5. Visualization of the ITAFA's average weight distribution across ViT encoder layers of our final model. (a): use of hidden layers as denoising progresses. The dotted arrow refers to denoising direction. (b, c, d): changes in the use of hidden layers as the sparsity, variance, and gradient magnitude of garment image changes. The dotted arrow refers to the increase in these image complexity components." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 429, + 403, + 553, + 493 + ], + "blocks": [ + { + "bbox": [ + 429, + 403, + 553, + 493 + ], + "lines": [ + { + "bbox": [ + 429, + 403, + 553, + 493 + ], + "spans": [ + { + "bbox": [ + 429, + 403, + 553, + 493 + ], + "type": "image", + "image_path": "8cd65d56bb3fa52d9f682e2699dcfaf20426fec0df2b454b3d5cdca43d824c11.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": "Fig. 5 visualizes how ITAFA leverages features from different layers of the ViT image encoder, DINOv2, across the denoising process. In Fig. 5 (a), where it observes the use of layers with respect to timesteps, the middle layers dominate the feature aggregation for much of the diffusion; however, a distinct shift in layer usage occurs from early to later layers as the timestep progresses. In the early timestep when the input is largely noise, the encoder's first layer is prominently utilized, and at the final timestep, the use of the" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28290" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 68, + 555, + 521 + ], + "blocks": [ + { + "bbox": [ + 55, + 68, + 555, + 521 + ], + "lines": [ + { + "bbox": [ + 55, + 68, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 68, + 555, + 521 + ], + "type": "image", + "image_path": "067efeaff30710115ae707c1c8b3e8b953aab6375882236c3bff8468ef12a788.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 526, + 555, + 549 + ], + "lines": [ + { + "bbox": [ + 55, + 526, + 555, + 549 + ], + "spans": [ + { + "bbox": [ + 55, + 526, + 555, + 549 + ], + "type": "text", + "content": "Figure 6. Qualitative comparison of ITA-MDT with prior methods on VITON-HD and DressCode Upper-body datasets. The first three rows show VITON-HD results; the next three show DressCode results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 570, + 297, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 297, + 701 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 297, + 701 + ], + "type": "text", + "content": "last layer markedly increases. This observation supports our hypothesis that dynamically aggregating features from different layers as the diffusion progresses effectively optimizes feature utilization across denoising stages. In Fig. 5 (b, c, d), we observe a clear pattern where specific layers are consistently favored depending on the intensity of particular image complexity (e.g., feature sparsity, variance, or gradient magnitude). This indicates that our model adaptively activates layers containing information most relevant to the type of visual complexity, highlighting the value and adaptability of our ITAFA." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 314, + 569, + 388, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 569, + 388, + 582 + ], + "spans": [ + { + "bbox": [ + 314, + 569, + 388, + 582 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "This paper presented ITA-MDT, a framework for the IVTON task that successfully leverages MDT, along with the proposed ITAFA and SRE modules, to address the challenges of transformer-based denoising diffusion models in preserving fine-grained details efficiently. Analysis confirmed that ITAFA and SRE effectively function as intended, enabling ITA-MDT to achieve state-of-the-art performance in several metrics with reduced model parameters and faster inference speed compared to previous methods." + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "28291" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 224 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 224 + ], + "type": "text", + "content": "This work was supported by Institute for Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No.RS-2021-II211381, Development of Causal AI through Video Understanding and Reinforcement Learning, and Its Applications to Real Environments) and partly supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No.RS-2022-II220184, Development and Study of AI Technologies to Inexpensively Conform to Evolving Policy on Ethics)" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 232, + 115, + 244 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 232, + 115, + 244 + ], + "spans": [ + { + "bbox": [ + 56, + 232, + 115, + 244 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 251, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 251, + 296, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 251, + 296, + 306 + ], + "spans": [ + { + "bbox": [ + 61, + 251, + 296, + 306 + ], + "type": "text", + "content": "[1] Seunghwan Choi, Sunghyun Park, Minsoo Lee, and Jaegul Choo. Viton-hd: High-resolution virtual try-on via misalignment-aware normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14131-14140, 2021. 1, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 307, + 296, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 307, + 296, + 339 + ], + "spans": [ + { + "bbox": [ + 61, + 307, + 296, + 339 + ], + "type": "text", + "content": "[2] Yisol Choi, Sangkyung Kwak, Kyungmin Lee, Hyungwon Choi, and Jinwoo Shin. Improving diffusion models for virtual try-on. arXiv preprint arXiv:2403.05139, 2024. 3, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 340, + 296, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 340, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 62, + 340, + 296, + 384 + ], + "type": "text", + "content": "[3] Lavinia De Divitiis, Federico Becattini, Claudio Baecchi, and Alberto Del Bimbo. Disentangling features for fashion recommendation. ACM Transactions on Multimedia Computing, Communications and Applications, 19(1s):1-21, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 384, + 296, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 384, + 296, + 449 + ], + "spans": [ + { + "bbox": [ + 62, + 384, + 296, + 449 + ], + "type": "text", + "content": "[4] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 449, + 296, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 449, + 296, + 503 + ], + "spans": [ + { + "bbox": [ + 62, + 449, + 296, + 503 + ], + "type": "text", + "content": "[5] Shanghua Gao, Pan Zhou, Ming-Ming Cheng, and Shuicheng Yan. Masked diffusion transformer is a strong image synthesizer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23164-23173, 2023. 2, 3, 4, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 505, + 296, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 505, + 296, + 559 + ], + "spans": [ + { + "bbox": [ + 62, + 505, + 296, + 559 + ], + "type": "text", + "content": "[6] Yuying Ge, Yibing Song, Ruimao Zhang, Chongjian Ge, Wei Liu, and Ping Luo. Parser-free virtual try-on via distilling appearance flows. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8485-8493, 2021. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 559, + 296, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 559, + 296, + 613 + ], + "spans": [ + { + "bbox": [ + 62, + 559, + 296, + 613 + ], + "type": "text", + "content": "[7] Junhong Gou, Siyu Sun, Jianfu Zhang, Jianlou Si, Chen Qian, and Liqing Zhang. Taming the power of diffusion models for high-quality virtual try-on with appearance flow. In Proceedings of the 31st ACM International Conference on Multimedia, pages 7599-7607, 2023. 3, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 615, + 296, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 615, + 296, + 658 + ], + "spans": [ + { + "bbox": [ + 62, + 615, + 296, + 658 + ], + "type": "text", + "content": "[8] Ríza Alp Güler, Natalia Neverova, and Iasonas Kokkinos. Denseseposé: Dense human pose estimation in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7297-7306, 2018. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 658, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 658, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 658, + 296, + 712 + ], + "type": "text", + "content": "[9] M Hadi Kiapour, Xufeng Han, Svetlana Lazebnik, Alexander C Berg, and Tamara L Berg. Where to buy it: Matching street clothing photos in online shops. In Proceedings of the IEEE international conference on computer vision, pages 3343-3351, 2015. 1" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 117 + ], + "type": "text", + "content": "[10] Xintong Han, Zuxuan Wu, Zhe Wu, Ruichi Yu, and Larry S Davis. Viton: An image-based virtual try-on network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7543-7552, 2018. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 118, + 555, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 555, + 171 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 555, + 171 + ], + "type": "text", + "content": "[11] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 175, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 196 + ], + "type": "text", + "content": "[12] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 198, + 555, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 555, + 230 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 555, + 230 + ], + "type": "text", + "content": "[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 232, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 232, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 317, + 232, + 553, + 274 + ], + "type": "text", + "content": "[14] Wei-Lin Hsiao and Kristen Grauman. Creating capsule wardrobes from fashion images. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7161-7170, 2018. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 277, + 555, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 277, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 277, + 555, + 331 + ], + "type": "text", + "content": "[15] Thibaut Issenhuth, Jérémie Mary, and Clément Calauzenes. Do not mask what you do not need to mask: a parser-free virtual try-on. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XX 16, pages 619–635. Springer, 2020. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 333, + 553, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 333, + 553, + 388 + ], + "spans": [ + { + "bbox": [ + 316, + 333, + 553, + 388 + ], + "type": "text", + "content": "[16] Jeongho Kim, Guojung Gu, Minho Park, Sunghyun Park, and Jaegul Choo. Stableviton: Learning semantic correspondence with latent diffusion model for virtual try-on. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8176-8185, 2024. 3, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 389, + 555, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 389, + 555, + 433 + ], + "spans": [ + { + "bbox": [ + 316, + 389, + 555, + 433 + ], + "type": "text", + "content": "[17] Gwanhyeong Koo, Sunjae Yoon, Ji Woo Hong, and Chang D Yoo. Flexied: Frequency-aware latent refinement for enhanced non-rigid editing. In European Conference on Computer Vision, pages 363-379. Springer, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 434, + 555, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 434, + 555, + 487 + ], + "spans": [ + { + "bbox": [ + 317, + 434, + 555, + 487 + ], + "type": "text", + "content": "[18] Sangyun Lee, Gyojung Gu, Sunghyun Park, Seunghwan Choi, and Jaegul Choo. High-resolution virtual try-on with misalignment and occlusion-handled conditions. In European Conference on Computer Vision, pages 204-219. Springer, 2022. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 491, + 555, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 491, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 317, + 491, + 555, + 544 + ], + "type": "text", + "content": "[19] Shugao Ma, Tomas Simon, Jason Saragih, Dawei Wang, Yuecheng Li, Fernando De La Torre, and Yaser Sheikh. Pixel codec avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 64-73, 2021. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 546, + 555, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 546, + 555, + 601 + ], + "spans": [ + { + "bbox": [ + 317, + 546, + 555, + 601 + ], + "type": "text", + "content": "[20] Davide Morelli, Matteo Fincato, Marcella Cornia, Federico Landi, Fabio Cesari, and Rita Cucchiara. Dress code: high-resolution multi-category virtual try-on. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2231-2235, 2022. 1, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 602, + 555, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 602, + 555, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 602, + 555, + 657 + ], + "type": "text", + "content": "[21] Davide Morelli, Alberto Baldrati, Giuseppe Cartella, Marcelia Cornia, Marco Bertini, and Rita Cucchiara. Ladi-vton: latent diffusion textual-inversion enhanced virtual try-on. In Proceedings of the 31st ACM International Conference on Multimedia, pages 8580-8589, 2023. 3, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 658, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 555, + 712 + ], + "type": "text", + "content": "[22] Shuliang Ning, Duomin Wang, Yipeng Qin, Zirong Jin, Baoyuan Wang, and Xiaoguang Han. Picture: Photorealistic virtual try-on from unconstrained designs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6976-6985, 2024. 3" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28292" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "text", + "content": "[23] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 296, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 296, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 296, + 173 + ], + "type": "text", + "content": "[24] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 2, 3, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 296, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 296, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 296, + 207 + ], + "type": "text", + "content": "[25] Trung X Pham, Zhang Kang, and Chang D Yoo. Cross-view masked diffusion transformers for person image synthesis. arXiv preprint arXiv:2402.01516, 2024. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 296, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 296, + 262 + ], + "type": "text", + "content": "[26] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 6, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 296, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 296, + 319 + ], + "type": "text", + "content": "[27] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM international conference on multimedia, pages 484-492, 2020. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 320, + 296, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 296, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 296, + 374 + ], + "type": "text", + "content": "[28] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International conference on machine learning, pages 8821-8831. Pmlr, 2021. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 376, + 296, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 296, + 432 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 296, + 432 + ], + "type": "text", + "content": "[29] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 434, + 296, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 296, + 498 + ], + "type": "text", + "content": "[30] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 500, + 296, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 500, + 296, + 565 + ], + "spans": [ + { + "bbox": [ + 56, + 500, + 296, + 565 + ], + "type": "text", + "content": "[31] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 567, + 296, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 296, + 632 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 296, + 632 + ], + "type": "text", + "content": "[32] Rohan Sarkar, Navaneeth Bodla, Mariya I Vasileva, Yen-Liang Lin, Anurag Beniwal, Alan Lu, and Gerard Medioni. Outfittransformer: Learning outfit representations for fashion recommendation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3601-3609, 2023. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 634, + 296, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 634, + 296, + 666 + ], + "spans": [ + { + "bbox": [ + 56, + 634, + 296, + 666 + ], + "type": "text", + "content": "[33] Claude Elwood Shannon. A mathematical theory of communication. The Bell system technical journal, 27(3):379-423, 1948. 6, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 668, + 296, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 668, + 296, + 700 + ], + "spans": [ + { + "bbox": [ + 56, + 668, + 296, + 700 + ], + "type": "text", + "content": "[34] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 702, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 702, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 702, + 296, + 714 + ], + "type": "text", + "content": "[35] Yang Song and Stefano Ermon. Generative modeling by" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "text", + "content": "estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 97, + 555, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 555, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 555, + 150 + ], + "type": "text", + "content": "[36] Bochao Wang, Huabin Zheng, Xiaodan Liang, Yimin Chen, Liang Lin, and Meng Yang. Toward characteristic-preserving image-based virtual try-on network. In Proceedings of the European conference on computer vision (ECCV), pages 589-604, 2018. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 152, + 555, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 555, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 555, + 195 + ], + "type": "text", + "content": "[37] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 198, + 555, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 555, + 262 + ], + "type": "text", + "content": "[38] Zhenyu Xie, Zaiyu Huang, Xin Dong, Fuwei Zhao, Haoye Dong, Xijin Zhang, Feida Zhu, and Xiaodan Liang. Gp-vton: Towards general purpose virtual try-on via collaborative localflow global-parsing learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23550-23559, 2023. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 264, + 555, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 555, + 319 + ], + "type": "text", + "content": "[39] Sicheng Xu, Guojun Chen, Yu-Xiao Guo, Jiaolong Yang, Chong Li, Zhenyu Zang, Yizhong Zhang, Xin Tong, and Baining Guo. Vasa-1: Lifelike audio-driven talking faces generated in real time. arXiv preprint arXiv:2404.10667, 2024. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 320, + 555, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 320, + 555, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 320, + 555, + 376 + ], + "type": "text", + "content": "[40] Han Yang, Ruimao Zhang, Xiaobao Guo, Wei Liu, Wangmeng Zuo, and Ping Luo. Towards photo-realistic virtual try-on by adaptively generating-preserving image content. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7850-7859, 2020. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 376, + 555, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 555, + 432 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 555, + 432 + ], + "type": "text", + "content": "[41] Xu Yang, Changxing Ding, Zhibin Hong, Junhao Huang, Jin Tao, and Xiangmin Xu. Texture-preserving diffusion models for high-fidelity virtual try-on. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7017-7026, 2024. 3, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 434, + 554, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 554, + 477 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 554, + 477 + ], + "type": "text", + "content": "[42] Sunjae Yoon, Gwanhyeong Koo, Ji Woo Hong, and Chang D Yoo. Dni: Dilutional noise initialization for diffusion video editing. In European Conference on Computer Vision, pages 180-195. Springer, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 478, + 554, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 554, + 511 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 554, + 511 + ], + "type": "text", + "content": "[43] Sunjae Yoon, Gwanhyeong Koo, Geonwoo Kim, and Chang D Yoo. Frag: Frequency adapting group for diffusion video editing. arXiv preprint arXiv:2406.06044, 2024." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 512, + 554, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 512, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 512, + 554, + 555 + ], + "type": "text", + "content": "[44] Sunjae Yoon, Gwanhyeong Koo, Younghwan Lee, and Chang Yoo. Tpc: Test-time procrustes calibration for diffusion-based human image animation. Advances in Neural Information Processing Systems, 37:118654-118677, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 557, + 555, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 557, + 555, + 612 + ], + "spans": [ + { + "bbox": [ + 316, + 557, + 555, + 612 + ], + "type": "text", + "content": "[45] Jianhao Zeng, Dan Song, Weizhi Nie, Hongshuo Tian, Tongtong Wang, and An-An Liu. Cat-dm: Controllable accelerated virtual try-on with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8372-8382, 2024. 3, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 613, + 555, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 555, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 555, + 667 + ], + "type": "text", + "content": "[46] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 669, + 555, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 555, + 714 + ], + "type": "text", + "content": "[47] Luyang Zhu, Dawei Yang, Tyler Zhu, Fitsum Reda, William Chan, Chitwan Sahara, Mohammad Norouzi, and Ira Kemelmacher-Shlizerman. Tryondiffusion: A tale of two unets. In Proceedings of the IEEE/CVF Conference on Com" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28293" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [], + "discarded_blocks": [ + { + "bbox": [ + 75, + 72, + 296, + 94 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 72, + 296, + 94 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 296, + 94 + ], + "type": "text", + "content": "puter Vision and Pattern Recognition, pages 4606-4615, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "28294" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/IceDiff_ High Resolution and High-Quality Arctic Sea Ice Forecasting with Generative Diffusion Prior/3f2f5e81-8b45-4fda-b37e-2ea2e3413d80_content_list.json b/2025/IceDiff_ High Resolution and High-Quality Arctic Sea Ice Forecasting with Generative Diffusion Prior/3f2f5e81-8b45-4fda-b37e-2ea2e3413d80_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..20d930a30789e495bd0ba633d8d6f4a4eeba63cb --- /dev/null +++ b/2025/IceDiff_ High Resolution and High-Quality Arctic Sea Ice Forecasting with Generative Diffusion Prior/3f2f5e81-8b45-4fda-b37e-2ea2e3413d80_content_list.json @@ -0,0 +1,1269 @@ +[ + { + "type": "text", + "text": "IceDiff: High Resolution and High-Quality Arctic Sea Ice Forecasting with Generative Diffusion Prior", + "text_level": 1, + "bbox": [ + 122, + 128, + 875, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jingyi Xu $^{1,\\ast}$ , Siwei Tu $^{1,\\ast}$ , Weidong Yang $^{1,\\dagger}$ , Ben Fei $^{3,\\dagger}$ , Shuhao Li $^{1}$ , Keyi Liu $^{1}$ , Yeqi Luo $^{1}$ , Lipeng Ma $^{1}$ , Lei Bai $^{2}$ $^{1}$ Fudan University, $^{2}$ Shanghai Artificial Intelligence Laboratory, $^{3}$ Chinese University of Hong Kong jyxu22@m.fudan.edu.cn, wdyang@fudan.edu.cn, benfei@cuhk.edu.hk, baisanshi@gmail.com * Equal Contributions, † Corresponding Authors", + "bbox": [ + 106, + 202, + 890, + 310 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 344, + 328, + 359 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Variation of Arctic sea ice has significant impacts on polar ecosystems, transporting routes, coastal communities, and global climate. Tracing the change of sea ice at a finer scale is paramount for both operational applications and scientific studies. Recent pan-Arctic sea ice forecasting methods that leverage advances in artificial intelligence have made promising progress over numerical models. However, forecasting sea ice at higher resolutions is still under-explored. To bridge the gap, we propose a two-module cooperative deep learning framework, IceDiff, to forecast sea ice concentration at finer scales. IceDiff first leverages a vision transformer to generate coarse yet superior forecasting results over previous methods at a regular $25\\mathrm{km}$ grid. This high-quality sea ice forecasting can be utilized as reliable guidance for the next module. Subsequently, an unconditional diffusion model pre-trained on low-resolution sea ice concentration maps is utilized for sampling down-scaled sea ice forecasting via a zero-shot guided sampling strategy and a patch-based method. For the first time, IceDiff demonstrates sea ice forecasting with a $6.25\\mathrm{km}$ resolution. IceDiff extends the boundary of existing sea ice forecasting models and more importantly, its capability to generate high-resolution sea ice concentration data is vital for pragmatic usages and research. Code is available at https://github.com/EtronTech/IceDiff.", + "bbox": [ + 88, + 376, + 485, + 755 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 782, + 223, + 799 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Arctic sea ice is essential for global climate change and local communities. Although the exact teleconnections between mid-latitude weather and the variation of Arctic sea ice are yet to be revealed, the reduction of sea ice during the last few decades clearly has impacts that transcend the Arctic region [3, 41]. Hence, establishing skillful forecasting", + "bbox": [ + 89, + 809, + 483, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "models at finer scales is paramount for both transportation in polar regions and scientific research in geoscience.", + "bbox": [ + 511, + 345, + 906, + 377 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing physics-based numerical models have demonstrated the ability to accurately forecast sea ice concentration (SIC, value ranges from $0\\%$ to $100\\%$ ) for short lead times of a few weeks. However, when forecasting at longer scales, their performance significantly decreases, and their forecasting skills are even worse than those of simple statistical models [3]. With the advent of deep learning methods, a series of works based on deep learning approaches have been committed to SIC forecasting with promising results. Notable studies cover forecasting leads range from daily to seasonal scale, e.g., IceNet [3] provides accurate sea ice extent (SIE, areas where the SIC value is greater than $15\\%$ and could be covered with sea ice) forecasts for 6 months lead time. SICNet [35] and $\\mathrm{SICNet}_{90}$ [34] are skillful for forecasting SIC in subsequent 7 and 90 days, respectively.", + "bbox": [ + 511, + 381, + 908, + 608 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite the improved forecasting skills of deep learning models over previous approaches, these models are still trained on sea ice products provided by the National Sea Ice Data Center (NSIDC) or the European Organisation for the Exploitation of Meteorological Satellites (EU-METSAT). These sea ice data products are primarily derived from passive microwave sensors with a resolution of $25\\mathrm{km}$ , which is considered too coarse for sea ice modeling. The in situ observations are costly to acquire and extremely spare to provide sufficient spatial coverage over pan-Arctic region. This limitation hinders the representation of detailed information on sea ice concentration and the accuracy of ice edge [10, 16]. The same challenge is also presented in weather forecasts. Therefore, several endeavors are proposed to work on downscaling atmospheric states [5, 11, 43]. For instance, SwinRDM [11] combines an improved SwinRNN [21] model with a denoising diffusion probabilistic model (DDPM) [20] to achieve superresolution (SR) in forecasting climate variables. However,", + "bbox": [ + 511, + 613, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "10567", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "forecasting sea ice at finer spatial scales, which is crucial, especially for operational applications is still under-explored.", + "bbox": [ + 93, + 90, + 480, + 135 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Due to the absence of high-resolution sea ice data products and sparse in situ observations, it is not feasible to train a forecast model over pan-Arctic region at a finer scale. Two primary solutions can address this challenge: (i) Downscale sea ice maps initially and then train a forecast model using these high-resolution SIC maps. (ii) Train a forecast model at the original resolution based on the sea ice products, then downscale the forecast results to a finer scale. In the first solution, downscaling sea ice maps may introduce slight errors, which could accumulate during the forecasting stage. Additionally, training a forecast model on high-resolution SIC maps also imposes a greater computational burden. Therefore, we opt for the latter solution and introduce IceDiff, a two-module cooperative framework. IceDiff aims to provide skillful forecasts at three different temporal scales in the forecasting module (FM) and produce spatially downscaled high-quality SIC maps in the guided diffusion module (GDM). Specifically, in the first module, IceDiff leverages U-Net-based vision transformer and provides forecasts that are superior to existing SIC forecasting models. Then, in the second module, we utilize a well-trained DDPM on the original resolution from the NSIDC production as effective prior and downscale forecasted SIC maps guided by forecasts from the previous module at the original scale. Our DDPM is pre-trained on daily SIC maps with a resolution of $25\\mathrm{km}$ , and it can also generalize well to weekly and monthly SIC maps. In every step $t$ of the sampling process of DDPM, the intermediate variable $x_0$ is estimated from the denoised SIC map $x_{t}$ . Then we propose a convolution kernel with estimizable parameters to simulate the up-scale degradation for the intermediate variable $x_0$ , which results in up-scaled $\\tilde{x}_0$ . A distance function between the low-resolution SIC map and up-scaled $\\tilde{x}_0$ is utilized to dynamically update the parameters of the convolution kernel in the reverse steps and guide the sampling of SIC map in the next time step. Furthermore, to make our IceDiff suitable for SIC maps with various resolutions, a patch-based method is also designed to downscale SIC maps at artistry resolution. We show in this work that using low-resolution SIC maps as guidance can better capture small-scale details in the generated high-resolution SIC maps. Our contributions are three folds:", + "bbox": [ + 93, + 140, + 480, + 773 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We propose a novel SIC forecasting framework IceDiff that demonstrates skills in three different temporal scales which cover subseasonal to seasonal forecasts, specifically, 7 days, weekly average of 8 weeks and monthly average of 6 months, that are commonly investigated by the community. IceDiff achieves state-of-the-art results in all three scales compared to existing deep learning models.", + "bbox": [ + 94, + 779, + 480, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- IceDiff is devised with an estimizable convolutional kernel to simulate the up-scaling degradation and designs a patch-based method for downscaling SIC maps at arbitrary resolution with a single and fixed-resolution unconditional diffusion model. By utilizing the generative diffusion prior inherent in this diffusion model, the high-resolution SIC maps can be obtained through zero-shot sampling. Moreover, our IceDiff is capable of capturing small-scale details and generating high-resolution SIC maps.", + "bbox": [ + 517, + 90, + 903, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Extensive experiments show that our IceDiff could provide accurate SIC forecasts and generate high-quality sea ice samples with finer details. IceDiff not only provides a new and comprehensive baseline for data-driven sea ice research but also extends the capability of existing SIC forecasting models and pushes this area of research towards more operational usage.", + "bbox": [ + 516, + 242, + 903, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 516, + 366, + 661, + 382 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Deep Learning-based Sea Ice Forecasting", + "text_level": 1, + "bbox": [ + 516, + 393, + 864, + 409 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Deep learning models have drawn the attention of sea ice research communities and have been widely investigated for Arctic sea ice forecasting [1, 2, 22, 33]. As a pioneering work, ConvLSTM [24] predicts SIC within the region of Barents Sea and has showcased comparable results to the state-of-the-art numerical climate models. IceNet [3] employs the U-Net architecture for merging pre-selected ERA5 variables and SIC data to improve the accuracy of seasonal SIE prediction up to six months. The experimental results have shown that the forecasting skill of IceNet in pan-Arctic region is superior to the numerical sea ice simulation model SEAS5. MT-IceNet [1] adopts a similar approach to IceNet and adds bi-monthly data to further improve forecasting SIC at seasonal scale. Mu et al. [30] propose to merge atmospheric and oceanographic data in a weighted manner through a variable selection network and provide nine months of forecasting leads. Considering the complex ice-ocean-atmosphere interactions have already embedded in the spatiotemporal correlations of consecutive SIC maps [33, 35], Ren et al. [35] proposed SICNet that solely rely on SIC data to predict concentration for the next seven days. Similar to IceNet, SICNet is built upon U-Net architecture and added with channel-wise dependency via temporal-spatial attention module (TSAM), which is essentially a revised version of the convolutional block attention module. By fusing the output of two SICNet, i.e. one for modeling SIC and the other for extracting climate information, and adjusting the number of output channels, Ren et al. [34] extends the lead time to 90 days. Besides CNN and LSTM-based models, recent works leverage the latest advances in artificial intelligence to further improve forecasting performance. For instance, Ice", + "bbox": [ + 514, + 417, + 903, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10568", + "bbox": [ + 480, + 946, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Former [42] jointly models SIC and sea ice thickness (SIT) based on Transformer backbone. It first decomposes high dimensional spatio-temporal data into sequences via multivariate empirical orthogonal functions [37, 38]. Then an encoder-decoder Transformer is utilized to predict SIC and SIT for up to 45 days. Despite the performance gain of the aforementioned deep learning models, none of them have addressed forecasting sea ice at scales lower than $25\\mathrm{km}$ .", + "bbox": [ + 89, + 90, + 480, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Super Resolution", + "text_level": 1, + "bbox": [ + 89, + 226, + 261, + 242 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Super-resolution aims to reconstruct high-resolution (HR) images from low-resolution (LR) counterparts. Several works [27, 28] explore the use of Variational Autoencoders for image SR. These methods usually improve the restoration quality by learning the conditional distribution of high-resolution images induced by low-resolution images. Although these methods are able to achieve realistic perceptual quality, VAE-based methods are often not as satisfactory due to their limited generation capabilities. To this end, other deep learning methods, such as Convolutional neural networks (CNNs), have been widely used for superresolution tasks in various domains. CNN-based methods typically use HR images corresponding to LR images to supervise training with reconstruction loss [4, 7, 36, 44]. However, these network trained solely on reconstruction loss makes it difficult to capture detailed information in images and generate diverse image results. Therefore, Generative Adversarial Networks (GANs) are proposed to solve these challenges. Stengel et al. [39] utilized GAN in a two-step SR process to achieve downscaling of climatological wind and solar irradiance. Leinonen et al. [23] introduced a recurrent, stochastic SR GAN for down-scaling time-evolving atmospheric fields, demonstrating the potential of GANs in improving the spatial resolution of LR images in atmospheric sciences applications. These models encourage the generator to generate high-quality images that the discriminator cannot distinguish from real images. Although GAN-based methods obtain satisfactory downscaled results, the diversity could be enhanced.", + "bbox": [ + 91, + 250, + 483, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Currently, diffusion model-based methods have been more widely used due to their diversity in generating images and the ability to generate high-quality images comparable to GANs [17, 40]. For instance, Chen et al. [11] utilize SR based on the conditional model to recover the high spatial resolution and finer-scale atmospheric details, unfolding the ability of diffusion models to generate high-quality and detailed images in SR tasks. The objective of this study is to forecast SIC maps at a high resolution of $6.25\\mathrm{km}$ . To achieve this, we have improved the SwinTransformer for predicting SIC at the standard resolution comparable to current methods. Additionally, we employ a diffusion-based SR model to generate high-resolution and high-quality results via zero-shot sampling.", + "bbox": [ + 89, + 688, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 513, + 89, + 648, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our IceDiff framework comprises two modules as depicted in Figure 1. The forecasting module provides skillful forecasts at three different temporal scales. The guided diffusion module produces spatially down-scaled high-quality SIC maps with a pre-trained DDPM via zero-shot sampling and the patch-based method. We elaborate on design details in the following sections.", + "bbox": [ + 511, + 114, + 905, + 220 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Forecasting Module", + "text_level": 1, + "bbox": [ + 511, + 229, + 702, + 244 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To achieve more accurate SIC forecasting at a finer resolution, a skillful forecasting module trained on original sea ice data products is primarily encouraged. In contrast to existing SIC forecasting methods that primarily concentrate on a single temporal scale [1, 3, 42], our objective is to enhance forecasting skill across various temporal scales and increase versatility. Encouraged by the success of Swin Transformer [25] in various data-driven weather forecasting methods [6, 11, 12], we propose to construct forecasting module by utilizing the U-Net architecture [8] and building on top of the Swin Transformer V2 backbone [26] with ResNet $_{TSAM}$ block [35] for better capturing both spatial and temporal features of SIC maps. As illustrated in Figure 1, the FM takes $N$ consecutive SIC maps, each of which has a shape of [448, 304], to predict subsequent maps with identical temporal length as the input at $25\\mathrm{km}$ spatial scale. It consists of encoder, bottleneck and decoder. The patch embedding layer first applies a $2\\times 2$ windows size to partition the original SIC input into patches and maps them into tokens via 2D convolution. Then, the temporal and spatial dependencies are modeled through the ResNet $_{TSAM}$ block. Both the encoder and the decoder have 3 layers of Swin Transformers that comprise a pre-configured even number of Swin Transformer Blocks for feature extraction and restoration. The calculation of two consecutive blocks can be described as follows:", + "bbox": [ + 511, + 252, + 906, + 643 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} z _ {s} ^ {b} = L N (W M S A (z ^ {b - 1})) + z ^ {b - 1}, \\\\ z ^ {b} = L N (M L P (z _ {s} ^ {b})) + z _ {s} ^ {b}, \\\\ z _ {s} ^ {b + 1} = L N (S W M S A (z ^ {b})) + z ^ {b}, \\\\ z ^ {b + 1} = L N \\left(M L P \\left(z _ {s} ^ {b + 1}\\right)\\right) + z _ {s} ^ {b}, \\tag {1} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 654, + 903, + 737 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $z_{s}^{b}$ and $z^{b}$ represents the output feature of the (Shifted) Window-Multi-head Self Attention module and the MLP module for block $b$ , respectively [26]. LN denotes the LayerNorm operation. The combination of such two consecutive blocks, i.e. the former for the employment of WMSA and the latter for SWMSA, ensures that local patterns of adjacent grids are sufficiently attended by the backbone network. We further adopt a shift window size of $28 \\times 19$ that is consistent with the ratio of SIC input map for facilitating the SWMHA mechanism. The output of the", + "bbox": [ + 511, + 748, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10569", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/63d1bfeb0826992771755d58394090e093f17358cb73eb1f8d63684a80886259.jpg", + "image_caption": [ + "Figure 1. Overview of IceDiff, a two-module cooperative framework to forecast sea ice concentration at a finer scale. IceDiff first employs the forecasting module (FM) that takes history SIC data as input to generate subsequent predictions at $25\\mathrm{km}$ resolution. $N$ denotes the number of sequential time stamps, for instance, set $N$ to 7 for seven days of forecasting leads, FM takes a tensor with the shape of $\\{7,448,304\\}$ as input and forecasts the SIC for the next 7 days. $C$ represents the quantity of embedded channels within each layer of FM, the original SIC map data has a channel of 1 rather than 3 for regular images. The guided diffusion module (GDM) then leverages accurate forecasts as guidance to generate downscaled samples at $6.25\\mathrm{km}$ resolution with high quality, see details in Figure 2." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 908, + 369 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "last block within each layer is delivered to a ResNet $_{TSAM}$ block for establishing channel-wise dependencies. Before the subsequent encoder and decoder layers, patch merging and patch expanding [8] are respectively performed to reshape the feature and fit it to the desired size. Following prior works [3, 8, 35], the skip connection is added between the corresponding encoder and decoder layers to improve the performance. The final prediction is generated via a patch expanding layer with ResNet $_{TSAM}$ that up-samples the decoded feature to the size of the original SIC input. Our meticulous designs enable the forecasting module in IceDiff to be applied across various temporal scales, enhancing forecasting skills, which serves as a more effective initiation for the guided diffusion module of IceDiff.", + "bbox": [ + 88, + 470, + 485, + 683 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Guided Diffusion Module", + "text_level": 1, + "bbox": [ + 89, + 696, + 323, + 709 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Off-the-shelf methods are limited to the forecasting of SIC maps at a scale of $25\\mathrm{km}$ from sea ice data products [1, 3, 35]. However, finer-scale SIC maps remain unexplored which is of significant value to ecosystems, transporting routes, coastal communities, and global climate. Therefore, after obtaining skillful forecasts, we exploit a downscaling module based on an unconditional diffusion model pre-trained on daily SIC maps to generate realistic high-resolution SIC maps with rich details. This guided diffusion module pre-trained on daily SIC maps could generalize well to SIC maps at various scales, such as weekly or monthly SIC maps. Moreover, our guided diffusion module", + "bbox": [ + 89, + 719, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "of IceDiff downscales the SIC forecasts from a forecasting module to a scale of $6.25\\mathrm{km}$ , further expanding the accuracy and practicality of sea ice prediction.", + "bbox": [ + 511, + 470, + 906, + 516 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, the reverse process of the diffusion model is conditioned on the low-resolution sea ice map $y$ , which transforms distribution $p_{\\theta}(x_{t-1}|x_t)$ into conditional distribution $p_{\\theta}(x_{t-1}|x_t,y)$ . Previous work [15] has derived the conditional transformation formula in the reverse process:", + "bbox": [ + 511, + 515, + 906, + 585 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\log p _ {\\theta} \\left(x _ {t} \\mid x _ {t + 1}, y\\right) = \\log \\left(p _ {\\theta} \\left(x _ {t} \\mid x _ {t + 1}\\right) p \\left(y \\mid x _ {t}\\right)\\right) + N _ {1} (2) \\\\ \\approx \\log p _ {\\theta} (z) + N _ {2}, z \\sim \\mathcal {N} (z; \\mu_ {\\theta} \\left(x _ {t}, t\\right) + \\Sigma g, \\Sigma I), (3) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 595, + 903, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $g$ refers to $\\nabla_{x_t}\\log p(y|x_t)$ . $N_{1} = \\log p_{\\theta}(y|x_{t + 1})$ , $N_{2}$ is a constant related to $g$ (proof in Appendix). The variance of the reverse process $\\Sigma = \\Sigma_{\\theta}(x_{t})$ is constant.", + "bbox": [ + 511, + 638, + 905, + 685 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Therefore, the sampling process of $p_{\\theta}(x_{t - 1}|x_t,y)$ integrates the gradient term $g$ with the diffusion model $(\\mu ,\\Sigma) = (\\mu_{\\theta}(x_{t}),\\Sigma_{\\theta}(x_{t}))$ , which plays a role in controlling the direction of map generation. We use a heuristic algorithm (proof in Appendix) to approximate the value of $g$ :", + "bbox": [ + 511, + 685, + 905, + 755 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\log p (y \\mid x _ {t}) = - \\log N - s \\mathcal {L} (\\mathcal {D} (\\tilde {x} _ {0}), y) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 765, + 906, + 781 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ng = \\nabla_ {x _ {t}} \\log p (y | x _ {t}) = - s \\nabla_ {x _ {t}} \\mathcal {L} (\\mathcal {D} (x _ {t}), y). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 784, + 906, + 799 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Among them, $s$ is the scaling factor used to control the degree of guidance. And $\\mathcal{L}$ is a distance function used to measure the distance deviation between two SIC maps. $\\mathcal{D}$ is an up-scale convolution function with estimizable parameters, where the parameters are randomly initialized and dynamically optimized with the gradient of distance function at", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "10570", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fd5f3d5c841bca26f4e9ae57dd1c219d066d56fc92b28951abff3474d14c4407.jpg", + "image_caption": [ + "(a)", + "(b)", + "(c)" + ], + "image_footnote": [], + "bbox": [ + 125, + 116, + 869, + 459 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f8f320287e357e63b912bd33dc4cbccc8d5dacdbf994dec8df6e639a57b2e7f9.jpg", + "image_caption": [ + "Figure 2. Overview of guided diffusion module in IceDiff for SIC map downscaling. (a) Pre-trained diffusion model is used to estimate $\\tilde{x}_0$ at reverse step $t$ . The estimable convolutional kernel of step $t$ reduces the resolution of $X_{t}$ so that guidance from low-resolution SIC maps can be incorporated to sample $X_{t-1}$ . (b) Low-resolution SIC map $Y$ is integrated during the sampling process. A distance function describes the distance between a low-resolution SIC map and the generated high-resolution SIC map $x$ after convolution. The gradient of distance function could be used to guide the sampling of $x_{t-1}$ and (c) dynamically update the parameters of the convolutional kernel.", + "Figure 3. Illustration of the patch-based method. We set patch size $p = 256$ and stride $r = 128$ to fit the size of pre-trained models. This setting will divide the example map into $K = 4$ sub-maps boxed with different colors. For each sub-map, an individualizable convolution kernel with adaptive parameters is set to add guidance from the corresponding low-resolution SIC sub-map. At each sampling step $t$ , the average estimated noise on four overlapping blocks will be utilized to update the mean $\\mu$ used to sample $x_{t-1}$ ." + ], + "image_footnote": [], + "bbox": [ + 101, + 554, + 893, + 675 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "every reverse step. $\\mathcal{D}$ connects two different resolution SIC maps, allowing low-resolution SIC maps to provide more effective guidance.", + "bbox": [ + 89, + 751, + 480, + 797 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Algorithm 1, GDM undergoes $T$ reverse steps to gradually restore pure Gaussian noise $x_{T} \\sim \\mathcal{N}(0, I)$ to high-resolution SIC maps. In each reverse step $t$ , the diffusion model calculates the estimated variable $\\tilde{x}_{0}$ which then undergoes the up-scale convolution function $\\mathcal{D}$ at time $t$ to establish a distance metric $\\mathcal{L}$ with the low-", + "bbox": [ + 89, + 809, + 483, + 898 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "resolution SIC map, where $\\mathcal{D}_t^\\varphi$ signifies that its convolutional kernel parameters are $\\varphi = \\varphi_t$ at the reverse steps $t$ . The $\\mathcal{L}$ subscript $\\varphi = \\varphi_t$ indicates that it is a function dependent on the convolutional kernel parameter $\\varphi$ at reverse step $t$ . The gradients of distance function $\\mathcal{L}$ with respect to $\\tilde{x}_0$ and the convolution kernel parameters of the up-scale convolution function are used to update $\\tilde{x}_0$ and the convolution kernel parameters, respectively. The update of convolution kernel parameters ensures the accuracy of the model's", + "bbox": [ + 511, + 751, + 906, + 888 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "10571", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/dd2d2bc103ea3c18c14546641ec03508720e98d2ce31462764c311e1b6c4d9bd.jpg", + "table_caption": [ + "Table 1. Quantitative results of forecasts. IceDiff is evaluated on three different temporal scales. The output length of IceDiff equals a pre-defined forecasting lead as depicted in Figure 1. The validated accuracy of IceDiff could provide reliable guidance for the guided diffusion module to generate high-quality downscaled SIC. BT represents that the SIC data [14] are generated by the Bootstrap retrieval algorithm [13] and the data distribution of CDR is considered to be closer to BT rather than NT [29]." + ], + "table_footnote": [], + "table_body": "
Temporal ScaleLead TimeDatasetMethodsRMSE↓MAE↓R²↑NSE↑IIEE↓SIEdif↓
Subseasonal7 Days (Daily)NTSICNet [35]0.04300.00910.9780.97511750.1127
NTIceDiff0.03240.00770.9870.9868990.0186
CDRSICNet [35]0.04900.01000.9820.9799760.0718
CDRIceDiff0.03960.00800.9890.9878350.0315
45 Days (Daily)BTIceFormer [42]0.06600.02010.960---
8 Weeks Average (Weekly)NTIceDiff0.04030.01080.9750.97211700.1051
CDRIceDiff0.05530.01120.9730.96913530.0871
90 Days (Daily)NTSICNet90 [34]-0.0512----
Seasonal6 Months Average (Monthly)NTIceNet [3]0.18200.09160.567---
NTMT-IceNet [1]0.07770.01970.915---
NTIceDiff0.05730.01880.9580.95220140.2419
CDRIceDiff0.06480.01680.9190.91320160.2657
", + "bbox": [ + 94, + 146, + 903, + 330 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Algorithm 1 Guided diffusion module in IceDiff with the guidance of low-resolution SIC map $y$ . An unconditional diffusion model pre-trained on SIC maps $\\epsilon_{\\theta}(x_t,t)$ is given.", + "bbox": [ + 89, + 342, + 480, + 388 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Input: Low-resolution SIC map $y$ , downscale convolution function $\\mathcal{D}$ composed of optimized convolutional kernels $\\mathcal{K}$ with parameters $\\varphi$ , learning rate $l$ , guidance scale $s$ , distant function $\\mathcal{L}$ .", + "bbox": [ + 91, + 392, + 482, + 446 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Output: Output high-resolution SIC map $x_0$ conditioned on $y$ . Sample $x_T$ from $\\mathcal{N}(0, I)$", + "bbox": [ + 91, + 448, + 480, + 474 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: for all t from T to 1 do", + "2: $\\tilde{x}_0 = \\frac{x_t}{\\sqrt{\\alpha_t}} - \\frac{\\sqrt{1 - \\alpha_t} \\epsilon_0(x_t, t)}{\\sqrt{\\alpha_t}}$", + "3: $\\mathcal{L}_{\\varphi, \\tilde{x}_0} = \\mathcal{L}(y, \\mathcal{D}^\\varphi(\\tilde{x}_0))$", + "4: $\\tilde{x}_0\\gets \\tilde{x}_0 - \\frac{s(1 - \\bar{\\alpha}_t)}{\\sqrt{\\bar{\\alpha}_{t - 1}\\beta_t}}\\nabla_{\\tilde{x}_0}\\mathcal{L}_{\\varphi ,\\tilde{x}_0}$", + "5: $\\tilde{\\mu}_{t} = \\frac{\\sqrt{\\bar{\\alpha}_{t - 1}}\\beta_{t}}{1 - \\bar{\\alpha}_{t}}\\tilde{x}_{0} + \\frac{\\sqrt{\\bar{\\alpha}_{t}}(1 - \\bar{\\alpha}_{t - 1})}{1 - \\bar{\\alpha}_{t}} x_{t}$", + "6: $\\tilde{\\beta}_{t} = \\frac{1 - \\bar{\\alpha}_{t - 1}}{1 - \\bar{\\alpha}_{t}}\\beta_{t}$", + "7: Sample $x_{t-1}$ from $\\mathcal{N}(\\tilde{\\mu}_t, \\tilde{\\beta}_t I)$", + "8: $\\varphi \\gets \\varphi - l \\nabla_{\\varphi} \\mathcal{L}_{\\varphi, \\bar{x}_0}$", + "9: end for", + "10: return $x_0$" + ], + "bbox": [ + 94, + 476, + 352, + 635 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "up-scaling process, making guidance of low-resolution SIC maps more precise and effective.", + "bbox": [ + 89, + 671, + 482, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Patch-based Strategy. Furthermore, pre-trained models can only generate fixed-size SIC maps. To tackle the problem of sampling downscaled SIC forecasts at any resolution, we employ a patch-based strategy. As shown in Figure 3, the patch-based strategy partitions SIC maps into several sub-regions based on a fixed stride. For each individual sub-region, the gradient terms of distance loss between SIC maps upscaled by the convolutional kernel and the corresponding low-resolution SIC sub-region maps are computed individually. Subsequently, the respective mean of Gaussian distribution in sub-regions (Eq. 3) and convolution kernel parameters are updated based on the gradient terms. The entire map is updated by taking the mean of", + "bbox": [ + 88, + 704, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ddc9098d71c0658ec8f78954159b79313823d13bd8efac4ac184dfd99cc707cd.jpg", + "table_caption": [ + "Table 2. Quantitative comparison between IceDiff for SIC downscaling and other methods. Const. and FID represents the Consistency and Fréchet Inception Distance metrics, respectively." + ], + "table_footnote": [], + "table_body": "
Method7 Days8 Week Avg.6 Month Avg.
Const.↓FID↓Const.↓FID↓Const.↓FID↓
Nearest20.0178.4521.9880.1723.0788.36
Bilinear17.8870.7520.0277.2918.3180.73
BiCubic13.1354.7113.4068.8510.0970.18
DGP [32]23.1875.4228.6281.4229.7485.47
GDP [17]8.2640.1510.5452.4010.9655.72
IceDiff6.8429.347.2531.296.6239.79
", + "bbox": [ + 517, + 383, + 903, + 507 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the updated values across all sub-regions, each weighted by a binary patch mask that serves as a quantifier of the regional scope, thereby refining the overall mean and convolution kernel parameters of the sampled high-resolution SIC maps. By leveraging this strategy, our model is capable of downscaling sea ice maps to the desired resolution and the practicality of the model has been further enhanced.", + "bbox": [ + 511, + 518, + 906, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 640, + 645, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Dataset and Baseline Setup", + "text_level": 1, + "bbox": [ + 511, + 666, + 758, + 683 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate our proposed IceDiff framework on the Climate Data Record of Passive Microwave Sea Ice Concentration Version 4 dataset from NSIDC [29] (CDR in Table 1). It records daily SIC data starting from October $25^{th}$ 1978 and provides the coverage of the pan-Arctic region $(\\mathrm{N} : -39.36^{\\circ}, \\mathrm{S} : -89.84^{\\circ}, \\mathrm{E} : 180^{\\circ}, \\mathrm{W} : -180^{\\circ})$ . The SIC map is formed of $448 \\times 304$ pixels and each of which represents an area of $25 \\mathrm{~km} \\times 25 \\mathrm{~km}$ grid and has a value ranging from $0\\%$ to $100\\%$ . We choose data from October $25^{th}$ 1978 to 2013 as the training dataset, the years 2014 and 2015 are selected as validation set, and data from 2016 to 2023 are used for testing. To comprehensively validate and verify the performance of our forecasting module, we utilize additional data [16] that are generated through the retrieval", + "bbox": [ + 509, + 688, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "10572", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0260bf368e92f1519117332c6f087d8bf05cd9404a00beb1af3752b427d0081a.jpg", + "table_caption": [ + "Table 3. Ablation study on TSAM. We compare the forecasting results between IceDiff with or without the TSAM." + ], + "table_footnote": [], + "table_body": "
Lead TimeTSAMRMSE↓MAE↓R²↑NSE↑IIEE↓SIEdif↓
7 DaysX0.04030.00820.98810.98608740.0336
✓0.03960.00800.98850.98658350.0315
8 Week Avg.X0.05350.01160.9750.97113860.1420
✓0.05530.01120.9730.96913530.0871
6 Month Avg.X0.07630.01940.8830.87526100.4919
✓0.06480.01680.9190.91320160.2657
", + "bbox": [ + 94, + 119, + 480, + 204 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "algorithm (NASA Team [9], NT in Table 1). The forecasting module is trained on three different temporal scales, i.e. the input of SIC maps are 7 days, 8 weeks average, and 6 months average, covering forecasting leads from subseasonal to seasonal scale. Corresponding deep learning-based models are selected for comparison, as illustrated in Table 1. Note that we follow the split of datasets in respective baselines to preprocess the additional data. We re-implemented SICNet (the source code is not available) based on the original paper and trained it under an environment identical to our forecasting module. To evaluate the performance of our downscaling, five methods are selected to set up baselines for downscaling quality comparison.", + "bbox": [ + 89, + 215, + 483, + 414 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 422, + 307, + 438 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We independently trained the forecasting module of IceDiff on SIC dataset of three aforementioned temporal scales for 60 epochs. For each temporal scale, the forecasting module outputs the same length of input data as forecasting leads, i.e. 7 days, 8 weeks average, and 6 months averages. The dimensions of embedding features of encoder, decoder, and bottleneck are all set to 32. We utilize the same hyperparameter of TSAM as in [35]. The forecasting module is optimized by AdamW under the mean square error (MSE) objective using Pytorch on one NVIDIA A100 80GB GPU for all experiments with a learning rate of 0.0005. All daily SIC maps are resized and cropped to $256 \\times 256$ to train our GDM of IceDiff. using AdamW with $\\beta_{1} = 0.9$ and $\\beta_{2} = 0.999$ . We train in 16-bit precision using loss-scaling, but maintain 32-bit weights, EMA, and optimizer state. We use an EMA rate of 0.9999. When sampling with 1000 timesteps, we use the same noise schedule as for training.", + "bbox": [ + 89, + 444, + 483, + 703 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Evaluation Metrics", + "text_level": 1, + "bbox": [ + 89, + 712, + 274, + 726 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To quantitatively evaluate our IceDff, we select commonly used root mean square error (RMSE) and mean absolute error (MAE) for the comparison of forecasting accuracy. $R^2$ score and Nash-Sutcliffe Efficiency [31] are utilized to analyze the similarity of spatial pattern between the ground truth and predicted SIC map. We adopt Integrated Ice-Edge Error score [18] and calculate the sea ice area differences $(SIE_{dif})$ , in millions of $km^2$ to further evaluate the prediction of SIE (where the SIC value is greater than 15%).", + "bbox": [ + 89, + 733, + 482, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As to the evaluation of downscaling quality, following [11] we utilize the Fréchet Inception Distance (FID)", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d9c3e5ac0cc768541af00ac8aa402ed29d25e59714e257166e1d2552224dc671.jpg", + "table_caption": [ + "Table 4. Ablation study on the ratio of Input/Forecast number of SIC maps. Hf., Db. and Eq. represent that the forecasting module takes input data which has half, double, and equal length of forecasting leads, respectively. For the Hf. of 7 Days, we set the input length to 3." + ], + "table_footnote": [], + "table_body": "
Lead TimeI/F.RMSE↓MAE↓R²↑NSE↑IIEE↓SIEdif↓
7 DaysHf.0.04160.00840.9870.9859050.0538
Db.0.04010.00840.9880.9868890.0599
Eq.0.03960.00800.9890.9878350.0315
8 Week Avg.Hf.0.05800.01180.9700.96614050.1176
Db.0.05940.01210.9690.96414350.1221
Eq.0.05530.01120.9730.96913530.0871
6 Month Avg.Hf.0.09040.02320.8420.83031920.6289
Db.0.07220.01790.8970.88920270.2795
Eq.0.06480.01680.9190.91320160.2657
", + "bbox": [ + 517, + 160, + 903, + 275 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "[19] for measuring the downscaling fidelity. In addition, we adopt Consistency to quantify the faithfulness between the downscaled map and the low-resolution map.", + "bbox": [ + 511, + 287, + 903, + 334 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The detailed equations for calculating the abovementioned metrics can be found in Appendix.", + "bbox": [ + 511, + 335, + 903, + 364 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Main Results", + "text_level": 1, + "bbox": [ + 511, + 377, + 651, + 391 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this subsection, we evaluate the proposed IceDiff with baseline methods for validation of the effectiveness.", + "bbox": [ + 511, + 400, + 903, + 429 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Set up new baselines for SIC prediction. The forecasting module of IceDiff is trained independently to predict SIC at 7 days, 8 weeks average, and 6 months average. As shown in Table 1, the performance of our proposed forecasting module of IceDiff is superior to SICNet in all evaluation metrics. The forecasting skill of FM across different SIC datasets is consistent in three temporal scales. Specifically, it outperforms SOTA seasonal forecasting baselines by a noticeable margin. Considering the overall performance of our FM, it could provide reliable SIC forecasts and sufficiently support the subsequent downscaling module.", + "bbox": [ + 511, + 430, + 906, + 597 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Analysis of the end of the melting season. To further verify the performance of IceDiff in extreme cases, we select two years of the end of the melting season (early September) when the mean Arctic SIC typically reaches its minimum for validation: (i) In 2016, it recorded historically low annual SIC; (ii) In 2022, the annual SIC has an abrupt increase compared to the previous year (annual trends are plotted in the Appendix). We present the SIC residuals (predicted SIC minus the ground truth) of September $1^{st}$ 2016 and 2022 in Figure 4(a) and Figure 4(b), respectively. The overall residuals are relatively low and the prediction errors mainly appear in part of the sea ice edge regions. The accurate forecasting results in temporal periods with highly varying SIC validate the effectiveness of IceDiff.", + "bbox": [ + 511, + 598, + 905, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "High downscaling quality. We systematically compare the results of IceDiff with other downscaling methods. In terms of interpolation-based methods, we select nearest, bilinear, and bicubic interpolation to compare with our diffusion-based GDM. Besides, we select DGP [32] as a GAN-based method and GDP [17] as a DDPM-based method for fur", + "bbox": [ + 511, + 810, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "10573", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ef9069d224d6630c2edfcec6deb52788010afeaca97aaac2ba5bdfbea255ab88.jpg", + "image_caption": [ + "(a)", + "Forecast Date Sep 1,2016" + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 263, + 281 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9e9a75ac043e9c73524086571625a79fb6b4b907dc4ce8dbc6aa3f8dbf1c3068.jpg", + "image_caption": [ + "(b)", + "Forecast Date Sep 1,2022", + "Figure 4. Qualitative analysis of the end of the melting season. The annual mean Arctic SIC has recorded a historical low in 2016 and it drastically increased in 2022 when compared to the previous annual mean. The residual plots of early September in (a)2016 and (b)2022 demonstrate where the FM overestimates (positive figures) or underestimates (negative figures) SIC." + ], + "image_footnote": [], + "bbox": [ + 263, + 87, + 480, + 282 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ther comparison. Interpolation-based methods upsample the low-resolution sea ice map, which uses existing pixel information to estimate missing pixel values.", + "bbox": [ + 89, + 407, + 483, + 453 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Figure 5, our GDM produces high-quality results and small-scale structures are successfully captured and rich details are generated through the guidance from the input map during each sampling step. By contrast, interpolation-based methods do not integrate details information of the low-resolution map, resulting in blurry edges and contours in the generated results. More quality analysis can be found in the appendix.", + "bbox": [ + 89, + 454, + 483, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To further validate the superiority of our method, we follow SwinRDM [11] and GDP [17] to utilize FID and Consistency to evaluate our method, respectively. As shown in Table 2, compared to interpolation-based methods, our IceDiff improves FID score by 31.11 on average, reflecting it has the capacity to help generate a more realistic high-resolution map with a satisfactory texture restoration effect. A lower Consistency demonstrates the structural difference between the downscaled map and the original SIC input is relatively imperceptible, which better preserves the spatial distribution of sea ice forecast.", + "bbox": [ + 89, + 575, + 483, + 742 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Ablation Studies on Design of FM", + "text_level": 1, + "bbox": [ + 89, + 756, + 387, + 772 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Table 3, by adopting TSAM to FM, the performance has improved for all metrics at 7 days and 6 months average lead times while the effects for 8 weeks average lead time are not all positive. Overall, incorporating TSAM to establish channel-wise dependency is favorable for SIC forecast models. To examine the impacts of input SIC data length, we trained FM for input lengths of half, double, and equal to the forecasting leads. The results in Table 4 show", + "bbox": [ + 89, + 779, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/03beacc4b8a477ee609539113f2f0d683f334a853d24a075ac30feb8c614c4e1.jpg", + "image_caption": [ + "(a) IceDiff", + "Figure 5. Qualitative analysis of downscaled SIC map. We visualize the SIC map by coloring the land and the ocean in yellow and blue, respectively. Since SIC has a range of $[0,1]$ , we map it to $[0,255]$ and replicate it for all 3 RGB channels. The darker the pixels appear, the SIC values are more closer to $0\\%$ . For better identification of the difference, we crop out a region and plot a grid size of $250\\mathrm{km}$ and further display details on $4\\times 4$ grid." + ], + "image_footnote": [], + "bbox": [ + 516, + 87, + 645, + 294 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e4875f857608b86f8a75df81d4c7333796c8907f80f19cb21b882bf8ade52022.jpg", + "image_caption": [ + "(b) Bicubic" + ], + "image_footnote": [], + "bbox": [ + 645, + 88, + 776, + 294 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/263520a0bc76ba2995ade0fcdc342302b2e7f246b2ee2dcb577fa8a485f9c13c.jpg", + "image_caption": [ + "(c) GDP" + ], + "image_footnote": [], + "bbox": [ + 774, + 88, + 905, + 294 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the peak of the performance at the ratio where the input length equals the forecasting leads.", + "bbox": [ + 511, + 417, + 906, + 450 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion and Limitations", + "text_level": 1, + "bbox": [ + 511, + 465, + 769, + 481 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose a novel pan-Arctic high-resolution SIC forecasting framework, IceDiff, which comprises a forecasting module and a guided diffusion module, these two modules closely cooperate with each other to fulfill accurate forecasting of SIC at a finer and more practical spatial scale. The forecasting module of IceDiff is capable of producing accurate forecasts at three different temporal scales and the guided diffusion module leverages reliable forecasts as guidance to generate high-quality down-scaled SIC maps. The forecasting skills of IceDiff at 7 days lead are superior to SOTA deep learning-based models. IceDiff also sets up a new baseline for weekly average forecasting of 8 weeks and provides competitive results in seasonal SIC forecasts. To the best of our knowledge, IceDiff is the first attempt to forecast SIC at $6.25\\mathrm{km}$ spatial scale which is a quarter of the grid size $(25\\mathrm{km})$ that current methods perform on. The experimental results show that our framework can generate down-scaled SIC superior to off-the-shelf methods and it is vital for pushing forecasting models towards operational usage and further promoting sea ice research.", + "bbox": [ + 511, + 491, + 906, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Existing sea ice concentration data products are all retrieved from satellite observations with spatial resolution limited to $25\\mathrm{km}$ . A finer-scale ground truth is missing compared with observations from stations in weather forecasting. Therefore, our limitation lies in the evaluations of IceDiff on these in situ observations. We will focus on evaluating our IceDiff on higher-resolution SIC observations.", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "10574", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sahara Ali and Jianwu Wang. Mt-icenet-a spatial and multi-temporal deep learning model for arctic sea ice forecasting. In 2022 IEEE/ACM International Conference on Big Data Computing, Applications and Technologies (BDCAT), pages 1–10. IEEE, 2022. 2, 3, 4, 6", + "[2] Sahara Ali, Yiyi Huang, Xin Huang, and Jianwu Wang. Sea ice forecasting using attention-based ensemble LSTM. arXiv preprint arXiv:2108.00853, 2021. 2", + "[3] Tom R Andersson, J Scott Hosking, María Pérez-Ortiz, Brooks Paige, Andrew Elliott, Chris Russell, Stephen Law, Daniel C Jones, Jeremy Wilkinson, Tony Phillips, et al. Seasonal arctic sea ice forecasting with probabilistic deep learning. Nature communications, 12(1):5124, 2021. 1, 2, 3, 4, 6", + "[4] Saeed Anwar, Salman Khan, and Nick Barnes. A deep journey into super-resolution: A survey. ACM Computing Surveys (CSUR), 53(3):1-34, 2020. 3", + "[5] Seth Bassetti, Brian Hutchinson, Claudia Tebaldi, and Ben Kravitz. Diffesm: Conditional emulation of temperature and precipitation in earth system models with 3d diffusion models. Journal of Advances in Modeling Earth Systems, 16(10): e2023MS004194, 2024. 1", + "[6] Kaifeng Bi, Lingxi Xie, Hengheng Zhang, Xin Chen, Xiaotao Gu, and Qi Tian. Accurate medium-range global weather forecasting with 3d neural networks. Nature, 619(7970): 533-538, 2023. 3", + "[7] Bruno Buongiorno Nardelli, Davide Cavaliere, Elodie Charles, and Daniele Ciani. Super-resolving ocean dynamics from space with computer vision algorithms. Remote Sensing, 14(5):1159, 2022. 3", + "[8] Hu Cao, Yueyue Wang, Joy Chen, Dongsheng Jiang, Xiaopeng Zhang, Qi Tian, and Manning Wang. Swin-unet: Unet-like pure transformer for medical image segmentation. In European conference on computer vision, pages 205-218. Springer, 2022. 3, 4", + "[9] D.J. Cavalieri, P. Gloersen, and W.J. Campbell. Determination of sea ice parameters with the nimbus-7 smmr. Journal of Geophysical Research, 89:5355-5369, 1984. 7", + "[10] D.J. Cavalieri, P. Gloersen, and W.J. Campbell. Deriving long-term time series of sea ice cover from satellite passive-microwave multisensor data sets. Journal of Geophysical Research, 104:15803-15814, 1999. 1", + "[11] Lei Chen, Fei Du, Yuan Hu, Zhibin Wang, and Fan Wang. Swinrdm: integrate swirrnn with diffusion model towards high-resolution and high-quality weather forecasting. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 322-330, 2023. 1, 3, 7, 8", + "[12] Lei Chen, Xiaohui Zhong, Feng Zhang, Yuan Cheng, Yinghui Xu, Yuan Qi, and Hao Li. Fuxi: A cascade machine learning forecasting system for 15-day global weather forecast. npj Climate and Atmospheric Science, 6(1):190, 2023. 3", + "[13] J.C. Comiso. Characteristics of arctic winter sea ice from satellite multispectral microwave observations. Journal of Geophysical Research, 91:975-994, 1986. 6" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Josefino C. Comiso. Bootstrap sea ice concentrations from nimbus-7 smmr and dmsp ssm/i-ssmis, version 3. 2017. 6", + "[15] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 4", + "[16] N. DiGirolamo, C.L. Parkinson, D.J. Cavalieri, P. Gloersen, and H.J. Zwally. Sea ice concentrations from nimbus-7 smmr //and dmsp ssm/i-ssmis passive microwave data. (nsidc-0051, version 1). National Snow and Ice Data Center Distributed Active Archive Center., 1996. 1, 6", + "[17] Ben Fei, Zhaoyang Lyu, Liang Pan, Junzhe Zhang, Weidong Yang, Tianyue Luo, Bo Zhang, and Bo Dai. Generative diffusion prior for unified image restoration and enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9935-9946, 2023. 3, 6, 7, 8", + "[18] Helge F Goessling, Steffen Tietsche, Jonathan J Day, Ed Hawkins, and Thomas Jung. Predictability of the arctic sea ice edge. Geophysical Research Letters, 43(4):1642-1650, 2016. 7", + "[19] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 7", + "[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1", + "[21] Yuan Hu, Lei Chen, Zhibin Wang, and Hao Li. Swinvrnn: A data-driven ensemble forecasting model via learned distribution perturbation. Journal of Advances in Modeling Earth Systems, 15(2):e2022MS003211, 2023. 1", + "[22] Young Jun Kim, Hyun-Cheol Kim, Daehyeon Han, Sanggyun Lee, and Jungho Im. Prediction of monthly arctic sea ice concentrations using satellite and reanalysis data based on convolutional neural networks. The Cryosphere, 14(3): 1083-1104, 2020. 2", + "[23] Jussi Leinonen, Daniele Nerini, and Alexis Berne. Stochastic super-resolution for downscaling time-evolving atmospheric fields with a generative adversarial network. IEEE Transactions on Geoscience and Remote Sensing, 59(9):7211-7223, 2020. 3", + "[24] Yang Liu, Laurens Bogaardt, Jisk Attema, and Wilco Hazeleger. Extended-range arctic sea ice forecast with convolutional long short-term memory networks. Monthly Weather Review, 149(6):1673-1693, 2021. 2", + "[25] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 3", + "[26] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12009-12019, 2022. 3", + "[27] Zhi-Song Liu, Wan-Chi Siu, and Yui-Lam Chan. Photorealistic image super-resolution via variational autoencoders." + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "10575", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE Transactions on Circuits and Systems for video Technology, 31(4):1351-1365, 2020. 3", + "[28] Zhi-Song Liu, Wan-Chi Siu, Li-Wen Wang, Chu-Tak Li, and Marie-Paule Cani. Unsupervised real image super-resolution via generative variational autoencoder. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 442-443, 2020. 3", + "[29] W.N. Meier, F. Fetterer, A.K. Windnagel, and J.S. Stewart. Noaa/nsidc climate data record of passive microwave sea ice concentration. (g02202, version 4). National Snow and Ice Data Center, 2021. 6", + "[30] Bin Mu, Xiaodan Luo, Shijin Yuan, and Xi Liang. Icetft v 1.0. 0: Interpretable long-term prediction of arctic sea ice extent with deep learning. Geoscientific Model Development Discussions, 2023:1-28, 2023. 2", + "[31] J.E. Nash and J.V. Sutcliffe. River flow forecasting through conceptual models part i — a discussion of principles. Journal of Hydrology, 10(3):282-290, 1970. 7", + "[32] Xingang Pan, Xiaohang Zhan, Bo Dai, Dahua Lin, Chen Change Loy, and Ping Luo. Exploiting deep generative prior for versatile image restoration and manipulation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):7474-7489, 2021. 6, 7", + "[33] Zisis I Petrou and Yingli Tian. Prediction of sea ice motion with convolutional long short-term memory networks. IEEE Transactions on Geoscience and Remote Sensing, 57 (9):6865-6876, 2019. 2", + "[34] Yibin Ren and Xiaofeng Li. Predicting the daily sea ice concentration on a subseasonal scale of the pan-arctic during the melting season by a deep learning model. IEEE Transactions on Geoscience and Remote Sensing, 61:1-15, 2023. 1, 2, 6", + "[35] Y Ren, X Li, and W Zhang. A data-driven deep learning model for weekly sea ice concentration prediction of the panarctic during the melting season, iee t. geosci. remote, 60, 4304819, 2022. 1, 2, 3, 4, 6, 7", + "[36] Eduardo Rocha Rodrigues, Igor Oliveira, Renato Cunha, and Marco Netto. Deepdownscale: A deep learning strategy for high-resolution weather forecast. In 2018 IEEE 14th International Conference on e-Science (e-Science), pages 415–422. IEEE, 2018. 3", + "[37] Qi Shao, Guangchao Hou, Wei Li, Guijun Han, Kangzhuang Liang, and Yang Bai. Ocean reanalysis data-driven deep learning forecast for sea surface multivariate in the south china sea. Earth and Space Science, 8(7):e2020EA001558, 2021. 3", + "[38] S Sparnocchia, N Pinardi, and E Demirov. Multivariate empirical orthogonal function analysis of the upper thermocline structure of the mediterranean sea from observations and model simulations. In Annales Geophysicae, pages 167-187. Copernicus GmbH, 2003. 3", + "[39] Karen Stengel, Andrew Glaws, Dylan Hettinger, and Ryan N King. Adversarial super-resolution of climatological wind and solar data. Proceedings of the National Academy of Sciences, 117(29):16805-16815, 2020. 3", + "[40] Siwei Tu, Weidong Yang, and Ben Fei. Taming generative diffusion for universal blind image restoration. arXiv preprint arXiv:2408.11287, 2024. 3" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Lei Wang, Xiaojun Yuan, and Cuihua Li. Subseasonal forecast of arctic sea ice concentration via statistical approaches. Climate Dynamics, 52:4953-4971, 2019. 1", + "[42] Qingyu Zheng, Ru Wang, Guijun Han, Wei Li, Xuan Wang, Qi Shao, Xiaobo Wu, Lige Cao, Gongfu Zhou, and Song Hu. A spatio-temporal multiscale deep learning model for subseasonal prediction of arctic sea ice. IEEE Transactions on Geoscience and Remote Sensing, 2024. 3, 6", + "[43] Xiaohui Zhong, Lei Chen, Jun Liu, Chensen Lin, Yuan Qi, and Hao Li. Fuxi-extreme: Improving extreme rainfall and wind forecasts with diffusion model. Science China Earth Sciences, pages 1-13, 2024. 1", + "[44] Ruofan Zhou and Sabine Susstrunk. Kernel modeling superresolution on real low-resolution images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2433-2443, 2019. 3" + ], + "bbox": [ + 514, + 90, + 903, + 316 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10576", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/IceDiff_ High Resolution and High-Quality Arctic Sea Ice Forecasting with Generative Diffusion Prior/3f2f5e81-8b45-4fda-b37e-2ea2e3413d80_model.json b/2025/IceDiff_ High Resolution and High-Quality Arctic Sea Ice Forecasting with Generative Diffusion Prior/3f2f5e81-8b45-4fda-b37e-2ea2e3413d80_model.json new file mode 100644 index 0000000000000000000000000000000000000000..89716d4364a260147d167386272512ae62988662 --- /dev/null +++ b/2025/IceDiff_ High Resolution and High-Quality Arctic Sea Ice Forecasting with Generative Diffusion Prior/3f2f5e81-8b45-4fda-b37e-2ea2e3413d80_model.json @@ -0,0 +1,1958 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.124, + 0.13, + 0.877, + 0.175 + ], + "angle": 0, + "content": "IceDiff: High Resolution and High-Quality Arctic Sea Ice Forecasting with Generative Diffusion Prior" + }, + { + "type": "text", + "bbox": [ + 0.107, + 0.203, + 0.892, + 0.311 + ], + "angle": 0, + "content": "Jingyi Xu\\(^{1,\\ast}\\), Siwei Tu\\(^{1,\\ast}\\), Weidong Yang\\(^{1,\\dagger}\\), Ben Fei\\(^{3,\\dagger}\\), Shuhao Li\\(^{1}\\), Keyi Liu\\(^{1}\\), Yeqi Luo\\(^{1}\\), Lipeng Ma\\(^{1}\\), Lei Bai\\(^{2}\\) \n\\(^{1}\\)Fudan University, \\(^{2}\\)Shanghai Artificial Intelligence Laboratory, \\(^{3}\\)Chinese University of Hong Kong jyxu22@m.fudan.edu.cn, wdyang@fudan.edu.cn, benfei@cuhk.edu.hk, baisanshi@gmail.com * Equal Contributions, † Corresponding Authors" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.345, + 0.329, + 0.361 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.377, + 0.486, + 0.756 + ], + "angle": 0, + "content": "Variation of Arctic sea ice has significant impacts on polar ecosystems, transporting routes, coastal communities, and global climate. Tracing the change of sea ice at a finer scale is paramount for both operational applications and scientific studies. Recent pan-Arctic sea ice forecasting methods that leverage advances in artificial intelligence have made promising progress over numerical models. However, forecasting sea ice at higher resolutions is still under-explored. To bridge the gap, we propose a two-module cooperative deep learning framework, IceDiff, to forecast sea ice concentration at finer scales. IceDiff first leverages a vision transformer to generate coarse yet superior forecasting results over previous methods at a regular \\(25\\mathrm{km}\\) grid. This high-quality sea ice forecasting can be utilized as reliable guidance for the next module. Subsequently, an unconditional diffusion model pre-trained on low-resolution sea ice concentration maps is utilized for sampling down-scaled sea ice forecasting via a zero-shot guided sampling strategy and a patch-based method. For the first time, IceDiff demonstrates sea ice forecasting with a \\(6.25\\mathrm{km}\\) resolution. IceDiff extends the boundary of existing sea ice forecasting models and more importantly, its capability to generate high-resolution sea ice concentration data is vital for pragmatic usages and research. Code is available at https://github.com/EtronTech/IceDiff." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.784, + 0.224, + 0.8 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Arctic sea ice is essential for global climate change and local communities. Although the exact teleconnections between mid-latitude weather and the variation of Arctic sea ice are yet to be revealed, the reduction of sea ice during the last few decades clearly has impacts that transcend the Arctic region [3, 41]. Hence, establishing skillful forecasting" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.346, + 0.907, + 0.378 + ], + "angle": 0, + "content": "models at finer scales is paramount for both transportation in polar regions and scientific research in geoscience." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.382, + 0.909, + 0.609 + ], + "angle": 0, + "content": "Existing physics-based numerical models have demonstrated the ability to accurately forecast sea ice concentration (SIC, value ranges from \\(0\\%\\) to \\(100\\%\\)) for short lead times of a few weeks. However, when forecasting at longer scales, their performance significantly decreases, and their forecasting skills are even worse than those of simple statistical models [3]. With the advent of deep learning methods, a series of works based on deep learning approaches have been committed to SIC forecasting with promising results. Notable studies cover forecasting leads range from daily to seasonal scale, e.g., IceNet [3] provides accurate sea ice extent (SIE, areas where the SIC value is greater than \\(15\\%\\) and could be covered with sea ice) forecasts for 6 months lead time. SICNet [35] and \\(\\mathrm{SICNet}_{90}\\) [34] are skillful for forecasting SIC in subsequent 7 and 90 days, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Despite the improved forecasting skills of deep learning models over previous approaches, these models are still trained on sea ice products provided by the National Sea Ice Data Center (NSIDC) or the European Organisation for the Exploitation of Meteorological Satellites (EU-METSAT). These sea ice data products are primarily derived from passive microwave sensors with a resolution of \\(25\\mathrm{km}\\), which is considered too coarse for sea ice modeling. The in situ observations are costly to acquire and extremely spare to provide sufficient spatial coverage over pan-Arctic region. This limitation hinders the representation of detailed information on sea ice concentration and the accuracy of ice edge [10, 16]. The same challenge is also presented in weather forecasts. Therefore, several endeavors are proposed to work on downscaling atmospheric states [5, 11, 43]. For instance, SwinRDM [11] combines an improved SwinRNN [21] model with a denoising diffusion probabilistic model (DDPM) [20] to achieve superresolution (SR) in forecasting climate variables. However," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10567" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.094, + 0.092, + 0.481, + 0.136 + ], + "angle": 0, + "content": "forecasting sea ice at finer spatial scales, which is crucial, especially for operational applications is still under-explored." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.141, + 0.482, + 0.775 + ], + "angle": 0, + "content": "Due to the absence of high-resolution sea ice data products and sparse in situ observations, it is not feasible to train a forecast model over pan-Arctic region at a finer scale. Two primary solutions can address this challenge: (i) Downscale sea ice maps initially and then train a forecast model using these high-resolution SIC maps. (ii) Train a forecast model at the original resolution based on the sea ice products, then downscale the forecast results to a finer scale. In the first solution, downscaling sea ice maps may introduce slight errors, which could accumulate during the forecasting stage. Additionally, training a forecast model on high-resolution SIC maps also imposes a greater computational burden. Therefore, we opt for the latter solution and introduce IceDiff, a two-module cooperative framework. IceDiff aims to provide skillful forecasts at three different temporal scales in the forecasting module (FM) and produce spatially downscaled high-quality SIC maps in the guided diffusion module (GDM). Specifically, in the first module, IceDiff leverages U-Net-based vision transformer and provides forecasts that are superior to existing SIC forecasting models. Then, in the second module, we utilize a well-trained DDPM on the original resolution from the NSIDC production as effective prior and downscale forecasted SIC maps guided by forecasts from the previous module at the original scale. Our DDPM is pre-trained on daily SIC maps with a resolution of \\(25\\mathrm{km}\\), and it can also generalize well to weekly and monthly SIC maps. In every step \\(t\\) of the sampling process of DDPM, the intermediate variable \\(x_0\\) is estimated from the denoised SIC map \\(x_{t}\\). Then we propose a convolution kernel with estimizable parameters to simulate the up-scale degradation for the intermediate variable \\(x_0\\), which results in up-scaled \\(\\tilde{x}_0\\). A distance function between the low-resolution SIC map and up-scaled \\(\\tilde{x}_0\\) is utilized to dynamically update the parameters of the convolution kernel in the reverse steps and guide the sampling of SIC map in the next time step. Furthermore, to make our IceDiff suitable for SIC maps with various resolutions, a patch-based method is also designed to downscale SIC maps at artistry resolution. We show in this work that using low-resolution SIC maps as guidance can better capture small-scale details in the generated high-resolution SIC maps. Our contributions are three folds:" + }, + { + "type": "text", + "bbox": [ + 0.095, + 0.78, + 0.482, + 0.9 + ], + "angle": 0, + "content": "- We propose a novel SIC forecasting framework IceDiff that demonstrates skills in three different temporal scales which cover subseasonal to seasonal forecasts, specifically, 7 days, weekly average of 8 weeks and monthly average of 6 months, that are commonly investigated by the community. IceDiff achieves state-of-the-art results in all three scales compared to existing deep learning models." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.092, + 0.905, + 0.242 + ], + "angle": 0, + "content": "- IceDiff is devised with an estimizable convolutional kernel to simulate the up-scaling degradation and designs a patch-based method for downscaling SIC maps at arbitrary resolution with a single and fixed-resolution unconditional diffusion model. By utilizing the generative diffusion prior inherent in this diffusion model, the high-resolution SIC maps can be obtained through zero-shot sampling. Moreover, our IceDiff is capable of capturing small-scale details and generating high-resolution SIC maps." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.243, + 0.905, + 0.348 + ], + "angle": 0, + "content": "- Extensive experiments show that our IceDiff could provide accurate SIC forecasts and generate high-quality sea ice samples with finer details. IceDiff not only provides a new and comprehensive baseline for data-driven sea ice research but also extends the capability of existing SIC forecasting models and pushes this area of research towards more operational usage." + }, + { + "type": "title", + "bbox": [ + 0.517, + 0.367, + 0.662, + 0.383 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.517, + 0.394, + 0.865, + 0.41 + ], + "angle": 0, + "content": "2.1. Deep Learning-based Sea Ice Forecasting" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.418, + 0.905, + 0.9 + ], + "angle": 0, + "content": "Deep learning models have drawn the attention of sea ice research communities and have been widely investigated for Arctic sea ice forecasting [1, 2, 22, 33]. As a pioneering work, ConvLSTM [24] predicts SIC within the region of Barents Sea and has showcased comparable results to the state-of-the-art numerical climate models. IceNet [3] employs the U-Net architecture for merging pre-selected ERA5 variables and SIC data to improve the accuracy of seasonal SIE prediction up to six months. The experimental results have shown that the forecasting skill of IceNet in pan-Arctic region is superior to the numerical sea ice simulation model SEAS5. MT-IceNet [1] adopts a similar approach to IceNet and adds bi-monthly data to further improve forecasting SIC at seasonal scale. Mu et al. [30] propose to merge atmospheric and oceanographic data in a weighted manner through a variable selection network and provide nine months of forecasting leads. Considering the complex ice-ocean-atmosphere interactions have already embedded in the spatiotemporal correlations of consecutive SIC maps [33, 35], Ren et al. [35] proposed SICNet that solely rely on SIC data to predict concentration for the next seven days. Similar to IceNet, SICNet is built upon U-Net architecture and added with channel-wise dependency via temporal-spatial attention module (TSAM), which is essentially a revised version of the convolutional block attention module. By fusing the output of two SICNet, i.e. one for modeling SIC and the other for extracting climate information, and adjusting the number of output channels, Ren et al. [34] extends the lead time to 90 days. Besides CNN and LSTM-based models, recent works leverage the latest advances in artificial intelligence to further improve forecasting performance. For instance, Ice" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.947, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10568" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.482, + 0.212 + ], + "angle": 0, + "content": "Former [42] jointly models SIC and sea ice thickness (SIT) based on Transformer backbone. It first decomposes high dimensional spatio-temporal data into sequences via multivariate empirical orthogonal functions [37, 38]. Then an encoder-decoder Transformer is utilized to predict SIC and SIT for up to 45 days. Despite the performance gain of the aforementioned deep learning models, none of them have addressed forecasting sea ice at scales lower than \\(25\\mathrm{km}\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.227, + 0.262, + 0.243 + ], + "angle": 0, + "content": "2.2. Super Resolution" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.25, + 0.485, + 0.688 + ], + "angle": 0, + "content": "Super-resolution aims to reconstruct high-resolution (HR) images from low-resolution (LR) counterparts. Several works [27, 28] explore the use of Variational Autoencoders for image SR. These methods usually improve the restoration quality by learning the conditional distribution of high-resolution images induced by low-resolution images. Although these methods are able to achieve realistic perceptual quality, VAE-based methods are often not as satisfactory due to their limited generation capabilities. To this end, other deep learning methods, such as Convolutional neural networks (CNNs), have been widely used for superresolution tasks in various domains. CNN-based methods typically use HR images corresponding to LR images to supervise training with reconstruction loss [4, 7, 36, 44]. However, these network trained solely on reconstruction loss makes it difficult to capture detailed information in images and generate diverse image results. Therefore, Generative Adversarial Networks (GANs) are proposed to solve these challenges. Stengel et al. [39] utilized GAN in a two-step SR process to achieve downscaling of climatological wind and solar irradiance. Leinonen et al. [23] introduced a recurrent, stochastic SR GAN for down-scaling time-evolving atmospheric fields, demonstrating the potential of GANs in improving the spatial resolution of LR images in atmospheric sciences applications. These models encourage the generator to generate high-quality images that the discriminator cannot distinguish from real images. Although GAN-based methods obtain satisfactory downscaled results, the diversity could be enhanced." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Currently, diffusion model-based methods have been more widely used due to their diversity in generating images and the ability to generate high-quality images comparable to GANs [17, 40]. For instance, Chen et al. [11] utilize SR based on the conditional model to recover the high spatial resolution and finer-scale atmospheric details, unfolding the ability of diffusion models to generate high-quality and detailed images in SR tasks. The objective of this study is to forecast SIC maps at a high resolution of \\(6.25\\mathrm{km}\\). To achieve this, we have improved the SwinTransformer for predicting SIC at the standard resolution comparable to current methods. Additionally, we employ a diffusion-based SR model to generate high-resolution and high-quality results via zero-shot sampling." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.65, + 0.108 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.115, + 0.906, + 0.221 + ], + "angle": 0, + "content": "Our IceDiff framework comprises two modules as depicted in Figure 1. The forecasting module provides skillful forecasts at three different temporal scales. The guided diffusion module produces spatially down-scaled high-quality SIC maps with a pre-trained DDPM via zero-shot sampling and the patch-based method. We elaborate on design details in the following sections." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.231, + 0.704, + 0.246 + ], + "angle": 0, + "content": "3.1. Forecasting Module" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.253, + 0.907, + 0.644 + ], + "angle": 0, + "content": "To achieve more accurate SIC forecasting at a finer resolution, a skillful forecasting module trained on original sea ice data products is primarily encouraged. In contrast to existing SIC forecasting methods that primarily concentrate on a single temporal scale [1, 3, 42], our objective is to enhance forecasting skill across various temporal scales and increase versatility. Encouraged by the success of Swin Transformer [25] in various data-driven weather forecasting methods [6, 11, 12], we propose to construct forecasting module by utilizing the U-Net architecture [8] and building on top of the Swin Transformer V2 backbone [26] with ResNet\\(_{TSAM}\\) block [35] for better capturing both spatial and temporal features of SIC maps. As illustrated in Figure 1, the FM takes \\(N\\) consecutive SIC maps, each of which has a shape of [448, 304], to predict subsequent maps with identical temporal length as the input at \\(25\\mathrm{km}\\) spatial scale. It consists of encoder, bottleneck and decoder. The patch embedding layer first applies a \\(2\\times 2\\) windows size to partition the original SIC input into patches and maps them into tokens via 2D convolution. Then, the temporal and spatial dependencies are modeled through the ResNet\\(_{TSAM}\\) block. Both the encoder and the decoder have 3 layers of Swin Transformers that comprise a pre-configured even number of Swin Transformer Blocks for feature extraction and restoration. The calculation of two consecutive blocks can be described as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.655, + 0.905, + 0.738 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} z _ {s} ^ {b} = L N (W M S A (z ^ {b - 1})) + z ^ {b - 1}, \\\\ z ^ {b} = L N (M L P (z _ {s} ^ {b})) + z _ {s} ^ {b}, \\\\ z _ {s} ^ {b + 1} = L N (S W M S A (z ^ {b})) + z ^ {b}, \\\\ z ^ {b + 1} = L N \\left(M L P \\left(z _ {s} ^ {b + 1}\\right)\\right) + z _ {s} ^ {b}, \\tag {1} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.749, + 0.907, + 0.901 + ], + "angle": 0, + "content": "where \\( z_{s}^{b} \\) and \\( z^{b} \\) represents the output feature of the (Shifted) Window-Multi-head Self Attention module and the MLP module for block \\( b \\), respectively [26]. LN denotes the LayerNorm operation. The combination of such two consecutive blocks, i.e. the former for the employment of WMSA and the latter for SWMSA, ensures that local patterns of adjacent grids are sufficiently attended by the backbone network. We further adopt a shift window size of \\( 28 \\times 19 \\) that is consistent with the ratio of SIC input map for facilitating the SWMHA mechanism. The output of the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10569" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.087, + 0.91, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.372, + 0.908, + 0.457 + ], + "angle": 0, + "content": "Figure 1. Overview of IceDiff, a two-module cooperative framework to forecast sea ice concentration at a finer scale. IceDiff first employs the forecasting module (FM) that takes history SIC data as input to generate subsequent predictions at \\(25\\mathrm{km}\\) resolution. \\(N\\) denotes the number of sequential time stamps, for instance, set \\(N\\) to 7 for seven days of forecasting leads, FM takes a tensor with the shape of \\(\\{7,448,304\\}\\) as input and forecasts the SIC for the next 7 days. \\(C\\) represents the quantity of embedded channels within each layer of FM, the original SIC map data has a channel of 1 rather than 3 for regular images. The guided diffusion module (GDM) then leverages accurate forecasts as guidance to generate downscaled samples at \\(6.25\\mathrm{km}\\) resolution with high quality, see details in Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.471, + 0.486, + 0.684 + ], + "angle": 0, + "content": "last block within each layer is delivered to a ResNet\\(_{TSAM}\\) block for establishing channel-wise dependencies. Before the subsequent encoder and decoder layers, patch merging and patch expanding [8] are respectively performed to reshape the feature and fit it to the desired size. Following prior works [3, 8, 35], the skip connection is added between the corresponding encoder and decoder layers to improve the performance. The final prediction is generated via a patch expanding layer with ResNet\\(_{TSAM}\\) that up-samples the decoded feature to the size of the original SIC input. Our meticulous designs enable the forecasting module in IceDiff to be applied across various temporal scales, enhancing forecasting skills, which serves as a more effective initiation for the guided diffusion module of IceDiff." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.697, + 0.325, + 0.71 + ], + "angle": 0, + "content": "3.2. Guided Diffusion Module" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Off-the-shelf methods are limited to the forecasting of SIC maps at a scale of \\(25\\mathrm{km}\\) from sea ice data products [1, 3, 35]. However, finer-scale SIC maps remain unexplored which is of significant value to ecosystems, transporting routes, coastal communities, and global climate. Therefore, after obtaining skillful forecasts, we exploit a downscaling module based on an unconditional diffusion model pre-trained on daily SIC maps to generate realistic high-resolution SIC maps with rich details. This guided diffusion module pre-trained on daily SIC maps could generalize well to SIC maps at various scales, such as weekly or monthly SIC maps. Moreover, our guided diffusion module" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.471, + 0.907, + 0.517 + ], + "angle": 0, + "content": "of IceDiff downscales the SIC forecasts from a forecasting module to a scale of \\(6.25\\mathrm{km}\\), further expanding the accuracy and practicality of sea ice prediction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.516, + 0.907, + 0.587 + ], + "angle": 0, + "content": "Specifically, the reverse process of the diffusion model is conditioned on the low-resolution sea ice map \\(y\\), which transforms distribution \\(p_{\\theta}(x_{t-1}|x_t)\\) into conditional distribution \\(p_{\\theta}(x_{t-1}|x_t,y)\\). Previous work [15] has derived the conditional transformation formula in the reverse process:" + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.597, + 0.905, + 0.63 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\log p _ {\\theta} \\left(x _ {t} \\mid x _ {t + 1}, y\\right) = \\log \\left(p _ {\\theta} \\left(x _ {t} \\mid x _ {t + 1}\\right) p \\left(y \\mid x _ {t}\\right)\\right) + N _ {1} (2) \\\\ \\approx \\log p _ {\\theta} (z) + N _ {2}, z \\sim \\mathcal {N} (z; \\mu_ {\\theta} \\left(x _ {t}, t\\right) + \\Sigma g, \\Sigma I), (3) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.64, + 0.906, + 0.686 + ], + "angle": 0, + "content": "where \\(g\\) refers to \\(\\nabla_{x_t}\\log p(y|x_t)\\). \\(N_{1} = \\log p_{\\theta}(y|x_{t + 1})\\), \\(N_{2}\\) is a constant related to \\(g\\) (proof in Appendix). The variance of the reverse process \\(\\Sigma = \\Sigma_{\\theta}(x_{t})\\) is constant." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.686, + 0.906, + 0.756 + ], + "angle": 0, + "content": "Therefore, the sampling process of \\( p_{\\theta}(x_{t - 1}|x_t,y) \\) integrates the gradient term \\( g \\) with the diffusion model \\( (\\mu ,\\Sigma) = (\\mu_{\\theta}(x_{t}),\\Sigma_{\\theta}(x_{t})) \\), which plays a role in controlling the direction of map generation. We use a heuristic algorithm (proof in Appendix) to approximate the value of \\( g \\):" + }, + { + "type": "equation", + "bbox": [ + 0.582, + 0.766, + 0.907, + 0.782 + ], + "angle": 0, + "content": "\\[\n\\log p (y \\mid x _ {t}) = - \\log N - s \\mathcal {L} (\\mathcal {D} (\\tilde {x} _ {0}), y) \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.785, + 0.907, + 0.8 + ], + "angle": 0, + "content": "\\[\ng = \\nabla_ {x _ {t}} \\log p (y | x _ {t}) = - s \\nabla_ {x _ {t}} \\mathcal {L} (\\mathcal {D} (x _ {t}), y). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Among them, \\( s \\) is the scaling factor used to control the degree of guidance. And \\( \\mathcal{L} \\) is a distance function used to measure the distance deviation between two SIC maps. \\( \\mathcal{D} \\) is an up-scale convolution function with estimizable parameters, where the parameters are randomly initialized and dynamically optimized with the gradient of distance function at" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10570" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.127, + 0.117, + 0.87, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.252, + 0.462, + 0.271, + 0.472 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image_caption", + "bbox": [ + 0.541, + 0.462, + 0.56, + 0.472 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.778, + 0.462, + 0.796, + 0.472 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.475, + 0.907, + 0.545 + ], + "angle": 0, + "content": "Figure 2. Overview of guided diffusion module in IceDiff for SIC map downscaling. (a) Pre-trained diffusion model is used to estimate \\(\\tilde{x}_0\\) at reverse step \\(t\\). The estimable convolutional kernel of step \\(t\\) reduces the resolution of \\(X_{t}\\) so that guidance from low-resolution SIC maps can be incorporated to sample \\(X_{t-1}\\). (b) Low-resolution SIC map \\(Y\\) is integrated during the sampling process. A distance function describes the distance between a low-resolution SIC map and the generated high-resolution SIC map \\(x\\) after convolution. The gradient of distance function could be used to guide the sampling of \\(x_{t-1}\\) and (c) dynamically update the parameters of the convolutional kernel." + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.555, + 0.895, + 0.676 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.681, + 0.907, + 0.737 + ], + "angle": 0, + "content": "Figure 3. Illustration of the patch-based method. We set patch size \\( p = 256 \\) and stride \\( r = 128 \\) to fit the size of pre-trained models. This setting will divide the example map into \\( K = 4 \\) sub-maps boxed with different colors. For each sub-map, an individualizable convolution kernel with adaptive parameters is set to add guidance from the corresponding low-resolution SIC sub-map. At each sampling step \\( t \\), the average estimated noise on four overlapping blocks will be utilized to update the mean \\( \\mu \\) used to sample \\( x_{t-1} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.752, + 0.482, + 0.798 + ], + "angle": 0, + "content": "every reverse step. \\(\\mathcal{D}\\) connects two different resolution SIC maps, allowing low-resolution SIC maps to provide more effective guidance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.9 + ], + "angle": 0, + "content": "As shown in Algorithm 1, GDM undergoes \\(T\\) reverse steps to gradually restore pure Gaussian noise \\(x_{T} \\sim \\mathcal{N}(0, I)\\) to high-resolution SIC maps. In each reverse step \\(t\\), the diffusion model calculates the estimated variable \\(\\tilde{x}_{0}\\) which then undergoes the up-scale convolution function \\(\\mathcal{D}\\) at time \\(t\\) to establish a distance metric \\(\\mathcal{L}\\) with the low-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.752, + 0.907, + 0.889 + ], + "angle": 0, + "content": "resolution SIC map, where \\(\\mathcal{D}_t^\\varphi\\) signifies that its convolutional kernel parameters are \\(\\varphi = \\varphi_t\\) at the reverse steps \\(t\\). The \\(\\mathcal{L}\\) subscript \\(\\varphi = \\varphi_t\\) indicates that it is a function dependent on the convolutional kernel parameter \\(\\varphi\\) at reverse step \\(t\\). The gradients of distance function \\(\\mathcal{L}\\) with respect to \\(\\tilde{x}_0\\) and the convolution kernel parameters of the up-scale convolution function are used to update \\(\\tilde{x}_0\\) and the convolution kernel parameters, respectively. The update of convolution kernel parameters ensures the accuracy of the model's" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10571" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.146 + ], + "angle": 0, + "content": "Table 1. Quantitative results of forecasts. IceDiff is evaluated on three different temporal scales. The output length of IceDiff equals a pre-defined forecasting lead as depicted in Figure 1. The validated accuracy of IceDiff could provide reliable guidance for the guided diffusion module to generate high-quality downscaled SIC. BT represents that the SIC data [14] are generated by the Bootstrap retrieval algorithm [13] and the data distribution of CDR is considered to be closer to BT rather than NT [29]." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.147, + 0.905, + 0.332 + ], + "angle": 0, + "content": "
Temporal ScaleLead TimeDatasetMethodsRMSE↓MAE↓R²↑NSE↑IIEE↓SIEdif↓
Subseasonal7 Days (Daily)NTSICNet [35]0.04300.00910.9780.97511750.1127
NTIceDiff0.03240.00770.9870.9868990.0186
CDRSICNet [35]0.04900.01000.9820.9799760.0718
CDRIceDiff0.03960.00800.9890.9878350.0315
45 Days (Daily)BTIceFormer [42]0.06600.02010.960---
8 Weeks Average (Weekly)NTIceDiff0.04030.01080.9750.97211700.1051
CDRIceDiff0.05530.01120.9730.96913530.0871
90 Days (Daily)NTSICNet90 [34]-0.0512----
Seasonal6 Months Average (Monthly)NTIceNet [3]0.18200.09160.567---
NTMT-IceNet [1]0.07770.01970.915---
NTIceDiff0.05730.01880.9580.95220140.2419
CDRIceDiff0.06480.01680.9190.91320160.2657
" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.343, + 0.482, + 0.389 + ], + "angle": 0, + "content": "Algorithm 1 Guided diffusion module in IceDiff with the guidance of low-resolution SIC map \\(y\\). An unconditional diffusion model pre-trained on SIC maps \\(\\epsilon_{\\theta}(x_t,t)\\) is given." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.393, + 0.483, + 0.447 + ], + "angle": 0, + "content": "Input: Low-resolution SIC map \\(y\\), downscale convolution function \\(\\mathcal{D}\\) composed of optimized convolutional kernels \\(\\mathcal{K}\\) with parameters \\(\\varphi\\), learning rate \\(l\\), guidance scale \\(s\\), distant function \\(\\mathcal{L}\\)." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.449, + 0.482, + 0.476 + ], + "angle": 0, + "content": "Output: Output high-resolution SIC map \\( x_0 \\) conditioned on \\( y \\). Sample \\( x_T \\) from \\( \\mathcal{N}(0, I) \\)" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.477, + 0.258, + 0.488 + ], + "angle": 0, + "content": "1: for all t from T to 1 do" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.488, + 0.3, + 0.506 + ], + "angle": 0, + "content": "2: \\(\\tilde{x}_0 = \\frac{x_t}{\\sqrt{\\alpha_t}} - \\frac{\\sqrt{1 - \\alpha_t} \\epsilon_0(x_t, t)}{\\sqrt{\\alpha_t}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.506, + 0.277, + 0.52 + ], + "angle": 0, + "content": "3: \\(\\mathcal{L}_{\\varphi, \\tilde{x}_0} = \\mathcal{L}(y, \\mathcal{D}^\\varphi(\\tilde{x}_0))\\)" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.52, + 0.327, + 0.541 + ], + "angle": 0, + "content": "4: \\(\\tilde{x}_0\\gets \\tilde{x}_0 - \\frac{s(1 - \\bar{\\alpha}_t)}{\\sqrt{\\bar{\\alpha}_{t - 1}\\beta_t}}\\nabla_{\\tilde{x}_0}\\mathcal{L}_{\\varphi ,\\tilde{x}_0}\\)" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.541, + 0.354, + 0.563 + ], + "angle": 0, + "content": "5: \\(\\tilde{\\mu}_{t} = \\frac{\\sqrt{\\bar{\\alpha}_{t - 1}}\\beta_{t}}{1 - \\bar{\\alpha}_{t}}\\tilde{x}_{0} + \\frac{\\sqrt{\\bar{\\alpha}_{t}}(1 - \\bar{\\alpha}_{t - 1})}{1 - \\bar{\\alpha}_{t}} x_{t}\\)" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.563, + 0.231, + 0.58 + ], + "angle": 0, + "content": "6: \\(\\tilde{\\beta}_{t} = \\frac{1 - \\bar{\\alpha}_{t - 1}}{1 - \\bar{\\alpha}_{t}}\\beta_{t}\\)" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.58, + 0.318, + 0.596 + ], + "angle": 0, + "content": "7: Sample \\( x_{t-1} \\) from \\( \\mathcal{N}(\\tilde{\\mu}_t, \\tilde{\\beta}_t I) \\)" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.596, + 0.26, + 0.61 + ], + "angle": 0, + "content": "8: \\(\\varphi \\gets \\varphi - l \\nabla_{\\varphi} \\mathcal{L}_{\\varphi, \\bar{x}_0}\\)" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.61, + 0.167, + 0.621 + ], + "angle": 0, + "content": "9: end for" + }, + { + "type": "text", + "bbox": [ + 0.095, + 0.624, + 0.18, + 0.636 + ], + "angle": 0, + "content": "10: return \\( x_0 \\)" + }, + { + "type": "list", + "bbox": [ + 0.095, + 0.477, + 0.354, + 0.636 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.672, + 0.483, + 0.701 + ], + "angle": 0, + "content": "up-scaling process, making guidance of low-resolution SIC maps more precise and effective." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.705, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Patch-based Strategy. Furthermore, pre-trained models can only generate fixed-size SIC maps. To tackle the problem of sampling downscaled SIC forecasts at any resolution, we employ a patch-based strategy. As shown in Figure 3, the patch-based strategy partitions SIC maps into several sub-regions based on a fixed stride. For each individual sub-region, the gradient terms of distance loss between SIC maps upscaled by the convolutional kernel and the corresponding low-resolution SIC sub-region maps are computed individually. Subsequently, the respective mean of Gaussian distribution in sub-regions (Eq. 3) and convolution kernel parameters are updated based on the gradient terms. The entire map is updated by taking the mean of" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.342, + 0.907, + 0.383 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison between IceDiff for SIC downscaling and other methods. Const. and FID represents the Consistency and Fréchet Inception Distance metrics, respectively." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.385, + 0.905, + 0.508 + ], + "angle": 0, + "content": "
Method7 Days8 Week Avg.6 Month Avg.
Const.↓FID↓Const.↓FID↓Const.↓FID↓
Nearest20.0178.4521.9880.1723.0788.36
Bilinear17.8870.7520.0277.2918.3180.73
BiCubic13.1354.7113.4068.8510.0970.18
DGP [32]23.1875.4228.6281.4229.7485.47
GDP [17]8.2640.1510.5452.4010.9655.72
IceDiff6.8429.347.2531.296.6239.79
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.52, + 0.907, + 0.626 + ], + "angle": 0, + "content": "the updated values across all sub-regions, each weighted by a binary patch mask that serves as a quantifier of the regional scope, thereby refining the overall mean and convolution kernel parameters of the sampled high-resolution SIC maps. By leveraging this strategy, our model is capable of downscaling sea ice maps to the desired resolution and the practicality of the model has been further enhanced." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.641, + 0.646, + 0.658 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.667, + 0.759, + 0.684 + ], + "angle": 0, + "content": "4.1. Dataset and Baseline Setup" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.689, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We evaluate our proposed IceDiff framework on the Climate Data Record of Passive Microwave Sea Ice Concentration Version 4 dataset from NSIDC [29] (CDR in Table 1). It records daily SIC data starting from October \\(25^{th}\\) 1978 and provides the coverage of the pan-Arctic region \\((\\mathrm{N} : -39.36^{\\circ}, \\mathrm{S} : -89.84^{\\circ}, \\mathrm{E} : 180^{\\circ}, \\mathrm{W} : -180^{\\circ})\\). The SIC map is formed of \\(448 \\times 304\\) pixels and each of which represents an area of \\(25 \\mathrm{~km} \\times 25 \\mathrm{~km}\\) grid and has a value ranging from \\(0\\%\\) to \\(100\\%\\). We choose data from October \\(25^{th}\\) 1978 to 2013 as the training dataset, the years 2014 and 2015 are selected as validation set, and data from 2016 to 2023 are used for testing. To comprehensively validate and verify the performance of our forecasting module, we utilize additional data [16] that are generated through the retrieval" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10572" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.09, + 0.483, + 0.117 + ], + "angle": 0, + "content": "Table 3. Ablation study on TSAM. We compare the forecasting results between IceDiff with or without the TSAM." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.12, + 0.481, + 0.205 + ], + "angle": 0, + "content": "
Lead TimeTSAMRMSE↓MAE↓R²↑NSE↑IIEE↓SIEdif↓
7 DaysX0.04030.00820.98810.98608740.0336
✓0.03960.00800.98850.98658350.0315
8 Week Avg.X0.05350.01160.9750.97113860.1420
✓0.05530.01120.9730.96913530.0871
6 Month Avg.X0.07630.01940.8830.87526100.4919
✓0.06480.01680.9190.91320160.2657
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.217, + 0.484, + 0.415 + ], + "angle": 0, + "content": "algorithm (NASA Team [9], NT in Table 1). The forecasting module is trained on three different temporal scales, i.e. the input of SIC maps are 7 days, 8 weeks average, and 6 months average, covering forecasting leads from subseasonal to seasonal scale. Corresponding deep learning-based models are selected for comparison, as illustrated in Table 1. Note that we follow the split of datasets in respective baselines to preprocess the additional data. We re-implemented SICNet (the source code is not available) based on the original paper and trained it under an environment identical to our forecasting module. To evaluate the performance of our downscaling, five methods are selected to set up baselines for downscaling quality comparison." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.424, + 0.308, + 0.439 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.445, + 0.484, + 0.704 + ], + "angle": 0, + "content": "We independently trained the forecasting module of IceDiff on SIC dataset of three aforementioned temporal scales for 60 epochs. For each temporal scale, the forecasting module outputs the same length of input data as forecasting leads, i.e. 7 days, 8 weeks average, and 6 months averages. The dimensions of embedding features of encoder, decoder, and bottleneck are all set to 32. We utilize the same hyperparameter of TSAM as in [35]. The forecasting module is optimized by AdamW under the mean square error (MSE) objective using Pytorch on one NVIDIA A100 80GB GPU for all experiments with a learning rate of 0.0005. All daily SIC maps are resized and cropped to \\(256 \\times 256\\) to train our GDM of IceDiff. using AdamW with \\(\\beta_{1} = 0.9\\) and \\(\\beta_{2} = 0.999\\). We train in 16-bit precision using loss-scaling, but maintain 32-bit weights, EMA, and optimizer state. We use an EMA rate of 0.9999. When sampling with 1000 timesteps, we use the same noise schedule as for training." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.713, + 0.275, + 0.727 + ], + "angle": 0, + "content": "4.3. Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.734, + 0.483, + 0.87 + ], + "angle": 0, + "content": "To quantitatively evaluate our IceDff, we select commonly used root mean square error (RMSE) and mean absolute error (MAE) for the comparison of forecasting accuracy. \\( R^2 \\) score and Nash-Sutcliffe Efficiency [31] are utilized to analyze the similarity of spatial pattern between the ground truth and predicted SIC map. We adopt Integrated Ice-Edge Error score [18] and calculate the sea ice area differences \\( (SIE_{dif}) \\), in millions of \\( km^2 \\) to further evaluate the prediction of SIE (where the SIC value is greater than 15%)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "As to the evaluation of downscaling quality, following [11] we utilize the Fréchet Inception Distance (FID)" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.907, + 0.159 + ], + "angle": 0, + "content": "Table 4. Ablation study on the ratio of Input/Forecast number of SIC maps. Hf., Db. and Eq. represent that the forecasting module takes input data which has half, double, and equal length of forecasting leads, respectively. For the Hf. of 7 Days, we set the input length to 3." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.161, + 0.905, + 0.276 + ], + "angle": 0, + "content": "
Lead TimeI/F.RMSE↓MAE↓R²↑NSE↑IIEE↓SIEdif↓
7 DaysHf.0.04160.00840.9870.9859050.0538
Db.0.04010.00840.9880.9868890.0599
Eq.0.03960.00800.9890.9878350.0315
8 Week Avg.Hf.0.05800.01180.9700.96614050.1176
Db.0.05940.01210.9690.96414350.1221
Eq.0.05530.01120.9730.96913530.0871
6 Month Avg.Hf.0.09040.02320.8420.83031920.6289
Db.0.07220.01790.8970.88920270.2795
Eq.0.06480.01680.9190.91320160.2657
" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.289, + 0.905, + 0.335 + ], + "angle": 0, + "content": "[19] for measuring the downscaling fidelity. In addition, we adopt Consistency to quantify the faithfulness between the downscaled map and the low-resolution map." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.336, + 0.905, + 0.366 + ], + "angle": 0, + "content": "The detailed equations for calculating the abovementioned metrics can be found in Appendix." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.378, + 0.652, + 0.392 + ], + "angle": 0, + "content": "4.4. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.401, + 0.905, + 0.43 + ], + "angle": 0, + "content": "In this subsection, we evaluate the proposed IceDiff with baseline methods for validation of the effectiveness." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.431, + 0.907, + 0.598 + ], + "angle": 0, + "content": "Set up new baselines for SIC prediction. The forecasting module of IceDiff is trained independently to predict SIC at 7 days, 8 weeks average, and 6 months average. As shown in Table 1, the performance of our proposed forecasting module of IceDiff is superior to SICNet in all evaluation metrics. The forecasting skill of FM across different SIC datasets is consistent in three temporal scales. Specifically, it outperforms SOTA seasonal forecasting baselines by a noticeable margin. Considering the overall performance of our FM, it could provide reliable SIC forecasts and sufficiently support the subsequent downscaling module." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.906, + 0.81 + ], + "angle": 0, + "content": "Analysis of the end of the melting season. To further verify the performance of IceDiff in extreme cases, we select two years of the end of the melting season (early September) when the mean Arctic SIC typically reaches its minimum for validation: (i) In 2016, it recorded historically low annual SIC; (ii) In 2022, the annual SIC has an abrupt increase compared to the previous year (annual trends are plotted in the Appendix). We present the SIC residuals (predicted SIC minus the ground truth) of September \\(1^{st}\\) 2016 and 2022 in Figure 4(a) and Figure 4(b), respectively. The overall residuals are relatively low and the prediction errors mainly appear in part of the sea ice edge regions. The accurate forecasting results in temporal periods with highly varying SIC validate the effectiveness of IceDiff." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.901 + ], + "angle": 0, + "content": "High downscaling quality. We systematically compare the results of IceDiff with other downscaling methods. In terms of interpolation-based methods, we select nearest, bilinear, and bicubic interpolation to compare with our diffusion-based GDM. Besides, we select DGP [32] as a GAN-based method and GDP [17] as a DDPM-based method for fur" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10573" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.088, + 0.264, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.1, + 0.289, + 0.117, + 0.298 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image_caption", + "bbox": [ + 0.142, + 0.285, + 0.209, + 0.306 + ], + "angle": 0, + "content": "Forecast Date Sep 1,2016" + }, + { + "type": "image", + "bbox": [ + 0.264, + 0.088, + 0.482, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.269, + 0.29, + 0.286, + 0.3 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.315, + 0.285, + 0.381, + 0.306 + ], + "angle": 0, + "content": "Forecast Date Sep 1,2022" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.308, + 0.484, + 0.391 + ], + "angle": 0, + "content": "Figure 4. Qualitative analysis of the end of the melting season. The annual mean Arctic SIC has recorded a historical low in 2016 and it drastically increased in 2022 when compared to the previous annual mean. The residual plots of early September in (a)2016 and (b)2022 demonstrate where the FM overestimates (positive figures) or underestimates (negative figures) SIC." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.408, + 0.484, + 0.454 + ], + "angle": 0, + "content": "ther comparison. Interpolation-based methods upsample the low-resolution sea ice map, which uses existing pixel information to estimate missing pixel values." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.455, + 0.484, + 0.576 + ], + "angle": 0, + "content": "As shown in Figure 5, our GDM produces high-quality results and small-scale structures are successfully captured and rich details are generated through the guidance from the input map during each sampling step. By contrast, interpolation-based methods do not integrate details information of the low-resolution map, resulting in blurry edges and contours in the generated results. More quality analysis can be found in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.577, + 0.484, + 0.743 + ], + "angle": 0, + "content": "To further validate the superiority of our method, we follow SwinRDM [11] and GDP [17] to utilize FID and Consistency to evaluate our method, respectively. As shown in Table 2, compared to interpolation-based methods, our IceDiff improves FID score by 31.11 on average, reflecting it has the capacity to help generate a more realistic high-resolution map with a satisfactory texture restoration effect. A lower Consistency demonstrates the structural difference between the downscaled map and the original SIC input is relatively imperceptible, which better preserves the spatial distribution of sea ice forecast." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.757, + 0.388, + 0.773 + ], + "angle": 0, + "content": "4.5. Ablation Studies on Design of FM" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.901 + ], + "angle": 0, + "content": "As shown in Table 3, by adopting TSAM to FM, the performance has improved for all metrics at 7 days and 6 months average lead times while the effects for 8 weeks average lead time are not all positive. Overall, incorporating TSAM to establish channel-wise dependency is favorable for SIC forecast models. To examine the impacts of input SIC data length, we trained FM for input lengths of half, double, and equal to the forecasting leads. The results in Table 4 show" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.088, + 0.646, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.556, + 0.295, + 0.605, + 0.304 + ], + "angle": 0, + "content": "(a) IceDiff" + }, + { + "type": "image", + "bbox": [ + 0.646, + 0.089, + 0.777, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.685, + 0.295, + 0.736, + 0.304 + ], + "angle": 0, + "content": "(b) Bicubic" + }, + { + "type": "image", + "bbox": [ + 0.776, + 0.089, + 0.906, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.821, + 0.295, + 0.859, + 0.304 + ], + "angle": 0, + "content": "(c) GDP" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.307, + 0.907, + 0.405 + ], + "angle": 0, + "content": "Figure 5. Qualitative analysis of downscaled SIC map. We visualize the SIC map by coloring the land and the ocean in yellow and blue, respectively. Since SIC has a range of \\([0,1]\\), we map it to \\([0,255]\\) and replicate it for all 3 RGB channels. The darker the pixels appear, the SIC values are more closer to \\(0\\%\\). For better identification of the difference, we crop out a region and plot a grid size of \\(250\\mathrm{km}\\) and further display details on \\(4\\times 4\\) grid." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.419, + 0.907, + 0.451 + ], + "angle": 0, + "content": "the peak of the performance at the ratio where the input length equals the forecasting leads." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.466, + 0.771, + 0.482 + ], + "angle": 0, + "content": "5. Conclusion and Limitations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.492, + 0.907, + 0.794 + ], + "angle": 0, + "content": "We propose a novel pan-Arctic high-resolution SIC forecasting framework, IceDiff, which comprises a forecasting module and a guided diffusion module, these two modules closely cooperate with each other to fulfill accurate forecasting of SIC at a finer and more practical spatial scale. The forecasting module of IceDiff is capable of producing accurate forecasts at three different temporal scales and the guided diffusion module leverages reliable forecasts as guidance to generate high-quality down-scaled SIC maps. The forecasting skills of IceDiff at 7 days lead are superior to SOTA deep learning-based models. IceDiff also sets up a new baseline for weekly average forecasting of 8 weeks and provides competitive results in seasonal SIC forecasts. To the best of our knowledge, IceDiff is the first attempt to forecast SIC at \\(6.25\\mathrm{km}\\) spatial scale which is a quarter of the grid size \\((25\\mathrm{km})\\) that current methods perform on. The experimental results show that our framework can generate down-scaled SIC superior to off-the-shelf methods and it is vital for pushing forecasting models towards operational usage and further promoting sea ice research." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Existing sea ice concentration data products are all retrieved from satellite observations with spatial resolution limited to \\(25\\mathrm{km}\\). A finer-scale ground truth is missing compared with observations from stations in weather forecasting. Therefore, our limitation lies in the evaluations of IceDiff on these in situ observations. We will focus on evaluating our IceDiff on higher-resolution SIC observations." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10574" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Sahara Ali and Jianwu Wang. Mt-icenet-a spatial and multi-temporal deep learning model for arctic sea ice forecasting. In 2022 IEEE/ACM International Conference on Big Data Computing, Applications and Technologies (BDCAT), pages 1–10. IEEE, 2022. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.187, + 0.484, + 0.228 + ], + "angle": 0, + "content": "[2] Sahara Ali, Yiyi Huang, Xin Huang, and Jianwu Wang. Sea ice forecasting using attention-based ensemble LSTM. arXiv preprint arXiv:2108.00853, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.231, + 0.483, + 0.313 + ], + "angle": 0, + "content": "[3] Tom R Andersson, J Scott Hosking, María Pérez-Ortiz, Brooks Paige, Andrew Elliott, Chris Russell, Stephen Law, Daniel C Jones, Jeremy Wilkinson, Tony Phillips, et al. Seasonal arctic sea ice forecasting with probabilistic deep learning. Nature communications, 12(1):5124, 2021. 1, 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.316, + 0.483, + 0.357 + ], + "angle": 0, + "content": "[4] Saeed Anwar, Salman Khan, and Nick Barnes. A deep journey into super-resolution: A survey. ACM Computing Surveys (CSUR), 53(3):1-34, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.36, + 0.482, + 0.427 + ], + "angle": 0, + "content": "[5] Seth Bassetti, Brian Hutchinson, Claudia Tebaldi, and Ben Kravitz. Diffesm: Conditional emulation of temperature and precipitation in earth system models with 3d diffusion models. Journal of Advances in Modeling Earth Systems, 16(10): e2023MS004194, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.431, + 0.482, + 0.485 + ], + "angle": 0, + "content": "[6] Kaifeng Bi, Lingxi Xie, Hengheng Zhang, Xin Chen, Xiaotao Gu, and Qi Tian. Accurate medium-range global weather forecasting with 3d neural networks. Nature, 619(7970): 533-538, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.488, + 0.482, + 0.543 + ], + "angle": 0, + "content": "[7] Bruno Buongiorno Nardelli, Davide Cavaliere, Elodie Charles, and Daniele Ciani. Super-resolving ocean dynamics from space with computer vision algorithms. Remote Sensing, 14(5):1159, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.546, + 0.482, + 0.613 + ], + "angle": 0, + "content": "[8] Hu Cao, Yueyue Wang, Joy Chen, Dongsheng Jiang, Xiaopeng Zhang, Qi Tian, and Manning Wang. Swin-unet: Unet-like pure transformer for medical image segmentation. In European conference on computer vision, pages 205-218. Springer, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.617, + 0.482, + 0.657 + ], + "angle": 0, + "content": "[9] D.J. Cavalieri, P. Gloersen, and W.J. Campbell. Determination of sea ice parameters with the nimbus-7 smmr. Journal of Geophysical Research, 89:5355-5369, 1984. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.661, + 0.482, + 0.715 + ], + "angle": 0, + "content": "[10] D.J. Cavalieri, P. Gloersen, and W.J. Campbell. Deriving long-term time series of sea ice cover from satellite passive-microwave multisensor data sets. Journal of Geophysical Research, 104:15803-15814, 1999. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.718, + 0.482, + 0.786 + ], + "angle": 0, + "content": "[11] Lei Chen, Fei Du, Yuan Hu, Zhibin Wang, and Fan Wang. Swinrdm: integrate swirrnn with diffusion model towards high-resolution and high-quality weather forecasting. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 322-330, 2023. 1, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.482, + 0.856 + ], + "angle": 0, + "content": "[12] Lei Chen, Xiaohui Zhong, Feng Zhang, Yuan Cheng, Yinghui Xu, Yuan Qi, and Hao Li. Fuxi: A cascade machine learning forecasting system for 15-day global weather forecast. npj Climate and Atmospheric Science, 6(1):190, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[13] J.C. Comiso. Characteristics of arctic winter sea ice from satellite multispectral microwave observations. Journal of Geophysical Research, 91:975-994, 1986. 6" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "[14] Josefino C. Comiso. Bootstrap sea ice concentrations from nimbus-7 smmr and dmsp ssm/i-ssmis, version 3. 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[15] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[16] N. DiGirolamo, C.L. Parkinson, D.J. Cavalieri, P. Gloersen, and H.J. Zwally. Sea ice concentrations from nimbus-7 smmr //and dmsp ssm/i-ssmis passive microwave data. (nsidc-0051, version 1). National Snow and Ice Data Center Distributed Active Archive Center., 1996. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.234, + 0.905, + 0.314 + ], + "angle": 0, + "content": "[17] Ben Fei, Zhaoyang Lyu, Liang Pan, Junzhe Zhang, Weidong Yang, Tianyue Luo, Bo Zhang, and Bo Dai. Generative diffusion prior for unified image restoration and enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9935-9946, 2023. 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.317, + 0.905, + 0.37 + ], + "angle": 0, + "content": "[18] Helge F Goessling, Steffen Tietsche, Jonathan J Day, Ed Hawkins, and Thomas Jung. Predictability of the arctic sea ice edge. Geophysical Research Letters, 43(4):1642-1650, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.373, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[19] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.443, + 0.905, + 0.483 + ], + "angle": 0, + "content": "[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.486, + 0.905, + 0.538 + ], + "angle": 0, + "content": "[21] Yuan Hu, Lei Chen, Zhibin Wang, and Hao Li. Swinvrnn: A data-driven ensemble forecasting model via learned distribution perturbation. Journal of Advances in Modeling Earth Systems, 15(2):e2022MS003211, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.54, + 0.905, + 0.607 + ], + "angle": 0, + "content": "[22] Young Jun Kim, Hyun-Cheol Kim, Daehyeon Han, Sanggyun Lee, and Jungho Im. Prediction of monthly arctic sea ice concentrations using satellite and reanalysis data based on convolutional neural networks. The Cryosphere, 14(3): 1083-1104, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.61, + 0.905, + 0.676 + ], + "angle": 0, + "content": "[23] Jussi Leinonen, Daniele Nerini, and Alexis Berne. Stochastic super-resolution for downscaling time-evolving atmospheric fields with a generative adversarial network. IEEE Transactions on Geoscience and Remote Sensing, 59(9):7211-7223, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.679, + 0.905, + 0.734 + ], + "angle": 0, + "content": "[24] Yang Liu, Laurens Bogaardt, Jisk Attema, and Wilco Hazeleger. Extended-range arctic sea ice forecast with convolutional long short-term memory networks. Monthly Weather Review, 149(6):1673-1693, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.735, + 0.905, + 0.803 + ], + "angle": 0, + "content": "[25] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.805, + 0.905, + 0.873 + ], + "angle": 0, + "content": "[26] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12009-12019, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.875, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[27] Zhi-Song Liu, Wan-Chi Siu, and Yui-Lam Chan. Photorealistic image super-resolution via variational autoencoders." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10575" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "IEEE Transactions on Circuits and Systems for video Technology, 31(4):1351-1365, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.191 + ], + "angle": 0, + "content": "[28] Zhi-Song Liu, Wan-Chi Siu, Li-Wen Wang, Chu-Tak Li, and Marie-Paule Cani. Unsupervised real image super-resolution via generative variational autoencoder. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 442-443, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[29] W.N. Meier, F. Fetterer, A.K. Windnagel, and J.S. Stewart. Noaa/nsidc climate data record of passive microwave sea ice concentration. (g02202, version 4). National Snow and Ice Data Center, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.305 + ], + "angle": 0, + "content": "[30] Bin Mu, Xiaodan Luo, Shijin Yuan, and Xi Liang. Icetft v 1.0. 0: Interpretable long-term prediction of arctic sea ice extent with deep learning. Geoscientific Model Development Discussions, 2023:1-28, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.306, + 0.482, + 0.348 + ], + "angle": 0, + "content": "[31] J.E. Nash and J.V. Sutcliffe. River flow forecasting through conceptual models part i — a discussion of principles. Journal of Hydrology, 10(3):282-290, 1970. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[32] Xingang Pan, Xiaohang Zhan, Bo Dai, Dahua Lin, Chen Change Loy, and Ping Luo. Exploiting deep generative prior for versatile image restoration and manipulation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):7474-7489, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.482, + 0.475 + ], + "angle": 0, + "content": "[33] Zisis I Petrou and Yingli Tian. Prediction of sea ice motion with convolutional long short-term memory networks. IEEE Transactions on Geoscience and Remote Sensing, 57 (9):6865-6876, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.477, + 0.482, + 0.533 + ], + "angle": 0, + "content": "[34] Yibin Ren and Xiaofeng Li. Predicting the daily sea ice concentration on a subseasonal scale of the pan-arctic during the melting season by a deep learning model. IEEE Transactions on Geoscience and Remote Sensing, 61:1-15, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.534, + 0.482, + 0.588 + ], + "angle": 0, + "content": "[35] Y Ren, X Li, and W Zhang. A data-driven deep learning model for weekly sea ice concentration prediction of the panarctic during the melting season, iee t. geosci. remote, 60, 4304819, 2022. 1, 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.591, + 0.482, + 0.659 + ], + "angle": 0, + "content": "[36] Eduardo Rocha Rodrigues, Igor Oliveira, Renato Cunha, and Marco Netto. Deepdownscale: A deep learning strategy for high-resolution weather forecast. In 2018 IEEE 14th International Conference on e-Science (e-Science), pages 415–422. IEEE, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.482, + 0.73 + ], + "angle": 0, + "content": "[37] Qi Shao, Guangchao Hou, Wei Li, Guijun Han, Kangzhuang Liang, and Yang Bai. Ocean reanalysis data-driven deep learning forecast for sea surface multivariate in the south china sea. Earth and Space Science, 8(7):e2020EA001558, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.482, + 0.8 + ], + "angle": 0, + "content": "[38] S Sparnocchia, N Pinardi, and E Demirov. Multivariate empirical orthogonal function analysis of the upper thermocline structure of the mediterranean sea from observations and model simulations. In Annales Geophysicae, pages 167-187. Copernicus GmbH, 2003. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.802, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[39] Karen Stengel, Andrew Glaws, Dylan Hettinger, and Ryan N King. Adversarial super-resolution of climatological wind and solar data. Proceedings of the National Academy of Sciences, 117(29):16805-16815, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[40] Siwei Tu, Weidong Yang, and Ben Fei. Taming generative diffusion for universal blind image restoration. arXiv preprint arXiv:2408.11287, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[41] Lei Wang, Xiaojun Yuan, and Cuihua Li. Subseasonal forecast of arctic sea ice concentration via statistical approaches. Climate Dynamics, 52:4953-4971, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[42] Qingyu Zheng, Ru Wang, Guijun Han, Wei Li, Xuan Wang, Qi Shao, Xiaobo Wu, Lige Cao, Gongfu Zhou, and Song Hu. A spatio-temporal multiscale deep learning model for subseasonal prediction of arctic sea ice. IEEE Transactions on Geoscience and Remote Sensing, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[43] Xiaohui Zhong, Lei Chen, Jun Liu, Chensen Lin, Yuan Qi, and Hao Li. Fuxi-extreme: Improving extreme rainfall and wind forecasts with diffusion model. Science China Earth Sciences, pages 1-13, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[44] Ruofan Zhou and Sabine Susstrunk. Kernel modeling superresolution on real low-resolution images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2433-2443, 2019. 3" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10576" + } + ] +] \ No newline at end of file diff --git a/2025/IceDiff_ High Resolution and High-Quality Arctic Sea Ice Forecasting with Generative Diffusion Prior/3f2f5e81-8b45-4fda-b37e-2ea2e3413d80_origin.pdf b/2025/IceDiff_ High Resolution and High-Quality Arctic Sea Ice Forecasting with Generative Diffusion Prior/3f2f5e81-8b45-4fda-b37e-2ea2e3413d80_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..bb2fe8f5e64e412c315021428d508c6b30b18957 --- /dev/null +++ b/2025/IceDiff_ High Resolution and High-Quality Arctic Sea Ice Forecasting with Generative Diffusion Prior/3f2f5e81-8b45-4fda-b37e-2ea2e3413d80_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eb1d9083d1ffc4da01a12e87d553bce2d63cb4216a6fb6e6c16c07099bc2ae00 +size 3112266 diff --git a/2025/IceDiff_ High Resolution and High-Quality Arctic Sea Ice Forecasting with Generative Diffusion Prior/full.md b/2025/IceDiff_ High Resolution and High-Quality Arctic Sea Ice Forecasting with Generative Diffusion Prior/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ec8543155b5ed32f24f296b1510139935812bdbd --- /dev/null +++ b/2025/IceDiff_ High Resolution and High-Quality Arctic Sea Ice Forecasting with Generative Diffusion Prior/full.md @@ -0,0 +1,260 @@ +# IceDiff: High Resolution and High-Quality Arctic Sea Ice Forecasting with Generative Diffusion Prior + +Jingyi Xu $^{1,\ast}$ , Siwei Tu $^{1,\ast}$ , Weidong Yang $^{1,\dagger}$ , Ben Fei $^{3,\dagger}$ , Shuhao Li $^{1}$ , Keyi Liu $^{1}$ , Yeqi Luo $^{1}$ , Lipeng Ma $^{1}$ , Lei Bai $^{2}$ $^{1}$ Fudan University, $^{2}$ Shanghai Artificial Intelligence Laboratory, $^{3}$ Chinese University of Hong Kong jyxu22@m.fudan.edu.cn, wdyang@fudan.edu.cn, benfei@cuhk.edu.hk, baisanshi@gmail.com * Equal Contributions, † Corresponding Authors + +# Abstract + +Variation of Arctic sea ice has significant impacts on polar ecosystems, transporting routes, coastal communities, and global climate. Tracing the change of sea ice at a finer scale is paramount for both operational applications and scientific studies. Recent pan-Arctic sea ice forecasting methods that leverage advances in artificial intelligence have made promising progress over numerical models. However, forecasting sea ice at higher resolutions is still under-explored. To bridge the gap, we propose a two-module cooperative deep learning framework, IceDiff, to forecast sea ice concentration at finer scales. IceDiff first leverages a vision transformer to generate coarse yet superior forecasting results over previous methods at a regular $25\mathrm{km}$ grid. This high-quality sea ice forecasting can be utilized as reliable guidance for the next module. Subsequently, an unconditional diffusion model pre-trained on low-resolution sea ice concentration maps is utilized for sampling down-scaled sea ice forecasting via a zero-shot guided sampling strategy and a patch-based method. For the first time, IceDiff demonstrates sea ice forecasting with a $6.25\mathrm{km}$ resolution. IceDiff extends the boundary of existing sea ice forecasting models and more importantly, its capability to generate high-resolution sea ice concentration data is vital for pragmatic usages and research. Code is available at https://github.com/EtronTech/IceDiff. + +# 1. Introduction + +Arctic sea ice is essential for global climate change and local communities. Although the exact teleconnections between mid-latitude weather and the variation of Arctic sea ice are yet to be revealed, the reduction of sea ice during the last few decades clearly has impacts that transcend the Arctic region [3, 41]. Hence, establishing skillful forecasting + +models at finer scales is paramount for both transportation in polar regions and scientific research in geoscience. + +Existing physics-based numerical models have demonstrated the ability to accurately forecast sea ice concentration (SIC, value ranges from $0\%$ to $100\%$ ) for short lead times of a few weeks. However, when forecasting at longer scales, their performance significantly decreases, and their forecasting skills are even worse than those of simple statistical models [3]. With the advent of deep learning methods, a series of works based on deep learning approaches have been committed to SIC forecasting with promising results. Notable studies cover forecasting leads range from daily to seasonal scale, e.g., IceNet [3] provides accurate sea ice extent (SIE, areas where the SIC value is greater than $15\%$ and could be covered with sea ice) forecasts for 6 months lead time. SICNet [35] and $\mathrm{SICNet}_{90}$ [34] are skillful for forecasting SIC in subsequent 7 and 90 days, respectively. + +Despite the improved forecasting skills of deep learning models over previous approaches, these models are still trained on sea ice products provided by the National Sea Ice Data Center (NSIDC) or the European Organisation for the Exploitation of Meteorological Satellites (EU-METSAT). These sea ice data products are primarily derived from passive microwave sensors with a resolution of $25\mathrm{km}$ , which is considered too coarse for sea ice modeling. The in situ observations are costly to acquire and extremely spare to provide sufficient spatial coverage over pan-Arctic region. This limitation hinders the representation of detailed information on sea ice concentration and the accuracy of ice edge [10, 16]. The same challenge is also presented in weather forecasts. Therefore, several endeavors are proposed to work on downscaling atmospheric states [5, 11, 43]. For instance, SwinRDM [11] combines an improved SwinRNN [21] model with a denoising diffusion probabilistic model (DDPM) [20] to achieve superresolution (SR) in forecasting climate variables. However, + +forecasting sea ice at finer spatial scales, which is crucial, especially for operational applications is still under-explored. + +Due to the absence of high-resolution sea ice data products and sparse in situ observations, it is not feasible to train a forecast model over pan-Arctic region at a finer scale. Two primary solutions can address this challenge: (i) Downscale sea ice maps initially and then train a forecast model using these high-resolution SIC maps. (ii) Train a forecast model at the original resolution based on the sea ice products, then downscale the forecast results to a finer scale. In the first solution, downscaling sea ice maps may introduce slight errors, which could accumulate during the forecasting stage. Additionally, training a forecast model on high-resolution SIC maps also imposes a greater computational burden. Therefore, we opt for the latter solution and introduce IceDiff, a two-module cooperative framework. IceDiff aims to provide skillful forecasts at three different temporal scales in the forecasting module (FM) and produce spatially downscaled high-quality SIC maps in the guided diffusion module (GDM). Specifically, in the first module, IceDiff leverages U-Net-based vision transformer and provides forecasts that are superior to existing SIC forecasting models. Then, in the second module, we utilize a well-trained DDPM on the original resolution from the NSIDC production as effective prior and downscale forecasted SIC maps guided by forecasts from the previous module at the original scale. Our DDPM is pre-trained on daily SIC maps with a resolution of $25\mathrm{km}$ , and it can also generalize well to weekly and monthly SIC maps. In every step $t$ of the sampling process of DDPM, the intermediate variable $x_0$ is estimated from the denoised SIC map $x_{t}$ . Then we propose a convolution kernel with estimizable parameters to simulate the up-scale degradation for the intermediate variable $x_0$ , which results in up-scaled $\tilde{x}_0$ . A distance function between the low-resolution SIC map and up-scaled $\tilde{x}_0$ is utilized to dynamically update the parameters of the convolution kernel in the reverse steps and guide the sampling of SIC map in the next time step. Furthermore, to make our IceDiff suitable for SIC maps with various resolutions, a patch-based method is also designed to downscale SIC maps at artistry resolution. We show in this work that using low-resolution SIC maps as guidance can better capture small-scale details in the generated high-resolution SIC maps. Our contributions are three folds: + +- We propose a novel SIC forecasting framework IceDiff that demonstrates skills in three different temporal scales which cover subseasonal to seasonal forecasts, specifically, 7 days, weekly average of 8 weeks and monthly average of 6 months, that are commonly investigated by the community. IceDiff achieves state-of-the-art results in all three scales compared to existing deep learning models. + +- IceDiff is devised with an estimizable convolutional kernel to simulate the up-scaling degradation and designs a patch-based method for downscaling SIC maps at arbitrary resolution with a single and fixed-resolution unconditional diffusion model. By utilizing the generative diffusion prior inherent in this diffusion model, the high-resolution SIC maps can be obtained through zero-shot sampling. Moreover, our IceDiff is capable of capturing small-scale details and generating high-resolution SIC maps. + +- Extensive experiments show that our IceDiff could provide accurate SIC forecasts and generate high-quality sea ice samples with finer details. IceDiff not only provides a new and comprehensive baseline for data-driven sea ice research but also extends the capability of existing SIC forecasting models and pushes this area of research towards more operational usage. + +# 2. Related Works + +# 2.1. Deep Learning-based Sea Ice Forecasting + +Deep learning models have drawn the attention of sea ice research communities and have been widely investigated for Arctic sea ice forecasting [1, 2, 22, 33]. As a pioneering work, ConvLSTM [24] predicts SIC within the region of Barents Sea and has showcased comparable results to the state-of-the-art numerical climate models. IceNet [3] employs the U-Net architecture for merging pre-selected ERA5 variables and SIC data to improve the accuracy of seasonal SIE prediction up to six months. The experimental results have shown that the forecasting skill of IceNet in pan-Arctic region is superior to the numerical sea ice simulation model SEAS5. MT-IceNet [1] adopts a similar approach to IceNet and adds bi-monthly data to further improve forecasting SIC at seasonal scale. Mu et al. [30] propose to merge atmospheric and oceanographic data in a weighted manner through a variable selection network and provide nine months of forecasting leads. Considering the complex ice-ocean-atmosphere interactions have already embedded in the spatiotemporal correlations of consecutive SIC maps [33, 35], Ren et al. [35] proposed SICNet that solely rely on SIC data to predict concentration for the next seven days. Similar to IceNet, SICNet is built upon U-Net architecture and added with channel-wise dependency via temporal-spatial attention module (TSAM), which is essentially a revised version of the convolutional block attention module. By fusing the output of two SICNet, i.e. one for modeling SIC and the other for extracting climate information, and adjusting the number of output channels, Ren et al. [34] extends the lead time to 90 days. Besides CNN and LSTM-based models, recent works leverage the latest advances in artificial intelligence to further improve forecasting performance. For instance, Ice + +Former [42] jointly models SIC and sea ice thickness (SIT) based on Transformer backbone. It first decomposes high dimensional spatio-temporal data into sequences via multivariate empirical orthogonal functions [37, 38]. Then an encoder-decoder Transformer is utilized to predict SIC and SIT for up to 45 days. Despite the performance gain of the aforementioned deep learning models, none of them have addressed forecasting sea ice at scales lower than $25\mathrm{km}$ . + +# 2.2. Super Resolution + +Super-resolution aims to reconstruct high-resolution (HR) images from low-resolution (LR) counterparts. Several works [27, 28] explore the use of Variational Autoencoders for image SR. These methods usually improve the restoration quality by learning the conditional distribution of high-resolution images induced by low-resolution images. Although these methods are able to achieve realistic perceptual quality, VAE-based methods are often not as satisfactory due to their limited generation capabilities. To this end, other deep learning methods, such as Convolutional neural networks (CNNs), have been widely used for superresolution tasks in various domains. CNN-based methods typically use HR images corresponding to LR images to supervise training with reconstruction loss [4, 7, 36, 44]. However, these network trained solely on reconstruction loss makes it difficult to capture detailed information in images and generate diverse image results. Therefore, Generative Adversarial Networks (GANs) are proposed to solve these challenges. Stengel et al. [39] utilized GAN in a two-step SR process to achieve downscaling of climatological wind and solar irradiance. Leinonen et al. [23] introduced a recurrent, stochastic SR GAN for down-scaling time-evolving atmospheric fields, demonstrating the potential of GANs in improving the spatial resolution of LR images in atmospheric sciences applications. These models encourage the generator to generate high-quality images that the discriminator cannot distinguish from real images. Although GAN-based methods obtain satisfactory downscaled results, the diversity could be enhanced. + +Currently, diffusion model-based methods have been more widely used due to their diversity in generating images and the ability to generate high-quality images comparable to GANs [17, 40]. For instance, Chen et al. [11] utilize SR based on the conditional model to recover the high spatial resolution and finer-scale atmospheric details, unfolding the ability of diffusion models to generate high-quality and detailed images in SR tasks. The objective of this study is to forecast SIC maps at a high resolution of $6.25\mathrm{km}$ . To achieve this, we have improved the SwinTransformer for predicting SIC at the standard resolution comparable to current methods. Additionally, we employ a diffusion-based SR model to generate high-resolution and high-quality results via zero-shot sampling. + +# 3. Methodology + +Our IceDiff framework comprises two modules as depicted in Figure 1. The forecasting module provides skillful forecasts at three different temporal scales. The guided diffusion module produces spatially down-scaled high-quality SIC maps with a pre-trained DDPM via zero-shot sampling and the patch-based method. We elaborate on design details in the following sections. + +# 3.1. Forecasting Module + +To achieve more accurate SIC forecasting at a finer resolution, a skillful forecasting module trained on original sea ice data products is primarily encouraged. In contrast to existing SIC forecasting methods that primarily concentrate on a single temporal scale [1, 3, 42], our objective is to enhance forecasting skill across various temporal scales and increase versatility. Encouraged by the success of Swin Transformer [25] in various data-driven weather forecasting methods [6, 11, 12], we propose to construct forecasting module by utilizing the U-Net architecture [8] and building on top of the Swin Transformer V2 backbone [26] with ResNet $_{TSAM}$ block [35] for better capturing both spatial and temporal features of SIC maps. As illustrated in Figure 1, the FM takes $N$ consecutive SIC maps, each of which has a shape of [448, 304], to predict subsequent maps with identical temporal length as the input at $25\mathrm{km}$ spatial scale. It consists of encoder, bottleneck and decoder. The patch embedding layer first applies a $2\times 2$ windows size to partition the original SIC input into patches and maps them into tokens via 2D convolution. Then, the temporal and spatial dependencies are modeled through the ResNet $_{TSAM}$ block. Both the encoder and the decoder have 3 layers of Swin Transformers that comprise a pre-configured even number of Swin Transformer Blocks for feature extraction and restoration. The calculation of two consecutive blocks can be described as follows: + +$$ +\begin{array}{l} z _ {s} ^ {b} = L N (W M S A (z ^ {b - 1})) + z ^ {b - 1}, \\ z ^ {b} = L N (M L P (z _ {s} ^ {b})) + z _ {s} ^ {b}, \\ z _ {s} ^ {b + 1} = L N (S W M S A (z ^ {b})) + z ^ {b}, \\ z ^ {b + 1} = L N \left(M L P \left(z _ {s} ^ {b + 1}\right)\right) + z _ {s} ^ {b}, \tag {1} \\ \end{array} +$$ + +where $z_{s}^{b}$ and $z^{b}$ represents the output feature of the (Shifted) Window-Multi-head Self Attention module and the MLP module for block $b$ , respectively [26]. LN denotes the LayerNorm operation. The combination of such two consecutive blocks, i.e. the former for the employment of WMSA and the latter for SWMSA, ensures that local patterns of adjacent grids are sufficiently attended by the backbone network. We further adopt a shift window size of $28 \times 19$ that is consistent with the ratio of SIC input map for facilitating the SWMHA mechanism. The output of the + +![](images/63d1bfeb0826992771755d58394090e093f17358cb73eb1f8d63684a80886259.jpg) +Figure 1. Overview of IceDiff, a two-module cooperative framework to forecast sea ice concentration at a finer scale. IceDiff first employs the forecasting module (FM) that takes history SIC data as input to generate subsequent predictions at $25\mathrm{km}$ resolution. $N$ denotes the number of sequential time stamps, for instance, set $N$ to 7 for seven days of forecasting leads, FM takes a tensor with the shape of $\{7,448,304\}$ as input and forecasts the SIC for the next 7 days. $C$ represents the quantity of embedded channels within each layer of FM, the original SIC map data has a channel of 1 rather than 3 for regular images. The guided diffusion module (GDM) then leverages accurate forecasts as guidance to generate downscaled samples at $6.25\mathrm{km}$ resolution with high quality, see details in Figure 2. + +last block within each layer is delivered to a ResNet $_{TSAM}$ block for establishing channel-wise dependencies. Before the subsequent encoder and decoder layers, patch merging and patch expanding [8] are respectively performed to reshape the feature and fit it to the desired size. Following prior works [3, 8, 35], the skip connection is added between the corresponding encoder and decoder layers to improve the performance. The final prediction is generated via a patch expanding layer with ResNet $_{TSAM}$ that up-samples the decoded feature to the size of the original SIC input. Our meticulous designs enable the forecasting module in IceDiff to be applied across various temporal scales, enhancing forecasting skills, which serves as a more effective initiation for the guided diffusion module of IceDiff. + +# 3.2. Guided Diffusion Module + +Off-the-shelf methods are limited to the forecasting of SIC maps at a scale of $25\mathrm{km}$ from sea ice data products [1, 3, 35]. However, finer-scale SIC maps remain unexplored which is of significant value to ecosystems, transporting routes, coastal communities, and global climate. Therefore, after obtaining skillful forecasts, we exploit a downscaling module based on an unconditional diffusion model pre-trained on daily SIC maps to generate realistic high-resolution SIC maps with rich details. This guided diffusion module pre-trained on daily SIC maps could generalize well to SIC maps at various scales, such as weekly or monthly SIC maps. Moreover, our guided diffusion module + +of IceDiff downscales the SIC forecasts from a forecasting module to a scale of $6.25\mathrm{km}$ , further expanding the accuracy and practicality of sea ice prediction. + +Specifically, the reverse process of the diffusion model is conditioned on the low-resolution sea ice map $y$ , which transforms distribution $p_{\theta}(x_{t-1}|x_t)$ into conditional distribution $p_{\theta}(x_{t-1}|x_t,y)$ . Previous work [15] has derived the conditional transformation formula in the reverse process: + +$$ +\begin{array}{l} \log p _ {\theta} \left(x _ {t} \mid x _ {t + 1}, y\right) = \log \left(p _ {\theta} \left(x _ {t} \mid x _ {t + 1}\right) p \left(y \mid x _ {t}\right)\right) + N _ {1} (2) \\ \approx \log p _ {\theta} (z) + N _ {2}, z \sim \mathcal {N} (z; \mu_ {\theta} \left(x _ {t}, t\right) + \Sigma g, \Sigma I), (3) \\ \end{array} +$$ + +where $g$ refers to $\nabla_{x_t}\log p(y|x_t)$ . $N_{1} = \log p_{\theta}(y|x_{t + 1})$ , $N_{2}$ is a constant related to $g$ (proof in Appendix). The variance of the reverse process $\Sigma = \Sigma_{\theta}(x_{t})$ is constant. + +Therefore, the sampling process of $p_{\theta}(x_{t - 1}|x_t,y)$ integrates the gradient term $g$ with the diffusion model $(\mu ,\Sigma) = (\mu_{\theta}(x_{t}),\Sigma_{\theta}(x_{t}))$ , which plays a role in controlling the direction of map generation. We use a heuristic algorithm (proof in Appendix) to approximate the value of $g$ : + +$$ +\log p (y \mid x _ {t}) = - \log N - s \mathcal {L} (\mathcal {D} (\tilde {x} _ {0}), y) \tag {4} +$$ + +$$ +g = \nabla_ {x _ {t}} \log p (y | x _ {t}) = - s \nabla_ {x _ {t}} \mathcal {L} (\mathcal {D} (x _ {t}), y). \tag {5} +$$ + +Among them, $s$ is the scaling factor used to control the degree of guidance. And $\mathcal{L}$ is a distance function used to measure the distance deviation between two SIC maps. $\mathcal{D}$ is an up-scale convolution function with estimizable parameters, where the parameters are randomly initialized and dynamically optimized with the gradient of distance function at + +![](images/fd5f3d5c841bca26f4e9ae57dd1c219d066d56fc92b28951abff3474d14c4407.jpg) +(a) +(b) +(c) + +![](images/f8f320287e357e63b912bd33dc4cbccc8d5dacdbf994dec8df6e639a57b2e7f9.jpg) +Figure 2. Overview of guided diffusion module in IceDiff for SIC map downscaling. (a) Pre-trained diffusion model is used to estimate $\tilde{x}_0$ at reverse step $t$ . The estimable convolutional kernel of step $t$ reduces the resolution of $X_{t}$ so that guidance from low-resolution SIC maps can be incorporated to sample $X_{t-1}$ . (b) Low-resolution SIC map $Y$ is integrated during the sampling process. A distance function describes the distance between a low-resolution SIC map and the generated high-resolution SIC map $x$ after convolution. The gradient of distance function could be used to guide the sampling of $x_{t-1}$ and (c) dynamically update the parameters of the convolutional kernel. +Figure 3. Illustration of the patch-based method. We set patch size $p = 256$ and stride $r = 128$ to fit the size of pre-trained models. This setting will divide the example map into $K = 4$ sub-maps boxed with different colors. For each sub-map, an individualizable convolution kernel with adaptive parameters is set to add guidance from the corresponding low-resolution SIC sub-map. At each sampling step $t$ , the average estimated noise on four overlapping blocks will be utilized to update the mean $\mu$ used to sample $x_{t-1}$ . + +every reverse step. $\mathcal{D}$ connects two different resolution SIC maps, allowing low-resolution SIC maps to provide more effective guidance. + +As shown in Algorithm 1, GDM undergoes $T$ reverse steps to gradually restore pure Gaussian noise $x_{T} \sim \mathcal{N}(0, I)$ to high-resolution SIC maps. In each reverse step $t$ , the diffusion model calculates the estimated variable $\tilde{x}_{0}$ which then undergoes the up-scale convolution function $\mathcal{D}$ at time $t$ to establish a distance metric $\mathcal{L}$ with the low- + +resolution SIC map, where $\mathcal{D}_t^\varphi$ signifies that its convolutional kernel parameters are $\varphi = \varphi_t$ at the reverse steps $t$ . The $\mathcal{L}$ subscript $\varphi = \varphi_t$ indicates that it is a function dependent on the convolutional kernel parameter $\varphi$ at reverse step $t$ . The gradients of distance function $\mathcal{L}$ with respect to $\tilde{x}_0$ and the convolution kernel parameters of the up-scale convolution function are used to update $\tilde{x}_0$ and the convolution kernel parameters, respectively. The update of convolution kernel parameters ensures the accuracy of the model's + +Table 1. Quantitative results of forecasts. IceDiff is evaluated on three different temporal scales. The output length of IceDiff equals a pre-defined forecasting lead as depicted in Figure 1. The validated accuracy of IceDiff could provide reliable guidance for the guided diffusion module to generate high-quality downscaled SIC. BT represents that the SIC data [14] are generated by the Bootstrap retrieval algorithm [13] and the data distribution of CDR is considered to be closer to BT rather than NT [29]. + +
Temporal ScaleLead TimeDatasetMethodsRMSE↓MAE↓R²↑NSE↑IIEE↓SIEdif↓
Subseasonal7 Days (Daily)NTSICNet [35]0.04300.00910.9780.97511750.1127
NTIceDiff0.03240.00770.9870.9868990.0186
CDRSICNet [35]0.04900.01000.9820.9799760.0718
CDRIceDiff0.03960.00800.9890.9878350.0315
45 Days (Daily)BTIceFormer [42]0.06600.02010.960---
8 Weeks Average (Weekly)NTIceDiff0.04030.01080.9750.97211700.1051
CDRIceDiff0.05530.01120.9730.96913530.0871
90 Days (Daily)NTSICNet90 [34]-0.0512----
Seasonal6 Months Average (Monthly)NTIceNet [3]0.18200.09160.567---
NTMT-IceNet [1]0.07770.01970.915---
NTIceDiff0.05730.01880.9580.95220140.2419
CDRIceDiff0.06480.01680.9190.91320160.2657
+ +Algorithm 1 Guided diffusion module in IceDiff with the guidance of low-resolution SIC map $y$ . An unconditional diffusion model pre-trained on SIC maps $\epsilon_{\theta}(x_t,t)$ is given. + +Input: Low-resolution SIC map $y$ , downscale convolution function $\mathcal{D}$ composed of optimized convolutional kernels $\mathcal{K}$ with parameters $\varphi$ , learning rate $l$ , guidance scale $s$ , distant function $\mathcal{L}$ . + +Output: Output high-resolution SIC map $x_0$ conditioned on $y$ . Sample $x_T$ from $\mathcal{N}(0, I)$ + +1: for all t from T to 1 do +2: $\tilde{x}_0 = \frac{x_t}{\sqrt{\alpha_t}} - \frac{\sqrt{1 - \alpha_t} \epsilon_0(x_t, t)}{\sqrt{\alpha_t}}$ +3: $\mathcal{L}_{\varphi, \tilde{x}_0} = \mathcal{L}(y, \mathcal{D}^\varphi(\tilde{x}_0))$ +4: $\tilde{x}_0\gets \tilde{x}_0 - \frac{s(1 - \bar{\alpha}_t)}{\sqrt{\bar{\alpha}_{t - 1}\beta_t}}\nabla_{\tilde{x}_0}\mathcal{L}_{\varphi ,\tilde{x}_0}$ +5: $\tilde{\mu}_{t} = \frac{\sqrt{\bar{\alpha}_{t - 1}}\beta_{t}}{1 - \bar{\alpha}_{t}}\tilde{x}_{0} + \frac{\sqrt{\bar{\alpha}_{t}}(1 - \bar{\alpha}_{t - 1})}{1 - \bar{\alpha}_{t}} x_{t}$ +6: $\tilde{\beta}_{t} = \frac{1 - \bar{\alpha}_{t - 1}}{1 - \bar{\alpha}_{t}}\beta_{t}$ +7: Sample $x_{t-1}$ from $\mathcal{N}(\tilde{\mu}_t, \tilde{\beta}_t I)$ +8: $\varphi \gets \varphi - l \nabla_{\varphi} \mathcal{L}_{\varphi, \bar{x}_0}$ +9: end for +10: return $x_0$ + +up-scaling process, making guidance of low-resolution SIC maps more precise and effective. + +Patch-based Strategy. Furthermore, pre-trained models can only generate fixed-size SIC maps. To tackle the problem of sampling downscaled SIC forecasts at any resolution, we employ a patch-based strategy. As shown in Figure 3, the patch-based strategy partitions SIC maps into several sub-regions based on a fixed stride. For each individual sub-region, the gradient terms of distance loss between SIC maps upscaled by the convolutional kernel and the corresponding low-resolution SIC sub-region maps are computed individually. Subsequently, the respective mean of Gaussian distribution in sub-regions (Eq. 3) and convolution kernel parameters are updated based on the gradient terms. The entire map is updated by taking the mean of + +Table 2. Quantitative comparison between IceDiff for SIC downscaling and other methods. Const. and FID represents the Consistency and Fréchet Inception Distance metrics, respectively. + +
Method7 Days8 Week Avg.6 Month Avg.
Const.↓FID↓Const.↓FID↓Const.↓FID↓
Nearest20.0178.4521.9880.1723.0788.36
Bilinear17.8870.7520.0277.2918.3180.73
BiCubic13.1354.7113.4068.8510.0970.18
DGP [32]23.1875.4228.6281.4229.7485.47
GDP [17]8.2640.1510.5452.4010.9655.72
IceDiff6.8429.347.2531.296.6239.79
+ +the updated values across all sub-regions, each weighted by a binary patch mask that serves as a quantifier of the regional scope, thereby refining the overall mean and convolution kernel parameters of the sampled high-resolution SIC maps. By leveraging this strategy, our model is capable of downscaling sea ice maps to the desired resolution and the practicality of the model has been further enhanced. + +# 4. Experiments + +# 4.1. Dataset and Baseline Setup + +We evaluate our proposed IceDiff framework on the Climate Data Record of Passive Microwave Sea Ice Concentration Version 4 dataset from NSIDC [29] (CDR in Table 1). It records daily SIC data starting from October $25^{th}$ 1978 and provides the coverage of the pan-Arctic region $(\mathrm{N} : -39.36^{\circ}, \mathrm{S} : -89.84^{\circ}, \mathrm{E} : 180^{\circ}, \mathrm{W} : -180^{\circ})$ . The SIC map is formed of $448 \times 304$ pixels and each of which represents an area of $25 \mathrm{~km} \times 25 \mathrm{~km}$ grid and has a value ranging from $0\%$ to $100\%$ . We choose data from October $25^{th}$ 1978 to 2013 as the training dataset, the years 2014 and 2015 are selected as validation set, and data from 2016 to 2023 are used for testing. To comprehensively validate and verify the performance of our forecasting module, we utilize additional data [16] that are generated through the retrieval + +Table 3. Ablation study on TSAM. We compare the forecasting results between IceDiff with or without the TSAM. + +
Lead TimeTSAMRMSE↓MAE↓R²↑NSE↑IIEE↓SIEdif↓
7 DaysX0.04030.00820.98810.98608740.0336
✓0.03960.00800.98850.98658350.0315
8 Week Avg.X0.05350.01160.9750.97113860.1420
✓0.05530.01120.9730.96913530.0871
6 Month Avg.X0.07630.01940.8830.87526100.4919
✓0.06480.01680.9190.91320160.2657
+ +algorithm (NASA Team [9], NT in Table 1). The forecasting module is trained on three different temporal scales, i.e. the input of SIC maps are 7 days, 8 weeks average, and 6 months average, covering forecasting leads from subseasonal to seasonal scale. Corresponding deep learning-based models are selected for comparison, as illustrated in Table 1. Note that we follow the split of datasets in respective baselines to preprocess the additional data. We re-implemented SICNet (the source code is not available) based on the original paper and trained it under an environment identical to our forecasting module. To evaluate the performance of our downscaling, five methods are selected to set up baselines for downscaling quality comparison. + +# 4.2. Implementation Details + +We independently trained the forecasting module of IceDiff on SIC dataset of three aforementioned temporal scales for 60 epochs. For each temporal scale, the forecasting module outputs the same length of input data as forecasting leads, i.e. 7 days, 8 weeks average, and 6 months averages. The dimensions of embedding features of encoder, decoder, and bottleneck are all set to 32. We utilize the same hyperparameter of TSAM as in [35]. The forecasting module is optimized by AdamW under the mean square error (MSE) objective using Pytorch on one NVIDIA A100 80GB GPU for all experiments with a learning rate of 0.0005. All daily SIC maps are resized and cropped to $256 \times 256$ to train our GDM of IceDiff. using AdamW with $\beta_{1} = 0.9$ and $\beta_{2} = 0.999$ . We train in 16-bit precision using loss-scaling, but maintain 32-bit weights, EMA, and optimizer state. We use an EMA rate of 0.9999. When sampling with 1000 timesteps, we use the same noise schedule as for training. + +# 4.3. Evaluation Metrics + +To quantitatively evaluate our IceDff, we select commonly used root mean square error (RMSE) and mean absolute error (MAE) for the comparison of forecasting accuracy. $R^2$ score and Nash-Sutcliffe Efficiency [31] are utilized to analyze the similarity of spatial pattern between the ground truth and predicted SIC map. We adopt Integrated Ice-Edge Error score [18] and calculate the sea ice area differences $(SIE_{dif})$ , in millions of $km^2$ to further evaluate the prediction of SIE (where the SIC value is greater than 15%). + +As to the evaluation of downscaling quality, following [11] we utilize the Fréchet Inception Distance (FID) + +Table 4. Ablation study on the ratio of Input/Forecast number of SIC maps. Hf., Db. and Eq. represent that the forecasting module takes input data which has half, double, and equal length of forecasting leads, respectively. For the Hf. of 7 Days, we set the input length to 3. + +
Lead TimeI/F.RMSE↓MAE↓R²↑NSE↑IIEE↓SIEdif↓
7 DaysHf.0.04160.00840.9870.9859050.0538
Db.0.04010.00840.9880.9868890.0599
Eq.0.03960.00800.9890.9878350.0315
8 Week Avg.Hf.0.05800.01180.9700.96614050.1176
Db.0.05940.01210.9690.96414350.1221
Eq.0.05530.01120.9730.96913530.0871
6 Month Avg.Hf.0.09040.02320.8420.83031920.6289
Db.0.07220.01790.8970.88920270.2795
Eq.0.06480.01680.9190.91320160.2657
+ +[19] for measuring the downscaling fidelity. In addition, we adopt Consistency to quantify the faithfulness between the downscaled map and the low-resolution map. + +The detailed equations for calculating the abovementioned metrics can be found in Appendix. + +# 4.4. Main Results + +In this subsection, we evaluate the proposed IceDiff with baseline methods for validation of the effectiveness. + +Set up new baselines for SIC prediction. The forecasting module of IceDiff is trained independently to predict SIC at 7 days, 8 weeks average, and 6 months average. As shown in Table 1, the performance of our proposed forecasting module of IceDiff is superior to SICNet in all evaluation metrics. The forecasting skill of FM across different SIC datasets is consistent in three temporal scales. Specifically, it outperforms SOTA seasonal forecasting baselines by a noticeable margin. Considering the overall performance of our FM, it could provide reliable SIC forecasts and sufficiently support the subsequent downscaling module. + +Analysis of the end of the melting season. To further verify the performance of IceDiff in extreme cases, we select two years of the end of the melting season (early September) when the mean Arctic SIC typically reaches its minimum for validation: (i) In 2016, it recorded historically low annual SIC; (ii) In 2022, the annual SIC has an abrupt increase compared to the previous year (annual trends are plotted in the Appendix). We present the SIC residuals (predicted SIC minus the ground truth) of September $1^{st}$ 2016 and 2022 in Figure 4(a) and Figure 4(b), respectively. The overall residuals are relatively low and the prediction errors mainly appear in part of the sea ice edge regions. The accurate forecasting results in temporal periods with highly varying SIC validate the effectiveness of IceDiff. + +High downscaling quality. We systematically compare the results of IceDiff with other downscaling methods. In terms of interpolation-based methods, we select nearest, bilinear, and bicubic interpolation to compare with our diffusion-based GDM. Besides, we select DGP [32] as a GAN-based method and GDP [17] as a DDPM-based method for fur + +![](images/ef9069d224d6630c2edfcec6deb52788010afeaca97aaac2ba5bdfbea255ab88.jpg) +(a) +Forecast Date Sep 1,2016 + +![](images/9e9a75ac043e9c73524086571625a79fb6b4b907dc4ce8dbc6aa3f8dbf1c3068.jpg) +(b) +Forecast Date Sep 1,2022 +Figure 4. Qualitative analysis of the end of the melting season. The annual mean Arctic SIC has recorded a historical low in 2016 and it drastically increased in 2022 when compared to the previous annual mean. The residual plots of early September in (a)2016 and (b)2022 demonstrate where the FM overestimates (positive figures) or underestimates (negative figures) SIC. + +ther comparison. Interpolation-based methods upsample the low-resolution sea ice map, which uses existing pixel information to estimate missing pixel values. + +As shown in Figure 5, our GDM produces high-quality results and small-scale structures are successfully captured and rich details are generated through the guidance from the input map during each sampling step. By contrast, interpolation-based methods do not integrate details information of the low-resolution map, resulting in blurry edges and contours in the generated results. More quality analysis can be found in the appendix. + +To further validate the superiority of our method, we follow SwinRDM [11] and GDP [17] to utilize FID and Consistency to evaluate our method, respectively. As shown in Table 2, compared to interpolation-based methods, our IceDiff improves FID score by 31.11 on average, reflecting it has the capacity to help generate a more realistic high-resolution map with a satisfactory texture restoration effect. A lower Consistency demonstrates the structural difference between the downscaled map and the original SIC input is relatively imperceptible, which better preserves the spatial distribution of sea ice forecast. + +# 4.5. Ablation Studies on Design of FM + +As shown in Table 3, by adopting TSAM to FM, the performance has improved for all metrics at 7 days and 6 months average lead times while the effects for 8 weeks average lead time are not all positive. Overall, incorporating TSAM to establish channel-wise dependency is favorable for SIC forecast models. To examine the impacts of input SIC data length, we trained FM for input lengths of half, double, and equal to the forecasting leads. The results in Table 4 show + +![](images/03beacc4b8a477ee609539113f2f0d683f334a853d24a075ac30feb8c614c4e1.jpg) +(a) IceDiff +Figure 5. Qualitative analysis of downscaled SIC map. We visualize the SIC map by coloring the land and the ocean in yellow and blue, respectively. Since SIC has a range of $[0,1]$ , we map it to $[0,255]$ and replicate it for all 3 RGB channels. The darker the pixels appear, the SIC values are more closer to $0\%$ . For better identification of the difference, we crop out a region and plot a grid size of $250\mathrm{km}$ and further display details on $4\times 4$ grid. + +![](images/e4875f857608b86f8a75df81d4c7333796c8907f80f19cb21b882bf8ade52022.jpg) +(b) Bicubic + +![](images/263520a0bc76ba2995ade0fcdc342302b2e7f246b2ee2dcb577fa8a485f9c13c.jpg) +(c) GDP + +the peak of the performance at the ratio where the input length equals the forecasting leads. + +# 5. Conclusion and Limitations + +We propose a novel pan-Arctic high-resolution SIC forecasting framework, IceDiff, which comprises a forecasting module and a guided diffusion module, these two modules closely cooperate with each other to fulfill accurate forecasting of SIC at a finer and more practical spatial scale. The forecasting module of IceDiff is capable of producing accurate forecasts at three different temporal scales and the guided diffusion module leverages reliable forecasts as guidance to generate high-quality down-scaled SIC maps. The forecasting skills of IceDiff at 7 days lead are superior to SOTA deep learning-based models. IceDiff also sets up a new baseline for weekly average forecasting of 8 weeks and provides competitive results in seasonal SIC forecasts. To the best of our knowledge, IceDiff is the first attempt to forecast SIC at $6.25\mathrm{km}$ spatial scale which is a quarter of the grid size $(25\mathrm{km})$ that current methods perform on. The experimental results show that our framework can generate down-scaled SIC superior to off-the-shelf methods and it is vital for pushing forecasting models towards operational usage and further promoting sea ice research. + +Existing sea ice concentration data products are all retrieved from satellite observations with spatial resolution limited to $25\mathrm{km}$ . A finer-scale ground truth is missing compared with observations from stations in weather forecasting. Therefore, our limitation lies in the evaluations of IceDiff on these in situ observations. We will focus on evaluating our IceDiff on higher-resolution SIC observations. + +# References + +[1] Sahara Ali and Jianwu Wang. Mt-icenet-a spatial and multi-temporal deep learning model for arctic sea ice forecasting. In 2022 IEEE/ACM International Conference on Big Data Computing, Applications and Technologies (BDCAT), pages 1–10. IEEE, 2022. 2, 3, 4, 6 +[2] Sahara Ali, Yiyi Huang, Xin Huang, and Jianwu Wang. Sea ice forecasting using attention-based ensemble LSTM. arXiv preprint arXiv:2108.00853, 2021. 2 +[3] Tom R Andersson, J Scott Hosking, María Pérez-Ortiz, Brooks Paige, Andrew Elliott, Chris Russell, Stephen Law, Daniel C Jones, Jeremy Wilkinson, Tony Phillips, et al. Seasonal arctic sea ice forecasting with probabilistic deep learning. Nature communications, 12(1):5124, 2021. 1, 2, 3, 4, 6 +[4] Saeed Anwar, Salman Khan, and Nick Barnes. A deep journey into super-resolution: A survey. ACM Computing Surveys (CSUR), 53(3):1-34, 2020. 3 +[5] Seth Bassetti, Brian Hutchinson, Claudia Tebaldi, and Ben Kravitz. Diffesm: Conditional emulation of temperature and precipitation in earth system models with 3d diffusion models. Journal of Advances in Modeling Earth Systems, 16(10): e2023MS004194, 2024. 1 +[6] Kaifeng Bi, Lingxi Xie, Hengheng Zhang, Xin Chen, Xiaotao Gu, and Qi Tian. Accurate medium-range global weather forecasting with 3d neural networks. Nature, 619(7970): 533-538, 2023. 3 +[7] Bruno Buongiorno Nardelli, Davide Cavaliere, Elodie Charles, and Daniele Ciani. Super-resolving ocean dynamics from space with computer vision algorithms. Remote Sensing, 14(5):1159, 2022. 3 +[8] Hu Cao, Yueyue Wang, Joy Chen, Dongsheng Jiang, Xiaopeng Zhang, Qi Tian, and Manning Wang. Swin-unet: Unet-like pure transformer for medical image segmentation. In European conference on computer vision, pages 205-218. Springer, 2022. 3, 4 +[9] D.J. Cavalieri, P. Gloersen, and W.J. Campbell. Determination of sea ice parameters with the nimbus-7 smmr. Journal of Geophysical Research, 89:5355-5369, 1984. 7 +[10] D.J. Cavalieri, P. Gloersen, and W.J. Campbell. Deriving long-term time series of sea ice cover from satellite passive-microwave multisensor data sets. Journal of Geophysical Research, 104:15803-15814, 1999. 1 +[11] Lei Chen, Fei Du, Yuan Hu, Zhibin Wang, and Fan Wang. Swinrdm: integrate swirrnn with diffusion model towards high-resolution and high-quality weather forecasting. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 322-330, 2023. 1, 3, 7, 8 +[12] Lei Chen, Xiaohui Zhong, Feng Zhang, Yuan Cheng, Yinghui Xu, Yuan Qi, and Hao Li. Fuxi: A cascade machine learning forecasting system for 15-day global weather forecast. npj Climate and Atmospheric Science, 6(1):190, 2023. 3 +[13] J.C. Comiso. Characteristics of arctic winter sea ice from satellite multispectral microwave observations. Journal of Geophysical Research, 91:975-994, 1986. 6 + +[14] Josefino C. Comiso. Bootstrap sea ice concentrations from nimbus-7 smmr and dmsp ssm/i-ssmis, version 3. 2017. 6 +[15] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 4 +[16] N. DiGirolamo, C.L. Parkinson, D.J. Cavalieri, P. Gloersen, and H.J. Zwally. Sea ice concentrations from nimbus-7 smmr //and dmsp ssm/i-ssmis passive microwave data. (nsidc-0051, version 1). National Snow and Ice Data Center Distributed Active Archive Center., 1996. 1, 6 +[17] Ben Fei, Zhaoyang Lyu, Liang Pan, Junzhe Zhang, Weidong Yang, Tianyue Luo, Bo Zhang, and Bo Dai. Generative diffusion prior for unified image restoration and enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9935-9946, 2023. 3, 6, 7, 8 +[18] Helge F Goessling, Steffen Tietsche, Jonathan J Day, Ed Hawkins, and Thomas Jung. Predictability of the arctic sea ice edge. Geophysical Research Letters, 43(4):1642-1650, 2016. 7 +[19] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 7 +[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1 +[21] Yuan Hu, Lei Chen, Zhibin Wang, and Hao Li. Swinvrnn: A data-driven ensemble forecasting model via learned distribution perturbation. Journal of Advances in Modeling Earth Systems, 15(2):e2022MS003211, 2023. 1 +[22] Young Jun Kim, Hyun-Cheol Kim, Daehyeon Han, Sanggyun Lee, and Jungho Im. Prediction of monthly arctic sea ice concentrations using satellite and reanalysis data based on convolutional neural networks. The Cryosphere, 14(3): 1083-1104, 2020. 2 +[23] Jussi Leinonen, Daniele Nerini, and Alexis Berne. Stochastic super-resolution for downscaling time-evolving atmospheric fields with a generative adversarial network. IEEE Transactions on Geoscience and Remote Sensing, 59(9):7211-7223, 2020. 3 +[24] Yang Liu, Laurens Bogaardt, Jisk Attema, and Wilco Hazeleger. Extended-range arctic sea ice forecast with convolutional long short-term memory networks. Monthly Weather Review, 149(6):1673-1693, 2021. 2 +[25] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 3 +[26] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12009-12019, 2022. 3 +[27] Zhi-Song Liu, Wan-Chi Siu, and Yui-Lam Chan. Photorealistic image super-resolution via variational autoencoders. + +IEEE Transactions on Circuits and Systems for video Technology, 31(4):1351-1365, 2020. 3 +[28] Zhi-Song Liu, Wan-Chi Siu, Li-Wen Wang, Chu-Tak Li, and Marie-Paule Cani. Unsupervised real image super-resolution via generative variational autoencoder. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 442-443, 2020. 3 +[29] W.N. Meier, F. Fetterer, A.K. Windnagel, and J.S. Stewart. Noaa/nsidc climate data record of passive microwave sea ice concentration. (g02202, version 4). National Snow and Ice Data Center, 2021. 6 +[30] Bin Mu, Xiaodan Luo, Shijin Yuan, and Xi Liang. Icetft v 1.0. 0: Interpretable long-term prediction of arctic sea ice extent with deep learning. Geoscientific Model Development Discussions, 2023:1-28, 2023. 2 +[31] J.E. Nash and J.V. Sutcliffe. River flow forecasting through conceptual models part i — a discussion of principles. Journal of Hydrology, 10(3):282-290, 1970. 7 +[32] Xingang Pan, Xiaohang Zhan, Bo Dai, Dahua Lin, Chen Change Loy, and Ping Luo. Exploiting deep generative prior for versatile image restoration and manipulation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):7474-7489, 2021. 6, 7 +[33] Zisis I Petrou and Yingli Tian. Prediction of sea ice motion with convolutional long short-term memory networks. IEEE Transactions on Geoscience and Remote Sensing, 57 (9):6865-6876, 2019. 2 +[34] Yibin Ren and Xiaofeng Li. Predicting the daily sea ice concentration on a subseasonal scale of the pan-arctic during the melting season by a deep learning model. IEEE Transactions on Geoscience and Remote Sensing, 61:1-15, 2023. 1, 2, 6 +[35] Y Ren, X Li, and W Zhang. A data-driven deep learning model for weekly sea ice concentration prediction of the panarctic during the melting season, iee t. geosci. remote, 60, 4304819, 2022. 1, 2, 3, 4, 6, 7 +[36] Eduardo Rocha Rodrigues, Igor Oliveira, Renato Cunha, and Marco Netto. Deepdownscale: A deep learning strategy for high-resolution weather forecast. In 2018 IEEE 14th International Conference on e-Science (e-Science), pages 415–422. IEEE, 2018. 3 +[37] Qi Shao, Guangchao Hou, Wei Li, Guijun Han, Kangzhuang Liang, and Yang Bai. Ocean reanalysis data-driven deep learning forecast for sea surface multivariate in the south china sea. Earth and Space Science, 8(7):e2020EA001558, 2021. 3 +[38] S Sparnocchia, N Pinardi, and E Demirov. Multivariate empirical orthogonal function analysis of the upper thermocline structure of the mediterranean sea from observations and model simulations. In Annales Geophysicae, pages 167-187. Copernicus GmbH, 2003. 3 +[39] Karen Stengel, Andrew Glaws, Dylan Hettinger, and Ryan N King. Adversarial super-resolution of climatological wind and solar data. Proceedings of the National Academy of Sciences, 117(29):16805-16815, 2020. 3 +[40] Siwei Tu, Weidong Yang, and Ben Fei. Taming generative diffusion for universal blind image restoration. arXiv preprint arXiv:2408.11287, 2024. 3 + +[41] Lei Wang, Xiaojun Yuan, and Cuihua Li. Subseasonal forecast of arctic sea ice concentration via statistical approaches. Climate Dynamics, 52:4953-4971, 2019. 1 +[42] Qingyu Zheng, Ru Wang, Guijun Han, Wei Li, Xuan Wang, Qi Shao, Xiaobo Wu, Lige Cao, Gongfu Zhou, and Song Hu. A spatio-temporal multiscale deep learning model for subseasonal prediction of arctic sea ice. IEEE Transactions on Geoscience and Remote Sensing, 2024. 3, 6 +[43] Xiaohui Zhong, Lei Chen, Jun Liu, Chensen Lin, Yuan Qi, and Hao Li. Fuxi-extreme: Improving extreme rainfall and wind forecasts with diffusion model. Science China Earth Sciences, pages 1-13, 2024. 1 +[44] Ruofan Zhou and Sabine Susstrunk. Kernel modeling superresolution on real low-resolution images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2433-2443, 2019. 3 \ No newline at end of file diff --git a/2025/IceDiff_ High Resolution and High-Quality Arctic Sea Ice Forecasting with Generative Diffusion Prior/images.zip b/2025/IceDiff_ High Resolution and High-Quality Arctic Sea Ice Forecasting with Generative Diffusion Prior/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..cf7877af87d2bf545dc577e5b3f02ed0f6ec9d04 --- /dev/null +++ b/2025/IceDiff_ High Resolution and High-Quality Arctic Sea Ice Forecasting with Generative Diffusion Prior/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:088267f3821800dab3419807d835b67ee442082fdf6179c6869612e45f08b1e7 +size 539491 diff --git a/2025/IceDiff_ High Resolution and High-Quality Arctic Sea Ice Forecasting with Generative Diffusion Prior/layout.json b/2025/IceDiff_ High Resolution and High-Quality Arctic Sea Ice Forecasting with Generative Diffusion Prior/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2d1a3f2497847649ddbb0bf78ad9851457563c78 --- /dev/null +++ b/2025/IceDiff_ High Resolution and High-Quality Arctic Sea Ice Forecasting with Generative Diffusion Prior/layout.json @@ -0,0 +1,8661 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 75, + 102, + 536, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 102, + 536, + 138 + ], + "spans": [ + { + "bbox": [ + 75, + 102, + 536, + 138 + ], + "type": "text", + "content": "IceDiff: High Resolution and High-Quality Arctic Sea Ice Forecasting with Generative Diffusion Prior" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 65, + 160, + 545, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 160, + 545, + 246 + ], + "spans": [ + { + "bbox": [ + 65, + 160, + 545, + 246 + ], + "type": "text", + "content": "Jingyi Xu" + }, + { + "bbox": [ + 65, + 160, + 545, + 246 + ], + "type": "inline_equation", + "content": "^{1,\\ast}" + }, + { + "bbox": [ + 65, + 160, + 545, + 246 + ], + "type": "text", + "content": ", Siwei Tu" + }, + { + "bbox": [ + 65, + 160, + 545, + 246 + ], + "type": "inline_equation", + "content": "^{1,\\ast}" + }, + { + "bbox": [ + 65, + 160, + 545, + 246 + ], + "type": "text", + "content": ", Weidong Yang" + }, + { + "bbox": [ + 65, + 160, + 545, + 246 + ], + "type": "inline_equation", + "content": "^{1,\\dagger}" + }, + { + "bbox": [ + 65, + 160, + 545, + 246 + ], + "type": "text", + "content": ", Ben Fei" + }, + { + "bbox": [ + 65, + 160, + 545, + 246 + ], + "type": "inline_equation", + "content": "^{3,\\dagger}" + }, + { + "bbox": [ + 65, + 160, + 545, + 246 + ], + "type": "text", + "content": ", Shuhao Li" + }, + { + "bbox": [ + 65, + 160, + 545, + 246 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 65, + 160, + 545, + 246 + ], + "type": "text", + "content": ", Keyi Liu" + }, + { + "bbox": [ + 65, + 160, + 545, + 246 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 65, + 160, + 545, + 246 + ], + "type": "text", + "content": ", Yeqi Luo" + }, + { + "bbox": [ + 65, + 160, + 545, + 246 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 65, + 160, + 545, + 246 + ], + "type": "text", + "content": ", Lipeng Ma" + }, + { + "bbox": [ + 65, + 160, + 545, + 246 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 65, + 160, + 545, + 246 + ], + "type": "text", + "content": ", Lei Bai" + }, + { + "bbox": [ + 65, + 160, + 545, + 246 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 65, + 160, + 545, + 246 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 65, + 160, + 545, + 246 + ], + "type": "text", + "content": "Fudan University, " + }, + { + "bbox": [ + 65, + 160, + 545, + 246 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 65, + 160, + 545, + 246 + ], + "type": "text", + "content": "Shanghai Artificial Intelligence Laboratory, " + }, + { + "bbox": [ + 65, + 160, + 545, + 246 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 65, + 160, + 545, + 246 + ], + "type": "text", + "content": "Chinese University of Hong Kong jyxu22@m.fudan.edu.cn, wdyang@fudan.edu.cn, benfei@cuhk.edu.hk, baisanshi@gmail.com * Equal Contributions, † Corresponding Authors" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 273, + 201, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 273, + 201, + 285 + ], + "spans": [ + { + "bbox": [ + 151, + 273, + 201, + 285 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 298, + 297, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 298, + 297, + 598 + ], + "spans": [ + { + "bbox": [ + 54, + 298, + 297, + 598 + ], + "type": "text", + "content": "Variation of Arctic sea ice has significant impacts on polar ecosystems, transporting routes, coastal communities, and global climate. Tracing the change of sea ice at a finer scale is paramount for both operational applications and scientific studies. Recent pan-Arctic sea ice forecasting methods that leverage advances in artificial intelligence have made promising progress over numerical models. However, forecasting sea ice at higher resolutions is still under-explored. To bridge the gap, we propose a two-module cooperative deep learning framework, IceDiff, to forecast sea ice concentration at finer scales. IceDiff first leverages a vision transformer to generate coarse yet superior forecasting results over previous methods at a regular " + }, + { + "bbox": [ + 54, + 298, + 297, + 598 + ], + "type": "inline_equation", + "content": "25\\mathrm{km}" + }, + { + "bbox": [ + 54, + 298, + 297, + 598 + ], + "type": "text", + "content": " grid. This high-quality sea ice forecasting can be utilized as reliable guidance for the next module. Subsequently, an unconditional diffusion model pre-trained on low-resolution sea ice concentration maps is utilized for sampling down-scaled sea ice forecasting via a zero-shot guided sampling strategy and a patch-based method. For the first time, IceDiff demonstrates sea ice forecasting with a " + }, + { + "bbox": [ + 54, + 298, + 297, + 598 + ], + "type": "inline_equation", + "content": "6.25\\mathrm{km}" + }, + { + "bbox": [ + 54, + 298, + 297, + 598 + ], + "type": "text", + "content": " resolution. IceDiff extends the boundary of existing sea ice forecasting models and more importantly, its capability to generate high-resolution sea ice concentration data is vital for pragmatic usages and research. Code is available at https://github.com/EtronTech/IceDiff." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 620, + 137, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 620, + 137, + 633 + ], + "spans": [ + { + "bbox": [ + 55, + 620, + 137, + 633 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "type": "text", + "content": "Arctic sea ice is essential for global climate change and local communities. Although the exact teleconnections between mid-latitude weather and the variation of Arctic sea ice are yet to be revealed, the reduction of sea ice during the last few decades clearly has impacts that transcend the Arctic region [3, 41]. Hence, establishing skillful forecasting" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 274, + 555, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 274, + 555, + 299 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 555, + 299 + ], + "type": "text", + "content": "models at finer scales is paramount for both transportation in polar regions and scientific research in geoscience." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 302, + 556, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 302, + 556, + 482 + ], + "spans": [ + { + "bbox": [ + 313, + 302, + 556, + 482 + ], + "type": "text", + "content": "Existing physics-based numerical models have demonstrated the ability to accurately forecast sea ice concentration (SIC, value ranges from " + }, + { + "bbox": [ + 313, + 302, + 556, + 482 + ], + "type": "inline_equation", + "content": "0\\%" + }, + { + "bbox": [ + 313, + 302, + 556, + 482 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 302, + 556, + 482 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 313, + 302, + 556, + 482 + ], + "type": "text", + "content": ") for short lead times of a few weeks. However, when forecasting at longer scales, their performance significantly decreases, and their forecasting skills are even worse than those of simple statistical models [3]. With the advent of deep learning methods, a series of works based on deep learning approaches have been committed to SIC forecasting with promising results. Notable studies cover forecasting leads range from daily to seasonal scale, e.g., IceNet [3] provides accurate sea ice extent (SIE, areas where the SIC value is greater than " + }, + { + "bbox": [ + 313, + 302, + 556, + 482 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 313, + 302, + 556, + 482 + ], + "type": "text", + "content": " and could be covered with sea ice) forecasts for 6 months lead time. SICNet [35] and " + }, + { + "bbox": [ + 313, + 302, + 556, + 482 + ], + "type": "inline_equation", + "content": "\\mathrm{SICNet}_{90}" + }, + { + "bbox": [ + 313, + 302, + 556, + 482 + ], + "type": "text", + "content": " [34] are skillful for forecasting SIC in subsequent 7 and 90 days, respectively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 486, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 556, + 714 + ], + "type": "text", + "content": "Despite the improved forecasting skills of deep learning models over previous approaches, these models are still trained on sea ice products provided by the National Sea Ice Data Center (NSIDC) or the European Organisation for the Exploitation of Meteorological Satellites (EU-METSAT). These sea ice data products are primarily derived from passive microwave sensors with a resolution of " + }, + { + "bbox": [ + 313, + 486, + 556, + 714 + ], + "type": "inline_equation", + "content": "25\\mathrm{km}" + }, + { + "bbox": [ + 313, + 486, + 556, + 714 + ], + "type": "text", + "content": ", which is considered too coarse for sea ice modeling. The in situ observations are costly to acquire and extremely spare to provide sufficient spatial coverage over pan-Arctic region. This limitation hinders the representation of detailed information on sea ice concentration and the accuracy of ice edge [10, 16]. The same challenge is also presented in weather forecasts. Therefore, several endeavors are proposed to work on downscaling atmospheric states [5, 11, 43]. For instance, SwinRDM [11] combines an improved SwinRNN [21] model with a denoising diffusion probabilistic model (DDPM) [20] to achieve superresolution (SR) in forecasting climate variables. However," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10567" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 72, + 294, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 72, + 294, + 107 + ], + "spans": [ + { + "bbox": [ + 57, + 72, + 294, + 107 + ], + "type": "text", + "content": "forecasting sea ice at finer spatial scales, which is crucial, especially for operational applications is still under-explored." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 111, + 294, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 111, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 57, + 111, + 294, + 613 + ], + "type": "text", + "content": "Due to the absence of high-resolution sea ice data products and sparse in situ observations, it is not feasible to train a forecast model over pan-Arctic region at a finer scale. Two primary solutions can address this challenge: (i) Downscale sea ice maps initially and then train a forecast model using these high-resolution SIC maps. (ii) Train a forecast model at the original resolution based on the sea ice products, then downscale the forecast results to a finer scale. In the first solution, downscaling sea ice maps may introduce slight errors, which could accumulate during the forecasting stage. Additionally, training a forecast model on high-resolution SIC maps also imposes a greater computational burden. Therefore, we opt for the latter solution and introduce IceDiff, a two-module cooperative framework. IceDiff aims to provide skillful forecasts at three different temporal scales in the forecasting module (FM) and produce spatially downscaled high-quality SIC maps in the guided diffusion module (GDM). Specifically, in the first module, IceDiff leverages U-Net-based vision transformer and provides forecasts that are superior to existing SIC forecasting models. Then, in the second module, we utilize a well-trained DDPM on the original resolution from the NSIDC production as effective prior and downscale forecasted SIC maps guided by forecasts from the previous module at the original scale. Our DDPM is pre-trained on daily SIC maps with a resolution of " + }, + { + "bbox": [ + 57, + 111, + 294, + 613 + ], + "type": "inline_equation", + "content": "25\\mathrm{km}" + }, + { + "bbox": [ + 57, + 111, + 294, + 613 + ], + "type": "text", + "content": ", and it can also generalize well to weekly and monthly SIC maps. In every step " + }, + { + "bbox": [ + 57, + 111, + 294, + 613 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 57, + 111, + 294, + 613 + ], + "type": "text", + "content": " of the sampling process of DDPM, the intermediate variable " + }, + { + "bbox": [ + 57, + 111, + 294, + 613 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 57, + 111, + 294, + 613 + ], + "type": "text", + "content": " is estimated from the denoised SIC map " + }, + { + "bbox": [ + 57, + 111, + 294, + 613 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 57, + 111, + 294, + 613 + ], + "type": "text", + "content": ". Then we propose a convolution kernel with estimizable parameters to simulate the up-scale degradation for the intermediate variable " + }, + { + "bbox": [ + 57, + 111, + 294, + 613 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 57, + 111, + 294, + 613 + ], + "type": "text", + "content": ", which results in up-scaled " + }, + { + "bbox": [ + 57, + 111, + 294, + 613 + ], + "type": "inline_equation", + "content": "\\tilde{x}_0" + }, + { + "bbox": [ + 57, + 111, + 294, + 613 + ], + "type": "text", + "content": ". A distance function between the low-resolution SIC map and up-scaled " + }, + { + "bbox": [ + 57, + 111, + 294, + 613 + ], + "type": "inline_equation", + "content": "\\tilde{x}_0" + }, + { + "bbox": [ + 57, + 111, + 294, + 613 + ], + "type": "text", + "content": " is utilized to dynamically update the parameters of the convolution kernel in the reverse steps and guide the sampling of SIC map in the next time step. Furthermore, to make our IceDiff suitable for SIC maps with various resolutions, a patch-based method is also designed to downscale SIC maps at artistry resolution. We show in this work that using low-resolution SIC maps as guidance can better capture small-scale details in the generated high-resolution SIC maps. Our contributions are three folds:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 617, + 294, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 617, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 617, + 294, + 712 + ], + "type": "text", + "content": "- We propose a novel SIC forecasting framework IceDiff that demonstrates skills in three different temporal scales which cover subseasonal to seasonal forecasts, specifically, 7 days, weekly average of 8 weeks and monthly average of 6 months, that are commonly investigated by the community. IceDiff achieves state-of-the-art results in all three scales compared to existing deep learning models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 317, + 72, + 553, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 72, + 553, + 191 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 553, + 191 + ], + "type": "text", + "content": "- IceDiff is devised with an estimizable convolutional kernel to simulate the up-scaling degradation and designs a patch-based method for downscaling SIC maps at arbitrary resolution with a single and fixed-resolution unconditional diffusion model. By utilizing the generative diffusion prior inherent in this diffusion model, the high-resolution SIC maps can be obtained through zero-shot sampling. Moreover, our IceDiff is capable of capturing small-scale details and generating high-resolution SIC maps." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 316, + 192, + 553, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 192, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 316, + 192, + 553, + 275 + ], + "type": "text", + "content": "- Extensive experiments show that our IceDiff could provide accurate SIC forecasts and generate high-quality sea ice samples with finer details. IceDiff not only provides a new and comprehensive baseline for data-driven sea ice research but also extends the capability of existing SIC forecasting models and pushes this area of research towards more operational usage." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 316, + 290, + 405, + 303 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 290, + 405, + 303 + ], + "spans": [ + { + "bbox": [ + 316, + 290, + 405, + 303 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 316, + 312, + 529, + 324 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 312, + 529, + 324 + ], + "spans": [ + { + "bbox": [ + 316, + 312, + 529, + 324 + ], + "type": "text", + "content": "2.1. Deep Learning-based Sea Ice Forecasting" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 315, + 331, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 331, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 315, + 331, + 553, + 712 + ], + "type": "text", + "content": "Deep learning models have drawn the attention of sea ice research communities and have been widely investigated for Arctic sea ice forecasting [1, 2, 22, 33]. As a pioneering work, ConvLSTM [24] predicts SIC within the region of Barents Sea and has showcased comparable results to the state-of-the-art numerical climate models. IceNet [3] employs the U-Net architecture for merging pre-selected ERA5 variables and SIC data to improve the accuracy of seasonal SIE prediction up to six months. The experimental results have shown that the forecasting skill of IceNet in pan-Arctic region is superior to the numerical sea ice simulation model SEAS5. MT-IceNet [1] adopts a similar approach to IceNet and adds bi-monthly data to further improve forecasting SIC at seasonal scale. Mu et al. [30] propose to merge atmospheric and oceanographic data in a weighted manner through a variable selection network and provide nine months of forecasting leads. Considering the complex ice-ocean-atmosphere interactions have already embedded in the spatiotemporal correlations of consecutive SIC maps [33, 35], Ren et al. [35] proposed SICNet that solely rely on SIC data to predict concentration for the next seven days. Similar to IceNet, SICNet is built upon U-Net architecture and added with channel-wise dependency via temporal-spatial attention module (TSAM), which is essentially a revised version of the convolutional block attention module. By fusing the output of two SICNet, i.e. one for modeling SIC and the other for extracting climate information, and adjusting the number of output channels, Ren et al. [34] extends the lead time to 90 days. Besides CNN and LSTM-based models, recent works leverage the latest advances in artificial intelligence to further improve forecasting performance. For instance, Ice" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "type": "text", + "content": "10568" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": "Former [42] jointly models SIC and sea ice thickness (SIT) based on Transformer backbone. It first decomposes high dimensional spatio-temporal data into sequences via multivariate empirical orthogonal functions [37, 38]. Then an encoder-decoder Transformer is utilized to predict SIC and SIT for up to 45 days. Despite the performance gain of the aforementioned deep learning models, none of them have addressed forecasting sea ice at scales lower than " + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "25\\mathrm{km}" + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 179, + 160, + 192 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 179, + 160, + 192 + ], + "spans": [ + { + "bbox": [ + 55, + 179, + 160, + 192 + ], + "type": "text", + "content": "2.2. Super Resolution" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 198, + 296, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 296, + 544 + ], + "type": "text", + "content": "Super-resolution aims to reconstruct high-resolution (HR) images from low-resolution (LR) counterparts. Several works [27, 28] explore the use of Variational Autoencoders for image SR. These methods usually improve the restoration quality by learning the conditional distribution of high-resolution images induced by low-resolution images. Although these methods are able to achieve realistic perceptual quality, VAE-based methods are often not as satisfactory due to their limited generation capabilities. To this end, other deep learning methods, such as Convolutional neural networks (CNNs), have been widely used for superresolution tasks in various domains. CNN-based methods typically use HR images corresponding to LR images to supervise training with reconstruction loss [4, 7, 36, 44]. However, these network trained solely on reconstruction loss makes it difficult to capture detailed information in images and generate diverse image results. Therefore, Generative Adversarial Networks (GANs) are proposed to solve these challenges. Stengel et al. [39] utilized GAN in a two-step SR process to achieve downscaling of climatological wind and solar irradiance. Leinonen et al. [23] introduced a recurrent, stochastic SR GAN for down-scaling time-evolving atmospheric fields, demonstrating the potential of GANs in improving the spatial resolution of LR images in atmospheric sciences applications. These models encourage the generator to generate high-quality images that the discriminator cannot distinguish from real images. Although GAN-based methods obtain satisfactory downscaled results, the diversity could be enhanced." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": "Currently, diffusion model-based methods have been more widely used due to their diversity in generating images and the ability to generate high-quality images comparable to GANs [17, 40]. For instance, Chen et al. [11] utilize SR based on the conditional model to recover the high spatial resolution and finer-scale atmospheric details, unfolding the ability of diffusion models to generate high-quality and detailed images in SR tasks. The objective of this study is to forecast SIC maps at a high resolution of " + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "inline_equation", + "content": "6.25\\mathrm{km}" + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": ". To achieve this, we have improved the SwinTransformer for predicting SIC at the standard resolution comparable to current methods. Additionally, we employ a diffusion-based SR model to generate high-resolution and high-quality results via zero-shot sampling." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 314, + 71, + 397, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 397, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 397, + 85 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 91, + 554, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 554, + 175 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 554, + 175 + ], + "type": "text", + "content": "Our IceDiff framework comprises two modules as depicted in Figure 1. The forecasting module provides skillful forecasts at three different temporal scales. The guided diffusion module produces spatially down-scaled high-quality SIC maps with a pre-trained DDPM via zero-shot sampling and the patch-based method. We elaborate on design details in the following sections." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 182, + 430, + 194 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 182, + 430, + 194 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 430, + 194 + ], + "type": "text", + "content": "3.1. Forecasting Module" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 200, + 555, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 200, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 200, + 555, + 510 + ], + "type": "text", + "content": "To achieve more accurate SIC forecasting at a finer resolution, a skillful forecasting module trained on original sea ice data products is primarily encouraged. In contrast to existing SIC forecasting methods that primarily concentrate on a single temporal scale [1, 3, 42], our objective is to enhance forecasting skill across various temporal scales and increase versatility. Encouraged by the success of Swin Transformer [25] in various data-driven weather forecasting methods [6, 11, 12], we propose to construct forecasting module by utilizing the U-Net architecture [8] and building on top of the Swin Transformer V2 backbone [26] with ResNet" + }, + { + "bbox": [ + 313, + 200, + 555, + 510 + ], + "type": "inline_equation", + "content": "_{TSAM}" + }, + { + "bbox": [ + 313, + 200, + 555, + 510 + ], + "type": "text", + "content": " block [35] for better capturing both spatial and temporal features of SIC maps. As illustrated in Figure 1, the FM takes " + }, + { + "bbox": [ + 313, + 200, + 555, + 510 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 200, + 555, + 510 + ], + "type": "text", + "content": " consecutive SIC maps, each of which has a shape of [448, 304], to predict subsequent maps with identical temporal length as the input at " + }, + { + "bbox": [ + 313, + 200, + 555, + 510 + ], + "type": "inline_equation", + "content": "25\\mathrm{km}" + }, + { + "bbox": [ + 313, + 200, + 555, + 510 + ], + "type": "text", + "content": " spatial scale. It consists of encoder, bottleneck and decoder. The patch embedding layer first applies a " + }, + { + "bbox": [ + 313, + 200, + 555, + 510 + ], + "type": "inline_equation", + "content": "2\\times 2" + }, + { + "bbox": [ + 313, + 200, + 555, + 510 + ], + "type": "text", + "content": " windows size to partition the original SIC input into patches and maps them into tokens via 2D convolution. Then, the temporal and spatial dependencies are modeled through the ResNet" + }, + { + "bbox": [ + 313, + 200, + 555, + 510 + ], + "type": "inline_equation", + "content": "_{TSAM}" + }, + { + "bbox": [ + 313, + 200, + 555, + 510 + ], + "type": "text", + "content": " block. Both the encoder and the decoder have 3 layers of Swin Transformers that comprise a pre-configured even number of Swin Transformer Blocks for feature extraction and restoration. The calculation of two consecutive blocks can be described as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 361, + 518, + 553, + 584 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 518, + 553, + 584 + ], + "spans": [ + { + "bbox": [ + 361, + 518, + 553, + 584 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} z _ {s} ^ {b} = L N (W M S A (z ^ {b - 1})) + z ^ {b - 1}, \\\\ z ^ {b} = L N (M L P (z _ {s} ^ {b})) + z _ {s} ^ {b}, \\\\ z _ {s} ^ {b + 1} = L N (S W M S A (z ^ {b})) + z ^ {b}, \\\\ z ^ {b + 1} = L N \\left(M L P \\left(z _ {s} ^ {b + 1}\\right)\\right) + z _ {s} ^ {b}, \\tag {1} \\\\ \\end{array}", + "image_path": "112b20c6ca048be0505887bc7b7909c4710792319d6a281b77d7ef6c8835d7be.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 593, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 593, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 593, + 555, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 593, + 555, + 713 + ], + "type": "inline_equation", + "content": "z_{s}^{b}" + }, + { + "bbox": [ + 313, + 593, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 593, + 555, + 713 + ], + "type": "inline_equation", + "content": "z^{b}" + }, + { + "bbox": [ + 313, + 593, + 555, + 713 + ], + "type": "text", + "content": " represents the output feature of the (Shifted) Window-Multi-head Self Attention module and the MLP module for block " + }, + { + "bbox": [ + 313, + 593, + 555, + 713 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 313, + 593, + 555, + 713 + ], + "type": "text", + "content": ", respectively [26]. LN denotes the LayerNorm operation. The combination of such two consecutive blocks, i.e. the former for the employment of WMSA and the latter for SWMSA, ensures that local patterns of adjacent grids are sufficiently attended by the backbone network. We further adopt a shift window size of " + }, + { + "bbox": [ + 313, + 593, + 555, + 713 + ], + "type": "inline_equation", + "content": "28 \\times 19" + }, + { + "bbox": [ + 313, + 593, + 555, + 713 + ], + "type": "text", + "content": " that is consistent with the ratio of SIC input map for facilitating the SWMHA mechanism. The output of the" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10569" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 556, + 293 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 556, + 293 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 556, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 556, + 293 + ], + "type": "image", + "image_path": "63d1bfeb0826992771755d58394090e093f17358cb73eb1f8d63684a80886259.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 294, + 555, + 361 + ], + "lines": [ + { + "bbox": [ + 55, + 294, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 294, + 555, + 361 + ], + "type": "text", + "content": "Figure 1. Overview of IceDiff, a two-module cooperative framework to forecast sea ice concentration at a finer scale. IceDiff first employs the forecasting module (FM) that takes history SIC data as input to generate subsequent predictions at " + }, + { + "bbox": [ + 55, + 294, + 555, + 361 + ], + "type": "inline_equation", + "content": "25\\mathrm{km}" + }, + { + "bbox": [ + 55, + 294, + 555, + 361 + ], + "type": "text", + "content": " resolution. " + }, + { + "bbox": [ + 55, + 294, + 555, + 361 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 294, + 555, + 361 + ], + "type": "text", + "content": " denotes the number of sequential time stamps, for instance, set " + }, + { + "bbox": [ + 55, + 294, + 555, + 361 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 294, + 555, + 361 + ], + "type": "text", + "content": " to 7 for seven days of forecasting leads, FM takes a tensor with the shape of " + }, + { + "bbox": [ + 55, + 294, + 555, + 361 + ], + "type": "inline_equation", + "content": "\\{7,448,304\\}" + }, + { + "bbox": [ + 55, + 294, + 555, + 361 + ], + "type": "text", + "content": " as input and forecasts the SIC for the next 7 days. " + }, + { + "bbox": [ + 55, + 294, + 555, + 361 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 294, + 555, + 361 + ], + "type": "text", + "content": " represents the quantity of embedded channels within each layer of FM, the original SIC map data has a channel of 1 rather than 3 for regular images. The guided diffusion module (GDM) then leverages accurate forecasts as guidance to generate downscaled samples at " + }, + { + "bbox": [ + 55, + 294, + 555, + 361 + ], + "type": "inline_equation", + "content": "6.25\\mathrm{km}" + }, + { + "bbox": [ + 55, + 294, + 555, + 361 + ], + "type": "text", + "content": " resolution with high quality, see details in Figure 2." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 373, + 297, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 373, + 297, + 541 + ], + "spans": [ + { + "bbox": [ + 54, + 373, + 297, + 541 + ], + "type": "text", + "content": "last block within each layer is delivered to a ResNet" + }, + { + "bbox": [ + 54, + 373, + 297, + 541 + ], + "type": "inline_equation", + "content": "_{TSAM}" + }, + { + "bbox": [ + 54, + 373, + 297, + 541 + ], + "type": "text", + "content": " block for establishing channel-wise dependencies. Before the subsequent encoder and decoder layers, patch merging and patch expanding [8] are respectively performed to reshape the feature and fit it to the desired size. Following prior works [3, 8, 35], the skip connection is added between the corresponding encoder and decoder layers to improve the performance. The final prediction is generated via a patch expanding layer with ResNet" + }, + { + "bbox": [ + 54, + 373, + 297, + 541 + ], + "type": "inline_equation", + "content": "_{TSAM}" + }, + { + "bbox": [ + 54, + 373, + 297, + 541 + ], + "type": "text", + "content": " that up-samples the decoded feature to the size of the original SIC input. Our meticulous designs enable the forecasting module in IceDiff to be applied across various temporal scales, enhancing forecasting skills, which serves as a more effective initiation for the guided diffusion module of IceDiff." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 552, + 198, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 198, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 198, + 562 + ], + "type": "text", + "content": "3.2. Guided Diffusion Module" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": "Off-the-shelf methods are limited to the forecasting of SIC maps at a scale of " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "25\\mathrm{km}" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " from sea ice data products [1, 3, 35]. However, finer-scale SIC maps remain unexplored which is of significant value to ecosystems, transporting routes, coastal communities, and global climate. Therefore, after obtaining skillful forecasts, we exploit a downscaling module based on an unconditional diffusion model pre-trained on daily SIC maps to generate realistic high-resolution SIC maps with rich details. This guided diffusion module pre-trained on daily SIC maps could generalize well to SIC maps at various scales, such as weekly or monthly SIC maps. Moreover, our guided diffusion module" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 373, + 555, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 555, + 409 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 555, + 409 + ], + "type": "text", + "content": "of IceDiff downscales the SIC forecasts from a forecasting module to a scale of " + }, + { + "bbox": [ + 313, + 373, + 555, + 409 + ], + "type": "inline_equation", + "content": "6.25\\mathrm{km}" + }, + { + "bbox": [ + 313, + 373, + 555, + 409 + ], + "type": "text", + "content": ", further expanding the accuracy and practicality of sea ice prediction." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 408, + 555, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 408, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 408, + 555, + 464 + ], + "type": "text", + "content": "Specifically, the reverse process of the diffusion model is conditioned on the low-resolution sea ice map " + }, + { + "bbox": [ + 313, + 408, + 555, + 464 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 408, + 555, + 464 + ], + "type": "text", + "content": ", which transforms distribution " + }, + { + "bbox": [ + 313, + 408, + 555, + 464 + ], + "type": "inline_equation", + "content": "p_{\\theta}(x_{t-1}|x_t)" + }, + { + "bbox": [ + 313, + 408, + 555, + 464 + ], + "type": "text", + "content": " into conditional distribution " + }, + { + "bbox": [ + 313, + 408, + 555, + 464 + ], + "type": "inline_equation", + "content": "p_{\\theta}(x_{t-1}|x_t,y)" + }, + { + "bbox": [ + 313, + 408, + 555, + 464 + ], + "type": "text", + "content": ". Previous work [15] has derived the conditional transformation formula in the reverse process:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 323, + 472, + 553, + 498 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 472, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 323, + 472, + 553, + 498 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\log p _ {\\theta} \\left(x _ {t} \\mid x _ {t + 1}, y\\right) = \\log \\left(p _ {\\theta} \\left(x _ {t} \\mid x _ {t + 1}\\right) p \\left(y \\mid x _ {t}\\right)\\right) + N _ {1} (2) \\\\ \\approx \\log p _ {\\theta} (z) + N _ {2}, z \\sim \\mathcal {N} (z; \\mu_ {\\theta} \\left(x _ {t}, t\\right) + \\Sigma g, \\Sigma I), (3) \\\\ \\end{array}", + "image_path": "c341d44eea884eb33e1b34a82a2519db32bf1f1c8c1483ce5c200c1bc5e6d12a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 506, + 554, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 554, + 543 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 554, + 543 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 506, + 554, + 543 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 313, + 506, + 554, + 543 + ], + "type": "text", + "content": " refers to " + }, + { + "bbox": [ + 313, + 506, + 554, + 543 + ], + "type": "inline_equation", + "content": "\\nabla_{x_t}\\log p(y|x_t)" + }, + { + "bbox": [ + 313, + 506, + 554, + 543 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 506, + 554, + 543 + ], + "type": "inline_equation", + "content": "N_{1} = \\log p_{\\theta}(y|x_{t + 1})" + }, + { + "bbox": [ + 313, + 506, + 554, + 543 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 506, + 554, + 543 + ], + "type": "inline_equation", + "content": "N_{2}" + }, + { + "bbox": [ + 313, + 506, + 554, + 543 + ], + "type": "text", + "content": " is a constant related to " + }, + { + "bbox": [ + 313, + 506, + 554, + 543 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 313, + 506, + 554, + 543 + ], + "type": "text", + "content": " (proof in Appendix). The variance of the reverse process " + }, + { + "bbox": [ + 313, + 506, + 554, + 543 + ], + "type": "inline_equation", + "content": "\\Sigma = \\Sigma_{\\theta}(x_{t})" + }, + { + "bbox": [ + 313, + 506, + 554, + 543 + ], + "type": "text", + "content": " is constant." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 543, + 554, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 543, + 554, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 543, + 554, + 598 + ], + "type": "text", + "content": "Therefore, the sampling process of " + }, + { + "bbox": [ + 313, + 543, + 554, + 598 + ], + "type": "inline_equation", + "content": "p_{\\theta}(x_{t - 1}|x_t,y)" + }, + { + "bbox": [ + 313, + 543, + 554, + 598 + ], + "type": "text", + "content": " integrates the gradient term " + }, + { + "bbox": [ + 313, + 543, + 554, + 598 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 313, + 543, + 554, + 598 + ], + "type": "text", + "content": " with the diffusion model " + }, + { + "bbox": [ + 313, + 543, + 554, + 598 + ], + "type": "inline_equation", + "content": "(\\mu ,\\Sigma) = (\\mu_{\\theta}(x_{t}),\\Sigma_{\\theta}(x_{t}))" + }, + { + "bbox": [ + 313, + 543, + 554, + 598 + ], + "type": "text", + "content": ", which plays a role in controlling the direction of map generation. We use a heuristic algorithm (proof in Appendix) to approximate the value of " + }, + { + "bbox": [ + 313, + 543, + 554, + 598 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 313, + 543, + 554, + 598 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 356, + 606, + 555, + 619 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 606, + 555, + 619 + ], + "spans": [ + { + "bbox": [ + 356, + 606, + 555, + 619 + ], + "type": "interline_equation", + "content": "\\log p (y \\mid x _ {t}) = - \\log N - s \\mathcal {L} (\\mathcal {D} (\\tilde {x} _ {0}), y) \\tag {4}", + "image_path": "56f2fd06a4f4a10ee9887826379bfe80885f2775814055da04daefd29ed6f1d4.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 350, + 621, + 555, + 633 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 621, + 555, + 633 + ], + "spans": [ + { + "bbox": [ + 350, + 621, + 555, + 633 + ], + "type": "interline_equation", + "content": "g = \\nabla_ {x _ {t}} \\log p (y | x _ {t}) = - s \\nabla_ {x _ {t}} \\mathcal {L} (\\mathcal {D} (x _ {t}), y). \\tag {5}", + "image_path": "6ade37edc4b0d227335fcde7e2c4dff66eb744f5e3b7c1e46df409fcec7a4499.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "Among them, " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": " is the scaling factor used to control the degree of guidance. And " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": " is a distance function used to measure the distance deviation between two SIC maps. " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": " is an up-scale convolution function with estimizable parameters, where the parameters are randomly initialized and dynamically optimized with the gradient of distance function at" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10570" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 77, + 92, + 532, + 364 + ], + "blocks": [ + { + "bbox": [ + 77, + 92, + 532, + 364 + ], + "lines": [ + { + "bbox": [ + 77, + 92, + 532, + 364 + ], + "spans": [ + { + "bbox": [ + 77, + 92, + 532, + 364 + ], + "type": "image", + "image_path": "fd5f3d5c841bca26f4e9ae57dd1c219d066d56fc92b28951abff3474d14c4407.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 154, + 365, + 165, + 373 + ], + "lines": [ + { + "bbox": [ + 154, + 365, + 165, + 373 + ], + "spans": [ + { + "bbox": [ + 154, + 365, + 165, + 373 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 331, + 365, + 342, + 373 + ], + "lines": [ + { + "bbox": [ + 331, + 365, + 342, + 373 + ], + "spans": [ + { + "bbox": [ + 331, + 365, + 342, + 373 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 476, + 365, + 487, + 373 + ], + "lines": [ + { + "bbox": [ + 476, + 365, + 487, + 373 + ], + "spans": [ + { + "bbox": [ + 476, + 365, + 487, + 373 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 62, + 439, + 547, + 535 + ], + "blocks": [ + { + "bbox": [ + 55, + 376, + 555, + 431 + ], + "lines": [ + { + "bbox": [ + 55, + 376, + 555, + 431 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 555, + 431 + ], + "type": "text", + "content": "Figure 2. Overview of guided diffusion module in IceDiff for SIC map downscaling. (a) Pre-trained diffusion model is used to estimate " + }, + { + "bbox": [ + 55, + 376, + 555, + 431 + ], + "type": "inline_equation", + "content": "\\tilde{x}_0" + }, + { + "bbox": [ + 55, + 376, + 555, + 431 + ], + "type": "text", + "content": " at reverse step " + }, + { + "bbox": [ + 55, + 376, + 555, + 431 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 376, + 555, + 431 + ], + "type": "text", + "content": ". The estimable convolutional kernel of step " + }, + { + "bbox": [ + 55, + 376, + 555, + 431 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 376, + 555, + 431 + ], + "type": "text", + "content": " reduces the resolution of " + }, + { + "bbox": [ + 55, + 376, + 555, + 431 + ], + "type": "inline_equation", + "content": "X_{t}" + }, + { + "bbox": [ + 55, + 376, + 555, + 431 + ], + "type": "text", + "content": " so that guidance from low-resolution SIC maps can be incorporated to sample " + }, + { + "bbox": [ + 55, + 376, + 555, + 431 + ], + "type": "inline_equation", + "content": "X_{t-1}" + }, + { + "bbox": [ + 55, + 376, + 555, + 431 + ], + "type": "text", + "content": ". (b) Low-resolution SIC map " + }, + { + "bbox": [ + 55, + 376, + 555, + 431 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 55, + 376, + 555, + 431 + ], + "type": "text", + "content": " is integrated during the sampling process. A distance function describes the distance between a low-resolution SIC map and the generated high-resolution SIC map " + }, + { + "bbox": [ + 55, + 376, + 555, + 431 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 376, + 555, + 431 + ], + "type": "text", + "content": " after convolution. The gradient of distance function could be used to guide the sampling of " + }, + { + "bbox": [ + 55, + 376, + 555, + 431 + ], + "type": "inline_equation", + "content": "x_{t-1}" + }, + { + "bbox": [ + 55, + 376, + 555, + 431 + ], + "type": "text", + "content": " and (c) dynamically update the parameters of the convolutional kernel." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 62, + 439, + 547, + 535 + ], + "lines": [ + { + "bbox": [ + 62, + 439, + 547, + 535 + ], + "spans": [ + { + "bbox": [ + 62, + 439, + 547, + 535 + ], + "type": "image", + "image_path": "f8f320287e357e63b912bd33dc4cbccc8d5dacdbf994dec8df6e639a57b2e7f9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 539, + 555, + 583 + ], + "lines": [ + { + "bbox": [ + 55, + 539, + 555, + 583 + ], + "spans": [ + { + "bbox": [ + 55, + 539, + 555, + 583 + ], + "type": "text", + "content": "Figure 3. Illustration of the patch-based method. We set patch size " + }, + { + "bbox": [ + 55, + 539, + 555, + 583 + ], + "type": "inline_equation", + "content": "p = 256" + }, + { + "bbox": [ + 55, + 539, + 555, + 583 + ], + "type": "text", + "content": " and stride " + }, + { + "bbox": [ + 55, + 539, + 555, + 583 + ], + "type": "inline_equation", + "content": "r = 128" + }, + { + "bbox": [ + 55, + 539, + 555, + 583 + ], + "type": "text", + "content": " to fit the size of pre-trained models. This setting will divide the example map into " + }, + { + "bbox": [ + 55, + 539, + 555, + 583 + ], + "type": "inline_equation", + "content": "K = 4" + }, + { + "bbox": [ + 55, + 539, + 555, + 583 + ], + "type": "text", + "content": " sub-maps boxed with different colors. For each sub-map, an individualizable convolution kernel with adaptive parameters is set to add guidance from the corresponding low-resolution SIC sub-map. At each sampling step " + }, + { + "bbox": [ + 55, + 539, + 555, + 583 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 539, + 555, + 583 + ], + "type": "text", + "content": ", the average estimated noise on four overlapping blocks will be utilized to update the mean " + }, + { + "bbox": [ + 55, + 539, + 555, + 583 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 55, + 539, + 555, + 583 + ], + "type": "text", + "content": " used to sample " + }, + { + "bbox": [ + 55, + 539, + 555, + 583 + ], + "type": "inline_equation", + "content": "x_{t-1}" + }, + { + "bbox": [ + 55, + 539, + 555, + 583 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 595, + 294, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 595, + 294, + 632 + ], + "spans": [ + { + "bbox": [ + 55, + 595, + 294, + 632 + ], + "type": "text", + "content": "every reverse step. " + }, + { + "bbox": [ + 55, + 595, + 294, + 632 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 595, + 294, + 632 + ], + "type": "text", + "content": " connects two different resolution SIC maps, allowing low-resolution SIC maps to provide more effective guidance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 641, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 712 + ], + "type": "text", + "content": "As shown in Algorithm 1, GDM undergoes " + }, + { + "bbox": [ + 55, + 641, + 296, + 712 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 641, + 296, + 712 + ], + "type": "text", + "content": " reverse steps to gradually restore pure Gaussian noise " + }, + { + "bbox": [ + 55, + 641, + 296, + 712 + ], + "type": "inline_equation", + "content": "x_{T} \\sim \\mathcal{N}(0, I)" + }, + { + "bbox": [ + 55, + 641, + 296, + 712 + ], + "type": "text", + "content": " to high-resolution SIC maps. In each reverse step " + }, + { + "bbox": [ + 55, + 641, + 296, + 712 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 641, + 296, + 712 + ], + "type": "text", + "content": ", the diffusion model calculates the estimated variable " + }, + { + "bbox": [ + 55, + 641, + 296, + 712 + ], + "type": "inline_equation", + "content": "\\tilde{x}_{0}" + }, + { + "bbox": [ + 55, + 641, + 296, + 712 + ], + "type": "text", + "content": " which then undergoes the up-scale convolution function " + }, + { + "bbox": [ + 55, + 641, + 296, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 641, + 296, + 712 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 55, + 641, + 296, + 712 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 641, + 296, + 712 + ], + "type": "text", + "content": " to establish a distance metric " + }, + { + "bbox": [ + 55, + 641, + 296, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 55, + 641, + 296, + 712 + ], + "type": "text", + "content": " with the low-" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 595, + 555, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 595, + 555, + 704 + ], + "spans": [ + { + "bbox": [ + 313, + 595, + 555, + 704 + ], + "type": "text", + "content": "resolution SIC map, where " + }, + { + "bbox": [ + 313, + 595, + 555, + 704 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_t^\\varphi" + }, + { + "bbox": [ + 313, + 595, + 555, + 704 + ], + "type": "text", + "content": " signifies that its convolutional kernel parameters are " + }, + { + "bbox": [ + 313, + 595, + 555, + 704 + ], + "type": "inline_equation", + "content": "\\varphi = \\varphi_t" + }, + { + "bbox": [ + 313, + 595, + 555, + 704 + ], + "type": "text", + "content": " at the reverse steps " + }, + { + "bbox": [ + 313, + 595, + 555, + 704 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 595, + 555, + 704 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 313, + 595, + 555, + 704 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 313, + 595, + 555, + 704 + ], + "type": "text", + "content": " subscript " + }, + { + "bbox": [ + 313, + 595, + 555, + 704 + ], + "type": "inline_equation", + "content": "\\varphi = \\varphi_t" + }, + { + "bbox": [ + 313, + 595, + 555, + 704 + ], + "type": "text", + "content": " indicates that it is a function dependent on the convolutional kernel parameter " + }, + { + "bbox": [ + 313, + 595, + 555, + 704 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 313, + 595, + 555, + 704 + ], + "type": "text", + "content": " at reverse step " + }, + { + "bbox": [ + 313, + 595, + 555, + 704 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 595, + 555, + 704 + ], + "type": "text", + "content": ". The gradients of distance function " + }, + { + "bbox": [ + 313, + 595, + 555, + 704 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 313, + 595, + 555, + 704 + ], + "type": "text", + "content": " with respect to " + }, + { + "bbox": [ + 313, + 595, + 555, + 704 + ], + "type": "inline_equation", + "content": "\\tilde{x}_0" + }, + { + "bbox": [ + 313, + 595, + 555, + 704 + ], + "type": "text", + "content": " and the convolution kernel parameters of the up-scale convolution function are used to update " + }, + { + "bbox": [ + 313, + 595, + 555, + 704 + ], + "type": "inline_equation", + "content": "\\tilde{x}_0" + }, + { + "bbox": [ + 313, + 595, + 555, + 704 + ], + "type": "text", + "content": " and the convolution kernel parameters, respectively. The update of convolution kernel parameters ensures the accuracy of the model's" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10571" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 116, + 553, + 262 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "type": "text", + "content": "Table 1. Quantitative results of forecasts. IceDiff is evaluated on three different temporal scales. The output length of IceDiff equals a pre-defined forecasting lead as depicted in Figure 1. The validated accuracy of IceDiff could provide reliable guidance for the guided diffusion module to generate high-quality downscaled SIC. BT represents that the SIC data [14] are generated by the Bootstrap retrieval algorithm [13] and the data distribution of CDR is considered to be closer to BT rather than NT [29]." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 116, + 553, + 262 + ], + "lines": [ + { + "bbox": [ + 58, + 116, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 58, + 116, + 553, + 262 + ], + "type": "table", + "html": "
Temporal ScaleLead TimeDatasetMethodsRMSE↓MAE↓R²↑NSE↑IIEE↓SIEdif↓
Subseasonal7 Days (Daily)NTSICNet [35]0.04300.00910.9780.97511750.1127
NTIceDiff0.03240.00770.9870.9868990.0186
CDRSICNet [35]0.04900.01000.9820.9799760.0718
CDRIceDiff0.03960.00800.9890.9878350.0315
45 Days (Daily)BTIceFormer [42]0.06600.02010.960---
8 Weeks Average (Weekly)NTIceDiff0.04030.01080.9750.97211700.1051
CDRIceDiff0.05530.01120.9730.96913530.0871
90 Days (Daily)NTSICNet90 [34]-0.0512----
Seasonal6 Months Average (Monthly)NTIceNet [3]0.18200.09160.567---
NTMT-IceNet [1]0.07770.01970.915---
NTIceDiff0.05730.01880.9580.95220140.2419
CDRIceDiff0.06480.01680.9190.91320160.2657
", + "image_path": "dd2d2bc103ea3c18c14546641ec03508720e98d2ce31462764c311e1b6c4d9bd.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 271, + 294, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 271, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 55, + 271, + 294, + 308 + ], + "type": "text", + "content": "Algorithm 1 Guided diffusion module in IceDiff with the guidance of low-resolution SIC map " + }, + { + "bbox": [ + 55, + 271, + 294, + 308 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 271, + 294, + 308 + ], + "type": "text", + "content": ". An unconditional diffusion model pre-trained on SIC maps " + }, + { + "bbox": [ + 55, + 271, + 294, + 308 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_t,t)" + }, + { + "bbox": [ + 55, + 271, + 294, + 308 + ], + "type": "text", + "content": " is given." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 311, + 295, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 311, + 295, + 354 + ], + "spans": [ + { + "bbox": [ + 56, + 311, + 295, + 354 + ], + "type": "text", + "content": "Input: Low-resolution SIC map " + }, + { + "bbox": [ + 56, + 311, + 295, + 354 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 56, + 311, + 295, + 354 + ], + "type": "text", + "content": ", downscale convolution function " + }, + { + "bbox": [ + 56, + 311, + 295, + 354 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 56, + 311, + 295, + 354 + ], + "type": "text", + "content": " composed of optimized convolutional kernels " + }, + { + "bbox": [ + 56, + 311, + 295, + 354 + ], + "type": "inline_equation", + "content": "\\mathcal{K}" + }, + { + "bbox": [ + 56, + 311, + 295, + 354 + ], + "type": "text", + "content": " with parameters " + }, + { + "bbox": [ + 56, + 311, + 295, + 354 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 56, + 311, + 295, + 354 + ], + "type": "text", + "content": ", learning rate " + }, + { + "bbox": [ + 56, + 311, + 295, + 354 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 56, + 311, + 295, + 354 + ], + "type": "text", + "content": ", guidance scale " + }, + { + "bbox": [ + 56, + 311, + 295, + 354 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 56, + 311, + 295, + 354 + ], + "type": "text", + "content": ", distant function " + }, + { + "bbox": [ + 56, + 311, + 295, + 354 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 56, + 311, + 295, + 354 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 355, + 294, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 355, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 355, + 294, + 376 + ], + "type": "text", + "content": "Output: Output high-resolution SIC map " + }, + { + "bbox": [ + 56, + 355, + 294, + 376 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 56, + 355, + 294, + 376 + ], + "type": "text", + "content": " conditioned on " + }, + { + "bbox": [ + 56, + 355, + 294, + 376 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 56, + 355, + 294, + 376 + ], + "type": "text", + "content": ". Sample " + }, + { + "bbox": [ + 56, + 355, + 294, + 376 + ], + "type": "inline_equation", + "content": "x_T" + }, + { + "bbox": [ + 56, + 355, + 294, + 376 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 56, + 355, + 294, + 376 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0, I)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 377, + 216, + 503 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 377, + 157, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 377, + 157, + 386 + ], + "spans": [ + { + "bbox": [ + 61, + 377, + 157, + 386 + ], + "type": "text", + "content": "1: for all t from T to 1 do" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 386, + 183, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 386, + 183, + 400 + ], + "spans": [ + { + "bbox": [ + 61, + 386, + 183, + 400 + ], + "type": "text", + "content": "2: " + }, + { + "bbox": [ + 61, + 386, + 183, + 400 + ], + "type": "inline_equation", + "content": "\\tilde{x}_0 = \\frac{x_t}{\\sqrt{\\alpha_t}} - \\frac{\\sqrt{1 - \\alpha_t} \\epsilon_0(x_t, t)}{\\sqrt{\\alpha_t}}" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 400, + 169, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 400, + 169, + 411 + ], + "spans": [ + { + "bbox": [ + 61, + 400, + 169, + 411 + ], + "type": "text", + "content": "3: " + }, + { + "bbox": [ + 61, + 400, + 169, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\varphi, \\tilde{x}_0} = \\mathcal{L}(y, \\mathcal{D}^\\varphi(\\tilde{x}_0))" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 411, + 200, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 411, + 200, + 428 + ], + "spans": [ + { + "bbox": [ + 61, + 411, + 200, + 428 + ], + "type": "text", + "content": "4: " + }, + { + "bbox": [ + 61, + 411, + 200, + 428 + ], + "type": "inline_equation", + "content": "\\tilde{x}_0\\gets \\tilde{x}_0 - \\frac{s(1 - \\bar{\\alpha}_t)}{\\sqrt{\\bar{\\alpha}_{t - 1}\\beta_t}}\\nabla_{\\tilde{x}_0}\\mathcal{L}_{\\varphi ,\\tilde{x}_0}" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 61, + 428, + 216, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 428, + 216, + 445 + ], + "spans": [ + { + "bbox": [ + 61, + 428, + 216, + 445 + ], + "type": "text", + "content": "5: " + }, + { + "bbox": [ + 61, + 428, + 216, + 445 + ], + "type": "inline_equation", + "content": "\\tilde{\\mu}_{t} = \\frac{\\sqrt{\\bar{\\alpha}_{t - 1}}\\beta_{t}}{1 - \\bar{\\alpha}_{t}}\\tilde{x}_{0} + \\frac{\\sqrt{\\bar{\\alpha}_{t}}(1 - \\bar{\\alpha}_{t - 1})}{1 - \\bar{\\alpha}_{t}} x_{t}" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 61, + 445, + 141, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 445, + 141, + 459 + ], + "spans": [ + { + "bbox": [ + 61, + 445, + 141, + 459 + ], + "type": "text", + "content": "6: " + }, + { + "bbox": [ + 61, + 445, + 141, + 459 + ], + "type": "inline_equation", + "content": "\\tilde{\\beta}_{t} = \\frac{1 - \\bar{\\alpha}_{t - 1}}{1 - \\bar{\\alpha}_{t}}\\beta_{t}" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 61, + 459, + 194, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 459, + 194, + 472 + ], + "spans": [ + { + "bbox": [ + 61, + 459, + 194, + 472 + ], + "type": "text", + "content": "7: Sample " + }, + { + "bbox": [ + 61, + 459, + 194, + 472 + ], + "type": "inline_equation", + "content": "x_{t-1}" + }, + { + "bbox": [ + 61, + 459, + 194, + 472 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 61, + 459, + 194, + 472 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\tilde{\\mu}_t, \\tilde{\\beta}_t I)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 61, + 472, + 159, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 472, + 159, + 483 + ], + "spans": [ + { + "bbox": [ + 61, + 472, + 159, + 483 + ], + "type": "text", + "content": "8: " + }, + { + "bbox": [ + 61, + 472, + 159, + 483 + ], + "type": "inline_equation", + "content": "\\varphi \\gets \\varphi - l \\nabla_{\\varphi} \\mathcal{L}_{\\varphi, \\bar{x}_0}" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 61, + 483, + 102, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 483, + 102, + 491 + ], + "spans": [ + { + "bbox": [ + 61, + 483, + 102, + 491 + ], + "type": "text", + "content": "9: end for" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 58, + 494, + 110, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 494, + 110, + 503 + ], + "spans": [ + { + "bbox": [ + 58, + 494, + 110, + 503 + ], + "type": "text", + "content": "10: return " + }, + { + "bbox": [ + 58, + 494, + 110, + 503 + ], + "type": "inline_equation", + "content": "x_0" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 532, + 295, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 532, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 532, + 295, + 555 + ], + "type": "text", + "content": "up-scaling process, making guidance of low-resolution SIC maps more precise and effective." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 54, + 558, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 558, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 558, + 296, + 713 + ], + "type": "text", + "content": "Patch-based Strategy. Furthermore, pre-trained models can only generate fixed-size SIC maps. To tackle the problem of sampling downscaled SIC forecasts at any resolution, we employ a patch-based strategy. As shown in Figure 3, the patch-based strategy partitions SIC maps into several sub-regions based on a fixed stride. For each individual sub-region, the gradient terms of distance loss between SIC maps upscaled by the convolutional kernel and the corresponding low-resolution SIC sub-region maps are computed individually. Subsequently, the respective mean of Gaussian distribution in sub-regions (Eq. 3) and convolution kernel parameters are updated based on the gradient terms. The entire map is updated by taking the mean of" + } + ] + } + ], + "index": 17 + }, + { + "type": "table", + "bbox": [ + 317, + 304, + 553, + 402 + ], + "blocks": [ + { + "bbox": [ + 313, + 270, + 555, + 303 + ], + "lines": [ + { + "bbox": [ + 313, + 270, + 555, + 303 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 555, + 303 + ], + "type": "text", + "content": "Table 2. Quantitative comparison between IceDiff for SIC downscaling and other methods. Const. and FID represents the Consistency and Fréchet Inception Distance metrics, respectively." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 304, + 553, + 402 + ], + "lines": [ + { + "bbox": [ + 317, + 304, + 553, + 402 + ], + "spans": [ + { + "bbox": [ + 317, + 304, + 553, + 402 + ], + "type": "table", + "html": "
Method7 Days8 Week Avg.6 Month Avg.
Const.↓FID↓Const.↓FID↓Const.↓FID↓
Nearest20.0178.4521.9880.1723.0788.36
Bilinear17.8870.7520.0277.2918.3180.73
BiCubic13.1354.7113.4068.8510.0970.18
DGP [32]23.1875.4228.6281.4229.7485.47
GDP [17]8.2640.1510.5452.4010.9655.72
IceDiff6.8429.347.2531.296.6239.79
", + "image_path": "ddc9098d71c0658ec8f78954159b79313823d13bd8efac4ac184dfd99cc707cd.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 411, + 555, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 555, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 555, + 495 + ], + "type": "text", + "content": "the updated values across all sub-regions, each weighted by a binary patch mask that serves as a quantifier of the regional scope, thereby refining the overall mean and convolution kernel parameters of the sampled high-resolution SIC maps. By leveraging this strategy, our model is capable of downscaling sea ice maps to the desired resolution and the practicality of the model has been further enhanced." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 507, + 395, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 507, + 395, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 507, + 395, + 521 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 528, + 464, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 528, + 464, + 541 + ], + "spans": [ + { + "bbox": [ + 313, + 528, + 464, + 541 + ], + "type": "text", + "content": "4.1. Dataset and Baseline Setup" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 545, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 545, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 545, + 555, + 713 + ], + "type": "text", + "content": "We evaluate our proposed IceDiff framework on the Climate Data Record of Passive Microwave Sea Ice Concentration Version 4 dataset from NSIDC [29] (CDR in Table 1). It records daily SIC data starting from October " + }, + { + "bbox": [ + 312, + 545, + 555, + 713 + ], + "type": "inline_equation", + "content": "25^{th}" + }, + { + "bbox": [ + 312, + 545, + 555, + 713 + ], + "type": "text", + "content": " 1978 and provides the coverage of the pan-Arctic region " + }, + { + "bbox": [ + 312, + 545, + 555, + 713 + ], + "type": "inline_equation", + "content": "(\\mathrm{N} : -39.36^{\\circ}, \\mathrm{S} : -89.84^{\\circ}, \\mathrm{E} : 180^{\\circ}, \\mathrm{W} : -180^{\\circ})" + }, + { + "bbox": [ + 312, + 545, + 555, + 713 + ], + "type": "text", + "content": ". The SIC map is formed of " + }, + { + "bbox": [ + 312, + 545, + 555, + 713 + ], + "type": "inline_equation", + "content": "448 \\times 304" + }, + { + "bbox": [ + 312, + 545, + 555, + 713 + ], + "type": "text", + "content": " pixels and each of which represents an area of " + }, + { + "bbox": [ + 312, + 545, + 555, + 713 + ], + "type": "inline_equation", + "content": "25 \\mathrm{~km} \\times 25 \\mathrm{~km}" + }, + { + "bbox": [ + 312, + 545, + 555, + 713 + ], + "type": "text", + "content": " grid and has a value ranging from " + }, + { + "bbox": [ + 312, + 545, + 555, + 713 + ], + "type": "inline_equation", + "content": "0\\%" + }, + { + "bbox": [ + 312, + 545, + 555, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 312, + 545, + 555, + 713 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 312, + 545, + 555, + 713 + ], + "type": "text", + "content": ". We choose data from October " + }, + { + "bbox": [ + 312, + 545, + 555, + 713 + ], + "type": "inline_equation", + "content": "25^{th}" + }, + { + "bbox": [ + 312, + 545, + 555, + 713 + ], + "type": "text", + "content": " 1978 to 2013 as the training dataset, the years 2014 and 2015 are selected as validation set, and data from 2016 to 2023 are used for testing. To comprehensively validate and verify the performance of our forecasting module, we utilize additional data [16] that are generated through the retrieval" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10572" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 95, + 294, + 162 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 295, + 92 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 295, + 92 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 295, + 92 + ], + "type": "text", + "content": "Table 3. Ablation study on TSAM. We compare the forecasting results between IceDiff with or without the TSAM." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 95, + 294, + 162 + ], + "lines": [ + { + "bbox": [ + 58, + 95, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 58, + 95, + 294, + 162 + ], + "type": "table", + "html": "
Lead TimeTSAMRMSE↓MAE↓R²↑NSE↑IIEE↓SIEdif↓
7 DaysX0.04030.00820.98810.98608740.0336
✓0.03960.00800.98850.98658350.0315
8 Week Avg.X0.05350.01160.9750.97113860.1420
✓0.05530.01120.9730.96913530.0871
6 Month Avg.X0.07630.01940.8830.87526100.4919
✓0.06480.01680.9190.91320160.2657
", + "image_path": "0260bf368e92f1519117332c6f087d8bf05cd9404a00beb1af3752b427d0081a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 171, + 296, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 171, + 296, + 328 + ], + "spans": [ + { + "bbox": [ + 55, + 171, + 296, + 328 + ], + "type": "text", + "content": "algorithm (NASA Team [9], NT in Table 1). The forecasting module is trained on three different temporal scales, i.e. the input of SIC maps are 7 days, 8 weeks average, and 6 months average, covering forecasting leads from subseasonal to seasonal scale. Corresponding deep learning-based models are selected for comparison, as illustrated in Table 1. Note that we follow the split of datasets in respective baselines to preprocess the additional data. We re-implemented SICNet (the source code is not available) based on the original paper and trained it under an environment identical to our forecasting module. To evaluate the performance of our downscaling, five methods are selected to set up baselines for downscaling quality comparison." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 335, + 188, + 347 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 335, + 188, + 347 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 188, + 347 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 352, + 296, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 352, + 296, + 557 + ], + "spans": [ + { + "bbox": [ + 55, + 352, + 296, + 557 + ], + "type": "text", + "content": "We independently trained the forecasting module of IceDiff on SIC dataset of three aforementioned temporal scales for 60 epochs. For each temporal scale, the forecasting module outputs the same length of input data as forecasting leads, i.e. 7 days, 8 weeks average, and 6 months averages. The dimensions of embedding features of encoder, decoder, and bottleneck are all set to 32. We utilize the same hyperparameter of TSAM as in [35]. The forecasting module is optimized by AdamW under the mean square error (MSE) objective using Pytorch on one NVIDIA A100 80GB GPU for all experiments with a learning rate of 0.0005. All daily SIC maps are resized and cropped to " + }, + { + "bbox": [ + 55, + 352, + 296, + 557 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 55, + 352, + 296, + 557 + ], + "type": "text", + "content": " to train our GDM of IceDiff. using AdamW with " + }, + { + "bbox": [ + 55, + 352, + 296, + 557 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.9" + }, + { + "bbox": [ + 55, + 352, + 296, + 557 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 352, + 296, + 557 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.999" + }, + { + "bbox": [ + 55, + 352, + 296, + 557 + ], + "type": "text", + "content": ". We train in 16-bit precision using loss-scaling, but maintain 32-bit weights, EMA, and optimizer state. We use an EMA rate of 0.9999. When sampling with 1000 timesteps, we use the same noise schedule as for training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 564, + 168, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 168, + 575 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 168, + 575 + ], + "type": "text", + "content": "4.3. Evaluation Metrics" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 581, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 581, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 581, + 295, + 689 + ], + "type": "text", + "content": "To quantitatively evaluate our IceDff, we select commonly used root mean square error (RMSE) and mean absolute error (MAE) for the comparison of forecasting accuracy. " + }, + { + "bbox": [ + 55, + 581, + 295, + 689 + ], + "type": "inline_equation", + "content": "R^2" + }, + { + "bbox": [ + 55, + 581, + 295, + 689 + ], + "type": "text", + "content": " score and Nash-Sutcliffe Efficiency [31] are utilized to analyze the similarity of spatial pattern between the ground truth and predicted SIC map. We adopt Integrated Ice-Edge Error score [18] and calculate the sea ice area differences " + }, + { + "bbox": [ + 55, + 581, + 295, + 689 + ], + "type": "inline_equation", + "content": "(SIE_{dif})" + }, + { + "bbox": [ + 55, + 581, + 295, + 689 + ], + "type": "text", + "content": ", in millions of " + }, + { + "bbox": [ + 55, + 581, + 295, + 689 + ], + "type": "inline_equation", + "content": "km^2" + }, + { + "bbox": [ + 55, + 581, + 295, + 689 + ], + "type": "text", + "content": " to further evaluate the prediction of SIE (where the SIC value is greater than 15%)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "As to the evaluation of downscaling quality, following [11] we utilize the Fréchet Inception Distance (FID)" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 317, + 127, + 553, + 218 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 555, + 125 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 555, + 125 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 555, + 125 + ], + "type": "text", + "content": "Table 4. Ablation study on the ratio of Input/Forecast number of SIC maps. Hf., Db. and Eq. represent that the forecasting module takes input data which has half, double, and equal length of forecasting leads, respectively. For the Hf. of 7 Days, we set the input length to 3." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 127, + 553, + 218 + ], + "lines": [ + { + "bbox": [ + 317, + 127, + 553, + 218 + ], + "spans": [ + { + "bbox": [ + 317, + 127, + 553, + 218 + ], + "type": "table", + "html": "
Lead TimeI/F.RMSE↓MAE↓R²↑NSE↑IIEE↓SIEdif↓
7 DaysHf.0.04160.00840.9870.9859050.0538
Db.0.04010.00840.9880.9868890.0599
Eq.0.03960.00800.9890.9878350.0315
8 Week Avg.Hf.0.05800.01180.9700.96614050.1176
Db.0.05940.01210.9690.96414350.1221
Eq.0.05530.01120.9730.96913530.0871
6 Month Avg.Hf.0.09040.02320.8420.83031920.6289
Db.0.07220.01790.8970.88920270.2795
Eq.0.06480.01680.9190.91320160.2657
", + "image_path": "d9c3e5ac0cc768541af00ac8aa402ed29d25e59714e257166e1d2552224dc671.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 228, + 553, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 553, + 265 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 553, + 265 + ], + "type": "text", + "content": "[19] for measuring the downscaling fidelity. In addition, we adopt Consistency to quantify the faithfulness between the downscaled map and the low-resolution map." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 266, + 553, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 553, + 289 + ], + "type": "text", + "content": "The detailed equations for calculating the abovementioned metrics can be found in Appendix." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 299, + 399, + 310 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 299, + 399, + 310 + ], + "spans": [ + { + "bbox": [ + 313, + 299, + 399, + 310 + ], + "type": "text", + "content": "4.4. Main Results" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 317, + 553, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 317, + 553, + 340 + ], + "spans": [ + { + "bbox": [ + 313, + 317, + 553, + 340 + ], + "type": "text", + "content": "In this subsection, we evaluate the proposed IceDiff with baseline methods for validation of the effectiveness." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 341, + 555, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 341, + 555, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 341, + 555, + 473 + ], + "type": "text", + "content": "Set up new baselines for SIC prediction. The forecasting module of IceDiff is trained independently to predict SIC at 7 days, 8 weeks average, and 6 months average. As shown in Table 1, the performance of our proposed forecasting module of IceDiff is superior to SICNet in all evaluation metrics. The forecasting skill of FM across different SIC datasets is consistent in three temporal scales. Specifically, it outperforms SOTA seasonal forecasting baselines by a noticeable margin. Considering the overall performance of our FM, it could provide reliable SIC forecasts and sufficiently support the subsequent downscaling module." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 474, + 554, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 554, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 554, + 641 + ], + "type": "text", + "content": "Analysis of the end of the melting season. To further verify the performance of IceDiff in extreme cases, we select two years of the end of the melting season (early September) when the mean Arctic SIC typically reaches its minimum for validation: (i) In 2016, it recorded historically low annual SIC; (ii) In 2022, the annual SIC has an abrupt increase compared to the previous year (annual trends are plotted in the Appendix). We present the SIC residuals (predicted SIC minus the ground truth) of September " + }, + { + "bbox": [ + 313, + 474, + 554, + 641 + ], + "type": "inline_equation", + "content": "1^{st}" + }, + { + "bbox": [ + 313, + 474, + 554, + 641 + ], + "type": "text", + "content": " 2016 and 2022 in Figure 4(a) and Figure 4(b), respectively. The overall residuals are relatively low and the prediction errors mainly appear in part of the sea ice edge regions. The accurate forecasting results in temporal periods with highly varying SIC validate the effectiveness of IceDiff." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "content": "High downscaling quality. We systematically compare the results of IceDiff with other downscaling methods. In terms of interpolation-based methods, we select nearest, bilinear, and bicubic interpolation to compare with our diffusion-based GDM. Besides, we select DGP [32] as a GAN-based method and GDP [17] as a DDPM-based method for fur" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10573" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 161, + 223 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 161, + 223 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 161, + 223 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 161, + 223 + ], + "type": "image", + "image_path": "ef9069d224d6630c2edfcec6deb52788010afeaca97aaac2ba5bdfbea255ab88.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 228, + 71, + 236 + ], + "lines": [ + { + "bbox": [ + 61, + 228, + 71, + 236 + ], + "spans": [ + { + "bbox": [ + 61, + 228, + 71, + 236 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 86, + 225, + 127, + 242 + ], + "lines": [ + { + "bbox": [ + 86, + 225, + 127, + 242 + ], + "spans": [ + { + "bbox": [ + 86, + 225, + 127, + 242 + ], + "type": "text", + "content": "Forecast Date Sep 1,2016" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 161, + 69, + 294, + 224 + ], + "blocks": [ + { + "bbox": [ + 161, + 69, + 294, + 224 + ], + "lines": [ + { + "bbox": [ + 161, + 69, + 294, + 224 + ], + "spans": [ + { + "bbox": [ + 161, + 69, + 294, + 224 + ], + "type": "image", + "image_path": "9e9a75ac043e9c73524086571625a79fb6b4b907dc4ce8dbc6aa3f8dbf1c3068.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 164, + 229, + 175, + 237 + ], + "lines": [ + { + "bbox": [ + 164, + 229, + 175, + 237 + ], + "spans": [ + { + "bbox": [ + 164, + 229, + 175, + 237 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 192, + 225, + 233, + 242 + ], + "lines": [ + { + "bbox": [ + 192, + 225, + 233, + 242 + ], + "spans": [ + { + "bbox": [ + 192, + 225, + 233, + 242 + ], + "type": "text", + "content": "Forecast Date Sep 1,2022" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 243, + 296, + 309 + ], + "lines": [ + { + "bbox": [ + 55, + 243, + 296, + 309 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 296, + 309 + ], + "type": "text", + "content": "Figure 4. Qualitative analysis of the end of the melting season. The annual mean Arctic SIC has recorded a historical low in 2016 and it drastically increased in 2022 when compared to the previous annual mean. The residual plots of early September in (a)2016 and (b)2022 demonstrate where the FM overestimates (positive figures) or underestimates (negative figures) SIC." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 323, + 296, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 323, + 296, + 359 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 296, + 359 + ], + "type": "text", + "content": "ther comparison. Interpolation-based methods upsample the low-resolution sea ice map, which uses existing pixel information to estimate missing pixel values." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 360, + 296, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 360, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 360, + 296, + 456 + ], + "type": "text", + "content": "As shown in Figure 5, our GDM produces high-quality results and small-scale structures are successfully captured and rich details are generated through the guidance from the input map during each sampling step. By contrast, interpolation-based methods do not integrate details information of the low-resolution map, resulting in blurry edges and contours in the generated results. More quality analysis can be found in the appendix." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 456, + 296, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 296, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 296, + 588 + ], + "type": "text", + "content": "To further validate the superiority of our method, we follow SwinRDM [11] and GDP [17] to utilize FID and Consistency to evaluate our method, respectively. As shown in Table 2, compared to interpolation-based methods, our IceDiff improves FID score by 31.11 on average, reflecting it has the capacity to help generate a more realistic high-resolution map with a satisfactory texture restoration effect. A lower Consistency demonstrates the structural difference between the downscaled map and the original SIC input is relatively imperceptible, which better preserves the spatial distribution of sea ice forecast." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 599, + 237, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 599, + 237, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 599, + 237, + 612 + ], + "type": "text", + "content": "4.5. Ablation Studies on Design of FM" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": "As shown in Table 3, by adopting TSAM to FM, the performance has improved for all metrics at 7 days and 6 months average lead times while the effects for 8 weeks average lead time are not all positive. Overall, incorporating TSAM to establish channel-wise dependency is favorable for SIC forecast models. To examine the impacts of input SIC data length, we trained FM for input lengths of half, double, and equal to the forecasting leads. The results in Table 4 show" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 316, + 69, + 395, + 233 + ], + "blocks": [ + { + "bbox": [ + 316, + 69, + 395, + 233 + ], + "lines": [ + { + "bbox": [ + 316, + 69, + 395, + 233 + ], + "spans": [ + { + "bbox": [ + 316, + 69, + 395, + 233 + ], + "type": "image", + "image_path": "03beacc4b8a477ee609539113f2f0d683f334a853d24a075ac30feb8c614c4e1.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 340, + 233, + 370, + 240 + ], + "lines": [ + { + "bbox": [ + 340, + 233, + 370, + 240 + ], + "spans": [ + { + "bbox": [ + 340, + 233, + 370, + 240 + ], + "type": "text", + "content": "(a) IceDiff" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 243, + 555, + 320 + ], + "lines": [ + { + "bbox": [ + 313, + 243, + 555, + 320 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 555, + 320 + ], + "type": "text", + "content": "Figure 5. Qualitative analysis of downscaled SIC map. We visualize the SIC map by coloring the land and the ocean in yellow and blue, respectively. Since SIC has a range of " + }, + { + "bbox": [ + 313, + 243, + 555, + 320 + ], + "type": "inline_equation", + "content": "[0,1]" + }, + { + "bbox": [ + 313, + 243, + 555, + 320 + ], + "type": "text", + "content": ", we map it to " + }, + { + "bbox": [ + 313, + 243, + 555, + 320 + ], + "type": "inline_equation", + "content": "[0,255]" + }, + { + "bbox": [ + 313, + 243, + 555, + 320 + ], + "type": "text", + "content": " and replicate it for all 3 RGB channels. The darker the pixels appear, the SIC values are more closer to " + }, + { + "bbox": [ + 313, + 243, + 555, + 320 + ], + "type": "inline_equation", + "content": "0\\%" + }, + { + "bbox": [ + 313, + 243, + 555, + 320 + ], + "type": "text", + "content": ". For better identification of the difference, we crop out a region and plot a grid size of " + }, + { + "bbox": [ + 313, + 243, + 555, + 320 + ], + "type": "inline_equation", + "content": "250\\mathrm{km}" + }, + { + "bbox": [ + 313, + 243, + 555, + 320 + ], + "type": "text", + "content": " and further display details on " + }, + { + "bbox": [ + 313, + 243, + 555, + 320 + ], + "type": "inline_equation", + "content": "4\\times 4" + }, + { + "bbox": [ + 313, + 243, + 555, + 320 + ], + "type": "text", + "content": " grid." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 395, + 70, + 475, + 233 + ], + "blocks": [ + { + "bbox": [ + 395, + 70, + 475, + 233 + ], + "lines": [ + { + "bbox": [ + 395, + 70, + 475, + 233 + ], + "spans": [ + { + "bbox": [ + 395, + 70, + 475, + 233 + ], + "type": "image", + "image_path": "e4875f857608b86f8a75df81d4c7333796c8907f80f19cb21b882bf8ade52022.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 419, + 233, + 450, + 240 + ], + "lines": [ + { + "bbox": [ + 419, + 233, + 450, + 240 + ], + "spans": [ + { + "bbox": [ + 419, + 233, + 450, + 240 + ], + "type": "text", + "content": "(b) Bicubic" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 474, + 70, + 554, + 233 + ], + "blocks": [ + { + "bbox": [ + 474, + 70, + 554, + 233 + ], + "lines": [ + { + "bbox": [ + 474, + 70, + 554, + 233 + ], + "spans": [ + { + "bbox": [ + 474, + 70, + 554, + 233 + ], + "type": "image", + "image_path": "263520a0bc76ba2995ade0fcdc342302b2e7f246b2ee2dcb577fa8a485f9c13c.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 502, + 233, + 525, + 240 + ], + "lines": [ + { + "bbox": [ + 502, + 233, + 525, + 240 + ], + "spans": [ + { + "bbox": [ + 502, + 233, + 525, + 240 + ], + "type": "text", + "content": "(c) GDP" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 331, + 555, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 555, + 357 + ], + "type": "text", + "content": "the peak of the performance at the ratio where the input length equals the forecasting leads." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 369, + 471, + 381 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 471, + 381 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 471, + 381 + ], + "type": "text", + "content": "5. Conclusion and Limitations" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 389, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 555, + 628 + ], + "type": "text", + "content": "We propose a novel pan-Arctic high-resolution SIC forecasting framework, IceDiff, which comprises a forecasting module and a guided diffusion module, these two modules closely cooperate with each other to fulfill accurate forecasting of SIC at a finer and more practical spatial scale. The forecasting module of IceDiff is capable of producing accurate forecasts at three different temporal scales and the guided diffusion module leverages reliable forecasts as guidance to generate high-quality down-scaled SIC maps. The forecasting skills of IceDiff at 7 days lead are superior to SOTA deep learning-based models. IceDiff also sets up a new baseline for weekly average forecasting of 8 weeks and provides competitive results in seasonal SIC forecasts. To the best of our knowledge, IceDiff is the first attempt to forecast SIC at " + }, + { + "bbox": [ + 313, + 389, + 555, + 628 + ], + "type": "inline_equation", + "content": "6.25\\mathrm{km}" + }, + { + "bbox": [ + 313, + 389, + 555, + 628 + ], + "type": "text", + "content": " spatial scale which is a quarter of the grid size " + }, + { + "bbox": [ + 313, + 389, + 555, + 628 + ], + "type": "inline_equation", + "content": "(25\\mathrm{km})" + }, + { + "bbox": [ + 313, + 389, + 555, + 628 + ], + "type": "text", + "content": " that current methods perform on. The experimental results show that our framework can generate down-scaled SIC superior to off-the-shelf methods and it is vital for pushing forecasting models towards operational usage and further promoting sea ice research." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "Existing sea ice concentration data products are all retrieved from satellite observations with spatial resolution limited to " + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "inline_equation", + "content": "25\\mathrm{km}" + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": ". A finer-scale ground truth is missing compared with observations from stations in weather forecasting. Therefore, our limitation lies in the evaluations of IceDiff on these in situ observations. We will focus on evaluating our IceDiff on higher-resolution SIC observations." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10574" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Sahara Ali and Jianwu Wang. Mt-icenet-a spatial and multi-temporal deep learning model for arctic sea ice forecasting. In 2022 IEEE/ACM International Conference on Big Data Computing, Applications and Technologies (BDCAT), pages 1–10. IEEE, 2022. 2, 3, 4, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 148, + 296, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 148, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 61, + 148, + 296, + 180 + ], + "type": "text", + "content": "[2] Sahara Ali, Yiyi Huang, Xin Huang, and Jianwu Wang. Sea ice forecasting using attention-based ensemble LSTM. arXiv preprint arXiv:2108.00853, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 182, + 295, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 182, + 295, + 247 + ], + "spans": [ + { + "bbox": [ + 62, + 182, + 295, + 247 + ], + "type": "text", + "content": "[3] Tom R Andersson, J Scott Hosking, María Pérez-Ortiz, Brooks Paige, Andrew Elliott, Chris Russell, Stephen Law, Daniel C Jones, Jeremy Wilkinson, Tony Phillips, et al. Seasonal arctic sea ice forecasting with probabilistic deep learning. Nature communications, 12(1):5124, 2021. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 250, + 295, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 250, + 295, + 282 + ], + "spans": [ + { + "bbox": [ + 62, + 250, + 295, + 282 + ], + "type": "text", + "content": "[4] Saeed Anwar, Salman Khan, and Nick Barnes. A deep journey into super-resolution: A survey. ACM Computing Surveys (CSUR), 53(3):1-34, 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 285, + 294, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 285, + 294, + 338 + ], + "spans": [ + { + "bbox": [ + 62, + 285, + 294, + 338 + ], + "type": "text", + "content": "[5] Seth Bassetti, Brian Hutchinson, Claudia Tebaldi, and Ben Kravitz. Diffesm: Conditional emulation of temperature and precipitation in earth system models with 3d diffusion models. Journal of Advances in Modeling Earth Systems, 16(10): e2023MS004194, 2024. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 341, + 294, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 341, + 294, + 384 + ], + "spans": [ + { + "bbox": [ + 62, + 341, + 294, + 384 + ], + "type": "text", + "content": "[6] Kaifeng Bi, Lingxi Xie, Hengheng Zhang, Xin Chen, Xiaotao Gu, and Qi Tian. Accurate medium-range global weather forecasting with 3d neural networks. Nature, 619(7970): 533-538, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 386, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 386, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 62, + 386, + 294, + 430 + ], + "type": "text", + "content": "[7] Bruno Buongiorno Nardelli, Davide Cavaliere, Elodie Charles, and Daniele Ciani. Super-resolving ocean dynamics from space with computer vision algorithms. Remote Sensing, 14(5):1159, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 432, + 294, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 432, + 294, + 485 + ], + "spans": [ + { + "bbox": [ + 62, + 432, + 294, + 485 + ], + "type": "text", + "content": "[8] Hu Cao, Yueyue Wang, Joy Chen, Dongsheng Jiang, Xiaopeng Zhang, Qi Tian, and Manning Wang. Swin-unet: Unet-like pure transformer for medical image segmentation. In European conference on computer vision, pages 205-218. Springer, 2022. 3, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 488, + 294, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 488, + 294, + 520 + ], + "spans": [ + { + "bbox": [ + 62, + 488, + 294, + 520 + ], + "type": "text", + "content": "[9] D.J. Cavalieri, P. Gloersen, and W.J. Campbell. Determination of sea ice parameters with the nimbus-7 smmr. Journal of Geophysical Research, 89:5355-5369, 1984. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 523, + 294, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 523, + 294, + 566 + ], + "spans": [ + { + "bbox": [ + 57, + 523, + 294, + 566 + ], + "type": "text", + "content": "[10] D.J. Cavalieri, P. Gloersen, and W.J. Campbell. Deriving long-term time series of sea ice cover from satellite passive-microwave multisensor data sets. Journal of Geophysical Research, 104:15803-15814, 1999. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 568, + 294, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 568, + 294, + 622 + ], + "spans": [ + { + "bbox": [ + 57, + 568, + 294, + 622 + ], + "type": "text", + "content": "[11] Lei Chen, Fei Du, Yuan Hu, Zhibin Wang, and Fan Wang. Swinrdm: integrate swirrnn with diffusion model towards high-resolution and high-quality weather forecasting. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 322-330, 2023. 1, 3, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 624, + 294, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 294, + 677 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 294, + 677 + ], + "type": "text", + "content": "[12] Lei Chen, Xiaohui Zhong, Feng Zhang, Yuan Cheng, Yinghui Xu, Yuan Qi, and Hao Li. Fuxi: A cascade machine learning forecasting system for 15-day global weather forecast. npj Climate and Atmospheric Science, 6(1):190, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "text", + "content": "[13] J.C. Comiso. Characteristics of arctic winter sea ice from satellite multispectral microwave observations. Journal of Geophysical Research, 91:975-994, 1986. 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "type": "text", + "content": "[14] Josefino C. Comiso. Bootstrap sea ice concentrations from nimbus-7 smmr and dmsp ssm/i-ssmis, version 3. 2017. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 96, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 553, + 128 + ], + "type": "text", + "content": "[15] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 129, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 553, + 183 + ], + "type": "text", + "content": "[16] N. DiGirolamo, C.L. Parkinson, D.J. Cavalieri, P. Gloersen, and H.J. Zwally. Sea ice concentrations from nimbus-7 smmr //and dmsp ssm/i-ssmis passive microwave data. (nsidc-0051, version 1). National Snow and Ice Data Center Distributed Active Archive Center., 1996. 1, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 185, + 553, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 185, + 553, + 248 + ], + "spans": [ + { + "bbox": [ + 317, + 185, + 553, + 248 + ], + "type": "text", + "content": "[17] Ben Fei, Zhaoyang Lyu, Liang Pan, Junzhe Zhang, Weidong Yang, Tianyue Luo, Bo Zhang, and Bo Dai. Generative diffusion prior for unified image restoration and enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9935-9946, 2023. 3, 6, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 251, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 251, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 317, + 251, + 553, + 293 + ], + "type": "text", + "content": "[18] Helge F Goessling, Steffen Tietsche, Jonathan J Day, Ed Hawkins, and Thomas Jung. Predictability of the arctic sea ice edge. Geophysical Research Letters, 43(4):1642-1650, 2016. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 295, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 295, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 317, + 295, + 553, + 348 + ], + "type": "text", + "content": "[19] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 350, + 553, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 350, + 553, + 382 + ], + "spans": [ + { + "bbox": [ + 317, + 350, + 553, + 382 + ], + "type": "text", + "content": "[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 384, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 384, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 317, + 384, + 553, + 426 + ], + "type": "text", + "content": "[21] Yuan Hu, Lei Chen, Zhibin Wang, and Hao Li. Swinvrnn: A data-driven ensemble forecasting model via learned distribution perturbation. Journal of Advances in Modeling Earth Systems, 15(2):e2022MS003211, 2023. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 427, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 427, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 317, + 427, + 553, + 480 + ], + "type": "text", + "content": "[22] Young Jun Kim, Hyun-Cheol Kim, Daehyeon Han, Sanggyun Lee, and Jungho Im. Prediction of monthly arctic sea ice concentrations using satellite and reanalysis data based on convolutional neural networks. The Cryosphere, 14(3): 1083-1104, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 483, + 553, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 483, + 553, + 535 + ], + "spans": [ + { + "bbox": [ + 317, + 483, + 553, + 535 + ], + "type": "text", + "content": "[23] Jussi Leinonen, Daniele Nerini, and Alexis Berne. Stochastic super-resolution for downscaling time-evolving atmospheric fields with a generative adversarial network. IEEE Transactions on Geoscience and Remote Sensing, 59(9):7211-7223, 2020. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 537, + 553, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 537, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 317, + 537, + 553, + 581 + ], + "type": "text", + "content": "[24] Yang Liu, Laurens Bogaardt, Jisk Attema, and Wilco Hazeleger. Extended-range arctic sea ice forecast with convolutional long short-term memory networks. Monthly Weather Review, 149(6):1673-1693, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 582, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 582, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 317, + 582, + 553, + 635 + ], + "type": "text", + "content": "[25] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 637, + 553, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 637, + 553, + 691 + ], + "spans": [ + { + "bbox": [ + 317, + 637, + 553, + 691 + ], + "type": "text", + "content": "[26] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12009-12019, 2022. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 693, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 693, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 693, + 553, + 713 + ], + "type": "text", + "content": "[27] Zhi-Song Liu, Wan-Chi Siu, and Yui-Lam Chan. Photorealistic image super-resolution via variational autoencoders." + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10575" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "IEEE Transactions on Circuits and Systems for video Technology, 31(4):1351-1365, 2020. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 151 + ], + "type": "text", + "content": "[28] Zhi-Song Liu, Wan-Chi Siu, Li-Wen Wang, Chu-Tak Li, and Marie-Paule Cani. Unsupervised real image super-resolution via generative variational autoencoder. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 442-443, 2020. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "text", + "content": "[29] W.N. Meier, F. Fetterer, A.K. Windnagel, and J.S. Stewart. Noaa/nsidc climate data record of passive microwave sea ice concentration. (g02202, version 4). National Snow and Ice Data Center, 2021. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 294, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 241 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 241 + ], + "type": "text", + "content": "[30] Bin Mu, Xiaodan Luo, Shijin Yuan, and Xi Liang. Icetft v 1.0. 0: Interpretable long-term prediction of arctic sea ice extent with deep learning. Geoscientific Model Development Discussions, 2023:1-28, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 242, + 294, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 242, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 56, + 242, + 294, + 275 + ], + "type": "text", + "content": "[31] J.E. Nash and J.V. Sutcliffe. River flow forecasting through conceptual models part i — a discussion of principles. Journal of Hydrology, 10(3):282-290, 1970. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 276, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 330 + ], + "type": "text", + "content": "[32] Xingang Pan, Xiaohang Zhan, Bo Dai, Dahua Lin, Chen Change Loy, and Ping Luo. Exploiting deep generative prior for versatile image restoration and manipulation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):7474-7489, 2021. 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 332, + 294, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 294, + 376 + ], + "type": "text", + "content": "[33] Zisis I Petrou and Yingli Tian. Prediction of sea ice motion with convolutional long short-term memory networks. IEEE Transactions on Geoscience and Remote Sensing, 57 (9):6865-6876, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 377, + 294, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 377, + 294, + 422 + ], + "spans": [ + { + "bbox": [ + 56, + 377, + 294, + 422 + ], + "type": "text", + "content": "[34] Yibin Ren and Xiaofeng Li. Predicting the daily sea ice concentration on a subseasonal scale of the pan-arctic during the melting season by a deep learning model. IEEE Transactions on Geoscience and Remote Sensing, 61:1-15, 2023. 1, 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 422, + 294, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 465 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 465 + ], + "type": "text", + "content": "[35] Y Ren, X Li, and W Zhang. A data-driven deep learning model for weekly sea ice concentration prediction of the panarctic during the melting season, iee t. geosci. remote, 60, 4304819, 2022. 1, 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 468, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 468, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 468, + 294, + 521 + ], + "type": "text", + "content": "[36] Eduardo Rocha Rodrigues, Igor Oliveira, Renato Cunha, and Marco Netto. Deepdownscale: A deep learning strategy for high-resolution weather forecast. In 2018 IEEE 14th International Conference on e-Science (e-Science), pages 415–422. IEEE, 2018. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 523, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 294, + 578 + ], + "type": "text", + "content": "[37] Qi Shao, Guangchao Hou, Wei Li, Guijun Han, Kangzhuang Liang, and Yang Bai. Ocean reanalysis data-driven deep learning forecast for sea surface multivariate in the south china sea. Earth and Space Science, 8(7):e2020EA001558, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 579, + 294, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 294, + 633 + ], + "type": "text", + "content": "[38] S Sparnocchia, N Pinardi, and E Demirov. Multivariate empirical orthogonal function analysis of the upper thermocline structure of the mediterranean sea from observations and model simulations. In Annales Geophysicae, pages 167-187. Copernicus GmbH, 2003. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 635, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 678 + ], + "type": "text", + "content": "[39] Karen Stengel, Andrew Glaws, Dylan Hettinger, and Ryan N King. Adversarial super-resolution of climatological wind and solar data. Proceedings of the National Academy of Sciences, 117(29):16805-16815, 2020. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "text", + "content": "[40] Siwei Tu, Weidong Yang, and Ben Fei. Taming generative diffusion for universal blind image restoration. arXiv preprint arXiv:2408.11287, 2024. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 553, + 251 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 315, + 72, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 553, + 106 + ], + "type": "text", + "content": "[41] Lei Wang, Xiaojun Yuan, and Cuihua Li. Subseasonal forecast of arctic sea ice concentration via statistical approaches. Climate Dynamics, 52:4953-4971, 2019. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "type": "text", + "content": "[42] Qingyu Zheng, Ru Wang, Guijun Han, Wei Li, Xuan Wang, Qi Shao, Xiaobo Wu, Lige Cao, Gongfu Zhou, and Song Hu. A spatio-temporal multiscale deep learning model for subseasonal prediction of arctic sea ice. IEEE Transactions on Geoscience and Remote Sensing, 2024. 3, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "type": "text", + "content": "[43] Xiaohui Zhong, Lei Chen, Jun Liu, Chensen Lin, Yuan Qi, and Hao Li. Fuxi-extreme: Improving extreme rainfall and wind forecasts with diffusion model. Science China Earth Sciences, pages 1-13, 2024. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "type": "text", + "content": "[44] Ruofan Zhou and Sabine Susstrunk. Kernel modeling superresolution on real low-resolution images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2433-2443, 2019. 3" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10576" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Identifying and Mitigating Position Bias of Multi-image Vision-Language Models/17ea1e19-b848-462e-8296-051308ed2dd4_content_list.json b/2025/Identifying and Mitigating Position Bias of Multi-image Vision-Language Models/17ea1e19-b848-462e-8296-051308ed2dd4_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f656d6eb93d0f88f99f8959d0ebd8f0fc052460a --- /dev/null +++ b/2025/Identifying and Mitigating Position Bias of Multi-image Vision-Language Models/17ea1e19-b848-462e-8296-051308ed2dd4_content_list.json @@ -0,0 +1,1659 @@ +[ + { + "type": "text", + "text": "Identifying and Mitigating Position Bias of Multi-image Vision-Language Models", + "text_level": 1, + "bbox": [ + 91, + 130, + 906, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xinyu Tian1 Shu Zou1 Zhaoyuan Yang2 Jing Zhang1 \n1Australian National University 2GE Research", + "bbox": [ + 230, + 179, + 764, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ firstname_lastname@anu.edu.au, $^{2}$ firstname_lastname@ge.com", + "bbox": [ + 240, + 217, + 750, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 268, + 328, + 286 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The evolution of Large Vision-Language Models (LVLMs) has progressed from single to multi-image reasoning. Despite this advancement, our findings indicate that LVLMs struggle to robustly utilize information across multiple images, with predictions significantly affected by the alteration of image positions. To further explore this issue, we introduce Position-wise Question Answering (PQA), a meticulously designed task to quantify reasoning capabilities at each position. Our analysis reveals a pronounced position bias in LVLMs: open-source models excel in reasoning with images positioned later but underperform with those in the middle or at the beginning, while proprietary models show improved comprehension for images at the beginning and end but struggle with those in the middle. Motivated by this, we propose SoFt Attention (SoFA), a simple, training-free approach that mitigates this bias by employing linear interpolation between inter-image causal attention and bidirectional counterparts. Experimental results demonstrate that SoFA reduces position bias and enhances the reasoning performance of existing LVLMs. The code will be available https://github.com/xytian1008/sofa.", + "bbox": [ + 89, + 300, + 485, + 618 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 647, + 223, + 664 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The advent of Large Vision-Language Models (LVLMs) [1, 2, 5, 24, 34] has catalyzed the development of various text-grounded visual applications, including visual question answering [3, 16, 36], code generation [8, 37, 52], and story telling [13, 26]. Open-source LVLMs such as LLaVA [34], BLIP2 [28], and InternVL [11] have demonstrated competitive performance on single-image tasks like VQAv2 [16], OKVQA [38], and MMMU [64]. However, these LVLMs are limited in their capacity to analyze interimage differences and relationships, as well as to process temporal information. Consequently, recent LVLMs, such as Flamingo [2, 4], Idefics [22-24], and Emu2 [51], have incorporated image-text interleaved data during training to enable ability to tackle multi-image tasks [15, 19, 50, 56].", + "bbox": [ + 89, + 672, + 482, + 885 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite the increasing scale and performance of LVLMs,", + "bbox": [ + 109, + 886, + 482, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the interpretation of how they reason across multiple images remains a mystery. In fact, we might find some hints in the topic of natural language processing (NLP). Recent research in NLP has highlighted retrieval-augmented generation (RAG) [25, 44, 71], which aims to enable large language models (LLMs) to reason across multiple documents to answer questions. However, [35] find that LLMs tend to effectively comprehend documents positioned at the beginning and end of the input, often neglecting the middle sections. Similarly, in LLM-as-a-judge [48, 55], [68] find that LLMs used to evaluate two model responses tend to favor the first one, regardless of content. This phenomenon, where the model's reasoning capability is influenced by the segment position, is referred to as position bias [35, 42, 60].", + "bbox": [ + 511, + 270, + 906, + 482 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Motivated by the research above, a natural question arises: does position bias exist in LVLMs? Specifically, we seek to determine whether LVLMs can robustly make use of information across multiple images. To investigate this, we begin by evaluating LVLMs on established multi-image benchmarks [15, 19, 50, 56]. We find that simply changing the order of presented images may significantly alter the predictions for most LVLMs, leading to considerable variation in benchmark accuracy. Fig. 1 provides examples of questions and model responses. For instance, in difference spotting [56], swapping two images causes OpenFlamingo [4] to miss distinctions, while in a simple counting task [19], reordering images leads Idefics2 [24] to incorrect reasoning. This positional sensitivity is apparent even in GPT-4o [1], a leading proprietary model.", + "bbox": [ + 511, + 487, + 908, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The above motivational study reveals an unfortunate fact: building upon LLMs, LVLMs may inherit the notorious position bias from their language component. This bias manifests as the model's tendency to excel at reasoning visual information from certain positions, while exhibiting weaknesses at others, compromising robustness and reliability in predictions. To further investigate which positions display strong understanding and which reveal weaknesses, i.e., identifying the pattern of such position bias, we introduce Position-wise Question Answering (PQA). Unlike most multi-image tasks that assess holistic comprehension, PQA requires LVLMs to reason consistently across each", + "bbox": [ + 511, + 719, + 910, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "10599", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/544f5f38c0745927eb794d5dd9747f98e1aa05cc89c160727c5081df15da5c58.jpg", + "image_caption": [ + "Figure 1. Examples where the predictions of LVLMs are influenced by changes in the image positions (Date accessed: Oct 25, 2024). The tasks, from left to right, are difference spotting [56], counting [19] and forensic detection [15], respectively. The correct answers are highlighted in green, while the incorrect ones are highlighted in red. The prompts and outputs are simplified here for clarity and aesthetics." + ], + "image_footnote": [], + "bbox": [ + 93, + 87, + 906, + 281 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "image and produce position-wise responses. This approach enables us to measure position-wise accuracy and analyze the model's reasoning capacity across different positions.", + "bbox": [ + 89, + 347, + 482, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our analysis reveals that most open-source models, e.g., OpenFlamingo [4] and Idefics2 [24], exhibit a pronounced recency bias, demonstrating stronger reasoning capabilities for images presented later while underperforming on those in the middle or at the beginning. In contrast, proprietary models like GPT-4o [1] show enhanced understanding of images at the beginning and end, while neglecting those in the middle, a pattern consistent with previous findings in LLMs [35]. This observation is further corroborated by Fig. 1, where LVLMs tend to reason incorrectly about images positioned in poor-performing areas.", + "bbox": [ + 89, + 397, + 483, + 564 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The existence of position bias significantly compromises model robustness and undermines accuracy across various multi-image tasks. Inspired by this, we aim to investigate the underlying mechanical causes and mitigate this issue. Specifically, we notice that in most LVLMs, the hidden states of image tokens are heavily influenced by their relative positions due to inter-image causal attention. In particular, image tokens are treated as a type of special text tokens where 1) images appearing later may interact with preceding ones; 2) images at the beginning are isolated, lacking access to global context. This autoregressive setup implicitly imposes a semantic order on the presented images, amplifying their positional information. To address this, we introduce SoFt Attention (SoFA), a training-free approach to mitigate position bias across various LVLMs. SoFA reduces the dependency of image hidden states on positional information by simply performing linear interpolation between inter-image causal attention and its bidirectional counterparts. Experimental results indicate that SoFA effectively equilibrates the reasoning capabilities of LVLMs across various positions, achieving average accuracy improvements of $2 \\sim 3\\%$ on existing multi-image benchmarks.", + "bbox": [ + 89, + 569, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are as follows:", + "bbox": [ + 511, + 347, + 813, + 361 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Despite the significant potential of LVLMs in multi-image applications, we find that position bias, a well-explored phenomenon in NLP, has emerged as a potential bottleneck in the development of LVLMs, substantially impacting model robustness and reliability.", + "- We further identify such position bias and discover that most open-source models demonstrate a pronounced recency bias, whereas proprietary models tend to struggle in middle positions.", + "- We introduce SoFA, a straightforward training-free approach designed to mitigate the position bias inherited from pre-trained LVLMs, and substantiate its effectiveness through experimental validation." + ], + "bbox": [ + 513, + 362, + 903, + 558 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 571, + 653, + 588 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Large Vision-Language Models. Existing LVLMs are generally composed of a vision encoder [43, 65], an LLM [12, 53, 54], and a bridging projector [2, 7, 27] that facilitates interaction between the two. Conventional LVLMs, pre-trained on image-text pairs [46] and fine-tuned through preference [33, 40] or instruction tuning [27, 41], demonstrate substantial success in applications like dialogue [5, 11], question answering [58, 69], and domain-specific tasks [39, 47]. However, constrained by their training data, these models primarily support single-image reasoning during inference. Recently, the integration of image-text interleaved data [23, 70] has enabled multi-image reasoning capabilities, encompassing image comparison [21, 56], association [15, 19], temporal relationship analysis [29], and emerging properties of multi-modal in-context learning [30, 67]. Nonetheless, limited research has delved into how LVLMs perform reasoning across multiple images or the underlying mechanisms driving this process. Position Bias in LLMs. Position bias, an inherent side effect of long-context input, has increasingly captured the at", + "bbox": [ + 511, + 598, + 906, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10600", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/181e6ae6c93931b3eeb9d154c5f1e22c5557bb7f10789500efd59f45f2ce356e.jpg", + "image_caption": [ + "Figure 2. The results of multiple evaluations while shuffling image positions. We record the minimum, maximum, and average accuracy (left vertical axis), along with the prediction inconsistency between the best and worst-performing evaluations (right vertical axis)." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 903, + 214 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tention of NLP researchers. The pioneering work on this issue, notably [35], reveals that when LLMs process multiple documents, they often concentrate on those at the beginning and end while disregarding much of the information in between, i.e., a phenomenon termed \"lost in the middle\". A related issue has also been identified in LLM-as-a-judge [48, 55], where [68] observe that LLMs tasked with evaluating multiple model responses tend to favor the first one, irrespective of its content. Recent studies have begun to explore methods to mitigate or circumvent position bias in LLMs [42, 45, 60, 63]. However, to the best of our knowledge, few studies have examined the effects of position bias within multi-modal applications, specifically regarding its implications for LVLMs.", + "bbox": [ + 88, + 263, + 485, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Position Bias in LVLMs. While systematic research on position bias in LVLMs is still lacking, recent studies have indirectly underscored the influence of positions on LVLMs through various findings. For instance, in multi-modal in-context learning, studies [9] and [6] reveal that demonstration images placed at the beginning contribute minimally to LVLM performance. Additionally, in assessing long-context capabilities, [61] reports that LVLMs attain particularly high accuracy when key information appears at the end of the input sequence, with similar results observed in [59] and [57]. However, previous research has neither thoroughly analyzed this position bias nor proposed solutions to address it. In contrast, this study systematically investigates position bias in both open-source and proprietary models and introduces an effective mitigation strategy.", + "bbox": [ + 89, + 478, + 483, + 705 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Understanding Position Bias in LVLMs", + "text_level": 1, + "bbox": [ + 89, + 722, + 444, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we investigate whether the position bias observed in LLMs extends to LVLMs. We start by empirically confirming the existence of position bias in LVLMs through a motivational study, revealing their poor robustness to image position changes (§3.1). Subsequently, we analyze the nature of this bias by evaluating and quantifying the reasoning capabilities of LVLMs regarding different positions (§3.2). These insights contribute to a deeper understanding and interpretation of LVLMs' reasoning behavior and common failure cases.", + "bbox": [ + 89, + 750, + 483, + 898 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. The Position Matters", + "text_level": 1, + "bbox": [ + 511, + 263, + 712, + 279 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To assess the existence of position bias, we conduct a simple motivational experiment, as shown in Fig. 2.", + "bbox": [ + 511, + 286, + 903, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Experimental Setup. We select four representative multi-image LVLMs: OpenFlamingo [4], Idefics2 [24], LLaVA-NeXT-Interleave1 [27], GPT-4o [1], and conduct evaluations across established multi-image benchmarks, including BLINK [15], Mantis-Eval [19], MuirBench [56], and MIRB [66]. To ensure clarity, we select partial tasks from these benchmarks that feature position-agnostic questions, wherein the positions of the images do not affect the answers to the questions. See §5.1 for the selected tasks and details. To explore the influence of image positions on the results, we perform multiple evaluations per benchmark, with the positions of the images shuffled in each iteration.", + "bbox": [ + 511, + 316, + 905, + 498 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Metric. We record the minimum, maximum, and average accuracy across these evaluations, which directly reflects the influence of positions on the performance of LVLMs. Furthermore, we assess the prediction inconsistency between the best and worst-performing evaluations, quantifying the proportion of examples exhibiting conflicting predictions across the two evaluations by changing positions.", + "bbox": [ + 511, + 500, + 903, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Results. As illustrated in Fig. 2, for most open-source LVLMs, the prediction inconsistency between the best and worst-performing evaluations approaches $30\\%$ , suggesting that merely altering the positions of the presented images may result in an amount of the model's predictions being changed. Such pronounced sensitivity to positional information results in highly unstable performance. For example, OpenFlamingo's accuracy on BLINK can vary significantly, reaching $38\\%$ at its peak and dropping to $30\\%$ at its lowest, indicating an accuracy span of up to $10\\%$ with position changes. GPT-4o, the state-of-the-art proprietary model, although showcasing superior performance, also displays a lack of robustness to positional variations, leading to a prediction inconsistency of around $25\\%$ .", + "bbox": [ + 511, + 604, + 905, + 815 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To qualitatively explore this issue, we present examples in Fig. 1 that illustrate prediction conflicts arising from alterations in image positions. For instance, when prompted to identify the differences between two images, Open-", + "bbox": [ + 511, + 816, + 903, + 878 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "1For brevity, we refer to LLaVA-NeXT-Interleave as LLaVA-NeXT.", + "bbox": [ + 531, + 886, + 890, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10601", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/77d93e66f4813cf90930cad4924e06f53ba8952058b77d12b00d011c9d6544a7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 88, + 303, + 213 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/39eddef4ed52608f0464f7bf9d90a1598601cc958984136679f48fda3834dce8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 310, + 87, + 504, + 213 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a13d5253052a0753a3be84751bde8a204485818747cf8ba13a2a370588e24dc4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 88, + 704, + 213 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/39f202d2bf24294a94d1e561929a75953d8a6e91798094ee0c9923cc90cc7b8e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 88, + 903, + 213 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e648dd5b0b16ce27c0de8461b5e20b4b1ae82f22893e4c098b4823dcdc0c4ea8.jpg", + "image_caption": [ + "Figure 3. The results on the PQA task. We report position-wise accuracy on four scenarios where the number of images is 5, 10, 15 and 20, respectively. A higher accuracy signifies strong reasoning at the specified position, whereas lower reflects poor-performing areas." + ], + "image_footnote": [], + "bbox": [ + 200, + 223, + 318, + 236 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/028a2c900e079d23a835b08d56212d9e899da15d69cda61c12a58f0ef8bd6ca5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 334, + 223, + 416, + 234 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/85791aa4be6580e62ea44bc5f0d9c0f705e6f140e452e37f4a9544776ce24477.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 437, + 223, + 550, + 234 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/351071bd03901fddff5c3585164da708ea4d291717f9011d5b1a3971f99e3521.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 568, + 223, + 648, + 234 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5e32f62575144608556a8130ccd261f44c99783fd4ec9db7cfef37144574e644.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 666, + 223, + 789, + 234 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Flamingo fails to extract fine-grained distinctions upon swapping their positions. Similarly, in the task of object counting, Idefics2 mistakenly identifies a dog as a cat due to the repositioning of images. GPT-4o demonstrates a comparable issue, where rearranging the image options alters both its answer and the accompanying rationale.", + "bbox": [ + 89, + 287, + 483, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Identifying Position Bias", + "text_level": 1, + "bbox": [ + 89, + 387, + 316, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The quantitative results and qualitative examples above imply biased reasoning across positions from LVLMs. To have a closer look at such position bias, we aim to identify models' reasoning capabilities at different image positions.", + "bbox": [ + 89, + 409, + 483, + 469 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, existing tasks have been restricted to evaluating models' holistic comprehension of multiple images, lacking the means to systematically analyze their fine-grained understanding of each individual position. Specifically, these tasks provide coarse ground truth answers, such as option symbols or numeric values, but do not offer individual rationales for each image, making it challenging to quantify the model's reasoning across different positions. To address this gap, we introduce Position-wise Question Answering (PQA), a carefully designed task that assesses the position-wise reasoning capabilities of LVLMs. In PQA, each question prompts LVLMs to reason independently across all images, generating a distinct response for each. This approach 1) ensures task-level neutrality with respect to any particular position and 2) enables quantification of LVLMs' reasoning performance for each position.", + "bbox": [ + 89, + 469, + 483, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, taking object counting as an example, we employ the following query template:", + "bbox": [ + 89, + 710, + 482, + 742 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "I: Please provide one answer for each image in the form like $[x, x, x, \\ldots]$ . \nQ: Among , ..., , how many cats can you find in these images? \nA: [3, 2, 0, ...].", + "bbox": [ + 89, + 748, + 482, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Each query consists of an instruction (I) and a question $(\\mathbb{Q})$ where the former prompts the model to generate position-wise responses (A) following a specified output format. We gather responses from LVLMs as a structured list, with each element representing a particular position. This enables us", + "bbox": [ + 89, + 825, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "to track position-wise accuracy, where a higher accuracy reflects stronger model capability at a given position, while lower indicates areas of relative weakness.", + "bbox": [ + 511, + 287, + 903, + 332 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Construction of position-wise examples. Since in PQA, each image is semantically independent of others, a PQA example essentially consists of several single-image instances that share the same question. This allows us to easily build position-wise examples from single-image benchmarks. Specifically, we select VQAv2 [16], a substantial single-image dataset that encompasses a diverse array of everyday topics and features both multiple-choice and open-ended questions. Initially, we identify and consolidate single-image examples with identical questions to create a question set, wherein each question corresponds to multiple images. Subsequently, we reformulate these questions into PQA format according to the above-mentioned template.", + "bbox": [ + 511, + 335, + 905, + 531 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Experimental Setup. We utilize the same LVLMs as previously mentioned and conduct evaluations on the proposed PQA task. Additionally, we also consider random guessing, where binary choice and numeric questions are answered randomly with yes or no, or with an integer value. We examine four scenarios with varying numbers of images: 5, 10, 15, and 20, collecting 1,000 examples for each scenario. Furthermore, to maintain fairness, we augment each example by shuffling the image positions four times, resulting in a total of 5,000 examples. In contrast to prior settings that evaluate holistic performance, we provide separate accuracy metrics for each position in this analysis.", + "bbox": [ + 511, + 534, + 905, + 715 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Results. Fig. 3 displays the results of LVLMs on the PQA task, from which we may derive the following insights:", + "bbox": [ + 511, + 717, + 903, + 747 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "1) Open-source LVLMs exhibit recency bias, whereas proprietary models struggle in middle positions. For open-source models, a stronger reasoning capability is observed for images positioned at the end, while significant underperformance is noted in the middle and beginning. For example, with 10 images, the accuracy of OpenFlamingo varies by nearly $9\\%$ between the far left and far right positions $(51.84\\% \\rightarrow 59.71\\%)$ . In contrast, proprietary models, i.e., GPT-40, reveal a typical U-shaped curve, highlighting pronounced weaknesses in the middle positions.", + "bbox": [ + 511, + 750, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "10602", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "2) Position bias becomes more pronounced as the number of images increases. For instance, with 20 images, the accuracy gap for OpenFlamingo between the two ends reaches $14\\%$ ( $36.79\\% \\rightarrow 50.17\\%$ ). Notably, at this stage, OpenFlamingo seems to nearly lost its reasoning capability for most of the beginning images, achieving results only marginally better than random guessing ( $34.49\\%$ ).", + "3) Relationship to position bias in LLMs. Previous research [35] shows that LLMs often overlook information in middle positions when processing multiple documents, a phenomenon termed \"lost in the middle\". Our findings suggest this effect persists for GPT in multi-image reasoning. However, for open-source models, the primacy position advantage appears to diminish, showing a marked decaying trend from the end to beginning. Whether the position bias in LVLMs is correlated with that of their LLM backbones remains an interesting question for further exploration." + ], + "bbox": [ + 89, + 90, + 480, + 348 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. The Proposed Method", + "text_level": 1, + "bbox": [ + 89, + 361, + 302, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we first analyze the underlying causes of position bias (§4.1). Then we introduce SoFA, a straightforward, training-free approach to mitigate this issue (§4.2).", + "bbox": [ + 89, + 386, + 482, + 433 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Position Mechanism in LVLMs", + "text_level": 1, + "bbox": [ + 89, + 441, + 366, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To investigate the origins of position bias, we need to first determine what contributes to the positional information in LVLMs. In NLP, it is generally believed that position embeddings [45, 49] and causal attention [42, 60] are the primary mechanisms enabling the sequential structure of input. Formally,", + "bbox": [ + 89, + 463, + 482, + 554 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} = \\mathcal {P} \\left(\\mathbf {X} \\mathbf {W} _ {\\mathbf {q}}, \\mathbf {p o s}\\right), \\quad \\mathbf {K} = \\mathcal {P} \\left(\\mathbf {X} \\mathbf {W} _ {\\mathbf {k}}, \\mathbf {p o s}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 564, + 482, + 590 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {H} = \\operatorname {S o f t m a x} \\left(\\mathbf {Q K} ^ {\\mathrm {T}} / \\sqrt {d}\\right) \\odot \\mathbb {1} _ {\\text {c a u s a l}} \\mathbf {V}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 585, + 380, + 602 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{Q}$ , $\\mathbf{K}$ , and $\\mathbf{V}$ represent the queries, keys, and values, respectively. $\\mathbf{X}$ refers to the input tokens, encompassing both image and text types, pos denotes the position embedding of $\\mathbf{X}$ , which is integrated into the attention via the position encoding function $\\mathcal{P}$ , $\\mathbb{1}_{\\mathrm{causal}}$ denotes the causal mask, which ensures autoregressive generation.", + "bbox": [ + 89, + 613, + 482, + 704 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This reveals two key factors influencing LVLMs' positional awareness: 1) position embedding explicitly initializes a sequential value to each token, and 2) causal attention implicitly enhances this positional characteristic through unidirectional information access. This presents us with two distinct solutions. One approach is to directly reassign position embeddings, a widely used strategy in NLP [10, 18, 45], as it fundamentally eliminates the influence of position on documents. However, this aggressive method is not suitable for LVLMs, since in many scenarios, such as video understanding and action recognition, we still expect images to retain positional information for temporal analysis. To clarify, our aim is to mitigate the imbalanced", + "bbox": [ + 89, + 704, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9ee0e20336200991d8dcd43aba63d83b448ed3b7a1e38bc0a19721f392c90721.jpg", + "image_caption": [ + "(A) Causal Attention" + ], + "image_footnote": [], + "bbox": [ + 516, + 88, + 624, + 172 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/954e46ff412889d6a74d516d7464db5e1041352826b2dbfc22ba8d57adbf24b9.jpg", + "image_caption": [ + "(B) Isolated Attention" + ], + "image_footnote": [], + "bbox": [ + 651, + 88, + 761, + 172 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a588def4b1cc8792d1edc54a90616a2a265609fef01b76f8ed1b2c324feaa086.jpg", + "image_caption": [ + "(C) Bidirectional Attention", + "Figure 4. The three inter-image attention mechanisms, where $I_{1}$ and $I_{2}$ represent the tokens of two images, $T_{1}$ and $T_{2}$ represent the tokens of two text segments. Note that typically each image involves numerous tokens, e.g., 576 for LLaVA. Here for clarity we simplify them to a single token. In (A), images interact in a unidirectional manner, allowing $I_{2}$ to attend to $I_{1}$ , while $I_{1}$ remains isolated. In (B), each image is isolated, indicating they can only attend to themselves. In (C), bidirectional interaction is enabled so that each image can attend to any other images. It is worth mentioning that we only alter the inter-image attention while preserving causal attention between the text segments." + ], + "image_footnote": [], + "bbox": [ + 785, + 88, + 897, + 172 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "reasoning capabilities of LVLMs across different positions, rather than to eliminate their positional awareness.", + "bbox": [ + 511, + 364, + 903, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Therefore, in this paper, we primarily investigate the relationship between causal attention and position bias. We notice that, while images do not exhibit a distinct semantic order as text does, most LVLMs, e.g., LLaVA series, treat image and text tokens in a uniform autoregressive manner. As shown in Fig. 4 (A), analogous to texts, there exists interimage causal attention which allows 1) later-positioned images to interact with those preceding them, while 2) earlier images remain isolated, lacking access to global contexts. From a mechanical perspective, this framework renders image hidden states highly dependent on their relative positions, thus implicitly amplifying their positional cues.", + "bbox": [ + 511, + 396, + 905, + 577 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To verify our hypothesis, we compare the inter-image causal attention against following two alternative variants:", + "bbox": [ + 511, + 579, + 903, + 609 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1) Isolated attention. As illustrated in Fig. 4 (B), we disable interactions between images, enforcing each image to attend solely to itself.", + "2) Bidirectional attention. Rather than restricting to unidirectional information flow, we permit each image to attend to any other, as shown in Fig. 4 (C)." + ], + "bbox": [ + 511, + 609, + 903, + 702 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We modify the default causal mask in the language decoder to its isolated and bidirectional counterparts and assess the impact of them on position bias in PQA, as shown in Fig. 5. It is evident that altering causal attention to either isolated or bidirectional forms significantly alleviates position bias, resulting in more balanced accuracy across different positions. This empirically confirms that inter-image causal attention is the primary source of position bias. However, these modifications lead to varying degrees of performance degradation, with isolated attention causing the most substantial decline. This outcome is reasonable, as such modifications may substantially deviate from the training framework, leading to out-of-distribution hidden states.", + "bbox": [ + 511, + 704, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "10603", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/33dc7e2e838ab5a2d2cd8bc50703beb61152d00f69180c48ff0582bb9ab7a12a.jpg", + "image_caption": [ + "(A) Idefics2", + "Figure 5. The position-wise accuracy of the three attention mechanisms on the PQA task, alongside our proposed SoFA method." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 289, + 220 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a77276c3fed29e2dfe69f3f12d0393328d128512c30a1c5591aeebed487915dd.jpg", + "image_caption": [ + "(B) LLaVA-NeXT" + ], + "image_footnote": [], + "bbox": [ + 290, + 88, + 480, + 220 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Mitigating Position Bias", + "text_level": 1, + "bbox": [ + 89, + 289, + 312, + 306 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The analysis above shows that inter-image causal attention is the main cause of position bias. Furthermore, it also reveals a trade-off between accuracy and robustness, a typical dilemma in machine learning. Building on this insight, we propose SoFt Attention (SoFA), a straightforward, training-free approach to mitigate position bias. SoFA employs linear interpolation between inter-image causal attention and its bidirectional counterpart. Formally,", + "bbox": [ + 89, + 311, + 483, + 433 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {H} _ {\\text {s o f t}} = \\operatorname {S o f t m a x} \\left(\\mathbf {Q K} ^ {\\mathrm {T}} / \\sqrt {d}\\right) \\odot \\mathbf {M} _ {\\text {s o f t}} \\mathbf {V} \\tag {2} \\\\ \\mathbf {M} _ {\\text {s o f t}} = (1 - \\sigma) \\mathbb {1} _ {\\text {c a u s a l}} + \\sigma \\mathbb {1} _ {\\text {b i d i r e c t i o n a l}} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 444, + 482, + 481 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbb{1}_{\\mathrm{causal}}$ and $\\mathbb{1}_{\\mathrm{bidual}}$ denote the two attention masks, with $\\sigma$ controlling the extent of bidirectional part. Note that since we are only interested in the visual positional relation, we only modify the inter-image attention, while preserving causal attention between text segments.", + "bbox": [ + 89, + 492, + 482, + 568 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SoFA as a position smoother. Intuitively, given that interimage causal attention excessively emphasizes positional information, SoFA integrates partial bidirectional attention to smooth the influence of position on the models at minimal cost, achieving a balanced trade-off between reasoning performance and position robustness.", + "bbox": [ + 89, + 568, + 482, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SoFA as an interaction enhancer. Practically, SoFA enhances the interaction between images, enriching the semantic information of earlier image features to enable more effective utilization by LVLMs. This aligns with the most cases where key information is typically distributed evenly across various input positions.", + "bbox": [ + 89, + 659, + 482, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For each task, we utilize a small validation set to determine the optimal values for $\\sigma$ . Besides, to better align with the training framework, we deploy SoFA every two layers rather than at every layer, while preserving causal attention in the remaining layers. We present SoFA's results in the PQA task as shown in Fig. 5. It is evident that SoFA significantly outperforms both isolated and bidirectional attention, while unlike causal attention in the default setting, SoFA demonstrates balanced performance and exhibits enhanced reasoning capability in most positions.", + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiment and Results", + "text_level": 1, + "bbox": [ + 511, + 90, + 740, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we empirically demonstrate that SoFA 1) effectively mitigates the position bias inherent in LVLMs and 2) significantly improves their overall reasoning capability across various multi-image applications.", + "bbox": [ + 511, + 114, + 903, + 175 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Models. We select a diverse set of open-source models, including Idefics2-8B [24], InternVL2-8B [11], VILA-13B [31], Mantis-8B [19], and LLaVA-NeXT-Interleave-7B [27].", + "bbox": [ + 511, + 175, + 903, + 234 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. For the input, we provide instructions that explicitly indicate the task type and output formats. For instance, for multiple-choice questions, we ask LVLMs to provide only the option symbol; for numeric questions, only the number; and for open-ended questions, the response is limited to short answers. For evaluation, to ensure fairness, we uniformly adopt GPT-4 [1], providing with the question, ground truth answer, candidate prediction, and it determines whether the prediction is correct or not. For SoFA, we determine $\\sigma$ using a 32-shot validation set per task and perform interpolation every two layers. Unless otherwise specified, all models are set to FP16 precision and utilize Flash Attention [14], with sub-image splitting [32] disabled for fairness.", + "bbox": [ + 511, + 236, + 906, + 448 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Position Robustness", + "text_level": 1, + "bbox": [ + 511, + 455, + 702, + 469 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Benchmark. We first evaluate whether SoFA may effectively mitigate position bias. To this end, similar to §3.1, we select position-agnostic tasks from existing multi-image benchmarks. Specifically, we exclude tasks where changing the order of images may alter the semantics of the question, rendering it unanswerable, e.g., video understanding. Here, we list some of the selected tasks: forensic detection and image jigsaw from BLINK [15], difference spotting and diagram understanding from MuirBench [56], visual analogy and attribute matching from MIRB [66]. We include all examples from Mantis-Eval [19] since the majority are position-agnostic. See the supplementary material for the details of more evaluated tasks. Our experimental setup and metrics are consistent with those described in §3.1.", + "bbox": [ + 511, + 477, + 905, + 688 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SoFA reduces the influence of positions on LVLMs. As demonstrated in Table 1, the integration of SoFA within the models markedly reduces prediction inconsistency, leading to enhanced stability in performance. For instance, Idefics2's inconsistency on BLINK declines from $41.55\\%$ to $12.36\\%$ , signifying that the responses for the majority of examples remain consistent despite variations in image positioning. This observation indicates that LVLMs exhibit more balanced reasoning capabilities across different positions, effectively mitigating the effects of position bias.", + "bbox": [ + 511, + 689, + 905, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SoFA eliminates the blind spots of LVLMs. In addition to the enhanced robustness, we observe a marked improvement in the performance of LVLMs in worst-case scenarios, particularly regarding minimum and average accuracy,", + "bbox": [ + 511, + 839, + 905, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "10604", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/67c6d57bc6564a00c3c00999f574970fe6668258306ca453b51ba3581fefed35.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBLINK [15]Mantis-Eval [19]MuirBench [56]MIRB [66]
MinAveMaxIncMinAveMaxIncMinAveMaxIncMinAveMaxInc
Idefics2 [24]43.2646.9348.6841.5543.9346.6251.1038.5743.2845.3047.8134.4645.2647.8849.6829.94
Idefics2 + SoFA47.1848.6349.1212.3648.7149.1850.838.6146.1247.5148.4513.2747.3948.3649.136.99
InternVL2 [11]38.8140.4543.7430.1845.3646.9249.5125.2449.2852.3356.1038.6541.8444.3846.3629.60
InternVL2 + SoFA41.6442.3243.517.1648.1349.2550.1112.0054.6955.9256.785.1644.2045.3545.876.64
VILA [31]45.9348.5951.4225.3348.4449.2051.8521.2941.6643.1248.2737.2647.1749.5952.3431.77
VILA + SoFA48.2750.8051.1710.6849.2951.6052.6812.1645.7646.5247.137.9248.2251.4351.9517.46
Mantis [19]48.3449.2451.5228.3556.4058.3863.4232.1247.6748.9451.1527.4553.1155.7157.4225.79
Mantis + SoFA49.2250.8752.3416.2360.4862.2164.6814.3049.8850.2650.795.6154.3956.3456.938.35
LLaVA-NeXT [27]49.3452.4055.4331.8954.5759.1062.2933.7551.9653.4555.7824.9255.4958.4060.2828.56
LLaVA-NeXT + SoFA53.2054.9156.0811.2058.8761.2361.9814.2053.6155.3756.499.1558.1359.8161.166.96
", + "bbox": [ + 91, + 88, + 903, + 280 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/94a53f18c9e8a5f84ccaacfaa178ebafd19d14929bd45a80a8865fa43fd29f15.jpg", + "table_caption": [ + "Table 1. The evaluation on position-agnostic tasks with and without SoFA. Similar to §3.1, we perform multiple evaluations and record minimum, maximum and average accuracy. We also report prediction inconsistency between best and worst-performing evaluations." + ], + "table_footnote": [], + "table_body": "
MethodBLINKM-EvalMuirMIRBNLVR2MVBench
Random Guessing39.2725.3121.4919.1350.7525.91
Idefics2 [24]45.7848.4142.8446.5881.9423.37
Idefics2 + SoFA48.3150.7747.3648.8385.1525.65
InternVL2 [11]38.9550.3054.5342.6685.5629.31
InternVL2 + SoFA43.2651.1157.1446.1988.1932.77
VILA [31]49.7952.6448.7350.3678.8225.81
VILA + SoFA50.2952.1352.8051.1580.3426.40
Mantis [19]51.4156.8544.2854.5990.2641.59
Mantis + SoFA53.1959.2347.6455.6889.9844.40
LLaVA-NeXT [27]53.3450.8348.2257.1587.2854.26
LLaVA-NeXT + SoFA55.9254.5150.4360.6789.4557.71
", + "bbox": [ + 91, + 328, + 486, + 520 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. The overall evaluation accuracy of LVLMs on various multi-image benchmarks with and without SoFA.", + "bbox": [ + 89, + 530, + 482, + 560 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "which both exhibit an increase of approximately $3\\%$ . This suggests that the model retains strong performance despite variations in the distribution of image positions, indicating that SoFA effectively addresses the deficiencies of LVLMs in positions where performance is typically suboptimal.", + "bbox": [ + 89, + 566, + 482, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Overall Results", + "text_level": 1, + "bbox": [ + 89, + 652, + 246, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Benchmark. Given the observation that SoFA effectively mitigates position bias, we proceed with a comprehensive evaluation to determine whether SoFA can enhance the overall reasoning capabilities of LVLMs. Specifically, we include all tasks from the selected benchmarks [15, 19, 56, 66]. In addition, we incorporate NLVR2 [50] and MVBench [29], which exclusively assess the models' capability in image-pair comparison and video understanding, respectively. We directly utilize the default image positions from the benchmarks and record their accuracy.", + "bbox": [ + 89, + 672, + 482, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "SoFA enhances reasoning capability of LVLMs. In Table 2, we observe that integrating SoFA leads to an average improvement of $2 \\sim 3\\%$ in the performance of LVLMs across various benchmarks. For instance, Idefics2 shows a $2.25\\%$ improvement on MIRB (46.58% → 48.83%),", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "while InternVL2 exhibits a $4.31\\%$ increase on BLINK $(38.95\\% \\rightarrow 43.26\\%)$ . This demonstrates that, despite the advancements in multi-image LVLMs, position bias poses a significant obstacle to their reasoning capability, whereas SoFA successfully addresses this limitation.", + "bbox": [ + 511, + 330, + 906, + 407 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Discussion", + "text_level": 1, + "bbox": [ + 511, + 421, + 632, + 436 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Attention distribution visualization. To gain a deeper understanding of SoFA's working mechanism, we visualize the attention distribution within LVLMs across different positions. Using PQA as an example, we derive the average attention value of the final layer's query token (whose hidden state is directly used for generation) for each image. As shown in Fig. 6, under the default setting, the attention distribution is dominated by the end positions, indicating that the model overly focuses on a few images while neglecting most others. This observation aligns with our findings in §3.2. The application of SoFA alters this trend by significantly smoothing the attention distribution, enabling the model to attend to images at all positions more evenly. This well explains how SoFA mitigates position bias.", + "bbox": [ + 511, + 445, + 906, + 656 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "SoFA improves in-context capability of LVLMs. Multimodal in-context learning is an emerging property recently acquired by LVLMs, wherein a few demonstrations allow them to recognize task formats, thereby significantly improving accuracy. However, recent studies [6, 9, 30] find that LVLMs tend to overlook demonstration images that appear earlier, leading to suboptimal performance. Here, we demonstrate that SoFA can improve the in-context learning capability of LVLMs by eliminating such blind spots. Following [30], We select three typical visual question answering benchmarks: VQAv2 [16], VizWiz [17], and OKVQA [38], and report the average accuracy of LVLMs across varying shot counts. As illustrated in Table 3, the application of SoFA leads to a further enhancement in LVLM performance, with this effect becoming increasingly pronounced as the number of shots rises. For instance, in the", + "bbox": [ + 511, + 657, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "10605", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a7f2cbbd62699c6b45a7bef744223616aa47aba15410d847c57b1da62ea8c5ae.jpg", + "image_caption": [ + "Figure 6. The attention distribution across positions on PQA." + ], + "image_footnote": [], + "bbox": [ + 102, + 90, + 472, + 224 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/824dcc3879ba9662972f9bca09375c0e417a86ebd0769e17bb13668493139531.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodVQAv2 [16]VizWiz [17]OK-VQA [38]
4-shot8-shot16-shot4-shot8-shot16-shot4-shot8-shot16-shot
w/o SoFA55.3459.3261.3741.5644.6545.3545.1948.3049.02
w/ SoFA56.1361.3064.8542.3046.9849.1745.9349.8752.36
", + "bbox": [ + 91, + 261, + 483, + 327 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e2484e051e858c3da715917df8fa8bdaf56d1733ff45d8acf3354f6c2a3a1c14.jpg", + "table_caption": [ + "Table 3. The average accuracy of LVLMs on visual question answering using in-context learning with and without SoFA." + ], + "table_footnote": [], + "table_body": "
# Images351020305075100
w/o SoFA82.1681.2777.1872.3668.5060.8055.1349.19
w/ SoFA83.9082.3378.9274.3070.6763.2759.3855.11
", + "bbox": [ + 102, + 377, + 472, + 428 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. The average accuracy of LVLMs on Multi-modal Haystack [59], where images range from a few to a hundred.", + "bbox": [ + 89, + 440, + 482, + 468 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "case of VizWiz, accuracy improves from $41.56\\%$ to $42.30\\%$ with 4 shots, and from $45.35\\%$ to $49.17\\%$ with 16 shots. This indicates that position bias undermines the effectiveness of demonstrations, while SoFA mitigates this issue.", + "bbox": [ + 89, + 477, + 482, + 537 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "SoFA exhibits advantages in long-context applications. Long-context scenarios have recently become a focus, where a single inference requires processing a substantial number of images. This is particularly beneficial for applications such as many-shot in-context learning [20] and multi-modal RAG [62]. Here, we show that SoFA can excel in these challenging scenarios. We select Visual Haystack [61], a benchmark designed to assess the long-context capabilities of LVLMs, and increase the input scale to encompass anywhere from a few images to a hundred. As shown in Table 4, the benefits of SoFA become increasingly evident as the number of images rises, achieving peak performance gains at 100 images (49.19% → 55.11%). This indicates that an increase in context length exacerbates the model's position bias, while SoFA compensates this loss.", + "bbox": [ + 89, + 537, + 482, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Contribution of SoFA to different tasks. Extending beyond the holistic evaluation, we categorize the benchmarks into various tasks to identify which ones benefit the most from SoFA. As depicted in Fig. 7, SoFA demonstrates the highest improvements in visual retrieval and analogy, with gains of $6.84\\%$ and $5.53\\%$ , respectively. In these tasks, the first image acts as a reference, and the models need to select a matching scene or similar style from the remaining images. This necessitates that the model pays extra attention", + "bbox": [ + 88, + 763, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9e6db53c170794c570aee95c360d8bb9f82bd3d31e2c300b24d7a531cbfb23f0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 526, + 88, + 895, + 203 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5b12e6e9bfe168598e7942ce232132937cc211feb8fb9a17ee9922c4c7da2d4a.jpg", + "table_caption": [ + "Figure 7. The performance gains of SoFA on different types of tasks. The results are averaged over selected models." + ], + "table_footnote": [], + "table_body": "
σBLINKMantis-EvalMuirBenchMIRBNLVR2MVBench
0.0047.8551.8147.7250.2784.7734.87
0.2549.2252.7049.3552.3485.4137.28
0.5050.0453.3350.9651.8686.3036.56
0.7549.6253.7448.2451.1486.8335.07
1.0048.2151.5847.3750.5385.3734.20
", + "bbox": [ + 521, + 251, + 898, + 340 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. The results on six benchmarks varying the value of $\\sigma$ .", + "bbox": [ + 522, + 351, + 893, + 364 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "to and adequately understands the first image. Intuitively, due to the recency bias inherent in open-source models, they often overlook earlier information, resulting in suboptimal performance. Thus, the significant gains brought by SoFA are understandable.", + "bbox": [ + 511, + 378, + 903, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The interpretation of $\\sigma$ . In Table 5, we conduct an ablation study on the effect of $\\sigma$ . It can be observed that the optimal $\\sigma$ varies significantly across benchmarks, for instance, with a value of 0.75 on NLVR2 and only 0.25 on MVBench. Intuitively, MVBench inputs involve video frames, which have a strong dependence on image positions, and an excessively high $\\sigma$ may compromise such positional information. In contrast, the examples in NLVR2 are image pairs, most of which are position-agnostic. This difference in task nature accounts for the variation in $\\sigma$ .", + "bbox": [ + 511, + 458, + 906, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 630, + 633, + 646 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this study, we introduce the concept of position bias, a well-known issue in NLP, into the realm of multi-modal applications. First, we reveal that LVLMs cannot robustly make use of information across multiple images, where merely altering the position of images may significantly change their predictions. Then we further identify this position bias, concluding that open-source models exhibit a strong recency bias, while proprietary models struggle with positions in the middle. Motivated by these findings, we propose SoFA, a simple, training-free approach that reduces the model's dependence on positional information through linear interpolation between inter-image causal attention and bidirectional counterparts. Experiments demonstrate that SoFA effectively mitigates position bias and significantly enhances the overall reasoning performance of existing models.", + "bbox": [ + 509, + 657, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "10606", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 225, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This research was, in part, funded by the U.S. Government - DARPA TIAMAT HR00112490421. The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the U.S. Government.", + "bbox": [ + 89, + 109, + 483, + 194 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 222, + 187, + 238 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1, 2, 3, 6", + "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob L. Menick, Sebastian Borgeaud, Andy Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karén Simonyan. Flamingo: a visual language model for few-shot learning. In NeurIPS, 2022. 1, 2", + "[3] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In CVPR, pages 2425-2433, 2015. 1", + "[4] Anas Awadalla, Irena Gao, Josh Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Shiori Sagawa, et al. Openflamingo: An opensource framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390, 2023. 1, 2, 3", + "[5] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 1, 2", + "[6] Folco Bertini Baldassini, Mustafa Shukor, Matthieu Cord, Laure Soulier, and Benjamin Piwowarski. What makes multimodal in-context learning work? In CVPR, pages 1539-1550, 2024. 3, 7", + "[7] Rohan Bavishi, Erich Elsen, Curtis Hawthorne, Maxwell Nye, Augustus Odena, Arushi Somani, and Săgnak Taşirlar. Introducing our multimodal models, 2023. 2", + "[8] Liangyu Chen, Bo Li, Sheng Shen, Jingkang Yang, Chunyuan Li, Kurt Keutzer, Trevor Darrell, and Ziwei Liu. Large language models are visual reasoning coordinators. In NeurIPS, 2024. 1", + "[9] Shuo Chen, Zhen Han, Bailan He, Mark Buckley, Philip Torr, Volker Tresp, and Jindong Gu. Understanding and improving in-context learning on vision-language models. In ICLR, 2023. 3, 7", + "[10] Yuhan Chen, Ang Lv, Ting-En Lin, Changyu Chen, Yuchuan Wu, Fei Huang, Yongbin Li, and Rui Yan. Fortify the short" + ], + "bbox": [ + 93, + 247, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "est stave in attention: Enhancing context awareness of large language models for effective tool use. In ACL, 2023. 5", + "[11] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In CVPR, pages 24185–24198, 2024. 1, 2, 6, 7", + "[12] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with $90\\%$ chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2(3):6, 2023. 2", + "[13] Neil Cohn, Ryan Taylor, and Kaitlin Pederson. A picture is worth more words over time: Multimodality and narrative structure across eight decades of american superhero comics. Multimodal Communication, 6(1):19-37, 2017. 1", + "[14] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. In NeurIPS, pages 16344-16359, 2022. 6", + "[15] Xingyu Fu, Yushi Hu, Bangzheng Li, Yu Feng, Haoyu Wang, Xudong Lin, Dan Roth, Noah A Smith, Wei-Chiu Ma, and Ranjay Krishna. Blink: Multimodal large language models can see but not perceive. In ECCV, 2024. 1, 2, 3, 6, 7", + "[16] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In CVPR, pages 6904-6913, 2017. 1, 4, 7, 8", + "[17] Danna Gurari, Qing Li, Abigale J Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo Luo, and Jeffrey P Bigham. Vizwiz grand challenge: Answering visual questions from blind people. In CVPR, pages 3608-3617, 2018. 7, 8", + "[18] Junqing He, Kunhao Pan, Xiaoqun Dong, Zhuoyang Song, LiuYiBo LiuYiBo, Qiangosun Qianguosun, Yuxin Liang, Hao Wang, Enming Zhang, and Jiaxing Zhang. Never lost in the middle: Mastering long-context question answering with position-agnostic decompositional training. In ACL, pages 13628–13642, 2024. 5", + "[19] Dongfu Jiang, Xuan He, Huaye Zeng, Cong Wei, Max Ku, Qian Liu, and Wenhu Chen. Mantis: Interleaved multi-image instruction tuning. arXiv preprint arXiv:2405.01483, 2024. 1, 2, 3, 6, 7", + "[20] Yixing Jiang, Jeremy Irvin, Ji Hun Wang, Muhammad Ahmed Chaudhry, Jonathan H Chen, and Andrew Y Ng. Many-shot in-context learning in multimodal foundation models. In ICML, 2024. 8", + "[21] Mehran Kazemi, Nishanth Dikkala, Ankit Anand, Petar Devic, Ishita Dasgupta, Fangyu Liu, Bahare Fatemi, Pranjal Awasthi, Dee Guo, Sreenivas Gollapudi, et al. Remi: A dataset for reasoning with multiple images. arXiv preprint arXiv:2406.09175, 2024. 2", + "[22] Hugo Laurençon, Andrés Marafioti, Victor Sanh, and Léo Tronchon. Building and better understanding vision-language models: insights and future directions. arXiv preprint arXiv:2408.12637, 2024. 1" + ], + "bbox": [ + 516, + 92, + 903, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "10607", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[23] Hugo Laurençon, Lucile Saulnier, Léo Tronchon, Stas Bekman, Amanpreet Singh, Anton Lozhkov, Thomas Wang, Siddharth Karamcheti, Alexander Rush, Douwe Kiela, et al. Obelics: An open web-scale filtered dataset of interleaved image-text documents. In NeurIPS, 2024. 2", + "[24] Hugo Laurençon, Léo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language models? arXiv preprint arXiv:2405.02246, 2024. 1, 2, 3, 6, 7", + "[25] Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Kuttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al. Retrieval-augmented generation for knowledge-intensive nlp tasks. In CVPR, pages 9459-9474, 2020. 1", + "[26] Bohao Li, Yuying Ge, Yixiao Ge, Guangzhi Wang, Rui Wang, Ruimao Zhang, and Ying Shan. Seed-bench: Benchmarking multimodal large language models. In CVPR, pages 13299-13308, 2024. 1", + "[27] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava next-interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895, 2024. 2, 3, 6, 7", + "[28] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, pages 19730–19742. PMLR, 2023. 1", + "[29] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mybench: A comprehensive multi-modal video understanding benchmark. In CVPR, pages 22195-22206, 2024. 2, 7", + "[30] Li Li, Jiawei Peng, Huiyi Chen, Chongyang Gao, and Xu Yang. How to configure good in-context sequence for visual question answering. In CVPR, pages 26710-26720, 2024. 2, 7", + "[31] Ji Lin, Hongxu Yin, Wei Ping, Pavlo Molchanov, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models. In CVPR, pages 26689-26699, 2024. 6, 7", + "[32] Ziyi Lin, Chris Liu, Renrui Zhang, Peng Gao, Longtian Qiu, Han Xiao, Han Qiu, Chen Lin, Wenqi Shao, Keqin Chen, et al. Sphinx: The joint mixing of weights, tasks, and visual embeddings for multi-modal large language models. arXiv preprint arXiv:2311.07575, 2023. 6", + "[33] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024. 2", + "[34] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2024. 1", + "[35] Nelson F Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. Lost in the middle: How language models use long contexts. In ACL, pages 157-173. MIT Press One Broadway, 12th Floor, Cambridge, Massachusetts 02142, USA ..., 2024. 1, 2, 3, 5", + "[36] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. In NeurIPS, pages 2507–2521, 2022. 1" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Pan Lu, Baolin Peng, Hao Cheng, Michel Galley, Kai-Wei Chang, Ying Nian Wu, Song-Chun Zhu, and Jianfeng Gao. Chameleon: Plug-and-play compositional reasoning with large language models. In NeurIPS, 2024. 1", + "[38] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In CVPR, pages 3195-3204, 2019. 1, 7, 8", + "[39] Duy MH Nguyen, Hoang Nguyen, Nghiem Diep, Tan Ngoc Pham, Tri Cao, Binh Nguyen, Paul Swoboda, Nhat Ho, Shadi Albarqouni, Pengtao Xie, et al. Lvm-med: Learning large-scale self-supervised vision models for medical imaging via second-order graph matching. In NeurIPS, 2024. 2", + "[40] Yassine Ouali, Adrian Bulat, Brais Martinez, and Georgios Tzimiropoulos. Clip-dpo: Vision-language models as a source of preference for fixing hallucinations in lvlms. In ECCV, 2024. 2", + "[41] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. In ICLR, 2023. 2", + "[42] Alexander Peysakhovich and Adam Lerer. Attention sorting combats recency bias in long context language models. arXiv preprint arXiv:2310.01427, 2023. 1, 3, 5", + "[43] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 2", + "[44] Ori Ram, Yoav Levine, Itay Dalmedigos, Dor Muhlgay, Amnon Shashua, Kevin Leyton-Brown, and Yoav Shoham. Incontext retrieval-augmented language models. In ACL, pages 1316-1331. MIT Press One Broadway, 12th Floor, Cambridge, Massachusetts 02142, USA . . ., 2023. 1", + "[45] Nir Ratner, Yoav Levine, Yonatan Belinkov, Ori Ram, Inbal Magar, Omri Abend, Ehud Karpas, Amnon Shashua, Kevin Leyton-Brown, and Yoav Shoham. Parallel context windows for large language models. In ACL, 2022. 3, 5", + "[46] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS, 2022. 2", + "[47] Junjie Shentu and Noura Al Moubayed. Cxr-irgen: An integrated vision and language model for the generation of clinically accurate chest x-ray image-report pairs. In WACV, pages 5212–5221, 2024. 2", + "[48] Lin Shi, Weicheng Ma, and Soroush Vosoughi. Judging the judges: A systematic investigation of position bias in pairwise comparative assessments by llms. arXiv preprint arXiv:2406.07791, 2024. 1, 3", + "[49] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 5" + ], + "bbox": [ + 516, + 90, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10608", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[50] Alane Suhr, Stephanie Zhou, Ally Zhang, Iris Zhang, Huajun Bai, and Yoav Artzi. A corpus for reasoning about natural language grounded in photographs. In ACL, 2018. 1, 7", + "[51] Quan Sun, Yufeng Cui, Xiaosong Zhang, Fan Zhang, Qiying Yu, Yueze Wang, Yongming Rao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative multimodal models are in-context learners. In CVPR, pages 14398-14409, 2024. 1", + "[52] Dídac Surís, Sachit Menon, and Carl Vondrick. Vipergpt: Visual inference via python execution for reasoning. In CVPR, pages 11888-11898, 2023. 1", + "[53] MosaicML NLP Team et al. Introducing mpt-7b: A new standard for open-source, commercially usable llms, 2023. 2", + "[54] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2", + "[55] Pat Verga, Sebastian Hofstatter, Sophia Althammer, Yixuan Su, Aleksandra Piktus, Arkady Arkhangorodsky, Minjie Xu, Naomi White, and Patrick Lewis. Replacing judges with juries: Evaluating llm generations with a panel of diverse models. arXiv preprint arXiv:2404.18796, 2024. 1, 3", + "[56] Fei Wang, Xingyu Fu, James Y Huang, Zekun Li, Qin Liu, Xiaogeng Liu, Mingyu Derek Ma, Nan Xu, Wenxuan Zhou, Kai Zhang, et al. Muirbench: A comprehensive benchmark for robust multi-image understanding. arXiv preprint arXiv:2406.09411, 2024. 1, 2, 3, 6, 7", + "[57] Hengyi Wang, Haizhou Shi, Shiwei Tan, Weiyi Qin, Wenyuan Wang, Tunyu Zhang, Akshay Nambi, Tanuja Ganu, and Hao Wang. Multimodal needle in a haystack: Benchmarking long-context capability of multimodal large language models. arXiv preprint arXiv:2406.11230, 2024. 3", + "[58] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, et al. Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079, 2023. 2", + "[59] Weiyun Wang, Shuibao Zhang, Yiming Ren, Yuchen Duan, Tiantong Li, Shuo Liu, Mengkang Hu, Zhe Chen, Kaipeng Zhang, Lewei Lu, et al. Needle in a multimodal haystack. In NeurIPS, 2024. 3, 8", + "[60] Ziqi Wang, Hanlin Zhang, Xiner Li, Kuan-Hao Huang, Chi Han, Shuiwang Ji, Sham M Kakade, Hao Peng, and Heng Ji. Eliminating position bias of language models: A mechanistic approach. arXiv preprint arXiv:2407.01100, 2024. 1, 3, 5", + "[61] Tsung-Han Wu, Giscard Biamby, Jerome Quenum, Ritwik Gupta, Joseph E Gonzalez, Trevor Darrell, and David M Chan. Visual haystacks: Answering harder questions about sets of images. arXiv preprint arXiv:2407.13766, 2024. 3, 8", + "[62] Michihiro Yasunaga, Armen Aghajanyan, Weijia Shi, Rich James, Jure Leskovec, Percy Liang, Mike Lewis, Luke Zettlemoyer, and Wen-tau Yih. Retrieval-augmented multimodal language modeling. In ICML, 2022. 8", + "[63] Yijiong Yu, Huiqiang Jiang, Xufang Luo, Qianhui Wu, ChinYew Lin, Dongsheng Li, Yuqing Yang, Yongfeng Huang, and Lili Qiu. Mitigate position bias in large language models via scaling a single dimension. arXiv preprint arXiv:2406.02536, 2024. 3" + ], + "bbox": [ + 91, + 92, + 480, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[64] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In CVPR, pages 9556-9567, 2024. 1", + "[65] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In CVPR, pages 11975-11986, 2023. 2", + "[66] Bingchen Zhao, Yongshuo Zong, Letian Zhang, and Timothy Hospedales. Benchmarking multi-image understanding in vision and language models: Perception, knowledge, reasoning, and multi-hop reasoning. arXiv preprint arXiv:2406.12742, 2024. 3, 6, 7", + "[67] Haozhe Zhao, Zefan Cai, Shuzheng Si, Xiaojian Ma, Kaikai An, Liang Chen, Zixuan Liu, Sheng Wang, Wenjuan Han, and Baobao Chang. Mmcl: Empowering vision-language model with multi-modal in-context learning. In ICLR, 2023. 2", + "[68] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. In NeurlPS, pages 46595-46623, 2023. 1, 3", + "[69] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. In ICLR, 2023. 2", + "[70] Wanrong Zhu, Jack Hessel, Anas Awadalla, Samir Yitzhak Gadre, Jesse Dodge, Alex Fang, Youngjae Yu, Ludwig Schmidt, William Yang Wang, and Yejin Choi. Multimodal c4: An open, billion-scale corpus of images interleaved with text. In NeurIPS, 2024. 2", + "[71] Yuchen Zhuang, Yue Yu, Kuan Wang, Haotian Sun, and Chao Zhang. Toolqa: A dataset for llm question answering with external tools. In NeurIPS, pages 50117-50143, 2023. 1" + ], + "bbox": [ + 516, + 92, + 903, + 597 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "10609", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Identifying and Mitigating Position Bias of Multi-image Vision-Language Models/17ea1e19-b848-462e-8296-051308ed2dd4_model.json b/2025/Identifying and Mitigating Position Bias of Multi-image Vision-Language Models/17ea1e19-b848-462e-8296-051308ed2dd4_model.json new file mode 100644 index 0000000000000000000000000000000000000000..323f692f5bbbb5f5981b543c607dbfd2703a08ea --- /dev/null +++ b/2025/Identifying and Mitigating Position Bias of Multi-image Vision-Language Models/17ea1e19-b848-462e-8296-051308ed2dd4_model.json @@ -0,0 +1,2510 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.131, + 0.908, + 0.155 + ], + "angle": 0, + "content": "Identifying and Mitigating Position Bias of Multi-image Vision-Language Models" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.18, + 0.765, + 0.217 + ], + "angle": 0, + "content": "Xinyu Tian1 Shu Zou1 Zhaoyuan Yang2 Jing Zhang1 \n1Australian National University 2GE Research" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.218, + 0.751, + 0.234 + ], + "angle": 0, + "content": "\\(^{1}\\)firstname_lastname@anu.edu.au, \\(^{2}\\)firstname_lastname@ge.com" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.269, + 0.329, + 0.287 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.301, + 0.486, + 0.619 + ], + "angle": 0, + "content": "The evolution of Large Vision-Language Models (LVLMs) has progressed from single to multi-image reasoning. Despite this advancement, our findings indicate that LVLMs struggle to robustly utilize information across multiple images, with predictions significantly affected by the alteration of image positions. To further explore this issue, we introduce Position-wise Question Answering (PQA), a meticulously designed task to quantify reasoning capabilities at each position. Our analysis reveals a pronounced position bias in LVLMs: open-source models excel in reasoning with images positioned later but underperform with those in the middle or at the beginning, while proprietary models show improved comprehension for images at the beginning and end but struggle with those in the middle. Motivated by this, we propose SoFt Attention (SoFA), a simple, training-free approach that mitigates this bias by employing linear interpolation between inter-image causal attention and bidirectional counterparts. Experimental results demonstrate that SoFA reduces position bias and enhances the reasoning performance of existing LVLMs. The code will be available https://github.com/xytian1008/sofa." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.648, + 0.224, + 0.665 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.483, + 0.886 + ], + "angle": 0, + "content": "The advent of Large Vision-Language Models (LVLMs) [1, 2, 5, 24, 34] has catalyzed the development of various text-grounded visual applications, including visual question answering [3, 16, 36], code generation [8, 37, 52], and story telling [13, 26]. Open-source LVLMs such as LLaVA [34], BLIP2 [28], and InternVL [11] have demonstrated competitive performance on single-image tasks like VQAv2 [16], OKVQA [38], and MMMU [64]. However, these LVLMs are limited in their capacity to analyze interimage differences and relationships, as well as to process temporal information. Consequently, recent LVLMs, such as Flamingo [2, 4], Idefics [22-24], and Emu2 [51], have incorporated image-text interleaved data during training to enable ability to tackle multi-image tasks [15, 19, 50, 56]." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.887, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Despite the increasing scale and performance of LVLMs," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.271, + 0.908, + 0.483 + ], + "angle": 0, + "content": "the interpretation of how they reason across multiple images remains a mystery. In fact, we might find some hints in the topic of natural language processing (NLP). Recent research in NLP has highlighted retrieval-augmented generation (RAG) [25, 44, 71], which aims to enable large language models (LLMs) to reason across multiple documents to answer questions. However, [35] find that LLMs tend to effectively comprehend documents positioned at the beginning and end of the input, often neglecting the middle sections. Similarly, in LLM-as-a-judge [48, 55], [68] find that LLMs used to evaluate two model responses tend to favor the first one, regardless of content. This phenomenon, where the model's reasoning capability is influenced by the segment position, is referred to as position bias [35, 42, 60]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.488, + 0.909, + 0.716 + ], + "angle": 0, + "content": "Motivated by the research above, a natural question arises: does position bias exist in LVLMs? Specifically, we seek to determine whether LVLMs can robustly make use of information across multiple images. To investigate this, we begin by evaluating LVLMs on established multi-image benchmarks [15, 19, 50, 56]. We find that simply changing the order of presented images may significantly alter the predictions for most LVLMs, leading to considerable variation in benchmark accuracy. Fig. 1 provides examples of questions and model responses. For instance, in difference spotting [56], swapping two images causes OpenFlamingo [4] to miss distinctions, while in a simple counting task [19], reordering images leads Idefics2 [24] to incorrect reasoning. This positional sensitivity is apparent even in GPT-4o [1], a leading proprietary model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.911, + 0.903 + ], + "angle": 0, + "content": "The above motivational study reveals an unfortunate fact: building upon LLMs, LVLMs may inherit the notorious position bias from their language component. This bias manifests as the model's tendency to excel at reasoning visual information from certain positions, while exhibiting weaknesses at others, compromising robustness and reliability in predictions. To further investigate which positions display strong understanding and which reveal weaknesses, i.e., identifying the pattern of such position bias, we introduce Position-wise Question Answering (PQA). Unlike most multi-image tasks that assess holistic comprehension, PQA requires LVLMs to reason consistently across each" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "10599" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.088, + 0.907, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.293, + 0.907, + 0.337 + ], + "angle": 0, + "content": "Figure 1. Examples where the predictions of LVLMs are influenced by changes in the image positions (Date accessed: Oct 25, 2024). The tasks, from left to right, are difference spotting [56], counting [19] and forensic detection [15], respectively. The correct answers are highlighted in green, while the incorrect ones are highlighted in red. The prompts and outputs are simplified here for clarity and aesthetics." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.348, + 0.483, + 0.394 + ], + "angle": 0, + "content": "image and produce position-wise responses. This approach enables us to measure position-wise accuracy and analyze the model's reasoning capacity across different positions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.398, + 0.484, + 0.565 + ], + "angle": 0, + "content": "Our analysis reveals that most open-source models, e.g., OpenFlamingo [4] and Idefics2 [24], exhibit a pronounced recency bias, demonstrating stronger reasoning capabilities for images presented later while underperforming on those in the middle or at the beginning. In contrast, proprietary models like GPT-4o [1] show enhanced understanding of images at the beginning and end, while neglecting those in the middle, a pattern consistent with previous findings in LLMs [35]. This observation is further corroborated by Fig. 1, where LVLMs tend to reason incorrectly about images positioned in poor-performing areas." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.57, + 0.484, + 0.902 + ], + "angle": 0, + "content": "The existence of position bias significantly compromises model robustness and undermines accuracy across various multi-image tasks. Inspired by this, we aim to investigate the underlying mechanical causes and mitigate this issue. Specifically, we notice that in most LVLMs, the hidden states of image tokens are heavily influenced by their relative positions due to inter-image causal attention. In particular, image tokens are treated as a type of special text tokens where 1) images appearing later may interact with preceding ones; 2) images at the beginning are isolated, lacking access to global context. This autoregressive setup implicitly imposes a semantic order on the presented images, amplifying their positional information. To address this, we introduce SoFt Attention (SoFA), a training-free approach to mitigate position bias across various LVLMs. SoFA reduces the dependency of image hidden states on positional information by simply performing linear interpolation between inter-image causal attention and its bidirectional counterparts. Experimental results indicate that SoFA effectively equilibrates the reasoning capabilities of LVLMs across various positions, achieving average accuracy improvements of \\(2 \\sim 3\\%\\) on existing multi-image benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.348, + 0.814, + 0.362 + ], + "angle": 0, + "content": "In summary, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.363, + 0.905, + 0.438 + ], + "angle": 0, + "content": "- Despite the significant potential of LVLMs in multi-image applications, we find that position bias, a well-explored phenomenon in NLP, has emerged as a potential bottleneck in the development of LVLMs, substantially impacting model robustness and reliability." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.438, + 0.905, + 0.498 + ], + "angle": 0, + "content": "- We further identify such position bias and discover that most open-source models demonstrate a pronounced recency bias, whereas proprietary models tend to struggle in middle positions." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.499, + 0.905, + 0.559 + ], + "angle": 0, + "content": "- We introduce SoFA, a straightforward training-free approach designed to mitigate the position bias inherited from pre-trained LVLMs, and substantiate its effectiveness through experimental validation." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.363, + 0.905, + 0.559 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.573, + 0.655, + 0.589 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Large Vision-Language Models. Existing LVLMs are generally composed of a vision encoder [43, 65], an LLM [12, 53, 54], and a bridging projector [2, 7, 27] that facilitates interaction between the two. Conventional LVLMs, pre-trained on image-text pairs [46] and fine-tuned through preference [33, 40] or instruction tuning [27, 41], demonstrate substantial success in applications like dialogue [5, 11], question answering [58, 69], and domain-specific tasks [39, 47]. However, constrained by their training data, these models primarily support single-image reasoning during inference. Recently, the integration of image-text interleaved data [23, 70] has enabled multi-image reasoning capabilities, encompassing image comparison [21, 56], association [15, 19], temporal relationship analysis [29], and emerging properties of multi-modal in-context learning [30, 67]. Nonetheless, limited research has delved into how LVLMs perform reasoning across multiple images or the underlying mechanisms driving this process. Position Bias in LLMs. Position bias, an inherent side effect of long-context input, has increasingly captured the at" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10600" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.905, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.225, + 0.908, + 0.253 + ], + "angle": 0, + "content": "Figure 2. The results of multiple evaluations while shuffling image positions. We record the minimum, maximum, and average accuracy (left vertical axis), along with the prediction inconsistency between the best and worst-performing evaluations (right vertical axis)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.265, + 0.486, + 0.476 + ], + "angle": 0, + "content": "tention of NLP researchers. The pioneering work on this issue, notably [35], reveals that when LLMs process multiple documents, they often concentrate on those at the beginning and end while disregarding much of the information in between, i.e., a phenomenon termed \"lost in the middle\". A related issue has also been identified in LLM-as-a-judge [48, 55], where [68] observe that LLMs tasked with evaluating multiple model responses tend to favor the first one, irrespective of its content. Recent studies have begun to explore methods to mitigate or circumvent position bias in LLMs [42, 45, 60, 63]. However, to the best of our knowledge, few studies have examined the effects of position bias within multi-modal applications, specifically regarding its implications for LVLMs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.479, + 0.485, + 0.706 + ], + "angle": 0, + "content": "Position Bias in LVLMs. While systematic research on position bias in LVLMs is still lacking, recent studies have indirectly underscored the influence of positions on LVLMs through various findings. For instance, in multi-modal in-context learning, studies [9] and [6] reveal that demonstration images placed at the beginning contribute minimally to LVLM performance. Additionally, in assessing long-context capabilities, [61] reports that LVLMs attain particularly high accuracy when key information appears at the end of the input sequence, with similar results observed in [59] and [57]. However, previous research has neither thoroughly analyzed this position bias nor proposed solutions to address it. In contrast, this study systematically investigates position bias in both open-source and proprietary models and introduces an effective mitigation strategy." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.723, + 0.445, + 0.74 + ], + "angle": 0, + "content": "3. Understanding Position Bias in LVLMs" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.9 + ], + "angle": 0, + "content": "In this section, we investigate whether the position bias observed in LLMs extends to LVLMs. We start by empirically confirming the existence of position bias in LVLMs through a motivational study, revealing their poor robustness to image position changes (§3.1). Subsequently, we analyze the nature of this bias by evaluating and quantifying the reasoning capabilities of LVLMs regarding different positions (§3.2). These insights contribute to a deeper understanding and interpretation of LVLMs' reasoning behavior and common failure cases." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.265, + 0.714, + 0.28 + ], + "angle": 0, + "content": "3.1. The Position Matters" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.287, + 0.905, + 0.317 + ], + "angle": 0, + "content": "To assess the existence of position bias, we conduct a simple motivational experiment, as shown in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.318, + 0.906, + 0.499 + ], + "angle": 0, + "content": "Experimental Setup. We select four representative multi-image LVLMs: OpenFlamingo [4], Idefics2 [24], LLaVA-NeXT-Interleave1 [27], GPT-4o [1], and conduct evaluations across established multi-image benchmarks, including BLINK [15], Mantis-Eval [19], MuirBench [56], and MIRB [66]. To ensure clarity, we select partial tasks from these benchmarks that feature position-agnostic questions, wherein the positions of the images do not affect the answers to the questions. See §5.1 for the selected tasks and details. To explore the influence of image positions on the results, we perform multiple evaluations per benchmark, with the positions of the images shuffled in each iteration." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.5, + 0.905, + 0.605 + ], + "angle": 0, + "content": "Metric. We record the minimum, maximum, and average accuracy across these evaluations, which directly reflects the influence of positions on the performance of LVLMs. Furthermore, we assess the prediction inconsistency between the best and worst-performing evaluations, quantifying the proportion of examples exhibiting conflicting predictions across the two evaluations by changing positions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.606, + 0.906, + 0.816 + ], + "angle": 0, + "content": "Results. As illustrated in Fig. 2, for most open-source LVLMs, the prediction inconsistency between the best and worst-performing evaluations approaches \\(30\\%\\), suggesting that merely altering the positions of the presented images may result in an amount of the model's predictions being changed. Such pronounced sensitivity to positional information results in highly unstable performance. For example, OpenFlamingo's accuracy on BLINK can vary significantly, reaching \\(38\\%\\) at its peak and dropping to \\(30\\%\\) at its lowest, indicating an accuracy span of up to \\(10\\%\\) with position changes. GPT-4o, the state-of-the-art proprietary model, although showcasing superior performance, also displays a lack of robustness to positional variations, leading to a prediction inconsistency of around \\(25\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.817, + 0.905, + 0.879 + ], + "angle": 0, + "content": "To qualitatively explore this issue, we present examples in Fig. 1 that illustrate prediction conflicts arising from alterations in image positions. For instance, when prompted to identify the differences between two images, Open-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.532, + 0.887, + 0.892, + 0.9 + ], + "angle": 0, + "content": "1For brevity, we refer to LLaVA-NeXT-Interleave as LLaVA-NeXT." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10601" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.305, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.312, + 0.088, + 0.505, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.089, + 0.705, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.089, + 0.905, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.202, + 0.224, + 0.319, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.335, + 0.224, + 0.417, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.439, + 0.224, + 0.552, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.569, + 0.224, + 0.65, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.224, + 0.79, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.247, + 0.908, + 0.277 + ], + "angle": 0, + "content": "Figure 3. The results on the PQA task. We report position-wise accuracy on four scenarios where the number of images is 5, 10, 15 and 20, respectively. A higher accuracy signifies strong reasoning at the specified position, whereas lower reflects poor-performing areas." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.288, + 0.484, + 0.38 + ], + "angle": 0, + "content": "Flamingo fails to extract fine-grained distinctions upon swapping their positions. Similarly, in the task of object counting, Idefics2 mistakenly identifies a dog as a cat due to the repositioning of images. GPT-4o demonstrates a comparable issue, where rearranging the image options alters both its answer and the accompanying rationale." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.388, + 0.317, + 0.403 + ], + "angle": 0, + "content": "3.2. Identifying Position Bias" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.41, + 0.484, + 0.47 + ], + "angle": 0, + "content": "The quantitative results and qualitative examples above imply biased reasoning across positions from LVLMs. To have a closer look at such position bias, we aim to identify models' reasoning capabilities at different image positions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.47, + 0.484, + 0.711 + ], + "angle": 0, + "content": "However, existing tasks have been restricted to evaluating models' holistic comprehension of multiple images, lacking the means to systematically analyze their fine-grained understanding of each individual position. Specifically, these tasks provide coarse ground truth answers, such as option symbols or numeric values, but do not offer individual rationales for each image, making it challenging to quantify the model's reasoning across different positions. To address this gap, we introduce Position-wise Question Answering (PQA), a carefully designed task that assesses the position-wise reasoning capabilities of LVLMs. In PQA, each question prompts LVLMs to reason independently across all images, generating a distinct response for each. This approach 1) ensures task-level neutrality with respect to any particular position and 2) enables quantification of LVLMs' reasoning performance for each position." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.712, + 0.483, + 0.743 + ], + "angle": 0, + "content": "Specifically, taking object counting as an example, we employ the following query template:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.749, + 0.483, + 0.817 + ], + "angle": 0, + "content": "I: Please provide one answer for each image in the form like \\([x, x, x, \\ldots]\\). \nQ: Among , ..., , how many cats can you find in these images? \nA: [3, 2, 0, ...]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Each query consists of an instruction (I) and a question \\((\\mathbb{Q})\\) where the former prompts the model to generate position-wise responses (A) following a specified output format. We gather responses from LVLMs as a structured list, with each element representing a particular position. This enables us" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.288, + 0.905, + 0.333 + ], + "angle": 0, + "content": "to track position-wise accuracy, where a higher accuracy reflects stronger model capability at a given position, while lower indicates areas of relative weakness." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.336, + 0.906, + 0.532 + ], + "angle": 0, + "content": "Construction of position-wise examples. Since in PQA, each image is semantically independent of others, a PQA example essentially consists of several single-image instances that share the same question. This allows us to easily build position-wise examples from single-image benchmarks. Specifically, we select VQAv2 [16], a substantial single-image dataset that encompasses a diverse array of everyday topics and features both multiple-choice and open-ended questions. Initially, we identify and consolidate single-image examples with identical questions to create a question set, wherein each question corresponds to multiple images. Subsequently, we reformulate these questions into PQA format according to the above-mentioned template." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.535, + 0.906, + 0.716 + ], + "angle": 0, + "content": "Experimental Setup. We utilize the same LVLMs as previously mentioned and conduct evaluations on the proposed PQA task. Additionally, we also consider random guessing, where binary choice and numeric questions are answered randomly with yes or no, or with an integer value. We examine four scenarios with varying numbers of images: 5, 10, 15, and 20, collecting 1,000 examples for each scenario. Furthermore, to maintain fairness, we augment each example by shuffling the image positions four times, resulting in a total of 5,000 examples. In contrast to prior settings that evaluate holistic performance, we provide separate accuracy metrics for each position in this analysis." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.718, + 0.905, + 0.748 + ], + "angle": 0, + "content": "Results. Fig. 3 displays the results of LVLMs on the PQA task, from which we may derive the following insights:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.902 + ], + "angle": 0, + "content": "1) Open-source LVLMs exhibit recency bias, whereas proprietary models struggle in middle positions. For open-source models, a stronger reasoning capability is observed for images positioned at the end, while significant underperformance is noted in the middle and beginning. For example, with 10 images, the accuracy of OpenFlamingo varies by nearly \\(9\\%\\) between the far left and far right positions \\((51.84\\% \\rightarrow 59.71\\%)\\). In contrast, proprietary models, i.e., GPT-40, reveal a typical U-shaped curve, highlighting pronounced weaknesses in the middle positions." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10602" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "2) Position bias becomes more pronounced as the number of images increases. For instance, with 20 images, the accuracy gap for OpenFlamingo between the two ends reaches \\(14\\%\\) (\\(36.79\\% \\rightarrow 50.17\\%\\)). Notably, at this stage, OpenFlamingo seems to nearly lost its reasoning capability for most of the beginning images, achieving results only marginally better than random guessing (\\(34.49\\%\\))." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.198, + 0.482, + 0.349 + ], + "angle": 0, + "content": "3) Relationship to position bias in LLMs. Previous research [35] shows that LLMs often overlook information in middle positions when processing multiple documents, a phenomenon termed \"lost in the middle\". Our findings suggest this effect persists for GPT in multi-image reasoning. However, for open-source models, the primacy position advantage appears to diminish, showing a marked decaying trend from the end to beginning. Whether the position bias in LVLMs is correlated with that of their LLM backbones remains an interesting question for further exploration." + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.362, + 0.303, + 0.379 + ], + "angle": 0, + "content": "4. The Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.387, + 0.483, + 0.434 + ], + "angle": 0, + "content": "In this section, we first analyze the underlying causes of position bias (§4.1). Then we introduce SoFA, a straightforward, training-free approach to mitigate this issue (§4.2)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.442, + 0.367, + 0.456 + ], + "angle": 0, + "content": "4.1. Position Mechanism in LVLMs" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.464, + 0.483, + 0.555 + ], + "angle": 0, + "content": "To investigate the origins of position bias, we need to first determine what contributes to the positional information in LVLMs. In NLP, it is generally believed that position embeddings [45, 49] and causal attention [42, 60] are the primary mechanisms enabling the sequential structure of input. Formally," + }, + { + "type": "equation", + "bbox": [ + 0.131, + 0.565, + 0.483, + 0.592 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} = \\mathcal {P} \\left(\\mathbf {X} \\mathbf {W} _ {\\mathbf {q}}, \\mathbf {p o s}\\right), \\quad \\mathbf {K} = \\mathcal {P} \\left(\\mathbf {X} \\mathbf {W} _ {\\mathbf {k}}, \\mathbf {p o s}\\right) \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.134, + 0.587, + 0.381, + 0.603 + ], + "angle": 0, + "content": "\\[\n\\mathbf {H} = \\operatorname {S o f t m a x} \\left(\\mathbf {Q K} ^ {\\mathrm {T}} / \\sqrt {d}\\right) \\odot \\mathbb {1} _ {\\text {c a u s a l}} \\mathbf {V}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.483, + 0.705 + ], + "angle": 0, + "content": "where \\(\\mathbf{Q}\\), \\(\\mathbf{K}\\), and \\(\\mathbf{V}\\) represent the queries, keys, and values, respectively. \\(\\mathbf{X}\\) refers to the input tokens, encompassing both image and text types, pos denotes the position embedding of \\(\\mathbf{X}\\), which is integrated into the attention via the position encoding function \\(\\mathcal{P}\\), \\(\\mathbb{1}_{\\mathrm{causal}}\\) denotes the causal mask, which ensures autoregressive generation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.901 + ], + "angle": 0, + "content": "This reveals two key factors influencing LVLMs' positional awareness: 1) position embedding explicitly initializes a sequential value to each token, and 2) causal attention implicitly enhances this positional characteristic through unidirectional information access. This presents us with two distinct solutions. One approach is to directly reassign position embeddings, a widely used strategy in NLP [10, 18, 45], as it fundamentally eliminates the influence of position on documents. However, this aggressive method is not suitable for LVLMs, since in many scenarios, such as video understanding and action recognition, we still expect images to retain positional information for temporal analysis. To clarify, our aim is to mitigate the imbalanced" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.089, + 0.625, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.177, + 0.616, + 0.188 + ], + "angle": 0, + "content": "(A) Causal Attention" + }, + { + "type": "image", + "bbox": [ + 0.653, + 0.089, + 0.762, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.657, + 0.177, + 0.756, + 0.188 + ], + "angle": 0, + "content": "(B) Isolated Attention" + }, + { + "type": "image", + "bbox": [ + 0.787, + 0.089, + 0.898, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.783, + 0.177, + 0.905, + 0.188 + ], + "angle": 0, + "content": "(C) Bidirectional Attention" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.2, + 0.905, + 0.352 + ], + "angle": 0, + "content": "Figure 4. The three inter-image attention mechanisms, where \\( I_{1} \\) and \\( I_{2} \\) represent the tokens of two images, \\( T_{1} \\) and \\( T_{2} \\) represent the tokens of two text segments. Note that typically each image involves numerous tokens, e.g., 576 for LLaVA. Here for clarity we simplify them to a single token. In (A), images interact in a unidirectional manner, allowing \\( I_{2} \\) to attend to \\( I_{1} \\), while \\( I_{1} \\) remains isolated. In (B), each image is isolated, indicating they can only attend to themselves. In (C), bidirectional interaction is enabled so that each image can attend to any other images. It is worth mentioning that we only alter the inter-image attention while preserving causal attention between the text segments." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.365, + 0.905, + 0.395 + ], + "angle": 0, + "content": "reasoning capabilities of LVLMs across different positions, rather than to eliminate their positional awareness." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.397, + 0.906, + 0.578 + ], + "angle": 0, + "content": "Therefore, in this paper, we primarily investigate the relationship between causal attention and position bias. We notice that, while images do not exhibit a distinct semantic order as text does, most LVLMs, e.g., LLaVA series, treat image and text tokens in a uniform autoregressive manner. As shown in Fig. 4 (A), analogous to texts, there exists interimage causal attention which allows 1) later-positioned images to interact with those preceding them, while 2) earlier images remain isolated, lacking access to global contexts. From a mechanical perspective, this framework renders image hidden states highly dependent on their relative positions, thus implicitly amplifying their positional cues." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.58, + 0.905, + 0.61 + ], + "angle": 0, + "content": "To verify our hypothesis, we compare the inter-image causal attention against following two alternative variants:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.611, + 0.905, + 0.656 + ], + "angle": 0, + "content": "1) Isolated attention. As illustrated in Fig. 4 (B), we disable interactions between images, enforcing each image to attend solely to itself." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.658, + 0.905, + 0.703 + ], + "angle": 0, + "content": "2) Bidirectional attention. Rather than restricting to unidirectional information flow, we permit each image to attend to any other, as shown in Fig. 4 (C)." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.611, + 0.905, + 0.703 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.906, + 0.901 + ], + "angle": 0, + "content": "We modify the default causal mask in the language decoder to its isolated and bidirectional counterparts and assess the impact of them on position bias in PQA, as shown in Fig. 5. It is evident that altering causal attention to either isolated or bidirectional forms significantly alleviates position bias, resulting in more balanced accuracy across different positions. This empirically confirms that inter-image causal attention is the primary source of position bias. However, these modifications lead to varying degrees of performance degradation, with isolated attention causing the most substantial decline. This outcome is reasonable, as such modifications may substantially deviate from the training framework, leading to out-of-distribution hidden states." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10603" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.29, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.171, + 0.225, + 0.236, + 0.237 + ], + "angle": 0, + "content": "(A) Idefics2" + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.089, + 0.482, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.343, + 0.225, + 0.438, + 0.237 + ], + "angle": 0, + "content": "(B) LLaVA-NeXT" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.25, + 0.483, + 0.278 + ], + "angle": 0, + "content": "Figure 5. The position-wise accuracy of the three attention mechanisms on the PQA task, alongside our proposed SoFA method." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.29, + 0.313, + 0.307 + ], + "angle": 0, + "content": "4.2. Mitigating Position Bias" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.312, + 0.484, + 0.434 + ], + "angle": 0, + "content": "The analysis above shows that inter-image causal attention is the main cause of position bias. Furthermore, it also reveals a trade-off between accuracy and robustness, a typical dilemma in machine learning. Building on this insight, we propose SoFt Attention (SoFA), a straightforward, training-free approach to mitigate position bias. SoFA employs linear interpolation between inter-image causal attention and its bidirectional counterpart. Formally," + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.445, + 0.483, + 0.482 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {H} _ {\\text {s o f t}} = \\operatorname {S o f t m a x} \\left(\\mathbf {Q K} ^ {\\mathrm {T}} / \\sqrt {d}\\right) \\odot \\mathbf {M} _ {\\text {s o f t}} \\mathbf {V} \\tag {2} \\\\ \\mathbf {M} _ {\\text {s o f t}} = (1 - \\sigma) \\mathbb {1} _ {\\text {c a u s a l}} + \\sigma \\mathbb {1} _ {\\text {b i d i r e c t i o n a l}} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.493, + 0.483, + 0.569 + ], + "angle": 0, + "content": "where \\(\\mathbb{1}_{\\mathrm{causal}}\\) and \\(\\mathbb{1}_{\\mathrm{bidual}}\\) denote the two attention masks, with \\(\\sigma\\) controlling the extent of bidirectional part. Note that since we are only interested in the visual positional relation, we only modify the inter-image attention, while preserving causal attention between text segments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.483, + 0.659 + ], + "angle": 0, + "content": "SoFA as a position smoother. Intuitively, given that interimage causal attention excessively emphasizes positional information, SoFA integrates partial bidirectional attention to smooth the influence of position on the models at minimal cost, achieving a balanced trade-off between reasoning performance and position robustness." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.75 + ], + "angle": 0, + "content": "SoFA as an interaction enhancer. Practically, SoFA enhances the interaction between images, enriching the semantic information of earlier image features to enable more effective utilization by LVLMs. This aligns with the most cases where key information is typically distributed evenly across various input positions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.751, + 0.483, + 0.902 + ], + "angle": 0, + "content": "For each task, we utilize a small validation set to determine the optimal values for \\(\\sigma\\). Besides, to better align with the training framework, we deploy SoFA every two layers rather than at every layer, while preserving causal attention in the remaining layers. We present SoFA's results in the PQA task as shown in Fig. 5. It is evident that SoFA significantly outperforms both isolated and bidirectional attention, while unlike causal attention in the default setting, SoFA demonstrates balanced performance and exhibits enhanced reasoning capability in most positions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.091, + 0.741, + 0.108 + ], + "angle": 0, + "content": "5. Experiment and Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.115, + 0.905, + 0.176 + ], + "angle": 0, + "content": "In this section, we empirically demonstrate that SoFA 1) effectively mitigates the position bias inherent in LVLMs and 2) significantly improves their overall reasoning capability across various multi-image applications." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.176, + 0.905, + 0.236 + ], + "angle": 0, + "content": "Models. We select a diverse set of open-source models, including Idefics2-8B [24], InternVL2-8B [11], VILA-13B [31], Mantis-8B [19], and LLaVA-NeXT-Interleave-7B [27]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.237, + 0.907, + 0.449 + ], + "angle": 0, + "content": "Implementation details. For the input, we provide instructions that explicitly indicate the task type and output formats. For instance, for multiple-choice questions, we ask LVLMs to provide only the option symbol; for numeric questions, only the number; and for open-ended questions, the response is limited to short answers. For evaluation, to ensure fairness, we uniformly adopt GPT-4 [1], providing with the question, ground truth answer, candidate prediction, and it determines whether the prediction is correct or not. For SoFA, we determine \\(\\sigma\\) using a 32-shot validation set per task and perform interpolation every two layers. Unless otherwise specified, all models are set to FP16 precision and utilize Flash Attention [14], with sub-image splitting [32] disabled for fairness." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.456, + 0.704, + 0.47 + ], + "angle": 0, + "content": "5.1. Position Robustness" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.478, + 0.906, + 0.689 + ], + "angle": 0, + "content": "Benchmark. We first evaluate whether SoFA may effectively mitigate position bias. To this end, similar to §3.1, we select position-agnostic tasks from existing multi-image benchmarks. Specifically, we exclude tasks where changing the order of images may alter the semantics of the question, rendering it unanswerable, e.g., video understanding. Here, we list some of the selected tasks: forensic detection and image jigsaw from BLINK [15], difference spotting and diagram understanding from MuirBench [56], visual analogy and attribute matching from MIRB [66]. We include all examples from Mantis-Eval [19] since the majority are position-agnostic. See the supplementary material for the details of more evaluated tasks. Our experimental setup and metrics are consistent with those described in §3.1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.906, + 0.84 + ], + "angle": 0, + "content": "SoFA reduces the influence of positions on LVLMs. As demonstrated in Table 1, the integration of SoFA within the models markedly reduces prediction inconsistency, leading to enhanced stability in performance. For instance, Idefics2's inconsistency on BLINK declines from \\(41.55\\%\\) to \\(12.36\\%\\), signifying that the responses for the majority of examples remain consistent despite variations in image positioning. This observation indicates that LVLMs exhibit more balanced reasoning capabilities across different positions, effectively mitigating the effects of position bias." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.902 + ], + "angle": 0, + "content": "SoFA eliminates the blind spots of LVLMs. In addition to the enhanced robustness, we observe a marked improvement in the performance of LVLMs in worst-case scenarios, particularly regarding minimum and average accuracy," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10604" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.089, + 0.905, + 0.281 + ], + "angle": 0, + "content": "
MethodBLINK [15]Mantis-Eval [19]MuirBench [56]MIRB [66]
MinAveMaxIncMinAveMaxIncMinAveMaxIncMinAveMaxInc
Idefics2 [24]43.2646.9348.6841.5543.9346.6251.1038.5743.2845.3047.8134.4645.2647.8849.6829.94
Idefics2 + SoFA47.1848.6349.1212.3648.7149.1850.838.6146.1247.5148.4513.2747.3948.3649.136.99
InternVL2 [11]38.8140.4543.7430.1845.3646.9249.5125.2449.2852.3356.1038.6541.8444.3846.3629.60
InternVL2 + SoFA41.6442.3243.517.1648.1349.2550.1112.0054.6955.9256.785.1644.2045.3545.876.64
VILA [31]45.9348.5951.4225.3348.4449.2051.8521.2941.6643.1248.2737.2647.1749.5952.3431.77
VILA + SoFA48.2750.8051.1710.6849.2951.6052.6812.1645.7646.5247.137.9248.2251.4351.9517.46
Mantis [19]48.3449.2451.5228.3556.4058.3863.4232.1247.6748.9451.1527.4553.1155.7157.4225.79
Mantis + SoFA49.2250.8752.3416.2360.4862.2164.6814.3049.8850.2650.795.6154.3956.3456.938.35
LLaVA-NeXT [27]49.3452.4055.4331.8954.5759.1062.2933.7551.9653.4555.7824.9255.4958.4060.2828.56
LLaVA-NeXT + SoFA53.2054.9156.0811.2058.8761.2361.9814.2053.6155.3756.499.1558.1359.8161.166.96
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.291, + 0.907, + 0.32 + ], + "angle": 0, + "content": "Table 1. The evaluation on position-agnostic tasks with and without SoFA. Similar to §3.1, we perform multiple evaluations and record minimum, maximum and average accuracy. We also report prediction inconsistency between best and worst-performing evaluations." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.329, + 0.487, + 0.521 + ], + "angle": 0, + "content": "
MethodBLINKM-EvalMuirMIRBNLVR2MVBench
Random Guessing39.2725.3121.4919.1350.7525.91
Idefics2 [24]45.7848.4142.8446.5881.9423.37
Idefics2 + SoFA48.3150.7747.3648.8385.1525.65
InternVL2 [11]38.9550.3054.5342.6685.5629.31
InternVL2 + SoFA43.2651.1157.1446.1988.1932.77
VILA [31]49.7952.6448.7350.3678.8225.81
VILA + SoFA50.2952.1352.8051.1580.3426.40
Mantis [19]51.4156.8544.2854.5990.2641.59
Mantis + SoFA53.1959.2347.6455.6889.9844.40
LLaVA-NeXT [27]53.3450.8348.2257.1587.2854.26
LLaVA-NeXT + SoFA55.9254.5150.4360.6789.4557.71
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.531, + 0.483, + 0.561 + ], + "angle": 0, + "content": "Table 2. The overall evaluation accuracy of LVLMs on various multi-image benchmarks with and without SoFA." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.567, + 0.483, + 0.645 + ], + "angle": 0, + "content": "which both exhibit an increase of approximately \\(3\\%\\). This suggests that the model retains strong performance despite variations in the distribution of image positions, indicating that SoFA effectively addresses the deficiencies of LVLMs in positions where performance is typically suboptimal." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.653, + 0.248, + 0.668 + ], + "angle": 0, + "content": "5.2. Overall Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.483, + 0.825 + ], + "angle": 0, + "content": "Benchmark. Given the observation that SoFA effectively mitigates position bias, we proceed with a comprehensive evaluation to determine whether SoFA can enhance the overall reasoning capabilities of LVLMs. Specifically, we include all tasks from the selected benchmarks [15, 19, 56, 66]. In addition, we incorporate NLVR2 [50] and MVBench [29], which exclusively assess the models' capability in image-pair comparison and video understanding, respectively. We directly utilize the default image positions from the benchmarks and record their accuracy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "SoFA enhances reasoning capability of LVLMs. In Table 2, we observe that integrating SoFA leads to an average improvement of \\(2 \\sim 3\\%\\) in the performance of LVLMs across various benchmarks. For instance, Idefics2 shows a \\(2.25\\%\\) improvement on MIRB (46.58% → 48.83%)," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.331, + 0.907, + 0.409 + ], + "angle": 0, + "content": "while InternVL2 exhibits a \\(4.31\\%\\) increase on BLINK \\((38.95\\% \\rightarrow 43.26\\%)\\). This demonstrates that, despite the advancements in multi-image LVLMs, position bias poses a significant obstacle to their reasoning capability, whereas SoFA successfully addresses this limitation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.422, + 0.633, + 0.437 + ], + "angle": 0, + "content": "5.3. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.446, + 0.907, + 0.657 + ], + "angle": 0, + "content": "Attention distribution visualization. To gain a deeper understanding of SoFA's working mechanism, we visualize the attention distribution within LVLMs across different positions. Using PQA as an example, we derive the average attention value of the final layer's query token (whose hidden state is directly used for generation) for each image. As shown in Fig. 6, under the default setting, the attention distribution is dominated by the end positions, indicating that the model overly focuses on a few images while neglecting most others. This observation aligns with our findings in §3.2. The application of SoFA alters this trend by significantly smoothing the attention distribution, enabling the model to attend to images at all positions more evenly. This well explains how SoFA mitigates position bias." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.909, + 0.901 + ], + "angle": 0, + "content": "SoFA improves in-context capability of LVLMs. Multimodal in-context learning is an emerging property recently acquired by LVLMs, wherein a few demonstrations allow them to recognize task formats, thereby significantly improving accuracy. However, recent studies [6, 9, 30] find that LVLMs tend to overlook demonstration images that appear earlier, leading to suboptimal performance. Here, we demonstrate that SoFA can improve the in-context learning capability of LVLMs by eliminating such blind spots. Following [30], We select three typical visual question answering benchmarks: VQAv2 [16], VizWiz [17], and OKVQA [38], and report the average accuracy of LVLMs across varying shot counts. As illustrated in Table 3, the application of SoFA leads to a further enhancement in LVLM performance, with this effect becoming increasingly pronounced as the number of shots rises. For instance, in the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10605" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.091, + 0.473, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.105, + 0.237, + 0.468, + 0.251 + ], + "angle": 0, + "content": "Figure 6. The attention distribution across positions on PQA." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.262, + 0.485, + 0.328 + ], + "angle": 0, + "content": "
MethodVQAv2 [16]VizWiz [17]OK-VQA [38]
4-shot8-shot16-shot4-shot8-shot16-shot4-shot8-shot16-shot
w/o SoFA55.3459.3261.3741.5644.6545.3545.1948.3049.02
w/ SoFA56.1361.3064.8542.3046.9849.1745.9349.8752.36
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.339, + 0.483, + 0.368 + ], + "angle": 0, + "content": "Table 3. The average accuracy of LVLMs on visual question answering using in-context learning with and without SoFA." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.378, + 0.473, + 0.429 + ], + "angle": 0, + "content": "
# Images351020305075100
w/o SoFA82.1681.2777.1872.3668.5060.8055.1349.19
w/ SoFA83.9082.3378.9274.3070.6763.2759.3855.11
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.441, + 0.483, + 0.469 + ], + "angle": 0, + "content": "Table 4. The average accuracy of LVLMs on Multi-modal Haystack [59], where images range from a few to a hundred." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.478, + 0.483, + 0.538 + ], + "angle": 0, + "content": "case of VizWiz, accuracy improves from \\(41.56\\%\\) to \\(42.30\\%\\) with 4 shots, and from \\(45.35\\%\\) to \\(49.17\\%\\) with 16 shots. This indicates that position bias undermines the effectiveness of demonstrations, while SoFA mitigates this issue." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.483, + 0.765 + ], + "angle": 0, + "content": "SoFA exhibits advantages in long-context applications. Long-context scenarios have recently become a focus, where a single inference requires processing a substantial number of images. This is particularly beneficial for applications such as many-shot in-context learning [20] and multi-modal RAG [62]. Here, we show that SoFA can excel in these challenging scenarios. We select Visual Haystack [61], a benchmark designed to assess the long-context capabilities of LVLMs, and increase the input scale to encompass anywhere from a few images to a hundred. As shown in Table 4, the benefits of SoFA become increasingly evident as the number of images rises, achieving peak performance gains at 100 images (49.19% → 55.11%). This indicates that an increase in context length exacerbates the model's position bias, while SoFA compensates this loss." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Contribution of SoFA to different tasks. Extending beyond the holistic evaluation, we categorize the benchmarks into various tasks to identify which ones benefit the most from SoFA. As depicted in Fig. 7, SoFA demonstrates the highest improvements in visual retrieval and analogy, with gains of \\(6.84\\%\\) and \\(5.53\\%\\), respectively. In these tasks, the first image acts as a reference, and the models need to select a matching scene or similar style from the remaining images. This necessitates that the model pays extra attention" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.089, + 0.897, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.212, + 0.907, + 0.24 + ], + "angle": 0, + "content": "Figure 7. The performance gains of SoFA on different types of tasks. The results are averaged over selected models." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.252, + 0.899, + 0.342 + ], + "angle": 0, + "content": "
σBLINKMantis-EvalMuirBenchMIRBNLVR2MVBench
0.0047.8551.8147.7250.2784.7734.87
0.2549.2252.7049.3552.3485.4137.28
0.5050.0453.3350.9651.8686.3036.56
0.7549.6253.7448.2451.1486.8335.07
1.0048.2151.5847.3750.5385.3734.20
" + }, + { + "type": "table_caption", + "bbox": [ + 0.523, + 0.352, + 0.895, + 0.366 + ], + "angle": 0, + "content": "Table 5. The results on six benchmarks varying the value of \\( \\sigma \\) ." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.38, + 0.905, + 0.456 + ], + "angle": 0, + "content": "to and adequately understands the first image. Intuitively, due to the recency bias inherent in open-source models, they often overlook earlier information, resulting in suboptimal performance. Thus, the significant gains brought by SoFA are understandable." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.459, + 0.907, + 0.61 + ], + "angle": 0, + "content": "The interpretation of \\(\\sigma\\). In Table 5, we conduct an ablation study on the effect of \\(\\sigma\\). It can be observed that the optimal \\(\\sigma\\) varies significantly across benchmarks, for instance, with a value of 0.75 on NLVR2 and only 0.25 on MVBench. Intuitively, MVBench inputs involve video frames, which have a strong dependence on image positions, and an excessively high \\(\\sigma\\) may compromise such positional information. In contrast, the examples in NLVR2 are image pairs, most of which are position-agnostic. This difference in task nature accounts for the variation in \\(\\sigma\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.631, + 0.634, + 0.647 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.659, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this study, we introduce the concept of position bias, a well-known issue in NLP, into the realm of multi-modal applications. First, we reveal that LVLMs cannot robustly make use of information across multiple images, where merely altering the position of images may significantly change their predictions. Then we further identify this position bias, concluding that open-source models exhibit a strong recency bias, while proprietary models struggle with positions in the middle. Motivated by these findings, we propose SoFA, a simple, training-free approach that reduces the model's dependence on positional information through linear interpolation between inter-image causal attention and bidirectional counterparts. Experiments demonstrate that SoFA effectively mitigates position bias and significantly enhances the overall reasoning performance of existing models." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10606" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.092, + 0.226, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.11, + 0.484, + 0.195 + ], + "angle": 0, + "content": "This research was, in part, funded by the U.S. Government - DARPA TIAMAT HR00112490421. The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the U.S. Government." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.223, + 0.188, + 0.239 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.248, + 0.484, + 0.317 + ], + "angle": 0, + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.319, + 0.484, + 0.457 + ], + "angle": 0, + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob L. Menick, Sebastian Borgeaud, Andy Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karén Simonyan. Flamingo: a visual language model for few-shot learning. In NeurIPS, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.459, + 0.483, + 0.514 + ], + "angle": 0, + "content": "[3] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In CVPR, pages 2425-2433, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.517, + 0.484, + 0.599 + ], + "angle": 0, + "content": "[4] Anas Awadalla, Irena Gao, Josh Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Shiori Sagawa, et al. Openflamingo: An opensource framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.602, + 0.484, + 0.657 + ], + "angle": 0, + "content": "[5] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.659, + 0.484, + 0.714 + ], + "angle": 0, + "content": "[6] Folco Bertini Baldassini, Mustafa Shukor, Matthieu Cord, Laure Soulier, and Benjamin Piwowarski. What makes multimodal in-context learning work? In CVPR, pages 1539-1550, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.716, + 0.484, + 0.758 + ], + "angle": 0, + "content": "[7] Rohan Bavishi, Erich Elsen, Curtis Hawthorne, Maxwell Nye, Augustus Odena, Arushi Somani, and Săgnak Taşirlar. Introducing our multimodal models, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.759, + 0.484, + 0.814 + ], + "angle": 0, + "content": "[8] Liangyu Chen, Bo Li, Sheng Shen, Jingkang Yang, Chunyuan Li, Kurt Keutzer, Trevor Darrell, and Ziwei Liu. Large language models are visual reasoning coordinators. In NeurIPS, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.817, + 0.484, + 0.871 + ], + "angle": 0, + "content": "[9] Shuo Chen, Zhen Han, Bailan He, Mark Buckley, Philip Torr, Volker Tresp, and Jindong Gu. Understanding and improving in-context learning on vision-language models. In ICLR, 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[10] Yuhan Chen, Ang Lv, Ting-En Lin, Changyu Chen, Yuchuan Wu, Fei Huang, Yongbin Li, and Rui Yan. Fortify the short" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.248, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "est stave in attention: Enhancing context awareness of large language models for effective tool use. In ACL, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.124, + 0.905, + 0.192 + ], + "angle": 0, + "content": "[11] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In CVPR, pages 24185–24198, 2024. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.195, + 0.905, + 0.277 + ], + "angle": 0, + "content": "[12] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with \\(90\\%\\) chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2(3):6, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.281, + 0.905, + 0.336 + ], + "angle": 0, + "content": "[13] Neil Cohn, Ryan Taylor, and Kaitlin Pederson. A picture is worth more words over time: Multimodality and narrative structure across eight decades of american superhero comics. Multimodal Communication, 6(1):19-37, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.339, + 0.905, + 0.394 + ], + "angle": 0, + "content": "[14] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. In NeurIPS, pages 16344-16359, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.397, + 0.905, + 0.453 + ], + "angle": 0, + "content": "[15] Xingyu Fu, Yushi Hu, Bangzheng Li, Yu Feng, Haoyu Wang, Xudong Lin, Dan Roth, Noah A Smith, Wei-Chiu Ma, and Ranjay Krishna. Blink: Multimodal large language models can see but not perceive. In ECCV, 2024. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.455, + 0.905, + 0.512 + ], + "angle": 0, + "content": "[16] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In CVPR, pages 6904-6913, 2017. 1, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.514, + 0.905, + 0.569 + ], + "angle": 0, + "content": "[17] Danna Gurari, Qing Li, Abigale J Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo Luo, and Jeffrey P Bigham. Vizwiz grand challenge: Answering visual questions from blind people. In CVPR, pages 3608-3617, 2018. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.572, + 0.905, + 0.655 + ], + "angle": 0, + "content": "[18] Junqing He, Kunhao Pan, Xiaoqun Dong, Zhuoyang Song, LiuYiBo LiuYiBo, Qiangosun Qianguosun, Yuxin Liang, Hao Wang, Enming Zhang, and Jiaxing Zhang. Never lost in the middle: Mastering long-context question answering with position-agnostic decompositional training. In ACL, pages 13628–13642, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.657, + 0.905, + 0.713 + ], + "angle": 0, + "content": "[19] Dongfu Jiang, Xuan He, Huaye Zeng, Cong Wei, Max Ku, Qian Liu, and Wenhu Chen. Mantis: Interleaved multi-image instruction tuning. arXiv preprint arXiv:2405.01483, 2024. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.716, + 0.905, + 0.77 + ], + "angle": 0, + "content": "[20] Yixing Jiang, Jeremy Irvin, Ji Hun Wang, Muhammad Ahmed Chaudhry, Jonathan H Chen, and Andrew Y Ng. Many-shot in-context learning in multimodal foundation models. In ICML, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.773, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[21] Mehran Kazemi, Nishanth Dikkala, Ankit Anand, Petar Devic, Ishita Dasgupta, Fangyu Liu, Bahare Fatemi, Pranjal Awasthi, Dee Guo, Sreenivas Gollapudi, et al. Remi: A dataset for reasoning with multiple images. arXiv preprint arXiv:2406.09175, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[22] Hugo Laurençon, Andrés Marafioti, Victor Sanh, and Léo Tronchon. Building and better understanding vision-language models: insights and future directions. arXiv preprint arXiv:2408.12637, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10607" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[23] Hugo Laurençon, Lucile Saulnier, Léo Tronchon, Stas Bekman, Amanpreet Singh, Anton Lozhkov, Thomas Wang, Siddharth Karamcheti, Alexander Rush, Douwe Kiela, et al. Obelics: An open web-scale filtered dataset of interleaved image-text documents. In NeurIPS, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[24] Hugo Laurençon, Léo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language models? arXiv preprint arXiv:2405.02246, 2024. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.275 + ], + "angle": 0, + "content": "[25] Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Kuttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al. Retrieval-augmented generation for knowledge-intensive nlp tasks. In CVPR, pages 9459-9474, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[26] Bohao Li, Yuying Ge, Yixiao Ge, Guangzhi Wang, Rui Wang, Ruimao Zhang, and Ying Shan. Seed-bench: Benchmarking multimodal large language models. In CVPR, pages 13299-13308, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.401 + ], + "angle": 0, + "content": "[27] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava next-interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.404, + 0.482, + 0.46 + ], + "angle": 0, + "content": "[28] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, pages 19730–19742. PMLR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.482, + 0.517 + ], + "angle": 0, + "content": "[29] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mybench: A comprehensive multi-modal video understanding benchmark. In CVPR, pages 22195-22206, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.482, + 0.572 + ], + "angle": 0, + "content": "[30] Li Li, Jiawei Peng, Huiyi Chen, Chongyang Gao, and Xu Yang. How to configure good in-context sequence for visual question answering. In CVPR, pages 26710-26720, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.575, + 0.482, + 0.617 + ], + "angle": 0, + "content": "[31] Ji Lin, Hongxu Yin, Wei Ping, Pavlo Molchanov, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models. In CVPR, pages 26689-26699, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.619, + 0.482, + 0.687 + ], + "angle": 0, + "content": "[32] Ziyi Lin, Chris Liu, Renrui Zhang, Peng Gao, Longtian Qiu, Han Xiao, Han Qiu, Chen Lin, Wenqi Shao, Keqin Chen, et al. Sphinx: The joint mixing of weights, tasks, and visual embeddings for multi-modal large language models. arXiv preprint arXiv:2311.07575, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.482, + 0.731 + ], + "angle": 0, + "content": "[33] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.733, + 0.482, + 0.759 + ], + "angle": 0, + "content": "[34] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[35] Nelson F Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. Lost in the middle: How language models use long contexts. In ACL, pages 157-173. MIT Press One Broadway, 12th Floor, Cambridge, Massachusetts 02142, USA ..., 2024. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[36] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. In NeurIPS, pages 2507–2521, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.148 + ], + "angle": 0, + "content": "[37] Pan Lu, Baolin Peng, Hao Cheng, Michel Galley, Kai-Wei Chang, Ying Nian Wu, Song-Chun Zhu, and Jianfeng Gao. Chameleon: Plug-and-play compositional reasoning with large language models. In NeurIPS, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.151, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[38] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In CVPR, pages 3195-3204, 2019. 1, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.209, + 0.905, + 0.278 + ], + "angle": 0, + "content": "[39] Duy MH Nguyen, Hoang Nguyen, Nghiem Diep, Tan Ngoc Pham, Tri Cao, Binh Nguyen, Paul Swoboda, Nhat Ho, Shadi Albarqouni, Pengtao Xie, et al. Lvm-med: Learning large-scale self-supervised vision models for medical imaging via second-order graph matching. In NeurIPS, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.281, + 0.905, + 0.335 + ], + "angle": 0, + "content": "[40] Yassine Ouali, Adrian Bulat, Brais Martinez, and Georgios Tzimiropoulos. Clip-dpo: Vision-language models as a source of preference for fixing hallucinations in lvlms. In ECCV, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.339, + 0.905, + 0.393 + ], + "angle": 0, + "content": "[41] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. In ICLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.397, + 0.905, + 0.438 + ], + "angle": 0, + "content": "[42] Alexander Peysakhovich and Adam Lerer. Attention sorting combats recency bias in long context language models. arXiv preprint arXiv:2310.01427, 2023. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.523 + ], + "angle": 0, + "content": "[43] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.527, + 0.905, + 0.596 + ], + "angle": 0, + "content": "[44] Ori Ram, Yoav Levine, Itay Dalmedigos, Dor Muhlgay, Amnon Shashua, Kevin Leyton-Brown, and Yoav Shoham. Incontext retrieval-augmented language models. In ACL, pages 1316-1331. MIT Press One Broadway, 12th Floor, Cambridge, Massachusetts 02142, USA . . ., 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.599, + 0.905, + 0.655 + ], + "angle": 0, + "content": "[45] Nir Ratner, Yoav Levine, Yonatan Belinkov, Ori Ram, Inbal Magar, Omri Abend, Ehud Karpas, Amnon Shashua, Kevin Leyton-Brown, and Yoav Shoham. Parallel context windows for large language models. In ACL, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.657, + 0.905, + 0.727 + ], + "angle": 0, + "content": "[46] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.729, + 0.905, + 0.784 + ], + "angle": 0, + "content": "[47] Junjie Shentu and Noura Al Moubayed. Cxr-irgen: An integrated vision and language model for the generation of clinically accurate chest x-ray image-report pairs. In WACV, pages 5212–5221, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.787, + 0.905, + 0.842 + ], + "angle": 0, + "content": "[48] Lin Shi, Weicheng Ma, and Soroush Vosoughi. Judging the judges: A systematic investigation of position bias in pairwise comparative assessments by llms. arXiv preprint arXiv:2406.07791, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[49] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10608" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[50] Alane Suhr, Stephanie Zhou, Ally Zhang, Iris Zhang, Huajun Bai, and Yoav Artzi. A corpus for reasoning about natural language grounded in photographs. In ACL, 2018. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.482, + 0.203 + ], + "angle": 0, + "content": "[51] Quan Sun, Yufeng Cui, Xiaosong Zhang, Fan Zhang, Qiying Yu, Yueze Wang, Yongming Rao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative multimodal models are in-context learners. In CVPR, pages 14398-14409, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.204, + 0.482, + 0.245 + ], + "angle": 0, + "content": "[52] Dídac Surís, Sachit Menon, and Carl Vondrick. Vipergpt: Visual inference via python execution for reasoning. In CVPR, pages 11888-11898, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.246, + 0.482, + 0.274 + ], + "angle": 0, + "content": "[53] MosaicML NLP Team et al. Introducing mpt-7b: A new standard for open-source, commercially usable llms, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.275, + 0.482, + 0.343 + ], + "angle": 0, + "content": "[54] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.482, + 0.412 + ], + "angle": 0, + "content": "[55] Pat Verga, Sebastian Hofstatter, Sophia Althammer, Yixuan Su, Aleksandra Piktus, Arkady Arkhangorodsky, Minjie Xu, Naomi White, and Patrick Lewis. Replacing judges with juries: Evaluating llm generations with a panel of diverse models. arXiv preprint arXiv:2404.18796, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.413, + 0.482, + 0.482 + ], + "angle": 0, + "content": "[56] Fei Wang, Xingyu Fu, James Y Huang, Zekun Li, Qin Liu, Xiaogeng Liu, Mingyu Derek Ma, Nan Xu, Wenxuan Zhou, Kai Zhang, et al. Muirbench: A comprehensive benchmark for robust multi-image understanding. arXiv preprint arXiv:2406.09411, 2024. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.482, + 0.552 + ], + "angle": 0, + "content": "[57] Hengyi Wang, Haizhou Shi, Shiwei Tan, Weiyi Qin, Wenyuan Wang, Tunyu Zhang, Akshay Nambi, Tanuja Ganu, and Hao Wang. Multimodal needle in a haystack: Benchmarking long-context capability of multimodal large language models. arXiv preprint arXiv:2406.11230, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.482, + 0.607 + ], + "angle": 0, + "content": "[58] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, et al. Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.608, + 0.482, + 0.663 + ], + "angle": 0, + "content": "[59] Weiyun Wang, Shuibao Zhang, Yiming Ren, Yuchen Duan, Tiantong Li, Shuo Liu, Mengkang Hu, Zhe Chen, Kaipeng Zhang, Lewei Lu, et al. Needle in a multimodal haystack. In NeurIPS, 2024. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.664, + 0.482, + 0.72 + ], + "angle": 0, + "content": "[60] Ziqi Wang, Hanlin Zhang, Xiner Li, Kuan-Hao Huang, Chi Han, Shuiwang Ji, Sham M Kakade, Hao Peng, and Heng Ji. Eliminating position bias of language models: A mechanistic approach. arXiv preprint arXiv:2407.01100, 2024. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.721, + 0.482, + 0.775 + ], + "angle": 0, + "content": "[61] Tsung-Han Wu, Giscard Biamby, Jerome Quenum, Ritwik Gupta, Joseph E Gonzalez, Trevor Darrell, and David M Chan. Visual haystacks: Answering harder questions about sets of images. arXiv preprint arXiv:2407.13766, 2024. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.482, + 0.831 + ], + "angle": 0, + "content": "[62] Michihiro Yasunaga, Armen Aghajanyan, Weijia Shi, Rich James, Jure Leskovec, Percy Liang, Mike Lewis, Luke Zettlemoyer, and Wen-tau Yih. Retrieval-augmented multimodal language modeling. In ICML, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[63] Yijiong Yu, Huiqiang Jiang, Xufang Luo, Qianhui Wu, ChinYew Lin, Dongsheng Li, Yuqing Yang, Yongfeng Huang, and Lili Qiu. Mitigate position bias in large language models via scaling a single dimension. arXiv preprint arXiv:2406.02536, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[64] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In CVPR, pages 9556-9567, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[65] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In CVPR, pages 11975-11986, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.206, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[66] Bingchen Zhao, Yongshuo Zong, Letian Zhang, and Timothy Hospedales. Benchmarking multi-image understanding in vision and language models: Perception, knowledge, reasoning, and multi-hop reasoning. arXiv preprint arXiv:2406.12742, 2024. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.276, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[67] Haozhe Zhao, Zefan Cai, Shuzheng Si, Xiaojian Ma, Kaikai An, Liang Chen, Zixuan Liu, Sheng Wang, Wenjuan Han, and Baobao Chang. Mmcl: Empowering vision-language model with multi-modal in-context learning. In ICLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.346, + 0.905, + 0.415 + ], + "angle": 0, + "content": "[68] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. In NeurlPS, pages 46595-46623, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.417, + 0.905, + 0.471 + ], + "angle": 0, + "content": "[69] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. In ICLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.473, + 0.905, + 0.543 + ], + "angle": 0, + "content": "[70] Wanrong Zhu, Jack Hessel, Anas Awadalla, Samir Yitzhak Gadre, Jesse Dodge, Alex Fang, Youngjae Yu, Ludwig Schmidt, William Yang Wang, and Yejin Choi. Multimodal c4: An open, billion-scale corpus of images interleaved with text. In NeurIPS, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.544, + 0.905, + 0.598 + ], + "angle": 0, + "content": "[71] Yuchen Zhuang, Yue Yu, Kuan Wang, Haotian Sun, and Chao Zhang. Toolqa: A dataset for llm question answering with external tools. In NeurIPS, pages 50117-50143, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.598 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10609" + } + ] +] \ No newline at end of file diff --git a/2025/Identifying and Mitigating Position Bias of Multi-image Vision-Language Models/17ea1e19-b848-462e-8296-051308ed2dd4_origin.pdf b/2025/Identifying and Mitigating Position Bias of Multi-image Vision-Language Models/17ea1e19-b848-462e-8296-051308ed2dd4_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..45b73668223990887705712b762d97028b33f48d --- /dev/null +++ b/2025/Identifying and Mitigating Position Bias of Multi-image Vision-Language Models/17ea1e19-b848-462e-8296-051308ed2dd4_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:14a6e6565828af5572a2ad7d09af747a5e3c9b019435b648160c086235827ee9 +size 1090757 diff --git a/2025/Identifying and Mitigating Position Bias of Multi-image Vision-Language Models/full.md b/2025/Identifying and Mitigating Position Bias of Multi-image Vision-Language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..772ad548bfface282166adda8926ec6a463f23fe --- /dev/null +++ b/2025/Identifying and Mitigating Position Bias of Multi-image Vision-Language Models/full.md @@ -0,0 +1,333 @@ +# Identifying and Mitigating Position Bias of Multi-image Vision-Language Models + +Xinyu Tian1 Shu Zou1 Zhaoyuan Yang2 Jing Zhang1 +1Australian National University 2GE Research + +$^{1}$ firstname_lastname@anu.edu.au, $^{2}$ firstname_lastname@ge.com + +# Abstract + +The evolution of Large Vision-Language Models (LVLMs) has progressed from single to multi-image reasoning. Despite this advancement, our findings indicate that LVLMs struggle to robustly utilize information across multiple images, with predictions significantly affected by the alteration of image positions. To further explore this issue, we introduce Position-wise Question Answering (PQA), a meticulously designed task to quantify reasoning capabilities at each position. Our analysis reveals a pronounced position bias in LVLMs: open-source models excel in reasoning with images positioned later but underperform with those in the middle or at the beginning, while proprietary models show improved comprehension for images at the beginning and end but struggle with those in the middle. Motivated by this, we propose SoFt Attention (SoFA), a simple, training-free approach that mitigates this bias by employing linear interpolation between inter-image causal attention and bidirectional counterparts. Experimental results demonstrate that SoFA reduces position bias and enhances the reasoning performance of existing LVLMs. The code will be available https://github.com/xytian1008/sofa. + +# 1. Introduction + +The advent of Large Vision-Language Models (LVLMs) [1, 2, 5, 24, 34] has catalyzed the development of various text-grounded visual applications, including visual question answering [3, 16, 36], code generation [8, 37, 52], and story telling [13, 26]. Open-source LVLMs such as LLaVA [34], BLIP2 [28], and InternVL [11] have demonstrated competitive performance on single-image tasks like VQAv2 [16], OKVQA [38], and MMMU [64]. However, these LVLMs are limited in their capacity to analyze interimage differences and relationships, as well as to process temporal information. Consequently, recent LVLMs, such as Flamingo [2, 4], Idefics [22-24], and Emu2 [51], have incorporated image-text interleaved data during training to enable ability to tackle multi-image tasks [15, 19, 50, 56]. + +Despite the increasing scale and performance of LVLMs, + +the interpretation of how they reason across multiple images remains a mystery. In fact, we might find some hints in the topic of natural language processing (NLP). Recent research in NLP has highlighted retrieval-augmented generation (RAG) [25, 44, 71], which aims to enable large language models (LLMs) to reason across multiple documents to answer questions. However, [35] find that LLMs tend to effectively comprehend documents positioned at the beginning and end of the input, often neglecting the middle sections. Similarly, in LLM-as-a-judge [48, 55], [68] find that LLMs used to evaluate two model responses tend to favor the first one, regardless of content. This phenomenon, where the model's reasoning capability is influenced by the segment position, is referred to as position bias [35, 42, 60]. + +Motivated by the research above, a natural question arises: does position bias exist in LVLMs? Specifically, we seek to determine whether LVLMs can robustly make use of information across multiple images. To investigate this, we begin by evaluating LVLMs on established multi-image benchmarks [15, 19, 50, 56]. We find that simply changing the order of presented images may significantly alter the predictions for most LVLMs, leading to considerable variation in benchmark accuracy. Fig. 1 provides examples of questions and model responses. For instance, in difference spotting [56], swapping two images causes OpenFlamingo [4] to miss distinctions, while in a simple counting task [19], reordering images leads Idefics2 [24] to incorrect reasoning. This positional sensitivity is apparent even in GPT-4o [1], a leading proprietary model. + +The above motivational study reveals an unfortunate fact: building upon LLMs, LVLMs may inherit the notorious position bias from their language component. This bias manifests as the model's tendency to excel at reasoning visual information from certain positions, while exhibiting weaknesses at others, compromising robustness and reliability in predictions. To further investigate which positions display strong understanding and which reveal weaknesses, i.e., identifying the pattern of such position bias, we introduce Position-wise Question Answering (PQA). Unlike most multi-image tasks that assess holistic comprehension, PQA requires LVLMs to reason consistently across each + +![](images/544f5f38c0745927eb794d5dd9747f98e1aa05cc89c160727c5081df15da5c58.jpg) +Figure 1. Examples where the predictions of LVLMs are influenced by changes in the image positions (Date accessed: Oct 25, 2024). The tasks, from left to right, are difference spotting [56], counting [19] and forensic detection [15], respectively. The correct answers are highlighted in green, while the incorrect ones are highlighted in red. The prompts and outputs are simplified here for clarity and aesthetics. + +image and produce position-wise responses. This approach enables us to measure position-wise accuracy and analyze the model's reasoning capacity across different positions. + +Our analysis reveals that most open-source models, e.g., OpenFlamingo [4] and Idefics2 [24], exhibit a pronounced recency bias, demonstrating stronger reasoning capabilities for images presented later while underperforming on those in the middle or at the beginning. In contrast, proprietary models like GPT-4o [1] show enhanced understanding of images at the beginning and end, while neglecting those in the middle, a pattern consistent with previous findings in LLMs [35]. This observation is further corroborated by Fig. 1, where LVLMs tend to reason incorrectly about images positioned in poor-performing areas. + +The existence of position bias significantly compromises model robustness and undermines accuracy across various multi-image tasks. Inspired by this, we aim to investigate the underlying mechanical causes and mitigate this issue. Specifically, we notice that in most LVLMs, the hidden states of image tokens are heavily influenced by their relative positions due to inter-image causal attention. In particular, image tokens are treated as a type of special text tokens where 1) images appearing later may interact with preceding ones; 2) images at the beginning are isolated, lacking access to global context. This autoregressive setup implicitly imposes a semantic order on the presented images, amplifying their positional information. To address this, we introduce SoFt Attention (SoFA), a training-free approach to mitigate position bias across various LVLMs. SoFA reduces the dependency of image hidden states on positional information by simply performing linear interpolation between inter-image causal attention and its bidirectional counterparts. Experimental results indicate that SoFA effectively equilibrates the reasoning capabilities of LVLMs across various positions, achieving average accuracy improvements of $2 \sim 3\%$ on existing multi-image benchmarks. + +In summary, our contributions are as follows: + +- Despite the significant potential of LVLMs in multi-image applications, we find that position bias, a well-explored phenomenon in NLP, has emerged as a potential bottleneck in the development of LVLMs, substantially impacting model robustness and reliability. +- We further identify such position bias and discover that most open-source models demonstrate a pronounced recency bias, whereas proprietary models tend to struggle in middle positions. +- We introduce SoFA, a straightforward training-free approach designed to mitigate the position bias inherited from pre-trained LVLMs, and substantiate its effectiveness through experimental validation. + +# 2. Related Work + +Large Vision-Language Models. Existing LVLMs are generally composed of a vision encoder [43, 65], an LLM [12, 53, 54], and a bridging projector [2, 7, 27] that facilitates interaction between the two. Conventional LVLMs, pre-trained on image-text pairs [46] and fine-tuned through preference [33, 40] or instruction tuning [27, 41], demonstrate substantial success in applications like dialogue [5, 11], question answering [58, 69], and domain-specific tasks [39, 47]. However, constrained by their training data, these models primarily support single-image reasoning during inference. Recently, the integration of image-text interleaved data [23, 70] has enabled multi-image reasoning capabilities, encompassing image comparison [21, 56], association [15, 19], temporal relationship analysis [29], and emerging properties of multi-modal in-context learning [30, 67]. Nonetheless, limited research has delved into how LVLMs perform reasoning across multiple images or the underlying mechanisms driving this process. Position Bias in LLMs. Position bias, an inherent side effect of long-context input, has increasingly captured the at + +![](images/181e6ae6c93931b3eeb9d154c5f1e22c5557bb7f10789500efd59f45f2ce356e.jpg) +Figure 2. The results of multiple evaluations while shuffling image positions. We record the minimum, maximum, and average accuracy (left vertical axis), along with the prediction inconsistency between the best and worst-performing evaluations (right vertical axis). + +tention of NLP researchers. The pioneering work on this issue, notably [35], reveals that when LLMs process multiple documents, they often concentrate on those at the beginning and end while disregarding much of the information in between, i.e., a phenomenon termed "lost in the middle". A related issue has also been identified in LLM-as-a-judge [48, 55], where [68] observe that LLMs tasked with evaluating multiple model responses tend to favor the first one, irrespective of its content. Recent studies have begun to explore methods to mitigate or circumvent position bias in LLMs [42, 45, 60, 63]. However, to the best of our knowledge, few studies have examined the effects of position bias within multi-modal applications, specifically regarding its implications for LVLMs. + +Position Bias in LVLMs. While systematic research on position bias in LVLMs is still lacking, recent studies have indirectly underscored the influence of positions on LVLMs through various findings. For instance, in multi-modal in-context learning, studies [9] and [6] reveal that demonstration images placed at the beginning contribute minimally to LVLM performance. Additionally, in assessing long-context capabilities, [61] reports that LVLMs attain particularly high accuracy when key information appears at the end of the input sequence, with similar results observed in [59] and [57]. However, previous research has neither thoroughly analyzed this position bias nor proposed solutions to address it. In contrast, this study systematically investigates position bias in both open-source and proprietary models and introduces an effective mitigation strategy. + +# 3. Understanding Position Bias in LVLMs + +In this section, we investigate whether the position bias observed in LLMs extends to LVLMs. We start by empirically confirming the existence of position bias in LVLMs through a motivational study, revealing their poor robustness to image position changes (§3.1). Subsequently, we analyze the nature of this bias by evaluating and quantifying the reasoning capabilities of LVLMs regarding different positions (§3.2). These insights contribute to a deeper understanding and interpretation of LVLMs' reasoning behavior and common failure cases. + +# 3.1. The Position Matters + +To assess the existence of position bias, we conduct a simple motivational experiment, as shown in Fig. 2. + +Experimental Setup. We select four representative multi-image LVLMs: OpenFlamingo [4], Idefics2 [24], LLaVA-NeXT-Interleave1 [27], GPT-4o [1], and conduct evaluations across established multi-image benchmarks, including BLINK [15], Mantis-Eval [19], MuirBench [56], and MIRB [66]. To ensure clarity, we select partial tasks from these benchmarks that feature position-agnostic questions, wherein the positions of the images do not affect the answers to the questions. See §5.1 for the selected tasks and details. To explore the influence of image positions on the results, we perform multiple evaluations per benchmark, with the positions of the images shuffled in each iteration. + +Metric. We record the minimum, maximum, and average accuracy across these evaluations, which directly reflects the influence of positions on the performance of LVLMs. Furthermore, we assess the prediction inconsistency between the best and worst-performing evaluations, quantifying the proportion of examples exhibiting conflicting predictions across the two evaluations by changing positions. + +Results. As illustrated in Fig. 2, for most open-source LVLMs, the prediction inconsistency between the best and worst-performing evaluations approaches $30\%$ , suggesting that merely altering the positions of the presented images may result in an amount of the model's predictions being changed. Such pronounced sensitivity to positional information results in highly unstable performance. For example, OpenFlamingo's accuracy on BLINK can vary significantly, reaching $38\%$ at its peak and dropping to $30\%$ at its lowest, indicating an accuracy span of up to $10\%$ with position changes. GPT-4o, the state-of-the-art proprietary model, although showcasing superior performance, also displays a lack of robustness to positional variations, leading to a prediction inconsistency of around $25\%$ . + +To qualitatively explore this issue, we present examples in Fig. 1 that illustrate prediction conflicts arising from alterations in image positions. For instance, when prompted to identify the differences between two images, Open- + +![](images/77d93e66f4813cf90930cad4924e06f53ba8952058b77d12b00d011c9d6544a7.jpg) + +![](images/39eddef4ed52608f0464f7bf9d90a1598601cc958984136679f48fda3834dce8.jpg) + +![](images/a13d5253052a0753a3be84751bde8a204485818747cf8ba13a2a370588e24dc4.jpg) + +![](images/39f202d2bf24294a94d1e561929a75953d8a6e91798094ee0c9923cc90cc7b8e.jpg) + +![](images/e648dd5b0b16ce27c0de8461b5e20b4b1ae82f22893e4c098b4823dcdc0c4ea8.jpg) +Figure 3. The results on the PQA task. We report position-wise accuracy on four scenarios where the number of images is 5, 10, 15 and 20, respectively. A higher accuracy signifies strong reasoning at the specified position, whereas lower reflects poor-performing areas. + +![](images/028a2c900e079d23a835b08d56212d9e899da15d69cda61c12a58f0ef8bd6ca5.jpg) + +![](images/85791aa4be6580e62ea44bc5f0d9c0f705e6f140e452e37f4a9544776ce24477.jpg) + +![](images/351071bd03901fddff5c3585164da708ea4d291717f9011d5b1a3971f99e3521.jpg) + +![](images/5e32f62575144608556a8130ccd261f44c99783fd4ec9db7cfef37144574e644.jpg) + +Flamingo fails to extract fine-grained distinctions upon swapping their positions. Similarly, in the task of object counting, Idefics2 mistakenly identifies a dog as a cat due to the repositioning of images. GPT-4o demonstrates a comparable issue, where rearranging the image options alters both its answer and the accompanying rationale. + +# 3.2. Identifying Position Bias + +The quantitative results and qualitative examples above imply biased reasoning across positions from LVLMs. To have a closer look at such position bias, we aim to identify models' reasoning capabilities at different image positions. + +However, existing tasks have been restricted to evaluating models' holistic comprehension of multiple images, lacking the means to systematically analyze their fine-grained understanding of each individual position. Specifically, these tasks provide coarse ground truth answers, such as option symbols or numeric values, but do not offer individual rationales for each image, making it challenging to quantify the model's reasoning across different positions. To address this gap, we introduce Position-wise Question Answering (PQA), a carefully designed task that assesses the position-wise reasoning capabilities of LVLMs. In PQA, each question prompts LVLMs to reason independently across all images, generating a distinct response for each. This approach 1) ensures task-level neutrality with respect to any particular position and 2) enables quantification of LVLMs' reasoning performance for each position. + +Specifically, taking object counting as an example, we employ the following query template: + +I: Please provide one answer for each image in the form like $[x, x, x, \ldots]$ . +Q: Among , ..., , how many cats can you find in these images? +A: [3, 2, 0, ...]. + +Each query consists of an instruction (I) and a question $(\mathbb{Q})$ where the former prompts the model to generate position-wise responses (A) following a specified output format. We gather responses from LVLMs as a structured list, with each element representing a particular position. This enables us + +to track position-wise accuracy, where a higher accuracy reflects stronger model capability at a given position, while lower indicates areas of relative weakness. + +Construction of position-wise examples. Since in PQA, each image is semantically independent of others, a PQA example essentially consists of several single-image instances that share the same question. This allows us to easily build position-wise examples from single-image benchmarks. Specifically, we select VQAv2 [16], a substantial single-image dataset that encompasses a diverse array of everyday topics and features both multiple-choice and open-ended questions. Initially, we identify and consolidate single-image examples with identical questions to create a question set, wherein each question corresponds to multiple images. Subsequently, we reformulate these questions into PQA format according to the above-mentioned template. + +Experimental Setup. We utilize the same LVLMs as previously mentioned and conduct evaluations on the proposed PQA task. Additionally, we also consider random guessing, where binary choice and numeric questions are answered randomly with yes or no, or with an integer value. We examine four scenarios with varying numbers of images: 5, 10, 15, and 20, collecting 1,000 examples for each scenario. Furthermore, to maintain fairness, we augment each example by shuffling the image positions four times, resulting in a total of 5,000 examples. In contrast to prior settings that evaluate holistic performance, we provide separate accuracy metrics for each position in this analysis. + +Results. Fig. 3 displays the results of LVLMs on the PQA task, from which we may derive the following insights: + +1) Open-source LVLMs exhibit recency bias, whereas proprietary models struggle in middle positions. For open-source models, a stronger reasoning capability is observed for images positioned at the end, while significant underperformance is noted in the middle and beginning. For example, with 10 images, the accuracy of OpenFlamingo varies by nearly $9\%$ between the far left and far right positions $(51.84\% \rightarrow 59.71\%)$ . In contrast, proprietary models, i.e., GPT-40, reveal a typical U-shaped curve, highlighting pronounced weaknesses in the middle positions. + +2) Position bias becomes more pronounced as the number of images increases. For instance, with 20 images, the accuracy gap for OpenFlamingo between the two ends reaches $14\%$ ( $36.79\% \rightarrow 50.17\%$ ). Notably, at this stage, OpenFlamingo seems to nearly lost its reasoning capability for most of the beginning images, achieving results only marginally better than random guessing ( $34.49\%$ ). +3) Relationship to position bias in LLMs. Previous research [35] shows that LLMs often overlook information in middle positions when processing multiple documents, a phenomenon termed "lost in the middle". Our findings suggest this effect persists for GPT in multi-image reasoning. However, for open-source models, the primacy position advantage appears to diminish, showing a marked decaying trend from the end to beginning. Whether the position bias in LVLMs is correlated with that of their LLM backbones remains an interesting question for further exploration. + +# 4. The Proposed Method + +In this section, we first analyze the underlying causes of position bias (§4.1). Then we introduce SoFA, a straightforward, training-free approach to mitigate this issue (§4.2). + +# 4.1. Position Mechanism in LVLMs + +To investigate the origins of position bias, we need to first determine what contributes to the positional information in LVLMs. In NLP, it is generally believed that position embeddings [45, 49] and causal attention [42, 60] are the primary mechanisms enabling the sequential structure of input. Formally, + +$$ +\mathbf {Q} = \mathcal {P} \left(\mathbf {X} \mathbf {W} _ {\mathbf {q}}, \mathbf {p o s}\right), \quad \mathbf {K} = \mathcal {P} \left(\mathbf {X} \mathbf {W} _ {\mathbf {k}}, \mathbf {p o s}\right) \tag {1} +$$ + +$$ +\mathbf {H} = \operatorname {S o f t m a x} \left(\mathbf {Q K} ^ {\mathrm {T}} / \sqrt {d}\right) \odot \mathbb {1} _ {\text {c a u s a l}} \mathbf {V} +$$ + +where $\mathbf{Q}$ , $\mathbf{K}$ , and $\mathbf{V}$ represent the queries, keys, and values, respectively. $\mathbf{X}$ refers to the input tokens, encompassing both image and text types, pos denotes the position embedding of $\mathbf{X}$ , which is integrated into the attention via the position encoding function $\mathcal{P}$ , $\mathbb{1}_{\mathrm{causal}}$ denotes the causal mask, which ensures autoregressive generation. + +This reveals two key factors influencing LVLMs' positional awareness: 1) position embedding explicitly initializes a sequential value to each token, and 2) causal attention implicitly enhances this positional characteristic through unidirectional information access. This presents us with two distinct solutions. One approach is to directly reassign position embeddings, a widely used strategy in NLP [10, 18, 45], as it fundamentally eliminates the influence of position on documents. However, this aggressive method is not suitable for LVLMs, since in many scenarios, such as video understanding and action recognition, we still expect images to retain positional information for temporal analysis. To clarify, our aim is to mitigate the imbalanced + +![](images/9ee0e20336200991d8dcd43aba63d83b448ed3b7a1e38bc0a19721f392c90721.jpg) +(A) Causal Attention + +![](images/954e46ff412889d6a74d516d7464db5e1041352826b2dbfc22ba8d57adbf24b9.jpg) +(B) Isolated Attention + +![](images/a588def4b1cc8792d1edc54a90616a2a265609fef01b76f8ed1b2c324feaa086.jpg) +(C) Bidirectional Attention +Figure 4. The three inter-image attention mechanisms, where $I_{1}$ and $I_{2}$ represent the tokens of two images, $T_{1}$ and $T_{2}$ represent the tokens of two text segments. Note that typically each image involves numerous tokens, e.g., 576 for LLaVA. Here for clarity we simplify them to a single token. In (A), images interact in a unidirectional manner, allowing $I_{2}$ to attend to $I_{1}$ , while $I_{1}$ remains isolated. In (B), each image is isolated, indicating they can only attend to themselves. In (C), bidirectional interaction is enabled so that each image can attend to any other images. It is worth mentioning that we only alter the inter-image attention while preserving causal attention between the text segments. + +reasoning capabilities of LVLMs across different positions, rather than to eliminate their positional awareness. + +Therefore, in this paper, we primarily investigate the relationship between causal attention and position bias. We notice that, while images do not exhibit a distinct semantic order as text does, most LVLMs, e.g., LLaVA series, treat image and text tokens in a uniform autoregressive manner. As shown in Fig. 4 (A), analogous to texts, there exists interimage causal attention which allows 1) later-positioned images to interact with those preceding them, while 2) earlier images remain isolated, lacking access to global contexts. From a mechanical perspective, this framework renders image hidden states highly dependent on their relative positions, thus implicitly amplifying their positional cues. + +To verify our hypothesis, we compare the inter-image causal attention against following two alternative variants: + +1) Isolated attention. As illustrated in Fig. 4 (B), we disable interactions between images, enforcing each image to attend solely to itself. +2) Bidirectional attention. Rather than restricting to unidirectional information flow, we permit each image to attend to any other, as shown in Fig. 4 (C). + +We modify the default causal mask in the language decoder to its isolated and bidirectional counterparts and assess the impact of them on position bias in PQA, as shown in Fig. 5. It is evident that altering causal attention to either isolated or bidirectional forms significantly alleviates position bias, resulting in more balanced accuracy across different positions. This empirically confirms that inter-image causal attention is the primary source of position bias. However, these modifications lead to varying degrees of performance degradation, with isolated attention causing the most substantial decline. This outcome is reasonable, as such modifications may substantially deviate from the training framework, leading to out-of-distribution hidden states. + +![](images/33dc7e2e838ab5a2d2cd8bc50703beb61152d00f69180c48ff0582bb9ab7a12a.jpg) +(A) Idefics2 +Figure 5. The position-wise accuracy of the three attention mechanisms on the PQA task, alongside our proposed SoFA method. + +![](images/a77276c3fed29e2dfe69f3f12d0393328d128512c30a1c5591aeebed487915dd.jpg) +(B) LLaVA-NeXT + +# 4.2. Mitigating Position Bias + +The analysis above shows that inter-image causal attention is the main cause of position bias. Furthermore, it also reveals a trade-off between accuracy and robustness, a typical dilemma in machine learning. Building on this insight, we propose SoFt Attention (SoFA), a straightforward, training-free approach to mitigate position bias. SoFA employs linear interpolation between inter-image causal attention and its bidirectional counterpart. Formally, + +$$ +\begin{array}{l} \mathbf {H} _ {\text {s o f t}} = \operatorname {S o f t m a x} \left(\mathbf {Q K} ^ {\mathrm {T}} / \sqrt {d}\right) \odot \mathbf {M} _ {\text {s o f t}} \mathbf {V} \tag {2} \\ \mathbf {M} _ {\text {s o f t}} = (1 - \sigma) \mathbb {1} _ {\text {c a u s a l}} + \sigma \mathbb {1} _ {\text {b i d i r e c t i o n a l}} \\ \end{array} +$$ + +where $\mathbb{1}_{\mathrm{causal}}$ and $\mathbb{1}_{\mathrm{bidual}}$ denote the two attention masks, with $\sigma$ controlling the extent of bidirectional part. Note that since we are only interested in the visual positional relation, we only modify the inter-image attention, while preserving causal attention between text segments. + +SoFA as a position smoother. Intuitively, given that interimage causal attention excessively emphasizes positional information, SoFA integrates partial bidirectional attention to smooth the influence of position on the models at minimal cost, achieving a balanced trade-off between reasoning performance and position robustness. + +SoFA as an interaction enhancer. Practically, SoFA enhances the interaction between images, enriching the semantic information of earlier image features to enable more effective utilization by LVLMs. This aligns with the most cases where key information is typically distributed evenly across various input positions. + +For each task, we utilize a small validation set to determine the optimal values for $\sigma$ . Besides, to better align with the training framework, we deploy SoFA every two layers rather than at every layer, while preserving causal attention in the remaining layers. We present SoFA's results in the PQA task as shown in Fig. 5. It is evident that SoFA significantly outperforms both isolated and bidirectional attention, while unlike causal attention in the default setting, SoFA demonstrates balanced performance and exhibits enhanced reasoning capability in most positions. + +# 5. Experiment and Results + +In this section, we empirically demonstrate that SoFA 1) effectively mitigates the position bias inherent in LVLMs and 2) significantly improves their overall reasoning capability across various multi-image applications. + +Models. We select a diverse set of open-source models, including Idefics2-8B [24], InternVL2-8B [11], VILA-13B [31], Mantis-8B [19], and LLaVA-NeXT-Interleave-7B [27]. + +Implementation details. For the input, we provide instructions that explicitly indicate the task type and output formats. For instance, for multiple-choice questions, we ask LVLMs to provide only the option symbol; for numeric questions, only the number; and for open-ended questions, the response is limited to short answers. For evaluation, to ensure fairness, we uniformly adopt GPT-4 [1], providing with the question, ground truth answer, candidate prediction, and it determines whether the prediction is correct or not. For SoFA, we determine $\sigma$ using a 32-shot validation set per task and perform interpolation every two layers. Unless otherwise specified, all models are set to FP16 precision and utilize Flash Attention [14], with sub-image splitting [32] disabled for fairness. + +# 5.1. Position Robustness + +Benchmark. We first evaluate whether SoFA may effectively mitigate position bias. To this end, similar to §3.1, we select position-agnostic tasks from existing multi-image benchmarks. Specifically, we exclude tasks where changing the order of images may alter the semantics of the question, rendering it unanswerable, e.g., video understanding. Here, we list some of the selected tasks: forensic detection and image jigsaw from BLINK [15], difference spotting and diagram understanding from MuirBench [56], visual analogy and attribute matching from MIRB [66]. We include all examples from Mantis-Eval [19] since the majority are position-agnostic. See the supplementary material for the details of more evaluated tasks. Our experimental setup and metrics are consistent with those described in §3.1. + +SoFA reduces the influence of positions on LVLMs. As demonstrated in Table 1, the integration of SoFA within the models markedly reduces prediction inconsistency, leading to enhanced stability in performance. For instance, Idefics2's inconsistency on BLINK declines from $41.55\%$ to $12.36\%$ , signifying that the responses for the majority of examples remain consistent despite variations in image positioning. This observation indicates that LVLMs exhibit more balanced reasoning capabilities across different positions, effectively mitigating the effects of position bias. + +SoFA eliminates the blind spots of LVLMs. In addition to the enhanced robustness, we observe a marked improvement in the performance of LVLMs in worst-case scenarios, particularly regarding minimum and average accuracy, + +
MethodBLINK [15]Mantis-Eval [19]MuirBench [56]MIRB [66]
MinAveMaxIncMinAveMaxIncMinAveMaxIncMinAveMaxInc
Idefics2 [24]43.2646.9348.6841.5543.9346.6251.1038.5743.2845.3047.8134.4645.2647.8849.6829.94
Idefics2 + SoFA47.1848.6349.1212.3648.7149.1850.838.6146.1247.5148.4513.2747.3948.3649.136.99
InternVL2 [11]38.8140.4543.7430.1845.3646.9249.5125.2449.2852.3356.1038.6541.8444.3846.3629.60
InternVL2 + SoFA41.6442.3243.517.1648.1349.2550.1112.0054.6955.9256.785.1644.2045.3545.876.64
VILA [31]45.9348.5951.4225.3348.4449.2051.8521.2941.6643.1248.2737.2647.1749.5952.3431.77
VILA + SoFA48.2750.8051.1710.6849.2951.6052.6812.1645.7646.5247.137.9248.2251.4351.9517.46
Mantis [19]48.3449.2451.5228.3556.4058.3863.4232.1247.6748.9451.1527.4553.1155.7157.4225.79
Mantis + SoFA49.2250.8752.3416.2360.4862.2164.6814.3049.8850.2650.795.6154.3956.3456.938.35
LLaVA-NeXT [27]49.3452.4055.4331.8954.5759.1062.2933.7551.9653.4555.7824.9255.4958.4060.2828.56
LLaVA-NeXT + SoFA53.2054.9156.0811.2058.8761.2361.9814.2053.6155.3756.499.1558.1359.8161.166.96
+ +Table 1. The evaluation on position-agnostic tasks with and without SoFA. Similar to §3.1, we perform multiple evaluations and record minimum, maximum and average accuracy. We also report prediction inconsistency between best and worst-performing evaluations. + +
MethodBLINKM-EvalMuirMIRBNLVR2MVBench
Random Guessing39.2725.3121.4919.1350.7525.91
Idefics2 [24]45.7848.4142.8446.5881.9423.37
Idefics2 + SoFA48.3150.7747.3648.8385.1525.65
InternVL2 [11]38.9550.3054.5342.6685.5629.31
InternVL2 + SoFA43.2651.1157.1446.1988.1932.77
VILA [31]49.7952.6448.7350.3678.8225.81
VILA + SoFA50.2952.1352.8051.1580.3426.40
Mantis [19]51.4156.8544.2854.5990.2641.59
Mantis + SoFA53.1959.2347.6455.6889.9844.40
LLaVA-NeXT [27]53.3450.8348.2257.1587.2854.26
LLaVA-NeXT + SoFA55.9254.5150.4360.6789.4557.71
+ +Table 2. The overall evaluation accuracy of LVLMs on various multi-image benchmarks with and without SoFA. + +which both exhibit an increase of approximately $3\%$ . This suggests that the model retains strong performance despite variations in the distribution of image positions, indicating that SoFA effectively addresses the deficiencies of LVLMs in positions where performance is typically suboptimal. + +# 5.2. Overall Results + +Benchmark. Given the observation that SoFA effectively mitigates position bias, we proceed with a comprehensive evaluation to determine whether SoFA can enhance the overall reasoning capabilities of LVLMs. Specifically, we include all tasks from the selected benchmarks [15, 19, 56, 66]. In addition, we incorporate NLVR2 [50] and MVBench [29], which exclusively assess the models' capability in image-pair comparison and video understanding, respectively. We directly utilize the default image positions from the benchmarks and record their accuracy. + +SoFA enhances reasoning capability of LVLMs. In Table 2, we observe that integrating SoFA leads to an average improvement of $2 \sim 3\%$ in the performance of LVLMs across various benchmarks. For instance, Idefics2 shows a $2.25\%$ improvement on MIRB (46.58% → 48.83%), + +while InternVL2 exhibits a $4.31\%$ increase on BLINK $(38.95\% \rightarrow 43.26\%)$ . This demonstrates that, despite the advancements in multi-image LVLMs, position bias poses a significant obstacle to their reasoning capability, whereas SoFA successfully addresses this limitation. + +# 5.3. Discussion + +Attention distribution visualization. To gain a deeper understanding of SoFA's working mechanism, we visualize the attention distribution within LVLMs across different positions. Using PQA as an example, we derive the average attention value of the final layer's query token (whose hidden state is directly used for generation) for each image. As shown in Fig. 6, under the default setting, the attention distribution is dominated by the end positions, indicating that the model overly focuses on a few images while neglecting most others. This observation aligns with our findings in §3.2. The application of SoFA alters this trend by significantly smoothing the attention distribution, enabling the model to attend to images at all positions more evenly. This well explains how SoFA mitigates position bias. + +SoFA improves in-context capability of LVLMs. Multimodal in-context learning is an emerging property recently acquired by LVLMs, wherein a few demonstrations allow them to recognize task formats, thereby significantly improving accuracy. However, recent studies [6, 9, 30] find that LVLMs tend to overlook demonstration images that appear earlier, leading to suboptimal performance. Here, we demonstrate that SoFA can improve the in-context learning capability of LVLMs by eliminating such blind spots. Following [30], We select three typical visual question answering benchmarks: VQAv2 [16], VizWiz [17], and OKVQA [38], and report the average accuracy of LVLMs across varying shot counts. As illustrated in Table 3, the application of SoFA leads to a further enhancement in LVLM performance, with this effect becoming increasingly pronounced as the number of shots rises. For instance, in the + +![](images/a7f2cbbd62699c6b45a7bef744223616aa47aba15410d847c57b1da62ea8c5ae.jpg) +Figure 6. The attention distribution across positions on PQA. + +
MethodVQAv2 [16]VizWiz [17]OK-VQA [38]
4-shot8-shot16-shot4-shot8-shot16-shot4-shot8-shot16-shot
w/o SoFA55.3459.3261.3741.5644.6545.3545.1948.3049.02
w/ SoFA56.1361.3064.8542.3046.9849.1745.9349.8752.36
+ +Table 3. The average accuracy of LVLMs on visual question answering using in-context learning with and without SoFA. + +
# Images351020305075100
w/o SoFA82.1681.2777.1872.3668.5060.8055.1349.19
w/ SoFA83.9082.3378.9274.3070.6763.2759.3855.11
+ +Table 4. The average accuracy of LVLMs on Multi-modal Haystack [59], where images range from a few to a hundred. + +case of VizWiz, accuracy improves from $41.56\%$ to $42.30\%$ with 4 shots, and from $45.35\%$ to $49.17\%$ with 16 shots. This indicates that position bias undermines the effectiveness of demonstrations, while SoFA mitigates this issue. + +SoFA exhibits advantages in long-context applications. Long-context scenarios have recently become a focus, where a single inference requires processing a substantial number of images. This is particularly beneficial for applications such as many-shot in-context learning [20] and multi-modal RAG [62]. Here, we show that SoFA can excel in these challenging scenarios. We select Visual Haystack [61], a benchmark designed to assess the long-context capabilities of LVLMs, and increase the input scale to encompass anywhere from a few images to a hundred. As shown in Table 4, the benefits of SoFA become increasingly evident as the number of images rises, achieving peak performance gains at 100 images (49.19% → 55.11%). This indicates that an increase in context length exacerbates the model's position bias, while SoFA compensates this loss. + +Contribution of SoFA to different tasks. Extending beyond the holistic evaluation, we categorize the benchmarks into various tasks to identify which ones benefit the most from SoFA. As depicted in Fig. 7, SoFA demonstrates the highest improvements in visual retrieval and analogy, with gains of $6.84\%$ and $5.53\%$ , respectively. In these tasks, the first image acts as a reference, and the models need to select a matching scene or similar style from the remaining images. This necessitates that the model pays extra attention + +![](images/9e6db53c170794c570aee95c360d8bb9f82bd3d31e2c300b24d7a531cbfb23f0.jpg) + +Figure 7. The performance gains of SoFA on different types of tasks. The results are averaged over selected models. + +
σBLINKMantis-EvalMuirBenchMIRBNLVR2MVBench
0.0047.8551.8147.7250.2784.7734.87
0.2549.2252.7049.3552.3485.4137.28
0.5050.0453.3350.9651.8686.3036.56
0.7549.6253.7448.2451.1486.8335.07
1.0048.2151.5847.3750.5385.3734.20
+ +Table 5. The results on six benchmarks varying the value of $\sigma$ . + +to and adequately understands the first image. Intuitively, due to the recency bias inherent in open-source models, they often overlook earlier information, resulting in suboptimal performance. Thus, the significant gains brought by SoFA are understandable. + +The interpretation of $\sigma$ . In Table 5, we conduct an ablation study on the effect of $\sigma$ . It can be observed that the optimal $\sigma$ varies significantly across benchmarks, for instance, with a value of 0.75 on NLVR2 and only 0.25 on MVBench. Intuitively, MVBench inputs involve video frames, which have a strong dependence on image positions, and an excessively high $\sigma$ may compromise such positional information. In contrast, the examples in NLVR2 are image pairs, most of which are position-agnostic. This difference in task nature accounts for the variation in $\sigma$ . + +# 6. Conclusion + +In this study, we introduce the concept of position bias, a well-known issue in NLP, into the realm of multi-modal applications. First, we reveal that LVLMs cannot robustly make use of information across multiple images, where merely altering the position of images may significantly change their predictions. Then we further identify this position bias, concluding that open-source models exhibit a strong recency bias, while proprietary models struggle with positions in the middle. Motivated by these findings, we propose SoFA, a simple, training-free approach that reduces the model's dependence on positional information through linear interpolation between inter-image causal attention and bidirectional counterparts. Experiments demonstrate that SoFA effectively mitigates position bias and significantly enhances the overall reasoning performance of existing models. + +# Acknowledgement + +This research was, in part, funded by the U.S. Government - DARPA TIAMAT HR00112490421. The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the U.S. Government. + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1, 2, 3, 6 +[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob L. Menick, Sebastian Borgeaud, Andy Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karén Simonyan. Flamingo: a visual language model for few-shot learning. In NeurIPS, 2022. 1, 2 +[3] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In CVPR, pages 2425-2433, 2015. 1 +[4] Anas Awadalla, Irena Gao, Josh Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Shiori Sagawa, et al. Openflamingo: An opensource framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390, 2023. 1, 2, 3 +[5] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 1, 2 +[6] Folco Bertini Baldassini, Mustafa Shukor, Matthieu Cord, Laure Soulier, and Benjamin Piwowarski. What makes multimodal in-context learning work? In CVPR, pages 1539-1550, 2024. 3, 7 +[7] Rohan Bavishi, Erich Elsen, Curtis Hawthorne, Maxwell Nye, Augustus Odena, Arushi Somani, and Săgnak Taşirlar. Introducing our multimodal models, 2023. 2 +[8] Liangyu Chen, Bo Li, Sheng Shen, Jingkang Yang, Chunyuan Li, Kurt Keutzer, Trevor Darrell, and Ziwei Liu. Large language models are visual reasoning coordinators. In NeurIPS, 2024. 1 +[9] Shuo Chen, Zhen Han, Bailan He, Mark Buckley, Philip Torr, Volker Tresp, and Jindong Gu. Understanding and improving in-context learning on vision-language models. In ICLR, 2023. 3, 7 +[10] Yuhan Chen, Ang Lv, Ting-En Lin, Changyu Chen, Yuchuan Wu, Fei Huang, Yongbin Li, and Rui Yan. Fortify the short + +est stave in attention: Enhancing context awareness of large language models for effective tool use. In ACL, 2023. 5 +[11] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In CVPR, pages 24185–24198, 2024. 1, 2, 6, 7 +[12] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with $90\%$ chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2(3):6, 2023. 2 +[13] Neil Cohn, Ryan Taylor, and Kaitlin Pederson. A picture is worth more words over time: Multimodality and narrative structure across eight decades of american superhero comics. Multimodal Communication, 6(1):19-37, 2017. 1 +[14] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. In NeurIPS, pages 16344-16359, 2022. 6 +[15] Xingyu Fu, Yushi Hu, Bangzheng Li, Yu Feng, Haoyu Wang, Xudong Lin, Dan Roth, Noah A Smith, Wei-Chiu Ma, and Ranjay Krishna. Blink: Multimodal large language models can see but not perceive. In ECCV, 2024. 1, 2, 3, 6, 7 +[16] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In CVPR, pages 6904-6913, 2017. 1, 4, 7, 8 +[17] Danna Gurari, Qing Li, Abigale J Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo Luo, and Jeffrey P Bigham. Vizwiz grand challenge: Answering visual questions from blind people. In CVPR, pages 3608-3617, 2018. 7, 8 +[18] Junqing He, Kunhao Pan, Xiaoqun Dong, Zhuoyang Song, LiuYiBo LiuYiBo, Qiangosun Qianguosun, Yuxin Liang, Hao Wang, Enming Zhang, and Jiaxing Zhang. Never lost in the middle: Mastering long-context question answering with position-agnostic decompositional training. In ACL, pages 13628–13642, 2024. 5 +[19] Dongfu Jiang, Xuan He, Huaye Zeng, Cong Wei, Max Ku, Qian Liu, and Wenhu Chen. Mantis: Interleaved multi-image instruction tuning. arXiv preprint arXiv:2405.01483, 2024. 1, 2, 3, 6, 7 +[20] Yixing Jiang, Jeremy Irvin, Ji Hun Wang, Muhammad Ahmed Chaudhry, Jonathan H Chen, and Andrew Y Ng. Many-shot in-context learning in multimodal foundation models. In ICML, 2024. 8 +[21] Mehran Kazemi, Nishanth Dikkala, Ankit Anand, Petar Devic, Ishita Dasgupta, Fangyu Liu, Bahare Fatemi, Pranjal Awasthi, Dee Guo, Sreenivas Gollapudi, et al. Remi: A dataset for reasoning with multiple images. arXiv preprint arXiv:2406.09175, 2024. 2 +[22] Hugo Laurençon, Andrés Marafioti, Victor Sanh, and Léo Tronchon. Building and better understanding vision-language models: insights and future directions. arXiv preprint arXiv:2408.12637, 2024. 1 + +[23] Hugo Laurençon, Lucile Saulnier, Léo Tronchon, Stas Bekman, Amanpreet Singh, Anton Lozhkov, Thomas Wang, Siddharth Karamcheti, Alexander Rush, Douwe Kiela, et al. Obelics: An open web-scale filtered dataset of interleaved image-text documents. In NeurIPS, 2024. 2 +[24] Hugo Laurençon, Léo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language models? arXiv preprint arXiv:2405.02246, 2024. 1, 2, 3, 6, 7 +[25] Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Kuttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al. Retrieval-augmented generation for knowledge-intensive nlp tasks. In CVPR, pages 9459-9474, 2020. 1 +[26] Bohao Li, Yuying Ge, Yixiao Ge, Guangzhi Wang, Rui Wang, Ruimao Zhang, and Ying Shan. Seed-bench: Benchmarking multimodal large language models. In CVPR, pages 13299-13308, 2024. 1 +[27] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava next-interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895, 2024. 2, 3, 6, 7 +[28] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, pages 19730–19742. PMLR, 2023. 1 +[29] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mybench: A comprehensive multi-modal video understanding benchmark. In CVPR, pages 22195-22206, 2024. 2, 7 +[30] Li Li, Jiawei Peng, Huiyi Chen, Chongyang Gao, and Xu Yang. How to configure good in-context sequence for visual question answering. In CVPR, pages 26710-26720, 2024. 2, 7 +[31] Ji Lin, Hongxu Yin, Wei Ping, Pavlo Molchanov, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models. In CVPR, pages 26689-26699, 2024. 6, 7 +[32] Ziyi Lin, Chris Liu, Renrui Zhang, Peng Gao, Longtian Qiu, Han Xiao, Han Qiu, Chen Lin, Wenqi Shao, Keqin Chen, et al. Sphinx: The joint mixing of weights, tasks, and visual embeddings for multi-modal large language models. arXiv preprint arXiv:2311.07575, 2023. 6 +[33] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024. 2 +[34] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2024. 1 +[35] Nelson F Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. Lost in the middle: How language models use long contexts. In ACL, pages 157-173. MIT Press One Broadway, 12th Floor, Cambridge, Massachusetts 02142, USA ..., 2024. 1, 2, 3, 5 +[36] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. In NeurIPS, pages 2507–2521, 2022. 1 + +[37] Pan Lu, Baolin Peng, Hao Cheng, Michel Galley, Kai-Wei Chang, Ying Nian Wu, Song-Chun Zhu, and Jianfeng Gao. Chameleon: Plug-and-play compositional reasoning with large language models. In NeurIPS, 2024. 1 +[38] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In CVPR, pages 3195-3204, 2019. 1, 7, 8 +[39] Duy MH Nguyen, Hoang Nguyen, Nghiem Diep, Tan Ngoc Pham, Tri Cao, Binh Nguyen, Paul Swoboda, Nhat Ho, Shadi Albarqouni, Pengtao Xie, et al. Lvm-med: Learning large-scale self-supervised vision models for medical imaging via second-order graph matching. In NeurIPS, 2024. 2 +[40] Yassine Ouali, Adrian Bulat, Brais Martinez, and Georgios Tzimiropoulos. Clip-dpo: Vision-language models as a source of preference for fixing hallucinations in lvlms. In ECCV, 2024. 2 +[41] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. In ICLR, 2023. 2 +[42] Alexander Peysakhovich and Adam Lerer. Attention sorting combats recency bias in long context language models. arXiv preprint arXiv:2310.01427, 2023. 1, 3, 5 +[43] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 2 +[44] Ori Ram, Yoav Levine, Itay Dalmedigos, Dor Muhlgay, Amnon Shashua, Kevin Leyton-Brown, and Yoav Shoham. Incontext retrieval-augmented language models. In ACL, pages 1316-1331. MIT Press One Broadway, 12th Floor, Cambridge, Massachusetts 02142, USA . . ., 2023. 1 +[45] Nir Ratner, Yoav Levine, Yonatan Belinkov, Ori Ram, Inbal Magar, Omri Abend, Ehud Karpas, Amnon Shashua, Kevin Leyton-Brown, and Yoav Shoham. Parallel context windows for large language models. In ACL, 2022. 3, 5 +[46] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS, 2022. 2 +[47] Junjie Shentu and Noura Al Moubayed. Cxr-irgen: An integrated vision and language model for the generation of clinically accurate chest x-ray image-report pairs. In WACV, pages 5212–5221, 2024. 2 +[48] Lin Shi, Weicheng Ma, and Soroush Vosoughi. Judging the judges: A systematic investigation of position bias in pairwise comparative assessments by llms. arXiv preprint arXiv:2406.07791, 2024. 1, 3 +[49] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 5 + +[50] Alane Suhr, Stephanie Zhou, Ally Zhang, Iris Zhang, Huajun Bai, and Yoav Artzi. A corpus for reasoning about natural language grounded in photographs. In ACL, 2018. 1, 7 +[51] Quan Sun, Yufeng Cui, Xiaosong Zhang, Fan Zhang, Qiying Yu, Yueze Wang, Yongming Rao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative multimodal models are in-context learners. In CVPR, pages 14398-14409, 2024. 1 +[52] Dídac Surís, Sachit Menon, and Carl Vondrick. Vipergpt: Visual inference via python execution for reasoning. In CVPR, pages 11888-11898, 2023. 1 +[53] MosaicML NLP Team et al. Introducing mpt-7b: A new standard for open-source, commercially usable llms, 2023. 2 +[54] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2 +[55] Pat Verga, Sebastian Hofstatter, Sophia Althammer, Yixuan Su, Aleksandra Piktus, Arkady Arkhangorodsky, Minjie Xu, Naomi White, and Patrick Lewis. Replacing judges with juries: Evaluating llm generations with a panel of diverse models. arXiv preprint arXiv:2404.18796, 2024. 1, 3 +[56] Fei Wang, Xingyu Fu, James Y Huang, Zekun Li, Qin Liu, Xiaogeng Liu, Mingyu Derek Ma, Nan Xu, Wenxuan Zhou, Kai Zhang, et al. Muirbench: A comprehensive benchmark for robust multi-image understanding. arXiv preprint arXiv:2406.09411, 2024. 1, 2, 3, 6, 7 +[57] Hengyi Wang, Haizhou Shi, Shiwei Tan, Weiyi Qin, Wenyuan Wang, Tunyu Zhang, Akshay Nambi, Tanuja Ganu, and Hao Wang. Multimodal needle in a haystack: Benchmarking long-context capability of multimodal large language models. arXiv preprint arXiv:2406.11230, 2024. 3 +[58] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, et al. Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079, 2023. 2 +[59] Weiyun Wang, Shuibao Zhang, Yiming Ren, Yuchen Duan, Tiantong Li, Shuo Liu, Mengkang Hu, Zhe Chen, Kaipeng Zhang, Lewei Lu, et al. Needle in a multimodal haystack. In NeurIPS, 2024. 3, 8 +[60] Ziqi Wang, Hanlin Zhang, Xiner Li, Kuan-Hao Huang, Chi Han, Shuiwang Ji, Sham M Kakade, Hao Peng, and Heng Ji. Eliminating position bias of language models: A mechanistic approach. arXiv preprint arXiv:2407.01100, 2024. 1, 3, 5 +[61] Tsung-Han Wu, Giscard Biamby, Jerome Quenum, Ritwik Gupta, Joseph E Gonzalez, Trevor Darrell, and David M Chan. Visual haystacks: Answering harder questions about sets of images. arXiv preprint arXiv:2407.13766, 2024. 3, 8 +[62] Michihiro Yasunaga, Armen Aghajanyan, Weijia Shi, Rich James, Jure Leskovec, Percy Liang, Mike Lewis, Luke Zettlemoyer, and Wen-tau Yih. Retrieval-augmented multimodal language modeling. In ICML, 2022. 8 +[63] Yijiong Yu, Huiqiang Jiang, Xufang Luo, Qianhui Wu, ChinYew Lin, Dongsheng Li, Yuqing Yang, Yongfeng Huang, and Lili Qiu. Mitigate position bias in large language models via scaling a single dimension. arXiv preprint arXiv:2406.02536, 2024. 3 + +[64] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In CVPR, pages 9556-9567, 2024. 1 +[65] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In CVPR, pages 11975-11986, 2023. 2 +[66] Bingchen Zhao, Yongshuo Zong, Letian Zhang, and Timothy Hospedales. Benchmarking multi-image understanding in vision and language models: Perception, knowledge, reasoning, and multi-hop reasoning. arXiv preprint arXiv:2406.12742, 2024. 3, 6, 7 +[67] Haozhe Zhao, Zefan Cai, Shuzheng Si, Xiaojian Ma, Kaikai An, Liang Chen, Zixuan Liu, Sheng Wang, Wenjuan Han, and Baobao Chang. Mmcl: Empowering vision-language model with multi-modal in-context learning. In ICLR, 2023. 2 +[68] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. In NeurlPS, pages 46595-46623, 2023. 1, 3 +[69] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. In ICLR, 2023. 2 +[70] Wanrong Zhu, Jack Hessel, Anas Awadalla, Samir Yitzhak Gadre, Jesse Dodge, Alex Fang, Youngjae Yu, Ludwig Schmidt, William Yang Wang, and Yejin Choi. Multimodal c4: An open, billion-scale corpus of images interleaved with text. In NeurIPS, 2024. 2 +[71] Yuchen Zhuang, Yue Yu, Kuan Wang, Haotian Sun, and Chao Zhang. Toolqa: A dataset for llm question answering with external tools. In NeurIPS, pages 50117-50143, 2023. 1 \ No newline at end of file diff --git a/2025/Identifying and Mitigating Position Bias of Multi-image Vision-Language Models/images.zip b/2025/Identifying and Mitigating Position Bias of Multi-image Vision-Language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e23d24e510bd52a749d785db3201cd83a13a6b7e --- /dev/null +++ b/2025/Identifying and Mitigating Position Bias of Multi-image Vision-Language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c0fe975d0a8a8360ce9cf302e20646868d8c0526b57015b0064daa22d5eb3929 +size 563219 diff --git a/2025/Identifying and Mitigating Position Bias of Multi-image Vision-Language Models/layout.json b/2025/Identifying and Mitigating Position Bias of Multi-image Vision-Language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9a52160cd3b51335c26badc4a650851385a443b1 --- /dev/null +++ b/2025/Identifying and Mitigating Position Bias of Multi-image Vision-Language Models/layout.json @@ -0,0 +1,8768 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 56, + 103, + 555, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 103, + 555, + 122 + ], + "spans": [ + { + "bbox": [ + 56, + 103, + 555, + 122 + ], + "type": "text", + "content": "Identifying and Mitigating Position Bias of Multi-image Vision-Language Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 141, + 142, + 468, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 142, + 468, + 171 + ], + "spans": [ + { + "bbox": [ + 141, + 142, + 468, + 171 + ], + "type": "text", + "content": "Xinyu Tian1 Shu Zou1 Zhaoyuan Yang2 Jing Zhang1 \n1Australian National University 2GE Research" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 147, + 172, + 459, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 172, + 459, + 185 + ], + "spans": [ + { + "bbox": [ + 147, + 172, + 459, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 147, + 172, + 459, + 185 + ], + "type": "text", + "content": "firstname_lastname@anu.edu.au, " + }, + { + "bbox": [ + 147, + 172, + 459, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 147, + 172, + 459, + 185 + ], + "type": "text", + "content": "firstname_lastname@ge.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 213, + 201, + 227 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 213, + 201, + 227 + ], + "spans": [ + { + "bbox": [ + 151, + 213, + 201, + 227 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 238, + 297, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 238, + 297, + 490 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 297, + 490 + ], + "type": "text", + "content": "The evolution of Large Vision-Language Models (LVLMs) has progressed from single to multi-image reasoning. Despite this advancement, our findings indicate that LVLMs struggle to robustly utilize information across multiple images, with predictions significantly affected by the alteration of image positions. To further explore this issue, we introduce Position-wise Question Answering (PQA), a meticulously designed task to quantify reasoning capabilities at each position. Our analysis reveals a pronounced position bias in LVLMs: open-source models excel in reasoning with images positioned later but underperform with those in the middle or at the beginning, while proprietary models show improved comprehension for images at the beginning and end but struggle with those in the middle. Motivated by this, we propose SoFt Attention (SoFA), a simple, training-free approach that mitigates this bias by employing linear interpolation between inter-image causal attention and bidirectional counterparts. Experimental results demonstrate that SoFA reduces position bias and enhances the reasoning performance of existing LVLMs. The code will be available https://github.com/xytian1008/sofa." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 513, + 137, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 513, + 137, + 526 + ], + "spans": [ + { + "bbox": [ + 55, + 513, + 137, + 526 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 533, + 295, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 701 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 701 + ], + "type": "text", + "content": "The advent of Large Vision-Language Models (LVLMs) [1, 2, 5, 24, 34] has catalyzed the development of various text-grounded visual applications, including visual question answering [3, 16, 36], code generation [8, 37, 52], and story telling [13, 26]. Open-source LVLMs such as LLaVA [34], BLIP2 [28], and InternVL [11] have demonstrated competitive performance on single-image tasks like VQAv2 [16], OKVQA [38], and MMMU [64]. However, these LVLMs are limited in their capacity to analyze interimage differences and relationships, as well as to process temporal information. Consequently, recent LVLMs, such as Flamingo [2, 4], Idefics [22-24], and Emu2 [51], have incorporated image-text interleaved data during training to enable ability to tackle multi-image tasks [15, 19, 50, 56]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 702, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 295, + 714 + ], + "type": "text", + "content": "Despite the increasing scale and performance of LVLMs," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 214, + 555, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 214, + 555, + 382 + ], + "spans": [ + { + "bbox": [ + 313, + 214, + 555, + 382 + ], + "type": "text", + "content": "the interpretation of how they reason across multiple images remains a mystery. In fact, we might find some hints in the topic of natural language processing (NLP). Recent research in NLP has highlighted retrieval-augmented generation (RAG) [25, 44, 71], which aims to enable large language models (LLMs) to reason across multiple documents to answer questions. However, [35] find that LLMs tend to effectively comprehend documents positioned at the beginning and end of the input, often neglecting the middle sections. Similarly, in LLM-as-a-judge [48, 55], [68] find that LLMs used to evaluate two model responses tend to favor the first one, regardless of content. This phenomenon, where the model's reasoning capability is influenced by the segment position, is referred to as position bias [35, 42, 60]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 386, + 556, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 386, + 556, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 386, + 556, + 567 + ], + "type": "text", + "content": "Motivated by the research above, a natural question arises: does position bias exist in LVLMs? Specifically, we seek to determine whether LVLMs can robustly make use of information across multiple images. To investigate this, we begin by evaluating LVLMs on established multi-image benchmarks [15, 19, 50, 56]. We find that simply changing the order of presented images may significantly alter the predictions for most LVLMs, leading to considerable variation in benchmark accuracy. Fig. 1 provides examples of questions and model responses. For instance, in difference spotting [56], swapping two images causes OpenFlamingo [4] to miss distinctions, while in a simple counting task [19], reordering images leads Idefics2 [24] to incorrect reasoning. This positional sensitivity is apparent even in GPT-4o [1], a leading proprietary model." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 570, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 557, + 715 + ], + "type": "text", + "content": "The above motivational study reveals an unfortunate fact: building upon LLMs, LVLMs may inherit the notorious position bias from their language component. This bias manifests as the model's tendency to excel at reasoning visual information from certain positions, while exhibiting weaknesses at others, compromising robustness and reliability in predictions. To further investigate which positions display strong understanding and which reveal weaknesses, i.e., identifying the pattern of such position bias, we introduce Position-wise Question Answering (PQA). Unlike most multi-image tasks that assess holistic comprehension, PQA requires LVLMs to reason consistently across each" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10599" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 69, + 555, + 223 + ], + "blocks": [ + { + "bbox": [ + 57, + 69, + 555, + 223 + ], + "lines": [ + { + "bbox": [ + 57, + 69, + 555, + 223 + ], + "spans": [ + { + "bbox": [ + 57, + 69, + 555, + 223 + ], + "type": "image", + "image_path": "544f5f38c0745927eb794d5dd9747f98e1aa05cc89c160727c5081df15da5c58.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 232, + 555, + 266 + ], + "lines": [ + { + "bbox": [ + 55, + 232, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 555, + 266 + ], + "type": "text", + "content": "Figure 1. Examples where the predictions of LVLMs are influenced by changes in the image positions (Date accessed: Oct 25, 2024). The tasks, from left to right, are difference spotting [56], counting [19] and forensic detection [15], respectively. The correct answers are highlighted in green, while the incorrect ones are highlighted in red. The prompts and outputs are simplified here for clarity and aesthetics." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 275, + 295, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 275, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 275, + 295, + 312 + ], + "type": "text", + "content": "image and produce position-wise responses. This approach enables us to measure position-wise accuracy and analyze the model's reasoning capacity across different positions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 315, + 296, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 315, + 296, + 447 + ], + "spans": [ + { + "bbox": [ + 55, + 315, + 296, + 447 + ], + "type": "text", + "content": "Our analysis reveals that most open-source models, e.g., OpenFlamingo [4] and Idefics2 [24], exhibit a pronounced recency bias, demonstrating stronger reasoning capabilities for images presented later while underperforming on those in the middle or at the beginning. In contrast, proprietary models like GPT-4o [1] show enhanced understanding of images at the beginning and end, while neglecting those in the middle, a pattern consistent with previous findings in LLMs [35]. This observation is further corroborated by Fig. 1, where LVLMs tend to reason incorrectly about images positioned in poor-performing areas." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 451, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 451, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 451, + 296, + 714 + ], + "type": "text", + "content": "The existence of position bias significantly compromises model robustness and undermines accuracy across various multi-image tasks. Inspired by this, we aim to investigate the underlying mechanical causes and mitigate this issue. Specifically, we notice that in most LVLMs, the hidden states of image tokens are heavily influenced by their relative positions due to inter-image causal attention. In particular, image tokens are treated as a type of special text tokens where 1) images appearing later may interact with preceding ones; 2) images at the beginning are isolated, lacking access to global context. This autoregressive setup implicitly imposes a semantic order on the presented images, amplifying their positional information. To address this, we introduce SoFt Attention (SoFA), a training-free approach to mitigate position bias across various LVLMs. SoFA reduces the dependency of image hidden states on positional information by simply performing linear interpolation between inter-image causal attention and its bidirectional counterparts. Experimental results indicate that SoFA effectively equilibrates the reasoning capabilities of LVLMs across various positions, achieving average accuracy improvements of " + }, + { + "bbox": [ + 55, + 451, + 296, + 714 + ], + "type": "inline_equation", + "content": "2 \\sim 3\\%" + }, + { + "bbox": [ + 55, + 451, + 296, + 714 + ], + "type": "text", + "content": " on existing multi-image benchmarks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 275, + 498, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 275, + 498, + 286 + ], + "spans": [ + { + "bbox": [ + 313, + 275, + 498, + 286 + ], + "type": "text", + "content": "In summary, our contributions are as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 287, + 553, + 442 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 287, + 553, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 287, + 553, + 346 + ], + "spans": [ + { + "bbox": [ + 314, + 287, + 553, + 346 + ], + "type": "text", + "content": "- Despite the significant potential of LVLMs in multi-image applications, we find that position bias, a well-explored phenomenon in NLP, has emerged as a potential bottleneck in the development of LVLMs, substantially impacting model robustness and reliability." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 346, + 553, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 346, + 553, + 394 + ], + "spans": [ + { + "bbox": [ + 314, + 346, + 553, + 394 + ], + "type": "text", + "content": "- We further identify such position bias and discover that most open-source models demonstrate a pronounced recency bias, whereas proprietary models tend to struggle in middle positions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 395, + 553, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 395, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 314, + 395, + 553, + 442 + ], + "type": "text", + "content": "- We introduce SoFA, a straightforward training-free approach designed to mitigate the position bias inherited from pre-trained LVLMs, and substantiate its effectiveness through experimental validation." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 453, + 400, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 453, + 400, + 466 + ], + "spans": [ + { + "bbox": [ + 314, + 453, + 400, + 466 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 474, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 555, + 714 + ], + "type": "text", + "content": "Large Vision-Language Models. Existing LVLMs are generally composed of a vision encoder [43, 65], an LLM [12, 53, 54], and a bridging projector [2, 7, 27] that facilitates interaction between the two. Conventional LVLMs, pre-trained on image-text pairs [46] and fine-tuned through preference [33, 40] or instruction tuning [27, 41], demonstrate substantial success in applications like dialogue [5, 11], question answering [58, 69], and domain-specific tasks [39, 47]. However, constrained by their training data, these models primarily support single-image reasoning during inference. Recently, the integration of image-text interleaved data [23, 70] has enabled multi-image reasoning capabilities, encompassing image comparison [21, 56], association [15, 19], temporal relationship analysis [29], and emerging properties of multi-modal in-context learning [30, 67]. Nonetheless, limited research has delved into how LVLMs perform reasoning across multiple images or the underlying mechanisms driving this process. Position Bias in LLMs. Position bias, an inherent side effect of long-context input, has increasingly captured the at" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10600" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 553, + 170 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 170 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 170 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 170 + ], + "type": "image", + "image_path": "181e6ae6c93931b3eeb9d154c5f1e22c5557bb7f10789500efd59f45f2ce356e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 178, + 555, + 200 + ], + "lines": [ + { + "bbox": [ + 55, + 178, + 555, + 200 + ], + "spans": [ + { + "bbox": [ + 55, + 178, + 555, + 200 + ], + "type": "text", + "content": "Figure 2. The results of multiple evaluations while shuffling image positions. We record the minimum, maximum, and average accuracy (left vertical axis), along with the prediction inconsistency between the best and worst-performing evaluations (right vertical axis)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 209, + 297, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 209, + 297, + 376 + ], + "spans": [ + { + "bbox": [ + 54, + 209, + 297, + 376 + ], + "type": "text", + "content": "tention of NLP researchers. The pioneering work on this issue, notably [35], reveals that when LLMs process multiple documents, they often concentrate on those at the beginning and end while disregarding much of the information in between, i.e., a phenomenon termed \"lost in the middle\". A related issue has also been identified in LLM-as-a-judge [48, 55], where [68] observe that LLMs tasked with evaluating multiple model responses tend to favor the first one, irrespective of its content. Recent studies have begun to explore methods to mitigate or circumvent position bias in LLMs [42, 45, 60, 63]. However, to the best of our knowledge, few studies have examined the effects of position bias within multi-modal applications, specifically regarding its implications for LVLMs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 379, + 296, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 379, + 296, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 379, + 296, + 559 + ], + "type": "text", + "content": "Position Bias in LVLMs. While systematic research on position bias in LVLMs is still lacking, recent studies have indirectly underscored the influence of positions on LVLMs through various findings. For instance, in multi-modal in-context learning, studies [9] and [6] reveal that demonstration images placed at the beginning contribute minimally to LVLM performance. Additionally, in assessing long-context capabilities, [61] reports that LVLMs attain particularly high accuracy when key information appears at the end of the input sequence, with similar results observed in [59] and [57]. However, previous research has neither thoroughly analyzed this position bias nor proposed solutions to address it. In contrast, this study systematically investigates position bias in both open-source and proprietary models and introduces an effective mitigation strategy." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 572, + 272, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 572, + 272, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 572, + 272, + 586 + ], + "type": "text", + "content": "3. Understanding Position Bias in LVLMs" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 594, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 712 + ], + "type": "text", + "content": "In this section, we investigate whether the position bias observed in LLMs extends to LVLMs. We start by empirically confirming the existence of position bias in LVLMs through a motivational study, revealing their poor robustness to image position changes (§3.1). Subsequently, we analyze the nature of this bias by evaluating and quantifying the reasoning capabilities of LVLMs regarding different positions (§3.2). These insights contribute to a deeper understanding and interpretation of LVLMs' reasoning behavior and common failure cases." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 209, + 436, + 221 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 209, + 436, + 221 + ], + "spans": [ + { + "bbox": [ + 313, + 209, + 436, + 221 + ], + "type": "text", + "content": "3.1. The Position Matters" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 227, + 553, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 227, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 227, + 553, + 251 + ], + "type": "text", + "content": "To assess the existence of position bias, we conduct a simple motivational experiment, as shown in Fig. 2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 251, + 554, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 554, + 395 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 554, + 395 + ], + "type": "text", + "content": "Experimental Setup. We select four representative multi-image LVLMs: OpenFlamingo [4], Idefics2 [24], LLaVA-NeXT-Interleave1 [27], GPT-4o [1], and conduct evaluations across established multi-image benchmarks, including BLINK [15], Mantis-Eval [19], MuirBench [56], and MIRB [66]. To ensure clarity, we select partial tasks from these benchmarks that feature position-agnostic questions, wherein the positions of the images do not affect the answers to the questions. See §5.1 for the selected tasks and details. To explore the influence of image positions on the results, we perform multiple evaluations per benchmark, with the positions of the images shuffled in each iteration." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 396, + 553, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 553, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 553, + 479 + ], + "type": "text", + "content": "Metric. We record the minimum, maximum, and average accuracy across these evaluations, which directly reflects the influence of positions on the performance of LVLMs. Furthermore, we assess the prediction inconsistency between the best and worst-performing evaluations, quantifying the proportion of examples exhibiting conflicting predictions across the two evaluations by changing positions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 479, + 554, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 554, + 646 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 554, + 646 + ], + "type": "text", + "content": "Results. As illustrated in Fig. 2, for most open-source LVLMs, the prediction inconsistency between the best and worst-performing evaluations approaches " + }, + { + "bbox": [ + 313, + 479, + 554, + 646 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 313, + 479, + 554, + 646 + ], + "type": "text", + "content": ", suggesting that merely altering the positions of the presented images may result in an amount of the model's predictions being changed. Such pronounced sensitivity to positional information results in highly unstable performance. For example, OpenFlamingo's accuracy on BLINK can vary significantly, reaching " + }, + { + "bbox": [ + 313, + 479, + 554, + 646 + ], + "type": "inline_equation", + "content": "38\\%" + }, + { + "bbox": [ + 313, + 479, + 554, + 646 + ], + "type": "text", + "content": " at its peak and dropping to " + }, + { + "bbox": [ + 313, + 479, + 554, + 646 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 313, + 479, + 554, + 646 + ], + "type": "text", + "content": " at its lowest, indicating an accuracy span of up to " + }, + { + "bbox": [ + 313, + 479, + 554, + 646 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 479, + 554, + 646 + ], + "type": "text", + "content": " with position changes. GPT-4o, the state-of-the-art proprietary model, although showcasing superior performance, also displays a lack of robustness to positional variations, leading to a prediction inconsistency of around " + }, + { + "bbox": [ + 313, + 479, + 554, + 646 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 313, + 479, + 554, + 646 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 647, + 553, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 553, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 553, + 696 + ], + "type": "text", + "content": "To qualitatively explore this issue, we present examples in Fig. 1 that illustrate prediction conflicts arising from alterations in image positions. For instance, when prompted to identify the differences between two images, Open-" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 325, + 702, + 545, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 702, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 325, + 702, + 545, + 712 + ], + "type": "text", + "content": "1For brevity, we refer to LLaVA-NeXT-Interleave as LLaVA-NeXT." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10601" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 186, + 169 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 186, + 169 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 186, + 169 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 186, + 169 + ], + "type": "image", + "image_path": "77d93e66f4813cf90930cad4924e06f53ba8952058b77d12b00d011c9d6544a7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 190, + 69, + 309, + 169 + ], + "blocks": [ + { + "bbox": [ + 190, + 69, + 309, + 169 + ], + "lines": [ + { + "bbox": [ + 190, + 69, + 309, + 169 + ], + "spans": [ + { + "bbox": [ + 190, + 69, + 309, + 169 + ], + "type": "image", + "image_path": "39eddef4ed52608f0464f7bf9d90a1598601cc958984136679f48fda3834dce8.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 312, + 70, + 431, + 169 + ], + "blocks": [ + { + "bbox": [ + 312, + 70, + 431, + 169 + ], + "lines": [ + { + "bbox": [ + 312, + 70, + 431, + 169 + ], + "spans": [ + { + "bbox": [ + 312, + 70, + 431, + 169 + ], + "type": "image", + "image_path": "a13d5253052a0753a3be84751bde8a204485818747cf8ba13a2a370588e24dc4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 434, + 70, + 553, + 169 + ], + "blocks": [ + { + "bbox": [ + 434, + 70, + 553, + 169 + ], + "lines": [ + { + "bbox": [ + 434, + 70, + 553, + 169 + ], + "spans": [ + { + "bbox": [ + 434, + 70, + 553, + 169 + ], + "type": "image", + "image_path": "39f202d2bf24294a94d1e561929a75953d8a6e91798094ee0c9923cc90cc7b8e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 123, + 177, + 195, + 187 + ], + "blocks": [ + { + "bbox": [ + 123, + 177, + 195, + 187 + ], + "lines": [ + { + "bbox": [ + 123, + 177, + 195, + 187 + ], + "spans": [ + { + "bbox": [ + 123, + 177, + 195, + 187 + ], + "type": "image", + "image_path": "e648dd5b0b16ce27c0de8461b5e20b4b1ae82f22893e4c098b4823dcdc0c4ea8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 195, + 555, + 219 + ], + "lines": [ + { + "bbox": [ + 55, + 195, + 555, + 219 + ], + "spans": [ + { + "bbox": [ + 55, + 195, + 555, + 219 + ], + "type": "text", + "content": "Figure 3. The results on the PQA task. We report position-wise accuracy on four scenarios where the number of images is 5, 10, 15 and 20, respectively. A higher accuracy signifies strong reasoning at the specified position, whereas lower reflects poor-performing areas." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 205, + 177, + 255, + 186 + ], + "blocks": [ + { + "bbox": [ + 205, + 177, + 255, + 186 + ], + "lines": [ + { + "bbox": [ + 205, + 177, + 255, + 186 + ], + "spans": [ + { + "bbox": [ + 205, + 177, + 255, + 186 + ], + "type": "image", + "image_path": "028a2c900e079d23a835b08d56212d9e899da15d69cda61c12a58f0ef8bd6ca5.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 268, + 177, + 337, + 186 + ], + "blocks": [ + { + "bbox": [ + 268, + 177, + 337, + 186 + ], + "lines": [ + { + "bbox": [ + 268, + 177, + 337, + 186 + ], + "spans": [ + { + "bbox": [ + 268, + 177, + 337, + 186 + ], + "type": "image", + "image_path": "85791aa4be6580e62ea44bc5f0d9c0f705e6f140e452e37f4a9544776ce24477.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 348, + 177, + 397, + 186 + ], + "blocks": [ + { + "bbox": [ + 348, + 177, + 397, + 186 + ], + "lines": [ + { + "bbox": [ + 348, + 177, + 397, + 186 + ], + "spans": [ + { + "bbox": [ + 348, + 177, + 397, + 186 + ], + "type": "image", + "image_path": "351071bd03901fddff5c3585164da708ea4d291717f9011d5b1a3971f99e3521.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 408, + 177, + 483, + 186 + ], + "blocks": [ + { + "bbox": [ + 408, + 177, + 483, + 186 + ], + "lines": [ + { + "bbox": [ + 408, + 177, + 483, + 186 + ], + "spans": [ + { + "bbox": [ + 408, + 177, + 483, + 186 + ], + "type": "image", + "image_path": "5e32f62575144608556a8130ccd261f44c99783fd4ec9db7cfef37144574e644.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 228, + 296, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 228, + 296, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 296, + 300 + ], + "type": "text", + "content": "Flamingo fails to extract fine-grained distinctions upon swapping their positions. Similarly, in the task of object counting, Idefics2 mistakenly identifies a dog as a cat due to the repositioning of images. GPT-4o demonstrates a comparable issue, where rearranging the image options alters both its answer and the accompanying rationale." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 307, + 194, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 307, + 194, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 307, + 194, + 319 + ], + "type": "text", + "content": "3.2. Identifying Position Bias" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 324, + 296, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 324, + 296, + 372 + ], + "spans": [ + { + "bbox": [ + 55, + 324, + 296, + 372 + ], + "type": "text", + "content": "The quantitative results and qualitative examples above imply biased reasoning across positions from LVLMs. To have a closer look at such position bias, we aim to identify models' reasoning capabilities at different image positions." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 372, + 296, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 372, + 296, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 372, + 296, + 563 + ], + "type": "text", + "content": "However, existing tasks have been restricted to evaluating models' holistic comprehension of multiple images, lacking the means to systematically analyze their fine-grained understanding of each individual position. Specifically, these tasks provide coarse ground truth answers, such as option symbols or numeric values, but do not offer individual rationales for each image, making it challenging to quantify the model's reasoning across different positions. To address this gap, we introduce Position-wise Question Answering (PQA), a carefully designed task that assesses the position-wise reasoning capabilities of LVLMs. In PQA, each question prompts LVLMs to reason independently across all images, generating a distinct response for each. This approach 1) ensures task-level neutrality with respect to any particular position and 2) enables quantification of LVLMs' reasoning performance for each position." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 563, + 295, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 295, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 295, + 588 + ], + "type": "text", + "content": "Specifically, taking object counting as an example, we employ the following query template:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 593, + 295, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 593, + 295, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 593, + 295, + 647 + ], + "type": "text", + "content": "I: Please provide one answer for each image in the form like " + }, + { + "bbox": [ + 55, + 593, + 295, + 647 + ], + "type": "inline_equation", + "content": "[x, x, x, \\ldots]" + }, + { + "bbox": [ + 55, + 593, + 295, + 647 + ], + "type": "text", + "content": ". \nQ: Among , ..., , how many cats can you find in these images? \nA: [3, 2, 0, ...]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": "Each query consists of an instruction (I) and a question " + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "inline_equation", + "content": "(\\mathbb{Q})" + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": " where the former prompts the model to generate position-wise responses (A) following a specified output format. We gather responses from LVLMs as a structured list, with each element representing a particular position. This enables us" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 228, + 553, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 553, + 263 + ], + "type": "text", + "content": "to track position-wise accuracy, where a higher accuracy reflects stronger model capability at a given position, while lower indicates areas of relative weakness." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 266, + 554, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 554, + 421 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 554, + 421 + ], + "type": "text", + "content": "Construction of position-wise examples. Since in PQA, each image is semantically independent of others, a PQA example essentially consists of several single-image instances that share the same question. This allows us to easily build position-wise examples from single-image benchmarks. Specifically, we select VQAv2 [16], a substantial single-image dataset that encompasses a diverse array of everyday topics and features both multiple-choice and open-ended questions. Initially, we identify and consolidate single-image examples with identical questions to create a question set, wherein each question corresponds to multiple images. Subsequently, we reformulate these questions into PQA format according to the above-mentioned template." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 423, + 554, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 554, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 554, + 567 + ], + "type": "text", + "content": "Experimental Setup. We utilize the same LVLMs as previously mentioned and conduct evaluations on the proposed PQA task. Additionally, we also consider random guessing, where binary choice and numeric questions are answered randomly with yes or no, or with an integer value. We examine four scenarios with varying numbers of images: 5, 10, 15, and 20, collecting 1,000 examples for each scenario. Furthermore, to maintain fairness, we augment each example by shuffling the image positions four times, resulting in a total of 5,000 examples. In contrast to prior settings that evaluate holistic performance, we provide separate accuracy metrics for each position in this analysis." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 568, + 553, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 568, + 553, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 568, + 553, + 592 + ], + "type": "text", + "content": "Results. Fig. 3 displays the results of LVLMs on the PQA task, from which we may derive the following insights:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": "1) Open-source LVLMs exhibit recency bias, whereas proprietary models struggle in middle positions. For open-source models, a stronger reasoning capability is observed for images positioned at the end, while significant underperformance is noted in the middle and beginning. For example, with 10 images, the accuracy of OpenFlamingo varies by nearly " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "9\\%" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " between the far left and far right positions " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "(51.84\\% \\rightarrow 59.71\\%)" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": ". In contrast, proprietary models, i.e., GPT-40, reveal a typical U-shaped curve, highlighting pronounced weaknesses in the middle positions." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10602" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 276 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "2) Position bias becomes more pronounced as the number of images increases. For instance, with 20 images, the accuracy gap for OpenFlamingo between the two ends reaches " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "14\\%" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "36.79\\% \\rightarrow 50.17\\%" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "). Notably, at this stage, OpenFlamingo seems to nearly lost its reasoning capability for most of the beginning images, achieving results only marginally better than random guessing (" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "34.49\\%" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 156, + 294, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 156, + 294, + 276 + ], + "spans": [ + { + "bbox": [ + 55, + 156, + 294, + 276 + ], + "type": "text", + "content": "3) Relationship to position bias in LLMs. Previous research [35] shows that LLMs often overlook information in middle positions when processing multiple documents, a phenomenon termed \"lost in the middle\". Our findings suggest this effect persists for GPT in multi-image reasoning. However, for open-source models, the primacy position advantage appears to diminish, showing a marked decaying trend from the end to beginning. Whether the position bias in LVLMs is correlated with that of their LLM backbones remains an interesting question for further exploration." + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 286, + 185, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 286, + 185, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 286, + 185, + 300 + ], + "type": "text", + "content": "4. The Proposed Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 306, + 295, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 306, + 295, + 343 + ], + "spans": [ + { + "bbox": [ + 55, + 306, + 295, + 343 + ], + "type": "text", + "content": "In this section, we first analyze the underlying causes of position bias (§4.1). Then we introduce SoFA, a straightforward, training-free approach to mitigate this issue (§4.2)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 350, + 224, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 350, + 224, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 224, + 361 + ], + "type": "text", + "content": "4.1. Position Mechanism in LVLMs" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 367, + 295, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 367, + 295, + 439 + ], + "spans": [ + { + "bbox": [ + 55, + 367, + 295, + 439 + ], + "type": "text", + "content": "To investigate the origins of position bias, we need to first determine what contributes to the positional information in LVLMs. In NLP, it is generally believed that position embeddings [45, 49] and causal attention [42, 60] are the primary mechanisms enabling the sequential structure of input. Formally," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 80, + 447, + 295, + 468 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 447, + 295, + 468 + ], + "spans": [ + { + "bbox": [ + 80, + 447, + 295, + 468 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} = \\mathcal {P} \\left(\\mathbf {X} \\mathbf {W} _ {\\mathbf {q}}, \\mathbf {p o s}\\right), \\quad \\mathbf {K} = \\mathcal {P} \\left(\\mathbf {X} \\mathbf {W} _ {\\mathbf {k}}, \\mathbf {p o s}\\right) \\tag {1}", + "image_path": "5b3132a028aae1978f6af0b998c1ab9d2cc76ef5dd056f779ca06c90c7e31a6d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 82, + 464, + 233, + 477 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 464, + 233, + 477 + ], + "spans": [ + { + "bbox": [ + 82, + 464, + 233, + 477 + ], + "type": "interline_equation", + "content": "\\mathbf {H} = \\operatorname {S o f t m a x} \\left(\\mathbf {Q K} ^ {\\mathrm {T}} / \\sqrt {d}\\right) \\odot \\mathbb {1} _ {\\text {c a u s a l}} \\mathbf {V}", + "image_path": "15cf94495dd7005330d686575df80910ced9a9ace701dbbc951d7262ff8fcf59.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 486, + 295, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 295, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 295, + 558 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 486, + 295, + 558 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 55, + 486, + 295, + 558 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 486, + 295, + 558 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 55, + 486, + 295, + 558 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 486, + 295, + 558 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 55, + 486, + 295, + 558 + ], + "type": "text", + "content": " represent the queries, keys, and values, respectively. " + }, + { + "bbox": [ + 55, + 486, + 295, + 558 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 55, + 486, + 295, + 558 + ], + "type": "text", + "content": " refers to the input tokens, encompassing both image and text types, pos denotes the position embedding of " + }, + { + "bbox": [ + 55, + 486, + 295, + 558 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 55, + 486, + 295, + 558 + ], + "type": "text", + "content": ", which is integrated into the attention via the position encoding function " + }, + { + "bbox": [ + 55, + 486, + 295, + 558 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 55, + 486, + 295, + 558 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 486, + 295, + 558 + ], + "type": "inline_equation", + "content": "\\mathbb{1}_{\\mathrm{causal}}" + }, + { + "bbox": [ + 55, + 486, + 295, + 558 + ], + "type": "text", + "content": " denotes the causal mask, which ensures autoregressive generation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": "This reveals two key factors influencing LVLMs' positional awareness: 1) position embedding explicitly initializes a sequential value to each token, and 2) causal attention implicitly enhances this positional characteristic through unidirectional information access. This presents us with two distinct solutions. One approach is to directly reassign position embeddings, a widely used strategy in NLP [10, 18, 45], as it fundamentally eliminates the influence of position on documents. However, this aggressive method is not suitable for LVLMs, since in many scenarios, such as video understanding and action recognition, we still expect images to retain positional information for temporal analysis. To clarify, our aim is to mitigate the imbalanced" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 316, + 70, + 382, + 137 + ], + "blocks": [ + { + "bbox": [ + 316, + 70, + 382, + 137 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 382, + 137 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 382, + 137 + ], + "type": "image", + "image_path": "9ee0e20336200991d8dcd43aba63d83b448ed3b7a1e38bc0a19721f392c90721.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 140, + 376, + 148 + ], + "lines": [ + { + "bbox": [ + 318, + 140, + 376, + 148 + ], + "spans": [ + { + "bbox": [ + 318, + 140, + 376, + 148 + ], + "type": "text", + "content": "(A) Causal Attention" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 399, + 70, + 466, + 137 + ], + "blocks": [ + { + "bbox": [ + 399, + 70, + 466, + 137 + ], + "lines": [ + { + "bbox": [ + 399, + 70, + 466, + 137 + ], + "spans": [ + { + "bbox": [ + 399, + 70, + 466, + 137 + ], + "type": "image", + "image_path": "954e46ff412889d6a74d516d7464db5e1041352826b2dbfc22ba8d57adbf24b9.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 402, + 140, + 462, + 148 + ], + "lines": [ + { + "bbox": [ + 402, + 140, + 462, + 148 + ], + "spans": [ + { + "bbox": [ + 402, + 140, + 462, + 148 + ], + "type": "text", + "content": "(B) Isolated Attention" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 481, + 70, + 549, + 137 + ], + "blocks": [ + { + "bbox": [ + 481, + 70, + 549, + 137 + ], + "lines": [ + { + "bbox": [ + 481, + 70, + 549, + 137 + ], + "spans": [ + { + "bbox": [ + 481, + 70, + 549, + 137 + ], + "type": "image", + "image_path": "a588def4b1cc8792d1edc54a90616a2a265609fef01b76f8ed1b2c324feaa086.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 479, + 140, + 553, + 148 + ], + "lines": [ + { + "bbox": [ + 479, + 140, + 553, + 148 + ], + "spans": [ + { + "bbox": [ + 479, + 140, + 553, + 148 + ], + "type": "text", + "content": "(C) Bidirectional Attention" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 158, + 553, + 278 + ], + "lines": [ + { + "bbox": [ + 313, + 158, + 553, + 278 + ], + "spans": [ + { + "bbox": [ + 313, + 158, + 553, + 278 + ], + "type": "text", + "content": "Figure 4. The three inter-image attention mechanisms, where " + }, + { + "bbox": [ + 313, + 158, + 553, + 278 + ], + "type": "inline_equation", + "content": "I_{1}" + }, + { + "bbox": [ + 313, + 158, + 553, + 278 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 158, + 553, + 278 + ], + "type": "inline_equation", + "content": "I_{2}" + }, + { + "bbox": [ + 313, + 158, + 553, + 278 + ], + "type": "text", + "content": " represent the tokens of two images, " + }, + { + "bbox": [ + 313, + 158, + 553, + 278 + ], + "type": "inline_equation", + "content": "T_{1}" + }, + { + "bbox": [ + 313, + 158, + 553, + 278 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 158, + 553, + 278 + ], + "type": "inline_equation", + "content": "T_{2}" + }, + { + "bbox": [ + 313, + 158, + 553, + 278 + ], + "type": "text", + "content": " represent the tokens of two text segments. Note that typically each image involves numerous tokens, e.g., 576 for LLaVA. Here for clarity we simplify them to a single token. In (A), images interact in a unidirectional manner, allowing " + }, + { + "bbox": [ + 313, + 158, + 553, + 278 + ], + "type": "inline_equation", + "content": "I_{2}" + }, + { + "bbox": [ + 313, + 158, + 553, + 278 + ], + "type": "text", + "content": " to attend to " + }, + { + "bbox": [ + 313, + 158, + 553, + 278 + ], + "type": "inline_equation", + "content": "I_{1}" + }, + { + "bbox": [ + 313, + 158, + 553, + 278 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 313, + 158, + 553, + 278 + ], + "type": "inline_equation", + "content": "I_{1}" + }, + { + "bbox": [ + 313, + 158, + 553, + 278 + ], + "type": "text", + "content": " remains isolated. In (B), each image is isolated, indicating they can only attend to themselves. In (C), bidirectional interaction is enabled so that each image can attend to any other images. It is worth mentioning that we only alter the inter-image attention while preserving causal attention between the text segments." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 289, + 553, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 553, + 312 + ], + "type": "text", + "content": "reasoning capabilities of LVLMs across different positions, rather than to eliminate their positional awareness." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 314, + 554, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 314, + 554, + 457 + ], + "spans": [ + { + "bbox": [ + 313, + 314, + 554, + 457 + ], + "type": "text", + "content": "Therefore, in this paper, we primarily investigate the relationship between causal attention and position bias. We notice that, while images do not exhibit a distinct semantic order as text does, most LVLMs, e.g., LLaVA series, treat image and text tokens in a uniform autoregressive manner. As shown in Fig. 4 (A), analogous to texts, there exists interimage causal attention which allows 1) later-positioned images to interact with those preceding them, while 2) earlier images remain isolated, lacking access to global contexts. From a mechanical perspective, this framework renders image hidden states highly dependent on their relative positions, thus implicitly amplifying their positional cues." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 459, + 553, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 459, + 553, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 459, + 553, + 483 + ], + "type": "text", + "content": "To verify our hypothesis, we compare the inter-image causal attention against following two alternative variants:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 483, + 553, + 556 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 314, + 483, + 553, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 483, + 553, + 519 + ], + "spans": [ + { + "bbox": [ + 314, + 483, + 553, + 519 + ], + "type": "text", + "content": "1) Isolated attention. As illustrated in Fig. 4 (B), we disable interactions between images, enforcing each image to attend solely to itself." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 521, + 553, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 553, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 553, + 556 + ], + "type": "text", + "content": "2) Bidirectional attention. Rather than restricting to unidirectional information flow, we permit each image to attend to any other, as shown in Fig. 4 (C)." + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 558, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 554, + 713 + ], + "type": "text", + "content": "We modify the default causal mask in the language decoder to its isolated and bidirectional counterparts and assess the impact of them on position bias in PQA, as shown in Fig. 5. It is evident that altering causal attention to either isolated or bidirectional forms significantly alleviates position bias, resulting in more balanced accuracy across different positions. This empirically confirms that inter-image causal attention is the primary source of position bias. However, these modifications lead to varying degrees of performance degradation, with isolated attention causing the most substantial decline. This outcome is reasonable, as such modifications may substantially deviate from the training framework, leading to out-of-distribution hidden states." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10603" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 177, + 175 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 177, + 175 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 177, + 175 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 177, + 175 + ], + "type": "image", + "image_path": "33dc7e2e838ab5a2d2cd8bc50703beb61152d00f69180c48ff0582bb9ab7a12a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 104, + 178, + 144, + 187 + ], + "lines": [ + { + "bbox": [ + 104, + 178, + 144, + 187 + ], + "spans": [ + { + "bbox": [ + 104, + 178, + 144, + 187 + ], + "type": "text", + "content": "(A) Idefics2" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 198, + 295, + 220 + ], + "lines": [ + { + "bbox": [ + 55, + 198, + 295, + 220 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 295, + 220 + ], + "type": "text", + "content": "Figure 5. The position-wise accuracy of the three attention mechanisms on the PQA task, alongside our proposed SoFA method." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 178, + 70, + 294, + 175 + ], + "blocks": [ + { + "bbox": [ + 178, + 70, + 294, + 175 + ], + "lines": [ + { + "bbox": [ + 178, + 70, + 294, + 175 + ], + "spans": [ + { + "bbox": [ + 178, + 70, + 294, + 175 + ], + "type": "image", + "image_path": "a77276c3fed29e2dfe69f3f12d0393328d128512c30a1c5591aeebed487915dd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 209, + 178, + 268, + 187 + ], + "lines": [ + { + "bbox": [ + 209, + 178, + 268, + 187 + ], + "spans": [ + { + "bbox": [ + 209, + 178, + 268, + 187 + ], + "type": "text", + "content": "(B) LLaVA-NeXT" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 229, + 191, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 229, + 191, + 243 + ], + "spans": [ + { + "bbox": [ + 55, + 229, + 191, + 243 + ], + "type": "text", + "content": "4.2. Mitigating Position Bias" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 247, + 296, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 247, + 296, + 343 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 296, + 343 + ], + "type": "text", + "content": "The analysis above shows that inter-image causal attention is the main cause of position bias. Furthermore, it also reveals a trade-off between accuracy and robustness, a typical dilemma in machine learning. Building on this insight, we propose SoFt Attention (SoFA), a straightforward, training-free approach to mitigate position bias. SoFA employs linear interpolation between inter-image causal attention and its bidirectional counterpart. Formally," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 99, + 352, + 295, + 381 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 352, + 295, + 381 + ], + "spans": [ + { + "bbox": [ + 99, + 352, + 295, + 381 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {H} _ {\\text {s o f t}} = \\operatorname {S o f t m a x} \\left(\\mathbf {Q K} ^ {\\mathrm {T}} / \\sqrt {d}\\right) \\odot \\mathbf {M} _ {\\text {s o f t}} \\mathbf {V} \\tag {2} \\\\ \\mathbf {M} _ {\\text {s o f t}} = (1 - \\sigma) \\mathbb {1} _ {\\text {c a u s a l}} + \\sigma \\mathbb {1} _ {\\text {b i d i r e c t i o n a l}} \\\\ \\end{array}", + "image_path": "b9bf4ec0160849103874b0bceacf001cba94777f367175a0f7ef341f9aed5db7.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 390, + 295, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 390, + 295, + 450 + ], + "spans": [ + { + "bbox": [ + 55, + 390, + 295, + 450 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 390, + 295, + 450 + ], + "type": "inline_equation", + "content": "\\mathbb{1}_{\\mathrm{causal}}" + }, + { + "bbox": [ + 55, + 390, + 295, + 450 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 390, + 295, + 450 + ], + "type": "inline_equation", + "content": "\\mathbb{1}_{\\mathrm{bidual}}" + }, + { + "bbox": [ + 55, + 390, + 295, + 450 + ], + "type": "text", + "content": " denote the two attention masks, with " + }, + { + "bbox": [ + 55, + 390, + 295, + 450 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 55, + 390, + 295, + 450 + ], + "type": "text", + "content": " controlling the extent of bidirectional part. Note that since we are only interested in the visual positional relation, we only modify the inter-image attention, while preserving causal attention between text segments." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 450, + 295, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 295, + 521 + ], + "type": "text", + "content": "SoFA as a position smoother. Intuitively, given that interimage causal attention excessively emphasizes positional information, SoFA integrates partial bidirectional attention to smooth the influence of position on the models at minimal cost, achieving a balanced trade-off between reasoning performance and position robustness." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 522, + 295, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 594 + ], + "type": "text", + "content": "SoFA as an interaction enhancer. Practically, SoFA enhances the interaction between images, enriching the semantic information of earlier image features to enable more effective utilization by LVLMs. This aligns with the most cases where key information is typically distributed evenly across various input positions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": "For each task, we utilize a small validation set to determine the optimal values for " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": ". Besides, to better align with the training framework, we deploy SoFA every two layers rather than at every layer, while preserving causal attention in the remaining layers. We present SoFA's results in the PQA task as shown in Fig. 5. It is evident that SoFA significantly outperforms both isolated and bidirectional attention, while unlike causal attention in the default setting, SoFA demonstrates balanced performance and exhibits enhanced reasoning capability in most positions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 453, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 453, + 85 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 453, + 85 + ], + "type": "text", + "content": "5. Experiment and Results" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 91, + 553, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 553, + 139 + ], + "type": "text", + "content": "In this section, we empirically demonstrate that SoFA 1) effectively mitigates the position bias inherent in LVLMs and 2) significantly improves their overall reasoning capability across various multi-image applications." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 139, + 553, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 139, + 553, + 186 + ], + "spans": [ + { + "bbox": [ + 313, + 139, + 553, + 186 + ], + "type": "text", + "content": "Models. We select a diverse set of open-source models, including Idefics2-8B [24], InternVL2-8B [11], VILA-13B [31], Mantis-8B [19], and LLaVA-NeXT-Interleave-7B [27]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 187, + 555, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 187, + 555, + 355 + ], + "spans": [ + { + "bbox": [ + 313, + 187, + 555, + 355 + ], + "type": "text", + "content": "Implementation details. For the input, we provide instructions that explicitly indicate the task type and output formats. For instance, for multiple-choice questions, we ask LVLMs to provide only the option symbol; for numeric questions, only the number; and for open-ended questions, the response is limited to short answers. For evaluation, to ensure fairness, we uniformly adopt GPT-4 [1], providing with the question, ground truth answer, candidate prediction, and it determines whether the prediction is correct or not. For SoFA, we determine " + }, + { + "bbox": [ + 313, + 187, + 555, + 355 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 187, + 555, + 355 + ], + "type": "text", + "content": " using a 32-shot validation set per task and perform interpolation every two layers. Unless otherwise specified, all models are set to FP16 precision and utilize Flash Attention [14], with sub-image splitting [32] disabled for fairness." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 361, + 430, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 430, + 372 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 430, + 372 + ], + "type": "text", + "content": "5.1. Position Robustness" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 378, + 554, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 554, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 554, + 545 + ], + "type": "text", + "content": "Benchmark. We first evaluate whether SoFA may effectively mitigate position bias. To this end, similar to §3.1, we select position-agnostic tasks from existing multi-image benchmarks. Specifically, we exclude tasks where changing the order of images may alter the semantics of the question, rendering it unanswerable, e.g., video understanding. Here, we list some of the selected tasks: forensic detection and image jigsaw from BLINK [15], difference spotting and diagram understanding from MuirBench [56], visual analogy and attribute matching from MIRB [66]. We include all examples from Mantis-Eval [19] since the majority are position-agnostic. See the supplementary material for the details of more evaluated tasks. Our experimental setup and metrics are consistent with those described in §3.1." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 546, + 554, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 554, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 554, + 665 + ], + "type": "text", + "content": "SoFA reduces the influence of positions on LVLMs. As demonstrated in Table 1, the integration of SoFA within the models markedly reduces prediction inconsistency, leading to enhanced stability in performance. For instance, Idefics2's inconsistency on BLINK declines from " + }, + { + "bbox": [ + 313, + 546, + 554, + 665 + ], + "type": "inline_equation", + "content": "41.55\\%" + }, + { + "bbox": [ + 313, + 546, + 554, + 665 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 546, + 554, + 665 + ], + "type": "inline_equation", + "content": "12.36\\%" + }, + { + "bbox": [ + 313, + 546, + 554, + 665 + ], + "type": "text", + "content": ", signifying that the responses for the majority of examples remain consistent despite variations in image positioning. This observation indicates that LVLMs exhibit more balanced reasoning capabilities across different positions, effectively mitigating the effects of position bias." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "content": "SoFA eliminates the blind spots of LVLMs. In addition to the enhanced robustness, we observe a marked improvement in the performance of LVLMs in worst-case scenarios, particularly regarding minimum and average accuracy," + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10604" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 553, + 222 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 553, + 222 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 553, + 222 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 553, + 222 + ], + "type": "table", + "html": "
MethodBLINK [15]Mantis-Eval [19]MuirBench [56]MIRB [66]
MinAveMaxIncMinAveMaxIncMinAveMaxIncMinAveMaxInc
Idefics2 [24]43.2646.9348.6841.5543.9346.6251.1038.5743.2845.3047.8134.4645.2647.8849.6829.94
Idefics2 + SoFA47.1848.6349.1212.3648.7149.1850.838.6146.1247.5148.4513.2747.3948.3649.136.99
InternVL2 [11]38.8140.4543.7430.1845.3646.9249.5125.2449.2852.3356.1038.6541.8444.3846.3629.60
InternVL2 + SoFA41.6442.3243.517.1648.1349.2550.1112.0054.6955.9256.785.1644.2045.3545.876.64
VILA [31]45.9348.5951.4225.3348.4449.2051.8521.2941.6643.1248.2737.2647.1749.5952.3431.77
VILA + SoFA48.2750.8051.1710.6849.2951.6052.6812.1645.7646.5247.137.9248.2251.4351.9517.46
Mantis [19]48.3449.2451.5228.3556.4058.3863.4232.1247.6748.9451.1527.4553.1155.7157.4225.79
Mantis + SoFA49.2250.8752.3416.2360.4862.2164.6814.3049.8850.2650.795.6154.3956.3456.938.35
LLaVA-NeXT [27]49.3452.4055.4331.8954.5759.1062.2933.7551.9653.4555.7824.9255.4958.4060.2828.56
LLaVA-NeXT + SoFA53.2054.9156.0811.2058.8761.2361.9814.2053.6155.3756.499.1558.1359.8161.166.96
", + "image_path": "67c6d57bc6564a00c3c00999f574970fe6668258306ca453b51ba3581fefed35.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 260, + 298, + 412 + ], + "blocks": [ + { + "bbox": [ + 55, + 230, + 555, + 253 + ], + "lines": [ + { + "bbox": [ + 55, + 230, + 555, + 253 + ], + "spans": [ + { + "bbox": [ + 55, + 230, + 555, + 253 + ], + "type": "text", + "content": "Table 1. The evaluation on position-agnostic tasks with and without SoFA. Similar to §3.1, we perform multiple evaluations and record minimum, maximum and average accuracy. We also report prediction inconsistency between best and worst-performing evaluations." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 260, + 298, + 412 + ], + "lines": [ + { + "bbox": [ + 56, + 260, + 298, + 412 + ], + "spans": [ + { + "bbox": [ + 56, + 260, + 298, + 412 + ], + "type": "table", + "html": "
MethodBLINKM-EvalMuirMIRBNLVR2MVBench
Random Guessing39.2725.3121.4919.1350.7525.91
Idefics2 [24]45.7848.4142.8446.5881.9423.37
Idefics2 + SoFA48.3150.7747.3648.8385.1525.65
InternVL2 [11]38.9550.3054.5342.6685.5629.31
InternVL2 + SoFA43.2651.1157.1446.1988.1932.77
VILA [31]49.7952.6448.7350.3678.8225.81
VILA + SoFA50.2952.1352.8051.1580.3426.40
Mantis [19]51.4156.8544.2854.5990.2641.59
Mantis + SoFA53.1959.2347.6455.6889.9844.40
LLaVA-NeXT [27]53.3450.8348.2257.1587.2854.26
LLaVA-NeXT + SoFA55.9254.5150.4360.6789.4557.71
", + "image_path": "94a53f18c9e8a5f84ccaacfaa178ebafd19d14929bd45a80a8865fa43fd29f15.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 420, + 295, + 444 + ], + "lines": [ + { + "bbox": [ + 55, + 420, + 295, + 444 + ], + "spans": [ + { + "bbox": [ + 55, + 420, + 295, + 444 + ], + "type": "text", + "content": "Table 2. The overall evaluation accuracy of LVLMs on various multi-image benchmarks with and without SoFA." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 449, + 295, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 295, + 510 + ], + "type": "text", + "content": "which both exhibit an increase of approximately " + }, + { + "bbox": [ + 55, + 449, + 295, + 510 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 55, + 449, + 295, + 510 + ], + "type": "text", + "content": ". This suggests that the model retains strong performance despite variations in the distribution of image positions, indicating that SoFA effectively addresses the deficiencies of LVLMs in positions where performance is typically suboptimal." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 517, + 151, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 151, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 151, + 529 + ], + "type": "text", + "content": "5.2. Overall Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 533, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 653 + ], + "type": "text", + "content": "Benchmark. Given the observation that SoFA effectively mitigates position bias, we proceed with a comprehensive evaluation to determine whether SoFA can enhance the overall reasoning capabilities of LVLMs. Specifically, we include all tasks from the selected benchmarks [15, 19, 56, 66]. In addition, we incorporate NLVR2 [50] and MVBench [29], which exclusively assess the models' capability in image-pair comparison and video understanding, respectively. We directly utilize the default image positions from the benchmarks and record their accuracy." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "SoFA enhances reasoning capability of LVLMs. In Table 2, we observe that integrating SoFA leads to an average improvement of " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "2 \\sim 3\\%" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": " in the performance of LVLMs across various benchmarks. For instance, Idefics2 shows a " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "2.25\\%" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": " improvement on MIRB (46.58% → 48.83%)," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 262, + 555, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 262, + 555, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 262, + 555, + 323 + ], + "type": "text", + "content": "while InternVL2 exhibits a " + }, + { + "bbox": [ + 313, + 262, + 555, + 323 + ], + "type": "inline_equation", + "content": "4.31\\%" + }, + { + "bbox": [ + 313, + 262, + 555, + 323 + ], + "type": "text", + "content": " increase on BLINK " + }, + { + "bbox": [ + 313, + 262, + 555, + 323 + ], + "type": "inline_equation", + "content": "(38.95\\% \\rightarrow 43.26\\%)" + }, + { + "bbox": [ + 313, + 262, + 555, + 323 + ], + "type": "text", + "content": ". This demonstrates that, despite the advancements in multi-image LVLMs, position bias poses a significant obstacle to their reasoning capability, whereas SoFA successfully addresses this limitation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 334, + 387, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 334, + 387, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 334, + 387, + 346 + ], + "type": "text", + "content": "5.3. Discussion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 353, + 555, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 353, + 555, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 353, + 555, + 520 + ], + "type": "text", + "content": "Attention distribution visualization. To gain a deeper understanding of SoFA's working mechanism, we visualize the attention distribution within LVLMs across different positions. Using PQA as an example, we derive the average attention value of the final layer's query token (whose hidden state is directly used for generation) for each image. As shown in Fig. 6, under the default setting, the attention distribution is dominated by the end positions, indicating that the model overly focuses on a few images while neglecting most others. This observation aligns with our findings in §3.2. The application of SoFA alters this trend by significantly smoothing the attention distribution, enabling the model to attend to images at all positions more evenly. This well explains how SoFA mitigates position bias." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "text", + "content": "SoFA improves in-context capability of LVLMs. Multimodal in-context learning is an emerging property recently acquired by LVLMs, wherein a few demonstrations allow them to recognize task formats, thereby significantly improving accuracy. However, recent studies [6, 9, 30] find that LVLMs tend to overlook demonstration images that appear earlier, leading to suboptimal performance. Here, we demonstrate that SoFA can improve the in-context learning capability of LVLMs by eliminating such blind spots. Following [30], We select three typical visual question answering benchmarks: VQAv2 [16], VizWiz [17], and OKVQA [38], and report the average accuracy of LVLMs across varying shot counts. As illustrated in Table 3, the application of SoFA leads to a further enhancement in LVLM performance, with this effect becoming increasingly pronounced as the number of shots rises. For instance, in the" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10605" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 72, + 289, + 178 + ], + "blocks": [ + { + "bbox": [ + 63, + 72, + 289, + 178 + ], + "lines": [ + { + "bbox": [ + 63, + 72, + 289, + 178 + ], + "spans": [ + { + "bbox": [ + 63, + 72, + 289, + 178 + ], + "type": "image", + "image_path": "a7f2cbbd62699c6b45a7bef744223616aa47aba15410d847c57b1da62ea8c5ae.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 187, + 286, + 198 + ], + "lines": [ + { + "bbox": [ + 64, + 187, + 286, + 198 + ], + "spans": [ + { + "bbox": [ + 64, + 187, + 286, + 198 + ], + "type": "text", + "content": "Figure 6. The attention distribution across positions on PQA." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 207, + 296, + 259 + ], + "blocks": [ + { + "bbox": [ + 56, + 207, + 296, + 259 + ], + "lines": [ + { + "bbox": [ + 56, + 207, + 296, + 259 + ], + "spans": [ + { + "bbox": [ + 56, + 207, + 296, + 259 + ], + "type": "table", + "html": "
MethodVQAv2 [16]VizWiz [17]OK-VQA [38]
4-shot8-shot16-shot4-shot8-shot16-shot4-shot8-shot16-shot
w/o SoFA55.3459.3261.3741.5644.6545.3545.1948.3049.02
w/ SoFA56.1361.3064.8542.3046.9849.1745.9349.8752.36
", + "image_path": "824dcc3879ba9662972f9bca09375c0e417a86ebd0769e17bb13668493139531.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 63, + 299, + 289, + 339 + ], + "blocks": [ + { + "bbox": [ + 55, + 268, + 295, + 291 + ], + "lines": [ + { + "bbox": [ + 55, + 268, + 295, + 291 + ], + "spans": [ + { + "bbox": [ + 55, + 268, + 295, + 291 + ], + "type": "text", + "content": "Table 3. The average accuracy of LVLMs on visual question answering using in-context learning with and without SoFA." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 299, + 289, + 339 + ], + "lines": [ + { + "bbox": [ + 63, + 299, + 289, + 339 + ], + "spans": [ + { + "bbox": [ + 63, + 299, + 289, + 339 + ], + "type": "table", + "html": "
# Images351020305075100
w/o SoFA82.1681.2777.1872.3668.5060.8055.1349.19
w/ SoFA83.9082.3378.9274.3070.6763.2759.3855.11
", + "image_path": "e2484e051e858c3da715917df8fa8bdaf56d1733ff45d8acf3354f6c2a3a1c14.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 349, + 295, + 371 + ], + "lines": [ + { + "bbox": [ + 55, + 349, + 295, + 371 + ], + "spans": [ + { + "bbox": [ + 55, + 349, + 295, + 371 + ], + "type": "text", + "content": "Table 4. The average accuracy of LVLMs on Multi-modal Haystack [59], where images range from a few to a hundred." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 378, + 295, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 378, + 295, + 426 + ], + "spans": [ + { + "bbox": [ + 55, + 378, + 295, + 426 + ], + "type": "text", + "content": "case of VizWiz, accuracy improves from " + }, + { + "bbox": [ + 55, + 378, + 295, + 426 + ], + "type": "inline_equation", + "content": "41.56\\%" + }, + { + "bbox": [ + 55, + 378, + 295, + 426 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 378, + 295, + 426 + ], + "type": "inline_equation", + "content": "42.30\\%" + }, + { + "bbox": [ + 55, + 378, + 295, + 426 + ], + "type": "text", + "content": " with 4 shots, and from " + }, + { + "bbox": [ + 55, + 378, + 295, + 426 + ], + "type": "inline_equation", + "content": "45.35\\%" + }, + { + "bbox": [ + 55, + 378, + 295, + 426 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 378, + 295, + 426 + ], + "type": "inline_equation", + "content": "49.17\\%" + }, + { + "bbox": [ + 55, + 378, + 295, + 426 + ], + "type": "text", + "content": " with 16 shots. This indicates that position bias undermines the effectiveness of demonstrations, while SoFA mitigates this issue." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 426, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 295, + 605 + ], + "type": "text", + "content": "SoFA exhibits advantages in long-context applications. Long-context scenarios have recently become a focus, where a single inference requires processing a substantial number of images. This is particularly beneficial for applications such as many-shot in-context learning [20] and multi-modal RAG [62]. Here, we show that SoFA can excel in these challenging scenarios. We select Visual Haystack [61], a benchmark designed to assess the long-context capabilities of LVLMs, and increase the input scale to encompass anywhere from a few images to a hundred. As shown in Table 4, the benefits of SoFA become increasingly evident as the number of images rises, achieving peak performance gains at 100 images (49.19% → 55.11%). This indicates that an increase in context length exacerbates the model's position bias, while SoFA compensates this loss." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "text", + "content": "Contribution of SoFA to different tasks. Extending beyond the holistic evaluation, we categorize the benchmarks into various tasks to identify which ones benefit the most from SoFA. As depicted in Fig. 7, SoFA demonstrates the highest improvements in visual retrieval and analogy, with gains of " + }, + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "6.84\\%" + }, + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "5.53\\%" + }, + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "text", + "content": ", respectively. In these tasks, the first image acts as a reference, and the models need to select a matching scene or similar style from the remaining images. This necessitates that the model pays extra attention" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 322, + 70, + 548, + 161 + ], + "blocks": [ + { + "bbox": [ + 322, + 70, + 548, + 161 + ], + "lines": [ + { + "bbox": [ + 322, + 70, + 548, + 161 + ], + "spans": [ + { + "bbox": [ + 322, + 70, + 548, + 161 + ], + "type": "image", + "image_path": "9e6db53c170794c570aee95c360d8bb9f82bd3d31e2c300b24d7a531cbfb23f0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 319, + 199, + 550, + 270 + ], + "blocks": [ + { + "bbox": [ + 313, + 167, + 555, + 190 + ], + "lines": [ + { + "bbox": [ + 313, + 167, + 555, + 190 + ], + "spans": [ + { + "bbox": [ + 313, + 167, + 555, + 190 + ], + "type": "text", + "content": "Figure 7. The performance gains of SoFA on different types of tasks. The results are averaged over selected models." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 199, + 550, + 270 + ], + "lines": [ + { + "bbox": [ + 319, + 199, + 550, + 270 + ], + "spans": [ + { + "bbox": [ + 319, + 199, + 550, + 270 + ], + "type": "table", + "html": "
σBLINKMantis-EvalMuirBenchMIRBNLVR2MVBench
0.0047.8551.8147.7250.2784.7734.87
0.2549.2252.7049.3552.3485.4137.28
0.5050.0453.3350.9651.8686.3036.56
0.7549.6253.7448.2451.1486.8335.07
1.0048.2151.5847.3750.5385.3734.20
", + "image_path": "5b12e6e9bfe168598e7942ce232132937cc211feb8fb9a17ee9922c4c7da2d4a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 320, + 278, + 547, + 289 + ], + "lines": [ + { + "bbox": [ + 320, + 278, + 547, + 289 + ], + "spans": [ + { + "bbox": [ + 320, + 278, + 547, + 289 + ], + "type": "text", + "content": "Table 5. The results on six benchmarks varying the value of " + }, + { + "bbox": [ + 320, + 278, + 547, + 289 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 320, + 278, + 547, + 289 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 300, + 553, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 553, + 361 + ], + "type": "text", + "content": "to and adequately understands the first image. Intuitively, due to the recency bias inherent in open-source models, they often overlook earlier information, resulting in suboptimal performance. Thus, the significant gains brought by SoFA are understandable." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 363, + 555, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 363, + 555, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 555, + 483 + ], + "type": "text", + "content": "The interpretation of " + }, + { + "bbox": [ + 313, + 363, + 555, + 483 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 363, + 555, + 483 + ], + "type": "text", + "content": ". In Table 5, we conduct an ablation study on the effect of " + }, + { + "bbox": [ + 313, + 363, + 555, + 483 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 363, + 555, + 483 + ], + "type": "text", + "content": ". It can be observed that the optimal " + }, + { + "bbox": [ + 313, + 363, + 555, + 483 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 363, + 555, + 483 + ], + "type": "text", + "content": " varies significantly across benchmarks, for instance, with a value of 0.75 on NLVR2 and only 0.25 on MVBench. Intuitively, MVBench inputs involve video frames, which have a strong dependence on image positions, and an excessively high " + }, + { + "bbox": [ + 313, + 363, + 555, + 483 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 363, + 555, + 483 + ], + "type": "text", + "content": " may compromise such positional information. In contrast, the examples in NLVR2 are image pairs, most of which are position-agnostic. This difference in task nature accounts for the variation in " + }, + { + "bbox": [ + 313, + 363, + 555, + 483 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 363, + 555, + 483 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 499, + 388, + 512 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 499, + 388, + 512 + ], + "spans": [ + { + "bbox": [ + 313, + 499, + 388, + 512 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 521, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 521, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 521, + 555, + 713 + ], + "type": "text", + "content": "In this study, we introduce the concept of position bias, a well-known issue in NLP, into the realm of multi-modal applications. First, we reveal that LVLMs cannot robustly make use of information across multiple images, where merely altering the position of images may significantly change their predictions. Then we further identify this position bias, concluding that open-source models exhibit a strong recency bias, while proprietary models struggle with positions in the middle. Motivated by these findings, we propose SoFA, a simple, training-free approach that reduces the model's dependence on positional information through linear interpolation between inter-image causal attention and bidirectional counterparts. Experiments demonstrate that SoFA effectively mitigates position bias and significantly enhances the overall reasoning performance of existing models." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10606" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 138, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 138, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 138, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 87, + 296, + 154 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 87, + 296, + 154 + ], + "spans": [ + { + "bbox": [ + 55, + 87, + 296, + 154 + ], + "type": "text", + "content": "This research was, in part, funded by the U.S. Government - DARPA TIAMAT HR00112490421. The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the U.S. Government." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 176, + 115, + 189 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 176, + 115, + 189 + ], + "spans": [ + { + "bbox": [ + 56, + 176, + 115, + 189 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 196, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 196, + 296, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 196, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 61, + 196, + 296, + 251 + ], + "type": "text", + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1, 2, 3, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 252, + 296, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 252, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 61, + 252, + 296, + 361 + ], + "type": "text", + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob L. Menick, Sebastian Borgeaud, Andy Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karén Simonyan. Flamingo: a visual language model for few-shot learning. In NeurIPS, 2022. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 363, + 295, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 363, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 61, + 363, + 295, + 407 + ], + "type": "text", + "content": "[3] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In CVPR, pages 2425-2433, 2015. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 409, + 296, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 409, + 296, + 474 + ], + "spans": [ + { + "bbox": [ + 62, + 409, + 296, + 474 + ], + "type": "text", + "content": "[4] Anas Awadalla, Irena Gao, Josh Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Shiori Sagawa, et al. Openflamingo: An opensource framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390, 2023. 1, 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 476, + 296, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 476, + 296, + 520 + ], + "spans": [ + { + "bbox": [ + 62, + 476, + 296, + 520 + ], + "type": "text", + "content": "[5] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 521, + 296, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 521, + 296, + 565 + ], + "spans": [ + { + "bbox": [ + 62, + 521, + 296, + 565 + ], + "type": "text", + "content": "[6] Folco Bertini Baldassini, Mustafa Shukor, Matthieu Cord, Laure Soulier, and Benjamin Piwowarski. What makes multimodal in-context learning work? In CVPR, pages 1539-1550, 2024. 3, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 567, + 296, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 567, + 296, + 600 + ], + "spans": [ + { + "bbox": [ + 62, + 567, + 296, + 600 + ], + "type": "text", + "content": "[7] Rohan Bavishi, Erich Elsen, Curtis Hawthorne, Maxwell Nye, Augustus Odena, Arushi Somani, and Săgnak Taşirlar. Introducing our multimodal models, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 601, + 296, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 601, + 296, + 644 + ], + "spans": [ + { + "bbox": [ + 62, + 601, + 296, + 644 + ], + "type": "text", + "content": "[8] Liangyu Chen, Bo Li, Sheng Shen, Jingkang Yang, Chunyuan Li, Kurt Keutzer, Trevor Darrell, and Ziwei Liu. Large language models are visual reasoning coordinators. In NeurIPS, 2024. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 647, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 647, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 62, + 647, + 296, + 689 + ], + "type": "text", + "content": "[9] Shuo Chen, Zhen Han, Bailan He, Mark Buckley, Philip Torr, Volker Tresp, and Jindong Gu. Understanding and improving in-context learning on vision-language models. In ICLR, 2023. 3, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "text", + "content": "[10] Yuhan Chen, Ang Lv, Ting-En Lin, Changyu Chen, Yuchuan Wu, Fei Huang, Yongbin Li, and Rui Yan. Fortify the short" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "est stave in attention: Enhancing context awareness of large language models for effective tool use. In ACL, 2023. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 98, + 553, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 98, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 316, + 98, + 553, + 152 + ], + "type": "text", + "content": "[11] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In CVPR, pages 24185–24198, 2024. 1, 2, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 154, + 553, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 154, + 553, + 219 + ], + "spans": [ + { + "bbox": [ + 316, + 154, + 553, + 219 + ], + "type": "text", + "content": "[12] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with " + }, + { + "bbox": [ + 316, + 154, + 553, + 219 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 316, + 154, + 553, + 219 + ], + "type": "text", + "content": " chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2(3):6, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 222, + 553, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 222, + 553, + 266 + ], + "spans": [ + { + "bbox": [ + 316, + 222, + 553, + 266 + ], + "type": "text", + "content": "[13] Neil Cohn, Ryan Taylor, and Kaitlin Pederson. A picture is worth more words over time: Multimodality and narrative structure across eight decades of american superhero comics. Multimodal Communication, 6(1):19-37, 2017. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 268, + 553, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 268, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 316, + 268, + 553, + 312 + ], + "type": "text", + "content": "[14] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. In NeurIPS, pages 16344-16359, 2022. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 314, + 553, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 314, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 316, + 314, + 553, + 358 + ], + "type": "text", + "content": "[15] Xingyu Fu, Yushi Hu, Bangzheng Li, Yu Feng, Haoyu Wang, Xudong Lin, Dan Roth, Noah A Smith, Wei-Chiu Ma, and Ranjay Krishna. Blink: Multimodal large language models can see but not perceive. In ECCV, 2024. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 360, + 553, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 360, + 553, + 405 + ], + "spans": [ + { + "bbox": [ + 316, + 360, + 553, + 405 + ], + "type": "text", + "content": "[16] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In CVPR, pages 6904-6913, 2017. 1, 4, 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 407, + 553, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 407, + 553, + 450 + ], + "spans": [ + { + "bbox": [ + 316, + 407, + 553, + 450 + ], + "type": "text", + "content": "[17] Danna Gurari, Qing Li, Abigale J Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo Luo, and Jeffrey P Bigham. Vizwiz grand challenge: Answering visual questions from blind people. In CVPR, pages 3608-3617, 2018. 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 453, + 553, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 453, + 553, + 518 + ], + "spans": [ + { + "bbox": [ + 316, + 453, + 553, + 518 + ], + "type": "text", + "content": "[18] Junqing He, Kunhao Pan, Xiaoqun Dong, Zhuoyang Song, LiuYiBo LiuYiBo, Qiangosun Qianguosun, Yuxin Liang, Hao Wang, Enming Zhang, and Jiaxing Zhang. Never lost in the middle: Mastering long-context question answering with position-agnostic decompositional training. In ACL, pages 13628–13642, 2024. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 520, + 553, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 520, + 553, + 564 + ], + "spans": [ + { + "bbox": [ + 316, + 520, + 553, + 564 + ], + "type": "text", + "content": "[19] Dongfu Jiang, Xuan He, Huaye Zeng, Cong Wei, Max Ku, Qian Liu, and Wenhu Chen. Mantis: Interleaved multi-image instruction tuning. arXiv preprint arXiv:2405.01483, 2024. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 567, + 553, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 567, + 553, + 609 + ], + "spans": [ + { + "bbox": [ + 316, + 567, + 553, + 609 + ], + "type": "text", + "content": "[20] Yixing Jiang, Jeremy Irvin, Ji Hun Wang, Muhammad Ahmed Chaudhry, Jonathan H Chen, and Andrew Y Ng. Many-shot in-context learning in multimodal foundation models. In ICML, 2024. 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 612, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 612, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 612, + 553, + 667 + ], + "type": "text", + "content": "[21] Mehran Kazemi, Nishanth Dikkala, Ankit Anand, Petar Devic, Ishita Dasgupta, Fangyu Liu, Bahare Fatemi, Pranjal Awasthi, Dee Guo, Sreenivas Gollapudi, et al. Remi: A dataset for reasoning with multiple images. arXiv preprint arXiv:2406.09175, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 670, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 714 + ], + "type": "text", + "content": "[22] Hugo Laurençon, Andrés Marafioti, Victor Sanh, and Léo Tronchon. Building and better understanding vision-language models: insights and future directions. arXiv preprint arXiv:2408.12637, 2024. 1" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10607" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[23] Hugo Laurençon, Lucile Saulnier, Léo Tronchon, Stas Bekman, Amanpreet Singh, Anton Lozhkov, Thomas Wang, Siddharth Karamcheti, Alexander Rush, Douwe Kiela, et al. Obelics: An open web-scale filtered dataset of interleaved image-text documents. In NeurIPS, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 295, + 162 + ], + "type": "text", + "content": "[24] Hugo Laurençon, Léo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language models? arXiv preprint arXiv:2405.02246, 2024. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "type": "text", + "content": "[25] Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Kuttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al. Retrieval-augmented generation for knowledge-intensive nlp tasks. In CVPR, pages 9459-9474, 2020. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 294, + 262 + ], + "type": "text", + "content": "[26] Bohao Li, Yuying Ge, Yixiao Ge, Guangzhi Wang, Rui Wang, Ruimao Zhang, and Ying Shan. Seed-bench: Benchmarking multimodal large language models. In CVPR, pages 13299-13308, 2024. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 294, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 317 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 317 + ], + "type": "text", + "content": "[27] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava next-interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 319, + 294, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 319, + 294, + 364 + ], + "spans": [ + { + "bbox": [ + 56, + 319, + 294, + 364 + ], + "type": "text", + "content": "[28] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, pages 19730–19742. PMLR, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 294, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 409 + ], + "type": "text", + "content": "[29] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mybench: A comprehensive multi-modal video understanding benchmark. In CVPR, pages 22195-22206, 2024. 2, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 411, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 453 + ], + "type": "text", + "content": "[30] Li Li, Jiawei Peng, Huiyi Chen, Chongyang Gao, and Xu Yang. How to configure good in-context sequence for visual question answering. In CVPR, pages 26710-26720, 2024. 2, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 455, + 294, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 455, + 294, + 488 + ], + "spans": [ + { + "bbox": [ + 56, + 455, + 294, + 488 + ], + "type": "text", + "content": "[31] Ji Lin, Hongxu Yin, Wei Ping, Pavlo Molchanov, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models. In CVPR, pages 26689-26699, 2024. 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 490, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 490, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 490, + 294, + 544 + ], + "type": "text", + "content": "[32] Ziyi Lin, Chris Liu, Renrui Zhang, Peng Gao, Longtian Qiu, Han Xiao, Han Qiu, Chen Lin, Wenqi Shao, Keqin Chen, et al. Sphinx: The joint mixing of weights, tasks, and visual embeddings for multi-modal large language models. arXiv preprint arXiv:2311.07575, 2023. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 545, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 294, + 578 + ], + "type": "text", + "content": "[33] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 580, + 294, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 580, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 580, + 294, + 601 + ], + "type": "text", + "content": "[34] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2024. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 603, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 294, + 657 + ], + "type": "text", + "content": "[35] Nelson F Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. Lost in the middle: How language models use long contexts. In ACL, pages 157-173. MIT Press One Broadway, 12th Floor, Cambridge, Massachusetts 02142, USA ..., 2024. 1, 2, 3, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "text", + "content": "[36] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. In NeurIPS, pages 2507–2521, 2022. 1" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 117 + ], + "type": "text", + "content": "[37] Pan Lu, Baolin Peng, Hao Cheng, Michel Galley, Kai-Wei Chang, Ying Nian Wu, Song-Chun Zhu, and Jianfeng Gao. Chameleon: Plug-and-play compositional reasoning with large language models. In NeurIPS, 2024. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 119, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 119, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 119, + 553, + 162 + ], + "type": "text", + "content": "[38] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In CVPR, pages 3195-3204, 2019. 1, 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 165, + 553, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 165, + 553, + 220 + ], + "spans": [ + { + "bbox": [ + 316, + 165, + 553, + 220 + ], + "type": "text", + "content": "[39] Duy MH Nguyen, Hoang Nguyen, Nghiem Diep, Tan Ngoc Pham, Tri Cao, Binh Nguyen, Paul Swoboda, Nhat Ho, Shadi Albarqouni, Pengtao Xie, et al. Lvm-med: Learning large-scale self-supervised vision models for medical imaging via second-order graph matching. In NeurIPS, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 222, + 553, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 222, + 553, + 265 + ], + "spans": [ + { + "bbox": [ + 316, + 222, + 553, + 265 + ], + "type": "text", + "content": "[40] Yassine Ouali, Adrian Bulat, Brais Martinez, and Georgios Tzimiropoulos. Clip-dpo: Vision-language models as a source of preference for fixing hallucinations in lvlms. In ECCV, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 268, + 553, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 268, + 553, + 311 + ], + "spans": [ + { + "bbox": [ + 316, + 268, + 553, + 311 + ], + "type": "text", + "content": "[41] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. In ICLR, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 314, + 553, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 314, + 553, + 346 + ], + "spans": [ + { + "bbox": [ + 316, + 314, + 553, + 346 + ], + "type": "text", + "content": "[42] Alexander Peysakhovich and Adam Lerer. Attention sorting combats recency bias in long context language models. arXiv preprint arXiv:2310.01427, 2023. 1, 3, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 350, + 553, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 414 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 414 + ], + "type": "text", + "content": "[43] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 417, + 553, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 417, + 553, + 472 + ], + "spans": [ + { + "bbox": [ + 316, + 417, + 553, + 472 + ], + "type": "text", + "content": "[44] Ori Ram, Yoav Levine, Itay Dalmedigos, Dor Muhlgay, Amnon Shashua, Kevin Leyton-Brown, and Yoav Shoham. Incontext retrieval-augmented language models. In ACL, pages 1316-1331. MIT Press One Broadway, 12th Floor, Cambridge, Massachusetts 02142, USA . . ., 2023. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 474, + 553, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 474, + 553, + 518 + ], + "spans": [ + { + "bbox": [ + 316, + 474, + 553, + 518 + ], + "type": "text", + "content": "[45] Nir Ratner, Yoav Levine, Yonatan Belinkov, Ori Ram, Inbal Magar, Omri Abend, Ehud Karpas, Amnon Shashua, Kevin Leyton-Brown, and Yoav Shoham. Parallel context windows for large language models. In ACL, 2022. 3, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 520, + 553, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 520, + 553, + 575 + ], + "spans": [ + { + "bbox": [ + 316, + 520, + 553, + 575 + ], + "type": "text", + "content": "[46] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 577, + 553, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 577, + 553, + 620 + ], + "spans": [ + { + "bbox": [ + 316, + 577, + 553, + 620 + ], + "type": "text", + "content": "[47] Junjie Shentu and Noura Al Moubayed. Cxr-irgen: An integrated vision and language model for the generation of clinically accurate chest x-ray image-report pairs. In WACV, pages 5212–5221, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 623, + 553, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 623, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 316, + 623, + 553, + 666 + ], + "type": "text", + "content": "[48] Lin Shi, Weicheng Ma, and Soroush Vosoughi. Judging the judges: A systematic investigation of position bias in pairwise comparative assessments by llms. arXiv preprint arXiv:2406.07791, 2024. 1, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[49] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 5" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "10608" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[50] Alane Suhr, Stephanie Zhou, Ally Zhang, Iris Zhang, Huajun Bai, and Yoav Artzi. A corpus for reasoning about natural language grounded in photographs. In ACL, 2018. 1, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 294, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 294, + 160 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 294, + 160 + ], + "type": "text", + "content": "[51] Quan Sun, Yufeng Cui, Xiaosong Zhang, Fan Zhang, Qiying Yu, Yueze Wang, Yongming Rao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative multimodal models are in-context learners. In CVPR, pages 14398-14409, 2024. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 161, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 161, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 161, + 294, + 194 + ], + "type": "text", + "content": "[52] Dídac Surís, Sachit Menon, and Carl Vondrick. Vipergpt: Visual inference via python execution for reasoning. In CVPR, pages 11888-11898, 2023. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 194, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 194, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 194, + 294, + 217 + ], + "type": "text", + "content": "[53] MosaicML NLP Team et al. Introducing mpt-7b: A new standard for open-source, commercially usable llms, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 217, + 294, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 271 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 271 + ], + "type": "text", + "content": "[54] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 272, + 294, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 294, + 326 + ], + "type": "text", + "content": "[55] Pat Verga, Sebastian Hofstatter, Sophia Althammer, Yixuan Su, Aleksandra Piktus, Arkady Arkhangorodsky, Minjie Xu, Naomi White, and Patrick Lewis. Replacing judges with juries: Evaluating llm generations with a panel of diverse models. arXiv preprint arXiv:2404.18796, 2024. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 327, + 294, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 381 + ], + "type": "text", + "content": "[56] Fei Wang, Xingyu Fu, James Y Huang, Zekun Li, Qin Liu, Xiaogeng Liu, Mingyu Derek Ma, Nan Xu, Wenxuan Zhou, Kai Zhang, et al. Muirbench: A comprehensive benchmark for robust multi-image understanding. arXiv preprint arXiv:2406.09411, 2024. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 383, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 294, + 437 + ], + "type": "text", + "content": "[57] Hengyi Wang, Haizhou Shi, Shiwei Tan, Weiyi Qin, Wenyuan Wang, Tunyu Zhang, Akshay Nambi, Tanuja Ganu, and Hao Wang. Multimodal needle in a haystack: Benchmarking long-context capability of multimodal large language models. arXiv preprint arXiv:2406.11230, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 437, + 294, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 480 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 480 + ], + "type": "text", + "content": "[58] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, et al. Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 481, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 481, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 481, + 294, + 525 + ], + "type": "text", + "content": "[59] Weiyun Wang, Shuibao Zhang, Yiming Ren, Yuchen Duan, Tiantong Li, Shuo Liu, Mengkang Hu, Zhe Chen, Kaipeng Zhang, Lewei Lu, et al. Needle in a multimodal haystack. In NeurIPS, 2024. 3, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 525, + 294, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 525, + 294, + 570 + ], + "spans": [ + { + "bbox": [ + 56, + 525, + 294, + 570 + ], + "type": "text", + "content": "[60] Ziqi Wang, Hanlin Zhang, Xiner Li, Kuan-Hao Huang, Chi Han, Shuiwang Ji, Sham M Kakade, Hao Peng, and Heng Ji. Eliminating position bias of language models: A mechanistic approach. arXiv preprint arXiv:2407.01100, 2024. 1, 3, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 571, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 571, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 571, + 294, + 613 + ], + "type": "text", + "content": "[61] Tsung-Han Wu, Giscard Biamby, Jerome Quenum, Ritwik Gupta, Joseph E Gonzalez, Trevor Darrell, and David M Chan. Visual haystacks: Answering harder questions about sets of images. arXiv preprint arXiv:2407.13766, 2024. 3, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 614, + 294, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 294, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 294, + 658 + ], + "type": "text", + "content": "[62] Michihiro Yasunaga, Armen Aghajanyan, Weijia Shi, Rich James, Jure Leskovec, Percy Liang, Mike Lewis, Luke Zettlemoyer, and Wen-tau Yih. Retrieval-augmented multimodal language modeling. In ICML, 2022. 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[63] Yijiong Yu, Huiqiang Jiang, Xufang Luo, Qianhui Wu, ChinYew Lin, Dongsheng Li, Yuqing Yang, Yongfeng Huang, and Lili Qiu. Mitigate position bias in large language models via scaling a single dimension. arXiv preprint arXiv:2406.02536, 2024. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 473 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "text", + "content": "[64] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In CVPR, pages 9556-9567, 2024. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 129, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 162 + ], + "type": "text", + "content": "[65] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In CVPR, pages 11975-11986, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "type": "text", + "content": "[66] Bingchen Zhao, Yongshuo Zong, Letian Zhang, and Timothy Hospedales. Benchmarking multi-image understanding in vision and language models: Perception, knowledge, reasoning, and multi-hop reasoning. arXiv preprint arXiv:2406.12742, 2024. 3, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 218, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 218, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 218, + 553, + 271 + ], + "type": "text", + "content": "[67] Haozhe Zhao, Zefan Cai, Shuzheng Si, Xiaojian Ma, Kaikai An, Liang Chen, Zixuan Liu, Sheng Wang, Wenjuan Han, and Baobao Chang. Mmcl: Empowering vision-language model with multi-modal in-context learning. In ICLR, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 274, + 553, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 274, + 553, + 328 + ], + "spans": [ + { + "bbox": [ + 316, + 274, + 553, + 328 + ], + "type": "text", + "content": "[68] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. In NeurlPS, pages 46595-46623, 2023. 1, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 330, + 553, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 330, + 553, + 373 + ], + "spans": [ + { + "bbox": [ + 316, + 330, + 553, + 373 + ], + "type": "text", + "content": "[69] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. In ICLR, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 374, + 553, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 374, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 374, + 553, + 430 + ], + "type": "text", + "content": "[70] Wanrong Zhu, Jack Hessel, Anas Awadalla, Samir Yitzhak Gadre, Jesse Dodge, Alex Fang, Youngjae Yu, Ludwig Schmidt, William Yang Wang, and Yejin Choi. Multimodal c4: An open, billion-scale corpus of images interleaved with text. In NeurIPS, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 430, + 553, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 430, + 553, + 473 + ], + "spans": [ + { + "bbox": [ + 316, + 430, + 553, + 473 + ], + "type": "text", + "content": "[71] Yuchen Zhuang, Yue Yu, Kuan Wang, Haotian Sun, and Chao Zhang. Toolqa: A dataset for llm question answering with external tools. In NeurIPS, pages 50117-50143, 2023. 1" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10609" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Identifying and Mitigating Spurious Correlation in Multi-Task Learning/67faa86d-7895-441b-b372-45b9fee04763_content_list.json b/2025/Identifying and Mitigating Spurious Correlation in Multi-Task Learning/67faa86d-7895-441b-b372-45b9fee04763_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f091329f7d0a81dab07a4a56c564d82dc18bc261 --- /dev/null +++ b/2025/Identifying and Mitigating Spurious Correlation in Multi-Task Learning/67faa86d-7895-441b-b372-45b9fee04763_content_list.json @@ -0,0 +1,1579 @@ +[ + { + "type": "text", + "text": "Identifying and Mitigating Spurious Correlation in Multi-Task Learning", + "text_level": 1, + "bbox": [ + 133, + 128, + 864, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Junyi Chai, Shenyu Lu & Xiaoqian Wang* \nElmore Family School of Electrical and Computer Engineering, Purdue University \nWest Lafayette, IN, USA", + "bbox": [ + 173, + 180, + 823, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{chai28,lu876,joywang}@purdue.edu", + "bbox": [ + 349, + 236, + 640, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 286, + 326, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multi-task learning (MTL) is a paradigm that aims to improve the generalization of models by simultaneously learning multiple related tasks, leveraging shared representations and task-specific information to enhance performance on individual tasks. However, existing work has shown that MTL can potentially hinder generalization, with one key factor being spurious correlations between tasks. Owing to the knowledge-sharing property, the per-task predictors are more likely to develop reliance on spurious features. Most existing approaches address this issue through distributional robustness, aiming to maintain consistent performance across different distributions under unknown covariate shifts. However, this formulation lacks theoretical guarantees and can be sensitive to the construction of covariate shifts. In this work, we propose a novel perspective, where we seek to identify spurious correlations between tasks. Drawing inspirations from conventional formulations on spurious correlation, for each task, we propose to distinguish its spurious tasks using the difference in correlation coefficients between the empirical distribution and class-wise resampled distributions, thereby capturing the correlations between task labels w.r.t. each class. We prove theoretically the feasibility of the resampling strategy in characterizing spurious correlations between tasks. Furthermore, we propose a simple fine-tuning strategy, debiased adversarial training, where the per-task predictors are adversarily trained to disregard information associated with their spurious tasks. Experimental results on six benchmark datasets show that our method effectively mitigates spurious correlations and outperforms state-of-the-art methods in improving generalization.", + "bbox": [ + 91, + 318, + 483, + 787 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 816, + 220, + 833 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Stemming from Stein's paradox [37] where the estimation error decreases under joint estimation, formalized in [4],", + "bbox": [ + 89, + 844, + 482, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "multi-task learning (MTL) has since become increasingly important in various applications, particularly in computer vision, demonstrating both improved accuracy and reduced computational overhead. MTL aims to jointly learn multiple tasks, allowing each task to leverages the shared representations and domain knowledge between tasks, with the goal of improving the generalization performance across all tasks [19, 46]. Compared with single-task learning, which focuses on optimizing over one specific objective, MTL extracts common patterns that are beneficial across different but related objectives, potentially leading to better performance on individual tasks. This collaborative framework mitigates overfitting and enhances model robustness through cross-task knowledge transfer [28, 29].", + "bbox": [ + 511, + 287, + 906, + 500 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite the benefits in MTL, existing work has demonstrated that under certain circumstances, MTL could lead to degraded performance in generalization, particularly when different tasks do not experience high knowledge-sharing [10, 36]. This abnormality has conventionally been attributed to the model capacity issue, where conflicting tasks may compete during optimization [42, 44]. However, Recent studies have revealed a counterintuitive phenomenon in MTL, where even under over-parametrization, a MTL model might excel in minimizing training loss, yet surprisingly underperform single-task models in terms of generalization [14]. Accordingly, spurious correlations between different tasks have been identified as an intrinsic problem that hurts generalization in MTL [14, 27]. Owing to the multi-tasking property, the shared representations are expected to encode knowledge across tasks; consequently, when two tasks exhibit high label correlations despite having unrelated underlying knowledge, the per-task predictors may erroneously leverage unrelated spurious knowledge, potentially compromising generalization.", + "bbox": [ + 511, + 503, + 908, + 806 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Most of existing work addressing spurious correlations in MTL approaches this issue through the lens of invariant learning, aiming to ensure that models perform consistently across various empirical distributions [11, 14]. Specifically, the shared representations are expected to recover the ground-truth latent causal features, and by en", + "bbox": [ + 511, + 809, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 886, + 236, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "25698", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "forcing distributional-invariance, the per-task predictors are expected to leverage features that are causally related to each task, instead of spurious features. Nonetheless, on one hand, counterfactual data is generally inaccessible on high-dimensional data, leading to the infeasibility in conducting causality discovery [2]. Consequently, the shared representations may fail to recover the underlying latent features, particularly when approximated by disentanglement across latent dimensions [11, 14]. Moreover, enforcing distributional invariance does not guarantee utilization of only the task-specific causal features, as the difference across empirical distributions does not necessarily break the spurious correlations between tasks [5]. In the absence of prior knowledge regarding spurious correlations between tasks, along with the potential lack of certain label pairings, it becomes difficult to construct the desired distributions. Alternatively, Makino et al. [27] proposes to mitigate the spurious correlations between tasks by generatively estimating the joint distribution of task labels. Despite theoretical soundness of the formulation, estimating joint distributions can be error-prone, especially when certain label combinations are missing. Additionally, estimating joint distributions involving labels of varied dimensions is often infeasible. This issue is particularly relevant for vision tasks, where labels from different tasks are likely to have differing dimensions.", + "bbox": [ + 93, + 90, + 480, + 482 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose an alternative formulation regarding spurious correlations in MTL. Instead of imposing regularizations on distributional invariance, we seek to directly identify the spurious correlations between tasks. Drawing inspirations from conventional formulations of spurious correlation, typically defined by the dependency between labels and spurious information, we propose to identify spurious correlations between tasks using correlation coefficients. Specifically, we detect spurious correlations between two tasks by examining differences in their correlation coefficients, calculated on both the original training data and class-wise resampled training data. We prove theoretically the feasibility of our formulation in identifying spurious correlations between tasks, as well as in distinguishing spurious correlations from genuine task relationships. Building upon the per-task identification of spurious correlations, we formulate the bias mitigation problem as a mini-max game, where for each task, we consider the predictors of its spurious tasks as the adversaries, such that by fine-tuning the task-specific projector and the task-specific predictor, we intend to minimize the training loss for this task, while maximizing the training loss for its spurious tasks under projection.", + "bbox": [ + 93, + 487, + 480, + 833 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We summarize our contribution as follows:", + "bbox": [ + 112, + 838, + 393, + 849 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We propose a theoretical framework for detecting spurious correlations in MTL by analyzing intra-task label dependencies, and we prove theoretically the feasibility of", + "bbox": [ + 94, + 856, + 480, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "our method in the identification.", + "bbox": [ + 529, + 92, + 740, + 104 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We address spurious correlations in multi-task learning by fine-tuning task-specific predictors, enhancing per-task performance through adversarial training that minimizes loss on the primary task while maximizing loss on its identified spurious tasks.", + "- We validate from experiments the effectiveness of our framework in identification of spurious correlations between tasks, as well as in improving generalizability on four benchmark datasets." + ], + "bbox": [ + 516, + 106, + 903, + 239 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 516, + 265, + 651, + 279 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Generalization in multi-task learning", + "text_level": 1, + "bbox": [ + 516, + 295, + 831, + 310 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The generalizability of deep neural networks has often been linked to the availability of large training datasets, making sample sparsity one of the fundamental limitations for model performance [1, 3, 8]. MTL addresses sample sparsity by leveraging labeled data across multiple related tasks, effectively expanding the available training data [13, 46]. Research on multi-task generalization predominantly assumes shared knowledge between tasks, where learning across diverse tasks enables more robust and transferable representations [12, 20, 39]. Despite empirical validations of the assumption [40], the fundamental assumption of task relatedness often fails in real-world applications, where competing tasks may vie for model capacity, potentially degrading MTL performance below that of single-task approaches [10, 36].", + "bbox": [ + 516, + 321, + 903, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Prior research has addressed this challenge through two distinct perspectives. One line of work has focused on the model capacity issue, seeking equilibrium between conflicting task objectives. Yang et al. [43] proposes to adaptively reweight per-task training loss, leveraging entropy minimization on unlabeled test samples. Jiang et al. [16] proposes to dynamically merges model branches while optimizing task weights based on the performance on validation set. Simliar reweighing strategies have also been applied in [21, 23]. Alternatively, Chen et al. [6] introduces a selective gradient filtering approach to resolve optimization conflicts, while Dimitriadis et al. [9] and Sener and Koltun [34] propose to formulate the problem from the perspective of Pareto optimality to achieve optimal trade-off between conflicting tasks. In contrast, recent work has focused on the exacerbation of spurious correlations under MTL, where the multi-tasking property could exacerbate reliance on spurious features for each task owing to the interaction between tasks [11, 14, 27]. Specifically, Hu et al. [14] and Fumero et al. [11] address intra-task spurious correlations through distributional invariance, while Makino et al. [27] employs generative modeling of joint distributions to mitigate these correlations.", + "bbox": [ + 516, + 554, + 903, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "25699", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Spurious correlation", + "text_level": 1, + "bbox": [ + 89, + 90, + 284, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Spurious correlation refers to misleading patterns that appear to hold in training data but fail to generalize to new, unseen data [32, 33]. Existing work on spurious correlations typically aims to reduce the error for the worst-performing group, characterized by both the task label and the associated spurious information. typically, these methods can be divided into two categories: data-augmentation, where the training distribution is modified to enhancing worst-group performance [5, 8, 15]; model-regularization, where regularizations are imposed during the training process to eliminate the affect of spurious information [24, 26, 32, 41, 47]. More recently, last-layer retraining has been proposed as an efficient alternative for addressing spurious correlations [17, 18]. In this approach, instead of updating all network parameters, only the classifier layer is fine-tuned on a balanced validation set, significantly reducing computational complexity.", + "bbox": [ + 89, + 112, + 483, + 369 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/89d6d7c033b92f9627bb3834710605157dc10b64834c7d193758bdf5f80afc67.jpg", + "image_caption": [ + "Figure 1. Our framework improves generalizability by identifying and correcting spurious correlations between tasks. For instance, on CelebA dataset, the task labels 'male' and 'attractiveness' are known to be spuriously correlated [26]. Compared with vanilla MTL where predictions for attractiveness falsely correlate with predictions for gender, our method effectively mitigates this bias." + ], + "image_footnote": [], + "bbox": [ + 142, + 378, + 433, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 664, + 228, + 680 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Consider $k$ tasks in total. Let $\\mathcal{D}^{\\mathrm{train}} := \\{(x_i, y_i^1, \\ldots, y_i^k), 1 \\leq i \\leq N\\}$ be the training set where for $i$ -th sample, $x_i \\in \\mathbb{R}^d$ is the input feature and $y_i^j \\in \\mathbb{R}^{d^j}$ is the label for $j$ -th task, let $\\pmb{f}: \\mathbb{R}^d \\to \\mathbb{R}^l$ be the function of shared encoder and $\\pmb{g}^j: \\mathbb{R}^l \\to \\mathbb{R}^{d^j}$ be the function of predictor for $j$ -th task, let $\\mathcal{L}^j$ be the training loss for $j$ -th task, a multi-task learning problem can be formulated as", + "bbox": [ + 89, + 689, + 483, + 813 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {\\boldsymbol {f}, \\boldsymbol {g} ^ {1}, \\dots , \\boldsymbol {g} ^ {k}} {\\arg \\min } \\sum_ {j = 1} ^ {k} w ^ {j} \\bar {\\mathcal {L}} ^ {j} (\\mathcal {D} ^ {\\text {t r a i n}}; \\boldsymbol {g} ^ {j}, \\boldsymbol {f}),\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 819, + 400, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\bar{\\mathcal{L}}^j (\\mathcal{D}^{\\mathrm{train}};\\pmb {g}^j,\\pmb {f})\\coloneqq \\frac{1}{N}\\sum_{i = 1}^N\\mathcal{L}^j (\\pmb {g}^j (\\pmb {f}(x_i)),y_i^j)$ is the average training loss for $j$ -th task, and $w^{j}$ specifies", + "bbox": [ + 89, + 867, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the weight of $\\bar{\\mathcal{L}}^j$ during optimization. Despite the benefit of exploring task-relatedness, which are expected to benefit the generalizability, recent work on MTL has pointed to the potential risk of spurious correlations between tasks owing to the confounder between task labels [14, 27]. Let $\\{\\mathcal{F}_i, 1 \\leq i \\leq K\\}$ be the set of underlying latent features such that $\\mathcal{F}_i \\perp \\mathcal{F}_j$ , $\\forall i \\neq j$ , let $\\mathcal{F}^j$ be the set of underlying latent features that are causally related with $j$ -th task, when $\\mathcal{F}^i \\neq \\mathcal{F}^j$ , optimizing $\\bar{\\mathcal{L}}^i$ and $\\bar{\\mathcal{L}}^j$ jointly could exacerbate reliance on spurious features from $\\mathcal{F}^j$ for $g^i$ , as the shared encoder needs to explicitly encode knowledge for both tasks. Such confounder brings in spurious correlations between tasks and can harm generalization performance.", + "bbox": [ + 511, + 90, + 903, + 286 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Nonetheless, compared with conventional formulations where the spurious information is assumed to be available on training data [32, 33], regarding MTL, we do not have the spurious information at hand, where each task may consider all other tasks as being either spurious or beneficial. Alternatively, most existing work [11, 14] has formulated the problem from the perspective of distributional robustness, where the model is expected to perform consistently on both the training data and the validation data $\\mathcal{D}^{\\mathrm{val}}$ under some unknown covariate shift:", + "bbox": [ + 511, + 287, + 906, + 436 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\underset {\\boldsymbol {f}, \\boldsymbol {g} ^ {1}, \\dots , \\boldsymbol {g} ^ {k}} {\\arg \\min } \\sum_ {j = 1} ^ {k} w ^ {j} \\bar {\\mathcal {L}} ^ {j} (\\mathcal {D} ^ {\\text {t r a i n}}; \\boldsymbol {g} ^ {j}, \\boldsymbol {f}) + \\sum_ {j = 1} ^ {k} w ^ {\\text {v a l}, j} \\bar {\\mathcal {L}} ^ {j} (\\mathcal {D} ^ {\\text {v a l}}; \\boldsymbol {g} ^ {j}, \\boldsymbol {f}), \\\\ s. t. R \\left(\\mathcal {D} ^ {\\text {t r a i n}}; f\\right) \\leq \\epsilon , R \\left(\\mathcal {D} ^ {\\text {v a l}}; f\\right) \\leq \\epsilon , \\tag {1} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 441, + 916, + 517 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $w^{\\mathrm{val},j}$ specifies the weight of validation loss for $j$ -th task, and $R(\\mathcal{D}^{\\mathrm{train}};f)$ and $R(\\mathcal{D}^{\\mathrm{val}};f)$ correspond to the regularization on $f$ such that the encoder is dimensional-wise disentangled. For example, MT-CRL [14] employs covariance between latent dimensions as the regularization, while Meta-learning [11] formulates $R(\\mathcal{D}^{\\mathrm{train}};f)$ as the entropy of normalized representations per dimension. Eq. (1) stems from the assumption that for $i$ -th task, the covariate shift on validation data effectively disrupts the correlation between its spurious features and its task labels, which could be infeasible in general. On one hand, without prior knowledge on the task-specific causal features, there's no guarantee that the covariate shift breaks the spurious correlation between tasks. Furthermore, the effectiveness of Eq. (1) may be influenced by the degree of bias in the validation set, introducing uncertainty in the mitigation spurious correlations.", + "bbox": [ + 511, + 518, + 905, + 772 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We use the following experiments on linear SEM [31] to validate the two perspectives. Following existing work [14], we construct two binary classification tasks where the causal features of each task are independent and follow a Gaussian distribution. Let $Y^{j}$ be the random variable of labels for $j$ -th task, we set the ratios between task labels on training data as $\\alpha_{y}^{\\text{train}} = \\frac{P(Y^{0} = y, Y^{1} = (1 - y))}{P(Y^{0} = y, Y^{1} = y)} = \\frac{1}{9}$ , $\\forall y \\in \\{0, 1\\}$ on training data. Regarding both tasks, we have", + "bbox": [ + 511, + 773, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "25700", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$P(Y^{j} = 0) = P(Y^{j} = 1) = 0.5, \\forall j \\in \\{0,1\\}$ , and the testing data is balanced sampled across each co-couurance of task labels such that $\\alpha_{y}^{\\mathrm{test}} = \\frac{P(Y^{0} = y, Y^{1} = (1 - y))}{P(Y^{0} = y, Y^{1} = y)} = 1$ , $\\forall y \\in \\{0,1\\}$ . Smaller $\\alpha_{y}$ indicates higher correlation between the two task labels.", + "bbox": [ + 89, + 90, + 483, + 172 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/57a342d01f5bd3767d4bd32dbfd1a473d84bf9d22973de2f0c49781cbcc019b9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TaskMTL (αtrainy = 1/9)MTL (αtrainy = 1)
Task 081.32±0.29%94.47±0.34%
Task 181.37±0.41%94.47±0.35%
", + "bbox": [ + 120, + 179, + 455, + 231 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/2b57347b9f224de43f6d9e45d74c46b49ae6cba18052cf852cc427b6eec192b2.jpg", + "table_caption": [ + "Table 1. Experimental results on synthetic data. We choose classification accuracy as the evaluation metric for both tasks. Under high correlation, the model performs remarkably worse than that under no spurious correlation." + ], + "table_footnote": [], + "table_body": "
TaskMT-CRL (αyval = 1/9)Meta-learning
Task 081.59±0.36%81.73±0.35%
Task 181.87±0.95%81.67±0.44%
", + "bbox": [ + 122, + 323, + 455, + 375 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/1305b7510dc94ad24754e0455560a01d4d6f4a8fb219527efe7bc49c1154c504.jpg", + "table_caption": [ + "Table 2. Experimental results on synthetic data. Both methods show marginal improvement in testing accuracy compared with MTL ( $\\alpha_{y}^{\\mathrm{train}} = \\frac{1}{9}$ ): $81.32 \\pm 0.29\\%$ for task 0 and $81.37 \\pm 0.41\\%$ for task 1, owing to the consistency in $\\alpha_{y}$ 's between training and validation data." + ], + "table_footnote": [], + "table_body": "
TaskMT-CRL (αyval = 1/5)MT-CRL (αyval = 1)
Task 083.74±0.65%87.59±0.74%
Task 183.32±0.59%88.16±0.46%
", + "bbox": [ + 102, + 479, + 470, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 3. Experimental results on synthetic data. The testing performance improves as $\\alpha_{y}^{\\mathrm{val}}$ 's decrease, substantiating the importance of mitigating bias on the validation data. Nonetheless, given the balanced validation set, MT-CRL still performs worse than the optimal baseline MTL ( $\\alpha_{y}^{\\mathrm{train}} = 1$ ): $94.47 \\pm 0.34\\%$ for task 0 and $94.47 \\pm 0.35\\%$ for task 1, highlighting the insufficiency in formulating spurious correlations via distributional robustness.", + "bbox": [ + 89, + 542, + 483, + 640 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We show results of under different constructions of validation sets in Tab. 1 - 2. Regarding MT-CRL [14], as the validation set is only assumed to contain unknown shift, we consider three different types of covariate shifts: The training-validation split where the validation set is sampled from the training distribution, as suggested by the authors as a general construction of covariate shift, where $\\alpha_{y}$ 's remain the same on both training and validation data; A less biased validation set where $\\alpha_{y}^{\\mathrm{val}} = \\frac{1}{5}, \\forall y \\in \\{0,1\\}$ ; A balanced validation set where $\\alpha_{y}^{\\mathrm{val}} = 1, \\forall y \\in \\{0,1\\}$ . While a balanced validation set shows improvement in the performance on testing data, the first two constructions of validation data show marginal improvement compared with vanilla MTL, validating our claim that the effectiveness of distributional invariance can be subject to the bias on validation data. Nonetheless, in reality, it is hard to obtain or construct a", + "bbox": [ + 89, + 657, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "perfectly balanced data owing to the uncertainty in the spurious correlations between tasks. Furthermore, it is possible that certain labeling pairs are not observed on training or validation data, leading to infeasibility of constructing such balanced data.", + "bbox": [ + 511, + 90, + 905, + 165 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Regarding Meta-learning [11], since the validation set is constructed per task such that for $j$ -th task, $P(Y^{j} = 0) = P(Y^{j} = 1) = 0.5$ , $\\forall j \\in \\{0,1\\}$ , we have $\\alpha_{y}$ 's remain the same on both training and validation data. As a result, such construction fails to show improvement in mitigating spurious information compared with the ERM baseline. This demonstrates the flaws in formulating the covariate shift via class-balance: the spurious correlation between tasks does not necessarily lead to imbalance in the marginal distribution of labels per task, where using class-balanced resampling can lead to marginal variations in the conditional distributions $P(Y_{i}|Y_{j})$ of task labels.", + "bbox": [ + 511, + 167, + 906, + 348 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 513, + 361, + 604, + 376 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Identification of spurious correlation between tasks", + "text_level": 1, + "bbox": [ + 511, + 386, + 905, + 416 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our preliminary discussion suggests the necessity of identifying spurious correlations between tasks. In light of the intractability in identifying the latent causal features for each task, particularly regarding high-dimensional vision data, we instead consider modeling the spurious correlations from the label perspective. Existing work [14, 26] has demonstrated that when two tasks are spuriously correlated, their corresponding task labels will be disproportionately paired, leading to an uneven joint distribution across task labels. Specifically, let $\\mathbb{S}^j$ be the set of values of labels for $j$ -th task, when spurious correlation exists between $i$ -th and $j$ -th tasks, there exist $y, y' \\in \\mathbb{S}^i$ and $\\tilde{y}, \\tilde{y}' \\in \\mathbb{S}^j$ , such that the following inequality regarding label distribution holds:", + "bbox": [ + 511, + 422, + 906, + 619 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP (Y ^ {i} = y, Y ^ {j} = \\tilde {y}) > P (Y ^ {i} = y ^ {\\prime}, Y ^ {j} = \\tilde {y} ^ {\\prime}). \\quad (2)\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 630, + 906, + 650 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To quantitatively evaluate Eq. (2), we observe that the correlation between task labels can be quantified using the Pearson correlation coefficient [30]:", + "bbox": [ + 511, + 660, + 905, + 705 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\rho_ {Y ^ {i}, Y ^ {j}} = \\frac {\\mathbb {E} [ Y ^ {i} Y ^ {j} ] - \\mathbb {E} [ Y ^ {i} ] \\mathbb {E} [ Y ^ {j} ]}{\\sqrt {\\mathbb {E} [ (Y ^ {i}) ^ {2} ] - (\\mathbb {E} [ Y ^ {i} ]) ^ {2}} \\sqrt {\\mathbb {E} [ (Y ^ {j}) ^ {2} ] - (\\mathbb {E} [ Y ^ {j} ]) ^ {2}}}.\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 715, + 906, + 753 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For simplicity of discussion, consider both tasks as binary classification where $Y^{i}$ and $Y^{j}$ follow Bernoulli distributions with expectations $p^{i}$ and $p^{j}$ , respectively. We refer to Algorithm 1 for extension to multi-class scenario. The following lemma states the close-form expression of $\\rho_{Y^{i},Y^{j}}$ :", + "bbox": [ + 511, + 762, + 906, + 853 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Lemma 1. Let $\\alpha_{y}^{ij} := \\frac{P(Y^i = y, Y^j = 0)}{P(Y^i = y, Y^j = 1)}$ be the ratio of joint probabilities of $Y^i$ and $Y^j$ evaluated w.r.t. class $y$ of $i$ -th", + "bbox": [ + 511, + 864, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "25701", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/78528c4fe56ed138b42b8dff1ffcff491b6b4a70bfcf4e82965c64fa2dd36c37.jpg", + "image_caption": [ + "Figure 2. Demonstration of our debiased adversarial training framework. For each task, the predictors within the blue rectangle represent those identified as spurious by our method. During optimization, the per-task projector and predictor are updated to ensure that the projected representation minimizes reliance on information of spurious tasks while maintaining performance on the target task." + ], + "image_footnote": [], + "bbox": [ + 277, + 89, + 743, + 354 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "task, we have $\\rho_{Y^i,Y^j}$ as", + "bbox": [ + 89, + 436, + 251, + 453 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\rho_ {Y ^ {i}, Y ^ {j}} = \\frac {\\left(\\frac {1}{(1 + \\alpha_ {1} ^ {i j})} - p ^ {j}\\right) p ^ {i}}{\\sqrt {p ^ {i} p ^ {j} (1 - p ^ {i}) (1 - p ^ {j})}}.\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 455, + 406, + 503 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We defer full proof to the Appendix. As shown in Lemma 1, the correlation coefficient between $Y^{i}$ and $Y^{j}$ is determined by the marginal distributions of $Y^{i}$ and $Y^{j}$ and the ratio of joint probabilities. However, the dependence of $\\rho_{Y^i,Y^j}$ on the marginal distributions could introduce potential errors in identifying spurious correlations, particularly in the presence of label imbalance, where the disproportionality between labeling pairs could be overlooked given small marginal probabilities of task labels. For example, even with a small $\\alpha_{1}^{ij}$ , when $p^i$ becomes trivial due to class imbalance, we may observe $\\rho_{Y^i,Y^j}\\approx 0$ , despite the presence of spurious correlations between the two tasks. We therefore consider the following class-wise resampling strategy to perturb the marginal distributions of task labels, such that when $p^i$ and $p^j$ change, we have $\\rho_{Y^i,Y^j}$ change, which enables us to recognize the spurious correlation between $Y^{i}$ and $Y^{j}$ while mitigating the influence of label imbalance. Specifically, let $Y^{\\prime i}$ and $Y^{\\prime j}$ be the random variable of task labels after resampling, let $\\mathcal{D}_{Y^j\\neq y}^{\\mathrm{train}}\\coloneqq \\{(x_i,y_i^1,\\dots ,y_i^k)|y_i^j\\neq y\\}$ be the subset of samples excluding class $y$ of $j$ -th task and $\\mathcal{D}_{Y^{\\prime j}\\neq y}^{\\mathrm{train}}$ be the corresponding subset after resampling, we consider the following down-sampling scheme within $\\mathcal{D}_{Y^j\\neq y}^{\\mathrm{train}}$ :", + "bbox": [ + 91, + 516, + 483, + 875 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left| \\mathcal {D} _ {Y ^ {\\prime j} \\neq y} ^ {' \\mathrm {t r a i n}} \\right| = \\xi \\left| \\mathcal {D} _ {Y ^ {j} \\neq y} ^ {\\mathrm {t r a i n}} \\right|,\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 877, + 370, + 905 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\xi < 1$ is the constant that determines the degree of resampling, and $\\mathcal{D}_{Y^{\\prime j}\\neq y}^{\\prime \\mathrm{train}}$ is constructed by random sampling. Under class-wise resampling, we intend to amplify the effect of correlations between task labels under different values of $Y^{\\prime j}$ within $\\mathbb{S}^j$ , so as to properly capture the dependence between task labels. Let $p^{\\prime i}$ be the expectation of $Y^{\\prime i}$ evaluated on the resampled set $\\mathcal{D}_{Y^{\\prime j}\\neq y}^{\\prime \\mathrm{train}}\\cup \\left(\\mathcal{D}^{\\mathrm{train}}\\backslash \\mathcal{D}_{Y^{\\prime j}\\neq y}^{\\mathrm{train}}\\right)$ , the following theorem states the feasibility of identifying spurious correlations between tasks based on the difference in correlation coefficients before and after resampling:", + "bbox": [ + 511, + 435, + 906, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Theorem 1. When spurious correlation exists between two tasks, the maximum difference in correlation coefficient before and after resampling is lower-bounded by the difference in expectations $p^i - p'^i$ , up to multiplicative constants:", + "bbox": [ + 511, + 603, + 906, + 666 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left| \\rho_ {Y ^ {i}, Y ^ {j}} - \\rho_ {Y ^ {\\prime i}, Y ^ {\\prime j}} \\right| \\geq \\mathcal {K} \\left| p ^ {' i} - p ^ {i} \\right|,\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 671, + 831, + 700 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nw h e r e \\mathcal {K} = \\frac {\\left(\\frac {1}{\\hat {\\alpha} _ {1} ^ {i j}} - \\hat {\\alpha} _ {1} ^ {i j}\\right)}{\\sqrt {(1 - p ^ {i}) (1 - (\\hat {\\alpha} _ {1} ^ {i j} - \\hat {\\alpha} _ {0} ^ {i j}) p ^ {i} - \\hat {\\alpha} _ {0} ^ {i j})}} a n d \\hat {\\alpha} _ {y} ^ {i j} = \\frac {1}{1 + \\alpha_ {y} ^ {i j}}.\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 707, + 898, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We defer full proof to the Appendix. As shown in Theorem 1, since $\\hat{\\alpha}_{y}^{ij}$ and $p^i$ remain constant before and after resampling, we have the lower-bound of difference in correlation coefficient before and after resampling monotonically increase as $p^{'i}$ increases. Accordingly, even though $p^i$ can be trivial, so long as $p^i$ and $p^{'i}$ differ significantly before and after resampling, we are guaranteed a substantial change in the correlation coefficient, which enables us to identify the spurious correlations among tasks. Furthermore, the following remark states the difference in correla", + "bbox": [ + 511, + 750, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "25702", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tion coefficient when spurious correlations do not exist between tasks:", + "bbox": [ + 89, + 90, + 482, + 119 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Remark 1. When two tasks are not spuriously correlated, we have $\\rho_{Y^i,Y^j} - \\rho_{Y'^i,Y'^j} = 0$ .", + "bbox": [ + 89, + 130, + 482, + 162 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Consequently, let $\\mathbb{Z}^j$ be the set of indices of identified spurious tasks for $j$ -th task, we have", + "bbox": [ + 89, + 171, + 483, + 202 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\left| \\rho_ {Y ^ {i}, Y ^ {j}} - \\rho_ {Y ^ {\\prime i}, Y ^ {\\prime j}} \\right| > \\delta , \\forall i \\in \\mathbb {Z} ^ {j},\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 210, + 403, + 238 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\delta$ is the threshold for distinguishing the spurious correlations between tasks. In practice, $\\delta$ can be trivially obtained owing to the distinctions in $\\left|\\rho_{Y^i,Y^j} - \\rho_{Y'^i,Y'^j}\\right|$ as stated in Theorem 1 and Remark 1. We refer to Sec. 5.3 for the detailed discussion.", + "bbox": [ + 89, + 247, + 483, + 328 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Debiased Adversarial Training", + "text_level": 1, + "bbox": [ + 89, + 338, + 362, + 354 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We move on to discuss the mitigation of spurious correlation between tasks. Due to the potential absence of certain label pairs in the training set, which impedes constructing balanced data even after identifying spurious tasks, we instead aim to ensure that the task-specific predictors effectively discard non-causal features for each task. In light of the multi-tasking property, we can readily formulate the objective as an adversarial learning problem w.r.t. the pertask predictors. Let $\\hat{f}$ be the function of pre-trained encoder which remains frozen during fine-tuning, regarding $j$ -th task, given its identified set of spurious tasks $\\mathbb{Z}^j$ , to utilize the per-task predictors as discriminators of non-causal features, we apply one extra projector $h^j: \\mathbb{R}^l \\to \\mathbb{R}^l$ before $g^j$ , such that the prediction for $j$ -th task is given by $g^i \\circ h^i \\circ \\hat{f}(x)$ . We consider the following objective for mitigating spurious correlations regarding $j$ -th task:", + "bbox": [ + 89, + 359, + 483, + 602 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname *{arg min}_{\\boldsymbol{h}^{j},\\boldsymbol{g}^{j}}\\bar{\\mathcal{L}}^{j}\\left(\\mathcal{D}^{\\text{train}};\\boldsymbol{h}^{j},\\boldsymbol{g}^{j}\\right) - \\sum_{t\\in \\mathbb{Z}^{j}}\\lambda^{j}\\bar{\\mathcal{L}}^{t}\\left(\\mathcal{D}^{\\text{train}};\\boldsymbol{h}^{j},\\boldsymbol{g}^{t}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 625, + 465, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where the optimization is performed over $\\pmb{h}^j, \\pmb{g}^j$ , while $\\pmb{g}^t$ remains frozen $\\forall t \\in \\mathbb{Z}^j$ , and $\\bar{\\mathcal{L}}^j(\\mathcal{D}^{\\mathrm{train}}; \\pmb{h}^j, \\pmb{g}^j) := \\frac{1}{N} \\sum_{i=1}^{N} \\mathcal{L}^j(\\pmb{g}^j(\\pmb{h}^j(\\hat{\\pmb{f}}(x_i))), y_i^j), \\quad \\bar{\\mathcal{L}}^t(\\mathcal{D}^{\\mathrm{train}}; \\pmb{h}^j, \\pmb{g}^t) := \\frac{1}{N} \\sum_{i=1}^{N} \\mathcal{L}^k(\\pmb{g}^k(\\pmb{h}^j(\\pmb{f}(x_i))), y_i^k)$ are the average training loss under projection. By minimizing the objective, we intend to ensure that the projected latent representation retains only the information pertinent to $j$ -th task while shedding the spurious information. We summarize our method in Algorithm 1.", + "bbox": [ + 89, + 669, + 483, + 810 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 823, + 223, + 840 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 848, + 281, + 864 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We validate our method on the following benchmark datasets: Multi-MNIST dataset [38], CelebA dataset [25],", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Algorithm 1 Debiased adversarial training", + "bbox": [ + 514, + 90, + 799, + 106 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Input: Pre-trained encoder $\\hat{f}$ , task-specific predictors $g^{j}$ 's s.t. $1 \\leq j \\leq k$ , task-specific projectors $h^{j}$ 's, training data $\\mathcal{D}^{\\mathrm{train}}$ , training epochs $E$ , learning rate $\\eta$ , threshold $\\delta$ , resampling constant $\\xi$ , hyper-parameter $\\lambda^{j}$ .", + "bbox": [ + 511, + 122, + 901, + 185 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: for $j = 1$ to $k$ do:", + "2: $\\mathbb{Z}^j = \\emptyset$", + "3: for $i = 1$ to $k$ s.t. $i \\neq j$ do:", + "4: Calculate $\\rho_{Y^i,Y^j}$ on $\\mathcal{D}^{\\mathrm{train}}$ .", + "5: for $y\\in \\mathbb{S}^j$ do:", + "6: Perform class-wise resampling:" + ], + "bbox": [ + 521, + 195, + 823, + 287 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\left| \\mathcal {D} _ {Y ^ {\\prime j} \\neq y} ^ {\\text {t r a i n}} \\right| = \\xi \\left| \\mathcal {D} _ {Y ^ {j} \\neq y} ^ {\\text {t r a i n}} \\right|.\n$$\n", + "text_format": "latex", + "bbox": [ + 640, + 297, + 805, + 324 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "7: Calculate $\\rho_{Y^{\\prime}i,Y^{\\prime}j}$ on", + "bbox": [ + 522, + 334, + 759, + 349 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {D} _ {Y ^ {\\prime j} \\neq y} ^ {\\text {t r a i n}} \\cup \\left(\\mathcal {D} ^ {\\text {t r a i n}} \\backslash \\mathcal {D} _ {Y ^ {j} \\neq y} ^ {\\text {t r a i n}}\\right).\n$$\n", + "text_format": "latex", + "bbox": [ + 627, + 362, + 818, + 388 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "8: if $\\rho_{Y^i,Y^j} - \\rho_{Y'^i,Y'^j}\\geq \\delta$ and $i\\notin \\mathbb{Z}^j$ then:", + "9: $\\mathbb{Z}^j = \\mathbb{Z}^j\\cup \\{i\\}$", + "10: end if", + "11: end for", + "12: end for", + "13: end for", + "14: for epoch $= 1$ to $E$ do:", + "15: for $\\mathbf{j} = 1$ to $k$ do:", + "16: Training loss:" + ], + "bbox": [ + 519, + 398, + 895, + 537 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\bar {\\mathcal {L}} ^ {j} \\left(\\mathcal {D} ^ {\\text {t r a i n}}; \\boldsymbol {h} ^ {j}, \\boldsymbol {g} ^ {j}\\right) - \\sum_ {t \\in \\mathbb {Z} ^ {j}} \\lambda^ {j} \\bar {\\mathcal {L}} ^ {t} \\left(\\mathcal {D} ^ {\\text {t r a i n}}; \\boldsymbol {h} ^ {j}, \\boldsymbol {g} ^ {t}\\right).\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 547, + 888, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "17: Update $h^j$ and $g^j$ by gradient descent:", + "bbox": [ + 517, + 593, + 846, + 609 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {h} ^ {j} = \\boldsymbol {h} ^ {j} - \\eta \\frac {\\partial \\mathcal {L}}{\\partial \\boldsymbol {h} ^ {j}}.\n$$\n", + "text_format": "latex", + "bbox": [ + 658, + 619, + 785, + 650 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {g} ^ {j} = \\boldsymbol {g} ^ {j} - \\eta \\frac {\\partial \\mathcal {L}}{\\partial \\boldsymbol {g} ^ {j}}.\n$$\n", + "text_format": "latex", + "bbox": [ + 661, + 657, + 782, + 691 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "18: end for", + "19: end for" + ], + "bbox": [ + 517, + 696, + 620, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Taskonomy dataset [45], MetaShift dataset [22], NYU-V2 dataset [35] and CityScape dataset [7]. We implement our method in PyTorch 2.0.0 on one RTX-3090 GPU1. We set $\\xi = 0.1$ during experiments and evaluate the difference in correlation coefficients between training set and class-wise resampled sets to construct $\\mathbb{Z}^j$ . We use accuracy as evaluation metric for classification tasks, cross-entropy loss and", + "bbox": [ + 511, + 755, + 906, + 862 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "$^{1}$ Code available at https://github.com/cjy24/Spurious-Correlation-MTL.", + "bbox": [ + 511, + 875, + 903, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "25703", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/efdd7f108465061e0e95b42ae60a670c281c753e5c344b0bce9657f3c3edeb87.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TaskSTLSubSelMT-CRLMeta-learningGMTLOurs
left digit (accuracy, ↑)+4.9%+3.4%+5.1%+0.2%+5.9%+9.3%
right digit (accuracy, ↑)+5.9%+2.9%+6.3%+0.3%+7.2%+9.6%
", + "bbox": [ + 184, + 88, + 812, + 138 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5003814a0fc6c32815285bf67a998140505277751c3d14e4503e9fe5e9ee332b.jpg", + "table_caption": [ + "Table 4. Experimental results on Multi-MNIST dataset. The relative improvement of each method is calculated in comparison to vanilla MTL." + ], + "table_footnote": [], + "table_body": "
TaskSTLSubSelMT-CRLMeta-learningGMTLOurs
Gender (accuracy, ↑)+2.2%+1.7%+1.3%+2.4%+2.6%+5.7%
Smiling (accuracy, ↑)+2.3%+2.5%+2.6%+3.4%+1.2%+4.6%
Age (accuracy, ↑)+1.2%+2.2%+0.1%+1.1%+1.5%+3.9%
Attractiveness (accuracy, ↑)+0.6%+0.4%+0.7%+0.2%+1.2%+2.4%
", + "bbox": [ + 171, + 189, + 826, + 271 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5a5cd60c3d51ff3483df878e47030fc558876241618463936db0159292f53bdf.jpg", + "table_caption": [ + "Table 5. Experimental results on CelebA dataset." + ], + "table_footnote": [], + "table_body": "
TaskSTLSubSelMT-CRLMeta-learningOurs
Scene (accuracy, ↑)+4.1%+3.2%+5.8%+5.3%+9.9%
Object (accuracy, ↑)+5.2%+4.1%+7.3%+8.1%+12.7%
Segmentation (cross entropy, ↓)-5.6%+1.7%+3.1%+6.9%+6.2%
", + "bbox": [ + 187, + 308, + 810, + 375 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0489d7171a07643534ead434fa4deb96b6e14fba597e1431c53fca7d52e4999d.jpg", + "table_caption": [ + "Table 6. Experimental results on Taskonomy dataset." + ], + "table_footnote": [], + "table_body": "
TaskSTLSubSelMT-CRLMeta-learningGMTLOurs
Animal (accuracy, ↑)+4.2%+2.7%+3.2%+1.5%+0.2%+5.4%
Scene (accuracy, ↑)+3.1%+2.1%+1.6%+0.1%+1.7%+4.0%
", + "bbox": [ + 192, + 411, + 805, + 462 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f4265470a7954471c3370878f7d374558aecea592b980dc3e69a6397ffad8f87.jpg", + "table_caption": [ + "Table 7. Experimental results on MetaShift dataset." + ], + "table_footnote": [], + "table_body": "
TaskSTLSubSelMT-CRLMeta-learningOurs
Segmentation (mean IoU, ↑)+2.7%+4.2%+6.9%+7.2%+11.3%
Depth (relative error, ↓)+3.1%+3.8%+7.5%+6.2%+10.2%
", + "bbox": [ + 197, + 500, + 799, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 8. Experimental results on NYU-V2 dataset.", + "bbox": [ + 346, + 560, + 650, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "mean IoU for segmentation tasks and relative error for depth estimation tasks. We defer details of each dataset and full results to the appendix.", + "bbox": [ + 89, + 601, + 482, + 647 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare our method with the following related methods on spurious correlation and generalization in MTL: Single-task learning (STL): the network is optimized over each task only without regularizations; MTL: vanilla MTL without regularizations; SubSel [21]: MTL with subset selection to address task conflicts; MT-CRL [14]: MTL with disentangled neural modules and invariant regularization on task-to-module routing; Meta-learning [11]: MTL with disentangled representation learning and bi-level optimization on per-task predictors; GMTL [27]: MTL with generative modeling of task labels.", + "bbox": [ + 89, + 652, + 482, + 819 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We repeat experiments on each dataset three times and report the average results. For MT-CRL [14], the covariate shift on training data is constructed by training-validation split, as suggested by the authors. Further, we do not include GMTL [27] on Taskonomy, NYU-V2 and CityScape", + "bbox": [ + 89, + 825, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "dataset, as it is infeasible to estimate the joint distribution involving high-dimensional masks. Following prior work [14], we report the relative performance improvement of each method over vanilla MTL for evaluation, as the scales of evaluation metrics differ significantly across tasks.", + "bbox": [ + 511, + 601, + 906, + 678 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Experimental Results", + "text_level": 1, + "bbox": [ + 511, + 694, + 718, + 710 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results are shown in Tab. 4 - 8. Compared with vanilla STL and MTL, Meta-learning [11] shows marginal improvement in testing performance, particularly when the marginal training distributions are evenly distributed per task. Regarding MT-CRL [14], the training-validation split does not appropriately capture the spurious correlations between tasks, where the method shows limited improvement on several tasks. Such observations are in line with our preliminary analysis: modeling spurious correlation between tasks using covariate shift is in general insufficient and lacks theoretical guarantee. Furthermore, GMTL [27] is limited to classification tasks, and its dependence on estimating", + "bbox": [ + 511, + 719, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "25704", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "joint distributions may introduce estimation errors, resulting in subpar performance on testing data. In comparison, our method shows better or comparable performance on all four datasets, validating the effectiveness of our method.", + "bbox": [ + 88, + 90, + 480, + 151 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Empirical Validation of Theoretical Results", + "text_level": 1, + "bbox": [ + 89, + 162, + 460, + 179 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We move on to discuss the validation of Theorem 1. Specifically, we consider three different variations of Multi-MNIST dataset: Spurious correlation (Sp.C.), with non-overlapping labeling pair between training and testing data, as in Sec. 5.1; Random assignment (R.A.), where the right digit is randomly assigned given the left digit on both training and testing data; Identical assignment (Id.A.), where the left digit is identical to the right digit on both training and testing data. Under identical assignment, the two tasks are no longer spuriously correlated.", + "bbox": [ + 89, + 184, + 483, + 335 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/102e74a9d2ed42b6a91bdb0f3e46cc15a63d4a7abf749e9fcfb082e4a141bea7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DataρYl,Yr - ρY′l,Y′r
Sp.C.0.29 ± 0.04
R.A.0.002 ± 0.0006
Id.A.0.001 ± 0.0003
", + "bbox": [ + 181, + 348, + 393, + 421 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Under each variation, we perform resampling on the left digit and report the maximum average change in correlation coefficient. The results are shown in Tab. 9. Compared with those under random assignment and identical assignment, when the two tasks are spuriously correlated, the difference in correlation coefficient shows a significant increase, substantiating the feasibility of identifying spurious correlations between tasks based on the difference in correlation coefficients, as proposed in Theorem 1.", + "bbox": [ + 89, + 520, + 483, + 656 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 666, + 243, + 683 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We further invest the effect of spurious correlation identification within our methodology under the three variations of Multi-MNIST dataset, and the necessity in distinguishing spurious correlation between tasks. Results are shown in Tab. 10. Under both spurious correlation and random assignment, the adversarial training framework shows similar performance, regardless of the identification of spurious correlations. Specifically, under random assignment, our framework demonstrates marginal improvement over STL and MTL baselines, irrespective of the construction of $\\mathbb{Z}^j$ ’s. This suggests that when the labels of two tasks are not correlated, the per-task predictors are also unlikely to be affected by the spurious features. Furthermore, under identical assignment, without correctly identifying the spurious", + "bbox": [ + 89, + 688, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c071c3f35470f71801473b13af357c95a3286ce63d48de58d840f9c5e23d81f5.jpg", + "table_caption": [ + "Table 9. Maximum average change in correlation coefficient under class-wise resampling on Multi-MNIST dataset. $Y^{l}$ and $Y^{r}$ refer to the random variables of left digit label and right digit label on training set, and $Y^{\\prime l}$ and $Y^{\\prime r}$ refer to the random variables of left digit label and right digit label on the resampled training set." + ], + "table_footnote": [], + "table_body": "
DataMethodleft digitright digit
Sp.C.STL88.26±1.34%88.43±1.54%
Sp.C.MTL84.14±1.19%83.57±1.27%
Sp.C.Ours (w/o Zj)91.59±1.47%90.74±1.22%
Sp.C.Ours91.63±1.46%90.78±1.26%
R.A.STL93.26±1.46%92.87±1.25%
R.A.MTL93.44±1.57%93.67±1.28%
R.A.Ours (w/o Zj)93.48±1.76%93.25±1.29%
R.A.Ours93.87±1.24%93.73±1.36%
Id.A.STL93.67±1.36%93.68±1.37%
Id.A.MTL93.81±1.55%93.49±1.32%
Id.A.Ours (w/o Zj)91.73±1.45%91.85±1.46%
Id.A.Ours93.84±1.43%93.52±1.19%
", + "bbox": [ + 514, + 87, + 915, + 290 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 10. Experimental results on Multi-MNIST dataset under random and identical assignments. Regarding 'Ours $(\\mathrm{w / o}\\mathbb{Z}^j)$ we apply the debiased adversarial training per-task by taking the other task as the spurious task. Particularly, under identical assignment, applying debiased adversarial training without identifying $\\mathbb{Z}^j$ ’s leads to slightly worse performance than the baselines, indicating the necessity of distinguishing the spurious correlations between tasks. In comparison, our method performs comparably with the baselines, owing to the identification of spurious correlation between the two tasks.", + "bbox": [ + 511, + 300, + 906, + 439 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "information, the adversarial training framework performs slightly worse than the STL and MTL baselines, underscoring the importance of correctly constructing $\\mathbb{Z}^j$ 's, within which our debiased adversarial training framework shall be applied to improve, rather than hinder, the generalizability.", + "bbox": [ + 511, + 467, + 906, + 544 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 556, + 633, + 573 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Spurious correlations in MTL present a significant but inadequately addressed challenge. In this paper, we experimentally show the limitations of existing formulations in mitigating spurious correlations. Motivated by the discussion, we propose a novel framework for identifying spurious correlations between tasks. Our approach formulates the problem from a label-based perspective, where we propose to quantify the spurious correlations between tasks based on the difference in correlation coefficients, evaluated on both the training data and the class-wise resampled training data. We show theoretically the feasibility of our framework in distinguishing spurious correlations between tasks. Based on our identification framework, we propose an adversarial training framework for mitigating between-task spurious correlations, which fine-tunes only the task-specific predictors to ensure that task-specific predictors sufficiently discard the spurious information. We validate from experiments on four benchmark datasets that our method outperforms existing baselines in mitigating spurious correlations between tasks, as well as the effectiveness of our method in identifying spurious correlations between tasks.", + "bbox": [ + 511, + 583, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "25705", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was partially supported by the EMBRIO Institute, contract #2120200, a National Science Foundation (NSF) Biology Integration Institute, and NSF IIS #1955890, IIS #2146091, IIS #2345235.", + "bbox": [ + 89, + 114, + 483, + 176 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 191, + 187, + 208 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Ibrahim M Alabdulmohsin, Behnam Neyshabur, and Xiaohua Zhai. Revisiting neural scaling laws in language and vision. Advances in Neural Information Processing Systems, 35:22300-22312, 2022. 2", + "[2] Simon Buchholz, Goutham Rajendran, Elan Rosenfeld, Bryon Aragam, Bernhard Schölkopf, and Pradeep Ravikumar. Learning linear causal representations from interventions under general nonlinear mixing. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[3] Yuan Cao and Quanquan Gu. Generalization bounds of stochastic gradient descent for wide and deep neural networks. Advances in neural information processing systems, 32, 2019. 2", + "[4] Rich Caruana. Multitask learning. Machine learning, 28: 41-75, 1997. 1", + "[5] Yimeng Chen, Ruibin Xiong, Zhi-Ming Ma, and Yanyan Lan. When does group invariant learning survive spurious correlations? Advances in Neural Information Processing Systems, 35:7038-7051, 2022. 2, 3", + "[6] Zhao Chen, Jiquan Ngiam, Yanping Huang, Thang Luong, Henrik Kretzschmar, Yuning Chai, and Dragomir Anguelov. Just pick a sign: Optimizing deep multitask models with gradient sign dropout. Advances in Neural Information Processing Systems, 33:2039-2050, 2020. 2", + "[7] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 6, 1", + "[8] Yuyang Deng, Junyuan Hong, Jiayu Zhou, and Mehrdad Mahdavi. On the generalization ability of unsupervised pretraining. In International Conference on Artificial Intelligence and Statistics, pages 4519-4527. PMLR, 2024. 2, 3", + "[9] Nikolaos Dimitriadis, Pascal Frossard, and François Fleuret. Pareto manifold learning: Tackling multiple tasks via ensembles of single-task models. In International Conference on Machine Learning, pages 8015-8052. PMLR, 2023. 2", + "[10] Chris Fifty, Ehsan Amid, Zhe Zhao, Tianhe Yu, Rohan Anil, and Chelsea Finn. Efficiently identifying task groupings for multi-task learning. Advances in Neural Information Processing Systems, 34:27503-27516, 2021. 1, 2", + "[11] Marco Fumero, Florian Wenzel, Luca Zancato, Alessandro Achille, Emanuele Rodolà, Stefano Soatto, Bernhard Schölkopf, and Francesco Locatello. Leveraging sparse and shared feature activations for disentangled representation learning. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 3, 4, 7" + ], + "bbox": [ + 93, + 215, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Jiechao Guan and Zhiwu Lu. Task relatedness-based generalization bounds for meta learning. In International Conference on Learning Representations, 2022. 2", + "[13] Yuzheng Hu, Ruicheng Xian, Qilong Wu, Qiuling Fan, Lang Yin, and Han Zhao. Revisiting scalarization in multi-task learning: A theoretical perspective. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[14] Ziniu Hu, Zhe Zhao, Xinyang Yi, Tiansheng Yao, Lichan Hong, Yizhou Sun, and Ed Chi. Improving multi-task generalization via regularizing spurious correlation. Advances in Neural Information Processing Systems, 35:11450-11466, 2022. 1, 2, 3, 4, 7", + "[15] Badr Youbi Idrissi, Martin Arjovsky, Mohammad Pezeshki, and David Lopez-Paz. Simple data balancing achieves competitive worst-group-accuracy. In Conference on Causal Learning and Reasoning, pages 336–351. PMLR, 2022. 3", + "[16] Junguang Jiang, Baixu Chen, Junwei Pan, Ximei Wang, Dapeng Liu, Jie Jiang, and Mingsheng Long. Forkmerge: Mitigating negative transfer in auxiliary-task learning. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[17] Polina Kirichenko, Pavel Izmailov, and Andrew Gordon Wilson. Last layer re-training is sufficient for robustness to spurious correlations. arXiv preprint arXiv:2204.02937, 2022. 3", + "[18] Tyler LaBonte, Vidya Muthukumar, and Abhishek Kumar. Towards last-layer retraining for group robustness with fewer annotations. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[19] Sébastien Lachapelle, Tristan Deleu, Divyat Mahajan, Ioannis Mitliagkas, Yoshua Bengio, Simon Lacoste-Julien, and Quentin Bertrand. Synergies between disentanglement and sparsity: Generalization and identifiability in multi-task learning. In International Conference on Machine Learning, pages 18171-18206. PMLR, 2023. 1", + "[20] Andrew K Lampinen and Surya Ganguli. An analytic theory of generalization dynamics and transfer learning in deep linear networks. arXiv preprint arXiv:1809.10374, 2018. 2", + "[21] Dongyue Li, Huy L Nguyen, and Hongyang R Zhang. Identification of negative transfers in multitask learning using surrogate models. arXiv preprint arXiv:2303.14582, 2023. 2, 7", + "[22] Weixin Liang and James Zou. Metashift: A dataset of datasets for evaluating contextual distribution shifts and training conflicts. arXiv preprint arXiv:2202.06523, 2022. 6, 1", + "[23] Bo Liu, Yihao Feng, Peter Stone, and Qiang Liu. Famo: Fast adaptive multitask optimization. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[24] Evan Z Liu, Behzad Haghgoo, Annie S Chen, Aditi Raghunathan, Pang Wei Koh, Shiori Sagawa, Percy Liang, and Chelsea Finn. Just train twice: Improving group robustness without training group information. In International Conference on Machine Learning, pages 6781-6792. PMLR, 2021. 3", + "[25] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of" + ], + "bbox": [ + 516, + 92, + 906, + 902 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "25706", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "International Conference on Computer Vision (ICCV), 2015. 6, 1", + "[26] Shenyu Lu, Junyi Chai, and Xiaoqian Wang. Neural collapse inspired debiased representation learning for min-max fairness. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 2048-2059, 2024. 3, 4", + "[27] Taro Makino, Krzysztof Geras, and Kyunghyun Cho. Generative multitask learning mitigates target-causing confounding. Advances in Neural Information Processing Systems, 35:36546-36558, 2022. 1, 2, 3, 7", + "[28] Michinari Momma, Chaosheng Dong, and Jia Liu. A multi-objective/multi-task learning framework induced by pareto stationarity. In International Conference on Machine Learning, pages 15895-15907. PMLR, 2022. 1", + "[29] Aviv Navon, Aviv Shamsian, Idan Achituve, Haggai Maron, Kenji Kawaguchi, Gal Chechik, and Ethan Fetaya. Multitask learning as a bargaining game. In International Conference on Machine Learning, pages 16428-16446. PMLR, 2022. 1", + "[30] Karl Pearson. Correlation coefficient. In *Royal Society Proceedings*, page 214, 1895. 4", + "[31] Elan Rosenfeld, Pradeep Ravikumar, and Andrej Risteski. The risks of invariant risk minimization. arXiv preprint arXiv:2010.05761, 2020. 3", + "[32] Shiori Sagawa, Pang Wei Koh, Tatsunori B Hashimoto, and Percy Liang. Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization. arXiv preprint arXiv:1911.08731, 2019. 3", + "[33] Shiori Sagawa, Aditi Raghunathan, Pang Wei Koh, and Percy Liang. An investigation of why overparameterization exacerbates spurious correlations. In International Conference on Machine Learning, pages 8346-8356. PMLR, 2020. 3", + "[34] Ozan Sener and Vladlen Koltun. Multi-task learning as multi-objective optimization. Advances in neural information processing systems, 31, 2018. 2", + "[35] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part V 12, pages 746-760. Springer, 2012. 6, 1", + "[36] Trevor Standley, Amir Zamir, Dawn Chen, Leonidas Guibas, Jitendra Malik, and Silvio Savarese. Which tasks should be learned together in multi-task learning? In International conference on machine learning, pages 9120-9132. PMLR, 2020. 1, 2", + "[37] Charles Stein et al. Inadmissibility of the usual estimator for the mean of a multivariate normal distribution. In Proceedings of the Third Berkeley symposium on mathematical statistics and probability, pages 197-206, 1956. 1", + "[38] Shao-Hua Sun. Multi-digit mnist for few-shot learning, 2019. 6, 1", + "[39] Nilesh Tripuraneni, Michael Jordan, and Chi Jin. On the theory of transfer learning: The importance of task diver" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "sity. Advances in neural information processing systems, 33: 7852-7862, 2020. 2", + "[40] Simon Vandenhende, Stamatios Georgoulis, Wouter Van Gansbeke, Marc Proesmans, Dengxin Dai, and Luc Van Gool. Multi-task learning for dense prediction tasks: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(7):3614-3633, 2021. 2", + "[41] Yipei Wang and Xiaoqian Wang. On the effect of key factors in spurious correlation: A theoretical perspective. In International Conference on Artificial Intelligence and Statistics, pages 3745-3753. PMLR, 2024. 3", + "[42] Zirui Wang, Yulia Tsvetkov, Orhan Firat, and Yuan Cao. Gradient vaccine: Investigating and improving multi-task optimization in massively multilingual models. arXiv preprint arXiv:2010.05874, 2020. 1", + "[43] Enneng Yang, Zhenyi Wang, Li Shen, Shiwei Liu, Guibing Guo, Xingwei Wang, and Dacheng Tao. Adamerging: Adaptive model merging for multi-task learning. arXiv preprint arXiv:2310.02575, 2023. 2", + "[44] Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient surgery for multi-task learning. Advances in Neural Information Processing Systems, 33:5824-5836, 2020. 1", + "[45] Amir R Zamir, Alexander Sax, William Shen, Leonidas J Guibas, Jitendra Malik, and Silvio Savarese. Taskonomy: Disentangling task transfer learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3712-3722, 2018. 6, 1", + "[46] Yu Zhang and Qiang Yang. A survey on multi-task learning. IEEE transactions on knowledge and data engineering, 34 (12):5586-5609, 2021. 1, 2", + "[47] Xiao Zhou, Yong Lin, Renjie Pi, Weizhong Zhang, Renzhe Xu, Peng Cui, and Tong Zhang. Model agnostic sample reweighting for out-of-distribution learning. In International Conference on Machine Learning, pages 27203-27221. PMLR, 2022. 3" + ], + "bbox": [ + 516, + 92, + 903, + 599 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "25707", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Identifying and Mitigating Spurious Correlation in Multi-Task Learning/67faa86d-7895-441b-b372-45b9fee04763_model.json b/2025/Identifying and Mitigating Spurious Correlation in Multi-Task Learning/67faa86d-7895-441b-b372-45b9fee04763_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3420e8be78b3350ac9ad58bae12b123da14f3b0c --- /dev/null +++ b/2025/Identifying and Mitigating Spurious Correlation in Multi-Task Learning/67faa86d-7895-441b-b372-45b9fee04763_model.json @@ -0,0 +1,2277 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.135, + 0.13, + 0.865, + 0.154 + ], + "angle": 0, + "content": "Identifying and Mitigating Spurious Correlation in Multi-Task Learning" + }, + { + "type": "text", + "bbox": [ + 0.174, + 0.181, + 0.825, + 0.233 + ], + "angle": 0, + "content": "Junyi Chai, Shenyu Lu & Xiaoqian Wang* \nElmore Family School of Electrical and Computer Engineering, Purdue University \nWest Lafayette, IN, USA" + }, + { + "type": "text", + "bbox": [ + 0.35, + 0.237, + 0.642, + 0.253 + ], + "angle": 0, + "content": "{chai28,lu876,joywang}@purdue.edu" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.287, + 0.327, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.319, + 0.485, + 0.789 + ], + "angle": 0, + "content": "Multi-task learning (MTL) is a paradigm that aims to improve the generalization of models by simultaneously learning multiple related tasks, leveraging shared representations and task-specific information to enhance performance on individual tasks. However, existing work has shown that MTL can potentially hinder generalization, with one key factor being spurious correlations between tasks. Owing to the knowledge-sharing property, the per-task predictors are more likely to develop reliance on spurious features. Most existing approaches address this issue through distributional robustness, aiming to maintain consistent performance across different distributions under unknown covariate shifts. However, this formulation lacks theoretical guarantees and can be sensitive to the construction of covariate shifts. In this work, we propose a novel perspective, where we seek to identify spurious correlations between tasks. Drawing inspirations from conventional formulations on spurious correlation, for each task, we propose to distinguish its spurious tasks using the difference in correlation coefficients between the empirical distribution and class-wise resampled distributions, thereby capturing the correlations between task labels w.r.t. each class. We prove theoretically the feasibility of the resampling strategy in characterizing spurious correlations between tasks. Furthermore, we propose a simple fine-tuning strategy, debiased adversarial training, where the per-task predictors are adversarily trained to disregard information associated with their spurious tasks. Experimental results on six benchmark datasets show that our method effectively mitigates spurious correlations and outperforms state-of-the-art methods in improving generalization." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.818, + 0.222, + 0.834 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.845, + 0.483, + 0.876 + ], + "angle": 0, + "content": "Stemming from Stein's paradox [37] where the estimation error decreases under joint estimation, formalized in [4]," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.288, + 0.908, + 0.501 + ], + "angle": 0, + "content": "multi-task learning (MTL) has since become increasingly important in various applications, particularly in computer vision, demonstrating both improved accuracy and reduced computational overhead. MTL aims to jointly learn multiple tasks, allowing each task to leverages the shared representations and domain knowledge between tasks, with the goal of improving the generalization performance across all tasks [19, 46]. Compared with single-task learning, which focuses on optimizing over one specific objective, MTL extracts common patterns that are beneficial across different but related objectives, potentially leading to better performance on individual tasks. This collaborative framework mitigates overfitting and enhances model robustness through cross-task knowledge transfer [28, 29]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.504, + 0.909, + 0.807 + ], + "angle": 0, + "content": "Despite the benefits in MTL, existing work has demonstrated that under certain circumstances, MTL could lead to degraded performance in generalization, particularly when different tasks do not experience high knowledge-sharing [10, 36]. This abnormality has conventionally been attributed to the model capacity issue, where conflicting tasks may compete during optimization [42, 44]. However, Recent studies have revealed a counterintuitive phenomenon in MTL, where even under over-parametrization, a MTL model might excel in minimizing training loss, yet surprisingly underperform single-task models in terms of generalization [14]. Accordingly, spurious correlations between different tasks have been identified as an intrinsic problem that hurts generalization in MTL [14, 27]. Owing to the multi-tasking property, the shared representations are expected to encode knowledge across tasks; consequently, when two tasks exhibit high label correlations despite having unrelated underlying knowledge, the per-task predictors may erroneously leverage unrelated spurious knowledge, potentially compromising generalization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Most of existing work addressing spurious correlations in MTL approaches this issue through the lens of invariant learning, aiming to ensure that models perform consistently across various empirical distributions [11, 14]. Specifically, the shared representations are expected to recover the ground-truth latent causal features, and by en" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.887, + 0.237, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25698" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.094, + 0.092, + 0.482, + 0.483 + ], + "angle": 0, + "content": "forcing distributional-invariance, the per-task predictors are expected to leverage features that are causally related to each task, instead of spurious features. Nonetheless, on one hand, counterfactual data is generally inaccessible on high-dimensional data, leading to the infeasibility in conducting causality discovery [2]. Consequently, the shared representations may fail to recover the underlying latent features, particularly when approximated by disentanglement across latent dimensions [11, 14]. Moreover, enforcing distributional invariance does not guarantee utilization of only the task-specific causal features, as the difference across empirical distributions does not necessarily break the spurious correlations between tasks [5]. In the absence of prior knowledge regarding spurious correlations between tasks, along with the potential lack of certain label pairings, it becomes difficult to construct the desired distributions. Alternatively, Makino et al. [27] proposes to mitigate the spurious correlations between tasks by generatively estimating the joint distribution of task labels. Despite theoretical soundness of the formulation, estimating joint distributions can be error-prone, especially when certain label combinations are missing. Additionally, estimating joint distributions involving labels of varied dimensions is often infeasible. This issue is particularly relevant for vision tasks, where labels from different tasks are likely to have differing dimensions." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.488, + 0.482, + 0.834 + ], + "angle": 0, + "content": "In this work, we propose an alternative formulation regarding spurious correlations in MTL. Instead of imposing regularizations on distributional invariance, we seek to directly identify the spurious correlations between tasks. Drawing inspirations from conventional formulations of spurious correlation, typically defined by the dependency between labels and spurious information, we propose to identify spurious correlations between tasks using correlation coefficients. Specifically, we detect spurious correlations between two tasks by examining differences in their correlation coefficients, calculated on both the original training data and class-wise resampled training data. We prove theoretically the feasibility of our formulation in identifying spurious correlations between tasks, as well as in distinguishing spurious correlations from genuine task relationships. Building upon the per-task identification of spurious correlations, we formulate the bias mitigation problem as a mini-max game, where for each task, we consider the predictors of its spurious tasks as the adversaries, such that by fine-tuning the task-specific projector and the task-specific predictor, we intend to minimize the training loss for this task, while maximizing the training loss for its spurious tasks under projection." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.839, + 0.395, + 0.851 + ], + "angle": 0, + "content": "We summarize our contribution as follows:" + }, + { + "type": "text", + "bbox": [ + 0.095, + 0.857, + 0.482, + 0.901 + ], + "angle": 0, + "content": "- We propose a theoretical framework for detecting spurious correlations in MTL by analyzing intra-task label dependencies, and we prove theoretically the feasibility of" + }, + { + "type": "text", + "bbox": [ + 0.53, + 0.093, + 0.741, + 0.106 + ], + "angle": 0, + "content": "our method in the identification." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.107, + 0.905, + 0.182 + ], + "angle": 0, + "content": "- We address spurious correlations in multi-task learning by fine-tuning task-specific predictors, enhancing per-task performance through adversarial training that minimizes loss on the primary task while maximizing loss on its identified spurious tasks." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.183, + 0.905, + 0.241 + ], + "angle": 0, + "content": "- We validate from experiments the effectiveness of our framework in identification of spurious correlations between tasks, as well as in improving generalizability on four benchmark datasets." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.107, + 0.905, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.517, + 0.266, + 0.653, + 0.28 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.517, + 0.296, + 0.833, + 0.311 + ], + "angle": 0, + "content": "2.1. Generalization in multi-task learning" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.323, + 0.905, + 0.549 + ], + "angle": 0, + "content": "The generalizability of deep neural networks has often been linked to the availability of large training datasets, making sample sparsity one of the fundamental limitations for model performance [1, 3, 8]. MTL addresses sample sparsity by leveraging labeled data across multiple related tasks, effectively expanding the available training data [13, 46]. Research on multi-task generalization predominantly assumes shared knowledge between tasks, where learning across diverse tasks enables more robust and transferable representations [12, 20, 39]. Despite empirical validations of the assumption [40], the fundamental assumption of task relatedness often fails in real-world applications, where competing tasks may vie for model capacity, potentially degrading MTL performance below that of single-task approaches [10, 36]." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.555, + 0.905, + 0.899 + ], + "angle": 0, + "content": "Prior research has addressed this challenge through two distinct perspectives. One line of work has focused on the model capacity issue, seeking equilibrium between conflicting task objectives. Yang et al. [43] proposes to adaptively reweight per-task training loss, leveraging entropy minimization on unlabeled test samples. Jiang et al. [16] proposes to dynamically merges model branches while optimizing task weights based on the performance on validation set. Simliar reweighing strategies have also been applied in [21, 23]. Alternatively, Chen et al. [6] introduces a selective gradient filtering approach to resolve optimization conflicts, while Dimitriadis et al. [9] and Sener and Koltun [34] propose to formulate the problem from the perspective of Pareto optimality to achieve optimal trade-off between conflicting tasks. In contrast, recent work has focused on the exacerbation of spurious correlations under MTL, where the multi-tasking property could exacerbate reliance on spurious features for each task owing to the interaction between tasks [11, 14, 27]. Specifically, Hu et al. [14] and Fumero et al. [11] address intra-task spurious correlations through distributional invariance, while Makino et al. [27] employs generative modeling of joint distributions to mitigate these correlations." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25699" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.285, + 0.107 + ], + "angle": 0, + "content": "2.2. Spurious correlation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.484, + 0.37 + ], + "angle": 0, + "content": "Spurious correlation refers to misleading patterns that appear to hold in training data but fail to generalize to new, unseen data [32, 33]. Existing work on spurious correlations typically aims to reduce the error for the worst-performing group, characterized by both the task label and the associated spurious information. typically, these methods can be divided into two categories: data-augmentation, where the training distribution is modified to enhancing worst-group performance [5, 8, 15]; model-regularization, where regularizations are imposed during the training process to eliminate the affect of spurious information [24, 26, 32, 41, 47]. More recently, last-layer retraining has been proposed as an efficient alternative for addressing spurious correlations [17, 18]. In this approach, instead of updating all network parameters, only the classifier layer is fine-tuned on a balanced validation set, significantly reducing computational complexity." + }, + { + "type": "image", + "bbox": [ + 0.143, + 0.38, + 0.434, + 0.548 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.562, + 0.484, + 0.646 + ], + "angle": 0, + "content": "Figure 1. Our framework improves generalizability by identifying and correcting spurious correlations between tasks. For instance, on CelebA dataset, the task labels 'male' and 'attractiveness' are known to be spuriously correlated [26]. Compared with vanilla MTL where predictions for attractiveness falsely correlate with predictions for gender, our method effectively mitigates this bias." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.665, + 0.229, + 0.681 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.484, + 0.814 + ], + "angle": 0, + "content": "Consider \\(k\\) tasks in total. Let \\(\\mathcal{D}^{\\mathrm{train}} := \\{(x_i, y_i^1, \\ldots, y_i^k), 1 \\leq i \\leq N\\}\\) be the training set where for \\(i\\)-th sample, \\(x_i \\in \\mathbb{R}^d\\) is the input feature and \\(y_i^j \\in \\mathbb{R}^{d^j}\\) is the label for \\(j\\)-th task, let \\(\\pmb{f}: \\mathbb{R}^d \\to \\mathbb{R}^l\\) be the function of shared encoder and \\(\\pmb{g}^j: \\mathbb{R}^l \\to \\mathbb{R}^{d^j}\\) be the function of predictor for \\(j\\)-th task, let \\(\\mathcal{L}^j\\) be the training loss for \\(j\\)-th task, a multi-task learning problem can be formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.82, + 0.401, + 0.861 + ], + "angle": 0, + "content": "\\[\n\\underset {\\boldsymbol {f}, \\boldsymbol {g} ^ {1}, \\dots , \\boldsymbol {g} ^ {k}} {\\arg \\min } \\sum_ {j = 1} ^ {k} w ^ {j} \\bar {\\mathcal {L}} ^ {j} (\\mathcal {D} ^ {\\text {t r a i n}}; \\boldsymbol {g} ^ {j}, \\boldsymbol {f}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.868, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\bar{\\mathcal{L}}^j (\\mathcal{D}^{\\mathrm{train}};\\pmb {g}^j,\\pmb {f})\\coloneqq \\frac{1}{N}\\sum_{i = 1}^N\\mathcal{L}^j (\\pmb {g}^j (\\pmb {f}(x_i)),y_i^j)\\) is the average training loss for \\(j\\) -th task, and \\(w^{j}\\) specifies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.091, + 0.905, + 0.287 + ], + "angle": 0, + "content": "the weight of \\(\\bar{\\mathcal{L}}^j\\) during optimization. Despite the benefit of exploring task-relatedness, which are expected to benefit the generalizability, recent work on MTL has pointed to the potential risk of spurious correlations between tasks owing to the confounder between task labels [14, 27]. Let \\(\\{\\mathcal{F}_i, 1 \\leq i \\leq K\\}\\) be the set of underlying latent features such that \\(\\mathcal{F}_i \\perp \\mathcal{F}_j\\), \\(\\forall i \\neq j\\), let \\(\\mathcal{F}^j\\) be the set of underlying latent features that are causally related with \\(j\\)-th task, when \\(\\mathcal{F}^i \\neq \\mathcal{F}^j\\), optimizing \\(\\bar{\\mathcal{L}}^i\\) and \\(\\bar{\\mathcal{L}}^j\\) jointly could exacerbate reliance on spurious features from \\(\\mathcal{F}^j\\) for \\(g^i\\), as the shared encoder needs to explicitly encode knowledge for both tasks. Such confounder brings in spurious correlations between tasks and can harm generalization performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.288, + 0.907, + 0.438 + ], + "angle": 0, + "content": "Nonetheless, compared with conventional formulations where the spurious information is assumed to be available on training data [32, 33], regarding MTL, we do not have the spurious information at hand, where each task may consider all other tasks as being either spurious or beneficial. Alternatively, most existing work [11, 14] has formulated the problem from the perspective of distributional robustness, where the model is expected to perform consistently on both the training data and the validation data \\(\\mathcal{D}^{\\mathrm{val}}\\) under some unknown covariate shift:" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.442, + 0.918, + 0.518 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\underset {\\boldsymbol {f}, \\boldsymbol {g} ^ {1}, \\dots , \\boldsymbol {g} ^ {k}} {\\arg \\min } \\sum_ {j = 1} ^ {k} w ^ {j} \\bar {\\mathcal {L}} ^ {j} (\\mathcal {D} ^ {\\text {t r a i n}}; \\boldsymbol {g} ^ {j}, \\boldsymbol {f}) + \\sum_ {j = 1} ^ {k} w ^ {\\text {v a l}, j} \\bar {\\mathcal {L}} ^ {j} (\\mathcal {D} ^ {\\text {v a l}}; \\boldsymbol {g} ^ {j}, \\boldsymbol {f}), \\\\ s. t. R \\left(\\mathcal {D} ^ {\\text {t r a i n}}; f\\right) \\leq \\epsilon , R \\left(\\mathcal {D} ^ {\\text {v a l}}; f\\right) \\leq \\epsilon , \\tag {1} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.519, + 0.906, + 0.773 + ], + "angle": 0, + "content": "where \\( w^{\\mathrm{val},j} \\) specifies the weight of validation loss for \\( j \\)-th task, and \\( R(\\mathcal{D}^{\\mathrm{train}};f) \\) and \\( R(\\mathcal{D}^{\\mathrm{val}};f) \\) correspond to the regularization on \\( f \\) such that the encoder is dimensional-wise disentangled. For example, MT-CRL [14] employs covariance between latent dimensions as the regularization, while Meta-learning [11] formulates \\( R(\\mathcal{D}^{\\mathrm{train}};f) \\) as the entropy of normalized representations per dimension. Eq. (1) stems from the assumption that for \\( i \\)-th task, the covariate shift on validation data effectively disrupts the correlation between its spurious features and its task labels, which could be infeasible in general. On one hand, without prior knowledge on the task-specific causal features, there's no guarantee that the covariate shift breaks the spurious correlation between tasks. Furthermore, the effectiveness of Eq. (1) may be influenced by the degree of bias in the validation set, introducing uncertainty in the mitigation spurious correlations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.775, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We use the following experiments on linear SEM [31] to validate the two perspectives. Following existing work [14], we construct two binary classification tasks where the causal features of each task are independent and follow a Gaussian distribution. Let \\( Y^{j} \\) be the random variable of labels for \\( j \\)-th task, we set the ratios between task labels on training data as \\( \\alpha_{y}^{\\text{train}} = \\frac{P(Y^{0} = y, Y^{1} = (1 - y))}{P(Y^{0} = y, Y^{1} = y)} = \\frac{1}{9} \\), \\( \\forall y \\in \\{0, 1\\} \\) on training data. Regarding both tasks, we have" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25700" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.485, + 0.173 + ], + "angle": 0, + "content": "\\(P(Y^{j} = 0) = P(Y^{j} = 1) = 0.5, \\forall j \\in \\{0,1\\}\\), and the testing data is balanced sampled across each co-couurance of task labels such that \\(\\alpha_{y}^{\\mathrm{test}} = \\frac{P(Y^{0} = y, Y^{1} = (1 - y))}{P(Y^{0} = y, Y^{1} = y)} = 1\\), \\(\\forall y \\in \\{0,1\\}\\). Smaller \\(\\alpha_{y}\\) indicates higher correlation between the two task labels." + }, + { + "type": "table", + "bbox": [ + 0.122, + 0.18, + 0.456, + 0.232 + ], + "angle": 0, + "content": "
TaskMTL (αtrainy = 1/9)MTL (αtrainy = 1)
Task 081.32±0.29%94.47±0.34%
Task 181.37±0.41%94.47±0.35%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.242, + 0.485, + 0.298 + ], + "angle": 0, + "content": "Table 1. Experimental results on synthetic data. We choose classification accuracy as the evaluation metric for both tasks. Under high correlation, the model performs remarkably worse than that under no spurious correlation." + }, + { + "type": "table", + "bbox": [ + 0.123, + 0.324, + 0.456, + 0.376 + ], + "angle": 0, + "content": "
TaskMT-CRL (αyval = 1/9)Meta-learning
Task 081.59±0.36%81.73±0.35%
Task 181.87±0.95%81.67±0.44%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.385, + 0.484, + 0.455 + ], + "angle": 0, + "content": "Table 2. Experimental results on synthetic data. Both methods show marginal improvement in testing accuracy compared with MTL (\\(\\alpha_{y}^{\\mathrm{train}} = \\frac{1}{9}\\)): \\(81.32 \\pm 0.29\\%\\) for task 0 and \\(81.37 \\pm 0.41\\%\\) for task 1, owing to the consistency in \\(\\alpha_{y}\\)'s between training and validation data." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.481, + 0.472, + 0.533 + ], + "angle": 0, + "content": "
TaskMT-CRL (αyval = 1/5)MT-CRL (αyval = 1)
Task 083.74±0.65%87.59±0.74%
Task 183.32±0.59%88.16±0.46%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.543, + 0.484, + 0.641 + ], + "angle": 0, + "content": "Table 3. Experimental results on synthetic data. The testing performance improves as \\(\\alpha_{y}^{\\mathrm{val}}\\)'s decrease, substantiating the importance of mitigating bias on the validation data. Nonetheless, given the balanced validation set, MT-CRL still performs worse than the optimal baseline MTL (\\(\\alpha_{y}^{\\mathrm{train}} = 1\\)): \\(94.47 \\pm 0.34\\%\\) for task 0 and \\(94.47 \\pm 0.35\\%\\) for task 1, highlighting the insufficiency in formulating spurious correlations via distributional robustness." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.658, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We show results of under different constructions of validation sets in Tab. 1 - 2. Regarding MT-CRL [14], as the validation set is only assumed to contain unknown shift, we consider three different types of covariate shifts: The training-validation split where the validation set is sampled from the training distribution, as suggested by the authors as a general construction of covariate shift, where \\(\\alpha_{y}\\)'s remain the same on both training and validation data; A less biased validation set where \\(\\alpha_{y}^{\\mathrm{val}} = \\frac{1}{5}, \\forall y \\in \\{0,1\\}\\); A balanced validation set where \\(\\alpha_{y}^{\\mathrm{val}} = 1, \\forall y \\in \\{0,1\\}\\). While a balanced validation set shows improvement in the performance on testing data, the first two constructions of validation data show marginal improvement compared with vanilla MTL, validating our claim that the effectiveness of distributional invariance can be subject to the bias on validation data. Nonetheless, in reality, it is hard to obtain or construct a" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.166 + ], + "angle": 0, + "content": "perfectly balanced data owing to the uncertainty in the spurious correlations between tasks. Furthermore, it is possible that certain labeling pairs are not observed on training or validation data, leading to infeasibility of constructing such balanced data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.168, + 0.907, + 0.349 + ], + "angle": 0, + "content": "Regarding Meta-learning [11], since the validation set is constructed per task such that for \\( j \\)-th task, \\( P(Y^{j} = 0) = P(Y^{j} = 1) = 0.5 \\), \\( \\forall j \\in \\{0,1\\} \\), we have \\( \\alpha_{y} \\)'s remain the same on both training and validation data. As a result, such construction fails to show improvement in mitigating spurious information compared with the ERM baseline. This demonstrates the flaws in formulating the covariate shift via class-balance: the spurious correlation between tasks does not necessarily lead to imbalance in the marginal distribution of labels per task, where using class-balanced resampling can lead to marginal variations in the conditional distributions \\( P(Y_{i}|Y_{j}) \\) of task labels." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.362, + 0.606, + 0.377 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.387, + 0.906, + 0.417 + ], + "angle": 0, + "content": "4.1. Identification of spurious correlation between tasks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.424, + 0.907, + 0.621 + ], + "angle": 0, + "content": "Our preliminary discussion suggests the necessity of identifying spurious correlations between tasks. In light of the intractability in identifying the latent causal features for each task, particularly regarding high-dimensional vision data, we instead consider modeling the spurious correlations from the label perspective. Existing work [14, 26] has demonstrated that when two tasks are spuriously correlated, their corresponding task labels will be disproportionately paired, leading to an uneven joint distribution across task labels. Specifically, let \\(\\mathbb{S}^j\\) be the set of values of labels for \\(j\\)-th task, when spurious correlation exists between \\(i\\)-th and \\(j\\)-th tasks, there exist \\(y, y' \\in \\mathbb{S}^i\\) and \\(\\tilde{y}, \\tilde{y}' \\in \\mathbb{S}^j\\), such that the following inequality regarding label distribution holds:" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.631, + 0.907, + 0.651 + ], + "angle": 0, + "content": "\\[\nP (Y ^ {i} = y, Y ^ {j} = \\tilde {y}) > P (Y ^ {i} = y ^ {\\prime}, Y ^ {j} = \\tilde {y} ^ {\\prime}). \\quad (2)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.661, + 0.906, + 0.707 + ], + "angle": 0, + "content": "To quantitatively evaluate Eq. (2), we observe that the correlation between task labels can be quantified using the Pearson correlation coefficient [30]:" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.716, + 0.908, + 0.755 + ], + "angle": 0, + "content": "\\[\n\\rho_ {Y ^ {i}, Y ^ {j}} = \\frac {\\mathbb {E} [ Y ^ {i} Y ^ {j} ] - \\mathbb {E} [ Y ^ {i} ] \\mathbb {E} [ Y ^ {j} ]}{\\sqrt {\\mathbb {E} [ (Y ^ {i}) ^ {2} ] - (\\mathbb {E} [ Y ^ {i} ]) ^ {2}} \\sqrt {\\mathbb {E} [ (Y ^ {j}) ^ {2} ] - (\\mathbb {E} [ Y ^ {j} ]) ^ {2}}}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.763, + 0.907, + 0.854 + ], + "angle": 0, + "content": "For simplicity of discussion, consider both tasks as binary classification where \\( Y^{i} \\) and \\( Y^{j} \\) follow Bernoulli distributions with expectations \\( p^{i} \\) and \\( p^{j} \\), respectively. We refer to Algorithm 1 for extension to multi-class scenario. The following lemma states the close-form expression of \\( \\rho_{Y^{i},Y^{j}} \\):" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.865, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Lemma 1. Let \\(\\alpha_{y}^{ij} := \\frac{P(Y^i = y, Y^j = 0)}{P(Y^i = y, Y^j = 1)}\\) be the ratio of joint probabilities of \\(Y^i\\) and \\(Y^j\\) evaluated w.r.t. class \\(y\\) of \\(i\\)-th" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25701" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.278, + 0.09, + 0.744, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.367, + 0.908, + 0.414 + ], + "angle": 0, + "content": "Figure 2. Demonstration of our debiased adversarial training framework. For each task, the predictors within the blue rectangle represent those identified as spurious by our method. During optimization, the per-task projector and predictor are updated to ensure that the projected representation minimizes reliance on information of spurious tasks while maintaining performance on the target task." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.437, + 0.253, + 0.454 + ], + "angle": 0, + "content": "task, we have \\(\\rho_{Y^i,Y^j}\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.457, + 0.407, + 0.505 + ], + "angle": 0, + "content": "\\[\n\\rho_ {Y ^ {i}, Y ^ {j}} = \\frac {\\left(\\frac {1}{(1 + \\alpha_ {1} ^ {i j})} - p ^ {j}\\right) p ^ {i}}{\\sqrt {p ^ {i} p ^ {j} (1 - p ^ {i}) (1 - p ^ {j})}}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.517, + 0.485, + 0.875 + ], + "angle": 0, + "content": "We defer full proof to the Appendix. As shown in Lemma 1, the correlation coefficient between \\( Y^{i} \\) and \\( Y^{j} \\) is determined by the marginal distributions of \\( Y^{i} \\) and \\( Y^{j} \\) and the ratio of joint probabilities. However, the dependence of \\( \\rho_{Y^i,Y^j} \\) on the marginal distributions could introduce potential errors in identifying spurious correlations, particularly in the presence of label imbalance, where the disproportionality between labeling pairs could be overlooked given small marginal probabilities of task labels. For example, even with a small \\( \\alpha_{1}^{ij} \\), when \\( p^i \\) becomes trivial due to class imbalance, we may observe \\( \\rho_{Y^i,Y^j}\\approx 0 \\), despite the presence of spurious correlations between the two tasks. We therefore consider the following class-wise resampling strategy to perturb the marginal distributions of task labels, such that when \\( p^i \\) and \\( p^j \\) change, we have \\( \\rho_{Y^i,Y^j} \\) change, which enables us to recognize the spurious correlation between \\( Y^{i} \\) and \\( Y^{j} \\) while mitigating the influence of label imbalance. Specifically, let \\( Y^{\\prime i} \\) and \\( Y^{\\prime j} \\) be the random variable of task labels after resampling, let \\( \\mathcal{D}_{Y^j\\neq y}^{\\mathrm{train}}\\coloneqq \\{(x_i,y_i^1,\\dots ,y_i^k)|y_i^j\\neq y\\} \\) be the subset of samples excluding class \\( y \\) of \\( j \\)-th task and \\( \\mathcal{D}_{Y^{\\prime j}\\neq y}^{\\mathrm{train}} \\) be the corresponding subset after resampling, we consider the following down-sampling scheme within \\( \\mathcal{D}_{Y^j\\neq y}^{\\mathrm{train}} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.878, + 0.371, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\left| \\mathcal {D} _ {Y ^ {\\prime j} \\neq y} ^ {' \\mathrm {t r a i n}} \\right| = \\xi \\left| \\mathcal {D} _ {Y ^ {j} \\neq y} ^ {\\mathrm {t r a i n}} \\right|,\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.436, + 0.907, + 0.596 + ], + "angle": 0, + "content": "where \\(\\xi < 1\\) is the constant that determines the degree of resampling, and \\(\\mathcal{D}_{Y^{\\prime j}\\neq y}^{\\prime \\mathrm{train}}\\) is constructed by random sampling. Under class-wise resampling, we intend to amplify the effect of correlations between task labels under different values of \\(Y^{\\prime j}\\) within \\(\\mathbb{S}^j\\), so as to properly capture the dependence between task labels. Let \\(p^{\\prime i}\\) be the expectation of \\(Y^{\\prime i}\\) evaluated on the resampled set \\(\\mathcal{D}_{Y^{\\prime j}\\neq y}^{\\prime \\mathrm{train}}\\cup \\left(\\mathcal{D}^{\\mathrm{train}}\\backslash \\mathcal{D}_{Y^{\\prime j}\\neq y}^{\\mathrm{train}}\\right)\\), the following theorem states the feasibility of identifying spurious correlations between tasks based on the difference in correlation coefficients before and after resampling:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.604, + 0.907, + 0.667 + ], + "angle": 0, + "content": "Theorem 1. When spurious correlation exists between two tasks, the maximum difference in correlation coefficient before and after resampling is lower-bounded by the difference in expectations \\( p^i - p'^i \\), up to multiplicative constants:" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.672, + 0.833, + 0.701 + ], + "angle": 0, + "content": "\\[\n\\left| \\rho_ {Y ^ {i}, Y ^ {j}} - \\rho_ {Y ^ {\\prime i}, Y ^ {\\prime j}} \\right| \\geq \\mathcal {K} \\left| p ^ {' i} - p ^ {i} \\right|,\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.708, + 0.899, + 0.746 + ], + "angle": 0, + "content": "\\[\nw h e r e \\mathcal {K} = \\frac {\\left(\\frac {1}{\\hat {\\alpha} _ {1} ^ {i j}} - \\hat {\\alpha} _ {1} ^ {i j}\\right)}{\\sqrt {(1 - p ^ {i}) (1 - (\\hat {\\alpha} _ {1} ^ {i j} - \\hat {\\alpha} _ {0} ^ {i j}) p ^ {i} - \\hat {\\alpha} _ {0} ^ {i j})}} a n d \\hat {\\alpha} _ {y} ^ {i j} = \\frac {1}{1 + \\alpha_ {y} ^ {i j}}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.902 + ], + "angle": 0, + "content": "We defer full proof to the Appendix. As shown in Theorem 1, since \\(\\hat{\\alpha}_{y}^{ij}\\) and \\(p^i\\) remain constant before and after resampling, we have the lower-bound of difference in correlation coefficient before and after resampling monotonically increase as \\(p^{'i}\\) increases. Accordingly, even though \\(p^i\\) can be trivial, so long as \\(p^i\\) and \\(p^{'i}\\) differ significantly before and after resampling, we are guaranteed a substantial change in the correlation coefficient, which enables us to identify the spurious correlations among tasks. Furthermore, the following remark states the difference in correla" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25702" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.121 + ], + "angle": 0, + "content": "tion coefficient when spurious correlations do not exist between tasks:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.131, + 0.483, + 0.164 + ], + "angle": 0, + "content": "Remark 1. When two tasks are not spuriously correlated, we have \\(\\rho_{Y^i,Y^j} - \\rho_{Y'^i,Y'^j} = 0\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.172, + 0.484, + 0.203 + ], + "angle": 0, + "content": "Consequently, let \\(\\mathbb{Z}^j\\) be the set of indices of identified spurious tasks for \\(j\\)-th task, we have" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.212, + 0.405, + 0.239 + ], + "angle": 0, + "content": "\\[\n\\left| \\rho_ {Y ^ {i}, Y ^ {j}} - \\rho_ {Y ^ {\\prime i}, Y ^ {\\prime j}} \\right| > \\delta , \\forall i \\in \\mathbb {Z} ^ {j},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.248, + 0.484, + 0.329 + ], + "angle": 0, + "content": "where \\(\\delta\\) is the threshold for distinguishing the spurious correlations between tasks. In practice, \\(\\delta\\) can be trivially obtained owing to the distinctions in \\(\\left|\\rho_{Y^i,Y^j} - \\rho_{Y'^i,Y'^j}\\right|\\) as stated in Theorem 1 and Remark 1. We refer to Sec. 5.3 for the detailed discussion." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.339, + 0.364, + 0.355 + ], + "angle": 0, + "content": "4.2. Debiased Adversarial Training" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.361, + 0.484, + 0.603 + ], + "angle": 0, + "content": "We move on to discuss the mitigation of spurious correlation between tasks. Due to the potential absence of certain label pairs in the training set, which impedes constructing balanced data even after identifying spurious tasks, we instead aim to ensure that the task-specific predictors effectively discard non-causal features for each task. In light of the multi-tasking property, we can readily formulate the objective as an adversarial learning problem w.r.t. the pertask predictors. Let \\(\\hat{f}\\) be the function of pre-trained encoder which remains frozen during fine-tuning, regarding \\(j\\)-th task, given its identified set of spurious tasks \\(\\mathbb{Z}^j\\), to utilize the per-task predictors as discriminators of non-causal features, we apply one extra projector \\(h^j: \\mathbb{R}^l \\to \\mathbb{R}^l\\) before \\(g^j\\), such that the prediction for \\(j\\)-th task is given by \\(g^i \\circ h^i \\circ \\hat{f}(x)\\). We consider the following objective for mitigating spurious correlations regarding \\(j\\)-th task:" + }, + { + "type": "equation", + "bbox": [ + 0.108, + 0.625, + 0.466, + 0.658 + ], + "angle": 0, + "content": "\\[\n\\operatorname *{arg min}_{\\boldsymbol{h}^{j},\\boldsymbol{g}^{j}}\\bar{\\mathcal{L}}^{j}\\left(\\mathcal{D}^{\\text{train}};\\boldsymbol{h}^{j},\\boldsymbol{g}^{j}\\right) - \\sum_{t\\in \\mathbb{Z}^{j}}\\lambda^{j}\\bar{\\mathcal{L}}^{t}\\left(\\mathcal{D}^{\\text{train}};\\boldsymbol{h}^{j},\\boldsymbol{g}^{t}\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.67, + 0.484, + 0.811 + ], + "angle": 0, + "content": "where the optimization is performed over \\( \\pmb{h}^j, \\pmb{g}^j \\), while \\( \\pmb{g}^t \\) remains frozen \\( \\forall t \\in \\mathbb{Z}^j \\), and \\( \\bar{\\mathcal{L}}^j(\\mathcal{D}^{\\mathrm{train}}; \\pmb{h}^j, \\pmb{g}^j) := \\frac{1}{N} \\sum_{i=1}^{N} \\mathcal{L}^j(\\pmb{g}^j(\\pmb{h}^j(\\hat{\\pmb{f}}(x_i))), y_i^j), \\quad \\bar{\\mathcal{L}}^t(\\mathcal{D}^{\\mathrm{train}}; \\pmb{h}^j, \\pmb{g}^t) := \\frac{1}{N} \\sum_{i=1}^{N} \\mathcal{L}^k(\\pmb{g}^k(\\pmb{h}^j(\\pmb{f}(x_i))), y_i^k) \\) are the average training loss under projection. By minimizing the objective, we intend to ensure that the projected latent representation retains only the information pertinent to \\( j \\)-th task while shedding the spurious information. We summarize our method in Algorithm 1." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.824, + 0.224, + 0.841 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.282, + 0.866 + ], + "angle": 0, + "content": "5.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We validate our method on the following benchmark datasets: Multi-MNIST dataset [38], CelebA dataset [25]," + }, + { + "type": "code_caption", + "bbox": [ + 0.515, + 0.092, + 0.8, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 Debiased adversarial training" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.123, + 0.903, + 0.186 + ], + "angle": 0, + "content": "Input: Pre-trained encoder \\(\\hat{f}\\), task-specific predictors \\(g^{j}\\)'s s.t. \\(1 \\leq j \\leq k\\), task-specific projectors \\(h^{j}\\)'s, training data \\(\\mathcal{D}^{\\mathrm{train}}\\), training epochs \\(E\\), learning rate \\(\\eta\\), threshold \\(\\delta\\), resampling constant \\(\\xi\\), hyper-parameter \\(\\lambda^{j}\\)." + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.196, + 0.66, + 0.212 + ], + "angle": 0, + "content": "1: for \\( j = 1 \\) to \\( k \\) do:" + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.213, + 0.619, + 0.226 + ], + "angle": 0, + "content": "2: \\(\\mathbb{Z}^j = \\emptyset\\)" + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.228, + 0.747, + 0.242 + ], + "angle": 0, + "content": "3: for \\( i = 1 \\) to \\( k \\) s.t. \\( i \\neq j \\) do:" + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.242, + 0.769, + 0.258 + ], + "angle": 0, + "content": "4: Calculate \\(\\rho_{Y^i,Y^j}\\) on \\(\\mathcal{D}^{\\mathrm{train}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.258, + 0.691, + 0.272 + ], + "angle": 0, + "content": "5: for \\(y\\in \\mathbb{S}^j\\) do:" + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.273, + 0.825, + 0.288 + ], + "angle": 0, + "content": "6: Perform class-wise resampling:" + }, + { + "type": "list", + "bbox": [ + 0.522, + 0.196, + 0.825, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "equation", + "bbox": [ + 0.642, + 0.298, + 0.806, + 0.325 + ], + "angle": 0, + "content": "\\[\n\\left| \\mathcal {D} _ {Y ^ {\\prime j} \\neq y} ^ {\\text {t r a i n}} \\right| = \\xi \\left| \\mathcal {D} _ {Y ^ {j} \\neq y} ^ {\\text {t r a i n}} \\right|.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.335, + 0.761, + 0.351 + ], + "angle": 0, + "content": "7: Calculate \\(\\rho_{Y^{\\prime}i,Y^{\\prime}j}\\) on" + }, + { + "type": "equation", + "bbox": [ + 0.629, + 0.363, + 0.819, + 0.389 + ], + "angle": 0, + "content": "\\[\n\\mathcal {D} _ {Y ^ {\\prime j} \\neq y} ^ {\\text {t r a i n}} \\cup \\left(\\mathcal {D} ^ {\\text {t r a i n}} \\backslash \\mathcal {D} _ {Y ^ {j} \\neq y} ^ {\\text {t r a i n}}\\right).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.4, + 0.897, + 0.416 + ], + "angle": 0, + "content": "8: if \\(\\rho_{Y^i,Y^j} - \\rho_{Y'^i,Y'^j}\\geq \\delta\\) and \\(i\\notin \\mathbb{Z}^j\\) then:" + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.418, + 0.746, + 0.432 + ], + "angle": 0, + "content": "9: \\(\\mathbb{Z}^j = \\mathbb{Z}^j\\cup \\{i\\}\\)" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.433, + 0.661, + 0.445 + ], + "angle": 0, + "content": "10: end if" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.449, + 0.646, + 0.461 + ], + "angle": 0, + "content": "11: end for" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.464, + 0.622, + 0.476 + ], + "angle": 0, + "content": "12: end for" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.479, + 0.598, + 0.49 + ], + "angle": 0, + "content": "13: end for" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.493, + 0.697, + 0.506 + ], + "angle": 0, + "content": "14: for epoch \\(= 1\\) to \\(E\\) do:" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.508, + 0.684, + 0.522 + ], + "angle": 0, + "content": "15: for \\(\\mathbf{j} = 1\\) to \\(k\\) do:" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.525, + 0.684, + 0.538 + ], + "angle": 0, + "content": "16: Training loss:" + }, + { + "type": "list", + "bbox": [ + 0.521, + 0.4, + 0.897, + 0.538 + ], + "angle": 0, + "content": null + }, + { + "type": "equation", + "bbox": [ + 0.556, + 0.549, + 0.89, + 0.582 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\bar {\\mathcal {L}} ^ {j} \\left(\\mathcal {D} ^ {\\text {t r a i n}}; \\boldsymbol {h} ^ {j}, \\boldsymbol {g} ^ {j}\\right) - \\sum_ {t \\in \\mathbb {Z} ^ {j}} \\lambda^ {j} \\bar {\\mathcal {L}} ^ {t} \\left(\\mathcal {D} ^ {\\text {t r a i n}}; \\boldsymbol {h} ^ {j}, \\boldsymbol {g} ^ {t}\\right).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.594, + 0.848, + 0.61 + ], + "angle": 0, + "content": "17: Update \\( h^j \\) and \\( g^j \\) by gradient descent:" + }, + { + "type": "equation", + "bbox": [ + 0.66, + 0.62, + 0.786, + 0.651 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {h} ^ {j} = \\boldsymbol {h} ^ {j} - \\eta \\frac {\\partial \\mathcal {L}}{\\partial \\boldsymbol {h} ^ {j}}.\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.663, + 0.659, + 0.784, + 0.692 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {g} ^ {j} = \\boldsymbol {g} ^ {j} - \\eta \\frac {\\partial \\mathcal {L}}{\\partial \\boldsymbol {g} ^ {j}}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.697, + 0.621, + 0.71 + ], + "angle": 0, + "content": "18: end for" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.713, + 0.597, + 0.725 + ], + "angle": 0, + "content": "19: end for" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.697, + 0.621, + 0.725 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.756, + 0.907, + 0.863 + ], + "angle": 0, + "content": "Taskonomy dataset [45], MetaShift dataset [22], NYU-V2 dataset [35] and CityScape dataset [7]. We implement our method in PyTorch 2.0.0 on one RTX-3090 GPU1. We set \\(\\xi = 0.1\\) during experiments and evaluate the difference in correlation coefficients between training set and class-wise resampled sets to construct \\(\\mathbb{Z}^j\\). We use accuracy as evaluation metric for classification tasks, cross-entropy loss and" + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.875, + 0.905, + 0.9 + ], + "angle": 0, + "content": "\\(^{1}\\)Code available at https://github.com/cjy24/Spurious-Correlation-MTL." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25703" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.185, + 0.089, + 0.813, + 0.139 + ], + "angle": 0, + "content": "
TaskSTLSubSelMT-CRLMeta-learningGMTLOurs
left digit (accuracy, ↑)+4.9%+3.4%+5.1%+0.2%+5.9%+9.3%
right digit (accuracy, ↑)+5.9%+2.9%+6.3%+0.3%+7.2%+9.6%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.149, + 0.908, + 0.178 + ], + "angle": 0, + "content": "Table 4. Experimental results on Multi-MNIST dataset. The relative improvement of each method is calculated in comparison to vanilla MTL." + }, + { + "type": "table", + "bbox": [ + 0.172, + 0.19, + 0.828, + 0.272 + ], + "angle": 0, + "content": "
TaskSTLSubSelMT-CRLMeta-learningGMTLOurs
Gender (accuracy, ↑)+2.2%+1.7%+1.3%+2.4%+2.6%+5.7%
Smiling (accuracy, ↑)+2.3%+2.5%+2.6%+3.4%+1.2%+4.6%
Age (accuracy, ↑)+1.2%+2.2%+0.1%+1.1%+1.5%+3.9%
Attractiveness (accuracy, ↑)+0.6%+0.4%+0.7%+0.2%+1.2%+2.4%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.352, + 0.281, + 0.645, + 0.295 + ], + "angle": 0, + "content": "Table 5. Experimental results on CelebA dataset." + }, + { + "type": "table", + "bbox": [ + 0.188, + 0.309, + 0.812, + 0.376 + ], + "angle": 0, + "content": "
TaskSTLSubSelMT-CRLMeta-learningOurs
Scene (accuracy, ↑)+4.1%+3.2%+5.8%+5.3%+9.9%
Object (accuracy, ↑)+5.2%+4.1%+7.3%+8.1%+12.7%
Segmentation (cross entropy, ↓)-5.6%+1.7%+3.1%+6.9%+6.2%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.341, + 0.385, + 0.656, + 0.4 + ], + "angle": 0, + "content": "Table 6. Experimental results on Taskonomy dataset." + }, + { + "type": "table", + "bbox": [ + 0.194, + 0.412, + 0.806, + 0.463 + ], + "angle": 0, + "content": "
TaskSTLSubSelMT-CRLMeta-learningGMTLOurs
Animal (accuracy, ↑)+4.2%+2.7%+3.2%+1.5%+0.2%+5.4%
Scene (accuracy, ↑)+3.1%+2.1%+1.6%+0.1%+1.7%+4.0%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.345, + 0.473, + 0.652, + 0.487 + ], + "angle": 0, + "content": "Table 7. Experimental results on MetaShift dataset." + }, + { + "type": "table", + "bbox": [ + 0.199, + 0.501, + 0.8, + 0.552 + ], + "angle": 0, + "content": "
TaskSTLSubSelMT-CRLMeta-learningOurs
Segmentation (mean IoU, ↑)+2.7%+4.2%+6.9%+7.2%+11.3%
Depth (relative error, ↓)+3.1%+3.8%+7.5%+6.2%+10.2%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.347, + 0.561, + 0.651, + 0.576 + ], + "angle": 0, + "content": "Table 8. Experimental results on NYU-V2 dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.602, + 0.483, + 0.648 + ], + "angle": 0, + "content": "mean IoU for segmentation tasks and relative error for depth estimation tasks. We defer details of each dataset and full results to the appendix." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.654, + 0.483, + 0.82 + ], + "angle": 0, + "content": "We compare our method with the following related methods on spurious correlation and generalization in MTL: Single-task learning (STL): the network is optimized over each task only without regularizations; MTL: vanilla MTL without regularizations; SubSel [21]: MTL with subset selection to address task conflicts; MT-CRL [14]: MTL with disentangled neural modules and invariant regularization on task-to-module routing; Meta-learning [11]: MTL with disentangled representation learning and bi-level optimization on per-task predictors; GMTL [27]: MTL with generative modeling of task labels." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.903 + ], + "angle": 0, + "content": "We repeat experiments on each dataset three times and report the average results. For MT-CRL [14], the covariate shift on training data is constructed by training-validation split, as suggested by the authors. Further, we do not include GMTL [27] on Taskonomy, NYU-V2 and CityScape" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.602, + 0.907, + 0.679 + ], + "angle": 0, + "content": "dataset, as it is infeasible to estimate the joint distribution involving high-dimensional masks. Following prior work [14], we report the relative performance improvement of each method over vanilla MTL for evaluation, as the scales of evaluation metrics differ significantly across tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.695, + 0.719, + 0.712 + ], + "angle": 0, + "content": "5.2. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Results are shown in Tab. 4 - 8. Compared with vanilla STL and MTL, Meta-learning [11] shows marginal improvement in testing performance, particularly when the marginal training distributions are evenly distributed per task. Regarding MT-CRL [14], the training-validation split does not appropriately capture the spurious correlations between tasks, where the method shows limited improvement on several tasks. Such observations are in line with our preliminary analysis: modeling spurious correlation between tasks using covariate shift is in general insufficient and lacks theoretical guarantee. Furthermore, GMTL [27] is limited to classification tasks, and its dependence on estimating" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25704" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "joint distributions may introduce estimation errors, resulting in subpar performance on testing data. In comparison, our method shows better or comparable performance on all four datasets, validating the effectiveness of our method." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.163, + 0.461, + 0.18 + ], + "angle": 0, + "content": "5.3. Empirical Validation of Theoretical Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.185, + 0.484, + 0.336 + ], + "angle": 0, + "content": "We move on to discuss the validation of Theorem 1. Specifically, we consider three different variations of Multi-MNIST dataset: Spurious correlation (Sp.C.), with non-overlapping labeling pair between training and testing data, as in Sec. 5.1; Random assignment (R.A.), where the right digit is randomly assigned given the left digit on both training and testing data; Identical assignment (Id.A.), where the left digit is identical to the right digit on both training and testing data. Under identical assignment, the two tasks are no longer spuriously correlated." + }, + { + "type": "table", + "bbox": [ + 0.182, + 0.349, + 0.394, + 0.422 + ], + "angle": 0, + "content": "
DataρYl,Yr - ρY′l,Y′r
Sp.C.0.29 ± 0.04
R.A.0.002 ± 0.0006
Id.A.0.001 ± 0.0003
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.432, + 0.485, + 0.504 + ], + "angle": 0, + "content": "Table 9. Maximum average change in correlation coefficient under class-wise resampling on Multi-MNIST dataset. \\( Y^{l} \\) and \\( Y^{r} \\) refer to the random variables of left digit label and right digit label on training set, and \\( Y^{\\prime l} \\) and \\( Y^{\\prime r} \\) refer to the random variables of left digit label and right digit label on the resampled training set." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.521, + 0.484, + 0.657 + ], + "angle": 0, + "content": "Under each variation, we perform resampling on the left digit and report the maximum average change in correlation coefficient. The results are shown in Tab. 9. Compared with those under random assignment and identical assignment, when the two tasks are spuriously correlated, the difference in correlation coefficient shows a significant increase, substantiating the feasibility of identifying spurious correlations between tasks based on the difference in correlation coefficients, as proposed in Theorem 1." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.667, + 0.245, + 0.684 + ], + "angle": 0, + "content": "5.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We further invest the effect of spurious correlation identification within our methodology under the three variations of Multi-MNIST dataset, and the necessity in distinguishing spurious correlation between tasks. Results are shown in Tab. 10. Under both spurious correlation and random assignment, the adversarial training framework shows similar performance, regardless of the identification of spurious correlations. Specifically, under random assignment, our framework demonstrates marginal improvement over STL and MTL baselines, irrespective of the construction of \\(\\mathbb{Z}^j\\)’s. This suggests that when the labels of two tasks are not correlated, the per-task predictors are also unlikely to be affected by the spurious features. Furthermore, under identical assignment, without correctly identifying the spurious" + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.088, + 0.916, + 0.291 + ], + "angle": 0, + "content": "
DataMethodleft digitright digit
Sp.C.STL88.26±1.34%88.43±1.54%
Sp.C.MTL84.14±1.19%83.57±1.27%
Sp.C.Ours (w/o Zj)91.59±1.47%90.74±1.22%
Sp.C.Ours91.63±1.46%90.78±1.26%
R.A.STL93.26±1.46%92.87±1.25%
R.A.MTL93.44±1.57%93.67±1.28%
R.A.Ours (w/o Zj)93.48±1.76%93.25±1.29%
R.A.Ours93.87±1.24%93.73±1.36%
Id.A.STL93.67±1.36%93.68±1.37%
Id.A.MTL93.81±1.55%93.49±1.32%
Id.A.Ours (w/o Zj)91.73±1.45%91.85±1.46%
Id.A.Ours93.84±1.43%93.52±1.19%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.301, + 0.907, + 0.44 + ], + "angle": 0, + "content": "Table 10. Experimental results on Multi-MNIST dataset under random and identical assignments. Regarding 'Ours \\((\\mathrm{w / o}\\mathbb{Z}^j)\\) we apply the debiased adversarial training per-task by taking the other task as the spurious task. Particularly, under identical assignment, applying debiased adversarial training without identifying \\(\\mathbb{Z}^j\\)’s leads to slightly worse performance than the baselines, indicating the necessity of distinguishing the spurious correlations between tasks. In comparison, our method performs comparably with the baselines, owing to the identification of spurious correlation between the two tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.468, + 0.907, + 0.545 + ], + "angle": 0, + "content": "information, the adversarial training framework performs slightly worse than the STL and MTL baselines, underscoring the importance of correctly constructing \\(\\mathbb{Z}^j\\)'s, within which our debiased adversarial training framework shall be applied to improve, rather than hinder, the generalizability." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.558, + 0.634, + 0.574 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Spurious correlations in MTL present a significant but inadequately addressed challenge. In this paper, we experimentally show the limitations of existing formulations in mitigating spurious correlations. Motivated by the discussion, we propose a novel framework for identifying spurious correlations between tasks. Our approach formulates the problem from a label-based perspective, where we propose to quantify the spurious correlations between tasks based on the difference in correlation coefficients, evaluated on both the training data and the class-wise resampled training data. We show theoretically the feasibility of our framework in distinguishing spurious correlations between tasks. Based on our identification framework, we propose an adversarial training framework for mitigating between-task spurious correlations, which fine-tunes only the task-specific predictors to ensure that task-specific predictors sufficiently discard the spurious information. We validate from experiments on four benchmark datasets that our method outperforms existing baselines in mitigating spurious correlations between tasks, as well as the effectiveness of our method in identifying spurious correlations between tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "25705" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.484, + 0.177 + ], + "angle": 0, + "content": "This work was partially supported by the EMBRIO Institute, contract #2120200, a National Science Foundation (NSF) Biology Integration Institute, and NSF IIS #1955890, IIS #2146091, IIS #2345235." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.192, + 0.188, + 0.209 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.217, + 0.484, + 0.272 + ], + "angle": 0, + "content": "[1] Ibrahim M Alabdulmohsin, Behnam Neyshabur, and Xiaohua Zhai. Revisiting neural scaling laws in language and vision. Advances in Neural Information Processing Systems, 35:22300-22312, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.274, + 0.484, + 0.344 + ], + "angle": 0, + "content": "[2] Simon Buchholz, Goutham Rajendran, Elan Rosenfeld, Bryon Aragam, Bernhard Schölkopf, and Pradeep Ravikumar. Learning linear causal representations from interventions under general nonlinear mixing. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.346, + 0.484, + 0.4 + ], + "angle": 0, + "content": "[3] Yuan Cao and Quanquan Gu. Generalization bounds of stochastic gradient descent for wide and deep neural networks. Advances in neural information processing systems, 32, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.403, + 0.482, + 0.431 + ], + "angle": 0, + "content": "[4] Rich Caruana. Multitask learning. Machine learning, 28: 41-75, 1997. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.433, + 0.483, + 0.488 + ], + "angle": 0, + "content": "[5] Yimeng Chen, Ruibin Xiong, Zhi-Ming Ma, and Yanyan Lan. When does group invariant learning survive spurious correlations? Advances in Neural Information Processing Systems, 35:7038-7051, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.49, + 0.484, + 0.559 + ], + "angle": 0, + "content": "[6] Zhao Chen, Jiquan Ngiam, Yanping Huang, Thang Luong, Henrik Kretzschmar, Yuning Chai, and Dragomir Anguelov. Just pick a sign: Optimizing deep multitask models with gradient sign dropout. Advances in Neural Information Processing Systems, 33:2039-2050, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.562, + 0.484, + 0.644 + ], + "angle": 0, + "content": "[7] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.647, + 0.484, + 0.702 + ], + "angle": 0, + "content": "[8] Yuyang Deng, Junyuan Hong, Jiayu Zhou, and Mehrdad Mahdavi. On the generalization ability of unsupervised pretraining. In International Conference on Artificial Intelligence and Statistics, pages 4519-4527. PMLR, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.704, + 0.484, + 0.759 + ], + "angle": 0, + "content": "[9] Nikolaos Dimitriadis, Pascal Frossard, and François Fleuret. Pareto manifold learning: Tackling multiple tasks via ensembles of single-task models. In International Conference on Machine Learning, pages 8015-8052. PMLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.761, + 0.483, + 0.816 + ], + "angle": 0, + "content": "[10] Chris Fifty, Ehsan Amid, Zhe Zhao, Tianhe Yu, Rohan Anil, and Chelsea Finn. Efficiently identifying task groupings for multi-task learning. Advances in Neural Information Processing Systems, 34:27503-27516, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[11] Marco Fumero, Florian Wenzel, Luca Zancato, Alessandro Achille, Emanuele Rodolà, Stefano Soatto, Bernhard Schölkopf, and Francesco Locatello. Leveraging sparse and shared feature activations for disentangled representation learning. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 3, 4, 7" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.217, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[12] Jiechao Guan and Zhiwu Lu. Task relatedness-based generalization bounds for meta learning. In International Conference on Learning Representations, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.905, + 0.192 + ], + "angle": 0, + "content": "[13] Yuzheng Hu, Ruicheng Xian, Qilong Wu, Qiuling Fan, Lang Yin, and Han Zhao. Revisiting scalarization in multi-task learning: A theoretical perspective. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.194, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[14] Ziniu Hu, Zhe Zhao, Xinyang Yi, Tiansheng Yao, Lichan Hong, Yizhou Sun, and Ed Chi. Improving multi-task generalization via regularizing spurious correlation. Advances in Neural Information Processing Systems, 35:11450-11466, 2022. 1, 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[15] Badr Youbi Idrissi, Martin Arjovsky, Mohammad Pezeshki, and David Lopez-Paz. Simple data balancing achieves competitive worst-group-accuracy. In Conference on Causal Learning and Reasoning, pages 336–351. PMLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.321, + 0.905, + 0.387 + ], + "angle": 0, + "content": "[16] Junguang Jiang, Baixu Chen, Junwei Pan, Ximei Wang, Dapeng Liu, Jie Jiang, and Mingsheng Long. Forkmerge: Mitigating negative transfer in auxiliary-task learning. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[17] Polina Kirichenko, Pavel Izmailov, and Andrew Gordon Wilson. Last layer re-training is sufficient for robustness to spurious correlations. arXiv preprint arXiv:2204.02937, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.448, + 0.905, + 0.503 + ], + "angle": 0, + "content": "[18] Tyler LaBonte, Vidya Muthukumar, and Abhishek Kumar. Towards last-layer retraining for group robustness with fewer annotations. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.505, + 0.905, + 0.588 + ], + "angle": 0, + "content": "[19] Sébastien Lachapelle, Tristan Deleu, Divyat Mahajan, Ioannis Mitliagkas, Yoshua Bengio, Simon Lacoste-Julien, and Quentin Bertrand. Synergies between disentanglement and sparsity: Generalization and identifiability in multi-task learning. In International Conference on Machine Learning, pages 18171-18206. PMLR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.59, + 0.905, + 0.631 + ], + "angle": 0, + "content": "[20] Andrew K Lampinen and Surya Ganguli. An analytic theory of generalization dynamics and transfer learning in deep linear networks. arXiv preprint arXiv:1809.10374, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.633, + 0.905, + 0.685 + ], + "angle": 0, + "content": "[21] Dongyue Li, Huy L Nguyen, and Hongyang R Zhang. Identification of negative transfers in multitask learning using surrogate models. arXiv preprint arXiv:2303.14582, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.689, + 0.905, + 0.744 + ], + "angle": 0, + "content": "[22] Weixin Liang and James Zou. Metashift: A dataset of datasets for evaluating contextual distribution shifts and training conflicts. arXiv preprint arXiv:2202.06523, 2022. 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.747, + 0.905, + 0.788 + ], + "angle": 0, + "content": "[23] Bo Liu, Yihao Feng, Peter Stone, and Qiang Liu. Famo: Fast adaptive multitask optimization. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.79, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[24] Evan Z Liu, Behzad Haghgoo, Annie S Chen, Aditi Raghunathan, Pang Wei Koh, Shiori Sagawa, Percy Liang, and Chelsea Finn. Just train twice: Improving group robustness without training group information. In International Conference on Machine Learning, pages 6781-6792. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.907, + 0.903 + ], + "angle": 0, + "content": "[25] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25706" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "International Conference on Computer Vision (ICCV), 2015. 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[26] Shenyu Lu, Junyi Chai, and Xiaoqian Wang. Neural collapse inspired debiased representation learning for min-max fairness. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 2048-2059, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[27] Taro Makino, Krzysztof Geras, and Kyunghyun Cho. Generative multitask learning mitigates target-causing confounding. Advances in Neural Information Processing Systems, 35:36546-36558, 2022. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.305 + ], + "angle": 0, + "content": "[28] Michinari Momma, Chaosheng Dong, and Jia Liu. A multi-objective/multi-task learning framework induced by pareto stationarity. In International Conference on Machine Learning, pages 15895-15907. PMLR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.306, + 0.482, + 0.374 + ], + "angle": 0, + "content": "[29] Aviv Navon, Aviv Shamsian, Idan Achituve, Haggai Maron, Kenji Kawaguchi, Gal Chechik, and Ethan Fetaya. Multitask learning as a bargaining game. In International Conference on Machine Learning, pages 16428-16446. PMLR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.376, + 0.482, + 0.404 + ], + "angle": 0, + "content": "[30] Karl Pearson. Correlation coefficient. In *Royal Society Proceedings*, page 214, 1895. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.482, + 0.446 + ], + "angle": 0, + "content": "[31] Elan Rosenfeld, Pradeep Ravikumar, and Andrej Risteski. The risks of invariant risk minimization. arXiv preprint arXiv:2010.05761, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.449, + 0.482, + 0.516 + ], + "angle": 0, + "content": "[32] Shiori Sagawa, Pang Wei Koh, Tatsunori B Hashimoto, and Percy Liang. Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization. arXiv preprint arXiv:1911.08731, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.482, + 0.586 + ], + "angle": 0, + "content": "[33] Shiori Sagawa, Aditi Raghunathan, Pang Wei Koh, and Percy Liang. An investigation of why overparameterization exacerbates spurious correlations. In International Conference on Machine Learning, pages 8346-8356. PMLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.482, + 0.631 + ], + "angle": 0, + "content": "[34] Ozan Sener and Vladlen Koltun. Multi-task learning as multi-objective optimization. Advances in neural information processing systems, 31, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.482, + 0.716 + ], + "angle": 0, + "content": "[35] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part V 12, pages 746-760. Springer, 2012. 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.482, + 0.785 + ], + "angle": 0, + "content": "[36] Trevor Standley, Amir Zamir, Dawn Chen, Leonidas Guibas, Jitendra Malik, and Silvio Savarese. Which tasks should be learned together in multi-task learning? In International conference on machine learning, pages 9120-9132. PMLR, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.788, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[37] Charles Stein et al. Inadmissibility of the usual estimator for the mean of a multivariate normal distribution. In Proceedings of the Third Berkeley symposium on mathematical statistics and probability, pages 197-206, 1956. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[38] Shao-Hua Sun. Multi-digit mnist for few-shot learning, 2019. 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[39] Nilesh Tripuraneni, Michael Jordan, and Chi Jin. On the theory of transfer learning: The importance of task diver" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "sity. Advances in neural information processing systems, 33: 7852-7862, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[40] Simon Vandenhende, Stamatios Georgoulis, Wouter Van Gansbeke, Marc Proesmans, Dengxin Dai, and Luc Van Gool. Multi-task learning for dense prediction tasks: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(7):3614-3633, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[41] Yipei Wang and Xiaoqian Wang. On the effect of key factors in spurious correlation: A theoretical perspective. In International Conference on Artificial Intelligence and Statistics, pages 3745-3753. PMLR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.304 + ], + "angle": 0, + "content": "[42] Zirui Wang, Yulia Tsvetkov, Orhan Firat, and Yuan Cao. Gradient vaccine: Investigating and improving multi-task optimization in massively multilingual models. arXiv preprint arXiv:2010.05874, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[43] Enneng Yang, Zhenyi Wang, Li Shen, Shiwei Liu, Guibing Guo, Xingwei Wang, and Dacheng Tao. Adamerging: Adaptive model merging for multi-task learning. arXiv preprint arXiv:2310.02575, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[44] Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient surgery for multi-task learning. Advances in Neural Information Processing Systems, 33:5824-5836, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.419, + 0.905, + 0.488 + ], + "angle": 0, + "content": "[45] Amir R Zamir, Alexander Sax, William Shen, Leonidas J Guibas, Jitendra Malik, and Silvio Savarese. Taskonomy: Disentangling task transfer learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3712-3722, 2018. 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.905, + 0.53 + ], + "angle": 0, + "content": "[46] Yu Zhang and Qiang Yang. A survey on multi-task learning. IEEE transactions on knowledge and data engineering, 34 (12):5586-5609, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.532, + 0.905, + 0.6 + ], + "angle": 0, + "content": "[47] Xiao Zhou, Yong Lin, Renjie Pi, Weizhong Zhang, Renzhe Xu, Peng Cui, and Tong Zhang. Model agnostic sample reweighting for out-of-distribution learning. In International Conference on Machine Learning, pages 27203-27221. PMLR, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.6 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25707" + } + ] +] \ No newline at end of file diff --git a/2025/Identifying and Mitigating Spurious Correlation in Multi-Task Learning/67faa86d-7895-441b-b372-45b9fee04763_origin.pdf b/2025/Identifying and Mitigating Spurious Correlation in Multi-Task Learning/67faa86d-7895-441b-b372-45b9fee04763_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6449a39da9e14f1d43319ddb808f4c6870b8c7d5 --- /dev/null +++ b/2025/Identifying and Mitigating Spurious Correlation in Multi-Task Learning/67faa86d-7895-441b-b372-45b9fee04763_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:498ae9a928c5a32b06f2bca54f61964dafbd54fcc8cbf5f2eb41b65cbacba6a5 +size 2003638 diff --git a/2025/Identifying and Mitigating Spurious Correlation in Multi-Task Learning/full.md b/2025/Identifying and Mitigating Spurious Correlation in Multi-Task Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..51edf60d8e90ff24ee81efd1b84828b5b0b65bfe --- /dev/null +++ b/2025/Identifying and Mitigating Spurious Correlation in Multi-Task Learning/full.md @@ -0,0 +1,335 @@ +# Identifying and Mitigating Spurious Correlation in Multi-Task Learning + +Junyi Chai, Shenyu Lu & Xiaoqian Wang* +Elmore Family School of Electrical and Computer Engineering, Purdue University +West Lafayette, IN, USA + +{chai28,lu876,joywang}@purdue.edu + +# Abstract + +Multi-task learning (MTL) is a paradigm that aims to improve the generalization of models by simultaneously learning multiple related tasks, leveraging shared representations and task-specific information to enhance performance on individual tasks. However, existing work has shown that MTL can potentially hinder generalization, with one key factor being spurious correlations between tasks. Owing to the knowledge-sharing property, the per-task predictors are more likely to develop reliance on spurious features. Most existing approaches address this issue through distributional robustness, aiming to maintain consistent performance across different distributions under unknown covariate shifts. However, this formulation lacks theoretical guarantees and can be sensitive to the construction of covariate shifts. In this work, we propose a novel perspective, where we seek to identify spurious correlations between tasks. Drawing inspirations from conventional formulations on spurious correlation, for each task, we propose to distinguish its spurious tasks using the difference in correlation coefficients between the empirical distribution and class-wise resampled distributions, thereby capturing the correlations between task labels w.r.t. each class. We prove theoretically the feasibility of the resampling strategy in characterizing spurious correlations between tasks. Furthermore, we propose a simple fine-tuning strategy, debiased adversarial training, where the per-task predictors are adversarily trained to disregard information associated with their spurious tasks. Experimental results on six benchmark datasets show that our method effectively mitigates spurious correlations and outperforms state-of-the-art methods in improving generalization. + +# 1. Introduction + +Stemming from Stein's paradox [37] where the estimation error decreases under joint estimation, formalized in [4], + +multi-task learning (MTL) has since become increasingly important in various applications, particularly in computer vision, demonstrating both improved accuracy and reduced computational overhead. MTL aims to jointly learn multiple tasks, allowing each task to leverages the shared representations and domain knowledge between tasks, with the goal of improving the generalization performance across all tasks [19, 46]. Compared with single-task learning, which focuses on optimizing over one specific objective, MTL extracts common patterns that are beneficial across different but related objectives, potentially leading to better performance on individual tasks. This collaborative framework mitigates overfitting and enhances model robustness through cross-task knowledge transfer [28, 29]. + +Despite the benefits in MTL, existing work has demonstrated that under certain circumstances, MTL could lead to degraded performance in generalization, particularly when different tasks do not experience high knowledge-sharing [10, 36]. This abnormality has conventionally been attributed to the model capacity issue, where conflicting tasks may compete during optimization [42, 44]. However, Recent studies have revealed a counterintuitive phenomenon in MTL, where even under over-parametrization, a MTL model might excel in minimizing training loss, yet surprisingly underperform single-task models in terms of generalization [14]. Accordingly, spurious correlations between different tasks have been identified as an intrinsic problem that hurts generalization in MTL [14, 27]. Owing to the multi-tasking property, the shared representations are expected to encode knowledge across tasks; consequently, when two tasks exhibit high label correlations despite having unrelated underlying knowledge, the per-task predictors may erroneously leverage unrelated spurious knowledge, potentially compromising generalization. + +Most of existing work addressing spurious correlations in MTL approaches this issue through the lens of invariant learning, aiming to ensure that models perform consistently across various empirical distributions [11, 14]. Specifically, the shared representations are expected to recover the ground-truth latent causal features, and by en + +forcing distributional-invariance, the per-task predictors are expected to leverage features that are causally related to each task, instead of spurious features. Nonetheless, on one hand, counterfactual data is generally inaccessible on high-dimensional data, leading to the infeasibility in conducting causality discovery [2]. Consequently, the shared representations may fail to recover the underlying latent features, particularly when approximated by disentanglement across latent dimensions [11, 14]. Moreover, enforcing distributional invariance does not guarantee utilization of only the task-specific causal features, as the difference across empirical distributions does not necessarily break the spurious correlations between tasks [5]. In the absence of prior knowledge regarding spurious correlations between tasks, along with the potential lack of certain label pairings, it becomes difficult to construct the desired distributions. Alternatively, Makino et al. [27] proposes to mitigate the spurious correlations between tasks by generatively estimating the joint distribution of task labels. Despite theoretical soundness of the formulation, estimating joint distributions can be error-prone, especially when certain label combinations are missing. Additionally, estimating joint distributions involving labels of varied dimensions is often infeasible. This issue is particularly relevant for vision tasks, where labels from different tasks are likely to have differing dimensions. + +In this work, we propose an alternative formulation regarding spurious correlations in MTL. Instead of imposing regularizations on distributional invariance, we seek to directly identify the spurious correlations between tasks. Drawing inspirations from conventional formulations of spurious correlation, typically defined by the dependency between labels and spurious information, we propose to identify spurious correlations between tasks using correlation coefficients. Specifically, we detect spurious correlations between two tasks by examining differences in their correlation coefficients, calculated on both the original training data and class-wise resampled training data. We prove theoretically the feasibility of our formulation in identifying spurious correlations between tasks, as well as in distinguishing spurious correlations from genuine task relationships. Building upon the per-task identification of spurious correlations, we formulate the bias mitigation problem as a mini-max game, where for each task, we consider the predictors of its spurious tasks as the adversaries, such that by fine-tuning the task-specific projector and the task-specific predictor, we intend to minimize the training loss for this task, while maximizing the training loss for its spurious tasks under projection. + +We summarize our contribution as follows: + +- We propose a theoretical framework for detecting spurious correlations in MTL by analyzing intra-task label dependencies, and we prove theoretically the feasibility of + +our method in the identification. + +- We address spurious correlations in multi-task learning by fine-tuning task-specific predictors, enhancing per-task performance through adversarial training that minimizes loss on the primary task while maximizing loss on its identified spurious tasks. +- We validate from experiments the effectiveness of our framework in identification of spurious correlations between tasks, as well as in improving generalizability on four benchmark datasets. + +# 2. Related Work + +# 2.1. Generalization in multi-task learning + +The generalizability of deep neural networks has often been linked to the availability of large training datasets, making sample sparsity one of the fundamental limitations for model performance [1, 3, 8]. MTL addresses sample sparsity by leveraging labeled data across multiple related tasks, effectively expanding the available training data [13, 46]. Research on multi-task generalization predominantly assumes shared knowledge between tasks, where learning across diverse tasks enables more robust and transferable representations [12, 20, 39]. Despite empirical validations of the assumption [40], the fundamental assumption of task relatedness often fails in real-world applications, where competing tasks may vie for model capacity, potentially degrading MTL performance below that of single-task approaches [10, 36]. + +Prior research has addressed this challenge through two distinct perspectives. One line of work has focused on the model capacity issue, seeking equilibrium between conflicting task objectives. Yang et al. [43] proposes to adaptively reweight per-task training loss, leveraging entropy minimization on unlabeled test samples. Jiang et al. [16] proposes to dynamically merges model branches while optimizing task weights based on the performance on validation set. Simliar reweighing strategies have also been applied in [21, 23]. Alternatively, Chen et al. [6] introduces a selective gradient filtering approach to resolve optimization conflicts, while Dimitriadis et al. [9] and Sener and Koltun [34] propose to formulate the problem from the perspective of Pareto optimality to achieve optimal trade-off between conflicting tasks. In contrast, recent work has focused on the exacerbation of spurious correlations under MTL, where the multi-tasking property could exacerbate reliance on spurious features for each task owing to the interaction between tasks [11, 14, 27]. Specifically, Hu et al. [14] and Fumero et al. [11] address intra-task spurious correlations through distributional invariance, while Makino et al. [27] employs generative modeling of joint distributions to mitigate these correlations. + +# 2.2. Spurious correlation + +Spurious correlation refers to misleading patterns that appear to hold in training data but fail to generalize to new, unseen data [32, 33]. Existing work on spurious correlations typically aims to reduce the error for the worst-performing group, characterized by both the task label and the associated spurious information. typically, these methods can be divided into two categories: data-augmentation, where the training distribution is modified to enhancing worst-group performance [5, 8, 15]; model-regularization, where regularizations are imposed during the training process to eliminate the affect of spurious information [24, 26, 32, 41, 47]. More recently, last-layer retraining has been proposed as an efficient alternative for addressing spurious correlations [17, 18]. In this approach, instead of updating all network parameters, only the classifier layer is fine-tuned on a balanced validation set, significantly reducing computational complexity. + +![](images/89d6d7c033b92f9627bb3834710605157dc10b64834c7d193758bdf5f80afc67.jpg) +Figure 1. Our framework improves generalizability by identifying and correcting spurious correlations between tasks. For instance, on CelebA dataset, the task labels 'male' and 'attractiveness' are known to be spuriously correlated [26]. Compared with vanilla MTL where predictions for attractiveness falsely correlate with predictions for gender, our method effectively mitigates this bias. + +# 3. Preliminaries + +Consider $k$ tasks in total. Let $\mathcal{D}^{\mathrm{train}} := \{(x_i, y_i^1, \ldots, y_i^k), 1 \leq i \leq N\}$ be the training set where for $i$ -th sample, $x_i \in \mathbb{R}^d$ is the input feature and $y_i^j \in \mathbb{R}^{d^j}$ is the label for $j$ -th task, let $\pmb{f}: \mathbb{R}^d \to \mathbb{R}^l$ be the function of shared encoder and $\pmb{g}^j: \mathbb{R}^l \to \mathbb{R}^{d^j}$ be the function of predictor for $j$ -th task, let $\mathcal{L}^j$ be the training loss for $j$ -th task, a multi-task learning problem can be formulated as + +$$ +\underset {\boldsymbol {f}, \boldsymbol {g} ^ {1}, \dots , \boldsymbol {g} ^ {k}} {\arg \min } \sum_ {j = 1} ^ {k} w ^ {j} \bar {\mathcal {L}} ^ {j} (\mathcal {D} ^ {\text {t r a i n}}; \boldsymbol {g} ^ {j}, \boldsymbol {f}), +$$ + +where $\bar{\mathcal{L}}^j (\mathcal{D}^{\mathrm{train}};\pmb {g}^j,\pmb {f})\coloneqq \frac{1}{N}\sum_{i = 1}^N\mathcal{L}^j (\pmb {g}^j (\pmb {f}(x_i)),y_i^j)$ is the average training loss for $j$ -th task, and $w^{j}$ specifies + +the weight of $\bar{\mathcal{L}}^j$ during optimization. Despite the benefit of exploring task-relatedness, which are expected to benefit the generalizability, recent work on MTL has pointed to the potential risk of spurious correlations between tasks owing to the confounder between task labels [14, 27]. Let $\{\mathcal{F}_i, 1 \leq i \leq K\}$ be the set of underlying latent features such that $\mathcal{F}_i \perp \mathcal{F}_j$ , $\forall i \neq j$ , let $\mathcal{F}^j$ be the set of underlying latent features that are causally related with $j$ -th task, when $\mathcal{F}^i \neq \mathcal{F}^j$ , optimizing $\bar{\mathcal{L}}^i$ and $\bar{\mathcal{L}}^j$ jointly could exacerbate reliance on spurious features from $\mathcal{F}^j$ for $g^i$ , as the shared encoder needs to explicitly encode knowledge for both tasks. Such confounder brings in spurious correlations between tasks and can harm generalization performance. + +Nonetheless, compared with conventional formulations where the spurious information is assumed to be available on training data [32, 33], regarding MTL, we do not have the spurious information at hand, where each task may consider all other tasks as being either spurious or beneficial. Alternatively, most existing work [11, 14] has formulated the problem from the perspective of distributional robustness, where the model is expected to perform consistently on both the training data and the validation data $\mathcal{D}^{\mathrm{val}}$ under some unknown covariate shift: + +$$ +\begin{array}{l} \underset {\boldsymbol {f}, \boldsymbol {g} ^ {1}, \dots , \boldsymbol {g} ^ {k}} {\arg \min } \sum_ {j = 1} ^ {k} w ^ {j} \bar {\mathcal {L}} ^ {j} (\mathcal {D} ^ {\text {t r a i n}}; \boldsymbol {g} ^ {j}, \boldsymbol {f}) + \sum_ {j = 1} ^ {k} w ^ {\text {v a l}, j} \bar {\mathcal {L}} ^ {j} (\mathcal {D} ^ {\text {v a l}}; \boldsymbol {g} ^ {j}, \boldsymbol {f}), \\ s. t. R \left(\mathcal {D} ^ {\text {t r a i n}}; f\right) \leq \epsilon , R \left(\mathcal {D} ^ {\text {v a l}}; f\right) \leq \epsilon , \tag {1} \\ \end{array} +$$ + +where $w^{\mathrm{val},j}$ specifies the weight of validation loss for $j$ -th task, and $R(\mathcal{D}^{\mathrm{train}};f)$ and $R(\mathcal{D}^{\mathrm{val}};f)$ correspond to the regularization on $f$ such that the encoder is dimensional-wise disentangled. For example, MT-CRL [14] employs covariance between latent dimensions as the regularization, while Meta-learning [11] formulates $R(\mathcal{D}^{\mathrm{train}};f)$ as the entropy of normalized representations per dimension. Eq. (1) stems from the assumption that for $i$ -th task, the covariate shift on validation data effectively disrupts the correlation between its spurious features and its task labels, which could be infeasible in general. On one hand, without prior knowledge on the task-specific causal features, there's no guarantee that the covariate shift breaks the spurious correlation between tasks. Furthermore, the effectiveness of Eq. (1) may be influenced by the degree of bias in the validation set, introducing uncertainty in the mitigation spurious correlations. + +We use the following experiments on linear SEM [31] to validate the two perspectives. Following existing work [14], we construct two binary classification tasks where the causal features of each task are independent and follow a Gaussian distribution. Let $Y^{j}$ be the random variable of labels for $j$ -th task, we set the ratios between task labels on training data as $\alpha_{y}^{\text{train}} = \frac{P(Y^{0} = y, Y^{1} = (1 - y))}{P(Y^{0} = y, Y^{1} = y)} = \frac{1}{9}$ , $\forall y \in \{0, 1\}$ on training data. Regarding both tasks, we have + +$P(Y^{j} = 0) = P(Y^{j} = 1) = 0.5, \forall j \in \{0,1\}$ , and the testing data is balanced sampled across each co-couurance of task labels such that $\alpha_{y}^{\mathrm{test}} = \frac{P(Y^{0} = y, Y^{1} = (1 - y))}{P(Y^{0} = y, Y^{1} = y)} = 1$ , $\forall y \in \{0,1\}$ . Smaller $\alpha_{y}$ indicates higher correlation between the two task labels. + +
TaskMTL (αtrainy = 1/9)MTL (αtrainy = 1)
Task 081.32±0.29%94.47±0.34%
Task 181.37±0.41%94.47±0.35%
+ +Table 1. Experimental results on synthetic data. We choose classification accuracy as the evaluation metric for both tasks. Under high correlation, the model performs remarkably worse than that under no spurious correlation. + +
TaskMT-CRL (αyval = 1/9)Meta-learning
Task 081.59±0.36%81.73±0.35%
Task 181.87±0.95%81.67±0.44%
+ +Table 2. Experimental results on synthetic data. Both methods show marginal improvement in testing accuracy compared with MTL ( $\alpha_{y}^{\mathrm{train}} = \frac{1}{9}$ ): $81.32 \pm 0.29\%$ for task 0 and $81.37 \pm 0.41\%$ for task 1, owing to the consistency in $\alpha_{y}$ 's between training and validation data. + +
TaskMT-CRL (αyval = 1/5)MT-CRL (αyval = 1)
Task 083.74±0.65%87.59±0.74%
Task 183.32±0.59%88.16±0.46%
+ +Table 3. Experimental results on synthetic data. The testing performance improves as $\alpha_{y}^{\mathrm{val}}$ 's decrease, substantiating the importance of mitigating bias on the validation data. Nonetheless, given the balanced validation set, MT-CRL still performs worse than the optimal baseline MTL ( $\alpha_{y}^{\mathrm{train}} = 1$ ): $94.47 \pm 0.34\%$ for task 0 and $94.47 \pm 0.35\%$ for task 1, highlighting the insufficiency in formulating spurious correlations via distributional robustness. + +We show results of under different constructions of validation sets in Tab. 1 - 2. Regarding MT-CRL [14], as the validation set is only assumed to contain unknown shift, we consider three different types of covariate shifts: The training-validation split where the validation set is sampled from the training distribution, as suggested by the authors as a general construction of covariate shift, where $\alpha_{y}$ 's remain the same on both training and validation data; A less biased validation set where $\alpha_{y}^{\mathrm{val}} = \frac{1}{5}, \forall y \in \{0,1\}$ ; A balanced validation set where $\alpha_{y}^{\mathrm{val}} = 1, \forall y \in \{0,1\}$ . While a balanced validation set shows improvement in the performance on testing data, the first two constructions of validation data show marginal improvement compared with vanilla MTL, validating our claim that the effectiveness of distributional invariance can be subject to the bias on validation data. Nonetheless, in reality, it is hard to obtain or construct a + +perfectly balanced data owing to the uncertainty in the spurious correlations between tasks. Furthermore, it is possible that certain labeling pairs are not observed on training or validation data, leading to infeasibility of constructing such balanced data. + +Regarding Meta-learning [11], since the validation set is constructed per task such that for $j$ -th task, $P(Y^{j} = 0) = P(Y^{j} = 1) = 0.5$ , $\forall j \in \{0,1\}$ , we have $\alpha_{y}$ 's remain the same on both training and validation data. As a result, such construction fails to show improvement in mitigating spurious information compared with the ERM baseline. This demonstrates the flaws in formulating the covariate shift via class-balance: the spurious correlation between tasks does not necessarily lead to imbalance in the marginal distribution of labels per task, where using class-balanced resampling can lead to marginal variations in the conditional distributions $P(Y_{i}|Y_{j})$ of task labels. + +# 4. Method + +# 4.1. Identification of spurious correlation between tasks + +Our preliminary discussion suggests the necessity of identifying spurious correlations between tasks. In light of the intractability in identifying the latent causal features for each task, particularly regarding high-dimensional vision data, we instead consider modeling the spurious correlations from the label perspective. Existing work [14, 26] has demonstrated that when two tasks are spuriously correlated, their corresponding task labels will be disproportionately paired, leading to an uneven joint distribution across task labels. Specifically, let $\mathbb{S}^j$ be the set of values of labels for $j$ -th task, when spurious correlation exists between $i$ -th and $j$ -th tasks, there exist $y, y' \in \mathbb{S}^i$ and $\tilde{y}, \tilde{y}' \in \mathbb{S}^j$ , such that the following inequality regarding label distribution holds: + +$$ +P (Y ^ {i} = y, Y ^ {j} = \tilde {y}) > P (Y ^ {i} = y ^ {\prime}, Y ^ {j} = \tilde {y} ^ {\prime}). \quad (2) +$$ + +To quantitatively evaluate Eq. (2), we observe that the correlation between task labels can be quantified using the Pearson correlation coefficient [30]: + +$$ +\rho_ {Y ^ {i}, Y ^ {j}} = \frac {\mathbb {E} [ Y ^ {i} Y ^ {j} ] - \mathbb {E} [ Y ^ {i} ] \mathbb {E} [ Y ^ {j} ]}{\sqrt {\mathbb {E} [ (Y ^ {i}) ^ {2} ] - (\mathbb {E} [ Y ^ {i} ]) ^ {2}} \sqrt {\mathbb {E} [ (Y ^ {j}) ^ {2} ] - (\mathbb {E} [ Y ^ {j} ]) ^ {2}}}. +$$ + +For simplicity of discussion, consider both tasks as binary classification where $Y^{i}$ and $Y^{j}$ follow Bernoulli distributions with expectations $p^{i}$ and $p^{j}$ , respectively. We refer to Algorithm 1 for extension to multi-class scenario. The following lemma states the close-form expression of $\rho_{Y^{i},Y^{j}}$ : + +Lemma 1. Let $\alpha_{y}^{ij} := \frac{P(Y^i = y, Y^j = 0)}{P(Y^i = y, Y^j = 1)}$ be the ratio of joint probabilities of $Y^i$ and $Y^j$ evaluated w.r.t. class $y$ of $i$ -th + +![](images/78528c4fe56ed138b42b8dff1ffcff491b6b4a70bfcf4e82965c64fa2dd36c37.jpg) +Figure 2. Demonstration of our debiased adversarial training framework. For each task, the predictors within the blue rectangle represent those identified as spurious by our method. During optimization, the per-task projector and predictor are updated to ensure that the projected representation minimizes reliance on information of spurious tasks while maintaining performance on the target task. + +task, we have $\rho_{Y^i,Y^j}$ as + +$$ +\rho_ {Y ^ {i}, Y ^ {j}} = \frac {\left(\frac {1}{(1 + \alpha_ {1} ^ {i j})} - p ^ {j}\right) p ^ {i}}{\sqrt {p ^ {i} p ^ {j} (1 - p ^ {i}) (1 - p ^ {j})}}. +$$ + +We defer full proof to the Appendix. As shown in Lemma 1, the correlation coefficient between $Y^{i}$ and $Y^{j}$ is determined by the marginal distributions of $Y^{i}$ and $Y^{j}$ and the ratio of joint probabilities. However, the dependence of $\rho_{Y^i,Y^j}$ on the marginal distributions could introduce potential errors in identifying spurious correlations, particularly in the presence of label imbalance, where the disproportionality between labeling pairs could be overlooked given small marginal probabilities of task labels. For example, even with a small $\alpha_{1}^{ij}$ , when $p^i$ becomes trivial due to class imbalance, we may observe $\rho_{Y^i,Y^j}\approx 0$ , despite the presence of spurious correlations between the two tasks. We therefore consider the following class-wise resampling strategy to perturb the marginal distributions of task labels, such that when $p^i$ and $p^j$ change, we have $\rho_{Y^i,Y^j}$ change, which enables us to recognize the spurious correlation between $Y^{i}$ and $Y^{j}$ while mitigating the influence of label imbalance. Specifically, let $Y^{\prime i}$ and $Y^{\prime j}$ be the random variable of task labels after resampling, let $\mathcal{D}_{Y^j\neq y}^{\mathrm{train}}\coloneqq \{(x_i,y_i^1,\dots ,y_i^k)|y_i^j\neq y\}$ be the subset of samples excluding class $y$ of $j$ -th task and $\mathcal{D}_{Y^{\prime j}\neq y}^{\mathrm{train}}$ be the corresponding subset after resampling, we consider the following down-sampling scheme within $\mathcal{D}_{Y^j\neq y}^{\mathrm{train}}$ : + +$$ +\left| \mathcal {D} _ {Y ^ {\prime j} \neq y} ^ {' \mathrm {t r a i n}} \right| = \xi \left| \mathcal {D} _ {Y ^ {j} \neq y} ^ {\mathrm {t r a i n}} \right|, +$$ + +where $\xi < 1$ is the constant that determines the degree of resampling, and $\mathcal{D}_{Y^{\prime j}\neq y}^{\prime \mathrm{train}}$ is constructed by random sampling. Under class-wise resampling, we intend to amplify the effect of correlations between task labels under different values of $Y^{\prime j}$ within $\mathbb{S}^j$ , so as to properly capture the dependence between task labels. Let $p^{\prime i}$ be the expectation of $Y^{\prime i}$ evaluated on the resampled set $\mathcal{D}_{Y^{\prime j}\neq y}^{\prime \mathrm{train}}\cup \left(\mathcal{D}^{\mathrm{train}}\backslash \mathcal{D}_{Y^{\prime j}\neq y}^{\mathrm{train}}\right)$ , the following theorem states the feasibility of identifying spurious correlations between tasks based on the difference in correlation coefficients before and after resampling: + +Theorem 1. When spurious correlation exists between two tasks, the maximum difference in correlation coefficient before and after resampling is lower-bounded by the difference in expectations $p^i - p'^i$ , up to multiplicative constants: + +$$ +\left| \rho_ {Y ^ {i}, Y ^ {j}} - \rho_ {Y ^ {\prime i}, Y ^ {\prime j}} \right| \geq \mathcal {K} \left| p ^ {' i} - p ^ {i} \right|, +$$ + +$$ +w h e r e \mathcal {K} = \frac {\left(\frac {1}{\hat {\alpha} _ {1} ^ {i j}} - \hat {\alpha} _ {1} ^ {i j}\right)}{\sqrt {(1 - p ^ {i}) (1 - (\hat {\alpha} _ {1} ^ {i j} - \hat {\alpha} _ {0} ^ {i j}) p ^ {i} - \hat {\alpha} _ {0} ^ {i j})}} a n d \hat {\alpha} _ {y} ^ {i j} = \frac {1}{1 + \alpha_ {y} ^ {i j}}. +$$ + +We defer full proof to the Appendix. As shown in Theorem 1, since $\hat{\alpha}_{y}^{ij}$ and $p^i$ remain constant before and after resampling, we have the lower-bound of difference in correlation coefficient before and after resampling monotonically increase as $p^{'i}$ increases. Accordingly, even though $p^i$ can be trivial, so long as $p^i$ and $p^{'i}$ differ significantly before and after resampling, we are guaranteed a substantial change in the correlation coefficient, which enables us to identify the spurious correlations among tasks. Furthermore, the following remark states the difference in correla + +tion coefficient when spurious correlations do not exist between tasks: + +Remark 1. When two tasks are not spuriously correlated, we have $\rho_{Y^i,Y^j} - \rho_{Y'^i,Y'^j} = 0$ . + +Consequently, let $\mathbb{Z}^j$ be the set of indices of identified spurious tasks for $j$ -th task, we have + +$$ +\left| \rho_ {Y ^ {i}, Y ^ {j}} - \rho_ {Y ^ {\prime i}, Y ^ {\prime j}} \right| > \delta , \forall i \in \mathbb {Z} ^ {j}, +$$ + +where $\delta$ is the threshold for distinguishing the spurious correlations between tasks. In practice, $\delta$ can be trivially obtained owing to the distinctions in $\left|\rho_{Y^i,Y^j} - \rho_{Y'^i,Y'^j}\right|$ as stated in Theorem 1 and Remark 1. We refer to Sec. 5.3 for the detailed discussion. + +# 4.2. Debiased Adversarial Training + +We move on to discuss the mitigation of spurious correlation between tasks. Due to the potential absence of certain label pairs in the training set, which impedes constructing balanced data even after identifying spurious tasks, we instead aim to ensure that the task-specific predictors effectively discard non-causal features for each task. In light of the multi-tasking property, we can readily formulate the objective as an adversarial learning problem w.r.t. the pertask predictors. Let $\hat{f}$ be the function of pre-trained encoder which remains frozen during fine-tuning, regarding $j$ -th task, given its identified set of spurious tasks $\mathbb{Z}^j$ , to utilize the per-task predictors as discriminators of non-causal features, we apply one extra projector $h^j: \mathbb{R}^l \to \mathbb{R}^l$ before $g^j$ , such that the prediction for $j$ -th task is given by $g^i \circ h^i \circ \hat{f}(x)$ . We consider the following objective for mitigating spurious correlations regarding $j$ -th task: + +$$ +\operatorname *{arg min}_{\boldsymbol{h}^{j},\boldsymbol{g}^{j}}\bar{\mathcal{L}}^{j}\left(\mathcal{D}^{\text{train}};\boldsymbol{h}^{j},\boldsymbol{g}^{j}\right) - \sum_{t\in \mathbb{Z}^{j}}\lambda^{j}\bar{\mathcal{L}}^{t}\left(\mathcal{D}^{\text{train}};\boldsymbol{h}^{j},\boldsymbol{g}^{t}\right), +$$ + +where the optimization is performed over $\pmb{h}^j, \pmb{g}^j$ , while $\pmb{g}^t$ remains frozen $\forall t \in \mathbb{Z}^j$ , and $\bar{\mathcal{L}}^j(\mathcal{D}^{\mathrm{train}}; \pmb{h}^j, \pmb{g}^j) := \frac{1}{N} \sum_{i=1}^{N} \mathcal{L}^j(\pmb{g}^j(\pmb{h}^j(\hat{\pmb{f}}(x_i))), y_i^j), \quad \bar{\mathcal{L}}^t(\mathcal{D}^{\mathrm{train}}; \pmb{h}^j, \pmb{g}^t) := \frac{1}{N} \sum_{i=1}^{N} \mathcal{L}^k(\pmb{g}^k(\pmb{h}^j(\pmb{f}(x_i))), y_i^k)$ are the average training loss under projection. By minimizing the objective, we intend to ensure that the projected latent representation retains only the information pertinent to $j$ -th task while shedding the spurious information. We summarize our method in Algorithm 1. + +# 5. Experiments + +# 5.1. Experimental Setup + +We validate our method on the following benchmark datasets: Multi-MNIST dataset [38], CelebA dataset [25], + +Algorithm 1 Debiased adversarial training + +Input: Pre-trained encoder $\hat{f}$ , task-specific predictors $g^{j}$ 's s.t. $1 \leq j \leq k$ , task-specific projectors $h^{j}$ 's, training data $\mathcal{D}^{\mathrm{train}}$ , training epochs $E$ , learning rate $\eta$ , threshold $\delta$ , resampling constant $\xi$ , hyper-parameter $\lambda^{j}$ . + +1: for $j = 1$ to $k$ do: +2: $\mathbb{Z}^j = \emptyset$ +3: for $i = 1$ to $k$ s.t. $i \neq j$ do: +4: Calculate $\rho_{Y^i,Y^j}$ on $\mathcal{D}^{\mathrm{train}}$ . +5: for $y\in \mathbb{S}^j$ do: +6: Perform class-wise resampling: + +$$ +\left| \mathcal {D} _ {Y ^ {\prime j} \neq y} ^ {\text {t r a i n}} \right| = \xi \left| \mathcal {D} _ {Y ^ {j} \neq y} ^ {\text {t r a i n}} \right|. +$$ + +7: Calculate $\rho_{Y^{\prime}i,Y^{\prime}j}$ on + +$$ +\mathcal {D} _ {Y ^ {\prime j} \neq y} ^ {\text {t r a i n}} \cup \left(\mathcal {D} ^ {\text {t r a i n}} \backslash \mathcal {D} _ {Y ^ {j} \neq y} ^ {\text {t r a i n}}\right). +$$ + +8: if $\rho_{Y^i,Y^j} - \rho_{Y'^i,Y'^j}\geq \delta$ and $i\notin \mathbb{Z}^j$ then: +9: $\mathbb{Z}^j = \mathbb{Z}^j\cup \{i\}$ +10: end if +11: end for +12: end for +13: end for +14: for epoch $= 1$ to $E$ do: +15: for $\mathbf{j} = 1$ to $k$ do: +16: Training loss: + +$$ +\mathcal {L} = \bar {\mathcal {L}} ^ {j} \left(\mathcal {D} ^ {\text {t r a i n}}; \boldsymbol {h} ^ {j}, \boldsymbol {g} ^ {j}\right) - \sum_ {t \in \mathbb {Z} ^ {j}} \lambda^ {j} \bar {\mathcal {L}} ^ {t} \left(\mathcal {D} ^ {\text {t r a i n}}; \boldsymbol {h} ^ {j}, \boldsymbol {g} ^ {t}\right). +$$ + +17: Update $h^j$ and $g^j$ by gradient descent: + +$$ +\boldsymbol {h} ^ {j} = \boldsymbol {h} ^ {j} - \eta \frac {\partial \mathcal {L}}{\partial \boldsymbol {h} ^ {j}}. +$$ + +$$ +\boldsymbol {g} ^ {j} = \boldsymbol {g} ^ {j} - \eta \frac {\partial \mathcal {L}}{\partial \boldsymbol {g} ^ {j}}. +$$ + +18: end for +19: end for + +Taskonomy dataset [45], MetaShift dataset [22], NYU-V2 dataset [35] and CityScape dataset [7]. We implement our method in PyTorch 2.0.0 on one RTX-3090 GPU1. We set $\xi = 0.1$ during experiments and evaluate the difference in correlation coefficients between training set and class-wise resampled sets to construct $\mathbb{Z}^j$ . We use accuracy as evaluation metric for classification tasks, cross-entropy loss and + +
TaskSTLSubSelMT-CRLMeta-learningGMTLOurs
left digit (accuracy, ↑)+4.9%+3.4%+5.1%+0.2%+5.9%+9.3%
right digit (accuracy, ↑)+5.9%+2.9%+6.3%+0.3%+7.2%+9.6%
+ +Table 4. Experimental results on Multi-MNIST dataset. The relative improvement of each method is calculated in comparison to vanilla MTL. + +
TaskSTLSubSelMT-CRLMeta-learningGMTLOurs
Gender (accuracy, ↑)+2.2%+1.7%+1.3%+2.4%+2.6%+5.7%
Smiling (accuracy, ↑)+2.3%+2.5%+2.6%+3.4%+1.2%+4.6%
Age (accuracy, ↑)+1.2%+2.2%+0.1%+1.1%+1.5%+3.9%
Attractiveness (accuracy, ↑)+0.6%+0.4%+0.7%+0.2%+1.2%+2.4%
+ +Table 5. Experimental results on CelebA dataset. + +
TaskSTLSubSelMT-CRLMeta-learningOurs
Scene (accuracy, ↑)+4.1%+3.2%+5.8%+5.3%+9.9%
Object (accuracy, ↑)+5.2%+4.1%+7.3%+8.1%+12.7%
Segmentation (cross entropy, ↓)-5.6%+1.7%+3.1%+6.9%+6.2%
+ +Table 6. Experimental results on Taskonomy dataset. + +
TaskSTLSubSelMT-CRLMeta-learningGMTLOurs
Animal (accuracy, ↑)+4.2%+2.7%+3.2%+1.5%+0.2%+5.4%
Scene (accuracy, ↑)+3.1%+2.1%+1.6%+0.1%+1.7%+4.0%
+ +Table 7. Experimental results on MetaShift dataset. + +
TaskSTLSubSelMT-CRLMeta-learningOurs
Segmentation (mean IoU, ↑)+2.7%+4.2%+6.9%+7.2%+11.3%
Depth (relative error, ↓)+3.1%+3.8%+7.5%+6.2%+10.2%
+ +Table 8. Experimental results on NYU-V2 dataset. + +mean IoU for segmentation tasks and relative error for depth estimation tasks. We defer details of each dataset and full results to the appendix. + +We compare our method with the following related methods on spurious correlation and generalization in MTL: Single-task learning (STL): the network is optimized over each task only without regularizations; MTL: vanilla MTL without regularizations; SubSel [21]: MTL with subset selection to address task conflicts; MT-CRL [14]: MTL with disentangled neural modules and invariant regularization on task-to-module routing; Meta-learning [11]: MTL with disentangled representation learning and bi-level optimization on per-task predictors; GMTL [27]: MTL with generative modeling of task labels. + +We repeat experiments on each dataset three times and report the average results. For MT-CRL [14], the covariate shift on training data is constructed by training-validation split, as suggested by the authors. Further, we do not include GMTL [27] on Taskonomy, NYU-V2 and CityScape + +dataset, as it is infeasible to estimate the joint distribution involving high-dimensional masks. Following prior work [14], we report the relative performance improvement of each method over vanilla MTL for evaluation, as the scales of evaluation metrics differ significantly across tasks. + +# 5.2. Experimental Results + +Results are shown in Tab. 4 - 8. Compared with vanilla STL and MTL, Meta-learning [11] shows marginal improvement in testing performance, particularly when the marginal training distributions are evenly distributed per task. Regarding MT-CRL [14], the training-validation split does not appropriately capture the spurious correlations between tasks, where the method shows limited improvement on several tasks. Such observations are in line with our preliminary analysis: modeling spurious correlation between tasks using covariate shift is in general insufficient and lacks theoretical guarantee. Furthermore, GMTL [27] is limited to classification tasks, and its dependence on estimating + +joint distributions may introduce estimation errors, resulting in subpar performance on testing data. In comparison, our method shows better or comparable performance on all four datasets, validating the effectiveness of our method. + +# 5.3. Empirical Validation of Theoretical Results + +We move on to discuss the validation of Theorem 1. Specifically, we consider three different variations of Multi-MNIST dataset: Spurious correlation (Sp.C.), with non-overlapping labeling pair between training and testing data, as in Sec. 5.1; Random assignment (R.A.), where the right digit is randomly assigned given the left digit on both training and testing data; Identical assignment (Id.A.), where the left digit is identical to the right digit on both training and testing data. Under identical assignment, the two tasks are no longer spuriously correlated. + +
DataρYl,Yr - ρY′l,Y′r
Sp.C.0.29 ± 0.04
R.A.0.002 ± 0.0006
Id.A.0.001 ± 0.0003
+ +Under each variation, we perform resampling on the left digit and report the maximum average change in correlation coefficient. The results are shown in Tab. 9. Compared with those under random assignment and identical assignment, when the two tasks are spuriously correlated, the difference in correlation coefficient shows a significant increase, substantiating the feasibility of identifying spurious correlations between tasks based on the difference in correlation coefficients, as proposed in Theorem 1. + +# 5.4. Ablation Study + +We further invest the effect of spurious correlation identification within our methodology under the three variations of Multi-MNIST dataset, and the necessity in distinguishing spurious correlation between tasks. Results are shown in Tab. 10. Under both spurious correlation and random assignment, the adversarial training framework shows similar performance, regardless of the identification of spurious correlations. Specifically, under random assignment, our framework demonstrates marginal improvement over STL and MTL baselines, irrespective of the construction of $\mathbb{Z}^j$ ’s. This suggests that when the labels of two tasks are not correlated, the per-task predictors are also unlikely to be affected by the spurious features. Furthermore, under identical assignment, without correctly identifying the spurious + +Table 9. Maximum average change in correlation coefficient under class-wise resampling on Multi-MNIST dataset. $Y^{l}$ and $Y^{r}$ refer to the random variables of left digit label and right digit label on training set, and $Y^{\prime l}$ and $Y^{\prime r}$ refer to the random variables of left digit label and right digit label on the resampled training set. + +
DataMethodleft digitright digit
Sp.C.STL88.26±1.34%88.43±1.54%
Sp.C.MTL84.14±1.19%83.57±1.27%
Sp.C.Ours (w/o Zj)91.59±1.47%90.74±1.22%
Sp.C.Ours91.63±1.46%90.78±1.26%
R.A.STL93.26±1.46%92.87±1.25%
R.A.MTL93.44±1.57%93.67±1.28%
R.A.Ours (w/o Zj)93.48±1.76%93.25±1.29%
R.A.Ours93.87±1.24%93.73±1.36%
Id.A.STL93.67±1.36%93.68±1.37%
Id.A.MTL93.81±1.55%93.49±1.32%
Id.A.Ours (w/o Zj)91.73±1.45%91.85±1.46%
Id.A.Ours93.84±1.43%93.52±1.19%
+ +Table 10. Experimental results on Multi-MNIST dataset under random and identical assignments. Regarding 'Ours $(\mathrm{w / o}\mathbb{Z}^j)$ we apply the debiased adversarial training per-task by taking the other task as the spurious task. Particularly, under identical assignment, applying debiased adversarial training without identifying $\mathbb{Z}^j$ ’s leads to slightly worse performance than the baselines, indicating the necessity of distinguishing the spurious correlations between tasks. In comparison, our method performs comparably with the baselines, owing to the identification of spurious correlation between the two tasks. + +information, the adversarial training framework performs slightly worse than the STL and MTL baselines, underscoring the importance of correctly constructing $\mathbb{Z}^j$ 's, within which our debiased adversarial training framework shall be applied to improve, rather than hinder, the generalizability. + +# 6. Conclusion + +Spurious correlations in MTL present a significant but inadequately addressed challenge. In this paper, we experimentally show the limitations of existing formulations in mitigating spurious correlations. Motivated by the discussion, we propose a novel framework for identifying spurious correlations between tasks. Our approach formulates the problem from a label-based perspective, where we propose to quantify the spurious correlations between tasks based on the difference in correlation coefficients, evaluated on both the training data and the class-wise resampled training data. We show theoretically the feasibility of our framework in distinguishing spurious correlations between tasks. Based on our identification framework, we propose an adversarial training framework for mitigating between-task spurious correlations, which fine-tunes only the task-specific predictors to ensure that task-specific predictors sufficiently discard the spurious information. We validate from experiments on four benchmark datasets that our method outperforms existing baselines in mitigating spurious correlations between tasks, as well as the effectiveness of our method in identifying spurious correlations between tasks. + +# Acknowledgements + +This work was partially supported by the EMBRIO Institute, contract #2120200, a National Science Foundation (NSF) Biology Integration Institute, and NSF IIS #1955890, IIS #2146091, IIS #2345235. + +# References + +[1] Ibrahim M Alabdulmohsin, Behnam Neyshabur, and Xiaohua Zhai. Revisiting neural scaling laws in language and vision. Advances in Neural Information Processing Systems, 35:22300-22312, 2022. 2 +[2] Simon Buchholz, Goutham Rajendran, Elan Rosenfeld, Bryon Aragam, Bernhard Schölkopf, and Pradeep Ravikumar. Learning linear causal representations from interventions under general nonlinear mixing. Advances in Neural Information Processing Systems, 36, 2024. 2 +[3] Yuan Cao and Quanquan Gu. Generalization bounds of stochastic gradient descent for wide and deep neural networks. Advances in neural information processing systems, 32, 2019. 2 +[4] Rich Caruana. Multitask learning. Machine learning, 28: 41-75, 1997. 1 +[5] Yimeng Chen, Ruibin Xiong, Zhi-Ming Ma, and Yanyan Lan. When does group invariant learning survive spurious correlations? Advances in Neural Information Processing Systems, 35:7038-7051, 2022. 2, 3 +[6] Zhao Chen, Jiquan Ngiam, Yanping Huang, Thang Luong, Henrik Kretzschmar, Yuning Chai, and Dragomir Anguelov. Just pick a sign: Optimizing deep multitask models with gradient sign dropout. Advances in Neural Information Processing Systems, 33:2039-2050, 2020. 2 +[7] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 6, 1 +[8] Yuyang Deng, Junyuan Hong, Jiayu Zhou, and Mehrdad Mahdavi. On the generalization ability of unsupervised pretraining. In International Conference on Artificial Intelligence and Statistics, pages 4519-4527. PMLR, 2024. 2, 3 +[9] Nikolaos Dimitriadis, Pascal Frossard, and François Fleuret. Pareto manifold learning: Tackling multiple tasks via ensembles of single-task models. In International Conference on Machine Learning, pages 8015-8052. PMLR, 2023. 2 +[10] Chris Fifty, Ehsan Amid, Zhe Zhao, Tianhe Yu, Rohan Anil, and Chelsea Finn. Efficiently identifying task groupings for multi-task learning. Advances in Neural Information Processing Systems, 34:27503-27516, 2021. 1, 2 +[11] Marco Fumero, Florian Wenzel, Luca Zancato, Alessandro Achille, Emanuele Rodolà, Stefano Soatto, Bernhard Schölkopf, and Francesco Locatello. Leveraging sparse and shared feature activations for disentangled representation learning. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 3, 4, 7 + +[12] Jiechao Guan and Zhiwu Lu. Task relatedness-based generalization bounds for meta learning. In International Conference on Learning Representations, 2022. 2 +[13] Yuzheng Hu, Ruicheng Xian, Qilong Wu, Qiuling Fan, Lang Yin, and Han Zhao. Revisiting scalarization in multi-task learning: A theoretical perspective. Advances in Neural Information Processing Systems, 36, 2024. 2 +[14] Ziniu Hu, Zhe Zhao, Xinyang Yi, Tiansheng Yao, Lichan Hong, Yizhou Sun, and Ed Chi. Improving multi-task generalization via regularizing spurious correlation. Advances in Neural Information Processing Systems, 35:11450-11466, 2022. 1, 2, 3, 4, 7 +[15] Badr Youbi Idrissi, Martin Arjovsky, Mohammad Pezeshki, and David Lopez-Paz. Simple data balancing achieves competitive worst-group-accuracy. In Conference on Causal Learning and Reasoning, pages 336–351. PMLR, 2022. 3 +[16] Junguang Jiang, Baixu Chen, Junwei Pan, Ximei Wang, Dapeng Liu, Jie Jiang, and Mingsheng Long. Forkmerge: Mitigating negative transfer in auxiliary-task learning. Advances in Neural Information Processing Systems, 36, 2024. 2 +[17] Polina Kirichenko, Pavel Izmailov, and Andrew Gordon Wilson. Last layer re-training is sufficient for robustness to spurious correlations. arXiv preprint arXiv:2204.02937, 2022. 3 +[18] Tyler LaBonte, Vidya Muthukumar, and Abhishek Kumar. Towards last-layer retraining for group robustness with fewer annotations. Advances in Neural Information Processing Systems, 36, 2024. 3 +[19] Sébastien Lachapelle, Tristan Deleu, Divyat Mahajan, Ioannis Mitliagkas, Yoshua Bengio, Simon Lacoste-Julien, and Quentin Bertrand. Synergies between disentanglement and sparsity: Generalization and identifiability in multi-task learning. In International Conference on Machine Learning, pages 18171-18206. PMLR, 2023. 1 +[20] Andrew K Lampinen and Surya Ganguli. An analytic theory of generalization dynamics and transfer learning in deep linear networks. arXiv preprint arXiv:1809.10374, 2018. 2 +[21] Dongyue Li, Huy L Nguyen, and Hongyang R Zhang. Identification of negative transfers in multitask learning using surrogate models. arXiv preprint arXiv:2303.14582, 2023. 2, 7 +[22] Weixin Liang and James Zou. Metashift: A dataset of datasets for evaluating contextual distribution shifts and training conflicts. arXiv preprint arXiv:2202.06523, 2022. 6, 1 +[23] Bo Liu, Yihao Feng, Peter Stone, and Qiang Liu. Famo: Fast adaptive multitask optimization. Advances in Neural Information Processing Systems, 36, 2024. 2 +[24] Evan Z Liu, Behzad Haghgoo, Annie S Chen, Aditi Raghunathan, Pang Wei Koh, Shiori Sagawa, Percy Liang, and Chelsea Finn. Just train twice: Improving group robustness without training group information. In International Conference on Machine Learning, pages 6781-6792. PMLR, 2021. 3 +[25] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of + +International Conference on Computer Vision (ICCV), 2015. 6, 1 +[26] Shenyu Lu, Junyi Chai, and Xiaoqian Wang. Neural collapse inspired debiased representation learning for min-max fairness. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 2048-2059, 2024. 3, 4 +[27] Taro Makino, Krzysztof Geras, and Kyunghyun Cho. Generative multitask learning mitigates target-causing confounding. Advances in Neural Information Processing Systems, 35:36546-36558, 2022. 1, 2, 3, 7 +[28] Michinari Momma, Chaosheng Dong, and Jia Liu. A multi-objective/multi-task learning framework induced by pareto stationarity. In International Conference on Machine Learning, pages 15895-15907. PMLR, 2022. 1 +[29] Aviv Navon, Aviv Shamsian, Idan Achituve, Haggai Maron, Kenji Kawaguchi, Gal Chechik, and Ethan Fetaya. Multitask learning as a bargaining game. In International Conference on Machine Learning, pages 16428-16446. PMLR, 2022. 1 +[30] Karl Pearson. Correlation coefficient. In *Royal Society Proceedings*, page 214, 1895. 4 +[31] Elan Rosenfeld, Pradeep Ravikumar, and Andrej Risteski. The risks of invariant risk minimization. arXiv preprint arXiv:2010.05761, 2020. 3 +[32] Shiori Sagawa, Pang Wei Koh, Tatsunori B Hashimoto, and Percy Liang. Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization. arXiv preprint arXiv:1911.08731, 2019. 3 +[33] Shiori Sagawa, Aditi Raghunathan, Pang Wei Koh, and Percy Liang. An investigation of why overparameterization exacerbates spurious correlations. In International Conference on Machine Learning, pages 8346-8356. PMLR, 2020. 3 +[34] Ozan Sener and Vladlen Koltun. Multi-task learning as multi-objective optimization. Advances in neural information processing systems, 31, 2018. 2 +[35] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part V 12, pages 746-760. Springer, 2012. 6, 1 +[36] Trevor Standley, Amir Zamir, Dawn Chen, Leonidas Guibas, Jitendra Malik, and Silvio Savarese. Which tasks should be learned together in multi-task learning? In International conference on machine learning, pages 9120-9132. PMLR, 2020. 1, 2 +[37] Charles Stein et al. Inadmissibility of the usual estimator for the mean of a multivariate normal distribution. In Proceedings of the Third Berkeley symposium on mathematical statistics and probability, pages 197-206, 1956. 1 +[38] Shao-Hua Sun. Multi-digit mnist for few-shot learning, 2019. 6, 1 +[39] Nilesh Tripuraneni, Michael Jordan, and Chi Jin. On the theory of transfer learning: The importance of task diver + +sity. Advances in neural information processing systems, 33: 7852-7862, 2020. 2 +[40] Simon Vandenhende, Stamatios Georgoulis, Wouter Van Gansbeke, Marc Proesmans, Dengxin Dai, and Luc Van Gool. Multi-task learning for dense prediction tasks: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(7):3614-3633, 2021. 2 +[41] Yipei Wang and Xiaoqian Wang. On the effect of key factors in spurious correlation: A theoretical perspective. In International Conference on Artificial Intelligence and Statistics, pages 3745-3753. PMLR, 2024. 3 +[42] Zirui Wang, Yulia Tsvetkov, Orhan Firat, and Yuan Cao. Gradient vaccine: Investigating and improving multi-task optimization in massively multilingual models. arXiv preprint arXiv:2010.05874, 2020. 1 +[43] Enneng Yang, Zhenyi Wang, Li Shen, Shiwei Liu, Guibing Guo, Xingwei Wang, and Dacheng Tao. Adamerging: Adaptive model merging for multi-task learning. arXiv preprint arXiv:2310.02575, 2023. 2 +[44] Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient surgery for multi-task learning. Advances in Neural Information Processing Systems, 33:5824-5836, 2020. 1 +[45] Amir R Zamir, Alexander Sax, William Shen, Leonidas J Guibas, Jitendra Malik, and Silvio Savarese. Taskonomy: Disentangling task transfer learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3712-3722, 2018. 6, 1 +[46] Yu Zhang and Qiang Yang. A survey on multi-task learning. IEEE transactions on knowledge and data engineering, 34 (12):5586-5609, 2021. 1, 2 +[47] Xiao Zhou, Yong Lin, Renjie Pi, Weizhong Zhang, Renzhe Xu, Peng Cui, and Tong Zhang. Model agnostic sample reweighting for out-of-distribution learning. In International Conference on Machine Learning, pages 27203-27221. PMLR, 2022. 3 \ No newline at end of file diff --git a/2025/Identifying and Mitigating Spurious Correlation in Multi-Task Learning/images.zip b/2025/Identifying and Mitigating Spurious Correlation in Multi-Task Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..404facd8a6772d9899f958d3754597e50f293bfc --- /dev/null +++ b/2025/Identifying and Mitigating Spurious Correlation in Multi-Task Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8d7c369e740251b0e6f09c81fdcc927580b166cb34af675d60f67f8fefd94c1b +size 408614 diff --git a/2025/Identifying and Mitigating Spurious Correlation in Multi-Task Learning/layout.json b/2025/Identifying and Mitigating Spurious Correlation in Multi-Task Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1d9d716cb1d6aaa4f24cdad874e09301d43a2782 --- /dev/null +++ b/2025/Identifying and Mitigating Spurious Correlation in Multi-Task Learning/layout.json @@ -0,0 +1,10198 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 82, + 102, + 529, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 102, + 529, + 121 + ], + "spans": [ + { + "bbox": [ + 82, + 102, + 529, + 121 + ], + "type": "text", + "content": "Identifying and Mitigating Spurious Correlation in Multi-Task Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 143, + 504, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 143, + 504, + 184 + ], + "spans": [ + { + "bbox": [ + 106, + 143, + 504, + 184 + ], + "type": "text", + "content": "Junyi Chai, Shenyu Lu & Xiaoqian Wang* \nElmore Family School of Electrical and Computer Engineering, Purdue University \nWest Lafayette, IN, USA" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 214, + 187, + 392, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 187, + 392, + 200 + ], + "spans": [ + { + "bbox": [ + 214, + 187, + 392, + 200 + ], + "type": "text", + "content": "{chai28,lu876,joywang}@purdue.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "spans": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 252, + 296, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 252, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 252, + 296, + 624 + ], + "type": "text", + "content": "Multi-task learning (MTL) is a paradigm that aims to improve the generalization of models by simultaneously learning multiple related tasks, leveraging shared representations and task-specific information to enhance performance on individual tasks. However, existing work has shown that MTL can potentially hinder generalization, with one key factor being spurious correlations between tasks. Owing to the knowledge-sharing property, the per-task predictors are more likely to develop reliance on spurious features. Most existing approaches address this issue through distributional robustness, aiming to maintain consistent performance across different distributions under unknown covariate shifts. However, this formulation lacks theoretical guarantees and can be sensitive to the construction of covariate shifts. In this work, we propose a novel perspective, where we seek to identify spurious correlations between tasks. Drawing inspirations from conventional formulations on spurious correlation, for each task, we propose to distinguish its spurious tasks using the difference in correlation coefficients between the empirical distribution and class-wise resampled distributions, thereby capturing the correlations between task labels w.r.t. each class. We prove theoretically the feasibility of the resampling strategy in characterizing spurious correlations between tasks. Furthermore, we propose a simple fine-tuning strategy, debiased adversarial training, where the per-task predictors are adversarily trained to disregard information associated with their spurious tasks. Experimental results on six benchmark datasets show that our method effectively mitigates spurious correlations and outperforms state-of-the-art methods in improving generalization." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 647, + 135, + 660 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 135, + 660 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 135, + 660 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 669, + 295, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 669, + 295, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 669, + 295, + 693 + ], + "type": "text", + "content": "Stemming from Stein's paradox [37] where the estimation error decreases under joint estimation, formalized in [4]," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 228, + 555, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 555, + 396 + ], + "type": "text", + "content": "multi-task learning (MTL) has since become increasingly important in various applications, particularly in computer vision, demonstrating both improved accuracy and reduced computational overhead. MTL aims to jointly learn multiple tasks, allowing each task to leverages the shared representations and domain knowledge between tasks, with the goal of improving the generalization performance across all tasks [19, 46]. Compared with single-task learning, which focuses on optimizing over one specific objective, MTL extracts common patterns that are beneficial across different but related objectives, potentially leading to better performance on individual tasks. This collaborative framework mitigates overfitting and enhances model robustness through cross-task knowledge transfer [28, 29]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 399, + 556, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 556, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 556, + 639 + ], + "type": "text", + "content": "Despite the benefits in MTL, existing work has demonstrated that under certain circumstances, MTL could lead to degraded performance in generalization, particularly when different tasks do not experience high knowledge-sharing [10, 36]. This abnormality has conventionally been attributed to the model capacity issue, where conflicting tasks may compete during optimization [42, 44]. However, Recent studies have revealed a counterintuitive phenomenon in MTL, where even under over-parametrization, a MTL model might excel in minimizing training loss, yet surprisingly underperform single-task models in terms of generalization [14]. Accordingly, spurious correlations between different tasks have been identified as an intrinsic problem that hurts generalization in MTL [14, 27]. Owing to the multi-tasking property, the shared representations are expected to encode knowledge across tasks; consequently, when two tasks exhibit high label correlations despite having unrelated underlying knowledge, the per-task predictors may erroneously leverage unrelated spurious knowledge, potentially compromising generalization." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": "Most of existing work addressing spurious correlations in MTL approaches this issue through the lens of invariant learning, aiming to ensure that models perform consistently across various empirical distributions [11, 14]. Specifically, the shared representations are expected to recover the ground-truth latent causal features, and by en" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 702, + 145, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 145, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 145, + 713 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25698" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 72, + 294, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 72, + 294, + 382 + ], + "spans": [ + { + "bbox": [ + 57, + 72, + 294, + 382 + ], + "type": "text", + "content": "forcing distributional-invariance, the per-task predictors are expected to leverage features that are causally related to each task, instead of spurious features. Nonetheless, on one hand, counterfactual data is generally inaccessible on high-dimensional data, leading to the infeasibility in conducting causality discovery [2]. Consequently, the shared representations may fail to recover the underlying latent features, particularly when approximated by disentanglement across latent dimensions [11, 14]. Moreover, enforcing distributional invariance does not guarantee utilization of only the task-specific causal features, as the difference across empirical distributions does not necessarily break the spurious correlations between tasks [5]. In the absence of prior knowledge regarding spurious correlations between tasks, along with the potential lack of certain label pairings, it becomes difficult to construct the desired distributions. Alternatively, Makino et al. [27] proposes to mitigate the spurious correlations between tasks by generatively estimating the joint distribution of task labels. Despite theoretical soundness of the formulation, estimating joint distributions can be error-prone, especially when certain label combinations are missing. Additionally, estimating joint distributions involving labels of varied dimensions is often infeasible. This issue is particularly relevant for vision tasks, where labels from different tasks are likely to have differing dimensions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 386, + 294, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 386, + 294, + 660 + ], + "spans": [ + { + "bbox": [ + 57, + 386, + 294, + 660 + ], + "type": "text", + "content": "In this work, we propose an alternative formulation regarding spurious correlations in MTL. Instead of imposing regularizations on distributional invariance, we seek to directly identify the spurious correlations between tasks. Drawing inspirations from conventional formulations of spurious correlation, typically defined by the dependency between labels and spurious information, we propose to identify spurious correlations between tasks using correlation coefficients. Specifically, we detect spurious correlations between two tasks by examining differences in their correlation coefficients, calculated on both the original training data and class-wise resampled training data. We prove theoretically the feasibility of our formulation in identifying spurious correlations between tasks, as well as in distinguishing spurious correlations from genuine task relationships. Building upon the per-task identification of spurious correlations, we formulate the bias mitigation problem as a mini-max game, where for each task, we consider the predictors of its spurious tasks as the adversaries, such that by fine-tuning the task-specific projector and the task-specific predictor, we intend to minimize the training loss for this task, while maximizing the training loss for its spurious tasks under projection." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 664, + 241, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 664, + 241, + 673 + ], + "spans": [ + { + "bbox": [ + 69, + 664, + 241, + 673 + ], + "type": "text", + "content": "We summarize our contribution as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 678, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 678, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 678, + 294, + 713 + ], + "type": "text", + "content": "- We propose a theoretical framework for detecting spurious correlations in MTL by analyzing intra-task label dependencies, and we prove theoretically the feasibility of" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 324, + 73, + 453, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 73, + 453, + 83 + ], + "spans": [ + { + "bbox": [ + 324, + 73, + 453, + 83 + ], + "type": "text", + "content": "our method in the identification." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 316, + 84, + 553, + 190 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 316, + 84, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 84, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 316, + 84, + 553, + 144 + ], + "type": "text", + "content": "- We address spurious correlations in multi-task learning by fine-tuning task-specific predictors, enhancing per-task performance through adversarial training that minimizes loss on the primary task while maximizing loss on its identified spurious tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 316, + 144, + 553, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 144, + 553, + 190 + ], + "spans": [ + { + "bbox": [ + 316, + 144, + 553, + 190 + ], + "type": "text", + "content": "- We validate from experiments the effectiveness of our framework in identification of spurious correlations between tasks, as well as in improving generalizability on four benchmark datasets." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 316, + 210, + 399, + 221 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 210, + 399, + 221 + ], + "spans": [ + { + "bbox": [ + 316, + 210, + 399, + 221 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 316, + 234, + 509, + 246 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 234, + 509, + 246 + ], + "spans": [ + { + "bbox": [ + 316, + 234, + 509, + 246 + ], + "type": "text", + "content": "2.1. Generalization in multi-task learning" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 316, + 255, + 553, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 255, + 553, + 434 + ], + "spans": [ + { + "bbox": [ + 316, + 255, + 553, + 434 + ], + "type": "text", + "content": "The generalizability of deep neural networks has often been linked to the availability of large training datasets, making sample sparsity one of the fundamental limitations for model performance [1, 3, 8]. MTL addresses sample sparsity by leveraging labeled data across multiple related tasks, effectively expanding the available training data [13, 46]. Research on multi-task generalization predominantly assumes shared knowledge between tasks, where learning across diverse tasks enables more robust and transferable representations [12, 20, 39]. Despite empirical validations of the assumption [40], the fundamental assumption of task relatedness often fails in real-world applications, where competing tasks may vie for model capacity, potentially degrading MTL performance below that of single-task approaches [10, 36]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 316, + 439, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 439, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 439, + 553, + 712 + ], + "type": "text", + "content": "Prior research has addressed this challenge through two distinct perspectives. One line of work has focused on the model capacity issue, seeking equilibrium between conflicting task objectives. Yang et al. [43] proposes to adaptively reweight per-task training loss, leveraging entropy minimization on unlabeled test samples. Jiang et al. [16] proposes to dynamically merges model branches while optimizing task weights based on the performance on validation set. Simliar reweighing strategies have also been applied in [21, 23]. Alternatively, Chen et al. [6] introduces a selective gradient filtering approach to resolve optimization conflicts, while Dimitriadis et al. [9] and Sener and Koltun [34] propose to formulate the problem from the perspective of Pareto optimality to achieve optimal trade-off between conflicting tasks. In contrast, recent work has focused on the exacerbation of spurious correlations under MTL, where the multi-tasking property could exacerbate reliance on spurious features for each task owing to the interaction between tasks [11, 14, 27]. Specifically, Hu et al. [14] and Fumero et al. [11] address intra-task spurious correlations through distributional invariance, while Makino et al. [27] employs generative modeling of joint distributions to mitigate these correlations." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "25699" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 174, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 174, + 84 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 174, + 84 + ], + "type": "text", + "content": "2.2. Spurious correlation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 296, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 296, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 296, + 293 + ], + "type": "text", + "content": "Spurious correlation refers to misleading patterns that appear to hold in training data but fail to generalize to new, unseen data [32, 33]. Existing work on spurious correlations typically aims to reduce the error for the worst-performing group, characterized by both the task label and the associated spurious information. typically, these methods can be divided into two categories: data-augmentation, where the training distribution is modified to enhancing worst-group performance [5, 8, 15]; model-regularization, where regularizations are imposed during the training process to eliminate the affect of spurious information [24, 26, 32, 41, 47]. More recently, last-layer retraining has been proposed as an efficient alternative for addressing spurious correlations [17, 18]. In this approach, instead of updating all network parameters, only the classifier layer is fine-tuned on a balanced validation set, significantly reducing computational complexity." + } + ] + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 87, + 300, + 265, + 434 + ], + "blocks": [ + { + "bbox": [ + 87, + 300, + 265, + 434 + ], + "lines": [ + { + "bbox": [ + 87, + 300, + 265, + 434 + ], + "spans": [ + { + "bbox": [ + 87, + 300, + 265, + 434 + ], + "type": "image", + "image_path": "89d6d7c033b92f9627bb3834710605157dc10b64834c7d193758bdf5f80afc67.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 445, + 296, + 511 + ], + "lines": [ + { + "bbox": [ + 55, + 445, + 296, + 511 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 296, + 511 + ], + "type": "text", + "content": "Figure 1. Our framework improves generalizability by identifying and correcting spurious correlations between tasks. For instance, on CelebA dataset, the task labels 'male' and 'attractiveness' are known to be spuriously correlated [26]. Compared with vanilla MTL where predictions for attractiveness falsely correlate with predictions for gender, our method effectively mitigates this bias." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 526, + 140, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 526, + 140, + 539 + ], + "spans": [ + { + "bbox": [ + 55, + 526, + 140, + 539 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 546, + 296, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 296, + 644 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 296, + 644 + ], + "type": "text", + "content": "Consider " + }, + { + "bbox": [ + 55, + 546, + 296, + 644 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 546, + 296, + 644 + ], + "type": "text", + "content": " tasks in total. Let " + }, + { + "bbox": [ + 55, + 546, + 296, + 644 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{train}} := \\{(x_i, y_i^1, \\ldots, y_i^k), 1 \\leq i \\leq N\\}" + }, + { + "bbox": [ + 55, + 546, + 296, + 644 + ], + "type": "text", + "content": " be the training set where for " + }, + { + "bbox": [ + 55, + 546, + 296, + 644 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 546, + 296, + 644 + ], + "type": "text", + "content": "-th sample, " + }, + { + "bbox": [ + 55, + 546, + 296, + 644 + ], + "type": "inline_equation", + "content": "x_i \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 55, + 546, + 296, + 644 + ], + "type": "text", + "content": " is the input feature and " + }, + { + "bbox": [ + 55, + 546, + 296, + 644 + ], + "type": "inline_equation", + "content": "y_i^j \\in \\mathbb{R}^{d^j}" + }, + { + "bbox": [ + 55, + 546, + 296, + 644 + ], + "type": "text", + "content": " is the label for " + }, + { + "bbox": [ + 55, + 546, + 296, + 644 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 546, + 296, + 644 + ], + "type": "text", + "content": "-th task, let " + }, + { + "bbox": [ + 55, + 546, + 296, + 644 + ], + "type": "inline_equation", + "content": "\\pmb{f}: \\mathbb{R}^d \\to \\mathbb{R}^l" + }, + { + "bbox": [ + 55, + 546, + 296, + 644 + ], + "type": "text", + "content": " be the function of shared encoder and " + }, + { + "bbox": [ + 55, + 546, + 296, + 644 + ], + "type": "inline_equation", + "content": "\\pmb{g}^j: \\mathbb{R}^l \\to \\mathbb{R}^{d^j}" + }, + { + "bbox": [ + 55, + 546, + 296, + 644 + ], + "type": "text", + "content": " be the function of predictor for " + }, + { + "bbox": [ + 55, + 546, + 296, + 644 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 546, + 296, + 644 + ], + "type": "text", + "content": "-th task, let " + }, + { + "bbox": [ + 55, + 546, + 296, + 644 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^j" + }, + { + "bbox": [ + 55, + 546, + 296, + 644 + ], + "type": "text", + "content": " be the training loss for " + }, + { + "bbox": [ + 55, + 546, + 296, + 644 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 546, + 296, + 644 + ], + "type": "text", + "content": "-th task, a multi-task learning problem can be formulated as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 649, + 245, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 649, + 245, + 681 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 245, + 681 + ], + "type": "interline_equation", + "content": "\\underset {\\boldsymbol {f}, \\boldsymbol {g} ^ {1}, \\dots , \\boldsymbol {g} ^ {k}} {\\arg \\min } \\sum_ {j = 1} ^ {k} w ^ {j} \\bar {\\mathcal {L}} ^ {j} (\\mathcal {D} ^ {\\text {t r a i n}}; \\boldsymbol {g} ^ {j}, \\boldsymbol {f}),", + "image_path": "94ed943a6a75bee7b0ce9538109672f44df248621e07f4e606ad6dd02e7f8860.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 687, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 687, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 687, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 687, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\bar{\\mathcal{L}}^j (\\mathcal{D}^{\\mathrm{train}};\\pmb {g}^j,\\pmb {f})\\coloneqq \\frac{1}{N}\\sum_{i = 1}^N\\mathcal{L}^j (\\pmb {g}^j (\\pmb {f}(x_i)),y_i^j)" + }, + { + "bbox": [ + 55, + 687, + 296, + 714 + ], + "type": "text", + "content": " is the average training loss for " + }, + { + "bbox": [ + 55, + 687, + 296, + 714 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 687, + 296, + 714 + ], + "type": "text", + "content": " -th task, and " + }, + { + "bbox": [ + 55, + 687, + 296, + 714 + ], + "type": "inline_equation", + "content": "w^{j}" + }, + { + "bbox": [ + 55, + 687, + 296, + 714 + ], + "type": "text", + "content": " specifies" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "content": "the weight of " + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "inline_equation", + "content": "\\bar{\\mathcal{L}}^j" + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "content": " during optimization. Despite the benefit of exploring task-relatedness, which are expected to benefit the generalizability, recent work on MTL has pointed to the potential risk of spurious correlations between tasks owing to the confounder between task labels [14, 27]. Let " + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{F}_i, 1 \\leq i \\leq K\\}" + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "content": " be the set of underlying latent features such that " + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_i \\perp \\mathcal{F}_j" + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "inline_equation", + "content": "\\forall i \\neq j" + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "content": ", let " + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^j" + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "content": " be the set of underlying latent features that are causally related with " + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "content": "-th task, when " + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^i \\neq \\mathcal{F}^j" + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "content": ", optimizing " + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "inline_equation", + "content": "\\bar{\\mathcal{L}}^i" + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "inline_equation", + "content": "\\bar{\\mathcal{L}}^j" + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "content": " jointly could exacerbate reliance on spurious features from " + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^j" + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "inline_equation", + "content": "g^i" + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "content": ", as the shared encoder needs to explicitly encode knowledge for both tasks. Such confounder brings in spurious correlations between tasks and can harm generalization performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 228, + 555, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 555, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 555, + 346 + ], + "type": "text", + "content": "Nonetheless, compared with conventional formulations where the spurious information is assumed to be available on training data [32, 33], regarding MTL, we do not have the spurious information at hand, where each task may consider all other tasks as being either spurious or beneficial. Alternatively, most existing work [11, 14] has formulated the problem from the perspective of distributional robustness, where the model is expected to perform consistently on both the training data and the validation data " + }, + { + "bbox": [ + 313, + 228, + 555, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{val}}" + }, + { + "bbox": [ + 313, + 228, + 555, + 346 + ], + "type": "text", + "content": " under some unknown covariate shift:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 350, + 561, + 410 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 561, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 561, + 410 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\underset {\\boldsymbol {f}, \\boldsymbol {g} ^ {1}, \\dots , \\boldsymbol {g} ^ {k}} {\\arg \\min } \\sum_ {j = 1} ^ {k} w ^ {j} \\bar {\\mathcal {L}} ^ {j} (\\mathcal {D} ^ {\\text {t r a i n}}; \\boldsymbol {g} ^ {j}, \\boldsymbol {f}) + \\sum_ {j = 1} ^ {k} w ^ {\\text {v a l}, j} \\bar {\\mathcal {L}} ^ {j} (\\mathcal {D} ^ {\\text {v a l}}; \\boldsymbol {g} ^ {j}, \\boldsymbol {f}), \\\\ s. t. R \\left(\\mathcal {D} ^ {\\text {t r a i n}}; f\\right) \\leq \\epsilon , R \\left(\\mathcal {D} ^ {\\text {v a l}}; f\\right) \\leq \\epsilon , \\tag {1} \\\\ \\end{array}", + "image_path": "c80e921d671ff628013a6985cba02ac6755bd65fc8020cfde0e9dbb3b32e7f71.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 411, + 554, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 554, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 554, + 612 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 411, + 554, + 612 + ], + "type": "inline_equation", + "content": "w^{\\mathrm{val},j}" + }, + { + "bbox": [ + 313, + 411, + 554, + 612 + ], + "type": "text", + "content": " specifies the weight of validation loss for " + }, + { + "bbox": [ + 313, + 411, + 554, + 612 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 411, + 554, + 612 + ], + "type": "text", + "content": "-th task, and " + }, + { + "bbox": [ + 313, + 411, + 554, + 612 + ], + "type": "inline_equation", + "content": "R(\\mathcal{D}^{\\mathrm{train}};f)" + }, + { + "bbox": [ + 313, + 411, + 554, + 612 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 411, + 554, + 612 + ], + "type": "inline_equation", + "content": "R(\\mathcal{D}^{\\mathrm{val}};f)" + }, + { + "bbox": [ + 313, + 411, + 554, + 612 + ], + "type": "text", + "content": " correspond to the regularization on " + }, + { + "bbox": [ + 313, + 411, + 554, + 612 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 411, + 554, + 612 + ], + "type": "text", + "content": " such that the encoder is dimensional-wise disentangled. For example, MT-CRL [14] employs covariance between latent dimensions as the regularization, while Meta-learning [11] formulates " + }, + { + "bbox": [ + 313, + 411, + 554, + 612 + ], + "type": "inline_equation", + "content": "R(\\mathcal{D}^{\\mathrm{train}};f)" + }, + { + "bbox": [ + 313, + 411, + 554, + 612 + ], + "type": "text", + "content": " as the entropy of normalized representations per dimension. Eq. (1) stems from the assumption that for " + }, + { + "bbox": [ + 313, + 411, + 554, + 612 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 411, + 554, + 612 + ], + "type": "text", + "content": "-th task, the covariate shift on validation data effectively disrupts the correlation between its spurious features and its task labels, which could be infeasible in general. On one hand, without prior knowledge on the task-specific causal features, there's no guarantee that the covariate shift breaks the spurious correlation between tasks. Furthermore, the effectiveness of Eq. (1) may be influenced by the degree of bias in the validation set, introducing uncertainty in the mitigation spurious correlations." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 613, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 613, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 613, + 556, + 715 + ], + "type": "text", + "content": "We use the following experiments on linear SEM [31] to validate the two perspectives. Following existing work [14], we construct two binary classification tasks where the causal features of each task are independent and follow a Gaussian distribution. Let " + }, + { + "bbox": [ + 313, + 613, + 556, + 715 + ], + "type": "inline_equation", + "content": "Y^{j}" + }, + { + "bbox": [ + 313, + 613, + 556, + 715 + ], + "type": "text", + "content": " be the random variable of labels for " + }, + { + "bbox": [ + 313, + 613, + 556, + 715 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 613, + 556, + 715 + ], + "type": "text", + "content": "-th task, we set the ratios between task labels on training data as " + }, + { + "bbox": [ + 313, + 613, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\alpha_{y}^{\\text{train}} = \\frac{P(Y^{0} = y, Y^{1} = (1 - y))}{P(Y^{0} = y, Y^{1} = y)} = \\frac{1}{9}" + }, + { + "bbox": [ + 313, + 613, + 556, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 613, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\forall y \\in \\{0, 1\\}" + }, + { + "bbox": [ + 313, + 613, + 556, + 715 + ], + "type": "text", + "content": " on training data. Regarding both tasks, we have" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25700" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 137 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 137 + ], + "type": "inline_equation", + "content": "P(Y^{j} = 0) = P(Y^{j} = 1) = 0.5, \\forall j \\in \\{0,1\\}" + }, + { + "bbox": [ + 55, + 72, + 296, + 137 + ], + "type": "text", + "content": ", and the testing data is balanced sampled across each co-couurance of task labels such that " + }, + { + "bbox": [ + 55, + 72, + 296, + 137 + ], + "type": "inline_equation", + "content": "\\alpha_{y}^{\\mathrm{test}} = \\frac{P(Y^{0} = y, Y^{1} = (1 - y))}{P(Y^{0} = y, Y^{1} = y)} = 1" + }, + { + "bbox": [ + 55, + 72, + 296, + 137 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 72, + 296, + 137 + ], + "type": "inline_equation", + "content": "\\forall y \\in \\{0,1\\}" + }, + { + "bbox": [ + 55, + 72, + 296, + 137 + ], + "type": "text", + "content": ". Smaller " + }, + { + "bbox": [ + 55, + 72, + 296, + 137 + ], + "type": "inline_equation", + "content": "\\alpha_{y}" + }, + { + "bbox": [ + 55, + 72, + 296, + 137 + ], + "type": "text", + "content": " indicates higher correlation between the two task labels." + } + ] + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 74, + 142, + 279, + 183 + ], + "blocks": [ + { + "bbox": [ + 74, + 142, + 279, + 183 + ], + "lines": [ + { + "bbox": [ + 74, + 142, + 279, + 183 + ], + "spans": [ + { + "bbox": [ + 74, + 142, + 279, + 183 + ], + "type": "table", + "html": "
TaskMTL (αtrainy = 1/9)MTL (αtrainy = 1)
Task 081.32±0.29%94.47±0.34%
Task 181.37±0.41%94.47±0.35%
", + "image_path": "57a342d01f5bd3767d4bd32dbfd1a473d84bf9d22973de2f0c49781cbcc019b9.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 75, + 256, + 279, + 297 + ], + "blocks": [ + { + "bbox": [ + 55, + 191, + 296, + 236 + ], + "lines": [ + { + "bbox": [ + 55, + 191, + 296, + 236 + ], + "spans": [ + { + "bbox": [ + 55, + 191, + 296, + 236 + ], + "type": "text", + "content": "Table 1. Experimental results on synthetic data. We choose classification accuracy as the evaluation metric for both tasks. Under high correlation, the model performs remarkably worse than that under no spurious correlation." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 75, + 256, + 279, + 297 + ], + "lines": [ + { + "bbox": [ + 75, + 256, + 279, + 297 + ], + "spans": [ + { + "bbox": [ + 75, + 256, + 279, + 297 + ], + "type": "table", + "html": "
TaskMT-CRL (αyval = 1/9)Meta-learning
Task 081.59±0.36%81.73±0.35%
Task 181.87±0.95%81.67±0.44%
", + "image_path": "2b57347b9f224de43f6d9e45d74c46b49ae6cba18052cf852cc427b6eec192b2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 63, + 380, + 288, + 422 + ], + "blocks": [ + { + "bbox": [ + 55, + 304, + 296, + 360 + ], + "lines": [ + { + "bbox": [ + 55, + 304, + 296, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 296, + 360 + ], + "type": "text", + "content": "Table 2. Experimental results on synthetic data. Both methods show marginal improvement in testing accuracy compared with MTL (" + }, + { + "bbox": [ + 55, + 304, + 296, + 360 + ], + "type": "inline_equation", + "content": "\\alpha_{y}^{\\mathrm{train}} = \\frac{1}{9}" + }, + { + "bbox": [ + 55, + 304, + 296, + 360 + ], + "type": "text", + "content": "): " + }, + { + "bbox": [ + 55, + 304, + 296, + 360 + ], + "type": "inline_equation", + "content": "81.32 \\pm 0.29\\%" + }, + { + "bbox": [ + 55, + 304, + 296, + 360 + ], + "type": "text", + "content": " for task 0 and " + }, + { + "bbox": [ + 55, + 304, + 296, + 360 + ], + "type": "inline_equation", + "content": "81.37 \\pm 0.41\\%" + }, + { + "bbox": [ + 55, + 304, + 296, + 360 + ], + "type": "text", + "content": " for task 1, owing to the consistency in " + }, + { + "bbox": [ + 55, + 304, + 296, + 360 + ], + "type": "inline_equation", + "content": "\\alpha_{y}" + }, + { + "bbox": [ + 55, + 304, + 296, + 360 + ], + "type": "text", + "content": "'s between training and validation data." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 380, + 288, + 422 + ], + "lines": [ + { + "bbox": [ + 63, + 380, + 288, + 422 + ], + "spans": [ + { + "bbox": [ + 63, + 380, + 288, + 422 + ], + "type": "table", + "html": "
TaskMT-CRL (αyval = 1/5)MT-CRL (αyval = 1)
Task 083.74±0.65%87.59±0.74%
Task 183.32±0.59%88.16±0.46%
", + "image_path": "1305b7510dc94ad24754e0455560a01d4d6f4a8fb219527efe7bc49c1154c504.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 430, + 296, + 507 + ], + "lines": [ + { + "bbox": [ + 55, + 430, + 296, + 507 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 296, + 507 + ], + "type": "text", + "content": "Table 3. Experimental results on synthetic data. The testing performance improves as " + }, + { + "bbox": [ + 55, + 430, + 296, + 507 + ], + "type": "inline_equation", + "content": "\\alpha_{y}^{\\mathrm{val}}" + }, + { + "bbox": [ + 55, + 430, + 296, + 507 + ], + "type": "text", + "content": "'s decrease, substantiating the importance of mitigating bias on the validation data. Nonetheless, given the balanced validation set, MT-CRL still performs worse than the optimal baseline MTL (" + }, + { + "bbox": [ + 55, + 430, + 296, + 507 + ], + "type": "inline_equation", + "content": "\\alpha_{y}^{\\mathrm{train}} = 1" + }, + { + "bbox": [ + 55, + 430, + 296, + 507 + ], + "type": "text", + "content": "): " + }, + { + "bbox": [ + 55, + 430, + 296, + 507 + ], + "type": "inline_equation", + "content": "94.47 \\pm 0.34\\%" + }, + { + "bbox": [ + 55, + 430, + 296, + 507 + ], + "type": "text", + "content": " for task 0 and " + }, + { + "bbox": [ + 55, + 430, + 296, + 507 + ], + "type": "inline_equation", + "content": "94.47 \\pm 0.35\\%" + }, + { + "bbox": [ + 55, + 430, + 296, + 507 + ], + "type": "text", + "content": " for task 1, highlighting the insufficiency in formulating spurious correlations via distributional robustness." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 521, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 296, + 713 + ], + "type": "text", + "content": "We show results of under different constructions of validation sets in Tab. 1 - 2. Regarding MT-CRL [14], as the validation set is only assumed to contain unknown shift, we consider three different types of covariate shifts: The training-validation split where the validation set is sampled from the training distribution, as suggested by the authors as a general construction of covariate shift, where " + }, + { + "bbox": [ + 55, + 521, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\alpha_{y}" + }, + { + "bbox": [ + 55, + 521, + 296, + 713 + ], + "type": "text", + "content": "'s remain the same on both training and validation data; A less biased validation set where " + }, + { + "bbox": [ + 55, + 521, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\alpha_{y}^{\\mathrm{val}} = \\frac{1}{5}, \\forall y \\in \\{0,1\\}" + }, + { + "bbox": [ + 55, + 521, + 296, + 713 + ], + "type": "text", + "content": "; A balanced validation set where " + }, + { + "bbox": [ + 55, + 521, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\alpha_{y}^{\\mathrm{val}} = 1, \\forall y \\in \\{0,1\\}" + }, + { + "bbox": [ + 55, + 521, + 296, + 713 + ], + "type": "text", + "content": ". While a balanced validation set shows improvement in the performance on testing data, the first two constructions of validation data show marginal improvement compared with vanilla MTL, validating our claim that the effectiveness of distributional invariance can be subject to the bias on validation data. Nonetheless, in reality, it is hard to obtain or construct a" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 554, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 131 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 131 + ], + "type": "text", + "content": "perfectly balanced data owing to the uncertainty in the spurious correlations between tasks. Furthermore, it is possible that certain labeling pairs are not observed on training or validation data, leading to infeasibility of constructing such balanced data." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 133, + 555, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 555, + 276 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 555, + 276 + ], + "type": "text", + "content": "Regarding Meta-learning [11], since the validation set is constructed per task such that for " + }, + { + "bbox": [ + 313, + 133, + 555, + 276 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 133, + 555, + 276 + ], + "type": "text", + "content": "-th task, " + }, + { + "bbox": [ + 313, + 133, + 555, + 276 + ], + "type": "inline_equation", + "content": "P(Y^{j} = 0) = P(Y^{j} = 1) = 0.5" + }, + { + "bbox": [ + 313, + 133, + 555, + 276 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 133, + 555, + 276 + ], + "type": "inline_equation", + "content": "\\forall j \\in \\{0,1\\}" + }, + { + "bbox": [ + 313, + 133, + 555, + 276 + ], + "type": "text", + "content": ", we have " + }, + { + "bbox": [ + 313, + 133, + 555, + 276 + ], + "type": "inline_equation", + "content": "\\alpha_{y}" + }, + { + "bbox": [ + 313, + 133, + 555, + 276 + ], + "type": "text", + "content": "'s remain the same on both training and validation data. As a result, such construction fails to show improvement in mitigating spurious information compared with the ERM baseline. This demonstrates the flaws in formulating the covariate shift via class-balance: the spurious correlation between tasks does not necessarily lead to imbalance in the marginal distribution of labels per task, where using class-balanced resampling can lead to marginal variations in the conditional distributions " + }, + { + "bbox": [ + 313, + 133, + 555, + 276 + ], + "type": "inline_equation", + "content": "P(Y_{i}|Y_{j})" + }, + { + "bbox": [ + 313, + 133, + 555, + 276 + ], + "type": "text", + "content": " of task labels." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 286, + 370, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 286, + 370, + 298 + ], + "spans": [ + { + "bbox": [ + 314, + 286, + 370, + 298 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 306, + 554, + 330 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 306, + 554, + 330 + ], + "spans": [ + { + "bbox": [ + 313, + 306, + 554, + 330 + ], + "type": "text", + "content": "4.1. Identification of spurious correlation between tasks" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 335, + 555, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 555, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 555, + 491 + ], + "type": "text", + "content": "Our preliminary discussion suggests the necessity of identifying spurious correlations between tasks. In light of the intractability in identifying the latent causal features for each task, particularly regarding high-dimensional vision data, we instead consider modeling the spurious correlations from the label perspective. Existing work [14, 26] has demonstrated that when two tasks are spuriously correlated, their corresponding task labels will be disproportionately paired, leading to an uneven joint distribution across task labels. Specifically, let " + }, + { + "bbox": [ + 313, + 335, + 555, + 491 + ], + "type": "inline_equation", + "content": "\\mathbb{S}^j" + }, + { + "bbox": [ + 313, + 335, + 555, + 491 + ], + "type": "text", + "content": " be the set of values of labels for " + }, + { + "bbox": [ + 313, + 335, + 555, + 491 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 335, + 555, + 491 + ], + "type": "text", + "content": "-th task, when spurious correlation exists between " + }, + { + "bbox": [ + 313, + 335, + 555, + 491 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 335, + 555, + 491 + ], + "type": "text", + "content": "-th and " + }, + { + "bbox": [ + 313, + 335, + 555, + 491 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 335, + 555, + 491 + ], + "type": "text", + "content": "-th tasks, there exist " + }, + { + "bbox": [ + 313, + 335, + 555, + 491 + ], + "type": "inline_equation", + "content": "y, y' \\in \\mathbb{S}^i" + }, + { + "bbox": [ + 313, + 335, + 555, + 491 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 335, + 555, + 491 + ], + "type": "inline_equation", + "content": "\\tilde{y}, \\tilde{y}' \\in \\mathbb{S}^j" + }, + { + "bbox": [ + 313, + 335, + 555, + 491 + ], + "type": "text", + "content": ", such that the following inequality regarding label distribution holds:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 341, + 499, + 555, + 515 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 499, + 555, + 515 + ], + "spans": [ + { + "bbox": [ + 341, + 499, + 555, + 515 + ], + "type": "interline_equation", + "content": "P (Y ^ {i} = y, Y ^ {j} = \\tilde {y}) > P (Y ^ {i} = y ^ {\\prime}, Y ^ {j} = \\tilde {y} ^ {\\prime}). \\quad (2)", + "image_path": "4c218366b094486f401c825c8c6ca62b7cfe61971badf8cbd66e4a29315374d4.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 523, + 554, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 523, + 554, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 523, + 554, + 559 + ], + "type": "text", + "content": "To quantitatively evaluate Eq. (2), we observe that the correlation between task labels can be quantified using the Pearson correlation coefficient [30]:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 567, + 555, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 555, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 555, + 597 + ], + "type": "interline_equation", + "content": "\\rho_ {Y ^ {i}, Y ^ {j}} = \\frac {\\mathbb {E} [ Y ^ {i} Y ^ {j} ] - \\mathbb {E} [ Y ^ {i} ] \\mathbb {E} [ Y ^ {j} ]}{\\sqrt {\\mathbb {E} [ (Y ^ {i}) ^ {2} ] - (\\mathbb {E} [ Y ^ {i} ]) ^ {2}} \\sqrt {\\mathbb {E} [ (Y ^ {j}) ^ {2} ] - (\\mathbb {E} [ Y ^ {j} ]) ^ {2}}}.", + "image_path": "82aa99fdcc27e93cbcf40762900ef287d9450ec69a298bbb9a2a854678a6885f.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 604, + 555, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 604, + 555, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 604, + 555, + 676 + ], + "type": "text", + "content": "For simplicity of discussion, consider both tasks as binary classification where " + }, + { + "bbox": [ + 313, + 604, + 555, + 676 + ], + "type": "inline_equation", + "content": "Y^{i}" + }, + { + "bbox": [ + 313, + 604, + 555, + 676 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 604, + 555, + 676 + ], + "type": "inline_equation", + "content": "Y^{j}" + }, + { + "bbox": [ + 313, + 604, + 555, + 676 + ], + "type": "text", + "content": " follow Bernoulli distributions with expectations " + }, + { + "bbox": [ + 313, + 604, + 555, + 676 + ], + "type": "inline_equation", + "content": "p^{i}" + }, + { + "bbox": [ + 313, + 604, + 555, + 676 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 604, + 555, + 676 + ], + "type": "inline_equation", + "content": "p^{j}" + }, + { + "bbox": [ + 313, + 604, + 555, + 676 + ], + "type": "text", + "content": ", respectively. We refer to Algorithm 1 for extension to multi-class scenario. The following lemma states the close-form expression of " + }, + { + "bbox": [ + 313, + 604, + 555, + 676 + ], + "type": "inline_equation", + "content": "\\rho_{Y^{i},Y^{j}}" + }, + { + "bbox": [ + 313, + 604, + 555, + 676 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 685, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 685, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 685, + 556, + 715 + ], + "type": "text", + "content": "Lemma 1. Let " + }, + { + "bbox": [ + 313, + 685, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\alpha_{y}^{ij} := \\frac{P(Y^i = y, Y^j = 0)}{P(Y^i = y, Y^j = 1)}" + }, + { + "bbox": [ + 313, + 685, + 556, + 715 + ], + "type": "text", + "content": " be the ratio of joint probabilities of " + }, + { + "bbox": [ + 313, + 685, + 556, + 715 + ], + "type": "inline_equation", + "content": "Y^i" + }, + { + "bbox": [ + 313, + 685, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 685, + 556, + 715 + ], + "type": "inline_equation", + "content": "Y^j" + }, + { + "bbox": [ + 313, + 685, + 556, + 715 + ], + "type": "text", + "content": " evaluated w.r.t. class " + }, + { + "bbox": [ + 313, + 685, + 556, + 715 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 685, + 556, + 715 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 685, + 556, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 685, + 556, + 715 + ], + "type": "text", + "content": "-th" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25701" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 170, + 71, + 455, + 281 + ], + "blocks": [ + { + "bbox": [ + 170, + 71, + 455, + 281 + ], + "lines": [ + { + "bbox": [ + 170, + 71, + 455, + 281 + ], + "spans": [ + { + "bbox": [ + 170, + 71, + 455, + 281 + ], + "type": "image", + "image_path": "78528c4fe56ed138b42b8dff1ffcff491b6b4a70bfcf4e82965c64fa2dd36c37.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 290, + 555, + 327 + ], + "lines": [ + { + "bbox": [ + 54, + 290, + 555, + 327 + ], + "spans": [ + { + "bbox": [ + 54, + 290, + 555, + 327 + ], + "type": "text", + "content": "Figure 2. Demonstration of our debiased adversarial training framework. For each task, the predictors within the blue rectangle represent those identified as spurious by our method. During optimization, the per-task projector and predictor are updated to ensure that the projected representation minimizes reliance on information of spurious tasks while maintaining performance on the target task." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 346, + 154, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 346, + 154, + 359 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 154, + 359 + ], + "type": "text", + "content": "task, we have " + }, + { + "bbox": [ + 55, + 346, + 154, + 359 + ], + "type": "inline_equation", + "content": "\\rho_{Y^i,Y^j}" + }, + { + "bbox": [ + 55, + 346, + 154, + 359 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 102, + 361, + 249, + 399 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 361, + 249, + 399 + ], + "spans": [ + { + "bbox": [ + 102, + 361, + 249, + 399 + ], + "type": "interline_equation", + "content": "\\rho_ {Y ^ {i}, Y ^ {j}} = \\frac {\\left(\\frac {1}{(1 + \\alpha_ {1} ^ {i j})} - p ^ {j}\\right) p ^ {i}}{\\sqrt {p ^ {i} p ^ {j} (1 - p ^ {i}) (1 - p ^ {j})}}.", + "image_path": "442ba4f83ad3003cf9424756b4afeb6ddd5c1eeced59a7eb0abda99ad022c5f5.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 409, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 409, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 56, + 409, + 296, + 693 + ], + "type": "text", + "content": "We defer full proof to the Appendix. As shown in Lemma 1, the correlation coefficient between " + }, + { + "bbox": [ + 56, + 409, + 296, + 693 + ], + "type": "inline_equation", + "content": "Y^{i}" + }, + { + "bbox": [ + 56, + 409, + 296, + 693 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 409, + 296, + 693 + ], + "type": "inline_equation", + "content": "Y^{j}" + }, + { + "bbox": [ + 56, + 409, + 296, + 693 + ], + "type": "text", + "content": " is determined by the marginal distributions of " + }, + { + "bbox": [ + 56, + 409, + 296, + 693 + ], + "type": "inline_equation", + "content": "Y^{i}" + }, + { + "bbox": [ + 56, + 409, + 296, + 693 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 409, + 296, + 693 + ], + "type": "inline_equation", + "content": "Y^{j}" + }, + { + "bbox": [ + 56, + 409, + 296, + 693 + ], + "type": "text", + "content": " and the ratio of joint probabilities. However, the dependence of " + }, + { + "bbox": [ + 56, + 409, + 296, + 693 + ], + "type": "inline_equation", + "content": "\\rho_{Y^i,Y^j}" + }, + { + "bbox": [ + 56, + 409, + 296, + 693 + ], + "type": "text", + "content": " on the marginal distributions could introduce potential errors in identifying spurious correlations, particularly in the presence of label imbalance, where the disproportionality between labeling pairs could be overlooked given small marginal probabilities of task labels. For example, even with a small " + }, + { + "bbox": [ + 56, + 409, + 296, + 693 + ], + "type": "inline_equation", + "content": "\\alpha_{1}^{ij}" + }, + { + "bbox": [ + 56, + 409, + 296, + 693 + ], + "type": "text", + "content": ", when " + }, + { + "bbox": [ + 56, + 409, + 296, + 693 + ], + "type": "inline_equation", + "content": "p^i" + }, + { + "bbox": [ + 56, + 409, + 296, + 693 + ], + "type": "text", + "content": " becomes trivial due to class imbalance, we may observe " + }, + { + "bbox": [ + 56, + 409, + 296, + 693 + ], + "type": "inline_equation", + "content": "\\rho_{Y^i,Y^j}\\approx 0" + }, + { + "bbox": [ + 56, + 409, + 296, + 693 + ], + "type": "text", + "content": ", despite the presence of spurious correlations between the two tasks. We therefore consider the following class-wise resampling strategy to perturb the marginal distributions of task labels, such that when " + }, + { + "bbox": [ + 56, + 409, + 296, + 693 + ], + "type": "inline_equation", + "content": "p^i" + }, + { + "bbox": [ + 56, + 409, + 296, + 693 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 409, + 296, + 693 + ], + "type": "inline_equation", + "content": "p^j" + }, + { + "bbox": [ + 56, + 409, + 296, + 693 + ], + "type": "text", + "content": " change, we have " + }, + { + "bbox": [ + 56, + 409, + 296, + 693 + ], + "type": "inline_equation", + "content": "\\rho_{Y^i,Y^j}" + }, + { + "bbox": [ + 56, + 409, + 296, + 693 + ], + "type": "text", + "content": " change, which enables us to recognize the spurious correlation between " + }, + { + "bbox": [ + 56, + 409, + 296, + 693 + ], + "type": "inline_equation", + "content": "Y^{i}" + }, + { + "bbox": [ + 56, + 409, + 296, + 693 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 409, + 296, + 693 + ], + "type": "inline_equation", + "content": "Y^{j}" + }, + { + "bbox": [ + 56, + 409, + 296, + 693 + ], + "type": "text", + "content": " while mitigating the influence of label imbalance. Specifically, let " + }, + { + "bbox": [ + 56, + 409, + 296, + 693 + ], + "type": "inline_equation", + "content": "Y^{\\prime i}" + }, + { + "bbox": [ + 56, + 409, + 296, + 693 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 409, + 296, + 693 + ], + "type": "inline_equation", + "content": "Y^{\\prime j}" + }, + { + "bbox": [ + 56, + 409, + 296, + 693 + ], + "type": "text", + "content": " be the random variable of task labels after resampling, let " + }, + { + "bbox": [ + 56, + 409, + 296, + 693 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{Y^j\\neq y}^{\\mathrm{train}}\\coloneqq \\{(x_i,y_i^1,\\dots ,y_i^k)|y_i^j\\neq y\\}" + }, + { + "bbox": [ + 56, + 409, + 296, + 693 + ], + "type": "text", + "content": " be the subset of samples excluding class " + }, + { + "bbox": [ + 56, + 409, + 296, + 693 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 56, + 409, + 296, + 693 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 56, + 409, + 296, + 693 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 56, + 409, + 296, + 693 + ], + "type": "text", + "content": "-th task and " + }, + { + "bbox": [ + 56, + 409, + 296, + 693 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{Y^{\\prime j}\\neq y}^{\\mathrm{train}}" + }, + { + "bbox": [ + 56, + 409, + 296, + 693 + ], + "type": "text", + "content": " be the corresponding subset after resampling, we consider the following down-sampling scheme within " + }, + { + "bbox": [ + 56, + 409, + 296, + 693 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{Y^j\\neq y}^{\\mathrm{train}}" + }, + { + "bbox": [ + 56, + 409, + 296, + 693 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 125, + 695, + 227, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 695, + 227, + 717 + ], + "spans": [ + { + "bbox": [ + 125, + 695, + 227, + 717 + ], + "type": "interline_equation", + "content": "\\left| \\mathcal {D} _ {Y ^ {\\prime j} \\neq y} ^ {' \\mathrm {t r a i n}} \\right| = \\xi \\left| \\mathcal {D} _ {Y ^ {j} \\neq y} ^ {\\mathrm {t r a i n}} \\right|,", + "image_path": "e561e3aecbbe00a1ec7dfc415f830440a9e8061f95dd90a7e5830c60d4ac31e3.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 345, + 555, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 345, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 345, + 555, + 472 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 345, + 555, + 472 + ], + "type": "inline_equation", + "content": "\\xi < 1" + }, + { + "bbox": [ + 313, + 345, + 555, + 472 + ], + "type": "text", + "content": " is the constant that determines the degree of resampling, and " + }, + { + "bbox": [ + 313, + 345, + 555, + 472 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{Y^{\\prime j}\\neq y}^{\\prime \\mathrm{train}}" + }, + { + "bbox": [ + 313, + 345, + 555, + 472 + ], + "type": "text", + "content": " is constructed by random sampling. Under class-wise resampling, we intend to amplify the effect of correlations between task labels under different values of " + }, + { + "bbox": [ + 313, + 345, + 555, + 472 + ], + "type": "inline_equation", + "content": "Y^{\\prime j}" + }, + { + "bbox": [ + 313, + 345, + 555, + 472 + ], + "type": "text", + "content": " within " + }, + { + "bbox": [ + 313, + 345, + 555, + 472 + ], + "type": "inline_equation", + "content": "\\mathbb{S}^j" + }, + { + "bbox": [ + 313, + 345, + 555, + 472 + ], + "type": "text", + "content": ", so as to properly capture the dependence between task labels. Let " + }, + { + "bbox": [ + 313, + 345, + 555, + 472 + ], + "type": "inline_equation", + "content": "p^{\\prime i}" + }, + { + "bbox": [ + 313, + 345, + 555, + 472 + ], + "type": "text", + "content": " be the expectation of " + }, + { + "bbox": [ + 313, + 345, + 555, + 472 + ], + "type": "inline_equation", + "content": "Y^{\\prime i}" + }, + { + "bbox": [ + 313, + 345, + 555, + 472 + ], + "type": "text", + "content": " evaluated on the resampled set " + }, + { + "bbox": [ + 313, + 345, + 555, + 472 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{Y^{\\prime j}\\neq y}^{\\prime \\mathrm{train}}\\cup \\left(\\mathcal{D}^{\\mathrm{train}}\\backslash \\mathcal{D}_{Y^{\\prime j}\\neq y}^{\\mathrm{train}}\\right)" + }, + { + "bbox": [ + 313, + 345, + 555, + 472 + ], + "type": "text", + "content": ", the following theorem states the feasibility of identifying spurious correlations between tasks based on the difference in correlation coefficients before and after resampling:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 478, + 555, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 478, + 555, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 478, + 555, + 528 + ], + "type": "text", + "content": "Theorem 1. When spurious correlation exists between two tasks, the maximum difference in correlation coefficient before and after resampling is lower-bounded by the difference in expectations " + }, + { + "bbox": [ + 313, + 478, + 555, + 528 + ], + "type": "inline_equation", + "content": "p^i - p'^i" + }, + { + "bbox": [ + 313, + 478, + 555, + 528 + ], + "type": "text", + "content": ", up to multiplicative constants:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 359, + 532, + 509, + 555 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 532, + 509, + 555 + ], + "spans": [ + { + "bbox": [ + 359, + 532, + 509, + 555 + ], + "type": "interline_equation", + "content": "\\left| \\rho_ {Y ^ {i}, Y ^ {j}} - \\rho_ {Y ^ {\\prime i}, Y ^ {\\prime j}} \\right| \\geq \\mathcal {K} \\left| p ^ {' i} - p ^ {i} \\right|,", + "image_path": "8ad2dbde0774b34675ad1fed6fc0dcb21e5bd671dce78f30a845dc4829cf7c0b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 560, + 550, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 560, + 550, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 560, + 550, + 590 + ], + "type": "interline_equation", + "content": "w h e r e \\mathcal {K} = \\frac {\\left(\\frac {1}{\\hat {\\alpha} _ {1} ^ {i j}} - \\hat {\\alpha} _ {1} ^ {i j}\\right)}{\\sqrt {(1 - p ^ {i}) (1 - (\\hat {\\alpha} _ {1} ^ {i j} - \\hat {\\alpha} _ {0} ^ {i j}) p ^ {i} - \\hat {\\alpha} _ {0} ^ {i j})}} a n d \\hat {\\alpha} _ {y} ^ {i j} = \\frac {1}{1 + \\alpha_ {y} ^ {i j}}.", + "image_path": "3226176ddc73eb5781b4aaee452993f43fb24054c4d09abcdbefa5a5fe3c3cea.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": "We defer full proof to the Appendix. As shown in Theorem 1, since " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\alpha}_{y}^{ij}" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "p^i" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " remain constant before and after resampling, we have the lower-bound of difference in correlation coefficient before and after resampling monotonically increase as " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "p^{'i}" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " increases. Accordingly, even though " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "p^i" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " can be trivial, so long as " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "p^i" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "p^{'i}" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " differ significantly before and after resampling, we are guaranteed a substantial change in the correlation coefficient, which enables us to identify the spurious correlations among tasks. Furthermore, the following remark states the difference in correla" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25702" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": "tion coefficient when spurious correlations do not exist between tasks:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 103, + 295, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 103, + 295, + 129 + ], + "spans": [ + { + "bbox": [ + 55, + 103, + 295, + 129 + ], + "type": "text", + "content": "Remark 1. When two tasks are not spuriously correlated, we have " + }, + { + "bbox": [ + 55, + 103, + 295, + 129 + ], + "type": "inline_equation", + "content": "\\rho_{Y^i,Y^j} - \\rho_{Y'^i,Y'^j} = 0" + }, + { + "bbox": [ + 55, + 103, + 295, + 129 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 136, + 296, + 160 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 136, + 296, + 160 + ], + "spans": [ + { + "bbox": [ + 55, + 136, + 296, + 160 + ], + "type": "text", + "content": "Consequently, let " + }, + { + "bbox": [ + 55, + 136, + 296, + 160 + ], + "type": "inline_equation", + "content": "\\mathbb{Z}^j" + }, + { + "bbox": [ + 55, + 136, + 296, + 160 + ], + "type": "text", + "content": " be the set of indices of identified spurious tasks for " + }, + { + "bbox": [ + 55, + 136, + 296, + 160 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 136, + 296, + 160 + ], + "type": "text", + "content": "-th task, we have" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 167, + 247, + 189 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 167, + 247, + 189 + ], + "spans": [ + { + "bbox": [ + 104, + 167, + 247, + 189 + ], + "type": "interline_equation", + "content": "\\left| \\rho_ {Y ^ {i}, Y ^ {j}} - \\rho_ {Y ^ {\\prime i}, Y ^ {\\prime j}} \\right| > \\delta , \\forall i \\in \\mathbb {Z} ^ {j},", + "image_path": "76c01c3cf1ce0569a9f6c7ac40d8d5687c4da5c1bf76cdbfecc528c423bdcf19.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 196, + 296, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 196, + 296, + 260 + ], + "spans": [ + { + "bbox": [ + 55, + 196, + 296, + 260 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 196, + 296, + 260 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 196, + 296, + 260 + ], + "type": "text", + "content": " is the threshold for distinguishing the spurious correlations between tasks. In practice, " + }, + { + "bbox": [ + 55, + 196, + 296, + 260 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 196, + 296, + 260 + ], + "type": "text", + "content": " can be trivially obtained owing to the distinctions in " + }, + { + "bbox": [ + 55, + 196, + 296, + 260 + ], + "type": "inline_equation", + "content": "\\left|\\rho_{Y^i,Y^j} - \\rho_{Y'^i,Y'^j}\\right|" + }, + { + "bbox": [ + 55, + 196, + 296, + 260 + ], + "type": "text", + "content": " as stated in Theorem 1 and Remark 1. We refer to Sec. 5.3 for the detailed discussion." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 268, + 222, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 268, + 222, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 268, + 222, + 281 + ], + "type": "text", + "content": "4.2. Debiased Adversarial Training" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 285, + 296, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 285, + 296, + 477 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 296, + 477 + ], + "type": "text", + "content": "We move on to discuss the mitigation of spurious correlation between tasks. Due to the potential absence of certain label pairs in the training set, which impedes constructing balanced data even after identifying spurious tasks, we instead aim to ensure that the task-specific predictors effectively discard non-causal features for each task. In light of the multi-tasking property, we can readily formulate the objective as an adversarial learning problem w.r.t. the pertask predictors. Let " + }, + { + "bbox": [ + 55, + 285, + 296, + 477 + ], + "type": "inline_equation", + "content": "\\hat{f}" + }, + { + "bbox": [ + 55, + 285, + 296, + 477 + ], + "type": "text", + "content": " be the function of pre-trained encoder which remains frozen during fine-tuning, regarding " + }, + { + "bbox": [ + 55, + 285, + 296, + 477 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 285, + 296, + 477 + ], + "type": "text", + "content": "-th task, given its identified set of spurious tasks " + }, + { + "bbox": [ + 55, + 285, + 296, + 477 + ], + "type": "inline_equation", + "content": "\\mathbb{Z}^j" + }, + { + "bbox": [ + 55, + 285, + 296, + 477 + ], + "type": "text", + "content": ", to utilize the per-task predictors as discriminators of non-causal features, we apply one extra projector " + }, + { + "bbox": [ + 55, + 285, + 296, + 477 + ], + "type": "inline_equation", + "content": "h^j: \\mathbb{R}^l \\to \\mathbb{R}^l" + }, + { + "bbox": [ + 55, + 285, + 296, + 477 + ], + "type": "text", + "content": " before " + }, + { + "bbox": [ + 55, + 285, + 296, + 477 + ], + "type": "inline_equation", + "content": "g^j" + }, + { + "bbox": [ + 55, + 285, + 296, + 477 + ], + "type": "text", + "content": ", such that the prediction for " + }, + { + "bbox": [ + 55, + 285, + 296, + 477 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 285, + 296, + 477 + ], + "type": "text", + "content": "-th task is given by " + }, + { + "bbox": [ + 55, + 285, + 296, + 477 + ], + "type": "inline_equation", + "content": "g^i \\circ h^i \\circ \\hat{f}(x)" + }, + { + "bbox": [ + 55, + 285, + 296, + 477 + ], + "type": "text", + "content": ". We consider the following objective for mitigating spurious correlations regarding " + }, + { + "bbox": [ + 55, + 285, + 296, + 477 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 285, + 296, + 477 + ], + "type": "text", + "content": "-th task:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 66, + 495, + 285, + 521 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 495, + 285, + 521 + ], + "spans": [ + { + "bbox": [ + 66, + 495, + 285, + 521 + ], + "type": "interline_equation", + "content": "\\operatorname *{arg min}_{\\boldsymbol{h}^{j},\\boldsymbol{g}^{j}}\\bar{\\mathcal{L}}^{j}\\left(\\mathcal{D}^{\\text{train}};\\boldsymbol{h}^{j},\\boldsymbol{g}^{j}\\right) - \\sum_{t\\in \\mathbb{Z}^{j}}\\lambda^{j}\\bar{\\mathcal{L}}^{t}\\left(\\mathcal{D}^{\\text{train}};\\boldsymbol{h}^{j},\\boldsymbol{g}^{t}\\right),", + "image_path": "fb4b6e3c3e135171142baebfeaf497371a89c67c780ad489b716bd97c5e67935.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 530, + 296, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 530, + 296, + 642 + ], + "spans": [ + { + "bbox": [ + 55, + 530, + 296, + 642 + ], + "type": "text", + "content": "where the optimization is performed over " + }, + { + "bbox": [ + 55, + 530, + 296, + 642 + ], + "type": "inline_equation", + "content": "\\pmb{h}^j, \\pmb{g}^j" + }, + { + "bbox": [ + 55, + 530, + 296, + 642 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 55, + 530, + 296, + 642 + ], + "type": "inline_equation", + "content": "\\pmb{g}^t" + }, + { + "bbox": [ + 55, + 530, + 296, + 642 + ], + "type": "text", + "content": " remains frozen " + }, + { + "bbox": [ + 55, + 530, + 296, + 642 + ], + "type": "inline_equation", + "content": "\\forall t \\in \\mathbb{Z}^j" + }, + { + "bbox": [ + 55, + 530, + 296, + 642 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 530, + 296, + 642 + ], + "type": "inline_equation", + "content": "\\bar{\\mathcal{L}}^j(\\mathcal{D}^{\\mathrm{train}}; \\pmb{h}^j, \\pmb{g}^j) := \\frac{1}{N} \\sum_{i=1}^{N} \\mathcal{L}^j(\\pmb{g}^j(\\pmb{h}^j(\\hat{\\pmb{f}}(x_i))), y_i^j), \\quad \\bar{\\mathcal{L}}^t(\\mathcal{D}^{\\mathrm{train}}; \\pmb{h}^j, \\pmb{g}^t) := \\frac{1}{N} \\sum_{i=1}^{N} \\mathcal{L}^k(\\pmb{g}^k(\\pmb{h}^j(\\pmb{f}(x_i))), y_i^k)" + }, + { + "bbox": [ + 55, + 530, + 296, + 642 + ], + "type": "text", + "content": " are the average training loss under projection. By minimizing the objective, we intend to ensure that the projected latent representation retains only the information pertinent to " + }, + { + "bbox": [ + 55, + 530, + 296, + 642 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 530, + 296, + 642 + ], + "type": "text", + "content": "-th task while shedding the spurious information. We summarize our method in Algorithm 1." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 652, + 137, + 666 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 652, + 137, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 652, + 137, + 666 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 672, + 172, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 172, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 172, + 685 + ], + "type": "text", + "content": "5.1. Experimental Setup" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "We validate our method on the following benchmark datasets: Multi-MNIST dataset [38], CelebA dataset [25]," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 72, + 489, + 84 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 489, + 84 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 489, + 84 + ], + "type": "text", + "content": "Algorithm 1 Debiased adversarial training" + } + ] + } + ], + "index": 12, + "type": "text" + }, + { + "bbox": [ + 313, + 97, + 552, + 147 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 97, + 552, + 147 + ], + "spans": [ + { + "bbox": [ + 313, + 97, + 552, + 147 + ], + "type": "text", + "content": "Input: Pre-trained encoder " + }, + { + "bbox": [ + 313, + 97, + 552, + 147 + ], + "type": "inline_equation", + "content": "\\hat{f}" + }, + { + "bbox": [ + 313, + 97, + 552, + 147 + ], + "type": "text", + "content": ", task-specific predictors " + }, + { + "bbox": [ + 313, + 97, + 552, + 147 + ], + "type": "inline_equation", + "content": "g^{j}" + }, + { + "bbox": [ + 313, + 97, + 552, + 147 + ], + "type": "text", + "content": "'s s.t. " + }, + { + "bbox": [ + 313, + 97, + 552, + 147 + ], + "type": "inline_equation", + "content": "1 \\leq j \\leq k" + }, + { + "bbox": [ + 313, + 97, + 552, + 147 + ], + "type": "text", + "content": ", task-specific projectors " + }, + { + "bbox": [ + 313, + 97, + 552, + 147 + ], + "type": "inline_equation", + "content": "h^{j}" + }, + { + "bbox": [ + 313, + 97, + 552, + 147 + ], + "type": "text", + "content": "'s, training data " + }, + { + "bbox": [ + 313, + 97, + 552, + 147 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{train}}" + }, + { + "bbox": [ + 313, + 97, + 552, + 147 + ], + "type": "text", + "content": ", training epochs " + }, + { + "bbox": [ + 313, + 97, + 552, + 147 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 313, + 97, + 552, + 147 + ], + "type": "text", + "content": ", learning rate " + }, + { + "bbox": [ + 313, + 97, + 552, + 147 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 97, + 552, + 147 + ], + "type": "text", + "content": ", threshold " + }, + { + "bbox": [ + 313, + 97, + 552, + 147 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 97, + 552, + 147 + ], + "type": "text", + "content": ", resampling constant " + }, + { + "bbox": [ + 313, + 97, + 552, + 147 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 313, + 97, + 552, + 147 + ], + "type": "text", + "content": ", hyper-parameter " + }, + { + "bbox": [ + 313, + 97, + 552, + 147 + ], + "type": "inline_equation", + "content": "\\lambda^{j}" + }, + { + "bbox": [ + 313, + 97, + 552, + 147 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 319, + 155, + 504, + 228 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 319, + 155, + 403, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 155, + 403, + 167 + ], + "spans": [ + { + "bbox": [ + 319, + 155, + 403, + 167 + ], + "type": "text", + "content": "1: for " + }, + { + "bbox": [ + 319, + 155, + 403, + 167 + ], + "type": "inline_equation", + "content": "j = 1" + }, + { + "bbox": [ + 319, + 155, + 403, + 167 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 319, + 155, + 403, + 167 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 319, + 155, + 403, + 167 + ], + "type": "text", + "content": " do:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 320, + 168, + 378, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 168, + 378, + 178 + ], + "spans": [ + { + "bbox": [ + 320, + 168, + 378, + 178 + ], + "type": "text", + "content": "2: " + }, + { + "bbox": [ + 320, + 168, + 378, + 178 + ], + "type": "inline_equation", + "content": "\\mathbb{Z}^j = \\emptyset" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 180, + 457, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 180, + 457, + 191 + ], + "spans": [ + { + "bbox": [ + 320, + 180, + 457, + 191 + ], + "type": "text", + "content": "3: for " + }, + { + "bbox": [ + 320, + 180, + 457, + 191 + ], + "type": "inline_equation", + "content": "i = 1" + }, + { + "bbox": [ + 320, + 180, + 457, + 191 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 320, + 180, + 457, + 191 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 320, + 180, + 457, + 191 + ], + "type": "text", + "content": " s.t. " + }, + { + "bbox": [ + 320, + 180, + 457, + 191 + ], + "type": "inline_equation", + "content": "i \\neq j" + }, + { + "bbox": [ + 320, + 180, + 457, + 191 + ], + "type": "text", + "content": " do:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 191, + 470, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 191, + 470, + 204 + ], + "spans": [ + { + "bbox": [ + 320, + 191, + 470, + 204 + ], + "type": "text", + "content": "4: Calculate " + }, + { + "bbox": [ + 320, + 191, + 470, + 204 + ], + "type": "inline_equation", + "content": "\\rho_{Y^i,Y^j}" + }, + { + "bbox": [ + 320, + 191, + 470, + 204 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 320, + 191, + 470, + 204 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{train}}" + }, + { + "bbox": [ + 320, + 191, + 470, + 204 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 204, + 422, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 204, + 422, + 215 + ], + "spans": [ + { + "bbox": [ + 320, + 204, + 422, + 215 + ], + "type": "text", + "content": "5: for " + }, + { + "bbox": [ + 320, + 204, + 422, + 215 + ], + "type": "inline_equation", + "content": "y\\in \\mathbb{S}^j" + }, + { + "bbox": [ + 320, + 204, + 422, + 215 + ], + "type": "text", + "content": " do:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 216, + 504, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 216, + 504, + 228 + ], + "spans": [ + { + "bbox": [ + 320, + 216, + 504, + 228 + ], + "type": "text", + "content": "6: Perform class-wise resampling:" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 392, + 236, + 493, + 257 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 236, + 493, + 257 + ], + "spans": [ + { + "bbox": [ + 392, + 236, + 493, + 257 + ], + "type": "interline_equation", + "content": "\\left| \\mathcal {D} _ {Y ^ {\\prime j} \\neq y} ^ {\\text {t r a i n}} \\right| = \\xi \\left| \\mathcal {D} _ {Y ^ {j} \\neq y} ^ {\\text {t r a i n}} \\right|.", + "image_path": "5cf7a780bfb7eef7d5c2dfaaaad8816e8afd69723c4aa5a29e99c1f2f73cfcbc.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 265, + 465, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 265, + 465, + 277 + ], + "spans": [ + { + "bbox": [ + 320, + 265, + 465, + 277 + ], + "type": "text", + "content": "7: Calculate " + }, + { + "bbox": [ + 320, + 265, + 465, + 277 + ], + "type": "inline_equation", + "content": "\\rho_{Y^{\\prime}i,Y^{\\prime}j}" + }, + { + "bbox": [ + 320, + 265, + 465, + 277 + ], + "type": "text", + "content": " on" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 384, + 287, + 501, + 308 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 287, + 501, + 308 + ], + "spans": [ + { + "bbox": [ + 384, + 287, + 501, + 308 + ], + "type": "interline_equation", + "content": "\\mathcal {D} _ {Y ^ {\\prime j} \\neq y} ^ {\\text {t r a i n}} \\cup \\left(\\mathcal {D} ^ {\\text {t r a i n}} \\backslash \\mathcal {D} _ {Y ^ {j} \\neq y} ^ {\\text {t r a i n}}\\right).", + "image_path": "a06a5551ea5eda276bf443c17e9c4e489eff9000a3ff0370614f87d55718ae5e.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 318, + 316, + 548, + 426 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 320, + 316, + 548, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 316, + 548, + 329 + ], + "spans": [ + { + "bbox": [ + 320, + 316, + 548, + 329 + ], + "type": "text", + "content": "8: if " + }, + { + "bbox": [ + 320, + 316, + 548, + 329 + ], + "type": "inline_equation", + "content": "\\rho_{Y^i,Y^j} - \\rho_{Y'^i,Y'^j}\\geq \\delta" + }, + { + "bbox": [ + 320, + 316, + 548, + 329 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 320, + 316, + 548, + 329 + ], + "type": "inline_equation", + "content": "i\\notin \\mathbb{Z}^j" + }, + { + "bbox": [ + 320, + 316, + 548, + 329 + ], + "type": "text", + "content": " then:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 331, + 456, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 331, + 456, + 342 + ], + "spans": [ + { + "bbox": [ + 320, + 331, + 456, + 342 + ], + "type": "text", + "content": "9: " + }, + { + "bbox": [ + 320, + 331, + 456, + 342 + ], + "type": "inline_equation", + "content": "\\mathbb{Z}^j = \\mathbb{Z}^j\\cup \\{i\\}" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 319, + 342, + 404, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 342, + 404, + 352 + ], + "spans": [ + { + "bbox": [ + 319, + 342, + 404, + 352 + ], + "type": "text", + "content": "10: end if" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 318, + 355, + 395, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 355, + 395, + 365 + ], + "spans": [ + { + "bbox": [ + 318, + 355, + 395, + 365 + ], + "type": "text", + "content": "11: end for" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 318, + 367, + 380, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 367, + 380, + 376 + ], + "spans": [ + { + "bbox": [ + 318, + 367, + 380, + 376 + ], + "type": "text", + "content": "12: end for" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 318, + 379, + 365, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 379, + 365, + 388 + ], + "spans": [ + { + "bbox": [ + 318, + 379, + 365, + 388 + ], + "type": "text", + "content": "13: end for" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 318, + 390, + 426, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 390, + 426, + 400 + ], + "spans": [ + { + "bbox": [ + 318, + 390, + 426, + 400 + ], + "type": "text", + "content": "14: for epoch " + }, + { + "bbox": [ + 318, + 390, + 426, + 400 + ], + "type": "inline_equation", + "content": "= 1" + }, + { + "bbox": [ + 318, + 390, + 426, + 400 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 318, + 390, + 426, + 400 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 318, + 390, + 426, + 400 + ], + "type": "text", + "content": " do:" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 318, + 402, + 418, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 402, + 418, + 413 + ], + "spans": [ + { + "bbox": [ + 318, + 402, + 418, + 413 + ], + "type": "text", + "content": "15: for " + }, + { + "bbox": [ + 318, + 402, + 418, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{j} = 1" + }, + { + "bbox": [ + 318, + 402, + 418, + 413 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 318, + 402, + 418, + 413 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 318, + 402, + 418, + 413 + ], + "type": "text", + "content": " do:" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 318, + 415, + 418, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 415, + 418, + 426 + ], + "spans": [ + { + "bbox": [ + 318, + 415, + 418, + 426 + ], + "type": "text", + "content": "16: Training loss:" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 340, + 434, + 544, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 434, + 544, + 460 + ], + "spans": [ + { + "bbox": [ + 340, + 434, + 544, + 460 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\bar {\\mathcal {L}} ^ {j} \\left(\\mathcal {D} ^ {\\text {t r a i n}}; \\boldsymbol {h} ^ {j}, \\boldsymbol {g} ^ {j}\\right) - \\sum_ {t \\in \\mathbb {Z} ^ {j}} \\lambda^ {j} \\bar {\\mathcal {L}} ^ {t} \\left(\\mathcal {D} ^ {\\text {t r a i n}}; \\boldsymbol {h} ^ {j}, \\boldsymbol {g} ^ {t}\\right).", + "image_path": "914aaba20cdea1c51562bc2f39585811c4233acfe559326c7fa48fbd5e5fee59.jpg" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 317, + 470, + 518, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 470, + 518, + 483 + ], + "spans": [ + { + "bbox": [ + 317, + 470, + 518, + 483 + ], + "type": "text", + "content": "17: Update " + }, + { + "bbox": [ + 317, + 470, + 518, + 483 + ], + "type": "inline_equation", + "content": "h^j" + }, + { + "bbox": [ + 317, + 470, + 518, + 483 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 317, + 470, + 518, + 483 + ], + "type": "inline_equation", + "content": "g^j" + }, + { + "bbox": [ + 317, + 470, + 518, + 483 + ], + "type": "text", + "content": " by gradient descent:" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 403, + 491, + 481, + 515 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 491, + 481, + 515 + ], + "spans": [ + { + "bbox": [ + 403, + 491, + 481, + 515 + ], + "type": "interline_equation", + "content": "\\boldsymbol {h} ^ {j} = \\boldsymbol {h} ^ {j} - \\eta \\frac {\\partial \\mathcal {L}}{\\partial \\boldsymbol {h} ^ {j}}.", + "image_path": "49cc6154baf26bbaaaa861f8e5ef47922e43a6ced120f0f456a568e6ddcb100e.jpg" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 405, + 521, + 479, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 521, + 479, + 548 + ], + "spans": [ + { + "bbox": [ + 405, + 521, + 479, + 548 + ], + "type": "interline_equation", + "content": "\\boldsymbol {g} ^ {j} = \\boldsymbol {g} ^ {j} - \\eta \\frac {\\partial \\mathcal {L}}{\\partial \\boldsymbol {g} ^ {j}}.", + "image_path": "6ad057ceb26c7559cc983ef779e4eef2a821138d91d4c7cf9493296df1c3a758.jpg" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 317, + 552, + 380, + 574 + ], + "type": "list", + "angle": 0, + "index": 40, + "blocks": [ + { + "bbox": [ + 317, + 552, + 380, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 552, + 380, + 562 + ], + "spans": [ + { + "bbox": [ + 317, + 552, + 380, + 562 + ], + "type": "text", + "content": "18: end for" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 317, + 564, + 365, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 564, + 365, + 574 + ], + "spans": [ + { + "bbox": [ + 317, + 564, + 365, + 574 + ], + "type": "text", + "content": "19: end for" + } + ] + } + ], + "index": 39 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 598, + 555, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 598, + 555, + 683 + ], + "spans": [ + { + "bbox": [ + 313, + 598, + 555, + 683 + ], + "type": "text", + "content": "Taskonomy dataset [45], MetaShift dataset [22], NYU-V2 dataset [35] and CityScape dataset [7]. We implement our method in PyTorch 2.0.0 on one RTX-3090 GPU1. We set " + }, + { + "bbox": [ + 313, + 598, + 555, + 683 + ], + "type": "inline_equation", + "content": "\\xi = 0.1" + }, + { + "bbox": [ + 313, + 598, + 555, + 683 + ], + "type": "text", + "content": " during experiments and evaluate the difference in correlation coefficients between training set and class-wise resampled sets to construct " + }, + { + "bbox": [ + 313, + 598, + 555, + 683 + ], + "type": "inline_equation", + "content": "\\mathbb{Z}^j" + }, + { + "bbox": [ + 313, + 598, + 555, + 683 + ], + "type": "text", + "content": ". We use accuracy as evaluation metric for classification tasks, cross-entropy loss and" + } + ] + } + ], + "index": 41 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 693, + 553, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 693, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 693, + 553, + 712 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 313, + 693, + 553, + 712 + ], + "type": "text", + "content": "Code available at https://github.com/cjy24/Spurious-Correlation-MTL." + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25703" + } + ] + } + ], + "index": 43 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 113, + 70, + 497, + 110 + ], + "blocks": [ + { + "bbox": [ + 113, + 70, + 497, + 110 + ], + "lines": [ + { + "bbox": [ + 113, + 70, + 497, + 110 + ], + "spans": [ + { + "bbox": [ + 113, + 70, + 497, + 110 + ], + "type": "table", + "html": "
TaskSTLSubSelMT-CRLMeta-learningGMTLOurs
left digit (accuracy, ↑)+4.9%+3.4%+5.1%+0.2%+5.9%+9.3%
right digit (accuracy, ↑)+5.9%+2.9%+6.3%+0.3%+7.2%+9.6%
", + "image_path": "efdd7f108465061e0e95b42ae60a670c281c753e5c344b0bce9657f3c3edeb87.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 105, + 150, + 506, + 215 + ], + "blocks": [ + { + "bbox": [ + 55, + 118, + 555, + 140 + ], + "lines": [ + { + "bbox": [ + 55, + 118, + 555, + 140 + ], + "spans": [ + { + "bbox": [ + 55, + 118, + 555, + 140 + ], + "type": "text", + "content": "Table 4. Experimental results on Multi-MNIST dataset. The relative improvement of each method is calculated in comparison to vanilla MTL." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 105, + 150, + 506, + 215 + ], + "lines": [ + { + "bbox": [ + 105, + 150, + 506, + 215 + ], + "spans": [ + { + "bbox": [ + 105, + 150, + 506, + 215 + ], + "type": "table", + "html": "
TaskSTLSubSelMT-CRLMeta-learningGMTLOurs
Gender (accuracy, ↑)+2.2%+1.7%+1.3%+2.4%+2.6%+5.7%
Smiling (accuracy, ↑)+2.3%+2.5%+2.6%+3.4%+1.2%+4.6%
Age (accuracy, ↑)+1.2%+2.2%+0.1%+1.1%+1.5%+3.9%
Attractiveness (accuracy, ↑)+0.6%+0.4%+0.7%+0.2%+1.2%+2.4%
", + "image_path": "5003814a0fc6c32815285bf67a998140505277751c3d14e4503e9fe5e9ee332b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 115, + 244, + 496, + 297 + ], + "blocks": [ + { + "bbox": [ + 215, + 222, + 394, + 233 + ], + "lines": [ + { + "bbox": [ + 215, + 222, + 394, + 233 + ], + "spans": [ + { + "bbox": [ + 215, + 222, + 394, + 233 + ], + "type": "text", + "content": "Table 5. Experimental results on CelebA dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 115, + 244, + 496, + 297 + ], + "lines": [ + { + "bbox": [ + 115, + 244, + 496, + 297 + ], + "spans": [ + { + "bbox": [ + 115, + 244, + 496, + 297 + ], + "type": "table", + "html": "
TaskSTLSubSelMT-CRLMeta-learningOurs
Scene (accuracy, ↑)+4.1%+3.2%+5.8%+5.3%+9.9%
Object (accuracy, ↑)+5.2%+4.1%+7.3%+8.1%+12.7%
Segmentation (cross entropy, ↓)-5.6%+1.7%+3.1%+6.9%+6.2%
", + "image_path": "5a5cd60c3d51ff3483df878e47030fc558876241618463936db0159292f53bdf.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 118, + 326, + 493, + 366 + ], + "blocks": [ + { + "bbox": [ + 208, + 304, + 401, + 316 + ], + "lines": [ + { + "bbox": [ + 208, + 304, + 401, + 316 + ], + "spans": [ + { + "bbox": [ + 208, + 304, + 401, + 316 + ], + "type": "text", + "content": "Table 6. Experimental results on Taskonomy dataset." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 118, + 326, + 493, + 366 + ], + "lines": [ + { + "bbox": [ + 118, + 326, + 493, + 366 + ], + "spans": [ + { + "bbox": [ + 118, + 326, + 493, + 366 + ], + "type": "table", + "html": "
TaskSTLSubSelMT-CRLMeta-learningGMTLOurs
Animal (accuracy, ↑)+4.2%+2.7%+3.2%+1.5%+0.2%+5.4%
Scene (accuracy, ↑)+3.1%+2.1%+1.6%+0.1%+1.7%+4.0%
", + "image_path": "0489d7171a07643534ead434fa4deb96b6e14fba597e1431c53fca7d52e4999d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 121, + 396, + 489, + 437 + ], + "blocks": [ + { + "bbox": [ + 211, + 374, + 399, + 385 + ], + "lines": [ + { + "bbox": [ + 211, + 374, + 399, + 385 + ], + "spans": [ + { + "bbox": [ + 211, + 374, + 399, + 385 + ], + "type": "text", + "content": "Table 7. Experimental results on MetaShift dataset." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 121, + 396, + 489, + 437 + ], + "lines": [ + { + "bbox": [ + 121, + 396, + 489, + 437 + ], + "spans": [ + { + "bbox": [ + 121, + 396, + 489, + 437 + ], + "type": "table", + "html": "
TaskSTLSubSelMT-CRLMeta-learningOurs
Segmentation (mean IoU, ↑)+2.7%+4.2%+6.9%+7.2%+11.3%
Depth (relative error, ↓)+3.1%+3.8%+7.5%+6.2%+10.2%
", + "image_path": "f4265470a7954471c3370878f7d374558aecea592b980dc3e69a6397ffad8f87.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 212, + 444, + 398, + 456 + ], + "lines": [ + { + "bbox": [ + 212, + 444, + 398, + 456 + ], + "spans": [ + { + "bbox": [ + 212, + 444, + 398, + 456 + ], + "type": "text", + "content": "Table 8. Experimental results on NYU-V2 dataset." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 476, + 295, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 476, + 295, + 513 + ], + "spans": [ + { + "bbox": [ + 55, + 476, + 295, + 513 + ], + "type": "text", + "content": "mean IoU for segmentation tasks and relative error for depth estimation tasks. We defer details of each dataset and full results to the appendix." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 517, + 295, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 295, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 295, + 649 + ], + "type": "text", + "content": "We compare our method with the following related methods on spurious correlation and generalization in MTL: Single-task learning (STL): the network is optimized over each task only without regularizations; MTL: vanilla MTL without regularizations; SubSel [21]: MTL with subset selection to address task conflicts; MT-CRL [14]: MTL with disentangled neural modules and invariant regularization on task-to-module routing; Meta-learning [11]: MTL with disentangled representation learning and bi-level optimization on per-task predictors; GMTL [27]: MTL with generative modeling of task labels." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "content": "We repeat experiments on each dataset three times and report the average results. For MT-CRL [14], the covariate shift on training data is constructed by training-validation split, as suggested by the authors. Further, we do not include GMTL [27] on Taskonomy, NYU-V2 and CityScape" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 476, + 555, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 476, + 555, + 537 + ], + "spans": [ + { + "bbox": [ + 313, + 476, + 555, + 537 + ], + "type": "text", + "content": "dataset, as it is infeasible to estimate the joint distribution involving high-dimensional masks. Following prior work [14], we report the relative performance improvement of each method over vanilla MTL for evaluation, as the scales of evaluation metrics differ significantly across tasks." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 550, + 440, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 550, + 440, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 550, + 440, + 563 + ], + "type": "text", + "content": "5.2. Experimental Results" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "type": "text", + "content": "Results are shown in Tab. 4 - 8. Compared with vanilla STL and MTL, Meta-learning [11] shows marginal improvement in testing performance, particularly when the marginal training distributions are evenly distributed per task. Regarding MT-CRL [14], the training-validation split does not appropriately capture the spurious correlations between tasks, where the method shows limited improvement on several tasks. Such observations are in line with our preliminary analysis: modeling spurious correlation between tasks using covariate shift is in general insufficient and lacks theoretical guarantee. Furthermore, GMTL [27] is limited to classification tasks, and its dependence on estimating" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25704" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 294, + 120 + ], + "type": "text", + "content": "joint distributions may introduce estimation errors, resulting in subpar performance on testing data. In comparison, our method shows better or comparable performance on all four datasets, validating the effectiveness of our method." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 129, + 282, + 142 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 129, + 282, + 142 + ], + "spans": [ + { + "bbox": [ + 55, + 129, + 282, + 142 + ], + "type": "text", + "content": "5.3. Empirical Validation of Theoretical Results" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 146, + 296, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 146, + 296, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 146, + 296, + 266 + ], + "type": "text", + "content": "We move on to discuss the validation of Theorem 1. Specifically, we consider three different variations of Multi-MNIST dataset: Spurious correlation (Sp.C.), with non-overlapping labeling pair between training and testing data, as in Sec. 5.1; Random assignment (R.A.), where the right digit is randomly assigned given the left digit on both training and testing data; Identical assignment (Id.A.), where the left digit is identical to the right digit on both training and testing data. Under identical assignment, the two tasks are no longer spuriously correlated." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 111, + 276, + 241, + 334 + ], + "blocks": [ + { + "bbox": [ + 111, + 276, + 241, + 334 + ], + "lines": [ + { + "bbox": [ + 111, + 276, + 241, + 334 + ], + "spans": [ + { + "bbox": [ + 111, + 276, + 241, + 334 + ], + "type": "table", + "html": "
DataρYl,Yr - ρY′l,Y′r
Sp.C.0.29 ± 0.04
R.A.0.002 ± 0.0006
Id.A.0.001 ± 0.0003
", + "image_path": "102e74a9d2ed42b6a91bdb0f3e46cc15a63d4a7abf749e9fcfb082e4a141bea7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 412, + 296, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 412, + 296, + 520 + ], + "spans": [ + { + "bbox": [ + 55, + 412, + 296, + 520 + ], + "type": "text", + "content": "Under each variation, we perform resampling on the left digit and report the maximum average change in correlation coefficient. The results are shown in Tab. 9. Compared with those under random assignment and identical assignment, when the two tasks are spuriously correlated, the difference in correlation coefficient shows a significant increase, substantiating the feasibility of identifying spurious correlations between tasks based on the difference in correlation coefficients, as proposed in Theorem 1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 528, + 149, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 528, + 149, + 541 + ], + "spans": [ + { + "bbox": [ + 55, + 528, + 149, + 541 + ], + "type": "text", + "content": "5.4. Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": "We further invest the effect of spurious correlation identification within our methodology under the three variations of Multi-MNIST dataset, and the necessity in distinguishing spurious correlation between tasks. Results are shown in Tab. 10. Under both spurious correlation and random assignment, the adversarial training framework shows similar performance, regardless of the identification of spurious correlations. Specifically, under random assignment, our framework demonstrates marginal improvement over STL and MTL baselines, irrespective of the construction of " + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbb{Z}^j" + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": "’s. This suggests that when the labels of two tasks are not correlated, the per-task predictors are also unlikely to be affected by the spurious features. Furthermore, under identical assignment, without correctly identifying the spurious" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 315, + 69, + 560, + 230 + ], + "blocks": [ + { + "bbox": [ + 55, + 342, + 296, + 399 + ], + "lines": [ + { + "bbox": [ + 55, + 342, + 296, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 296, + 399 + ], + "type": "text", + "content": "Table 9. Maximum average change in correlation coefficient under class-wise resampling on Multi-MNIST dataset. " + }, + { + "bbox": [ + 55, + 342, + 296, + 399 + ], + "type": "inline_equation", + "content": "Y^{l}" + }, + { + "bbox": [ + 55, + 342, + 296, + 399 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 342, + 296, + 399 + ], + "type": "inline_equation", + "content": "Y^{r}" + }, + { + "bbox": [ + 55, + 342, + 296, + 399 + ], + "type": "text", + "content": " refer to the random variables of left digit label and right digit label on training set, and " + }, + { + "bbox": [ + 55, + 342, + 296, + 399 + ], + "type": "inline_equation", + "content": "Y^{\\prime l}" + }, + { + "bbox": [ + 55, + 342, + 296, + 399 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 342, + 296, + 399 + ], + "type": "inline_equation", + "content": "Y^{\\prime r}" + }, + { + "bbox": [ + 55, + 342, + 296, + 399 + ], + "type": "text", + "content": " refer to the random variables of left digit label and right digit label on the resampled training set." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 69, + 560, + 230 + ], + "lines": [ + { + "bbox": [ + 315, + 69, + 560, + 230 + ], + "spans": [ + { + "bbox": [ + 315, + 69, + 560, + 230 + ], + "type": "table", + "html": "
DataMethodleft digitright digit
Sp.C.STL88.26±1.34%88.43±1.54%
Sp.C.MTL84.14±1.19%83.57±1.27%
Sp.C.Ours (w/o Zj)91.59±1.47%90.74±1.22%
Sp.C.Ours91.63±1.46%90.78±1.26%
R.A.STL93.26±1.46%92.87±1.25%
R.A.MTL93.44±1.57%93.67±1.28%
R.A.Ours (w/o Zj)93.48±1.76%93.25±1.29%
R.A.Ours93.87±1.24%93.73±1.36%
Id.A.STL93.67±1.36%93.68±1.37%
Id.A.MTL93.81±1.55%93.49±1.32%
Id.A.Ours (w/o Zj)91.73±1.45%91.85±1.46%
Id.A.Ours93.84±1.43%93.52±1.19%
", + "image_path": "c071c3f35470f71801473b13af357c95a3286ce63d48de58d840f9c5e23d81f5.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 238, + 555, + 348 + ], + "lines": [ + { + "bbox": [ + 313, + 238, + 555, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 238, + 555, + 348 + ], + "type": "text", + "content": "Table 10. Experimental results on Multi-MNIST dataset under random and identical assignments. Regarding 'Ours " + }, + { + "bbox": [ + 313, + 238, + 555, + 348 + ], + "type": "inline_equation", + "content": "(\\mathrm{w / o}\\mathbb{Z}^j)" + }, + { + "bbox": [ + 313, + 238, + 555, + 348 + ], + "type": "text", + "content": " we apply the debiased adversarial training per-task by taking the other task as the spurious task. Particularly, under identical assignment, applying debiased adversarial training without identifying " + }, + { + "bbox": [ + 313, + 238, + 555, + 348 + ], + "type": "inline_equation", + "content": "\\mathbb{Z}^j" + }, + { + "bbox": [ + 313, + 238, + 555, + 348 + ], + "type": "text", + "content": "’s leads to slightly worse performance than the baselines, indicating the necessity of distinguishing the spurious correlations between tasks. In comparison, our method performs comparably with the baselines, owing to the identification of spurious correlation between the two tasks." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 370, + 555, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 370, + 555, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 370, + 555, + 431 + ], + "type": "text", + "content": "information, the adversarial training framework performs slightly worse than the STL and MTL baselines, underscoring the importance of correctly constructing " + }, + { + "bbox": [ + 313, + 370, + 555, + 431 + ], + "type": "inline_equation", + "content": "\\mathbb{Z}^j" + }, + { + "bbox": [ + 313, + 370, + 555, + 431 + ], + "type": "text", + "content": "'s, within which our debiased adversarial training framework shall be applied to improve, rather than hinder, the generalizability." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 441, + 388, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 388, + 454 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 388, + 454 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 462, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 555, + 713 + ], + "type": "text", + "content": "Spurious correlations in MTL present a significant but inadequately addressed challenge. In this paper, we experimentally show the limitations of existing formulations in mitigating spurious correlations. Motivated by the discussion, we propose a novel framework for identifying spurious correlations between tasks. Our approach formulates the problem from a label-based perspective, where we propose to quantify the spurious correlations between tasks based on the difference in correlation coefficients, evaluated on both the training data and the class-wise resampled training data. We show theoretically the feasibility of our framework in distinguishing spurious correlations between tasks. Based on our identification framework, we propose an adversarial training framework for mitigating between-task spurious correlations, which fine-tunes only the task-specific predictors to ensure that task-specific predictors sufficiently discard the spurious information. We validate from experiments on four benchmark datasets that our method outperforms existing baselines in mitigating spurious correlations between tasks, as well as the effectiveness of our method in identifying spurious correlations between tasks." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "25705" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 140 + ], + "type": "text", + "content": "This work was partially supported by the EMBRIO Institute, contract #2120200, a National Science Foundation (NSF) Biology Integration Institute, and NSF IIS #1955890, IIS #2146091, IIS #2345235." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 115, + 165 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 115, + 165 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 115, + 165 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 171, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 171, + 296, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 171, + 296, + 215 + ], + "spans": [ + { + "bbox": [ + 61, + 171, + 296, + 215 + ], + "type": "text", + "content": "[1] Ibrahim M Alabdulmohsin, Behnam Neyshabur, and Xiaohua Zhai. Revisiting neural scaling laws in language and vision. Advances in Neural Information Processing Systems, 35:22300-22312, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 217, + 296, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 217, + 296, + 272 + ], + "spans": [ + { + "bbox": [ + 61, + 217, + 296, + 272 + ], + "type": "text", + "content": "[2] Simon Buchholz, Goutham Rajendran, Elan Rosenfeld, Bryon Aragam, Bernhard Schölkopf, and Pradeep Ravikumar. Learning linear causal representations from interventions under general nonlinear mixing. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 274, + 296, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 274, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 61, + 274, + 296, + 316 + ], + "type": "text", + "content": "[3] Yuan Cao and Quanquan Gu. Generalization bounds of stochastic gradient descent for wide and deep neural networks. Advances in neural information processing systems, 32, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 319, + 294, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 319, + 294, + 341 + ], + "spans": [ + { + "bbox": [ + 62, + 319, + 294, + 341 + ], + "type": "text", + "content": "[4] Rich Caruana. Multitask learning. Machine learning, 28: 41-75, 1997. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 342, + 295, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 342, + 295, + 386 + ], + "spans": [ + { + "bbox": [ + 62, + 342, + 295, + 386 + ], + "type": "text", + "content": "[5] Yimeng Chen, Ruibin Xiong, Zhi-Ming Ma, and Yanyan Lan. When does group invariant learning survive spurious correlations? Advances in Neural Information Processing Systems, 35:7038-7051, 2022. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 388, + 296, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 388, + 296, + 442 + ], + "spans": [ + { + "bbox": [ + 62, + 388, + 296, + 442 + ], + "type": "text", + "content": "[6] Zhao Chen, Jiquan Ngiam, Yanping Huang, Thang Luong, Henrik Kretzschmar, Yuning Chai, and Dragomir Anguelov. Just pick a sign: Optimizing deep multitask models with gradient sign dropout. Advances in Neural Information Processing Systems, 33:2039-2050, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 445, + 296, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 445, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 62, + 445, + 296, + 510 + ], + "type": "text", + "content": "[7] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 6, 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 512, + 296, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 512, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 62, + 512, + 296, + 555 + ], + "type": "text", + "content": "[8] Yuyang Deng, Junyuan Hong, Jiayu Zhou, and Mehrdad Mahdavi. On the generalization ability of unsupervised pretraining. In International Conference on Artificial Intelligence and Statistics, pages 4519-4527. PMLR, 2024. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 557, + 296, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 557, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 62, + 557, + 296, + 601 + ], + "type": "text", + "content": "[9] Nikolaos Dimitriadis, Pascal Frossard, and François Fleuret. Pareto manifold learning: Tackling multiple tasks via ensembles of single-task models. In International Conference on Machine Learning, pages 8015-8052. PMLR, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 602, + 295, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 602, + 295, + 646 + ], + "spans": [ + { + "bbox": [ + 57, + 602, + 295, + 646 + ], + "type": "text", + "content": "[10] Chris Fifty, Ehsan Amid, Zhe Zhao, Tianhe Yu, Rohan Anil, and Chelsea Finn. Efficiently identifying task groupings for multi-task learning. Advances in Neural Information Processing Systems, 34:27503-27516, 2021. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 647, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 296, + 713 + ], + "type": "text", + "content": "[11] Marco Fumero, Florian Wenzel, Luca Zancato, Alessandro Achille, Emanuele Rodolà, Stefano Soatto, Bernhard Schölkopf, and Francesco Locatello. Leveraging sparse and shared feature activations for disentangled representation learning. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 3, 4, 7" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 715 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[12] Jiechao Guan and Zhiwu Lu. Task relatedness-based generalization bounds for meta learning. In International Conference on Learning Representations, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 108, + 553, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 553, + 152 + ], + "type": "text", + "content": "[13] Yuzheng Hu, Ruicheng Xian, Qilong Wu, Qiuling Fan, Lang Yin, and Han Zhao. Revisiting scalarization in multi-task learning: A theoretical perspective. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 153, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 153, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 317, + 153, + 553, + 206 + ], + "type": "text", + "content": "[14] Ziniu Hu, Zhe Zhao, Xinyang Yi, Tiansheng Yao, Lichan Hong, Yizhou Sun, and Ed Chi. Improving multi-task generalization via regularizing spurious correlation. Advances in Neural Information Processing Systems, 35:11450-11466, 2022. 1, 2, 3, 4, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 209, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 553, + 252 + ], + "type": "text", + "content": "[15] Badr Youbi Idrissi, Martin Arjovsky, Mohammad Pezeshki, and David Lopez-Paz. Simple data balancing achieves competitive worst-group-accuracy. In Conference on Causal Learning and Reasoning, pages 336–351. PMLR, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 254, + 553, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 254, + 553, + 306 + ], + "spans": [ + { + "bbox": [ + 317, + 254, + 553, + 306 + ], + "type": "text", + "content": "[16] Junguang Jiang, Baixu Chen, Junwei Pan, Ximei Wang, Dapeng Liu, Jie Jiang, and Mingsheng Long. Forkmerge: Mitigating negative transfer in auxiliary-task learning. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 309, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 352 + ], + "type": "text", + "content": "[17] Polina Kirichenko, Pavel Izmailov, and Andrew Gordon Wilson. Last layer re-training is sufficient for robustness to spurious correlations. arXiv preprint arXiv:2204.02937, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 354, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 398 + ], + "type": "text", + "content": "[18] Tyler LaBonte, Vidya Muthukumar, and Abhishek Kumar. Towards last-layer retraining for group robustness with fewer annotations. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 399, + 553, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 399, + 553, + 465 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 553, + 465 + ], + "type": "text", + "content": "[19] Sébastien Lachapelle, Tristan Deleu, Divyat Mahajan, Ioannis Mitliagkas, Yoshua Bengio, Simon Lacoste-Julien, and Quentin Bertrand. Synergies between disentanglement and sparsity: Generalization and identifiability in multi-task learning. In International Conference on Machine Learning, pages 18171-18206. PMLR, 2023. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 467, + 553, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 467, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 316, + 467, + 553, + 499 + ], + "type": "text", + "content": "[20] Andrew K Lampinen and Surya Ganguli. An analytic theory of generalization dynamics and transfer learning in deep linear networks. arXiv preprint arXiv:1809.10374, 2018. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 501, + 553, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 501, + 553, + 542 + ], + "spans": [ + { + "bbox": [ + 316, + 501, + 553, + 542 + ], + "type": "text", + "content": "[21] Dongyue Li, Huy L Nguyen, and Hongyang R Zhang. Identification of negative transfers in multitask learning using surrogate models. arXiv preprint arXiv:2303.14582, 2023. 2, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 545, + 553, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 545, + 553, + 589 + ], + "spans": [ + { + "bbox": [ + 317, + 545, + 553, + 589 + ], + "type": "text", + "content": "[22] Weixin Liang and James Zou. Metashift: A dataset of datasets for evaluating contextual distribution shifts and training conflicts. arXiv preprint arXiv:2202.06523, 2022. 6, 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 591, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 591, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 591, + 553, + 624 + ], + "type": "text", + "content": "[23] Bo Liu, Yihao Feng, Peter Stone, and Qiang Liu. Famo: Fast adaptive multitask optimization. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 625, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 625, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 625, + 553, + 689 + ], + "type": "text", + "content": "[24] Evan Z Liu, Behzad Haghgoo, Annie S Chen, Aditi Raghunathan, Pang Wei Koh, Shiori Sagawa, Percy Liang, and Chelsea Finn. Just train twice: Improving group robustness without training group information. In International Conference on Machine Learning, pages 6781-6792. PMLR, 2021. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 691, + 555, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 555, + 715 + ], + "type": "text", + "content": "[25] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25706" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "type": "text", + "content": "International Conference on Computer Vision (ICCV), 2015. 6, 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 150 + ], + "type": "text", + "content": "[26] Shenyu Lu, Junyi Chai, and Xiaoqian Wang. Neural collapse inspired debiased representation learning for min-max fairness. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 2048-2059, 2024. 3, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "text", + "content": "[27] Taro Makino, Krzysztof Geras, and Kyunghyun Cho. Generative multitask learning mitigates target-causing confounding. Advances in Neural Information Processing Systems, 35:36546-36558, 2022. 1, 2, 3, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 294, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 241 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 241 + ], + "type": "text", + "content": "[28] Michinari Momma, Chaosheng Dong, and Jia Liu. A multi-objective/multi-task learning framework induced by pareto stationarity. In International Conference on Machine Learning, pages 15895-15907. PMLR, 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 242, + 294, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 242, + 294, + 296 + ], + "spans": [ + { + "bbox": [ + 56, + 242, + 294, + 296 + ], + "type": "text", + "content": "[29] Aviv Navon, Aviv Shamsian, Idan Achituve, Haggai Maron, Kenji Kawaguchi, Gal Chechik, and Ethan Fetaya. Multitask learning as a bargaining game. In International Conference on Machine Learning, pages 16428-16446. PMLR, 2022. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 297, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 294, + 319 + ], + "type": "text", + "content": "[30] Karl Pearson. Correlation coefficient. In *Royal Society Proceedings*, page 214, 1895. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 321, + 294, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 294, + 353 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 294, + 353 + ], + "type": "text", + "content": "[31] Elan Rosenfeld, Pradeep Ravikumar, and Andrej Risteski. The risks of invariant risk minimization. arXiv preprint arXiv:2010.05761, 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 355, + 294, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 355, + 294, + 408 + ], + "spans": [ + { + "bbox": [ + 56, + 355, + 294, + 408 + ], + "type": "text", + "content": "[32] Shiori Sagawa, Pang Wei Koh, Tatsunori B Hashimoto, and Percy Liang. Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization. arXiv preprint arXiv:1911.08731, 2019. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 411, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 464 + ], + "type": "text", + "content": "[33] Shiori Sagawa, Aditi Raghunathan, Pang Wei Koh, and Percy Liang. An investigation of why overparameterization exacerbates spurious correlations. In International Conference on Machine Learning, pages 8346-8356. PMLR, 2020. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 467, + 294, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 294, + 499 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 294, + 499 + ], + "type": "text", + "content": "[34] Ozan Sener and Vladlen Koltun. Multi-task learning as multi-objective optimization. Advances in neural information processing systems, 31, 2018. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 501, + 294, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 294, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 294, + 567 + ], + "type": "text", + "content": "[35] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part V 12, pages 746-760. Springer, 2012. 6, 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 568, + 294, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 294, + 621 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 294, + 621 + ], + "type": "text", + "content": "[36] Trevor Standley, Amir Zamir, Dawn Chen, Leonidas Guibas, Jitendra Malik, and Silvio Savarese. Which tasks should be learned together in multi-task learning? In International conference on machine learning, pages 9120-9132. PMLR, 2020. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 624, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 667 + ], + "type": "text", + "content": "[37] Charles Stein et al. Inadmissibility of the usual estimator for the mean of a multivariate normal distribution. In Proceedings of the Third Berkeley symposium on mathematical statistics and probability, pages 197-206, 1956. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 669, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 294, + 689 + ], + "type": "text", + "content": "[38] Shao-Hua Sun. Multi-digit mnist for few-shot learning, 2019. 6, 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[39] Nilesh Tripuraneni, Michael Jordan, and Chi Jin. On the theory of transfer learning: The importance of task diver" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 475 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "sity. Advances in neural information processing systems, 33: 7852-7862, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "text", + "content": "[40] Simon Vandenhende, Stamatios Georgoulis, Wouter Van Gansbeke, Marc Proesmans, Dengxin Dai, and Luc Van Gool. Multi-task learning for dense prediction tasks: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(7):3614-3633, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "text", + "content": "[41] Yipei Wang and Xiaoqian Wang. On the effect of key factors in spurious correlation: A theoretical perspective. In International Conference on Artificial Intelligence and Statistics, pages 3745-3753. PMLR, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 197, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 240 + ], + "type": "text", + "content": "[42] Zirui Wang, Yulia Tsvetkov, Orhan Firat, and Yuan Cao. Gradient vaccine: Investigating and improving multi-task optimization in massively multilingual models. arXiv preprint arXiv:2010.05874, 2020. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 241, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 285 + ], + "type": "text", + "content": "[43] Enneng Yang, Zhenyi Wang, Li Shen, Shiwei Liu, Guibing Guo, Xingwei Wang, and Dacheng Tao. Adamerging: Adaptive model merging for multi-task learning. arXiv preprint arXiv:2310.02575, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 286, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 330 + ], + "type": "text", + "content": "[44] Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient surgery for multi-task learning. Advances in Neural Information Processing Systems, 33:5824-5836, 2020. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 331, + 553, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 386 + ], + "type": "text", + "content": "[45] Amir R Zamir, Alexander Sax, William Shen, Leonidas J Guibas, Jitendra Malik, and Silvio Savarese. Taskonomy: Disentangling task transfer learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3712-3722, 2018. 6, 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 388, + 553, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 419 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 419 + ], + "type": "text", + "content": "[46] Yu Zhang and Qiang Yang. A survey on multi-task learning. IEEE transactions on knowledge and data engineering, 34 (12):5586-5609, 2021. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 421, + 553, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 421, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 316, + 421, + 553, + 475 + ], + "type": "text", + "content": "[47] Xiao Zhou, Yong Lin, Renjie Pi, Weizhong Zhang, Renzhe Xu, Peng Cui, and Tong Zhang. Model agnostic sample reweighting for out-of-distribution learning. In International Conference on Machine Learning, pages 27203-27221. PMLR, 2022. 3" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25707" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Identity-Clothing Similarity Modeling for Unsupervised Clothing Change Person Re-Identification/42258662-090f-4600-a2a7-5bf4fc16321d_content_list.json b/2025/Identity-Clothing Similarity Modeling for Unsupervised Clothing Change Person Re-Identification/42258662-090f-4600-a2a7-5bf4fc16321d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c047a7011b52312a3cc5008c9ec53c1d0d3e643f --- /dev/null +++ b/2025/Identity-Clothing Similarity Modeling for Unsupervised Clothing Change Person Re-Identification/42258662-090f-4600-a2a7-5bf4fc16321d_content_list.json @@ -0,0 +1,1785 @@ +[ + { + "type": "text", + "text": "Identity-Clothing Similarity Modeling for Unsupervised Clothing Change Person Re-Identification", + "text_level": 1, + "bbox": [ + 91, + 128, + 906, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhiqi Pang $^{1}$ Junjie Wang $^{2}$ Lingling Zhao $^{1}$ Chunyu Wang $^{1*}$", + "bbox": [ + 217, + 202, + 784, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Harbin Institute of Technology, China", + "bbox": [ + 343, + 222, + 653, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Nanjing Medical University, China", + "bbox": [ + 352, + 238, + 642, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zqpang98@gmail.com, junjiehit@163.com, zhaoll@hit.edu.cn, chunyu@hit.edu.cn", + "bbox": [ + 165, + 258, + 831, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 308, + 326, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Clothing change person re-identification (CC-ReID) aims to match different images of the same person, even when the clothing varies across images. To reduce manual labeling costs, existing unsupervised CC-ReID methods employ clustering algorithms to generate pseudo-labels. However, they often fail to assign the same pseudo-label to two images with the same identity but different clothing—referred to as a clothing change positive pair—thus hindering clothing-invariant feature learning. To address this issue, we propose the identity-clothing similarity modeling (ICSM) framework. To effectively connect clothing change positive pairs, ICSM first performs clothing-aware learning to leverage all discriminative information, including clothing, to obtain compact clusters. It then extracts cluster-level identity and clothing features and performs inter-cluster similarity estimation to identify clothing change positive clusters, reliable negative clusters, and hard negative clusters for each compact cluster. During optimization, we design an adaptive version of existing optimization methods to enhance similarities of clothing change positive pairs, while also introducing text semantics as a supervisory signal to further promote clothing invariance. Extensive experimental results across multiple datasets validate the effectiveness of the proposed framework, demonstrating its superiority over existing unsupervised methods and its competitiveness with some supervised approaches.", + "bbox": [ + 88, + 340, + 485, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 762, + 220, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Person re-identification (ReID) [32, 41, 43, 47, 48] aims to match images of the same individual captured in multicamera scenarios. Compared to traditional ReID, clothing change ReID (CC-ReID) [7, 23, 24, 26, 27] assumes that individuals may change clothing and focuses on matching images of the same person with different outfits. Although", + "bbox": [ + 89, + 786, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b7b099dab8d6d5ccc079d2912428f6cd91294679fa533a817e61a32e2d6f7185.jpg", + "image_caption": [ + "(a) Original images" + ], + "image_footnote": [], + "bbox": [ + 532, + 311, + 705, + 401 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f31bcb5e52e2ed94b5e7c21da8877061b20058b837b355e824e3113e0db2910f.jpg", + "image_caption": [ + "(b) Clothing-occluded images", + "Figure 1. Visualization of features extracted from original images and clothing-occluded images. Different colors represent features with different identities, while the same color but different shapes indicate features with the same identity but different clothing. Image pairs that are correctly (incorrectly) clustered by clustering algorithms are connected with green (red) arrows." + ], + "image_footnote": [], + "bbox": [ + 712, + 310, + 888, + 401 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "existing CC-ReID methods have achieved promising performance, they largely depend on large-scale manual annotations. However, manual labeling for retrieval tasks is an extremely labor-intensive process [1, 3]. Consequently, researchers have begun to focus on unsupervised CC-ReID (UCC-ReID) [20, 29]. In UCC-ReID, two images of the same person with different clothing—referred to as a clothing change positive pair—often exhibit significant appearance differences. This makes it challenging for clustering algorithms [4] to group their features into the same cluster, thus preventing them from obtaining the same pseudo-label. This phenomenon severely hinders the performance improvement of UCC-ReID methods. We refer to each clothing change positive pair that receives the same pseudo-label as a connected positive pair, while those that receive different pseudo-labels are termed unconnected positive pairs.", + "bbox": [ + 511, + 537, + 906, + 779 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To obtain robust pseudo-labels for clothing changes, existing UCC-ReID methods [20, 29] leverage body shapes extracted by human parsing models [21], which are typically clothing-invariant and identity-related. For example, CICL [29] clusters the fused features of original and body shape images. However, the unavoidable presence of clothing information in the fused features limits the number of connected positive pairs, while generating a large number", + "bbox": [ + 511, + 780, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author: Chunyu Wang", + "bbox": [ + 107, + 886, + 316, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19251", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "of unconnected positive pairs. Additionally, existing UCC-ReID methods [20, 29] typically enhance intra-class compactness and inter-class separability indiscriminately during the optimization phase, similar to unsupervised traditional ReID methods [38, 46]. However, such optimization methods reduce the similarity of each unconnected positive pair, i.e., they decrease the feature similarity between images of the same person with different clothing, thus hindering clothing-invariance learning.", + "bbox": [ + 89, + 90, + 480, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the above issues, we propose the identityclothing similarity modeling (ICSM) framework. An intuitive idea for obtaining robust pseudo-labels for clothing changes is to cluster the features of clothing-occluded images. However, we find this approach to be unfeasible. Specifically, clothing information is not entirely detrimental in the process of obtaining accurate pseudo-labels. As shown in Fig. 1a, the features of original images with the same identity and clothing—referred to as a same clothing positive pair—demonstrate high compactness. In contrast, as depicted in Fig. 1b, the clothing-occluded images lose some clothing-related discriminative information, resulting in reduced compactness for same clothing positive pairs. This significantly hinders clustering algorithms [4] from assigning the same pseudo-label to each same clothing positive pair. To this end, ICSM first performs clothing-aware learning to utilize all discriminative information, including clothing, to obtain compact clusters. It then extracts cluster-level identity and clothing features using visual-language models [22, 31, 49], and finally introduces an inter-cluster similarity estimation method to integrate identity and clothing similarities, identifying clothing change positive clusters, reliable negative clusters, and hard negative clusters for each compact cluster.", + "bbox": [ + 91, + 229, + 483, + 592 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "During the optimization phase, we enhance the model's robustness to clothing changes from both image semantics and text semantics. On one hand, we develop an adaptive version of existing optimization methods that focuses solely on reliable positive and negative pairs, thereby improving similarities of connected positive pairs while maintaining similarities of unconnected positive pairs. On the other hand, we introduce text semantics as a supervisory signal and design identity-oriented semantic contrastive loss and clothing-invariant semantic contrastive loss to further promote clothing invariance learning.", + "bbox": [ + 89, + 594, + 482, + 760 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contributions are summarized as follows:", + "bbox": [ + 109, + 762, + 450, + 776 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose an identity-clothing similarity modeling (ICSM) framework, which aims to connect clothing change positive pairs and promote clothing invariance learning to address clothing variation.", + "- We develop an inter-cluster similarity estimation method that not only fully utilizes clothing-related discriminative information but also integrates identity and clothing similarities to identify clothing change positive clusters, reli" + ], + "bbox": [ + 89, + 780, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "able negative clusters, and hard negative clusters for each compact cluster.", + "bbox": [ + 526, + 90, + 903, + 119 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We not only introduce an adaptive version of traditional optimization methods for UCC-ReID, but also design identity-oriented semantic contrastive loss and clothing-invariant semantic contrastive loss to leverage text semantics in promoting clothing invariance learning.", + "- Experimental results on multiple CC-ReID datasets validate the effectiveness of the proposed framework. ICSM not only outperforms existing UCC-ReID methods but also achieves competitive performance compared to earlier supervised methods." + ], + "bbox": [ + 511, + 121, + 903, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 287, + 653, + 304 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Unsupervised Traditional ReID", + "text_level": 1, + "bbox": [ + 513, + 314, + 790, + 330 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unsupervised traditional ReID (UT-ReID) methods [1, 3, 18, 45] assume that individuals do not change clothing, iterating between clustering and optimization. In early research, both BUC [25] and HCT [44] employ hierarchical clustering to obtain identity pseudo-labels, with HCT [44] also incorporating a batch-hard triplet loss [12] to address hard samples. Subsequent studies, such as CAP [34] and ICE [1], introduce camera invariance learning to mitigate the impact of varying camera conditions. In contrast to these approaches, IICS [36] and CIFL [28] focus on obtaining pseudo-labels robust to camera discrepancies. To alleviate the influence of noisy labels, both PPLR [2] and SE-CRET [10] assess the consistency between global and local features to enhance the overall reliability of training pseudolabels. Purification [18] supervises the student model using a teacher model to reduce the interference of noisy labels. Although existing UT-ReID methods face challenges in adapting to CC-ReID scenarios, their contributions to unsupervised learning provide valuable insights for our work.", + "bbox": [ + 511, + 335, + 906, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Clothing Change ReID", + "text_level": 1, + "bbox": [ + 513, + 636, + 725, + 652 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Clothing change ReID (CC-ReID) methods [7, 13, 16] typically assumes that individuals may change their clothing and emphasizes the importance of clothing-invariant and identity-related features. Existing methods often introduce additional identity information (e.g., keypoints [30], shape [13], and gait [16]) or clothing information (e.g., status [15] and category [7]) to aid the optimization process, achieving promising performance. However, most existing methods rely heavily on manual identity labels or even fine-grained clothing labels, which significantly limits flexibility and scalability. To address these issues, researchers explore several unsupervised CC-ReID methods. For instance, SiCL [20] and CICL [29] focus on utilizing body shape to obtain pseudo-labels that are robust to clothing variations. However, the presence of numerous unconnected positive pairs severely restricts their performance, a factor they overlook.", + "bbox": [ + 511, + 657, + 906, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19252", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this challenge, we integrates identity and clothing similarities to reduce the number of unconnected positive pairs. Additionally, we limit the optimization scope to reliable positive and negative pairs, thereby maintaining similarities of unconnected positive pairs.", + "bbox": [ + 89, + 90, + 483, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Framework", + "text_level": 1, + "bbox": [ + 89, + 180, + 297, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "During the training phase, we are provided with an unlabeled training image set $\\{x_{i}\\}_{i = 1}^{N}$ , where $N$ denotes the number of images. Images from the same individual may exhibit different clothing. As illustrated in Fig. 2, our proposed identity-clothing similarity modeling (ICSM) consists of three stages. Specifically, stage I (clothing-aware learning) iterates through the following two processes: 1) clustering image features to generate pseudo-labels, and 2) optimizing the image encoder $E_{\\mathrm{I}}$ based on the pseudolabels. The final pseudo-labels are saved, and the clusters corresponding to these pseudo-labels are referred to as compact clusters. Stage II (cluster-level feature learning) first extracts identity-related image set $\\{x_{i}^{s}\\}_{i = 1}^{N}$ and clothing-related image set $\\{x_{i}^{c}\\}_{i = 1}^{N}$ from $\\{x_{i}\\}_{i = 1}^{N}$ , and then introduces two CLIP models [31]: CLIP-S and CLIP-C. These models are used to learn cluster-level identity features and cluster-level clothing features, respectively, for each compact cluster. Stage III (clothing-invariant learning) first performs inter-cluster similarity estimation (ISE) to identify clothing change positive clusters, reliable negative clusters, and hard negative clusters for each compact cluster. Subsequently, we not only introduce an adaptive version of traditional optimization methods to help the image encoder $E_{\\mathrm{III}}$ learn features that are identity-relevant and clothing-invariant, but also design identity-oriented semantic contrastive loss and clothing-invariant semantic contrastive loss to further promote clothing invariance learning.", + "bbox": [ + 91, + 205, + 483, + 613 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "It is important to note that although ICSM involves a three-stage training process, the optimization in each stage is efficient and does not significantly increase the overall training time. Detailed analysis is provided in Sec. 4.3.", + "bbox": [ + 89, + 613, + 483, + 674 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Clothing-Aware Learning", + "text_level": 1, + "bbox": [ + 89, + 683, + 326, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this stage, we iteratively train the image encoder $E_{\\mathrm{I}}$ following the traditional unsupervised ReID approaches [1, 3]. Specifically, in each iteration, we first use $E_{\\mathrm{I}}$ to extract features from all training images. Then, we apply the DBSCAN algorithm [4] to all image features to obtain pseudolabels, and subsequently compute the centroid for each cluster. For example, the centroid of the $p$ -th cluster is defined as:", + "bbox": [ + 89, + 705, + 483, + 824 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nv _ {p} = \\frac {1}{N _ {p}} \\sum_ {i = 1} ^ {N _ {p}} f _ {i}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 228, + 823, + 482, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $N_{p}$ denotes the number of images in the $p$ -th cluster, and $f_{i}$ represents the features of the image $x_{i}$ within that", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "cluster. Then, we introduce the identity classification loss and hard contrastive loss to optimize $E_{\\mathrm{I}}$ . For any image $x_{i}$ , the identity classification loss is defined as:", + "bbox": [ + 511, + 90, + 905, + 137 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {i d} = - \\log \\frac {\\exp \\left(f _ {i} \\cdot v _ {p} ^ {\\mathrm {T}} / \\tau_ {i d}\\right)}{\\sum_ {q = 1} ^ {| V |} \\exp \\left(f _ {i} \\cdot v _ {q} ^ {\\mathrm {T}} / \\tau_ {i d}\\right)}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 146, + 906, + 186 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $f_{i}$ is the feature of image $x_{i}$ , $v_{p}$ is the centroid of $f_{i}$ , $|V|$ is the number of centroids in the current iteration, and $\\tau_{id}$ is the temperature hyperparameter. For any image $x_{i}$ , the hard contrastive loss is defined as:", + "bbox": [ + 511, + 196, + 906, + 256 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {h c} = - \\log \\frac {\\exp \\left(f _ {i} \\cdot f _ {p} ^ {\\mathrm {T}} / \\tau_ {h c}\\right)}{\\sum_ {f _ {q} \\in F _ {i} ^ {n} \\cup f _ {p}} \\exp \\left(f _ {i} \\cdot f _ {q} ^ {\\mathrm {T}} / \\tau_ {h c}\\right)}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 263, + 906, + 306 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $f_{i}^{p}$ is the feature of the most distant positive sample from $f_{i}$ in a minibatch of size $P\\times K$ (where $P$ is the number of sampled clusters and $K$ is the number of images sampled from each cluster), $F_{i}^{n}$ denotes the set of features corresponding to all negative samples of $f_{i}$ within the minibatch, and $\\tau_{hc}$ is the temperature hyperparameter. In summary, the overall loss function of stage I is defined as:", + "bbox": [ + 511, + 315, + 905, + 422 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {s 1} = L _ {i d} + L _ {h c}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 647, + 434, + 906, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The clustering and optimization process is iteratively performed, and the final pseudo-labels are saved. Due to the discriminative power brought by clothing, the clusters corresponding to these pseudo-labels exhibit high compactness, which we refer to as \"compact clusters\". In the following, the term \"cluster\" typically refers to these \"compact clusters.\" The images within a cluster usually share not only the same identity but also the same clothing.", + "bbox": [ + 511, + 460, + 906, + 583 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Cluster-Level Feature Learning", + "text_level": 1, + "bbox": [ + 511, + 590, + 794, + 607 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Due to the lack of connected positive pairs, it is challenging to improve the model's robustness to clothing changes based solely on compact clusters. Therefore, before training, we should attempt to connect two clusters that each contain images of the same person but with different clothing, i.e., a clothing change positive cluster pair. To this end, stage II focuses on obtaining cluster-level identity feature and cluster-level clothing feature for each compact cluster, thereby laying the foundation for connecting clothing change positive cluster pairs in the stage III.", + "bbox": [ + 511, + 613, + 905, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Stage II first obtains identity-related images and clothing-related images. As shown in Fig. 3, for any given original image $x_{i}$ , we first use a human parsing model [21] to obtain the parsed image $\\hat{x_i}$ . Based on $\\hat{x_i}$ , we then extract the identity-related image $x_{i}^{s}$ and the clothing-related image $x_{i}^{c}$ from the original image $x_{i}$ . Specifically, in $x_{i}^{s}$ , the pixel values of the clothing region are set to 0, while the pixel values of the other body regions are kept the same as in $x_{i}$ and the background pixel values are set to 255. In $x_{i}^{c}$ , the", + "bbox": [ + 511, + 765, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19253", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/33ecfe168bbae0008ddc730268e2eb08017100b3bdc0c9c65847a34b12b0f790.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 132, + 85, + 624, + 191 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/904cb22b1a57f2a82b7a89e3effedf03524df5b9170f02563c73bce00c53bdf7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 132, + 193, + 624, + 330 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c26c17bd2533c59a7acbc07e77dbf9a3b622df04e846eb37c36c73208b6a5f09.jpg", + "image_caption": [ + "Figure 2. The proposed identity-clothing similarity modeling (ICSM) framework for UCC-ReID as a three-stage pipeline. The different colors of the clusters represent different pseudo-labels. The image encoders $E_{\\mathrm{I}}$ and $E_{\\mathrm{III}}$ are optimized, while all image and text encoders in the stage II are frozen. See description in the text for additional detail." + ], + "image_footnote": [], + "bbox": [ + 133, + 333, + 624, + 386 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/376754699728b591620cd4deb53ebb733908dfbf42154a53564d402b87563afe.jpg", + "image_caption": [ + "Stage II (cluster-level feature learning)" + ], + "image_footnote": [], + "bbox": [ + 635, + 107, + 862, + 383 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/58b41685c572d3ba54d38c95d29a78245a2b5c0931382d4ea83c40590f6b80a7.jpg", + "image_caption": [ + "Figure 3. Example of obtaining identity-related image $x_{i}^{s}$ and clothing-related image $x_{i}^{c}$ ." + ], + "image_footnote": [], + "bbox": [ + 133, + 464, + 442, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "pixel values of the clothing region remain the same as in $x_{i}$ , while all other regions are set to 255. We refer to the clusters composed of identity-related images as identity-related clusters, and the clusters composed of clothing-related images as clothing-related clusters.", + "bbox": [ + 89, + 595, + 482, + 670 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Subsequently, stage II leverages the strong summarizing and descriptive ability of text for person images [5, 6, 37], along with the powerful text-image understanding capabilities of vision-language models [22, 31, 49], to obtain cluster-level features in text form. Specifically, we introduce two CLIP models [31]: CLIP-S and CLIP-C, which share the same pre-trained parameters. CLIP-S is used to learn cluster-level identity features from identity-related clusters, while CLIP-C is used to learn cluster-level clothing features from clothing-related clusters.", + "bbox": [ + 89, + 672, + 482, + 823 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For CLIP-S, we first set an identity-related text for each identity-related cluster as \"A photo of a $[S_1][S_2]\\ldots [S_M]$ person,\" where $\\left[[S_1][S_2]\\ldots [S_M]\\right]$ represents a set of learnable text embeddings, and $M$ is the number of learnable text embeddings. For any identity-related image $x_{i}^{s}$ , we input it", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "along with the corresponding identity-related text into the image encoder and text encoder of CLIP-S, respectively. This produces the image feature $f_{i}^{s}$ and text feature $t_{i}^{s}$ . Afterward, we freeze the image and text encoders of CLIP-S and optimize the learnable text embeddings, as follows:", + "bbox": [ + 511, + 468, + 906, + 544 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL ^ {s} = L _ {i t} ^ {s} + L _ {t i} ^ {s}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 653, + 554, + 903, + 570 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $L_{it}^{s}$ and $L_{ti}^{s}$ are defined as:", + "bbox": [ + 511, + 580, + 736, + 597 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {i t} ^ {s} = - \\log \\frac {\\exp \\left(f _ {i} ^ {s} \\cdot t _ {i} ^ {s ^ {\\mathrm {T}}}\\right)}{\\sum_ {j = 1} ^ {B} \\exp \\left(f _ {i} ^ {s} \\cdot t _ {j} ^ {s ^ {\\mathrm {T}}}\\right)}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 604, + 903, + 646 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {t i} ^ {s} = \\frac {- 1}{| \\psi_ {i} ^ {s} |} \\sum_ {f _ {p} ^ {s} \\in \\psi_ {i} ^ {s}} \\log \\frac {\\exp \\left(t _ {i} ^ {s} \\cdot f _ {p} ^ {s ^ {\\mathrm {T}}}\\right)}{\\sum_ {j = 1} ^ {B} \\exp \\left(t _ {i} ^ {s} \\cdot f _ {j} ^ {s ^ {\\mathrm {T}}}\\right)}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 565, + 655, + 903, + 699 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\psi_i^s$ represents the set of image features that share the same pseudo-label as $t_i^s$ , and $B$ is the batch size.", + "bbox": [ + 511, + 704, + 903, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, we use the optimized text feature $t_i^s$ as the cluster-level identity feature for the corresponding compact cluster $u_i$ . Similarly, for CLIP-C, we first set a clothing-related text for each clothing-related cluster: \"A photo of $[C_1][C_2]\\ldots [C_M]$ clothing\", where $\"[C_1][C_2]\\ldots [C_M]\"$ represents a set of learnable text embeddings. Then, we follow the same optimization scheme as for $L_{it}^s$ and $L_{ti}^s$ to set $L_{it}^c$ and $L_{ti}^c$ for CLIP-C. Based on CLIP-C and the clothing-related text, we obtain the cluster-level clothing feature for each compact cluster. For example, the cluster-level clothing feature for the $i$ -th cluster $u_i$ is represented as $t_i^c$ .", + "bbox": [ + 511, + 734, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19254", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Clothing-Invariant Learning", + "text_level": 1, + "bbox": [ + 89, + 90, + 346, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inter-cluster similarity estimation. In theory, the higher the similarity between the cluster-level identity features of two compact clusters, the more likely it is that they contain images of the same person. Conversely, the lower the similarity between the cluster-level clothing features of two clusters, the more likely it is that they contain images with different clothing. Therefore, in stage III, we propose an inter-cluster similarity estimation (ISE) method, which seeks clothing change positive clusters, reliable negative clusters, and hard negative clusters for each cluster.", + "bbox": [ + 89, + 113, + 483, + 263 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For any cluster $u_{i}$ , ISE first calculates similarities between its cluster-level identity feature and clothing feature with those of other clusters, and then ranks all other clusters in descending order of similarity, resulting in the corresponding identity similarity sequence and clothing similarity sequence for $u_{i}$ . For any cluster $u_{i}$ , the set consisting of all clothing change positive clusters is defined as:", + "bbox": [ + 89, + 266, + 483, + 371 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nU _ {i} ^ {c} = \\left\\{u _ {j} \\mid t _ {i} ^ {s} \\cdot t _ {j} ^ {s ^ {\\mathrm {T}}} > \\eta_ {a} \\text {a n d} t _ {i} ^ {c} \\cdot t _ {j} ^ {c ^ {\\mathrm {T}}} \\leq \\eta_ {c} \\right\\}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 383, + 482, + 405 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\eta_{a}$ denotes the similarity value ranked at $\\gamma_{a}$ in the identity similarity sequence, and $\\eta_{c}$ denotes the similarity value ranked at $\\gamma_{c}$ in the clothing similarity sequence. Additionally, for any cluster $u_{i}$ , we define the positive cluster set as $U_{i}^{a} = U_{i}^{c} \\cup u_{i}$ . For any cluster $u_{i}$ , the set of its reliable negative clusters is defined as:", + "bbox": [ + 89, + 416, + 483, + 506 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nU _ {i} ^ {b} = \\left\\{u _ {j} \\mid t _ {i} ^ {s} \\cdot t _ {j} ^ {s T} \\leq \\eta_ {b} \\right\\}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 518, + 482, + 541 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\eta_{b}$ denotes the similarity value ranked at $\\gamma_{b}$ in the identity similarity sequence. For any cluster $u_{i}$ , we define the clusters with low cluster-level identity feature similarity and high cluster-level clothing feature similarity as hard negative clusters. Therefore, for any cluster $u_{i}$ , the set of its hard negative clusters is defined as:", + "bbox": [ + 89, + 551, + 483, + 643 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nU _ {i} ^ {h} = \\left\\{u _ {j} \\mid t _ {i} ^ {s} \\cdot t _ {j} ^ {s T} \\leq \\eta_ {b} \\text {a n d} t _ {i} ^ {c} \\cdot t _ {j} ^ {c T} > \\eta_ {c} \\right\\}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 120, + 655, + 482, + 676 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, for any cluster $u_{i}$ , we save the reliable positive cluster set $U_{i}^{a}$ , clothing change positive cluster set $U_{i}^{c}$ , reliable negative cluster set $U_{i}^{b}$ , and hard negative cluster set $U_{i}^{h}$ .", + "bbox": [ + 89, + 688, + 482, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Image semantic contrastive learning. Due to the difficulty in fully connecting all clothing change positive cluster pairs, there may be additional clothing change positive clusters beyond the clothing change positive cluster set $U_{i}^{c}$ for any cluster $u_{i}$ . In this scenario, traditional optimization methods [1, 3] aimed at increasing inter-class separability would reduce the similarities between unconnected positive pairs, thereby hindering clothing-invariant learning. To address this, we design an adaptive version of the traditional optimization method for the UCC-ReID scenario. Specifically, for any image feature $f_{i}$ , the adaptive identity classification", + "bbox": [ + 89, + 734, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "loss is defined as:", + "bbox": [ + 513, + 90, + 633, + 104 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {a i d} = \\frac {- 1}{\\left| V _ {i} ^ {a} \\right|} \\sum_ {v _ {p} \\in V _ {i} ^ {a}} \\log \\frac {\\exp \\left(f _ {i} \\cdot v _ {p} ^ {\\mathrm {T}} / \\tau_ {i d}\\right)}{\\sum_ {v _ {q} \\in V _ {i} ^ {b} \\cup v _ {p}} \\exp \\left(f _ {i} \\cdot v _ {q} ^ {\\mathrm {T}} / \\tau_ {i d}\\right)}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 109, + 903, + 162 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $V_{i}^{a}$ and $V_{i}^{b}$ respectively represent the set of centroids of all clusters in $U_{i}^{a}$ and $U_{i}^{b}$ corresponding to $f_{i}$ . For any image feature $f_{i}$ , the adaptive hard contrastive loss is defined as:", + "bbox": [ + 511, + 164, + 906, + 223 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {a h c} = \\frac {- 1}{\\left| V _ {i} ^ {c} \\right|} \\sum_ {v _ {p} \\in V _ {i} ^ {c}} \\log \\frac {\\exp \\left(f _ {i} \\cdot v _ {p} ^ {\\mathrm {T}} / \\tau_ {h c}\\right)}{\\sum_ {v _ {q} \\in V _ {i} ^ {h} \\cup v _ {p}} \\exp \\left(f _ {i} \\cdot v _ {q} ^ {\\mathrm {T}} / \\tau_ {h c}\\right)}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 228, + 903, + 281 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $V_{i}^{c}$ and $V_{i}^{h}$ respectively represent the set of centroids of all clusters in $U_{i}^{c}$ and $U_{i}^{h}$ corresponding to $f_{i}$ . Since the above process is optimized only for reliable clusters, it maintains the similarities of unconnected positive pairs.", + "bbox": [ + 511, + 282, + 906, + 343 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Text semantic contrastive learning. To fully leverage identity-related and clothing-related text semantics to assist model optimization, we introduce the identity-oriented semantic contrastive loss and clothing-invariant semantic contrastive loss. Firstly, for each image $x_{i}$ , we compute the similarity distribution $\\alpha_{i}$ between the cluster-level identity feature $t_{i}^{s}$ of $x_{i}$ and all other cluster-level identity features. The $k$ -th element of $\\alpha_{i}$ , denoted as $\\alpha_{i}^{(k)}$ , is defined as:", + "bbox": [ + 511, + 343, + 905, + 467 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {i} ^ {(k)} = \\frac {\\exp \\left(t _ {i} ^ {s} \\cdot t _ {k} ^ {s T} / \\tau_ {i s}\\right)}{\\sum t _ {j} ^ {s} \\in T _ {i} ^ {s}} \\exp \\left(t _ {i} ^ {s} \\cdot t _ {j} ^ {s T} / \\tau_ {i s}\\right), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 474, + 903, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $T_{i}^{s}$ represents the set of all cluster-level identity features except for $t_{i}^{s}$ , and $\\tau_{is}$ is the temperature hyperparameter. Next, we compute the similarity distribution $\\beta_{i}$ between the image feature $f_{i}$ of $x_{i}$ and the centroids of all other clusters. The $k$ -th element of $\\beta_{i}$ , denoted as $\\beta_{i}^{(k)}$ , is defined as:", + "bbox": [ + 511, + 520, + 905, + 599 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\beta_ {i} ^ {(k)} = \\frac {\\exp \\left(f _ {i} \\cdot v _ {k} ^ {\\mathrm {T}} / \\tau_ {i s}\\right)}{\\sum_ {v _ {j} \\in V _ {i}} \\exp \\left(f _ {i} \\cdot v _ {j} ^ {\\mathrm {T}} / \\tau_ {i s}\\right)}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 608, + 903, + 646 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $V_{i}$ represents the set of all centroids except for the centroid of $x_{i}$ . In theory, if all image features exhibit high robustness to clothing changes, then $\\beta_{i}$ should exhibit high similarity to $\\alpha_{i}$ . Therefore, to leverage identity-related text semantics to ensure high robustness of image features to clothing changes, we introduce the identity-oriented semantic contrastive loss:", + "bbox": [ + 511, + 648, + 905, + 753 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {i s} = \\sum_ {k = 1} ^ {| V | - 1} \\alpha_ {i} ^ {(k)} \\log \\left(\\alpha_ {i} ^ {(k)} / \\beta_ {i} ^ {(k)}\\right), \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 758, + 903, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $|V|$ is the number of centroids in the current iteration. Additionally, to further promote clothing-invariant learning by leveraging clothing-related text semantics, we introduce the clothing-invariant semantic contrastive loss:", + "bbox": [ + 511, + 809, + 905, + 869 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {c s} = \\left[ f _ {i} \\cdot t _ {p} ^ {c \\mathrm {T}} - \\min _ {t _ {q} ^ {c} \\in T _ {i} ^ {a}} \\left(f _ {i} \\cdot t _ {q} ^ {c \\mathrm {T}}\\right) \\right] _ {+}, \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 875, + 903, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19255", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $t_p^c$ represents the cluster-level clothing feature of image feature $f_{i}$ , and $T_{i}^{a}$ is the set of cluster-level clothing features of all clusters in $U_{i}^{a}$ corresponding to $f_{i}$ .", + "bbox": [ + 89, + 90, + 482, + 136 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In summary, the overall loss function of stage III is defined as:", + "bbox": [ + 89, + 138, + 482, + 167 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {s 3} = L _ {a i d} + L _ {a h c} + \\lambda_ {i s} L _ {i s} + \\lambda_ {c s} L _ {c s}, \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 185, + 482, + 202 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\lambda_{is}$ and $\\lambda_{cs}$ are the weighting coefficients for $L_{is}$ and $L_{cs}$ , respectively.", + "bbox": [ + 89, + 215, + 482, + 247 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 89, + 266, + 215, + 282 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets and Evaluation Metrics", + "text_level": 1, + "bbox": [ + 89, + 292, + 377, + 308 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate the proposed framework on the PRCC [39], VC-Clothes [33], LTCC [30], Celeb-reID [14], and Celeb-reID-light [14] datasets, with statistical results presented in Tab. 1. Following existing works [20, 29], we employ three testing schemes: (1) clothing change (CC): the gallery contains only images with clothing changes; (2) same clothing (SC): the gallery includes only images with the same clothing; (3) general (G): the gallery consists of images with both clothing changes and the same clothing. For the PRCC [39] dataset, we utilize the CC and SC testing schemes; for VC-Clothes [33] and LTCC [30], we use the CC and G schemes; and for Celeb-reID [14] and Celeb-reID-light [14], we implement the CC testing scheme. For each testing scheme, we assess recognition performance using mean average precision (mAP) and cumulative matching characteristic (CMC).", + "bbox": [ + 89, + 316, + 483, + 559 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 574, + 307, + 590 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "All images are resized to $256\\times 128$ pixels. During the training phase, random horizontal flipping and random cropping are employed for data augmentation. All image encoders utilize ResNet-50 [9] pre-trained on CLIP [31] as the backbone. For DBSCAN [4], the distance threshold and minimum sample size are set to 0.55 and 4, respectively. The pseudo-labels from stage I are saved and passed on to stage II and stage III, where no further clustering or pseudo-label updates are performed. For each set of learnable text embeddings, we set $M = 4$ . For ISE, we set $\\gamma_{a} = 2\\%$ , $\\gamma_{c} = 50\\%$ , and $\\gamma_{b} = 60\\%$ . For stage I and stage III, we configure each minibatch to contain $P = 16$ identities, with each identity comprising $K = 4$ images. For stage II, we set $B = 64$ . For temperature hyperparameters, we set $\\tau_{id} = 0.5$ , $\\tau_{hc} = 0.1$ , and $\\tau_{is} = 0.5$ . The model is optimized using the Adam optimizer [17], with an initial learning rate set to 0.00035. Both stage I and stage III consist of 50 epochs. In stage II, the learnable text embeddings in CLIP-S and CLIP-C are optimized for 120 epochs, respectively. During the testing phase, only the optimized image encoder $E_{\\mathrm{III}}$ is used for inference.", + "bbox": [ + 89, + 598, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0bef2302128dc802ff052dd8e1f46029228523d2633c36a34b4f5292d40e8673.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetsIdentitiesTrainingQueryGallery
PRCC22117,8963,54312,259
VC-Clothes2569,4491,0208,591
LTCC1529,5764937,050
Celeb-reID105220,2082,97211,006
Celeb-reID-light5909,021887934
", + "bbox": [ + 529, + 88, + 890, + 167 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Statistical results of PRCC, VC-Clothes, LTCC, CelebreID, and CelebreID-light.", + "bbox": [ + 511, + 176, + 903, + 205 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Comparison with State-of-the-Arts", + "text_level": 1, + "bbox": [ + 511, + 234, + 816, + 251 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate the proposed method on PRCC [39], VC-Clothes (VC) [33], LTCC [30], Celeb-reID (CR) [14], and Celeb-reID-light (CRL) [14]. As shown in Tab. 2, we first report the performance of seven existing UT-ReID methods, including ICE [1], CC [3], CACL [19], PPLR [2], RTMem [42], Purification [18], and DCCC [11]. ICE [1], CACL [19], and Purification [18] achieve superior performance in these testing scenarios. We find that ICSM significantly outperforms all of the aforementioned methods, particularly in clothing change scenarios. This validates the effectiveness of ICSM's unique design for addressing clothing change scenarios.", + "bbox": [ + 511, + 258, + 906, + 439 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Subsequently, we compare advanced UCC-ReID methods: SiCL [20] and CICL [29]. We observe that ICSM's overall performance surpasses these methods, particularly achieving significant advantages in PRCC(CC) and VC(CC). This can be attributed to two main factors: first, ICSM combines identity and clothing similarities to identify clothing change positive clusters, reliable negative clusters, and hard negative clusters for each compact cluster. Second, ICSM not only incorporates an adaptive version of traditional optimization methods to maintain similarities of unconnected positive pairs but also leverages text semantics to further enhance clothing invariance learning.", + "bbox": [ + 511, + 443, + 908, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finally, we compare ICSM with existing supervised CC-ReID (SCC-ReID) methods, including RCSANet [15], GI-ReID [16], CAL [7], AIM [40], CCFA [8], and $\\mathrm{FIRe}^2$ [35]. We observe that ICSM achieves competitive performance on PRCC(CC) compared to earlier SCC-ReID methods (RCSANet [15] and CAL [7]), and surpasses GI-ReID [16] on VC(CC) and LTCC(CC). Moreover, ICSM achieves an mAP of approximately $70\\%$ compared to advanced methods (AIM [40], CCFA [8], and $\\mathrm{FIRe}^2$ [35]) under the two testing settings of LTCC. These experimental results validate the superiority of ICSM and the potential of UCC-ReID.", + "bbox": [ + 511, + 626, + 908, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Additionally, we report the training and test times for SiCL [20], CICL [29], and ICSM on VC(CC). As shown in Tab. 3, the training time for ICSM is slightly higher than that of SiCL and significantly lower than that of CICL. This is because, despite ICSM's three-stage training, the optimization in each stage is efficient and does not significantly", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19256", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/bef7b65818872d4467c1340e0d95f565c263dad6fb2189fcaa40a63ff4b6c114.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsReferencePRCC(CC)PRCC(SC)VC(CC)VC(G)LTCC(CC)LTCC(G)CRCRL
mAPR-1mAPR-1mAPR-1mAPR-1mAPR-1mAPR-1mAPR-1mAPR-1
UT-ReID
ICEICCV2145.532.695.793.328.531.451.870.110.116.322.644.04.940.75.07.1
CCACCV2246.334.494.490.225.731.045.162.46.07.411.017.03.432.83.24.4
CACLTIP2252.141.794.790.949.758.968.082.46.29.822.345.65.142.33.64.3
PPLRCVPR2251.440.091.787.423.132.547.768.14.44.86.011.24.841.34.36.2
RTMemTIP2347.235.893.692.737.546.159.775.35.87.518.743.15.039.84.96.8
PurificationTIP2352.341.995.293.047.656.866.881.68.915.623.347.65.242.64.97.0
DCCCICASSP2447.035.390.689.539.350.458.976.17.812.719.644.93.935.73.35.0
UCC-ReID
SiCLarXiv2455.443.296.293.863.971.777.287.910.120.727.657.65.645.45.610.5
CICLNN2455.345.096.795.369.372.876.787.313.025.729.660.36.749.37.812.5
ICSMOurs57.847.997.597.173.575.178.789.613.526.230.160.87.049.67.712.6
SCC-ReID
RCSANetICCV2148.650.297.2100.0--------11.955.616.729.5
GI-ReIDCVPR22-33.3--57.864.5--10.423.729.463.2----
CALCVPR2255.855.299.8100.081.781.487.292.9--------
AIMCVPR2358.357.999.9100.0----19.140.641.176.3----
CCFACVPR2358.461.298.799.6----22.145.342.575.8----
FIRe2TIFS2463.165.099.5100.0----19.144.639.975.918.264.0--
", + "bbox": [ + 109, + 87, + 890, + 354 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3544167821ebb9f64c47df66c2b7e5ff4988f514bc2f19f322c449a102f5203a.jpg", + "table_caption": [ + "Table 2. Comparison of ICSM with existing methods. Optimal performance under two unsupervised settings is highlighted in bold. \"R-1\" denotes \"Rank-1\"." + ], + "table_footnote": [], + "table_body": "
MethodsmAPRank-1Training timeTest time
SiCL63.971.76798.2514.20
CICL69.372.810569.9314.16
ICSM73.575.17236.5113.97
", + "bbox": [ + 125, + 414, + 450, + 469 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Statistics on performance and efficiency (seconds).", + "bbox": [ + 106, + 479, + 460, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "increase the overall training time. In contrast, while existing UCC-ReID methods conduct only one stage of training, they often need to optimize three neural network modules simultaneously, which can require even more training time than ICSM. Additionally, ICSM has a similar testing time compared to existing UCC-ReID methods. These results demonstrate that ICSM not only achieves superior performance but also maintains efficiency.", + "bbox": [ + 89, + 521, + 483, + 642 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 651, + 243, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, the baseline we tested optimizes the image encoder using only $L_{id}$ and $L_{hc}$ from stage I. As shown in Tab. 4, M1 corresponds to the baseline, while M2 introduces ISE on top of M1. Additionally, to accommodate ISE, M2 replaces $L_{id}$ and $L_{hc}$ with $L_{aid}$ and $L_{ahc}$ , respectively. M3 and M4 introduce $L_{is}$ and $L_{cs}$ , respectively, building upon M2. M5, which represents ICSM, incorporates both $L_{is}$ and $L_{cs}$ on top of M2.", + "bbox": [ + 89, + 674, + 482, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of ISE. M2 introduces ISE on top of M1 to identify clothing change positive clusters, reliable negative clusters, and hard negative clusters for each compact cluster. As shown in Tab. 4, M2 achieves significant performance improvements over M1 in PRCC(CC), VC-Clothes(CC), and VC-Clothes(G). To further explore the nature of ISE, we visualize the features of several clusters. As shown", + "bbox": [ + 89, + 795, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9e6992493cfb051bf3fa3f9ba6cfabe5ef0eb7849803e1d43faa5d964fd0192e.jpg", + "image_caption": [ + "(a) M1" + ], + "image_footnote": [], + "bbox": [ + 521, + 417, + 705, + 531 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d4e3fff688845ba21e030142c68905ad5e153d93c28056e13c5ab214fecf8a6c.jpg", + "image_caption": [ + "(b) M2" + ], + "image_footnote": [], + "bbox": [ + 709, + 417, + 895, + 530 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4881f7b6374435d20b7cbc8a60ee3df331963f681f14c3a97083d1157924b2db.jpg", + "image_caption": [ + "Figure 4. Visualization of the clusters corresponding to M1 and M2. Clusters are labeled with black dashed lines. For any given cluster, the tail and head of the red arrow connect it to its clothing change positive cluster identified by ISE.", + "Figure 5. Feature maps of M1 and M2. Each triplet, from left to right, consists of the original image, the feature map of M1, and the feature map of M2." + ], + "image_footnote": [], + "bbox": [ + 524, + 627, + 895, + 684 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "in Fig. 4, M2 is able to identify clothing change positive clusters based on M1, significantly reducing the distances between the positive cluster pairs. For the unconnected clothing change positive clusters, the distances are not significantly increased because $L_{aid}$ and $L_{ahc}$ typically prevent these clusters from being included in the optimization scope. This process undoubtedly enhances the model's robustness to clothing changes. As shown in Fig. 5, we observe that M2 indeed focuses more on clothing-invariant re", + "bbox": [ + 511, + 763, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19257", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/fd9e942d8c4a785a8f969cb37ed4f81379a1f5113d933af7af4cfe829d5f6472.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsBaselineISE\\(L_{is}\\)\\(L_{cs}\\)PRCC(CC)PRCC(SC)VC-Clothes(CC)VC-Clothes(G)
mAPRank-1mAPRank-1mAPRank-1mAPRank-1
M1✓×××43.931.592.289.023.630.846.667.8
M2✓✓××53.643.095.192.857.664.969.983.5
M3✓✓✓×56.847.197.095.668.672.575.286.3
M4✓✓×✓55.344.695.393.263.869.272.785.0
M5/ICSM✓✓✓✓57.847.997.597.173.575.178.789.6
", + "bbox": [ + 151, + 88, + 848, + 179 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Results of the ablation study for the proposed method.", + "bbox": [ + 307, + 189, + 683, + 204 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/43ff2dac8fdf843e1bc425dfbc051ccbe48b109e78f8ce5f334f1cde2ff467f2.jpg", + "image_caption": [ + "(a) M2" + ], + "image_footnote": [], + "bbox": [ + 109, + 227, + 279, + 306 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6f6af1072a563bb0999401f82de4fcf3dacf5e45f59fd42c935009a37e9b372f.jpg", + "image_caption": [ + "(b) M3" + ], + "image_footnote": [], + "bbox": [ + 292, + 228, + 464, + 306 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ad57bd73604052a7d00d7a56eaf6799e0f7b57096f0f6704e27c85aabe999d4c.jpg", + "image_caption": [ + "(c) M4" + ], + "image_footnote": [], + "bbox": [ + 109, + 321, + 279, + 401 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/980d21180ca32db2b08bda720a8d72d048ba1ee391d1e3b1389ff64d45b95987.jpg", + "image_caption": [ + "(d) M5" + ], + "image_footnote": [], + "bbox": [ + 292, + 321, + 464, + 401 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "gions such as the head and arms, or areas like shoes that are less frequently changed, compared to M1. The above results validate the effectiveness of ISE.", + "bbox": [ + 89, + 484, + 483, + 529 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of $L_{is}$ and $L_{cs}$ . As shown in Tab. 4, M3 and M4 introduce $L_{is}$ and $L_{cs}$ into M2, respectively, both achieving significant performance improvements. Additionally, we visualize the feature distance distributions of same clothing positive pairs and clothing change positive pairs for M2, M3, M4, and M5. As shown in Fig. 6, compared to M2, both M3 and M4 significantly reduce the feature distance of clothing change positive pairs while also increasing the overlap between the feature distance distributions of these two types. This validates that both $L_{is}$ and $L_{cs}$ can enhance model performance by promoting the learning of clothing-invariant features. Furthermore, we observe that M5 further reduces the feature distance of clothing change positive pairs compared to M3 and M4, achieving superior performance over all previously mentioned methods.", + "bbox": [ + 89, + 531, + 483, + 772 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Sensitivity Analysis", + "text_level": 1, + "bbox": [ + 89, + 786, + 276, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Fig. 7, when $\\lambda_{is} = 2$ and $\\lambda_{cs} = 1$ , the model achieves optimal performance on both PRCC(CC) and VC-Clothes(CC). This confirms the generalizability of $\\lambda_{is}$ and $\\lambda_{cs}$ . We also observe that when $\\lambda_{is}$ is set too high or too low, model performance deteriorates. This is mainly because a small $\\lambda_{is}$ reduces the contribution of identity-", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/951e7a7b2133b430221dbc9b99d72572eabbf49db71b58dea56907ec9d0e3beb.jpg", + "image_caption": [ + "(a) PRCC(CC)" + ], + "image_footnote": [], + "bbox": [ + 532, + 227, + 702, + 308 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6cd129c725b65f8f567db020e8c0ba5e9ae107ff377bb42eee141c6c638fac32.jpg", + "image_caption": [ + "(b) VC-Clothes(CC)" + ], + "image_footnote": [], + "bbox": [ + 715, + 227, + 885, + 308 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/31b39396ff259a14bd491747ad000c2d21c829585af3c708aedba6f2042a418c.jpg", + "image_caption": [ + "Figure 6. Feature distance distributions of same clothing positive pairs and clothing change positive pairs for M2, M3, M4, and M5.", + "(a) PRCC(CC)", + "Figure 8. Impact of $(\\gamma_{a},\\gamma_{c})$ on performance." + ], + "image_footnote": [], + "bbox": [ + 532, + 364, + 702, + 452 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0496e457536f5e0396c8d355a28bda3261b1497c8d5c137246591961c255899e.jpg", + "image_caption": [ + "Figure 7. Impact of $(\\lambda_{is},\\lambda_{cs})$ on performance.", + "(b) VC-Clothes(CC)" + ], + "image_footnote": [], + "bbox": [ + 715, + 364, + 887, + 452 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "related text semantics, while a large $\\lambda_{is}$ negatively impacts the effectiveness of $L_{aid}$ and $L_{ahc}$ . As shown in Fig. 8, when $\\gamma_{a} = 2\\%$ and $\\gamma_{c} = 50\\%$ , the model achieves optimal performance on both datasets. This not only validates the generalization capability of ISE but also verifies the model's robustness to $\\gamma_{a}$ and $\\gamma_{c}$ .", + "bbox": [ + 511, + 523, + 906, + 614 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 632, + 633, + 647 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose an identity-clothing similarity modeling (ICSM) framework for UCC-ReID. To facilitate clothing-invariant feature learning, we introduce intercluster similarity estimation (ISE), identity-oriented semantic contrastive loss, and clothing-invariant semantic contrastive loss. The results of the ablation study validate their effectiveness, demonstrating that ISE enhances the model's robustness to clothing changes by identifying reliable clothing change positive clusters, while identity-oriented semantic contrastive loss and clothing-invariant semantic contrastive loss further promote clothing invariance learning. Experimental results across eight testing scenarios validate the effectiveness and superiority of the proposed ICSM, which not only outperforms existing unsupervised methods but also achieves competitive performance compared to certain supervised methods.", + "bbox": [ + 511, + 659, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19258", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 240, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is supported by National Key Research and Development Program of China (Grant no. 2023YFC3305005).", + "bbox": [ + 89, + 114, + 480, + 160 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 176, + 186, + 191 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Hao Chen, Benoit Lagadec, and Francois Bremond. ICE: Inter-instance contrastive encoding for unsupervised person re-identification. In ICCV, pages 14960-14969, 2021. 1, 2, 3, 5, 6", + "[2] Yoonki Cho, Woo Jae Kim, Seunghoon Hong, and Sung-Eui Yoon. Part-based pseudo label refinement for unsupervised person re-identification. In CVPR, pages 7308-7318, 2022. 2, 6", + "[3] Zuozhuo Dai, Guangyuan Wang, Weihao Yuan, Siyu Zhu, and Ping Tan. Cluster contrast for unsupervised person re-identification. In ACCV, pages 1142–1160, 2022. 1, 2, 3, 5, 6", + "[4] Martin Ester, Hans-Peter Kriegel, Jörg Sander, Xiaowei Xu, et al. A density-based algorithm for discovering clusters in large spatial databases with noise. In KDD, pages 226-231, 1996. 1, 2, 3, 6", + "[5] Tiantian Gong, Guodong Du, Junsheng Wang, Yongkang Ding, and Liyan Zhang. Prototype-guided cross-modal completion and alignment for incomplete text-based person re-identification. In ACM MM, pages 5253-5261, 2023. 4", + "[6] Tiantian Gong, Junsheng Wang, and Liyan Zhang. Cross-modal semantic aligning and neighbor-aware completing for robust text-image person retrieval. Information Fusion, 112: 102544, 2024. 4", + "[7] Xinqian Gu, Hong Chang, Bingpeng Ma, Shutao Bai, Shiguang Shan, and Xilin Chen. Clothes-changing person re-identification with rgb modality only. In CVPR, pages 1060–1069, 2022. 1, 2, 6", + "[8] Ke Han, Shaogang Gong, Yan Huang, Liang Wang, and Ti-niu Tan. Clothing-change feature augmentation for person re-identification. In CVPR, pages 22066–22075, 2023. 6", + "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 6", + "[10] Tao He, Leqi Shen, Yuchen Guo, Guiguang Ding, and Zhenhua Guo. Secret: Self-consistent pseudo label refinement for unsupervised domain adaptive person re-identification. In AAAI, pages 879–887, 2022. 2", + "[11] Ziqi He, Mengjia Xue, Yunhao Du, Zhicheng Zhao, and Fei Su. Dynamic clustering and cluster contrastive learning for unsupervised person re-id with feature distribution alignment. In ICASSP, pages 3610-3614, 2024. 6", + "[12] Alexander Hermans, Lucas Beyer, and Bastian Leibe. In defense of the triplet loss for person re-identification. arXiv preprint arXiv:1703.07737, 2017. 2", + "[13] Peixian Hong, Tao Wu, Ancong Wu, Xintong Han, and Wei-Shi Zheng. Fine-grained shape-appearance mutual learning for cloth-changing person re-identification. In CVPR, pages 10513–10522, 2021. 2" + ], + "bbox": [ + 93, + 200, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Yan Huang, Jingsong Xu, Qiang Wu, Yi Zhong, Peng Zhang, and Zhaoxiang Zhang. Beyond scalar neuron: Adopting vector-neuron capsules for long-term person re-identification. IEEE TCSVT, 30(10):3459-3471, 2019. 6", + "[15] Yan Huang, Qiang Wu, JingSong Xu, Yi Zhong, and ZhaoXiang Zhang. Clothing status awareness for long-term person re-identification. In ICCV, pages 11895–11904, 2021. 2, 6", + "[16] Xin Jin, Tianyu He, Kecheng Zheng, Zhiheng Yin, Xu Shen, Zhen Huang, Ruoyu Feng, Jianqiang Huang, Zhibo Chen, and Xian-Sheng Hua. Cloth-changing person re-identification from a single image with gait prediction and regularization. In CVPR, pages 14278–14287, 2022. 2, 6", + "[17] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[18] Long Lan, Xiao Teng, Jing Zhang, Xiang Zhang, and Dacheng Tao. Learning to purification for unsupervised person re-identification. IEEE TIP, 32:3338-3353, 2023. 2, 6", + "[19] Mingkun Li, Chun-Guang Li, and Jun Guo. Cluster-guided asymmetric contrastive learning for unsupervised person re-identification. IEEE TIP, 31:3606-3617, 2022. 6", + "[20] Mingkun Li, Peng Xu, Chun-Guang Li, and Jun Guo. SiCL: Silhouette-driven contrastive learning for unsupervised person re-identification with clothes change. arXiv preprint arXiv:2305.13600, 2023. 1, 2, 6", + "[21] Peike Li, Yunqiu Xu, Yunchao Wei, and Yi Yang. Self-correction for human parsing. IEEE TPAMI, 44(6):3260-3271, 2020. 1, 3", + "[22] Siyuan Li, Li Sun, and Qingli Li. Clip-reid: exploiting vision-language model for image re-identification without concrete text labels. In AAAI, pages 1405–1413, 2023. 2, 4", + "[23] Shuang Li, Jiaxu Leng, Guozhang Li, Ji Gan, Xinbo Gao, et al. Clip-driven cloth-agnostic feature learning for cloth-changing person re-identification. arXiv preprint arXiv:2406.09198, 2024. 1", + "[24] Weijia Li, Saihui Hou, Chunjie Zhang, Chunshui Cao, Xu Liu, Yongzhen Huang, and Yao Zhao. An in-depth exploration of person re-identification and gait recognition in cloth-changing conditions. In CVPR, pages 13824–13833, 2023. 1", + "[25] Yutian Lin, Xuanyi Dong, Liang Zheng, Yan Yan, and Yi Yang. A bottom-up clustering approach to unsupervised person re-identification. In AAAI, pages 8738–8745, 2019. 2", + "[26] Fangyi Liu, Mang Ye, and Bo Du. Cloth-aware augmentation for cloth-generalized person re-identification. In ACM MM, 2024. 1", + "[27] Vuong D Nguyen, Pranav Mantini, and Shishir K Shah. Contrastive clothing and pose generation for cloth-changing person re-identification. In CVPR, pages 7541–7549, 2024. 1", + "[28] Zhiqi Pang, Lingling Zhao, Qiuyang Liu, and Chunyu Wang. Camera invariant feature learning for unsupervised person re-identification. IEEE TMM, 25:6171-6182, 2022. 2", + "[29] Zhiqi Pang, Lingling Zhao, and Chunyu Wang. Clothing-invariant contrastive learning for unsupervised person re-identification. Neural Networks, page 106477, 2024. 1, 2, 6" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19259", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Xuelin Qian, Wenxuan Wang, Li Zhang, Fangrui Zhu, Yanwei Fu, Tao Xiang, Yu-Gang Jiang, and Xiangyang Xue. Long-term cloth-changing person re-identification. In ACCV, 2020. 2, 6", + "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 2, 3, 4, 6", + "[32] Jiangming Shi, Yachao Zhang, Xiangbo Yin, Yuan Xie, Zhizhong Zhang, Jianping Fan, Zhongchao Shi, and Yanyun Qu. Dual pseudo-labels interactive self-training for semi-supervised visible-infrared person re-identification. In ICCV, pages 11218-11228, 2023. 1", + "[33] Fangbin Wan, Yang Wu, Xuelin Qian, Yixiong Chen, and Yanwei Fu. When person re-identification meets changing clothes. In CVPRW, pages 830–831, 2020. 6", + "[34] Menglin Wang, Baisheng Lai, Jianqiang Huang, Xiaojin Gong, and Xian-Sheng Hua. Camera-aware proxies for unsupervised person re-identification. In AAAI, pages 2764–2772, 2021. 2", + "[35] Qizao Wang, Xuelin Qian, Bin Li, Xiangyang Xue, and Yanwei Fu. Exploring fine-grained representation and decomposition for cloth-changing person re-identification. IEEE TIFS, 2024. 6", + "[36] Shiyu Xuan and Shiliang Zhang. Intra-inter camera similarity for unsupervised person re-identification. In CVPR, pages 11926–11935, 2021. 2", + "[37] Shuanglin Yan, Neng Dong, Jun Liu, Liyan Zhang, and Jinhui Tang. Learning comprehensive representations with richer self for text-to-image person re-identification. In ACM MM, pages 6202–6211, 2023. 4", + "[38] Fengxiang Yang, Zhun Zhong, Zhiming Luo, Yuanzheng Cai, Yaojin Lin, Shaozi Li, and Nicu Sebe. Joint noisetolerant learning and meta camera shift adaptation for unsupervised person re-identification. In CVPR, pages 4855–4864, 2021. 2", + "[39] Qize Yang, Ancong Wu, and Wei-Shi Zheng. Person re-identification by contour sketch under moderate clothing change. IEEE TPAMI, 43(6):2029-2046, 2019. 6", + "[40] Zhengwei Yang, Meng Lin, Xian Zhong, Yu Wu, and Zheng Wang. Good is bad: Causality inspired cloth-debiasing for cloth-changing person re-identification. In CVPR, pages 1472–1481, 2023. 6", + "[41] Mang Ye, Jianbing Shen, Gaojie Lin, Tao Xiang, Ling Shao, and Steven CH Hoi. Deep learning for person re-identification: A survey and outlook. IEEE TPAMI, 44(6): 2872-2893, 2021. 1", + "[42] Junhui Yin, Xinyu Zhang, Zhanyu Ma, Jun Guo, and Yifan Liu. A real-time memory updating strategy for unsupervised person re-identification. IEEE TIP, 32:2309–2321, 2023. 6", + "[43] Xiangbo Yin, Jiangming Shi, Yachao Zhang, Yang Lu, Zhizhong Zhang, Yuan Xie, and Yanyun Qu. Robust pseudo-label learning with neighbor relation for unsupervised visible-infrared person re-identification. In ACM MM, pages 2242–2251, 2024. 1" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[44] Kaiwei Zeng, Munan Ning, Yaohua Wang, and Yang Guo. Hierarchical clustering with hard-batch triplet loss for person re-identification. In CVPR, pages 13657–13665, 2020. 2", + "[45] Xiao Zhang, Yixiao Ge, Yu Qiao, and Hongsheng Li. Refining pseudo labels with clustering consensus over generations for unsupervised object re-identification. In CVPR, pages 3436–3445, 2021. 2", + "[46] Xinyu Zhang, Dongdong Li, Zhigang Wang, Jian Wang, Errui Ding, Javen Qinfeng Shi, Zhaoxiang Zhang, and Jingdong Wang. Implicit sample extension for unsupervised person re-identification. In CVPR, pages 7369-7378, 2022. 2", + "[47] Yukang Zhang and Hanzi Wang. Diverse embedding expansion network and low-light cross-modality benchmark for visible-infrared person re-identification. In CVPR, pages 2153–2162, 2023. 1", + "[48] Yukang Zhang, Yan Yan, Jie Li, and Hanzi Wang. Mrcn: A novel modality restitution and compensation network for visible-infrared person re-identification. In AAAI, pages 3498–3506, 2023. 1", + "[49] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 130(9):2337-2348, 2022. 2, 4" + ], + "bbox": [ + 516, + 92, + 903, + 402 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19260", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Identity-Clothing Similarity Modeling for Unsupervised Clothing Change Person Re-Identification/42258662-090f-4600-a2a7-5bf4fc16321d_model.json b/2025/Identity-Clothing Similarity Modeling for Unsupervised Clothing Change Person Re-Identification/42258662-090f-4600-a2a7-5bf4fc16321d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..876b66c342c5c61a43c1f774656f9bfa553bfe4b --- /dev/null +++ b/2025/Identity-Clothing Similarity Modeling for Unsupervised Clothing Change Person Re-Identification/42258662-090f-4600-a2a7-5bf4fc16321d_model.json @@ -0,0 +1,2442 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.13, + 0.908, + 0.175 + ], + "angle": 0, + "content": "Identity-Clothing Similarity Modeling for Unsupervised Clothing Change Person Re-Identification" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.203, + 0.785, + 0.223 + ], + "angle": 0, + "content": "Zhiqi Pang\\(^{1}\\) Junjie Wang\\(^{2}\\) Lingling Zhao\\(^{1}\\) Chunyu Wang\\(^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.344, + 0.223, + 0.655, + 0.239 + ], + "angle": 0, + "content": "1 Harbin Institute of Technology, China" + }, + { + "type": "text", + "bbox": [ + 0.353, + 0.239, + 0.643, + 0.257 + ], + "angle": 0, + "content": "\\(^{2}\\) Nanjing Medical University, China" + }, + { + "type": "text", + "bbox": [ + 0.166, + 0.26, + 0.832, + 0.275 + ], + "angle": 0, + "content": "zqpang98@gmail.com, junjiehit@163.com, zhaoll@hit.edu.cn, chunyu@hit.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.309, + 0.327, + 0.325 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.341, + 0.486, + 0.735 + ], + "angle": 0, + "content": "Clothing change person re-identification (CC-ReID) aims to match different images of the same person, even when the clothing varies across images. To reduce manual labeling costs, existing unsupervised CC-ReID methods employ clustering algorithms to generate pseudo-labels. However, they often fail to assign the same pseudo-label to two images with the same identity but different clothing—referred to as a clothing change positive pair—thus hindering clothing-invariant feature learning. To address this issue, we propose the identity-clothing similarity modeling (ICSM) framework. To effectively connect clothing change positive pairs, ICSM first performs clothing-aware learning to leverage all discriminative information, including clothing, to obtain compact clusters. It then extracts cluster-level identity and clothing features and performs inter-cluster similarity estimation to identify clothing change positive clusters, reliable negative clusters, and hard negative clusters for each compact cluster. During optimization, we design an adaptive version of existing optimization methods to enhance similarities of clothing change positive pairs, while also introducing text semantics as a supervisory signal to further promote clothing invariance. Extensive experimental results across multiple datasets validate the effectiveness of the proposed framework, demonstrating its superiority over existing unsupervised methods and its competitiveness with some supervised approaches." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.763, + 0.222, + 0.779 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.787, + 0.484, + 0.879 + ], + "angle": 0, + "content": "Person re-identification (ReID) [32, 41, 43, 47, 48] aims to match images of the same individual captured in multicamera scenarios. Compared to traditional ReID, clothing change ReID (CC-ReID) [7, 23, 24, 26, 27] assumes that individuals may change clothing and focuses on matching images of the same person with different outfits. Although" + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.312, + 0.706, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.565, + 0.403, + 0.669, + 0.416 + ], + "angle": 0, + "content": "(a) Original images" + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.311, + 0.889, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.725, + 0.403, + 0.883, + 0.416 + ], + "angle": 0, + "content": "(b) Clothing-occluded images" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.428, + 0.907, + 0.512 + ], + "angle": 0, + "content": "Figure 1. Visualization of features extracted from original images and clothing-occluded images. Different colors represent features with different identities, while the same color but different shapes indicate features with the same identity but different clothing. Image pairs that are correctly (incorrectly) clustered by clustering algorithms are connected with green (red) arrows." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.907, + 0.78 + ], + "angle": 0, + "content": "existing CC-ReID methods have achieved promising performance, they largely depend on large-scale manual annotations. However, manual labeling for retrieval tasks is an extremely labor-intensive process [1, 3]. Consequently, researchers have begun to focus on unsupervised CC-ReID (UCC-ReID) [20, 29]. In UCC-ReID, two images of the same person with different clothing—referred to as a clothing change positive pair—often exhibit significant appearance differences. This makes it challenging for clustering algorithms [4] to group their features into the same cluster, thus preventing them from obtaining the same pseudo-label. This phenomenon severely hinders the performance improvement of UCC-ReID methods. We refer to each clothing change positive pair that receives the same pseudo-label as a connected positive pair, while those that receive different pseudo-labels are termed unconnected positive pairs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.909, + 0.902 + ], + "angle": 0, + "content": "To obtain robust pseudo-labels for clothing changes, existing UCC-ReID methods [20, 29] leverage body shapes extracted by human parsing models [21], which are typically clothing-invariant and identity-related. For example, CICL [29] clusters the fused features of original and body shape images. However, the unavoidable presence of clothing information in the fused features limits the number of connected positive pairs, while generating a large number" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.318, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author: Chunyu Wang" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19251" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.228 + ], + "angle": 0, + "content": "of unconnected positive pairs. Additionally, existing UCC-ReID methods [20, 29] typically enhance intra-class compactness and inter-class separability indiscriminately during the optimization phase, similar to unsupervised traditional ReID methods [38, 46]. However, such optimization methods reduce the similarity of each unconnected positive pair, i.e., they decrease the feature similarity between images of the same person with different clothing, thus hindering clothing-invariance learning." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.23, + 0.485, + 0.593 + ], + "angle": 0, + "content": "To address the above issues, we propose the identityclothing similarity modeling (ICSM) framework. An intuitive idea for obtaining robust pseudo-labels for clothing changes is to cluster the features of clothing-occluded images. However, we find this approach to be unfeasible. Specifically, clothing information is not entirely detrimental in the process of obtaining accurate pseudo-labels. As shown in Fig. 1a, the features of original images with the same identity and clothing—referred to as a same clothing positive pair—demonstrate high compactness. In contrast, as depicted in Fig. 1b, the clothing-occluded images lose some clothing-related discriminative information, resulting in reduced compactness for same clothing positive pairs. This significantly hinders clustering algorithms [4] from assigning the same pseudo-label to each same clothing positive pair. To this end, ICSM first performs clothing-aware learning to utilize all discriminative information, including clothing, to obtain compact clusters. It then extracts cluster-level identity and clothing features using visual-language models [22, 31, 49], and finally introduces an inter-cluster similarity estimation method to integrate identity and clothing similarities, identifying clothing change positive clusters, reliable negative clusters, and hard negative clusters for each compact cluster." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.595, + 0.483, + 0.761 + ], + "angle": 0, + "content": "During the optimization phase, we enhance the model's robustness to clothing changes from both image semantics and text semantics. On one hand, we develop an adaptive version of existing optimization methods that focuses solely on reliable positive and negative pairs, thereby improving similarities of connected positive pairs while maintaining similarities of unconnected positive pairs. On the other hand, we introduce text semantics as a supervisory signal and design identity-oriented semantic contrastive loss and clothing-invariant semantic contrastive loss to further promote clothing invariance learning." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.763, + 0.451, + 0.777 + ], + "angle": 0, + "content": "The main contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.841 + ], + "angle": 0, + "content": "- We propose an identity-clothing similarity modeling (ICSM) framework, which aims to connect clothing change positive pairs and promote clothing invariance learning to address clothing variation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.484, + 0.902 + ], + "angle": 0, + "content": "- We develop an inter-cluster similarity estimation method that not only fully utilizes clothing-related discriminative information but also integrates identity and clothing similarities to identify clothing change positive clusters, reli" + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.781, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "able negative clusters, and hard negative clusters for each compact cluster." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.122, + 0.905, + 0.197 + ], + "angle": 0, + "content": "- We not only introduce an adaptive version of traditional optimization methods for UCC-ReID, but also design identity-oriented semantic contrastive loss and clothing-invariant semantic contrastive loss to leverage text semantics in promoting clothing invariance learning." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.198, + 0.905, + 0.273 + ], + "angle": 0, + "content": "- Experimental results on multiple CC-ReID datasets validate the effectiveness of the proposed framework. ICSM not only outperforms existing UCC-ReID methods but also achieves competitive performance compared to earlier supervised methods." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.122, + 0.905, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.289, + 0.655, + 0.305 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.315, + 0.792, + 0.331 + ], + "angle": 0, + "content": "2.1. Unsupervised Traditional ReID" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.337, + 0.907, + 0.626 + ], + "angle": 0, + "content": "Unsupervised traditional ReID (UT-ReID) methods [1, 3, 18, 45] assume that individuals do not change clothing, iterating between clustering and optimization. In early research, both BUC [25] and HCT [44] employ hierarchical clustering to obtain identity pseudo-labels, with HCT [44] also incorporating a batch-hard triplet loss [12] to address hard samples. Subsequent studies, such as CAP [34] and ICE [1], introduce camera invariance learning to mitigate the impact of varying camera conditions. In contrast to these approaches, IICS [36] and CIFL [28] focus on obtaining pseudo-labels robust to camera discrepancies. To alleviate the influence of noisy labels, both PPLR [2] and SE-CRET [10] assess the consistency between global and local features to enhance the overall reliability of training pseudolabels. Purification [18] supervises the student model using a teacher model to reduce the interference of noisy labels. Although existing UT-ReID methods face challenges in adapting to CC-ReID scenarios, their contributions to unsupervised learning provide valuable insights for our work." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.637, + 0.727, + 0.653 + ], + "angle": 0, + "content": "2.2. Clothing Change ReID" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Clothing change ReID (CC-ReID) methods [7, 13, 16] typically assumes that individuals may change their clothing and emphasizes the importance of clothing-invariant and identity-related features. Existing methods often introduce additional identity information (e.g., keypoints [30], shape [13], and gait [16]) or clothing information (e.g., status [15] and category [7]) to aid the optimization process, achieving promising performance. However, most existing methods rely heavily on manual identity labels or even fine-grained clothing labels, which significantly limits flexibility and scalability. To address these issues, researchers explore several unsupervised CC-ReID methods. For instance, SiCL [20] and CICL [29] focus on utilizing body shape to obtain pseudo-labels that are robust to clothing variations. However, the presence of numerous unconnected positive pairs severely restricts their performance, a factor they overlook." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19252" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.168 + ], + "angle": 0, + "content": "To address this challenge, we integrates identity and clothing similarities to reduce the number of unconnected positive pairs. Additionally, we limit the optimization scope to reliable positive and negative pairs, thereby maintaining similarities of unconnected positive pairs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.181, + 0.299, + 0.198 + ], + "angle": 0, + "content": "3. Proposed Framework" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.206, + 0.485, + 0.614 + ], + "angle": 0, + "content": "During the training phase, we are provided with an unlabeled training image set \\(\\{x_{i}\\}_{i = 1}^{N}\\), where \\(N\\) denotes the number of images. Images from the same individual may exhibit different clothing. As illustrated in Fig. 2, our proposed identity-clothing similarity modeling (ICSM) consists of three stages. Specifically, stage I (clothing-aware learning) iterates through the following two processes: 1) clustering image features to generate pseudo-labels, and 2) optimizing the image encoder \\(E_{\\mathrm{I}}\\) based on the pseudolabels. The final pseudo-labels are saved, and the clusters corresponding to these pseudo-labels are referred to as compact clusters. Stage II (cluster-level feature learning) first extracts identity-related image set \\(\\{x_{i}^{s}\\}_{i = 1}^{N}\\) and clothing-related image set \\(\\{x_{i}^{c}\\}_{i = 1}^{N}\\) from \\(\\{x_{i}\\}_{i = 1}^{N}\\), and then introduces two CLIP models [31]: CLIP-S and CLIP-C. These models are used to learn cluster-level identity features and cluster-level clothing features, respectively, for each compact cluster. Stage III (clothing-invariant learning) first performs inter-cluster similarity estimation (ISE) to identify clothing change positive clusters, reliable negative clusters, and hard negative clusters for each compact cluster. Subsequently, we not only introduce an adaptive version of traditional optimization methods to help the image encoder \\(E_{\\mathrm{III}}\\) learn features that are identity-relevant and clothing-invariant, but also design identity-oriented semantic contrastive loss and clothing-invariant semantic contrastive loss to further promote clothing invariance learning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.484, + 0.675 + ], + "angle": 0, + "content": "It is important to note that although ICSM involves a three-stage training process, the optimization in each stage is efficient and does not significantly increase the overall training time. Detailed analysis is provided in Sec. 4.3." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.684, + 0.327, + 0.701 + ], + "angle": 0, + "content": "3.1. Clothing-Aware Learning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.706, + 0.484, + 0.825 + ], + "angle": 0, + "content": "In this stage, we iteratively train the image encoder \\( E_{\\mathrm{I}} \\) following the traditional unsupervised ReID approaches [1, 3]. Specifically, in each iteration, we first use \\( E_{\\mathrm{I}} \\) to extract features from all training images. Then, we apply the DBSCAN algorithm [4] to all image features to obtain pseudolabels, and subsequently compute the centroid for each cluster. For example, the centroid of the \\( p \\)-th cluster is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.23, + 0.824, + 0.483, + 0.866 + ], + "angle": 0, + "content": "\\[\nv _ {p} = \\frac {1}{N _ {p}} \\sum_ {i = 1} ^ {N _ {p}} f _ {i}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "where \\(N_{p}\\) denotes the number of images in the \\(p\\)-th cluster, and \\(f_{i}\\) represents the features of the image \\(x_{i}\\) within that" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.138 + ], + "angle": 0, + "content": "cluster. Then, we introduce the identity classification loss and hard contrastive loss to optimize \\( E_{\\mathrm{I}} \\). For any image \\( x_{i} \\), the identity classification loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.147, + 0.907, + 0.188 + ], + "angle": 0, + "content": "\\[\nL _ {i d} = - \\log \\frac {\\exp \\left(f _ {i} \\cdot v _ {p} ^ {\\mathrm {T}} / \\tau_ {i d}\\right)}{\\sum_ {q = 1} ^ {| V |} \\exp \\left(f _ {i} \\cdot v _ {q} ^ {\\mathrm {T}} / \\tau_ {i d}\\right)}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.197, + 0.907, + 0.257 + ], + "angle": 0, + "content": "where \\( f_{i} \\) is the feature of image \\( x_{i} \\), \\( v_{p} \\) is the centroid of \\( f_{i} \\), \\( |V| \\) is the number of centroids in the current iteration, and \\( \\tau_{id} \\) is the temperature hyperparameter. For any image \\( x_{i} \\), the hard contrastive loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.56, + 0.265, + 0.907, + 0.307 + ], + "angle": 0, + "content": "\\[\nL _ {h c} = - \\log \\frac {\\exp \\left(f _ {i} \\cdot f _ {p} ^ {\\mathrm {T}} / \\tau_ {h c}\\right)}{\\sum_ {f _ {q} \\in F _ {i} ^ {n} \\cup f _ {p}} \\exp \\left(f _ {i} \\cdot f _ {q} ^ {\\mathrm {T}} / \\tau_ {h c}\\right)}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.316, + 0.906, + 0.423 + ], + "angle": 0, + "content": "where \\( f_{i}^{p} \\) is the feature of the most distant positive sample from \\( f_{i} \\) in a minibatch of size \\( P\\times K \\) (where \\( P \\) is the number of sampled clusters and \\( K \\) is the number of images sampled from each cluster), \\( F_{i}^{n} \\) denotes the set of features corresponding to all negative samples of \\( f_{i} \\) within the minibatch, and \\( \\tau_{hc} \\) is the temperature hyperparameter. In summary, the overall loss function of stage I is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.648, + 0.435, + 0.907, + 0.452 + ], + "angle": 0, + "content": "\\[\nL _ {s 1} = L _ {i d} + L _ {h c}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.462, + 0.907, + 0.584 + ], + "angle": 0, + "content": "The clustering and optimization process is iteratively performed, and the final pseudo-labels are saved. Due to the discriminative power brought by clothing, the clusters corresponding to these pseudo-labels exhibit high compactness, which we refer to as \"compact clusters\". In the following, the term \"cluster\" typically refers to these \"compact clusters.\" The images within a cluster usually share not only the same identity but also the same clothing." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.592, + 0.795, + 0.608 + ], + "angle": 0, + "content": "3.2. Cluster-Level Feature Learning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.906, + 0.765 + ], + "angle": 0, + "content": "Due to the lack of connected positive pairs, it is challenging to improve the model's robustness to clothing changes based solely on compact clusters. Therefore, before training, we should attempt to connect two clusters that each contain images of the same person but with different clothing, i.e., a clothing change positive cluster pair. To this end, stage II focuses on obtaining cluster-level identity feature and cluster-level clothing feature for each compact cluster, thereby laying the foundation for connecting clothing change positive cluster pairs in the stage III." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Stage II first obtains identity-related images and clothing-related images. As shown in Fig. 3, for any given original image \\( x_{i} \\), we first use a human parsing model [21] to obtain the parsed image \\( \\hat{x_i} \\). Based on \\( \\hat{x_i} \\), we then extract the identity-related image \\( x_{i}^{s} \\) and the clothing-related image \\( x_{i}^{c} \\) from the original image \\( x_{i} \\). Specifically, in \\( x_{i}^{s} \\), the pixel values of the clothing region are set to 0, while the pixel values of the other body regions are kept the same as in \\( x_{i} \\) and the background pixel values are set to 255. In \\( x_{i}^{c} \\), the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19253" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.133, + 0.087, + 0.625, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.133, + 0.194, + 0.625, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.134, + 0.334, + 0.625, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.636, + 0.089, + 0.864, + 0.108 + ], + "angle": 0, + "content": "Stage II (cluster-level feature learning)" + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.108, + 0.864, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.4, + 0.908, + 0.442 + ], + "angle": 0, + "content": "Figure 2. The proposed identity-clothing similarity modeling (ICSM) framework for UCC-ReID as a three-stage pipeline. The different colors of the clusters represent different pseudo-labels. The image encoders \\( E_{\\mathrm{I}} \\) and \\( E_{\\mathrm{III}} \\) are optimized, while all image and text encoders in the stage II are frozen. See description in the text for additional detail." + }, + { + "type": "image", + "bbox": [ + 0.134, + 0.465, + 0.443, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.539, + 0.483, + 0.569 + ], + "angle": 0, + "content": "Figure 3. Example of obtaining identity-related image \\( x_{i}^{s} \\) and clothing-related image \\( x_{i}^{c} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.597, + 0.483, + 0.671 + ], + "angle": 0, + "content": "pixel values of the clothing region remain the same as in \\( x_{i} \\), while all other regions are set to 255. We refer to the clusters composed of identity-related images as identity-related clusters, and the clusters composed of clothing-related images as clothing-related clusters." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.673, + 0.483, + 0.824 + ], + "angle": 0, + "content": "Subsequently, stage II leverages the strong summarizing and descriptive ability of text for person images [5, 6, 37], along with the powerful text-image understanding capabilities of vision-language models [22, 31, 49], to obtain cluster-level features in text form. Specifically, we introduce two CLIP models [31]: CLIP-S and CLIP-C, which share the same pre-trained parameters. CLIP-S is used to learn cluster-level identity features from identity-related clusters, while CLIP-C is used to learn cluster-level clothing features from clothing-related clusters." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "For CLIP-S, we first set an identity-related text for each identity-related cluster as \"A photo of a \\([S_1][S_2]\\ldots [S_M]\\) person,\" where \\(\\left[[S_1][S_2]\\ldots [S_M]\\right]\\) represents a set of learnable text embeddings, and \\(M\\) is the number of learnable text embeddings. For any identity-related image \\(x_{i}^{s}\\), we input it" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.469, + 0.907, + 0.545 + ], + "angle": 0, + "content": "along with the corresponding identity-related text into the image encoder and text encoder of CLIP-S, respectively. This produces the image feature \\( f_{i}^{s} \\) and text feature \\( t_{i}^{s} \\). Afterward, we freeze the image and text encoders of CLIP-S and optimize the learnable text embeddings, as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.654, + 0.555, + 0.905, + 0.571 + ], + "angle": 0, + "content": "\\[\nL ^ {s} = L _ {i t} ^ {s} + L _ {t i} ^ {s}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.581, + 0.738, + 0.598 + ], + "angle": 0, + "content": "where \\(L_{it}^{s}\\) and \\(L_{ti}^{s}\\) are defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.595, + 0.606, + 0.905, + 0.647 + ], + "angle": 0, + "content": "\\[\nL _ {i t} ^ {s} = - \\log \\frac {\\exp \\left(f _ {i} ^ {s} \\cdot t _ {i} ^ {s ^ {\\mathrm {T}}}\\right)}{\\sum_ {j = 1} ^ {B} \\exp \\left(f _ {i} ^ {s} \\cdot t _ {j} ^ {s ^ {\\mathrm {T}}}\\right)}, \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.566, + 0.656, + 0.905, + 0.7 + ], + "angle": 0, + "content": "\\[\nL _ {t i} ^ {s} = \\frac {- 1}{| \\psi_ {i} ^ {s} |} \\sum_ {f _ {p} ^ {s} \\in \\psi_ {i} ^ {s}} \\log \\frac {\\exp \\left(t _ {i} ^ {s} \\cdot f _ {p} ^ {s ^ {\\mathrm {T}}}\\right)}{\\sum_ {j = 1} ^ {B} \\exp \\left(t _ {i} ^ {s} \\cdot f _ {j} ^ {s ^ {\\mathrm {T}}}\\right)}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.905, + 0.734 + ], + "angle": 0, + "content": "where \\(\\psi_i^s\\) represents the set of image features that share the same pseudo-label as \\(t_i^s\\), and \\(B\\) is the batch size." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Finally, we use the optimized text feature \\( t_i^s \\) as the cluster-level identity feature for the corresponding compact cluster \\( u_i \\). Similarly, for CLIP-C, we first set a clothing-related text for each clothing-related cluster: \"A photo of \\( [C_1][C_2]\\ldots [C_M] \\) clothing\", where \\( \"[C_1][C_2]\\ldots [C_M]\" \\) represents a set of learnable text embeddings. Then, we follow the same optimization scheme as for \\( L_{it}^s \\) and \\( L_{ti}^s \\) to set \\( L_{it}^c \\) and \\( L_{ti}^c \\) for CLIP-C. Based on CLIP-C and the clothing-related text, we obtain the cluster-level clothing feature for each compact cluster. For example, the cluster-level clothing feature for the \\( i \\)-th cluster \\( u_i \\) is represented as \\( t_i^c \\)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19254" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.348, + 0.108 + ], + "angle": 0, + "content": "3.3. Clothing-Invariant Learning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.484, + 0.265 + ], + "angle": 0, + "content": "Inter-cluster similarity estimation. In theory, the higher the similarity between the cluster-level identity features of two compact clusters, the more likely it is that they contain images of the same person. Conversely, the lower the similarity between the cluster-level clothing features of two clusters, the more likely it is that they contain images with different clothing. Therefore, in stage III, we propose an inter-cluster similarity estimation (ISE) method, which seeks clothing change positive clusters, reliable negative clusters, and hard negative clusters for each cluster." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.267, + 0.484, + 0.372 + ], + "angle": 0, + "content": "For any cluster \\( u_{i} \\), ISE first calculates similarities between its cluster-level identity feature and clothing feature with those of other clusters, and then ranks all other clusters in descending order of similarity, resulting in the corresponding identity similarity sequence and clothing similarity sequence for \\( u_{i} \\). For any cluster \\( u_{i} \\), the set consisting of all clothing change positive clusters is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.385, + 0.483, + 0.406 + ], + "angle": 0, + "content": "\\[\nU _ {i} ^ {c} = \\left\\{u _ {j} \\mid t _ {i} ^ {s} \\cdot t _ {j} ^ {s ^ {\\mathrm {T}}} > \\eta_ {a} \\text {a n d} t _ {i} ^ {c} \\cdot t _ {j} ^ {c ^ {\\mathrm {T}}} \\leq \\eta_ {c} \\right\\}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.417, + 0.484, + 0.507 + ], + "angle": 0, + "content": "where \\(\\eta_{a}\\) denotes the similarity value ranked at \\(\\gamma_{a}\\) in the identity similarity sequence, and \\(\\eta_{c}\\) denotes the similarity value ranked at \\(\\gamma_{c}\\) in the clothing similarity sequence. Additionally, for any cluster \\(u_{i}\\), we define the positive cluster set as \\(U_{i}^{a} = U_{i}^{c} \\cup u_{i}\\). For any cluster \\(u_{i}\\), the set of its reliable negative clusters is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.201, + 0.52, + 0.483, + 0.542 + ], + "angle": 0, + "content": "\\[\nU _ {i} ^ {b} = \\left\\{u _ {j} \\mid t _ {i} ^ {s} \\cdot t _ {j} ^ {s T} \\leq \\eta_ {b} \\right\\}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.553, + 0.484, + 0.645 + ], + "angle": 0, + "content": "where \\(\\eta_{b}\\) denotes the similarity value ranked at \\(\\gamma_{b}\\) in the identity similarity sequence. For any cluster \\(u_{i}\\), we define the clusters with low cluster-level identity feature similarity and high cluster-level clothing feature similarity as hard negative clusters. Therefore, for any cluster \\(u_{i}\\), the set of its hard negative clusters is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.122, + 0.656, + 0.483, + 0.677 + ], + "angle": 0, + "content": "\\[\nU _ {i} ^ {h} = \\left\\{u _ {j} \\mid t _ {i} ^ {s} \\cdot t _ {j} ^ {s T} \\leq \\eta_ {b} \\text {a n d} t _ {i} ^ {c} \\cdot t _ {j} ^ {c T} > \\eta_ {c} \\right\\}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.734 + ], + "angle": 0, + "content": "Finally, for any cluster \\( u_{i} \\), we save the reliable positive cluster set \\( U_{i}^{a} \\), clothing change positive cluster set \\( U_{i}^{c} \\), reliable negative cluster set \\( U_{i}^{b} \\), and hard negative cluster set \\( U_{i}^{h} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Image semantic contrastive learning. Due to the difficulty in fully connecting all clothing change positive cluster pairs, there may be additional clothing change positive clusters beyond the clothing change positive cluster set \\( U_{i}^{c} \\) for any cluster \\( u_{i} \\). In this scenario, traditional optimization methods [1, 3] aimed at increasing inter-class separability would reduce the similarities between unconnected positive pairs, thereby hindering clothing-invariant learning. To address this, we design an adaptive version of the traditional optimization method for the UCC-ReID scenario. Specifically, for any image feature \\( f_{i} \\), the adaptive identity classification" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.634, + 0.106 + ], + "angle": 0, + "content": "loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.526, + 0.111, + 0.905, + 0.164 + ], + "angle": 0, + "content": "\\[\nL _ {a i d} = \\frac {- 1}{\\left| V _ {i} ^ {a} \\right|} \\sum_ {v _ {p} \\in V _ {i} ^ {a}} \\log \\frac {\\exp \\left(f _ {i} \\cdot v _ {p} ^ {\\mathrm {T}} / \\tau_ {i d}\\right)}{\\sum_ {v _ {q} \\in V _ {i} ^ {b} \\cup v _ {p}} \\exp \\left(f _ {i} \\cdot v _ {q} ^ {\\mathrm {T}} / \\tau_ {i d}\\right)}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.165, + 0.907, + 0.224 + ], + "angle": 0, + "content": "where \\( V_{i}^{a} \\) and \\( V_{i}^{b} \\) respectively represent the set of centroids of all clusters in \\( U_{i}^{a} \\) and \\( U_{i}^{b} \\) corresponding to \\( f_{i} \\). For any image feature \\( f_{i} \\), the adaptive hard contrastive loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.229, + 0.905, + 0.282 + ], + "angle": 0, + "content": "\\[\nL _ {a h c} = \\frac {- 1}{\\left| V _ {i} ^ {c} \\right|} \\sum_ {v _ {p} \\in V _ {i} ^ {c}} \\log \\frac {\\exp \\left(f _ {i} \\cdot v _ {p} ^ {\\mathrm {T}} / \\tau_ {h c}\\right)}{\\sum_ {v _ {q} \\in V _ {i} ^ {h} \\cup v _ {p}} \\exp \\left(f _ {i} \\cdot v _ {q} ^ {\\mathrm {T}} / \\tau_ {h c}\\right)}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.283, + 0.907, + 0.344 + ], + "angle": 0, + "content": "where \\( V_{i}^{c} \\) and \\( V_{i}^{h} \\) respectively represent the set of centroids of all clusters in \\( U_{i}^{c} \\) and \\( U_{i}^{h} \\) corresponding to \\( f_{i} \\). Since the above process is optimized only for reliable clusters, it maintains the similarities of unconnected positive pairs." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.344, + 0.906, + 0.468 + ], + "angle": 0, + "content": "Text semantic contrastive learning. To fully leverage identity-related and clothing-related text semantics to assist model optimization, we introduce the identity-oriented semantic contrastive loss and clothing-invariant semantic contrastive loss. Firstly, for each image \\( x_{i} \\), we compute the similarity distribution \\( \\alpha_{i} \\) between the cluster-level identity feature \\( t_{i}^{s} \\) of \\( x_{i} \\) and all other cluster-level identity features. The \\( k \\)-th element of \\( \\alpha_{i} \\), denoted as \\( \\alpha_{i}^{(k)} \\), is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.593, + 0.475, + 0.905, + 0.515 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {i} ^ {(k)} = \\frac {\\exp \\left(t _ {i} ^ {s} \\cdot t _ {k} ^ {s T} / \\tau_ {i s}\\right)}{\\sum t _ {j} ^ {s} \\in T _ {i} ^ {s}} \\exp \\left(t _ {i} ^ {s} \\cdot t _ {j} ^ {s T} / \\tau_ {i s}\\right), \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.521, + 0.906, + 0.6 + ], + "angle": 0, + "content": "where \\( T_{i}^{s} \\) represents the set of all cluster-level identity features except for \\( t_{i}^{s} \\), and \\( \\tau_{is} \\) is the temperature hyperparameter. Next, we compute the similarity distribution \\( \\beta_{i} \\) between the image feature \\( f_{i} \\) of \\( x_{i} \\) and the centroids of all other clusters. The \\( k \\)-th element of \\( \\beta_{i} \\), denoted as \\( \\beta_{i}^{(k)} \\), is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.592, + 0.609, + 0.905, + 0.647 + ], + "angle": 0, + "content": "\\[\n\\beta_ {i} ^ {(k)} = \\frac {\\exp \\left(f _ {i} \\cdot v _ {k} ^ {\\mathrm {T}} / \\tau_ {i s}\\right)}{\\sum_ {v _ {j} \\in V _ {i}} \\exp \\left(f _ {i} \\cdot v _ {j} ^ {\\mathrm {T}} / \\tau_ {i s}\\right)}, \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.649, + 0.906, + 0.754 + ], + "angle": 0, + "content": "where \\( V_{i} \\) represents the set of all centroids except for the centroid of \\( x_{i} \\). In theory, if all image features exhibit high robustness to clothing changes, then \\( \\beta_{i} \\) should exhibit high similarity to \\( \\alpha_{i} \\). Therefore, to leverage identity-related text semantics to ensure high robustness of image features to clothing changes, we introduce the identity-oriented semantic contrastive loss:" + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.76, + 0.905, + 0.803 + ], + "angle": 0, + "content": "\\[\nL _ {i s} = \\sum_ {k = 1} ^ {| V | - 1} \\alpha_ {i} ^ {(k)} \\log \\left(\\alpha_ {i} ^ {(k)} / \\beta_ {i} ^ {(k)}\\right), \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.81, + 0.906, + 0.87 + ], + "angle": 0, + "content": "where \\( |V| \\) is the number of centroids in the current iteration. Additionally, to further promote clothing-invariant learning by leveraging clothing-related text semantics, we introduce the clothing-invariant semantic contrastive loss:" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.876, + 0.905, + 0.904 + ], + "angle": 0, + "content": "\\[\nL _ {c s} = \\left[ f _ {i} \\cdot t _ {p} ^ {c \\mathrm {T}} - \\min _ {t _ {q} ^ {c} \\in T _ {i} ^ {a}} \\left(f _ {i} \\cdot t _ {q} ^ {c \\mathrm {T}}\\right) \\right] _ {+}, \\tag {16}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19255" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.137 + ], + "angle": 0, + "content": "where \\(t_p^c\\) represents the cluster-level clothing feature of image feature \\(f_{i}\\), and \\(T_{i}^{a}\\) is the set of cluster-level clothing features of all clusters in \\(U_{i}^{a}\\) corresponding to \\(f_{i}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.139, + 0.483, + 0.169 + ], + "angle": 0, + "content": "In summary, the overall loss function of stage III is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.153, + 0.186, + 0.483, + 0.203 + ], + "angle": 0, + "content": "\\[\nL _ {s 3} = L _ {a i d} + L _ {a h c} + \\lambda_ {i s} L _ {i s} + \\lambda_ {c s} L _ {c s}, \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.217, + 0.483, + 0.248 + ], + "angle": 0, + "content": "where \\(\\lambda_{is}\\) and \\(\\lambda_{cs}\\) are the weighting coefficients for \\(L_{is}\\) and \\(L_{cs}\\), respectively." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.267, + 0.216, + 0.284 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.294, + 0.379, + 0.309 + ], + "angle": 0, + "content": "4.1. Datasets and Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.318, + 0.484, + 0.56 + ], + "angle": 0, + "content": "We evaluate the proposed framework on the PRCC [39], VC-Clothes [33], LTCC [30], Celeb-reID [14], and Celeb-reID-light [14] datasets, with statistical results presented in Tab. 1. Following existing works [20, 29], we employ three testing schemes: (1) clothing change (CC): the gallery contains only images with clothing changes; (2) same clothing (SC): the gallery includes only images with the same clothing; (3) general (G): the gallery consists of images with both clothing changes and the same clothing. For the PRCC [39] dataset, we utilize the CC and SC testing schemes; for VC-Clothes [33] and LTCC [30], we use the CC and G schemes; and for Celeb-reID [14] and Celeb-reID-light [14], we implement the CC testing scheme. For each testing scheme, we assess recognition performance using mean average precision (mAP) and cumulative matching characteristic (CMC)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.575, + 0.308, + 0.591 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.484, + 0.901 + ], + "angle": 0, + "content": "All images are resized to \\(256\\times 128\\) pixels. During the training phase, random horizontal flipping and random cropping are employed for data augmentation. All image encoders utilize ResNet-50 [9] pre-trained on CLIP [31] as the backbone. For DBSCAN [4], the distance threshold and minimum sample size are set to 0.55 and 4, respectively. The pseudo-labels from stage I are saved and passed on to stage II and stage III, where no further clustering or pseudo-label updates are performed. For each set of learnable text embeddings, we set \\(M = 4\\). For ISE, we set \\(\\gamma_{a} = 2\\%\\), \\(\\gamma_{c} = 50\\%\\), and \\(\\gamma_{b} = 60\\%\\). For stage I and stage III, we configure each minibatch to contain \\(P = 16\\) identities, with each identity comprising \\(K = 4\\) images. For stage II, we set \\(B = 64\\). For temperature hyperparameters, we set \\(\\tau_{id} = 0.5\\), \\(\\tau_{hc} = 0.1\\), and \\(\\tau_{is} = 0.5\\). The model is optimized using the Adam optimizer [17], with an initial learning rate set to 0.00035. Both stage I and stage III consist of 50 epochs. In stage II, the learnable text embeddings in CLIP-S and CLIP-C are optimized for 120 epochs, respectively. During the testing phase, only the optimized image encoder \\(E_{\\mathrm{III}}\\) is used for inference." + }, + { + "type": "table", + "bbox": [ + 0.531, + 0.089, + 0.892, + 0.168 + ], + "angle": 0, + "content": "
DatasetsIdentitiesTrainingQueryGallery
PRCC22117,8963,54312,259
VC-Clothes2569,4491,0208,591
LTCC1529,5764937,050
Celeb-reID105220,2082,97211,006
Celeb-reID-light5909,021887934
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.177, + 0.905, + 0.206 + ], + "angle": 0, + "content": "Table 1. Statistical results of PRCC, VC-Clothes, LTCC, CelebreID, and CelebreID-light." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.236, + 0.818, + 0.252 + ], + "angle": 0, + "content": "4.3. Comparison with State-of-the-Arts" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.26, + 0.907, + 0.44 + ], + "angle": 0, + "content": "We evaluate the proposed method on PRCC [39], VC-Clothes (VC) [33], LTCC [30], Celeb-reID (CR) [14], and Celeb-reID-light (CRL) [14]. As shown in Tab. 2, we first report the performance of seven existing UT-ReID methods, including ICE [1], CC [3], CACL [19], PPLR [2], RTMem [42], Purification [18], and DCCC [11]. ICE [1], CACL [19], and Purification [18] achieve superior performance in these testing scenarios. We find that ICSM significantly outperforms all of the aforementioned methods, particularly in clothing change scenarios. This validates the effectiveness of ICSM's unique design for addressing clothing change scenarios." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.444, + 0.909, + 0.626 + ], + "angle": 0, + "content": "Subsequently, we compare advanced UCC-ReID methods: SiCL [20] and CICL [29]. We observe that ICSM's overall performance surpasses these methods, particularly achieving significant advantages in PRCC(CC) and VC(CC). This can be attributed to two main factors: first, ICSM combines identity and clothing similarities to identify clothing change positive clusters, reliable negative clusters, and hard negative clusters for each compact cluster. Second, ICSM not only incorporates an adaptive version of traditional optimization methods to maintain similarities of unconnected positive pairs but also leverages text semantics to further enhance clothing invariance learning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.627, + 0.909, + 0.807 + ], + "angle": 0, + "content": "Finally, we compare ICSM with existing supervised CC-ReID (SCC-ReID) methods, including RCSANet [15], GI-ReID [16], CAL [7], AIM [40], CCFA [8], and \\(\\mathrm{FIRe}^2\\) [35]. We observe that ICSM achieves competitive performance on PRCC(CC) compared to earlier SCC-ReID methods (RCSANet [15] and CAL [7]), and surpasses GI-ReID [16] on VC(CC) and LTCC(CC). Moreover, ICSM achieves an mAP of approximately \\(70\\%\\) compared to advanced methods (AIM [40], CCFA [8], and \\(\\mathrm{FIRe}^2\\) [35]) under the two testing settings of LTCC. These experimental results validate the superiority of ICSM and the potential of UCC-ReID." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Additionally, we report the training and test times for SiCL [20], CICL [29], and ICSM on VC(CC). As shown in Tab. 3, the training time for ICSM is slightly higher than that of SiCL and significantly lower than that of CICL. This is because, despite ICSM's three-stage training, the optimization in each stage is efficient and does not significantly" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19256" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.111, + 0.088, + 0.891, + 0.355 + ], + "angle": 0, + "content": "
MethodsReferencePRCC(CC)PRCC(SC)VC(CC)VC(G)LTCC(CC)LTCC(G)CRCRL
mAPR-1mAPR-1mAPR-1mAPR-1mAPR-1mAPR-1mAPR-1mAPR-1
UT-ReID
ICEICCV2145.532.695.793.328.531.451.870.110.116.322.644.04.940.75.07.1
CCACCV2246.334.494.490.225.731.045.162.46.07.411.017.03.432.83.24.4
CACLTIP2252.141.794.790.949.758.968.082.46.29.822.345.65.142.33.64.3
PPLRCVPR2251.440.091.787.423.132.547.768.14.44.86.011.24.841.34.36.2
RTMemTIP2347.235.893.692.737.546.159.775.35.87.518.743.15.039.84.96.8
PurificationTIP2352.341.995.293.047.656.866.881.68.915.623.347.65.242.64.97.0
DCCCICASSP2447.035.390.689.539.350.458.976.17.812.719.644.93.935.73.35.0
UCC-ReID
SiCLarXiv2455.443.296.293.863.971.777.287.910.120.727.657.65.645.45.610.5
CICLNN2455.345.096.795.369.372.876.787.313.025.729.660.36.749.37.812.5
ICSMOurs57.847.997.597.173.575.178.789.613.526.230.160.87.049.67.712.6
SCC-ReID
RCSANetICCV2148.650.297.2100.0--------11.955.616.729.5
GI-ReIDCVPR22-33.3--57.864.5--10.423.729.463.2----
CALCVPR2255.855.299.8100.081.781.487.292.9--------
AIMCVPR2358.357.999.9100.0----19.140.641.176.3----
CCFACVPR2358.461.298.799.6----22.145.342.575.8----
FIRe2TIFS2463.165.099.5100.0----19.144.639.975.918.264.0--
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.364, + 0.908, + 0.391 + ], + "angle": 0, + "content": "Table 2. Comparison of ICSM with existing methods. Optimal performance under two unsupervised settings is highlighted in bold. \"R-1\" denotes \"Rank-1\"." + }, + { + "type": "table", + "bbox": [ + 0.127, + 0.415, + 0.451, + 0.47 + ], + "angle": 0, + "content": "
MethodsmAPRank-1Training timeTest time
SiCL63.971.76798.2514.20
CICL69.372.810569.9314.16
ICSM73.575.17236.5113.97
" + }, + { + "type": "table_caption", + "bbox": [ + 0.107, + 0.48, + 0.461, + 0.495 + ], + "angle": 0, + "content": "Table 3. Statistics on performance and efficiency (seconds)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.522, + 0.484, + 0.643 + ], + "angle": 0, + "content": "increase the overall training time. In contrast, while existing UCC-ReID methods conduct only one stage of training, they often need to optimize three neural network modules simultaneously, which can require even more training time than ICSM. Additionally, ICSM has a similar testing time compared to existing UCC-ReID methods. These results demonstrate that ICSM not only achieves superior performance but also maintains efficiency." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.652, + 0.245, + 0.669 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.795 + ], + "angle": 0, + "content": "In this section, the baseline we tested optimizes the image encoder using only \\( L_{id} \\) and \\( L_{hc} \\) from stage I. As shown in Tab. 4, M1 corresponds to the baseline, while M2 introduces ISE on top of M1. Additionally, to accommodate ISE, M2 replaces \\( L_{id} \\) and \\( L_{hc} \\) with \\( L_{aid} \\) and \\( L_{ahc} \\), respectively. M3 and M4 introduce \\( L_{is} \\) and \\( L_{cs} \\), respectively, building upon M2. M5, which represents ICSM, incorporates both \\( L_{is} \\) and \\( L_{cs} \\) on top of M2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Effectiveness of ISE. M2 introduces ISE on top of M1 to identify clothing change positive clusters, reliable negative clusters, and hard negative clusters for each compact cluster. As shown in Tab. 4, M2 achieves significant performance improvements over M1 in PRCC(CC), VC-Clothes(CC), and VC-Clothes(G). To further explore the nature of ISE, we visualize the features of several clusters. As shown" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.419, + 0.707, + 0.532 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.597, + 0.533, + 0.635, + 0.545 + ], + "angle": 0, + "content": "(a) M1" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.419, + 0.897, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.786, + 0.533, + 0.825, + 0.545 + ], + "angle": 0, + "content": "(b) M2" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.557, + 0.908, + 0.614 + ], + "angle": 0, + "content": "Figure 4. Visualization of the clusters corresponding to M1 and M2. Clusters are labeled with black dashed lines. For any given cluster, the tail and head of the red arrow connect it to its clothing change positive cluster identified by ISE." + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.628, + 0.897, + 0.685 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.695, + 0.908, + 0.738 + ], + "angle": 0, + "content": "Figure 5. Feature maps of M1 and M2. Each triplet, from left to right, consists of the original image, the feature map of M1, and the feature map of M2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.903 + ], + "angle": 0, + "content": "in Fig. 4, M2 is able to identify clothing change positive clusters based on M1, significantly reducing the distances between the positive cluster pairs. For the unconnected clothing change positive clusters, the distances are not significantly increased because \\( L_{aid} \\) and \\( L_{ahc} \\) typically prevent these clusters from being included in the optimization scope. This process undoubtedly enhances the model's robustness to clothing changes. As shown in Fig. 5, we observe that M2 indeed focuses more on clothing-invariant re" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19257" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.152, + 0.089, + 0.849, + 0.18 + ], + "angle": 0, + "content": "
MethodsBaselineISE\\(L_{is}\\)\\(L_{cs}\\)PRCC(CC)PRCC(SC)VC-Clothes(CC)VC-Clothes(G)
mAPRank-1mAPRank-1mAPRank-1mAPRank-1
M1✓×××43.931.592.289.023.630.846.667.8
M2✓✓××53.643.095.192.857.664.969.983.5
M3✓✓✓×56.847.197.095.668.672.575.286.3
M4✓✓×✓55.344.695.393.263.869.272.785.0
M5/ICSM✓✓✓✓57.847.997.597.173.575.178.789.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.308, + 0.19, + 0.684, + 0.205 + ], + "angle": 0, + "content": "Table 4. Results of the ablation study for the proposed method." + }, + { + "type": "image", + "bbox": [ + 0.11, + 0.228, + 0.28, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.177, + 0.31, + 0.215, + 0.321 + ], + "angle": 0, + "content": "(a) M2" + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.229, + 0.465, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.361, + 0.31, + 0.399, + 0.321 + ], + "angle": 0, + "content": "(b) M3" + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.323, + 0.28, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.177, + 0.404, + 0.215, + 0.415 + ], + "angle": 0, + "content": "(c) M4" + }, + { + "type": "image", + "bbox": [ + 0.294, + 0.323, + 0.465, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.361, + 0.404, + 0.399, + 0.415 + ], + "angle": 0, + "content": "(d) M5" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.428, + 0.484, + 0.457 + ], + "angle": 0, + "content": "Figure 6. Feature distance distributions of same clothing positive pairs and clothing change positive pairs for M2, M3, M4, and M5." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.485, + 0.484, + 0.53 + ], + "angle": 0, + "content": "gions such as the head and arms, or areas like shoes that are less frequently changed, compared to M1. The above results validate the effectiveness of ISE." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.532, + 0.484, + 0.773 + ], + "angle": 0, + "content": "Effectiveness of \\( L_{is} \\) and \\( L_{cs} \\). As shown in Tab. 4, M3 and M4 introduce \\( L_{is} \\) and \\( L_{cs} \\) into M2, respectively, both achieving significant performance improvements. Additionally, we visualize the feature distance distributions of same clothing positive pairs and clothing change positive pairs for M2, M3, M4, and M5. As shown in Fig. 6, compared to M2, both M3 and M4 significantly reduce the feature distance of clothing change positive pairs while also increasing the overlap between the feature distance distributions of these two types. This validates that both \\( L_{is} \\) and \\( L_{cs} \\) can enhance model performance by promoting the learning of clothing-invariant features. Furthermore, we observe that M5 further reduces the feature distance of clothing change positive pairs compared to M3 and M4, achieving superior performance over all previously mentioned methods." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.787, + 0.277, + 0.803 + ], + "angle": 0, + "content": "4.5. Sensitivity Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.902 + ], + "angle": 0, + "content": "As shown in Fig. 7, when \\(\\lambda_{is} = 2\\) and \\(\\lambda_{cs} = 1\\), the model achieves optimal performance on both PRCC(CC) and VC-Clothes(CC). This confirms the generalizability of \\(\\lambda_{is}\\) and \\(\\lambda_{cs}\\). We also observe that when \\(\\lambda_{is}\\) is set too high or too low, model performance deteriorates. This is mainly because a small \\(\\lambda_{is}\\) reduces the contribution of identity-" + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.228, + 0.703, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.578, + 0.311, + 0.658, + 0.323 + ], + "angle": 0, + "content": "(a) PRCC(CC)" + }, + { + "type": "image", + "bbox": [ + 0.717, + 0.228, + 0.886, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.749, + 0.311, + 0.857, + 0.323 + ], + "angle": 0, + "content": "(b) VC-Clothes(CC)" + }, + { + "type": "image_caption", + "bbox": [ + 0.569, + 0.335, + 0.851, + 0.351 + ], + "angle": 0, + "content": "Figure 7. Impact of \\((\\lambda_{is},\\lambda_{cs})\\) on performance." + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.366, + 0.703, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.578, + 0.456, + 0.658, + 0.467 + ], + "angle": 0, + "content": "(a) PRCC(CC)" + }, + { + "type": "image", + "bbox": [ + 0.717, + 0.366, + 0.888, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.749, + 0.456, + 0.857, + 0.467 + ], + "angle": 0, + "content": "(b) VC-Clothes(CC)" + }, + { + "type": "image_caption", + "bbox": [ + 0.574, + 0.48, + 0.844, + 0.495 + ], + "angle": 0, + "content": "Figure 8. Impact of \\((\\gamma_{a},\\gamma_{c})\\) on performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.907, + 0.616 + ], + "angle": 0, + "content": "related text semantics, while a large \\(\\lambda_{is}\\) negatively impacts the effectiveness of \\(L_{aid}\\) and \\(L_{ahc}\\). As shown in Fig. 8, when \\(\\gamma_{a} = 2\\%\\) and \\(\\gamma_{c} = 50\\%\\), the model achieves optimal performance on both datasets. This not only validates the generalization capability of ISE but also verifies the model's robustness to \\(\\gamma_{a}\\) and \\(\\gamma_{c}\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.633, + 0.634, + 0.648 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we propose an identity-clothing similarity modeling (ICSM) framework for UCC-ReID. To facilitate clothing-invariant feature learning, we introduce intercluster similarity estimation (ISE), identity-oriented semantic contrastive loss, and clothing-invariant semantic contrastive loss. The results of the ablation study validate their effectiveness, demonstrating that ISE enhances the model's robustness to clothing changes by identifying reliable clothing change positive clusters, while identity-oriented semantic contrastive loss and clothing-invariant semantic contrastive loss further promote clothing invariance learning. Experimental results across eight testing scenarios validate the effectiveness and superiority of the proposed ICSM, which not only outperforms existing unsupervised methods but also achieves competitive performance compared to certain supervised methods." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19258" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.241, + 0.108 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.116, + 0.482, + 0.161 + ], + "angle": 0, + "content": "This work is supported by National Key Research and Development Program of China (Grant no. 2023YFC3305005)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.177, + 0.187, + 0.192 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.201, + 0.484, + 0.256 + ], + "angle": 0, + "content": "[1] Hao Chen, Benoit Lagadec, and Francois Bremond. ICE: Inter-instance contrastive encoding for unsupervised person re-identification. In ICCV, pages 14960-14969, 2021. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.259, + 0.484, + 0.313 + ], + "angle": 0, + "content": "[2] Yoonki Cho, Woo Jae Kim, Seunghoon Hong, and Sung-Eui Yoon. Part-based pseudo label refinement for unsupervised person re-identification. In CVPR, pages 7308-7318, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.316, + 0.484, + 0.37 + ], + "angle": 0, + "content": "[3] Zuozhuo Dai, Guangyuan Wang, Weihao Yuan, Siyu Zhu, and Ping Tan. Cluster contrast for unsupervised person re-identification. In ACCV, pages 1142–1160, 2022. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.373, + 0.482, + 0.427 + ], + "angle": 0, + "content": "[4] Martin Ester, Hans-Peter Kriegel, Jörg Sander, Xiaowei Xu, et al. A density-based algorithm for discovering clusters in large spatial databases with noise. In KDD, pages 226-231, 1996. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.431, + 0.482, + 0.485 + ], + "angle": 0, + "content": "[5] Tiantian Gong, Guodong Du, Junsheng Wang, Yongkang Ding, and Liyan Zhang. Prototype-guided cross-modal completion and alignment for incomplete text-based person re-identification. In ACM MM, pages 5253-5261, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.488, + 0.482, + 0.541 + ], + "angle": 0, + "content": "[6] Tiantian Gong, Junsheng Wang, and Liyan Zhang. Cross-modal semantic aligning and neighbor-aware completing for robust text-image person retrieval. Information Fusion, 112: 102544, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.545, + 0.482, + 0.599 + ], + "angle": 0, + "content": "[7] Xinqian Gu, Hong Chang, Bingpeng Ma, Shutao Bai, Shiguang Shan, and Xilin Chen. Clothes-changing person re-identification with rgb modality only. In CVPR, pages 1060–1069, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.602, + 0.482, + 0.643 + ], + "angle": 0, + "content": "[8] Ke Han, Shaogang Gong, Yan Huang, Liang Wang, and Ti-niu Tan. Clothing-change feature augmentation for person re-identification. In CVPR, pages 22066–22075, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.646, + 0.482, + 0.686 + ], + "angle": 0, + "content": "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.688, + 0.482, + 0.743 + ], + "angle": 0, + "content": "[10] Tao He, Leqi Shen, Yuchen Guo, Guiguang Ding, and Zhenhua Guo. Secret: Self-consistent pseudo label refinement for unsupervised domain adaptive person re-identification. In AAAI, pages 879–887, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.746, + 0.482, + 0.8 + ], + "angle": 0, + "content": "[11] Ziqi He, Mengjia Xue, Yunhao Du, Zhicheng Zhao, and Fei Su. Dynamic clustering and cluster contrastive learning for unsupervised person re-id with feature distribution alignment. In ICASSP, pages 3610-3614, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.802, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[12] Alexander Hermans, Lucas Beyer, and Bastian Leibe. In defense of the triplet loss for person re-identification. arXiv preprint arXiv:1703.07737, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[13] Peixian Hong, Tao Wu, Ancong Wu, Xintong Han, and Wei-Shi Zheng. Fine-grained shape-appearance mutual learning for cloth-changing person re-identification. In CVPR, pages 10513–10522, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.201, + 0.484, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[14] Yan Huang, Jingsong Xu, Qiang Wu, Yi Zhong, Peng Zhang, and Zhaoxiang Zhang. Beyond scalar neuron: Adopting vector-neuron capsules for long-term person re-identification. IEEE TCSVT, 30(10):3459-3471, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[15] Yan Huang, Qiang Wu, JingSong Xu, Yi Zhong, and ZhaoXiang Zhang. Clothing status awareness for long-term person re-identification. In ICCV, pages 11895–11904, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.193, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[16] Xin Jin, Tianyu He, Kecheng Zheng, Zhiheng Yin, Xu Shen, Zhen Huang, Ruoyu Feng, Jianqiang Huang, Zhibo Chen, and Xian-Sheng Hua. Cloth-changing person re-identification from a single image with gait prediction and regularization. In CVPR, pages 14278–14287, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.263, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[17] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.306, + 0.905, + 0.346 + ], + "angle": 0, + "content": "[18] Long Lan, Xiao Teng, Jing Zhang, Xiang Zhang, and Dacheng Tao. Learning to purification for unsupervised person re-identification. IEEE TIP, 32:3338-3353, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.349, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[19] Mingkun Li, Chun-Guang Li, and Jun Guo. Cluster-guided asymmetric contrastive learning for unsupervised person re-identification. IEEE TIP, 31:3606-3617, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.391, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[20] Mingkun Li, Peng Xu, Chun-Guang Li, and Jun Guo. SiCL: Silhouette-driven contrastive learning for unsupervised person re-identification with clothes change. arXiv preprint arXiv:2305.13600, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.448, + 0.905, + 0.488 + ], + "angle": 0, + "content": "[21] Peike Li, Yunqiu Xu, Yunchao Wei, and Yi Yang. Self-correction for human parsing. IEEE TPAMI, 44(6):3260-3271, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.491, + 0.905, + 0.543 + ], + "angle": 0, + "content": "[22] Siyuan Li, Li Sun, and Qingli Li. Clip-reid: exploiting vision-language model for image re-identification without concrete text labels. In AAAI, pages 1405–1413, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.547, + 0.905, + 0.601 + ], + "angle": 0, + "content": "[23] Shuang Li, Jiaxu Leng, Guozhang Li, Ji Gan, Xinbo Gao, et al. Clip-driven cloth-agnostic feature learning for cloth-changing person re-identification. arXiv preprint arXiv:2406.09198, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.604, + 0.905, + 0.671 + ], + "angle": 0, + "content": "[24] Weijia Li, Saihui Hou, Chunjie Zhang, Chunshui Cao, Xu Liu, Yongzhen Huang, and Yao Zhao. An in-depth exploration of person re-identification and gait recognition in cloth-changing conditions. In CVPR, pages 13824–13833, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.675, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[25] Yutian Lin, Xuanyi Dong, Liang Zheng, Yan Yan, and Yi Yang. A bottom-up clustering approach to unsupervised person re-identification. In AAAI, pages 8738–8745, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.718, + 0.905, + 0.757 + ], + "angle": 0, + "content": "[26] Fangyi Liu, Mang Ye, and Bo Du. Cloth-aware augmentation for cloth-generalized person re-identification. In ACM MM, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.76, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[27] Vuong D Nguyen, Pranav Mantini, and Shishir K Shah. Contrastive clothing and pose generation for cloth-changing person re-identification. In CVPR, pages 7541–7549, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.803, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[28] Zhiqi Pang, Lingling Zhao, Qiuyang Liu, and Chunyu Wang. Camera invariant feature learning for unsupervised person re-identification. IEEE TMM, 25:6171-6182, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[29] Zhiqi Pang, Lingling Zhao, and Chunyu Wang. Clothing-invariant contrastive learning for unsupervised person re-identification. Neural Networks, page 106477, 2024. 1, 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19259" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[30] Xuelin Qian, Wenxuan Wang, Li Zhang, Fangrui Zhu, Yanwei Fu, Tao Xiang, Yu-Gang Jiang, and Xiangyang Xue. Long-term cloth-changing person re-identification. In ACCV, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.151, + 0.482, + 0.22 + ], + "angle": 0, + "content": "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.223, + 0.482, + 0.292 + ], + "angle": 0, + "content": "[32] Jiangming Shi, Yachao Zhang, Xiangbo Yin, Yuan Xie, Zhizhong Zhang, Jianping Fan, Zhongchao Shi, and Yanyun Qu. Dual pseudo-labels interactive self-training for semi-supervised visible-infrared person re-identification. In ICCV, pages 11218-11228, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.295, + 0.482, + 0.335 + ], + "angle": 0, + "content": "[33] Fangbin Wan, Yang Wu, Xuelin Qian, Yixiong Chen, and Yanwei Fu. When person re-identification meets changing clothes. In CVPRW, pages 830–831, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.339, + 0.482, + 0.392 + ], + "angle": 0, + "content": "[34] Menglin Wang, Baisheng Lai, Jianqiang Huang, Xiaojin Gong, and Xian-Sheng Hua. Camera-aware proxies for unsupervised person re-identification. In AAAI, pages 2764–2772, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.397, + 0.482, + 0.45 + ], + "angle": 0, + "content": "[35] Qizao Wang, Xuelin Qian, Bin Li, Xiangyang Xue, and Yanwei Fu. Exploring fine-grained representation and decomposition for cloth-changing person re-identification. IEEE TIFS, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.454, + 0.482, + 0.494 + ], + "angle": 0, + "content": "[36] Shiyu Xuan and Shiliang Zhang. Intra-inter camera similarity for unsupervised person re-identification. In CVPR, pages 11926–11935, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.498, + 0.482, + 0.553 + ], + "angle": 0, + "content": "[37] Shuanglin Yan, Neng Dong, Jun Liu, Liyan Zhang, and Jinhui Tang. Learning comprehensive representations with richer self for text-to-image person re-identification. In ACM MM, pages 6202–6211, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.557, + 0.482, + 0.623 + ], + "angle": 0, + "content": "[38] Fengxiang Yang, Zhun Zhong, Zhiming Luo, Yuanzheng Cai, Yaojin Lin, Shaozi Li, and Nicu Sebe. Joint noisetolerant learning and meta camera shift adaptation for unsupervised person re-identification. In CVPR, pages 4855–4864, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.629, + 0.482, + 0.668 + ], + "angle": 0, + "content": "[39] Qize Yang, Ancong Wu, and Wei-Shi Zheng. Person re-identification by contour sketch under moderate clothing change. IEEE TPAMI, 43(6):2029-2046, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.672, + 0.482, + 0.726 + ], + "angle": 0, + "content": "[40] Zhengwei Yang, Meng Lin, Xian Zhong, Yu Wu, and Zheng Wang. Good is bad: Causality inspired cloth-debiasing for cloth-changing person re-identification. In CVPR, pages 1472–1481, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.73, + 0.482, + 0.784 + ], + "angle": 0, + "content": "[41] Mang Ye, Jianbing Shen, Gaojie Lin, Tao Xiang, Ling Shao, and Steven CH Hoi. Deep learning for person re-identification: A survey and outlook. IEEE TPAMI, 44(6): 2872-2893, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.788, + 0.482, + 0.828 + ], + "angle": 0, + "content": "[42] Junhui Yin, Xinyu Zhang, Zhanyu Ma, Jun Guo, and Yifan Liu. A real-time memory updating strategy for unsupervised person re-identification. IEEE TIP, 32:2309–2321, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[43] Xiangbo Yin, Jiangming Shi, Yachao Zhang, Yang Lu, Zhizhong Zhang, Yuan Xie, and Yanyun Qu. Robust pseudo-label learning with neighbor relation for unsupervised visible-infrared person re-identification. In ACM MM, pages 2242–2251, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[44] Kaiwei Zeng, Munan Ning, Yaohua Wang, and Yang Guo. Hierarchical clustering with hard-batch triplet loss for person re-identification. In CVPR, pages 13657–13665, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[45] Xiao Zhang, Yixiao Ge, Yu Qiao, and Hongsheng Li. Refining pseudo labels with clustering consensus over generations for unsupervised object re-identification. In CVPR, pages 3436–3445, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[46] Xinyu Zhang, Dongdong Li, Zhigang Wang, Jian Wang, Errui Ding, Javen Qinfeng Shi, Zhaoxiang Zhang, and Jingdong Wang. Implicit sample extension for unsupervised person re-identification. In CVPR, pages 7369-7378, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[47] Yukang Zhang and Hanzi Wang. Diverse embedding expansion network and low-light cross-modality benchmark for visible-infrared person re-identification. In CVPR, pages 2153–2162, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.306, + 0.905, + 0.359 + ], + "angle": 0, + "content": "[48] Yukang Zhang, Yan Yan, Jie Li, and Hanzi Wang. Mrcn: A novel modality restitution and compensation network for visible-infrared person re-identification. In AAAI, pages 3498–3506, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[49] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 130(9):2337-2348, 2022. 2, 4" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19260" + } + ] +] \ No newline at end of file diff --git a/2025/Identity-Clothing Similarity Modeling for Unsupervised Clothing Change Person Re-Identification/42258662-090f-4600-a2a7-5bf4fc16321d_origin.pdf b/2025/Identity-Clothing Similarity Modeling for Unsupervised Clothing Change Person Re-Identification/42258662-090f-4600-a2a7-5bf4fc16321d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..af63fc0e8d4a039b05fba3ea885da793add1f945 --- /dev/null +++ b/2025/Identity-Clothing Similarity Modeling for Unsupervised Clothing Change Person Re-Identification/42258662-090f-4600-a2a7-5bf4fc16321d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c32004588913ec5859fae6331dd8e940b6e91735153578cd87526d4fe38eb0c1 +size 1810314 diff --git a/2025/Identity-Clothing Similarity Modeling for Unsupervised Clothing Change Person Re-Identification/full.md b/2025/Identity-Clothing Similarity Modeling for Unsupervised Clothing Change Person Re-Identification/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2aa87ac0bd408895c0986588bcb2dfc97550b30f --- /dev/null +++ b/2025/Identity-Clothing Similarity Modeling for Unsupervised Clothing Change Person Re-Identification/full.md @@ -0,0 +1,367 @@ +# Identity-Clothing Similarity Modeling for Unsupervised Clothing Change Person Re-Identification + +Zhiqi Pang $^{1}$ Junjie Wang $^{2}$ Lingling Zhao $^{1}$ Chunyu Wang $^{1*}$ + +1 Harbin Institute of Technology, China + +$^{2}$ Nanjing Medical University, China + +zqpang98@gmail.com, junjiehit@163.com, zhaoll@hit.edu.cn, chunyu@hit.edu.cn + +# Abstract + +Clothing change person re-identification (CC-ReID) aims to match different images of the same person, even when the clothing varies across images. To reduce manual labeling costs, existing unsupervised CC-ReID methods employ clustering algorithms to generate pseudo-labels. However, they often fail to assign the same pseudo-label to two images with the same identity but different clothing—referred to as a clothing change positive pair—thus hindering clothing-invariant feature learning. To address this issue, we propose the identity-clothing similarity modeling (ICSM) framework. To effectively connect clothing change positive pairs, ICSM first performs clothing-aware learning to leverage all discriminative information, including clothing, to obtain compact clusters. It then extracts cluster-level identity and clothing features and performs inter-cluster similarity estimation to identify clothing change positive clusters, reliable negative clusters, and hard negative clusters for each compact cluster. During optimization, we design an adaptive version of existing optimization methods to enhance similarities of clothing change positive pairs, while also introducing text semantics as a supervisory signal to further promote clothing invariance. Extensive experimental results across multiple datasets validate the effectiveness of the proposed framework, demonstrating its superiority over existing unsupervised methods and its competitiveness with some supervised approaches. + +# 1. Introduction + +Person re-identification (ReID) [32, 41, 43, 47, 48] aims to match images of the same individual captured in multicamera scenarios. Compared to traditional ReID, clothing change ReID (CC-ReID) [7, 23, 24, 26, 27] assumes that individuals may change clothing and focuses on matching images of the same person with different outfits. Although + +![](images/b7b099dab8d6d5ccc079d2912428f6cd91294679fa533a817e61a32e2d6f7185.jpg) +(a) Original images + +![](images/f31bcb5e52e2ed94b5e7c21da8877061b20058b837b355e824e3113e0db2910f.jpg) +(b) Clothing-occluded images +Figure 1. Visualization of features extracted from original images and clothing-occluded images. Different colors represent features with different identities, while the same color but different shapes indicate features with the same identity but different clothing. Image pairs that are correctly (incorrectly) clustered by clustering algorithms are connected with green (red) arrows. + +existing CC-ReID methods have achieved promising performance, they largely depend on large-scale manual annotations. However, manual labeling for retrieval tasks is an extremely labor-intensive process [1, 3]. Consequently, researchers have begun to focus on unsupervised CC-ReID (UCC-ReID) [20, 29]. In UCC-ReID, two images of the same person with different clothing—referred to as a clothing change positive pair—often exhibit significant appearance differences. This makes it challenging for clustering algorithms [4] to group their features into the same cluster, thus preventing them from obtaining the same pseudo-label. This phenomenon severely hinders the performance improvement of UCC-ReID methods. We refer to each clothing change positive pair that receives the same pseudo-label as a connected positive pair, while those that receive different pseudo-labels are termed unconnected positive pairs. + +To obtain robust pseudo-labels for clothing changes, existing UCC-ReID methods [20, 29] leverage body shapes extracted by human parsing models [21], which are typically clothing-invariant and identity-related. For example, CICL [29] clusters the fused features of original and body shape images. However, the unavoidable presence of clothing information in the fused features limits the number of connected positive pairs, while generating a large number + +of unconnected positive pairs. Additionally, existing UCC-ReID methods [20, 29] typically enhance intra-class compactness and inter-class separability indiscriminately during the optimization phase, similar to unsupervised traditional ReID methods [38, 46]. However, such optimization methods reduce the similarity of each unconnected positive pair, i.e., they decrease the feature similarity between images of the same person with different clothing, thus hindering clothing-invariance learning. + +To address the above issues, we propose the identityclothing similarity modeling (ICSM) framework. An intuitive idea for obtaining robust pseudo-labels for clothing changes is to cluster the features of clothing-occluded images. However, we find this approach to be unfeasible. Specifically, clothing information is not entirely detrimental in the process of obtaining accurate pseudo-labels. As shown in Fig. 1a, the features of original images with the same identity and clothing—referred to as a same clothing positive pair—demonstrate high compactness. In contrast, as depicted in Fig. 1b, the clothing-occluded images lose some clothing-related discriminative information, resulting in reduced compactness for same clothing positive pairs. This significantly hinders clustering algorithms [4] from assigning the same pseudo-label to each same clothing positive pair. To this end, ICSM first performs clothing-aware learning to utilize all discriminative information, including clothing, to obtain compact clusters. It then extracts cluster-level identity and clothing features using visual-language models [22, 31, 49], and finally introduces an inter-cluster similarity estimation method to integrate identity and clothing similarities, identifying clothing change positive clusters, reliable negative clusters, and hard negative clusters for each compact cluster. + +During the optimization phase, we enhance the model's robustness to clothing changes from both image semantics and text semantics. On one hand, we develop an adaptive version of existing optimization methods that focuses solely on reliable positive and negative pairs, thereby improving similarities of connected positive pairs while maintaining similarities of unconnected positive pairs. On the other hand, we introduce text semantics as a supervisory signal and design identity-oriented semantic contrastive loss and clothing-invariant semantic contrastive loss to further promote clothing invariance learning. + +The main contributions are summarized as follows: + +- We propose an identity-clothing similarity modeling (ICSM) framework, which aims to connect clothing change positive pairs and promote clothing invariance learning to address clothing variation. +- We develop an inter-cluster similarity estimation method that not only fully utilizes clothing-related discriminative information but also integrates identity and clothing similarities to identify clothing change positive clusters, reli + +able negative clusters, and hard negative clusters for each compact cluster. + +- We not only introduce an adaptive version of traditional optimization methods for UCC-ReID, but also design identity-oriented semantic contrastive loss and clothing-invariant semantic contrastive loss to leverage text semantics in promoting clothing invariance learning. +- Experimental results on multiple CC-ReID datasets validate the effectiveness of the proposed framework. ICSM not only outperforms existing UCC-ReID methods but also achieves competitive performance compared to earlier supervised methods. + +# 2. Related Work + +# 2.1. Unsupervised Traditional ReID + +Unsupervised traditional ReID (UT-ReID) methods [1, 3, 18, 45] assume that individuals do not change clothing, iterating between clustering and optimization. In early research, both BUC [25] and HCT [44] employ hierarchical clustering to obtain identity pseudo-labels, with HCT [44] also incorporating a batch-hard triplet loss [12] to address hard samples. Subsequent studies, such as CAP [34] and ICE [1], introduce camera invariance learning to mitigate the impact of varying camera conditions. In contrast to these approaches, IICS [36] and CIFL [28] focus on obtaining pseudo-labels robust to camera discrepancies. To alleviate the influence of noisy labels, both PPLR [2] and SE-CRET [10] assess the consistency between global and local features to enhance the overall reliability of training pseudolabels. Purification [18] supervises the student model using a teacher model to reduce the interference of noisy labels. Although existing UT-ReID methods face challenges in adapting to CC-ReID scenarios, their contributions to unsupervised learning provide valuable insights for our work. + +# 2.2. Clothing Change ReID + +Clothing change ReID (CC-ReID) methods [7, 13, 16] typically assumes that individuals may change their clothing and emphasizes the importance of clothing-invariant and identity-related features. Existing methods often introduce additional identity information (e.g., keypoints [30], shape [13], and gait [16]) or clothing information (e.g., status [15] and category [7]) to aid the optimization process, achieving promising performance. However, most existing methods rely heavily on manual identity labels or even fine-grained clothing labels, which significantly limits flexibility and scalability. To address these issues, researchers explore several unsupervised CC-ReID methods. For instance, SiCL [20] and CICL [29] focus on utilizing body shape to obtain pseudo-labels that are robust to clothing variations. However, the presence of numerous unconnected positive pairs severely restricts their performance, a factor they overlook. + +To address this challenge, we integrates identity and clothing similarities to reduce the number of unconnected positive pairs. Additionally, we limit the optimization scope to reliable positive and negative pairs, thereby maintaining similarities of unconnected positive pairs. + +# 3. Proposed Framework + +During the training phase, we are provided with an unlabeled training image set $\{x_{i}\}_{i = 1}^{N}$ , where $N$ denotes the number of images. Images from the same individual may exhibit different clothing. As illustrated in Fig. 2, our proposed identity-clothing similarity modeling (ICSM) consists of three stages. Specifically, stage I (clothing-aware learning) iterates through the following two processes: 1) clustering image features to generate pseudo-labels, and 2) optimizing the image encoder $E_{\mathrm{I}}$ based on the pseudolabels. The final pseudo-labels are saved, and the clusters corresponding to these pseudo-labels are referred to as compact clusters. Stage II (cluster-level feature learning) first extracts identity-related image set $\{x_{i}^{s}\}_{i = 1}^{N}$ and clothing-related image set $\{x_{i}^{c}\}_{i = 1}^{N}$ from $\{x_{i}\}_{i = 1}^{N}$ , and then introduces two CLIP models [31]: CLIP-S and CLIP-C. These models are used to learn cluster-level identity features and cluster-level clothing features, respectively, for each compact cluster. Stage III (clothing-invariant learning) first performs inter-cluster similarity estimation (ISE) to identify clothing change positive clusters, reliable negative clusters, and hard negative clusters for each compact cluster. Subsequently, we not only introduce an adaptive version of traditional optimization methods to help the image encoder $E_{\mathrm{III}}$ learn features that are identity-relevant and clothing-invariant, but also design identity-oriented semantic contrastive loss and clothing-invariant semantic contrastive loss to further promote clothing invariance learning. + +It is important to note that although ICSM involves a three-stage training process, the optimization in each stage is efficient and does not significantly increase the overall training time. Detailed analysis is provided in Sec. 4.3. + +# 3.1. Clothing-Aware Learning + +In this stage, we iteratively train the image encoder $E_{\mathrm{I}}$ following the traditional unsupervised ReID approaches [1, 3]. Specifically, in each iteration, we first use $E_{\mathrm{I}}$ to extract features from all training images. Then, we apply the DBSCAN algorithm [4] to all image features to obtain pseudolabels, and subsequently compute the centroid for each cluster. For example, the centroid of the $p$ -th cluster is defined as: + +$$ +v _ {p} = \frac {1}{N _ {p}} \sum_ {i = 1} ^ {N _ {p}} f _ {i}, \tag {1} +$$ + +where $N_{p}$ denotes the number of images in the $p$ -th cluster, and $f_{i}$ represents the features of the image $x_{i}$ within that + +cluster. Then, we introduce the identity classification loss and hard contrastive loss to optimize $E_{\mathrm{I}}$ . For any image $x_{i}$ , the identity classification loss is defined as: + +$$ +L _ {i d} = - \log \frac {\exp \left(f _ {i} \cdot v _ {p} ^ {\mathrm {T}} / \tau_ {i d}\right)}{\sum_ {q = 1} ^ {| V |} \exp \left(f _ {i} \cdot v _ {q} ^ {\mathrm {T}} / \tau_ {i d}\right)}, \tag {2} +$$ + +where $f_{i}$ is the feature of image $x_{i}$ , $v_{p}$ is the centroid of $f_{i}$ , $|V|$ is the number of centroids in the current iteration, and $\tau_{id}$ is the temperature hyperparameter. For any image $x_{i}$ , the hard contrastive loss is defined as: + +$$ +L _ {h c} = - \log \frac {\exp \left(f _ {i} \cdot f _ {p} ^ {\mathrm {T}} / \tau_ {h c}\right)}{\sum_ {f _ {q} \in F _ {i} ^ {n} \cup f _ {p}} \exp \left(f _ {i} \cdot f _ {q} ^ {\mathrm {T}} / \tau_ {h c}\right)}, \tag {3} +$$ + +where $f_{i}^{p}$ is the feature of the most distant positive sample from $f_{i}$ in a minibatch of size $P\times K$ (where $P$ is the number of sampled clusters and $K$ is the number of images sampled from each cluster), $F_{i}^{n}$ denotes the set of features corresponding to all negative samples of $f_{i}$ within the minibatch, and $\tau_{hc}$ is the temperature hyperparameter. In summary, the overall loss function of stage I is defined as: + +$$ +L _ {s 1} = L _ {i d} + L _ {h c}. \tag {4} +$$ + +The clustering and optimization process is iteratively performed, and the final pseudo-labels are saved. Due to the discriminative power brought by clothing, the clusters corresponding to these pseudo-labels exhibit high compactness, which we refer to as "compact clusters". In the following, the term "cluster" typically refers to these "compact clusters." The images within a cluster usually share not only the same identity but also the same clothing. + +# 3.2. Cluster-Level Feature Learning + +Due to the lack of connected positive pairs, it is challenging to improve the model's robustness to clothing changes based solely on compact clusters. Therefore, before training, we should attempt to connect two clusters that each contain images of the same person but with different clothing, i.e., a clothing change positive cluster pair. To this end, stage II focuses on obtaining cluster-level identity feature and cluster-level clothing feature for each compact cluster, thereby laying the foundation for connecting clothing change positive cluster pairs in the stage III. + +Stage II first obtains identity-related images and clothing-related images. As shown in Fig. 3, for any given original image $x_{i}$ , we first use a human parsing model [21] to obtain the parsed image $\hat{x_i}$ . Based on $\hat{x_i}$ , we then extract the identity-related image $x_{i}^{s}$ and the clothing-related image $x_{i}^{c}$ from the original image $x_{i}$ . Specifically, in $x_{i}^{s}$ , the pixel values of the clothing region are set to 0, while the pixel values of the other body regions are kept the same as in $x_{i}$ and the background pixel values are set to 255. In $x_{i}^{c}$ , the + +![](images/33ecfe168bbae0008ddc730268e2eb08017100b3bdc0c9c65847a34b12b0f790.jpg) + +![](images/904cb22b1a57f2a82b7a89e3effedf03524df5b9170f02563c73bce00c53bdf7.jpg) + +![](images/c26c17bd2533c59a7acbc07e77dbf9a3b622df04e846eb37c36c73208b6a5f09.jpg) +Figure 2. The proposed identity-clothing similarity modeling (ICSM) framework for UCC-ReID as a three-stage pipeline. The different colors of the clusters represent different pseudo-labels. The image encoders $E_{\mathrm{I}}$ and $E_{\mathrm{III}}$ are optimized, while all image and text encoders in the stage II are frozen. See description in the text for additional detail. + +![](images/376754699728b591620cd4deb53ebb733908dfbf42154a53564d402b87563afe.jpg) +Stage II (cluster-level feature learning) + +![](images/58b41685c572d3ba54d38c95d29a78245a2b5c0931382d4ea83c40590f6b80a7.jpg) +Figure 3. Example of obtaining identity-related image $x_{i}^{s}$ and clothing-related image $x_{i}^{c}$ . + +pixel values of the clothing region remain the same as in $x_{i}$ , while all other regions are set to 255. We refer to the clusters composed of identity-related images as identity-related clusters, and the clusters composed of clothing-related images as clothing-related clusters. + +Subsequently, stage II leverages the strong summarizing and descriptive ability of text for person images [5, 6, 37], along with the powerful text-image understanding capabilities of vision-language models [22, 31, 49], to obtain cluster-level features in text form. Specifically, we introduce two CLIP models [31]: CLIP-S and CLIP-C, which share the same pre-trained parameters. CLIP-S is used to learn cluster-level identity features from identity-related clusters, while CLIP-C is used to learn cluster-level clothing features from clothing-related clusters. + +For CLIP-S, we first set an identity-related text for each identity-related cluster as "A photo of a $[S_1][S_2]\ldots [S_M]$ person," where $\left[[S_1][S_2]\ldots [S_M]\right]$ represents a set of learnable text embeddings, and $M$ is the number of learnable text embeddings. For any identity-related image $x_{i}^{s}$ , we input it + +along with the corresponding identity-related text into the image encoder and text encoder of CLIP-S, respectively. This produces the image feature $f_{i}^{s}$ and text feature $t_{i}^{s}$ . Afterward, we freeze the image and text encoders of CLIP-S and optimize the learnable text embeddings, as follows: + +$$ +L ^ {s} = L _ {i t} ^ {s} + L _ {t i} ^ {s}, \tag {5} +$$ + +where $L_{it}^{s}$ and $L_{ti}^{s}$ are defined as: + +$$ +L _ {i t} ^ {s} = - \log \frac {\exp \left(f _ {i} ^ {s} \cdot t _ {i} ^ {s ^ {\mathrm {T}}}\right)}{\sum_ {j = 1} ^ {B} \exp \left(f _ {i} ^ {s} \cdot t _ {j} ^ {s ^ {\mathrm {T}}}\right)}, \tag {6} +$$ + +$$ +L _ {t i} ^ {s} = \frac {- 1}{| \psi_ {i} ^ {s} |} \sum_ {f _ {p} ^ {s} \in \psi_ {i} ^ {s}} \log \frac {\exp \left(t _ {i} ^ {s} \cdot f _ {p} ^ {s ^ {\mathrm {T}}}\right)}{\sum_ {j = 1} ^ {B} \exp \left(t _ {i} ^ {s} \cdot f _ {j} ^ {s ^ {\mathrm {T}}}\right)}, \tag {7} +$$ + +where $\psi_i^s$ represents the set of image features that share the same pseudo-label as $t_i^s$ , and $B$ is the batch size. + +Finally, we use the optimized text feature $t_i^s$ as the cluster-level identity feature for the corresponding compact cluster $u_i$ . Similarly, for CLIP-C, we first set a clothing-related text for each clothing-related cluster: "A photo of $[C_1][C_2]\ldots [C_M]$ clothing", where $"[C_1][C_2]\ldots [C_M]"$ represents a set of learnable text embeddings. Then, we follow the same optimization scheme as for $L_{it}^s$ and $L_{ti}^s$ to set $L_{it}^c$ and $L_{ti}^c$ for CLIP-C. Based on CLIP-C and the clothing-related text, we obtain the cluster-level clothing feature for each compact cluster. For example, the cluster-level clothing feature for the $i$ -th cluster $u_i$ is represented as $t_i^c$ . + +# 3.3. Clothing-Invariant Learning + +Inter-cluster similarity estimation. In theory, the higher the similarity between the cluster-level identity features of two compact clusters, the more likely it is that they contain images of the same person. Conversely, the lower the similarity between the cluster-level clothing features of two clusters, the more likely it is that they contain images with different clothing. Therefore, in stage III, we propose an inter-cluster similarity estimation (ISE) method, which seeks clothing change positive clusters, reliable negative clusters, and hard negative clusters for each cluster. + +For any cluster $u_{i}$ , ISE first calculates similarities between its cluster-level identity feature and clothing feature with those of other clusters, and then ranks all other clusters in descending order of similarity, resulting in the corresponding identity similarity sequence and clothing similarity sequence for $u_{i}$ . For any cluster $u_{i}$ , the set consisting of all clothing change positive clusters is defined as: + +$$ +U _ {i} ^ {c} = \left\{u _ {j} \mid t _ {i} ^ {s} \cdot t _ {j} ^ {s ^ {\mathrm {T}}} > \eta_ {a} \text {a n d} t _ {i} ^ {c} \cdot t _ {j} ^ {c ^ {\mathrm {T}}} \leq \eta_ {c} \right\}, \tag {8} +$$ + +where $\eta_{a}$ denotes the similarity value ranked at $\gamma_{a}$ in the identity similarity sequence, and $\eta_{c}$ denotes the similarity value ranked at $\gamma_{c}$ in the clothing similarity sequence. Additionally, for any cluster $u_{i}$ , we define the positive cluster set as $U_{i}^{a} = U_{i}^{c} \cup u_{i}$ . For any cluster $u_{i}$ , the set of its reliable negative clusters is defined as: + +$$ +U _ {i} ^ {b} = \left\{u _ {j} \mid t _ {i} ^ {s} \cdot t _ {j} ^ {s T} \leq \eta_ {b} \right\}, \tag {9} +$$ + +where $\eta_{b}$ denotes the similarity value ranked at $\gamma_{b}$ in the identity similarity sequence. For any cluster $u_{i}$ , we define the clusters with low cluster-level identity feature similarity and high cluster-level clothing feature similarity as hard negative clusters. Therefore, for any cluster $u_{i}$ , the set of its hard negative clusters is defined as: + +$$ +U _ {i} ^ {h} = \left\{u _ {j} \mid t _ {i} ^ {s} \cdot t _ {j} ^ {s T} \leq \eta_ {b} \text {a n d} t _ {i} ^ {c} \cdot t _ {j} ^ {c T} > \eta_ {c} \right\}, \tag {10} +$$ + +Finally, for any cluster $u_{i}$ , we save the reliable positive cluster set $U_{i}^{a}$ , clothing change positive cluster set $U_{i}^{c}$ , reliable negative cluster set $U_{i}^{b}$ , and hard negative cluster set $U_{i}^{h}$ . + +Image semantic contrastive learning. Due to the difficulty in fully connecting all clothing change positive cluster pairs, there may be additional clothing change positive clusters beyond the clothing change positive cluster set $U_{i}^{c}$ for any cluster $u_{i}$ . In this scenario, traditional optimization methods [1, 3] aimed at increasing inter-class separability would reduce the similarities between unconnected positive pairs, thereby hindering clothing-invariant learning. To address this, we design an adaptive version of the traditional optimization method for the UCC-ReID scenario. Specifically, for any image feature $f_{i}$ , the adaptive identity classification + +loss is defined as: + +$$ +L _ {a i d} = \frac {- 1}{\left| V _ {i} ^ {a} \right|} \sum_ {v _ {p} \in V _ {i} ^ {a}} \log \frac {\exp \left(f _ {i} \cdot v _ {p} ^ {\mathrm {T}} / \tau_ {i d}\right)}{\sum_ {v _ {q} \in V _ {i} ^ {b} \cup v _ {p}} \exp \left(f _ {i} \cdot v _ {q} ^ {\mathrm {T}} / \tau_ {i d}\right)}, \tag {11} +$$ + +where $V_{i}^{a}$ and $V_{i}^{b}$ respectively represent the set of centroids of all clusters in $U_{i}^{a}$ and $U_{i}^{b}$ corresponding to $f_{i}$ . For any image feature $f_{i}$ , the adaptive hard contrastive loss is defined as: + +$$ +L _ {a h c} = \frac {- 1}{\left| V _ {i} ^ {c} \right|} \sum_ {v _ {p} \in V _ {i} ^ {c}} \log \frac {\exp \left(f _ {i} \cdot v _ {p} ^ {\mathrm {T}} / \tau_ {h c}\right)}{\sum_ {v _ {q} \in V _ {i} ^ {h} \cup v _ {p}} \exp \left(f _ {i} \cdot v _ {q} ^ {\mathrm {T}} / \tau_ {h c}\right)}, \tag {12} +$$ + +where $V_{i}^{c}$ and $V_{i}^{h}$ respectively represent the set of centroids of all clusters in $U_{i}^{c}$ and $U_{i}^{h}$ corresponding to $f_{i}$ . Since the above process is optimized only for reliable clusters, it maintains the similarities of unconnected positive pairs. + +Text semantic contrastive learning. To fully leverage identity-related and clothing-related text semantics to assist model optimization, we introduce the identity-oriented semantic contrastive loss and clothing-invariant semantic contrastive loss. Firstly, for each image $x_{i}$ , we compute the similarity distribution $\alpha_{i}$ between the cluster-level identity feature $t_{i}^{s}$ of $x_{i}$ and all other cluster-level identity features. The $k$ -th element of $\alpha_{i}$ , denoted as $\alpha_{i}^{(k)}$ , is defined as: + +$$ +\alpha_ {i} ^ {(k)} = \frac {\exp \left(t _ {i} ^ {s} \cdot t _ {k} ^ {s T} / \tau_ {i s}\right)}{\sum t _ {j} ^ {s} \in T _ {i} ^ {s}} \exp \left(t _ {i} ^ {s} \cdot t _ {j} ^ {s T} / \tau_ {i s}\right), \tag {13} +$$ + +where $T_{i}^{s}$ represents the set of all cluster-level identity features except for $t_{i}^{s}$ , and $\tau_{is}$ is the temperature hyperparameter. Next, we compute the similarity distribution $\beta_{i}$ between the image feature $f_{i}$ of $x_{i}$ and the centroids of all other clusters. The $k$ -th element of $\beta_{i}$ , denoted as $\beta_{i}^{(k)}$ , is defined as: + +$$ +\beta_ {i} ^ {(k)} = \frac {\exp \left(f _ {i} \cdot v _ {k} ^ {\mathrm {T}} / \tau_ {i s}\right)}{\sum_ {v _ {j} \in V _ {i}} \exp \left(f _ {i} \cdot v _ {j} ^ {\mathrm {T}} / \tau_ {i s}\right)}, \tag {14} +$$ + +where $V_{i}$ represents the set of all centroids except for the centroid of $x_{i}$ . In theory, if all image features exhibit high robustness to clothing changes, then $\beta_{i}$ should exhibit high similarity to $\alpha_{i}$ . Therefore, to leverage identity-related text semantics to ensure high robustness of image features to clothing changes, we introduce the identity-oriented semantic contrastive loss: + +$$ +L _ {i s} = \sum_ {k = 1} ^ {| V | - 1} \alpha_ {i} ^ {(k)} \log \left(\alpha_ {i} ^ {(k)} / \beta_ {i} ^ {(k)}\right), \tag {15} +$$ + +where $|V|$ is the number of centroids in the current iteration. Additionally, to further promote clothing-invariant learning by leveraging clothing-related text semantics, we introduce the clothing-invariant semantic contrastive loss: + +$$ +L _ {c s} = \left[ f _ {i} \cdot t _ {p} ^ {c \mathrm {T}} - \min _ {t _ {q} ^ {c} \in T _ {i} ^ {a}} \left(f _ {i} \cdot t _ {q} ^ {c \mathrm {T}}\right) \right] _ {+}, \tag {16} +$$ + +where $t_p^c$ represents the cluster-level clothing feature of image feature $f_{i}$ , and $T_{i}^{a}$ is the set of cluster-level clothing features of all clusters in $U_{i}^{a}$ corresponding to $f_{i}$ . + +In summary, the overall loss function of stage III is defined as: + +$$ +L _ {s 3} = L _ {a i d} + L _ {a h c} + \lambda_ {i s} L _ {i s} + \lambda_ {c s} L _ {c s}, \tag {17} +$$ + +where $\lambda_{is}$ and $\lambda_{cs}$ are the weighting coefficients for $L_{is}$ and $L_{cs}$ , respectively. + +# 4. Experiment + +# 4.1. Datasets and Evaluation Metrics + +We evaluate the proposed framework on the PRCC [39], VC-Clothes [33], LTCC [30], Celeb-reID [14], and Celeb-reID-light [14] datasets, with statistical results presented in Tab. 1. Following existing works [20, 29], we employ three testing schemes: (1) clothing change (CC): the gallery contains only images with clothing changes; (2) same clothing (SC): the gallery includes only images with the same clothing; (3) general (G): the gallery consists of images with both clothing changes and the same clothing. For the PRCC [39] dataset, we utilize the CC and SC testing schemes; for VC-Clothes [33] and LTCC [30], we use the CC and G schemes; and for Celeb-reID [14] and Celeb-reID-light [14], we implement the CC testing scheme. For each testing scheme, we assess recognition performance using mean average precision (mAP) and cumulative matching characteristic (CMC). + +# 4.2. Implementation Details + +All images are resized to $256\times 128$ pixels. During the training phase, random horizontal flipping and random cropping are employed for data augmentation. All image encoders utilize ResNet-50 [9] pre-trained on CLIP [31] as the backbone. For DBSCAN [4], the distance threshold and minimum sample size are set to 0.55 and 4, respectively. The pseudo-labels from stage I are saved and passed on to stage II and stage III, where no further clustering or pseudo-label updates are performed. For each set of learnable text embeddings, we set $M = 4$ . For ISE, we set $\gamma_{a} = 2\%$ , $\gamma_{c} = 50\%$ , and $\gamma_{b} = 60\%$ . For stage I and stage III, we configure each minibatch to contain $P = 16$ identities, with each identity comprising $K = 4$ images. For stage II, we set $B = 64$ . For temperature hyperparameters, we set $\tau_{id} = 0.5$ , $\tau_{hc} = 0.1$ , and $\tau_{is} = 0.5$ . The model is optimized using the Adam optimizer [17], with an initial learning rate set to 0.00035. Both stage I and stage III consist of 50 epochs. In stage II, the learnable text embeddings in CLIP-S and CLIP-C are optimized for 120 epochs, respectively. During the testing phase, only the optimized image encoder $E_{\mathrm{III}}$ is used for inference. + +
DatasetsIdentitiesTrainingQueryGallery
PRCC22117,8963,54312,259
VC-Clothes2569,4491,0208,591
LTCC1529,5764937,050
Celeb-reID105220,2082,97211,006
Celeb-reID-light5909,021887934
+ +Table 1. Statistical results of PRCC, VC-Clothes, LTCC, CelebreID, and CelebreID-light. + +# 4.3. Comparison with State-of-the-Arts + +We evaluate the proposed method on PRCC [39], VC-Clothes (VC) [33], LTCC [30], Celeb-reID (CR) [14], and Celeb-reID-light (CRL) [14]. As shown in Tab. 2, we first report the performance of seven existing UT-ReID methods, including ICE [1], CC [3], CACL [19], PPLR [2], RTMem [42], Purification [18], and DCCC [11]. ICE [1], CACL [19], and Purification [18] achieve superior performance in these testing scenarios. We find that ICSM significantly outperforms all of the aforementioned methods, particularly in clothing change scenarios. This validates the effectiveness of ICSM's unique design for addressing clothing change scenarios. + +Subsequently, we compare advanced UCC-ReID methods: SiCL [20] and CICL [29]. We observe that ICSM's overall performance surpasses these methods, particularly achieving significant advantages in PRCC(CC) and VC(CC). This can be attributed to two main factors: first, ICSM combines identity and clothing similarities to identify clothing change positive clusters, reliable negative clusters, and hard negative clusters for each compact cluster. Second, ICSM not only incorporates an adaptive version of traditional optimization methods to maintain similarities of unconnected positive pairs but also leverages text semantics to further enhance clothing invariance learning. + +Finally, we compare ICSM with existing supervised CC-ReID (SCC-ReID) methods, including RCSANet [15], GI-ReID [16], CAL [7], AIM [40], CCFA [8], and $\mathrm{FIRe}^2$ [35]. We observe that ICSM achieves competitive performance on PRCC(CC) compared to earlier SCC-ReID methods (RCSANet [15] and CAL [7]), and surpasses GI-ReID [16] on VC(CC) and LTCC(CC). Moreover, ICSM achieves an mAP of approximately $70\%$ compared to advanced methods (AIM [40], CCFA [8], and $\mathrm{FIRe}^2$ [35]) under the two testing settings of LTCC. These experimental results validate the superiority of ICSM and the potential of UCC-ReID. + +Additionally, we report the training and test times for SiCL [20], CICL [29], and ICSM on VC(CC). As shown in Tab. 3, the training time for ICSM is slightly higher than that of SiCL and significantly lower than that of CICL. This is because, despite ICSM's three-stage training, the optimization in each stage is efficient and does not significantly + +
MethodsReferencePRCC(CC)PRCC(SC)VC(CC)VC(G)LTCC(CC)LTCC(G)CRCRL
mAPR-1mAPR-1mAPR-1mAPR-1mAPR-1mAPR-1mAPR-1mAPR-1
UT-ReID
ICEICCV2145.532.695.793.328.531.451.870.110.116.322.644.04.940.75.07.1
CCACCV2246.334.494.490.225.731.045.162.46.07.411.017.03.432.83.24.4
CACLTIP2252.141.794.790.949.758.968.082.46.29.822.345.65.142.33.64.3
PPLRCVPR2251.440.091.787.423.132.547.768.14.44.86.011.24.841.34.36.2
RTMemTIP2347.235.893.692.737.546.159.775.35.87.518.743.15.039.84.96.8
PurificationTIP2352.341.995.293.047.656.866.881.68.915.623.347.65.242.64.97.0
DCCCICASSP2447.035.390.689.539.350.458.976.17.812.719.644.93.935.73.35.0
UCC-ReID
SiCLarXiv2455.443.296.293.863.971.777.287.910.120.727.657.65.645.45.610.5
CICLNN2455.345.096.795.369.372.876.787.313.025.729.660.36.749.37.812.5
ICSMOurs57.847.997.597.173.575.178.789.613.526.230.160.87.049.67.712.6
SCC-ReID
RCSANetICCV2148.650.297.2100.0--------11.955.616.729.5
GI-ReIDCVPR22-33.3--57.864.5--10.423.729.463.2----
CALCVPR2255.855.299.8100.081.781.487.292.9--------
AIMCVPR2358.357.999.9100.0----19.140.641.176.3----
CCFACVPR2358.461.298.799.6----22.145.342.575.8----
FIRe2TIFS2463.165.099.5100.0----19.144.639.975.918.264.0--
+ +Table 2. Comparison of ICSM with existing methods. Optimal performance under two unsupervised settings is highlighted in bold. "R-1" denotes "Rank-1". + +
MethodsmAPRank-1Training timeTest time
SiCL63.971.76798.2514.20
CICL69.372.810569.9314.16
ICSM73.575.17236.5113.97
+ +Table 3. Statistics on performance and efficiency (seconds). + +increase the overall training time. In contrast, while existing UCC-ReID methods conduct only one stage of training, they often need to optimize three neural network modules simultaneously, which can require even more training time than ICSM. Additionally, ICSM has a similar testing time compared to existing UCC-ReID methods. These results demonstrate that ICSM not only achieves superior performance but also maintains efficiency. + +# 4.4. Ablation Study + +In this section, the baseline we tested optimizes the image encoder using only $L_{id}$ and $L_{hc}$ from stage I. As shown in Tab. 4, M1 corresponds to the baseline, while M2 introduces ISE on top of M1. Additionally, to accommodate ISE, M2 replaces $L_{id}$ and $L_{hc}$ with $L_{aid}$ and $L_{ahc}$ , respectively. M3 and M4 introduce $L_{is}$ and $L_{cs}$ , respectively, building upon M2. M5, which represents ICSM, incorporates both $L_{is}$ and $L_{cs}$ on top of M2. + +Effectiveness of ISE. M2 introduces ISE on top of M1 to identify clothing change positive clusters, reliable negative clusters, and hard negative clusters for each compact cluster. As shown in Tab. 4, M2 achieves significant performance improvements over M1 in PRCC(CC), VC-Clothes(CC), and VC-Clothes(G). To further explore the nature of ISE, we visualize the features of several clusters. As shown + +![](images/9e6992493cfb051bf3fa3f9ba6cfabe5ef0eb7849803e1d43faa5d964fd0192e.jpg) +(a) M1 + +![](images/d4e3fff688845ba21e030142c68905ad5e153d93c28056e13c5ab214fecf8a6c.jpg) +(b) M2 + +![](images/4881f7b6374435d20b7cbc8a60ee3df331963f681f14c3a97083d1157924b2db.jpg) +Figure 4. Visualization of the clusters corresponding to M1 and M2. Clusters are labeled with black dashed lines. For any given cluster, the tail and head of the red arrow connect it to its clothing change positive cluster identified by ISE. +Figure 5. Feature maps of M1 and M2. Each triplet, from left to right, consists of the original image, the feature map of M1, and the feature map of M2. + +in Fig. 4, M2 is able to identify clothing change positive clusters based on M1, significantly reducing the distances between the positive cluster pairs. For the unconnected clothing change positive clusters, the distances are not significantly increased because $L_{aid}$ and $L_{ahc}$ typically prevent these clusters from being included in the optimization scope. This process undoubtedly enhances the model's robustness to clothing changes. As shown in Fig. 5, we observe that M2 indeed focuses more on clothing-invariant re + +
MethodsBaselineISE\(L_{is}\)\(L_{cs}\)PRCC(CC)PRCC(SC)VC-Clothes(CC)VC-Clothes(G)
mAPRank-1mAPRank-1mAPRank-1mAPRank-1
M1✓×××43.931.592.289.023.630.846.667.8
M2✓✓××53.643.095.192.857.664.969.983.5
M3✓✓✓×56.847.197.095.668.672.575.286.3
M4✓✓×✓55.344.695.393.263.869.272.785.0
M5/ICSM✓✓✓✓57.847.997.597.173.575.178.789.6
+ +Table 4. Results of the ablation study for the proposed method. + +![](images/43ff2dac8fdf843e1bc425dfbc051ccbe48b109e78f8ce5f334f1cde2ff467f2.jpg) +(a) M2 + +![](images/6f6af1072a563bb0999401f82de4fcf3dacf5e45f59fd42c935009a37e9b372f.jpg) +(b) M3 + +![](images/ad57bd73604052a7d00d7a56eaf6799e0f7b57096f0f6704e27c85aabe999d4c.jpg) +(c) M4 + +![](images/980d21180ca32db2b08bda720a8d72d048ba1ee391d1e3b1389ff64d45b95987.jpg) +(d) M5 + +gions such as the head and arms, or areas like shoes that are less frequently changed, compared to M1. The above results validate the effectiveness of ISE. + +Effectiveness of $L_{is}$ and $L_{cs}$ . As shown in Tab. 4, M3 and M4 introduce $L_{is}$ and $L_{cs}$ into M2, respectively, both achieving significant performance improvements. Additionally, we visualize the feature distance distributions of same clothing positive pairs and clothing change positive pairs for M2, M3, M4, and M5. As shown in Fig. 6, compared to M2, both M3 and M4 significantly reduce the feature distance of clothing change positive pairs while also increasing the overlap between the feature distance distributions of these two types. This validates that both $L_{is}$ and $L_{cs}$ can enhance model performance by promoting the learning of clothing-invariant features. Furthermore, we observe that M5 further reduces the feature distance of clothing change positive pairs compared to M3 and M4, achieving superior performance over all previously mentioned methods. + +# 4.5. Sensitivity Analysis + +As shown in Fig. 7, when $\lambda_{is} = 2$ and $\lambda_{cs} = 1$ , the model achieves optimal performance on both PRCC(CC) and VC-Clothes(CC). This confirms the generalizability of $\lambda_{is}$ and $\lambda_{cs}$ . We also observe that when $\lambda_{is}$ is set too high or too low, model performance deteriorates. This is mainly because a small $\lambda_{is}$ reduces the contribution of identity- + +![](images/951e7a7b2133b430221dbc9b99d72572eabbf49db71b58dea56907ec9d0e3beb.jpg) +(a) PRCC(CC) + +![](images/6cd129c725b65f8f567db020e8c0ba5e9ae107ff377bb42eee141c6c638fac32.jpg) +(b) VC-Clothes(CC) + +![](images/31b39396ff259a14bd491747ad000c2d21c829585af3c708aedba6f2042a418c.jpg) +Figure 6. Feature distance distributions of same clothing positive pairs and clothing change positive pairs for M2, M3, M4, and M5. +(a) PRCC(CC) +Figure 8. Impact of $(\gamma_{a},\gamma_{c})$ on performance. + +![](images/0496e457536f5e0396c8d355a28bda3261b1497c8d5c137246591961c255899e.jpg) +Figure 7. Impact of $(\lambda_{is},\lambda_{cs})$ on performance. +(b) VC-Clothes(CC) + +related text semantics, while a large $\lambda_{is}$ negatively impacts the effectiveness of $L_{aid}$ and $L_{ahc}$ . As shown in Fig. 8, when $\gamma_{a} = 2\%$ and $\gamma_{c} = 50\%$ , the model achieves optimal performance on both datasets. This not only validates the generalization capability of ISE but also verifies the model's robustness to $\gamma_{a}$ and $\gamma_{c}$ . + +# 5. Conclusion + +In this paper, we propose an identity-clothing similarity modeling (ICSM) framework for UCC-ReID. To facilitate clothing-invariant feature learning, we introduce intercluster similarity estimation (ISE), identity-oriented semantic contrastive loss, and clothing-invariant semantic contrastive loss. The results of the ablation study validate their effectiveness, demonstrating that ISE enhances the model's robustness to clothing changes by identifying reliable clothing change positive clusters, while identity-oriented semantic contrastive loss and clothing-invariant semantic contrastive loss further promote clothing invariance learning. Experimental results across eight testing scenarios validate the effectiveness and superiority of the proposed ICSM, which not only outperforms existing unsupervised methods but also achieves competitive performance compared to certain supervised methods. + +# Acknowledgment + +This work is supported by National Key Research and Development Program of China (Grant no. 2023YFC3305005). + +# References + +[1] Hao Chen, Benoit Lagadec, and Francois Bremond. ICE: Inter-instance contrastive encoding for unsupervised person re-identification. In ICCV, pages 14960-14969, 2021. 1, 2, 3, 5, 6 +[2] Yoonki Cho, Woo Jae Kim, Seunghoon Hong, and Sung-Eui Yoon. Part-based pseudo label refinement for unsupervised person re-identification. In CVPR, pages 7308-7318, 2022. 2, 6 +[3] Zuozhuo Dai, Guangyuan Wang, Weihao Yuan, Siyu Zhu, and Ping Tan. Cluster contrast for unsupervised person re-identification. In ACCV, pages 1142–1160, 2022. 1, 2, 3, 5, 6 +[4] Martin Ester, Hans-Peter Kriegel, Jörg Sander, Xiaowei Xu, et al. A density-based algorithm for discovering clusters in large spatial databases with noise. In KDD, pages 226-231, 1996. 1, 2, 3, 6 +[5] Tiantian Gong, Guodong Du, Junsheng Wang, Yongkang Ding, and Liyan Zhang. Prototype-guided cross-modal completion and alignment for incomplete text-based person re-identification. In ACM MM, pages 5253-5261, 2023. 4 +[6] Tiantian Gong, Junsheng Wang, and Liyan Zhang. Cross-modal semantic aligning and neighbor-aware completing for robust text-image person retrieval. Information Fusion, 112: 102544, 2024. 4 +[7] Xinqian Gu, Hong Chang, Bingpeng Ma, Shutao Bai, Shiguang Shan, and Xilin Chen. Clothes-changing person re-identification with rgb modality only. In CVPR, pages 1060–1069, 2022. 1, 2, 6 +[8] Ke Han, Shaogang Gong, Yan Huang, Liang Wang, and Ti-niu Tan. Clothing-change feature augmentation for person re-identification. In CVPR, pages 22066–22075, 2023. 6 +[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 6 +[10] Tao He, Leqi Shen, Yuchen Guo, Guiguang Ding, and Zhenhua Guo. Secret: Self-consistent pseudo label refinement for unsupervised domain adaptive person re-identification. In AAAI, pages 879–887, 2022. 2 +[11] Ziqi He, Mengjia Xue, Yunhao Du, Zhicheng Zhao, and Fei Su. Dynamic clustering and cluster contrastive learning for unsupervised person re-id with feature distribution alignment. In ICASSP, pages 3610-3614, 2024. 6 +[12] Alexander Hermans, Lucas Beyer, and Bastian Leibe. In defense of the triplet loss for person re-identification. arXiv preprint arXiv:1703.07737, 2017. 2 +[13] Peixian Hong, Tao Wu, Ancong Wu, Xintong Han, and Wei-Shi Zheng. Fine-grained shape-appearance mutual learning for cloth-changing person re-identification. In CVPR, pages 10513–10522, 2021. 2 + +[14] Yan Huang, Jingsong Xu, Qiang Wu, Yi Zhong, Peng Zhang, and Zhaoxiang Zhang. Beyond scalar neuron: Adopting vector-neuron capsules for long-term person re-identification. IEEE TCSVT, 30(10):3459-3471, 2019. 6 +[15] Yan Huang, Qiang Wu, JingSong Xu, Yi Zhong, and ZhaoXiang Zhang. Clothing status awareness for long-term person re-identification. In ICCV, pages 11895–11904, 2021. 2, 6 +[16] Xin Jin, Tianyu He, Kecheng Zheng, Zhiheng Yin, Xu Shen, Zhen Huang, Ruoyu Feng, Jianqiang Huang, Zhibo Chen, and Xian-Sheng Hua. Cloth-changing person re-identification from a single image with gait prediction and regularization. In CVPR, pages 14278–14287, 2022. 2, 6 +[17] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[18] Long Lan, Xiao Teng, Jing Zhang, Xiang Zhang, and Dacheng Tao. Learning to purification for unsupervised person re-identification. IEEE TIP, 32:3338-3353, 2023. 2, 6 +[19] Mingkun Li, Chun-Guang Li, and Jun Guo. Cluster-guided asymmetric contrastive learning for unsupervised person re-identification. IEEE TIP, 31:3606-3617, 2022. 6 +[20] Mingkun Li, Peng Xu, Chun-Guang Li, and Jun Guo. SiCL: Silhouette-driven contrastive learning for unsupervised person re-identification with clothes change. arXiv preprint arXiv:2305.13600, 2023. 1, 2, 6 +[21] Peike Li, Yunqiu Xu, Yunchao Wei, and Yi Yang. Self-correction for human parsing. IEEE TPAMI, 44(6):3260-3271, 2020. 1, 3 +[22] Siyuan Li, Li Sun, and Qingli Li. Clip-reid: exploiting vision-language model for image re-identification without concrete text labels. In AAAI, pages 1405–1413, 2023. 2, 4 +[23] Shuang Li, Jiaxu Leng, Guozhang Li, Ji Gan, Xinbo Gao, et al. Clip-driven cloth-agnostic feature learning for cloth-changing person re-identification. arXiv preprint arXiv:2406.09198, 2024. 1 +[24] Weijia Li, Saihui Hou, Chunjie Zhang, Chunshui Cao, Xu Liu, Yongzhen Huang, and Yao Zhao. An in-depth exploration of person re-identification and gait recognition in cloth-changing conditions. In CVPR, pages 13824–13833, 2023. 1 +[25] Yutian Lin, Xuanyi Dong, Liang Zheng, Yan Yan, and Yi Yang. A bottom-up clustering approach to unsupervised person re-identification. In AAAI, pages 8738–8745, 2019. 2 +[26] Fangyi Liu, Mang Ye, and Bo Du. Cloth-aware augmentation for cloth-generalized person re-identification. In ACM MM, 2024. 1 +[27] Vuong D Nguyen, Pranav Mantini, and Shishir K Shah. Contrastive clothing and pose generation for cloth-changing person re-identification. In CVPR, pages 7541–7549, 2024. 1 +[28] Zhiqi Pang, Lingling Zhao, Qiuyang Liu, and Chunyu Wang. Camera invariant feature learning for unsupervised person re-identification. IEEE TMM, 25:6171-6182, 2022. 2 +[29] Zhiqi Pang, Lingling Zhao, and Chunyu Wang. Clothing-invariant contrastive learning for unsupervised person re-identification. Neural Networks, page 106477, 2024. 1, 2, 6 + +[30] Xuelin Qian, Wenxuan Wang, Li Zhang, Fangrui Zhu, Yanwei Fu, Tao Xiang, Yu-Gang Jiang, and Xiangyang Xue. Long-term cloth-changing person re-identification. In ACCV, 2020. 2, 6 +[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 2, 3, 4, 6 +[32] Jiangming Shi, Yachao Zhang, Xiangbo Yin, Yuan Xie, Zhizhong Zhang, Jianping Fan, Zhongchao Shi, and Yanyun Qu. Dual pseudo-labels interactive self-training for semi-supervised visible-infrared person re-identification. In ICCV, pages 11218-11228, 2023. 1 +[33] Fangbin Wan, Yang Wu, Xuelin Qian, Yixiong Chen, and Yanwei Fu. When person re-identification meets changing clothes. In CVPRW, pages 830–831, 2020. 6 +[34] Menglin Wang, Baisheng Lai, Jianqiang Huang, Xiaojin Gong, and Xian-Sheng Hua. Camera-aware proxies for unsupervised person re-identification. In AAAI, pages 2764–2772, 2021. 2 +[35] Qizao Wang, Xuelin Qian, Bin Li, Xiangyang Xue, and Yanwei Fu. Exploring fine-grained representation and decomposition for cloth-changing person re-identification. IEEE TIFS, 2024. 6 +[36] Shiyu Xuan and Shiliang Zhang. Intra-inter camera similarity for unsupervised person re-identification. In CVPR, pages 11926–11935, 2021. 2 +[37] Shuanglin Yan, Neng Dong, Jun Liu, Liyan Zhang, and Jinhui Tang. Learning comprehensive representations with richer self for text-to-image person re-identification. In ACM MM, pages 6202–6211, 2023. 4 +[38] Fengxiang Yang, Zhun Zhong, Zhiming Luo, Yuanzheng Cai, Yaojin Lin, Shaozi Li, and Nicu Sebe. Joint noisetolerant learning and meta camera shift adaptation for unsupervised person re-identification. In CVPR, pages 4855–4864, 2021. 2 +[39] Qize Yang, Ancong Wu, and Wei-Shi Zheng. Person re-identification by contour sketch under moderate clothing change. IEEE TPAMI, 43(6):2029-2046, 2019. 6 +[40] Zhengwei Yang, Meng Lin, Xian Zhong, Yu Wu, and Zheng Wang. Good is bad: Causality inspired cloth-debiasing for cloth-changing person re-identification. In CVPR, pages 1472–1481, 2023. 6 +[41] Mang Ye, Jianbing Shen, Gaojie Lin, Tao Xiang, Ling Shao, and Steven CH Hoi. Deep learning for person re-identification: A survey and outlook. IEEE TPAMI, 44(6): 2872-2893, 2021. 1 +[42] Junhui Yin, Xinyu Zhang, Zhanyu Ma, Jun Guo, and Yifan Liu. A real-time memory updating strategy for unsupervised person re-identification. IEEE TIP, 32:2309–2321, 2023. 6 +[43] Xiangbo Yin, Jiangming Shi, Yachao Zhang, Yang Lu, Zhizhong Zhang, Yuan Xie, and Yanyun Qu. Robust pseudo-label learning with neighbor relation for unsupervised visible-infrared person re-identification. In ACM MM, pages 2242–2251, 2024. 1 + +[44] Kaiwei Zeng, Munan Ning, Yaohua Wang, and Yang Guo. Hierarchical clustering with hard-batch triplet loss for person re-identification. In CVPR, pages 13657–13665, 2020. 2 +[45] Xiao Zhang, Yixiao Ge, Yu Qiao, and Hongsheng Li. Refining pseudo labels with clustering consensus over generations for unsupervised object re-identification. In CVPR, pages 3436–3445, 2021. 2 +[46] Xinyu Zhang, Dongdong Li, Zhigang Wang, Jian Wang, Errui Ding, Javen Qinfeng Shi, Zhaoxiang Zhang, and Jingdong Wang. Implicit sample extension for unsupervised person re-identification. In CVPR, pages 7369-7378, 2022. 2 +[47] Yukang Zhang and Hanzi Wang. Diverse embedding expansion network and low-light cross-modality benchmark for visible-infrared person re-identification. In CVPR, pages 2153–2162, 2023. 1 +[48] Yukang Zhang, Yan Yan, Jie Li, and Hanzi Wang. Mrcn: A novel modality restitution and compensation network for visible-infrared person re-identification. In AAAI, pages 3498–3506, 2023. 1 +[49] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 130(9):2337-2348, 2022. 2, 4 \ No newline at end of file diff --git a/2025/Identity-Clothing Similarity Modeling for Unsupervised Clothing Change Person Re-Identification/images.zip b/2025/Identity-Clothing Similarity Modeling for Unsupervised Clothing Change Person Re-Identification/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..00ad9c3902b03fb7f28bf4189e3fa70442f18b0e --- /dev/null +++ b/2025/Identity-Clothing Similarity Modeling for Unsupervised Clothing Change Person Re-Identification/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2f3b79a7aecd7e2b3a6c88039dbece01efaaef67f9dcba3db7ba328f2e7ba893 +size 576564 diff --git a/2025/Identity-Clothing Similarity Modeling for Unsupervised Clothing Change Person Re-Identification/layout.json b/2025/Identity-Clothing Similarity Modeling for Unsupervised Clothing Change Person Re-Identification/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..873ca737fa073d6836dc327ea19fa274c7c956db --- /dev/null +++ b/2025/Identity-Clothing Similarity Modeling for Unsupervised Clothing Change Person Re-Identification/layout.json @@ -0,0 +1,11187 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 56, + 102, + 555, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 102, + 555, + 138 + ], + "spans": [ + { + "bbox": [ + 56, + 102, + 555, + 138 + ], + "type": "text", + "content": "Identity-Clothing Similarity Modeling for Unsupervised Clothing Change Person Re-Identification" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 160, + 480, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 160, + 480, + 176 + ], + "spans": [ + { + "bbox": [ + 133, + 160, + 480, + 176 + ], + "type": "text", + "content": "Zhiqi Pang" + }, + { + "bbox": [ + 133, + 160, + 480, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 133, + 160, + 480, + 176 + ], + "type": "text", + "content": " Junjie Wang" + }, + { + "bbox": [ + 133, + 160, + 480, + 176 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 133, + 160, + 480, + 176 + ], + "type": "text", + "content": " Lingling Zhao" + }, + { + "bbox": [ + 133, + 160, + 480, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 133, + 160, + 480, + 176 + ], + "type": "text", + "content": " Chunyu Wang" + }, + { + "bbox": [ + 133, + 160, + 480, + 176 + ], + "type": "inline_equation", + "content": "^{1*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 210, + 176, + 400, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 210, + 176, + 400, + 189 + ], + "spans": [ + { + "bbox": [ + 210, + 176, + 400, + 189 + ], + "type": "text", + "content": "1 Harbin Institute of Technology, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 216, + 189, + 393, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 189, + 393, + 203 + ], + "spans": [ + { + "bbox": [ + 216, + 189, + 393, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 216, + 189, + 393, + 203 + ], + "type": "text", + "content": " Nanjing Medical University, China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 101, + 205, + 509, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 205, + 509, + 217 + ], + "spans": [ + { + "bbox": [ + 101, + 205, + 509, + 217 + ], + "type": "text", + "content": "zqpang98@gmail.com, junjiehit@163.com, zhaoll@hit.edu.cn, chunyu@hit.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 244, + 200, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 244, + 200, + 257 + ], + "spans": [ + { + "bbox": [ + 151, + 244, + 200, + 257 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 270, + 297, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 270, + 297, + 582 + ], + "spans": [ + { + "bbox": [ + 54, + 270, + 297, + 582 + ], + "type": "text", + "content": "Clothing change person re-identification (CC-ReID) aims to match different images of the same person, even when the clothing varies across images. To reduce manual labeling costs, existing unsupervised CC-ReID methods employ clustering algorithms to generate pseudo-labels. However, they often fail to assign the same pseudo-label to two images with the same identity but different clothing—referred to as a clothing change positive pair—thus hindering clothing-invariant feature learning. To address this issue, we propose the identity-clothing similarity modeling (ICSM) framework. To effectively connect clothing change positive pairs, ICSM first performs clothing-aware learning to leverage all discriminative information, including clothing, to obtain compact clusters. It then extracts cluster-level identity and clothing features and performs inter-cluster similarity estimation to identify clothing change positive clusters, reliable negative clusters, and hard negative clusters for each compact cluster. During optimization, we design an adaptive version of existing optimization methods to enhance similarities of clothing change positive pairs, while also introducing text semantics as a supervisory signal to further promote clothing invariance. Extensive experimental results across multiple datasets validate the effectiveness of the proposed framework, demonstrating its superiority over existing unsupervised methods and its competitiveness with some supervised approaches." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 604, + 135, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 604, + 135, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 604, + 135, + 616 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 623, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 623, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 623, + 296, + 696 + ], + "type": "text", + "content": "Person re-identification (ReID) [32, 41, 43, 47, 48] aims to match images of the same individual captured in multicamera scenarios. Compared to traditional ReID, clothing change ReID (CC-ReID) [7, 23, 24, 26, 27] assumes that individuals may change clothing and focuses on matching images of the same person with different outfits. Although" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 326, + 247, + 432, + 318 + ], + "blocks": [ + { + "bbox": [ + 326, + 247, + 432, + 318 + ], + "lines": [ + { + "bbox": [ + 326, + 247, + 432, + 318 + ], + "spans": [ + { + "bbox": [ + 326, + 247, + 432, + 318 + ], + "type": "image", + "image_path": "b7b099dab8d6d5ccc079d2912428f6cd91294679fa533a817e61a32e2d6f7185.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 319, + 409, + 329 + ], + "lines": [ + { + "bbox": [ + 345, + 319, + 409, + 329 + ], + "spans": [ + { + "bbox": [ + 345, + 319, + 409, + 329 + ], + "type": "text", + "content": "(a) Original images" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 436, + 246, + 544, + 318 + ], + "blocks": [ + { + "bbox": [ + 436, + 246, + 544, + 318 + ], + "lines": [ + { + "bbox": [ + 436, + 246, + 544, + 318 + ], + "spans": [ + { + "bbox": [ + 436, + 246, + 544, + 318 + ], + "type": "image", + "image_path": "f31bcb5e52e2ed94b5e7c21da8877061b20058b837b355e824e3113e0db2910f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 443, + 319, + 540, + 329 + ], + "lines": [ + { + "bbox": [ + 443, + 319, + 540, + 329 + ], + "spans": [ + { + "bbox": [ + 443, + 319, + 540, + 329 + ], + "type": "text", + "content": "(b) Clothing-occluded images" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 338, + 555, + 405 + ], + "lines": [ + { + "bbox": [ + 313, + 338, + 555, + 405 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 555, + 405 + ], + "type": "text", + "content": "Figure 1. Visualization of features extracted from original images and clothing-occluded images. Different colors represent features with different identities, while the same color but different shapes indicate features with the same identity but different clothing. Image pairs that are correctly (incorrectly) clustered by clustering algorithms are connected with green (red) arrows." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 426, + 555, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 617 + ], + "type": "text", + "content": "existing CC-ReID methods have achieved promising performance, they largely depend on large-scale manual annotations. However, manual labeling for retrieval tasks is an extremely labor-intensive process [1, 3]. Consequently, researchers have begun to focus on unsupervised CC-ReID (UCC-ReID) [20, 29]. In UCC-ReID, two images of the same person with different clothing—referred to as a clothing change positive pair—often exhibit significant appearance differences. This makes it challenging for clustering algorithms [4] to group their features into the same cluster, thus preventing them from obtaining the same pseudo-label. This phenomenon severely hinders the performance improvement of UCC-ReID methods. We refer to each clothing change positive pair that receives the same pseudo-label as a connected positive pair, while those that receive different pseudo-labels are termed unconnected positive pairs." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "text", + "content": "To obtain robust pseudo-labels for clothing changes, existing UCC-ReID methods [20, 29] leverage body shapes extracted by human parsing models [21], which are typically clothing-invariant and identity-related. For example, CICL [29] clusters the fused features of original and body shape images. However, the unavoidable presence of clothing information in the fused features limits the number of connected positive pairs, while generating a large number" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 194, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 194, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 194, + 713 + ], + "type": "text", + "content": "*Corresponding author: Chunyu Wang" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19251" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "content": "of unconnected positive pairs. Additionally, existing UCC-ReID methods [20, 29] typically enhance intra-class compactness and inter-class separability indiscriminately during the optimization phase, similar to unsupervised traditional ReID methods [38, 46]. However, such optimization methods reduce the similarity of each unconnected positive pair, i.e., they decrease the feature similarity between images of the same person with different clothing, thus hindering clothing-invariance learning." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 182, + 296, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 182, + 296, + 469 + ], + "spans": [ + { + "bbox": [ + 56, + 182, + 296, + 469 + ], + "type": "text", + "content": "To address the above issues, we propose the identityclothing similarity modeling (ICSM) framework. An intuitive idea for obtaining robust pseudo-labels for clothing changes is to cluster the features of clothing-occluded images. However, we find this approach to be unfeasible. Specifically, clothing information is not entirely detrimental in the process of obtaining accurate pseudo-labels. As shown in Fig. 1a, the features of original images with the same identity and clothing—referred to as a same clothing positive pair—demonstrate high compactness. In contrast, as depicted in Fig. 1b, the clothing-occluded images lose some clothing-related discriminative information, resulting in reduced compactness for same clothing positive pairs. This significantly hinders clustering algorithms [4] from assigning the same pseudo-label to each same clothing positive pair. To this end, ICSM first performs clothing-aware learning to utilize all discriminative information, including clothing, to obtain compact clusters. It then extracts cluster-level identity and clothing features using visual-language models [22, 31, 49], and finally introduces an inter-cluster similarity estimation method to integrate identity and clothing similarities, identifying clothing change positive clusters, reliable negative clusters, and hard negative clusters for each compact cluster." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 471, + 295, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 471, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 471, + 295, + 602 + ], + "type": "text", + "content": "During the optimization phase, we enhance the model's robustness to clothing changes from both image semantics and text semantics. On one hand, we develop an adaptive version of existing optimization methods that focuses solely on reliable positive and negative pairs, thereby improving similarities of connected positive pairs while maintaining similarities of unconnected positive pairs. On the other hand, we introduce text semantics as a supervisory signal and design identity-oriented semantic contrastive loss and clothing-invariant semantic contrastive loss to further promote clothing invariance learning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 604, + 276, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 604, + 276, + 615 + ], + "spans": [ + { + "bbox": [ + 67, + 604, + 276, + 615 + ], + "type": "text", + "content": "The main contributions are summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 618, + 295, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 666 + ], + "type": "text", + "content": "- We propose an identity-clothing similarity modeling (ICSM) framework, which aims to connect clothing change positive pairs and promote clothing invariance learning to address clothing variation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "content": "- We develop an inter-cluster similarity estimation method that not only fully utilizes clothing-related discriminative information but also integrates identity and clothing similarities to identify clothing change positive clusters, reli" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 322, + 72, + 553, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 553, + 95 + ], + "type": "text", + "content": "able negative clusters, and hard negative clusters for each compact cluster." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 96, + 553, + 216 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 313, + 96, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 96, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 96, + 553, + 156 + ], + "type": "text", + "content": "- We not only introduce an adaptive version of traditional optimization methods for UCC-ReID, but also design identity-oriented semantic contrastive loss and clothing-invariant semantic contrastive loss to leverage text semantics in promoting clothing invariance learning." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 156, + 553, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 156, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 553, + 216 + ], + "type": "text", + "content": "- Experimental results on multiple CC-ReID datasets validate the effectiveness of the proposed framework. ICSM not only outperforms existing UCC-ReID methods but also achieves competitive performance compared to earlier supervised methods." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 228, + 400, + 241 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 228, + 400, + 241 + ], + "spans": [ + { + "bbox": [ + 314, + 228, + 400, + 241 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 249, + 484, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 249, + 484, + 262 + ], + "spans": [ + { + "bbox": [ + 314, + 249, + 484, + 262 + ], + "type": "text", + "content": "2.1. Unsupervised Traditional ReID" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 266, + 555, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 555, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 555, + 495 + ], + "type": "text", + "content": "Unsupervised traditional ReID (UT-ReID) methods [1, 3, 18, 45] assume that individuals do not change clothing, iterating between clustering and optimization. In early research, both BUC [25] and HCT [44] employ hierarchical clustering to obtain identity pseudo-labels, with HCT [44] also incorporating a batch-hard triplet loss [12] to address hard samples. Subsequent studies, such as CAP [34] and ICE [1], introduce camera invariance learning to mitigate the impact of varying camera conditions. In contrast to these approaches, IICS [36] and CIFL [28] focus on obtaining pseudo-labels robust to camera discrepancies. To alleviate the influence of noisy labels, both PPLR [2] and SE-CRET [10] assess the consistency between global and local features to enhance the overall reliability of training pseudolabels. Purification [18] supervises the student model using a teacher model to reduce the interference of noisy labels. Although existing UT-ReID methods face challenges in adapting to CC-ReID scenarios, their contributions to unsupervised learning provide valuable insights for our work." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 504, + 444, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 504, + 444, + 517 + ], + "spans": [ + { + "bbox": [ + 314, + 504, + 444, + 517 + ], + "type": "text", + "content": "2.2. Clothing Change ReID" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 521, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 714 + ], + "type": "text", + "content": "Clothing change ReID (CC-ReID) methods [7, 13, 16] typically assumes that individuals may change their clothing and emphasizes the importance of clothing-invariant and identity-related features. Existing methods often introduce additional identity information (e.g., keypoints [30], shape [13], and gait [16]) or clothing information (e.g., status [15] and category [7]) to aid the optimization process, achieving promising performance. However, most existing methods rely heavily on manual identity labels or even fine-grained clothing labels, which significantly limits flexibility and scalability. To address these issues, researchers explore several unsupervised CC-ReID methods. For instance, SiCL [20] and CICL [29] focus on utilizing body shape to obtain pseudo-labels that are robust to clothing variations. However, the presence of numerous unconnected positive pairs severely restricts their performance, a factor they overlook." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19252" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": "To address this challenge, we integrates identity and clothing similarities to reduce the number of unconnected positive pairs. Additionally, we limit the optimization scope to reliable positive and negative pairs, thereby maintaining similarities of unconnected positive pairs." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 143, + 182, + 156 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 143, + 182, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 143, + 182, + 156 + ], + "type": "text", + "content": "3. Proposed Framework" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 296, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 296, + 486 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 296, + 486 + ], + "type": "text", + "content": "During the training phase, we are provided with an unlabeled training image set " + }, + { + "bbox": [ + 56, + 163, + 296, + 486 + ], + "type": "inline_equation", + "content": "\\{x_{i}\\}_{i = 1}^{N}" + }, + { + "bbox": [ + 56, + 163, + 296, + 486 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 56, + 163, + 296, + 486 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 56, + 163, + 296, + 486 + ], + "type": "text", + "content": " denotes the number of images. Images from the same individual may exhibit different clothing. As illustrated in Fig. 2, our proposed identity-clothing similarity modeling (ICSM) consists of three stages. Specifically, stage I (clothing-aware learning) iterates through the following two processes: 1) clustering image features to generate pseudo-labels, and 2) optimizing the image encoder " + }, + { + "bbox": [ + 56, + 163, + 296, + 486 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{I}}" + }, + { + "bbox": [ + 56, + 163, + 296, + 486 + ], + "type": "text", + "content": " based on the pseudolabels. The final pseudo-labels are saved, and the clusters corresponding to these pseudo-labels are referred to as compact clusters. Stage II (cluster-level feature learning) first extracts identity-related image set " + }, + { + "bbox": [ + 56, + 163, + 296, + 486 + ], + "type": "inline_equation", + "content": "\\{x_{i}^{s}\\}_{i = 1}^{N}" + }, + { + "bbox": [ + 56, + 163, + 296, + 486 + ], + "type": "text", + "content": " and clothing-related image set " + }, + { + "bbox": [ + 56, + 163, + 296, + 486 + ], + "type": "inline_equation", + "content": "\\{x_{i}^{c}\\}_{i = 1}^{N}" + }, + { + "bbox": [ + 56, + 163, + 296, + 486 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 56, + 163, + 296, + 486 + ], + "type": "inline_equation", + "content": "\\{x_{i}\\}_{i = 1}^{N}" + }, + { + "bbox": [ + 56, + 163, + 296, + 486 + ], + "type": "text", + "content": ", and then introduces two CLIP models [31]: CLIP-S and CLIP-C. These models are used to learn cluster-level identity features and cluster-level clothing features, respectively, for each compact cluster. Stage III (clothing-invariant learning) first performs inter-cluster similarity estimation (ISE) to identify clothing change positive clusters, reliable negative clusters, and hard negative clusters for each compact cluster. Subsequently, we not only introduce an adaptive version of traditional optimization methods to help the image encoder " + }, + { + "bbox": [ + 56, + 163, + 296, + 486 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{III}}" + }, + { + "bbox": [ + 56, + 163, + 296, + 486 + ], + "type": "text", + "content": " learn features that are identity-relevant and clothing-invariant, but also design identity-oriented semantic contrastive loss and clothing-invariant semantic contrastive loss to further promote clothing invariance learning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 486, + 296, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 296, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 296, + 534 + ], + "type": "text", + "content": "It is important to note that although ICSM involves a three-stage training process, the optimization in each stage is efficient and does not significantly increase the overall training time. Detailed analysis is provided in Sec. 4.3." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 541, + 200, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 541, + 200, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 541, + 200, + 555 + ], + "type": "text", + "content": "3.1. Clothing-Aware Learning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 559, + 296, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 559, + 296, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 296, + 653 + ], + "type": "text", + "content": "In this stage, we iteratively train the image encoder " + }, + { + "bbox": [ + 55, + 559, + 296, + 653 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{I}}" + }, + { + "bbox": [ + 55, + 559, + 296, + 653 + ], + "type": "text", + "content": " following the traditional unsupervised ReID approaches [1, 3]. Specifically, in each iteration, we first use " + }, + { + "bbox": [ + 55, + 559, + 296, + 653 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{I}}" + }, + { + "bbox": [ + 55, + 559, + 296, + 653 + ], + "type": "text", + "content": " to extract features from all training images. Then, we apply the DBSCAN algorithm [4] to all image features to obtain pseudolabels, and subsequently compute the centroid for each cluster. For example, the centroid of the " + }, + { + "bbox": [ + 55, + 559, + 296, + 653 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 559, + 296, + 653 + ], + "type": "text", + "content": "-th cluster is defined as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 140, + 652, + 295, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 652, + 295, + 685 + ], + "spans": [ + { + "bbox": [ + 140, + 652, + 295, + 685 + ], + "type": "interline_equation", + "content": "v _ {p} = \\frac {1}{N _ {p}} \\sum_ {i = 1} ^ {N _ {p}} f _ {i}, \\tag {1}", + "image_path": "7130ebac2cb3dec84f6d732be26b195e364c5a3d7477c4831d6fa6dc77055531.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "N_{p}" + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": " denotes the number of images in the " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "-th cluster, and " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": " represents the features of the image " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": " within that" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": "cluster. Then, we introduce the identity classification loss and hard contrastive loss to optimize " + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{I}}" + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": ". For any image " + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": ", the identity classification loss is defined as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 355, + 116, + 555, + 148 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 116, + 555, + 148 + ], + "spans": [ + { + "bbox": [ + 355, + 116, + 555, + 148 + ], + "type": "interline_equation", + "content": "L _ {i d} = - \\log \\frac {\\exp \\left(f _ {i} \\cdot v _ {p} ^ {\\mathrm {T}} / \\tau_ {i d}\\right)}{\\sum_ {q = 1} ^ {| V |} \\exp \\left(f _ {i} \\cdot v _ {q} ^ {\\mathrm {T}} / \\tau_ {i d}\\right)}, \\tag {2}", + "image_path": "90cd2a45f7d861009d138c8bfc4dded91ad24be17bd36627d56117c17cc0019d.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 156, + 555, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 156, + 555, + 203 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 555, + 203 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 156, + 555, + 203 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 313, + 156, + 555, + 203 + ], + "type": "text", + "content": " is the feature of image " + }, + { + "bbox": [ + 313, + 156, + 555, + 203 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 313, + 156, + 555, + 203 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 156, + 555, + 203 + ], + "type": "inline_equation", + "content": "v_{p}" + }, + { + "bbox": [ + 313, + 156, + 555, + 203 + ], + "type": "text", + "content": " is the centroid of " + }, + { + "bbox": [ + 313, + 156, + 555, + 203 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 313, + 156, + 555, + 203 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 156, + 555, + 203 + ], + "type": "inline_equation", + "content": "|V|" + }, + { + "bbox": [ + 313, + 156, + 555, + 203 + ], + "type": "text", + "content": " is the number of centroids in the current iteration, and " + }, + { + "bbox": [ + 313, + 156, + 555, + 203 + ], + "type": "inline_equation", + "content": "\\tau_{id}" + }, + { + "bbox": [ + 313, + 156, + 555, + 203 + ], + "type": "text", + "content": " is the temperature hyperparameter. For any image " + }, + { + "bbox": [ + 313, + 156, + 555, + 203 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 313, + 156, + 555, + 203 + ], + "type": "text", + "content": ", the hard contrastive loss is defined as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 342, + 209, + 555, + 243 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 209, + 555, + 243 + ], + "spans": [ + { + "bbox": [ + 342, + 209, + 555, + 243 + ], + "type": "interline_equation", + "content": "L _ {h c} = - \\log \\frac {\\exp \\left(f _ {i} \\cdot f _ {p} ^ {\\mathrm {T}} / \\tau_ {h c}\\right)}{\\sum_ {f _ {q} \\in F _ {i} ^ {n} \\cup f _ {p}} \\exp \\left(f _ {i} \\cdot f _ {q} ^ {\\mathrm {T}} / \\tau_ {h c}\\right)}, \\tag {3}", + "image_path": "2808bf572ce3ed456093d2ad2773b0706bedd52c30aabbe744f5060100ac83ba.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 250, + 554, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 250, + 554, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 250, + 554, + 335 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 250, + 554, + 335 + ], + "type": "inline_equation", + "content": "f_{i}^{p}" + }, + { + "bbox": [ + 313, + 250, + 554, + 335 + ], + "type": "text", + "content": " is the feature of the most distant positive sample from " + }, + { + "bbox": [ + 313, + 250, + 554, + 335 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 313, + 250, + 554, + 335 + ], + "type": "text", + "content": " in a minibatch of size " + }, + { + "bbox": [ + 313, + 250, + 554, + 335 + ], + "type": "inline_equation", + "content": "P\\times K" + }, + { + "bbox": [ + 313, + 250, + 554, + 335 + ], + "type": "text", + "content": " (where " + }, + { + "bbox": [ + 313, + 250, + 554, + 335 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 250, + 554, + 335 + ], + "type": "text", + "content": " is the number of sampled clusters and " + }, + { + "bbox": [ + 313, + 250, + 554, + 335 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 250, + 554, + 335 + ], + "type": "text", + "content": " is the number of images sampled from each cluster), " + }, + { + "bbox": [ + 313, + 250, + 554, + 335 + ], + "type": "inline_equation", + "content": "F_{i}^{n}" + }, + { + "bbox": [ + 313, + 250, + 554, + 335 + ], + "type": "text", + "content": " denotes the set of features corresponding to all negative samples of " + }, + { + "bbox": [ + 313, + 250, + 554, + 335 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 313, + 250, + 554, + 335 + ], + "type": "text", + "content": " within the minibatch, and " + }, + { + "bbox": [ + 313, + 250, + 554, + 335 + ], + "type": "inline_equation", + "content": "\\tau_{hc}" + }, + { + "bbox": [ + 313, + 250, + 554, + 335 + ], + "type": "text", + "content": " is the temperature hyperparameter. In summary, the overall loss function of stage I is defined as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 396, + 344, + 555, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 396, + 344, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 396, + 344, + 555, + 357 + ], + "type": "interline_equation", + "content": "L _ {s 1} = L _ {i d} + L _ {h c}. \\tag {4}", + "image_path": "b480faa9ba8ec199452d8e23ef00ab89b9c342f522ea3304735be5a3a1681725.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 365, + 555, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 555, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 555, + 462 + ], + "type": "text", + "content": "The clustering and optimization process is iteratively performed, and the final pseudo-labels are saved. Due to the discriminative power brought by clothing, the clusters corresponding to these pseudo-labels exhibit high compactness, which we refer to as \"compact clusters\". In the following, the term \"cluster\" typically refers to these \"compact clusters.\" The images within a cluster usually share not only the same identity but also the same clothing." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 468, + 486, + 481 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 486, + 481 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 486, + 481 + ], + "type": "text", + "content": "3.2. Cluster-Level Feature Learning" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 486, + 554, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 554, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 554, + 605 + ], + "type": "text", + "content": "Due to the lack of connected positive pairs, it is challenging to improve the model's robustness to clothing changes based solely on compact clusters. Therefore, before training, we should attempt to connect two clusters that each contain images of the same person but with different clothing, i.e., a clothing change positive cluster pair. To this end, stage II focuses on obtaining cluster-level identity feature and cluster-level clothing feature for each compact cluster, thereby laying the foundation for connecting clothing change positive cluster pairs in the stage III." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "text", + "content": "Stage II first obtains identity-related images and clothing-related images. As shown in Fig. 3, for any given original image " + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "text", + "content": ", we first use a human parsing model [21] to obtain the parsed image " + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\hat{x_i}" + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "text", + "content": ". Based on " + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\hat{x_i}" + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "text", + "content": ", we then extract the identity-related image " + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "inline_equation", + "content": "x_{i}^{s}" + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "text", + "content": " and the clothing-related image " + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "inline_equation", + "content": "x_{i}^{c}" + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "text", + "content": " from the original image " + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "text", + "content": ". Specifically, in " + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "inline_equation", + "content": "x_{i}^{s}" + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "text", + "content": ", the pixel values of the clothing region are set to 0, while the pixel values of the other body regions are kept the same as in " + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "text", + "content": " and the background pixel values are set to 255. In " + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "inline_equation", + "content": "x_{i}^{c}" + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "text", + "content": ", the" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19253" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 81, + 68, + 382, + 152 + ], + "blocks": [ + { + "bbox": [ + 81, + 68, + 382, + 152 + ], + "lines": [ + { + "bbox": [ + 81, + 68, + 382, + 152 + ], + "spans": [ + { + "bbox": [ + 81, + 68, + 382, + 152 + ], + "type": "image", + "image_path": "33ecfe168bbae0008ddc730268e2eb08017100b3bdc0c9c65847a34b12b0f790.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 81, + 153, + 382, + 262 + ], + "blocks": [ + { + "bbox": [ + 81, + 153, + 382, + 262 + ], + "lines": [ + { + "bbox": [ + 81, + 153, + 382, + 262 + ], + "spans": [ + { + "bbox": [ + 81, + 153, + 382, + 262 + ], + "type": "image", + "image_path": "904cb22b1a57f2a82b7a89e3effedf03524df5b9170f02563c73bce00c53bdf7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 82, + 264, + 382, + 306 + ], + "blocks": [ + { + "bbox": [ + 82, + 264, + 382, + 306 + ], + "lines": [ + { + "bbox": [ + 82, + 264, + 382, + 306 + ], + "spans": [ + { + "bbox": [ + 82, + 264, + 382, + 306 + ], + "type": "image", + "image_path": "c26c17bd2533c59a7acbc07e77dbf9a3b622df04e846eb37c36c73208b6a5f09.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 316, + 555, + 350 + ], + "lines": [ + { + "bbox": [ + 54, + 316, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 54, + 316, + 555, + 350 + ], + "type": "text", + "content": "Figure 2. The proposed identity-clothing similarity modeling (ICSM) framework for UCC-ReID as a three-stage pipeline. The different colors of the clusters represent different pseudo-labels. The image encoders " + }, + { + "bbox": [ + 54, + 316, + 555, + 350 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{I}}" + }, + { + "bbox": [ + 54, + 316, + 555, + 350 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 316, + 555, + 350 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{III}}" + }, + { + "bbox": [ + 54, + 316, + 555, + 350 + ], + "type": "text", + "content": " are optimized, while all image and text encoders in the stage II are frozen. See description in the text for additional detail." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 389, + 85, + 528, + 304 + ], + "blocks": [ + { + "bbox": [ + 389, + 70, + 528, + 85 + ], + "lines": [ + { + "bbox": [ + 389, + 70, + 528, + 85 + ], + "spans": [ + { + "bbox": [ + 389, + 70, + 528, + 85 + ], + "type": "text", + "content": "Stage II (cluster-level feature learning)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 389, + 85, + 528, + 304 + ], + "lines": [ + { + "bbox": [ + 389, + 85, + 528, + 304 + ], + "spans": [ + { + "bbox": [ + 389, + 85, + 528, + 304 + ], + "type": "image", + "image_path": "376754699728b591620cd4deb53ebb733908dfbf42154a53564d402b87563afe.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 82, + 368, + 271, + 418 + ], + "blocks": [ + { + "bbox": [ + 82, + 368, + 271, + 418 + ], + "lines": [ + { + "bbox": [ + 82, + 368, + 271, + 418 + ], + "spans": [ + { + "bbox": [ + 82, + 368, + 271, + 418 + ], + "type": "image", + "image_path": "58b41685c572d3ba54d38c95d29a78245a2b5c0931382d4ea83c40590f6b80a7.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 426, + 295, + 450 + ], + "lines": [ + { + "bbox": [ + 55, + 426, + 295, + 450 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 295, + 450 + ], + "type": "text", + "content": "Figure 3. Example of obtaining identity-related image " + }, + { + "bbox": [ + 55, + 426, + 295, + 450 + ], + "type": "inline_equation", + "content": "x_{i}^{s}" + }, + { + "bbox": [ + 55, + 426, + 295, + 450 + ], + "type": "text", + "content": " and clothing-related image " + }, + { + "bbox": [ + 55, + 426, + 295, + 450 + ], + "type": "inline_equation", + "content": "x_{i}^{c}" + }, + { + "bbox": [ + 55, + 426, + 295, + 450 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 472, + 295, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 295, + 531 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 295, + 531 + ], + "type": "text", + "content": "pixel values of the clothing region remain the same as in " + }, + { + "bbox": [ + 55, + 472, + 295, + 531 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 55, + 472, + 295, + 531 + ], + "type": "text", + "content": ", while all other regions are set to 255. We refer to the clusters composed of identity-related images as identity-related clusters, and the clusters composed of clothing-related images as clothing-related clusters." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 533, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 652 + ], + "type": "text", + "content": "Subsequently, stage II leverages the strong summarizing and descriptive ability of text for person images [5, 6, 37], along with the powerful text-image understanding capabilities of vision-language models [22, 31, 49], to obtain cluster-level features in text form. Specifically, we introduce two CLIP models [31]: CLIP-S and CLIP-C, which share the same pre-trained parameters. CLIP-S is used to learn cluster-level identity features from identity-related clusters, while CLIP-C is used to learn cluster-level clothing features from clothing-related clusters." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "For CLIP-S, we first set an identity-related text for each identity-related cluster as \"A photo of a " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "[S_1][S_2]\\ldots [S_M]" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": " person,\" where " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\left[[S_1][S_2]\\ldots [S_M]\\right]" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": " represents a set of learnable text embeddings, and " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": " is the number of learnable text embeddings. For any identity-related image " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "x_{i}^{s}" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": ", we input it" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 371, + 555, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 371, + 555, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 555, + 431 + ], + "type": "text", + "content": "along with the corresponding identity-related text into the image encoder and text encoder of CLIP-S, respectively. This produces the image feature " + }, + { + "bbox": [ + 313, + 371, + 555, + 431 + ], + "type": "inline_equation", + "content": "f_{i}^{s}" + }, + { + "bbox": [ + 313, + 371, + 555, + 431 + ], + "type": "text", + "content": " and text feature " + }, + { + "bbox": [ + 313, + 371, + 555, + 431 + ], + "type": "inline_equation", + "content": "t_{i}^{s}" + }, + { + "bbox": [ + 313, + 371, + 555, + 431 + ], + "type": "text", + "content": ". Afterward, we freeze the image and text encoders of CLIP-S and optimize the learnable text embeddings, as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 400, + 439, + 553, + 452 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 439, + 553, + 452 + ], + "spans": [ + { + "bbox": [ + 400, + 439, + 553, + 452 + ], + "type": "interline_equation", + "content": "L ^ {s} = L _ {i t} ^ {s} + L _ {t i} ^ {s}, \\tag {5}", + "image_path": "eb680b0e8905d2c5a24cb03e13b8a0145ece7c96c6dd4703d1adccb37538b443.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 460, + 451, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 451, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 451, + 473 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 460, + 451, + 473 + ], + "type": "inline_equation", + "content": "L_{it}^{s}" + }, + { + "bbox": [ + 313, + 460, + 451, + 473 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 460, + 451, + 473 + ], + "type": "inline_equation", + "content": "L_{ti}^{s}" + }, + { + "bbox": [ + 313, + 460, + 451, + 473 + ], + "type": "text", + "content": " are defined as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 364, + 479, + 553, + 512 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 479, + 553, + 512 + ], + "spans": [ + { + "bbox": [ + 364, + 479, + 553, + 512 + ], + "type": "interline_equation", + "content": "L _ {i t} ^ {s} = - \\log \\frac {\\exp \\left(f _ {i} ^ {s} \\cdot t _ {i} ^ {s ^ {\\mathrm {T}}}\\right)}{\\sum_ {j = 1} ^ {B} \\exp \\left(f _ {i} ^ {s} \\cdot t _ {j} ^ {s ^ {\\mathrm {T}}}\\right)}, \\tag {6}", + "image_path": "a517a76937c568e5fb049dc38efb193c742b93d6fff4844527dfd1160981f508.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 346, + 519, + 553, + 554 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 519, + 553, + 554 + ], + "spans": [ + { + "bbox": [ + 346, + 519, + 553, + 554 + ], + "type": "interline_equation", + "content": "L _ {t i} ^ {s} = \\frac {- 1}{| \\psi_ {i} ^ {s} |} \\sum_ {f _ {p} ^ {s} \\in \\psi_ {i} ^ {s}} \\log \\frac {\\exp \\left(t _ {i} ^ {s} \\cdot f _ {p} ^ {s ^ {\\mathrm {T}}}\\right)}{\\sum_ {j = 1} ^ {B} \\exp \\left(t _ {i} ^ {s} \\cdot f _ {j} ^ {s ^ {\\mathrm {T}}}\\right)}, \\tag {7}", + "image_path": "972eebdfd8eede6a19381253fa91b082f79b1ba7ee0bf3ca2df9ba177326b1aa.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 558, + 553, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 553, + 581 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 558, + 553, + 581 + ], + "type": "inline_equation", + "content": "\\psi_i^s" + }, + { + "bbox": [ + 313, + 558, + 553, + 581 + ], + "type": "text", + "content": " represents the set of image features that share the same pseudo-label as " + }, + { + "bbox": [ + 313, + 558, + 553, + 581 + ], + "type": "inline_equation", + "content": "t_i^s" + }, + { + "bbox": [ + 313, + 558, + 553, + 581 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 558, + 553, + 581 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 558, + 553, + 581 + ], + "type": "text", + "content": " is the batch size." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "content": "Finally, we use the optimized text feature " + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "inline_equation", + "content": "t_i^s" + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "content": " as the cluster-level identity feature for the corresponding compact cluster " + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "inline_equation", + "content": "u_i" + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "content": ". Similarly, for CLIP-C, we first set a clothing-related text for each clothing-related cluster: \"A photo of " + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "inline_equation", + "content": "[C_1][C_2]\\ldots [C_M]" + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "content": " clothing\", where " + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "inline_equation", + "content": "\"[C_1][C_2]\\ldots [C_M]\"" + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "content": " represents a set of learnable text embeddings. Then, we follow the same optimization scheme as for " + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "inline_equation", + "content": "L_{it}^s" + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "inline_equation", + "content": "L_{ti}^s" + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "content": " to set " + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "inline_equation", + "content": "L_{it}^c" + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "inline_equation", + "content": "L_{ti}^c" + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "content": " for CLIP-C. Based on CLIP-C and the clothing-related text, we obtain the cluster-level clothing feature for each compact cluster. For example, the cluster-level clothing feature for the " + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "content": "-th cluster " + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "inline_equation", + "content": "u_i" + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "content": " is represented as " + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "inline_equation", + "content": "t_i^c" + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19254" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 212, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 212, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 212, + 85 + ], + "type": "text", + "content": "3.3. Clothing-Invariant Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 296, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 296, + 209 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 296, + 209 + ], + "type": "text", + "content": "Inter-cluster similarity estimation. In theory, the higher the similarity between the cluster-level identity features of two compact clusters, the more likely it is that they contain images of the same person. Conversely, the lower the similarity between the cluster-level clothing features of two clusters, the more likely it is that they contain images with different clothing. Therefore, in stage III, we propose an inter-cluster similarity estimation (ISE) method, which seeks clothing change positive clusters, reliable negative clusters, and hard negative clusters for each cluster." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 211, + 296, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 211, + 296, + 294 + ], + "spans": [ + { + "bbox": [ + 55, + 211, + 296, + 294 + ], + "type": "text", + "content": "For any cluster " + }, + { + "bbox": [ + 55, + 211, + 296, + 294 + ], + "type": "inline_equation", + "content": "u_{i}" + }, + { + "bbox": [ + 55, + 211, + 296, + 294 + ], + "type": "text", + "content": ", ISE first calculates similarities between its cluster-level identity feature and clothing feature with those of other clusters, and then ranks all other clusters in descending order of similarity, resulting in the corresponding identity similarity sequence and clothing similarity sequence for " + }, + { + "bbox": [ + 55, + 211, + 296, + 294 + ], + "type": "inline_equation", + "content": "u_{i}" + }, + { + "bbox": [ + 55, + 211, + 296, + 294 + ], + "type": "text", + "content": ". For any cluster " + }, + { + "bbox": [ + 55, + 211, + 296, + 294 + ], + "type": "inline_equation", + "content": "u_{i}" + }, + { + "bbox": [ + 55, + 211, + 296, + 294 + ], + "type": "text", + "content": ", the set consisting of all clothing change positive clusters is defined as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 83, + 304, + 295, + 321 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 304, + 295, + 321 + ], + "spans": [ + { + "bbox": [ + 83, + 304, + 295, + 321 + ], + "type": "interline_equation", + "content": "U _ {i} ^ {c} = \\left\\{u _ {j} \\mid t _ {i} ^ {s} \\cdot t _ {j} ^ {s ^ {\\mathrm {T}}} > \\eta_ {a} \\text {a n d} t _ {i} ^ {c} \\cdot t _ {j} ^ {c ^ {\\mathrm {T}}} \\leq \\eta_ {c} \\right\\}, \\tag {8}", + "image_path": "cd7daacdfa4eb1e12f46170fb404ff18a0b427a00b8f250389651e55818efd58.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 330, + 296, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 330, + 296, + 401 + ], + "spans": [ + { + "bbox": [ + 55, + 330, + 296, + 401 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 330, + 296, + 401 + ], + "type": "inline_equation", + "content": "\\eta_{a}" + }, + { + "bbox": [ + 55, + 330, + 296, + 401 + ], + "type": "text", + "content": " denotes the similarity value ranked at " + }, + { + "bbox": [ + 55, + 330, + 296, + 401 + ], + "type": "inline_equation", + "content": "\\gamma_{a}" + }, + { + "bbox": [ + 55, + 330, + 296, + 401 + ], + "type": "text", + "content": " in the identity similarity sequence, and " + }, + { + "bbox": [ + 55, + 330, + 296, + 401 + ], + "type": "inline_equation", + "content": "\\eta_{c}" + }, + { + "bbox": [ + 55, + 330, + 296, + 401 + ], + "type": "text", + "content": " denotes the similarity value ranked at " + }, + { + "bbox": [ + 55, + 330, + 296, + 401 + ], + "type": "inline_equation", + "content": "\\gamma_{c}" + }, + { + "bbox": [ + 55, + 330, + 296, + 401 + ], + "type": "text", + "content": " in the clothing similarity sequence. Additionally, for any cluster " + }, + { + "bbox": [ + 55, + 330, + 296, + 401 + ], + "type": "inline_equation", + "content": "u_{i}" + }, + { + "bbox": [ + 55, + 330, + 296, + 401 + ], + "type": "text", + "content": ", we define the positive cluster set as " + }, + { + "bbox": [ + 55, + 330, + 296, + 401 + ], + "type": "inline_equation", + "content": "U_{i}^{a} = U_{i}^{c} \\cup u_{i}" + }, + { + "bbox": [ + 55, + 330, + 296, + 401 + ], + "type": "text", + "content": ". For any cluster " + }, + { + "bbox": [ + 55, + 330, + 296, + 401 + ], + "type": "inline_equation", + "content": "u_{i}" + }, + { + "bbox": [ + 55, + 330, + 296, + 401 + ], + "type": "text", + "content": ", the set of its reliable negative clusters is defined as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 123, + 411, + 295, + 429 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 411, + 295, + 429 + ], + "spans": [ + { + "bbox": [ + 123, + 411, + 295, + 429 + ], + "type": "interline_equation", + "content": "U _ {i} ^ {b} = \\left\\{u _ {j} \\mid t _ {i} ^ {s} \\cdot t _ {j} ^ {s T} \\leq \\eta_ {b} \\right\\}, \\tag {9}", + "image_path": "17f32adceb6cb085afb4a8de65be6ade0e4c9331463c52ee69fa076e88bf5d1c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 437, + 296, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 296, + 510 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 437, + 296, + 510 + ], + "type": "inline_equation", + "content": "\\eta_{b}" + }, + { + "bbox": [ + 55, + 437, + 296, + 510 + ], + "type": "text", + "content": " denotes the similarity value ranked at " + }, + { + "bbox": [ + 55, + 437, + 296, + 510 + ], + "type": "inline_equation", + "content": "\\gamma_{b}" + }, + { + "bbox": [ + 55, + 437, + 296, + 510 + ], + "type": "text", + "content": " in the identity similarity sequence. For any cluster " + }, + { + "bbox": [ + 55, + 437, + 296, + 510 + ], + "type": "inline_equation", + "content": "u_{i}" + }, + { + "bbox": [ + 55, + 437, + 296, + 510 + ], + "type": "text", + "content": ", we define the clusters with low cluster-level identity feature similarity and high cluster-level clothing feature similarity as hard negative clusters. Therefore, for any cluster " + }, + { + "bbox": [ + 55, + 437, + 296, + 510 + ], + "type": "inline_equation", + "content": "u_{i}" + }, + { + "bbox": [ + 55, + 437, + 296, + 510 + ], + "type": "text", + "content": ", the set of its hard negative clusters is defined as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 74, + 519, + 295, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 519, + 295, + 536 + ], + "spans": [ + { + "bbox": [ + 74, + 519, + 295, + 536 + ], + "type": "interline_equation", + "content": "U _ {i} ^ {h} = \\left\\{u _ {j} \\mid t _ {i} ^ {s} \\cdot t _ {j} ^ {s T} \\leq \\eta_ {b} \\text {a n d} t _ {i} ^ {c} \\cdot t _ {j} ^ {c T} > \\eta_ {c} \\right\\}, \\tag {10}", + "image_path": "d671e93b2c0f41a393b6edb6967795f830e48c80c917919786d29f76b2ab6d34.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 545, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 581 + ], + "type": "text", + "content": "Finally, for any cluster " + }, + { + "bbox": [ + 55, + 545, + 295, + 581 + ], + "type": "inline_equation", + "content": "u_{i}" + }, + { + "bbox": [ + 55, + 545, + 295, + 581 + ], + "type": "text", + "content": ", we save the reliable positive cluster set " + }, + { + "bbox": [ + 55, + 545, + 295, + 581 + ], + "type": "inline_equation", + "content": "U_{i}^{a}" + }, + { + "bbox": [ + 55, + 545, + 295, + 581 + ], + "type": "text", + "content": ", clothing change positive cluster set " + }, + { + "bbox": [ + 55, + 545, + 295, + 581 + ], + "type": "inline_equation", + "content": "U_{i}^{c}" + }, + { + "bbox": [ + 55, + 545, + 295, + 581 + ], + "type": "text", + "content": ", reliable negative cluster set " + }, + { + "bbox": [ + 55, + 545, + 295, + 581 + ], + "type": "inline_equation", + "content": "U_{i}^{b}" + }, + { + "bbox": [ + 55, + 545, + 295, + 581 + ], + "type": "text", + "content": ", and hard negative cluster set " + }, + { + "bbox": [ + 55, + 545, + 295, + 581 + ], + "type": "inline_equation", + "content": "U_{i}^{h}" + }, + { + "bbox": [ + 55, + 545, + 295, + 581 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": "Image semantic contrastive learning. Due to the difficulty in fully connecting all clothing change positive cluster pairs, there may be additional clothing change positive clusters beyond the clothing change positive cluster set " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "U_{i}^{c}" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": " for any cluster " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "u_{i}" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": ". In this scenario, traditional optimization methods [1, 3] aimed at increasing inter-class separability would reduce the similarities between unconnected positive pairs, thereby hindering clothing-invariant learning. To address this, we design an adaptive version of the traditional optimization method for the UCC-ReID scenario. Specifically, for any image feature " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": ", the adaptive identity classification" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 72, + 388, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 388, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 388, + 83 + ], + "type": "text", + "content": "loss is defined as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 321, + 87, + 553, + 129 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 87, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 321, + 87, + 553, + 129 + ], + "type": "interline_equation", + "content": "L _ {a i d} = \\frac {- 1}{\\left| V _ {i} ^ {a} \\right|} \\sum_ {v _ {p} \\in V _ {i} ^ {a}} \\log \\frac {\\exp \\left(f _ {i} \\cdot v _ {p} ^ {\\mathrm {T}} / \\tau_ {i d}\\right)}{\\sum_ {v _ {q} \\in V _ {i} ^ {b} \\cup v _ {p}} \\exp \\left(f _ {i} \\cdot v _ {q} ^ {\\mathrm {T}} / \\tau_ {i d}\\right)}, \\tag {11}", + "image_path": "9d0a0c17ae08982c81c86c421d3420cc98cb48caa507ea489c1c97a14d5ae83a.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 130, + 555, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 130, + 555, + 177 + ], + "spans": [ + { + "bbox": [ + 313, + 130, + 555, + 177 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 130, + 555, + 177 + ], + "type": "inline_equation", + "content": "V_{i}^{a}" + }, + { + "bbox": [ + 313, + 130, + 555, + 177 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 130, + 555, + 177 + ], + "type": "inline_equation", + "content": "V_{i}^{b}" + }, + { + "bbox": [ + 313, + 130, + 555, + 177 + ], + "type": "text", + "content": " respectively represent the set of centroids of all clusters in " + }, + { + "bbox": [ + 313, + 130, + 555, + 177 + ], + "type": "inline_equation", + "content": "U_{i}^{a}" + }, + { + "bbox": [ + 313, + 130, + 555, + 177 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 130, + 555, + 177 + ], + "type": "inline_equation", + "content": "U_{i}^{b}" + }, + { + "bbox": [ + 313, + 130, + 555, + 177 + ], + "type": "text", + "content": " corresponding to " + }, + { + "bbox": [ + 313, + 130, + 555, + 177 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 313, + 130, + 555, + 177 + ], + "type": "text", + "content": ". For any image feature " + }, + { + "bbox": [ + 313, + 130, + 555, + 177 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 313, + 130, + 555, + 177 + ], + "type": "text", + "content": ", the adaptive hard contrastive loss is defined as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 321, + 181, + 553, + 223 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 181, + 553, + 223 + ], + "spans": [ + { + "bbox": [ + 321, + 181, + 553, + 223 + ], + "type": "interline_equation", + "content": "L _ {a h c} = \\frac {- 1}{\\left| V _ {i} ^ {c} \\right|} \\sum_ {v _ {p} \\in V _ {i} ^ {c}} \\log \\frac {\\exp \\left(f _ {i} \\cdot v _ {p} ^ {\\mathrm {T}} / \\tau_ {h c}\\right)}{\\sum_ {v _ {q} \\in V _ {i} ^ {h} \\cup v _ {p}} \\exp \\left(f _ {i} \\cdot v _ {q} ^ {\\mathrm {T}} / \\tau_ {h c}\\right)}, \\tag {12}", + "image_path": "a2060a2b36a68e78f425835c98f18c46afadddf7cb63399e60c51529e619d80f.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 224, + 555, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 224, + 555, + 272 + ], + "spans": [ + { + "bbox": [ + 313, + 224, + 555, + 272 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 224, + 555, + 272 + ], + "type": "inline_equation", + "content": "V_{i}^{c}" + }, + { + "bbox": [ + 313, + 224, + 555, + 272 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 224, + 555, + 272 + ], + "type": "inline_equation", + "content": "V_{i}^{h}" + }, + { + "bbox": [ + 313, + 224, + 555, + 272 + ], + "type": "text", + "content": " respectively represent the set of centroids of all clusters in " + }, + { + "bbox": [ + 313, + 224, + 555, + 272 + ], + "type": "inline_equation", + "content": "U_{i}^{c}" + }, + { + "bbox": [ + 313, + 224, + 555, + 272 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 224, + 555, + 272 + ], + "type": "inline_equation", + "content": "U_{i}^{h}" + }, + { + "bbox": [ + 313, + 224, + 555, + 272 + ], + "type": "text", + "content": " corresponding to " + }, + { + "bbox": [ + 313, + 224, + 555, + 272 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 313, + 224, + 555, + 272 + ], + "type": "text", + "content": ". Since the above process is optimized only for reliable clusters, it maintains the similarities of unconnected positive pairs." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 272, + 554, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 272, + 554, + 370 + ], + "spans": [ + { + "bbox": [ + 313, + 272, + 554, + 370 + ], + "type": "text", + "content": "Text semantic contrastive learning. To fully leverage identity-related and clothing-related text semantics to assist model optimization, we introduce the identity-oriented semantic contrastive loss and clothing-invariant semantic contrastive loss. Firstly, for each image " + }, + { + "bbox": [ + 313, + 272, + 554, + 370 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 313, + 272, + 554, + 370 + ], + "type": "text", + "content": ", we compute the similarity distribution " + }, + { + "bbox": [ + 313, + 272, + 554, + 370 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 313, + 272, + 554, + 370 + ], + "type": "text", + "content": " between the cluster-level identity feature " + }, + { + "bbox": [ + 313, + 272, + 554, + 370 + ], + "type": "inline_equation", + "content": "t_{i}^{s}" + }, + { + "bbox": [ + 313, + 272, + 554, + 370 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 272, + 554, + 370 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 313, + 272, + 554, + 370 + ], + "type": "text", + "content": " and all other cluster-level identity features. The " + }, + { + "bbox": [ + 313, + 272, + 554, + 370 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 272, + 554, + 370 + ], + "type": "text", + "content": "-th element of " + }, + { + "bbox": [ + 313, + 272, + 554, + 370 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 313, + 272, + 554, + 370 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 313, + 272, + 554, + 370 + ], + "type": "inline_equation", + "content": "\\alpha_{i}^{(k)}" + }, + { + "bbox": [ + 313, + 272, + 554, + 370 + ], + "type": "text", + "content": ", is defined as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 362, + 376, + 553, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 376, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 362, + 376, + 553, + 407 + ], + "type": "interline_equation", + "content": "\\alpha_ {i} ^ {(k)} = \\frac {\\exp \\left(t _ {i} ^ {s} \\cdot t _ {k} ^ {s T} / \\tau_ {i s}\\right)}{\\sum t _ {j} ^ {s} \\in T _ {i} ^ {s}} \\exp \\left(t _ {i} ^ {s} \\cdot t _ {j} ^ {s T} / \\tau_ {i s}\\right), \\tag {13}", + "image_path": "6404876ca6028c1d0cb6bac446aaae0900f4fd51b30fe4699b6fe3a319b86c89.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 412, + 554, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 412, + 554, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 412, + 554, + 475 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 412, + 554, + 475 + ], + "type": "inline_equation", + "content": "T_{i}^{s}" + }, + { + "bbox": [ + 313, + 412, + 554, + 475 + ], + "type": "text", + "content": " represents the set of all cluster-level identity features except for " + }, + { + "bbox": [ + 313, + 412, + 554, + 475 + ], + "type": "inline_equation", + "content": "t_{i}^{s}" + }, + { + "bbox": [ + 313, + 412, + 554, + 475 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 412, + 554, + 475 + ], + "type": "inline_equation", + "content": "\\tau_{is}" + }, + { + "bbox": [ + 313, + 412, + 554, + 475 + ], + "type": "text", + "content": " is the temperature hyperparameter. Next, we compute the similarity distribution " + }, + { + "bbox": [ + 313, + 412, + 554, + 475 + ], + "type": "inline_equation", + "content": "\\beta_{i}" + }, + { + "bbox": [ + 313, + 412, + 554, + 475 + ], + "type": "text", + "content": " between the image feature " + }, + { + "bbox": [ + 313, + 412, + 554, + 475 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 313, + 412, + 554, + 475 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 412, + 554, + 475 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 313, + 412, + 554, + 475 + ], + "type": "text", + "content": " and the centroids of all other clusters. The " + }, + { + "bbox": [ + 313, + 412, + 554, + 475 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 412, + 554, + 475 + ], + "type": "text", + "content": "-th element of " + }, + { + "bbox": [ + 313, + 412, + 554, + 475 + ], + "type": "inline_equation", + "content": "\\beta_{i}" + }, + { + "bbox": [ + 313, + 412, + 554, + 475 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 313, + 412, + 554, + 475 + ], + "type": "inline_equation", + "content": "\\beta_{i}^{(k)}" + }, + { + "bbox": [ + 313, + 412, + 554, + 475 + ], + "type": "text", + "content": ", is defined as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 362, + 482, + 553, + 512 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 482, + 553, + 512 + ], + "spans": [ + { + "bbox": [ + 362, + 482, + 553, + 512 + ], + "type": "interline_equation", + "content": "\\beta_ {i} ^ {(k)} = \\frac {\\exp \\left(f _ {i} \\cdot v _ {k} ^ {\\mathrm {T}} / \\tau_ {i s}\\right)}{\\sum_ {v _ {j} \\in V _ {i}} \\exp \\left(f _ {i} \\cdot v _ {j} ^ {\\mathrm {T}} / \\tau_ {i s}\\right)}, \\tag {14}", + "image_path": "d92f77537139f38b3d00043e843197bf8d19b001c8f92279c5a277fa3e73b6d0.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 514, + 554, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 554, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 554, + 597 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 514, + 554, + 597 + ], + "type": "inline_equation", + "content": "V_{i}" + }, + { + "bbox": [ + 313, + 514, + 554, + 597 + ], + "type": "text", + "content": " represents the set of all centroids except for the centroid of " + }, + { + "bbox": [ + 313, + 514, + 554, + 597 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 313, + 514, + 554, + 597 + ], + "type": "text", + "content": ". In theory, if all image features exhibit high robustness to clothing changes, then " + }, + { + "bbox": [ + 313, + 514, + 554, + 597 + ], + "type": "inline_equation", + "content": "\\beta_{i}" + }, + { + "bbox": [ + 313, + 514, + 554, + 597 + ], + "type": "text", + "content": " should exhibit high similarity to " + }, + { + "bbox": [ + 313, + 514, + 554, + 597 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 313, + 514, + 554, + 597 + ], + "type": "text", + "content": ". Therefore, to leverage identity-related text semantics to ensure high robustness of image features to clothing changes, we introduce the identity-oriented semantic contrastive loss:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 366, + 601, + 553, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 601, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 366, + 601, + 553, + 635 + ], + "type": "interline_equation", + "content": "L _ {i s} = \\sum_ {k = 1} ^ {| V | - 1} \\alpha_ {i} ^ {(k)} \\log \\left(\\alpha_ {i} ^ {(k)} / \\beta_ {i} ^ {(k)}\\right), \\tag {15}", + "image_path": "cafce34eae5173482fe17105b10ffdda63d2e1e370852ebf42f808056a84aa5d.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 641, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 554, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 641, + 554, + 689 + ], + "type": "inline_equation", + "content": "|V|" + }, + { + "bbox": [ + 313, + 641, + 554, + 689 + ], + "type": "text", + "content": " is the number of centroids in the current iteration. Additionally, to further promote clothing-invariant learning by leveraging clothing-related text semantics, we introduce the clothing-invariant semantic contrastive loss:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 359, + 693, + 553, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 693, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 359, + 693, + 553, + 715 + ], + "type": "interline_equation", + "content": "L _ {c s} = \\left[ f _ {i} \\cdot t _ {p} ^ {c \\mathrm {T}} - \\min _ {t _ {q} ^ {c} \\in T _ {i} ^ {a}} \\left(f _ {i} \\cdot t _ {q} ^ {c \\mathrm {T}}\\right) \\right] _ {+}, \\tag {16}", + "image_path": "50913fad2ca6c740c74aebe4d698d7f6cc6502b3527714167b35544724344efb.jpg" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19255" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "inline_equation", + "content": "t_p^c" + }, + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "content": " represents the cluster-level clothing feature of image feature " + }, + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "inline_equation", + "content": "T_{i}^{a}" + }, + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "content": " is the set of cluster-level clothing features of all clusters in " + }, + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "inline_equation", + "content": "U_{i}^{a}" + }, + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "content": " corresponding to " + }, + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 110, + 295, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 110, + 295, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 110, + 295, + 133 + ], + "type": "text", + "content": "In summary, the overall loss function of stage III is defined as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 93, + 147, + 295, + 160 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 147, + 295, + 160 + ], + "spans": [ + { + "bbox": [ + 93, + 147, + 295, + 160 + ], + "type": "interline_equation", + "content": "L _ {s 3} = L _ {a i d} + L _ {a h c} + \\lambda_ {i s} L _ {i s} + \\lambda_ {c s} L _ {c s}, \\tag {17}", + "image_path": "be3b6e1dbff27e2edc4cce021f20f12ffb7da0cfc758aa310e9e516218eada4d.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 171, + 295, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 171, + 295, + 196 + ], + "spans": [ + { + "bbox": [ + 55, + 171, + 295, + 196 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 171, + 295, + 196 + ], + "type": "inline_equation", + "content": "\\lambda_{is}" + }, + { + "bbox": [ + 55, + 171, + 295, + 196 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 171, + 295, + 196 + ], + "type": "inline_equation", + "content": "\\lambda_{cs}" + }, + { + "bbox": [ + 55, + 171, + 295, + 196 + ], + "type": "text", + "content": " are the weighting coefficients for " + }, + { + "bbox": [ + 55, + 171, + 295, + 196 + ], + "type": "inline_equation", + "content": "L_{is}" + }, + { + "bbox": [ + 55, + 171, + 295, + 196 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 171, + 295, + 196 + ], + "type": "inline_equation", + "content": "L_{cs}" + }, + { + "bbox": [ + 55, + 171, + 295, + 196 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 211, + 132, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 211, + 132, + 224 + ], + "spans": [ + { + "bbox": [ + 55, + 211, + 132, + 224 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 232, + 231, + 244 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 232, + 231, + 244 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 231, + 244 + ], + "type": "text", + "content": "4.1. Datasets and Evaluation Metrics" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 251, + 296, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 296, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 296, + 443 + ], + "type": "text", + "content": "We evaluate the proposed framework on the PRCC [39], VC-Clothes [33], LTCC [30], Celeb-reID [14], and Celeb-reID-light [14] datasets, with statistical results presented in Tab. 1. Following existing works [20, 29], we employ three testing schemes: (1) clothing change (CC): the gallery contains only images with clothing changes; (2) same clothing (SC): the gallery includes only images with the same clothing; (3) general (G): the gallery consists of images with both clothing changes and the same clothing. For the PRCC [39] dataset, we utilize the CC and SC testing schemes; for VC-Clothes [33] and LTCC [30], we use the CC and G schemes; and for Celeb-reID [14] and Celeb-reID-light [14], we implement the CC testing scheme. For each testing scheme, we assess recognition performance using mean average precision (mAP) and cumulative matching characteristic (CMC)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 455, + 188, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 455, + 188, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 455, + 188, + 468 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "text", + "content": "All images are resized to " + }, + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "inline_equation", + "content": "256\\times 128" + }, + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "text", + "content": " pixels. During the training phase, random horizontal flipping and random cropping are employed for data augmentation. All image encoders utilize ResNet-50 [9] pre-trained on CLIP [31] as the backbone. For DBSCAN [4], the distance threshold and minimum sample size are set to 0.55 and 4, respectively. The pseudo-labels from stage I are saved and passed on to stage II and stage III, where no further clustering or pseudo-label updates are performed. For each set of learnable text embeddings, we set " + }, + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "inline_equation", + "content": "M = 4" + }, + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "text", + "content": ". For ISE, we set " + }, + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\gamma_{a} = 2\\%" + }, + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\gamma_{c} = 50\\%" + }, + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\gamma_{b} = 60\\%" + }, + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "text", + "content": ". For stage I and stage III, we configure each minibatch to contain " + }, + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "inline_equation", + "content": "P = 16" + }, + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "text", + "content": " identities, with each identity comprising " + }, + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "inline_equation", + "content": "K = 4" + }, + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "text", + "content": " images. For stage II, we set " + }, + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "inline_equation", + "content": "B = 64" + }, + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "text", + "content": ". For temperature hyperparameters, we set " + }, + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\tau_{id} = 0.5" + }, + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\tau_{hc} = 0.1" + }, + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\tau_{is} = 0.5" + }, + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "text", + "content": ". The model is optimized using the Adam optimizer [17], with an initial learning rate set to 0.00035. Both stage I and stage III consist of 50 epochs. In stage II, the learnable text embeddings in CLIP-S and CLIP-C are optimized for 120 epochs, respectively. During the testing phase, only the optimized image encoder " + }, + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{III}}" + }, + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "text", + "content": " is used for inference." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 324, + 70, + 545, + 133 + ], + "blocks": [ + { + "bbox": [ + 324, + 70, + 545, + 133 + ], + "lines": [ + { + "bbox": [ + 324, + 70, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 324, + 70, + 545, + 133 + ], + "type": "table", + "html": "
DatasetsIdentitiesTrainingQueryGallery
PRCC22117,8963,54312,259
VC-Clothes2569,4491,0208,591
LTCC1529,5764937,050
Celeb-reID105220,2082,97211,006
Celeb-reID-light5909,021887934
", + "image_path": "0bef2302128dc802ff052dd8e1f46029228523d2633c36a34b4f5292d40e8673.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 140, + 553, + 163 + ], + "lines": [ + { + "bbox": [ + 313, + 140, + 553, + 163 + ], + "spans": [ + { + "bbox": [ + 313, + 140, + 553, + 163 + ], + "type": "text", + "content": "Table 1. Statistical results of PRCC, VC-Clothes, LTCC, CelebreID, and CelebreID-light." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 186, + 500, + 199 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 186, + 500, + 199 + ], + "spans": [ + { + "bbox": [ + 313, + 186, + 500, + 199 + ], + "type": "text", + "content": "4.3. Comparison with State-of-the-Arts" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 205, + 555, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 555, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 555, + 348 + ], + "type": "text", + "content": "We evaluate the proposed method on PRCC [39], VC-Clothes (VC) [33], LTCC [30], Celeb-reID (CR) [14], and Celeb-reID-light (CRL) [14]. As shown in Tab. 2, we first report the performance of seven existing UT-ReID methods, including ICE [1], CC [3], CACL [19], PPLR [2], RTMem [42], Purification [18], and DCCC [11]. ICE [1], CACL [19], and Purification [18] achieve superior performance in these testing scenarios. We find that ICSM significantly outperforms all of the aforementioned methods, particularly in clothing change scenarios. This validates the effectiveness of ICSM's unique design for addressing clothing change scenarios." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 351, + 556, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 351, + 556, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 351, + 556, + 495 + ], + "type": "text", + "content": "Subsequently, we compare advanced UCC-ReID methods: SiCL [20] and CICL [29]. We observe that ICSM's overall performance surpasses these methods, particularly achieving significant advantages in PRCC(CC) and VC(CC). This can be attributed to two main factors: first, ICSM combines identity and clothing similarities to identify clothing change positive clusters, reliable negative clusters, and hard negative clusters for each compact cluster. Second, ICSM not only incorporates an adaptive version of traditional optimization methods to maintain similarities of unconnected positive pairs but also leverages text semantics to further enhance clothing invariance learning." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 496, + 556, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 496, + 556, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 496, + 556, + 639 + ], + "type": "text", + "content": "Finally, we compare ICSM with existing supervised CC-ReID (SCC-ReID) methods, including RCSANet [15], GI-ReID [16], CAL [7], AIM [40], CCFA [8], and " + }, + { + "bbox": [ + 313, + 496, + 556, + 639 + ], + "type": "inline_equation", + "content": "\\mathrm{FIRe}^2" + }, + { + "bbox": [ + 313, + 496, + 556, + 639 + ], + "type": "text", + "content": " [35]. We observe that ICSM achieves competitive performance on PRCC(CC) compared to earlier SCC-ReID methods (RCSANet [15] and CAL [7]), and surpasses GI-ReID [16] on VC(CC) and LTCC(CC). Moreover, ICSM achieves an mAP of approximately " + }, + { + "bbox": [ + 313, + 496, + 556, + 639 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 313, + 496, + 556, + 639 + ], + "type": "text", + "content": " compared to advanced methods (AIM [40], CCFA [8], and " + }, + { + "bbox": [ + 313, + 496, + 556, + 639 + ], + "type": "inline_equation", + "content": "\\mathrm{FIRe}^2" + }, + { + "bbox": [ + 313, + 496, + 556, + 639 + ], + "type": "text", + "content": " [35]) under the two testing settings of LTCC. These experimental results validate the superiority of ICSM and the potential of UCC-ReID." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "Additionally, we report the training and test times for SiCL [20], CICL [29], and ICSM on VC(CC). As shown in Tab. 3, the training time for ICSM is slightly higher than that of SiCL and significantly lower than that of CICL. This is because, despite ICSM's three-stage training, the optimization in each stage is efficient and does not significantly" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19256" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 67, + 69, + 545, + 281 + ], + "blocks": [ + { + "bbox": [ + 67, + 69, + 545, + 281 + ], + "lines": [ + { + "bbox": [ + 67, + 69, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 67, + 69, + 545, + 281 + ], + "type": "table", + "html": "
MethodsReferencePRCC(CC)PRCC(SC)VC(CC)VC(G)LTCC(CC)LTCC(G)CRCRL
mAPR-1mAPR-1mAPR-1mAPR-1mAPR-1mAPR-1mAPR-1mAPR-1
UT-ReID
ICEICCV2145.532.695.793.328.531.451.870.110.116.322.644.04.940.75.07.1
CCACCV2246.334.494.490.225.731.045.162.46.07.411.017.03.432.83.24.4
CACLTIP2252.141.794.790.949.758.968.082.46.29.822.345.65.142.33.64.3
PPLRCVPR2251.440.091.787.423.132.547.768.14.44.86.011.24.841.34.36.2
RTMemTIP2347.235.893.692.737.546.159.775.35.87.518.743.15.039.84.96.8
PurificationTIP2352.341.995.293.047.656.866.881.68.915.623.347.65.242.64.97.0
DCCCICASSP2447.035.390.689.539.350.458.976.17.812.719.644.93.935.73.35.0
UCC-ReID
SiCLarXiv2455.443.296.293.863.971.777.287.910.120.727.657.65.645.45.610.5
CICLNN2455.345.096.795.369.372.876.787.313.025.729.660.36.749.37.812.5
ICSMOurs57.847.997.597.173.575.178.789.613.526.230.160.87.049.67.712.6
SCC-ReID
RCSANetICCV2148.650.297.2100.0--------11.955.616.729.5
GI-ReIDCVPR22-33.3--57.864.5--10.423.729.463.2----
CALCVPR2255.855.299.8100.081.781.487.292.9--------
AIMCVPR2358.357.999.9100.0----19.140.641.176.3----
CCFACVPR2358.461.298.799.6----22.145.342.575.8----
FIRe2TIFS2463.165.099.5100.0----19.144.639.975.918.264.0--
", + "image_path": "bef7b65818872d4467c1340e0d95f565c263dad6fb2189fcaa40a63ff4b6c114.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 77, + 328, + 276, + 372 + ], + "blocks": [ + { + "bbox": [ + 55, + 288, + 555, + 309 + ], + "lines": [ + { + "bbox": [ + 55, + 288, + 555, + 309 + ], + "spans": [ + { + "bbox": [ + 55, + 288, + 555, + 309 + ], + "type": "text", + "content": "Table 2. Comparison of ICSM with existing methods. Optimal performance under two unsupervised settings is highlighted in bold. \"R-1\" denotes \"Rank-1\"." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 77, + 328, + 276, + 372 + ], + "lines": [ + { + "bbox": [ + 77, + 328, + 276, + 372 + ], + "spans": [ + { + "bbox": [ + 77, + 328, + 276, + 372 + ], + "type": "table", + "html": "
MethodsmAPRank-1Training timeTest time
SiCL63.971.76798.2514.20
CICL69.372.810569.9314.16
ICSM73.575.17236.5113.97
", + "image_path": "3544167821ebb9f64c47df66c2b7e5ff4988f514bc2f19f322c449a102f5203a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 65, + 380, + 282, + 392 + ], + "lines": [ + { + "bbox": [ + 65, + 380, + 282, + 392 + ], + "spans": [ + { + "bbox": [ + 65, + 380, + 282, + 392 + ], + "type": "text", + "content": "Table 3. Statistics on performance and efficiency (seconds)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 413, + 296, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 413, + 296, + 509 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 296, + 509 + ], + "type": "text", + "content": "increase the overall training time. In contrast, while existing UCC-ReID methods conduct only one stage of training, they often need to optimize three neural network modules simultaneously, which can require even more training time than ICSM. Additionally, ICSM has a similar testing time compared to existing UCC-ReID methods. These results demonstrate that ICSM not only achieves superior performance but also maintains efficiency." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 516, + 149, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 516, + 149, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 516, + 149, + 529 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 534, + 295, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 629 + ], + "type": "text", + "content": "In this section, the baseline we tested optimizes the image encoder using only " + }, + { + "bbox": [ + 55, + 534, + 295, + 629 + ], + "type": "inline_equation", + "content": "L_{id}" + }, + { + "bbox": [ + 55, + 534, + 295, + 629 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 534, + 295, + 629 + ], + "type": "inline_equation", + "content": "L_{hc}" + }, + { + "bbox": [ + 55, + 534, + 295, + 629 + ], + "type": "text", + "content": " from stage I. As shown in Tab. 4, M1 corresponds to the baseline, while M2 introduces ISE on top of M1. Additionally, to accommodate ISE, M2 replaces " + }, + { + "bbox": [ + 55, + 534, + 295, + 629 + ], + "type": "inline_equation", + "content": "L_{id}" + }, + { + "bbox": [ + 55, + 534, + 295, + 629 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 534, + 295, + 629 + ], + "type": "inline_equation", + "content": "L_{hc}" + }, + { + "bbox": [ + 55, + 534, + 295, + 629 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 534, + 295, + 629 + ], + "type": "inline_equation", + "content": "L_{aid}" + }, + { + "bbox": [ + 55, + 534, + 295, + 629 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 534, + 295, + 629 + ], + "type": "inline_equation", + "content": "L_{ahc}" + }, + { + "bbox": [ + 55, + 534, + 295, + 629 + ], + "type": "text", + "content": ", respectively. M3 and M4 introduce " + }, + { + "bbox": [ + 55, + 534, + 295, + 629 + ], + "type": "inline_equation", + "content": "L_{is}" + }, + { + "bbox": [ + 55, + 534, + 295, + 629 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 534, + 295, + 629 + ], + "type": "inline_equation", + "content": "L_{cs}" + }, + { + "bbox": [ + 55, + 534, + 295, + 629 + ], + "type": "text", + "content": ", respectively, building upon M2. M5, which represents ICSM, incorporates both " + }, + { + "bbox": [ + 55, + 534, + 295, + 629 + ], + "type": "inline_equation", + "content": "L_{is}" + }, + { + "bbox": [ + 55, + 534, + 295, + 629 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 534, + 295, + 629 + ], + "type": "inline_equation", + "content": "L_{cs}" + }, + { + "bbox": [ + 55, + 534, + 295, + 629 + ], + "type": "text", + "content": " on top of M2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": "Effectiveness of ISE. M2 introduces ISE on top of M1 to identify clothing change positive clusters, reliable negative clusters, and hard negative clusters for each compact cluster. As shown in Tab. 4, M2 achieves significant performance improvements over M1 in PRCC(CC), VC-Clothes(CC), and VC-Clothes(G). To further explore the nature of ISE, we visualize the features of several clusters. As shown" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 319, + 331, + 432, + 421 + ], + "blocks": [ + { + "bbox": [ + 319, + 331, + 432, + 421 + ], + "lines": [ + { + "bbox": [ + 319, + 331, + 432, + 421 + ], + "spans": [ + { + "bbox": [ + 319, + 331, + 432, + 421 + ], + "type": "image", + "image_path": "9e6992493cfb051bf3fa3f9ba6cfabe5ef0eb7849803e1d43faa5d964fd0192e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 365, + 422, + 388, + 431 + ], + "lines": [ + { + "bbox": [ + 365, + 422, + 388, + 431 + ], + "spans": [ + { + "bbox": [ + 365, + 422, + 388, + 431 + ], + "type": "text", + "content": "(a) M1" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 434, + 331, + 548, + 420 + ], + "blocks": [ + { + "bbox": [ + 434, + 331, + 548, + 420 + ], + "lines": [ + { + "bbox": [ + 434, + 331, + 548, + 420 + ], + "spans": [ + { + "bbox": [ + 434, + 331, + 548, + 420 + ], + "type": "image", + "image_path": "d4e3fff688845ba21e030142c68905ad5e153d93c28056e13c5ab214fecf8a6c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 481, + 422, + 504, + 431 + ], + "lines": [ + { + "bbox": [ + 481, + 422, + 504, + 431 + ], + "spans": [ + { + "bbox": [ + 481, + 422, + 504, + 431 + ], + "type": "text", + "content": "(b) M2" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 321, + 497, + 548, + 542 + ], + "blocks": [ + { + "bbox": [ + 313, + 441, + 555, + 486 + ], + "lines": [ + { + "bbox": [ + 313, + 441, + 555, + 486 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 555, + 486 + ], + "type": "text", + "content": "Figure 4. Visualization of the clusters corresponding to M1 and M2. Clusters are labeled with black dashed lines. For any given cluster, the tail and head of the red arrow connect it to its clothing change positive cluster identified by ISE." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 321, + 497, + 548, + 542 + ], + "lines": [ + { + "bbox": [ + 321, + 497, + 548, + 542 + ], + "spans": [ + { + "bbox": [ + 321, + 497, + 548, + 542 + ], + "type": "image", + "image_path": "4881f7b6374435d20b7cbc8a60ee3df331963f681f14c3a97083d1157924b2db.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 550, + 555, + 584 + ], + "lines": [ + { + "bbox": [ + 313, + 550, + 555, + 584 + ], + "spans": [ + { + "bbox": [ + 313, + 550, + 555, + 584 + ], + "type": "text", + "content": "Figure 5. Feature maps of M1 and M2. Each triplet, from left to right, consists of the original image, the feature map of M1, and the feature map of M2." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": "in Fig. 4, M2 is able to identify clothing change positive clusters based on M1, significantly reducing the distances between the positive cluster pairs. For the unconnected clothing change positive clusters, the distances are not significantly increased because " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "L_{aid}" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "L_{ahc}" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": " typically prevent these clusters from being included in the optimization scope. This process undoubtedly enhances the model's robustness to clothing changes. As shown in Fig. 5, we observe that M2 indeed focuses more on clothing-invariant re" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19257" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 93, + 70, + 519, + 142 + ], + "blocks": [ + { + "bbox": [ + 93, + 70, + 519, + 142 + ], + "lines": [ + { + "bbox": [ + 93, + 70, + 519, + 142 + ], + "spans": [ + { + "bbox": [ + 93, + 70, + 519, + 142 + ], + "type": "table", + "html": "
MethodsBaselineISE\\(L_{is}\\)\\(L_{cs}\\)PRCC(CC)PRCC(SC)VC-Clothes(CC)VC-Clothes(G)
mAPRank-1mAPRank-1mAPRank-1mAPRank-1
M1✓×××43.931.592.289.023.630.846.667.8
M2✓✓××53.643.095.192.857.664.969.983.5
M3✓✓✓×56.847.197.095.668.672.575.286.3
M4✓✓×✓55.344.695.393.263.869.272.785.0
M5/ICSM✓✓✓✓57.847.997.597.173.575.178.789.6
", + "image_path": "fd9e942d8c4a785a8f969cb37ed4f81379a1f5113d933af7af4cfe829d5f6472.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 188, + 150, + 418, + 162 + ], + "lines": [ + { + "bbox": [ + 188, + 150, + 418, + 162 + ], + "spans": [ + { + "bbox": [ + 188, + 150, + 418, + 162 + ], + "type": "text", + "content": "Table 4. Results of the ablation study for the proposed method." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 67, + 180, + 171, + 243 + ], + "blocks": [ + { + "bbox": [ + 67, + 180, + 171, + 243 + ], + "lines": [ + { + "bbox": [ + 67, + 180, + 171, + 243 + ], + "spans": [ + { + "bbox": [ + 67, + 180, + 171, + 243 + ], + "type": "image", + "image_path": "43ff2dac8fdf843e1bc425dfbc051ccbe48b109e78f8ce5f334f1cde2ff467f2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 108, + 245, + 131, + 254 + ], + "lines": [ + { + "bbox": [ + 108, + 245, + 131, + 254 + ], + "spans": [ + { + "bbox": [ + 108, + 245, + 131, + 254 + ], + "type": "text", + "content": "(a) M2" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 179, + 181, + 284, + 243 + ], + "blocks": [ + { + "bbox": [ + 179, + 181, + 284, + 243 + ], + "lines": [ + { + "bbox": [ + 179, + 181, + 284, + 243 + ], + "spans": [ + { + "bbox": [ + 179, + 181, + 284, + 243 + ], + "type": "image", + "image_path": "6f6af1072a563bb0999401f82de4fcf3dacf5e45f59fd42c935009a37e9b372f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 245, + 244, + 254 + ], + "lines": [ + { + "bbox": [ + 220, + 245, + 244, + 254 + ], + "spans": [ + { + "bbox": [ + 220, + 245, + 244, + 254 + ], + "type": "text", + "content": "(b) M3" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 67, + 255, + 171, + 318 + ], + "blocks": [ + { + "bbox": [ + 67, + 255, + 171, + 318 + ], + "lines": [ + { + "bbox": [ + 67, + 255, + 171, + 318 + ], + "spans": [ + { + "bbox": [ + 67, + 255, + 171, + 318 + ], + "type": "image", + "image_path": "ad57bd73604052a7d00d7a56eaf6799e0f7b57096f0f6704e27c85aabe999d4c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 108, + 319, + 131, + 328 + ], + "lines": [ + { + "bbox": [ + 108, + 319, + 131, + 328 + ], + "spans": [ + { + "bbox": [ + 108, + 319, + 131, + 328 + ], + "type": "text", + "content": "(c) M4" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 179, + 255, + 284, + 318 + ], + "blocks": [ + { + "bbox": [ + 179, + 255, + 284, + 318 + ], + "lines": [ + { + "bbox": [ + 179, + 255, + 284, + 318 + ], + "spans": [ + { + "bbox": [ + 179, + 255, + 284, + 318 + ], + "type": "image", + "image_path": "980d21180ca32db2b08bda720a8d72d048ba1ee391d1e3b1389ff64d45b95987.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 319, + 244, + 328 + ], + "lines": [ + { + "bbox": [ + 220, + 319, + 244, + 328 + ], + "spans": [ + { + "bbox": [ + 220, + 319, + 244, + 328 + ], + "type": "text", + "content": "(d) M5" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 384, + 296, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 296, + 419 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 296, + 419 + ], + "type": "text", + "content": "gions such as the head and arms, or areas like shoes that are less frequently changed, compared to M1. The above results validate the effectiveness of ISE." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 421, + 296, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 421, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 421, + 296, + 612 + ], + "type": "text", + "content": "Effectiveness of " + }, + { + "bbox": [ + 55, + 421, + 296, + 612 + ], + "type": "inline_equation", + "content": "L_{is}" + }, + { + "bbox": [ + 55, + 421, + 296, + 612 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 421, + 296, + 612 + ], + "type": "inline_equation", + "content": "L_{cs}" + }, + { + "bbox": [ + 55, + 421, + 296, + 612 + ], + "type": "text", + "content": ". As shown in Tab. 4, M3 and M4 introduce " + }, + { + "bbox": [ + 55, + 421, + 296, + 612 + ], + "type": "inline_equation", + "content": "L_{is}" + }, + { + "bbox": [ + 55, + 421, + 296, + 612 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 421, + 296, + 612 + ], + "type": "inline_equation", + "content": "L_{cs}" + }, + { + "bbox": [ + 55, + 421, + 296, + 612 + ], + "type": "text", + "content": " into M2, respectively, both achieving significant performance improvements. Additionally, we visualize the feature distance distributions of same clothing positive pairs and clothing change positive pairs for M2, M3, M4, and M5. As shown in Fig. 6, compared to M2, both M3 and M4 significantly reduce the feature distance of clothing change positive pairs while also increasing the overlap between the feature distance distributions of these two types. This validates that both " + }, + { + "bbox": [ + 55, + 421, + 296, + 612 + ], + "type": "inline_equation", + "content": "L_{is}" + }, + { + "bbox": [ + 55, + 421, + 296, + 612 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 421, + 296, + 612 + ], + "type": "inline_equation", + "content": "L_{cs}" + }, + { + "bbox": [ + 55, + 421, + 296, + 612 + ], + "type": "text", + "content": " can enhance model performance by promoting the learning of clothing-invariant features. Furthermore, we observe that M5 further reduces the feature distance of clothing change positive pairs compared to M3 and M4, achieving superior performance over all previously mentioned methods." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 623, + 169, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 623, + 169, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 623, + 169, + 635 + ], + "type": "text", + "content": "4.5. Sensitivity Analysis" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "As shown in Fig. 7, when " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{is} = 2" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{cs} = 1" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": ", the model achieves optimal performance on both PRCC(CC) and VC-Clothes(CC). This confirms the generalizability of " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{is}" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{cs}" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": ". We also observe that when " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{is}" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " is set too high or too low, model performance deteriorates. This is mainly because a small " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{is}" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " reduces the contribution of identity-" + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 326, + 180, + 430, + 244 + ], + "blocks": [ + { + "bbox": [ + 326, + 180, + 430, + 244 + ], + "lines": [ + { + "bbox": [ + 326, + 180, + 430, + 244 + ], + "spans": [ + { + "bbox": [ + 326, + 180, + 430, + 244 + ], + "type": "image", + "image_path": "951e7a7b2133b430221dbc9b99d72572eabbf49db71b58dea56907ec9d0e3beb.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 353, + 246, + 402, + 255 + ], + "lines": [ + { + "bbox": [ + 353, + 246, + 402, + 255 + ], + "spans": [ + { + "bbox": [ + 353, + 246, + 402, + 255 + ], + "type": "text", + "content": "(a) PRCC(CC)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 438, + 180, + 542, + 244 + ], + "blocks": [ + { + "bbox": [ + 438, + 180, + 542, + 244 + ], + "lines": [ + { + "bbox": [ + 438, + 180, + 542, + 244 + ], + "spans": [ + { + "bbox": [ + 438, + 180, + 542, + 244 + ], + "type": "image", + "image_path": "6cd129c725b65f8f567db020e8c0ba5e9ae107ff377bb42eee141c6c638fac32.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 458, + 246, + 524, + 255 + ], + "lines": [ + { + "bbox": [ + 458, + 246, + 524, + 255 + ], + "spans": [ + { + "bbox": [ + 458, + 246, + 524, + 255 + ], + "type": "text", + "content": "(b) VC-Clothes(CC)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 326, + 289, + 430, + 358 + ], + "blocks": [ + { + "bbox": [ + 55, + 338, + 296, + 361 + ], + "lines": [ + { + "bbox": [ + 55, + 338, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 296, + 361 + ], + "type": "text", + "content": "Figure 6. Feature distance distributions of same clothing positive pairs and clothing change positive pairs for M2, M3, M4, and M5." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 326, + 289, + 430, + 358 + ], + "lines": [ + { + "bbox": [ + 326, + 289, + 430, + 358 + ], + "spans": [ + { + "bbox": [ + 326, + 289, + 430, + 358 + ], + "type": "image", + "image_path": "31b39396ff259a14bd491747ad000c2d21c829585af3c708aedba6f2042a418c.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 353, + 361, + 402, + 369 + ], + "lines": [ + { + "bbox": [ + 353, + 361, + 402, + 369 + ], + "spans": [ + { + "bbox": [ + 353, + 361, + 402, + 369 + ], + "type": "text", + "content": "(a) PRCC(CC)" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 351, + 380, + 516, + 392 + ], + "lines": [ + { + "bbox": [ + 351, + 380, + 516, + 392 + ], + "spans": [ + { + "bbox": [ + 351, + 380, + 516, + 392 + ], + "type": "text", + "content": "Figure 8. Impact of " + }, + { + "bbox": [ + 351, + 380, + 516, + 392 + ], + "type": "inline_equation", + "content": "(\\gamma_{a},\\gamma_{c})" + }, + { + "bbox": [ + 351, + 380, + 516, + 392 + ], + "type": "text", + "content": " on performance." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 438, + 289, + 543, + 358 + ], + "blocks": [ + { + "bbox": [ + 348, + 265, + 520, + 277 + ], + "lines": [ + { + "bbox": [ + 348, + 265, + 520, + 277 + ], + "spans": [ + { + "bbox": [ + 348, + 265, + 520, + 277 + ], + "type": "text", + "content": "Figure 7. Impact of " + }, + { + "bbox": [ + 348, + 265, + 520, + 277 + ], + "type": "inline_equation", + "content": "(\\lambda_{is},\\lambda_{cs})" + }, + { + "bbox": [ + 348, + 265, + 520, + 277 + ], + "type": "text", + "content": " on performance." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 438, + 289, + 543, + 358 + ], + "lines": [ + { + "bbox": [ + 438, + 289, + 543, + 358 + ], + "spans": [ + { + "bbox": [ + 438, + 289, + 543, + 358 + ], + "type": "image", + "image_path": "0496e457536f5e0396c8d355a28bda3261b1497c8d5c137246591961c255899e.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 458, + 361, + 524, + 369 + ], + "lines": [ + { + "bbox": [ + 458, + 361, + 524, + 369 + ], + "spans": [ + { + "bbox": [ + 458, + 361, + 524, + 369 + ], + "type": "text", + "content": "(b) VC-Clothes(CC)" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 415, + 555, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 555, + 487 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 555, + 487 + ], + "type": "text", + "content": "related text semantics, while a large " + }, + { + "bbox": [ + 313, + 415, + 555, + 487 + ], + "type": "inline_equation", + "content": "\\lambda_{is}" + }, + { + "bbox": [ + 313, + 415, + 555, + 487 + ], + "type": "text", + "content": " negatively impacts the effectiveness of " + }, + { + "bbox": [ + 313, + 415, + 555, + 487 + ], + "type": "inline_equation", + "content": "L_{aid}" + }, + { + "bbox": [ + 313, + 415, + 555, + 487 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 415, + 555, + 487 + ], + "type": "inline_equation", + "content": "L_{ahc}" + }, + { + "bbox": [ + 313, + 415, + 555, + 487 + ], + "type": "text", + "content": ". As shown in Fig. 8, when " + }, + { + "bbox": [ + 313, + 415, + 555, + 487 + ], + "type": "inline_equation", + "content": "\\gamma_{a} = 2\\%" + }, + { + "bbox": [ + 313, + 415, + 555, + 487 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 415, + 555, + 487 + ], + "type": "inline_equation", + "content": "\\gamma_{c} = 50\\%" + }, + { + "bbox": [ + 313, + 415, + 555, + 487 + ], + "type": "text", + "content": ", the model achieves optimal performance on both datasets. This not only validates the generalization capability of ISE but also verifies the model's robustness to " + }, + { + "bbox": [ + 313, + 415, + 555, + 487 + ], + "type": "inline_equation", + "content": "\\gamma_{a}" + }, + { + "bbox": [ + 313, + 415, + 555, + 487 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 415, + 555, + 487 + ], + "type": "inline_equation", + "content": "\\gamma_{c}" + }, + { + "bbox": [ + 313, + 415, + 555, + 487 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 501, + 388, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 501, + 388, + 513 + ], + "spans": [ + { + "bbox": [ + 313, + 501, + 388, + 513 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we propose an identity-clothing similarity modeling (ICSM) framework for UCC-ReID. To facilitate clothing-invariant feature learning, we introduce intercluster similarity estimation (ISE), identity-oriented semantic contrastive loss, and clothing-invariant semantic contrastive loss. The results of the ablation study validate their effectiveness, demonstrating that ISE enhances the model's robustness to clothing changes by identifying reliable clothing change positive clusters, while identity-oriented semantic contrastive loss and clothing-invariant semantic contrastive loss further promote clothing invariance learning. Experimental results across eight testing scenarios validate the effectiveness and superiority of the proposed ICSM, which not only outperforms existing unsupervised methods but also achieves competitive performance compared to certain supervised methods." + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19258" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 294, + 127 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 294, + 127 + ], + "type": "text", + "content": "This work is supported by National Key Research and Development Program of China (Grant no. 2023YFC3305005)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 114, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 114, + 152 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 114, + 152 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 159, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 159, + 296, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 159, + 296, + 202 + ], + "spans": [ + { + "bbox": [ + 61, + 159, + 296, + 202 + ], + "type": "text", + "content": "[1] Hao Chen, Benoit Lagadec, and Francois Bremond. ICE: Inter-instance contrastive encoding for unsupervised person re-identification. In ICCV, pages 14960-14969, 2021. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 205, + 296, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 205, + 296, + 247 + ], + "spans": [ + { + "bbox": [ + 61, + 205, + 296, + 247 + ], + "type": "text", + "content": "[2] Yoonki Cho, Woo Jae Kim, Seunghoon Hong, and Sung-Eui Yoon. Part-based pseudo label refinement for unsupervised person re-identification. In CVPR, pages 7308-7318, 2022. 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 250, + 296, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 250, + 296, + 293 + ], + "spans": [ + { + "bbox": [ + 62, + 250, + 296, + 293 + ], + "type": "text", + "content": "[3] Zuozhuo Dai, Guangyuan Wang, Weihao Yuan, Siyu Zhu, and Ping Tan. Cluster contrast for unsupervised person re-identification. In ACCV, pages 1142–1160, 2022. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 295, + 294, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 295, + 294, + 338 + ], + "spans": [ + { + "bbox": [ + 62, + 295, + 294, + 338 + ], + "type": "text", + "content": "[4] Martin Ester, Hans-Peter Kriegel, Jörg Sander, Xiaowei Xu, et al. A density-based algorithm for discovering clusters in large spatial databases with noise. In KDD, pages 226-231, 1996. 1, 2, 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 341, + 294, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 341, + 294, + 384 + ], + "spans": [ + { + "bbox": [ + 62, + 341, + 294, + 384 + ], + "type": "text", + "content": "[5] Tiantian Gong, Guodong Du, Junsheng Wang, Yongkang Ding, and Liyan Zhang. Prototype-guided cross-modal completion and alignment for incomplete text-based person re-identification. In ACM MM, pages 5253-5261, 2023. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 386, + 294, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 386, + 294, + 428 + ], + "spans": [ + { + "bbox": [ + 62, + 386, + 294, + 428 + ], + "type": "text", + "content": "[6] Tiantian Gong, Junsheng Wang, and Liyan Zhang. Cross-modal semantic aligning and neighbor-aware completing for robust text-image person retrieval. Information Fusion, 112: 102544, 2024. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 431, + 294, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 431, + 294, + 474 + ], + "spans": [ + { + "bbox": [ + 62, + 431, + 294, + 474 + ], + "type": "text", + "content": "[7] Xinqian Gu, Hong Chang, Bingpeng Ma, Shutao Bai, Shiguang Shan, and Xilin Chen. Clothes-changing person re-identification with rgb modality only. In CVPR, pages 1060–1069, 2022. 1, 2, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 476, + 294, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 476, + 294, + 509 + ], + "spans": [ + { + "bbox": [ + 62, + 476, + 294, + 509 + ], + "type": "text", + "content": "[8] Ke Han, Shaogang Gong, Yan Huang, Liang Wang, and Ti-niu Tan. Clothing-change feature augmentation for person re-identification. In CVPR, pages 22066–22075, 2023. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 511, + 294, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 511, + 294, + 543 + ], + "spans": [ + { + "bbox": [ + 62, + 511, + 294, + 543 + ], + "type": "text", + "content": "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 544, + 294, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 544, + 294, + 588 + ], + "spans": [ + { + "bbox": [ + 57, + 544, + 294, + 588 + ], + "type": "text", + "content": "[10] Tao He, Leqi Shen, Yuchen Guo, Guiguang Ding, and Zhenhua Guo. Secret: Self-consistent pseudo label refinement for unsupervised domain adaptive person re-identification. In AAAI, pages 879–887, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 590, + 294, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 590, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 57, + 590, + 294, + 633 + ], + "type": "text", + "content": "[11] Ziqi He, Mengjia Xue, Yunhao Du, Zhicheng Zhao, and Fei Su. Dynamic clustering and cluster contrastive learning for unsupervised person re-id with feature distribution alignment. In ICASSP, pages 3610-3614, 2024. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 635, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 294, + 667 + ], + "type": "text", + "content": "[12] Alexander Hermans, Lucas Beyer, and Bastian Leibe. In defense of the triplet loss for person re-identification. arXiv preprint arXiv:1703.07737, 2017. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "type": "text", + "content": "[13] Peixian Hong, Tao Wu, Ancong Wu, Xintong Han, and Wei-Shi Zheng. Fine-grained shape-appearance mutual learning for cloth-changing person re-identification. In CVPR, pages 10513–10522, 2021. 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[14] Yan Huang, Jingsong Xu, Qiang Wu, Yi Zhong, Peng Zhang, and Zhaoxiang Zhang. Beyond scalar neuron: Adopting vector-neuron capsules for long-term person re-identification. IEEE TCSVT, 30(10):3459-3471, 2019. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 118, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 150 + ], + "type": "text", + "content": "[15] Yan Huang, Qiang Wu, JingSong Xu, Yi Zhong, and ZhaoXiang Zhang. Clothing status awareness for long-term person re-identification. In ICCV, pages 11895–11904, 2021. 2, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 152, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 152, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 317, + 152, + 553, + 206 + ], + "type": "text", + "content": "[16] Xin Jin, Tianyu He, Kecheng Zheng, Zhiheng Yin, Xu Shen, Zhen Huang, Ruoyu Feng, Jianqiang Huang, Zhibo Chen, and Xian-Sheng Hua. Cloth-changing person re-identification from a single image with gait prediction and regularization. In CVPR, pages 14278–14287, 2022. 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 208, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 208, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 317, + 208, + 553, + 239 + ], + "type": "text", + "content": "[17] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 242, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 242, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 317, + 242, + 553, + 274 + ], + "type": "text", + "content": "[18] Long Lan, Xiao Teng, Jing Zhang, Xiang Zhang, and Dacheng Tao. Learning to purification for unsupervised person re-identification. IEEE TIP, 32:3338-3353, 2023. 2, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 276, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 276, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 317, + 276, + 553, + 308 + ], + "type": "text", + "content": "[19] Mingkun Li, Chun-Guang Li, and Jun Guo. Cluster-guided asymmetric contrastive learning for unsupervised person re-identification. IEEE TIP, 31:3606-3617, 2022. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 309, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 309, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 317, + 309, + 553, + 352 + ], + "type": "text", + "content": "[20] Mingkun Li, Peng Xu, Chun-Guang Li, and Jun Guo. SiCL: Silhouette-driven contrastive learning for unsupervised person re-identification with clothes change. arXiv preprint arXiv:2305.13600, 2023. 1, 2, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 354, + 553, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 386 + ], + "type": "text", + "content": "[21] Peike Li, Yunqiu Xu, Yunchao Wei, and Yi Yang. Self-correction for human parsing. IEEE TPAMI, 44(6):3260-3271, 2020. 1, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 388, + 553, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 430 + ], + "type": "text", + "content": "[22] Siyuan Li, Li Sun, and Qingli Li. Clip-reid: exploiting vision-language model for image re-identification without concrete text labels. In AAAI, pages 1405–1413, 2023. 2, 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 433, + 553, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 433, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 316, + 433, + 553, + 475 + ], + "type": "text", + "content": "[23] Shuang Li, Jiaxu Leng, Guozhang Li, Ji Gan, Xinbo Gao, et al. Clip-driven cloth-agnostic feature learning for cloth-changing person re-identification. arXiv preprint arXiv:2406.09198, 2024. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 478, + 553, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 478, + 553, + 531 + ], + "spans": [ + { + "bbox": [ + 317, + 478, + 553, + 531 + ], + "type": "text", + "content": "[24] Weijia Li, Saihui Hou, Chunjie Zhang, Chunshui Cao, Xu Liu, Yongzhen Huang, and Yao Zhao. An in-depth exploration of person re-identification and gait recognition in cloth-changing conditions. In CVPR, pages 13824–13833, 2023. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 534, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 534, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 317, + 534, + 553, + 567 + ], + "type": "text", + "content": "[25] Yutian Lin, Xuanyi Dong, Liang Zheng, Yan Yan, and Yi Yang. A bottom-up clustering approach to unsupervised person re-identification. In AAAI, pages 8738–8745, 2019. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 568, + 553, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 568, + 553, + 599 + ], + "spans": [ + { + "bbox": [ + 317, + 568, + 553, + 599 + ], + "type": "text", + "content": "[26] Fangyi Liu, Mang Ye, and Bo Du. Cloth-aware augmentation for cloth-generalized person re-identification. In ACM MM, 2024. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 601, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 601, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 317, + 601, + 553, + 634 + ], + "type": "text", + "content": "[27] Vuong D Nguyen, Pranav Mantini, and Shishir K Shah. Contrastive clothing and pose generation for cloth-changing person re-identification. In CVPR, pages 7541–7549, 2024. 1" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 635, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 553, + 667 + ], + "type": "text", + "content": "[28] Zhiqi Pang, Lingling Zhao, Qiuyang Liu, and Chunyu Wang. Camera invariant feature learning for unsupervised person re-identification. IEEE TMM, 25:6171-6182, 2022. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[29] Zhiqi Pang, Lingling Zhao, and Chunyu Wang. Clothing-invariant contrastive learning for unsupervised person re-identification. Neural Networks, page 106477, 2024. 1, 2, 6" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19259" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[30] Xuelin Qian, Wenxuan Wang, Li Zhang, Fangrui Zhu, Yanwei Fu, Tao Xiang, Yu-Gang Jiang, and Xiangyang Xue. Long-term cloth-changing person re-identification. In ACCV, 2020. 2, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 119, + 294, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 119, + 294, + 174 + ], + "spans": [ + { + "bbox": [ + 56, + 119, + 294, + 174 + ], + "type": "text", + "content": "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 2, 3, 4, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 176, + 294, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 176, + 294, + 231 + ], + "spans": [ + { + "bbox": [ + 56, + 176, + 294, + 231 + ], + "type": "text", + "content": "[32] Jiangming Shi, Yachao Zhang, Xiangbo Yin, Yuan Xie, Zhizhong Zhang, Jianping Fan, Zhongchao Shi, and Yanyun Qu. Dual pseudo-labels interactive self-training for semi-supervised visible-infrared person re-identification. In ICCV, pages 11218-11228, 2023. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 233, + 294, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 233, + 294, + 265 + ], + "spans": [ + { + "bbox": [ + 56, + 233, + 294, + 265 + ], + "type": "text", + "content": "[33] Fangbin Wan, Yang Wu, Xuelin Qian, Yixiong Chen, and Yanwei Fu. When person re-identification meets changing clothes. In CVPRW, pages 830–831, 2020. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 268, + 294, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 268, + 294, + 310 + ], + "spans": [ + { + "bbox": [ + 56, + 268, + 294, + 310 + ], + "type": "text", + "content": "[34] Menglin Wang, Baisheng Lai, Jianqiang Huang, Xiaojin Gong, and Xian-Sheng Hua. Camera-aware proxies for unsupervised person re-identification. In AAAI, pages 2764–2772, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 314, + 294, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 314, + 294, + 356 + ], + "spans": [ + { + "bbox": [ + 56, + 314, + 294, + 356 + ], + "type": "text", + "content": "[35] Qizao Wang, Xuelin Qian, Bin Li, Xiangyang Xue, and Yanwei Fu. Exploring fine-grained representation and decomposition for cloth-changing person re-identification. IEEE TIFS, 2024. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 359, + 294, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 359, + 294, + 391 + ], + "spans": [ + { + "bbox": [ + 56, + 359, + 294, + 391 + ], + "type": "text", + "content": "[36] Shiyu Xuan and Shiliang Zhang. Intra-inter camera similarity for unsupervised person re-identification. In CVPR, pages 11926–11935, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 394, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 394, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 394, + 294, + 437 + ], + "type": "text", + "content": "[37] Shuanglin Yan, Neng Dong, Jun Liu, Liyan Zhang, and Jinhui Tang. Learning comprehensive representations with richer self for text-to-image person re-identification. In ACM MM, pages 6202–6211, 2023. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 441, + 294, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 441, + 294, + 493 + ], + "spans": [ + { + "bbox": [ + 56, + 441, + 294, + 493 + ], + "type": "text", + "content": "[38] Fengxiang Yang, Zhun Zhong, Zhiming Luo, Yuanzheng Cai, Yaojin Lin, Shaozi Li, and Nicu Sebe. Joint noisetolerant learning and meta camera shift adaptation for unsupervised person re-identification. In CVPR, pages 4855–4864, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 498, + 294, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 498, + 294, + 529 + ], + "spans": [ + { + "bbox": [ + 56, + 498, + 294, + 529 + ], + "type": "text", + "content": "[39] Qize Yang, Ancong Wu, and Wei-Shi Zheng. Person re-identification by contour sketch under moderate clothing change. IEEE TPAMI, 43(6):2029-2046, 2019. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 532, + 294, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 532, + 294, + 574 + ], + "spans": [ + { + "bbox": [ + 56, + 532, + 294, + 574 + ], + "type": "text", + "content": "[40] Zhengwei Yang, Meng Lin, Xian Zhong, Yu Wu, and Zheng Wang. Good is bad: Causality inspired cloth-debiasing for cloth-changing person re-identification. In CVPR, pages 1472–1481, 2023. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 578, + 294, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 578, + 294, + 620 + ], + "spans": [ + { + "bbox": [ + 56, + 578, + 294, + 620 + ], + "type": "text", + "content": "[41] Mang Ye, Jianbing Shen, Gaojie Lin, Tao Xiang, Ling Shao, and Steven CH Hoi. Deep learning for person re-identification: A survey and outlook. IEEE TPAMI, 44(6): 2872-2893, 2021. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 624, + 294, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 655 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 655 + ], + "type": "text", + "content": "[42] Junhui Yin, Xinyu Zhang, Zhanyu Ma, Jun Guo, and Yifan Liu. A real-time memory updating strategy for unsupervised person re-identification. IEEE TIP, 32:2309–2321, 2023. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "text", + "content": "[43] Xiangbo Yin, Jiangming Shi, Yachao Zhang, Yang Lu, Zhizhong Zhang, Yuan Xie, and Yanyun Qu. Robust pseudo-label learning with neighbor relation for unsupervised visible-infrared person re-identification. In ACM MM, pages 2242–2251, 2024. 1" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 319 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[44] Kaiwei Zeng, Munan Ning, Yaohua Wang, and Yang Guo. Hierarchical clustering with hard-batch triplet loss for person re-identification. In CVPR, pages 13657–13665, 2020. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 107, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 149 + ], + "type": "text", + "content": "[45] Xiao Zhang, Yixiao Ge, Yu Qiao, and Hongsheng Li. Refining pseudo labels with clustering consensus over generations for unsupervised object re-identification. In CVPR, pages 3436–3445, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "text", + "content": "[46] Xinyu Zhang, Dongdong Li, Zhigang Wang, Jian Wang, Errui Ding, Javen Qinfeng Shi, Zhaoxiang Zhang, and Jingdong Wang. Implicit sample extension for unsupervised person re-identification. In CVPR, pages 7369-7378, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 198, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 239 + ], + "type": "text", + "content": "[47] Yukang Zhang and Hanzi Wang. Diverse embedding expansion network and low-light cross-modality benchmark for visible-infrared person re-identification. In CVPR, pages 2153–2162, 2023. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 242, + 553, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 242, + 553, + 284 + ], + "spans": [ + { + "bbox": [ + 316, + 242, + 553, + 284 + ], + "type": "text", + "content": "[48] Yukang Zhang, Yan Yan, Jie Li, and Hanzi Wang. Mrcn: A novel modality restitution and compensation network for visible-infrared person re-identification. In AAAI, pages 3498–3506, 2023. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 287, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 553, + 319 + ], + "type": "text", + "content": "[49] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 130(9):2337-2348, 2022. 2, 4" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19260" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Identity-Preserving Text-to-Video Generation by Frequency Decomposition/16b4d9e1-d6f9-4df5-9c6a-3949d63cad2d_content_list.json b/2025/Identity-Preserving Text-to-Video Generation by Frequency Decomposition/16b4d9e1-d6f9-4df5-9c6a-3949d63cad2d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0f529070376f39cc591cb609c770bfcb809b8867 --- /dev/null +++ b/2025/Identity-Preserving Text-to-Video Generation by Frequency Decomposition/16b4d9e1-d6f9-4df5-9c6a-3949d63cad2d_content_list.json @@ -0,0 +1,1813 @@ +[ + { + "type": "text", + "text": "Identity-Preserving Text-to-Video Generation by Frequency Decomposition", + "text_level": 1, + "bbox": [ + 119, + 130, + 879, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shenghai Yuan $^{1,4}$ , Jina Fu Huang $^{3}$ , Xianyi He $^{1,4}$ , Yunyang Ge $^{1,4}$ , Yujun Shi $^{5}$ , Liuhan Chen $^{1,4}$ , Jiebo Luo $^{3}$ , Li Yuan $^{1,2,\\dagger}$", + "bbox": [ + 254, + 179, + 743, + 218 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Peking University, 2 Peng Cheng Laboratory, 3 University of Rochester, 4 Rabbitpre Intelligence, 5 National University of Singapore", + "bbox": [ + 210, + 232, + 785, + 270 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{yuanshenghai, HeXianyi, yunyang, chenliuhan}@stu.pku.edu.cn, shi.yujun@u.nus.edu, yuanli-ece@pku.edu.cn, {jhuang90@ur, jluo@cs}.rochester.edu", + "bbox": [ + 148, + 289, + 844, + 323 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://oku-yuangroup.github.io/ConsisID", + "text_level": 1, + "bbox": [ + 259, + 340, + 736, + 358 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/34b54a7e24fa53311fda96cdaa729a4fdf59d6896abd5e3cc5055d42acb33f7a.jpg", + "image_caption": [ + "Reference Image" + ], + "image_footnote": [], + "bbox": [ + 125, + 388, + 214, + 458 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "a man standing next to an airplane, engaged in a conversation on his cell phone ... wearings sunglasses and a black top ... talking seriously ... airplane has a green striperunning along its side, and there is a large engine visible behind his ...", + "bbox": [ + 230, + 369, + 861, + 388 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8c8d0e12247599b8644cccecb1e8528bb9339a8aa8432162a342213397f31948.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 220, + 388, + 483, + 458 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/59fc6dd022d7e21adb1bcf77fba633461a578cf874bfa8a97dda2da1a0741a47.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 483, + 388, + 612, + 458 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1d77819c653754433091187f4d5b7b4c77832ce6411034518c8b17f37798c1fe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 612, + 388, + 741, + 458 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f5e00868e7d1b23805e951a3339837a845c521dfdc07e08e3d5e752c51ec2348.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 741, + 388, + 872, + 458 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f029aaf3ef14bd5451a239f376da728354404bd4b192f53420e07333f5312550.jpg", + "image_caption": [ + "Reference Image" + ], + "image_footnote": [], + "bbox": [ + 125, + 484, + 214, + 553 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "a young woman with long blonde hair ... front of a lush, green bush...white flowers ... wearing a black top ... smiling and looking at the camera while gently touching the flowers on the bush ... then bends downslightly and smells one of the flower...", + "bbox": [ + 222, + 468, + 870, + 484 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d4fc29abb568f78cc3c1aae6314f1ac4aa5f9ad86d00bbd5c0c82a6f50ebe922.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 220, + 484, + 351, + 554 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4ae97d4fe9b05646a907ba30623396ef9828e41dc2587d415cf74afaaa34a64a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 351, + 484, + 478, + 554 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bb0f02d3e50739dfe45ecf9ae29d058c0db21b2e9648245af32f966e2c62f110.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 480, + 484, + 609, + 554 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2a0dc2c59d8a00a6a74a988c1c62816a5d893d5dcef5710281cc958ca9e768c3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 611, + 484, + 741, + 554 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e676eccfc9c1c145bf46da5125eed96a72a038c9afa02862f13ad3adcf002d92.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 741, + 484, + 872, + 554 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d98e7dada3c07beb38d5374f3537080db851d998e1da545db5a313a0a64df515.jpg", + "image_caption": [ + "Reference Image" + ], + "image_footnote": [], + "bbox": [ + 125, + 575, + 214, + 647 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "... a man ... his gaze focused and intense as he holds the basketball ... dressed in athletic gear ... The basketball is positioned firmly in his hands as he gradually puts it down ... urban street court, with the fading light of dusk casting a soft glow over the scene. The man's serious expression and steady grip on the ball ...", + "bbox": [ + 222, + 560, + 867, + 577 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b76fd513743a1e45fba7555bc7cbab5498badd0ab55ab6b030c354e996af2e43.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 220, + 577, + 351, + 646 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/213b985ccecdf65154508d4fef0220dbefae56257961671ca24a9a8563c25bb5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 351, + 577, + 480, + 646 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/074fa865a382c4e2a6d8163a3bb00355d485522016f22ee0d6dc70c35c6f379c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 480, + 577, + 609, + 647 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/40a242efbdff3f99b2d0804ed1e3fd81236be0add5f95a9f51e3840423625347.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 611, + 577, + 741, + 647 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/02d55525786d8af3fedaa0e7f6418306a745c2ebfc18b217fdae8eedf308e0de.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 741, + 577, + 872, + 647 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1171cbd67bc0542229ff2e85b5b63a0b39018981ac1042d8ca7deb3769ec8082.jpg", + "image_caption": [ + "Reference Image" + ], + "image_footnote": [], + "bbox": [ + 125, + 670, + 214, + 738 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "a young boy sitting at a table, eating a piece of food. He appears to be enjoying his meal, as he takes a bite and chews it. The boy is wearing a blue shirt and has short hair...background is dark, with some light coming from the left side of the frame. There is a straw visible on the right side of the frame ...", + "bbox": [ + 220, + 650, + 870, + 670 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/71329db7f406c6a1885d06dbb74fd5e2e04095284f1cbe247d9d96be9178c128.jpg", + "image_caption": [ + "Figure 1. Examples of identity-preserving video generation (IPT2V) by our ConsisID. Given a reference image, our method can generate realistic and personalized human-centered videos while preserving identity. Red indicates that attributes in long instructions." + ], + "image_footnote": [], + "bbox": [ + 220, + 670, + 351, + 738 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bb7842d95518c076af165e834e81f2c4c5fce71c46dad32125dccaddd5d05696.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 351, + 670, + 478, + 738 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bc38c9aec80ace97a6161457a904a8479fad07c0ef314884c34f32619601e83d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 480, + 670, + 609, + 738 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f6e448c4e9073d84f80d5cc990c987cd994c547581405da1d441134acdd42bb6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 611, + 670, + 741, + 738 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4f122182d385fe095ea3583f65c56f81bc0a05b5c1015e4f108f7797a8315b06.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 741, + 670, + 872, + 738 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 792, + 326, + 808 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Identity-preserving text-to-video (IPT2V) generation aims to create high-fidelity videos with consistent human identity. It is an important task in video generation but remains an open problem for generative models. This paper pushes the technical frontier of IPT2V in two directions that", + "bbox": [ + 89, + 825, + 485, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "have not been resolved in the literature: (1) A tuning-free pipeline without tedious case-by-case finetuning, and (2) A frequency-aware heuristic identity-preserving Diffusion Transformer (DiT)-based control scheme. To achieve these goals, we propose ConsisID, a tuning-free DiT-based controllable IPT2V model to keep human-identity consistent in the generated video. Inspired by prior findings in fre", + "bbox": [ + 511, + 794, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation.", + "bbox": [ + 236, + 0, + 810, + 18 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "Except for this watermark, it is identical to the accepted version;", + "bbox": [ + 323, + 16, + 723, + 31 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 292, + 31, + 754, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12978", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "quency analysis of vision/diffusion transformers, it employs identity-control signals base on frequency domain, since facial features can be decomposed into low-frequency global features (e.g., profile, proportions) and high-frequency intrinsic features (e.g., identity markers that remain unaffected by pose changes). Extensive experiments demonstrate that our frequency-aware heuristic scheme provides an optimal control solution for DiT-based models, making strides towards more effective IPT2V.", + "bbox": [ + 88, + 90, + 482, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 256, + 222, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Large-scale pre-trained video diffusion models [30, 66, 77, 78] have facilitated a variety of downstream applications [49, 54, 69, 70, 72, 74], particularly in identity-preserving text-to-video (IPT2V) [7, 38, 58, 60, 61]. However, existing methods face significant challenges, particularly the high overhead associated with the need for case-by-case finetuning, which diminishes their applicability. Within the open-source community, only the ID-Animator [15] can implement tuning-free IPT2V, but it can only generate videos similar to talking head [57] and has poor id preservation.", + "bbox": [ + 91, + 281, + 482, + 431 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Additionally, the above efforts are predominantly based on U-Net and cannot be adapted to the emerging DiT-based video model [30, 63, 66, 77, 78]. This challenge may stem from the inherent limitations of DiT compared to U-Net, including greater difficulty in training convergence and weakness in perceiving facial details. From some prior findings in frequency analysis of vision/diffusion transformers [2-4, 45, 50, 55, 71], we can know that the reason is: Finding 1: Shallow (e.g., low-level, low-frequency) features are essential for pixel-level prediction tasks in diffusion models, as they ease model training. U-Net facilitates model convergence by aggregating shallow features to the decoder via long skip connections, a mechanism that DiT does not incorporate; Finding 2: Transformers have limited perception of high-frequency information, which is important for preserving facial features. The encoder-decoder architecture of U-Net naturally possesses multi-scale features (e.g., richness in high-frequency), while DiT lacks a comparable structure. To develop a DiT-based control model, these must be addressed first. Please see Appendix for more details.", + "bbox": [ + 91, + 433, + 482, + 734 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For ID-preserving video generation, the challenges stem from the requirement for each frame to incorporate both high-frequency (e.g., age- and make-up-independent identity markers) and low-frequency information (e.g., facial shape) derived from the reference image, which can just be used to make up for the DiT defects mentioned above. Therefore, we propose ConsisID, to keep the identity consistency in video generation by frequency decomposition, based on the previously Findings of DiT in frequency analysis. Thanks to the large-scale pre-trained DiT, we can use its powerful capabilities to achieve tuning-free effects.", + "bbox": [ + 89, + 734, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ConsisID decouples identity features into high- and low-frequency signals, which are injected into specific locations within the DiT, facilitating efficient IPT2V generation. Specifically, in line with Finding 1, we first convert the reference image and the facial key points to the low-frequency signal, then concatenate them with input noise latent to ease the training. Following Finding 2, we utilize a dual-tower feature extractor to capture high-frequency facial information, which is integrated with vision tokens within the transformer block, thereby enhancing the DiT's high-frequency perception capabilities. Finally, to transform the pre-trained model into an IPT2V model and improve its generalization, we further introduce a hierarchical training strategy.", + "bbox": [ + 511, + 90, + 903, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows:", + "bbox": [ + 532, + 287, + 857, + 301 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce ConsisID, a tuning-free identity-preserving DiT-based IPT2V model, which preserves the identity of the main subject of the video using control signals from frequency decomposition.", + "- We propose a hierarchical training strategy, including coarse-to-fine training, dynamic mask loss, and dynamic cross-face loss, which work together to facilitate training and enhance generalization effectively.", + "- Extensive experiments demonstrate that our ConsisID can generate high-quality, editable, consistent identity-preserving videos, benefiting from our frequency-aware identity-preserving T2V DiT-based control scheme." + ], + "bbox": [ + 511, + 304, + 903, + 484 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 497, + 653, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Tuning-based Identity-preserving T2V Models. Diffusion models are widely recognized for their strong generative capabilities [19, 34-37, 44, 72, 73], significantly advancing the development of identity-preserving generative models [8, 39, 59, 74]. Initially, the researchers used tuning-based methods to generate content that matched the input ID. This process requires finetuning pretrained model for each new person during inference. For example, DreamBooth [47] introduced a novel loss function to fine-tune the entire network, embedding identity information while preserving the original generative capabilities. LoRA [21], similar to DreamBooth [47], requires training only a small subset of network parameters. In contrast, Textual Inversion [11] freezes the pretrained network and embeds identity information into a trainable word embedding. Subsequent tuning-based methods, including both image and video models based on U-Net or DiT architectures [7, 26, 48, 58, 60-62, 65, 76], generally follow three main approaches. While these models demonstrate substantial effectiveness, the requirement to fine-tune for each new identity restricts their practical applicability.", + "bbox": [ + 511, + 523, + 906, + 840 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Tuning-free Identity-preserving T2V Models. To address the issue of high resource consumption, several tuning-free diffusion models have recently emerged in the field of image generation [13, 14, 29, 56, 67]. These models", + "bbox": [ + 511, + 840, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12979", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/77d7dd582ca5e7bd8036fda0935f30af10ef8bd107c4c743d66195a068f5cca1.jpg", + "image_caption": [ + "Figure 2. Overview of the proposed ConsisID. Based on Findings of DiT, low-frequency facial information is embedded into the shallow layers, while high-frequency information is incorporated into the vision tokens within the attention blocks. The ID-preserving Recipe is applied to ease training and improve generalization. The cross face, DropToken and Dropout are executed based on probability." + ], + "image_footnote": [], + "bbox": [ + 133, + 88, + 867, + 324 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "do not require finetuning parameters for newly introduced IDs during inference. For instance, IP-Adapter [67] utilizes the CLIP [43] features of the identity image through cross-attention to guide the pretrained model in generating identity-preserving images. InstantID [56] extends this approach by replacing CLIP [43] features with Arcface [10] features and integrating a pose network to adjust facial proportions. Unlike these initial methods, which introduce control signals via visual tokens, PhotoMaker [29] and Imagine Yourself [16] leverage text tokens. Specifically, PhotoMaker [29] concatenates identity features obtained from the CLIP encoder [43] to the text embedding, while Imagine Yourself [16] uses element-wise addition for feature fusion. In the domain of video generation, only MovieGen [42] and ID-Animator [15] currently support ID-preserving text-to-video (IPT2V) generation. MovieGen is closed-source, whereas ID-Animator is open-source but uses a methodology similar to image models, leading to lower-quality identity preservation in the generated videos. We select the emerging DiT architecture [30, 33, 66, 77] and optimize it for IPT2V, drawing on conclusions from prior frequency analyses [2-4, 45, 50, 55]. This enables high-quality, editable, and consistent ID-preserving video generation.", + "bbox": [ + 91, + 404, + 483, + 752 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 762, + 225, + 780 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 787, + 232, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion Model. Text-to-video generation models usually utilize the diffusion paradigm, which gradually transforms noise $\\epsilon$ into a video $x_0$ . Originally, denoising was conducted directly within the pixel space [20, 51, 52]; however, due to significant computational overheads, recent methods predominantly employ latent space [12, 25, 46,", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "72]. The optimization process is defined:", + "bbox": [ + 513, + 404, + 785, + 419 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {a} = \\mathbb {E} _ {x _ {0}, t, y, \\epsilon} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (x _ {0}, t, \\tau_ {\\theta} (y)) \\| _ {2} ^ {2} \\right], \\qquad (1)\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 431, + 906, + 460 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $y$ is text condition, $\\epsilon$ is sampled from a standard normal distribution (e.g., $\\epsilon \\sim \\mathcal{N}(0,1)$ ), and $\\tau_{\\theta}(\\cdot)$ is the text encoder. By replacing $x_0$ with $\\mathcal{E}(x_0)$ , the latent diffusion is derived, which is used by ConsisID.", + "bbox": [ + 511, + 470, + 906, + 531 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion Transformer. The DiT-based video generation model shows significant potential in simulating the physical world [6, 66, 78]. Despite being a novel architecture, research on controllable generation has been limited, and current methods [9, 13, 42, 75] largely resemble U-Net based approaches [11, 39, 74]. However, no study has yet examined why this approach works with DiT. Drawing from prior analyses of Diffusion and Transformer from a frequency domain perspective [2-4, 45, 50, 55], we conclude that: (1) Low-frequency (e.g., shallow-layer) features are essential for pixel-level prediction tasks in diffusion models, which helps facilitate model training; (2) Transformers have limited perception for high-frequency information, which is important for controllable generation. Based on these, we decouple ID features into high- and low-frequency parts and inject them into specific locations, achieving effective identity-preserving text-to-video generation.", + "bbox": [ + 511, + 532, + 908, + 790 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. ConsisID: Keep Your Identity Consistent", + "text_level": 1, + "bbox": [ + 511, + 801, + 864, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overview is illustrated in Figure 2. Given a reference image, the global facial extractor and local facial extractor inject both high- and low-frequency facial information into model, which then generates identity-preserving videos with the assistance of the consistency training strategy.", + "bbox": [ + 511, + 824, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12980", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2.1. Low-frequency View: Global Facial Extractor", + "text_level": 1, + "bbox": [ + 89, + 90, + 457, + 104 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In light of Finding 1, enhancing low-level (e.g., shallow, low-frequency) features accelerates model convergence. To easily adapt a pre-trained model for the IPT2V task, the most direct approach is concatenating the reference face with the noise input latent [5]. However, the reference face contains both high-frequency details (e.g., eye and lip textures) and low-frequency information (e.g., facial proportions and contours). From Finding 2, prematurely injecting high-frequency information into the Transformer is inefficient and may hinder the model's processing of low-frequency information, as the Transformer focuses primarily on low-frequency features. In addition, feeding the reference face directly into the model could introduce irrelevant noise such as lighting and shadows. To mitigate this, we extract facial key points, convert them to an RGB image, and then concatenate it with the reference image, as shown in Figure 2. This strategy focuses the model's attention on the low-frequency signals in the face, while minimizing the impact of extraneous features. We found that when this component is discarded, the model is difficult to converge. The objective function is changed to:", + "bbox": [ + 89, + 108, + 483, + 400 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {b} = \\mathbb {E} _ {x _ {0}, t, y, f, \\epsilon} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (x _ {0}, t, \\tau_ {\\theta} (y), \\psi_ {\\theta} (f)) \\| _ {2} ^ {2} \\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 409, + 482, + 428 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\psi_{\\theta}(\\cdot)$ is a variational autoencoder, $f$ represents the reference image, we ignore key points here for simplicity.", + "bbox": [ + 89, + 436, + 482, + 467 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2. High-frequency View: Local Facial Extractor", + "text_level": 1, + "bbox": [ + 89, + 474, + 454, + 489 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In light of Finding 2, we recognize that Transformers have limited sensitivity to high-frequency information. It can be concluded that relying solely on global facial extractor is insufficient for IPT2V generation, as global facial features lack of high-frequency information. So we use a face recognition backbone [10] to extract high-frequency features, as these are invariant to non-ID attributes (e.g., expression, posture, and shape). We refer to these features as intrinsic identity features (e.g., high-frequency), since age and makeup do not alter an individual's core identity. Following [14], we utilize the penultimate layer of the backbone, rather than its output, as it retains more spatial information pertinent to identity. However, our experiments reveal that while the face recognition backbone improves identity consistency, it lacks the semantic features required for editing. This task demands not only maintaining identity consistency but also incorporating the ability to edit, such as generating videos of faces with the same identity but varying age and makeup. Previous research [15, 16] relies solely on the CLIP encoder [43] to enable editing capabilities. However, since CLIP is not specifically trained on face datasets, the features it extracts include irrelevant non-face information, which can compromise identity fidelity [29, 56, 67].", + "bbox": [ + 89, + 492, + 482, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address these challenges, we first use a facial recognition backbone to extract features that strongly represent intrinsic identity, and a CLIP image encoder to capture semantically rich features. We then employ the Q-Former", + "bbox": [ + 89, + 840, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "[27, 28, 64] to fuse these features, resulting in intrinsic identity representations enriched with high-frequency semantic information. To mitigate the impact of irrelevant features from CLIP, dropout [1, 22] is applied post-processing. Additionally, we follow [14] to concatenate the shallow, multiscale features from the facial recognition backbone (after interpolation) with the CLIP features. This approach ensures that the model captures essential intrinsic identity features while filtering out extraneous noise unrelated to identity. Finally, we apply cross-attention to facilitate interaction between this feature set and the visual tokens produced by each attention block of the pre-trained model, thereby enhancing the high-frequency information in the DiT:", + "bbox": [ + 511, + 90, + 906, + 287 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nZ _ {i} ^ {\\prime} = Z _ {i} + \\text {A t t e n t i o n} \\left(Q _ {i} ^ {v}, K _ {i} ^ {f}, V _ {i} ^ {f}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 297, + 906, + 316 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $i$ represents the layer number of the attention block, $Q^{v} = Z_{i}W_{i}^{q}$ , $K^{f} = FW_{i}^{k}$ , and $V^{f} = FW_{i}^{v}$ , where $Z_{i}$ is the visual token, $F$ represents the intrinsic identity features, and $W_{q}, W_{k}$ , and $W_{v}$ are trainable parameters. The objective function is changed to:", + "bbox": [ + 511, + 327, + 906, + 398 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c} = \\mathbb {E} _ {x _ {0}, t, y, f, \\epsilon} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(x _ {0}, t, \\tau_ {\\theta} (y), \\psi_ {\\theta} (f), \\varphi_ {\\theta} (f)\\right) \\| _ {2} ^ {2} \\right], \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 407, + 906, + 426 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\varphi_{\\theta}(\\cdot)$ is the local facial extractor.", + "bbox": [ + 511, + 434, + 777, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.3. Consistency Training Strategy", + "text_level": 1, + "bbox": [ + 511, + 458, + 769, + 473 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During training, we randomly select a frame from the training frames and apply the Crop & Align [10] to extract the facial region as reference images, which is subsequently used as an identity-control signal, alongside the text as control.", + "bbox": [ + 511, + 477, + 905, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Coarse-to-Fine Training. Compared to Identity-preserving image generation, video generation requires maintaining consistency in both spatial and temporal dimensions, ensuring that high and low-frequency facial information matches the reference image. To mitigate the complexity of training, we propose a hierarchical strategy where the model learns information globally before refining it locally. In the coarse-grained phase (e.g., corresponding Finding 1), we employ the global facial extractor, enabling the model to prioritize low-frequency features, such as facial contours and proportions, thereby ensuring rapid acquisition of identity information from the reference image and consistency across the video sequence. In the fine-grained phase (e.g. corresponding to Finding 2), the local facial extractor shifts the model's focus to high-frequency details, such as the texture details of eyes and lips (e.g., intrinsic identification), improving the fidelity of facial expressions and the overall similarity of the generated face.", + "bbox": [ + 511, + 537, + 906, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Dynamic Mask Loss. The objective of our task is to ensure that the identity of the person in the generated video remains consistent with the input reference image. However, Equation 1 considers the entire scene, encompassing both high- and low-frequency identity information as well as redundant background content, which introduces noise that", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12981", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/237290a6b297b7c75a39cec96650554bfd2c89fc1bf7c86b8c925d0f983ba3d2.jpg", + "image_caption": [ + "Figure 3. User Study between ConsisID and state-of-the-art methods. ConsisID is preferred by voters in all dimensions." + ], + "image_footnote": [], + "bbox": [ + 119, + 88, + 457, + 256 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "interferes with model training. To address this, we propose to focus the model's attention on face regions. Specifically, we first extract the facial mask from the video, apply trilinear interpolation to map it to the latent space, and finally use this mask to constrain the computation of $\\mathcal{L}_c$ :", + "bbox": [ + 89, + 315, + 483, + 391 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {d} = M \\odot \\mathcal {L} _ {c}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 233, + 398, + 482, + 414 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $M$ represents a mask with the same shape as $\\epsilon$ . However, if Equation 5 is used as the supervisory signal for all training data, the model may fail to generate a natural background during inference. To mitigate this issue, we apply Equation 5 with a probability $p$ of $\\alpha$ , resulting in:", + "bbox": [ + 89, + 421, + 483, + 497 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {e} = \\left\\{ \\begin{array}{l l} \\mathcal {L} _ {d}, & \\text {i f} p > \\alpha \\\\ \\mathcal {L} _ {c}, & \\text {i f} p \\leq \\alpha \\end{array} \\right. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 503, + 482, + 544 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dynamic Cross-Face Loss. After training with Equation 6, we observed that the model struggled to generate satisfactory results for persons not present in the data domain during inference. This issue arises because the model, trained exclusively on faces from the training frames, tends to overfit by adopting a \"copy-paste\" shortcut—essentially replicating the reference image without alteration. To improve the model's generalization capability, we introduce slight Gaussian noise $\\zeta$ to the reference images and use cross-face (e.g., reference images are sourced from video frames outside the training frames) as inputs with probability $\\beta$ :", + "bbox": [ + 89, + 547, + 483, + 715 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {f} = \\left\\{ \\begin{array}{l l} \\mathcal {L} _ {e} & \\text {w h e r e} x _ {0} \\cdot \\zeta , \\quad \\text {i f} p > \\beta \\\\ \\mathcal {L} _ {e} & \\text {w h e r e} x _ {c} \\cdot \\zeta , \\quad \\text {i f} p \\leq \\beta \\end{array} \\right. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 720, + 482, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $x_0$ is the reference image extracted from the training frames, and $x_{c}$ is extracted from outside the training frames.", + "bbox": [ + 89, + 767, + 482, + 797 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 808, + 223, + 825 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Setup", + "text_level": 1, + "bbox": [ + 89, + 833, + 173, + 849 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details. ConsisID selects DiT-based generation architectures CogVideoX-5B [66] as our baseline for validation. We use an in-house human-centric", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/8175fcae909c6e9ebb18b4a75ac0860eaaacd961bb39c585fdff38493f9aade9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
FaceSim-Arc ↑FaceSim-Cur ↑CLIPScore ↑FID ↓
ID-Animator [15]0.320.3324.97117.46
ConsisID0.580.6027.93151.82
", + "bbox": [ + 517, + 88, + 903, + 133 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Quantitative comparison with state-of-the-art methods. ConsID achieve well-aligned results across most metrics. \"↓\" denotes lower is better. \"↑\" denotes higher is better.", + "bbox": [ + 511, + 143, + 906, + 186 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "dataset for training, which differs from previous datasets [40, 57, 68] that focus only on the face. In the training phase, we set the resolution to $480 \\times 720$ and extracted 49 consecutive frames at a stride of 3 from each video as training data. We set the batch size to 80, the learning rate to $3 \\times 10^{-6}$ , and the total number of training steps to $1.8\\mathrm{k}$ . The randomly discarded text rate is set to 0.1, with AdamW serving as an optimizer and cosine_with_restarts as a learning rate scheduler. The training strategy is the same as Section 3.2.3. We set $\\alpha$ and $\\beta$ in the dynamic cross-face loss $(\\mathcal{L}_e)$ and dynamic mask loss $(\\mathcal{L}_f)$ to 0.5, respectively. In the inference phase, we employ DPM [51] with a sampling step of 50, and a classifier free guidance ratio of 6.0. For more details and results, please refer to Appendix.", + "bbox": [ + 511, + 205, + 906, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Benchmark. Since there is an absence of an evaluation dataset, we select 30 persons who were not included in the training data and sourced five high-quality images for each ID from the internet. We then design 90 distinct prompts, encompassing a variety of expressions, actions, and backgrounds for evaluation. Building on previous works [15, 42], we evaluate four dimensions: (1). Identity Preservation: We use FaceSim-Arc [10] and introduce FaceSim-Cur, which assesses identity preservation by measuring feature differences between face regions in the generated videos and those in real face images within the ArcFace [10] and CurricularFace [23] feature spaces. (2) Visual Quality: We utilize FID [18] by calculating feature differences in the face regions between the generated frames and real face images within the InceptionV3 [53] feature space. (3) Text Relevance: We utilize CLIPScore [17] to measure the similarity between the generated videos and the input prompts. (4). Motion Amplitude: Due to the lack of reliable metrics [24, 73], we evaluate through the user study.", + "bbox": [ + 511, + 417, + 908, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Qualitative Analysis", + "text_level": 1, + "bbox": [ + 511, + 712, + 705, + 728 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we compare our method, ConsisID, with ID-Animator [15] (e.g., the only available open-source model) for tuning-free IPT2V tasks. We randomly select images and text prompts of four individuals for qualitative analysis, all of which are absent from the training data. As shown in Figure 4, ID-Animator cannot generate human body parts beyond the face and is unable to generate complex actions or backgrounds in response to text prompts (e.g., action, attribute, background), which significantly limits its practical application. In addition, the preservation of the identity is inadequate; for example, in case 1, the reference image", + "bbox": [ + 511, + 734, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12982", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f2181427b3d85c505fa48e2347daf297da2ffd5a21badf93ca0264b3f610026d.jpg", + "image_caption": [ + "Figure 4. Qualitative analysis between ConsisID and ID-Animator [15]. ID-Animator can only generate videos of the face region, and the identity Preservation is poor (e.g., shape, texture). Additionally, it cannot generate specified content according to the text prompt (e.g., action, decoration, background). ConsisID achieves advantages in identity preservation, visual quality, motion amplitude, and text relevance. Moreover, our ConsistID can generate more frames rather than ID-Animator (49 $480 \\times 720$ p frames v.s. 16 $512 \\times 512$ p frames)." + ], + "image_footnote": [], + "bbox": [ + 163, + 87, + 834, + 599 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "appears to be processed with skin smoothing. In case 2, wrinkles have been introduced which detract from the aesthetic quality. In cases 3 and 4, the face is distorted due to the lack of low frequency information, which compromises identity consistency. In contrast, the proposed ConsisID consistently produces high-quality, realistic videos that accurately match the reference identity and adhere to prompt.", + "bbox": [ + 89, + 686, + 482, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Quantitative Analysis", + "text_level": 1, + "bbox": [ + 89, + 801, + 294, + 819 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We present a comprehensive quantitative evaluation of different methods, with results displayed in Table 1. Consistent with Figure 4, our method outperforms state-of-the-art methods across five metrics. For identity preservation, ConsisID achieves a higher score by designing appropri", + "bbox": [ + 89, + 825, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ate identity signals for DiT from a frequency perspective. By contrast, ID-Animator [15] is not optimized for IPT2V and only partially retains facial features, resulting in lower FaceSim-Arc [10] and FaceSim-Cur scores. For Text Relevance, ConsisID not only controls expressions via prompts but also adjusts actions and backgrounds, achieving higher CLIPScore [17]. Regarding visual quality, the FID is presented solely as a reference due to its limited alignment [31, 32, 41, 73] with human perception. Please refer to Figure 4 and 3 for qualitative analysis of the visual quality.", + "bbox": [ + 511, + 686, + 906, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. User Study", + "text_level": 1, + "bbox": [ + 511, + 848, + 635, + 864 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Building on previous work, we conduct a human evaluation using a binary voting strategy, with each questionnaire con", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12983", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bbc2af07b9696d973d07ea1692bd0d7a753454559e973d94a8eed714283f165e.jpg", + "image_caption": [ + "Figure 5. Effect of Different Components via Qualitative Analysis. Removing any component may result in the loss of high- or low-frequency facial information, or hinder the ability to modify video content based on the text prompt." + ], + "image_footnote": [], + "bbox": [ + 99, + 88, + 473, + 489 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "taining only 80 questions. Participants are required to view 40 video clips, a setup designed to improve both engagement and questionnaire validity. For the IPT2V task, each question requires participants to separately judge which option performs better in terms of Identity Preservation, Visual Quality, Text Alignment, and Motion Amplitude. This composition ensures the accuracy of the human evaluation. Owing to the extensive participant base required for this evaluation, we successfully gathered 103 valid questionnaires. The results, depicted in Figure 3, demonstrate a significant superiority of our method over ID-Animator [15], verifying the effectiveness of the designed DiT for IPT2V generation.", + "bbox": [ + 89, + 582, + 482, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Effect of the Identity Signal Injection in DiT", + "text_level": 1, + "bbox": [ + 89, + 772, + 465, + 789 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To assess the effectiveness of Finding 1 and Finding 2, we perform ablation studies on different methods of injecting control signals into DiT. Specifically, these experiments involved (a) injecting only low-frequency face information with key points into the noise latent, (b) injecting only high-frequency face signals within the attention block, (c) combining (a) and (b), (d) based on (c), but the low-frequency", + "bbox": [ + 89, + 794, + 482, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a4e4ec88eb66f94d25bba40665477d32a1841b3a6105327195353b440640a845.jpg", + "image_caption": [ + "Figure 6. Effect of Different Control Signal Injection Way via Qualitative Analysis. Only (c), which injects both high & low-freq face signals into the suitable location, performs best." + ], + "image_footnote": [], + "bbox": [ + 521, + 87, + 897, + 502 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ecd90b4dcd2df5ca06fab6db8ddc6c83f769c445101a2993969049e171352c63.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
FaceSim-Arc ↑FaceSim-Cur ↑CLIPScore ↑FID ↓
w/o GFE0.050.0534.86269.88
w/o LFE0.660.6834.48104.34
w/o CFT0.540.5834.47144.62
w/o DML0.620.6734.23187.78
w/o DCL0.650.6932.21117.80
ConsisID0.730.7536.77127.42
", + "bbox": [ + 517, + 571, + 903, + 669 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Effect of Local Facial Extractor (LFE), Global Facial Extractor (GFE), coarse-to-fine training (CFT), dynamic mask loss (DML) and dynamic cross-face loss (DCL) by Automatic Metrics. Removing any of the above methods significantly reduces identity preservation, text relevance, and visual quality.", + "bbox": [ + 511, + 678, + 906, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "face information does not contain key points, and $(\\mathbf{e} - \\mathbf{f})$ based on (c), but the high-frequency signal is injected at the output or input of the attention block. (g) injecting only high-frequency face signals before the attention block. To reduce overhead, for each identity, we only select 2 reference images each with 90 text prompts for the evaluation. The results are shown in Figure 6 and Table 3. For Finding $I$ , we observe that only injecting high-frequency signals (a)", + "bbox": [ + 511, + 779, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12984", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d0a636a1eb123b39bfe870fe69bbc7d637cda5a173efb0abb6ae04f544db96aa.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
PlanFaceSim-Arc ↑FaceSim-Cur ↑CLIPScore ↑FID ↓
a0.050.0534.86269.88
b0.660.6834.48104.34
c0.730.7536.77127.42
d0.640.6830.69177.65
e0.620.6633.61164.15
funstable training process
gunstable training process
", + "bbox": [ + 94, + 88, + 480, + 204 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Effect of Different Control Signal Injection Way via Quantitative Analysis. Only plan c, which injects both high and low-frequency face information into the model, performs best.", + "bbox": [ + 89, + 214, + 482, + 257 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "greatly increases the training difficulty, causing the model to fail to converge due to the lack of low-frequency signal injection. In addition, the inclusion of facial key points (d) allows a greater focus on low-frequency information, thereby facilitating training and improving model performance. For Finding 2, when only low-frequency signals are injected (b), the model lacks high-frequency information. This reliance on low-frequency signals causes the generated face in the video to copy the reference image, making it difficult to control facial expressions, movements, and other features through prompts. Furthermore, injecting identity signals into the attention block input (f - g) disrupts the intended frequency domain distribution of DiT, resulting in a gradient explosion. Embedding control signals in the attention block (c) is preferable to embedding them in the output (e) because attention block processes predominantly low-frequency information. By embedding high-frequency information internally, the attention block is guided to highlight intrinsic facial features, whereas injecting it into the output merely concatenates features without directing focus, reducing DiT's modeling capacity. Moreover, we apply a Fourier transform to the generated videos (only the face region) to visually compare the influence of different components to extract facial information. As shown in Figure 7, the Fourier spectrum and the log amplitude of the Fourier transform reveal that injecting high or low-frequency signals can indeed enhance the corresponding frequency information of the generated face. Moreover, the low-frequency signal can be further enhanced by matching with the face key points, and injecting the high-frequency signal into the attention block has the highest feature utilization rate. Our method (c) shows strongest high and low frequency, further validating the efficiency benefit from Findings 1 and 2.", + "bbox": [ + 91, + 280, + 483, + 781 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Ablation on the Consistency Training Strategy", + "text_level": 1, + "bbox": [ + 89, + 787, + 482, + 805 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To reduce overhead, for each identity, we only select 2 reference images for the following experiments. To demonstrate the benefits of the proposed consistency training strategy, we perform ablation experiments on coarse-to-fine training (CFT), dynamic mask loss $\\mathcal{L}_e$ (DML), and dynamic crossface loss $\\mathcal{L}_f$ (DCL), with the results presented in Figure", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/65f5d2294d04dd308479a5f1c85c79ba40e4ffc3e649d056ce20ff6373cad2f4.jpg", + "image_caption": [ + "(a) only high-freq" + ], + "image_footnote": [], + "bbox": [ + 532, + 88, + 630, + 160 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/00d4ad5dfc734192981137bf8c8ffcc15c55685a44ab405736cb53e07c0448bd.jpg", + "image_caption": [ + "(b) only low-freq" + ], + "image_footnote": [], + "bbox": [ + 656, + 89, + 754, + 161 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7ae2c20bdfd58558e5b664029efe9b17a3c80dae5f5588743999e2841a4c32e0.jpg", + "image_caption": [ + "(c) high & low freq" + ], + "image_footnote": [], + "bbox": [ + 782, + 89, + 877, + 161 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2497ec5f968a9e5b83550d1f63698a3c2b580d2106247ad2b398c9c5a9e3d7aa.jpg", + "image_caption": [ + "(d) high & low† freq" + ], + "image_footnote": [], + "bbox": [ + 532, + 172, + 630, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5650e498907d8bcca60ce89715d5620898a31a3ea6a7ea27f5eb5b79cd51311c.jpg", + "image_caption": [ + "(f) Relative log amplitudes of Fourier transformed videos" + ], + "image_footnote": [], + "bbox": [ + 640, + 176, + 880, + 327 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/64bb93253f7a3f9033c3b75436b212918d74de990232fcfa5ac589ca23e6c566.jpg", + "image_caption": [ + "(e) high+ & low freq", + "Figure 7. (a - e) Fourier spectrum of different id signal injection. The center area represents low frequencies and the surrounding area represents high frequencies. (f) Relative log amplitudes of Fourier transformed generated videos. A larger response value indicates a higher inclusion of frequency information. (a - f) verify the effect of our frequency decomposition." + ], + "image_footnote": [], + "bbox": [ + 532, + 258, + 630, + 330 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5 and Table 2. When CFT is removed, GFE and LFE exhibit competing behaviors, complicating the model's ability to prioritize high and low-frequency information accurately, leading to convergence at suboptimal points. Removing DML required the model to simultaneously focus on both foreground and background elements, with background noise negatively affecting training and reducing facial consistency. Similarly, the exclusion of DCL impaired the generalization capability, reducing fidelity for faces, not in the training set and reducing its effectiveness in generating identity-preserving videos as intended.", + "bbox": [ + 511, + 460, + 906, + 627 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 640, + 633, + 656 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present ConsisID, a unified framework for keeping faces consistent in video generation by frequency decomposition. It can seamlessly integrate into existing DiT-based text-to-video models, for generating high-quality, editable, consistent identity-preserving videos. Extensive experiments show that ConsisID outperforms the current state-of-the-art identity-preserving T2V models. It reveals that our frequency-aware heuristic DiT-based control scheme is an optimal solution for IPT2V generation.", + "bbox": [ + 511, + 665, + 906, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgments", + "text_level": 1, + "bbox": [ + 511, + 815, + 692, + 832 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We thank all the anonymous reviewers for their constructive comments. This work was supported in part by the Natural Science Foundation of China (No. 62202014, 62332002, 62425101, 62088102).", + "bbox": [ + 511, + 840, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12985", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Hassan Akbari, Liangzhe Yuan, Rui Qian, Wei-Hong Chuang, Shih-Fu Chang, Yin Cui, and Boqing Gong. Vatt: Transformers for multimodal self-supervised learning from raw video, audio and text. NeurIPS, 34:24206-24221, 2021.", + "[2] Dosovitskiy Alexey. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv: 2010.11929, 2020.", + "[3] Jiawang Bai, Li Yuan, Shu-Tao Xia, Shuicheng Yan, Zhifeng Li, and Wei Liu. Improving vision transformers by revisiting high-frequency components. In ECCV, pages 1-18. Springer, 2022.", + "[4] Fan Bao, Shen Nie, Kaiwen Xue, Yue Cao, Chongxuan Li, Hang Su, and Jun Zhu. All are worth words: A vit backbone for diffusion models. In CVPR, pages 22669-22679, 2023.", + "[5] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023.", + "[6] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. In openai, 2024.", + "[7] Hila Chefer, Shiran Zada, Roni Paiss, Ariel Ephrat, Omer Tov, Michael Rubinstein, Lior Wolf, Tali Dekel, Tomer Michaeli, and Inbar Mosseri. Still-moving: Customized video generation without customized video data. arXiv preprint arXiv:2407.08674, 2024.", + "[8] Li Chen, Mengyi Zhao, Yiheng Liu, Mingxu Ding, Yangyang Song, Shizun Wang, Xu Wang, Hao Yang, Jing Liu, Kang Du, et al. Photovverse: Tuning-free image customization with text-to-image diffusion models. arXiv preprint arXiv:2309.05793, 2023.", + "[9] Soon Yau Cheong, Duygu Ceylan, Armin Mustafa, Andrew Gilbert, and Chun-Hao Paul Huang. Boosting camera motion control for video diffusion transformers. arXiv preprint arXiv:2410.10802, 2024.", + "[10] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In CVPR, pages 4690-4699, 2019.", + "[11] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022.", + "[12] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. ICLR, 2024.", + "[13] Zinan Guo, Yanze Wu, Zhuowei Chen, Lang Chen, and Qian He. Pulid: Pure and lightning id customization via contrastive alignment. arXiv preprint arXiv:2404.16022, 2024.", + "[14] Junjie He, Yifeng Geng, and Liefeng Bo. Unipotrait: A unified framework for identity-preserving single-" + ], + "bbox": [ + 93, + 114, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and multi-human image personalization. arXiv preprint arXiv:2408.05939, 2024.", + "[15] Xuanhua He, Quande Liu, Shengju Qian, Xin Wang, Tao Hu, Ke Cao, Keyu Yan, Man Zhou, and Jie Zhang. Id-Animator: Zero-shot identity-preserving human video generation. arXiv preprint arXiv:2404.15275, 2024.", + "[16] Zecheng He, Bo Sun, Felix Juefei-Xu, Haoyu Ma, Ankit Ramchandani, Vincent Cheung, Siddharth Shah, Anmol Kalia, Harihar Subramanyam, Alireza Zareian, et al. Imagine yourself: Tuning-free personalized image generation. arXiv preprint arXiv:2409.13346, 2024.", + "[17] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021.", + "[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS, 2017.", + "[19] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022.", + "[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 33:6840-6851, 2020.", + "[21] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021.", + "[22] Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Q Weinberger. Deep networks with stochastic depth. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 646-661. Springer, 2016.", + "[23] Yuge Huang, Yuhan Wang, Ying Tai, Xiaoming Liu, Pengcheng Shen, Shaoxin Li, Jilin Li, and Feiyue Huang. Curricularface: adaptive curriculum learning loss for deep face recognition. In CVPR, pages 5901-5910, 2020.", + "[24] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024.", + "[25] Dan Kondratyuk, Lijun Yu, Xiuye Gu, José Lezama, Jonathan Huang, Rachel Hornung, Hartwig Adam, Hassan Akbari, Yair Alon, Vighnesh Birodkar, et al. Videopoet: A large language model for zero-shot video generation. arXiv preprint arXiv:2312.14125, 2023.", + "[26] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In CVPR, pages 1931-1941, 2023.", + "[27] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, pages 12888-12900. PMLR, 2022.", + "[28] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with" + ], + "bbox": [ + 516, + 93, + 903, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12986", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "frozen image encoders and large language models. In ICML, pages 19730-19742. PMLR, 2023.", + "[29] Zhen Li, Mingdeng Cao, Xintao Wang, Zhongang Qi, Ming-Ming Cheng, and Ying Shan. Photomaker: Customizing realistic human photos via stacked id embedding. In CVPR, pages 8640-8650, 2024.", + "[30] Bin Lin, Yunyang Ge, Xinhua Cheng, Zongjian Li, Bin Zhu, Shaodong Wang, Xianyi He, Yang Ye, Shenghai Yuan, Liuhan Chen, et al. Open-sora plan: Open-source large video generation model. arXiv preprint arXiv:2412.00131, 2024.", + "[31] Yaofang Liu, Xiaodong Cun, Xuebo Liu, Xintao Wang, Yong Zhang, Haoxin Chen, Yang Liu, Tieyong Zeng, Raymond Chan, and Ying Shan. Evalcrafter: Benchmarking and evaluating large video generation models. arXiv preprint arXiv:2310.11440, 2023.", + "[32] Yuanxin Liu, Lei Li, Shuhuai Ren, Rundong Gao, Shicheng Li, Sishuo Chen, Xu Sun, and Lu Hou. Fetv: A benchmark for fine-grained evaluation of open-domain text-to-video generation. NeurIPS, 36, 2024.", + "[33] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024.", + "[34] Yue Ma, Xiaodong Cun, Yingqing He, Chenyang Qi, Xintao Wang, Ying Shan, Xiu Li, and Qifeng Chen. Magic-stick: Controllable video editing via control handle transformations. arXiv preprint arXiv:2312.03047, 2023.", + "[35] Yue Ma, Yingqing He, Xiaodong Cun, Xintao Wang, Siran Chen, Xiu Li, and Qifeng Chen. Follow your pose: Pose-guided text-to-video generation using pose-free videos. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4117-4125, 2024.", + "[36] Yue Ma, Yingqing He, Hongfa Wang, Andong Wang, Chenyang Qi, Chengfei Cai, Xiu Li, Zhifeng Li, Heung-Yeung Shum, Wei Liu, et al. Follow-your-click: Open-domain regional image animation via short prompts. arXiv preprint arXiv:2403.08268, 2024.", + "[37] Yue Ma, Hongyu Liu, Hongfa Wang, Heng Pan, Yingqing He, Junkun Yuan, Ailing Zeng, Chengfei Cai, Heung-Yeung Shum, Wei Liu, et al. Follow-your-emoji: Fine-controllable and expressive freestyle portrait animation. In SIGGRAPH Asia 2024 Conference Papers, pages 1–12, 2024.", + "[38] Ze Ma, Daquan Zhou, Chun-Hsiao Yeh, Xue-She Wang, Xiuyu Li, Huanrui Yang, Zhen Dong, Kurt Keutzer, and Jiashi Feng. Magic-me: Identity-specific video customized diffusion. arXiv preprint arXiv:2402.09368, 2024.", + "[39] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453, 2023.", + "[40] Arsha Nagrani, Joon Son Chung, Weidi Xie, and Andrew Zisserman. Voxceleb: Large-scale speaker verification in the wild. Computer Speech & Language, 60:101027, 2020.", + "[41] Mayu Otani, Riku Togashi, Yu Sawai, Ryosuke Ishigami, Yuta Nakashima, Esa Rahtu, Janne Heikkila, and Shin'ichi Satoh. Toward verifiable and reproducible human evaluation" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "for text-to-image generation. In CVPR, pages 14277-14286, 2023.", + "[42] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, et al. Movie gen: A cast of media foundation models. arXiv preprint arXiv:2410.13720, 2024.", + "[43] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021.", + "[44] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, pages 8821-8831, 2021.", + "[45] Chen Rao, Guangyuan Li, Zehua Lan, Jiakai Sun, Junsheng Luan, Wei Xing, Lei Zhao, Huaizhong Lin, Jianfeng Dong, and Dalong Zhang. Rethinking video deblurring with wavelet-aware dynamic transformer and diffusion model. arXiv preprint arXiv:2408.13459, 2024.", + "[46] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bjorn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022.", + "[47] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, pages 22500-22510, 2023.", + "[48] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdream booth: Hypernetworks for fast personalization of text-to-image models. In CVPR, pages 6527-6536, 2024.", + "[49] Yujun Shi, Jun Hao Liew, Hanshu Yan, Vincent YF Tan, and Jiashi Feng. Instadrag: Lightning fast and accurate drag-based image editing emerging from videos. arXiv preprint arXiv:2405.13722, 2024.", + "[50] Chenyang Si, Ziqi Huang, Yuming Jiang, and Ziwei Liu. Freeu: Free lunch in diffusion u-net. In CVPR, pages 4733-4743, 2024.", + "[51] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, pages 2256-2265. PMLR, 2015.", + "[52] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv, 2020.", + "[53] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, pages 2818-2826, 2016.", + "[54] Zhenyu Tang, Junwu Zhang, Xinhua Cheng, Wangbo Yu, Chaoran Feng, Yatian Pang, Bin Lin, and Li Yuan. Cycle3d: High-quality and consistent image-to-3d generation via generation-reconstruction cycle. arXiv preprint arXiv:2407.19548, 2024.", + "[55] Yuchuan Tian, Zhijun Tu, Hanting Chen, Jie Hu, Chao Xu, and Yunhe Wang. U-dits: Downsample tokens in u-shaped" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12987", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "diffusion transformers. arXiv preprint arXiv:2405.02730, 2024.", + "[56] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, and Anthony Chen. Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519, 2024.", + "[57] Ting-Chun Wang, Arun Mallya, and Ming-Yu Liu. One-shot free-view neural talking-head synthesis for video conferencing. In CVPR, pages 10039-10049, 2021.", + "[58] Zhao Wang, Aoxue Li, Enze Xie, Lingting Zhu, Yong Guo, Qi Dou, and Zhenguo Li. Customvideo: Customizing text-to-video generation with multiple subjects. arXiv preprint arXiv:2401.09962, 2024.", + "[59] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. In ICCV, pages 15943-15953, 2023.", + "[60] Yujie Wei, Shiwei Zhang, Zhiwu Qing, Hangjie Yuan, Zhiheng Liu, Yu Liu, Yingya Zhang, Jingren Zhou, and Hongming Shan. Dreamvideo: Composing your dream videos with customized subject and motion. In CVPR, pages 6537-6549, 2024.", + "[61] Jianzong Wu, Xiangtai Li, Yanhong Zeng, Jiangning Zhang, Qianyu Zhou, Yining Li, Yunhai Tong, and Kai Chen. Motionbooth: Motion-aware customized text-to-video generation. arXiv preprint arXiv:2406.17758, 2024.", + "[62] Jing Xiong, Gongye Liu, Lun Huang, Chengyue Wu, Taiqiang Wu, Yao Mu, Yuan Yao, Hui Shen, Zhongwei Wan, Jinfa Huang, et al. Autoregressive models in vision: A survey. arXiv preprint arXiv:2411.05902, 2024.", + "[63] Jiaqi Xu, Xinyi Zou, Kunzhe Huang, Yunkuo Chen, Bo Liu, MengLi Cheng, Xing Shi, and Jun Huang. Easyanimate: A high-performance long video generation method based on transformer architecture. arXiv preprint arXiv:2405.18991, 2024.", + "[64] Le Xue, Manli Shu, Anas Awadalla, Jun Wang, An Yan, Senthil Purushwalkam, Honglu Zhou, Viraj Prabhu, Yutong Dai, Michael S Ryoo, et al. xgen-mm (blip-3): A family of open large multimodal models. arXiv preprint arXiv:2408.08872, 2024.", + "[65] Shuo Yang, Kun-Peng Ning, Yu-Yang Liu, Jia-Yu Yao, Yong-Hong Tian, Yi-Bing Song, and Li Yuan. Is parameter collision hindering continual learning in llms? arXiv preprint arXiv:2410.10179, 2024.", + "[66] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024.", + "[67] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023.", + "[68] Jianhui Yu, Hao Zhu, Liming Jiang, Chen Change Loy, Weidong Cai, and Wayne Wu. Celebv-text: A large-scale facial text-video dataset. In CVPR, pages 14805-14814, 2023.", + "[69] Wangbo Yu, Chaoran Feng, Jiye Tang, Xu Jia, Li Yuan, and Yonghong Tian. Evagaussians: Event stream assisted" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "gaussian splatting from blurry images. arXiv preprint arXiv:2405.20224, 2024.", + "[70] Wangbo Yu, Jinbo Xing, Li Yuan, Wenbo Hu, Xiaoyu Li, Zhipeng Huang, Xiangjun Gao, Tien-Tsin Wong, Ying Shan, and Yonghong Tian. Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis. arXiv preprint arXiv:2409.02048, 2024.", + "[71] Yongsheng Yu, Ziyun Zeng, Hang Hua, Jianlong Fu, and Jiebo Luo. Promptfix: You prompt and we fix the photo. arXiv preprint arXiv:2405.16785, 2024.", + "[72] Shenghai Yuan, Jinfa Huang, Yujun Shi, Yongqi Xu, Ruijie Zhu, Bin Lin, Xinhua Cheng, Li Yuan, and Jiebo Luo. Mag-ictime: Time-lapse video generation models as metamorphic simulators. arXiv preprint arXiv:2404.05014, 2024.", + "[73] Shenghai Yuan, Jina Huang, Yongqi Xu, Yaoyang Liu, Shaofeng Zhang, Yujun Shi, Rui-Jie Zhu, Xinhua Cheng, Jiebo Luo, and Li Yuan. Chronomagic-bench: A benchmark for metamorphic evaluation of text-to-time-lapse video generation. NeurIPS, 37:21236-21270, 2025.", + "[74] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, pages 3836-3847, 2023.", + "[75] Zhenghao Zhang, Junchao Liao, Menghao Li, Long Qin, and Weizhi Wang. Tora: Trajectory-oriented diffusion transformer for video generation. arXiv preprint arXiv:2407.21705, 2024.", + "[76] Mingzhe Zheng, Yongqi Xu, Haojian Huang, Xuran Ma, Yexin Liu, Wenjie Shu, Yatian Pang, Feilong Tang, Qifeng Chen, Harry Yang, et al. Videogen-of-thought: A collaborative framework for multi-shot video generation. arXiv preprint arXiv:2412.02259, 2024.", + "[77] Zangwei Zheng, Xiangyu Peng, and Yang You. Open-sora: Democratizing efficient video production for all. In *Github*, 2024.", + "[78] Yuan Zhou, Qiuyue Wang, Yuxuan Cai, and Huan Yang. Allegro: Open the black box of commercial-level video generation model. arXiv preprint arXiv:2410.15458, 2024." + ], + "bbox": [ + 516, + 92, + 903, + 614 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "12988", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Identity-Preserving Text-to-Video Generation by Frequency Decomposition/16b4d9e1-d6f9-4df5-9c6a-3949d63cad2d_model.json b/2025/Identity-Preserving Text-to-Video Generation by Frequency Decomposition/16b4d9e1-d6f9-4df5-9c6a-3949d63cad2d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..39599c9d4dc43337f7f790d96ec2ff3607226d61 --- /dev/null +++ b/2025/Identity-Preserving Text-to-Video Generation by Frequency Decomposition/16b4d9e1-d6f9-4df5-9c6a-3949d63cad2d_model.json @@ -0,0 +1,2730 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.02 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + }, + { + "type": "header", + "bbox": [ + 0.325, + 0.017, + 0.725, + 0.032 + ], + "angle": 0, + "content": "Except for this watermark, it is identical to the accepted version;" + }, + { + "type": "header", + "bbox": [ + 0.294, + 0.032, + 0.756, + 0.047 + ], + "angle": 0, + "content": "the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.12, + 0.131, + 0.88, + 0.154 + ], + "angle": 0, + "content": "Identity-Preserving Text-to-Video Generation by Frequency Decomposition" + }, + { + "type": "text", + "bbox": [ + 0.255, + 0.18, + 0.744, + 0.219 + ], + "angle": 0, + "content": "Shenghai Yuan\\(^{1,4}\\), Jina Fu Huang\\(^{3}\\), Xianyi He\\(^{1,4}\\), Yunyang Ge\\(^{1,4}\\), Yujun Shi\\(^{5}\\), Liuhan Chen\\(^{1,4}\\), Jiebo Luo\\(^{3}\\), Li Yuan\\(^{1,2,\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.211, + 0.233, + 0.787, + 0.271 + ], + "angle": 0, + "content": "1 Peking University, 2 Peng Cheng Laboratory, 3 University of Rochester, 4 Rabbitpre Intelligence, 5 National University of Singapore" + }, + { + "type": "text", + "bbox": [ + 0.149, + 0.29, + 0.845, + 0.324 + ], + "angle": 0, + "content": "{yuanshenghai, HeXianyi, yunyang, chenliuhan}@stu.pku.edu.cn, shi.yujun@u.nus.edu, yuanli-ece@pku.edu.cn, {jhuang90@ur, jluo@cs}.rochester.edu" + }, + { + "type": "title", + "bbox": [ + 0.26, + 0.341, + 0.738, + 0.359 + ], + "angle": 0, + "content": "https://oku-yuangroup.github.io/ConsisID" + }, + { + "type": "image_caption", + "bbox": [ + 0.133, + 0.376, + 0.207, + 0.387 + ], + "angle": 0, + "content": "Reference Image" + }, + { + "type": "image", + "bbox": [ + 0.126, + 0.389, + 0.215, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.371, + 0.862, + 0.389 + ], + "angle": 0, + "content": "a man standing next to an airplane, engaged in a conversation on his cell phone ... wearings sunglasses and a black top ... talking seriously ... airplane has a green striperunning along its side, and there is a large engine visible behind his ..." + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.39, + 0.484, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.484, + 0.39, + 0.613, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.613, + 0.39, + 0.743, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.743, + 0.39, + 0.874, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.133, + 0.471, + 0.207, + 0.482 + ], + "angle": 0, + "content": "Reference Image" + }, + { + "type": "image", + "bbox": [ + 0.126, + 0.485, + 0.215, + 0.554 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.469, + 0.871, + 0.485 + ], + "angle": 0, + "content": "a young woman with long blonde hair ... front of a lush, green bush...white flowers ... wearing a black top ... smiling and looking at the camera while gently touching the flowers on the bush ... then bends downslightly and smells one of the flower..." + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.486, + 0.352, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.486, + 0.48, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.481, + 0.486, + 0.611, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.486, + 0.742, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.743, + 0.486, + 0.874, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.134, + 0.565, + 0.207, + 0.575 + ], + "angle": 0, + "content": "Reference Image" + }, + { + "type": "image", + "bbox": [ + 0.126, + 0.577, + 0.215, + 0.648 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.561, + 0.868, + 0.578 + ], + "angle": 0, + "content": "... a man ... his gaze focused and intense as he holds the basketball ... dressed in athletic gear ... The basketball is positioned firmly in his hands as he gradually puts it down ... urban street court, with the fading light of dusk casting a soft glow over the scene. The man's serious expression and steady grip on the ball ..." + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.578, + 0.352, + 0.647 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.578, + 0.481, + 0.647 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.481, + 0.578, + 0.611, + 0.648 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.578, + 0.742, + 0.648 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.743, + 0.578, + 0.874, + 0.648 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.133, + 0.657, + 0.207, + 0.667 + ], + "angle": 0, + "content": "Reference Image" + }, + { + "type": "image", + "bbox": [ + 0.126, + 0.671, + 0.215, + 0.739 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.651, + 0.871, + 0.671 + ], + "angle": 0, + "content": "a young boy sitting at a table, eating a piece of food. He appears to be enjoying his meal, as he takes a bite and chews it. The boy is wearing a blue shirt and has short hair...background is dark, with some light coming from the left side of the frame. There is a straw visible on the right side of the frame ..." + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.671, + 0.352, + 0.739 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.671, + 0.48, + 0.739 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.481, + 0.671, + 0.611, + 0.739 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.671, + 0.742, + 0.739 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.743, + 0.671, + 0.874, + 0.739 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.75, + 0.908, + 0.78 + ], + "angle": 0, + "content": "Figure 1. Examples of identity-preserving video generation (IPT2V) by our ConsisID. Given a reference image, our method can generate realistic and personalized human-centered videos while preserving identity. Red indicates that attributes in long instructions." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.794, + 0.327, + 0.809 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Identity-preserving text-to-video (IPT2V) generation aims to create high-fidelity videos with consistent human identity. It is an important task in video generation but remains an open problem for generative models. This paper pushes the technical frontier of IPT2V in two directions that" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.903 + ], + "angle": 0, + "content": "have not been resolved in the literature: (1) A tuning-free pipeline without tedious case-by-case finetuning, and (2) A frequency-aware heuristic identity-preserving Diffusion Transformer (DiT)-based control scheme. To achieve these goals, we propose ConsisID, a tuning-free DiT-based controllable IPT2V model to keep human-identity consistent in the generated video. Inspired by prior findings in fre" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12978" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.092, + 0.483, + 0.228 + ], + "angle": 0, + "content": "quency analysis of vision/diffusion transformers, it employs identity-control signals base on frequency domain, since facial features can be decomposed into low-frequency global features (e.g., profile, proportions) and high-frequency intrinsic features (e.g., identity markers that remain unaffected by pose changes). Extensive experiments demonstrate that our frequency-aware heuristic scheme provides an optimal control solution for DiT-based models, making strides towards more effective IPT2V." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.257, + 0.223, + 0.273 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.282, + 0.483, + 0.433 + ], + "angle": 0, + "content": "Large-scale pre-trained video diffusion models [30, 66, 77, 78] have facilitated a variety of downstream applications [49, 54, 69, 70, 72, 74], particularly in identity-preserving text-to-video (IPT2V) [7, 38, 58, 60, 61]. However, existing methods face significant challenges, particularly the high overhead associated with the need for case-by-case finetuning, which diminishes their applicability. Within the open-source community, only the ID-Animator [15] can implement tuning-free IPT2V, but it can only generate videos similar to talking head [57] and has poor id preservation." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.434, + 0.483, + 0.735 + ], + "angle": 0, + "content": "Additionally, the above efforts are predominantly based on U-Net and cannot be adapted to the emerging DiT-based video model [30, 63, 66, 77, 78]. This challenge may stem from the inherent limitations of DiT compared to U-Net, including greater difficulty in training convergence and weakness in perceiving facial details. From some prior findings in frequency analysis of vision/diffusion transformers [2-4, 45, 50, 55, 71], we can know that the reason is: Finding 1: Shallow (e.g., low-level, low-frequency) features are essential for pixel-level prediction tasks in diffusion models, as they ease model training. U-Net facilitates model convergence by aggregating shallow features to the decoder via long skip connections, a mechanism that DiT does not incorporate; Finding 2: Transformers have limited perception of high-frequency information, which is important for preserving facial features. The encoder-decoder architecture of U-Net naturally possesses multi-scale features (e.g., richness in high-frequency), while DiT lacks a comparable structure. To develop a DiT-based control model, these must be addressed first. Please see Appendix for more details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.736, + 0.483, + 0.902 + ], + "angle": 0, + "content": "For ID-preserving video generation, the challenges stem from the requirement for each frame to incorporate both high-frequency (e.g., age- and make-up-independent identity markers) and low-frequency information (e.g., facial shape) derived from the reference image, which can just be used to make up for the DiT defects mentioned above. Therefore, we propose ConsisID, to keep the identity consistency in video generation by frequency decomposition, based on the previously Findings of DiT in frequency analysis. Thanks to the large-scale pre-trained DiT, we can use its powerful capabilities to achieve tuning-free effects." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.288 + ], + "angle": 0, + "content": "ConsisID decouples identity features into high- and low-frequency signals, which are injected into specific locations within the DiT, facilitating efficient IPT2V generation. Specifically, in line with Finding 1, we first convert the reference image and the facial key points to the low-frequency signal, then concatenate them with input noise latent to ease the training. Following Finding 2, we utilize a dual-tower feature extractor to capture high-frequency facial information, which is integrated with vision tokens within the transformer block, thereby enhancing the DiT's high-frequency perception capabilities. Finally, to transform the pre-trained model into an IPT2V model and improve its generalization, we further introduce a hierarchical training strategy." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.289, + 0.858, + 0.302 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.305, + 0.905, + 0.365 + ], + "angle": 0, + "content": "- We introduce ConsisID, a tuning-free identity-preserving DiT-based IPT2V model, which preserves the identity of the main subject of the video using control signals from frequency decomposition." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.365, + 0.905, + 0.425 + ], + "angle": 0, + "content": "- We propose a hierarchical training strategy, including coarse-to-fine training, dynamic mask loss, and dynamic cross-face loss, which work together to facilitate training and enhance generalization effectively." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.426, + 0.905, + 0.486 + ], + "angle": 0, + "content": "- Extensive experiments demonstrate that our ConsisID can generate high-quality, editable, consistent identity-preserving videos, benefiting from our frequency-aware identity-preserving T2V DiT-based control scheme." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.305, + 0.905, + 0.486 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.498, + 0.655, + 0.514 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.907, + 0.841 + ], + "angle": 0, + "content": "Tuning-based Identity-preserving T2V Models. Diffusion models are widely recognized for their strong generative capabilities [19, 34-37, 44, 72, 73], significantly advancing the development of identity-preserving generative models [8, 39, 59, 74]. Initially, the researchers used tuning-based methods to generate content that matched the input ID. This process requires finetuning pretrained model for each new person during inference. For example, DreamBooth [47] introduced a novel loss function to fine-tune the entire network, embedding identity information while preserving the original generative capabilities. LoRA [21], similar to DreamBooth [47], requires training only a small subset of network parameters. In contrast, Textual Inversion [11] freezes the pretrained network and embeds identity information into a trainable word embedding. Subsequent tuning-based methods, including both image and video models based on U-Net or DiT architectures [7, 26, 48, 58, 60-62, 65, 76], generally follow three main approaches. While these models demonstrate substantial effectiveness, the requirement to fine-tune for each new identity restricts their practical applicability." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Tuning-free Identity-preserving T2V Models. To address the issue of high resource consumption, several tuning-free diffusion models have recently emerged in the field of image generation [13, 14, 29, 56, 67]. These models" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12979" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.089, + 0.868, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.335, + 0.908, + 0.381 + ], + "angle": 0, + "content": "Figure 2. Overview of the proposed ConsisID. Based on Findings of DiT, low-frequency facial information is embedded into the shallow layers, while high-frequency information is incorporated into the vision tokens within the attention blocks. The ID-preserving Recipe is applied to ease training and improve generalization. The cross face, DropToken and Dropout are executed based on probability." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.405, + 0.485, + 0.753 + ], + "angle": 0, + "content": "do not require finetuning parameters for newly introduced IDs during inference. For instance, IP-Adapter [67] utilizes the CLIP [43] features of the identity image through cross-attention to guide the pretrained model in generating identity-preserving images. InstantID [56] extends this approach by replacing CLIP [43] features with Arcface [10] features and integrating a pose network to adjust facial proportions. Unlike these initial methods, which introduce control signals via visual tokens, PhotoMaker [29] and Imagine Yourself [16] leverage text tokens. Specifically, PhotoMaker [29] concatenates identity features obtained from the CLIP encoder [43] to the text embedding, while Imagine Yourself [16] uses element-wise addition for feature fusion. In the domain of video generation, only MovieGen [42] and ID-Animator [15] currently support ID-preserving text-to-video (IPT2V) generation. MovieGen is closed-source, whereas ID-Animator is open-source but uses a methodology similar to image models, leading to lower-quality identity preservation in the generated videos. We select the emerging DiT architecture [30, 33, 66, 77] and optimize it for IPT2V, drawing on conclusions from prior frequency analyses [2-4, 45, 50, 55]. This enables high-quality, editable, and consistent ID-preserving video generation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.763, + 0.226, + 0.781 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.788, + 0.233, + 0.803 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Diffusion Model. Text-to-video generation models usually utilize the diffusion paradigm, which gradually transforms noise \\(\\epsilon\\) into a video \\(x_0\\). Originally, denoising was conducted directly within the pixel space [20, 51, 52]; however, due to significant computational overheads, recent methods predominantly employ latent space [12, 25, 46," + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.405, + 0.787, + 0.42 + ], + "angle": 0, + "content": "72]. The optimization process is defined:" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.433, + 0.907, + 0.462 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {a} = \\mathbb {E} _ {x _ {0}, t, y, \\epsilon} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (x _ {0}, t, \\tau_ {\\theta} (y)) \\| _ {2} ^ {2} \\right], \\qquad (1)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.471, + 0.907, + 0.532 + ], + "angle": 0, + "content": "where \\(y\\) is text condition, \\(\\epsilon\\) is sampled from a standard normal distribution (e.g., \\(\\epsilon \\sim \\mathcal{N}(0,1)\\)), and \\(\\tau_{\\theta}(\\cdot)\\) is the text encoder. By replacing \\(x_0\\) with \\(\\mathcal{E}(x_0)\\), the latent diffusion is derived, which is used by ConsisID." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.533, + 0.909, + 0.791 + ], + "angle": 0, + "content": "Diffusion Transformer. The DiT-based video generation model shows significant potential in simulating the physical world [6, 66, 78]. Despite being a novel architecture, research on controllable generation has been limited, and current methods [9, 13, 42, 75] largely resemble U-Net based approaches [11, 39, 74]. However, no study has yet examined why this approach works with DiT. Drawing from prior analyses of Diffusion and Transformer from a frequency domain perspective [2-4, 45, 50, 55], we conclude that: (1) Low-frequency (e.g., shallow-layer) features are essential for pixel-level prediction tasks in diffusion models, which helps facilitate model training; (2) Transformers have limited perception for high-frequency information, which is important for controllable generation. Based on these, we decouple ID features into high- and low-frequency parts and inject them into specific locations, achieving effective identity-preserving text-to-video generation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.802, + 0.866, + 0.818 + ], + "angle": 0, + "content": "3.2. ConsisID: Keep Your Identity Consistent" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.908, + 0.903 + ], + "angle": 0, + "content": "The overview is illustrated in Figure 2. Given a reference image, the global facial extractor and local facial extractor inject both high- and low-frequency facial information into model, which then generates identity-preserving videos with the assistance of the consistency training strategy." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12980" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.092, + 0.459, + 0.106 + ], + "angle": 0, + "content": "3.2.1. Low-frequency View: Global Facial Extractor" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.109, + 0.484, + 0.401 + ], + "angle": 0, + "content": "In light of Finding 1, enhancing low-level (e.g., shallow, low-frequency) features accelerates model convergence. To easily adapt a pre-trained model for the IPT2V task, the most direct approach is concatenating the reference face with the noise input latent [5]. However, the reference face contains both high-frequency details (e.g., eye and lip textures) and low-frequency information (e.g., facial proportions and contours). From Finding 2, prematurely injecting high-frequency information into the Transformer is inefficient and may hinder the model's processing of low-frequency information, as the Transformer focuses primarily on low-frequency features. In addition, feeding the reference face directly into the model could introduce irrelevant noise such as lighting and shadows. To mitigate this, we extract facial key points, convert them to an RGB image, and then concatenate it with the reference image, as shown in Figure 2. This strategy focuses the model's attention on the low-frequency signals in the face, while minimizing the impact of extraneous features. We found that when this component is discarded, the model is difficult to converge. The objective function is changed to:" + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.41, + 0.483, + 0.429 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {b} = \\mathbb {E} _ {x _ {0}, t, y, f, \\epsilon} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (x _ {0}, t, \\tau_ {\\theta} (y), \\psi_ {\\theta} (f)) \\| _ {2} ^ {2} \\right], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.437, + 0.483, + 0.468 + ], + "angle": 0, + "content": "where \\(\\psi_{\\theta}(\\cdot)\\) is a variational autoencoder, \\(f\\) represents the reference image, we ignore key points here for simplicity." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.475, + 0.455, + 0.49 + ], + "angle": 0, + "content": "3.2.2. High-frequency View: Local Facial Extractor" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.493, + 0.483, + 0.84 + ], + "angle": 0, + "content": "In light of Finding 2, we recognize that Transformers have limited sensitivity to high-frequency information. It can be concluded that relying solely on global facial extractor is insufficient for IPT2V generation, as global facial features lack of high-frequency information. So we use a face recognition backbone [10] to extract high-frequency features, as these are invariant to non-ID attributes (e.g., expression, posture, and shape). We refer to these features as intrinsic identity features (e.g., high-frequency), since age and makeup do not alter an individual's core identity. Following [14], we utilize the penultimate layer of the backbone, rather than its output, as it retains more spatial information pertinent to identity. However, our experiments reveal that while the face recognition backbone improves identity consistency, it lacks the semantic features required for editing. This task demands not only maintaining identity consistency but also incorporating the ability to edit, such as generating videos of faces with the same identity but varying age and makeup. Previous research [15, 16] relies solely on the CLIP encoder [43] to enable editing capabilities. However, since CLIP is not specifically trained on face datasets, the features it extracts include irrelevant non-face information, which can compromise identity fidelity [29, 56, 67]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To address these challenges, we first use a facial recognition backbone to extract features that strongly represent intrinsic identity, and a CLIP image encoder to capture semantically rich features. We then employ the Q-Former" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.289 + ], + "angle": 0, + "content": "[27, 28, 64] to fuse these features, resulting in intrinsic identity representations enriched with high-frequency semantic information. To mitigate the impact of irrelevant features from CLIP, dropout [1, 22] is applied post-processing. Additionally, we follow [14] to concatenate the shallow, multiscale features from the facial recognition backbone (after interpolation) with the CLIP features. This approach ensures that the model captures essential intrinsic identity features while filtering out extraneous noise unrelated to identity. Finally, we apply cross-attention to facilitate interaction between this feature set and the visual tokens produced by each attention block of the pre-trained model, thereby enhancing the high-frequency information in the DiT:" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.299, + 0.907, + 0.318 + ], + "angle": 0, + "content": "\\[\nZ _ {i} ^ {\\prime} = Z _ {i} + \\text {A t t e n t i o n} \\left(Q _ {i} ^ {v}, K _ {i} ^ {f}, V _ {i} ^ {f}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.328, + 0.907, + 0.399 + ], + "angle": 0, + "content": "where \\(i\\) represents the layer number of the attention block, \\(Q^{v} = Z_{i}W_{i}^{q}\\), \\(K^{f} = FW_{i}^{k}\\), and \\(V^{f} = FW_{i}^{v}\\), where \\(Z_{i}\\) is the visual token, \\(F\\) represents the intrinsic identity features, and \\(W_{q}, W_{k}\\), and \\(W_{v}\\) are trainable parameters. The objective function is changed to:" + }, + { + "type": "equation", + "bbox": [ + 0.524, + 0.408, + 0.907, + 0.427 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c} = \\mathbb {E} _ {x _ {0}, t, y, f, \\epsilon} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(x _ {0}, t, \\tau_ {\\theta} (y), \\psi_ {\\theta} (f), \\varphi_ {\\theta} (f)\\right) \\| _ {2} ^ {2} \\right], \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.435, + 0.779, + 0.452 + ], + "angle": 0, + "content": "where \\(\\varphi_{\\theta}(\\cdot)\\) is the local facial extractor." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.459, + 0.771, + 0.474 + ], + "angle": 0, + "content": "3.2.3. Consistency Training Strategy" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.478, + 0.906, + 0.538 + ], + "angle": 0, + "content": "During training, we randomly select a frame from the training frames and apply the Crop & Align [10] to extract the facial region as reference images, which is subsequently used as an identity-control signal, alongside the text as control." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.907, + 0.81 + ], + "angle": 0, + "content": "Coarse-to-Fine Training. Compared to Identity-preserving image generation, video generation requires maintaining consistency in both spatial and temporal dimensions, ensuring that high and low-frequency facial information matches the reference image. To mitigate the complexity of training, we propose a hierarchical strategy where the model learns information globally before refining it locally. In the coarse-grained phase (e.g., corresponding Finding 1), we employ the global facial extractor, enabling the model to prioritize low-frequency features, such as facial contours and proportions, thereby ensuring rapid acquisition of identity information from the reference image and consistency across the video sequence. In the fine-grained phase (e.g. corresponding to Finding 2), the local facial extractor shifts the model's focus to high-frequency details, such as the texture details of eyes and lips (e.g., intrinsic identification), improving the fidelity of facial expressions and the overall similarity of the generated face." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Dynamic Mask Loss. The objective of our task is to ensure that the identity of the person in the generated video remains consistent with the input reference image. However, Equation 1 considers the entire scene, encompassing both high- and low-frequency identity information as well as redundant background content, which introduces noise that" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12981" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.12, + 0.089, + 0.458, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.265, + 0.486, + 0.295 + ], + "angle": 0, + "content": "Figure 3. User Study between ConsisID and state-of-the-art methods. ConsisID is preferred by voters in all dimensions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.316, + 0.484, + 0.392 + ], + "angle": 0, + "content": "interferes with model training. To address this, we propose to focus the model's attention on face regions. Specifically, we first extract the facial mask from the video, apply trilinear interpolation to map it to the latent space, and finally use this mask to constrain the computation of \\(\\mathcal{L}_c\\):" + }, + { + "type": "equation", + "bbox": [ + 0.235, + 0.399, + 0.483, + 0.415 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {d} = M \\odot \\mathcal {L} _ {c}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.422, + 0.484, + 0.498 + ], + "angle": 0, + "content": "where \\(M\\) represents a mask with the same shape as \\(\\epsilon\\). However, if Equation 5 is used as the supervisory signal for all training data, the model may fail to generate a natural background during inference. To mitigate this issue, we apply Equation 5 with a probability \\(p\\) of \\(\\alpha\\), resulting in:" + }, + { + "type": "equation", + "bbox": [ + 0.209, + 0.504, + 0.483, + 0.545 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {e} = \\left\\{ \\begin{array}{l l} \\mathcal {L} _ {d}, & \\text {i f} p > \\alpha \\\\ \\mathcal {L} _ {c}, & \\text {i f} p \\leq \\alpha \\end{array} \\right. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.549, + 0.484, + 0.716 + ], + "angle": 0, + "content": "Dynamic Cross-Face Loss. After training with Equation 6, we observed that the model struggled to generate satisfactory results for persons not present in the data domain during inference. This issue arises because the model, trained exclusively on faces from the training frames, tends to overfit by adopting a \"copy-paste\" shortcut—essentially replicating the reference image without alteration. To improve the model's generalization capability, we introduce slight Gaussian noise \\(\\zeta\\) to the reference images and use cross-face (e.g., reference images are sourced from video frames outside the training frames) as inputs with probability \\(\\beta\\):" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.722, + 0.483, + 0.763 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {f} = \\left\\{ \\begin{array}{l l} \\mathcal {L} _ {e} & \\text {w h e r e} x _ {0} \\cdot \\zeta , \\quad \\text {i f} p > \\beta \\\\ \\mathcal {L} _ {e} & \\text {w h e r e} x _ {c} \\cdot \\zeta , \\quad \\text {i f} p \\leq \\beta \\end{array} \\right. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.768, + 0.483, + 0.799 + ], + "angle": 0, + "content": "where \\(x_0\\) is the reference image extracted from the training frames, and \\(x_{c}\\) is extracted from outside the training frames." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.809, + 0.224, + 0.827 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.834, + 0.174, + 0.851 + ], + "angle": 0, + "content": "4.1. Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Implementation details. ConsisID selects DiT-based generation architectures CogVideoX-5B [66] as our baseline for validation. We use an in-house human-centric" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.089, + 0.905, + 0.134 + ], + "angle": 0, + "content": "
FaceSim-Arc ↑FaceSim-Cur ↑CLIPScore ↑FID ↓
ID-Animator [15]0.320.3324.97117.46
ConsisID0.580.6027.93151.82
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.144, + 0.907, + 0.187 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison with state-of-the-art methods. ConsID achieve well-aligned results across most metrics. \"↓\" denotes lower is better. \"↑\" denotes higher is better." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.207, + 0.907, + 0.418 + ], + "angle": 0, + "content": "dataset for training, which differs from previous datasets [40, 57, 68] that focus only on the face. In the training phase, we set the resolution to \\(480 \\times 720\\) and extracted 49 consecutive frames at a stride of 3 from each video as training data. We set the batch size to 80, the learning rate to \\(3 \\times 10^{-6}\\), and the total number of training steps to \\(1.8\\mathrm{k}\\). The randomly discarded text rate is set to 0.1, with AdamW serving as an optimizer and cosine_with_restarts as a learning rate scheduler. The training strategy is the same as Section 3.2.3. We set \\(\\alpha\\) and \\(\\beta\\) in the dynamic cross-face loss \\((\\mathcal{L}_e)\\) and dynamic mask loss \\((\\mathcal{L}_f)\\) to 0.5, respectively. In the inference phase, we employ DPM [51] with a sampling step of 50, and a classifier free guidance ratio of 6.0. For more details and results, please refer to Appendix." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.418, + 0.909, + 0.706 + ], + "angle": 0, + "content": "Benchmark. Since there is an absence of an evaluation dataset, we select 30 persons who were not included in the training data and sourced five high-quality images for each ID from the internet. We then design 90 distinct prompts, encompassing a variety of expressions, actions, and backgrounds for evaluation. Building on previous works [15, 42], we evaluate four dimensions: (1). Identity Preservation: We use FaceSim-Arc [10] and introduce FaceSim-Cur, which assesses identity preservation by measuring feature differences between face regions in the generated videos and those in real face images within the ArcFace [10] and CurricularFace [23] feature spaces. (2) Visual Quality: We utilize FID [18] by calculating feature differences in the face regions between the generated frames and real face images within the InceptionV3 [53] feature space. (3) Text Relevance: We utilize CLIPScore [17] to measure the similarity between the generated videos and the input prompts. (4). Motion Amplitude: Due to the lack of reliable metrics [24, 73], we evaluate through the user study." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.713, + 0.707, + 0.729 + ], + "angle": 0, + "content": "4.2. Qualitative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In this section, we compare our method, ConsisID, with ID-Animator [15] (e.g., the only available open-source model) for tuning-free IPT2V tasks. We randomly select images and text prompts of four individuals for qualitative analysis, all of which are absent from the training data. As shown in Figure 4, ID-Animator cannot generate human body parts beyond the face and is unable to generate complex actions or backgrounds in response to text prompts (e.g., action, attribute, background), which significantly limits its practical application. In addition, the preservation of the identity is inadequate; for example, in case 1, the reference image" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12982" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.164, + 0.088, + 0.836, + 0.601 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.611, + 0.907, + 0.669 + ], + "angle": 0, + "content": "Figure 4. Qualitative analysis between ConsisID and ID-Animator [15]. ID-Animator can only generate videos of the face region, and the identity Preservation is poor (e.g., shape, texture). Additionally, it cannot generate specified content according to the text prompt (e.g., action, decoration, background). ConsisID achieves advantages in identity preservation, visual quality, motion amplitude, and text relevance. Moreover, our ConsistID can generate more frames rather than ID-Animator (49 \\(480 \\times 720\\)p frames v.s. 16 \\(512 \\times 512\\)p frames)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.687, + 0.483, + 0.794 + ], + "angle": 0, + "content": "appears to be processed with skin smoothing. In case 2, wrinkles have been introduced which detract from the aesthetic quality. In cases 3 and 4, the face is distorted due to the lack of low frequency information, which compromises identity consistency. In contrast, the proposed ConsisID consistently produces high-quality, realistic videos that accurately match the reference identity and adhere to prompt." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.803, + 0.295, + 0.82 + ], + "angle": 0, + "content": "4.3. Quantitative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.903 + ], + "angle": 0, + "content": "We present a comprehensive quantitative evaluation of different methods, with results displayed in Table 1. Consistent with Figure 4, our method outperforms state-of-the-art methods across five metrics. For identity preservation, ConsisID achieves a higher score by designing appropri" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.687, + 0.907, + 0.84 + ], + "angle": 0, + "content": "ate identity signals for DiT from a frequency perspective. By contrast, ID-Animator [15] is not optimized for IPT2V and only partially retains facial features, resulting in lower FaceSim-Arc [10] and FaceSim-Cur scores. For Text Relevance, ConsisID not only controls expressions via prompts but also adjusts actions and backgrounds, achieving higher CLIPScore [17]. Regarding visual quality, the FID is presented solely as a reference due to its limited alignment [31, 32, 41, 73] with human perception. Please refer to Figure 4 and 3 for qualitative analysis of the visual quality." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.637, + 0.865 + ], + "angle": 0, + "content": "4.4. User Study" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Building on previous work, we conduct a human evaluation using a binary voting strategy, with each questionnaire con" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12983" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.089, + 0.475, + 0.491 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.501, + 0.483, + 0.557 + ], + "angle": 0, + "content": "Figure 5. Effect of Different Components via Qualitative Analysis. Removing any component may result in the loss of high- or low-frequency facial information, or hinder the ability to modify video content based on the text prompt." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.583, + 0.483, + 0.765 + ], + "angle": 0, + "content": "taining only 80 questions. Participants are required to view 40 video clips, a setup designed to improve both engagement and questionnaire validity. For the IPT2V task, each question requires participants to separately judge which option performs better in terms of Identity Preservation, Visual Quality, Text Alignment, and Motion Amplitude. This composition ensures the accuracy of the human evaluation. Owing to the extensive participant base required for this evaluation, we successfully gathered 103 valid questionnaires. The results, depicted in Figure 3, demonstrate a significant superiority of our method over ID-Animator [15], verifying the effectiveness of the designed DiT for IPT2V generation." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.773, + 0.466, + 0.79 + ], + "angle": 0, + "content": "4.5. Effect of the Identity Signal Injection in DiT" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.903 + ], + "angle": 0, + "content": "To assess the effectiveness of Finding 1 and Finding 2, we perform ablation studies on different methods of injecting control signals into DiT. Specifically, these experiments involved (a) injecting only low-frequency face information with key points into the noise latent, (b) injecting only high-frequency face signals within the attention block, (c) combining (a) and (b), (d) based on (c), but the low-frequency" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.088, + 0.898, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.513, + 0.907, + 0.556 + ], + "angle": 0, + "content": "Figure 6. Effect of Different Control Signal Injection Way via Qualitative Analysis. Only (c), which injects both high & low-freq face signals into the suitable location, performs best." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.573, + 0.905, + 0.67 + ], + "angle": 0, + "content": "
FaceSim-Arc ↑FaceSim-Cur ↑CLIPScore ↑FID ↓
w/o GFE0.050.0534.86269.88
w/o LFE0.660.6834.48104.34
w/o CFT0.540.5834.47144.62
w/o DML0.620.6734.23187.78
w/o DCL0.650.6932.21117.80
ConsisID0.730.7536.77127.42
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.679, + 0.907, + 0.751 + ], + "angle": 0, + "content": "Table 2. Effect of Local Facial Extractor (LFE), Global Facial Extractor (GFE), coarse-to-fine training (CFT), dynamic mask loss (DML) and dynamic cross-face loss (DCL) by Automatic Metrics. Removing any of the above methods significantly reduces identity preservation, text relevance, and visual quality." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.903 + ], + "angle": 0, + "content": "face information does not contain key points, and \\((\\mathbf{e} - \\mathbf{f})\\) based on (c), but the high-frequency signal is injected at the output or input of the attention block. (g) injecting only high-frequency face signals before the attention block. To reduce overhead, for each identity, we only select 2 reference images each with 90 text prompts for the evaluation. The results are shown in Figure 6 and Table 3. For Finding \\(I\\), we observe that only injecting high-frequency signals (a)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12984" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.482, + 0.205 + ], + "angle": 0, + "content": "
PlanFaceSim-Arc ↑FaceSim-Cur ↑CLIPScore ↑FID ↓
a0.050.0534.86269.88
b0.660.6834.48104.34
c0.730.7536.77127.42
d0.640.6830.69177.65
e0.620.6633.61164.15
funstable training process
gunstable training process
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.215, + 0.483, + 0.258 + ], + "angle": 0, + "content": "Table 3. Effect of Different Control Signal Injection Way via Quantitative Analysis. Only plan c, which injects both high and low-frequency face information into the model, performs best." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.281, + 0.485, + 0.782 + ], + "angle": 0, + "content": "greatly increases the training difficulty, causing the model to fail to converge due to the lack of low-frequency signal injection. In addition, the inclusion of facial key points (d) allows a greater focus on low-frequency information, thereby facilitating training and improving model performance. For Finding 2, when only low-frequency signals are injected (b), the model lacks high-frequency information. This reliance on low-frequency signals causes the generated face in the video to copy the reference image, making it difficult to control facial expressions, movements, and other features through prompts. Furthermore, injecting identity signals into the attention block input (f - g) disrupts the intended frequency domain distribution of DiT, resulting in a gradient explosion. Embedding control signals in the attention block (c) is preferable to embedding them in the output (e) because attention block processes predominantly low-frequency information. By embedding high-frequency information internally, the attention block is guided to highlight intrinsic facial features, whereas injecting it into the output merely concatenates features without directing focus, reducing DiT's modeling capacity. Moreover, we apply a Fourier transform to the generated videos (only the face region) to visually compare the influence of different components to extract facial information. As shown in Figure 7, the Fourier spectrum and the log amplitude of the Fourier transform reveal that injecting high or low-frequency signals can indeed enhance the corresponding frequency information of the generated face. Moreover, the low-frequency signal can be further enhanced by matching with the face key points, and injecting the high-frequency signal into the attention block has the highest feature utilization rate. Our method (c) shows strongest high and low frequency, further validating the efficiency benefit from Findings 1 and 2." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.789, + 0.483, + 0.806 + ], + "angle": 0, + "content": "4.6. Ablation on the Consistency Training Strategy" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To reduce overhead, for each identity, we only select 2 reference images for the following experiments. To demonstrate the benefits of the proposed consistency training strategy, we perform ablation experiments on coarse-to-fine training (CFT), dynamic mask loss \\(\\mathcal{L}_e\\) (DML), and dynamic crossface loss \\(\\mathcal{L}_f\\) (DCL), with the results presented in Figure" + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.089, + 0.631, + 0.161 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.162, + 0.618, + 0.171 + ], + "angle": 0, + "content": "(a) only high-freq" + }, + { + "type": "image", + "bbox": [ + 0.658, + 0.09, + 0.755, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.671, + 0.162, + 0.742, + 0.171 + ], + "angle": 0, + "content": "(b) only low-freq" + }, + { + "type": "image", + "bbox": [ + 0.783, + 0.09, + 0.878, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.789, + 0.162, + 0.872, + 0.171 + ], + "angle": 0, + "content": "(c) high & low freq" + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.174, + 0.631, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.537, + 0.247, + 0.625, + 0.256 + ], + "angle": 0, + "content": "(d) high & low† freq" + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.178, + 0.882, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.653, + 0.33, + 0.887, + 0.341 + ], + "angle": 0, + "content": "(f) Relative log amplitudes of Fourier transformed videos" + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.26, + 0.631, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.538, + 0.332, + 0.625, + 0.34 + ], + "angle": 0, + "content": "(e) high+ & low freq" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.352, + 0.907, + 0.435 + ], + "angle": 0, + "content": "Figure 7. (a - e) Fourier spectrum of different id signal injection. The center area represents low frequencies and the surrounding area represents high frequencies. (f) Relative log amplitudes of Fourier transformed generated videos. A larger response value indicates a higher inclusion of frequency information. (a - f) verify the effect of our frequency decomposition." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.461, + 0.907, + 0.628 + ], + "angle": 0, + "content": "5 and Table 2. When CFT is removed, GFE and LFE exhibit competing behaviors, complicating the model's ability to prioritize high and low-frequency information accurately, leading to convergence at suboptimal points. Removing DML required the model to simultaneously focus on both foreground and background elements, with background noise negatively affecting training and reducing facial consistency. Similarly, the exclusion of DCL impaired the generalization capability, reducing fidelity for faces, not in the training set and reducing its effectiveness in generating identity-preserving videos as intended." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.641, + 0.634, + 0.657 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.666, + 0.907, + 0.803 + ], + "angle": 0, + "content": "In this paper, we present ConsisID, a unified framework for keeping faces consistent in video generation by frequency decomposition. It can seamlessly integrate into existing DiT-based text-to-video models, for generating high-quality, editable, consistent identity-preserving videos. Extensive experiments show that ConsisID outperforms the current state-of-the-art identity-preserving T2V models. It reveals that our frequency-aware heuristic DiT-based control scheme is an optimal solution for IPT2V generation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.816, + 0.693, + 0.833 + ], + "angle": 0, + "content": "6. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.906, + 0.901 + ], + "angle": 0, + "content": "We thank all the anonymous reviewers for their constructive comments. This work was supported in part by the Natural Science Foundation of China (No. 62202014, 62332002, 62425101, 62088102)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12985" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.484, + 0.171 + ], + "angle": 0, + "content": "[1] Hassan Akbari, Liangzhe Yuan, Rui Qian, Wei-Hong Chuang, Shih-Fu Chang, Yin Cui, and Boqing Gong. Vatt: Transformers for multimodal self-supervised learning from raw video, audio and text. NeurIPS, 34:24206-24221, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.173, + 0.484, + 0.213 + ], + "angle": 0, + "content": "[2] Dosovitskiy Alexey. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv: 2010.11929, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.216, + 0.483, + 0.271 + ], + "angle": 0, + "content": "[3] Jiawang Bai, Li Yuan, Shu-Tao Xia, Shuicheng Yan, Zhifeng Li, and Wei Liu. Improving vision transformers by revisiting high-frequency components. In ECCV, pages 1-18. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.273, + 0.482, + 0.315 + ], + "angle": 0, + "content": "[4] Fan Bao, Shen Nie, Kaiwen Xue, Yue Cao, Chongxuan Li, Hang Su, and Jun Zhu. All are worth words: A vit backbone for diffusion models. In CVPR, pages 22669-22679, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.317, + 0.483, + 0.385 + ], + "angle": 0, + "content": "[5] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.388, + 0.482, + 0.456 + ], + "angle": 0, + "content": "[6] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. In openai, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.459, + 0.483, + 0.528 + ], + "angle": 0, + "content": "[7] Hila Chefer, Shiran Zada, Roni Paiss, Ariel Ephrat, Omer Tov, Michael Rubinstein, Lior Wolf, Tali Dekel, Tomer Michaeli, and Inbar Mosseri. Still-moving: Customized video generation without customized video data. arXiv preprint arXiv:2407.08674, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.53, + 0.483, + 0.599 + ], + "angle": 0, + "content": "[8] Li Chen, Mengyi Zhao, Yiheng Liu, Mingxu Ding, Yangyang Song, Shizun Wang, Xu Wang, Hao Yang, Jing Liu, Kang Du, et al. Photovverse: Tuning-free image customization with text-to-image diffusion models. arXiv preprint arXiv:2309.05793, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.601, + 0.483, + 0.656 + ], + "angle": 0, + "content": "[9] Soon Yau Cheong, Duygu Ceylan, Armin Mustafa, Andrew Gilbert, and Chun-Hao Paul Huang. Boosting camera motion control for video diffusion transformers. arXiv preprint arXiv:2410.10802, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.658, + 0.483, + 0.699 + ], + "angle": 0, + "content": "[10] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In CVPR, pages 4690-4699, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.702, + 0.483, + 0.77 + ], + "angle": 0, + "content": "[11] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.772, + 0.483, + 0.828 + ], + "angle": 0, + "content": "[12] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. ICLR, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.83, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[13] Zinan Guo, Yanze Wu, Zhuowei Chen, Lang Chen, and Qian He. Pulid: Pure and lightning id customization via contrastive alignment. arXiv preprint arXiv:2404.16022, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[14] Junjie He, Yifeng Geng, and Liefeng Bo. Unipotrait: A unified framework for identity-preserving single-" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.094, + 0.905, + 0.12 + ], + "angle": 0, + "content": "and multi-human image personalization. arXiv preprint arXiv:2408.05939, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[15] Xuanhua He, Quande Liu, Shengju Qian, Xin Wang, Tao Hu, Ke Cao, Keyu Yan, Man Zhou, and Jie Zhang. Id-Animator: Zero-shot identity-preserving human video generation. arXiv preprint arXiv:2404.15275, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.179, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[16] Zecheng He, Bo Sun, Felix Juefei-Xu, Haoyu Ma, Ankit Ramchandani, Vincent Cheung, Siddharth Shah, Anmol Kalia, Harihar Subramanyam, Alireza Zareian, et al. Imagine yourself: Tuning-free personalized image generation. arXiv preprint arXiv:2409.13346, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.25, + 0.904, + 0.303 + ], + "angle": 0, + "content": "[17] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.306, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.904, + 0.39 + ], + "angle": 0, + "content": "[19] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.392, + 0.904, + 0.419 + ], + "angle": 0, + "content": "[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 33:6840-6851, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.421, + 0.905, + 0.476 + ], + "angle": 0, + "content": "[21] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.478, + 0.905, + 0.547 + ], + "angle": 0, + "content": "[22] Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Q Weinberger. Deep networks with stochastic depth. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 646-661. Springer, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.905, + 0.604 + ], + "angle": 0, + "content": "[23] Yuge Huang, Yuhan Wang, Ying Tai, Xiaoming Liu, Pengcheng Shen, Shaoxin Li, Jilin Li, and Feiyue Huang. Curricularface: adaptive curriculum learning loss for deep face recognition. In CVPR, pages 5901-5910, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.605, + 0.905, + 0.687 + ], + "angle": 0, + "content": "[24] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.689, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[25] Dan Kondratyuk, Lijun Yu, Xiuye Gu, José Lezama, Jonathan Huang, Rachel Hornung, Hartwig Adam, Hassan Akbari, Yair Alon, Vighnesh Birodkar, et al. Videopoet: A large language model for zero-shot video generation. arXiv preprint arXiv:2312.14125, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.76, + 0.905, + 0.813 + ], + "angle": 0, + "content": "[26] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In CVPR, pages 1931-1941, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.816, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[27] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, pages 12888-12900. PMLR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[28] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.094, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12986" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "frozen image encoders and large language models. In ICML, pages 19730-19742. PMLR, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.178 + ], + "angle": 0, + "content": "[29] Zhen Li, Mingdeng Cao, Xintao Wang, Zhongang Qi, Ming-Ming Cheng, and Ying Shan. Photomaker: Customizing realistic human photos via stacked id embedding. In CVPR, pages 8640-8650, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.234 + ], + "angle": 0, + "content": "[30] Bin Lin, Yunyang Ge, Xinhua Cheng, Zongjian Li, Bin Zhu, Shaodong Wang, Xianyi He, Yang Ye, Shenghai Yuan, Liuhan Chen, et al. Open-sora plan: Open-source large video generation model. arXiv preprint arXiv:2412.00131, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.236, + 0.483, + 0.303 + ], + "angle": 0, + "content": "[31] Yaofang Liu, Xiaodong Cun, Xuebo Liu, Xintao Wang, Yong Zhang, Haoxin Chen, Yang Liu, Tieyong Zeng, Raymond Chan, and Ying Shan. Evalcrafter: Benchmarking and evaluating large video generation models. arXiv preprint arXiv:2310.11440, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.305, + 0.482, + 0.361 + ], + "angle": 0, + "content": "[32] Yuanxin Liu, Lei Li, Shuhuai Ren, Rundong Gao, Shicheng Li, Sishuo Chen, Xu Sun, and Lu Hou. Fetv: A benchmark for fine-grained evaluation of open-domain text-to-video generation. NeurIPS, 36, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.482, + 0.419 + ], + "angle": 0, + "content": "[33] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.482, + 0.476 + ], + "angle": 0, + "content": "[34] Yue Ma, Xiaodong Cun, Yingqing He, Chenyang Qi, Xintao Wang, Ying Shan, Xiu Li, and Qifeng Chen. Magic-stick: Controllable video editing via control handle transformations. arXiv preprint arXiv:2312.03047, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.477, + 0.482, + 0.546 + ], + "angle": 0, + "content": "[35] Yue Ma, Yingqing He, Xiaodong Cun, Xintao Wang, Siran Chen, Xiu Li, and Qifeng Chen. Follow your pose: Pose-guided text-to-video generation using pose-free videos. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4117-4125, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.482, + 0.616 + ], + "angle": 0, + "content": "[36] Yue Ma, Yingqing He, Hongfa Wang, Andong Wang, Chenyang Qi, Chengfei Cai, Xiu Li, Zhifeng Li, Heung-Yeung Shum, Wei Liu, et al. Follow-your-click: Open-domain regional image animation via short prompts. arXiv preprint arXiv:2403.08268, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.618, + 0.482, + 0.687 + ], + "angle": 0, + "content": "[37] Yue Ma, Hongyu Liu, Hongfa Wang, Heng Pan, Yingqing He, Junkun Yuan, Ailing Zeng, Chengfei Cai, Heung-Yeung Shum, Wei Liu, et al. Follow-your-emoji: Fine-controllable and expressive freestyle portrait animation. In SIGGRAPH Asia 2024 Conference Papers, pages 1–12, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.482, + 0.744 + ], + "angle": 0, + "content": "[38] Ze Ma, Daquan Zhou, Chun-Hsiao Yeh, Xue-She Wang, Xiuyu Li, Huanrui Yang, Zhen Dong, Kurt Keutzer, and Jiashi Feng. Magic-me: Identity-specific video customized diffusion. arXiv preprint arXiv:2402.09368, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[39] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.816, + 0.482, + 0.858 + ], + "angle": 0, + "content": "[40] Arsha Nagrani, Joon Son Chung, Weidi Xie, and Andrew Zisserman. Voxceleb: Large-scale speaker verification in the wild. Computer Speech & Language, 60:101027, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[41] Mayu Otani, Riku Togashi, Yu Sawai, Ryosuke Ishigami, Yuta Nakashima, Esa Rahtu, Janne Heikkila, and Shin'ichi Satoh. Toward verifiable and reproducible human evaluation" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "for text-to-image generation. In CVPR, pages 14277-14286, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[42] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, et al. Movie gen: A cast of media foundation models. arXiv preprint arXiv:2410.13720, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[43] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.317 + ], + "angle": 0, + "content": "[44] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, pages 8821-8831, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[45] Chen Rao, Guangyuan Li, Zehua Lan, Jiakai Sun, Junsheng Luan, Wei Xing, Lei Zhao, Huaizhong Lin, Jianfeng Dong, and Dalong Zhang. Rethinking video deblurring with wavelet-aware dynamic transformer and diffusion model. arXiv preprint arXiv:2408.13459, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[46] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bjorn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.433, + 0.905, + 0.489 + ], + "angle": 0, + "content": "[47] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, pages 22500-22510, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.905, + 0.558 + ], + "angle": 0, + "content": "[48] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdream booth: Hypernetworks for fast personalization of text-to-image models. In CVPR, pages 6527-6536, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.56, + 0.905, + 0.615 + ], + "angle": 0, + "content": "[49] Yujun Shi, Jun Hao Liew, Hanshu Yan, Vincent YF Tan, and Jiashi Feng. Instadrag: Lightning fast and accurate drag-based image editing emerging from videos. arXiv preprint arXiv:2405.13722, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.617, + 0.905, + 0.657 + ], + "angle": 0, + "content": "[50] Chenyang Si, Ziqi Huang, Yuming Jiang, and Ziwei Liu. Freeu: Free lunch in diffusion u-net. In CVPR, pages 4733-4743, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.66, + 0.905, + 0.715 + ], + "angle": 0, + "content": "[51] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, pages 2256-2265. PMLR, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.717, + 0.905, + 0.744 + ], + "angle": 0, + "content": "[52] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[53] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, pages 2818-2826, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[54] Zhenyu Tang, Junwu Zhang, Xinhua Cheng, Wangbo Yu, Chaoran Feng, Yatian Pang, Bin Lin, and Li Yuan. Cycle3d: High-quality and consistent image-to-3d generation via generation-reconstruction cycle. arXiv preprint arXiv:2407.19548, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[55] Yuchuan Tian, Zhijun Tu, Hanting Chen, Jie Hu, Chao Xu, and Yunhe Wang. U-dits: Downsample tokens in u-shaped" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12987" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "diffusion transformers. arXiv preprint arXiv:2405.02730, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.163 + ], + "angle": 0, + "content": "[56] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, and Anthony Chen. Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.165, + 0.482, + 0.207 + ], + "angle": 0, + "content": "[57] Ting-Chun Wang, Arun Mallya, and Ming-Yu Liu. One-shot free-view neural talking-head synthesis for video conferencing. In CVPR, pages 10039-10049, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.208, + 0.483, + 0.262 + ], + "angle": 0, + "content": "[58] Zhao Wang, Aoxue Li, Enze Xie, Lingting Zhu, Yong Guo, Qi Dou, and Zhenguo Li. Customvideo: Customizing text-to-video generation with multiple subjects. arXiv preprint arXiv:2401.09962, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.483, + 0.32 + ], + "angle": 0, + "content": "[59] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. In ICCV, pages 15943-15953, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.483, + 0.39 + ], + "angle": 0, + "content": "[60] Yujie Wei, Shiwei Zhang, Zhiwu Qing, Hangjie Yuan, Zhiheng Liu, Yu Liu, Yingya Zhang, Jingren Zhou, and Hongming Shan. Dreamvideo: Composing your dream videos with customized subject and motion. In CVPR, pages 6537-6549, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.392, + 0.483, + 0.447 + ], + "angle": 0, + "content": "[61] Jianzong Wu, Xiangtai Li, Yanhong Zeng, Jiangning Zhang, Qianyu Zhou, Yining Li, Yunhai Tong, and Kai Chen. Motionbooth: Motion-aware customized text-to-video generation. arXiv preprint arXiv:2406.17758, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.449, + 0.483, + 0.504 + ], + "angle": 0, + "content": "[62] Jing Xiong, Gongye Liu, Lun Huang, Chengyue Wu, Taiqiang Wu, Yao Mu, Yuan Yao, Hui Shen, Zhongwei Wan, Jinfa Huang, et al. Autoregressive models in vision: A survey. arXiv preprint arXiv:2411.05902, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.505, + 0.483, + 0.573 + ], + "angle": 0, + "content": "[63] Jiaqi Xu, Xinyi Zou, Kunzhe Huang, Yunkuo Chen, Bo Liu, MengLi Cheng, Xing Shi, and Jun Huang. Easyanimate: A high-performance long video generation method based on transformer architecture. arXiv preprint arXiv:2405.18991, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.576, + 0.483, + 0.645 + ], + "angle": 0, + "content": "[64] Le Xue, Manli Shu, Anas Awadalla, Jun Wang, An Yan, Senthil Purushwalkam, Honglu Zhou, Viraj Prabhu, Yutong Dai, Michael S Ryoo, et al. xgen-mm (blip-3): A family of open large multimodal models. arXiv preprint arXiv:2408.08872, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.483, + 0.702 + ], + "angle": 0, + "content": "[65] Shuo Yang, Kun-Peng Ning, Yu-Yang Liu, Jia-Yu Yao, Yong-Hong Tian, Yi-Bing Song, and Li Yuan. Is parameter collision hindering continual learning in llms? arXiv preprint arXiv:2410.10179, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.483, + 0.771 + ], + "angle": 0, + "content": "[66] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.773, + 0.483, + 0.829 + ], + "angle": 0, + "content": "[67] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.831, + 0.483, + 0.872 + ], + "angle": 0, + "content": "[68] Jianhui Yu, Hao Zhu, Liming Jiang, Chen Change Loy, Weidong Cai, and Wayne Wu. Celebv-text: A large-scale facial text-video dataset. In CVPR, pages 14805-14814, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[69] Wangbo Yu, Chaoran Feng, Jiye Tang, Xu Jia, Li Yuan, and Yonghong Tian. Evagaussians: Event stream assisted" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "gaussian splatting from blurry images. arXiv preprint arXiv:2405.20224, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[70] Wangbo Yu, Jinbo Xing, Li Yuan, Wenbo Hu, Xiaoyu Li, Zhipeng Huang, Xiangjun Gao, Tien-Tsin Wong, Ying Shan, and Yonghong Tian. Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis. arXiv preprint arXiv:2409.02048, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[71] Yongsheng Yu, Ziyun Zeng, Hang Hua, Jianlong Fu, and Jiebo Luo. Promptfix: You prompt and we fix the photo. arXiv preprint arXiv:2405.16785, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.905, + 0.291 + ], + "angle": 0, + "content": "[72] Shenghai Yuan, Jinfa Huang, Yujun Shi, Yongqi Xu, Ruijie Zhu, Bin Lin, Xinhua Cheng, Li Yuan, and Jiebo Luo. Mag-ictime: Time-lapse video generation models as metamorphic simulators. arXiv preprint arXiv:2404.05014, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.292, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[73] Shenghai Yuan, Jina Huang, Yongqi Xu, Yaoyang Liu, Shaofeng Zhang, Yujun Shi, Rui-Jie Zhu, Xinhua Cheng, Jiebo Luo, and Li Yuan. Chronomagic-bench: A benchmark for metamorphic evaluation of text-to-time-lapse video generation. NeurIPS, 37:21236-21270, 2025." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.404 + ], + "angle": 0, + "content": "[74] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, pages 3836-3847, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.405, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[75] Zhenghao Zhang, Junchao Liao, Menghao Li, Long Qin, and Weizhi Wang. Tora: Trajectory-oriented diffusion transformer for video generation. arXiv preprint arXiv:2407.21705, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.461, + 0.905, + 0.531 + ], + "angle": 0, + "content": "[76] Mingzhe Zheng, Yongqi Xu, Haojian Huang, Xuran Ma, Yexin Liu, Wenjie Shu, Yatian Pang, Feilong Tang, Qifeng Chen, Harry Yang, et al. Videogen-of-thought: A collaborative framework for multi-shot video generation. arXiv preprint arXiv:2412.02259, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.532, + 0.905, + 0.572 + ], + "angle": 0, + "content": "[77] Zangwei Zheng, Xiangyu Peng, and Yang You. Open-sora: Democratizing efficient video production for all. In *Github*, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.574, + 0.905, + 0.616 + ], + "angle": 0, + "content": "[78] Yuan Zhou, Qiuyue Wang, Yuxuan Cai, and Huan Yang. Allegro: Open the black box of commercial-level video generation model. arXiv preprint arXiv:2410.15458, 2024." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.616 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12988" + } + ] +] \ No newline at end of file diff --git a/2025/Identity-Preserving Text-to-Video Generation by Frequency Decomposition/16b4d9e1-d6f9-4df5-9c6a-3949d63cad2d_origin.pdf b/2025/Identity-Preserving Text-to-Video Generation by Frequency Decomposition/16b4d9e1-d6f9-4df5-9c6a-3949d63cad2d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..eab03f36317742faf88ab1a324b1424e5b1fafe1 --- /dev/null +++ b/2025/Identity-Preserving Text-to-Video Generation by Frequency Decomposition/16b4d9e1-d6f9-4df5-9c6a-3949d63cad2d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:be71d85be3860e0cb0f32792bd332eba6d488c479dd3dcac12b46ad8e21e6d91 +size 6565106 diff --git a/2025/Identity-Preserving Text-to-Video Generation by Frequency Decomposition/full.md b/2025/Identity-Preserving Text-to-Video Generation by Frequency Decomposition/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7f4effd8c650f0974b7172941f4c6d06bd350e07 --- /dev/null +++ b/2025/Identity-Preserving Text-to-Video Generation by Frequency Decomposition/full.md @@ -0,0 +1,365 @@ +# Identity-Preserving Text-to-Video Generation by Frequency Decomposition + +Shenghai Yuan $^{1,4}$ , Jina Fu Huang $^{3}$ , Xianyi He $^{1,4}$ , Yunyang Ge $^{1,4}$ , Yujun Shi $^{5}$ , Liuhan Chen $^{1,4}$ , Jiebo Luo $^{3}$ , Li Yuan $^{1,2,\dagger}$ + +1 Peking University, 2 Peng Cheng Laboratory, 3 University of Rochester, 4 Rabbitpre Intelligence, 5 National University of Singapore + +{yuanshenghai, HeXianyi, yunyang, chenliuhan}@stu.pku.edu.cn, shi.yujun@u.nus.edu, yuanli-ece@pku.edu.cn, {jhuang90@ur, jluo@cs}.rochester.edu + +# https://oku-yuangroup.github.io/ConsisID + +![](images/34b54a7e24fa53311fda96cdaa729a4fdf59d6896abd5e3cc5055d42acb33f7a.jpg) +Reference Image + +a man standing next to an airplane, engaged in a conversation on his cell phone ... wearings sunglasses and a black top ... talking seriously ... airplane has a green striperunning along its side, and there is a large engine visible behind his ... + +![](images/8c8d0e12247599b8644cccecb1e8528bb9339a8aa8432162a342213397f31948.jpg) + +![](images/59fc6dd022d7e21adb1bcf77fba633461a578cf874bfa8a97dda2da1a0741a47.jpg) + +![](images/1d77819c653754433091187f4d5b7b4c77832ce6411034518c8b17f37798c1fe.jpg) + +![](images/f5e00868e7d1b23805e951a3339837a845c521dfdc07e08e3d5e752c51ec2348.jpg) + +![](images/f029aaf3ef14bd5451a239f376da728354404bd4b192f53420e07333f5312550.jpg) +Reference Image + +a young woman with long blonde hair ... front of a lush, green bush...white flowers ... wearing a black top ... smiling and looking at the camera while gently touching the flowers on the bush ... then bends downslightly and smells one of the flower... + +![](images/d4fc29abb568f78cc3c1aae6314f1ac4aa5f9ad86d00bbd5c0c82a6f50ebe922.jpg) + +![](images/4ae97d4fe9b05646a907ba30623396ef9828e41dc2587d415cf74afaaa34a64a.jpg) + +![](images/bb0f02d3e50739dfe45ecf9ae29d058c0db21b2e9648245af32f966e2c62f110.jpg) + +![](images/2a0dc2c59d8a00a6a74a988c1c62816a5d893d5dcef5710281cc958ca9e768c3.jpg) + +![](images/e676eccfc9c1c145bf46da5125eed96a72a038c9afa02862f13ad3adcf002d92.jpg) + +![](images/d98e7dada3c07beb38d5374f3537080db851d998e1da545db5a313a0a64df515.jpg) +Reference Image + +... a man ... his gaze focused and intense as he holds the basketball ... dressed in athletic gear ... The basketball is positioned firmly in his hands as he gradually puts it down ... urban street court, with the fading light of dusk casting a soft glow over the scene. The man's serious expression and steady grip on the ball ... + +![](images/b76fd513743a1e45fba7555bc7cbab5498badd0ab55ab6b030c354e996af2e43.jpg) + +![](images/213b985ccecdf65154508d4fef0220dbefae56257961671ca24a9a8563c25bb5.jpg) + +![](images/074fa865a382c4e2a6d8163a3bb00355d485522016f22ee0d6dc70c35c6f379c.jpg) + +![](images/40a242efbdff3f99b2d0804ed1e3fd81236be0add5f95a9f51e3840423625347.jpg) + +![](images/02d55525786d8af3fedaa0e7f6418306a745c2ebfc18b217fdae8eedf308e0de.jpg) + +![](images/1171cbd67bc0542229ff2e85b5b63a0b39018981ac1042d8ca7deb3769ec8082.jpg) +Reference Image + +a young boy sitting at a table, eating a piece of food. He appears to be enjoying his meal, as he takes a bite and chews it. The boy is wearing a blue shirt and has short hair...background is dark, with some light coming from the left side of the frame. There is a straw visible on the right side of the frame ... + +![](images/71329db7f406c6a1885d06dbb74fd5e2e04095284f1cbe247d9d96be9178c128.jpg) +Figure 1. Examples of identity-preserving video generation (IPT2V) by our ConsisID. Given a reference image, our method can generate realistic and personalized human-centered videos while preserving identity. Red indicates that attributes in long instructions. + +![](images/bb7842d95518c076af165e834e81f2c4c5fce71c46dad32125dccaddd5d05696.jpg) + +![](images/bc38c9aec80ace97a6161457a904a8479fad07c0ef314884c34f32619601e83d.jpg) + +![](images/f6e448c4e9073d84f80d5cc990c987cd994c547581405da1d441134acdd42bb6.jpg) + +![](images/4f122182d385fe095ea3583f65c56f81bc0a05b5c1015e4f108f7797a8315b06.jpg) + +# Abstract + +Identity-preserving text-to-video (IPT2V) generation aims to create high-fidelity videos with consistent human identity. It is an important task in video generation but remains an open problem for generative models. This paper pushes the technical frontier of IPT2V in two directions that + +have not been resolved in the literature: (1) A tuning-free pipeline without tedious case-by-case finetuning, and (2) A frequency-aware heuristic identity-preserving Diffusion Transformer (DiT)-based control scheme. To achieve these goals, we propose ConsisID, a tuning-free DiT-based controllable IPT2V model to keep human-identity consistent in the generated video. Inspired by prior findings in fre + +quency analysis of vision/diffusion transformers, it employs identity-control signals base on frequency domain, since facial features can be decomposed into low-frequency global features (e.g., profile, proportions) and high-frequency intrinsic features (e.g., identity markers that remain unaffected by pose changes). Extensive experiments demonstrate that our frequency-aware heuristic scheme provides an optimal control solution for DiT-based models, making strides towards more effective IPT2V. + +# 1. Introduction + +Large-scale pre-trained video diffusion models [30, 66, 77, 78] have facilitated a variety of downstream applications [49, 54, 69, 70, 72, 74], particularly in identity-preserving text-to-video (IPT2V) [7, 38, 58, 60, 61]. However, existing methods face significant challenges, particularly the high overhead associated with the need for case-by-case finetuning, which diminishes their applicability. Within the open-source community, only the ID-Animator [15] can implement tuning-free IPT2V, but it can only generate videos similar to talking head [57] and has poor id preservation. + +Additionally, the above efforts are predominantly based on U-Net and cannot be adapted to the emerging DiT-based video model [30, 63, 66, 77, 78]. This challenge may stem from the inherent limitations of DiT compared to U-Net, including greater difficulty in training convergence and weakness in perceiving facial details. From some prior findings in frequency analysis of vision/diffusion transformers [2-4, 45, 50, 55, 71], we can know that the reason is: Finding 1: Shallow (e.g., low-level, low-frequency) features are essential for pixel-level prediction tasks in diffusion models, as they ease model training. U-Net facilitates model convergence by aggregating shallow features to the decoder via long skip connections, a mechanism that DiT does not incorporate; Finding 2: Transformers have limited perception of high-frequency information, which is important for preserving facial features. The encoder-decoder architecture of U-Net naturally possesses multi-scale features (e.g., richness in high-frequency), while DiT lacks a comparable structure. To develop a DiT-based control model, these must be addressed first. Please see Appendix for more details. + +For ID-preserving video generation, the challenges stem from the requirement for each frame to incorporate both high-frequency (e.g., age- and make-up-independent identity markers) and low-frequency information (e.g., facial shape) derived from the reference image, which can just be used to make up for the DiT defects mentioned above. Therefore, we propose ConsisID, to keep the identity consistency in video generation by frequency decomposition, based on the previously Findings of DiT in frequency analysis. Thanks to the large-scale pre-trained DiT, we can use its powerful capabilities to achieve tuning-free effects. + +ConsisID decouples identity features into high- and low-frequency signals, which are injected into specific locations within the DiT, facilitating efficient IPT2V generation. Specifically, in line with Finding 1, we first convert the reference image and the facial key points to the low-frequency signal, then concatenate them with input noise latent to ease the training. Following Finding 2, we utilize a dual-tower feature extractor to capture high-frequency facial information, which is integrated with vision tokens within the transformer block, thereby enhancing the DiT's high-frequency perception capabilities. Finally, to transform the pre-trained model into an IPT2V model and improve its generalization, we further introduce a hierarchical training strategy. + +Our contributions can be summarized as follows: + +- We introduce ConsisID, a tuning-free identity-preserving DiT-based IPT2V model, which preserves the identity of the main subject of the video using control signals from frequency decomposition. +- We propose a hierarchical training strategy, including coarse-to-fine training, dynamic mask loss, and dynamic cross-face loss, which work together to facilitate training and enhance generalization effectively. +- Extensive experiments demonstrate that our ConsisID can generate high-quality, editable, consistent identity-preserving videos, benefiting from our frequency-aware identity-preserving T2V DiT-based control scheme. + +# 2. Related Work + +Tuning-based Identity-preserving T2V Models. Diffusion models are widely recognized for their strong generative capabilities [19, 34-37, 44, 72, 73], significantly advancing the development of identity-preserving generative models [8, 39, 59, 74]. Initially, the researchers used tuning-based methods to generate content that matched the input ID. This process requires finetuning pretrained model for each new person during inference. For example, DreamBooth [47] introduced a novel loss function to fine-tune the entire network, embedding identity information while preserving the original generative capabilities. LoRA [21], similar to DreamBooth [47], requires training only a small subset of network parameters. In contrast, Textual Inversion [11] freezes the pretrained network and embeds identity information into a trainable word embedding. Subsequent tuning-based methods, including both image and video models based on U-Net or DiT architectures [7, 26, 48, 58, 60-62, 65, 76], generally follow three main approaches. While these models demonstrate substantial effectiveness, the requirement to fine-tune for each new identity restricts their practical applicability. + +Tuning-free Identity-preserving T2V Models. To address the issue of high resource consumption, several tuning-free diffusion models have recently emerged in the field of image generation [13, 14, 29, 56, 67]. These models + +![](images/77d7dd582ca5e7bd8036fda0935f30af10ef8bd107c4c743d66195a068f5cca1.jpg) +Figure 2. Overview of the proposed ConsisID. Based on Findings of DiT, low-frequency facial information is embedded into the shallow layers, while high-frequency information is incorporated into the vision tokens within the attention blocks. The ID-preserving Recipe is applied to ease training and improve generalization. The cross face, DropToken and Dropout are executed based on probability. + +do not require finetuning parameters for newly introduced IDs during inference. For instance, IP-Adapter [67] utilizes the CLIP [43] features of the identity image through cross-attention to guide the pretrained model in generating identity-preserving images. InstantID [56] extends this approach by replacing CLIP [43] features with Arcface [10] features and integrating a pose network to adjust facial proportions. Unlike these initial methods, which introduce control signals via visual tokens, PhotoMaker [29] and Imagine Yourself [16] leverage text tokens. Specifically, PhotoMaker [29] concatenates identity features obtained from the CLIP encoder [43] to the text embedding, while Imagine Yourself [16] uses element-wise addition for feature fusion. In the domain of video generation, only MovieGen [42] and ID-Animator [15] currently support ID-preserving text-to-video (IPT2V) generation. MovieGen is closed-source, whereas ID-Animator is open-source but uses a methodology similar to image models, leading to lower-quality identity preservation in the generated videos. We select the emerging DiT architecture [30, 33, 66, 77] and optimize it for IPT2V, drawing on conclusions from prior frequency analyses [2-4, 45, 50, 55]. This enables high-quality, editable, and consistent ID-preserving video generation. + +# 3. Methodology + +# 3.1. Preliminaries + +Diffusion Model. Text-to-video generation models usually utilize the diffusion paradigm, which gradually transforms noise $\epsilon$ into a video $x_0$ . Originally, denoising was conducted directly within the pixel space [20, 51, 52]; however, due to significant computational overheads, recent methods predominantly employ latent space [12, 25, 46, + +72]. The optimization process is defined: + +$$ +\mathcal {L} _ {a} = \mathbb {E} _ {x _ {0}, t, y, \epsilon} \left[ \| \epsilon - \epsilon_ {\theta} (x _ {0}, t, \tau_ {\theta} (y)) \| _ {2} ^ {2} \right], \qquad (1) +$$ + +where $y$ is text condition, $\epsilon$ is sampled from a standard normal distribution (e.g., $\epsilon \sim \mathcal{N}(0,1)$ ), and $\tau_{\theta}(\cdot)$ is the text encoder. By replacing $x_0$ with $\mathcal{E}(x_0)$ , the latent diffusion is derived, which is used by ConsisID. + +Diffusion Transformer. The DiT-based video generation model shows significant potential in simulating the physical world [6, 66, 78]. Despite being a novel architecture, research on controllable generation has been limited, and current methods [9, 13, 42, 75] largely resemble U-Net based approaches [11, 39, 74]. However, no study has yet examined why this approach works with DiT. Drawing from prior analyses of Diffusion and Transformer from a frequency domain perspective [2-4, 45, 50, 55], we conclude that: (1) Low-frequency (e.g., shallow-layer) features are essential for pixel-level prediction tasks in diffusion models, which helps facilitate model training; (2) Transformers have limited perception for high-frequency information, which is important for controllable generation. Based on these, we decouple ID features into high- and low-frequency parts and inject them into specific locations, achieving effective identity-preserving text-to-video generation. + +# 3.2. ConsisID: Keep Your Identity Consistent + +The overview is illustrated in Figure 2. Given a reference image, the global facial extractor and local facial extractor inject both high- and low-frequency facial information into model, which then generates identity-preserving videos with the assistance of the consistency training strategy. + +# 3.2.1. Low-frequency View: Global Facial Extractor + +In light of Finding 1, enhancing low-level (e.g., shallow, low-frequency) features accelerates model convergence. To easily adapt a pre-trained model for the IPT2V task, the most direct approach is concatenating the reference face with the noise input latent [5]. However, the reference face contains both high-frequency details (e.g., eye and lip textures) and low-frequency information (e.g., facial proportions and contours). From Finding 2, prematurely injecting high-frequency information into the Transformer is inefficient and may hinder the model's processing of low-frequency information, as the Transformer focuses primarily on low-frequency features. In addition, feeding the reference face directly into the model could introduce irrelevant noise such as lighting and shadows. To mitigate this, we extract facial key points, convert them to an RGB image, and then concatenate it with the reference image, as shown in Figure 2. This strategy focuses the model's attention on the low-frequency signals in the face, while minimizing the impact of extraneous features. We found that when this component is discarded, the model is difficult to converge. The objective function is changed to: + +$$ +\mathcal {L} _ {b} = \mathbb {E} _ {x _ {0}, t, y, f, \epsilon} \left[ \| \epsilon - \epsilon_ {\theta} (x _ {0}, t, \tau_ {\theta} (y), \psi_ {\theta} (f)) \| _ {2} ^ {2} \right], \tag {2} +$$ + +where $\psi_{\theta}(\cdot)$ is a variational autoencoder, $f$ represents the reference image, we ignore key points here for simplicity. + +# 3.2.2. High-frequency View: Local Facial Extractor + +In light of Finding 2, we recognize that Transformers have limited sensitivity to high-frequency information. It can be concluded that relying solely on global facial extractor is insufficient for IPT2V generation, as global facial features lack of high-frequency information. So we use a face recognition backbone [10] to extract high-frequency features, as these are invariant to non-ID attributes (e.g., expression, posture, and shape). We refer to these features as intrinsic identity features (e.g., high-frequency), since age and makeup do not alter an individual's core identity. Following [14], we utilize the penultimate layer of the backbone, rather than its output, as it retains more spatial information pertinent to identity. However, our experiments reveal that while the face recognition backbone improves identity consistency, it lacks the semantic features required for editing. This task demands not only maintaining identity consistency but also incorporating the ability to edit, such as generating videos of faces with the same identity but varying age and makeup. Previous research [15, 16] relies solely on the CLIP encoder [43] to enable editing capabilities. However, since CLIP is not specifically trained on face datasets, the features it extracts include irrelevant non-face information, which can compromise identity fidelity [29, 56, 67]. + +To address these challenges, we first use a facial recognition backbone to extract features that strongly represent intrinsic identity, and a CLIP image encoder to capture semantically rich features. We then employ the Q-Former + +[27, 28, 64] to fuse these features, resulting in intrinsic identity representations enriched with high-frequency semantic information. To mitigate the impact of irrelevant features from CLIP, dropout [1, 22] is applied post-processing. Additionally, we follow [14] to concatenate the shallow, multiscale features from the facial recognition backbone (after interpolation) with the CLIP features. This approach ensures that the model captures essential intrinsic identity features while filtering out extraneous noise unrelated to identity. Finally, we apply cross-attention to facilitate interaction between this feature set and the visual tokens produced by each attention block of the pre-trained model, thereby enhancing the high-frequency information in the DiT: + +$$ +Z _ {i} ^ {\prime} = Z _ {i} + \text {A t t e n t i o n} \left(Q _ {i} ^ {v}, K _ {i} ^ {f}, V _ {i} ^ {f}\right), \tag {3} +$$ + +where $i$ represents the layer number of the attention block, $Q^{v} = Z_{i}W_{i}^{q}$ , $K^{f} = FW_{i}^{k}$ , and $V^{f} = FW_{i}^{v}$ , where $Z_{i}$ is the visual token, $F$ represents the intrinsic identity features, and $W_{q}, W_{k}$ , and $W_{v}$ are trainable parameters. The objective function is changed to: + +$$ +\mathcal {L} _ {c} = \mathbb {E} _ {x _ {0}, t, y, f, \epsilon} \left[ \| \epsilon - \epsilon_ {\theta} \left(x _ {0}, t, \tau_ {\theta} (y), \psi_ {\theta} (f), \varphi_ {\theta} (f)\right) \| _ {2} ^ {2} \right], \tag {4} +$$ + +where $\varphi_{\theta}(\cdot)$ is the local facial extractor. + +# 3.2.3. Consistency Training Strategy + +During training, we randomly select a frame from the training frames and apply the Crop & Align [10] to extract the facial region as reference images, which is subsequently used as an identity-control signal, alongside the text as control. + +Coarse-to-Fine Training. Compared to Identity-preserving image generation, video generation requires maintaining consistency in both spatial and temporal dimensions, ensuring that high and low-frequency facial information matches the reference image. To mitigate the complexity of training, we propose a hierarchical strategy where the model learns information globally before refining it locally. In the coarse-grained phase (e.g., corresponding Finding 1), we employ the global facial extractor, enabling the model to prioritize low-frequency features, such as facial contours and proportions, thereby ensuring rapid acquisition of identity information from the reference image and consistency across the video sequence. In the fine-grained phase (e.g. corresponding to Finding 2), the local facial extractor shifts the model's focus to high-frequency details, such as the texture details of eyes and lips (e.g., intrinsic identification), improving the fidelity of facial expressions and the overall similarity of the generated face. + +Dynamic Mask Loss. The objective of our task is to ensure that the identity of the person in the generated video remains consistent with the input reference image. However, Equation 1 considers the entire scene, encompassing both high- and low-frequency identity information as well as redundant background content, which introduces noise that + +![](images/237290a6b297b7c75a39cec96650554bfd2c89fc1bf7c86b8c925d0f983ba3d2.jpg) +Figure 3. User Study between ConsisID and state-of-the-art methods. ConsisID is preferred by voters in all dimensions. + +interferes with model training. To address this, we propose to focus the model's attention on face regions. Specifically, we first extract the facial mask from the video, apply trilinear interpolation to map it to the latent space, and finally use this mask to constrain the computation of $\mathcal{L}_c$ : + +$$ +\mathcal {L} _ {d} = M \odot \mathcal {L} _ {c}, \tag {5} +$$ + +where $M$ represents a mask with the same shape as $\epsilon$ . However, if Equation 5 is used as the supervisory signal for all training data, the model may fail to generate a natural background during inference. To mitigate this issue, we apply Equation 5 with a probability $p$ of $\alpha$ , resulting in: + +$$ +\mathcal {L} _ {e} = \left\{ \begin{array}{l l} \mathcal {L} _ {d}, & \text {i f} p > \alpha \\ \mathcal {L} _ {c}, & \text {i f} p \leq \alpha \end{array} \right. \tag {6} +$$ + +Dynamic Cross-Face Loss. After training with Equation 6, we observed that the model struggled to generate satisfactory results for persons not present in the data domain during inference. This issue arises because the model, trained exclusively on faces from the training frames, tends to overfit by adopting a "copy-paste" shortcut—essentially replicating the reference image without alteration. To improve the model's generalization capability, we introduce slight Gaussian noise $\zeta$ to the reference images and use cross-face (e.g., reference images are sourced from video frames outside the training frames) as inputs with probability $\beta$ : + +$$ +\mathcal {L} _ {f} = \left\{ \begin{array}{l l} \mathcal {L} _ {e} & \text {w h e r e} x _ {0} \cdot \zeta , \quad \text {i f} p > \beta \\ \mathcal {L} _ {e} & \text {w h e r e} x _ {c} \cdot \zeta , \quad \text {i f} p \leq \beta \end{array} \right. \tag {7} +$$ + +where $x_0$ is the reference image extracted from the training frames, and $x_{c}$ is extracted from outside the training frames. + +# 4. Experiments + +# 4.1. Setup + +Implementation details. ConsisID selects DiT-based generation architectures CogVideoX-5B [66] as our baseline for validation. We use an in-house human-centric + +
FaceSim-Arc ↑FaceSim-Cur ↑CLIPScore ↑FID ↓
ID-Animator [15]0.320.3324.97117.46
ConsisID0.580.6027.93151.82
+ +Table 1. Quantitative comparison with state-of-the-art methods. ConsID achieve well-aligned results across most metrics. "↓" denotes lower is better. "↑" denotes higher is better. + +dataset for training, which differs from previous datasets [40, 57, 68] that focus only on the face. In the training phase, we set the resolution to $480 \times 720$ and extracted 49 consecutive frames at a stride of 3 from each video as training data. We set the batch size to 80, the learning rate to $3 \times 10^{-6}$ , and the total number of training steps to $1.8\mathrm{k}$ . The randomly discarded text rate is set to 0.1, with AdamW serving as an optimizer and cosine_with_restarts as a learning rate scheduler. The training strategy is the same as Section 3.2.3. We set $\alpha$ and $\beta$ in the dynamic cross-face loss $(\mathcal{L}_e)$ and dynamic mask loss $(\mathcal{L}_f)$ to 0.5, respectively. In the inference phase, we employ DPM [51] with a sampling step of 50, and a classifier free guidance ratio of 6.0. For more details and results, please refer to Appendix. + +Benchmark. Since there is an absence of an evaluation dataset, we select 30 persons who were not included in the training data and sourced five high-quality images for each ID from the internet. We then design 90 distinct prompts, encompassing a variety of expressions, actions, and backgrounds for evaluation. Building on previous works [15, 42], we evaluate four dimensions: (1). Identity Preservation: We use FaceSim-Arc [10] and introduce FaceSim-Cur, which assesses identity preservation by measuring feature differences between face regions in the generated videos and those in real face images within the ArcFace [10] and CurricularFace [23] feature spaces. (2) Visual Quality: We utilize FID [18] by calculating feature differences in the face regions between the generated frames and real face images within the InceptionV3 [53] feature space. (3) Text Relevance: We utilize CLIPScore [17] to measure the similarity between the generated videos and the input prompts. (4). Motion Amplitude: Due to the lack of reliable metrics [24, 73], we evaluate through the user study. + +# 4.2. Qualitative Analysis + +In this section, we compare our method, ConsisID, with ID-Animator [15] (e.g., the only available open-source model) for tuning-free IPT2V tasks. We randomly select images and text prompts of four individuals for qualitative analysis, all of which are absent from the training data. As shown in Figure 4, ID-Animator cannot generate human body parts beyond the face and is unable to generate complex actions or backgrounds in response to text prompts (e.g., action, attribute, background), which significantly limits its practical application. In addition, the preservation of the identity is inadequate; for example, in case 1, the reference image + +![](images/f2181427b3d85c505fa48e2347daf297da2ffd5a21badf93ca0264b3f610026d.jpg) +Figure 4. Qualitative analysis between ConsisID and ID-Animator [15]. ID-Animator can only generate videos of the face region, and the identity Preservation is poor (e.g., shape, texture). Additionally, it cannot generate specified content according to the text prompt (e.g., action, decoration, background). ConsisID achieves advantages in identity preservation, visual quality, motion amplitude, and text relevance. Moreover, our ConsistID can generate more frames rather than ID-Animator (49 $480 \times 720$ p frames v.s. 16 $512 \times 512$ p frames). + +appears to be processed with skin smoothing. In case 2, wrinkles have been introduced which detract from the aesthetic quality. In cases 3 and 4, the face is distorted due to the lack of low frequency information, which compromises identity consistency. In contrast, the proposed ConsisID consistently produces high-quality, realistic videos that accurately match the reference identity and adhere to prompt. + +# 4.3. Quantitative Analysis + +We present a comprehensive quantitative evaluation of different methods, with results displayed in Table 1. Consistent with Figure 4, our method outperforms state-of-the-art methods across five metrics. For identity preservation, ConsisID achieves a higher score by designing appropri + +ate identity signals for DiT from a frequency perspective. By contrast, ID-Animator [15] is not optimized for IPT2V and only partially retains facial features, resulting in lower FaceSim-Arc [10] and FaceSim-Cur scores. For Text Relevance, ConsisID not only controls expressions via prompts but also adjusts actions and backgrounds, achieving higher CLIPScore [17]. Regarding visual quality, the FID is presented solely as a reference due to its limited alignment [31, 32, 41, 73] with human perception. Please refer to Figure 4 and 3 for qualitative analysis of the visual quality. + +# 4.4. User Study + +Building on previous work, we conduct a human evaluation using a binary voting strategy, with each questionnaire con + +![](images/bbc2af07b9696d973d07ea1692bd0d7a753454559e973d94a8eed714283f165e.jpg) +Figure 5. Effect of Different Components via Qualitative Analysis. Removing any component may result in the loss of high- or low-frequency facial information, or hinder the ability to modify video content based on the text prompt. + +taining only 80 questions. Participants are required to view 40 video clips, a setup designed to improve both engagement and questionnaire validity. For the IPT2V task, each question requires participants to separately judge which option performs better in terms of Identity Preservation, Visual Quality, Text Alignment, and Motion Amplitude. This composition ensures the accuracy of the human evaluation. Owing to the extensive participant base required for this evaluation, we successfully gathered 103 valid questionnaires. The results, depicted in Figure 3, demonstrate a significant superiority of our method over ID-Animator [15], verifying the effectiveness of the designed DiT for IPT2V generation. + +# 4.5. Effect of the Identity Signal Injection in DiT + +To assess the effectiveness of Finding 1 and Finding 2, we perform ablation studies on different methods of injecting control signals into DiT. Specifically, these experiments involved (a) injecting only low-frequency face information with key points into the noise latent, (b) injecting only high-frequency face signals within the attention block, (c) combining (a) and (b), (d) based on (c), but the low-frequency + +![](images/a4e4ec88eb66f94d25bba40665477d32a1841b3a6105327195353b440640a845.jpg) +Figure 6. Effect of Different Control Signal Injection Way via Qualitative Analysis. Only (c), which injects both high & low-freq face signals into the suitable location, performs best. + +
FaceSim-Arc ↑FaceSim-Cur ↑CLIPScore ↑FID ↓
w/o GFE0.050.0534.86269.88
w/o LFE0.660.6834.48104.34
w/o CFT0.540.5834.47144.62
w/o DML0.620.6734.23187.78
w/o DCL0.650.6932.21117.80
ConsisID0.730.7536.77127.42
+ +Table 2. Effect of Local Facial Extractor (LFE), Global Facial Extractor (GFE), coarse-to-fine training (CFT), dynamic mask loss (DML) and dynamic cross-face loss (DCL) by Automatic Metrics. Removing any of the above methods significantly reduces identity preservation, text relevance, and visual quality. + +face information does not contain key points, and $(\mathbf{e} - \mathbf{f})$ based on (c), but the high-frequency signal is injected at the output or input of the attention block. (g) injecting only high-frequency face signals before the attention block. To reduce overhead, for each identity, we only select 2 reference images each with 90 text prompts for the evaluation. The results are shown in Figure 6 and Table 3. For Finding $I$ , we observe that only injecting high-frequency signals (a) + +
PlanFaceSim-Arc ↑FaceSim-Cur ↑CLIPScore ↑FID ↓
a0.050.0534.86269.88
b0.660.6834.48104.34
c0.730.7536.77127.42
d0.640.6830.69177.65
e0.620.6633.61164.15
funstable training process
gunstable training process
+ +Table 3. Effect of Different Control Signal Injection Way via Quantitative Analysis. Only plan c, which injects both high and low-frequency face information into the model, performs best. + +greatly increases the training difficulty, causing the model to fail to converge due to the lack of low-frequency signal injection. In addition, the inclusion of facial key points (d) allows a greater focus on low-frequency information, thereby facilitating training and improving model performance. For Finding 2, when only low-frequency signals are injected (b), the model lacks high-frequency information. This reliance on low-frequency signals causes the generated face in the video to copy the reference image, making it difficult to control facial expressions, movements, and other features through prompts. Furthermore, injecting identity signals into the attention block input (f - g) disrupts the intended frequency domain distribution of DiT, resulting in a gradient explosion. Embedding control signals in the attention block (c) is preferable to embedding them in the output (e) because attention block processes predominantly low-frequency information. By embedding high-frequency information internally, the attention block is guided to highlight intrinsic facial features, whereas injecting it into the output merely concatenates features without directing focus, reducing DiT's modeling capacity. Moreover, we apply a Fourier transform to the generated videos (only the face region) to visually compare the influence of different components to extract facial information. As shown in Figure 7, the Fourier spectrum and the log amplitude of the Fourier transform reveal that injecting high or low-frequency signals can indeed enhance the corresponding frequency information of the generated face. Moreover, the low-frequency signal can be further enhanced by matching with the face key points, and injecting the high-frequency signal into the attention block has the highest feature utilization rate. Our method (c) shows strongest high and low frequency, further validating the efficiency benefit from Findings 1 and 2. + +# 4.6. Ablation on the Consistency Training Strategy + +To reduce overhead, for each identity, we only select 2 reference images for the following experiments. To demonstrate the benefits of the proposed consistency training strategy, we perform ablation experiments on coarse-to-fine training (CFT), dynamic mask loss $\mathcal{L}_e$ (DML), and dynamic crossface loss $\mathcal{L}_f$ (DCL), with the results presented in Figure + +![](images/65f5d2294d04dd308479a5f1c85c79ba40e4ffc3e649d056ce20ff6373cad2f4.jpg) +(a) only high-freq + +![](images/00d4ad5dfc734192981137bf8c8ffcc15c55685a44ab405736cb53e07c0448bd.jpg) +(b) only low-freq + +![](images/7ae2c20bdfd58558e5b664029efe9b17a3c80dae5f5588743999e2841a4c32e0.jpg) +(c) high & low freq + +![](images/2497ec5f968a9e5b83550d1f63698a3c2b580d2106247ad2b398c9c5a9e3d7aa.jpg) +(d) high & low† freq + +![](images/5650e498907d8bcca60ce89715d5620898a31a3ea6a7ea27f5eb5b79cd51311c.jpg) +(f) Relative log amplitudes of Fourier transformed videos + +![](images/64bb93253f7a3f9033c3b75436b212918d74de990232fcfa5ac589ca23e6c566.jpg) +(e) high+ & low freq +Figure 7. (a - e) Fourier spectrum of different id signal injection. The center area represents low frequencies and the surrounding area represents high frequencies. (f) Relative log amplitudes of Fourier transformed generated videos. A larger response value indicates a higher inclusion of frequency information. (a - f) verify the effect of our frequency decomposition. + +5 and Table 2. When CFT is removed, GFE and LFE exhibit competing behaviors, complicating the model's ability to prioritize high and low-frequency information accurately, leading to convergence at suboptimal points. Removing DML required the model to simultaneously focus on both foreground and background elements, with background noise negatively affecting training and reducing facial consistency. Similarly, the exclusion of DCL impaired the generalization capability, reducing fidelity for faces, not in the training set and reducing its effectiveness in generating identity-preserving videos as intended. + +# 5. Conclusion + +In this paper, we present ConsisID, a unified framework for keeping faces consistent in video generation by frequency decomposition. It can seamlessly integrate into existing DiT-based text-to-video models, for generating high-quality, editable, consistent identity-preserving videos. Extensive experiments show that ConsisID outperforms the current state-of-the-art identity-preserving T2V models. It reveals that our frequency-aware heuristic DiT-based control scheme is an optimal solution for IPT2V generation. + +# 6. Acknowledgments + +We thank all the anonymous reviewers for their constructive comments. This work was supported in part by the Natural Science Foundation of China (No. 62202014, 62332002, 62425101, 62088102). + +# References + +[1] Hassan Akbari, Liangzhe Yuan, Rui Qian, Wei-Hong Chuang, Shih-Fu Chang, Yin Cui, and Boqing Gong. Vatt: Transformers for multimodal self-supervised learning from raw video, audio and text. NeurIPS, 34:24206-24221, 2021. +[2] Dosovitskiy Alexey. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv: 2010.11929, 2020. +[3] Jiawang Bai, Li Yuan, Shu-Tao Xia, Shuicheng Yan, Zhifeng Li, and Wei Liu. Improving vision transformers by revisiting high-frequency components. In ECCV, pages 1-18. Springer, 2022. +[4] Fan Bao, Shen Nie, Kaiwen Xue, Yue Cao, Chongxuan Li, Hang Su, and Jun Zhu. All are worth words: A vit backbone for diffusion models. In CVPR, pages 22669-22679, 2023. +[5] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. +[6] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. In openai, 2024. +[7] Hila Chefer, Shiran Zada, Roni Paiss, Ariel Ephrat, Omer Tov, Michael Rubinstein, Lior Wolf, Tali Dekel, Tomer Michaeli, and Inbar Mosseri. Still-moving: Customized video generation without customized video data. arXiv preprint arXiv:2407.08674, 2024. +[8] Li Chen, Mengyi Zhao, Yiheng Liu, Mingxu Ding, Yangyang Song, Shizun Wang, Xu Wang, Hao Yang, Jing Liu, Kang Du, et al. Photovverse: Tuning-free image customization with text-to-image diffusion models. arXiv preprint arXiv:2309.05793, 2023. +[9] Soon Yau Cheong, Duygu Ceylan, Armin Mustafa, Andrew Gilbert, and Chun-Hao Paul Huang. Boosting camera motion control for video diffusion transformers. arXiv preprint arXiv:2410.10802, 2024. +[10] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In CVPR, pages 4690-4699, 2019. +[11] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. +[12] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. ICLR, 2024. +[13] Zinan Guo, Yanze Wu, Zhuowei Chen, Lang Chen, and Qian He. Pulid: Pure and lightning id customization via contrastive alignment. arXiv preprint arXiv:2404.16022, 2024. +[14] Junjie He, Yifeng Geng, and Liefeng Bo. Unipotrait: A unified framework for identity-preserving single- + +and multi-human image personalization. arXiv preprint arXiv:2408.05939, 2024. +[15] Xuanhua He, Quande Liu, Shengju Qian, Xin Wang, Tao Hu, Ke Cao, Keyu Yan, Man Zhou, and Jie Zhang. Id-Animator: Zero-shot identity-preserving human video generation. arXiv preprint arXiv:2404.15275, 2024. +[16] Zecheng He, Bo Sun, Felix Juefei-Xu, Haoyu Ma, Ankit Ramchandani, Vincent Cheung, Siddharth Shah, Anmol Kalia, Harihar Subramanyam, Alireza Zareian, et al. Imagine yourself: Tuning-free personalized image generation. arXiv preprint arXiv:2409.13346, 2024. +[17] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. +[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS, 2017. +[19] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. +[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 33:6840-6851, 2020. +[21] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. +[22] Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Q Weinberger. Deep networks with stochastic depth. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 646-661. Springer, 2016. +[23] Yuge Huang, Yuhan Wang, Ying Tai, Xiaoming Liu, Pengcheng Shen, Shaoxin Li, Jilin Li, and Feiyue Huang. Curricularface: adaptive curriculum learning loss for deep face recognition. In CVPR, pages 5901-5910, 2020. +[24] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024. +[25] Dan Kondratyuk, Lijun Yu, Xiuye Gu, José Lezama, Jonathan Huang, Rachel Hornung, Hartwig Adam, Hassan Akbari, Yair Alon, Vighnesh Birodkar, et al. Videopoet: A large language model for zero-shot video generation. arXiv preprint arXiv:2312.14125, 2023. +[26] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In CVPR, pages 1931-1941, 2023. +[27] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, pages 12888-12900. PMLR, 2022. +[28] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with + +frozen image encoders and large language models. In ICML, pages 19730-19742. PMLR, 2023. +[29] Zhen Li, Mingdeng Cao, Xintao Wang, Zhongang Qi, Ming-Ming Cheng, and Ying Shan. Photomaker: Customizing realistic human photos via stacked id embedding. In CVPR, pages 8640-8650, 2024. +[30] Bin Lin, Yunyang Ge, Xinhua Cheng, Zongjian Li, Bin Zhu, Shaodong Wang, Xianyi He, Yang Ye, Shenghai Yuan, Liuhan Chen, et al. Open-sora plan: Open-source large video generation model. arXiv preprint arXiv:2412.00131, 2024. +[31] Yaofang Liu, Xiaodong Cun, Xuebo Liu, Xintao Wang, Yong Zhang, Haoxin Chen, Yang Liu, Tieyong Zeng, Raymond Chan, and Ying Shan. Evalcrafter: Benchmarking and evaluating large video generation models. arXiv preprint arXiv:2310.11440, 2023. +[32] Yuanxin Liu, Lei Li, Shuhuai Ren, Rundong Gao, Shicheng Li, Sishuo Chen, Xu Sun, and Lu Hou. Fetv: A benchmark for fine-grained evaluation of open-domain text-to-video generation. NeurIPS, 36, 2024. +[33] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. +[34] Yue Ma, Xiaodong Cun, Yingqing He, Chenyang Qi, Xintao Wang, Ying Shan, Xiu Li, and Qifeng Chen. Magic-stick: Controllable video editing via control handle transformations. arXiv preprint arXiv:2312.03047, 2023. +[35] Yue Ma, Yingqing He, Xiaodong Cun, Xintao Wang, Siran Chen, Xiu Li, and Qifeng Chen. Follow your pose: Pose-guided text-to-video generation using pose-free videos. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4117-4125, 2024. +[36] Yue Ma, Yingqing He, Hongfa Wang, Andong Wang, Chenyang Qi, Chengfei Cai, Xiu Li, Zhifeng Li, Heung-Yeung Shum, Wei Liu, et al. Follow-your-click: Open-domain regional image animation via short prompts. arXiv preprint arXiv:2403.08268, 2024. +[37] Yue Ma, Hongyu Liu, Hongfa Wang, Heng Pan, Yingqing He, Junkun Yuan, Ailing Zeng, Chengfei Cai, Heung-Yeung Shum, Wei Liu, et al. Follow-your-emoji: Fine-controllable and expressive freestyle portrait animation. In SIGGRAPH Asia 2024 Conference Papers, pages 1–12, 2024. +[38] Ze Ma, Daquan Zhou, Chun-Hsiao Yeh, Xue-She Wang, Xiuyu Li, Huanrui Yang, Zhen Dong, Kurt Keutzer, and Jiashi Feng. Magic-me: Identity-specific video customized diffusion. arXiv preprint arXiv:2402.09368, 2024. +[39] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453, 2023. +[40] Arsha Nagrani, Joon Son Chung, Weidi Xie, and Andrew Zisserman. Voxceleb: Large-scale speaker verification in the wild. Computer Speech & Language, 60:101027, 2020. +[41] Mayu Otani, Riku Togashi, Yu Sawai, Ryosuke Ishigami, Yuta Nakashima, Esa Rahtu, Janne Heikkila, and Shin'ichi Satoh. Toward verifiable and reproducible human evaluation + +for text-to-image generation. In CVPR, pages 14277-14286, 2023. +[42] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, et al. Movie gen: A cast of media foundation models. arXiv preprint arXiv:2410.13720, 2024. +[43] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. +[44] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, pages 8821-8831, 2021. +[45] Chen Rao, Guangyuan Li, Zehua Lan, Jiakai Sun, Junsheng Luan, Wei Xing, Lei Zhao, Huaizhong Lin, Jianfeng Dong, and Dalong Zhang. Rethinking video deblurring with wavelet-aware dynamic transformer and diffusion model. arXiv preprint arXiv:2408.13459, 2024. +[46] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bjorn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. +[47] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, pages 22500-22510, 2023. +[48] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdream booth: Hypernetworks for fast personalization of text-to-image models. In CVPR, pages 6527-6536, 2024. +[49] Yujun Shi, Jun Hao Liew, Hanshu Yan, Vincent YF Tan, and Jiashi Feng. Instadrag: Lightning fast and accurate drag-based image editing emerging from videos. arXiv preprint arXiv:2405.13722, 2024. +[50] Chenyang Si, Ziqi Huang, Yuming Jiang, and Ziwei Liu. Freeu: Free lunch in diffusion u-net. In CVPR, pages 4733-4743, 2024. +[51] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, pages 2256-2265. PMLR, 2015. +[52] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv, 2020. +[53] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, pages 2818-2826, 2016. +[54] Zhenyu Tang, Junwu Zhang, Xinhua Cheng, Wangbo Yu, Chaoran Feng, Yatian Pang, Bin Lin, and Li Yuan. Cycle3d: High-quality and consistent image-to-3d generation via generation-reconstruction cycle. arXiv preprint arXiv:2407.19548, 2024. +[55] Yuchuan Tian, Zhijun Tu, Hanting Chen, Jie Hu, Chao Xu, and Yunhe Wang. U-dits: Downsample tokens in u-shaped + +diffusion transformers. arXiv preprint arXiv:2405.02730, 2024. +[56] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, and Anthony Chen. Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519, 2024. +[57] Ting-Chun Wang, Arun Mallya, and Ming-Yu Liu. One-shot free-view neural talking-head synthesis for video conferencing. In CVPR, pages 10039-10049, 2021. +[58] Zhao Wang, Aoxue Li, Enze Xie, Lingting Zhu, Yong Guo, Qi Dou, and Zhenguo Li. Customvideo: Customizing text-to-video generation with multiple subjects. arXiv preprint arXiv:2401.09962, 2024. +[59] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. In ICCV, pages 15943-15953, 2023. +[60] Yujie Wei, Shiwei Zhang, Zhiwu Qing, Hangjie Yuan, Zhiheng Liu, Yu Liu, Yingya Zhang, Jingren Zhou, and Hongming Shan. Dreamvideo: Composing your dream videos with customized subject and motion. In CVPR, pages 6537-6549, 2024. +[61] Jianzong Wu, Xiangtai Li, Yanhong Zeng, Jiangning Zhang, Qianyu Zhou, Yining Li, Yunhai Tong, and Kai Chen. Motionbooth: Motion-aware customized text-to-video generation. arXiv preprint arXiv:2406.17758, 2024. +[62] Jing Xiong, Gongye Liu, Lun Huang, Chengyue Wu, Taiqiang Wu, Yao Mu, Yuan Yao, Hui Shen, Zhongwei Wan, Jinfa Huang, et al. Autoregressive models in vision: A survey. arXiv preprint arXiv:2411.05902, 2024. +[63] Jiaqi Xu, Xinyi Zou, Kunzhe Huang, Yunkuo Chen, Bo Liu, MengLi Cheng, Xing Shi, and Jun Huang. Easyanimate: A high-performance long video generation method based on transformer architecture. arXiv preprint arXiv:2405.18991, 2024. +[64] Le Xue, Manli Shu, Anas Awadalla, Jun Wang, An Yan, Senthil Purushwalkam, Honglu Zhou, Viraj Prabhu, Yutong Dai, Michael S Ryoo, et al. xgen-mm (blip-3): A family of open large multimodal models. arXiv preprint arXiv:2408.08872, 2024. +[65] Shuo Yang, Kun-Peng Ning, Yu-Yang Liu, Jia-Yu Yao, Yong-Hong Tian, Yi-Bing Song, and Li Yuan. Is parameter collision hindering continual learning in llms? arXiv preprint arXiv:2410.10179, 2024. +[66] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. +[67] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. +[68] Jianhui Yu, Hao Zhu, Liming Jiang, Chen Change Loy, Weidong Cai, and Wayne Wu. Celebv-text: A large-scale facial text-video dataset. In CVPR, pages 14805-14814, 2023. +[69] Wangbo Yu, Chaoran Feng, Jiye Tang, Xu Jia, Li Yuan, and Yonghong Tian. Evagaussians: Event stream assisted + +gaussian splatting from blurry images. arXiv preprint arXiv:2405.20224, 2024. +[70] Wangbo Yu, Jinbo Xing, Li Yuan, Wenbo Hu, Xiaoyu Li, Zhipeng Huang, Xiangjun Gao, Tien-Tsin Wong, Ying Shan, and Yonghong Tian. Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis. arXiv preprint arXiv:2409.02048, 2024. +[71] Yongsheng Yu, Ziyun Zeng, Hang Hua, Jianlong Fu, and Jiebo Luo. Promptfix: You prompt and we fix the photo. arXiv preprint arXiv:2405.16785, 2024. +[72] Shenghai Yuan, Jinfa Huang, Yujun Shi, Yongqi Xu, Ruijie Zhu, Bin Lin, Xinhua Cheng, Li Yuan, and Jiebo Luo. Mag-ictime: Time-lapse video generation models as metamorphic simulators. arXiv preprint arXiv:2404.05014, 2024. +[73] Shenghai Yuan, Jina Huang, Yongqi Xu, Yaoyang Liu, Shaofeng Zhang, Yujun Shi, Rui-Jie Zhu, Xinhua Cheng, Jiebo Luo, and Li Yuan. Chronomagic-bench: A benchmark for metamorphic evaluation of text-to-time-lapse video generation. NeurIPS, 37:21236-21270, 2025. +[74] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, pages 3836-3847, 2023. +[75] Zhenghao Zhang, Junchao Liao, Menghao Li, Long Qin, and Weizhi Wang. Tora: Trajectory-oriented diffusion transformer for video generation. arXiv preprint arXiv:2407.21705, 2024. +[76] Mingzhe Zheng, Yongqi Xu, Haojian Huang, Xuran Ma, Yexin Liu, Wenjie Shu, Yatian Pang, Feilong Tang, Qifeng Chen, Harry Yang, et al. Videogen-of-thought: A collaborative framework for multi-shot video generation. arXiv preprint arXiv:2412.02259, 2024. +[77] Zangwei Zheng, Xiangyu Peng, and Yang You. Open-sora: Democratizing efficient video production for all. In *Github*, 2024. +[78] Yuan Zhou, Qiuyue Wang, Yuxuan Cai, and Huan Yang. Allegro: Open the black box of commercial-level video generation model. arXiv preprint arXiv:2410.15458, 2024. \ No newline at end of file diff --git a/2025/Identity-Preserving Text-to-Video Generation by Frequency Decomposition/images.zip b/2025/Identity-Preserving Text-to-Video Generation by Frequency Decomposition/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c4ae5e0ad3aaad4764674799d8c6827c9dd42f83 --- /dev/null +++ b/2025/Identity-Preserving Text-to-Video Generation by Frequency Decomposition/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8963d2e7d9c49894e8ad67681a13728aca5a3d2bac769045da9407025c1d19d3 +size 885064 diff --git a/2025/Identity-Preserving Text-to-Video Generation by Frequency Decomposition/layout.json b/2025/Identity-Preserving Text-to-Video Generation by Frequency Decomposition/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2d7ae1448982db5785301e51d0f56f58877d195e --- /dev/null +++ b/2025/Identity-Preserving Text-to-Video Generation by Frequency Decomposition/layout.json @@ -0,0 +1,9536 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 73, + 103, + 538, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 103, + 538, + 121 + ], + "spans": [ + { + "bbox": [ + 73, + 103, + 538, + 121 + ], + "type": "text", + "content": "Identity-Preserving Text-to-Video Generation by Frequency Decomposition" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 156, + 142, + 455, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 142, + 455, + 173 + ], + "spans": [ + { + "bbox": [ + 156, + 142, + 455, + 173 + ], + "type": "text", + "content": "Shenghai Yuan" + }, + { + "bbox": [ + 156, + 142, + 455, + 173 + ], + "type": "inline_equation", + "content": "^{1,4}" + }, + { + "bbox": [ + 156, + 142, + 455, + 173 + ], + "type": "text", + "content": ", Jina Fu Huang" + }, + { + "bbox": [ + 156, + 142, + 455, + 173 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 156, + 142, + 455, + 173 + ], + "type": "text", + "content": ", Xianyi He" + }, + { + "bbox": [ + 156, + 142, + 455, + 173 + ], + "type": "inline_equation", + "content": "^{1,4}" + }, + { + "bbox": [ + 156, + 142, + 455, + 173 + ], + "type": "text", + "content": ", Yunyang Ge" + }, + { + "bbox": [ + 156, + 142, + 455, + 173 + ], + "type": "inline_equation", + "content": "^{1,4}" + }, + { + "bbox": [ + 156, + 142, + 455, + 173 + ], + "type": "text", + "content": ", Yujun Shi" + }, + { + "bbox": [ + 156, + 142, + 455, + 173 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 156, + 142, + 455, + 173 + ], + "type": "text", + "content": ", Liuhan Chen" + }, + { + "bbox": [ + 156, + 142, + 455, + 173 + ], + "type": "inline_equation", + "content": "^{1,4}" + }, + { + "bbox": [ + 156, + 142, + 455, + 173 + ], + "type": "text", + "content": ", Jiebo Luo" + }, + { + "bbox": [ + 156, + 142, + 455, + 173 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 156, + 142, + 455, + 173 + ], + "type": "text", + "content": ", Li Yuan" + }, + { + "bbox": [ + 156, + 142, + 455, + 173 + ], + "type": "inline_equation", + "content": "^{1,2,\\dagger}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 129, + 184, + 481, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 184, + 481, + 214 + ], + "spans": [ + { + "bbox": [ + 129, + 184, + 481, + 214 + ], + "type": "text", + "content": "1 Peking University, 2 Peng Cheng Laboratory, 3 University of Rochester, 4 Rabbitpre Intelligence, 5 National University of Singapore" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 91, + 229, + 517, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 229, + 517, + 256 + ], + "spans": [ + { + "bbox": [ + 91, + 229, + 517, + 256 + ], + "type": "text", + "content": "{yuanshenghai, HeXianyi, yunyang, chenliuhan}@stu.pku.edu.cn, shi.yujun@u.nus.edu, yuanli-ece@pku.edu.cn, {jhuang90@ur, jluo@cs}.rochester.edu" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 159, + 270, + 451, + 284 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 270, + 451, + 284 + ], + "spans": [ + { + "bbox": [ + 159, + 270, + 451, + 284 + ], + "type": "text", + "content": "https://oku-yuangroup.github.io/ConsisID" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 77, + 308, + 131, + 363 + ], + "blocks": [ + { + "bbox": [ + 81, + 297, + 126, + 306 + ], + "lines": [ + { + "bbox": [ + 81, + 297, + 126, + 306 + ], + "spans": [ + { + "bbox": [ + 81, + 297, + 126, + 306 + ], + "type": "text", + "content": "Reference Image" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 77, + 308, + 131, + 363 + ], + "lines": [ + { + "bbox": [ + 77, + 308, + 131, + 363 + ], + "spans": [ + { + "bbox": [ + 77, + 308, + 131, + 363 + ], + "type": "image", + "image_path": "34b54a7e24fa53311fda96cdaa729a4fdf59d6896abd5e3cc5055d42acb33f7a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 293, + 527, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 293, + 527, + 308 + ], + "spans": [ + { + "bbox": [ + 141, + 293, + 527, + 308 + ], + "type": "text", + "content": "a man standing next to an airplane, engaged in a conversation on his cell phone ... wearings sunglasses and a black top ... talking seriously ... airplane has a green striperunning along its side, and there is a large engine visible behind his ..." + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 135, + 308, + 296, + 363 + ], + "blocks": [ + { + "bbox": [ + 135, + 308, + 296, + 363 + ], + "lines": [ + { + "bbox": [ + 135, + 308, + 296, + 363 + ], + "spans": [ + { + "bbox": [ + 135, + 308, + 296, + 363 + ], + "type": "image", + "image_path": "8c8d0e12247599b8644cccecb1e8528bb9339a8aa8432162a342213397f31948.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 296, + 308, + 375, + 363 + ], + "blocks": [ + { + "bbox": [ + 296, + 308, + 375, + 363 + ], + "lines": [ + { + "bbox": [ + 296, + 308, + 375, + 363 + ], + "spans": [ + { + "bbox": [ + 296, + 308, + 375, + 363 + ], + "type": "image", + "image_path": "59fc6dd022d7e21adb1bcf77fba633461a578cf874bfa8a97dda2da1a0741a47.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 375, + 308, + 454, + 363 + ], + "blocks": [ + { + "bbox": [ + 375, + 308, + 454, + 363 + ], + "lines": [ + { + "bbox": [ + 375, + 308, + 454, + 363 + ], + "spans": [ + { + "bbox": [ + 375, + 308, + 454, + 363 + ], + "type": "image", + "image_path": "1d77819c653754433091187f4d5b7b4c77832ce6411034518c8b17f37798c1fe.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 454, + 308, + 534, + 363 + ], + "blocks": [ + { + "bbox": [ + 454, + 308, + 534, + 363 + ], + "lines": [ + { + "bbox": [ + 454, + 308, + 534, + 363 + ], + "spans": [ + { + "bbox": [ + 454, + 308, + 534, + 363 + ], + "type": "image", + "image_path": "f5e00868e7d1b23805e951a3339837a845c521dfdc07e08e3d5e752c51ec2348.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 77, + 384, + 131, + 438 + ], + "blocks": [ + { + "bbox": [ + 81, + 373, + 126, + 381 + ], + "lines": [ + { + "bbox": [ + 81, + 373, + 126, + 381 + ], + "spans": [ + { + "bbox": [ + 81, + 373, + 126, + 381 + ], + "type": "text", + "content": "Reference Image" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 77, + 384, + 131, + 438 + ], + "lines": [ + { + "bbox": [ + 77, + 384, + 131, + 438 + ], + "spans": [ + { + "bbox": [ + 77, + 384, + 131, + 438 + ], + "type": "image", + "image_path": "f029aaf3ef14bd5451a239f376da728354404bd4b192f53420e07333f5312550.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 136, + 371, + 533, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 371, + 533, + 384 + ], + "spans": [ + { + "bbox": [ + 136, + 371, + 533, + 384 + ], + "type": "text", + "content": "a young woman with long blonde hair ... front of a lush, green bush...white flowers ... wearing a black top ... smiling and looking at the camera while gently touching the flowers on the bush ... then bends downslightly and smells one of the flower..." + } + ] + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 135, + 384, + 215, + 439 + ], + "blocks": [ + { + "bbox": [ + 135, + 384, + 215, + 439 + ], + "lines": [ + { + "bbox": [ + 135, + 384, + 215, + 439 + ], + "spans": [ + { + "bbox": [ + 135, + 384, + 215, + 439 + ], + "type": "image", + "image_path": "d4fc29abb568f78cc3c1aae6314f1ac4aa5f9ad86d00bbd5c0c82a6f50ebe922.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 215, + 384, + 293, + 439 + ], + "blocks": [ + { + "bbox": [ + 215, + 384, + 293, + 439 + ], + "lines": [ + { + "bbox": [ + 215, + 384, + 293, + 439 + ], + "spans": [ + { + "bbox": [ + 215, + 384, + 293, + 439 + ], + "type": "image", + "image_path": "4ae97d4fe9b05646a907ba30623396ef9828e41dc2587d415cf74afaaa34a64a.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 294, + 384, + 373, + 439 + ], + "blocks": [ + { + "bbox": [ + 294, + 384, + 373, + 439 + ], + "lines": [ + { + "bbox": [ + 294, + 384, + 373, + 439 + ], + "spans": [ + { + "bbox": [ + 294, + 384, + 373, + 439 + ], + "type": "image", + "image_path": "bb0f02d3e50739dfe45ecf9ae29d058c0db21b2e9648245af32f966e2c62f110.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 374, + 384, + 454, + 439 + ], + "blocks": [ + { + "bbox": [ + 374, + 384, + 454, + 439 + ], + "lines": [ + { + "bbox": [ + 374, + 384, + 454, + 439 + ], + "spans": [ + { + "bbox": [ + 374, + 384, + 454, + 439 + ], + "type": "image", + "image_path": "2a0dc2c59d8a00a6a74a988c1c62816a5d893d5dcef5710281cc958ca9e768c3.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 454, + 384, + 534, + 439 + ], + "blocks": [ + { + "bbox": [ + 454, + 384, + 534, + 439 + ], + "lines": [ + { + "bbox": [ + 454, + 384, + 534, + 439 + ], + "spans": [ + { + "bbox": [ + 454, + 384, + 534, + 439 + ], + "type": "image", + "image_path": "e676eccfc9c1c145bf46da5125eed96a72a038c9afa02862f13ad3adcf002d92.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 77, + 456, + 131, + 513 + ], + "blocks": [ + { + "bbox": [ + 82, + 447, + 126, + 455 + ], + "lines": [ + { + "bbox": [ + 82, + 447, + 126, + 455 + ], + "spans": [ + { + "bbox": [ + 82, + 447, + 126, + 455 + ], + "type": "text", + "content": "Reference Image" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 77, + 456, + 131, + 513 + ], + "lines": [ + { + "bbox": [ + 77, + 456, + 131, + 513 + ], + "spans": [ + { + "bbox": [ + 77, + 456, + 131, + 513 + ], + "type": "image", + "image_path": "d98e7dada3c07beb38d5374f3537080db851d998e1da545db5a313a0a64df515.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "bbox": [ + 136, + 444, + 531, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 444, + 531, + 457 + ], + "spans": [ + { + "bbox": [ + 136, + 444, + 531, + 457 + ], + "type": "text", + "content": "... a man ... his gaze focused and intense as he holds the basketball ... dressed in athletic gear ... The basketball is positioned firmly in his hands as he gradually puts it down ... urban street court, with the fading light of dusk casting a soft glow over the scene. The man's serious expression and steady grip on the ball ..." + } + ] + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 135, + 457, + 215, + 512 + ], + "blocks": [ + { + "bbox": [ + 135, + 457, + 215, + 512 + ], + "lines": [ + { + "bbox": [ + 135, + 457, + 215, + 512 + ], + "spans": [ + { + "bbox": [ + 135, + 457, + 215, + 512 + ], + "type": "image", + "image_path": "b76fd513743a1e45fba7555bc7cbab5498badd0ab55ab6b030c354e996af2e43.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 215, + 457, + 294, + 512 + ], + "blocks": [ + { + "bbox": [ + 215, + 457, + 294, + 512 + ], + "lines": [ + { + "bbox": [ + 215, + 457, + 294, + 512 + ], + "spans": [ + { + "bbox": [ + 215, + 457, + 294, + 512 + ], + "type": "image", + "image_path": "213b985ccecdf65154508d4fef0220dbefae56257961671ca24a9a8563c25bb5.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 294, + 457, + 373, + 513 + ], + "blocks": [ + { + "bbox": [ + 294, + 457, + 373, + 513 + ], + "lines": [ + { + "bbox": [ + 294, + 457, + 373, + 513 + ], + "spans": [ + { + "bbox": [ + 294, + 457, + 373, + 513 + ], + "type": "image", + "image_path": "074fa865a382c4e2a6d8163a3bb00355d485522016f22ee0d6dc70c35c6f379c.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 374, + 457, + 454, + 513 + ], + "blocks": [ + { + "bbox": [ + 374, + 457, + 454, + 513 + ], + "lines": [ + { + "bbox": [ + 374, + 457, + 454, + 513 + ], + "spans": [ + { + "bbox": [ + 374, + 457, + 454, + 513 + ], + "type": "image", + "image_path": "40a242efbdff3f99b2d0804ed1e3fd81236be0add5f95a9f51e3840423625347.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 454, + 457, + 534, + 513 + ], + "blocks": [ + { + "bbox": [ + 454, + 457, + 534, + 513 + ], + "lines": [ + { + "bbox": [ + 454, + 457, + 534, + 513 + ], + "spans": [ + { + "bbox": [ + 454, + 457, + 534, + 513 + ], + "type": "image", + "image_path": "02d55525786d8af3fedaa0e7f6418306a745c2ebfc18b217fdae8eedf308e0de.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 77, + 531, + 131, + 585 + ], + "blocks": [ + { + "bbox": [ + 81, + 520, + 126, + 528 + ], + "lines": [ + { + "bbox": [ + 81, + 520, + 126, + 528 + ], + "spans": [ + { + "bbox": [ + 81, + 520, + 126, + 528 + ], + "type": "text", + "content": "Reference Image" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 77, + 531, + 131, + 585 + ], + "lines": [ + { + "bbox": [ + 77, + 531, + 131, + 585 + ], + "spans": [ + { + "bbox": [ + 77, + 531, + 131, + 585 + ], + "type": "image", + "image_path": "1171cbd67bc0542229ff2e85b5b63a0b39018981ac1042d8ca7deb3769ec8082.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "bbox": [ + 135, + 515, + 533, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 515, + 533, + 531 + ], + "spans": [ + { + "bbox": [ + 135, + 515, + 533, + 531 + ], + "type": "text", + "content": "a young boy sitting at a table, eating a piece of food. He appears to be enjoying his meal, as he takes a bite and chews it. The boy is wearing a blue shirt and has short hair...background is dark, with some light coming from the left side of the frame. There is a straw visible on the right side of the frame ..." + } + ] + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 135, + 531, + 215, + 585 + ], + "blocks": [ + { + "bbox": [ + 135, + 531, + 215, + 585 + ], + "lines": [ + { + "bbox": [ + 135, + 531, + 215, + 585 + ], + "spans": [ + { + "bbox": [ + 135, + 531, + 215, + 585 + ], + "type": "image", + "image_path": "71329db7f406c6a1885d06dbb74fd5e2e04095284f1cbe247d9d96be9178c128.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 594, + 555, + 617 + ], + "lines": [ + { + "bbox": [ + 55, + 594, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 555, + 617 + ], + "type": "text", + "content": "Figure 1. Examples of identity-preserving video generation (IPT2V) by our ConsisID. Given a reference image, our method can generate realistic and personalized human-centered videos while preserving identity. Red indicates that attributes in long instructions." + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 215, + 531, + 293, + 585 + ], + "blocks": [ + { + "bbox": [ + 215, + 531, + 293, + 585 + ], + "lines": [ + { + "bbox": [ + 215, + 531, + 293, + 585 + ], + "spans": [ + { + "bbox": [ + 215, + 531, + 293, + 585 + ], + "type": "image", + "image_path": "bb7842d95518c076af165e834e81f2c4c5fce71c46dad32125dccaddd5d05696.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 294, + 531, + 373, + 585 + ], + "blocks": [ + { + "bbox": [ + 294, + 531, + 373, + 585 + ], + "lines": [ + { + "bbox": [ + 294, + 531, + 373, + 585 + ], + "spans": [ + { + "bbox": [ + 294, + 531, + 373, + 585 + ], + "type": "image", + "image_path": "bc38c9aec80ace97a6161457a904a8479fad07c0ef314884c34f32619601e83d.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 374, + 531, + 454, + 585 + ], + "blocks": [ + { + "bbox": [ + 374, + 531, + 454, + 585 + ], + "lines": [ + { + "bbox": [ + 374, + 531, + 454, + 585 + ], + "spans": [ + { + "bbox": [ + 374, + 531, + 454, + 585 + ], + "type": "image", + "image_path": "f6e448c4e9073d84f80d5cc990c987cd994c547581405da1d441134acdd42bb6.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 454, + 531, + 534, + 585 + ], + "blocks": [ + { + "bbox": [ + 454, + 531, + 534, + 585 + ], + "lines": [ + { + "bbox": [ + 454, + 531, + 534, + 585 + ], + "spans": [ + { + "bbox": [ + 454, + 531, + 534, + 585 + ], + "type": "image", + "image_path": "4f122182d385fe095ea3583f65c56f81bc0a05b5c1015e4f108f7797a8315b06.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "bbox": [ + 151, + 628, + 200, + 640 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 628, + 200, + 640 + ], + "spans": [ + { + "bbox": [ + 151, + 628, + 200, + 640 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 55, + 654, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 297, + 715 + ], + "type": "text", + "content": "Identity-preserving text-to-video (IPT2V) generation aims to create high-fidelity videos with consistent human identity. It is an important task in video generation but remains an open problem for generative models. This paper pushes the technical frontier of IPT2V in two directions that" + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": "have not been resolved in the literature: (1) A tuning-free pipeline without tedious case-by-case finetuning, and (2) A frequency-aware heuristic identity-preserving Diffusion Transformer (DiT)-based control scheme. To achieve these goals, we propose ConsisID, a tuning-free DiT-based controllable IPT2V model to keep human-identity consistent in the generated video. Inspired by prior findings in fre" + } + ] + } + ], + "index": 43 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 15 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 15 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 15 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 198, + 13, + 443, + 25 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 13, + 443, + 25 + ], + "spans": [ + { + "bbox": [ + 198, + 13, + 443, + 25 + ], + "type": "text", + "content": "Except for this watermark, it is identical to the accepted version;" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "spans": [ + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "type": "text", + "content": "the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "12978" + } + ] + } + ], + "index": 44 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 295, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 295, + 180 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 295, + 180 + ], + "type": "text", + "content": "quency analysis of vision/diffusion transformers, it employs identity-control signals base on frequency domain, since facial features can be decomposed into low-frequency global features (e.g., profile, proportions) and high-frequency intrinsic features (e.g., identity markers that remain unaffected by pose changes). Extensive experiments demonstrate that our frequency-aware heuristic scheme provides an optimal control solution for DiT-based models, making strides towards more effective IPT2V." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 203, + 136, + 216 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 203, + 136, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 203, + 136, + 216 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 223, + 295, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 223, + 295, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 223, + 295, + 342 + ], + "type": "text", + "content": "Large-scale pre-trained video diffusion models [30, 66, 77, 78] have facilitated a variety of downstream applications [49, 54, 69, 70, 72, 74], particularly in identity-preserving text-to-video (IPT2V) [7, 38, 58, 60, 61]. However, existing methods face significant challenges, particularly the high overhead associated with the need for case-by-case finetuning, which diminishes their applicability. Within the open-source community, only the ID-Animator [15] can implement tuning-free IPT2V, but it can only generate videos similar to talking head [57] and has poor id preservation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 343, + 295, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 295, + 582 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 295, + 582 + ], + "type": "text", + "content": "Additionally, the above efforts are predominantly based on U-Net and cannot be adapted to the emerging DiT-based video model [30, 63, 66, 77, 78]. This challenge may stem from the inherent limitations of DiT compared to U-Net, including greater difficulty in training convergence and weakness in perceiving facial details. From some prior findings in frequency analysis of vision/diffusion transformers [2-4, 45, 50, 55, 71], we can know that the reason is: Finding 1: Shallow (e.g., low-level, low-frequency) features are essential for pixel-level prediction tasks in diffusion models, as they ease model training. U-Net facilitates model convergence by aggregating shallow features to the decoder via long skip connections, a mechanism that DiT does not incorporate; Finding 2: Transformers have limited perception of high-frequency information, which is important for preserving facial features. The encoder-decoder architecture of U-Net naturally possesses multi-scale features (e.g., richness in high-frequency), while DiT lacks a comparable structure. To develop a DiT-based control model, these must be addressed first. Please see Appendix for more details." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": "For ID-preserving video generation, the challenges stem from the requirement for each frame to incorporate both high-frequency (e.g., age- and make-up-independent identity markers) and low-frequency information (e.g., facial shape) derived from the reference image, which can just be used to make up for the DiT defects mentioned above. Therefore, we propose ConsisID, to keep the identity consistency in video generation by frequency decomposition, based on the previously Findings of DiT in frequency analysis. Thanks to the large-scale pre-trained DiT, we can use its powerful capabilities to achieve tuning-free effects." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "content": "ConsisID decouples identity features into high- and low-frequency signals, which are injected into specific locations within the DiT, facilitating efficient IPT2V generation. Specifically, in line with Finding 1, we first convert the reference image and the facial key points to the low-frequency signal, then concatenate them with input noise latent to ease the training. Following Finding 2, we utilize a dual-tower feature extractor to capture high-frequency facial information, which is integrated with vision tokens within the transformer block, thereby enhancing the DiT's high-frequency perception capabilities. Finally, to transform the pre-trained model into an IPT2V model and improve its generalization, we further introduce a hierarchical training strategy." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 326, + 228, + 525, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 228, + 525, + 239 + ], + "spans": [ + { + "bbox": [ + 326, + 228, + 525, + 239 + ], + "type": "text", + "content": "Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 241, + 553, + 384 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 314, + 241, + 553, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 241, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 314, + 241, + 553, + 289 + ], + "type": "text", + "content": "- We introduce ConsisID, a tuning-free identity-preserving DiT-based IPT2V model, which preserves the identity of the main subject of the video using control signals from frequency decomposition." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 289, + 553, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 553, + 336 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 553, + 336 + ], + "type": "text", + "content": "- We propose a hierarchical training strategy, including coarse-to-fine training, dynamic mask loss, and dynamic cross-face loss, which work together to facilitate training and enhance generalization effectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 337, + 553, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 337, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 337, + 553, + 384 + ], + "type": "text", + "content": "- Extensive experiments demonstrate that our ConsisID can generate high-quality, editable, consistent identity-preserving videos, benefiting from our frequency-aware identity-preserving T2V DiT-based control scheme." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 394, + 400, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 394, + 400, + 407 + ], + "spans": [ + { + "bbox": [ + 314, + 394, + 400, + 407 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 415, + 555, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 555, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 555, + 666 + ], + "type": "text", + "content": "Tuning-based Identity-preserving T2V Models. Diffusion models are widely recognized for their strong generative capabilities [19, 34-37, 44, 72, 73], significantly advancing the development of identity-preserving generative models [8, 39, 59, 74]. Initially, the researchers used tuning-based methods to generate content that matched the input ID. This process requires finetuning pretrained model for each new person during inference. For example, DreamBooth [47] introduced a novel loss function to fine-tune the entire network, embedding identity information while preserving the original generative capabilities. LoRA [21], similar to DreamBooth [47], requires training only a small subset of network parameters. In contrast, Textual Inversion [11] freezes the pretrained network and embeds identity information into a trainable word embedding. Subsequent tuning-based methods, including both image and video models based on U-Net or DiT architectures [7, 26, 48, 58, 60-62, 65, 76], generally follow three main approaches. While these models demonstrate substantial effectiveness, the requirement to fine-tune for each new identity restricts their practical applicability." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "type": "text", + "content": "Tuning-free Identity-preserving T2V Models. To address the issue of high resource consumption, several tuning-free diffusion models have recently emerged in the field of image generation [13, 14, 29, 56, 67]. These models" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12979" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 70, + 531, + 257 + ], + "blocks": [ + { + "bbox": [ + 82, + 70, + 531, + 257 + ], + "lines": [ + { + "bbox": [ + 82, + 70, + 531, + 257 + ], + "spans": [ + { + "bbox": [ + 82, + 70, + 531, + 257 + ], + "type": "image", + "image_path": "77d7dd582ca5e7bd8036fda0935f30af10ef8bd107c4c743d66195a068f5cca1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 265, + 555, + 301 + ], + "lines": [ + { + "bbox": [ + 54, + 265, + 555, + 301 + ], + "spans": [ + { + "bbox": [ + 54, + 265, + 555, + 301 + ], + "type": "text", + "content": "Figure 2. Overview of the proposed ConsisID. Based on Findings of DiT, low-frequency facial information is embedded into the shallow layers, while high-frequency information is incorporated into the vision tokens within the attention blocks. The ID-preserving Recipe is applied to ease training and improve generalization. The cross face, DropToken and Dropout are executed based on probability." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 320, + 296, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 296, + 596 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 296, + 596 + ], + "type": "text", + "content": "do not require finetuning parameters for newly introduced IDs during inference. For instance, IP-Adapter [67] utilizes the CLIP [43] features of the identity image through cross-attention to guide the pretrained model in generating identity-preserving images. InstantID [56] extends this approach by replacing CLIP [43] features with Arcface [10] features and integrating a pose network to adjust facial proportions. Unlike these initial methods, which introduce control signals via visual tokens, PhotoMaker [29] and Imagine Yourself [16] leverage text tokens. Specifically, PhotoMaker [29] concatenates identity features obtained from the CLIP encoder [43] to the text embedding, while Imagine Yourself [16] uses element-wise addition for feature fusion. In the domain of video generation, only MovieGen [42] and ID-Animator [15] currently support ID-preserving text-to-video (IPT2V) generation. MovieGen is closed-source, whereas ID-Animator is open-source but uses a methodology similar to image models, leading to lower-quality identity preservation in the generated videos. We select the emerging DiT architecture [30, 33, 66, 77] and optimize it for IPT2V, drawing on conclusions from prior frequency analyses [2-4, 45, 50, 55]. This enables high-quality, editable, and consistent ID-preserving video generation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 604, + 138, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 604, + 138, + 618 + ], + "spans": [ + { + "bbox": [ + 55, + 604, + 138, + 618 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 624, + 142, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 142, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 142, + 635 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "Diffusion Model. Text-to-video generation models usually utilize the diffusion paradigm, which gradually transforms noise " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " into a video " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": ". Originally, denoising was conducted directly within the pixel space [20, 51, 52]; however, due to significant computational overheads, recent methods predominantly employ latent space [12, 25, 46," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 320, + 481, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 320, + 481, + 332 + ], + "spans": [ + { + "bbox": [ + 314, + 320, + 481, + 332 + ], + "type": "text", + "content": "72]. The optimization process is defined:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 348, + 342, + 555, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 342, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 348, + 342, + 555, + 365 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {a} = \\mathbb {E} _ {x _ {0}, t, y, \\epsilon} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (x _ {0}, t, \\tau_ {\\theta} (y)) \\| _ {2} ^ {2} \\right], \\qquad (1)", + "image_path": "438ccaefcae5f3bed1e17ced3f73bd64b0488afc995151ee74c7513cf1606f61.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 373, + 555, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 555, + 421 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 555, + 421 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 373, + 555, + 421 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 373, + 555, + 421 + ], + "type": "text", + "content": " is text condition, " + }, + { + "bbox": [ + 313, + 373, + 555, + 421 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 373, + 555, + 421 + ], + "type": "text", + "content": " is sampled from a standard normal distribution (e.g., " + }, + { + "bbox": [ + 313, + 373, + 555, + 421 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0,1)" + }, + { + "bbox": [ + 313, + 373, + 555, + 421 + ], + "type": "text", + "content": "), and " + }, + { + "bbox": [ + 313, + 373, + 555, + 421 + ], + "type": "inline_equation", + "content": "\\tau_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 313, + 373, + 555, + 421 + ], + "type": "text", + "content": " is the text encoder. By replacing " + }, + { + "bbox": [ + 313, + 373, + 555, + 421 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 313, + 373, + 555, + 421 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 373, + 555, + 421 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(x_0)" + }, + { + "bbox": [ + 313, + 373, + 555, + 421 + ], + "type": "text", + "content": ", the latent diffusion is derived, which is used by ConsisID." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 422, + 556, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 556, + 626 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 556, + 626 + ], + "type": "text", + "content": "Diffusion Transformer. The DiT-based video generation model shows significant potential in simulating the physical world [6, 66, 78]. Despite being a novel architecture, research on controllable generation has been limited, and current methods [9, 13, 42, 75] largely resemble U-Net based approaches [11, 39, 74]. However, no study has yet examined why this approach works with DiT. Drawing from prior analyses of Diffusion and Transformer from a frequency domain perspective [2-4, 45, 50, 55], we conclude that: (1) Low-frequency (e.g., shallow-layer) features are essential for pixel-level prediction tasks in diffusion models, which helps facilitate model training; (2) Transformers have limited perception for high-frequency information, which is important for controllable generation. Based on these, we decouple ID features into high- and low-frequency parts and inject them into specific locations, achieving effective identity-preserving text-to-video generation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 635, + 529, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 529, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 529, + 647 + ], + "type": "text", + "content": "3.2. ConsisID: Keep Your Identity Consistent" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "text", + "content": "The overview is illustrated in Figure 2. Given a reference image, the global facial extractor and local facial extractor inject both high- and low-frequency facial information into model, which then generates identity-preserving videos with the assistance of the consistency training strategy." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12980" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 280, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 280, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 280, + 83 + ], + "type": "text", + "content": "3.2.1. Low-frequency View: Global Facial Extractor" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 86, + 296, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 86, + 296, + 317 + ], + "spans": [ + { + "bbox": [ + 55, + 86, + 296, + 317 + ], + "type": "text", + "content": "In light of Finding 1, enhancing low-level (e.g., shallow, low-frequency) features accelerates model convergence. To easily adapt a pre-trained model for the IPT2V task, the most direct approach is concatenating the reference face with the noise input latent [5]. However, the reference face contains both high-frequency details (e.g., eye and lip textures) and low-frequency information (e.g., facial proportions and contours). From Finding 2, prematurely injecting high-frequency information into the Transformer is inefficient and may hinder the model's processing of low-frequency information, as the Transformer focuses primarily on low-frequency features. In addition, feeding the reference face directly into the model could introduce irrelevant noise such as lighting and shadows. To mitigate this, we extract facial key points, convert them to an RGB image, and then concatenate it with the reference image, as shown in Figure 2. This strategy focuses the model's attention on the low-frequency signals in the face, while minimizing the impact of extraneous features. We found that when this component is discarded, the model is difficult to converge. The objective function is changed to:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 80, + 324, + 295, + 339 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 324, + 295, + 339 + ], + "spans": [ + { + "bbox": [ + 80, + 324, + 295, + 339 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {b} = \\mathbb {E} _ {x _ {0}, t, y, f, \\epsilon} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (x _ {0}, t, \\tau_ {\\theta} (y), \\psi_ {\\theta} (f)) \\| _ {2} ^ {2} \\right], \\tag {2}", + "image_path": "31d71ec9a1a22698d3e2b833893898ec9e5cfeb8a190473c92a4969b0527ed49.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 346, + 295, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 346, + 295, + 370 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 295, + 370 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 346, + 295, + 370 + ], + "type": "inline_equation", + "content": "\\psi_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 55, + 346, + 295, + 370 + ], + "type": "text", + "content": " is a variational autoencoder, " + }, + { + "bbox": [ + 55, + 346, + 295, + 370 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 346, + 295, + 370 + ], + "type": "text", + "content": " represents the reference image, we ignore key points here for simplicity." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 376, + 278, + 388 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 278, + 388 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 278, + 388 + ], + "type": "text", + "content": "3.2.2. High-frequency View: Local Facial Extractor" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 390, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 390, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 390, + 295, + 665 + ], + "type": "text", + "content": "In light of Finding 2, we recognize that Transformers have limited sensitivity to high-frequency information. It can be concluded that relying solely on global facial extractor is insufficient for IPT2V generation, as global facial features lack of high-frequency information. So we use a face recognition backbone [10] to extract high-frequency features, as these are invariant to non-ID attributes (e.g., expression, posture, and shape). We refer to these features as intrinsic identity features (e.g., high-frequency), since age and makeup do not alter an individual's core identity. Following [14], we utilize the penultimate layer of the backbone, rather than its output, as it retains more spatial information pertinent to identity. However, our experiments reveal that while the face recognition backbone improves identity consistency, it lacks the semantic features required for editing. This task demands not only maintaining identity consistency but also incorporating the ability to edit, such as generating videos of faces with the same identity but varying age and makeup. Previous research [15, 16] relies solely on the CLIP encoder [43] to enable editing capabilities. However, since CLIP is not specifically trained on face datasets, the features it extracts include irrelevant non-face information, which can compromise identity fidelity [29, 56, 67]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "content": "To address these challenges, we first use a facial recognition backbone to extract features that strongly represent intrinsic identity, and a CLIP image encoder to capture semantically rich features. We then employ the Q-Former" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "text", + "content": "[27, 28, 64] to fuse these features, resulting in intrinsic identity representations enriched with high-frequency semantic information. To mitigate the impact of irrelevant features from CLIP, dropout [1, 22] is applied post-processing. Additionally, we follow [14] to concatenate the shallow, multiscale features from the facial recognition backbone (after interpolation) with the CLIP features. This approach ensures that the model captures essential intrinsic identity features while filtering out extraneous noise unrelated to identity. Finally, we apply cross-attention to facilitate interaction between this feature set and the visual tokens produced by each attention block of the pre-trained model, thereby enhancing the high-frequency information in the DiT:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 361, + 236, + 555, + 251 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 236, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 361, + 236, + 555, + 251 + ], + "type": "interline_equation", + "content": "Z _ {i} ^ {\\prime} = Z _ {i} + \\text {A t t e n t i o n} \\left(Q _ {i} ^ {v}, K _ {i} ^ {f}, V _ {i} ^ {f}\\right), \\tag {3}", + "image_path": "0731f043ebc7f9a4c2ff76b66677ce205379ef8aef498274efc755aca9e5dedb.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 259, + 555, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 259, + 555, + 316 + ], + "spans": [ + { + "bbox": [ + 313, + 259, + 555, + 316 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 259, + 555, + 316 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 259, + 555, + 316 + ], + "type": "text", + "content": " represents the layer number of the attention block, " + }, + { + "bbox": [ + 313, + 259, + 555, + 316 + ], + "type": "inline_equation", + "content": "Q^{v} = Z_{i}W_{i}^{q}" + }, + { + "bbox": [ + 313, + 259, + 555, + 316 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 259, + 555, + 316 + ], + "type": "inline_equation", + "content": "K^{f} = FW_{i}^{k}" + }, + { + "bbox": [ + 313, + 259, + 555, + 316 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 259, + 555, + 316 + ], + "type": "inline_equation", + "content": "V^{f} = FW_{i}^{v}" + }, + { + "bbox": [ + 313, + 259, + 555, + 316 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 259, + 555, + 316 + ], + "type": "inline_equation", + "content": "Z_{i}" + }, + { + "bbox": [ + 313, + 259, + 555, + 316 + ], + "type": "text", + "content": " is the visual token, " + }, + { + "bbox": [ + 313, + 259, + 555, + 316 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 313, + 259, + 555, + 316 + ], + "type": "text", + "content": " represents the intrinsic identity features, and " + }, + { + "bbox": [ + 313, + 259, + 555, + 316 + ], + "type": "inline_equation", + "content": "W_{q}, W_{k}" + }, + { + "bbox": [ + 313, + 259, + 555, + 316 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 259, + 555, + 316 + ], + "type": "inline_equation", + "content": "W_{v}" + }, + { + "bbox": [ + 313, + 259, + 555, + 316 + ], + "type": "text", + "content": " are trainable parameters. The objective function is changed to:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 320, + 323, + 555, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 323, + 555, + 338 + ], + "spans": [ + { + "bbox": [ + 320, + 323, + 555, + 338 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c} = \\mathbb {E} _ {x _ {0}, t, y, f, \\epsilon} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(x _ {0}, t, \\tau_ {\\theta} (y), \\psi_ {\\theta} (f), \\varphi_ {\\theta} (f)\\right) \\| _ {2} ^ {2} \\right], \\tag {4}", + "image_path": "8b91d0f88b7008a9353f7896c5b183a3e9601e5bc2412d593966d7a743fd0443.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 344, + 476, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 344, + 476, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 344, + 476, + 357 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 344, + 476, + 357 + ], + "type": "inline_equation", + "content": "\\varphi_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 313, + 344, + 476, + 357 + ], + "type": "text", + "content": " is the local facial extractor." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 363, + 471, + 375 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 363, + 471, + 375 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 471, + 375 + ], + "type": "text", + "content": "3.2.3. Consistency Training Strategy" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 378, + 554, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 554, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 554, + 426 + ], + "type": "text", + "content": "During training, we randomly select a frame from the training frames and apply the Crop & Align [10] to extract the facial region as reference images, which is subsequently used as an identity-control signal, alongside the text as control." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 426, + 555, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 641 + ], + "type": "text", + "content": "Coarse-to-Fine Training. Compared to Identity-preserving image generation, video generation requires maintaining consistency in both spatial and temporal dimensions, ensuring that high and low-frequency facial information matches the reference image. To mitigate the complexity of training, we propose a hierarchical strategy where the model learns information globally before refining it locally. In the coarse-grained phase (e.g., corresponding Finding 1), we employ the global facial extractor, enabling the model to prioritize low-frequency features, such as facial contours and proportions, thereby ensuring rapid acquisition of identity information from the reference image and consistency across the video sequence. In the fine-grained phase (e.g. corresponding to Finding 2), the local facial extractor shifts the model's focus to high-frequency details, such as the texture details of eyes and lips (e.g., intrinsic identification), improving the fidelity of facial expressions and the overall similarity of the generated face." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "Dynamic Mask Loss. The objective of our task is to ensure that the identity of the person in the generated video remains consistent with the input reference image. However, Equation 1 considers the entire scene, encompassing both high- and low-frequency identity information as well as redundant background content, which introduces noise that" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12981" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 73, + 70, + 280, + 203 + ], + "blocks": [ + { + "bbox": [ + 73, + 70, + 280, + 203 + ], + "lines": [ + { + "bbox": [ + 73, + 70, + 280, + 203 + ], + "spans": [ + { + "bbox": [ + 73, + 70, + 280, + 203 + ], + "type": "image", + "image_path": "237290a6b297b7c75a39cec96650554bfd2c89fc1bf7c86b8c925d0f983ba3d2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 209, + 297, + 233 + ], + "lines": [ + { + "bbox": [ + 55, + 209, + 297, + 233 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 297, + 233 + ], + "type": "text", + "content": "Figure 3. User Study between ConsisID and state-of-the-art methods. ConsisID is preferred by voters in all dimensions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 250, + 296, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 250, + 296, + 310 + ], + "spans": [ + { + "bbox": [ + 55, + 250, + 296, + 310 + ], + "type": "text", + "content": "interferes with model training. To address this, we propose to focus the model's attention on face regions. Specifically, we first extract the facial mask from the video, apply trilinear interpolation to map it to the latent space, and finally use this mask to constrain the computation of " + }, + { + "bbox": [ + 55, + 250, + 296, + 310 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 55, + 250, + 296, + 310 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 143, + 316, + 295, + 328 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 316, + 295, + 328 + ], + "spans": [ + { + "bbox": [ + 143, + 316, + 295, + 328 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {d} = M \\odot \\mathcal {L} _ {c}, \\tag {5}", + "image_path": "a7c49098a870fb6d44039679a89f85413a5add20dc134ae83dd7973d6f4624c1.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 334, + 296, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 334, + 296, + 394 + ], + "spans": [ + { + "bbox": [ + 55, + 334, + 296, + 394 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 334, + 296, + 394 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 334, + 296, + 394 + ], + "type": "text", + "content": " represents a mask with the same shape as " + }, + { + "bbox": [ + 55, + 334, + 296, + 394 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 334, + 296, + 394 + ], + "type": "text", + "content": ". However, if Equation 5 is used as the supervisory signal for all training data, the model may fail to generate a natural background during inference. To mitigate this issue, we apply Equation 5 with a probability " + }, + { + "bbox": [ + 55, + 334, + 296, + 394 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 334, + 296, + 394 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 55, + 334, + 296, + 394 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 334, + 296, + 394 + ], + "type": "text", + "content": ", resulting in:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 127, + 399, + 295, + 431 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 399, + 295, + 431 + ], + "spans": [ + { + "bbox": [ + 127, + 399, + 295, + 431 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {e} = \\left\\{ \\begin{array}{l l} \\mathcal {L} _ {d}, & \\text {i f} p > \\alpha \\\\ \\mathcal {L} _ {c}, & \\text {i f} p \\leq \\alpha \\end{array} \\right. \\tag {6}", + "image_path": "a1201e2136637cc32e2b170a2d177d95fc92cb1097d93d0e167cc2f3c047128f.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 434, + 296, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 296, + 567 + ], + "type": "text", + "content": "Dynamic Cross-Face Loss. After training with Equation 6, we observed that the model struggled to generate satisfactory results for persons not present in the data domain during inference. This issue arises because the model, trained exclusively on faces from the training frames, tends to overfit by adopting a \"copy-paste\" shortcut—essentially replicating the reference image without alteration. To improve the model's generalization capability, we introduce slight Gaussian noise " + }, + { + "bbox": [ + 55, + 434, + 296, + 567 + ], + "type": "inline_equation", + "content": "\\zeta" + }, + { + "bbox": [ + 55, + 434, + 296, + 567 + ], + "type": "text", + "content": " to the reference images and use cross-face (e.g., reference images are sourced from video frames outside the training frames) as inputs with probability " + }, + { + "bbox": [ + 55, + 434, + 296, + 567 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 434, + 296, + 567 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 94, + 571, + 295, + 604 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 571, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 94, + 571, + 295, + 604 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {f} = \\left\\{ \\begin{array}{l l} \\mathcal {L} _ {e} & \\text {w h e r e} x _ {0} \\cdot \\zeta , \\quad \\text {i f} p > \\beta \\\\ \\mathcal {L} _ {e} & \\text {w h e r e} x _ {c} \\cdot \\zeta , \\quad \\text {i f} p \\leq \\beta \\end{array} \\right. \\tag {7}", + "image_path": "ef81b26a39a168d6c3a250cd563a8ff01a69861ca44a9cf403782ed2eec53da4.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 608, + 295, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 608, + 295, + 632 + ], + "spans": [ + { + "bbox": [ + 55, + 608, + 295, + 632 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 608, + 295, + 632 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 55, + 608, + 295, + 632 + ], + "type": "text", + "content": " is the reference image extracted from the training frames, and " + }, + { + "bbox": [ + 55, + 608, + 295, + 632 + ], + "type": "inline_equation", + "content": "x_{c}" + }, + { + "bbox": [ + 55, + 608, + 295, + 632 + ], + "type": "text", + "content": " is extracted from outside the training frames." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 640, + 137, + 654 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 640, + 137, + 654 + ], + "spans": [ + { + "bbox": [ + 55, + 640, + 137, + 654 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 660, + 106, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 660, + 106, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 660, + 106, + 673 + ], + "type": "text", + "content": "4.1. Setup" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Implementation details. ConsisID selects DiT-based generation architectures CogVideoX-5B [66] as our baseline for validation. We use an in-house human-centric" + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 317, + 70, + 553, + 106 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 553, + 106 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 553, + 106 + ], + "type": "table", + "html": "
FaceSim-Arc ↑FaceSim-Cur ↑CLIPScore ↑FID ↓
ID-Animator [15]0.320.3324.97117.46
ConsisID0.580.6027.93151.82
", + "image_path": "8175fcae909c6e9ebb18b4a75ac0860eaaacd961bb39c585fdff38493f9aade9.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 114, + 555, + 148 + ], + "lines": [ + { + "bbox": [ + 313, + 114, + 555, + 148 + ], + "spans": [ + { + "bbox": [ + 313, + 114, + 555, + 148 + ], + "type": "text", + "content": "Table 1. Quantitative comparison with state-of-the-art methods. ConsID achieve well-aligned results across most metrics. \"↓\" denotes lower is better. \"↑\" denotes higher is better." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 163, + 555, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 163, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 163, + 555, + 331 + ], + "type": "text", + "content": "dataset for training, which differs from previous datasets [40, 57, 68] that focus only on the face. In the training phase, we set the resolution to " + }, + { + "bbox": [ + 313, + 163, + 555, + 331 + ], + "type": "inline_equation", + "content": "480 \\times 720" + }, + { + "bbox": [ + 313, + 163, + 555, + 331 + ], + "type": "text", + "content": " and extracted 49 consecutive frames at a stride of 3 from each video as training data. We set the batch size to 80, the learning rate to " + }, + { + "bbox": [ + 313, + 163, + 555, + 331 + ], + "type": "inline_equation", + "content": "3 \\times 10^{-6}" + }, + { + "bbox": [ + 313, + 163, + 555, + 331 + ], + "type": "text", + "content": ", and the total number of training steps to " + }, + { + "bbox": [ + 313, + 163, + 555, + 331 + ], + "type": "inline_equation", + "content": "1.8\\mathrm{k}" + }, + { + "bbox": [ + 313, + 163, + 555, + 331 + ], + "type": "text", + "content": ". The randomly discarded text rate is set to 0.1, with AdamW serving as an optimizer and cosine_with_restarts as a learning rate scheduler. The training strategy is the same as Section 3.2.3. We set " + }, + { + "bbox": [ + 313, + 163, + 555, + 331 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 163, + 555, + 331 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 163, + 555, + 331 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 163, + 555, + 331 + ], + "type": "text", + "content": " in the dynamic cross-face loss " + }, + { + "bbox": [ + 313, + 163, + 555, + 331 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_e)" + }, + { + "bbox": [ + 313, + 163, + 555, + 331 + ], + "type": "text", + "content": " and dynamic mask loss " + }, + { + "bbox": [ + 313, + 163, + 555, + 331 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_f)" + }, + { + "bbox": [ + 313, + 163, + 555, + 331 + ], + "type": "text", + "content": " to 0.5, respectively. In the inference phase, we employ DPM [51] with a sampling step of 50, and a classifier free guidance ratio of 6.0. For more details and results, please refer to Appendix." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 331, + 556, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 556, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 556, + 559 + ], + "type": "text", + "content": "Benchmark. Since there is an absence of an evaluation dataset, we select 30 persons who were not included in the training data and sourced five high-quality images for each ID from the internet. We then design 90 distinct prompts, encompassing a variety of expressions, actions, and backgrounds for evaluation. Building on previous works [15, 42], we evaluate four dimensions: (1). Identity Preservation: We use FaceSim-Arc [10] and introduce FaceSim-Cur, which assesses identity preservation by measuring feature differences between face regions in the generated videos and those in real face images within the ArcFace [10] and CurricularFace [23] feature spaces. (2) Visual Quality: We utilize FID [18] by calculating feature differences in the face regions between the generated frames and real face images within the InceptionV3 [53] feature space. (3) Text Relevance: We utilize CLIPScore [17] to measure the similarity between the generated videos and the input prompts. (4). Motion Amplitude: Due to the lack of reliable metrics [24, 73], we evaluate through the user study." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 564, + 432, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 564, + 432, + 577 + ], + "spans": [ + { + "bbox": [ + 313, + 564, + 432, + 577 + ], + "type": "text", + "content": "4.2. Qualitative Analysis" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": "In this section, we compare our method, ConsisID, with ID-Animator [15] (e.g., the only available open-source model) for tuning-free IPT2V tasks. We randomly select images and text prompts of four individuals for qualitative analysis, all of which are absent from the training data. As shown in Figure 4, ID-Animator cannot generate human body parts beyond the face and is unable to generate complex actions or backgrounds in response to text prompts (e.g., action, attribute, background), which significantly limits its practical application. In addition, the preservation of the identity is inadequate; for example, in case 1, the reference image" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12982" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 100, + 69, + 511, + 475 + ], + "blocks": [ + { + "bbox": [ + 100, + 69, + 511, + 475 + ], + "lines": [ + { + "bbox": [ + 100, + 69, + 511, + 475 + ], + "spans": [ + { + "bbox": [ + 100, + 69, + 511, + 475 + ], + "type": "image", + "image_path": "f2181427b3d85c505fa48e2347daf297da2ffd5a21badf93ca0264b3f610026d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 483, + 555, + 529 + ], + "lines": [ + { + "bbox": [ + 55, + 483, + 555, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 555, + 529 + ], + "type": "text", + "content": "Figure 4. Qualitative analysis between ConsisID and ID-Animator [15]. ID-Animator can only generate videos of the face region, and the identity Preservation is poor (e.g., shape, texture). Additionally, it cannot generate specified content according to the text prompt (e.g., action, decoration, background). ConsisID achieves advantages in identity preservation, visual quality, motion amplitude, and text relevance. Moreover, our ConsistID can generate more frames rather than ID-Animator (49 " + }, + { + "bbox": [ + 55, + 483, + 555, + 529 + ], + "type": "inline_equation", + "content": "480 \\times 720" + }, + { + "bbox": [ + 55, + 483, + 555, + 529 + ], + "type": "text", + "content": "p frames v.s. 16 " + }, + { + "bbox": [ + 55, + 483, + 555, + 529 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 55, + 483, + 555, + 529 + ], + "type": "text", + "content": "p frames)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 544, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 628 + ], + "type": "text", + "content": "appears to be processed with skin smoothing. In case 2, wrinkles have been introduced which detract from the aesthetic quality. In cases 3 and 4, the face is distorted due to the lack of low frequency information, which compromises identity consistency. In contrast, the proposed ConsisID consistently produces high-quality, realistic videos that accurately match the reference identity and adhere to prompt." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 635, + 180, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 180, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 180, + 649 + ], + "type": "text", + "content": "4.3. Quantitative Analysis" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "content": "We present a comprehensive quantitative evaluation of different methods, with results displayed in Table 1. Consistent with Figure 4, our method outperforms state-of-the-art methods across five metrics. For identity preservation, ConsisID achieves a higher score by designing appropri" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 544, + 555, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 555, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 555, + 665 + ], + "type": "text", + "content": "ate identity signals for DiT from a frequency perspective. By contrast, ID-Animator [15] is not optimized for IPT2V and only partially retains facial features, resulting in lower FaceSim-Arc [10] and FaceSim-Cur scores. For Text Relevance, ConsisID not only controls expressions via prompts but also adjusts actions and backgrounds, achieving higher CLIPScore [17]. Regarding visual quality, the FID is presented solely as a reference due to its limited alignment [31, 32, 41, 73] with human perception. Please refer to Figure 4 and 3 for qualitative analysis of the visual quality." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 672, + 389, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 389, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 389, + 685 + ], + "type": "text", + "content": "4.4. User Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "Building on previous work, we conduct a human evaluation using a binary voting strategy, with each questionnaire con" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12983" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 290, + 388 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 290, + 388 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 290, + 388 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 290, + 388 + ], + "type": "image", + "image_path": "bbc2af07b9696d973d07ea1692bd0d7a753454559e973d94a8eed714283f165e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 396, + 295, + 441 + ], + "lines": [ + { + "bbox": [ + 55, + 396, + 295, + 441 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 295, + 441 + ], + "type": "text", + "content": "Figure 5. Effect of Different Components via Qualitative Analysis. Removing any component may result in the loss of high- or low-frequency facial information, or hinder the ability to modify video content based on the text prompt." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 461, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 461, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 461, + 295, + 605 + ], + "type": "text", + "content": "taining only 80 questions. Participants are required to view 40 video clips, a setup designed to improve both engagement and questionnaire validity. For the IPT2V task, each question requires participants to separately judge which option performs better in terms of Identity Preservation, Visual Quality, Text Alignment, and Motion Amplitude. This composition ensures the accuracy of the human evaluation. Owing to the extensive participant base required for this evaluation, we successfully gathered 103 valid questionnaires. The results, depicted in Figure 3, demonstrate a significant superiority of our method over ID-Animator [15], verifying the effectiveness of the designed DiT for IPT2V generation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 612, + 285, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 285, + 625 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 285, + 625 + ], + "type": "text", + "content": "4.5. Effect of the Identity Signal Injection in DiT" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "text", + "content": "To assess the effectiveness of Finding 1 and Finding 2, we perform ablation studies on different methods of injecting control signals into DiT. Specifically, these experiments involved (a) injecting only low-frequency face information with key points into the noise latent, (b) injecting only high-frequency face signals within the attention block, (c) combining (a) and (b), (d) based on (c), but the low-frequency" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 319, + 69, + 549, + 398 + ], + "blocks": [ + { + "bbox": [ + 319, + 69, + 549, + 398 + ], + "lines": [ + { + "bbox": [ + 319, + 69, + 549, + 398 + ], + "spans": [ + { + "bbox": [ + 319, + 69, + 549, + 398 + ], + "type": "image", + "image_path": "a4e4ec88eb66f94d25bba40665477d32a1841b3a6105327195353b440640a845.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 406, + 555, + 440 + ], + "lines": [ + { + "bbox": [ + 313, + 406, + 555, + 440 + ], + "spans": [ + { + "bbox": [ + 313, + 406, + 555, + 440 + ], + "type": "text", + "content": "Figure 6. Effect of Different Control Signal Injection Way via Qualitative Analysis. Only (c), which injects both high & low-freq face signals into the suitable location, performs best." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 317, + 453, + 553, + 530 + ], + "blocks": [ + { + "bbox": [ + 317, + 453, + 553, + 530 + ], + "lines": [ + { + "bbox": [ + 317, + 453, + 553, + 530 + ], + "spans": [ + { + "bbox": [ + 317, + 453, + 553, + 530 + ], + "type": "table", + "html": "
FaceSim-Arc ↑FaceSim-Cur ↑CLIPScore ↑FID ↓
w/o GFE0.050.0534.86269.88
w/o LFE0.660.6834.48104.34
w/o CFT0.540.5834.47144.62
w/o DML0.620.6734.23187.78
w/o DCL0.650.6932.21117.80
ConsisID0.730.7536.77127.42
", + "image_path": "ecd90b4dcd2df5ca06fab6db8ddc6c83f769c445101a2993969049e171352c63.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 537, + 555, + 594 + ], + "lines": [ + { + "bbox": [ + 313, + 537, + 555, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 537, + 555, + 594 + ], + "type": "text", + "content": "Table 2. Effect of Local Facial Extractor (LFE), Global Facial Extractor (GFE), coarse-to-fine training (CFT), dynamic mask loss (DML) and dynamic cross-face loss (DCL) by Automatic Metrics. Removing any of the above methods significantly reduces identity preservation, text relevance, and visual quality." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "content": "face information does not contain key points, and " + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "inline_equation", + "content": "(\\mathbf{e} - \\mathbf{f})" + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "content": " based on (c), but the high-frequency signal is injected at the output or input of the attention block. (g) injecting only high-frequency face signals before the attention block. To reduce overhead, for each identity, we only select 2 reference images each with 90 text prompts for the evaluation. The results are shown in Figure 6 and Table 3. For Finding " + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "content": ", we observe that only injecting high-frequency signals (a)" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12984" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 294, + 162 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 294, + 162 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 294, + 162 + ], + "type": "table", + "html": "
PlanFaceSim-Arc ↑FaceSim-Cur ↑CLIPScore ↑FID ↓
a0.050.0534.86269.88
b0.660.6834.48104.34
c0.730.7536.77127.42
d0.640.6830.69177.65
e0.620.6633.61164.15
funstable training process
gunstable training process
", + "image_path": "d0a636a1eb123b39bfe870fe69bbc7d637cda5a173efb0abb6ae04f544db96aa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 170, + 295, + 204 + ], + "lines": [ + { + "bbox": [ + 55, + 170, + 295, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 170, + 295, + 204 + ], + "type": "text", + "content": "Table 3. Effect of Different Control Signal Injection Way via Quantitative Analysis. Only plan c, which injects both high and low-frequency face information into the model, performs best." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 56, + 222, + 296, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 222, + 296, + 619 + ], + "spans": [ + { + "bbox": [ + 56, + 222, + 296, + 619 + ], + "type": "text", + "content": "greatly increases the training difficulty, causing the model to fail to converge due to the lack of low-frequency signal injection. In addition, the inclusion of facial key points (d) allows a greater focus on low-frequency information, thereby facilitating training and improving model performance. For Finding 2, when only low-frequency signals are injected (b), the model lacks high-frequency information. This reliance on low-frequency signals causes the generated face in the video to copy the reference image, making it difficult to control facial expressions, movements, and other features through prompts. Furthermore, injecting identity signals into the attention block input (f - g) disrupts the intended frequency domain distribution of DiT, resulting in a gradient explosion. Embedding control signals in the attention block (c) is preferable to embedding them in the output (e) because attention block processes predominantly low-frequency information. By embedding high-frequency information internally, the attention block is guided to highlight intrinsic facial features, whereas injecting it into the output merely concatenates features without directing focus, reducing DiT's modeling capacity. Moreover, we apply a Fourier transform to the generated videos (only the face region) to visually compare the influence of different components to extract facial information. As shown in Figure 7, the Fourier spectrum and the log amplitude of the Fourier transform reveal that injecting high or low-frequency signals can indeed enhance the corresponding frequency information of the generated face. Moreover, the low-frequency signal can be further enhanced by matching with the face key points, and injecting the high-frequency signal into the attention block has the highest feature utilization rate. Our method (c) shows strongest high and low frequency, further validating the efficiency benefit from Findings 1 and 2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 624, + 295, + 638 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 295, + 638 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 295, + 638 + ], + "type": "text", + "content": "4.6. Ablation on the Consistency Training Strategy" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "To reduce overhead, for each identity, we only select 2 reference images for the following experiments. To demonstrate the benefits of the proposed consistency training strategy, we perform ablation experiments on coarse-to-fine training (CFT), dynamic mask loss " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_e" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": " (DML), and dynamic crossface loss " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_f" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": " (DCL), with the results presented in Figure" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 326, + 70, + 386, + 127 + ], + "blocks": [ + { + "bbox": [ + 326, + 70, + 386, + 127 + ], + "lines": [ + { + "bbox": [ + 326, + 70, + 386, + 127 + ], + "spans": [ + { + "bbox": [ + 326, + 70, + 386, + 127 + ], + "type": "image", + "image_path": "65f5d2294d04dd308479a5f1c85c79ba40e4ffc3e649d056ce20ff6373cad2f4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 333, + 128, + 378, + 135 + ], + "lines": [ + { + "bbox": [ + 333, + 128, + 378, + 135 + ], + "spans": [ + { + "bbox": [ + 333, + 128, + 378, + 135 + ], + "type": "text", + "content": "(a) only high-freq" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 402, + 71, + 462, + 128 + ], + "blocks": [ + { + "bbox": [ + 402, + 71, + 462, + 128 + ], + "lines": [ + { + "bbox": [ + 402, + 71, + 462, + 128 + ], + "spans": [ + { + "bbox": [ + 402, + 71, + 462, + 128 + ], + "type": "image", + "image_path": "00d4ad5dfc734192981137bf8c8ffcc15c55685a44ab405736cb53e07c0448bd.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 410, + 128, + 454, + 135 + ], + "lines": [ + { + "bbox": [ + 410, + 128, + 454, + 135 + ], + "spans": [ + { + "bbox": [ + 410, + 128, + 454, + 135 + ], + "type": "text", + "content": "(b) only low-freq" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 479, + 71, + 537, + 128 + ], + "blocks": [ + { + "bbox": [ + 479, + 71, + 537, + 128 + ], + "lines": [ + { + "bbox": [ + 479, + 71, + 537, + 128 + ], + "spans": [ + { + "bbox": [ + 479, + 71, + 537, + 128 + ], + "type": "image", + "image_path": "7ae2c20bdfd58558e5b664029efe9b17a3c80dae5f5588743999e2841a4c32e0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 482, + 128, + 533, + 135 + ], + "lines": [ + { + "bbox": [ + 482, + 128, + 533, + 135 + ], + "spans": [ + { + "bbox": [ + 482, + 128, + 533, + 135 + ], + "type": "text", + "content": "(c) high & low freq" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 326, + 137, + 386, + 194 + ], + "blocks": [ + { + "bbox": [ + 326, + 137, + 386, + 194 + ], + "lines": [ + { + "bbox": [ + 326, + 137, + 386, + 194 + ], + "spans": [ + { + "bbox": [ + 326, + 137, + 386, + 194 + ], + "type": "image", + "image_path": "2497ec5f968a9e5b83550d1f63698a3c2b580d2106247ad2b398c9c5a9e3d7aa.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 328, + 195, + 382, + 202 + ], + "lines": [ + { + "bbox": [ + 328, + 195, + 382, + 202 + ], + "spans": [ + { + "bbox": [ + 328, + 195, + 382, + 202 + ], + "type": "text", + "content": "(d) high & low† freq" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 392, + 140, + 539, + 259 + ], + "blocks": [ + { + "bbox": [ + 392, + 140, + 539, + 259 + ], + "lines": [ + { + "bbox": [ + 392, + 140, + 539, + 259 + ], + "spans": [ + { + "bbox": [ + 392, + 140, + 539, + 259 + ], + "type": "image", + "image_path": "5650e498907d8bcca60ce89715d5620898a31a3ea6a7ea27f5eb5b79cd51311c.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 399, + 261, + 542, + 270 + ], + "lines": [ + { + "bbox": [ + 399, + 261, + 542, + 270 + ], + "spans": [ + { + "bbox": [ + 399, + 261, + 542, + 270 + ], + "type": "text", + "content": "(f) Relative log amplitudes of Fourier transformed videos" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 326, + 205, + 386, + 262 + ], + "blocks": [ + { + "bbox": [ + 326, + 205, + 386, + 262 + ], + "lines": [ + { + "bbox": [ + 326, + 205, + 386, + 262 + ], + "spans": [ + { + "bbox": [ + 326, + 205, + 386, + 262 + ], + "type": "image", + "image_path": "64bb93253f7a3f9033c3b75436b212918d74de990232fcfa5ac589ca23e6c566.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 262, + 382, + 269 + ], + "lines": [ + { + "bbox": [ + 329, + 262, + 382, + 269 + ], + "spans": [ + { + "bbox": [ + 329, + 262, + 382, + 269 + ], + "type": "text", + "content": "(e) high+ & low freq" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 278, + 555, + 344 + ], + "lines": [ + { + "bbox": [ + 313, + 278, + 555, + 344 + ], + "spans": [ + { + "bbox": [ + 313, + 278, + 555, + 344 + ], + "type": "text", + "content": "Figure 7. (a - e) Fourier spectrum of different id signal injection. The center area represents low frequencies and the surrounding area represents high frequencies. (f) Relative log amplitudes of Fourier transformed generated videos. A larger response value indicates a higher inclusion of frequency information. (a - f) verify the effect of our frequency decomposition." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 365, + 555, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 555, + 497 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 555, + 497 + ], + "type": "text", + "content": "5 and Table 2. When CFT is removed, GFE and LFE exhibit competing behaviors, complicating the model's ability to prioritize high and low-frequency information accurately, leading to convergence at suboptimal points. Removing DML required the model to simultaneously focus on both foreground and background elements, with background noise negatively affecting training and reducing facial consistency. Similarly, the exclusion of DCL impaired the generalization capability, reducing fidelity for faces, not in the training set and reducing its effectiveness in generating identity-preserving videos as intended." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 507, + 388, + 520 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 507, + 388, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 507, + 388, + 520 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 527, + 555, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 527, + 555, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 527, + 555, + 635 + ], + "type": "text", + "content": "In this paper, we present ConsisID, a unified framework for keeping faces consistent in video generation by frequency decomposition. It can seamlessly integrate into existing DiT-based text-to-video models, for generating high-quality, editable, consistent identity-preserving videos. Extensive experiments show that ConsisID outperforms the current state-of-the-art identity-preserving T2V models. It reveals that our frequency-aware heuristic DiT-based control scheme is an optimal solution for IPT2V generation." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 646, + 424, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 646, + 424, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 646, + 424, + 659 + ], + "type": "text", + "content": "6. Acknowledgments" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "text", + "content": "We thank all the anonymous reviewers for their constructive comments. This work was supported in part by the Natural Science Foundation of China (No. 62202014, 62332002, 62425101, 62088102)." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12985" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "text", + "content": "[1] Hassan Akbari, Liangzhe Yuan, Rui Qian, Wei-Hong Chuang, Shih-Fu Chang, Yin Cui, and Boqing Gong. Vatt: Transformers for multimodal self-supervised learning from raw video, audio and text. NeurIPS, 34:24206-24221, 2021." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 137, + 296, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 296, + 168 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 296, + 168 + ], + "type": "text", + "content": "[2] Dosovitskiy Alexey. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv: 2010.11929, 2020." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 171, + 295, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 171, + 295, + 214 + ], + "spans": [ + { + "bbox": [ + 62, + 171, + 295, + 214 + ], + "type": "text", + "content": "[3] Jiawang Bai, Li Yuan, Shu-Tao Xia, Shuicheng Yan, Zhifeng Li, and Wei Liu. Improving vision transformers by revisiting high-frequency components. In ECCV, pages 1-18. Springer, 2022." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 216, + 294, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 216, + 294, + 249 + ], + "spans": [ + { + "bbox": [ + 62, + 216, + 294, + 249 + ], + "type": "text", + "content": "[4] Fan Bao, Shen Nie, Kaiwen Xue, Yue Cao, Chongxuan Li, Hang Su, and Jun Zhu. All are worth words: A vit backbone for diffusion models. In CVPR, pages 22669-22679, 2023." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 251, + 295, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 251, + 295, + 304 + ], + "spans": [ + { + "bbox": [ + 62, + 251, + 295, + 304 + ], + "type": "text", + "content": "[5] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 307, + 294, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 307, + 294, + 361 + ], + "spans": [ + { + "bbox": [ + 62, + 307, + 294, + 361 + ], + "type": "text", + "content": "[6] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. In openai, 2024." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 363, + 295, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 363, + 295, + 418 + ], + "spans": [ + { + "bbox": [ + 62, + 363, + 295, + 418 + ], + "type": "text", + "content": "[7] Hila Chefer, Shiran Zada, Roni Paiss, Ariel Ephrat, Omer Tov, Michael Rubinstein, Lior Wolf, Tali Dekel, Tomer Michaeli, and Inbar Mosseri. Still-moving: Customized video generation without customized video data. arXiv preprint arXiv:2407.08674, 2024." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 419, + 295, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 419, + 295, + 474 + ], + "spans": [ + { + "bbox": [ + 62, + 419, + 295, + 474 + ], + "type": "text", + "content": "[8] Li Chen, Mengyi Zhao, Yiheng Liu, Mingxu Ding, Yangyang Song, Shizun Wang, Xu Wang, Hao Yang, Jing Liu, Kang Du, et al. Photovverse: Tuning-free image customization with text-to-image diffusion models. arXiv preprint arXiv:2309.05793, 2023." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 475, + 295, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 475, + 295, + 519 + ], + "spans": [ + { + "bbox": [ + 62, + 475, + 295, + 519 + ], + "type": "text", + "content": "[9] Soon Yau Cheong, Duygu Ceylan, Armin Mustafa, Andrew Gilbert, and Chun-Hao Paul Huang. Boosting camera motion control for video diffusion transformers. arXiv preprint arXiv:2410.10802, 2024." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 521, + 295, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 521, + 295, + 553 + ], + "spans": [ + { + "bbox": [ + 57, + 521, + 295, + 553 + ], + "type": "text", + "content": "[10] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In CVPR, pages 4690-4699, 2019." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 555, + 295, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 555, + 295, + 609 + ], + "spans": [ + { + "bbox": [ + 57, + 555, + 295, + 609 + ], + "type": "text", + "content": "[11] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 611, + 295, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 611, + 295, + 655 + ], + "spans": [ + { + "bbox": [ + 57, + 611, + 295, + 655 + ], + "type": "text", + "content": "[12] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. ICLR, 2024." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 657, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 657, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 657, + 295, + 689 + ], + "type": "text", + "content": "[13] Zinan Guo, Yanze Wu, Zhuowei Chen, Lang Chen, and Qian He. Pulid: Pure and lightning id customization via contrastive alignment. arXiv preprint arXiv:2404.16022, 2024." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "text", + "content": "[14] Junjie He, Yifeng Geng, and Liefeng Bo. Unipotrait: A unified framework for identity-preserving single-" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 74, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "type": "text", + "content": "and multi-human image personalization. arXiv preprint arXiv:2408.05939, 2024." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 553, + 140 + ], + "type": "text", + "content": "[15] Xuanhua He, Quande Liu, Shengju Qian, Xin Wang, Tao Hu, Ke Cao, Keyu Yan, Man Zhou, and Jie Zhang. Id-Animator: Zero-shot identity-preserving human video generation. arXiv preprint arXiv:2404.15275, 2024." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 141, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 141, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 317, + 141, + 553, + 195 + ], + "type": "text", + "content": "[16] Zecheng He, Bo Sun, Felix Juefei-Xu, Haoyu Ma, Ankit Ramchandani, Vincent Cheung, Siddharth Shah, Anmol Kalia, Harihar Subramanyam, Alireza Zareian, et al. Imagine yourself: Tuning-free personalized image generation. arXiv preprint arXiv:2409.13346, 2024." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 198, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 198, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 317, + 198, + 553, + 239 + ], + "type": "text", + "content": "[17] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 242, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 242, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 317, + 242, + 553, + 285 + ], + "type": "text", + "content": "[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS, 2017." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 286, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 308 + ], + "type": "text", + "content": "[19] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 310, + 553, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 310, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 310, + 553, + 331 + ], + "type": "text", + "content": "[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 33:6840-6851, 2020." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 333, + 553, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 333, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 333, + 553, + 376 + ], + "type": "text", + "content": "[21] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 378, + 553, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 378, + 553, + 433 + ], + "spans": [ + { + "bbox": [ + 316, + 378, + 553, + 433 + ], + "type": "text", + "content": "[22] Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Q Weinberger. Deep networks with stochastic depth. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 646-661. Springer, 2016." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 434, + 553, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 553, + 478 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 553, + 478 + ], + "type": "text", + "content": "[23] Yuge Huang, Yuhan Wang, Ying Tai, Xiaoming Liu, Pengcheng Shen, Shaoxin Li, Jilin Li, and Feiyue Huang. Curricularface: adaptive curriculum learning loss for deep face recognition. In CVPR, pages 5901-5910, 2020." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 479, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 479, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 317, + 479, + 553, + 544 + ], + "type": "text", + "content": "[24] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 545, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 545, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 317, + 545, + 553, + 600 + ], + "type": "text", + "content": "[25] Dan Kondratyuk, Lijun Yu, Xiuye Gu, José Lezama, Jonathan Huang, Rachel Hornung, Hartwig Adam, Hassan Akbari, Yair Alon, Vighnesh Birodkar, et al. Videopoet: A large language model for zero-shot video generation. arXiv preprint arXiv:2312.14125, 2023." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 601, + 553, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 601, + 553, + 643 + ], + "spans": [ + { + "bbox": [ + 317, + 601, + 553, + 643 + ], + "type": "text", + "content": "[26] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In CVPR, pages 1931-1941, 2023." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 646, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 646, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 646, + 553, + 689 + ], + "type": "text", + "content": "[27] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, pages 12888-12900. PMLR, 2022." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "text", + "content": "[28] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "12986" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "frozen image encoders and large language models. In ICML, pages 19730-19742. PMLR, 2023." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "text", + "content": "[29] Zhen Li, Mingdeng Cao, Xintao Wang, Zhongang Qi, Ming-Ming Cheng, and Ying Shan. Photomaker: Customizing realistic human photos via stacked id embedding. In CVPR, pages 8640-8650, 2024." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 185 + ], + "type": "text", + "content": "[30] Bin Lin, Yunyang Ge, Xinhua Cheng, Zongjian Li, Bin Zhu, Shaodong Wang, Xianyi He, Yang Ye, Shenghai Yuan, Liuhan Chen, et al. Open-sora plan: Open-source large video generation model. arXiv preprint arXiv:2412.00131, 2024." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 295, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 295, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 295, + 239 + ], + "type": "text", + "content": "[31] Yaofang Liu, Xiaodong Cun, Xuebo Liu, Xintao Wang, Yong Zhang, Haoxin Chen, Yang Liu, Tieyong Zeng, Raymond Chan, and Ying Shan. Evalcrafter: Benchmarking and evaluating large video generation models. arXiv preprint arXiv:2310.11440, 2023." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 241, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 294, + 285 + ], + "type": "text", + "content": "[32] Yuanxin Liu, Lei Li, Shuhuai Ren, Rundong Gao, Shicheng Li, Sishuo Chen, Xu Sun, and Lu Hou. Fetv: A benchmark for fine-grained evaluation of open-domain text-to-video generation. NeurIPS, 36, 2024." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 286, + 294, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 294, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 294, + 331 + ], + "type": "text", + "content": "[33] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 332, + 294, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 294, + 376 + ], + "type": "text", + "content": "[34] Yue Ma, Xiaodong Cun, Yingqing He, Chenyang Qi, Xintao Wang, Ying Shan, Xiu Li, and Qifeng Chen. Magic-stick: Controllable video editing via control handle transformations. arXiv preprint arXiv:2312.03047, 2023." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 377, + 294, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 377, + 294, + 432 + ], + "spans": [ + { + "bbox": [ + 56, + 377, + 294, + 432 + ], + "type": "text", + "content": "[35] Yue Ma, Yingqing He, Xiaodong Cun, Xintao Wang, Siran Chen, Xiu Li, and Qifeng Chen. Follow your pose: Pose-guided text-to-video generation using pose-free videos. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4117-4125, 2024." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 434, + 294, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 294, + 487 + ], + "type": "text", + "content": "[36] Yue Ma, Yingqing He, Hongfa Wang, Andong Wang, Chenyang Qi, Chengfei Cai, Xiu Li, Zhifeng Li, Heung-Yeung Shum, Wei Liu, et al. Follow-your-click: Open-domain regional image animation via short prompts. arXiv preprint arXiv:2403.08268, 2024." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 489, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 489, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 489, + 294, + 544 + ], + "type": "text", + "content": "[37] Yue Ma, Hongyu Liu, Hongfa Wang, Heng Pan, Yingqing He, Junkun Yuan, Ailing Zeng, Chengfei Cai, Heung-Yeung Shum, Wei Liu, et al. Follow-your-emoji: Fine-controllable and expressive freestyle portrait animation. In SIGGRAPH Asia 2024 Conference Papers, pages 1–12, 2024." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 545, + 294, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 294, + 589 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 294, + 589 + ], + "type": "text", + "content": "[38] Ze Ma, Daquan Zhou, Chun-Hsiao Yeh, Xue-She Wang, Xiuyu Li, Huanrui Yang, Zhen Dong, Kurt Keutzer, and Jiashi Feng. Magic-me: Identity-specific video customized diffusion. arXiv preprint arXiv:2402.09368, 2024." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 590, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 644 + ], + "type": "text", + "content": "[39] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453, 2023." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 646, + 294, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 646, + 294, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 646, + 294, + 679 + ], + "type": "text", + "content": "[40] Arsha Nagrani, Joon Son Chung, Weidi Xie, and Andrew Zisserman. Voxceleb: Large-scale speaker verification in the wild. Computer Speech & Language, 60:101027, 2020." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "text", + "content": "[41] Mayu Otani, Riku Togashi, Yu Sawai, Ryosuke Ishigami, Yuta Nakashima, Esa Rahtu, Janne Heikkila, and Shin'ichi Satoh. Toward verifiable and reproducible human evaluation" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "for text-to-image generation. In CVPR, pages 14277-14286, 2023." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "type": "text", + "content": "[42] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, et al. Movie gen: A cast of media foundation models. arXiv preprint arXiv:2410.13720, 2024." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "type": "text", + "content": "[43] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "type": "text", + "content": "[44] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, pages 8821-8831, 2021." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 253, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 553, + 308 + ], + "type": "text", + "content": "[45] Chen Rao, Guangyuan Li, Zehua Lan, Jiakai Sun, Junsheng Luan, Wei Xing, Lei Zhao, Huaizhong Lin, Jianfeng Dong, and Dalong Zhang. Rethinking video deblurring with wavelet-aware dynamic transformer and diffusion model. arXiv preprint arXiv:2408.13459, 2024." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 308, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 553, + 341 + ], + "type": "text", + "content": "[46] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bjorn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 342, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 553, + 387 + ], + "type": "text", + "content": "[47] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, pages 22500-22510, 2023." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 388, + 553, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 441 + ], + "type": "text", + "content": "[48] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdream booth: Hypernetworks for fast personalization of text-to-image models. In CVPR, pages 6527-6536, 2024." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 443, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 443, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 443, + 553, + 487 + ], + "type": "text", + "content": "[49] Yujun Shi, Jun Hao Liew, Hanshu Yan, Vincent YF Tan, and Jiashi Feng. Instadrag: Lightning fast and accurate drag-based image editing emerging from videos. arXiv preprint arXiv:2405.13722, 2024." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 488, + 553, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 488, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 316, + 488, + 553, + 520 + ], + "type": "text", + "content": "[50] Chenyang Si, Ziqi Huang, Yuming Jiang, and Ziwei Liu. Freeu: Free lunch in diffusion u-net. In CVPR, pages 4733-4743, 2024." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 522, + 553, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 522, + 553, + 566 + ], + "spans": [ + { + "bbox": [ + 316, + 522, + 553, + 566 + ], + "type": "text", + "content": "[51] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, pages 2256-2265. PMLR, 2015." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 567, + 553, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 567, + 553, + 589 + ], + "spans": [ + { + "bbox": [ + 316, + 567, + 553, + 589 + ], + "type": "text", + "content": "[52] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv, 2020." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 590, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 553, + 633 + ], + "type": "text", + "content": "[53] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, pages 2818-2826, 2016." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "text", + "content": "[54] Zhenyu Tang, Junwu Zhang, Xinhua Cheng, Wangbo Yu, Chaoran Feng, Yatian Pang, Bin Lin, and Li Yuan. Cycle3d: High-quality and consistent image-to-3d generation via generation-reconstruction cycle. arXiv preprint arXiv:2407.19548, 2024." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[55] Yuchuan Tian, Zhijun Tu, Hanting Chen, Jie Hu, Chao Xu, and Yunhe Wang. U-dits: Downsample tokens in u-shaped" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12987" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "diffusion transformers. arXiv preprint arXiv:2405.02730, 2024." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 129 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 129 + ], + "type": "text", + "content": "[56] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, and Anthony Chen. Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519, 2024." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 130, + 294, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 130, + 294, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 130, + 294, + 163 + ], + "type": "text", + "content": "[57] Ting-Chun Wang, Arun Mallya, and Ming-Yu Liu. One-shot free-view neural talking-head synthesis for video conferencing. In CVPR, pages 10039-10049, 2021." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 164, + 295, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 164, + 295, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 295, + 207 + ], + "type": "text", + "content": "[58] Zhao Wang, Aoxue Li, Enze Xie, Lingting Zhu, Yong Guo, Qi Dou, and Zhenguo Li. Customvideo: Customizing text-to-video generation with multiple subjects. arXiv preprint arXiv:2401.09962, 2024." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 209, + 295, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 295, + 253 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 295, + 253 + ], + "type": "text", + "content": "[59] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. In ICCV, pages 15943-15953, 2023." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 254, + 295, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 295, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 295, + 308 + ], + "type": "text", + "content": "[60] Yujie Wei, Shiwei Zhang, Zhiwu Qing, Hangjie Yuan, Zhiheng Liu, Yu Liu, Yingya Zhang, Jingren Zhou, and Hongming Shan. Dreamvideo: Composing your dream videos with customized subject and motion. In CVPR, pages 6537-6549, 2024." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 310, + 295, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 310, + 295, + 354 + ], + "spans": [ + { + "bbox": [ + 56, + 310, + 295, + 354 + ], + "type": "text", + "content": "[61] Jianzong Wu, Xiangtai Li, Yanhong Zeng, Jiangning Zhang, Qianyu Zhou, Yining Li, Yunhai Tong, and Kai Chen. Motionbooth: Motion-aware customized text-to-video generation. arXiv preprint arXiv:2406.17758, 2024." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 355, + 295, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 355, + 295, + 399 + ], + "spans": [ + { + "bbox": [ + 56, + 355, + 295, + 399 + ], + "type": "text", + "content": "[62] Jing Xiong, Gongye Liu, Lun Huang, Chengyue Wu, Taiqiang Wu, Yao Mu, Yuan Yao, Hui Shen, Zhongwei Wan, Jinfa Huang, et al. Autoregressive models in vision: A survey. arXiv preprint arXiv:2411.05902, 2024." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 399, + 295, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 295, + 453 + ], + "type": "text", + "content": "[63] Jiaqi Xu, Xinyi Zou, Kunzhe Huang, Yunkuo Chen, Bo Liu, MengLi Cheng, Xing Shi, and Jun Huang. Easyanimate: A high-performance long video generation method based on transformer architecture. arXiv preprint arXiv:2405.18991, 2024." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 456, + 295, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 295, + 510 + ], + "type": "text", + "content": "[64] Le Xue, Manli Shu, Anas Awadalla, Jun Wang, An Yan, Senthil Purushwalkam, Honglu Zhou, Viraj Prabhu, Yutong Dai, Michael S Ryoo, et al. xgen-mm (blip-3): A family of open large multimodal models. arXiv preprint arXiv:2408.08872, 2024." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 512, + 295, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 295, + 555 + ], + "type": "text", + "content": "[65] Shuo Yang, Kun-Peng Ning, Yu-Yang Liu, Jia-Yu Yao, Yong-Hong Tian, Yi-Bing Song, and Li Yuan. Is parameter collision hindering continual learning in llms? arXiv preprint arXiv:2410.10179, 2024." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 557, + 295, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 295, + 610 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 295, + 610 + ], + "type": "text", + "content": "[66] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 612, + 295, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 612, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 612, + 295, + 656 + ], + "type": "text", + "content": "[67] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 658, + 295, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 295, + 690 + ], + "type": "text", + "content": "[68] Jianhui Yu, Hao Zhu, Liming Jiang, Chen Change Loy, Weidong Cai, and Wayne Wu. Celebv-text: A large-scale facial text-video dataset. In CVPR, pages 14805-14814, 2023." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 692, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 295, + 713 + ], + "type": "text", + "content": "[69] Wangbo Yu, Chaoran Feng, Jiye Tang, Xu Jia, Li Yuan, and Yonghong Tian. Evagaussians: Event stream assisted" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 487 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "gaussian splatting from blurry images. arXiv preprint arXiv:2405.20224, 2024." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "text", + "content": "[70] Wangbo Yu, Jinbo Xing, Li Yuan, Wenbo Hu, Xiaoyu Li, Zhipeng Huang, Xiangjun Gao, Tien-Tsin Wong, Ying Shan, and Yonghong Tian. Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis. arXiv preprint arXiv:2409.02048, 2024." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 152, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 185 + ], + "type": "text", + "content": "[71] Yongsheng Yu, Ziyun Zeng, Hang Hua, Jianlong Fu, and Jiebo Luo. Promptfix: You prompt and we fix the photo. arXiv preprint arXiv:2405.16785, 2024." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 186, + 553, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 230 + ], + "type": "text", + "content": "[72] Shenghai Yuan, Jinfa Huang, Yujun Shi, Yongqi Xu, Ruijie Zhu, Bin Lin, Xinhua Cheng, Li Yuan, and Jiebo Luo. Mag-ictime: Time-lapse video generation models as metamorphic simulators. arXiv preprint arXiv:2404.05014, 2024." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 231, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 231, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 231, + 553, + 285 + ], + "type": "text", + "content": "[73] Shenghai Yuan, Jina Huang, Yongqi Xu, Yaoyang Liu, Shaofeng Zhang, Yujun Shi, Rui-Jie Zhu, Xinhua Cheng, Jiebo Luo, and Li Yuan. Chronomagic-bench: A benchmark for metamorphic evaluation of text-to-time-lapse video generation. NeurIPS, 37:21236-21270, 2025." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 286, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 319 + ], + "type": "text", + "content": "[74] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, pages 3836-3847, 2023." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 320, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 320, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 320, + 553, + 363 + ], + "type": "text", + "content": "[75] Zhenghao Zhang, Junchao Liao, Menghao Li, Long Qin, and Weizhi Wang. Tora: Trajectory-oriented diffusion transformer for video generation. arXiv preprint arXiv:2407.21705, 2024." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 365, + 553, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 420 + ], + "type": "text", + "content": "[76] Mingzhe Zheng, Yongqi Xu, Haojian Huang, Xuran Ma, Yexin Liu, Wenjie Shu, Yatian Pang, Feilong Tang, Qifeng Chen, Harry Yang, et al. Videogen-of-thought: A collaborative framework for multi-shot video generation. arXiv preprint arXiv:2412.02259, 2024." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 421, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 421, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 421, + 553, + 453 + ], + "type": "text", + "content": "[77] Zangwei Zheng, Xiangyu Peng, and Yang You. Open-sora: Democratizing efficient video production for all. In *Github*, 2024." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 454, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 454, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 454, + 553, + 487 + ], + "type": "text", + "content": "[78] Yuan Zhou, Qiuyue Wang, Yuxuan Cai, and Huan Yang. Allegro: Open the black box of commercial-level video generation model. arXiv preprint arXiv:2410.15458, 2024." + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12988" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Identity-preserving Distillation Sampling by Fixed-Point Iterator/8dc89aa9-e649-490d-ba0b-e87a936deb98_content_list.json b/2025/Identity-preserving Distillation Sampling by Fixed-Point Iterator/8dc89aa9-e649-490d-ba0b-e87a936deb98_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..bd183e0f72c2c678ca0a182f6739b055f190505b --- /dev/null +++ b/2025/Identity-preserving Distillation Sampling by Fixed-Point Iterator/8dc89aa9-e649-490d-ba0b-e87a936deb98_content_list.json @@ -0,0 +1,1497 @@ +[ + { + "type": "text", + "text": "Identity-preserving Distillation Sampling by Fixed-Point Iterator", + "text_level": 1, + "bbox": [ + 171, + 130, + 826, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "SeonHwa Kim $^{1}$ Jiwon Kim $^{1}$ Soobin Park $^{2}$ Donghoon Ahn $^{1}$ Jiwon Kang $^{1}$ Seungryong Kim $^{3*}$ Kyong Hwan Jin $^{1*}$ Eunju Cha $^{2*}$ $^{1}$ Korea University $^{2}$ Sookmyung Women's University $^{3}$ KAIST", + "bbox": [ + 169, + 178, + 825, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{sunkim0062, jwonkim, dhahn99, jiwon7258, kyong_jin}@korea.ac.kr, {psb1219j, eunju.cha}@sookmyung.ac.kr, seungryong.kim@kaist.ac.kr", + "bbox": [ + 210, + 234, + 782, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 304, + 326, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Score distillation sampling (SDS) demonstrates a powerful capability for text-conditioned 2D image and 3D object generation by distilling the knowledge from learned score functions. However, SDS often suffers from blurriness caused by noisy gradients. When SDS meets the image editing, such degradations can be reduced by adjusting bias shifts using reference pairs, but the de-biasing techniques are still corrupted by erroneous gradients. To this end, we introduce Identity-preserving Distillation Sampling (IDS), which compensates for the gradient leading to undesired changes in the results. Based on the analysis that these errors come from the text-conditioned scores, a new regularization technique, called fixed-point iterative regularization (FPR), is proposed to modify the score itself, driving the preservation of the identity even including poses and structures. Thanks to a self-correction by FPR, the proposed method provides clear and unambiguous representations corresponding to the given prompts in image-to-image editing and editable neural radiance field (NeRF). The structural consistency between the source and the edited data is obviously maintained compared to other state-of-the-art methods. Our code is https://github.com/shhh0620/IDS", + "bbox": [ + 88, + 337, + 485, + 684 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 717, + 220, + 732 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion models [4, 8, 9, 22, 25] have shown powerful representations on text-to-image (T2I) generative tasks. With the advance of classifier guidance (CG) and classifier-free guidance (CFG) paradigms [1, 4, 8, 10], diffusion models improve the quality of generated samples [9, 25]. Such high-quality image generators can be easily extended to image editing by simply modifying forward/reverse iterations [14], applying CFG with a target prompt [2, 7] or interchanging attention layers [26].", + "bbox": [ + 89, + 734, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6a18606f6793dfc1980d978ecb2ab4577f6e06385b0fef0d0185ce334224edb2.jpg", + "image_caption": [ + "Figure 1. Trace of guided updating from source to target images using delta denoising score (DDS) and identity-preserving distillation sampling (IDS). DDS moves a gradient of score function toward $\\mathcal{M}_{\\mathbf{z}}$ manifold directed by stochastic direction $\\epsilon$ . In contrast, IDS moves a gradient with a corrected direction by a fixed-point regularization." + ], + "image_footnote": [], + "bbox": [ + 514, + 301, + 906, + 580 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, Delta Denoising Score (DDS) [6] is proposed to edit a source image by distilling the rich generative prior of T2I diffusion models. It is based on the analysis of Score Distillation Sampling (SDS) [20], originally developed to optimize a parametric generator such as Neural Radiance Fields (NeRFs) [16] by exploiting the learned score of the diffusion models. Even though SDS offers remarkable performance in synthesizing 3D scenes, noisy gradients from stochastic perturbations lead to significantly over-saturated results that faithfully follow the given text prompts. In the context of image editing, text prompts do not often include information about the identity of the source image, such as the background, the object's pose, or the structure of the", + "bbox": [ + 511, + 703, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 887, + 236, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "11115", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/738d635f30a3df04304cc174651668ddc18fbbd91b32a508accacc2c253472d9.jpg", + "image_caption": [ + "Figure 2. Flowchart of IDS. The backbone of our algorithm employs DDS [6] framework to distill score function into a target image. Our fixed-point regularization (FPR) obtains a guided noise, $\\epsilon^{*}$ , from iterative updates using posterior mean computed by Tweedie's formula. When distilling the score function to a target image, the guided noise is updated while maintaining the identity of the source." + ], + "image_footnote": [], + "bbox": [ + 155, + 88, + 844, + 327 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "content, which should be retained during updates. Thus, DDS is designed to resolve such blurriness by erasing gradients of non-text-aligned features from SDS gradients. There is no explicit procedure to preserve the source's identity in DDS updates because the fine gradient may provide the conserved identity. However, this cannot be guaranteed if many variations in the structure are possible, such as editing the image of a cat into a pig, as shown in Fig. 1. To alleviate this problem, Contrastive Denoising Score (CDS) [17] and Posterior Distillation Sampling (PDS) [12] are introduced to maximize the mutual information of the source image and edited image. Although such algorithms rely heavily on text prompts, the algorithms have yet to analyze the inherent error caused by text-conditioned scores.", + "bbox": [ + 88, + 409, + 480, + 619 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we investigate the underlying meaning of text-conditioned score. The gradient maps the stochastic latent, generated by applying the forward diffusion process to the given image, to one of the possible images described by the prompt, including the original image. Simply, the score obtained from the latent of the source image ('source latent') and the source prompt can be a gradient to another image represented by the identical text. Based on this interpretation, the accumulation of misaligned directions causes the loss of the source's identity, leading to structural changes in the result with DDS, as shown in Fig. 1.", + "bbox": [ + 88, + 625, + 482, + 790 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this issue, we propose a novel score distillation sampling to effectively preserve the identity of the original image by self-correcting the misaligned gradients, called Identity-preserving Distillation Sampling (IDS). The key insight is that if the score is precisely adjusted to the source image, the conditional expectation of the source image given the source latent contains meaningful information", + "bbox": [ + 88, + 795, + 482, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "that should be preserved during the editing. This conditional expectation corresponds to the posterior mean computed by Tweedie's formula using the learned score [3, 5]. The source latent is iteratively updated to make the posterior mean similar to the source image. This procedure, named a fixed-point iterative regularization (FPR), results in the aligned score with the source that provides reliable gradients for editing, as illustrated in Fig. 1. Following IDS update is performed using guided noise extracted from the refined source latent, rather than random Gaussian noise. This further ensures the identity preservation. Our method demonstrated superior results compared with baselines in two tasks: editing images by prompts and editing NeRF.", + "bbox": [ + 511, + 409, + 906, + 604 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our main contributions are as follows:", + "bbox": [ + 531, + 607, + 870, + 619 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We obtain reliable gradients for the score distillation function by a fixed-point iterator with respect to posterior means. The iterator corrects the text-conditioned score, guiding SDS gradients toward reliable pre-trained manifolds.", + "- Our fixed-point regularization preserves the identities of sources such as structures and poses in edited targets for 2D and 3D editing. Such preservation is well demonstrated in qualitative and quantitative results." + ], + "bbox": [ + 511, + 623, + 906, + 758 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related works", + "text_level": 1, + "bbox": [ + 511, + 773, + 660, + 789 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Image Editing with Diffusion Models", + "text_level": 1, + "bbox": [ + 511, + 800, + 833, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With the great success of image generation using diffusion models, the pre-trained diffusion models have been recently employed for image editing tasks, demonstrating significant advancements in the quality and flexibility of generated edits [2, 7, 12, 14, 17, 26]. Stochastic Differential", + "bbox": [ + 511, + 825, + 905, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "11116", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Editing (SDEdit) [14] is a pioneering work in which the source image was modified by adding noise and solving reverse stochastic differential equations. Thanks to the text-conditional Latent Diffusion Model (LDM), a.k.a. Stable Diffusion [22], text-driven editing approaches have been introduced. Specifically, the text embedding was injected through the cross-attention layer of the model for image editing and translation, while retaining the structure of the original image [7, 26]. The editing was further controlled by rescaling the attention of the specific word [7] or by manipulating the self-attention features [26]. These approaches provide greater control by balancing fidelity between the edited prompt and the source image without the need for model training, fine-tuning, additional data, or optimization. However, the current DDIM-based inversion [24] can lead to unsatisfactory reconstructions for real images, and the cross-attention bottleneck limits its effectiveness for broader edits. Crafting suitable prompts also remains challenging for complex compositions.", + "bbox": [ + 89, + 90, + 483, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Score distillation sampling", + "text_level": 1, + "bbox": [ + 89, + 386, + 331, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Score Distillation Sampling (SDS) [20] enables text-driven 3D synthesis by leveraging probability density distillation loss to distill knowledge from 2D diffusion models, allowing high-quality 3D scene generation based on textual prompts without 3D training data. However, SDS has limitations, often producing oversaturated and overly smooth 3D models, and lacking diversity across initializations. To address these limitations of SDS, various models have been proposed based on exploiting multi-step denoising [30], a variation approach [27], negative conditioning [11], and ordinary differential equation trajectory [28]. To mitigate the limitation of noisy gradients in SDS, which hampers precise image editing, DDS [6] was introduced. By computing the delta between the derived gradient and the target pair, DDS effectively isolates and removes unwanted noise in the gradient direction. Despite these advancements, DDS still faces challenges in preserving the complete structural consistency of the source image's identity.", + "bbox": [ + 89, + 409, + 483, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 695, + 230, + 709 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Diffusion Model and Sampling Guidance", + "text_level": 1, + "bbox": [ + 89, + 719, + 441, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text-to-image diffusion models $\\epsilon_{\\phi}(\\cdot)$ are based on diffusion probabilistic models (DPMs) [9, 22, 25]. The models are trained to estimate the denoising score when the original image $\\mathbf{z}_0$ and the text condition $y$ are given:", + "bbox": [ + 89, + 742, + 483, + 804 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (\\phi) = \\mathbb {E} _ {t, \\epsilon} [ \\| \\epsilon_ {\\phi} (\\mathbf {z} _ {t}, y, t) - \\epsilon \\| _ {2} ^ {2} ],\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 811, + 397, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\epsilon \\sim \\mathcal{N}(0,\\mathbf{I})$ and $t\\sim \\mathcal{U}(0,1)$ $\\mathbf{z}_t$ refers to the stochastic latent of $\\mathbf{z}_0$ via the forward diffusion process as follows:", + "bbox": [ + 89, + 839, + 483, + 869 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} _ {t} = \\sqrt {\\alpha_ {t}} \\mathbf {z} _ {0} + \\sqrt {1 - \\alpha_ {t}} \\epsilon , \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 883, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\alpha_{t}$ is noise schedule. With the trained $\\epsilon_{\\phi}(\\cdot)$ , high-quality samples can be generated using the classifier-free guidance (CFG) [8] by subtracting unconditioned denoising score from the conditioned score with guidance scale $\\omega$ :", + "bbox": [ + 511, + 90, + 905, + 151 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\epsilon_ {\\phi} ^ {\\omega} \\left(\\mathbf {z} _ {t}, y, t\\right) = (1 + \\omega) \\epsilon_ {\\phi} \\left(\\mathbf {z} _ {t}, y, t\\right) - \\omega \\epsilon_ {\\phi} \\left(\\mathbf {z} _ {t}, \\varnothing , t\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 165, + 906, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Score Distillation Sampling (SDS)", + "text_level": 1, + "bbox": [ + 511, + 194, + 808, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With pretrained text-to-image diffusion models $\\epsilon_{\\phi}(\\cdot)$ , SDS [20] synthesizes 3D data $\\mathbf{z}$ for a given text prompt $y$ by optimizing the differentiable rendering function paretrized by $\\theta$ , where $\\mathbf{z} = g(\\theta)$ :", + "bbox": [ + 511, + 217, + 906, + 273 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\theta} \\mathcal {L} _ {\\mathrm {S D S}} (\\mathbf {z}, y) = \\mathbb {E} _ {t, \\epsilon} \\left[ \\omega (t) \\left(\\epsilon_ {\\phi} ^ {\\omega} \\left(\\mathbf {z} _ {t}, y, t\\right) - \\epsilon\\right) \\frac {\\partial \\mathbf {z}}{\\partial \\theta} \\right]. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 284, + 906, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The optimized parameters $\\theta^{*}$ provide the text-conditioned 3D volume that follows the diffusion prior [20]. However, a single text prompt $y$ can refer to many different 3D volumes, each with diverse backgrounds or structural details of the object. Therefore, an inherent limitation of SDS [20] is that the score conditioned by the prompt $y$ does not always provide the diffusion prior to the identical object during the optimization process, leading to blurry and unclear results.", + "bbox": [ + 511, + 325, + 906, + 446 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Delta Denoising Score (DDS)", + "text_level": 1, + "bbox": [ + 511, + 457, + 769, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "DDS [6] is proposed to synthesize the image $\\mathbf{z}^{\\mathrm{trg}}$ from the given source image $\\mathbf{z}^{\\mathrm{src}}$ and its corresponding prompt $y^{\\mathrm{src}}$ , which is aligned to the target prompt $y^{\\mathrm{trg}}$ . Based on the insight that the gradient should be zero if $y^{\\mathrm{trg}}$ matches $y^{\\mathrm{src}}$ , DDS minimizes the identity change of $\\mathbf{z}^{\\mathrm{src}}$ by simple replacing $\\epsilon$ in (3) with the score $\\epsilon_{\\phi}^{\\omega}(\\mathbf{z}_{t}^{\\mathrm{src}}, y^{\\mathrm{src}}, t)$ as follows:", + "bbox": [ + 511, + 479, + 905, + 566 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\theta} \\mathcal {L} _ {\\mathrm {D D S}} = \\mathbb {E} _ {t, \\epsilon} \\left[ \\left(\\epsilon_ {\\phi} ^ {\\omega} \\left(\\mathbf {z} _ {t} ^ {\\operatorname {t r g}}, y ^ {\\operatorname {t r g}}, t\\right) - \\epsilon_ {\\phi} ^ {\\omega} \\left(\\mathbf {z} _ {t} ^ {\\operatorname {s r c}}, y ^ {\\operatorname {s r c}}, t\\right)\\right) \\frac {\\partial \\mathbf {z} ^ {\\operatorname {t r g}}}{\\partial \\theta} \\right]. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 587, + 906, + 618 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For simplicity, we denote $\\epsilon_{\\phi}^{\\mathrm{trg}} = \\epsilon_{\\phi}^{\\omega}(\\mathbf{z}^{\\mathrm{trg}},y^{\\mathrm{trg}},t)$ and $\\epsilon_{\\phi}^{\\mathrm{src}} = \\epsilon_{\\phi}^{\\omega}(\\mathbf{z}^{\\mathrm{src}},y^{\\mathrm{src}},t)$ . Here, $\\epsilon_{\\phi}^{\\mathrm{trg}}$ and $\\epsilon_{\\phi}^{\\mathrm{src}}$ can be interpreted as the gradients representing the direction from $\\mathbf{z}_t^{\\mathrm{trg}}$ to $\\mathbf{z}^{\\mathrm{trg}}$ and the direction from $\\mathbf{z}_t^{\\mathrm{src}}$ to $\\mathbf{z}^{\\mathrm{src}}$ , respectively. $\\theta = \\mathbf{z}^{\\mathrm{trg}}$ is thus gradually optimized along the direction from $\\mathbf{z}^{\\mathrm{src}}$ to $\\mathbf{z}^{\\mathrm{trg}}$ , as shown in Fig. 1. It is worth noting that the guidance of the update can be calculated at the same point $\\mathbf{z}_t^{\\mathrm{trg}}$ , thanks to the shared $\\epsilon$ . However, the slight error in the gradient caused by the score $\\epsilon_{\\phi}^{\\mathrm{src}}$ still leads to the incorrect direction for the optimization.", + "bbox": [ + 511, + 628, + 906, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.4. Fixed-point Iteration", + "text_level": 1, + "bbox": [ + 511, + 799, + 710, + 814 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In numerical analysis, a fixed-point iteration [18] is an iterative method to find fixed points of a function $f$ , where $f(x) = x$ . Given an initial point $x_0$ , the iteration is defined as:", + "bbox": [ + 511, + 823, + 905, + 881 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nx _ {n + 1} = f (x _ {n}), \\quad n = 0, 1, 2, \\dots\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 885, + 820, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "11117", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Under appropriate conditions, this sequence converges to a fixed point. Thanks to its applicability to non-linear problems with low computational costs, fixed-point iteration is widely used in optimization, including applications in the context of diffusion models [13].", + "bbox": [ + 89, + 90, + 482, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 89, + 180, + 181, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the source pair $\\{\\mathbf{z}^{\\mathrm{src}}, y^{\\mathrm{src}}\\}$ , the aim of our work is to provide an edited result $\\mathbf{z}^{\\mathrm{trg}}$ that is aligned with $y^{\\mathrm{trg}}$ while maintaining the source's identity. To this end, we introduce a novel approach called Identity-preserving Distillation Sampling (IDS), which (1) corrects the error of the gradient aligned with the text prompt by the fixed-point iterator and (2) provides the result $\\mathbf{z}^{\\mathrm{trg}}$ using the guided noise.", + "bbox": [ + 89, + 205, + 483, + 313 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Motivation", + "text_level": 1, + "bbox": [ + 89, + 321, + 210, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Analysis of the text-conditioned score. We first investigated how much identity of the given image $\\mathbf{z}^{\\mathrm{src}}$ could be contained in the text-conditioned score $\\epsilon_{\\phi}^{\\mathrm{src}}$ . To do this, we conducted the experiment to compare the original image $\\mathbf{z}^{\\mathrm{src}}$ and the posterior mean $\\mathbf{z}_{0|t}^{\\mathrm{src}} = \\mathbb{E}[\\mathbf{z}^{\\mathrm{src}}|\\mathbf{z}_t^{\\mathrm{src}}]$ , which is given by:", + "bbox": [ + 89, + 344, + 483, + 430 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} _ {0 | t} ^ {\\mathrm {s r c}} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\mathbf {z} _ {t} ^ {\\mathrm {s r c}} - \\sqrt {1 - \\alpha_ {t}} \\epsilon_ {\\phi} ^ {\\mathrm {s r c}}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 431, + 482, + 459 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{z}_t^{\\mathrm{src}}$ denotes the source latent generated by (1). As shown in the first row of the supplementary Fig. S1, it is difficult to recognize the features of $\\mathbf{z}^{\\mathrm{src}}$ in $\\mathbf{z}_{0|t}^{\\mathrm{src}}$ , such as hairstyle, details of eyes, and background. This demonstrates that the score $\\epsilon_{\\phi}^{\\mathrm{src}}$ is not exactly adjusted to the given image $\\mathbf{z}^{\\mathrm{src}}$ . This deformation becomes more pronounced with increasing $t$ . The experiment confirms that $\\epsilon_{\\phi}^{\\mathrm{src}}$ may not be a precise guidance to the source image $\\mathbf{z}^{\\mathrm{src}}$ . Therefore, the text-conditioned score $\\epsilon_{\\phi}^{\\mathrm{src}}$ needs to be modified to maintain the identity of the source image $\\mathbf{z}^{\\mathrm{src}}$ in the edited result $\\mathbf{z}^{\\mathrm{trg}}$ .", + "bbox": [ + 89, + 462, + 482, + 626 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Accumulated error in DDS. The transformed image $\\mathbf{z}^{\\mathrm{trg}}$ can be converted back to the original image $\\mathbf{z}^{\\mathrm{src}}$ by reversing the set of $\\epsilon$ used to synthesize $\\mathbf{z}^{\\mathrm{trg}}$ from $\\mathbf{z}^{\\mathrm{src}}$ and swapping $\\{\\mathbf{z}^{\\mathrm{src}}, y^{\\mathrm{src}}\\}$ and $\\{\\mathbf{z}^{\\mathrm{trg}}, y^{\\mathrm{trg}}\\}$ to calculate the DDS loss in (4). If the guidance from $\\mathbf{z}^{\\mathrm{src}}$ to $\\mathbf{z}^{\\mathrm{trg}}$ is computed exactly, the perfect reconstruction can be achieved. Nevertheless, as can be seen from the second row in Fig. 3, DDS [6] fails to restore the original image $\\mathbf{z}^{\\mathrm{src}}$ from the edited image $\\mathbf{z}^{\\mathrm{trg}}$ , which implies that the direction from $\\mathbf{z}^{\\mathrm{src}}$ to $\\mathbf{z}^{\\mathrm{trg}}$ is calculated incorrectly. Based on our analysis, this error is because the text-conditioned score $\\epsilon_{\\phi}^{\\mathrm{src}}$ do not refer to the source $\\mathbf{z}^{\\mathrm{src}}$ , which can be explicitly expressed as the difference between the injected noise $\\epsilon$ and the score $\\epsilon_{\\phi}^{\\mathrm{src}}$ . While the optimization is being processed, the error inevitably accumulates, leading to the undesirable change to the structure and the pose. To address these issues, we investigated whether the guidance from $\\mathbf{z}^{\\mathrm{src}}$ to $\\mathbf{z}^{\\mathrm{trg}}$ can be properly provided while preserving the source's identity, when the timestep $t$ is con", + "bbox": [ + 89, + 628, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3aa8a2e8e325d12c6f49923f8ac4a7c927f92d57dab571a6b15aa14a2c0240e0.jpg", + "image_caption": [ + "Figure 3. Accumulated error in DDS. $\\mathbf{z}^{\\mathrm{trg}}$ is edited image of source image $\\mathbf{z}^{\\mathrm{src}}$ by prompt $y^{\\mathrm{src}}\\rightarrow y^{\\mathrm{trg}}$ . $\\mathbf{z}^{\\mathrm{src\\dagger}}$ is the inverted image of $\\mathbf{z}^{\\mathrm{trg}}$ by prompt $y^{\\mathrm{trg}}\\rightarrow y^{\\mathrm{src}}$ . (First row) Inversion result of DDS with timestep $t\\sim \\mathcal{U}(0,0.2)$ . (Second row) Inversion result of DDS with $t\\sim \\mathcal{U}(0,1)$ . (Third row) Inversion result of ours." + ], + "image_footnote": [], + "bbox": [ + 509, + 82, + 893, + 396 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "strained by $t \\sim \\mathcal{U}(0,0.2)$ . This is because the posterior mean $\\mathbf{z}_{0|t}^{\\mathrm{src}}$ and the source image $\\mathbf{z}^{\\mathrm{src}}$ are similar for small timestep $t$ , as illustrated in the first row of supplementary Fig. S1. However, as depicted in the first row of Fig. 3, DDS yields unrealistic result with this setting, whereby the structure of the given image $\\mathbf{z}^{\\mathrm{src}}$ is overemphasized. This implies that it is not sufficient to simply limit the timestep $t$ to prevent the score from deviating too far from $\\mathbf{z}^{\\mathrm{src}}$ to correct the misalignment of the score to $\\mathbf{z}^{\\mathrm{src}}$ . Hence, we propose a fundamental approach to refine the gradient to achieve identity consistency without unwanted overemphasis on details.", + "bbox": [ + 511, + 484, + 906, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Identity-preserving Distillation Sampling (IDS)", + "text_level": 1, + "bbox": [ + 511, + 674, + 903, + 691 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fixed-point Regularization (FPR). Here, we introduce a Fixed-point Regularization (FPR) method that adjusts the text-conditioned score $\\epsilon_{\\phi}^{\\mathrm{src}}$ to the source image $\\mathbf{z}^{\\mathrm{src}}$ . Our key premise is that if the score $\\epsilon_{\\phi}^{\\mathrm{src}}$ is rightly estimated as a gradient to $\\mathbf{z}^{\\mathrm{src}}$ , the posterior mean $\\mathbf{z}_{0|t}^{\\mathrm{src}}$ also contains sufficient information about $\\mathbf{z}^{\\mathrm{src}}$ . Therefore, FPR loss is designed to minimize the difference between $\\mathbf{z}^{\\mathrm{src}}$ and $\\mathbf{z}_{0|t}^{\\mathrm{src}}$ as follows:", + "bbox": [ + 511, + 698, + 905, + 806 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {F P R}} = d \\left(\\mathbf {z} ^ {\\mathrm {s r c}}, \\mathbf {z} _ {0 | t} ^ {\\mathrm {s c}}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 640, + 814, + 903, + 832 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $d(\\mathbf{x}_1,\\mathbf{x}_2)$ can be any metric to compare $\\mathbf{x}_1$ and $\\mathbf{x}_2$ . Here, we employed the Euclidean loss, and further investigations using various metrics are provided in Supplementary Materials.", + "bbox": [ + 511, + 840, + 903, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "11118", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The score $\\epsilon_{\\phi}^{\\mathrm{src}}$ needs to be modified to minimize the FPR loss before obtaining the updated direction. There are two ways to control the score $\\epsilon_{\\phi}^{\\mathrm{src}}$ by altering the injection noise $\\epsilon$ or the source latent $\\mathbf{z}_t^{\\mathrm{src}}$ . As illustrated in supplementary Fig. S1, the proposed FPR revises the score $\\epsilon_{\\phi}^{\\mathrm{src}}$ to serve the source's identity for both approaches. Note that the score incorporates the content details, with the updates being performed with respect to the source latent $\\mathbf{z}_t^{\\mathrm{src}}$ compared to the noise $\\epsilon$ . Thus, $\\mathbf{z}_t^{\\mathrm{src}}$ is updated to minimize the FPR loss as follows:", + "bbox": [ + 89, + 90, + 483, + 239 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} _ {t} ^ {\\mathrm {s r c}} \\leftarrow \\mathbf {z} _ {t} ^ {\\mathrm {s r c}} - \\lambda \\nabla_ {\\mathbf {z} _ {t} ^ {\\mathrm {s r c}}} \\mathcal {L} _ {\\mathrm {F P R}}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 199, + 241, + 482, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda$ and $N$ denote a regularization scale and the number of iterations, respectively.", + "bbox": [ + 89, + 262, + 483, + 294 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Fixed-point Regularization (FPR)" + ], + "code_body": "Require: $\\mathbf{z}^{\\mathrm{src}},y^{\\mathrm{src}},\\epsilon_{\\phi},\\omega ,\\lambda ,N$ \n1: $\\epsilon \\sim \\mathcal{N}(0,\\mathbf{I})$ \n2: $t\\sim \\mathcal{U}(0,1)$ \n3: $\\mathbf{z}_t^{\\mathrm{src}}\\gets \\sqrt{\\alpha_t}\\mathbf{z}^{\\mathrm{src}} + \\sqrt{1 - \\alpha_t}\\epsilon$ \n4: for $\\mathrm{i} = 1,\\dots ,\\mathrm{N}$ do \n5: $\\epsilon_{\\phi}^{\\mathrm{src}}\\gets (1 + \\omega)\\epsilon_{\\phi}(\\mathbf{z}_{t}^{\\mathrm{src}},y^{\\mathrm{src}},t) - \\omega \\epsilon_{\\phi}(\\mathbf{z}_{t}^{\\mathrm{src}},\\emptyset ,t)$ \n6: $\\mathbf{z}_{0|t}^{\\mathrm{src}}\\gets \\frac{1}{\\sqrt{\\alpha_t}} (\\mathbf{z}_t^{\\mathrm{src}} - \\sqrt{1 - \\alpha_t}\\epsilon_\\phi^{\\mathrm{src}})$ \n7: $\\mathcal{L}_{\\mathrm{FPR}}\\gets d(\\mathbf{z}_{0|t}^{\\mathrm{src}},\\mathbf{z}^{\\mathrm{src}})$ \n8: $\\mathbf{z}_t^{\\mathrm{src}}\\gets \\mathbf{z}_t^{\\mathrm{src}} - \\lambda \\nabla_{\\mathbf{z}_t^{\\mathrm{src}}}\\mathcal{L}_{\\mathrm{FPR}}$ \n9: end for \n10: $\\epsilon^{*}\\gets \\frac{1}{\\sqrt{1 - \\alpha_{t}}} (\\mathbf{z}_{t}^{\\mathrm{src}} - \\sqrt{\\alpha_{t}}\\mathbf{z}^{\\mathrm{src}})$ \n11: return $\\epsilon^{*}$", + "bbox": [ + 91, + 324, + 473, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Editing with guided noise. Thanks to the proposed FPR, the optimized source latent $\\mathbf{z}_t^{\\mathrm{sfc*}}$ containing the source's identity can be obtained. Then, the guided noise $\\epsilon^{*}$ is extracted as follows:", + "bbox": [ + 89, + 523, + 483, + 579 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\epsilon^ {*} = \\frac {1}{\\sqrt {1 - \\alpha_ {t}}} \\left(\\mathbf {z} _ {t} ^ {\\mathrm {s r c} *} - \\sqrt {\\alpha_ {t}} \\mathbf {z} ^ {\\mathrm {s r c}}\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 583, + 482, + 612 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\epsilon^{*}$ is utilized to produce the stochastic latent $\\mathbf{z}_t^{\\mathrm{trg}*}$ by applying the forward diffusion process to the target image $\\mathbf{z}^{\\mathrm{trg}}$ . With $\\mathbf{z}_t^{\\mathrm{src}*}$ and $\\mathbf{z}_t^{\\mathrm{trg}*}$ , the updated direction is given by:", + "bbox": [ + 89, + 619, + 483, + 665 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\theta} \\mathcal {L} _ {\\mathrm {I D S}} = \\mathbb {E} _ {t, \\epsilon} \\left[ \\left(\\epsilon_ {\\phi} ^ {\\omega} \\left(\\mathbf {z} _ {t} ^ {\\operatorname {t r g} *}, y ^ {\\operatorname {t r g}}, t\\right) - \\epsilon_ {\\phi} ^ {\\omega} \\left(\\mathbf {z} _ {t} ^ {\\operatorname {s r c} *}, y ^ {\\operatorname {s r c}}, t\\right)\\right) \\frac {\\partial \\mathbf {z} ^ {\\operatorname {t r g}}}{\\partial \\theta} \\right]. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 669, + 482, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "It is worth noting that $\\epsilon^{*}$ guides the appropriate gradients for editing while conserving the source's identity. In contrast to DDS, the proposed IDS perfectly reconstructs the source from the edited result $\\mathbf{z}^{\\mathrm{trg}}$ , as shown in the third row of Fig. 3. This confirms that the correct score and the corresponding injection noise can preserve the identity without further consideration of mutual information. The flowchart of our IDS is illustrated in Fig. 2.", + "bbox": [ + 89, + 712, + 483, + 833 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Results", + "text_level": 1, + "bbox": [ + 89, + 844, + 176, + 859 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate our method through editing experiments conducted on two experiments. In Sec. 5.1, we perform a com", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "parison on image-to-image editing across several datasets. In Sec. 5.2, we extend our evaluation to editable Neural Radiance Fields (NeRF) [16].", + "bbox": [ + 511, + 90, + 905, + 137 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Text-guided image editing", + "text_level": 1, + "bbox": [ + 511, + 145, + 750, + 162 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. To evaluate our method, we conduct comparative experiments against four state-of-the-art image editing models: Prompt-to-Prompt (P2P) [7], Plug-and-Play (PNP) [26], DDS [6], and CDS [17]. The implementations of the baselines are carried out by referencing the official source code for each method. More details are provided in Supplementary Materials.", + "bbox": [ + 511, + 167, + 906, + 273 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Qualitative Results. We present the qualitative results comparing our method with the baselines in Fig. 4. Prompt-to-Prompt (P2P) [7] performs image editing after applying DDIM inversion [4, 24] to the source image, leading to disregarding the structural components of the source image and following the target prompt excessively. Plug-and-Play (PnP) [26] has limitations in object recognition, as seen in the fourth row of Fig. 4. The third row of Fig. 4 demonstrates that DDS [6] and CDS [17] exhibited limitations, particularly in preserving the structural characteristics of the source image. In contrast, our method successfully edits the image while preserving the structural integrity of the source image.", + "bbox": [ + 511, + 273, + 906, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Quantitative Results. To measure the identity-preserving performance, we utilize two datasets. First, we collect 250 cat images from the LAION 5B dataset [23] based on [17] for Cat-to-Others task and measure Intersection over Union (IoU). Second, we gather 28 images from the InstructPix2Pix (IP2P) dataset [2], which contains the pairs of source and target images and corresponding prompts and calculate the background Peak-Signal-to-Noise-Ratio (PSNR). Details of the metrics are provided in Supplementary Materials. In addition, we use the LPIPS score [29] for each experiment to quantify the similarity between source and target images. The results are presented in Tab. 1. Our method consistently achieves the lowest LPIPS score across all datasets, indicating that it best preserves the structural semantics of the source images.", + "bbox": [ + 511, + 470, + 908, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For user evaluation, we present 35 comparison sets for four baselines and our method, gathering responses from 47 participants. Participants are asked to choose the most appropriate image for the following three questions: 1. Which", + "bbox": [ + 511, + 696, + 908, + 758 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/15fe5b8572880ac3feb279b2ebe9fcaced4158f082a514ea1208d5e23f324fdb.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
cat2pigcat2squirrelIp2p
MetricIoU (↑)LPIPS (↓)IoU (↑)LPIPS (↓)PSNR (↑)LPIPS (↓)
P2P [7]0.580.420.520.4620.880.47
PnP [26]0.550.520.530.5223.810.39
DDS [6]0.690.280.650.3026.020.24
CDS [17]0.720.250.710.2627.350.21
IDS (Ours)0.740.220.710.2429.250.19
", + "bbox": [ + 521, + 777, + 895, + 849 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Quantitative results for image editing. LPIPS [29] and IoU was measured on LAION 5B [23], while LPIPS and background PSNR was measured on InstructPix2Pix [2].", + "bbox": [ + 511, + 854, + 906, + 897 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "11119", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8829e7eac60bf87688e6d33af64d90f86b791ee186d114ffb93e1f244f670a6a.jpg", + "image_caption": [ + "Figure 4. Qualitative results of InstructPix2Pix dataset [2]. Our method successfully edits the image aligning with the target text prompt while preserving the structural integrity of the source image." + ], + "image_footnote": [], + "bbox": [ + 99, + 90, + 901, + 685 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "image best fits the text condition? 2. Which image best preserves the structural information of the original image?", + "bbox": [ + 89, + 744, + 482, + 776 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ac70e3f3afce906e5b70b9ba46bc0bb8b7f09584c200d8ca2e1e6aaa7acb1eec.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
User Preference Rate (%)GPT score [19]
MetricText (↑)Preserving (↑)Quality (↑)Text (↑)Preserving (↑)Quality (↑)
P2P [7]11.134.808.095.665.375.77
PnP [26]7.727.176.936.546.776.74
DDS [6]20.3010.8216.237.607.517.37
CDS [17]17.0216.7217.088.268.008.09
IDS (Ours)43.8360.4951.678.979.008.80
", + "bbox": [ + 93, + 795, + 480, + 862 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. User study and GPT scores [19] show that our method achieved the highest scores across all questions for image editing.", + "bbox": [ + 89, + 868, + 482, + 896 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3. Which image has the best quality for text-based image editing? Additionally, we measure the GPT score using the Dreambench++ [19] method, which generates human-aligned assessments for the same questions by refining the scoring into ten distinct levels. As shown in Tab. 2, our method receives the highest ratings for all questions.", + "bbox": [ + 511, + 744, + 906, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Editing NeRF", + "text_level": 1, + "bbox": [ + 511, + 845, + 658, + 862 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct experiments involving 3D rendering of edited images to demonstrate the effectiveness of our method in", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "11120", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/19031eeb194960ee28999139602c6543b87400b6b7058ba872122eecd641d5ae.jpg", + "image_caption": [ + "Figure 5. Qualitative results on Synthetic $360^{\\circ}$ and LLFF datasets. IDS outperforms the baselines by preserving the structural consistency of the source image and maintaining the integrity of regions that should remain unchanged, while precisely editing only the areas specified by the target prompt. Furthermore, comparisons of the depth map results also highlight the superior consistency of our method over other baseline models." + ], + "image_footnote": [], + "bbox": [ + 99, + 88, + 883, + 492 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "maintaining structural consistency. This approach is particularly relevant as consistency has an even greater impact on outcomes in 3D environments.", + "bbox": [ + 88, + 579, + 482, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Datasets. We evaluated our method on widely used NeRF datasets: Synthetic NeRF [16] and LLFF [15]. Since NeRF datasets have no given pairs of source and target prompts, we manually composed image descriptions.", + "bbox": [ + 88, + 626, + 482, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Results. Fig. 5 illustrates the qualitative results of our method compared with NeRF editing baselines. In the first row, the target prompt specifies a precise part of the image for fine-grained editing. DDS [6] and CDS [17] fail to differentiate and edit the specific area. At the same time, our method accurately identifies the region indicated by the target prompt in the image and performs detailed editing exclusively on that part. The second row demonstrates a scenario in which the target prompt is designed to edit the mood of the image. Our approach adjusts the colors associated with \"autumn\" and \"leaves\" throughout the image while maintaining consistency in the \"trunk\" whereas DDS and CDS also changed the \"trunk\". In terms of depth maps, our method generates clean depth maps with minimal noise", + "bbox": [ + 88, + 689, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "after image editing, whereas DDS and CDS introduce noticeable noise into the depth maps.", + "bbox": [ + 511, + 579, + 906, + 609 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e1574aa11de15ee47c700b943628f0276879f9982ddd1262805c8effe443e77d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MetricCLIP [21] (↑)User Preference Rate (%)
Text (↑)Preserving (↑)Quality (↑)
DDS [6]0.159636.8828.3732.62
CDS [17]0.159722.7023.4021.28
IDS (Ours)0.162640.4248.2346.10
", + "bbox": [ + 527, + 619, + 893, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Quantitative results of NeRF editing with respect to CLIP score and User Preference Rate. IDS demonstrates superior quantitative performance compared to the baselines.", + "bbox": [ + 511, + 691, + 906, + 736 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Results. Based on edited images, we performed 3D rendering and subsequently conducted quantitative evaluations provided in Tab. 3. To assess whether the edited 3D images are precisely aligned with the target prompts, we measured the CLIP [21] scores at $200\\mathrm{k}$ iterations of training on the LLFF dataset. We additionally present a user evaluation conducted under the same setup in Sec. 5.1. Consistent with the trends observed in the qualitative results, our method demonstrates superior performance in the quantitative evaluations compared to other baselines.", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "11121", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7b5891dbf5e397e1ca07b53b8994a2ac0a5308f781b9eead31346704728e805b.jpg", + "image_caption": [ + "\"A photo of shark\" $\\rightarrow$ \"A photo of dolphin\"" + ], + "image_footnote": [], + "bbox": [ + 101, + 89, + 470, + 157 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fe19d7b84027058dec2d3b1a9a25b31d8ffdc7676f00da67c3d8b1e8ee24a9f0.jpg", + "image_caption": [ + "Figure 6. Ablation study on scale $\\lambda$ . To show the effect of scale more extremely, the number of iterations of FPR is set as 1. The result of DDS is the same as $\\lambda = 0$ since it means no update for source identity.", + "\"The green leaves\" $\\rightarrow$ \"Yellow and red leaves in autumn\"", + "Figure 7. Optimization steps. Our method effectively preserves the consistency of the source image, even as the number of iteration steps increases up to 400." + ], + "image_footnote": [], + "bbox": [ + 91, + 234, + 472, + 388 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/36f56cc4f718574e8dbf5a489f143d797734d8c8a408e6d526194c1131974d5c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
optim stepFPR iterLPIPS(↓)CLIP(↑)time (sec/img)Memory (GB)
DDS200-0.2400.29322.456.27
CDS200-0.2100.28759.318.83
IDS20010.1990.28550.808.63
30.1900.277107.77
10030.1650.26554.04
1500.1800.27281.25
", + "bbox": [ + 98, + 459, + 472, + 527 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Discussions", + "text_level": 1, + "bbox": [ + 89, + 602, + 210, + 617 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.1. Ablation studies on FPR", + "text_level": 1, + "bbox": [ + 89, + 627, + 313, + 642 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "FPR iteration $N$ . We conduct experiments on FPR iterations $N$ to evaluate its impact and determine the optimal iteration count. Although performing just one iteration of FPR is sufficient to preserve the source identity, as shown in lower LPIPS score than baselines of Tab. 4, we set $N = 3$ to emphasize the purpose of our method.", + "bbox": [ + 89, + 651, + 482, + 742 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Scale $\\lambda$ . The scaling factor $\\lambda$ of FPR determines how much information of source latent $\\mathbf{z}^{\\mathrm{src}}$ is kept. As shown in Fig. 6, increasing the scale preserves the attributes of the source image, resulting in more successful editing when it is hard to translate due to the structural mismatch between the source and the target prompt.", + "bbox": [ + 89, + 743, + 482, + 834 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.2. Optimization steps", + "text_level": 1, + "bbox": [ + 89, + 845, + 272, + 863 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To show that our method can prevent error accumulation during translation, we set the experiment to extend the num", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ber of optimization steps from 200 to 400. In the results of DDS and CDS, there is color boosting or loss of details due to the cumulated error. In contrast, IDS maintains the characteristics of the original images, such as the color of the pumpkin in the first row of Fig. 7 and the shape of the leaf in the second row of Fig. 7, better than other methods.", + "bbox": [ + 511, + 90, + 905, + 181 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Limitation", + "text_level": 1, + "bbox": [ + 511, + 195, + 627, + 210 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The proposed IDS demonstrates outstanding performance across evaluation metrics assessing consistency between source and target images. However, during FPR process, IDS relies solely on information from the source $\\{\\mathbf{z}^{\\mathrm{src}}, y^{\\mathrm{src}}\\}$ without incorporating target-side information. This results in comparatively lower CLIP scores [21] than other baselines (Tab. 5) and failure cases for more complex translations (Fig. 8). In addition, our method requires additional computational overhead (Tab. 4) since FPR is applied to each optimization iteration. Detailed discussion about our limitation is provided in Sec. G of supplementary. Our future direction will explore changing the score conditioned by the target prompt $y^{\\mathrm{trg}}$ , leading to a better alignment with $y^{\\mathrm{trg}}$ .", + "bbox": [ + 511, + 220, + 906, + 431 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d22728cccfb101b796396b7df7b6351448c5eb01178554b145caf0f250f436b0.jpg", + "table_caption": [ + "Table 4. Computational complexity on 28 images of InstructPix2Pix [2] for various settings. Lower LPIPS and higher CLIP scores mean better quality." + ], + "table_footnote": [], + "table_body": "
P2P [7]PnP [26]DDS [6]CDS [17]IDS (Ours)
cat2lion0.290.210.300.290.29
cat2dog0.270.260.270.270.26
Ip2p0.280.300.290.290.28
", + "bbox": [ + 527, + 443, + 888, + 493 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Limitation of IDS with respect to CLIP score[21] for image editing on LAION 5B [23] and InstructPix2Pix [2].", + "bbox": [ + 511, + 497, + 905, + 526 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3414560104d3a705bce3ba19d768d8f5bbee176d715cb4fab571847e93062713.jpg", + "image_caption": [ + "\"Photo free night, house, aurora\" → \"... with two dogs\"", + "Figure 8. Failure case for complex text prompt." + ], + "image_footnote": [], + "bbox": [ + 519, + 532, + 898, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "8. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 656, + 633, + 671 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We proposed a new distillation sampling method using a fixed-point regularization which aligns the text-conditioned score towards identity-preserved manifolds. The proposed fixed-point regularization preserves the source's identity by re-projecting the intermediate score status on posterior means. In this manner, corrected noises guide a gradient of distilled score toward identity-consistent manifolds. Owing to self-correction by a fixed-point iterator and guided injection noise, the proposed identity-preserving distillation sampling provides clear and unambiguous representations corresponding to the given prompts in text-guided image editing and editable neural radiance field (NeRF). Furthermore, our model can be utilized as a universal module in addition to the existing score-sampling processes.", + "bbox": [ + 511, + 681, + 906, + 893 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "11122", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement This work was supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (RS-2024-00335741, RS-2024-00357197).", + "bbox": [ + 89, + 90, + 485, + 151 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 165, + 187, + 181 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Donghoon Ahn, Hyoungwon Cho, Jaewon Min, Wooseok Jang, Jungwoo Kim, SeonHwa Kim, Hyun Hee Park, Kyong Hwan Jin, and Seungryong Kim. Self-rectifying diffusion sampling with perturbed-attention guidance. arXiv preprint arXiv:2403.17377, 2024. 1", + "[2] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 1, 2, 5, 6, 8", + "[3] Hyungjin Chung, Jeongsol Kim, Michael Thompson McCann, Marc Louis Klasky, and Jong Chul Ye. Diffusion posterior sampling for general noisy inverse problems. In The Eleventh International Conference on Learning Representations, 2023. 2", + "[4] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 1, 5", + "[5] Bradley Efron. Tweedie's formula and selection bias. Journal of the American Statistical Association, 106(496):1602-1614, 2011. 2", + "[6] Amir Hertz, Kfir Aberman, and Daniel Cohen-Or. Delta denoising score. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2328-2337, 2023. 1, 2, 3, 4, 5, 6, 7, 8", + "[7] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-or. Prompt-to-prompt image editing with cross-attention control. In The Eleventh International Conference on Learning Representations, 2023. 1, 2, 3, 5, 6, 8", + "[8] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 1, 3", + "[9] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 3", + "[10] Susung Hong, Gyuseong Lee, Wooseok Jang, and Seungryong Kim. Improving sample quality of diffusion models using self-attention guidance. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7462–7471, 2023. 1", + "[11] Oren Katzir, Or Patashnik, Daniel Cohen-Or, and Dani Lischinski. Noise-free score distillation. arXiv preprint arXiv:2310.17590, 2023. 3", + "[12] Juil Koo, Chanho Park, and Minhyuk Sung. Posterior distillation sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13352-13361, 2024. 2", + "[13] Barak Meiri, Dvir Samuel, Nir Darshan, Gal Chechik, Shai Avidan, and Rami Ben-Ari. Fixed-point inversion for text-" + ], + "bbox": [ + 91, + 189, + 483, + 902 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "to-image diffusion models. arXiv preprint arXiv:2312.12540, 2023. 4", + "[14] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations, 2022. 1, 2, 3", + "[15] Ben Mildenhall, Pratul P Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (ToG), 38(4):1-14, 2019. 7", + "[16] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 5, 7", + "[17] Hyelin Nam, Gihyun Kwon, Geon Yeong Park, and Jong Chul Ye. Contrastive denoising score for text-guided latent diffusion image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9192-9201, 2024. 2, 5, 6, 7, 8", + "[18] Neal Parikh, Stephen Boyd, et al. Proximal algorithms. Foundations and trends® in Optimization, 1(3):127-239, 2014. 3", + "[19] Yuang Peng, Yuxin Cui, Haomiao Tang, Zekun Qi, Runpei Dong, Jing Bai, Chunrui Han, Zheng Ge, Xiangyu Zhang, and Shu-Tao Xia. Dreambench++: A human-aligned benchmark for personalized image generation. arXiv preprint arXiv:2406.16855, 2024. 6", + "[20] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 1, 3", + "[21] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 7, 8", + "[22] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 3", + "[23] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 5, 8", + "[24] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3, 5", + "[25] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 1, 3" + ], + "bbox": [ + 514, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "11123", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023. 1, 2, 3, 5, 6, 8", + "[27] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[28] Zike Wu, Pan Zhou, Xuanyu Yi, Xiaoding Yuan, and Hanwang Zhang. Consistent3d: Towards consistent high-fidelity text-to-3d generation with deterministic sampling prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9892–9902, 2024. 3", + "[29] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5", + "[30] Zhizhuo Zhou and Shubham Tulsiani. Sparsefusion: Distilling view-conditioned diffusion for 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12588-12597, 2023. 3" + ], + "bbox": [ + 89, + 90, + 482, + 429 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "11124", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Identity-preserving Distillation Sampling by Fixed-Point Iterator/8dc89aa9-e649-490d-ba0b-e87a936deb98_model.json b/2025/Identity-preserving Distillation Sampling by Fixed-Point Iterator/8dc89aa9-e649-490d-ba0b-e87a936deb98_model.json new file mode 100644 index 0000000000000000000000000000000000000000..24307400521a24c65193c4badd4e7d550cd3fda2 --- /dev/null +++ b/2025/Identity-preserving Distillation Sampling by Fixed-Point Iterator/8dc89aa9-e649-490d-ba0b-e87a936deb98_model.json @@ -0,0 +1,1892 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.172, + 0.131, + 0.828, + 0.154 + ], + "angle": 0, + "content": "Identity-preserving Distillation Sampling by Fixed-Point Iterator" + }, + { + "type": "text", + "bbox": [ + 0.171, + 0.179, + 0.826, + 0.234 + ], + "angle": 0, + "content": "SeonHwa Kim\\(^{1}\\) Jiwon Kim\\(^{1}\\) Soobin Park\\(^{2}\\) Donghoon Ahn\\(^{1}\\) Jiwon Kang\\(^{1}\\) Seungryong Kim\\(^{3*}\\) Kyong Hwan Jin\\(^{1*}\\) Eunju Cha\\(^{2*}\\) \n\\(^{1}\\)Korea University \\(^{2}\\)Sookmyung Women's University \\(^{3}\\)KAIST" + }, + { + "type": "text", + "bbox": [ + 0.211, + 0.236, + 0.783, + 0.27 + ], + "angle": 0, + "content": "{sunkim0062, jwonkim, dhahn99, jiwon7258, kyong_jin}@korea.ac.kr, {psb1219j, eunju.cha}@sookmyung.ac.kr, seungryong.kim@kaist.ac.kr" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.305, + 0.327, + 0.32 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.338, + 0.486, + 0.685 + ], + "angle": 0, + "content": "Score distillation sampling (SDS) demonstrates a powerful capability for text-conditioned 2D image and 3D object generation by distilling the knowledge from learned score functions. However, SDS often suffers from blurriness caused by noisy gradients. When SDS meets the image editing, such degradations can be reduced by adjusting bias shifts using reference pairs, but the de-biasing techniques are still corrupted by erroneous gradients. To this end, we introduce Identity-preserving Distillation Sampling (IDS), which compensates for the gradient leading to undesired changes in the results. Based on the analysis that these errors come from the text-conditioned scores, a new regularization technique, called fixed-point iterative regularization (FPR), is proposed to modify the score itself, driving the preservation of the identity even including poses and structures. Thanks to a self-correction by FPR, the proposed method provides clear and unambiguous representations corresponding to the given prompts in image-to-image editing and editable neural radiance field (NeRF). The structural consistency between the source and the edited data is obviously maintained compared to other state-of-the-art methods. Our code is https://github.com/shhh0620/IDS" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.718, + 0.222, + 0.733 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.736, + 0.485, + 0.875 + ], + "angle": 0, + "content": "Diffusion models [4, 8, 9, 22, 25] have shown powerful representations on text-to-image (T2I) generative tasks. With the advance of classifier guidance (CG) and classifier-free guidance (CFG) paradigms [1, 4, 8, 10], diffusion models improve the quality of generated samples [9, 25]. Such high-quality image generators can be easily extended to image editing by simply modifying forward/reverse iterations [14], applying CFG with a target prompt [2, 7] or interchanging attention layers [26]." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.302, + 0.907, + 0.582 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.591, + 0.909, + 0.677 + ], + "angle": 0, + "content": "Figure 1. Trace of guided updating from source to target images using delta denoising score (DDS) and identity-preserving distillation sampling (IDS). DDS moves a gradient of score function toward \\(\\mathcal{M}_{\\mathbf{z}}\\) manifold directed by stochastic direction \\(\\epsilon\\). In contrast, IDS moves a gradient with a corrected direction by a fixed-point regularization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Recently, Delta Denoising Score (DDS) [6] is proposed to edit a source image by distilling the rich generative prior of T2I diffusion models. It is based on the analysis of Score Distillation Sampling (SDS) [20], originally developed to optimize a parametric generator such as Neural Radiance Fields (NeRFs) [16] by exploiting the learned score of the diffusion models. Even though SDS offers remarkable performance in synthesizing 3D scenes, noisy gradients from stochastic perturbations lead to significantly over-saturated results that faithfully follow the given text prompts. In the context of image editing, text prompts do not often include information about the identity of the source image, such as the background, the object's pose, or the structure of the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.237, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11115" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.156, + 0.089, + 0.845, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.34, + 0.908, + 0.385 + ], + "angle": 0, + "content": "Figure 2. Flowchart of IDS. The backbone of our algorithm employs DDS [6] framework to distill score function into a target image. Our fixed-point regularization (FPR) obtains a guided noise, \\(\\epsilon^{*}\\), from iterative updates using posterior mean computed by Tweedie's formula. When distilling the score function to a target image, the guided noise is updated while maintaining the identity of the source." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.41, + 0.482, + 0.621 + ], + "angle": 0, + "content": "content, which should be retained during updates. Thus, DDS is designed to resolve such blurriness by erasing gradients of non-text-aligned features from SDS gradients. There is no explicit procedure to preserve the source's identity in DDS updates because the fine gradient may provide the conserved identity. However, this cannot be guaranteed if many variations in the structure are possible, such as editing the image of a cat into a pig, as shown in Fig. 1. To alleviate this problem, Contrastive Denoising Score (CDS) [17] and Posterior Distillation Sampling (PDS) [12] are introduced to maximize the mutual information of the source image and edited image. Although such algorithms rely heavily on text prompts, the algorithms have yet to analyze the inherent error caused by text-conditioned scores." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.625, + 0.483, + 0.791 + ], + "angle": 0, + "content": "To this end, we investigate the underlying meaning of text-conditioned score. The gradient maps the stochastic latent, generated by applying the forward diffusion process to the given image, to one of the possible images described by the prompt, including the original image. Simply, the score obtained from the latent of the source image ('source latent') and the source prompt can be a gradient to another image represented by the identical text. Based on this interpretation, the accumulation of misaligned directions causes the loss of the source's identity, leading to structural changes in the result with DDS, as shown in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.796, + 0.483, + 0.903 + ], + "angle": 0, + "content": "To address this issue, we propose a novel score distillation sampling to effectively preserve the identity of the original image by self-correcting the misaligned gradients, called Identity-preserving Distillation Sampling (IDS). The key insight is that if the score is precisely adjusted to the source image, the conditional expectation of the source image given the source latent contains meaningful information" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.41, + 0.907, + 0.606 + ], + "angle": 0, + "content": "that should be preserved during the editing. This conditional expectation corresponds to the posterior mean computed by Tweedie's formula using the learned score [3, 5]. The source latent is iteratively updated to make the posterior mean similar to the source image. This procedure, named a fixed-point iterative regularization (FPR), results in the aligned score with the source that provides reliable gradients for editing, as illustrated in Fig. 1. Following IDS update is performed using guided noise extracted from the refined source latent, rather than random Gaussian noise. This further ensures the identity preservation. Our method demonstrated superior results compared with baselines in two tasks: editing images by prompts and editing NeRF." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.608, + 0.871, + 0.621 + ], + "angle": 0, + "content": "In summary, our main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.624, + 0.905, + 0.697 + ], + "angle": 0, + "content": "- We obtain reliable gradients for the score distillation function by a fixed-point iterator with respect to posterior means. The iterator corrects the text-conditioned score, guiding SDS gradients toward reliable pre-trained manifolds." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.699, + 0.907, + 0.759 + ], + "angle": 0, + "content": "- Our fixed-point regularization preserves the identities of sources such as structures and poses in edited targets for 2D and 3D editing. Such preservation is well demonstrated in qualitative and quantitative results." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.624, + 0.907, + 0.759 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.775, + 0.661, + 0.79 + ], + "angle": 0, + "content": "2. Related works" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.801, + 0.834, + 0.817 + ], + "angle": 0, + "content": "2.1. Image Editing with Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.901 + ], + "angle": 0, + "content": "With the great success of image generation using diffusion models, the pre-trained diffusion models have been recently employed for image editing tasks, demonstrating significant advancements in the quality and flexibility of generated edits [2, 7, 12, 14, 17, 26]. Stochastic Differential" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11116" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.485, + 0.38 + ], + "angle": 0, + "content": "Editing (SDEdit) [14] is a pioneering work in which the source image was modified by adding noise and solving reverse stochastic differential equations. Thanks to the text-conditional Latent Diffusion Model (LDM), a.k.a. Stable Diffusion [22], text-driven editing approaches have been introduced. Specifically, the text embedding was injected through the cross-attention layer of the model for image editing and translation, while retaining the structure of the original image [7, 26]. The editing was further controlled by rescaling the attention of the specific word [7] or by manipulating the self-attention features [26]. These approaches provide greater control by balancing fidelity between the edited prompt and the source image without the need for model training, fine-tuning, additional data, or optimization. However, the current DDIM-based inversion [24] can lead to unsatisfactory reconstructions for real images, and the cross-attention bottleneck limits its effectiveness for broader edits. Crafting suitable prompts also remains challenging for complex compositions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.387, + 0.332, + 0.403 + ], + "angle": 0, + "content": "2.2. Score distillation sampling" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.41, + 0.484, + 0.683 + ], + "angle": 0, + "content": "Score Distillation Sampling (SDS) [20] enables text-driven 3D synthesis by leveraging probability density distillation loss to distill knowledge from 2D diffusion models, allowing high-quality 3D scene generation based on textual prompts without 3D training data. However, SDS has limitations, often producing oversaturated and overly smooth 3D models, and lacking diversity across initializations. To address these limitations of SDS, various models have been proposed based on exploiting multi-step denoising [30], a variation approach [27], negative conditioning [11], and ordinary differential equation trajectory [28]. To mitigate the limitation of noisy gradients in SDS, which hampers precise image editing, DDS [6] was introduced. By computing the delta between the derived gradient and the target pair, DDS effectively isolates and removes unwanted noise in the gradient direction. Despite these advancements, DDS still faces challenges in preserving the complete structural consistency of the source image's identity." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.696, + 0.231, + 0.71 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.72, + 0.442, + 0.736 + ], + "angle": 0, + "content": "3.1. Diffusion Model and Sampling Guidance" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.743, + 0.484, + 0.805 + ], + "angle": 0, + "content": "Text-to-image diffusion models \\(\\epsilon_{\\phi}(\\cdot)\\) are based on diffusion probabilistic models (DPMs) [9, 22, 25]. The models are trained to estimate the denoising score when the original image \\(\\mathbf{z}_0\\) and the text condition \\(y\\) are given:" + }, + { + "type": "equation", + "bbox": [ + 0.176, + 0.813, + 0.398, + 0.831 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (\\phi) = \\mathbb {E} _ {t, \\epsilon} [ \\| \\epsilon_ {\\phi} (\\mathbf {z} _ {t}, y, t) - \\epsilon \\| _ {2} ^ {2} ],\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.871 + ], + "angle": 0, + "content": "where \\(\\epsilon \\sim \\mathcal{N}(0,\\mathbf{I})\\) and \\(t\\sim \\mathcal{U}(0,1)\\) \\(\\mathbf{z}_t\\) refers to the stochastic latent of \\(\\mathbf{z}_0\\) via the forward diffusion process as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.198, + 0.885, + 0.484, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} _ {t} = \\sqrt {\\alpha_ {t}} \\mathbf {z} _ {0} + \\sqrt {1 - \\alpha_ {t}} \\epsilon , \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.152 + ], + "angle": 0, + "content": "where \\(\\alpha_{t}\\) is noise schedule. With the trained \\(\\epsilon_{\\phi}(\\cdot)\\), high-quality samples can be generated using the classifier-free guidance (CFG) [8] by subtracting unconditioned denoising score from the conditioned score with guidance scale \\(\\omega\\):" + }, + { + "type": "equation", + "bbox": [ + 0.534, + 0.166, + 0.907, + 0.183 + ], + "angle": 0, + "content": "\\[\n\\epsilon_ {\\phi} ^ {\\omega} \\left(\\mathbf {z} _ {t}, y, t\\right) = (1 + \\omega) \\epsilon_ {\\phi} \\left(\\mathbf {z} _ {t}, y, t\\right) - \\omega \\epsilon_ {\\phi} \\left(\\mathbf {z} _ {t}, \\varnothing , t\\right). \\tag {2}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.195, + 0.81, + 0.212 + ], + "angle": 0, + "content": "3.2. Score Distillation Sampling (SDS)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.218, + 0.907, + 0.275 + ], + "angle": 0, + "content": "With pretrained text-to-image diffusion models \\(\\epsilon_{\\phi}(\\cdot)\\), SDS [20] synthesizes 3D data \\(\\mathbf{z}\\) for a given text prompt \\(y\\) by optimizing the differentiable rendering function paretrized by \\(\\theta\\), where \\(\\mathbf{z} = g(\\theta)\\):" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.285, + 0.907, + 0.317 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\theta} \\mathcal {L} _ {\\mathrm {S D S}} (\\mathbf {z}, y) = \\mathbb {E} _ {t, \\epsilon} \\left[ \\omega (t) \\left(\\epsilon_ {\\phi} ^ {\\omega} \\left(\\mathbf {z} _ {t}, y, t\\right) - \\epsilon\\right) \\frac {\\partial \\mathbf {z}}{\\partial \\theta} \\right]. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.326, + 0.907, + 0.448 + ], + "angle": 0, + "content": "The optimized parameters \\(\\theta^{*}\\) provide the text-conditioned 3D volume that follows the diffusion prior [20]. However, a single text prompt \\(y\\) can refer to many different 3D volumes, each with diverse backgrounds or structural details of the object. Therefore, an inherent limitation of SDS [20] is that the score conditioned by the prompt \\(y\\) does not always provide the diffusion prior to the identical object during the optimization process, leading to blurry and unclear results." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.458, + 0.771, + 0.475 + ], + "angle": 0, + "content": "3.3. Delta Denoising Score (DDS)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.481, + 0.906, + 0.568 + ], + "angle": 0, + "content": "DDS [6] is proposed to synthesize the image \\(\\mathbf{z}^{\\mathrm{trg}}\\) from the given source image \\(\\mathbf{z}^{\\mathrm{src}}\\) and its corresponding prompt \\(y^{\\mathrm{src}}\\), which is aligned to the target prompt \\(y^{\\mathrm{trg}}\\). Based on the insight that the gradient should be zero if \\(y^{\\mathrm{trg}}\\) matches \\(y^{\\mathrm{src}}\\), DDS minimizes the identity change of \\(\\mathbf{z}^{\\mathrm{src}}\\) by simple replacing \\(\\epsilon\\) in (3) with the score \\(\\epsilon_{\\phi}^{\\omega}(\\mathbf{z}_{t}^{\\mathrm{src}}, y^{\\mathrm{src}}, t)\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.588, + 0.907, + 0.619 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\theta} \\mathcal {L} _ {\\mathrm {D D S}} = \\mathbb {E} _ {t, \\epsilon} \\left[ \\left(\\epsilon_ {\\phi} ^ {\\omega} \\left(\\mathbf {z} _ {t} ^ {\\operatorname {t r g}}, y ^ {\\operatorname {t r g}}, t\\right) - \\epsilon_ {\\phi} ^ {\\omega} \\left(\\mathbf {z} _ {t} ^ {\\operatorname {s r c}}, y ^ {\\operatorname {s r c}}, t\\right)\\right) \\frac {\\partial \\mathbf {z} ^ {\\operatorname {t r g}}}{\\partial \\theta} \\right]. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.63, + 0.907, + 0.789 + ], + "angle": 0, + "content": "For simplicity, we denote \\(\\epsilon_{\\phi}^{\\mathrm{trg}} = \\epsilon_{\\phi}^{\\omega}(\\mathbf{z}^{\\mathrm{trg}},y^{\\mathrm{trg}},t)\\) and \\(\\epsilon_{\\phi}^{\\mathrm{src}} = \\epsilon_{\\phi}^{\\omega}(\\mathbf{z}^{\\mathrm{src}},y^{\\mathrm{src}},t)\\). Here, \\(\\epsilon_{\\phi}^{\\mathrm{trg}}\\) and \\(\\epsilon_{\\phi}^{\\mathrm{src}}\\) can be interpreted as the gradients representing the direction from \\(\\mathbf{z}_t^{\\mathrm{trg}}\\) to \\(\\mathbf{z}^{\\mathrm{trg}}\\) and the direction from \\(\\mathbf{z}_t^{\\mathrm{src}}\\) to \\(\\mathbf{z}^{\\mathrm{src}}\\), respectively. \\(\\theta = \\mathbf{z}^{\\mathrm{trg}}\\) is thus gradually optimized along the direction from \\(\\mathbf{z}^{\\mathrm{src}}\\) to \\(\\mathbf{z}^{\\mathrm{trg}}\\), as shown in Fig. 1. It is worth noting that the guidance of the update can be calculated at the same point \\(\\mathbf{z}_t^{\\mathrm{trg}}\\), thanks to the shared \\(\\epsilon\\). However, the slight error in the gradient caused by the score \\(\\epsilon_{\\phi}^{\\mathrm{src}}\\) still leads to the incorrect direction for the optimization." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.8, + 0.712, + 0.815 + ], + "angle": 0, + "content": "3.4. Fixed-point Iteration" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.824, + 0.906, + 0.882 + ], + "angle": 0, + "content": "In numerical analysis, a fixed-point iteration [18] is an iterative method to find fixed points of a function \\( f \\), where \\( f(x) = x \\). Given an initial point \\( x_0 \\), the iteration is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.886, + 0.821, + 0.902 + ], + "angle": 0, + "content": "\\[\nx _ {n + 1} = f (x _ {n}), \\quad n = 0, 1, 2, \\dots\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11117" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.168 + ], + "angle": 0, + "content": "Under appropriate conditions, this sequence converges to a fixed point. Thanks to its applicability to non-linear problems with low computational costs, fixed-point iteration is widely used in optimization, including applications in the context of diffusion models [13]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.181, + 0.182, + 0.197 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.206, + 0.484, + 0.314 + ], + "angle": 0, + "content": "Given the source pair \\(\\{\\mathbf{z}^{\\mathrm{src}}, y^{\\mathrm{src}}\\}\\), the aim of our work is to provide an edited result \\(\\mathbf{z}^{\\mathrm{trg}}\\) that is aligned with \\(y^{\\mathrm{trg}}\\) while maintaining the source's identity. To this end, we introduce a novel approach called Identity-preserving Distillation Sampling (IDS), which (1) corrects the error of the gradient aligned with the text prompt by the fixed-point iterator and (2) provides the result \\(\\mathbf{z}^{\\mathrm{trg}}\\) using the guided noise." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.323, + 0.212, + 0.337 + ], + "angle": 0, + "content": "4.1. Motivation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.345, + 0.484, + 0.431 + ], + "angle": 0, + "content": "Analysis of the text-conditioned score. We first investigated how much identity of the given image \\(\\mathbf{z}^{\\mathrm{src}}\\) could be contained in the text-conditioned score \\(\\epsilon_{\\phi}^{\\mathrm{src}}\\). To do this, we conducted the experiment to compare the original image \\(\\mathbf{z}^{\\mathrm{src}}\\) and the posterior mean \\(\\mathbf{z}_{0|t}^{\\mathrm{src}} = \\mathbb{E}[\\mathbf{z}^{\\mathrm{src}}|\\mathbf{z}_t^{\\mathrm{src}}]\\), which is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.432, + 0.483, + 0.46 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} _ {0 | t} ^ {\\mathrm {s r c}} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\mathbf {z} _ {t} ^ {\\mathrm {s r c}} - \\sqrt {1 - \\alpha_ {t}} \\epsilon_ {\\phi} ^ {\\mathrm {s r c}}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.463, + 0.483, + 0.627 + ], + "angle": 0, + "content": "where \\(\\mathbf{z}_t^{\\mathrm{src}}\\) denotes the source latent generated by (1). As shown in the first row of the supplementary Fig. S1, it is difficult to recognize the features of \\(\\mathbf{z}^{\\mathrm{src}}\\) in \\(\\mathbf{z}_{0|t}^{\\mathrm{src}}\\), such as hairstyle, details of eyes, and background. This demonstrates that the score \\(\\epsilon_{\\phi}^{\\mathrm{src}}\\) is not exactly adjusted to the given image \\(\\mathbf{z}^{\\mathrm{src}}\\). This deformation becomes more pronounced with increasing \\(t\\). The experiment confirms that \\(\\epsilon_{\\phi}^{\\mathrm{src}}\\) may not be a precise guidance to the source image \\(\\mathbf{z}^{\\mathrm{src}}\\). Therefore, the text-conditioned score \\(\\epsilon_{\\phi}^{\\mathrm{src}}\\) needs to be modified to maintain the identity of the source image \\(\\mathbf{z}^{\\mathrm{src}}\\) in the edited result \\(\\mathbf{z}^{\\mathrm{trg}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Accumulated error in DDS. The transformed image \\(\\mathbf{z}^{\\mathrm{trg}}\\) can be converted back to the original image \\(\\mathbf{z}^{\\mathrm{src}}\\) by reversing the set of \\(\\epsilon\\) used to synthesize \\(\\mathbf{z}^{\\mathrm{trg}}\\) from \\(\\mathbf{z}^{\\mathrm{src}}\\) and swapping \\(\\{\\mathbf{z}^{\\mathrm{src}}, y^{\\mathrm{src}}\\}\\) and \\(\\{\\mathbf{z}^{\\mathrm{trg}}, y^{\\mathrm{trg}}\\}\\) to calculate the DDS loss in (4). If the guidance from \\(\\mathbf{z}^{\\mathrm{src}}\\) to \\(\\mathbf{z}^{\\mathrm{trg}}\\) is computed exactly, the perfect reconstruction can be achieved. Nevertheless, as can be seen from the second row in Fig. 3, DDS [6] fails to restore the original image \\(\\mathbf{z}^{\\mathrm{src}}\\) from the edited image \\(\\mathbf{z}^{\\mathrm{trg}}\\), which implies that the direction from \\(\\mathbf{z}^{\\mathrm{src}}\\) to \\(\\mathbf{z}^{\\mathrm{trg}}\\) is calculated incorrectly. Based on our analysis, this error is because the text-conditioned score \\(\\epsilon_{\\phi}^{\\mathrm{src}}\\) do not refer to the source \\(\\mathbf{z}^{\\mathrm{src}}\\), which can be explicitly expressed as the difference between the injected noise \\(\\epsilon\\) and the score \\(\\epsilon_{\\phi}^{\\mathrm{src}}\\). While the optimization is being processed, the error inevitably accumulates, leading to the undesirable change to the structure and the pose. To address these issues, we investigated whether the guidance from \\(\\mathbf{z}^{\\mathrm{src}}\\) to \\(\\mathbf{z}^{\\mathrm{trg}}\\) can be properly provided while preserving the source's identity, when the timestep \\(t\\) is con" + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.083, + 0.895, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.4, + 0.907, + 0.47 + ], + "angle": 0, + "content": "Figure 3. Accumulated error in DDS. \\(\\mathbf{z}^{\\mathrm{trg}}\\) is edited image of source image \\(\\mathbf{z}^{\\mathrm{src}}\\) by prompt \\(y^{\\mathrm{src}}\\rightarrow y^{\\mathrm{trg}}\\). \\(\\mathbf{z}^{\\mathrm{src\\dagger}}\\) is the inverted image of \\(\\mathbf{z}^{\\mathrm{trg}}\\) by prompt \\(y^{\\mathrm{trg}}\\rightarrow y^{\\mathrm{src}}\\). (First row) Inversion result of DDS with timestep \\(t\\sim \\mathcal{U}(0,0.2)\\). (Second row) Inversion result of DDS with \\(t\\sim \\mathcal{U}(0,1)\\). (Third row) Inversion result of ours." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.486, + 0.907, + 0.668 + ], + "angle": 0, + "content": "strained by \\( t \\sim \\mathcal{U}(0,0.2) \\). This is because the posterior mean \\( \\mathbf{z}_{0|t}^{\\mathrm{src}} \\) and the source image \\( \\mathbf{z}^{\\mathrm{src}} \\) are similar for small timestep \\( t \\), as illustrated in the first row of supplementary Fig. S1. However, as depicted in the first row of Fig. 3, DDS yields unrealistic result with this setting, whereby the structure of the given image \\( \\mathbf{z}^{\\mathrm{src}} \\) is overemphasized. This implies that it is not sufficient to simply limit the timestep \\( t \\) to prevent the score from deviating too far from \\( \\mathbf{z}^{\\mathrm{src}} \\) to correct the misalignment of the score to \\( \\mathbf{z}^{\\mathrm{src}} \\). Hence, we propose a fundamental approach to refine the gradient to achieve identity consistency without unwanted overemphasis on details." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.675, + 0.905, + 0.692 + ], + "angle": 0, + "content": "4.2. Identity-preserving Distillation Sampling (IDS)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.699, + 0.906, + 0.807 + ], + "angle": 0, + "content": "Fixed-point Regularization (FPR). Here, we introduce a Fixed-point Regularization (FPR) method that adjusts the text-conditioned score \\(\\epsilon_{\\phi}^{\\mathrm{src}}\\) to the source image \\(\\mathbf{z}^{\\mathrm{src}}\\). Our key premise is that if the score \\(\\epsilon_{\\phi}^{\\mathrm{src}}\\) is rightly estimated as a gradient to \\(\\mathbf{z}^{\\mathrm{src}}\\), the posterior mean \\(\\mathbf{z}_{0|t}^{\\mathrm{src}}\\) also contains sufficient information about \\(\\mathbf{z}^{\\mathrm{src}}\\). Therefore, FPR loss is designed to minimize the difference between \\(\\mathbf{z}^{\\mathrm{src}}\\) and \\(\\mathbf{z}_{0|t}^{\\mathrm{src}}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.641, + 0.815, + 0.905, + 0.833 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {F P R}} = d \\left(\\mathbf {z} ^ {\\mathrm {s r c}}, \\mathbf {z} _ {0 | t} ^ {\\mathrm {s c}}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.905, + 0.901 + ], + "angle": 0, + "content": "where \\(d(\\mathbf{x}_1,\\mathbf{x}_2)\\) can be any metric to compare \\(\\mathbf{x}_1\\) and \\(\\mathbf{x}_2\\). Here, we employed the Euclidean loss, and further investigations using various metrics are provided in Supplementary Materials." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11118" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.241 + ], + "angle": 0, + "content": "The score \\(\\epsilon_{\\phi}^{\\mathrm{src}}\\) needs to be modified to minimize the FPR loss before obtaining the updated direction. There are two ways to control the score \\(\\epsilon_{\\phi}^{\\mathrm{src}}\\) by altering the injection noise \\(\\epsilon\\) or the source latent \\(\\mathbf{z}_t^{\\mathrm{src}}\\). As illustrated in supplementary Fig. S1, the proposed FPR revises the score \\(\\epsilon_{\\phi}^{\\mathrm{src}}\\) to serve the source's identity for both approaches. Note that the score incorporates the content details, with the updates being performed with respect to the source latent \\(\\mathbf{z}_t^{\\mathrm{src}}\\) compared to the noise \\(\\epsilon\\). Thus, \\(\\mathbf{z}_t^{\\mathrm{src}}\\) is updated to minimize the FPR loss as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.2, + 0.242, + 0.483, + 0.26 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} _ {t} ^ {\\mathrm {s r c}} \\leftarrow \\mathbf {z} _ {t} ^ {\\mathrm {s r c}} - \\lambda \\nabla_ {\\mathbf {z} _ {t} ^ {\\mathrm {s r c}}} \\mathcal {L} _ {\\mathrm {F P R}}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.263, + 0.484, + 0.295 + ], + "angle": 0, + "content": "where \\(\\lambda\\) and \\(N\\) denote a regularization scale and the number of iterations, respectively." + }, + { + "type": "code_caption", + "bbox": [ + 0.093, + 0.306, + 0.406, + 0.322 + ], + "angle": 0, + "content": "Algorithm 1 Fixed-point Regularization (FPR)" + }, + { + "type": "algorithm", + "bbox": [ + 0.093, + 0.325, + 0.475, + 0.515 + ], + "angle": 0, + "content": "Require: \\(\\mathbf{z}^{\\mathrm{src}},y^{\\mathrm{src}},\\epsilon_{\\phi},\\omega ,\\lambda ,N\\) \n1: \\(\\epsilon \\sim \\mathcal{N}(0,\\mathbf{I})\\) \n2: \\(t\\sim \\mathcal{U}(0,1)\\) \n3: \\(\\mathbf{z}_t^{\\mathrm{src}}\\gets \\sqrt{\\alpha_t}\\mathbf{z}^{\\mathrm{src}} + \\sqrt{1 - \\alpha_t}\\epsilon\\) \n4: for \\(\\mathrm{i} = 1,\\dots ,\\mathrm{N}\\) do \n5: \\(\\epsilon_{\\phi}^{\\mathrm{src}}\\gets (1 + \\omega)\\epsilon_{\\phi}(\\mathbf{z}_{t}^{\\mathrm{src}},y^{\\mathrm{src}},t) - \\omega \\epsilon_{\\phi}(\\mathbf{z}_{t}^{\\mathrm{src}},\\emptyset ,t)\\) \n6: \\(\\mathbf{z}_{0|t}^{\\mathrm{src}}\\gets \\frac{1}{\\sqrt{\\alpha_t}} (\\mathbf{z}_t^{\\mathrm{src}} - \\sqrt{1 - \\alpha_t}\\epsilon_\\phi^{\\mathrm{src}})\\) \n7: \\(\\mathcal{L}_{\\mathrm{FPR}}\\gets d(\\mathbf{z}_{0|t}^{\\mathrm{src}},\\mathbf{z}^{\\mathrm{src}})\\) \n8: \\(\\mathbf{z}_t^{\\mathrm{src}}\\gets \\mathbf{z}_t^{\\mathrm{src}} - \\lambda \\nabla_{\\mathbf{z}_t^{\\mathrm{src}}}\\mathcal{L}_{\\mathrm{FPR}}\\) \n9: end for \n10: \\(\\epsilon^{*}\\gets \\frac{1}{\\sqrt{1 - \\alpha_{t}}} (\\mathbf{z}_{t}^{\\mathrm{src}} - \\sqrt{\\alpha_{t}}\\mathbf{z}^{\\mathrm{src}})\\) \n11: return \\(\\epsilon^{*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.525, + 0.484, + 0.58 + ], + "angle": 0, + "content": "Editing with guided noise. Thanks to the proposed FPR, the optimized source latent \\(\\mathbf{z}_t^{\\mathrm{sfc*}}\\) containing the source's identity can be obtained. Then, the guided noise \\(\\epsilon^{*}\\) is extracted as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.584, + 0.483, + 0.613 + ], + "angle": 0, + "content": "\\[\n\\epsilon^ {*} = \\frac {1}{\\sqrt {1 - \\alpha_ {t}}} \\left(\\mathbf {z} _ {t} ^ {\\mathrm {s r c} *} - \\sqrt {\\alpha_ {t}} \\mathbf {z} ^ {\\mathrm {s r c}}\\right). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.62, + 0.484, + 0.666 + ], + "angle": 0, + "content": "\\(\\epsilon^{*}\\) is utilized to produce the stochastic latent \\(\\mathbf{z}_t^{\\mathrm{trg}*}\\) by applying the forward diffusion process to the target image \\(\\mathbf{z}^{\\mathrm{trg}}\\). With \\(\\mathbf{z}_t^{\\mathrm{src}*}\\) and \\(\\mathbf{z}_t^{\\mathrm{trg}*}\\), the updated direction is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.67, + 0.483, + 0.712 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\theta} \\mathcal {L} _ {\\mathrm {I D S}} = \\mathbb {E} _ {t, \\epsilon} \\left[ \\left(\\epsilon_ {\\phi} ^ {\\omega} \\left(\\mathbf {z} _ {t} ^ {\\operatorname {t r g} *}, y ^ {\\operatorname {t r g}}, t\\right) - \\epsilon_ {\\phi} ^ {\\omega} \\left(\\mathbf {z} _ {t} ^ {\\operatorname {s r c} *}, y ^ {\\operatorname {s r c}}, t\\right)\\right) \\frac {\\partial \\mathbf {z} ^ {\\operatorname {t r g}}}{\\partial \\theta} \\right]. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.713, + 0.484, + 0.834 + ], + "angle": 0, + "content": "It is worth noting that \\(\\epsilon^{*}\\) guides the appropriate gradients for editing while conserving the source's identity. In contrast to DDS, the proposed IDS perfectly reconstructs the source from the edited result \\(\\mathbf{z}^{\\mathrm{trg}}\\), as shown in the third row of Fig. 3. This confirms that the correct score and the corresponding injection noise can preserve the identity without further consideration of mutual information. The flowchart of our IDS is illustrated in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.845, + 0.178, + 0.861 + ], + "angle": 0, + "content": "5. Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "We evaluate our method through editing experiments conducted on two experiments. In Sec. 5.1, we perform a com" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.138 + ], + "angle": 0, + "content": "parison on image-to-image editing across several datasets. In Sec. 5.2, we extend our evaluation to editable Neural Radiance Fields (NeRF) [16]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.146, + 0.75, + 0.163 + ], + "angle": 0, + "content": "5.1. Text-guided image editing" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.169, + 0.907, + 0.274 + ], + "angle": 0, + "content": "Baselines. To evaluate our method, we conduct comparative experiments against four state-of-the-art image editing models: Prompt-to-Prompt (P2P) [7], Plug-and-Play (PNP) [26], DDS [6], and CDS [17]. The implementations of the baselines are carried out by referencing the official source code for each method. More details are provided in Supplementary Materials." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.275, + 0.908, + 0.472 + ], + "angle": 0, + "content": "Qualitative Results. We present the qualitative results comparing our method with the baselines in Fig. 4. Prompt-to-Prompt (P2P) [7] performs image editing after applying DDIM inversion [4, 24] to the source image, leading to disregarding the structural components of the source image and following the target prompt excessively. Plug-and-Play (PnP) [26] has limitations in object recognition, as seen in the fourth row of Fig. 4. The third row of Fig. 4 demonstrates that DDS [6] and CDS [17] exhibited limitations, particularly in preserving the structural characteristics of the source image. In contrast, our method successfully edits the image while preserving the structural integrity of the source image." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.471, + 0.909, + 0.698 + ], + "angle": 0, + "content": "Quantitative Results. To measure the identity-preserving performance, we utilize two datasets. First, we collect 250 cat images from the LAION 5B dataset [23] based on [17] for Cat-to-Others task and measure Intersection over Union (IoU). Second, we gather 28 images from the InstructPix2Pix (IP2P) dataset [2], which contains the pairs of source and target images and corresponding prompts and calculate the background Peak-Signal-to-Noise-Ratio (PSNR). Details of the metrics are provided in Supplementary Materials. In addition, we use the LPIPS score [29] for each experiment to quantify the similarity between source and target images. The results are presented in Tab. 1. Our method consistently achieves the lowest LPIPS score across all datasets, indicating that it best preserves the structural semantics of the source images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.698, + 0.909, + 0.759 + ], + "angle": 0, + "content": "For user evaluation, we present 35 comparison sets for four baselines and our method, gathering responses from 47 participants. Participants are asked to choose the most appropriate image for the following three questions: 1. Which" + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.779, + 0.897, + 0.851 + ], + "angle": 0, + "content": "
cat2pigcat2squirrelIp2p
MetricIoU (↑)LPIPS (↓)IoU (↑)LPIPS (↓)PSNR (↑)LPIPS (↓)
P2P [7]0.580.420.520.4620.880.47
PnP [26]0.550.520.530.5223.810.39
DDS [6]0.690.280.650.3026.020.24
CDS [17]0.720.250.710.2627.350.21
IDS (Ours)0.740.220.710.2429.250.19
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.855, + 0.907, + 0.898 + ], + "angle": 0, + "content": "Table 1. Quantitative results for image editing. LPIPS [29] and IoU was measured on LAION 5B [23], while LPIPS and background PSNR was measured on InstructPix2Pix [2]." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11119" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.092, + 0.902, + 0.686 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.69, + 0.908, + 0.72 + ], + "angle": 0, + "content": "Figure 4. Qualitative results of InstructPix2Pix dataset [2]. Our method successfully edits the image aligning with the target text prompt while preserving the structural integrity of the source image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.745, + 0.483, + 0.777 + ], + "angle": 0, + "content": "image best fits the text condition? 2. Which image best preserves the structural information of the original image?" + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.796, + 0.481, + 0.863 + ], + "angle": 0, + "content": "
User Preference Rate (%)GPT score [19]
MetricText (↑)Preserving (↑)Quality (↑)Text (↑)Preserving (↑)Quality (↑)
P2P [7]11.134.808.095.665.375.77
PnP [26]7.727.176.936.546.776.74
DDS [6]20.3010.8216.237.607.517.37
CDS [17]17.0216.7217.088.268.008.09
IDS (Ours)43.8360.4951.678.979.008.80
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.869, + 0.483, + 0.897 + ], + "angle": 0, + "content": "Table 2. User study and GPT scores [19] show that our method achieved the highest scores across all questions for image editing." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.745, + 0.907, + 0.837 + ], + "angle": 0, + "content": "3. Which image has the best quality for text-based image editing? Additionally, we measure the GPT score using the Dreambench++ [19] method, which generates human-aligned assessments for the same questions by refining the scoring into ten distinct levels. As shown in Tab. 2, our method receives the highest ratings for all questions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.847, + 0.659, + 0.863 + ], + "angle": 0, + "content": "5.2. Editing NeRF" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We conduct experiments involving 3D rendering of edited images to demonstrate the effectiveness of our method in" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11120" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.089, + 0.885, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.497, + 0.908, + 0.555 + ], + "angle": 0, + "content": "Figure 5. Qualitative results on Synthetic \\(360^{\\circ}\\) and LLFF datasets. IDS outperforms the baselines by preserving the structural consistency of the source image and maintaining the integrity of regions that should remain unchanged, while precisely editing only the areas specified by the target prompt. Furthermore, comparisons of the depth map results also highlight the superior consistency of our method over other baseline models." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.58, + 0.483, + 0.625 + ], + "angle": 0, + "content": "maintaining structural consistency. This approach is particularly relevant as consistency has an even greater impact on outcomes in 3D environments." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.627, + 0.483, + 0.688 + ], + "angle": 0, + "content": "Datasets. We evaluated our method on widely used NeRF datasets: Synthetic NeRF [16] and LLFF [15]. Since NeRF datasets have no given pairs of source and target prompts, we manually composed image descriptions." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.69, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Qualitative Results. Fig. 5 illustrates the qualitative results of our method compared with NeRF editing baselines. In the first row, the target prompt specifies a precise part of the image for fine-grained editing. DDS [6] and CDS [17] fail to differentiate and edit the specific area. At the same time, our method accurately identifies the region indicated by the target prompt in the image and performs detailed editing exclusively on that part. The second row demonstrates a scenario in which the target prompt is designed to edit the mood of the image. Our approach adjusts the colors associated with \"autumn\" and \"leaves\" throughout the image while maintaining consistency in the \"trunk\" whereas DDS and CDS also changed the \"trunk\". In terms of depth maps, our method generates clean depth maps with minimal noise" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.58, + 0.907, + 0.611 + ], + "angle": 0, + "content": "after image editing, whereas DDS and CDS introduce noticeable noise into the depth maps." + }, + { + "type": "table", + "bbox": [ + 0.528, + 0.621, + 0.895, + 0.683 + ], + "angle": 0, + "content": "
MetricCLIP [21] (↑)User Preference Rate (%)
Text (↑)Preserving (↑)Quality (↑)
DDS [6]0.159636.8828.3732.62
CDS [17]0.159722.7023.4021.28
IDS (Ours)0.162640.4248.2346.10
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.693, + 0.908, + 0.737 + ], + "angle": 0, + "content": "Table 3. Quantitative results of NeRF editing with respect to CLIP score and User Preference Rate. IDS demonstrates superior quantitative performance compared to the baselines." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Quantitative Results. Based on edited images, we performed 3D rendering and subsequently conducted quantitative evaluations provided in Tab. 3. To assess whether the edited 3D images are precisely aligned with the target prompts, we measured the CLIP [21] scores at \\(200\\mathrm{k}\\) iterations of training on the LLFF dataset. We additionally present a user evaluation conducted under the same setup in Sec. 5.1. Consistent with the trends observed in the qualitative results, our method demonstrates superior performance in the quantitative evaluations compared to other baselines." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "11121" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.09, + 0.472, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.171, + 0.162, + 0.408, + 0.175 + ], + "angle": 0, + "content": "\"A photo of shark\" \\(\\rightarrow\\) \"A photo of dolphin\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.175, + 0.483, + 0.23 + ], + "angle": 0, + "content": "Figure 6. Ablation study on scale \\(\\lambda\\). To show the effect of scale more extremely, the number of iterations of FPR is set as 1. The result of DDS is the same as \\(\\lambda = 0\\) since it means no update for source identity." + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.236, + 0.473, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.135, + 0.393, + 0.443, + 0.407 + ], + "angle": 0, + "content": "\"The green leaves\" \\(\\rightarrow\\) \"Yellow and red leaves in autumn\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.408, + 0.483, + 0.45 + ], + "angle": 0, + "content": "Figure 7. Optimization steps. Our method effectively preserves the consistency of the source image, even as the number of iteration steps increases up to 400." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.46, + 0.473, + 0.529 + ], + "angle": 0, + "content": "
optim stepFPR iterLPIPS(↓)CLIP(↑)time (sec/img)Memory (GB)
DDS200-0.2400.29322.456.27
CDS200-0.2100.28759.318.83
IDS20010.1990.28550.808.63
30.1900.277107.77
10030.1650.26554.04
1500.1800.27281.25
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.534, + 0.483, + 0.576 + ], + "angle": 0, + "content": "Table 4. Computational complexity on 28 images of InstructPix2Pix [2] for various settings. Lower LPIPS and higher CLIP scores mean better quality." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.603, + 0.212, + 0.618 + ], + "angle": 0, + "content": "6. Discussions" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.628, + 0.315, + 0.643 + ], + "angle": 0, + "content": "6.1. Ablation studies on FPR" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.652, + 0.483, + 0.743 + ], + "angle": 0, + "content": "FPR iteration \\(N\\). We conduct experiments on FPR iterations \\(N\\) to evaluate its impact and determine the optimal iteration count. Although performing just one iteration of FPR is sufficient to preserve the source identity, as shown in lower LPIPS score than baselines of Tab. 4, we set \\(N = 3\\) to emphasize the purpose of our method." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.744, + 0.483, + 0.835 + ], + "angle": 0, + "content": "Scale \\(\\lambda\\). The scaling factor \\(\\lambda\\) of FPR determines how much information of source latent \\(\\mathbf{z}^{\\mathrm{src}}\\) is kept. As shown in Fig. 6, increasing the scale preserves the attributes of the source image, resulting in more successful editing when it is hard to translate due to the structural mismatch between the source and the target prompt." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.847, + 0.273, + 0.864 + ], + "angle": 0, + "content": "6.2. Optimization steps" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To show that our method can prevent error accumulation during translation, we set the experiment to extend the num" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.182 + ], + "angle": 0, + "content": "ber of optimization steps from 200 to 400. In the results of DDS and CDS, there is color boosting or loss of details due to the cumulated error. In contrast, IDS maintains the characteristics of the original images, such as the color of the pumpkin in the first row of Fig. 7 and the shape of the leaf in the second row of Fig. 7, better than other methods." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.196, + 0.628, + 0.211 + ], + "angle": 0, + "content": "7. Limitation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.221, + 0.907, + 0.433 + ], + "angle": 0, + "content": "The proposed IDS demonstrates outstanding performance across evaluation metrics assessing consistency between source and target images. However, during FPR process, IDS relies solely on information from the source \\(\\{\\mathbf{z}^{\\mathrm{src}}, y^{\\mathrm{src}}\\}\\) without incorporating target-side information. This results in comparatively lower CLIP scores [21] than other baselines (Tab. 5) and failure cases for more complex translations (Fig. 8). In addition, our method requires additional computational overhead (Tab. 4) since FPR is applied to each optimization iteration. Detailed discussion about our limitation is provided in Sec. G of supplementary. Our future direction will explore changing the score conditioned by the target prompt \\(y^{\\mathrm{trg}}\\), leading to a better alignment with \\(y^{\\mathrm{trg}}\\)." + }, + { + "type": "table", + "bbox": [ + 0.529, + 0.444, + 0.889, + 0.494 + ], + "angle": 0, + "content": "
P2P [7]PnP [26]DDS [6]CDS [17]IDS (Ours)
cat2lion0.290.210.300.290.29
cat2dog0.270.260.270.270.26
Ip2p0.280.300.290.290.28
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.498, + 0.906, + 0.527 + ], + "angle": 0, + "content": "Table 5. Limitation of IDS with respect to CLIP score[21] for image editing on LAION 5B [23] and InstructPix2Pix [2]." + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.534, + 0.9, + 0.62 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.558, + 0.621, + 0.859, + 0.633 + ], + "angle": 0, + "content": "\"Photo free night, house, aurora\" → \"... with two dogs\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.565, + 0.634, + 0.853, + 0.647 + ], + "angle": 0, + "content": "Figure 8. Failure case for complex text prompt." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.657, + 0.634, + 0.672 + ], + "angle": 0, + "content": "8. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.682, + 0.907, + 0.894 + ], + "angle": 0, + "content": "We proposed a new distillation sampling method using a fixed-point regularization which aligns the text-conditioned score towards identity-preserved manifolds. The proposed fixed-point regularization preserves the source's identity by re-projecting the intermediate score status on posterior means. In this manner, corrected noises guide a gradient of distilled score toward identity-consistent manifolds. Owing to self-correction by a fixed-point iterator and guided injection noise, the proposed identity-preserving distillation sampling provides clear and unambiguous representations corresponding to the given prompts in text-guided image editing and editable neural radiance field (NeRF). Furthermore, our model can be utilized as a universal module in addition to the existing score-sampling processes." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11122" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.486, + 0.152 + ], + "angle": 0, + "content": "Acknowledgement This work was supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (RS-2024-00335741, RS-2024-00357197)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.166, + 0.188, + 0.182 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.484, + 0.261 + ], + "angle": 0, + "content": "[1] Donghoon Ahn, Hyoungwon Cho, Jaewon Min, Wooseok Jang, Jungwoo Kim, SeonHwa Kim, Hyun Hee Park, Kyong Hwan Jin, and Seungryong Kim. Self-rectifying diffusion sampling with perturbed-attention guidance. arXiv preprint arXiv:2403.17377, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.262, + 0.484, + 0.331 + ], + "angle": 0, + "content": "[2] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 1, 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.333, + 0.484, + 0.401 + ], + "angle": 0, + "content": "[3] Hyungjin Chung, Jeongsol Kim, Michael Thompson McCann, Marc Louis Klasky, and Jong Chul Ye. Diffusion posterior sampling for general noisy inverse problems. In The Eleventh International Conference on Learning Representations, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.403, + 0.484, + 0.445 + ], + "angle": 0, + "content": "[4] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.447, + 0.484, + 0.488 + ], + "angle": 0, + "content": "[5] Bradley Efron. Tweedie's formula and selection bias. Journal of the American Statistical Association, 106(496):1602-1614, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.484, + 0.545 + ], + "angle": 0, + "content": "[6] Amir Hertz, Kfir Aberman, and Daniel Cohen-Or. Delta denoising score. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2328-2337, 2023. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.547, + 0.484, + 0.615 + ], + "angle": 0, + "content": "[7] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-or. Prompt-to-prompt image editing with cross-attention control. In The Eleventh International Conference on Learning Representations, 2023. 1, 2, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.617, + 0.484, + 0.659 + ], + "angle": 0, + "content": "[8] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.66, + 0.484, + 0.702 + ], + "angle": 0, + "content": "[9] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.703, + 0.484, + 0.771 + ], + "angle": 0, + "content": "[10] Susung Hong, Gyuseong Lee, Wooseok Jang, and Seungryong Kim. Improving sample quality of diffusion models using self-attention guidance. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7462–7471, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.773, + 0.484, + 0.815 + ], + "angle": 0, + "content": "[11] Oren Katzir, Or Patashnik, Daniel Cohen-Or, and Dani Lischinski. Noise-free score distillation. arXiv preprint arXiv:2310.17590, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.484, + 0.871 + ], + "angle": 0, + "content": "[12] Juil Koo, Chanho Park, and Minhyuk Sung. Posterior distillation sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13352-13361, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.484, + 0.903 + ], + "angle": 0, + "content": "[13] Barak Meiri, Dvir Samuel, Nir Darshan, Gal Chechik, Shai Avidan, and Rami Ben-Ari. Fixed-point inversion for text-" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.19, + 0.484, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.093, + 0.907, + 0.12 + ], + "angle": 0, + "content": "to-image diffusion models. arXiv preprint arXiv:2312.12540, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.907, + 0.19 + ], + "angle": 0, + "content": "[14] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.907, + 0.262 + ], + "angle": 0, + "content": "[15] Ben Mildenhall, Pratul P Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (ToG), 38(4):1-14, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.906, + 0.331 + ], + "angle": 0, + "content": "[16] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.907, + 0.404 + ], + "angle": 0, + "content": "[17] Hyelin Nam, Gihyun Kwon, Geon Yeong Park, and Jong Chul Ye. Contrastive denoising score for text-guided latent diffusion image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9192-9201, 2024. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.406, + 0.906, + 0.447 + ], + "angle": 0, + "content": "[18] Neal Parikh, Stephen Boyd, et al. Proximal algorithms. Foundations and trends® in Optimization, 1(3):127-239, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.449, + 0.907, + 0.517 + ], + "angle": 0, + "content": "[19] Yuang Peng, Yuxin Cui, Haomiao Tang, Zekun Qi, Runpei Dong, Jing Bai, Chunrui Han, Zheng Ge, Xiangyu Zhang, and Shu-Tao Xia. Dreambench++: A human-aligned benchmark for personalized image generation. arXiv preprint arXiv:2406.16855, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.906, + 0.561 + ], + "angle": 0, + "content": "[20] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.563, + 0.907, + 0.645 + ], + "angle": 0, + "content": "[21] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.647, + 0.907, + 0.716 + ], + "angle": 0, + "content": "[22] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.907, + 0.8 + ], + "angle": 0, + "content": "[23] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.907, + 0.844 + ], + "angle": 0, + "content": "[24] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[25] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.515, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11123" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[26] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023. 1, 2, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.163, + 0.483, + 0.232 + ], + "angle": 0, + "content": "[27] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.234, + 0.483, + 0.303 + ], + "angle": 0, + "content": "[28] Zike Wu, Pan Zhou, Xuanyu Yi, Xiaoding Yuan, and Hanwang Zhang. Consistent3d: Towards consistent high-fidelity text-to-3d generation with deterministic sampling prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9892–9902, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.304, + 0.483, + 0.373 + ], + "angle": 0, + "content": "[29] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.375, + 0.482, + 0.43 + ], + "angle": 0, + "content": "[30] Zhizhuo Zhou and Shubham Tulsiani. Sparsefusion: Distilling view-conditioned diffusion for 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12588-12597, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.092, + 0.483, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "11124" + } + ] +] \ No newline at end of file diff --git a/2025/Identity-preserving Distillation Sampling by Fixed-Point Iterator/8dc89aa9-e649-490d-ba0b-e87a936deb98_origin.pdf b/2025/Identity-preserving Distillation Sampling by Fixed-Point Iterator/8dc89aa9-e649-490d-ba0b-e87a936deb98_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b098bbaa949569b2f863870e1c3a793964c367fd --- /dev/null +++ b/2025/Identity-preserving Distillation Sampling by Fixed-Point Iterator/8dc89aa9-e649-490d-ba0b-e87a936deb98_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a9c9c83ab9ccde98e601ab0d8f7ca88fc474c90792cfd3f5e5a7c6adfb6886fd +size 7693665 diff --git a/2025/Identity-preserving Distillation Sampling by Fixed-Point Iterator/full.md b/2025/Identity-preserving Distillation Sampling by Fixed-Point Iterator/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c5e0c8cd572b32b4eee25b9f2aa99df57bdd4b67 --- /dev/null +++ b/2025/Identity-preserving Distillation Sampling by Fixed-Point Iterator/full.md @@ -0,0 +1,297 @@ +# Identity-preserving Distillation Sampling by Fixed-Point Iterator + +SeonHwa Kim $^{1}$ Jiwon Kim $^{1}$ Soobin Park $^{2}$ Donghoon Ahn $^{1}$ Jiwon Kang $^{1}$ Seungryong Kim $^{3*}$ Kyong Hwan Jin $^{1*}$ Eunju Cha $^{2*}$ $^{1}$ Korea University $^{2}$ Sookmyung Women's University $^{3}$ KAIST + +{sunkim0062, jwonkim, dhahn99, jiwon7258, kyong_jin}@korea.ac.kr, {psb1219j, eunju.cha}@sookmyung.ac.kr, seungryong.kim@kaist.ac.kr + +# Abstract + +Score distillation sampling (SDS) demonstrates a powerful capability for text-conditioned 2D image and 3D object generation by distilling the knowledge from learned score functions. However, SDS often suffers from blurriness caused by noisy gradients. When SDS meets the image editing, such degradations can be reduced by adjusting bias shifts using reference pairs, but the de-biasing techniques are still corrupted by erroneous gradients. To this end, we introduce Identity-preserving Distillation Sampling (IDS), which compensates for the gradient leading to undesired changes in the results. Based on the analysis that these errors come from the text-conditioned scores, a new regularization technique, called fixed-point iterative regularization (FPR), is proposed to modify the score itself, driving the preservation of the identity even including poses and structures. Thanks to a self-correction by FPR, the proposed method provides clear and unambiguous representations corresponding to the given prompts in image-to-image editing and editable neural radiance field (NeRF). The structural consistency between the source and the edited data is obviously maintained compared to other state-of-the-art methods. Our code is https://github.com/shhh0620/IDS + +# 1. Introduction + +Diffusion models [4, 8, 9, 22, 25] have shown powerful representations on text-to-image (T2I) generative tasks. With the advance of classifier guidance (CG) and classifier-free guidance (CFG) paradigms [1, 4, 8, 10], diffusion models improve the quality of generated samples [9, 25]. Such high-quality image generators can be easily extended to image editing by simply modifying forward/reverse iterations [14], applying CFG with a target prompt [2, 7] or interchanging attention layers [26]. + +![](images/6a18606f6793dfc1980d978ecb2ab4577f6e06385b0fef0d0185ce334224edb2.jpg) +Figure 1. Trace of guided updating from source to target images using delta denoising score (DDS) and identity-preserving distillation sampling (IDS). DDS moves a gradient of score function toward $\mathcal{M}_{\mathbf{z}}$ manifold directed by stochastic direction $\epsilon$ . In contrast, IDS moves a gradient with a corrected direction by a fixed-point regularization. + +Recently, Delta Denoising Score (DDS) [6] is proposed to edit a source image by distilling the rich generative prior of T2I diffusion models. It is based on the analysis of Score Distillation Sampling (SDS) [20], originally developed to optimize a parametric generator such as Neural Radiance Fields (NeRFs) [16] by exploiting the learned score of the diffusion models. Even though SDS offers remarkable performance in synthesizing 3D scenes, noisy gradients from stochastic perturbations lead to significantly over-saturated results that faithfully follow the given text prompts. In the context of image editing, text prompts do not often include information about the identity of the source image, such as the background, the object's pose, or the structure of the + +![](images/738d635f30a3df04304cc174651668ddc18fbbd91b32a508accacc2c253472d9.jpg) +Figure 2. Flowchart of IDS. The backbone of our algorithm employs DDS [6] framework to distill score function into a target image. Our fixed-point regularization (FPR) obtains a guided noise, $\epsilon^{*}$ , from iterative updates using posterior mean computed by Tweedie's formula. When distilling the score function to a target image, the guided noise is updated while maintaining the identity of the source. + +content, which should be retained during updates. Thus, DDS is designed to resolve such blurriness by erasing gradients of non-text-aligned features from SDS gradients. There is no explicit procedure to preserve the source's identity in DDS updates because the fine gradient may provide the conserved identity. However, this cannot be guaranteed if many variations in the structure are possible, such as editing the image of a cat into a pig, as shown in Fig. 1. To alleviate this problem, Contrastive Denoising Score (CDS) [17] and Posterior Distillation Sampling (PDS) [12] are introduced to maximize the mutual information of the source image and edited image. Although such algorithms rely heavily on text prompts, the algorithms have yet to analyze the inherent error caused by text-conditioned scores. + +To this end, we investigate the underlying meaning of text-conditioned score. The gradient maps the stochastic latent, generated by applying the forward diffusion process to the given image, to one of the possible images described by the prompt, including the original image. Simply, the score obtained from the latent of the source image ('source latent') and the source prompt can be a gradient to another image represented by the identical text. Based on this interpretation, the accumulation of misaligned directions causes the loss of the source's identity, leading to structural changes in the result with DDS, as shown in Fig. 1. + +To address this issue, we propose a novel score distillation sampling to effectively preserve the identity of the original image by self-correcting the misaligned gradients, called Identity-preserving Distillation Sampling (IDS). The key insight is that if the score is precisely adjusted to the source image, the conditional expectation of the source image given the source latent contains meaningful information + +that should be preserved during the editing. This conditional expectation corresponds to the posterior mean computed by Tweedie's formula using the learned score [3, 5]. The source latent is iteratively updated to make the posterior mean similar to the source image. This procedure, named a fixed-point iterative regularization (FPR), results in the aligned score with the source that provides reliable gradients for editing, as illustrated in Fig. 1. Following IDS update is performed using guided noise extracted from the refined source latent, rather than random Gaussian noise. This further ensures the identity preservation. Our method demonstrated superior results compared with baselines in two tasks: editing images by prompts and editing NeRF. + +In summary, our main contributions are as follows: + +- We obtain reliable gradients for the score distillation function by a fixed-point iterator with respect to posterior means. The iterator corrects the text-conditioned score, guiding SDS gradients toward reliable pre-trained manifolds. +- Our fixed-point regularization preserves the identities of sources such as structures and poses in edited targets for 2D and 3D editing. Such preservation is well demonstrated in qualitative and quantitative results. + +# 2. Related works + +# 2.1. Image Editing with Diffusion Models + +With the great success of image generation using diffusion models, the pre-trained diffusion models have been recently employed for image editing tasks, demonstrating significant advancements in the quality and flexibility of generated edits [2, 7, 12, 14, 17, 26]. Stochastic Differential + +Editing (SDEdit) [14] is a pioneering work in which the source image was modified by adding noise and solving reverse stochastic differential equations. Thanks to the text-conditional Latent Diffusion Model (LDM), a.k.a. Stable Diffusion [22], text-driven editing approaches have been introduced. Specifically, the text embedding was injected through the cross-attention layer of the model for image editing and translation, while retaining the structure of the original image [7, 26]. The editing was further controlled by rescaling the attention of the specific word [7] or by manipulating the self-attention features [26]. These approaches provide greater control by balancing fidelity between the edited prompt and the source image without the need for model training, fine-tuning, additional data, or optimization. However, the current DDIM-based inversion [24] can lead to unsatisfactory reconstructions for real images, and the cross-attention bottleneck limits its effectiveness for broader edits. Crafting suitable prompts also remains challenging for complex compositions. + +# 2.2. Score distillation sampling + +Score Distillation Sampling (SDS) [20] enables text-driven 3D synthesis by leveraging probability density distillation loss to distill knowledge from 2D diffusion models, allowing high-quality 3D scene generation based on textual prompts without 3D training data. However, SDS has limitations, often producing oversaturated and overly smooth 3D models, and lacking diversity across initializations. To address these limitations of SDS, various models have been proposed based on exploiting multi-step denoising [30], a variation approach [27], negative conditioning [11], and ordinary differential equation trajectory [28]. To mitigate the limitation of noisy gradients in SDS, which hampers precise image editing, DDS [6] was introduced. By computing the delta between the derived gradient and the target pair, DDS effectively isolates and removes unwanted noise in the gradient direction. Despite these advancements, DDS still faces challenges in preserving the complete structural consistency of the source image's identity. + +# 3. Preliminaries + +# 3.1. Diffusion Model and Sampling Guidance + +Text-to-image diffusion models $\epsilon_{\phi}(\cdot)$ are based on diffusion probabilistic models (DPMs) [9, 22, 25]. The models are trained to estimate the denoising score when the original image $\mathbf{z}_0$ and the text condition $y$ are given: + +$$ +\mathcal {L} (\phi) = \mathbb {E} _ {t, \epsilon} [ \| \epsilon_ {\phi} (\mathbf {z} _ {t}, y, t) - \epsilon \| _ {2} ^ {2} ], +$$ + +where $\epsilon \sim \mathcal{N}(0,\mathbf{I})$ and $t\sim \mathcal{U}(0,1)$ $\mathbf{z}_t$ refers to the stochastic latent of $\mathbf{z}_0$ via the forward diffusion process as follows: + +$$ +\mathbf {z} _ {t} = \sqrt {\alpha_ {t}} \mathbf {z} _ {0} + \sqrt {1 - \alpha_ {t}} \epsilon , \tag {1} +$$ + +where $\alpha_{t}$ is noise schedule. With the trained $\epsilon_{\phi}(\cdot)$ , high-quality samples can be generated using the classifier-free guidance (CFG) [8] by subtracting unconditioned denoising score from the conditioned score with guidance scale $\omega$ : + +$$ +\epsilon_ {\phi} ^ {\omega} \left(\mathbf {z} _ {t}, y, t\right) = (1 + \omega) \epsilon_ {\phi} \left(\mathbf {z} _ {t}, y, t\right) - \omega \epsilon_ {\phi} \left(\mathbf {z} _ {t}, \varnothing , t\right). \tag {2} +$$ + +# 3.2. Score Distillation Sampling (SDS) + +With pretrained text-to-image diffusion models $\epsilon_{\phi}(\cdot)$ , SDS [20] synthesizes 3D data $\mathbf{z}$ for a given text prompt $y$ by optimizing the differentiable rendering function paretrized by $\theta$ , where $\mathbf{z} = g(\theta)$ : + +$$ +\nabla_ {\theta} \mathcal {L} _ {\mathrm {S D S}} (\mathbf {z}, y) = \mathbb {E} _ {t, \epsilon} \left[ \omega (t) \left(\epsilon_ {\phi} ^ {\omega} \left(\mathbf {z} _ {t}, y, t\right) - \epsilon\right) \frac {\partial \mathbf {z}}{\partial \theta} \right]. \tag {3} +$$ + +The optimized parameters $\theta^{*}$ provide the text-conditioned 3D volume that follows the diffusion prior [20]. However, a single text prompt $y$ can refer to many different 3D volumes, each with diverse backgrounds or structural details of the object. Therefore, an inherent limitation of SDS [20] is that the score conditioned by the prompt $y$ does not always provide the diffusion prior to the identical object during the optimization process, leading to blurry and unclear results. + +# 3.3. Delta Denoising Score (DDS) + +DDS [6] is proposed to synthesize the image $\mathbf{z}^{\mathrm{trg}}$ from the given source image $\mathbf{z}^{\mathrm{src}}$ and its corresponding prompt $y^{\mathrm{src}}$ , which is aligned to the target prompt $y^{\mathrm{trg}}$ . Based on the insight that the gradient should be zero if $y^{\mathrm{trg}}$ matches $y^{\mathrm{src}}$ , DDS minimizes the identity change of $\mathbf{z}^{\mathrm{src}}$ by simple replacing $\epsilon$ in (3) with the score $\epsilon_{\phi}^{\omega}(\mathbf{z}_{t}^{\mathrm{src}}, y^{\mathrm{src}}, t)$ as follows: + +$$ +\nabla_ {\theta} \mathcal {L} _ {\mathrm {D D S}} = \mathbb {E} _ {t, \epsilon} \left[ \left(\epsilon_ {\phi} ^ {\omega} \left(\mathbf {z} _ {t} ^ {\operatorname {t r g}}, y ^ {\operatorname {t r g}}, t\right) - \epsilon_ {\phi} ^ {\omega} \left(\mathbf {z} _ {t} ^ {\operatorname {s r c}}, y ^ {\operatorname {s r c}}, t\right)\right) \frac {\partial \mathbf {z} ^ {\operatorname {t r g}}}{\partial \theta} \right]. \tag {4} +$$ + +For simplicity, we denote $\epsilon_{\phi}^{\mathrm{trg}} = \epsilon_{\phi}^{\omega}(\mathbf{z}^{\mathrm{trg}},y^{\mathrm{trg}},t)$ and $\epsilon_{\phi}^{\mathrm{src}} = \epsilon_{\phi}^{\omega}(\mathbf{z}^{\mathrm{src}},y^{\mathrm{src}},t)$ . Here, $\epsilon_{\phi}^{\mathrm{trg}}$ and $\epsilon_{\phi}^{\mathrm{src}}$ can be interpreted as the gradients representing the direction from $\mathbf{z}_t^{\mathrm{trg}}$ to $\mathbf{z}^{\mathrm{trg}}$ and the direction from $\mathbf{z}_t^{\mathrm{src}}$ to $\mathbf{z}^{\mathrm{src}}$ , respectively. $\theta = \mathbf{z}^{\mathrm{trg}}$ is thus gradually optimized along the direction from $\mathbf{z}^{\mathrm{src}}$ to $\mathbf{z}^{\mathrm{trg}}$ , as shown in Fig. 1. It is worth noting that the guidance of the update can be calculated at the same point $\mathbf{z}_t^{\mathrm{trg}}$ , thanks to the shared $\epsilon$ . However, the slight error in the gradient caused by the score $\epsilon_{\phi}^{\mathrm{src}}$ still leads to the incorrect direction for the optimization. + +# 3.4. Fixed-point Iteration + +In numerical analysis, a fixed-point iteration [18] is an iterative method to find fixed points of a function $f$ , where $f(x) = x$ . Given an initial point $x_0$ , the iteration is defined as: + +$$ +x _ {n + 1} = f (x _ {n}), \quad n = 0, 1, 2, \dots +$$ + +Under appropriate conditions, this sequence converges to a fixed point. Thanks to its applicability to non-linear problems with low computational costs, fixed-point iteration is widely used in optimization, including applications in the context of diffusion models [13]. + +# 4. Method + +Given the source pair $\{\mathbf{z}^{\mathrm{src}}, y^{\mathrm{src}}\}$ , the aim of our work is to provide an edited result $\mathbf{z}^{\mathrm{trg}}$ that is aligned with $y^{\mathrm{trg}}$ while maintaining the source's identity. To this end, we introduce a novel approach called Identity-preserving Distillation Sampling (IDS), which (1) corrects the error of the gradient aligned with the text prompt by the fixed-point iterator and (2) provides the result $\mathbf{z}^{\mathrm{trg}}$ using the guided noise. + +# 4.1. Motivation + +Analysis of the text-conditioned score. We first investigated how much identity of the given image $\mathbf{z}^{\mathrm{src}}$ could be contained in the text-conditioned score $\epsilon_{\phi}^{\mathrm{src}}$ . To do this, we conducted the experiment to compare the original image $\mathbf{z}^{\mathrm{src}}$ and the posterior mean $\mathbf{z}_{0|t}^{\mathrm{src}} = \mathbb{E}[\mathbf{z}^{\mathrm{src}}|\mathbf{z}_t^{\mathrm{src}}]$ , which is given by: + +$$ +\mathbf {z} _ {0 | t} ^ {\mathrm {s r c}} = \frac {1}{\sqrt {\alpha_ {t}}} \left(\mathbf {z} _ {t} ^ {\mathrm {s r c}} - \sqrt {1 - \alpha_ {t}} \epsilon_ {\phi} ^ {\mathrm {s r c}}\right), \tag {5} +$$ + +where $\mathbf{z}_t^{\mathrm{src}}$ denotes the source latent generated by (1). As shown in the first row of the supplementary Fig. S1, it is difficult to recognize the features of $\mathbf{z}^{\mathrm{src}}$ in $\mathbf{z}_{0|t}^{\mathrm{src}}$ , such as hairstyle, details of eyes, and background. This demonstrates that the score $\epsilon_{\phi}^{\mathrm{src}}$ is not exactly adjusted to the given image $\mathbf{z}^{\mathrm{src}}$ . This deformation becomes more pronounced with increasing $t$ . The experiment confirms that $\epsilon_{\phi}^{\mathrm{src}}$ may not be a precise guidance to the source image $\mathbf{z}^{\mathrm{src}}$ . Therefore, the text-conditioned score $\epsilon_{\phi}^{\mathrm{src}}$ needs to be modified to maintain the identity of the source image $\mathbf{z}^{\mathrm{src}}$ in the edited result $\mathbf{z}^{\mathrm{trg}}$ . + +Accumulated error in DDS. The transformed image $\mathbf{z}^{\mathrm{trg}}$ can be converted back to the original image $\mathbf{z}^{\mathrm{src}}$ by reversing the set of $\epsilon$ used to synthesize $\mathbf{z}^{\mathrm{trg}}$ from $\mathbf{z}^{\mathrm{src}}$ and swapping $\{\mathbf{z}^{\mathrm{src}}, y^{\mathrm{src}}\}$ and $\{\mathbf{z}^{\mathrm{trg}}, y^{\mathrm{trg}}\}$ to calculate the DDS loss in (4). If the guidance from $\mathbf{z}^{\mathrm{src}}$ to $\mathbf{z}^{\mathrm{trg}}$ is computed exactly, the perfect reconstruction can be achieved. Nevertheless, as can be seen from the second row in Fig. 3, DDS [6] fails to restore the original image $\mathbf{z}^{\mathrm{src}}$ from the edited image $\mathbf{z}^{\mathrm{trg}}$ , which implies that the direction from $\mathbf{z}^{\mathrm{src}}$ to $\mathbf{z}^{\mathrm{trg}}$ is calculated incorrectly. Based on our analysis, this error is because the text-conditioned score $\epsilon_{\phi}^{\mathrm{src}}$ do not refer to the source $\mathbf{z}^{\mathrm{src}}$ , which can be explicitly expressed as the difference between the injected noise $\epsilon$ and the score $\epsilon_{\phi}^{\mathrm{src}}$ . While the optimization is being processed, the error inevitably accumulates, leading to the undesirable change to the structure and the pose. To address these issues, we investigated whether the guidance from $\mathbf{z}^{\mathrm{src}}$ to $\mathbf{z}^{\mathrm{trg}}$ can be properly provided while preserving the source's identity, when the timestep $t$ is con + +![](images/3aa8a2e8e325d12c6f49923f8ac4a7c927f92d57dab571a6b15aa14a2c0240e0.jpg) +Figure 3. Accumulated error in DDS. $\mathbf{z}^{\mathrm{trg}}$ is edited image of source image $\mathbf{z}^{\mathrm{src}}$ by prompt $y^{\mathrm{src}}\rightarrow y^{\mathrm{trg}}$ . $\mathbf{z}^{\mathrm{src\dagger}}$ is the inverted image of $\mathbf{z}^{\mathrm{trg}}$ by prompt $y^{\mathrm{trg}}\rightarrow y^{\mathrm{src}}$ . (First row) Inversion result of DDS with timestep $t\sim \mathcal{U}(0,0.2)$ . (Second row) Inversion result of DDS with $t\sim \mathcal{U}(0,1)$ . (Third row) Inversion result of ours. + +strained by $t \sim \mathcal{U}(0,0.2)$ . This is because the posterior mean $\mathbf{z}_{0|t}^{\mathrm{src}}$ and the source image $\mathbf{z}^{\mathrm{src}}$ are similar for small timestep $t$ , as illustrated in the first row of supplementary Fig. S1. However, as depicted in the first row of Fig. 3, DDS yields unrealistic result with this setting, whereby the structure of the given image $\mathbf{z}^{\mathrm{src}}$ is overemphasized. This implies that it is not sufficient to simply limit the timestep $t$ to prevent the score from deviating too far from $\mathbf{z}^{\mathrm{src}}$ to correct the misalignment of the score to $\mathbf{z}^{\mathrm{src}}$ . Hence, we propose a fundamental approach to refine the gradient to achieve identity consistency without unwanted overemphasis on details. + +# 4.2. Identity-preserving Distillation Sampling (IDS) + +Fixed-point Regularization (FPR). Here, we introduce a Fixed-point Regularization (FPR) method that adjusts the text-conditioned score $\epsilon_{\phi}^{\mathrm{src}}$ to the source image $\mathbf{z}^{\mathrm{src}}$ . Our key premise is that if the score $\epsilon_{\phi}^{\mathrm{src}}$ is rightly estimated as a gradient to $\mathbf{z}^{\mathrm{src}}$ , the posterior mean $\mathbf{z}_{0|t}^{\mathrm{src}}$ also contains sufficient information about $\mathbf{z}^{\mathrm{src}}$ . Therefore, FPR loss is designed to minimize the difference between $\mathbf{z}^{\mathrm{src}}$ and $\mathbf{z}_{0|t}^{\mathrm{src}}$ as follows: + +$$ +\mathcal {L} _ {\mathrm {F P R}} = d \left(\mathbf {z} ^ {\mathrm {s r c}}, \mathbf {z} _ {0 | t} ^ {\mathrm {s c}}\right), \tag {6} +$$ + +where $d(\mathbf{x}_1,\mathbf{x}_2)$ can be any metric to compare $\mathbf{x}_1$ and $\mathbf{x}_2$ . Here, we employed the Euclidean loss, and further investigations using various metrics are provided in Supplementary Materials. + +The score $\epsilon_{\phi}^{\mathrm{src}}$ needs to be modified to minimize the FPR loss before obtaining the updated direction. There are two ways to control the score $\epsilon_{\phi}^{\mathrm{src}}$ by altering the injection noise $\epsilon$ or the source latent $\mathbf{z}_t^{\mathrm{src}}$ . As illustrated in supplementary Fig. S1, the proposed FPR revises the score $\epsilon_{\phi}^{\mathrm{src}}$ to serve the source's identity for both approaches. Note that the score incorporates the content details, with the updates being performed with respect to the source latent $\mathbf{z}_t^{\mathrm{src}}$ compared to the noise $\epsilon$ . Thus, $\mathbf{z}_t^{\mathrm{src}}$ is updated to minimize the FPR loss as follows: + +$$ +\mathbf {z} _ {t} ^ {\mathrm {s r c}} \leftarrow \mathbf {z} _ {t} ^ {\mathrm {s r c}} - \lambda \nabla_ {\mathbf {z} _ {t} ^ {\mathrm {s r c}}} \mathcal {L} _ {\mathrm {F P R}}, \tag {7} +$$ + +where $\lambda$ and $N$ denote a regularization scale and the number of iterations, respectively. + +Algorithm 1 Fixed-point Regularization (FPR) +Require: $\mathbf{z}^{\mathrm{src}},y^{\mathrm{src}},\epsilon_{\phi},\omega ,\lambda ,N$ +1: $\epsilon \sim \mathcal{N}(0,\mathbf{I})$ +2: $t\sim \mathcal{U}(0,1)$ +3: $\mathbf{z}_t^{\mathrm{src}}\gets \sqrt{\alpha_t}\mathbf{z}^{\mathrm{src}} + \sqrt{1 - \alpha_t}\epsilon$ +4: for $\mathrm{i} = 1,\dots ,\mathrm{N}$ do +5: $\epsilon_{\phi}^{\mathrm{src}}\gets (1 + \omega)\epsilon_{\phi}(\mathbf{z}_{t}^{\mathrm{src}},y^{\mathrm{src}},t) - \omega \epsilon_{\phi}(\mathbf{z}_{t}^{\mathrm{src}},\emptyset ,t)$ +6: $\mathbf{z}_{0|t}^{\mathrm{src}}\gets \frac{1}{\sqrt{\alpha_t}} (\mathbf{z}_t^{\mathrm{src}} - \sqrt{1 - \alpha_t}\epsilon_\phi^{\mathrm{src}})$ +7: $\mathcal{L}_{\mathrm{FPR}}\gets d(\mathbf{z}_{0|t}^{\mathrm{src}},\mathbf{z}^{\mathrm{src}})$ +8: $\mathbf{z}_t^{\mathrm{src}}\gets \mathbf{z}_t^{\mathrm{src}} - \lambda \nabla_{\mathbf{z}_t^{\mathrm{src}}}\mathcal{L}_{\mathrm{FPR}}$ +9: end for +10: $\epsilon^{*}\gets \frac{1}{\sqrt{1 - \alpha_{t}}} (\mathbf{z}_{t}^{\mathrm{src}} - \sqrt{\alpha_{t}}\mathbf{z}^{\mathrm{src}})$ +11: return $\epsilon^{*}$ + +Editing with guided noise. Thanks to the proposed FPR, the optimized source latent $\mathbf{z}_t^{\mathrm{sfc*}}$ containing the source's identity can be obtained. Then, the guided noise $\epsilon^{*}$ is extracted as follows: + +$$ +\epsilon^ {*} = \frac {1}{\sqrt {1 - \alpha_ {t}}} \left(\mathbf {z} _ {t} ^ {\mathrm {s r c} *} - \sqrt {\alpha_ {t}} \mathbf {z} ^ {\mathrm {s r c}}\right). \tag {8} +$$ + +$\epsilon^{*}$ is utilized to produce the stochastic latent $\mathbf{z}_t^{\mathrm{trg}*}$ by applying the forward diffusion process to the target image $\mathbf{z}^{\mathrm{trg}}$ . With $\mathbf{z}_t^{\mathrm{src}*}$ and $\mathbf{z}_t^{\mathrm{trg}*}$ , the updated direction is given by: + +$$ +\nabla_ {\theta} \mathcal {L} _ {\mathrm {I D S}} = \mathbb {E} _ {t, \epsilon} \left[ \left(\epsilon_ {\phi} ^ {\omega} \left(\mathbf {z} _ {t} ^ {\operatorname {t r g} *}, y ^ {\operatorname {t r g}}, t\right) - \epsilon_ {\phi} ^ {\omega} \left(\mathbf {z} _ {t} ^ {\operatorname {s r c} *}, y ^ {\operatorname {s r c}}, t\right)\right) \frac {\partial \mathbf {z} ^ {\operatorname {t r g}}}{\partial \theta} \right]. \tag {9} +$$ + +It is worth noting that $\epsilon^{*}$ guides the appropriate gradients for editing while conserving the source's identity. In contrast to DDS, the proposed IDS perfectly reconstructs the source from the edited result $\mathbf{z}^{\mathrm{trg}}$ , as shown in the third row of Fig. 3. This confirms that the correct score and the corresponding injection noise can preserve the identity without further consideration of mutual information. The flowchart of our IDS is illustrated in Fig. 2. + +# 5. Results + +We evaluate our method through editing experiments conducted on two experiments. In Sec. 5.1, we perform a com + +parison on image-to-image editing across several datasets. In Sec. 5.2, we extend our evaluation to editable Neural Radiance Fields (NeRF) [16]. + +# 5.1. Text-guided image editing + +Baselines. To evaluate our method, we conduct comparative experiments against four state-of-the-art image editing models: Prompt-to-Prompt (P2P) [7], Plug-and-Play (PNP) [26], DDS [6], and CDS [17]. The implementations of the baselines are carried out by referencing the official source code for each method. More details are provided in Supplementary Materials. + +Qualitative Results. We present the qualitative results comparing our method with the baselines in Fig. 4. Prompt-to-Prompt (P2P) [7] performs image editing after applying DDIM inversion [4, 24] to the source image, leading to disregarding the structural components of the source image and following the target prompt excessively. Plug-and-Play (PnP) [26] has limitations in object recognition, as seen in the fourth row of Fig. 4. The third row of Fig. 4 demonstrates that DDS [6] and CDS [17] exhibited limitations, particularly in preserving the structural characteristics of the source image. In contrast, our method successfully edits the image while preserving the structural integrity of the source image. + +Quantitative Results. To measure the identity-preserving performance, we utilize two datasets. First, we collect 250 cat images from the LAION 5B dataset [23] based on [17] for Cat-to-Others task and measure Intersection over Union (IoU). Second, we gather 28 images from the InstructPix2Pix (IP2P) dataset [2], which contains the pairs of source and target images and corresponding prompts and calculate the background Peak-Signal-to-Noise-Ratio (PSNR). Details of the metrics are provided in Supplementary Materials. In addition, we use the LPIPS score [29] for each experiment to quantify the similarity between source and target images. The results are presented in Tab. 1. Our method consistently achieves the lowest LPIPS score across all datasets, indicating that it best preserves the structural semantics of the source images. + +For user evaluation, we present 35 comparison sets for four baselines and our method, gathering responses from 47 participants. Participants are asked to choose the most appropriate image for the following three questions: 1. Which + +
cat2pigcat2squirrelIp2p
MetricIoU (↑)LPIPS (↓)IoU (↑)LPIPS (↓)PSNR (↑)LPIPS (↓)
P2P [7]0.580.420.520.4620.880.47
PnP [26]0.550.520.530.5223.810.39
DDS [6]0.690.280.650.3026.020.24
CDS [17]0.720.250.710.2627.350.21
IDS (Ours)0.740.220.710.2429.250.19
+ +Table 1. Quantitative results for image editing. LPIPS [29] and IoU was measured on LAION 5B [23], while LPIPS and background PSNR was measured on InstructPix2Pix [2]. + +![](images/8829e7eac60bf87688e6d33af64d90f86b791ee186d114ffb93e1f244f670a6a.jpg) +Figure 4. Qualitative results of InstructPix2Pix dataset [2]. Our method successfully edits the image aligning with the target text prompt while preserving the structural integrity of the source image. + +image best fits the text condition? 2. Which image best preserves the structural information of the original image? + +
User Preference Rate (%)GPT score [19]
MetricText (↑)Preserving (↑)Quality (↑)Text (↑)Preserving (↑)Quality (↑)
P2P [7]11.134.808.095.665.375.77
PnP [26]7.727.176.936.546.776.74
DDS [6]20.3010.8216.237.607.517.37
CDS [17]17.0216.7217.088.268.008.09
IDS (Ours)43.8360.4951.678.979.008.80
+ +Table 2. User study and GPT scores [19] show that our method achieved the highest scores across all questions for image editing. + +3. Which image has the best quality for text-based image editing? Additionally, we measure the GPT score using the Dreambench++ [19] method, which generates human-aligned assessments for the same questions by refining the scoring into ten distinct levels. As shown in Tab. 2, our method receives the highest ratings for all questions. + +# 5.2. Editing NeRF + +We conduct experiments involving 3D rendering of edited images to demonstrate the effectiveness of our method in + +![](images/19031eeb194960ee28999139602c6543b87400b6b7058ba872122eecd641d5ae.jpg) +Figure 5. Qualitative results on Synthetic $360^{\circ}$ and LLFF datasets. IDS outperforms the baselines by preserving the structural consistency of the source image and maintaining the integrity of regions that should remain unchanged, while precisely editing only the areas specified by the target prompt. Furthermore, comparisons of the depth map results also highlight the superior consistency of our method over other baseline models. + +maintaining structural consistency. This approach is particularly relevant as consistency has an even greater impact on outcomes in 3D environments. + +Datasets. We evaluated our method on widely used NeRF datasets: Synthetic NeRF [16] and LLFF [15]. Since NeRF datasets have no given pairs of source and target prompts, we manually composed image descriptions. + +Qualitative Results. Fig. 5 illustrates the qualitative results of our method compared with NeRF editing baselines. In the first row, the target prompt specifies a precise part of the image for fine-grained editing. DDS [6] and CDS [17] fail to differentiate and edit the specific area. At the same time, our method accurately identifies the region indicated by the target prompt in the image and performs detailed editing exclusively on that part. The second row demonstrates a scenario in which the target prompt is designed to edit the mood of the image. Our approach adjusts the colors associated with "autumn" and "leaves" throughout the image while maintaining consistency in the "trunk" whereas DDS and CDS also changed the "trunk". In terms of depth maps, our method generates clean depth maps with minimal noise + +after image editing, whereas DDS and CDS introduce noticeable noise into the depth maps. + +
MetricCLIP [21] (↑)User Preference Rate (%)
Text (↑)Preserving (↑)Quality (↑)
DDS [6]0.159636.8828.3732.62
CDS [17]0.159722.7023.4021.28
IDS (Ours)0.162640.4248.2346.10
+ +Table 3. Quantitative results of NeRF editing with respect to CLIP score and User Preference Rate. IDS demonstrates superior quantitative performance compared to the baselines. + +Quantitative Results. Based on edited images, we performed 3D rendering and subsequently conducted quantitative evaluations provided in Tab. 3. To assess whether the edited 3D images are precisely aligned with the target prompts, we measured the CLIP [21] scores at $200\mathrm{k}$ iterations of training on the LLFF dataset. We additionally present a user evaluation conducted under the same setup in Sec. 5.1. Consistent with the trends observed in the qualitative results, our method demonstrates superior performance in the quantitative evaluations compared to other baselines. + +![](images/7b5891dbf5e397e1ca07b53b8994a2ac0a5308f781b9eead31346704728e805b.jpg) +"A photo of shark" $\rightarrow$ "A photo of dolphin" + +![](images/fe19d7b84027058dec2d3b1a9a25b31d8ffdc7676f00da67c3d8b1e8ee24a9f0.jpg) +Figure 6. Ablation study on scale $\lambda$ . To show the effect of scale more extremely, the number of iterations of FPR is set as 1. The result of DDS is the same as $\lambda = 0$ since it means no update for source identity. +"The green leaves" $\rightarrow$ "Yellow and red leaves in autumn" +Figure 7. Optimization steps. Our method effectively preserves the consistency of the source image, even as the number of iteration steps increases up to 400. + +
optim stepFPR iterLPIPS(↓)CLIP(↑)time (sec/img)Memory (GB)
DDS200-0.2400.29322.456.27
CDS200-0.2100.28759.318.83
IDS20010.1990.28550.808.63
30.1900.277107.77
10030.1650.26554.04
1500.1800.27281.25
+ +# 6. Discussions + +# 6.1. Ablation studies on FPR + +FPR iteration $N$ . We conduct experiments on FPR iterations $N$ to evaluate its impact and determine the optimal iteration count. Although performing just one iteration of FPR is sufficient to preserve the source identity, as shown in lower LPIPS score than baselines of Tab. 4, we set $N = 3$ to emphasize the purpose of our method. + +Scale $\lambda$ . The scaling factor $\lambda$ of FPR determines how much information of source latent $\mathbf{z}^{\mathrm{src}}$ is kept. As shown in Fig. 6, increasing the scale preserves the attributes of the source image, resulting in more successful editing when it is hard to translate due to the structural mismatch between the source and the target prompt. + +# 6.2. Optimization steps + +To show that our method can prevent error accumulation during translation, we set the experiment to extend the num + +ber of optimization steps from 200 to 400. In the results of DDS and CDS, there is color boosting or loss of details due to the cumulated error. In contrast, IDS maintains the characteristics of the original images, such as the color of the pumpkin in the first row of Fig. 7 and the shape of the leaf in the second row of Fig. 7, better than other methods. + +# 7. Limitation + +The proposed IDS demonstrates outstanding performance across evaluation metrics assessing consistency between source and target images. However, during FPR process, IDS relies solely on information from the source $\{\mathbf{z}^{\mathrm{src}}, y^{\mathrm{src}}\}$ without incorporating target-side information. This results in comparatively lower CLIP scores [21] than other baselines (Tab. 5) and failure cases for more complex translations (Fig. 8). In addition, our method requires additional computational overhead (Tab. 4) since FPR is applied to each optimization iteration. Detailed discussion about our limitation is provided in Sec. G of supplementary. Our future direction will explore changing the score conditioned by the target prompt $y^{\mathrm{trg}}$ , leading to a better alignment with $y^{\mathrm{trg}}$ . + +Table 4. Computational complexity on 28 images of InstructPix2Pix [2] for various settings. Lower LPIPS and higher CLIP scores mean better quality. + +
P2P [7]PnP [26]DDS [6]CDS [17]IDS (Ours)
cat2lion0.290.210.300.290.29
cat2dog0.270.260.270.270.26
Ip2p0.280.300.290.290.28
+ +Table 5. Limitation of IDS with respect to CLIP score[21] for image editing on LAION 5B [23] and InstructPix2Pix [2]. + +![](images/3414560104d3a705bce3ba19d768d8f5bbee176d715cb4fab571847e93062713.jpg) +"Photo free night, house, aurora" → "... with two dogs" +Figure 8. Failure case for complex text prompt. + +# 8. Conclusion + +We proposed a new distillation sampling method using a fixed-point regularization which aligns the text-conditioned score towards identity-preserved manifolds. The proposed fixed-point regularization preserves the source's identity by re-projecting the intermediate score status on posterior means. In this manner, corrected noises guide a gradient of distilled score toward identity-consistent manifolds. Owing to self-correction by a fixed-point iterator and guided injection noise, the proposed identity-preserving distillation sampling provides clear and unambiguous representations corresponding to the given prompts in text-guided image editing and editable neural radiance field (NeRF). Furthermore, our model can be utilized as a universal module in addition to the existing score-sampling processes. + +Acknowledgement This work was supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (RS-2024-00335741, RS-2024-00357197). + +# References + +[1] Donghoon Ahn, Hyoungwon Cho, Jaewon Min, Wooseok Jang, Jungwoo Kim, SeonHwa Kim, Hyun Hee Park, Kyong Hwan Jin, and Seungryong Kim. Self-rectifying diffusion sampling with perturbed-attention guidance. arXiv preprint arXiv:2403.17377, 2024. 1 +[2] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 1, 2, 5, 6, 8 +[3] Hyungjin Chung, Jeongsol Kim, Michael Thompson McCann, Marc Louis Klasky, and Jong Chul Ye. Diffusion posterior sampling for general noisy inverse problems. In The Eleventh International Conference on Learning Representations, 2023. 2 +[4] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 1, 5 +[5] Bradley Efron. Tweedie's formula and selection bias. Journal of the American Statistical Association, 106(496):1602-1614, 2011. 2 +[6] Amir Hertz, Kfir Aberman, and Daniel Cohen-Or. Delta denoising score. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2328-2337, 2023. 1, 2, 3, 4, 5, 6, 7, 8 +[7] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-or. Prompt-to-prompt image editing with cross-attention control. In The Eleventh International Conference on Learning Representations, 2023. 1, 2, 3, 5, 6, 8 +[8] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 1, 3 +[9] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 3 +[10] Susung Hong, Gyuseong Lee, Wooseok Jang, and Seungryong Kim. Improving sample quality of diffusion models using self-attention guidance. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7462–7471, 2023. 1 +[11] Oren Katzir, Or Patashnik, Daniel Cohen-Or, and Dani Lischinski. Noise-free score distillation. arXiv preprint arXiv:2310.17590, 2023. 3 +[12] Juil Koo, Chanho Park, and Minhyuk Sung. Posterior distillation sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13352-13361, 2024. 2 +[13] Barak Meiri, Dvir Samuel, Nir Darshan, Gal Chechik, Shai Avidan, and Rami Ben-Ari. Fixed-point inversion for text- + +to-image diffusion models. arXiv preprint arXiv:2312.12540, 2023. 4 +[14] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations, 2022. 1, 2, 3 +[15] Ben Mildenhall, Pratul P Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (ToG), 38(4):1-14, 2019. 7 +[16] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 5, 7 +[17] Hyelin Nam, Gihyun Kwon, Geon Yeong Park, and Jong Chul Ye. Contrastive denoising score for text-guided latent diffusion image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9192-9201, 2024. 2, 5, 6, 7, 8 +[18] Neal Parikh, Stephen Boyd, et al. Proximal algorithms. Foundations and trends® in Optimization, 1(3):127-239, 2014. 3 +[19] Yuang Peng, Yuxin Cui, Haomiao Tang, Zekun Qi, Runpei Dong, Jing Bai, Chunrui Han, Zheng Ge, Xiangyu Zhang, and Shu-Tao Xia. Dreambench++: A human-aligned benchmark for personalized image generation. arXiv preprint arXiv:2406.16855, 2024. 6 +[20] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 1, 3 +[21] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 7, 8 +[22] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 3 +[23] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 5, 8 +[24] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3, 5 +[25] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 1, 3 + +[26] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023. 1, 2, 3, 5, 6, 8 +[27] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems, 36, 2024. 3 +[28] Zike Wu, Pan Zhou, Xuanyu Yi, Xiaoding Yuan, and Hanwang Zhang. Consistent3d: Towards consistent high-fidelity text-to-3d generation with deterministic sampling prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9892–9902, 2024. 3 +[29] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5 +[30] Zhizhuo Zhou and Shubham Tulsiani. Sparsefusion: Distilling view-conditioned diffusion for 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12588-12597, 2023. 3 \ No newline at end of file diff --git a/2025/Identity-preserving Distillation Sampling by Fixed-Point Iterator/images.zip b/2025/Identity-preserving Distillation Sampling by Fixed-Point Iterator/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c4c81d89d986ee69eb9a6cf6969984c5979bd784 --- /dev/null +++ b/2025/Identity-preserving Distillation Sampling by Fixed-Point Iterator/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a3c6634a7685f8ab96959784a56c999744e07659cd0edc96e1577a03eb3d5060 +size 981038 diff --git a/2025/Identity-preserving Distillation Sampling by Fixed-Point Iterator/layout.json b/2025/Identity-preserving Distillation Sampling by Fixed-Point Iterator/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b93efa23190e09967cfc38074252cda599be8472 --- /dev/null +++ b/2025/Identity-preserving Distillation Sampling by Fixed-Point Iterator/layout.json @@ -0,0 +1,9175 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 105, + 103, + 506, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 103, + 506, + 121 + ], + "spans": [ + { + "bbox": [ + 105, + 103, + 506, + 121 + ], + "type": "text", + "content": "Identity-preserving Distillation Sampling by Fixed-Point Iterator" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 141, + 505, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 141, + 505, + 185 + ], + "spans": [ + { + "bbox": [ + 104, + 141, + 505, + 185 + ], + "type": "text", + "content": "SeonHwa Kim" + }, + { + "bbox": [ + 104, + 141, + 505, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 104, + 141, + 505, + 185 + ], + "type": "text", + "content": " Jiwon Kim" + }, + { + "bbox": [ + 104, + 141, + 505, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 104, + 141, + 505, + 185 + ], + "type": "text", + "content": " Soobin Park" + }, + { + "bbox": [ + 104, + 141, + 505, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 104, + 141, + 505, + 185 + ], + "type": "text", + "content": " Donghoon Ahn" + }, + { + "bbox": [ + 104, + 141, + 505, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 104, + 141, + 505, + 185 + ], + "type": "text", + "content": " Jiwon Kang" + }, + { + "bbox": [ + 104, + 141, + 505, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 104, + 141, + 505, + 185 + ], + "type": "text", + "content": " Seungryong Kim" + }, + { + "bbox": [ + 104, + 141, + 505, + 185 + ], + "type": "inline_equation", + "content": "^{3*}" + }, + { + "bbox": [ + 104, + 141, + 505, + 185 + ], + "type": "text", + "content": " Kyong Hwan Jin" + }, + { + "bbox": [ + 104, + 141, + 505, + 185 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 104, + 141, + 505, + 185 + ], + "type": "text", + "content": " Eunju Cha" + }, + { + "bbox": [ + 104, + 141, + 505, + 185 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 104, + 141, + 505, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 104, + 141, + 505, + 185 + ], + "type": "text", + "content": "Korea University " + }, + { + "bbox": [ + 104, + 141, + 505, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 104, + 141, + 505, + 185 + ], + "type": "text", + "content": "Sookmyung Women's University " + }, + { + "bbox": [ + 104, + 141, + 505, + 185 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 104, + 141, + 505, + 185 + ], + "type": "text", + "content": "KAIST" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 129, + 186, + 479, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 186, + 479, + 213 + ], + "spans": [ + { + "bbox": [ + 129, + 186, + 479, + 213 + ], + "type": "text", + "content": "{sunkim0062, jwonkim, dhahn99, jiwon7258, kyong_jin}@korea.ac.kr, {psb1219j, eunju.cha}@sookmyung.ac.kr, seungryong.kim@kaist.ac.kr" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "spans": [ + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 267, + 297, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 267, + 297, + 542 + ], + "spans": [ + { + "bbox": [ + 54, + 267, + 297, + 542 + ], + "type": "text", + "content": "Score distillation sampling (SDS) demonstrates a powerful capability for text-conditioned 2D image and 3D object generation by distilling the knowledge from learned score functions. However, SDS often suffers from blurriness caused by noisy gradients. When SDS meets the image editing, such degradations can be reduced by adjusting bias shifts using reference pairs, but the de-biasing techniques are still corrupted by erroneous gradients. To this end, we introduce Identity-preserving Distillation Sampling (IDS), which compensates for the gradient leading to undesired changes in the results. Based on the analysis that these errors come from the text-conditioned scores, a new regularization technique, called fixed-point iterative regularization (FPR), is proposed to modify the score itself, driving the preservation of the identity even including poses and structures. Thanks to a self-correction by FPR, the proposed method provides clear and unambiguous representations corresponding to the given prompts in image-to-image editing and editable neural radiance field (NeRF). The structural consistency between the source and the edited data is obviously maintained compared to other state-of-the-art methods. Our code is https://github.com/shhh0620/IDS" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 568, + 135, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 568, + 135, + 580 + ], + "spans": [ + { + "bbox": [ + 55, + 568, + 135, + 580 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 582, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 693 + ], + "type": "text", + "content": "Diffusion models [4, 8, 9, 22, 25] have shown powerful representations on text-to-image (T2I) generative tasks. With the advance of classifier guidance (CG) and classifier-free guidance (CFG) paradigms [1, 4, 8, 10], diffusion models improve the quality of generated samples [9, 25]. Such high-quality image generators can be easily extended to image editing by simply modifying forward/reverse iterations [14], applying CFG with a target prompt [2, 7] or interchanging attention layers [26]." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 315, + 239, + 555, + 460 + ], + "blocks": [ + { + "bbox": [ + 315, + 239, + 555, + 460 + ], + "lines": [ + { + "bbox": [ + 315, + 239, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 315, + 239, + 555, + 460 + ], + "type": "image", + "image_path": "6a18606f6793dfc1980d978ecb2ab4577f6e06385b0fef0d0185ce334224edb2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 468, + 556, + 536 + ], + "lines": [ + { + "bbox": [ + 313, + 468, + 556, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 556, + 536 + ], + "type": "text", + "content": "Figure 1. Trace of guided updating from source to target images using delta denoising score (DDS) and identity-preserving distillation sampling (IDS). DDS moves a gradient of score function toward " + }, + { + "bbox": [ + 313, + 468, + 556, + 536 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\mathbf{z}}" + }, + { + "bbox": [ + 313, + 468, + 556, + 536 + ], + "type": "text", + "content": " manifold directed by stochastic direction " + }, + { + "bbox": [ + 313, + 468, + 556, + 536 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 468, + 556, + 536 + ], + "type": "text", + "content": ". In contrast, IDS moves a gradient with a corrected direction by a fixed-point regularization." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 557, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 556, + 715 + ], + "type": "text", + "content": "Recently, Delta Denoising Score (DDS) [6] is proposed to edit a source image by distilling the rich generative prior of T2I diffusion models. It is based on the analysis of Score Distillation Sampling (SDS) [20], originally developed to optimize a parametric generator such as Neural Radiance Fields (NeRFs) [16] by exploiting the learned score of the diffusion models. Even though SDS offers remarkable performance in synthesizing 3D scenes, noisy gradients from stochastic perturbations lead to significantly over-saturated results that faithfully follow the given text prompts. In the context of image editing, text prompts do not often include information about the identity of the source image, such as the background, the object's pose, or the structure of the" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 145, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 145, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 145, + 713 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11115" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 95, + 70, + 517, + 259 + ], + "blocks": [ + { + "bbox": [ + 95, + 70, + 517, + 259 + ], + "lines": [ + { + "bbox": [ + 95, + 70, + 517, + 259 + ], + "spans": [ + { + "bbox": [ + 95, + 70, + 517, + 259 + ], + "type": "image", + "image_path": "738d635f30a3df04304cc174651668ddc18fbbd91b32a508accacc2c253472d9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 269, + 555, + 304 + ], + "lines": [ + { + "bbox": [ + 54, + 269, + 555, + 304 + ], + "spans": [ + { + "bbox": [ + 54, + 269, + 555, + 304 + ], + "type": "text", + "content": "Figure 2. Flowchart of IDS. The backbone of our algorithm employs DDS [6] framework to distill score function into a target image. Our fixed-point regularization (FPR) obtains a guided noise, " + }, + { + "bbox": [ + 54, + 269, + 555, + 304 + ], + "type": "inline_equation", + "content": "\\epsilon^{*}" + }, + { + "bbox": [ + 54, + 269, + 555, + 304 + ], + "type": "text", + "content": ", from iterative updates using posterior mean computed by Tweedie's formula. When distilling the score function to a target image, the guided noise is updated while maintaining the identity of the source." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 324, + 294, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 324, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 54, + 324, + 294, + 491 + ], + "type": "text", + "content": "content, which should be retained during updates. Thus, DDS is designed to resolve such blurriness by erasing gradients of non-text-aligned features from SDS gradients. There is no explicit procedure to preserve the source's identity in DDS updates because the fine gradient may provide the conserved identity. However, this cannot be guaranteed if many variations in the structure are possible, such as editing the image of a cat into a pig, as shown in Fig. 1. To alleviate this problem, Contrastive Denoising Score (CDS) [17] and Posterior Distillation Sampling (PDS) [12] are introduced to maximize the mutual information of the source image and edited image. Although such algorithms rely heavily on text prompts, the algorithms have yet to analyze the inherent error caused by text-conditioned scores." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 495, + 295, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 495, + 295, + 626 + ], + "spans": [ + { + "bbox": [ + 54, + 495, + 295, + 626 + ], + "type": "text", + "content": "To this end, we investigate the underlying meaning of text-conditioned score. The gradient maps the stochastic latent, generated by applying the forward diffusion process to the given image, to one of the possible images described by the prompt, including the original image. Simply, the score obtained from the latent of the source image ('source latent') and the source prompt can be a gradient to another image represented by the identical text. Based on this interpretation, the accumulation of misaligned directions causes the loss of the source's identity, leading to structural changes in the result with DDS, as shown in Fig. 1." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 630, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 630, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 630, + 295, + 715 + ], + "type": "text", + "content": "To address this issue, we propose a novel score distillation sampling to effectively preserve the identity of the original image by self-correcting the misaligned gradients, called Identity-preserving Distillation Sampling (IDS). The key insight is that if the score is precisely adjusted to the source image, the conditional expectation of the source image given the source latent contains meaningful information" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 324, + 555, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 555, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 555, + 479 + ], + "type": "text", + "content": "that should be preserved during the editing. This conditional expectation corresponds to the posterior mean computed by Tweedie's formula using the learned score [3, 5]. The source latent is iteratively updated to make the posterior mean similar to the source image. This procedure, named a fixed-point iterative regularization (FPR), results in the aligned score with the source that provides reliable gradients for editing, as illustrated in Fig. 1. Following IDS update is performed using guided noise extracted from the refined source latent, rather than random Gaussian noise. This further ensures the identity preservation. Our method demonstrated superior results compared with baselines in two tasks: editing images by prompts and editing NeRF." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 325, + 481, + 533, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 481, + 533, + 491 + ], + "spans": [ + { + "bbox": [ + 325, + 481, + 533, + 491 + ], + "type": "text", + "content": "In summary, our main contributions are as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 494, + 555, + 601 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 494, + 553, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 494, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 314, + 494, + 553, + 552 + ], + "type": "text", + "content": "- We obtain reliable gradients for the score distillation function by a fixed-point iterator with respect to posterior means. The iterator corrects the text-conditioned score, guiding SDS gradients toward reliable pre-trained manifolds." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 553, + 555, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 553, + 555, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 553, + 555, + 601 + ], + "type": "text", + "content": "- Our fixed-point regularization preserves the identities of sources such as structures and poses in edited targets for 2D and 3D editing. Such preservation is well demonstrated in qualitative and quantitative results." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 613, + 404, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 613, + 404, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 613, + 404, + 625 + ], + "type": "text", + "content": "2. Related works" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 634, + 510, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 634, + 510, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 634, + 510, + 647 + ], + "type": "text", + "content": "2.1. Image Editing with Diffusion Models" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": "With the great success of image generation using diffusion models, the pre-trained diffusion models have been recently employed for image editing tasks, demonstrating significant advancements in the quality and flexibility of generated edits [2, 7, 12, 14, 17, 26]. Stochastic Differential" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11116" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 300 + ], + "type": "text", + "content": "Editing (SDEdit) [14] is a pioneering work in which the source image was modified by adding noise and solving reverse stochastic differential equations. Thanks to the text-conditional Latent Diffusion Model (LDM), a.k.a. Stable Diffusion [22], text-driven editing approaches have been introduced. Specifically, the text embedding was injected through the cross-attention layer of the model for image editing and translation, while retaining the structure of the original image [7, 26]. The editing was further controlled by rescaling the attention of the specific word [7] or by manipulating the self-attention features [26]. These approaches provide greater control by balancing fidelity between the edited prompt and the source image without the need for model training, fine-tuning, additional data, or optimization. However, the current DDIM-based inversion [24] can lead to unsatisfactory reconstructions for real images, and the cross-attention bottleneck limits its effectiveness for broader edits. Crafting suitable prompts also remains challenging for complex compositions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 306, + 203, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 306, + 203, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 306, + 203, + 319 + ], + "type": "text", + "content": "2.2. Score distillation sampling" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 324, + 296, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 324, + 296, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 324, + 296, + 540 + ], + "type": "text", + "content": "Score Distillation Sampling (SDS) [20] enables text-driven 3D synthesis by leveraging probability density distillation loss to distill knowledge from 2D diffusion models, allowing high-quality 3D scene generation based on textual prompts without 3D training data. However, SDS has limitations, often producing oversaturated and overly smooth 3D models, and lacking diversity across initializations. To address these limitations of SDS, various models have been proposed based on exploiting multi-step denoising [30], a variation approach [27], negative conditioning [11], and ordinary differential equation trajectory [28]. To mitigate the limitation of noisy gradients in SDS, which hampers precise image editing, DDS [6] was introduced. By computing the delta between the derived gradient and the target pair, DDS effectively isolates and removes unwanted noise in the gradient direction. Despite these advancements, DDS still faces challenges in preserving the complete structural consistency of the source image's identity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 551, + 141, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 551, + 141, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 551, + 141, + 562 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 570, + 270, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 270, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 270, + 582 + ], + "type": "text", + "content": "3.1. Diffusion Model and Sampling Guidance" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 588, + 296, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 296, + 637 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 296, + 637 + ], + "type": "text", + "content": "Text-to-image diffusion models " + }, + { + "bbox": [ + 55, + 588, + 296, + 637 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}(\\cdot)" + }, + { + "bbox": [ + 55, + 588, + 296, + 637 + ], + "type": "text", + "content": " are based on diffusion probabilistic models (DPMs) [9, 22, 25]. The models are trained to estimate the denoising score when the original image " + }, + { + "bbox": [ + 55, + 588, + 296, + 637 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0" + }, + { + "bbox": [ + 55, + 588, + 296, + 637 + ], + "type": "text", + "content": " and the text condition " + }, + { + "bbox": [ + 55, + 588, + 296, + 637 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 588, + 296, + 637 + ], + "type": "text", + "content": " are given:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 107, + 643, + 243, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 643, + 243, + 658 + ], + "spans": [ + { + "bbox": [ + 107, + 643, + 243, + 658 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (\\phi) = \\mathbb {E} _ {t, \\epsilon} [ \\| \\epsilon_ {\\phi} (\\mathbf {z} _ {t}, y, t) - \\epsilon \\| _ {2} ^ {2} ],", + "image_path": "15dbd144b788b2c7fb27b1f1a047547f3f730f0eb1dd12da60efd22751bd7514.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 665, + 296, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 665, + 296, + 689 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0,\\mathbf{I})" + }, + { + "bbox": [ + 55, + 665, + 296, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 665, + 296, + 689 + ], + "type": "inline_equation", + "content": "t\\sim \\mathcal{U}(0,1)" + }, + { + "bbox": [ + 55, + 665, + 296, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t" + }, + { + "bbox": [ + 55, + 665, + 296, + 689 + ], + "type": "text", + "content": " refers to the stochastic latent of " + }, + { + "bbox": [ + 55, + 665, + 296, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0" + }, + { + "bbox": [ + 55, + 665, + 296, + 689 + ], + "type": "text", + "content": " via the forward diffusion process as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 121, + 700, + 296, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 700, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 121, + 700, + 296, + 715 + ], + "type": "interline_equation", + "content": "\\mathbf {z} _ {t} = \\sqrt {\\alpha_ {t}} \\mathbf {z} _ {0} + \\sqrt {1 - \\alpha_ {t}} \\epsilon , \\tag {1}", + "image_path": "9aff40cd506d3be03c3008bfa0d4a3d951b72ac12fd1b57659999775ee11d598.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": " is noise schedule. With the trained " + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}(\\cdot)" + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": ", high-quality samples can be generated using the classifier-free guidance (CFG) [8] by subtracting unconditioned denoising score from the conditioned score with guidance scale " + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 326, + 131, + 555, + 144 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 131, + 555, + 144 + ], + "spans": [ + { + "bbox": [ + 326, + 131, + 555, + 144 + ], + "type": "interline_equation", + "content": "\\epsilon_ {\\phi} ^ {\\omega} \\left(\\mathbf {z} _ {t}, y, t\\right) = (1 + \\omega) \\epsilon_ {\\phi} \\left(\\mathbf {z} _ {t}, y, t\\right) - \\omega \\epsilon_ {\\phi} \\left(\\mathbf {z} _ {t}, \\varnothing , t\\right). \\tag {2}", + "image_path": "ffa841a692c4ebeb1162d5ef17e97f044db43e8348ab23936887fa1437672e0b.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 154, + 495, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 154, + 495, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 154, + 495, + 167 + ], + "type": "text", + "content": "3.2. Score Distillation Sampling (SDS)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 172, + 555, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 172, + 555, + 217 + ], + "spans": [ + { + "bbox": [ + 313, + 172, + 555, + 217 + ], + "type": "text", + "content": "With pretrained text-to-image diffusion models " + }, + { + "bbox": [ + 313, + 172, + 555, + 217 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}(\\cdot)" + }, + { + "bbox": [ + 313, + 172, + 555, + 217 + ], + "type": "text", + "content": ", SDS [20] synthesizes 3D data " + }, + { + "bbox": [ + 313, + 172, + 555, + 217 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 313, + 172, + 555, + 217 + ], + "type": "text", + "content": " for a given text prompt " + }, + { + "bbox": [ + 313, + 172, + 555, + 217 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 172, + 555, + 217 + ], + "type": "text", + "content": " by optimizing the differentiable rendering function paretrized by " + }, + { + "bbox": [ + 313, + 172, + 555, + 217 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 172, + 555, + 217 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 172, + 555, + 217 + ], + "type": "inline_equation", + "content": "\\mathbf{z} = g(\\theta)" + }, + { + "bbox": [ + 313, + 172, + 555, + 217 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 339, + 225, + 555, + 251 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 225, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 339, + 225, + 555, + 251 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\theta} \\mathcal {L} _ {\\mathrm {S D S}} (\\mathbf {z}, y) = \\mathbb {E} _ {t, \\epsilon} \\left[ \\omega (t) \\left(\\epsilon_ {\\phi} ^ {\\omega} \\left(\\mathbf {z} _ {t}, y, t\\right) - \\epsilon\\right) \\frac {\\partial \\mathbf {z}}{\\partial \\theta} \\right]. \\tag {3}", + "image_path": "1165e613b2884d2898280c7fadc95c3ab3a5622e6ea52a40b084e936cd753e2e.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 258, + 555, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 555, + 354 + ], + "type": "text", + "content": "The optimized parameters " + }, + { + "bbox": [ + 313, + 258, + 555, + 354 + ], + "type": "inline_equation", + "content": "\\theta^{*}" + }, + { + "bbox": [ + 313, + 258, + 555, + 354 + ], + "type": "text", + "content": " provide the text-conditioned 3D volume that follows the diffusion prior [20]. However, a single text prompt " + }, + { + "bbox": [ + 313, + 258, + 555, + 354 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 258, + 555, + 354 + ], + "type": "text", + "content": " can refer to many different 3D volumes, each with diverse backgrounds or structural details of the object. Therefore, an inherent limitation of SDS [20] is that the score conditioned by the prompt " + }, + { + "bbox": [ + 313, + 258, + 555, + 354 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 258, + 555, + 354 + ], + "type": "text", + "content": " does not always provide the diffusion prior to the identical object during the optimization process, leading to blurry and unclear results." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 362, + 471, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 362, + 471, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 362, + 471, + 376 + ], + "type": "text", + "content": "3.3. Delta Denoising Score (DDS)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 380, + 554, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 554, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 554, + 449 + ], + "type": "text", + "content": "DDS [6] is proposed to synthesize the image " + }, + { + "bbox": [ + 313, + 380, + 554, + 449 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\mathrm{trg}}" + }, + { + "bbox": [ + 313, + 380, + 554, + 449 + ], + "type": "text", + "content": " from the given source image " + }, + { + "bbox": [ + 313, + 380, + 554, + 449 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\mathrm{src}}" + }, + { + "bbox": [ + 313, + 380, + 554, + 449 + ], + "type": "text", + "content": " and its corresponding prompt " + }, + { + "bbox": [ + 313, + 380, + 554, + 449 + ], + "type": "inline_equation", + "content": "y^{\\mathrm{src}}" + }, + { + "bbox": [ + 313, + 380, + 554, + 449 + ], + "type": "text", + "content": ", which is aligned to the target prompt " + }, + { + "bbox": [ + 313, + 380, + 554, + 449 + ], + "type": "inline_equation", + "content": "y^{\\mathrm{trg}}" + }, + { + "bbox": [ + 313, + 380, + 554, + 449 + ], + "type": "text", + "content": ". Based on the insight that the gradient should be zero if " + }, + { + "bbox": [ + 313, + 380, + 554, + 449 + ], + "type": "inline_equation", + "content": "y^{\\mathrm{trg}}" + }, + { + "bbox": [ + 313, + 380, + 554, + 449 + ], + "type": "text", + "content": " matches " + }, + { + "bbox": [ + 313, + 380, + 554, + 449 + ], + "type": "inline_equation", + "content": "y^{\\mathrm{src}}" + }, + { + "bbox": [ + 313, + 380, + 554, + 449 + ], + "type": "text", + "content": ", DDS minimizes the identity change of " + }, + { + "bbox": [ + 313, + 380, + 554, + 449 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\mathrm{src}}" + }, + { + "bbox": [ + 313, + 380, + 554, + 449 + ], + "type": "text", + "content": " by simple replacing " + }, + { + "bbox": [ + 313, + 380, + 554, + 449 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 380, + 554, + 449 + ], + "type": "text", + "content": " in (3) with the score " + }, + { + "bbox": [ + 313, + 380, + 554, + 449 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}^{\\omega}(\\mathbf{z}_{t}^{\\mathrm{src}}, y^{\\mathrm{src}}, t)" + }, + { + "bbox": [ + 313, + 380, + 554, + 449 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 319, + 465, + 555, + 490 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 465, + 555, + 490 + ], + "spans": [ + { + "bbox": [ + 319, + 465, + 555, + 490 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\theta} \\mathcal {L} _ {\\mathrm {D D S}} = \\mathbb {E} _ {t, \\epsilon} \\left[ \\left(\\epsilon_ {\\phi} ^ {\\omega} \\left(\\mathbf {z} _ {t} ^ {\\operatorname {t r g}}, y ^ {\\operatorname {t r g}}, t\\right) - \\epsilon_ {\\phi} ^ {\\omega} \\left(\\mathbf {z} _ {t} ^ {\\operatorname {s r c}}, y ^ {\\operatorname {s r c}}, t\\right)\\right) \\frac {\\partial \\mathbf {z} ^ {\\operatorname {t r g}}}{\\partial \\theta} \\right]. \\tag {4}", + "image_path": "6e6795ae7a13a835234c01852ba6f56c88a57c78fc272ad389670c5f82a7ae8e.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 498, + 555, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 624 + ], + "type": "text", + "content": "For simplicity, we denote " + }, + { + "bbox": [ + 313, + 498, + 555, + 624 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}^{\\mathrm{trg}} = \\epsilon_{\\phi}^{\\omega}(\\mathbf{z}^{\\mathrm{trg}},y^{\\mathrm{trg}},t)" + }, + { + "bbox": [ + 313, + 498, + 555, + 624 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 498, + 555, + 624 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}^{\\mathrm{src}} = \\epsilon_{\\phi}^{\\omega}(\\mathbf{z}^{\\mathrm{src}},y^{\\mathrm{src}},t)" + }, + { + "bbox": [ + 313, + 498, + 555, + 624 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 313, + 498, + 555, + 624 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}^{\\mathrm{trg}}" + }, + { + "bbox": [ + 313, + 498, + 555, + 624 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 498, + 555, + 624 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}^{\\mathrm{src}}" + }, + { + "bbox": [ + 313, + 498, + 555, + 624 + ], + "type": "text", + "content": " can be interpreted as the gradients representing the direction from " + }, + { + "bbox": [ + 313, + 498, + 555, + 624 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t^{\\mathrm{trg}}" + }, + { + "bbox": [ + 313, + 498, + 555, + 624 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 498, + 555, + 624 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\mathrm{trg}}" + }, + { + "bbox": [ + 313, + 498, + 555, + 624 + ], + "type": "text", + "content": " and the direction from " + }, + { + "bbox": [ + 313, + 498, + 555, + 624 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t^{\\mathrm{src}}" + }, + { + "bbox": [ + 313, + 498, + 555, + 624 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 498, + 555, + 624 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\mathrm{src}}" + }, + { + "bbox": [ + 313, + 498, + 555, + 624 + ], + "type": "text", + "content": ", respectively. " + }, + { + "bbox": [ + 313, + 498, + 555, + 624 + ], + "type": "inline_equation", + "content": "\\theta = \\mathbf{z}^{\\mathrm{trg}}" + }, + { + "bbox": [ + 313, + 498, + 555, + 624 + ], + "type": "text", + "content": " is thus gradually optimized along the direction from " + }, + { + "bbox": [ + 313, + 498, + 555, + 624 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\mathrm{src}}" + }, + { + "bbox": [ + 313, + 498, + 555, + 624 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 498, + 555, + 624 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\mathrm{trg}}" + }, + { + "bbox": [ + 313, + 498, + 555, + 624 + ], + "type": "text", + "content": ", as shown in Fig. 1. It is worth noting that the guidance of the update can be calculated at the same point " + }, + { + "bbox": [ + 313, + 498, + 555, + 624 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t^{\\mathrm{trg}}" + }, + { + "bbox": [ + 313, + 498, + 555, + 624 + ], + "type": "text", + "content": ", thanks to the shared " + }, + { + "bbox": [ + 313, + 498, + 555, + 624 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 498, + 555, + 624 + ], + "type": "text", + "content": ". However, the slight error in the gradient caused by the score " + }, + { + "bbox": [ + 313, + 498, + 555, + 624 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}^{\\mathrm{src}}" + }, + { + "bbox": [ + 313, + 498, + 555, + 624 + ], + "type": "text", + "content": " still leads to the incorrect direction for the optimization." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 633, + 435, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 633, + 435, + 645 + ], + "spans": [ + { + "bbox": [ + 313, + 633, + 435, + 645 + ], + "type": "text", + "content": "3.4. Fixed-point Iteration" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 652, + 554, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 652, + 554, + 698 + ], + "spans": [ + { + "bbox": [ + 313, + 652, + 554, + 698 + ], + "type": "text", + "content": "In numerical analysis, a fixed-point iteration [18] is an iterative method to find fixed points of a function " + }, + { + "bbox": [ + 313, + 652, + 554, + 698 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 652, + 554, + 698 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 652, + 554, + 698 + ], + "type": "inline_equation", + "content": "f(x) = x" + }, + { + "bbox": [ + 313, + 652, + 554, + 698 + ], + "type": "text", + "content": ". Given an initial point " + }, + { + "bbox": [ + 313, + 652, + 554, + 698 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 313, + 652, + 554, + 698 + ], + "type": "text", + "content": ", the iteration is defined as:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 365, + 701, + 502, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 701, + 502, + 714 + ], + "spans": [ + { + "bbox": [ + 365, + 701, + 502, + 714 + ], + "type": "interline_equation", + "content": "x _ {n + 1} = f (x _ {n}), \\quad n = 0, 1, 2, \\dots", + "image_path": "1e749ddc0628b0f5f6d42607c66b28df9b1aee20469f6f472d21d46dfbd4fab0.jpg" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "11117" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "type": "text", + "content": "Under appropriate conditions, this sequence converges to a fixed point. Thanks to its applicability to non-linear problems with low computational costs, fixed-point iteration is widely used in optimization, including applications in the context of diffusion models [13]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 143, + 111, + 156 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 143, + 111, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 143, + 111, + 156 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 163, + 296, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 163, + 296, + 248 + ], + "spans": [ + { + "bbox": [ + 55, + 163, + 296, + 248 + ], + "type": "text", + "content": "Given the source pair " + }, + { + "bbox": [ + 55, + 163, + 296, + 248 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{z}^{\\mathrm{src}}, y^{\\mathrm{src}}\\}" + }, + { + "bbox": [ + 55, + 163, + 296, + 248 + ], + "type": "text", + "content": ", the aim of our work is to provide an edited result " + }, + { + "bbox": [ + 55, + 163, + 296, + 248 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\mathrm{trg}}" + }, + { + "bbox": [ + 55, + 163, + 296, + 248 + ], + "type": "text", + "content": " that is aligned with " + }, + { + "bbox": [ + 55, + 163, + 296, + 248 + ], + "type": "inline_equation", + "content": "y^{\\mathrm{trg}}" + }, + { + "bbox": [ + 55, + 163, + 296, + 248 + ], + "type": "text", + "content": " while maintaining the source's identity. To this end, we introduce a novel approach called Identity-preserving Distillation Sampling (IDS), which (1) corrects the error of the gradient aligned with the text prompt by the fixed-point iterator and (2) provides the result " + }, + { + "bbox": [ + 55, + 163, + 296, + 248 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\mathrm{trg}}" + }, + { + "bbox": [ + 55, + 163, + 296, + 248 + ], + "type": "text", + "content": " using the guided noise." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 255, + 129, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 129, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 129, + 266 + ], + "type": "text", + "content": "4.1. Motivation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 273, + 296, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 273, + 296, + 341 + ], + "spans": [ + { + "bbox": [ + 55, + 273, + 296, + 341 + ], + "type": "text", + "content": "Analysis of the text-conditioned score. We first investigated how much identity of the given image " + }, + { + "bbox": [ + 55, + 273, + 296, + 341 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\mathrm{src}}" + }, + { + "bbox": [ + 55, + 273, + 296, + 341 + ], + "type": "text", + "content": " could be contained in the text-conditioned score " + }, + { + "bbox": [ + 55, + 273, + 296, + 341 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}^{\\mathrm{src}}" + }, + { + "bbox": [ + 55, + 273, + 296, + 341 + ], + "type": "text", + "content": ". To do this, we conducted the experiment to compare the original image " + }, + { + "bbox": [ + 55, + 273, + 296, + 341 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\mathrm{src}}" + }, + { + "bbox": [ + 55, + 273, + 296, + 341 + ], + "type": "text", + "content": " and the posterior mean " + }, + { + "bbox": [ + 55, + 273, + 296, + 341 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{0|t}^{\\mathrm{src}} = \\mathbb{E}[\\mathbf{z}^{\\mathrm{src}}|\\mathbf{z}_t^{\\mathrm{src}}]" + }, + { + "bbox": [ + 55, + 273, + 296, + 341 + ], + "type": "text", + "content": ", which is given by:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 110, + 342, + 295, + 364 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 342, + 295, + 364 + ], + "spans": [ + { + "bbox": [ + 110, + 342, + 295, + 364 + ], + "type": "interline_equation", + "content": "\\mathbf {z} _ {0 | t} ^ {\\mathrm {s r c}} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\mathbf {z} _ {t} ^ {\\mathrm {s r c}} - \\sqrt {1 - \\alpha_ {t}} \\epsilon_ {\\phi} ^ {\\mathrm {s r c}}\\right), \\tag {5}", + "image_path": "bee62ad204459aee70eccb255249b53168a04e7071d2b883319dffd6e9eda5f4.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 366, + 295, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 366, + 295, + 496 + ], + "spans": [ + { + "bbox": [ + 55, + 366, + 295, + 496 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 366, + 295, + 496 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t^{\\mathrm{src}}" + }, + { + "bbox": [ + 55, + 366, + 295, + 496 + ], + "type": "text", + "content": " denotes the source latent generated by (1). As shown in the first row of the supplementary Fig. S1, it is difficult to recognize the features of " + }, + { + "bbox": [ + 55, + 366, + 295, + 496 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\mathrm{src}}" + }, + { + "bbox": [ + 55, + 366, + 295, + 496 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 55, + 366, + 295, + 496 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{0|t}^{\\mathrm{src}}" + }, + { + "bbox": [ + 55, + 366, + 295, + 496 + ], + "type": "text", + "content": ", such as hairstyle, details of eyes, and background. This demonstrates that the score " + }, + { + "bbox": [ + 55, + 366, + 295, + 496 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}^{\\mathrm{src}}" + }, + { + "bbox": [ + 55, + 366, + 295, + 496 + ], + "type": "text", + "content": " is not exactly adjusted to the given image " + }, + { + "bbox": [ + 55, + 366, + 295, + 496 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\mathrm{src}}" + }, + { + "bbox": [ + 55, + 366, + 295, + 496 + ], + "type": "text", + "content": ". This deformation becomes more pronounced with increasing " + }, + { + "bbox": [ + 55, + 366, + 295, + 496 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 366, + 295, + 496 + ], + "type": "text", + "content": ". The experiment confirms that " + }, + { + "bbox": [ + 55, + 366, + 295, + 496 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}^{\\mathrm{src}}" + }, + { + "bbox": [ + 55, + 366, + 295, + 496 + ], + "type": "text", + "content": " may not be a precise guidance to the source image " + }, + { + "bbox": [ + 55, + 366, + 295, + 496 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\mathrm{src}}" + }, + { + "bbox": [ + 55, + 366, + 295, + 496 + ], + "type": "text", + "content": ". Therefore, the text-conditioned score " + }, + { + "bbox": [ + 55, + 366, + 295, + 496 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}^{\\mathrm{src}}" + }, + { + "bbox": [ + 55, + 366, + 295, + 496 + ], + "type": "text", + "content": " needs to be modified to maintain the identity of the source image " + }, + { + "bbox": [ + 55, + 366, + 295, + 496 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\mathrm{src}}" + }, + { + "bbox": [ + 55, + 366, + 295, + 496 + ], + "type": "text", + "content": " in the edited result " + }, + { + "bbox": [ + 55, + 366, + 295, + 496 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\mathrm{trg}}" + }, + { + "bbox": [ + 55, + 366, + 295, + 496 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "type": "text", + "content": "Accumulated error in DDS. The transformed image " + }, + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\mathrm{trg}}" + }, + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "type": "text", + "content": " can be converted back to the original image " + }, + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\mathrm{src}}" + }, + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "type": "text", + "content": " by reversing the set of " + }, + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "type": "text", + "content": " used to synthesize " + }, + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\mathrm{trg}}" + }, + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\mathrm{src}}" + }, + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "type": "text", + "content": " and swapping " + }, + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{z}^{\\mathrm{src}}, y^{\\mathrm{src}}\\}" + }, + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{z}^{\\mathrm{trg}}, y^{\\mathrm{trg}}\\}" + }, + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "type": "text", + "content": " to calculate the DDS loss in (4). If the guidance from " + }, + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\mathrm{src}}" + }, + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\mathrm{trg}}" + }, + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "type": "text", + "content": " is computed exactly, the perfect reconstruction can be achieved. Nevertheless, as can be seen from the second row in Fig. 3, DDS [6] fails to restore the original image " + }, + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\mathrm{src}}" + }, + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "type": "text", + "content": " from the edited image " + }, + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\mathrm{trg}}" + }, + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "type": "text", + "content": ", which implies that the direction from " + }, + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\mathrm{src}}" + }, + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\mathrm{trg}}" + }, + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "type": "text", + "content": " is calculated incorrectly. Based on our analysis, this error is because the text-conditioned score " + }, + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}^{\\mathrm{src}}" + }, + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "type": "text", + "content": " do not refer to the source " + }, + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\mathrm{src}}" + }, + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "type": "text", + "content": ", which can be explicitly expressed as the difference between the injected noise " + }, + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "type": "text", + "content": " and the score " + }, + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}^{\\mathrm{src}}" + }, + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "type": "text", + "content": ". While the optimization is being processed, the error inevitably accumulates, leading to the undesirable change to the structure and the pose. To address these issues, we investigated whether the guidance from " + }, + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\mathrm{src}}" + }, + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\mathrm{trg}}" + }, + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "type": "text", + "content": " can be properly provided while preserving the source's identity, when the timestep " + }, + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "type": "text", + "content": " is con" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 312, + 65, + 547, + 314 + ], + "blocks": [ + { + "bbox": [ + 312, + 65, + 547, + 314 + ], + "lines": [ + { + "bbox": [ + 312, + 65, + 547, + 314 + ], + "spans": [ + { + "bbox": [ + 312, + 65, + 547, + 314 + ], + "type": "image", + "image_path": "3aa8a2e8e325d12c6f49923f8ac4a7c927f92d57dab571a6b15aa14a2c0240e0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 316, + 555, + 372 + ], + "lines": [ + { + "bbox": [ + 313, + 316, + 555, + 372 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 555, + 372 + ], + "type": "text", + "content": "Figure 3. Accumulated error in DDS. " + }, + { + "bbox": [ + 313, + 316, + 555, + 372 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\mathrm{trg}}" + }, + { + "bbox": [ + 313, + 316, + 555, + 372 + ], + "type": "text", + "content": " is edited image of source image " + }, + { + "bbox": [ + 313, + 316, + 555, + 372 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\mathrm{src}}" + }, + { + "bbox": [ + 313, + 316, + 555, + 372 + ], + "type": "text", + "content": " by prompt " + }, + { + "bbox": [ + 313, + 316, + 555, + 372 + ], + "type": "inline_equation", + "content": "y^{\\mathrm{src}}\\rightarrow y^{\\mathrm{trg}}" + }, + { + "bbox": [ + 313, + 316, + 555, + 372 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 316, + 555, + 372 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\mathrm{src\\dagger}}" + }, + { + "bbox": [ + 313, + 316, + 555, + 372 + ], + "type": "text", + "content": " is the inverted image of " + }, + { + "bbox": [ + 313, + 316, + 555, + 372 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\mathrm{trg}}" + }, + { + "bbox": [ + 313, + 316, + 555, + 372 + ], + "type": "text", + "content": " by prompt " + }, + { + "bbox": [ + 313, + 316, + 555, + 372 + ], + "type": "inline_equation", + "content": "y^{\\mathrm{trg}}\\rightarrow y^{\\mathrm{src}}" + }, + { + "bbox": [ + 313, + 316, + 555, + 372 + ], + "type": "text", + "content": ". (First row) Inversion result of DDS with timestep " + }, + { + "bbox": [ + 313, + 316, + 555, + 372 + ], + "type": "inline_equation", + "content": "t\\sim \\mathcal{U}(0,0.2)" + }, + { + "bbox": [ + 313, + 316, + 555, + 372 + ], + "type": "text", + "content": ". (Second row) Inversion result of DDS with " + }, + { + "bbox": [ + 313, + 316, + 555, + 372 + ], + "type": "inline_equation", + "content": "t\\sim \\mathcal{U}(0,1)" + }, + { + "bbox": [ + 313, + 316, + 555, + 372 + ], + "type": "text", + "content": ". (Third row) Inversion result of ours." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 384, + 555, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 555, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 555, + 529 + ], + "type": "text", + "content": "strained by " + }, + { + "bbox": [ + 313, + 384, + 555, + 529 + ], + "type": "inline_equation", + "content": "t \\sim \\mathcal{U}(0,0.2)" + }, + { + "bbox": [ + 313, + 384, + 555, + 529 + ], + "type": "text", + "content": ". This is because the posterior mean " + }, + { + "bbox": [ + 313, + 384, + 555, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{0|t}^{\\mathrm{src}}" + }, + { + "bbox": [ + 313, + 384, + 555, + 529 + ], + "type": "text", + "content": " and the source image " + }, + { + "bbox": [ + 313, + 384, + 555, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\mathrm{src}}" + }, + { + "bbox": [ + 313, + 384, + 555, + 529 + ], + "type": "text", + "content": " are similar for small timestep " + }, + { + "bbox": [ + 313, + 384, + 555, + 529 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 384, + 555, + 529 + ], + "type": "text", + "content": ", as illustrated in the first row of supplementary Fig. S1. However, as depicted in the first row of Fig. 3, DDS yields unrealistic result with this setting, whereby the structure of the given image " + }, + { + "bbox": [ + 313, + 384, + 555, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\mathrm{src}}" + }, + { + "bbox": [ + 313, + 384, + 555, + 529 + ], + "type": "text", + "content": " is overemphasized. This implies that it is not sufficient to simply limit the timestep " + }, + { + "bbox": [ + 313, + 384, + 555, + 529 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 384, + 555, + 529 + ], + "type": "text", + "content": " to prevent the score from deviating too far from " + }, + { + "bbox": [ + 313, + 384, + 555, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\mathrm{src}}" + }, + { + "bbox": [ + 313, + 384, + 555, + 529 + ], + "type": "text", + "content": " to correct the misalignment of the score to " + }, + { + "bbox": [ + 313, + 384, + 555, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\mathrm{src}}" + }, + { + "bbox": [ + 313, + 384, + 555, + 529 + ], + "type": "text", + "content": ". Hence, we propose a fundamental approach to refine the gradient to achieve identity consistency without unwanted overemphasis on details." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 534, + 553, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 553, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 553, + 548 + ], + "type": "text", + "content": "4.2. Identity-preserving Distillation Sampling (IDS)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 553, + 554, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 553, + 554, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 553, + 554, + 639 + ], + "type": "text", + "content": "Fixed-point Regularization (FPR). Here, we introduce a Fixed-point Regularization (FPR) method that adjusts the text-conditioned score " + }, + { + "bbox": [ + 313, + 553, + 554, + 639 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}^{\\mathrm{src}}" + }, + { + "bbox": [ + 313, + 553, + 554, + 639 + ], + "type": "text", + "content": " to the source image " + }, + { + "bbox": [ + 313, + 553, + 554, + 639 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\mathrm{src}}" + }, + { + "bbox": [ + 313, + 553, + 554, + 639 + ], + "type": "text", + "content": ". Our key premise is that if the score " + }, + { + "bbox": [ + 313, + 553, + 554, + 639 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}^{\\mathrm{src}}" + }, + { + "bbox": [ + 313, + 553, + 554, + 639 + ], + "type": "text", + "content": " is rightly estimated as a gradient to " + }, + { + "bbox": [ + 313, + 553, + 554, + 639 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\mathrm{src}}" + }, + { + "bbox": [ + 313, + 553, + 554, + 639 + ], + "type": "text", + "content": ", the posterior mean " + }, + { + "bbox": [ + 313, + 553, + 554, + 639 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{0|t}^{\\mathrm{src}}" + }, + { + "bbox": [ + 313, + 553, + 554, + 639 + ], + "type": "text", + "content": " also contains sufficient information about " + }, + { + "bbox": [ + 313, + 553, + 554, + 639 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\mathrm{src}}" + }, + { + "bbox": [ + 313, + 553, + 554, + 639 + ], + "type": "text", + "content": ". Therefore, FPR loss is designed to minimize the difference between " + }, + { + "bbox": [ + 313, + 553, + 554, + 639 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\mathrm{src}}" + }, + { + "bbox": [ + 313, + 553, + 554, + 639 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 553, + 554, + 639 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{0|t}^{\\mathrm{src}}" + }, + { + "bbox": [ + 313, + 553, + 554, + 639 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 392, + 645, + 553, + 659 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 645, + 553, + 659 + ], + "spans": [ + { + "bbox": [ + 392, + 645, + 553, + 659 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {F P R}} = d \\left(\\mathbf {z} ^ {\\mathrm {s r c}}, \\mathbf {z} _ {0 | t} ^ {\\mathrm {s c}}\\right), \\tag {6}", + "image_path": "39f2cf5b740cdf0c0536c605ee98365e1c1bc4c5f45274d32837dd2353692b9e.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "type": "inline_equation", + "content": "d(\\mathbf{x}_1,\\mathbf{x}_2)" + }, + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "type": "text", + "content": " can be any metric to compare " + }, + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_1" + }, + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_2" + }, + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "type": "text", + "content": ". Here, we employed the Euclidean loss, and further investigations using various metrics are provided in Supplementary Materials." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11118" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 190 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 190 + ], + "type": "text", + "content": "The score " + }, + { + "bbox": [ + 55, + 72, + 296, + 190 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}^{\\mathrm{src}}" + }, + { + "bbox": [ + 55, + 72, + 296, + 190 + ], + "type": "text", + "content": " needs to be modified to minimize the FPR loss before obtaining the updated direction. There are two ways to control the score " + }, + { + "bbox": [ + 55, + 72, + 296, + 190 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}^{\\mathrm{src}}" + }, + { + "bbox": [ + 55, + 72, + 296, + 190 + ], + "type": "text", + "content": " by altering the injection noise " + }, + { + "bbox": [ + 55, + 72, + 296, + 190 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 72, + 296, + 190 + ], + "type": "text", + "content": " or the source latent " + }, + { + "bbox": [ + 55, + 72, + 296, + 190 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t^{\\mathrm{src}}" + }, + { + "bbox": [ + 55, + 72, + 296, + 190 + ], + "type": "text", + "content": ". As illustrated in supplementary Fig. S1, the proposed FPR revises the score " + }, + { + "bbox": [ + 55, + 72, + 296, + 190 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}^{\\mathrm{src}}" + }, + { + "bbox": [ + 55, + 72, + 296, + 190 + ], + "type": "text", + "content": " to serve the source's identity for both approaches. Note that the score incorporates the content details, with the updates being performed with respect to the source latent " + }, + { + "bbox": [ + 55, + 72, + 296, + 190 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t^{\\mathrm{src}}" + }, + { + "bbox": [ + 55, + 72, + 296, + 190 + ], + "type": "text", + "content": " compared to the noise " + }, + { + "bbox": [ + 55, + 72, + 296, + 190 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 72, + 296, + 190 + ], + "type": "text", + "content": ". Thus, " + }, + { + "bbox": [ + 55, + 72, + 296, + 190 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t^{\\mathrm{src}}" + }, + { + "bbox": [ + 55, + 72, + 296, + 190 + ], + "type": "text", + "content": " is updated to minimize the FPR loss as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 122, + 191, + 295, + 205 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 191, + 295, + 205 + ], + "spans": [ + { + "bbox": [ + 122, + 191, + 295, + 205 + ], + "type": "interline_equation", + "content": "\\mathbf {z} _ {t} ^ {\\mathrm {s r c}} \\leftarrow \\mathbf {z} _ {t} ^ {\\mathrm {s r c}} - \\lambda \\nabla_ {\\mathbf {z} _ {t} ^ {\\mathrm {s r c}}} \\mathcal {L} _ {\\mathrm {F P R}}, \\tag {7}", + "image_path": "9f50a35bfb53d93840d6b641733d083b49534a27c1d1a3362568b077467e23a9.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 208, + 296, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 208, + 296, + 233 + ], + "spans": [ + { + "bbox": [ + 55, + 208, + 296, + 233 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 208, + 296, + 233 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 208, + 296, + 233 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 208, + 296, + 233 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 208, + 296, + 233 + ], + "type": "text", + "content": " denote a regularization scale and the number of iterations, respectively." + } + ] + } + ], + "index": 2 + }, + { + "type": "code", + "bbox": [ + 56, + 257, + 290, + 407 + ], + "blocks": [ + { + "bbox": [ + 56, + 242, + 248, + 255 + ], + "lines": [ + { + "bbox": [ + 56, + 242, + 248, + 255 + ], + "spans": [ + { + "bbox": [ + 56, + 242, + 248, + 255 + ], + "type": "text", + "content": "Algorithm 1 Fixed-point Regularization (FPR)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 56, + 257, + 290, + 407 + ], + "lines": [ + { + "bbox": [ + 56, + 257, + 290, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 257, + 290, + 407 + ], + "type": "text", + "content": "Require: " + }, + { + "bbox": [ + 56, + 257, + 290, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\mathrm{src}},y^{\\mathrm{src}},\\epsilon_{\\phi},\\omega ,\\lambda ,N" + }, + { + "bbox": [ + 56, + 257, + 290, + 407 + ], + "type": "text", + "content": " \n1: " + }, + { + "bbox": [ + 56, + 257, + 290, + 407 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0,\\mathbf{I})" + }, + { + "bbox": [ + 56, + 257, + 290, + 407 + ], + "type": "text", + "content": " \n2: " + }, + { + "bbox": [ + 56, + 257, + 290, + 407 + ], + "type": "inline_equation", + "content": "t\\sim \\mathcal{U}(0,1)" + }, + { + "bbox": [ + 56, + 257, + 290, + 407 + ], + "type": "text", + "content": " \n3: " + }, + { + "bbox": [ + 56, + 257, + 290, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t^{\\mathrm{src}}\\gets \\sqrt{\\alpha_t}\\mathbf{z}^{\\mathrm{src}} + \\sqrt{1 - \\alpha_t}\\epsilon" + }, + { + "bbox": [ + 56, + 257, + 290, + 407 + ], + "type": "text", + "content": " \n4: for " + }, + { + "bbox": [ + 56, + 257, + 290, + 407 + ], + "type": "inline_equation", + "content": "\\mathrm{i} = 1,\\dots ,\\mathrm{N}" + }, + { + "bbox": [ + 56, + 257, + 290, + 407 + ], + "type": "text", + "content": " do \n5: " + }, + { + "bbox": [ + 56, + 257, + 290, + 407 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}^{\\mathrm{src}}\\gets (1 + \\omega)\\epsilon_{\\phi}(\\mathbf{z}_{t}^{\\mathrm{src}},y^{\\mathrm{src}},t) - \\omega \\epsilon_{\\phi}(\\mathbf{z}_{t}^{\\mathrm{src}},\\emptyset ,t)" + }, + { + "bbox": [ + 56, + 257, + 290, + 407 + ], + "type": "text", + "content": " \n6: " + }, + { + "bbox": [ + 56, + 257, + 290, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{0|t}^{\\mathrm{src}}\\gets \\frac{1}{\\sqrt{\\alpha_t}} (\\mathbf{z}_t^{\\mathrm{src}} - \\sqrt{1 - \\alpha_t}\\epsilon_\\phi^{\\mathrm{src}})" + }, + { + "bbox": [ + 56, + 257, + 290, + 407 + ], + "type": "text", + "content": " \n7: " + }, + { + "bbox": [ + 56, + 257, + 290, + 407 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{FPR}}\\gets d(\\mathbf{z}_{0|t}^{\\mathrm{src}},\\mathbf{z}^{\\mathrm{src}})" + }, + { + "bbox": [ + 56, + 257, + 290, + 407 + ], + "type": "text", + "content": " \n8: " + }, + { + "bbox": [ + 56, + 257, + 290, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t^{\\mathrm{src}}\\gets \\mathbf{z}_t^{\\mathrm{src}} - \\lambda \\nabla_{\\mathbf{z}_t^{\\mathrm{src}}}\\mathcal{L}_{\\mathrm{FPR}}" + }, + { + "bbox": [ + 56, + 257, + 290, + 407 + ], + "type": "text", + "content": " \n9: end for \n10: " + }, + { + "bbox": [ + 56, + 257, + 290, + 407 + ], + "type": "inline_equation", + "content": "\\epsilon^{*}\\gets \\frac{1}{\\sqrt{1 - \\alpha_{t}}} (\\mathbf{z}_{t}^{\\mathrm{src}} - \\sqrt{\\alpha_{t}}\\mathbf{z}^{\\mathrm{src}})" + }, + { + "bbox": [ + 56, + 257, + 290, + 407 + ], + "type": "text", + "content": " \n11: return " + }, + { + "bbox": [ + 56, + 257, + 290, + 407 + ], + "type": "inline_equation", + "content": "\\epsilon^{*}" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "code_body" + } + ], + "index": 4, + "sub_type": "algorithm" + }, + { + "bbox": [ + 55, + 415, + 296, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 296, + 459 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 296, + 459 + ], + "type": "text", + "content": "Editing with guided noise. Thanks to the proposed FPR, the optimized source latent " + }, + { + "bbox": [ + 55, + 415, + 296, + 459 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t^{\\mathrm{sfc*}}" + }, + { + "bbox": [ + 55, + 415, + 296, + 459 + ], + "type": "text", + "content": " containing the source's identity can be obtained. Then, the guided noise " + }, + { + "bbox": [ + 55, + 415, + 296, + 459 + ], + "type": "inline_equation", + "content": "\\epsilon^{*}" + }, + { + "bbox": [ + 55, + 415, + 296, + 459 + ], + "type": "text", + "content": " is extracted as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 113, + 462, + 295, + 485 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 462, + 295, + 485 + ], + "spans": [ + { + "bbox": [ + 113, + 462, + 295, + 485 + ], + "type": "interline_equation", + "content": "\\epsilon^ {*} = \\frac {1}{\\sqrt {1 - \\alpha_ {t}}} \\left(\\mathbf {z} _ {t} ^ {\\mathrm {s r c} *} - \\sqrt {\\alpha_ {t}} \\mathbf {z} ^ {\\mathrm {s r c}}\\right). \\tag {8}", + "image_path": "b5a74d1e3872cb99f212a381cbab4c68d7c50c681d39801c4304ec1e397483df.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 491, + 296, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 296, + 527 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 296, + 527 + ], + "type": "inline_equation", + "content": "\\epsilon^{*}" + }, + { + "bbox": [ + 55, + 491, + 296, + 527 + ], + "type": "text", + "content": " is utilized to produce the stochastic latent " + }, + { + "bbox": [ + 55, + 491, + 296, + 527 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t^{\\mathrm{trg}*}" + }, + { + "bbox": [ + 55, + 491, + 296, + 527 + ], + "type": "text", + "content": " by applying the forward diffusion process to the target image " + }, + { + "bbox": [ + 55, + 491, + 296, + 527 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\mathrm{trg}}" + }, + { + "bbox": [ + 55, + 491, + 296, + 527 + ], + "type": "text", + "content": ". With " + }, + { + "bbox": [ + 55, + 491, + 296, + 527 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t^{\\mathrm{src}*}" + }, + { + "bbox": [ + 55, + 491, + 296, + 527 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 491, + 296, + 527 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t^{\\mathrm{trg}*}" + }, + { + "bbox": [ + 55, + 491, + 296, + 527 + ], + "type": "text", + "content": ", the updated direction is given by:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 63, + 530, + 295, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 530, + 295, + 563 + ], + "spans": [ + { + "bbox": [ + 63, + 530, + 295, + 563 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\theta} \\mathcal {L} _ {\\mathrm {I D S}} = \\mathbb {E} _ {t, \\epsilon} \\left[ \\left(\\epsilon_ {\\phi} ^ {\\omega} \\left(\\mathbf {z} _ {t} ^ {\\operatorname {t r g} *}, y ^ {\\operatorname {t r g}}, t\\right) - \\epsilon_ {\\phi} ^ {\\omega} \\left(\\mathbf {z} _ {t} ^ {\\operatorname {s r c} *}, y ^ {\\operatorname {s r c}}, t\\right)\\right) \\frac {\\partial \\mathbf {z} ^ {\\operatorname {t r g}}}{\\partial \\theta} \\right]. \\tag {9}", + "image_path": "0e4fa0496fae20ba18f35b3662195b164add4a14a4b614838fb09099611013a3.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 564, + 296, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 296, + 660 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 296, + 660 + ], + "type": "text", + "content": "It is worth noting that " + }, + { + "bbox": [ + 55, + 564, + 296, + 660 + ], + "type": "inline_equation", + "content": "\\epsilon^{*}" + }, + { + "bbox": [ + 55, + 564, + 296, + 660 + ], + "type": "text", + "content": " guides the appropriate gradients for editing while conserving the source's identity. In contrast to DDS, the proposed IDS perfectly reconstructs the source from the edited result " + }, + { + "bbox": [ + 55, + 564, + 296, + 660 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\mathrm{trg}}" + }, + { + "bbox": [ + 55, + 564, + 296, + 660 + ], + "type": "text", + "content": ", as shown in the third row of Fig. 3. This confirms that the correct score and the corresponding injection noise can preserve the identity without further consideration of mutual information. The flowchart of our IDS is illustrated in Fig. 2." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 669, + 108, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 669, + 108, + 681 + ], + "spans": [ + { + "bbox": [ + 55, + 669, + 108, + 681 + ], + "type": "text", + "content": "5. Results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "We evaluate our method through editing experiments conducted on two experiments. In Sec. 5.1, we perform a com" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": "parison on image-to-image editing across several datasets. In Sec. 5.2, we extend our evaluation to editable Neural Radiance Fields (NeRF) [16]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 115, + 459, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 115, + 459, + 129 + ], + "spans": [ + { + "bbox": [ + 313, + 115, + 459, + 129 + ], + "type": "text", + "content": "5.1. Text-guided image editing" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 133, + 555, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 555, + 217 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 555, + 217 + ], + "type": "text", + "content": "Baselines. To evaluate our method, we conduct comparative experiments against four state-of-the-art image editing models: Prompt-to-Prompt (P2P) [7], Plug-and-Play (PNP) [26], DDS [6], and CDS [17]. The implementations of the baselines are carried out by referencing the official source code for each method. More details are provided in Supplementary Materials." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 217, + 555, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 555, + 373 + ], + "type": "text", + "content": "Qualitative Results. We present the qualitative results comparing our method with the baselines in Fig. 4. Prompt-to-Prompt (P2P) [7] performs image editing after applying DDIM inversion [4, 24] to the source image, leading to disregarding the structural components of the source image and following the target prompt excessively. Plug-and-Play (PnP) [26] has limitations in object recognition, as seen in the fourth row of Fig. 4. The third row of Fig. 4 demonstrates that DDS [6] and CDS [17] exhibited limitations, particularly in preserving the structural characteristics of the source image. In contrast, our method successfully edits the image while preserving the structural integrity of the source image." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 373, + 556, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 556, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 556, + 552 + ], + "type": "text", + "content": "Quantitative Results. To measure the identity-preserving performance, we utilize two datasets. First, we collect 250 cat images from the LAION 5B dataset [23] based on [17] for Cat-to-Others task and measure Intersection over Union (IoU). Second, we gather 28 images from the InstructPix2Pix (IP2P) dataset [2], which contains the pairs of source and target images and corresponding prompts and calculate the background Peak-Signal-to-Noise-Ratio (PSNR). Details of the metrics are provided in Supplementary Materials. In addition, we use the LPIPS score [29] for each experiment to quantify the similarity between source and target images. The results are presented in Tab. 1. Our method consistently achieves the lowest LPIPS score across all datasets, indicating that it best preserves the structural semantics of the source images." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 552, + 556, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 556, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 556, + 601 + ], + "type": "text", + "content": "For user evaluation, we present 35 comparison sets for four baselines and our method, gathering responses from 47 participants. Participants are asked to choose the most appropriate image for the following three questions: 1. Which" + } + ] + } + ], + "index": 17 + }, + { + "type": "table", + "bbox": [ + 319, + 616, + 548, + 673 + ], + "blocks": [ + { + "bbox": [ + 319, + 616, + 548, + 673 + ], + "lines": [ + { + "bbox": [ + 319, + 616, + 548, + 673 + ], + "spans": [ + { + "bbox": [ + 319, + 616, + 548, + 673 + ], + "type": "table", + "html": "
cat2pigcat2squirrelIp2p
MetricIoU (↑)LPIPS (↓)IoU (↑)LPIPS (↓)PSNR (↑)LPIPS (↓)
P2P [7]0.580.420.520.4620.880.47
PnP [26]0.550.520.530.5223.810.39
DDS [6]0.690.280.650.3026.020.24
CDS [17]0.720.250.710.2627.350.21
IDS (Ours)0.740.220.710.2429.250.19
", + "image_path": "15fe5b8572880ac3feb279b2ebe9fcaced4158f082a514ea1208d5e23f324fdb.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 677, + 555, + 711 + ], + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 711 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 711 + ], + "type": "text", + "content": "Table 1. Quantitative results for image editing. LPIPS [29] and IoU was measured on LAION 5B [23], while LPIPS and background PSNR was measured on InstructPix2Pix [2]." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11119" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 72, + 552, + 543 + ], + "blocks": [ + { + "bbox": [ + 61, + 72, + 552, + 543 + ], + "lines": [ + { + "bbox": [ + 61, + 72, + 552, + 543 + ], + "spans": [ + { + "bbox": [ + 61, + 72, + 552, + 543 + ], + "type": "image", + "image_path": "8829e7eac60bf87688e6d33af64d90f86b791ee186d114ffb93e1f244f670a6a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 546, + 555, + 570 + ], + "lines": [ + { + "bbox": [ + 55, + 546, + 555, + 570 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 555, + 570 + ], + "type": "text", + "content": "Figure 4. Qualitative results of InstructPix2Pix dataset [2]. Our method successfully edits the image aligning with the target text prompt while preserving the structural integrity of the source image." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 590, + 295, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 295, + 615 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 295, + 615 + ], + "type": "text", + "content": "image best fits the text condition? 2. Which image best preserves the structural information of the original image?" + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 57, + 630, + 294, + 683 + ], + "blocks": [ + { + "bbox": [ + 57, + 630, + 294, + 683 + ], + "lines": [ + { + "bbox": [ + 57, + 630, + 294, + 683 + ], + "spans": [ + { + "bbox": [ + 57, + 630, + 294, + 683 + ], + "type": "table", + "html": "
User Preference Rate (%)GPT score [19]
MetricText (↑)Preserving (↑)Quality (↑)Text (↑)Preserving (↑)Quality (↑)
P2P [7]11.134.808.095.665.375.77
PnP [26]7.727.176.936.546.776.74
DDS [6]20.3010.8216.237.607.517.37
CDS [17]17.0216.7217.088.268.008.09
IDS (Ours)43.8360.4951.678.979.008.80
", + "image_path": "ac70e3f3afce906e5b70b9ba46bc0bb8b7f09584c200d8ca2e1e6aaa7acb1eec.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 688, + 295, + 710 + ], + "lines": [ + { + "bbox": [ + 55, + 688, + 295, + 710 + ], + "spans": [ + { + "bbox": [ + 55, + 688, + 295, + 710 + ], + "type": "text", + "content": "Table 2. User study and GPT scores [19] show that our method achieved the highest scores across all questions for image editing." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 590, + 555, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 590, + 555, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 590, + 555, + 662 + ], + "type": "text", + "content": "3. Which image has the best quality for text-based image editing? Additionally, we measure the GPT score using the Dreambench++ [19] method, which generates human-aligned assessments for the same questions by refining the scoring into ten distinct levels. As shown in Tab. 2, our method receives the highest ratings for all questions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 670, + 403, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 670, + 403, + 683 + ], + "spans": [ + { + "bbox": [ + 313, + 670, + 403, + 683 + ], + "type": "text", + "content": "5.2. Editing NeRF" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "We conduct experiments involving 3D rendering of edited images to demonstrate the effectiveness of our method in" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11120" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 541, + 390 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 541, + 390 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 541, + 390 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 541, + 390 + ], + "type": "image", + "image_path": "19031eeb194960ee28999139602c6543b87400b6b7058ba872122eecd641d5ae.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 393, + 555, + 439 + ], + "lines": [ + { + "bbox": [ + 54, + 393, + 555, + 439 + ], + "spans": [ + { + "bbox": [ + 54, + 393, + 555, + 439 + ], + "type": "text", + "content": "Figure 5. Qualitative results on Synthetic " + }, + { + "bbox": [ + 54, + 393, + 555, + 439 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 54, + 393, + 555, + 439 + ], + "type": "text", + "content": " and LLFF datasets. IDS outperforms the baselines by preserving the structural consistency of the source image and maintaining the integrity of regions that should remain unchanged, while precisely editing only the areas specified by the target prompt. Furthermore, comparisons of the depth map results also highlight the superior consistency of our method over other baseline models." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 459, + 295, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 459, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 54, + 459, + 295, + 495 + ], + "type": "text", + "content": "maintaining structural consistency. This approach is particularly relevant as consistency has an even greater impact on outcomes in 3D environments." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 496, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 496, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 54, + 496, + 295, + 544 + ], + "type": "text", + "content": "Datasets. We evaluated our method on widely used NeRF datasets: Synthetic NeRF [16] and LLFF [15]. Since NeRF datasets have no given pairs of source and target prompts, we manually composed image descriptions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 546, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 546, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 546, + 295, + 714 + ], + "type": "text", + "content": "Qualitative Results. Fig. 5 illustrates the qualitative results of our method compared with NeRF editing baselines. In the first row, the target prompt specifies a precise part of the image for fine-grained editing. DDS [6] and CDS [17] fail to differentiate and edit the specific area. At the same time, our method accurately identifies the region indicated by the target prompt in the image and performs detailed editing exclusively on that part. The second row demonstrates a scenario in which the target prompt is designed to edit the mood of the image. Our approach adjusts the colors associated with \"autumn\" and \"leaves\" throughout the image while maintaining consistency in the \"trunk\" whereas DDS and CDS also changed the \"trunk\". In terms of depth maps, our method generates clean depth maps with minimal noise" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 459, + 555, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 459, + 555, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 459, + 555, + 483 + ], + "type": "text", + "content": "after image editing, whereas DDS and CDS introduce noticeable noise into the depth maps." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 323, + 491, + 547, + 540 + ], + "blocks": [ + { + "bbox": [ + 323, + 491, + 547, + 540 + ], + "lines": [ + { + "bbox": [ + 323, + 491, + 547, + 540 + ], + "spans": [ + { + "bbox": [ + 323, + 491, + 547, + 540 + ], + "type": "table", + "html": "
MetricCLIP [21] (↑)User Preference Rate (%)
Text (↑)Preserving (↑)Quality (↑)
DDS [6]0.159636.8828.3732.62
CDS [17]0.159722.7023.4021.28
IDS (Ours)0.162640.4248.2346.10
", + "image_path": "e1574aa11de15ee47c700b943628f0276879f9982ddd1262805c8effe443e77d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 548, + 555, + 583 + ], + "lines": [ + { + "bbox": [ + 313, + 548, + 555, + 583 + ], + "spans": [ + { + "bbox": [ + 313, + 548, + 555, + 583 + ], + "type": "text", + "content": "Table 3. Quantitative results of NeRF editing with respect to CLIP score and User Preference Rate. IDS demonstrates superior quantitative performance compared to the baselines." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "Quantitative Results. Based on edited images, we performed 3D rendering and subsequently conducted quantitative evaluations provided in Tab. 3. To assess whether the edited 3D images are precisely aligned with the target prompts, we measured the CLIP [21] scores at " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "200\\mathrm{k}" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": " iterations of training on the LLFF dataset. We additionally present a user evaluation conducted under the same setup in Sec. 5.1. Consistent with the trends observed in the qualitative results, our method demonstrates superior performance in the quantitative evaluations compared to other baselines." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11121" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 71, + 288, + 125 + ], + "blocks": [ + { + "bbox": [ + 62, + 71, + 288, + 125 + ], + "lines": [ + { + "bbox": [ + 62, + 71, + 288, + 125 + ], + "spans": [ + { + "bbox": [ + 62, + 71, + 288, + 125 + ], + "type": "image", + "image_path": "7b5891dbf5e397e1ca07b53b8994a2ac0a5308f781b9eead31346704728e805b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 104, + 128, + 249, + 138 + ], + "lines": [ + { + "bbox": [ + 104, + 128, + 249, + 138 + ], + "spans": [ + { + "bbox": [ + 104, + 128, + 249, + 138 + ], + "type": "text", + "content": "\"A photo of shark\" " + }, + { + "bbox": [ + 104, + 128, + 249, + 138 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 104, + 128, + 249, + 138 + ], + "type": "text", + "content": " \"A photo of dolphin\"" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 56, + 186, + 289, + 308 + ], + "blocks": [ + { + "bbox": [ + 55, + 138, + 295, + 182 + ], + "lines": [ + { + "bbox": [ + 55, + 138, + 295, + 182 + ], + "spans": [ + { + "bbox": [ + 55, + 138, + 295, + 182 + ], + "type": "text", + "content": "Figure 6. Ablation study on scale " + }, + { + "bbox": [ + 55, + 138, + 295, + 182 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 138, + 295, + 182 + ], + "type": "text", + "content": ". To show the effect of scale more extremely, the number of iterations of FPR is set as 1. The result of DDS is the same as " + }, + { + "bbox": [ + 55, + 138, + 295, + 182 + ], + "type": "inline_equation", + "content": "\\lambda = 0" + }, + { + "bbox": [ + 55, + 138, + 295, + 182 + ], + "type": "text", + "content": " since it means no update for source identity." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 56, + 186, + 289, + 308 + ], + "lines": [ + { + "bbox": [ + 56, + 186, + 289, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 289, + 308 + ], + "type": "image", + "image_path": "fe19d7b84027058dec2d3b1a9a25b31d8ffdc7676f00da67c3d8b1e8ee24a9f0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 82, + 311, + 271, + 322 + ], + "lines": [ + { + "bbox": [ + 82, + 311, + 271, + 322 + ], + "spans": [ + { + "bbox": [ + 82, + 311, + 271, + 322 + ], + "type": "text", + "content": "\"The green leaves\" " + }, + { + "bbox": [ + 82, + 311, + 271, + 322 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 82, + 311, + 271, + 322 + ], + "type": "text", + "content": " \"Yellow and red leaves in autumn\"" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 323, + 295, + 356 + ], + "lines": [ + { + "bbox": [ + 55, + 323, + 295, + 356 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 295, + 356 + ], + "type": "text", + "content": "Figure 7. Optimization steps. Our method effectively preserves the consistency of the source image, even as the number of iteration steps increases up to 400." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 60, + 364, + 289, + 418 + ], + "blocks": [ + { + "bbox": [ + 60, + 364, + 289, + 418 + ], + "lines": [ + { + "bbox": [ + 60, + 364, + 289, + 418 + ], + "spans": [ + { + "bbox": [ + 60, + 364, + 289, + 418 + ], + "type": "table", + "html": "
optim stepFPR iterLPIPS(↓)CLIP(↑)time (sec/img)Memory (GB)
DDS200-0.2400.29322.456.27
CDS200-0.2100.28759.318.83
IDS20010.1990.28550.808.63
30.1900.277107.77
10030.1650.26554.04
1500.1800.27281.25
", + "image_path": "36f56cc4f718574e8dbf5a489f143d797734d8c8a408e6d526194c1131974d5c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 477, + 129, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 477, + 129, + 489 + ], + "spans": [ + { + "bbox": [ + 55, + 477, + 129, + 489 + ], + "type": "text", + "content": "6. Discussions" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 497, + 192, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 497, + 192, + 509 + ], + "spans": [ + { + "bbox": [ + 55, + 497, + 192, + 509 + ], + "type": "text", + "content": "6.1. Ablation studies on FPR" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 516, + 295, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 516, + 295, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 516, + 295, + 588 + ], + "type": "text", + "content": "FPR iteration " + }, + { + "bbox": [ + 55, + 516, + 295, + 588 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 516, + 295, + 588 + ], + "type": "text", + "content": ". We conduct experiments on FPR iterations " + }, + { + "bbox": [ + 55, + 516, + 295, + 588 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 516, + 295, + 588 + ], + "type": "text", + "content": " to evaluate its impact and determine the optimal iteration count. Although performing just one iteration of FPR is sufficient to preserve the source identity, as shown in lower LPIPS score than baselines of Tab. 4, we set " + }, + { + "bbox": [ + 55, + 516, + 295, + 588 + ], + "type": "inline_equation", + "content": "N = 3" + }, + { + "bbox": [ + 55, + 516, + 295, + 588 + ], + "type": "text", + "content": " to emphasize the purpose of our method." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 589, + 295, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 589, + 295, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 589, + 295, + 661 + ], + "type": "text", + "content": "Scale " + }, + { + "bbox": [ + 55, + 589, + 295, + 661 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 589, + 295, + 661 + ], + "type": "text", + "content": ". The scaling factor " + }, + { + "bbox": [ + 55, + 589, + 295, + 661 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 589, + 295, + 661 + ], + "type": "text", + "content": " of FPR determines how much information of source latent " + }, + { + "bbox": [ + 55, + 589, + 295, + 661 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\mathrm{src}}" + }, + { + "bbox": [ + 55, + 589, + 295, + 661 + ], + "type": "text", + "content": " is kept. As shown in Fig. 6, increasing the scale preserves the attributes of the source image, resulting in more successful editing when it is hard to translate due to the structural mismatch between the source and the target prompt." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 670, + 167, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 670, + 167, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 670, + 167, + 684 + ], + "type": "text", + "content": "6.2. Optimization steps" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "To show that our method can prevent error accumulation during translation, we set the experiment to extend the num" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": "ber of optimization steps from 200 to 400. In the results of DDS and CDS, there is color boosting or loss of details due to the cumulated error. In contrast, IDS maintains the characteristics of the original images, such as the color of the pumpkin in the first row of Fig. 7 and the shape of the leaf in the second row of Fig. 7, better than other methods." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 155, + 384, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 155, + 384, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 155, + 384, + 167 + ], + "type": "text", + "content": "7. Limitation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 175, + 555, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 175, + 555, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 175, + 555, + 342 + ], + "type": "text", + "content": "The proposed IDS demonstrates outstanding performance across evaluation metrics assessing consistency between source and target images. However, during FPR process, IDS relies solely on information from the source " + }, + { + "bbox": [ + 313, + 175, + 555, + 342 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{z}^{\\mathrm{src}}, y^{\\mathrm{src}}\\}" + }, + { + "bbox": [ + 313, + 175, + 555, + 342 + ], + "type": "text", + "content": " without incorporating target-side information. This results in comparatively lower CLIP scores [21] than other baselines (Tab. 5) and failure cases for more complex translations (Fig. 8). In addition, our method requires additional computational overhead (Tab. 4) since FPR is applied to each optimization iteration. Detailed discussion about our limitation is provided in Sec. G of supplementary. Our future direction will explore changing the score conditioned by the target prompt " + }, + { + "bbox": [ + 313, + 175, + 555, + 342 + ], + "type": "inline_equation", + "content": "y^{\\mathrm{trg}}" + }, + { + "bbox": [ + 313, + 175, + 555, + 342 + ], + "type": "text", + "content": ", leading to a better alignment with " + }, + { + "bbox": [ + 313, + 175, + 555, + 342 + ], + "type": "inline_equation", + "content": "y^{\\mathrm{trg}}" + }, + { + "bbox": [ + 313, + 175, + 555, + 342 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "type": "table", + "bbox": [ + 323, + 351, + 544, + 391 + ], + "blocks": [ + { + "bbox": [ + 55, + 422, + 295, + 456 + ], + "lines": [ + { + "bbox": [ + 55, + 422, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 295, + 456 + ], + "type": "text", + "content": "Table 4. Computational complexity on 28 images of InstructPix2Pix [2] for various settings. Lower LPIPS and higher CLIP scores mean better quality." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 323, + 351, + 544, + 391 + ], + "lines": [ + { + "bbox": [ + 323, + 351, + 544, + 391 + ], + "spans": [ + { + "bbox": [ + 323, + 351, + 544, + 391 + ], + "type": "table", + "html": "
P2P [7]PnP [26]DDS [6]CDS [17]IDS (Ours)
cat2lion0.290.210.300.290.29
cat2dog0.270.260.270.270.26
Ip2p0.280.300.290.290.28
", + "image_path": "d22728cccfb101b796396b7df7b6351448c5eb01178554b145caf0f250f436b0.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 394, + 554, + 417 + ], + "lines": [ + { + "bbox": [ + 313, + 394, + 554, + 417 + ], + "spans": [ + { + "bbox": [ + 313, + 394, + 554, + 417 + ], + "type": "text", + "content": "Table 5. Limitation of IDS with respect to CLIP score[21] for image editing on LAION 5B [23] and InstructPix2Pix [2]." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 318, + 422, + 550, + 491 + ], + "blocks": [ + { + "bbox": [ + 318, + 422, + 550, + 491 + ], + "lines": [ + { + "bbox": [ + 318, + 422, + 550, + 491 + ], + "spans": [ + { + "bbox": [ + 318, + 422, + 550, + 491 + ], + "type": "image", + "image_path": "3414560104d3a705bce3ba19d768d8f5bbee176d715cb4fab571847e93062713.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 341, + 491, + 525, + 501 + ], + "lines": [ + { + "bbox": [ + 341, + 491, + 525, + 501 + ], + "spans": [ + { + "bbox": [ + 341, + 491, + 525, + 501 + ], + "type": "text", + "content": "\"Photo free night, house, aurora\" → \"... with two dogs\"" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 345, + 502, + 522, + 512 + ], + "lines": [ + { + "bbox": [ + 345, + 502, + 522, + 512 + ], + "spans": [ + { + "bbox": [ + 345, + 502, + 522, + 512 + ], + "type": "text", + "content": "Figure 8. Failure case for complex text prompt." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 520, + 388, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 520, + 388, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 520, + 388, + 532 + ], + "type": "text", + "content": "8. Conclusion" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 540, + 555, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 555, + 708 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 555, + 708 + ], + "type": "text", + "content": "We proposed a new distillation sampling method using a fixed-point regularization which aligns the text-conditioned score towards identity-preserved manifolds. The proposed fixed-point regularization preserves the source's identity by re-projecting the intermediate score status on posterior means. In this manner, corrected noises guide a gradient of distilled score toward identity-consistent manifolds. Owing to self-correction by a fixed-point iterator and guided injection noise, the proposed identity-preserving distillation sampling provides clear and unambiguous representations corresponding to the given prompts in text-guided image editing and editable neural radiance field (NeRF). Furthermore, our model can be utilized as a universal module in addition to the existing score-sampling processes." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11122" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 120 + ], + "type": "text", + "content": "Acknowledgement This work was supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (RS-2024-00335741, RS-2024-00357197)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 131, + 115, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 131, + 115, + 144 + ], + "spans": [ + { + "bbox": [ + 56, + 131, + 115, + 144 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 296, + 715 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 150, + 296, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 296, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 296, + 206 + ], + "type": "text", + "content": "[1] Donghoon Ahn, Hyoungwon Cho, Jaewon Min, Wooseok Jang, Jungwoo Kim, SeonHwa Kim, Hyun Hee Park, Kyong Hwan Jin, and Seungryong Kim. Self-rectifying diffusion sampling with perturbed-attention guidance. arXiv preprint arXiv:2403.17377, 2024. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 207, + 296, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 207, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 207, + 296, + 262 + ], + "type": "text", + "content": "[2] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 1, 2, 5, 6, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 263, + 296, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 263, + 296, + 317 + ], + "spans": [ + { + "bbox": [ + 56, + 263, + 296, + 317 + ], + "type": "text", + "content": "[3] Hyungjin Chung, Jeongsol Kim, Michael Thompson McCann, Marc Louis Klasky, and Jong Chul Ye. Diffusion posterior sampling for general noisy inverse problems. In The Eleventh International Conference on Learning Representations, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 319, + 296, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 319, + 296, + 352 + ], + "spans": [ + { + "bbox": [ + 56, + 319, + 296, + 352 + ], + "type": "text", + "content": "[4] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 1, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 354, + 296, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 296, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 296, + 386 + ], + "type": "text", + "content": "[5] Bradley Efron. Tweedie's formula and selection bias. Journal of the American Statistical Association, 106(496):1602-1614, 2011. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 388, + 296, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 296, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 296, + 431 + ], + "type": "text", + "content": "[6] Amir Hertz, Kfir Aberman, and Daniel Cohen-Or. Delta denoising score. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2328-2337, 2023. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 433, + 296, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 433, + 296, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 433, + 296, + 487 + ], + "type": "text", + "content": "[7] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-or. Prompt-to-prompt image editing with cross-attention control. In The Eleventh International Conference on Learning Representations, 2023. 1, 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 488, + 296, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 488, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 488, + 296, + 521 + ], + "type": "text", + "content": "[8] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 1, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 522, + 296, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 522, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 522, + 296, + 555 + ], + "type": "text", + "content": "[9] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 556, + 296, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 556, + 296, + 610 + ], + "spans": [ + { + "bbox": [ + 56, + 556, + 296, + 610 + ], + "type": "text", + "content": "[10] Susung Hong, Gyuseong Lee, Wooseok Jang, and Seungryong Kim. Improving sample quality of diffusion models using self-attention guidance. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7462–7471, 2023. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 612, + 296, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 612, + 296, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 612, + 296, + 645 + ], + "type": "text", + "content": "[11] Oren Katzir, Or Patashnik, Daniel Cohen-Or, and Dani Lischinski. Noise-free score distillation. arXiv preprint arXiv:2310.17590, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 647, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 296, + 689 + ], + "type": "text", + "content": "[12] Juil Koo, Chanho Park, and Minhyuk Sung. Posterior distillation sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13352-13361, 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 691, + 296, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 296, + 715 + ], + "type": "text", + "content": "[13] Barak Meiri, Dvir Samuel, Nir Darshan, Gal Chechik, Shai Avidan, and Rami Ben-Ari. Fixed-point inversion for text-" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 315, + 73, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 73, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 315, + 73, + 555, + 95 + ], + "type": "text", + "content": "to-image diffusion models. arXiv preprint arXiv:2312.12540, 2023. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 96, + 555, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 555, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 555, + 150 + ], + "type": "text", + "content": "[14] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations, 2022. 1, 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 152, + 555, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 555, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 555, + 207 + ], + "type": "text", + "content": "[15] Ben Mildenhall, Pratul P Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (ToG), 38(4):1-14, 2019. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 209, + 554, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 554, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 554, + 262 + ], + "type": "text", + "content": "[16] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 5, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 264, + 555, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 555, + 319 + ], + "type": "text", + "content": "[17] Hyelin Nam, Gihyun Kwon, Geon Yeong Park, and Jong Chul Ye. Contrastive denoising score for text-guided latent diffusion image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9192-9201, 2024. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 321, + 554, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 321, + 554, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 321, + 554, + 354 + ], + "type": "text", + "content": "[18] Neal Parikh, Stephen Boyd, et al. Proximal algorithms. Foundations and trends® in Optimization, 1(3):127-239, 2014. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 355, + 555, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 355, + 555, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 355, + 555, + 409 + ], + "type": "text", + "content": "[19] Yuang Peng, Yuxin Cui, Haomiao Tang, Zekun Qi, Runpei Dong, Jing Bai, Chunrui Han, Zheng Ge, Xiangyu Zhang, and Shu-Tao Xia. Dreambench++: A human-aligned benchmark for personalized image generation. arXiv preprint arXiv:2406.16855, 2024. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 411, + 554, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 554, + 444 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 554, + 444 + ], + "type": "text", + "content": "[20] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 1, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 445, + 555, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 445, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 445, + 555, + 510 + ], + "type": "text", + "content": "[21] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 512, + 555, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 512, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 512, + 555, + 567 + ], + "type": "text", + "content": "[22] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 568, + 555, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 555, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 555, + 633 + ], + "type": "text", + "content": "[23] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 5, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 635, + 555, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 555, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 555, + 668 + ], + "type": "text", + "content": "[24] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3, 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "type": "text", + "content": "[25] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 1, 3" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11123" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 340 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[26] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023. 1, 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 129, + 295, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 129, + 295, + 183 + ], + "spans": [ + { + "bbox": [ + 55, + 129, + 295, + 183 + ], + "type": "text", + "content": "[27] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 185, + 295, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 185, + 295, + 239 + ], + "spans": [ + { + "bbox": [ + 57, + 185, + 295, + 239 + ], + "type": "text", + "content": "[28] Zike Wu, Pan Zhou, Xuanyu Yi, Xiaoding Yuan, and Hanwang Zhang. Consistent3d: Towards consistent high-fidelity text-to-3d generation with deterministic sampling prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9892–9902, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 240, + 295, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 240, + 295, + 295 + ], + "spans": [ + { + "bbox": [ + 57, + 240, + 295, + 295 + ], + "type": "text", + "content": "[29] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 297, + 294, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 297, + 294, + 340 + ], + "spans": [ + { + "bbox": [ + 57, + 297, + 294, + 340 + ], + "type": "text", + "content": "[30] Zhizhuo Zhou and Shubham Tulsiani. Sparsefusion: Distilling view-conditioned diffusion for 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12588-12597, 2023. 3" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "11124" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Illumination Spectrum Estimation for Multispectral Images via Surface Reflectance Modeling and Spatial-Spectral Feature Generation/afff26f0-40a0-4b11-8350-630ae1c85fad_content_list.json b/2025/Illumination Spectrum Estimation for Multispectral Images via Surface Reflectance Modeling and Spatial-Spectral Feature Generation/afff26f0-40a0-4b11-8350-630ae1c85fad_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9816c0b04a6c8bab8f557c675225b1d26750c60f --- /dev/null +++ b/2025/Illumination Spectrum Estimation for Multispectral Images via Surface Reflectance Modeling and Spatial-Spectral Feature Generation/afff26f0-40a0-4b11-8350-630ae1c85fad_content_list.json @@ -0,0 +1,2204 @@ +[ + { + "type": "text", + "text": "Illumination Spectrum Estimation for Multispectral Images via Surface Reflectance Modeling and Spatial-Spectral Feature Generation", + "text_level": 1, + "bbox": [ + 138, + 130, + 859, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hyejin Oh $^{1}$ Woo-Shik Kim $^{2}$ Sangyoon Lee $^{2}$ YungKyung Park $^{3}$ Je-Won Kang $^{1*}$ $^{1}$ Dept. of Electronic and Electrical Engineering, Ewha W. University, Seoul, South Korea \n $^{2}$ Samsung Advanced Institute of Technology, South Korea \n $^{3}$ Dept.of Color Design, Ewha W. University, Seoul, South Korea", + "bbox": [ + 143, + 202, + 852, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{hyejin5, yungkyung PARK, jewonk}@ewha.ac.kr {wooshik.kim, sangy00n.lee}@samsung.com", + "bbox": [ + 116, + 276, + 875, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 325, + 326, + 342 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multispectral (MS) images contain richer spectral information than RGB images due to their increased number of channels and are widely used for various applications. However, achieving accurate estimation in MS images remains challenging, as previous studies have struggled with spectral diversity and the inherent entanglement between the illuminant and surface reflectance spectra. To tackle these challenges, in this paper, we propose a novel Illumination spectrum estimation technique for MS images via Surface reflectance modeling and Spatial-spectral feature generation (ISS). The proposed technique employs a learnable spectral unmixing (SU) block to enhance surface reflectance modeling, which was unattempted in the illumination spectrum estimation, and a feature mixing block to fuse spectral and spatial features of MS images with cross-attention. The features are refined iteratively and processed through a decoder to produce an illumination spectrum estimator. Experimental results demonstrate that the proposed technique achieves state-of-the-art performance in illumination spectrum estimation in various MS image datasets. The code is available at https://github.com/heyjinnii/ISS-MSI.git.", + "bbox": [ + 89, + 357, + 483, + 691 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 717, + 220, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multispectral (MS) imaging sensors capture a broader range of spectral bands and provide richer information than in RGB images [29, 35]. As these imaging sensors become compact and cost-effective [14], they are actively used in various image processing applications [3, 42, 44, 60, 70]. In computer vision, computational color constancy is used to replicate the perceptual stability of the human visual system, enabling reliable color rendering under varying lighting conditions [7, 34]. In MS imaging, the color constancy", + "bbox": [ + 89, + 742, + 482, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/636e9b20a64b06d4bffd846412b0cef04956e6e9f155016d30eaba5c6d550ec1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 325, + 640, + 422 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4032483957fb5fc17224a778d34dcbfce0e6fafb93ac627bbca60205383ccb79.jpg", + "image_caption": [ + "Figure 1. Our model synthesizes spectral features derived from surface reflectance modeling with CNN features, iteratively refining the fused representation in an IRSS framework. The features enhance the accuracy of illumination spectrum estimation by effectively accounting for the reflectance characteristics of objects within the image, as demonstrated in the last column, in contrast to the conventional deep features shown in the middle column." + ], + "image_footnote": [], + "bbox": [ + 516, + 422, + 640, + 521 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/395f66e907a50ffd92103965006be6acee628de20f13a870873079cf6be1ce7c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 325, + 766, + 422 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c38296a263f3f74f1cc44f28ba0bf934f870481f798eac706dcb643336ed849d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 422, + 766, + 521 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/db3cceb5902486ee1e51ed8dcd1b618b7d0d9343d378eba952e7ba023ed83109.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 767, + 325, + 903, + 422 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/da430e92130ddefa98959edb2e902938ab304e53bbb66958339fcc1e0e80553f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 767, + 422, + 903, + 521 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "becomes even more crucial than in the conventional RGB imaging [14, 26, 28, 40, 54, 72], because MS images with the increased number of channels are highly sensitive to illumination changes. The color constancy is achieved by solving an illumination and reflectance spectra separation (IRSS) problem [48], which is an underconstrained problem, requiring the simultaneous recovery of both the illumination spectrum and surface reflectance [72]. Accordingly, achieving accurate estimation in MS images remains a challenge [9, 21, 72].", + "bbox": [ + 511, + 657, + 906, + 808 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advances in sensor technology have enabled the acquisition of high-resolution, high-quality MS images [14]. The separation is more challenging in the real-world MS images, because both illumination and reflectance contribute more intricately to the observed spectrum and can vary more dynamically over their channels [30, 31, 71].", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Je-Won Kang is a corresponding author.", + "bbox": [ + 107, + 886, + 328, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2215", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Previous studies have attempted to address the IRSS problem by imposing several constraints to MS images. In [2, 10, 72], scene illumination was separated from surface reflectance using low-rank matrix factorizations. However, relying on low-dimensional subspaces of surface reflectance would incur instability when the number of channels is insufficient relative to the basis functions [72]. Despite the limited research to date, deep learning techniques have been applied to MS images for illumination spectrum estimation [14, 40]. However, the previous deep models have largely overlooked the joint analysis of the illumination spectrum and surface reflectance.", + "bbox": [ + 89, + 90, + 480, + 270 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To overcome the limitations of existing methods, we propose a comprehensive deep learning-based IRSS framework, dubbed Illumination spectrum estimation for MS images via Surface reflectance modeling and Spatial-spectral feature generation (ISS). Spectral unmixing (SU) [25] is employed in this framework to decompose the measured spectrum into a set of endmembers, representing the spectral reflectance characteristics of specific surfaces, and abundances, which indicate the proportion of each endmember. SU has been widely used to analyze MS images in various fields, such as material science [68] and remote sensing [22, 64], yet few studies have been conducted on illumination spectrum estimation. In the proposed technique, rather than deriving the exact surface reflectance of objects, which is practically challenging, we focus on integrating refined reflectance surface modeling into a deep learning-based illumination spectrum estimation approach. Gaussian Mixture Model (GMM) was applied to generate robust endmembers and abundances [46, 53, 74]. We extend this approach by incorporating learnable parameters for endmembers and abundances, providing a robust and flexible decomposition method.", + "bbox": [ + 91, + 276, + 483, + 608 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing learning-based methods have attempted to build an IRSS framework in their models. However, they may suffer from a limited amount of reflectance data [39] or rely on image prior such as segmentation masks [63] and clusters [15] that are not closely related to illumination spectrum estimation. Their methods were challenging to address diverse illumination spectrum and complicated scenes. In this paper, we propose a feature mixing block that combines the results of endmembers and abundance maps with deep visual features, fully exploiting both spatial information and spectral characteristics of the surface reflectance in the input MS image. Conventional deep learning-based feature estimation methods, commonly employed for illumination estimation, often focus on specific bright areas within a single object and misclassify irrelevant regions as distinct features. As shown in Fig. 1, even in shaded areas where previous methods would struggle, the proposed ISS technique appropriately captures the relevant regions for illumination spectrum estimation, treating both bright and shaded areas", + "bbox": [ + 89, + 613, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "as the parts of the same material. The features generated by the feature mixing block are fed into a decoder to reconstruct a reflectance map and into an illumination estimator to establish the IRSS framework, finally estimating illumination spectrum of an MS image in the output.", + "bbox": [ + 511, + 90, + 903, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The contributions of this paper are as follows:", + "bbox": [ + 532, + 167, + 836, + 181 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel deep learning-based IRSS framework to accurately estimate illumination spectrum, addressing the challenges of spectral diversity in MS images.", + "- We introduce a learnable SU module and a feature mixing block, which combines refined spectral features from surface reflectance modeling with deep visual features using cross-attention, effectively capturing both spatial and spectral characteristics of MS images.", + "- We evaluate the performance of the proposed technique on various image datasets, demonstrating its superiority to existing studies." + ], + "bbox": [ + 511, + 184, + 903, + 349 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 364, + 663, + 381 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Illumination Spectrum Estimation", + "text_level": 1, + "bbox": [ + 511, + 391, + 812, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Color constancy is achieved through two steps, including illumination estimation and chromatic adaptation. Early color constancy studies have focused on illumination estimation from RGB images. Statistics-based methods [8, 33, 62] and deep learning-based methods [1, 3, 4, 20, 36, 43, 56, 67] were developed to improve the accuracy of illumination estimation. These methods only used RGB images to separate illumination and reflectance.", + "bbox": [ + 511, + 415, + 903, + 535 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In recent studies, MS images were used to address color constancy by exploiting rich spectral information, which allows for more accurate results. Previous methods extended well-established statistical models and relevant assumptions from RGB to MS spectral domains [13, 27, 61]. In [2, 10, 72], they separated illumination spectrum from surface reflectance using low-rank matrix factorizations. This approach enables the decomposition of the scene with few target objects.", + "bbox": [ + 511, + 537, + 905, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Various learning-based illumination estimation techniques were studied for MS images. In [40], illumination spectrum estimation was modeled as a constrained matrix factorization problem. Spectral reflectance images required per-image calibration using a known reference or measurement device [39]. They used a synthesized training set with simulated lighting conditions. However, the limited number of true reflectance data relative to diverse lighting conditions would lead to performance degradation in complex scenes. In [15], an unsupervised method for segmenting illuminant regions was developed. They assumed the illumination spectrum to be piecewise constant as a probabilistic clustering problem. However, this assumption was relatively simple to address various illumination spectrum. In [63], an image segmentation method was used to lo", + "bbox": [ + 511, + 674, + 905, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2216", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/591824fc857145bd88ff8c36e1088314c48025d3841c0e8909b4979a62f60478.jpg", + "image_caption": [ + "Figure 2. The proposed network uses several sub-networks including a spectral feature extractor, spectral unmixing block, and feature mixing block to generate spatial and spectral features, which are iteratively refined. The decoder uses reflectance decoder and illumination estimator, consisting of an IRSS framework, and estimates illumination spectrum as a result." + ], + "image_footnote": [], + "bbox": [ + 99, + 97, + 898, + 294 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "calize regions dominantly affected by illumination within the image, which were then used for illumination estimation. However, the gaps between the segmentation results and key regions needed for accurate illumination estimation were problematic. In [14], an input MS image was divided into multiple triplets, processed using a pyramid-based convolutional color constancy block. However, the spectral information of illumination and surface reflectance among different channels was not fully exploited.", + "bbox": [ + 88, + 375, + 483, + 512 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Spectral Unmixing", + "text_level": 1, + "bbox": [ + 89, + 535, + 272, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The SU [25] has been used to extract unique spectrum from various materials in MS images. Previous SU methods used linear [5] and non-linear mixing models [19]. These methods are based on the assumption that each pixel can be expressed as a combination of endmember spectra and corresponding abundances. GMM [23, 46, 74] and deep learning-based SU methods [6, 16, 50, 52, 57, 59, 69] have been studied, extensively. GMM-based models [53] address spectral variability by modeling each endmember as a probabilistic distribution rather than a fixed value, tailored to real-world applications. Deep learning approaches, on the other hand, utilize neural networks to automatically learn the underlying structure of MS data. CNN-based [69] and autoencoder-based models [16, 50, 52, 59] have demonstrated effectiveness in enhancing unmixing accuracy at the cost of computational complexity.", + "bbox": [ + 88, + 563, + 482, + 805 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our method introduces a more refined, learnable SU model based on the GMM, which can model reflectance surface with high spectral variability. The spectral features is combined with a deep model and is further processed, which can adaptively address the spatial and spectral characteristics of input MS images.", + "bbox": [ + 88, + 809, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Attention Mechanism in Spectral Imaging", + "text_level": 1, + "bbox": [ + 511, + 375, + 872, + 392 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Attention mechanisms of transformers have been widely used to capture spectral characteristics in spectral imaging [11, 18, 37, 38, 55, 73]. Early studies employed spectral and spatial features independently. In [11, 55, 73], the spectral and spatial features are addressed with sequential processing. In [18, 37, 38], the features are processed in parallel using separated branches and combined with late fusion. While these methods were computationally efficient, they failed to effectively fuse the features for illumination spectrum estimation. Recently, cross-attention [51, 65] and 3D convolution [17, 24] were adopted to fuse the spectral and spatial features of MS images. More recently, transformer architectures have been more actively employed for this purpose [32, 41, 45]. Cross-attention mechanisms facilitate the modeling of global dependencies between spectral and spatial features, enabling the effective learning of complex feature relation.", + "bbox": [ + 511, + 398, + 906, + 654 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Method", + "text_level": 1, + "bbox": [ + 511, + 667, + 687, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. MS Image Formulation Model", + "text_level": 1, + "bbox": [ + 511, + 693, + 782, + 708 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In an IRSS framework [48], an MS image $\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times C}$ with height $H$ , width $W$ , and channel $C$ is modeled as", + "bbox": [ + 511, + 714, + 905, + 746 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {I} (x) = \\int_ {\\Lambda} \\mathbf {R} (x, \\lambda) \\cdot \\mathbf {L} (\\lambda) \\cdot \\mathbf {S} (\\lambda) d \\lambda , \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 757, + 906, + 790 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{R}(x,\\lambda)$ represents surface reflectance of a spatial coordinate $x$ and wavelength $\\lambda$ ranged within $\\Lambda$ . $\\mathbf{L}(\\lambda)$ and $\\mathbf{S}(\\lambda) \\in \\mathbb{R}^C$ represent illumination spectrum and camera sensitivity, respectively.", + "bbox": [ + 511, + 794, + 905, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the joint IRSS problem, given camera sensitivity, it is necessary to accurately model reflectance surface to produce $\\mathbf{L}(\\lambda)$ . In our approach, the reflectance is modeled as", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2217", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the product of abundances and endmembers through the SU, representing a detailed spectral characteristics, given as,", + "bbox": [ + 89, + 90, + 482, + 122 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {R} (x, \\lambda) = \\sum_ {k = 1} ^ {K} \\mathbf {U} _ {k} (x) \\cdot \\mathbf {E} _ {k} (\\lambda), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 133, + 482, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{E}_k(\\lambda)$ and $\\mathbf{U}_k(x)$ denote the $k$ -th endmember and its abundance map, respectively. $\\mathbf{I}(x)$ is modeled by incorporating Eq. (2) as follows:", + "bbox": [ + 89, + 184, + 483, + 231 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {I} (x) = \\int_ {\\Lambda} \\left(\\sum_ {k = 1} ^ {K} \\mathbf {U} _ {k} (x) \\cdot \\mathbf {E} _ {k} (\\lambda)\\right) \\cdot \\mathbf {L} (\\lambda) \\cdot \\mathbf {S} (\\lambda) d \\lambda , \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 247, + 482, + 289 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the SU-based reflectance surface model is used to provide more robust and flexible approach to accurately decompose the components. Fig. 2 presents this process with further details in the following subsections.", + "bbox": [ + 89, + 301, + 483, + 362 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Surface Reflectance Representation with Learnable SU", + "text_level": 1, + "bbox": [ + 89, + 372, + 483, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the proposed technique, the SU is used to provide detailed spectrum profile of reflectance surface, and the GMM is employed to decompose observed spectrum into a set of endmember spectra and their corresponding abundances. The components are transformed to learnable tensors to accomplish a deep IRSS framework in the following.", + "bbox": [ + 89, + 411, + 483, + 502 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the GMM, a probability density of an observed spectrum $\\mathbf{I}_x\\in \\mathbb{R}^C$ at each pixel $x$ is modeled as:", + "bbox": [ + 89, + 503, + 483, + 534 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np \\left(\\mathbf {I} _ {x}\\right) = \\sum_ {k = 1} ^ {K} \\pi_ {k} \\mathcal {N} \\left(\\mathbf {I} _ {x} \\mid \\boldsymbol {\\mu} _ {k}, \\boldsymbol {\\Sigma} _ {k}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 547, + 482, + 588 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $K$ is the number of normal distributions, and $\\pmb{\\mu}_k$ and $\\pmb{\\Sigma}_k$ are the mean vector and covariance matrix. $\\pi_k$ is a weighting coefficient, in which $\\sum_{k=1}^{K} \\pi_k = 1$ .", + "bbox": [ + 89, + 595, + 482, + 641 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In Eq. (4), $p(\\mathbf{I}_x)$ represents a likelihood that all $K$ spectrum can explain the observed spectrum, quantified by the linear combination of $\\mathcal{N}(\\mathbf{I}_x|\\boldsymbol{\\mu}_k,\\boldsymbol{\\Sigma}_k)$ and its proportion $\\pi_{k}$ and thus $\\pmb{\\mu}_{k}$ and $\\pmb{\\Sigma}_{k}$ represent the $k$ -th endmember spectrum and the spectral variability within the endmember. To derive the optimal parameter set $\\{\\pi_k^*,\\pmb {\\mu}_k^*,\\pmb {\\Sigma}_k^*\\}$ , given a set of observed spectra $\\mathbf{I}_x\\in \\mathbf{I}$ of training samples, we apply an Expectation-Maximization (EM) algorithm [12] to iteratively refine the parameters.", + "bbox": [ + 89, + 642, + 482, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the results, we directly obtain the end-member spectra from the mean vectors, i.e., $\\pmb{\\mu}^{*} = \\{\\pmb{\\mu}_{1}^{*},\\pmb{\\mu}_{2}^{*},\\dots ,\\pmb{\\mu}_{K}^{*}\\}$ . Then, we calculate a set of abundance $\\gamma^{*} = \\{\\gamma_{1}^{*},\\gamma_{2}^{*},\\dots ,\\gamma_{K}^{*}\\}$ . $\\gamma_k^* (x)$ is computed at each pixel coordinate $x$ as follows:", + "bbox": [ + 89, + 777, + 483, + 853 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\gamma} _ {k} ^ {*} (x) = \\frac {\\pi_ {k} ^ {*} \\mathcal {N} \\left(\\mathbf {I} _ {x} \\mid \\boldsymbol {\\mu} _ {k} ^ {*} , \\boldsymbol {\\Sigma} _ {k} ^ {*}\\right)}{\\sum_ {j = 1} ^ {K} \\pi_ {j} ^ {*} \\mathcal {N} \\left(\\mathbf {I} _ {x} \\mid \\boldsymbol {\\mu} _ {j} ^ {*} , \\boldsymbol {\\Sigma} _ {j} ^ {*}\\right)}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 864, + 482, + 905 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\gamma_k^*(x)$ consists of an abundance map for the $k$ -th end-member spectrum.", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In previous studies [47, 72], surface reflectance has been approximated using a low-dimensional model based on a subspace assumption, where a small number of spectral bases (e.g., 3) is sufficient for most natural images. These findings could lead to a tight subspace expression of the surface reflectance for practical use cases. In our experiments, we empirically selected $K$ and observed best results with $K = 5$ , aligning with the subspace assumption.", + "bbox": [ + 511, + 121, + 903, + 241 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the proposed model, the parameters pass through learnable layers, allowing for a more flexible representation that adapts to real-world images, mathematically expressed,", + "bbox": [ + 511, + 242, + 903, + 287 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {E} = W _ {e} \\boldsymbol {\\mu} ^ {*} \\in \\mathbb {R} ^ {K \\times C}, \\mathbf {U} = W _ {u} \\gamma^ {*} \\in \\mathbb {R} ^ {K \\times H \\times W}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 311, + 903, + 330 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $W_{e}$ and $W_{u}$ are convolution layers. $\\mathbf{E}$ and $\\mathbf{U}$ are the tensors of endmembers and abundance maps, respectively, as the final output of the SU block.", + "bbox": [ + 511, + 340, + 903, + 387 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/10755f8188b178b33a2e1d42fea7d6230d7354d8e393e4edbffbcadf40a40232.jpg", + "image_caption": [ + "Figure 3. Cross-attention operations in a feature mixing block. It uses input spectral and visual features from a SU block and CNN, respectively, and produces the spatial-spectral features as output, iteratively refined with an $N$ -step GRU." + ], + "image_footnote": [], + "bbox": [ + 527, + 400, + 895, + 583 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Feature Mixing Block with Cross-Attention", + "text_level": 1, + "bbox": [ + 511, + 665, + 880, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Feature mixing block (FMB) uses both the SU features and CNN features as input features. The CNN features are from a spatial feature extractor (SFE) which consists of 3 layers of convolutions, as illustrated in Fig. 2. Previous model-based methods neglected spatial information [2, 10, 72], degrading performance in complex scenes. In contrast, deep learning-based methods captured visual features but struggled to integrate physical modeling [14, 40]. In the proposed method, the SU features and CNN features are used complementarily to generate spatial-spectral features tailored for MS images. Fig. 3 presents the detailed operations in the FMB. The spectral and spatial features are fused using cross-attention and iteratively refined through a gated recurrent unit (GRU).", + "bbox": [ + 511, + 688, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2218", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The FMB employs a cross-attention mechanism of a transformer to fuse spectral and spatial features as shown in Fig. 3. In the process, given $\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times C}$ , the spatial features derived from CNN are used as queries, and the end-members from a spectral unmixing block are used as keys and values as follows:", + "bbox": [ + 89, + 90, + 483, + 181 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} = W _ {q} \\mathbf {F} _ {n}, \\quad \\mathbf {K} = W _ {k} \\mathbf {E}, \\quad \\mathbf {V} = W _ {v} \\mathbf {E}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 198, + 482, + 215 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{Q} \\in \\mathbb{R}^{D \\times H \\times W}$ is the query, and $\\mathbf{K}, \\mathbf{V} \\in \\mathbb{R}^{K \\times D}$ are the key and value. $\\mathbf{F}_n$ is a visual feature, which is iteratively refined using a $n$ -step GRU. $\\mathbf{F}_0$ is initialized to $\\mathbf{F}_{cnn}$ as visual features of a SFE network. $D$ is its spectral dimension. $W_q \\in \\mathbb{R}^{D \\times D}$ , and $W_k, W_v \\in \\mathbb{R}^{D \\times C}$ are learnable linear transformations to embed the features to a cross-attention block.", + "bbox": [ + 89, + 223, + 483, + 328 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Spectral-wise correlations $\\mathbf{S} \\in \\mathbb{R}^{K \\times H \\times W}$ between a query of spatial locations in $\\mathbf{F}_n$ and a key of spectral characteristics of $\\mathbf{E}$ are computed as follows:", + "bbox": [ + 89, + 330, + 483, + 378 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {S} = \\mathbf {Q} \\cdot \\mathbf {K} ^ {T}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 238, + 392, + 482, + 411 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the attention scores calculate the relevance between the visual features and spectral features.", + "bbox": [ + 89, + 421, + 482, + 450 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "These spectral correlations are modified by adding abundance map $\\mathbf{U} \\in \\mathbb{R}^{K \\times H \\times W}$ obtained from the SU block, which generating an attention weight $\\mathbf{A} \\in \\mathbb{R}^{K \\times H \\times W}$ as follows:", + "bbox": [ + 89, + 452, + 483, + 513 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {A} = \\operatorname {s o f t m a x} (\\mathbf {S} + \\mathbf {U}), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 526, + 482, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where softmax is a softmax function. Then, we compute an attention output through a weighted sum between attention weights and values, given as:", + "bbox": [ + 89, + 553, + 483, + 599 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {O} = \\mathbf {A} \\cdot \\mathbf {V}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 243, + 614, + 482, + 631 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{O} \\in \\mathbb{R}^{D \\times H \\times W}$ is the attention output.", + "bbox": [ + 89, + 641, + 393, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the FMB, the output undergoes an update process using a GRU followed by a two-layer fully connected network (MLP), which iteratively refines the output feature with the query features. Mathematically, this can be expressed as:", + "bbox": [ + 89, + 659, + 483, + 720 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {n + 1} = \\operatorname {M L P} \\left(\\operatorname {G R U} \\left(\\mathbf {F} _ {n}, \\mathbf {O}\\right)\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 734, + 482, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{F}_n$ is updated for $N$ iterations to fuse the spectral and spatial features, until the final output $\\mathbf{F}_N$ is generated. $\\mathbf{F}_N$ represents a refined reflectance feature, maintaining the same dimensions as the input CNN features while incorporating spectral information from the SU method. $N$ is set to 6 in our experiments. The Fig. 4 shows the gradually refined features during the iterative attention process of FMB.", + "bbox": [ + 89, + 762, + 483, + 868 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We present a pseudo-code implementation in Algorithm 1.", + "bbox": [ + 89, + 869, + 483, + 898 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c0988fd10e984b7d9b6b59d3f0d366cfdaaf6b8ed0bcb9e1b88f56cabb5c219d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 90, + 566, + 157 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/431289af9814dc71ef82f6ea1cd1be6029701b2d8a88f17d12ead44af8b73b12.jpg", + "image_caption": [ + "Step 0" + ], + "image_footnote": [], + "bbox": [ + 566, + 98, + 668, + 157 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ab919df20e4a64a859b9bbad83f919bbafd1c70d982ae1ef0338c2ced7165700.jpg", + "image_caption": [ + "Step 2" + ], + "image_footnote": [], + "bbox": [ + 669, + 98, + 743, + 157 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b55cfaaa440ecd621c7f69e081b013ef357b119573c56294e185337f319709c5.jpg", + "image_caption": [ + "Step 4" + ], + "image_footnote": [], + "bbox": [ + 745, + 98, + 818, + 157 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/18b5f54d93b0fa10f0ed547ec3dfd8d91dda5e3b50b0fcdcd607a6b144bc0788.jpg", + "image_caption": [ + "Step 6" + ], + "image_footnote": [], + "bbox": [ + 820, + 98, + 897, + 157 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2ac9985a1f676375c4cc41e7db2192299294c12b3fa4ad17c56033f0a5aeebba.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 159, + 589, + 217 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4749dbe1c0ad4f659388b234c7d3a97332abe5f4cd8da8bc9d65ec60f1a3a3de.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 591, + 159, + 666, + 217 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1692a0d17ba44fcda318e61d76919a70f6913785c478d4ce0b57f43883f912c7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 159, + 743, + 217 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/00a10150855f298dbf6bd286ed4b2cf3e313babd847e444a2776baaa3434a8eb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 745, + 159, + 818, + 217 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/81ee8b27128cbd811d6e1e256737e34687487b6e25dfe3ee7d514f97716d5dcf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 821, + 159, + 897, + 217 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/693f936579181e53516e3ab149d5052e388e28dfe3e0b2d0c78b7d7d97501cfd.jpg", + "image_caption": [ + "Figure 4. Progressive feature refinement visualization through iterative stages in FMB. The leftmost image shows a rendered RGB of input MS image. From the initial query feature of FMB, the features are gradually refined from step 0 to step 6." + ], + "image_footnote": [], + "bbox": [ + 516, + 219, + 589, + 277 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5b2aa6b73ee539e167fae80cf8a024eace372fd6a5d61bc1ecb93a471c11e3c8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 591, + 219, + 666, + 277 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f32866ff76c3416531b348f48bed73547720ac9a596110c4f98d403a676d3b77.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 219, + 743, + 277 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d4709571de54ef66bc34414813c6840056323c653589e63e7003797d8746feba.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 745, + 219, + 818, + 277 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f6fd1d05109772f3914010b5f5557ea004db5f47a1b9ca46fed30988e80c1083.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 821, + 219, + 897, + 277 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1: SU-CNN Feature Mixing Block" + ], + "code_body": "Input: $\\mathbf{F}_0\\gets \\mathbf{F}_{cnn}\\in \\mathbb{R}^{D\\times H\\times W}$ : CNN feature \n $\\mathbf{E}\\in \\mathbb{R}^{K\\times C}$ : Endmember from SU block \n $\\mathbf{U}\\in \\mathbb{R}^{K\\times H\\times W}$ : Abundance map from SU block \nOutput: $\\mathbf{F}_{\\mathrm{ref}}\\leftarrow \\mathbf{F}_N\\in \\mathbb{R}^{D\\times H\\times W}$ : Updated feature map for N iterations \n $\\mathbf{K}\\leftarrow W_{k}\\mathbf{E}$ //Key projection \n $\\mathbf{V}\\leftarrow W_{v}\\mathbf{E}$ //Value projection \nFunction FeatureMixing $(\\mathbf{F}_0,\\mathbf{K},\\mathbf{V},\\mathbf{U})$ .. \nfor $n = 0$ to $N - 1$ do \nQ←WqFn //Query projection \nS←Q·KT// Spectral correlation \nA←softmax(S+U) //Attention weights \nO←A·V //Attention output \nFn+1←MLP(GRU(Fn,O)) //Feature update with GRU and MLP \nend \nreturn FN \n $\\mathbf{F}_N\\gets$ FeatureMixing(F0,K,V,U)", + "bbox": [ + 532, + 388, + 883, + 676 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Illumination Spectrum Estimation in Decoder Blocks", + "text_level": 1, + "bbox": [ + 513, + 709, + 906, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In decoder blocks, the proposed network generates an estimation of the illumination spectrum, which is denoted by $\\hat{\\mathbf{L}}$ , at the end. The network uses two branches to calculate not only the illumination spectrum estimation but also the surface reflectance approximation, which is denoted by $\\hat{\\mathbf{R}}$ . The approximation of the surface reflectance in the decoder enables the network to use a joint IRSS framework, rather than producing an exact surface reflectance in the absence of GT. As shown in Fig. 2, $\\mathbf{F}_N$ is fed into a reflectance decoder, consisting of three transposed convolution layers, to", + "bbox": [ + 511, + 750, + 906, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2219", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "calculate $\\hat{\\mathbf{R}}$ as intermediate results in the IRSS. Specifically, we use $3\\times 3$ convolution kernels in the implementation. The reflectance decoder uses the only $\\mathbf{F}_N$ as input to facilitate the optimization and fully focus on the reflectance prediction.", + "bbox": [ + 89, + 90, + 480, + 165 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The illumination spectrum block is built with a series of convolutional and linear layers, in which an input is constructed by concatenating $\\mathbf{F}_{cnn}$ and $\\mathbf{F}_N$ . The subsequent convolutional layers are employed to extract spatially localized spectral features, effectively capturing illumination in the target scene. We employ two layers of convolution kernel with stride 2. Rectified linear unit (ReLU) activation function is applied after all convolution and linear layers in the block, except for the final output linear layer.", + "bbox": [ + 89, + 166, + 482, + 303 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.5. Loss function", + "text_level": 1, + "bbox": [ + 89, + 310, + 230, + 324 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For training the proposed architecture, we employ a reconstruction loss $\\mathcal{L}_{\\mathrm{RC}}$ and an illumination estimation loss $\\mathcal{L}_{\\mathrm{AE}}$ . $\\mathcal{L}_{\\mathrm{RC}}$ is computed using an $\\mathcal{L}_1$ distance between $\\mathbf{I}$ and $\\hat{\\mathbf{I}}$ :", + "bbox": [ + 89, + 334, + 482, + 378 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {R C}} = \\sum_ {x \\in (h, w)} \\sum_ {\\lambda \\in \\Lambda} \\| \\hat {\\mathbf {I}} (x, \\lambda) - \\mathbf {I} (x, \\lambda) \\| _ {1}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 390, + 482, + 424 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where the loss is computed over all spatial spectral locations, i.e., $h = 1 \\cdots H$ and $w = 1 \\cdots W$ , and all channels $\\Lambda$ . $\\hat{\\mathbf{I}}$ is reconstructed using an integral of $\\hat{\\mathbf{L}}$ and $\\hat{\\mathbf{R}}$ ,", + "bbox": [ + 89, + 429, + 482, + 473 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also utilize an angular error (AE), $\\mathcal{L}_{\\mathrm{AE}}$ , defined as:", + "bbox": [ + 109, + 474, + 470, + 488 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {A E}} = \\arccos \\left(\\frac {\\hat {\\mathbf {L}} \\cdot \\mathbf {L}}{\\| \\hat {\\mathbf {L}} \\| \\| \\mathbf {L} \\|}\\right), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 497, + 482, + 537 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{L}$ and $\\hat{\\mathbf{L}}$ are the ground-truth (GT) and predicted illumination spectrum vector, respectively. For illumination spectrum estimation, capturing relative differences of illuminant among multi-channels is crucial. Spectral orientation defines its color characteristics, which would affect color constancy, while the intensity of illuminant reflects brightness. Thus, we use an angular error to focus on the orientation by measuring the angle between vectors.", + "bbox": [ + 89, + 547, + 482, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The total loss function $\\mathcal{L}$ is defined as:", + "bbox": [ + 109, + 669, + 367, + 681 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\alpha_ {1} \\mathcal {L} _ {\\mathrm {R C}} + \\alpha_ {2} \\mathcal {L} _ {\\mathrm {A E}}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 698, + 480, + 714 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\alpha_{1}$ and $\\alpha_{2}$ are weighting factors, which are set to 0.5 and 1, respectively, in our experiments.", + "bbox": [ + 89, + 719, + 482, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experimental Results", + "text_level": 1, + "bbox": [ + 89, + 762, + 295, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Spectral Reflectance Datasets", + "text_level": 1, + "bbox": [ + 89, + 786, + 352, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate the proposed technique using two spectral reflectance datasets: KAUST [40] and CAVE [66]. In the datasets, we use 31 channels ranging from $400nm$ to $700nm$ at $10nm$ intervals. We relight the scenes by multiplying the surface reflectance of each pixel with an illuminant spectra matrix with 59 standard illuminants from", + "bbox": [ + 89, + 809, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "[49] resulting in a 31-channel spectral image. We randomly select $80\\%$ of the KAUST reflectance and illuminant data for training and the remaining $20\\%$ used for testing as in [40]. CAVE dataset is utilized only for testing the models. We evaluate the performance of the proposed technique with existing methods including both statistic-based and learning-based methods, GrayEdge [27], LRMF [72], ISNL [58], PWIR [54], DUN [40] and BeyondRGB [14]. We used mean- $\\Delta \\mathrm{A}_{\\mathrm{MS}}$ and std- $\\Delta \\mathrm{A}_{\\mathrm{MS}}$ for evaluation metrics, which are the average and standard deviation of an angular error measured between the GT and predicted illumination spectrum.", + "bbox": [ + 511, + 90, + 903, + 271 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tab. 1 exhibits the performance comparisons, showing that the proposed technique achieves the best results on both datasets. The proposed technique produces 5.55 and 3.29 in terms of mean- $\\Delta \\mathrm{A}_{\\mathrm{MS}}$ and std- $\\Delta \\mathrm{A}_{\\mathrm{MS}}$ , respectively, in KAUST dataset. For CAVE dataset, which is not used for model training, it shows reliable performance achieving 5.93 and 3.90 in mean- $\\Delta \\mathrm{A}_{\\mathrm{MS}}$ and std- $\\Delta \\mathrm{A}_{\\mathrm{MS}}$ , respectively. These results demonstrate that the proposed technique holds stable performance on unseen reflectance data.", + "bbox": [ + 511, + 272, + 906, + 409 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1209e9800feb7f1aad746f87a533cc7269144fc3925f65ed55b5761f2319fa2b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetMethodmean-ΔAMS↓std-ΔAMS↓
KAUSTGrayEdge [27]11.468.02
LRMF [72]24.0623.49
ISNL [58]20.0511.46
PWIR [54]28.0721.20
DUN [40]9.176.30
BeyondRGB [14]7.304.45
Ours5.553.29
CAVEGrayEdge [27]18.3314.89
LRMF [72]20.0515.47
ISNL [58]22.9216.62
PWIR [54]34.9521.20
DUN [40]16.0413.75
BeyondRGB [14]6.643.62
Ours5.933.90
", + "bbox": [ + 522, + 421, + 897, + 640 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Performance comparisons of the proposed technique and existing methods using KAUST and CAVE datasets.", + "bbox": [ + 511, + 650, + 903, + 678 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Beyond RGB Dataset", + "text_level": 1, + "bbox": [ + 511, + 695, + 714, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The BeyondRGB [14] dataset is a comprehensive, real-world dataset, designed for MS image-related studies, focused on mobile devices. The dataset is organized into \"Lab\" and \"Field,\" which include a variety of objects with diverse colors and textures in indoor scenes and real-world data in outdoor scenes, both encompassing complex scenes with rich object content. For experimental results, we use the same data-split protocol of [14]. For comparisons, we used PWIR and BeyondRGB, because the other methods required reflectance data for training [40] or the input dimension needs to match the number of spectral bands in GT spectra [27, 58, 72]", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2220", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/11ad1a0741846c3c4dddda4357a2a6b7032b08cab2a80e65759fc0899361d933.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 133, + 88, + 277, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/289bf1a424a0c8c9de12839718b97ac3c8e1e846ee1b68c4cf628ea545c429e8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 133, + 172, + 274, + 252 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f661da813ea0cf9f722597cc1f1fb0c96868fb167a3d7f526caecda7c525e8de.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 279, + 88, + 424, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a835ebdab787fe677dee1f9ef4bb28f36287917235550f78ca03a03fc79b93ec.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 172, + 424, + 252 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3e3075dc6170a67dc305057d1765627955c5590c0338b4e1e3371737f1561f28.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 426, + 89, + 570, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/00f1eeb682dee850d3576b3c10190465f74e5ada1808e4267b3a1a336a12bc16.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 426, + 172, + 568, + 252 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4738b24ec9c10711e8f5d0ff4e77cdd3f3ba0601e15974f358505f0e678f78bd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 571, + 89, + 715, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/19b4f5c6332d03a7c0d84b2ad6d3b73a73d224b9eae1fd2e58bf53769e0e54ca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 573, + 172, + 715, + 252 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/00d323253bdfa5b31541e95b7aa01b063e72b9b47bf6b50c56e84d3f8bc4e8ed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 717, + 89, + 861, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/02bc1a3ecc451ab18c90ddc1a3354e43a83e5521e8db8f0bcad47fba7d3c3924.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 720, + 172, + 864, + 252 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6984c8bc99acbf5a3f996854968a22675fe48dce008ffc68fc67e39f939c1df9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 133, + 257, + 277, + 340 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a2de56cb48871d04b489e78d3f9a08c4d746c0f9e00178f8f7417b0f50248fe5.jpg", + "image_caption": [ + "Figure 5. Each of two rows shows MS images of BeyondRGB dataset in the first row and their estimated illuminant spectrum in the second row. The results of our model and PWIR are plotted with red and green dashed lines, respectively. The GT is shown with black solid lines." + ], + "image_footnote": [], + "bbox": [ + 133, + 342, + 276, + 422 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f0ebc663ad294d2015e2f9c4e6f8b801bc9b1522436280fc4b8700287e64a0c8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 279, + 258, + 424, + 340 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9f5e41a04383d260e3e73a0160c4d88d896bac651b42a2f9d87fd3e2949d9454.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 342, + 423, + 422 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0e9e1f18ec475138c42ee36c28aa4a619997552c0aa325074701da5fa3597636.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 426, + 258, + 570, + 340 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3e7c87d9958a4254da26edbeef788fc5d58139ca853997ba4dbc32d89f9b17ca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 426, + 342, + 568, + 422 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/016805b07df054c1585d499879e09c69147bd087a78164f2e4ed7ef364fb7599.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 571, + 258, + 715, + 340 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8001f9d055b0095d0df4209f3faa3ccae167aa815ad35be722bc14bfe7d8f2c5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 573, + 342, + 715, + 422 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fb6766194498200ae07dd9a0553cb47d241c0998567ba9e77f545b79265244b0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 717, + 258, + 861, + 340 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0aefebea1e66a61519e73628cdacf1172cfdd0e6c4c28b0ca8e9f664dc052743.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 720, + 342, + 864, + 422 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 2 shows the accuracy of the tested methods. It is clearly demonstrated with the results that the proposed technique provides superior accuracy compared to all the tested methods. Our technique significantly reduces the angular errors, achieving values of 3.65 and 5.97 in the Lab and Field datasets, respectively, in mean- $\\Delta \\mathrm{A}_{\\mathrm{MS}}$ . In the BeyondRGB dataset, the proposed technique presents a greater performance gap compared to the second-best algorithm [14] than in KAUST and CAVE datasets, which can present the effectiveness of our approach in complex scenes.", + "bbox": [ + 89, + 476, + 483, + 628 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/81f340794992dc37565ad63c4c1645888b60c90752cff54aaec730ad6ffce552.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetMethodmean-ΔAMS↓std-ΔAMS↓
LabPWIR [54]27.1010.86
BeyondRGB [14]5.922.92
Ours3.652.72
FieldPWIR [54]16.316.07
BeyondRGB [14]7.225.54
Ours5.973.90
", + "bbox": [ + 101, + 638, + 473, + 755 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Performance comparisons of the proposed technique and existing methods using real-world MS dataset, BeyondRGB.", + "bbox": [ + 89, + 766, + 482, + 795 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 5 shows the results of illumination spectrum using the proposed technique in comparison with PWIR [54]. The top row of the image consists of Lab images, captured with controlled and artificial illuminations. From left to right, the illuminations are the spectrum of fluorescent, mixed LED, halogen lamp, incandescent light bulb, and LED illumi", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "nant, which are found in daily environments. The bottom row contains real-world Field images, where outdoor images are illuminated by natural sunlight. The estimated illumination spectrum of our model, represented by the red dashed line, demonstrate a significantly more detailed alignment with the shape and peaks of the GT compared to PWIR. In the first image of the Lab set, the proposed technique shows a superior performance in estimating complex illumination conditions, such as those with multiple peaks characteristic of fluorescent lighting. Even for the bottom row comprising a wider variety of scenes, our model demonstrates robust and accurate illumination estimation performance.", + "bbox": [ + 511, + 476, + 906, + 674 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Surface Reflectance Modeling", + "text_level": 1, + "bbox": [ + 511, + 681, + 779, + 698 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We investigate the reconstruction results of surface reflectance on the KAUST and CAVE datasets, in which GT reflectance data is available. As shown in Fig. 6, we compare the estimated reflectance from our reflectance decoder and corresponding GT reflectance. From the results, we observe that even though the network has not directly seen GT reflectance during the training process, the surface reflectance is sufficiently considered while estimating illumination from images. Even in complex optical conditions like refractive surfaces with shadows shown in the window and glasses scenes, and reflection and diffraction effects demonstrated by the CD, our method effectively reconstruct the surface reflectance preserving these challenges.", + "bbox": [ + 511, + 704, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2221", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2275750dc7a9756e73f8c594d34c2cc84a2b14abfac46e21d69663875e741b08.jpg", + "image_caption": [ + "Figure 6. Visualization of reconstructed surface reflectance for KAUST and CAVE datasets. The first and second columns display GT and reconstructed surface reflectance rendered in RGB. The following columns display the estimated surface reflectance at each wavelength, respectively." + ], + "image_footnote": [], + "bbox": [ + 93, + 87, + 480, + 272 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ing visual properties. The results align with our motivation for effective disentanglement of illumination and reflectance through physics-based approach and validate the effectiveness of our proposed method.", + "bbox": [ + 89, + 366, + 483, + 428 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 435, + 243, + 452 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Network architecture: Tab. 3 presents the effectiveness of the proposed modules consisting of the network architecture. When the modules are turned off one-by-one, we observe a considerable reduction in accuracy. First, without SU block, the network relies solely on visual features without using spectral features, resulting in an approximate increase of 1.12 in angular error. Second, when disabling FMB, the spatial and spectral features are not fused appropriately but used independently to estimate the illumination spectrum. It increases approximately 3.65 of an angular error. Lastly, when the SFE module is turned off, we observe a dramatic degradation in accuracy. This result is expected, as the network is unable to exploit any visual features from the input image without the SFE. Fig. 7 shows the output features of the FMB, when disabling SFE or SU. Compared to the rendered RGB image on the left, the synthesized feature can be effectively used for illumination estimation by localizing regions relevant to the task.", + "bbox": [ + 88, + 458, + 482, + 731 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/134416055ec8fde66a6c84fe0cf44a1c1784e18878bd45cf1b73ddbd5d748658.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SUFMBSFEmean-ΔAMS ↓
X✓✓5.52
✓X✓8.05
✓✓X24.64
✓✓✓4.40
", + "bbox": [ + 91, + 741, + 318, + 823 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/999f464e4f5b3be7143dfdfea7843f4c9b32ee361132bc23df1ae9358744cb70.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
LAELCRcmean-ΔAMS↓
✓X4.72
X✓73.59
✓✓4.40
", + "bbox": [ + 330, + 741, + 483, + 823 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Ablation studies of the sub-modules of the proposed network (left) and training loss functions (right).", + "bbox": [ + 89, + 833, + 482, + 862 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Loss function: We conducted ablation tests in the loss terms. In the proposed method, using both $\\mathcal{L}_{\\mathrm{AE}}$ and $\\mathcal{L}_{\\mathrm{RC}}$", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ca96fd750fbeb4f7bb731ecdd738b71dfb77ebc4c013cb56d5766cabcde34e3b.jpg", + "image_caption": [ + "Figure 7. Output features from an FMB block in the ablation test. The second and the third columns display the features, when disabling SFE and SU, respectively. Our features in the last column highlight more relevant regions including high spectral diversity in local surfaces to achieve more enhanced accuracy." + ], + "image_footnote": [], + "bbox": [ + 516, + 90, + 903, + 321 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "together achieves the best performance. When $\\mathcal{L}_{\\mathrm{RC}}$ is not used, the performance decreases by approximately 0.32 compared to using both the losses. By incorporating $\\mathcal{L}_{\\mathrm{RC}}$ , the network can further learn the impact of a global illumination spectrum on the reconstructed image in terms of spatial representation. When excluding $\\mathcal{L}_{\\mathrm{AE}}$ , we observed substantial performance drops (e.g. 73 in mean- $\\Delta \\mathrm{A}_{\\mathrm{MS}}$ ). These results are expected, since relying solely on the reflectance for the image reconstruction may provide insufficient constraints for illumination estimation in the absence of GT. $\\mathcal{L}_{\\mathrm{AE}}$ plays a crucial role in the estimation.", + "bbox": [ + 511, + 420, + 906, + 587 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 609, + 633, + 626 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we proposed illumination spectrum estimation technique for MS images. The proposed ISS technique used a learnable SU block which modeled surface reflectance, and a feature mixing block which fused spatial and spectral information using cross-attention operations. The spatial-spectral features were efficiently used to focus on important regions in complex scenes to enhance the accuracy of illumination estimation. Experimental results demonstrated that the proposed ISS technique provided a superiority to existing studies in various MS image datasets.", + "bbox": [ + 511, + 638, + 906, + 790 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In our future study, we will investigate illumination estimation under locally varying and multi-illuminant scenarios. While the current model is designed to estimate a global illumination spectrum, the proposed ISS technique can be extended to improve accuracy in estimating multiple illuminant by efficiently disentangling reflectance properties under locally varying lighting conditions.", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2222", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 240, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was partly supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (RS-2021-II212068, Artificial Intelligence Innovation Hub, $40\\%$ ) and was partly supported by the NRF grant funded by MSIT (No.NRF-2022R1A2C4002052, $60\\%$ ).", + "bbox": [ + 89, + 114, + 483, + 200 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 229, + 187, + 244 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Mahmoud Affifi and Michael S Brown. Sensor-independent illumination estimation for dnn models. arXiv preprint arXiv:1912.06888, 2019. 2", + "[2] Dongsheng An, Jinli Suo, Haoqian Wang, and Qionghai Dai. Illumination estimation from specular highlight in a multispectral image. Optics Express, 23(13):17008-17023, 2015. 2, 4", + "[3] Simone Bianco, Claudio Cusano, and Raimondo Schettini. Color constancy using cnns. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 81-89, 2015. 1, 2", + "[4] Simone Bianco, Claudio Cusano, and Raimondo Schettini. Single and multiple illuminant estimation using convolutional neural networks. IEEE Transactions on Image Processing, 26(9):4347-4362, 2017. 2", + "[5] José M Bioucas-Dias, Antonio Plaza, Nicolas Dobigeon, Mario Parente, Qian Du, Paul Gader, and Jocelyn Chanussot. Hyperspectral unmixing overview: Geometrical, statistical, and sparse regression-based approaches. IEEE journal of selected topics in applied earth observations and remote sensing, 5(2):354-379, 2012. 3", + "[6] Ricardo Augusto Borsoi, Tales Imbiriba, and José Carlos Moreira Bermudez. Deep generative endmember modeling: An application to unsupervised spectral unmixing. IEEE Transactions on Computational Imaging, 6:374-384, 2019. 3", + "[7] David H Brainard and Brian A Wandell. Analysis of the retina theory of color vision. JOSA A, 3(10):1651-1661, 1986. 1", + "[8] Gershon Buchsbaum. A spatial processor model for object colour perception. Journal of the Franklin institute, 310(1): 1-26, 1980. 2", + "[9] Ayan Chakrabarti, Keigo Hirakawa, and Todd Zickler. Color constancy with spatio-spectral statistics. IEEE Transactions on Pattern Analysis and Machine Intelligence, 34(8):1509-1519, 2011. 1", + "[10] Xiaochuan Chen, Mark S Drew, and Ze-Nian Li. Illumination and reflectance spectra separation of hyperspectral image data under multiple illumination conditions. *Electronic Imaging*, 29:194–199, 2017. 2, 4", + "[11] Yushi Chen, Hanlu Jiang, Chunyang Li, Xiuping Jia, and Pedram Ghamisi. Deep feature extraction and classification of hyperspectral images based on convolutional neural networks. IEEE transactions on geoscience and remote sensing, 54(10):6232-6251, 2016. 3" + ], + "bbox": [ + 93, + 253, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Arthur P Dempster, Nan M Laird, and Donald B Rubin. Maximum likelihood from incomplete data via the em algorithm. Journal of the royal statistical society: series B (methodological), 39(1):1-22, 1977. 4", + "[13] Arjan Gijsenij, Theo Gevers, and Joost Van De Weijer. Improving color constancy by photometric edge weighting. IEEE Transactions on Pattern Analysis and Machine Intelligence, 34(5):918-929, 2011. 2", + "[14] Ortal Glatt, Yotam Ater, Woo-Shik Kim, Shira Werman, Oded Berby, Yael Zini, Shay Zelinger, Sangyoon Lee, Heejin Choi, and Evgeny Soloveichik. Beyond rgb: A real world dataset for multispectral imaging in mobile devices. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4344-4354, 2024. 1, 2, 3, 4, 6, 7", + "[15] Lin Gu, Cong Phuoc Huynh, and Antonio Robles-Kelly. Segmentation and estimation of spatially varying illumination. IEEE transactions on image processing, 23(8):3478-3489, 2014. 2", + "[16] Rui Guo, Wei Wang, and Hairong Qi. Hyperspectral image unmixing using autoencoder cascade. In 2015 7th Workshop on Hyperspectral Image and Signal Processing: Evolution in Remote Sensing (WHISPERS), pages 1-4. IEEE, 2015. 3", + "[17] Amina Ben Hamida, Alexandre Benoit, Patrick Lambert, and Chokri Ben Amar. 3-d deep learning approach for remote sensing image classification. IEEE Transactions on geoscience and remote sensing, 56(8):4420-4434, 2018. 3", + "[18] Siyuan Hao, Wei Wang, Yuanxin Ye, Tingyuan Nie, and Lorenzo Bruzzone. Two-stream deep architecture for hyperspectral image classification. IEEE Transactions on Geoscience and Remote Sensing, 56(4):2349-2361, 2017. 3", + "[19] Rob Heylen, Mario Parente, and Paul Gader. A review of nonlinear hyperspectral unmixing methods. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 7(6):1844-1868, 2014. 3", + "[20] Yuanming Hu, Baoyuan Wang, and Stephen Lin. Fc4: Fully convolutional color constancy with confidence-weighted pooling. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4085-4094, 2017. 2", + "[21] Anya Hurlbert. Challenges to color constancy in a contemporary light. Current Opinion in Behavioral Sciences, 30: 186-193, 2019. 1", + "[22] Marian-Daniel Iordache, José M Bioucas-Dias, and Antonio Plaza. Sparse unmixing of hyperspectral data. IEEE Transactions on Geoscience and Remote Sensing, 49(6):2014-2039, 2011. 2", + "[23] Qiwen Jin, Yong Ma, Erting Pan, Fan Fan, Jun Huang, Hao Li, Chenhong Sui, and Xiaoguang Mei. Hyperspectral unmixing with gaussian mixture model and spatial group sparsity. Remote Sensing, 11(20):2434, 2019. 3", + "[24] Murali Kanthi, T Hitendra Sarma, and C Shobha Bindu. A 3d-deep cnn based feature extraction and hyperspectral image classification. In 2020 IEEE India Geoscience and Remote Sensing Symposium (InGARSS), pages 229-232. IEEE, 2020. 3", + "[25] Nirmal Keshava and John F Mustard. Spectral unmixing. IEEE signal processing magazine, 19(1):44-57, 2002. 2, 3" + ], + "bbox": [ + 516, + 92, + 906, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2223", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Haris Ahmad Khan. Multispectral constancy for illuminant invariant representation of multispectral images. PhD thesis, Université Bourgogne Franche-Comté; Norwegian University of Science and ..., 2018. 1", + "[27] Haris Ahmad Khan, Jean-Baptiste Thomas, Jon Yngve Hardeberg, and Olivier Lalignant. Illuminant estimation in multispectral imaging. JOSA A, 34(7):1085-1098, 2017. 2, 6", + "[28] Haris Ahmad Khan, Jean-Baptiste Thomas, and Jon Yngve Hardeberg. Towards highlight based illuminant estimation in multispectral images. In Image and Signal Processing: 8th International Conference, ICISP 2018, Cherbourg, France, July 2-4, 2018, Proceedings 8, pages 517-525. Springer, 2018. 1", + "[29] Muhammad Jaleed Khan, Hamid Saeed Khan, Adeel Yousaf, Khurram Khurshid, and Asad Abbas. Modern trends in hyperspectral image analysis: A review. *IEEE Access*, 6:14118-14129, 2018. 1", + "[30] Dongyoung Kim, Jinwoo Kim, Junsang Yu, and Seon Joo Kim. Attentive illumination decomposition model for multi-illuminant white balancing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25512–25521, 2024. 1", + "[31] Furkan Klinl, Doga Yilmaz, Baris Özcan, and Furkan Kiraç. Deterministic neural illumination mapping for efficient autowhite balance correction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1139-1147, 2023. 1", + "[32] Fanqiang Kong, Tongbo Cao, Yunsong Li, Dan Li, and Kedi Hu. Multi-scale spatial-spectral attention network for multispectral image compression based on variational autoencoder. Signal Processing, 198:108589, 2022. 3", + "[33] Edwin H Land. The retina theory of color vision. Scientific american, 237(6):108-129, 1977. 2", + "[34] Edwin H Land. Recent advances in retina theory. In Central and Peripheral Mechanisms of Colour Vision: Proceedings of an International Symposium Held at The Wenner-Gren Center Stockholm, June 14–15, 1984, pages 5–17. Springer, 1985. 1", + "[35] David Landgrebe. Hyperspectral image data analysis. IEEE Signal processing magazine, 19(1):17-28, 2002. 1", + "[36] Bing Li, Haina Qin, Weihua Xiong, Yangxi Li, Songhe Feng, Weiming Hu, and Stephen Maybank. Ranking-based color constancy with limited training samples. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(10): 12304-12320, 2023. 2", + "[37] Rui Li, Shunyi Zheng, Chenxi Duan, Yang Yang, and Xiqi Wang. Classification of hyperspectral image based on double-branch dual-attention mechanism network. Remote Sensing, 12(3):582, 2020. 3", + "[38] Xian Li, Mingli Ding, and Aleksandra Pizurica. Deep feature fusion via two-stream convolutional neural network for hyperspectral image classification. IEEE Transactions on Geoscience and Remote Sensing, 58(4):2615-2629, 2019. 3", + "[39] Yuqi Li, Chong Wang, Jieyu Zhao, and Qingshu Yuan. Efficient spectral reconstruction using a trichromatic camera via sample optimization. The Visual Computer, 34:1773-1783, 2018. 2" + ], + "bbox": [ + 91, + 92, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Yuqi Li, Qiang Fu, and Wolfgang Heidrich. Multispectral illumination estimation using deep unrolling network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2672-2681, 2021. 1, 2, 4, 6", + "[41] Shuaiqi Liu, Siyuan Liu, Shichong Zhang, Bing Li, Weiming Hu, and Yu-Dong Zhang. Ssau-net: A spectral-spatial attention-based u-net for hyperspectral image fusion. IEEE Transactions on Geoscience and Remote Sensing, 60:1-16, 2022. 3", + "[42] Zhenqi Liu, Xinyu Wang, Yanfei Zhong, Meng Shu, and Chen Sun. Siamhyper: Learning a hyperspectral object tracker from an rgb-based tracker. IEEE Transactions on Image Processing, 31:7116-7129, 2022. 1", + "[43] Yi-Chen Lo, Chia-Che Chang, Hsuan-Chao Chiu, Yu-Hao Huang, Chia-Ping Chen, Yu-Lin Chang, and Kevin Jou. Clcc: Contrastive learning for color constancy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8053-8063, 2021. 2", + "[44] Zhongyu Lou, Theo Gevers, Ninghang Hu, Marcel P Lucassen, et al. Color constancy by deep learning. In BMVC, pages 76-1, 2015. 1", + "[45] Qing Ma, Junjun Jiang, Xianming Liu, and Jiayi Ma. Reciprocal transformer for hyperspectral and multispectral image fusion. Information Fusion, 104:102148, 2024. 3", + "[46] Yong Ma, Qiwen Jin, Xiaoguang Mei, Xiaobing Dai, Fan Fan, Hao Li, and Jun Huang. Hyperspectral unmixing with gaussian mixture model and low-rank representation. Remote Sensing, 11(8):911, 2019. 2, 3", + "[47] Laurence Maloney. Evaluation of linear models of surface spectral reflectance with small numbers of parameters. In Physics-based vision: Principles and practice. Jones and Bartlett, 1992. 4", + "[48] Laurence T Maloney and Brian A Wandell. Color constancy: a method for recovering surface spectral reflectance. In Readings in Computer Vision, pages 293-297. Elsevier, 1987. 1, 3", + "[49] Thomas Mansencal, Michael Mauderer, Michael Parsons, Nick Shaw, Kevin Wheatley, Sean Cooper, Jean D. Vandenberg, Luke Canavan, Katherine Crowson, Ofek Lev, Katrin Leinweber, Shriramana Sharma, Troy James Sobotka, Dominik Moritz, Matt Pppp, Chinmay Rane, Pavithra Eswaramoorthy, John Metic, Ben Pearlstine, Manuel Leonhardt, Olli Niemitalo, Marek Szymanski, Maximilian Schambach, Sianyi Huang, Mike Wei, Nishant Joywardhan, Omar Wagih, Pawel Redman, Joseph Goldstone, Stephen Hill, Jedediah Smith, Frederic Savoir, Geetansh Saxena, Saransh Chopra, Ilia Sibiryakov, Tim Gates, Gajendra Pal, Nicolas Tessore, Aurélien Pierre, François-Xavier Thomas, Sabarish Srinivasan, and Tucker Downs. Colour 0.4.2, 2022. 6", + "[50] Savas Ozkan, Berk Kaya, and Gozde Bozdagi Akar. Endnet: Sparse autoencoder network for endmember extraction and hyperspectral unmixing. IEEE Transactions on Geoscience and Remote Sensing, 57(1):482-496, 2018. 3", + "[51] Yishu Peng, Yuwen Zhang, Bing Tu, Qianming Li, and Wujing Li. Spatial-spectral transformer with cross-attention for hyperspectral image classification. IEEE Transactions on Geoscience and Remote Sensing, 60:1–15, 2022. 3" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2224", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[52] Ying Qu and Hairong Qi. udas: An untied denoising autoencoder with sparsity for spectral unmixing. IEEE Transactions on Geoscience and Remote Sensing, 57(3):1698-1712, 2018. 3", + "[53] Douglas A Reynolds et al. Gaussian mixture models. Encyclopedia of biometrics, 741(659-663), 2009. 2, 3", + "[54] Antonio Robles-Kelly and Ran Wei. A convolutional neural network for pixelwise illuminant recovery in colour and spectral images. In 2018 24th International Conference on Pattern Recognition (ICPR), pages 109-114. IEEE, 2018. 1, 6, 7", + "[55] Swalpa Kumar Roy, Gopal Krishna, Shiv Ram Dubey, and Bidyut B Chaudhuri. Hybridsn: Exploring 3-d-2-d cnn feature hierarchy for hyperspectral image classification. IEEE Geoscience and Remote Sensing Letters, 17(2):277-281, 2019. 3", + "[56] Wu Shi, Chen Change Loy, and Xiaou Tang. Deep specialized network for illuminant estimation. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 371-387. Springer, 2016. 2", + "[57] Alberto Signoroni, Mattia Savardi, Annalisa Baronio, and Sergio Benini. Deep learning meets hyperspectral image analysis: A multidisciplinary review. Journal of imaging, 5(5):52, 2019. 3", + "[58] Tong Su, Yu Zhou, Yao Yu, Xun Cao, and Sidan Du. Illumination separation of non-lambertian scenes from a single hyperspectral image. Optics express, 26(20):26167-26178, 2018. 6", + "[59] Yuanchao Su, Jun Li, Antonio Plaza, Andrea Marinoni, Paolo Gamba, and Somdatta Chakravortty. Daen: Deep autoencoder networks for hyperspectral unmixing. IEEE Transactions on Geoscience and Remote Sensing, 57(7): 4309-4321, 2019. 3", + "[60] Hao Sun, Xiangtao Zheng, and Xiaogiang Lu. A supervised segmentation network for hyperspectral image classification. IEEE Transactions on Image Processing, 30:2810-2825, 2021. 1", + "[61] Jean-Baptiste Thomas. Illuminant estimation from uncalibrated multispectral images. In 2015 Colour and Visual Computing Symposium (CVCS), pages 1-6. IEEE, 2015. 2", + "[62] Joost Van De Weijer, Theo Gevers, and Arjan Gijsenij. Edge-based color constancy. IEEE Transactions on image processing, 16(9):2207-2214, 2007. 2", + "[63] Donik Vršnak, Ilija Domislovic, Marko Subasic, and Sven Lončarić. Framework for illumination estimation and segmentation in multi-illuminant scenes. IEEE access, 11: 2128-2137, 2023. 2", + "[64] Xia Xu and Zhenwei Shi. Multi-objective based spectral unmixing for hyperspectral images. ISPRS Journal of Photogrammetry and Remote Sensing, 124:54-69, 2017. 2", + "[65] Kai Yang, Hao Sun, Chunbo Zou, and Xiaoqiang Lu. Cross-attention spectral-spatial network for hyperspectral image classification. IEEE Transactions on Geoscience and Remote Sensing, 60:1-14, 2021. 3", + "[66] Fumihito Yasuma, Tomoo Mitsunaga, Daisuke Iso, and Shree K Nayar. Generalized assorted pixel camera: post-capture control of resolution, dynamic range, and spectrum." + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE transactions on image processing, 19(9):2241-2253, 2010.6", + "[67] Huanglin Yu, Ke Chen, Kaiqi Wang, Yanlin Qian, Zhaoxiang Zhang, and Kui Jia. Cascading convolutional color constancy. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 12725-12732, 2020. 2", + "[68] Jinkai Zhang, Benoit Rivard, and Arturo Sanchez-Azofeifa. Derivative spectral unmixing of hyperspectral data applied to mixtures of lichen and rock. IEEE Transactions on Geoscience and Remote Sensing, 42(9):1934-1940, 2004. 2", + "[69] Xiangrong Zhang, Yujia Sun, Jingyan Zhang, Peng Wu, and Licheng Jiao. Hyperspectral unmixing via deep convolutional neural networks. IEEE Geoscience and Remote Sensing Letters, 15(11):1755-1759, 2018. 3", + "[70] Chunhui Zhao, Boao Qin, Shou Feng, Wenxiang Zhu, Weiwei Sun, Wei Li, and Xiuping Jia. Hyperspectral image classification with multi-attention transformer and adaptive superpixel segmentation-based active learning. IEEE Transactions on Image Processing, 32:3606-3621, 2023. 1", + "[71] Runxing Zhao, Zhiwen Wang, Wuyuan Guo, and Canlong Zhang. Multi-scene image enhancement based on multi-channel illumination estimation. Expert Systems with Applications, 226:120271, 2023. 1", + "[72] Yinqiang Zheng, Imari Sato, and Yoichi Sato. Illumination and reflectance spectra separation of a hyperspectral image meets low-rank matrix factorization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1779-1787, 2015. 1, 2, 4, 6", + "[73] Zilong Zhong, Jonathan Li, Zhiming Luo, and Michael Chapman. Spectral-spatial residual network for hyperspectral image classification: A 3-d deep learning framework. IEEE Transactions on Geoscience and Remote Sensing, 56 (2):847-858, 2017. 3", + "[74] Yuan Zhou, Anand Rangarajan, and Paul D Gader. A gaussian mixture model representation of endmember variability in hyperspectral unmixing. IEEE Transactions on Image Processing, 27(5):2242-2256, 2018. 2, 3" + ], + "bbox": [ + 516, + 92, + 903, + 614 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "2225", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Illumination Spectrum Estimation for Multispectral Images via Surface Reflectance Modeling and Spatial-Spectral Feature Generation/afff26f0-40a0-4b11-8350-630ae1c85fad_model.json b/2025/Illumination Spectrum Estimation for Multispectral Images via Surface Reflectance Modeling and Spatial-Spectral Feature Generation/afff26f0-40a0-4b11-8350-630ae1c85fad_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7e5c75289600e2648ae78f8fc89a5132be19e480 --- /dev/null +++ b/2025/Illumination Spectrum Estimation for Multispectral Images via Surface Reflectance Modeling and Spatial-Spectral Feature Generation/afff26f0-40a0-4b11-8350-630ae1c85fad_model.json @@ -0,0 +1,2961 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.14, + 0.131, + 0.861, + 0.177 + ], + "angle": 0, + "content": "Illumination Spectrum Estimation for Multispectral Images via Surface Reflectance Modeling and Spatial-Spectral Feature Generation" + }, + { + "type": "text", + "bbox": [ + 0.145, + 0.203, + 0.854, + 0.275 + ], + "angle": 0, + "content": "Hyejin Oh\\(^{1}\\) Woo-Shik Kim\\(^{2}\\) Sangyoon Lee\\(^{2}\\) YungKyung Park\\(^{3}\\) Je-Won Kang\\(^{1*}\\) \n\\(^{1}\\)Dept. of Electronic and Electrical Engineering, Ewha W. University, Seoul, South Korea \n\\(^{2}\\)Samsung Advanced Institute of Technology, South Korea \n\\(^{3}\\)Dept.of Color Design, Ewha W. University, Seoul, South Korea" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.277, + 0.876, + 0.293 + ], + "angle": 0, + "content": "{hyejin5, yungkyung PARK, jewonk}@ewha.ac.kr {wooshik.kim, sangy00n.lee}@samsung.com" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.327, + 0.327, + 0.343 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.358, + 0.485, + 0.692 + ], + "angle": 0, + "content": "Multispectral (MS) images contain richer spectral information than RGB images due to their increased number of channels and are widely used for various applications. However, achieving accurate estimation in MS images remains challenging, as previous studies have struggled with spectral diversity and the inherent entanglement between the illuminant and surface reflectance spectra. To tackle these challenges, in this paper, we propose a novel Illumination spectrum estimation technique for MS images via Surface reflectance modeling and Spatial-spectral feature generation (ISS). The proposed technique employs a learnable spectral unmixing (SU) block to enhance surface reflectance modeling, which was unattempted in the illumination spectrum estimation, and a feature mixing block to fuse spectral and spatial features of MS images with cross-attention. The features are refined iteratively and processed through a decoder to produce an illumination spectrum estimator. Experimental results demonstrate that the proposed technique achieves state-of-the-art performance in illumination spectrum estimation in various MS image datasets. The code is available at https://github.com/heyjinnii/ISS-MSI.git." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.718, + 0.222, + 0.735 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.743, + 0.483, + 0.88 + ], + "angle": 0, + "content": "Multispectral (MS) imaging sensors capture a broader range of spectral bands and provide richer information than in RGB images [29, 35]. As these imaging sensors become compact and cost-effective [14], they are actively used in various image processing applications [3, 42, 44, 60, 70]. In computer vision, computational color constancy is used to replicate the perceptual stability of the human visual system, enabling reliable color rendering under varying lighting conditions [7, 34]. In MS imaging, the color constancy" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.327, + 0.642, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.424, + 0.642, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.327, + 0.767, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.424, + 0.767, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.768, + 0.327, + 0.904, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.768, + 0.424, + 0.904, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.533, + 0.907, + 0.631 + ], + "angle": 0, + "content": "Figure 1. Our model synthesizes spectral features derived from surface reflectance modeling with CNN features, iteratively refining the fused representation in an IRSS framework. The features enhance the accuracy of illumination spectrum estimation by effectively accounting for the reflectance characteristics of objects within the image, as demonstrated in the last column, in contrast to the conventional deep features shown in the middle column." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.658, + 0.907, + 0.809 + ], + "angle": 0, + "content": "becomes even more crucial than in the conventional RGB imaging [14, 26, 28, 40, 54, 72], because MS images with the increased number of channels are highly sensitive to illumination changes. The color constancy is achieved by solving an illumination and reflectance spectra separation (IRSS) problem [48], which is an underconstrained problem, requiring the simultaneous recovery of both the illumination spectrum and surface reflectance [72]. Accordingly, achieving accurate estimation in MS images remains a challenge [9, 21, 72]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Recent advances in sensor technology have enabled the acquisition of high-resolution, high-quality MS images [14]. The separation is more challenging in the real-world MS images, because both illumination and reflectance contribute more intricately to the observed spectrum and can vary more dynamically over their channels [30, 31, 71]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.329, + 0.901 + ], + "angle": 0, + "content": "* Je-Won Kang is a corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2215" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.271 + ], + "angle": 0, + "content": "Previous studies have attempted to address the IRSS problem by imposing several constraints to MS images. In [2, 10, 72], scene illumination was separated from surface reflectance using low-rank matrix factorizations. However, relying on low-dimensional subspaces of surface reflectance would incur instability when the number of channels is insufficient relative to the basis functions [72]. Despite the limited research to date, deep learning techniques have been applied to MS images for illumination spectrum estimation [14, 40]. However, the previous deep models have largely overlooked the joint analysis of the illumination spectrum and surface reflectance." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.277, + 0.485, + 0.609 + ], + "angle": 0, + "content": "To overcome the limitations of existing methods, we propose a comprehensive deep learning-based IRSS framework, dubbed Illumination spectrum estimation for MS images via Surface reflectance modeling and Spatial-spectral feature generation (ISS). Spectral unmixing (SU) [25] is employed in this framework to decompose the measured spectrum into a set of endmembers, representing the spectral reflectance characteristics of specific surfaces, and abundances, which indicate the proportion of each endmember. SU has been widely used to analyze MS images in various fields, such as material science [68] and remote sensing [22, 64], yet few studies have been conducted on illumination spectrum estimation. In the proposed technique, rather than deriving the exact surface reflectance of objects, which is practically challenging, we focus on integrating refined reflectance surface modeling into a deep learning-based illumination spectrum estimation approach. Gaussian Mixture Model (GMM) was applied to generate robust endmembers and abundances [46, 53, 74]. We extend this approach by incorporating learnable parameters for endmembers and abundances, providing a robust and flexible decomposition method." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Existing learning-based methods have attempted to build an IRSS framework in their models. However, they may suffer from a limited amount of reflectance data [39] or rely on image prior such as segmentation masks [63] and clusters [15] that are not closely related to illumination spectrum estimation. Their methods were challenging to address diverse illumination spectrum and complicated scenes. In this paper, we propose a feature mixing block that combines the results of endmembers and abundance maps with deep visual features, fully exploiting both spatial information and spectral characteristics of the surface reflectance in the input MS image. Conventional deep learning-based feature estimation methods, commonly employed for illumination estimation, often focus on specific bright areas within a single object and misclassify irrelevant regions as distinct features. As shown in Fig. 1, even in shaded areas where previous methods would struggle, the proposed ISS technique appropriately captures the relevant regions for illumination spectrum estimation, treating both bright and shaded areas" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.168 + ], + "angle": 0, + "content": "as the parts of the same material. The features generated by the feature mixing block are fed into a decoder to reconstruct a reflectance map and into an illumination estimator to establish the IRSS framework, finally estimating illumination spectrum of an MS image in the output." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.169, + 0.838, + 0.183 + ], + "angle": 0, + "content": "The contributions of this paper are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.185, + 0.905, + 0.229 + ], + "angle": 0, + "content": "- We propose a novel deep learning-based IRSS framework to accurately estimate illumination spectrum, addressing the challenges of spectral diversity in MS images." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.23, + 0.905, + 0.305 + ], + "angle": 0, + "content": "- We introduce a learnable SU module and a feature mixing block, which combines refined spectral features from surface reflectance modeling with deep visual features using cross-attention, effectively capturing both spatial and spectral characteristics of MS images." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.306, + 0.905, + 0.351 + ], + "angle": 0, + "content": "- We evaluate the performance of the proposed technique on various image datasets, demonstrating its superiority to existing studies." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.185, + 0.905, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.366, + 0.664, + 0.382 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.392, + 0.813, + 0.408 + ], + "angle": 0, + "content": "2.1. Illumination Spectrum Estimation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.416, + 0.905, + 0.536 + ], + "angle": 0, + "content": "Color constancy is achieved through two steps, including illumination estimation and chromatic adaptation. Early color constancy studies have focused on illumination estimation from RGB images. Statistics-based methods [8, 33, 62] and deep learning-based methods [1, 3, 4, 20, 36, 43, 56, 67] were developed to improve the accuracy of illumination estimation. These methods only used RGB images to separate illumination and reflectance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.538, + 0.906, + 0.674 + ], + "angle": 0, + "content": "In recent studies, MS images were used to address color constancy by exploiting rich spectral information, which allows for more accurate results. Previous methods extended well-established statistical models and relevant assumptions from RGB to MS spectral domains [13, 27, 61]. In [2, 10, 72], they separated illumination spectrum from surface reflectance using low-rank matrix factorizations. This approach enables the decomposition of the scene with few target objects." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Various learning-based illumination estimation techniques were studied for MS images. In [40], illumination spectrum estimation was modeled as a constrained matrix factorization problem. Spectral reflectance images required per-image calibration using a known reference or measurement device [39]. They used a synthesized training set with simulated lighting conditions. However, the limited number of true reflectance data relative to diverse lighting conditions would lead to performance degradation in complex scenes. In [15], an unsupervised method for segmenting illuminant regions was developed. They assumed the illumination spectrum to be piecewise constant as a probabilistic clustering problem. However, this assumption was relatively simple to address various illumination spectrum. In [63], an image segmentation method was used to lo" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2216" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.098, + 0.9, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.307, + 0.908, + 0.351 + ], + "angle": 0, + "content": "Figure 2. The proposed network uses several sub-networks including a spectral feature extractor, spectral unmixing block, and feature mixing block to generate spatial and spectral features, which are iteratively refined. The decoder uses reflectance decoder and illumination estimator, consisting of an IRSS framework, and estimates illumination spectrum as a result." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.376, + 0.484, + 0.513 + ], + "angle": 0, + "content": "calize regions dominantly affected by illumination within the image, which were then used for illumination estimation. However, the gaps between the segmentation results and key regions needed for accurate illumination estimation were problematic. In [14], an input MS image was divided into multiple triplets, processed using a pyramid-based convolutional color constancy block. However, the spectral information of illumination and surface reflectance among different channels was not fully exploited." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.536, + 0.273, + 0.552 + ], + "angle": 0, + "content": "2.2. Spectral Unmixing" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.564, + 0.483, + 0.806 + ], + "angle": 0, + "content": "The SU [25] has been used to extract unique spectrum from various materials in MS images. Previous SU methods used linear [5] and non-linear mixing models [19]. These methods are based on the assumption that each pixel can be expressed as a combination of endmember spectra and corresponding abundances. GMM [23, 46, 74] and deep learning-based SU methods [6, 16, 50, 52, 57, 59, 69] have been studied, extensively. GMM-based models [53] address spectral variability by modeling each endmember as a probabilistic distribution rather than a fixed value, tailored to real-world applications. Deep learning approaches, on the other hand, utilize neural networks to automatically learn the underlying structure of MS data. CNN-based [69] and autoencoder-based models [16, 50, 52, 59] have demonstrated effectiveness in enhancing unmixing accuracy at the cost of computational complexity." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.81, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Our method introduces a more refined, learnable SU model based on the GMM, which can model reflectance surface with high spectral variability. The spectral features is combined with a deep model and is further processed, which can adaptively address the spatial and spectral characteristics of input MS images." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.376, + 0.874, + 0.393 + ], + "angle": 0, + "content": "2.3. Attention Mechanism in Spectral Imaging" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.399, + 0.907, + 0.655 + ], + "angle": 0, + "content": "Attention mechanisms of transformers have been widely used to capture spectral characteristics in spectral imaging [11, 18, 37, 38, 55, 73]. Early studies employed spectral and spatial features independently. In [11, 55, 73], the spectral and spatial features are addressed with sequential processing. In [18, 37, 38], the features are processed in parallel using separated branches and combined with late fusion. While these methods were computationally efficient, they failed to effectively fuse the features for illumination spectrum estimation. Recently, cross-attention [51, 65] and 3D convolution [17, 24] were adopted to fuse the spectral and spatial features of MS images. More recently, transformer architectures have been more actively employed for this purpose [32, 41, 45]. Cross-attention mechanisms facilitate the modeling of global dependencies between spectral and spatial features, enabling the effective learning of complex feature relation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.669, + 0.688, + 0.687 + ], + "angle": 0, + "content": "3. Proposed Method" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.694, + 0.783, + 0.709 + ], + "angle": 0, + "content": "3.1. MS Image Formulation Model" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.715, + 0.906, + 0.747 + ], + "angle": 0, + "content": "In an IRSS framework [48], an MS image \\(\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times C}\\) with height \\(H\\), width \\(W\\), and channel \\(C\\) is modeled as" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.758, + 0.907, + 0.791 + ], + "angle": 0, + "content": "\\[\n\\mathbf {I} (x) = \\int_ {\\Lambda} \\mathbf {R} (x, \\lambda) \\cdot \\mathbf {L} (\\lambda) \\cdot \\mathbf {S} (\\lambda) d \\lambda , \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.855 + ], + "angle": 0, + "content": "where \\(\\mathbf{R}(x,\\lambda)\\) represents surface reflectance of a spatial coordinate \\(x\\) and wavelength \\(\\lambda\\) ranged within \\(\\Lambda\\). \\(\\mathbf{L}(\\lambda)\\) and \\(\\mathbf{S}(\\lambda) \\in \\mathbb{R}^C\\) represent illumination spectrum and camera sensitivity, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "In the joint IRSS problem, given camera sensitivity, it is necessary to accurately model reflectance surface to produce \\(\\mathbf{L}(\\lambda)\\). In our approach, the reflectance is modeled as" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2217" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.123 + ], + "angle": 0, + "content": "the product of abundances and endmembers through the SU, representing a detailed spectral characteristics, given as," + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.135, + 0.483, + 0.178 + ], + "angle": 0, + "content": "\\[\n\\mathbf {R} (x, \\lambda) = \\sum_ {k = 1} ^ {K} \\mathbf {U} _ {k} (x) \\cdot \\mathbf {E} _ {k} (\\lambda), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.185, + 0.485, + 0.232 + ], + "angle": 0, + "content": "where \\(\\mathbf{E}_k(\\lambda)\\) and \\(\\mathbf{U}_k(x)\\) denote the \\(k\\)-th endmember and its abundance map, respectively. \\(\\mathbf{I}(x)\\) is modeled by incorporating Eq. (2) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.114, + 0.248, + 0.483, + 0.29 + ], + "angle": 0, + "content": "\\[\n\\mathbf {I} (x) = \\int_ {\\Lambda} \\left(\\sum_ {k = 1} ^ {K} \\mathbf {U} _ {k} (x) \\cdot \\mathbf {E} _ {k} (\\lambda)\\right) \\cdot \\mathbf {L} (\\lambda) \\cdot \\mathbf {S} (\\lambda) d \\lambda , \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.302, + 0.484, + 0.363 + ], + "angle": 0, + "content": "where the SU-based reflectance surface model is used to provide more robust and flexible approach to accurately decompose the components. Fig. 2 presents this process with further details in the following subsections." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.373, + 0.484, + 0.403 + ], + "angle": 0, + "content": "3.2. Surface Reflectance Representation with Learnable SU" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.412, + 0.484, + 0.503 + ], + "angle": 0, + "content": "In the proposed technique, the SU is used to provide detailed spectrum profile of reflectance surface, and the GMM is employed to decompose observed spectrum into a set of endmember spectra and their corresponding abundances. The components are transformed to learnable tensors to accomplish a deep IRSS framework in the following." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.504, + 0.484, + 0.535 + ], + "angle": 0, + "content": "Based on the GMM, a probability density of an observed spectrum \\(\\mathbf{I}_x\\in \\mathbb{R}^C\\) at each pixel \\(x\\) is modeled as:" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.548, + 0.483, + 0.589 + ], + "angle": 0, + "content": "\\[\np \\left(\\mathbf {I} _ {x}\\right) = \\sum_ {k = 1} ^ {K} \\pi_ {k} \\mathcal {N} \\left(\\mathbf {I} _ {x} \\mid \\boldsymbol {\\mu} _ {k}, \\boldsymbol {\\Sigma} _ {k}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.596, + 0.483, + 0.642 + ], + "angle": 0, + "content": "where \\(K\\) is the number of normal distributions, and \\(\\pmb{\\mu}_k\\) and \\(\\pmb{\\Sigma}_k\\) are the mean vector and covariance matrix. \\(\\pi_k\\) is a weighting coefficient, in which \\(\\sum_{k=1}^{K} \\pi_k = 1\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.483, + 0.778 + ], + "angle": 0, + "content": "In Eq. (4), \\(p(\\mathbf{I}_x)\\) represents a likelihood that all \\(K\\) spectrum can explain the observed spectrum, quantified by the linear combination of \\(\\mathcal{N}(\\mathbf{I}_x|\\boldsymbol{\\mu}_k,\\boldsymbol{\\Sigma}_k)\\) and its proportion \\(\\pi_{k}\\) and thus \\(\\pmb{\\mu}_{k}\\) and \\(\\pmb{\\Sigma}_{k}\\) represent the \\(k\\)-th endmember spectrum and the spectral variability within the endmember. To derive the optimal parameter set \\(\\{\\pi_k^*,\\pmb {\\mu}_k^*,\\pmb {\\Sigma}_k^*\\}\\), given a set of observed spectra \\(\\mathbf{I}_x\\in \\mathbf{I}\\) of training samples, we apply an Expectation-Maximization (EM) algorithm [12] to iteratively refine the parameters." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.779, + 0.484, + 0.854 + ], + "angle": 0, + "content": "Based on the results, we directly obtain the end-member spectra from the mean vectors, i.e., \\(\\pmb{\\mu}^{*} = \\{\\pmb{\\mu}_{1}^{*},\\pmb{\\mu}_{2}^{*},\\dots ,\\pmb{\\mu}_{K}^{*}\\}\\). Then, we calculate a set of abundance \\(\\gamma^{*} = \\{\\gamma_{1}^{*},\\gamma_{2}^{*},\\dots ,\\gamma_{K}^{*}\\}\\). \\(\\gamma_k^* (x)\\) is computed at each pixel coordinate \\(x\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.866, + 0.483, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\gamma} _ {k} ^ {*} (x) = \\frac {\\pi_ {k} ^ {*} \\mathcal {N} \\left(\\mathbf {I} _ {x} \\mid \\boldsymbol {\\mu} _ {k} ^ {*} , \\boldsymbol {\\Sigma} _ {k} ^ {*}\\right)}{\\sum_ {j = 1} ^ {K} \\pi_ {j} ^ {*} \\mathcal {N} \\left(\\mathbf {I} _ {x} \\mid \\boldsymbol {\\mu} _ {j} ^ {*} , \\boldsymbol {\\Sigma} _ {j} ^ {*}\\right)}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "where \\(\\gamma_k^*(x)\\) consists of an abundance map for the \\(k\\)-th end-member spectrum." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.122, + 0.905, + 0.242 + ], + "angle": 0, + "content": "In previous studies [47, 72], surface reflectance has been approximated using a low-dimensional model based on a subspace assumption, where a small number of spectral bases (e.g., 3) is sufficient for most natural images. These findings could lead to a tight subspace expression of the surface reflectance for practical use cases. In our experiments, we empirically selected \\(K\\) and observed best results with \\(K = 5\\), aligning with the subspace assumption." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.243, + 0.905, + 0.289 + ], + "angle": 0, + "content": "In the proposed model, the parameters pass through learnable layers, allowing for a more flexible representation that adapts to real-world images, mathematically expressed," + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.313, + 0.905, + 0.332 + ], + "angle": 0, + "content": "\\[\n\\mathbf {E} = W _ {e} \\boldsymbol {\\mu} ^ {*} \\in \\mathbb {R} ^ {K \\times C}, \\mathbf {U} = W _ {u} \\gamma^ {*} \\in \\mathbb {R} ^ {K \\times H \\times W}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.342, + 0.905, + 0.388 + ], + "angle": 0, + "content": "where \\( W_{e} \\) and \\( W_{u} \\) are convolution layers. \\( \\mathbf{E} \\) and \\( \\mathbf{U} \\) are the tensors of endmembers and abundance maps, respectively, as the final output of the SU block." + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.401, + 0.897, + 0.584 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.596, + 0.907, + 0.652 + ], + "angle": 0, + "content": "Figure 3. Cross-attention operations in a feature mixing block. It uses input spectral and visual features from a SU block and CNN, respectively, and produces the spatial-spectral features as output, iteratively refined with an \\(N\\)-step GRU." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.666, + 0.882, + 0.683 + ], + "angle": 0, + "content": "3.3. Feature Mixing Block with Cross-Attention" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Feature mixing block (FMB) uses both the SU features and CNN features as input features. The CNN features are from a spatial feature extractor (SFE) which consists of 3 layers of convolutions, as illustrated in Fig. 2. Previous model-based methods neglected spatial information [2, 10, 72], degrading performance in complex scenes. In contrast, deep learning-based methods captured visual features but struggled to integrate physical modeling [14, 40]. In the proposed method, the SU features and CNN features are used complementarily to generate spatial-spectral features tailored for MS images. Fig. 3 presents the detailed operations in the FMB. The spectral and spatial features are fused using cross-attention and iteratively refined through a gated recurrent unit (GRU)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2218" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.182 + ], + "angle": 0, + "content": "The FMB employs a cross-attention mechanism of a transformer to fuse spectral and spatial features as shown in Fig. 3. In the process, given \\(\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times C}\\), the spatial features derived from CNN are used as queries, and the end-members from a spectral unmixing block are used as keys and values as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.144, + 0.199, + 0.483, + 0.216 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} = W _ {q} \\mathbf {F} _ {n}, \\quad \\mathbf {K} = W _ {k} \\mathbf {E}, \\quad \\mathbf {V} = W _ {v} \\mathbf {E}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.224, + 0.484, + 0.329 + ], + "angle": 0, + "content": "where \\(\\mathbf{Q} \\in \\mathbb{R}^{D \\times H \\times W}\\) is the query, and \\(\\mathbf{K}, \\mathbf{V} \\in \\mathbb{R}^{K \\times D}\\) are the key and value. \\(\\mathbf{F}_n\\) is a visual feature, which is iteratively refined using a \\(n\\)-step GRU. \\(\\mathbf{F}_0\\) is initialized to \\(\\mathbf{F}_{cnn}\\) as visual features of a SFE network. \\(D\\) is its spectral dimension. \\(W_q \\in \\mathbb{R}^{D \\times D}\\), and \\(W_k, W_v \\in \\mathbb{R}^{D \\times C}\\) are learnable linear transformations to embed the features to a cross-attention block." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.332, + 0.484, + 0.379 + ], + "angle": 0, + "content": "Spectral-wise correlations \\(\\mathbf{S} \\in \\mathbb{R}^{K \\times H \\times W}\\) between a query of spatial locations in \\(\\mathbf{F}_n\\) and a key of spectral characteristics of \\(\\mathbf{E}\\) are computed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.24, + 0.393, + 0.483, + 0.412 + ], + "angle": 0, + "content": "\\[\n\\mathbf {S} = \\mathbf {Q} \\cdot \\mathbf {K} ^ {T}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.422, + 0.483, + 0.452 + ], + "angle": 0, + "content": "where the attention scores calculate the relevance between the visual features and spectral features." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.453, + 0.484, + 0.514 + ], + "angle": 0, + "content": "These spectral correlations are modified by adding abundance map \\(\\mathbf{U} \\in \\mathbb{R}^{K \\times H \\times W}\\) obtained from the SU block, which generating an attention weight \\(\\mathbf{A} \\in \\mathbb{R}^{K \\times H \\times W}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.209, + 0.527, + 0.483, + 0.544 + ], + "angle": 0, + "content": "\\[\n\\mathbf {A} = \\operatorname {s o f t m a x} (\\mathbf {S} + \\mathbf {U}), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.554, + 0.484, + 0.601 + ], + "angle": 0, + "content": "where softmax is a softmax function. Then, we compute an attention output through a weighted sum between attention weights and values, given as:" + }, + { + "type": "equation", + "bbox": [ + 0.244, + 0.616, + 0.483, + 0.632 + ], + "angle": 0, + "content": "\\[\n\\mathbf {O} = \\mathbf {A} \\cdot \\mathbf {V}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.642, + 0.395, + 0.659 + ], + "angle": 0, + "content": "where \\(\\mathbf{O} \\in \\mathbb{R}^{D \\times H \\times W}\\) is the attention output." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.484, + 0.721 + ], + "angle": 0, + "content": "In the FMB, the output undergoes an update process using a GRU followed by a two-layer fully connected network (MLP), which iteratively refines the output feature with the query features. Mathematically, this can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.736, + 0.483, + 0.754 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {n + 1} = \\operatorname {M L P} \\left(\\operatorname {G R U} \\left(\\mathbf {F} _ {n}, \\mathbf {O}\\right)\\right), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.763, + 0.484, + 0.869 + ], + "angle": 0, + "content": "where \\(\\mathbf{F}_n\\) is updated for \\(N\\) iterations to fuse the spectral and spatial features, until the final output \\(\\mathbf{F}_N\\) is generated. \\(\\mathbf{F}_N\\) represents a refined reflectance feature, maintaining the same dimensions as the input CNN features while incorporating spectral information from the SU method. \\(N\\) is set to 6 in our experiments. The Fig. 4 shows the gradually refined features during the iterative attention process of FMB." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.9 + ], + "angle": 0, + "content": "We present a pseudo-code implementation in Algorithm 1." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.092, + 0.568, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.617, + 0.093, + 0.644, + 0.101 + ], + "angle": 0, + "content": "Step 0" + }, + { + "type": "image", + "bbox": [ + 0.568, + 0.099, + 0.669, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.695, + 0.093, + 0.72, + 0.101 + ], + "angle": 0, + "content": "Step 2" + }, + { + "type": "image", + "bbox": [ + 0.67, + 0.099, + 0.744, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.772, + 0.093, + 0.797, + 0.101 + ], + "angle": 0, + "content": "Step 4" + }, + { + "type": "image", + "bbox": [ + 0.746, + 0.099, + 0.82, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.846, + 0.093, + 0.871, + 0.101 + ], + "angle": 0, + "content": "Step 6" + }, + { + "type": "image", + "bbox": [ + 0.821, + 0.099, + 0.898, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.16, + 0.591, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.593, + 0.16, + 0.667, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.67, + 0.16, + 0.744, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.746, + 0.16, + 0.82, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.822, + 0.16, + 0.898, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.22, + 0.591, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.593, + 0.22, + 0.667, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.67, + 0.22, + 0.744, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.746, + 0.22, + 0.82, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.822, + 0.22, + 0.898, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.292, + 0.907, + 0.349 + ], + "angle": 0, + "content": "Figure 4. Progressive feature refinement visualization through iterative stages in FMB. The leftmost image shows a rendered RGB of input MS image. From the initial query feature of FMB, the features are gradually refined from step 0 to step 6." + }, + { + "type": "code_caption", + "bbox": [ + 0.526, + 0.371, + 0.832, + 0.387 + ], + "angle": 0, + "content": "Algorithm 1: SU-CNN Feature Mixing Block" + }, + { + "type": "algorithm", + "bbox": [ + 0.533, + 0.389, + 0.884, + 0.677 + ], + "angle": 0, + "content": "Input: \\(\\mathbf{F}_0\\gets \\mathbf{F}_{cnn}\\in \\mathbb{R}^{D\\times H\\times W}\\) : CNN feature \n\\(\\mathbf{E}\\in \\mathbb{R}^{K\\times C}\\) : Endmember from SU block \n\\(\\mathbf{U}\\in \\mathbb{R}^{K\\times H\\times W}\\) : Abundance map from SU block \nOutput: \\(\\mathbf{F}_{\\mathrm{ref}}\\leftarrow \\mathbf{F}_N\\in \\mathbb{R}^{D\\times H\\times W}\\) : Updated feature map for N iterations \n\\(\\mathbf{K}\\leftarrow W_{k}\\mathbf{E}\\) //Key projection \n\\(\\mathbf{V}\\leftarrow W_{v}\\mathbf{E}\\) //Value projection \nFunction FeatureMixing \\((\\mathbf{F}_0,\\mathbf{K},\\mathbf{V},\\mathbf{U})\\) .. \nfor \\(n = 0\\) to \\(N - 1\\) do \nQ←WqFn //Query projection \nS←Q·KT// Spectral correlation \nA←softmax(S+U) //Attention weights \nO←A·V //Attention output \nFn+1←MLP(GRU(Fn,O)) //Feature update with GRU and MLP \nend \nreturn FN \n\\(\\mathbf{F}_N\\gets\\) FeatureMixing(F0,K,V,U)" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.71, + 0.907, + 0.74 + ], + "angle": 0, + "content": "3.4. Illumination Spectrum Estimation in Decoder Blocks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.903 + ], + "angle": 0, + "content": "In decoder blocks, the proposed network generates an estimation of the illumination spectrum, which is denoted by \\(\\hat{\\mathbf{L}}\\), at the end. The network uses two branches to calculate not only the illumination spectrum estimation but also the surface reflectance approximation, which is denoted by \\(\\hat{\\mathbf{R}}\\). The approximation of the surface reflectance in the decoder enables the network to use a joint IRSS framework, rather than producing an exact surface reflectance in the absence of GT. As shown in Fig. 2, \\(\\mathbf{F}_N\\) is fed into a reflectance decoder, consisting of three transposed convolution layers, to" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2219" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.482, + 0.166 + ], + "angle": 0, + "content": "calculate \\(\\hat{\\mathbf{R}}\\) as intermediate results in the IRSS. Specifically, we use \\(3\\times 3\\) convolution kernels in the implementation. The reflectance decoder uses the only \\(\\mathbf{F}_N\\) as input to facilitate the optimization and fully focus on the reflectance prediction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.167, + 0.483, + 0.304 + ], + "angle": 0, + "content": "The illumination spectrum block is built with a series of convolutional and linear layers, in which an input is constructed by concatenating \\(\\mathbf{F}_{cnn}\\) and \\(\\mathbf{F}_N\\). The subsequent convolutional layers are employed to extract spatially localized spectral features, effectively capturing illumination in the target scene. We employ two layers of convolution kernel with stride 2. Rectified linear unit (ReLU) activation function is applied after all convolution and linear layers in the block, except for the final output linear layer." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.311, + 0.232, + 0.325 + ], + "angle": 0, + "content": "3.5. Loss function" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.335, + 0.483, + 0.38 + ], + "angle": 0, + "content": "For training the proposed architecture, we employ a reconstruction loss \\(\\mathcal{L}_{\\mathrm{RC}}\\) and an illumination estimation loss \\(\\mathcal{L}_{\\mathrm{AE}}\\). \\(\\mathcal{L}_{\\mathrm{RC}}\\) is computed using an \\(\\mathcal{L}_1\\) distance between \\(\\mathbf{I}\\) and \\(\\hat{\\mathbf{I}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.391, + 0.483, + 0.425 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {R C}} = \\sum_ {x \\in (h, w)} \\sum_ {\\lambda \\in \\Lambda} \\| \\hat {\\mathbf {I}} (x, \\lambda) - \\mathbf {I} (x, \\lambda) \\| _ {1}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.43, + 0.483, + 0.474 + ], + "angle": 0, + "content": "where the loss is computed over all spatial spectral locations, i.e., \\( h = 1 \\cdots H \\) and \\( w = 1 \\cdots W \\), and all channels \\( \\Lambda \\). \\( \\hat{\\mathbf{I}} \\) is reconstructed using an integral of \\( \\hat{\\mathbf{L}} \\) and \\( \\hat{\\mathbf{R}} \\)," + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.475, + 0.471, + 0.489 + ], + "angle": 0, + "content": "We also utilize an angular error (AE), \\(\\mathcal{L}_{\\mathrm{AE}}\\), defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.498, + 0.483, + 0.539 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {A E}} = \\arccos \\left(\\frac {\\hat {\\mathbf {L}} \\cdot \\mathbf {L}}{\\| \\hat {\\mathbf {L}} \\| \\| \\mathbf {L} \\|}\\right), \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.548, + 0.483, + 0.669 + ], + "angle": 0, + "content": "where \\(\\mathbf{L}\\) and \\(\\hat{\\mathbf{L}}\\) are the ground-truth (GT) and predicted illumination spectrum vector, respectively. For illumination spectrum estimation, capturing relative differences of illuminant among multi-channels is crucial. Spectral orientation defines its color characteristics, which would affect color constancy, while the intensity of illuminant reflects brightness. Thus, we use an angular error to focus on the orientation by measuring the angle between vectors." + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.67, + 0.368, + 0.683 + ], + "angle": 0, + "content": "The total loss function \\(\\mathcal{L}\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.211, + 0.699, + 0.482, + 0.715 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\alpha_ {1} \\mathcal {L} _ {\\mathrm {R C}} + \\alpha_ {2} \\mathcal {L} _ {\\mathrm {A E}}, \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.751 + ], + "angle": 0, + "content": "where \\(\\alpha_{1}\\) and \\(\\alpha_{2}\\) are weighting factors, which are set to 0.5 and 1, respectively, in our experiments." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.763, + 0.296, + 0.779 + ], + "angle": 0, + "content": "4. Experimental Results" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.787, + 0.353, + 0.803 + ], + "angle": 0, + "content": "4.1. Spectral Reflectance Datasets" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We evaluate the proposed technique using two spectral reflectance datasets: KAUST [40] and CAVE [66]. In the datasets, we use 31 channels ranging from \\(400nm\\) to \\(700nm\\) at \\(10nm\\) intervals. We relight the scenes by multiplying the surface reflectance of each pixel with an illuminant spectra matrix with 59 standard illuminants from" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.272 + ], + "angle": 0, + "content": "[49] resulting in a 31-channel spectral image. We randomly select \\(80\\%\\) of the KAUST reflectance and illuminant data for training and the remaining \\(20\\%\\) used for testing as in [40]. CAVE dataset is utilized only for testing the models. We evaluate the performance of the proposed technique with existing methods including both statistic-based and learning-based methods, GrayEdge [27], LRMF [72], ISNL [58], PWIR [54], DUN [40] and BeyondRGB [14]. We used mean- \\(\\Delta \\mathrm{A}_{\\mathrm{MS}}\\) and std- \\(\\Delta \\mathrm{A}_{\\mathrm{MS}}\\) for evaluation metrics, which are the average and standard deviation of an angular error measured between the GT and predicted illumination spectrum." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.273, + 0.907, + 0.41 + ], + "angle": 0, + "content": "Tab. 1 exhibits the performance comparisons, showing that the proposed technique achieves the best results on both datasets. The proposed technique produces 5.55 and 3.29 in terms of mean- \\(\\Delta \\mathrm{A}_{\\mathrm{MS}}\\) and std- \\(\\Delta \\mathrm{A}_{\\mathrm{MS}}\\), respectively, in KAUST dataset. For CAVE dataset, which is not used for model training, it shows reliable performance achieving 5.93 and 3.90 in mean- \\(\\Delta \\mathrm{A}_{\\mathrm{MS}}\\) and std- \\(\\Delta \\mathrm{A}_{\\mathrm{MS}}\\), respectively. These results demonstrate that the proposed technique holds stable performance on unseen reflectance data." + }, + { + "type": "table", + "bbox": [ + 0.524, + 0.422, + 0.898, + 0.641 + ], + "angle": 0, + "content": "
DatasetMethodmean-ΔAMS↓std-ΔAMS↓
KAUSTGrayEdge [27]11.468.02
LRMF [72]24.0623.49
ISNL [58]20.0511.46
PWIR [54]28.0721.20
DUN [40]9.176.30
BeyondRGB [14]7.304.45
Ours5.553.29
CAVEGrayEdge [27]18.3314.89
LRMF [72]20.0515.47
ISNL [58]22.9216.62
PWIR [54]34.9521.20
DUN [40]16.0413.75
BeyondRGB [14]6.643.62
Ours5.933.90
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.651, + 0.905, + 0.679 + ], + "angle": 0, + "content": "Table 1. Performance comparisons of the proposed technique and existing methods using KAUST and CAVE datasets." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.696, + 0.715, + 0.712 + ], + "angle": 0, + "content": "4.2. Beyond RGB Dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "The BeyondRGB [14] dataset is a comprehensive, real-world dataset, designed for MS image-related studies, focused on mobile devices. The dataset is organized into \"Lab\" and \"Field,\" which include a variety of objects with diverse colors and textures in indoor scenes and real-world data in outdoor scenes, both encompassing complex scenes with rich object content. For experimental results, we use the same data-split protocol of [14]. For comparisons, we used PWIR and BeyondRGB, because the other methods required reflectance data for training [40] or the input dimension needs to match the number of spectral bands in GT spectra [27, 58, 72]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2220" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.089, + 0.278, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.134, + 0.174, + 0.276, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.089, + 0.425, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.283, + 0.174, + 0.425, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.427, + 0.09, + 0.571, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.428, + 0.174, + 0.57, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.572, + 0.09, + 0.716, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.575, + 0.174, + 0.716, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.718, + 0.09, + 0.862, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.174, + 0.865, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.134, + 0.258, + 0.278, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.135, + 0.343, + 0.277, + 0.423 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.259, + 0.425, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.283, + 0.343, + 0.424, + 0.423 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.427, + 0.259, + 0.571, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.428, + 0.343, + 0.57, + 0.423 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.572, + 0.259, + 0.716, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.574, + 0.343, + 0.717, + 0.423 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.718, + 0.259, + 0.862, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.343, + 0.865, + 0.423 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.433, + 0.907, + 0.463 + ], + "angle": 0, + "content": "Figure 5. Each of two rows shows MS images of BeyondRGB dataset in the first row and their estimated illuminant spectrum in the second row. The results of our model and PWIR are plotted with red and green dashed lines, respectively. The GT is shown with black solid lines." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.477, + 0.484, + 0.63 + ], + "angle": 0, + "content": "Tab. 2 shows the accuracy of the tested methods. It is clearly demonstrated with the results that the proposed technique provides superior accuracy compared to all the tested methods. Our technique significantly reduces the angular errors, achieving values of 3.65 and 5.97 in the Lab and Field datasets, respectively, in mean- \\(\\Delta \\mathrm{A}_{\\mathrm{MS}}\\). In the BeyondRGB dataset, the proposed technique presents a greater performance gap compared to the second-best algorithm [14] than in KAUST and CAVE datasets, which can present the effectiveness of our approach in complex scenes." + }, + { + "type": "table", + "bbox": [ + 0.102, + 0.64, + 0.475, + 0.756 + ], + "angle": 0, + "content": "
DatasetMethodmean-ΔAMS↓std-ΔAMS↓
LabPWIR [54]27.1010.86
BeyondRGB [14]5.922.92
Ours3.652.72
FieldPWIR [54]16.316.07
BeyondRGB [14]7.225.54
Ours5.973.90
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.767, + 0.483, + 0.796 + ], + "angle": 0, + "content": "Table 2. Performance comparisons of the proposed technique and existing methods using real-world MS dataset, BeyondRGB." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Fig. 5 shows the results of illumination spectrum using the proposed technique in comparison with PWIR [54]. The top row of the image consists of Lab images, captured with controlled and artificial illuminations. From left to right, the illuminations are the spectrum of fluorescent, mixed LED, halogen lamp, incandescent light bulb, and LED illumi" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.477, + 0.907, + 0.675 + ], + "angle": 0, + "content": "nant, which are found in daily environments. The bottom row contains real-world Field images, where outdoor images are illuminated by natural sunlight. The estimated illumination spectrum of our model, represented by the red dashed line, demonstrate a significantly more detailed alignment with the shape and peaks of the GT compared to PWIR. In the first image of the Lab set, the proposed technique shows a superior performance in estimating complex illumination conditions, such as those with multiple peaks characteristic of fluorescent lighting. Even for the bottom row comprising a wider variety of scenes, our model demonstrates robust and accurate illumination estimation performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.682, + 0.78, + 0.699 + ], + "angle": 0, + "content": "4.3. Surface Reflectance Modeling" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We investigate the reconstruction results of surface reflectance on the KAUST and CAVE datasets, in which GT reflectance data is available. As shown in Fig. 6, we compare the estimated reflectance from our reflectance decoder and corresponding GT reflectance. From the results, we observe that even though the network has not directly seen GT reflectance during the training process, the surface reflectance is sufficiently considered while estimating illumination from images. Even in complex optical conditions like refractive surfaces with shadows shown in the window and glasses scenes, and reflection and diffraction effects demonstrated by the CD, our method effectively reconstruct the surface reflectance preserving these challenges." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "2221" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.088, + 0.482, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.283, + 0.485, + 0.354 + ], + "angle": 0, + "content": "Figure 6. Visualization of reconstructed surface reflectance for KAUST and CAVE datasets. The first and second columns display GT and reconstructed surface reflectance rendered in RGB. The following columns display the estimated surface reflectance at each wavelength, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.367, + 0.484, + 0.429 + ], + "angle": 0, + "content": "ing visual properties. The results align with our motivation for effective disentanglement of illumination and reflectance through physics-based approach and validate the effectiveness of our proposed method." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.436, + 0.245, + 0.453 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.459, + 0.483, + 0.732 + ], + "angle": 0, + "content": "Network architecture: Tab. 3 presents the effectiveness of the proposed modules consisting of the network architecture. When the modules are turned off one-by-one, we observe a considerable reduction in accuracy. First, without SU block, the network relies solely on visual features without using spectral features, resulting in an approximate increase of 1.12 in angular error. Second, when disabling FMB, the spatial and spectral features are not fused appropriately but used independently to estimate the illumination spectrum. It increases approximately 3.65 of an angular error. Lastly, when the SFE module is turned off, we observe a dramatic degradation in accuracy. This result is expected, as the network is unable to exploit any visual features from the input image without the SFE. Fig. 7 shows the output features of the FMB, when disabling SFE or SU. Compared to the rendered RGB image on the left, the synthesized feature can be effectively used for illumination estimation by localizing regions relevant to the task." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.742, + 0.319, + 0.824 + ], + "angle": 0, + "content": "
SUFMBSFEmean-ΔAMS ↓
X✓✓5.52
✓X✓8.05
✓✓X24.64
✓✓✓4.40
" + }, + { + "type": "table", + "bbox": [ + 0.331, + 0.742, + 0.484, + 0.824 + ], + "angle": 0, + "content": "
LAELCRcmean-ΔAMS↓
✓X4.72
X✓73.59
✓✓4.40
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.834, + 0.483, + 0.863 + ], + "angle": 0, + "content": "Table 3. Ablation studies of the sub-modules of the proposed network (left) and training loss functions (right)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Loss function: We conducted ablation tests in the loss terms. In the proposed method, using both \\(\\mathcal{L}_{\\mathrm{AE}}\\) and \\(\\mathcal{L}_{\\mathrm{RC}}\\)" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.091, + 0.905, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.334, + 0.908, + 0.406 + ], + "angle": 0, + "content": "Figure 7. Output features from an FMB block in the ablation test. The second and the third columns display the features, when disabling SFE and SU, respectively. Our features in the last column highlight more relevant regions including high spectral diversity in local surfaces to achieve more enhanced accuracy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.421, + 0.907, + 0.588 + ], + "angle": 0, + "content": "together achieves the best performance. When \\(\\mathcal{L}_{\\mathrm{RC}}\\) is not used, the performance decreases by approximately 0.32 compared to using both the losses. By incorporating \\(\\mathcal{L}_{\\mathrm{RC}}\\), the network can further learn the impact of a global illumination spectrum on the reconstructed image in terms of spatial representation. When excluding \\(\\mathcal{L}_{\\mathrm{AE}}\\), we observed substantial performance drops (e.g. 73 in mean-\\(\\Delta \\mathrm{A}_{\\mathrm{MS}}\\)). These results are expected, since relying solely on the reflectance for the image reconstruction may provide insufficient constraints for illumination estimation in the absence of GT. \\(\\mathcal{L}_{\\mathrm{AE}}\\) plays a crucial role in the estimation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.61, + 0.634, + 0.627 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.64, + 0.907, + 0.791 + ], + "angle": 0, + "content": "In this paper, we proposed illumination spectrum estimation technique for MS images. The proposed ISS technique used a learnable SU block which modeled surface reflectance, and a feature mixing block which fused spatial and spectral information using cross-attention operations. The spatial-spectral features were efficiently used to focus on important regions in complex scenes to enhance the accuracy of illumination estimation. Experimental results demonstrated that the proposed ISS technique provided a superiority to existing studies in various MS image datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In our future study, we will investigate illumination estimation under locally varying and multi-illuminant scenarios. While the current model is designed to estimate a global illumination spectrum, the proposed ISS technique can be extended to improve accuracy in estimating multiple illuminant by efficiently disentangling reflectance properties under locally varying lighting conditions." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2222" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.241, + 0.108 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.484, + 0.201 + ], + "angle": 0, + "content": "This work was partly supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (RS-2021-II212068, Artificial Intelligence Innovation Hub, \\(40\\%\\) ) and was partly supported by the NRF grant funded by MSIT (No.NRF-2022R1A2C4002052, \\(60\\%\\))." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.23, + 0.188, + 0.246 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.255, + 0.484, + 0.297 + ], + "angle": 0, + "content": "[1] Mahmoud Affifi and Michael S Brown. Sensor-independent illumination estimation for dnn models. arXiv preprint arXiv:1912.06888, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.299, + 0.484, + 0.354 + ], + "angle": 0, + "content": "[2] Dongsheng An, Jinli Suo, Haoqian Wang, and Qionghai Dai. Illumination estimation from specular highlight in a multispectral image. Optics Express, 23(13):17008-17023, 2015. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.357, + 0.483, + 0.412 + ], + "angle": 0, + "content": "[3] Simone Bianco, Claudio Cusano, and Raimondo Schettini. Color constancy using cnns. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 81-89, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.414, + 0.484, + 0.47 + ], + "angle": 0, + "content": "[4] Simone Bianco, Claudio Cusano, and Raimondo Schettini. Single and multiple illuminant estimation using convolutional neural networks. IEEE Transactions on Image Processing, 26(9):4347-4362, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.472, + 0.484, + 0.555 + ], + "angle": 0, + "content": "[5] José M Bioucas-Dias, Antonio Plaza, Nicolas Dobigeon, Mario Parente, Qian Du, Paul Gader, and Jocelyn Chanussot. Hyperspectral unmixing overview: Geometrical, statistical, and sparse regression-based approaches. IEEE journal of selected topics in applied earth observations and remote sensing, 5(2):354-379, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.557, + 0.484, + 0.625 + ], + "angle": 0, + "content": "[6] Ricardo Augusto Borsoi, Tales Imbiriba, and José Carlos Moreira Bermudez. Deep generative endmember modeling: An application to unsupervised spectral unmixing. IEEE Transactions on Computational Imaging, 6:374-384, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.629, + 0.484, + 0.669 + ], + "angle": 0, + "content": "[7] David H Brainard and Brian A Wandell. Analysis of the retina theory of color vision. JOSA A, 3(10):1651-1661, 1986. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.673, + 0.484, + 0.713 + ], + "angle": 0, + "content": "[8] Gershon Buchsbaum. A spatial processor model for object colour perception. Journal of the Franklin institute, 310(1): 1-26, 1980. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.716, + 0.484, + 0.771 + ], + "angle": 0, + "content": "[9] Ayan Chakrabarti, Keigo Hirakawa, and Todd Zickler. Color constancy with spatio-spectral statistics. IEEE Transactions on Pattern Analysis and Machine Intelligence, 34(8):1509-1519, 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.774, + 0.483, + 0.829 + ], + "angle": 0, + "content": "[10] Xiaochuan Chen, Mark S Drew, and Ze-Nian Li. Illumination and reflectance spectra separation of hyperspectral image data under multiple illumination conditions. *Electronic Imaging*, 29:194–199, 2017. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.831, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[11] Yushi Chen, Hanlu Jiang, Chunyang Li, Xiuping Jia, and Pedram Ghamisi. Deep feature extraction and classification of hyperspectral images based on convolutional neural networks. IEEE transactions on geoscience and remote sensing, 54(10):6232-6251, 2016. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.255, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.147 + ], + "angle": 0, + "content": "[12] Arthur P Dempster, Nan M Laird, and Donald B Rubin. Maximum likelihood from incomplete data via the em algorithm. Journal of the royal statistical society: series B (methodological), 39(1):1-22, 1977. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.906, + 0.205 + ], + "angle": 0, + "content": "[13] Arjan Gijsenij, Theo Gevers, and Joost Van De Weijer. Improving color constancy by photometric edge weighting. IEEE Transactions on Pattern Analysis and Machine Intelligence, 34(5):918-929, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.207, + 0.906, + 0.302 + ], + "angle": 0, + "content": "[14] Ortal Glatt, Yotam Ater, Woo-Shik Kim, Shira Werman, Oded Berby, Yael Zini, Shay Zelinger, Sangyoon Lee, Heejin Choi, and Evgeny Soloveichik. Beyond rgb: A real world dataset for multispectral imaging in mobile devices. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4344-4354, 2024. 1, 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.304, + 0.906, + 0.358 + ], + "angle": 0, + "content": "[15] Lin Gu, Cong Phuoc Huynh, and Antonio Robles-Kelly. Segmentation and estimation of spatially varying illumination. IEEE transactions on image processing, 23(8):3478-3489, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.361, + 0.906, + 0.417 + ], + "angle": 0, + "content": "[16] Rui Guo, Wei Wang, and Hairong Qi. Hyperspectral image unmixing using autoencoder cascade. In 2015 7th Workshop on Hyperspectral Image and Signal Processing: Evolution in Remote Sensing (WHISPERS), pages 1-4. IEEE, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.419, + 0.906, + 0.473 + ], + "angle": 0, + "content": "[17] Amina Ben Hamida, Alexandre Benoit, Patrick Lambert, and Chokri Ben Amar. 3-d deep learning approach for remote sensing image classification. IEEE Transactions on geoscience and remote sensing, 56(8):4420-4434, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.906, + 0.531 + ], + "angle": 0, + "content": "[18] Siyuan Hao, Wei Wang, Yuanxin Ye, Tingyuan Nie, and Lorenzo Bruzzone. Two-stream deep architecture for hyperspectral image classification. IEEE Transactions on Geoscience and Remote Sensing, 56(4):2349-2361, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.532, + 0.908, + 0.587 + ], + "angle": 0, + "content": "[19] Rob Heylen, Mario Parente, and Paul Gader. A review of nonlinear hyperspectral unmixing methods. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 7(6):1844-1868, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.589, + 0.906, + 0.644 + ], + "angle": 0, + "content": "[20] Yuanming Hu, Baoyuan Wang, and Stephen Lin. Fc4: Fully convolutional color constancy with confidence-weighted pooling. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4085-4094, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.646, + 0.906, + 0.686 + ], + "angle": 0, + "content": "[21] Anya Hurlbert. Challenges to color constancy in a contemporary light. Current Opinion in Behavioral Sciences, 30: 186-193, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.906, + 0.743 + ], + "angle": 0, + "content": "[22] Marian-Daniel Iordache, José M Bioucas-Dias, and Antonio Plaza. Sparse unmixing of hyperspectral data. IEEE Transactions on Geoscience and Remote Sensing, 49(6):2014-2039, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.906, + 0.801 + ], + "angle": 0, + "content": "[23] Qiwen Jin, Yong Ma, Erting Pan, Fan Fan, Jun Huang, Hao Li, Chenhong Sui, and Xiaoguang Mei. Hyperspectral unmixing with gaussian mixture model and spatial group sparsity. Remote Sensing, 11(20):2434, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.906, + 0.87 + ], + "angle": 0, + "content": "[24] Murali Kanthi, T Hitendra Sarma, and C Shobha Bindu. A 3d-deep cnn based feature extraction and hyperspectral image classification. In 2020 IEEE India Geoscience and Remote Sensing Symposium (InGARSS), pages 229-232. IEEE, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.872, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[25] Nirmal Keshava and John F Mustard. Spectral unmixing. IEEE signal processing magazine, 19(1):44-57, 2002. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2223" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[26] Haris Ahmad Khan. Multispectral constancy for illuminant invariant representation of multispectral images. PhD thesis, Université Bourgogne Franche-Comté; Norwegian University of Science and ..., 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.15, + 0.482, + 0.191 + ], + "angle": 0, + "content": "[27] Haris Ahmad Khan, Jean-Baptiste Thomas, Jon Yngve Hardeberg, and Olivier Lalignant. Illuminant estimation in multispectral imaging. JOSA A, 34(7):1085-1098, 2017. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.482, + 0.274 + ], + "angle": 0, + "content": "[28] Haris Ahmad Khan, Jean-Baptiste Thomas, and Jon Yngve Hardeberg. Towards highlight based illuminant estimation in multispectral images. In Image and Signal Processing: 8th International Conference, ICISP 2018, Cherbourg, France, July 2-4, 2018, Proceedings 8, pages 517-525. Springer, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.482, + 0.331 + ], + "angle": 0, + "content": "[29] Muhammad Jaleed Khan, Hamid Saeed Khan, Adeel Yousaf, Khurram Khurshid, and Asad Abbas. Modern trends in hyperspectral image analysis: A review. *IEEE Access*, 6:14118-14129, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.403 + ], + "angle": 0, + "content": "[30] Dongyoung Kim, Jinwoo Kim, Junsang Yu, and Seon Joo Kim. Attentive illumination decomposition model for multi-illuminant white balancing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25512–25521, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[31] Furkan Klinl, Doga Yilmaz, Baris Özcan, and Furkan Kiraç. Deterministic neural illumination mapping for efficient autowhite balance correction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1139-1147, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.482, + 0.531 + ], + "angle": 0, + "content": "[32] Fanqiang Kong, Tongbo Cao, Yunsong Li, Dan Li, and Kedi Hu. Multi-scale spatial-spectral attention network for multispectral image compression based on variational autoencoder. Signal Processing, 198:108589, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.482, + 0.559 + ], + "angle": 0, + "content": "[33] Edwin H Land. The retina theory of color vision. Scientific american, 237(6):108-129, 1977. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.482, + 0.629 + ], + "angle": 0, + "content": "[34] Edwin H Land. Recent advances in retina theory. In Central and Peripheral Mechanisms of Colour Vision: Proceedings of an International Symposium Held at The Wenner-Gren Center Stockholm, June 14–15, 1984, pages 5–17. Springer, 1985. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.482, + 0.659 + ], + "angle": 0, + "content": "[35] David Landgrebe. Hyperspectral image data analysis. IEEE Signal processing magazine, 19(1):17-28, 2002. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.482, + 0.729 + ], + "angle": 0, + "content": "[36] Bing Li, Haina Qin, Weihua Xiong, Yangxi Li, Songhe Feng, Weiming Hu, and Stephen Maybank. Ranking-based color constancy with limited training samples. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(10): 12304-12320, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.733, + 0.482, + 0.787 + ], + "angle": 0, + "content": "[37] Rui Li, Shunyi Zheng, Chenxi Duan, Yang Yang, and Xiqi Wang. Classification of hyperspectral image based on double-branch dual-attention mechanism network. Remote Sensing, 12(3):582, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[38] Xian Li, Mingli Ding, and Aleksandra Pizurica. Deep feature fusion via two-stream convolutional neural network for hyperspectral image classification. IEEE Transactions on Geoscience and Remote Sensing, 58(4):2615-2629, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[39] Yuqi Li, Chong Wang, Jieyu Zhao, and Qingshu Yuan. Efficient spectral reconstruction using a trichromatic camera via sample optimization. The Visual Computer, 34:1773-1783, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[40] Yuqi Li, Qiang Fu, and Wolfgang Heidrich. Multispectral illumination estimation using deep unrolling network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2672-2681, 2021. 1, 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.216 + ], + "angle": 0, + "content": "[41] Shuaiqi Liu, Siyuan Liu, Shichong Zhang, Bing Li, Weiming Hu, and Yu-Dong Zhang. Ssau-net: A spectral-spatial attention-based u-net for hyperspectral image fusion. IEEE Transactions on Geoscience and Remote Sensing, 60:1-16, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.219, + 0.905, + 0.273 + ], + "angle": 0, + "content": "[42] Zhenqi Liu, Xinyu Wang, Yanfei Zhong, Meng Shu, and Chen Sun. Siamhyper: Learning a hyperspectral object tracker from an rgb-based tracker. IEEE Transactions on Image Processing, 31:7116-7129, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.275, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[43] Yi-Chen Lo, Chia-Che Chang, Hsuan-Chao Chiu, Yu-Hao Huang, Chia-Ping Chen, Yu-Lin Chang, and Kevin Jou. Clcc: Contrastive learning for color constancy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8053-8063, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.345, + 0.905, + 0.385 + ], + "angle": 0, + "content": "[44] Zhongyu Lou, Theo Gevers, Ninghang Hu, Marcel P Lucassen, et al. Color constancy by deep learning. In BMVC, pages 76-1, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.387, + 0.905, + 0.427 + ], + "angle": 0, + "content": "[45] Qing Ma, Junjun Jiang, Xianming Liu, and Jiayi Ma. Reciprocal transformer for hyperspectral and multispectral image fusion. Information Fusion, 104:102148, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.429, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[46] Yong Ma, Qiwen Jin, Xiaoguang Mei, Xiaobing Dai, Fan Fan, Hao Li, and Jun Huang. Hyperspectral unmixing with gaussian mixture model and low-rank representation. Remote Sensing, 11(8):911, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.485, + 0.905, + 0.538 + ], + "angle": 0, + "content": "[47] Laurence Maloney. Evaluation of linear models of surface spectral reflectance with small numbers of parameters. In Physics-based vision: Principles and practice. Jones and Bartlett, 1992. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.541, + 0.905, + 0.594 + ], + "angle": 0, + "content": "[48] Laurence T Maloney and Brian A Wandell. Color constancy: a method for recovering surface spectral reflectance. In Readings in Computer Vision, pages 293-297. Elsevier, 1987. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.597, + 0.905, + 0.787 + ], + "angle": 0, + "content": "[49] Thomas Mansencal, Michael Mauderer, Michael Parsons, Nick Shaw, Kevin Wheatley, Sean Cooper, Jean D. Vandenberg, Luke Canavan, Katherine Crowson, Ofek Lev, Katrin Leinweber, Shriramana Sharma, Troy James Sobotka, Dominik Moritz, Matt Pppp, Chinmay Rane, Pavithra Eswaramoorthy, John Metic, Ben Pearlstine, Manuel Leonhardt, Olli Niemitalo, Marek Szymanski, Maximilian Schambach, Sianyi Huang, Mike Wei, Nishant Joywardhan, Omar Wagih, Pawel Redman, Joseph Goldstone, Stephen Hill, Jedediah Smith, Frederic Savoir, Geetansh Saxena, Saransh Chopra, Ilia Sibiryakov, Tim Gates, Gajendra Pal, Nicolas Tessore, Aurélien Pierre, François-Xavier Thomas, Sabarish Srinivasan, and Tucker Downs. Colour 0.4.2, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[50] Savas Ozkan, Berk Kaya, and Gozde Bozdagi Akar. Endnet: Sparse autoencoder network for endmember extraction and hyperspectral unmixing. IEEE Transactions on Geoscience and Remote Sensing, 57(1):482-496, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[51] Yishu Peng, Yuwen Zhang, Bing Tu, Qianming Li, and Wujing Li. Spatial-spectral transformer with cross-attention for hyperspectral image classification. IEEE Transactions on Geoscience and Remote Sensing, 60:1–15, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2224" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.146 + ], + "angle": 0, + "content": "[52] Ying Qu and Hairong Qi. udas: An untied denoising autoencoder with sparsity for spectral unmixing. IEEE Transactions on Geoscience and Remote Sensing, 57(3):1698-1712, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.175 + ], + "angle": 0, + "content": "[53] Douglas A Reynolds et al. Gaussian mixture models. Encyclopedia of biometrics, 741(659-663), 2009. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.482, + 0.245 + ], + "angle": 0, + "content": "[54] Antonio Robles-Kelly and Ran Wei. A convolutional neural network for pixelwise illuminant recovery in colour and spectral images. In 2018 24th International Conference on Pattern Recognition (ICPR), pages 109-114. IEEE, 2018. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.247, + 0.482, + 0.314 + ], + "angle": 0, + "content": "[55] Swalpa Kumar Roy, Gopal Krishna, Shiv Ram Dubey, and Bidyut B Chaudhuri. Hybridsn: Exploring 3-d-2-d cnn feature hierarchy for hyperspectral image classification. IEEE Geoscience and Remote Sensing Letters, 17(2):277-281, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.482, + 0.384 + ], + "angle": 0, + "content": "[56] Wu Shi, Chen Change Loy, and Xiaou Tang. Deep specialized network for illuminant estimation. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 371-387. Springer, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.385, + 0.482, + 0.44 + ], + "angle": 0, + "content": "[57] Alberto Signoroni, Mattia Savardi, Annalisa Baronio, and Sergio Benini. Deep learning meets hyperspectral image analysis: A multidisciplinary review. Journal of imaging, 5(5):52, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.482, + 0.495 + ], + "angle": 0, + "content": "[58] Tong Su, Yu Zhou, Yao Yu, Xun Cao, and Sidan Du. Illumination separation of non-lambertian scenes from a single hyperspectral image. Optics express, 26(20):26167-26178, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.497, + 0.482, + 0.564 + ], + "angle": 0, + "content": "[59] Yuanchao Su, Jun Li, Antonio Plaza, Andrea Marinoni, Paolo Gamba, and Somdatta Chakravortty. Daen: Deep autoencoder networks for hyperspectral unmixing. IEEE Transactions on Geoscience and Remote Sensing, 57(7): 4309-4321, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.482, + 0.62 + ], + "angle": 0, + "content": "[60] Hao Sun, Xiangtao Zheng, and Xiaogiang Lu. A supervised segmentation network for hyperspectral image classification. IEEE Transactions on Image Processing, 30:2810-2825, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.623, + 0.482, + 0.663 + ], + "angle": 0, + "content": "[61] Jean-Baptiste Thomas. Illuminant estimation from uncalibrated multispectral images. In 2015 Colour and Visual Computing Symposium (CVCS), pages 1-6. IEEE, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.665, + 0.482, + 0.704 + ], + "angle": 0, + "content": "[62] Joost Van De Weijer, Theo Gevers, and Arjan Gijsenij. Edge-based color constancy. IEEE Transactions on image processing, 16(9):2207-2214, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.707, + 0.482, + 0.76 + ], + "angle": 0, + "content": "[63] Donik Vršnak, Ilija Domislovic, Marko Subasic, and Sven Lončarić. Framework for illumination estimation and segmentation in multi-illuminant scenes. IEEE access, 11: 2128-2137, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.482, + 0.803 + ], + "angle": 0, + "content": "[64] Xia Xu and Zhenwei Shi. Multi-objective based spectral unmixing for hyperspectral images. ISPRS Journal of Photogrammetry and Remote Sensing, 124:54-69, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.805, + 0.482, + 0.858 + ], + "angle": 0, + "content": "[65] Kai Yang, Hao Sun, Chunbo Zou, and Xiaoqiang Lu. Cross-attention spectral-spatial network for hyperspectral image classification. IEEE Transactions on Geoscience and Remote Sensing, 60:1-14, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[66] Fumihito Yasuma, Tomoo Mitsunaga, Daisuke Iso, and Shree K Nayar. Generalized assorted pixel camera: post-capture control of resolution, dynamic range, and spectrum." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "IEEE transactions on image processing, 19(9):2241-2253, 2010.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[67] Huanglin Yu, Ke Chen, Kaiqi Wang, Yanlin Qian, Zhaoxiang Zhang, and Kui Jia. Cascading convolutional color constancy. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 12725-12732, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[68] Jinkai Zhang, Benoit Rivard, and Arturo Sanchez-Azofeifa. Derivative spectral unmixing of hyperspectral data applied to mixtures of lichen and rock. IEEE Transactions on Geoscience and Remote Sensing, 42(9):1934-1940, 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.905, + 0.29 + ], + "angle": 0, + "content": "[69] Xiangrong Zhang, Yujia Sun, Jingyan Zhang, Peng Wu, and Licheng Jiao. Hyperspectral unmixing via deep convolutional neural networks. IEEE Geoscience and Remote Sensing Letters, 15(11):1755-1759, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.292, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[70] Chunhui Zhao, Boao Qin, Shou Feng, Wenxiang Zhu, Weiwei Sun, Wei Li, and Xiuping Jia. Hyperspectral image classification with multi-attention transformer and adaptive superpixel segmentation-based active learning. IEEE Transactions on Image Processing, 32:3606-3621, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[71] Runxing Zhao, Zhiwen Wang, Wuyuan Guo, and Canlong Zhang. Multi-scene image enhancement based on multi-channel illumination estimation. Expert Systems with Applications, 226:120271, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.419, + 0.905, + 0.487 + ], + "angle": 0, + "content": "[72] Yinqiang Zheng, Imari Sato, and Yoichi Sato. Illumination and reflectance spectra separation of a hyperspectral image meets low-rank matrix factorization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1779-1787, 2015. 1, 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.489, + 0.905, + 0.558 + ], + "angle": 0, + "content": "[73] Zilong Zhong, Jonathan Li, Zhiming Luo, and Michael Chapman. Spectral-spatial residual network for hyperspectral image classification: A 3-d deep learning framework. IEEE Transactions on Geoscience and Remote Sensing, 56 (2):847-858, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.56, + 0.905, + 0.615 + ], + "angle": 0, + "content": "[74] Yuan Zhou, Anand Rangarajan, and Paul D Gader. A gaussian mixture model representation of endmember variability in hyperspectral unmixing. IEEE Transactions on Image Processing, 27(5):2242-2256, 2018. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.615 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2225" + } + ] +] \ No newline at end of file diff --git a/2025/Illumination Spectrum Estimation for Multispectral Images via Surface Reflectance Modeling and Spatial-Spectral Feature Generation/afff26f0-40a0-4b11-8350-630ae1c85fad_origin.pdf b/2025/Illumination Spectrum Estimation for Multispectral Images via Surface Reflectance Modeling and Spatial-Spectral Feature Generation/afff26f0-40a0-4b11-8350-630ae1c85fad_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1e4b785aae964ab40cf672a2de833100e22eb4f8 --- /dev/null +++ b/2025/Illumination Spectrum Estimation for Multispectral Images via Surface Reflectance Modeling and Spatial-Spectral Feature Generation/afff26f0-40a0-4b11-8350-630ae1c85fad_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:66774efc9fee1a281bddea1d17e7d51e3998d1dd55f8d9bfa34ea2379e6e2f71 +size 3002949 diff --git a/2025/Illumination Spectrum Estimation for Multispectral Images via Surface Reflectance Modeling and Spatial-Spectral Feature Generation/full.md b/2025/Illumination Spectrum Estimation for Multispectral Images via Surface Reflectance Modeling and Spatial-Spectral Feature Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2b20210cf64c393b748c8d0336d91c69fb789476 --- /dev/null +++ b/2025/Illumination Spectrum Estimation for Multispectral Images via Surface Reflectance Modeling and Spatial-Spectral Feature Generation/full.md @@ -0,0 +1,453 @@ +# Illumination Spectrum Estimation for Multispectral Images via Surface Reflectance Modeling and Spatial-Spectral Feature Generation + +Hyejin Oh $^{1}$ Woo-Shik Kim $^{2}$ Sangyoon Lee $^{2}$ YungKyung Park $^{3}$ Je-Won Kang $^{1*}$ $^{1}$ Dept. of Electronic and Electrical Engineering, Ewha W. University, Seoul, South Korea + $^{2}$ Samsung Advanced Institute of Technology, South Korea + $^{3}$ Dept.of Color Design, Ewha W. University, Seoul, South Korea + +{hyejin5, yungkyung PARK, jewonk}@ewha.ac.kr {wooshik.kim, sangy00n.lee}@samsung.com + +# Abstract + +Multispectral (MS) images contain richer spectral information than RGB images due to their increased number of channels and are widely used for various applications. However, achieving accurate estimation in MS images remains challenging, as previous studies have struggled with spectral diversity and the inherent entanglement between the illuminant and surface reflectance spectra. To tackle these challenges, in this paper, we propose a novel Illumination spectrum estimation technique for MS images via Surface reflectance modeling and Spatial-spectral feature generation (ISS). The proposed technique employs a learnable spectral unmixing (SU) block to enhance surface reflectance modeling, which was unattempted in the illumination spectrum estimation, and a feature mixing block to fuse spectral and spatial features of MS images with cross-attention. The features are refined iteratively and processed through a decoder to produce an illumination spectrum estimator. Experimental results demonstrate that the proposed technique achieves state-of-the-art performance in illumination spectrum estimation in various MS image datasets. The code is available at https://github.com/heyjinnii/ISS-MSI.git. + +# 1. Introduction + +Multispectral (MS) imaging sensors capture a broader range of spectral bands and provide richer information than in RGB images [29, 35]. As these imaging sensors become compact and cost-effective [14], they are actively used in various image processing applications [3, 42, 44, 60, 70]. In computer vision, computational color constancy is used to replicate the perceptual stability of the human visual system, enabling reliable color rendering under varying lighting conditions [7, 34]. In MS imaging, the color constancy + +![](images/636e9b20a64b06d4bffd846412b0cef04956e6e9f155016d30eaba5c6d550ec1.jpg) + +![](images/4032483957fb5fc17224a778d34dcbfce0e6fafb93ac627bbca60205383ccb79.jpg) +Figure 1. Our model synthesizes spectral features derived from surface reflectance modeling with CNN features, iteratively refining the fused representation in an IRSS framework. The features enhance the accuracy of illumination spectrum estimation by effectively accounting for the reflectance characteristics of objects within the image, as demonstrated in the last column, in contrast to the conventional deep features shown in the middle column. + +![](images/395f66e907a50ffd92103965006be6acee628de20f13a870873079cf6be1ce7c.jpg) + +![](images/c38296a263f3f74f1cc44f28ba0bf934f870481f798eac706dcb643336ed849d.jpg) + +![](images/db3cceb5902486ee1e51ed8dcd1b618b7d0d9343d378eba952e7ba023ed83109.jpg) + +![](images/da430e92130ddefa98959edb2e902938ab304e53bbb66958339fcc1e0e80553f.jpg) + +becomes even more crucial than in the conventional RGB imaging [14, 26, 28, 40, 54, 72], because MS images with the increased number of channels are highly sensitive to illumination changes. The color constancy is achieved by solving an illumination and reflectance spectra separation (IRSS) problem [48], which is an underconstrained problem, requiring the simultaneous recovery of both the illumination spectrum and surface reflectance [72]. Accordingly, achieving accurate estimation in MS images remains a challenge [9, 21, 72]. + +Recent advances in sensor technology have enabled the acquisition of high-resolution, high-quality MS images [14]. The separation is more challenging in the real-world MS images, because both illumination and reflectance contribute more intricately to the observed spectrum and can vary more dynamically over their channels [30, 31, 71]. + +Previous studies have attempted to address the IRSS problem by imposing several constraints to MS images. In [2, 10, 72], scene illumination was separated from surface reflectance using low-rank matrix factorizations. However, relying on low-dimensional subspaces of surface reflectance would incur instability when the number of channels is insufficient relative to the basis functions [72]. Despite the limited research to date, deep learning techniques have been applied to MS images for illumination spectrum estimation [14, 40]. However, the previous deep models have largely overlooked the joint analysis of the illumination spectrum and surface reflectance. + +To overcome the limitations of existing methods, we propose a comprehensive deep learning-based IRSS framework, dubbed Illumination spectrum estimation for MS images via Surface reflectance modeling and Spatial-spectral feature generation (ISS). Spectral unmixing (SU) [25] is employed in this framework to decompose the measured spectrum into a set of endmembers, representing the spectral reflectance characteristics of specific surfaces, and abundances, which indicate the proportion of each endmember. SU has been widely used to analyze MS images in various fields, such as material science [68] and remote sensing [22, 64], yet few studies have been conducted on illumination spectrum estimation. In the proposed technique, rather than deriving the exact surface reflectance of objects, which is practically challenging, we focus on integrating refined reflectance surface modeling into a deep learning-based illumination spectrum estimation approach. Gaussian Mixture Model (GMM) was applied to generate robust endmembers and abundances [46, 53, 74]. We extend this approach by incorporating learnable parameters for endmembers and abundances, providing a robust and flexible decomposition method. + +Existing learning-based methods have attempted to build an IRSS framework in their models. However, they may suffer from a limited amount of reflectance data [39] or rely on image prior such as segmentation masks [63] and clusters [15] that are not closely related to illumination spectrum estimation. Their methods were challenging to address diverse illumination spectrum and complicated scenes. In this paper, we propose a feature mixing block that combines the results of endmembers and abundance maps with deep visual features, fully exploiting both spatial information and spectral characteristics of the surface reflectance in the input MS image. Conventional deep learning-based feature estimation methods, commonly employed for illumination estimation, often focus on specific bright areas within a single object and misclassify irrelevant regions as distinct features. As shown in Fig. 1, even in shaded areas where previous methods would struggle, the proposed ISS technique appropriately captures the relevant regions for illumination spectrum estimation, treating both bright and shaded areas + +as the parts of the same material. The features generated by the feature mixing block are fed into a decoder to reconstruct a reflectance map and into an illumination estimator to establish the IRSS framework, finally estimating illumination spectrum of an MS image in the output. + +The contributions of this paper are as follows: + +- We propose a novel deep learning-based IRSS framework to accurately estimate illumination spectrum, addressing the challenges of spectral diversity in MS images. +- We introduce a learnable SU module and a feature mixing block, which combines refined spectral features from surface reflectance modeling with deep visual features using cross-attention, effectively capturing both spatial and spectral characteristics of MS images. +- We evaluate the performance of the proposed technique on various image datasets, demonstrating its superiority to existing studies. + +# 2. Related Works + +# 2.1. Illumination Spectrum Estimation + +Color constancy is achieved through two steps, including illumination estimation and chromatic adaptation. Early color constancy studies have focused on illumination estimation from RGB images. Statistics-based methods [8, 33, 62] and deep learning-based methods [1, 3, 4, 20, 36, 43, 56, 67] were developed to improve the accuracy of illumination estimation. These methods only used RGB images to separate illumination and reflectance. + +In recent studies, MS images were used to address color constancy by exploiting rich spectral information, which allows for more accurate results. Previous methods extended well-established statistical models and relevant assumptions from RGB to MS spectral domains [13, 27, 61]. In [2, 10, 72], they separated illumination spectrum from surface reflectance using low-rank matrix factorizations. This approach enables the decomposition of the scene with few target objects. + +Various learning-based illumination estimation techniques were studied for MS images. In [40], illumination spectrum estimation was modeled as a constrained matrix factorization problem. Spectral reflectance images required per-image calibration using a known reference or measurement device [39]. They used a synthesized training set with simulated lighting conditions. However, the limited number of true reflectance data relative to diverse lighting conditions would lead to performance degradation in complex scenes. In [15], an unsupervised method for segmenting illuminant regions was developed. They assumed the illumination spectrum to be piecewise constant as a probabilistic clustering problem. However, this assumption was relatively simple to address various illumination spectrum. In [63], an image segmentation method was used to lo + +![](images/591824fc857145bd88ff8c36e1088314c48025d3841c0e8909b4979a62f60478.jpg) +Figure 2. The proposed network uses several sub-networks including a spectral feature extractor, spectral unmixing block, and feature mixing block to generate spatial and spectral features, which are iteratively refined. The decoder uses reflectance decoder and illumination estimator, consisting of an IRSS framework, and estimates illumination spectrum as a result. + +calize regions dominantly affected by illumination within the image, which were then used for illumination estimation. However, the gaps between the segmentation results and key regions needed for accurate illumination estimation were problematic. In [14], an input MS image was divided into multiple triplets, processed using a pyramid-based convolutional color constancy block. However, the spectral information of illumination and surface reflectance among different channels was not fully exploited. + +# 2.2. Spectral Unmixing + +The SU [25] has been used to extract unique spectrum from various materials in MS images. Previous SU methods used linear [5] and non-linear mixing models [19]. These methods are based on the assumption that each pixel can be expressed as a combination of endmember spectra and corresponding abundances. GMM [23, 46, 74] and deep learning-based SU methods [6, 16, 50, 52, 57, 59, 69] have been studied, extensively. GMM-based models [53] address spectral variability by modeling each endmember as a probabilistic distribution rather than a fixed value, tailored to real-world applications. Deep learning approaches, on the other hand, utilize neural networks to automatically learn the underlying structure of MS data. CNN-based [69] and autoencoder-based models [16, 50, 52, 59] have demonstrated effectiveness in enhancing unmixing accuracy at the cost of computational complexity. + +Our method introduces a more refined, learnable SU model based on the GMM, which can model reflectance surface with high spectral variability. The spectral features is combined with a deep model and is further processed, which can adaptively address the spatial and spectral characteristics of input MS images. + +# 2.3. Attention Mechanism in Spectral Imaging + +Attention mechanisms of transformers have been widely used to capture spectral characteristics in spectral imaging [11, 18, 37, 38, 55, 73]. Early studies employed spectral and spatial features independently. In [11, 55, 73], the spectral and spatial features are addressed with sequential processing. In [18, 37, 38], the features are processed in parallel using separated branches and combined with late fusion. While these methods were computationally efficient, they failed to effectively fuse the features for illumination spectrum estimation. Recently, cross-attention [51, 65] and 3D convolution [17, 24] were adopted to fuse the spectral and spatial features of MS images. More recently, transformer architectures have been more actively employed for this purpose [32, 41, 45]. Cross-attention mechanisms facilitate the modeling of global dependencies between spectral and spatial features, enabling the effective learning of complex feature relation. + +# 3. Proposed Method + +# 3.1. MS Image Formulation Model + +In an IRSS framework [48], an MS image $\mathbf{I} \in \mathbb{R}^{H \times W \times C}$ with height $H$ , width $W$ , and channel $C$ is modeled as + +$$ +\mathbf {I} (x) = \int_ {\Lambda} \mathbf {R} (x, \lambda) \cdot \mathbf {L} (\lambda) \cdot \mathbf {S} (\lambda) d \lambda , \tag {1} +$$ + +where $\mathbf{R}(x,\lambda)$ represents surface reflectance of a spatial coordinate $x$ and wavelength $\lambda$ ranged within $\Lambda$ . $\mathbf{L}(\lambda)$ and $\mathbf{S}(\lambda) \in \mathbb{R}^C$ represent illumination spectrum and camera sensitivity, respectively. + +In the joint IRSS problem, given camera sensitivity, it is necessary to accurately model reflectance surface to produce $\mathbf{L}(\lambda)$ . In our approach, the reflectance is modeled as + +the product of abundances and endmembers through the SU, representing a detailed spectral characteristics, given as, + +$$ +\mathbf {R} (x, \lambda) = \sum_ {k = 1} ^ {K} \mathbf {U} _ {k} (x) \cdot \mathbf {E} _ {k} (\lambda), \tag {2} +$$ + +where $\mathbf{E}_k(\lambda)$ and $\mathbf{U}_k(x)$ denote the $k$ -th endmember and its abundance map, respectively. $\mathbf{I}(x)$ is modeled by incorporating Eq. (2) as follows: + +$$ +\mathbf {I} (x) = \int_ {\Lambda} \left(\sum_ {k = 1} ^ {K} \mathbf {U} _ {k} (x) \cdot \mathbf {E} _ {k} (\lambda)\right) \cdot \mathbf {L} (\lambda) \cdot \mathbf {S} (\lambda) d \lambda , \tag {3} +$$ + +where the SU-based reflectance surface model is used to provide more robust and flexible approach to accurately decompose the components. Fig. 2 presents this process with further details in the following subsections. + +# 3.2. Surface Reflectance Representation with Learnable SU + +In the proposed technique, the SU is used to provide detailed spectrum profile of reflectance surface, and the GMM is employed to decompose observed spectrum into a set of endmember spectra and their corresponding abundances. The components are transformed to learnable tensors to accomplish a deep IRSS framework in the following. + +Based on the GMM, a probability density of an observed spectrum $\mathbf{I}_x\in \mathbb{R}^C$ at each pixel $x$ is modeled as: + +$$ +p \left(\mathbf {I} _ {x}\right) = \sum_ {k = 1} ^ {K} \pi_ {k} \mathcal {N} \left(\mathbf {I} _ {x} \mid \boldsymbol {\mu} _ {k}, \boldsymbol {\Sigma} _ {k}\right), \tag {4} +$$ + +where $K$ is the number of normal distributions, and $\pmb{\mu}_k$ and $\pmb{\Sigma}_k$ are the mean vector and covariance matrix. $\pi_k$ is a weighting coefficient, in which $\sum_{k=1}^{K} \pi_k = 1$ . + +In Eq. (4), $p(\mathbf{I}_x)$ represents a likelihood that all $K$ spectrum can explain the observed spectrum, quantified by the linear combination of $\mathcal{N}(\mathbf{I}_x|\boldsymbol{\mu}_k,\boldsymbol{\Sigma}_k)$ and its proportion $\pi_{k}$ and thus $\pmb{\mu}_{k}$ and $\pmb{\Sigma}_{k}$ represent the $k$ -th endmember spectrum and the spectral variability within the endmember. To derive the optimal parameter set $\{\pi_k^*,\pmb {\mu}_k^*,\pmb {\Sigma}_k^*\}$ , given a set of observed spectra $\mathbf{I}_x\in \mathbf{I}$ of training samples, we apply an Expectation-Maximization (EM) algorithm [12] to iteratively refine the parameters. + +Based on the results, we directly obtain the end-member spectra from the mean vectors, i.e., $\pmb{\mu}^{*} = \{\pmb{\mu}_{1}^{*},\pmb{\mu}_{2}^{*},\dots ,\pmb{\mu}_{K}^{*}\}$ . Then, we calculate a set of abundance $\gamma^{*} = \{\gamma_{1}^{*},\gamma_{2}^{*},\dots ,\gamma_{K}^{*}\}$ . $\gamma_k^* (x)$ is computed at each pixel coordinate $x$ as follows: + +$$ +\boldsymbol {\gamma} _ {k} ^ {*} (x) = \frac {\pi_ {k} ^ {*} \mathcal {N} \left(\mathbf {I} _ {x} \mid \boldsymbol {\mu} _ {k} ^ {*} , \boldsymbol {\Sigma} _ {k} ^ {*}\right)}{\sum_ {j = 1} ^ {K} \pi_ {j} ^ {*} \mathcal {N} \left(\mathbf {I} _ {x} \mid \boldsymbol {\mu} _ {j} ^ {*} , \boldsymbol {\Sigma} _ {j} ^ {*}\right)}, \tag {5} +$$ + +where $\gamma_k^*(x)$ consists of an abundance map for the $k$ -th end-member spectrum. + +In previous studies [47, 72], surface reflectance has been approximated using a low-dimensional model based on a subspace assumption, where a small number of spectral bases (e.g., 3) is sufficient for most natural images. These findings could lead to a tight subspace expression of the surface reflectance for practical use cases. In our experiments, we empirically selected $K$ and observed best results with $K = 5$ , aligning with the subspace assumption. + +In the proposed model, the parameters pass through learnable layers, allowing for a more flexible representation that adapts to real-world images, mathematically expressed, + +$$ +\mathbf {E} = W _ {e} \boldsymbol {\mu} ^ {*} \in \mathbb {R} ^ {K \times C}, \mathbf {U} = W _ {u} \gamma^ {*} \in \mathbb {R} ^ {K \times H \times W}, \tag {6} +$$ + +where $W_{e}$ and $W_{u}$ are convolution layers. $\mathbf{E}$ and $\mathbf{U}$ are the tensors of endmembers and abundance maps, respectively, as the final output of the SU block. + +![](images/10755f8188b178b33a2e1d42fea7d6230d7354d8e393e4edbffbcadf40a40232.jpg) +Figure 3. Cross-attention operations in a feature mixing block. It uses input spectral and visual features from a SU block and CNN, respectively, and produces the spatial-spectral features as output, iteratively refined with an $N$ -step GRU. + +# 3.3. Feature Mixing Block with Cross-Attention + +Feature mixing block (FMB) uses both the SU features and CNN features as input features. The CNN features are from a spatial feature extractor (SFE) which consists of 3 layers of convolutions, as illustrated in Fig. 2. Previous model-based methods neglected spatial information [2, 10, 72], degrading performance in complex scenes. In contrast, deep learning-based methods captured visual features but struggled to integrate physical modeling [14, 40]. In the proposed method, the SU features and CNN features are used complementarily to generate spatial-spectral features tailored for MS images. Fig. 3 presents the detailed operations in the FMB. The spectral and spatial features are fused using cross-attention and iteratively refined through a gated recurrent unit (GRU). + +The FMB employs a cross-attention mechanism of a transformer to fuse spectral and spatial features as shown in Fig. 3. In the process, given $\mathbf{I} \in \mathbb{R}^{H \times W \times C}$ , the spatial features derived from CNN are used as queries, and the end-members from a spectral unmixing block are used as keys and values as follows: + +$$ +\mathbf {Q} = W _ {q} \mathbf {F} _ {n}, \quad \mathbf {K} = W _ {k} \mathbf {E}, \quad \mathbf {V} = W _ {v} \mathbf {E}, \tag {7} +$$ + +where $\mathbf{Q} \in \mathbb{R}^{D \times H \times W}$ is the query, and $\mathbf{K}, \mathbf{V} \in \mathbb{R}^{K \times D}$ are the key and value. $\mathbf{F}_n$ is a visual feature, which is iteratively refined using a $n$ -step GRU. $\mathbf{F}_0$ is initialized to $\mathbf{F}_{cnn}$ as visual features of a SFE network. $D$ is its spectral dimension. $W_q \in \mathbb{R}^{D \times D}$ , and $W_k, W_v \in \mathbb{R}^{D \times C}$ are learnable linear transformations to embed the features to a cross-attention block. + +Spectral-wise correlations $\mathbf{S} \in \mathbb{R}^{K \times H \times W}$ between a query of spatial locations in $\mathbf{F}_n$ and a key of spectral characteristics of $\mathbf{E}$ are computed as follows: + +$$ +\mathbf {S} = \mathbf {Q} \cdot \mathbf {K} ^ {T}, \tag {8} +$$ + +where the attention scores calculate the relevance between the visual features and spectral features. + +These spectral correlations are modified by adding abundance map $\mathbf{U} \in \mathbb{R}^{K \times H \times W}$ obtained from the SU block, which generating an attention weight $\mathbf{A} \in \mathbb{R}^{K \times H \times W}$ as follows: + +$$ +\mathbf {A} = \operatorname {s o f t m a x} (\mathbf {S} + \mathbf {U}), \tag {9} +$$ + +where softmax is a softmax function. Then, we compute an attention output through a weighted sum between attention weights and values, given as: + +$$ +\mathbf {O} = \mathbf {A} \cdot \mathbf {V}, \tag {10} +$$ + +where $\mathbf{O} \in \mathbb{R}^{D \times H \times W}$ is the attention output. + +In the FMB, the output undergoes an update process using a GRU followed by a two-layer fully connected network (MLP), which iteratively refines the output feature with the query features. Mathematically, this can be expressed as: + +$$ +\mathbf {F} _ {n + 1} = \operatorname {M L P} \left(\operatorname {G R U} \left(\mathbf {F} _ {n}, \mathbf {O}\right)\right), \tag {11} +$$ + +where $\mathbf{F}_n$ is updated for $N$ iterations to fuse the spectral and spatial features, until the final output $\mathbf{F}_N$ is generated. $\mathbf{F}_N$ represents a refined reflectance feature, maintaining the same dimensions as the input CNN features while incorporating spectral information from the SU method. $N$ is set to 6 in our experiments. The Fig. 4 shows the gradually refined features during the iterative attention process of FMB. + +We present a pseudo-code implementation in Algorithm 1. + +![](images/c0988fd10e984b7d9b6b59d3f0d366cfdaaf6b8ed0bcb9e1b88f56cabb5c219d.jpg) + +![](images/431289af9814dc71ef82f6ea1cd1be6029701b2d8a88f17d12ead44af8b73b12.jpg) +Step 0 + +![](images/ab919df20e4a64a859b9bbad83f919bbafd1c70d982ae1ef0338c2ced7165700.jpg) +Step 2 + +![](images/b55cfaaa440ecd621c7f69e081b013ef357b119573c56294e185337f319709c5.jpg) +Step 4 + +![](images/18b5f54d93b0fa10f0ed547ec3dfd8d91dda5e3b50b0fcdcd607a6b144bc0788.jpg) +Step 6 + +![](images/2ac9985a1f676375c4cc41e7db2192299294c12b3fa4ad17c56033f0a5aeebba.jpg) + +![](images/4749dbe1c0ad4f659388b234c7d3a97332abe5f4cd8da8bc9d65ec60f1a3a3de.jpg) + +![](images/1692a0d17ba44fcda318e61d76919a70f6913785c478d4ce0b57f43883f912c7.jpg) + +![](images/00a10150855f298dbf6bd286ed4b2cf3e313babd847e444a2776baaa3434a8eb.jpg) + +![](images/81ee8b27128cbd811d6e1e256737e34687487b6e25dfe3ee7d514f97716d5dcf.jpg) + +![](images/693f936579181e53516e3ab149d5052e388e28dfe3e0b2d0c78b7d7d97501cfd.jpg) +Figure 4. Progressive feature refinement visualization through iterative stages in FMB. The leftmost image shows a rendered RGB of input MS image. From the initial query feature of FMB, the features are gradually refined from step 0 to step 6. + +![](images/5b2aa6b73ee539e167fae80cf8a024eace372fd6a5d61bc1ecb93a471c11e3c8.jpg) + +![](images/f32866ff76c3416531b348f48bed73547720ac9a596110c4f98d403a676d3b77.jpg) + +![](images/d4709571de54ef66bc34414813c6840056323c653589e63e7003797d8746feba.jpg) + +![](images/f6fd1d05109772f3914010b5f5557ea004db5f47a1b9ca46fed30988e80c1083.jpg) + +Algorithm 1: SU-CNN Feature Mixing Block +Input: $\mathbf{F}_0\gets \mathbf{F}_{cnn}\in \mathbb{R}^{D\times H\times W}$ : CNN feature + $\mathbf{E}\in \mathbb{R}^{K\times C}$ : Endmember from SU block + $\mathbf{U}\in \mathbb{R}^{K\times H\times W}$ : Abundance map from SU block +Output: $\mathbf{F}_{\mathrm{ref}}\leftarrow \mathbf{F}_N\in \mathbb{R}^{D\times H\times W}$ : Updated feature map for N iterations + $\mathbf{K}\leftarrow W_{k}\mathbf{E}$ //Key projection + $\mathbf{V}\leftarrow W_{v}\mathbf{E}$ //Value projection +Function FeatureMixing $(\mathbf{F}_0,\mathbf{K},\mathbf{V},\mathbf{U})$ .. +for $n = 0$ to $N - 1$ do +Q←WqFn //Query projection +S←Q·KT// Spectral correlation +A←softmax(S+U) //Attention weights +O←A·V //Attention output +Fn+1←MLP(GRU(Fn,O)) //Feature update with GRU and MLP +end +return FN + $\mathbf{F}_N\gets$ FeatureMixing(F0,K,V,U) + +# 3.4. Illumination Spectrum Estimation in Decoder Blocks + +In decoder blocks, the proposed network generates an estimation of the illumination spectrum, which is denoted by $\hat{\mathbf{L}}$ , at the end. The network uses two branches to calculate not only the illumination spectrum estimation but also the surface reflectance approximation, which is denoted by $\hat{\mathbf{R}}$ . The approximation of the surface reflectance in the decoder enables the network to use a joint IRSS framework, rather than producing an exact surface reflectance in the absence of GT. As shown in Fig. 2, $\mathbf{F}_N$ is fed into a reflectance decoder, consisting of three transposed convolution layers, to + +calculate $\hat{\mathbf{R}}$ as intermediate results in the IRSS. Specifically, we use $3\times 3$ convolution kernels in the implementation. The reflectance decoder uses the only $\mathbf{F}_N$ as input to facilitate the optimization and fully focus on the reflectance prediction. + +The illumination spectrum block is built with a series of convolutional and linear layers, in which an input is constructed by concatenating $\mathbf{F}_{cnn}$ and $\mathbf{F}_N$ . The subsequent convolutional layers are employed to extract spatially localized spectral features, effectively capturing illumination in the target scene. We employ two layers of convolution kernel with stride 2. Rectified linear unit (ReLU) activation function is applied after all convolution and linear layers in the block, except for the final output linear layer. + +# 3.5. Loss function + +For training the proposed architecture, we employ a reconstruction loss $\mathcal{L}_{\mathrm{RC}}$ and an illumination estimation loss $\mathcal{L}_{\mathrm{AE}}$ . $\mathcal{L}_{\mathrm{RC}}$ is computed using an $\mathcal{L}_1$ distance between $\mathbf{I}$ and $\hat{\mathbf{I}}$ : + +$$ +\mathcal {L} _ {\mathrm {R C}} = \sum_ {x \in (h, w)} \sum_ {\lambda \in \Lambda} \| \hat {\mathbf {I}} (x, \lambda) - \mathbf {I} (x, \lambda) \| _ {1}, \tag {12} +$$ + +where the loss is computed over all spatial spectral locations, i.e., $h = 1 \cdots H$ and $w = 1 \cdots W$ , and all channels $\Lambda$ . $\hat{\mathbf{I}}$ is reconstructed using an integral of $\hat{\mathbf{L}}$ and $\hat{\mathbf{R}}$ , + +We also utilize an angular error (AE), $\mathcal{L}_{\mathrm{AE}}$ , defined as: + +$$ +\mathcal {L} _ {\mathrm {A E}} = \arccos \left(\frac {\hat {\mathbf {L}} \cdot \mathbf {L}}{\| \hat {\mathbf {L}} \| \| \mathbf {L} \|}\right), \tag {13} +$$ + +where $\mathbf{L}$ and $\hat{\mathbf{L}}$ are the ground-truth (GT) and predicted illumination spectrum vector, respectively. For illumination spectrum estimation, capturing relative differences of illuminant among multi-channels is crucial. Spectral orientation defines its color characteristics, which would affect color constancy, while the intensity of illuminant reflects brightness. Thus, we use an angular error to focus on the orientation by measuring the angle between vectors. + +The total loss function $\mathcal{L}$ is defined as: + +$$ +\mathcal {L} = \alpha_ {1} \mathcal {L} _ {\mathrm {R C}} + \alpha_ {2} \mathcal {L} _ {\mathrm {A E}}, \tag {14} +$$ + +where $\alpha_{1}$ and $\alpha_{2}$ are weighting factors, which are set to 0.5 and 1, respectively, in our experiments. + +# 4. Experimental Results + +# 4.1. Spectral Reflectance Datasets + +We evaluate the proposed technique using two spectral reflectance datasets: KAUST [40] and CAVE [66]. In the datasets, we use 31 channels ranging from $400nm$ to $700nm$ at $10nm$ intervals. We relight the scenes by multiplying the surface reflectance of each pixel with an illuminant spectra matrix with 59 standard illuminants from + +[49] resulting in a 31-channel spectral image. We randomly select $80\%$ of the KAUST reflectance and illuminant data for training and the remaining $20\%$ used for testing as in [40]. CAVE dataset is utilized only for testing the models. We evaluate the performance of the proposed technique with existing methods including both statistic-based and learning-based methods, GrayEdge [27], LRMF [72], ISNL [58], PWIR [54], DUN [40] and BeyondRGB [14]. We used mean- $\Delta \mathrm{A}_{\mathrm{MS}}$ and std- $\Delta \mathrm{A}_{\mathrm{MS}}$ for evaluation metrics, which are the average and standard deviation of an angular error measured between the GT and predicted illumination spectrum. + +Tab. 1 exhibits the performance comparisons, showing that the proposed technique achieves the best results on both datasets. The proposed technique produces 5.55 and 3.29 in terms of mean- $\Delta \mathrm{A}_{\mathrm{MS}}$ and std- $\Delta \mathrm{A}_{\mathrm{MS}}$ , respectively, in KAUST dataset. For CAVE dataset, which is not used for model training, it shows reliable performance achieving 5.93 and 3.90 in mean- $\Delta \mathrm{A}_{\mathrm{MS}}$ and std- $\Delta \mathrm{A}_{\mathrm{MS}}$ , respectively. These results demonstrate that the proposed technique holds stable performance on unseen reflectance data. + +
DatasetMethodmean-ΔAMS↓std-ΔAMS↓
KAUSTGrayEdge [27]11.468.02
LRMF [72]24.0623.49
ISNL [58]20.0511.46
PWIR [54]28.0721.20
DUN [40]9.176.30
BeyondRGB [14]7.304.45
Ours5.553.29
CAVEGrayEdge [27]18.3314.89
LRMF [72]20.0515.47
ISNL [58]22.9216.62
PWIR [54]34.9521.20
DUN [40]16.0413.75
BeyondRGB [14]6.643.62
Ours5.933.90
+ +Table 1. Performance comparisons of the proposed technique and existing methods using KAUST and CAVE datasets. + +# 4.2. Beyond RGB Dataset + +The BeyondRGB [14] dataset is a comprehensive, real-world dataset, designed for MS image-related studies, focused on mobile devices. The dataset is organized into "Lab" and "Field," which include a variety of objects with diverse colors and textures in indoor scenes and real-world data in outdoor scenes, both encompassing complex scenes with rich object content. For experimental results, we use the same data-split protocol of [14]. For comparisons, we used PWIR and BeyondRGB, because the other methods required reflectance data for training [40] or the input dimension needs to match the number of spectral bands in GT spectra [27, 58, 72] + +![](images/11ad1a0741846c3c4dddda4357a2a6b7032b08cab2a80e65759fc0899361d933.jpg) + +![](images/289bf1a424a0c8c9de12839718b97ac3c8e1e846ee1b68c4cf628ea545c429e8.jpg) + +![](images/f661da813ea0cf9f722597cc1f1fb0c96868fb167a3d7f526caecda7c525e8de.jpg) + +![](images/a835ebdab787fe677dee1f9ef4bb28f36287917235550f78ca03a03fc79b93ec.jpg) + +![](images/3e3075dc6170a67dc305057d1765627955c5590c0338b4e1e3371737f1561f28.jpg) + +![](images/00f1eeb682dee850d3576b3c10190465f74e5ada1808e4267b3a1a336a12bc16.jpg) + +![](images/4738b24ec9c10711e8f5d0ff4e77cdd3f3ba0601e15974f358505f0e678f78bd.jpg) + +![](images/19b4f5c6332d03a7c0d84b2ad6d3b73a73d224b9eae1fd2e58bf53769e0e54ca.jpg) + +![](images/00d323253bdfa5b31541e95b7aa01b063e72b9b47bf6b50c56e84d3f8bc4e8ed.jpg) + +![](images/02bc1a3ecc451ab18c90ddc1a3354e43a83e5521e8db8f0bcad47fba7d3c3924.jpg) + +![](images/6984c8bc99acbf5a3f996854968a22675fe48dce008ffc68fc67e39f939c1df9.jpg) + +![](images/a2de56cb48871d04b489e78d3f9a08c4d746c0f9e00178f8f7417b0f50248fe5.jpg) +Figure 5. Each of two rows shows MS images of BeyondRGB dataset in the first row and their estimated illuminant spectrum in the second row. The results of our model and PWIR are plotted with red and green dashed lines, respectively. The GT is shown with black solid lines. + +![](images/f0ebc663ad294d2015e2f9c4e6f8b801bc9b1522436280fc4b8700287e64a0c8.jpg) + +![](images/9f5e41a04383d260e3e73a0160c4d88d896bac651b42a2f9d87fd3e2949d9454.jpg) + +![](images/0e9e1f18ec475138c42ee36c28aa4a619997552c0aa325074701da5fa3597636.jpg) + +![](images/3e7c87d9958a4254da26edbeef788fc5d58139ca853997ba4dbc32d89f9b17ca.jpg) + +![](images/016805b07df054c1585d499879e09c69147bd087a78164f2e4ed7ef364fb7599.jpg) + +![](images/8001f9d055b0095d0df4209f3faa3ccae167aa815ad35be722bc14bfe7d8f2c5.jpg) + +![](images/fb6766194498200ae07dd9a0553cb47d241c0998567ba9e77f545b79265244b0.jpg) + +![](images/0aefebea1e66a61519e73628cdacf1172cfdd0e6c4c28b0ca8e9f664dc052743.jpg) + +Tab. 2 shows the accuracy of the tested methods. It is clearly demonstrated with the results that the proposed technique provides superior accuracy compared to all the tested methods. Our technique significantly reduces the angular errors, achieving values of 3.65 and 5.97 in the Lab and Field datasets, respectively, in mean- $\Delta \mathrm{A}_{\mathrm{MS}}$ . In the BeyondRGB dataset, the proposed technique presents a greater performance gap compared to the second-best algorithm [14] than in KAUST and CAVE datasets, which can present the effectiveness of our approach in complex scenes. + +
DatasetMethodmean-ΔAMS↓std-ΔAMS↓
LabPWIR [54]27.1010.86
BeyondRGB [14]5.922.92
Ours3.652.72
FieldPWIR [54]16.316.07
BeyondRGB [14]7.225.54
Ours5.973.90
+ +Table 2. Performance comparisons of the proposed technique and existing methods using real-world MS dataset, BeyondRGB. + +Fig. 5 shows the results of illumination spectrum using the proposed technique in comparison with PWIR [54]. The top row of the image consists of Lab images, captured with controlled and artificial illuminations. From left to right, the illuminations are the spectrum of fluorescent, mixed LED, halogen lamp, incandescent light bulb, and LED illumi + +nant, which are found in daily environments. The bottom row contains real-world Field images, where outdoor images are illuminated by natural sunlight. The estimated illumination spectrum of our model, represented by the red dashed line, demonstrate a significantly more detailed alignment with the shape and peaks of the GT compared to PWIR. In the first image of the Lab set, the proposed technique shows a superior performance in estimating complex illumination conditions, such as those with multiple peaks characteristic of fluorescent lighting. Even for the bottom row comprising a wider variety of scenes, our model demonstrates robust and accurate illumination estimation performance. + +# 4.3. Surface Reflectance Modeling + +We investigate the reconstruction results of surface reflectance on the KAUST and CAVE datasets, in which GT reflectance data is available. As shown in Fig. 6, we compare the estimated reflectance from our reflectance decoder and corresponding GT reflectance. From the results, we observe that even though the network has not directly seen GT reflectance during the training process, the surface reflectance is sufficiently considered while estimating illumination from images. Even in complex optical conditions like refractive surfaces with shadows shown in the window and glasses scenes, and reflection and diffraction effects demonstrated by the CD, our method effectively reconstruct the surface reflectance preserving these challenges. + +![](images/2275750dc7a9756e73f8c594d34c2cc84a2b14abfac46e21d69663875e741b08.jpg) +Figure 6. Visualization of reconstructed surface reflectance for KAUST and CAVE datasets. The first and second columns display GT and reconstructed surface reflectance rendered in RGB. The following columns display the estimated surface reflectance at each wavelength, respectively. + +ing visual properties. The results align with our motivation for effective disentanglement of illumination and reflectance through physics-based approach and validate the effectiveness of our proposed method. + +# 4.4. Ablation Study + +Network architecture: Tab. 3 presents the effectiveness of the proposed modules consisting of the network architecture. When the modules are turned off one-by-one, we observe a considerable reduction in accuracy. First, without SU block, the network relies solely on visual features without using spectral features, resulting in an approximate increase of 1.12 in angular error. Second, when disabling FMB, the spatial and spectral features are not fused appropriately but used independently to estimate the illumination spectrum. It increases approximately 3.65 of an angular error. Lastly, when the SFE module is turned off, we observe a dramatic degradation in accuracy. This result is expected, as the network is unable to exploit any visual features from the input image without the SFE. Fig. 7 shows the output features of the FMB, when disabling SFE or SU. Compared to the rendered RGB image on the left, the synthesized feature can be effectively used for illumination estimation by localizing regions relevant to the task. + +
SUFMBSFEmean-ΔAMS ↓
X✓✓5.52
✓X✓8.05
✓✓X24.64
✓✓✓4.40
+ +
LAELCRcmean-ΔAMS↓
✓X4.72
X✓73.59
✓✓4.40
+ +Table 3. Ablation studies of the sub-modules of the proposed network (left) and training loss functions (right). + +Loss function: We conducted ablation tests in the loss terms. In the proposed method, using both $\mathcal{L}_{\mathrm{AE}}$ and $\mathcal{L}_{\mathrm{RC}}$ + +![](images/ca96fd750fbeb4f7bb731ecdd738b71dfb77ebc4c013cb56d5766cabcde34e3b.jpg) +Figure 7. Output features from an FMB block in the ablation test. The second and the third columns display the features, when disabling SFE and SU, respectively. Our features in the last column highlight more relevant regions including high spectral diversity in local surfaces to achieve more enhanced accuracy. + +together achieves the best performance. When $\mathcal{L}_{\mathrm{RC}}$ is not used, the performance decreases by approximately 0.32 compared to using both the losses. By incorporating $\mathcal{L}_{\mathrm{RC}}$ , the network can further learn the impact of a global illumination spectrum on the reconstructed image in terms of spatial representation. When excluding $\mathcal{L}_{\mathrm{AE}}$ , we observed substantial performance drops (e.g. 73 in mean- $\Delta \mathrm{A}_{\mathrm{MS}}$ ). These results are expected, since relying solely on the reflectance for the image reconstruction may provide insufficient constraints for illumination estimation in the absence of GT. $\mathcal{L}_{\mathrm{AE}}$ plays a crucial role in the estimation. + +# 5. Conclusion + +In this paper, we proposed illumination spectrum estimation technique for MS images. The proposed ISS technique used a learnable SU block which modeled surface reflectance, and a feature mixing block which fused spatial and spectral information using cross-attention operations. The spatial-spectral features were efficiently used to focus on important regions in complex scenes to enhance the accuracy of illumination estimation. Experimental results demonstrated that the proposed ISS technique provided a superiority to existing studies in various MS image datasets. + +In our future study, we will investigate illumination estimation under locally varying and multi-illuminant scenarios. While the current model is designed to estimate a global illumination spectrum, the proposed ISS technique can be extended to improve accuracy in estimating multiple illuminant by efficiently disentangling reflectance properties under locally varying lighting conditions. + +# Acknowledgment + +This work was partly supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (RS-2021-II212068, Artificial Intelligence Innovation Hub, $40\%$ ) and was partly supported by the NRF grant funded by MSIT (No.NRF-2022R1A2C4002052, $60\%$ ). + +# References + +[1] Mahmoud Affifi and Michael S Brown. Sensor-independent illumination estimation for dnn models. arXiv preprint arXiv:1912.06888, 2019. 2 +[2] Dongsheng An, Jinli Suo, Haoqian Wang, and Qionghai Dai. Illumination estimation from specular highlight in a multispectral image. Optics Express, 23(13):17008-17023, 2015. 2, 4 +[3] Simone Bianco, Claudio Cusano, and Raimondo Schettini. Color constancy using cnns. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 81-89, 2015. 1, 2 +[4] Simone Bianco, Claudio Cusano, and Raimondo Schettini. Single and multiple illuminant estimation using convolutional neural networks. IEEE Transactions on Image Processing, 26(9):4347-4362, 2017. 2 +[5] José M Bioucas-Dias, Antonio Plaza, Nicolas Dobigeon, Mario Parente, Qian Du, Paul Gader, and Jocelyn Chanussot. Hyperspectral unmixing overview: Geometrical, statistical, and sparse regression-based approaches. IEEE journal of selected topics in applied earth observations and remote sensing, 5(2):354-379, 2012. 3 +[6] Ricardo Augusto Borsoi, Tales Imbiriba, and José Carlos Moreira Bermudez. Deep generative endmember modeling: An application to unsupervised spectral unmixing. IEEE Transactions on Computational Imaging, 6:374-384, 2019. 3 +[7] David H Brainard and Brian A Wandell. Analysis of the retina theory of color vision. JOSA A, 3(10):1651-1661, 1986. 1 +[8] Gershon Buchsbaum. A spatial processor model for object colour perception. Journal of the Franklin institute, 310(1): 1-26, 1980. 2 +[9] Ayan Chakrabarti, Keigo Hirakawa, and Todd Zickler. Color constancy with spatio-spectral statistics. IEEE Transactions on Pattern Analysis and Machine Intelligence, 34(8):1509-1519, 2011. 1 +[10] Xiaochuan Chen, Mark S Drew, and Ze-Nian Li. Illumination and reflectance spectra separation of hyperspectral image data under multiple illumination conditions. *Electronic Imaging*, 29:194–199, 2017. 2, 4 +[11] Yushi Chen, Hanlu Jiang, Chunyang Li, Xiuping Jia, and Pedram Ghamisi. Deep feature extraction and classification of hyperspectral images based on convolutional neural networks. IEEE transactions on geoscience and remote sensing, 54(10):6232-6251, 2016. 3 + +[12] Arthur P Dempster, Nan M Laird, and Donald B Rubin. Maximum likelihood from incomplete data via the em algorithm. Journal of the royal statistical society: series B (methodological), 39(1):1-22, 1977. 4 +[13] Arjan Gijsenij, Theo Gevers, and Joost Van De Weijer. Improving color constancy by photometric edge weighting. IEEE Transactions on Pattern Analysis and Machine Intelligence, 34(5):918-929, 2011. 2 +[14] Ortal Glatt, Yotam Ater, Woo-Shik Kim, Shira Werman, Oded Berby, Yael Zini, Shay Zelinger, Sangyoon Lee, Heejin Choi, and Evgeny Soloveichik. Beyond rgb: A real world dataset for multispectral imaging in mobile devices. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4344-4354, 2024. 1, 2, 3, 4, 6, 7 +[15] Lin Gu, Cong Phuoc Huynh, and Antonio Robles-Kelly. Segmentation and estimation of spatially varying illumination. IEEE transactions on image processing, 23(8):3478-3489, 2014. 2 +[16] Rui Guo, Wei Wang, and Hairong Qi. Hyperspectral image unmixing using autoencoder cascade. In 2015 7th Workshop on Hyperspectral Image and Signal Processing: Evolution in Remote Sensing (WHISPERS), pages 1-4. IEEE, 2015. 3 +[17] Amina Ben Hamida, Alexandre Benoit, Patrick Lambert, and Chokri Ben Amar. 3-d deep learning approach for remote sensing image classification. IEEE Transactions on geoscience and remote sensing, 56(8):4420-4434, 2018. 3 +[18] Siyuan Hao, Wei Wang, Yuanxin Ye, Tingyuan Nie, and Lorenzo Bruzzone. Two-stream deep architecture for hyperspectral image classification. IEEE Transactions on Geoscience and Remote Sensing, 56(4):2349-2361, 2017. 3 +[19] Rob Heylen, Mario Parente, and Paul Gader. A review of nonlinear hyperspectral unmixing methods. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 7(6):1844-1868, 2014. 3 +[20] Yuanming Hu, Baoyuan Wang, and Stephen Lin. Fc4: Fully convolutional color constancy with confidence-weighted pooling. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4085-4094, 2017. 2 +[21] Anya Hurlbert. Challenges to color constancy in a contemporary light. Current Opinion in Behavioral Sciences, 30: 186-193, 2019. 1 +[22] Marian-Daniel Iordache, José M Bioucas-Dias, and Antonio Plaza. Sparse unmixing of hyperspectral data. IEEE Transactions on Geoscience and Remote Sensing, 49(6):2014-2039, 2011. 2 +[23] Qiwen Jin, Yong Ma, Erting Pan, Fan Fan, Jun Huang, Hao Li, Chenhong Sui, and Xiaoguang Mei. Hyperspectral unmixing with gaussian mixture model and spatial group sparsity. Remote Sensing, 11(20):2434, 2019. 3 +[24] Murali Kanthi, T Hitendra Sarma, and C Shobha Bindu. A 3d-deep cnn based feature extraction and hyperspectral image classification. In 2020 IEEE India Geoscience and Remote Sensing Symposium (InGARSS), pages 229-232. IEEE, 2020. 3 +[25] Nirmal Keshava and John F Mustard. Spectral unmixing. IEEE signal processing magazine, 19(1):44-57, 2002. 2, 3 + +[26] Haris Ahmad Khan. Multispectral constancy for illuminant invariant representation of multispectral images. PhD thesis, Université Bourgogne Franche-Comté; Norwegian University of Science and ..., 2018. 1 +[27] Haris Ahmad Khan, Jean-Baptiste Thomas, Jon Yngve Hardeberg, and Olivier Lalignant. Illuminant estimation in multispectral imaging. JOSA A, 34(7):1085-1098, 2017. 2, 6 +[28] Haris Ahmad Khan, Jean-Baptiste Thomas, and Jon Yngve Hardeberg. Towards highlight based illuminant estimation in multispectral images. In Image and Signal Processing: 8th International Conference, ICISP 2018, Cherbourg, France, July 2-4, 2018, Proceedings 8, pages 517-525. Springer, 2018. 1 +[29] Muhammad Jaleed Khan, Hamid Saeed Khan, Adeel Yousaf, Khurram Khurshid, and Asad Abbas. Modern trends in hyperspectral image analysis: A review. *IEEE Access*, 6:14118-14129, 2018. 1 +[30] Dongyoung Kim, Jinwoo Kim, Junsang Yu, and Seon Joo Kim. Attentive illumination decomposition model for multi-illuminant white balancing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25512–25521, 2024. 1 +[31] Furkan Klinl, Doga Yilmaz, Baris Özcan, and Furkan Kiraç. Deterministic neural illumination mapping for efficient autowhite balance correction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1139-1147, 2023. 1 +[32] Fanqiang Kong, Tongbo Cao, Yunsong Li, Dan Li, and Kedi Hu. Multi-scale spatial-spectral attention network for multispectral image compression based on variational autoencoder. Signal Processing, 198:108589, 2022. 3 +[33] Edwin H Land. The retina theory of color vision. Scientific american, 237(6):108-129, 1977. 2 +[34] Edwin H Land. Recent advances in retina theory. In Central and Peripheral Mechanisms of Colour Vision: Proceedings of an International Symposium Held at The Wenner-Gren Center Stockholm, June 14–15, 1984, pages 5–17. Springer, 1985. 1 +[35] David Landgrebe. Hyperspectral image data analysis. IEEE Signal processing magazine, 19(1):17-28, 2002. 1 +[36] Bing Li, Haina Qin, Weihua Xiong, Yangxi Li, Songhe Feng, Weiming Hu, and Stephen Maybank. Ranking-based color constancy with limited training samples. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(10): 12304-12320, 2023. 2 +[37] Rui Li, Shunyi Zheng, Chenxi Duan, Yang Yang, and Xiqi Wang. Classification of hyperspectral image based on double-branch dual-attention mechanism network. Remote Sensing, 12(3):582, 2020. 3 +[38] Xian Li, Mingli Ding, and Aleksandra Pizurica. Deep feature fusion via two-stream convolutional neural network for hyperspectral image classification. IEEE Transactions on Geoscience and Remote Sensing, 58(4):2615-2629, 2019. 3 +[39] Yuqi Li, Chong Wang, Jieyu Zhao, and Qingshu Yuan. Efficient spectral reconstruction using a trichromatic camera via sample optimization. The Visual Computer, 34:1773-1783, 2018. 2 + +[40] Yuqi Li, Qiang Fu, and Wolfgang Heidrich. Multispectral illumination estimation using deep unrolling network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2672-2681, 2021. 1, 2, 4, 6 +[41] Shuaiqi Liu, Siyuan Liu, Shichong Zhang, Bing Li, Weiming Hu, and Yu-Dong Zhang. Ssau-net: A spectral-spatial attention-based u-net for hyperspectral image fusion. IEEE Transactions on Geoscience and Remote Sensing, 60:1-16, 2022. 3 +[42] Zhenqi Liu, Xinyu Wang, Yanfei Zhong, Meng Shu, and Chen Sun. Siamhyper: Learning a hyperspectral object tracker from an rgb-based tracker. IEEE Transactions on Image Processing, 31:7116-7129, 2022. 1 +[43] Yi-Chen Lo, Chia-Che Chang, Hsuan-Chao Chiu, Yu-Hao Huang, Chia-Ping Chen, Yu-Lin Chang, and Kevin Jou. Clcc: Contrastive learning for color constancy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8053-8063, 2021. 2 +[44] Zhongyu Lou, Theo Gevers, Ninghang Hu, Marcel P Lucassen, et al. Color constancy by deep learning. In BMVC, pages 76-1, 2015. 1 +[45] Qing Ma, Junjun Jiang, Xianming Liu, and Jiayi Ma. Reciprocal transformer for hyperspectral and multispectral image fusion. Information Fusion, 104:102148, 2024. 3 +[46] Yong Ma, Qiwen Jin, Xiaoguang Mei, Xiaobing Dai, Fan Fan, Hao Li, and Jun Huang. Hyperspectral unmixing with gaussian mixture model and low-rank representation. Remote Sensing, 11(8):911, 2019. 2, 3 +[47] Laurence Maloney. Evaluation of linear models of surface spectral reflectance with small numbers of parameters. In Physics-based vision: Principles and practice. Jones and Bartlett, 1992. 4 +[48] Laurence T Maloney and Brian A Wandell. Color constancy: a method for recovering surface spectral reflectance. In Readings in Computer Vision, pages 293-297. Elsevier, 1987. 1, 3 +[49] Thomas Mansencal, Michael Mauderer, Michael Parsons, Nick Shaw, Kevin Wheatley, Sean Cooper, Jean D. Vandenberg, Luke Canavan, Katherine Crowson, Ofek Lev, Katrin Leinweber, Shriramana Sharma, Troy James Sobotka, Dominik Moritz, Matt Pppp, Chinmay Rane, Pavithra Eswaramoorthy, John Metic, Ben Pearlstine, Manuel Leonhardt, Olli Niemitalo, Marek Szymanski, Maximilian Schambach, Sianyi Huang, Mike Wei, Nishant Joywardhan, Omar Wagih, Pawel Redman, Joseph Goldstone, Stephen Hill, Jedediah Smith, Frederic Savoir, Geetansh Saxena, Saransh Chopra, Ilia Sibiryakov, Tim Gates, Gajendra Pal, Nicolas Tessore, Aurélien Pierre, François-Xavier Thomas, Sabarish Srinivasan, and Tucker Downs. Colour 0.4.2, 2022. 6 +[50] Savas Ozkan, Berk Kaya, and Gozde Bozdagi Akar. Endnet: Sparse autoencoder network for endmember extraction and hyperspectral unmixing. IEEE Transactions on Geoscience and Remote Sensing, 57(1):482-496, 2018. 3 +[51] Yishu Peng, Yuwen Zhang, Bing Tu, Qianming Li, and Wujing Li. Spatial-spectral transformer with cross-attention for hyperspectral image classification. IEEE Transactions on Geoscience and Remote Sensing, 60:1–15, 2022. 3 + +[52] Ying Qu and Hairong Qi. udas: An untied denoising autoencoder with sparsity for spectral unmixing. IEEE Transactions on Geoscience and Remote Sensing, 57(3):1698-1712, 2018. 3 +[53] Douglas A Reynolds et al. Gaussian mixture models. Encyclopedia of biometrics, 741(659-663), 2009. 2, 3 +[54] Antonio Robles-Kelly and Ran Wei. A convolutional neural network for pixelwise illuminant recovery in colour and spectral images. In 2018 24th International Conference on Pattern Recognition (ICPR), pages 109-114. IEEE, 2018. 1, 6, 7 +[55] Swalpa Kumar Roy, Gopal Krishna, Shiv Ram Dubey, and Bidyut B Chaudhuri. Hybridsn: Exploring 3-d-2-d cnn feature hierarchy for hyperspectral image classification. IEEE Geoscience and Remote Sensing Letters, 17(2):277-281, 2019. 3 +[56] Wu Shi, Chen Change Loy, and Xiaou Tang. Deep specialized network for illuminant estimation. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 371-387. Springer, 2016. 2 +[57] Alberto Signoroni, Mattia Savardi, Annalisa Baronio, and Sergio Benini. Deep learning meets hyperspectral image analysis: A multidisciplinary review. Journal of imaging, 5(5):52, 2019. 3 +[58] Tong Su, Yu Zhou, Yao Yu, Xun Cao, and Sidan Du. Illumination separation of non-lambertian scenes from a single hyperspectral image. Optics express, 26(20):26167-26178, 2018. 6 +[59] Yuanchao Su, Jun Li, Antonio Plaza, Andrea Marinoni, Paolo Gamba, and Somdatta Chakravortty. Daen: Deep autoencoder networks for hyperspectral unmixing. IEEE Transactions on Geoscience and Remote Sensing, 57(7): 4309-4321, 2019. 3 +[60] Hao Sun, Xiangtao Zheng, and Xiaogiang Lu. A supervised segmentation network for hyperspectral image classification. IEEE Transactions on Image Processing, 30:2810-2825, 2021. 1 +[61] Jean-Baptiste Thomas. Illuminant estimation from uncalibrated multispectral images. In 2015 Colour and Visual Computing Symposium (CVCS), pages 1-6. IEEE, 2015. 2 +[62] Joost Van De Weijer, Theo Gevers, and Arjan Gijsenij. Edge-based color constancy. IEEE Transactions on image processing, 16(9):2207-2214, 2007. 2 +[63] Donik Vršnak, Ilija Domislovic, Marko Subasic, and Sven Lončarić. Framework for illumination estimation and segmentation in multi-illuminant scenes. IEEE access, 11: 2128-2137, 2023. 2 +[64] Xia Xu and Zhenwei Shi. Multi-objective based spectral unmixing for hyperspectral images. ISPRS Journal of Photogrammetry and Remote Sensing, 124:54-69, 2017. 2 +[65] Kai Yang, Hao Sun, Chunbo Zou, and Xiaoqiang Lu. Cross-attention spectral-spatial network for hyperspectral image classification. IEEE Transactions on Geoscience and Remote Sensing, 60:1-14, 2021. 3 +[66] Fumihito Yasuma, Tomoo Mitsunaga, Daisuke Iso, and Shree K Nayar. Generalized assorted pixel camera: post-capture control of resolution, dynamic range, and spectrum. + +IEEE transactions on image processing, 19(9):2241-2253, 2010.6 +[67] Huanglin Yu, Ke Chen, Kaiqi Wang, Yanlin Qian, Zhaoxiang Zhang, and Kui Jia. Cascading convolutional color constancy. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 12725-12732, 2020. 2 +[68] Jinkai Zhang, Benoit Rivard, and Arturo Sanchez-Azofeifa. Derivative spectral unmixing of hyperspectral data applied to mixtures of lichen and rock. IEEE Transactions on Geoscience and Remote Sensing, 42(9):1934-1940, 2004. 2 +[69] Xiangrong Zhang, Yujia Sun, Jingyan Zhang, Peng Wu, and Licheng Jiao. Hyperspectral unmixing via deep convolutional neural networks. IEEE Geoscience and Remote Sensing Letters, 15(11):1755-1759, 2018. 3 +[70] Chunhui Zhao, Boao Qin, Shou Feng, Wenxiang Zhu, Weiwei Sun, Wei Li, and Xiuping Jia. Hyperspectral image classification with multi-attention transformer and adaptive superpixel segmentation-based active learning. IEEE Transactions on Image Processing, 32:3606-3621, 2023. 1 +[71] Runxing Zhao, Zhiwen Wang, Wuyuan Guo, and Canlong Zhang. Multi-scene image enhancement based on multi-channel illumination estimation. Expert Systems with Applications, 226:120271, 2023. 1 +[72] Yinqiang Zheng, Imari Sato, and Yoichi Sato. Illumination and reflectance spectra separation of a hyperspectral image meets low-rank matrix factorization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1779-1787, 2015. 1, 2, 4, 6 +[73] Zilong Zhong, Jonathan Li, Zhiming Luo, and Michael Chapman. Spectral-spatial residual network for hyperspectral image classification: A 3-d deep learning framework. IEEE Transactions on Geoscience and Remote Sensing, 56 (2):847-858, 2017. 3 +[74] Yuan Zhou, Anand Rangarajan, and Paul D Gader. A gaussian mixture model representation of endmember variability in hyperspectral unmixing. IEEE Transactions on Image Processing, 27(5):2242-2256, 2018. 2, 3 \ No newline at end of file diff --git a/2025/Illumination Spectrum Estimation for Multispectral Images via Surface Reflectance Modeling and Spatial-Spectral Feature Generation/images.zip b/2025/Illumination Spectrum Estimation for Multispectral Images via Surface Reflectance Modeling and Spatial-Spectral Feature Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..dc4a3106fd931687218061429467c8c6e9a3d7e5 --- /dev/null +++ b/2025/Illumination Spectrum Estimation for Multispectral Images via Surface Reflectance Modeling and Spatial-Spectral Feature Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:da0237c84409d88ad0d83fa0c9d39cd26c0a08aa7373ca27072fb0eb0f1c0cff +size 636303 diff --git a/2025/Illumination Spectrum Estimation for Multispectral Images via Surface Reflectance Modeling and Spatial-Spectral Feature Generation/layout.json b/2025/Illumination Spectrum Estimation for Multispectral Images via Surface Reflectance Modeling and Spatial-Spectral Feature Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6012f696ea99efe3cd7ed442ba626204d16de362 --- /dev/null +++ b/2025/Illumination Spectrum Estimation for Multispectral Images via Surface Reflectance Modeling and Spatial-Spectral Feature Generation/layout.json @@ -0,0 +1,12004 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 85, + 103, + 526, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 103, + 526, + 140 + ], + "spans": [ + { + "bbox": [ + 85, + 103, + 526, + 140 + ], + "type": "text", + "content": "Illumination Spectrum Estimation for Multispectral Images via Surface Reflectance Modeling and Spatial-Spectral Feature Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 88, + 160, + 522, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 160, + 522, + 217 + ], + "spans": [ + { + "bbox": [ + 88, + 160, + 522, + 217 + ], + "type": "text", + "content": "Hyejin Oh" + }, + { + "bbox": [ + 88, + 160, + 522, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 88, + 160, + 522, + 217 + ], + "type": "text", + "content": " Woo-Shik Kim" + }, + { + "bbox": [ + 88, + 160, + 522, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 88, + 160, + 522, + 217 + ], + "type": "text", + "content": " Sangyoon Lee" + }, + { + "bbox": [ + 88, + 160, + 522, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 88, + 160, + 522, + 217 + ], + "type": "text", + "content": " YungKyung Park" + }, + { + "bbox": [ + 88, + 160, + 522, + 217 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 88, + 160, + 522, + 217 + ], + "type": "text", + "content": " Je-Won Kang" + }, + { + "bbox": [ + 88, + 160, + 522, + 217 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 88, + 160, + 522, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 88, + 160, + 522, + 217 + ], + "type": "text", + "content": "Dept. of Electronic and Electrical Engineering, Ewha W. University, Seoul, South Korea \n" + }, + { + "bbox": [ + 88, + 160, + 522, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 88, + 160, + 522, + 217 + ], + "type": "text", + "content": "Samsung Advanced Institute of Technology, South Korea \n" + }, + { + "bbox": [ + 88, + 160, + 522, + 217 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 88, + 160, + 522, + 217 + ], + "type": "text", + "content": "Dept.of Color Design, Ewha W. University, Seoul, South Korea" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 71, + 219, + 536, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 219, + 536, + 232 + ], + "spans": [ + { + "bbox": [ + 71, + 219, + 536, + 232 + ], + "type": "text", + "content": "{hyejin5, yungkyung PARK, jewonk}@ewha.ac.kr {wooshik.kim, sangy00n.lee}@samsung.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "spans": [ + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 283, + 296, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 283, + 296, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 283, + 296, + 548 + ], + "type": "text", + "content": "Multispectral (MS) images contain richer spectral information than RGB images due to their increased number of channels and are widely used for various applications. However, achieving accurate estimation in MS images remains challenging, as previous studies have struggled with spectral diversity and the inherent entanglement between the illuminant and surface reflectance spectra. To tackle these challenges, in this paper, we propose a novel Illumination spectrum estimation technique for MS images via Surface reflectance modeling and Spatial-spectral feature generation (ISS). The proposed technique employs a learnable spectral unmixing (SU) block to enhance surface reflectance modeling, which was unattempted in the illumination spectrum estimation, and a feature mixing block to fuse spectral and spatial features of MS images with cross-attention. The features are refined iteratively and processed through a decoder to produce an illumination spectrum estimator. Experimental results demonstrate that the proposed technique achieves state-of-the-art performance in illumination spectrum estimation in various MS image datasets. The code is available at https://github.com/heyjinnii/ISS-MSI.git." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 568, + 135, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 135, + 582 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 135, + 582 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 588, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 295, + 696 + ], + "type": "text", + "content": "Multispectral (MS) imaging sensors capture a broader range of spectral bands and provide richer information than in RGB images [29, 35]. As these imaging sensors become compact and cost-effective [14], they are actively used in various image processing applications [3, 42, 44, 60, 70]. In computer vision, computational color constancy is used to replicate the perceptual stability of the human visual system, enabling reliable color rendering under varying lighting conditions [7, 34]. In MS imaging, the color constancy" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 258, + 392, + 335 + ], + "blocks": [ + { + "bbox": [ + 316, + 258, + 392, + 335 + ], + "lines": [ + { + "bbox": [ + 316, + 258, + 392, + 335 + ], + "spans": [ + { + "bbox": [ + 316, + 258, + 392, + 335 + ], + "type": "image", + "image_path": "636e9b20a64b06d4bffd846412b0cef04956e6e9f155016d30eaba5c6d550ec1.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 335, + 392, + 413 + ], + "blocks": [ + { + "bbox": [ + 316, + 335, + 392, + 413 + ], + "lines": [ + { + "bbox": [ + 316, + 335, + 392, + 413 + ], + "spans": [ + { + "bbox": [ + 316, + 335, + 392, + 413 + ], + "type": "image", + "image_path": "4032483957fb5fc17224a778d34dcbfce0e6fafb93ac627bbca60205383ccb79.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 422, + 555, + 499 + ], + "lines": [ + { + "bbox": [ + 313, + 422, + 555, + 499 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 555, + 499 + ], + "type": "text", + "content": "Figure 1. Our model synthesizes spectral features derived from surface reflectance modeling with CNN features, iteratively refining the fused representation in an IRSS framework. The features enhance the accuracy of illumination spectrum estimation by effectively accounting for the reflectance characteristics of objects within the image, as demonstrated in the last column, in contrast to the conventional deep features shown in the middle column." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 392, + 258, + 469, + 335 + ], + "blocks": [ + { + "bbox": [ + 392, + 258, + 469, + 335 + ], + "lines": [ + { + "bbox": [ + 392, + 258, + 469, + 335 + ], + "spans": [ + { + "bbox": [ + 392, + 258, + 469, + 335 + ], + "type": "image", + "image_path": "395f66e907a50ffd92103965006be6acee628de20f13a870873079cf6be1ce7c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 392, + 335, + 469, + 413 + ], + "blocks": [ + { + "bbox": [ + 392, + 335, + 469, + 413 + ], + "lines": [ + { + "bbox": [ + 392, + 335, + 469, + 413 + ], + "spans": [ + { + "bbox": [ + 392, + 335, + 469, + 413 + ], + "type": "image", + "image_path": "c38296a263f3f74f1cc44f28ba0bf934f870481f798eac706dcb643336ed849d.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 470, + 258, + 553, + 335 + ], + "blocks": [ + { + "bbox": [ + 470, + 258, + 553, + 335 + ], + "lines": [ + { + "bbox": [ + 470, + 258, + 553, + 335 + ], + "spans": [ + { + "bbox": [ + 470, + 258, + 553, + 335 + ], + "type": "image", + "image_path": "db3cceb5902486ee1e51ed8dcd1b618b7d0d9343d378eba952e7ba023ed83109.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 470, + 335, + 553, + 413 + ], + "blocks": [ + { + "bbox": [ + 470, + 335, + 553, + 413 + ], + "lines": [ + { + "bbox": [ + 470, + 335, + 553, + 413 + ], + "spans": [ + { + "bbox": [ + 470, + 335, + 553, + 413 + ], + "type": "image", + "image_path": "da430e92130ddefa98959edb2e902938ab304e53bbb66958339fcc1e0e80553f.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 521, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 640 + ], + "type": "text", + "content": "becomes even more crucial than in the conventional RGB imaging [14, 26, 28, 40, 54, 72], because MS images with the increased number of channels are highly sensitive to illumination changes. The color constancy is achieved by solving an illumination and reflectance spectra separation (IRSS) problem [48], which is an underconstrained problem, requiring the simultaneous recovery of both the illumination spectrum and surface reflectance [72]. Accordingly, achieving accurate estimation in MS images remains a challenge [9, 21, 72]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "Recent advances in sensor technology have enabled the acquisition of high-resolution, high-quality MS images [14]. The separation is more challenging in the real-world MS images, because both illumination and reflectance contribute more intricately to the observed spectrum and can vary more dynamically over their channels [30, 31, 71]." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 201, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 201, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 201, + 713 + ], + "type": "text", + "content": "* Je-Won Kang is a corresponding author." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2215" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 214 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 214 + ], + "type": "text", + "content": "Previous studies have attempted to address the IRSS problem by imposing several constraints to MS images. In [2, 10, 72], scene illumination was separated from surface reflectance using low-rank matrix factorizations. However, relying on low-dimensional subspaces of surface reflectance would incur instability when the number of channels is insufficient relative to the basis functions [72]. Despite the limited research to date, deep learning techniques have been applied to MS images for illumination spectrum estimation [14, 40]. However, the previous deep models have largely overlooked the joint analysis of the illumination spectrum and surface reflectance." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 219, + 296, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 296, + 482 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 296, + 482 + ], + "type": "text", + "content": "To overcome the limitations of existing methods, we propose a comprehensive deep learning-based IRSS framework, dubbed Illumination spectrum estimation for MS images via Surface reflectance modeling and Spatial-spectral feature generation (ISS). Spectral unmixing (SU) [25] is employed in this framework to decompose the measured spectrum into a set of endmembers, representing the spectral reflectance characteristics of specific surfaces, and abundances, which indicate the proportion of each endmember. SU has been widely used to analyze MS images in various fields, such as material science [68] and remote sensing [22, 64], yet few studies have been conducted on illumination spectrum estimation. In the proposed technique, rather than deriving the exact surface reflectance of objects, which is practically challenging, we focus on integrating refined reflectance surface modeling into a deep learning-based illumination spectrum estimation approach. Gaussian Mixture Model (GMM) was applied to generate robust endmembers and abundances [46, 53, 74]. We extend this approach by incorporating learnable parameters for endmembers and abundances, providing a robust and flexible decomposition method." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "type": "text", + "content": "Existing learning-based methods have attempted to build an IRSS framework in their models. However, they may suffer from a limited amount of reflectance data [39] or rely on image prior such as segmentation masks [63] and clusters [15] that are not closely related to illumination spectrum estimation. Their methods were challenging to address diverse illumination spectrum and complicated scenes. In this paper, we propose a feature mixing block that combines the results of endmembers and abundance maps with deep visual features, fully exploiting both spatial information and spectral characteristics of the surface reflectance in the input MS image. Conventional deep learning-based feature estimation methods, commonly employed for illumination estimation, often focus on specific bright areas within a single object and misclassify irrelevant regions as distinct features. As shown in Fig. 1, even in shaded areas where previous methods would struggle, the proposed ISS technique appropriately captures the relevant regions for illumination spectrum estimation, treating both bright and shaded areas" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "as the parts of the same material. The features generated by the feature mixing block are fed into a decoder to reconstruct a reflectance map and into an illumination estimator to establish the IRSS framework, finally estimating illumination spectrum of an MS image in the output." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 326, + 133, + 512, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 133, + 512, + 144 + ], + "spans": [ + { + "bbox": [ + 326, + 133, + 512, + 144 + ], + "type": "text", + "content": "The contributions of this paper are as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 146, + 553, + 277 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 314, + 146, + 553, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 146, + 553, + 181 + ], + "spans": [ + { + "bbox": [ + 314, + 146, + 553, + 181 + ], + "type": "text", + "content": "- We propose a novel deep learning-based IRSS framework to accurately estimate illumination spectrum, addressing the challenges of spectral diversity in MS images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 182, + 553, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 182, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 553, + 241 + ], + "type": "text", + "content": "- We introduce a learnable SU module and a feature mixing block, which combines refined spectral features from surface reflectance modeling with deep visual features using cross-attention, effectively capturing both spatial and spectral characteristics of MS images." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 242, + 553, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 242, + 553, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 242, + 553, + 277 + ], + "type": "text", + "content": "- We evaluate the performance of the proposed technique on various image datasets, demonstrating its superiority to existing studies." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 289, + 406, + 302 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 289, + 406, + 302 + ], + "spans": [ + { + "bbox": [ + 314, + 289, + 406, + 302 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 310, + 497, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 310, + 497, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 310, + 497, + 323 + ], + "type": "text", + "content": "2.1. Illumination Spectrum Estimation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 329, + 553, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 329, + 553, + 424 + ], + "spans": [ + { + "bbox": [ + 313, + 329, + 553, + 424 + ], + "type": "text", + "content": "Color constancy is achieved through two steps, including illumination estimation and chromatic adaptation. Early color constancy studies have focused on illumination estimation from RGB images. Statistics-based methods [8, 33, 62] and deep learning-based methods [1, 3, 4, 20, 36, 43, 56, 67] were developed to improve the accuracy of illumination estimation. These methods only used RGB images to separate illumination and reflectance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 426, + 554, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 554, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 554, + 533 + ], + "type": "text", + "content": "In recent studies, MS images were used to address color constancy by exploiting rich spectral information, which allows for more accurate results. Previous methods extended well-established statistical models and relevant assumptions from RGB to MS spectral domains [13, 27, 61]. In [2, 10, 72], they separated illumination spectrum from surface reflectance using low-rank matrix factorizations. This approach enables the decomposition of the scene with few target objects." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 534, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 554, + 713 + ], + "type": "text", + "content": "Various learning-based illumination estimation techniques were studied for MS images. In [40], illumination spectrum estimation was modeled as a constrained matrix factorization problem. Spectral reflectance images required per-image calibration using a known reference or measurement device [39]. They used a synthesized training set with simulated lighting conditions. However, the limited number of true reflectance data relative to diverse lighting conditions would lead to performance degradation in complex scenes. In [15], an unsupervised method for segmenting illuminant regions was developed. They assumed the illumination spectrum to be piecewise constant as a probabilistic clustering problem. However, this assumption was relatively simple to address various illumination spectrum. In [63], an image segmentation method was used to lo" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2216" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 77, + 550, + 233 + ], + "blocks": [ + { + "bbox": [ + 61, + 77, + 550, + 233 + ], + "lines": [ + { + "bbox": [ + 61, + 77, + 550, + 233 + ], + "spans": [ + { + "bbox": [ + 61, + 77, + 550, + 233 + ], + "type": "image", + "image_path": "591824fc857145bd88ff8c36e1088314c48025d3841c0e8909b4979a62f60478.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 243, + 555, + 277 + ], + "lines": [ + { + "bbox": [ + 54, + 243, + 555, + 277 + ], + "spans": [ + { + "bbox": [ + 54, + 243, + 555, + 277 + ], + "type": "text", + "content": "Figure 2. The proposed network uses several sub-networks including a spectral feature extractor, spectral unmixing block, and feature mixing block to generate spatial and spectral features, which are iteratively refined. The decoder uses reflectance decoder and illumination estimator, consisting of an IRSS framework, and estimates illumination spectrum as a result." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 297, + 296, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 297, + 296, + 406 + ], + "spans": [ + { + "bbox": [ + 54, + 297, + 296, + 406 + ], + "type": "text", + "content": "calize regions dominantly affected by illumination within the image, which were then used for illumination estimation. However, the gaps between the segmentation results and key regions needed for accurate illumination estimation were problematic. In [14], an input MS image was divided into multiple triplets, processed using a pyramid-based convolutional color constancy block. However, the spectral information of illumination and surface reflectance among different channels was not fully exploited." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 424, + 167, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 424, + 167, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 424, + 167, + 437 + ], + "type": "text", + "content": "2.2. Spectral Unmixing" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 446, + 295, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 446, + 295, + 638 + ], + "spans": [ + { + "bbox": [ + 54, + 446, + 295, + 638 + ], + "type": "text", + "content": "The SU [25] has been used to extract unique spectrum from various materials in MS images. Previous SU methods used linear [5] and non-linear mixing models [19]. These methods are based on the assumption that each pixel can be expressed as a combination of endmember spectra and corresponding abundances. GMM [23, 46, 74] and deep learning-based SU methods [6, 16, 50, 52, 57, 59, 69] have been studied, extensively. GMM-based models [53] address spectral variability by modeling each endmember as a probabilistic distribution rather than a fixed value, tailored to real-world applications. Deep learning approaches, on the other hand, utilize neural networks to automatically learn the underlying structure of MS data. CNN-based [69] and autoencoder-based models [16, 50, 52, 59] have demonstrated effectiveness in enhancing unmixing accuracy at the cost of computational complexity." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 641, + 296, + 714 + ], + "type": "text", + "content": "Our method introduces a more refined, learnable SU model based on the GMM, which can model reflectance surface with high spectral variability. The spectral features is combined with a deep model and is further processed, which can adaptively address the spatial and spectral characteristics of input MS images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 297, + 534, + 311 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 534, + 311 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 534, + 311 + ], + "type": "text", + "content": "2.3. Attention Mechanism in Spectral Imaging" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 316, + 555, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 555, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 555, + 518 + ], + "type": "text", + "content": "Attention mechanisms of transformers have been widely used to capture spectral characteristics in spectral imaging [11, 18, 37, 38, 55, 73]. Early studies employed spectral and spatial features independently. In [11, 55, 73], the spectral and spatial features are addressed with sequential processing. In [18, 37, 38], the features are processed in parallel using separated branches and combined with late fusion. While these methods were computationally efficient, they failed to effectively fuse the features for illumination spectrum estimation. Recently, cross-attention [51, 65] and 3D convolution [17, 24] were adopted to fuse the spectral and spatial features of MS images. More recently, transformer architectures have been more actively employed for this purpose [32, 41, 45]. Cross-attention mechanisms facilitate the modeling of global dependencies between spectral and spatial features, enabling the effective learning of complex feature relation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 529, + 421, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 421, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 421, + 544 + ], + "type": "text", + "content": "3. Proposed Method" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 549, + 479, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 549, + 479, + 561 + ], + "spans": [ + { + "bbox": [ + 313, + 549, + 479, + 561 + ], + "type": "text", + "content": "3.1. MS Image Formulation Model" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 566, + 554, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 566, + 554, + 591 + ], + "spans": [ + { + "bbox": [ + 313, + 566, + 554, + 591 + ], + "type": "text", + "content": "In an IRSS framework [48], an MS image " + }, + { + "bbox": [ + 313, + 566, + 554, + 591 + ], + "type": "inline_equation", + "content": "\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 313, + 566, + 554, + 591 + ], + "type": "text", + "content": " with height " + }, + { + "bbox": [ + 313, + 566, + 554, + 591 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 566, + 554, + 591 + ], + "type": "text", + "content": ", width " + }, + { + "bbox": [ + 313, + 566, + 554, + 591 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 313, + 566, + 554, + 591 + ], + "type": "text", + "content": ", and channel " + }, + { + "bbox": [ + 313, + 566, + 554, + 591 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 566, + 554, + 591 + ], + "type": "text", + "content": " is modeled as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 359, + 600, + 555, + 626 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 600, + 555, + 626 + ], + "spans": [ + { + "bbox": [ + 359, + 600, + 555, + 626 + ], + "type": "interline_equation", + "content": "\\mathbf {I} (x) = \\int_ {\\Lambda} \\mathbf {R} (x, \\lambda) \\cdot \\mathbf {L} (\\lambda) \\cdot \\mathbf {S} (\\lambda) d \\lambda , \\tag {1}", + "image_path": "6971305d7aa5232a383bd235272aef625fc2e1a8c6af05ae48d83df137adcb50.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 629, + 554, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 677 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 629, + 554, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{R}(x,\\lambda)" + }, + { + "bbox": [ + 313, + 629, + 554, + 677 + ], + "type": "text", + "content": " represents surface reflectance of a spatial coordinate " + }, + { + "bbox": [ + 313, + 629, + 554, + 677 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 629, + 554, + 677 + ], + "type": "text", + "content": " and wavelength " + }, + { + "bbox": [ + 313, + 629, + 554, + 677 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 629, + 554, + 677 + ], + "type": "text", + "content": " ranged within " + }, + { + "bbox": [ + 313, + 629, + 554, + 677 + ], + "type": "inline_equation", + "content": "\\Lambda" + }, + { + "bbox": [ + 313, + 629, + 554, + 677 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 629, + 554, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{L}(\\lambda)" + }, + { + "bbox": [ + 313, + 629, + 554, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 629, + 554, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{S}(\\lambda) \\in \\mathbb{R}^C" + }, + { + "bbox": [ + 313, + 629, + 554, + 677 + ], + "type": "text", + "content": " represent illumination spectrum and camera sensitivity, respectively." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "In the joint IRSS problem, given camera sensitivity, it is necessary to accurately model reflectance surface to produce " + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{L}(\\lambda)" + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": ". In our approach, the reflectance is modeled as" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2217" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": "the product of abundances and endmembers through the SU, representing a detailed spectral characteristics, given as," + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 111, + 106, + 295, + 140 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 106, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 111, + 106, + 295, + 140 + ], + "type": "interline_equation", + "content": "\\mathbf {R} (x, \\lambda) = \\sum_ {k = 1} ^ {K} \\mathbf {U} _ {k} (x) \\cdot \\mathbf {E} _ {k} (\\lambda), \\tag {2}", + "image_path": "7dd365922b882ca681dbd8360a6fd3e3bc14dee428481d5335e5fa8784823714.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 146, + 296, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 146, + 296, + 183 + ], + "spans": [ + { + "bbox": [ + 55, + 146, + 296, + 183 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 146, + 296, + 183 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_k(\\lambda)" + }, + { + "bbox": [ + 55, + 146, + 296, + 183 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 146, + 296, + 183 + ], + "type": "inline_equation", + "content": "\\mathbf{U}_k(x)" + }, + { + "bbox": [ + 55, + 146, + 296, + 183 + ], + "type": "text", + "content": " denote the " + }, + { + "bbox": [ + 55, + 146, + 296, + 183 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 146, + 296, + 183 + ], + "type": "text", + "content": "-th endmember and its abundance map, respectively. " + }, + { + "bbox": [ + 55, + 146, + 296, + 183 + ], + "type": "inline_equation", + "content": "\\mathbf{I}(x)" + }, + { + "bbox": [ + 55, + 146, + 296, + 183 + ], + "type": "text", + "content": " is modeled by incorporating Eq. (2) as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 196, + 295, + 229 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 196, + 295, + 229 + ], + "spans": [ + { + "bbox": [ + 69, + 196, + 295, + 229 + ], + "type": "interline_equation", + "content": "\\mathbf {I} (x) = \\int_ {\\Lambda} \\left(\\sum_ {k = 1} ^ {K} \\mathbf {U} _ {k} (x) \\cdot \\mathbf {E} _ {k} (\\lambda)\\right) \\cdot \\mathbf {L} (\\lambda) \\cdot \\mathbf {S} (\\lambda) d \\lambda , \\tag {3}", + "image_path": "b9de59e1095246c20f53a63d52b0a2f606395ca7ff8be33935e85a50a6f98fa9.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 239, + 296, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 296, + 287 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 296, + 287 + ], + "type": "text", + "content": "where the SU-based reflectance surface model is used to provide more robust and flexible approach to accurately decompose the components. Fig. 2 presents this process with further details in the following subsections." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 295, + 296, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 295, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 295, + 296, + 319 + ], + "type": "text", + "content": "3.2. Surface Reflectance Representation with Learnable SU" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 326, + 296, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 326, + 296, + 398 + ], + "spans": [ + { + "bbox": [ + 55, + 326, + 296, + 398 + ], + "type": "text", + "content": "In the proposed technique, the SU is used to provide detailed spectrum profile of reflectance surface, and the GMM is employed to decompose observed spectrum into a set of endmember spectra and their corresponding abundances. The components are transformed to learnable tensors to accomplish a deep IRSS framework in the following." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 399, + 296, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 399, + 296, + 423 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 296, + 423 + ], + "type": "text", + "content": "Based on the GMM, a probability density of an observed spectrum " + }, + { + "bbox": [ + 55, + 399, + 296, + 423 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_x\\in \\mathbb{R}^C" + }, + { + "bbox": [ + 55, + 399, + 296, + 423 + ], + "type": "text", + "content": " at each pixel " + }, + { + "bbox": [ + 55, + 399, + 296, + 423 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 399, + 296, + 423 + ], + "type": "text", + "content": " is modeled as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 111, + 434, + 295, + 466 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 434, + 295, + 466 + ], + "spans": [ + { + "bbox": [ + 111, + 434, + 295, + 466 + ], + "type": "interline_equation", + "content": "p \\left(\\mathbf {I} _ {x}\\right) = \\sum_ {k = 1} ^ {K} \\pi_ {k} \\mathcal {N} \\left(\\mathbf {I} _ {x} \\mid \\boldsymbol {\\mu} _ {k}, \\boldsymbol {\\Sigma} _ {k}\\right), \\tag {4}", + "image_path": "b6de30968af52c7e87b9dfd97e067565dcce4cb228e814a7229669aba78a50a5.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 472, + 295, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 295, + 508 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 295, + 508 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 472, + 295, + 508 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 472, + 295, + 508 + ], + "type": "text", + "content": " is the number of normal distributions, and " + }, + { + "bbox": [ + 55, + 472, + 295, + 508 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}_k" + }, + { + "bbox": [ + 55, + 472, + 295, + 508 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 472, + 295, + 508 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}_k" + }, + { + "bbox": [ + 55, + 472, + 295, + 508 + ], + "type": "text", + "content": " are the mean vector and covariance matrix. " + }, + { + "bbox": [ + 55, + 472, + 295, + 508 + ], + "type": "inline_equation", + "content": "\\pi_k" + }, + { + "bbox": [ + 55, + 472, + 295, + 508 + ], + "type": "text", + "content": " is a weighting coefficient, in which " + }, + { + "bbox": [ + 55, + 472, + 295, + 508 + ], + "type": "inline_equation", + "content": "\\sum_{k=1}^{K} \\pi_k = 1" + }, + { + "bbox": [ + 55, + 472, + 295, + 508 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 509, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 295, + 616 + ], + "type": "text", + "content": "In Eq. (4), " + }, + { + "bbox": [ + 55, + 509, + 295, + 616 + ], + "type": "inline_equation", + "content": "p(\\mathbf{I}_x)" + }, + { + "bbox": [ + 55, + 509, + 295, + 616 + ], + "type": "text", + "content": " represents a likelihood that all " + }, + { + "bbox": [ + 55, + 509, + 295, + 616 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 509, + 295, + 616 + ], + "type": "text", + "content": " spectrum can explain the observed spectrum, quantified by the linear combination of " + }, + { + "bbox": [ + 55, + 509, + 295, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\mathbf{I}_x|\\boldsymbol{\\mu}_k,\\boldsymbol{\\Sigma}_k)" + }, + { + "bbox": [ + 55, + 509, + 295, + 616 + ], + "type": "text", + "content": " and its proportion " + }, + { + "bbox": [ + 55, + 509, + 295, + 616 + ], + "type": "inline_equation", + "content": "\\pi_{k}" + }, + { + "bbox": [ + 55, + 509, + 295, + 616 + ], + "type": "text", + "content": " and thus " + }, + { + "bbox": [ + 55, + 509, + 295, + 616 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}_{k}" + }, + { + "bbox": [ + 55, + 509, + 295, + 616 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 509, + 295, + 616 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}_{k}" + }, + { + "bbox": [ + 55, + 509, + 295, + 616 + ], + "type": "text", + "content": " represent the " + }, + { + "bbox": [ + 55, + 509, + 295, + 616 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 509, + 295, + 616 + ], + "type": "text", + "content": "-th endmember spectrum and the spectral variability within the endmember. To derive the optimal parameter set " + }, + { + "bbox": [ + 55, + 509, + 295, + 616 + ], + "type": "inline_equation", + "content": "\\{\\pi_k^*,\\pmb {\\mu}_k^*,\\pmb {\\Sigma}_k^*\\}" + }, + { + "bbox": [ + 55, + 509, + 295, + 616 + ], + "type": "text", + "content": ", given a set of observed spectra " + }, + { + "bbox": [ + 55, + 509, + 295, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_x\\in \\mathbf{I}" + }, + { + "bbox": [ + 55, + 509, + 295, + 616 + ], + "type": "text", + "content": " of training samples, we apply an Expectation-Maximization (EM) algorithm [12] to iteratively refine the parameters." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 616, + 296, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 296, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 296, + 676 + ], + "type": "text", + "content": "Based on the results, we directly obtain the end-member spectra from the mean vectors, i.e., " + }, + { + "bbox": [ + 55, + 616, + 296, + 676 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}^{*} = \\{\\pmb{\\mu}_{1}^{*},\\pmb{\\mu}_{2}^{*},\\dots ,\\pmb{\\mu}_{K}^{*}\\}" + }, + { + "bbox": [ + 55, + 616, + 296, + 676 + ], + "type": "text", + "content": ". Then, we calculate a set of abundance " + }, + { + "bbox": [ + 55, + 616, + 296, + 676 + ], + "type": "inline_equation", + "content": "\\gamma^{*} = \\{\\gamma_{1}^{*},\\gamma_{2}^{*},\\dots ,\\gamma_{K}^{*}\\}" + }, + { + "bbox": [ + 55, + 616, + 296, + 676 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 616, + 296, + 676 + ], + "type": "inline_equation", + "content": "\\gamma_k^* (x)" + }, + { + "bbox": [ + 55, + 616, + 296, + 676 + ], + "type": "text", + "content": " is computed at each pixel coordinate " + }, + { + "bbox": [ + 55, + 616, + 296, + 676 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 616, + 296, + 676 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 104, + 685, + 295, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 685, + 295, + 717 + ], + "spans": [ + { + "bbox": [ + 104, + 685, + 295, + 717 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\gamma} _ {k} ^ {*} (x) = \\frac {\\pi_ {k} ^ {*} \\mathcal {N} \\left(\\mathbf {I} _ {x} \\mid \\boldsymbol {\\mu} _ {k} ^ {*} , \\boldsymbol {\\Sigma} _ {k} ^ {*}\\right)}{\\sum_ {j = 1} ^ {K} \\pi_ {j} ^ {*} \\mathcal {N} \\left(\\mathbf {I} _ {x} \\mid \\boldsymbol {\\mu} _ {j} ^ {*} , \\boldsymbol {\\Sigma} _ {j} ^ {*}\\right)}, \\tag {5}", + "image_path": "07a20392e0c30a430cd11620c3d3bd942c0e1de9a88707d5448267f9275f6e42.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "inline_equation", + "content": "\\gamma_k^*(x)" + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": " consists of an abundance map for the " + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "-th end-member spectrum." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 96, + 553, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 96, + 553, + 191 + ], + "spans": [ + { + "bbox": [ + 313, + 96, + 553, + 191 + ], + "type": "text", + "content": "In previous studies [47, 72], surface reflectance has been approximated using a low-dimensional model based on a subspace assumption, where a small number of spectral bases (e.g., 3) is sufficient for most natural images. These findings could lead to a tight subspace expression of the surface reflectance for practical use cases. In our experiments, we empirically selected " + }, + { + "bbox": [ + 313, + 96, + 553, + 191 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 96, + 553, + 191 + ], + "type": "text", + "content": " and observed best results with " + }, + { + "bbox": [ + 313, + 96, + 553, + 191 + ], + "type": "inline_equation", + "content": "K = 5" + }, + { + "bbox": [ + 313, + 96, + 553, + 191 + ], + "type": "text", + "content": ", aligning with the subspace assumption." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 192, + 553, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 192, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 192, + 553, + 228 + ], + "type": "text", + "content": "In the proposed model, the parameters pass through learnable layers, allowing for a more flexible representation that adapts to real-world images, mathematically expressed," + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 325, + 247, + 553, + 262 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 247, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 325, + 247, + 553, + 262 + ], + "type": "interline_equation", + "content": "\\mathbf {E} = W _ {e} \\boldsymbol {\\mu} ^ {*} \\in \\mathbb {R} ^ {K \\times C}, \\mathbf {U} = W _ {u} \\gamma^ {*} \\in \\mathbb {R} ^ {K \\times H \\times W}, \\tag {6}", + "image_path": "2b7e384af60f255377cbad30a21cd8c1dfd837aa0c0d29442faea4ffdfe4deb9.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 270, + 553, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 553, + 307 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 553, + 307 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 270, + 553, + 307 + ], + "type": "inline_equation", + "content": "W_{e}" + }, + { + "bbox": [ + 313, + 270, + 553, + 307 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 270, + 553, + 307 + ], + "type": "inline_equation", + "content": "W_{u}" + }, + { + "bbox": [ + 313, + 270, + 553, + 307 + ], + "type": "text", + "content": " are convolution layers. " + }, + { + "bbox": [ + 313, + 270, + 553, + 307 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 313, + 270, + 553, + 307 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 270, + 553, + 307 + ], + "type": "inline_equation", + "content": "\\mathbf{U}" + }, + { + "bbox": [ + 313, + 270, + 553, + 307 + ], + "type": "text", + "content": " are the tensors of endmembers and abundance maps, respectively, as the final output of the SU block." + } + ] + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 323, + 317, + 548, + 462 + ], + "blocks": [ + { + "bbox": [ + 323, + 317, + 548, + 462 + ], + "lines": [ + { + "bbox": [ + 323, + 317, + 548, + 462 + ], + "spans": [ + { + "bbox": [ + 323, + 317, + 548, + 462 + ], + "type": "image", + "image_path": "10755f8188b178b33a2e1d42fea7d6230d7354d8e393e4edbffbcadf40a40232.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 472, + 555, + 516 + ], + "lines": [ + { + "bbox": [ + 313, + 472, + 555, + 516 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 555, + 516 + ], + "type": "text", + "content": "Figure 3. Cross-attention operations in a feature mixing block. It uses input spectral and visual features from a SU block and CNN, respectively, and produces the spatial-spectral features as output, iteratively refined with an " + }, + { + "bbox": [ + 313, + 472, + 555, + 516 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 472, + 555, + 516 + ], + "type": "text", + "content": "-step GRU." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 527, + 539, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 527, + 539, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 527, + 539, + 540 + ], + "type": "text", + "content": "3.3. Feature Mixing Block with Cross-Attention" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "content": "Feature mixing block (FMB) uses both the SU features and CNN features as input features. The CNN features are from a spatial feature extractor (SFE) which consists of 3 layers of convolutions, as illustrated in Fig. 2. Previous model-based methods neglected spatial information [2, 10, 72], degrading performance in complex scenes. In contrast, deep learning-based methods captured visual features but struggled to integrate physical modeling [14, 40]. In the proposed method, the SU features and CNN features are used complementarily to generate spatial-spectral features tailored for MS images. Fig. 3 presents the detailed operations in the FMB. The spectral and spatial features are fused using cross-attention and iteratively refined through a gated recurrent unit (GRU)." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2218" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "content": "The FMB employs a cross-attention mechanism of a transformer to fuse spectral and spatial features as shown in Fig. 3. In the process, given " + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "content": ", the spatial features derived from CNN are used as queries, and the end-members from a spectral unmixing block are used as keys and values as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 88, + 157, + 295, + 171 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 157, + 295, + 171 + ], + "spans": [ + { + "bbox": [ + 88, + 157, + 295, + 171 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} = W _ {q} \\mathbf {F} _ {n}, \\quad \\mathbf {K} = W _ {k} \\mathbf {E}, \\quad \\mathbf {V} = W _ {v} \\mathbf {E}, \\tag {7}", + "image_path": "63f0ca00b74c486f33212959ae218fbe2ad5ddee42da25599f698849740260c1.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 177, + 296, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 177, + 296, + 260 + ], + "spans": [ + { + "bbox": [ + 55, + 177, + 296, + 260 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 177, + 296, + 260 + ], + "type": "inline_equation", + "content": "\\mathbf{Q} \\in \\mathbb{R}^{D \\times H \\times W}" + }, + { + "bbox": [ + 55, + 177, + 296, + 260 + ], + "type": "text", + "content": " is the query, and " + }, + { + "bbox": [ + 55, + 177, + 296, + 260 + ], + "type": "inline_equation", + "content": "\\mathbf{K}, \\mathbf{V} \\in \\mathbb{R}^{K \\times D}" + }, + { + "bbox": [ + 55, + 177, + 296, + 260 + ], + "type": "text", + "content": " are the key and value. " + }, + { + "bbox": [ + 55, + 177, + 296, + 260 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_n" + }, + { + "bbox": [ + 55, + 177, + 296, + 260 + ], + "type": "text", + "content": " is a visual feature, which is iteratively refined using a " + }, + { + "bbox": [ + 55, + 177, + 296, + 260 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 177, + 296, + 260 + ], + "type": "text", + "content": "-step GRU. " + }, + { + "bbox": [ + 55, + 177, + 296, + 260 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_0" + }, + { + "bbox": [ + 55, + 177, + 296, + 260 + ], + "type": "text", + "content": " is initialized to " + }, + { + "bbox": [ + 55, + 177, + 296, + 260 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{cnn}" + }, + { + "bbox": [ + 55, + 177, + 296, + 260 + ], + "type": "text", + "content": " as visual features of a SFE network. " + }, + { + "bbox": [ + 55, + 177, + 296, + 260 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 177, + 296, + 260 + ], + "type": "text", + "content": " is its spectral dimension. " + }, + { + "bbox": [ + 55, + 177, + 296, + 260 + ], + "type": "inline_equation", + "content": "W_q \\in \\mathbb{R}^{D \\times D}" + }, + { + "bbox": [ + 55, + 177, + 296, + 260 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 177, + 296, + 260 + ], + "type": "inline_equation", + "content": "W_k, W_v \\in \\mathbb{R}^{D \\times C}" + }, + { + "bbox": [ + 55, + 177, + 296, + 260 + ], + "type": "text", + "content": " are learnable linear transformations to embed the features to a cross-attention block." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 262, + 296, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 262, + 296, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 296, + 300 + ], + "type": "text", + "content": "Spectral-wise correlations " + }, + { + "bbox": [ + 55, + 262, + 296, + 300 + ], + "type": "inline_equation", + "content": "\\mathbf{S} \\in \\mathbb{R}^{K \\times H \\times W}" + }, + { + "bbox": [ + 55, + 262, + 296, + 300 + ], + "type": "text", + "content": " between a query of spatial locations in " + }, + { + "bbox": [ + 55, + 262, + 296, + 300 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_n" + }, + { + "bbox": [ + 55, + 262, + 296, + 300 + ], + "type": "text", + "content": " and a key of spectral characteristics of " + }, + { + "bbox": [ + 55, + 262, + 296, + 300 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 55, + 262, + 296, + 300 + ], + "type": "text", + "content": " are computed as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 311, + 295, + 326 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 311, + 295, + 326 + ], + "spans": [ + { + "bbox": [ + 146, + 311, + 295, + 326 + ], + "type": "interline_equation", + "content": "\\mathbf {S} = \\mathbf {Q} \\cdot \\mathbf {K} ^ {T}, \\tag {8}", + "image_path": "c5813fc3c5322decb198bf19077a5043b9064274e9fc459e58600cec9da587aa.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 334, + 295, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 334, + 295, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 334, + 295, + 357 + ], + "type": "text", + "content": "where the attention scores calculate the relevance between the visual features and spectral features." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 358, + 296, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 358, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 358, + 296, + 407 + ], + "type": "text", + "content": "These spectral correlations are modified by adding abundance map " + }, + { + "bbox": [ + 55, + 358, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{U} \\in \\mathbb{R}^{K \\times H \\times W}" + }, + { + "bbox": [ + 55, + 358, + 296, + 407 + ], + "type": "text", + "content": " obtained from the SU block, which generating an attention weight " + }, + { + "bbox": [ + 55, + 358, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{A} \\in \\mathbb{R}^{K \\times H \\times W}" + }, + { + "bbox": [ + 55, + 358, + 296, + 407 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 127, + 417, + 295, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 417, + 295, + 430 + ], + "spans": [ + { + "bbox": [ + 127, + 417, + 295, + 430 + ], + "type": "interline_equation", + "content": "\\mathbf {A} = \\operatorname {s o f t m a x} (\\mathbf {S} + \\mathbf {U}), \\tag {9}", + "image_path": "186d4ed6f63cce6dae5d8cca061663b82a68413df1b42788bb9ff4239ddacc57.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 438, + 296, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 438, + 296, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 438, + 296, + 475 + ], + "type": "text", + "content": "where softmax is a softmax function. Then, we compute an attention output through a weighted sum between attention weights and values, given as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 149, + 487, + 295, + 500 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 487, + 295, + 500 + ], + "spans": [ + { + "bbox": [ + 149, + 487, + 295, + 500 + ], + "type": "interline_equation", + "content": "\\mathbf {O} = \\mathbf {A} \\cdot \\mathbf {V}, \\tag {10}", + "image_path": "ae0e19879a7750cb9c4eb84ef5da3367c273491685180f642a01d22080afc60c.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 508, + 241, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 508, + 241, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 508, + 241, + 521 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 508, + 241, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{O} \\in \\mathbb{R}^{D \\times H \\times W}" + }, + { + "bbox": [ + 55, + 508, + 241, + 521 + ], + "type": "text", + "content": " is the attention output." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 522, + 296, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 296, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 296, + 571 + ], + "type": "text", + "content": "In the FMB, the output undergoes an update process using a GRU followed by a two-layer fully connected network (MLP), which iteratively refines the output feature with the query features. Mathematically, this can be expressed as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 113, + 582, + 295, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 582, + 295, + 597 + ], + "spans": [ + { + "bbox": [ + 113, + 582, + 295, + 597 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {n + 1} = \\operatorname {M L P} \\left(\\operatorname {G R U} \\left(\\mathbf {F} _ {n}, \\mathbf {O}\\right)\\right), \\tag {11}", + "image_path": "fe4bc03ad2a105face3701354c8eee02f36157ea92bfad3154225c7e9702f94d.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 604, + 296, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 604, + 296, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 604, + 296, + 688 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 604, + 296, + 688 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_n" + }, + { + "bbox": [ + 55, + 604, + 296, + 688 + ], + "type": "text", + "content": " is updated for " + }, + { + "bbox": [ + 55, + 604, + 296, + 688 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 604, + 296, + 688 + ], + "type": "text", + "content": " iterations to fuse the spectral and spatial features, until the final output " + }, + { + "bbox": [ + 55, + 604, + 296, + 688 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_N" + }, + { + "bbox": [ + 55, + 604, + 296, + 688 + ], + "type": "text", + "content": " is generated. " + }, + { + "bbox": [ + 55, + 604, + 296, + 688 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_N" + }, + { + "bbox": [ + 55, + 604, + 296, + 688 + ], + "type": "text", + "content": " represents a refined reflectance feature, maintaining the same dimensions as the input CNN features while incorporating spectral information from the SU method. " + }, + { + "bbox": [ + 55, + 604, + 296, + 688 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 604, + 296, + 688 + ], + "type": "text", + "content": " is set to 6 in our experiments. The Fig. 4 shows the gradually refined features during the iterative attention process of FMB." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 689, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 712 + ], + "type": "text", + "content": "We present a pseudo-code implementation in Algorithm 1." + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 316, + 72, + 347, + 125 + ], + "blocks": [ + { + "bbox": [ + 316, + 72, + 347, + 125 + ], + "lines": [ + { + "bbox": [ + 316, + 72, + 347, + 125 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 347, + 125 + ], + "type": "image", + "image_path": "c0988fd10e984b7d9b6b59d3f0d366cfdaaf6b8ed0bcb9e1b88f56cabb5c219d.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 347, + 78, + 409, + 125 + ], + "blocks": [ + { + "bbox": [ + 377, + 73, + 394, + 79 + ], + "lines": [ + { + "bbox": [ + 377, + 73, + 394, + 79 + ], + "spans": [ + { + "bbox": [ + 377, + 73, + 394, + 79 + ], + "type": "text", + "content": "Step 0" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 347, + 78, + 409, + 125 + ], + "lines": [ + { + "bbox": [ + 347, + 78, + 409, + 125 + ], + "spans": [ + { + "bbox": [ + 347, + 78, + 409, + 125 + ], + "type": "image", + "image_path": "431289af9814dc71ef82f6ea1cd1be6029701b2d8a88f17d12ead44af8b73b12.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 410, + 78, + 455, + 125 + ], + "blocks": [ + { + "bbox": [ + 425, + 73, + 440, + 79 + ], + "lines": [ + { + "bbox": [ + 425, + 73, + 440, + 79 + ], + "spans": [ + { + "bbox": [ + 425, + 73, + 440, + 79 + ], + "type": "text", + "content": "Step 2" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 410, + 78, + 455, + 125 + ], + "lines": [ + { + "bbox": [ + 410, + 78, + 455, + 125 + ], + "spans": [ + { + "bbox": [ + 410, + 78, + 455, + 125 + ], + "type": "image", + "image_path": "ab919df20e4a64a859b9bbad83f919bbafd1c70d982ae1ef0338c2ced7165700.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 456, + 78, + 501, + 125 + ], + "blocks": [ + { + "bbox": [ + 472, + 73, + 487, + 79 + ], + "lines": [ + { + "bbox": [ + 472, + 73, + 487, + 79 + ], + "spans": [ + { + "bbox": [ + 472, + 73, + 487, + 79 + ], + "type": "text", + "content": "Step 4" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 456, + 78, + 501, + 125 + ], + "lines": [ + { + "bbox": [ + 456, + 78, + 501, + 125 + ], + "spans": [ + { + "bbox": [ + 456, + 78, + 501, + 125 + ], + "type": "image", + "image_path": "b55cfaaa440ecd621c7f69e081b013ef357b119573c56294e185337f319709c5.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 502, + 78, + 549, + 125 + ], + "blocks": [ + { + "bbox": [ + 517, + 73, + 533, + 79 + ], + "lines": [ + { + "bbox": [ + 517, + 73, + 533, + 79 + ], + "spans": [ + { + "bbox": [ + 517, + 73, + 533, + 79 + ], + "type": "text", + "content": "Step 6" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 502, + 78, + 549, + 125 + ], + "lines": [ + { + "bbox": [ + 502, + 78, + 549, + 125 + ], + "spans": [ + { + "bbox": [ + 502, + 78, + 549, + 125 + ], + "type": "image", + "image_path": "18b5f54d93b0fa10f0ed547ec3dfd8d91dda5e3b50b0fcdcd607a6b144bc0788.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 316, + 126, + 361, + 172 + ], + "blocks": [ + { + "bbox": [ + 316, + 126, + 361, + 172 + ], + "lines": [ + { + "bbox": [ + 316, + 126, + 361, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 126, + 361, + 172 + ], + "type": "image", + "image_path": "2ac9985a1f676375c4cc41e7db2192299294c12b3fa4ad17c56033f0a5aeebba.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 362, + 126, + 408, + 172 + ], + "blocks": [ + { + "bbox": [ + 362, + 126, + 408, + 172 + ], + "lines": [ + { + "bbox": [ + 362, + 126, + 408, + 172 + ], + "spans": [ + { + "bbox": [ + 362, + 126, + 408, + 172 + ], + "type": "image", + "image_path": "4749dbe1c0ad4f659388b234c7d3a97332abe5f4cd8da8bc9d65ec60f1a3a3de.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 410, + 126, + 455, + 172 + ], + "blocks": [ + { + "bbox": [ + 410, + 126, + 455, + 172 + ], + "lines": [ + { + "bbox": [ + 410, + 126, + 455, + 172 + ], + "spans": [ + { + "bbox": [ + 410, + 126, + 455, + 172 + ], + "type": "image", + "image_path": "1692a0d17ba44fcda318e61d76919a70f6913785c478d4ce0b57f43883f912c7.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 456, + 126, + 501, + 172 + ], + "blocks": [ + { + "bbox": [ + 456, + 126, + 501, + 172 + ], + "lines": [ + { + "bbox": [ + 456, + 126, + 501, + 172 + ], + "spans": [ + { + "bbox": [ + 456, + 126, + 501, + 172 + ], + "type": "image", + "image_path": "00a10150855f298dbf6bd286ed4b2cf3e313babd847e444a2776baaa3434a8eb.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 503, + 126, + 549, + 172 + ], + "blocks": [ + { + "bbox": [ + 503, + 126, + 549, + 172 + ], + "lines": [ + { + "bbox": [ + 503, + 126, + 549, + 172 + ], + "spans": [ + { + "bbox": [ + 503, + 126, + 549, + 172 + ], + "type": "image", + "image_path": "81ee8b27128cbd811d6e1e256737e34687487b6e25dfe3ee7d514f97716d5dcf.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 316, + 174, + 361, + 220 + ], + "blocks": [ + { + "bbox": [ + 316, + 174, + 361, + 220 + ], + "lines": [ + { + "bbox": [ + 316, + 174, + 361, + 220 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 361, + 220 + ], + "type": "image", + "image_path": "693f936579181e53516e3ab149d5052e388e28dfe3e0b2d0c78b7d7d97501cfd.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 231, + 555, + 276 + ], + "lines": [ + { + "bbox": [ + 313, + 231, + 555, + 276 + ], + "spans": [ + { + "bbox": [ + 313, + 231, + 555, + 276 + ], + "type": "text", + "content": "Figure 4. Progressive feature refinement visualization through iterative stages in FMB. The leftmost image shows a rendered RGB of input MS image. From the initial query feature of FMB, the features are gradually refined from step 0 to step 6." + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 362, + 174, + 408, + 220 + ], + "blocks": [ + { + "bbox": [ + 362, + 174, + 408, + 220 + ], + "lines": [ + { + "bbox": [ + 362, + 174, + 408, + 220 + ], + "spans": [ + { + "bbox": [ + 362, + 174, + 408, + 220 + ], + "type": "image", + "image_path": "5b2aa6b73ee539e167fae80cf8a024eace372fd6a5d61bc1ecb93a471c11e3c8.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 410, + 174, + 455, + 220 + ], + "blocks": [ + { + "bbox": [ + 410, + 174, + 455, + 220 + ], + "lines": [ + { + "bbox": [ + 410, + 174, + 455, + 220 + ], + "spans": [ + { + "bbox": [ + 410, + 174, + 455, + 220 + ], + "type": "image", + "image_path": "f32866ff76c3416531b348f48bed73547720ac9a596110c4f98d403a676d3b77.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 456, + 174, + 501, + 220 + ], + "blocks": [ + { + "bbox": [ + 456, + 174, + 501, + 220 + ], + "lines": [ + { + "bbox": [ + 456, + 174, + 501, + 220 + ], + "spans": [ + { + "bbox": [ + 456, + 174, + 501, + 220 + ], + "type": "image", + "image_path": "d4709571de54ef66bc34414813c6840056323c653589e63e7003797d8746feba.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 503, + 174, + 549, + 220 + ], + "blocks": [ + { + "bbox": [ + 503, + 174, + 549, + 220 + ], + "lines": [ + { + "bbox": [ + 503, + 174, + 549, + 220 + ], + "spans": [ + { + "bbox": [ + 503, + 174, + 549, + 220 + ], + "type": "image", + "image_path": "f6fd1d05109772f3914010b5f5557ea004db5f47a1b9ca46fed30988e80c1083.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "code", + "bbox": [ + 326, + 308, + 541, + 536 + ], + "blocks": [ + { + "bbox": [ + 321, + 293, + 509, + 306 + ], + "lines": [ + { + "bbox": [ + 321, + 293, + 509, + 306 + ], + "spans": [ + { + "bbox": [ + 321, + 293, + 509, + 306 + ], + "type": "text", + "content": "Algorithm 1: SU-CNN Feature Mixing Block" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 326, + 308, + 541, + 536 + ], + "lines": [ + { + "bbox": [ + 326, + 308, + 541, + 536 + ], + "spans": [ + { + "bbox": [ + 326, + 308, + 541, + 536 + ], + "type": "text", + "content": "Input: " + }, + { + "bbox": [ + 326, + 308, + 541, + 536 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_0\\gets \\mathbf{F}_{cnn}\\in \\mathbb{R}^{D\\times H\\times W}" + }, + { + "bbox": [ + 326, + 308, + 541, + 536 + ], + "type": "text", + "content": " : CNN feature \n" + }, + { + "bbox": [ + 326, + 308, + 541, + 536 + ], + "type": "inline_equation", + "content": "\\mathbf{E}\\in \\mathbb{R}^{K\\times C}" + }, + { + "bbox": [ + 326, + 308, + 541, + 536 + ], + "type": "text", + "content": " : Endmember from SU block \n" + }, + { + "bbox": [ + 326, + 308, + 541, + 536 + ], + "type": "inline_equation", + "content": "\\mathbf{U}\\in \\mathbb{R}^{K\\times H\\times W}" + }, + { + "bbox": [ + 326, + 308, + 541, + 536 + ], + "type": "text", + "content": " : Abundance map from SU block \nOutput: " + }, + { + "bbox": [ + 326, + 308, + 541, + 536 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{ref}}\\leftarrow \\mathbf{F}_N\\in \\mathbb{R}^{D\\times H\\times W}" + }, + { + "bbox": [ + 326, + 308, + 541, + 536 + ], + "type": "text", + "content": " : Updated feature map for N iterations \n" + }, + { + "bbox": [ + 326, + 308, + 541, + 536 + ], + "type": "inline_equation", + "content": "\\mathbf{K}\\leftarrow W_{k}\\mathbf{E}" + }, + { + "bbox": [ + 326, + 308, + 541, + 536 + ], + "type": "text", + "content": " //Key projection \n" + }, + { + "bbox": [ + 326, + 308, + 541, + 536 + ], + "type": "inline_equation", + "content": "\\mathbf{V}\\leftarrow W_{v}\\mathbf{E}" + }, + { + "bbox": [ + 326, + 308, + 541, + 536 + ], + "type": "text", + "content": " //Value projection \nFunction FeatureMixing " + }, + { + "bbox": [ + 326, + 308, + 541, + 536 + ], + "type": "inline_equation", + "content": "(\\mathbf{F}_0,\\mathbf{K},\\mathbf{V},\\mathbf{U})" + }, + { + "bbox": [ + 326, + 308, + 541, + 536 + ], + "type": "text", + "content": " .. \nfor " + }, + { + "bbox": [ + 326, + 308, + 541, + 536 + ], + "type": "inline_equation", + "content": "n = 0" + }, + { + "bbox": [ + 326, + 308, + 541, + 536 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 326, + 308, + 541, + 536 + ], + "type": "inline_equation", + "content": "N - 1" + }, + { + "bbox": [ + 326, + 308, + 541, + 536 + ], + "type": "text", + "content": " do \nQ←WqFn //Query projection \nS←Q·KT// Spectral correlation \nA←softmax(S+U) //Attention weights \nO←A·V //Attention output \nFn+1←MLP(GRU(Fn,O)) //Feature update with GRU and MLP \nend \nreturn FN \n" + }, + { + "bbox": [ + 326, + 308, + 541, + 536 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_N\\gets" + }, + { + "bbox": [ + 326, + 308, + 541, + 536 + ], + "type": "text", + "content": " FeatureMixing(F0,K,V,U)" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "code_body" + } + ], + "index": 36, + "sub_type": "algorithm" + }, + { + "bbox": [ + 314, + 562, + 555, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 562, + 555, + 586 + ], + "spans": [ + { + "bbox": [ + 314, + 562, + 555, + 586 + ], + "type": "text", + "content": "3.4. Illumination Spectrum Estimation in Decoder Blocks" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": "In decoder blocks, the proposed network generates an estimation of the illumination spectrum, which is denoted by " + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{L}}" + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": ", at the end. The network uses two branches to calculate not only the illumination spectrum estimation but also the surface reflectance approximation, which is denoted by " + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{R}}" + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": ". The approximation of the surface reflectance in the decoder enables the network to use a joint IRSS framework, rather than producing an exact surface reflectance in the absence of GT. As shown in Fig. 2, " + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_N" + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": " is fed into a reflectance decoder, consisting of three transposed convolution layers, to" + } + ] + } + ], + "index": 38 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2219" + } + ] + } + ], + "index": 39 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "type": "text", + "content": "calculate " + }, + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{R}}" + }, + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "type": "text", + "content": " as intermediate results in the IRSS. Specifically, we use " + }, + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "type": "inline_equation", + "content": "3\\times 3" + }, + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "type": "text", + "content": " convolution kernels in the implementation. The reflectance decoder uses the only " + }, + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_N" + }, + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "type": "text", + "content": " as input to facilitate the optimization and fully focus on the reflectance prediction." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 132, + 295, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 132, + 295, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 132, + 295, + 240 + ], + "type": "text", + "content": "The illumination spectrum block is built with a series of convolutional and linear layers, in which an input is constructed by concatenating " + }, + { + "bbox": [ + 55, + 132, + 295, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{cnn}" + }, + { + "bbox": [ + 55, + 132, + 295, + 240 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 132, + 295, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_N" + }, + { + "bbox": [ + 55, + 132, + 295, + 240 + ], + "type": "text", + "content": ". The subsequent convolutional layers are employed to extract spatially localized spectral features, effectively capturing illumination in the target scene. We employ two layers of convolution kernel with stride 2. Rectified linear unit (ReLU) activation function is applied after all convolution and linear layers in the block, except for the final output linear layer." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 246, + 141, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 246, + 141, + 257 + ], + "spans": [ + { + "bbox": [ + 55, + 246, + 141, + 257 + ], + "type": "text", + "content": "3.5. Loss function" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 265, + 295, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 265, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 265, + 295, + 300 + ], + "type": "text", + "content": "For training the proposed architecture, we employ a reconstruction loss " + }, + { + "bbox": [ + 55, + 265, + 295, + 300 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{RC}}" + }, + { + "bbox": [ + 55, + 265, + 295, + 300 + ], + "type": "text", + "content": " and an illumination estimation loss " + }, + { + "bbox": [ + 55, + 265, + 295, + 300 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{AE}}" + }, + { + "bbox": [ + 55, + 265, + 295, + 300 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 265, + 295, + 300 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{RC}}" + }, + { + "bbox": [ + 55, + 265, + 295, + 300 + ], + "type": "text", + "content": " is computed using an " + }, + { + "bbox": [ + 55, + 265, + 295, + 300 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 55, + 265, + 295, + 300 + ], + "type": "text", + "content": " distance between " + }, + { + "bbox": [ + 55, + 265, + 295, + 300 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 55, + 265, + 295, + 300 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 265, + 295, + 300 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}" + }, + { + "bbox": [ + 55, + 265, + 295, + 300 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 91, + 309, + 295, + 336 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 309, + 295, + 336 + ], + "spans": [ + { + "bbox": [ + 91, + 309, + 295, + 336 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {R C}} = \\sum_ {x \\in (h, w)} \\sum_ {\\lambda \\in \\Lambda} \\| \\hat {\\mathbf {I}} (x, \\lambda) - \\mathbf {I} (x, \\lambda) \\| _ {1}, \\tag {12}", + "image_path": "676997ff9ca15176e1f4b92769455bcbfcc837afb2abd1ec19bf8437d81b030b.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 340, + 295, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 340, + 295, + 375 + ], + "spans": [ + { + "bbox": [ + 55, + 340, + 295, + 375 + ], + "type": "text", + "content": "where the loss is computed over all spatial spectral locations, i.e., " + }, + { + "bbox": [ + 55, + 340, + 295, + 375 + ], + "type": "inline_equation", + "content": "h = 1 \\cdots H" + }, + { + "bbox": [ + 55, + 340, + 295, + 375 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 340, + 295, + 375 + ], + "type": "inline_equation", + "content": "w = 1 \\cdots W" + }, + { + "bbox": [ + 55, + 340, + 295, + 375 + ], + "type": "text", + "content": ", and all channels " + }, + { + "bbox": [ + 55, + 340, + 295, + 375 + ], + "type": "inline_equation", + "content": "\\Lambda" + }, + { + "bbox": [ + 55, + 340, + 295, + 375 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 340, + 295, + 375 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}" + }, + { + "bbox": [ + 55, + 340, + 295, + 375 + ], + "type": "text", + "content": " is reconstructed using an integral of " + }, + { + "bbox": [ + 55, + 340, + 295, + 375 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{L}}" + }, + { + "bbox": [ + 55, + 340, + 295, + 375 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 340, + 295, + 375 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{R}}" + }, + { + "bbox": [ + 55, + 340, + 295, + 375 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 376, + 288, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 376, + 288, + 387 + ], + "spans": [ + { + "bbox": [ + 67, + 376, + 288, + 387 + ], + "type": "text", + "content": "We also utilize an angular error (AE), " + }, + { + "bbox": [ + 67, + 376, + 288, + 387 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{AE}}" + }, + { + "bbox": [ + 67, + 376, + 288, + 387 + ], + "type": "text", + "content": ", defined as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 117, + 394, + 295, + 426 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 394, + 295, + 426 + ], + "spans": [ + { + "bbox": [ + 117, + 394, + 295, + 426 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {A E}} = \\arccos \\left(\\frac {\\hat {\\mathbf {L}} \\cdot \\mathbf {L}}{\\| \\hat {\\mathbf {L}} \\| \\| \\mathbf {L} \\|}\\right), \\tag {13}", + "image_path": "81282ba67eee31dec5d12547afde437ae554d0a3b4d20e4520ba080a925e7561.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 434, + 295, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 295, + 529 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 434, + 295, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{L}" + }, + { + "bbox": [ + 55, + 434, + 295, + 529 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 434, + 295, + 529 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{L}}" + }, + { + "bbox": [ + 55, + 434, + 295, + 529 + ], + "type": "text", + "content": " are the ground-truth (GT) and predicted illumination spectrum vector, respectively. For illumination spectrum estimation, capturing relative differences of illuminant among multi-channels is crucial. Spectral orientation defines its color characteristics, which would affect color constancy, while the intensity of illuminant reflects brightness. Thus, we use an angular error to focus on the orientation by measuring the angle between vectors." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 530, + 225, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 530, + 225, + 540 + ], + "spans": [ + { + "bbox": [ + 67, + 530, + 225, + 540 + ], + "type": "text", + "content": "The total loss function " + }, + { + "bbox": [ + 67, + 530, + 225, + 540 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 67, + 530, + 225, + 540 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 129, + 553, + 294, + 566 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 553, + 294, + 566 + ], + "spans": [ + { + "bbox": [ + 129, + 553, + 294, + 566 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\alpha_ {1} \\mathcal {L} _ {\\mathrm {R C}} + \\alpha_ {2} \\mathcal {L} _ {\\mathrm {A E}}, \\tag {14}", + "image_path": "cb923ed11b6279b3a309e0b3a92cdc153c99ad591998eef1253284ad24f163f7.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 570, + 295, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 594 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 570, + 295, + 594 + ], + "type": "inline_equation", + "content": "\\alpha_{1}" + }, + { + "bbox": [ + 55, + 570, + 295, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 570, + 295, + 594 + ], + "type": "inline_equation", + "content": "\\alpha_{2}" + }, + { + "bbox": [ + 55, + 570, + 295, + 594 + ], + "type": "text", + "content": " are weighting factors, which are set to 0.5 and 1, respectively, in our experiments." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 604, + 181, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 604, + 181, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 604, + 181, + 616 + ], + "type": "text", + "content": "4. Experimental Results" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 623, + 216, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 623, + 216, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 623, + 216, + 635 + ], + "type": "text", + "content": "4.1. Spectral Reflectance Datasets" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 641, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 295, + 713 + ], + "type": "text", + "content": "We evaluate the proposed technique using two spectral reflectance datasets: KAUST [40] and CAVE [66]. In the datasets, we use 31 channels ranging from " + }, + { + "bbox": [ + 55, + 641, + 295, + 713 + ], + "type": "inline_equation", + "content": "400nm" + }, + { + "bbox": [ + 55, + 641, + 295, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 641, + 295, + 713 + ], + "type": "inline_equation", + "content": "700nm" + }, + { + "bbox": [ + 55, + 641, + 295, + 713 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 55, + 641, + 295, + 713 + ], + "type": "inline_equation", + "content": "10nm" + }, + { + "bbox": [ + 55, + 641, + 295, + 713 + ], + "type": "text", + "content": " intervals. We relight the scenes by multiplying the surface reflectance of each pixel with an illuminant spectra matrix with 59 standard illuminants from" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "content": "[49] resulting in a 31-channel spectral image. We randomly select " + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "content": " of the KAUST reflectance and illuminant data for training and the remaining " + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "content": " used for testing as in [40]. CAVE dataset is utilized only for testing the models. We evaluate the performance of the proposed technique with existing methods including both statistic-based and learning-based methods, GrayEdge [27], LRMF [72], ISNL [58], PWIR [54], DUN [40] and BeyondRGB [14]. We used mean- " + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "inline_equation", + "content": "\\Delta \\mathrm{A}_{\\mathrm{MS}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "content": " and std- " + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "inline_equation", + "content": "\\Delta \\mathrm{A}_{\\mathrm{MS}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "content": " for evaluation metrics, which are the average and standard deviation of an angular error measured between the GT and predicted illumination spectrum." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 216, + 555, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 216, + 555, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 555, + 324 + ], + "type": "text", + "content": "Tab. 1 exhibits the performance comparisons, showing that the proposed technique achieves the best results on both datasets. The proposed technique produces 5.55 and 3.29 in terms of mean- " + }, + { + "bbox": [ + 313, + 216, + 555, + 324 + ], + "type": "inline_equation", + "content": "\\Delta \\mathrm{A}_{\\mathrm{MS}}" + }, + { + "bbox": [ + 313, + 216, + 555, + 324 + ], + "type": "text", + "content": " and std- " + }, + { + "bbox": [ + 313, + 216, + 555, + 324 + ], + "type": "inline_equation", + "content": "\\Delta \\mathrm{A}_{\\mathrm{MS}}" + }, + { + "bbox": [ + 313, + 216, + 555, + 324 + ], + "type": "text", + "content": ", respectively, in KAUST dataset. For CAVE dataset, which is not used for model training, it shows reliable performance achieving 5.93 and 3.90 in mean- " + }, + { + "bbox": [ + 313, + 216, + 555, + 324 + ], + "type": "inline_equation", + "content": "\\Delta \\mathrm{A}_{\\mathrm{MS}}" + }, + { + "bbox": [ + 313, + 216, + 555, + 324 + ], + "type": "text", + "content": " and std- " + }, + { + "bbox": [ + 313, + 216, + 555, + 324 + ], + "type": "inline_equation", + "content": "\\Delta \\mathrm{A}_{\\mathrm{MS}}" + }, + { + "bbox": [ + 313, + 216, + 555, + 324 + ], + "type": "text", + "content": ", respectively. These results demonstrate that the proposed technique holds stable performance on unseen reflectance data." + } + ] + } + ], + "index": 16 + }, + { + "type": "table", + "bbox": [ + 320, + 334, + 549, + 507 + ], + "blocks": [ + { + "bbox": [ + 320, + 334, + 549, + 507 + ], + "lines": [ + { + "bbox": [ + 320, + 334, + 549, + 507 + ], + "spans": [ + { + "bbox": [ + 320, + 334, + 549, + 507 + ], + "type": "table", + "html": "
DatasetMethodmean-ΔAMS↓std-ΔAMS↓
KAUSTGrayEdge [27]11.468.02
LRMF [72]24.0623.49
ISNL [58]20.0511.46
PWIR [54]28.0721.20
DUN [40]9.176.30
BeyondRGB [14]7.304.45
Ours5.553.29
CAVEGrayEdge [27]18.3314.89
LRMF [72]20.0515.47
ISNL [58]22.9216.62
PWIR [54]34.9521.20
DUN [40]16.0413.75
BeyondRGB [14]6.643.62
Ours5.933.90
", + "image_path": "1209e9800feb7f1aad746f87a533cc7269144fc3925f65ed55b5761f2319fa2b.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 515, + 553, + 537 + ], + "lines": [ + { + "bbox": [ + 313, + 515, + 553, + 537 + ], + "spans": [ + { + "bbox": [ + 313, + 515, + 553, + 537 + ], + "type": "text", + "content": "Table 1. Performance comparisons of the proposed technique and existing methods using KAUST and CAVE datasets." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 551, + 437, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 551, + 437, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 551, + 437, + 563 + ], + "type": "text", + "content": "4.2. Beyond RGB Dataset" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "The BeyondRGB [14] dataset is a comprehensive, real-world dataset, designed for MS image-related studies, focused on mobile devices. The dataset is organized into \"Lab\" and \"Field,\" which include a variety of objects with diverse colors and textures in indoor scenes and real-world data in outdoor scenes, both encompassing complex scenes with rich object content. For experimental results, we use the same data-split protocol of [14]. For comparisons, we used PWIR and BeyondRGB, because the other methods required reflectance data for training [40] or the input dimension needs to match the number of spectral bands in GT spectra [27, 58, 72]" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2220" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 70, + 170, + 137 + ], + "blocks": [ + { + "bbox": [ + 82, + 70, + 170, + 137 + ], + "lines": [ + { + "bbox": [ + 82, + 70, + 170, + 137 + ], + "spans": [ + { + "bbox": [ + 82, + 70, + 170, + 137 + ], + "type": "image", + "image_path": "11ad1a0741846c3c4dddda4357a2a6b7032b08cab2a80e65759fc0899361d933.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 82, + 137, + 168, + 200 + ], + "blocks": [ + { + "bbox": [ + 82, + 137, + 168, + 200 + ], + "lines": [ + { + "bbox": [ + 82, + 137, + 168, + 200 + ], + "spans": [ + { + "bbox": [ + 82, + 137, + 168, + 200 + ], + "type": "image", + "image_path": "289bf1a424a0c8c9de12839718b97ac3c8e1e846ee1b68c4cf628ea545c429e8.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 171, + 70, + 260, + 137 + ], + "blocks": [ + { + "bbox": [ + 171, + 70, + 260, + 137 + ], + "lines": [ + { + "bbox": [ + 171, + 70, + 260, + 137 + ], + "spans": [ + { + "bbox": [ + 171, + 70, + 260, + 137 + ], + "type": "image", + "image_path": "f661da813ea0cf9f722597cc1f1fb0c96868fb167a3d7f526caecda7c525e8de.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 173, + 137, + 260, + 200 + ], + "blocks": [ + { + "bbox": [ + 173, + 137, + 260, + 200 + ], + "lines": [ + { + "bbox": [ + 173, + 137, + 260, + 200 + ], + "spans": [ + { + "bbox": [ + 173, + 137, + 260, + 200 + ], + "type": "image", + "image_path": "a835ebdab787fe677dee1f9ef4bb28f36287917235550f78ca03a03fc79b93ec.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 261, + 71, + 349, + 137 + ], + "blocks": [ + { + "bbox": [ + 261, + 71, + 349, + 137 + ], + "lines": [ + { + "bbox": [ + 261, + 71, + 349, + 137 + ], + "spans": [ + { + "bbox": [ + 261, + 71, + 349, + 137 + ], + "type": "image", + "image_path": "3e3075dc6170a67dc305057d1765627955c5590c0338b4e1e3371737f1561f28.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 261, + 137, + 348, + 200 + ], + "blocks": [ + { + "bbox": [ + 261, + 137, + 348, + 200 + ], + "lines": [ + { + "bbox": [ + 261, + 137, + 348, + 200 + ], + "spans": [ + { + "bbox": [ + 261, + 137, + 348, + 200 + ], + "type": "image", + "image_path": "00f1eeb682dee850d3576b3c10190465f74e5ada1808e4267b3a1a336a12bc16.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 350, + 71, + 438, + 137 + ], + "blocks": [ + { + "bbox": [ + 350, + 71, + 438, + 137 + ], + "lines": [ + { + "bbox": [ + 350, + 71, + 438, + 137 + ], + "spans": [ + { + "bbox": [ + 350, + 71, + 438, + 137 + ], + "type": "image", + "image_path": "4738b24ec9c10711e8f5d0ff4e77cdd3f3ba0601e15974f358505f0e678f78bd.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 351, + 137, + 438, + 200 + ], + "blocks": [ + { + "bbox": [ + 351, + 137, + 438, + 200 + ], + "lines": [ + { + "bbox": [ + 351, + 137, + 438, + 200 + ], + "spans": [ + { + "bbox": [ + 351, + 137, + 438, + 200 + ], + "type": "image", + "image_path": "19b4f5c6332d03a7c0d84b2ad6d3b73a73d224b9eae1fd2e58bf53769e0e54ca.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 439, + 71, + 527, + 137 + ], + "blocks": [ + { + "bbox": [ + 439, + 71, + 527, + 137 + ], + "lines": [ + { + "bbox": [ + 439, + 71, + 527, + 137 + ], + "spans": [ + { + "bbox": [ + 439, + 71, + 527, + 137 + ], + "type": "image", + "image_path": "00d323253bdfa5b31541e95b7aa01b063e72b9b47bf6b50c56e84d3f8bc4e8ed.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 441, + 137, + 529, + 200 + ], + "blocks": [ + { + "bbox": [ + 441, + 137, + 529, + 200 + ], + "lines": [ + { + "bbox": [ + 441, + 137, + 529, + 200 + ], + "spans": [ + { + "bbox": [ + 441, + 137, + 529, + 200 + ], + "type": "image", + "image_path": "02bc1a3ecc451ab18c90ddc1a3354e43a83e5521e8db8f0bcad47fba7d3c3924.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 82, + 204, + 170, + 270 + ], + "blocks": [ + { + "bbox": [ + 82, + 204, + 170, + 270 + ], + "lines": [ + { + "bbox": [ + 82, + 204, + 170, + 270 + ], + "spans": [ + { + "bbox": [ + 82, + 204, + 170, + 270 + ], + "type": "image", + "image_path": "6984c8bc99acbf5a3f996854968a22675fe48dce008ffc68fc67e39f939c1df9.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 82, + 271, + 169, + 335 + ], + "blocks": [ + { + "bbox": [ + 82, + 271, + 169, + 335 + ], + "lines": [ + { + "bbox": [ + 82, + 271, + 169, + 335 + ], + "spans": [ + { + "bbox": [ + 82, + 271, + 169, + 335 + ], + "type": "image", + "image_path": "a2de56cb48871d04b489e78d3f9a08c4d746c0f9e00178f8f7417b0f50248fe5.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 342, + 555, + 366 + ], + "lines": [ + { + "bbox": [ + 55, + 342, + 555, + 366 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 555, + 366 + ], + "type": "text", + "content": "Figure 5. Each of two rows shows MS images of BeyondRGB dataset in the first row and their estimated illuminant spectrum in the second row. The results of our model and PWIR are plotted with red and green dashed lines, respectively. The GT is shown with black solid lines." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 171, + 205, + 260, + 270 + ], + "blocks": [ + { + "bbox": [ + 171, + 205, + 260, + 270 + ], + "lines": [ + { + "bbox": [ + 171, + 205, + 260, + 270 + ], + "spans": [ + { + "bbox": [ + 171, + 205, + 260, + 270 + ], + "type": "image", + "image_path": "f0ebc663ad294d2015e2f9c4e6f8b801bc9b1522436280fc4b8700287e64a0c8.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 173, + 271, + 259, + 335 + ], + "blocks": [ + { + "bbox": [ + 173, + 271, + 259, + 335 + ], + "lines": [ + { + "bbox": [ + 173, + 271, + 259, + 335 + ], + "spans": [ + { + "bbox": [ + 173, + 271, + 259, + 335 + ], + "type": "image", + "image_path": "9f5e41a04383d260e3e73a0160c4d88d896bac651b42a2f9d87fd3e2949d9454.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 261, + 205, + 349, + 270 + ], + "blocks": [ + { + "bbox": [ + 261, + 205, + 349, + 270 + ], + "lines": [ + { + "bbox": [ + 261, + 205, + 349, + 270 + ], + "spans": [ + { + "bbox": [ + 261, + 205, + 349, + 270 + ], + "type": "image", + "image_path": "0e9e1f18ec475138c42ee36c28aa4a619997552c0aa325074701da5fa3597636.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 261, + 271, + 348, + 335 + ], + "blocks": [ + { + "bbox": [ + 261, + 271, + 348, + 335 + ], + "lines": [ + { + "bbox": [ + 261, + 271, + 348, + 335 + ], + "spans": [ + { + "bbox": [ + 261, + 271, + 348, + 335 + ], + "type": "image", + "image_path": "3e7c87d9958a4254da26edbeef788fc5d58139ca853997ba4dbc32d89f9b17ca.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 350, + 205, + 438, + 270 + ], + "blocks": [ + { + "bbox": [ + 350, + 205, + 438, + 270 + ], + "lines": [ + { + "bbox": [ + 350, + 205, + 438, + 270 + ], + "spans": [ + { + "bbox": [ + 350, + 205, + 438, + 270 + ], + "type": "image", + "image_path": "016805b07df054c1585d499879e09c69147bd087a78164f2e4ed7ef364fb7599.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 351, + 271, + 438, + 335 + ], + "blocks": [ + { + "bbox": [ + 351, + 271, + 438, + 335 + ], + "lines": [ + { + "bbox": [ + 351, + 271, + 438, + 335 + ], + "spans": [ + { + "bbox": [ + 351, + 271, + 438, + 335 + ], + "type": "image", + "image_path": "8001f9d055b0095d0df4209f3faa3ccae167aa815ad35be722bc14bfe7d8f2c5.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 439, + 205, + 527, + 270 + ], + "blocks": [ + { + "bbox": [ + 439, + 205, + 527, + 270 + ], + "lines": [ + { + "bbox": [ + 439, + 205, + 527, + 270 + ], + "spans": [ + { + "bbox": [ + 439, + 205, + 527, + 270 + ], + "type": "image", + "image_path": "fb6766194498200ae07dd9a0553cb47d241c0998567ba9e77f545b79265244b0.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 441, + 271, + 529, + 335 + ], + "blocks": [ + { + "bbox": [ + 441, + 271, + 529, + 335 + ], + "lines": [ + { + "bbox": [ + 441, + 271, + 529, + 335 + ], + "spans": [ + { + "bbox": [ + 441, + 271, + 529, + 335 + ], + "type": "image", + "image_path": "0aefebea1e66a61519e73628cdacf1172cfdd0e6c4c28b0ca8e9f664dc052743.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 377, + 296, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 377, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 377, + 296, + 498 + ], + "type": "text", + "content": "Tab. 2 shows the accuracy of the tested methods. It is clearly demonstrated with the results that the proposed technique provides superior accuracy compared to all the tested methods. Our technique significantly reduces the angular errors, achieving values of 3.65 and 5.97 in the Lab and Field datasets, respectively, in mean- " + }, + { + "bbox": [ + 55, + 377, + 296, + 498 + ], + "type": "inline_equation", + "content": "\\Delta \\mathrm{A}_{\\mathrm{MS}}" + }, + { + "bbox": [ + 55, + 377, + 296, + 498 + ], + "type": "text", + "content": ". In the BeyondRGB dataset, the proposed technique presents a greater performance gap compared to the second-best algorithm [14] than in KAUST and CAVE datasets, which can present the effectiveness of our approach in complex scenes." + } + ] + } + ], + "index": 21 + }, + { + "type": "table", + "bbox": [ + 62, + 506, + 290, + 598 + ], + "blocks": [ + { + "bbox": [ + 62, + 506, + 290, + 598 + ], + "lines": [ + { + "bbox": [ + 62, + 506, + 290, + 598 + ], + "spans": [ + { + "bbox": [ + 62, + 506, + 290, + 598 + ], + "type": "table", + "html": "
DatasetMethodmean-ΔAMS↓std-ΔAMS↓
LabPWIR [54]27.1010.86
BeyondRGB [14]5.922.92
Ours3.652.72
FieldPWIR [54]16.316.07
BeyondRGB [14]7.225.54
Ours5.973.90
", + "image_path": "81f340794992dc37565ad63c4c1645888b60c90752cff54aaec730ad6ffce552.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "table_body" + } + ], + "index": 22 + }, + { + "bbox": [ + 55, + 607, + 295, + 630 + ], + "lines": [ + { + "bbox": [ + 55, + 607, + 295, + 630 + ], + "spans": [ + { + "bbox": [ + 55, + 607, + 295, + 630 + ], + "type": "text", + "content": "Table 2. Performance comparisons of the proposed technique and existing methods using real-world MS dataset, BeyondRGB." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "Fig. 5 shows the results of illumination spectrum using the proposed technique in comparison with PWIR [54]. The top row of the image consists of Lab images, captured with controlled and artificial illuminations. From left to right, the illuminations are the spectrum of fluorescent, mixed LED, halogen lamp, incandescent light bulb, and LED illumi" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 377, + 555, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 377, + 555, + 534 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 555, + 534 + ], + "type": "text", + "content": "nant, which are found in daily environments. The bottom row contains real-world Field images, where outdoor images are illuminated by natural sunlight. The estimated illumination spectrum of our model, represented by the red dashed line, demonstrate a significantly more detailed alignment with the shape and peaks of the GT compared to PWIR. In the first image of the Lab set, the proposed technique shows a superior performance in estimating complex illumination conditions, such as those with multiple peaks characteristic of fluorescent lighting. Even for the bottom row comprising a wider variety of scenes, our model demonstrates robust and accurate illumination estimation performance." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 540, + 477, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 477, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 477, + 553 + ], + "type": "text", + "content": "4.3. Surface Reflectance Modeling" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 558, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 715 + ], + "type": "text", + "content": "We investigate the reconstruction results of surface reflectance on the KAUST and CAVE datasets, in which GT reflectance data is available. As shown in Fig. 6, we compare the estimated reflectance from our reflectance decoder and corresponding GT reflectance. From the results, we observe that even though the network has not directly seen GT reflectance during the training process, the surface reflectance is sufficiently considered while estimating illumination from images. Even in complex optical conditions like refractive surfaces with shadows shown in the window and glasses scenes, and reflection and diffraction effects demonstrated by the CD, our method effectively reconstruct the surface reflectance preserving these challenges." + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "2221" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 69, + 294, + 216 + ], + "blocks": [ + { + "bbox": [ + 57, + 69, + 294, + 216 + ], + "lines": [ + { + "bbox": [ + 57, + 69, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 57, + 69, + 294, + 216 + ], + "type": "image", + "image_path": "2275750dc7a9756e73f8c594d34c2cc84a2b14abfac46e21d69663875e741b08.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 224, + 296, + 280 + ], + "lines": [ + { + "bbox": [ + 55, + 224, + 296, + 280 + ], + "spans": [ + { + "bbox": [ + 55, + 224, + 296, + 280 + ], + "type": "text", + "content": "Figure 6. Visualization of reconstructed surface reflectance for KAUST and CAVE datasets. The first and second columns display GT and reconstructed surface reflectance rendered in RGB. The following columns display the estimated surface reflectance at each wavelength, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 290, + 296, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 290, + 296, + 339 + ], + "spans": [ + { + "bbox": [ + 55, + 290, + 296, + 339 + ], + "type": "text", + "content": "ing visual properties. The results align with our motivation for effective disentanglement of illumination and reflectance through physics-based approach and validate the effectiveness of our proposed method." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 345, + 149, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 345, + 149, + 358 + ], + "spans": [ + { + "bbox": [ + 55, + 345, + 149, + 358 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 363, + 295, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 363, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 54, + 363, + 295, + 579 + ], + "type": "text", + "content": "Network architecture: Tab. 3 presents the effectiveness of the proposed modules consisting of the network architecture. When the modules are turned off one-by-one, we observe a considerable reduction in accuracy. First, without SU block, the network relies solely on visual features without using spectral features, resulting in an approximate increase of 1.12 in angular error. Second, when disabling FMB, the spatial and spectral features are not fused appropriately but used independently to estimate the illumination spectrum. It increases approximately 3.65 of an angular error. Lastly, when the SFE module is turned off, we observe a dramatic degradation in accuracy. This result is expected, as the network is unable to exploit any visual features from the input image without the SFE. Fig. 7 shows the output features of the FMB, when disabling SFE or SU. Compared to the rendered RGB image on the left, the synthesized feature can be effectively used for illumination estimation by localizing regions relevant to the task." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 56, + 587, + 195, + 652 + ], + "blocks": [ + { + "bbox": [ + 56, + 587, + 195, + 652 + ], + "lines": [ + { + "bbox": [ + 56, + 587, + 195, + 652 + ], + "spans": [ + { + "bbox": [ + 56, + 587, + 195, + 652 + ], + "type": "table", + "html": "
SUFMBSFEmean-ΔAMS ↓
X✓✓5.52
✓X✓8.05
✓✓X24.64
✓✓✓4.40
", + "image_path": "134416055ec8fde66a6c84fe0cf44a1c1784e18878bd45cf1b73ddbd5d748658.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 202, + 587, + 296, + 652 + ], + "blocks": [ + { + "bbox": [ + 202, + 587, + 296, + 652 + ], + "lines": [ + { + "bbox": [ + 202, + 587, + 296, + 652 + ], + "spans": [ + { + "bbox": [ + 202, + 587, + 296, + 652 + ], + "type": "table", + "html": "
LAELCRcmean-ΔAMS↓
✓X4.72
X✓73.59
✓✓4.40
", + "image_path": "999f464e4f5b3be7143dfdfea7843f4c9b32ee361132bc23df1ae9358744cb70.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 660, + 295, + 683 + ], + "lines": [ + { + "bbox": [ + 55, + 660, + 295, + 683 + ], + "spans": [ + { + "bbox": [ + 55, + 660, + 295, + 683 + ], + "type": "text", + "content": "Table 3. Ablation studies of the sub-modules of the proposed network (left) and training loss functions (right)." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "Loss function: We conducted ablation tests in the loss terms. In the proposed method, using both " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{AE}}" + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{RC}}" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 72, + 553, + 255 + ], + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 255 + ], + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 255 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 255 + ], + "type": "image", + "image_path": "ca96fd750fbeb4f7bb731ecdd738b71dfb77ebc4c013cb56d5766cabcde34e3b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 264, + 555, + 321 + ], + "lines": [ + { + "bbox": [ + 313, + 264, + 555, + 321 + ], + "spans": [ + { + "bbox": [ + 313, + 264, + 555, + 321 + ], + "type": "text", + "content": "Figure 7. Output features from an FMB block in the ablation test. The second and the third columns display the features, when disabling SFE and SU, respectively. Our features in the last column highlight more relevant regions including high spectral diversity in local surfaces to achieve more enhanced accuracy." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 333, + 555, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 333, + 555, + 465 + ], + "spans": [ + { + "bbox": [ + 313, + 333, + 555, + 465 + ], + "type": "text", + "content": "together achieves the best performance. When " + }, + { + "bbox": [ + 313, + 333, + 555, + 465 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{RC}}" + }, + { + "bbox": [ + 313, + 333, + 555, + 465 + ], + "type": "text", + "content": " is not used, the performance decreases by approximately 0.32 compared to using both the losses. By incorporating " + }, + { + "bbox": [ + 313, + 333, + 555, + 465 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{RC}}" + }, + { + "bbox": [ + 313, + 333, + 555, + 465 + ], + "type": "text", + "content": ", the network can further learn the impact of a global illumination spectrum on the reconstructed image in terms of spatial representation. When excluding " + }, + { + "bbox": [ + 313, + 333, + 555, + 465 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{AE}}" + }, + { + "bbox": [ + 313, + 333, + 555, + 465 + ], + "type": "text", + "content": ", we observed substantial performance drops (e.g. 73 in mean-" + }, + { + "bbox": [ + 313, + 333, + 555, + 465 + ], + "type": "inline_equation", + "content": "\\Delta \\mathrm{A}_{\\mathrm{MS}}" + }, + { + "bbox": [ + 313, + 333, + 555, + 465 + ], + "type": "text", + "content": "). These results are expected, since relying solely on the reflectance for the image reconstruction may provide insufficient constraints for illumination estimation in the absence of GT. " + }, + { + "bbox": [ + 313, + 333, + 555, + 465 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{AE}}" + }, + { + "bbox": [ + 313, + 333, + 555, + 465 + ], + "type": "text", + "content": " plays a crucial role in the estimation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 483, + 388, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 388, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 388, + 496 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 506, + 555, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 555, + 626 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 555, + 626 + ], + "type": "text", + "content": "In this paper, we proposed illumination spectrum estimation technique for MS images. The proposed ISS technique used a learnable SU block which modeled surface reflectance, and a feature mixing block which fused spatial and spectral information using cross-attention operations. The spatial-spectral features were efficiently used to focus on important regions in complex scenes to enhance the accuracy of illumination estimation. Experimental results demonstrated that the proposed ISS technique provided a superiority to existing studies in various MS image datasets." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "In our future study, we will investigate illumination estimation under locally varying and multi-illuminant scenarios. While the current model is designed to estimate a global illumination spectrum, the proposed ISS technique can be extended to improve accuracy in estimating multiple illuminant by efficiently disentangling reflectance properties under locally varying lighting conditions." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2222" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 159 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 159 + ], + "type": "text", + "content": "This work was partly supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (RS-2021-II212068, Artificial Intelligence Innovation Hub, " + }, + { + "bbox": [ + 55, + 91, + 296, + 159 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 55, + 91, + 296, + 159 + ], + "type": "text", + "content": " ) and was partly supported by the NRF grant funded by MSIT (No.NRF-2022R1A2C4002052, " + }, + { + "bbox": [ + 55, + 91, + 296, + 159 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 55, + 91, + 296, + 159 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 182, + 115, + 194 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 182, + 115, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 182, + 115, + 194 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 201, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 201, + 296, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 201, + 296, + 235 + ], + "spans": [ + { + "bbox": [ + 61, + 201, + 296, + 235 + ], + "type": "text", + "content": "[1] Mahmoud Affifi and Michael S Brown. Sensor-independent illumination estimation for dnn models. arXiv preprint arXiv:1912.06888, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 236, + 296, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 236, + 296, + 280 + ], + "spans": [ + { + "bbox": [ + 61, + 236, + 296, + 280 + ], + "type": "text", + "content": "[2] Dongsheng An, Jinli Suo, Haoqian Wang, and Qionghai Dai. Illumination estimation from specular highlight in a multispectral image. Optics Express, 23(13):17008-17023, 2015. 2, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 282, + 295, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 282, + 295, + 326 + ], + "spans": [ + { + "bbox": [ + 61, + 282, + 295, + 326 + ], + "type": "text", + "content": "[3] Simone Bianco, Claudio Cusano, and Raimondo Schettini. Color constancy using cnns. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 81-89, 2015. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 327, + 296, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 327, + 296, + 372 + ], + "spans": [ + { + "bbox": [ + 62, + 327, + 296, + 372 + ], + "type": "text", + "content": "[4] Simone Bianco, Claudio Cusano, and Raimondo Schettini. Single and multiple illuminant estimation using convolutional neural networks. IEEE Transactions on Image Processing, 26(9):4347-4362, 2017. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 373, + 296, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 373, + 296, + 439 + ], + "spans": [ + { + "bbox": [ + 62, + 373, + 296, + 439 + ], + "type": "text", + "content": "[5] José M Bioucas-Dias, Antonio Plaza, Nicolas Dobigeon, Mario Parente, Qian Du, Paul Gader, and Jocelyn Chanussot. Hyperspectral unmixing overview: Geometrical, statistical, and sparse regression-based approaches. IEEE journal of selected topics in applied earth observations and remote sensing, 5(2):354-379, 2012. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 441, + 296, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 441, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 62, + 441, + 296, + 495 + ], + "type": "text", + "content": "[6] Ricardo Augusto Borsoi, Tales Imbiriba, and José Carlos Moreira Bermudez. Deep generative endmember modeling: An application to unsupervised spectral unmixing. IEEE Transactions on Computational Imaging, 6:374-384, 2019. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 498, + 296, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 498, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 62, + 498, + 296, + 529 + ], + "type": "text", + "content": "[7] David H Brainard and Brian A Wandell. Analysis of the retina theory of color vision. JOSA A, 3(10):1651-1661, 1986. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 533, + 296, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 533, + 296, + 564 + ], + "spans": [ + { + "bbox": [ + 62, + 533, + 296, + 564 + ], + "type": "text", + "content": "[8] Gershon Buchsbaum. A spatial processor model for object colour perception. Journal of the Franklin institute, 310(1): 1-26, 1980. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 567, + 296, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 567, + 296, + 610 + ], + "spans": [ + { + "bbox": [ + 62, + 567, + 296, + 610 + ], + "type": "text", + "content": "[9] Ayan Chakrabarti, Keigo Hirakawa, and Todd Zickler. Color constancy with spatio-spectral statistics. IEEE Transactions on Pattern Analysis and Machine Intelligence, 34(8):1509-1519, 2011. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 613, + 295, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 295, + 656 + ], + "type": "text", + "content": "[10] Xiaochuan Chen, Mark S Drew, and Ze-Nian Li. Illumination and reflectance spectra separation of hyperspectral image data under multiple illumination conditions. *Electronic Imaging*, 29:194–199, 2017. 2, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 658, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 296, + 712 + ], + "type": "text", + "content": "[11] Yushi Chen, Hanlu Jiang, Chunyang Li, Xiuping Jia, and Pedram Ghamisi. Deep feature extraction and classification of hyperspectral images based on convolutional neural networks. IEEE transactions on geoscience and remote sensing, 54(10):6232-6251, 2016. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "type": "text", + "content": "[12] Arthur P Dempster, Nan M Laird, and Donald B Rubin. Maximum likelihood from incomplete data via the em algorithm. Journal of the royal statistical society: series B (methodological), 39(1):1-22, 1977. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 118, + 554, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 554, + 162 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 554, + 162 + ], + "type": "text", + "content": "[13] Arjan Gijsenij, Theo Gevers, and Joost Van De Weijer. Improving color constancy by photometric edge weighting. IEEE Transactions on Pattern Analysis and Machine Intelligence, 34(5):918-929, 2011. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 163, + 554, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 554, + 239 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 554, + 239 + ], + "type": "text", + "content": "[14] Ortal Glatt, Yotam Ater, Woo-Shik Kim, Shira Werman, Oded Berby, Yael Zini, Shay Zelinger, Sangyoon Lee, Heejin Choi, and Evgeny Soloveichik. Beyond rgb: A real world dataset for multispectral imaging in mobile devices. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4344-4354, 2024. 1, 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 240, + 554, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 240, + 554, + 283 + ], + "spans": [ + { + "bbox": [ + 316, + 240, + 554, + 283 + ], + "type": "text", + "content": "[15] Lin Gu, Cong Phuoc Huynh, and Antonio Robles-Kelly. Segmentation and estimation of spatially varying illumination. IEEE transactions on image processing, 23(8):3478-3489, 2014. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 285, + 554, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 285, + 554, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 285, + 554, + 330 + ], + "type": "text", + "content": "[16] Rui Guo, Wei Wang, and Hairong Qi. Hyperspectral image unmixing using autoencoder cascade. In 2015 7th Workshop on Hyperspectral Image and Signal Processing: Evolution in Remote Sensing (WHISPERS), pages 1-4. IEEE, 2015. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 331, + 554, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 554, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 554, + 374 + ], + "type": "text", + "content": "[17] Amina Ben Hamida, Alexandre Benoit, Patrick Lambert, and Chokri Ben Amar. 3-d deep learning approach for remote sensing image classification. IEEE Transactions on geoscience and remote sensing, 56(8):4420-4434, 2018. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 376, + 554, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 554, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 554, + 420 + ], + "type": "text", + "content": "[18] Siyuan Hao, Wei Wang, Yuanxin Ye, Tingyuan Nie, and Lorenzo Bruzzone. Two-stream deep architecture for hyperspectral image classification. IEEE Transactions on Geoscience and Remote Sensing, 56(4):2349-2361, 2017. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 421, + 555, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 421, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 421, + 555, + 464 + ], + "type": "text", + "content": "[19] Rob Heylen, Mario Parente, and Paul Gader. A review of nonlinear hyperspectral unmixing methods. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 7(6):1844-1868, 2014. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 466, + 554, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 466, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 466, + 554, + 510 + ], + "type": "text", + "content": "[20] Yuanming Hu, Baoyuan Wang, and Stephen Lin. Fc4: Fully convolutional color constancy with confidence-weighted pooling. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4085-4094, 2017. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 511, + 554, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 511, + 554, + 543 + ], + "spans": [ + { + "bbox": [ + 316, + 511, + 554, + 543 + ], + "type": "text", + "content": "[21] Anya Hurlbert. Challenges to color constancy in a contemporary light. Current Opinion in Behavioral Sciences, 30: 186-193, 2019. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 545, + 554, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 554, + 588 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 554, + 588 + ], + "type": "text", + "content": "[22] Marian-Daniel Iordache, José M Bioucas-Dias, and Antonio Plaza. Sparse unmixing of hyperspectral data. IEEE Transactions on Geoscience and Remote Sensing, 49(6):2014-2039, 2011. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 590, + 554, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 554, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 554, + 634 + ], + "type": "text", + "content": "[23] Qiwen Jin, Yong Ma, Erting Pan, Fan Fan, Jun Huang, Hao Li, Chenhong Sui, and Xiaoguang Mei. Hyperspectral unmixing with gaussian mixture model and spatial group sparsity. Remote Sensing, 11(20):2434, 2019. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 635, + 554, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 554, + 689 + ], + "type": "text", + "content": "[24] Murali Kanthi, T Hitendra Sarma, and C Shobha Bindu. A 3d-deep cnn based feature extraction and hyperspectral image classification. In 2020 IEEE India Geoscience and Remote Sensing Symposium (InGARSS), pages 229-232. IEEE, 2020. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 690, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 690, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 690, + 553, + 714 + ], + "type": "text", + "content": "[25] Nirmal Keshava and John F Mustard. Spectral unmixing. IEEE signal processing magazine, 19(1):44-57, 2002. 2, 3" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2223" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[26] Haris Ahmad Khan. Multispectral constancy for illuminant invariant representation of multispectral images. PhD thesis, Université Bourgogne Franche-Comté; Norwegian University of Science and ..., 2018. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 151 + ], + "type": "text", + "content": "[27] Haris Ahmad Khan, Jean-Baptiste Thomas, Jon Yngve Hardeberg, and Olivier Lalignant. Illuminant estimation in multispectral imaging. JOSA A, 34(7):1085-1098, 2017. 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 217 + ], + "type": "text", + "content": "[28] Haris Ahmad Khan, Jean-Baptiste Thomas, and Jon Yngve Hardeberg. Towards highlight based illuminant estimation in multispectral images. In Image and Signal Processing: 8th International Conference, ICISP 2018, Cherbourg, France, July 2-4, 2018, Proceedings 8, pages 517-525. Springer, 2018. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 294, + 262 + ], + "type": "text", + "content": "[29] Muhammad Jaleed Khan, Hamid Saeed Khan, Adeel Yousaf, Khurram Khurshid, and Asad Abbas. Modern trends in hyperspectral image analysis: A review. *IEEE Access*, 6:14118-14129, 2018. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 319 + ], + "type": "text", + "content": "[30] Dongyoung Kim, Jinwoo Kim, Junsang Yu, and Seon Joo Kim. Attentive illumination decomposition model for multi-illuminant white balancing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25512–25521, 2024. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 320, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 294, + 374 + ], + "type": "text", + "content": "[31] Furkan Klinl, Doga Yilmaz, Baris Özcan, and Furkan Kiraç. Deterministic neural illumination mapping for efficient autowhite balance correction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1139-1147, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 376, + 294, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 294, + 420 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 294, + 420 + ], + "type": "text", + "content": "[32] Fanqiang Kong, Tongbo Cao, Yunsong Li, Dan Li, and Kedi Hu. Multi-scale spatial-spectral attention network for multispectral image compression based on variational autoencoder. Signal Processing, 198:108589, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 422, + 294, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 442 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 442 + ], + "type": "text", + "content": "[33] Edwin H Land. The retina theory of color vision. Scientific american, 237(6):108-129, 1977. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 445, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 498 + ], + "type": "text", + "content": "[34] Edwin H Land. Recent advances in retina theory. In Central and Peripheral Mechanisms of Colour Vision: Proceedings of an International Symposium Held at The Wenner-Gren Center Stockholm, June 14–15, 1984, pages 5–17. Springer, 1985. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 501, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 294, + 521 + ], + "type": "text", + "content": "[35] David Landgrebe. Hyperspectral image data analysis. IEEE Signal processing magazine, 19(1):17-28, 2002. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 523, + 294, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 294, + 577 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 294, + 577 + ], + "type": "text", + "content": "[36] Bing Li, Haina Qin, Weihua Xiong, Yangxi Li, Songhe Feng, Weiming Hu, and Stephen Maybank. Ranking-based color constancy with limited training samples. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(10): 12304-12320, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 580, + 294, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 580, + 294, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 580, + 294, + 623 + ], + "type": "text", + "content": "[37] Rui Li, Shunyi Zheng, Chenxi Duan, Yang Yang, and Xiqi Wang. Classification of hyperspectral image based on double-branch dual-attention mechanism network. Remote Sensing, 12(3):582, 2020. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 624, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 667 + ], + "type": "text", + "content": "[38] Xian Li, Mingli Ding, and Aleksandra Pizurica. Deep feature fusion via two-stream convolutional neural network for hyperspectral image classification. IEEE Transactions on Geoscience and Remote Sensing, 58(4):2615-2629, 2019. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[39] Yuqi Li, Chong Wang, Jieyu Zhao, and Qingshu Yuan. Efficient spectral reconstruction using a trichromatic camera via sample optimization. The Visual Computer, 34:1773-1783, 2018. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[40] Yuqi Li, Qiang Fu, and Wolfgang Heidrich. Multispectral illumination estimation using deep unrolling network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2672-2681, 2021. 1, 2, 4, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 171 + ], + "type": "text", + "content": "[41] Shuaiqi Liu, Siyuan Liu, Shichong Zhang, Bing Li, Weiming Hu, and Yu-Dong Zhang. Ssau-net: A spectral-spatial attention-based u-net for hyperspectral image fusion. IEEE Transactions on Geoscience and Remote Sensing, 60:1-16, 2022. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 173, + 553, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 173, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 316, + 173, + 553, + 216 + ], + "type": "text", + "content": "[42] Zhenqi Liu, Xinyu Wang, Yanfei Zhong, Meng Shu, and Chen Sun. Siamhyper: Learning a hyperspectral object tracker from an rgb-based tracker. IEEE Transactions on Image Processing, 31:7116-7129, 2022. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 217, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 271 + ], + "type": "text", + "content": "[43] Yi-Chen Lo, Chia-Che Chang, Hsuan-Chao Chiu, Yu-Hao Huang, Chia-Ping Chen, Yu-Lin Chang, and Kevin Jou. Clcc: Contrastive learning for color constancy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8053-8063, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 273, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 273, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 273, + 553, + 304 + ], + "type": "text", + "content": "[44] Zhongyu Lou, Theo Gevers, Ninghang Hu, Marcel P Lucassen, et al. Color constancy by deep learning. In BMVC, pages 76-1, 2015. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 306, + 553, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 306, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 316, + 306, + 553, + 338 + ], + "type": "text", + "content": "[45] Qing Ma, Junjun Jiang, Xianming Liu, and Jiayi Ma. Reciprocal transformer for hyperspectral and multispectral image fusion. Information Fusion, 104:102148, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 339, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 339, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 339, + 553, + 381 + ], + "type": "text", + "content": "[46] Yong Ma, Qiwen Jin, Xiaoguang Mei, Xiaobing Dai, Fan Fan, Hao Li, and Jun Huang. Hyperspectral unmixing with gaussian mixture model and low-rank representation. Remote Sensing, 11(8):911, 2019. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 384, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 384, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 384, + 553, + 426 + ], + "type": "text", + "content": "[47] Laurence Maloney. Evaluation of linear models of surface spectral reflectance with small numbers of parameters. In Physics-based vision: Principles and practice. Jones and Bartlett, 1992. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 428, + 553, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 428, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 316, + 428, + 553, + 470 + ], + "type": "text", + "content": "[48] Laurence T Maloney and Brian A Wandell. Color constancy: a method for recovering surface spectral reflectance. In Readings in Computer Vision, pages 293-297. Elsevier, 1987. 1, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 472, + 553, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 472, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 316, + 472, + 553, + 623 + ], + "type": "text", + "content": "[49] Thomas Mansencal, Michael Mauderer, Michael Parsons, Nick Shaw, Kevin Wheatley, Sean Cooper, Jean D. Vandenberg, Luke Canavan, Katherine Crowson, Ofek Lev, Katrin Leinweber, Shriramana Sharma, Troy James Sobotka, Dominik Moritz, Matt Pppp, Chinmay Rane, Pavithra Eswaramoorthy, John Metic, Ben Pearlstine, Manuel Leonhardt, Olli Niemitalo, Marek Szymanski, Maximilian Schambach, Sianyi Huang, Mike Wei, Nishant Joywardhan, Omar Wagih, Pawel Redman, Joseph Goldstone, Stephen Hill, Jedediah Smith, Frederic Savoir, Geetansh Saxena, Saransh Chopra, Ilia Sibiryakov, Tim Gates, Gajendra Pal, Nicolas Tessore, Aurélien Pierre, François-Xavier Thomas, Sabarish Srinivasan, and Tucker Downs. Colour 0.4.2, 2022. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 625, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 668 + ], + "type": "text", + "content": "[50] Savas Ozkan, Berk Kaya, and Gozde Bozdagi Akar. Endnet: Sparse autoencoder network for endmember extraction and hyperspectral unmixing. IEEE Transactions on Geoscience and Remote Sensing, 57(1):482-496, 2018. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[51] Yishu Peng, Yuwen Zhang, Bing Tu, Qianming Li, and Wujing Li. Spatial-spectral transformer with cross-attention for hyperspectral image classification. IEEE Transactions on Geoscience and Remote Sensing, 60:1–15, 2022. 3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2224" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "text", + "content": "[52] Ying Qu and Hairong Qi. udas: An untied denoising autoencoder with sparsity for spectral unmixing. IEEE Transactions on Geoscience and Remote Sensing, 57(3):1698-1712, 2018. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 138 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 138 + ], + "type": "text", + "content": "[53] Douglas A Reynolds et al. Gaussian mixture models. Encyclopedia of biometrics, 741(659-663), 2009. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 194 + ], + "type": "text", + "content": "[54] Antonio Robles-Kelly and Ran Wei. A convolutional neural network for pixelwise illuminant recovery in colour and spectral images. In 2018 24th International Conference on Pattern Recognition (ICPR), pages 109-114. IEEE, 2018. 1, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 195, + 294, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 195, + 294, + 248 + ], + "spans": [ + { + "bbox": [ + 56, + 195, + 294, + 248 + ], + "type": "text", + "content": "[55] Swalpa Kumar Roy, Gopal Krishna, Shiv Ram Dubey, and Bidyut B Chaudhuri. Hybridsn: Exploring 3-d-2-d cnn feature hierarchy for hyperspectral image classification. IEEE Geoscience and Remote Sensing Letters, 17(2):277-281, 2019. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 250, + 294, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 294, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 294, + 304 + ], + "type": "text", + "content": "[56] Wu Shi, Chen Change Loy, and Xiaou Tang. Deep specialized network for illuminant estimation. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 371-387. Springer, 2016. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 304, + 294, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 304, + 294, + 348 + ], + "spans": [ + { + "bbox": [ + 56, + 304, + 294, + 348 + ], + "type": "text", + "content": "[57] Alberto Signoroni, Mattia Savardi, Annalisa Baronio, and Sergio Benini. Deep learning meets hyperspectral image analysis: A multidisciplinary review. Journal of imaging, 5(5):52, 2019. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 350, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 294, + 392 + ], + "type": "text", + "content": "[58] Tong Su, Yu Zhou, Yao Yu, Xun Cao, and Sidan Du. Illumination separation of non-lambertian scenes from a single hyperspectral image. Optics express, 26(20):26167-26178, 2018. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 393, + 294, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 393, + 294, + 446 + ], + "spans": [ + { + "bbox": [ + 56, + 393, + 294, + 446 + ], + "type": "text", + "content": "[59] Yuanchao Su, Jun Li, Antonio Plaza, Andrea Marinoni, Paolo Gamba, and Somdatta Chakravortty. Daen: Deep autoencoder networks for hyperspectral unmixing. IEEE Transactions on Geoscience and Remote Sensing, 57(7): 4309-4321, 2019. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 449, + 294, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 294, + 491 + ], + "type": "text", + "content": "[60] Hao Sun, Xiangtao Zheng, and Xiaogiang Lu. A supervised segmentation network for hyperspectral image classification. IEEE Transactions on Image Processing, 30:2810-2825, 2021. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 493, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 493, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 493, + 294, + 525 + ], + "type": "text", + "content": "[61] Jean-Baptiste Thomas. Illuminant estimation from uncalibrated multispectral images. In 2015 Colour and Visual Computing Symposium (CVCS), pages 1-6. IEEE, 2015. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 526, + 294, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 526, + 294, + 557 + ], + "spans": [ + { + "bbox": [ + 56, + 526, + 294, + 557 + ], + "type": "text", + "content": "[62] Joost Van De Weijer, Theo Gevers, and Arjan Gijsenij. Edge-based color constancy. IEEE Transactions on image processing, 16(9):2207-2214, 2007. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 559, + 294, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 294, + 601 + ], + "type": "text", + "content": "[63] Donik Vršnak, Ilija Domislovic, Marko Subasic, and Sven Lončarić. Framework for illumination estimation and segmentation in multi-illuminant scenes. IEEE access, 11: 2128-2137, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 603, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 294, + 635 + ], + "type": "text", + "content": "[64] Xia Xu and Zhenwei Shi. Multi-objective based spectral unmixing for hyperspectral images. ISPRS Journal of Photogrammetry and Remote Sensing, 124:54-69, 2017. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 637, + 294, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 637, + 294, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 637, + 294, + 679 + ], + "type": "text", + "content": "[65] Kai Yang, Hao Sun, Chunbo Zou, and Xiaoqiang Lu. Cross-attention spectral-spatial network for hyperspectral image classification. IEEE Transactions on Geoscience and Remote Sensing, 60:1-14, 2021. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "text", + "content": "[66] Fumihito Yasuma, Tomoo Mitsunaga, Daisuke Iso, and Shree K Nayar. Generalized assorted pixel camera: post-capture control of resolution, dynamic range, and spectrum." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 487 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "type": "text", + "content": "IEEE transactions on image processing, 19(9):2241-2253, 2010.6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "text", + "content": "[67] Huanglin Yu, Ke Chen, Kaiqi Wang, Yanlin Qian, Zhaoxiang Zhang, and Kui Jia. Cascading convolutional color constancy. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 12725-12732, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "type": "text", + "content": "[68] Jinkai Zhang, Benoit Rivard, and Arturo Sanchez-Azofeifa. Derivative spectral unmixing of hyperspectral data applied to mixtures of lichen and rock. IEEE Transactions on Geoscience and Remote Sensing, 42(9):1934-1940, 2004. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 186, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 229 + ], + "type": "text", + "content": "[69] Xiangrong Zhang, Yujia Sun, Jingyan Zhang, Peng Wu, and Licheng Jiao. Hyperspectral unmixing via deep convolutional neural networks. IEEE Geoscience and Remote Sensing Letters, 15(11):1755-1759, 2018. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 231, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 231, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 231, + 553, + 285 + ], + "type": "text", + "content": "[70] Chunhui Zhao, Boao Qin, Shou Feng, Wenxiang Zhu, Weiwei Sun, Wei Li, and Xiuping Jia. Hyperspectral image classification with multi-attention transformer and adaptive superpixel segmentation-based active learning. IEEE Transactions on Image Processing, 32:3606-3621, 2023. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 286, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 330 + ], + "type": "text", + "content": "[71] Runxing Zhao, Zhiwen Wang, Wuyuan Guo, and Canlong Zhang. Multi-scene image enhancement based on multi-channel illumination estimation. Expert Systems with Applications, 226:120271, 2023. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 331, + 553, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 385 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 385 + ], + "type": "text", + "content": "[72] Yinqiang Zheng, Imari Sato, and Yoichi Sato. Illumination and reflectance spectra separation of a hyperspectral image meets low-rank matrix factorization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1779-1787, 2015. 1, 2, 4, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 387, + 553, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 387, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 316, + 387, + 553, + 441 + ], + "type": "text", + "content": "[73] Zilong Zhong, Jonathan Li, Zhiming Luo, and Michael Chapman. Spectral-spatial residual network for hyperspectral image classification: A 3-d deep learning framework. IEEE Transactions on Geoscience and Remote Sensing, 56 (2):847-858, 2017. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 443, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 443, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 443, + 553, + 487 + ], + "type": "text", + "content": "[74] Yuan Zhou, Anand Rangarajan, and Paul D Gader. A gaussian mixture model representation of endmember variability in hyperspectral unmixing. IEEE Transactions on Image Processing, 27(5):2242-2256, 2018. 2, 3" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2225" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/ImViD_ Immersive Volumetric Videos for Enhanced VR Engagement/6da376dc-6436-436c-b26c-23ba201328ef_content_list.json b/2025/ImViD_ Immersive Volumetric Videos for Enhanced VR Engagement/6da376dc-6436-436c-b26c-23ba201328ef_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ee3f69a2551f5e1a013296ff82c4999e5df00168 --- /dev/null +++ b/2025/ImViD_ Immersive Volumetric Videos for Enhanced VR Engagement/6da376dc-6436-436c-b26c-23ba201328ef_content_list.json @@ -0,0 +1,1627 @@ +[ + { + "type": "text", + "text": "ImViD: Immersive Volumetric Videos for Enhanced VR Engagement", + "text_level": 1, + "bbox": [ + 151, + 130, + 846, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhengxian Yang $^{1*}$ , Shi Pan $^{1*}$ , Shengqi Wang $^{1*}$ , Haoxiang Wang $^{1}$ , Li Lin $^{2}$ , Guanjun Li $^{3}$ , Zhengqi Wen $^{1\\dagger}$ , Borong Lin $^{1\\dagger}$ , Jianhua Tao $^{1\\dagger}$ , Tao Yu $^{1\\dagger}$ , Tsinghua University ${}^{2}$ Migu Beijing Research Institute ${}^{3}$ Institute of Automation, Chinese Academy of Science $\\{\\mathrm{zx - yang23, ps23, shengqi - 21, whx22}\\} @$ mails.tsinghua.edu.cn lilin@migu.chinamobile.com guanjun.li@nlpr ia.ac.cn $\\{\\mathrm{zqwen, linbr, jhtao, ytrock}\\} @$ tsinghua.edu.cn", + "bbox": [ + 117, + 179, + 880, + 287 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/29df6dfe8c358de09148a8c338fa2c40e7df876d550543ebbb5b53dcecaae38c.jpg", + "image_caption": [ + "Figure 1. We introduce ImViD, a dataset for immersive volumetric videos. ImViD records dynamic scenes using a multi-view audio-video capture rig moving in a space-oriented manner, which provides a new benchmark for volumetric video reconstruction and its application." + ], + "image_footnote": [], + "bbox": [ + 93, + 309, + 305, + 481 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/54cea3e825ce7d714e26bf4ae2061d89768861e2c9285a1234747496b4cf9a61.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 305, + 309, + 691, + 479 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7620de647f0ca8575e9ca85febef46564a18a61207b78bd20144f82cf7d2d859.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 692, + 309, + 903, + 479 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 539, + 326, + 555 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "User engagement is greatly enhanced by fully immersive multi-modal experiences that combine visual and auditory stimuli. Consequently, the next frontier in VR/AR technologies lies in immersive volumetric videos with complete scene capture, large 6-DoF interaction space, multimodal feedback, and high resolution&frame-rate contents. To stimulate the reconstruction of immersive volumetric videos, we introduce ImViD, a multi-view, multi-modal dataset featuring complete space-oriented data capture and various indoor/outdoor scenarios. Our capture rig supports multi-view video-audio capture while on the move, a capability absent in existing datasets, significantly enhancing the completeness, flexibility, and efficiency of data capture.", + "bbox": [ + 88, + 571, + 482, + 767 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The captured multi-view videos (with synchronized audi- 15 minutes, and include rich foreground-background elements, and complex dynamics. We benchmark existing methods using our dataset and establish a base pipeline for construct", + "bbox": [ + 88, + 768, + 482, + 843 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing immersive volumetric videos from multi-view audiovisual inputs for 6-DoF multi-modal immersive VR experiences. The benchmark and the reconstruction and interaction results demonstrate the effectiveness of our dataset and baseline method, which we believe will stimulate future research on immersive volumetric video production. Project Page: https://yzxqh.github.io/ImViD/", + "bbox": [ + 511, + 541, + 906, + 647 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 662, + 643, + 679 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "High-quality and high-fidelity digital modeling of the real world has long been a topic of significant interest. With breakthroughs in 3D reconstruction algorithms and virtual reality technology, the generation of such visually compelling content is gradually becoming feasible, whether for reproducing past events and memories or remotely viewing current occurrences. Achieving complete user immersion is a goal pursued by many researchers.", + "bbox": [ + 509, + 688, + 906, + 809 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "What defines an immersive volumetric video?", + "bbox": [ + 532, + 810, + 834, + 824 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dwelling on previous tests [1, 6] and user feedback [2, 42], we believe that immersive media should have the following four characteristics: i) Full $360^{\\circ}$ foreground and background, ii) High-quality 6-DoF interaction, which current volumetric video can achieve, iii) Multimodal experi", + "bbox": [ + 511, + 824, + 905, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contributions. † Corresponding author. \nZhengxian Yang, Shengqi Wang, Haoxiang Wang, Jianhua Tao and Tao Yu are at BNRist&Department of Automation; Shi Pan and Borong Lin are at Department of Building Science; Zhengqi Wen is at BNRist.", + "bbox": [ + 89, + 849, + 482, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16554", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ence with light and sound, and iv) High frame rate and long-duration dynamic content. We refer to a media format that meets all these criteria as immersive volumetric video.", + "bbox": [ + 89, + 90, + 480, + 135 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, with the development of camera equipment, despite many efforts using monocular handheld cameras or fixed camera arrays to build datasets, none of these solutions fully address all four requirements. For example, Panoptic Sports [19], ZJU-Mocap [44] and DiVa-360 [37] achieved human/object-centric $360^{\\circ}$ capture but lack complex backgrounds and multimodal data; Google's immersive light field [6] achieves 6-DoF volumetric video in an inside-looking-out manner but lacks multimodality and high frame rate, long-duration dynamic content; Replay [49] introduces sound, but the camera setup is not suitable for human viewing habits thus cannot achieve high-quality 6-DoF interaction. Both academia and industry require an appropriate capture method to support the development of immersive volumetric video technology.", + "bbox": [ + 88, + 137, + 482, + 364 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Additionally, NeRF [41]'s emergence has increased attention to reconstruction and rendering techniques, making it easier to generate high-fidelity interactive 3D scenes from multiple views. 3DGS [23] and its subsequent improvements [10, 29, 38, 40, 57, 59, 61] have further boosted rendering speed and quality, fueling peoples' expectations for rapid real-world modeling and its applications in industries such as education, healthcare, and entertainment. This can be seen as an evolutionary opportunity for immersive volumetric video. However, current technologies face challenges. In dynamic light field reconstruction, neither explicit [22, 24, 30, 35, 55] nor implicit neural representation based [4, 8, 14, 28, 48, 51, 53] methods have been able to achieve a practical balance between foreground-background reconstruction accuracy, rendering speed, spatiotemporal consistency, and storage efficiency in dynamic scenes; In the realm of sound field reconstruction, it is typically necessary to collect a large amount of binaural audio or room impulse response data from different locations for model training, which is user-unfriendly in practical use.", + "bbox": [ + 91, + 364, + 482, + 667 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To better support the development of immersive volumetric video technology, we have developed a moving capture rig (with 46 synchronized cameras, as shown in Fig.3) and designed two space-oriented data capture strategies. We have constructed an audiovisual dataset containing 7 indoor and outdoor scenes, with a $360^{\\circ}$ field of view, large 6-DoF interaction space, and high-resolution, long-duration dynamic contents. We also validate this dataset using the latest novel view synthesis and 3D reconstruction methods. Finally, a complete pipeline for constructing light and sound fields from this type of data was proposed, which initially achieved immersive volume video.", + "bbox": [ + 89, + 667, + 482, + 848 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions can be summarized as follows:", + "bbox": [ + 109, + 849, + 470, + 864 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We propose to our knowledge the first and probably the highest-quality multi-view dataset ImViD for immersive", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "volumetric video. It features $360^{\\circ}$ foreground and background capture, long-duration video&audio data under high resolution@5K, high frame rate@60FPS.", + "bbox": [ + 526, + 90, + 903, + 136 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We conduct extensive experiments to validate the effectiveness of the proposed datasets by evaluating and analyzing the performance of several current baselines.", + "- Follow the proposed pipeline shown in Figure 2, We finally achieve the reconstruction of immersive volumetric video and integrate it for immersive VR experience." + ], + "bbox": [ + 513, + 136, + 903, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 242, + 653, + 258 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Datasets for Dynamic Novel View Synthesis. The datasets for dynamic scene reconstruction can be classified as monocular-based and multi-view-based.", + "bbox": [ + 511, + 268, + 903, + 313 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Monocular acquisition systems are popular due to their low cost and ease of construction. such as HyperNeRF [43], Dynamic Scene Dataset [60], and D2NeRF [56]. However, these datasets suffer from low resolution, limited capture space, and the limited durations. Although NeRF On-the-go [45] allows for larger capture ranges by walking while shooting, high-quality reconstructions are confined to the capture path, and the small field of view (FOV) limits prolonged observations of specific scene positions.", + "bbox": [ + 511, + 315, + 903, + 450 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-camera data collection has gained significant attention due to its ability to provide a larger FOV and richer details, such as Immersive Light Field dataset [6], Technicolor [46], UCSD Dynamic Scene Dataset [34] and The Plenoptic Dataset [28], followed by [27, 33, 54]. However, all these setups remain static during capture, limiting them to frontal views and hindering $360^{\\circ}$ reconstruction. Additionally, the video sequences are typically short, with a maximum duration of 2 minutes (often less than 30 seconds) and a maximum resolution of 4K, which is insufficient for immersive VR experiences.", + "bbox": [ + 511, + 452, + 905, + 618 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Moreover, the previously mentioned datasets lack sound recordings, leading to the loss of immersion without multimodality. The Replay dataset [49] addresses this by focusing on long sequences with professional actors in familiar settings. However, aside from the head-mounted cameras that enables rotation with head, all other cameras remain static. Furthermore, the DSLR arrangement does not align with human viewing habits in VR, making them unsuitable as benchmarks for novel view synthesis tasks. Recent work [9] presents a dataset with a $360^{\\circ}$ camera and audio capture device but constrained by a fixed-point shooting strategy, resulting in sparse viewpoints that hinder the reconstruction of high-quality dynamic scenes. More details about datasets can be found in Tab. 1.", + "bbox": [ + 511, + 619, + 905, + 830 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Dynamic Light&Sound Field Reconstruction. Temporal variations in input images make reconstructing dynamic scenes more challenging than static ones. The key issue is efficiently representing 4D scenes while ensuring spa", + "bbox": [ + 511, + 839, + 903, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16555", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/3ce26600105276a6e8d0df49247fddf8a82e5c664d4a8c754c8c914b854f33d0.jpg", + "table_caption": [ + "Table 1. Existing real-world datasets for dynamic novel view synthesis (full version in the supplementary materials)." + ], + "table_footnote": [], + "table_body": "
DatasetsNo SceneOutdoor/IndoorCamerasMobilityResolutionAnglesDurationFPSMultimodalityContent
PanopticSports [19]65Indoor480 camerasStatic640×480360°5mins25XHuman-centric actions
Technicolor [46]12Indoor16 camerasStatic2048×1088Frontal2s30XHas a number of close-ups sequences, captured medium angle scenes and other animated scenes where the movement does not come from a human
Immersive-Lightfield [6]15both46 camerasStatic2560×1920Frontal10-30s30XSimple and slow motion of human, animals, objects
Dynamic Scene Datasets (NVIDIA) [60]8Outdoor1 Mobile phone /12 camerasFixed-point Waving /Static1920×1080Frontal5s60XSimple body motions (facial, jump, etc)
UCSD Dynamic [34]96Outdoor10 camerasStatic1920×1080Frontal1-2mins120XVarious visual effects and human interactions
ZJU-Mocap [44]10Indoor21 camerasStatic1024×1024360°20s50XSimple body motions (punch, kick, etc.)
Plenoptic Dataset (DyN-eRF/Neural 3D) [28]6Indoor21 camerasStatic2704×2028Frontal10-30s30XContains high specularity, translucency and transparency objects, motions with changing topology, selfcast moving shadows, various lighting conditions and multiple people moving around in open living room space
iPhone Datasets [15]14both1 hand-held phone /2 camerasFixed-point Waving /Static640×480Frontal8-15s30/60XFeaturing non-repetitive motion, from various categories such as generic objects, humans, and pets
ENeRF-Outdoor [33]4Outdoor18 camerasStatic1920×1080Frontal20-40s30XComplex human motions
Replay [49]46Indoor12 camerasStatic3840×2160360°5mins30✓(Audio)Dancing, chatting, playing video games, unwrapping presents, playing ping pong
360+X [9]28both1 360°cameras and 1 Spectacles camerasStatic5760×2880360°10s (2152 sequence)30✓(Audio)Capture in 17 cities across 5 countries. Panoptic perspective to scene understanding with audio
Im-ViD(Ours)7both46 camerasMoveable5312×2988Frontal and 360°1-5mins60✓(Audio)Seven common indoor and outdoor scenes in daily life, including opera, face-to-face communication, teaching, discussion, music performance, interaction with pets, and playing. Each scene has high-quality synchronized multi-view video and audio with a duration of more than 1 minute, and contains rich elements such as various small objects, glass, and changes in light and shadow
", + "bbox": [ + 93, + 114, + 906, + 426 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tiotemporal consistency and high accuracy. Traditional methods estimate varying geometries like surfaces [11] and depth maps [21] to handle dynamics. Recently, efforts have shifted towards adapting neural radiance field representations for dynamic scenes. DyNeRF [28] pioneer in the dynamic reconstruction with Neural radiance field. Followed works such as [4, 8, 14, 27, 48, 51, 53] make efforts to improve the high-fidelity and efficiency in the 4D neural representation. However, the common challenges these methods face are memory cost and the modeling of spatiotemporal complexity. The latest work have extended 3DGS to dynamic scenes for high-quality real-time rendering. While modeling dynamic scenes as static ones is straightforward, methods relying solely on monocular data [20, 32] is not suitable for this paradigm, and face challenges of adding constraints or priors to better supervise training. Other works support multi-view data as input. [39] and [52] initialize each frame from the previous one to build Gaussian point clouds per-frame, leading to high memory and suddenly appearing or disappearing artifacts. 4DGS [55] introduces Hexplane to encode spatiotemporal information, using MLPs to predict Gaussian properties change. Subsequent works [13, 17, 26] further add geometric and optical flow constraints to improve frame-to-frame motion but perform well only on short sequences with small actions. Other approaches [22, 24, 30, 35] employ the explicit parametric form (such as radial basis functions) of 3DGS to enhance spatiotemporal consistency. Some methods [12, 58] focus on representation with dual quaternions, 4D spheri", + "bbox": [ + 89, + 450, + 485, + 891 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "cal harmonics, or rotors for end-to-end training. More improvements such as parallel training [50] and using codebooks and masks to enhance scene compactness [25], show the potential in efficient and high-quality representation of long-term complex dynamic scenes.", + "bbox": [ + 511, + 450, + 906, + 527 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In addition to light field reconstruction, reconstructing the sound field from multimodal data is equally important, often referred to as novel-view acoustic synthesis. For the novel-view acoustic synthesis task, AV-NeRF [31] synthesizes spatial audio by leveraging visual features from rendered images. NeRAF [7] further decouples the camera and microphone poses used for training. AV-GS [5] incorporates a 3D scene representation during sound field synthesis, addressing the issue where previous methods overlooked the contribution of the broader 3D scene geometry. SOAF [16] considers occlusion in the light field when synthesizing the sound field. However, the above novel-view acoustic synthesis tasks typically require a large amount of multimodal training data collected from different locations, which is often user-unfriendly in practical applications. Moreover, these methods generally cannot solve the problem of sound field synthesis for moving sound sources. This paper proposes a training-free novel-view acoustic synthesis approach for our dataset, which also supports sound field synthesis for moving sound sources.", + "bbox": [ + 511, + 527, + 908, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. ImViD Dataset", + "text_level": 1, + "bbox": [ + 511, + 844, + 663, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our goal is to build an immersive volumetric video dataset of real world that integrates both foreground and back-", + "bbox": [ + 511, + 869, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16556", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/fc1cdf68e993250542181382f874bc20780f5c2496c601e9da23071faafea384.jpg", + "image_caption": [ + "Figure 2. The pipeline to realize the multimodal 6-DoF immersive VR experiences. We applied a carefully designed rig to (a) simultaneously capture multi-view video and audio. The (b1) presents our reconstruction of dynamic light field based on STG [30] while (b2) demonstrates the construction process of sound field. We have achieved better results than the original algorithm in long-term dynamic scenes by incorporating affine color transformation and t-dimensional density control. Ultimately, we achieve a 6-DoF immersive experience in both light and sound fields, and also benchmark on recent representative methods like 4DGS [55] and 4Drotor [12] to demonstrate the effectiveness of both our dataset and baseline method." + ], + "image_footnote": [], + "bbox": [ + 122, + 88, + 877, + 417 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ground, providing assistance for the development of related research such as spatial video and VR/AR application. Therefore, we have meticulously refined every aspect of the process, from equipment setup and scene selection to shooting strategies and data processing, ensuring maintaining the elements of interest to researchers to the greatest extent. The released ImViD dataset comprises a total of 7 large scenes, including 5 indoor and 2 outdoor environments. The content features a diverse range of common activities, such as opera, meeting, teaching, and playing on the grass. Each scene contains a set of static photos, multi-view time-synchronized video&audio sequences at 5K resolution and 60 FPS, employing two different collection strategies, along with calibrated camera poses.", + "bbox": [ + 88, + 539, + 485, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Data Acquisition", + "text_level": 1, + "bbox": [ + 89, + 757, + 256, + 773 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The handheld monocular camera is easy to move, providing limited perspectives from various locations. In contrast, the fixed camera arrays, while stationary, offers dense perspectives within a limited range. We aim to combine the advantages of both to design an effective data collection system and strategy for fully immersive VR experience—a feature not found in existing datasets. As shown in Figure 3, our capture system consists of more than 40 GoPro cameras in", + "bbox": [ + 89, + 780, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "stalled on a hemispherical surface mounted on a remotely controllable mobile car, all of which are connected to a PC to receive synchronous control commands. The system is carefully designed to be at a height similar to that of a human, and the camera arrangement better aligns with human viewing habits compared to other existing devices. This rig can synchronously collect about 1,000 images at 5K resolution within a space of at least $6\\mathrm{m}^3$ in 2 min. It is also capable of synchronously capturing video and audio at 60 FPS and 5K resolution for 30 min (limited by heat dissipation). Thus, our capture strategy is as follows:", + "bbox": [ + 511, + 539, + 906, + 705 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Step1: Efficient high-quality, high-density $360^{\\circ}$ image acquisition of static environments.", + "bbox": [ + 511, + 707, + 906, + 738 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Step2: Continuous synchronous collection of dynamic scenes under two different strategies.", + "bbox": [ + 511, + 739, + 903, + 771 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "i) Fixed-point shooting. The cart remains stationary during the capture, focusing on capturing dynamic details.", + "ii) Moving shooting. The cart slowly moves while the scene is in motion, providing additional level of details and larger exploratory space." + ], + "bbox": [ + 511, + 773, + 903, + 852 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Throughout the capture process, we ensure strict synchronization of the world time codes across all cameras. We carefully maintained consistent focal lengths, exposure,", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16557", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/93ee5eea297500e1b2858df8345906a2384b52ee47b14b1c1729d23290d6a3b0.jpg", + "image_caption": [ + "Figure 3. Our rig support two kinds of capturing strategies for high resolution, high frame rate and $360^{\\circ}$ dynamic data acquisition." + ], + "image_footnote": [], + "bbox": [ + 101, + 90, + 289, + 272 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/787b9f96ad10d46945cdaa06015c5884217d06c098fcd13a5f97cef15e89b9eb.jpg", + "image_caption": [ + "Figure 4. Calculation method for spatiotemporal capture density. 1) the capture strategy of handheld monocular camera 2) represents the fixed camera array 3) Our rig covers a volume of $0.6 \\times \\pi \\times 0.5^2 \\, \\mathrm{m}^3$ over 5 seconds." + ], + "image_footnote": [], + "bbox": [ + 295, + 90, + 467, + 272 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and white balance settings for each scene, based on the results of manual iterative testing of camera settings, thereby minimizing hardware-induced variations in data quality. To achieve high-quality audio, we increased the noise reduction coefficient of the cameras and minimized noise generation as much as possible.", + "bbox": [ + 89, + 340, + 483, + 434 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Data Processing", + "text_level": 1, + "bbox": [ + 89, + 440, + 251, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the photos from Step1, we completed the calibration in COLMAP [47] by combining the cameras' prior parameters and their relative positions. For the video&audio from Step2, we first aligned the multi-view sequences for each scene using the world time codes. This step is crucial as the quality of the alignment directly determines the quality of dynamic scene reconstruction and significantly impacts whether the multi-view audio can be used for constructing a 6-DoF sound field. Our time code alignment error is at the level of approximately 2 milliseconds. Subsequently, we performed frame extraction from the videos, again utilizing the GoPro camera parameters as priors to complete the calibration.", + "bbox": [ + 88, + 462, + 482, + 656 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "It is important to note that for the dynamic scene data captured with a moving rig, we also performed time code alignment. However, this data poses significant challenges for existing calibration methods, often resulting in errors and floaters. There are also no solutions for reconstruction algorithms specifically addressing dynamic scenes from moving rig currently. Therefore, we do not provide the calibrated results for this type of data at this time. But we believe that this data will greatly contribute to the advancement of the field, and thus we will also publicly releasing the data which has been time-synchronized.", + "bbox": [ + 89, + 659, + 482, + 825 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Dataset Analysis", + "text_level": 1, + "bbox": [ + 89, + 833, + 256, + 849 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Diversity. Our dataset, as depicted in Table 2, includes 7 common indoor and outdoor scenes, each comprising one or more takes, totaling 16 video sequences. Each scene is", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/db07f5f4f664c27676e09bbac34031a49fb20ca9a54d32481c233c0850fb1735.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 526, + 90, + 898, + 256 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "rich in foreground and background elements, meticulously crafted to preserve the authenticity of the environment. The characters, clothing, and motions within each scene exhibit a diverse range, set against varying ambient lighting conditions. We also included areas of interest like tables, chairs, windows as much as possible for researchers.", + "bbox": [ + 511, + 354, + 906, + 445 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Quality. Leveraging a total of 39 cameras, we captured dynamic scenes at a resolution of $5312 \\times 2988$ and 60 FPS. The images of static environment is $5568 \\times 4176$ .", + "bbox": [ + 511, + 446, + 906, + 491 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Duration. As outlined in Table 2, each scene in our dataset contains at least 1 minute of footage. The longest sequence in our forthcoming public dataset spans approximately 5 minutes. The total processed data duration amounts to 38 minutes 46 seconds, including 139,560 frames per camera.", + "bbox": [ + 511, + 493, + 905, + 568 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Coverage. As highlighted in Table 2, we conducted image acquisition from over 1,000 viewpoints before recording dynamic content, thereby offering a $180^{\\circ} - 360^{\\circ}$ viewing space in VR. For scenes employing the mobile shooting strategy, we introduced a metric termed 'spatiotemporal capture density,' quantifying the captured volume covered by our rig per second, as shown in Figure 4. This metric, unavailable in current datasets (where static camera arrays are deemed 0 and handheld monocular cameras follow a single trajectory without capturing volume), registers at approximately $0.10~m^3 /s$ in our dataset.", + "bbox": [ + 511, + 569, + 906, + 736 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Benchmarks and Experiments", + "text_level": 1, + "bbox": [ + 511, + 753, + 792, + 771 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To comprehensively assess the effectiveness of the dataset, we selected three scenes (2 indoor and 1 outdoor) to test the latest dynamic light field reconstruction algorithms, providing a benchmark for research in areas such as novel view synthesis. We also proposed an improvement strategy based on one of the methods. Furthermore, we presented and validated a feasible pipeline for constructing 6-DoF sound fields from multi-view audio sequences. In this section, we", + "bbox": [ + 511, + 780, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16558", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/13c7d7d3108da5e76b322c474f06734c8e86d96140d1e0ec46f93bb6fa5ef87e.jpg", + "table_caption": [ + "Table 2. The detailed statistics of our ImViD dataset. #Take represents the number of different episodes we collected in this scene. #Strategy represents the strategies used in Step2. 1 indicates only (i) was used in Step 2, 2 indicates that both were used. Avg.S-T Density indicates the spatiotemporal coverage density of our collection method, measured by the captured volume size per second." + ], + "table_footnote": [], + "table_body": "
Name#Camera#Static viewpoints#Take#StrategyAvg.S-T Density(m3/s)Viewing SpaceAvg.DurationStorage(GB)
Scene1 Opera39115221-180°3min22s226
Scene2 Laboratory391225220.10360°1min42s137.3
Scene3 Classroom391223220.10360°4min42s497
Scene4 Meeting39122311-360°3min16s114
Scene5 Rendition391620420.10360°2min02s516
Scene6 Puppy391404320.10360°1min50s359
Scene7 Playing391224220.10360°1min10s220
Total--16---38min46s2069.3
", + "bbox": [ + 91, + 141, + 903, + 267 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "will first introduce the selected light field pipeline and our proposed sound field pipeline (Sec 4.1), then provide details on the experimental setup (Sec 4.2).", + "bbox": [ + 89, + 291, + 483, + 338 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Baseline", + "text_level": 1, + "bbox": [ + 89, + 345, + 192, + 359 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dynamic Light Field Reconstruction. We conducted a thorough survey of the current dynamic reconstruction algorithms, prioritizing rendering speed and support for multiview data input. We finally decided to establish a benchmark on 3DGS-based method from three paradigms.", + "bbox": [ + 89, + 367, + 483, + 444 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4DGS [55] introduces Hexplane to encode spatiotemporal information to feature $f_{d}$ and use MLPs decoders $D = \\{\\phi_{x},\\phi_{r},\\phi_{s}\\}$ to predict gaussians' changes in position $\\Delta \\chi = \\phi_{x}(f_{d})$ , rotation $\\Delta r = \\phi_r(f_d)$ , and scaling $\\Delta s = \\phi_s(f_d)$ . Utilizing the sparse point cloud from frame 0 as initialization and training static gaussians for the first 3000 iterations, the dynamic scene $G$ is represented as:", + "bbox": [ + 89, + 443, + 483, + 549 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nG \\left(\\chi^ {\\prime}, r ^ {\\prime}, s ^ {\\prime}\\right) = G \\left(\\chi + \\Delta \\chi , r + \\Delta r, s + \\Delta s\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 559, + 480, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4DrotorGS [12] is a representative method of dimensionality expansion. It extends 3DGS to a four-dimensional Gaussian ellipsoid by introducing the concept of rotor, expanding the covariance matrix to four dimensions, i.e.:", + "bbox": [ + 89, + 585, + 483, + 647 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nG _ {4 D} (x) = e ^ {- \\frac {1}{2} (x - \\mu_ {4 D}) ^ {T} \\Sigma_ {4 D} ^ {- 1} (x - \\mu_ {4 D})} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 656, + 480, + 676 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Each frame of the scene corresponds to a time slice of the 4D Gaussian ellipsoids followed by $\\alpha$ -blending.", + "bbox": [ + 89, + 686, + 482, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "STG [30], on the other hand, use polynomial fitting for the motion and rotation of the 3DGS, and employs an RBF function to model the opacity changes over time. It concatenates the sparse point clouds from all frames as the initialization, and the final representation of each point in the scene is:", + "bbox": [ + 89, + 715, + 483, + 806 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {i} (t) = \\sigma_ {i} (t) \\exp \\left(- \\frac {1}{2} \\left(x - \\mu_ {i} (t)\\right) ^ {T} \\Sigma_ {i} (t) ^ {- 1} \\left(x - \\mu_ {i} (t)\\right)\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 98, + 814, + 480, + 854 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Although we ensured consistent white balance and other settings across cameras based on the characteristics of scenes during shooting, color differences still exist due to", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "factors like occlusions when capturing from different angles. During experiments, we found that while existing methods fit well to the training views, there is a noticeable color shift during viewpoint transitions in the reconstructed light field, manifesting as flickering and floaters. This is especially apparent when training in segments, where the differences between segments are more pronounced, significantly reducing immersion. We improved the original STG (named as $\\mathrm{STG}++$ ) by introducing global, time-invariant color mapping during training. Detailed implementation can be seen in supplementary materials.", + "bbox": [ + 511, + 291, + 906, + 458 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Sound Field Reconstruction. Taking the recording microphone as the origin of the coordinate system, we denote the coordinates of the sound source as $s(t) = \\{x_s(t), y_s(t)\\}$ , the user wearing the VR device as $l(t) = \\{x_l(t), y_l(t)\\}$ , and the angle of deviation from the y-axis in the counterclockwise direction as $\\theta_l(t)$ , where $t$ is the time index. Based on these premises, generating the user's binaural spatial audio can be divided into three parts: sound direction mapping, sound distance mapping, and sound auralization. Note that we assume there is only one dominating sound source (omnidirectional emitting) to simplify the sound field reconstruction.", + "bbox": [ + 511, + 462, + 906, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the sound direction mapping part, we calculate the direction of the sound source relative to the user's ears, $\\theta_{s}(t)$", + "bbox": [ + 511, + 643, + 905, + 675 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {s} (t) = \\operatorname {a r c c o s} \\frac {\\mathbf {v} _ {1} (t) \\mathbf {v} _ {2} (t)}{| \\mathbf {v} _ {1} (t) | | \\mathbf {v} _ {2} (t) |}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 606, + 685, + 903, + 718 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where", + "bbox": [ + 514, + 720, + 558, + 734 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} _ {1} (t) = \\left[ x _ {s} (t) - x _ {l} (t), y _ {s} (t) - y _ {l} (t) \\right] ^ {T}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 744, + 903, + 762 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} _ {2} (t) = \\left[ - \\sin \\left(\\theta_ {l} (t)\\right), \\cos \\left(\\theta_ {l} (t)\\right) \\right] ^ {T}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 763, + 903, + 782 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the sound distance mapping part, we calculate the scaling of the sound source when it reaches the user's ears, relative to the original recording. We denote this scaling parameter as $\\lambda(t)$ ,", + "bbox": [ + 511, + 791, + 905, + 852 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\lambda (t) = \\frac {\\sqrt {x _ {s} ^ {2} (t) + y _ {s} ^ {2} (t)}}{\\sqrt {(x _ {s} (t) - x _ {l} (t)) ^ {2} + (y _ {s} (t) - y _ {l} (t)) ^ {2}}} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 859, + 903, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16559", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the sound auralization part, we convert the original recording into spatial audio based on $\\theta_s(t)$ and $\\lambda (t)$ . In the short-time Fourier transformation (STFT) domain, we denote the audio from the left and right ears as $A_{L}(t,f)$ and $A_{R}(t,f)$ , respectively, where $f$ is the frequency index. Then $A_{L}(t,f)$ and $A_{R}(t,f)$ can be written as", + "bbox": [ + 89, + 90, + 483, + 183 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nA _ {L} (t, f) = F _ {L} \\left(\\theta_ {s} (t)\\right) F _ {l} (t, f) A _ {O} (t, f) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 194, + 482, + 210 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nA _ {R} (t, f) = F _ {R} \\left(\\theta_ {s} (t)\\right) F _ {l} (t, f) A _ {O} (t, f), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 212, + 482, + 229 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $F_{L}(\\theta_{s}(t))$ and $F_{R}(\\theta_{s}(t))$ are head-related transfer functions (HRTF) based on $\\theta_s(t)$ , $F_{l}(t)$ denotes the room impulse response (RIR) at the user's location and $A_{O}(t,f)$ is the sound emitted from the speaker in the STFT domain. We utilize $\\theta_{s}(t)$ to retrieve $F_{L}(\\theta_{s}(t))$ and $F_{R}(\\theta_{s}(t))$ from the SADIE II dataset [3]. In order to model $F_{l}(t)$ and $A_{O}(t,f)$ , we make the following assumptions: $F_{l}(t)$ change between different locations at the same time index $t$ only depending on the parameter $\\lambda (t)$ , i.e.,", + "bbox": [ + 89, + 239, + 483, + 378 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nF _ {l} (t, f) = \\lambda (t) F (t, f), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 388, + 482, + 406 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $F(t)$ is the RIR at the recording microphone. Using the above assumption, Eq.(8) and Eq.(9) can be rewritten as", + "bbox": [ + 89, + 416, + 483, + 446 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nA _ {L} (t, f) = \\lambda (t) F _ {L} \\left(\\theta_ {s} (t)\\right) A (t, f), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 474, + 482, + 491 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nA _ {R} (t, f) = \\lambda (t) F _ {R} \\left(\\theta_ {s} (t)\\right) A (t, f), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 493, + 482, + 510 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $A(t,f) = F(t,f)A_O(t,f)$ is the audio at the recording microphone.", + "bbox": [ + 89, + 521, + 483, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Experimental Settings", + "text_level": 1, + "bbox": [ + 89, + 561, + 297, + 577 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Training setups. The training for all light field data was completed on a single NVIDIA A100 GPU. Due to memory limitations, we split the video and trained a model every 60 frames, total 300 frames. We also followed the original train-test-split strategy of the baseline for our experiments.", + "bbox": [ + 89, + 583, + 483, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Metric. We analyzed the quality of dynamic light field reconstruction using the most common evaluation metrics PSNR, SSIM, and LPIPS(alex).", + "bbox": [ + 89, + 659, + 483, + 704 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Results and Analysis", + "text_level": 1, + "bbox": [ + 89, + 718, + 289, + 734 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Validation of Light Field Data", + "text_level": 1, + "bbox": [ + 89, + 742, + 356, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Results. As illustrated in Table 3, 4 and Figure 5, our indoor and outdoor data both yield correct results across three 3DGS-based paradigms. STG++ outperforms all baselines in all scenes, delivering slightly better results, the highest rendering speed, and a moderate model size. We also found that, since 4DGS trains using the sparse point cloud from the 0th frame and limits the number of final points, its performance is significantly lower than other pipelines on texture-richer outdoor scenes, such as Scene6", + "bbox": [ + 89, + 763, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e4ee11162c2bc5a61527694bed825afd2cbfbb2f99583afd623fda4886d179c6.jpg", + "table_caption": [ + "Table 3. Test views performance of 3DGS-based dynamic scene reconstruction method on ImViD dataset." + ], + "table_footnote": [], + "table_body": "
MethodScene1 Opera_girlScene2 LaboratoryScene6 Puppy
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
4DGS23.2270.7530.41025.7980.8900.17618.1210.2220.711
4DRotor27.2630.7750.32828.0070.9170.09817.9160.2980.331
STG28.4820.7860.28726.3060.9100.11420.4970.5940.211
STG++31.2400.7990.27727.5810.9160.10720.5330.5970.202
", + "bbox": [ + 514, + 128, + 903, + 189 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ceff96a6c471a2195677a2040d93fe1f4d9f2a35b7204a003797fcd1f700f62c.jpg", + "table_caption": [ + "Table 4. The storage size of training results from different methods and their rendering speed on a single A100 GPU. *CPU version." + ], + "table_footnote": [], + "table_body": "
Scene1 Opera_girlScene2 LaboratoryScene6 Puppy
Mem(MB)FPSMem(MB)FPSMem(MB)FPS
4DGS101.0546.22105.0244.93101.6541.86
4DRotor*5818.611.006867.180.674415.820.56
STG382.4196.10318.41133.651287.55100.00
STG++387.17108.89329.07137.781276.21110.47
", + "bbox": [ + 514, + 243, + 903, + 320 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5ad60dc675aff4495b9e2776d12cd328b8843c0438b35785af5148d24eced6c8.jpg", + "table_caption": [ + "Table 5. User study for the sound field construction." + ], + "table_footnote": [], + "table_body": "
Auditory Spatial PerceptionSound QualityImmersiveness
Very poor0.00%0.00%0.00%
Poor0.00%0.00%0.00%
Fair14.28%19.04%9.52%
Good23.80%33.33%42.85%
Excellent61.90%47.61%47.61%
", + "bbox": [ + 514, + 358, + 903, + 441 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "puppy. Additionally, due to the density control in the time dimension implemented by 4Drotor, it performs better in frames and regions with more significant motion, which is an aspect worth referencing and adopting in future work. More results can be seen in the supplementary material.", + "bbox": [ + 511, + 465, + 906, + 542 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Results. As shown in Figure 6, Our STG++ achieves better temporal continuity in the color of pixels at the same location. There are almost no abrupt transitions both within and between segments. Details are showed in the supplementary for a more intuitive experience.", + "bbox": [ + 511, + 549, + 905, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Effectiveness of Sound Field Data", + "text_level": 1, + "bbox": [ + 511, + 635, + 808, + 650 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Results. Since our sound field reconstruction algorithm is non-trainable, the reconstructed spatial audio lacks a corresponding ground truth, making it difficult to evaluate the results using quantitative metrics. Therefore, we conducted a user study with 21 experts to assess the performance of the sound field reconstruction.", + "bbox": [ + 511, + 657, + 905, + 747 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Each participant rated the sound field reconstruction on three dimensions: auditory spatial perception, sound quality, and immersiveness. Auditory spatial perception refers to the listener's ability to perceive the distribution and localization of sound in space. Sound quality refers to the audio quality of the spatial audio compared to the corresponding microphone-recorded signal. Immersiveness refers to the listener's sense of being in a space where the sound source genuinely exists.", + "bbox": [ + 511, + 750, + 906, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The results of the user study are shown in Table 5, with", + "bbox": [ + 532, + 885, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16560", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d53e3b037fabb6a88cf96a14c862440b4201c59e4bcbde094403f8d8efcff0c0.jpg", + "image_caption": [ + "Figure 5. Comparison of the rendering results of four baselines on Scene 1 Opera, Scene 2 Laboratory, and Scene 6 Puppy." + ], + "image_footnote": [], + "bbox": [ + 99, + 90, + 888, + 305 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/da0ead87a4c48d8d85119c17848f637e6c43694718d0bc7e740bd8bab431e61c.jpg", + "image_caption": [ + "Figure 6. The continuity of pixels at the same location across different frames and segments. The visualizations present the RGB (middle) and brightness (right) variations." + ], + "image_footnote": [], + "bbox": [ + 94, + 361, + 480, + 536 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "numbers as percentages of participants. The majority of the participants $(61.90\\%)$ rated the auditory spatial perception as excellent, $80.94\\%$ participants felt that the generated spatial audio did not exhibit significant degradation in quality, and $90.46\\%$ found the audio immersive, which demonstrates the effectiveness of our data capture and sound field reconstruction methods.", + "bbox": [ + 89, + 626, + 482, + 731 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Multimodal VR Experiences.", + "text_level": 1, + "bbox": [ + 89, + 753, + 349, + 768 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We finally integrated light field and sound field reconstruction results to achieve a 6-DoF multimodal immersive VR experience with real-time rendering speed at 60 FPS on a single 3090 GPU. As shown in Figure 7, the visual and auditory experiences correspond uniquely to the user's position and viewing direction. The experience can be further explored by listening through headphones with the supplementary video provided.", + "bbox": [ + 89, + 779, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2d974acfdb4ab91af8a93b9c3e0ea54ca4f3aec0a5de2203e0c84586ec2a33b5.jpg", + "image_caption": [ + "Figure 7. Visualization of the interaction trajectory and corresponding visual&auditory results. Note that the multi-modal feedback changes consistently with the positions and orientations." + ], + "image_footnote": [], + "bbox": [ + 532, + 364, + 882, + 583 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 662, + 633, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we introduce ImViD, to our knowledge, the first dataset for immersive volumetric video. This dataset encompasses seven large indoor and outdoor scenes, featuring rich foreground and background content that showcases engaging and commonly encountered daily scenarios, along with high-quality synchronized audio sequences. We validated the effectiveness of the dataset using the latest dynamic light field reconstruction algorithms, providing both quantitative and qualitative analyses of current limitations and potential improvements. Moreover, we propose a base pipeline for constructing immersive volumetric video from this data. We believe ImViD will facilitate the exploration of the limitations of current volumetric video reconstruction algorithms and drive immersive VR/AR experiences.", + "bbox": [ + 509, + 688, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16561", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 240, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 224, + 187, + 242 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] https://www.apple.com/newsroom/2024/07/new-apple-immersive-video-series-and-films-premiere-on-vision-pro/.1", + "[2] https://www.gracia.ai/.1", + "[3] Cal Armstrong, Lewis Thresh, Damian Murphy, and Gavin Kearney. A perceptual evaluation of individual and non-individual hrfts: A case study of the sadie ii database. Applied Sciences, 8(11):2029, 2018. 7", + "[4] Benjamin Attal, Jia-Bin Huang, Christian Richardt, Michael Zollhoefer, Johannes Kopf, Matthew O'Toole, and Changil Kim. Hyperreel: High-fidelity 6-dof video with ray-conditioned sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16610-16620, 2023. 2, 3", + "[5] Swapnil Bhosale, Haosen Yang, Diptesh Kanojia, Jiankang Deng, and Xiatian Zhu. Av-gs: Learning material and geometry aware priors for novel view acoustic synthesis. arXiv preprint arXiv:2406.08920, 2024. 3", + "[6] Michael Broxton, John Flynn, Ryan Overbeck, Daniel Erickson, Peter Hedman, Matthew Duvall, Jason Dourgarian, Jay Busch, Matt Whalen, and Paul Debevec. Immersive light field video with a layered mesh representation. ACM Transactions on Graphics (TOG), 39(4):86-1, 2020. 1, 2, 3, 4, 7", + "[7] Amandine Brunetto, Sascha Hornauer, and Fabien Moutarde. Neraf: 3d scene infused neural radiance and acoustic fields. arXiv preprint arXiv:2405.18213, 2024. 3", + "[8] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 130-141, 2023. 2, 3", + "[9] Hao Chen, Yuqi Hou, Chenyuan Qu, Irene Testini, Xiaohan Hong, and Jianbo Jiao. $360+\\mathrm{x}$ : A panoptic multimodal scene understanding dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19373-19382, 2024. 2, 3, 1, 7", + "[10] Kai Cheng, Xiaoxiao Long, Kaizhi Yang, Yao Yao, Wei Yin, Yuexin Ma, Wenping Wang, and Xuejin Chen. Gaussianpro: 3d gaussian splatting with progressive propagation. arXiv preprint arXiv:2402.14650, 2024. 2", + "[11] Alvaro Collet, Ming Chuang, Pat Sweeney, Don Gillett, Dennis Evseev, David Calabrese, Hugues Hoppe, Adam Kirk, and Steve Sullivan. High-quality streamable free-viewpoint video. ACM Transactions on Graphics (ToG), 34(4):1-13, 2015. 3" + ], + "bbox": [ + 93, + 250, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "This work was supported by National Key R&D Program of China (No.2022YFF0902204), NSFC (No.62171255, 52425801), Tsinghua University-Migu Xinkong Culture Technology (Xiamen) Co.Ltd. Joint Research Center for Intelligent Light Field and Interaction Technology, Guoqiang Institute of Tsinghua University (No.2021GQG0001).", + "[12] Yuanxing Duan, Fangyin Wei, Qiyu Dai, Yuhang He, Wenzheng Chen, and Baoquan Chen. 4d-rotor gaussian splatting: Towards efficient novel view synthesis for dynamic scenes. In ACM SIGGRAPH 2024 Conference Papers, pages 1–11, 2024. 3, 4, 6", + "[13] Bardienus P Duisterhof, Zhao Mandi, Yunchao Yao, Jia-Wei Liu, Mike Zheng Shou, Shuran Song, and Jeffrey Ichnowski. Md-splating: Learning metric deformation from 4d gaussians in highly deformable scenes. arXiv preprint arXiv:2312.00583, 2023. 3", + "[14] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12479–12488, 2023. 2, 3", + "[15] Hang Gao, Ruilong Li, Shubham Tulsiani, Bryan Russell, and Angjoo Kanazawa. Monocular dynamic view synthesis: A reality check. Advances in Neural Information Processing Systems, 35:33768-33780, 2022. 3, 7", + "[16] Huiyu Gao, Jiahao Ma, David Ahmedt-Aristizabal, Chuong Nguyen, and Miaomiao Liu. Soaf: Scene occlusion-aware neural acoustic field. arXiv preprint arXiv:2407.02264, 2024.3", + "[17] Zhiyang Guo, Wengang Zhou, Li Li, Min Wang, and Houqiang Li. Motion-aware 3d gaussian splatting for efficient dynamic scene reconstruction. arXiv preprint arXiv:2403.11447, 2024. 3", + "[18] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE transactions on pattern analysis and machine intelligence, 36(7):1325-1339, 2013. 1", + "[19] Hanbyul Joo, Tomas Simon, Xulong Li, Hao Liu, Lei Tan, Lin Gui, Sean Banerjee, Timothy Scott Godisart, Bart Nabbe, Iain Matthews, Takeo Kanade, Shohei Nobuhara, and Yaser Sheikh. Panoptic studio: A massively multiview system for social interaction capture. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017. 2, 3, 1, 7", + "[20] HyunJun Jung, Nikolas Brasch, Jifei Song, Eduardo Perez-Pellitero, Yiren Zhou, Zhihao Li, Nassir Navab, and Benjamin Busam. Deformable 3d gaussian splatting for animatable human avatars. arXiv preprint arXiv:2312.15059, 2023. 3", + "[21] Takeo Kanade, Peter Rander, and PJ Narayanan. Virtualized reality: Constructing virtual worlds from real scenes. IEEE multimedia, 4(1):34-47, 1997. 3", + "[22] Kai Katsumata, Duc Minh Vo, and Hideki Nakayama. A compact dynamic 3d gaussian representation for real-time dynamic view synthesis, 2024. 2, 3", + "[23] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuehler, and George Drettakis. 3D Gaussian Splatting for Real-Time Radiance Field Rendering. ACM Transactions on Graphics, 42(4):1-14, 2023. 2", + "[24] Agelos Kratimenos, Jiahui Lei, and Kostas Daniilidis. Dynmf: Neural motion factorization for real-time dynamic view synthesis with 3d gaussian splatting. arXiv preprint arXiv:2312.00112, 2023. 2, 3" + ], + "bbox": [ + 93, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16562", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Joo Chan Lee, Daniel Rho, Xiangyu Sun, Jong Hwan Ko, and Eunbyung Park. Compact 3d gaussian splatting for static and dynamic radiance fields. arXiv preprint arXiv:2408.03822, 2024. 3", + "[26] Deqi Li, Shi-Sheng Huang, Zhiyuan Lu, Xinran Duan, and Hua Huang. St-4dgs: Spatial-temporally consistent 4d gaussian splatting for efficient dynamic scene rendering. In ACM SIGGRAPH 2024 Conference Papers, pages 1–11, 2024. 3", + "[27] Lingzhi Li, Zhen Shen, Zhongshu Wang, Li Shen, and Ping Tan. Streaming radiance fields for 3d video synthesis. Advances in Neural Information Processing Systems, 35: 13485-13498, 2022. 2, 3, 1, 7", + "[28] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5521-5531, 2022. 2, 3, 1, 7", + "[29] Yanyan Li, Chenyu Lyu, Yan Di, Guangyao Zhai, Gim Hee Lee, and Federico Tombari. Geogaussian: Geometry-aware gaussian splatting for scene rendering. arXiv preprint arXiv:2403.11324, 2024. 2", + "[30] Zhan Li, Zhang Chen, Zhong Li, and Yi Xu. Spacetime gaussian feature splatting for real-time dynamic view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8508-8520, 2024. 2, 3, 4, 6", + "[31] Susan Liang, Chao Huang, Yapeng Tian, Anurag Kumar, and Chenliang Xu. Av-nerf: Learning neural fields for real-world audio-visual scene synthesis. Advances in Neural Information Processing Systems, 36:37472-37490, 2023. 3, 5", + "[32] Yiqing Liang, Numair Khan, Zhengqin Li, Thu Nguyen-Phuoc, Douglas Lanman, James Tompkin, and Lei Xiao. Gaufre: Gaussian deformation fields for real-time dynamic novel view synthesis. arXiv preprint arXiv:2312.11458, 2023. 3", + "[33] Haotong Lin, Sida Peng, Zhen Xu, Yunzhi Yan, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Efficient neural radiance fields for interactive free-viewpoint video. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 2, 3, 1, 7", + "[34] Kai-En Lin, Lei Xiao, Feng Liu, Guowei Yang, and Ravi Ramamoorthi. Deep 3d mask volume for view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1749–1758, 2021. 2, 3, 1, 7", + "[35] Youtian Lin, Zuozhuo Dai, Siyu Zhu, and Yao Yao. Gaussian-flow: 4d reconstruction with dynamic 3d gaussian particle. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21136-21145, 2024. 2, 3", + "[36] Qingming Liu, Yuan Liu, Jiepeng Wang, Xianqiang Lv, Peng Wang, Wenping Wang, and Junhui Hou. Modgs: Dynamic gaussian splatting from causally-captured monocular videos. arXiv preprint arXiv:2406.00434, 2024. 7", + "[37] Cheng-You Lu, Peisen Zhou, Angela Xing, Chandradeep Pokhariya, Arnab Dey, Ishaan Nikhil Shah, Rugved Mavidi" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "palli, Dylan Hu, Andrew I Comport, Kefan Chen, et al. Diva360: The dynamic visual dataset for immersive neural fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22466-22476, 2024. 2, 7", + "[38] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. arXiv preprint arXiv:2312.00109, 2023. 2", + "[39] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713, 2023. 3", + "[40] Dawid Malarz, Weronika Smolak, Jacek Tabor, Sławomir Tadeja, and Przemysław Spurek. Gaussian splatting with nef-based color and opacity. 2", + "[41] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2", + "[42] David Narciso, Maximino Bessa, Miguel Melo, Antonio Coelho, and José Vasconcelos-Raposo. Immersive $360^{\\circ}$ video user experience: impact of different variables in the sense of presence and cybersickness. Universal Access in the Information Society, 18:77-87, 2019. 1", + "[43] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228, 2021. 2, 1, 7", + "[44] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In CVPR, 2021. 2, 3, 1, 7", + "[45] Weining Ren, Zihan Zhu, Boyang Sun, Jiaqi Chen, Marc Pollefeys, and Songyou Peng. Nerf on-the-go: Exploiting uncertainty for distractor-free nerfs in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8931-8940, 2024. 2, 1, 7", + "[46] Neus Sabater, Guillaume Boisson, Benoit Vandame, Paul Kerbiriou, Frederic Babon, Matthieu Hog, Remy Gendrot, Tristan Langlois, Olivier Bureller, Arno Schubert, et al. Dataset and pipeline for multi-view light-field video. In Proceedings of the IEEE conference on computer vision and pattern recognition Workshops, pages 30-40, 2017. 2, 3, 1, 7", + "[47] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 5", + "[48] Ruizhi Shao, Zerong Zheng, Hanzhang Tu, Boring Liu, Hongwen Zhang, and Yebin Liu. Tensor4d: Efficient neural 4d decomposition for high-fidelity dynamic reconstruction and rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16632-16642, 2023. 2, 3, 1" + ], + "bbox": [ + 516, + 90, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16563", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Roman Shapovalov, Yanir Kleiman, Ignacio Rocco, David Novotny, Andrea Vedaldi, Changan Chen, Filippos Kokkinos, Ben Graham, and Natalia Neverova. Replay: Multi-modal multi-view acted videos for casual holography. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20338-20348, 2023. 2, 3, 1, 7", + "[50] Richard Shaw, Michal Nazarczuk, Jifei Song, Arthur Moreau, Sibi Catley-Chandar, Helisa Dhamo, and Eduardo Perez-Pellitero. Swings: Sliding windows for dynamic 3d gaussian splatting, 2024. 3", + "[51] Liangchen Song, Anpei Chen, Zhong Li, Zhang Chen, Lele Chen, Junsong Yuan, Yi Xu, and Andreas Geiger. Nerfplayer: A streamable dynamic scene representation with decomposed neural radiance fields. IEEE Transactions on Visualization and Computer Graphics, 29(5):2732-2742, 2023. 2, 3", + "[52] Jiakai Sun, Han Jiao, Guangyuan Li, Zhanjie Zhang, Lei Zhao, and Wei Xing. 3dgstream: On-the-fly training of 3d gaussians for efficient streaming of photo-realistic free-viewpoint videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20675-20685, 2024. 3", + "[53] Feng Wang, Sinan Tan, Xinghang Li, Zeyue Tian, Yafei Song, and Huaping Liu. Mixed neural voxels for fast multiview video synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19706-19716, 2023. 2, 3", + "[54] Feng Wang, Zilong Chen, Guokang Wang, Yafei Song, and Huaping Liu. Masked space-time hash encoding for efficient dynamic scene reconstruction. Advances in Neural Information Processing Systems, 36, 2024. 2, 1, 7", + "[55] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20310-20320, 2024. 2, 3, 4, 6", + "[56] Tianhao Wu, Fangcheng Zhong, Andrea Tagliasacchi, Forrester Cole, and Cengiz Oztireli. D^2nerf: Self-supervised decoupling of dynamic and static objects from a monocular video. Advances in neural information processing systems, 35:32653-32666, 2022. 2, 1, 7", + "[57] Yunzhi Yan, Haotong Lin, Chenxu Zhou, Weijie Wang, Haiyang Sun, Kun Zhan, Xianpeng Lang, Xiaowei Zhou, and Sida Peng. Street gaussians for modeling dynamic urban scenes. arXiv preprint arXiv:2401.01339, 2024. 2", + "[58] Zeyu Yang, Hongye Yang, Zijie Pan, Xiatian Zhu, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. arXiv preprint arXiv:2310.10642, 2023. 3", + "[59] Ziyi Yang, Xinyu Gao, Yangtian Sun, Yihua Huang, Xiaoyang Lyu, Wen Zhou, Shaohui Jiao, Xiaojuan Qi, and Xiaogang Jin. Spec-gaussian: Anisotropic view-dependent appearance for 3d gaussian splatting. arXiv preprint arXiv:2402.15870, 2024. 2", + "[60] Jae Shin Yoon, Kihwan Kim, Orazio Gallo, Hyun Soo Park, and Jan Kautz. Novel view synthesis of dynamic scenes with" + ], + "bbox": [ + 89, + 92, + 482, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "globally coherent depths from a monocular camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5336-5345, 2020. 2, 3, 1, 7", + "[61] Dongbin Zhang, Chuming Wang, Weitao Wang, Peihao Li, Minghan Qin, and Haoqian Wang. Gaussian in the wild: 3d gaussian splatting for unconstrained image collections. arXiv preprint arXiv:2403.15704, 2024. 2" + ], + "bbox": [ + 516, + 92, + 903, + 191 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "16564", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/ImViD_ Immersive Volumetric Videos for Enhanced VR Engagement/6da376dc-6436-436c-b26c-23ba201328ef_model.json b/2025/ImViD_ Immersive Volumetric Videos for Enhanced VR Engagement/6da376dc-6436-436c-b26c-23ba201328ef_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6f08b6599c67a8450b8514241d2fac986cf4b4c6 --- /dev/null +++ b/2025/ImViD_ Immersive Volumetric Videos for Enhanced VR Engagement/6da376dc-6436-436c-b26c-23ba201328ef_model.json @@ -0,0 +1,2356 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.152, + 0.131, + 0.848, + 0.154 + ], + "angle": 0, + "content": "ImViD: Immersive Volumetric Videos for Enhanced VR Engagement" + }, + { + "type": "text", + "bbox": [ + 0.119, + 0.18, + 0.882, + 0.289 + ], + "angle": 0, + "content": "Zhengxian Yang\\(^{1*}\\), Shi Pan\\(^{1*}\\), Shengqi Wang\\(^{1*}\\), Haoxiang Wang\\(^{1}\\), Li Lin\\(^{2}\\), Guanjun Li\\(^{3}\\), Zhengqi Wen\\(^{1\\dagger}\\), Borong Lin\\(^{1\\dagger}\\), Jianhua Tao\\(^{1\\dagger}\\), Tao Yu\\(^{1\\dagger}\\), Tsinghua University \\({}^{2}\\)Migu Beijing Research Institute \\({}^{3}\\)Institute of Automation, Chinese Academy of Science \\(\\{\\mathrm{zx - yang23, ps23, shengqi - 21, whx22}\\} @\\)mails.tsinghua.edu.cn lilin@migu.chinamobile.com guanjun.li@nlpr ia.ac.cn \\(\\{\\mathrm{zqwen, linbr, jhtao, ytrock}\\} @\\)tsinghua.edu.cn" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.31, + 0.306, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.306, + 0.31, + 0.692, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.693, + 0.31, + 0.905, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.497, + 0.908, + 0.526 + ], + "angle": 0, + "content": "Figure 1. We introduce ImViD, a dataset for immersive volumetric videos. ImViD records dynamic scenes using a multi-view audio-video capture rig moving in a space-oriented manner, which provides a new benchmark for volumetric video reconstruction and its application." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.54, + 0.327, + 0.556 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.572, + 0.483, + 0.768 + ], + "angle": 0, + "content": "User engagement is greatly enhanced by fully immersive multi-modal experiences that combine visual and auditory stimuli. Consequently, the next frontier in VR/AR technologies lies in immersive volumetric videos with complete scene capture, large 6-DoF interaction space, multimodal feedback, and high resolution&frame-rate contents. To stimulate the reconstruction of immersive volumetric videos, we introduce ImViD, a multi-view, multi-modal dataset featuring complete space-oriented data capture and various indoor/outdoor scenarios. Our capture rig supports multi-view video-audio capture while on the move, a capability absent in existing datasets, significantly enhancing the completeness, flexibility, and efficiency of data capture." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.769, + 0.483, + 0.844 + ], + "angle": 0, + "content": "The captured multi-view videos (with synchronized audi- 15 minutes, and include rich foreground-background elements, and complex dynamics. We benchmark existing methods using our dataset and establish a base pipeline for construct" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.542, + 0.907, + 0.648 + ], + "angle": 0, + "content": "ing immersive volumetric videos from multi-view audiovisual inputs for 6-DoF multi-modal immersive VR experiences. The benchmark and the reconstruction and interaction results demonstrate the effectiveness of our dataset and baseline method, which we believe will stimulate future research on immersive volumetric video production. Project Page: https://yzxqh.github.io/ImViD/" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.664, + 0.645, + 0.68 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.689, + 0.907, + 0.81 + ], + "angle": 0, + "content": "High-quality and high-fidelity digital modeling of the real world has long been a topic of significant interest. With breakthroughs in 3D reconstruction algorithms and virtual reality technology, the generation of such visually compelling content is gradually becoming feasible, whether for reproducing past events and memories or remotely viewing current occurrences. Achieving complete user immersion is a goal pursued by many researchers." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.811, + 0.836, + 0.825 + ], + "angle": 0, + "content": "What defines an immersive volumetric video?" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Dwelling on previous tests [1, 6] and user feedback [2, 42], we believe that immersive media should have the following four characteristics: i) Full \\(360^{\\circ}\\) foreground and background, ii) High-quality 6-DoF interaction, which current volumetric video can achieve, iii) Multimodal experi" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.851, + 0.483, + 0.901 + ], + "angle": 0, + "content": "*Equal contributions. † Corresponding author. \nZhengxian Yang, Shengqi Wang, Haoxiang Wang, Jianhua Tao and Tao Yu are at BNRist&Department of Automation; Shi Pan and Borong Lin are at Department of Building Science; Zhengqi Wen is at BNRist." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16554" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.136 + ], + "angle": 0, + "content": "ence with light and sound, and iv) High frame rate and long-duration dynamic content. We refer to a media format that meets all these criteria as immersive volumetric video." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.138, + 0.483, + 0.365 + ], + "angle": 0, + "content": "However, with the development of camera equipment, despite many efforts using monocular handheld cameras or fixed camera arrays to build datasets, none of these solutions fully address all four requirements. For example, Panoptic Sports [19], ZJU-Mocap [44] and DiVa-360 [37] achieved human/object-centric \\(360^{\\circ}\\) capture but lack complex backgrounds and multimodal data; Google's immersive light field [6] achieves 6-DoF volumetric video in an inside-looking-out manner but lacks multimodality and high frame rate, long-duration dynamic content; Replay [49] introduces sound, but the camera setup is not suitable for human viewing habits thus cannot achieve high-quality 6-DoF interaction. Both academia and industry require an appropriate capture method to support the development of immersive volumetric video technology." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.366, + 0.483, + 0.668 + ], + "angle": 0, + "content": "Additionally, NeRF [41]'s emergence has increased attention to reconstruction and rendering techniques, making it easier to generate high-fidelity interactive 3D scenes from multiple views. 3DGS [23] and its subsequent improvements [10, 29, 38, 40, 57, 59, 61] have further boosted rendering speed and quality, fueling peoples' expectations for rapid real-world modeling and its applications in industries such as education, healthcare, and entertainment. This can be seen as an evolutionary opportunity for immersive volumetric video. However, current technologies face challenges. In dynamic light field reconstruction, neither explicit [22, 24, 30, 35, 55] nor implicit neural representation based [4, 8, 14, 28, 48, 51, 53] methods have been able to achieve a practical balance between foreground-background reconstruction accuracy, rendering speed, spatiotemporal consistency, and storage efficiency in dynamic scenes; In the realm of sound field reconstruction, it is typically necessary to collect a large amount of binaural audio or room impulse response data from different locations for model training, which is user-unfriendly in practical use." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.669, + 0.483, + 0.849 + ], + "angle": 0, + "content": "To better support the development of immersive volumetric video technology, we have developed a moving capture rig (with 46 synchronized cameras, as shown in Fig.3) and designed two space-oriented data capture strategies. We have constructed an audiovisual dataset containing 7 indoor and outdoor scenes, with a \\(360^{\\circ}\\) field of view, large 6-DoF interaction space, and high-resolution, long-duration dynamic contents. We also validate this dataset using the latest novel view synthesis and 3D reconstruction methods. Finally, a complete pipeline for constructing light and sound fields from this type of data was proposed, which initially achieved immersive volume video." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.851, + 0.472, + 0.866 + ], + "angle": 0, + "content": "Our main contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "- We propose to our knowledge the first and probably the highest-quality multi-view dataset ImViD for immersive" + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "volumetric video. It features \\(360^{\\circ}\\) foreground and background capture, long-duration video&audio data under high resolution@5K, high frame rate@60FPS." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.137, + 0.905, + 0.182 + ], + "angle": 0, + "content": "- We conduct extensive experiments to validate the effectiveness of the proposed datasets by evaluating and analyzing the performance of several current baselines." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.182, + 0.905, + 0.228 + ], + "angle": 0, + "content": "- Follow the proposed pipeline shown in Figure 2, We finally achieve the reconstruction of immersive volumetric video and integrate it for immersive VR experience." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.137, + 0.905, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.243, + 0.655, + 0.259 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.269, + 0.905, + 0.314 + ], + "angle": 0, + "content": "Datasets for Dynamic Novel View Synthesis. The datasets for dynamic scene reconstruction can be classified as monocular-based and multi-view-based." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.316, + 0.905, + 0.451 + ], + "angle": 0, + "content": "Monocular acquisition systems are popular due to their low cost and ease of construction. such as HyperNeRF [43], Dynamic Scene Dataset [60], and D2NeRF [56]. However, these datasets suffer from low resolution, limited capture space, and the limited durations. Although NeRF On-the-go [45] allows for larger capture ranges by walking while shooting, high-quality reconstructions are confined to the capture path, and the small field of view (FOV) limits prolonged observations of specific scene positions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.453, + 0.906, + 0.619 + ], + "angle": 0, + "content": "Multi-camera data collection has gained significant attention due to its ability to provide a larger FOV and richer details, such as Immersive Light Field dataset [6], Technicolor [46], UCSD Dynamic Scene Dataset [34] and The Plenoptic Dataset [28], followed by [27, 33, 54]. However, all these setups remain static during capture, limiting them to frontal views and hindering \\(360^{\\circ}\\) reconstruction. Additionally, the video sequences are typically short, with a maximum duration of 2 minutes (often less than 30 seconds) and a maximum resolution of 4K, which is insufficient for immersive VR experiences." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.62, + 0.906, + 0.831 + ], + "angle": 0, + "content": "Moreover, the previously mentioned datasets lack sound recordings, leading to the loss of immersion without multimodality. The Replay dataset [49] addresses this by focusing on long sequences with professional actors in familiar settings. However, aside from the head-mounted cameras that enables rotation with head, all other cameras remain static. Furthermore, the DSLR arrangement does not align with human viewing habits in VR, making them unsuitable as benchmarks for novel view synthesis tasks. Recent work [9] presents a dataset with a \\(360^{\\circ}\\) camera and audio capture device but constrained by a fixed-point shooting strategy, resulting in sparse viewpoints that hinder the reconstruction of high-quality dynamic scenes. More details about datasets can be found in Tab. 1." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.84, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Dynamic Light&Sound Field Reconstruction. Temporal variations in input images make reconstructing dynamic scenes more challenging than static ones. The key issue is efficiently representing 4D scenes while ensuring spa" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16555" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.155, + 0.09, + 0.843, + 0.105 + ], + "angle": 0, + "content": "Table 1. Existing real-world datasets for dynamic novel view synthesis (full version in the supplementary materials)." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.115, + 0.907, + 0.428 + ], + "angle": 0, + "content": "
DatasetsNo SceneOutdoor/IndoorCamerasMobilityResolutionAnglesDurationFPSMultimodalityContent
PanopticSports [19]65Indoor480 camerasStatic640×480360°5mins25XHuman-centric actions
Technicolor [46]12Indoor16 camerasStatic2048×1088Frontal2s30XHas a number of close-ups sequences, captured medium angle scenes and other animated scenes where the movement does not come from a human
Immersive-Lightfield [6]15both46 camerasStatic2560×1920Frontal10-30s30XSimple and slow motion of human, animals, objects
Dynamic Scene Datasets (NVIDIA) [60]8Outdoor1 Mobile phone /12 camerasFixed-point Waving /Static1920×1080Frontal5s60XSimple body motions (facial, jump, etc)
UCSD Dynamic [34]96Outdoor10 camerasStatic1920×1080Frontal1-2mins120XVarious visual effects and human interactions
ZJU-Mocap [44]10Indoor21 camerasStatic1024×1024360°20s50XSimple body motions (punch, kick, etc.)
Plenoptic Dataset (DyN-eRF/Neural 3D) [28]6Indoor21 camerasStatic2704×2028Frontal10-30s30XContains high specularity, translucency and transparency objects, motions with changing topology, selfcast moving shadows, various lighting conditions and multiple people moving around in open living room space
iPhone Datasets [15]14both1 hand-held phone /2 camerasFixed-point Waving /Static640×480Frontal8-15s30/60XFeaturing non-repetitive motion, from various categories such as generic objects, humans, and pets
ENeRF-Outdoor [33]4Outdoor18 camerasStatic1920×1080Frontal20-40s30XComplex human motions
Replay [49]46Indoor12 camerasStatic3840×2160360°5mins30✓(Audio)Dancing, chatting, playing video games, unwrapping presents, playing ping pong
360+X [9]28both1 360°cameras and 1 Spectacles camerasStatic5760×2880360°10s (2152 sequence)30✓(Audio)Capture in 17 cities across 5 countries. Panoptic perspective to scene understanding with audio
Im-ViD(Ours)7both46 camerasMoveable5312×2988Frontal and 360°1-5mins60✓(Audio)Seven common indoor and outdoor scenes in daily life, including opera, face-to-face communication, teaching, discussion, music performance, interaction with pets, and playing. Each scene has high-quality synchronized multi-view video and audio with a duration of more than 1 minute, and contains rich elements such as various small objects, glass, and changes in light and shadow
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.452, + 0.486, + 0.892 + ], + "angle": 0, + "content": "tiotemporal consistency and high accuracy. Traditional methods estimate varying geometries like surfaces [11] and depth maps [21] to handle dynamics. Recently, efforts have shifted towards adapting neural radiance field representations for dynamic scenes. DyNeRF [28] pioneer in the dynamic reconstruction with Neural radiance field. Followed works such as [4, 8, 14, 27, 48, 51, 53] make efforts to improve the high-fidelity and efficiency in the 4D neural representation. However, the common challenges these methods face are memory cost and the modeling of spatiotemporal complexity. The latest work have extended 3DGS to dynamic scenes for high-quality real-time rendering. While modeling dynamic scenes as static ones is straightforward, methods relying solely on monocular data [20, 32] is not suitable for this paradigm, and face challenges of adding constraints or priors to better supervise training. Other works support multi-view data as input. [39] and [52] initialize each frame from the previous one to build Gaussian point clouds per-frame, leading to high memory and suddenly appearing or disappearing artifacts. 4DGS [55] introduces Hexplane to encode spatiotemporal information, using MLPs to predict Gaussian properties change. Subsequent works [13, 17, 26] further add geometric and optical flow constraints to improve frame-to-frame motion but perform well only on short sequences with small actions. Other approaches [22, 24, 30, 35] employ the explicit parametric form (such as radial basis functions) of 3DGS to enhance spatiotemporal consistency. Some methods [12, 58] focus on representation with dual quaternions, 4D spheri" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.452, + 0.907, + 0.528 + ], + "angle": 0, + "content": "cal harmonics, or rotors for end-to-end training. More improvements such as parallel training [50] and using codebooks and masks to enhance scene compactness [25], show the potential in efficient and high-quality representation of long-term complex dynamic scenes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.529, + 0.909, + 0.832 + ], + "angle": 0, + "content": "In addition to light field reconstruction, reconstructing the sound field from multimodal data is equally important, often referred to as novel-view acoustic synthesis. For the novel-view acoustic synthesis task, AV-NeRF [31] synthesizes spatial audio by leveraging visual features from rendered images. NeRAF [7] further decouples the camera and microphone poses used for training. AV-GS [5] incorporates a 3D scene representation during sound field synthesis, addressing the issue where previous methods overlooked the contribution of the broader 3D scene geometry. SOAF [16] considers occlusion in the light field when synthesizing the sound field. However, the above novel-view acoustic synthesis tasks typically require a large amount of multimodal training data collected from different locations, which is often user-unfriendly in practical applications. Moreover, these methods generally cannot solve the problem of sound field synthesis for moving sound sources. This paper proposes a training-free novel-view acoustic synthesis approach for our dataset, which also supports sound field synthesis for moving sound sources." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.845, + 0.665, + 0.861 + ], + "angle": 0, + "content": "3. ImViD Dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Our goal is to build an immersive volumetric video dataset of real world that integrates both foreground and back-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16556" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.123, + 0.089, + 0.878, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.429, + 0.908, + 0.513 + ], + "angle": 0, + "content": "Figure 2. The pipeline to realize the multimodal 6-DoF immersive VR experiences. We applied a carefully designed rig to (a) simultaneously capture multi-view video and audio. The (b1) presents our reconstruction of dynamic light field based on STG [30] while (b2) demonstrates the construction process of sound field. We have achieved better results than the original algorithm in long-term dynamic scenes by incorporating affine color transformation and t-dimensional density control. Ultimately, we achieve a 6-DoF immersive experience in both light and sound fields, and also benchmark on recent representative methods like 4DGS [55] and 4Drotor [12] to demonstrate the effectiveness of both our dataset and baseline method." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.54, + 0.486, + 0.751 + ], + "angle": 0, + "content": "ground, providing assistance for the development of related research such as spatial video and VR/AR application. Therefore, we have meticulously refined every aspect of the process, from equipment setup and scene selection to shooting strategies and data processing, ensuring maintaining the elements of interest to researchers to the greatest extent. The released ImViD dataset comprises a total of 7 large scenes, including 5 indoor and 2 outdoor environments. The content features a diverse range of common activities, such as opera, meeting, teaching, and playing on the grass. Each scene contains a set of static photos, multi-view time-synchronized video&audio sequences at 5K resolution and 60 FPS, employing two different collection strategies, along with calibrated camera poses." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.758, + 0.258, + 0.775 + ], + "angle": 0, + "content": "3.1. Data Acquisition" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.485, + 0.901 + ], + "angle": 0, + "content": "The handheld monocular camera is easy to move, providing limited perspectives from various locations. In contrast, the fixed camera arrays, while stationary, offers dense perspectives within a limited range. We aim to combine the advantages of both to design an effective data collection system and strategy for fully immersive VR experience—a feature not found in existing datasets. As shown in Figure 3, our capture system consists of more than 40 GoPro cameras in" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.54, + 0.907, + 0.706 + ], + "angle": 0, + "content": "stalled on a hemispherical surface mounted on a remotely controllable mobile car, all of which are connected to a PC to receive synchronous control commands. The system is carefully designed to be at a height similar to that of a human, and the camera arrangement better aligns with human viewing habits compared to other existing devices. This rig can synchronously collect about 1,000 images at 5K resolution within a space of at least \\(6\\mathrm{m}^3\\) in 2 min. It is also capable of synchronously capturing video and audio at 60 FPS and 5K resolution for 30 min (limited by heat dissipation). Thus, our capture strategy is as follows:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.708, + 0.907, + 0.739 + ], + "angle": 0, + "content": "Step1: Efficient high-quality, high-density \\(360^{\\circ}\\) image acquisition of static environments." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.741, + 0.905, + 0.772 + ], + "angle": 0, + "content": "Step2: Continuous synchronous collection of dynamic scenes under two different strategies." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.775, + 0.905, + 0.804 + ], + "angle": 0, + "content": "i) Fixed-point shooting. The cart remains stationary during the capture, focusing on capturing dynamic details." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.808, + 0.905, + 0.853 + ], + "angle": 0, + "content": "ii) Moving shooting. The cart slowly moves while the scene is in motion, providing additional level of details and larger exploratory space." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.775, + 0.905, + 0.853 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Throughout the capture process, we ensure strict synchronization of the world time codes across all cameras. We carefully maintained consistent focal lengths, exposure," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16557" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.092, + 0.29, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.289, + 0.483, + 0.318 + ], + "angle": 0, + "content": "Figure 3. Our rig support two kinds of capturing strategies for high resolution, high frame rate and \\(360^{\\circ}\\) dynamic data acquisition." + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.092, + 0.468, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.342, + 0.484, + 0.435 + ], + "angle": 0, + "content": "and white balance settings for each scene, based on the results of manual iterative testing of camera settings, thereby minimizing hardware-induced variations in data quality. To achieve high-quality audio, we increased the noise reduction coefficient of the cameras and minimized noise generation as much as possible." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.441, + 0.253, + 0.457 + ], + "angle": 0, + "content": "3.2. Data Processing" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.463, + 0.483, + 0.657 + ], + "angle": 0, + "content": "For the photos from Step1, we completed the calibration in COLMAP [47] by combining the cameras' prior parameters and their relative positions. For the video&audio from Step2, we first aligned the multi-view sequences for each scene using the world time codes. This step is crucial as the quality of the alignment directly determines the quality of dynamic scene reconstruction and significantly impacts whether the multi-view audio can be used for constructing a 6-DoF sound field. Our time code alignment error is at the level of approximately 2 milliseconds. Subsequently, we performed frame extraction from the videos, again utilizing the GoPro camera parameters as priors to complete the calibration." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.826 + ], + "angle": 0, + "content": "It is important to note that for the dynamic scene data captured with a moving rig, we also performed time code alignment. However, this data poses significant challenges for existing calibration methods, often resulting in errors and floaters. There are also no solutions for reconstruction algorithms specifically addressing dynamic scenes from moving rig currently. Therefore, we do not provide the calibrated results for this type of data at this time. But we believe that this data will greatly contribute to the advancement of the field, and thus we will also publicly releasing the data which has been time-synchronized." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.834, + 0.258, + 0.851 + ], + "angle": 0, + "content": "3.3. Dataset Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Diversity. Our dataset, as depicted in Table 2, includes 7 common indoor and outdoor scenes, each comprising one or more takes, totaling 16 video sequences. Each scene is" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.092, + 0.9, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.27, + 0.908, + 0.326 + ], + "angle": 0, + "content": "Figure 4. Calculation method for spatiotemporal capture density. 1) the capture strategy of handheld monocular camera 2) represents the fixed camera array 3) Our rig covers a volume of \\(0.6 \\times \\pi \\times 0.5^2 \\, \\mathrm{m}^3\\) over 5 seconds." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.355, + 0.907, + 0.446 + ], + "angle": 0, + "content": "rich in foreground and background elements, meticulously crafted to preserve the authenticity of the environment. The characters, clothing, and motions within each scene exhibit a diverse range, set against varying ambient lighting conditions. We also included areas of interest like tables, chairs, windows as much as possible for researchers." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.447, + 0.907, + 0.492 + ], + "angle": 0, + "content": "Quality. Leveraging a total of 39 cameras, we captured dynamic scenes at a resolution of \\(5312 \\times 2988\\) and 60 FPS. The images of static environment is \\(5568 \\times 4176\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.494, + 0.906, + 0.569 + ], + "angle": 0, + "content": "Duration. As outlined in Table 2, each scene in our dataset contains at least 1 minute of footage. The longest sequence in our forthcoming public dataset spans approximately 5 minutes. The total processed data duration amounts to 38 minutes 46 seconds, including 139,560 frames per camera." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.57, + 0.907, + 0.737 + ], + "angle": 0, + "content": "Coverage. As highlighted in Table 2, we conducted image acquisition from over 1,000 viewpoints before recording dynamic content, thereby offering a \\(180^{\\circ} - 360^{\\circ}\\) viewing space in VR. For scenes employing the mobile shooting strategy, we introduced a metric termed 'spatiotemporal capture density,' quantifying the captured volume covered by our rig per second, as shown in Figure 4. This metric, unavailable in current datasets (where static camera arrays are deemed 0 and handheld monocular cameras follow a single trajectory without capturing volume), registers at approximately \\(0.10~m^3 /s\\) in our dataset." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.754, + 0.794, + 0.772 + ], + "angle": 0, + "content": "4. Benchmarks and Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.907, + 0.901 + ], + "angle": 0, + "content": "To comprehensively assess the effectiveness of the dataset, we selected three scenes (2 indoor and 1 outdoor) to test the latest dynamic light field reconstruction algorithms, providing a benchmark for research in areas such as novel view synthesis. We also proposed an improvement strategy based on one of the methods. Furthermore, we presented and validated a feasible pipeline for constructing 6-DoF sound fields from multi-view audio sequences. In this section, we" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16558" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.133 + ], + "angle": 0, + "content": "Table 2. The detailed statistics of our ImViD dataset. #Take represents the number of different episodes we collected in this scene. #Strategy represents the strategies used in Step2. 1 indicates only (i) was used in Step 2, 2 indicates that both were used. Avg.S-T Density indicates the spatiotemporal coverage density of our collection method, measured by the captured volume size per second." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.142, + 0.905, + 0.268 + ], + "angle": 0, + "content": "
Name#Camera#Static viewpoints#Take#StrategyAvg.S-T Density(m3/s)Viewing SpaceAvg.DurationStorage(GB)
Scene1 Opera39115221-180°3min22s226
Scene2 Laboratory391225220.10360°1min42s137.3
Scene3 Classroom391223220.10360°4min42s497
Scene4 Meeting39122311-360°3min16s114
Scene5 Rendition391620420.10360°2min02s516
Scene6 Puppy391404320.10360°1min50s359
Scene7 Playing391224220.10360°1min10s220
Total--16---38min46s2069.3
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.292, + 0.484, + 0.339 + ], + "angle": 0, + "content": "will first introduce the selected light field pipeline and our proposed sound field pipeline (Sec 4.1), then provide details on the experimental setup (Sec 4.2)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.347, + 0.193, + 0.361 + ], + "angle": 0, + "content": "4.1. Baseline" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.368, + 0.484, + 0.445 + ], + "angle": 0, + "content": "Dynamic Light Field Reconstruction. We conducted a thorough survey of the current dynamic reconstruction algorithms, prioritizing rendering speed and support for multiview data input. We finally decided to establish a benchmark on 3DGS-based method from three paradigms." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.444, + 0.484, + 0.55 + ], + "angle": 0, + "content": "4DGS [55] introduces Hexplane to encode spatiotemporal information to feature \\( f_{d} \\) and use MLPs decoders \\( D = \\{\\phi_{x},\\phi_{r},\\phi_{s}\\} \\) to predict gaussians' changes in position \\( \\Delta \\chi = \\phi_{x}(f_{d}) \\), rotation \\( \\Delta r = \\phi_r(f_d) \\), and scaling \\( \\Delta s = \\phi_s(f_d) \\). Utilizing the sparse point cloud from frame 0 as initialization and training static gaussians for the first 3000 iterations, the dynamic scene \\( G \\) is represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.56, + 0.482, + 0.577 + ], + "angle": 0, + "content": "\\[\nG \\left(\\chi^ {\\prime}, r ^ {\\prime}, s ^ {\\prime}\\right) = G \\left(\\chi + \\Delta \\chi , r + \\Delta r, s + \\Delta s\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.587, + 0.484, + 0.648 + ], + "angle": 0, + "content": "4DrotorGS [12] is a representative method of dimensionality expansion. It extends 3DGS to a four-dimensional Gaussian ellipsoid by introducing the concept of rotor, expanding the covariance matrix to four dimensions, i.e.:" + }, + { + "type": "equation", + "bbox": [ + 0.164, + 0.657, + 0.482, + 0.677 + ], + "angle": 0, + "content": "\\[\nG _ {4 D} (x) = e ^ {- \\frac {1}{2} (x - \\mu_ {4 D}) ^ {T} \\Sigma_ {4 D} ^ {- 1} (x - \\mu_ {4 D})} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.687, + 0.483, + 0.717 + ], + "angle": 0, + "content": "Each frame of the scene corresponds to a time slice of the 4D Gaussian ellipsoids followed by \\(\\alpha\\)-blending." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.717, + 0.484, + 0.807 + ], + "angle": 0, + "content": "STG [30], on the other hand, use polynomial fitting for the motion and rotation of the 3DGS, and employs an RBF function to model the opacity changes over time. It concatenates the sparse point clouds from all frames as the initialization, and the final representation of each point in the scene is:" + }, + { + "type": "equation", + "bbox": [ + 0.099, + 0.815, + 0.482, + 0.855 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {i} (t) = \\sigma_ {i} (t) \\exp \\left(- \\frac {1}{2} \\left(x - \\mu_ {i} (t)\\right) ^ {T} \\Sigma_ {i} (t) ^ {- 1} \\left(x - \\mu_ {i} (t)\\right)\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Although we ensured consistent white balance and other settings across cameras based on the characteristics of scenes during shooting, color differences still exist due to" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.292, + 0.907, + 0.459 + ], + "angle": 0, + "content": "factors like occlusions when capturing from different angles. During experiments, we found that while existing methods fit well to the training views, there is a noticeable color shift during viewpoint transitions in the reconstructed light field, manifesting as flickering and floaters. This is especially apparent when training in segments, where the differences between segments are more pronounced, significantly reducing immersion. We improved the original STG (named as \\(\\mathrm{STG}++\\)) by introducing global, time-invariant color mapping during training. Detailed implementation can be seen in supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.463, + 0.907, + 0.644 + ], + "angle": 0, + "content": "Sound Field Reconstruction. Taking the recording microphone as the origin of the coordinate system, we denote the coordinates of the sound source as \\( s(t) = \\{x_s(t), y_s(t)\\} \\), the user wearing the VR device as \\( l(t) = \\{x_l(t), y_l(t)\\} \\), and the angle of deviation from the y-axis in the counterclockwise direction as \\( \\theta_l(t) \\), where \\( t \\) is the time index. Based on these premises, generating the user's binaural spatial audio can be divided into three parts: sound direction mapping, sound distance mapping, and sound auralization. Note that we assume there is only one dominating sound source (omnidirectional emitting) to simplify the sound field reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.645, + 0.906, + 0.676 + ], + "angle": 0, + "content": "In the sound direction mapping part, we calculate the direction of the sound source relative to the user's ears, \\(\\theta_{s}(t)\\)" + }, + { + "type": "equation", + "bbox": [ + 0.607, + 0.686, + 0.905, + 0.719 + ], + "angle": 0, + "content": "\\[\n\\theta_ {s} (t) = \\operatorname {a r c c o s} \\frac {\\mathbf {v} _ {1} (t) \\mathbf {v} _ {2} (t)}{| \\mathbf {v} _ {1} (t) | | \\mathbf {v} _ {2} (t) |}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.722, + 0.56, + 0.735 + ], + "angle": 0, + "content": "where" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.745, + 0.905, + 0.763 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} _ {1} (t) = \\left[ x _ {s} (t) - x _ {l} (t), y _ {s} (t) - y _ {l} (t) \\right] ^ {T}, \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.765, + 0.905, + 0.784 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} _ {2} (t) = \\left[ - \\sin \\left(\\theta_ {l} (t)\\right), \\cos \\left(\\theta_ {l} (t)\\right) \\right] ^ {T}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.792, + 0.906, + 0.853 + ], + "angle": 0, + "content": "In the sound distance mapping part, we calculate the scaling of the sound source when it reaches the user's ears, relative to the original recording. We denote this scaling parameter as \\(\\lambda(t)\\)," + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.861, + 0.905, + 0.9 + ], + "angle": 0, + "content": "\\[\n\\lambda (t) = \\frac {\\sqrt {x _ {s} ^ {2} (t) + y _ {s} ^ {2} (t)}}{\\sqrt {(x _ {s} (t) - x _ {l} (t)) ^ {2} + (y _ {s} (t) - y _ {l} (t)) ^ {2}}} \\tag {7}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16559" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.184 + ], + "angle": 0, + "content": "In the sound auralization part, we convert the original recording into spatial audio based on \\(\\theta_s(t)\\) and \\(\\lambda (t)\\). In the short-time Fourier transformation (STFT) domain, we denote the audio from the left and right ears as \\(A_{L}(t,f)\\) and \\(A_{R}(t,f)\\), respectively, where \\(f\\) is the frequency index. Then \\(A_{L}(t,f)\\) and \\(A_{R}(t,f)\\) can be written as" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.195, + 0.483, + 0.211 + ], + "angle": 0, + "content": "\\[\nA _ {L} (t, f) = F _ {L} \\left(\\theta_ {s} (t)\\right) F _ {l} (t, f) A _ {O} (t, f) \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.155, + 0.213, + 0.483, + 0.23 + ], + "angle": 0, + "content": "\\[\nA _ {R} (t, f) = F _ {R} \\left(\\theta_ {s} (t)\\right) F _ {l} (t, f) A _ {O} (t, f), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.241, + 0.485, + 0.379 + ], + "angle": 0, + "content": "where \\(F_{L}(\\theta_{s}(t))\\) and \\(F_{R}(\\theta_{s}(t))\\) are head-related transfer functions (HRTF) based on \\(\\theta_s(t)\\), \\(F_{l}(t)\\) denotes the room impulse response (RIR) at the user's location and \\(A_{O}(t,f)\\) is the sound emitted from the speaker in the STFT domain. We utilize \\(\\theta_{s}(t)\\) to retrieve \\(F_{L}(\\theta_{s}(t))\\) and \\(F_{R}(\\theta_{s}(t))\\) from the SADIE II dataset [3]. In order to model \\(F_{l}(t)\\) and \\(A_{O}(t,f)\\), we make the following assumptions: \\(F_{l}(t)\\) change between different locations at the same time index \\(t\\) only depending on the parameter \\(\\lambda (t)\\), i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.208, + 0.39, + 0.483, + 0.407 + ], + "angle": 0, + "content": "\\[\nF _ {l} (t, f) = \\lambda (t) F (t, f), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.417, + 0.484, + 0.448 + ], + "angle": 0, + "content": "where \\( F(t) \\) is the RIR at the recording microphone. Using the above assumption, Eq.(8) and Eq.(9) can be rewritten as" + }, + { + "type": "equation", + "bbox": [ + 0.171, + 0.476, + 0.483, + 0.492 + ], + "angle": 0, + "content": "\\[\nA _ {L} (t, f) = \\lambda (t) F _ {L} \\left(\\theta_ {s} (t)\\right) A (t, f), \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.173, + 0.494, + 0.483, + 0.511 + ], + "angle": 0, + "content": "\\[\nA _ {R} (t, f) = \\lambda (t) F _ {R} \\left(\\theta_ {s} (t)\\right) A (t, f), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.522, + 0.484, + 0.553 + ], + "angle": 0, + "content": "where \\( A(t,f) = F(t,f)A_O(t,f) \\) is the audio at the recording microphone." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.562, + 0.299, + 0.578 + ], + "angle": 0, + "content": "4.2. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.484, + 0.658 + ], + "angle": 0, + "content": "Training setups. The training for all light field data was completed on a single NVIDIA A100 GPU. Due to memory limitations, we split the video and trained a model every 60 frames, total 300 frames. We also followed the original train-test-split strategy of the baseline for our experiments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.484, + 0.705 + ], + "angle": 0, + "content": "Metric. We analyzed the quality of dynamic light field reconstruction using the most common evaluation metrics PSNR, SSIM, and LPIPS(alex)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.719, + 0.29, + 0.735 + ], + "angle": 0, + "content": "5. Results and Analysis" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.743, + 0.357, + 0.759 + ], + "angle": 0, + "content": "5.1. Validation of Light Field Data" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Quantitative Results. As illustrated in Table 3, 4 and Figure 5, our indoor and outdoor data both yield correct results across three 3DGS-based paradigms. STG++ outperforms all baselines in all scenes, delivering slightly better results, the highest rendering speed, and a moderate model size. We also found that, since 4DGS trains using the sparse point cloud from the 0th frame and limits the number of final points, its performance is significantly lower than other pipelines on texture-richer outdoor scenes, such as Scene6" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.907, + 0.117 + ], + "angle": 0, + "content": "Table 3. Test views performance of 3DGS-based dynamic scene reconstruction method on ImViD dataset." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.129, + 0.905, + 0.19 + ], + "angle": 0, + "content": "
MethodScene1 Opera_girlScene2 LaboratoryScene6 Puppy
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
4DGS23.2270.7530.41025.7980.8900.17618.1210.2220.711
4DRotor27.2630.7750.32828.0070.9170.09817.9160.2980.331
STG28.4820.7860.28726.3060.9100.11420.4970.5940.211
STG++31.2400.7990.27727.5810.9160.10720.5330.5970.202
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.205, + 0.907, + 0.233 + ], + "angle": 0, + "content": "Table 4. The storage size of training results from different methods and their rendering speed on a single A100 GPU. *CPU version." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.244, + 0.905, + 0.321 + ], + "angle": 0, + "content": "
Scene1 Opera_girlScene2 LaboratoryScene6 Puppy
Mem(MB)FPSMem(MB)FPSMem(MB)FPS
4DGS101.0546.22105.0244.93101.6541.86
4DRotor*5818.611.006867.180.674415.820.56
STG382.4196.10318.41133.651287.55100.00
STG++387.17108.89329.07137.781276.21110.47
" + }, + { + "type": "table_caption", + "bbox": [ + 0.555, + 0.334, + 0.866, + 0.348 + ], + "angle": 0, + "content": "Table 5. User study for the sound field construction." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.359, + 0.905, + 0.442 + ], + "angle": 0, + "content": "
Auditory Spatial PerceptionSound QualityImmersiveness
Very poor0.00%0.00%0.00%
Poor0.00%0.00%0.00%
Fair14.28%19.04%9.52%
Good23.80%33.33%42.85%
Excellent61.90%47.61%47.61%
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.467, + 0.907, + 0.543 + ], + "angle": 0, + "content": "puppy. Additionally, due to the density control in the time dimension implemented by 4Drotor, it performs better in frames and regions with more significant motion, which is an aspect worth referencing and adopting in future work. More results can be seen in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.55, + 0.906, + 0.626 + ], + "angle": 0, + "content": "Qualitative Results. As shown in Figure 6, Our STG++ achieves better temporal continuity in the color of pixels at the same location. There are almost no abrupt transitions both within and between segments. Details are showed in the supplementary for a more intuitive experience." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.636, + 0.809, + 0.651 + ], + "angle": 0, + "content": "5.2. Effectiveness of Sound Field Data" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.906, + 0.748 + ], + "angle": 0, + "content": "Quantitative Results. Since our sound field reconstruction algorithm is non-trainable, the reconstructed spatial audio lacks a corresponding ground truth, making it difficult to evaluate the results using quantitative metrics. Therefore, we conducted a user study with 21 experts to assess the performance of the sound field reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.885 + ], + "angle": 0, + "content": "Each participant rated the sound field reconstruction on three dimensions: auditory spatial perception, sound quality, and immersiveness. Auditory spatial perception refers to the listener's ability to perceive the distribution and localization of sound in space. Sound quality refers to the audio quality of the spatial audio compared to the corresponding microphone-recorded signal. Immersiveness refers to the listener's sense of being in a space where the sound source genuinely exists." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.886, + 0.907, + 0.901 + ], + "angle": 0, + "content": "The results of the user study are shown in Table 5, with" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16560" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.092, + 0.889, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.134, + 0.323, + 0.861, + 0.337 + ], + "angle": 0, + "content": "Figure 5. Comparison of the rendering results of four baselines on Scene 1 Opera, Scene 2 Laboratory, and Scene 6 Puppy." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.362, + 0.482, + 0.537 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.55, + 0.483, + 0.592 + ], + "angle": 0, + "content": "Figure 6. The continuity of pixels at the same location across different frames and segments. The visualizations present the RGB (middle) and brightness (right) variations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.627, + 0.483, + 0.732 + ], + "angle": 0, + "content": "numbers as percentages of participants. The majority of the participants \\((61.90\\%)\\) rated the auditory spatial perception as excellent, \\(80.94\\%\\) participants felt that the generated spatial audio did not exhibit significant degradation in quality, and \\(90.46\\%\\) found the audio immersive, which demonstrates the effectiveness of our data capture and sound field reconstruction methods." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.754, + 0.35, + 0.77 + ], + "angle": 0, + "content": "5.3. Multimodal VR Experiences." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We finally integrated light field and sound field reconstruction results to achieve a 6-DoF multimodal immersive VR experience with real-time rendering speed at 60 FPS on a single 3090 GPU. As shown in Figure 7, the visual and auditory experiences correspond uniquely to the user's position and viewing direction. The experience can be further explored by listening through headphones with the supplementary video provided." + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.365, + 0.883, + 0.584 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.596, + 0.907, + 0.639 + ], + "angle": 0, + "content": "Figure 7. Visualization of the interaction trajectory and corresponding visual&auditory results. Note that the multi-modal feedback changes consistently with the positions and orientations." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.664, + 0.634, + 0.68 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.689, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this work, we introduce ImViD, to our knowledge, the first dataset for immersive volumetric video. This dataset encompasses seven large indoor and outdoor scenes, featuring rich foreground and background content that showcases engaging and commonly encountered daily scenarios, along with high-quality synchronized audio sequences. We validated the effectiveness of the dataset using the latest dynamic light field reconstruction algorithms, providing both quantitative and qualitative analyses of current limitations and potential improvements. Moreover, we propose a base pipeline for constructing immersive volumetric video from this data. We believe ImViD will facilitate the exploration of the limitations of current volumetric video reconstruction algorithms and drive immersive VR/AR experiences." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16561" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.241, + 0.108 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.484, + 0.201 + ], + "angle": 0, + "content": "This work was supported by National Key R&D Program of China (No.2022YFF0902204), NSFC (No.62171255, 52425801), Tsinghua University-Migu Xinkong Culture Technology (Xiamen) Co.Ltd. Joint Research Center for Intelligent Light Field and Interaction Technology, Guoqiang Institute of Tsinghua University (No.2021GQG0001)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.226, + 0.188, + 0.243 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.251, + 0.484, + 0.294 + ], + "angle": 0, + "content": "[1] https://www.apple.com/newsroom/2024/07/new-apple-immersive-video-series-and-films-premiere-on-vision-pro/.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.295, + 0.34, + 0.308 + ], + "angle": 0, + "content": "[2] https://www.gracia.ai/.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.31, + 0.483, + 0.365 + ], + "angle": 0, + "content": "[3] Cal Armstrong, Lewis Thresh, Damian Murphy, and Gavin Kearney. A perceptual evaluation of individual and non-individual hrfts: A case study of the sadie ii database. Applied Sciences, 8(11):2029, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.367, + 0.484, + 0.449 + ], + "angle": 0, + "content": "[4] Benjamin Attal, Jia-Bin Huang, Christian Richardt, Michael Zollhoefer, Johannes Kopf, Matthew O'Toole, and Changil Kim. Hyperreel: High-fidelity 6-dof video with ray-conditioned sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16610-16620, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.451, + 0.483, + 0.506 + ], + "angle": 0, + "content": "[5] Swapnil Bhosale, Haosen Yang, Diptesh Kanojia, Jiankang Deng, and Xiatian Zhu. Av-gs: Learning material and geometry aware priors for novel view acoustic synthesis. arXiv preprint arXiv:2406.08920, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.508, + 0.484, + 0.588 + ], + "angle": 0, + "content": "[6] Michael Broxton, John Flynn, Ryan Overbeck, Daniel Erickson, Peter Hedman, Matthew Duvall, Jason Dourgarian, Jay Busch, Matt Whalen, and Paul Debevec. Immersive light field video with a layered mesh representation. ACM Transactions on Graphics (TOG), 39(4):86-1, 2020. 1, 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.591, + 0.484, + 0.645 + ], + "angle": 0, + "content": "[7] Amandine Brunetto, Sascha Hornauer, and Fabien Moutarde. Neraf: 3d scene infused neural radiance and acoustic fields. arXiv preprint arXiv:2405.18213, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.649, + 0.483, + 0.704 + ], + "angle": 0, + "content": "[8] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 130-141, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.705, + 0.484, + 0.773 + ], + "angle": 0, + "content": "[9] Hao Chen, Yuqi Hou, Chenyuan Qu, Irene Testini, Xiaohan Hong, and Jianbo Jiao. \\(360+\\mathrm{x}\\): A panoptic multimodal scene understanding dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19373-19382, 2024. 2, 3, 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.776, + 0.483, + 0.831 + ], + "angle": 0, + "content": "[10] Kai Cheng, Xiaoxiao Long, Kaizhi Yang, Yao Yao, Wei Yin, Yuexin Ma, Wenping Wang, and Xuejin Chen. Gaussianpro: 3d gaussian splatting with progressive propagation. arXiv preprint arXiv:2402.14650, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[11] Alvaro Collet, Ming Chuang, Pat Sweeney, Don Gillett, Dennis Evseev, David Calabrese, Hugues Hoppe, Adam Kirk, and Steve Sullivan. High-quality streamable free-viewpoint video. ACM Transactions on Graphics (ToG), 34(4):1-13, 2015. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.251, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[12] Yuanxing Duan, Fangyin Wei, Qiyu Dai, Yuhang He, Wenzheng Chen, and Baoquan Chen. 4d-rotor gaussian splatting: Towards efficient novel view synthesis for dynamic scenes. In ACM SIGGRAPH 2024 Conference Papers, pages 1–11, 2024. 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[13] Bardienus P Duisterhof, Zhao Mandi, Yunchao Yao, Jia-Wei Liu, Mike Zheng Shou, Shuran Song, and Jeffrey Ichnowski. Md-splating: Learning metric deformation from 4d gaussians in highly deformable scenes. arXiv preprint arXiv:2312.00583, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.233, + 0.905, + 0.313 + ], + "angle": 0, + "content": "[14] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12479–12488, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.316, + 0.905, + 0.37 + ], + "angle": 0, + "content": "[15] Hang Gao, Ruilong Li, Shubham Tulsiani, Bryan Russell, and Angjoo Kanazawa. Monocular dynamic view synthesis: A reality check. Advances in Neural Information Processing Systems, 35:33768-33780, 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.905, + 0.425 + ], + "angle": 0, + "content": "[16] Huiyu Gao, Jiahao Ma, David Ahmedt-Aristizabal, Chuong Nguyen, and Miaomiao Liu. Soaf: Scene occlusion-aware neural acoustic field. arXiv preprint arXiv:2407.02264, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.427, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[17] Zhiyang Guo, Wengang Zhou, Li Li, Min Wang, and Houqiang Li. Motion-aware 3d gaussian splatting for efficient dynamic scene reconstruction. arXiv preprint arXiv:2403.11447, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.484, + 0.905, + 0.552 + ], + "angle": 0, + "content": "[18] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE transactions on pattern analysis and machine intelligence, 36(7):1325-1339, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.554, + 0.905, + 0.636 + ], + "angle": 0, + "content": "[19] Hanbyul Joo, Tomas Simon, Xulong Li, Hao Liu, Lei Tan, Lin Gui, Sean Banerjee, Timothy Scott Godisart, Bart Nabbe, Iain Matthews, Takeo Kanade, Shohei Nobuhara, and Yaser Sheikh. Panoptic studio: A massively multiview system for social interaction capture. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017. 2, 3, 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.637, + 0.905, + 0.704 + ], + "angle": 0, + "content": "[20] HyunJun Jung, Nikolas Brasch, Jifei Song, Eduardo Perez-Pellitero, Yiren Zhou, Zhihao Li, Nassir Navab, and Benjamin Busam. Deformable 3d gaussian splatting for animatable human avatars. arXiv preprint arXiv:2312.15059, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.706, + 0.905, + 0.748 + ], + "angle": 0, + "content": "[21] Takeo Kanade, Peter Rander, and PJ Narayanan. Virtualized reality: Constructing virtual worlds from real scenes. IEEE multimedia, 4(1):34-47, 1997. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.749, + 0.905, + 0.789 + ], + "angle": 0, + "content": "[22] Kai Katsumata, Duc Minh Vo, and Hideki Nakayama. A compact dynamic 3d gaussian representation for real-time dynamic view synthesis, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[23] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuehler, and George Drettakis. 3D Gaussian Splatting for Real-Time Radiance Field Rendering. ACM Transactions on Graphics, 42(4):1-14, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[24] Agelos Kratimenos, Jiahui Lei, and Kostas Daniilidis. Dynmf: Neural motion factorization for real-time dynamic view synthesis with 3d gaussian splatting. arXiv preprint arXiv:2312.00112, 2023. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16562" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[25] Joo Chan Lee, Daniel Rho, Xiangyu Sun, Jong Hwan Ko, and Eunbyung Park. Compact 3d gaussian splatting for static and dynamic radiance fields. arXiv preprint arXiv:2408.03822, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[26] Deqi Li, Shi-Sheng Huang, Zhiyuan Lu, Xinran Duan, and Hua Huang. St-4dgs: Spatial-temporally consistent 4d gaussian splatting for efficient dynamic scene rendering. In ACM SIGGRAPH 2024 Conference Papers, pages 1–11, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[27] Lingzhi Li, Zhen Shen, Zhongshu Wang, Li Shen, and Ping Tan. Streaming radiance fields for 3d video synthesis. Advances in Neural Information Processing Systems, 35: 13485-13498, 2022. 2, 3, 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.482, + 0.358 + ], + "angle": 0, + "content": "[28] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5521-5531, 2022. 2, 3, 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.482, + 0.416 + ], + "angle": 0, + "content": "[29] Yanyan Li, Chenyu Lyu, Yan Di, Guangyao Zhai, Gim Hee Lee, and Federico Tombari. Geogaussian: Geometry-aware gaussian splatting for scene rendering. arXiv preprint arXiv:2403.11324, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.419, + 0.482, + 0.487 + ], + "angle": 0, + "content": "[30] Zhan Li, Zhang Chen, Zhong Li, and Yi Xu. Spacetime gaussian feature splatting for real-time dynamic view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8508-8520, 2024. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.482, + 0.544 + ], + "angle": 0, + "content": "[31] Susan Liang, Chao Huang, Yapeng Tian, Anurag Kumar, and Chenliang Xu. Av-nerf: Learning neural fields for real-world audio-visual scene synthesis. Advances in Neural Information Processing Systems, 36:37472-37490, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.547, + 0.482, + 0.613 + ], + "angle": 0, + "content": "[32] Yiqing Liang, Numair Khan, Zhengqin Li, Thu Nguyen-Phuoc, Douglas Lanman, James Tompkin, and Lei Xiao. Gaufre: Gaussian deformation fields for real-time dynamic novel view synthesis. arXiv preprint arXiv:2312.11458, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.617, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[33] Haotong Lin, Sida Peng, Zhen Xu, Yunzhi Yan, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Efficient neural radiance fields for interactive free-viewpoint video. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 2, 3, 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.675, + 0.482, + 0.741 + ], + "angle": 0, + "content": "[34] Kai-En Lin, Lei Xiao, Feng Liu, Guowei Yang, and Ravi Ramamoorthi. Deep 3d mask volume for view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1749–1758, 2021. 2, 3, 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.745, + 0.482, + 0.813 + ], + "angle": 0, + "content": "[35] Youtian Lin, Zuozhuo Dai, Siyu Zhu, and Yao Yao. Gaussian-flow: 4d reconstruction with dynamic 3d gaussian particle. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21136-21145, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.816, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[36] Qingming Liu, Yuan Liu, Jiepeng Wang, Xianqiang Lv, Peng Wang, Wenping Wang, and Junhui Hou. Modgs: Dynamic gaussian splatting from causally-captured monocular videos. arXiv preprint arXiv:2406.00434, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[37] Cheng-You Lu, Peisen Zhou, Angela Xing, Chandradeep Pokhariya, Arnab Dey, Ishaan Nikhil Shah, Rugved Mavidi" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.092, + 0.905, + 0.16 + ], + "angle": 0, + "content": "palli, Dylan Hu, Andrew I Comport, Kefan Chen, et al. Diva360: The dynamic visual dataset for immersive neural fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22466-22476, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[38] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. arXiv preprint arXiv:2312.00109, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[39] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[40] Dawid Malarz, Weronika Smolak, Jacek Tabor, Sławomir Tadeja, and Przemysław Spurek. Gaussian splatting with nef-based color and opacity. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.905, + 0.388 + ], + "angle": 0, + "content": "[41] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.392, + 0.905, + 0.461 + ], + "angle": 0, + "content": "[42] David Narciso, Maximino Bessa, Miguel Melo, Antonio Coelho, and José Vasconcelos-Raposo. Immersive \\(360^{\\circ}\\) video user experience: impact of different variables in the sense of presence and cybersickness. Universal Access in the Information Society, 18:77-87, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.463, + 0.905, + 0.545 + ], + "angle": 0, + "content": "[43] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228, 2021. 2, 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.905, + 0.616 + ], + "angle": 0, + "content": "[44] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In CVPR, 2021. 2, 3, 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.619, + 0.905, + 0.687 + ], + "angle": 0, + "content": "[45] Weining Ren, Zihan Zhu, Boyang Sun, Jiaqi Chen, Marc Pollefeys, and Songyou Peng. Nerf on-the-go: Exploiting uncertainty for distractor-free nerfs in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8931-8940, 2024. 2, 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.69, + 0.905, + 0.772 + ], + "angle": 0, + "content": "[46] Neus Sabater, Guillaume Boisson, Benoit Vandame, Paul Kerbiriou, Frederic Babon, Matthieu Hog, Remy Gendrot, Tristan Langlois, Olivier Bureller, Arno Schubert, et al. Dataset and pipeline for multi-view light-field video. In Proceedings of the IEEE conference on computer vision and pattern recognition Workshops, pages 30-40, 2017. 2, 3, 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.775, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[47] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[48] Ruizhi Shao, Zerong Zheng, Hanzhang Tu, Boring Liu, Hongwen Zhang, and Yebin Liu. Tensor4d: Efficient neural 4d decomposition for high-fidelity dynamic reconstruction and rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16632-16642, 2023. 2, 3, 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16563" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.093, + 0.482, + 0.176 + ], + "angle": 0, + "content": "[49] Roman Shapovalov, Yanir Kleiman, Ignacio Rocco, David Novotny, Andrea Vedaldi, Changan Chen, Filippos Kokkinos, Ben Graham, and Natalia Neverova. Replay: Multi-modal multi-view acted videos for casual holography. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20338-20348, 2023. 2, 3, 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.178, + 0.483, + 0.234 + ], + "angle": 0, + "content": "[50] Richard Shaw, Michal Nazarczuk, Jifei Song, Arthur Moreau, Sibi Catley-Chandar, Helisa Dhamo, and Eduardo Perez-Pellitero. Swings: Sliding windows for dynamic 3d gaussian splatting, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.235, + 0.483, + 0.317 + ], + "angle": 0, + "content": "[51] Liangchen Song, Anpei Chen, Zhong Li, Zhang Chen, Lele Chen, Junsong Yuan, Yi Xu, and Andreas Geiger. Nerfplayer: A streamable dynamic scene representation with decomposed neural radiance fields. IEEE Transactions on Visualization and Computer Graphics, 29(5):2732-2742, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.319, + 0.483, + 0.401 + ], + "angle": 0, + "content": "[52] Jiakai Sun, Han Jiao, Guangyuan Li, Zhanjie Zhang, Lei Zhao, and Wei Xing. 3dgstream: On-the-fly training of 3d gaussians for efficient streaming of photo-realistic free-viewpoint videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20675-20685, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.404, + 0.483, + 0.473 + ], + "angle": 0, + "content": "[53] Feng Wang, Sinan Tan, Xinghang Li, Zeyue Tian, Yafei Song, and Huaping Liu. Mixed neural voxels for fast multiview video synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19706-19716, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.476, + 0.483, + 0.531 + ], + "angle": 0, + "content": "[54] Feng Wang, Zilong Chen, Guokang Wang, Yafei Song, and Huaping Liu. Masked space-time hash encoding for efficient dynamic scene reconstruction. Advances in Neural Information Processing Systems, 36, 2024. 2, 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.533, + 0.483, + 0.615 + ], + "angle": 0, + "content": "[55] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20310-20320, 2024. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.617, + 0.483, + 0.685 + ], + "angle": 0, + "content": "[56] Tianhao Wu, Fangcheng Zhong, Andrea Tagliasacchi, Forrester Cole, and Cengiz Oztireli. D^2nerf: Self-supervised decoupling of dynamic and static objects from a monocular video. Advances in neural information processing systems, 35:32653-32666, 2022. 2, 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.688, + 0.483, + 0.744 + ], + "angle": 0, + "content": "[57] Yunzhi Yan, Haotong Lin, Chenxu Zhou, Weijie Wang, Haiyang Sun, Kun Zhan, Xianpeng Lang, Xiaowei Zhou, and Sida Peng. Street gaussians for modeling dynamic urban scenes. arXiv preprint arXiv:2401.01339, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.746, + 0.483, + 0.8 + ], + "angle": 0, + "content": "[58] Zeyu Yang, Hongye Yang, Zijie Pan, Xiatian Zhu, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. arXiv preprint arXiv:2310.10642, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.802, + 0.483, + 0.87 + ], + "angle": 0, + "content": "[59] Ziyi Yang, Xinyu Gao, Yangtian Sun, Yihua Huang, Xiaoyang Lyu, Wen Zhou, Shaohui Jiao, Xiaojuan Qi, and Xiaogang Jin. Spec-gaussian: Anisotropic view-dependent appearance for 3d gaussian splatting. arXiv preprint arXiv:2402.15870, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[60] Jae Shin Yoon, Kihwan Kim, Orazio Gallo, Hyun Soo Park, and Jan Kautz. Novel view synthesis of dynamic scenes with" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "globally coherent depths from a monocular camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5336-5345, 2020. 2, 3, 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.192 + ], + "angle": 0, + "content": "[61] Dongbin Zhang, Chuming Wang, Weitao Wang, Peihao Li, Minghan Qin, and Haoqian Wang. Gaussian in the wild: 3d gaussian splatting for unconstrained image collections. arXiv preprint arXiv:2403.15704, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16564" + } + ] +] \ No newline at end of file diff --git a/2025/ImViD_ Immersive Volumetric Videos for Enhanced VR Engagement/6da376dc-6436-436c-b26c-23ba201328ef_origin.pdf b/2025/ImViD_ Immersive Volumetric Videos for Enhanced VR Engagement/6da376dc-6436-436c-b26c-23ba201328ef_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d9fff9f859968e97f85a34aafbbadf14bad60f6f --- /dev/null +++ b/2025/ImViD_ Immersive Volumetric Videos for Enhanced VR Engagement/6da376dc-6436-436c-b26c-23ba201328ef_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2364dca07dda67fe8a4aef9c69bbab534aab2fd172d94ed846feba684e60691c +size 9523846 diff --git a/2025/ImViD_ Immersive Volumetric Videos for Enhanced VR Engagement/full.md b/2025/ImViD_ Immersive Volumetric Videos for Enhanced VR Engagement/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e285c29432b24b2f9d7ab7e3d33ce2b55b4c99c2 --- /dev/null +++ b/2025/ImViD_ Immersive Volumetric Videos for Enhanced VR Engagement/full.md @@ -0,0 +1,336 @@ +# ImViD: Immersive Volumetric Videos for Enhanced VR Engagement + +Zhengxian Yang $^{1*}$ , Shi Pan $^{1*}$ , Shengqi Wang $^{1*}$ , Haoxiang Wang $^{1}$ , Li Lin $^{2}$ , Guanjun Li $^{3}$ , Zhengqi Wen $^{1\dagger}$ , Borong Lin $^{1\dagger}$ , Jianhua Tao $^{1\dagger}$ , Tao Yu $^{1\dagger}$ , Tsinghua University ${}^{2}$ Migu Beijing Research Institute ${}^{3}$ Institute of Automation, Chinese Academy of Science $\{\mathrm{zx - yang23, ps23, shengqi - 21, whx22}\} @$ mails.tsinghua.edu.cn lilin@migu.chinamobile.com guanjun.li@nlpr ia.ac.cn $\{\mathrm{zqwen, linbr, jhtao, ytrock}\} @$ tsinghua.edu.cn + +![](images/29df6dfe8c358de09148a8c338fa2c40e7df876d550543ebbb5b53dcecaae38c.jpg) +Figure 1. We introduce ImViD, a dataset for immersive volumetric videos. ImViD records dynamic scenes using a multi-view audio-video capture rig moving in a space-oriented manner, which provides a new benchmark for volumetric video reconstruction and its application. + +![](images/54cea3e825ce7d714e26bf4ae2061d89768861e2c9285a1234747496b4cf9a61.jpg) + +![](images/7620de647f0ca8575e9ca85febef46564a18a61207b78bd20144f82cf7d2d859.jpg) + +# Abstract + +User engagement is greatly enhanced by fully immersive multi-modal experiences that combine visual and auditory stimuli. Consequently, the next frontier in VR/AR technologies lies in immersive volumetric videos with complete scene capture, large 6-DoF interaction space, multimodal feedback, and high resolution&frame-rate contents. To stimulate the reconstruction of immersive volumetric videos, we introduce ImViD, a multi-view, multi-modal dataset featuring complete space-oriented data capture and various indoor/outdoor scenarios. Our capture rig supports multi-view video-audio capture while on the move, a capability absent in existing datasets, significantly enhancing the completeness, flexibility, and efficiency of data capture. + +The captured multi-view videos (with synchronized audi- 15 minutes, and include rich foreground-background elements, and complex dynamics. We benchmark existing methods using our dataset and establish a base pipeline for construct + +ing immersive volumetric videos from multi-view audiovisual inputs for 6-DoF multi-modal immersive VR experiences. The benchmark and the reconstruction and interaction results demonstrate the effectiveness of our dataset and baseline method, which we believe will stimulate future research on immersive volumetric video production. Project Page: https://yzxqh.github.io/ImViD/ + +# 1. Introduction + +High-quality and high-fidelity digital modeling of the real world has long been a topic of significant interest. With breakthroughs in 3D reconstruction algorithms and virtual reality technology, the generation of such visually compelling content is gradually becoming feasible, whether for reproducing past events and memories or remotely viewing current occurrences. Achieving complete user immersion is a goal pursued by many researchers. + +What defines an immersive volumetric video? + +Dwelling on previous tests [1, 6] and user feedback [2, 42], we believe that immersive media should have the following four characteristics: i) Full $360^{\circ}$ foreground and background, ii) High-quality 6-DoF interaction, which current volumetric video can achieve, iii) Multimodal experi + +ence with light and sound, and iv) High frame rate and long-duration dynamic content. We refer to a media format that meets all these criteria as immersive volumetric video. + +However, with the development of camera equipment, despite many efforts using monocular handheld cameras or fixed camera arrays to build datasets, none of these solutions fully address all four requirements. For example, Panoptic Sports [19], ZJU-Mocap [44] and DiVa-360 [37] achieved human/object-centric $360^{\circ}$ capture but lack complex backgrounds and multimodal data; Google's immersive light field [6] achieves 6-DoF volumetric video in an inside-looking-out manner but lacks multimodality and high frame rate, long-duration dynamic content; Replay [49] introduces sound, but the camera setup is not suitable for human viewing habits thus cannot achieve high-quality 6-DoF interaction. Both academia and industry require an appropriate capture method to support the development of immersive volumetric video technology. + +Additionally, NeRF [41]'s emergence has increased attention to reconstruction and rendering techniques, making it easier to generate high-fidelity interactive 3D scenes from multiple views. 3DGS [23] and its subsequent improvements [10, 29, 38, 40, 57, 59, 61] have further boosted rendering speed and quality, fueling peoples' expectations for rapid real-world modeling and its applications in industries such as education, healthcare, and entertainment. This can be seen as an evolutionary opportunity for immersive volumetric video. However, current technologies face challenges. In dynamic light field reconstruction, neither explicit [22, 24, 30, 35, 55] nor implicit neural representation based [4, 8, 14, 28, 48, 51, 53] methods have been able to achieve a practical balance between foreground-background reconstruction accuracy, rendering speed, spatiotemporal consistency, and storage efficiency in dynamic scenes; In the realm of sound field reconstruction, it is typically necessary to collect a large amount of binaural audio or room impulse response data from different locations for model training, which is user-unfriendly in practical use. + +To better support the development of immersive volumetric video technology, we have developed a moving capture rig (with 46 synchronized cameras, as shown in Fig.3) and designed two space-oriented data capture strategies. We have constructed an audiovisual dataset containing 7 indoor and outdoor scenes, with a $360^{\circ}$ field of view, large 6-DoF interaction space, and high-resolution, long-duration dynamic contents. We also validate this dataset using the latest novel view synthesis and 3D reconstruction methods. Finally, a complete pipeline for constructing light and sound fields from this type of data was proposed, which initially achieved immersive volume video. + +Our main contributions can be summarized as follows: + +- We propose to our knowledge the first and probably the highest-quality multi-view dataset ImViD for immersive + +volumetric video. It features $360^{\circ}$ foreground and background capture, long-duration video&audio data under high resolution@5K, high frame rate@60FPS. + +- We conduct extensive experiments to validate the effectiveness of the proposed datasets by evaluating and analyzing the performance of several current baselines. +- Follow the proposed pipeline shown in Figure 2, We finally achieve the reconstruction of immersive volumetric video and integrate it for immersive VR experience. + +# 2. Related Work + +Datasets for Dynamic Novel View Synthesis. The datasets for dynamic scene reconstruction can be classified as monocular-based and multi-view-based. + +Monocular acquisition systems are popular due to their low cost and ease of construction. such as HyperNeRF [43], Dynamic Scene Dataset [60], and D2NeRF [56]. However, these datasets suffer from low resolution, limited capture space, and the limited durations. Although NeRF On-the-go [45] allows for larger capture ranges by walking while shooting, high-quality reconstructions are confined to the capture path, and the small field of view (FOV) limits prolonged observations of specific scene positions. + +Multi-camera data collection has gained significant attention due to its ability to provide a larger FOV and richer details, such as Immersive Light Field dataset [6], Technicolor [46], UCSD Dynamic Scene Dataset [34] and The Plenoptic Dataset [28], followed by [27, 33, 54]. However, all these setups remain static during capture, limiting them to frontal views and hindering $360^{\circ}$ reconstruction. Additionally, the video sequences are typically short, with a maximum duration of 2 minutes (often less than 30 seconds) and a maximum resolution of 4K, which is insufficient for immersive VR experiences. + +Moreover, the previously mentioned datasets lack sound recordings, leading to the loss of immersion without multimodality. The Replay dataset [49] addresses this by focusing on long sequences with professional actors in familiar settings. However, aside from the head-mounted cameras that enables rotation with head, all other cameras remain static. Furthermore, the DSLR arrangement does not align with human viewing habits in VR, making them unsuitable as benchmarks for novel view synthesis tasks. Recent work [9] presents a dataset with a $360^{\circ}$ camera and audio capture device but constrained by a fixed-point shooting strategy, resulting in sparse viewpoints that hinder the reconstruction of high-quality dynamic scenes. More details about datasets can be found in Tab. 1. + +Dynamic Light&Sound Field Reconstruction. Temporal variations in input images make reconstructing dynamic scenes more challenging than static ones. The key issue is efficiently representing 4D scenes while ensuring spa + +Table 1. Existing real-world datasets for dynamic novel view synthesis (full version in the supplementary materials). + +
DatasetsNo SceneOutdoor/IndoorCamerasMobilityResolutionAnglesDurationFPSMultimodalityContent
PanopticSports [19]65Indoor480 camerasStatic640×480360°5mins25XHuman-centric actions
Technicolor [46]12Indoor16 camerasStatic2048×1088Frontal2s30XHas a number of close-ups sequences, captured medium angle scenes and other animated scenes where the movement does not come from a human
Immersive-Lightfield [6]15both46 camerasStatic2560×1920Frontal10-30s30XSimple and slow motion of human, animals, objects
Dynamic Scene Datasets (NVIDIA) [60]8Outdoor1 Mobile phone /12 camerasFixed-point Waving /Static1920×1080Frontal5s60XSimple body motions (facial, jump, etc)
UCSD Dynamic [34]96Outdoor10 camerasStatic1920×1080Frontal1-2mins120XVarious visual effects and human interactions
ZJU-Mocap [44]10Indoor21 camerasStatic1024×1024360°20s50XSimple body motions (punch, kick, etc.)
Plenoptic Dataset (DyN-eRF/Neural 3D) [28]6Indoor21 camerasStatic2704×2028Frontal10-30s30XContains high specularity, translucency and transparency objects, motions with changing topology, selfcast moving shadows, various lighting conditions and multiple people moving around in open living room space
iPhone Datasets [15]14both1 hand-held phone /2 camerasFixed-point Waving /Static640×480Frontal8-15s30/60XFeaturing non-repetitive motion, from various categories such as generic objects, humans, and pets
ENeRF-Outdoor [33]4Outdoor18 camerasStatic1920×1080Frontal20-40s30XComplex human motions
Replay [49]46Indoor12 camerasStatic3840×2160360°5mins30✓(Audio)Dancing, chatting, playing video games, unwrapping presents, playing ping pong
360+X [9]28both1 360°cameras and 1 Spectacles camerasStatic5760×2880360°10s (2152 sequence)30✓(Audio)Capture in 17 cities across 5 countries. Panoptic perspective to scene understanding with audio
Im-ViD(Ours)7both46 camerasMoveable5312×2988Frontal and 360°1-5mins60✓(Audio)Seven common indoor and outdoor scenes in daily life, including opera, face-to-face communication, teaching, discussion, music performance, interaction with pets, and playing. Each scene has high-quality synchronized multi-view video and audio with a duration of more than 1 minute, and contains rich elements such as various small objects, glass, and changes in light and shadow
+ +tiotemporal consistency and high accuracy. Traditional methods estimate varying geometries like surfaces [11] and depth maps [21] to handle dynamics. Recently, efforts have shifted towards adapting neural radiance field representations for dynamic scenes. DyNeRF [28] pioneer in the dynamic reconstruction with Neural radiance field. Followed works such as [4, 8, 14, 27, 48, 51, 53] make efforts to improve the high-fidelity and efficiency in the 4D neural representation. However, the common challenges these methods face are memory cost and the modeling of spatiotemporal complexity. The latest work have extended 3DGS to dynamic scenes for high-quality real-time rendering. While modeling dynamic scenes as static ones is straightforward, methods relying solely on monocular data [20, 32] is not suitable for this paradigm, and face challenges of adding constraints or priors to better supervise training. Other works support multi-view data as input. [39] and [52] initialize each frame from the previous one to build Gaussian point clouds per-frame, leading to high memory and suddenly appearing or disappearing artifacts. 4DGS [55] introduces Hexplane to encode spatiotemporal information, using MLPs to predict Gaussian properties change. Subsequent works [13, 17, 26] further add geometric and optical flow constraints to improve frame-to-frame motion but perform well only on short sequences with small actions. Other approaches [22, 24, 30, 35] employ the explicit parametric form (such as radial basis functions) of 3DGS to enhance spatiotemporal consistency. Some methods [12, 58] focus on representation with dual quaternions, 4D spheri + +cal harmonics, or rotors for end-to-end training. More improvements such as parallel training [50] and using codebooks and masks to enhance scene compactness [25], show the potential in efficient and high-quality representation of long-term complex dynamic scenes. + +In addition to light field reconstruction, reconstructing the sound field from multimodal data is equally important, often referred to as novel-view acoustic synthesis. For the novel-view acoustic synthesis task, AV-NeRF [31] synthesizes spatial audio by leveraging visual features from rendered images. NeRAF [7] further decouples the camera and microphone poses used for training. AV-GS [5] incorporates a 3D scene representation during sound field synthesis, addressing the issue where previous methods overlooked the contribution of the broader 3D scene geometry. SOAF [16] considers occlusion in the light field when synthesizing the sound field. However, the above novel-view acoustic synthesis tasks typically require a large amount of multimodal training data collected from different locations, which is often user-unfriendly in practical applications. Moreover, these methods generally cannot solve the problem of sound field synthesis for moving sound sources. This paper proposes a training-free novel-view acoustic synthesis approach for our dataset, which also supports sound field synthesis for moving sound sources. + +# 3. ImViD Dataset + +Our goal is to build an immersive volumetric video dataset of real world that integrates both foreground and back- + +![](images/fc1cdf68e993250542181382f874bc20780f5c2496c601e9da23071faafea384.jpg) +Figure 2. The pipeline to realize the multimodal 6-DoF immersive VR experiences. We applied a carefully designed rig to (a) simultaneously capture multi-view video and audio. The (b1) presents our reconstruction of dynamic light field based on STG [30] while (b2) demonstrates the construction process of sound field. We have achieved better results than the original algorithm in long-term dynamic scenes by incorporating affine color transformation and t-dimensional density control. Ultimately, we achieve a 6-DoF immersive experience in both light and sound fields, and also benchmark on recent representative methods like 4DGS [55] and 4Drotor [12] to demonstrate the effectiveness of both our dataset and baseline method. + +ground, providing assistance for the development of related research such as spatial video and VR/AR application. Therefore, we have meticulously refined every aspect of the process, from equipment setup and scene selection to shooting strategies and data processing, ensuring maintaining the elements of interest to researchers to the greatest extent. The released ImViD dataset comprises a total of 7 large scenes, including 5 indoor and 2 outdoor environments. The content features a diverse range of common activities, such as opera, meeting, teaching, and playing on the grass. Each scene contains a set of static photos, multi-view time-synchronized video&audio sequences at 5K resolution and 60 FPS, employing two different collection strategies, along with calibrated camera poses. + +# 3.1. Data Acquisition + +The handheld monocular camera is easy to move, providing limited perspectives from various locations. In contrast, the fixed camera arrays, while stationary, offers dense perspectives within a limited range. We aim to combine the advantages of both to design an effective data collection system and strategy for fully immersive VR experience—a feature not found in existing datasets. As shown in Figure 3, our capture system consists of more than 40 GoPro cameras in + +stalled on a hemispherical surface mounted on a remotely controllable mobile car, all of which are connected to a PC to receive synchronous control commands. The system is carefully designed to be at a height similar to that of a human, and the camera arrangement better aligns with human viewing habits compared to other existing devices. This rig can synchronously collect about 1,000 images at 5K resolution within a space of at least $6\mathrm{m}^3$ in 2 min. It is also capable of synchronously capturing video and audio at 60 FPS and 5K resolution for 30 min (limited by heat dissipation). Thus, our capture strategy is as follows: + +Step1: Efficient high-quality, high-density $360^{\circ}$ image acquisition of static environments. + +Step2: Continuous synchronous collection of dynamic scenes under two different strategies. + +i) Fixed-point shooting. The cart remains stationary during the capture, focusing on capturing dynamic details. +ii) Moving shooting. The cart slowly moves while the scene is in motion, providing additional level of details and larger exploratory space. + +Throughout the capture process, we ensure strict synchronization of the world time codes across all cameras. We carefully maintained consistent focal lengths, exposure, + +![](images/93ee5eea297500e1b2858df8345906a2384b52ee47b14b1c1729d23290d6a3b0.jpg) +Figure 3. Our rig support two kinds of capturing strategies for high resolution, high frame rate and $360^{\circ}$ dynamic data acquisition. + +![](images/787b9f96ad10d46945cdaa06015c5884217d06c098fcd13a5f97cef15e89b9eb.jpg) +Figure 4. Calculation method for spatiotemporal capture density. 1) the capture strategy of handheld monocular camera 2) represents the fixed camera array 3) Our rig covers a volume of $0.6 \times \pi \times 0.5^2 \, \mathrm{m}^3$ over 5 seconds. + +and white balance settings for each scene, based on the results of manual iterative testing of camera settings, thereby minimizing hardware-induced variations in data quality. To achieve high-quality audio, we increased the noise reduction coefficient of the cameras and minimized noise generation as much as possible. + +# 3.2. Data Processing + +For the photos from Step1, we completed the calibration in COLMAP [47] by combining the cameras' prior parameters and their relative positions. For the video&audio from Step2, we first aligned the multi-view sequences for each scene using the world time codes. This step is crucial as the quality of the alignment directly determines the quality of dynamic scene reconstruction and significantly impacts whether the multi-view audio can be used for constructing a 6-DoF sound field. Our time code alignment error is at the level of approximately 2 milliseconds. Subsequently, we performed frame extraction from the videos, again utilizing the GoPro camera parameters as priors to complete the calibration. + +It is important to note that for the dynamic scene data captured with a moving rig, we also performed time code alignment. However, this data poses significant challenges for existing calibration methods, often resulting in errors and floaters. There are also no solutions for reconstruction algorithms specifically addressing dynamic scenes from moving rig currently. Therefore, we do not provide the calibrated results for this type of data at this time. But we believe that this data will greatly contribute to the advancement of the field, and thus we will also publicly releasing the data which has been time-synchronized. + +# 3.3. Dataset Analysis + +Diversity. Our dataset, as depicted in Table 2, includes 7 common indoor and outdoor scenes, each comprising one or more takes, totaling 16 video sequences. Each scene is + +![](images/db07f5f4f664c27676e09bbac34031a49fb20ca9a54d32481c233c0850fb1735.jpg) + +rich in foreground and background elements, meticulously crafted to preserve the authenticity of the environment. The characters, clothing, and motions within each scene exhibit a diverse range, set against varying ambient lighting conditions. We also included areas of interest like tables, chairs, windows as much as possible for researchers. + +Quality. Leveraging a total of 39 cameras, we captured dynamic scenes at a resolution of $5312 \times 2988$ and 60 FPS. The images of static environment is $5568 \times 4176$ . + +Duration. As outlined in Table 2, each scene in our dataset contains at least 1 minute of footage. The longest sequence in our forthcoming public dataset spans approximately 5 minutes. The total processed data duration amounts to 38 minutes 46 seconds, including 139,560 frames per camera. + +Coverage. As highlighted in Table 2, we conducted image acquisition from over 1,000 viewpoints before recording dynamic content, thereby offering a $180^{\circ} - 360^{\circ}$ viewing space in VR. For scenes employing the mobile shooting strategy, we introduced a metric termed 'spatiotemporal capture density,' quantifying the captured volume covered by our rig per second, as shown in Figure 4. This metric, unavailable in current datasets (where static camera arrays are deemed 0 and handheld monocular cameras follow a single trajectory without capturing volume), registers at approximately $0.10~m^3 /s$ in our dataset. + +# 4. Benchmarks and Experiments + +To comprehensively assess the effectiveness of the dataset, we selected three scenes (2 indoor and 1 outdoor) to test the latest dynamic light field reconstruction algorithms, providing a benchmark for research in areas such as novel view synthesis. We also proposed an improvement strategy based on one of the methods. Furthermore, we presented and validated a feasible pipeline for constructing 6-DoF sound fields from multi-view audio sequences. In this section, we + +Table 2. The detailed statistics of our ImViD dataset. #Take represents the number of different episodes we collected in this scene. #Strategy represents the strategies used in Step2. 1 indicates only (i) was used in Step 2, 2 indicates that both were used. Avg.S-T Density indicates the spatiotemporal coverage density of our collection method, measured by the captured volume size per second. + +
Name#Camera#Static viewpoints#Take#StrategyAvg.S-T Density(m3/s)Viewing SpaceAvg.DurationStorage(GB)
Scene1 Opera39115221-180°3min22s226
Scene2 Laboratory391225220.10360°1min42s137.3
Scene3 Classroom391223220.10360°4min42s497
Scene4 Meeting39122311-360°3min16s114
Scene5 Rendition391620420.10360°2min02s516
Scene6 Puppy391404320.10360°1min50s359
Scene7 Playing391224220.10360°1min10s220
Total--16---38min46s2069.3
+ +will first introduce the selected light field pipeline and our proposed sound field pipeline (Sec 4.1), then provide details on the experimental setup (Sec 4.2). + +# 4.1. Baseline + +Dynamic Light Field Reconstruction. We conducted a thorough survey of the current dynamic reconstruction algorithms, prioritizing rendering speed and support for multiview data input. We finally decided to establish a benchmark on 3DGS-based method from three paradigms. + +4DGS [55] introduces Hexplane to encode spatiotemporal information to feature $f_{d}$ and use MLPs decoders $D = \{\phi_{x},\phi_{r},\phi_{s}\}$ to predict gaussians' changes in position $\Delta \chi = \phi_{x}(f_{d})$ , rotation $\Delta r = \phi_r(f_d)$ , and scaling $\Delta s = \phi_s(f_d)$ . Utilizing the sparse point cloud from frame 0 as initialization and training static gaussians for the first 3000 iterations, the dynamic scene $G$ is represented as: + +$$ +G \left(\chi^ {\prime}, r ^ {\prime}, s ^ {\prime}\right) = G \left(\chi + \Delta \chi , r + \Delta r, s + \Delta s\right) \tag {1} +$$ + +4DrotorGS [12] is a representative method of dimensionality expansion. It extends 3DGS to a four-dimensional Gaussian ellipsoid by introducing the concept of rotor, expanding the covariance matrix to four dimensions, i.e.: + +$$ +G _ {4 D} (x) = e ^ {- \frac {1}{2} (x - \mu_ {4 D}) ^ {T} \Sigma_ {4 D} ^ {- 1} (x - \mu_ {4 D})} \tag {2} +$$ + +Each frame of the scene corresponds to a time slice of the 4D Gaussian ellipsoids followed by $\alpha$ -blending. + +STG [30], on the other hand, use polynomial fitting for the motion and rotation of the 3DGS, and employs an RBF function to model the opacity changes over time. It concatenates the sparse point clouds from all frames as the initialization, and the final representation of each point in the scene is: + +$$ +\alpha_ {i} (t) = \sigma_ {i} (t) \exp \left(- \frac {1}{2} \left(x - \mu_ {i} (t)\right) ^ {T} \Sigma_ {i} (t) ^ {- 1} \left(x - \mu_ {i} (t)\right)\right) \tag {3} +$$ + +Although we ensured consistent white balance and other settings across cameras based on the characteristics of scenes during shooting, color differences still exist due to + +factors like occlusions when capturing from different angles. During experiments, we found that while existing methods fit well to the training views, there is a noticeable color shift during viewpoint transitions in the reconstructed light field, manifesting as flickering and floaters. This is especially apparent when training in segments, where the differences between segments are more pronounced, significantly reducing immersion. We improved the original STG (named as $\mathrm{STG}++$ ) by introducing global, time-invariant color mapping during training. Detailed implementation can be seen in supplementary materials. + +Sound Field Reconstruction. Taking the recording microphone as the origin of the coordinate system, we denote the coordinates of the sound source as $s(t) = \{x_s(t), y_s(t)\}$ , the user wearing the VR device as $l(t) = \{x_l(t), y_l(t)\}$ , and the angle of deviation from the y-axis in the counterclockwise direction as $\theta_l(t)$ , where $t$ is the time index. Based on these premises, generating the user's binaural spatial audio can be divided into three parts: sound direction mapping, sound distance mapping, and sound auralization. Note that we assume there is only one dominating sound source (omnidirectional emitting) to simplify the sound field reconstruction. + +In the sound direction mapping part, we calculate the direction of the sound source relative to the user's ears, $\theta_{s}(t)$ + +$$ +\theta_ {s} (t) = \operatorname {a r c c o s} \frac {\mathbf {v} _ {1} (t) \mathbf {v} _ {2} (t)}{| \mathbf {v} _ {1} (t) | | \mathbf {v} _ {2} (t) |}, \tag {4} +$$ + +where + +$$ +\mathbf {v} _ {1} (t) = \left[ x _ {s} (t) - x _ {l} (t), y _ {s} (t) - y _ {l} (t) \right] ^ {T}, \tag {5} +$$ + +$$ +\mathbf {v} _ {2} (t) = \left[ - \sin \left(\theta_ {l} (t)\right), \cos \left(\theta_ {l} (t)\right) \right] ^ {T}, \tag {6} +$$ + +In the sound distance mapping part, we calculate the scaling of the sound source when it reaches the user's ears, relative to the original recording. We denote this scaling parameter as $\lambda(t)$ , + +$$ +\lambda (t) = \frac {\sqrt {x _ {s} ^ {2} (t) + y _ {s} ^ {2} (t)}}{\sqrt {(x _ {s} (t) - x _ {l} (t)) ^ {2} + (y _ {s} (t) - y _ {l} (t)) ^ {2}}} \tag {7} +$$ + +In the sound auralization part, we convert the original recording into spatial audio based on $\theta_s(t)$ and $\lambda (t)$ . In the short-time Fourier transformation (STFT) domain, we denote the audio from the left and right ears as $A_{L}(t,f)$ and $A_{R}(t,f)$ , respectively, where $f$ is the frequency index. Then $A_{L}(t,f)$ and $A_{R}(t,f)$ can be written as + +$$ +A _ {L} (t, f) = F _ {L} \left(\theta_ {s} (t)\right) F _ {l} (t, f) A _ {O} (t, f) \tag {8} +$$ + +$$ +A _ {R} (t, f) = F _ {R} \left(\theta_ {s} (t)\right) F _ {l} (t, f) A _ {O} (t, f), \tag {9} +$$ + +where $F_{L}(\theta_{s}(t))$ and $F_{R}(\theta_{s}(t))$ are head-related transfer functions (HRTF) based on $\theta_s(t)$ , $F_{l}(t)$ denotes the room impulse response (RIR) at the user's location and $A_{O}(t,f)$ is the sound emitted from the speaker in the STFT domain. We utilize $\theta_{s}(t)$ to retrieve $F_{L}(\theta_{s}(t))$ and $F_{R}(\theta_{s}(t))$ from the SADIE II dataset [3]. In order to model $F_{l}(t)$ and $A_{O}(t,f)$ , we make the following assumptions: $F_{l}(t)$ change between different locations at the same time index $t$ only depending on the parameter $\lambda (t)$ , i.e., + +$$ +F _ {l} (t, f) = \lambda (t) F (t, f), \tag {10} +$$ + +where $F(t)$ is the RIR at the recording microphone. Using the above assumption, Eq.(8) and Eq.(9) can be rewritten as + +$$ +A _ {L} (t, f) = \lambda (t) F _ {L} \left(\theta_ {s} (t)\right) A (t, f), \tag {11} +$$ + +$$ +A _ {R} (t, f) = \lambda (t) F _ {R} \left(\theta_ {s} (t)\right) A (t, f), \tag {12} +$$ + +where $A(t,f) = F(t,f)A_O(t,f)$ is the audio at the recording microphone. + +# 4.2. Experimental Settings + +Training setups. The training for all light field data was completed on a single NVIDIA A100 GPU. Due to memory limitations, we split the video and trained a model every 60 frames, total 300 frames. We also followed the original train-test-split strategy of the baseline for our experiments. + +Metric. We analyzed the quality of dynamic light field reconstruction using the most common evaluation metrics PSNR, SSIM, and LPIPS(alex). + +# 5. Results and Analysis + +# 5.1. Validation of Light Field Data + +Quantitative Results. As illustrated in Table 3, 4 and Figure 5, our indoor and outdoor data both yield correct results across three 3DGS-based paradigms. STG++ outperforms all baselines in all scenes, delivering slightly better results, the highest rendering speed, and a moderate model size. We also found that, since 4DGS trains using the sparse point cloud from the 0th frame and limits the number of final points, its performance is significantly lower than other pipelines on texture-richer outdoor scenes, such as Scene6 + +Table 3. Test views performance of 3DGS-based dynamic scene reconstruction method on ImViD dataset. + +
MethodScene1 Opera_girlScene2 LaboratoryScene6 Puppy
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
4DGS23.2270.7530.41025.7980.8900.17618.1210.2220.711
4DRotor27.2630.7750.32828.0070.9170.09817.9160.2980.331
STG28.4820.7860.28726.3060.9100.11420.4970.5940.211
STG++31.2400.7990.27727.5810.9160.10720.5330.5970.202
+ +Table 4. The storage size of training results from different methods and their rendering speed on a single A100 GPU. *CPU version. + +
Scene1 Opera_girlScene2 LaboratoryScene6 Puppy
Mem(MB)FPSMem(MB)FPSMem(MB)FPS
4DGS101.0546.22105.0244.93101.6541.86
4DRotor*5818.611.006867.180.674415.820.56
STG382.4196.10318.41133.651287.55100.00
STG++387.17108.89329.07137.781276.21110.47
+ +Table 5. User study for the sound field construction. + +
Auditory Spatial PerceptionSound QualityImmersiveness
Very poor0.00%0.00%0.00%
Poor0.00%0.00%0.00%
Fair14.28%19.04%9.52%
Good23.80%33.33%42.85%
Excellent61.90%47.61%47.61%
+ +puppy. Additionally, due to the density control in the time dimension implemented by 4Drotor, it performs better in frames and regions with more significant motion, which is an aspect worth referencing and adopting in future work. More results can be seen in the supplementary material. + +Qualitative Results. As shown in Figure 6, Our STG++ achieves better temporal continuity in the color of pixels at the same location. There are almost no abrupt transitions both within and between segments. Details are showed in the supplementary for a more intuitive experience. + +# 5.2. Effectiveness of Sound Field Data + +Quantitative Results. Since our sound field reconstruction algorithm is non-trainable, the reconstructed spatial audio lacks a corresponding ground truth, making it difficult to evaluate the results using quantitative metrics. Therefore, we conducted a user study with 21 experts to assess the performance of the sound field reconstruction. + +Each participant rated the sound field reconstruction on three dimensions: auditory spatial perception, sound quality, and immersiveness. Auditory spatial perception refers to the listener's ability to perceive the distribution and localization of sound in space. Sound quality refers to the audio quality of the spatial audio compared to the corresponding microphone-recorded signal. Immersiveness refers to the listener's sense of being in a space where the sound source genuinely exists. + +The results of the user study are shown in Table 5, with + +![](images/d53e3b037fabb6a88cf96a14c862440b4201c59e4bcbde094403f8d8efcff0c0.jpg) +Figure 5. Comparison of the rendering results of four baselines on Scene 1 Opera, Scene 2 Laboratory, and Scene 6 Puppy. + +![](images/da0ead87a4c48d8d85119c17848f637e6c43694718d0bc7e740bd8bab431e61c.jpg) +Figure 6. The continuity of pixels at the same location across different frames and segments. The visualizations present the RGB (middle) and brightness (right) variations. + +numbers as percentages of participants. The majority of the participants $(61.90\%)$ rated the auditory spatial perception as excellent, $80.94\%$ participants felt that the generated spatial audio did not exhibit significant degradation in quality, and $90.46\%$ found the audio immersive, which demonstrates the effectiveness of our data capture and sound field reconstruction methods. + +# 5.3. Multimodal VR Experiences. + +We finally integrated light field and sound field reconstruction results to achieve a 6-DoF multimodal immersive VR experience with real-time rendering speed at 60 FPS on a single 3090 GPU. As shown in Figure 7, the visual and auditory experiences correspond uniquely to the user's position and viewing direction. The experience can be further explored by listening through headphones with the supplementary video provided. + +![](images/2d974acfdb4ab91af8a93b9c3e0ea54ca4f3aec0a5de2203e0c84586ec2a33b5.jpg) +Figure 7. Visualization of the interaction trajectory and corresponding visual&auditory results. Note that the multi-modal feedback changes consistently with the positions and orientations. + +# 6. Conclusion + +In this work, we introduce ImViD, to our knowledge, the first dataset for immersive volumetric video. This dataset encompasses seven large indoor and outdoor scenes, featuring rich foreground and background content that showcases engaging and commonly encountered daily scenarios, along with high-quality synchronized audio sequences. We validated the effectiveness of the dataset using the latest dynamic light field reconstruction algorithms, providing both quantitative and qualitative analyses of current limitations and potential improvements. Moreover, we propose a base pipeline for constructing immersive volumetric video from this data. We believe ImViD will facilitate the exploration of the limitations of current volumetric video reconstruction algorithms and drive immersive VR/AR experiences. + +# Acknowledgment + +# References + +[1] https://www.apple.com/newsroom/2024/07/new-apple-immersive-video-series-and-films-premiere-on-vision-pro/.1 +[2] https://www.gracia.ai/.1 +[3] Cal Armstrong, Lewis Thresh, Damian Murphy, and Gavin Kearney. A perceptual evaluation of individual and non-individual hrfts: A case study of the sadie ii database. Applied Sciences, 8(11):2029, 2018. 7 +[4] Benjamin Attal, Jia-Bin Huang, Christian Richardt, Michael Zollhoefer, Johannes Kopf, Matthew O'Toole, and Changil Kim. Hyperreel: High-fidelity 6-dof video with ray-conditioned sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16610-16620, 2023. 2, 3 +[5] Swapnil Bhosale, Haosen Yang, Diptesh Kanojia, Jiankang Deng, and Xiatian Zhu. Av-gs: Learning material and geometry aware priors for novel view acoustic synthesis. arXiv preprint arXiv:2406.08920, 2024. 3 +[6] Michael Broxton, John Flynn, Ryan Overbeck, Daniel Erickson, Peter Hedman, Matthew Duvall, Jason Dourgarian, Jay Busch, Matt Whalen, and Paul Debevec. Immersive light field video with a layered mesh representation. ACM Transactions on Graphics (TOG), 39(4):86-1, 2020. 1, 2, 3, 4, 7 +[7] Amandine Brunetto, Sascha Hornauer, and Fabien Moutarde. Neraf: 3d scene infused neural radiance and acoustic fields. arXiv preprint arXiv:2405.18213, 2024. 3 +[8] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 130-141, 2023. 2, 3 +[9] Hao Chen, Yuqi Hou, Chenyuan Qu, Irene Testini, Xiaohan Hong, and Jianbo Jiao. $360+\mathrm{x}$ : A panoptic multimodal scene understanding dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19373-19382, 2024. 2, 3, 1, 7 +[10] Kai Cheng, Xiaoxiao Long, Kaizhi Yang, Yao Yao, Wei Yin, Yuexin Ma, Wenping Wang, and Xuejin Chen. Gaussianpro: 3d gaussian splatting with progressive propagation. arXiv preprint arXiv:2402.14650, 2024. 2 +[11] Alvaro Collet, Ming Chuang, Pat Sweeney, Don Gillett, Dennis Evseev, David Calabrese, Hugues Hoppe, Adam Kirk, and Steve Sullivan. High-quality streamable free-viewpoint video. ACM Transactions on Graphics (ToG), 34(4):1-13, 2015. 3 + +This work was supported by National Key R&D Program of China (No.2022YFF0902204), NSFC (No.62171255, 52425801), Tsinghua University-Migu Xinkong Culture Technology (Xiamen) Co.Ltd. Joint Research Center for Intelligent Light Field and Interaction Technology, Guoqiang Institute of Tsinghua University (No.2021GQG0001). +[12] Yuanxing Duan, Fangyin Wei, Qiyu Dai, Yuhang He, Wenzheng Chen, and Baoquan Chen. 4d-rotor gaussian splatting: Towards efficient novel view synthesis for dynamic scenes. In ACM SIGGRAPH 2024 Conference Papers, pages 1–11, 2024. 3, 4, 6 +[13] Bardienus P Duisterhof, Zhao Mandi, Yunchao Yao, Jia-Wei Liu, Mike Zheng Shou, Shuran Song, and Jeffrey Ichnowski. Md-splating: Learning metric deformation from 4d gaussians in highly deformable scenes. arXiv preprint arXiv:2312.00583, 2023. 3 +[14] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12479–12488, 2023. 2, 3 +[15] Hang Gao, Ruilong Li, Shubham Tulsiani, Bryan Russell, and Angjoo Kanazawa. Monocular dynamic view synthesis: A reality check. Advances in Neural Information Processing Systems, 35:33768-33780, 2022. 3, 7 +[16] Huiyu Gao, Jiahao Ma, David Ahmedt-Aristizabal, Chuong Nguyen, and Miaomiao Liu. Soaf: Scene occlusion-aware neural acoustic field. arXiv preprint arXiv:2407.02264, 2024.3 +[17] Zhiyang Guo, Wengang Zhou, Li Li, Min Wang, and Houqiang Li. Motion-aware 3d gaussian splatting for efficient dynamic scene reconstruction. arXiv preprint arXiv:2403.11447, 2024. 3 +[18] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE transactions on pattern analysis and machine intelligence, 36(7):1325-1339, 2013. 1 +[19] Hanbyul Joo, Tomas Simon, Xulong Li, Hao Liu, Lei Tan, Lin Gui, Sean Banerjee, Timothy Scott Godisart, Bart Nabbe, Iain Matthews, Takeo Kanade, Shohei Nobuhara, and Yaser Sheikh. Panoptic studio: A massively multiview system for social interaction capture. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017. 2, 3, 1, 7 +[20] HyunJun Jung, Nikolas Brasch, Jifei Song, Eduardo Perez-Pellitero, Yiren Zhou, Zhihao Li, Nassir Navab, and Benjamin Busam. Deformable 3d gaussian splatting for animatable human avatars. arXiv preprint arXiv:2312.15059, 2023. 3 +[21] Takeo Kanade, Peter Rander, and PJ Narayanan. Virtualized reality: Constructing virtual worlds from real scenes. IEEE multimedia, 4(1):34-47, 1997. 3 +[22] Kai Katsumata, Duc Minh Vo, and Hideki Nakayama. A compact dynamic 3d gaussian representation for real-time dynamic view synthesis, 2024. 2, 3 +[23] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuehler, and George Drettakis. 3D Gaussian Splatting for Real-Time Radiance Field Rendering. ACM Transactions on Graphics, 42(4):1-14, 2023. 2 +[24] Agelos Kratimenos, Jiahui Lei, and Kostas Daniilidis. Dynmf: Neural motion factorization for real-time dynamic view synthesis with 3d gaussian splatting. arXiv preprint arXiv:2312.00112, 2023. 2, 3 + +[25] Joo Chan Lee, Daniel Rho, Xiangyu Sun, Jong Hwan Ko, and Eunbyung Park. Compact 3d gaussian splatting for static and dynamic radiance fields. arXiv preprint arXiv:2408.03822, 2024. 3 +[26] Deqi Li, Shi-Sheng Huang, Zhiyuan Lu, Xinran Duan, and Hua Huang. St-4dgs: Spatial-temporally consistent 4d gaussian splatting for efficient dynamic scene rendering. In ACM SIGGRAPH 2024 Conference Papers, pages 1–11, 2024. 3 +[27] Lingzhi Li, Zhen Shen, Zhongshu Wang, Li Shen, and Ping Tan. Streaming radiance fields for 3d video synthesis. Advances in Neural Information Processing Systems, 35: 13485-13498, 2022. 2, 3, 1, 7 +[28] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5521-5531, 2022. 2, 3, 1, 7 +[29] Yanyan Li, Chenyu Lyu, Yan Di, Guangyao Zhai, Gim Hee Lee, and Federico Tombari. Geogaussian: Geometry-aware gaussian splatting for scene rendering. arXiv preprint arXiv:2403.11324, 2024. 2 +[30] Zhan Li, Zhang Chen, Zhong Li, and Yi Xu. Spacetime gaussian feature splatting for real-time dynamic view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8508-8520, 2024. 2, 3, 4, 6 +[31] Susan Liang, Chao Huang, Yapeng Tian, Anurag Kumar, and Chenliang Xu. Av-nerf: Learning neural fields for real-world audio-visual scene synthesis. Advances in Neural Information Processing Systems, 36:37472-37490, 2023. 3, 5 +[32] Yiqing Liang, Numair Khan, Zhengqin Li, Thu Nguyen-Phuoc, Douglas Lanman, James Tompkin, and Lei Xiao. Gaufre: Gaussian deformation fields for real-time dynamic novel view synthesis. arXiv preprint arXiv:2312.11458, 2023. 3 +[33] Haotong Lin, Sida Peng, Zhen Xu, Yunzhi Yan, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Efficient neural radiance fields for interactive free-viewpoint video. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 2, 3, 1, 7 +[34] Kai-En Lin, Lei Xiao, Feng Liu, Guowei Yang, and Ravi Ramamoorthi. Deep 3d mask volume for view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1749–1758, 2021. 2, 3, 1, 7 +[35] Youtian Lin, Zuozhuo Dai, Siyu Zhu, and Yao Yao. Gaussian-flow: 4d reconstruction with dynamic 3d gaussian particle. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21136-21145, 2024. 2, 3 +[36] Qingming Liu, Yuan Liu, Jiepeng Wang, Xianqiang Lv, Peng Wang, Wenping Wang, and Junhui Hou. Modgs: Dynamic gaussian splatting from causally-captured monocular videos. arXiv preprint arXiv:2406.00434, 2024. 7 +[37] Cheng-You Lu, Peisen Zhou, Angela Xing, Chandradeep Pokhariya, Arnab Dey, Ishaan Nikhil Shah, Rugved Mavidi + +palli, Dylan Hu, Andrew I Comport, Kefan Chen, et al. Diva360: The dynamic visual dataset for immersive neural fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22466-22476, 2024. 2, 7 +[38] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. arXiv preprint arXiv:2312.00109, 2023. 2 +[39] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713, 2023. 3 +[40] Dawid Malarz, Weronika Smolak, Jacek Tabor, Sławomir Tadeja, and Przemysław Spurek. Gaussian splatting with nef-based color and opacity. 2 +[41] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2 +[42] David Narciso, Maximino Bessa, Miguel Melo, Antonio Coelho, and José Vasconcelos-Raposo. Immersive $360^{\circ}$ video user experience: impact of different variables in the sense of presence and cybersickness. Universal Access in the Information Society, 18:77-87, 2019. 1 +[43] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228, 2021. 2, 1, 7 +[44] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In CVPR, 2021. 2, 3, 1, 7 +[45] Weining Ren, Zihan Zhu, Boyang Sun, Jiaqi Chen, Marc Pollefeys, and Songyou Peng. Nerf on-the-go: Exploiting uncertainty for distractor-free nerfs in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8931-8940, 2024. 2, 1, 7 +[46] Neus Sabater, Guillaume Boisson, Benoit Vandame, Paul Kerbiriou, Frederic Babon, Matthieu Hog, Remy Gendrot, Tristan Langlois, Olivier Bureller, Arno Schubert, et al. Dataset and pipeline for multi-view light-field video. In Proceedings of the IEEE conference on computer vision and pattern recognition Workshops, pages 30-40, 2017. 2, 3, 1, 7 +[47] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 5 +[48] Ruizhi Shao, Zerong Zheng, Hanzhang Tu, Boring Liu, Hongwen Zhang, and Yebin Liu. Tensor4d: Efficient neural 4d decomposition for high-fidelity dynamic reconstruction and rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16632-16642, 2023. 2, 3, 1 + +[49] Roman Shapovalov, Yanir Kleiman, Ignacio Rocco, David Novotny, Andrea Vedaldi, Changan Chen, Filippos Kokkinos, Ben Graham, and Natalia Neverova. Replay: Multi-modal multi-view acted videos for casual holography. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20338-20348, 2023. 2, 3, 1, 7 +[50] Richard Shaw, Michal Nazarczuk, Jifei Song, Arthur Moreau, Sibi Catley-Chandar, Helisa Dhamo, and Eduardo Perez-Pellitero. Swings: Sliding windows for dynamic 3d gaussian splatting, 2024. 3 +[51] Liangchen Song, Anpei Chen, Zhong Li, Zhang Chen, Lele Chen, Junsong Yuan, Yi Xu, and Andreas Geiger. Nerfplayer: A streamable dynamic scene representation with decomposed neural radiance fields. IEEE Transactions on Visualization and Computer Graphics, 29(5):2732-2742, 2023. 2, 3 +[52] Jiakai Sun, Han Jiao, Guangyuan Li, Zhanjie Zhang, Lei Zhao, and Wei Xing. 3dgstream: On-the-fly training of 3d gaussians for efficient streaming of photo-realistic free-viewpoint videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20675-20685, 2024. 3 +[53] Feng Wang, Sinan Tan, Xinghang Li, Zeyue Tian, Yafei Song, and Huaping Liu. Mixed neural voxels for fast multiview video synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19706-19716, 2023. 2, 3 +[54] Feng Wang, Zilong Chen, Guokang Wang, Yafei Song, and Huaping Liu. Masked space-time hash encoding for efficient dynamic scene reconstruction. Advances in Neural Information Processing Systems, 36, 2024. 2, 1, 7 +[55] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20310-20320, 2024. 2, 3, 4, 6 +[56] Tianhao Wu, Fangcheng Zhong, Andrea Tagliasacchi, Forrester Cole, and Cengiz Oztireli. D^2nerf: Self-supervised decoupling of dynamic and static objects from a monocular video. Advances in neural information processing systems, 35:32653-32666, 2022. 2, 1, 7 +[57] Yunzhi Yan, Haotong Lin, Chenxu Zhou, Weijie Wang, Haiyang Sun, Kun Zhan, Xianpeng Lang, Xiaowei Zhou, and Sida Peng. Street gaussians for modeling dynamic urban scenes. arXiv preprint arXiv:2401.01339, 2024. 2 +[58] Zeyu Yang, Hongye Yang, Zijie Pan, Xiatian Zhu, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. arXiv preprint arXiv:2310.10642, 2023. 3 +[59] Ziyi Yang, Xinyu Gao, Yangtian Sun, Yihua Huang, Xiaoyang Lyu, Wen Zhou, Shaohui Jiao, Xiaojuan Qi, and Xiaogang Jin. Spec-gaussian: Anisotropic view-dependent appearance for 3d gaussian splatting. arXiv preprint arXiv:2402.15870, 2024. 2 +[60] Jae Shin Yoon, Kihwan Kim, Orazio Gallo, Hyun Soo Park, and Jan Kautz. Novel view synthesis of dynamic scenes with + +globally coherent depths from a monocular camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5336-5345, 2020. 2, 3, 1, 7 +[61] Dongbin Zhang, Chuming Wang, Weitao Wang, Peihao Li, Minghan Qin, and Haoqian Wang. Gaussian in the wild: 3d gaussian splatting for unconstrained image collections. arXiv preprint arXiv:2403.15704, 2024. 2 \ No newline at end of file diff --git a/2025/ImViD_ Immersive Volumetric Videos for Enhanced VR Engagement/images.zip b/2025/ImViD_ Immersive Volumetric Videos for Enhanced VR Engagement/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d7181a72a89f197edd60f7a6e382bd2e8dd2b684 --- /dev/null +++ b/2025/ImViD_ Immersive Volumetric Videos for Enhanced VR Engagement/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1efa9f8ac6d6c5d0bdad5ea3aec04b679da028ac0ac0acfe1e988a906b4dfe6f +size 845034 diff --git a/2025/ImViD_ Immersive Volumetric Videos for Enhanced VR Engagement/layout.json b/2025/ImViD_ Immersive Volumetric Videos for Enhanced VR Engagement/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..bd65afe55b090a161287528955cda6761d65bcee --- /dev/null +++ b/2025/ImViD_ Immersive Volumetric Videos for Enhanced VR Engagement/layout.json @@ -0,0 +1,8538 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 93, + 103, + 518, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 103, + 518, + 121 + ], + "spans": [ + { + "bbox": [ + 93, + 103, + 518, + 121 + ], + "type": "text", + "content": "ImViD: Immersive Volumetric Videos for Enhanced VR Engagement" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 72, + 142, + 539, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 142, + 539, + 228 + ], + "spans": [ + { + "bbox": [ + 72, + 142, + 539, + 228 + ], + "type": "text", + "content": "Zhengxian Yang" + }, + { + "bbox": [ + 72, + 142, + 539, + 228 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 72, + 142, + 539, + 228 + ], + "type": "text", + "content": ", Shi Pan" + }, + { + "bbox": [ + 72, + 142, + 539, + 228 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 72, + 142, + 539, + 228 + ], + "type": "text", + "content": ", Shengqi Wang" + }, + { + "bbox": [ + 72, + 142, + 539, + 228 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 72, + 142, + 539, + 228 + ], + "type": "text", + "content": ", Haoxiang Wang" + }, + { + "bbox": [ + 72, + 142, + 539, + 228 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 72, + 142, + 539, + 228 + ], + "type": "text", + "content": ", Li Lin" + }, + { + "bbox": [ + 72, + 142, + 539, + 228 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 72, + 142, + 539, + 228 + ], + "type": "text", + "content": ", Guanjun Li" + }, + { + "bbox": [ + 72, + 142, + 539, + 228 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 72, + 142, + 539, + 228 + ], + "type": "text", + "content": ", Zhengqi Wen" + }, + { + "bbox": [ + 72, + 142, + 539, + 228 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 72, + 142, + 539, + 228 + ], + "type": "text", + "content": ", Borong Lin" + }, + { + "bbox": [ + 72, + 142, + 539, + 228 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 72, + 142, + 539, + 228 + ], + "type": "text", + "content": ", Jianhua Tao" + }, + { + "bbox": [ + 72, + 142, + 539, + 228 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 72, + 142, + 539, + 228 + ], + "type": "text", + "content": ", Tao Yu" + }, + { + "bbox": [ + 72, + 142, + 539, + 228 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 72, + 142, + 539, + 228 + ], + "type": "text", + "content": ", Tsinghua University " + }, + { + "bbox": [ + 72, + 142, + 539, + 228 + ], + "type": "inline_equation", + "content": "{}^{2}" + }, + { + "bbox": [ + 72, + 142, + 539, + 228 + ], + "type": "text", + "content": "Migu Beijing Research Institute " + }, + { + "bbox": [ + 72, + 142, + 539, + 228 + ], + "type": "inline_equation", + "content": "{}^{3}" + }, + { + "bbox": [ + 72, + 142, + 539, + 228 + ], + "type": "text", + "content": "Institute of Automation, Chinese Academy of Science " + }, + { + "bbox": [ + 72, + 142, + 539, + 228 + ], + "type": "inline_equation", + "content": "\\{\\mathrm{zx - yang23, ps23, shengqi - 21, whx22}\\} @" + }, + { + "bbox": [ + 72, + 142, + 539, + 228 + ], + "type": "text", + "content": "mails.tsinghua.edu.cn lilin@migu.chinamobile.com guanjun.li@nlpr ia.ac.cn " + }, + { + "bbox": [ + 72, + 142, + 539, + 228 + ], + "type": "inline_equation", + "content": "\\{\\mathrm{zqwen, linbr, jhtao, ytrock}\\} @" + }, + { + "bbox": [ + 72, + 142, + 539, + 228 + ], + "type": "text", + "content": "tsinghua.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 57, + 245, + 187, + 381 + ], + "blocks": [ + { + "bbox": [ + 57, + 245, + 187, + 381 + ], + "lines": [ + { + "bbox": [ + 57, + 245, + 187, + 381 + ], + "spans": [ + { + "bbox": [ + 57, + 245, + 187, + 381 + ], + "type": "image", + "image_path": "29df6dfe8c358de09148a8c338fa2c40e7df876d550543ebbb5b53dcecaae38c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 393, + 555, + 416 + ], + "lines": [ + { + "bbox": [ + 55, + 393, + 555, + 416 + ], + "spans": [ + { + "bbox": [ + 55, + 393, + 555, + 416 + ], + "type": "text", + "content": "Figure 1. We introduce ImViD, a dataset for immersive volumetric videos. ImViD records dynamic scenes using a multi-view audio-video capture rig moving in a space-oriented manner, which provides a new benchmark for volumetric video reconstruction and its application." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 187, + 245, + 423, + 380 + ], + "blocks": [ + { + "bbox": [ + 187, + 245, + 423, + 380 + ], + "lines": [ + { + "bbox": [ + 187, + 245, + 423, + 380 + ], + "spans": [ + { + "bbox": [ + 187, + 245, + 423, + 380 + ], + "type": "image", + "image_path": "54cea3e825ce7d714e26bf4ae2061d89768861e2c9285a1234747496b4cf9a61.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 424, + 245, + 553, + 380 + ], + "blocks": [ + { + "bbox": [ + 424, + 245, + 553, + 380 + ], + "lines": [ + { + "bbox": [ + 424, + 245, + 553, + 380 + ], + "spans": [ + { + "bbox": [ + 424, + 245, + 553, + 380 + ], + "type": "image", + "image_path": "7620de647f0ca8575e9ca85febef46564a18a61207b78bd20144f82cf7d2d859.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 152, + 427, + 200, + 440 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 427, + 200, + 440 + ], + "spans": [ + { + "bbox": [ + 152, + 427, + 200, + 440 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 453, + 295, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 453, + 295, + 608 + ], + "spans": [ + { + "bbox": [ + 54, + 453, + 295, + 608 + ], + "type": "text", + "content": "User engagement is greatly enhanced by fully immersive multi-modal experiences that combine visual and auditory stimuli. Consequently, the next frontier in VR/AR technologies lies in immersive volumetric videos with complete scene capture, large 6-DoF interaction space, multimodal feedback, and high resolution&frame-rate contents. To stimulate the reconstruction of immersive volumetric videos, we introduce ImViD, a multi-view, multi-modal dataset featuring complete space-oriented data capture and various indoor/outdoor scenarios. Our capture rig supports multi-view video-audio capture while on the move, a capability absent in existing datasets, significantly enhancing the completeness, flexibility, and efficiency of data capture." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 609, + 295, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 609, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 54, + 609, + 295, + 668 + ], + "type": "text", + "content": "The captured multi-view videos (with synchronized audi- 15 minutes, and include rich foreground-background elements, and complex dynamics. We benchmark existing methods using our dataset and establish a base pipeline for construct" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 429, + 555, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 429, + 555, + 513 + ], + "spans": [ + { + "bbox": [ + 313, + 429, + 555, + 513 + ], + "type": "text", + "content": "ing immersive volumetric videos from multi-view audiovisual inputs for 6-DoF multi-modal immersive VR experiences. The benchmark and the reconstruction and interaction results demonstrate the effectiveness of our dataset and baseline method, which we believe will stimulate future research on immersive volumetric video production. Project Page: https://yzxqh.github.io/ImViD/" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 525, + 394, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 525, + 394, + 538 + ], + "spans": [ + { + "bbox": [ + 314, + 525, + 394, + 538 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 545, + 555, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 545, + 555, + 641 + ], + "spans": [ + { + "bbox": [ + 312, + 545, + 555, + 641 + ], + "type": "text", + "content": "High-quality and high-fidelity digital modeling of the real world has long been a topic of significant interest. With breakthroughs in 3D reconstruction algorithms and virtual reality technology, the generation of such visually compelling content is gradually becoming feasible, whether for reproducing past events and memories or remotely viewing current occurrences. Achieving complete user immersion is a goal pursued by many researchers." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 326, + 642, + 511, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 642, + 511, + 653 + ], + "spans": [ + { + "bbox": [ + 326, + 642, + 511, + 653 + ], + "type": "text", + "content": "What defines an immersive volumetric video?" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "content": "Dwelling on previous tests [1, 6] and user feedback [2, 42], we believe that immersive media should have the following four characteristics: i) Full " + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "content": " foreground and background, ii) High-quality 6-DoF interaction, which current volumetric video can achieve, iii) Multimodal experi" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 673, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 673, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 673, + 295, + 713 + ], + "type": "text", + "content": "*Equal contributions. † Corresponding author. \nZhengxian Yang, Shengqi Wang, Haoxiang Wang, Jianhua Tao and Tao Yu are at BNRist&Department of Automation; Shi Pan and Borong Lin are at Department of Building Science; Zhengqi Wen is at BNRist." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16554" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 107 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 107 + ], + "type": "text", + "content": "ence with light and sound, and iv) High frame rate and long-duration dynamic content. We refer to a media format that meets all these criteria as immersive volumetric video." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 109, + 295, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 109, + 295, + 289 + ], + "spans": [ + { + "bbox": [ + 54, + 109, + 295, + 289 + ], + "type": "text", + "content": "However, with the development of camera equipment, despite many efforts using monocular handheld cameras or fixed camera arrays to build datasets, none of these solutions fully address all four requirements. For example, Panoptic Sports [19], ZJU-Mocap [44] and DiVa-360 [37] achieved human/object-centric " + }, + { + "bbox": [ + 54, + 109, + 295, + 289 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 54, + 109, + 295, + 289 + ], + "type": "text", + "content": " capture but lack complex backgrounds and multimodal data; Google's immersive light field [6] achieves 6-DoF volumetric video in an inside-looking-out manner but lacks multimodality and high frame rate, long-duration dynamic content; Replay [49] introduces sound, but the camera setup is not suitable for human viewing habits thus cannot achieve high-quality 6-DoF interaction. Both academia and industry require an appropriate capture method to support the development of immersive volumetric video technology." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 289, + 295, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 289, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 56, + 289, + 295, + 529 + ], + "type": "text", + "content": "Additionally, NeRF [41]'s emergence has increased attention to reconstruction and rendering techniques, making it easier to generate high-fidelity interactive 3D scenes from multiple views. 3DGS [23] and its subsequent improvements [10, 29, 38, 40, 57, 59, 61] have further boosted rendering speed and quality, fueling peoples' expectations for rapid real-world modeling and its applications in industries such as education, healthcare, and entertainment. This can be seen as an evolutionary opportunity for immersive volumetric video. However, current technologies face challenges. In dynamic light field reconstruction, neither explicit [22, 24, 30, 35, 55] nor implicit neural representation based [4, 8, 14, 28, 48, 51, 53] methods have been able to achieve a practical balance between foreground-background reconstruction accuracy, rendering speed, spatiotemporal consistency, and storage efficiency in dynamic scenes; In the realm of sound field reconstruction, it is typically necessary to collect a large amount of binaural audio or room impulse response data from different locations for model training, which is user-unfriendly in practical use." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 529, + 295, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 295, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 295, + 672 + ], + "type": "text", + "content": "To better support the development of immersive volumetric video technology, we have developed a moving capture rig (with 46 synchronized cameras, as shown in Fig.3) and designed two space-oriented data capture strategies. We have constructed an audiovisual dataset containing 7 indoor and outdoor scenes, with a " + }, + { + "bbox": [ + 55, + 529, + 295, + 672 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 55, + 529, + 295, + 672 + ], + "type": "text", + "content": " field of view, large 6-DoF interaction space, and high-resolution, long-duration dynamic contents. We also validate this dataset using the latest novel view synthesis and 3D reconstruction methods. Finally, a complete pipeline for constructing light and sound fields from this type of data was proposed, which initially achieved immersive volume video." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 673, + 288, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 673, + 288, + 685 + ], + "spans": [ + { + "bbox": [ + 67, + 673, + 288, + 685 + ], + "type": "text", + "content": "Our main contributions can be summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "- We propose to our knowledge the first and probably the highest-quality multi-view dataset ImViD for immersive" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 322, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 553, + 108 + ], + "type": "text", + "content": "volumetric video. It features " + }, + { + "bbox": [ + 322, + 72, + 553, + 108 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 322, + 72, + 553, + 108 + ], + "type": "text", + "content": " foreground and background capture, long-duration video&audio data under high resolution@5K, high frame rate@60FPS." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 108, + 553, + 180 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 108, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 108, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 314, + 108, + 553, + 144 + ], + "type": "text", + "content": "- We conduct extensive experiments to validate the effectiveness of the proposed datasets by evaluating and analyzing the performance of several current baselines." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 144, + 553, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 144, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 314, + 144, + 553, + 180 + ], + "type": "text", + "content": "- Follow the proposed pipeline shown in Figure 2, We finally achieve the reconstruction of immersive volumetric video and integrate it for immersive VR experience." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 192, + 400, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 192, + 400, + 205 + ], + "spans": [ + { + "bbox": [ + 314, + 192, + 400, + 205 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 213, + 553, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 553, + 248 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 553, + 248 + ], + "type": "text", + "content": "Datasets for Dynamic Novel View Synthesis. The datasets for dynamic scene reconstruction can be classified as monocular-based and multi-view-based." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 250, + 553, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 250, + 553, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 250, + 553, + 357 + ], + "type": "text", + "content": "Monocular acquisition systems are popular due to their low cost and ease of construction. such as HyperNeRF [43], Dynamic Scene Dataset [60], and D2NeRF [56]. However, these datasets suffer from low resolution, limited capture space, and the limited durations. Although NeRF On-the-go [45] allows for larger capture ranges by walking while shooting, high-quality reconstructions are confined to the capture path, and the small field of view (FOV) limits prolonged observations of specific scene positions." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 358, + 554, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 358, + 554, + 490 + ], + "spans": [ + { + "bbox": [ + 313, + 358, + 554, + 490 + ], + "type": "text", + "content": "Multi-camera data collection has gained significant attention due to its ability to provide a larger FOV and richer details, such as Immersive Light Field dataset [6], Technicolor [46], UCSD Dynamic Scene Dataset [34] and The Plenoptic Dataset [28], followed by [27, 33, 54]. However, all these setups remain static during capture, limiting them to frontal views and hindering " + }, + { + "bbox": [ + 313, + 358, + 554, + 490 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 313, + 358, + 554, + 490 + ], + "type": "text", + "content": " reconstruction. Additionally, the video sequences are typically short, with a maximum duration of 2 minutes (often less than 30 seconds) and a maximum resolution of 4K, which is insufficient for immersive VR experiences." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 491, + 554, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 554, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 554, + 658 + ], + "type": "text", + "content": "Moreover, the previously mentioned datasets lack sound recordings, leading to the loss of immersion without multimodality. The Replay dataset [49] addresses this by focusing on long sequences with professional actors in familiar settings. However, aside from the head-mounted cameras that enables rotation with head, all other cameras remain static. Furthermore, the DSLR arrangement does not align with human viewing habits in VR, making them unsuitable as benchmarks for novel view synthesis tasks. Recent work [9] presents a dataset with a " + }, + { + "bbox": [ + 313, + 491, + 554, + 658 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 313, + 491, + 554, + 658 + ], + "type": "text", + "content": " camera and audio capture device but constrained by a fixed-point shooting strategy, resulting in sparse viewpoints that hinder the reconstruction of high-quality dynamic scenes. More details about datasets can be found in Tab. 1." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 665, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 714 + ], + "type": "text", + "content": "Dynamic Light&Sound Field Reconstruction. Temporal variations in input images make reconstructing dynamic scenes more challenging than static ones. The key issue is efficiently representing 4D scenes while ensuring spa" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16555" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 91, + 555, + 338 + ], + "blocks": [ + { + "bbox": [ + 94, + 71, + 515, + 83 + ], + "lines": [ + { + "bbox": [ + 94, + 71, + 515, + 83 + ], + "spans": [ + { + "bbox": [ + 94, + 71, + 515, + 83 + ], + "type": "text", + "content": "Table 1. Existing real-world datasets for dynamic novel view synthesis (full version in the supplementary materials)." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 91, + 555, + 338 + ], + "lines": [ + { + "bbox": [ + 57, + 91, + 555, + 338 + ], + "spans": [ + { + "bbox": [ + 57, + 91, + 555, + 338 + ], + "type": "table", + "html": "
DatasetsNo SceneOutdoor/IndoorCamerasMobilityResolutionAnglesDurationFPSMultimodalityContent
PanopticSports [19]65Indoor480 camerasStatic640×480360°5mins25XHuman-centric actions
Technicolor [46]12Indoor16 camerasStatic2048×1088Frontal2s30XHas a number of close-ups sequences, captured medium angle scenes and other animated scenes where the movement does not come from a human
Immersive-Lightfield [6]15both46 camerasStatic2560×1920Frontal10-30s30XSimple and slow motion of human, animals, objects
Dynamic Scene Datasets (NVIDIA) [60]8Outdoor1 Mobile phone /12 camerasFixed-point Waving /Static1920×1080Frontal5s60XSimple body motions (facial, jump, etc)
UCSD Dynamic [34]96Outdoor10 camerasStatic1920×1080Frontal1-2mins120XVarious visual effects and human interactions
ZJU-Mocap [44]10Indoor21 camerasStatic1024×1024360°20s50XSimple body motions (punch, kick, etc.)
Plenoptic Dataset (DyN-eRF/Neural 3D) [28]6Indoor21 camerasStatic2704×2028Frontal10-30s30XContains high specularity, translucency and transparency objects, motions with changing topology, selfcast moving shadows, various lighting conditions and multiple people moving around in open living room space
iPhone Datasets [15]14both1 hand-held phone /2 camerasFixed-point Waving /Static640×480Frontal8-15s30/60XFeaturing non-repetitive motion, from various categories such as generic objects, humans, and pets
ENeRF-Outdoor [33]4Outdoor18 camerasStatic1920×1080Frontal20-40s30XComplex human motions
Replay [49]46Indoor12 camerasStatic3840×2160360°5mins30✓(Audio)Dancing, chatting, playing video games, unwrapping presents, playing ping pong
360+X [9]28both1 360°cameras and 1 Spectacles camerasStatic5760×2880360°10s (2152 sequence)30✓(Audio)Capture in 17 cities across 5 countries. Panoptic perspective to scene understanding with audio
Im-ViD(Ours)7both46 camerasMoveable5312×2988Frontal and 360°1-5mins60✓(Audio)Seven common indoor and outdoor scenes in daily life, including opera, face-to-face communication, teaching, discussion, music performance, interaction with pets, and playing. Each scene has high-quality synchronized multi-view video and audio with a duration of more than 1 minute, and contains rich elements such as various small objects, glass, and changes in light and shadow
", + "image_path": "3ce26600105276a6e8d0df49247fddf8a82e5c664d4a8c754c8c914b854f33d0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 357, + 297, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 357, + 297, + 706 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 297, + 706 + ], + "type": "text", + "content": "tiotemporal consistency and high accuracy. Traditional methods estimate varying geometries like surfaces [11] and depth maps [21] to handle dynamics. Recently, efforts have shifted towards adapting neural radiance field representations for dynamic scenes. DyNeRF [28] pioneer in the dynamic reconstruction with Neural radiance field. Followed works such as [4, 8, 14, 27, 48, 51, 53] make efforts to improve the high-fidelity and efficiency in the 4D neural representation. However, the common challenges these methods face are memory cost and the modeling of spatiotemporal complexity. The latest work have extended 3DGS to dynamic scenes for high-quality real-time rendering. While modeling dynamic scenes as static ones is straightforward, methods relying solely on monocular data [20, 32] is not suitable for this paradigm, and face challenges of adding constraints or priors to better supervise training. Other works support multi-view data as input. [39] and [52] initialize each frame from the previous one to build Gaussian point clouds per-frame, leading to high memory and suddenly appearing or disappearing artifacts. 4DGS [55] introduces Hexplane to encode spatiotemporal information, using MLPs to predict Gaussian properties change. Subsequent works [13, 17, 26] further add geometric and optical flow constraints to improve frame-to-frame motion but perform well only on short sequences with small actions. Other approaches [22, 24, 30, 35] employ the explicit parametric form (such as radial basis functions) of 3DGS to enhance spatiotemporal consistency. Some methods [12, 58] focus on representation with dual quaternions, 4D spheri" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 357, + 555, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 555, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 555, + 418 + ], + "type": "text", + "content": "cal harmonics, or rotors for end-to-end training. More improvements such as parallel training [50] and using codebooks and masks to enhance scene compactness [25], show the potential in efficient and high-quality representation of long-term complex dynamic scenes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 418, + 556, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 556, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 556, + 658 + ], + "type": "text", + "content": "In addition to light field reconstruction, reconstructing the sound field from multimodal data is equally important, often referred to as novel-view acoustic synthesis. For the novel-view acoustic synthesis task, AV-NeRF [31] synthesizes spatial audio by leveraging visual features from rendered images. NeRAF [7] further decouples the camera and microphone poses used for training. AV-GS [5] incorporates a 3D scene representation during sound field synthesis, addressing the issue where previous methods overlooked the contribution of the broader 3D scene geometry. SOAF [16] considers occlusion in the light field when synthesizing the sound field. However, the above novel-view acoustic synthesis tasks typically require a large amount of multimodal training data collected from different locations, which is often user-unfriendly in practical applications. Moreover, these methods generally cannot solve the problem of sound field synthesis for moving sound sources. This paper proposes a training-free novel-view acoustic synthesis approach for our dataset, which also supports sound field synthesis for moving sound sources." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 669, + 406, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 669, + 406, + 681 + ], + "spans": [ + { + "bbox": [ + 313, + 669, + 406, + 681 + ], + "type": "text", + "content": "3. ImViD Dataset" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "content": "Our goal is to build an immersive volumetric video dataset of real world that integrates both foreground and back-" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "16556" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 75, + 70, + 537, + 331 + ], + "blocks": [ + { + "bbox": [ + 75, + 70, + 537, + 331 + ], + "lines": [ + { + "bbox": [ + 75, + 70, + 537, + 331 + ], + "spans": [ + { + "bbox": [ + 75, + 70, + 537, + 331 + ], + "type": "image", + "image_path": "fc1cdf68e993250542181382f874bc20780f5c2496c601e9da23071faafea384.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 339, + 555, + 406 + ], + "lines": [ + { + "bbox": [ + 54, + 339, + 555, + 406 + ], + "spans": [ + { + "bbox": [ + 54, + 339, + 555, + 406 + ], + "type": "text", + "content": "Figure 2. The pipeline to realize the multimodal 6-DoF immersive VR experiences. We applied a carefully designed rig to (a) simultaneously capture multi-view video and audio. The (b1) presents our reconstruction of dynamic light field based on STG [30] while (b2) demonstrates the construction process of sound field. We have achieved better results than the original algorithm in long-term dynamic scenes by incorporating affine color transformation and t-dimensional density control. Ultimately, we achieve a 6-DoF immersive experience in both light and sound fields, and also benchmark on recent representative methods like 4DGS [55] and 4Drotor [12] to demonstrate the effectiveness of both our dataset and baseline method." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 427, + 297, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 427, + 297, + 594 + ], + "spans": [ + { + "bbox": [ + 54, + 427, + 297, + 594 + ], + "type": "text", + "content": "ground, providing assistance for the development of related research such as spatial video and VR/AR application. Therefore, we have meticulously refined every aspect of the process, from equipment setup and scene selection to shooting strategies and data processing, ensuring maintaining the elements of interest to researchers to the greatest extent. The released ImViD dataset comprises a total of 7 large scenes, including 5 indoor and 2 outdoor environments. The content features a diverse range of common activities, such as opera, meeting, teaching, and playing on the grass. Each scene contains a set of static photos, multi-view time-synchronized video&audio sequences at 5K resolution and 60 FPS, employing two different collection strategies, along with calibrated camera poses." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 600, + 157, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 157, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 157, + 613 + ], + "type": "text", + "content": "3.1. Data Acquisition" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "text", + "content": "The handheld monocular camera is easy to move, providing limited perspectives from various locations. In contrast, the fixed camera arrays, while stationary, offers dense perspectives within a limited range. We aim to combine the advantages of both to design an effective data collection system and strategy for fully immersive VR experience—a feature not found in existing datasets. As shown in Figure 3, our capture system consists of more than 40 GoPro cameras in" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 427, + 555, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 427, + 555, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 427, + 555, + 559 + ], + "type": "text", + "content": "stalled on a hemispherical surface mounted on a remotely controllable mobile car, all of which are connected to a PC to receive synchronous control commands. The system is carefully designed to be at a height similar to that of a human, and the camera arrangement better aligns with human viewing habits compared to other existing devices. This rig can synchronously collect about 1,000 images at 5K resolution within a space of at least " + }, + { + "bbox": [ + 313, + 427, + 555, + 559 + ], + "type": "inline_equation", + "content": "6\\mathrm{m}^3" + }, + { + "bbox": [ + 313, + 427, + 555, + 559 + ], + "type": "text", + "content": " in 2 min. It is also capable of synchronously capturing video and audio at 60 FPS and 5K resolution for 30 min (limited by heat dissipation). Thus, our capture strategy is as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 560, + 555, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 560, + 555, + 585 + ], + "spans": [ + { + "bbox": [ + 313, + 560, + 555, + 585 + ], + "type": "text", + "content": "Step1: Efficient high-quality, high-density " + }, + { + "bbox": [ + 313, + 560, + 555, + 585 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 313, + 560, + 555, + 585 + ], + "type": "text", + "content": " image acquisition of static environments." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 586, + 553, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 553, + 611 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 553, + 611 + ], + "type": "text", + "content": "Step2: Continuous synchronous collection of dynamic scenes under two different strategies." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 613, + 553, + 675 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 313, + 613, + 553, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 613, + 553, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 613, + 553, + 636 + ], + "type": "text", + "content": "i) Fixed-point shooting. The cart remains stationary during the capture, focusing on capturing dynamic details." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 639, + 553, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 639, + 553, + 675 + ], + "spans": [ + { + "bbox": [ + 313, + 639, + 553, + 675 + ], + "type": "text", + "content": "ii) Moving shooting. The cart slowly moves while the scene is in motion, providing additional level of details and larger exploratory space." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "Throughout the capture process, we ensure strict synchronization of the world time codes across all cameras. We carefully maintained consistent focal lengths, exposure," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16557" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 72, + 177, + 216 + ], + "blocks": [ + { + "bbox": [ + 62, + 72, + 177, + 216 + ], + "lines": [ + { + "bbox": [ + 62, + 72, + 177, + 216 + ], + "spans": [ + { + "bbox": [ + 62, + 72, + 177, + 216 + ], + "type": "image", + "image_path": "93ee5eea297500e1b2858df8345906a2384b52ee47b14b1c1729d23290d6a3b0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 228, + 295, + 251 + ], + "lines": [ + { + "bbox": [ + 55, + 228, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 295, + 251 + ], + "type": "text", + "content": "Figure 3. Our rig support two kinds of capturing strategies for high resolution, high frame rate and " + }, + { + "bbox": [ + 55, + 228, + 295, + 251 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 55, + 228, + 295, + 251 + ], + "type": "text", + "content": " dynamic data acquisition." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 181, + 72, + 286, + 216 + ], + "blocks": [ + { + "bbox": [ + 181, + 72, + 286, + 216 + ], + "lines": [ + { + "bbox": [ + 181, + 72, + 286, + 216 + ], + "spans": [ + { + "bbox": [ + 181, + 72, + 286, + 216 + ], + "type": "image", + "image_path": "787b9f96ad10d46945cdaa06015c5884217d06c098fcd13a5f97cef15e89b9eb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 213, + 555, + 258 + ], + "lines": [ + { + "bbox": [ + 313, + 213, + 555, + 258 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 555, + 258 + ], + "type": "text", + "content": "Figure 4. Calculation method for spatiotemporal capture density. 1) the capture strategy of handheld monocular camera 2) represents the fixed camera array 3) Our rig covers a volume of " + }, + { + "bbox": [ + 313, + 213, + 555, + 258 + ], + "type": "inline_equation", + "content": "0.6 \\times \\pi \\times 0.5^2 \\, \\mathrm{m}^3" + }, + { + "bbox": [ + 313, + 213, + 555, + 258 + ], + "type": "text", + "content": " over 5 seconds." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 270, + 296, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 296, + 344 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 296, + 344 + ], + "type": "text", + "content": "and white balance settings for each scene, based on the results of manual iterative testing of camera settings, thereby minimizing hardware-induced variations in data quality. To achieve high-quality audio, we increased the noise reduction coefficient of the cameras and minimized noise generation as much as possible." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 349, + 154, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 349, + 154, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 349, + 154, + 361 + ], + "type": "text", + "content": "3.2. Data Processing" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 366, + 295, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 366, + 295, + 520 + ], + "spans": [ + { + "bbox": [ + 54, + 366, + 295, + 520 + ], + "type": "text", + "content": "For the photos from Step1, we completed the calibration in COLMAP [47] by combining the cameras' prior parameters and their relative positions. For the video&audio from Step2, we first aligned the multi-view sequences for each scene using the world time codes. This step is crucial as the quality of the alignment directly determines the quality of dynamic scene reconstruction and significantly impacts whether the multi-view audio can be used for constructing a 6-DoF sound field. Our time code alignment error is at the level of approximately 2 milliseconds. Subsequently, we performed frame extraction from the videos, again utilizing the GoPro camera parameters as priors to complete the calibration." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 522, + 295, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 654 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 654 + ], + "type": "text", + "content": "It is important to note that for the dynamic scene data captured with a moving rig, we also performed time code alignment. However, this data poses significant challenges for existing calibration methods, often resulting in errors and floaters. There are also no solutions for reconstruction algorithms specifically addressing dynamic scenes from moving rig currently. Therefore, we do not provide the calibrated results for this type of data at this time. But we believe that this data will greatly contribute to the advancement of the field, and thus we will also publicly releasing the data which has been time-synchronized." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 660, + 157, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 660, + 157, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 660, + 157, + 673 + ], + "type": "text", + "content": "3.3. Dataset Analysis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "Diversity. Our dataset, as depicted in Table 2, includes 7 common indoor and outdoor scenes, each comprising one or more takes, totaling 16 video sequences. Each scene is" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 322, + 72, + 550, + 203 + ], + "blocks": [ + { + "bbox": [ + 322, + 72, + 550, + 203 + ], + "lines": [ + { + "bbox": [ + 322, + 72, + 550, + 203 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 550, + 203 + ], + "type": "image", + "image_path": "db07f5f4f664c27676e09bbac34031a49fb20ca9a54d32481c233c0850fb1735.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 281, + 555, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 555, + 353 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 555, + 353 + ], + "type": "text", + "content": "rich in foreground and background elements, meticulously crafted to preserve the authenticity of the environment. The characters, clothing, and motions within each scene exhibit a diverse range, set against varying ambient lighting conditions. We also included areas of interest like tables, chairs, windows as much as possible for researchers." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 354, + 555, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 555, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 555, + 389 + ], + "type": "text", + "content": "Quality. Leveraging a total of 39 cameras, we captured dynamic scenes at a resolution of " + }, + { + "bbox": [ + 313, + 354, + 555, + 389 + ], + "type": "inline_equation", + "content": "5312 \\times 2988" + }, + { + "bbox": [ + 313, + 354, + 555, + 389 + ], + "type": "text", + "content": " and 60 FPS. The images of static environment is " + }, + { + "bbox": [ + 313, + 354, + 555, + 389 + ], + "type": "inline_equation", + "content": "5568 \\times 4176" + }, + { + "bbox": [ + 313, + 354, + 555, + 389 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 391, + 554, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 391, + 554, + 450 + ], + "spans": [ + { + "bbox": [ + 313, + 391, + 554, + 450 + ], + "type": "text", + "content": "Duration. As outlined in Table 2, each scene in our dataset contains at least 1 minute of footage. The longest sequence in our forthcoming public dataset spans approximately 5 minutes. The total processed data duration amounts to 38 minutes 46 seconds, including 139,560 frames per camera." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 451, + 555, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 451, + 555, + 583 + ], + "spans": [ + { + "bbox": [ + 313, + 451, + 555, + 583 + ], + "type": "text", + "content": "Coverage. As highlighted in Table 2, we conducted image acquisition from over 1,000 viewpoints before recording dynamic content, thereby offering a " + }, + { + "bbox": [ + 313, + 451, + 555, + 583 + ], + "type": "inline_equation", + "content": "180^{\\circ} - 360^{\\circ}" + }, + { + "bbox": [ + 313, + 451, + 555, + 583 + ], + "type": "text", + "content": " viewing space in VR. For scenes employing the mobile shooting strategy, we introduced a metric termed 'spatiotemporal capture density,' quantifying the captured volume covered by our rig per second, as shown in Figure 4. This metric, unavailable in current datasets (where static camera arrays are deemed 0 and handheld monocular cameras follow a single trajectory without capturing volume), registers at approximately " + }, + { + "bbox": [ + 313, + 451, + 555, + 583 + ], + "type": "inline_equation", + "content": "0.10~m^3 /s" + }, + { + "bbox": [ + 313, + 451, + 555, + 583 + ], + "type": "text", + "content": " in our dataset." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 597, + 485, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 597, + 485, + 611 + ], + "spans": [ + { + "bbox": [ + 313, + 597, + 485, + 611 + ], + "type": "text", + "content": "4. Benchmarks and Experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "text", + "content": "To comprehensively assess the effectiveness of the dataset, we selected three scenes (2 indoor and 1 outdoor) to test the latest dynamic light field reconstruction algorithms, providing a benchmark for research in areas such as novel view synthesis. We also proposed an improvement strategy based on one of the methods. Furthermore, we presented and validated a feasible pipeline for constructing 6-DoF sound fields from multi-view audio sequences. In this section, we" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16558" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 112, + 553, + 212 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "text", + "content": "Table 2. The detailed statistics of our ImViD dataset. #Take represents the number of different episodes we collected in this scene. #Strategy represents the strategies used in Step2. 1 indicates only (i) was used in Step 2, 2 indicates that both were used. Avg.S-T Density indicates the spatiotemporal coverage density of our collection method, measured by the captured volume size per second." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 112, + 553, + 212 + ], + "lines": [ + { + "bbox": [ + 56, + 112, + 553, + 212 + ], + "spans": [ + { + "bbox": [ + 56, + 112, + 553, + 212 + ], + "type": "table", + "html": "
Name#Camera#Static viewpoints#Take#StrategyAvg.S-T Density(m3/s)Viewing SpaceAvg.DurationStorage(GB)
Scene1 Opera39115221-180°3min22s226
Scene2 Laboratory391225220.10360°1min42s137.3
Scene3 Classroom391223220.10360°4min42s497
Scene4 Meeting39122311-360°3min16s114
Scene5 Rendition391620420.10360°2min02s516
Scene6 Puppy391404320.10360°1min50s359
Scene7 Playing391224220.10360°1min10s220
Total--16---38min46s2069.3
", + "image_path": "13c7d7d3108da5e76b322c474f06734c8e86d96140d1e0ec46f93bb6fa5ef87e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 231, + 296, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 231, + 296, + 268 + ], + "spans": [ + { + "bbox": [ + 55, + 231, + 296, + 268 + ], + "type": "text", + "content": "will first introduce the selected light field pipeline and our proposed sound field pipeline (Sec 4.1), then provide details on the experimental setup (Sec 4.2)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 274, + 118, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 274, + 118, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 274, + 118, + 285 + ], + "type": "text", + "content": "4.1. Baseline" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 291, + 296, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 291, + 296, + 352 + ], + "spans": [ + { + "bbox": [ + 55, + 291, + 296, + 352 + ], + "type": "text", + "content": "Dynamic Light Field Reconstruction. We conducted a thorough survey of the current dynamic reconstruction algorithms, prioritizing rendering speed and support for multiview data input. We finally decided to establish a benchmark on 3DGS-based method from three paradigms." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 351, + 296, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 351, + 296, + 435 + ], + "spans": [ + { + "bbox": [ + 55, + 351, + 296, + 435 + ], + "type": "text", + "content": "4DGS [55] introduces Hexplane to encode spatiotemporal information to feature " + }, + { + "bbox": [ + 55, + 351, + 296, + 435 + ], + "type": "inline_equation", + "content": "f_{d}" + }, + { + "bbox": [ + 55, + 351, + 296, + 435 + ], + "type": "text", + "content": " and use MLPs decoders " + }, + { + "bbox": [ + 55, + 351, + 296, + 435 + ], + "type": "inline_equation", + "content": "D = \\{\\phi_{x},\\phi_{r},\\phi_{s}\\}" + }, + { + "bbox": [ + 55, + 351, + 296, + 435 + ], + "type": "text", + "content": " to predict gaussians' changes in position " + }, + { + "bbox": [ + 55, + 351, + 296, + 435 + ], + "type": "inline_equation", + "content": "\\Delta \\chi = \\phi_{x}(f_{d})" + }, + { + "bbox": [ + 55, + 351, + 296, + 435 + ], + "type": "text", + "content": ", rotation " + }, + { + "bbox": [ + 55, + 351, + 296, + 435 + ], + "type": "inline_equation", + "content": "\\Delta r = \\phi_r(f_d)" + }, + { + "bbox": [ + 55, + 351, + 296, + 435 + ], + "type": "text", + "content": ", and scaling " + }, + { + "bbox": [ + 55, + 351, + 296, + 435 + ], + "type": "inline_equation", + "content": "\\Delta s = \\phi_s(f_d)" + }, + { + "bbox": [ + 55, + 351, + 296, + 435 + ], + "type": "text", + "content": ". Utilizing the sparse point cloud from frame 0 as initialization and training static gaussians for the first 3000 iterations, the dynamic scene " + }, + { + "bbox": [ + 55, + 351, + 296, + 435 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 55, + 351, + 296, + 435 + ], + "type": "text", + "content": " is represented as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 84, + 443, + 294, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 443, + 294, + 456 + ], + "spans": [ + { + "bbox": [ + 84, + 443, + 294, + 456 + ], + "type": "interline_equation", + "content": "G \\left(\\chi^ {\\prime}, r ^ {\\prime}, s ^ {\\prime}\\right) = G \\left(\\chi + \\Delta \\chi , r + \\Delta r, s + \\Delta s\\right) \\tag {1}", + "image_path": "bf8a471cb2cc249814f0e0596951ce0fb1e3e78ae33689bc22b7157be55f2617.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 464, + 296, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 464, + 296, + 513 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 296, + 513 + ], + "type": "text", + "content": "4DrotorGS [12] is a representative method of dimensionality expansion. It extends 3DGS to a four-dimensional Gaussian ellipsoid by introducing the concept of rotor, expanding the covariance matrix to four dimensions, i.e.:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 100, + 520, + 294, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 520, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 100, + 520, + 294, + 536 + ], + "type": "interline_equation", + "content": "G _ {4 D} (x) = e ^ {- \\frac {1}{2} (x - \\mu_ {4 D}) ^ {T} \\Sigma_ {4 D} ^ {- 1} (x - \\mu_ {4 D})} \\tag {2}", + "image_path": "7f65a4422d79727e0b91c445e6ebb644b388f7768c21558cf36af446d6d569ef.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 544, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 567 + ], + "type": "text", + "content": "Each frame of the scene corresponds to a time slice of the 4D Gaussian ellipsoids followed by " + }, + { + "bbox": [ + 55, + 544, + 295, + 567 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 544, + 295, + 567 + ], + "type": "text", + "content": "-blending." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 567, + 296, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 567, + 296, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 296, + 639 + ], + "type": "text", + "content": "STG [30], on the other hand, use polynomial fitting for the motion and rotation of the 3DGS, and employs an RBF function to model the opacity changes over time. It concatenates the sparse point clouds from all frames as the initialization, and the final representation of each point in the scene is:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 60, + 645, + 294, + 677 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 645, + 294, + 677 + ], + "spans": [ + { + "bbox": [ + 60, + 645, + 294, + 677 + ], + "type": "interline_equation", + "content": "\\alpha_ {i} (t) = \\sigma_ {i} (t) \\exp \\left(- \\frac {1}{2} \\left(x - \\mu_ {i} (t)\\right) ^ {T} \\Sigma_ {i} (t) ^ {- 1} \\left(x - \\mu_ {i} (t)\\right)\\right) \\tag {3}", + "image_path": "f2e2f667c9e10fa49f93c7335eb44d4f8602418e807cc0fdbfe9facf2b49205d.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Although we ensured consistent white balance and other settings across cameras based on the characteristics of scenes during shooting, color differences still exist due to" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 231, + 555, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 231, + 555, + 363 + ], + "spans": [ + { + "bbox": [ + 313, + 231, + 555, + 363 + ], + "type": "text", + "content": "factors like occlusions when capturing from different angles. During experiments, we found that while existing methods fit well to the training views, there is a noticeable color shift during viewpoint transitions in the reconstructed light field, manifesting as flickering and floaters. This is especially apparent when training in segments, where the differences between segments are more pronounced, significantly reducing immersion. We improved the original STG (named as " + }, + { + "bbox": [ + 313, + 231, + 555, + 363 + ], + "type": "inline_equation", + "content": "\\mathrm{STG}++" + }, + { + "bbox": [ + 313, + 231, + 555, + 363 + ], + "type": "text", + "content": ") by introducing global, time-invariant color mapping during training. Detailed implementation can be seen in supplementary materials." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 366, + 555, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 366, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 366, + 555, + 510 + ], + "type": "text", + "content": "Sound Field Reconstruction. Taking the recording microphone as the origin of the coordinate system, we denote the coordinates of the sound source as " + }, + { + "bbox": [ + 313, + 366, + 555, + 510 + ], + "type": "inline_equation", + "content": "s(t) = \\{x_s(t), y_s(t)\\}" + }, + { + "bbox": [ + 313, + 366, + 555, + 510 + ], + "type": "text", + "content": ", the user wearing the VR device as " + }, + { + "bbox": [ + 313, + 366, + 555, + 510 + ], + "type": "inline_equation", + "content": "l(t) = \\{x_l(t), y_l(t)\\}" + }, + { + "bbox": [ + 313, + 366, + 555, + 510 + ], + "type": "text", + "content": ", and the angle of deviation from the y-axis in the counterclockwise direction as " + }, + { + "bbox": [ + 313, + 366, + 555, + 510 + ], + "type": "inline_equation", + "content": "\\theta_l(t)" + }, + { + "bbox": [ + 313, + 366, + 555, + 510 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 366, + 555, + 510 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 366, + 555, + 510 + ], + "type": "text", + "content": " is the time index. Based on these premises, generating the user's binaural spatial audio can be divided into three parts: sound direction mapping, sound distance mapping, and sound auralization. Note that we assume there is only one dominating sound source (omnidirectional emitting) to simplify the sound field reconstruction." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 510, + 554, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 554, + 535 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 554, + 535 + ], + "type": "text", + "content": "In the sound direction mapping part, we calculate the direction of the sound source relative to the user's ears, " + }, + { + "bbox": [ + 313, + 510, + 554, + 535 + ], + "type": "inline_equation", + "content": "\\theta_{s}(t)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 371, + 543, + 553, + 569 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 371, + 543, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 371, + 543, + 553, + 569 + ], + "type": "interline_equation", + "content": "\\theta_ {s} (t) = \\operatorname {a r c c o s} \\frac {\\mathbf {v} _ {1} (t) \\mathbf {v} _ {2} (t)}{| \\mathbf {v} _ {1} (t) | | \\mathbf {v} _ {2} (t) |}, \\tag {4}", + "image_path": "c74f8b538536264186edd5fc28d34bbb86ad2fe862534729c2eb2b7d85ac7fa5.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 571, + 342, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 571, + 342, + 582 + ], + "spans": [ + { + "bbox": [ + 315, + 571, + 342, + 582 + ], + "type": "text", + "content": "where" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 352, + 590, + 553, + 604 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 590, + 553, + 604 + ], + "spans": [ + { + "bbox": [ + 352, + 590, + 553, + 604 + ], + "type": "interline_equation", + "content": "\\mathbf {v} _ {1} (t) = \\left[ x _ {s} (t) - x _ {l} (t), y _ {s} (t) - y _ {l} (t) \\right] ^ {T}, \\tag {5}", + "image_path": "4b3305a23c93174cc28e93fb9fc2c438deb915f443e04a388ec31c4a3a5812cd.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 353, + 605, + 553, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 605, + 553, + 620 + ], + "spans": [ + { + "bbox": [ + 353, + 605, + 553, + 620 + ], + "type": "interline_equation", + "content": "\\mathbf {v} _ {2} (t) = \\left[ - \\sin \\left(\\theta_ {l} (t)\\right), \\cos \\left(\\theta_ {l} (t)\\right) \\right] ^ {T}, \\tag {6}", + "image_path": "4bc350af5730e7f02fd476d9676867a55add4964d9d45e6007535916707e93e0.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 627, + 554, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 627, + 554, + 675 + ], + "spans": [ + { + "bbox": [ + 313, + 627, + 554, + 675 + ], + "type": "text", + "content": "In the sound distance mapping part, we calculate the scaling of the sound source when it reaches the user's ears, relative to the original recording. We denote this scaling parameter as " + }, + { + "bbox": [ + 313, + 627, + 554, + 675 + ], + "type": "inline_equation", + "content": "\\lambda(t)" + }, + { + "bbox": [ + 313, + 627, + 554, + 675 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 339, + 681, + 553, + 712 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 681, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 339, + 681, + 553, + 712 + ], + "type": "interline_equation", + "content": "\\lambda (t) = \\frac {\\sqrt {x _ {s} ^ {2} (t) + y _ {s} ^ {2} (t)}}{\\sqrt {(x _ {s} (t) - x _ {l} (t)) ^ {2} + (y _ {s} (t) - y _ {l} (t)) ^ {2}}} \\tag {7}", + "image_path": "8467b3b3d529527286e14738ee1e8da5dbf525451118757275118af231c5f8c1.jpg" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16559" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "text", + "content": "In the sound auralization part, we convert the original recording into spatial audio based on " + }, + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "inline_equation", + "content": "\\theta_s(t)" + }, + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "inline_equation", + "content": "\\lambda (t)" + }, + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "text", + "content": ". In the short-time Fourier transformation (STFT) domain, we denote the audio from the left and right ears as " + }, + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "inline_equation", + "content": "A_{L}(t,f)" + }, + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "inline_equation", + "content": "A_{R}(t,f)" + }, + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "text", + "content": ", respectively, where " + }, + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "text", + "content": " is the frequency index. Then " + }, + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "inline_equation", + "content": "A_{L}(t,f)" + }, + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "inline_equation", + "content": "A_{R}(t,f)" + }, + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "text", + "content": " can be written as" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 94, + 154, + 295, + 167 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 154, + 295, + 167 + ], + "spans": [ + { + "bbox": [ + 94, + 154, + 295, + 167 + ], + "type": "interline_equation", + "content": "A _ {L} (t, f) = F _ {L} \\left(\\theta_ {s} (t)\\right) F _ {l} (t, f) A _ {O} (t, f) \\tag {8}", + "image_path": "9d34684608e7534f89552ea7e708befcf04a9d096a743bf6de6a805614009ca7.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 94, + 168, + 295, + 182 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 168, + 295, + 182 + ], + "spans": [ + { + "bbox": [ + 94, + 168, + 295, + 182 + ], + "type": "interline_equation", + "content": "A _ {R} (t, f) = F _ {R} \\left(\\theta_ {s} (t)\\right) F _ {l} (t, f) A _ {O} (t, f), \\tag {9}", + "image_path": "d12855877d7dd4ee25612dbe0f7bf8623e02d024200bd5bf78ab3121da5c6d9e.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 190, + 296, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 190, + 296, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 190, + 296, + 300 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 190, + 296, + 300 + ], + "type": "inline_equation", + "content": "F_{L}(\\theta_{s}(t))" + }, + { + "bbox": [ + 55, + 190, + 296, + 300 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 190, + 296, + 300 + ], + "type": "inline_equation", + "content": "F_{R}(\\theta_{s}(t))" + }, + { + "bbox": [ + 55, + 190, + 296, + 300 + ], + "type": "text", + "content": " are head-related transfer functions (HRTF) based on " + }, + { + "bbox": [ + 55, + 190, + 296, + 300 + ], + "type": "inline_equation", + "content": "\\theta_s(t)" + }, + { + "bbox": [ + 55, + 190, + 296, + 300 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 190, + 296, + 300 + ], + "type": "inline_equation", + "content": "F_{l}(t)" + }, + { + "bbox": [ + 55, + 190, + 296, + 300 + ], + "type": "text", + "content": " denotes the room impulse response (RIR) at the user's location and " + }, + { + "bbox": [ + 55, + 190, + 296, + 300 + ], + "type": "inline_equation", + "content": "A_{O}(t,f)" + }, + { + "bbox": [ + 55, + 190, + 296, + 300 + ], + "type": "text", + "content": " is the sound emitted from the speaker in the STFT domain. We utilize " + }, + { + "bbox": [ + 55, + 190, + 296, + 300 + ], + "type": "inline_equation", + "content": "\\theta_{s}(t)" + }, + { + "bbox": [ + 55, + 190, + 296, + 300 + ], + "type": "text", + "content": " to retrieve " + }, + { + "bbox": [ + 55, + 190, + 296, + 300 + ], + "type": "inline_equation", + "content": "F_{L}(\\theta_{s}(t))" + }, + { + "bbox": [ + 55, + 190, + 296, + 300 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 190, + 296, + 300 + ], + "type": "inline_equation", + "content": "F_{R}(\\theta_{s}(t))" + }, + { + "bbox": [ + 55, + 190, + 296, + 300 + ], + "type": "text", + "content": " from the SADIE II dataset [3]. In order to model " + }, + { + "bbox": [ + 55, + 190, + 296, + 300 + ], + "type": "inline_equation", + "content": "F_{l}(t)" + }, + { + "bbox": [ + 55, + 190, + 296, + 300 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 190, + 296, + 300 + ], + "type": "inline_equation", + "content": "A_{O}(t,f)" + }, + { + "bbox": [ + 55, + 190, + 296, + 300 + ], + "type": "text", + "content": ", we make the following assumptions: " + }, + { + "bbox": [ + 55, + 190, + 296, + 300 + ], + "type": "inline_equation", + "content": "F_{l}(t)" + }, + { + "bbox": [ + 55, + 190, + 296, + 300 + ], + "type": "text", + "content": " change between different locations at the same time index " + }, + { + "bbox": [ + 55, + 190, + 296, + 300 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 190, + 296, + 300 + ], + "type": "text", + "content": " only depending on the parameter " + }, + { + "bbox": [ + 55, + 190, + 296, + 300 + ], + "type": "inline_equation", + "content": "\\lambda (t)" + }, + { + "bbox": [ + 55, + 190, + 296, + 300 + ], + "type": "text", + "content": ", i.e.," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 127, + 308, + 295, + 322 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 308, + 295, + 322 + ], + "spans": [ + { + "bbox": [ + 127, + 308, + 295, + 322 + ], + "type": "interline_equation", + "content": "F _ {l} (t, f) = \\lambda (t) F (t, f), \\tag {10}", + "image_path": "1dfb7da7de97acba26d9214e1ba076ed148905f41733f06209fa8456e41c61c5.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 330, + 296, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 330, + 296, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 330, + 296, + 354 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 330, + 296, + 354 + ], + "type": "inline_equation", + "content": "F(t)" + }, + { + "bbox": [ + 55, + 330, + 296, + 354 + ], + "type": "text", + "content": " is the RIR at the recording microphone. Using the above assumption, Eq.(8) and Eq.(9) can be rewritten as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 104, + 376, + 295, + 389 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 376, + 295, + 389 + ], + "spans": [ + { + "bbox": [ + 104, + 376, + 295, + 389 + ], + "type": "interline_equation", + "content": "A _ {L} (t, f) = \\lambda (t) F _ {L} \\left(\\theta_ {s} (t)\\right) A (t, f), \\tag {11}", + "image_path": "6426c4c1d09cf974dd7eb93e068da89f186f403920c2b40dbc62c8ed00d9b32d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 391, + 295, + 404 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 391, + 295, + 404 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 295, + 404 + ], + "type": "interline_equation", + "content": "A _ {R} (t, f) = \\lambda (t) F _ {R} \\left(\\theta_ {s} (t)\\right) A (t, f), \\tag {12}", + "image_path": "3dc4a817b8601e2a2308ed9a5c3f231debc0cd90b5fbcdcddc79c14f560f0adc.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 413, + 296, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 413, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 296, + 437 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 413, + 296, + 437 + ], + "type": "inline_equation", + "content": "A(t,f) = F(t,f)A_O(t,f)" + }, + { + "bbox": [ + 55, + 413, + 296, + 437 + ], + "type": "text", + "content": " is the audio at the recording microphone." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 445, + 182, + 457 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 182, + 457 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 182, + 457 + ], + "type": "text", + "content": "4.2. Experimental Settings" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 462, + 296, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 296, + 521 + ], + "type": "text", + "content": "Training setups. The training for all light field data was completed on a single NVIDIA A100 GPU. Due to memory limitations, we split the video and trained a model every 60 frames, total 300 frames. We also followed the original train-test-split strategy of the baseline for our experiments." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 522, + 296, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 296, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 296, + 558 + ], + "type": "text", + "content": "Metric. We analyzed the quality of dynamic light field reconstruction using the most common evaluation metrics PSNR, SSIM, and LPIPS(alex)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 569, + 177, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 569, + 177, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 569, + 177, + 582 + ], + "type": "text", + "content": "5. Results and Analysis" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 588, + 218, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 218, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 218, + 601 + ], + "type": "text", + "content": "5.1. Validation of Light Field Data" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": "Quantitative Results. As illustrated in Table 3, 4 and Figure 5, our indoor and outdoor data both yield correct results across three 3DGS-based paradigms. STG++ outperforms all baselines in all scenes, delivering slightly better results, the highest rendering speed, and a moderate model size. We also found that, since 4DGS trains using the sparse point cloud from the 0th frame and limits the number of final points, its performance is significantly lower than other pipelines on texture-richer outdoor scenes, such as Scene6" + } + ] + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 315, + 102, + 553, + 150 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 555, + 92 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 555, + 92 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 555, + 92 + ], + "type": "text", + "content": "Table 3. Test views performance of 3DGS-based dynamic scene reconstruction method on ImViD dataset." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 102, + 553, + 150 + ], + "lines": [ + { + "bbox": [ + 315, + 102, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 315, + 102, + 553, + 150 + ], + "type": "table", + "html": "
MethodScene1 Opera_girlScene2 LaboratoryScene6 Puppy
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
4DGS23.2270.7530.41025.7980.8900.17618.1210.2220.711
4DRotor27.2630.7750.32828.0070.9170.09817.9160.2980.331
STG28.4820.7860.28726.3060.9100.11420.4970.5940.211
STG++31.2400.7990.27727.5810.9160.10720.5330.5970.202
", + "image_path": "e4ee11162c2bc5a61527694bed825afd2cbfbb2f99583afd623fda4886d179c6.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "type": "table", + "bbox": [ + 315, + 193, + 553, + 254 + ], + "blocks": [ + { + "bbox": [ + 313, + 162, + 555, + 184 + ], + "lines": [ + { + "bbox": [ + 313, + 162, + 555, + 184 + ], + "spans": [ + { + "bbox": [ + 313, + 162, + 555, + 184 + ], + "type": "text", + "content": "Table 4. The storage size of training results from different methods and their rendering speed on a single A100 GPU. *CPU version." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 193, + 553, + 254 + ], + "lines": [ + { + "bbox": [ + 315, + 193, + 553, + 254 + ], + "spans": [ + { + "bbox": [ + 315, + 193, + 553, + 254 + ], + "type": "table", + "html": "
Scene1 Opera_girlScene2 LaboratoryScene6 Puppy
Mem(MB)FPSMem(MB)FPSMem(MB)FPS
4DGS101.0546.22105.0244.93101.6541.86
4DRotor*5818.611.006867.180.674415.820.56
STG382.4196.10318.41133.651287.55100.00
STG++387.17108.89329.07137.781276.21110.47
", + "image_path": "ceff96a6c471a2195677a2040d93fe1f4d9f2a35b7204a003797fcd1f700f62c.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_body" + } + ], + "index": 18 + }, + { + "type": "table", + "bbox": [ + 315, + 284, + 553, + 350 + ], + "blocks": [ + { + "bbox": [ + 339, + 264, + 529, + 275 + ], + "lines": [ + { + "bbox": [ + 339, + 264, + 529, + 275 + ], + "spans": [ + { + "bbox": [ + 339, + 264, + 529, + 275 + ], + "type": "text", + "content": "Table 5. User study for the sound field construction." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 284, + 553, + 350 + ], + "lines": [ + { + "bbox": [ + 315, + 284, + 553, + 350 + ], + "spans": [ + { + "bbox": [ + 315, + 284, + 553, + 350 + ], + "type": "table", + "html": "
Auditory Spatial PerceptionSound QualityImmersiveness
Very poor0.00%0.00%0.00%
Poor0.00%0.00%0.00%
Fair14.28%19.04%9.52%
Good23.80%33.33%42.85%
Excellent61.90%47.61%47.61%
", + "image_path": "5ad60dc675aff4495b9e2776d12cd328b8843c0438b35785af5148d24eced6c8.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "table_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 369, + 555, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 555, + 430 + ], + "type": "text", + "content": "puppy. Additionally, due to the density control in the time dimension implemented by 4Drotor, it performs better in frames and regions with more significant motion, which is an aspect worth referencing and adopting in future work. More results can be seen in the supplementary material." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 435, + 554, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 435, + 554, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 435, + 554, + 495 + ], + "type": "text", + "content": "Qualitative Results. As shown in Figure 6, Our STG++ achieves better temporal continuity in the color of pixels at the same location. There are almost no abrupt transitions both within and between segments. Details are showed in the supplementary for a more intuitive experience." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 503, + 495, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 503, + 495, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 503, + 495, + 515 + ], + "type": "text", + "content": "5.2. Effectiveness of Sound Field Data" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 521, + 554, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 554, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 554, + 592 + ], + "type": "text", + "content": "Quantitative Results. Since our sound field reconstruction algorithm is non-trainable, the reconstructed spatial audio lacks a corresponding ground truth, making it difficult to evaluate the results using quantitative metrics. Therefore, we conducted a user study with 21 experts to assess the performance of the sound field reconstruction." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 594, + 555, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 700 + ], + "type": "text", + "content": "Each participant rated the sound field reconstruction on three dimensions: auditory spatial perception, sound quality, and immersiveness. Auditory spatial perception refers to the listener's ability to perceive the distribution and localization of sound in space. Sound quality refers to the audio quality of the spatial audio compared to the corresponding microphone-recorded signal. Immersiveness refers to the listener's sense of being in a space where the sound source genuinely exists." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 326, + 701, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 555, + 713 + ], + "type": "text", + "content": "The results of the user study are shown in Table 5, with" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16560" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 72, + 544, + 242 + ], + "blocks": [ + { + "bbox": [ + 61, + 72, + 544, + 242 + ], + "lines": [ + { + "bbox": [ + 61, + 72, + 544, + 242 + ], + "spans": [ + { + "bbox": [ + 61, + 72, + 544, + 242 + ], + "type": "image", + "image_path": "d53e3b037fabb6a88cf96a14c862440b4201c59e4bcbde094403f8d8efcff0c0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 82, + 255, + 526, + 266 + ], + "lines": [ + { + "bbox": [ + 82, + 255, + 526, + 266 + ], + "spans": [ + { + "bbox": [ + 82, + 255, + 526, + 266 + ], + "type": "text", + "content": "Figure 5. Comparison of the rendering results of four baselines on Scene 1 Opera, Scene 2 Laboratory, and Scene 6 Puppy." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 286, + 294, + 425 + ], + "blocks": [ + { + "bbox": [ + 58, + 286, + 294, + 425 + ], + "lines": [ + { + "bbox": [ + 58, + 286, + 294, + 425 + ], + "spans": [ + { + "bbox": [ + 58, + 286, + 294, + 425 + ], + "type": "image", + "image_path": "da0ead87a4c48d8d85119c17848f637e6c43694718d0bc7e740bd8bab431e61c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 435, + 295, + 468 + ], + "lines": [ + { + "bbox": [ + 55, + 435, + 295, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 435, + 295, + 468 + ], + "type": "text", + "content": "Figure 6. The continuity of pixels at the same location across different frames and segments. The visualizations present the RGB (middle) and brightness (right) variations." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 496, + 295, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 496, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 496, + 295, + 579 + ], + "type": "text", + "content": "numbers as percentages of participants. The majority of the participants " + }, + { + "bbox": [ + 55, + 496, + 295, + 579 + ], + "type": "inline_equation", + "content": "(61.90\\%)" + }, + { + "bbox": [ + 55, + 496, + 295, + 579 + ], + "type": "text", + "content": " rated the auditory spatial perception as excellent, " + }, + { + "bbox": [ + 55, + 496, + 295, + 579 + ], + "type": "inline_equation", + "content": "80.94\\%" + }, + { + "bbox": [ + 55, + 496, + 295, + 579 + ], + "type": "text", + "content": " participants felt that the generated spatial audio did not exhibit significant degradation in quality, and " + }, + { + "bbox": [ + 55, + 496, + 295, + 579 + ], + "type": "inline_equation", + "content": "90.46\\%" + }, + { + "bbox": [ + 55, + 496, + 295, + 579 + ], + "type": "text", + "content": " found the audio immersive, which demonstrates the effectiveness of our data capture and sound field reconstruction methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 597, + 214, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 214, + 609 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 214, + 609 + ], + "type": "text", + "content": "5.3. Multimodal VR Experiences." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": "We finally integrated light field and sound field reconstruction results to achieve a 6-DoF multimodal immersive VR experience with real-time rendering speed at 60 FPS on a single 3090 GPU. As shown in Figure 7, the visual and auditory experiences correspond uniquely to the user's position and viewing direction. The experience can be further explored by listening through headphones with the supplementary video provided." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 326, + 289, + 540, + 462 + ], + "blocks": [ + { + "bbox": [ + 326, + 289, + 540, + 462 + ], + "lines": [ + { + "bbox": [ + 326, + 289, + 540, + 462 + ], + "spans": [ + { + "bbox": [ + 326, + 289, + 540, + 462 + ], + "type": "image", + "image_path": "2d974acfdb4ab91af8a93b9c3e0ea54ca4f3aec0a5de2203e0c84586ec2a33b5.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 472, + 555, + 506 + ], + "lines": [ + { + "bbox": [ + 313, + 472, + 555, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 555, + 506 + ], + "type": "text", + "content": "Figure 7. Visualization of the interaction trajectory and corresponding visual&auditory results. Note that the multi-modal feedback changes consistently with the positions and orientations." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 525, + 388, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 388, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 388, + 538 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 545, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 545, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 545, + 555, + 713 + ], + "type": "text", + "content": "In this work, we introduce ImViD, to our knowledge, the first dataset for immersive volumetric video. This dataset encompasses seven large indoor and outdoor scenes, featuring rich foreground and background content that showcases engaging and commonly encountered daily scenarios, along with high-quality synchronized audio sequences. We validated the effectiveness of the dataset using the latest dynamic light field reconstruction algorithms, providing both quantitative and qualitative analyses of current limitations and potential improvements. Moreover, we propose a base pipeline for constructing immersive volumetric video from this data. We believe ImViD will facilitate the exploration of the limitations of current volumetric video reconstruction algorithms and drive immersive VR/AR experiences." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16561" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 178, + 115, + 192 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 178, + 115, + 192 + ], + "spans": [ + { + "bbox": [ + 56, + 178, + 115, + 192 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 198, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 198, + 296, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 198, + 296, + 232 + ], + "spans": [ + { + "bbox": [ + 61, + 198, + 296, + 232 + ], + "type": "text", + "content": "[1] https://www.apple.com/newsroom/2024/07/new-apple-immersive-video-series-and-films-premiere-on-vision-pro/.1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 233, + 208, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 233, + 208, + 243 + ], + "spans": [ + { + "bbox": [ + 61, + 233, + 208, + 243 + ], + "type": "text", + "content": "[2] https://www.gracia.ai/.1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 245, + 295, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 245, + 295, + 289 + ], + "spans": [ + { + "bbox": [ + 61, + 245, + 295, + 289 + ], + "type": "text", + "content": "[3] Cal Armstrong, Lewis Thresh, Damian Murphy, and Gavin Kearney. A perceptual evaluation of individual and non-individual hrfts: A case study of the sadie ii database. Applied Sciences, 8(11):2029, 2018. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 290, + 296, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 290, + 296, + 355 + ], + "spans": [ + { + "bbox": [ + 61, + 290, + 296, + 355 + ], + "type": "text", + "content": "[4] Benjamin Attal, Jia-Bin Huang, Christian Richardt, Michael Zollhoefer, Johannes Kopf, Matthew O'Toole, and Changil Kim. Hyperreel: High-fidelity 6-dof video with ray-conditioned sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16610-16620, 2023. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 357, + 295, + 400 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 357, + 295, + 400 + ], + "spans": [ + { + "bbox": [ + 62, + 357, + 295, + 400 + ], + "type": "text", + "content": "[5] Swapnil Bhosale, Haosen Yang, Diptesh Kanojia, Jiankang Deng, and Xiatian Zhu. Av-gs: Learning material and geometry aware priors for novel view acoustic synthesis. arXiv preprint arXiv:2406.08920, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 402, + 296, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 402, + 296, + 465 + ], + "spans": [ + { + "bbox": [ + 62, + 402, + 296, + 465 + ], + "type": "text", + "content": "[6] Michael Broxton, John Flynn, Ryan Overbeck, Daniel Erickson, Peter Hedman, Matthew Duvall, Jason Dourgarian, Jay Busch, Matt Whalen, and Paul Debevec. Immersive light field video with a layered mesh representation. ACM Transactions on Graphics (TOG), 39(4):86-1, 2020. 1, 2, 3, 4, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 468, + 296, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 468, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 62, + 468, + 296, + 510 + ], + "type": "text", + "content": "[7] Amandine Brunetto, Sascha Hornauer, and Fabien Moutarde. Neraf: 3d scene infused neural radiance and acoustic fields. arXiv preprint arXiv:2405.18213, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 514, + 295, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 514, + 295, + 557 + ], + "spans": [ + { + "bbox": [ + 62, + 514, + 295, + 557 + ], + "type": "text", + "content": "[8] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 130-141, 2023. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 558, + 296, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 558, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 62, + 558, + 296, + 612 + ], + "type": "text", + "content": "[9] Hao Chen, Yuqi Hou, Chenyuan Qu, Irene Testini, Xiaohan Hong, and Jianbo Jiao. " + }, + { + "bbox": [ + 62, + 558, + 296, + 612 + ], + "type": "inline_equation", + "content": "360+\\mathrm{x}" + }, + { + "bbox": [ + 62, + 558, + 296, + 612 + ], + "type": "text", + "content": ": A panoptic multimodal scene understanding dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19373-19382, 2024. 2, 3, 1, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 614, + 295, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 614, + 295, + 658 + ], + "spans": [ + { + "bbox": [ + 57, + 614, + 295, + 658 + ], + "type": "text", + "content": "[10] Kai Cheng, Xiaoxiao Long, Kaizhi Yang, Yao Yao, Wei Yin, Yuexin Ma, Wenping Wang, and Xuejin Chen. Gaussianpro: 3d gaussian splatting with progressive propagation. arXiv preprint arXiv:2402.14650, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 658, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 296, + 712 + ], + "type": "text", + "content": "[11] Alvaro Collet, Ming Chuang, Pat Sweeney, Don Gillett, Dennis Evseev, David Calabrese, Hugues Hoppe, Adam Kirk, and Steve Sullivan. High-quality streamable free-viewpoint video. ACM Transactions on Graphics (ToG), 34(4):1-13, 2015. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 57, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 55, + 90, + 296, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 296, + 159 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 296, + 159 + ], + "type": "text", + "content": "This work was supported by National Key R&D Program of China (No.2022YFF0902204), NSFC (No.62171255, 52425801), Tsinghua University-Migu Xinkong Culture Technology (Xiamen) Co.Ltd. Joint Research Center for Intelligent Light Field and Interaction Technology, Guoqiang Institute of Tsinghua University (No.2021GQG0001)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "text", + "content": "[12] Yuanxing Duan, Fangyin Wei, Qiyu Dai, Yuhang He, Wenzheng Chen, and Baoquan Chen. 4d-rotor gaussian splatting: Towards efficient novel view synthesis for dynamic scenes. In ACM SIGGRAPH 2024 Conference Papers, pages 1–11, 2024. 3, 4, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 129, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 183 + ], + "type": "text", + "content": "[13] Bardienus P Duisterhof, Zhao Mandi, Yunchao Yao, Jia-Wei Liu, Mike Zheng Shou, Shuran Song, and Jeffrey Ichnowski. Md-splating: Learning metric deformation from 4d gaussians in highly deformable scenes. arXiv preprint arXiv:2312.00583, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 184, + 553, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 184, + 553, + 247 + ], + "spans": [ + { + "bbox": [ + 317, + 184, + 553, + 247 + ], + "type": "text", + "content": "[14] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12479–12488, 2023. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 250, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 250, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 250, + 553, + 293 + ], + "type": "text", + "content": "[15] Hang Gao, Ruilong Li, Shubham Tulsiani, Bryan Russell, and Angjoo Kanazawa. Monocular dynamic view synthesis: A reality check. Advances in Neural Information Processing Systems, 35:33768-33780, 2022. 3, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 294, + 553, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 553, + 336 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 553, + 336 + ], + "type": "text", + "content": "[16] Huiyu Gao, Jiahao Ma, David Ahmedt-Aristizabal, Chuong Nguyen, and Miaomiao Liu. Soaf: Scene occlusion-aware neural acoustic field. arXiv preprint arXiv:2407.02264, 2024.3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 338, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 553, + 381 + ], + "type": "text", + "content": "[17] Zhiyang Guo, Wengang Zhou, Li Li, Min Wang, and Houqiang Li. Motion-aware 3d gaussian splatting for efficient dynamic scene reconstruction. arXiv preprint arXiv:2403.11447, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 383, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 383, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 317, + 383, + 553, + 437 + ], + "type": "text", + "content": "[18] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE transactions on pattern analysis and machine intelligence, 36(7):1325-1339, 2013. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 438, + 553, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 438, + 553, + 503 + ], + "spans": [ + { + "bbox": [ + 317, + 438, + 553, + 503 + ], + "type": "text", + "content": "[19] Hanbyul Joo, Tomas Simon, Xulong Li, Hao Liu, Lei Tan, Lin Gui, Sean Banerjee, Timothy Scott Godisart, Bart Nabbe, Iain Matthews, Takeo Kanade, Shohei Nobuhara, and Yaser Sheikh. Panoptic studio: A massively multiview system for social interaction capture. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017. 2, 3, 1, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 504, + 553, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 504, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 317, + 504, + 553, + 557 + ], + "type": "text", + "content": "[20] HyunJun Jung, Nikolas Brasch, Jifei Song, Eduardo Perez-Pellitero, Yiren Zhou, Zhihao Li, Nassir Navab, and Benjamin Busam. Deformable 3d gaussian splatting for animatable human avatars. arXiv preprint arXiv:2312.15059, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 559, + 553, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 553, + 592 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 553, + 592 + ], + "type": "text", + "content": "[21] Takeo Kanade, Peter Rander, and PJ Narayanan. Virtualized reality: Constructing virtual worlds from real scenes. IEEE multimedia, 4(1):34-47, 1997. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 593, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 593, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 593, + 553, + 624 + ], + "type": "text", + "content": "[22] Kai Katsumata, Duc Minh Vo, and Hideki Nakayama. A compact dynamic 3d gaussian representation for real-time dynamic view synthesis, 2024. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 625, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 669 + ], + "type": "text", + "content": "[23] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuehler, and George Drettakis. 3D Gaussian Splatting for Real-Time Radiance Field Rendering. ACM Transactions on Graphics, 42(4):1-14, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[24] Agelos Kratimenos, Jiahui Lei, and Kostas Daniilidis. Dynmf: Neural motion factorization for real-time dynamic view synthesis with 3d gaussian splatting. arXiv preprint arXiv:2312.00112, 2023. 2, 3" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16562" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[25] Joo Chan Lee, Daniel Rho, Xiangyu Sun, Jong Hwan Ko, and Eunbyung Park. Compact 3d gaussian splatting for static and dynamic radiance fields. arXiv preprint arXiv:2408.03822, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 162 + ], + "type": "text", + "content": "[26] Deqi Li, Shi-Sheng Huang, Zhiyuan Lu, Xinran Duan, and Hua Huang. St-4dgs: Spatial-temporally consistent 4d gaussian splatting for efficient dynamic scene rendering. In ACM SIGGRAPH 2024 Conference Papers, pages 1–11, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "type": "text", + "content": "[27] Lingzhi Li, Zhen Shen, Zhongshu Wang, Li Shen, and Ping Tan. Streaming radiance fields for 3d video synthesis. Advances in Neural Information Processing Systems, 35: 13485-13498, 2022. 2, 3, 1, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 294, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 283 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 283 + ], + "type": "text", + "content": "[28] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5521-5531, 2022. 2, 3, 1, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 286, + 294, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 294, + 329 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 294, + 329 + ], + "type": "text", + "content": "[29] Yanyan Li, Chenyu Lyu, Yan Di, Guangyao Zhai, Gim Hee Lee, and Federico Tombari. Geogaussian: Geometry-aware gaussian splatting for scene rendering. arXiv preprint arXiv:2403.11324, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 331, + 294, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 294, + 385 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 294, + 385 + ], + "type": "text", + "content": "[30] Zhan Li, Zhang Chen, Zhong Li, and Yi Xu. Spacetime gaussian feature splatting for real-time dynamic view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8508-8520, 2024. 2, 3, 4, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 388, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 430 + ], + "type": "text", + "content": "[31] Susan Liang, Chao Huang, Yapeng Tian, Anurag Kumar, and Chenliang Xu. Av-nerf: Learning neural fields for real-world audio-visual scene synthesis. Advances in Neural Information Processing Systems, 36:37472-37490, 2023. 3, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 433, + 294, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 433, + 294, + 485 + ], + "spans": [ + { + "bbox": [ + 56, + 433, + 294, + 485 + ], + "type": "text", + "content": "[32] Yiqing Liang, Numair Khan, Zhengqin Li, Thu Nguyen-Phuoc, Douglas Lanman, James Tompkin, and Lei Xiao. Gaufre: Gaussian deformation fields for real-time dynamic novel view synthesis. arXiv preprint arXiv:2312.11458, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 488, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 488, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 488, + 294, + 533 + ], + "type": "text", + "content": "[33] Haotong Lin, Sida Peng, Zhen Xu, Yunzhi Yan, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Efficient neural radiance fields for interactive free-viewpoint video. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 2, 3, 1, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 534, + 294, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 534, + 294, + 586 + ], + "spans": [ + { + "bbox": [ + 56, + 534, + 294, + 586 + ], + "type": "text", + "content": "[34] Kai-En Lin, Lei Xiao, Feng Liu, Guowei Yang, and Ravi Ramamoorthi. Deep 3d mask volume for view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1749–1758, 2021. 2, 3, 1, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 590, + 294, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 643 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 643 + ], + "type": "text", + "content": "[35] Youtian Lin, Zuozhuo Dai, Siyu Zhu, and Yao Yao. Gaussian-flow: 4d reconstruction with dynamic 3d gaussian particle. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21136-21145, 2024. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 646, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 646, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 646, + 294, + 689 + ], + "type": "text", + "content": "[36] Qingming Liu, Yuan Liu, Jiepeng Wang, Xianqiang Lv, Peng Wang, Wenping Wang, and Junhui Hou. Modgs: Dynamic gaussian splatting from causally-captured monocular videos. arXiv preprint arXiv:2406.00434, 2024. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[37] Cheng-You Lu, Peisen Zhou, Angela Xing, Chandradeep Pokhariya, Arnab Dey, Ishaan Nikhil Shah, Rugved Mavidi" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 72, + 553, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 72, + 553, + 126 + ], + "spans": [ + { + "bbox": [ + 333, + 72, + 553, + 126 + ], + "type": "text", + "content": "palli, Dylan Hu, Andrew I Comport, Kefan Chen, et al. Diva360: The dynamic visual dataset for immersive neural fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22466-22476, 2024. 2, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "type": "text", + "content": "[38] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. arXiv preprint arXiv:2312.00109, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 175, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 217 + ], + "type": "text", + "content": "[39] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 220, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 252 + ], + "type": "text", + "content": "[40] Dawid Malarz, Weronika Smolak, Jacek Tabor, Sławomir Tadeja, and Przemysław Spurek. Gaussian splatting with nef-based color and opacity. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 254, + 553, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 553, + 307 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 553, + 307 + ], + "type": "text", + "content": "[41] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 310, + 553, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 310, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 316, + 310, + 553, + 365 + ], + "type": "text", + "content": "[42] David Narciso, Maximino Bessa, Miguel Melo, Antonio Coelho, and José Vasconcelos-Raposo. Immersive " + }, + { + "bbox": [ + 316, + 310, + 553, + 365 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 316, + 310, + 553, + 365 + ], + "type": "text", + "content": " video user experience: impact of different variables in the sense of presence and cybersickness. Universal Access in the Information Society, 18:77-87, 2019. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 366, + 553, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 366, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 316, + 366, + 553, + 431 + ], + "type": "text", + "content": "[43] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228, 2021. 2, 1, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 434, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 553, + 487 + ], + "type": "text", + "content": "[44] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In CVPR, 2021. 2, 3, 1, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 490, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 490, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 490, + 553, + 544 + ], + "type": "text", + "content": "[45] Weining Ren, Zihan Zhu, Boyang Sun, Jiaqi Chen, Marc Pollefeys, and Songyou Peng. Nerf on-the-go: Exploiting uncertainty for distractor-free nerfs in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8931-8940, 2024. 2, 1, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 546, + 553, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 546, + 553, + 611 + ], + "spans": [ + { + "bbox": [ + 316, + 546, + 553, + 611 + ], + "type": "text", + "content": "[46] Neus Sabater, Guillaume Boisson, Benoit Vandame, Paul Kerbiriou, Frederic Babon, Matthieu Hog, Remy Gendrot, Tristan Langlois, Olivier Bureller, Arno Schubert, et al. Dataset and pipeline for multi-view light-field video. In Proceedings of the IEEE conference on computer vision and pattern recognition Workshops, pages 30-40, 2017. 2, 3, 1, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 613, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 645 + ], + "type": "text", + "content": "[47] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "type": "text", + "content": "[48] Ruizhi Shao, Zerong Zheng, Hanzhang Tu, Boring Liu, Hongwen Zhang, and Yebin Liu. Tensor4d: Efficient neural 4d decomposition for high-fidelity dynamic reconstruction and rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16632-16642, 2023. 2, 3, 1" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16563" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 139 + ], + "type": "text", + "content": "[49] Roman Shapovalov, Yanir Kleiman, Ignacio Rocco, David Novotny, Andrea Vedaldi, Changan Chen, Filippos Kokkinos, Ben Graham, and Natalia Neverova. Replay: Multi-modal multi-view acted videos for casual holography. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20338-20348, 2023. 2, 3, 1, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 140, + 295, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 140, + 295, + 185 + ], + "spans": [ + { + "bbox": [ + 55, + 140, + 295, + 185 + ], + "type": "text", + "content": "[50] Richard Shaw, Michal Nazarczuk, Jifei Song, Arthur Moreau, Sibi Catley-Chandar, Helisa Dhamo, and Eduardo Perez-Pellitero. Swings: Sliding windows for dynamic 3d gaussian splatting, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 186, + 295, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 186, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 57, + 186, + 295, + 251 + ], + "type": "text", + "content": "[51] Liangchen Song, Anpei Chen, Zhong Li, Zhang Chen, Lele Chen, Junsong Yuan, Yi Xu, and Andreas Geiger. Nerfplayer: A streamable dynamic scene representation with decomposed neural radiance fields. IEEE Transactions on Visualization and Computer Graphics, 29(5):2732-2742, 2023. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 252, + 295, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 252, + 295, + 317 + ], + "spans": [ + { + "bbox": [ + 57, + 252, + 295, + 317 + ], + "type": "text", + "content": "[52] Jiakai Sun, Han Jiao, Guangyuan Li, Zhanjie Zhang, Lei Zhao, and Wei Xing. 3dgstream: On-the-fly training of 3d gaussians for efficient streaming of photo-realistic free-viewpoint videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20675-20685, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 319, + 295, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 319, + 295, + 374 + ], + "spans": [ + { + "bbox": [ + 57, + 319, + 295, + 374 + ], + "type": "text", + "content": "[53] Feng Wang, Sinan Tan, Xinghang Li, Zeyue Tian, Yafei Song, and Huaping Liu. Mixed neural voxels for fast multiview video synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19706-19716, 2023. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 376, + 295, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 376, + 295, + 420 + ], + "spans": [ + { + "bbox": [ + 57, + 376, + 295, + 420 + ], + "type": "text", + "content": "[54] Feng Wang, Zilong Chen, Guokang Wang, Yafei Song, and Huaping Liu. Masked space-time hash encoding for efficient dynamic scene reconstruction. Advances in Neural Information Processing Systems, 36, 2024. 2, 1, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 422, + 295, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 422, + 295, + 487 + ], + "spans": [ + { + "bbox": [ + 57, + 422, + 295, + 487 + ], + "type": "text", + "content": "[55] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20310-20320, 2024. 2, 3, 4, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 488, + 295, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 488, + 295, + 542 + ], + "spans": [ + { + "bbox": [ + 57, + 488, + 295, + 542 + ], + "type": "text", + "content": "[56] Tianhao Wu, Fangcheng Zhong, Andrea Tagliasacchi, Forrester Cole, and Cengiz Oztireli. D^2nerf: Self-supervised decoupling of dynamic and static objects from a monocular video. Advances in neural information processing systems, 35:32653-32666, 2022. 2, 1, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 544, + 295, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 544, + 295, + 589 + ], + "spans": [ + { + "bbox": [ + 57, + 544, + 295, + 589 + ], + "type": "text", + "content": "[57] Yunzhi Yan, Haotong Lin, Chenxu Zhou, Weijie Wang, Haiyang Sun, Kun Zhan, Xianpeng Lang, Xiaowei Zhou, and Sida Peng. Street gaussians for modeling dynamic urban scenes. arXiv preprint arXiv:2401.01339, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 590, + 295, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 590, + 295, + 633 + ], + "spans": [ + { + "bbox": [ + 57, + 590, + 295, + 633 + ], + "type": "text", + "content": "[58] Zeyu Yang, Hongye Yang, Zijie Pan, Xiatian Zhu, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. arXiv preprint arXiv:2310.10642, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "type": "text", + "content": "[59] Ziyi Yang, Xinyu Gao, Yangtian Sun, Yihua Huang, Xiaoyang Lyu, Wen Zhou, Shaohui Jiao, Xiaojuan Qi, and Xiaogang Jin. Spec-gaussian: Anisotropic view-dependent appearance for 3d gaussian splatting. arXiv preprint arXiv:2402.15870, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "text", + "content": "[60] Jae Shin Yoon, Kihwan Kim, Orazio Gallo, Hyun Soo Park, and Jan Kautz. Novel view synthesis of dynamic scenes with" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 152 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "text", + "content": "globally coherent depths from a monocular camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5336-5345, 2020. 2, 3, 1, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 107, + 553, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 152 + ], + "type": "text", + "content": "[61] Dongbin Zhang, Chuming Wang, Weitao Wang, Peihao Li, Minghan Qin, and Haoqian Wang. Gaussian in the wild: 3d gaussian splatting for unconstrained image collections. arXiv preprint arXiv:2403.15704, 2024. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16564" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Image Generation Diversity Issues and How to Tame Them/7493d326-112b-4a6d-8c2c-7e4449ad1384_content_list.json b/2025/Image Generation Diversity Issues and How to Tame Them/7493d326-112b-4a6d-8c2c-7e4449ad1384_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9cf20f7a8de84e167a821ac376b2ddbb0de43b8f --- /dev/null +++ b/2025/Image Generation Diversity Issues and How to Tame Them/7493d326-112b-4a6d-8c2c-7e4449ad1384_content_list.json @@ -0,0 +1,1379 @@ +[ + { + "type": "text", + "text": "Image Generation Diversity Issues and How to Tame Them", + "text_level": 1, + "bbox": [ + 200, + 130, + 795, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mischa Dombrowski1 Weitong Zhang2 Sarah Cechnicka2", + "bbox": [ + 236, + 179, + 759, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hadrien Reynaud² Bernhard Kainz1,2", + "bbox": [ + 331, + 199, + 666, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Friedrich-Alexander-Universität Erlangen-Nürnberg $^{2}$ Imperial College London", + "bbox": [ + 160, + 215, + 836, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "mischa.dombrowski@fau.de", + "bbox": [ + 390, + 237, + 604, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 286, + 326, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generative methods have reached a level of quality that is almost indistinguishable from real data. However, while individual samples may appear unique, generative models often exhibit limitations in covering the full data distribution. Unlike quality issues, diversity problems within generative models are not easily detected by simply observing single images or generated datasets, which means we need a specific measure to assess the diversity of these models. In this paper, we draw attention to the current lack of diversity in generative models and the inability of common metrics to measure this. We achieve this by framing diversity as an image retrieval problem, where we measure how many real images can be retrieved using synthetic data as queries. This yields the Image Retrieval Score (IRS), an interpretable, hyperparameter-free metric that quantifies the diversity of a generative model's output. IRS requires only a subset of synthetic samples and provides a statistical measure of confidence. Our experiments indicate that current feature extractors commonly used in generative model assessment are inadequate for evaluating diversity effectively. Consequently, we perform an extensive search for the best feature extractors to assess diversity. Evaluation reveals that current diffusion models converge to limited subsets of the real distribution, with no current state-of-the-art models superpassing $77\\%$ of the diversity of the training data. To address this limitation, we introduce Diversity-Aware Diffusion Models (DiADM), a novel approach that improves diversity of unconditional diffusion models without loss of image quality. We do this by disentangling diversity from image quality by using a diversity aware module that uses pseudo-unconditional features as input. We provide a Python package offering unified feature extraction and metric computation to further facilitate the evaluation of generative models https://github.com/MischaD/beyondfid.", + "bbox": [ + 86, + 369, + 483, + 883 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4167b51f1a8905e82a2b7e30a60ce5a15048409adb695873d68a5fa0959e4183.jpg", + "image_caption": [ + "Low Diversity" + ], + "image_footnote": [], + "bbox": [ + 516, + 285, + 604, + 434 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/29ef5003bf151ccf25c9045c6134b6c99418f4e1c76d718d1ed9f44ec54fac28.jpg", + "image_caption": [ + "High Diversity" + ], + "image_footnote": [], + "bbox": [ + 607, + 285, + 691, + 434 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c8e1ad8d8111c37e0a5d63bfeaea045fec25ae6b303cf5504ad267bf538b045e.jpg", + "image_caption": [ + "Predicted Diversity $\\mathrm{IRS}_{\\infty}$ $(\\%)$", + "Figure 1. Predicted gender diversity - as one possible example for diversity - after sampling pre-trained text-to-image diffusion models infinitely often. Visually it is apparent that most terms are inherently biased toward one gender. Our proposed IRS score predicts that off-the-shelf models will only ever reach 50 percent diversity, which is equivalent to one gender being perfectly represented while the other gender is missing entirely. Using diversification strategies increases diversity. Details about this can be found in Sec. 4.4." + ], + "image_footnote": [], + "bbox": [ + 696, + 286, + 903, + 436 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 614, + 643, + 630 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With advancements in large-scale models and datasets, diffusion models have demonstrated substantial capacity to capture high-dimensional, complex distributions, enabling photo-realistic conditional image generation across diverse guidance types, including image-based [56], classifier-based [11], and text-based guidance [42]. However, diffusion models frequently inherit and amplify data biases due to their reliance on large-scale, naturally biased datasets [4, 53], limiting comprehensive distributional coverage in generated samples. For instance, text-to-image diffusion models often exhibit gender bias for certain occupations. When prompted with \"a teacher\", models predominantly generate females, with male representations notably underrepresented, as shown in Fig. 1. The \"conditional mode collapse\" [1, 31] also affects image- [56] and class-conditional [11] diffusion models, where the guidance often produces overly uniform outputs, leading to a loss in diver", + "bbox": [ + 509, + 643, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3029", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sity in the sampled distribution. Dataset diversity remains an unresolved issue, even for real datasets [58], due to ambiguous definitions and concepts. In contrast, synthetic data can leverage its training data as a reference. To avoid problems like mode drift [2], synthetic data should be at least as diverse as its training set. Key motivations for ensuring diversity include content creation [44], addressing privacy and memorization concerns [6], mitigating the lack of diversity in synthetic datasets [14], data augmentation [15], promoting algorithmic fairness [17], and enhancing downstream performance [8]. To improve diversity in conditional image generation, [28] introduces noise perturbations to enhance diversity, though at the expense of image quality and certainty. [30] balances a trade-off between faithfulness and realism in diverse image synthesis, where diversity comes at the cost of fidelity. These diversity-oriented approaches highlight an inherent compromise: diversity is achieved by sacrificing generation quality, with randomness that lacks interpretability. In parallel, classifier-based diversity guidance [47] and text-based methods [13, 55] have been developed to reduce bias in text-to-image generation. However, these techniques are task-specific and lack generalizability across different types of diffusion models for diversity enhancement. Thus, a fundamental challenge emerges: a universal, interpretable metric to evaluate diversity without bias or conditional dependence, and without compromising generation quality.", + "bbox": [ + 89, + 90, + 483, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this, we propose the Image Retrieval Score (IRS), an interpretable metric for quantifying diversity in generative models. IRS assesses how effectively synthetic data retrieves real images as queries, providing a hyperparameter-free, statistically grounded evaluation based on minimal samples. Our approach is motivated by a thought experiment that illustrates how individual samples reveal model diversity. Using IRS as a core, generic metric for diversity, we quantify and reveal two key findings: (1) existing feature extractors used to evaluate generative models lack the capacity to measure diversity effectively, and (2) current diffusion models converge on limited subsets of the real distribution, with no state-of-the-art models surpassing $77\\%$ of the diversity in training data.", + "bbox": [ + 89, + 500, + 483, + 712 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This further motivates the use of IRS as an unbiased guidance to drive a comprehensive search for optimal feature extractors, enhancing diversity evaluation without compromising generation quality. Our selection is based on [49] who experimented on the impact of feature extractors on existing generative metrics. Building on these insights, we introduce Diversity-Aware Diffusion Models (DiADM), an approach that improves diversity across the full data distri", + "bbox": [ + 89, + 713, + 483, + 835 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7dd0ecfe12433cd49daa81d1380b264c8580ecc2a72f7ca7c82077827aa0842d.jpg", + "image_caption": [ + "Figure 2. We model finding image pairs (image retrieval) as randomly drawing from urns with replacement (Ideal). We observe that all used feature extraction models exhibit performance issues by collapsing in the feature space for real images resulting in a measurement gap affecting currently used metrics such as $\\mathrm{FID}_{\\mathrm{Inception}}$ , $\\mathrm{FID}_{\\mathrm{DINOv2}}$ , Precision and Recall. Comparing the results of synthetic images to real data shows that datasets generated by generative models show even stronger diversity issues which result in a synthetic data distribution gap. Our proposed metric leverages a real reference dataset to remove this measurement gap." + ], + "image_footnote": [], + "bbox": [ + 542, + 95, + 880, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "bution in unconditional diffusion models without compromising image quality. DiADM accomplishes this by disentangling diversity from image quality through a diversity-aware module utilizing pseudo-unconditional features. Our contributions can be summarized as follows:", + "bbox": [ + 511, + 468, + 906, + 542 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We demonstrate the limitations of widely-used feature extractors for diversity metrics, highlighting their poor performance on real datasets.", + "- We propose an intuitive framework for assessing dataset diversity in unconditional diffusion models.", + "- We identify and quantify the reduction in diversity resulting from current training methodologies.", + "- We provide a comprehensive analysis showing limited image diversity across a wide range of state-of-the-art diffusion models.", + "- We introduce a diversity-aware plug-in module, which utilizes pseudo-labels to separate diversity from fidelity, thereby enhancing generative diversity in unconditional diffusion models." + ], + "bbox": [ + 514, + 544, + 903, + 753 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 768, + 656, + 784 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image Generation: Diffusion models [48] progressively map complex distributions to a standard Gaussian. Conditional diffusion models have demonstrated a strong ability to generate highly realistic images that align closely with specified conditions, excelling in tasks such as image synthesis [54], image restoration [57], and image reconstruction [56]. Additionally, [11] enhances diffusion", + "bbox": [ + 511, + 794, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1Consider a die purported to have 1000 sides. If rolled 1000 times but yielding only 10 distinct outcomes, one would naturally question the claim. Similarly, in generative models, each sample serves as a \"roll\", reflecting the model's diversity in an unconditional manner.", + "bbox": [ + 89, + 849, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3030", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "models with an auxiliary classifier, enabling high-fidelity, class-conditional generation, while text-conditional models generate images based on text prompts [13, 55]. However, diffusion models often struggle to achieve sufficient diversity in task-agnostic generation, especially when trained on large-scale datasets that carry inherent biases [4, 53]. For example, Stable Diffusion, based on latent diffusion models (LDM) [39, 42], produces high-quality images aligned with text prompts, yet frequently reflects bias (e.g., age, gender) for certain terms. These limitations lead to restricted distributional coverage and biased outputs.", + "bbox": [ + 89, + 90, + 480, + 257 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diversity in Generation: Research on diversity in diffusion models is limited, with most existing approaches focused on GAN-based models [9, 25]. Current methods for diversity in diffusion models fall into the primary categories: bias mitigation and diversity enhancement. In bias mitigation, [37] uses a re-weighted loss to address long-tailed distributions, promoting balanced class representation. For text-conditional models, [13] attempts to mitigate bias by adding attribute words to prompts, though the added prompts may be ambiguous. [31] extends this to both class- and text-conditional models, using reinforcement learning to match the diversity ratio in reference images but relying on task-specific diversity reward functions. For diversity enhancement, [28] introduces noise perturbations, which increase diversity at the expense of image quality and certainty. Similarly, [30] balances diversity with realism, where gains in diversity tend to compromise fidelity.", + "bbox": [ + 89, + 284, + 482, + 541 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Metrics of Diversity: Existing metrics in diffusion models remain limited and are unsuitable for accurate and interpretable diversity assessment. Traditional image quality metrics like FID [20] require extensive sampling (e.g., 50k images) to be reliable, making them computationally intensive and challenging to interpret in practical applications. Precision and recall [26] are most established in literature [12, 36, 41, 46]. They add supplementary insights but are hindered by dependency on dataset size, hyperparameters, and interpretability issues. Credibility and interpretability of metrics are highly underestimated properties. For example, a comparison of three years of advancements in diffusion models, as seen in [12] and [27], shows an order-of-magnitude improvement in FID but a one percentage point reduction in precision. Coverage and density [32] also measure diversity but are dependant on the ratio between train and test set, highly depend on hyperparameters, lack evaluation on real datasets, and density was designed to ignore real outliers which are important in the context of diversity assessment. Vendi Score [16], which operates entirely without a real reference set, sacrifices fidelity and is computationally prohibitive for large datasets.", + "bbox": [ + 89, + 568, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 89, + 604, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $\\mathbf{x_t}$ be image $t$ from a dataset consisting of $\\mathrm{N}_{train}$ real images residing in image space $\\mathcal{X} \\in \\mathbb{R}^{x\\times h\\times w}$ . Unconditional generative models aim to learn the distribution $p_{data}(\\mathbf{X})$ and sample $\\mathrm{N}_{\\text{sample}}$ synthetic images from it.", + "bbox": [ + 511, + 114, + 903, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Coupon Collector Problem", + "text_level": 1, + "bbox": [ + 511, + 183, + 756, + 199 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The performance of generative models is generally measured by assessing the similarity of the real and the synthetic datasets. It is assumed that the training samples represent discrete observations from a continuous unknown distribution. We suggest viewing all synthetic images as a composition of real images, which holds true as long as there is no training objective that enforces out-of-distribution generation. To assess diversity, we argue that every training sample should be the main component of at least one synthetic sample. To measure this we use the fact that the observed training distribution is discrete. A model that has perfectly memorized every training sample would generate all images with equal probability $\\frac{1}{\\mathrm{N}_{train}}$ . Consequently, we can model sampling synthetic images with an urn experiment. Each synthetic sample consists mainly of one component which is then compounded with other training images. A real sample is considered learned if it serves as the main component for at least one synthetic image. Evaluating diversity in this context can be framed as the coupon collector problem [5]. Let $P_{n}(N_{t} > 0)$ denote the probability of generating $n$ samples where at least one corresponds to a training image $t$ , with $N_{t}$ denoting the number of generated images a training image is associated with. This probability can be computed by considering the compound probability of not sampling this training image which is $\\frac{\\mathrm{N}_{train} - 1}{\\mathrm{N}_{train}}$ . Since the same probability applies to each image the expected value of the diversity can be calculated as", + "bbox": [ + 511, + 205, + 906, + 613 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} [ \\text {D i v e r s i t y} ] = \\left(1 - \\left(\\frac {\\mathrm {N} _ {\\text {t r a i n}} - 1}{\\mathrm {N} _ {\\text {t r a i n}}}\\right) ^ {n}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 632, + 906, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this definition, diversity is equal to the percentage of learned real images. This idealized scenario is shown as ideal in Fig. 2. Therefore, assessing diversity comes down to predicting which training image corresponds to which synthetic image.", + "bbox": [ + 511, + 674, + 908, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Image Retrieval", + "text_level": 1, + "bbox": [ + 511, + 757, + 671, + 773 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To assess diversity we first need to define how to decide on what constitutes a main component for each synthetic image. For each synthetic image instance $\\mathbf{x}_{\\mathrm{t}}^{\\prime}$ we aim to identify the real image $\\mathbf{x}_{\\mathrm{t}}$ that is closest to it according to a measurement $\\mathcal{P}$ (e.g., Euclidean distance) on the feature of the images extracted by a pretrained feature extractor $\\mathcal{F}$ . Then, to assess the diversity we have to compute the size of the dataset of learned images:", + "bbox": [ + 511, + 779, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3031", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {X} _ {\\text {l e a r n e d}} = \\left\\{\\mathbf {x} _ {t} \\in \\mathcal {X} \\mid \\exists \\mathbf {x} _ {t} ^ {\\prime} \\in \\mathcal {X} ^ {\\prime}: \\mathbf {x} _ {t} = \\arg \\min _ {\\mathbf {x} _ {t}} \\mathcal {P} \\left(\\mathbf {x} _ {t}, \\mathbf {x} _ {t} ^ {\\prime}\\right) \\right\\}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 114, + 495, + 159 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We define the cardinality of this set as $\\mathbf{N}_{\\text{learned}}$ .", + "bbox": [ + 89, + 161, + 410, + 175 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We then call the trained image instance $\\mathbf{x}_{\\mathrm{t}}$ \"learned\" and compute the diversity of the synthetic dataset as:", + "bbox": [ + 89, + 176, + 482, + 205 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {I R S} _ {\\alpha} = \\frac {\\mathrm {N} _ {\\text {l e a r n e d}}}{\\mathrm {N} _ {\\text {t r a i n}}} \\in [ 0, 1 ] \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 217, + 482, + 250 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We term this quantity as the image retrieval score (IRS). To emphasize that $\\mathrm{IRS}_{\\alpha}$ depends on the number of images sampled we add $\\alpha := \\frac{\\mathrm{N}_{\\text{sample}}}{\\mathrm{N}_{\\text{train}}}$ as a subscript. Note that all other generative metrics also depend on $\\alpha$ but often ignore this dependency. The name image retrieval score derives from the fact that computing IRS also works when comparing two disjoint real datasets. In this context the problem can be thought of as an image retrieval task, where a real image serves as the query image and the goal is to find the most similar image in a comparison dataset. Intuitively, IRS measures the ratio of images in a training dataset that are retrievable by another dataset. If the comparison dataset consists of real images we say that an image is \"retrievable\". If the comparison dataset consists of synthetic images we say that an image is \"learned\".", + "bbox": [ + 89, + 258, + 483, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The key advantage of IRS is that we can give a reasonably good estimate of dataset diversity with only a minimal number of samples. For example, if the same image appears three times within just ten samples, it is a strong indicator of overfitting. Consequently, we aim to estimate the diversity after $n << \\mathrm{N}_{train}$ samples. To compute the probability of retrieving exactly $k = \\mathcal{X}_{\\text{learned}}$ images we use standard combinatorics. This problem can be split into two parts. Firstly, we compute the number of possibilities of splitting $n$ into $k$ different subsets. Instinctively, we are computing the number of combinations in which different synthetic samples derive from the same retrieved (still unassigned) image $\\mathbf{x}_t$ according to equation Eq. (2). The computation involves using Stirling's number of the second kind, $\\operatorname{Stir}(n, k)$ , which should be multiplied by the total number of possible assignments for each subset, given by $\\frac{\\mathrm{N}_{train}!}{(\\mathrm{N}_{train} - \\mathrm{N}_{\\text{learned}})!}$ . This will become more clear with an example. Say, the first subset consists of three different images $x_3', x_{10}', x_{11}'$ , all of which are learned from the same real image. Then for the first of $k$ subsets, there are $\\mathrm{N}_{train}$ different choices. For the second subset, we have $\\mathrm{N}_{train} - 1$ choices and so on. Dividing by all choices leaves us with the probability for generating $k = \\mathrm{N}_{\\text{learned}}$ different images:", + "bbox": [ + 89, + 486, + 483, + 835 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {P} (k, n, s) = \\frac {\\operatorname {S t i r} (n , k) * s !}{(s - k) ! * (s) ^ {n}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 845, + 482, + 878 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with $k = \\mathrm{N}_{\\text{learned}}$ and $s = \\mathrm{N}_{\\text{train}}$ and $n = \\mathrm{N}_{\\text{sample}}$ .", + "bbox": [ + 89, + 885, + 482, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For large values, we estimate this by computing the log of Eq. (4) instead. [51] showed that", + "bbox": [ + 511, + 90, + 906, + 121 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S t i r} (n, k) = \\sqrt {\\frac {v - 1}{v (1 - G)}} \\left(\\frac {v - 1}{v - G}\\right) ^ {n - k} \\frac {k ^ {n}}{n ^ {k}} e ^ {k (1 - G)} \\binom {n} {k}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 146, + 903, + 198 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $v = \\frac{n}{k}$ and $G \\in (0,1)$ is the unique solution to $G = v e^{G - v}$ . Asymptotically, this estimate is close, reaching a maximum relative error of roughly $\\frac{0.066}{n}$ if $n$ and $k$ are small, e.g., (10, 3) [51]. Computing this in log-space for ImageNet roughly takes 3 seconds and using memorization techniques reduces consecutive computations to 1.3 seconds.", + "bbox": [ + 511, + 199, + 906, + 304 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Image Retrieval Score", + "text_level": 1, + "bbox": [ + 513, + 315, + 720, + 330 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given Eq. (5) we can now estimate IRS as well as a confidence interval for it. We assume that $\\mathrm{N}_{\\text {learned }}$ and $\\mathrm{N}_{\\text {sample }}$ are fixed and we aim to estimate the model diversity. To compute the maximum likelihood estimate for IRS, we optimize over values of $s \\in \\mathbb{N} \\cap [\\mathrm{N}_{\\text {learned }}, \\mathrm{N}_{\\text {train }}}$ :", + "bbox": [ + 511, + 335, + 905, + 414 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {I R S} _ {\\infty} = \\frac {1}{\\mathrm {N} _ {t r a i n}} \\arg \\max _ {s} \\mathrm {P} (\\mathrm {N} _ {l e a r n e d}, \\mathrm {N} _ {s a m p l e}, s), (6)\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 424, + 905, + 457 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which means we expect $\\mathrm{N}_{\\mathrm{learned},\\infty} = \\mathrm{N}_{\\mathrm{train}}*\\mathrm{IRS}_{\\infty}$ different images to be sampled eventually. To compute the lower bound we want to minimize the probability of the real diversity being lower but sampling abnormally many different images early. Mathematically, we estimate this as:", + "bbox": [ + 511, + 465, + 906, + 541 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathrm {I R S} _ {\\infty , L} = \\frac {1}{N _ {\\text {t r a i n}}} \\arg \\max _ {s \\in \\mathbb {N} \\cap [ N _ {\\text {l e a n e d}}, \\infty , N _ {\\text {t r a i n}} ]} \\\\ \\min \\left(N _ {\\text {t r a i n}}, N _ {\\text {s a m p l e}}\\right) \\tag {7} \\\\ \\sum_ {k = N _ {\\text {l e a n e d}}} \\quad \\mathrm {P} (k, N _ {\\text {s a m p l e}}, s) > \\alpha_ {e} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 564, + 903, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Accordingly, we can compute the estimate for the upper boundary:", + "bbox": [ + 511, + 656, + 905, + 686 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathrm {I R S} _ {\\infty , U} = \\frac {1}{N _ {\\text {t r a i n}}} \\arg \\min _ {s \\in \\mathbb {N} \\cap \\left[ N _ {\\text {l e a n e d}}, N _ {\\text {l e a n e d}, \\infty} \\right]} \\\\ \\sum_ {k = 1} ^ {s} \\mathrm {P} (k, N _ {\\text {s a m p l e}}, s) > \\alpha_ {e} \\tag {8} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 709, + 903, + 782 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "for a desired probability of error $\\alpha_{e}$ . Intuitively, this value represents the most likely estimate of $N_{\\mathrm{learned},\\infty}$ that would produce $N_{\\mathrm{learned}}$ unique samples after drawing $N_{\\mathrm{sample}}$ images. We use the infinity symbol to indicate that this estimate represents the potential diversity if we sampled an infinite number of images. This value also reflects the model's diversity, representing the percentage of samples that the", + "bbox": [ + 511, + 794, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3032", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/98cd23f6996e94800cea916e18829910f058f34d2571bf6a62ba02da8062a0af.jpg", + "image_caption": [ + "Figure 3. Overview of our proposed DiADM model. Instead of using placeholder labels for unconditional generation, we propose to leverage precomputed features from $\\mathcal{F}$ instead. That way we can disentangle fidelity (FID) from diversity (IRS)." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 277, + 209 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/56fece750a9c67e2dbcad5057031e9c8129d139664512ce735e0ffd2c56509e5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 285, + 88, + 478, + 209 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "model can generate at its limit. In Appx. 7 we show the dependency of IRS and the confidence intervals on the number of samples showing that at $\\alpha = 0.1$ , the estimates start to get very confident around the ground truth value.", + "bbox": [ + 89, + 301, + 483, + 363 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Adjusting for Measurement Gap", + "text_level": 1, + "bbox": [ + 89, + 373, + 377, + 390 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Sec. 4 we will show the measurement gap stemming from feature extractors collapsing to smaller representation spaces. This results in the underestimation of diversity for real data which in turn also influences the measured diversity for synthetic data. To eliminate the influence of the measurement gap we adjust for the reduced amount of diversity in the feature space of the extractor by normalizing by the measured diversity of real data according to the feature extractor. Effectively, we compute $\\mathrm{IRS}_{\\infty ,real}$ on real data using a reference dataset. Then we compute $\\mathrm{IRS}_{\\infty ,snth}$ on synthetic data and report the adjusted $\\mathrm{IRS}_{\\infty ,a} = \\frac{\\mathrm{IRS}_{\\infty,snth}}{\\mathrm{IRS}_{\\infty,real}}$ instead. Intuitively, $\\mathrm{IRS}_{\\infty ,a}$ computes the ratio of diversity in the feature space of the synthetic images compared to the real images. Note that this implies that the adjusted score can be above one for low $\\alpha$ values (see Fig. 4). We provide additional theoretical analysis in Appx. 15.1 and more quantitative results in Appx. 15.2, to demonstrate effectiveness in enhancing sensitivity and reliability for measuring diversity in generative models compared to other metrics.", + "bbox": [ + 89, + 395, + 483, + 702 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Diversity-Aware Diffusion Models", + "text_level": 1, + "bbox": [ + 89, + 712, + 387, + 728 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address diversity issues, we propose DiADMs. These models leverage the ability to extract meaningful representations using Inception v3 as feature extractor without relying on labels. Rather than employing placeholder labels, as is common in unconditional image generation, we use these pre-computed image features and feed them directly into the model. We term this approach 'pseudo-unconditional' because, while the architecture resembles a conditional setup, no explicit labels are needed. In principle, any feature extractor or custom self-supervised model can be used. However, we use the pre-trained ImageNet model, which is", + "bbox": [ + 89, + 734, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "widely recognized as an effective feature extractor. The core idea is that, with proper training, the model behaves as if each training instance represents its own class, allowing for direct diversification of training instances during sampling time. Crucially, this approach decouples the diversity of the generated images from their fidelity as shown in Fig. 3. The diffusion model is responsible for producing high-fidelity images, while the diversity module ensures comprehensive coverage of the training distribution's diversity. We directly utilize features extracted from $\\mathcal{F}$ of the training dataset to generate synthetic data, resulting in a synthetic dataset that should maintain diversity if trained properly. The backbone model employed is the XS architecture as proposed by [24], with modifications to adapt the label dimensionality to match the feature dimensionality of $\\mathcal{F}$ . Inception is used as the feature extraction model for all datasets to maintain generalizability.", + "bbox": [ + 511, + 90, + 906, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 513, + 361, + 643, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Image Generation Models: Since diversity is mostly underexplored we aim to compare a series of different state-of-the-art methods against one another. [12] introduced classifier guidance and other modifications to Diffusion Models that first lead to competitive performance. Unlike all the current methods, they operate in image space. [41] were the first to introduce a model that performs diffusion in latent space. [24] use several improved training and design strategies for the U-Net backbone. [36] propose to use a Transformer architecture. [27] use autoregressive prediction and leverage masking for efficient training. Our experiments focus on both label-conditional and unconditional generative models, with an emphasis on enhancing diversity in unconditional image generation. This focus is motivated by the observation that any form of conditioning ultimately reduces to unconditional generation over the marginal distribution defined by the conditioning. We use pre-trained models directly and if no unconditional implementation was provided we mimic it by setting the class conditioning to the placeholder value used for unconditional training [22]. For our experiment with DiADM we train models ourselves using a fixed compute budget of 574 A40 GPU hours, which is only one-tenth of the training length in [24].", + "bbox": [ + 511, + 387, + 906, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics: We sample each model 50000 times which is a common choice to get a good estimate for FID [23, 49]. To mitigate the influence of distribution shift to the test datasets, which is the case for example in ImageNet, we report all metrics on the entire training set. To get a reference dataset for the adjusted IRS score, we take away 50000 samples for the reference dataset. If the datasets are smaller than 100000 samples we split the dataset in half and average the IRS results for as many runs as we have synthetic images.", + "bbox": [ + 511, + 734, + 908, + 885 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets: For our primary experiments, we use ImageNet", + "bbox": [ + 511, + 885, + 903, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3033", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/10ee7e34c66f64ebe7c2ee4d9ad46af4cc97bffd6329ff49843bceb5f20c1a41.jpg", + "image_caption": [ + "Figure 4. Visualization of the measurement gap across a diverse set of feature extractors by computing the unadjusted diversity of real data. The theoretical limit would be sampling $\\mathrm{N}_{train}$ images where all of them correspond to a different image in the training dataset. The idealized scenario follows Eq. (1)." + ], + "image_footnote": [], + "bbox": [ + 135, + 92, + 439, + 258 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "512 [10]. To explore diversity within a more densely populated image space, we also include FFHQ [23], a benchmark dataset focused on facial images, and Chest-Xray14, which comprises medical chest X-ray images [52]. Additionally, we evaluate CelebV-HQ, a dataset containing face video frames [59], and Dynamic, a medical video dataset [35], treating the data as individual frames as done in [40]. The spatial resolution is fixed to 512 pixels and videos are additionally limited to the first 60 frames.", + "bbox": [ + 89, + 368, + 482, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Encoders: To find the best feature extractor $\\mathcal{F}$ we compute the features using a large set of publicly available feature extractors following the selection of [49]. This includes BYOL [18], CLIP [38], a ConvNeXT based architecture [29], data2vec [3], DINOv2 [34], Inception [21, 45, 50], MAE [19], SwAV [7]. Additionally we add \"Random\" which uses a randomly initialized Inception-v3 [50] which has been shown to already extract meaningful features [33].", + "bbox": [ + 89, + 505, + 482, + 626 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Measurement Gap", + "text_level": 1, + "bbox": [ + 89, + 636, + 272, + 652 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "First we need to decide for a feature extractor $\\mathcal{F}$ that decides whether two samples correspond to each other. We choose to limit ourselves to large, pre-trained and publicly available feature extractors due to their universal applicability and established use. Custom feature extractors are also possible and could be used to increase interpretability even further. In Appx. 13 for example we compute IRS using a re-identification model trained to detect memorized samples unveiling diversity issues. To remain generalizable our goal is to find the most meaningful feature space in terms of image retrieval, that can be generally applied to different datasets. We experiment with a vast set of feature extractors $\\mathcal{F}$ , measurements $\\mathcal{P}$ , and datasets. Importantly we limit the experiments to evaluation on real data. In Sec. 3.1 we made the assumption, that all samples of the training dataset should be the main component of at", + "bbox": [ + 89, + 657, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "least one synthetic image. In the context of only comparing real images, this means that every train image should have at least one test image it is more similar to than all other train images. However, as we show in Fig. 13 and Tab. 6 of the supplementary material, all feature extractors are drastically worse than the expected value. This is a clear indicator for the insufficient representative nature of feature extractors currently used to compute all metrics in the realm of image generation. We refer to this observation as measurement gap. In the supplements we present examples how this measurement gap can be reduced largely if tailord towards certain datasets (e.g., Tab. 5). However, to remain dataset agnostic we propose an adjustment step that uses the measured IRS on real data as reference value to make up for this measurement gap. Visually this gap results in a lower apparent diversity as shown in Fig. 4. There are two key takeaways from this experiment. First, not even real data is diverse according to commonly used feature extractors which severely reduces the interpretability of currently used diversity metrics. Secondly, it could be worthwhile for special tasks, to train feature extractors specifically to assess diversity as done in Tab. 5.", + "bbox": [ + 511, + 90, + 903, + 422 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "However, thanks to the adjustment step described in Sec. 3.4, for our metric the absolute value of diversity of real data does not matter. For synthetic data it just matters how diverse the synthetic data is in comparison to the real data. This adjustment leads to direct interpretability of the metric as diversity of images. In Fig. 5 we visualize this interpretability of IRS by balancing and splitting ImageNet into train, reference and test datasets. We use the labels to manually reduce the diversity in the test dataset by successively removing classes. We use 500 images of each class for train and 100 image of each class for reference and test data. We can see a clear linear correlation between the number of classes and our proposed IRS score. Additionally, it is the only diversity metric that achieves $100\\%$ diversity when all classes are present in the test dataset. The slight overestimation of IRS in the mid-diversity-range is likely to be caused by label ambiguity. Coverage also correlates well but severely overestimates diversity when the number of classes is very low. It saturates at 0.97 and completely fails for different hyperparameter settings as we show in Sec. 15.2. Recall, which is the most commonly used metric to assess synthetic diversity, saturates at about 0.8. From our point of view, this is one of the key reasons why diversity has not yet established itself as the key metric that should be optimized.", + "bbox": [ + 511, + 428, + 903, + 805 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "While IRS works with any kind of feature extractor, we choose to use the one that has the best visual interpretability when it comes to image retrieval to make the results even more meaningful. Hence, we choose $\\mathcal{F}$ based on the meaningfulness of the image retrieval. To avoid the necessity of a ground truth image retrieval dataset from the training", + "bbox": [ + 511, + 809, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3034", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2321420e196cac40dd722752745f1c82ffa348e6538fade110bd31e8da47b3b0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetNConsensusBYOLCLIPConvNeXtdata2vecDINOv2InceptionMAERandomSwAV
ImageNet35142 (3.43%)21.7289.5695.512.6592.7074.2087.1712.3489.39
FFHQ1862 (3.33%)14.1893.0297.852.4292.1675.3586.7921.4398.60
ChestX-ray141389 (1.55%)40.3245.9377.751.9483.1575.9589.7033.7796.11
CelebV-HQ5329 (19.47%)31.8389.4093.716.6891.3379.9494.8651.7996.70
Dynamic23 (0.29%)47.8334.7895.658.7078.2665.2295.6565.22100.00
", + "bbox": [ + 96, + 88, + 901, + 186 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Measures the agreement of each $\\mathcal{P}$ with the ensemble if five or more agree on the same retrieved image $(\\%)$ .", + "bbox": [ + 151, + 198, + 841, + 213 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b9fe34051ebb70ed2da3e7a317a5cb11085a75d88afb99974c5b607ae18a6c84.jpg", + "image_caption": [ + "Figure 5. Measuring diversity of datasets by removing classes and computing IRS. If only 800 out of 1000 classes are part of the test set say that the diversity is at $80\\%$ . By manually removing ImageNet classes we can assess how good commonly used metrics are at measuring diversity compared to ours. To improve visual interpretability we normalize FID and Vendi to be within 0 and 1 with 1 being best. IRS correlates best with the real diversity measured as fraction between number of classes in test and train dataset." + ], + "image_footnote": [], + "bbox": [ + 101, + 243, + 473, + 431 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "dataset, we ensemble all feature extractors to compute a consensus on the image retrieval of real images. Details and qualitative examples can be found in Sec. 9. Consensus is reached if five or more models agree on the same image pair. For every image pair where consensus is reached we compute how often a single model agrees with the ensemble and report the results in Tab. 1. As expected the percentage of cases where consensus is reached is higher for natural images. Overall we see that SwAV generally agrees the most with the ensemble. Unless mentioned otherwise, this is the feature extractor we use to compute IRS scores.", + "bbox": [ + 89, + 604, + 483, + 771 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Diversity Issue in State-of-the-art Diffusion models", + "text_level": 1, + "bbox": [ + 89, + 785, + 483, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Recently, [40] published a synthetic ultrasound video dataset based on the EchoNet dataset [35]. It is accompanied with a feature extraction model used to privatize the dataset using re-identification. Analyzing these features, [14] were the first to notice a diversity issue. Us", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ing our proposed IRS score we are now able to quantify it. First we try retrieve training samples using real frames from the test dataset. The dataset has 7465 training videos and we can see that 1277 test frames are able to retrieve 1159 distinct train videos. Following Eqs. (6) to (8) this means real data has a diversity of $\\mathrm{IRS}_{\\infty,real} = 86\\%$ [75%, 100%]. Contrary to that, 1159 different synthetic frames are only able to retrieve 692 out of 7465 samples. This is equivalent to an $\\mathrm{IRS}_{\\infty,snth} = 12.3\\%$ [11.9%, 12.9%]. Following the adjustment step introduced in Sec. 3.4 we conclude that the model has only learned $\\mathrm{IRS}_{\\infty,a} = 14.3\\%$ of the data. Next, we compare multiple state-of-the-art approaches on conditional and unconditional image generation for ImageNet. We show their diversity according to IRS accompanied with other common metrics in Sec. 4.2. We initially expected that models for lower resolution outperform those with higher resolution. But overall this does not seem to be the case. Our experiments on MAR and EDM nicely confirm the feasibility of IRS as metrics as the score increases with the model size. Other metrics like Recall and Coverage confirm this observation but the difference seems marginal. Using IRS we can quantize and interpret the size of the gap between the smallest and largest models. The smallest model only reached a diversity of $46\\%$ whereas the largest model reached $75\\%$ . Unlike the relatively small gap of 0.9 points in FID, this indicates a large gap in sampling diversity. Crucially, following our experiments from Fig. 5 our proposed IRS score can be interpreted. Even the best model only properly learned $77\\%$ of the data, irrespective of the number of samples generated. Note that this is for conditional image generation, so generation specifically asks the model to generate an equal number of images from all classes. Generally, the conditional models performed better than the unconditional model which leads us to believe that the conditioning needs to be improved. Due to limited resources, we will continue our experiments by fixing the compute and only use the best performing method in terms of FID and IRS which is EDM. To keep the training time reasonable we restrict our experiments on the XS variant.", + "bbox": [ + 511, + 239, + 906, + 830 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Improving Diversity Using Pseudo Labels", + "text_level": 1, + "bbox": [ + 511, + 845, + 870, + 863 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Now that we have established EDM as the best method for unconditional image generation and at the same time have", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3035", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/dcdba2cc70e8f380a0576c0d877f5c02f69a6484383cf1f42f24e568a8a6c22a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelImage ResolutionFID ↓Prec. ↑Rec. ↑Dens. ↑Cov. ↑Vendi ↑IRS∞,a ↑ (Ours)
Pixel diffusion
ADM-256 [12]2566.01 (30.30)0.82 (0.57)0.62 (0.73)1.08 (0.41)0.91 (0.40)70.94 (36.18)0.44 (0.20)
Transformer
DiT-XL/2-256 [36]25622.15 (8.72)0.94 (0.69)0.34 (0.76)1.58 (0.70)0.85 (0.84)126.96 (58.15)0.23 (0.33)
DiT-XL/2-512 [36]51222.99 (9.54)0.96 (0.70)0.27 (0.73)1.90 (0.72)0.86 (0.82)128.98 (55.17)0.21 (0.34)
MAR-B-256 [27]2563.79 (10.36)0.83 (0.72)0.67 (0.71)1.18 (0.72)0.96 (0.75)83.03 (55.78)0.45 (0.38)
MAR-L-256 [27]2563.30 (10.36)0.82 (0.72)0.71 (0.71)1.10 (0.73)0.96 (0.75)81.80 (55.95)0.56 (0.38)
MAR-H-256 [27]2563.11 (10.36)0.82 (0.72)0.72 (0.71)1.07 (0.74)0.96 (0.76)81.37 (55.82)0.64 (0.38)
Latent diffusion, U-Net
LDM-256 [43]25626.09 (37.39)0.96 (0.61)0.21 (0.68)1.80 (0.45)0.83 (0.28)126.94 (30.83)0.16 (0.16)
EDM-2-XS-512 [24]5123.79 (75.02)0.83 (0.42)0.65 (0.63)1.22 (0.25)0.95 (0.13)72.41 (26.95)0.46 (0.09)
EDM-2-S-512 [24]5123.33 (122.48)0.85 (0.33)0.67 (0.42)1.26 (0.16)0.97 (0.07)80.25 (21.34)0.59 (0.04)
EDM-2-M-512 [24]5123.30 (107.45)0.85 (0.36)0.69 (0.61)1.24 (0.19)0.97 (0.09)82.99 (22.19)0.65 (0.06)
EDM-2-L-512 [24]5122.90 (118.87)0.84 (0.23)0.70 (0.51)1.22 (0.11)0.97 (0.06)82.10 (22.89)0.71 (0.03)
EDM-2-XL-512 [24]5122.92 (141.74)0.84 (0.25)0.71 (0.45)1.21 (0.12)0.97 (0.06)83.23 (20.04)0.77 (0.03)
EDM-2-XXL-512 [24]5122.87 (124.29)0.84 (0.33)0.71 (0.60)1.22 (0.17)0.97 (0.07)82.45 (21.52)0.75 (0.05)
", + "bbox": [ + 91, + 87, + 906, + 295 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "shown that this method still lacks diversity, we continue with EDM-2 as baseline for unconditional image generation and compare the trained models with a fixed compute budget. Quantitative results are shown in Tab. 3. The results show that the leveraging pseudo labels improves the sampling diversity compared to EDM-2 in all cases. In three cases the diversity is even better than that of the real reference dataset proofing that the pseudo conditioning works well and we can specifically query the model to generate diverse data. It also improves FID scores compared to the unconditional trainings.", + "bbox": [ + 88, + 387, + 480, + 554 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4d2257386b6993efeb16d1851bb7037dce7e851a99c365b9da215a043efedcad.jpg", + "table_caption": [ + "Table 2. Image generation of ImageNet across different generative models. Numbers in brackets indicate results when sampling the model without conditioning. Inception is used as feature extractor for all metrics except for IRS where we use SwAV. Precision, recall, density, coverage and Vendi score had to be computed on a subset of the training data of 50000 and 10000 for Vendi score. Pretrained conditional models were used as unconditional models by setting conditional guidance to 0." + ], + "table_footnote": [], + "table_body": "
FID ↓IRS∞,a ↑
EDM [24]DiADM (Ours)EDM [24]DiADM (Ours)
ImageNet-51251.5922.280.090.15
FFHQ40.926.240.231.51
ChestX-ray1424.296.760.191.08
CelebV-HQ68.4113.640.180.69
Dynamic13.825.560.601.04
", + "bbox": [ + 91, + 565, + 480, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. FID and IRS results with and without our proposed diversity awareness module", + "bbox": [ + 89, + 666, + 482, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Text-to-image example", + "text_level": 1, + "bbox": [ + 89, + 719, + 302, + 736 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Extending IRS to text-conditional image generation is straightforward. The evaluation of diversity depends on the chosen perspective, which is defined by the reference datasets. By changing the reference dataset, we can address various diversity-related issues, such as fairness. One notable example is the gender bias observed in text-to-image models. For our experiments, we use Deepfloyd (IF-I-XLv1.0) $^2$ . To assess gender diversity, we create a reference dataset with a balanced gender distribution for specific job", + "bbox": [ + 88, + 741, + 482, + 878 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "roles by prompting the model to generate 100 images each of males and females. Half of these images are reserved as a balanced test dataset. We then prompt the model without specifying gender. The results are presented in Sec. 4.2. All groups, including the general term human, show gender bias, with the latter generating only female images. IRS estimates that in all cases, the generated diversity reaches only about $50\\%$ of that of the balanced reference dataset.", + "bbox": [ + 511, + 387, + 906, + 508 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Limitations", + "text_level": 1, + "bbox": [ + 511, + 518, + 638, + 534 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The stochasticity involved for a low number of samples means that minor IRS differences are not meaningful for a low number of samples. However, this is also reflected in the high values of uncertainty. Additionally, we believe that finding a better feature extractor that maximizes measured diversity (unadjusted IRS) on real datasets should be the focus. Experiments on improving diversity are limited to using real features for generation and restrictive computational budget. Furthermore, increasing the conditioning requires similar experiments analyzing memorization issues as text-conditioning [6].", + "bbox": [ + 511, + 542, + 908, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion and Outlook", + "text_level": 1, + "bbox": [ + 511, + 723, + 741, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we reveal that all current methods for evaluating diversity and the feature extractors they are based on are inadequate. Using our proposed IRS metric, we demonstrate that these shortcomings result in all state-of-the-art image generation methods facing challenges with diversity. To address this issue for unconditional diffusion models, we propose DiADMs, which enhance the performance of current methods by separating diversity from fidelity. In future work, we aim to extend our efforts to improve diversity in conditional image generation as well.", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "$^{2}$ https://huggingface.co/DeepFloyd/IF-I-XL-v1.0", + "bbox": [ + 107, + 886, + 370, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3036", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements: This work was supported by the High-Tech-Agenda Bavaria. HPC resources were provided by the Erlangen National High Performance Computing Center (NHR@FAU) of the Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU) under the NHR project b143dc and b180dc. NHR funding is provided by federal and Bavarian state authorities. NHR@FAU hardware is partially funded by the German Research Foundation (DFG) - 440719683. Support was also received by the ERC - projects MIA-NORMAL 101083647 as well as DFG 513220538, 512819079 and DFG large scale infrastructure funding Art 91b GG. This work was supported by the UKRI Centre for Doctoral Training in Artificial Intelligence for Healthcare (EP/S023283/1) and Ultromics Ltd.", + "bbox": [ + 89, + 90, + 485, + 303 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 319, + 187, + 335 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sumukh K Aithal, Pratyush Maini, Zachary C Lipton, and J Zico Kolter. Understanding hallucinations in diffusion models through mode interpolation. arXiv preprint arXiv:2406.09358, 2024. 1", + "[2] Sina Alemohammad, Josue Casco-Rodriguez, Lorenzo Luzi, Ahmed Imtiaz Humayun, Hossein Babaei, Daniel LeJeune, Ali Siahkoohi, and Richard G. Baraniuk. Self-Consuming Generative Models Go MAD, 2023. arXiv:2307.01850 [cs]. 2", + "[3] Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, and Michael Auli. data2vec: A General Framework for Self-supervised Learning in Speech, Vision and Language. In Proceedings of the 39th International Conference on Machine Learning, pages 1298-1312. PMLR, 2022. ISSN: 2640-3498. 6", + "[4] Hritik Bansal, Da Yin, Masoud Monajatipoor, and Kai-Wei Chang. How well can text-to-image generative models understand ethical natural language interventions? arXiv preprint arXiv:2210.15230, 2022. 1, 3", + "[5] Leonard E. Baum and Patrick Billingsley. Asymptotic Distributions for the Coupon Collector's Problem. The Annals of Mathematical Statistics, 36(6):1835-1839, 1965. Publisher: Institute of Mathematical Statistics. 3", + "[6] Nicholas Carlini, Jamie Hayes, Milad Nasr, Matthew Jagielski, Vikash Sehwag, Florian Tramér, Borja Balle, Daphne Ippolito, and Eric Wallace. Extracting Training Data from Diffusion Models, 2023. arXiv:2301.13188 [cs]. 2, 8", + "[7] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised Learning of Visual Features by Contrasting Cluster Assignments. In Advances in Neural Information Processing Systems, pages 9912-9924. Curran Associates, Inc., 2020. 6", + "[8] Tingxiu Chen, Yilei Shi, Zixuan Zheng, Bingcong Yan, Jingliang Hu, Xiao Xiang Zhu, and Lichao Mou. Ultrasound Image-to-Video Synthesis via Latent Dynamic Diffusion Models. In Medical Image Computing and Computer Assisted Intervention - MICCAI 2024, pages 764-774. Springer Nature Switzerland, Cham, 2024. Series Title: Lecture Notes in Computer Science. 2" + ], + "bbox": [ + 99, + 344, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[9] Ching-Yao Chuang, Varun Jampani, Yuanzhen Li, Antonio Torralba, and Stefanie Jegelka. Debiasing vision-language models via biased prompts. arXiv preprint arXiv:2302.00070, 2023. 3", + "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. 2015. 6", + "[11] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 1, 2", + "[12] Prafulla Dhariwal and Alex Nichol. Diffusion Models Beat GANs on Image Synthesis, 2021. arXiv:2105.05233 [cs, stat]. 3, 5, 8", + "[13] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. Cogview: Mastering text-to-image generation via transformers. Advances in neural information processing systems, 34:19822-19835, 2021. 2, 3", + "[14] Mischa Dombrowski, Hadrien Reynaud, and Bernhard Kainz. Uncovering Hidden Subspaces in Video Diffusion Models Using Re-Identification, 2024. arXiv:2411.04956. 2, 7, 5, 6", + "[15] Chun-Mei Feng, Kai Yu, Yong Liu, Salman Khan, and Wangmeng Zuo. Diverse Data Augmentation with Diffusions for Effective Test-time Prompt Tuning. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 2704-2714, Paris, France, 2023. IEEE. 2", + "[16] Dan Friedman and Adji Bousso Dieng. The Vendi Score: A Diversity Evaluation Metric for Machine Learning, 2023. arXiv:2210.02410. 3", + "[17] Felix Friedrich, Manuel Brack, Lukas Struppek, Dominik Hintersdorf, Patrick Schramowski, Sasha Luccioni, and Kristian Kersting. Fair Diffusion: Instructing Text-to-Image Generation Models on Fairness, 2023. arXiv:2302.10893. 2", + "[18] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre H. Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, Bilal Piot, Koray Kavukcuoglu, Rémi Munos, and Michal Valko. Bootstrap your own latent: A new approach to self-supervised Learning, 2020. arXiv:2006.07733 [cs, stat]. 6", + "[19] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked Autoencoders Are Scalable Vision Learners. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15979-15988, New Orleans, LA, USA, 2022. IEEE. 6", + "[20] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 3, 6", + "[21] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 6", + "[22] Jonathan Ho and Tim Salimans. Classifier-Free Diffusion Guidance. 2022. 5" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3037", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[23] Tero Karras, Samuli Laine, and Timo Aila. A Style-Based Generator Architecture for Generative Adversarial Networks. 2019. 5, 6", + "[24] Tero Karras, Miika Aittala, Jaakko Lehtinen, Janne Hellsten, Timo Aila, and Samuli Laine. Analyzing and Improving the Training Dynamics of Diffusion Models, 2024. arXiv:2312.02696 [cs, stat]. 5, 8", + "[25] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 3", + "[26] Tuomas Kynkänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved Precision and Recall Metric for Assessing Generative Models. 2019. 3", + "[27] Tianhong Li, Yonglong Tian, He Li, Mingyang Deng, and Kaiming He. Autoregressive Image Generation without Vector Quantization, 2024. arXiv:2406.11838 [cs]. 3, 5, 8", + "[28] Jiawei Liu, Qiang Wang, Huijie Fan, Yinong Wang, Yandong Tang, and Liangqiong Qu. Residual denoising diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2773-2783, 2024. 2, 3", + "[29] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A ConvNet for the 2020s, 2022. arXiv:2201.03545. 6", + "[30] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations, 2022. 2, 3", + "[31] Zichen Miao, Jiang Wang, Ze Wang, Zhengyuan Yang, Lijuan Wang, Qiang Qiu, and Zicheng Liu. Training diffusion models towards diverse image generation with reinforcement learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10844-10853, 2024. 1, 3", + "[32] Muhammad Ferjad Naeem, Seong Joon Oh, Youngjung Uh, Yunjey Choi, and Jaejun Yoo. Reliable Fidelity and Diversity Metrics for Generative Models. 2020. 3", + "[33] Jamie A. OaReilly and Fawad Asadi. Pre-trained vs. Random Weights for Calculating Fréchet Inception Distance in Medical Imaging. In 2021 13th Biomedical Engineering International Conference (BMEiCON), pages 1-4, Ayutthaya, Thailand, 2021. IEEE. 6, 5", + "[34] Maxime Oquab, Timothee Darct, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINoV2: Learning Robust Visual Features without Supervision, 2024. arXiv:2304.07193 [cs]. 6, 2", + "[35] David Ouyang, Bryan He, Amirata Ghorbani, Neal Yuan, Joseph Ebinger, Curtis P. Langlotz, Paul A. Heidenreich, Robert A. Harrington, David H. Liang, Euan A. Ashley, and" + ], + "bbox": [ + 91, + 92, + 480, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "James Y. Zou. Video-based AI for beat-to-beat assessment of cardiac function. Nature, 580(7802):252-256, 2020. Publisher: Springer US. 6, 7, 5", + "[36] William Peebles and Saining Xie. Scalable Diffusion Models with Transformers. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 4172-4182, Paris, France, 2023. IEEE. 3, 5, 8", + "[37] Yiming Qin, Huangjie Zheng, Jiangchao Yao, Mingyuan Zhou, and Ya Zhang. Class-balancing diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18434-18443, 2023. 3", + "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning Transferable Visual Models From Natural Language Supervision, 2021. arXiv:2103.00020. 6", + "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3", + "[40] Hadrien Reynaud, Qingjie Meng, Mischa Dombrowski, Arijit Ghosh, Thomas Day, Alberto Gomez, Paul Leeson, and Bernhard Kainz. EchoNet-Synthetic: Privacy-preserving Video Generation for Safe Medical Data Sharing, 2024. arXiv:2406.00808 [cs]. 6, 7, 5", + "[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-Resolution Image Synthesis with Latent Diffusion Models. 2021. 3, 5", + "[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 3", + "[43] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-Resolution Image Synthesis with Latent Diffusion Models, 2022. arXiv:2112.10752 [cs]. 8", + "[44] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. DreamBooth: Fine Tuning Text-to-Image Diffusion Models for Subject-Driven Generation. 2022. 2", + "[45] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, Xi Chen, and Xi Chen. Improved Techniques for Training GANs. 2016. 6", + "[46] Axel Sauer, Katja Schwarz, and Andreas Geiger. StyleGANXL: Scaling StyleGAN to Large Diverse Datasets, 2022. arXiv:2202.00273 [cs]. 3", + "[47] Vikash Sehwag, Caner Hazirbas, Albert Gordo, First Ozgenel, and Cristian Canton. Generating high fidelity data from low-density regions using diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11492-11501, 2022. 2", + "[48] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3038", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] George Stein, Jesse C. Cresswell, Rasa Hosseinzadeh, Yi Sui, Brendan Leigh Ross, Valentin Villecloze, Zhaoyan Liu, Anthony L. Caterini, J. Eric T. Taylor, and Gabriel Loaiza-Ganem. Exposing flaws of generative model evaluation metrics and their unfair treatment of diffusion models, 2023. arXiv:2306.04675 [cs, stat]. 2, 5, 6", + "[50] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the Inception Architecture for Computer Vision. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2818-2826, Las Vegas, NV, USA, 2016. IEEE. 6", + "[51] N. M. Temme. Asymptotic Estimates of Stirling Numbers. Studies in Applied Mathematics, 89(3):233-243, 1993. 4", + "[52] Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, Mohammadhadi Bagheri, and Ronald M Summers. ChestX-ray8: Hospital-Scale Chest X-Ray Database and Benchmarks on Weakly-Supervised Classification and Localization of Common Thorax Diseases. 2017. 6", + "[53] Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang. Diffusion models: A comprehensive survey of methods and applications. ACM Computing Surveys, 56(4): 1-39, 2023. 1, 3", + "[54] Ling Yang, Jingwei Liu, Shenda Hong, Zhilong Zhang, Zhilin Huang, Zheming Cai, Wentao Zhang, and Bin Cui. Improving diffusion-based image synthesis with context prediction. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[55] Cheng Zhang, Xuanbai Chen, Siqi Chai, Chen Henry Wu, Dmitry Lagun, Thabo Beeler, and Fernando De la Torre. Itigen: Inclusive text-to-image generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3969-3980, 2023. 2, 3", + "[56] Weitong Zhang, Chengqi Zang, Liu Li, Sarah Cechnicka, Cheng Ouyang, and Bernhard Kainz. Stability and generalizability in sde diffusion models with measure-preserving dynamics. In NeurIPS, 2024. 1, 2", + "[57] Yi Zhang, Xiaoyu Shi, Dasong Li, Xiaogang Wang, Jian Wang, and Hongsheng Li. A unified conditional framework for diffusion-based image restoration. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[58] Dora Zhao, Jerone T. A. Andrews, Orestis Papakyriakopoulos, and Alice Xiang. Position: Measure Dataset Diversity, Don't Just Claim It, 2024. arXiv:2407.08188. 2", + "[59] Hao Zhu, Wayne Wu, Wentao Zhu, Liming Jiang, Siwei Tang, Li Zhang, Ziwei Liu, and Chen Change Loy. CelebV-HQ: A Large-Scale Video Facial Attributes Dataset, 2022. arXiv:2207.12393 [cs]. 6" + ], + "bbox": [ + 91, + 90, + 480, + 768 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "3039", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Image Generation Diversity Issues and How to Tame Them/7493d326-112b-4a6d-8c2c-7e4449ad1384_model.json b/2025/Image Generation Diversity Issues and How to Tame Them/7493d326-112b-4a6d-8c2c-7e4449ad1384_model.json new file mode 100644 index 0000000000000000000000000000000000000000..de33cf3205b1b3f401fe4a41e029cd55878e586b --- /dev/null +++ b/2025/Image Generation Diversity Issues and How to Tame Them/7493d326-112b-4a6d-8c2c-7e4449ad1384_model.json @@ -0,0 +1,2070 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.202, + 0.131, + 0.797, + 0.152 + ], + "angle": 0, + "content": "Image Generation Diversity Issues and How to Tame Them" + }, + { + "type": "text", + "bbox": [ + 0.238, + 0.18, + 0.76, + 0.2 + ], + "angle": 0, + "content": "Mischa Dombrowski1 Weitong Zhang2 Sarah Cechnicka2" + }, + { + "type": "text", + "bbox": [ + 0.332, + 0.2, + 0.667, + 0.216 + ], + "angle": 0, + "content": "Hadrien Reynaud² Bernhard Kainz1,2" + }, + { + "type": "text", + "bbox": [ + 0.161, + 0.216, + 0.838, + 0.235 + ], + "angle": 0, + "content": "\\(^{1}\\)Friedrich-Alexander-Universität Erlangen-Nürnberg \\(^{2}\\)Imperial College London" + }, + { + "type": "text", + "bbox": [ + 0.391, + 0.238, + 0.605, + 0.251 + ], + "angle": 0, + "content": "mischa.dombrowski@fau.de" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.287, + 0.327, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.37, + 0.485, + 0.884 + ], + "angle": 0, + "content": "Generative methods have reached a level of quality that is almost indistinguishable from real data. However, while individual samples may appear unique, generative models often exhibit limitations in covering the full data distribution. Unlike quality issues, diversity problems within generative models are not easily detected by simply observing single images or generated datasets, which means we need a specific measure to assess the diversity of these models. In this paper, we draw attention to the current lack of diversity in generative models and the inability of common metrics to measure this. We achieve this by framing diversity as an image retrieval problem, where we measure how many real images can be retrieved using synthetic data as queries. This yields the Image Retrieval Score (IRS), an interpretable, hyperparameter-free metric that quantifies the diversity of a generative model's output. IRS requires only a subset of synthetic samples and provides a statistical measure of confidence. Our experiments indicate that current feature extractors commonly used in generative model assessment are inadequate for evaluating diversity effectively. Consequently, we perform an extensive search for the best feature extractors to assess diversity. Evaluation reveals that current diffusion models converge to limited subsets of the real distribution, with no current state-of-the-art models superpassing \\(77\\%\\) of the diversity of the training data. To address this limitation, we introduce Diversity-Aware Diffusion Models (DiADM), a novel approach that improves diversity of unconditional diffusion models without loss of image quality. We do this by disentangling diversity from image quality by using a diversity aware module that uses pseudo-unconditional features as input. We provide a Python package offering unified feature extraction and metric computation to further facilitate the evaluation of generative models https://github.com/MischaD/beyondfid." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.286, + 0.605, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.538, + 0.438, + 0.583, + 0.446 + ], + "angle": 0, + "content": "Low Diversity" + }, + { + "type": "image", + "bbox": [ + 0.608, + 0.286, + 0.692, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.624, + 0.438, + 0.67, + 0.446 + ], + "angle": 0, + "content": "High Diversity" + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.287, + 0.905, + 0.437 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.767, + 0.438, + 0.862, + 0.447 + ], + "angle": 0, + "content": "Predicted Diversity \\(\\mathrm{IRS}_{\\infty}\\) \\((\\%)\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.458, + 0.907, + 0.582 + ], + "angle": 0, + "content": "Figure 1. Predicted gender diversity - as one possible example for diversity - after sampling pre-trained text-to-image diffusion models infinitely often. Visually it is apparent that most terms are inherently biased toward one gender. Our proposed IRS score predicts that off-the-shelf models will only ever reach 50 percent diversity, which is equivalent to one gender being perfectly represented while the other gender is missing entirely. Using diversification strategies increases diversity. Details about this can be found in Sec. 4.4." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.615, + 0.645, + 0.631 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.644, + 0.907, + 0.903 + ], + "angle": 0, + "content": "With advancements in large-scale models and datasets, diffusion models have demonstrated substantial capacity to capture high-dimensional, complex distributions, enabling photo-realistic conditional image generation across diverse guidance types, including image-based [56], classifier-based [11], and text-based guidance [42]. However, diffusion models frequently inherit and amplify data biases due to their reliance on large-scale, naturally biased datasets [4, 53], limiting comprehensive distributional coverage in generated samples. For instance, text-to-image diffusion models often exhibit gender bias for certain occupations. When prompted with \"a teacher\", models predominantly generate females, with male representations notably underrepresented, as shown in Fig. 1. The \"conditional mode collapse\" [1, 31] also affects image- [56] and class-conditional [11] diffusion models, where the guidance often produces overly uniform outputs, leading to a loss in diver" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3029" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.5 + ], + "angle": 0, + "content": "sity in the sampled distribution. Dataset diversity remains an unresolved issue, even for real datasets [58], due to ambiguous definitions and concepts. In contrast, synthetic data can leverage its training data as a reference. To avoid problems like mode drift [2], synthetic data should be at least as diverse as its training set. Key motivations for ensuring diversity include content creation [44], addressing privacy and memorization concerns [6], mitigating the lack of diversity in synthetic datasets [14], data augmentation [15], promoting algorithmic fairness [17], and enhancing downstream performance [8]. To improve diversity in conditional image generation, [28] introduces noise perturbations to enhance diversity, though at the expense of image quality and certainty. [30] balances a trade-off between faithfulness and realism in diverse image synthesis, where diversity comes at the cost of fidelity. These diversity-oriented approaches highlight an inherent compromise: diversity is achieved by sacrificing generation quality, with randomness that lacks interpretability. In parallel, classifier-based diversity guidance [47] and text-based methods [13, 55] have been developed to reduce bias in text-to-image generation. However, these techniques are task-specific and lack generalizability across different types of diffusion models for diversity enhancement. Thus, a fundamental challenge emerges: a universal, interpretable metric to evaluate diversity without bias or conditional dependence, and without compromising generation quality." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.501, + 0.484, + 0.713 + ], + "angle": 0, + "content": "To address this, we propose the Image Retrieval Score (IRS), an interpretable metric for quantifying diversity in generative models. IRS assesses how effectively synthetic data retrieves real images as queries, providing a hyperparameter-free, statistically grounded evaluation based on minimal samples. Our approach is motivated by a thought experiment that illustrates how individual samples reveal model diversity. Using IRS as a core, generic metric for diversity, we quantify and reveal two key findings: (1) existing feature extractors used to evaluate generative models lack the capacity to measure diversity effectively, and (2) current diffusion models converge on limited subsets of the real distribution, with no state-of-the-art models surpassing \\(77\\%\\) of the diversity in training data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.714, + 0.484, + 0.837 + ], + "angle": 0, + "content": "This further motivates the use of IRS as an unbiased guidance to drive a comprehensive search for optimal feature extractors, enhancing diversity evaluation without compromising generation quality. Our selection is based on [49] who experimented on the impact of feature extractors on existing generative metrics. Building on these insights, we introduce Diversity-Aware Diffusion Models (DiADM), an approach that improves diversity across the full data distri" + }, + { + "type": "image", + "bbox": [ + 0.543, + 0.096, + 0.882, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.303, + 0.908, + 0.444 + ], + "angle": 0, + "content": "Figure 2. We model finding image pairs (image retrieval) as randomly drawing from urns with replacement (Ideal). We observe that all used feature extraction models exhibit performance issues by collapsing in the feature space for real images resulting in a measurement gap affecting currently used metrics such as \\(\\mathrm{FID}_{\\mathrm{Inception}}\\), \\(\\mathrm{FID}_{\\mathrm{DINOv2}}\\), Precision and Recall. Comparing the results of synthetic images to real data shows that datasets generated by generative models show even stronger diversity issues which result in a synthetic data distribution gap. Our proposed metric leverages a real reference dataset to remove this measurement gap." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.469, + 0.907, + 0.544 + ], + "angle": 0, + "content": "bution in unconditional diffusion models without compromising image quality. DiADM accomplishes this by disentangling diversity from image quality through a diversity-aware module utilizing pseudo-unconditional features. Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.545, + 0.905, + 0.589 + ], + "angle": 0, + "content": "- We demonstrate the limitations of widely-used feature extractors for diversity metrics, highlighting their poor performance on real datasets." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.59, + 0.905, + 0.62 + ], + "angle": 0, + "content": "- We propose an intuitive framework for assessing dataset diversity in unconditional diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.621, + 0.905, + 0.65 + ], + "angle": 0, + "content": "- We identify and quantify the reduction in diversity resulting from current training methodologies." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.651, + 0.905, + 0.694 + ], + "angle": 0, + "content": "- We provide a comprehensive analysis showing limited image diversity across a wide range of state-of-the-art diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.696, + 0.905, + 0.755 + ], + "angle": 0, + "content": "- We introduce a diversity-aware plug-in module, which utilizes pseudo-labels to separate diversity from fidelity, thereby enhancing generative diversity in unconditional diffusion models." + }, + { + "type": "list", + "bbox": [ + 0.515, + 0.545, + 0.905, + 0.755 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.77, + 0.657, + 0.785 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Image Generation: Diffusion models [48] progressively map complex distributions to a standard Gaussian. Conditional diffusion models have demonstrated a strong ability to generate highly realistic images that align closely with specified conditions, excelling in tasks such as image synthesis [54], image restoration [57], and image reconstruction [56]. Additionally, [11] enhances diffusion" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.851, + 0.483, + 0.901 + ], + "angle": 0, + "content": "1Consider a die purported to have 1000 sides. If rolled 1000 times but yielding only 10 distinct outcomes, one would naturally question the claim. Similarly, in generative models, each sample serves as a \"roll\", reflecting the model's diversity in an unconditional manner." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3030" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.258 + ], + "angle": 0, + "content": "models with an auxiliary classifier, enabling high-fidelity, class-conditional generation, while text-conditional models generate images based on text prompts [13, 55]. However, diffusion models often struggle to achieve sufficient diversity in task-agnostic generation, especially when trained on large-scale datasets that carry inherent biases [4, 53]. For example, Stable Diffusion, based on latent diffusion models (LDM) [39, 42], produces high-quality images aligned with text prompts, yet frequently reflects bias (e.g., age, gender) for certain terms. These limitations lead to restricted distributional coverage and biased outputs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.285, + 0.483, + 0.542 + ], + "angle": 0, + "content": "Diversity in Generation: Research on diversity in diffusion models is limited, with most existing approaches focused on GAN-based models [9, 25]. Current methods for diversity in diffusion models fall into the primary categories: bias mitigation and diversity enhancement. In bias mitigation, [37] uses a re-weighted loss to address long-tailed distributions, promoting balanced class representation. For text-conditional models, [13] attempts to mitigate bias by adding attribute words to prompts, though the added prompts may be ambiguous. [31] extends this to both class- and text-conditional models, using reinforcement learning to match the diversity ratio in reference images but relying on task-specific diversity reward functions. For diversity enhancement, [28] introduces noise perturbations, which increase diversity at the expense of image quality and certainty. Similarly, [30] balances diversity with realism, where gains in diversity tend to compromise fidelity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Metrics of Diversity: Existing metrics in diffusion models remain limited and are unsuitable for accurate and interpretable diversity assessment. Traditional image quality metrics like FID [20] require extensive sampling (e.g., 50k images) to be reliable, making them computationally intensive and challenging to interpret in practical applications. Precision and recall [26] are most established in literature [12, 36, 41, 46]. They add supplementary insights but are hindered by dependency on dataset size, hyperparameters, and interpretability issues. Credibility and interpretability of metrics are highly underestimated properties. For example, a comparison of three years of advancements in diffusion models, as seen in [12] and [27], shows an order-of-magnitude improvement in FID but a one percentage point reduction in precision. Coverage and density [32] also measure diversity but are dependant on the ratio between train and test set, highly depend on hyperparameters, lack evaluation on real datasets, and density was designed to ignore real outliers which are important in the context of diversity assessment. Vendi Score [16], which operates entirely without a real reference set, sacrifices fidelity and is computationally prohibitive for large datasets." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.605, + 0.106 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.115, + 0.905, + 0.178 + ], + "angle": 0, + "content": "Let \\(\\mathbf{x_t}\\) be image \\(t\\) from a dataset consisting of \\(\\mathrm{N}_{train}\\) real images residing in image space \\(\\mathcal{X} \\in \\mathbb{R}^{x\\times h\\times w}\\). Unconditional generative models aim to learn the distribution \\(p_{data}(\\mathbf{X})\\) and sample \\(\\mathrm{N}_{\\text{sample}}\\) synthetic images from it." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.184, + 0.757, + 0.2 + ], + "angle": 0, + "content": "3.1. Coupon Collector Problem" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.206, + 0.907, + 0.614 + ], + "angle": 0, + "content": "The performance of generative models is generally measured by assessing the similarity of the real and the synthetic datasets. It is assumed that the training samples represent discrete observations from a continuous unknown distribution. We suggest viewing all synthetic images as a composition of real images, which holds true as long as there is no training objective that enforces out-of-distribution generation. To assess diversity, we argue that every training sample should be the main component of at least one synthetic sample. To measure this we use the fact that the observed training distribution is discrete. A model that has perfectly memorized every training sample would generate all images with equal probability \\(\\frac{1}{\\mathrm{N}_{train}}\\). Consequently, we can model sampling synthetic images with an urn experiment. Each synthetic sample consists mainly of one component which is then compounded with other training images. A real sample is considered learned if it serves as the main component for at least one synthetic image. Evaluating diversity in this context can be framed as the coupon collector problem [5]. Let \\(P_{n}(N_{t} > 0)\\) denote the probability of generating \\(n\\) samples where at least one corresponds to a training image \\(t\\), with \\(N_{t}\\) denoting the number of generated images a training image is associated with. This probability can be computed by considering the compound probability of not sampling this training image which is \\(\\frac{\\mathrm{N}_{train} - 1}{\\mathrm{N}_{train}}\\). Since the same probability applies to each image the expected value of the diversity can be calculated as" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.633, + 0.907, + 0.669 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} [ \\text {D i v e r s i t y} ] = \\left(1 - \\left(\\frac {\\mathrm {N} _ {\\text {t r a i n}} - 1}{\\mathrm {N} _ {\\text {t r a i n}}}\\right) ^ {n}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.751 + ], + "angle": 0, + "content": "In this definition, diversity is equal to the percentage of learned real images. This idealized scenario is shown as ideal in Fig. 2. Therefore, assessing diversity comes down to predicting which training image corresponds to which synthetic image." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.758, + 0.673, + 0.774 + ], + "angle": 0, + "content": "3.2. Image Retrieval" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.903 + ], + "angle": 0, + "content": "To assess diversity we first need to define how to decide on what constitutes a main component for each synthetic image. For each synthetic image instance \\(\\mathbf{x}_{\\mathrm{t}}^{\\prime}\\) we aim to identify the real image \\(\\mathbf{x}_{\\mathrm{t}}\\) that is closest to it according to a measurement \\(\\mathcal{P}\\) (e.g., Euclidean distance) on the feature of the images extracted by a pretrained feature extractor \\(\\mathcal{F}\\). Then, to assess the diversity we have to compute the size of the dataset of learned images:" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "3031" + } + ], + [ + { + "type": "equation", + "bbox": [ + 0.091, + 0.115, + 0.496, + 0.16 + ], + "angle": 0, + "content": "\\[\n\\mathcal {X} _ {\\text {l e a r n e d}} = \\left\\{\\mathbf {x} _ {t} \\in \\mathcal {X} \\mid \\exists \\mathbf {x} _ {t} ^ {\\prime} \\in \\mathcal {X} ^ {\\prime}: \\mathbf {x} _ {t} = \\arg \\min _ {\\mathbf {x} _ {t}} \\mathcal {P} \\left(\\mathbf {x} _ {t}, \\mathbf {x} _ {t} ^ {\\prime}\\right) \\right\\}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.162, + 0.411, + 0.176 + ], + "angle": 0, + "content": "We define the cardinality of this set as \\(\\mathbf{N}_{\\text{learned}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.177, + 0.483, + 0.207 + ], + "angle": 0, + "content": "We then call the trained image instance \\(\\mathbf{x}_{\\mathrm{t}}\\) \"learned\" and compute the diversity of the synthetic dataset as:" + }, + { + "type": "equation", + "bbox": [ + 0.199, + 0.218, + 0.483, + 0.25 + ], + "angle": 0, + "content": "\\[\n\\mathrm {I R S} _ {\\alpha} = \\frac {\\mathrm {N} _ {\\text {l e a r n e d}}}{\\mathrm {N} _ {\\text {t r a i n}}} \\in [ 0, 1 ] \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.259, + 0.484, + 0.486 + ], + "angle": 0, + "content": "We term this quantity as the image retrieval score (IRS). To emphasize that \\(\\mathrm{IRS}_{\\alpha}\\) depends on the number of images sampled we add \\(\\alpha := \\frac{\\mathrm{N}_{\\text{sample}}}{\\mathrm{N}_{\\text{train}}}\\) as a subscript. Note that all other generative metrics also depend on \\(\\alpha\\) but often ignore this dependency. The name image retrieval score derives from the fact that computing IRS also works when comparing two disjoint real datasets. In this context the problem can be thought of as an image retrieval task, where a real image serves as the query image and the goal is to find the most similar image in a comparison dataset. Intuitively, IRS measures the ratio of images in a training dataset that are retrievable by another dataset. If the comparison dataset consists of real images we say that an image is \"retrievable\". If the comparison dataset consists of synthetic images we say that an image is \"learned\"." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.487, + 0.484, + 0.836 + ], + "angle": 0, + "content": "The key advantage of IRS is that we can give a reasonably good estimate of dataset diversity with only a minimal number of samples. For example, if the same image appears three times within just ten samples, it is a strong indicator of overfitting. Consequently, we aim to estimate the diversity after \\( n << \\mathrm{N}_{train} \\) samples. To compute the probability of retrieving exactly \\( k = \\mathcal{X}_{\\text{learned}} \\) images we use standard combinatorics. This problem can be split into two parts. Firstly, we compute the number of possibilities of splitting \\( n \\) into \\( k \\) different subsets. Instinctively, we are computing the number of combinations in which different synthetic samples derive from the same retrieved (still unassigned) image \\( \\mathbf{x}_t \\) according to equation Eq. (2). The computation involves using Stirling's number of the second kind, \\( \\operatorname{Stir}(n, k) \\), which should be multiplied by the total number of possible assignments for each subset, given by \\( \\frac{\\mathrm{N}_{train}!}{(\\mathrm{N}_{train} - \\mathrm{N}_{\\text{learned}})!} \\). This will become more clear with an example. Say, the first subset consists of three different images \\( x_3', x_{10}', x_{11}' \\), all of which are learned from the same real image. Then for the first of \\( k \\) subsets, there are \\( \\mathrm{N}_{train} \\) different choices. For the second subset, we have \\( \\mathrm{N}_{train} - 1 \\) choices and so on. Dividing by all choices leaves us with the probability for generating \\( k = \\mathrm{N}_{\\text{learned}} \\) different images:" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.847, + 0.483, + 0.88 + ], + "angle": 0, + "content": "\\[\n\\mathrm {P} (k, n, s) = \\frac {\\operatorname {S t i r} (n , k) * s !}{(s - k) ! * (s) ^ {n}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.903 + ], + "angle": 0, + "content": "with \\( k = \\mathrm{N}_{\\text{learned}} \\) and \\( s = \\mathrm{N}_{\\text{train}} \\) and \\( n = \\mathrm{N}_{\\text{sample}} \\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.122 + ], + "angle": 0, + "content": "For large values, we estimate this by computing the log of Eq. (4) instead. [51] showed that" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.147, + 0.905, + 0.199 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S t i r} (n, k) = \\sqrt {\\frac {v - 1}{v (1 - G)}} \\left(\\frac {v - 1}{v - G}\\right) ^ {n - k} \\frac {k ^ {n}}{n ^ {k}} e ^ {k (1 - G)} \\binom {n} {k}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.2, + 0.907, + 0.305 + ], + "angle": 0, + "content": "where \\( v = \\frac{n}{k} \\) and \\( G \\in (0,1) \\) is the unique solution to \\( G = v e^{G - v} \\). Asymptotically, this estimate is close, reaching a maximum relative error of roughly \\( \\frac{0.066}{n} \\) if \\( n \\) and \\( k \\) are small, e.g., (10, 3) [51]. Computing this in log-space for ImageNet roughly takes 3 seconds and using memorization techniques reduces consecutive computations to 1.3 seconds." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.316, + 0.722, + 0.332 + ], + "angle": 0, + "content": "3.3. Image Retrieval Score" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.337, + 0.906, + 0.415 + ], + "angle": 0, + "content": "Given Eq. (5) we can now estimate IRS as well as a confidence interval for it. We assume that \\(\\mathrm{N}_{\\text {learned }}\\) and \\(\\mathrm{N}_{\\text {sample }}\\) are fixed and we aim to estimate the model diversity. To compute the maximum likelihood estimate for IRS, we optimize over values of \\(s \\in \\mathbb{N} \\cap [\\mathrm{N}_{\\text {learned }}, \\mathrm{N}_{\\text {train }}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.425, + 0.906, + 0.458 + ], + "angle": 0, + "content": "\\[\n\\mathrm {I R S} _ {\\infty} = \\frac {1}{\\mathrm {N} _ {t r a i n}} \\arg \\max _ {s} \\mathrm {P} (\\mathrm {N} _ {l e a r n e d}, \\mathrm {N} _ {s a m p l e}, s), (6)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.466, + 0.907, + 0.542 + ], + "angle": 0, + "content": "which means we expect \\(\\mathrm{N}_{\\mathrm{learned},\\infty} = \\mathrm{N}_{\\mathrm{train}}*\\mathrm{IRS}_{\\infty}\\) different images to be sampled eventually. To compute the lower bound we want to minimize the probability of the real diversity being lower but sampling abnormally many different images early. Mathematically, we estimate this as:" + }, + { + "type": "equation", + "bbox": [ + 0.552, + 0.565, + 0.905, + 0.645 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathrm {I R S} _ {\\infty , L} = \\frac {1}{N _ {\\text {t r a i n}}} \\arg \\max _ {s \\in \\mathbb {N} \\cap [ N _ {\\text {l e a n e d}}, \\infty , N _ {\\text {t r a i n}} ]} \\\\ \\min \\left(N _ {\\text {t r a i n}}, N _ {\\text {s a m p l e}}\\right) \\tag {7} \\\\ \\sum_ {k = N _ {\\text {l e a n e d}}} \\quad \\mathrm {P} (k, N _ {\\text {s a m p l e}}, s) > \\alpha_ {e} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.657, + 0.906, + 0.687 + ], + "angle": 0, + "content": "Accordingly, we can compute the estimate for the upper boundary:" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.71, + 0.905, + 0.784 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathrm {I R S} _ {\\infty , U} = \\frac {1}{N _ {\\text {t r a i n}}} \\arg \\min _ {s \\in \\mathbb {N} \\cap \\left[ N _ {\\text {l e a n e d}}, N _ {\\text {l e a n e d}, \\infty} \\right]} \\\\ \\sum_ {k = 1} ^ {s} \\mathrm {P} (k, N _ {\\text {s a m p l e}}, s) > \\alpha_ {e} \\tag {8} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.907, + 0.902 + ], + "angle": 0, + "content": "for a desired probability of error \\(\\alpha_{e}\\). Intuitively, this value represents the most likely estimate of \\(N_{\\mathrm{learned},\\infty}\\) that would produce \\(N_{\\mathrm{learned}}\\) unique samples after drawing \\(N_{\\mathrm{sample}}\\) images. We use the infinity symbol to indicate that this estimate represents the potential diversity if we sampled an infinite number of images. This value also reflects the model's diversity, representing the percentage of samples that the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3032" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.089, + 0.278, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.287, + 0.089, + 0.48, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.219, + 0.483, + 0.276 + ], + "angle": 0, + "content": "Figure 3. Overview of our proposed DiADM model. Instead of using placeholder labels for unconditional generation, we propose to leverage precomputed features from \\(\\mathcal{F}\\) instead. That way we can disentangle fidelity (FID) from diversity (IRS)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.303, + 0.484, + 0.364 + ], + "angle": 0, + "content": "model can generate at its limit. In Appx. 7 we show the dependency of IRS and the confidence intervals on the number of samples showing that at \\(\\alpha = 0.1\\), the estimates start to get very confident around the ground truth value." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.374, + 0.379, + 0.391 + ], + "angle": 0, + "content": "3.4. Adjusting for Measurement Gap" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.396, + 0.484, + 0.703 + ], + "angle": 0, + "content": "In Sec. 4 we will show the measurement gap stemming from feature extractors collapsing to smaller representation spaces. This results in the underestimation of diversity for real data which in turn also influences the measured diversity for synthetic data. To eliminate the influence of the measurement gap we adjust for the reduced amount of diversity in the feature space of the extractor by normalizing by the measured diversity of real data according to the feature extractor. Effectively, we compute \\(\\mathrm{IRS}_{\\infty ,real}\\) on real data using a reference dataset. Then we compute \\(\\mathrm{IRS}_{\\infty ,snth}\\) on synthetic data and report the adjusted \\(\\mathrm{IRS}_{\\infty ,a} = \\frac{\\mathrm{IRS}_{\\infty,snth}}{\\mathrm{IRS}_{\\infty,real}}\\) instead. Intuitively, \\(\\mathrm{IRS}_{\\infty ,a}\\) computes the ratio of diversity in the feature space of the synthetic images compared to the real images. Note that this implies that the adjusted score can be above one for low \\(\\alpha\\) values (see Fig. 4). We provide additional theoretical analysis in Appx. 15.1 and more quantitative results in Appx. 15.2, to demonstrate effectiveness in enhancing sensitivity and reliability for measuring diversity in generative models compared to other metrics." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.713, + 0.388, + 0.729 + ], + "angle": 0, + "content": "3.5. Diversity-Aware Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To address diversity issues, we propose DiADMs. These models leverage the ability to extract meaningful representations using Inception v3 as feature extractor without relying on labels. Rather than employing placeholder labels, as is common in unconditional image generation, we use these pre-computed image features and feed them directly into the model. We term this approach 'pseudo-unconditional' because, while the architecture resembles a conditional setup, no explicit labels are needed. In principle, any feature extractor or custom self-supervised model can be used. However, we use the pre-trained ImageNet model, which is" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.35 + ], + "angle": 0, + "content": "widely recognized as an effective feature extractor. The core idea is that, with proper training, the model behaves as if each training instance represents its own class, allowing for direct diversification of training instances during sampling time. Crucially, this approach decouples the diversity of the generated images from their fidelity as shown in Fig. 3. The diffusion model is responsible for producing high-fidelity images, while the diversity module ensures comprehensive coverage of the training distribution's diversity. We directly utilize features extracted from \\(\\mathcal{F}\\) of the training dataset to generate synthetic data, resulting in a synthetic dataset that should maintain diversity if trained properly. The backbone model employed is the XS architecture as proposed by [24], with modifications to adapt the label dimensionality to match the feature dimensionality of \\(\\mathcal{F}\\). Inception is used as the feature extraction model for all datasets to maintain generalizability." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.362, + 0.645, + 0.379 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.388, + 0.907, + 0.735 + ], + "angle": 0, + "content": "Image Generation Models: Since diversity is mostly underexplored we aim to compare a series of different state-of-the-art methods against one another. [12] introduced classifier guidance and other modifications to Diffusion Models that first lead to competitive performance. Unlike all the current methods, they operate in image space. [41] were the first to introduce a model that performs diffusion in latent space. [24] use several improved training and design strategies for the U-Net backbone. [36] propose to use a Transformer architecture. [27] use autoregressive prediction and leverage masking for efficient training. Our experiments focus on both label-conditional and unconditional generative models, with an emphasis on enhancing diversity in unconditional image generation. This focus is motivated by the observation that any form of conditioning ultimately reduces to unconditional generation over the marginal distribution defined by the conditioning. We use pre-trained models directly and if no unconditional implementation was provided we mimic it by setting the class conditioning to the placeholder value used for unconditional training [22]. For our experiment with DiADM we train models ourselves using a fixed compute budget of 574 A40 GPU hours, which is only one-tenth of the training length in [24]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.886 + ], + "angle": 0, + "content": "Metrics: We sample each model 50000 times which is a common choice to get a good estimate for FID [23, 49]. To mitigate the influence of distribution shift to the test datasets, which is the case for example in ImageNet, we report all metrics on the entire training set. To get a reference dataset for the adjusted IRS score, we take away 50000 samples for the reference dataset. If the datasets are smaller than 100000 samples we split the dataset in half and average the IRS results for as many runs as we have synthetic images." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Datasets: For our primary experiments, we use ImageNet" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3033" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.137, + 0.093, + 0.44, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.273, + 0.487, + 0.345 + ], + "angle": 0, + "content": "Figure 4. Visualization of the measurement gap across a diverse set of feature extractors by computing the unadjusted diversity of real data. The theoretical limit would be sampling \\( \\mathrm{N}_{train} \\) images where all of them correspond to a different image in the training dataset. The idealized scenario follows Eq. (1)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.369, + 0.483, + 0.505 + ], + "angle": 0, + "content": "512 [10]. To explore diversity within a more densely populated image space, we also include FFHQ [23], a benchmark dataset focused on facial images, and Chest-Xray14, which comprises medical chest X-ray images [52]. Additionally, we evaluate CelebV-HQ, a dataset containing face video frames [59], and Dynamic, a medical video dataset [35], treating the data as individual frames as done in [40]. The spatial resolution is fixed to 512 pixels and videos are additionally limited to the first 60 frames." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.506, + 0.483, + 0.627 + ], + "angle": 0, + "content": "Encoders: To find the best feature extractor \\(\\mathcal{F}\\) we compute the features using a large set of publicly available feature extractors following the selection of [49]. This includes BYOL [18], CLIP [38], a ConvNeXT based architecture [29], data2vec [3], DINOv2 [34], Inception [21, 45, 50], MAE [19], SwAV [7]. Additionally we add \"Random\" which uses a randomly initialized Inception-v3 [50] which has been shown to already extract meaningful features [33]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.637, + 0.273, + 0.654 + ], + "angle": 0, + "content": "4.1. Measurement Gap" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.483, + 0.901 + ], + "angle": 0, + "content": "First we need to decide for a feature extractor \\(\\mathcal{F}\\) that decides whether two samples correspond to each other. We choose to limit ourselves to large, pre-trained and publicly available feature extractors due to their universal applicability and established use. Custom feature extractors are also possible and could be used to increase interpretability even further. In Appx. 13 for example we compute IRS using a re-identification model trained to detect memorized samples unveiling diversity issues. To remain generalizable our goal is to find the most meaningful feature space in terms of image retrieval, that can be generally applied to different datasets. We experiment with a vast set of feature extractors \\(\\mathcal{F}\\), measurements \\(\\mathcal{P}\\), and datasets. Importantly we limit the experiments to evaluation on real data. In Sec. 3.1 we made the assumption, that all samples of the training dataset should be the main component of at" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.424 + ], + "angle": 0, + "content": "least one synthetic image. In the context of only comparing real images, this means that every train image should have at least one test image it is more similar to than all other train images. However, as we show in Fig. 13 and Tab. 6 of the supplementary material, all feature extractors are drastically worse than the expected value. This is a clear indicator for the insufficient representative nature of feature extractors currently used to compute all metrics in the realm of image generation. We refer to this observation as measurement gap. In the supplements we present examples how this measurement gap can be reduced largely if tailord towards certain datasets (e.g., Tab. 5). However, to remain dataset agnostic we propose an adjustment step that uses the measured IRS on real data as reference value to make up for this measurement gap. Visually this gap results in a lower apparent diversity as shown in Fig. 4. There are two key takeaways from this experiment. First, not even real data is diverse according to commonly used feature extractors which severely reduces the interpretability of currently used diversity metrics. Secondly, it could be worthwhile for special tasks, to train feature extractors specifically to assess diversity as done in Tab. 5." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.429, + 0.905, + 0.806 + ], + "angle": 0, + "content": "However, thanks to the adjustment step described in Sec. 3.4, for our metric the absolute value of diversity of real data does not matter. For synthetic data it just matters how diverse the synthetic data is in comparison to the real data. This adjustment leads to direct interpretability of the metric as diversity of images. In Fig. 5 we visualize this interpretability of IRS by balancing and splitting ImageNet into train, reference and test datasets. We use the labels to manually reduce the diversity in the test dataset by successively removing classes. We use 500 images of each class for train and 100 image of each class for reference and test data. We can see a clear linear correlation between the number of classes and our proposed IRS score. Additionally, it is the only diversity metric that achieves \\(100\\%\\) diversity when all classes are present in the test dataset. The slight overestimation of IRS in the mid-diversity-range is likely to be caused by label ambiguity. Coverage also correlates well but severely overestimates diversity when the number of classes is very low. It saturates at 0.97 and completely fails for different hyperparameter settings as we show in Sec. 15.2. Recall, which is the most commonly used metric to assess synthetic diversity, saturates at about 0.8. From our point of view, this is one of the key reasons why diversity has not yet established itself as the key metric that should be optimized." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.903 + ], + "angle": 0, + "content": "While IRS works with any kind of feature extractor, we choose to use the one that has the best visual interpretability when it comes to image retrieval to make the results even more meaningful. Hence, we choose \\(\\mathcal{F}\\) based on the meaningfulness of the image retrieval. To avoid the necessity of a ground truth image retrieval dataset from the training" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "3034" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.097, + 0.089, + 0.902, + 0.188 + ], + "angle": 0, + "content": "
DatasetNConsensusBYOLCLIPConvNeXtdata2vecDINOv2InceptionMAERandomSwAV
ImageNet35142 (3.43%)21.7289.5695.512.6592.7074.2087.1712.3489.39
FFHQ1862 (3.33%)14.1893.0297.852.4292.1675.3586.7921.4398.60
ChestX-ray141389 (1.55%)40.3245.9377.751.9483.1575.9589.7033.7796.11
CelebV-HQ5329 (19.47%)31.8389.4093.716.6891.3379.9494.8651.7996.70
Dynamic23 (0.29%)47.8334.7895.658.7078.2665.2295.6565.22100.00
" + }, + { + "type": "table_caption", + "bbox": [ + 0.153, + 0.199, + 0.843, + 0.214 + ], + "angle": 0, + "content": "Table 1. Measures the agreement of each \\(\\mathcal{P}\\) with the ensemble if five or more agree on the same retrieved image \\((\\%)\\)." + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.244, + 0.475, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.449, + 0.486, + 0.575 + ], + "angle": 0, + "content": "Figure 5. Measuring diversity of datasets by removing classes and computing IRS. If only 800 out of 1000 classes are part of the test set say that the diversity is at \\(80\\%\\). By manually removing ImageNet classes we can assess how good commonly used metrics are at measuring diversity compared to ours. To improve visual interpretability we normalize FID and Vendi to be within 0 and 1 with 1 being best. IRS correlates best with the real diversity measured as fraction between number of classes in test and train dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.605, + 0.484, + 0.772 + ], + "angle": 0, + "content": "dataset, we ensemble all feature extractors to compute a consensus on the image retrieval of real images. Details and qualitative examples can be found in Sec. 9. Consensus is reached if five or more models agree on the same image pair. For every image pair where consensus is reached we compute how often a single model agrees with the ensemble and report the results in Tab. 1. As expected the percentage of cases where consensus is reached is higher for natural images. Overall we see that SwAV generally agrees the most with the ensemble. Unless mentioned otherwise, this is the feature extractor we use to compute IRS scores." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.786, + 0.484, + 0.817 + ], + "angle": 0, + "content": "4.2. Diversity Issue in State-of-the-art Diffusion models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Recently, [40] published a synthetic ultrasound video dataset based on the EchoNet dataset [35]. It is accompanied with a feature extraction model used to privatize the dataset using re-identification. Analyzing these features, [14] were the first to notice a diversity issue. Us" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.24, + 0.908, + 0.832 + ], + "angle": 0, + "content": "ing our proposed IRS score we are now able to quantify it. First we try retrieve training samples using real frames from the test dataset. The dataset has 7465 training videos and we can see that 1277 test frames are able to retrieve 1159 distinct train videos. Following Eqs. (6) to (8) this means real data has a diversity of \\(\\mathrm{IRS}_{\\infty,real} = 86\\%\\) [75%, 100%]. Contrary to that, 1159 different synthetic frames are only able to retrieve 692 out of 7465 samples. This is equivalent to an \\(\\mathrm{IRS}_{\\infty,snth} = 12.3\\%\\) [11.9%, 12.9%]. Following the adjustment step introduced in Sec. 3.4 we conclude that the model has only learned \\(\\mathrm{IRS}_{\\infty,a} = 14.3\\%\\) of the data. Next, we compare multiple state-of-the-art approaches on conditional and unconditional image generation for ImageNet. We show their diversity according to IRS accompanied with other common metrics in Sec. 4.2. We initially expected that models for lower resolution outperform those with higher resolution. But overall this does not seem to be the case. Our experiments on MAR and EDM nicely confirm the feasibility of IRS as metrics as the score increases with the model size. Other metrics like Recall and Coverage confirm this observation but the difference seems marginal. Using IRS we can quantize and interpret the size of the gap between the smallest and largest models. The smallest model only reached a diversity of \\(46\\%\\) whereas the largest model reached \\(75\\%\\). Unlike the relatively small gap of 0.9 points in FID, this indicates a large gap in sampling diversity. Crucially, following our experiments from Fig. 5 our proposed IRS score can be interpreted. Even the best model only properly learned \\(77\\%\\) of the data, irrespective of the number of samples generated. Note that this is for conditional image generation, so generation specifically asks the model to generate an equal number of images from all classes. Generally, the conditional models performed better than the unconditional model which leads us to believe that the conditioning needs to be improved. Due to limited resources, we will continue our experiments by fixing the compute and only use the best performing method in terms of FID and IRS which is EDM. To keep the training time reasonable we restrict our experiments on the XS variant." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.846, + 0.871, + 0.864 + ], + "angle": 0, + "content": "4.3. Improving Diversity Using Pseudo Labels" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.908, + 0.903 + ], + "angle": 0, + "content": "Now that we have established EDM as the best method for unconditional image generation and at the same time have" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "3035" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.088, + 0.907, + 0.296 + ], + "angle": 0, + "content": "
ModelImage ResolutionFID ↓Prec. ↑Rec. ↑Dens. ↑Cov. ↑Vendi ↑IRS∞,a ↑ (Ours)
Pixel diffusion
ADM-256 [12]2566.01 (30.30)0.82 (0.57)0.62 (0.73)1.08 (0.41)0.91 (0.40)70.94 (36.18)0.44 (0.20)
Transformer
DiT-XL/2-256 [36]25622.15 (8.72)0.94 (0.69)0.34 (0.76)1.58 (0.70)0.85 (0.84)126.96 (58.15)0.23 (0.33)
DiT-XL/2-512 [36]51222.99 (9.54)0.96 (0.70)0.27 (0.73)1.90 (0.72)0.86 (0.82)128.98 (55.17)0.21 (0.34)
MAR-B-256 [27]2563.79 (10.36)0.83 (0.72)0.67 (0.71)1.18 (0.72)0.96 (0.75)83.03 (55.78)0.45 (0.38)
MAR-L-256 [27]2563.30 (10.36)0.82 (0.72)0.71 (0.71)1.10 (0.73)0.96 (0.75)81.80 (55.95)0.56 (0.38)
MAR-H-256 [27]2563.11 (10.36)0.82 (0.72)0.72 (0.71)1.07 (0.74)0.96 (0.76)81.37 (55.82)0.64 (0.38)
Latent diffusion, U-Net
LDM-256 [43]25626.09 (37.39)0.96 (0.61)0.21 (0.68)1.80 (0.45)0.83 (0.28)126.94 (30.83)0.16 (0.16)
EDM-2-XS-512 [24]5123.79 (75.02)0.83 (0.42)0.65 (0.63)1.22 (0.25)0.95 (0.13)72.41 (26.95)0.46 (0.09)
EDM-2-S-512 [24]5123.33 (122.48)0.85 (0.33)0.67 (0.42)1.26 (0.16)0.97 (0.07)80.25 (21.34)0.59 (0.04)
EDM-2-M-512 [24]5123.30 (107.45)0.85 (0.36)0.69 (0.61)1.24 (0.19)0.97 (0.09)82.99 (22.19)0.65 (0.06)
EDM-2-L-512 [24]5122.90 (118.87)0.84 (0.23)0.70 (0.51)1.22 (0.11)0.97 (0.06)82.10 (22.89)0.71 (0.03)
EDM-2-XL-512 [24]5122.92 (141.74)0.84 (0.25)0.71 (0.45)1.21 (0.12)0.97 (0.06)83.23 (20.04)0.77 (0.03)
EDM-2-XXL-512 [24]5122.87 (124.29)0.84 (0.33)0.71 (0.60)1.22 (0.17)0.97 (0.07)82.45 (21.52)0.75 (0.05)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.306, + 0.907, + 0.364 + ], + "angle": 0, + "content": "Table 2. Image generation of ImageNet across different generative models. Numbers in brackets indicate results when sampling the model without conditioning. Inception is used as feature extractor for all metrics except for IRS where we use SwAV. Precision, recall, density, coverage and Vendi score had to be computed on a subset of the training data of 50000 and 10000 for Vendi score. Pretrained conditional models were used as unconditional models by setting conditional guidance to 0." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.388, + 0.482, + 0.555 + ], + "angle": 0, + "content": "shown that this method still lacks diversity, we continue with EDM-2 as baseline for unconditional image generation and compare the trained models with a fixed compute budget. Quantitative results are shown in Tab. 3. The results show that the leveraging pseudo labels improves the sampling diversity compared to EDM-2 in all cases. In three cases the diversity is even better than that of the real reference dataset proofing that the pseudo conditioning works well and we can specifically query the model to generate diverse data. It also improves FID scores compared to the unconditional trainings." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.566, + 0.482, + 0.658 + ], + "angle": 0, + "content": "
FID ↓IRS∞,a ↑
EDM [24]DiADM (Ours)EDM [24]DiADM (Ours)
ImageNet-51251.5922.280.090.15
FFHQ40.926.240.231.51
ChestX-ray1424.296.760.191.08
CelebV-HQ68.4113.640.180.69
Dynamic13.825.560.601.04
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.667, + 0.483, + 0.697 + ], + "angle": 0, + "content": "Table 3. FID and IRS results with and without our proposed diversity awareness module" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.72, + 0.303, + 0.737 + ], + "angle": 0, + "content": "4.4. Text-to-image example" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.742, + 0.483, + 0.88 + ], + "angle": 0, + "content": "Extending IRS to text-conditional image generation is straightforward. The evaluation of diversity depends on the chosen perspective, which is defined by the reference datasets. By changing the reference dataset, we can address various diversity-related issues, such as fairness. One notable example is the gender bias observed in text-to-image models. For our experiments, we use Deepfloyd (IF-I-XLv1.0)\\(^2\\). To assess gender diversity, we create a reference dataset with a balanced gender distribution for specific job" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.388, + 0.907, + 0.51 + ], + "angle": 0, + "content": "roles by prompting the model to generate 100 images each of males and females. Half of these images are reserved as a balanced test dataset. We then prompt the model without specifying gender. The results are presented in Sec. 4.2. All groups, including the general term human, show gender bias, with the latter generating only female images. IRS estimates that in all cases, the generated diversity reaches only about \\(50\\%\\) of that of the balanced reference dataset." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.52, + 0.64, + 0.535 + ], + "angle": 0, + "content": "4.5. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.543, + 0.909, + 0.71 + ], + "angle": 0, + "content": "The stochasticity involved for a low number of samples means that minor IRS differences are not meaningful for a low number of samples. However, this is also reflected in the high values of uncertainty. Additionally, we believe that finding a better feature extractor that maximizes measured diversity (unadjusted IRS) on real datasets should be the focus. Experiments on improving diversity are limited to using real features for generation and restrictive computational budget. Furthermore, increasing the conditioning requires similar experiments analyzing memorization issues as text-conditioning [6]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.724, + 0.743, + 0.74 + ], + "angle": 0, + "content": "5. Conclusion and Outlook" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In this paper, we reveal that all current methods for evaluating diversity and the feature extractors they are based on are inadequate. Using our proposed IRS metric, we demonstrate that these shortcomings result in all state-of-the-art image generation methods facing challenges with diversity. To address this issue for unconditional diffusion models, we propose DiADMs, which enhance the performance of current methods by separating diversity from fidelity. In future work, we aim to extend our efforts to improve diversity in conditional image generation as well." + }, + { + "type": "page_footnote", + "bbox": [ + 0.108, + 0.887, + 0.372, + 0.902 + ], + "angle": 0, + "content": "\\(^{2}\\)https://huggingface.co/DeepFloyd/IF-I-XL-v1.0" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "3036" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.304 + ], + "angle": 0, + "content": "Acknowledgements: This work was supported by the High-Tech-Agenda Bavaria. HPC resources were provided by the Erlangen National High Performance Computing Center (NHR@FAU) of the Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU) under the NHR project b143dc and b180dc. NHR funding is provided by federal and Bavarian state authorities. NHR@FAU hardware is partially funded by the German Research Foundation (DFG) - 440719683. Support was also received by the ERC - projects MIA-NORMAL 101083647 as well as DFG 513220538, 512819079 and DFG large scale infrastructure funding Art 91b GG. This work was supported by the UKRI Centre for Doctoral Training in Artificial Intelligence for Healthcare (EP/S023283/1) and Ultromics Ltd." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.32, + 0.188, + 0.336 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.345, + 0.484, + 0.4 + ], + "angle": 0, + "content": "[1] Sumukh K Aithal, Pratyush Maini, Zachary C Lipton, and J Zico Kolter. Understanding hallucinations in diffusion models through mode interpolation. arXiv preprint arXiv:2406.09358, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.402, + 0.484, + 0.471 + ], + "angle": 0, + "content": "[2] Sina Alemohammad, Josue Casco-Rodriguez, Lorenzo Luzi, Ahmed Imtiaz Humayun, Hossein Babaei, Daniel LeJeune, Ali Siahkoohi, and Richard G. Baraniuk. Self-Consuming Generative Models Go MAD, 2023. arXiv:2307.01850 [cs]. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.474, + 0.484, + 0.557 + ], + "angle": 0, + "content": "[3] Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, and Michael Auli. data2vec: A General Framework for Self-supervised Learning in Speech, Vision and Language. In Proceedings of the 39th International Conference on Machine Learning, pages 1298-1312. PMLR, 2022. ISSN: 2640-3498. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.559, + 0.484, + 0.615 + ], + "angle": 0, + "content": "[4] Hritik Bansal, Da Yin, Masoud Monajatipoor, and Kai-Wei Chang. How well can text-to-image generative models understand ethical natural language interventions? arXiv preprint arXiv:2210.15230, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.617, + 0.485, + 0.672 + ], + "angle": 0, + "content": "[5] Leonard E. Baum and Patrick Billingsley. Asymptotic Distributions for the Coupon Collector's Problem. The Annals of Mathematical Statistics, 36(6):1835-1839, 1965. Publisher: Institute of Mathematical Statistics. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.675, + 0.484, + 0.731 + ], + "angle": 0, + "content": "[6] Nicholas Carlini, Jamie Hayes, Milad Nasr, Matthew Jagielski, Vikash Sehwag, Florian Tramér, Borja Balle, Daphne Ippolito, and Eric Wallace. Extracting Training Data from Diffusion Models, 2023. arXiv:2301.13188 [cs]. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.733, + 0.484, + 0.801 + ], + "angle": 0, + "content": "[7] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised Learning of Visual Features by Contrasting Cluster Assignments. In Advances in Neural Information Processing Systems, pages 9912-9924. Curran Associates, Inc., 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.804, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[8] Tingxiu Chen, Yilei Shi, Zixuan Zheng, Bingcong Yan, Jingliang Hu, Xiao Xiang Zhu, and Lichao Mou. Ultrasound Image-to-Video Synthesis via Latent Dynamic Diffusion Models. In Medical Image Computing and Computer Assisted Intervention - MICCAI 2024, pages 764-774. Springer Nature Switzerland, Cham, 2024. Series Title: Lecture Notes in Computer Science. 2" + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.345, + 0.485, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.093, + 0.907, + 0.147 + ], + "angle": 0, + "content": "[9] Ching-Yao Chuang, Varun Jampani, Yuanzhen Li, Antonio Torralba, and Stefanie Jegelka. Debiasing vision-language models via biased prompts. arXiv preprint arXiv:2302.00070, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.906, + 0.188 + ], + "angle": 0, + "content": "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.191, + 0.906, + 0.233 + ], + "angle": 0, + "content": "[11] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.905, + 0.274 + ], + "angle": 0, + "content": "[12] Prafulla Dhariwal and Alex Nichol. Diffusion Models Beat GANs on Image Synthesis, 2021. arXiv:2105.05233 [cs, stat]. 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.275, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[13] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. Cogview: Mastering text-to-image generation via transformers. Advances in neural information processing systems, 34:19822-19835, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.345, + 0.905, + 0.398 + ], + "angle": 0, + "content": "[14] Mischa Dombrowski, Hadrien Reynaud, and Bernhard Kainz. Uncovering Hidden Subspaces in Video Diffusion Models Using Re-Identification, 2024. arXiv:2411.04956. 2, 7, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.906, + 0.469 + ], + "angle": 0, + "content": "[15] Chun-Mei Feng, Kai Yu, Yong Liu, Salman Khan, and Wangmeng Zuo. Diverse Data Augmentation with Diffusions for Effective Test-time Prompt Tuning. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 2704-2714, Paris, France, 2023. IEEE. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.905, + 0.51 + ], + "angle": 0, + "content": "[16] Dan Friedman and Adji Bousso Dieng. The Vendi Score: A Diversity Evaluation Metric for Machine Learning, 2023. arXiv:2210.02410. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.511, + 0.906, + 0.567 + ], + "angle": 0, + "content": "[17] Felix Friedrich, Manuel Brack, Lukas Struppek, Dominik Hintersdorf, Patrick Schramowski, Sasha Luccioni, and Kristian Kersting. Fair Diffusion: Instructing Text-to-Image Generation Models on Fairness, 2023. arXiv:2302.10893. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.568, + 0.906, + 0.663 + ], + "angle": 0, + "content": "[18] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre H. Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, Bilal Piot, Koray Kavukcuoglu, Rémi Munos, and Michal Valko. Bootstrap your own latent: A new approach to self-supervised Learning, 2020. arXiv:2006.07733 [cs, stat]. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.665, + 0.906, + 0.733 + ], + "angle": 0, + "content": "[19] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked Autoencoders Are Scalable Vision Learners. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15979-15988, New Orleans, LA, USA, 2022. IEEE. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.734, + 0.906, + 0.801 + ], + "angle": 0, + "content": "[20] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.906, + 0.872 + ], + "angle": 0, + "content": "[21] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[22] Jonathan Ho and Tim Salimans. Classifier-Free Diffusion Guidance. 2022. 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3037" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[23] Tero Karras, Samuli Laine, and Timo Aila. A Style-Based Generator Architecture for Generative Adversarial Networks. 2019. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.192 + ], + "angle": 0, + "content": "[24] Tero Karras, Miika Aittala, Jaakko Lehtinen, Janne Hellsten, Timo Aila, and Samuli Laine. Analyzing and Improving the Training Dynamics of Diffusion Models, 2024. arXiv:2312.02696 [cs, stat]. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.263 + ], + "angle": 0, + "content": "[25] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.482, + 0.305 + ], + "angle": 0, + "content": "[26] Tuomas Kynkänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved Precision and Recall Metric for Assessing Generative Models. 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.307, + 0.482, + 0.348 + ], + "angle": 0, + "content": "[27] Tianhong Li, Yonglong Tian, He Li, Mingyang Deng, and Kaiming He. Autoregressive Image Generation without Vector Quantization, 2024. arXiv:2406.11838 [cs]. 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.482, + 0.419 + ], + "angle": 0, + "content": "[28] Jiawei Liu, Qiang Wang, Huijie Fan, Yinong Wang, Yandong Tang, and Liangqiong Qu. Residual denoising diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2773-2783, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.482, + 0.462 + ], + "angle": 0, + "content": "[29] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A ConvNet for the 2020s, 2022. arXiv:2201.03545. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.463, + 0.482, + 0.533 + ], + "angle": 0, + "content": "[30] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.534, + 0.482, + 0.617 + ], + "angle": 0, + "content": "[31] Zichen Miao, Jiang Wang, Ze Wang, Zhengyuan Yang, Lijuan Wang, Qiang Qiu, and Zicheng Liu. Training diffusion models towards diverse image generation with reinforcement learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10844-10853, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.619, + 0.482, + 0.66 + ], + "angle": 0, + "content": "[32] Muhammad Ferjad Naeem, Seong Joon Oh, Youngjung Uh, Yunjey Choi, and Jaejun Yoo. Reliable Fidelity and Diversity Metrics for Generative Models. 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.662, + 0.482, + 0.732 + ], + "angle": 0, + "content": "[33] Jamie A. OaReilly and Fawad Asadi. Pre-trained vs. Random Weights for Calculating Fréchet Inception Distance in Medical Imaging. In 2021 13th Biomedical Engineering International Conference (BMEiCON), pages 1-4, Ayutthaya, Thailand, 2021. IEEE. 6, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.733, + 0.482, + 0.858 + ], + "angle": 0, + "content": "[34] Maxime Oquab, Timothee Darct, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINoV2: Learning Robust Visual Features without Supervision, 2024. arXiv:2304.07193 [cs]. 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[35] David Ouyang, Bryan He, Amirata Ghorbani, Neal Yuan, Joseph Ebinger, Curtis P. Langlotz, Paul A. Heidenreich, Robert A. Harrington, David H. Liang, Euan A. Ashley, and" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "James Y. Zou. Video-based AI for beat-to-beat assessment of cardiac function. Nature, 580(7802):252-256, 2020. Publisher: Springer US. 6, 7, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[36] William Peebles and Saining Xie. Scalable Diffusion Models with Transformers. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 4172-4182, Paris, France, 2023. IEEE. 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.19, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[37] Yiming Qin, Huangjie Zheng, Jiangchao Yao, Mingyuan Zhou, and Ya Zhang. Class-balancing diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18434-18443, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.246, + 0.905, + 0.328 + ], + "angle": 0, + "content": "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning Transferable Visual Models From Natural Language Supervision, 2021. arXiv:2103.00020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.329, + 0.905, + 0.412 + ], + "angle": 0, + "content": "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.413, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[40] Hadrien Reynaud, Qingjie Meng, Mischa Dombrowski, Arijit Ghosh, Thomas Day, Alberto Gomez, Paul Leeson, and Bernhard Kainz. EchoNet-Synthetic: Privacy-preserving Video Generation for Safe Medical Data Sharing, 2024. arXiv:2406.00808 [cs]. 6, 7, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.483, + 0.905, + 0.525 + ], + "angle": 0, + "content": "[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-Resolution Image Synthesis with Latent Diffusion Models. 2021. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.905, + 0.594 + ], + "angle": 0, + "content": "[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.595, + 0.905, + 0.65 + ], + "angle": 0, + "content": "[43] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-Resolution Image Synthesis with Latent Diffusion Models, 2022. arXiv:2112.10752 [cs]. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.65, + 0.905, + 0.704 + ], + "angle": 0, + "content": "[44] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. DreamBooth: Fine Tuning Text-to-Image Diffusion Models for Subject-Driven Generation. 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.705, + 0.905, + 0.748 + ], + "angle": 0, + "content": "[45] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, Xi Chen, and Xi Chen. Improved Techniques for Training GANs. 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.748, + 0.905, + 0.79 + ], + "angle": 0, + "content": "[46] Axel Sauer, Katja Schwarz, and Andreas Geiger. StyleGANXL: Scaling StyleGAN to Large Diverse Datasets, 2022. arXiv:2202.00273 [cs]. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.859 + ], + "angle": 0, + "content": "[47] Vikash Sehwag, Caner Hazirbas, Albert Gordo, First Ozgenel, and Cristian Canton. Generating high fidelity data from low-density regions using diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11492-11501, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[48] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3038" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.176 + ], + "angle": 0, + "content": "[49] George Stein, Jesse C. Cresswell, Rasa Hosseinzadeh, Yi Sui, Brendan Leigh Ross, Valentin Villecloze, Zhaoyan Liu, Anthony L. Caterini, J. Eric T. Taylor, and Gabriel Loaiza-Ganem. Exposing flaws of generative model evaluation metrics and their unfair treatment of diffusion models, 2023. arXiv:2306.04675 [cs, stat]. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.178, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[50] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the Inception Architecture for Computer Vision. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2818-2826, Las Vegas, NV, USA, 2016. IEEE. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.275 + ], + "angle": 0, + "content": "[51] N. M. Temme. Asymptotic Estimates of Stirling Numbers. Studies in Applied Mathematics, 89(3):233-243, 1993. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.482, + 0.344 + ], + "angle": 0, + "content": "[52] Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, Mohammadhadi Bagheri, and Ronald M Summers. ChestX-ray8: Hospital-Scale Chest X-Ray Database and Benchmarks on Weakly-Supervised Classification and Localization of Common Thorax Diseases. 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.347, + 0.482, + 0.415 + ], + "angle": 0, + "content": "[53] Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang. Diffusion models: A comprehensive survey of methods and applications. ACM Computing Surveys, 56(4): 1-39, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.418, + 0.482, + 0.485 + ], + "angle": 0, + "content": "[54] Ling Yang, Jingwei Liu, Shenda Hong, Zhilong Zhang, Zhilin Huang, Zheming Cai, Wentao Zhang, and Bin Cui. Improving diffusion-based image synthesis with context prediction. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.488, + 0.482, + 0.557 + ], + "angle": 0, + "content": "[55] Cheng Zhang, Xuanbai Chen, Siqi Chai, Chen Henry Wu, Dmitry Lagun, Thabo Beeler, and Fernando De la Torre. Itigen: Inclusive text-to-image generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3969-3980, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.559, + 0.482, + 0.613 + ], + "angle": 0, + "content": "[56] Weitong Zhang, Chengqi Zang, Liu Li, Sarah Cechnicka, Cheng Ouyang, and Bernhard Kainz. Stability and generalizability in sde diffusion models with measure-preserving dynamics. In NeurIPS, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.616, + 0.482, + 0.669 + ], + "angle": 0, + "content": "[57] Yi Zhang, Xiaoyu Shi, Dasong Li, Xiaogang Wang, Jian Wang, and Hongsheng Li. A unified conditional framework for diffusion-based image restoration. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.672, + 0.482, + 0.712 + ], + "angle": 0, + "content": "[58] Dora Zhao, Jerone T. A. Andrews, Orestis Papakyriakopoulos, and Alice Xiang. Position: Measure Dataset Diversity, Don't Just Claim It, 2024. arXiv:2407.08188. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.715, + 0.482, + 0.769 + ], + "angle": 0, + "content": "[59] Hao Zhu, Wayne Wu, Wentao Zhu, Liming Jiang, Siwei Tang, Li Zhang, Ziwei Liu, and Chen Change Loy. CelebV-HQ: A Large-Scale Video Facial Attributes Dataset, 2022. arXiv:2207.12393 [cs]. 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.769 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3039" + } + ] +] \ No newline at end of file diff --git a/2025/Image Generation Diversity Issues and How to Tame Them/7493d326-112b-4a6d-8c2c-7e4449ad1384_origin.pdf b/2025/Image Generation Diversity Issues and How to Tame Them/7493d326-112b-4a6d-8c2c-7e4449ad1384_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7c4ae357f5d8bf9f00f5874c91b9c85de820f720 --- /dev/null +++ b/2025/Image Generation Diversity Issues and How to Tame Them/7493d326-112b-4a6d-8c2c-7e4449ad1384_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1733f572944b232ad5c93dd964eda8cef5b3a568f4d3629d33764ae0549c62d7 +size 3562971 diff --git a/2025/Image Generation Diversity Issues and How to Tame Them/full.md b/2025/Image Generation Diversity Issues and How to Tame Them/full.md new file mode 100644 index 0000000000000000000000000000000000000000..bfb53bf34ad10608c9d416b303ee0979a62da1c1 --- /dev/null +++ b/2025/Image Generation Diversity Issues and How to Tame Them/full.md @@ -0,0 +1,278 @@ +# Image Generation Diversity Issues and How to Tame Them + +Mischa Dombrowski1 Weitong Zhang2 Sarah Cechnicka2 + +Hadrien Reynaud² Bernhard Kainz1,2 + +$^{1}$ Friedrich-Alexander-Universität Erlangen-Nürnberg $^{2}$ Imperial College London + +mischa.dombrowski@fau.de + +# Abstract + +Generative methods have reached a level of quality that is almost indistinguishable from real data. However, while individual samples may appear unique, generative models often exhibit limitations in covering the full data distribution. Unlike quality issues, diversity problems within generative models are not easily detected by simply observing single images or generated datasets, which means we need a specific measure to assess the diversity of these models. In this paper, we draw attention to the current lack of diversity in generative models and the inability of common metrics to measure this. We achieve this by framing diversity as an image retrieval problem, where we measure how many real images can be retrieved using synthetic data as queries. This yields the Image Retrieval Score (IRS), an interpretable, hyperparameter-free metric that quantifies the diversity of a generative model's output. IRS requires only a subset of synthetic samples and provides a statistical measure of confidence. Our experiments indicate that current feature extractors commonly used in generative model assessment are inadequate for evaluating diversity effectively. Consequently, we perform an extensive search for the best feature extractors to assess diversity. Evaluation reveals that current diffusion models converge to limited subsets of the real distribution, with no current state-of-the-art models superpassing $77\%$ of the diversity of the training data. To address this limitation, we introduce Diversity-Aware Diffusion Models (DiADM), a novel approach that improves diversity of unconditional diffusion models without loss of image quality. We do this by disentangling diversity from image quality by using a diversity aware module that uses pseudo-unconditional features as input. We provide a Python package offering unified feature extraction and metric computation to further facilitate the evaluation of generative models https://github.com/MischaD/beyondfid. + +![](images/4167b51f1a8905e82a2b7e30a60ce5a15048409adb695873d68a5fa0959e4183.jpg) +Low Diversity + +![](images/29ef5003bf151ccf25c9045c6134b6c99418f4e1c76d718d1ed9f44ec54fac28.jpg) +High Diversity + +![](images/c8e1ad8d8111c37e0a5d63bfeaea045fec25ae6b303cf5504ad267bf538b045e.jpg) +Predicted Diversity $\mathrm{IRS}_{\infty}$ $(\%)$ +Figure 1. Predicted gender diversity - as one possible example for diversity - after sampling pre-trained text-to-image diffusion models infinitely often. Visually it is apparent that most terms are inherently biased toward one gender. Our proposed IRS score predicts that off-the-shelf models will only ever reach 50 percent diversity, which is equivalent to one gender being perfectly represented while the other gender is missing entirely. Using diversification strategies increases diversity. Details about this can be found in Sec. 4.4. + +# 1. Introduction + +With advancements in large-scale models and datasets, diffusion models have demonstrated substantial capacity to capture high-dimensional, complex distributions, enabling photo-realistic conditional image generation across diverse guidance types, including image-based [56], classifier-based [11], and text-based guidance [42]. However, diffusion models frequently inherit and amplify data biases due to their reliance on large-scale, naturally biased datasets [4, 53], limiting comprehensive distributional coverage in generated samples. For instance, text-to-image diffusion models often exhibit gender bias for certain occupations. When prompted with "a teacher", models predominantly generate females, with male representations notably underrepresented, as shown in Fig. 1. The "conditional mode collapse" [1, 31] also affects image- [56] and class-conditional [11] diffusion models, where the guidance often produces overly uniform outputs, leading to a loss in diver + +sity in the sampled distribution. Dataset diversity remains an unresolved issue, even for real datasets [58], due to ambiguous definitions and concepts. In contrast, synthetic data can leverage its training data as a reference. To avoid problems like mode drift [2], synthetic data should be at least as diverse as its training set. Key motivations for ensuring diversity include content creation [44], addressing privacy and memorization concerns [6], mitigating the lack of diversity in synthetic datasets [14], data augmentation [15], promoting algorithmic fairness [17], and enhancing downstream performance [8]. To improve diversity in conditional image generation, [28] introduces noise perturbations to enhance diversity, though at the expense of image quality and certainty. [30] balances a trade-off between faithfulness and realism in diverse image synthesis, where diversity comes at the cost of fidelity. These diversity-oriented approaches highlight an inherent compromise: diversity is achieved by sacrificing generation quality, with randomness that lacks interpretability. In parallel, classifier-based diversity guidance [47] and text-based methods [13, 55] have been developed to reduce bias in text-to-image generation. However, these techniques are task-specific and lack generalizability across different types of diffusion models for diversity enhancement. Thus, a fundamental challenge emerges: a universal, interpretable metric to evaluate diversity without bias or conditional dependence, and without compromising generation quality. + +To address this, we propose the Image Retrieval Score (IRS), an interpretable metric for quantifying diversity in generative models. IRS assesses how effectively synthetic data retrieves real images as queries, providing a hyperparameter-free, statistically grounded evaluation based on minimal samples. Our approach is motivated by a thought experiment that illustrates how individual samples reveal model diversity. Using IRS as a core, generic metric for diversity, we quantify and reveal two key findings: (1) existing feature extractors used to evaluate generative models lack the capacity to measure diversity effectively, and (2) current diffusion models converge on limited subsets of the real distribution, with no state-of-the-art models surpassing $77\%$ of the diversity in training data. + +This further motivates the use of IRS as an unbiased guidance to drive a comprehensive search for optimal feature extractors, enhancing diversity evaluation without compromising generation quality. Our selection is based on [49] who experimented on the impact of feature extractors on existing generative metrics. Building on these insights, we introduce Diversity-Aware Diffusion Models (DiADM), an approach that improves diversity across the full data distri + +![](images/7dd0ecfe12433cd49daa81d1380b264c8580ecc2a72f7ca7c82077827aa0842d.jpg) +Figure 2. We model finding image pairs (image retrieval) as randomly drawing from urns with replacement (Ideal). We observe that all used feature extraction models exhibit performance issues by collapsing in the feature space for real images resulting in a measurement gap affecting currently used metrics such as $\mathrm{FID}_{\mathrm{Inception}}$ , $\mathrm{FID}_{\mathrm{DINOv2}}$ , Precision and Recall. Comparing the results of synthetic images to real data shows that datasets generated by generative models show even stronger diversity issues which result in a synthetic data distribution gap. Our proposed metric leverages a real reference dataset to remove this measurement gap. + +bution in unconditional diffusion models without compromising image quality. DiADM accomplishes this by disentangling diversity from image quality through a diversity-aware module utilizing pseudo-unconditional features. Our contributions can be summarized as follows: + +- We demonstrate the limitations of widely-used feature extractors for diversity metrics, highlighting their poor performance on real datasets. +- We propose an intuitive framework for assessing dataset diversity in unconditional diffusion models. +- We identify and quantify the reduction in diversity resulting from current training methodologies. +- We provide a comprehensive analysis showing limited image diversity across a wide range of state-of-the-art diffusion models. +- We introduce a diversity-aware plug-in module, which utilizes pseudo-labels to separate diversity from fidelity, thereby enhancing generative diversity in unconditional diffusion models. + +# 2. Related Work + +Image Generation: Diffusion models [48] progressively map complex distributions to a standard Gaussian. Conditional diffusion models have demonstrated a strong ability to generate highly realistic images that align closely with specified conditions, excelling in tasks such as image synthesis [54], image restoration [57], and image reconstruction [56]. Additionally, [11] enhances diffusion + +models with an auxiliary classifier, enabling high-fidelity, class-conditional generation, while text-conditional models generate images based on text prompts [13, 55]. However, diffusion models often struggle to achieve sufficient diversity in task-agnostic generation, especially when trained on large-scale datasets that carry inherent biases [4, 53]. For example, Stable Diffusion, based on latent diffusion models (LDM) [39, 42], produces high-quality images aligned with text prompts, yet frequently reflects bias (e.g., age, gender) for certain terms. These limitations lead to restricted distributional coverage and biased outputs. + +Diversity in Generation: Research on diversity in diffusion models is limited, with most existing approaches focused on GAN-based models [9, 25]. Current methods for diversity in diffusion models fall into the primary categories: bias mitigation and diversity enhancement. In bias mitigation, [37] uses a re-weighted loss to address long-tailed distributions, promoting balanced class representation. For text-conditional models, [13] attempts to mitigate bias by adding attribute words to prompts, though the added prompts may be ambiguous. [31] extends this to both class- and text-conditional models, using reinforcement learning to match the diversity ratio in reference images but relying on task-specific diversity reward functions. For diversity enhancement, [28] introduces noise perturbations, which increase diversity at the expense of image quality and certainty. Similarly, [30] balances diversity with realism, where gains in diversity tend to compromise fidelity. + +Metrics of Diversity: Existing metrics in diffusion models remain limited and are unsuitable for accurate and interpretable diversity assessment. Traditional image quality metrics like FID [20] require extensive sampling (e.g., 50k images) to be reliable, making them computationally intensive and challenging to interpret in practical applications. Precision and recall [26] are most established in literature [12, 36, 41, 46]. They add supplementary insights but are hindered by dependency on dataset size, hyperparameters, and interpretability issues. Credibility and interpretability of metrics are highly underestimated properties. For example, a comparison of three years of advancements in diffusion models, as seen in [12] and [27], shows an order-of-magnitude improvement in FID but a one percentage point reduction in precision. Coverage and density [32] also measure diversity but are dependant on the ratio between train and test set, highly depend on hyperparameters, lack evaluation on real datasets, and density was designed to ignore real outliers which are important in the context of diversity assessment. Vendi Score [16], which operates entirely without a real reference set, sacrifices fidelity and is computationally prohibitive for large datasets. + +# 3. Method + +Let $\mathbf{x_t}$ be image $t$ from a dataset consisting of $\mathrm{N}_{train}$ real images residing in image space $\mathcal{X} \in \mathbb{R}^{x\times h\times w}$ . Unconditional generative models aim to learn the distribution $p_{data}(\mathbf{X})$ and sample $\mathrm{N}_{\text{sample}}$ synthetic images from it. + +# 3.1. Coupon Collector Problem + +The performance of generative models is generally measured by assessing the similarity of the real and the synthetic datasets. It is assumed that the training samples represent discrete observations from a continuous unknown distribution. We suggest viewing all synthetic images as a composition of real images, which holds true as long as there is no training objective that enforces out-of-distribution generation. To assess diversity, we argue that every training sample should be the main component of at least one synthetic sample. To measure this we use the fact that the observed training distribution is discrete. A model that has perfectly memorized every training sample would generate all images with equal probability $\frac{1}{\mathrm{N}_{train}}$ . Consequently, we can model sampling synthetic images with an urn experiment. Each synthetic sample consists mainly of one component which is then compounded with other training images. A real sample is considered learned if it serves as the main component for at least one synthetic image. Evaluating diversity in this context can be framed as the coupon collector problem [5]. Let $P_{n}(N_{t} > 0)$ denote the probability of generating $n$ samples where at least one corresponds to a training image $t$ , with $N_{t}$ denoting the number of generated images a training image is associated with. This probability can be computed by considering the compound probability of not sampling this training image which is $\frac{\mathrm{N}_{train} - 1}{\mathrm{N}_{train}}$ . Since the same probability applies to each image the expected value of the diversity can be calculated as + +$$ +\mathbb {E} [ \text {D i v e r s i t y} ] = \left(1 - \left(\frac {\mathrm {N} _ {\text {t r a i n}} - 1}{\mathrm {N} _ {\text {t r a i n}}}\right) ^ {n}\right). \tag {1} +$$ + +In this definition, diversity is equal to the percentage of learned real images. This idealized scenario is shown as ideal in Fig. 2. Therefore, assessing diversity comes down to predicting which training image corresponds to which synthetic image. + +# 3.2. Image Retrieval + +To assess diversity we first need to define how to decide on what constitutes a main component for each synthetic image. For each synthetic image instance $\mathbf{x}_{\mathrm{t}}^{\prime}$ we aim to identify the real image $\mathbf{x}_{\mathrm{t}}$ that is closest to it according to a measurement $\mathcal{P}$ (e.g., Euclidean distance) on the feature of the images extracted by a pretrained feature extractor $\mathcal{F}$ . Then, to assess the diversity we have to compute the size of the dataset of learned images: + +$$ +\mathcal {X} _ {\text {l e a r n e d}} = \left\{\mathbf {x} _ {t} \in \mathcal {X} \mid \exists \mathbf {x} _ {t} ^ {\prime} \in \mathcal {X} ^ {\prime}: \mathbf {x} _ {t} = \arg \min _ {\mathbf {x} _ {t}} \mathcal {P} \left(\mathbf {x} _ {t}, \mathbf {x} _ {t} ^ {\prime}\right) \right\}. \tag {2} +$$ + +We define the cardinality of this set as $\mathbf{N}_{\text{learned}}$ . + +We then call the trained image instance $\mathbf{x}_{\mathrm{t}}$ "learned" and compute the diversity of the synthetic dataset as: + +$$ +\mathrm {I R S} _ {\alpha} = \frac {\mathrm {N} _ {\text {l e a r n e d}}}{\mathrm {N} _ {\text {t r a i n}}} \in [ 0, 1 ] \tag {3} +$$ + +We term this quantity as the image retrieval score (IRS). To emphasize that $\mathrm{IRS}_{\alpha}$ depends on the number of images sampled we add $\alpha := \frac{\mathrm{N}_{\text{sample}}}{\mathrm{N}_{\text{train}}}$ as a subscript. Note that all other generative metrics also depend on $\alpha$ but often ignore this dependency. The name image retrieval score derives from the fact that computing IRS also works when comparing two disjoint real datasets. In this context the problem can be thought of as an image retrieval task, where a real image serves as the query image and the goal is to find the most similar image in a comparison dataset. Intuitively, IRS measures the ratio of images in a training dataset that are retrievable by another dataset. If the comparison dataset consists of real images we say that an image is "retrievable". If the comparison dataset consists of synthetic images we say that an image is "learned". + +The key advantage of IRS is that we can give a reasonably good estimate of dataset diversity with only a minimal number of samples. For example, if the same image appears three times within just ten samples, it is a strong indicator of overfitting. Consequently, we aim to estimate the diversity after $n << \mathrm{N}_{train}$ samples. To compute the probability of retrieving exactly $k = \mathcal{X}_{\text{learned}}$ images we use standard combinatorics. This problem can be split into two parts. Firstly, we compute the number of possibilities of splitting $n$ into $k$ different subsets. Instinctively, we are computing the number of combinations in which different synthetic samples derive from the same retrieved (still unassigned) image $\mathbf{x}_t$ according to equation Eq. (2). The computation involves using Stirling's number of the second kind, $\operatorname{Stir}(n, k)$ , which should be multiplied by the total number of possible assignments for each subset, given by $\frac{\mathrm{N}_{train}!}{(\mathrm{N}_{train} - \mathrm{N}_{\text{learned}})!}$ . This will become more clear with an example. Say, the first subset consists of three different images $x_3', x_{10}', x_{11}'$ , all of which are learned from the same real image. Then for the first of $k$ subsets, there are $\mathrm{N}_{train}$ different choices. For the second subset, we have $\mathrm{N}_{train} - 1$ choices and so on. Dividing by all choices leaves us with the probability for generating $k = \mathrm{N}_{\text{learned}}$ different images: + +$$ +\mathrm {P} (k, n, s) = \frac {\operatorname {S t i r} (n , k) * s !}{(s - k) ! * (s) ^ {n}}, \tag {4} +$$ + +with $k = \mathrm{N}_{\text{learned}}$ and $s = \mathrm{N}_{\text{train}}$ and $n = \mathrm{N}_{\text{sample}}$ . + +For large values, we estimate this by computing the log of Eq. (4) instead. [51] showed that + +$$ +\operatorname {S t i r} (n, k) = \sqrt {\frac {v - 1}{v (1 - G)}} \left(\frac {v - 1}{v - G}\right) ^ {n - k} \frac {k ^ {n}}{n ^ {k}} e ^ {k (1 - G)} \binom {n} {k}, \tag {5} +$$ + +where $v = \frac{n}{k}$ and $G \in (0,1)$ is the unique solution to $G = v e^{G - v}$ . Asymptotically, this estimate is close, reaching a maximum relative error of roughly $\frac{0.066}{n}$ if $n$ and $k$ are small, e.g., (10, 3) [51]. Computing this in log-space for ImageNet roughly takes 3 seconds and using memorization techniques reduces consecutive computations to 1.3 seconds. + +# 3.3. Image Retrieval Score + +Given Eq. (5) we can now estimate IRS as well as a confidence interval for it. We assume that $\mathrm{N}_{\text {learned }}$ and $\mathrm{N}_{\text {sample }}$ are fixed and we aim to estimate the model diversity. To compute the maximum likelihood estimate for IRS, we optimize over values of $s \in \mathbb{N} \cap [\mathrm{N}_{\text {learned }}, \mathrm{N}_{\text {train }}}$ : + +$$ +\mathrm {I R S} _ {\infty} = \frac {1}{\mathrm {N} _ {t r a i n}} \arg \max _ {s} \mathrm {P} (\mathrm {N} _ {l e a r n e d}, \mathrm {N} _ {s a m p l e}, s), (6) +$$ + +which means we expect $\mathrm{N}_{\mathrm{learned},\infty} = \mathrm{N}_{\mathrm{train}}*\mathrm{IRS}_{\infty}$ different images to be sampled eventually. To compute the lower bound we want to minimize the probability of the real diversity being lower but sampling abnormally many different images early. Mathematically, we estimate this as: + +$$ +\begin{array}{l} \mathrm {I R S} _ {\infty , L} = \frac {1}{N _ {\text {t r a i n}}} \arg \max _ {s \in \mathbb {N} \cap [ N _ {\text {l e a n e d}}, \infty , N _ {\text {t r a i n}} ]} \\ \min \left(N _ {\text {t r a i n}}, N _ {\text {s a m p l e}}\right) \tag {7} \\ \sum_ {k = N _ {\text {l e a n e d}}} \quad \mathrm {P} (k, N _ {\text {s a m p l e}}, s) > \alpha_ {e} \\ \end{array} +$$ + +Accordingly, we can compute the estimate for the upper boundary: + +$$ +\begin{array}{l} \mathrm {I R S} _ {\infty , U} = \frac {1}{N _ {\text {t r a i n}}} \arg \min _ {s \in \mathbb {N} \cap \left[ N _ {\text {l e a n e d}}, N _ {\text {l e a n e d}, \infty} \right]} \\ \sum_ {k = 1} ^ {s} \mathrm {P} (k, N _ {\text {s a m p l e}}, s) > \alpha_ {e} \tag {8} \\ \end{array} +$$ + +for a desired probability of error $\alpha_{e}$ . Intuitively, this value represents the most likely estimate of $N_{\mathrm{learned},\infty}$ that would produce $N_{\mathrm{learned}}$ unique samples after drawing $N_{\mathrm{sample}}$ images. We use the infinity symbol to indicate that this estimate represents the potential diversity if we sampled an infinite number of images. This value also reflects the model's diversity, representing the percentage of samples that the + +![](images/98cd23f6996e94800cea916e18829910f058f34d2571bf6a62ba02da8062a0af.jpg) +Figure 3. Overview of our proposed DiADM model. Instead of using placeholder labels for unconditional generation, we propose to leverage precomputed features from $\mathcal{F}$ instead. That way we can disentangle fidelity (FID) from diversity (IRS). + +![](images/56fece750a9c67e2dbcad5057031e9c8129d139664512ce735e0ffd2c56509e5.jpg) + +model can generate at its limit. In Appx. 7 we show the dependency of IRS and the confidence intervals on the number of samples showing that at $\alpha = 0.1$ , the estimates start to get very confident around the ground truth value. + +# 3.4. Adjusting for Measurement Gap + +In Sec. 4 we will show the measurement gap stemming from feature extractors collapsing to smaller representation spaces. This results in the underestimation of diversity for real data which in turn also influences the measured diversity for synthetic data. To eliminate the influence of the measurement gap we adjust for the reduced amount of diversity in the feature space of the extractor by normalizing by the measured diversity of real data according to the feature extractor. Effectively, we compute $\mathrm{IRS}_{\infty ,real}$ on real data using a reference dataset. Then we compute $\mathrm{IRS}_{\infty ,snth}$ on synthetic data and report the adjusted $\mathrm{IRS}_{\infty ,a} = \frac{\mathrm{IRS}_{\infty,snth}}{\mathrm{IRS}_{\infty,real}}$ instead. Intuitively, $\mathrm{IRS}_{\infty ,a}$ computes the ratio of diversity in the feature space of the synthetic images compared to the real images. Note that this implies that the adjusted score can be above one for low $\alpha$ values (see Fig. 4). We provide additional theoretical analysis in Appx. 15.1 and more quantitative results in Appx. 15.2, to demonstrate effectiveness in enhancing sensitivity and reliability for measuring diversity in generative models compared to other metrics. + +# 3.5. Diversity-Aware Diffusion Models + +To address diversity issues, we propose DiADMs. These models leverage the ability to extract meaningful representations using Inception v3 as feature extractor without relying on labels. Rather than employing placeholder labels, as is common in unconditional image generation, we use these pre-computed image features and feed them directly into the model. We term this approach 'pseudo-unconditional' because, while the architecture resembles a conditional setup, no explicit labels are needed. In principle, any feature extractor or custom self-supervised model can be used. However, we use the pre-trained ImageNet model, which is + +widely recognized as an effective feature extractor. The core idea is that, with proper training, the model behaves as if each training instance represents its own class, allowing for direct diversification of training instances during sampling time. Crucially, this approach decouples the diversity of the generated images from their fidelity as shown in Fig. 3. The diffusion model is responsible for producing high-fidelity images, while the diversity module ensures comprehensive coverage of the training distribution's diversity. We directly utilize features extracted from $\mathcal{F}$ of the training dataset to generate synthetic data, resulting in a synthetic dataset that should maintain diversity if trained properly. The backbone model employed is the XS architecture as proposed by [24], with modifications to adapt the label dimensionality to match the feature dimensionality of $\mathcal{F}$ . Inception is used as the feature extraction model for all datasets to maintain generalizability. + +# 4. Experiments + +Image Generation Models: Since diversity is mostly underexplored we aim to compare a series of different state-of-the-art methods against one another. [12] introduced classifier guidance and other modifications to Diffusion Models that first lead to competitive performance. Unlike all the current methods, they operate in image space. [41] were the first to introduce a model that performs diffusion in latent space. [24] use several improved training and design strategies for the U-Net backbone. [36] propose to use a Transformer architecture. [27] use autoregressive prediction and leverage masking for efficient training. Our experiments focus on both label-conditional and unconditional generative models, with an emphasis on enhancing diversity in unconditional image generation. This focus is motivated by the observation that any form of conditioning ultimately reduces to unconditional generation over the marginal distribution defined by the conditioning. We use pre-trained models directly and if no unconditional implementation was provided we mimic it by setting the class conditioning to the placeholder value used for unconditional training [22]. For our experiment with DiADM we train models ourselves using a fixed compute budget of 574 A40 GPU hours, which is only one-tenth of the training length in [24]. + +Metrics: We sample each model 50000 times which is a common choice to get a good estimate for FID [23, 49]. To mitigate the influence of distribution shift to the test datasets, which is the case for example in ImageNet, we report all metrics on the entire training set. To get a reference dataset for the adjusted IRS score, we take away 50000 samples for the reference dataset. If the datasets are smaller than 100000 samples we split the dataset in half and average the IRS results for as many runs as we have synthetic images. + +Datasets: For our primary experiments, we use ImageNet + +![](images/10ee7e34c66f64ebe7c2ee4d9ad46af4cc97bffd6329ff49843bceb5f20c1a41.jpg) +Figure 4. Visualization of the measurement gap across a diverse set of feature extractors by computing the unadjusted diversity of real data. The theoretical limit would be sampling $\mathrm{N}_{train}$ images where all of them correspond to a different image in the training dataset. The idealized scenario follows Eq. (1). + +512 [10]. To explore diversity within a more densely populated image space, we also include FFHQ [23], a benchmark dataset focused on facial images, and Chest-Xray14, which comprises medical chest X-ray images [52]. Additionally, we evaluate CelebV-HQ, a dataset containing face video frames [59], and Dynamic, a medical video dataset [35], treating the data as individual frames as done in [40]. The spatial resolution is fixed to 512 pixels and videos are additionally limited to the first 60 frames. + +Encoders: To find the best feature extractor $\mathcal{F}$ we compute the features using a large set of publicly available feature extractors following the selection of [49]. This includes BYOL [18], CLIP [38], a ConvNeXT based architecture [29], data2vec [3], DINOv2 [34], Inception [21, 45, 50], MAE [19], SwAV [7]. Additionally we add "Random" which uses a randomly initialized Inception-v3 [50] which has been shown to already extract meaningful features [33]. + +# 4.1. Measurement Gap + +First we need to decide for a feature extractor $\mathcal{F}$ that decides whether two samples correspond to each other. We choose to limit ourselves to large, pre-trained and publicly available feature extractors due to their universal applicability and established use. Custom feature extractors are also possible and could be used to increase interpretability even further. In Appx. 13 for example we compute IRS using a re-identification model trained to detect memorized samples unveiling diversity issues. To remain generalizable our goal is to find the most meaningful feature space in terms of image retrieval, that can be generally applied to different datasets. We experiment with a vast set of feature extractors $\mathcal{F}$ , measurements $\mathcal{P}$ , and datasets. Importantly we limit the experiments to evaluation on real data. In Sec. 3.1 we made the assumption, that all samples of the training dataset should be the main component of at + +least one synthetic image. In the context of only comparing real images, this means that every train image should have at least one test image it is more similar to than all other train images. However, as we show in Fig. 13 and Tab. 6 of the supplementary material, all feature extractors are drastically worse than the expected value. This is a clear indicator for the insufficient representative nature of feature extractors currently used to compute all metrics in the realm of image generation. We refer to this observation as measurement gap. In the supplements we present examples how this measurement gap can be reduced largely if tailord towards certain datasets (e.g., Tab. 5). However, to remain dataset agnostic we propose an adjustment step that uses the measured IRS on real data as reference value to make up for this measurement gap. Visually this gap results in a lower apparent diversity as shown in Fig. 4. There are two key takeaways from this experiment. First, not even real data is diverse according to commonly used feature extractors which severely reduces the interpretability of currently used diversity metrics. Secondly, it could be worthwhile for special tasks, to train feature extractors specifically to assess diversity as done in Tab. 5. + +However, thanks to the adjustment step described in Sec. 3.4, for our metric the absolute value of diversity of real data does not matter. For synthetic data it just matters how diverse the synthetic data is in comparison to the real data. This adjustment leads to direct interpretability of the metric as diversity of images. In Fig. 5 we visualize this interpretability of IRS by balancing and splitting ImageNet into train, reference and test datasets. We use the labels to manually reduce the diversity in the test dataset by successively removing classes. We use 500 images of each class for train and 100 image of each class for reference and test data. We can see a clear linear correlation between the number of classes and our proposed IRS score. Additionally, it is the only diversity metric that achieves $100\%$ diversity when all classes are present in the test dataset. The slight overestimation of IRS in the mid-diversity-range is likely to be caused by label ambiguity. Coverage also correlates well but severely overestimates diversity when the number of classes is very low. It saturates at 0.97 and completely fails for different hyperparameter settings as we show in Sec. 15.2. Recall, which is the most commonly used metric to assess synthetic diversity, saturates at about 0.8. From our point of view, this is one of the key reasons why diversity has not yet established itself as the key metric that should be optimized. + +While IRS works with any kind of feature extractor, we choose to use the one that has the best visual interpretability when it comes to image retrieval to make the results even more meaningful. Hence, we choose $\mathcal{F}$ based on the meaningfulness of the image retrieval. To avoid the necessity of a ground truth image retrieval dataset from the training + +
DatasetNConsensusBYOLCLIPConvNeXtdata2vecDINOv2InceptionMAERandomSwAV
ImageNet35142 (3.43%)21.7289.5695.512.6592.7074.2087.1712.3489.39
FFHQ1862 (3.33%)14.1893.0297.852.4292.1675.3586.7921.4398.60
ChestX-ray141389 (1.55%)40.3245.9377.751.9483.1575.9589.7033.7796.11
CelebV-HQ5329 (19.47%)31.8389.4093.716.6891.3379.9494.8651.7996.70
Dynamic23 (0.29%)47.8334.7895.658.7078.2665.2295.6565.22100.00
+ +Table 1. Measures the agreement of each $\mathcal{P}$ with the ensemble if five or more agree on the same retrieved image $(\%)$ . + +![](images/b9fe34051ebb70ed2da3e7a317a5cb11085a75d88afb99974c5b607ae18a6c84.jpg) +Figure 5. Measuring diversity of datasets by removing classes and computing IRS. If only 800 out of 1000 classes are part of the test set say that the diversity is at $80\%$ . By manually removing ImageNet classes we can assess how good commonly used metrics are at measuring diversity compared to ours. To improve visual interpretability we normalize FID and Vendi to be within 0 and 1 with 1 being best. IRS correlates best with the real diversity measured as fraction between number of classes in test and train dataset. + +dataset, we ensemble all feature extractors to compute a consensus on the image retrieval of real images. Details and qualitative examples can be found in Sec. 9. Consensus is reached if five or more models agree on the same image pair. For every image pair where consensus is reached we compute how often a single model agrees with the ensemble and report the results in Tab. 1. As expected the percentage of cases where consensus is reached is higher for natural images. Overall we see that SwAV generally agrees the most with the ensemble. Unless mentioned otherwise, this is the feature extractor we use to compute IRS scores. + +# 4.2. Diversity Issue in State-of-the-art Diffusion models + +Recently, [40] published a synthetic ultrasound video dataset based on the EchoNet dataset [35]. It is accompanied with a feature extraction model used to privatize the dataset using re-identification. Analyzing these features, [14] were the first to notice a diversity issue. Us + +ing our proposed IRS score we are now able to quantify it. First we try retrieve training samples using real frames from the test dataset. The dataset has 7465 training videos and we can see that 1277 test frames are able to retrieve 1159 distinct train videos. Following Eqs. (6) to (8) this means real data has a diversity of $\mathrm{IRS}_{\infty,real} = 86\%$ [75%, 100%]. Contrary to that, 1159 different synthetic frames are only able to retrieve 692 out of 7465 samples. This is equivalent to an $\mathrm{IRS}_{\infty,snth} = 12.3\%$ [11.9%, 12.9%]. Following the adjustment step introduced in Sec. 3.4 we conclude that the model has only learned $\mathrm{IRS}_{\infty,a} = 14.3\%$ of the data. Next, we compare multiple state-of-the-art approaches on conditional and unconditional image generation for ImageNet. We show their diversity according to IRS accompanied with other common metrics in Sec. 4.2. We initially expected that models for lower resolution outperform those with higher resolution. But overall this does not seem to be the case. Our experiments on MAR and EDM nicely confirm the feasibility of IRS as metrics as the score increases with the model size. Other metrics like Recall and Coverage confirm this observation but the difference seems marginal. Using IRS we can quantize and interpret the size of the gap between the smallest and largest models. The smallest model only reached a diversity of $46\%$ whereas the largest model reached $75\%$ . Unlike the relatively small gap of 0.9 points in FID, this indicates a large gap in sampling diversity. Crucially, following our experiments from Fig. 5 our proposed IRS score can be interpreted. Even the best model only properly learned $77\%$ of the data, irrespective of the number of samples generated. Note that this is for conditional image generation, so generation specifically asks the model to generate an equal number of images from all classes. Generally, the conditional models performed better than the unconditional model which leads us to believe that the conditioning needs to be improved. Due to limited resources, we will continue our experiments by fixing the compute and only use the best performing method in terms of FID and IRS which is EDM. To keep the training time reasonable we restrict our experiments on the XS variant. + +# 4.3. Improving Diversity Using Pseudo Labels + +Now that we have established EDM as the best method for unconditional image generation and at the same time have + +
ModelImage ResolutionFID ↓Prec. ↑Rec. ↑Dens. ↑Cov. ↑Vendi ↑IRS∞,a ↑ (Ours)
Pixel diffusion
ADM-256 [12]2566.01 (30.30)0.82 (0.57)0.62 (0.73)1.08 (0.41)0.91 (0.40)70.94 (36.18)0.44 (0.20)
Transformer
DiT-XL/2-256 [36]25622.15 (8.72)0.94 (0.69)0.34 (0.76)1.58 (0.70)0.85 (0.84)126.96 (58.15)0.23 (0.33)
DiT-XL/2-512 [36]51222.99 (9.54)0.96 (0.70)0.27 (0.73)1.90 (0.72)0.86 (0.82)128.98 (55.17)0.21 (0.34)
MAR-B-256 [27]2563.79 (10.36)0.83 (0.72)0.67 (0.71)1.18 (0.72)0.96 (0.75)83.03 (55.78)0.45 (0.38)
MAR-L-256 [27]2563.30 (10.36)0.82 (0.72)0.71 (0.71)1.10 (0.73)0.96 (0.75)81.80 (55.95)0.56 (0.38)
MAR-H-256 [27]2563.11 (10.36)0.82 (0.72)0.72 (0.71)1.07 (0.74)0.96 (0.76)81.37 (55.82)0.64 (0.38)
Latent diffusion, U-Net
LDM-256 [43]25626.09 (37.39)0.96 (0.61)0.21 (0.68)1.80 (0.45)0.83 (0.28)126.94 (30.83)0.16 (0.16)
EDM-2-XS-512 [24]5123.79 (75.02)0.83 (0.42)0.65 (0.63)1.22 (0.25)0.95 (0.13)72.41 (26.95)0.46 (0.09)
EDM-2-S-512 [24]5123.33 (122.48)0.85 (0.33)0.67 (0.42)1.26 (0.16)0.97 (0.07)80.25 (21.34)0.59 (0.04)
EDM-2-M-512 [24]5123.30 (107.45)0.85 (0.36)0.69 (0.61)1.24 (0.19)0.97 (0.09)82.99 (22.19)0.65 (0.06)
EDM-2-L-512 [24]5122.90 (118.87)0.84 (0.23)0.70 (0.51)1.22 (0.11)0.97 (0.06)82.10 (22.89)0.71 (0.03)
EDM-2-XL-512 [24]5122.92 (141.74)0.84 (0.25)0.71 (0.45)1.21 (0.12)0.97 (0.06)83.23 (20.04)0.77 (0.03)
EDM-2-XXL-512 [24]5122.87 (124.29)0.84 (0.33)0.71 (0.60)1.22 (0.17)0.97 (0.07)82.45 (21.52)0.75 (0.05)
+ +shown that this method still lacks diversity, we continue with EDM-2 as baseline for unconditional image generation and compare the trained models with a fixed compute budget. Quantitative results are shown in Tab. 3. The results show that the leveraging pseudo labels improves the sampling diversity compared to EDM-2 in all cases. In three cases the diversity is even better than that of the real reference dataset proofing that the pseudo conditioning works well and we can specifically query the model to generate diverse data. It also improves FID scores compared to the unconditional trainings. + +Table 2. Image generation of ImageNet across different generative models. Numbers in brackets indicate results when sampling the model without conditioning. Inception is used as feature extractor for all metrics except for IRS where we use SwAV. Precision, recall, density, coverage and Vendi score had to be computed on a subset of the training data of 50000 and 10000 for Vendi score. Pretrained conditional models were used as unconditional models by setting conditional guidance to 0. + +
FID ↓IRS∞,a ↑
EDM [24]DiADM (Ours)EDM [24]DiADM (Ours)
ImageNet-51251.5922.280.090.15
FFHQ40.926.240.231.51
ChestX-ray1424.296.760.191.08
CelebV-HQ68.4113.640.180.69
Dynamic13.825.560.601.04
+ +Table 3. FID and IRS results with and without our proposed diversity awareness module + +# 4.4. Text-to-image example + +Extending IRS to text-conditional image generation is straightforward. The evaluation of diversity depends on the chosen perspective, which is defined by the reference datasets. By changing the reference dataset, we can address various diversity-related issues, such as fairness. One notable example is the gender bias observed in text-to-image models. For our experiments, we use Deepfloyd (IF-I-XLv1.0) $^2$ . To assess gender diversity, we create a reference dataset with a balanced gender distribution for specific job + +roles by prompting the model to generate 100 images each of males and females. Half of these images are reserved as a balanced test dataset. We then prompt the model without specifying gender. The results are presented in Sec. 4.2. All groups, including the general term human, show gender bias, with the latter generating only female images. IRS estimates that in all cases, the generated diversity reaches only about $50\%$ of that of the balanced reference dataset. + +# 4.5. Limitations + +The stochasticity involved for a low number of samples means that minor IRS differences are not meaningful for a low number of samples. However, this is also reflected in the high values of uncertainty. Additionally, we believe that finding a better feature extractor that maximizes measured diversity (unadjusted IRS) on real datasets should be the focus. Experiments on improving diversity are limited to using real features for generation and restrictive computational budget. Furthermore, increasing the conditioning requires similar experiments analyzing memorization issues as text-conditioning [6]. + +# 5. Conclusion and Outlook + +In this paper, we reveal that all current methods for evaluating diversity and the feature extractors they are based on are inadequate. Using our proposed IRS metric, we demonstrate that these shortcomings result in all state-of-the-art image generation methods facing challenges with diversity. To address this issue for unconditional diffusion models, we propose DiADMs, which enhance the performance of current methods by separating diversity from fidelity. In future work, we aim to extend our efforts to improve diversity in conditional image generation as well. + +Acknowledgements: This work was supported by the High-Tech-Agenda Bavaria. HPC resources were provided by the Erlangen National High Performance Computing Center (NHR@FAU) of the Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU) under the NHR project b143dc and b180dc. NHR funding is provided by federal and Bavarian state authorities. NHR@FAU hardware is partially funded by the German Research Foundation (DFG) - 440719683. Support was also received by the ERC - projects MIA-NORMAL 101083647 as well as DFG 513220538, 512819079 and DFG large scale infrastructure funding Art 91b GG. This work was supported by the UKRI Centre for Doctoral Training in Artificial Intelligence for Healthcare (EP/S023283/1) and Ultromics Ltd. + +# References + +[1] Sumukh K Aithal, Pratyush Maini, Zachary C Lipton, and J Zico Kolter. Understanding hallucinations in diffusion models through mode interpolation. arXiv preprint arXiv:2406.09358, 2024. 1 +[2] Sina Alemohammad, Josue Casco-Rodriguez, Lorenzo Luzi, Ahmed Imtiaz Humayun, Hossein Babaei, Daniel LeJeune, Ali Siahkoohi, and Richard G. Baraniuk. Self-Consuming Generative Models Go MAD, 2023. arXiv:2307.01850 [cs]. 2 +[3] Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, and Michael Auli. data2vec: A General Framework for Self-supervised Learning in Speech, Vision and Language. In Proceedings of the 39th International Conference on Machine Learning, pages 1298-1312. PMLR, 2022. ISSN: 2640-3498. 6 +[4] Hritik Bansal, Da Yin, Masoud Monajatipoor, and Kai-Wei Chang. How well can text-to-image generative models understand ethical natural language interventions? arXiv preprint arXiv:2210.15230, 2022. 1, 3 +[5] Leonard E. Baum and Patrick Billingsley. Asymptotic Distributions for the Coupon Collector's Problem. The Annals of Mathematical Statistics, 36(6):1835-1839, 1965. Publisher: Institute of Mathematical Statistics. 3 +[6] Nicholas Carlini, Jamie Hayes, Milad Nasr, Matthew Jagielski, Vikash Sehwag, Florian Tramér, Borja Balle, Daphne Ippolito, and Eric Wallace. Extracting Training Data from Diffusion Models, 2023. arXiv:2301.13188 [cs]. 2, 8 +[7] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised Learning of Visual Features by Contrasting Cluster Assignments. In Advances in Neural Information Processing Systems, pages 9912-9924. Curran Associates, Inc., 2020. 6 +[8] Tingxiu Chen, Yilei Shi, Zixuan Zheng, Bingcong Yan, Jingliang Hu, Xiao Xiang Zhu, and Lichao Mou. Ultrasound Image-to-Video Synthesis via Latent Dynamic Diffusion Models. In Medical Image Computing and Computer Assisted Intervention - MICCAI 2024, pages 764-774. Springer Nature Switzerland, Cham, 2024. Series Title: Lecture Notes in Computer Science. 2 + +[9] Ching-Yao Chuang, Varun Jampani, Yuanzhen Li, Antonio Torralba, and Stefanie Jegelka. Debiasing vision-language models via biased prompts. arXiv preprint arXiv:2302.00070, 2023. 3 +[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. 2015. 6 +[11] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 1, 2 +[12] Prafulla Dhariwal and Alex Nichol. Diffusion Models Beat GANs on Image Synthesis, 2021. arXiv:2105.05233 [cs, stat]. 3, 5, 8 +[13] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. Cogview: Mastering text-to-image generation via transformers. Advances in neural information processing systems, 34:19822-19835, 2021. 2, 3 +[14] Mischa Dombrowski, Hadrien Reynaud, and Bernhard Kainz. Uncovering Hidden Subspaces in Video Diffusion Models Using Re-Identification, 2024. arXiv:2411.04956. 2, 7, 5, 6 +[15] Chun-Mei Feng, Kai Yu, Yong Liu, Salman Khan, and Wangmeng Zuo. Diverse Data Augmentation with Diffusions for Effective Test-time Prompt Tuning. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 2704-2714, Paris, France, 2023. IEEE. 2 +[16] Dan Friedman and Adji Bousso Dieng. The Vendi Score: A Diversity Evaluation Metric for Machine Learning, 2023. arXiv:2210.02410. 3 +[17] Felix Friedrich, Manuel Brack, Lukas Struppek, Dominik Hintersdorf, Patrick Schramowski, Sasha Luccioni, and Kristian Kersting. Fair Diffusion: Instructing Text-to-Image Generation Models on Fairness, 2023. arXiv:2302.10893. 2 +[18] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre H. Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, Bilal Piot, Koray Kavukcuoglu, Rémi Munos, and Michal Valko. Bootstrap your own latent: A new approach to self-supervised Learning, 2020. arXiv:2006.07733 [cs, stat]. 6 +[19] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked Autoencoders Are Scalable Vision Learners. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15979-15988, New Orleans, LA, USA, 2022. IEEE. 6 +[20] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 3, 6 +[21] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 6 +[22] Jonathan Ho and Tim Salimans. Classifier-Free Diffusion Guidance. 2022. 5 + +[23] Tero Karras, Samuli Laine, and Timo Aila. A Style-Based Generator Architecture for Generative Adversarial Networks. 2019. 5, 6 +[24] Tero Karras, Miika Aittala, Jaakko Lehtinen, Janne Hellsten, Timo Aila, and Samuli Laine. Analyzing and Improving the Training Dynamics of Diffusion Models, 2024. arXiv:2312.02696 [cs, stat]. 5, 8 +[25] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 3 +[26] Tuomas Kynkänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved Precision and Recall Metric for Assessing Generative Models. 2019. 3 +[27] Tianhong Li, Yonglong Tian, He Li, Mingyang Deng, and Kaiming He. Autoregressive Image Generation without Vector Quantization, 2024. arXiv:2406.11838 [cs]. 3, 5, 8 +[28] Jiawei Liu, Qiang Wang, Huijie Fan, Yinong Wang, Yandong Tang, and Liangqiong Qu. Residual denoising diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2773-2783, 2024. 2, 3 +[29] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A ConvNet for the 2020s, 2022. arXiv:2201.03545. 6 +[30] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations, 2022. 2, 3 +[31] Zichen Miao, Jiang Wang, Ze Wang, Zhengyuan Yang, Lijuan Wang, Qiang Qiu, and Zicheng Liu. Training diffusion models towards diverse image generation with reinforcement learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10844-10853, 2024. 1, 3 +[32] Muhammad Ferjad Naeem, Seong Joon Oh, Youngjung Uh, Yunjey Choi, and Jaejun Yoo. Reliable Fidelity and Diversity Metrics for Generative Models. 2020. 3 +[33] Jamie A. OaReilly and Fawad Asadi. Pre-trained vs. Random Weights for Calculating Fréchet Inception Distance in Medical Imaging. In 2021 13th Biomedical Engineering International Conference (BMEiCON), pages 1-4, Ayutthaya, Thailand, 2021. IEEE. 6, 5 +[34] Maxime Oquab, Timothee Darct, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINoV2: Learning Robust Visual Features without Supervision, 2024. arXiv:2304.07193 [cs]. 6, 2 +[35] David Ouyang, Bryan He, Amirata Ghorbani, Neal Yuan, Joseph Ebinger, Curtis P. Langlotz, Paul A. Heidenreich, Robert A. Harrington, David H. Liang, Euan A. Ashley, and + +James Y. Zou. Video-based AI for beat-to-beat assessment of cardiac function. Nature, 580(7802):252-256, 2020. Publisher: Springer US. 6, 7, 5 +[36] William Peebles and Saining Xie. Scalable Diffusion Models with Transformers. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 4172-4182, Paris, France, 2023. IEEE. 3, 5, 8 +[37] Yiming Qin, Huangjie Zheng, Jiangchao Yao, Mingyuan Zhou, and Ya Zhang. Class-balancing diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18434-18443, 2023. 3 +[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning Transferable Visual Models From Natural Language Supervision, 2021. arXiv:2103.00020. 6 +[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3 +[40] Hadrien Reynaud, Qingjie Meng, Mischa Dombrowski, Arijit Ghosh, Thomas Day, Alberto Gomez, Paul Leeson, and Bernhard Kainz. EchoNet-Synthetic: Privacy-preserving Video Generation for Safe Medical Data Sharing, 2024. arXiv:2406.00808 [cs]. 6, 7, 5 +[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-Resolution Image Synthesis with Latent Diffusion Models. 2021. 3, 5 +[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 3 +[43] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-Resolution Image Synthesis with Latent Diffusion Models, 2022. arXiv:2112.10752 [cs]. 8 +[44] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. DreamBooth: Fine Tuning Text-to-Image Diffusion Models for Subject-Driven Generation. 2022. 2 +[45] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, Xi Chen, and Xi Chen. Improved Techniques for Training GANs. 2016. 6 +[46] Axel Sauer, Katja Schwarz, and Andreas Geiger. StyleGANXL: Scaling StyleGAN to Large Diverse Datasets, 2022. arXiv:2202.00273 [cs]. 3 +[47] Vikash Sehwag, Caner Hazirbas, Albert Gordo, First Ozgenel, and Cristian Canton. Generating high fidelity data from low-density regions using diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11492-11501, 2022. 2 +[48] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2 + +[49] George Stein, Jesse C. Cresswell, Rasa Hosseinzadeh, Yi Sui, Brendan Leigh Ross, Valentin Villecloze, Zhaoyan Liu, Anthony L. Caterini, J. Eric T. Taylor, and Gabriel Loaiza-Ganem. Exposing flaws of generative model evaluation metrics and their unfair treatment of diffusion models, 2023. arXiv:2306.04675 [cs, stat]. 2, 5, 6 +[50] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the Inception Architecture for Computer Vision. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2818-2826, Las Vegas, NV, USA, 2016. IEEE. 6 +[51] N. M. Temme. Asymptotic Estimates of Stirling Numbers. Studies in Applied Mathematics, 89(3):233-243, 1993. 4 +[52] Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, Mohammadhadi Bagheri, and Ronald M Summers. ChestX-ray8: Hospital-Scale Chest X-Ray Database and Benchmarks on Weakly-Supervised Classification and Localization of Common Thorax Diseases. 2017. 6 +[53] Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang. Diffusion models: A comprehensive survey of methods and applications. ACM Computing Surveys, 56(4): 1-39, 2023. 1, 3 +[54] Ling Yang, Jingwei Liu, Shenda Hong, Zhilong Zhang, Zhilin Huang, Zheming Cai, Wentao Zhang, and Bin Cui. Improving diffusion-based image synthesis with context prediction. Advances in Neural Information Processing Systems, 36, 2024. 2 +[55] Cheng Zhang, Xuanbai Chen, Siqi Chai, Chen Henry Wu, Dmitry Lagun, Thabo Beeler, and Fernando De la Torre. Itigen: Inclusive text-to-image generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3969-3980, 2023. 2, 3 +[56] Weitong Zhang, Chengqi Zang, Liu Li, Sarah Cechnicka, Cheng Ouyang, and Bernhard Kainz. Stability and generalizability in sde diffusion models with measure-preserving dynamics. In NeurIPS, 2024. 1, 2 +[57] Yi Zhang, Xiaoyu Shi, Dasong Li, Xiaogang Wang, Jian Wang, and Hongsheng Li. A unified conditional framework for diffusion-based image restoration. Advances in Neural Information Processing Systems, 36, 2024. 2 +[58] Dora Zhao, Jerone T. A. Andrews, Orestis Papakyriakopoulos, and Alice Xiang. Position: Measure Dataset Diversity, Don't Just Claim It, 2024. arXiv:2407.08188. 2 +[59] Hao Zhu, Wayne Wu, Wentao Zhu, Liming Jiang, Siwei Tang, Li Zhang, Ziwei Liu, and Chen Change Loy. CelebV-HQ: A Large-Scale Video Facial Attributes Dataset, 2022. arXiv:2207.12393 [cs]. 6 \ No newline at end of file diff --git a/2025/Image Generation Diversity Issues and How to Tame Them/images.zip b/2025/Image Generation Diversity Issues and How to Tame Them/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8ad2f52c3d77aea5ec971a97764449b6f1380000 --- /dev/null +++ b/2025/Image Generation Diversity Issues and How to Tame Them/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bc260c2c82aae02b74f0ec14dc3a4584a6d9f382f710f014de312bf8d0839ecf +size 432516 diff --git a/2025/Image Generation Diversity Issues and How to Tame Them/layout.json b/2025/Image Generation Diversity Issues and How to Tame Them/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..de50b018e9aebc433e99bdd3dc4d7cf7f489e641 --- /dev/null +++ b/2025/Image Generation Diversity Issues and How to Tame Them/layout.json @@ -0,0 +1,7974 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 123, + 103, + 487, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 103, + 487, + 120 + ], + "spans": [ + { + "bbox": [ + 123, + 103, + 487, + 120 + ], + "type": "text", + "content": "Image Generation Diversity Issues and How to Tame Them" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 145, + 142, + 465, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 142, + 465, + 158 + ], + "spans": [ + { + "bbox": [ + 145, + 142, + 465, + 158 + ], + "type": "text", + "content": "Mischa Dombrowski1 Weitong Zhang2 Sarah Cechnicka2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 203, + 158, + 408, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 158, + 408, + 171 + ], + "spans": [ + { + "bbox": [ + 203, + 158, + 408, + 171 + ], + "type": "text", + "content": "Hadrien Reynaud² Bernhard Kainz1,2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 98, + 171, + 512, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 98, + 171, + 512, + 186 + ], + "spans": [ + { + "bbox": [ + 98, + 171, + 512, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 98, + 171, + 512, + 186 + ], + "type": "text", + "content": "Friedrich-Alexander-Universität Erlangen-Nürnberg " + }, + { + "bbox": [ + 98, + 171, + 512, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 98, + 171, + 512, + 186 + ], + "type": "text", + "content": "Imperial College London" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 239, + 188, + 370, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 188, + 370, + 198 + ], + "spans": [ + { + "bbox": [ + 239, + 188, + 370, + 198 + ], + "type": "text", + "content": "mischa.dombrowski@fau.de" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 152, + 227, + 200, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 227, + 200, + 239 + ], + "spans": [ + { + "bbox": [ + 152, + 227, + 200, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 293, + 296, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 293, + 296, + 700 + ], + "spans": [ + { + "bbox": [ + 53, + 293, + 296, + 700 + ], + "type": "text", + "content": "Generative methods have reached a level of quality that is almost indistinguishable from real data. However, while individual samples may appear unique, generative models often exhibit limitations in covering the full data distribution. Unlike quality issues, diversity problems within generative models are not easily detected by simply observing single images or generated datasets, which means we need a specific measure to assess the diversity of these models. In this paper, we draw attention to the current lack of diversity in generative models and the inability of common metrics to measure this. We achieve this by framing diversity as an image retrieval problem, where we measure how many real images can be retrieved using synthetic data as queries. This yields the Image Retrieval Score (IRS), an interpretable, hyperparameter-free metric that quantifies the diversity of a generative model's output. IRS requires only a subset of synthetic samples and provides a statistical measure of confidence. Our experiments indicate that current feature extractors commonly used in generative model assessment are inadequate for evaluating diversity effectively. Consequently, we perform an extensive search for the best feature extractors to assess diversity. Evaluation reveals that current diffusion models converge to limited subsets of the real distribution, with no current state-of-the-art models superpassing " + }, + { + "bbox": [ + 53, + 293, + 296, + 700 + ], + "type": "inline_equation", + "content": "77\\%" + }, + { + "bbox": [ + 53, + 293, + 296, + 700 + ], + "type": "text", + "content": " of the diversity of the training data. To address this limitation, we introduce Diversity-Aware Diffusion Models (DiADM), a novel approach that improves diversity of unconditional diffusion models without loss of image quality. We do this by disentangling diversity from image quality by using a diversity aware module that uses pseudo-unconditional features as input. We provide a Python package offering unified feature extraction and metric computation to further facilitate the evaluation of generative models https://github.com/MischaD/beyondfid." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 226, + 370, + 344 + ], + "blocks": [ + { + "bbox": [ + 316, + 226, + 370, + 344 + ], + "lines": [ + { + "bbox": [ + 316, + 226, + 370, + 344 + ], + "spans": [ + { + "bbox": [ + 316, + 226, + 370, + 344 + ], + "type": "image", + "image_path": "4167b51f1a8905e82a2b7e30a60ce5a15048409adb695873d68a5fa0959e4183.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 346, + 356, + 353 + ], + "lines": [ + { + "bbox": [ + 329, + 346, + 356, + 353 + ], + "spans": [ + { + "bbox": [ + 329, + 346, + 356, + 353 + ], + "type": "text", + "content": "Low Diversity" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 372, + 226, + 423, + 344 + ], + "blocks": [ + { + "bbox": [ + 372, + 226, + 423, + 344 + ], + "lines": [ + { + "bbox": [ + 372, + 226, + 423, + 344 + ], + "spans": [ + { + "bbox": [ + 372, + 226, + 423, + 344 + ], + "type": "image", + "image_path": "29ef5003bf151ccf25c9045c6134b6c99418f4e1c76d718d1ed9f44ec54fac28.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 381, + 346, + 410, + 353 + ], + "lines": [ + { + "bbox": [ + 381, + 346, + 410, + 353 + ], + "spans": [ + { + "bbox": [ + 381, + 346, + 410, + 353 + ], + "type": "text", + "content": "High Diversity" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 426, + 227, + 553, + 346 + ], + "blocks": [ + { + "bbox": [ + 426, + 227, + 553, + 346 + ], + "lines": [ + { + "bbox": [ + 426, + 227, + 553, + 346 + ], + "spans": [ + { + "bbox": [ + 426, + 227, + 553, + 346 + ], + "type": "image", + "image_path": "c8e1ad8d8111c37e0a5d63bfeaea045fec25ae6b303cf5504ad267bf538b045e.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 469, + 346, + 527, + 354 + ], + "lines": [ + { + "bbox": [ + 469, + 346, + 527, + 354 + ], + "spans": [ + { + "bbox": [ + 469, + 346, + 527, + 354 + ], + "type": "text", + "content": "Predicted Diversity " + }, + { + "bbox": [ + 469, + 346, + 527, + 354 + ], + "type": "inline_equation", + "content": "\\mathrm{IRS}_{\\infty}" + }, + { + "bbox": [ + 469, + 346, + 527, + 354 + ], + "type": "inline_equation", + "content": "(\\%)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 362, + 555, + 460 + ], + "lines": [ + { + "bbox": [ + 313, + 362, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 362, + 555, + 460 + ], + "type": "text", + "content": "Figure 1. Predicted gender diversity - as one possible example for diversity - after sampling pre-trained text-to-image diffusion models infinitely often. Visually it is apparent that most terms are inherently biased toward one gender. Our proposed IRS score predicts that off-the-shelf models will only ever reach 50 percent diversity, which is equivalent to one gender being perfectly represented while the other gender is missing entirely. Using diversification strategies increases diversity. Details about this can be found in Sec. 4.4." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 487, + 394, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 487, + 394, + 499 + ], + "spans": [ + { + "bbox": [ + 314, + 487, + 394, + 499 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 510, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 510, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 510, + 555, + 715 + ], + "type": "text", + "content": "With advancements in large-scale models and datasets, diffusion models have demonstrated substantial capacity to capture high-dimensional, complex distributions, enabling photo-realistic conditional image generation across diverse guidance types, including image-based [56], classifier-based [11], and text-based guidance [42]. However, diffusion models frequently inherit and amplify data biases due to their reliance on large-scale, naturally biased datasets [4, 53], limiting comprehensive distributional coverage in generated samples. For instance, text-to-image diffusion models often exhibit gender bias for certain occupations. When prompted with \"a teacher\", models predominantly generate females, with male representations notably underrepresented, as shown in Fig. 1. The \"conditional mode collapse\" [1, 31] also affects image- [56] and class-conditional [11] diffusion models, where the guidance often produces overly uniform outputs, leading to a loss in diver" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3029" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 396 + ], + "type": "text", + "content": "sity in the sampled distribution. Dataset diversity remains an unresolved issue, even for real datasets [58], due to ambiguous definitions and concepts. In contrast, synthetic data can leverage its training data as a reference. To avoid problems like mode drift [2], synthetic data should be at least as diverse as its training set. Key motivations for ensuring diversity include content creation [44], addressing privacy and memorization concerns [6], mitigating the lack of diversity in synthetic datasets [14], data augmentation [15], promoting algorithmic fairness [17], and enhancing downstream performance [8]. To improve diversity in conditional image generation, [28] introduces noise perturbations to enhance diversity, though at the expense of image quality and certainty. [30] balances a trade-off between faithfulness and realism in diverse image synthesis, where diversity comes at the cost of fidelity. These diversity-oriented approaches highlight an inherent compromise: diversity is achieved by sacrificing generation quality, with randomness that lacks interpretability. In parallel, classifier-based diversity guidance [47] and text-based methods [13, 55] have been developed to reduce bias in text-to-image generation. However, these techniques are task-specific and lack generalizability across different types of diffusion models for diversity enhancement. Thus, a fundamental challenge emerges: a universal, interpretable metric to evaluate diversity without bias or conditional dependence, and without compromising generation quality." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 396, + 296, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 296, + 564 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 296, + 564 + ], + "type": "text", + "content": "To address this, we propose the Image Retrieval Score (IRS), an interpretable metric for quantifying diversity in generative models. IRS assesses how effectively synthetic data retrieves real images as queries, providing a hyperparameter-free, statistically grounded evaluation based on minimal samples. Our approach is motivated by a thought experiment that illustrates how individual samples reveal model diversity. Using IRS as a core, generic metric for diversity, we quantify and reveal two key findings: (1) existing feature extractors used to evaluate generative models lack the capacity to measure diversity effectively, and (2) current diffusion models converge on limited subsets of the real distribution, with no state-of-the-art models surpassing " + }, + { + "bbox": [ + 55, + 396, + 296, + 564 + ], + "type": "inline_equation", + "content": "77\\%" + }, + { + "bbox": [ + 55, + 396, + 296, + 564 + ], + "type": "text", + "content": " of the diversity in training data." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 565, + 296, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 565, + 296, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 565, + 296, + 662 + ], + "type": "text", + "content": "This further motivates the use of IRS as an unbiased guidance to drive a comprehensive search for optimal feature extractors, enhancing diversity evaluation without compromising generation quality. Our selection is based on [49] who experimented on the impact of feature extractors on existing generative metrics. Building on these insights, we introduce Diversity-Aware Diffusion Models (DiADM), an approach that improves diversity across the full data distri" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 332, + 76, + 539, + 228 + ], + "blocks": [ + { + "bbox": [ + 332, + 76, + 539, + 228 + ], + "lines": [ + { + "bbox": [ + 332, + 76, + 539, + 228 + ], + "spans": [ + { + "bbox": [ + 332, + 76, + 539, + 228 + ], + "type": "image", + "image_path": "7dd0ecfe12433cd49daa81d1380b264c8580ecc2a72f7ca7c82077827aa0842d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 239, + 555, + 351 + ], + "lines": [ + { + "bbox": [ + 313, + 239, + 555, + 351 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 555, + 351 + ], + "type": "text", + "content": "Figure 2. We model finding image pairs (image retrieval) as randomly drawing from urns with replacement (Ideal). We observe that all used feature extraction models exhibit performance issues by collapsing in the feature space for real images resulting in a measurement gap affecting currently used metrics such as " + }, + { + "bbox": [ + 313, + 239, + 555, + 351 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_{\\mathrm{Inception}}" + }, + { + "bbox": [ + 313, + 239, + 555, + 351 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 239, + 555, + 351 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_{\\mathrm{DINOv2}}" + }, + { + "bbox": [ + 313, + 239, + 555, + 351 + ], + "type": "text", + "content": ", Precision and Recall. Comparing the results of synthetic images to real data shows that datasets generated by generative models show even stronger diversity issues which result in a synthetic data distribution gap. Our proposed metric leverages a real reference dataset to remove this measurement gap." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 371, + 555, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 371, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 555, + 430 + ], + "type": "text", + "content": "bution in unconditional diffusion models without compromising image quality. DiADM accomplishes this by disentangling diversity from image quality through a diversity-aware module utilizing pseudo-unconditional features. Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 315, + 431, + 553, + 597 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 315, + 431, + 553, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 431, + 553, + 466 + ], + "spans": [ + { + "bbox": [ + 315, + 431, + 553, + 466 + ], + "type": "text", + "content": "- We demonstrate the limitations of widely-used feature extractors for diversity metrics, highlighting their poor performance on real datasets." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 315, + 467, + 553, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 467, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 315, + 467, + 553, + 491 + ], + "type": "text", + "content": "- We propose an intuitive framework for assessing dataset diversity in unconditional diffusion models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 315, + 491, + 553, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 491, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 315, + 491, + 553, + 514 + ], + "type": "text", + "content": "- We identify and quantify the reduction in diversity resulting from current training methodologies." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 315, + 515, + 553, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 515, + 553, + 549 + ], + "spans": [ + { + "bbox": [ + 315, + 515, + 553, + 549 + ], + "type": "text", + "content": "- We provide a comprehensive analysis showing limited image diversity across a wide range of state-of-the-art diffusion models." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 315, + 551, + 553, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 551, + 553, + 597 + ], + "spans": [ + { + "bbox": [ + 315, + 551, + 553, + 597 + ], + "type": "text", + "content": "- We introduce a diversity-aware plug-in module, which utilizes pseudo-labels to separate diversity from fidelity, thereby enhancing generative diversity in unconditional diffusion models." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 609, + 402, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 609, + 402, + 621 + ], + "spans": [ + { + "bbox": [ + 314, + 609, + 402, + 621 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": "Image Generation: Diffusion models [48] progressively map complex distributions to a standard Gaussian. Conditional diffusion models have demonstrated a strong ability to generate highly realistic images that align closely with specified conditions, excelling in tasks such as image synthesis [54], image restoration [57], and image reconstruction [56]. Additionally, [11] enhances diffusion" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 673, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 673, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 673, + 295, + 713 + ], + "type": "text", + "content": "1Consider a die purported to have 1000 sides. If rolled 1000 times but yielding only 10 distinct outcomes, one would naturally question the claim. Similarly, in generative models, each sample serves as a \"roll\", reflecting the model's diversity in an unconditional manner." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3030" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "content": "models with an auxiliary classifier, enabling high-fidelity, class-conditional generation, while text-conditional models generate images based on text prompts [13, 55]. However, diffusion models often struggle to achieve sufficient diversity in task-agnostic generation, especially when trained on large-scale datasets that carry inherent biases [4, 53]. For example, Stable Diffusion, based on latent diffusion models (LDM) [39, 42], produces high-quality images aligned with text prompts, yet frequently reflects bias (e.g., age, gender) for certain terms. These limitations lead to restricted distributional coverage and biased outputs." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 225, + 295, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 225, + 295, + 429 + ], + "spans": [ + { + "bbox": [ + 55, + 225, + 295, + 429 + ], + "type": "text", + "content": "Diversity in Generation: Research on diversity in diffusion models is limited, with most existing approaches focused on GAN-based models [9, 25]. Current methods for diversity in diffusion models fall into the primary categories: bias mitigation and diversity enhancement. In bias mitigation, [37] uses a re-weighted loss to address long-tailed distributions, promoting balanced class representation. For text-conditional models, [13] attempts to mitigate bias by adding attribute words to prompts, though the added prompts may be ambiguous. [31] extends this to both class- and text-conditional models, using reinforcement learning to match the diversity ratio in reference images but relying on task-specific diversity reward functions. For diversity enhancement, [28] introduces noise perturbations, which increase diversity at the expense of image quality and certainty. Similarly, [30] balances diversity with realism, where gains in diversity tend to compromise fidelity." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 450, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 295, + 713 + ], + "type": "text", + "content": "Metrics of Diversity: Existing metrics in diffusion models remain limited and are unsuitable for accurate and interpretable diversity assessment. Traditional image quality metrics like FID [20] require extensive sampling (e.g., 50k images) to be reliable, making them computationally intensive and challenging to interpret in practical applications. Precision and recall [26] are most established in literature [12, 36, 41, 46]. They add supplementary insights but are hindered by dependency on dataset size, hyperparameters, and interpretability issues. Credibility and interpretability of metrics are highly underestimated properties. For example, a comparison of three years of advancements in diffusion models, as seen in [12] and [27], shows an order-of-magnitude improvement in FID but a one percentage point reduction in precision. Coverage and density [32] also measure diversity but are dependant on the ratio between train and test set, highly depend on hyperparameters, lack evaluation on real datasets, and density was designed to ignore real outliers which are important in the context of diversity assessment. Vendi Score [16], which operates entirely without a real reference set, sacrifices fidelity and is computationally prohibitive for large datasets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 314, + 71, + 370, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 370, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 370, + 83 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 91, + 553, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 553, + 140 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 313, + 91, + 553, + 140 + ], + "type": "inline_equation", + "content": "\\mathbf{x_t}" + }, + { + "bbox": [ + 313, + 91, + 553, + 140 + ], + "type": "text", + "content": " be image " + }, + { + "bbox": [ + 313, + 91, + 553, + 140 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 91, + 553, + 140 + ], + "type": "text", + "content": " from a dataset consisting of " + }, + { + "bbox": [ + 313, + 91, + 553, + 140 + ], + "type": "inline_equation", + "content": "\\mathrm{N}_{train}" + }, + { + "bbox": [ + 313, + 91, + 553, + 140 + ], + "type": "text", + "content": " real images residing in image space " + }, + { + "bbox": [ + 313, + 91, + 553, + 140 + ], + "type": "inline_equation", + "content": "\\mathcal{X} \\in \\mathbb{R}^{x\\times h\\times w}" + }, + { + "bbox": [ + 313, + 91, + 553, + 140 + ], + "type": "text", + "content": ". Unconditional generative models aim to learn the distribution " + }, + { + "bbox": [ + 313, + 91, + 553, + 140 + ], + "type": "inline_equation", + "content": "p_{data}(\\mathbf{X})" + }, + { + "bbox": [ + 313, + 91, + 553, + 140 + ], + "type": "text", + "content": " and sample " + }, + { + "bbox": [ + 313, + 91, + 553, + 140 + ], + "type": "inline_equation", + "content": "\\mathrm{N}_{\\text{sample}}" + }, + { + "bbox": [ + 313, + 91, + 553, + 140 + ], + "type": "text", + "content": " synthetic images from it." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 145, + 463, + 158 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 145, + 463, + 158 + ], + "spans": [ + { + "bbox": [ + 313, + 145, + 463, + 158 + ], + "type": "text", + "content": "3.1. Coupon Collector Problem" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 163, + 555, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 163, + 555, + 486 + ], + "spans": [ + { + "bbox": [ + 313, + 163, + 555, + 486 + ], + "type": "text", + "content": "The performance of generative models is generally measured by assessing the similarity of the real and the synthetic datasets. It is assumed that the training samples represent discrete observations from a continuous unknown distribution. We suggest viewing all synthetic images as a composition of real images, which holds true as long as there is no training objective that enforces out-of-distribution generation. To assess diversity, we argue that every training sample should be the main component of at least one synthetic sample. To measure this we use the fact that the observed training distribution is discrete. A model that has perfectly memorized every training sample would generate all images with equal probability " + }, + { + "bbox": [ + 313, + 163, + 555, + 486 + ], + "type": "inline_equation", + "content": "\\frac{1}{\\mathrm{N}_{train}}" + }, + { + "bbox": [ + 313, + 163, + 555, + 486 + ], + "type": "text", + "content": ". Consequently, we can model sampling synthetic images with an urn experiment. Each synthetic sample consists mainly of one component which is then compounded with other training images. A real sample is considered learned if it serves as the main component for at least one synthetic image. Evaluating diversity in this context can be framed as the coupon collector problem [5]. Let " + }, + { + "bbox": [ + 313, + 163, + 555, + 486 + ], + "type": "inline_equation", + "content": "P_{n}(N_{t} > 0)" + }, + { + "bbox": [ + 313, + 163, + 555, + 486 + ], + "type": "text", + "content": " denote the probability of generating " + }, + { + "bbox": [ + 313, + 163, + 555, + 486 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 163, + 555, + 486 + ], + "type": "text", + "content": " samples where at least one corresponds to a training image " + }, + { + "bbox": [ + 313, + 163, + 555, + 486 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 163, + 555, + 486 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 313, + 163, + 555, + 486 + ], + "type": "inline_equation", + "content": "N_{t}" + }, + { + "bbox": [ + 313, + 163, + 555, + 486 + ], + "type": "text", + "content": " denoting the number of generated images a training image is associated with. This probability can be computed by considering the compound probability of not sampling this training image which is " + }, + { + "bbox": [ + 313, + 163, + 555, + 486 + ], + "type": "inline_equation", + "content": "\\frac{\\mathrm{N}_{train} - 1}{\\mathrm{N}_{train}}" + }, + { + "bbox": [ + 313, + 163, + 555, + 486 + ], + "type": "text", + "content": ". Since the same probability applies to each image the expected value of the diversity can be calculated as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 349, + 501, + 555, + 529 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 501, + 555, + 529 + ], + "spans": [ + { + "bbox": [ + 349, + 501, + 555, + 529 + ], + "type": "interline_equation", + "content": "\\mathbb {E} [ \\text {D i v e r s i t y} ] = \\left(1 - \\left(\\frac {\\mathrm {N} _ {\\text {t r a i n}} - 1}{\\mathrm {N} _ {\\text {t r a i n}}}\\right) ^ {n}\\right). \\tag {1}", + "image_path": "40ebd3bdfe0059012b0556123cec30094995ddbfb70091b8c471abb1e7224f74.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 534, + 556, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 594 + ], + "type": "text", + "content": "In this definition, diversity is equal to the percentage of learned real images. This idealized scenario is shown as ideal in Fig. 2. Therefore, assessing diversity comes down to predicting which training image corresponds to which synthetic image." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 600, + 411, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 411, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 411, + 613 + ], + "type": "text", + "content": "3.2. Image Retrieval" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "content": "To assess diversity we first need to define how to decide on what constitutes a main component for each synthetic image. For each synthetic image instance " + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{t}}^{\\prime}" + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "content": " we aim to identify the real image " + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{t}}" + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "content": " that is closest to it according to a measurement " + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "content": " (e.g., Euclidean distance) on the feature of the images extracted by a pretrained feature extractor " + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "content": ". Then, to assess the diversity we have to compute the size of the dataset of learned images:" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "3031" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 91, + 303, + 126 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 303, + 126 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 303, + 126 + ], + "type": "interline_equation", + "content": "\\mathcal {X} _ {\\text {l e a r n e d}} = \\left\\{\\mathbf {x} _ {t} \\in \\mathcal {X} \\mid \\exists \\mathbf {x} _ {t} ^ {\\prime} \\in \\mathcal {X} ^ {\\prime}: \\mathbf {x} _ {t} = \\arg \\min _ {\\mathbf {x} _ {t}} \\mathcal {P} \\left(\\mathbf {x} _ {t}, \\mathbf {x} _ {t} ^ {\\prime}\\right) \\right\\}. \\tag {2}", + "image_path": "d5d1b29876767fa2d53923767f7d2b8583497d386573c5e1019433b78d4316fb.jpg" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 128, + 251, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 128, + 251, + 139 + ], + "spans": [ + { + "bbox": [ + 55, + 128, + 251, + 139 + ], + "type": "text", + "content": "We define the cardinality of this set as " + }, + { + "bbox": [ + 55, + 128, + 251, + 139 + ], + "type": "inline_equation", + "content": "\\mathbf{N}_{\\text{learned}}" + }, + { + "bbox": [ + 55, + 128, + 251, + 139 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 140, + 295, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 140, + 295, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 140, + 295, + 163 + ], + "type": "text", + "content": "We then call the trained image instance " + }, + { + "bbox": [ + 55, + 140, + 295, + 163 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{t}}" + }, + { + "bbox": [ + 55, + 140, + 295, + 163 + ], + "type": "text", + "content": " \"learned\" and compute the diversity of the synthetic dataset as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 121, + 172, + 295, + 198 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 172, + 295, + 198 + ], + "spans": [ + { + "bbox": [ + 121, + 172, + 295, + 198 + ], + "type": "interline_equation", + "content": "\\mathrm {I R S} _ {\\alpha} = \\frac {\\mathrm {N} _ {\\text {l e a r n e d}}}{\\mathrm {N} _ {\\text {t r a i n}}} \\in [ 0, 1 ] \\tag {3}", + "image_path": "34882acc1eacf2cb43fafc2e1babc4bcd956ef994d8d616493b77fd77a58d3d5.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 205, + 296, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 205, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 296, + 384 + ], + "type": "text", + "content": "We term this quantity as the image retrieval score (IRS). To emphasize that " + }, + { + "bbox": [ + 55, + 205, + 296, + 384 + ], + "type": "inline_equation", + "content": "\\mathrm{IRS}_{\\alpha}" + }, + { + "bbox": [ + 55, + 205, + 296, + 384 + ], + "type": "text", + "content": " depends on the number of images sampled we add " + }, + { + "bbox": [ + 55, + 205, + 296, + 384 + ], + "type": "inline_equation", + "content": "\\alpha := \\frac{\\mathrm{N}_{\\text{sample}}}{\\mathrm{N}_{\\text{train}}}" + }, + { + "bbox": [ + 55, + 205, + 296, + 384 + ], + "type": "text", + "content": " as a subscript. Note that all other generative metrics also depend on " + }, + { + "bbox": [ + 55, + 205, + 296, + 384 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 205, + 296, + 384 + ], + "type": "text", + "content": " but often ignore this dependency. The name image retrieval score derives from the fact that computing IRS also works when comparing two disjoint real datasets. In this context the problem can be thought of as an image retrieval task, where a real image serves as the query image and the goal is to find the most similar image in a comparison dataset. Intuitively, IRS measures the ratio of images in a training dataset that are retrievable by another dataset. If the comparison dataset consists of real images we say that an image is \"retrievable\". If the comparison dataset consists of synthetic images we say that an image is \"learned\"." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 385, + 296, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 385, + 296, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 385, + 296, + 662 + ], + "type": "text", + "content": "The key advantage of IRS is that we can give a reasonably good estimate of dataset diversity with only a minimal number of samples. For example, if the same image appears three times within just ten samples, it is a strong indicator of overfitting. Consequently, we aim to estimate the diversity after " + }, + { + "bbox": [ + 55, + 385, + 296, + 662 + ], + "type": "inline_equation", + "content": "n << \\mathrm{N}_{train}" + }, + { + "bbox": [ + 55, + 385, + 296, + 662 + ], + "type": "text", + "content": " samples. To compute the probability of retrieving exactly " + }, + { + "bbox": [ + 55, + 385, + 296, + 662 + ], + "type": "inline_equation", + "content": "k = \\mathcal{X}_{\\text{learned}}" + }, + { + "bbox": [ + 55, + 385, + 296, + 662 + ], + "type": "text", + "content": " images we use standard combinatorics. This problem can be split into two parts. Firstly, we compute the number of possibilities of splitting " + }, + { + "bbox": [ + 55, + 385, + 296, + 662 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 385, + 296, + 662 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 55, + 385, + 296, + 662 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 385, + 296, + 662 + ], + "type": "text", + "content": " different subsets. Instinctively, we are computing the number of combinations in which different synthetic samples derive from the same retrieved (still unassigned) image " + }, + { + "bbox": [ + 55, + 385, + 296, + 662 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 55, + 385, + 296, + 662 + ], + "type": "text", + "content": " according to equation Eq. (2). The computation involves using Stirling's number of the second kind, " + }, + { + "bbox": [ + 55, + 385, + 296, + 662 + ], + "type": "inline_equation", + "content": "\\operatorname{Stir}(n, k)" + }, + { + "bbox": [ + 55, + 385, + 296, + 662 + ], + "type": "text", + "content": ", which should be multiplied by the total number of possible assignments for each subset, given by " + }, + { + "bbox": [ + 55, + 385, + 296, + 662 + ], + "type": "inline_equation", + "content": "\\frac{\\mathrm{N}_{train}!}{(\\mathrm{N}_{train} - \\mathrm{N}_{\\text{learned}})!}" + }, + { + "bbox": [ + 55, + 385, + 296, + 662 + ], + "type": "text", + "content": ". This will become more clear with an example. Say, the first subset consists of three different images " + }, + { + "bbox": [ + 55, + 385, + 296, + 662 + ], + "type": "inline_equation", + "content": "x_3', x_{10}', x_{11}'" + }, + { + "bbox": [ + 55, + 385, + 296, + 662 + ], + "type": "text", + "content": ", all of which are learned from the same real image. Then for the first of " + }, + { + "bbox": [ + 55, + 385, + 296, + 662 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 385, + 296, + 662 + ], + "type": "text", + "content": " subsets, there are " + }, + { + "bbox": [ + 55, + 385, + 296, + 662 + ], + "type": "inline_equation", + "content": "\\mathrm{N}_{train}" + }, + { + "bbox": [ + 55, + 385, + 296, + 662 + ], + "type": "text", + "content": " different choices. For the second subset, we have " + }, + { + "bbox": [ + 55, + 385, + 296, + 662 + ], + "type": "inline_equation", + "content": "\\mathrm{N}_{train} - 1" + }, + { + "bbox": [ + 55, + 385, + 296, + 662 + ], + "type": "text", + "content": " choices and so on. Dividing by all choices leaves us with the probability for generating " + }, + { + "bbox": [ + 55, + 385, + 296, + 662 + ], + "type": "inline_equation", + "content": "k = \\mathrm{N}_{\\text{learned}}" + }, + { + "bbox": [ + 55, + 385, + 296, + 662 + ], + "type": "text", + "content": " different images:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 113, + 670, + 295, + 696 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 670, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 113, + 670, + 295, + 696 + ], + "type": "interline_equation", + "content": "\\mathrm {P} (k, n, s) = \\frac {\\operatorname {S t i r} (n , k) * s !}{(s - k) ! * (s) ^ {n}}, \\tag {4}", + "image_path": "6159af709a62292cd3ad49799ac0c8b5d81b1bb816ddd1912f06a7ef5f2039e6.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 701, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 715 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 55, + 701, + 295, + 715 + ], + "type": "inline_equation", + "content": "k = \\mathrm{N}_{\\text{learned}}" + }, + { + "bbox": [ + 55, + 701, + 295, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 701, + 295, + 715 + ], + "type": "inline_equation", + "content": "s = \\mathrm{N}_{\\text{train}}" + }, + { + "bbox": [ + 55, + 701, + 295, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 701, + 295, + 715 + ], + "type": "inline_equation", + "content": "n = \\mathrm{N}_{\\text{sample}}" + }, + { + "bbox": [ + 55, + 701, + 295, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": "For large values, we estimate this by computing the log of Eq. (4) instead. [51] showed that" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 315, + 116, + 553, + 157 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 116, + 553, + 157 + ], + "spans": [ + { + "bbox": [ + 315, + 116, + 553, + 157 + ], + "type": "interline_equation", + "content": "\\operatorname {S t i r} (n, k) = \\sqrt {\\frac {v - 1}{v (1 - G)}} \\left(\\frac {v - 1}{v - G}\\right) ^ {n - k} \\frac {k ^ {n}}{n ^ {k}} e ^ {k (1 - G)} \\binom {n} {k}, \\tag {5}", + "image_path": "e83b826cd9d8c903ab3181bdf01f1a7958b0f0fae4c02ac46722d6df0b85c2c1.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 158, + 555, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 158, + 555, + 241 + ], + "spans": [ + { + "bbox": [ + 313, + 158, + 555, + 241 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 158, + 555, + 241 + ], + "type": "inline_equation", + "content": "v = \\frac{n}{k}" + }, + { + "bbox": [ + 313, + 158, + 555, + 241 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 158, + 555, + 241 + ], + "type": "inline_equation", + "content": "G \\in (0,1)" + }, + { + "bbox": [ + 313, + 158, + 555, + 241 + ], + "type": "text", + "content": " is the unique solution to " + }, + { + "bbox": [ + 313, + 158, + 555, + 241 + ], + "type": "inline_equation", + "content": "G = v e^{G - v}" + }, + { + "bbox": [ + 313, + 158, + 555, + 241 + ], + "type": "text", + "content": ". Asymptotically, this estimate is close, reaching a maximum relative error of roughly " + }, + { + "bbox": [ + 313, + 158, + 555, + 241 + ], + "type": "inline_equation", + "content": "\\frac{0.066}{n}" + }, + { + "bbox": [ + 313, + 158, + 555, + 241 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 313, + 158, + 555, + 241 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 158, + 555, + 241 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 158, + 555, + 241 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 158, + 555, + 241 + ], + "type": "text", + "content": " are small, e.g., (10, 3) [51]. Computing this in log-space for ImageNet roughly takes 3 seconds and using memorization techniques reduces consecutive computations to 1.3 seconds." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 250, + 441, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 250, + 441, + 262 + ], + "spans": [ + { + "bbox": [ + 314, + 250, + 441, + 262 + ], + "type": "text", + "content": "3.3. Image Retrieval Score" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 266, + 554, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 554, + 328 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 554, + 328 + ], + "type": "text", + "content": "Given Eq. (5) we can now estimate IRS as well as a confidence interval for it. We assume that " + }, + { + "bbox": [ + 313, + 266, + 554, + 328 + ], + "type": "inline_equation", + "content": "\\mathrm{N}_{\\text {learned }}" + }, + { + "bbox": [ + 313, + 266, + 554, + 328 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 266, + 554, + 328 + ], + "type": "inline_equation", + "content": "\\mathrm{N}_{\\text {sample }}" + }, + { + "bbox": [ + 313, + 266, + 554, + 328 + ], + "type": "text", + "content": " are fixed and we aim to estimate the model diversity. To compute the maximum likelihood estimate for IRS, we optimize over values of " + }, + { + "bbox": [ + 313, + 266, + 554, + 328 + ], + "type": "inline_equation", + "content": "s \\in \\mathbb{N} \\cap [\\mathrm{N}_{\\text {learned }}, \\mathrm{N}_{\\text {train }}}" + }, + { + "bbox": [ + 313, + 266, + 554, + 328 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 325, + 336, + 554, + 362 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 336, + 554, + 362 + ], + "spans": [ + { + "bbox": [ + 325, + 336, + 554, + 362 + ], + "type": "interline_equation", + "content": "\\mathrm {I R S} _ {\\infty} = \\frac {1}{\\mathrm {N} _ {t r a i n}} \\arg \\max _ {s} \\mathrm {P} (\\mathrm {N} _ {l e a r n e d}, \\mathrm {N} _ {s a m p l e}, s), (6)", + "image_path": "940c2d8978e13299016453e5fe2afee51bc07dab3413f6a9ccca4dc1c2d0beba.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 369, + 555, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 555, + 429 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 555, + 429 + ], + "type": "text", + "content": "which means we expect " + }, + { + "bbox": [ + 313, + 369, + 555, + 429 + ], + "type": "inline_equation", + "content": "\\mathrm{N}_{\\mathrm{learned},\\infty} = \\mathrm{N}_{\\mathrm{train}}*\\mathrm{IRS}_{\\infty}" + }, + { + "bbox": [ + 313, + 369, + 555, + 429 + ], + "type": "text", + "content": " different images to be sampled eventually. To compute the lower bound we want to minimize the probability of the real diversity being lower but sampling abnormally many different images early. Mathematically, we estimate this as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 337, + 447, + 553, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 447, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 337, + 447, + 553, + 510 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathrm {I R S} _ {\\infty , L} = \\frac {1}{N _ {\\text {t r a i n}}} \\arg \\max _ {s \\in \\mathbb {N} \\cap [ N _ {\\text {l e a n e d}}, \\infty , N _ {\\text {t r a i n}} ]} \\\\ \\min \\left(N _ {\\text {t r a i n}}, N _ {\\text {s a m p l e}}\\right) \\tag {7} \\\\ \\sum_ {k = N _ {\\text {l e a n e d}}} \\quad \\mathrm {P} (k, N _ {\\text {s a m p l e}}, s) > \\alpha_ {e} \\\\ \\end{array}", + "image_path": "e61f1477dd3815061a190bedabd3c15a02079c3d944d20db046af51736db259e.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 520, + 554, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 520, + 554, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 520, + 554, + 544 + ], + "type": "text", + "content": "Accordingly, we can compute the estimate for the upper boundary:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 349, + 562, + 553, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 562, + 553, + 620 + ], + "spans": [ + { + "bbox": [ + 349, + 562, + 553, + 620 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathrm {I R S} _ {\\infty , U} = \\frac {1}{N _ {\\text {t r a i n}}} \\arg \\min _ {s \\in \\mathbb {N} \\cap \\left[ N _ {\\text {l e a n e d}}, N _ {\\text {l e a n e d}, \\infty} \\right]} \\\\ \\sum_ {k = 1} ^ {s} \\mathrm {P} (k, N _ {\\text {s a m p l e}}, s) > \\alpha_ {e} \\tag {8} \\\\ \\end{array}", + "image_path": "b085805f216f7720112f6f741c20619ae2a1dc5b70e706f3888b1feac33a7654.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 629, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 555, + 714 + ], + "type": "text", + "content": "for a desired probability of error " + }, + { + "bbox": [ + 313, + 629, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\alpha_{e}" + }, + { + "bbox": [ + 313, + 629, + 555, + 714 + ], + "type": "text", + "content": ". Intuitively, this value represents the most likely estimate of " + }, + { + "bbox": [ + 313, + 629, + 555, + 714 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{learned},\\infty}" + }, + { + "bbox": [ + 313, + 629, + 555, + 714 + ], + "type": "text", + "content": " that would produce " + }, + { + "bbox": [ + 313, + 629, + 555, + 714 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{learned}}" + }, + { + "bbox": [ + 313, + 629, + 555, + 714 + ], + "type": "text", + "content": " unique samples after drawing " + }, + { + "bbox": [ + 313, + 629, + 555, + 714 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{sample}}" + }, + { + "bbox": [ + 313, + 629, + 555, + 714 + ], + "type": "text", + "content": " images. We use the infinity symbol to indicate that this estimate represents the potential diversity if we sampled an infinite number of images. This value also reflects the model's diversity, representing the percentage of samples that the" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3032" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 170, + 166 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 170, + 166 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 170, + 166 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 170, + 166 + ], + "type": "image", + "image_path": "98cd23f6996e94800cea916e18829910f058f34d2571bf6a62ba02da8062a0af.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 173, + 295, + 218 + ], + "lines": [ + { + "bbox": [ + 55, + 173, + 295, + 218 + ], + "spans": [ + { + "bbox": [ + 55, + 173, + 295, + 218 + ], + "type": "text", + "content": "Figure 3. Overview of our proposed DiADM model. Instead of using placeholder labels for unconditional generation, we propose to leverage precomputed features from " + }, + { + "bbox": [ + 55, + 173, + 295, + 218 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 55, + 173, + 295, + 218 + ], + "type": "text", + "content": " instead. That way we can disentangle fidelity (FID) from diversity (IRS)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 175, + 70, + 293, + 166 + ], + "blocks": [ + { + "bbox": [ + 175, + 70, + 293, + 166 + ], + "lines": [ + { + "bbox": [ + 175, + 70, + 293, + 166 + ], + "spans": [ + { + "bbox": [ + 175, + 70, + 293, + 166 + ], + "type": "image", + "image_path": "56fece750a9c67e2dbcad5057031e9c8129d139664512ce735e0ffd2c56509e5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 239, + 296, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 296, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 296, + 288 + ], + "type": "text", + "content": "model can generate at its limit. In Appx. 7 we show the dependency of IRS and the confidence intervals on the number of samples showing that at " + }, + { + "bbox": [ + 55, + 239, + 296, + 288 + ], + "type": "inline_equation", + "content": "\\alpha = 0.1" + }, + { + "bbox": [ + 55, + 239, + 296, + 288 + ], + "type": "text", + "content": ", the estimates start to get very confident around the ground truth value." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 296, + 231, + 309 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 296, + 231, + 309 + ], + "spans": [ + { + "bbox": [ + 55, + 296, + 231, + 309 + ], + "type": "text", + "content": "3.4. Adjusting for Measurement Gap" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 313, + 296, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 313, + 296, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 313, + 296, + 556 + ], + "type": "text", + "content": "In Sec. 4 we will show the measurement gap stemming from feature extractors collapsing to smaller representation spaces. This results in the underestimation of diversity for real data which in turn also influences the measured diversity for synthetic data. To eliminate the influence of the measurement gap we adjust for the reduced amount of diversity in the feature space of the extractor by normalizing by the measured diversity of real data according to the feature extractor. Effectively, we compute " + }, + { + "bbox": [ + 55, + 313, + 296, + 556 + ], + "type": "inline_equation", + "content": "\\mathrm{IRS}_{\\infty ,real}" + }, + { + "bbox": [ + 55, + 313, + 296, + 556 + ], + "type": "text", + "content": " on real data using a reference dataset. Then we compute " + }, + { + "bbox": [ + 55, + 313, + 296, + 556 + ], + "type": "inline_equation", + "content": "\\mathrm{IRS}_{\\infty ,snth}" + }, + { + "bbox": [ + 55, + 313, + 296, + 556 + ], + "type": "text", + "content": " on synthetic data and report the adjusted " + }, + { + "bbox": [ + 55, + 313, + 296, + 556 + ], + "type": "inline_equation", + "content": "\\mathrm{IRS}_{\\infty ,a} = \\frac{\\mathrm{IRS}_{\\infty,snth}}{\\mathrm{IRS}_{\\infty,real}}" + }, + { + "bbox": [ + 55, + 313, + 296, + 556 + ], + "type": "text", + "content": " instead. Intuitively, " + }, + { + "bbox": [ + 55, + 313, + 296, + 556 + ], + "type": "inline_equation", + "content": "\\mathrm{IRS}_{\\infty ,a}" + }, + { + "bbox": [ + 55, + 313, + 296, + 556 + ], + "type": "text", + "content": " computes the ratio of diversity in the feature space of the synthetic images compared to the real images. Note that this implies that the adjusted score can be above one for low " + }, + { + "bbox": [ + 55, + 313, + 296, + 556 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 313, + 296, + 556 + ], + "type": "text", + "content": " values (see Fig. 4). We provide additional theoretical analysis in Appx. 15.1 and more quantitative results in Appx. 15.2, to demonstrate effectiveness in enhancing sensitivity and reliability for measuring diversity in generative models compared to other metrics." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 564, + 237, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 237, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 237, + 577 + ], + "type": "text", + "content": "3.5. Diversity-Aware Diffusion Models" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": "To address diversity issues, we propose DiADMs. These models leverage the ability to extract meaningful representations using Inception v3 as feature extractor without relying on labels. Rather than employing placeholder labels, as is common in unconditional image generation, we use these pre-computed image features and feed them directly into the model. We term this approach 'pseudo-unconditional' because, while the architecture resembles a conditional setup, no explicit labels are needed. In principle, any feature extractor or custom self-supervised model can be used. However, we use the pre-trained ImageNet model, which is" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 555, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 277 + ], + "type": "text", + "content": "widely recognized as an effective feature extractor. The core idea is that, with proper training, the model behaves as if each training instance represents its own class, allowing for direct diversification of training instances during sampling time. Crucially, this approach decouples the diversity of the generated images from their fidelity as shown in Fig. 3. The diffusion model is responsible for producing high-fidelity images, while the diversity module ensures comprehensive coverage of the training distribution's diversity. We directly utilize features extracted from " + }, + { + "bbox": [ + 313, + 72, + 555, + 277 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 313, + 72, + 555, + 277 + ], + "type": "text", + "content": " of the training dataset to generate synthetic data, resulting in a synthetic dataset that should maintain diversity if trained properly. The backbone model employed is the XS architecture as proposed by [24], with modifications to adapt the label dimensionality to match the feature dimensionality of " + }, + { + "bbox": [ + 313, + 72, + 555, + 277 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 313, + 72, + 555, + 277 + ], + "type": "text", + "content": ". Inception is used as the feature extraction model for all datasets to maintain generalizability." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 286, + 394, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 286, + 394, + 300 + ], + "spans": [ + { + "bbox": [ + 314, + 286, + 394, + 300 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 307, + 555, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 307, + 555, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 307, + 555, + 582 + ], + "type": "text", + "content": "Image Generation Models: Since diversity is mostly underexplored we aim to compare a series of different state-of-the-art methods against one another. [12] introduced classifier guidance and other modifications to Diffusion Models that first lead to competitive performance. Unlike all the current methods, they operate in image space. [41] were the first to introduce a model that performs diffusion in latent space. [24] use several improved training and design strategies for the U-Net backbone. [36] propose to use a Transformer architecture. [27] use autoregressive prediction and leverage masking for efficient training. Our experiments focus on both label-conditional and unconditional generative models, with an emphasis on enhancing diversity in unconditional image generation. This focus is motivated by the observation that any form of conditioning ultimately reduces to unconditional generation over the marginal distribution defined by the conditioning. We use pre-trained models directly and if no unconditional implementation was provided we mimic it by setting the class conditioning to the placeholder value used for unconditional training [22]. For our experiment with DiADM we train models ourselves using a fixed compute budget of 574 A40 GPU hours, which is only one-tenth of the training length in [24]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 582, + 556, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 701 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 701 + ], + "type": "text", + "content": "Metrics: We sample each model 50000 times which is a common choice to get a good estimate for FID [23, 49]. To mitigate the influence of distribution shift to the test datasets, which is the case for example in ImageNet, we report all metrics on the entire training set. To get a reference dataset for the adjusted IRS score, we take away 50000 samples for the reference dataset. If the datasets are smaller than 100000 samples we split the dataset in half and average the IRS results for as many runs as we have synthetic images." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "type": "text", + "content": "Datasets: For our primary experiments, we use ImageNet" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3033" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 83, + 73, + 269, + 205 + ], + "blocks": [ + { + "bbox": [ + 83, + 73, + 269, + 205 + ], + "lines": [ + { + "bbox": [ + 83, + 73, + 269, + 205 + ], + "spans": [ + { + "bbox": [ + 83, + 73, + 269, + 205 + ], + "type": "image", + "image_path": "10ee7e34c66f64ebe7c2ee4d9ad46af4cc97bffd6329ff49843bceb5f20c1a41.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 216, + 298, + 273 + ], + "lines": [ + { + "bbox": [ + 55, + 216, + 298, + 273 + ], + "spans": [ + { + "bbox": [ + 55, + 216, + 298, + 273 + ], + "type": "text", + "content": "Figure 4. Visualization of the measurement gap across a diverse set of feature extractors by computing the unadjusted diversity of real data. The theoretical limit would be sampling " + }, + { + "bbox": [ + 55, + 216, + 298, + 273 + ], + "type": "inline_equation", + "content": "\\mathrm{N}_{train}" + }, + { + "bbox": [ + 55, + 216, + 298, + 273 + ], + "type": "text", + "content": " images where all of them correspond to a different image in the training dataset. The idealized scenario follows Eq. (1)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 292, + 295, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 292, + 295, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 292, + 295, + 399 + ], + "type": "text", + "content": "512 [10]. To explore diversity within a more densely populated image space, we also include FFHQ [23], a benchmark dataset focused on facial images, and Chest-Xray14, which comprises medical chest X-ray images [52]. Additionally, we evaluate CelebV-HQ, a dataset containing face video frames [59], and Dynamic, a medical video dataset [35], treating the data as individual frames as done in [40]. The spatial resolution is fixed to 512 pixels and videos are additionally limited to the first 60 frames." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 400, + 295, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 400, + 295, + 496 + ], + "spans": [ + { + "bbox": [ + 55, + 400, + 295, + 496 + ], + "type": "text", + "content": "Encoders: To find the best feature extractor " + }, + { + "bbox": [ + 55, + 400, + 295, + 496 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 55, + 400, + 295, + 496 + ], + "type": "text", + "content": " we compute the features using a large set of publicly available feature extractors following the selection of [49]. This includes BYOL [18], CLIP [38], a ConvNeXT based architecture [29], data2vec [3], DINOv2 [34], Inception [21, 45, 50], MAE [19], SwAV [7]. Additionally we add \"Random\" which uses a randomly initialized Inception-v3 [50] which has been shown to already extract meaningful features [33]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 504, + 167, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 504, + 167, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 504, + 167, + 517 + ], + "type": "text", + "content": "4.1. Measurement Gap" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "text", + "content": "First we need to decide for a feature extractor " + }, + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "text", + "content": " that decides whether two samples correspond to each other. We choose to limit ourselves to large, pre-trained and publicly available feature extractors due to their universal applicability and established use. Custom feature extractors are also possible and could be used to increase interpretability even further. In Appx. 13 for example we compute IRS using a re-identification model trained to detect memorized samples unveiling diversity issues. To remain generalizable our goal is to find the most meaningful feature space in terms of image retrieval, that can be generally applied to different datasets. We experiment with a vast set of feature extractors " + }, + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "text", + "content": ", measurements " + }, + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "text", + "content": ", and datasets. Importantly we limit the experiments to evaluation on real data. In Sec. 3.1 we made the assumption, that all samples of the training dataset should be the main component of at" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 335 + ], + "type": "text", + "content": "least one synthetic image. In the context of only comparing real images, this means that every train image should have at least one test image it is more similar to than all other train images. However, as we show in Fig. 13 and Tab. 6 of the supplementary material, all feature extractors are drastically worse than the expected value. This is a clear indicator for the insufficient representative nature of feature extractors currently used to compute all metrics in the realm of image generation. We refer to this observation as measurement gap. In the supplements we present examples how this measurement gap can be reduced largely if tailord towards certain datasets (e.g., Tab. 5). However, to remain dataset agnostic we propose an adjustment step that uses the measured IRS on real data as reference value to make up for this measurement gap. Visually this gap results in a lower apparent diversity as shown in Fig. 4. There are two key takeaways from this experiment. First, not even real data is diverse according to commonly used feature extractors which severely reduces the interpretability of currently used diversity metrics. Secondly, it could be worthwhile for special tasks, to train feature extractors specifically to assess diversity as done in Tab. 5." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 339, + 553, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 339, + 553, + 638 + ], + "spans": [ + { + "bbox": [ + 313, + 339, + 553, + 638 + ], + "type": "text", + "content": "However, thanks to the adjustment step described in Sec. 3.4, for our metric the absolute value of diversity of real data does not matter. For synthetic data it just matters how diverse the synthetic data is in comparison to the real data. This adjustment leads to direct interpretability of the metric as diversity of images. In Fig. 5 we visualize this interpretability of IRS by balancing and splitting ImageNet into train, reference and test datasets. We use the labels to manually reduce the diversity in the test dataset by successively removing classes. We use 500 images of each class for train and 100 image of each class for reference and test data. We can see a clear linear correlation between the number of classes and our proposed IRS score. Additionally, it is the only diversity metric that achieves " + }, + { + "bbox": [ + 313, + 339, + 553, + 638 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 313, + 339, + 553, + 638 + ], + "type": "text", + "content": " diversity when all classes are present in the test dataset. The slight overestimation of IRS in the mid-diversity-range is likely to be caused by label ambiguity. Coverage also correlates well but severely overestimates diversity when the number of classes is very low. It saturates at 0.97 and completely fails for different hyperparameter settings as we show in Sec. 15.2. Recall, which is the most commonly used metric to assess synthetic diversity, saturates at about 0.8. From our point of view, this is one of the key reasons why diversity has not yet established itself as the key metric that should be optimized." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": "While IRS works with any kind of feature extractor, we choose to use the one that has the best visual interpretability when it comes to image retrieval to make the results even more meaningful. Hence, we choose " + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": " based on the meaningfulness of the image retrieval. To avoid the necessity of a ground truth image retrieval dataset from the training" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "3034" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 70, + 552, + 148 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 552, + 148 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 552, + 148 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 552, + 148 + ], + "type": "table", + "html": "
DatasetNConsensusBYOLCLIPConvNeXtdata2vecDINOv2InceptionMAERandomSwAV
ImageNet35142 (3.43%)21.7289.5695.512.6592.7074.2087.1712.3489.39
FFHQ1862 (3.33%)14.1893.0297.852.4292.1675.3586.7921.4398.60
ChestX-ray141389 (1.55%)40.3245.9377.751.9483.1575.9589.7033.7796.11
CelebV-HQ5329 (19.47%)31.8389.4093.716.6891.3379.9494.8651.7996.70
Dynamic23 (0.29%)47.8334.7895.658.7078.2665.2295.6565.22100.00
", + "image_path": "2321420e196cac40dd722752745f1c82ffa348e6538fade110bd31e8da47b3b0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 93, + 157, + 515, + 169 + ], + "lines": [ + { + "bbox": [ + 93, + 157, + 515, + 169 + ], + "spans": [ + { + "bbox": [ + 93, + 157, + 515, + 169 + ], + "type": "text", + "content": "Table 1. Measures the agreement of each " + }, + { + "bbox": [ + 93, + 157, + 515, + 169 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 93, + 157, + 515, + 169 + ], + "type": "text", + "content": " with the ensemble if five or more agree on the same retrieved image " + }, + { + "bbox": [ + 93, + 157, + 515, + 169 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 93, + 157, + 515, + 169 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 62, + 193, + 290, + 342 + ], + "blocks": [ + { + "bbox": [ + 62, + 193, + 290, + 342 + ], + "lines": [ + { + "bbox": [ + 62, + 193, + 290, + 342 + ], + "spans": [ + { + "bbox": [ + 62, + 193, + 290, + 342 + ], + "type": "image", + "image_path": "b9fe34051ebb70ed2da3e7a317a5cb11085a75d88afb99974c5b607ae18a6c84.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 355, + 297, + 455 + ], + "lines": [ + { + "bbox": [ + 55, + 355, + 297, + 455 + ], + "spans": [ + { + "bbox": [ + 55, + 355, + 297, + 455 + ], + "type": "text", + "content": "Figure 5. Measuring diversity of datasets by removing classes and computing IRS. If only 800 out of 1000 classes are part of the test set say that the diversity is at " + }, + { + "bbox": [ + 55, + 355, + 297, + 455 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 55, + 355, + 297, + 455 + ], + "type": "text", + "content": ". By manually removing ImageNet classes we can assess how good commonly used metrics are at measuring diversity compared to ours. To improve visual interpretability we normalize FID and Vendi to be within 0 and 1 with 1 being best. IRS correlates best with the real diversity measured as fraction between number of classes in test and train dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 479, + 296, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 296, + 611 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 296, + 611 + ], + "type": "text", + "content": "dataset, we ensemble all feature extractors to compute a consensus on the image retrieval of real images. Details and qualitative examples can be found in Sec. 9. Consensus is reached if five or more models agree on the same image pair. For every image pair where consensus is reached we compute how often a single model agrees with the ensemble and report the results in Tab. 1. As expected the percentage of cases where consensus is reached is higher for natural images. Overall we see that SwAV generally agrees the most with the ensemble. Unless mentioned otherwise, this is the feature extractor we use to compute IRS scores." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 622, + 296, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 622, + 296, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 622, + 296, + 647 + ], + "type": "text", + "content": "4.2. Diversity Issue in State-of-the-art Diffusion models" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "Recently, [40] published a synthetic ultrasound video dataset based on the EchoNet dataset [35]. It is accompanied with a feature extraction model used to privatize the dataset using re-identification. Analyzing these features, [14] were the first to notice a diversity issue. Us" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 190, + 555, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 190, + 555, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 190, + 555, + 658 + ], + "type": "text", + "content": "ing our proposed IRS score we are now able to quantify it. First we try retrieve training samples using real frames from the test dataset. The dataset has 7465 training videos and we can see that 1277 test frames are able to retrieve 1159 distinct train videos. Following Eqs. (6) to (8) this means real data has a diversity of " + }, + { + "bbox": [ + 313, + 190, + 555, + 658 + ], + "type": "inline_equation", + "content": "\\mathrm{IRS}_{\\infty,real} = 86\\%" + }, + { + "bbox": [ + 313, + 190, + 555, + 658 + ], + "type": "text", + "content": " [75%, 100%]. Contrary to that, 1159 different synthetic frames are only able to retrieve 692 out of 7465 samples. This is equivalent to an " + }, + { + "bbox": [ + 313, + 190, + 555, + 658 + ], + "type": "inline_equation", + "content": "\\mathrm{IRS}_{\\infty,snth} = 12.3\\%" + }, + { + "bbox": [ + 313, + 190, + 555, + 658 + ], + "type": "text", + "content": " [11.9%, 12.9%]. Following the adjustment step introduced in Sec. 3.4 we conclude that the model has only learned " + }, + { + "bbox": [ + 313, + 190, + 555, + 658 + ], + "type": "inline_equation", + "content": "\\mathrm{IRS}_{\\infty,a} = 14.3\\%" + }, + { + "bbox": [ + 313, + 190, + 555, + 658 + ], + "type": "text", + "content": " of the data. Next, we compare multiple state-of-the-art approaches on conditional and unconditional image generation for ImageNet. We show their diversity according to IRS accompanied with other common metrics in Sec. 4.2. We initially expected that models for lower resolution outperform those with higher resolution. But overall this does not seem to be the case. Our experiments on MAR and EDM nicely confirm the feasibility of IRS as metrics as the score increases with the model size. Other metrics like Recall and Coverage confirm this observation but the difference seems marginal. Using IRS we can quantize and interpret the size of the gap between the smallest and largest models. The smallest model only reached a diversity of " + }, + { + "bbox": [ + 313, + 190, + 555, + 658 + ], + "type": "inline_equation", + "content": "46\\%" + }, + { + "bbox": [ + 313, + 190, + 555, + 658 + ], + "type": "text", + "content": " whereas the largest model reached " + }, + { + "bbox": [ + 313, + 190, + 555, + 658 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 313, + 190, + 555, + 658 + ], + "type": "text", + "content": ". Unlike the relatively small gap of 0.9 points in FID, this indicates a large gap in sampling diversity. Crucially, following our experiments from Fig. 5 our proposed IRS score can be interpreted. Even the best model only properly learned " + }, + { + "bbox": [ + 313, + 190, + 555, + 658 + ], + "type": "inline_equation", + "content": "77\\%" + }, + { + "bbox": [ + 313, + 190, + 555, + 658 + ], + "type": "text", + "content": " of the data, irrespective of the number of samples generated. Note that this is for conditional image generation, so generation specifically asks the model to generate an equal number of images from all classes. Generally, the conditional models performed better than the unconditional model which leads us to believe that the conditioning needs to be improved. Due to limited resources, we will continue our experiments by fixing the compute and only use the best performing method in terms of FID and IRS which is EDM. To keep the training time reasonable we restrict our experiments on the XS variant." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 670, + 533, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 670, + 533, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 670, + 533, + 684 + ], + "type": "text", + "content": "4.3. Improving Diversity Using Pseudo Labels" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "Now that we have established EDM as the best method for unconditional image generation and at the same time have" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "3035" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 69, + 555, + 234 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 555, + 234 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 555, + 234 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 555, + 234 + ], + "type": "table", + "html": "
ModelImage ResolutionFID ↓Prec. ↑Rec. ↑Dens. ↑Cov. ↑Vendi ↑IRS∞,a ↑ (Ours)
Pixel diffusion
ADM-256 [12]2566.01 (30.30)0.82 (0.57)0.62 (0.73)1.08 (0.41)0.91 (0.40)70.94 (36.18)0.44 (0.20)
Transformer
DiT-XL/2-256 [36]25622.15 (8.72)0.94 (0.69)0.34 (0.76)1.58 (0.70)0.85 (0.84)126.96 (58.15)0.23 (0.33)
DiT-XL/2-512 [36]51222.99 (9.54)0.96 (0.70)0.27 (0.73)1.90 (0.72)0.86 (0.82)128.98 (55.17)0.21 (0.34)
MAR-B-256 [27]2563.79 (10.36)0.83 (0.72)0.67 (0.71)1.18 (0.72)0.96 (0.75)83.03 (55.78)0.45 (0.38)
MAR-L-256 [27]2563.30 (10.36)0.82 (0.72)0.71 (0.71)1.10 (0.73)0.96 (0.75)81.80 (55.95)0.56 (0.38)
MAR-H-256 [27]2563.11 (10.36)0.82 (0.72)0.72 (0.71)1.07 (0.74)0.96 (0.76)81.37 (55.82)0.64 (0.38)
Latent diffusion, U-Net
LDM-256 [43]25626.09 (37.39)0.96 (0.61)0.21 (0.68)1.80 (0.45)0.83 (0.28)126.94 (30.83)0.16 (0.16)
EDM-2-XS-512 [24]5123.79 (75.02)0.83 (0.42)0.65 (0.63)1.22 (0.25)0.95 (0.13)72.41 (26.95)0.46 (0.09)
EDM-2-S-512 [24]5123.33 (122.48)0.85 (0.33)0.67 (0.42)1.26 (0.16)0.97 (0.07)80.25 (21.34)0.59 (0.04)
EDM-2-M-512 [24]5123.30 (107.45)0.85 (0.36)0.69 (0.61)1.24 (0.19)0.97 (0.09)82.99 (22.19)0.65 (0.06)
EDM-2-L-512 [24]5122.90 (118.87)0.84 (0.23)0.70 (0.51)1.22 (0.11)0.97 (0.06)82.10 (22.89)0.71 (0.03)
EDM-2-XL-512 [24]5122.92 (141.74)0.84 (0.25)0.71 (0.45)1.21 (0.12)0.97 (0.06)83.23 (20.04)0.77 (0.03)
EDM-2-XXL-512 [24]5122.87 (124.29)0.84 (0.33)0.71 (0.60)1.22 (0.17)0.97 (0.07)82.45 (21.52)0.75 (0.05)
", + "image_path": "dcdba2cc70e8f380a0576c0d877f5c02f69a6484383cf1f42f24e568a8a6c22a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 307, + 294, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 307, + 294, + 439 + ], + "spans": [ + { + "bbox": [ + 54, + 307, + 294, + 439 + ], + "type": "text", + "content": "shown that this method still lacks diversity, we continue with EDM-2 as baseline for unconditional image generation and compare the trained models with a fixed compute budget. Quantitative results are shown in Tab. 3. The results show that the leveraging pseudo labels improves the sampling diversity compared to EDM-2 in all cases. In three cases the diversity is even better than that of the real reference dataset proofing that the pseudo conditioning works well and we can specifically query the model to generate diverse data. It also improves FID scores compared to the unconditional trainings." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 56, + 448, + 294, + 521 + ], + "blocks": [ + { + "bbox": [ + 55, + 242, + 555, + 288 + ], + "lines": [ + { + "bbox": [ + 55, + 242, + 555, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 242, + 555, + 288 + ], + "type": "text", + "content": "Table 2. Image generation of ImageNet across different generative models. Numbers in brackets indicate results when sampling the model without conditioning. Inception is used as feature extractor for all metrics except for IRS where we use SwAV. Precision, recall, density, coverage and Vendi score had to be computed on a subset of the training data of 50000 and 10000 for Vendi score. Pretrained conditional models were used as unconditional models by setting conditional guidance to 0." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 448, + 294, + 521 + ], + "lines": [ + { + "bbox": [ + 56, + 448, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 448, + 294, + 521 + ], + "type": "table", + "html": "
FID ↓IRS∞,a ↑
EDM [24]DiADM (Ours)EDM [24]DiADM (Ours)
ImageNet-51251.5922.280.090.15
FFHQ40.926.240.231.51
ChestX-ray1424.296.760.191.08
CelebV-HQ68.4113.640.180.69
Dynamic13.825.560.601.04
", + "image_path": "4d2257386b6993efeb16d1851bb7037dce7e851a99c365b9da215a043efedcad.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 528, + 295, + 552 + ], + "lines": [ + { + "bbox": [ + 55, + 528, + 295, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 528, + 295, + 552 + ], + "type": "text", + "content": "Table 3. FID and IRS results with and without our proposed diversity awareness module" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 570, + 185, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 185, + 583 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 185, + 583 + ], + "type": "text", + "content": "4.4. Text-to-image example" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 587, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 587, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 54, + 587, + 295, + 696 + ], + "type": "text", + "content": "Extending IRS to text-conditional image generation is straightforward. The evaluation of diversity depends on the chosen perspective, which is defined by the reference datasets. By changing the reference dataset, we can address various diversity-related issues, such as fairness. One notable example is the gender bias observed in text-to-image models. For our experiments, we use Deepfloyd (IF-I-XLv1.0)" + }, + { + "bbox": [ + 54, + 587, + 295, + 696 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 54, + 587, + 295, + 696 + ], + "type": "text", + "content": ". To assess gender diversity, we create a reference dataset with a balanced gender distribution for specific job" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 307, + 555, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 307, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 307, + 555, + 403 + ], + "type": "text", + "content": "roles by prompting the model to generate 100 images each of males and females. Half of these images are reserved as a balanced test dataset. We then prompt the model without specifying gender. The results are presented in Sec. 4.2. All groups, including the general term human, show gender bias, with the latter generating only female images. IRS estimates that in all cases, the generated diversity reaches only about " + }, + { + "bbox": [ + 313, + 307, + 555, + 403 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 307, + 555, + 403 + ], + "type": "text", + "content": " of that of the balanced reference dataset." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 411, + 391, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 391, + 423 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 391, + 423 + ], + "type": "text", + "content": "4.5. Limitations" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 430, + 556, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 556, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 556, + 562 + ], + "type": "text", + "content": "The stochasticity involved for a low number of samples means that minor IRS differences are not meaningful for a low number of samples. However, this is also reflected in the high values of uncertainty. Additionally, we believe that finding a better feature extractor that maximizes measured diversity (unadjusted IRS) on real datasets should be the focus. Experiments on improving diversity are limited to using real features for generation and restrictive computational budget. Furthermore, increasing the conditioning requires similar experiments analyzing memorization issues as text-conditioning [6]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 573, + 454, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 573, + 454, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 573, + 454, + 586 + ], + "type": "text", + "content": "5. Conclusion and Outlook" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "In this paper, we reveal that all current methods for evaluating diversity and the feature extractors they are based on are inadequate. Using our proposed IRS metric, we demonstrate that these shortcomings result in all state-of-the-art image generation methods facing challenges with diversity. To address this issue for unconditional diffusion models, we propose DiADMs, which enhance the performance of current methods by separating diversity from fidelity. In future work, we aim to extend our efforts to improve diversity in conditional image generation as well." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 66, + 702, + 227, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 227, + 714 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 227, + 714 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 66, + 702, + 227, + 714 + ], + "type": "text", + "content": "https://huggingface.co/DeepFloyd/IF-I-XL-v1.0" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "3036" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 240 + ], + "type": "text", + "content": "Acknowledgements: This work was supported by the High-Tech-Agenda Bavaria. HPC resources were provided by the Erlangen National High Performance Computing Center (NHR@FAU) of the Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU) under the NHR project b143dc and b180dc. NHR funding is provided by federal and Bavarian state authorities. NHR@FAU hardware is partially funded by the German Research Foundation (DFG) - 440719683. Support was also received by the ERC - projects MIA-NORMAL 101083647 as well as DFG 513220538, 512819079 and DFG large scale infrastructure funding Art 91b GG. This work was supported by the UKRI Centre for Doctoral Training in Artificial Intelligence for Healthcare (EP/S023283/1) and Ultromics Ltd." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 253, + 115, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 115, + 266 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 115, + 266 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 273, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 61, + 273, + 296, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 273, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 61, + 273, + 296, + 316 + ], + "type": "text", + "content": "[1] Sumukh K Aithal, Pratyush Maini, Zachary C Lipton, and J Zico Kolter. Understanding hallucinations in diffusion models through mode interpolation. arXiv preprint arXiv:2406.09358, 2024. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 318, + 296, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 318, + 296, + 373 + ], + "spans": [ + { + "bbox": [ + 61, + 318, + 296, + 373 + ], + "type": "text", + "content": "[2] Sina Alemohammad, Josue Casco-Rodriguez, Lorenzo Luzi, Ahmed Imtiaz Humayun, Hossein Babaei, Daniel LeJeune, Ali Siahkoohi, and Richard G. Baraniuk. Self-Consuming Generative Models Go MAD, 2023. arXiv:2307.01850 [cs]. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 375, + 296, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 375, + 296, + 441 + ], + "spans": [ + { + "bbox": [ + 61, + 375, + 296, + 441 + ], + "type": "text", + "content": "[3] Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, and Michael Auli. data2vec: A General Framework for Self-supervised Learning in Speech, Vision and Language. In Proceedings of the 39th International Conference on Machine Learning, pages 1298-1312. PMLR, 2022. ISSN: 2640-3498. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 442, + 296, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 442, + 296, + 487 + ], + "spans": [ + { + "bbox": [ + 61, + 442, + 296, + 487 + ], + "type": "text", + "content": "[4] Hritik Bansal, Da Yin, Masoud Monajatipoor, and Kai-Wei Chang. How well can text-to-image generative models understand ethical natural language interventions? arXiv preprint arXiv:2210.15230, 2022. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 488, + 296, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 488, + 296, + 532 + ], + "spans": [ + { + "bbox": [ + 62, + 488, + 296, + 532 + ], + "type": "text", + "content": "[5] Leonard E. Baum and Patrick Billingsley. Asymptotic Distributions for the Coupon Collector's Problem. The Annals of Mathematical Statistics, 36(6):1835-1839, 1965. Publisher: Institute of Mathematical Statistics. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 534, + 296, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 534, + 296, + 578 + ], + "spans": [ + { + "bbox": [ + 62, + 534, + 296, + 578 + ], + "type": "text", + "content": "[6] Nicholas Carlini, Jamie Hayes, Milad Nasr, Matthew Jagielski, Vikash Sehwag, Florian Tramér, Borja Balle, Daphne Ippolito, and Eric Wallace. Extracting Training Data from Diffusion Models, 2023. arXiv:2301.13188 [cs]. 2, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 580, + 296, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 580, + 296, + 634 + ], + "spans": [ + { + "bbox": [ + 62, + 580, + 296, + 634 + ], + "type": "text", + "content": "[7] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised Learning of Visual Features by Contrasting Cluster Assignments. In Advances in Neural Information Processing Systems, pages 9912-9924. Curran Associates, Inc., 2020. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 636, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 636, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 636, + 296, + 713 + ], + "type": "text", + "content": "[8] Tingxiu Chen, Yilei Shi, Zixuan Zheng, Bingcong Yan, Jingliang Hu, Xiao Xiang Zhu, and Lichao Mou. Ultrasound Image-to-Video Synthesis via Latent Dynamic Diffusion Models. In Medical Image Computing and Computer Assisted Intervention - MICCAI 2024, pages 764-774. Springer Nature Switzerland, Cham, 2024. Series Title: Lecture Notes in Computer Science. 2" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 320, + 73, + 555, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 555, + 116 + ], + "type": "text", + "content": "[9] Ching-Yao Chuang, Varun Jampani, Yuanzhen Li, Antonio Torralba, and Stefanie Jegelka. Debiasing vision-language models via biased prompts. arXiv preprint arXiv:2302.00070, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 316, + 118, + 554, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 554, + 148 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 554, + 148 + ], + "type": "text", + "content": "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. 2015. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 151, + 554, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 151, + 554, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 151, + 554, + 184 + ], + "type": "text", + "content": "[11] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 185, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 553, + 217 + ], + "type": "text", + "content": "[12] Prafulla Dhariwal and Alex Nichol. Diffusion Models Beat GANs on Image Synthesis, 2021. arXiv:2105.05233 [cs, stat]. 3, 5, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 217, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 271 + ], + "type": "text", + "content": "[13] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. Cogview: Mastering text-to-image generation via transformers. Advances in neural information processing systems, 34:19822-19835, 2021. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 273, + 553, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 273, + 553, + 315 + ], + "spans": [ + { + "bbox": [ + 316, + 273, + 553, + 315 + ], + "type": "text", + "content": "[14] Mischa Dombrowski, Hadrien Reynaud, and Bernhard Kainz. Uncovering Hidden Subspaces in Video Diffusion Models Using Re-Identification, 2024. arXiv:2411.04956. 2, 7, 5, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 316, + 554, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 554, + 371 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 554, + 371 + ], + "type": "text", + "content": "[15] Chun-Mei Feng, Kai Yu, Yong Liu, Salman Khan, and Wangmeng Zuo. Diverse Data Augmentation with Diffusions for Effective Test-time Prompt Tuning. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 2704-2714, Paris, France, 2023. IEEE. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 372, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 553, + 403 + ], + "type": "text", + "content": "[16] Dan Friedman and Adji Bousso Dieng. The Vendi Score: A Diversity Evaluation Metric for Machine Learning, 2023. arXiv:2210.02410. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 404, + 554, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 404, + 554, + 449 + ], + "spans": [ + { + "bbox": [ + 316, + 404, + 554, + 449 + ], + "type": "text", + "content": "[17] Felix Friedrich, Manuel Brack, Lukas Struppek, Dominik Hintersdorf, Patrick Schramowski, Sasha Luccioni, and Kristian Kersting. Fair Diffusion: Instructing Text-to-Image Generation Models on Fairness, 2023. arXiv:2302.10893. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 449, + 554, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 449, + 554, + 525 + ], + "spans": [ + { + "bbox": [ + 317, + 449, + 554, + 525 + ], + "type": "text", + "content": "[18] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre H. Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, Bilal Piot, Koray Kavukcuoglu, Rémi Munos, and Michal Valko. Bootstrap your own latent: A new approach to self-supervised Learning, 2020. arXiv:2006.07733 [cs, stat]. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 526, + 554, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 526, + 554, + 580 + ], + "spans": [ + { + "bbox": [ + 317, + 526, + 554, + 580 + ], + "type": "text", + "content": "[19] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked Autoencoders Are Scalable Vision Learners. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15979-15988, New Orleans, LA, USA, 2022. IEEE. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 581, + 554, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 581, + 554, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 581, + 554, + 634 + ], + "type": "text", + "content": "[20] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 3, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 636, + 554, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 554, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 554, + 690 + ], + "type": "text", + "content": "[21] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 691, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 712 + ], + "type": "text", + "content": "[22] Jonathan Ho and Tim Salimans. Classifier-Free Diffusion Guidance. 2022. 5" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3037" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[23] Tero Karras, Samuli Laine, and Timo Aila. A Style-Based Generator Architecture for Generative Adversarial Networks. 2019. 5, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 152 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 152 + ], + "type": "text", + "content": "[24] Tero Karras, Miika Aittala, Jaakko Lehtinen, Janne Hellsten, Timo Aila, and Samuli Laine. Analyzing and Improving the Training Dynamics of Diffusion Models, 2024. arXiv:2312.02696 [cs, stat]. 5, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 208 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 208 + ], + "type": "text", + "content": "[25] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 294, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 241 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 241 + ], + "type": "text", + "content": "[26] Tuomas Kynkänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved Precision and Recall Metric for Assessing Generative Models. 2019. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 243, + 294, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 294, + 275 + ], + "type": "text", + "content": "[27] Tianhong Li, Yonglong Tian, He Li, Mingyang Deng, and Kaiming He. Autoregressive Image Generation without Vector Quantization, 2024. arXiv:2406.11838 [cs]. 3, 5, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 276, + 294, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 331 + ], + "type": "text", + "content": "[28] Jiawei Liu, Qiang Wang, Huijie Fan, Yinong Wang, Yandong Tang, and Liangqiong Qu. Residual denoising diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2773-2783, 2024. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 332, + 294, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 294, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 294, + 365 + ], + "type": "text", + "content": "[29] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A ConvNet for the 2020s, 2022. arXiv:2201.03545. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 366, + 294, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 366, + 294, + 422 + ], + "spans": [ + { + "bbox": [ + 56, + 366, + 294, + 422 + ], + "type": "text", + "content": "[30] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations, 2022. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 422, + 294, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 488 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 488 + ], + "type": "text", + "content": "[31] Zichen Miao, Jiang Wang, Ze Wang, Zhengyuan Yang, Lijuan Wang, Qiang Qiu, and Zicheng Liu. Training diffusion models towards diverse image generation with reinforcement learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10844-10853, 2024. 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 490, + 294, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 490, + 294, + 522 + ], + "spans": [ + { + "bbox": [ + 56, + 490, + 294, + 522 + ], + "type": "text", + "content": "[32] Muhammad Ferjad Naeem, Seong Joon Oh, Youngjung Uh, Yunjey Choi, and Jaejun Yoo. Reliable Fidelity and Diversity Metrics for Generative Models. 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 524, + 294, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 524, + 294, + 579 + ], + "spans": [ + { + "bbox": [ + 56, + 524, + 294, + 579 + ], + "type": "text", + "content": "[33] Jamie A. OaReilly and Fawad Asadi. Pre-trained vs. Random Weights for Calculating Fréchet Inception Distance in Medical Imaging. In 2021 13th Biomedical Engineering International Conference (BMEiCON), pages 1-4, Ayutthaya, Thailand, 2021. IEEE. 6, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 580, + 294, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 580, + 294, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 580, + 294, + 679 + ], + "type": "text", + "content": "[34] Maxime Oquab, Timothee Darct, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINoV2: Learning Robust Visual Features without Supervision, 2024. arXiv:2304.07193 [cs]. 6, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 680, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 714 + ], + "type": "text", + "content": "[35] David Ouyang, Bryan He, Amirata Ghorbani, Neal Yuan, Joseph Ebinger, Curtis P. Langlotz, Paul A. Heidenreich, Robert A. Harrington, David H. Liang, Euan A. Ashley, and" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "James Y. Zou. Video-based AI for beat-to-beat assessment of cardiac function. Nature, 580(7802):252-256, 2020. Publisher: Springer US. 6, 7, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "text", + "content": "[36] William Peebles and Saining Xie. Scalable Diffusion Models with Transformers. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 4172-4182, Paris, France, 2023. IEEE. 3, 5, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 150, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 150, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 150, + 553, + 194 + ], + "type": "text", + "content": "[37] Yiming Qin, Huangjie Zheng, Jiangchao Yao, Mingyuan Zhou, and Ya Zhang. Class-balancing diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18434-18443, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 194, + 553, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 194, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 316, + 194, + 553, + 259 + ], + "type": "text", + "content": "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning Transferable Visual Models From Natural Language Supervision, 2021. arXiv:2103.00020. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 260, + 553, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 260, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 316, + 260, + 553, + 326 + ], + "type": "text", + "content": "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 327, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 381 + ], + "type": "text", + "content": "[40] Hadrien Reynaud, Qingjie Meng, Mischa Dombrowski, Arijit Ghosh, Thomas Day, Alberto Gomez, Paul Leeson, and Bernhard Kainz. EchoNet-Synthetic: Privacy-preserving Video Generation for Safe Medical Data Sharing, 2024. arXiv:2406.00808 [cs]. 6, 7, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 382, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 382, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 382, + 553, + 415 + ], + "type": "text", + "content": "[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-Resolution Image Synthesis with Latent Diffusion Models. 2021. 3, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 415, + 553, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 553, + 470 + ], + "type": "text", + "content": "[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 471, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 471, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 471, + 553, + 514 + ], + "type": "text", + "content": "[43] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-Resolution Image Synthesis with Latent Diffusion Models, 2022. arXiv:2112.10752 [cs]. 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 514, + 553, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 514, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 316, + 514, + 553, + 557 + ], + "type": "text", + "content": "[44] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. DreamBooth: Fine Tuning Text-to-Image Diffusion Models for Subject-Driven Generation. 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 558, + 553, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 558, + 553, + 592 + ], + "spans": [ + { + "bbox": [ + 316, + 558, + 553, + 592 + ], + "type": "text", + "content": "[45] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, Xi Chen, and Xi Chen. Improved Techniques for Training GANs. 2016. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 592, + 553, + 625 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 592, + 553, + 625 + ], + "spans": [ + { + "bbox": [ + 316, + 592, + 553, + 625 + ], + "type": "text", + "content": "[46] Axel Sauer, Katja Schwarz, and Andreas Geiger. StyleGANXL: Scaling StyleGAN to Large Diverse Datasets, 2022. arXiv:2202.00273 [cs]. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 625, + 553, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 680 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 680 + ], + "type": "text", + "content": "[47] Vikash Sehwag, Caner Hazirbas, Albert Gordo, First Ozgenel, and Cristian Canton. Generating high fidelity data from low-density regions using diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11492-11501, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[48] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3038" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 609 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 139 + ], + "type": "text", + "content": "[49] George Stein, Jesse C. Cresswell, Rasa Hosseinzadeh, Yi Sui, Brendan Leigh Ross, Valentin Villecloze, Zhaoyan Liu, Anthony L. Caterini, J. Eric T. Taylor, and Gabriel Loaiza-Ganem. Exposing flaws of generative model evaluation metrics and their unfair treatment of diffusion models, 2023. arXiv:2306.04675 [cs, stat]. 2, 5, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 195 + ], + "type": "text", + "content": "[50] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the Inception Architecture for Computer Vision. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2818-2826, Las Vegas, NV, USA, 2016. IEEE. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 197, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 217 + ], + "type": "text", + "content": "[51] N. M. Temme. Asymptotic Estimates of Stirling Numbers. Studies in Applied Mathematics, 89(3):233-243, 1993. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 294, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 294, + 272 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 294, + 272 + ], + "type": "text", + "content": "[52] Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, Mohammadhadi Bagheri, and Ronald M Summers. ChestX-ray8: Hospital-Scale Chest X-Ray Database and Benchmarks on Weakly-Supervised Classification and Localization of Common Thorax Diseases. 2017. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 274, + 294, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 274, + 294, + 328 + ], + "spans": [ + { + "bbox": [ + 56, + 274, + 294, + 328 + ], + "type": "text", + "content": "[53] Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang. Diffusion models: A comprehensive survey of methods and applications. ACM Computing Surveys, 56(4): 1-39, 2023. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 331, + 294, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 294, + 384 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 294, + 384 + ], + "type": "text", + "content": "[54] Ling Yang, Jingwei Liu, Shenda Hong, Zhilong Zhang, Zhilin Huang, Zheming Cai, Wentao Zhang, and Bin Cui. Improving diffusion-based image synthesis with context prediction. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 386, + 294, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 386, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 56, + 386, + 294, + 441 + ], + "type": "text", + "content": "[55] Cheng Zhang, Xuanbai Chen, Siqi Chai, Chen Henry Wu, Dmitry Lagun, Thabo Beeler, and Fernando De la Torre. Itigen: Inclusive text-to-image generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3969-3980, 2023. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 442, + 294, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 442, + 294, + 485 + ], + "spans": [ + { + "bbox": [ + 56, + 442, + 294, + 485 + ], + "type": "text", + "content": "[56] Weitong Zhang, Chengqi Zang, Liu Li, Sarah Cechnicka, Cheng Ouyang, and Bernhard Kainz. Stability and generalizability in sde diffusion models with measure-preserving dynamics. In NeurIPS, 2024. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 487, + 294, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 487, + 294, + 529 + ], + "spans": [ + { + "bbox": [ + 56, + 487, + 294, + 529 + ], + "type": "text", + "content": "[57] Yi Zhang, Xiaoyu Shi, Dasong Li, Xiaogang Wang, Jian Wang, and Hongsheng Li. A unified conditional framework for diffusion-based image restoration. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 532, + 294, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 532, + 294, + 563 + ], + "spans": [ + { + "bbox": [ + 56, + 532, + 294, + 563 + ], + "type": "text", + "content": "[58] Dora Zhao, Jerone T. A. Andrews, Orestis Papakyriakopoulos, and Alice Xiang. Position: Measure Dataset Diversity, Don't Just Claim It, 2024. arXiv:2407.08188. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 566, + 294, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 566, + 294, + 609 + ], + "spans": [ + { + "bbox": [ + 56, + 566, + 294, + 609 + ], + "type": "text", + "content": "[59] Hao Zhu, Wayne Wu, Wentao Zhu, Liming Jiang, Siwei Tang, Li Zhang, Ziwei Liu, and Chen Change Loy. CelebV-HQ: A Large-Scale Video Facial Attributes Dataset, 2022. arXiv:2207.12393 [cs]. 6" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3039" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Image Over Text_ Transforming Formula Recognition Evaluation with Character Detection Matching/e8290078-d6a1-49c9-b28d-cb4de981b7cd_content_list.json b/2025/Image Over Text_ Transforming Formula Recognition Evaluation with Character Detection Matching/e8290078-d6a1-49c9-b28d-cb4de981b7cd_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b35874e415c59a5eac2c421d1a2ae9b7385daae5 --- /dev/null +++ b/2025/Image Over Text_ Transforming Formula Recognition Evaluation with Character Detection Matching/e8290078-d6a1-49c9-b28d-cb4de981b7cd_content_list.json @@ -0,0 +1,1996 @@ +[ + { + "type": "text", + "text": "Image Over Text: Transforming Formula Recognition Evaluation with Character Detection Matching", + "text_level": 1, + "bbox": [ + 143, + 128, + 854, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Bin Wang $^{1*}$", + "bbox": [ + 215, + 202, + 310, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rui Zhang", + "bbox": [ + 184, + 220, + 277, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fan $\\mathbf{W}\\mathbf{u}^{1*}$", + "bbox": [ + 339, + 203, + 419, + 219 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Renqiu Xia1,2", + "bbox": [ + 315, + 220, + 424, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Linke Ouyang $^{1*}$", + "bbox": [ + 442, + 203, + 571, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Botian $\\mathbf{Shi}^1$", + "bbox": [ + 455, + 220, + 547, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhuangcheng Gu", + "bbox": [ + 601, + 203, + 745, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Bo Zhang", + "bbox": [ + 580, + 220, + 666, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Conghui $\\mathrm{He}^{1\\dagger}$", + "bbox": [ + 699, + 220, + 810, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Shanghai Artificial Intelligence Laboratory $^{2}$ Shanghai Jiao Tong University", + "bbox": [ + 181, + 255, + 815, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 308, + 326, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Formula recognition presents significant challenges due to the complicated structure and varied notation of mathematical expressions. Despite continuous advancements in formula recognition models, the evaluation metrics employed by these models, such as BLEU and Edit Distance, still exhibit notable limitations. They overlook the fact that the same formula has diverse representations and is highly sensitive to the distribution of training data, thereby causing unfairness in formula recognition evaluation. To this end, we propose a Character Detection Matching (CDM) metric, ensuring the evaluation objectivity by designing an image-level rather than a LaTeX-level metric score. Specifically, CDM renders both the model-predicted LaTeX and the ground-truth LaTeX formulas into image-formatted formulas, then employs visual feature extraction and localization techniques for precise character-level matching, incorporating spatial position information. Such a spatially-aware and character-matching method offers a more accurate and equitable evaluation compared with previous BLEU and Edit Distance metrics that rely solely on text-based character matching. Experimentally, we evaluated various formula recognition models using CDM, BLEU, and ExpRate metrics. Their results demonstrate that the CDM aligns more closely with human evaluation standards and provides a fairer comparison across different models by eliminating discrepancies caused by diverse formula representations. Code is available at https://github.com/opendatalab/UniMERNet/tree/main/cdm", + "bbox": [ + 89, + 340, + 483, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 806, + 220, + 821 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mathematical formula recognition is crucial in document analysis as it directly impacts the scientific rigor and accuracy of the document content [7, 19, 31, 37]. Unlike", + "bbox": [ + 89, + 832, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "standard Optical Character Recognition (OCR) technique, formula recognition presents unique challenges. Formulas often encompass multi-level symbols, subscripts, fractions, and other complicated structures, requiring models to comprehend spatial and structural relationships rather than just linear, sequential text [14, 19]. Besides, formulas exhibit representational diversity, meaning that the same formula can be expressed in multiple valid ways.", + "bbox": [ + 511, + 310, + 906, + 431 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, significant advancements in formula recognition [4, 8, 20, 28, 39, 41, 42] have been primarily driven by the continuous development of artificial intelligence technique [5, 35, 40]. Besides, commercial formula recognition software like Mathpix1 and the recently proposed UniMERNet [30] model have achieved impressive results in diverse real-world settings. Despite these advancements, the existing evaluation metrics [17, 27] for formula recognition still face some challenges. Commonly-used metrics such as BLEU [27] and Edit Distance [17] primarily rely on text-based character matching, which introduces several limitations as follows:", + "bbox": [ + 511, + 431, + 908, + 613 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(1) Low Metric Reliability. BLEU and Edit Distance are reliable for evaluating the quality of text-level similarity. However, the diversity in formula representations makes these text-level evaluation metrics inadequate for precisely reflecting formula recognition quality. For example, as shown in Figure 1 (Case 1), a model's prediction might render an image identical to the ground truth formula. However, due to the variations in formula expression styles, the evaluation results obtained using the ExpRate [8], BLEU, and Edit Distance may be somewhat misleading.", + "(2) Unfair Model Comparison. Current metrics may be susceptible to discrepancies between the distributions of training and testing data. As illustrated in Figure 1 (Case 1 and Case 2), a model may produce a correct prediction but score poorly due to representational differences from the ground truth, while an incorrect prediction might score higher if its representation aligns more closely with the test" + ], + "bbox": [ + 511, + 614, + 908, + 873 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution. †Corresponding author: heconghui@pjlab.org.cn", + "bbox": [ + 107, + 886, + 482, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "r={\\frac{1}{2}{\\alpha}{\\beta}}|\\sin\\beta\\left( \\right)\\sigma\\sigma\\sigma-{2}\\sigma_{1}\\left( \\right)1", + "bbox": [ + 119, + 212, + 316, + 271 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2. Element Region Localization", + "bbox": [ + 117, + 282, + 271, + 294 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{c c} \\frac {\\text {r e n d e r e a c h}}{\\text {t o k e n w i t h a}} & r = \\frac {\\alpha}{\\beta} \\big | \\sin \\beta \\left(\\sigma_ {2} + \\sigma_ {1}\\right) \\big | \\\\ \\text {d i f f e r e n t c o l o r} & \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 306, + 323, + 334 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{c} r = \\frac {\\alpha}{\\beta} | \\sin \\beta (\\sigma_ {2} + \\sigma_ {1}) | \\\\ \\text {e x t r a c t e a c h} \\\\ \\text {c o l o r a n d l o c a t e} \\\\ \\text {t h e b o u n d i n g b o x} \\end{array} r = \\frac {\\alpha}{\\beta} | \\sin \\beta (\\sigma_ {2} + \\sigma_ {1}) |\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 339, + 323, + 387 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/476845664a3358db013068b11bce6a22519ceb0eaea91ffb09aed9e18eb92e3e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 120, + 401, + 310, + 470 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fef9e67c4918b4b39f8cf2ed6809bb9b033dad3e1bb7223a4568b9655df9b2db.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 356, + 148, + 464, + 185 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/309337165261ce94e6e8adc3a54c0274b598009d569e43604e9e0ce91d3480b6.jpg", + "image_caption": [ + "Step 3: Invalid Match Elimination" + ], + "image_footnote": [], + "bbox": [ + 356, + 200, + 465, + 237 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a6bed2a0a6ca774c3e62edf7709e80e9b09f3360c8ef2bd0145a54fd5192eaff.jpg", + "image_caption": [ + "Step 2: Element Region Matching", + "1. Calculate Match Cost Between Every Element Pairs", + "Step 4: Metric Calculation" + ], + "image_footnote": [], + "bbox": [ + 467, + 151, + 689, + 239 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/44520171490c921152c241eb9a56460ee74d0c32aa01941c7d3bb3a0e0ec2288.jpg", + "image_caption": [ + "2. Find the Best Match by Hungarian Algorithm", + "Evaluation Metric" + ], + "image_footnote": [], + "bbox": [ + 661, + 151, + 859, + 236 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2e6767b0c3ac6e1015c28c40261f78ef5437b67a707c7dff6097d8061027362d.jpg", + "image_caption": [ + "1. Token Consistency Check" + ], + "image_footnote": [], + "bbox": [ + 369, + 292, + 570, + 376 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9b656f304f8b314129cfa7edfe5f8d6c233268eba74e037704e963b8acfa559e.jpg", + "image_caption": [ + "2. Position Relationship Consistency Check", + "Figure 2. Overview of the Character Detection Matching (CDM), consisting of four main stages. (1) Element Localization, where bounding boxes of individual elements are extracted. (2) Element Region Matching, which employs a bipartite graph matching method to pair prediction with ground truth elements. (3) Invalid Match Elimination, where inconsistent matches are discarded through token and positional relationship checks. (4) Metric Calculation, where matching accuracy is evaluated using the F1-Score and ExpRate@CDM." + ], + "image_footnote": [], + "bbox": [ + 370, + 398, + 570, + 479 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF 1 \\text {s c o r e} = \\frac {2 T P}{2 T P + F P + F N}\n$$\n", + "text_format": "latex", + "bbox": [ + 704, + 306, + 816, + 325 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- TP: number of matched bbox pairs", + "- FP: number of unmatched bboxes in prediction", + "FN: number of unmatched bboxes in GT" + ], + "bbox": [ + 681, + 335, + 843, + 363 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nT P = \\left(\\bigtriangledown\\right) _ {n u m} = 1 3\n$$\n", + "text_format": "latex", + "bbox": [ + 651, + 383, + 741, + 425 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF P = \\left(\\square\\right) _ {\\text {n u m}} ^ {\\text {P r e d}} = 3\n$$\n", + "text_format": "latex", + "bbox": [ + 774, + 383, + 859, + 402 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF N = \\left( \\begin{array}{l l} \\square & T \\\\ 0 & n u m \\end{array} \\right) ^ {G T} = 3\n$$\n", + "text_format": "latex", + "bbox": [ + 774, + 415, + 859, + 433 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF 1 s c o r e = \\frac {2 \\times 1 3}{2 \\times 1 3 + 3 + 3} = 0. 8 1 2 5\n$$\n", + "text_format": "latex", + "bbox": [ + 676, + 443, + 831, + 467 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "of formula recognition across diverse representations.", + "bbox": [ + 89, + 566, + 444, + 583 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Character Detection Matching", + "text_level": 1, + "bbox": [ + 89, + 599, + 370, + 618 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Due to the diversity of LaTeX expressions, text-based character-matching methods are unreliable for formula recognition evaluation. The basic idea of CDM is to compare the rendered images from LaTex text. If the image rendered from the predicted LaTeX source code matches the image rendered from the ground truth LaTeX source code, the formula is considered entirely correct. However, directly comparing the pixel values of the original and predicted formulas is not ideal. Any error or extra/missing character in the prediction can cause subsequent characters to be mismatched. Additionally, two similar formulas might have different layouts, with one being a single-line formula and the other a multi-line formula due to line breaks. Therefore, a more robust algorithm is needed to calculate the match between the predicted result and the ground truth image.", + "bbox": [ + 89, + 627, + 482, + 867 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To this end, we propose a metric that incorporates a bipartite matching step for element-level matching in images,", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "providing a more intuitive assessment. As shown in Figure 2, the algorithm consists of four stages as follows.", + "bbox": [ + 511, + 566, + 903, + 598 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Element Localization", + "text_level": 1, + "bbox": [ + 511, + 606, + 712, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "First, the bounding boxes (bboxes) of each individual element in the rendered image are extracted, followed by the subsequent steps:", + "bbox": [ + 511, + 628, + 903, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "LaTeX Source Normalization. LaTeX source codes of both the ground truth and predicted formulas are normalized, breaking them down into individual tokens such as \"2\", \"a\", \"A\", \"\\alpha\", \"\\sin\". Composite elements are decomposed into individual characters, e.g., \"\\frac{ab}{\\alpha} is decomposed into \\frac{a}{b}\\)\".", + "bbox": [ + 511, + 674, + 905, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Element Region Localization. To accurately detect character positions, we render each token in a unique color. We construct an RGB color list with a fixed interval of 15, creating a sequence from $(0,0,15)$ to $(255,255,255)$ . This generates $(255 / 15 + 1)^{3} = 5832$ distinct colors, sufficient for complex formulas. Each token is assigned a unique color using the LaTeX command \"\\mathsf{\\\\mathsf{m}}\\mathsf{a}\\mathsf{h}\\mathsf{o}\\mathsf{c}\\mathsf{l}\\mathsf{r}\\mathsf{R}\\mathsf{B}\\mathsf{]} \\{\\mathsf{r},\\mathsf{g},\\mathsf{b}\\} \"$. After rendering, we apply image processing techniques to identify pixels match-", + "bbox": [ + 511, + 763, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19684", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ing each color, thereby determining the exact coordinates of each token.", + "bbox": [ + 89, + 90, + 485, + 122 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Element Region Matching", + "text_level": 1, + "bbox": [ + 89, + 128, + 330, + 146 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this stage, a bipartite matching method pairs the predicted elements with the corresponding ground truth elements. Based on the element localization, two sets are obtained for each formula: one for the ground truth independent elements $y$ and one for the predicted independent elements $\\hat{y}$ . The number of independent elements in each set is $N_y$ and $N_{\\hat{y}}$ , respectively, with $N = \\min(N_y, N_{\\hat{y}})$ being the number of elements in the smaller set.", + "bbox": [ + 89, + 150, + 483, + 271 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To measure the similarity between $y$ and $\\hat{y}$ , we match elements in the two sets by identifying the corresponding ground truth element for each predicted element. We use the bipartite matching Hungarian algorithm [13], as described in DETR [6], to find a permutation $\\hat{\\sigma}$ that minimizes the total matching cost:", + "bbox": [ + 89, + 272, + 483, + 363 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\sigma} = \\arg \\min _ {\\sigma \\in S _ {N}} \\sum_ {i = 1} ^ {N} L _ {\\text {m a t c h}} \\left(y _ {i}, \\hat {y} _ {\\sigma (i)}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 369, + 483, + 411 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {m a t c h}} = W _ {t} \\times L _ {t} + W _ {p} \\times L _ {p} + W _ {o} \\times L _ {o}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 422, + 482, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the matching cost $L_{match}$ is defined as a weighted sum of three components as introduced as follows:", + "bbox": [ + 89, + 443, + 483, + 472 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- Token Matching Cost $L_{t}$ : This component measures whether the tokens corresponding to two bounding boxes are the same. If they are identical, the cost is 0; if they are different, the cost is 1. For tokens that render identically but are different, such as \" (\" , \"\\leftarrow\" , and \"\\bigig( \", the cost is 0.05 , which can be formulated as follows:", + "bbox": [ + 89, + 474, + 483, + 577 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {t}} = \\left\\{ \\begin{array}{l l} 0, & \\text {i f} t _ {i} = \\hat {t} _ {\\sigma (i)}; \\\\ 0. 0 5, & \\text {i f} t _ {i} \\approx \\hat {t} _ {\\sigma (i)}; \\\\ 1, & \\text {o t h e r w i s e}; \\end{array} \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 199, + 577, + 482, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\approx$ denotes tokens that differ but render identically.", + "bbox": [ + 104, + 636, + 480, + 650 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- Positional Proximity Cost $L_{p}$ : This component measures the proximity of the two bounding boxes' positions using the L1 norm of their coordinates, which can be formulated as follows:", + "bbox": [ + 91, + 651, + 482, + 710 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {p}} = \\frac {1}{D _ {b}} \\times \\| b _ {i} - \\hat {b} _ {\\sigma (i)} \\| _ {1}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 202, + 714, + 482, + 746 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $b = [x_1, y_1, x_2, y_2]$ , and $D_b$ is the dimension of the bounding box coordinates.", + "bbox": [ + 102, + 751, + 482, + 781 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- Order Similarity Cost $L_{o}$ : This measures the similarity of the token order in the original LaTeX source (an approximation of reading order). The order is normalized to the range [0, 1], and the L1 norm can be calculated as follows:", + "bbox": [ + 91, + 781, + 483, + 854 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {o}} = \\frac {1}{D _ {o}} \\times \\left\\| o _ {i} - \\hat {o} _ {\\sigma (i)} \\right\\| _ {1}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 202, + 853, + 482, + 883 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $o_i$ means order of $i$ -th token, with dimension $D_o$ .", + "bbox": [ + 104, + 885, + 475, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Overall, the weights $W_{t}, W_{p}, W_{o}$ are used to balance the contributions of the three components. By employing this comprehensive matching strategy, we ensure a more accurate and robust evaluation of the correspondence between the predicted and ground truth elements, thereby improving the overall assessment of formula recognition quality.", + "bbox": [ + 511, + 90, + 905, + 183 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Invalid Match Elimination", + "text_level": 1, + "bbox": [ + 511, + 191, + 756, + 207 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After pairing the individual elements of the predicted result with the ground truth using the Hungarian matching algorithm, we need to verify these pairs and eliminate invalid matches. This process involves two steps:", + "bbox": [ + 511, + 214, + 905, + 273 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Token Consistency Check. Check whether the elements in each matched pair are consistent in terms of characters. If they are inconsistent, discard the match.", + "bbox": [ + 511, + 275, + 905, + 320 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Position Relationship Consistency Check. The relative positions of elements in mathematical formulas are crucial. For instance, in the expressions $2^3$ and $3^2$ , bipartite matching might pair 2 with 2 and 3 with 3, but their meanings and visual representations are entirely different. Thus, we need to check the consistency of the positional relationships within the matched pairs. We treat each element in the matched pair as a bounding box and analyze their relative positions. Specifically, we assume an affine transformation between the ground truth and predicted elements:", + "bbox": [ + 511, + 321, + 905, + 470 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {b} _ {\\sigma (i)} = \\mathbf {A} \\left(b _ {i}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 656, + 482, + 905, + 502 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{A}$ is the affine transformation matrix. To identify inconsistent match pairs, we detect pairs that do not conform to this transformation relationship. We employ the RANSAC algorithm [9] for this purpose. RANSAC can determine the optimal transformation matrix $\\mathbf{A}$ in the presence of noise. Given that formulas are usually horizontally arranged during rendering, we fix the rotation angle in the transformation matrix to 0, considering only translation and scaling. This approach not only improves the convergence speed of the RANSAC algorithm but also enhances the final matching accuracy.", + "bbox": [ + 511, + 512, + 905, + 679 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To account for line-breaking effects in formulas, we perform multiple rounds of RANSAC iterations to ensure that as many matched pairs as possible conform to the transformation relationship. After several iterations, matched pairs that still do not conform to the transformation relationship are considered incorrect and are eliminated.", + "bbox": [ + 511, + 679, + 905, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The above two steps effectively eliminate invalid match pairs, ensuring more accurate final matching results.", + "bbox": [ + 511, + 770, + 905, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Metric Calculation", + "text_level": 1, + "bbox": [ + 511, + 810, + 694, + 825 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use the F1-Score as the default metric for evaluating CDM (Character Detection Metric), defined as:", + "bbox": [ + 511, + 833, + 905, + 863 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nC D M = \\frac {2 \\times T P}{2 \\times T P + F P + F N}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 873, + 903, + 905 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19685", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d8ff569f6b6f08e795bdf71a951aa6d5af2d34f01a954ab7c802c34bf22c3cb8.jpg", + "image_caption": [ + "(a) CDM distribution on UniMER-Test dataset" + ], + "image_footnote": [], + "bbox": [ + 169, + 78, + 537, + 277 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/acfb733409878c8353891713f6b8d6cbda9bd3dfabcd24eddb073964b56ddf48.jpg", + "image_caption": [ + "(b) Examples by CDM score range", + "Figure 3. CDM Score Distribution and Example Analysis. (a) CDM score distribution for Mathpix, UniMERNet, Texify, and Pix2Tex on the UniMER-Test dataset. (b) Example analysis of Pix2Tex predictions across different CDM score ranges." + ], + "image_footnote": [], + "bbox": [ + 540, + 77, + 826, + 277 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $TP$ denotes true positives, $FP$ denotes false positives, and $FN$ denotes false negatives.", + "bbox": [ + 89, + 349, + 482, + 380 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To further evaluate the accuracy of formula recognition, we introduce the ExpRate@CDM metric, defined as:", + "bbox": [ + 89, + 380, + 482, + 411 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {E x p R a t e} @ C D M = \\frac {\\sum_ {i = 1} ^ {N} \\mathbb {I} \\left(C D M _ {i} = 1\\right)}{N}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 422, + 482, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbb{I}$ is the indicator function that equals 1 if $CDM_{i} = 1$ and 0 otherwise, and $N$ is the total number of formulas. This metric represents the proportion of formulas for which the model's prediction results are perfectly matched. Essentially, ExpRate@CDM serves as a precise version of the ExpRate metric specifically for formula recognition.", + "bbox": [ + 89, + 465, + 483, + 556 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 580, + 223, + 597 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Models and Data", + "text_level": 1, + "bbox": [ + 89, + 604, + 258, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We validate the CDM metric by evaluating several mainstream formula recognition models using both subjective impressions and objective metrics. The models include open-source UniMERNet [30], Texify [28], Pix2tex [4], and the commercial Mathpix API, all tested on the UniMERTest dataset. Besides, we evaluate document-level models, such as the open-source Nougat [5] and the commercial GPT-4o [10]. Vary [34] and StrucTexTv3 [21] are excluded as they are currently unavailable.", + "bbox": [ + 89, + 628, + 482, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "UniMER-Test Dataset. The dataset includes 23,757 formula samples, categorized into Simple Printed Expressions (SPE), Complex Printed Expressions (CPE), Screenshot Expressions (SCE), and Handwritten Expressions (HWE). We use these categories to conduct the model evaluation.", + "bbox": [ + 89, + 763, + 482, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tiny-Doc-Math Dataset. To evaluate document-level recognition, we construct the Tiny-Doc-Math dataset, consisting of arXiv papers in mathematics and computer science, published after June 2024, to ensure that they are not", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "in the training data of the compared models. We obtain Latex code and corresponding PDFs, match displayed equations using regular expressions, and manually verify them. Overall, the dataset includes 12 PDFs, totaling 196 pages and 437 formulas.", + "bbox": [ + 511, + 349, + 906, + 424 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This validation set includes both formula-level and document-level evaluations:", + "bbox": [ + 511, + 426, + 906, + 455 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Formula-level: Using single rendered formula images as input, we evaluate Mathpix, Pix2Tex, and UniMERNet. These models accept cropped formula images as input, and we compare the model outputs with the ground truth to compute relevant metrics.", + "- Document-level: Using PDFs or images as input, we evaluate Nougat, GPT-4o, and Mathpix, which can convert entire PDF pages into Markdown format. We match the displayed equations in the model outputs using regular expressions and compare them with the ground truth LaTeX formulas to compute relevant metrics." + ], + "bbox": [ + 513, + 455, + 905, + 623 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Credibility Assessment of CDM", + "text_level": 1, + "bbox": [ + 511, + 633, + 792, + 648 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2.1. Rendering Success Rate", + "text_level": 1, + "bbox": [ + 511, + 656, + 725, + 671 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The CDM metric relies on the successful rendering of formula images. For models that fail to render images, we assign a CDM score of 0, as rendering failures indicate that the predicted LaTeX code lacks critical elements. The rendering success rates on the UniMER-Test dataset for Pix2tex, Texify, UniMERNet, and Mathpix are $96.63\\%$ , $94.77\\%$ , $99.71\\%$ , and $97.82\\%$ , respectively, ensuring the applicability and reliability of the CDM metric.", + "bbox": [ + 511, + 675, + 906, + 796 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2.2. User Preference Evaluation", + "text_level": 1, + "bbox": [ + 511, + 806, + 751, + 819 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We analyze the distribution of CDM scores for four models on the UniMER-Test dataset. As shown in Figure 3a, Math-pix and UniMERNet perform well in terms of CDM scores. We conduct a detailed analysis of the Pix2Tex model by randomly selecting samples from different score ranges to", + "bbox": [ + 511, + 824, + 905, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19686", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bd60bd570cb6355ae0a2f0e3e187fd7db70f4f52099ef7aab3291bafd9d71a37.jpg", + "image_caption": [ + "(a) Human Evaluation Preferences" + ], + "image_footnote": [], + "bbox": [ + 107, + 87, + 321, + 262 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2e973c968c75f8b1b534c2d12b568ff1507ad9f36f14a1a99310c9ce6eba7abb.jpg", + "image_caption": [ + "(b) BLEU vs.CDM Score Distribution" + ], + "image_footnote": [], + "bbox": [ + 328, + 88, + 629, + 262 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e69b68ade6a5bf1d4311efebde234b734e4b690aca3b0a6b464e40808d14460d.jpg", + "image_caption": [ + "(c) Impact of Formula Writing Styles" + ], + "image_footnote": [], + "bbox": [ + 635, + 89, + 887, + 261 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7314099a976ec80cb10c028ef7556bb1382df3225a593df61b5979804041de0f.jpg", + "image_caption": [ + "Figure 4. Human evaluation and metric comparison. (a) Distribution of human preferences. (b) Distribution of cases preferring BLEU or CDM across different score ranges. (c) Impact of formula writing styles on BLEU, ExpRate and CDM scores.", + "Figure 5. Annotation interface for evaluating user preferences between CDM and BLEU scores. Annotators chose from four options: \"Score A is better\", \"Score B is better\", \"Both scores are equally good\", or \"Neither score is good\"." + ], + "image_footnote": [], + "bbox": [ + 102, + 334, + 472, + 534 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "evaluate if the prediction quality corresponds to the CDM scores. The analysis in Figure 3b shows that the CDM scores effectively reflect formula quality, with higher scores indicating fewer errors.", + "bbox": [ + 89, + 625, + 482, + 684 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To verify the consistency between the CDM metric and human evaluation, we select 1,008 CDM scores from Pix2Tex predictions, ensuring a balanced score distribution. We design an annotation interface (shown in Figure 5), displaying a ground truth label and the corresponding predicted LaTeX rendered image. Annotators choose between ScoreA, ScoreB, Both (credible), and Neither (credible). ScoreA and ScoreB correspond to the BLEU and CDM scores, respectively, but their order is randomized. For more details, please refer to the supplementary materials.", + "bbox": [ + 89, + 686, + 482, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The results in Figure 4a show that $64\\%$ of participants prefer the CDM metric, and $32\\%$ consider both metrics good. This indicates a $96\\%$ consistency between the CDM metric and human evaluation, demonstrating its reliability.", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6359a83996b672aaa760963763cb4d4e27b70d1d7049cd5c1aa25a22ac1cd655.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelExpRateExpRate@CDMBLEUCDM
Pix2tex0.12370.29100.40800.6360
Texify0.22880.49500.58900.7550
Matphix0.26100.50000.80670.9510
UniMERNet0.47990.81100.84250.9680
", + "bbox": [ + 517, + 333, + 911, + 414 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. UniMER-Test evaluation results. Comparison of mainstream models using different metrics.", + "bbox": [ + 511, + 424, + 906, + 452 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 4b compares the number of cases where BLEU or CDM is preferred across different score ranges, showing that CDM consistently outperforms BLEU.", + "bbox": [ + 511, + 457, + 906, + 503 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2.3. Objective Stability Assessment", + "text_level": 1, + "bbox": [ + 511, + 508, + 772, + 523 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To evaluate the impact of formula writing styles on the CDM and BLEU metrics, we randomly select 50 formulas with LaTeX source code and rewrite each formula five times using GPT-4, generating 250 additional formulas. We manually verify these formulas to ensure their rendered results are identical to the original 50 formulas. Using the initial LaTeX source code as the ground truth, we analyze the score distribution of the BLEU and CDM metrics. As shown in Figure 4c, the CDM metric is unaffected by style changes, with all samples scoring 1. In contrast, the BLEU metric's scores are dispersed, making it unsuitable for formula evaluation. The CDM metric remains robust and reliable despite formatting changes.", + "bbox": [ + 511, + 527, + 906, + 724 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Evaluation of Mainstream Models", + "text_level": 1, + "bbox": [ + 511, + 732, + 810, + 747 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct a detailed evaluation of mainstream models using both the CDM and BLEU metric. Note that all BLEU metric in this paper have been normalized [4, 8]. However, as discussed in the limitation section, normalization operations cannot address all issues, which will be evident in the following experiments.", + "bbox": [ + 511, + 753, + 905, + 845 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3.1. UniMER-Test Evaluation", + "text_level": 1, + "bbox": [ + 511, + 851, + 736, + 864 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Table 1, the evaluation results of the four models on the UniMER-Test dataset indicate that the quality", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19687", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/639a37c515982557d6ee702a836a2d0075b51ab004419d8796129e2ff8fd5c63.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodSPECPEHWESCE
BLEU ↑CDM ↑BLEU ↑CDM ↑BLEU ↑CDM ↑BLEU ↑CDM ↑
Pix2tex [4]0.87300.96190.65500.64890.01200.24530.09200.6762
Texify [28]0.90600.98520.69000.70410.34100.52690.42000.7932
Mathpix0.79200.97290.80610.96710.80600.93180.81820.9238
UniMERNet [30]0.91700.99460.91600.97070.92100.95300.61600.9461
", + "bbox": [ + 189, + 78, + 810, + 186 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4470f1eb323c60d3e0a121c2ff66fd41fe85108446e59dce47aab77bb7559e4a.jpg", + "table_caption": [ + "Table 2. UniMER-Test subset evaluation results. Analysis of anomalies in BLEU and CDM metrics across different subsets." + ], + "table_footnote": [], + "table_body": "
Image TypeModelBLEUCDMExpRate@CDM
FormulaPix2tex0.46480.744400.3684
GPT-4o0.64310.73300.4324
UniMERNet0.60560.93960.6887
Mathpix0.61120.94800.2105
DocumentGPT-4o0.34110.65020.1670
Nougat0.58970.83260.6086
Mathpix0.59390.95670.6292
", + "bbox": [ + 94, + 229, + 498, + 351 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Tiny-Doc-Math Evaluation Results. Performance of mainstream models on cropped formula inputs and document-level screenshots using BLEU and CDM metrics.", + "bbox": [ + 89, + 361, + 485, + 405 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ranges from low to high as follows: Pix2Tex, Texify, Mathpix, and UniMERNet, based on both BLEU and CDM metrics. ExpRate@CDM clearly shows the proportion of completely correct predictions for each model, indicating that the text character-based ExpRate is unreliable.", + "bbox": [ + 89, + 438, + 483, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "From the results in Table 1, it appears that the trends of the BLEU and CDM metrics are consistent. To verify the reliability of using the BLEU metric for model comparison, we further present evaluation results on the UniMERTest subsets. As shown in Table 2, we observe two notable anomalies: Firstly, in the SCE subset, when comparing the quality of the Mathpix and UniMERNet models, the BLEU and CDM metrics provide opposite conclusions. A detailed review of the UniMERNet paper reveals that the SCE subset was annotated based on Mathpix and then manually corrected. This means that the expression style of the SCE formulas is more consistent with Mathpix. Consequently, even though the CDM metric indicates that UniMERNet has better actual model quality, the BLEU metric, influenced by the expression style, suggests that Mathpix is superior. Secondly, for the Pix2Tex model, the BLEU metric is very low on the HWE and SCE subsets but performs well on the SPE and CPE subsets. This discrepancy arises because the Pix2Tex training set includes a large number of printed formulas from arXiv and lacks data in the HWE and SCE styles.", + "bbox": [ + 89, + 518, + 482, + 835 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "These anomalies clearly illustrate the limitations of the BLEU metric in evaluating the quality of formula recognition models. In contrast, the CDM metric proposed in this paper is fair and intuitive.", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3.2. Tiny-Doc-Math Evaluation", + "text_level": 1, + "bbox": [ + 511, + 233, + 750, + 247 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The evaluation results of Tiny-Doc-Math are shown in Table 3. For cropped formula inputs (formula-level), all four models perform reasonably well, with CDM scores above 0.7. Notably, the current leading multimodal large model GPT-4o has the highest BLEU score among the four models but the lowest CDM score. This discrepancy indicates that the BLEU metric may not be reliable, suggesting that the formula recognition accuracy of GPT-4o still has room for improvement, lagging behind traditional SOTA models. Additionally, although Mathpix has the highest CDM score, only $21.05\\%$ of the formulas are completely accurate. Manual verification revealed that many formulas are missing commas or periods at the end.", + "bbox": [ + 511, + 255, + 906, + 450 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "When the input is document-level screenshots, the models output the recognition results for the entire document (not just the formulas). Evaluation is conducted by matching the recognized block formulas. In this scenario, it can be observed that the accuracy of GPT-4o further decreases. In contrast, Mathpix and Nougat perform better, but even the document multimodal large model Nougat only achieves a CDM score of 0.8326. This indicates that there is still significant room for improvement in document-level recognition models. Mathpix remains the best performer, with a fully correct formula rate of $62.92\\%$ . The accuracy of document-level recognition is crucial for advanced document understanding tasks like scientific knowledge Q&A, and CDM provides an excellent standard for selecting formula models and offers direction for improving formula recognition.", + "bbox": [ + 511, + 455, + 908, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 720, + 633, + 736 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduced Character Detection Matching (CDM), a novel evaluation metric for formula recognition task. CDM addresses the shortcomings of the existing metrics in formula recognition by utilizing spatial character matching, overcoming the challenges posed by diverse formula representations. Comprehensive evaluations on different models and datasets demonstrate CDM's superiority in precisely reflecting recognition quality, showing more intuitive assessment and paving the way for future research and improvements in the formula recognition field.", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19688", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 269, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This project was supported by National Key R&D Program of China (NO.2022ZD0160102) and Shanghai Artificial Intelligence Laboratory.", + "bbox": [ + 89, + 114, + 483, + 161 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 176, + 187, + 191 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Francisco Álvaro, Joan-Andreu Sánchez, and José-Miguel Benédí. An integrated grammar-based approach for mathematical expression recognition. Pattern Recognition, 51: 135–147, 2016.", + "[2] Ahmad-Montaser Awal, Harold Mouchere, and Christian Viard-Gaudin. A global learning approach for an online handwritten mathematical expression recognition system. Pattern Recognition Letters, 35:68-77, 2014.", + "[3] Xiaohang Bian, Bo Qin, Xiaozhe Xin, Jianwu Li, Xuefeng Su, and Yanfeng Wang. Handwritten mathematical expression recognition via attention aggregation based bidirectional mutual learning. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 113-121, 2022.", + "[4] Lukas Blecher. pix2tex - latexOCR. https://github.com/lukas-blecher/LaTeX-OCR, 2022. Accessed: 2024-2-29.", + "[5] Lukas Blecher, Guillem Cucurull, Thomas Scialom, and Robert Stojnic. Nougat: Neural optical understanding for academic documents. arXiv.org, 2308.13418, 2023.", + "[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision (ECCV), pages 213-229. Springer, 2020.", + "[7] Hiuyi Cheng, Peirong Zhang, Sihang Wu, Jiaxin Zhang, Qiyuan Zhu, Zecheng Xie, Jing Li, Kai Ding, and Lianwen Jin. M6doc: A large-scale multi-format, multi-type, multi-Layout, multi-language, multi-annotation category dataset for modern document layout analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15138–15147, 2023.", + "[8] Yuntian Deng, Anssi Kanervisto, Jeffrey Ling, and Alexander M Rush. Image-to-markup generation with coarse-to-fine attention. In International Conference on Machine Learning (ICML), pages 980-989. PMLR, 2017.", + "[9] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981.", + "[10] GPT-4o. Gpt-4o. https://openai.com/index/hello-gpt-4o/, 2024. Accessed: 2024-08-15.", + "[11] Mingxin Huang, Yuliang Liu, Dingkang Liang, Lianwen Jin, and Xiang Bai. Mini-monkey: Alleviate the sawtooth effect by multi-scale adaptive cropping. arXiv.org, 2024.", + "[12] Geewook Kim, Teakgyu Hong, Moonbin Yim, JeongYeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sangdoo Yun, Dongyoon Han, and Seunghyun Park. Ocr-free document understanding transformer. In European Confer" + ], + "bbox": [ + 93, + 200, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ence on Computer Vision (ECCV), pages 498-517. Springer, 2022.", + "[13] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955.", + "[14] Vinay Kukreja et al. Recent trends in mathematical expressions recognition: An lda-based analysis. Expert Systems with Applications, 213:119028, 2023.", + "[15] Stéphane Lavirotte and Loïc Pottier. Mathematical formula recognition using graph grammar. In Document Recognition V, San Jose, CA, USA, January 24, 1998, 1998.", + "[16] Anh Duc Le, Bipin Indurkhya, and Masaki Nakagawa. Pattern generation strategies for improving recognition of handwritten mathematical expressions. Pattern Recognition Letters, 128:255-262, 2019.", + "[17] Vladimir I Levenshtein et al. Binary codes capable of correcting deletions, insertions, and reversals. In Soviet physics doklady, pages 707-710. Soviet Union, 1966.", + "[18] Bohan Li, Ye Yuan, Dingkang Liang, Xiao Liu, Zhilong Ji, Jinfeng Bai, Wenyu Liu, and Xiang Bai. When counting meets hmer: counting-aware network for handwritten mathematical expression recognition. In European Conference on Computer Vision (ECCV), pages 197-214. Springer, 2022.", + "[19] Minghao Li, Yiheng Xu, Lei Cui, Shaohan Huang, Furu Wei, Zhoujun Li, and Ming Zhou. Docbank: A benchmark dataset for document layout analysis. arXiv preprint arXiv:2006.01038, 2020.", + "[20] Minghui Liao, Guan Pang, Jing Huang, Tal Hassner, and Xiang Bai. Mask textspotter v3: Segmentation proposal network for robust scene text spotting. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XI 16, pages 706-722. Springer, 2020.", + "[21] Pengyuan Lyu, Yulin Li, Hao Zhou, Weihong Ma, Xingyu Wan, Qunyi Xie, Liang Wu, Chengquan Zhang, Kun Yao, Errui Ding, et al. Structextv3: An efficient vision-language model for text-rich image perception, comprehension, and beyond. arXiv.org, 2024.", + "[22] Scott MacLean and George Labahn. A new approach for recognizing handwritten mathematics using relational grammars and fuzzy sets. International Journal on Document Analysis and Recognition (IJDAR), 16:139-163, 2013.", + "[23] Mahshad Mahdavi, Richard Zanibbi, Harold Mouchere, Christian Viard-Gaudin, and Utpal Garain. Icdar 2019 crohme+ tfd: Competition on recognition of handwritten mathematical expressions and typeset formula detection. In International Conference on Document Analysis and Recognition (ICDAR), pages 1533-1538. IEEE, 2019.", + "[24] Harold Mouchere, Christian Viard-Gaudin, Richard Zanibbi, Utpal Garain, Dae Hwan Kim, and Jin Hyung Kim. Icdar 2013 crohme: Third international competition on recognition of online handwritten mathematical expressions. In International Conference on Document Analysis and Recognition (ICDAR), pages 1428-1432. IEEE, 2013.", + "[25] Harold Mouchere, Christian Viard-Gaudin, Richard Zanibbi, and Utpal Garain. Icfhr 2014 competition on recognition of on-line handwritten mathematical expressions (crohme" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19689", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "2014). In International Conference on Frontiers in Hand-writing Recognition (ICFHR), pages 791-796. IEEE, 2014.", + "[26] Harold Mouchère, Christian Viard-Gaudin, Richard Zanibbi, and Utpal Garain. Icfhr2016 crohme: Competition on recognition of online handwritten mathematical expressions. In International Conference on Frontiers in Handwriting Recognition (ICFHR), pages 607-612. IEEE, 2016.", + "[27] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002.", + "[28] Vik Paruchuri. Texify. https://github.com/VikParuchuri/texify, 2023. Accessed: 2024-2-29.", + "[29] Josselin Somerville Roberts, Tony Lee, Chi Heem Wong, Michihiro Yasunaga, Yifan Mai, and Percy Liang. Image2struct: Benchmarking structure extraction for vision-language models. arXiv preprint arXiv:2410.22456, 2024.", + "[30] Bin Wang, Zhuangcheng Gu, Chao Xu, Bo Zhang, Botian Shi, and Conghui He. Unimernet: A universal network for real-world mathematical expression recognition. arXiv.org, 2024.", + "[31] Bin Wang, Chao Xu, Xiaomeng Zhao, Linke Ouyang, Fan Wu, Zhiyuan Zhao, Rui Xu, Kaiwen Liu, Yuan Qu, Fukai Shang, et al. Mineru: An open-source solution for precise document content extraction. arXiv preprint arXiv:2409.18839, 2024.", + "[32] Zelun Wang and Jyh-Charn Liu. Translating math formula images to latex sequences using deep neural networks with sequence-level training. International Journal on Document Analysis and Recognition (IJDAR), 24(1):63-75, 2021.", + "[33] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE TIP, 13(4):600-612, 2004.", + "[34] Haoran Wei, Lingyu Kong, Jinyue Chen, Liang Zhao, Zheng Ge, Jinrong Yang, Jianjian Sun, Chunrui Han, and Xiangyu Zhang. Vary: Scaling up the vision vocabulary for large vision-language models. arXiv.org, 2312.06109, 2023.", + "[35] Haoran Wei, Lingyu Kong, Jinyue Chen, Liang Zhao, Zheng Ge, Jinrong Yang, Jianjian Sun, Chunrui Han, and Xiangyu Zhang. Vary: Scaling up the vision vocabulary for large vision-language model. In European Conference on Computer Vision, pages 408-424. Springer, 2025.", + "[36] Renqiu Xia, Bo Zhang, Haoyang Peng, Ning Liao, Peng Ye, Botian Shi, Junchi Yan, and Yu Qiao. Structchart: Perception, structuring, reasoning for visual chart understanding. arXiv preprint arXiv:2309.11268, 2023.", + "[37] Renqiu Xia, Song Mao, Xiangchao Yan, Hongbin Zhou, Bo Zhang, Haoyang Peng, Jiahao Pi, Daocheng Fu, Wenjie Wu, Hancheng Ye, et al. Docgenome: An open largescale scientific document benchmark for training and testing multi-modal large language models. arXiv preprint arXiv:2406.11633, 2024.", + "[38] Renqiu Xia, Bo Zhang, Hancheng Ye, Xiangchao Yan, Qi Liu, Hongbin Zhou, Zijun Chen, Min Dou, Botian Shi, Junchi Yan, et al. Chartx & chartvlm: A versatile benchmark and foundation model for complicated chart reasoning. arXiv preprint arXiv:2402.12185, 2024." + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Mingkun Yang, Yushuo Guan, Minghui Liao, Xin He, Kaigui Bian, Song Bai, Cong Yao, and Xiang Bai. Symmetry-constrained rectification network for scene text recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9147-9156, 2019.", + "[40] Ya-Qi Yu, Minghui Liao, Jiwen Zhang, and Jihao Wu. Texthawk2: A large vision-language model excels in bilingualOCR and grounding with 16x fewer tokens. arXiv preprint arXiv:2410.05261, 2024.", + "[41] Ye Yuan, Xiao Liu, Wondimu Dikubab, Hui Liu, Zhilong Ji, Zhongqin Wu, and Xiang Bai. Syntax-aware network for handwritten mathematical expression recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4553-4562, 2022.", + "[42] Jianshu Zhang, Jun Du, Yongxin Yang, Yi-Zhe Song, Si Wei, and Lirong Dai. A tree-structured decoder for image-to-markup generation. In International Conference on Machine Learning (ICML), pages 11076–11085. PMLR, 2020.", + "[43] Wenqi Zhao, Liangcai Gao, Zuoyu Yan, Shuai Peng, Lin Du, and Ziyin Zhang. Handwritten mathematical expression recognition with bidirectionally trained transformer. In International Conference on Document Analysis and Recognition (ICDAR), pages 570-584. Springer, 2021." + ], + "bbox": [ + 516, + 90, + 905, + 416 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19690", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Image Over Text_ Transforming Formula Recognition Evaluation with Character Detection Matching/e8290078-d6a1-49c9-b28d-cb4de981b7cd_model.json b/2025/Image Over Text_ Transforming Formula Recognition Evaluation with Character Detection Matching/e8290078-d6a1-49c9-b28d-cb4de981b7cd_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8a0756bcda5b90a696a60424fe4985179006e957 --- /dev/null +++ b/2025/Image Over Text_ Transforming Formula Recognition Evaluation with Character Detection Matching/e8290078-d6a1-49c9-b28d-cb4de981b7cd_model.json @@ -0,0 +1,2662 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.145, + 0.13, + 0.856, + 0.177 + ], + "angle": 0, + "content": "Image Over Text: Transforming Formula Recognition Evaluation with Character Detection Matching" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.203, + 0.312, + 0.222 + ], + "angle": 0, + "content": "Bin Wang\\(^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.186, + 0.222, + 0.278, + 0.24 + ], + "angle": 0, + "content": "Rui Zhang" + }, + { + "type": "text", + "bbox": [ + 0.34, + 0.204, + 0.42, + 0.22 + ], + "angle": 0, + "content": "Fan \\(\\mathbf{W}\\mathbf{u}^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.316, + 0.222, + 0.425, + 0.24 + ], + "angle": 0, + "content": "Renqiu Xia1,2" + }, + { + "type": "text", + "bbox": [ + 0.443, + 0.204, + 0.573, + 0.221 + ], + "angle": 0, + "content": "Linke Ouyang\\(^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.457, + 0.222, + 0.548, + 0.238 + ], + "angle": 0, + "content": "Botian \\(\\mathbf{Shi}^1\\)" + }, + { + "type": "text", + "bbox": [ + 0.602, + 0.204, + 0.746, + 0.221 + ], + "angle": 0, + "content": "Zhuangcheng Gu" + }, + { + "type": "text", + "bbox": [ + 0.581, + 0.222, + 0.667, + 0.24 + ], + "angle": 0, + "content": "Bo Zhang" + }, + { + "type": "text", + "bbox": [ + 0.7, + 0.222, + 0.812, + 0.24 + ], + "angle": 0, + "content": "Conghui \\(\\mathrm{He}^{1\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.183, + 0.256, + 0.816, + 0.275 + ], + "angle": 0, + "content": "\\(^{1}\\)Shanghai Artificial Intelligence Laboratory \\(^{2}\\)Shanghai Jiao Tong University" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.309, + 0.327, + 0.326 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.341, + 0.485, + 0.764 + ], + "angle": 0, + "content": "Formula recognition presents significant challenges due to the complicated structure and varied notation of mathematical expressions. Despite continuous advancements in formula recognition models, the evaluation metrics employed by these models, such as BLEU and Edit Distance, still exhibit notable limitations. They overlook the fact that the same formula has diverse representations and is highly sensitive to the distribution of training data, thereby causing unfairness in formula recognition evaluation. To this end, we propose a Character Detection Matching (CDM) metric, ensuring the evaluation objectivity by designing an image-level rather than a LaTeX-level metric score. Specifically, CDM renders both the model-predicted LaTeX and the ground-truth LaTeX formulas into image-formatted formulas, then employs visual feature extraction and localization techniques for precise character-level matching, incorporating spatial position information. Such a spatially-aware and character-matching method offers a more accurate and equitable evaluation compared with previous BLEU and Edit Distance metrics that rely solely on text-based character matching. Experimentally, we evaluated various formula recognition models using CDM, BLEU, and ExpRate metrics. Their results demonstrate that the CDM aligns more closely with human evaluation standards and provides a fairer comparison across different models by eliminating discrepancies caused by diverse formula representations. Code is available at https://github.com/opendatalab/UniMERNet/tree/main/cdm" + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.807, + 0.222, + 0.823 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.833, + 0.484, + 0.879 + ], + "angle": 0, + "content": "Mathematical formula recognition is crucial in document analysis as it directly impacts the scientific rigor and accuracy of the document content [7, 19, 31, 37]. Unlike" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.311, + 0.907, + 0.432 + ], + "angle": 0, + "content": "standard Optical Character Recognition (OCR) technique, formula recognition presents unique challenges. Formulas often encompass multi-level symbols, subscripts, fractions, and other complicated structures, requiring models to comprehend spatial and structural relationships rather than just linear, sequential text [14, 19]. Besides, formulas exhibit representational diversity, meaning that the same formula can be expressed in multiple valid ways." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.433, + 0.909, + 0.614 + ], + "angle": 0, + "content": "In recent years, significant advancements in formula recognition [4, 8, 20, 28, 39, 41, 42] have been primarily driven by the continuous development of artificial intelligence technique [5, 35, 40]. Besides, commercial formula recognition software like Mathpix1 and the recently proposed UniMERNet [30] model have achieved impressive results in diverse real-world settings. Despite these advancements, the existing evaluation metrics [17, 27] for formula recognition still face some challenges. Commonly-used metrics such as BLEU [27] and Edit Distance [17] primarily rely on text-based character matching, which introduces several limitations as follows:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.615, + 0.909, + 0.767 + ], + "angle": 0, + "content": "(1) Low Metric Reliability. BLEU and Edit Distance are reliable for evaluating the quality of text-level similarity. However, the diversity in formula representations makes these text-level evaluation metrics inadequate for precisely reflecting formula recognition quality. For example, as shown in Figure 1 (Case 1), a model's prediction might render an image identical to the ground truth formula. However, due to the variations in formula expression styles, the evaluation results obtained using the ExpRate [8], BLEU, and Edit Distance may be somewhat misleading." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.768, + 0.909, + 0.874 + ], + "angle": 0, + "content": "(2) Unfair Model Comparison. Current metrics may be susceptible to discrepancies between the distributions of training and testing data. As illustrated in Figure 1 (Case 1 and Case 2), a model may produce a correct prediction but score poorly due to representational differences from the ground truth, while an incorrect prediction might score higher if its representation aligns more closely with the test" + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.615, + 0.909, + 0.874 + ], + "angle": 0, + "content": null + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.483, + 0.901 + ], + "angle": 0, + "content": "*Equal contribution. †Corresponding author: heconghui@pjlab.org.cn" + }, + { + "type": "page_footnote", + "bbox": [ + 0.532, + 0.887, + 0.835, + 0.9 + ], + "angle": 0, + "content": "r={\\frac{1}{2}{\\alpha}{\\beta}}|\\sin\\beta\\left( \\right)\\sigma\\sigma\\sigma-{2}\\sigma_{1}\\left( \\right)1" + }, + { + "type": "table_caption", + "bbox": [ + 0.119, + 0.283, + 0.272, + 0.295 + ], + "angle": 0, + "content": "2. Element Region Localization" + }, + { + "type": "equation", + "bbox": [ + 0.116, + 0.308, + 0.325, + 0.335 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{c c} \\frac {\\text {r e n d e r e a c h}}{\\text {t o k e n w i t h a}} & r = \\frac {\\alpha}{\\beta} \\big | \\sin \\beta \\left(\\sigma_ {2} + \\sigma_ {1}\\right) \\big | \\\\ \\text {d i f f e r e n t c o l o r} & \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.115, + 0.34, + 0.325, + 0.388 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{c} r = \\frac {\\alpha}{\\beta} | \\sin \\beta (\\sigma_ {2} + \\sigma_ {1}) | \\\\ \\text {e x t r a c t e a c h} \\\\ \\text {c o l o r a n d l o c a t e} \\\\ \\text {t h e b o u n d i n g b o x} \\end{array} r = \\frac {\\alpha}{\\beta} | \\sin \\beta (\\sigma_ {2} + \\sigma_ {1}) |\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.122, + 0.402, + 0.311, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.101, + 0.715, + 0.114 + ], + "angle": 0, + "content": "Step 2: Element Region Matching" + }, + { + "type": "image_caption", + "bbox": [ + 0.394, + 0.125, + 0.426, + 0.137 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.357, + 0.15, + 0.465, + 0.186 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.357, + 0.201, + 0.466, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.124, + 0.663, + 0.145 + ], + "angle": 0, + "content": "1. Calculate Match Cost Between Every Element Pairs" + }, + { + "type": "image", + "bbox": [ + 0.468, + 0.152, + 0.69, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.694, + 0.124, + 0.825, + 0.145 + ], + "angle": 0, + "content": "2. Find the Best Match by Hungarian Algorithm" + }, + { + "type": "image", + "bbox": [ + 0.663, + 0.152, + 0.861, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.39, + 0.259, + 0.566, + 0.271 + ], + "angle": 0, + "content": "Step 3: Invalid Match Elimination" + }, + { + "type": "image_caption", + "bbox": [ + 0.346, + 0.283, + 0.484, + 0.295 + ], + "angle": 0, + "content": "1. Token Consistency Check" + }, + { + "type": "image", + "bbox": [ + 0.37, + 0.294, + 0.571, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.347, + 0.385, + 0.558, + 0.397 + ], + "angle": 0, + "content": "2. Position Relationship Consistency Check" + }, + { + "type": "image", + "bbox": [ + 0.371, + 0.399, + 0.571, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.681, + 0.259, + 0.835, + 0.271 + ], + "angle": 0, + "content": "Step 4: Metric Calculation" + }, + { + "type": "image_caption", + "bbox": [ + 0.717, + 0.29, + 0.798, + 0.3 + ], + "angle": 0, + "content": "Evaluation Metric" + }, + { + "type": "equation", + "bbox": [ + 0.705, + 0.307, + 0.817, + 0.327 + ], + "angle": 0, + "content": "\\[\nF 1 \\text {s c o r e} = \\frac {2 T P}{2 T P + F P + F N}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.683, + 0.336, + 0.808, + 0.344 + ], + "angle": 0, + "content": "- TP: number of matched bbox pairs" + }, + { + "type": "text", + "bbox": [ + 0.683, + 0.346, + 0.844, + 0.354 + ], + "angle": 0, + "content": "- FP: number of unmatched bboxes in prediction" + }, + { + "type": "text", + "bbox": [ + 0.683, + 0.356, + 0.823, + 0.364 + ], + "angle": 0, + "content": "FN: number of unmatched bboxes in GT" + }, + { + "type": "list", + "bbox": [ + 0.683, + 0.336, + 0.844, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "equation", + "bbox": [ + 0.653, + 0.385, + 0.743, + 0.426 + ], + "angle": 0, + "content": "\\[\nT P = \\left(\\bigtriangledown\\right) _ {n u m} = 1 3\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.775, + 0.384, + 0.861, + 0.403 + ], + "angle": 0, + "content": "\\[\nF P = \\left(\\square\\right) _ {\\text {n u m}} ^ {\\text {P r e d}} = 3\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.775, + 0.416, + 0.861, + 0.434 + ], + "angle": 0, + "content": "\\[\nF N = \\left( \\begin{array}{l l} \\square & T \\\\ 0 & n u m \\end{array} \\right) ^ {G T} = 3\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.678, + 0.444, + 0.833, + 0.468 + ], + "angle": 0, + "content": "\\[\nF 1 s c o r e = \\frac {2 \\times 1 3}{2 \\times 1 3 + 3 + 3} = 0. 8 1 2 5\n\\]" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.498, + 0.907, + 0.555 + ], + "angle": 0, + "content": "Figure 2. Overview of the Character Detection Matching (CDM), consisting of four main stages. (1) Element Localization, where bounding boxes of individual elements are extracted. (2) Element Region Matching, which employs a bipartite graph matching method to pair prediction with ground truth elements. (3) Invalid Match Elimination, where inconsistent matches are discarded through token and positional relationship checks. (4) Metric Calculation, where matching accuracy is evaluated using the F1-Score and ExpRate@CDM." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.568, + 0.445, + 0.584 + ], + "angle": 0, + "content": "of formula recognition across diverse representations." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.601, + 0.372, + 0.619 + ], + "angle": 0, + "content": "4. Character Detection Matching" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.628, + 0.483, + 0.868 + ], + "angle": 0, + "content": "Due to the diversity of LaTeX expressions, text-based character-matching methods are unreliable for formula recognition evaluation. The basic idea of CDM is to compare the rendered images from LaTex text. If the image rendered from the predicted LaTeX source code matches the image rendered from the ground truth LaTeX source code, the formula is considered entirely correct. However, directly comparing the pixel values of the original and predicted formulas is not ideal. Any error or extra/missing character in the prediction can cause subsequent characters to be mismatched. Additionally, two similar formulas might have different layouts, with one being a single-line formula and the other a multi-line formula due to line breaks. Therefore, a more robust algorithm is needed to calculate the match between the predicted result and the ground truth image." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "To this end, we propose a metric that incorporates a bipartite matching step for element-level matching in images," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.905, + 0.599 + ], + "angle": 0, + "content": "providing a more intuitive assessment. As shown in Figure 2, the algorithm consists of four stages as follows." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.607, + 0.714, + 0.621 + ], + "angle": 0, + "content": "4.1. Element Localization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.905, + 0.674 + ], + "angle": 0, + "content": "First, the bounding boxes (bboxes) of each individual element in the rendered image are extracted, followed by the subsequent steps:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.675, + 0.906, + 0.765 + ], + "angle": 0, + "content": "LaTeX Source Normalization. LaTeX source codes of both the ground truth and predicted formulas are normalized, breaking them down into individual tokens such as \"2\", \"a\", \"A\", \"\\alpha\", \"\\sin\". Composite elements are decomposed into individual characters, e.g., \"\\frac{ab}{\\alpha} is decomposed into \\frac{a}{b}\\)\"." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Element Region Localization. To accurately detect character positions, we render each token in a unique color. We construct an RGB color list with a fixed interval of 15, creating a sequence from \\((0,0,15)\\) to \\((255,255,255)\\). This generates \\((255 / 15 + 1)^{3} = 5832\\) distinct colors, sufficient for complex formulas. Each token is assigned a unique color using the LaTeX command \"\\mathsf{\\\\mathsf{m}}\\mathsf{a}\\mathsf{h}\\mathsf{o}\\mathsf{c}\\mathsf{l}\\mathsf{r}\\mathsf{R}\\mathsf{B}\\mathsf{]} \\{\\mathsf{r},\\mathsf{g},\\mathsf{b}\\} \"$. After rendering, we apply image processing techniques to identify pixels match-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19684" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.123 + ], + "angle": 0, + "content": "ing each color, thereby determining the exact coordinates of each token." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.129, + 0.331, + 0.147 + ], + "angle": 0, + "content": "4.2. Element Region Matching" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.151, + 0.484, + 0.272 + ], + "angle": 0, + "content": "In this stage, a bipartite matching method pairs the predicted elements with the corresponding ground truth elements. Based on the element localization, two sets are obtained for each formula: one for the ground truth independent elements \\(y\\) and one for the predicted independent elements \\(\\hat{y}\\). The number of independent elements in each set is \\(N_y\\) and \\(N_{\\hat{y}}\\), respectively, with \\(N = \\min(N_y, N_{\\hat{y}})\\) being the number of elements in the smaller set." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.273, + 0.484, + 0.364 + ], + "angle": 0, + "content": "To measure the similarity between \\( y \\) and \\( \\hat{y} \\), we match elements in the two sets by identifying the corresponding ground truth element for each predicted element. We use the bipartite matching Hungarian algorithm [13], as described in DETR [6], to find a permutation \\( \\hat{\\sigma} \\) that minimizes the total matching cost:" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.37, + 0.484, + 0.412 + ], + "angle": 0, + "content": "\\[\n\\hat {\\sigma} = \\arg \\min _ {\\sigma \\in S _ {N}} \\sum_ {i = 1} ^ {N} L _ {\\text {m a t c h}} \\left(y _ {i}, \\hat {y} _ {\\sigma (i)}\\right), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.142, + 0.423, + 0.483, + 0.44 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {m a t c h}} = W _ {t} \\times L _ {t} + W _ {p} \\times L _ {p} + W _ {o} \\times L _ {o}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.444, + 0.484, + 0.473 + ], + "angle": 0, + "content": "where the matching cost \\( L_{match} \\) is defined as a weighted sum of three components as introduced as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.475, + 0.484, + 0.578 + ], + "angle": 0, + "content": "- Token Matching Cost \\( L_{t} \\): This component measures whether the tokens corresponding to two bounding boxes are the same. If they are identical, the cost is 0; if they are different, the cost is 1. For tokens that render identically but are different, such as \" (\" , \"\\leftarrow\" , and \"\\bigig( \", the cost is 0.05 , which can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.2, + 0.578, + 0.483, + 0.634 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {t}} = \\left\\{ \\begin{array}{l l} 0, & \\text {i f} t _ {i} = \\hat {t} _ {\\sigma (i)}; \\\\ 0. 0 5, & \\text {i f} t _ {i} \\approx \\hat {t} _ {\\sigma (i)}; \\\\ 1, & \\text {o t h e r w i s e}; \\end{array} \\right. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.637, + 0.481, + 0.651 + ], + "angle": 0, + "content": "where \\(\\approx\\) denotes tokens that differ but render identically." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.652, + 0.483, + 0.711 + ], + "angle": 0, + "content": "- Positional Proximity Cost \\(L_{p}\\): This component measures the proximity of the two bounding boxes' positions using the L1 norm of their coordinates, which can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.204, + 0.715, + 0.483, + 0.747 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {p}} = \\frac {1}{D _ {b}} \\times \\| b _ {i} - \\hat {b} _ {\\sigma (i)} \\| _ {1}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.752, + 0.483, + 0.782 + ], + "angle": 0, + "content": "where \\(b = [x_1, y_1, x_2, y_2]\\), and \\(D_b\\) is the dimension of the bounding box coordinates." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.782, + 0.484, + 0.855 + ], + "angle": 0, + "content": "- Order Similarity Cost \\(L_{o}\\): This measures the similarity of the token order in the original LaTeX source (an approximation of reading order). The order is normalized to the range [0, 1], and the L1 norm can be calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.203, + 0.854, + 0.483, + 0.884 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {o}} = \\frac {1}{D _ {o}} \\times \\left\\| o _ {i} - \\hat {o} _ {\\sigma (i)} \\right\\| _ {1}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.886, + 0.476, + 0.902 + ], + "angle": 0, + "content": "where \\(o_i\\) means order of \\(i\\)-th token, with dimension \\(D_o\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.184 + ], + "angle": 0, + "content": "Overall, the weights \\( W_{t}, W_{p}, W_{o} \\) are used to balance the contributions of the three components. By employing this comprehensive matching strategy, we ensure a more accurate and robust evaluation of the correspondence between the predicted and ground truth elements, thereby improving the overall assessment of formula recognition quality." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.193, + 0.757, + 0.208 + ], + "angle": 0, + "content": "4.3. Invalid Match Elimination" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.215, + 0.906, + 0.275 + ], + "angle": 0, + "content": "After pairing the individual elements of the predicted result with the ground truth using the Hungarian matching algorithm, we need to verify these pairs and eliminate invalid matches. This process involves two steps:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.276, + 0.906, + 0.321 + ], + "angle": 0, + "content": "Token Consistency Check. Check whether the elements in each matched pair are consistent in terms of characters. If they are inconsistent, discard the match." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.322, + 0.906, + 0.472 + ], + "angle": 0, + "content": "Position Relationship Consistency Check. The relative positions of elements in mathematical formulas are crucial. For instance, in the expressions \\( 2^3 \\) and \\( 3^2 \\), bipartite matching might pair 2 with 2 and 3 with 3, but their meanings and visual representations are entirely different. Thus, we need to check the consistency of the positional relationships within the matched pairs. We treat each element in the matched pair as a bounding box and analyze their relative positions. Specifically, we assume an affine transformation between the ground truth and predicted elements:" + }, + { + "type": "equation", + "bbox": [ + 0.658, + 0.483, + 0.906, + 0.503 + ], + "angle": 0, + "content": "\\[\n\\hat {b} _ {\\sigma (i)} = \\mathbf {A} \\left(b _ {i}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.513, + 0.906, + 0.68 + ], + "angle": 0, + "content": "where \\(\\mathbf{A}\\) is the affine transformation matrix. To identify inconsistent match pairs, we detect pairs that do not conform to this transformation relationship. We employ the RANSAC algorithm [9] for this purpose. RANSAC can determine the optimal transformation matrix \\(\\mathbf{A}\\) in the presence of noise. Given that formulas are usually horizontally arranged during rendering, we fix the rotation angle in the transformation matrix to 0, considering only translation and scaling. This approach not only improves the convergence speed of the RANSAC algorithm but also enhances the final matching accuracy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.68, + 0.906, + 0.77 + ], + "angle": 0, + "content": "To account for line-breaking effects in formulas, we perform multiple rounds of RANSAC iterations to ensure that as many matched pairs as possible conform to the transformation relationship. After several iterations, matched pairs that still do not conform to the transformation relationship are considered incorrect and are eliminated." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.771, + 0.906, + 0.802 + ], + "angle": 0, + "content": "The above two steps effectively eliminate invalid match pairs, ensuring more accurate final matching results." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.811, + 0.696, + 0.826 + ], + "angle": 0, + "content": "4.4. Metric Calculation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.834, + 0.906, + 0.864 + ], + "angle": 0, + "content": "We use the F1-Score as the default metric for evaluating CDM (Character Detection Metric), defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.874, + 0.905, + 0.906 + ], + "angle": 0, + "content": "\\[\nC D M = \\frac {2 \\times T P}{2 \\times T P + F P + F N}, \\tag {8}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19685" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.171, + 0.079, + 0.538, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.228, + 0.295, + 0.476, + 0.307 + ], + "angle": 0, + "content": "(a) CDM distribution on UniMER-Test dataset" + }, + { + "type": "image", + "bbox": [ + 0.542, + 0.078, + 0.828, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.59, + 0.295, + 0.776, + 0.308 + ], + "angle": 0, + "content": "(b) Examples by CDM score range" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.315, + 0.907, + 0.345 + ], + "angle": 0, + "content": "Figure 3. CDM Score Distribution and Example Analysis. (a) CDM score distribution for Mathpix, UniMERNet, Texify, and Pix2Tex on the UniMER-Test dataset. (b) Example analysis of Pix2Tex predictions across different CDM score ranges." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.351, + 0.483, + 0.381 + ], + "angle": 0, + "content": "where \\( TP \\) denotes true positives, \\( FP \\) denotes false positives, and \\( FN \\) denotes false negatives." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.381, + 0.483, + 0.412 + ], + "angle": 0, + "content": "To further evaluate the accuracy of formula recognition, we introduce the ExpRate@CDM metric, defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.139, + 0.423, + 0.483, + 0.456 + ], + "angle": 0, + "content": "\\[\n\\operatorname {E x p R a t e} @ C D M = \\frac {\\sum_ {i = 1} ^ {N} \\mathbb {I} \\left(C D M _ {i} = 1\\right)}{N}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.467, + 0.484, + 0.558 + ], + "angle": 0, + "content": "where \\(\\mathbb{I}\\) is the indicator function that equals 1 if \\(CDM_{i} = 1\\) and 0 otherwise, and \\(N\\) is the total number of formulas. This metric represents the proportion of formulas for which the model's prediction results are perfectly matched. Essentially, ExpRate@CDM serves as a precise version of the ExpRate metric specifically for formula recognition." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.581, + 0.224, + 0.598 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.606, + 0.259, + 0.621 + ], + "angle": 0, + "content": "5.1. Models and Data" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.764 + ], + "angle": 0, + "content": "We validate the CDM metric by evaluating several mainstream formula recognition models using both subjective impressions and objective metrics. The models include open-source UniMERNet [30], Texify [28], Pix2tex [4], and the commercial Mathpix API, all tested on the UniMERTest dataset. Besides, we evaluate document-level models, such as the open-source Nougat [5] and the commercial GPT-4o [10]. Vary [34] and StrucTexTv3 [21] are excluded as they are currently unavailable." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.84 + ], + "angle": 0, + "content": "UniMER-Test Dataset. The dataset includes 23,757 formula samples, categorized into Simple Printed Expressions (SPE), Complex Printed Expressions (CPE), Screenshot Expressions (SCE), and Handwritten Expressions (HWE). We use these categories to conduct the model evaluation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Tiny-Doc-Math Dataset. To evaluate document-level recognition, we construct the Tiny-Doc-Math dataset, consisting of arXiv papers in mathematics and computer science, published after June 2024, to ensure that they are not" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.351, + 0.907, + 0.425 + ], + "angle": 0, + "content": "in the training data of the compared models. We obtain Latex code and corresponding PDFs, match displayed equations using regular expressions, and manually verify them. Overall, the dataset includes 12 PDFs, totaling 196 pages and 437 formulas." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.427, + 0.907, + 0.456 + ], + "angle": 0, + "content": "This validation set includes both formula-level and document-level evaluations:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.457, + 0.906, + 0.532 + ], + "angle": 0, + "content": "- Formula-level: Using single rendered formula images as input, we evaluate Mathpix, Pix2Tex, and UniMERNet. These models accept cropped formula images as input, and we compare the model outputs with the ground truth to compute relevant metrics." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.533, + 0.906, + 0.624 + ], + "angle": 0, + "content": "- Document-level: Using PDFs or images as input, we evaluate Nougat, GPT-4o, and Mathpix, which can convert entire PDF pages into Markdown format. We match the displayed equations in the model outputs using regular expressions and compare them with the ground truth LaTeX formulas to compute relevant metrics." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.457, + 0.906, + 0.624 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.634, + 0.794, + 0.65 + ], + "angle": 0, + "content": "5.2. Credibility Assessment of CDM" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.657, + 0.727, + 0.672 + ], + "angle": 0, + "content": "5.2.1. Rendering Success Rate" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.676, + 0.907, + 0.797 + ], + "angle": 0, + "content": "The CDM metric relies on the successful rendering of formula images. For models that fail to render images, we assign a CDM score of 0, as rendering failures indicate that the predicted LaTeX code lacks critical elements. The rendering success rates on the UniMER-Test dataset for Pix2tex, Texify, UniMERNet, and Mathpix are \\(96.63\\%\\), \\(94.77\\%\\), \\(99.71\\%\\), and \\(97.82\\%\\), respectively, ensuring the applicability and reliability of the CDM metric." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.807, + 0.753, + 0.82 + ], + "angle": 0, + "content": "5.2.2. User Preference Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.906, + 0.902 + ], + "angle": 0, + "content": "We analyze the distribution of CDM scores for four models on the UniMER-Test dataset. As shown in Figure 3a, Math-pix and UniMERNet perform well in terms of CDM scores. We conduct a detailed analysis of the Pix2Tex model by randomly selecting samples from different score ranges to" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19686" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.109, + 0.088, + 0.322, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.122, + 0.279, + 0.306, + 0.292 + ], + "angle": 0, + "content": "(a) Human Evaluation Preferences" + }, + { + "type": "image", + "bbox": [ + 0.329, + 0.089, + 0.63, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.373, + 0.279, + 0.581, + 0.292 + ], + "angle": 0, + "content": "(b) BLEU vs.CDM Score Distribution" + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.09, + 0.888, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.661, + 0.279, + 0.86, + 0.293 + ], + "angle": 0, + "content": "(c) Impact of Formula Writing Styles" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.296, + 0.908, + 0.325 + ], + "angle": 0, + "content": "Figure 4. Human evaluation and metric comparison. (a) Distribution of human preferences. (b) Distribution of cases preferring BLEU or CDM across different score ranges. (c) Impact of formula writing styles on BLEU, ExpRate and CDM scores." + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.335, + 0.473, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.55, + 0.483, + 0.607 + ], + "angle": 0, + "content": "Figure 5. Annotation interface for evaluating user preferences between CDM and BLEU scores. Annotators chose from four options: \"Score A is better\", \"Score B is better\", \"Both scores are equally good\", or \"Neither score is good\"." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.625, + 0.483, + 0.685 + ], + "angle": 0, + "content": "evaluate if the prediction quality corresponds to the CDM scores. The analysis in Figure 3b shows that the CDM scores effectively reflect formula quality, with higher scores indicating fewer errors." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.687, + 0.483, + 0.839 + ], + "angle": 0, + "content": "To verify the consistency between the CDM metric and human evaluation, we select 1,008 CDM scores from Pix2Tex predictions, ensuring a balanced score distribution. We design an annotation interface (shown in Figure 5), displaying a ground truth label and the corresponding predicted LaTeX rendered image. Annotators choose between ScoreA, ScoreB, Both (credible), and Neither (credible). ScoreA and ScoreB correspond to the BLEU and CDM scores, respectively, but their order is randomized. For more details, please refer to the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "The results in Figure 4a show that \\(64\\%\\) of participants prefer the CDM metric, and \\(32\\%\\) consider both metrics good. This indicates a \\(96\\%\\) consistency between the CDM metric and human evaluation, demonstrating its reliability." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.334, + 0.912, + 0.415 + ], + "angle": 0, + "content": "
ModelExpRateExpRate@CDMBLEUCDM
Pix2tex0.12370.29100.40800.6360
Texify0.22880.49500.58900.7550
Matphix0.26100.50000.80670.9510
UniMERNet0.47990.81100.84250.9680
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.425, + 0.907, + 0.453 + ], + "angle": 0, + "content": "Table 1. UniMER-Test evaluation results. Comparison of mainstream models using different metrics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.458, + 0.907, + 0.504 + ], + "angle": 0, + "content": "Figure 4b compares the number of cases where BLEU or CDM is preferred across different score ranges, showing that CDM consistently outperforms BLEU." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.51, + 0.774, + 0.525 + ], + "angle": 0, + "content": "5.2.3. Objective Stability Assessment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.529, + 0.907, + 0.726 + ], + "angle": 0, + "content": "To evaluate the impact of formula writing styles on the CDM and BLEU metrics, we randomly select 50 formulas with LaTeX source code and rewrite each formula five times using GPT-4, generating 250 additional formulas. We manually verify these formulas to ensure their rendered results are identical to the original 50 formulas. Using the initial LaTeX source code as the ground truth, we analyze the score distribution of the BLEU and CDM metrics. As shown in Figure 4c, the CDM metric is unaffected by style changes, with all samples scoring 1. In contrast, the BLEU metric's scores are dispersed, making it unsuitable for formula evaluation. The CDM metric remains robust and reliable despite formatting changes." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.733, + 0.812, + 0.748 + ], + "angle": 0, + "content": "5.3. Evaluation of Mainstream Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.755, + 0.906, + 0.846 + ], + "angle": 0, + "content": "We conduct a detailed evaluation of mainstream models using both the CDM and BLEU metric. Note that all BLEU metric in this paper have been normalized [4, 8]. However, as discussed in the limitation section, normalization operations cannot address all issues, which will be evident in the following experiments." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.852, + 0.738, + 0.866 + ], + "angle": 0, + "content": "5.3.1. UniMER-Test Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "As shown in Table 1, the evaluation results of the four models on the UniMER-Test dataset indicate that the quality" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19687" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.191, + 0.079, + 0.812, + 0.187 + ], + "angle": 0, + "content": "
MethodSPECPEHWESCE
BLEU ↑CDM ↑BLEU ↑CDM ↑BLEU ↑CDM ↑BLEU ↑CDM ↑
Pix2tex [4]0.87300.96190.65500.64890.01200.24530.09200.6762
Texify [28]0.90600.98520.69000.70410.34100.52690.42000.7932
Mathpix0.79200.97290.80610.96710.80600.93180.81820.9238
UniMERNet [30]0.91700.99460.91600.97070.92100.95300.61600.9461
" + }, + { + "type": "table_caption", + "bbox": [ + 0.131, + 0.193, + 0.865, + 0.207 + ], + "angle": 0, + "content": "Table 2. UniMER-Test subset evaluation results. Analysis of anomalies in BLEU and CDM metrics across different subsets." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.23, + 0.499, + 0.352 + ], + "angle": 0, + "content": "
Image TypeModelBLEUCDMExpRate@CDM
FormulaPix2tex0.46480.744400.3684
GPT-4o0.64310.73300.4324
UniMERNet0.60560.93960.6887
Mathpix0.61120.94800.2105
DocumentGPT-4o0.34110.65020.1670
Nougat0.58970.83260.6086
Mathpix0.59390.95670.6292
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.362, + 0.486, + 0.406 + ], + "angle": 0, + "content": "Table 3. Tiny-Doc-Math Evaluation Results. Performance of mainstream models on cropped formula inputs and document-level screenshots using BLEU and CDM metrics." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.439, + 0.484, + 0.515 + ], + "angle": 0, + "content": "ranges from low to high as follows: Pix2Tex, Texify, Mathpix, and UniMERNet, based on both BLEU and CDM metrics. ExpRate@CDM clearly shows the proportion of completely correct predictions for each model, indicating that the text character-based ExpRate is unreliable." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.519, + 0.483, + 0.837 + ], + "angle": 0, + "content": "From the results in Table 1, it appears that the trends of the BLEU and CDM metrics are consistent. To verify the reliability of using the BLEU metric for model comparison, we further present evaluation results on the UniMERTest subsets. As shown in Table 2, we observe two notable anomalies: Firstly, in the SCE subset, when comparing the quality of the Mathpix and UniMERNet models, the BLEU and CDM metrics provide opposite conclusions. A detailed review of the UniMERNet paper reveals that the SCE subset was annotated based on Mathpix and then manually corrected. This means that the expression style of the SCE formulas is more consistent with Mathpix. Consequently, even though the CDM metric indicates that UniMERNet has better actual model quality, the BLEU metric, influenced by the expression style, suggests that Mathpix is superior. Secondly, for the Pix2Tex model, the BLEU metric is very low on the HWE and SCE subsets but performs well on the SPE and CPE subsets. This discrepancy arises because the Pix2Tex training set includes a large number of printed formulas from arXiv and lacks data in the HWE and SCE styles." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.903 + ], + "angle": 0, + "content": "These anomalies clearly illustrate the limitations of the BLEU metric in evaluating the quality of formula recognition models. In contrast, the CDM metric proposed in this paper is fair and intuitive." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.234, + 0.75, + 0.248 + ], + "angle": 0, + "content": "5.3.2. Tiny-Doc-Math Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.256, + 0.907, + 0.452 + ], + "angle": 0, + "content": "The evaluation results of Tiny-Doc-Math are shown in Table 3. For cropped formula inputs (formula-level), all four models perform reasonably well, with CDM scores above 0.7. Notably, the current leading multimodal large model GPT-4o has the highest BLEU score among the four models but the lowest CDM score. This discrepancy indicates that the BLEU metric may not be reliable, suggesting that the formula recognition accuracy of GPT-4o still has room for improvement, lagging behind traditional SOTA models. Additionally, although Mathpix has the highest CDM score, only \\(21.05\\%\\) of the formulas are completely accurate. Manual verification revealed that many formulas are missing commas or periods at the end." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.456, + 0.909, + 0.698 + ], + "angle": 0, + "content": "When the input is document-level screenshots, the models output the recognition results for the entire document (not just the formulas). Evaluation is conducted by matching the recognized block formulas. In this scenario, it can be observed that the accuracy of GPT-4o further decreases. In contrast, Mathpix and Nougat perform better, but even the document multimodal large model Nougat only achieves a CDM score of 0.8326. This indicates that there is still significant room for improvement in document-level recognition models. Mathpix remains the best performer, with a fully correct formula rate of \\(62.92\\%\\). The accuracy of document-level recognition is crucial for advanced document understanding tasks like scientific knowledge Q&A, and CDM provides an excellent standard for selecting formula models and offers direction for improving formula recognition." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.721, + 0.634, + 0.737 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In this paper, we introduced Character Detection Matching (CDM), a novel evaluation metric for formula recognition task. CDM addresses the shortcomings of the existing metrics in formula recognition by utilizing spatial character matching, overcoming the challenges posed by diverse formula representations. Comprehensive evaluations on different models and datasets demonstrate CDM's superiority in precisely reflecting recognition quality, showing more intuitive assessment and paving the way for future research and improvements in the formula recognition field." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19688" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.27, + 0.108 + ], + "angle": 0, + "content": "7. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.116, + 0.484, + 0.162 + ], + "angle": 0, + "content": "This project was supported by National Key R&D Program of China (NO.2022ZD0160102) and Shanghai Artificial Intelligence Laboratory." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.177, + 0.188, + 0.193 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.202, + 0.484, + 0.257 + ], + "angle": 0, + "content": "[1] Francisco Álvaro, Joan-Andreu Sánchez, and José-Miguel Benédí. An integrated grammar-based approach for mathematical expression recognition. Pattern Recognition, 51: 135–147, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.259, + 0.484, + 0.315 + ], + "angle": 0, + "content": "[2] Ahmad-Montaser Awal, Harold Mouchere, and Christian Viard-Gaudin. A global learning approach for an online handwritten mathematical expression recognition system. Pattern Recognition Letters, 35:68-77, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.317, + 0.484, + 0.398 + ], + "angle": 0, + "content": "[3] Xiaohang Bian, Bo Qin, Xiaozhe Xin, Jianwu Li, Xuefeng Su, and Yanfeng Wang. Handwritten mathematical expression recognition via attention aggregation based bidirectional mutual learning. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 113-121, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.402, + 0.482, + 0.441 + ], + "angle": 0, + "content": "[4] Lukas Blecher. pix2tex - latexOCR. https://github.com/lukas-blecher/LaTeX-OCR, 2022. Accessed: 2024-2-29." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.445, + 0.484, + 0.486 + ], + "angle": 0, + "content": "[5] Lukas Blecher, Guillem Cucurull, Thomas Scialom, and Robert Stojnic. Nougat: Neural optical understanding for academic documents. arXiv.org, 2308.13418, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.489, + 0.484, + 0.557 + ], + "angle": 0, + "content": "[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision (ECCV), pages 213-229. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.56, + 0.485, + 0.656 + ], + "angle": 0, + "content": "[7] Hiuyi Cheng, Peirong Zhang, Sihang Wu, Jiaxin Zhang, Qiyuan Zhu, Zecheng Xie, Jing Li, Kai Ding, and Lianwen Jin. M6doc: A large-scale multi-format, multi-type, multi-Layout, multi-language, multi-annotation category dataset for modern document layout analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15138–15147, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.659, + 0.484, + 0.713 + ], + "angle": 0, + "content": "[8] Yuntian Deng, Anssi Kanervisto, Jeffrey Ling, and Alexander M Rush. Image-to-markup generation with coarse-to-fine attention. In International Conference on Machine Learning (ICML), pages 980-989. PMLR, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.716, + 0.484, + 0.77 + ], + "angle": 0, + "content": "[9] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.773, + 0.482, + 0.8 + ], + "angle": 0, + "content": "[10] GPT-4o. Gpt-4o. https://openai.com/index/hello-gpt-4o/, 2024. Accessed: 2024-08-15." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.483, + 0.843 + ], + "angle": 0, + "content": "[11] Mingxin Huang, Yuliang Liu, Dingkang Liang, Lianwen Jin, and Xiang Bai. Mini-monkey: Alleviate the sawtooth effect by multi-scale adaptive cropping. arXiv.org, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[12] Geewook Kim, Teakgyu Hong, Moonbin Yim, JeongYeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sangdoo Yun, Dongyoon Han, and Seunghyun Park. Ocr-free document understanding transformer. In European Confer" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.202, + 0.485, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "ence on Computer Vision (ECCV), pages 498-517. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[13] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.905, + 0.207 + ], + "angle": 0, + "content": "[14] Vinay Kukreja et al. Recent trends in mathematical expressions recognition: An lda-based analysis. Expert Systems with Applications, 213:119028, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.209, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[15] Stéphane Lavirotte and Loïc Pottier. Mathematical formula recognition using graph grammar. In Document Recognition V, San Jose, CA, USA, January 24, 1998, 1998." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.251, + 0.905, + 0.305 + ], + "angle": 0, + "content": "[16] Anh Duc Le, Bipin Indurkhya, and Masaki Nakagawa. Pattern generation strategies for improving recognition of handwritten mathematical expressions. Pattern Recognition Letters, 128:255-262, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.308, + 0.905, + 0.35 + ], + "angle": 0, + "content": "[17] Vladimir I Levenshtein et al. Binary codes capable of correcting deletions, insertions, and reversals. In Soviet physics doklady, pages 707-710. Soviet Union, 1966." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.351, + 0.905, + 0.42 + ], + "angle": 0, + "content": "[18] Bohan Li, Ye Yuan, Dingkang Liang, Xiao Liu, Zhilong Ji, Jinfeng Bai, Wenyu Liu, and Xiang Bai. When counting meets hmer: counting-aware network for handwritten mathematical expression recognition. In European Conference on Computer Vision (ECCV), pages 197-214. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.422, + 0.905, + 0.476 + ], + "angle": 0, + "content": "[19] Minghao Li, Yiheng Xu, Lei Cui, Shaohan Huang, Furu Wei, Zhoujun Li, and Ming Zhou. Docbank: A benchmark dataset for document layout analysis. arXiv preprint arXiv:2006.01038, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.479, + 0.905, + 0.561 + ], + "angle": 0, + "content": "[20] Minghui Liao, Guan Pang, Jing Huang, Tal Hassner, and Xiang Bai. Mask textspotter v3: Segmentation proposal network for robust scene text spotting. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XI 16, pages 706-722. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.564, + 0.905, + 0.632 + ], + "angle": 0, + "content": "[21] Pengyuan Lyu, Yulin Li, Hao Zhou, Weihong Ma, Xingyu Wan, Qunyi Xie, Liang Wu, Chengquan Zhang, Kun Yao, Errui Ding, et al. Structextv3: An efficient vision-language model for text-rich image perception, comprehension, and beyond. arXiv.org, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.634, + 0.905, + 0.689 + ], + "angle": 0, + "content": "[22] Scott MacLean and George Labahn. A new approach for recognizing handwritten mathematics using relational grammars and fuzzy sets. International Journal on Document Analysis and Recognition (IJDAR), 16:139-163, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.691, + 0.905, + 0.773 + ], + "angle": 0, + "content": "[23] Mahshad Mahdavi, Richard Zanibbi, Harold Mouchere, Christian Viard-Gaudin, and Utpal Garain. Icdar 2019 crohme+ tfd: Competition on recognition of handwritten mathematical expressions and typeset formula detection. In International Conference on Document Analysis and Recognition (ICDAR), pages 1533-1538. IEEE, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.775, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[24] Harold Mouchere, Christian Viard-Gaudin, Richard Zanibbi, Utpal Garain, Dae Hwan Kim, and Jin Hyung Kim. Icdar 2013 crohme: Third international competition on recognition of online handwritten mathematical expressions. In International Conference on Document Analysis and Recognition (ICDAR), pages 1428-1432. IEEE, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[25] Harold Mouchere, Christian Viard-Gaudin, Richard Zanibbi, and Utpal Garain. Icfhr 2014 competition on recognition of on-line handwritten mathematical expressions (crohme" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19689" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "2014). In International Conference on Frontiers in Hand-writing Recognition (ICFHR), pages 791-796. IEEE, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.483, + 0.189 + ], + "angle": 0, + "content": "[26] Harold Mouchère, Christian Viard-Gaudin, Richard Zanibbi, and Utpal Garain. Icfhr2016 crohme: Competition on recognition of online handwritten mathematical expressions. In International Conference on Frontiers in Handwriting Recognition (ICFHR), pages 607-612. IEEE, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.482, + 0.258 + ], + "angle": 0, + "content": "[27] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.26, + 0.482, + 0.287 + ], + "angle": 0, + "content": "[28] Vik Paruchuri. Texify. https://github.com/VikParuchuri/texify, 2023. Accessed: 2024-2-29." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.289, + 0.483, + 0.343 + ], + "angle": 0, + "content": "[29] Josselin Somerville Roberts, Tony Lee, Chi Heem Wong, Michihiro Yasunaga, Yifan Mai, and Percy Liang. Image2struct: Benchmarking structure extraction for vision-language models. arXiv preprint arXiv:2410.22456, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.483, + 0.398 + ], + "angle": 0, + "content": "[30] Bin Wang, Zhuangcheng Gu, Chao Xu, Bo Zhang, Botian Shi, and Conghui He. Unimernet: A universal network for real-world mathematical expression recognition. arXiv.org, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.483, + 0.468 + ], + "angle": 0, + "content": "[31] Bin Wang, Chao Xu, Xiaomeng Zhao, Linke Ouyang, Fan Wu, Zhiyuan Zhao, Rui Xu, Kaiwen Liu, Yuan Qu, Fukai Shang, et al. Mineru: An open-source solution for precise document content extraction. arXiv preprint arXiv:2409.18839, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.483, + 0.525 + ], + "angle": 0, + "content": "[32] Zelun Wang and Jyh-Charn Liu. Translating math formula images to latex sequences using deep neural networks with sequence-level training. International Journal on Document Analysis and Recognition (IJDAR), 24(1):63-75, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.526, + 0.483, + 0.567 + ], + "angle": 0, + "content": "[33] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE TIP, 13(4):600-612, 2004." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.568, + 0.483, + 0.622 + ], + "angle": 0, + "content": "[34] Haoran Wei, Lingyu Kong, Jinyue Chen, Liang Zhao, Zheng Ge, Jinrong Yang, Jianjian Sun, Chunrui Han, and Xiangyu Zhang. Vary: Scaling up the vision vocabulary for large vision-language models. arXiv.org, 2312.06109, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.623, + 0.483, + 0.692 + ], + "angle": 0, + "content": "[35] Haoran Wei, Lingyu Kong, Jinyue Chen, Liang Zhao, Zheng Ge, Jinrong Yang, Jianjian Sun, Chunrui Han, and Xiangyu Zhang. Vary: Scaling up the vision vocabulary for large vision-language model. In European Conference on Computer Vision, pages 408-424. Springer, 2025." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.693, + 0.483, + 0.748 + ], + "angle": 0, + "content": "[36] Renqiu Xia, Bo Zhang, Haoyang Peng, Ning Liao, Peng Ye, Botian Shi, Junchi Yan, and Yu Qiao. Structchart: Perception, structuring, reasoning for visual chart understanding. arXiv preprint arXiv:2309.11268, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.749, + 0.483, + 0.83 + ], + "angle": 0, + "content": "[37] Renqiu Xia, Song Mao, Xiangchao Yan, Hongbin Zhou, Bo Zhang, Haoyang Peng, Jiahao Pi, Daocheng Fu, Wenjie Wu, Hancheng Ye, et al. Docgenome: An open largescale scientific document benchmark for training and testing multi-modal large language models. arXiv preprint arXiv:2406.11633, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[38] Renqiu Xia, Bo Zhang, Hancheng Ye, Xiangchao Yan, Qi Liu, Hongbin Zhou, Zijun Chen, Min Dou, Botian Shi, Junchi Yan, et al. Chartx & chartvlm: A versatile benchmark and foundation model for complicated chart reasoning. arXiv preprint arXiv:2402.12185, 2024." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.162 + ], + "angle": 0, + "content": "[39] Mingkun Yang, Yushuo Guan, Minghui Liao, Xin He, Kaigui Bian, Song Bai, Cong Yao, and Xiang Bai. Symmetry-constrained rectification network for scene text recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9147-9156, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.906, + 0.218 + ], + "angle": 0, + "content": "[40] Ya-Qi Yu, Minghui Liao, Jiwen Zhang, and Jihao Wu. Texthawk2: A large vision-language model excels in bilingualOCR and grounding with 16x fewer tokens. arXiv preprint arXiv:2410.05261, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.906, + 0.289 + ], + "angle": 0, + "content": "[41] Ye Yuan, Xiao Liu, Wondimu Dikubab, Hui Liu, Zhilong Ji, Zhongqin Wu, and Xiang Bai. Syntax-aware network for handwritten mathematical expression recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4553-4562, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.29, + 0.906, + 0.346 + ], + "angle": 0, + "content": "[42] Jianshu Zhang, Jun Du, Yongxin Yang, Yi-Zhe Song, Si Wei, and Lirong Dai. A tree-structured decoder for image-to-markup generation. In International Conference on Machine Learning (ICML), pages 11076–11085. PMLR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.347, + 0.906, + 0.417 + ], + "angle": 0, + "content": "[43] Wenqi Zhao, Liangcai Gao, Zuoyu Yan, Shuai Peng, Lin Du, and Ziyin Zhang. Handwritten mathematical expression recognition with bidirectionally trained transformer. In International Conference on Document Analysis and Recognition (ICDAR), pages 570-584. Springer, 2021." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19690" + } + ] +] \ No newline at end of file diff --git a/2025/Image Over Text_ Transforming Formula Recognition Evaluation with Character Detection Matching/e8290078-d6a1-49c9-b28d-cb4de981b7cd_origin.pdf b/2025/Image Over Text_ Transforming Formula Recognition Evaluation with Character Detection Matching/e8290078-d6a1-49c9-b28d-cb4de981b7cd_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ef4b5f9f106659559d6ba09c6eab118f5fe8f138 --- /dev/null +++ b/2025/Image Over Text_ Transforming Formula Recognition Evaluation with Character Detection Matching/e8290078-d6a1-49c9-b28d-cb4de981b7cd_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:38e2f67be44a5bcd5695d5284da0cb771ea372bfabc775a53ca740c653f95211 +size 916560 diff --git a/2025/Image Over Text_ Transforming Formula Recognition Evaluation with Character Detection Matching/full.md b/2025/Image Over Text_ Transforming Formula Recognition Evaluation with Character Detection Matching/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b80914b771d3c7074937e85ceecdc5d9a5365fba --- /dev/null +++ b/2025/Image Over Text_ Transforming Formula Recognition Evaluation with Character Detection Matching/full.md @@ -0,0 +1,401 @@ +# Image Over Text: Transforming Formula Recognition Evaluation with Character Detection Matching + +Bin Wang $^{1*}$ + +Rui Zhang + +Fan $\mathbf{W}\mathbf{u}^{1*}$ + +Renqiu Xia1,2 + +Linke Ouyang $^{1*}$ + +Botian $\mathbf{Shi}^1$ + +Zhuangcheng Gu + +Bo Zhang + +Conghui $\mathrm{He}^{1\dagger}$ + +$^{1}$ Shanghai Artificial Intelligence Laboratory $^{2}$ Shanghai Jiao Tong University + +# Abstract + +Formula recognition presents significant challenges due to the complicated structure and varied notation of mathematical expressions. Despite continuous advancements in formula recognition models, the evaluation metrics employed by these models, such as BLEU and Edit Distance, still exhibit notable limitations. They overlook the fact that the same formula has diverse representations and is highly sensitive to the distribution of training data, thereby causing unfairness in formula recognition evaluation. To this end, we propose a Character Detection Matching (CDM) metric, ensuring the evaluation objectivity by designing an image-level rather than a LaTeX-level metric score. Specifically, CDM renders both the model-predicted LaTeX and the ground-truth LaTeX formulas into image-formatted formulas, then employs visual feature extraction and localization techniques for precise character-level matching, incorporating spatial position information. Such a spatially-aware and character-matching method offers a more accurate and equitable evaluation compared with previous BLEU and Edit Distance metrics that rely solely on text-based character matching. Experimentally, we evaluated various formula recognition models using CDM, BLEU, and ExpRate metrics. Their results demonstrate that the CDM aligns more closely with human evaluation standards and provides a fairer comparison across different models by eliminating discrepancies caused by diverse formula representations. Code is available at https://github.com/opendatalab/UniMERNet/tree/main/cdm + +# 1. Introduction + +Mathematical formula recognition is crucial in document analysis as it directly impacts the scientific rigor and accuracy of the document content [7, 19, 31, 37]. Unlike + +standard Optical Character Recognition (OCR) technique, formula recognition presents unique challenges. Formulas often encompass multi-level symbols, subscripts, fractions, and other complicated structures, requiring models to comprehend spatial and structural relationships rather than just linear, sequential text [14, 19]. Besides, formulas exhibit representational diversity, meaning that the same formula can be expressed in multiple valid ways. + +In recent years, significant advancements in formula recognition [4, 8, 20, 28, 39, 41, 42] have been primarily driven by the continuous development of artificial intelligence technique [5, 35, 40]. Besides, commercial formula recognition software like Mathpix1 and the recently proposed UniMERNet [30] model have achieved impressive results in diverse real-world settings. Despite these advancements, the existing evaluation metrics [17, 27] for formula recognition still face some challenges. Commonly-used metrics such as BLEU [27] and Edit Distance [17] primarily rely on text-based character matching, which introduces several limitations as follows: + +(1) Low Metric Reliability. BLEU and Edit Distance are reliable for evaluating the quality of text-level similarity. However, the diversity in formula representations makes these text-level evaluation metrics inadequate for precisely reflecting formula recognition quality. For example, as shown in Figure 1 (Case 1), a model's prediction might render an image identical to the ground truth formula. However, due to the variations in formula expression styles, the evaluation results obtained using the ExpRate [8], BLEU, and Edit Distance may be somewhat misleading. +(2) Unfair Model Comparison. Current metrics may be susceptible to discrepancies between the distributions of training and testing data. As illustrated in Figure 1 (Case 1 and Case 2), a model may produce a correct prediction but score poorly due to representational differences from the ground truth, while an incorrect prediction might score higher if its representation aligns more closely with the test + +![](images/54908e7bb795b7b29794bef89c537ab213ad6dab3dcd000ef1bbf226b40119bd.jpg) +Figure 1. Illustration of the limitations of the existing metrics (ExpRate [8], BLEU [27], and Edit Distance [17]) and the advantages of the proposed CDM. Case 1: Due to different expression styles, the model prediction and Ground Truth (GT) may appear visually similar, but they receive low scores when evaluated using traditional metrics such as BLEU. Case 2: Errors in numeric predictions within formulas are essentially prediction errors. However, when evaluated using BLEU, such errors may receive higher metric scores due to the model predicting an expression style close to the GT. Case 3: Model predictions that are visually obviously incorrect may receive high scores using BLEU metric, which is inconsistent with human judgment standards. + +data distribution. + +(3) Lack of Intuitive Scoring. There can be a significant discrepancy between BLEU scores and human perception. For instance, in Figure 1 (Case 3), a model's prediction contains many errors, yet the BLEU score is as high as 0.907, which does not align with human judgment. + +To address these issues, we propose a novel evaluation metric for formula recognition: Character Detection Matching (CDM). The proposed CDM regards the formula recognition evaluation as an image-based object detection task, by converting both the predicted LaTeX and the groundtruth LaTeX formulas into the image-formatted formulas and treating each character as an independent target. This approach overcomes the challenges posed by the diverse expression styles of formulas and aligns more closely with human subjective evaluation standards. CDM offers the advantages as follows: 1) Accuracy and Reliable. By calculating metrics in the image space, CDM eliminates issues caused by different valid representations of the same formula, directly reflecting recognition accuracy and aligning more closely with human intuitive perception. 2) Fairness. CDM removes the high dependency on consistent data distribution between training and evaluation task, allowing for a fair comparison of different models based on their true recognition capabilities. Our contributions can be summarized as follows: + +- We perform a detailed analysis of the existing formula recognition evaluation metrics, highlighting the limitations of ExpRate and BLEU and their unreliability specifically for evaluating formula recognition tasks. + +- We introduce a novel evaluation metric, CDM, which assesses formula recognition quality by performing visual character matching between rendered images of predicted and ground-truth formulas, providing an intuitive and fair evaluation standard. +- We validate CDM's effectiveness through extensive experiments on various mainstream models and datasets, demonstrating its superiority over traditional metrics like BLEU in assessing formula recognition performance. + +# 2. Related Work + +# 2.1. Formula Recognition Algorithms + +Initially, researchers employ specific grammar rules to represent the spatial structure of formulas, including graph grammars [15], relational grammars [22], and probabilistic grammars [1, 2]. Besides, the CROHME competitions [16, 23-26] have promoted the development of handwritten formula recognition by incorporating deep learning algorithms. Key contributions include a neural encoder-decoder model with coarse-to-fine attention [8], a tree-structured decoder [42], and the Counting-Aware Network [18], which integrates a weakly-supervised counting module. The ABM network [3] employs mutual distillation and an Attention Aggregation Module, while a transformer-based decoder [43] simplifies model architecture. The Syntax-Aware Network (SAN) [41] models recognition as a tree traversal process, significantly improving accuracy for complex expressions. Overall, these models employ ExpRate [8] for formula recognition evaluation. + +In document information extraction [18, 29, 37, 38], Donut [12] directly converts input documents into structured outputs without using traditional OCR tools. Textify [28] and UniMERNet [30] are designed using Donut [12], utilizing more diverse datasets and data augmentation operations. Nougat [5] is designed to convert PDF documents from screenshot to Markdown format, making the document content (e.g. table and formula) easier to edit. These methods use BLEU [27] and Edit Distance [17] metrics for formula recognition evaluation. + +# 2.2. Formula Recognition Evaluation Metrics + +BLEU is initially proposed for machine translation tasks, matching standard and machine-translated texts using N-grams (sequences of N words) between the generated and the reference texts. It applies a brevity penalty factor to produce the final BLEU score [27]: + +$$ +B L E U = B P \cdot \exp \left(\sum_ {n = 1} ^ {N} w _ {n} \log p _ {n}\right), \tag {1} +$$ + +where BP is the brevity penalty factor, and $p_n$ is the N-gram match result, with $n$ ranging from 1 to 4. + +Edit Distance is also commonly-used metric to assess the similarity between the generated and the reference texts. It measures the number of insertions, deletions, or substitutions needed to transform one text into another, with a smaller Edit Distance indicating higher similarity [17]. + +ExpRate refers to the proportion of samples where the texts are exactly matched out of the total number of samples. Compared to BLEU and Edit Distance, ExpRate is coarser and more stringent in evaluation [18]. + +The above three metrics can effectively evaluate the textual differences between ground truth and reference, making them suitable for tasks requiring strict matches. BLEU and Edit Distance, in particular, provide a finer evaluation of text recognition capabilities compared to ExpRate, making them widely used in extensive text recognition tasks such as document recognition [5, 11]. These metrics are also applied to formula recognition, with most open-source models, such as Pix2Tex [4] and Texify, adopting them for evaluation and comparison. + +In addition to text-based metrics, image edit distance has been explored to measure the accuracy of predicted formulas [32]. Image processing metrics like MSE (Mean Squared Error) and SSIM [33] have also been considered. Structuring Chart-oriented Representation Metric (SCRM) [36] is designed to comprehensively evaluate the information represented by structured triplet representations. However, these metrics are better suited for natural images. For document images such as formula images, even slight character misalignments can result in significant penalties, making these metrics less suitable for formula recognition. + +# 3. Limitations of Current Metrics + +Although ExpRate, BLEU, and Edit Distance are widely used in formula evaluation tasks, they exhibit significant limitations in accurately reflecting formula recognition performance, particularly in scenarios where there are domain gaps between training and testing data distributions. The main reason is that a single formula can have multiple valid LaTeX representations, making the Ground Truth (GT) LaTeX non-unique, which introduces inherent flaws for the formula evaluation. + +As illustrated in Case 1 of Figure 1 earlier, the formula $(x + y) + z = x + (y + z)$ corresponds to the GT annotation " $\backslash$ left(x+y\rightleftarrows right)+z=x+ $\backslash$ left(y+z\rightleftarrows right)" When the model's prediction is "(x+y)+z=x+(y+z) ", the prediction is correct because the rendered formula image matches the GT image, despite different LaTeX syntax. Theoretically, the ExpRate/BLEU/Edit Distance results should be $1 / 1 / 0$ , indicating a correct instance. However, in practice, ExpRate is 0, BLEU is 0.449,and Edit Distance is 0.571,failing to accurately assess the formula's quality. + +The aforementioned issues make it challenging to objectively evaluate the performance of different formula recognition models. For instance, as illustrated in Case 2 of Figure 1, one character "z" is misrecognized as "2". The prediction is incorrect, and the ExpRate, BLEU, and Edit Distance metrics reflect this error. However, when compared to Case 1 where the model prediction is correct, the BLEU and Edit Distance metrics for the incorrect prediction in Case 2 are better than those for the correct prediction in Case 1. + +A LaTeX regularization method, which abstracts LaTeX code into a tree structure and standardizes elements, addresses LaTeX syntax diversity [8]. Pix2tex [4], Texify [28], and UniMERNet [30] use such regularization method as a preprocessing step before evaluation, which can solve part of the syntax inconsistency issue. For instance, "x^b_a", "x^{\hat{}}\{b\}_{-}\{a\} ", and "x_{-}\{a\}^{\hat{}}\{b\}" all compile to $x_{a}^{b}$ . Directly calculating BLEU scores would not correctly assess the model's prediction quality. Regularized code unifies these into a consistent format, such as always adding curly braces and arranging superscripts before subscripts, contributing to the fairness of subsequent metric calculations. However, regularization does not solve all LaTeX syntax diversity issues. Some symbols have multiple representations, such as " $\backslash$ leq" and " $\backslash$ le" both representing $\leq$ . Exhaustively listing these representations is challenging due to the huge LaTeX symbol library and many additional symbols provided by extension packages (e.g., ammath, amssymb). + +Overall, while regularization mitigates some issues, it does not fully address the inherent limitations of current metrics in evaluating formula recognition performance. This highlights the need for a more robust and comprehensive evaluation metric that can accurately reflect the quality + +![](images/135a3740aefa0df9fb3f27b344f99453f16cada6e7a97a334ed5dc1309427679.jpg) +Step 1: Element Localization + +1. LaTeX Source Normalization + +
r={\frac{1}{2}{\alpha}
{\beta}}|\sin\beta
\left( \right)\sigma\sigma\sigma-{2}
\sigma_{1}\left( \right)1
+ +2. Element Region Localization + +$$ +\begin{array}{c c} \frac {\text {r e n d e r e a c h}}{\text {t o k e n w i t h a}} & r = \frac {\alpha}{\beta} \big | \sin \beta \left(\sigma_ {2} + \sigma_ {1}\right) \big | \\ \text {d i f f e r e n t c o l o r} & \end{array} +$$ + +$$ +\begin{array}{c} r = \frac {\alpha}{\beta} | \sin \beta (\sigma_ {2} + \sigma_ {1}) | \\ \text {e x t r a c t e a c h} \\ \text {c o l o r a n d l o c a t e} \\ \text {t h e b o u n d i n g b o x} \end{array} r = \frac {\alpha}{\beta} | \sin \beta (\sigma_ {2} + \sigma_ {1}) | +$$ + +![](images/476845664a3358db013068b11bce6a22519ceb0eaea91ffb09aed9e18eb92e3e.jpg) + +![](images/fef9e67c4918b4b39f8cf2ed6809bb9b033dad3e1bb7223a4568b9655df9b2db.jpg) +Input + +![](images/309337165261ce94e6e8adc3a54c0274b598009d569e43604e9e0ce91d3480b6.jpg) +Step 3: Invalid Match Elimination + +![](images/a6bed2a0a6ca774c3e62edf7709e80e9b09f3360c8ef2bd0145a54fd5192eaff.jpg) +Step 2: Element Region Matching +1. Calculate Match Cost Between Every Element Pairs +Step 4: Metric Calculation + +![](images/44520171490c921152c241eb9a56460ee74d0c32aa01941c7d3bb3a0e0ec2288.jpg) +2. Find the Best Match by Hungarian Algorithm +Evaluation Metric + +![](images/2e6767b0c3ac6e1015c28c40261f78ef5437b67a707c7dff6097d8061027362d.jpg) +1. Token Consistency Check + +![](images/9b656f304f8b314129cfa7edfe5f8d6c233268eba74e037704e963b8acfa559e.jpg) +2. Position Relationship Consistency Check +Figure 2. Overview of the Character Detection Matching (CDM), consisting of four main stages. (1) Element Localization, where bounding boxes of individual elements are extracted. (2) Element Region Matching, which employs a bipartite graph matching method to pair prediction with ground truth elements. (3) Invalid Match Elimination, where inconsistent matches are discarded through token and positional relationship checks. (4) Metric Calculation, where matching accuracy is evaluated using the F1-Score and ExpRate@CDM. + +$$ +F 1 \text {s c o r e} = \frac {2 T P}{2 T P + F P + F N} +$$ + +- TP: number of matched bbox pairs +- FP: number of unmatched bboxes in prediction +FN: number of unmatched bboxes in GT + +$$ +T P = \left(\bigtriangledown\right) _ {n u m} = 1 3 +$$ + +$$ +F P = \left(\square\right) _ {\text {n u m}} ^ {\text {P r e d}} = 3 +$$ + +$$ +F N = \left( \begin{array}{l l} \square & T \\ 0 & n u m \end{array} \right) ^ {G T} = 3 +$$ + +$$ +F 1 s c o r e = \frac {2 \times 1 3}{2 \times 1 3 + 3 + 3} = 0. 8 1 2 5 +$$ + +of formula recognition across diverse representations. + +# 4. Character Detection Matching + +Due to the diversity of LaTeX expressions, text-based character-matching methods are unreliable for formula recognition evaluation. The basic idea of CDM is to compare the rendered images from LaTex text. If the image rendered from the predicted LaTeX source code matches the image rendered from the ground truth LaTeX source code, the formula is considered entirely correct. However, directly comparing the pixel values of the original and predicted formulas is not ideal. Any error or extra/missing character in the prediction can cause subsequent characters to be mismatched. Additionally, two similar formulas might have different layouts, with one being a single-line formula and the other a multi-line formula due to line breaks. Therefore, a more robust algorithm is needed to calculate the match between the predicted result and the ground truth image. + +To this end, we propose a metric that incorporates a bipartite matching step for element-level matching in images, + +providing a more intuitive assessment. As shown in Figure 2, the algorithm consists of four stages as follows. + +# 4.1. Element Localization + +First, the bounding boxes (bboxes) of each individual element in the rendered image are extracted, followed by the subsequent steps: + +LaTeX Source Normalization. LaTeX source codes of both the ground truth and predicted formulas are normalized, breaking them down into individual tokens such as "2", "a", "A", "\alpha", "\sin". Composite elements are decomposed into individual characters, e.g., "\frac{ab}{\alpha} is decomposed into \frac{a}{b}\)". + +Element Region Localization. To accurately detect character positions, we render each token in a unique color. We construct an RGB color list with a fixed interval of 15, creating a sequence from $(0,0,15)$ to $(255,255,255)$ . This generates $(255 / 15 + 1)^{3} = 5832$ distinct colors, sufficient for complex formulas. Each token is assigned a unique color using the LaTeX command "\mathsf{\\mathsf{m}}\mathsf{a}\mathsf{h}\mathsf{o}\mathsf{c}\mathsf{l}\mathsf{r}\mathsf{R}\mathsf{B}\mathsf{]} \{\mathsf{r},\mathsf{g},\mathsf{b}\} "$. After rendering, we apply image processing techniques to identify pixels match- + +ing each color, thereby determining the exact coordinates of each token. + +# 4.2. Element Region Matching + +In this stage, a bipartite matching method pairs the predicted elements with the corresponding ground truth elements. Based on the element localization, two sets are obtained for each formula: one for the ground truth independent elements $y$ and one for the predicted independent elements $\hat{y}$ . The number of independent elements in each set is $N_y$ and $N_{\hat{y}}$ , respectively, with $N = \min(N_y, N_{\hat{y}})$ being the number of elements in the smaller set. + +To measure the similarity between $y$ and $\hat{y}$ , we match elements in the two sets by identifying the corresponding ground truth element for each predicted element. We use the bipartite matching Hungarian algorithm [13], as described in DETR [6], to find a permutation $\hat{\sigma}$ that minimizes the total matching cost: + +$$ +\hat {\sigma} = \arg \min _ {\sigma \in S _ {N}} \sum_ {i = 1} ^ {N} L _ {\text {m a t c h}} \left(y _ {i}, \hat {y} _ {\sigma (i)}\right), \tag {2} +$$ + +$$ +L _ {\text {m a t c h}} = W _ {t} \times L _ {t} + W _ {p} \times L _ {p} + W _ {o} \times L _ {o}, \tag {3} +$$ + +where the matching cost $L_{match}$ is defined as a weighted sum of three components as introduced as follows: + +- Token Matching Cost $L_{t}$ : This component measures whether the tokens corresponding to two bounding boxes are the same. If they are identical, the cost is 0; if they are different, the cost is 1. For tokens that render identically but are different, such as " (" , "\leftarrow" , and "\bigig( ", the cost is 0.05 , which can be formulated as follows: + +$$ +L _ {\mathrm {t}} = \left\{ \begin{array}{l l} 0, & \text {i f} t _ {i} = \hat {t} _ {\sigma (i)}; \\ 0. 0 5, & \text {i f} t _ {i} \approx \hat {t} _ {\sigma (i)}; \\ 1, & \text {o t h e r w i s e}; \end{array} \right. \tag {4} +$$ + +where $\approx$ denotes tokens that differ but render identically. + +- Positional Proximity Cost $L_{p}$ : This component measures the proximity of the two bounding boxes' positions using the L1 norm of their coordinates, which can be formulated as follows: + +$$ +L _ {\mathrm {p}} = \frac {1}{D _ {b}} \times \| b _ {i} - \hat {b} _ {\sigma (i)} \| _ {1}, \tag {5} +$$ + +where $b = [x_1, y_1, x_2, y_2]$ , and $D_b$ is the dimension of the bounding box coordinates. + +- Order Similarity Cost $L_{o}$ : This measures the similarity of the token order in the original LaTeX source (an approximation of reading order). The order is normalized to the range [0, 1], and the L1 norm can be calculated as follows: + +$$ +L _ {\mathrm {o}} = \frac {1}{D _ {o}} \times \left\| o _ {i} - \hat {o} _ {\sigma (i)} \right\| _ {1}, \tag {6} +$$ + +where $o_i$ means order of $i$ -th token, with dimension $D_o$ . + +Overall, the weights $W_{t}, W_{p}, W_{o}$ are used to balance the contributions of the three components. By employing this comprehensive matching strategy, we ensure a more accurate and robust evaluation of the correspondence between the predicted and ground truth elements, thereby improving the overall assessment of formula recognition quality. + +# 4.3. Invalid Match Elimination + +After pairing the individual elements of the predicted result with the ground truth using the Hungarian matching algorithm, we need to verify these pairs and eliminate invalid matches. This process involves two steps: + +Token Consistency Check. Check whether the elements in each matched pair are consistent in terms of characters. If they are inconsistent, discard the match. + +Position Relationship Consistency Check. The relative positions of elements in mathematical formulas are crucial. For instance, in the expressions $2^3$ and $3^2$ , bipartite matching might pair 2 with 2 and 3 with 3, but their meanings and visual representations are entirely different. Thus, we need to check the consistency of the positional relationships within the matched pairs. We treat each element in the matched pair as a bounding box and analyze their relative positions. Specifically, we assume an affine transformation between the ground truth and predicted elements: + +$$ +\hat {b} _ {\sigma (i)} = \mathbf {A} \left(b _ {i}\right), \tag {7} +$$ + +where $\mathbf{A}$ is the affine transformation matrix. To identify inconsistent match pairs, we detect pairs that do not conform to this transformation relationship. We employ the RANSAC algorithm [9] for this purpose. RANSAC can determine the optimal transformation matrix $\mathbf{A}$ in the presence of noise. Given that formulas are usually horizontally arranged during rendering, we fix the rotation angle in the transformation matrix to 0, considering only translation and scaling. This approach not only improves the convergence speed of the RANSAC algorithm but also enhances the final matching accuracy. + +To account for line-breaking effects in formulas, we perform multiple rounds of RANSAC iterations to ensure that as many matched pairs as possible conform to the transformation relationship. After several iterations, matched pairs that still do not conform to the transformation relationship are considered incorrect and are eliminated. + +The above two steps effectively eliminate invalid match pairs, ensuring more accurate final matching results. + +# 4.4. Metric Calculation + +We use the F1-Score as the default metric for evaluating CDM (Character Detection Metric), defined as: + +$$ +C D M = \frac {2 \times T P}{2 \times T P + F P + F N}, \tag {8} +$$ + +![](images/d8ff569f6b6f08e795bdf71a951aa6d5af2d34f01a954ab7c802c34bf22c3cb8.jpg) +(a) CDM distribution on UniMER-Test dataset + +![](images/acfb733409878c8353891713f6b8d6cbda9bd3dfabcd24eddb073964b56ddf48.jpg) +(b) Examples by CDM score range +Figure 3. CDM Score Distribution and Example Analysis. (a) CDM score distribution for Mathpix, UniMERNet, Texify, and Pix2Tex on the UniMER-Test dataset. (b) Example analysis of Pix2Tex predictions across different CDM score ranges. + +where $TP$ denotes true positives, $FP$ denotes false positives, and $FN$ denotes false negatives. + +To further evaluate the accuracy of formula recognition, we introduce the ExpRate@CDM metric, defined as: + +$$ +\operatorname {E x p R a t e} @ C D M = \frac {\sum_ {i = 1} ^ {N} \mathbb {I} \left(C D M _ {i} = 1\right)}{N}, \tag {9} +$$ + +where $\mathbb{I}$ is the indicator function that equals 1 if $CDM_{i} = 1$ and 0 otherwise, and $N$ is the total number of formulas. This metric represents the proportion of formulas for which the model's prediction results are perfectly matched. Essentially, ExpRate@CDM serves as a precise version of the ExpRate metric specifically for formula recognition. + +# 5. Experiments + +# 5.1. Models and Data + +We validate the CDM metric by evaluating several mainstream formula recognition models using both subjective impressions and objective metrics. The models include open-source UniMERNet [30], Texify [28], Pix2tex [4], and the commercial Mathpix API, all tested on the UniMERTest dataset. Besides, we evaluate document-level models, such as the open-source Nougat [5] and the commercial GPT-4o [10]. Vary [34] and StrucTexTv3 [21] are excluded as they are currently unavailable. + +UniMER-Test Dataset. The dataset includes 23,757 formula samples, categorized into Simple Printed Expressions (SPE), Complex Printed Expressions (CPE), Screenshot Expressions (SCE), and Handwritten Expressions (HWE). We use these categories to conduct the model evaluation. + +Tiny-Doc-Math Dataset. To evaluate document-level recognition, we construct the Tiny-Doc-Math dataset, consisting of arXiv papers in mathematics and computer science, published after June 2024, to ensure that they are not + +in the training data of the compared models. We obtain Latex code and corresponding PDFs, match displayed equations using regular expressions, and manually verify them. Overall, the dataset includes 12 PDFs, totaling 196 pages and 437 formulas. + +This validation set includes both formula-level and document-level evaluations: + +- Formula-level: Using single rendered formula images as input, we evaluate Mathpix, Pix2Tex, and UniMERNet. These models accept cropped formula images as input, and we compare the model outputs with the ground truth to compute relevant metrics. +- Document-level: Using PDFs or images as input, we evaluate Nougat, GPT-4o, and Mathpix, which can convert entire PDF pages into Markdown format. We match the displayed equations in the model outputs using regular expressions and compare them with the ground truth LaTeX formulas to compute relevant metrics. + +# 5.2. Credibility Assessment of CDM + +# 5.2.1. Rendering Success Rate + +The CDM metric relies on the successful rendering of formula images. For models that fail to render images, we assign a CDM score of 0, as rendering failures indicate that the predicted LaTeX code lacks critical elements. The rendering success rates on the UniMER-Test dataset for Pix2tex, Texify, UniMERNet, and Mathpix are $96.63\%$ , $94.77\%$ , $99.71\%$ , and $97.82\%$ , respectively, ensuring the applicability and reliability of the CDM metric. + +# 5.2.2. User Preference Evaluation + +We analyze the distribution of CDM scores for four models on the UniMER-Test dataset. As shown in Figure 3a, Math-pix and UniMERNet perform well in terms of CDM scores. We conduct a detailed analysis of the Pix2Tex model by randomly selecting samples from different score ranges to + +![](images/bd60bd570cb6355ae0a2f0e3e187fd7db70f4f52099ef7aab3291bafd9d71a37.jpg) +(a) Human Evaluation Preferences + +![](images/2e973c968c75f8b1b534c2d12b568ff1507ad9f36f14a1a99310c9ce6eba7abb.jpg) +(b) BLEU vs.CDM Score Distribution + +![](images/e69b68ade6a5bf1d4311efebde234b734e4b690aca3b0a6b464e40808d14460d.jpg) +(c) Impact of Formula Writing Styles + +![](images/7314099a976ec80cb10c028ef7556bb1382df3225a593df61b5979804041de0f.jpg) +Figure 4. Human evaluation and metric comparison. (a) Distribution of human preferences. (b) Distribution of cases preferring BLEU or CDM across different score ranges. (c) Impact of formula writing styles on BLEU, ExpRate and CDM scores. +Figure 5. Annotation interface for evaluating user preferences between CDM and BLEU scores. Annotators chose from four options: "Score A is better", "Score B is better", "Both scores are equally good", or "Neither score is good". + +evaluate if the prediction quality corresponds to the CDM scores. The analysis in Figure 3b shows that the CDM scores effectively reflect formula quality, with higher scores indicating fewer errors. + +To verify the consistency between the CDM metric and human evaluation, we select 1,008 CDM scores from Pix2Tex predictions, ensuring a balanced score distribution. We design an annotation interface (shown in Figure 5), displaying a ground truth label and the corresponding predicted LaTeX rendered image. Annotators choose between ScoreA, ScoreB, Both (credible), and Neither (credible). ScoreA and ScoreB correspond to the BLEU and CDM scores, respectively, but their order is randomized. For more details, please refer to the supplementary materials. + +The results in Figure 4a show that $64\%$ of participants prefer the CDM metric, and $32\%$ consider both metrics good. This indicates a $96\%$ consistency between the CDM metric and human evaluation, demonstrating its reliability. + +
ModelExpRateExpRate@CDMBLEUCDM
Pix2tex0.12370.29100.40800.6360
Texify0.22880.49500.58900.7550
Matphix0.26100.50000.80670.9510
UniMERNet0.47990.81100.84250.9680
+ +Table 1. UniMER-Test evaluation results. Comparison of mainstream models using different metrics. + +Figure 4b compares the number of cases where BLEU or CDM is preferred across different score ranges, showing that CDM consistently outperforms BLEU. + +# 5.2.3. Objective Stability Assessment + +To evaluate the impact of formula writing styles on the CDM and BLEU metrics, we randomly select 50 formulas with LaTeX source code and rewrite each formula five times using GPT-4, generating 250 additional formulas. We manually verify these formulas to ensure their rendered results are identical to the original 50 formulas. Using the initial LaTeX source code as the ground truth, we analyze the score distribution of the BLEU and CDM metrics. As shown in Figure 4c, the CDM metric is unaffected by style changes, with all samples scoring 1. In contrast, the BLEU metric's scores are dispersed, making it unsuitable for formula evaluation. The CDM metric remains robust and reliable despite formatting changes. + +# 5.3. Evaluation of Mainstream Models + +We conduct a detailed evaluation of mainstream models using both the CDM and BLEU metric. Note that all BLEU metric in this paper have been normalized [4, 8]. However, as discussed in the limitation section, normalization operations cannot address all issues, which will be evident in the following experiments. + +# 5.3.1. UniMER-Test Evaluation + +As shown in Table 1, the evaluation results of the four models on the UniMER-Test dataset indicate that the quality + +
MethodSPECPEHWESCE
BLEU ↑CDM ↑BLEU ↑CDM ↑BLEU ↑CDM ↑BLEU ↑CDM ↑
Pix2tex [4]0.87300.96190.65500.64890.01200.24530.09200.6762
Texify [28]0.90600.98520.69000.70410.34100.52690.42000.7932
Mathpix0.79200.97290.80610.96710.80600.93180.81820.9238
UniMERNet [30]0.91700.99460.91600.97070.92100.95300.61600.9461
+ +Table 2. UniMER-Test subset evaluation results. Analysis of anomalies in BLEU and CDM metrics across different subsets. + +
Image TypeModelBLEUCDMExpRate@CDM
FormulaPix2tex0.46480.744400.3684
GPT-4o0.64310.73300.4324
UniMERNet0.60560.93960.6887
Mathpix0.61120.94800.2105
DocumentGPT-4o0.34110.65020.1670
Nougat0.58970.83260.6086
Mathpix0.59390.95670.6292
+ +Table 3. Tiny-Doc-Math Evaluation Results. Performance of mainstream models on cropped formula inputs and document-level screenshots using BLEU and CDM metrics. + +ranges from low to high as follows: Pix2Tex, Texify, Mathpix, and UniMERNet, based on both BLEU and CDM metrics. ExpRate@CDM clearly shows the proportion of completely correct predictions for each model, indicating that the text character-based ExpRate is unreliable. + +From the results in Table 1, it appears that the trends of the BLEU and CDM metrics are consistent. To verify the reliability of using the BLEU metric for model comparison, we further present evaluation results on the UniMERTest subsets. As shown in Table 2, we observe two notable anomalies: Firstly, in the SCE subset, when comparing the quality of the Mathpix and UniMERNet models, the BLEU and CDM metrics provide opposite conclusions. A detailed review of the UniMERNet paper reveals that the SCE subset was annotated based on Mathpix and then manually corrected. This means that the expression style of the SCE formulas is more consistent with Mathpix. Consequently, even though the CDM metric indicates that UniMERNet has better actual model quality, the BLEU metric, influenced by the expression style, suggests that Mathpix is superior. Secondly, for the Pix2Tex model, the BLEU metric is very low on the HWE and SCE subsets but performs well on the SPE and CPE subsets. This discrepancy arises because the Pix2Tex training set includes a large number of printed formulas from arXiv and lacks data in the HWE and SCE styles. + +These anomalies clearly illustrate the limitations of the BLEU metric in evaluating the quality of formula recognition models. In contrast, the CDM metric proposed in this paper is fair and intuitive. + +# 5.3.2. Tiny-Doc-Math Evaluation + +The evaluation results of Tiny-Doc-Math are shown in Table 3. For cropped formula inputs (formula-level), all four models perform reasonably well, with CDM scores above 0.7. Notably, the current leading multimodal large model GPT-4o has the highest BLEU score among the four models but the lowest CDM score. This discrepancy indicates that the BLEU metric may not be reliable, suggesting that the formula recognition accuracy of GPT-4o still has room for improvement, lagging behind traditional SOTA models. Additionally, although Mathpix has the highest CDM score, only $21.05\%$ of the formulas are completely accurate. Manual verification revealed that many formulas are missing commas or periods at the end. + +When the input is document-level screenshots, the models output the recognition results for the entire document (not just the formulas). Evaluation is conducted by matching the recognized block formulas. In this scenario, it can be observed that the accuracy of GPT-4o further decreases. In contrast, Mathpix and Nougat perform better, but even the document multimodal large model Nougat only achieves a CDM score of 0.8326. This indicates that there is still significant room for improvement in document-level recognition models. Mathpix remains the best performer, with a fully correct formula rate of $62.92\%$ . The accuracy of document-level recognition is crucial for advanced document understanding tasks like scientific knowledge Q&A, and CDM provides an excellent standard for selecting formula models and offers direction for improving formula recognition. + +# 6. Conclusion + +In this paper, we introduced Character Detection Matching (CDM), a novel evaluation metric for formula recognition task. CDM addresses the shortcomings of the existing metrics in formula recognition by utilizing spatial character matching, overcoming the challenges posed by diverse formula representations. Comprehensive evaluations on different models and datasets demonstrate CDM's superiority in precisely reflecting recognition quality, showing more intuitive assessment and paving the way for future research and improvements in the formula recognition field. + +# 7. Acknowledgments + +This project was supported by National Key R&D Program of China (NO.2022ZD0160102) and Shanghai Artificial Intelligence Laboratory. + +# References + +[1] Francisco Álvaro, Joan-Andreu Sánchez, and José-Miguel Benédí. An integrated grammar-based approach for mathematical expression recognition. Pattern Recognition, 51: 135–147, 2016. +[2] Ahmad-Montaser Awal, Harold Mouchere, and Christian Viard-Gaudin. A global learning approach for an online handwritten mathematical expression recognition system. Pattern Recognition Letters, 35:68-77, 2014. +[3] Xiaohang Bian, Bo Qin, Xiaozhe Xin, Jianwu Li, Xuefeng Su, and Yanfeng Wang. Handwritten mathematical expression recognition via attention aggregation based bidirectional mutual learning. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 113-121, 2022. +[4] Lukas Blecher. pix2tex - latexOCR. https://github.com/lukas-blecher/LaTeX-OCR, 2022. Accessed: 2024-2-29. +[5] Lukas Blecher, Guillem Cucurull, Thomas Scialom, and Robert Stojnic. Nougat: Neural optical understanding for academic documents. arXiv.org, 2308.13418, 2023. +[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision (ECCV), pages 213-229. Springer, 2020. +[7] Hiuyi Cheng, Peirong Zhang, Sihang Wu, Jiaxin Zhang, Qiyuan Zhu, Zecheng Xie, Jing Li, Kai Ding, and Lianwen Jin. M6doc: A large-scale multi-format, multi-type, multi-Layout, multi-language, multi-annotation category dataset for modern document layout analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15138–15147, 2023. +[8] Yuntian Deng, Anssi Kanervisto, Jeffrey Ling, and Alexander M Rush. Image-to-markup generation with coarse-to-fine attention. In International Conference on Machine Learning (ICML), pages 980-989. PMLR, 2017. +[9] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. +[10] GPT-4o. Gpt-4o. https://openai.com/index/hello-gpt-4o/, 2024. Accessed: 2024-08-15. +[11] Mingxin Huang, Yuliang Liu, Dingkang Liang, Lianwen Jin, and Xiang Bai. Mini-monkey: Alleviate the sawtooth effect by multi-scale adaptive cropping. arXiv.org, 2024. +[12] Geewook Kim, Teakgyu Hong, Moonbin Yim, JeongYeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sangdoo Yun, Dongyoon Han, and Seunghyun Park. Ocr-free document understanding transformer. In European Confer + +ence on Computer Vision (ECCV), pages 498-517. Springer, 2022. +[13] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. +[14] Vinay Kukreja et al. Recent trends in mathematical expressions recognition: An lda-based analysis. Expert Systems with Applications, 213:119028, 2023. +[15] Stéphane Lavirotte and Loïc Pottier. Mathematical formula recognition using graph grammar. In Document Recognition V, San Jose, CA, USA, January 24, 1998, 1998. +[16] Anh Duc Le, Bipin Indurkhya, and Masaki Nakagawa. Pattern generation strategies for improving recognition of handwritten mathematical expressions. Pattern Recognition Letters, 128:255-262, 2019. +[17] Vladimir I Levenshtein et al. Binary codes capable of correcting deletions, insertions, and reversals. In Soviet physics doklady, pages 707-710. Soviet Union, 1966. +[18] Bohan Li, Ye Yuan, Dingkang Liang, Xiao Liu, Zhilong Ji, Jinfeng Bai, Wenyu Liu, and Xiang Bai. When counting meets hmer: counting-aware network for handwritten mathematical expression recognition. In European Conference on Computer Vision (ECCV), pages 197-214. Springer, 2022. +[19] Minghao Li, Yiheng Xu, Lei Cui, Shaohan Huang, Furu Wei, Zhoujun Li, and Ming Zhou. Docbank: A benchmark dataset for document layout analysis. arXiv preprint arXiv:2006.01038, 2020. +[20] Minghui Liao, Guan Pang, Jing Huang, Tal Hassner, and Xiang Bai. Mask textspotter v3: Segmentation proposal network for robust scene text spotting. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XI 16, pages 706-722. Springer, 2020. +[21] Pengyuan Lyu, Yulin Li, Hao Zhou, Weihong Ma, Xingyu Wan, Qunyi Xie, Liang Wu, Chengquan Zhang, Kun Yao, Errui Ding, et al. Structextv3: An efficient vision-language model for text-rich image perception, comprehension, and beyond. arXiv.org, 2024. +[22] Scott MacLean and George Labahn. A new approach for recognizing handwritten mathematics using relational grammars and fuzzy sets. International Journal on Document Analysis and Recognition (IJDAR), 16:139-163, 2013. +[23] Mahshad Mahdavi, Richard Zanibbi, Harold Mouchere, Christian Viard-Gaudin, and Utpal Garain. Icdar 2019 crohme+ tfd: Competition on recognition of handwritten mathematical expressions and typeset formula detection. In International Conference on Document Analysis and Recognition (ICDAR), pages 1533-1538. IEEE, 2019. +[24] Harold Mouchere, Christian Viard-Gaudin, Richard Zanibbi, Utpal Garain, Dae Hwan Kim, and Jin Hyung Kim. Icdar 2013 crohme: Third international competition on recognition of online handwritten mathematical expressions. In International Conference on Document Analysis and Recognition (ICDAR), pages 1428-1432. IEEE, 2013. +[25] Harold Mouchere, Christian Viard-Gaudin, Richard Zanibbi, and Utpal Garain. Icfhr 2014 competition on recognition of on-line handwritten mathematical expressions (crohme + +2014). In International Conference on Frontiers in Hand-writing Recognition (ICFHR), pages 791-796. IEEE, 2014. +[26] Harold Mouchère, Christian Viard-Gaudin, Richard Zanibbi, and Utpal Garain. Icfhr2016 crohme: Competition on recognition of online handwritten mathematical expressions. In International Conference on Frontiers in Handwriting Recognition (ICFHR), pages 607-612. IEEE, 2016. +[27] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002. +[28] Vik Paruchuri. Texify. https://github.com/VikParuchuri/texify, 2023. Accessed: 2024-2-29. +[29] Josselin Somerville Roberts, Tony Lee, Chi Heem Wong, Michihiro Yasunaga, Yifan Mai, and Percy Liang. Image2struct: Benchmarking structure extraction for vision-language models. arXiv preprint arXiv:2410.22456, 2024. +[30] Bin Wang, Zhuangcheng Gu, Chao Xu, Bo Zhang, Botian Shi, and Conghui He. Unimernet: A universal network for real-world mathematical expression recognition. arXiv.org, 2024. +[31] Bin Wang, Chao Xu, Xiaomeng Zhao, Linke Ouyang, Fan Wu, Zhiyuan Zhao, Rui Xu, Kaiwen Liu, Yuan Qu, Fukai Shang, et al. Mineru: An open-source solution for precise document content extraction. arXiv preprint arXiv:2409.18839, 2024. +[32] Zelun Wang and Jyh-Charn Liu. Translating math formula images to latex sequences using deep neural networks with sequence-level training. International Journal on Document Analysis and Recognition (IJDAR), 24(1):63-75, 2021. +[33] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE TIP, 13(4):600-612, 2004. +[34] Haoran Wei, Lingyu Kong, Jinyue Chen, Liang Zhao, Zheng Ge, Jinrong Yang, Jianjian Sun, Chunrui Han, and Xiangyu Zhang. Vary: Scaling up the vision vocabulary for large vision-language models. arXiv.org, 2312.06109, 2023. +[35] Haoran Wei, Lingyu Kong, Jinyue Chen, Liang Zhao, Zheng Ge, Jinrong Yang, Jianjian Sun, Chunrui Han, and Xiangyu Zhang. Vary: Scaling up the vision vocabulary for large vision-language model. In European Conference on Computer Vision, pages 408-424. Springer, 2025. +[36] Renqiu Xia, Bo Zhang, Haoyang Peng, Ning Liao, Peng Ye, Botian Shi, Junchi Yan, and Yu Qiao. Structchart: Perception, structuring, reasoning for visual chart understanding. arXiv preprint arXiv:2309.11268, 2023. +[37] Renqiu Xia, Song Mao, Xiangchao Yan, Hongbin Zhou, Bo Zhang, Haoyang Peng, Jiahao Pi, Daocheng Fu, Wenjie Wu, Hancheng Ye, et al. Docgenome: An open largescale scientific document benchmark for training and testing multi-modal large language models. arXiv preprint arXiv:2406.11633, 2024. +[38] Renqiu Xia, Bo Zhang, Hancheng Ye, Xiangchao Yan, Qi Liu, Hongbin Zhou, Zijun Chen, Min Dou, Botian Shi, Junchi Yan, et al. Chartx & chartvlm: A versatile benchmark and foundation model for complicated chart reasoning. arXiv preprint arXiv:2402.12185, 2024. + +[39] Mingkun Yang, Yushuo Guan, Minghui Liao, Xin He, Kaigui Bian, Song Bai, Cong Yao, and Xiang Bai. Symmetry-constrained rectification network for scene text recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9147-9156, 2019. +[40] Ya-Qi Yu, Minghui Liao, Jiwen Zhang, and Jihao Wu. Texthawk2: A large vision-language model excels in bilingualOCR and grounding with 16x fewer tokens. arXiv preprint arXiv:2410.05261, 2024. +[41] Ye Yuan, Xiao Liu, Wondimu Dikubab, Hui Liu, Zhilong Ji, Zhongqin Wu, and Xiang Bai. Syntax-aware network for handwritten mathematical expression recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4553-4562, 2022. +[42] Jianshu Zhang, Jun Du, Yongxin Yang, Yi-Zhe Song, Si Wei, and Lirong Dai. A tree-structured decoder for image-to-markup generation. In International Conference on Machine Learning (ICML), pages 11076–11085. PMLR, 2020. +[43] Wenqi Zhao, Liangcai Gao, Zuoyu Yan, Shuai Peng, Lin Du, and Ziyin Zhang. Handwritten mathematical expression recognition with bidirectionally trained transformer. In International Conference on Document Analysis and Recognition (ICDAR), pages 570-584. Springer, 2021. \ No newline at end of file diff --git a/2025/Image Over Text_ Transforming Formula Recognition Evaluation with Character Detection Matching/images.zip b/2025/Image Over Text_ Transforming Formula Recognition Evaluation with Character Detection Matching/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..03f544e9124690f91fe08d07403273c01d864240 --- /dev/null +++ b/2025/Image Over Text_ Transforming Formula Recognition Evaluation with Character Detection Matching/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f48b46aca0ce96c377655ee2040fe39ae077958b620a2a370a21f46fcc43a68d +size 435037 diff --git a/2025/Image Over Text_ Transforming Formula Recognition Evaluation with Character Detection Matching/layout.json b/2025/Image Over Text_ Transforming Formula Recognition Evaluation with Character Detection Matching/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f15e9828e0cb2684393932135acc6db84e53b8cd --- /dev/null +++ b/2025/Image Over Text_ Transforming Formula Recognition Evaluation with Character Detection Matching/layout.json @@ -0,0 +1,9232 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 88, + 102, + 523, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 102, + 523, + 140 + ], + "spans": [ + { + "bbox": [ + 88, + 102, + 523, + 140 + ], + "type": "text", + "content": "Image Over Text: Transforming Formula Recognition Evaluation with Character Detection Matching" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 160, + 190, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 190, + 175 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 190, + 175 + ], + "type": "text", + "content": "Bin Wang" + }, + { + "bbox": [ + 132, + 160, + 190, + 175 + ], + "type": "inline_equation", + "content": "^{1*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 113, + 175, + 170, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 175, + 170, + 190 + ], + "spans": [ + { + "bbox": [ + 113, + 175, + 170, + 190 + ], + "type": "text", + "content": "Rui Zhang" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 208, + 161, + 257, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 161, + 257, + 174 + ], + "spans": [ + { + "bbox": [ + 208, + 161, + 257, + 174 + ], + "type": "text", + "content": "Fan " + }, + { + "bbox": [ + 208, + 161, + 257, + 174 + ], + "type": "inline_equation", + "content": "\\mathbf{W}\\mathbf{u}^{1*}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 193, + 175, + 260, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 193, + 175, + 260, + 190 + ], + "spans": [ + { + "bbox": [ + 193, + 175, + 260, + 190 + ], + "type": "text", + "content": "Renqiu Xia1,2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 271, + 161, + 350, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 271, + 161, + 350, + 175 + ], + "spans": [ + { + "bbox": [ + 271, + 161, + 350, + 175 + ], + "type": "text", + "content": "Linke Ouyang" + }, + { + "bbox": [ + 271, + 161, + 350, + 175 + ], + "type": "inline_equation", + "content": "^{1*}" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 279, + 175, + 335, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 175, + 335, + 188 + ], + "spans": [ + { + "bbox": [ + 279, + 175, + 335, + 188 + ], + "type": "text", + "content": "Botian " + }, + { + "bbox": [ + 279, + 175, + 335, + 188 + ], + "type": "inline_equation", + "content": "\\mathbf{Shi}^1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 368, + 161, + 456, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 161, + 456, + 175 + ], + "spans": [ + { + "bbox": [ + 368, + 161, + 456, + 175 + ], + "type": "text", + "content": "Zhuangcheng Gu" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 355, + 175, + 408, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 175, + 408, + 190 + ], + "spans": [ + { + "bbox": [ + 355, + 175, + 408, + 190 + ], + "type": "text", + "content": "Bo Zhang" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 428, + 175, + 496, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 428, + 175, + 496, + 190 + ], + "spans": [ + { + "bbox": [ + 428, + 175, + 496, + 190 + ], + "type": "text", + "content": "Conghui " + }, + { + "bbox": [ + 428, + 175, + 496, + 190 + ], + "type": "inline_equation", + "content": "\\mathrm{He}^{1\\dagger}" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 111, + 202, + 499, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 202, + 499, + 217 + ], + "spans": [ + { + "bbox": [ + 111, + 202, + 499, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 111, + 202, + 499, + 217 + ], + "type": "text", + "content": "Shanghai Artificial Intelligence Laboratory " + }, + { + "bbox": [ + 111, + 202, + 499, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 111, + 202, + 499, + 217 + ], + "type": "text", + "content": "Shanghai Jiao Tong University" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "spans": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 270, + 296, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 296, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 296, + 605 + ], + "type": "text", + "content": "Formula recognition presents significant challenges due to the complicated structure and varied notation of mathematical expressions. Despite continuous advancements in formula recognition models, the evaluation metrics employed by these models, such as BLEU and Edit Distance, still exhibit notable limitations. They overlook the fact that the same formula has diverse representations and is highly sensitive to the distribution of training data, thereby causing unfairness in formula recognition evaluation. To this end, we propose a Character Detection Matching (CDM) metric, ensuring the evaluation objectivity by designing an image-level rather than a LaTeX-level metric score. Specifically, CDM renders both the model-predicted LaTeX and the ground-truth LaTeX formulas into image-formatted formulas, then employs visual feature extraction and localization techniques for precise character-level matching, incorporating spatial position information. Such a spatially-aware and character-matching method offers a more accurate and equitable evaluation compared with previous BLEU and Edit Distance metrics that rely solely on text-based character matching. Experimentally, we evaluated various formula recognition models using CDM, BLEU, and ExpRate metrics. Their results demonstrate that the CDM aligns more closely with human evaluation standards and provides a fairer comparison across different models by eliminating discrepancies caused by diverse formula representations. Code is available at https://github.com/opendatalab/UniMERNet/tree/main/cdm" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 639, + 135, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 639, + 135, + 651 + ], + "spans": [ + { + "bbox": [ + 56, + 639, + 135, + 651 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 659, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 296, + 696 + ], + "type": "text", + "content": "Mathematical formula recognition is crucial in document analysis as it directly impacts the scientific rigor and accuracy of the document content [7, 19, 31, 37]. Unlike" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 246, + 555, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 246, + 555, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 246, + 555, + 342 + ], + "type": "text", + "content": "standard Optical Character Recognition (OCR) technique, formula recognition presents unique challenges. Formulas often encompass multi-level symbols, subscripts, fractions, and other complicated structures, requiring models to comprehend spatial and structural relationships rather than just linear, sequential text [14, 19]. Besides, formulas exhibit representational diversity, meaning that the same formula can be expressed in multiple valid ways." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 342, + 556, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 556, + 486 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 556, + 486 + ], + "type": "text", + "content": "In recent years, significant advancements in formula recognition [4, 8, 20, 28, 39, 41, 42] have been primarily driven by the continuous development of artificial intelligence technique [5, 35, 40]. Besides, commercial formula recognition software like Mathpix1 and the recently proposed UniMERNet [30] model have achieved impressive results in diverse real-world settings. Despite these advancements, the existing evaluation metrics [17, 27] for formula recognition still face some challenges. Commonly-used metrics such as BLEU [27] and Edit Distance [17] primarily rely on text-based character matching, which introduces several limitations as follows:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 487, + 556, + 692 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 313, + 487, + 556, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 556, + 607 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 556, + 607 + ], + "type": "text", + "content": "(1) Low Metric Reliability. BLEU and Edit Distance are reliable for evaluating the quality of text-level similarity. However, the diversity in formula representations makes these text-level evaluation metrics inadequate for precisely reflecting formula recognition quality. For example, as shown in Figure 1 (Case 1), a model's prediction might render an image identical to the ground truth formula. However, due to the variations in formula expression styles, the evaluation results obtained using the ExpRate [8], BLEU, and Edit Distance may be somewhat misleading." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 608, + 556, + 692 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 608, + 556, + 692 + ], + "spans": [ + { + "bbox": [ + 313, + 608, + 556, + 692 + ], + "type": "text", + "content": "(2) Unfair Model Comparison. Current metrics may be susceptible to discrepancies between the distributions of training and testing data. As illustrated in Figure 1 (Case 1 and Case 2), a model may produce a correct prediction but score poorly due to representational differences from the ground truth, while an incorrect prediction might score higher if its representation aligns more closely with the test" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 295, + 713 + ], + "type": "text", + "content": "*Equal contribution. †Corresponding author: heconghui@pjlab.org.cn" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 325, + 702, + 511, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 702, + 511, + 712 + ], + "spans": [ + { + "bbox": [ + 325, + 702, + 511, + 712 + ], + "type": "text", + "content": "r={\\frac{1}{2}{\\alpha}{\\beta}}|\\sin\\beta\\left( \\right)\\sigma\\sigma\\sigma-{2}\\sigma_{1}\\left( \\right)1", + "image_path": "6918082fca78ec0652b05161fad99521bd6814cbb3d07cfe1b642b4f573b3cfb.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 72, + 224, + 166, + 233 + ], + "lines": [ + { + "bbox": [ + 72, + 224, + 166, + 233 + ], + "spans": [ + { + "bbox": [ + 72, + 224, + 166, + 233 + ], + "type": "text", + "content": "2. Element Region Localization" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 70, + 243, + 198, + 265 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 243, + 198, + 265 + ], + "spans": [ + { + "bbox": [ + 70, + 243, + 198, + 265 + ], + "type": "interline_equation", + "content": "\\begin{array}{c c} \\frac {\\text {r e n d e r e a c h}}{\\text {t o k e n w i t h a}} & r = \\frac {\\alpha}{\\beta} \\big | \\sin \\beta \\left(\\sigma_ {2} + \\sigma_ {1}\\right) \\big | \\\\ \\text {d i f f e r e n t c o l o r} & \\end{array}", + "image_path": "0b4ce9c5b90ef7a1d8f50a8b2310773558a4f65fa92748150daff2619437836d.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 70, + 269, + 198, + 307 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 269, + 198, + 307 + ], + "spans": [ + { + "bbox": [ + 70, + 269, + 198, + 307 + ], + "type": "interline_equation", + "content": "\\begin{array}{c} r = \\frac {\\alpha}{\\beta} | \\sin \\beta (\\sigma_ {2} + \\sigma_ {1}) | \\\\ \\text {e x t r a c t e a c h} \\\\ \\text {c o l o r a n d l o c a t e} \\\\ \\text {t h e b o u n d i n g b o x} \\end{array} r = \\frac {\\alpha}{\\beta} | \\sin \\beta (\\sigma_ {2} + \\sigma_ {1}) |", + "image_path": "af2a657360d16dc129df8ff58cad81f08eca75e2e909c60a5877bf0cd9233aec.jpg" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 74, + 318, + 190, + 373 + ], + "blocks": [ + { + "bbox": [ + 74, + 318, + 190, + 373 + ], + "lines": [ + { + "bbox": [ + 74, + 318, + 190, + 373 + ], + "spans": [ + { + "bbox": [ + 74, + 318, + 190, + 373 + ], + "type": "image", + "image_path": "476845664a3358db013068b11bce6a22519ceb0eaea91ffb09aed9e18eb92e3e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 218, + 118, + 284, + 147 + ], + "blocks": [ + { + "bbox": [ + 241, + 99, + 260, + 108 + ], + "lines": [ + { + "bbox": [ + 241, + 99, + 260, + 108 + ], + "spans": [ + { + "bbox": [ + 241, + 99, + 260, + 108 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 218, + 118, + 284, + 147 + ], + "lines": [ + { + "bbox": [ + 218, + 118, + 284, + 147 + ], + "spans": [ + { + "bbox": [ + 218, + 118, + 284, + 147 + ], + "type": "image", + "image_path": "fef9e67c4918b4b39f8cf2ed6809bb9b033dad3e1bb7223a4568b9655df9b2db.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 218, + 159, + 285, + 188 + ], + "blocks": [ + { + "bbox": [ + 218, + 159, + 285, + 188 + ], + "lines": [ + { + "bbox": [ + 218, + 159, + 285, + 188 + ], + "spans": [ + { + "bbox": [ + 218, + 159, + 285, + 188 + ], + "type": "image", + "image_path": "309337165261ce94e6e8adc3a54c0274b598009d569e43604e9e0ce91d3480b6.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 238, + 205, + 346, + 214 + ], + "lines": [ + { + "bbox": [ + 238, + 205, + 346, + 214 + ], + "spans": [ + { + "bbox": [ + 238, + 205, + 346, + 214 + ], + "type": "text", + "content": "Step 3: Invalid Match Elimination" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 286, + 120, + 422, + 190 + ], + "blocks": [ + { + "bbox": [ + 318, + 79, + 437, + 90 + ], + "lines": [ + { + "bbox": [ + 318, + 79, + 437, + 90 + ], + "spans": [ + { + "bbox": [ + 318, + 79, + 437, + 90 + ], + "type": "text", + "content": "Step 2: Element Region Matching" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 98, + 405, + 114 + ], + "lines": [ + { + "bbox": [ + 304, + 98, + 405, + 114 + ], + "spans": [ + { + "bbox": [ + 304, + 98, + 405, + 114 + ], + "type": "text", + "content": "1. Calculate Match Cost Between Every Element Pairs" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 286, + 120, + 422, + 190 + ], + "lines": [ + { + "bbox": [ + 286, + 120, + 422, + 190 + ], + "spans": [ + { + "bbox": [ + 286, + 120, + 422, + 190 + ], + "type": "image", + "image_path": "a6bed2a0a6ca774c3e62edf7709e80e9b09f3360c8ef2bd0145a54fd5192eaff.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 416, + 205, + 511, + 214 + ], + "lines": [ + { + "bbox": [ + 416, + 205, + 511, + 214 + ], + "spans": [ + { + "bbox": [ + 416, + 205, + 511, + 214 + ], + "type": "text", + "content": "Step 4: Metric Calculation" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 405, + 120, + 526, + 187 + ], + "blocks": [ + { + "bbox": [ + 424, + 98, + 504, + 114 + ], + "lines": [ + { + "bbox": [ + 424, + 98, + 504, + 114 + ], + "spans": [ + { + "bbox": [ + 424, + 98, + 504, + 114 + ], + "type": "text", + "content": "2. Find the Best Match by Hungarian Algorithm" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 405, + 120, + 526, + 187 + ], + "lines": [ + { + "bbox": [ + 405, + 120, + 526, + 187 + ], + "spans": [ + { + "bbox": [ + 405, + 120, + 526, + 187 + ], + "type": "image", + "image_path": "44520171490c921152c241eb9a56460ee74d0c32aa01941c7d3bb3a0e0ec2288.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 438, + 229, + 488, + 237 + ], + "lines": [ + { + "bbox": [ + 438, + 229, + 488, + 237 + ], + "spans": [ + { + "bbox": [ + 438, + 229, + 488, + 237 + ], + "type": "text", + "content": "Evaluation Metric" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 226, + 232, + 349, + 298 + ], + "blocks": [ + { + "bbox": [ + 211, + 224, + 296, + 233 + ], + "lines": [ + { + "bbox": [ + 211, + 224, + 296, + 233 + ], + "spans": [ + { + "bbox": [ + 211, + 224, + 296, + 233 + ], + "type": "text", + "content": "1. Token Consistency Check" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 226, + 232, + 349, + 298 + ], + "lines": [ + { + "bbox": [ + 226, + 232, + 349, + 298 + ], + "spans": [ + { + "bbox": [ + 226, + 232, + 349, + 298 + ], + "type": "image", + "image_path": "2e6767b0c3ac6e1015c28c40261f78ef5437b67a707c7dff6097d8061027362d.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 227, + 316, + 349, + 380 + ], + "blocks": [ + { + "bbox": [ + 212, + 304, + 341, + 314 + ], + "lines": [ + { + "bbox": [ + 212, + 304, + 341, + 314 + ], + "spans": [ + { + "bbox": [ + 212, + 304, + 341, + 314 + ], + "type": "text", + "content": "2. Position Relationship Consistency Check" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 227, + 316, + 349, + 380 + ], + "lines": [ + { + "bbox": [ + 227, + 316, + 349, + 380 + ], + "spans": [ + { + "bbox": [ + 227, + 316, + 349, + 380 + ], + "type": "image", + "image_path": "9b656f304f8b314129cfa7edfe5f8d6c233268eba74e037704e963b8acfa559e.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 394, + 555, + 439 + ], + "lines": [ + { + "bbox": [ + 55, + 394, + 555, + 439 + ], + "spans": [ + { + "bbox": [ + 55, + 394, + 555, + 439 + ], + "type": "text", + "content": "Figure 2. Overview of the Character Detection Matching (CDM), consisting of four main stages. (1) Element Localization, where bounding boxes of individual elements are extracted. (2) Element Region Matching, which employs a bipartite graph matching method to pair prediction with ground truth elements. (3) Invalid Match Elimination, where inconsistent matches are discarded through token and positional relationship checks. (4) Metric Calculation, where matching accuracy is evaluated using the F1-Score and ExpRate@CDM." + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 431, + 243, + 500, + 258 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 431, + 243, + 500, + 258 + ], + "spans": [ + { + "bbox": [ + 431, + 243, + 500, + 258 + ], + "type": "interline_equation", + "content": "F 1 \\text {s c o r e} = \\frac {2 T P}{2 T P + F P + F N}", + "image_path": "33e259a818f6ca1bbbfdffe4515a90744ec2d9ebb910cf358eeb3cacb22e94fa.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 417, + 266, + 516, + 288 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 417, + 266, + 494, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 417, + 266, + 494, + 272 + ], + "spans": [ + { + "bbox": [ + 417, + 266, + 494, + 272 + ], + "type": "text", + "content": "- TP: number of matched bbox pairs" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 417, + 274, + 516, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 417, + 274, + 516, + 280 + ], + "spans": [ + { + "bbox": [ + 417, + 274, + 516, + 280 + ], + "type": "text", + "content": "- FP: number of unmatched bboxes in prediction" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 417, + 281, + 503, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 417, + 281, + 503, + 288 + ], + "spans": [ + { + "bbox": [ + 417, + 281, + 503, + 288 + ], + "type": "text", + "content": "FN: number of unmatched bboxes in GT" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 399, + 304, + 454, + 337 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 304, + 454, + 337 + ], + "spans": [ + { + "bbox": [ + 399, + 304, + 454, + 337 + ], + "type": "interline_equation", + "content": "T P = \\left(\\bigtriangledown\\right) _ {n u m} = 1 3", + "image_path": "8419c0b4f82a9f5713fc80e517c51aba383a9c48324512c735931fcaf6cd5d52.jpg" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 474, + 304, + 526, + 319 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 304, + 526, + 319 + ], + "spans": [ + { + "bbox": [ + 474, + 304, + 526, + 319 + ], + "type": "interline_equation", + "content": "F P = \\left(\\square\\right) _ {\\text {n u m}} ^ {\\text {P r e d}} = 3", + "image_path": "dc61bd043523b6784f3d85d520c347b8abcdf1e47f268ad258fb30d4dc5cbfc3.jpg" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 474, + 329, + 526, + 343 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 329, + 526, + 343 + ], + "spans": [ + { + "bbox": [ + 474, + 329, + 526, + 343 + ], + "type": "interline_equation", + "content": "F N = \\left( \\begin{array}{l l} \\square & T \\\\ 0 & n u m \\end{array} \\right) ^ {G T} = 3", + "image_path": "4f8f071a59da18705a766ebb10bff0ebac2a2ead0e85e28ec6a160bbe4d1efac.jpg" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 414, + 351, + 509, + 370 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 414, + 351, + 509, + 370 + ], + "spans": [ + { + "bbox": [ + 414, + 351, + 509, + 370 + ], + "type": "interline_equation", + "content": "F 1 s c o r e = \\frac {2 \\times 1 3}{2 \\times 1 3 + 3 + 3} = 0. 8 1 2 5", + "image_path": "9abb4124fc3fd1cebaf4631fc9231bcf9c717d6f64c16682c6ca46d44014e10f.jpg" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 55, + 449, + 272, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 272, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 272, + 462 + ], + "type": "text", + "content": "of formula recognition across diverse representations." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 55, + 475, + 227, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 227, + 490 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 227, + 490 + ], + "type": "text", + "content": "4. Character Detection Matching" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 55, + 497, + 295, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 497, + 295, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 497, + 295, + 687 + ], + "type": "text", + "content": "Due to the diversity of LaTeX expressions, text-based character-matching methods are unreliable for formula recognition evaluation. The basic idea of CDM is to compare the rendered images from LaTex text. If the image rendered from the predicted LaTeX source code matches the image rendered from the ground truth LaTeX source code, the formula is considered entirely correct. However, directly comparing the pixel values of the original and predicted formulas is not ideal. Any error or extra/missing character in the prediction can cause subsequent characters to be mismatched. Additionally, two similar formulas might have different layouts, with one being a single-line formula and the other a multi-line formula due to line breaks. Therefore, a more robust algorithm is needed to calculate the match between the predicted result and the ground truth image." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "To this end, we propose a metric that incorporates a bipartite matching step for element-level matching in images," + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 313, + 449, + 553, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 553, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 553, + 474 + ], + "type": "text", + "content": "providing a more intuitive assessment. As shown in Figure 2, the algorithm consists of four stages as follows." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 313, + 480, + 436, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 436, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 436, + 491 + ], + "type": "text", + "content": "4.1. Element Localization" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 313, + 498, + 553, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 553, + 533 + ], + "type": "text", + "content": "First, the bounding boxes (bboxes) of each individual element in the rendered image are extracted, followed by the subsequent steps:" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 313, + 534, + 554, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 554, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 554, + 605 + ], + "type": "text", + "content": "LaTeX Source Normalization. LaTeX source codes of both the ground truth and predicted formulas are normalized, breaking them down into individual tokens such as \"2\", \"a\", \"A\", \"\\alpha\", \"\\sin\". Composite elements are decomposed into individual characters, e.g., \"\\frac{ab}{\\alpha} is decomposed into \\frac{a}{b}\\)\"." + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": "Element Region Localization. To accurately detect character positions, we render each token in a unique color. We construct an RGB color list with a fixed interval of 15, creating a sequence from " + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "(0,0,15)" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "(255,255,255)" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": ". This generates " + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "(255 / 15 + 1)^{3} = 5832" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": " distinct colors, sufficient for complex formulas. Each token is assigned a unique color using the LaTeX command \"\\mathsf{\\\\mathsf{m}}\\mathsf{a}\\mathsf{h}\\mathsf{o}\\mathsf{c}\\mathsf{l}\\mathsf{r}\\mathsf{R}\\mathsf{B}\\mathsf{]} \\{\\mathsf{r},\\mathsf{g},\\mathsf{b}\\} \"$. After rendering, we apply image processing techniques to identify pixels match-" + } + ] + } + ], + "index": 41 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19684" + } + ] + } + ], + "index": 42 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "content": "ing each color, thereby determining the exact coordinates of each token." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 102, + 202, + 116 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 102, + 202, + 116 + ], + "spans": [ + { + "bbox": [ + 55, + 102, + 202, + 116 + ], + "type": "text", + "content": "4.2. Element Region Matching" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 119, + 296, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 119, + 296, + 215 + ], + "spans": [ + { + "bbox": [ + 55, + 119, + 296, + 215 + ], + "type": "text", + "content": "In this stage, a bipartite matching method pairs the predicted elements with the corresponding ground truth elements. Based on the element localization, two sets are obtained for each formula: one for the ground truth independent elements " + }, + { + "bbox": [ + 55, + 119, + 296, + 215 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 119, + 296, + 215 + ], + "type": "text", + "content": " and one for the predicted independent elements " + }, + { + "bbox": [ + 55, + 119, + 296, + 215 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 55, + 119, + 296, + 215 + ], + "type": "text", + "content": ". The number of independent elements in each set is " + }, + { + "bbox": [ + 55, + 119, + 296, + 215 + ], + "type": "inline_equation", + "content": "N_y" + }, + { + "bbox": [ + 55, + 119, + 296, + 215 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 119, + 296, + 215 + ], + "type": "inline_equation", + "content": "N_{\\hat{y}}" + }, + { + "bbox": [ + 55, + 119, + 296, + 215 + ], + "type": "text", + "content": ", respectively, with " + }, + { + "bbox": [ + 55, + 119, + 296, + 215 + ], + "type": "inline_equation", + "content": "N = \\min(N_y, N_{\\hat{y}})" + }, + { + "bbox": [ + 55, + 119, + 296, + 215 + ], + "type": "text", + "content": " being the number of elements in the smaller set." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 216, + 296, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 216, + 296, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 216, + 296, + 288 + ], + "type": "text", + "content": "To measure the similarity between " + }, + { + "bbox": [ + 55, + 216, + 296, + 288 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 216, + 296, + 288 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 216, + 296, + 288 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 55, + 216, + 296, + 288 + ], + "type": "text", + "content": ", we match elements in the two sets by identifying the corresponding ground truth element for each predicted element. We use the bipartite matching Hungarian algorithm [13], as described in DETR [6], to find a permutation " + }, + { + "bbox": [ + 55, + 216, + 296, + 288 + ], + "type": "inline_equation", + "content": "\\hat{\\sigma}" + }, + { + "bbox": [ + 55, + 216, + 296, + 288 + ], + "type": "text", + "content": " that minimizes the total matching cost:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 102, + 293, + 296, + 326 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 293, + 296, + 326 + ], + "spans": [ + { + "bbox": [ + 102, + 293, + 296, + 326 + ], + "type": "interline_equation", + "content": "\\hat {\\sigma} = \\arg \\min _ {\\sigma \\in S _ {N}} \\sum_ {i = 1} ^ {N} L _ {\\text {m a t c h}} \\left(y _ {i}, \\hat {y} _ {\\sigma (i)}\\right), \\tag {2}", + "image_path": "49622efa476cea47f7352fac50c2761aac2283eeeca049496ae92d7da2f1acd5.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 86, + 335, + 295, + 348 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 335, + 295, + 348 + ], + "spans": [ + { + "bbox": [ + 86, + 335, + 295, + 348 + ], + "type": "interline_equation", + "content": "L _ {\\text {m a t c h}} = W _ {t} \\times L _ {t} + W _ {p} \\times L _ {p} + W _ {o} \\times L _ {o}, \\tag {3}", + "image_path": "33c81faa962d96deb076de6eb2e99c7e9b214d9d65618c82547836eb3bb551f9.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 351, + 296, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 351, + 296, + 374 + ], + "spans": [ + { + "bbox": [ + 55, + 351, + 296, + 374 + ], + "type": "text", + "content": "where the matching cost " + }, + { + "bbox": [ + 55, + 351, + 296, + 374 + ], + "type": "inline_equation", + "content": "L_{match}" + }, + { + "bbox": [ + 55, + 351, + 296, + 374 + ], + "type": "text", + "content": " is defined as a weighted sum of three components as introduced as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 376, + 296, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 296, + 457 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 296, + 457 + ], + "type": "text", + "content": "- Token Matching Cost " + }, + { + "bbox": [ + 55, + 376, + 296, + 457 + ], + "type": "inline_equation", + "content": "L_{t}" + }, + { + "bbox": [ + 55, + 376, + 296, + 457 + ], + "type": "text", + "content": ": This component measures whether the tokens corresponding to two bounding boxes are the same. If they are identical, the cost is 0; if they are different, the cost is 1. For tokens that render identically but are different, such as \" (\" , \"\\leftarrow\" , and \"\\bigig( \", the cost is 0.05 , which can be formulated as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 122, + 457, + 295, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 457, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 122, + 457, + 295, + 502 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {t}} = \\left\\{ \\begin{array}{l l} 0, & \\text {i f} t _ {i} = \\hat {t} _ {\\sigma (i)}; \\\\ 0. 0 5, & \\text {i f} t _ {i} \\approx \\hat {t} _ {\\sigma (i)}; \\\\ 1, & \\text {o t h e r w i s e}; \\end{array} \\right. \\tag {4}", + "image_path": "6ff4baf12ce566105f3f37b1c6092fec48a3a9faf80dc0ac4d09cf93372eddeb.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 64, + 504, + 294, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 504, + 294, + 515 + ], + "spans": [ + { + "bbox": [ + 64, + 504, + 294, + 515 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 64, + 504, + 294, + 515 + ], + "type": "inline_equation", + "content": "\\approx" + }, + { + "bbox": [ + 64, + 504, + 294, + 515 + ], + "type": "text", + "content": " denotes tokens that differ but render identically." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 516, + 295, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 516, + 295, + 563 + ], + "spans": [ + { + "bbox": [ + 56, + 516, + 295, + 563 + ], + "type": "text", + "content": "- Positional Proximity Cost " + }, + { + "bbox": [ + 56, + 516, + 295, + 563 + ], + "type": "inline_equation", + "content": "L_{p}" + }, + { + "bbox": [ + 56, + 516, + 295, + 563 + ], + "type": "text", + "content": ": This component measures the proximity of the two bounding boxes' positions using the L1 norm of their coordinates, which can be formulated as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 124, + 566, + 295, + 591 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 566, + 295, + 591 + ], + "spans": [ + { + "bbox": [ + 124, + 566, + 295, + 591 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {p}} = \\frac {1}{D _ {b}} \\times \\| b _ {i} - \\hat {b} _ {\\sigma (i)} \\| _ {1}, \\tag {5}", + "image_path": "1fbf6938183d24c2069f5fcb72a26a81be2c7758fd57c6ff528111fc8c69d0c3.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 63, + 595, + 295, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 595, + 295, + 619 + ], + "spans": [ + { + "bbox": [ + 63, + 595, + 295, + 619 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 63, + 595, + 295, + 619 + ], + "type": "inline_equation", + "content": "b = [x_1, y_1, x_2, y_2]" + }, + { + "bbox": [ + 63, + 595, + 295, + 619 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 63, + 595, + 295, + 619 + ], + "type": "inline_equation", + "content": "D_b" + }, + { + "bbox": [ + 63, + 595, + 295, + 619 + ], + "type": "text", + "content": " is the dimension of the bounding box coordinates." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 619, + 296, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 619, + 296, + 677 + ], + "spans": [ + { + "bbox": [ + 56, + 619, + 296, + 677 + ], + "type": "text", + "content": "- Order Similarity Cost " + }, + { + "bbox": [ + 56, + 619, + 296, + 677 + ], + "type": "inline_equation", + "content": "L_{o}" + }, + { + "bbox": [ + 56, + 619, + 296, + 677 + ], + "type": "text", + "content": ": This measures the similarity of the token order in the original LaTeX source (an approximation of reading order). The order is normalized to the range [0, 1], and the L1 norm can be calculated as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 124, + 676, + 295, + 700 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 676, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 124, + 676, + 295, + 700 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {o}} = \\frac {1}{D _ {o}} \\times \\left\\| o _ {i} - \\hat {o} _ {\\sigma (i)} \\right\\| _ {1}, \\tag {6}", + "image_path": "be7f57e8a6d59a409477f22fc945d74497b5fc1c6c95f6d7ef6817259381f3dc.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 64, + 701, + 291, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 701, + 291, + 714 + ], + "spans": [ + { + "bbox": [ + 64, + 701, + 291, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 64, + 701, + 291, + 714 + ], + "type": "inline_equation", + "content": "o_i" + }, + { + "bbox": [ + 64, + 701, + 291, + 714 + ], + "type": "text", + "content": " means order of " + }, + { + "bbox": [ + 64, + 701, + 291, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 64, + 701, + 291, + 714 + ], + "type": "text", + "content": "-th token, with dimension " + }, + { + "bbox": [ + 64, + 701, + 291, + 714 + ], + "type": "inline_equation", + "content": "D_o" + }, + { + "bbox": [ + 64, + 701, + 291, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 72, + 554, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 145 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 145 + ], + "type": "text", + "content": "Overall, the weights " + }, + { + "bbox": [ + 313, + 72, + 554, + 145 + ], + "type": "inline_equation", + "content": "W_{t}, W_{p}, W_{o}" + }, + { + "bbox": [ + 313, + 72, + 554, + 145 + ], + "type": "text", + "content": " are used to balance the contributions of the three components. By employing this comprehensive matching strategy, we ensure a more accurate and robust evaluation of the correspondence between the predicted and ground truth elements, thereby improving the overall assessment of formula recognition quality." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 152, + 463, + 164 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 152, + 463, + 164 + ], + "spans": [ + { + "bbox": [ + 313, + 152, + 463, + 164 + ], + "type": "text", + "content": "4.3. Invalid Match Elimination" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 170, + 554, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 170, + 554, + 217 + ], + "spans": [ + { + "bbox": [ + 313, + 170, + 554, + 217 + ], + "type": "text", + "content": "After pairing the individual elements of the predicted result with the ground truth using the Hungarian matching algorithm, we need to verify these pairs and eliminate invalid matches. This process involves two steps:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 218, + 554, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 218, + 554, + 254 + ], + "spans": [ + { + "bbox": [ + 313, + 218, + 554, + 254 + ], + "type": "text", + "content": "Token Consistency Check. Check whether the elements in each matched pair are consistent in terms of characters. If they are inconsistent, discard the match." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 255, + 554, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 255, + 554, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 554, + 373 + ], + "type": "text", + "content": "Position Relationship Consistency Check. The relative positions of elements in mathematical formulas are crucial. For instance, in the expressions " + }, + { + "bbox": [ + 313, + 255, + 554, + 373 + ], + "type": "inline_equation", + "content": "2^3" + }, + { + "bbox": [ + 313, + 255, + 554, + 373 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 255, + 554, + 373 + ], + "type": "inline_equation", + "content": "3^2" + }, + { + "bbox": [ + 313, + 255, + 554, + 373 + ], + "type": "text", + "content": ", bipartite matching might pair 2 with 2 and 3 with 3, but their meanings and visual representations are entirely different. Thus, we need to check the consistency of the positional relationships within the matched pairs. We treat each element in the matched pair as a bounding box and analyze their relative positions. Specifically, we assume an affine transformation between the ground truth and predicted elements:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 402, + 382, + 554, + 398 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 382, + 554, + 398 + ], + "spans": [ + { + "bbox": [ + 402, + 382, + 554, + 398 + ], + "type": "interline_equation", + "content": "\\hat {b} _ {\\sigma (i)} = \\mathbf {A} \\left(b _ {i}\\right), \\tag {7}", + "image_path": "d89c2e823244512f03953143db2083438033b6c3e26b7220e69427f56bbbbcb9.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 406, + 554, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 406, + 554, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 406, + 554, + 538 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 406, + 554, + 538 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 313, + 406, + 554, + 538 + ], + "type": "text", + "content": " is the affine transformation matrix. To identify inconsistent match pairs, we detect pairs that do not conform to this transformation relationship. We employ the RANSAC algorithm [9] for this purpose. RANSAC can determine the optimal transformation matrix " + }, + { + "bbox": [ + 313, + 406, + 554, + 538 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 313, + 406, + 554, + 538 + ], + "type": "text", + "content": " in the presence of noise. Given that formulas are usually horizontally arranged during rendering, we fix the rotation angle in the transformation matrix to 0, considering only translation and scaling. This approach not only improves the convergence speed of the RANSAC algorithm but also enhances the final matching accuracy." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 538, + 554, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 538, + 554, + 609 + ], + "spans": [ + { + "bbox": [ + 313, + 538, + 554, + 609 + ], + "type": "text", + "content": "To account for line-breaking effects in formulas, we perform multiple rounds of RANSAC iterations to ensure that as many matched pairs as possible conform to the transformation relationship. After several iterations, matched pairs that still do not conform to the transformation relationship are considered incorrect and are eliminated." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 610, + 554, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 610, + 554, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 610, + 554, + 635 + ], + "type": "text", + "content": "The above two steps effectively eliminate invalid match pairs, ensuring more accurate final matching results." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 642, + 425, + 654 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 425, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 425, + 654 + ], + "type": "text", + "content": "4.4. Metric Calculation" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 660, + 554, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 660, + 554, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 660, + 554, + 684 + ], + "type": "text", + "content": "We use the F1-Score as the default metric for evaluating CDM (Character Detection Metric), defined as:" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 365, + 692, + 553, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 692, + 553, + 717 + ], + "spans": [ + { + "bbox": [ + 365, + 692, + 553, + 717 + ], + "type": "interline_equation", + "content": "C D M = \\frac {2 \\times T P}{2 \\times T P + F P + F N}, \\tag {8}", + "image_path": "b3ec1360b9e46d2ec91b9c192ef8ab446d2d89ec047143c1ad2b2ebbec02f5d6.jpg" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19685" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 104, + 62, + 329, + 220 + ], + "blocks": [ + { + "bbox": [ + 104, + 62, + 329, + 220 + ], + "lines": [ + { + "bbox": [ + 104, + 62, + 329, + 220 + ], + "spans": [ + { + "bbox": [ + 104, + 62, + 329, + 220 + ], + "type": "image", + "image_path": "d8ff569f6b6f08e795bdf71a951aa6d5af2d34f01a954ab7c802c34bf22c3cb8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 139, + 233, + 291, + 243 + ], + "lines": [ + { + "bbox": [ + 139, + 233, + 291, + 243 + ], + "spans": [ + { + "bbox": [ + 139, + 233, + 291, + 243 + ], + "type": "text", + "content": "(a) CDM distribution on UniMER-Test dataset" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 331, + 61, + 506, + 220 + ], + "blocks": [ + { + "bbox": [ + 331, + 61, + 506, + 220 + ], + "lines": [ + { + "bbox": [ + 331, + 61, + 506, + 220 + ], + "spans": [ + { + "bbox": [ + 331, + 61, + 506, + 220 + ], + "type": "image", + "image_path": "acfb733409878c8353891713f6b8d6cbda9bd3dfabcd24eddb073964b56ddf48.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 361, + 233, + 474, + 243 + ], + "lines": [ + { + "bbox": [ + 361, + 233, + 474, + 243 + ], + "spans": [ + { + "bbox": [ + 361, + 233, + 474, + 243 + ], + "type": "text", + "content": "(b) Examples by CDM score range" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 249, + 555, + 273 + ], + "lines": [ + { + "bbox": [ + 54, + 249, + 555, + 273 + ], + "spans": [ + { + "bbox": [ + 54, + 249, + 555, + 273 + ], + "type": "text", + "content": "Figure 3. CDM Score Distribution and Example Analysis. (a) CDM score distribution for Mathpix, UniMERNet, Texify, and Pix2Tex on the UniMER-Test dataset. (b) Example analysis of Pix2Tex predictions across different CDM score ranges." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 277, + 295, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 295, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 295, + 301 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 277, + 295, + 301 + ], + "type": "inline_equation", + "content": "TP" + }, + { + "bbox": [ + 55, + 277, + 295, + 301 + ], + "type": "text", + "content": " denotes true positives, " + }, + { + "bbox": [ + 55, + 277, + 295, + 301 + ], + "type": "inline_equation", + "content": "FP" + }, + { + "bbox": [ + 55, + 277, + 295, + 301 + ], + "type": "text", + "content": " denotes false positives, and " + }, + { + "bbox": [ + 55, + 277, + 295, + 301 + ], + "type": "inline_equation", + "content": "FN" + }, + { + "bbox": [ + 55, + 277, + 295, + 301 + ], + "type": "text", + "content": " denotes false negatives." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 301, + 295, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 301, + 295, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 301, + 295, + 326 + ], + "type": "text", + "content": "To further evaluate the accuracy of formula recognition, we introduce the ExpRate@CDM metric, defined as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 85, + 335, + 295, + 361 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 335, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 85, + 335, + 295, + 361 + ], + "type": "interline_equation", + "content": "\\operatorname {E x p R a t e} @ C D M = \\frac {\\sum_ {i = 1} ^ {N} \\mathbb {I} \\left(C D M _ {i} = 1\\right)}{N}, \\tag {9}", + "image_path": "179c84c1135dbfb48bdcba53e06e16fb56c9a423da5a8c001c95d86f8fffa5b1.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 369, + 296, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 369, + 296, + 441 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 296, + 441 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 369, + 296, + 441 + ], + "type": "inline_equation", + "content": "\\mathbb{I}" + }, + { + "bbox": [ + 55, + 369, + 296, + 441 + ], + "type": "text", + "content": " is the indicator function that equals 1 if " + }, + { + "bbox": [ + 55, + 369, + 296, + 441 + ], + "type": "inline_equation", + "content": "CDM_{i} = 1" + }, + { + "bbox": [ + 55, + 369, + 296, + 441 + ], + "type": "text", + "content": " and 0 otherwise, and " + }, + { + "bbox": [ + 55, + 369, + 296, + 441 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 369, + 296, + 441 + ], + "type": "text", + "content": " is the total number of formulas. This metric represents the proportion of formulas for which the model's prediction results are perfectly matched. Essentially, ExpRate@CDM serves as a precise version of the ExpRate metric specifically for formula recognition." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 460, + 137, + 473 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 137, + 473 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 137, + 473 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 479, + 158, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 158, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 158, + 491 + ], + "type": "text", + "content": "5.1. Models and Data" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 498, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 605 + ], + "type": "text", + "content": "We validate the CDM metric by evaluating several mainstream formula recognition models using both subjective impressions and objective metrics. The models include open-source UniMERNet [30], Texify [28], Pix2tex [4], and the commercial Mathpix API, all tested on the UniMERTest dataset. Besides, we evaluate document-level models, such as the open-source Nougat [5] and the commercial GPT-4o [10]. Vary [34] and StrucTexTv3 [21] are excluded as they are currently unavailable." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 605, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 665 + ], + "type": "text", + "content": "UniMER-Test Dataset. The dataset includes 23,757 formula samples, categorized into Simple Printed Expressions (SPE), Complex Printed Expressions (CPE), Screenshot Expressions (SCE), and Handwritten Expressions (HWE). We use these categories to conduct the model evaluation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Tiny-Doc-Math Dataset. To evaluate document-level recognition, we construct the Tiny-Doc-Math dataset, consisting of arXiv papers in mathematics and computer science, published after June 2024, to ensure that they are not" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 277, + 555, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 555, + 336 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 555, + 336 + ], + "type": "text", + "content": "in the training data of the compared models. We obtain Latex code and corresponding PDFs, match displayed equations using regular expressions, and manually verify them. Overall, the dataset includes 12 PDFs, totaling 196 pages and 437 formulas." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 338, + 555, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 555, + 361 + ], + "type": "text", + "content": "This validation set includes both formula-level and document-level evaluations:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 361, + 554, + 494 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 314, + 361, + 554, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 361, + 554, + 421 + ], + "spans": [ + { + "bbox": [ + 314, + 361, + 554, + 421 + ], + "type": "text", + "content": "- Formula-level: Using single rendered formula images as input, we evaluate Mathpix, Pix2Tex, and UniMERNet. These models accept cropped formula images as input, and we compare the model outputs with the ground truth to compute relevant metrics." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 422, + 554, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 422, + 554, + 494 + ], + "spans": [ + { + "bbox": [ + 314, + 422, + 554, + 494 + ], + "type": "text", + "content": "- Document-level: Using PDFs or images as input, we evaluate Nougat, GPT-4o, and Mathpix, which can convert entire PDF pages into Markdown format. We match the displayed equations in the model outputs using regular expressions and compare them with the ground truth LaTeX formulas to compute relevant metrics." + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 502, + 485, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 485, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 485, + 514 + ], + "type": "text", + "content": "5.2. Credibility Assessment of CDM" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 520, + 444, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 520, + 444, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 520, + 444, + 532 + ], + "type": "text", + "content": "5.2.1. Rendering Success Rate" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 535, + 555, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 535, + 555, + 631 + ], + "spans": [ + { + "bbox": [ + 313, + 535, + 555, + 631 + ], + "type": "text", + "content": "The CDM metric relies on the successful rendering of formula images. For models that fail to render images, we assign a CDM score of 0, as rendering failures indicate that the predicted LaTeX code lacks critical elements. The rendering success rates on the UniMER-Test dataset for Pix2tex, Texify, UniMERNet, and Mathpix are " + }, + { + "bbox": [ + 313, + 535, + 555, + 631 + ], + "type": "inline_equation", + "content": "96.63\\%" + }, + { + "bbox": [ + 313, + 535, + 555, + 631 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 535, + 555, + 631 + ], + "type": "inline_equation", + "content": "94.77\\%" + }, + { + "bbox": [ + 313, + 535, + 555, + 631 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 535, + 555, + 631 + ], + "type": "inline_equation", + "content": "99.71\\%" + }, + { + "bbox": [ + 313, + 535, + 555, + 631 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 535, + 555, + 631 + ], + "type": "inline_equation", + "content": "97.82\\%" + }, + { + "bbox": [ + 313, + 535, + 555, + 631 + ], + "type": "text", + "content": ", respectively, ensuring the applicability and reliability of the CDM metric." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 639, + 460, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 639, + 460, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 639, + 460, + 649 + ], + "type": "text", + "content": "5.2.2. User Preference Evaluation" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 653, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 554, + 714 + ], + "type": "text", + "content": "We analyze the distribution of CDM scores for four models on the UniMER-Test dataset. As shown in Figure 3a, Math-pix and UniMERNet perform well in terms of CDM scores. We conduct a detailed analysis of the Pix2Tex model by randomly selecting samples from different score ranges to" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19686" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 69, + 197, + 208 + ], + "blocks": [ + { + "bbox": [ + 66, + 69, + 197, + 208 + ], + "lines": [ + { + "bbox": [ + 66, + 69, + 197, + 208 + ], + "spans": [ + { + "bbox": [ + 66, + 69, + 197, + 208 + ], + "type": "image", + "image_path": "bd60bd570cb6355ae0a2f0e3e187fd7db70f4f52099ef7aab3291bafd9d71a37.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 74, + 220, + 187, + 231 + ], + "lines": [ + { + "bbox": [ + 74, + 220, + 187, + 231 + ], + "spans": [ + { + "bbox": [ + 74, + 220, + 187, + 231 + ], + "type": "text", + "content": "(a) Human Evaluation Preferences" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 201, + 70, + 385, + 208 + ], + "blocks": [ + { + "bbox": [ + 201, + 70, + 385, + 208 + ], + "lines": [ + { + "bbox": [ + 201, + 70, + 385, + 208 + ], + "spans": [ + { + "bbox": [ + 201, + 70, + 385, + 208 + ], + "type": "image", + "image_path": "2e973c968c75f8b1b534c2d12b568ff1507ad9f36f14a1a99310c9ce6eba7abb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 228, + 220, + 355, + 231 + ], + "lines": [ + { + "bbox": [ + 228, + 220, + 355, + 231 + ], + "spans": [ + { + "bbox": [ + 228, + 220, + 355, + 231 + ], + "type": "text", + "content": "(b) BLEU vs.CDM Score Distribution" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 389, + 71, + 543, + 207 + ], + "blocks": [ + { + "bbox": [ + 389, + 71, + 543, + 207 + ], + "lines": [ + { + "bbox": [ + 389, + 71, + 543, + 207 + ], + "spans": [ + { + "bbox": [ + 389, + 71, + 543, + 207 + ], + "type": "image", + "image_path": "e69b68ade6a5bf1d4311efebde234b734e4b690aca3b0a6b464e40808d14460d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 404, + 220, + 526, + 232 + ], + "lines": [ + { + "bbox": [ + 404, + 220, + 526, + 232 + ], + "spans": [ + { + "bbox": [ + 404, + 220, + 526, + 232 + ], + "type": "text", + "content": "(c) Impact of Formula Writing Styles" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 63, + 265, + 289, + 423 + ], + "blocks": [ + { + "bbox": [ + 55, + 234, + 555, + 257 + ], + "lines": [ + { + "bbox": [ + 55, + 234, + 555, + 257 + ], + "spans": [ + { + "bbox": [ + 55, + 234, + 555, + 257 + ], + "type": "text", + "content": "Figure 4. Human evaluation and metric comparison. (a) Distribution of human preferences. (b) Distribution of cases preferring BLEU or CDM across different score ranges. (c) Impact of formula writing styles on BLEU, ExpRate and CDM scores." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 63, + 265, + 289, + 423 + ], + "lines": [ + { + "bbox": [ + 63, + 265, + 289, + 423 + ], + "spans": [ + { + "bbox": [ + 63, + 265, + 289, + 423 + ], + "type": "image", + "image_path": "7314099a976ec80cb10c028ef7556bb1382df3225a593df61b5979804041de0f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 435, + 295, + 480 + ], + "lines": [ + { + "bbox": [ + 55, + 435, + 295, + 480 + ], + "spans": [ + { + "bbox": [ + 55, + 435, + 295, + 480 + ], + "type": "text", + "content": "Figure 5. Annotation interface for evaluating user preferences between CDM and BLEU scores. Annotators chose from four options: \"Score A is better\", \"Score B is better\", \"Both scores are equally good\", or \"Neither score is good\"." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 495, + 295, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 295, + 542 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 295, + 542 + ], + "type": "text", + "content": "evaluate if the prediction quality corresponds to the CDM scores. The analysis in Figure 3b shows that the CDM scores effectively reflect formula quality, with higher scores indicating fewer errors." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 544, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 664 + ], + "type": "text", + "content": "To verify the consistency between the CDM metric and human evaluation, we select 1,008 CDM scores from Pix2Tex predictions, ensuring a balanced score distribution. We design an annotation interface (shown in Figure 5), displaying a ground truth label and the corresponding predicted LaTeX rendered image. Annotators choose between ScoreA, ScoreB, Both (credible), and Neither (credible). ScoreA and ScoreB correspond to the BLEU and CDM scores, respectively, but their order is randomized. For more details, please refer to the supplementary materials." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "The results in Figure 4a show that " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "64\\%" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " of participants prefer the CDM metric, and " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "32\\%" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " consider both metrics good. This indicates a " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "96\\%" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " consistency between the CDM metric and human evaluation, demonstrating its reliability." + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 317, + 264, + 558, + 328 + ], + "blocks": [ + { + "bbox": [ + 317, + 264, + 558, + 328 + ], + "lines": [ + { + "bbox": [ + 317, + 264, + 558, + 328 + ], + "spans": [ + { + "bbox": [ + 317, + 264, + 558, + 328 + ], + "type": "table", + "html": "
ModelExpRateExpRate@CDMBLEUCDM
Pix2tex0.12370.29100.40800.6360
Texify0.22880.49500.58900.7550
Matphix0.26100.50000.80670.9510
UniMERNet0.47990.81100.84250.9680
", + "image_path": "6359a83996b672aaa760963763cb4d4e27b70d1d7049cd5c1aa25a22ac1cd655.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 336, + 555, + 358 + ], + "lines": [ + { + "bbox": [ + 313, + 336, + 555, + 358 + ], + "spans": [ + { + "bbox": [ + 313, + 336, + 555, + 358 + ], + "type": "text", + "content": "Table 1. UniMER-Test evaluation results. Comparison of mainstream models using different metrics." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 362, + 555, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 362, + 555, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 362, + 555, + 399 + ], + "type": "text", + "content": "Figure 4b compares the number of cases where BLEU or CDM is preferred across different score ranges, showing that CDM consistently outperforms BLEU." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 403, + 473, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 473, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 473, + 415 + ], + "type": "text", + "content": "5.2.3. Objective Stability Assessment" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 418, + 555, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 555, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 555, + 574 + ], + "type": "text", + "content": "To evaluate the impact of formula writing styles on the CDM and BLEU metrics, we randomly select 50 formulas with LaTeX source code and rewrite each formula five times using GPT-4, generating 250 additional formulas. We manually verify these formulas to ensure their rendered results are identical to the original 50 formulas. Using the initial LaTeX source code as the ground truth, we analyze the score distribution of the BLEU and CDM metrics. As shown in Figure 4c, the CDM metric is unaffected by style changes, with all samples scoring 1. In contrast, the BLEU metric's scores are dispersed, making it unsuitable for formula evaluation. The CDM metric remains robust and reliable despite formatting changes." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 580, + 496, + 592 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 580, + 496, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 580, + 496, + 592 + ], + "type": "text", + "content": "5.3. Evaluation of Mainstream Models" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 597, + 554, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 597, + 554, + 670 + ], + "spans": [ + { + "bbox": [ + 313, + 597, + 554, + 670 + ], + "type": "text", + "content": "We conduct a detailed evaluation of mainstream models using both the CDM and BLEU metric. Note that all BLEU metric in this paper have been normalized [4, 8]. However, as discussed in the limitation section, normalization operations cannot address all issues, which will be evident in the following experiments." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 674, + 451, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 674, + 451, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 674, + 451, + 685 + ], + "type": "text", + "content": "5.3.1. UniMER-Test Evaluation" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "As shown in Table 1, the evaluation results of the four models on the UniMER-Test dataset indicate that the quality" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19687" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 116, + 62, + 496, + 148 + ], + "blocks": [ + { + "bbox": [ + 116, + 62, + 496, + 148 + ], + "lines": [ + { + "bbox": [ + 116, + 62, + 496, + 148 + ], + "spans": [ + { + "bbox": [ + 116, + 62, + 496, + 148 + ], + "type": "table", + "html": "
MethodSPECPEHWESCE
BLEU ↑CDM ↑BLEU ↑CDM ↑BLEU ↑CDM ↑BLEU ↑CDM ↑
Pix2tex [4]0.87300.96190.65500.64890.01200.24530.09200.6762
Texify [28]0.90600.98520.69000.70410.34100.52690.42000.7932
Mathpix0.79200.97290.80610.96710.80600.93180.81820.9238
UniMERNet [30]0.91700.99460.91600.97070.92100.95300.61600.9461
", + "image_path": "639a37c515982557d6ee702a836a2d0075b51ab004419d8796129e2ff8fd5c63.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 182, + 305, + 278 + ], + "blocks": [ + { + "bbox": [ + 80, + 152, + 529, + 163 + ], + "lines": [ + { + "bbox": [ + 80, + 152, + 529, + 163 + ], + "spans": [ + { + "bbox": [ + 80, + 152, + 529, + 163 + ], + "type": "text", + "content": "Table 2. UniMER-Test subset evaluation results. Analysis of anomalies in BLEU and CDM metrics across different subsets." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 182, + 305, + 278 + ], + "lines": [ + { + "bbox": [ + 58, + 182, + 305, + 278 + ], + "spans": [ + { + "bbox": [ + 58, + 182, + 305, + 278 + ], + "type": "table", + "html": "
Image TypeModelBLEUCDMExpRate@CDM
FormulaPix2tex0.46480.744400.3684
GPT-4o0.64310.73300.4324
UniMERNet0.60560.93960.6887
Mathpix0.61120.94800.2105
DocumentGPT-4o0.34110.65020.1670
Nougat0.58970.83260.6086
Mathpix0.59390.95670.6292
", + "image_path": "4470f1eb323c60d3e0a121c2ff66fd41fe85108446e59dce47aab77bb7559e4a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 286, + 297, + 321 + ], + "lines": [ + { + "bbox": [ + 55, + 286, + 297, + 321 + ], + "spans": [ + { + "bbox": [ + 55, + 286, + 297, + 321 + ], + "type": "text", + "content": "Table 3. Tiny-Doc-Math Evaluation Results. Performance of mainstream models on cropped formula inputs and document-level screenshots using BLEU and CDM metrics." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 347, + 296, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 347, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 347, + 296, + 407 + ], + "type": "text", + "content": "ranges from low to high as follows: Pix2Tex, Texify, Mathpix, and UniMERNet, based on both BLEU and CDM metrics. ExpRate@CDM clearly shows the proportion of completely correct predictions for each model, indicating that the text character-based ExpRate is unreliable." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 411, + 295, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 295, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 295, + 662 + ], + "type": "text", + "content": "From the results in Table 1, it appears that the trends of the BLEU and CDM metrics are consistent. To verify the reliability of using the BLEU metric for model comparison, we further present evaluation results on the UniMERTest subsets. As shown in Table 2, we observe two notable anomalies: Firstly, in the SCE subset, when comparing the quality of the Mathpix and UniMERNet models, the BLEU and CDM metrics provide opposite conclusions. A detailed review of the UniMERNet paper reveals that the SCE subset was annotated based on Mathpix and then manually corrected. This means that the expression style of the SCE formulas is more consistent with Mathpix. Consequently, even though the CDM metric indicates that UniMERNet has better actual model quality, the BLEU metric, influenced by the expression style, suggests that Mathpix is superior. Secondly, for the Pix2Tex model, the BLEU metric is very low on the HWE and SCE subsets but performs well on the SPE and CPE subsets. This discrepancy arises because the Pix2Tex training set includes a large number of printed formulas from arXiv and lacks data in the HWE and SCE styles." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "These anomalies clearly illustrate the limitations of the BLEU metric in evaluating the quality of formula recognition models. In contrast, the CDM metric proposed in this paper is fair and intuitive." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 185, + 459, + 196 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 185, + 459, + 196 + ], + "spans": [ + { + "bbox": [ + 313, + 185, + 459, + 196 + ], + "type": "text", + "content": "5.3.2. Tiny-Doc-Math Evaluation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 202, + 555, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 202, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 202, + 555, + 357 + ], + "type": "text", + "content": "The evaluation results of Tiny-Doc-Math are shown in Table 3. For cropped formula inputs (formula-level), all four models perform reasonably well, with CDM scores above 0.7. Notably, the current leading multimodal large model GPT-4o has the highest BLEU score among the four models but the lowest CDM score. This discrepancy indicates that the BLEU metric may not be reliable, suggesting that the formula recognition accuracy of GPT-4o still has room for improvement, lagging behind traditional SOTA models. Additionally, although Mathpix has the highest CDM score, only " + }, + { + "bbox": [ + 313, + 202, + 555, + 357 + ], + "type": "inline_equation", + "content": "21.05\\%" + }, + { + "bbox": [ + 313, + 202, + 555, + 357 + ], + "type": "text", + "content": " of the formulas are completely accurate. Manual verification revealed that many formulas are missing commas or periods at the end." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 361, + 556, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 556, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 556, + 552 + ], + "type": "text", + "content": "When the input is document-level screenshots, the models output the recognition results for the entire document (not just the formulas). Evaluation is conducted by matching the recognized block formulas. In this scenario, it can be observed that the accuracy of GPT-4o further decreases. In contrast, Mathpix and Nougat perform better, but even the document multimodal large model Nougat only achieves a CDM score of 0.8326. This indicates that there is still significant room for improvement in document-level recognition models. Mathpix remains the best performer, with a fully correct formula rate of " + }, + { + "bbox": [ + 313, + 361, + 556, + 552 + ], + "type": "inline_equation", + "content": "62.92\\%" + }, + { + "bbox": [ + 313, + 361, + 556, + 552 + ], + "type": "text", + "content": ". The accuracy of document-level recognition is crucial for advanced document understanding tasks like scientific knowledge Q&A, and CDM provides an excellent standard for selecting formula models and offers direction for improving formula recognition." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 571, + 388, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 571, + 388, + 583 + ], + "spans": [ + { + "bbox": [ + 313, + 571, + 388, + 583 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "In this paper, we introduced Character Detection Matching (CDM), a novel evaluation metric for formula recognition task. CDM addresses the shortcomings of the existing metrics in formula recognition by utilizing spatial character matching, overcoming the challenges posed by diverse formula representations. Comprehensive evaluations on different models and datasets demonstrate CDM's superiority in precisely reflecting recognition quality, showing more intuitive assessment and paving the way for future research and improvements in the formula recognition field." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19688" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "text", + "content": "7. Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 128 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 128 + ], + "type": "text", + "content": "This project was supported by National Key R&D Program of China (NO.2022ZD0160102) and Shanghai Artificial Intelligence Laboratory." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 115, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 115, + 152 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 115, + 152 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 159, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 159, + 296, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 159, + 296, + 203 + ], + "spans": [ + { + "bbox": [ + 61, + 159, + 296, + 203 + ], + "type": "text", + "content": "[1] Francisco Álvaro, Joan-Andreu Sánchez, and José-Miguel Benédí. An integrated grammar-based approach for mathematical expression recognition. Pattern Recognition, 51: 135–147, 2016." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 205, + 296, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 205, + 296, + 249 + ], + "spans": [ + { + "bbox": [ + 61, + 205, + 296, + 249 + ], + "type": "text", + "content": "[2] Ahmad-Montaser Awal, Harold Mouchere, and Christian Viard-Gaudin. A global learning approach for an online handwritten mathematical expression recognition system. Pattern Recognition Letters, 35:68-77, 2014." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 251, + 296, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 251, + 296, + 315 + ], + "spans": [ + { + "bbox": [ + 62, + 251, + 296, + 315 + ], + "type": "text", + "content": "[3] Xiaohang Bian, Bo Qin, Xiaozhe Xin, Jianwu Li, Xuefeng Su, and Yanfeng Wang. Handwritten mathematical expression recognition via attention aggregation based bidirectional mutual learning. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 113-121, 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 318, + 294, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 318, + 294, + 349 + ], + "spans": [ + { + "bbox": [ + 62, + 318, + 294, + 349 + ], + "type": "text", + "content": "[4] Lukas Blecher. pix2tex - latexOCR. https://github.com/lukas-blecher/LaTeX-OCR, 2022. Accessed: 2024-2-29." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 352, + 296, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 352, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 62, + 352, + 296, + 384 + ], + "type": "text", + "content": "[5] Lukas Blecher, Guillem Cucurull, Thomas Scialom, and Robert Stojnic. Nougat: Neural optical understanding for academic documents. arXiv.org, 2308.13418, 2023." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 387, + 296, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 387, + 296, + 441 + ], + "spans": [ + { + "bbox": [ + 62, + 387, + 296, + 441 + ], + "type": "text", + "content": "[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision (ECCV), pages 213-229. Springer, 2020." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 443, + 296, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 443, + 296, + 519 + ], + "spans": [ + { + "bbox": [ + 62, + 443, + 296, + 519 + ], + "type": "text", + "content": "[7] Hiuyi Cheng, Peirong Zhang, Sihang Wu, Jiaxin Zhang, Qiyuan Zhu, Zecheng Xie, Jing Li, Kai Ding, and Lianwen Jin. M6doc: A large-scale multi-format, multi-type, multi-Layout, multi-language, multi-annotation category dataset for modern document layout analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15138–15147, 2023." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 521, + 296, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 521, + 296, + 564 + ], + "spans": [ + { + "bbox": [ + 62, + 521, + 296, + 564 + ], + "type": "text", + "content": "[8] Yuntian Deng, Anssi Kanervisto, Jeffrey Ling, and Alexander M Rush. Image-to-markup generation with coarse-to-fine attention. In International Conference on Machine Learning (ICML), pages 980-989. PMLR, 2017." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 567, + 296, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 567, + 296, + 609 + ], + "spans": [ + { + "bbox": [ + 62, + 567, + 296, + 609 + ], + "type": "text", + "content": "[9] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 612, + 294, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 612, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 57, + 612, + 294, + 633 + ], + "type": "text", + "content": "[10] GPT-4o. Gpt-4o. https://openai.com/index/hello-gpt-4o/, 2024. Accessed: 2024-08-15." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 635, + 295, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 295, + 667 + ], + "type": "text", + "content": "[11] Mingxin Huang, Yuliang Liu, Dingkang Liang, Lianwen Jin, and Xiang Bai. Mini-monkey: Alleviate the sawtooth effect by multi-scale adaptive cropping. arXiv.org, 2024." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "text", + "content": "[12] Geewook Kim, Teakgyu Hong, Moonbin Yim, JeongYeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sangdoo Yun, Dongyoon Han, and Seunghyun Park. Ocr-free document understanding transformer. In European Confer" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "type": "text", + "content": "ence on Computer Vision (ECCV), pages 498-517. Springer, 2022." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 96, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 128 + ], + "type": "text", + "content": "[13] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 130, + 553, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 553, + 163 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 553, + 163 + ], + "type": "text", + "content": "[14] Vinay Kukreja et al. Recent trends in mathematical expressions recognition: An lda-based analysis. Expert Systems with Applications, 213:119028, 2023." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 165, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 165, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 165, + 553, + 196 + ], + "type": "text", + "content": "[15] Stéphane Lavirotte and Loïc Pottier. Mathematical formula recognition using graph grammar. In Document Recognition V, San Jose, CA, USA, January 24, 1998, 1998." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 198, + 553, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 241 + ], + "type": "text", + "content": "[16] Anh Duc Le, Bipin Indurkhya, and Masaki Nakagawa. Pattern generation strategies for improving recognition of handwritten mathematical expressions. Pattern Recognition Letters, 128:255-262, 2019." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 243, + 553, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 553, + 277 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 553, + 277 + ], + "type": "text", + "content": "[17] Vladimir I Levenshtein et al. Binary codes capable of correcting deletions, insertions, and reversals. In Soviet physics doklady, pages 707-710. Soviet Union, 1966." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 277, + 553, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 277, + 553, + 332 + ], + "spans": [ + { + "bbox": [ + 317, + 277, + 553, + 332 + ], + "type": "text", + "content": "[18] Bohan Li, Ye Yuan, Dingkang Liang, Xiao Liu, Zhilong Ji, Jinfeng Bai, Wenyu Liu, and Xiang Bai. When counting meets hmer: counting-aware network for handwritten mathematical expression recognition. In European Conference on Computer Vision (ECCV), pages 197-214. Springer, 2022." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 334, + 553, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 334, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 334, + 553, + 376 + ], + "type": "text", + "content": "[19] Minghao Li, Yiheng Xu, Lei Cui, Shaohan Huang, Furu Wei, Zhoujun Li, and Ming Zhou. Docbank: A benchmark dataset for document layout analysis. arXiv preprint arXiv:2006.01038, 2020." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 379, + 553, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 379, + 553, + 444 + ], + "spans": [ + { + "bbox": [ + 317, + 379, + 553, + 444 + ], + "type": "text", + "content": "[20] Minghui Liao, Guan Pang, Jing Huang, Tal Hassner, and Xiang Bai. Mask textspotter v3: Segmentation proposal network for robust scene text spotting. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XI 16, pages 706-722. Springer, 2020." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 446, + 553, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 446, + 553, + 500 + ], + "spans": [ + { + "bbox": [ + 317, + 446, + 553, + 500 + ], + "type": "text", + "content": "[21] Pengyuan Lyu, Yulin Li, Hao Zhou, Weihong Ma, Xingyu Wan, Qunyi Xie, Liang Wu, Chengquan Zhang, Kun Yao, Errui Ding, et al. Structextv3: An efficient vision-language model for text-rich image perception, comprehension, and beyond. arXiv.org, 2024." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 502, + 553, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 502, + 553, + 545 + ], + "spans": [ + { + "bbox": [ + 316, + 502, + 553, + 545 + ], + "type": "text", + "content": "[22] Scott MacLean and George Labahn. A new approach for recognizing handwritten mathematics using relational grammars and fuzzy sets. International Journal on Document Analysis and Recognition (IJDAR), 16:139-163, 2013." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 547, + 553, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 547, + 553, + 612 + ], + "spans": [ + { + "bbox": [ + 317, + 547, + 553, + 612 + ], + "type": "text", + "content": "[23] Mahshad Mahdavi, Richard Zanibbi, Harold Mouchere, Christian Viard-Gaudin, and Utpal Garain. Icdar 2019 crohme+ tfd: Competition on recognition of handwritten mathematical expressions and typeset formula detection. In International Conference on Document Analysis and Recognition (ICDAR), pages 1533-1538. IEEE, 2019." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 613, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 613, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 317, + 613, + 553, + 678 + ], + "type": "text", + "content": "[24] Harold Mouchere, Christian Viard-Gaudin, Richard Zanibbi, Utpal Garain, Dae Hwan Kim, and Jin Hyung Kim. Icdar 2013 crohme: Third international competition on recognition of online handwritten mathematical expressions. In International Conference on Document Analysis and Recognition (ICDAR), pages 1428-1432. IEEE, 2013." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[25] Harold Mouchere, Christian Viard-Gaudin, Richard Zanibbi, and Utpal Garain. Icfhr 2014 competition on recognition of on-line handwritten mathematical expressions (crohme" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19689" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "2014). In International Conference on Frontiers in Hand-writing Recognition (ICFHR), pages 791-796. IEEE, 2014." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 295, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 295, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 295, + 149 + ], + "type": "text", + "content": "[26] Harold Mouchère, Christian Viard-Gaudin, Richard Zanibbi, and Utpal Garain. Icfhr2016 crohme: Competition on recognition of online handwritten mathematical expressions. In International Conference on Frontiers in Handwriting Recognition (ICFHR), pages 607-612. IEEE, 2016." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 294, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 294, + 204 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 294, + 204 + ], + "type": "text", + "content": "[27] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 205, + 294, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 205, + 294, + 227 + ], + "spans": [ + { + "bbox": [ + 57, + 205, + 294, + 227 + ], + "type": "text", + "content": "[28] Vik Paruchuri. Texify. https://github.com/VikParuchuri/texify, 2023. Accessed: 2024-2-29." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 228, + 295, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 295, + 271 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 295, + 271 + ], + "type": "text", + "content": "[29] Josselin Somerville Roberts, Tony Lee, Chi Heem Wong, Michihiro Yasunaga, Yifan Mai, and Percy Liang. Image2struct: Benchmarking structure extraction for vision-language models. arXiv preprint arXiv:2410.22456, 2024." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 272, + 295, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 295, + 315 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 295, + 315 + ], + "type": "text", + "content": "[30] Bin Wang, Zhuangcheng Gu, Chao Xu, Bo Zhang, Botian Shi, and Conghui He. Unimernet: A universal network for real-world mathematical expression recognition. arXiv.org, 2024." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 316, + 295, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 295, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 295, + 370 + ], + "type": "text", + "content": "[31] Bin Wang, Chao Xu, Xiaomeng Zhao, Linke Ouyang, Fan Wu, Zhiyuan Zhao, Rui Xu, Kaiwen Liu, Yuan Qu, Fukai Shang, et al. Mineru: An open-source solution for precise document content extraction. arXiv preprint arXiv:2409.18839, 2024." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 371, + 295, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 295, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 295, + 415 + ], + "type": "text", + "content": "[32] Zelun Wang and Jyh-Charn Liu. Translating math formula images to latex sequences using deep neural networks with sequence-level training. International Journal on Document Analysis and Recognition (IJDAR), 24(1):63-75, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 416, + 295, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 416, + 295, + 449 + ], + "spans": [ + { + "bbox": [ + 56, + 416, + 295, + 449 + ], + "type": "text", + "content": "[33] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE TIP, 13(4):600-612, 2004." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 449, + 295, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 295, + 492 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 295, + 492 + ], + "type": "text", + "content": "[34] Haoran Wei, Lingyu Kong, Jinyue Chen, Liang Zhao, Zheng Ge, Jinrong Yang, Jianjian Sun, Chunrui Han, and Xiangyu Zhang. Vary: Scaling up the vision vocabulary for large vision-language models. arXiv.org, 2312.06109, 2023." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 493, + 295, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 493, + 295, + 548 + ], + "spans": [ + { + "bbox": [ + 56, + 493, + 295, + 548 + ], + "type": "text", + "content": "[35] Haoran Wei, Lingyu Kong, Jinyue Chen, Liang Zhao, Zheng Ge, Jinrong Yang, Jianjian Sun, Chunrui Han, and Xiangyu Zhang. Vary: Scaling up the vision vocabulary for large vision-language model. In European Conference on Computer Vision, pages 408-424. Springer, 2025." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 548, + 295, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 295, + 592 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 295, + 592 + ], + "type": "text", + "content": "[36] Renqiu Xia, Bo Zhang, Haoyang Peng, Ning Liao, Peng Ye, Botian Shi, Junchi Yan, and Yu Qiao. Structchart: Perception, structuring, reasoning for visual chart understanding. arXiv preprint arXiv:2309.11268, 2023." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 593, + 295, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 593, + 295, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 593, + 295, + 657 + ], + "type": "text", + "content": "[37] Renqiu Xia, Song Mao, Xiangchao Yan, Hongbin Zhou, Bo Zhang, Haoyang Peng, Jiahao Pi, Daocheng Fu, Wenjie Wu, Hancheng Ye, et al. Docgenome: An open largescale scientific document benchmark for training and testing multi-modal large language models. arXiv preprint arXiv:2406.11633, 2024." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "type": "text", + "content": "[38] Renqiu Xia, Bo Zhang, Hancheng Ye, Xiangchao Yan, Qi Liu, Hongbin Zhou, Zijun Chen, Min Dou, Botian Shi, Junchi Yan, et al. Chartx & chartvlm: A versatile benchmark and foundation model for complicated chart reasoning. arXiv preprint arXiv:2402.12185, 2024." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 554, + 330 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 316, + 72, + 554, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 554, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 554, + 128 + ], + "type": "text", + "content": "[39] Mingkun Yang, Yushuo Guan, Minghui Liao, Xin He, Kaigui Bian, Song Bai, Cong Yao, and Xiang Bai. Symmetry-constrained rectification network for scene text recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9147-9156, 2019." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 129, + 554, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 554, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 554, + 172 + ], + "type": "text", + "content": "[40] Ya-Qi Yu, Minghui Liao, Jiwen Zhang, and Jihao Wu. Texthawk2: A large vision-language model excels in bilingualOCR and grounding with 16x fewer tokens. arXiv preprint arXiv:2410.05261, 2024." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 174, + 554, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 554, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 554, + 228 + ], + "type": "text", + "content": "[41] Ye Yuan, Xiao Liu, Wondimu Dikubab, Hui Liu, Zhilong Ji, Zhongqin Wu, and Xiang Bai. Syntax-aware network for handwritten mathematical expression recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4553-4562, 2022." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 229, + 554, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 229, + 554, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 229, + 554, + 274 + ], + "type": "text", + "content": "[42] Jianshu Zhang, Jun Du, Yongxin Yang, Yi-Zhe Song, Si Wei, and Lirong Dai. A tree-structured decoder for image-to-markup generation. In International Conference on Machine Learning (ICML), pages 11076–11085. PMLR, 2020." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 274, + 554, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 274, + 554, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 274, + 554, + 330 + ], + "type": "text", + "content": "[43] Wenqi Zhao, Liangcai Gao, Zuoyu Yan, Shuai Peng, Lin Du, and Ziyin Zhang. Handwritten mathematical expression recognition with bidirectionally trained transformer. In International Conference on Document Analysis and Recognition (ICDAR), pages 570-584. Springer, 2021." + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19690" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file